989 files changed, 470949 insertions, 0 deletions

diff --git a/third_party/aom/.clang-format b/third_party/aom/.clang-format
new file mode 100644
index 0000000000..7837b7704d
--- /dev/null
+++ b/third_party/aom/.clang-format
@@ -0,0 +1,94 @@
+---
+Language:        Cpp
+# BasedOnStyle:  Google
+# Generated with clang-format 3.9.1
+AccessModifierOffset: -1
+AlignAfterOpenBracket: Align
+AlignConsecutiveAssignments: false
+AlignConsecutiveDeclarations: false
+AlignEscapedNewlinesLeft: true
+AlignOperands:   true
+AlignTrailingComments: true
+AllowAllParametersOfDeclarationOnNextLine: true
+AllowShortBlocksOnASingleLine: false
+AllowShortCaseLabelsOnASingleLine: true
+AllowShortFunctionsOnASingleLine: All
+AllowShortIfStatementsOnASingleLine: true
+AllowShortLoopsOnASingleLine: true
+AlwaysBreakAfterDefinitionReturnType: None
+AlwaysBreakAfterReturnType: None
+AlwaysBreakBeforeMultilineStrings: true
+AlwaysBreakTemplateDeclarations: true
+BinPackArguments: true
+BinPackParameters: true
+BraceWrapping:
+  AfterClass:      false
+  AfterControlStatement: false
+  AfterEnum:       false
+  AfterFunction:   false
+  AfterNamespace:  false
+  AfterObjCDeclaration: false
+  AfterStruct:     false
+  AfterUnion:      false
+  BeforeCatch:     false
+  BeforeElse:      false
+  IndentBraces:    false
+BreakBeforeBinaryOperators: None
+BreakBeforeBraces: Attach
+BreakBeforeTernaryOperators: true
+BreakConstructorInitializersBeforeComma: false
+BreakAfterJavaFieldAnnotations: false
+BreakStringLiterals: true
+ColumnLimit:     80
+CommentPragmas:  '^ IWYU pragma:'
+ConstructorInitializerAllOnOneLineOrOnePerLine: false
+ConstructorInitializerIndentWidth: 4
+ContinuationIndentWidth: 4
+Cpp11BracedListStyle: false
+DerivePointerAlignment: false
+DisableFormat:   false
+ExperimentalAutoDetectBinPacking: false
+ForEachMacros:   [ foreach, Q_FOREACH, BOOST_FOREACH ]
+IncludeCategories:
+  - Regex:           '^<.*\.h>'
+    Priority:        1
+  - Regex:           '^<.*'
+    Priority:        2
+  - Regex:           '.*'
+    Priority:        3
+IncludeIsMainRegex: '([-_](test|unittest))?$'
+IndentCaseLabels: true
+IndentWidth:     2
+IndentWrappedFunctionNames: false
+KeepEmptyLinesAtTheStartOfBlocks: false
+MacroBlockBegin: ''
+MacroBlockEnd:   ''
+MaxEmptyLinesToKeep: 1
+NamespaceIndentation: None
+ObjCBlockIndentWidth: 2
+ObjCSpaceAfterProperty: false
+ObjCSpaceBeforeProtocolList: false
+PenaltyBreakBeforeFirstCallParameter: 1
+PenaltyBreakComment: 300
+PenaltyBreakFirstLessLess: 120
+PenaltyBreakString: 1000
+PenaltyExcessCharacter: 1000000
+PenaltyReturnTypeOnItsOwnLine: 200
+PointerAlignment: Right
+ReflowComments:  true
+SortIncludes:    false
+SpaceAfterCStyleCast: false
+SpaceBeforeAssignmentOperators: true
+SpaceBeforeParens: ControlStatements
+SpaceInEmptyParentheses: false
+SpacesBeforeTrailingComments: 2
+SpacesInAngles:  false
+SpacesInContainerLiterals: true
+SpacesInCStyleCastParentheses: false
+SpacesInParentheses: false
+SpacesInSquareBrackets: false
+Standard:        Auto
+TabWidth:        8
+UseTab:          Never
+...
+
diff --git a/third_party/aom/.mailmap b/third_party/aom/.mailmap
new file mode 100644
index 0000000000..4672e5ccb8
--- /dev/null
+++ b/third_party/aom/.mailmap
@@ -0,0 +1,32 @@
+Adrian Grange <agrange@google.com>
+Aℓex Converse <aconverse@google.com>
+Aℓex Converse <aconverse@google.com> <alex.converse@gmail.com>
+Alexis Ballier <aballier@gentoo.org> <alexis.ballier@gmail.com>
+Alpha Lam <hclam@google.com> <hclam@chromium.org>
+Deb Mukherjee <debargha@google.com>
+Erik Niemeyer <erik.a.niemeyer@intel.com> <erik.a.niemeyer@gmail.com>
+Guillaume Martres <gmartres@google.com> <smarter3@gmail.com>
+Hangyu Kuang <hkuang@google.com>
+Hui Su <huisu@google.com>
+Jacky Chen <jackychen@google.com>
+Jim Bankoski <jimbankoski@google.com>
+Johann Koenig <johannkoenig@google.com>
+Johann Koenig <johannkoenig@google.com> <johann.koenig@duck.com>
+Johann Koenig <johannkoenig@google.com> <johann.koenig@gmail.com>
+John Koleszar <jkoleszar@google.com>
+Joshua Litt <joshualitt@google.com> <joshualitt@chromium.org>
+Marco Paniconi <marpan@google.com>
+Marco Paniconi <marpan@google.com> <marpan@chromium.org>
+Pascal Massimino <pascal.massimino@gmail.com>
+Paul Wilkins <paulwilkins@google.com>
+Ralph Giles <giles@xiph.org> <giles@entropywave.com>
+Ralph Giles <giles@xiph.org> <giles@mozilla.com>
+Ronald S. Bultje <rsbultje@gmail.com> <rbultje@google.com>
+Sami Pietilä <samipietila@google.com>
+Tamar Levy <tamar.levy@intel.com>
+Tamar Levy <tamar.levy@intel.com> <levytamar82@gmail.com>
+Tero Rintaluoma <teror@google.com> <tero.rintaluoma@on2.com>
+Timothy B. Terriberry <tterribe@xiph.org> Tim Terriberry <tterriberry@mozilla.com>
+Tom Finegan <tomfinegan@google.com>
+Tom Finegan <tomfinegan@google.com> <tomfinegan@chromium.org>
+Yaowu Xu <yaowu@google.com> <yaowu@xuyaowu.com>
diff --git a/third_party/aom/AUTHORS b/third_party/aom/AUTHORS
new file mode 100644
index 0000000000..95c3c8bf2b
--- /dev/null
+++ b/third_party/aom/AUTHORS
@@ -0,0 +1,144 @@
+# This file is automatically generated from the git commit history
+# by tools/gen_authors.sh.
+
+Aaron Watry <awatry@gmail.com>
+Abo Talib Mahfoodh <ab.mahfoodh@gmail.com>
+Adam Xu <adam@xuyaowu.com>
+Adrian Grange <agrange@google.com>
+Aℓex Converse <aconverse@google.com>
+Ahmad Sharif <asharif@google.com>
+Alexander Voronov <avoronov@graphics.cs.msu.ru>
+Alexis Ballier <aballier@gentoo.org>
+Alok Ahuja <waveletcoeff@gmail.com>
+Alpha Lam <hclam@google.com>
+A.Mahfoodh <ab.mahfoodh@gmail.com>
+Ami Fischman <fischman@chromium.org>
+Andoni Morales Alastruey <ylatuya@gmail.com>
+Andres Mejia <mcitadel@gmail.com>
+Andrew Russell <anrussell@google.com>
+Angie Chiang <angiebird@google.com>
+Aron Rosenberg <arosenberg@logitech.com>
+Attila Nagy <attilanagy@google.com>
+Brion Vibber <bvibber@wikimedia.org>
+changjun.yang <changjun.yang@intel.com>
+Charles 'Buck' Krasic <ckrasic@google.com>
+chm <chm@rock-chips.com>
+Christian Duvivier <cduvivier@google.com>
+Daniel Kang <ddkang@google.com>
+Deb Mukherjee <debargha@google.com>
+Dim Temp <dimtemp0@gmail.com>
+Dmitry Kovalev <dkovalev@google.com>
+Dragan Mrdjan <dmrdjan@mips.com>
+Ed Baker <edward.baker@intel.com>
+Ehsan Akhgari <ehsan.akhgari@gmail.com>
+Erik Niemeyer <erik.a.niemeyer@intel.com>
+Fabio Pedretti <fabio.ped@libero.it>
+Frank Galligan <fgalligan@google.com>
+Fredrik Söderquist <fs@opera.com>
+Fritz Koenig <frkoenig@google.com>
+Gaute Strokkenes <gaute.strokkenes@broadcom.com>
+Geza Lore <gezalore@gmail.com>
+Ghislain MARY <ghislainmary2@gmail.com>
+Giuseppe Scrivano <gscrivano@gnu.org>
+Gordana Cmiljanovic <gordana.cmiljanovic@imgtec.com>
+Guillaume Martres <gmartres@google.com>
+Guillermo Ballester Valor <gbvalor@gmail.com>
+Hangyu Kuang <hkuang@google.com>
+Hanno Böck <hanno@hboeck.de>
+Henrik Lundin <hlundin@google.com>
+Hui Su <huisu@google.com>
+Ivan Maltz <ivanmaltz@google.com>
+Jacek Caban <cjacek@gmail.com>
+Jacky Chen <jackychen@google.com>
+James Berry <jamesberry@google.com>
+James Yu <james.yu@linaro.org>
+James Zern <jzern@google.com>
+Jan Gerber <j@mailb.org>
+Jan Kratochvil <jan.kratochvil@redhat.com>
+Janne Salonen <jsalonen@google.com>
+Jean-Marc Valin <jmvalin@jmvalin.ca>
+Jeff Faust <jfaust@google.com>
+Jeff Muizelaar <jmuizelaar@mozilla.com>
+Jeff Petkau <jpet@chromium.org>
+Jia Jia <jia.jia@linaro.org>
+Jian Zhou <zhoujian@google.com>
+Jim Bankoski <jimbankoski@google.com>
+Jingning Han <jingning@google.com>
+Joey Parrish <joeyparrish@google.com>
+Johann Koenig <johannkoenig@chromium.org>
+Johann Koenig <johannkoenig@google.com>
+John Koleszar <jkoleszar@google.com>
+Johnny Klonaris <google@jawknee.com>
+John Stark <jhnstrk@gmail.com>
+Joshua Bleecher Snyder <josh@treelinelabs.com>
+Joshua Litt <joshualitt@google.com>
+Julia Robson <juliamrobson@gmail.com>
+Justin Clift <justin@salasaga.org>
+Justin Lebar <justin.lebar@gmail.com>
+KO Myung-Hun <komh@chollian.net>
+Lawrence Velázquez <larryv@macports.org>
+Lou Quillio <louquillio@google.com>
+Luca Barbato <lu_zero@gentoo.org>
+Makoto Kato <makoto.kt@gmail.com>
+Mans Rullgard <mans@mansr.com>
+Marco Paniconi <marpan@google.com>
+Mark Mentovai <mark@chromium.org>
+Martin Ettl <ettl.martin78@googlemail.com>
+Martin Storsjo <martin@martin.st>
+Matthew Heaney <matthewjheaney@chromium.org>
+Michael Kohler <michaelkohler@live.com>
+Mike Frysinger <vapier@chromium.org>
+Mike Hommey <mhommey@mozilla.com>
+Mikhal Shemer <mikhal@google.com>
+Minghai Shang <minghai@google.com>
+Morton Jonuschat <yabawock@gmail.com>
+Nathan E. Egge <negge@dgql.org>
+Nico Weber <thakis@chromium.org>
+Parag Salasakar <img.mips1@gmail.com>
+Pascal Massimino <pascal.massimino@gmail.com>
+Patrik Westin <patrik.westin@gmail.com>
+Paul Wilkins <paulwilkins@google.com>
+Pavol Rusnak <stick@gk2.sk>
+Paweł Hajdan <phajdan@google.com>
+Pengchong Jin <pengchong@google.com>
+Peter de Rivaz <peter.derivaz@argondesign.com>
+Peter de Rivaz <peter.derivaz@gmail.com>
+Philip Jägenstedt <philipj@opera.com>
+Priit Laes <plaes@plaes.org>
+Rafael Ávila de Espíndola <rafael.espindola@gmail.com>
+Rafaël Carré <funman@videolan.org>
+Ralph Giles <giles@xiph.org>
+Rob Bradford <rob@linux.intel.com>
+Ronald S. Bultje <rsbultje@gmail.com>
+Rui Ueyama <ruiu@google.com>
+Sami Pietilä <samipietila@google.com>
+Sasi Inguva <isasi@google.com>
+Scott Graham <scottmg@chromium.org>
+Scott LaVarnway <slavarnway@google.com>
+Sean McGovern <gseanmcg@gmail.com>
+Sergey Kolomenkin <kolomenkin@gmail.com>
+Sergey Ulanov <sergeyu@chromium.org>
+Shimon Doodkin <helpmepro1@gmail.com>
+Shunyao Li <shunyaoli@google.com>
+Stefan Holmer <holmer@google.com>
+Steinar Midtskogen <stemidts@cisco.com>
+Suman Sunkara <sunkaras@google.com>
+Taekhyun Kim <takim@nvidia.com>
+Takanori MATSUURA <t.matsuu@gmail.com>
+Tamar Levy <tamar.levy@intel.com>
+Tao Bai <michaelbai@chromium.org>
+Tero Rintaluoma <teror@google.com>
+Thijs Vermeir <thijsvermeir@gmail.com>
+Thomas Daede <tdaede@mozilla.com>
+Thomas Davies <thdavies@cisco.com>
+Thomas <thdavies@cisco.com>
+Tim Kopp <tkopp@google.com>
+Timothy B. Terriberry <tterribe@xiph.org>
+Tom Finegan <tomfinegan@google.com>
+Tristan Matthews <le.businessman@gmail.com>
+Tristan Matthews <tmatth@videolan.org>
+Vignesh Venkatasubramanian <vigneshv@google.com>
+Yaowu Xu <yaowu@google.com>
+Yongzhe Wang <yongzhe@google.com>
+Yunqing Wang <yunqingwang@google.com>
+Zoe Liu <zoeliu@google.com>
diff --git a/third_party/aom/CHANGELOG b/third_party/aom/CHANGELOG
new file mode 100644
index 0000000000..dad0ea1d29
--- /dev/null
+++ b/third_party/aom/CHANGELOG
@@ -0,0 +1,630 @@
+Next Release
+  - Incompatible changes:
+    The AV1 encoder's default keyframe interval changed to 128 from 9999.
+
+2016-04-07 v0.1.0 "AOMedia Codec 1"
+  This release is the first Alliance for Open Media codec.
+2015-11-09 v1.5.0 "Javan Whistling Duck"
+  This release improves upon the VP9 encoder and speeds up the encoding and
+  decoding processes.
+
+  - Upgrading:
+    This release is ABI incompatible with 1.4.0. It drops deprecated VP8
+    controls and adds a variety of VP9 controls for testing.
+
+    The vpxenc utility now prefers VP9 by default.
+
+  - Enhancements:
+    Faster VP9 encoding and decoding
+    Smaller library size by combining functions used by VP8 and VP9
+
+  - Bug Fixes:
+    A variety of fuzzing issues
+
+2015-04-03 v1.4.0 "Indian Runner Duck"
+  This release includes significant improvements to the VP9 codec.
+
+  - Upgrading:
+    This release is ABI incompatible with 1.3.0. It drops the compatibility
+    layer, requiring VPX_IMG_FMT_* instead of IMG_FMT_*, and adds several codec
+    controls for VP9.
+
+  - Enhancements:
+    Faster VP9 encoding and decoding
+    Multithreaded VP9 decoding (tile and frame-based)
+    Multithreaded VP9 encoding - on by default
+    YUV 4:2:2 and 4:4:4 support in VP9
+    10 and 12bit support in VP9
+    64bit ARM support by replacing ARM assembly with intrinsics
+
+  - Bug Fixes:
+    Fixes a VP9 bitstream issue in Profile 1. This only affected non-YUV 4:2:0
+    files.
+
+  - Known Issues:
+    Frame Parallel decoding fails for segmented and non-420 files.
+
+2013-11-15 v1.3.0 "Forest"
+  This release introduces the VP9 codec in a backward-compatible way.
+  All existing users of VP8 can continue to use the library without
+  modification. However, some VP8 options do not map to VP9 in the same manner.
+
+  The VP9 encoder in this release is not feature complete. Users interested in
+  the encoder are advised to use the git master branch and discuss issues on
+  libvpx mailing lists.
+
+  - Upgrading:
+    This release is ABI and API compatible with Duclair (v1.0.0). Users
+    of older releases should refer to the Upgrading notes in this document
+    for that release.
+
+  - Enhancements:
+      Get rid of bashisms in the main build scripts
+      Added usage info on command line options
+      Add lossless compression mode
+      Dll build of libvpx
+      Add additional Mac OS X targets: 10.7, 10.8 and 10.9 (darwin11-13)
+      Add option to disable documentation
+      configure: add --enable-external-build support
+      make: support V=1 as short form of verbose=yes
+      configure: support mingw-w64
+      configure: support hardfloat armv7 CHOSTS
+      configure: add support for android x86
+      Add estimated completion time to vpxenc
+      Don't exit on decode errors in vpxenc
+      vpxenc: support scaling prior to encoding
+      vpxdec: support scaling output
+      vpxenc: improve progress indicators with --skip
+      msvs: Don't link to winmm.lib
+      Add a new script for producing vcxproj files
+      Produce Visual Studio 10 and 11 project files
+      Produce Windows Phone project files
+      msvs-build: use msbuild for vs >= 2005
+      configure: default configure log to config.log
+      Add encoding option --static-thresh
+
+  - Speed:
+      Miscellaneous speed optimizations for VP8 and VP9.
+
+  - Quality:
+      In general, quality is consistent with the Eider release.
+
+  - Bug Fixes:
+      This release represents approximately a year of engineering effort,
+      and contains multiple bug fixes. Please refer to git history for details.
+
+
+2012-12-21 v1.2.0
+  This release acts as a checkpoint for a large amount of internal refactoring
+  and testing. It also contains a number of small bugfixes, so all users are
+  encouraged to upgrade.
+
+  - Upgrading:
+    This release is ABI and API compatible with Duclair (v1.0.0). Users
+    of older releases should refer to the Upgrading notes in this
+    document for that release.
+
+  - Enhancements:
+      VP8 optimizations for MIPS dspr2
+      vpxenc: add -quiet option
+
+  - Speed:
+      Encoder and decoder speed is consistent with the Eider release.
+
+  - Quality:
+      In general, quality is consistent with the Eider release.
+
+      Minor tweaks to ARNR filtering
+      Minor improvements to real time encoding with multiple temporal layers
+
+  - Bug Fixes:
+      Fixes multithreaded encoder race condition in loopfilter
+      Fixes multi-resolution threaded encoding
+      Fix potential encoder dead-lock after picture resize
+
+
+2012-05-09 v1.1.0 "Eider"
+  This introduces a number of enhancements, mostly focused on real-time
+  encoding. In addition, it fixes a decoder bug (first introduced in
+  Duclair) so all users of that release are encouraged to upgrade.
+
+  - Upgrading:
+    This release is ABI and API compatible with Duclair (v1.0.0). Users
+    of older releases should refer to the Upgrading notes in this
+    document for that release.
+
+    This release introduces a new temporal denoiser, controlled by the
+    VP8E_SET_NOISE_SENSITIVITY control. The temporal denoiser does not
+    currently take a strength parameter, so the control is effectively
+    a boolean - zero (off) or non-zero (on). For compatibility with
+    existing applications, the values accepted are the same as those
+    for the spatial denoiser (0-6). The temporal denoiser is enabled
+    by default, and the older spatial denoiser may be restored by
+    configuring with --disable-temporal-denoising. The temporal denoiser
+    is more computationally intensive than the spatial one.
+
+    This release removes support for a legacy, decode only API that was
+    supported, but deprecated, at the initial release of libvpx
+    (v0.9.0). This is not expected to have any impact. If you are
+    impacted, you can apply a reversion to commit 2bf8fb58 locally.
+    Please update to the latest libvpx API if you are affected.
+
+  - Enhancements:
+      Adds a motion compensated temporal denoiser to the encoder, which
+      gives higher quality than the older spatial denoiser. (See above
+      for notes on upgrading).
+
+      In addition, support for new compilers and platforms were added,
+      including:
+        improved support for XCode
+        Android x86 NDK build
+        OS/2 support
+        SunCC support
+
+      Changing resolution with vpx_codec_enc_config_set() is now
+      supported. Previously, reinitializing the codec was required to
+      change the input resolution.
+
+      The vpxenc application has initial support for producing multiple
+      encodes from the same input in one call. Resizing is not yet
+      supported, but varying other codec parameters is. Use -- to
+      delineate output streams. Options persist from one stream to the
+      next.
+
+      Also, the vpxenc application will now use a keyframe interval of
+      5 seconds by default. Use the --kf-max-dist option to override.
+
+  - Speed:
+      Decoder performance improved 2.5% versus Duclair. Encoder speed is
+      consistent with Duclair for most material. Two pass encoding of
+      slideshow-like material will see significant improvements.
+
+      Large realtime encoding speed gains at a small quality expense are
+      possible by configuring the on-the-fly bitpacking experiment with
+      --enable-onthefly-bitpacking. Realtime encoder can be up to 13%
+      faster (ARM) depending on the number of threads and bitrate
+      settings. This technique sees constant gain over the 5-16 speed
+      range. For VC style input the loss seen is up to 0.2dB. See commit
+      52cf4dca for further details.
+
+  - Quality:
+      On the whole, quality is consistent with the Duclair release. Some
+      tweaks:
+
+        Reduced blockiness in easy sections by applying a penalty to
+        intra modes.
+
+        Improved quality of static sections (like slideshows) with
+        two pass encoding.
+
+        Improved keyframe sizing with multiple temporal layers
+
+  - Bug Fixes:
+      Corrected alt-ref contribution to frame rate for visible updates
+      to the alt-ref buffer. This affected applications making manual
+      usage of the frame reference flags, or temporal layers.
+
+      Additional constraints were added to disable multi-frame quality
+      enhancement (MFQE) in sections of the frame where there is motion.
+      (#392)
+
+      Fixed corruption issues when vpx_codec_enc_config_set() was called
+      with spatial resampling enabled.
+
+      Fixed a decoder error introduced in Duclair where the segmentation
+      map was not being reinitialized on keyframes (#378)
+
+
+2012-01-27 v1.0.0 "Duclair"
+  Our fourth named release, focused on performance and features related to
+  real-time encoding. It also fixes a decoder crash bug introduced in
+  v0.9.7, so all users of that release are encouraged to upgrade.
+
+  - Upgrading:
+      This release is ABI incompatible with prior releases of libvpx, so the
+      "major" version number has been bumped to 1. You must recompile your
+      applications against the latest version of the libvpx headers. The
+      API remains compatible, and this should not require code changes in most
+      applications.
+
+  - Enhancements:
+      This release introduces several substantial new features to the encoder,
+      of particular interest to real time streaming applications.
+
+      Temporal scalability allows the encoder to produce a stream that can
+      be decimated to different frame rates, with independent rate targetting
+      for each substream.
+
+      Multiframe quality enhancement postprocessing can make visual quality
+      more consistent in the presence of frames that are substantially
+      different quality than the surrounding frames, as in the temporal
+      scalability case and in some forced keyframe scenarios.
+
+      Multiple-resolution encoding support allows the encoding of the
+      same content at different resolutions faster than encoding them
+      separately.
+
+  - Speed:
+      Optimization targets for this release included the decoder and the real-
+      time modes of the encoder. Decoder speed on x86 has improved 10.5% with
+      this release. Encoder improvements followed a curve where speeds 1-3
+      improved 4.0%-1.5%, speeds 4-8 improved <1%, and speeds 9-16 improved
+      1.5% to 10.5%, respectively. "Best" mode speed is consistent with the
+      Cayuga release.
+
+  - Quality:
+      Encoder quality in the single stream case is consistent with the Cayuga
+      release.
+
+  - Bug Fixes:
+      This release fixes an OOB read decoder crash bug present in v0.9.7
+      related to the clamping of motion vectors in SPLITMV blocks. This
+      behavior could be triggered by corrupt input or by starting
+      decoding from a P-frame.
+
+
+2011-08-15 v0.9.7-p1 "Cayuga" patch 1
+  This is an incremental bugfix release against Cayuga. All users of that
+  release are strongly encouraged to upgrade.
+
+    - Fix potential OOB reads (cdae03a)
+
+          An unbounded out of bounds read was discovered when the
+          decoder was requested to perform error concealment (new in
+          Cayuga) given a frame with corrupt partition sizes.
+
+          A bounded out of bounds read was discovered affecting all
+          versions of libvpx. Given an multipartition input frame that
+          is truncated between the mode/mv partition and the first
+          residiual paritition (in the block of partition offsets), up
+          to 3 extra bytes could have been read from the source buffer.
+          The code will not take any action regardless of the contents
+          of these undefined bytes, as the truncated buffer is detected
+          immediately following the read based on the calculated
+          starting position of the coefficient partition.
+
+    - Fix potential error concealment crash when the very first frame
+      is missing or corrupt (a609be5)
+
+    - Fix significant artifacts in error concealment (a4c2211, 99d870a)
+
+    - Revert 1-pass CBR rate control changes (e961317)
+      Further testing showed this change produced undesirable visual
+      artifacts, rolling back for now.
+
+
+2011-08-02 v0.9.7 "Cayuga"
+  Our third named release, focused on a faster, higher quality, encoder.
+
+  - Upgrading:
+    This release is backwards compatible with Aylesbury (v0.9.5) and
+    Bali (v0.9.6). Users of older releases should refer to the Upgrading
+    notes in this document for that release.
+
+  - Enhancements:
+          Stereo 3D format support for vpxenc
+          Runtime detection of available processor cores.
+          Allow specifying --end-usage by enum name
+          vpxdec: test for frame corruption
+          vpxenc: add quantizer histogram display
+          vpxenc: add rate histogram display
+          Set VPX_FRAME_IS_DROPPABLE
+          update configure for ios sdk 4.3
+          Avoid text relocations in ARM vp8 decoder
+          Generate a vpx.pc file for pkg-config.
+          New ways of passing encoded data between encoder and decoder.
+
+  - Speed:
+      This release includes across-the-board speed improvements to the
+      encoder. On x86, these measure at approximately 11.5% in Best mode,
+      21.5% in Good mode (speed 0), and 22.5% in Realtime mode (speed 6).
+      On ARM Cortex A9 with Neon extensions, real-time encoding of video
+      telephony content is 35% faster than Bali on single core and 48%
+      faster on multi-core. On the NVidia Tegra2 platform, real time
+      encoding is 40% faster than Bali.
+
+      Decoder speed was not a priority for this release, but improved
+      approximately 8.4% on x86.
+
+          Reduce motion vector search on alt-ref frame.
+          Encoder loopfilter running in its own thread
+          Reworked loopfilter to precalculate more parameters
+          SSE2/SSSE3 optimizations for build_predictors_mbuv{,_s}().
+          Make hor UV predict ~2x faster (73 vs 132 cycles) using SSSE3.
+          Removed redundant checks
+          Reduced structure sizes
+          utilize preload in ARMv6 MC/LPF/Copy routines
+          ARM optimized quantization, dfct, variance, subtract
+          Increase chrow row alignment to 16 bytes.
+          disable trellis optimization for first pass
+          Write SSSE3 sub-pixel filter function
+          Improve SSE2 half-pixel filter funtions
+          Add vp8_sub_pixel_variance16x8_ssse3 function
+          Reduce unnecessary distortion computation
+          Use diamond search to replace full search
+          Preload reference area in sub-pixel motion search (real-time mode)
+
+  - Quality:
+      This release focused primarily on one-pass use cases, including
+      video conferencing. Low latency data rate control was significantly
+      improved, improving streamability over bandwidth constrained links.
+      Added support for error concealment, allowing frames to maintain
+      visual quality in the presence of substantial packet loss.
+
+          Add rc_max_intra_bitrate_pct control
+          Limit size of initial keyframe in one-pass.
+          Improve framerate adaptation
+          Improved 1-pass CBR rate control
+          Improved KF insertion after fades to still.
+          Improved key frame detection.
+          Improved activity masking (lower PSNR impact for same SSIM boost)
+          Improved interaction between GF and ARFs
+          Adding error-concealment to the decoder.
+          Adding support for independent partitions
+          Adjusted rate-distortion constants
+
+
+  - Bug Fixes:
+          Removed firstpass motion map
+          Fix parallel make install
+          Fix multithreaded encoding for 1 MB wide frame
+          Fixed iwalsh_neon build problems with RVDS4.1
+          Fix semaphore emulation, spin-wait intrinsics on Windows
+          Fix build with xcode4 and simplify GLOBAL.
+          Mark ARM asm objects as allowing a non-executable stack.
+          Fix vpxenc encoding incorrect webm file header on big endian
+
+
+2011-03-07 v0.9.6 "Bali"
+  Our second named release, focused on a faster, higher quality, encoder.
+
+  - Upgrading:
+    This release is backwards compatible with Aylesbury (v0.9.5). Users
+    of older releases should refer to the Upgrading notes in this
+    document for that release.
+
+  - Enhancements:
+      vpxenc --psnr shows a summary when encode completes
+      --tune=ssim option to enable activity masking
+      improved postproc visualizations for development
+      updated support for Apple iOS to SDK 4.2
+      query decoder to determine which reference frames were updated
+      implemented error tracking in the decoder
+      fix pipe support on windows
+
+  - Speed:
+      Primary focus was on good quality mode, speed 0. Average improvement
+      on x86 about 40%, up to 100% on user-generated content at that speed.
+      Best quality mode speed improved 35%, and realtime speed 10-20%. This
+      release also saw significant improvement in realtime encoding speed
+      on ARM platforms.
+
+        Improved encoder threading
+        Dont pick encoder filter level when loopfilter is disabled.
+        Avoid double copying of key frames into alt and golden buffer
+        FDCT optimizations.
+        x86 sse2 temporal filter
+        SSSE3 version of fast quantizer
+        vp8_rd_pick_best_mbsegmentation code restructure
+        Adjusted breakout RD for SPLITMV
+        Changed segmentation check order
+        Improved rd_pick_intra4x4block
+        Adds armv6 optimized variance calculation
+        ARMv6 optimized sad16x16
+        ARMv6 optimized half pixel variance calculations
+        Full search SAD function optimization in SSE4.1
+        Improve MV prediction accuracy to achieve performance gain
+        Improve MV prediction in vp8_pick_inter_mode() for speed>3
+
+  - Quality:
+      Best quality mode improved PSNR 6.3%, and SSIM 6.1%. This release
+      also includes support for "activity masking," which greatly improves
+      SSIM at the expense of PSNR. For now, this feature is available with
+      the --tune=ssim option. Further experimentation in this area
+      is ongoing. This release also introduces a new rate control mode
+      called "CQ," which changes the allocation of bits within a clip to
+      the sections where they will have the most visual impact.
+
+        Tuning for the more exact quantizer.
+        Relax rate control for last few frames
+        CQ Mode
+        Limit key frame quantizer for forced key frames.
+        KF/GF Pulsing
+        Add simple version of activity masking.
+        make rdmult adaptive for intra in quantizer RDO
+        cap the best quantizer for 2nd order DC
+        change the threshold of DC check for encode breakout
+
+  - Bug Fixes:
+      Fix crash on Sparc Solaris.
+      Fix counter of fixed keyframe distance
+      ARNR filter pointer update bug fix
+      Fixed use of motion percentage in KF/GF group calc
+      Changed condition for using RD in Intra Mode
+      Fix encoder real-time only configuration.
+      Fix ARM encoder crash with multiple token partitions
+      Fixed bug first cluster timecode of webm file is wrong.
+      Fixed various encoder bugs with odd-sized images
+      vp8e_get_preview fixed when spatial resampling enabled
+      quantizer: fix assertion in fast quantizer path
+      Allocate source buffers to be multiples of 16
+      Fix for manual Golden frame frequency
+      Fix drastic undershoot in long form content
+
+
+2010-10-28 v0.9.5 "Aylesbury"
+  Our first named release, focused on a faster decoder, and a better encoder.
+
+  - Upgrading:
+    This release incorporates backwards-incompatible changes to the
+    ivfenc and ivfdec tools. These tools are now called vpxenc and vpxdec.
+
+    vpxdec
+      * the -q (quiet) option has been removed, and replaced with
+        -v (verbose). the output is quiet by default. Use -v to see
+        the version number of the binary.
+
+      * The default behavior is now to write output to a single file
+        instead of individual frames. The -y option has been removed.
+        Y4M output is the default.
+
+      * For raw I420/YV12 output instead of Y4M, the --i420 or --yv12
+        options must be specified.
+
+          $ ivfdec -o OUTPUT INPUT
+          $ vpxdec --i420 -o OUTPUT INPUT
+
+      * If an output file is not specified, the default is to write
+        Y4M to stdout. This makes piping more natural.
+
+          $ ivfdec -y -o - INPUT | ...
+          $ vpxdec INPUT | ...
+
+      * The output file has additional flexibility for formatting the
+        filename. It supports escape characters for constructing a
+        filename from the width, height, and sequence number. This
+        replaces the -p option. To get the equivalent:
+
+          $ ivfdec -p frame INPUT
+          $ vpxdec --i420 -o frame-%wx%h-%4.i420 INPUT
+
+    vpxenc
+      * The output file must be specified with -o, rather than as the
+        last argument.
+
+          $ ivfenc <options> INPUT OUTPUT
+          $ vpxenc <options> -o OUTPUT INPUT
+
+      * The output defaults to webm. To get IVF output, use the --ivf
+        option.
+
+          $ ivfenc <options> INPUT OUTPUT.ivf
+          $ vpxenc <options> -o OUTPUT.ivf --ivf INPUT
+
+
+  - Enhancements:
+      ivfenc and ivfdec have been renamed to vpxenc, vpxdec.
+      vpxdec supports .webm input
+      vpxdec writes .y4m by default
+      vpxenc writes .webm output by default
+      vpxenc --psnr now shows the average/overall PSNR at the end
+      ARM platforms now support runtime cpu detection
+      vpxdec visualizations added for motion vectors, block modes, references
+      vpxdec now silent by default
+      vpxdec --progress shows frame-by-frame timing information
+      vpxenc supports the distinction between --fps and --timebase
+      NASM is now a supported assembler
+      configure: enable PIC for shared libs by default
+      configure: add --enable-small
+      configure: support for ppc32-linux-gcc
+      configure: support for sparc-solaris-gcc
+
+  - Bugs:
+      Improve handling of invalid frames
+      Fix valgrind errors in the NEON loop filters.
+      Fix loopfilter delta zero transitions
+      Fix valgrind errors in vp8_sixtap_predict8x4_armv6().
+      Build fixes for darwin-icc
+
+  - Speed:
+      20-40% (average 28%) improvement in libvpx decoder speed,
+      including:
+        Rewrite vp8_short_walsh4x4_sse2()
+        Optimizations on the loopfilters.
+        Miscellaneous improvements for Atom
+        Add 4-tap version of 2nd-pass ARMv6 MC filter.
+        Improved multithread utilization
+        Better instruction choices on x86
+        reorder data to use wider instructions
+        Update NEON wide idcts
+        Make block access to frame buffer sequential
+        Improved subset block search
+        Bilinear subpixel optimizations for ssse3.
+        Decrease memory footprint
+
+      Encoder speed improvements (percentage gain not measured):
+        Skip unnecessary search of identical frames
+        Add SSE2 subtract functions
+        Improve bounds checking in vp8_diamond_search_sadx4()
+        Added vp8_fast_quantize_b_sse2
+
+  - Quality:
+      Over 7% overall PSNR improvement (6.3% SSIM) in "best" quality
+      encoding mode, and up to 60% improvement on very noisy, still
+      or slow moving source video
+
+        Motion compensated temporal filter for Alt-Ref Noise Reduction
+        Improved use of trellis quantization on 2nd order Y blocks
+        Tune effect of motion on KF/GF boost in two pass
+        Allow coefficient optimization for good quality speed 0.
+        Improved control of active min quantizer for two pass.
+        Enable ARFs for non-lagged compress
+
+2010-09-02 v0.9.2
+  - Enhancements:
+      Disable frame dropping by default
+      Improved multithreaded performance
+      Improved Force Key Frame Behaviour
+      Increased rate control buffer level precision
+      Fix bug in 1st pass motion compensation
+      ivfenc: correct fixed kf interval, --disable-kf
+  - Speed:
+      Changed above and left context data layout
+      Rework idct calling structure.
+      Removed unnecessary MB_MODE_INFO copies
+      x86: SSSE3 sixtap prediction
+      Reworked IDCT to include reconstruction (add) step
+      Swap alt/gold/new/last frame buffer ptrs instead of copying.
+      Improve SSE2 loopfilter functions
+      Change bitreader to use a larger window.
+      Avoid loopfilter reinitialization when possible
+  - Quality:
+      Normalize quantizer's zero bin and rounding factors
+      Add trellis quantization.
+      Make the quantizer exact.
+      Updates to ARNR filtering algorithm
+      Fix breakout thresh computation for golden & AltRef frames
+      Redo the forward 4x4 dct
+      Improve the accuracy of forward walsh-hadamard transform
+      Further adjustment of RD behaviour with Q and Zbin.
+  - Build System:
+      Allow linking of libs built with MinGW to MSVC
+      Fix target auto-detection on mingw32
+      Allow --cpu= to work for x86.
+      configure: pass original arguments through to make dist
+      Fix builds without runtime CPU detection
+      msvs: fix install of codec sources
+      msvs: Change devenv.com command line for better msys support
+      msvs: Add vs9 targets.
+      Add x86_64-linux-icc target
+  - Bugs:
+      Potential crashes on older MinGW builds
+      Fix two-pass framrate for Y4M input.
+      Fixed simple loop filter, other crashes on ARM v6
+      arm: fix missing dependency with --enable-shared
+      configure: support directories containing .o
+      Replace pinsrw (SSE) with MMX instructions
+      apple: include proper mach primatives
+      Fixed rate control bug with long key frame interval.
+      Fix DSO link errors on x86-64 when not using a version script
+      Fixed buffer selection for UV in AltRef filtering
+
+
+2010-06-17 v0.9.1
+  - Enhancements:
+      * ivfenc/ivfdec now support YUV4MPEG2 input and pipe I/O
+      * Speed optimizations
+  - Bugfixes:
+      * Rate control
+      * Prevent out-of-bounds accesses on invalid data
+  - Build system updates:
+      * Detect toolchain to be used automatically for native builds
+      * Support building shared libraries
+      * Better autotools emulation (--prefix, --libdir, DESTDIR)
+  - Updated LICENSE
+      * http://webmproject.blogspot.com/2010/06/changes-to-webm-open-source-license.html
+
+
+2010-05-18 v0.9.0
+  - Initial open source release. Welcome to WebM and VP8!
+
diff --git a/third_party/aom/CMakeLists.txt b/third_party/aom/CMakeLists.txt
new file mode 100644
index 0000000000..4ba2d40aaf
--- /dev/null
+++ b/third_party/aom/CMakeLists.txt
@@ -0,0 +1,359 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+cmake_minimum_required(VERSION 3.5)
+project(AOM C CXX)
+
+set(AOM_ROOT "${CMAKE_CURRENT_SOURCE_DIR}")
+set(AOM_CONFIG_DIR "${CMAKE_CURRENT_BINARY_DIR}")
+
+if ("${AOM_ROOT}" STREQUAL "${AOM_CONFIG_DIR}")
+  message(FATAL_ERROR
+          "Building from within the aom source tree is not supported.\n"
+          "Hint: mkdir -p ../aom_build && cd ../aom_build\n"
+          "Run cmake from there.")
+endif ()
+
+include("${AOM_ROOT}/build/cmake/aom_configure.cmake")
+include("${AOM_ROOT}/build/cmake/aom_optimization.cmake")
+include("${AOM_ROOT}/aom_dsp/aom_dsp.cmake")
+include("${AOM_ROOT}/aom_mem/aom_mem.cmake")
+include("${AOM_ROOT}/aom_ports/aom_ports.cmake")
+include("${AOM_ROOT}/aom_scale/aom_scale.cmake")
+include("${AOM_ROOT}/aom_util/aom_util.cmake")
+include("${AOM_ROOT}/av1/av1.cmake")
+include("${AOM_ROOT}/test/test.cmake")
+
+set(AOM_RTCD_SOURCES
+    "${AOM_CONFIG_DIR}/aom_dsp_rtcd.h"
+    "${AOM_CONFIG_DIR}/aom_scale_rtcd.h"
+    "${AOM_CONFIG_DIR}/av1_rtcd.h"
+    "${AOM_ROOT}/aom_dsp/aom_dsp_rtcd_defs.pl"
+    "${AOM_ROOT}/aom_dsp/aom_dsp_rtcd.c"
+    "${AOM_ROOT}/aom_scale/aom_scale_rtcd.pl"
+    "${AOM_ROOT}/aom_scale/aom_scale_rtcd.c"
+    "${AOM_ROOT}/av1/common/av1_rtcd_defs.pl"
+    "${AOM_ROOT}/av1/common/av1_rtcd.c"
+    "${AOM_ROOT}/build/make/rtcd.pl")
+
+# TODO(tomfinegan): Use libwebm's cmake support directly.
+set(AOM_LIBWEBM_SOURCES
+    "${AOM_ROOT}/third_party/libwebm/common/hdr_util.cc"
+    "${AOM_ROOT}/third_party/libwebm/common/hdr_util.h"
+    "${AOM_ROOT}/third_party/libwebm/common/webmids.h"
+    "${AOM_ROOT}/third_party/libwebm/mkvmuxer/mkvmuxer.cc"
+    "${AOM_ROOT}/third_party/libwebm/mkvmuxer/mkvmuxer.h"
+    "${AOM_ROOT}/third_party/libwebm/mkvmuxer/mkvmuxertypes.h"
+    "${AOM_ROOT}/third_party/libwebm/mkvmuxer/mkvmuxerutil.cc"
+    "${AOM_ROOT}/third_party/libwebm/mkvmuxer/mkvmuxerutil.h"
+    "${AOM_ROOT}/third_party/libwebm/mkvmuxer/mkvwriter.cc"
+    "${AOM_ROOT}/third_party/libwebm/mkvmuxer/mkvwriter.h"
+    "${AOM_ROOT}/third_party/libwebm/mkvparser/mkvparser.cc"
+    "${AOM_ROOT}/third_party/libwebm/mkvparser/mkvparser.h"
+    "${AOM_ROOT}/third_party/libwebm/mkvparser/mkvreader.cc"
+    "${AOM_ROOT}/third_party/libwebm/mkvparser/mkvreader.h")
+
+set(AOM_LIBYUV_SOURCES
+    "${AOM_ROOT}/third_party/libyuv/include/libyuv/basic_types.h"
+    "${AOM_ROOT}/third_party/libyuv/include/libyuv/convert.h"
+    "${AOM_ROOT}/third_party/libyuv/include/libyuv/convert_argb.h"
+    "${AOM_ROOT}/third_party/libyuv/include/libyuv/convert_from.h"
+    "${AOM_ROOT}/third_party/libyuv/include/libyuv/cpu_id.h"
+    "${AOM_ROOT}/third_party/libyuv/include/libyuv/planar_functions.h"
+    "${AOM_ROOT}/third_party/libyuv/include/libyuv/rotate.h"
+    "${AOM_ROOT}/third_party/libyuv/include/libyuv/row.h"
+    "${AOM_ROOT}/third_party/libyuv/include/libyuv/scale.h"
+    "${AOM_ROOT}/third_party/libyuv/include/libyuv/scale_row.h"
+    "${AOM_ROOT}/third_party/libyuv/source/cpu_id.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/planar_functions.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/row_any.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/row_common.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/row_gcc.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/row_mips.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/row_neon.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/row_neon64.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/row_win.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/scale.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/scale_any.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/scale_common.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/scale_gcc.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/scale_mips.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/scale_neon.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/scale_neon64.cc"
+    "${AOM_ROOT}/third_party/libyuv/source/scale_win.cc")
+
+set(AOM_SOURCES
+    "${AOM_CONFIG_DIR}/aom_config.c"
+    "${AOM_CONFIG_DIR}/aom_config.h"
+    "${AOM_ROOT}/aom/aom.h"
+    "${AOM_ROOT}/aom/aom_codec.h"
+    "${AOM_ROOT}/aom/aom_decoder.h"
+    "${AOM_ROOT}/aom/aom_encoder.h"
+    "${AOM_ROOT}/aom/aom_frame_buffer.h"
+    "${AOM_ROOT}/aom/aom_image.h"
+    "${AOM_ROOT}/aom/aom_integer.h"
+    "${AOM_ROOT}/aom/aomcx.h"
+    "${AOM_ROOT}/aom/aomdx.h"
+    "${AOM_ROOT}/aom/internal/aom_codec_internal.h"
+    "${AOM_ROOT}/aom/src/aom_codec.c"
+    "${AOM_ROOT}/aom/src/aom_decoder.c"
+    "${AOM_ROOT}/aom/src/aom_encoder.c"
+    "${AOM_ROOT}/aom/src/aom_image.c")
+
+set(AOM_COMMON_APP_UTIL_SOURCES
+    "${AOM_ROOT}/args.c"
+    "${AOM_ROOT}/args.h"
+    "${AOM_ROOT}/md5_utils.c"
+    "${AOM_ROOT}/md5_utils.h"
+    "${AOM_ROOT}/tools_common.c"
+    "${AOM_ROOT}/tools_common.h"
+    "${AOM_ROOT}/video_common.h")
+
+set(AOM_DECODER_APP_UTIL_SOURCES
+    "${AOM_ROOT}/ivfdec.c"
+    "${AOM_ROOT}/ivfdec.h"
+    "${AOM_ROOT}/video_reader.c"
+    "${AOM_ROOT}/video_reader.h"
+    "${AOM_ROOT}/y4menc.c"
+    "${AOM_ROOT}/y4menc.h")
+
+set(AOM_ENCODER_APP_UTIL_SOURCES
+    "${AOM_ROOT}/ivfenc.c"
+    "${AOM_ROOT}/ivfenc.h"
+    "${AOM_ROOT}/video_writer.c"
+    "${AOM_ROOT}/video_writer.h"
+    "${AOM_ROOT}/warnings.c"
+    "${AOM_ROOT}/warnings.h"
+    "${AOM_ROOT}/y4minput.c"
+    "${AOM_ROOT}/y4minput.h"
+    "${AOM_ROOT}/examples/encoder_util.h"
+    "${AOM_ROOT}/examples/encoder_util.c")
+
+set(AOM_ENCODER_STATS_SOURCES
+    "${AOM_ROOT}/aomstats.c"
+    "${AOM_ROOT}/aomstats.h"
+    "${AOM_ROOT}/rate_hist.c"
+    "${AOM_ROOT}/rate_hist.h")
+
+set(AOM_WEBM_DECODER_SOURCES
+    "${AOM_ROOT}/webmdec.cc"
+    "${AOM_ROOT}/webmdec.h")
+
+set(AOM_WEBM_ENCODER_SOURCES
+    "${AOM_ROOT}/webmenc.cc"
+    "${AOM_ROOT}/webmenc.h")
+
+include_directories(${AOM_ROOT} ${AOM_CONFIG_DIR})
+
+# Targets
+# TODO(tomfinegan): Move rtcd target setup where it belongs for each rtcd
+# source.
+add_rtcd_build_step("${AOM_ROOT}/aom_dsp/aom_dsp_rtcd_defs.pl"
+                    "${AOM_CONFIG_DIR}/aom_dsp_rtcd.h"
+                    "${AOM_ROOT}/aom_dsp/aom_dsp_rtcd.c"
+                    "aom_dsp_rtcd")
+add_rtcd_build_step("${AOM_ROOT}/aom_scale/aom_scale_rtcd.pl"
+                    "${AOM_CONFIG_DIR}/aom_scale_rtcd.h"
+                    "${AOM_ROOT}/aom_scale/aom_scale_rtcd.c"
+                    "aom_scale_rtcd")
+add_rtcd_build_step("${AOM_ROOT}/av1/common/av1_rtcd_defs.pl"
+                    "${AOM_CONFIG_DIR}/av1_rtcd.h"
+                    "${AOM_ROOT}/av1/common/av1_rtcd.c"
+                    "av1_rtcd")
+
+add_library(aom_rtcd OBJECT ${AOM_RTCD_SOURCES})
+add_library(aom_encoder_stats OBJECT ${AOM_ENCODER_STATS_SOURCES})
+add_library(aom ${AOM_SOURCES} $<TARGET_OBJECTS:aom_rtcd>)
+
+# List of object and static library targets.
+set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_rtcd aom_encoder_stats aom_mem
+    aom_scale aom)
+
+# Setup dependencies.
+setup_aom_dsp_targets()
+setup_aom_mem_targets()
+setup_aom_ports_targets()
+setup_aom_util_targets()
+setup_aom_scale_targets()
+setup_av1_targets()
+
+# Make all library targets depend on aom_rtcd to make sure it builds first.
+foreach (aom_lib ${AOM_LIB_TARGETS})
+  if (NOT "${aom_lib}" STREQUAL "aom_rtcd")
+    add_dependencies(${aom_lib} aom_rtcd)
+  endif ()
+endforeach ()
+#
+# Application and application support targets.
+#
+add_library(aom_common_app_util OBJECT ${AOM_COMMON_APP_UTIL_SOURCES})
+add_library(aom_decoder_app_util OBJECT ${AOM_DECODER_APP_UTIL_SOURCES})
+add_library(aom_encoder_app_util OBJECT ${AOM_ENCODER_APP_UTIL_SOURCES})
+
+add_executable(aom_cx_set_ref
+               "${AOM_ROOT}/examples/aom_cx_set_ref.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_encoder_app_util>)
+
+add_executable(aomdec
+               "${AOM_ROOT}/aomdec.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_decoder_app_util>)
+
+add_executable(aomenc
+               "${AOM_ROOT}/aomenc.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_encoder_app_util>
+               $<TARGET_OBJECTS:aom_encoder_stats>)
+
+if (CONFIG_ANALYZER)
+  add_executable(analyzer
+                 "${AOM_ROOT}/examples/analyzer.cc"
+                 $<TARGET_OBJECTS:aom_common_app_util>
+                 $<TARGET_OBJECTS:aom_decoder_app_util>)
+  target_link_libraries(analyzer PUBLIC ${wxWidgets_LIBRARIES})
+  set(AOM_APP_TARGETS ${AOM_APP_TARGETS} analyzer)
+endif ()
+
+add_executable(decode_to_md5
+               "${AOM_ROOT}/examples/decode_to_md5.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_decoder_app_util>)
+
+add_executable(decode_with_drops
+               "${AOM_ROOT}/examples/decode_with_drops.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_decoder_app_util>)
+
+add_executable(lossless_encoder
+               "${AOM_ROOT}/examples/lossless_encoder.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_encoder_app_util>)
+
+add_executable(set_maps
+               "${AOM_ROOT}/examples/set_maps.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_encoder_app_util>)
+
+add_executable(simple_decoder
+               "${AOM_ROOT}/examples/simple_decoder.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_decoder_app_util>)
+
+if (CONFIG_INSPECTION)
+add_executable(inspect
+               "${AOM_ROOT}/examples/inspect.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_decoder_app_util>)
+set(AOM_APP_TARGETS ${AOM_APP_TARGETS} inspect)
+endif ()
+
+add_executable(simple_encoder
+               "${AOM_ROOT}/examples/simple_encoder.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_encoder_app_util>)
+
+add_executable(twopass_encoder
+               "${AOM_ROOT}/examples/twopass_encoder.c"
+               $<TARGET_OBJECTS:aom_common_app_util>
+               $<TARGET_OBJECTS:aom_encoder_app_util>)
+
+# List of app targets.
+set(AOM_APP_TARGETS ${AOM_APP_TARGETS} aom_cx_set_ref aomdec aomenc
+    decode_to_md5 decode_with_drops lossless_encoder set_maps simple_decoder
+    simple_encoder twopass_encoder)
+
+foreach (aom_app ${AOM_APP_TARGETS})
+  target_link_libraries(${aom_app} PUBLIC aom)
+endforeach ()
+
+if (CONFIG_LIBYUV)
+  add_library(yuv OBJECT ${AOM_LIBYUV_SOURCES})
+  if (NOT MSVC)
+    target_compile_options(yuv PRIVATE -Wno-unused-parameter)
+  endif ()
+  include_directories("${AOM_ROOT}/third_party/libyuv/include")
+
+  # Add to existing targets.
+  foreach (aom_app ${AOM_APP_TARGETS})
+    target_sources(${aom_app} PUBLIC $<TARGET_OBJECTS:yuv>)
+    set_property(TARGET ${aom_app} PROPERTY LINKER_LANGUAGE CXX)
+  endforeach ()
+endif ()
+
+if (CONFIG_WEBM_IO)
+  add_library(webm OBJECT ${AOM_LIBWEBM_SOURCES})
+  include_directories("${AOM_ROOT}/third_party/libwebm")
+
+  if (NOT MSVC)
+    target_compile_options(webm PRIVATE -Wno-shadow)
+  endif ()
+
+  # Add to existing targets.
+  target_sources(aom_decoder_app_util PUBLIC ${AOM_WEBM_DECODER_SOURCES})
+  target_sources(aom_encoder_app_util PUBLIC ${AOM_WEBM_ENCODER_SOURCES})
+
+  foreach (aom_app ${AOM_APP_TARGETS})
+    target_sources(${aom_app} PUBLIC $<TARGET_OBJECTS:webm>)
+    set_property(TARGET ${aom_app} PROPERTY LINKER_LANGUAGE CXX)
+   endforeach ()
+endif ()
+
+if (CONFIG_UNIT_TESTS)
+  # Create test_libaom target and the targets it depends on.
+  setup_aom_test_targets()
+endif ()
+
+if (HAVE_PTHREAD_H AND CONFIG_MULTITHREAD)
+  find_package(Threads)
+  foreach (app_target ${AOM_APP_TARGETS})
+    target_link_libraries(${app_target} PUBLIC Threads::Threads)
+  endforeach ()
+endif ()
+
+if (XCODE)
+  if (CONFIG_LIBYUV OR CONFIG_WEBM_IO)
+    # The Xcode generator does not obey LINKER_LANGUAGE. Because of the issue
+    # what looks like a C++ file needs to be in any target that Xcode will link
+    # when the target contains a C++ dependency.
+    # Without this Xcode will try to link with the C linker, which always ends
+    # badly when a dependency actually includes C++.
+    # Note: LINKER_LANGUAGE is explicitly set to C++ for all targets touched
+    # here, it really is the Xcode generator's fault, or just a deficiency in
+    # Xcode itself.
+    set(XCODE_DUMMY_CXX_FILE "${AOM_CONFIG_DIR}/dummy.cc")
+    file(WRITE "${XCODE_DUMMY_CXX_FILE}"
+         "// Xcode needs a C++ file to link, ignore this file.")
+    foreach (aom_app ${AOM_APP_TARGETS})
+      target_sources(${aom_app} PUBLIC "${XCODE_DUMMY_CXX_FILE}")
+    endforeach ()
+  endif ()
+endif ()
+
+# Aomedia install rule.
+# TODO(tomfinegan): Add the missing pkg-config related bits.
+set(AOM_INSTALL_INCS
+    "${AOM_ROOT}/aom/aom.h"
+    "${AOM_ROOT}/aom/aomcx.h"
+    "${AOM_ROOT}/aom/aomdx.h"
+    "${AOM_ROOT}/aom/aom_codec.h"
+    "${AOM_ROOT}/aom/aom_frame_buffer.h"
+    "${AOM_ROOT}/aom/aom_image.h"
+    "${AOM_ROOT}/aom/aom_integer.h"
+    "${AOM_ROOT}/aom/aom_decoder.h"
+    "${AOM_ROOT}/aom/aom_encoder.h"
+    "${AOM_ROOT}/aom/aom.h")
+set(AOM_INSTALL_BINS aomdec aomenc)
+set(AOM_INSTALL_LIBS aom)
+
+install(FILES ${AOM_INSTALL_INCS}
+        DESTINATION "${CMAKE_INSTALL_PREFIX}/include/aom")
+install(TARGETS ${AOM_INSTALL_LIBS} DESTINATION "${CMAKE_INSTALL_PREFIX}/lib")
+install(TARGETS ${AOM_INSTALL_BINS} DESTINATION "${CMAKE_INSTALL_PREFIX}/bin")
diff --git a/third_party/aom/LICENSE b/third_party/aom/LICENSE
new file mode 100644
index 0000000000..fc340c3764
--- /dev/null
+++ b/third_party/aom/LICENSE
@@ -0,0 +1,27 @@
+Copyright (c) 2016, Alliance for Open Media. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+
+1. Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in
+   the documentation and/or other materials provided with the
+   distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
+LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
+ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+
diff --git a/third_party/aom/PATENTS b/third_party/aom/PATENTS
new file mode 100644
index 0000000000..97842e02f1
--- /dev/null
+++ b/third_party/aom/PATENTS
@@ -0,0 +1,108 @@
+Alliance for Open Media Patent License 1.0
+
+1. License Terms.
+
+1.1. Patent License. Subject to the terms and conditions of this License, each
+     Licensor, on behalf of itself and successors in interest and assigns,
+     grants Licensee a non-sublicensable, perpetual, worldwide, non-exclusive,
+     no-charge, royalty-free, irrevocable (except as expressly stated in this
+     License) patent license to its Necessary Claims to make, use, sell, offer
+     for sale, import or distribute any Implementation.
+
+1.2. Conditions.
+
+1.2.1. Availability. As a condition to the grant of rights to Licensee to make,
+       sell, offer for sale, import or distribute an Implementation under
+       Section 1.1, Licensee must make its Necessary Claims available under
+       this License, and must reproduce this License with any Implementation
+       as follows:
+
+       a. For distribution in source code, by including this License in the
+          root directory of the source code with its Implementation.
+
+       b. For distribution in any other form (including binary, object form,
+          and/or hardware description code (e.g., HDL, RTL, Gate Level Netlist,
+          GDSII, etc.)), by including this License in the documentation, legal
+          notices, and/or other written materials provided with the
+          Implementation.
+
+1.2.2. Additional Conditions. This license is directly from Licensor to
+       Licensee.  Licensee acknowledges as a condition of benefiting from it
+       that no rights from Licensor are received from suppliers, distributors,
+       or otherwise in connection with this License.
+
+1.3. Defensive Termination. If any Licensee, its Affiliates, or its agents
+     initiates patent litigation or files, maintains, or voluntarily
+     participates in a lawsuit against another entity or any person asserting
+     that any Implementation infringes Necessary Claims, any patent licenses
+     granted under this License directly to the Licensee are immediately
+     terminated as of the date of the initiation of action unless 1) that suit
+     was in response to a corresponding suit regarding an Implementation first
+     brought against an initiating entity, or 2) that suit was brought to
+     enforce the terms of this License (including intervention in a third-party
+     action by a Licensee).
+
+1.4. Disclaimers. The Reference Implementation and Specification are provided
+     "AS IS" and without warranty. The entire risk as to implementing or
+     otherwise using the Reference Implementation or Specification is assumed
+     by the implementer and user. Licensor expressly disclaims any warranties
+     (express, implied, or otherwise), including implied warranties of
+     merchantability, non-infringement, fitness for a particular purpose, or
+     title, related to the material. IN NO EVENT WILL LICENSOR BE LIABLE TO
+     ANY OTHER PARTY FOR LOST PROFITS OR ANY FORM OF INDIRECT, SPECIAL,
+     INCIDENTAL, OR CONSEQUENTIAL DAMAGES OF ANY CHARACTER FROM ANY CAUSES OF
+     ACTION OF ANY KIND WITH RESPECT TO THIS LICENSE, WHETHER BASED ON BREACH
+     OF CONTRACT, TORT (INCLUDING NEGLIGENCE), OR OTHERWISE, AND WHETHER OR
+     NOT THE OTHER PARTRY HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+2. Definitions.
+
+2.1. Affiliate.  �Affiliate� means an entity that directly or indirectly
+     Controls, is Controlled by, or is under common Control of that party.
+
+2.2. Control. �Control� means direct or indirect control of more than 50% of
+     the voting power to elect directors of that corporation, or for any other
+     entity, the power to direct management of such entity.
+
+2.3. Decoder.  "Decoder" means any decoder that conforms fully with all
+     non-optional portions of the Specification.
+
+2.4. Encoder.  "Encoder" means any encoder that produces a bitstream that can
+     be decoded by a Decoder only to the extent it produces such a bitstream.
+
+2.5. Final Deliverable.  �Final Deliverable� means the final version of a
+     deliverable approved by the Alliance for Open Media as a Final
+     Deliverable.
+
+2.6. Implementation.  "Implementation" means any implementation, including the
+     Reference Implementation, that is an Encoder and/or a Decoder. An
+     Implementation also includes components of an Implementation only to the
+     extent they are used as part of an Implementation.
+
+2.7. License. �License� means this license.
+
+2.8. Licensee. �Licensee� means any person or entity who exercises patent
+     rights granted under this License.
+
+2.9. Licensor.  "Licensor" means (i) any Licensee that makes, sells, offers
+     for sale, imports or distributes any Implementation, or (ii) a person
+     or entity that has a licensing obligation to the Implementation as a
+     result of its membership and/or participation in the Alliance for Open
+     Media working group that developed the Specification.
+
+2.10. Necessary Claims.  "Necessary Claims" means all claims of patents or
+      patent applications, (a) that currently or at any time in the future,
+      are owned or controlled by the Licensor, and (b) (i) would be an
+      Essential Claim as defined by the W3C Policy as of February 5, 2004
+      (https://www.w3.org/Consortium/Patent-Policy-20040205/#def-essential)
+      as if the Specification was a W3C Recommendation; or (ii) are infringed
+      by the Reference Implementation.
+
+2.11. Reference Implementation. �Reference Implementation� means an Encoder
+      and/or Decoder released by the Alliance for Open Media as a Final
+      Deliverable.
+
+2.12. Specification. �Specification� means the specification designated by
+      the Alliance for Open Media as a Final Deliverable for which this
+      License was issued.
+
diff --git a/third_party/aom/README b/third_party/aom/README
new file mode 100644
index 0000000000..9aa30daa13
--- /dev/null
+++ b/third_party/aom/README
@@ -0,0 +1,168 @@
+README - 9 March 2017
+
+Welcome to the AV1 Codec SDK!
+
+COMPILING THE APPLICATIONS/LIBRARIES:
+  The build system used is similar to autotools. Building generally consists of
+  "configuring" with your desired build options, then using GNU make to build
+  the application.
+
+  1. Prerequisites
+
+    * All x86 targets require the Yasm[1] assembler be installed.
+    * All Windows builds require that Cygwin[2] be installed.
+    * Building the documentation requires Doxygen[3]. If you do not
+      have this package, the install-docs option will be disabled.
+    * Downloading the data for the unit tests requires curl[4] and sha1sum.
+      sha1sum is provided via the GNU coreutils, installed by default on
+      many *nix platforms, as well as MinGW and Cygwin. If coreutils is not
+      available, a compatible version of sha1sum can be built from
+      source[5]. These requirements are optional if not running the unit
+      tests.
+
+    [1]: http://www.tortall.net/projects/yasm
+    [2]: http://www.cygwin.com
+    [3]: http://www.doxygen.org
+    [4]: http://curl.haxx.se
+    [5]: http://www.microbrew.org/tools/md5sha1sum/
+
+  2. Out-of-tree builds
+  Out of tree builds are a supported method of building the application. For
+  an out of tree build, the source tree is kept separate from the object
+  files produced during compilation. For instance:
+
+    $ mkdir build
+    $ cd build
+    $ ../libaom/configure <options>
+    $ make
+
+  3. Configuration options
+  The 'configure' script supports a number of options. The --help option can be
+  used to get a list of supported options:
+    $ ../libaom/configure --help
+
+  4. Cross development
+  For cross development, the most notable option is the --target option. The
+  most up-to-date list of supported targets can be found at the bottom of the
+  --help output of the configure script. As of this writing, the list of
+  available targets is:
+
+    armv6-linux-rvct
+    armv6-linux-gcc
+    armv6-none-rvct
+    arm64-darwin-gcc
+    armv7-android-gcc
+    armv7-darwin-gcc
+    armv7-linux-rvct
+    armv7-linux-gcc
+    armv7-none-rvct
+    armv7-win32-vs12
+    armv7-win32-vs14
+    armv7s-darwin-gcc
+    mips32-linux-gcc
+    mips64-linux-gcc
+    sparc-solaris-gcc
+    x86-android-gcc
+    x86-darwin8-gcc
+    x86-darwin8-icc
+    x86-darwin9-gcc
+    x86-darwin9-icc
+    x86-darwin10-gcc
+    x86-darwin11-gcc
+    x86-darwin12-gcc
+    x86-darwin13-gcc
+    x86-darwin14-gcc
+    x86-darwin15-gcc
+    x86-darwin16-gcc
+    x86-iphonesimulator-gcc
+    x86-linux-gcc
+    x86-linux-icc
+    x86-os2-gcc
+    x86-solaris-gcc
+    x86-win32-gcc
+    x86-win32-vs12
+    x86-win32-vs14
+    x86_64-android-gcc
+    x86_64-darwin9-gcc
+    x86_64-darwin10-gcc
+    x86_64-darwin11-gcc
+    x86_64-darwin12-gcc
+    x86_64-darwin13-gcc
+    x86_64-darwin14-gcc
+    x86_64-darwin15-gcc
+    x86_64-darwin16-gcc
+    x86_64-iphonesimulator-gcc
+    x86_64-linux-gcc
+    x86_64-linux-icc
+    x86_64-solaris-gcc
+    x86_64-win64-gcc
+    x86_64-win64-vs12
+    x86_64-win64-vs14
+    generic-gnu
+
+  The generic-gnu target, in conjunction with the CROSS environment variable,
+  can be used to cross compile architectures that aren't explicitly listed, if
+  the toolchain is a cross GNU (gcc/binutils) toolchain. Other POSIX toolchains
+  will likely work as well. For instance, to build using the mipsel-linux-uclibc
+  toolchain, the following command could be used (note, POSIX SH syntax, adapt
+  to your shell as necessary):
+
+    $ CROSS=mipsel-linux-uclibc- ../libaom/configure
+
+  In addition, the executables to be invoked can be overridden by specifying the
+  environment variables: CC, AR, LD, AS, STRIP, NM. Additional flags can be
+  passed to these executables with CFLAGS, LDFLAGS, and ASFLAGS.
+
+  5. Configuration errors
+  If the configuration step fails, the first step is to look in the error log.
+  This defaults to config.log. This should give a good indication of what went
+  wrong. If not, contact us for support.
+
+AV1 TEST VECTORS:
+  The test vectors can be downloaded and verified using the build system after
+  running configure. To specify an alternate directory the
+  LIBAOM_TEST_DATA_PATH environment variable can be used.
+
+  $ ./configure --enable-unit-tests
+  $ LIBAOM_TEST_DATA_PATH=../-test-data make testdata
+
+UNIT TESTS:
+  The unit tests (consisting mainly of the test_libaom binary) can be run using
+  make. This will download the test data if necessary.
+
+  $ ../libaom/configure --enable-unit-tests
+  $ make test
+
+  Test may be run in parallel using make -j which supports up to 10 shards by
+  default.
+  $ make -j10 test
+
+  If you have additional cores you can scale the tests to match:
+  $ shards=$(nproc); \
+    make -j$shards test \
+    NUM_SHARDS=$shards SHARDS="$(seq -s' ' 0 $(( shards - 1 )))" \
+    && echo "success"
+
+  The GTEST_FILTER environment variable (equivalent to --gtest_filter) can be
+  used to control which tests are run while sharding:
+  $ GTEST_FILTER='SSE2*' make -j10 test
+
+CODE STYLE:
+  The coding style used by this project is enforced with clang-format using the
+  configuration contained in the .clang-format file in the root of the
+  repository.
+
+  Before pushing changes for review you can format your code with:
+  # Apply clang-format to modified .c, .h and .cc files
+  $ clang-format -i --style=file \
+    $(git diff --name-only --diff-filter=ACMR '*.[hc]' '*.cc')
+
+  Check the .clang-format file for the version used to generate it if there is
+  any difference between your local formatting and the review system.
+
+  See also: http://clang.llvm.org/docs/ClangFormat.html
+
+SUPPORT
+  This library is an open source project supported by its community. Please
+  please email webm-discuss@webmproject.org for help.
+
diff --git a/third_party/aom/aom/aom.h b/third_party/aom/aom/aom.h
new file mode 100644
index 0000000000..98366b8709
--- /dev/null
+++ b/third_party/aom/aom/aom.h
@@ -0,0 +1,160 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/*!\defgroup aom AOM
+ * \ingroup codecs
+ * AOM is aom's newest video compression algorithm that uses motion
+ * compensated prediction, Discrete Cosine Transform (DCT) coding of the
+ * prediction error signal and context dependent entropy coding techniques
+ * based on arithmetic principles. It features:
+ *  - YUV 4:2:0 image format
+ *  - Macro-block based coding (16x16 luma plus two 8x8 chroma)
+ *  - 1/4 (1/8) pixel accuracy motion compensated prediction
+ *  - 4x4 DCT transform
+ *  - 128 level linear quantizer
+ *  - In loop deblocking filter
+ *  - Context-based entropy coding
+ *
+ * @{
+ */
+/*!\file
+ * \brief Provides controls common to both the AOM encoder and decoder.
+ */
+#ifndef AOM_AOM_H_
+#define AOM_AOM_H_
+
+#include "./aom_codec.h"
+#include "./aom_image.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*!\brief Control functions
+ *
+ * The set of macros define the control functions of AOM interface
+ */
+enum aom_com_control_id {
+  /*!\brief pass in an external frame into decoder to be used as reference frame
+   */
+  AOM_SET_REFERENCE = 1,
+  AOM_COPY_REFERENCE = 2, /**< get a copy of reference frame from the decoder */
+  AOM_SET_POSTPROC = 3,   /**< set the decoder's post processing settings  */
+  AOM_SET_DBG_COLOR_REF_FRAME =
+      4, /**< set the reference frames to color for each macroblock */
+  AOM_SET_DBG_COLOR_MB_MODES = 5, /**< set which macro block modes to color */
+  AOM_SET_DBG_COLOR_B_MODES = 6,  /**< set which blocks modes to color */
+  AOM_SET_DBG_DISPLAY_MV = 7,     /**< set which motion vector modes to draw */
+
+  /* TODO(jkoleszar): The encoder incorrectly reuses some of these values (5+)
+   * for its control ids. These should be migrated to something like the
+   * AOM_DECODER_CTRL_ID_START range next time we're ready to break the ABI.
+   */
+  AV1_GET_REFERENCE = 128, /**< get a pointer to a reference frame */
+  AOM_COMMON_CTRL_ID_MAX,
+
+  AV1_GET_NEW_FRAME_IMAGE = 192, /**< get a pointer to the new frame */
+
+  AOM_DECODER_CTRL_ID_START = 256
+};
+
+/*!\brief post process flags
+ *
+ * The set of macros define AOM decoder post processing flags
+ */
+enum aom_postproc_level {
+  AOM_NOFILTERING = 0,
+  AOM_DEBLOCK = 1 << 0,
+  AOM_DEMACROBLOCK = 1 << 1,
+  AOM_ADDNOISE = 1 << 2,
+  AOM_DEBUG_TXT_FRAME_INFO = 1 << 3, /**< print frame information */
+  AOM_DEBUG_TXT_MBLK_MODES =
+      1 << 4, /**< print macro block modes over each macro block */
+  AOM_DEBUG_TXT_DC_DIFF = 1 << 5,   /**< print dc diff for each macro block */
+  AOM_DEBUG_TXT_RATE_INFO = 1 << 6, /**< print video rate info (encoder only) */
+  AOM_MFQE = 1 << 10
+};
+
+/*!\brief post process flags
+ *
+ * This define a structure that describe the post processing settings. For
+ * the best objective measure (using the PSNR metric) set post_proc_flag
+ * to AOM_DEBLOCK and deblocking_level to 1.
+ */
+
+typedef struct aom_postproc_cfg {
+  /*!\brief the types of post processing to be done, should be combination of
+   * "aom_postproc_level" */
+  int post_proc_flag;
+  int deblocking_level; /**< the strength of deblocking, valid range [0, 16] */
+  int noise_level; /**< the strength of additive noise, valid range [0, 16] */
+} aom_postproc_cfg_t;
+
+/*!\brief reference frame type
+ *
+ * The set of macros define the type of AOM reference frames
+ */
+typedef enum aom_ref_frame_type {
+  AOM_LAST_FRAME = 1,
+  AOM_GOLD_FRAME = 2,
+  AOM_ALTR_FRAME = 4
+} aom_ref_frame_type_t;
+
+/*!\brief reference frame data struct
+ *
+ * Define the data struct to access aom reference frames.
+ */
+typedef struct aom_ref_frame {
+  aom_ref_frame_type_t frame_type; /**< which reference frame */
+  aom_image_t img;                 /**< reference frame data in image format */
+} aom_ref_frame_t;
+
+/*!\brief AV1 specific reference frame data struct
+ *
+ * Define the data struct to access av1 reference frames.
+ */
+typedef struct av1_ref_frame {
+  int idx;         /**< frame index to get (input) */
+  aom_image_t img; /**< img structure to populate (output) */
+} av1_ref_frame_t;
+
+/*!\cond */
+/*!\brief aom decoder control function parameter type
+ *
+ * defines the data type for each of AOM decoder control function requires
+ */
+AOM_CTRL_USE_TYPE(AOM_SET_REFERENCE, aom_ref_frame_t *)
+#define AOM_CTRL_AOM_SET_REFERENCE
+AOM_CTRL_USE_TYPE(AOM_COPY_REFERENCE, aom_ref_frame_t *)
+#define AOM_CTRL_AOM_COPY_REFERENCE
+AOM_CTRL_USE_TYPE(AOM_SET_POSTPROC, aom_postproc_cfg_t *)
+#define AOM_CTRL_AOM_SET_POSTPROC
+AOM_CTRL_USE_TYPE(AOM_SET_DBG_COLOR_REF_FRAME, int)
+#define AOM_CTRL_AOM_SET_DBG_COLOR_REF_FRAME
+AOM_CTRL_USE_TYPE(AOM_SET_DBG_COLOR_MB_MODES, int)
+#define AOM_CTRL_AOM_SET_DBG_COLOR_MB_MODES
+AOM_CTRL_USE_TYPE(AOM_SET_DBG_COLOR_B_MODES, int)
+#define AOM_CTRL_AOM_SET_DBG_COLOR_B_MODES
+AOM_CTRL_USE_TYPE(AOM_SET_DBG_DISPLAY_MV, int)
+#define AOM_CTRL_AOM_SET_DBG_DISPLAY_MV
+AOM_CTRL_USE_TYPE(AV1_GET_REFERENCE, av1_ref_frame_t *)
+#define AOM_CTRL_AV1_GET_REFERENCE
+AOM_CTRL_USE_TYPE(AV1_GET_NEW_FRAME_IMAGE, aom_image_t *)
+#define AOM_CTRL_AV1_GET_NEW_FRAME_IMAGE
+
+/*!\endcond */
+/*! @} - end defgroup aom */
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_AOM_H_
diff --git a/third_party/aom/aom/aom_codec.h b/third_party/aom/aom/aom_codec.h
new file mode 100644
index 0000000000..1d301d16b2
--- /dev/null
+++ b/third_party/aom/aom/aom_codec.h
@@ -0,0 +1,487 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/*!\defgroup codec Common Algorithm Interface
+ * This abstraction allows applications to easily support multiple video
+ * formats with minimal code duplication. This section describes the interface
+ * common to all codecs (both encoders and decoders).
+ * @{
+ */
+
+/*!\file
+ * \brief Describes the codec algorithm interface to applications.
+ *
+ * This file describes the interface between an application and a
+ * video codec algorithm.
+ *
+ * An application instantiates a specific codec instance by using
+ * aom_codec_init() and a pointer to the algorithm's interface structure:
+ *     <pre>
+ *     my_app.c:
+ *       extern aom_codec_iface_t my_codec;
+ *       {
+ *           aom_codec_ctx_t algo;
+ *           res = aom_codec_init(&algo, &my_codec);
+ *       }
+ *     </pre>
+ *
+ * Once initialized, the instance is manged using other functions from
+ * the aom_codec_* family.
+ */
+#ifndef AOM_AOM_CODEC_H_
+#define AOM_AOM_CODEC_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./aom_integer.h"
+#include "./aom_image.h"
+
+/*!\brief Decorator indicating a function is deprecated */
+#ifndef DEPRECATED
+#if defined(__GNUC__) && __GNUC__
+#define DEPRECATED __attribute__((deprecated))
+#elif defined(_MSC_VER)
+#define DEPRECATED
+#else
+#define DEPRECATED
+#endif
+#endif /* DEPRECATED */
+
+#ifndef DECLSPEC_DEPRECATED
+#if defined(__GNUC__) && __GNUC__
+#define DECLSPEC_DEPRECATED /**< \copydoc #DEPRECATED */
+#elif defined(_MSC_VER)
+/*!\brief \copydoc #DEPRECATED */
+#define DECLSPEC_DEPRECATED __declspec(deprecated)
+#else
+#define DECLSPEC_DEPRECATED /**< \copydoc #DEPRECATED */
+#endif
+#endif /* DECLSPEC_DEPRECATED */
+
+/*!\brief Decorator indicating a function is potentially unused */
+#ifdef UNUSED
+#elif defined(__GNUC__) || defined(__clang__)
+#define UNUSED __attribute__((unused))
+#else
+#define UNUSED
+#endif
+
+/*!\brief Decorator indicating that given struct/union/enum is packed */
+#ifndef ATTRIBUTE_PACKED
+#if defined(__GNUC__) && __GNUC__
+#define ATTRIBUTE_PACKED __attribute__((packed))
+#elif defined(_MSC_VER)
+#define ATTRIBUTE_PACKED
+#else
+#define ATTRIBUTE_PACKED
+#endif
+#endif /* ATTRIBUTE_PACKED */
+
+/*!\brief Current ABI version number
+ *
+ * \internal
+ * If this file is altered in any way that changes the ABI, this value
+ * must be bumped.  Examples include, but are not limited to, changing
+ * types, removing or reassigning enums, adding/removing/rearranging
+ * fields to structures
+ */
+#define AOM_CODEC_ABI_VERSION (3 + AOM_IMAGE_ABI_VERSION) /**<\hideinitializer*/
+
+/*!\brief Algorithm return codes */
+typedef enum {
+  /*!\brief Operation completed without error */
+  AOM_CODEC_OK,
+
+  /*!\brief Unspecified error */
+  AOM_CODEC_ERROR,
+
+  /*!\brief Memory operation failed */
+  AOM_CODEC_MEM_ERROR,
+
+  /*!\brief ABI version mismatch */
+  AOM_CODEC_ABI_MISMATCH,
+
+  /*!\brief Algorithm does not have required capability */
+  AOM_CODEC_INCAPABLE,
+
+  /*!\brief The given bitstream is not supported.
+   *
+   * The bitstream was unable to be parsed at the highest level. The decoder
+   * is unable to proceed. This error \ref SHOULD be treated as fatal to the
+   * stream. */
+  AOM_CODEC_UNSUP_BITSTREAM,
+
+  /*!\brief Encoded bitstream uses an unsupported feature
+   *
+   * The decoder does not implement a feature required by the encoder. This
+   * return code should only be used for features that prevent future
+   * pictures from being properly decoded. This error \ref MAY be treated as
+   * fatal to the stream or \ref MAY be treated as fatal to the current GOP.
+   */
+  AOM_CODEC_UNSUP_FEATURE,
+
+  /*!\brief The coded data for this stream is corrupt or incomplete
+   *
+   * There was a problem decoding the current frame.  This return code
+   * should only be used for failures that prevent future pictures from
+   * being properly decoded. This error \ref MAY be treated as fatal to the
+   * stream or \ref MAY be treated as fatal to the current GOP. If decoding
+   * is continued for the current GOP, artifacts may be present.
+   */
+  AOM_CODEC_CORRUPT_FRAME,
+
+  /*!\brief An application-supplied parameter is not valid.
+   *
+   */
+  AOM_CODEC_INVALID_PARAM,
+
+  /*!\brief An iterator reached the end of list.
+   *
+   */
+  AOM_CODEC_LIST_END
+
+} aom_codec_err_t;
+
+/*! \brief Codec capabilities bitfield
+ *
+ *  Each codec advertises the capabilities it supports as part of its
+ *  ::aom_codec_iface_t interface structure. Capabilities are extra interfaces
+ *  or functionality, and are not required to be supported.
+ *
+ *  The available flags are specified by AOM_CODEC_CAP_* defines.
+ */
+typedef long aom_codec_caps_t;
+#define AOM_CODEC_CAP_DECODER 0x1 /**< Is a decoder */
+#define AOM_CODEC_CAP_ENCODER 0x2 /**< Is an encoder */
+
+/*! \brief Initialization-time Feature Enabling
+ *
+ *  Certain codec features must be known at initialization time, to allow for
+ *  proper memory allocation.
+ *
+ *  The available flags are specified by AOM_CODEC_USE_* defines.
+ */
+typedef long aom_codec_flags_t;
+
+/*!\brief Codec interface structure.
+ *
+ * Contains function pointers and other data private to the codec
+ * implementation. This structure is opaque to the application.
+ */
+typedef const struct aom_codec_iface aom_codec_iface_t;
+
+/*!\brief Codec private data structure.
+ *
+ * Contains data private to the codec implementation. This structure is opaque
+ * to the application.
+ */
+typedef struct aom_codec_priv aom_codec_priv_t;
+
+/*!\brief Iterator
+ *
+ * Opaque storage used for iterating over lists.
+ */
+typedef const void *aom_codec_iter_t;
+
+/*!\brief Codec context structure
+ *
+ * All codecs \ref MUST support this context structure fully. In general,
+ * this data should be considered private to the codec algorithm, and
+ * not be manipulated or examined by the calling application. Applications
+ * may reference the 'name' member to get a printable description of the
+ * algorithm.
+ */
+typedef struct aom_codec_ctx {
+  const char *name;             /**< Printable interface name */
+  aom_codec_iface_t *iface;     /**< Interface pointers */
+  aom_codec_err_t err;          /**< Last returned error */
+  const char *err_detail;       /**< Detailed info, if available */
+  aom_codec_flags_t init_flags; /**< Flags passed at init time */
+  union {
+    /**< Decoder Configuration Pointer */
+    const struct aom_codec_dec_cfg *dec;
+    /**< Encoder Configuration Pointer */
+    const struct aom_codec_enc_cfg *enc;
+    const void *raw;
+  } config;               /**< Configuration pointer aliasing union */
+  aom_codec_priv_t *priv; /**< Algorithm private storage */
+} aom_codec_ctx_t;
+
+/*!\brief Bit depth for codec
+ * *
+ * This enumeration determines the bit depth of the codec.
+ */
+typedef enum aom_bit_depth {
+  AOM_BITS_8 = 8,   /**<  8 bits */
+  AOM_BITS_10 = 10, /**< 10 bits */
+  AOM_BITS_12 = 12, /**< 12 bits */
+} aom_bit_depth_t;
+
+/*!\brief Superblock size selection.
+ *
+ * Defines the superblock size used for encoding. The superblock size can
+ * either be fixed at 64x64 or 128x128 pixels, or it can be dynamically
+ * selected by the encoder for each frame.
+ */
+typedef enum aom_superblock_size {
+  AOM_SUPERBLOCK_SIZE_64X64,   /**< Always use 64x64 superblocks. */
+  AOM_SUPERBLOCK_SIZE_128X128, /**< Always use 128x128 superblocks. */
+  AOM_SUPERBLOCK_SIZE_DYNAMIC  /**< Select superblock size dynamically. */
+} aom_superblock_size_t;
+
+/*
+ * Library Version Number Interface
+ *
+ * For example, see the following sample return values:
+ *     aom_codec_version()           (1<<16 | 2<<8 | 3)
+ *     aom_codec_version_str()       "v1.2.3-rc1-16-gec6a1ba"
+ *     aom_codec_version_extra_str() "rc1-16-gec6a1ba"
+ */
+
+/*!\brief Return the version information (as an integer)
+ *
+ * Returns a packed encoding of the library version number. This will only
+ * include
+ * the major.minor.patch component of the version number. Note that this encoded
+ * value should be accessed through the macros provided, as the encoding may
+ * change
+ * in the future.
+ *
+ */
+int aom_codec_version(void);
+#define AOM_VERSION_MAJOR(v) \
+  ((v >> 16) & 0xff) /**< extract major from packed version */
+#define AOM_VERSION_MINOR(v) \
+  ((v >> 8) & 0xff) /**< extract minor from packed version */
+#define AOM_VERSION_PATCH(v) \
+  ((v >> 0) & 0xff) /**< extract patch from packed version */
+
+/*!\brief Return the version major number */
+#define aom_codec_version_major() ((aom_codec_version() >> 16) & 0xff)
+
+/*!\brief Return the version minor number */
+#define aom_codec_version_minor() ((aom_codec_version() >> 8) & 0xff)
+
+/*!\brief Return the version patch number */
+#define aom_codec_version_patch() ((aom_codec_version() >> 0) & 0xff)
+
+/*!\brief Return the version information (as a string)
+ *
+ * Returns a printable string containing the full library version number. This
+ * may
+ * contain additional text following the three digit version number, as to
+ * indicate
+ * release candidates, prerelease versions, etc.
+ *
+ */
+const char *aom_codec_version_str(void);
+
+/*!\brief Return the version information (as a string)
+ *
+ * Returns a printable "extra string". This is the component of the string
+ * returned
+ * by aom_codec_version_str() following the three digit version number.
+ *
+ */
+const char *aom_codec_version_extra_str(void);
+
+/*!\brief Return the build configuration
+ *
+ * Returns a printable string containing an encoded version of the build
+ * configuration. This may be useful to aom support.
+ *
+ */
+const char *aom_codec_build_config(void);
+
+/*!\brief Return the name for a given interface
+ *
+ * Returns a human readable string for name of the given codec interface.
+ *
+ * \param[in]    iface     Interface pointer
+ *
+ */
+const char *aom_codec_iface_name(aom_codec_iface_t *iface);
+
+/*!\brief Convert error number to printable string
+ *
+ * Returns a human readable string for the last error returned by the
+ * algorithm. The returned error will be one line and will not contain
+ * any newline characters.
+ *
+ *
+ * \param[in]    err     Error number.
+ *
+ */
+const char *aom_codec_err_to_string(aom_codec_err_t err);
+
+/*!\brief Retrieve error synopsis for codec context
+ *
+ * Returns a human readable string for the last error returned by the
+ * algorithm. The returned error will be one line and will not contain
+ * any newline characters.
+ *
+ *
+ * \param[in]    ctx     Pointer to this instance's context.
+ *
+ */
+const char *aom_codec_error(aom_codec_ctx_t *ctx);
+
+/*!\brief Retrieve detailed error information for codec context
+ *
+ * Returns a human readable string providing detailed information about
+ * the last error.
+ *
+ * \param[in]    ctx     Pointer to this instance's context.
+ *
+ * \retval NULL
+ *     No detailed information is available.
+ */
+const char *aom_codec_error_detail(aom_codec_ctx_t *ctx);
+
+/* REQUIRED FUNCTIONS
+ *
+ * The following functions are required to be implemented for all codecs.
+ * They represent the base case functionality expected of all codecs.
+ */
+
+/*!\brief Destroy a codec instance
+ *
+ * Destroys a codec context, freeing any associated memory buffers.
+ *
+ * \param[in] ctx   Pointer to this instance's context
+ *
+ * \retval #AOM_CODEC_OK
+ *     The codec algorithm initialized.
+ * \retval #AOM_CODEC_MEM_ERROR
+ *     Memory allocation failed.
+ */
+aom_codec_err_t aom_codec_destroy(aom_codec_ctx_t *ctx);
+
+/*!\brief Get the capabilities of an algorithm.
+ *
+ * Retrieves the capabilities bitfield from the algorithm's interface.
+ *
+ * \param[in] iface   Pointer to the algorithm interface
+ *
+ */
+aom_codec_caps_t aom_codec_get_caps(aom_codec_iface_t *iface);
+
+/*!\brief Control algorithm
+ *
+ * This function is used to exchange algorithm specific data with the codec
+ * instance. This can be used to implement features specific to a particular
+ * algorithm.
+ *
+ * This wrapper function dispatches the request to the helper function
+ * associated with the given ctrl_id. It tries to call this function
+ * transparently, but will return #AOM_CODEC_ERROR if the request could not
+ * be dispatched.
+ *
+ * Note that this function should not be used directly. Call the
+ * #aom_codec_control wrapper macro instead.
+ *
+ * \param[in]     ctx              Pointer to this instance's context
+ * \param[in]     ctrl_id          Algorithm specific control identifier
+ *
+ * \retval #AOM_CODEC_OK
+ *     The control request was processed.
+ * \retval #AOM_CODEC_ERROR
+ *     The control request was not processed.
+ * \retval #AOM_CODEC_INVALID_PARAM
+ *     The data was not valid.
+ */
+aom_codec_err_t aom_codec_control_(aom_codec_ctx_t *ctx, int ctrl_id, ...);
+#if defined(AOM_DISABLE_CTRL_TYPECHECKS) && AOM_DISABLE_CTRL_TYPECHECKS
+#define aom_codec_control(ctx, id, data) aom_codec_control_(ctx, id, data)
+#define AOM_CTRL_USE_TYPE(id, typ)
+#define AOM_CTRL_USE_TYPE_DEPRECATED(id, typ)
+#define AOM_CTRL_VOID(id, typ)
+
+#else
+/*!\brief aom_codec_control wrapper macro
+ *
+ * This macro allows for type safe conversions across the variadic parameter
+ * to aom_codec_control_().
+ *
+ * \internal
+ * It works by dispatching the call to the control function through a wrapper
+ * function named with the id parameter.
+ */
+#define aom_codec_control(ctx, id, data) \
+  aom_codec_control_##id(ctx, id, data) /**<\hideinitializer*/
+
+/*!\brief aom_codec_control type definition macro
+ *
+ * This macro allows for type safe conversions across the variadic parameter
+ * to aom_codec_control_(). It defines the type of the argument for a given
+ * control identifier.
+ *
+ * \internal
+ * It defines a static function with
+ * the correctly typed arguments as a wrapper to the type-unsafe internal
+ * function.
+ */
+#define AOM_CTRL_USE_TYPE(id, typ)                                           \
+  static aom_codec_err_t aom_codec_control_##id(aom_codec_ctx_t *, int, typ) \
+      UNUSED;                                                                \
+                                                                             \
+  static aom_codec_err_t aom_codec_control_##id(aom_codec_ctx_t *ctx,        \
+                                                int ctrl_id, typ data) {     \
+    return aom_codec_control_(ctx, ctrl_id, data);                           \
+  } /**<\hideinitializer*/
+
+/*!\brief aom_codec_control deprecated type definition macro
+ *
+ * Like #AOM_CTRL_USE_TYPE, but indicates that the specified control is
+ * deprecated and should not be used. Consult the documentation for your
+ * codec for more information.
+ *
+ * \internal
+ * It defines a static function with the correctly typed arguments as a
+ * wrapper to the type-unsafe internal function.
+ */
+#define AOM_CTRL_USE_TYPE_DEPRECATED(id, typ)                        \
+  DECLSPEC_DEPRECATED static aom_codec_err_t aom_codec_control_##id( \
+      aom_codec_ctx_t *, int, typ) DEPRECATED UNUSED;                \
+                                                                     \
+  DECLSPEC_DEPRECATED static aom_codec_err_t aom_codec_control_##id( \
+      aom_codec_ctx_t *ctx, int ctrl_id, typ data) {                 \
+    return aom_codec_control_(ctx, ctrl_id, data);                   \
+  } /**<\hideinitializer*/
+
+/*!\brief aom_codec_control void type definition macro
+ *
+ * This macro allows for type safe conversions across the variadic parameter
+ * to aom_codec_control_(). It indicates that a given control identifier takes
+ * no argument.
+ *
+ * \internal
+ * It defines a static function without a data argument as a wrapper to the
+ * type-unsafe internal function.
+ */
+#define AOM_CTRL_VOID(id)                                               \
+  static aom_codec_err_t aom_codec_control_##id(aom_codec_ctx_t *, int) \
+      UNUSED;                                                           \
+                                                                        \
+  static aom_codec_err_t aom_codec_control_##id(aom_codec_ctx_t *ctx,   \
+                                                int ctrl_id) {          \
+    return aom_codec_control_(ctx, ctrl_id);                            \
+  } /**<\hideinitializer*/
+
+#endif
+
+/*!@} - end defgroup codec*/
+#ifdef __cplusplus
+}
+#endif
+#endif  // AOM_AOM_CODEC_H_
diff --git a/third_party/aom/aom/aom_codec.mk b/third_party/aom/aom/aom_codec.mk
new file mode 100644
index 0000000000..33bd3fe3be
--- /dev/null
+++ b/third_party/aom/aom/aom_codec.mk
@@ -0,0 +1,42 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+API_EXPORTS += exports
+
+API_SRCS-$(CONFIG_AV1_ENCODER) += aom.h
+API_SRCS-$(CONFIG_AV1_ENCODER) += aomcx.h
+API_DOC_SRCS-$(CONFIG_AV1_ENCODER) += aom.h
+API_DOC_SRCS-$(CONFIG_AV1_ENCODER) += aomcx.h
+
+API_SRCS-$(CONFIG_AV1_DECODER) += aom.h
+API_SRCS-$(CONFIG_AV1_DECODER) += aomdx.h
+API_DOC_SRCS-$(CONFIG_AV1_DECODER) += aom.h
+API_DOC_SRCS-$(CONFIG_AV1_DECODER) += aomdx.h
+
+API_DOC_SRCS-yes += aom_codec.h
+API_DOC_SRCS-yes += aom_decoder.h
+API_DOC_SRCS-yes += aom_encoder.h
+API_DOC_SRCS-yes += aom_frame_buffer.h
+API_DOC_SRCS-yes += aom_image.h
+
+API_SRCS-yes += src/aom_decoder.c
+API_SRCS-yes += aom_decoder.h
+API_SRCS-yes += src/aom_encoder.c
+API_SRCS-yes += aom_encoder.h
+API_SRCS-yes += internal/aom_codec_internal.h
+API_SRCS-yes += src/aom_codec.c
+API_SRCS-yes += src/aom_image.c
+API_SRCS-yes += aom_codec.h
+API_SRCS-yes += aom_codec.mk
+API_SRCS-yes += aom_frame_buffer.h
+API_SRCS-yes += aom_image.h
+API_SRCS-yes += aom_integer.h
diff --git a/third_party/aom/aom/aom_decoder.h b/third_party/aom/aom/aom_decoder.h
new file mode 100644
index 0000000000..e6f05048a9
--- /dev/null
+++ b/third_party/aom/aom/aom_decoder.h
@@ -0,0 +1,366 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_AOM_DECODER_H_
+#define AOM_AOM_DECODER_H_
+
+/*!\defgroup decoder Decoder Algorithm Interface
+ * \ingroup codec
+ * This abstraction allows applications using this decoder to easily support
+ * multiple video formats with minimal code duplication. This section describes
+ * the interface common to all decoders.
+ * @{
+ */
+
+/*!\file
+ * \brief Describes the decoder algorithm interface to applications.
+ *
+ * This file describes the interface between an application and a
+ * video decoder algorithm.
+ *
+ */
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./aom_codec.h"
+#include "./aom_frame_buffer.h"
+
+/*!\brief Current ABI version number
+ *
+ * \internal
+ * If this file is altered in any way that changes the ABI, this value
+ * must be bumped.  Examples include, but are not limited to, changing
+ * types, removing or reassigning enums, adding/removing/rearranging
+ * fields to structures
+ */
+#define AOM_DECODER_ABI_VERSION \
+  (3 + AOM_CODEC_ABI_VERSION) /**<\hideinitializer*/
+
+/*! \brief Decoder capabilities bitfield
+ *
+ *  Each decoder advertises the capabilities it supports as part of its
+ *  ::aom_codec_iface_t interface structure. Capabilities are extra interfaces
+ *  or functionality, and are not required to be supported by a decoder.
+ *
+ *  The available flags are specified by AOM_CODEC_CAP_* defines.
+ */
+#define AOM_CODEC_CAP_PUT_SLICE 0x10000 /**< Will issue put_slice callbacks */
+#define AOM_CODEC_CAP_PUT_FRAME 0x20000 /**< Will issue put_frame callbacks */
+#define AOM_CODEC_CAP_POSTPROC 0x40000  /**< Can postprocess decoded frame */
+/*!\brief Can conceal errors due to packet loss */
+#define AOM_CODEC_CAP_ERROR_CONCEALMENT 0x80000
+/*!\brief Can receive encoded frames one fragment at a time */
+#define AOM_CODEC_CAP_INPUT_FRAGMENTS 0x100000
+
+/*! \brief Initialization-time Feature Enabling
+ *
+ *  Certain codec features must be known at initialization time, to allow for
+ *  proper memory allocation.
+ *
+ *  The available flags are specified by AOM_CODEC_USE_* defines.
+ */
+/*!\brief Can support frame-based multi-threading */
+#define AOM_CODEC_CAP_FRAME_THREADING 0x200000
+/*!brief Can support external frame buffers */
+#define AOM_CODEC_CAP_EXTERNAL_FRAME_BUFFER 0x400000
+
+#define AOM_CODEC_USE_POSTPROC 0x10000 /**< Postprocess decoded frame */
+/*!\brief Conceal errors in decoded frames */
+#define AOM_CODEC_USE_ERROR_CONCEALMENT 0x20000
+/*!\brief The input frame should be passed to the decoder one fragment at a
+ * time */
+#define AOM_CODEC_USE_INPUT_FRAGMENTS 0x40000
+/*!\brief Enable frame-based multi-threading */
+#define AOM_CODEC_USE_FRAME_THREADING 0x80000
+
+/*!\brief Stream properties
+ *
+ * This structure is used to query or set properties of the decoded
+ * stream. Algorithms may extend this structure with data specific
+ * to their bitstream by setting the sz member appropriately.
+ */
+typedef struct aom_codec_stream_info {
+  unsigned int sz;    /**< Size of this structure */
+  unsigned int w;     /**< Width (or 0 for unknown/default) */
+  unsigned int h;     /**< Height (or 0 for unknown/default) */
+  unsigned int is_kf; /**< Current frame is a keyframe */
+} aom_codec_stream_info_t;
+
+/* REQUIRED FUNCTIONS
+ *
+ * The following functions are required to be implemented for all decoders.
+ * They represent the base case functionality expected of all decoders.
+ */
+
+/*!\brief Initialization Configurations
+ *
+ * This structure is used to pass init time configuration options to the
+ * decoder.
+ */
+typedef struct aom_codec_dec_cfg {
+  unsigned int threads; /**< Maximum number of threads to use, default 1 */
+  unsigned int w;       /**< Width */
+  unsigned int h;       /**< Height */
+} aom_codec_dec_cfg_t;  /**< alias for struct aom_codec_dec_cfg */
+
+/*!\brief Initialize a decoder instance
+ *
+ * Initializes a decoder context using the given interface. Applications
+ * should call the aom_codec_dec_init convenience macro instead of this
+ * function directly, to ensure that the ABI version number parameter
+ * is properly initialized.
+ *
+ * If the library was configured with --disable-multithread, this call
+ * is not thread safe and should be guarded with a lock if being used
+ * in a multithreaded context.
+ *
+ * \param[in]    ctx     Pointer to this instance's context.
+ * \param[in]    iface   Pointer to the algorithm interface to use.
+ * \param[in]    cfg     Configuration to use, if known. May be NULL.
+ * \param[in]    flags   Bitfield of AOM_CODEC_USE_* flags
+ * \param[in]    ver     ABI version number. Must be set to
+ *                       AOM_DECODER_ABI_VERSION
+ * \retval #AOM_CODEC_OK
+ *     The decoder algorithm initialized.
+ * \retval #AOM_CODEC_MEM_ERROR
+ *     Memory allocation failed.
+ */
+aom_codec_err_t aom_codec_dec_init_ver(aom_codec_ctx_t *ctx,
+                                       aom_codec_iface_t *iface,
+                                       const aom_codec_dec_cfg_t *cfg,
+                                       aom_codec_flags_t flags, int ver);
+
+/*!\brief Convenience macro for aom_codec_dec_init_ver()
+ *
+ * Ensures the ABI version parameter is properly set.
+ */
+#define aom_codec_dec_init(ctx, iface, cfg, flags) \
+  aom_codec_dec_init_ver(ctx, iface, cfg, flags, AOM_DECODER_ABI_VERSION)
+
+/*!\brief Parse stream info from a buffer
+ *
+ * Performs high level parsing of the bitstream. Construction of a decoder
+ * context is not necessary. Can be used to determine if the bitstream is
+ * of the proper format, and to extract information from the stream.
+ *
+ * \param[in]      iface   Pointer to the algorithm interface
+ * \param[in]      data    Pointer to a block of data to parse
+ * \param[in]      data_sz Size of the data buffer
+ * \param[in,out]  si      Pointer to stream info to update. The size member
+ *                         \ref MUST be properly initialized, but \ref MAY be
+ *                         clobbered by the algorithm. This parameter \ref MAY
+ *                         be NULL.
+ *
+ * \retval #AOM_CODEC_OK
+ *     Bitstream is parsable and stream information updated
+ */
+aom_codec_err_t aom_codec_peek_stream_info(aom_codec_iface_t *iface,
+                                           const uint8_t *data,
+                                           unsigned int data_sz,
+                                           aom_codec_stream_info_t *si);
+
+/*!\brief Return information about the current stream.
+ *
+ * Returns information about the stream that has been parsed during decoding.
+ *
+ * \param[in]      ctx     Pointer to this instance's context
+ * \param[in,out]  si      Pointer to stream info to update. The size member
+ *                         \ref MUST be properly initialized, but \ref MAY be
+ *                         clobbered by the algorithm. This parameter \ref MAY
+ *                         be NULL.
+ *
+ * \retval #AOM_CODEC_OK
+ *     Bitstream is parsable and stream information updated
+ */
+aom_codec_err_t aom_codec_get_stream_info(aom_codec_ctx_t *ctx,
+                                          aom_codec_stream_info_t *si);
+
+/*!\brief Decode data
+ *
+ * Processes a buffer of coded data. If the processing results in a new
+ * decoded frame becoming available, PUT_SLICE and PUT_FRAME events may be
+ * generated, as appropriate. Encoded data \ref MUST be passed in DTS (decode
+ * time stamp) order. Frames produced will always be in PTS (presentation
+ * time stamp) order.
+ * If the decoder is configured with AOM_CODEC_USE_INPUT_FRAGMENTS enabled,
+ * data and data_sz can contain a fragment of the encoded frame. Fragment
+ * \#n must contain at least partition \#n, but can also contain subsequent
+ * partitions (\#n+1 - \#n+i), and if so, fragments \#n+1, .., \#n+i must
+ * be empty. When no more data is available, this function should be called
+ * with NULL as data and 0 as data_sz. The memory passed to this function
+ * must be available until the frame has been decoded.
+ *
+ * \param[in] ctx          Pointer to this instance's context
+ * \param[in] data         Pointer to this block of new coded data. If
+ *                         NULL, a AOM_CODEC_CB_PUT_FRAME event is posted
+ *                         for the previously decoded frame.
+ * \param[in] data_sz      Size of the coded data, in bytes.
+ * \param[in] user_priv    Application specific data to associate with
+ *                         this frame.
+ * \param[in] deadline     Soft deadline the decoder should attempt to meet,
+ *                         in us. Set to zero for unlimited.
+ *
+ * \return Returns #AOM_CODEC_OK if the coded data was processed completely
+ *         and future pictures can be decoded without error. Otherwise,
+ *         see the descriptions of the other error codes in ::aom_codec_err_t
+ *         for recoverability capabilities.
+ */
+aom_codec_err_t aom_codec_decode(aom_codec_ctx_t *ctx, const uint8_t *data,
+                                 unsigned int data_sz, void *user_priv,
+                                 long deadline);
+
+/*!\brief Decoded frames iterator
+ *
+ * Iterates over a list of the frames available for display. The iterator
+ * storage should be initialized to NULL to start the iteration. Iteration is
+ * complete when this function returns NULL.
+ *
+ * The list of available frames becomes valid upon completion of the
+ * aom_codec_decode call, and remains valid until the next call to
+ * aom_codec_decode.
+ *
+ * \param[in]     ctx      Pointer to this instance's context
+ * \param[in,out] iter     Iterator storage, initialized to NULL
+ *
+ * \return Returns a pointer to an image, if one is ready for display. Frames
+ *         produced will always be in PTS (presentation time stamp) order.
+ */
+aom_image_t *aom_codec_get_frame(aom_codec_ctx_t *ctx, aom_codec_iter_t *iter);
+
+/*!\defgroup cap_put_frame Frame-Based Decoding Functions
+ *
+ * The following functions are required to be implemented for all decoders
+ * that advertise the AOM_CODEC_CAP_PUT_FRAME capability. Calling these
+ * functions
+ * for codecs that don't advertise this capability will result in an error
+ * code being returned, usually AOM_CODEC_ERROR
+ * @{
+ */
+
+/*!\brief put frame callback prototype
+ *
+ * This callback is invoked by the decoder to notify the application of
+ * the availability of decoded image data.
+ */
+typedef void (*aom_codec_put_frame_cb_fn_t)(void *user_priv,
+                                            const aom_image_t *img);
+
+/*!\brief Register for notification of frame completion.
+ *
+ * Registers a given function to be called when a decoded frame is
+ * available.
+ *
+ * \param[in] ctx          Pointer to this instance's context
+ * \param[in] cb           Pointer to the callback function
+ * \param[in] user_priv    User's private data
+ *
+ * \retval #AOM_CODEC_OK
+ *     Callback successfully registered.
+ * \retval #AOM_CODEC_ERROR
+ *     Decoder context not initialized, or algorithm not capable of
+ *     posting slice completion.
+ */
+aom_codec_err_t aom_codec_register_put_frame_cb(aom_codec_ctx_t *ctx,
+                                                aom_codec_put_frame_cb_fn_t cb,
+                                                void *user_priv);
+
+/*!@} - end defgroup cap_put_frame */
+
+/*!\defgroup cap_put_slice Slice-Based Decoding Functions
+ *
+ * The following functions are required to be implemented for all decoders
+ * that advertise the AOM_CODEC_CAP_PUT_SLICE capability. Calling these
+ * functions
+ * for codecs that don't advertise this capability will result in an error
+ * code being returned, usually AOM_CODEC_ERROR
+ * @{
+ */
+
+/*!\brief put slice callback prototype
+ *
+ * This callback is invoked by the decoder to notify the application of
+ * the availability of partially decoded image data. The
+ */
+typedef void (*aom_codec_put_slice_cb_fn_t)(void *user_priv,
+                                            const aom_image_t *img,
+                                            const aom_image_rect_t *valid,
+                                            const aom_image_rect_t *update);
+
+/*!\brief Register for notification of slice completion.
+ *
+ * Registers a given function to be called when a decoded slice is
+ * available.
+ *
+ * \param[in] ctx          Pointer to this instance's context
+ * \param[in] cb           Pointer to the callback function
+ * \param[in] user_priv    User's private data
+ *
+ * \retval #AOM_CODEC_OK
+ *     Callback successfully registered.
+ * \retval #AOM_CODEC_ERROR
+ *     Decoder context not initialized, or algorithm not capable of
+ *     posting slice completion.
+ */
+aom_codec_err_t aom_codec_register_put_slice_cb(aom_codec_ctx_t *ctx,
+                                                aom_codec_put_slice_cb_fn_t cb,
+                                                void *user_priv);
+
+/*!@} - end defgroup cap_put_slice*/
+
+/*!\defgroup cap_external_frame_buffer External Frame Buffer Functions
+ *
+ * The following section is required to be implemented for all decoders
+ * that advertise the AOM_CODEC_CAP_EXTERNAL_FRAME_BUFFER capability.
+ * Calling this function for codecs that don't advertise this capability
+ * will result in an error code being returned, usually AOM_CODEC_ERROR.
+ *
+ * \note
+ * Currently this only works with AV1.
+ * @{
+ */
+
+/*!\brief Pass in external frame buffers for the decoder to use.
+ *
+ * Registers functions to be called when libaom needs a frame buffer
+ * to decode the current frame and a function to be called when libaom does
+ * not internally reference the frame buffer. This set function must
+ * be called before the first call to decode or libaom will assume the
+ * default behavior of allocating frame buffers internally.
+ *
+ * \param[in] ctx          Pointer to this instance's context
+ * \param[in] cb_get       Pointer to the get callback function
+ * \param[in] cb_release   Pointer to the release callback function
+ * \param[in] cb_priv      Callback's private data
+ *
+ * \retval #AOM_CODEC_OK
+ *     External frame buffers will be used by libaom.
+ * \retval #AOM_CODEC_INVALID_PARAM
+ *     One or more of the callbacks were NULL.
+ * \retval #AOM_CODEC_ERROR
+ *     Decoder context not initialized, or algorithm not capable of
+ *     using external frame buffers.
+ *
+ * \note
+ * When decoding AV1, the application may be required to pass in at least
+ * #AOM_MAXIMUM_WORK_BUFFERS external frame
+ * buffers.
+ */
+aom_codec_err_t aom_codec_set_frame_buffer_functions(
+    aom_codec_ctx_t *ctx, aom_get_frame_buffer_cb_fn_t cb_get,
+    aom_release_frame_buffer_cb_fn_t cb_release, void *cb_priv);
+
+/*!@} - end defgroup cap_external_frame_buffer */
+
+/*!@} - end defgroup decoder*/
+#ifdef __cplusplus
+}
+#endif
+#endif  // AOM_AOM_DECODER_H_
diff --git a/third_party/aom/aom/aom_encoder.h b/third_party/aom/aom/aom_encoder.h
new file mode 100644
index 0000000000..14c9f0c263
--- /dev/null
+++ b/third_party/aom/aom/aom_encoder.h
@@ -0,0 +1,832 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_AOM_ENCODER_H_
+#define AOM_AOM_ENCODER_H_
+
+/*!\defgroup encoder Encoder Algorithm Interface
+ * \ingroup codec
+ * This abstraction allows applications using this encoder to easily support
+ * multiple video formats with minimal code duplication. This section describes
+ * the interface common to all encoders.
+ * @{
+ */
+
+/*!\file
+ * \brief Describes the encoder algorithm interface to applications.
+ *
+ * This file describes the interface between an application and a
+ * video encoder algorithm.
+ *
+ */
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./aom_codec.h"
+
+/*!\brief Current ABI version number
+ *
+ * \internal
+ * If this file is altered in any way that changes the ABI, this value
+ * must be bumped.  Examples include, but are not limited to, changing
+ * types, removing or reassigning enums, adding/removing/rearranging
+ * fields to structures
+ */
+#define AOM_ENCODER_ABI_VERSION \
+  (5 + AOM_CODEC_ABI_VERSION) /**<\hideinitializer*/
+
+/*! \brief Encoder capabilities bitfield
+ *
+ *  Each encoder advertises the capabilities it supports as part of its
+ *  ::aom_codec_iface_t interface structure. Capabilities are extra
+ *  interfaces or functionality, and are not required to be supported
+ *  by an encoder.
+ *
+ *  The available flags are specified by AOM_CODEC_CAP_* defines.
+ */
+#define AOM_CODEC_CAP_PSNR 0x10000 /**< Can issue PSNR packets */
+
+/*! Can output one partition at a time. Each partition is returned in its
+ *  own AOM_CODEC_CX_FRAME_PKT, with the FRAME_IS_FRAGMENT flag set for
+ *  every partition but the last. In this mode all frames are always
+ *  returned partition by partition.
+ */
+#define AOM_CODEC_CAP_OUTPUT_PARTITION 0x20000
+
+/*! Can support input images at greater than 8 bitdepth.
+ */
+#define AOM_CODEC_CAP_HIGHBITDEPTH 0x40000
+
+/*! \brief Initialization-time Feature Enabling
+ *
+ *  Certain codec features must be known at initialization time, to allow
+ *  for proper memory allocation.
+ *
+ *  The available flags are specified by AOM_CODEC_USE_* defines.
+ */
+#define AOM_CODEC_USE_PSNR 0x10000 /**< Calculate PSNR on each frame */
+/*!\brief Make the encoder output one  partition at a time. */
+#define AOM_CODEC_USE_OUTPUT_PARTITION 0x20000
+#define AOM_CODEC_USE_HIGHBITDEPTH 0x40000 /**< Use high bitdepth */
+
+/*!\brief Generic fixed size buffer structure
+ *
+ * This structure is able to hold a reference to any fixed size buffer.
+ */
+typedef struct aom_fixed_buf {
+  void *buf;       /**< Pointer to the data */
+  size_t sz;       /**< Length of the buffer, in chars */
+} aom_fixed_buf_t; /**< alias for struct aom_fixed_buf */
+
+/*!\brief Time Stamp Type
+ *
+ * An integer, which when multiplied by the stream's time base, provides
+ * the absolute time of a sample.
+ */
+typedef int64_t aom_codec_pts_t;
+
+/*!\brief Compressed Frame Flags
+ *
+ * This type represents a bitfield containing information about a compressed
+ * frame that may be useful to an application. The most significant 16 bits
+ * can be used by an algorithm to provide additional detail, for example to
+ * support frame types that are codec specific (MPEG-1 D-frames for example)
+ */
+typedef uint32_t aom_codec_frame_flags_t;
+#define AOM_FRAME_IS_KEY 0x1 /**< frame is the start of a GOP */
+/*!\brief frame can be dropped without affecting the stream (no future frame
+ * depends on this one) */
+#define AOM_FRAME_IS_DROPPABLE 0x2
+/*!\brief frame should be decoded but will not be shown */
+#define AOM_FRAME_IS_INVISIBLE 0x4
+/*!\brief this is a fragment of the encoded frame */
+#define AOM_FRAME_IS_FRAGMENT 0x8
+
+/*!\brief Error Resilient flags
+ *
+ * These flags define which error resilient features to enable in the
+ * encoder. The flags are specified through the
+ * aom_codec_enc_cfg::g_error_resilient variable.
+ */
+typedef uint32_t aom_codec_er_flags_t;
+/*!\brief Improve resiliency against losses of whole frames */
+#define AOM_ERROR_RESILIENT_DEFAULT 0x1
+/*!\brief The frame partitions are independently decodable by the bool decoder,
+ * meaning that partitions can be decoded even though earlier partitions have
+ * been lost. Note that intra prediction is still done over the partition
+ * boundary. */
+#define AOM_ERROR_RESILIENT_PARTITIONS 0x2
+
+/*!\brief Encoder output packet variants
+ *
+ * This enumeration lists the different kinds of data packets that can be
+ * returned by calls to aom_codec_get_cx_data(). Algorithms \ref MAY
+ * extend this list to provide additional functionality.
+ */
+enum aom_codec_cx_pkt_kind {
+  AOM_CODEC_CX_FRAME_PKT,    /**< Compressed video frame */
+  AOM_CODEC_STATS_PKT,       /**< Two-pass statistics for this frame */
+  AOM_CODEC_FPMB_STATS_PKT,  /**< first pass mb statistics for this frame */
+  AOM_CODEC_PSNR_PKT,        /**< PSNR statistics for this frame */
+  AOM_CODEC_CUSTOM_PKT = 256 /**< Algorithm extensions  */
+};
+
+/*!\brief Encoder output packet
+ *
+ * This structure contains the different kinds of output data the encoder
+ * may produce while compressing a frame.
+ */
+typedef struct aom_codec_cx_pkt {
+  enum aom_codec_cx_pkt_kind kind; /**< packet variant */
+  union {
+    struct {
+      void *buf; /**< compressed data buffer */
+      size_t sz; /**< length of compressed data */
+      /*!\brief time stamp to show frame (in timebase units) */
+      aom_codec_pts_t pts;
+      /*!\brief duration to show frame (in timebase units) */
+      unsigned long duration;
+      aom_codec_frame_flags_t flags; /**< flags for this frame */
+      /*!\brief the partition id defines the decoding order of the partitions.
+       * Only applicable when "output partition" mode is enabled. First
+       * partition has id 0.*/
+      int partition_id;
+    } frame;                            /**< data for compressed frame packet */
+    aom_fixed_buf_t twopass_stats;      /**< data for two-pass packet */
+    aom_fixed_buf_t firstpass_mb_stats; /**< first pass mb packet */
+    struct aom_psnr_pkt {
+      unsigned int samples[4]; /**< Number of samples, total/y/u/v */
+      uint64_t sse[4];         /**< sum squared error, total/y/u/v */
+      double psnr[4];          /**< PSNR, total/y/u/v */
+    } psnr;                    /**< data for PSNR packet */
+    aom_fixed_buf_t raw;       /**< data for arbitrary packets */
+
+    /* This packet size is fixed to allow codecs to extend this
+     * interface without having to manage storage for raw packets,
+     * i.e., if it's smaller than 128 bytes, you can store in the
+     * packet list directly.
+     */
+    char pad[128 - sizeof(enum aom_codec_cx_pkt_kind)]; /**< fixed sz */
+  } data;                                               /**< packet data */
+} aom_codec_cx_pkt_t; /**< alias for struct aom_codec_cx_pkt */
+
+/*!\brief Rational Number
+ *
+ * This structure holds a fractional value.
+ */
+typedef struct aom_rational {
+  int num;        /**< fraction numerator */
+  int den;        /**< fraction denominator */
+} aom_rational_t; /**< alias for struct aom_rational */
+
+/*!\brief Multi-pass Encoding Pass */
+enum aom_enc_pass {
+  AOM_RC_ONE_PASS,   /**< Single pass mode */
+  AOM_RC_FIRST_PASS, /**< First pass of multi-pass mode */
+  AOM_RC_LAST_PASS   /**< Final pass of multi-pass mode */
+};
+
+/*!\brief Rate control mode */
+enum aom_rc_mode {
+  AOM_VBR, /**< Variable Bit Rate (VBR) mode */
+  AOM_CBR, /**< Constant Bit Rate (CBR) mode */
+  AOM_CQ,  /**< Constrained Quality (CQ)  mode */
+  AOM_Q,   /**< Constant Quality (Q) mode */
+};
+
+/*!\brief Keyframe placement mode.
+ *
+ * This enumeration determines whether keyframes are placed automatically by
+ * the encoder or whether this behavior is disabled. Older releases of this
+ * SDK were implemented such that AOM_KF_FIXED meant keyframes were disabled.
+ * This name is confusing for this behavior, so the new symbols to be used
+ * are AOM_KF_AUTO and AOM_KF_DISABLED.
+ */
+enum aom_kf_mode {
+  AOM_KF_FIXED,       /**< deprecated, implies AOM_KF_DISABLED */
+  AOM_KF_AUTO,        /**< Encoder determines optimal placement automatically */
+  AOM_KF_DISABLED = 0 /**< Encoder does not place keyframes. */
+};
+
+/*!\brief Encoded Frame Flags
+ *
+ * This type indicates a bitfield to be passed to aom_codec_encode(), defining
+ * per-frame boolean values. By convention, bits common to all codecs will be
+ * named AOM_EFLAG_*, and bits specific to an algorithm will be named
+ * /algo/_eflag_*. The lower order 16 bits are reserved for common use.
+ */
+typedef long aom_enc_frame_flags_t;
+#define AOM_EFLAG_FORCE_KF (1 << 0) /**< Force this frame to be a keyframe */
+
+/*!\brief Encoder configuration structure
+ *
+ * This structure contains the encoder settings that have common representations
+ * across all codecs. This doesn't imply that all codecs support all features,
+ * however.
+ */
+typedef struct aom_codec_enc_cfg {
+  /*
+   * generic settings (g)
+   */
+
+  /*!\brief Algorithm specific "usage" value
+   *
+   * Algorithms may define multiple values for usage, which may convey the
+   * intent of how the application intends to use the stream. If this value
+   * is non-zero, consult the documentation for the codec to determine its
+   * meaning.
+   */
+  unsigned int g_usage;
+
+  /*!\brief Maximum number of threads to use
+   *
+   * For multi-threaded implementations, use no more than this number of
+   * threads. The codec may use fewer threads than allowed. The value
+   * 0 is equivalent to the value 1.
+   */
+  unsigned int g_threads;
+
+  /*!\brief Bitstream profile to use
+   *
+   * Some codecs support a notion of multiple bitstream profiles. Typically
+   * this maps to a set of features that are turned on or off. Often the
+   * profile to use is determined by the features of the intended decoder.
+   * Consult the documentation for the codec to determine the valid values
+   * for this parameter, or set to zero for a sane default.
+   */
+  unsigned int g_profile; /**< profile of bitstream to use */
+
+  /*!\brief Width of the frame
+   *
+   * This value identifies the presentation resolution of the frame,
+   * in pixels. Note that the frames passed as input to the encoder must
+   * have this resolution. Frames will be presented by the decoder in this
+   * resolution, independent of any spatial resampling the encoder may do.
+   */
+  unsigned int g_w;
+
+  /*!\brief Height of the frame
+   *
+   * This value identifies the presentation resolution of the frame,
+   * in pixels. Note that the frames passed as input to the encoder must
+   * have this resolution. Frames will be presented by the decoder in this
+   * resolution, independent of any spatial resampling the encoder may do.
+   */
+  unsigned int g_h;
+
+  /*!\brief Bit-depth of the codec
+   *
+   * This value identifies the bit_depth of the codec,
+   * Only certain bit-depths are supported as identified in the
+   * aom_bit_depth_t enum.
+   */
+  aom_bit_depth_t g_bit_depth;
+
+  /*!\brief Bit-depth of the input frames
+   *
+   * This value identifies the bit_depth of the input frames in bits.
+   * Note that the frames passed as input to the encoder must have
+   * this bit-depth.
+   */
+  unsigned int g_input_bit_depth;
+
+  /*!\brief Stream timebase units
+   *
+   * Indicates the smallest interval of time, in seconds, used by the stream.
+   * For fixed frame rate material, or variable frame rate material where
+   * frames are timed at a multiple of a given clock (ex: video capture),
+   * the \ref RECOMMENDED method is to set the timebase to the reciprocal
+   * of the frame rate (ex: 1001/30000 for 29.970 Hz NTSC). This allows the
+   * pts to correspond to the frame number, which can be handy. For
+   * re-encoding video from containers with absolute time timestamps, the
+   * \ref RECOMMENDED method is to set the timebase to that of the parent
+   * container or multimedia framework (ex: 1/1000 for ms, as in FLV).
+   */
+  struct aom_rational g_timebase;
+
+  /*!\brief Enable error resilient modes.
+   *
+   * The error resilient bitfield indicates to the encoder which features
+   * it should enable to take measures for streaming over lossy or noisy
+   * links.
+   */
+  aom_codec_er_flags_t g_error_resilient;
+
+  /*!\brief Multi-pass Encoding Mode
+   *
+   * This value should be set to the current phase for multi-pass encoding.
+   * For single pass, set to #AOM_RC_ONE_PASS.
+   */
+  enum aom_enc_pass g_pass;
+
+  /*!\brief Allow lagged encoding
+   *
+   * If set, this value allows the encoder to consume a number of input
+   * frames before producing output frames. This allows the encoder to
+   * base decisions for the current frame on future frames. This does
+   * increase the latency of the encoding pipeline, so it is not appropriate
+   * in all situations (ex: realtime encoding).
+   *
+   * Note that this is a maximum value -- the encoder may produce frames
+   * sooner than the given limit. Set this value to 0 to disable this
+   * feature.
+   */
+  unsigned int g_lag_in_frames;
+
+  /*
+   * rate control settings (rc)
+   */
+
+  /*!\brief Temporal resampling configuration, if supported by the codec.
+   *
+   * Temporal resampling allows the codec to "drop" frames as a strategy to
+   * meet its target data rate. This can cause temporal discontinuities in
+   * the encoded video, which may appear as stuttering during playback. This
+   * trade-off is often acceptable, but for many applications is not. It can
+   * be disabled in these cases.
+   *
+   * Note that not all codecs support this feature. All aom AVx codecs do.
+   * For other codecs, consult the documentation for that algorithm.
+   *
+   * This threshold is described as a percentage of the target data buffer.
+   * When the data buffer falls below this percentage of fullness, a
+   * dropped frame is indicated. Set the threshold to zero (0) to disable
+   * this feature.
+   */
+  unsigned int rc_dropframe_thresh;
+
+  /*!\brief Enable/disable spatial resampling, if supported by the codec.
+   *
+   * Spatial resampling allows the codec to compress a lower resolution
+   * version of the frame, which is then upscaled by the encoder to the
+   * correct presentation resolution. This increases visual quality at
+   * low data rates, at the expense of CPU time on the encoder/decoder.
+   */
+  unsigned int rc_resize_allowed;
+
+  /*!\brief Internal coded frame width.
+   *
+   * If spatial resampling is enabled this specifies the width of the
+   * encoded frame.
+   */
+  unsigned int rc_scaled_width;
+
+  /*!\brief Internal coded frame height.
+   *
+   * If spatial resampling is enabled this specifies the height of the
+   * encoded frame.
+   */
+  unsigned int rc_scaled_height;
+
+  /*!\brief Spatial resampling up watermark.
+   *
+   * This threshold is described as a percentage of the target data buffer.
+   * When the data buffer rises above this percentage of fullness, the
+   * encoder will step up to a higher resolution version of the frame.
+   */
+  unsigned int rc_resize_up_thresh;
+
+  /*!\brief Spatial resampling down watermark.
+   *
+   * This threshold is described as a percentage of the target data buffer.
+   * When the data buffer falls below this percentage of fullness, the
+   * encoder will step down to a lower resolution version of the frame.
+   */
+  unsigned int rc_resize_down_thresh;
+
+  /*!\brief Rate control algorithm to use.
+   *
+   * Indicates whether the end usage of this stream is to be streamed over
+   * a bandwidth constrained link, indicating that Constant Bit Rate (CBR)
+   * mode should be used, or whether it will be played back on a high
+   * bandwidth link, as from a local disk, where higher variations in
+   * bitrate are acceptable.
+   */
+  enum aom_rc_mode rc_end_usage;
+
+  /*!\brief Two-pass stats buffer.
+   *
+   * A buffer containing all of the stats packets produced in the first
+   * pass, concatenated.
+   */
+  aom_fixed_buf_t rc_twopass_stats_in;
+
+  /*!\brief first pass mb stats buffer.
+   *
+   * A buffer containing all of the first pass mb stats packets produced
+   * in the first pass, concatenated.
+   */
+  aom_fixed_buf_t rc_firstpass_mb_stats_in;
+
+  /*!\brief Target data rate
+   *
+   * Target bandwidth to use for this stream, in kilobits per second.
+   */
+  unsigned int rc_target_bitrate;
+
+  /*
+   * quantizer settings
+   */
+
+  /*!\brief Minimum (Best Quality) Quantizer
+   *
+   * The quantizer is the most direct control over the quality of the
+   * encoded image. The range of valid values for the quantizer is codec
+   * specific. Consult the documentation for the codec to determine the
+   * values to use. To determine the range programmatically, call
+   * aom_codec_enc_config_default() with a usage value of 0.
+   */
+  unsigned int rc_min_quantizer;
+
+  /*!\brief Maximum (Worst Quality) Quantizer
+   *
+   * The quantizer is the most direct control over the quality of the
+   * encoded image. The range of valid values for the quantizer is codec
+   * specific. Consult the documentation for the codec to determine the
+   * values to use. To determine the range programmatically, call
+   * aom_codec_enc_config_default() with a usage value of 0.
+   */
+  unsigned int rc_max_quantizer;
+
+  /*
+   * bitrate tolerance
+   */
+
+  /*!\brief Rate control adaptation undershoot control
+   *
+   * This value, expressed as a percentage of the target bitrate,
+   * controls the maximum allowed adaptation speed of the codec.
+   * This factor controls the maximum amount of bits that can
+   * be subtracted from the target bitrate in order to compensate
+   * for prior overshoot.
+   *
+   * Valid values in the range 0-1000.
+   */
+  unsigned int rc_undershoot_pct;
+
+  /*!\brief Rate control adaptation overshoot control
+   *
+   * This value, expressed as a percentage of the target bitrate,
+   * controls the maximum allowed adaptation speed of the codec.
+   * This factor controls the maximum amount of bits that can
+   * be added to the target bitrate in order to compensate for
+   * prior undershoot.
+   *
+   * Valid values in the range 0-1000.
+   */
+  unsigned int rc_overshoot_pct;
+
+  /*
+   * decoder buffer model parameters
+   */
+
+  /*!\brief Decoder Buffer Size
+   *
+   * This value indicates the amount of data that may be buffered by the
+   * decoding application. Note that this value is expressed in units of
+   * time (milliseconds). For example, a value of 5000 indicates that the
+   * client will buffer (at least) 5000ms worth of encoded data. Use the
+   * target bitrate (#rc_target_bitrate) to convert to bits/bytes, if
+   * necessary.
+   */
+  unsigned int rc_buf_sz;
+
+  /*!\brief Decoder Buffer Initial Size
+   *
+   * This value indicates the amount of data that will be buffered by the
+   * decoding application prior to beginning playback. This value is
+   * expressed in units of time (milliseconds). Use the target bitrate
+   * (#rc_target_bitrate) to convert to bits/bytes, if necessary.
+   */
+  unsigned int rc_buf_initial_sz;
+
+  /*!\brief Decoder Buffer Optimal Size
+   *
+   * This value indicates the amount of data that the encoder should try
+   * to maintain in the decoder's buffer. This value is expressed in units
+   * of time (milliseconds). Use the target bitrate (#rc_target_bitrate)
+   * to convert to bits/bytes, if necessary.
+   */
+  unsigned int rc_buf_optimal_sz;
+
+  /*
+   * 2 pass rate control parameters
+   */
+
+  /*!\brief Two-pass mode CBR/VBR bias
+   *
+   * Bias, expressed on a scale of 0 to 100, for determining target size
+   * for the current frame. The value 0 indicates the optimal CBR mode
+   * value should be used. The value 100 indicates the optimal VBR mode
+   * value should be used. Values in between indicate which way the
+   * encoder should "lean."
+   */
+  unsigned int rc_2pass_vbr_bias_pct;
+
+  /*!\brief Two-pass mode per-GOP minimum bitrate
+   *
+   * This value, expressed as a percentage of the target bitrate, indicates
+   * the minimum bitrate to be used for a single GOP (aka "section")
+   */
+  unsigned int rc_2pass_vbr_minsection_pct;
+
+  /*!\brief Two-pass mode per-GOP maximum bitrate
+   *
+   * This value, expressed as a percentage of the target bitrate, indicates
+   * the maximum bitrate to be used for a single GOP (aka "section")
+   */
+  unsigned int rc_2pass_vbr_maxsection_pct;
+
+  /*
+   * keyframing settings (kf)
+   */
+
+  /*!\brief Keyframe placement mode
+   *
+   * This value indicates whether the encoder should place keyframes at a
+   * fixed interval, or determine the optimal placement automatically
+   * (as governed by the #kf_min_dist and #kf_max_dist parameters)
+   */
+  enum aom_kf_mode kf_mode;
+
+  /*!\brief Keyframe minimum interval
+   *
+   * This value, expressed as a number of frames, prevents the encoder from
+   * placing a keyframe nearer than kf_min_dist to the previous keyframe. At
+   * least kf_min_dist frames non-keyframes will be coded before the next
+   * keyframe. Set kf_min_dist equal to kf_max_dist for a fixed interval.
+   */
+  unsigned int kf_min_dist;
+
+  /*!\brief Keyframe maximum interval
+   *
+   * This value, expressed as a number of frames, forces the encoder to code
+   * a keyframe if one has not been coded in the last kf_max_dist frames.
+   * A value of 0 implies all frames will be keyframes. Set kf_min_dist
+   * equal to kf_max_dist for a fixed interval.
+   */
+  unsigned int kf_max_dist;
+} aom_codec_enc_cfg_t; /**< alias for struct aom_codec_enc_cfg */
+
+/*!\brief Initialize an encoder instance
+ *
+ * Initializes a encoder context using the given interface. Applications
+ * should call the aom_codec_enc_init convenience macro instead of this
+ * function directly, to ensure that the ABI version number parameter
+ * is properly initialized.
+ *
+ * If the library was configured with --disable-multithread, this call
+ * is not thread safe and should be guarded with a lock if being used
+ * in a multithreaded context.
+ *
+ * \param[in]    ctx     Pointer to this instance's context.
+ * \param[in]    iface   Pointer to the algorithm interface to use.
+ * \param[in]    cfg     Configuration to use, if known. May be NULL.
+ * \param[in]    flags   Bitfield of AOM_CODEC_USE_* flags
+ * \param[in]    ver     ABI version number. Must be set to
+ *                       AOM_ENCODER_ABI_VERSION
+ * \retval #AOM_CODEC_OK
+ *     The decoder algorithm initialized.
+ * \retval #AOM_CODEC_MEM_ERROR
+ *     Memory allocation failed.
+ */
+aom_codec_err_t aom_codec_enc_init_ver(aom_codec_ctx_t *ctx,
+                                       aom_codec_iface_t *iface,
+                                       const aom_codec_enc_cfg_t *cfg,
+                                       aom_codec_flags_t flags, int ver);
+
+/*!\brief Convenience macro for aom_codec_enc_init_ver()
+ *
+ * Ensures the ABI version parameter is properly set.
+ */
+#define aom_codec_enc_init(ctx, iface, cfg, flags) \
+  aom_codec_enc_init_ver(ctx, iface, cfg, flags, AOM_ENCODER_ABI_VERSION)
+
+/*!\brief Initialize multi-encoder instance
+ *
+ * Initializes multi-encoder context using the given interface.
+ * Applications should call the aom_codec_enc_init_multi convenience macro
+ * instead of this function directly, to ensure that the ABI version number
+ * parameter is properly initialized.
+ *
+ * \param[in]    ctx     Pointer to this instance's context.
+ * \param[in]    iface   Pointer to the algorithm interface to use.
+ * \param[in]    cfg     Configuration to use, if known. May be NULL.
+ * \param[in]    num_enc Total number of encoders.
+ * \param[in]    flags   Bitfield of AOM_CODEC_USE_* flags
+ * \param[in]    dsf     Pointer to down-sampling factors.
+ * \param[in]    ver     ABI version number. Must be set to
+ *                       AOM_ENCODER_ABI_VERSION
+ * \retval #AOM_CODEC_OK
+ *     The decoder algorithm initialized.
+ * \retval #AOM_CODEC_MEM_ERROR
+ *     Memory allocation failed.
+ */
+aom_codec_err_t aom_codec_enc_init_multi_ver(
+    aom_codec_ctx_t *ctx, aom_codec_iface_t *iface, aom_codec_enc_cfg_t *cfg,
+    int num_enc, aom_codec_flags_t flags, aom_rational_t *dsf, int ver);
+
+/*!\brief Convenience macro for aom_codec_enc_init_multi_ver()
+ *
+ * Ensures the ABI version parameter is properly set.
+ */
+#define aom_codec_enc_init_multi(ctx, iface, cfg, num_enc, flags, dsf) \
+  aom_codec_enc_init_multi_ver(ctx, iface, cfg, num_enc, flags, dsf,   \
+                               AOM_ENCODER_ABI_VERSION)
+
+/*!\brief Get a default configuration
+ *
+ * Initializes a encoder configuration structure with default values. Supports
+ * the notion of "usages" so that an algorithm may offer different default
+ * settings depending on the user's intended goal. This function \ref SHOULD
+ * be called by all applications to initialize the configuration structure
+ * before specializing the configuration with application specific values.
+ *
+ * \param[in]    iface     Pointer to the algorithm interface to use.
+ * \param[out]   cfg       Configuration buffer to populate.
+ * \param[in]    reserved  Must set to 0.
+ *
+ * \retval #AOM_CODEC_OK
+ *     The configuration was populated.
+ * \retval #AOM_CODEC_INCAPABLE
+ *     Interface is not an encoder interface.
+ * \retval #AOM_CODEC_INVALID_PARAM
+ *     A parameter was NULL, or the usage value was not recognized.
+ */
+aom_codec_err_t aom_codec_enc_config_default(aom_codec_iface_t *iface,
+                                             aom_codec_enc_cfg_t *cfg,
+                                             unsigned int reserved);
+
+/*!\brief Set or change configuration
+ *
+ * Reconfigures an encoder instance according to the given configuration.
+ *
+ * \param[in]    ctx     Pointer to this instance's context
+ * \param[in]    cfg     Configuration buffer to use
+ *
+ * \retval #AOM_CODEC_OK
+ *     The configuration was populated.
+ * \retval #AOM_CODEC_INCAPABLE
+ *     Interface is not an encoder interface.
+ * \retval #AOM_CODEC_INVALID_PARAM
+ *     A parameter was NULL, or the usage value was not recognized.
+ */
+aom_codec_err_t aom_codec_enc_config_set(aom_codec_ctx_t *ctx,
+                                         const aom_codec_enc_cfg_t *cfg);
+
+/*!\brief Get global stream headers
+ *
+ * Retrieves a stream level global header packet, if supported by the codec.
+ *
+ * \param[in]    ctx     Pointer to this instance's context
+ *
+ * \retval NULL
+ *     Encoder does not support global header
+ * \retval Non-NULL
+ *     Pointer to buffer containing global header packet
+ */
+aom_fixed_buf_t *aom_codec_get_global_headers(aom_codec_ctx_t *ctx);
+
+/*!\brief deadline parameter analogous to  AVx GOOD QUALITY mode. */
+#define AOM_DL_GOOD_QUALITY (1000000)
+/*!\brief Encode a frame
+ *
+ * Encodes a video frame at the given "presentation time." The presentation
+ * time stamp (PTS) \ref MUST be strictly increasing.
+ *
+ * The encoder supports the notion of a soft real-time deadline. Given a
+ * non-zero value to the deadline parameter, the encoder will make a "best
+ * effort" guarantee to  return before the given time slice expires. It is
+ * implicit that limiting the available time to encode will degrade the
+ * output quality. The encoder can be given an unlimited time to produce the
+ * best possible frame by specifying a deadline of '0'. This deadline
+ * supercedes the AVx notion of "best quality, good quality, realtime".
+ * Applications that wish to map these former settings to the new deadline
+ * based system can use the symbol #AOM_DL_GOOD_QUALITY.
+ *
+ * When the last frame has been passed to the encoder, this function should
+ * continue to be called, with the img parameter set to NULL. This will
+ * signal the end-of-stream condition to the encoder and allow it to encode
+ * any held buffers. Encoding is complete when aom_codec_encode() is called
+ * and aom_codec_get_cx_data() returns no data.
+ *
+ * \param[in]    ctx       Pointer to this instance's context
+ * \param[in]    img       Image data to encode, NULL to flush.
+ * \param[in]    pts       Presentation time stamp, in timebase units.
+ * \param[in]    duration  Duration to show frame, in timebase units.
+ * \param[in]    flags     Flags to use for encoding this frame.
+ * \param[in]    deadline  Time to spend encoding, in microseconds. (0=infinite)
+ *
+ * \retval #AOM_CODEC_OK
+ *     The configuration was populated.
+ * \retval #AOM_CODEC_INCAPABLE
+ *     Interface is not an encoder interface.
+ * \retval #AOM_CODEC_INVALID_PARAM
+ *     A parameter was NULL, the image format is unsupported, etc.
+ */
+aom_codec_err_t aom_codec_encode(aom_codec_ctx_t *ctx, const aom_image_t *img,
+                                 aom_codec_pts_t pts, unsigned long duration,
+                                 aom_enc_frame_flags_t flags,
+                                 unsigned long deadline);
+
+/*!\brief Set compressed data output buffer
+ *
+ * Sets the buffer that the codec should output the compressed data
+ * into. This call effectively sets the buffer pointer returned in the
+ * next AOM_CODEC_CX_FRAME_PKT packet. Subsequent packets will be
+ * appended into this buffer. The buffer is preserved across frames,
+ * so applications must periodically call this function after flushing
+ * the accumulated compressed data to disk or to the network to reset
+ * the pointer to the buffer's head.
+ *
+ * `pad_before` bytes will be skipped before writing the compressed
+ * data, and `pad_after` bytes will be appended to the packet. The size
+ * of the packet will be the sum of the size of the actual compressed
+ * data, pad_before, and pad_after. The padding bytes will be preserved
+ * (not overwritten).
+ *
+ * Note that calling this function does not guarantee that the returned
+ * compressed data will be placed into the specified buffer. In the
+ * event that the encoded data will not fit into the buffer provided,
+ * the returned packet \ref MAY point to an internal buffer, as it would
+ * if this call were never used. In this event, the output packet will
+ * NOT have any padding, and the application must free space and copy it
+ * to the proper place. This is of particular note in configurations
+ * that may output multiple packets for a single encoded frame (e.g., lagged
+ * encoding) or if the application does not reset the buffer periodically.
+ *
+ * Applications may restore the default behavior of the codec providing
+ * the compressed data buffer by calling this function with a NULL
+ * buffer.
+ *
+ * Applications \ref MUSTNOT call this function during iteration of
+ * aom_codec_get_cx_data().
+ *
+ * \param[in]    ctx         Pointer to this instance's context
+ * \param[in]    buf         Buffer to store compressed data into
+ * \param[in]    pad_before  Bytes to skip before writing compressed data
+ * \param[in]    pad_after   Bytes to skip after writing compressed data
+ *
+ * \retval #AOM_CODEC_OK
+ *     The buffer was set successfully.
+ * \retval #AOM_CODEC_INVALID_PARAM
+ *     A parameter was NULL, the image format is unsupported, etc.
+ */
+aom_codec_err_t aom_codec_set_cx_data_buf(aom_codec_ctx_t *ctx,
+                                          const aom_fixed_buf_t *buf,
+                                          unsigned int pad_before,
+                                          unsigned int pad_after);
+
+/*!\brief Encoded data iterator
+ *
+ * Iterates over a list of data packets to be passed from the encoder to the
+ * application. The different kinds of packets available are enumerated in
+ * #aom_codec_cx_pkt_kind.
+ *
+ * #AOM_CODEC_CX_FRAME_PKT packets should be passed to the application's
+ * muxer. Multiple compressed frames may be in the list.
+ * #AOM_CODEC_STATS_PKT packets should be appended to a global buffer.
+ *
+ * The application \ref MUST silently ignore any packet kinds that it does
+ * not recognize or support.
+ *
+ * The data buffers returned from this function are only guaranteed to be
+ * valid until the application makes another call to any aom_codec_* function.
+ *
+ * \param[in]     ctx      Pointer to this instance's context
+ * \param[in,out] iter     Iterator storage, initialized to NULL
+ *
+ * \return Returns a pointer to an output data packet (compressed frame data,
+ *         two-pass statistics, etc.) or NULL to signal end-of-list.
+ *
+ */
+const aom_codec_cx_pkt_t *aom_codec_get_cx_data(aom_codec_ctx_t *ctx,
+                                                aom_codec_iter_t *iter);
+
+/*!\brief Get Preview Frame
+ *
+ * Returns an image that can be used as a preview. Shows the image as it would
+ * exist at the decompressor. The application \ref MUST NOT write into this
+ * image buffer.
+ *
+ * \param[in]     ctx      Pointer to this instance's context
+ *
+ * \return Returns a pointer to a preview image, or NULL if no image is
+ *         available.
+ *
+ */
+const aom_image_t *aom_codec_get_preview_frame(aom_codec_ctx_t *ctx);
+
+/*!@} - end defgroup encoder*/
+#ifdef __cplusplus
+}
+#endif
+#endif  // AOM_AOM_ENCODER_H_
diff --git a/third_party/aom/aom/aom_frame_buffer.h b/third_party/aom/aom/aom_frame_buffer.h
new file mode 100644
index 0000000000..c87cf749b6
--- /dev/null
+++ b/third_party/aom/aom/aom_frame_buffer.h
@@ -0,0 +1,84 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_AOM_FRAME_BUFFER_H_
+#define AOM_AOM_FRAME_BUFFER_H_
+
+/*!\file
+ * \brief Describes the decoder external frame buffer interface.
+ */
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./aom_integer.h"
+
+/*!\brief The maximum number of work buffers used by libaom.
+ *  Support maximum 4 threads to decode video in parallel.
+ *  Each thread will use one work buffer.
+ * TODO(hkuang): Add support to set number of worker threads dynamically.
+ */
+#define AOM_MAXIMUM_WORK_BUFFERS 8
+
+/*!\brief The maximum number of reference buffers that a AV1 encoder may use.
+ */
+#define AOM_MAXIMUM_REF_BUFFERS 8
+
+/*!\brief External frame buffer
+ *
+ * This structure holds allocated frame buffers used by the decoder.
+ */
+typedef struct aom_codec_frame_buffer {
+  uint8_t *data; /**< Pointer to the data buffer */
+  size_t size;   /**< Size of data in bytes */
+  void *priv;    /**< Frame's private data */
+} aom_codec_frame_buffer_t;
+
+/*!\brief get frame buffer callback prototype
+ *
+ * This callback is invoked by the decoder to retrieve data for the frame
+ * buffer in order for the decode call to complete. The callback must
+ * allocate at least min_size in bytes and assign it to fb->data. The callback
+ * must zero out all the data allocated. Then the callback must set fb->size
+ * to the allocated size. The application does not need to align the allocated
+ * data. The callback is triggered when the decoder needs a frame buffer to
+ * decode a compressed image into. This function may be called more than once
+ * for every call to aom_codec_decode. The application may set fb->priv to
+ * some data which will be passed back in the ximage and the release function
+ * call. |fb| is guaranteed to not be NULL. On success the callback must
+ * return 0. Any failure the callback must return a value less than 0.
+ *
+ * \param[in] priv         Callback's private data
+ * \param[in] new_size     Size in bytes needed by the buffer
+ * \param[in,out] fb       Pointer to aom_codec_frame_buffer_t
+ */
+typedef int (*aom_get_frame_buffer_cb_fn_t)(void *priv, size_t min_size,
+                                            aom_codec_frame_buffer_t *fb);
+
+/*!\brief release frame buffer callback prototype
+ *
+ * This callback is invoked by the decoder when the frame buffer is not
+ * referenced by any other buffers. |fb| is guaranteed to not be NULL. On
+ * success the callback must return 0. Any failure the callback must return
+ * a value less than 0.
+ *
+ * \param[in] priv         Callback's private data
+ * \param[in] fb           Pointer to aom_codec_frame_buffer_t
+ */
+typedef int (*aom_release_frame_buffer_cb_fn_t)(void *priv,
+                                                aom_codec_frame_buffer_t *fb);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_AOM_FRAME_BUFFER_H_
diff --git a/third_party/aom/aom/aom_image.h b/third_party/aom/aom/aom_image.h
new file mode 100644
index 0000000000..b2f75e6391
--- /dev/null
+++ b/third_party/aom/aom/aom_image.h
@@ -0,0 +1,225 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/*!\file
+ * \brief Describes the aom image descriptor and associated operations
+ *
+ */
+#ifndef AOM_AOM_IMAGE_H_
+#define AOM_AOM_IMAGE_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*!\brief Current ABI version number
+ *
+ * \internal
+ * If this file is altered in any way that changes the ABI, this value
+ * must be bumped.  Examples include, but are not limited to, changing
+ * types, removing or reassigning enums, adding/removing/rearranging
+ * fields to structures
+ */
+#define AOM_IMAGE_ABI_VERSION (4) /**<\hideinitializer*/
+
+#define AOM_IMG_FMT_PLANAR 0x100       /**< Image is a planar format. */
+#define AOM_IMG_FMT_UV_FLIP 0x200      /**< V plane precedes U in memory. */
+#define AOM_IMG_FMT_HAS_ALPHA 0x400    /**< Image has an alpha channel. */
+#define AOM_IMG_FMT_HIGHBITDEPTH 0x800 /**< Image uses 16bit framebuffer. */
+
+/*!\brief List of supported image formats */
+typedef enum aom_img_fmt {
+  AOM_IMG_FMT_NONE,
+  AOM_IMG_FMT_RGB24,     /**< 24 bit per pixel packed RGB */
+  AOM_IMG_FMT_RGB32,     /**< 32 bit per pixel packed 0RGB */
+  AOM_IMG_FMT_RGB565,    /**< 16 bit per pixel, 565 */
+  AOM_IMG_FMT_RGB555,    /**< 16 bit per pixel, 555 */
+  AOM_IMG_FMT_UYVY,      /**< UYVY packed YUV */
+  AOM_IMG_FMT_YUY2,      /**< YUYV packed YUV */
+  AOM_IMG_FMT_YVYU,      /**< YVYU packed YUV */
+  AOM_IMG_FMT_BGR24,     /**< 24 bit per pixel packed BGR */
+  AOM_IMG_FMT_RGB32_LE,  /**< 32 bit packed BGR0 */
+  AOM_IMG_FMT_ARGB,      /**< 32 bit packed ARGB, alpha=255 */
+  AOM_IMG_FMT_ARGB_LE,   /**< 32 bit packed BGRA, alpha=255 */
+  AOM_IMG_FMT_RGB565_LE, /**< 16 bit per pixel, gggbbbbb rrrrrggg */
+  AOM_IMG_FMT_RGB555_LE, /**< 16 bit per pixel, gggbbbbb 0rrrrrgg */
+  AOM_IMG_FMT_YV12 =
+      AOM_IMG_FMT_PLANAR | AOM_IMG_FMT_UV_FLIP | 1, /**< planar YVU */
+  AOM_IMG_FMT_I420 = AOM_IMG_FMT_PLANAR | 2,
+  AOM_IMG_FMT_AOMYV12 = AOM_IMG_FMT_PLANAR | AOM_IMG_FMT_UV_FLIP |
+                        3, /** < planar 4:2:0 format with aom color space */
+  AOM_IMG_FMT_AOMI420 = AOM_IMG_FMT_PLANAR | 4,
+  AOM_IMG_FMT_I422 = AOM_IMG_FMT_PLANAR | 5,
+  AOM_IMG_FMT_I444 = AOM_IMG_FMT_PLANAR | 6,
+  AOM_IMG_FMT_I440 = AOM_IMG_FMT_PLANAR | 7,
+  AOM_IMG_FMT_444A = AOM_IMG_FMT_PLANAR | AOM_IMG_FMT_HAS_ALPHA | 6,
+  AOM_IMG_FMT_I42016 = AOM_IMG_FMT_I420 | AOM_IMG_FMT_HIGHBITDEPTH,
+  AOM_IMG_FMT_I42216 = AOM_IMG_FMT_I422 | AOM_IMG_FMT_HIGHBITDEPTH,
+  AOM_IMG_FMT_I44416 = AOM_IMG_FMT_I444 | AOM_IMG_FMT_HIGHBITDEPTH,
+  AOM_IMG_FMT_I44016 = AOM_IMG_FMT_I440 | AOM_IMG_FMT_HIGHBITDEPTH
+} aom_img_fmt_t; /**< alias for enum aom_img_fmt */
+
+/*!\brief List of supported color spaces */
+typedef enum aom_color_space {
+  AOM_CS_UNKNOWN = 0,   /**< Unknown */
+  AOM_CS_BT_601 = 1,    /**< BT.601 */
+  AOM_CS_BT_709 = 2,    /**< BT.709 */
+  AOM_CS_SMPTE_170 = 3, /**< SMPTE.170 */
+  AOM_CS_SMPTE_240 = 4, /**< SMPTE.240 */
+  AOM_CS_BT_2020 = 5,   /**< BT.2020 */
+  AOM_CS_RESERVED = 6,  /**< Reserved */
+  AOM_CS_SRGB = 7       /**< sRGB */
+} aom_color_space_t;    /**< alias for enum aom_color_space */
+
+/*!\brief List of supported color range */
+typedef enum aom_color_range {
+  AOM_CR_STUDIO_RANGE = 0, /**< Y [16..235], UV [16..240] */
+  AOM_CR_FULL_RANGE = 1    /**< YUV/RGB [0..255] */
+} aom_color_range_t;       /**< alias for enum aom_color_range */
+
+/**\brief Image Descriptor */
+typedef struct aom_image {
+  aom_img_fmt_t fmt;       /**< Image Format */
+  aom_color_space_t cs;    /**< Color Space */
+  aom_color_range_t range; /**< Color Range */
+
+  /* Image storage dimensions */
+  unsigned int w;         /**< Stored image width */
+  unsigned int h;         /**< Stored image height */
+  unsigned int bit_depth; /**< Stored image bit-depth */
+
+  /* Image display dimensions */
+  unsigned int d_w; /**< Displayed image width */
+  unsigned int d_h; /**< Displayed image height */
+
+  /* Image intended rendering dimensions */
+  unsigned int r_w; /**< Intended rendering image width */
+  unsigned int r_h; /**< Intended rendering image height */
+
+  /* Chroma subsampling info */
+  unsigned int x_chroma_shift; /**< subsampling order, X */
+  unsigned int y_chroma_shift; /**< subsampling order, Y */
+
+/* Image data pointers. */
+#define AOM_PLANE_PACKED 0  /**< To be used for all packed formats */
+#define AOM_PLANE_Y 0       /**< Y (Luminance) plane */
+#define AOM_PLANE_U 1       /**< U (Chroma) plane */
+#define AOM_PLANE_V 2       /**< V (Chroma) plane */
+#define AOM_PLANE_ALPHA 3   /**< A (Transparency) plane */
+  unsigned char *planes[4]; /**< pointer to the top left pixel for each plane */
+  int stride[4];            /**< stride between rows for each plane */
+
+  int bps; /**< bits per sample (for packed formats) */
+
+  /*!\brief The following member may be set by the application to associate
+   * data with this image.
+   */
+  void *user_priv;
+
+  /* The following members should be treated as private. */
+  unsigned char *img_data; /**< private */
+  int img_data_owner;      /**< private */
+  int self_allocd;         /**< private */
+
+  void *fb_priv; /**< Frame buffer data associated with the image. */
+} aom_image_t;   /**< alias for struct aom_image */
+
+/**\brief Representation of a rectangle on a surface */
+typedef struct aom_image_rect {
+  unsigned int x;   /**< leftmost column */
+  unsigned int y;   /**< topmost row */
+  unsigned int w;   /**< width */
+  unsigned int h;   /**< height */
+} aom_image_rect_t; /**< alias for struct aom_image_rect */
+
+/*!\brief Open a descriptor, allocating storage for the underlying image
+ *
+ * Returns a descriptor for storing an image of the given format. The
+ * storage for the descriptor is allocated on the heap.
+ *
+ * \param[in]    img       Pointer to storage for descriptor. If this parameter
+ *                         is NULL, the storage for the descriptor will be
+ *                         allocated on the heap.
+ * \param[in]    fmt       Format for the image
+ * \param[in]    d_w       Width of the image
+ * \param[in]    d_h       Height of the image
+ * \param[in]    align     Alignment, in bytes, of the image buffer and
+ *                         each row in the image(stride).
+ *
+ * \return Returns a pointer to the initialized image descriptor. If the img
+ *         parameter is non-null, the value of the img parameter will be
+ *         returned.
+ */
+aom_image_t *aom_img_alloc(aom_image_t *img, aom_img_fmt_t fmt,
+                           unsigned int d_w, unsigned int d_h,
+                           unsigned int align);
+
+/*!\brief Open a descriptor, using existing storage for the underlying image
+ *
+ * Returns a descriptor for storing an image of the given format. The
+ * storage for descriptor has been allocated elsewhere, and a descriptor is
+ * desired to "wrap" that storage.
+ *
+ * \param[in]    img       Pointer to storage for descriptor. If this parameter
+ *                         is NULL, the storage for the descriptor will be
+ *                         allocated on the heap.
+ * \param[in]    fmt       Format for the image
+ * \param[in]    d_w       Width of the image
+ * \param[in]    d_h       Height of the image
+ * \param[in]    align     Alignment, in bytes, of each row in the image.
+ * \param[in]    img_data  Storage to use for the image
+ *
+ * \return Returns a pointer to the initialized image descriptor. If the img
+ *         parameter is non-null, the value of the img parameter will be
+ *         returned.
+ */
+aom_image_t *aom_img_wrap(aom_image_t *img, aom_img_fmt_t fmt, unsigned int d_w,
+                          unsigned int d_h, unsigned int align,
+                          unsigned char *img_data);
+
+/*!\brief Set the rectangle identifying the displayed portion of the image
+ *
+ * Updates the displayed rectangle (aka viewport) on the image surface to
+ * match the specified coordinates and size.
+ *
+ * \param[in]    img       Image descriptor
+ * \param[in]    x         leftmost column
+ * \param[in]    y         topmost row
+ * \param[in]    w         width
+ * \param[in]    h         height
+ *
+ * \return 0 if the requested rectangle is valid, nonzero otherwise.
+ */
+int aom_img_set_rect(aom_image_t *img, unsigned int x, unsigned int y,
+                     unsigned int w, unsigned int h);
+
+/*!\brief Flip the image vertically (top for bottom)
+ *
+ * Adjusts the image descriptor's pointers and strides to make the image
+ * be referenced upside-down.
+ *
+ * \param[in]    img       Image descriptor
+ */
+void aom_img_flip(aom_image_t *img);
+
+/*!\brief Close an image descriptor
+ *
+ * Frees all allocated storage associated with an image descriptor.
+ *
+ * \param[in]    img       Image descriptor
+ */
+void aom_img_free(aom_image_t *img);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_AOM_IMAGE_H_
diff --git a/third_party/aom/aom/aom_integer.h b/third_party/aom/aom/aom_integer.h
new file mode 100644
index 0000000000..2e8f23f459
--- /dev/null
+++ b/third_party/aom/aom/aom_integer.h
@@ -0,0 +1,64 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_AOM_INTEGER_H_
+#define AOM_AOM_INTEGER_H_
+
+/* get ptrdiff_t, size_t, wchar_t, NULL */
+#include <stddef.h>
+
+#if defined(_MSC_VER)
+#define AOM_FORCE_INLINE __forceinline
+#define AOM_INLINE __inline
+#else
+#define AOM_FORCE_INLINE __inline__ __attribute__((always_inline))
+// TODO(jbb): Allow a way to force inline off for older compilers.
+#define AOM_INLINE inline
+#endif
+
+#if defined(AOM_EMULATE_INTTYPES)
+typedef signed char int8_t;
+typedef signed short int16_t;
+typedef signed int int32_t;
+
+typedef unsigned char uint8_t;
+typedef unsigned short uint16_t;
+typedef unsigned int uint32_t;
+
+#ifndef _UINTPTR_T_DEFINED
+typedef size_t uintptr_t;
+#endif
+
+#else
+
+/* Most platforms have the C99 standard integer types. */
+
+#if defined(__cplusplus)
+#if !defined(__STDC_FORMAT_MACROS)
+#define __STDC_FORMAT_MACROS
+#endif
+#if !defined(__STDC_LIMIT_MACROS)
+#define __STDC_LIMIT_MACROS
+#endif
+#endif  // __cplusplus
+
+#include <stdint.h>
+
+#endif
+
+/* VS2010 defines stdint.h, but not inttypes.h */
+#if defined(_MSC_VER) && _MSC_VER < 1800
+#define PRId64 "I64d"
+#else
+#include <inttypes.h>
+#endif
+
+#endif  // AOM_AOM_INTEGER_H_
diff --git a/third_party/aom/aom/aomcx.h b/third_party/aom/aom/aomcx.h
new file mode 100644
index 0000000000..debec21b09
--- /dev/null
+++ b/third_party/aom/aom/aomcx.h
@@ -0,0 +1,750 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_AOMCX_H_
+#define AOM_AOMCX_H_
+
+/*!\defgroup aom_encoder AOMedia AOM/AV1 Encoder
+ * \ingroup aom
+ *
+ * @{
+ */
+#include "./aom.h"
+#include "./aom_encoder.h"
+
+/*!\file
+ * \brief Provides definitions for using AOM or AV1 encoder algorithm within the
+ *        aom Codec Interface.
+ */
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*!\name Algorithm interface for AV1
+ *
+ * This interface provides the capability to encode raw AV1 streams.
+ * @{
+ */
+extern aom_codec_iface_t aom_codec_av1_cx_algo;
+extern aom_codec_iface_t *aom_codec_av1_cx(void);
+/*!@} - end algorithm interface member group*/
+
+/*
+ * Algorithm Flags
+ */
+
+/*!\brief Don't reference the last frame
+ *
+ * When this flag is set, the encoder will not use the last frame as a
+ * predictor. When not set, the encoder will choose whether to use the
+ * last frame or not automatically.
+ */
+#define AOM_EFLAG_NO_REF_LAST (1 << 16)
+
+/*!\brief Don't reference the golden frame
+ *
+ * When this flag is set, the encoder will not use the golden frame as a
+ * predictor. When not set, the encoder will choose whether to use the
+ * golden frame or not automatically.
+ */
+#define AOM_EFLAG_NO_REF_GF (1 << 17)
+
+/*!\brief Don't reference the alternate reference frame
+ *
+ * When this flag is set, the encoder will not use the alt ref frame as a
+ * predictor. When not set, the encoder will choose whether to use the
+ * alt ref frame or not automatically.
+ */
+#define AOM_EFLAG_NO_REF_ARF (1 << 21)
+
+/*!\brief Don't update the last frame
+ *
+ * When this flag is set, the encoder will not update the last frame with
+ * the contents of the current frame.
+ */
+#define AOM_EFLAG_NO_UPD_LAST (1 << 18)
+
+/*!\brief Don't update the golden frame
+ *
+ * When this flag is set, the encoder will not update the golden frame with
+ * the contents of the current frame.
+ */
+#define AOM_EFLAG_NO_UPD_GF (1 << 22)
+
+/*!\brief Don't update the alternate reference frame
+ *
+ * When this flag is set, the encoder will not update the alt ref frame with
+ * the contents of the current frame.
+ */
+#define AOM_EFLAG_NO_UPD_ARF (1 << 23)
+
+/*!\brief Force golden frame update
+ *
+ * When this flag is set, the encoder copy the contents of the current frame
+ * to the golden frame buffer.
+ */
+#define AOM_EFLAG_FORCE_GF (1 << 19)
+
+/*!\brief Force alternate reference frame update
+ *
+ * When this flag is set, the encoder copy the contents of the current frame
+ * to the alternate reference frame buffer.
+ */
+#define AOM_EFLAG_FORCE_ARF (1 << 24)
+
+/*!\brief Disable entropy update
+ *
+ * When this flag is set, the encoder will not update its internal entropy
+ * model based on the entropy of this frame.
+ */
+#define AOM_EFLAG_NO_UPD_ENTROPY (1 << 20)
+
+/*!\brief AVx encoder control functions
+ *
+ * This set of macros define the control functions available for AVx
+ * encoder interface.
+ *
+ * \sa #aom_codec_control
+ */
+enum aome_enc_control_id {
+  /*!\brief Codec control function to set which reference frame encoder can use.
+   */
+  AOME_USE_REFERENCE = 7,
+
+  /*!\brief Codec control function to pass an ROI map to encoder.
+   */
+  AOME_SET_ROI_MAP = 8,
+
+  /*!\brief Codec control function to pass an Active map to encoder.
+   */
+  AOME_SET_ACTIVEMAP,
+
+  /*!\brief Codec control function to set encoder scaling mode.
+   */
+  AOME_SET_SCALEMODE = 11,
+
+  /*!\brief Codec control function to set encoder internal speed settings.
+   *
+   * Changes in this value influences, among others, the encoder's selection
+   * of motion estimation methods. Values greater than 0 will increase encoder
+   * speed at the expense of quality.
+   *
+   * \note Valid range: 0..8
+   */
+  AOME_SET_CPUUSED = 13,
+
+  /*!\brief Codec control function to enable automatic set and use alf frames.
+   */
+  AOME_SET_ENABLEAUTOALTREF,
+
+  /*!\brief Codec control function to set sharpness.
+   */
+  AOME_SET_SHARPNESS = AOME_SET_ENABLEAUTOALTREF + 2,
+
+  /*!\brief Codec control function to set the threshold for MBs treated static.
+   */
+  AOME_SET_STATIC_THRESHOLD,
+
+  /*!\brief Codec control function to get last quantizer chosen by the encoder.
+   *
+   * Return value uses internal quantizer scale defined by the codec.
+   */
+  AOME_GET_LAST_QUANTIZER = AOME_SET_STATIC_THRESHOLD + 2,
+
+  /*!\brief Codec control function to get last quantizer chosen by the encoder.
+   *
+   * Return value uses the 0..63 scale as used by the rc_*_quantizer config
+   * parameters.
+   */
+  AOME_GET_LAST_QUANTIZER_64,
+
+  /*!\brief Codec control function to set the max no of frames to create arf.
+   */
+  AOME_SET_ARNR_MAXFRAMES,
+
+  /*!\brief Codec control function to set the filter strength for the arf.
+   */
+  AOME_SET_ARNR_STRENGTH,
+
+  /*!\brief Codec control function to set visual tuning.
+   */
+  AOME_SET_TUNING = AOME_SET_ARNR_STRENGTH + 2,
+
+  /*!\brief Codec control function to set constrained quality level.
+   *
+   * \attention For this value to be used aom_codec_enc_cfg_t::g_usage must be
+   *            set to #AOM_CQ.
+   * \note Valid range: 0..63
+   */
+  AOME_SET_CQ_LEVEL,
+
+  /*!\brief Codec control function to set Max data rate for Intra frames.
+   *
+   * This value controls additional clamping on the maximum size of a
+   * keyframe. It is expressed as a percentage of the average
+   * per-frame bitrate, with the special (and default) value 0 meaning
+   * unlimited, or no additional clamping beyond the codec's built-in
+   * algorithm.
+   *
+   * For example, to allocate no more than 4.5 frames worth of bitrate
+   * to a keyframe, set this to 450.
+   */
+  AOME_SET_MAX_INTRA_BITRATE_PCT,
+
+  /*!\brief Codec control function to set max data rate for Inter frames.
+   *
+   * This value controls additional clamping on the maximum size of an
+   * inter frame. It is expressed as a percentage of the average
+   * per-frame bitrate, with the special (and default) value 0 meaning
+   * unlimited, or no additional clamping beyond the codec's built-in
+   * algorithm.
+   *
+   * For example, to allow no more than 4.5 frames worth of bitrate
+   * to an inter frame, set this to 450.
+   */
+  AV1E_SET_MAX_INTER_BITRATE_PCT = AOME_SET_MAX_INTRA_BITRATE_PCT + 2,
+
+  /*!\brief Boost percentage for Golden Frame in CBR mode.
+   *
+   * This value controls the amount of boost given to Golden Frame in
+   * CBR mode. It is expressed as a percentage of the average
+   * per-frame bitrate, with the special (and default) value 0 meaning
+   * the feature is off, i.e., no golden frame boost in CBR mode and
+   * average bitrate target is used.
+   *
+   * For example, to allow 100% more bits, i.e, 2X, in a golden frame
+   * than average frame, set this to 100.
+   */
+  AV1E_SET_GF_CBR_BOOST_PCT,
+
+  /*!\brief Codec control function to set lossless encoding mode.
+   *
+   * AV1 can operate in lossless encoding mode, in which the bitstream
+   * produced will be able to decode and reconstruct a perfect copy of
+   * input source. This control function provides a mean to switch encoder
+   * into lossless coding mode(1) or normal coding mode(0) that may be lossy.
+   *                          0 = lossy coding mode
+   *                          1 = lossless coding mode
+   *
+   *  By default, encoder operates in normal coding mode (maybe lossy).
+   */
+  AV1E_SET_LOSSLESS = AV1E_SET_GF_CBR_BOOST_PCT + 2,
+
+  /*!\brief Codec control function to set number of tile columns.
+   *
+   * In encoding and decoding, AV1 allows an input image frame be partitioned
+   * into separated vertical tile columns, which can be encoded or decoded
+   * independently. This enables easy implementation of parallel encoding and
+   * decoding. This control requests the encoder to use column tiles in
+   * encoding an input frame, with number of tile columns (in Log2 unit) as
+   * the parameter:
+   *             0 = 1 tile column
+   *             1 = 2 tile columns
+   *             2 = 4 tile columns
+   *             .....
+   *             n = 2**n tile columns
+   * The requested tile columns will be capped by encoder based on image size
+   * limitation (The minimum width of a tile column is 256 pixel, the maximum
+   * is 4096).
+   *
+   * By default, the value is 0, i.e. one single column tile for entire image.
+   */
+  AV1E_SET_TILE_COLUMNS,
+
+  /*!\brief Codec control function to set number of tile rows.
+   *
+   * In encoding and decoding, AV1 allows an input image frame be partitioned
+   * into separated horizontal tile rows. Tile rows are encoded or decoded
+   * sequentially. Even though encoding/decoding of later tile rows depends on
+   * earlier ones, this allows the encoder to output data packets for tile rows
+   * prior to completely processing all tile rows in a frame, thereby reducing
+   * the latency in processing between input and output. The parameter
+   * for this control describes the number of tile rows, which has a valid
+   * range [0, 2]:
+   *            0 = 1 tile row
+   *            1 = 2 tile rows
+   *            2 = 4 tile rows
+   *
+   * By default, the value is 0, i.e. one single row tile for entire image.
+   */
+  AV1E_SET_TILE_ROWS,
+
+  /*!\brief Codec control function to enable frame parallel decoding feature.
+   *
+   * AV1 has a bitstream feature to reduce decoding dependency between frames
+   * by turning off backward update of probability context used in encoding
+   * and decoding. This allows staged parallel processing of more than one
+   * video frames in the decoder. This control function provides a mean to
+   * turn this feature on or off for bitstreams produced by encoder.
+   *
+   * By default, this feature is off.
+   */
+  AV1E_SET_FRAME_PARALLEL_DECODING,
+
+  /*!\brief Codec control function to set adaptive quantization mode.
+   *
+   * AV1 has a segment based feature that allows encoder to adaptively change
+   * quantization parameter for each segment within a frame to improve the
+   * subjective quality. This control makes encoder operate in one of the
+   * several AQ_modes supported.
+   *
+   * By default, encoder operates with AQ_Mode 0(adaptive quantization off).
+   */
+  AV1E_SET_AQ_MODE,
+
+  /*!\brief Codec control function to enable/disable periodic Q boost.
+   *
+   * One AV1 encoder speed feature is to enable quality boost by lowering
+   * frame level Q periodically. This control function provides a mean to
+   * turn on/off this feature.
+   *               0 = off
+   *               1 = on
+   *
+   * By default, the encoder is allowed to use this feature for appropriate
+   * encoding modes.
+   */
+  AV1E_SET_FRAME_PERIODIC_BOOST,
+
+  /*!\brief Codec control function to set noise sensitivity.
+   *
+   *  0: off, 1: On(YOnly)
+   */
+  AV1E_SET_NOISE_SENSITIVITY,
+
+  /*!\brief Codec control function to set content type.
+   * \note Valid parameter range:
+   *              AOM_CONTENT_DEFAULT = Regular video content (Default)
+   *              AOM_CONTENT_SCREEN  = Screen capture content
+   */
+  AV1E_SET_TUNE_CONTENT,
+
+  /*!\brief Codec control function to set color space info.
+   * \note Valid ranges: 0..7, default is "UNKNOWN".
+   *                     0 = UNKNOWN,
+   *                     1 = BT_601
+   *                     2 = BT_709
+   *                     3 = SMPTE_170
+   *                     4 = SMPTE_240
+   *                     5 = BT_2020
+   *                     6 = RESERVED
+   *                     7 = SRGB
+   */
+  AV1E_SET_COLOR_SPACE,
+
+  /*!\brief Codec control function to set minimum interval between GF/ARF frames
+   *
+   * By default the value is set as 4.
+   */
+  AV1E_SET_MIN_GF_INTERVAL,
+
+  /*!\brief Codec control function to set minimum interval between GF/ARF frames
+   *
+   * By default the value is set as 16.
+   */
+  AV1E_SET_MAX_GF_INTERVAL,
+
+  /*!\brief Codec control function to get an Active map back from the encoder.
+   */
+  AV1E_GET_ACTIVEMAP,
+
+  /*!\brief Codec control function to set color range bit.
+   * \note Valid ranges: 0..1, default is 0
+   *                     0 = Limited range (16..235 or HBD equivalent)
+   *                     1 = Full range (0..255 or HBD equivalent)
+   */
+  AV1E_SET_COLOR_RANGE,
+
+  /*!\brief Codec control function to set intended rendering image size.
+   *
+   * By default, this is identical to the image size in pixels.
+   */
+  AV1E_SET_RENDER_SIZE,
+
+  /*!\brief Codec control function to set target level.
+   *
+   * 255: off (default); 0: only keep level stats; 10: target for level 1.0;
+   * 11: target for level 1.1; ... 62: target for level 6.2
+   */
+  AV1E_SET_TARGET_LEVEL,
+
+  /*!\brief Codec control function to get bitstream level.
+   */
+  AV1E_GET_LEVEL,
+
+  /*!\brief Codec control function to set intended superblock size.
+   *
+   * By default, the superblock size is determined separately for each
+   * frame by the encoder.
+   *
+   * Experiment: EXT_PARTITION
+   */
+  AV1E_SET_SUPERBLOCK_SIZE,
+
+  /*!\brief Codec control function to enable automatic set and use
+   * bwd-pred frames.
+   *
+   * Experiment: EXT_REFS
+   */
+  AOME_SET_ENABLEAUTOBWDREF,
+
+  /*!\brief Codec control function to encode with quantisation matrices.
+   *
+   * AOM can operate with default quantisation matrices dependent on
+   * quantisation level and block type.
+   *                          0 = do not use quantisation matrices
+   *                          1 = use quantisation matrices
+   *
+   *  By default, the encoder operates without quantisation matrices.
+   *
+   * Experiment: AOM_QM
+   */
+  AV1E_SET_ENABLE_QM,
+
+  /*!\brief Codec control function to set the min quant matrix flatness.
+   *
+   * AOM can operate with different ranges of quantisation matrices.
+   * As quantisation levels increase, the matrices get flatter. This
+   * control sets the minimum level of flatness from which the matrices
+   * are determined.
+   *
+   *  By default, the encoder sets this minimum at half the available
+   *  range.
+   *
+   * Experiment: AOM_QM
+   */
+  AV1E_SET_QM_MIN,
+
+  /*!\brief Codec control function to set the max quant matrix flatness.
+   *
+   * AOM can operate with different ranges of quantisation matrices.
+   * As quantisation levels increase, the matrices get flatter. This
+   * control sets the maximum level of flatness possible.
+   *
+   * By default, the encoder sets this maximum at the top of the
+   * available range.
+   *
+   * Experiment: AOM_QM
+   */
+  AV1E_SET_QM_MAX,
+
+  /*!\brief Codec control function to set a maximum number of tile groups.
+   *
+   * This will set the maximum number of tile groups. This will be
+   * overridden if an MTU size is set. The default value is 1.
+   *
+   * Experiment: TILE_GROUPS
+   */
+  AV1E_SET_NUM_TG,
+
+  /*!\brief Codec control function to set an MTU size for a tile group.
+   *
+   * This will set the maximum number of bytes in a tile group. This can be
+   * exceeded only if a single tile is larger than this amount.
+   *
+   * By default, the value is 0, in which case a fixed number of tile groups
+   * is used.
+   *
+   * Experiment: TILE_GROUPS
+   */
+  AV1E_SET_MTU,
+
+  /*!\brief Codec control function to set dependent_horz_tiles.
+  *
+  * In encoding and decoding, AV1 allows enabling dependent horizontal tile
+  * The parameter for this control describes the value of this flag,
+  * which has a valid range [0, 1]:
+  *            0 = disable dependent horizontal tile
+  *            1 = enable dependent horizontal tile,
+  *
+  * By default, the value is 0, i.e. disable dependent horizontal tile.
+  */
+  AV1E_SET_TILE_DEPENDENT_ROWS,
+
+  /*!\brief Codec control function to set the number of symbols in an ANS data
+   * window.
+   *
+   * The number of ANS symbols (both boolean and non-booleans alphabets) in an
+   * ANS data window is set to 1 << value.
+   *
+   * \note Valid range: [8, 23]
+   *
+   * Experiment: ANS
+   */
+  AV1E_SET_ANS_WINDOW_SIZE_LOG2,
+
+  /*!\brief Codec control function to set temporal mv prediction
+  * enabling/disabling.
+  *
+  * This will enable or disable temporal mv predicton. The default value is 0.
+  *
+  * Experiment: TEMPMV_SIGNALING
+  */
+  AV1E_SET_DISABLE_TEMPMV,
+
+  /*!\brief Codec control function to set loop_filter_across_tiles_enabled.
+   *
+   * In encoding and decoding, AV1 allows disabling loop filter across tile
+   * boundary The parameter for this control describes the value of this flag,
+   * which has a valid range [0, 1]:
+   *            0 = disable loop filter across tile boundary
+   *            1 = enable loop filter across tile boundary
+   *
+   * By default, the value is 1, i.e. enable loop filter across tile boundary.
+   *
+   * Experiment: LOOPFILTERING_ACROSS_TILES
+   */
+  AV1E_SET_TILE_LOOPFILTER,
+
+  /*!\brief Codec control function to set the delta q mode
+  *
+  * AV1 has a segment based feature that allows encoder to adaptively change
+  * quantization parameter for each segment within a frame to improve the
+  * subjective quality. the delta q mode is added on top of segment based
+  * feature, and allows control per 64x64 q and lf delta.This control makes
+  * encoder operate in one of the several DELTA_Q_modes supported.
+  *
+  * By default, encoder operates with DELTAQ_Mode 0(deltaq signaling off).
+  */
+  AV1E_SET_DELTAQ_MODE,
+
+  /*!\brief Codec control function to set the tile encoding mode to 0 or 1.
+   *
+   * 0 means that the tile encoding mode is TILE_NORMAL, and 1 means that the
+   * tile encoding mode is TILE_VR.
+   *
+   * Experiment: EXT_TILE
+   */
+  AV1E_SET_TILE_ENCODING_MODE,
+
+  /*!\brief Codec control function to enable the extreme motion vector unit test
+   * in AV1. Please note that this is only used in motion vector unit test.
+   *
+   * 0 : off, 1 : MAX_EXTREME_MV, 2 : MIN_EXTREME_MV
+   */
+  AV1E_ENABLE_MOTION_VECTOR_UNIT_TEST,
+};
+
+/*!\brief aom 1-D scaling mode
+ *
+ * This set of constants define 1-D aom scaling modes
+ */
+typedef enum aom_scaling_mode_1d {
+  AOME_NORMAL = 0,
+  AOME_FOURFIVE = 1,
+  AOME_THREEFIVE = 2,
+  AOME_ONETWO = 3
+} AOM_SCALING_MODE;
+
+/*!\brief  aom region of interest map
+ *
+ * These defines the data structures for the region of interest map
+ *
+ */
+
+typedef struct aom_roi_map {
+  /*! An id between 0 and 3 for each 16x16 region within a frame. */
+  unsigned char *roi_map;
+  unsigned int rows; /**< Number of rows. */
+  unsigned int cols; /**< Number of columns. */
+  // TODO(paulwilkins): broken for AV1 which has 8 segments
+  // q and loop filter deltas for each segment
+  // (see MAX_MB_SEGMENTS)
+  int delta_q[4];  /**< Quantizer deltas. */
+  int delta_lf[4]; /**< Loop filter deltas. */
+  /*! Static breakout threshold for each segment. */
+  unsigned int static_threshold[4];
+} aom_roi_map_t;
+
+/*!\brief  aom active region map
+ *
+ * These defines the data structures for active region map
+ *
+ */
+
+typedef struct aom_active_map {
+  /*!\brief specify an on (1) or off (0) each 16x16 region within a frame */
+  unsigned char *active_map;
+  unsigned int rows; /**< number of rows */
+  unsigned int cols; /**< number of cols */
+} aom_active_map_t;
+
+/*!\brief  aom image scaling mode
+ *
+ * This defines the data structure for image scaling mode
+ *
+ */
+typedef struct aom_scaling_mode {
+  AOM_SCALING_MODE h_scaling_mode; /**< horizontal scaling mode */
+  AOM_SCALING_MODE v_scaling_mode; /**< vertical scaling mode   */
+} aom_scaling_mode_t;
+
+/*!brief AV1 encoder content type */
+typedef enum {
+  AOM_CONTENT_DEFAULT,
+  AOM_CONTENT_SCREEN,
+  AOM_CONTENT_INVALID
+} aom_tune_content;
+
+/*!\brief Model tuning parameters
+ *
+ * Changes the encoder to tune for certain types of input material.
+ *
+ */
+typedef enum { AOM_TUNE_PSNR, AOM_TUNE_SSIM } aom_tune_metric;
+
+/*!\cond */
+/*!\brief Encoder control function parameter type
+ *
+ * Defines the data types that AOME/AV1E control functions take. Note that
+ * additional common controls are defined in aom.h
+ *
+ */
+
+AOM_CTRL_USE_TYPE_DEPRECATED(AOME_USE_REFERENCE, int)
+#define AOM_CTRL_AOME_USE_REFERENCE
+AOM_CTRL_USE_TYPE(AOME_SET_ROI_MAP, aom_roi_map_t *)
+#define AOM_CTRL_AOME_SET_ROI_MAP
+AOM_CTRL_USE_TYPE(AOME_SET_ACTIVEMAP, aom_active_map_t *)
+#define AOM_CTRL_AOME_SET_ACTIVEMAP
+AOM_CTRL_USE_TYPE(AOME_SET_SCALEMODE, aom_scaling_mode_t *)
+#define AOM_CTRL_AOME_SET_SCALEMODE
+
+AOM_CTRL_USE_TYPE(AOME_SET_CPUUSED, int)
+#define AOM_CTRL_AOME_SET_CPUUSED
+AOM_CTRL_USE_TYPE(AOME_SET_ENABLEAUTOALTREF, unsigned int)
+#define AOM_CTRL_AOME_SET_ENABLEAUTOALTREF
+
+AOM_CTRL_USE_TYPE(AOME_SET_ENABLEAUTOBWDREF, unsigned int)
+#define AOM_CTRL_AOME_SET_ENABLEAUTOBWDREF
+
+AOM_CTRL_USE_TYPE(AOME_SET_SHARPNESS, unsigned int)
+#define AOM_CTRL_AOME_SET_SHARPNESS
+AOM_CTRL_USE_TYPE(AOME_SET_STATIC_THRESHOLD, unsigned int)
+#define AOM_CTRL_AOME_SET_STATIC_THRESHOLD
+
+AOM_CTRL_USE_TYPE(AOME_SET_ARNR_MAXFRAMES, unsigned int)
+#define AOM_CTRL_AOME_SET_ARNR_MAXFRAMES
+AOM_CTRL_USE_TYPE(AOME_SET_ARNR_STRENGTH, unsigned int)
+#define AOM_CTRL_AOME_SET_ARNR_STRENGTH
+AOM_CTRL_USE_TYPE(AOME_SET_TUNING, int) /* aom_tune_metric */
+#define AOM_CTRL_AOME_SET_TUNING
+AOM_CTRL_USE_TYPE(AOME_SET_CQ_LEVEL, unsigned int)
+#define AOM_CTRL_AOME_SET_CQ_LEVEL
+
+AOM_CTRL_USE_TYPE(AV1E_SET_TILE_COLUMNS, int)
+#define AOM_CTRL_AV1E_SET_TILE_COLUMNS
+AOM_CTRL_USE_TYPE(AV1E_SET_TILE_ROWS, int)
+#define AOM_CTRL_AV1E_SET_TILE_ROWS
+
+AOM_CTRL_USE_TYPE(AV1E_SET_TILE_DEPENDENT_ROWS, int)
+#define AOM_CTRL_AV1E_SET_TILE_DEPENDENT_ROWS
+
+AOM_CTRL_USE_TYPE(AV1E_SET_TILE_LOOPFILTER, int)
+#define AOM_CTRL_AV1E_SET_TILE_LOOPFILTER
+
+AOM_CTRL_USE_TYPE(AOME_GET_LAST_QUANTIZER, int *)
+#define AOM_CTRL_AOME_GET_LAST_QUANTIZER
+AOM_CTRL_USE_TYPE(AOME_GET_LAST_QUANTIZER_64, int *)
+#define AOM_CTRL_AOME_GET_LAST_QUANTIZER_64
+
+AOM_CTRL_USE_TYPE(AOME_SET_MAX_INTRA_BITRATE_PCT, unsigned int)
+#define AOM_CTRL_AOME_SET_MAX_INTRA_BITRATE_PCT
+AOM_CTRL_USE_TYPE(AOME_SET_MAX_INTER_BITRATE_PCT, unsigned int)
+#define AOM_CTRL_AOME_SET_MAX_INTER_BITRATE_PCT
+
+AOM_CTRL_USE_TYPE(AV1E_SET_GF_CBR_BOOST_PCT, unsigned int)
+#define AOM_CTRL_AV1E_SET_GF_CBR_BOOST_PCT
+
+AOM_CTRL_USE_TYPE(AV1E_SET_LOSSLESS, unsigned int)
+#define AOM_CTRL_AV1E_SET_LOSSLESS
+
+AOM_CTRL_USE_TYPE(AV1E_SET_ENABLE_QM, unsigned int)
+#define AOM_CTRL_AV1E_SET_ENABLE_QM
+
+AOM_CTRL_USE_TYPE(AV1E_SET_QM_MIN, unsigned int)
+#define AOM_CTRL_AV1E_SET_QM_MIN
+
+AOM_CTRL_USE_TYPE(AV1E_SET_QM_MAX, unsigned int)
+#define AOM_CTRL_AV1E_SET_QM_MAX
+
+AOM_CTRL_USE_TYPE(AV1E_SET_NUM_TG, unsigned int)
+#define AOM_CTRL_AV1E_SET_NUM_TG
+AOM_CTRL_USE_TYPE(AV1E_SET_MTU, unsigned int)
+#define AOM_CTRL_AV1E_SET_MTU
+
+AOM_CTRL_USE_TYPE(AV1E_SET_DISABLE_TEMPMV, unsigned int)
+#define AOM_CTRL_AV1E_SET_DISABLE_TEMPMV
+
+AOM_CTRL_USE_TYPE(AV1E_SET_FRAME_PARALLEL_DECODING, unsigned int)
+#define AOM_CTRL_AV1E_SET_FRAME_PARALLEL_DECODING
+
+AOM_CTRL_USE_TYPE(AV1E_SET_AQ_MODE, unsigned int)
+#define AOM_CTRL_AV1E_SET_AQ_MODE
+
+AOM_CTRL_USE_TYPE(AV1E_SET_DELTAQ_MODE, unsigned int)
+#define AOM_CTRL_AV1E_SET_DELTAQ_MODE
+
+AOM_CTRL_USE_TYPE(AV1E_SET_FRAME_PERIODIC_BOOST, unsigned int)
+#define AOM_CTRL_AV1E_SET_FRAME_PERIODIC_BOOST
+
+AOM_CTRL_USE_TYPE(AV1E_SET_NOISE_SENSITIVITY, unsigned int)
+#define AOM_CTRL_AV1E_SET_NOISE_SENSITIVITY
+
+AOM_CTRL_USE_TYPE(AV1E_SET_TUNE_CONTENT, int) /* aom_tune_content */
+#define AOM_CTRL_AV1E_SET_TUNE_CONTENT
+
+AOM_CTRL_USE_TYPE(AV1E_SET_COLOR_SPACE, int)
+#define AOM_CTRL_AV1E_SET_COLOR_SPACE
+
+AOM_CTRL_USE_TYPE(AV1E_SET_MIN_GF_INTERVAL, unsigned int)
+#define AOM_CTRL_AV1E_SET_MIN_GF_INTERVAL
+
+AOM_CTRL_USE_TYPE(AV1E_SET_MAX_GF_INTERVAL, unsigned int)
+#define AOM_CTRL_AV1E_SET_MAX_GF_INTERVAL
+
+AOM_CTRL_USE_TYPE(AV1E_GET_ACTIVEMAP, aom_active_map_t *)
+#define AOM_CTRL_AV1E_GET_ACTIVEMAP
+
+AOM_CTRL_USE_TYPE(AV1E_SET_COLOR_RANGE, int)
+#define AOM_CTRL_AV1E_SET_COLOR_RANGE
+
+/*!\brief
+ *
+ * TODO(rbultje) : add support of the control in ffmpeg
+ */
+#define AOM_CTRL_AV1E_SET_RENDER_SIZE
+AOM_CTRL_USE_TYPE(AV1E_SET_RENDER_SIZE, int *)
+
+AOM_CTRL_USE_TYPE(AV1E_SET_SUPERBLOCK_SIZE, unsigned int)
+#define AOM_CTRL_AV1E_SET_SUPERBLOCK_SIZE
+
+AOM_CTRL_USE_TYPE(AV1E_SET_TARGET_LEVEL, unsigned int)
+#define AOM_CTRL_AV1E_SET_TARGET_LEVEL
+
+AOM_CTRL_USE_TYPE(AV1E_GET_LEVEL, int *)
+#define AOM_CTRL_AV1E_GET_LEVEL
+
+AOM_CTRL_USE_TYPE(AV1E_SET_ANS_WINDOW_SIZE_LOG2, unsigned int)
+#define AOM_CTRL_AV1E_SET_ANS_WINDOW_SIZE_LOG2
+
+AOM_CTRL_USE_TYPE(AV1E_SET_TILE_ENCODING_MODE, unsigned int)
+#define AOM_CTRL_AV1E_SET_TILE_ENCODING_MODE
+
+AOM_CTRL_USE_TYPE(AV1E_ENABLE_MOTION_VECTOR_UNIT_TEST, unsigned int)
+#define AOM_CTRL_AV1E_ENABLE_MOTION_VECTOR_UNIT_TEST
+
+/*!\endcond */
+/*! @} - end defgroup aom_encoder */
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_AOMCX_H_
diff --git a/third_party/aom/aom/aomdx.h b/third_party/aom/aom/aomdx.h
new file mode 100644
index 0000000000..99f50c17ea
--- /dev/null
+++ b/third_party/aom/aom/aomdx.h
@@ -0,0 +1,219 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/*!\defgroup aom_decoder AOMedia AOM/AV1 Decoder
+ * \ingroup aom
+ *
+ * @{
+ */
+/*!\file
+ * \brief Provides definitions for using AOM or AV1 within the aom Decoder
+ *        interface.
+ */
+#ifndef AOM_AOMDX_H_
+#define AOM_AOMDX_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Include controls common to both the encoder and decoder */
+#include "./aom.h"
+
+/*!\name Algorithm interface for AV1
+ *
+ * This interface provides the capability to decode AV1 streams.
+ * @{
+ */
+extern aom_codec_iface_t aom_codec_av1_dx_algo;
+extern aom_codec_iface_t *aom_codec_av1_dx(void);
+/*!@} - end algorithm interface member group*/
+
+/** Data structure that stores bit accounting for debug
+ */
+typedef struct Accounting Accounting;
+
+/** Callback that inspects decoder frame data.
+ */
+typedef void (*aom_inspect_cb)(void *decoder, void *ctx);
+
+/*!\brief Structure to hold inspection callback and context.
+ *
+ * Defines a structure to hold the inspection callback function and calling
+ * context.
+ */
+typedef struct aom_inspect_init {
+  /*! Inspection callback. */
+  aom_inspect_cb inspect_cb;
+
+  /*! Inspection context. */
+  void *inspect_ctx;
+} aom_inspect_init;
+
+/*!\enum aom_dec_control_id
+ * \brief AOM decoder control functions
+ *
+ * This set of macros define the control functions available for the AOM
+ * decoder interface.
+ *
+ * \sa #aom_codec_control
+ */
+enum aom_dec_control_id {
+  /** control function to get info on which reference frames were updated
+   *  by the last decode
+   */
+  AOMD_GET_LAST_REF_UPDATES = AOM_DECODER_CTRL_ID_START,
+
+  /** check if the indicated frame is corrupted */
+  AOMD_GET_FRAME_CORRUPTED,
+
+  /** control function to get info on which reference frames were used
+   *  by the last decode
+   */
+  AOMD_GET_LAST_REF_USED,
+
+  /** decryption function to decrypt encoded buffer data immediately
+   * before decoding. Takes a aom_decrypt_init, which contains
+   * a callback function and opaque context pointer.
+   */
+  AOMD_SET_DECRYPTOR,
+  // AOMD_SET_DECRYPTOR = AOMD_SET_DECRYPTOR,
+
+  /** control function to get the dimensions that the current frame is decoded
+   * at. This may be different to the intended display size for the frame as
+   * specified in the wrapper or frame header (see AV1D_GET_DISPLAY_SIZE). */
+  AV1D_GET_FRAME_SIZE,
+
+  /** control function to get the current frame's intended display dimensions
+   * (as specified in the wrapper or frame header). This may be different to
+   * the decoded dimensions of this frame (see AV1D_GET_FRAME_SIZE). */
+  AV1D_GET_DISPLAY_SIZE,
+
+  /** control function to get the bit depth of the stream. */
+  AV1D_GET_BIT_DEPTH,
+
+  /** control function to set the byte alignment of the planes in the reference
+   * buffers. Valid values are power of 2, from 32 to 1024. A value of 0 sets
+   * legacy alignment. I.e. Y plane is aligned to 32 bytes, U plane directly
+   * follows Y plane, and V plane directly follows U plane. Default value is 0.
+   */
+  AV1_SET_BYTE_ALIGNMENT,
+
+  /** control function to invert the decoding order to from right to left. The
+   * function is used in a test to confirm the decoding independence of tile
+   * columns. The function may be used in application where this order
+   * of decoding is desired.
+   *
+   * TODO(yaowu): Rework the unit test that uses this control, and in a future
+   *              release, this test-only control shall be removed.
+   */
+  AV1_INVERT_TILE_DECODE_ORDER,
+
+  /** control function to set the skip loop filter flag. Valid values are
+   * integers. The decoder will skip the loop filter when its value is set to
+   * nonzero. If the loop filter is skipped the decoder may accumulate decode
+   * artifacts. The default value is 0.
+   */
+  AV1_SET_SKIP_LOOP_FILTER,
+
+  /** control function to retrieve a pointer to the Accounting struct.  When
+   * compiled without --enable-accounting, this returns AOM_CODEC_INCAPABLE.
+   * If called before a frame has been decoded, this returns AOM_CODEC_ERROR.
+   * The caller should ensure that AOM_CODEC_OK is returned before attempting
+   * to dereference the Accounting pointer.
+   */
+  AV1_GET_ACCOUNTING,
+
+  /** control function to get last decoded frame quantizer. Returned value uses
+   * internal quantizer scale defined by the codec.
+   */
+  AOMD_GET_LAST_QUANTIZER,
+
+  /** control function to set the range of tile decoding. A value that is
+   * greater and equal to zero indicates only the specific row/column is
+   * decoded. A value that is -1 indicates the whole row/column is decoded.
+   * A special case is both values are -1 that means the whole frame is
+   * decoded.
+   */
+  AV1_SET_DECODE_TILE_ROW,
+  AV1_SET_DECODE_TILE_COL,
+
+  /** control function to set an aom_inspect_cb callback that is invoked each
+   * time a frame is decoded.  When compiled without --enable-inspection, this
+   * returns AOM_CODEC_INCAPABLE.
+   */
+  AV1_SET_INSPECTION_CALLBACK,
+
+  AOM_DECODER_CTRL_ID_MAX,
+};
+
+/** Decrypt n bytes of data from input -> output, using the decrypt_state
+ *  passed in AOMD_SET_DECRYPTOR.
+ */
+typedef void (*aom_decrypt_cb)(void *decrypt_state, const unsigned char *input,
+                               unsigned char *output, int count);
+
+/*!\brief Structure to hold decryption state
+ *
+ * Defines a structure to hold the decryption state and access function.
+ */
+typedef struct aom_decrypt_init {
+  /*! Decrypt callback. */
+  aom_decrypt_cb decrypt_cb;
+
+  /*! Decryption state. */
+  void *decrypt_state;
+} aom_decrypt_init;
+
+/*!\cond */
+/*!\brief AOM decoder control function parameter type
+ *
+ * Defines the data types that AOMD control functions take. Note that
+ * additional common controls are defined in aom.h
+ *
+ */
+
+AOM_CTRL_USE_TYPE(AOMD_GET_LAST_REF_UPDATES, int *)
+#define AOM_CTRL_AOMD_GET_LAST_REF_UPDATES
+AOM_CTRL_USE_TYPE(AOMD_GET_FRAME_CORRUPTED, int *)
+#define AOM_CTRL_AOMD_GET_FRAME_CORRUPTED
+AOM_CTRL_USE_TYPE(AOMD_GET_LAST_REF_USED, int *)
+#define AOM_CTRL_AOMD_GET_LAST_REF_USED
+AOM_CTRL_USE_TYPE(AOMD_GET_LAST_QUANTIZER, int *)
+#define AOM_CTRL_AOMD_GET_LAST_QUANTIZER
+AOM_CTRL_USE_TYPE(AOMD_SET_DECRYPTOR, aom_decrypt_init *)
+#define AOM_CTRL_AOMD_SET_DECRYPTOR
+// AOM_CTRL_USE_TYPE(AOMD_SET_DECRYPTOR, aom_decrypt_init *)
+//#define AOM_CTRL_AOMD_SET_DECRYPTOR
+AOM_CTRL_USE_TYPE(AV1D_GET_DISPLAY_SIZE, int *)
+#define AOM_CTRL_AV1D_GET_DISPLAY_SIZE
+AOM_CTRL_USE_TYPE(AV1D_GET_BIT_DEPTH, unsigned int *)
+#define AOM_CTRL_AV1D_GET_BIT_DEPTH
+AOM_CTRL_USE_TYPE(AV1D_GET_FRAME_SIZE, int *)
+#define AOM_CTRL_AV1D_GET_FRAME_SIZE
+AOM_CTRL_USE_TYPE(AV1_INVERT_TILE_DECODE_ORDER, int)
+#define AOM_CTRL_AV1_INVERT_TILE_DECODE_ORDER
+AOM_CTRL_USE_TYPE(AV1_GET_ACCOUNTING, Accounting **)
+#define AOM_CTRL_AV1_GET_ACCOUNTING
+AOM_CTRL_USE_TYPE(AV1_SET_DECODE_TILE_ROW, int)
+#define AOM_CTRL_AV1_SET_DECODE_TILE_ROW
+AOM_CTRL_USE_TYPE(AV1_SET_DECODE_TILE_COL, int)
+#define AOM_CTRL_AV1_SET_DECODE_TILE_COL
+AOM_CTRL_USE_TYPE(AV1_SET_INSPECTION_CALLBACK, aom_inspect_init *)
+#define AOM_CTRL_AV1_SET_INSPECTION_CALLBACK
+/*!\endcond */
+/*! @} - end defgroup aom_decoder */
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_AOMDX_H_
diff --git a/third_party/aom/aom/exports_com b/third_party/aom/aom/exports_com
new file mode 100644
index 0000000000..0c79fa124b
--- /dev/null
+++ b/third_party/aom/aom/exports_com
@@ -0,0 +1,16 @@
+text aom_codec_build_config
+text aom_codec_control_
+text aom_codec_destroy
+text aom_codec_err_to_string
+text aom_codec_error
+text aom_codec_error_detail
+text aom_codec_get_caps
+text aom_codec_iface_name
+text aom_codec_version
+text aom_codec_version_extra_str
+text aom_codec_version_str
+text aom_img_alloc
+text aom_img_flip
+text aom_img_free
+text aom_img_set_rect
+text aom_img_wrap
diff --git a/third_party/aom/aom/exports_dec b/third_party/aom/aom/exports_dec
new file mode 100644
index 0000000000..de8fe449a9
--- /dev/null
+++ b/third_party/aom/aom/exports_dec
@@ -0,0 +1,8 @@
+text aom_codec_dec_init_ver
+text aom_codec_decode
+text aom_codec_get_frame
+text aom_codec_get_stream_info
+text aom_codec_peek_stream_info
+text aom_codec_register_put_frame_cb
+text aom_codec_register_put_slice_cb
+text aom_codec_set_frame_buffer_functions
diff --git a/third_party/aom/aom/exports_enc b/third_party/aom/aom/exports_enc
new file mode 100644
index 0000000000..0dcca7da35
--- /dev/null
+++ b/third_party/aom/aom/exports_enc
@@ -0,0 +1,9 @@
+text aom_codec_enc_config_default
+text aom_codec_enc_config_set
+text aom_codec_enc_init_multi_ver
+text aom_codec_enc_init_ver
+text aom_codec_encode
+text aom_codec_get_cx_data
+text aom_codec_get_global_headers
+text aom_codec_get_preview_frame
+text aom_codec_set_cx_data_buf
diff --git a/third_party/aom/aom/internal/aom_codec_internal.h b/third_party/aom/aom/internal/aom_codec_internal.h
new file mode 100644
index 0000000000..7f9df7a722
--- /dev/null
+++ b/third_party/aom/aom/internal/aom_codec_internal.h
@@ -0,0 +1,467 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/*!\file
+ * \brief Describes the decoder algorithm interface for algorithm
+ *        implementations.
+ *
+ * This file defines the private structures and data types that are only
+ * relevant to implementing an algorithm, as opposed to using it.
+ *
+ * To create a decoder algorithm class, an interface structure is put
+ * into the global namespace:
+ *     <pre>
+ *     my_codec.c:
+ *       aom_codec_iface_t my_codec = {
+ *           "My Codec v1.0",
+ *           AOM_CODEC_ALG_ABI_VERSION,
+ *           ...
+ *       };
+ *     </pre>
+ *
+ * An application instantiates a specific decoder instance by using
+ * aom_codec_init() and a pointer to the algorithm's interface structure:
+ *     <pre>
+ *     my_app.c:
+ *       extern aom_codec_iface_t my_codec;
+ *       {
+ *           aom_codec_ctx_t algo;
+ *           res = aom_codec_init(&algo, &my_codec);
+ *       }
+ *     </pre>
+ *
+ * Once initialized, the instance is manged using other functions from
+ * the aom_codec_* family.
+ */
+#ifndef AOM_INTERNAL_AOM_CODEC_INTERNAL_H_
+#define AOM_INTERNAL_AOM_CODEC_INTERNAL_H_
+#include "./aom_config.h"
+#include "../aom_decoder.h"
+#include "../aom_encoder.h"
+#include <stdarg.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*!\brief Current ABI version number
+ *
+ * \internal
+ * If this file is altered in any way that changes the ABI, this value
+ * must be bumped.  Examples include, but are not limited to, changing
+ * types, removing or reassigning enums, adding/removing/rearranging
+ * fields to structures
+ */
+#define AOM_CODEC_INTERNAL_ABI_VERSION (5) /**<\hideinitializer*/
+
+typedef struct aom_codec_alg_priv aom_codec_alg_priv_t;
+typedef struct aom_codec_priv_enc_mr_cfg aom_codec_priv_enc_mr_cfg_t;
+
+/*!\brief init function pointer prototype
+ *
+ * Performs algorithm-specific initialization of the decoder context. This
+ * function is called by the generic aom_codec_init() wrapper function, so
+ * plugins implementing this interface may trust the input parameters to be
+ * properly initialized.
+ *
+ * \param[in] ctx   Pointer to this instance's context
+ * \retval #AOM_CODEC_OK
+ *     The input stream was recognized and decoder initialized.
+ * \retval #AOM_CODEC_MEM_ERROR
+ *     Memory operation failed.
+ */
+typedef aom_codec_err_t (*aom_codec_init_fn_t)(
+    aom_codec_ctx_t *ctx, aom_codec_priv_enc_mr_cfg_t *data);
+
+/*!\brief destroy function pointer prototype
+ *
+ * Performs algorithm-specific destruction of the decoder context. This
+ * function is called by the generic aom_codec_destroy() wrapper function,
+ * so plugins implementing this interface may trust the input parameters
+ * to be properly initialized.
+ *
+ * \param[in] ctx   Pointer to this instance's context
+ * \retval #AOM_CODEC_OK
+ *     The input stream was recognized and decoder initialized.
+ * \retval #AOM_CODEC_MEM_ERROR
+ *     Memory operation failed.
+ */
+typedef aom_codec_err_t (*aom_codec_destroy_fn_t)(aom_codec_alg_priv_t *ctx);
+
+/*!\brief parse stream info function pointer prototype
+ *
+ * Performs high level parsing of the bitstream. This function is called by the
+ * generic aom_codec_peek_stream_info() wrapper function, so plugins
+ * implementing this interface may trust the input parameters to be properly
+ * initialized.
+ *
+ * \param[in]      data    Pointer to a block of data to parse
+ * \param[in]      data_sz Size of the data buffer
+ * \param[in,out]  si      Pointer to stream info to update. The size member
+ *                         \ref MUST be properly initialized, but \ref MAY be
+ *                         clobbered by the algorithm. This parameter \ref MAY
+ *                         be NULL.
+ *
+ * \retval #AOM_CODEC_OK
+ *     Bitstream is parsable and stream information updated
+ */
+typedef aom_codec_err_t (*aom_codec_peek_si_fn_t)(const uint8_t *data,
+                                                  unsigned int data_sz,
+                                                  aom_codec_stream_info_t *si);
+
+/*!\brief Return information about the current stream.
+ *
+ * Returns information about the stream that has been parsed during decoding.
+ *
+ * \param[in]      ctx     Pointer to this instance's context
+ * \param[in,out]  si      Pointer to stream info to update. The size member
+ *                         \ref MUST be properly initialized, but \ref MAY be
+ *                         clobbered by the algorithm. This parameter \ref MAY
+ *                         be NULL.
+ *
+ * \retval #AOM_CODEC_OK
+ *     Bitstream is parsable and stream information updated
+ */
+typedef aom_codec_err_t (*aom_codec_get_si_fn_t)(aom_codec_alg_priv_t *ctx,
+                                                 aom_codec_stream_info_t *si);
+
+/*!\brief control function pointer prototype
+ *
+ * This function is used to exchange algorithm specific data with the decoder
+ * instance. This can be used to implement features specific to a particular
+ * algorithm.
+ *
+ * This function is called by the generic aom_codec_control() wrapper
+ * function, so plugins implementing this interface may trust the input
+ * parameters to be properly initialized. However,  this interface does not
+ * provide type safety for the exchanged data or assign meanings to the
+ * control codes. Those details should be specified in the algorithm's
+ * header file. In particular, the ctrl_id parameter is guaranteed to exist
+ * in the algorithm's control mapping table, and the data parameter may be NULL.
+ *
+ *
+ * \param[in]     ctx              Pointer to this instance's context
+ * \param[in]     ctrl_id          Algorithm specific control identifier
+ * \param[in,out] data             Data to exchange with algorithm instance.
+ *
+ * \retval #AOM_CODEC_OK
+ *     The internal state data was deserialized.
+ */
+typedef aom_codec_err_t (*aom_codec_control_fn_t)(aom_codec_alg_priv_t *ctx,
+                                                  va_list ap);
+
+/*!\brief control function pointer mapping
+ *
+ * This structure stores the mapping between control identifiers and
+ * implementing functions. Each algorithm provides a list of these
+ * mappings. This list is searched by the aom_codec_control() wrapper
+ * function to determine which function to invoke. The special
+ * value {0, NULL} is used to indicate end-of-list, and must be
+ * present. The special value {0, <non-null>} can be used as a catch-all
+ * mapping. This implies that ctrl_id values chosen by the algorithm
+ * \ref MUST be non-zero.
+ */
+typedef const struct aom_codec_ctrl_fn_map {
+  int ctrl_id;
+  aom_codec_control_fn_t fn;
+} aom_codec_ctrl_fn_map_t;
+
+/*!\brief decode data function pointer prototype
+ *
+ * Processes a buffer of coded data. If the processing results in a new
+ * decoded frame becoming available, #AOM_CODEC_CB_PUT_SLICE and
+ * #AOM_CODEC_CB_PUT_FRAME events are generated as appropriate. This
+ * function is called by the generic aom_codec_decode() wrapper function,
+ * so plugins implementing this interface may trust the input parameters
+ * to be properly initialized.
+ *
+ * \param[in] ctx          Pointer to this instance's context
+ * \param[in] data         Pointer to this block of new coded data. If
+ *                         NULL, a #AOM_CODEC_CB_PUT_FRAME event is posted
+ *                         for the previously decoded frame.
+ * \param[in] data_sz      Size of the coded data, in bytes.
+ *
+ * \return Returns #AOM_CODEC_OK if the coded data was processed completely
+ *         and future pictures can be decoded without error. Otherwise,
+ *         see the descriptions of the other error codes in ::aom_codec_err_t
+ *         for recoverability capabilities.
+ */
+typedef aom_codec_err_t (*aom_codec_decode_fn_t)(aom_codec_alg_priv_t *ctx,
+                                                 const uint8_t *data,
+                                                 unsigned int data_sz,
+                                                 void *user_priv,
+                                                 long deadline);
+
+/*!\brief Decoded frames iterator
+ *
+ * Iterates over a list of the frames available for display. The iterator
+ * storage should be initialized to NULL to start the iteration. Iteration is
+ * complete when this function returns NULL.
+ *
+ * The list of available frames becomes valid upon completion of the
+ * aom_codec_decode call, and remains valid until the next call to
+ * aom_codec_decode.
+ *
+ * \param[in]     ctx      Pointer to this instance's context
+ * \param[in out] iter     Iterator storage, initialized to NULL
+ *
+ * \return Returns a pointer to an image, if one is ready for display. Frames
+ *         produced will always be in PTS (presentation time stamp) order.
+ */
+typedef aom_image_t *(*aom_codec_get_frame_fn_t)(aom_codec_alg_priv_t *ctx,
+                                                 aom_codec_iter_t *iter);
+
+/*!\brief Pass in external frame buffers for the decoder to use.
+ *
+ * Registers functions to be called when libaom needs a frame buffer
+ * to decode the current frame and a function to be called when libaom does
+ * not internally reference the frame buffer. This set function must
+ * be called before the first call to decode or libaom will assume the
+ * default behavior of allocating frame buffers internally.
+ *
+ * \param[in] ctx          Pointer to this instance's context
+ * \param[in] cb_get       Pointer to the get callback function
+ * \param[in] cb_release   Pointer to the release callback function
+ * \param[in] cb_priv      Callback's private data
+ *
+ * \retval #AOM_CODEC_OK
+ *     External frame buffers will be used by libaom.
+ * \retval #AOM_CODEC_INVALID_PARAM
+ *     One or more of the callbacks were NULL.
+ * \retval #AOM_CODEC_ERROR
+ *     Decoder context not initialized, or algorithm not capable of
+ *     using external frame buffers.
+ *
+ * \note
+ * When decoding AV1, the application may be required to pass in at least
+ * #AOM_MAXIMUM_WORK_BUFFERS external frame
+ * buffers.
+ */
+typedef aom_codec_err_t (*aom_codec_set_fb_fn_t)(
+    aom_codec_alg_priv_t *ctx, aom_get_frame_buffer_cb_fn_t cb_get,
+    aom_release_frame_buffer_cb_fn_t cb_release, void *cb_priv);
+
+typedef aom_codec_err_t (*aom_codec_encode_fn_t)(aom_codec_alg_priv_t *ctx,
+                                                 const aom_image_t *img,
+                                                 aom_codec_pts_t pts,
+                                                 unsigned long duration,
+                                                 aom_enc_frame_flags_t flags,
+                                                 unsigned long deadline);
+typedef const aom_codec_cx_pkt_t *(*aom_codec_get_cx_data_fn_t)(
+    aom_codec_alg_priv_t *ctx, aom_codec_iter_t *iter);
+
+typedef aom_codec_err_t (*aom_codec_enc_config_set_fn_t)(
+    aom_codec_alg_priv_t *ctx, const aom_codec_enc_cfg_t *cfg);
+typedef aom_fixed_buf_t *(*aom_codec_get_global_headers_fn_t)(
+    aom_codec_alg_priv_t *ctx);
+
+typedef aom_image_t *(*aom_codec_get_preview_frame_fn_t)(
+    aom_codec_alg_priv_t *ctx);
+
+typedef aom_codec_err_t (*aom_codec_enc_mr_get_mem_loc_fn_t)(
+    const aom_codec_enc_cfg_t *cfg, void **mem_loc);
+
+/*!\brief usage configuration mapping
+ *
+ * This structure stores the mapping between usage identifiers and
+ * configuration structures. Each algorithm provides a list of these
+ * mappings. This list is searched by the aom_codec_enc_config_default()
+ * wrapper function to determine which config to return. The special value
+ * {-1, {0}} is used to indicate end-of-list, and must be present. At least
+ * one mapping must be present, in addition to the end-of-list.
+ *
+ */
+typedef const struct aom_codec_enc_cfg_map {
+  int usage;
+  aom_codec_enc_cfg_t cfg;
+} aom_codec_enc_cfg_map_t;
+
+/*!\brief Decoder algorithm interface interface
+ *
+ * All decoders \ref MUST expose a variable of this type.
+ */
+struct aom_codec_iface {
+  const char *name;                   /**< Identification String  */
+  int abi_version;                    /**< Implemented ABI version */
+  aom_codec_caps_t caps;              /**< Decoder capabilities */
+  aom_codec_init_fn_t init;           /**< \copydoc ::aom_codec_init_fn_t */
+  aom_codec_destroy_fn_t destroy;     /**< \copydoc ::aom_codec_destroy_fn_t */
+  aom_codec_ctrl_fn_map_t *ctrl_maps; /**< \copydoc ::aom_codec_ctrl_fn_map_t */
+  struct aom_codec_dec_iface {
+    aom_codec_peek_si_fn_t peek_si; /**< \copydoc ::aom_codec_peek_si_fn_t */
+    aom_codec_get_si_fn_t get_si;   /**< \copydoc ::aom_codec_get_si_fn_t */
+    aom_codec_decode_fn_t decode;   /**< \copydoc ::aom_codec_decode_fn_t */
+    aom_codec_get_frame_fn_t
+        get_frame;                   /**< \copydoc ::aom_codec_get_frame_fn_t */
+    aom_codec_set_fb_fn_t set_fb_fn; /**< \copydoc ::aom_codec_set_fb_fn_t */
+  } dec;
+  struct aom_codec_enc_iface {
+    int cfg_map_count;
+    aom_codec_enc_cfg_map_t
+        *cfg_maps;                /**< \copydoc ::aom_codec_enc_cfg_map_t */
+    aom_codec_encode_fn_t encode; /**< \copydoc ::aom_codec_encode_fn_t */
+    aom_codec_get_cx_data_fn_t
+        get_cx_data; /**< \copydoc ::aom_codec_get_cx_data_fn_t */
+    aom_codec_enc_config_set_fn_t
+        cfg_set; /**< \copydoc ::aom_codec_enc_config_set_fn_t */
+    aom_codec_get_global_headers_fn_t
+        get_glob_hdrs; /**< \copydoc ::aom_codec_get_global_headers_fn_t */
+    aom_codec_get_preview_frame_fn_t
+        get_preview; /**< \copydoc ::aom_codec_get_preview_frame_fn_t */
+    aom_codec_enc_mr_get_mem_loc_fn_t
+        mr_get_mem_loc; /**< \copydoc ::aom_codec_enc_mr_get_mem_loc_fn_t */
+  } enc;
+};
+
+/*!\brief Callback function pointer / user data pair storage */
+typedef struct aom_codec_priv_cb_pair {
+  union {
+    aom_codec_put_frame_cb_fn_t put_frame;
+    aom_codec_put_slice_cb_fn_t put_slice;
+  } u;
+  void *user_priv;
+} aom_codec_priv_cb_pair_t;
+
+/*!\brief Instance private storage
+ *
+ * This structure is allocated by the algorithm's init function. It can be
+ * extended in one of two ways. First, a second, algorithm specific structure
+ * can be allocated and the priv member pointed to it. Alternatively, this
+ * structure can be made the first member of the algorithm specific structure,
+ * and the pointer cast to the proper type.
+ */
+struct aom_codec_priv {
+  const char *err_detail;
+  aom_codec_flags_t init_flags;
+  struct {
+    aom_codec_priv_cb_pair_t put_frame_cb;
+    aom_codec_priv_cb_pair_t put_slice_cb;
+  } dec;
+  struct {
+    aom_fixed_buf_t cx_data_dst_buf;
+    unsigned int cx_data_pad_before;
+    unsigned int cx_data_pad_after;
+    aom_codec_cx_pkt_t cx_data_pkt;
+    unsigned int total_encoders;
+  } enc;
+};
+
+/*
+ * Multi-resolution encoding internal configuration
+ */
+struct aom_codec_priv_enc_mr_cfg {
+  unsigned int mr_total_resolutions;
+  unsigned int mr_encoder_id;
+  struct aom_rational mr_down_sampling_factor;
+  void *mr_low_res_mode_info;
+};
+
+#undef AOM_CTRL_USE_TYPE
+#define AOM_CTRL_USE_TYPE(id, typ) \
+  static AOM_INLINE typ id##__value(va_list args) { return va_arg(args, typ); }
+
+#undef AOM_CTRL_USE_TYPE_DEPRECATED
+#define AOM_CTRL_USE_TYPE_DEPRECATED(id, typ) \
+  static AOM_INLINE typ id##__value(va_list args) { return va_arg(args, typ); }
+
+#define CAST(id, arg) id##__value(arg)
+
+/* CODEC_INTERFACE convenience macro
+ *
+ * By convention, each codec interface is a struct with extern linkage, where
+ * the symbol is suffixed with _algo. A getter function is also defined to
+ * return a pointer to the struct, since in some cases it's easier to work
+ * with text symbols than data symbols (see issue #169). This function has
+ * the same name as the struct, less the _algo suffix. The CODEC_INTERFACE
+ * macro is provided to define this getter function automatically.
+ */
+#define CODEC_INTERFACE(id)                          \
+  aom_codec_iface_t *id(void) { return &id##_algo; } \
+  aom_codec_iface_t id##_algo
+
+/* Internal Utility Functions
+ *
+ * The following functions are intended to be used inside algorithms as
+ * utilities for manipulating aom_codec_* data structures.
+ */
+struct aom_codec_pkt_list {
+  unsigned int cnt;
+  unsigned int max;
+  struct aom_codec_cx_pkt pkts[1];
+};
+
+#define aom_codec_pkt_list_decl(n)     \
+  union {                              \
+    struct aom_codec_pkt_list head;    \
+    struct {                           \
+      struct aom_codec_pkt_list head;  \
+      struct aom_codec_cx_pkt pkts[n]; \
+    } alloc;                           \
+  }
+
+#define aom_codec_pkt_list_init(m) \
+  (m)->alloc.head.cnt = 0,         \
+  (m)->alloc.head.max = sizeof((m)->alloc.pkts) / sizeof((m)->alloc.pkts[0])
+
+int aom_codec_pkt_list_add(struct aom_codec_pkt_list *,
+                           const struct aom_codec_cx_pkt *);
+
+const aom_codec_cx_pkt_t *aom_codec_pkt_list_get(
+    struct aom_codec_pkt_list *list, aom_codec_iter_t *iter);
+
+#include <stdio.h>
+#include <setjmp.h>
+
+struct aom_internal_error_info {
+  aom_codec_err_t error_code;
+  int has_detail;
+  char detail[80];
+  int setjmp;
+  jmp_buf jmp;
+};
+
+#define CLANG_ANALYZER_NORETURN
+#if defined(__has_feature)
+#if __has_feature(attribute_analyzer_noreturn)
+#undef CLANG_ANALYZER_NORETURN
+#define CLANG_ANALYZER_NORETURN __attribute__((analyzer_noreturn))
+#endif
+#endif
+
+void aom_internal_error(struct aom_internal_error_info *info,
+                        aom_codec_err_t error, const char *fmt,
+                        ...) CLANG_ANALYZER_NORETURN;
+
+void aom_merge_corrupted_flag(int *corrupted, int value);
+
+#if CONFIG_DEBUG
+#define AOM_CHECK_MEM_ERROR(error_info, lval, expr)                         \
+  do {                                                                      \
+    lval = (expr);                                                          \
+    if (!lval)                                                              \
+      aom_internal_error(error_info, AOM_CODEC_MEM_ERROR,                   \
+                         "Failed to allocate " #lval " at %s:%d", __FILE__, \
+                         __LINE__);                                         \
+  } while (0)
+#else
+#define AOM_CHECK_MEM_ERROR(error_info, lval, expr)       \
+  do {                                                    \
+    lval = (expr);                                        \
+    if (!lval)                                            \
+      aom_internal_error(error_info, AOM_CODEC_MEM_ERROR, \
+                         "Failed to allocate " #lval);    \
+  } while (0)
+#endif
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_INTERNAL_AOM_CODEC_INTERNAL_H_
diff --git a/third_party/aom/aom/src/aom_codec.c b/third_party/aom/aom/src/aom_codec.c
new file mode 100644
index 0000000000..873d758768
--- /dev/null
+++ b/third_party/aom/aom/src/aom_codec.c
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/*!\file
+ * \brief Provides the high level interface to wrap decoder algorithms.
+ *
+ */
+#include <stdarg.h>
+#include <stdlib.h>
+#include "aom/aom_integer.h"
+#include "aom/internal/aom_codec_internal.h"
+#include "aom_version.h"
+
+#define SAVE_STATUS(ctx, var) (ctx ? (ctx->err = var) : var)
+
+int aom_codec_version(void) { return VERSION_PACKED; }
+
+const char *aom_codec_version_str(void) { return VERSION_STRING_NOSP; }
+
+const char *aom_codec_version_extra_str(void) { return VERSION_EXTRA; }
+
+const char *aom_codec_iface_name(aom_codec_iface_t *iface) {
+  return iface ? iface->name : "<invalid interface>";
+}
+
+const char *aom_codec_err_to_string(aom_codec_err_t err) {
+  switch (err) {
+    case AOM_CODEC_OK: return "Success";
+    case AOM_CODEC_ERROR: return "Unspecified internal error";
+    case AOM_CODEC_MEM_ERROR: return "Memory allocation error";
+    case AOM_CODEC_ABI_MISMATCH: return "ABI version mismatch";
+    case AOM_CODEC_INCAPABLE:
+      return "Codec does not implement requested capability";
+    case AOM_CODEC_UNSUP_BITSTREAM:
+      return "Bitstream not supported by this decoder";
+    case AOM_CODEC_UNSUP_FEATURE:
+      return "Bitstream required feature not supported by this decoder";
+    case AOM_CODEC_CORRUPT_FRAME: return "Corrupt frame detected";
+    case AOM_CODEC_INVALID_PARAM: return "Invalid parameter";
+    case AOM_CODEC_LIST_END: return "End of iterated list";
+  }
+
+  return "Unrecognized error code";
+}
+
+const char *aom_codec_error(aom_codec_ctx_t *ctx) {
+  return (ctx) ? aom_codec_err_to_string(ctx->err)
+               : aom_codec_err_to_string(AOM_CODEC_INVALID_PARAM);
+}
+
+const char *aom_codec_error_detail(aom_codec_ctx_t *ctx) {
+  if (ctx && ctx->err)
+    return ctx->priv ? ctx->priv->err_detail : ctx->err_detail;
+
+  return NULL;
+}
+
+aom_codec_err_t aom_codec_destroy(aom_codec_ctx_t *ctx) {
+  aom_codec_err_t res;
+
+  if (!ctx)
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv)
+    res = AOM_CODEC_ERROR;
+  else {
+    ctx->iface->destroy((aom_codec_alg_priv_t *)ctx->priv);
+
+    ctx->iface = NULL;
+    ctx->name = NULL;
+    ctx->priv = NULL;
+    res = AOM_CODEC_OK;
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+aom_codec_caps_t aom_codec_get_caps(aom_codec_iface_t *iface) {
+  return (iface) ? iface->caps : 0;
+}
+
+aom_codec_err_t aom_codec_control_(aom_codec_ctx_t *ctx, int ctrl_id, ...) {
+  aom_codec_err_t res;
+
+  if (!ctx || !ctrl_id)
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv || !ctx->iface->ctrl_maps)
+    res = AOM_CODEC_ERROR;
+  else {
+    aom_codec_ctrl_fn_map_t *entry;
+
+    res = AOM_CODEC_ERROR;
+
+    for (entry = ctx->iface->ctrl_maps; entry && entry->fn; entry++) {
+      if (!entry->ctrl_id || entry->ctrl_id == ctrl_id) {
+        va_list ap;
+
+        va_start(ap, ctrl_id);
+        res = entry->fn((aom_codec_alg_priv_t *)ctx->priv, ap);
+        va_end(ap);
+        break;
+      }
+    }
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+void aom_internal_error(struct aom_internal_error_info *info,
+                        aom_codec_err_t error, const char *fmt, ...) {
+  va_list ap;
+
+  info->error_code = error;
+  info->has_detail = 0;
+
+  if (fmt) {
+    size_t sz = sizeof(info->detail);
+
+    info->has_detail = 1;
+    va_start(ap, fmt);
+    vsnprintf(info->detail, sz - 1, fmt, ap);
+    va_end(ap);
+    info->detail[sz - 1] = '\0';
+  }
+
+  if (info->setjmp) longjmp(info->jmp, info->error_code);
+}
+
+void aom_merge_corrupted_flag(int *corrupted, int value) {
+  *corrupted |= value;
+}
diff --git a/third_party/aom/aom/src/aom_decoder.c b/third_party/aom/aom/src/aom_decoder.c
new file mode 100644
index 0000000000..c6cec6d79c
--- /dev/null
+++ b/third_party/aom/aom/src/aom_decoder.c
@@ -0,0 +1,189 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/*!\file
+ * \brief Provides the high level interface to wrap decoder algorithms.
+ *
+ */
+#include <string.h>
+#include "aom/internal/aom_codec_internal.h"
+
+#define SAVE_STATUS(ctx, var) (ctx ? (ctx->err = var) : var)
+
+static aom_codec_alg_priv_t *get_alg_priv(aom_codec_ctx_t *ctx) {
+  return (aom_codec_alg_priv_t *)ctx->priv;
+}
+
+aom_codec_err_t aom_codec_dec_init_ver(aom_codec_ctx_t *ctx,
+                                       aom_codec_iface_t *iface,
+                                       const aom_codec_dec_cfg_t *cfg,
+                                       aom_codec_flags_t flags, int ver) {
+  aom_codec_err_t res;
+
+  if (ver != AOM_DECODER_ABI_VERSION)
+    res = AOM_CODEC_ABI_MISMATCH;
+  else if (!ctx || !iface)
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (iface->abi_version != AOM_CODEC_INTERNAL_ABI_VERSION)
+    res = AOM_CODEC_ABI_MISMATCH;
+  else if ((flags & AOM_CODEC_USE_POSTPROC) &&
+           !(iface->caps & AOM_CODEC_CAP_POSTPROC))
+    res = AOM_CODEC_INCAPABLE;
+  else if ((flags & AOM_CODEC_USE_ERROR_CONCEALMENT) &&
+           !(iface->caps & AOM_CODEC_CAP_ERROR_CONCEALMENT))
+    res = AOM_CODEC_INCAPABLE;
+  else if ((flags & AOM_CODEC_USE_INPUT_FRAGMENTS) &&
+           !(iface->caps & AOM_CODEC_CAP_INPUT_FRAGMENTS))
+    res = AOM_CODEC_INCAPABLE;
+  else if (!(iface->caps & AOM_CODEC_CAP_DECODER))
+    res = AOM_CODEC_INCAPABLE;
+  else {
+    memset(ctx, 0, sizeof(*ctx));
+    ctx->iface = iface;
+    ctx->name = iface->name;
+    ctx->priv = NULL;
+    ctx->init_flags = flags;
+    ctx->config.dec = cfg;
+
+    res = ctx->iface->init(ctx, NULL);
+    if (res) {
+      ctx->err_detail = ctx->priv ? ctx->priv->err_detail : NULL;
+      aom_codec_destroy(ctx);
+    }
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+aom_codec_err_t aom_codec_peek_stream_info(aom_codec_iface_t *iface,
+                                           const uint8_t *data,
+                                           unsigned int data_sz,
+                                           aom_codec_stream_info_t *si) {
+  aom_codec_err_t res;
+
+  if (!iface || !data || !data_sz || !si ||
+      si->sz < sizeof(aom_codec_stream_info_t))
+    res = AOM_CODEC_INVALID_PARAM;
+  else {
+    /* Set default/unknown values */
+    si->w = 0;
+    si->h = 0;
+
+    res = iface->dec.peek_si(data, data_sz, si);
+  }
+
+  return res;
+}
+
+aom_codec_err_t aom_codec_get_stream_info(aom_codec_ctx_t *ctx,
+                                          aom_codec_stream_info_t *si) {
+  aom_codec_err_t res;
+
+  if (!ctx || !si || si->sz < sizeof(aom_codec_stream_info_t))
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv)
+    res = AOM_CODEC_ERROR;
+  else {
+    /* Set default/unknown values */
+    si->w = 0;
+    si->h = 0;
+
+    res = ctx->iface->dec.get_si(get_alg_priv(ctx), si);
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+aom_codec_err_t aom_codec_decode(aom_codec_ctx_t *ctx, const uint8_t *data,
+                                 unsigned int data_sz, void *user_priv,
+                                 long deadline) {
+  aom_codec_err_t res;
+
+  /* Sanity checks */
+  /* NULL data ptr allowed if data_sz is 0 too */
+  if (!ctx || (!data && data_sz) || (data && !data_sz))
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv)
+    res = AOM_CODEC_ERROR;
+  else {
+    res = ctx->iface->dec.decode(get_alg_priv(ctx), data, data_sz, user_priv,
+                                 deadline);
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+aom_image_t *aom_codec_get_frame(aom_codec_ctx_t *ctx, aom_codec_iter_t *iter) {
+  aom_image_t *img;
+
+  if (!ctx || !iter || !ctx->iface || !ctx->priv)
+    img = NULL;
+  else
+    img = ctx->iface->dec.get_frame(get_alg_priv(ctx), iter);
+
+  return img;
+}
+
+aom_codec_err_t aom_codec_register_put_frame_cb(aom_codec_ctx_t *ctx,
+                                                aom_codec_put_frame_cb_fn_t cb,
+                                                void *user_priv) {
+  aom_codec_err_t res;
+
+  if (!ctx || !cb)
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv ||
+           !(ctx->iface->caps & AOM_CODEC_CAP_PUT_FRAME))
+    res = AOM_CODEC_ERROR;
+  else {
+    ctx->priv->dec.put_frame_cb.u.put_frame = cb;
+    ctx->priv->dec.put_frame_cb.user_priv = user_priv;
+    res = AOM_CODEC_OK;
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+aom_codec_err_t aom_codec_register_put_slice_cb(aom_codec_ctx_t *ctx,
+                                                aom_codec_put_slice_cb_fn_t cb,
+                                                void *user_priv) {
+  aom_codec_err_t res;
+
+  if (!ctx || !cb)
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv ||
+           !(ctx->iface->caps & AOM_CODEC_CAP_PUT_SLICE))
+    res = AOM_CODEC_ERROR;
+  else {
+    ctx->priv->dec.put_slice_cb.u.put_slice = cb;
+    ctx->priv->dec.put_slice_cb.user_priv = user_priv;
+    res = AOM_CODEC_OK;
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+aom_codec_err_t aom_codec_set_frame_buffer_functions(
+    aom_codec_ctx_t *ctx, aom_get_frame_buffer_cb_fn_t cb_get,
+    aom_release_frame_buffer_cb_fn_t cb_release, void *cb_priv) {
+  aom_codec_err_t res;
+
+  if (!ctx || !cb_get || !cb_release) {
+    res = AOM_CODEC_INVALID_PARAM;
+  } else if (!ctx->iface || !ctx->priv ||
+             !(ctx->iface->caps & AOM_CODEC_CAP_EXTERNAL_FRAME_BUFFER)) {
+    res = AOM_CODEC_ERROR;
+  } else {
+    res = ctx->iface->dec.set_fb_fn(get_alg_priv(ctx), cb_get, cb_release,
+                                    cb_priv);
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
diff --git a/third_party/aom/aom/src/aom_encoder.c b/third_party/aom/aom/src/aom_encoder.c
new file mode 100644
index 0000000000..ac84c888a4
--- /dev/null
+++ b/third_party/aom/aom/src/aom_encoder.c
@@ -0,0 +1,405 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/*!\file
+ * \brief Provides the high level interface to wrap encoder algorithms.
+ *
+ */
+#include "./aom_config.h"
+
+#if HAVE_FEXCEPT
+#define _GNU_SOURCE
+#include <fenv.h>
+#endif
+
+#include <limits.h>
+#include <string.h>
+#include "aom/internal/aom_codec_internal.h"
+
+#define SAVE_STATUS(ctx, var) (ctx ? (ctx->err = var) : var)
+
+static aom_codec_alg_priv_t *get_alg_priv(aom_codec_ctx_t *ctx) {
+  return (aom_codec_alg_priv_t *)ctx->priv;
+}
+
+aom_codec_err_t aom_codec_enc_init_ver(aom_codec_ctx_t *ctx,
+                                       aom_codec_iface_t *iface,
+                                       const aom_codec_enc_cfg_t *cfg,
+                                       aom_codec_flags_t flags, int ver) {
+  aom_codec_err_t res;
+
+  if (ver != AOM_ENCODER_ABI_VERSION)
+    res = AOM_CODEC_ABI_MISMATCH;
+  else if (!ctx || !iface || !cfg)
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (iface->abi_version != AOM_CODEC_INTERNAL_ABI_VERSION)
+    res = AOM_CODEC_ABI_MISMATCH;
+  else if (!(iface->caps & AOM_CODEC_CAP_ENCODER))
+    res = AOM_CODEC_INCAPABLE;
+  else if ((flags & AOM_CODEC_USE_PSNR) && !(iface->caps & AOM_CODEC_CAP_PSNR))
+    res = AOM_CODEC_INCAPABLE;
+  else if ((flags & AOM_CODEC_USE_OUTPUT_PARTITION) &&
+           !(iface->caps & AOM_CODEC_CAP_OUTPUT_PARTITION))
+    res = AOM_CODEC_INCAPABLE;
+  else {
+    ctx->iface = iface;
+    ctx->name = iface->name;
+    ctx->priv = NULL;
+    ctx->init_flags = flags;
+    ctx->config.enc = cfg;
+    res = ctx->iface->init(ctx, NULL);
+
+    if (res) {
+      ctx->err_detail = ctx->priv ? ctx->priv->err_detail : NULL;
+      aom_codec_destroy(ctx);
+    }
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+aom_codec_err_t aom_codec_enc_init_multi_ver(
+    aom_codec_ctx_t *ctx, aom_codec_iface_t *iface, aom_codec_enc_cfg_t *cfg,
+    int num_enc, aom_codec_flags_t flags, aom_rational_t *dsf, int ver) {
+  aom_codec_err_t res = AOM_CODEC_OK;
+
+  if (ver != AOM_ENCODER_ABI_VERSION)
+    res = AOM_CODEC_ABI_MISMATCH;
+  else if (!ctx || !iface || !cfg || (num_enc > 16 || num_enc < 1))
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (iface->abi_version != AOM_CODEC_INTERNAL_ABI_VERSION)
+    res = AOM_CODEC_ABI_MISMATCH;
+  else if (!(iface->caps & AOM_CODEC_CAP_ENCODER))
+    res = AOM_CODEC_INCAPABLE;
+  else if ((flags & AOM_CODEC_USE_PSNR) && !(iface->caps & AOM_CODEC_CAP_PSNR))
+    res = AOM_CODEC_INCAPABLE;
+  else if ((flags & AOM_CODEC_USE_OUTPUT_PARTITION) &&
+           !(iface->caps & AOM_CODEC_CAP_OUTPUT_PARTITION))
+    res = AOM_CODEC_INCAPABLE;
+  else {
+    int i;
+    void *mem_loc = NULL;
+
+    if (!(res = iface->enc.mr_get_mem_loc(cfg, &mem_loc))) {
+      for (i = 0; i < num_enc; i++) {
+        aom_codec_priv_enc_mr_cfg_t mr_cfg;
+
+        /* Validate down-sampling factor. */
+        if (dsf->num < 1 || dsf->num > 4096 || dsf->den < 1 ||
+            dsf->den > dsf->num) {
+          res = AOM_CODEC_INVALID_PARAM;
+          break;
+        }
+
+        mr_cfg.mr_low_res_mode_info = mem_loc;
+        mr_cfg.mr_total_resolutions = num_enc;
+        mr_cfg.mr_encoder_id = num_enc - 1 - i;
+        mr_cfg.mr_down_sampling_factor.num = dsf->num;
+        mr_cfg.mr_down_sampling_factor.den = dsf->den;
+
+        /* Force Key-frame synchronization. Namely, encoder at higher
+         * resolution always use the same frame_type chosen by the
+         * lowest-resolution encoder.
+         */
+        if (mr_cfg.mr_encoder_id) cfg->kf_mode = AOM_KF_DISABLED;
+
+        ctx->iface = iface;
+        ctx->name = iface->name;
+        ctx->priv = NULL;
+        ctx->init_flags = flags;
+        ctx->config.enc = cfg;
+        res = ctx->iface->init(ctx, &mr_cfg);
+
+        if (res) {
+          const char *error_detail = ctx->priv ? ctx->priv->err_detail : NULL;
+          /* Destroy current ctx */
+          ctx->err_detail = error_detail;
+          aom_codec_destroy(ctx);
+
+          /* Destroy already allocated high-level ctx */
+          while (i) {
+            ctx--;
+            ctx->err_detail = error_detail;
+            aom_codec_destroy(ctx);
+            i--;
+          }
+        }
+
+        if (res) break;
+
+        ctx++;
+        cfg++;
+        dsf++;
+      }
+      ctx--;
+    }
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+aom_codec_err_t aom_codec_enc_config_default(aom_codec_iface_t *iface,
+                                             aom_codec_enc_cfg_t *cfg,
+                                             unsigned int usage) {
+  aom_codec_err_t res;
+  aom_codec_enc_cfg_map_t *map;
+  int i;
+
+  if (!iface || !cfg || usage > INT_MAX)
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (!(iface->caps & AOM_CODEC_CAP_ENCODER))
+    res = AOM_CODEC_INCAPABLE;
+  else {
+    res = AOM_CODEC_INVALID_PARAM;
+
+    for (i = 0; i < iface->enc.cfg_map_count; ++i) {
+      map = iface->enc.cfg_maps + i;
+      if (map->usage == (int)usage) {
+        *cfg = map->cfg;
+        cfg->g_usage = usage;
+        res = AOM_CODEC_OK;
+        break;
+      }
+    }
+  }
+
+  return res;
+}
+
+/* clang-format off */
+#define FLOATING_POINT_BEGIN_SCOPE do {
+#define FLOATING_POINT_END_SCOPE } while (0);
+/* clang-format on */
+
+#if ARCH_X86 || ARCH_X86_64
+/* On X86, disable the x87 unit's internal 80 bit precision for better
+ * consistency with the SSE unit's 64 bit precision.
+ */
+#include "aom_ports/x86.h"
+#define FLOATING_POINT_SET_PRECISION \
+  unsigned short x87_orig_mode = x87_set_double_precision();
+#define FLOATING_POINT_RESTORE_PRECISION x87_set_control_word(x87_orig_mode);
+#else
+#define FLOATING_POINT_SET_PRECISION
+#define FLOATING_POINT_RESTORE_PRECISION
+#endif  // ARCH_X86 || ARCH_X86_64
+
+#if HAVE_FEXCEPT && CONFIG_DEBUG
+#define FLOATING_POINT_SET_EXCEPTIONS \
+  const int float_excepts = feenableexcept(FE_DIVBYZERO);
+#define FLOATING_POINT_RESTORE_EXCEPTIONS feenableexcept(float_excepts);
+#else
+#define FLOATING_POINT_SET_EXCEPTIONS
+#define FLOATING_POINT_RESTORE_EXCEPTIONS
+#endif  // HAVE_FEXCEPT && CONFIG_DEBUG
+
+#define FLOATING_POINT_INIT    \
+  FLOATING_POINT_BEGIN_SCOPE   \
+  FLOATING_POINT_SET_PRECISION \
+  FLOATING_POINT_SET_EXCEPTIONS
+
+#define FLOATING_POINT_RESTORE      \
+  FLOATING_POINT_RESTORE_EXCEPTIONS \
+  FLOATING_POINT_RESTORE_PRECISION  \
+  FLOATING_POINT_END_SCOPE
+
+aom_codec_err_t aom_codec_encode(aom_codec_ctx_t *ctx, const aom_image_t *img,
+                                 aom_codec_pts_t pts, unsigned long duration,
+                                 aom_enc_frame_flags_t flags,
+                                 unsigned long deadline) {
+  aom_codec_err_t res = AOM_CODEC_OK;
+
+  if (!ctx || (img && !duration))
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (!ctx->iface || !ctx->priv)
+    res = AOM_CODEC_ERROR;
+  else if (!(ctx->iface->caps & AOM_CODEC_CAP_ENCODER))
+    res = AOM_CODEC_INCAPABLE;
+  else {
+    unsigned int num_enc = ctx->priv->enc.total_encoders;
+
+    /* Execute in a normalized floating point environment, if the platform
+     * requires it.
+     */
+    FLOATING_POINT_INIT
+
+    if (num_enc == 1)
+      res = ctx->iface->enc.encode(get_alg_priv(ctx), img, pts, duration, flags,
+                                   deadline);
+    else {
+      /* Multi-resolution encoding:
+       * Encode multi-levels in reverse order. For example,
+       * if mr_total_resolutions = 3, first encode level 2,
+       * then encode level 1, and finally encode level 0.
+       */
+      int i;
+
+      ctx += num_enc - 1;
+      if (img) img += num_enc - 1;
+
+      for (i = num_enc - 1; i >= 0; i--) {
+        if ((res = ctx->iface->enc.encode(get_alg_priv(ctx), img, pts, duration,
+                                          flags, deadline)))
+          break;
+
+        ctx--;
+        if (img) img--;
+      }
+      ctx++;
+    }
+
+    FLOATING_POINT_RESTORE
+  }
+
+  return SAVE_STATUS(ctx, res);
+}
+
+const aom_codec_cx_pkt_t *aom_codec_get_cx_data(aom_codec_ctx_t *ctx,
+                                                aom_codec_iter_t *iter) {
+  const aom_codec_cx_pkt_t *pkt = NULL;
+
+  if (ctx) {
+    if (!iter)
+      ctx->err = AOM_CODEC_INVALID_PARAM;
+    else if (!ctx->iface || !ctx->priv)
+      ctx->err = AOM_CODEC_ERROR;
+    else if (!(ctx->iface->caps & AOM_CODEC_CAP_ENCODER))
+      ctx->err = AOM_CODEC_INCAPABLE;
+    else
+      pkt = ctx->iface->enc.get_cx_data(get_alg_priv(ctx), iter);
+  }
+
+  if (pkt && pkt->kind == AOM_CODEC_CX_FRAME_PKT) {
+    // If the application has specified a destination area for the
+    // compressed data, and the codec has not placed the data there,
+    // and it fits, copy it.
+    aom_codec_priv_t *const priv = ctx->priv;
+    char *const dst_buf = (char *)priv->enc.cx_data_dst_buf.buf;
+
+    if (dst_buf && pkt->data.raw.buf != dst_buf &&
+        pkt->data.raw.sz + priv->enc.cx_data_pad_before +
+                priv->enc.cx_data_pad_after <=
+            priv->enc.cx_data_dst_buf.sz) {
+      aom_codec_cx_pkt_t *modified_pkt = &priv->enc.cx_data_pkt;
+
+      memcpy(dst_buf + priv->enc.cx_data_pad_before, pkt->data.raw.buf,
+             pkt->data.raw.sz);
+      *modified_pkt = *pkt;
+      modified_pkt->data.raw.buf = dst_buf;
+      modified_pkt->data.raw.sz +=
+          priv->enc.cx_data_pad_before + priv->enc.cx_data_pad_after;
+      pkt = modified_pkt;
+    }
+
+    if (dst_buf == pkt->data.raw.buf) {
+      priv->enc.cx_data_dst_buf.buf = dst_buf + pkt->data.raw.sz;
+      priv->enc.cx_data_dst_buf.sz -= pkt->data.raw.sz;
+    }
+  }
+
+  return pkt;
+}
+
+aom_codec_err_t aom_codec_set_cx_data_buf(aom_codec_ctx_t *ctx,
+                                          const aom_fixed_buf_t *buf,
+                                          unsigned int pad_before,
+                                          unsigned int pad_after) {
+  if (!ctx || !ctx->priv) return AOM_CODEC_INVALID_PARAM;
+
+  if (buf) {
+    ctx->priv->enc.cx_data_dst_buf = *buf;
+    ctx->priv->enc.cx_data_pad_before = pad_before;
+    ctx->priv->enc.cx_data_pad_after = pad_after;
+  } else {
+    ctx->priv->enc.cx_data_dst_buf.buf = NULL;
+    ctx->priv->enc.cx_data_dst_buf.sz = 0;
+    ctx->priv->enc.cx_data_pad_before = 0;
+    ctx->priv->enc.cx_data_pad_after = 0;
+  }
+
+  return AOM_CODEC_OK;
+}
+
+const aom_image_t *aom_codec_get_preview_frame(aom_codec_ctx_t *ctx) {
+  aom_image_t *img = NULL;
+
+  if (ctx) {
+    if (!ctx->iface || !ctx->priv)
+      ctx->err = AOM_CODEC_ERROR;
+    else if (!(ctx->iface->caps & AOM_CODEC_CAP_ENCODER))
+      ctx->err = AOM_CODEC_INCAPABLE;
+    else if (!ctx->iface->enc.get_preview)
+      ctx->err = AOM_CODEC_INCAPABLE;
+    else
+      img = ctx->iface->enc.get_preview(get_alg_priv(ctx));
+  }
+
+  return img;
+}
+
+aom_fixed_buf_t *aom_codec_get_global_headers(aom_codec_ctx_t *ctx) {
+  aom_fixed_buf_t *buf = NULL;
+
+  if (ctx) {
+    if (!ctx->iface || !ctx->priv)
+      ctx->err = AOM_CODEC_ERROR;
+    else if (!(ctx->iface->caps & AOM_CODEC_CAP_ENCODER))
+      ctx->err = AOM_CODEC_INCAPABLE;
+    else if (!ctx->iface->enc.get_glob_hdrs)
+      ctx->err = AOM_CODEC_INCAPABLE;
+    else
+      buf = ctx->iface->enc.get_glob_hdrs(get_alg_priv(ctx));
+  }
+
+  return buf;
+}
+
+aom_codec_err_t aom_codec_enc_config_set(aom_codec_ctx_t *ctx,
+                                         const aom_codec_enc_cfg_t *cfg) {
+  aom_codec_err_t res;
+
+  if (!ctx || !ctx->iface || !ctx->priv || !cfg)
+    res = AOM_CODEC_INVALID_PARAM;
+  else if (!(ctx->iface->caps & AOM_CODEC_CAP_ENCODER))
+    res = AOM_CODEC_INCAPABLE;
+  else
+    res = ctx->iface->enc.cfg_set(get_alg_priv(ctx), cfg);
+
+  return SAVE_STATUS(ctx, res);
+}
+
+int aom_codec_pkt_list_add(struct aom_codec_pkt_list *list,
+                           const struct aom_codec_cx_pkt *pkt) {
+  if (list->cnt < list->max) {
+    list->pkts[list->cnt++] = *pkt;
+    return 0;
+  }
+
+  return 1;
+}
+
+const aom_codec_cx_pkt_t *aom_codec_pkt_list_get(
+    struct aom_codec_pkt_list *list, aom_codec_iter_t *iter) {
+  const aom_codec_cx_pkt_t *pkt;
+
+  if (!(*iter)) {
+    *iter = list->pkts;
+  }
+
+  pkt = (const aom_codec_cx_pkt_t *)*iter;
+
+  if ((size_t)(pkt - list->pkts) < list->cnt)
+    *iter = pkt + 1;
+  else
+    pkt = NULL;
+
+  return pkt;
+}
diff --git a/third_party/aom/aom/src/aom_image.c b/third_party/aom/aom/src/aom_image.c
new file mode 100644
index 0000000000..0d54fd46db
--- /dev/null
+++ b/third_party/aom/aom/src/aom_image.c
@@ -0,0 +1,240 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aom_image.h"
+#include "aom/aom_integer.h"
+#include "aom_mem/aom_mem.h"
+
+static aom_image_t *img_alloc_helper(aom_image_t *img, aom_img_fmt_t fmt,
+                                     unsigned int d_w, unsigned int d_h,
+                                     unsigned int buf_align,
+                                     unsigned int stride_align,
+                                     unsigned char *img_data) {
+  unsigned int h, w, s, xcs, ycs, bps;
+  unsigned int stride_in_bytes;
+  int align;
+
+  /* Treat align==0 like align==1 */
+  if (!buf_align) buf_align = 1;
+
+  /* Validate alignment (must be power of 2) */
+  if (buf_align & (buf_align - 1)) goto fail;
+
+  /* Treat align==0 like align==1 */
+  if (!stride_align) stride_align = 1;
+
+  /* Validate alignment (must be power of 2) */
+  if (stride_align & (stride_align - 1)) goto fail;
+
+  /* Get sample size for this format */
+  switch (fmt) {
+    case AOM_IMG_FMT_RGB32:
+    case AOM_IMG_FMT_RGB32_LE:
+    case AOM_IMG_FMT_ARGB:
+    case AOM_IMG_FMT_ARGB_LE: bps = 32; break;
+    case AOM_IMG_FMT_RGB24:
+    case AOM_IMG_FMT_BGR24: bps = 24; break;
+    case AOM_IMG_FMT_RGB565:
+    case AOM_IMG_FMT_RGB565_LE:
+    case AOM_IMG_FMT_RGB555:
+    case AOM_IMG_FMT_RGB555_LE:
+    case AOM_IMG_FMT_UYVY:
+    case AOM_IMG_FMT_YUY2:
+    case AOM_IMG_FMT_YVYU: bps = 16; break;
+    case AOM_IMG_FMT_I420:
+    case AOM_IMG_FMT_YV12:
+    case AOM_IMG_FMT_AOMI420:
+    case AOM_IMG_FMT_AOMYV12: bps = 12; break;
+    case AOM_IMG_FMT_I422:
+    case AOM_IMG_FMT_I440: bps = 16; break;
+    case AOM_IMG_FMT_I444: bps = 24; break;
+    case AOM_IMG_FMT_I42016: bps = 24; break;
+    case AOM_IMG_FMT_I42216:
+    case AOM_IMG_FMT_I44016: bps = 32; break;
+    case AOM_IMG_FMT_I44416: bps = 48; break;
+    default: bps = 16; break;
+  }
+
+  /* Get chroma shift values for this format */
+  switch (fmt) {
+    case AOM_IMG_FMT_I420:
+    case AOM_IMG_FMT_YV12:
+    case AOM_IMG_FMT_AOMI420:
+    case AOM_IMG_FMT_AOMYV12:
+    case AOM_IMG_FMT_I422:
+    case AOM_IMG_FMT_I42016:
+    case AOM_IMG_FMT_I42216: xcs = 1; break;
+    default: xcs = 0; break;
+  }
+
+  switch (fmt) {
+    case AOM_IMG_FMT_I420:
+    case AOM_IMG_FMT_I440:
+    case AOM_IMG_FMT_YV12:
+    case AOM_IMG_FMT_AOMI420:
+    case AOM_IMG_FMT_AOMYV12:
+    case AOM_IMG_FMT_I42016:
+    case AOM_IMG_FMT_I44016: ycs = 1; break;
+    default: ycs = 0; break;
+  }
+
+  /* Calculate storage sizes given the chroma subsampling */
+  align = (1 << xcs) - 1;
+  w = (d_w + align) & ~align;
+  align = (1 << ycs) - 1;
+  h = (d_h + align) & ~align;
+  s = (fmt & AOM_IMG_FMT_PLANAR) ? w : bps * w / 8;
+  s = (s + stride_align - 1) & ~(stride_align - 1);
+  stride_in_bytes = (fmt & AOM_IMG_FMT_HIGHBITDEPTH) ? s * 2 : s;
+
+  /* Allocate the new image */
+  if (!img) {
+    img = (aom_image_t *)calloc(1, sizeof(aom_image_t));
+
+    if (!img) goto fail;
+
+    img->self_allocd = 1;
+  } else {
+    memset(img, 0, sizeof(aom_image_t));
+  }
+
+  img->img_data = img_data;
+
+  if (!img_data) {
+    const uint64_t alloc_size = (fmt & AOM_IMG_FMT_PLANAR)
+                                    ? (uint64_t)h * s * bps / 8
+                                    : (uint64_t)h * s;
+
+    if (alloc_size != (size_t)alloc_size) goto fail;
+
+    img->img_data = (uint8_t *)aom_memalign(buf_align, (size_t)alloc_size);
+    img->img_data_owner = 1;
+  }
+
+  if (!img->img_data) goto fail;
+
+  img->fmt = fmt;
+  img->bit_depth = (fmt & AOM_IMG_FMT_HIGHBITDEPTH) ? 16 : 8;
+  img->w = w;
+  img->h = h;
+  img->x_chroma_shift = xcs;
+  img->y_chroma_shift = ycs;
+  img->bps = bps;
+
+  /* Calculate strides */
+  img->stride[AOM_PLANE_Y] = img->stride[AOM_PLANE_ALPHA] = stride_in_bytes;
+  img->stride[AOM_PLANE_U] = img->stride[AOM_PLANE_V] = stride_in_bytes >> xcs;
+
+  /* Default viewport to entire image */
+  if (!aom_img_set_rect(img, 0, 0, d_w, d_h)) return img;
+
+fail:
+  aom_img_free(img);
+  return NULL;
+}
+
+aom_image_t *aom_img_alloc(aom_image_t *img, aom_img_fmt_t fmt,
+                           unsigned int d_w, unsigned int d_h,
+                           unsigned int align) {
+  return img_alloc_helper(img, fmt, d_w, d_h, align, align, NULL);
+}
+
+aom_image_t *aom_img_wrap(aom_image_t *img, aom_img_fmt_t fmt, unsigned int d_w,
+                          unsigned int d_h, unsigned int stride_align,
+                          unsigned char *img_data) {
+  /* By setting buf_align = 1, we don't change buffer alignment in this
+   * function. */
+  return img_alloc_helper(img, fmt, d_w, d_h, 1, stride_align, img_data);
+}
+
+int aom_img_set_rect(aom_image_t *img, unsigned int x, unsigned int y,
+                     unsigned int w, unsigned int h) {
+  unsigned char *data;
+
+  if (x + w <= img->w && y + h <= img->h) {
+    img->d_w = w;
+    img->d_h = h;
+
+    /* Calculate plane pointers */
+    if (!(img->fmt & AOM_IMG_FMT_PLANAR)) {
+      img->planes[AOM_PLANE_PACKED] =
+          img->img_data + x * img->bps / 8 + y * img->stride[AOM_PLANE_PACKED];
+    } else {
+      const int bytes_per_sample =
+          (img->fmt & AOM_IMG_FMT_HIGHBITDEPTH) ? 2 : 1;
+      data = img->img_data;
+
+      if (img->fmt & AOM_IMG_FMT_HAS_ALPHA) {
+        img->planes[AOM_PLANE_ALPHA] =
+            data + x * bytes_per_sample + y * img->stride[AOM_PLANE_ALPHA];
+        data += img->h * img->stride[AOM_PLANE_ALPHA];
+      }
+
+      img->planes[AOM_PLANE_Y] =
+          data + x * bytes_per_sample + y * img->stride[AOM_PLANE_Y];
+      data += img->h * img->stride[AOM_PLANE_Y];
+
+      if (!(img->fmt & AOM_IMG_FMT_UV_FLIP)) {
+        img->planes[AOM_PLANE_U] =
+            data + (x >> img->x_chroma_shift) * bytes_per_sample +
+            (y >> img->y_chroma_shift) * img->stride[AOM_PLANE_U];
+        data += (img->h >> img->y_chroma_shift) * img->stride[AOM_PLANE_U];
+        img->planes[AOM_PLANE_V] =
+            data + (x >> img->x_chroma_shift) * bytes_per_sample +
+            (y >> img->y_chroma_shift) * img->stride[AOM_PLANE_V];
+      } else {
+        img->planes[AOM_PLANE_V] =
+            data + (x >> img->x_chroma_shift) * bytes_per_sample +
+            (y >> img->y_chroma_shift) * img->stride[AOM_PLANE_V];
+        data += (img->h >> img->y_chroma_shift) * img->stride[AOM_PLANE_V];
+        img->planes[AOM_PLANE_U] =
+            data + (x >> img->x_chroma_shift) * bytes_per_sample +
+            (y >> img->y_chroma_shift) * img->stride[AOM_PLANE_U];
+      }
+    }
+    return 0;
+  }
+  return -1;
+}
+
+void aom_img_flip(aom_image_t *img) {
+  /* Note: In the calculation pointer adjustment calculation, we want the
+   * rhs to be promoted to a signed type. Section 6.3.1.8 of the ISO C99
+   * standard indicates that if the adjustment parameter is unsigned, the
+   * stride parameter will be promoted to unsigned, causing errors when
+   * the lhs is a larger type than the rhs.
+   */
+  img->planes[AOM_PLANE_Y] += (signed)(img->d_h - 1) * img->stride[AOM_PLANE_Y];
+  img->stride[AOM_PLANE_Y] = -img->stride[AOM_PLANE_Y];
+
+  img->planes[AOM_PLANE_U] += (signed)((img->d_h >> img->y_chroma_shift) - 1) *
+                              img->stride[AOM_PLANE_U];
+  img->stride[AOM_PLANE_U] = -img->stride[AOM_PLANE_U];
+
+  img->planes[AOM_PLANE_V] += (signed)((img->d_h >> img->y_chroma_shift) - 1) *
+                              img->stride[AOM_PLANE_V];
+  img->stride[AOM_PLANE_V] = -img->stride[AOM_PLANE_V];
+
+  img->planes[AOM_PLANE_ALPHA] +=
+      (signed)(img->d_h - 1) * img->stride[AOM_PLANE_ALPHA];
+  img->stride[AOM_PLANE_ALPHA] = -img->stride[AOM_PLANE_ALPHA];
+}
+
+void aom_img_free(aom_image_t *img) {
+  if (img) {
+    if (img->img_data && img->img_data_owner) aom_free(img->img_data);
+
+    if (img->self_allocd) free(img);
+  }
+}
diff --git a/third_party/aom/aom_dsp/add_noise.c b/third_party/aom/aom_dsp/add_noise.c
new file mode 100644
index 0000000000..389cf2049f
--- /dev/null
+++ b/third_party/aom/aom_dsp/add_noise.c
@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+void aom_plane_add_noise_c(uint8_t *start, char *noise, char blackclamp[16],
+                           char whiteclamp[16], char bothclamp[16],
+                           unsigned int width, unsigned int height, int pitch) {
+  unsigned int i, j;
+
+  for (i = 0; i < height; ++i) {
+    uint8_t *pos = start + i * pitch;
+    char *ref = (char *)(noise + (rand() & 0xff));  // NOLINT
+
+    for (j = 0; j < width; ++j) {
+      int v = pos[j];
+
+      v = clamp(v - blackclamp[0], 0, 255);
+      v = clamp(v + bothclamp[0], 0, 255);
+      v = clamp(v - whiteclamp[0], 0, 255);
+
+      pos[j] = v + ref[j];
+    }
+  }
+}
+
+static double gaussian(double sigma, double mu, double x) {
+  return 1 / (sigma * sqrt(2.0 * 3.14159265)) *
+         (exp(-(x - mu) * (x - mu) / (2 * sigma * sigma)));
+}
+
+int aom_setup_noise(double sigma, int size, char *noise) {
+  char char_dist[256];
+  int next = 0, i, j;
+
+  // set up a 256 entry lookup that matches gaussian distribution
+  for (i = -32; i < 32; ++i) {
+    const int a_i = (int)(0.5 + 256 * gaussian(sigma, 0, i));
+    if (a_i) {
+      for (j = 0; j < a_i; ++j) {
+        char_dist[next + j] = (char)i;
+      }
+      next = next + j;
+    }
+  }
+
+  // Rounding error - might mean we have less than 256.
+  for (; next < 256; ++next) {
+    char_dist[next] = 0;
+  }
+
+  for (i = 0; i < size; ++i) {
+    noise[i] = char_dist[rand() & 0xff];  // NOLINT
+  }
+
+  // Returns the highest non 0 value used in distribution.
+  return -char_dist[0];
+}
diff --git a/third_party/aom/aom_dsp/ans.h b/third_party/aom/aom_dsp/ans.h
new file mode 100644
index 0000000000..a7a2f0eab3
--- /dev/null
+++ b/third_party/aom/aom_dsp/ans.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_ANS_H_
+#define AOM_DSP_ANS_H_
+// Constants, types and utilities for Asymmetric Numeral Systems
+// http://arxiv.org/abs/1311.2540v2
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+// Use windowed ANS, size is passed in at initialization
+#define ANS_MAX_SYMBOLS 1
+#define ANS_REVERSE 1
+
+typedef uint8_t AnsP8;
+#define ANS_P8_PRECISION 256u
+#define ANS_P8_SHIFT 8
+#define RANS_PROB_BITS 15
+#define RANS_PRECISION (1u << RANS_PROB_BITS)
+
+// L_BASE is the ANS base state. L_BASE % PRECISION must be 0.
+#define L_BASE (1u << 17)
+#define IO_BASE 256
+// Range I = { L_BASE, L_BASE + 1, ..., L_BASE * IO_BASE - 1 }
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_DSP_ANS_H_
diff --git a/third_party/aom/aom_dsp/ansreader.h b/third_party/aom/aom_dsp/ansreader.h
new file mode 100644
index 0000000000..e50c63b2d1
--- /dev/null
+++ b/third_party/aom/aom_dsp/ansreader.h
@@ -0,0 +1,214 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_ANSREADER_H_
+#define AOM_DSP_ANSREADER_H_
+// An implementation of Asymmetric Numeral Systems
+// http://arxiv.org/abs/1311.2540v2
+// Implements decoding of:
+// * rABS (range Asymmetric Binary Systems), a boolean coder
+// * rANS (range Asymmetric Numeral Systems), a multi-symbol coder
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/prob.h"
+#include "aom_dsp/ans.h"
+#include "aom_ports/mem_ops.h"
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+struct AnsDecoder {
+  const uint8_t *buf;
+  int buf_offset;
+  uint32_t state;
+#if ANS_MAX_SYMBOLS
+  int symbols_left;
+  int window_size;
+#endif
+#if CONFIG_ACCOUNTING
+  Accounting *accounting;
+#endif
+};
+
+static INLINE int ans_read_reinit(struct AnsDecoder *const ans);
+
+static INLINE unsigned refill_state(struct AnsDecoder *const ans,
+                                    unsigned state) {
+#if ANS_REVERSE
+  while (state < L_BASE && ans->buf_offset < 0) {
+    state = state * IO_BASE + ans->buf[ans->buf_offset++];
+  }
+#else
+  while (state < L_BASE && ans->buf_offset > 0) {
+    state = state * IO_BASE + ans->buf[--ans->buf_offset];
+  }
+#endif
+  return state;
+}
+
+// Decode one rABS encoded boolean where the probability of the value being zero
+// is p0.
+static INLINE int rabs_read(struct AnsDecoder *ans, AnsP8 p0) {
+#if ANS_MAX_SYMBOLS
+  if (ans->symbols_left-- == 0) {
+    ans_read_reinit(ans);
+    ans->symbols_left--;
+  }
+#endif
+  unsigned state = refill_state(ans, ans->state);
+  const unsigned quotient = state / ANS_P8_PRECISION;
+  const unsigned remainder = state % ANS_P8_PRECISION;
+  const int value = remainder >= p0;
+  const unsigned qp0 = quotient * p0;
+  if (value)
+    state = state - qp0 - p0;
+  else
+    state = qp0 + remainder;
+  ans->state = state;
+  return value;
+}
+
+// Decode one rABS encoded boolean where the probability of the value being zero
+// is one half.
+static INLINE int rabs_read_bit(struct AnsDecoder *ans) {
+#if ANS_MAX_SYMBOLS
+  if (ans->symbols_left-- == 0) {
+    ans_read_reinit(ans);
+    ans->symbols_left--;
+  }
+#endif
+  unsigned state = refill_state(ans, ans->state);
+  const int value = !!(state & 0x80);
+  ans->state = ((state >> 1) & ~0x7F) | (state & 0x7F);
+  return value;
+}
+
+struct rans_dec_sym {
+  uint8_t val;
+  aom_cdf_prob prob;
+  aom_cdf_prob cum_prob;  // not-inclusive
+};
+
+static INLINE void fetch_sym(struct rans_dec_sym *out, const aom_cdf_prob *cdf,
+                             aom_cdf_prob rem) {
+  int i;
+  aom_cdf_prob cum_prob = 0, top_prob;
+  // TODO(skal): if critical, could be a binary search.
+  // Or, better, an O(1) alias-table.
+  for (i = 0; rem >= (top_prob = cdf[i]); ++i) {
+    cum_prob = top_prob;
+  }
+  out->val = i;
+  out->prob = top_prob - cum_prob;
+  out->cum_prob = cum_prob;
+}
+
+static INLINE int rans_read(struct AnsDecoder *ans, const aom_cdf_prob *tab) {
+  unsigned rem;
+  unsigned quo;
+  struct rans_dec_sym sym;
+#if ANS_MAX_SYMBOLS
+  if (ans->symbols_left-- == 0) {
+    ans_read_reinit(ans);
+    ans->symbols_left--;
+  }
+#endif
+  ans->state = refill_state(ans, ans->state);
+  quo = ans->state / RANS_PRECISION;
+  rem = ans->state % RANS_PRECISION;
+  fetch_sym(&sym, tab, rem);
+  ans->state = quo * sym.prob + rem - sym.cum_prob;
+  return sym.val;
+}
+
+static INLINE int ans_read_init(struct AnsDecoder *const ans,
+                                const uint8_t *const buf, int offset) {
+  unsigned x;
+  if (offset < 1) return 1;
+#if ANS_REVERSE
+  ans->buf = buf + offset;
+  ans->buf_offset = -offset;
+  x = buf[0];
+  if ((x & 0x80) == 0) {  // Marker is 0xxx xxxx
+    if (offset < 2) return 1;
+    ans->buf_offset += 2;
+    ans->state = mem_get_be16(buf) & 0x7FFF;
+#if L_BASE * IO_BASE > (1 << 23)
+  } else if ((x & 0xC0) == 0x80) {  // Marker is 10xx xxxx
+    if (offset < 3) return 1;
+    ans->buf_offset += 3;
+    ans->state = mem_get_be24(buf) & 0x3FFFFF;
+  } else {  // Marker is 11xx xxxx
+    if (offset < 4) return 1;
+    ans->buf_offset += 4;
+    ans->state = mem_get_be32(buf) & 0x3FFFFFFF;
+#else
+  } else {  // Marker is 1xxx xxxx
+    if (offset < 3) return 1;
+    ans->buf_offset += 3;
+    ans->state = mem_get_be24(buf) & 0x7FFFFF;
+#endif
+  }
+#else
+  ans->buf = buf;
+  x = buf[offset - 1];
+  if ((x & 0x80) == 0) {  // Marker is 0xxx xxxx
+    if (offset < 2) return 1;
+    ans->buf_offset = offset - 2;
+    ans->state = mem_get_le16(buf + offset - 2) & 0x7FFF;
+  } else if ((x & 0xC0) == 0x80) {  // Marker is 10xx xxxx
+    if (offset < 3) return 1;
+    ans->buf_offset = offset - 3;
+    ans->state = mem_get_le24(buf + offset - 3) & 0x3FFFFF;
+  } else if ((x & 0xE0) == 0xE0) {  // Marker is 111x xxxx
+    if (offset < 4) return 1;
+    ans->buf_offset = offset - 4;
+    ans->state = mem_get_le32(buf + offset - 4) & 0x1FFFFFFF;
+  } else {
+    // Marker 110x xxxx implies this byte is a superframe marker
+    return 1;
+  }
+#endif  // ANS_REVERSE
+#if CONFIG_ACCOUNTING
+  ans->accounting = NULL;
+#endif
+  ans->state += L_BASE;
+  if (ans->state >= L_BASE * IO_BASE) return 1;
+#if ANS_MAX_SYMBOLS
+  assert(ans->window_size > 1);
+  ans->symbols_left = ans->window_size;
+#endif
+  return 0;
+}
+
+#if ANS_REVERSE
+static INLINE int ans_read_reinit(struct AnsDecoder *const ans) {
+  return ans_read_init(ans, ans->buf + ans->buf_offset, -ans->buf_offset);
+}
+#endif
+
+static INLINE int ans_read_end(const struct AnsDecoder *const ans) {
+  return ans->buf_offset == 0 && ans->state < L_BASE;
+}
+
+static INLINE int ans_reader_has_error(const struct AnsDecoder *const ans) {
+  return ans->state < L_BASE / RANS_PRECISION;
+}
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_DSP_ANSREADER_H_
diff --git a/third_party/aom/aom_dsp/answriter.h b/third_party/aom/aom_dsp/answriter.h
new file mode 100644
index 0000000000..353acf1a94
--- /dev/null
+++ b/third_party/aom/aom_dsp/answriter.h
@@ -0,0 +1,148 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_ANSWRITER_H_
+#define AOM_DSP_ANSWRITER_H_
+// An implementation of Asymmetric Numeral Systems
+// http://arxiv.org/abs/1311.2540v2
+// Implements encoding of:
+// * rABS (range Asymmetric Binary Systems), a boolean coder
+// * rANS (range Asymmetric Numeral Systems), a multi-symbol coder
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/ans.h"
+#include "aom_dsp/prob.h"
+#include "aom_ports/mem_ops.h"
+#include "av1/common/odintrin.h"
+
+#if RANS_PRECISION <= OD_DIVU_DMAX
+#define ANS_DIVREM(quotient, remainder, dividend, divisor) \
+  do {                                                     \
+    quotient = OD_DIVU_SMALL((dividend), (divisor));       \
+    remainder = (dividend) - (quotient) * (divisor);       \
+  } while (0)
+#else
+#define ANS_DIVREM(quotient, remainder, dividend, divisor) \
+  do {                                                     \
+    quotient = (dividend) / (divisor);                     \
+    remainder = (dividend) % (divisor);                    \
+  } while (0)
+#endif
+
+#define ANS_DIV8(dividend, divisor) OD_DIVU_SMALL((dividend), (divisor))
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+struct AnsCoder {
+  uint8_t *buf;
+  int buf_offset;
+  uint32_t state;
+};
+
+static INLINE void ans_write_init(struct AnsCoder *const ans,
+                                  uint8_t *const buf) {
+  ans->buf = buf;
+  ans->buf_offset = 0;
+  ans->state = L_BASE;
+}
+
+static INLINE int ans_write_end(struct AnsCoder *const ans) {
+  uint32_t state;
+  int ans_size;
+  assert(ans->state >= L_BASE);
+  assert(ans->state < L_BASE * IO_BASE);
+  state = ans->state - L_BASE;
+  if (state < (1u << 15)) {
+    mem_put_le16(ans->buf + ans->buf_offset, (0x00u << 15) + state);
+    ans_size = ans->buf_offset + 2;
+#if ANS_REVERSE
+#if L_BASE * IO_BASE > (1 << 23)
+  } else if (state < (1u << 22)) {
+    mem_put_le24(ans->buf + ans->buf_offset, (0x02u << 22) + state);
+    ans_size = ans->buf_offset + 3;
+  } else if (state < (1u << 30)) {
+    mem_put_le32(ans->buf + ans->buf_offset, (0x03u << 30) + state);
+    ans_size = ans->buf_offset + 4;
+#else
+  } else if (state < (1u << 23)) {
+    mem_put_le24(ans->buf + ans->buf_offset, (0x01u << 23) + state);
+    ans_size = ans->buf_offset + 3;
+#endif
+#else
+  } else if (state < (1u << 22)) {
+    mem_put_le24(ans->buf + ans->buf_offset, (0x02u << 22) + state);
+    ans_size = ans->buf_offset + 3;
+  } else if (state < (1u << 29)) {
+    mem_put_le32(ans->buf + ans->buf_offset, (0x07u << 29) + state);
+    ans_size = ans->buf_offset + 4;
+#endif
+  } else {
+    assert(0 && "State is too large to be serialized");
+    return ans->buf_offset;
+  }
+#if ANS_REVERSE
+  {
+    int i;
+    uint8_t tmp;
+    for (i = 0; i < (ans_size >> 1); i++) {
+      tmp = ans->buf[i];
+      ans->buf[i] = ans->buf[ans_size - 1 - i];
+      ans->buf[ans_size - 1 - i] = tmp;
+    }
+    ans->buf += ans_size;
+    ans->buf_offset = 0;
+    ans->state = L_BASE;
+  }
+#endif
+  return ans_size;
+}
+
+// Write one boolean using rABS where p0 is the probability of the value being
+// zero.
+static INLINE void rabs_write(struct AnsCoder *ans, int value, AnsP8 p0) {
+  const AnsP8 p = ANS_P8_PRECISION - p0;
+  const unsigned l_s = value ? p : p0;
+  unsigned state = ans->state;
+  while (state >= L_BASE / ANS_P8_PRECISION * IO_BASE * l_s) {
+    ans->buf[ans->buf_offset++] = state % IO_BASE;
+    state /= IO_BASE;
+  }
+  const unsigned quotient = ANS_DIV8(state, l_s);
+  const unsigned remainder = state - quotient * l_s;
+  ans->state = quotient * ANS_P8_PRECISION + remainder + (value ? p0 : 0);
+}
+
+// Encode one symbol using rANS.
+// cum_prob: The cumulative probability before this symbol (the offset of
+// the symbol in the symbol cycle)
+// prob: The probability of this symbol (l_s from the paper)
+// RANS_PRECISION takes the place of m from the paper.
+static INLINE void rans_write(struct AnsCoder *ans, aom_cdf_prob cum_prob,
+                              aom_cdf_prob prob) {
+  unsigned quotient, remainder;
+  while (ans->state >= L_BASE / RANS_PRECISION * IO_BASE * prob) {
+    ans->buf[ans->buf_offset++] = ans->state % IO_BASE;
+    ans->state /= IO_BASE;
+  }
+  ANS_DIVREM(quotient, remainder, ans->state, prob);
+  ans->state = quotient * RANS_PRECISION + remainder + cum_prob;
+}
+
+#undef ANS_DIV8
+#undef ANS_DIVREM
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_DSP_ANSWRITER_H_
diff --git a/third_party/aom/aom_dsp/aom_convolve.c b/third_party/aom/aom_dsp/aom_convolve.c
new file mode 100644
index 0000000000..74f4c00fbb
--- /dev/null
+++ b/third_party/aom/aom_dsp/aom_convolve.c
@@ -0,0 +1,854 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <string.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_ports/mem.h"
+
+static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const InterpKernel *x_filters, int x0_q4,
+                           int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const InterpKernel *x_filters, int x0_q4,
+                               int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = ROUND_POWER_OF_TWO(
+          dst[x] + clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const InterpKernel *y_filters, int y0_q4,
+                          int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve_avg_vert(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const InterpKernel *y_filters, int y0_q4,
+                              int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = ROUND_POWER_OF_TWO(
+          dst[y * dst_stride] +
+              clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)),
+          1);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                     ptrdiff_t dst_stride, const InterpKernel *const x_filters,
+                     int x0_q4, int x_step_q4,
+                     const InterpKernel *const y_filters, int y0_q4,
+                     int y_step_q4, int w, int h) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint8_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
+  int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp,
+                 MAX_SB_SIZE, x_filters, x0_q4, x_step_q4, w,
+                 intermediate_height);
+  convolve_vert(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), MAX_SB_SIZE, dst,
+                dst_stride, y_filters, y0_q4, y_step_q4, w, h);
+}
+
+static const InterpKernel *get_filter_base(const int16_t *filter) {
+  // NOTE: This assumes that the filter table is 256-byte aligned.
+  // TODO(agrange) Modify to make independent of table alignment.
+  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
+}
+
+static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
+  return (int)((const InterpKernel *)(intptr_t)f - base);
+}
+
+void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4, int w,
+                           int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
+                 w, h);
+}
+
+void aom_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4, int w,
+                               int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                     x_step_q4, w, h);
+}
+
+void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const int16_t *filter_x, int x_step_q4,
+                          const int16_t *filter_y, int y_step_q4, int w,
+                          int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4, y_step_q4,
+                w, h);
+}
+
+void aom_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4, int w,
+                              int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
+                    y_step_q4, w, h);
+}
+
+void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                     ptrdiff_t dst_stride, const int16_t *filter_x,
+                     int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                     int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
+           filters_y, y0_q4, y_step_q4, w, h);
+}
+
+void aom_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const int16_t *filter_x,
+                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                         int w, int h) {
+  /* Fixed size intermediate buffer places limits on parameters. */
+  DECLARE_ALIGNED(16, uint8_t, temp[MAX_SB_SIZE * MAX_SB_SIZE]);
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+
+  aom_convolve8_c(src, src_stride, temp, MAX_SB_SIZE, filter_x, x_step_q4,
+                  filter_y, y_step_q4, w, h);
+  aom_convolve_avg_c(temp, MAX_SB_SIZE, dst, dst_stride, NULL, 0, NULL, 0, w,
+                     h);
+}
+
+void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const int16_t *filter_x,
+                         int filter_x_stride, const int16_t *filter_y,
+                         int filter_y_stride, int w, int h) {
+  int r;
+
+  (void)filter_x;
+  (void)filter_x_stride;
+  (void)filter_y;
+  (void)filter_y_stride;
+
+  for (r = h; r > 0; --r) {
+    memcpy(dst, src, w);
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void aom_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                        ptrdiff_t dst_stride, const int16_t *filter_x,
+                        int filter_x_stride, const int16_t *filter_y,
+                        int filter_y_stride, int w, int h) {
+  int x, y;
+
+  (void)filter_x;
+  (void)filter_x_stride;
+  (void)filter_y;
+  (void)filter_y_stride;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
+
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void aom_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                        ptrdiff_t dst_stride, const int16_t *filter_x,
+                        int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                        int w, int h) {
+  aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                        filter_y, y_step_q4, w, h);
+}
+
+void aom_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                       ptrdiff_t dst_stride, const int16_t *filter_x,
+                       int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                       int w, int h) {
+  aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                       filter_y, y_step_q4, w, h);
+}
+
+void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                     ptrdiff_t dst_stride, const int16_t *filter_x,
+                     int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                     int w, int h) {
+  aom_convolve8_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                  filter_y, y_step_q4, w, h);
+}
+
+void aom_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4, int w,
+                            int h) {
+  aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                            x_step_q4, filter_y, y_step_q4, w, h);
+}
+
+void aom_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4, int w,
+                           int h) {
+  aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                           x_step_q4, filter_y, y_step_q4, w, h);
+}
+
+void aom_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const int16_t *filter_x,
+                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                         int w, int h) {
+  aom_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                      filter_y, y_step_q4, w, h);
+}
+
+#if CONFIG_LOOP_RESTORATION
+static void convolve_add_src_horiz(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const InterpKernel *x_filters, int x0_q4,
+                                   int x_step_q4, int w, int h) {
+  int x, y;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS) +
+                          src_x[SUBPEL_TAPS / 2 - 1]);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_add_src_vert(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const InterpKernel *y_filters, int y0_q4,
+                                  int y_step_q4, int w, int h) {
+  int x, y;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] =
+          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS) +
+                     src_y[(SUBPEL_TAPS / 2 - 1) * src_stride]);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve_add_src(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const InterpKernel *const x_filters, int x0_q4,
+                             int x_step_q4, const InterpKernel *const y_filters,
+                             int y0_q4, int y_step_q4, int w, int h) {
+  uint8_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
+  int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  convolve_add_src_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride,
+                         temp, MAX_SB_SIZE, x_filters, x0_q4, x_step_q4, w,
+                         intermediate_height);
+  convolve_add_src_vert(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), MAX_SB_SIZE,
+                        dst, dst_stride, y_filters, y0_q4, y_step_q4, w, h);
+}
+
+void aom_convolve8_add_src_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int x_step_q4,
+                                   const int16_t *filter_y, int y_step_q4,
+                                   int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  (void)filter_y;
+  (void)y_step_q4;
+
+  convolve_add_src_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                         x_step_q4, w, h);
+}
+
+void aom_convolve8_add_src_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4, int w,
+                                  int h) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  (void)filter_x;
+  (void)x_step_q4;
+
+  convolve_add_src_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
+                        y_step_q4, w, h);
+}
+
+void aom_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int x_step_q4,
+                             const int16_t *filter_y, int y_step_q4, int w,
+                             int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  convolve_add_src(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                   x_step_q4, filters_y, y0_q4, y_step_q4, w, h);
+}
+#endif  // CONFIG_LOOP_RESTORATION
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+                                  uint8_t *dst8, ptrdiff_t dst_stride,
+                                  const InterpKernel *x_filters, int x0_q4,
+                                  int x_step_q4, int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_avg_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+                                      uint8_t *dst8, ptrdiff_t dst_stride,
+                                      const InterpKernel *x_filters, int x0_q4,
+                                      int x_step_q4, int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = ROUND_POWER_OF_TWO(
+          dst[x] + clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
+          1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride,
+                                 uint8_t *dst8, ptrdiff_t dst_stride,
+                                 const InterpKernel *y_filters, int y0_q4,
+                                 int y_step_q4, int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] =
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve_avg_vert(const uint8_t *src8, ptrdiff_t src_stride,
+                                     uint8_t *dst8, ptrdiff_t dst_stride,
+                                     const InterpKernel *y_filters, int y0_q4,
+                                     int y_step_q4, int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = ROUND_POWER_OF_TWO(
+          dst[y * dst_stride] +
+              clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
+          1);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const InterpKernel *const x_filters, int x0_q4,
+                            int x_step_q4, const InterpKernel *const y_filters,
+                            int y0_q4, int y_step_q4, int w, int h, int bd) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
+  int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  highbd_convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride,
+                        CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE, x_filters, x0_q4,
+                        x_step_q4, w, intermediate_height, bd);
+  highbd_convolve_vert(
+      CONVERT_TO_BYTEPTR(temp) + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
+      MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4, int w,
+                                  int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                        x_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                      uint8_t *dst, ptrdiff_t dst_stride,
+                                      const int16_t *filter_x, int x_step_q4,
+                                      const int16_t *filter_y, int y_step_q4,
+                                      int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                            x_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const int16_t *filter_x, int x_step_q4,
+                                 const int16_t *filter_y, int y_step_q4, int w,
+                                 int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
+                       y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                     uint8_t *dst, ptrdiff_t dst_stride,
+                                     const int16_t *filter_x, int x_step_q4,
+                                     const int16_t *filter_y, int y_step_q4,
+                                     int w, int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
+                           y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4, int w,
+                            int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  highbd_convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
+                  filters_y, y0_q4, y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int x_step_q4,
+                                const int16_t *filter_y, int y_step_q4, int w,
+                                int h, int bd) {
+  // Fixed size intermediate buffer places limits on parameters.
+  DECLARE_ALIGNED(16, uint16_t, temp[MAX_SB_SIZE * MAX_SB_SIZE]);
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+
+  aom_highbd_convolve8_c(src, src_stride, CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE,
+                         filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);
+  aom_highbd_convolve_avg_c(CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE, dst,
+                            dst_stride, NULL, 0, NULL, 0, w, h, bd);
+}
+
+void aom_highbd_convolve_copy_c(const uint8_t *src8, ptrdiff_t src_stride,
+                                uint8_t *dst8, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int filter_x_stride,
+                                const int16_t *filter_y, int filter_y_stride,
+                                int w, int h, int bd) {
+  int r;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  for (r = h; r > 0; --r) {
+    memcpy(dst, src, w * sizeof(uint16_t));
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void aom_highbd_convolve_avg_c(const uint8_t *src8, ptrdiff_t src_stride,
+                               uint8_t *dst8, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int filter_x_stride,
+                               const int16_t *filter_y, int filter_y_stride,
+                               int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+#if CONFIG_LOOP_RESTORATION
+static void highbd_convolve_add_src_horiz(const uint8_t *src8,
+                                          ptrdiff_t src_stride, uint8_t *dst8,
+                                          ptrdiff_t dst_stride,
+                                          const InterpKernel *x_filters,
+                                          int x0_q4, int x_step_q4, int w,
+                                          int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel_highbd(
+          ROUND_POWER_OF_TWO(sum, FILTER_BITS) + src_x[SUBPEL_TAPS / 2 - 1],
+          bd);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void highbd_convolve_add_src_vert(const uint8_t *src8,
+                                         ptrdiff_t src_stride, uint8_t *dst8,
+                                         ptrdiff_t dst_stride,
+                                         const InterpKernel *y_filters,
+                                         int y0_q4, int y_step_q4, int w, int h,
+                                         int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] =
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS) +
+                                src_y[(SUBPEL_TAPS / 2 - 1) * src_stride],
+                            bd);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve_add_src(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const InterpKernel *const x_filters,
+                                    int x0_q4, int x_step_q4,
+                                    const InterpKernel *const y_filters,
+                                    int y0_q4, int y_step_q4, int w, int h,
+                                    int bd) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
+  int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  highbd_convolve_add_src_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                                src_stride, CONVERT_TO_BYTEPTR(temp),
+                                MAX_SB_SIZE, x_filters, x0_q4, x_step_q4, w,
+                                intermediate_height, bd);
+  highbd_convolve_add_src_vert(
+      CONVERT_TO_BYTEPTR(temp) + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
+      MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_add_src_horiz_c(
+    const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+    ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,
+    const int16_t *filter_y, int y_step_q4, int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  highbd_convolve_add_src_horiz(src, src_stride, dst, dst_stride, filters_x,
+                                x0_q4, x_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_add_src_vert_c(const uint8_t *src,
+                                         ptrdiff_t src_stride, uint8_t *dst,
+                                         ptrdiff_t dst_stride,
+                                         const int16_t *filter_x, int x_step_q4,
+                                         const int16_t *filter_y, int y_step_q4,
+                                         int w, int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  highbd_convolve_add_src_vert(src, src_stride, dst, dst_stride, filters_y,
+                               y0_q4, y_step_q4, w, h, bd);
+}
+
+void aom_highbd_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const int16_t *filter_x, int x_step_q4,
+                                    const int16_t *filter_y, int y_step_q4,
+                                    int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  highbd_convolve_add_src(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                          x_step_q4, filters_y, y0_q4, y_step_q4, w, h, bd);
+}
+#endif  // CONFIG_LOOP_RESTORATION
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/aom_convolve.h b/third_party/aom/aom_dsp/aom_convolve.h
new file mode 100644
index 0000000000..d0de6c5d20
--- /dev/null
+++ b/third_party/aom/aom_dsp/aom_convolve.h
@@ -0,0 +1,57 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_DSP_AOM_CONVOLVE_H_
+#define AOM_DSP_AOM_CONVOLVE_H_
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Note: Fixed size intermediate buffers, place limits on parameters
+// of some functions. 2d filtering proceeds in 2 steps:
+//   (1) Interpolate horizontally into an intermediate buffer, temp.
+//   (2) Interpolate temp vertically to derive the sub-pixel result.
+// Deriving the maximum number of rows in the temp buffer (135):
+// --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+// --Largest block size is 64x64 pixels.
+// --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+//   original frame (in 1/16th pixel units).
+// --Must round-up because block may be located at sub-pixel position.
+// --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+// --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+#define MAX_EXT_SIZE 263
+#else
+#define MAX_EXT_SIZE 135
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+
+typedef void (*convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4, int w,
+                              int h);
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*highbd_convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
+                                     uint8_t *dst, ptrdiff_t dst_stride,
+                                     const int16_t *filter_x, int x_step_q4,
+                                     const int16_t *filter_y, int y_step_q4,
+                                     int w, int h, int bd);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_AOM_CONVOLVE_H_
diff --git a/third_party/aom/aom_dsp/aom_dsp.cmake b/third_party/aom/aom_dsp/aom_dsp.cmake
new file mode 100644
index 0000000000..f00348cbcf
--- /dev/null
+++ b/third_party/aom/aom_dsp/aom_dsp.cmake
@@ -0,0 +1,509 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+set(AOM_DSP_COMMON_SOURCES
+    "${AOM_ROOT}/aom_dsp/aom_convolve.c"
+    "${AOM_ROOT}/aom_dsp/aom_convolve.h"
+    "${AOM_ROOT}/aom_dsp/aom_dsp_common.h"
+    "${AOM_ROOT}/aom_dsp/aom_filter.h"
+    "${AOM_ROOT}/aom_dsp/aom_simd.h"
+    "${AOM_ROOT}/aom_dsp/aom_simd_inline.h"
+    "${AOM_ROOT}/aom_dsp/blend.h"
+    "${AOM_ROOT}/aom_dsp/blend_a64_hmask.c"
+    "${AOM_ROOT}/aom_dsp/blend_a64_mask.c"
+    "${AOM_ROOT}/aom_dsp/blend_a64_vmask.c"
+    "${AOM_ROOT}/aom_dsp/intrapred.c"
+    "${AOM_ROOT}/aom_dsp/loopfilter.c"
+    "${AOM_ROOT}/aom_dsp/prob.c"
+    "${AOM_ROOT}/aom_dsp/prob.h"
+    "${AOM_ROOT}/aom_dsp/sad.c"
+    "${AOM_ROOT}/aom_dsp/simd/v128_intrinsics.h"
+    "${AOM_ROOT}/aom_dsp/simd/v128_intrinsics_c.h"
+    "${AOM_ROOT}/aom_dsp/simd/v256_intrinsics.h"
+    "${AOM_ROOT}/aom_dsp/simd/v256_intrinsics_c.h"
+    "${AOM_ROOT}/aom_dsp/simd/v64_intrinsics.h"
+    "${AOM_ROOT}/aom_dsp/simd/v64_intrinsics_c.h"
+    "${AOM_ROOT}/aom_dsp/subtract.c"
+    "${AOM_ROOT}/aom_dsp/txfm_common.h"
+    "${AOM_ROOT}/aom_dsp/x86/txfm_common_intrin.h")
+
+set(AOM_DSP_COMMON_ASM_SSE2
+    "${AOM_ROOT}/aom_dsp/x86/aom_convolve_copy_sse2.asm"
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_sse2.asm"
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm"
+    "${AOM_ROOT}/aom_dsp/x86/intrapred_sse2.asm")
+
+set(AOM_DSP_COMMON_INTRIN_SSE2
+    "${AOM_ROOT}/aom_dsp/x86/aom_asm_stubs.c"
+    "${AOM_ROOT}/aom_dsp/x86/convolve.h"
+    "${AOM_ROOT}/aom_dsp/x86/txfm_common_sse2.h"
+    "${AOM_ROOT}/aom_dsp/x86/loopfilter_sse2.c")
+
+set(AOM_DSP_COMMON_ASM_SSSE3
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_ssse3.asm"
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm"
+    "${AOM_ROOT}/aom_dsp/x86/intrapred_ssse3.asm")
+
+set(AOM_DSP_COMMON_INTRIN_SSSE3
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c"
+    "${AOM_ROOT}/aom_dsp/x86/inv_txfm_ssse3.c")
+
+set(AOM_DSP_COMMON_INTRIN_SSE4_1
+    "${AOM_ROOT}/aom_dsp/x86/blend_a64_hmask_sse4.c"
+    "${AOM_ROOT}/aom_dsp/x86/blend_a64_mask_sse4.c"
+    "${AOM_ROOT}/aom_dsp/x86/blend_a64_vmask_sse4.c")
+
+set(AOM_DSP_COMMON_INTRIN_AVX2
+    "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c"
+    "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_avx2.c"
+    "${AOM_ROOT}/aom_dsp/x86/loopfilter_avx2.c")
+
+set(AOM_DSP_COMMON_ASM_NEON
+    "${AOM_ROOT}/aom_dsp/arm/aom_convolve8_avg_neon_asm.asm"
+    "${AOM_ROOT}/aom_dsp/arm/aom_convolve8_neon_asm.asm"
+    "${AOM_ROOT}/aom_dsp/arm/aom_convolve_avg_neon_asm.asm"
+    "${AOM_ROOT}/aom_dsp/arm/aom_convolve_copy_neon_asm.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct16x16_1_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct16x16_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct32x32_1_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct32x32_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct4x4_1_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct4x4_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct8x8_1_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/idct8x8_add_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/intrapred_neon_asm.asm"
+    "${AOM_ROOT}/aom_dsp/arm/loopfilter_16_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/loopfilter_4_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/loopfilter_8_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/loopfilter_mb_neon.asm"
+    "${AOM_ROOT}/aom_dsp/arm/save_reg_neon.asm")
+
+set(AOM_DSP_COMMON_INTRIN_NEON
+    "${AOM_ROOT}/aom_dsp/arm/aom_convolve_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/avg_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/fwd_txfm_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/hadamard_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/idct16x16_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/intrapred_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/loopfilter_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/sad4d_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/sad_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/subpel_variance_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/subtract_neon.c"
+    "${AOM_ROOT}/aom_dsp/arm/variance_neon.c")
+
+if ("${AOM_TARGET_CPU}" STREQUAL "arm64")
+  set(AOM_DSP_COMMON_INTRIN_NEON
+      ${AOM_DSP_COMMON_INTRIN_NEON}
+      "${AOM_ROOT}/aom_dsp/arm/aom_convolve8_avg_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/aom_convolve8_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/aom_convolve_avg_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/aom_convolve_copy_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct16x16_1_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct16x16_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct32x32_1_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct32x32_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct4x4_1_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct4x4_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct8x8_1_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/idct8x8_add_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/intrapred_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/loopfilter_16_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/loopfilter_4_neon.c"
+      "${AOM_ROOT}/aom_dsp/arm/loopfilter_8_neon.c")
+endif ()
+
+set(AOM_DSP_COMMON_INTRIN_DSPR2
+    "${AOM_ROOT}/aom_dsp/mips/common_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/common_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/convolve2_avg_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve2_avg_horiz_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve2_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve2_horiz_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve2_vert_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve8_avg_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve8_avg_horiz_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve8_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve8_horiz_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve8_vert_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/convolve_common_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/intrapred16_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/intrapred4_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/intrapred8_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/inv_txfm_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_filters_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_filters_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_macros_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_masks_dspr2.h"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_mb_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_mb_horiz_dspr2.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_mb_vert_dspr2.c")
+
+set(AOM_DSP_COMMON_INTRIN_MSA
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_avg_horiz_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_avg_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_avg_vert_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_horiz_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve8_vert_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve_avg_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve_copy_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/aom_convolve_msa.h"
+    "${AOM_ROOT}/aom_dsp/mips/fwd_dct32x32_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/fwd_txfm_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/fwd_txfm_msa.h"
+    "${AOM_ROOT}/aom_dsp/mips/idct16x16_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/idct32x32_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/idct4x4_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/idct8x8_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/intrapred_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/inv_txfm_msa.h"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_16_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_4_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_8_msa.c"
+    "${AOM_ROOT}/aom_dsp/mips/loopfilter_msa.h"
+    "${AOM_ROOT}/aom_dsp/mips/macros_msa.h"
+    "${AOM_ROOT}/aom_dsp/mips/txfm_macros_msa.h")
+
+if (CONFIG_HIGHBITDEPTH)
+  set(AOM_DSP_COMMON_ASM_SSE2
+      ${AOM_DSP_COMMON_ASM_SSE2}
+      "${AOM_ROOT}/aom_dsp/x86/highbd_intrapred_sse2.asm")
+
+  set(AOM_DSP_COMMON_INTRIN_SSE2
+      ${AOM_DSP_COMMON_INTRIN_SSE2}
+      "${AOM_ROOT}/aom_dsp/x86/highbd_loopfilter_sse2.c")
+
+  set(AOM_DSP_COMMON_INTRIN_AVX2
+      ${AOM_DSP_COMMON_INTRIN_AVX2}
+      "${AOM_ROOT}/aom_dsp/x86/highbd_convolve_avx2.c")
+else ()
+  set(AOM_DSP_COMMON_INTRIN_DSPR2
+      ${AOM_DSP_COMMON_INTRIN_DSPR2}
+      "${AOM_ROOT}/aom_dsp/mips/itrans16_dspr2.c"
+      "${AOM_ROOT}/aom_dsp/mips/itrans32_cols_dspr2.c"
+      "${AOM_ROOT}/aom_dsp/mips/itrans32_dspr2.c"
+      "${AOM_ROOT}/aom_dsp/mips/itrans4_dspr2.c"
+      "${AOM_ROOT}/aom_dsp/mips/itrans8_dspr2.c")
+endif ()
+
+if (CONFIG_ANS)
+  set(AOM_DSP_COMMON_SOURCES
+      ${AOM_DSP_COMMON_SOURCES}
+      "${AOM_ROOT}/aom_dsp/ans.h")
+elseif (CONFIG_DAALA_EC)
+  set(AOM_DSP_COMMON_SOURCES
+      ${AOM_DSP_COMMON_SOURCES}
+      "${AOM_ROOT}/aom_dsp/entcode.c"
+      "${AOM_ROOT}/aom_dsp/entcode.h")
+endif ()
+
+if (CONFIG_AV1)
+  set(AOM_DSP_COMMON_SOURCES
+      ${AOM_DSP_COMMON_SOURCES}
+      "${AOM_ROOT}/aom_dsp/inv_txfm.c"
+      "${AOM_ROOT}/aom_dsp/inv_txfm.h")
+
+  set(AOM_DSP_COMMON_ASM_SSE2
+      ${AOM_DSP_COMMON_ASM_SSE2}
+      "${AOM_ROOT}/aom_dsp/x86/inv_wht_sse2.asm")
+
+  set(AOM_DSP_COMMON_INTRIN_SSE2
+      ${AOM_DSP_COMMON_INTRIN_SSE2}
+      "${AOM_ROOT}/aom_dsp/x86/inv_txfm_sse2.c"
+      "${AOM_ROOT}/aom_dsp/x86/inv_txfm_sse2.h")
+endif ()
+
+if (CONFIG_DECODERS)
+  set(AOM_DSP_DECODER_SOURCES
+      "${AOM_ROOT}/aom_dsp/binary_codes_reader.c"
+      "${AOM_ROOT}/aom_dsp/binary_codes_reader.h"
+      "${AOM_ROOT}/aom_dsp/bitreader.h"
+      "${AOM_ROOT}/aom_dsp/bitreader_buffer.c"
+      "${AOM_ROOT}/aom_dsp/bitreader_buffer.h")
+
+  if (CONFIG_ANS)
+    set(AOM_DSP_DECODER_SOURCES
+        ${AOM_DSP_DECODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/ansreader.h")
+  elseif (CONFIG_DAALA_EC)
+    set(AOM_DSP_DECODER_SOURCES
+        ${AOM_DSP_DECODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/daalaboolreader.c"
+        "${AOM_ROOT}/aom_dsp/daalaboolreader.h"
+        "${AOM_ROOT}/aom_dsp/entdec.c"
+        "${AOM_ROOT}/aom_dsp/entdec.h")
+  else ()
+    set(AOM_DSP_DECODER_SOURCES
+        ${AOM_DSP_DECODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/dkboolreader.c"
+        "${AOM_ROOT}/aom_dsp/dkboolreader.h")
+  endif ()
+endif ()
+
+if (CONFIG_ENCODERS)
+  set(AOM_DSP_ENCODER_SOURCES
+      "${AOM_ROOT}/aom_dsp/binary_codes_writer.c"
+      "${AOM_ROOT}/aom_dsp/binary_codes_writer.h"
+      "${AOM_ROOT}/aom_dsp/bitwriter.h"
+      "${AOM_ROOT}/aom_dsp/bitwriter_buffer.c"
+      "${AOM_ROOT}/aom_dsp/bitwriter_buffer.h"
+      "${AOM_ROOT}/aom_dsp/psnr.c"
+      "${AOM_ROOT}/aom_dsp/psnr.h"
+      "${AOM_ROOT}/aom_dsp/variance.c"
+      "${AOM_ROOT}/aom_dsp/variance.h")
+
+  set(AOM_DSP_ENCODER_ASM_SSE2
+      ${AOM_DSP_ENCODER_ASM_SSE2}
+      "${AOM_ROOT}/aom_dsp/x86/halfpix_variance_impl_sse2.asm"
+      "${AOM_ROOT}/aom_dsp/x86/sad4d_sse2.asm"
+      "${AOM_ROOT}/aom_dsp/x86/sad_sse2.asm"
+      "${AOM_ROOT}/aom_dsp/x86/subtract_sse2.asm"
+      "${AOM_ROOT}/aom_dsp/x86/subpel_variance_sse2.asm")
+
+  set(AOM_DSP_ENCODER_INTRIN_SSE2
+      "${AOM_ROOT}/aom_dsp/x86/quantize_sse2.c")
+
+  set(AOM_DSP_ENCODER_ASM_SSSE3
+      "${AOM_ROOT}/aom_dsp/x86/sad_ssse3.asm")
+
+  set(AOM_DSP_ENCODER_ASM_SSSE3_X86_64
+      "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm"
+      "${AOM_ROOT}/aom_dsp/x86/ssim_opt_x86_64.asm")
+
+  set(AOM_DSP_ENCODER_INTRIN_SSE3 "${AOM_ROOT}/aom_dsp/x86/sad_sse3.asm")
+  set(AOM_DSP_ENCODER_ASM_SSE4_1 "${AOM_ROOT}/aom_dsp/x86/sad_sse4.asm")
+
+  set(AOM_DSP_ENCODER_INTRIN_AVX2
+      "${AOM_ROOT}/aom_dsp/x86/sad4d_avx2.c"
+      "${AOM_ROOT}/aom_dsp/x86/sad_avx2.c"
+      "${AOM_ROOT}/aom_dsp/x86/sad_impl_avx2.c"
+      "${AOM_ROOT}/aom_dsp/x86/variance_avx2.c"
+      "${AOM_ROOT}/aom_dsp/x86/variance_impl_avx2.c")
+
+  if (CONFIG_AV1_ENCODER)
+    set(AOM_DSP_ENCODER_SOURCES
+        ${AOM_DSP_ENCODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/avg.c"
+        "${AOM_ROOT}/aom_dsp/fwd_txfm.c"
+        "${AOM_ROOT}/aom_dsp/fwd_txfm.h"
+        "${AOM_ROOT}/aom_dsp/quantize.c"
+        "${AOM_ROOT}/aom_dsp/quantize.h"
+        "${AOM_ROOT}/aom_dsp/sum_squares.c")
+
+    set(AOM_DSP_ENCODER_INTRIN_SSE2
+        ${AOM_DSP_ENCODER_INTRIN_SSE2}
+        "${AOM_ROOT}/aom_dsp/x86/avg_intrin_sse2.c"
+        "${AOM_ROOT}/aom_dsp/x86/fwd_dct32_8cols_sse2.c"
+        "${AOM_ROOT}/aom_dsp/x86/fwd_dct32x32_impl_sse2.h"
+        "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_impl_sse2.h"
+        "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_sse2.c"
+        "${AOM_ROOT}/aom_dsp/x86/fwd_txfm_sse2.h"
+        "${AOM_ROOT}/aom_dsp/x86/halfpix_variance_sse2.c"
+        "${AOM_ROOT}/aom_dsp/x86/variance_sse2.c"
+        "${AOM_ROOT}/aom_dsp/x86/sum_squares_sse2.c")
+
+    set(AOM_DSP_ENCODER_INTRIN_SSSE3
+        ${AOM_DSP_ENCODER_INTRIN_SSSE3}
+        "${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_ssse3.c"
+        "${AOM_ROOT}/aom_dsp/x86/masked_variance_intrin_ssse3.c")
+
+    set(AOM_DSP_ENCODER_ASM_SSSE3_X86_64
+        ${AOM_DSP_ENCODER_ASM_SSSE3_X86_64}
+        "${AOM_ROOT}/aom_dsp/x86/avg_ssse3_x86_64.asm"
+        "${AOM_ROOT}/aom_dsp/x86/quantize_ssse3_x86_64.asm")
+
+    set(AOM_DSP_ENCODER_AVX_ASM_X86_64
+        ${AOM_DSP_ENCODER_AVX_ASM_X86_64}
+        "${AOM_ROOT}/aom_dsp/x86/quantize_avx_x86_64.asm")
+
+    set(AOM_DSP_ENCODER_INTRIN_MSA
+        "${AOM_ROOT}/aom_dsp/mips/avg_msa.c"
+        "${AOM_ROOT}/aom_dsp/mips/sad_msa.c"
+        "${AOM_ROOT}/aom_dsp/mips/subtract_msa.c"
+        "${AOM_ROOT}/aom_dsp/mips/variance_msa.c"
+        "${AOM_ROOT}/aom_dsp/mips/sub_pixel_variance_msa.c")
+
+    if (CONFIG_HIGHBITDEPTH)
+      set(AOM_DSP_ENCODER_INTRIN_SSE2
+          ${AOM_DSP_ENCODER_INTRIN_SSE2}
+          "${AOM_ROOT}/aom_dsp/x86/highbd_quantize_intrin_sse2.c"
+          "${AOM_ROOT}/aom_dsp/x86/highbd_subtract_sse2.c")
+    endif ()
+  endif ()
+
+  if (CONFIG_HIGHBITDEPTH)
+    set(AOM_DSP_ENCODER_ASM_SSE2
+        ${AOM_DSP_ENCODER_ASM_SSE2}
+        "${AOM_ROOT}/aom_dsp/x86/highbd_sad4d_sse2.asm"
+        "${AOM_ROOT}/aom_dsp/x86/highbd_sad_sse2.asm"
+        "${AOM_ROOT}/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm"
+        "${AOM_ROOT}/aom_dsp/x86/highbd_variance_impl_sse2.asm"
+        "${AOM_ROOT}/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm"
+        "${AOM_ROOT}/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm")
+
+    set(AOM_DSP_ENCODER_INTRIN_SSE2
+        ${AOM_DSP_ENCODER_INTRIN_SSE2}
+        "${AOM_ROOT}/aom_dsp/x86/highbd_variance_sse2.c")
+
+    set(AOM_DSP_ENCODER_INTRIN_SSE4_1
+        ${AOM_DSP_ENCODER_INTRIN_SSE4_1}
+        "${AOM_ROOT}/aom_dsp/x86/highbd_variance_sse4.c")
+
+    set(AOM_DSP_ENCODER_INTRIN_AVX2
+        ${AOM_DSP_ENCODER_INTRIN_AVX2}
+        "${AOM_ROOT}/aom_dsp/x86/sad_highbd_avx2.c")
+  endif ()
+
+  if (CONFIG_ANS)
+    set(AOM_DSP_ENCODER_SOURCES
+        ${AOM_DSP_ENCODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/answriter.h"
+        "${AOM_ROOT}/aom_dsp/buf_ans.c"
+        "${AOM_ROOT}/aom_dsp/buf_ans.h")
+  elseif (CONFIG_DAALA_EC)
+    set(AOM_DSP_ENCODER_SOURCES
+        ${AOM_DSP_ENCODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/daalaboolwriter.c"
+        "${AOM_ROOT}/aom_dsp/daalaboolwriter.h"
+        "${AOM_ROOT}/aom_dsp/entenc.c"
+        "${AOM_ROOT}/aom_dsp/entenc.h")
+  else ()
+    set(AOM_DSP_ENCODER_SOURCES
+        ${AOM_DSP_ENCODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/dkboolwriter.c"
+        "${AOM_ROOT}/aom_dsp/dkboolwriter.h")
+  endif ()
+
+  if (CONFIG_INTERNAL_STATS)
+    set(AOM_DSP_ENCODER_SOURCES
+        ${AOM_DSP_ENCODER_SOURCES}
+        "${AOM_ROOT}/aom_dsp/fastssim.c"
+        "${AOM_ROOT}/aom_dsp/psnrhvs.c"
+        "${AOM_ROOT}/aom_dsp/ssim.c"
+        "${AOM_ROOT}/aom_dsp/ssim.h")
+  endif ()
+endif ()
+
+if (CONFIG_MOTION_VAR)
+  set(AOM_DSP_ENCODER_INTRIN_SSE4_1
+      ${AOM_DSP_ENCODER_INTRIN_SSE4_1}
+      "${AOM_ROOT}/aom_dsp/x86/obmc_sad_sse4.c"
+      "${AOM_ROOT}/aom_dsp/x86/obmc_variance_sse4.c")
+endif ()
+
+# Creates aom_dsp build targets. Must not be called until after libaom target
+# has been created.
+function (setup_aom_dsp_targets)
+  add_library(aom_dsp_common OBJECT ${AOM_DSP_COMMON_SOURCES})
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_dsp_common)
+  target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_dsp_common>)
+
+  if (CONFIG_DECODERS)
+    add_library(aom_dsp_decoder OBJECT ${AOM_DSP_DECODER_SOURCES})
+    set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_dsp_decoder)
+    target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_dsp_decoder>)
+  endif ()
+
+  if (CONFIG_ENCODERS)
+    add_library(aom_dsp_encoder OBJECT ${AOM_DSP_ENCODER_SOURCES})
+    set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_dsp_encoder)
+    target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_dsp_encoder>)
+  endif ()
+
+  if (HAVE_SSE2)
+    add_asm_library("aom_dsp_common_sse2" "AOM_DSP_COMMON_ASM_SSE2" "aom")
+    add_intrinsics_object_library("-msse2" "sse2" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_SSE2")
+    if (CONFIG_ENCODERS)
+      add_asm_library("aom_dsp_encoder_sse2" "AOM_DSP_ENCODER_ASM_SSE2" "aom")
+      add_intrinsics_object_library("-msse2" "sse2" "aom_dsp_encoder"
+                                    "AOM_DSP_ENCODER_INTRIN_SSE2")
+    endif()
+  endif ()
+
+  if (HAVE_SSE3 AND CONFIG_ENCODERS)
+    add_asm_library("aom_dsp_encoder_sse3" "AOM_DSP_ENCODER_INTRIN_SSE3" "aom")
+  endif ()
+
+  if (HAVE_SSSE3)
+    add_asm_library("aom_dsp_common_ssse3" "AOM_DSP_COMMON_ASM_SSSE3" "aom")
+    add_intrinsics_object_library("-mssse3" "ssse3" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_SSSE3")
+
+    if (CONFIG_ENCODERS)
+      if ("${AOM_TARGET_CPU}" STREQUAL "x86_64")
+        list(APPEND AOM_DSP_ENCODER_ASM_SSSE3
+             ${AOM_DSP_ENCODER_ASM_SSSE3_X86_64})
+      endif ()
+      add_asm_library("aom_dsp_encoder_ssse3" "AOM_DSP_ENCODER_ASM_SSSE3" "aom")
+      add_intrinsics_object_library("-mssse3" "ssse3" "aom_dsp_encoder"
+                                    "AOM_DSP_ENCODER_INTRIN_SSSE3")
+    endif ()
+  endif ()
+
+  if (HAVE_SSE4_1)
+    add_intrinsics_object_library("-msse4.1" "sse4_1" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_SSE4_1")
+    if (CONFIG_ENCODERS)
+      if (AOM_DSP_ENCODER_INTRIN_SSE4_1)
+        add_intrinsics_object_library("-msse4.1" "sse4_1" "aom_dsp_encoder"
+                                      "AOM_DSP_ENCODER_INTRIN_SSE4_1")
+      endif ()
+      add_asm_library("aom_dsp_encoder_sse4_1" "AOM_DSP_ENCODER_ASM_SSE4_1"
+                      "aom")
+    endif ()
+  endif ()
+
+  if (HAVE_AVX AND "${AOM_TARGET_CPU}" STREQUAL "x86_64")
+    add_asm_library("aom_dsp_encoder_avx" "AOM_DSP_ENCODER_AVX_ASM_X86_64"
+                    "aom")
+  endif ()
+
+  if (HAVE_AVX2)
+    add_intrinsics_object_library("-mavx2" "avx2" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_AVX2")
+    if (CONFIG_ENCODERS)
+      add_intrinsics_object_library("-mavx2" "avx2" "aom_dsp_encoder"
+                                    "AOM_DSP_ENCODER_INTRIN_AVX2")
+    endif ()
+  endif ()
+
+  if (HAVE_NEON_ASM)
+    if (AOM_ADS2GAS_REQUIRED)
+      add_gas_asm_library("aom_dsp_common_neon" "AOM_DSP_COMMON_ASM_NEON" "aom")
+    else ()
+      add_asm_library("aom_dsp_common_neon" "AOM_DSP_COMMON_ASM_NEON" "aom")
+    endif ()
+  endif ()
+
+  if (HAVE_NEON)
+    add_intrinsics_object_library("${AOM_NEON_INTRIN_FLAG}" "neon"
+                                  "aom_dsp_common" "AOM_DSP_COMMON_INTRIN_NEON")
+  endif ()
+
+  if (HAVE_DSPR2)
+    add_intrinsics_object_library("" "dspr2" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_DSPR2")
+  endif ()
+
+  if (HAVE_MSA)
+    add_intrinsics_object_library("" "msa" "aom_dsp_common"
+                                  "AOM_DSP_COMMON_INTRIN_MSA")
+    if (CONFIG_ENCODERS)
+      add_intrinsics_object_library("" "msa" "aom_dsp_encoder"
+                                    "AOM_DSP_ENCODER_INTRIN_MSA")
+    endif ()
+  endif ()
+
+  # Pass the new lib targets up to the parent scope instance of
+  # $AOM_LIB_TARGETS.
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} PARENT_SCOPE)
+endfunction ()
diff --git a/third_party/aom/aom_dsp/aom_dsp.mk b/third_party/aom/aom_dsp/aom_dsp.mk
new file mode 100644
index 0000000000..8c7241b831
--- /dev/null
+++ b/third_party/aom/aom_dsp/aom_dsp.mk
@@ -0,0 +1,428 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+DSP_SRCS-yes += aom_dsp.mk
+DSP_SRCS-yes += aom_dsp_common.h
+
+DSP_SRCS-$(HAVE_MSA)    += mips/macros_msa.h
+
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64)   += x86/synonyms.h
+
+# bit reader
+DSP_SRCS-yes += prob.h
+DSP_SRCS-yes += prob.c
+DSP_SRCS-$(CONFIG_ANS) += ans.h
+
+ifeq ($(CONFIG_ENCODERS),yes)
+ifeq ($(CONFIG_ANS),yes)
+DSP_SRCS-yes += answriter.h
+DSP_SRCS-yes += buf_ans.h
+DSP_SRCS-yes += buf_ans.c
+else ifeq ($(CONFIG_DAALA_EC),yes)
+DSP_SRCS-yes += entenc.c
+DSP_SRCS-yes += entenc.h
+DSP_SRCS-yes += daalaboolwriter.c
+DSP_SRCS-yes += daalaboolwriter.h
+else
+DSP_SRCS-yes += dkboolwriter.h
+DSP_SRCS-yes += dkboolwriter.c
+endif
+DSP_SRCS-yes += bitwriter.h
+DSP_SRCS-yes += bitwriter_buffer.c
+DSP_SRCS-yes += bitwriter_buffer.h
+DSP_SRCS-yes += binary_codes_writer.c
+DSP_SRCS-yes += binary_codes_writer.h
+DSP_SRCS-yes += psnr.c
+DSP_SRCS-yes += psnr.h
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += ssim.c
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += ssim.h
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += psnrhvs.c
+DSP_SRCS-$(CONFIG_INTERNAL_STATS) += fastssim.c
+endif
+
+ifeq ($(CONFIG_DECODERS),yes)
+ifeq ($(CONFIG_ANS),yes)
+DSP_SRCS-yes += ansreader.h
+else ifeq ($(CONFIG_DAALA_EC),yes)
+DSP_SRCS-yes += entdec.c
+DSP_SRCS-yes += entdec.h
+DSP_SRCS-yes += daalaboolreader.c
+DSP_SRCS-yes += daalaboolreader.h
+else
+DSP_SRCS-yes += dkboolreader.h
+DSP_SRCS-yes += dkboolreader.c
+endif
+DSP_SRCS-yes += bitreader.h
+DSP_SRCS-yes += bitreader_buffer.c
+DSP_SRCS-yes += bitreader_buffer.h
+DSP_SRCS-yes += binary_codes_reader.c
+DSP_SRCS-yes += binary_codes_reader.h
+endif
+
+# intra predictions
+DSP_SRCS-yes += intrapred.c
+
+ifeq ($(CONFIG_DAALA_EC),yes)
+DSP_SRCS-yes += entcode.c
+DSP_SRCS-yes += entcode.h
+endif
+
+DSP_SRCS-$(HAVE_SSE) += x86/intrapred_sse2.asm
+DSP_SRCS-$(HAVE_SSE2) += x86/intrapred_sse2.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/intrapred_ssse3.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/aom_subpixel_8t_ssse3.asm
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE)  += x86/highbd_intrapred_sse2.asm
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_intrapred_sse2.asm
+endif  # CONFIG_HIGHBITDEPTH
+
+DSP_SRCS-$(HAVE_NEON_ASM) += arm/intrapred_neon_asm$(ASM)
+DSP_SRCS-$(HAVE_NEON) += arm/intrapred_neon.c
+DSP_SRCS-$(HAVE_MSA) += mips/intrapred_msa.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/intrapred4_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/intrapred8_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/intrapred16_dspr2.c
+
+DSP_SRCS-$(HAVE_DSPR2)  += mips/common_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/common_dspr2.c
+
+# inter predictions
+DSP_SRCS-yes            += blend.h
+DSP_SRCS-yes            += blend_a64_mask.c
+DSP_SRCS-yes            += blend_a64_hmask.c
+DSP_SRCS-yes            += blend_a64_vmask.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_sse4.h
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_a64_mask_sse4.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_a64_hmask_sse4.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/blend_a64_vmask_sse4.c
+
+# interpolation filters
+DSP_SRCS-yes += aom_convolve.c
+DSP_SRCS-yes += aom_convolve.h
+DSP_SRCS-yes += aom_filter.h
+
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/convolve.h
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/aom_asm_stubs.c
+DSP_SRCS-$(HAVE_SSE2)  += x86/aom_subpixel_8t_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)  += x86/aom_subpixel_bilinear_sse2.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/aom_subpixel_8t_ssse3.asm
+DSP_SRCS-$(HAVE_SSSE3) += x86/aom_subpixel_bilinear_ssse3.asm
+DSP_SRCS-$(HAVE_AVX2)  += x86/aom_subpixel_8t_intrin_avx2.c
+DSP_SRCS-$(HAVE_SSSE3) += x86/aom_subpixel_8t_intrin_ssse3.c
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)  += x86/aom_high_subpixel_8t_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)  += x86/aom_high_subpixel_bilinear_sse2.asm
+DSP_SRCS-$(HAVE_AVX2)  += x86/highbd_convolve_avx2.c
+endif
+DSP_SRCS-$(HAVE_SSE2)  += x86/aom_convolve_copy_sse2.asm
+
+ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes += arm/aom_convolve_copy_neon_asm$(ASM)
+DSP_SRCS-yes += arm/aom_convolve8_avg_neon_asm$(ASM)
+DSP_SRCS-yes += arm/aom_convolve8_neon_asm$(ASM)
+DSP_SRCS-yes += arm/aom_convolve_avg_neon_asm$(ASM)
+DSP_SRCS-yes += arm/aom_convolve_neon.c
+else
+ifeq ($(HAVE_NEON),yes)
+DSP_SRCS-yes += arm/aom_convolve_copy_neon.c
+DSP_SRCS-yes += arm/aom_convolve8_avg_neon.c
+DSP_SRCS-yes += arm/aom_convolve8_neon.c
+DSP_SRCS-yes += arm/aom_convolve_avg_neon.c
+DSP_SRCS-yes += arm/aom_convolve_neon.c
+endif  # HAVE_NEON
+endif  # HAVE_NEON_ASM
+
+# common (msa)
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_avg_horiz_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_avg_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_avg_vert_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_horiz_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve8_vert_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve_avg_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve_copy_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/aom_convolve_msa.h
+
+# common (dspr2)
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve_common_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_avg_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_avg_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve2_vert_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_avg_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_avg_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/convolve8_vert_dspr2.c
+
+# loop filters
+DSP_SRCS-yes += loopfilter.c
+
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64)   += x86/loopfilter_sse2.c
+DSP_SRCS-$(HAVE_AVX2)                += x86/loopfilter_avx2.c
+
+DSP_SRCS-$(HAVE_NEON)   += arm/loopfilter_neon.c
+ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes  += arm/loopfilter_mb_neon$(ASM)
+DSP_SRCS-yes  += arm/loopfilter_16_neon$(ASM)
+DSP_SRCS-yes  += arm/loopfilter_8_neon$(ASM)
+DSP_SRCS-yes  += arm/loopfilter_4_neon$(ASM)
+else
+ifeq ($(HAVE_NEON),yes)
+DSP_SRCS-yes   += arm/loopfilter_16_neon.c
+DSP_SRCS-yes   += arm/loopfilter_8_neon.c
+DSP_SRCS-yes   += arm/loopfilter_4_neon.c
+endif  # HAVE_NEON
+endif  # HAVE_NEON_ASM
+
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_msa.h
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_16_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_8_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/loopfilter_4_msa.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_filters_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_filters_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_macros_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_masks_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_mb_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_mb_horiz_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2)  += mips/loopfilter_mb_vert_dspr2.c
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_loopfilter_sse2.c
+endif  # CONFIG_HIGHBITDEPTH
+
+DSP_SRCS-yes            += txfm_common.h
+DSP_SRCS-yes            += x86/txfm_common_intrin.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/txfm_common_sse2.h
+DSP_SRCS-$(HAVE_MSA)    += mips/txfm_macros_msa.h
+
+# forward transform
+ifneq ($(findstring yes,$(CONFIG_AV1)$(CONFIG_PVQ)),)
+DSP_SRCS-$(HAVE_AVX2)   += x86/txfm_common_avx2.h
+ifeq ($(CONFIG_AV1_ENCODER),yes)
+DSP_SRCS-yes            += fwd_txfm.c
+DSP_SRCS-yes            += fwd_txfm.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_txfm_sse2.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_txfm_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_dct32_8cols_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_txfm_impl_sse2.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/fwd_dct32x32_impl_sse2.h
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_SSSE3)  += x86/fwd_txfm_ssse3_x86_64.asm
+endif
+DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_txfm_avx2.h
+DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_txfm_avx2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/fwd_dct32x32_impl_avx2.h
+DSP_SRCS-$(HAVE_NEON)   += arm/fwd_txfm_neon.c
+DSP_SRCS-$(HAVE_MSA)    += mips/fwd_txfm_msa.h
+DSP_SRCS-$(HAVE_MSA)    += mips/fwd_txfm_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/fwd_dct32x32_msa.c
+endif  # CONFIG_AV1_ENCODER
+endif  # CONFIG_AV1
+
+# inverse transform
+ifeq ($(CONFIG_AV1), yes)
+DSP_SRCS-yes            += inv_txfm.h
+DSP_SRCS-yes            += inv_txfm.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/inv_txfm_sse2.h
+DSP_SRCS-$(HAVE_SSE2)   += x86/inv_txfm_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/inv_wht_sse2.asm
+DSP_SRCS-$(HAVE_SSSE3)  += x86/inv_txfm_ssse3.c
+
+ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes  += arm/save_reg_neon$(ASM)
+DSP_SRCS-yes  += arm/idct4x4_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct4x4_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct8x8_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct8x8_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct16x16_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct16x16_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct32x32_1_add_neon$(ASM)
+DSP_SRCS-yes  += arm/idct32x32_add_neon$(ASM)
+else
+ifeq ($(HAVE_NEON),yes)
+DSP_SRCS-yes  += arm/idct4x4_1_add_neon.c
+DSP_SRCS-yes  += arm/idct4x4_add_neon.c
+DSP_SRCS-yes  += arm/idct8x8_1_add_neon.c
+DSP_SRCS-yes  += arm/idct8x8_add_neon.c
+DSP_SRCS-yes  += arm/idct16x16_1_add_neon.c
+DSP_SRCS-yes  += arm/idct16x16_add_neon.c
+DSP_SRCS-yes  += arm/idct32x32_1_add_neon.c
+DSP_SRCS-yes  += arm/idct32x32_add_neon.c
+endif  # HAVE_NEON
+endif  # HAVE_NEON_ASM
+DSP_SRCS-$(HAVE_NEON)  += arm/idct16x16_neon.c
+
+DSP_SRCS-$(HAVE_MSA)   += mips/inv_txfm_msa.h
+DSP_SRCS-$(HAVE_MSA)   += mips/idct4x4_msa.c
+DSP_SRCS-$(HAVE_MSA)   += mips/idct8x8_msa.c
+DSP_SRCS-$(HAVE_MSA)   += mips/idct16x16_msa.c
+DSP_SRCS-$(HAVE_MSA)   += mips/idct32x32_msa.c
+
+ifneq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_DSPR2) += mips/inv_txfm_dspr2.h
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans4_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans8_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans16_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans32_dspr2.c
+DSP_SRCS-$(HAVE_DSPR2) += mips/itrans32_cols_dspr2.c
+endif  # CONFIG_HIGHBITDEPTH
+endif  # CONFIG_AV1
+
+# quantization
+ifneq ($(filter yes,$(CONFIG_AV1_ENCODER)),)
+DSP_SRCS-yes            += quantize.c
+DSP_SRCS-yes            += quantize.h
+
+DSP_SRCS-$(HAVE_SSE2)   += x86/quantize_sse2.c
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_quantize_intrin_sse2.c
+endif
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_SSSE3)  += x86/quantize_ssse3_x86_64.asm
+DSP_SRCS-$(HAVE_AVX)    += x86/quantize_avx_x86_64.asm
+endif
+
+# avg
+DSP_SRCS-yes           += avg.c
+DSP_SRCS-$(HAVE_SSE2)  += x86/avg_intrin_sse2.c
+DSP_SRCS-$(HAVE_NEON)  += arm/avg_neon.c
+DSP_SRCS-$(HAVE_MSA)   += mips/avg_msa.c
+DSP_SRCS-$(HAVE_NEON)  += arm/hadamard_neon.c
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_SSSE3) += x86/avg_ssse3_x86_64.asm
+endif
+
+# high bit depth subtract
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)  += x86/highbd_subtract_sse2.c
+endif
+
+endif  # CONFIG_AV1_ENCODER
+
+ifeq ($(CONFIG_AV1_ENCODER),yes)
+DSP_SRCS-yes            += sum_squares.c
+
+DSP_SRCS-$(HAVE_SSE2)   += x86/sum_squares_sse2.c
+endif # CONFIG_AV1_ENCODER
+
+ifeq ($(CONFIG_ENCODERS),yes)
+DSP_SRCS-yes            += sad.c
+DSP_SRCS-yes            += subtract.c
+
+DSP_SRCS-$(HAVE_MEDIA)  += arm/sad_media$(ASM)
+DSP_SRCS-$(HAVE_NEON)   += arm/sad4d_neon.c
+DSP_SRCS-$(HAVE_NEON)   += arm/sad_neon.c
+DSP_SRCS-$(HAVE_NEON)   += arm/subtract_neon.c
+
+DSP_SRCS-$(HAVE_MSA)    += mips/sad_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/subtract_msa.c
+
+DSP_SRCS-$(HAVE_SSE3)   += x86/sad_sse3.asm
+DSP_SRCS-$(HAVE_SSSE3)  += x86/sad_ssse3.asm
+DSP_SRCS-$(HAVE_SSE4_1) += x86/sad_sse4.asm
+DSP_SRCS-$(HAVE_AVX2)   += x86/sad4d_avx2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/sad_avx2.c
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_AVX2)   += x86/sad_highbd_avx2.c
+endif
+
+ifeq ($(CONFIG_AV1_ENCODER),yes)
+ifeq ($(CONFIG_EXT_INTER),yes)
+DSP_SRCS-$(HAVE_SSSE3)  += x86/masked_sad_intrin_ssse3.c
+DSP_SRCS-$(HAVE_SSSE3)  += x86/masked_variance_intrin_ssse3.c
+endif  #CONFIG_EXT_INTER
+ifeq ($(CONFIG_MOTION_VAR),yes)
+DSP_SRCS-$(HAVE_SSE4_1) += x86/obmc_sad_sse4.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/obmc_variance_sse4.c
+endif  #CONFIG_MOTION_VAR
+ifeq ($(CONFIG_EXT_PARTITION),yes)
+DSP_SRCS-$(HAVE_AVX2) += x86/sad_impl_avx2.c
+endif
+endif  #CONFIG_AV1_ENCODER
+
+DSP_SRCS-$(HAVE_SSE)    += x86/sad4d_sse2.asm
+DSP_SRCS-$(HAVE_SSE)    += x86/sad_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/sad4d_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/sad_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/subtract_sse2.asm
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad4d_sse2.asm
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_sad_sse2.asm
+endif  # CONFIG_HIGHBITDEPTH
+
+endif  # CONFIG_ENCODERS
+
+ifneq ($(filter yes,$(CONFIG_ENCODERS)),)
+DSP_SRCS-yes            += variance.c
+DSP_SRCS-yes            += variance.h
+
+DSP_SRCS-$(HAVE_MEDIA)  += arm/bilinear_filter_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/subpel_variance_media.c
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_halfpixvar16x16_h_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_halfpixvar16x16_hv_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_halfpixvar16x16_v_media$(ASM)
+DSP_SRCS-$(HAVE_MEDIA)  += arm/variance_media$(ASM)
+DSP_SRCS-$(HAVE_NEON)   += arm/subpel_variance_neon.c
+DSP_SRCS-$(HAVE_NEON)   += arm/variance_neon.c
+
+DSP_SRCS-$(HAVE_MSA)    += mips/variance_msa.c
+DSP_SRCS-$(HAVE_MSA)    += mips/sub_pixel_variance_msa.c
+
+DSP_SRCS-$(HAVE_SSE)    += x86/variance_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/variance_sse2.c  # Contains SSE2 and SSSE3
+DSP_SRCS-$(HAVE_SSE2)   += x86/halfpix_variance_sse2.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/halfpix_variance_impl_sse2.asm
+DSP_SRCS-$(HAVE_AVX2)   += x86/variance_avx2.c
+DSP_SRCS-$(HAVE_AVX2)   += x86/variance_impl_avx2.c
+
+ifeq ($(ARCH_X86_64),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/ssim_opt_x86_64.asm
+endif  # ARCH_X86_64
+
+DSP_SRCS-$(HAVE_SSE)    += x86/subpel_variance_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/subpel_variance_sse2.asm  # Contains SSE2 and SSSE3
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_variance_sse2.c
+DSP_SRCS-$(HAVE_SSE4_1) += x86/highbd_variance_sse4.c
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_variance_impl_sse2.asm
+DSP_SRCS-$(HAVE_SSE2)   += x86/highbd_subpel_variance_impl_sse2.asm
+endif  # CONFIG_HIGHBITDEPTH
+endif  # CONFIG_ENCODERS
+
+DSP_SRCS-no += $(DSP_SRCS_REMOVE-yes)
+
+DSP_SRCS-yes += aom_dsp_rtcd.c
+DSP_SRCS-yes += aom_dsp_rtcd_defs.pl
+
+DSP_SRCS-yes += aom_simd.h
+DSP_SRCS-yes += aom_simd_inline.h
+DSP_SRCS-yes += simd/v64_intrinsics.h
+DSP_SRCS-yes += simd/v64_intrinsics_c.h
+DSP_SRCS-yes += simd/v128_intrinsics.h
+DSP_SRCS-yes += simd/v128_intrinsics_c.h
+DSP_SRCS-yes += simd/v256_intrinsics.h
+DSP_SRCS-yes += simd/v256_intrinsics_c.h
+DSP_SRCS-yes += simd/v256_intrinsics_v128.h
+DSP_SRCS-$(HAVE_SSE2) += simd/v64_intrinsics_x86.h
+DSP_SRCS-$(HAVE_SSE2) += simd/v128_intrinsics_x86.h
+DSP_SRCS-$(HAVE_SSE2) += simd/v256_intrinsics_x86.h
+DSP_SRCS-$(HAVE_NEON) += simd/v64_intrinsics_arm.h
+DSP_SRCS-$(HAVE_NEON) += simd/v128_intrinsics_arm.h
+DSP_SRCS-$(HAVE_NEON) += simd/v256_intrinsics_arm.h
+
+$(eval $(call rtcd_h_template,aom_dsp_rtcd,aom_dsp/aom_dsp_rtcd_defs.pl))
diff --git a/third_party/aom/aom_dsp/aom_dsp_common.h b/third_party/aom/aom_dsp/aom_dsp_common.h
new file mode 100644
index 0000000000..47ffbeb6cc
--- /dev/null
+++ b/third_party/aom/aom_dsp/aom_dsp_common.h
@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_AOM_DSP_COMMON_H_
+#define AOM_DSP_AOM_DSP_COMMON_H_
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef MAX_SB_SIZE
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+#define MAX_SB_SIZE 128
+#else
+#define MAX_SB_SIZE 64
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+#endif  // ndef MAX_SB_SIZE
+
+#define AOMMIN(x, y) (((x) < (y)) ? (x) : (y))
+#define AOMMAX(x, y) (((x) > (y)) ? (x) : (y))
+
+#define IMPLIES(a, b) (!(a) || (b))  //  Logical 'a implies b' (or 'a -> b')
+
+#define IS_POWER_OF_TWO(x) (((x) & ((x)-1)) == 0)
+
+/* Left shifting a negative value became undefined behavior in C99 (downgraded
+   from merely implementation-defined in C89). This should still compile to the
+   correct thing on any two's-complement machine, but avoid ubsan warnings.*/
+#define AOM_SIGNED_SHL(x, shift) ((x) * (((x)*0 + 1) << (shift)))
+
+// These can be used to give a hint about branch outcomes.
+// This can have an effect, even if your target processor has a
+// good branch predictor, as these hints can affect basic block
+// ordering by the compiler.
+#ifdef __GNUC__
+#define LIKELY(v) __builtin_expect(v, 1)
+#define UNLIKELY(v) __builtin_expect(v, 0)
+#else
+#define LIKELY(v) (v)
+#define UNLIKELY(v) (v)
+#endif
+
+#define AOM_SWAP(type, a, b) \
+  do {                       \
+    type c = (b);            \
+    b = a;                   \
+    a = c;                   \
+  } while (0)
+
+#if CONFIG_AOM_QM
+typedef uint16_t qm_val_t;
+#define AOM_QM_BITS 6
+#endif
+#if CONFIG_HIGHBITDEPTH
+// Note:
+// tran_low_t  is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+typedef int64_t tran_high_t;
+typedef int32_t tran_low_t;
+#else
+// Note:
+// tran_low_t  is the datatype used for final transform coefficients.
+// tran_high_t is the datatype used for intermediate transform stages.
+typedef int32_t tran_high_t;
+typedef int16_t tran_low_t;
+#endif  // CONFIG_HIGHBITDEPTH
+
+static INLINE uint8_t clip_pixel(int val) {
+  return (val > 255) ? 255 : (val < 0) ? 0 : val;
+}
+
+static INLINE int clamp(int value, int low, int high) {
+  return value < low ? low : (value > high ? high : value);
+}
+
+static INLINE double fclamp(double value, double low, double high) {
+  return value < low ? low : (value > high ? high : value);
+}
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE uint16_t clip_pixel_highbd(int val, int bd) {
+  switch (bd) {
+    case 8:
+    default: return (uint16_t)clamp(val, 0, 255);
+    case 10: return (uint16_t)clamp(val, 0, 1023);
+    case 12: return (uint16_t)clamp(val, 0, 4095);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_AOM_DSP_COMMON_H_
diff --git a/third_party/aom/aom_dsp/aom_dsp_rtcd.c b/third_party/aom/aom_dsp/aom_dsp_rtcd.c
new file mode 100644
index 0000000000..11a57d3822
--- /dev/null
+++ b/third_party/aom/aom_dsp/aom_dsp_rtcd.c
@@ -0,0 +1,16 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "./aom_config.h"
+#define RTCD_C
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/aom_once.h"
+
+void aom_dsp_rtcd() { once(setup_rtcd_internal); }
diff --git a/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl b/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl
new file mode 100755
index 0000000000..b4ef0d92f3
--- /dev/null
+++ b/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -0,0 +1,1495 @@
+sub aom_dsp_forward_decls() {
+print <<EOF
+/*
+ * DSP
+ */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/enums.h"
+
+EOF
+}
+forward_decls qw/aom_dsp_forward_decls/;
+
+# optimizations which depend on multiple features
+$avx2_ssse3 = '';
+if ((aom_config("HAVE_AVX2") eq "yes") && (aom_config("HAVE_SSSE3") eq "yes")) {
+  $avx2_ssse3 = 'avx2';
+}
+
+# functions that are 64 bit only.
+$mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = '';
+if ($opts{arch} eq "x86_64") {
+  $mmx_x86_64 = 'mmx';
+  $sse2_x86_64 = 'sse2';
+  $ssse3_x86_64 = 'ssse3';
+  $avx_x86_64 = 'avx';
+  $avx2_x86_64 = 'avx2';
+}
+
+if (aom_config("CONFIG_EXT_PARTITION") eq "yes") {
+  @block_widths = (4, 8, 16, 32, 64, 128)
+} else {
+  @block_widths = (4, 8, 16, 32, 64)
+}
+
+@block_sizes = ();
+foreach $w (@block_widths) {
+  foreach $h (@block_widths) {
+    push @block_sizes, [$w, $h] if ($w <= 2*$h && $h <= 2*$w) ;
+  }
+}
+
+@tx_dims = (2, 4, 8, 16, 32);
+if (aom_config("CONFIG_TX64X64") eq "yes") {
+  push @tx_dims, '64';
+}
+
+@pred_names = qw/dc dc_top dc_left dc_128 v h d207e d63e d45e d117 d135 d153/;
+if (aom_config("CONFIG_ALT_INTRA") eq "yes") {
+  push @pred_names, qw/paeth smooth/;
+} else {
+  push @pred_names, 'tm';
+}
+
+#
+# Intra prediction
+#
+
+foreach $dim (@tx_dims) {
+  $w = ${dim};
+  $h = ${dim};
+  foreach $pred_name (@pred_names) {
+    add_proto "void", "aom_${pred_name}_predictor_${w}x${h}",
+              "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
+    if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+      add_proto "void", "aom_highbd_${pred_name}_predictor_${w}x${h}",
+                "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd";
+    }
+  }
+}
+
+specialize qw/aom_d63e_predictor_4x4 ssse3/;
+specialize qw/aom_h_predictor_4x4 neon dspr2 msa sse2/;
+specialize qw/aom_d135_predictor_4x4 neon/;
+specialize qw/aom_d153_predictor_4x4 ssse3/;
+specialize qw/aom_v_predictor_4x4 neon msa sse2/;
+if (aom_config("CONFIG_ALT_INTRA") eq "") {
+  specialize qw/aom_tm_predictor_4x4 neon dspr2 msa sse2/;
+}  # CONFIG_ALT_INTRA
+specialize qw/aom_dc_predictor_4x4 dspr2 msa neon sse2/;
+specialize qw/aom_dc_top_predictor_4x4 msa neon sse2/;
+specialize qw/aom_dc_left_predictor_4x4 msa neon sse2/;
+specialize qw/aom_dc_128_predictor_4x4 msa neon sse2/;
+specialize qw/aom_h_predictor_8x8 neon dspr2 msa sse2/;
+specialize qw/aom_d153_predictor_8x8 ssse3/;
+specialize qw/aom_v_predictor_8x8 neon msa sse2/;
+if (aom_config("CONFIG_ALT_INTRA") eq "") {
+  specialize qw/aom_tm_predictor_8x8 neon dspr2 msa sse2/;
+}  # CONFIG_ALT_INTRA
+specialize qw/aom_dc_predictor_8x8 dspr2 neon msa sse2/;
+specialize qw/aom_dc_top_predictor_8x8 neon msa sse2/;
+specialize qw/aom_dc_left_predictor_8x8 neon msa sse2/;
+specialize qw/aom_dc_128_predictor_8x8 neon msa sse2/;
+specialize qw/aom_h_predictor_16x16 neon dspr2 msa sse2/;
+specialize qw/aom_d153_predictor_16x16 ssse3/;
+specialize qw/aom_v_predictor_16x16 neon msa sse2/;
+if (aom_config("CONFIG_ALT_INTRA") eq "") {
+  specialize qw/aom_tm_predictor_16x16 neon msa sse2/;
+}  # CONFIG_ALT_INTRA
+specialize qw/aom_dc_predictor_16x16 dspr2 neon msa sse2/;
+specialize qw/aom_dc_top_predictor_16x16 neon msa sse2/;
+specialize qw/aom_dc_left_predictor_16x16 neon msa sse2/;
+specialize qw/aom_dc_128_predictor_16x16 neon msa sse2/;
+specialize qw/aom_h_predictor_32x32 neon msa sse2/;
+specialize qw/aom_d153_predictor_32x32 ssse3/;
+specialize qw/aom_v_predictor_32x32 neon msa sse2/;
+if (aom_config("CONFIG_ALT_INTRA") eq "") {
+  specialize qw/aom_tm_predictor_32x32 neon msa sse2/;
+}  # CONFIG_ALT_INTRA
+specialize qw/aom_dc_predictor_32x32 msa neon sse2/;
+specialize qw/aom_dc_top_predictor_32x32 msa neon sse2/;
+specialize qw/aom_dc_left_predictor_32x32 msa neon sse2/;
+specialize qw/aom_dc_128_predictor_32x32 msa neon sse2/;
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  specialize qw/aom_highbd_v_predictor_4x4 sse2/;
+  if (aom_config("CONFIG_ALT_INTRA") eq "") {
+    specialize qw/aom_highbd_tm_predictor_4x4 sse2/;
+  }  # CONFIG_ALT_INTRA
+  specialize qw/aom_highbd_dc_predictor_4x4 sse2/;
+  specialize qw/aom_highbd_v_predictor_8x8 sse2/;
+  if (aom_config("CONFIG_ALT_INTRA") eq "") {
+    specialize qw/aom_highbd_tm_predictor_8x8 sse2/;
+  }  # CONFIG_ALT_INTRA
+  specialize qw/aom_highbd_dc_predictor_8x8 sse2/;;
+  specialize qw/aom_highbd_v_predictor_16x16 sse2/;
+  if (aom_config("CONFIG_ALT_INTRA") eq "") {
+    specialize qw/aom_highbd_tm_predictor_16x16 sse2/;
+  }  # CONFIG_ALT_INTRA
+  specialize qw/aom_highbd_dc_predictor_16x16 sse2/;
+  specialize qw/aom_highbd_v_predictor_32x32 sse2/;
+  if (aom_config("CONFIG_ALT_INTRA") eq "") {
+    specialize qw/aom_highbd_tm_predictor_32x32 sse2/;
+  }  # CONFIG_ALT_INTRA
+  specialize qw/aom_highbd_dc_predictor_32x32 sse2/;
+}  # CONFIG_HIGHBITDEPTH
+
+#
+# Sub Pixel Filters
+#
+add_proto qw/void aom_convolve_copy/,       "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_convolve_avg/,        "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_convolve8/,           "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_convolve8_horiz/,     "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_convolve8_vert/,      "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_convolve8_avg/,       "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_convolve8_avg_vert/,  "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_scaled_2d/,           "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_scaled_horiz/,        "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_scaled_vert/,         "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_scaled_avg_2d/,       "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_scaled_avg_horiz/,    "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+add_proto qw/void aom_scaled_avg_vert/,     "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+
+specialize qw/aom_convolve_copy       sse2      /;
+specialize qw/aom_convolve_avg        sse2      /;
+specialize qw/aom_convolve8           sse2 ssse3/, "$avx2_ssse3";
+specialize qw/aom_convolve8_horiz     sse2 ssse3/, "$avx2_ssse3";
+specialize qw/aom_convolve8_vert      sse2 ssse3/, "$avx2_ssse3";
+specialize qw/aom_convolve8_avg       sse2 ssse3/;
+specialize qw/aom_convolve8_avg_horiz sse2 ssse3/;
+specialize qw/aom_convolve8_avg_vert  sse2 ssse3/;
+specialize qw/aom_scaled_2d                ssse3/;
+
+if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
+  add_proto qw/void aom_convolve8_add_src/,       "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+  add_proto qw/void aom_convolve8_add_src_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+  add_proto qw/void aom_convolve8_add_src_vert/,  "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
+
+  specialize qw/aom_convolve8_add_src ssse3/;
+  specialize qw/aom_convolve8_add_src_horiz ssse3/;
+  specialize qw/aom_convolve8_add_src_vert ssse3/;
+}  # CONFIG_LOOP_RESTORATION
+
+# TODO(any): These need to be extended to up to 128x128 block sizes
+if (!(aom_config("CONFIG_AV1") eq "yes" && aom_config("CONFIG_EXT_PARTITION") eq "yes")) {
+  specialize qw/aom_convolve_copy       neon dspr2 msa/;
+  specialize qw/aom_convolve_avg        neon dspr2 msa/;
+  specialize qw/aom_convolve8           neon dspr2 msa/;
+  specialize qw/aom_convolve8_horiz     neon dspr2 msa/;
+  specialize qw/aom_convolve8_vert      neon dspr2 msa/;
+  specialize qw/aom_convolve8_avg       neon dspr2 msa/;
+  specialize qw/aom_convolve8_avg_horiz neon dspr2 msa/;
+  specialize qw/aom_convolve8_avg_vert  neon dspr2 msa/;
+}
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void aom_highbd_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/aom_highbd_convolve_copy sse2 avx2/;
+
+  add_proto qw/void aom_highbd_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/aom_highbd_convolve_avg sse2 avx2/;
+
+  add_proto qw/void aom_highbd_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/aom_highbd_convolve8 avx2/, "$sse2_x86_64";
+
+  add_proto qw/void aom_highbd_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/aom_highbd_convolve8_horiz avx2/, "$sse2_x86_64";
+
+  add_proto qw/void aom_highbd_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/aom_highbd_convolve8_vert avx2/, "$sse2_x86_64";
+
+  add_proto qw/void aom_highbd_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/aom_highbd_convolve8_avg avx2/, "$sse2_x86_64";
+
+  add_proto qw/void aom_highbd_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/aom_highbd_convolve8_avg_horiz avx2/, "$sse2_x86_64";
+
+  add_proto qw/void aom_highbd_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/aom_highbd_convolve8_avg_vert avx2/, "$sse2_x86_64";
+
+  if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
+    add_proto qw/void aom_highbd_convolve8_add_src/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+    add_proto qw/void aom_highbd_convolve8_add_src_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+    add_proto qw/void aom_highbd_convolve8_add_src_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+
+    specialize qw/aom_highbd_convolve8_add_src/, "$sse2_x86_64";
+    # The _horiz/_vert functions are currently unused, so we don't bother
+    # specialising them.
+  }  # CONFIG_LOOP_RESTORATION
+}  # CONFIG_HIGHBITDEPTH
+
+#
+# Loopfilter
+#
+add_proto qw/void aom_lpf_vertical_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/aom_lpf_vertical_16 sse2 neon_asm dspr2 msa/;
+$aom_lpf_vertical_16_neon_asm=aom_lpf_vertical_16_neon;
+
+add_proto qw/void aom_lpf_vertical_16_dual/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/aom_lpf_vertical_16_dual sse2 neon_asm dspr2 msa/;
+$aom_lpf_vertical_16_dual_neon_asm=aom_lpf_vertical_16_dual_neon;
+
+add_proto qw/void aom_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/aom_lpf_vertical_8 sse2 neon dspr2 msa/;
+
+add_proto qw/void aom_lpf_vertical_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/aom_lpf_vertical_8_dual sse2 neon_asm dspr2 msa/;
+$aom_lpf_vertical_8_dual_neon_asm=aom_lpf_vertical_8_dual_neon;
+
+add_proto qw/void aom_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/aom_lpf_vertical_4 sse2 neon dspr2 msa/;
+
+add_proto qw/void aom_lpf_vertical_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/aom_lpf_vertical_4_dual sse2 neon dspr2 msa/;
+
+add_proto qw/void aom_lpf_horizontal_edge_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/aom_lpf_horizontal_edge_8 sse2 avx2 neon_asm dspr2 msa/;
+$aom_lpf_horizontal_edge_8_neon_asm=aom_lpf_horizontal_edge_8_neon;
+
+add_proto qw/void aom_lpf_horizontal_edge_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/aom_lpf_horizontal_edge_16 sse2 avx2 neon_asm dspr2 msa/;
+$aom_lpf_horizontal_edge_16_neon_asm=aom_lpf_horizontal_edge_16_neon;
+
+add_proto qw/void aom_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/aom_lpf_horizontal_8 sse2 neon dspr2 msa/;
+
+add_proto qw/void aom_lpf_horizontal_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/aom_lpf_horizontal_8_dual sse2 neon_asm dspr2 msa/;
+$aom_lpf_horizontal_8_dual_neon_asm=aom_lpf_horizontal_8_dual_neon;
+
+add_proto qw/void aom_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/aom_lpf_horizontal_4 sse2 neon dspr2 msa/;
+
+add_proto qw/void aom_lpf_horizontal_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/aom_lpf_horizontal_4_dual sse2 neon dspr2 msa/;
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void aom_highbd_lpf_vertical_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/aom_highbd_lpf_vertical_16 sse2/;
+
+  add_proto qw/void aom_highbd_lpf_vertical_16_dual/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/aom_highbd_lpf_vertical_16_dual sse2/;
+
+  add_proto qw/void aom_highbd_lpf_vertical_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/aom_highbd_lpf_vertical_8 sse2/;
+
+  add_proto qw/void aom_highbd_lpf_vertical_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+  specialize qw/aom_highbd_lpf_vertical_8_dual sse2/;
+
+  add_proto qw/void aom_highbd_lpf_vertical_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/aom_highbd_lpf_vertical_4 sse2/;
+
+  add_proto qw/void aom_highbd_lpf_vertical_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+  specialize qw/aom_highbd_lpf_vertical_4_dual sse2/;
+
+  add_proto qw/void aom_highbd_lpf_horizontal_edge_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/aom_highbd_lpf_horizontal_edge_8 sse2/;
+
+  add_proto qw/void aom_highbd_lpf_horizontal_edge_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/aom_highbd_lpf_horizontal_edge_16 sse2/;
+
+  add_proto qw/void aom_highbd_lpf_horizontal_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/aom_highbd_lpf_horizontal_8 sse2/;
+
+  add_proto qw/void aom_highbd_lpf_horizontal_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+  specialize qw/aom_highbd_lpf_horizontal_8_dual sse2/;
+
+  add_proto qw/void aom_highbd_lpf_horizontal_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+  specialize qw/aom_highbd_lpf_horizontal_4 sse2/;
+
+  add_proto qw/void aom_highbd_lpf_horizontal_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+  specialize qw/aom_highbd_lpf_horizontal_4_dual sse2/;
+}  # CONFIG_HIGHBITDEPTH
+
+#
+# Encoder functions.
+#
+
+#
+# Forward transform
+#
+if ((aom_config("CONFIG_AV1_ENCODER") eq "yes") || (aom_config("CONFIG_PVQ") eq "yes")){
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    add_proto qw/void aom_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct4x4 sse2/;
+
+    add_proto qw/void aom_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct4x4_1 sse2/;
+
+    add_proto qw/void aom_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct8x8 sse2/, "$ssse3_x86_64";
+
+    add_proto qw/void aom_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct8x8_1 sse2/;
+
+    add_proto qw/void aom_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct16x16 sse2/;
+
+    add_proto qw/void aom_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct16x16_1 sse2 avx2/;
+
+    add_proto qw/void aom_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct32x32 sse2 avx2/;
+
+    add_proto qw/void aom_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct32x32_rd sse2 avx2/;
+
+    add_proto qw/void aom_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct32x32_1 sse2 avx2/;
+
+    # High bit depth
+    add_proto qw/void aom_highbd_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_highbd_fdct4x4 sse2/;
+
+    add_proto qw/void aom_highbd_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_highbd_fdct8x8 sse2/;
+
+    add_proto qw/void aom_highbd_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+
+    add_proto qw/void aom_highbd_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_highbd_fdct16x16 sse2/;
+
+    add_proto qw/void aom_highbd_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+
+    add_proto qw/void aom_highbd_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_highbd_fdct32x32 sse2/;
+
+    add_proto qw/void aom_highbd_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_highbd_fdct32x32_rd sse2/;
+
+    add_proto qw/void aom_highbd_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+  } else {
+    add_proto qw/void aom_fdct4x4/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct4x4 sse2 msa/;
+
+    add_proto qw/void aom_fdct4x4_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct4x4_1 sse2/;
+
+    add_proto qw/void aom_fdct8x8/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct8x8 sse2 neon msa/, "$ssse3_x86_64";
+
+    add_proto qw/void aom_fdct8x8_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct8x8_1 sse2 neon msa/;
+
+    add_proto qw/void aom_fdct16x16/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct16x16 sse2 msa/;
+
+    add_proto qw/void aom_fdct16x16_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct16x16_1 sse2 avx2 msa/;
+
+    add_proto qw/void aom_fdct32x32/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct32x32 sse2 avx2 msa/;
+
+    add_proto qw/void aom_fdct32x32_rd/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct32x32_rd sse2 avx2 msa/;
+
+    add_proto qw/void aom_fdct32x32_1/, "const int16_t *input, tran_low_t *output, int stride";
+    specialize qw/aom_fdct32x32_1 sse2 avx2 msa/;
+  }  # CONFIG_HIGHBITDEPTH
+}  # CONFIG_AV1_ENCODER
+
+#
+# Inverse transform
+if (aom_config("CONFIG_AV1") eq "yes") {
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void aom_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+
+  add_proto qw/void aom_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+  specialize qw/aom_iwht4x4_16_add sse2/;
+
+  add_proto qw/void aom_highbd_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+
+  add_proto qw/void aom_highbd_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+
+  add_proto qw/void aom_highbd_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+
+  add_proto qw/void aom_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+
+  add_proto qw/void aom_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+
+  add_proto qw/void aom_highbd_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+
+  add_proto qw/void aom_highbd_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+
+  add_proto qw/void aom_highbd_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+
+  {
+    add_proto qw/void aom_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct4x4_16_add sse2/;
+
+    add_proto qw/void aom_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct4x4_1_add sse2/;
+
+    add_proto qw/void aom_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct8x8_64_add sse2 ssse3/;
+
+    add_proto qw/void aom_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct8x8_12_add sse2 ssse3/;
+
+    add_proto qw/void aom_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct8x8_1_add sse2/;
+
+    add_proto qw/void aom_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct16x16_256_add sse2/;
+
+    add_proto qw/void aom_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+
+    add_proto qw/void aom_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct16x16_10_add sse2/;
+
+    add_proto qw/void aom_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct16x16_1_add sse2/;
+
+    add_proto qw/void aom_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct32x32_1024_add sse2 ssse3/;
+
+    add_proto qw/void aom_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct32x32_135_add sse2 ssse3/;
+    # Need to add 135 eob idct32x32 implementations.
+    $aom_idct32x32_135_add_sse2=aom_idct32x32_1024_add_sse2;
+
+    add_proto qw/void aom_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct32x32_34_add sse2 ssse3/;
+
+    add_proto qw/void aom_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct32x32_1_add sse2/;
+
+    add_proto qw/void aom_highbd_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
+    specialize qw/aom_highbd_idct4x4_16_add sse2/;
+  }
+} else {
+  {
+    add_proto qw/void aom_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct4x4_1_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void aom_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct4x4_16_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void aom_idct8x8_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct8x8_1_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void aom_idct8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct8x8_64_add sse2 ssse3 neon dspr2 msa/;
+
+    add_proto qw/void aom_idct8x8_12_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct8x8_12_add sse2 ssse3 neon dspr2 msa/;
+
+    add_proto qw/void aom_idct16x16_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct16x16_1_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void aom_idct16x16_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct16x16_256_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void aom_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+
+    add_proto qw/void aom_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct16x16_10_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void aom_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct32x32_1024_add sse2 ssse3 neon dspr2 msa/;
+
+    add_proto qw/void aom_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct32x32_135_add sse2 ssse3 neon dspr2 msa/;
+    # Need to add 135 eob idct32x32 implementations.
+    $aom_idct32x32_135_add_sse2=aom_idct32x32_1024_add_sse2;
+    $aom_idct32x32_135_add_neon=aom_idct32x32_1024_add_neon;
+    $aom_idct32x32_135_add_dspr2=aom_idct32x32_1024_add_dspr2;
+    $aom_idct32x32_135_add_msa=aom_idct32x32_1024_add_msa;
+
+    add_proto qw/void aom_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct32x32_34_add sse2 ssse3 neon dspr2 msa/;
+    # Need to add 34 eob idct32x32 neon implementation.
+    $aom_idct32x32_34_add_neon=aom_idct32x32_1024_add_neon;
+
+    add_proto qw/void aom_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_idct32x32_1_add sse2 neon dspr2 msa/;
+
+    add_proto qw/void aom_iwht4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_iwht4x4_1_add msa/;
+
+    add_proto qw/void aom_iwht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride";
+    specialize qw/aom_iwht4x4_16_add msa sse2/;
+  }
+}  # CONFIG_HIGHBITDEPTH
+}  # CONFIG_AV1
+
+#
+# Quantization
+#
+if (aom_config("CONFIG_AOM_QM") eq "yes") {
+  if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
+    add_proto qw/void aom_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr";
+
+    add_proto qw/void aom_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr";
+
+    add_proto qw/void aom_quantize_b_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr";
+
+    if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+      add_proto qw/void aom_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr";
+
+      add_proto qw/void aom_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr";
+
+      add_proto qw/void aom_highbd_quantize_b_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr";
+    }  # CONFIG_HIGHBITDEPTH
+  }  # CONFIG_AV1_ENCODER
+} else {
+  if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
+    add_proto qw/void aom_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    specialize qw/aom_quantize_b sse2/, "$ssse3_x86_64", "$avx_x86_64";
+
+    add_proto qw/void aom_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    specialize qw/aom_quantize_b_32x32/, "$ssse3_x86_64", "$avx_x86_64";
+
+    add_proto qw/void aom_quantize_b_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+
+    if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+      add_proto qw/void aom_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+      specialize qw/aom_highbd_quantize_b sse2/;
+
+      add_proto qw/void aom_highbd_quantize_b_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+      specialize qw/aom_highbd_quantize_b_32x32 sse2/;
+
+      add_proto qw/void aom_highbd_quantize_b_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    }  # CONFIG_HIGHBITDEPTH
+  }  # CONFIG_AV1_ENCODER
+} # CONFIG_AOM_QM
+if (aom_config("CONFIG_AV1") eq "yes") {
+  #
+  # Alpha blending with mask
+  #
+  add_proto qw/void aom_blend_a64_mask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx";
+  add_proto qw/void aom_blend_a64_hmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w";
+  add_proto qw/void aom_blend_a64_vmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w";
+  specialize "aom_blend_a64_mask", qw/sse4_1/;
+  specialize "aom_blend_a64_hmask", qw/sse4_1/;
+  specialize "aom_blend_a64_vmask", qw/sse4_1/;
+
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    add_proto qw/void aom_highbd_blend_a64_mask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx, int bd";
+    add_proto qw/void aom_highbd_blend_a64_hmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd";
+    add_proto qw/void aom_highbd_blend_a64_vmask/, "uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd";
+    specialize "aom_highbd_blend_a64_mask", qw/sse4_1/;
+    specialize "aom_highbd_blend_a64_hmask", qw/sse4_1/;
+    specialize "aom_highbd_blend_a64_vmask", qw/sse4_1/;
+  }
+}  # CONFIG_AV1
+
+if (aom_config("CONFIG_ENCODERS") eq "yes") {
+#
+# Block subtraction
+#
+add_proto qw/void aom_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
+specialize qw/aom_subtract_block neon msa sse2/;
+
+if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
+#
+# Sum of Squares
+#
+add_proto qw/uint64_t aom_sum_squares_2d_i16/, "const int16_t *src, int stride, int width, int height";
+specialize qw/aom_sum_squares_2d_i16 sse2/;
+
+add_proto qw/uint64_t aom_sum_squares_i16/, "const int16_t *src, uint32_t N";
+specialize qw/aom_sum_squares_i16 sse2/;
+}
+
+
+#
+# Avg
+#
+if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
+  #
+  # Avg
+  #
+  add_proto qw/unsigned int aom_avg_8x8/, "const uint8_t *, int p";
+  specialize qw/aom_avg_8x8 sse2 neon msa/;
+  add_proto qw/unsigned int aom_avg_4x4/, "const uint8_t *, int p";
+  specialize qw/aom_avg_4x4 sse2 neon msa/;
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    add_proto qw/unsigned int aom_highbd_avg_8x8/, "const uint8_t *, int p";
+    add_proto qw/unsigned int aom_highbd_avg_4x4/, "const uint8_t *, int p";
+    add_proto qw/void aom_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd";
+    specialize qw/aom_highbd_subtract_block sse2/;
+  }
+
+  #
+  # Minmax
+  #
+  add_proto qw/void aom_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
+  specialize qw/aom_minmax_8x8 sse2 neon/;
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    add_proto qw/void aom_highbd_minmax_8x8/, "const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max";
+  }
+
+  add_proto qw/void aom_hadamard_8x8/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
+  specialize qw/aom_hadamard_8x8 sse2 neon/, "$ssse3_x86_64";
+
+  add_proto qw/void aom_hadamard_16x16/, "const int16_t *src_diff, int src_stride, int16_t *coeff";
+  specialize qw/aom_hadamard_16x16 sse2 neon/;
+
+  add_proto qw/int aom_satd/, "const int16_t *coeff, int length";
+  specialize qw/aom_satd sse2 neon/;
+
+  add_proto qw/void aom_int_pro_row/, "int16_t *hbuf, const uint8_t *ref, int ref_stride, int height";
+  specialize qw/aom_int_pro_row sse2 neon/;
+
+  add_proto qw/int16_t aom_int_pro_col/, "const uint8_t *ref, int width";
+  specialize qw/aom_int_pro_col sse2 neon/;
+
+  add_proto qw/int aom_vector_var/, "const int16_t *ref, const int16_t *src, int bwl";
+  specialize qw/aom_vector_var neon sse2/;
+}  # CONFIG_AV1_ENCODER
+
+#
+# Single block SAD / Single block Avg SAD
+#
+foreach (@block_sizes) {
+  ($w, $h) = @$_;
+  add_proto qw/unsigned int/, "aom_sad${w}x${h}", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+  add_proto qw/unsigned int/, "aom_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+}
+
+specialize qw/aom_sad128x128    avx2                sse2/;
+specialize qw/aom_sad128x64     avx2                sse2/;
+specialize qw/aom_sad64x128     avx2                sse2/;
+specialize qw/aom_sad64x64      avx2       neon msa sse2/;
+specialize qw/aom_sad64x32      avx2            msa sse2/;
+specialize qw/aom_sad32x64      avx2            msa sse2/;
+specialize qw/aom_sad32x32      avx2       neon msa sse2/;
+specialize qw/aom_sad32x16      avx2            msa sse2/;
+specialize qw/aom_sad16x32                      msa sse2/;
+specialize qw/aom_sad16x16           media neon msa sse2/;
+specialize qw/aom_sad16x8                  neon msa sse2/;
+specialize qw/aom_sad8x16                  neon msa sse2/;
+specialize qw/aom_sad8x8                   neon msa sse2/;
+specialize qw/aom_sad8x4                        msa sse2/;
+specialize qw/aom_sad4x8                        msa sse2/;
+specialize qw/aom_sad4x4                   neon msa sse2/;
+
+specialize qw/aom_sad128x128_avg avx2     sse2/;
+specialize qw/aom_sad128x64_avg  avx2     sse2/;
+specialize qw/aom_sad64x128_avg  avx2     sse2/;
+specialize qw/aom_sad64x64_avg   avx2 msa sse2/;
+specialize qw/aom_sad64x32_avg   avx2 msa sse2/;
+specialize qw/aom_sad32x64_avg   avx2 msa sse2/;
+specialize qw/aom_sad32x32_avg   avx2 msa sse2/;
+specialize qw/aom_sad32x16_avg   avx2 msa sse2/;
+specialize qw/aom_sad16x32_avg        msa sse2/;
+specialize qw/aom_sad16x16_avg        msa sse2/;
+specialize qw/aom_sad16x8_avg         msa sse2/;
+specialize qw/aom_sad8x16_avg         msa sse2/;
+specialize qw/aom_sad8x8_avg          msa sse2/;
+specialize qw/aom_sad8x4_avg          msa sse2/;
+specialize qw/aom_sad4x8_avg          msa sse2/;
+specialize qw/aom_sad4x4_avg          msa sse2/;
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  foreach (@block_sizes) {
+    ($w, $h) = @$_;
+    add_proto qw/unsigned int/, "aom_highbd_sad${w}x${h}", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+    add_proto qw/unsigned int/, "aom_highbd_sad${w}x${h}_avg", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
+    if ($w != 128 && $h != 128 && $w != 4) {
+      specialize "aom_highbd_sad${w}x${h}", qw/sse2/;
+      specialize "aom_highbd_sad${w}x${h}_avg", qw/sse2/;
+    }
+  }
+  specialize qw/aom_highbd_sad128x128 avx2/;
+  specialize qw/aom_highbd_sad128x64  avx2/;
+  specialize qw/aom_highbd_sad64x128  avx2/;
+  specialize qw/aom_highbd_sad64x64   avx2/;
+  specialize qw/aom_highbd_sad64x32   avx2/;
+  specialize qw/aom_highbd_sad32x64   avx2/;
+  specialize qw/aom_highbd_sad32x32   avx2/;
+  specialize qw/aom_highbd_sad32x16   avx2/;
+  specialize qw/aom_highbd_sad16x32   avx2/;
+  specialize qw/aom_highbd_sad16x16   avx2/;
+  specialize qw/aom_highbd_sad16x8    avx2/;
+
+  specialize qw/aom_highbd_sad128x128_avg avx2/;
+  specialize qw/aom_highbd_sad128x64_avg  avx2/;
+  specialize qw/aom_highbd_sad64x128_avg  avx2/;
+  specialize qw/aom_highbd_sad64x64_avg   avx2/;
+  specialize qw/aom_highbd_sad64x32_avg   avx2/;
+  specialize qw/aom_highbd_sad32x64_avg   avx2/;
+  specialize qw/aom_highbd_sad32x32_avg   avx2/;
+  specialize qw/aom_highbd_sad32x16_avg   avx2/;
+  specialize qw/aom_highbd_sad16x32_avg   avx2/;
+  specialize qw/aom_highbd_sad16x16_avg   avx2/;
+  specialize qw/aom_highbd_sad16x8_avg    avx2/;
+}
+
+#
+# Masked SAD
+#
+if (aom_config("CONFIG_EXT_INTER") eq "yes") {
+  foreach (@block_sizes) {
+    ($w, $h) = @$_;
+    add_proto qw/unsigned int/, "aom_masked_sad${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *mask, int mask_stride";
+    specialize "aom_masked_sad${w}x${h}", qw/ssse3/;
+  }
+
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    foreach (@block_sizes) {
+      ($w, $h) = @$_;
+      add_proto qw/unsigned int/, "aom_highbd_masked_sad${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *mask, int mask_stride";
+      specialize "aom_highbd_masked_sad${w}x${h}", qw/ssse3/;
+    }
+  }
+}
+
+#
+# OBMC SAD
+#
+if (aom_config("CONFIG_MOTION_VAR") eq "yes") {
+  foreach (@block_sizes) {
+    ($w, $h) = @$_;
+    add_proto qw/unsigned int/, "aom_obmc_sad${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask";
+    specialize "aom_obmc_sad${w}x${h}", qw/sse4_1/;
+  }
+
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    foreach (@block_sizes) {
+      ($w, $h) = @$_;
+      add_proto qw/unsigned int/, "aom_highbd_obmc_sad${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask";
+      specialize "aom_highbd_obmc_sad${w}x${h}", qw/sse4_1/;
+    }
+  }
+}
+
+#
+# Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally
+#
+# Blocks of 3
+foreach $s (@block_widths) {
+  add_proto qw/void/, "aom_sad${s}x${s}x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+}
+specialize qw/aom_sad64x64x3            msa/;
+specialize qw/aom_sad32x32x3            msa/;
+specialize qw/aom_sad16x16x3 sse3 ssse3 msa/;
+specialize qw/aom_sad8x8x3   sse3       msa/;
+specialize qw/aom_sad4x4x3   sse3       msa/;
+
+add_proto qw/void/, "aom_sad16x8x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/aom_sad16x8x3 sse3 ssse3 msa/;
+add_proto qw/void/, "aom_sad8x16x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/aom_sad8x16x3 sse3 msa/;
+
+# Blocks of 8
+foreach $s (@block_widths) {
+  add_proto qw/void/, "aom_sad${s}x${s}x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+}
+specialize qw/aom_sad64x64x8        msa/;
+specialize qw/aom_sad32x32x8        msa/;
+specialize qw/aom_sad16x16x8 sse4_1 msa/;
+specialize qw/aom_sad8x8x8   sse4_1 msa/;
+specialize qw/aom_sad4x4x8   sse4_1 msa/;
+
+add_proto qw/void/, "aom_sad16x8x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/aom_sad16x8x8 sse4_1 msa/;
+add_proto qw/void/, "aom_sad8x16x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/aom_sad8x16x8 sse4_1 msa/;
+add_proto qw/void/, "aom_sad8x4x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/aom_sad8x4x8 msa/;
+add_proto qw/void/, "aom_sad4x8x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+specialize qw/aom_sad4x8x8 msa/;
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  foreach $s (@block_widths) {
+    # Blocks of 3
+    add_proto qw/void/, "aom_highbd_sad${s}x${s}x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+    # Blocks of 8
+    add_proto qw/void/, "aom_highbd_sad${s}x${s}x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  }
+  # Blocks of 3
+  add_proto qw/void/, "aom_highbd_sad16x8x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  add_proto qw/void/, "aom_highbd_sad8x16x3", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  # Blocks of 8
+  add_proto qw/void/, "aom_highbd_sad16x8x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  add_proto qw/void/, "aom_highbd_sad8x16x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  add_proto qw/void/, "aom_highbd_sad8x4x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+  add_proto qw/void/, "aom_highbd_sad4x8x8", "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array";
+}
+
+#
+# Multi-block SAD, comparing a reference to N independent blocks
+#
+foreach (@block_sizes) {
+  ($w, $h) = @$_;
+  add_proto qw/void/, "aom_sad${w}x${h}x4d", "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+}
+
+specialize qw/aom_sad128x128x4d avx2          sse2/;
+specialize qw/aom_sad128x64x4d  avx2          sse2/;
+specialize qw/aom_sad64x128x4d  avx2          sse2/;
+specialize qw/aom_sad64x64x4d   avx2 neon msa sse2/;
+specialize qw/aom_sad64x32x4d   avx2      msa sse2/;
+specialize qw/aom_sad32x64x4d   avx2      msa sse2/;
+specialize qw/aom_sad32x32x4d   avx2 neon msa sse2/;
+specialize qw/aom_sad32x16x4d             msa sse2/;
+specialize qw/aom_sad16x32x4d             msa sse2/;
+specialize qw/aom_sad16x16x4d        neon msa sse2/;
+specialize qw/aom_sad16x8x4d              msa sse2/;
+specialize qw/aom_sad8x16x4d              msa sse2/;
+specialize qw/aom_sad8x8x4d               msa sse2/;
+specialize qw/aom_sad8x4x4d               msa sse2/;
+specialize qw/aom_sad4x8x4d               msa sse2/;
+specialize qw/aom_sad4x4x4d               msa sse2/;
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  #
+  # Multi-block SAD, comparing a reference to N independent blocks
+  #
+  foreach (@block_sizes) {
+    ($w, $h) = @$_;
+    add_proto qw/void/, "aom_highbd_sad${w}x${h}x4d", "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
+    if ($w != 128 && $h != 128) {
+      specialize "aom_highbd_sad${w}x${h}x4d", qw/sse2/;
+    }
+  }
+  specialize qw/aom_highbd_sad128x128x4d avx2/;
+  specialize qw/aom_highbd_sad128x64x4d  avx2/;
+  specialize qw/aom_highbd_sad64x128x4d  avx2/;
+  specialize qw/aom_highbd_sad64x64x4d   avx2/;
+  specialize qw/aom_highbd_sad64x32x4d   avx2/;
+  specialize qw/aom_highbd_sad32x64x4d   avx2/;
+  specialize qw/aom_highbd_sad32x32x4d   avx2/;
+  specialize qw/aom_highbd_sad32x16x4d   avx2/;
+  specialize qw/aom_highbd_sad16x32x4d   avx2/;
+  specialize qw/aom_highbd_sad16x16x4d   avx2/;
+  specialize qw/aom_highbd_sad16x8x4d    avx2/;
+}
+
+#
+# Structured Similarity (SSIM)
+#
+if (aom_config("CONFIG_INTERNAL_STATS") eq "yes") {
+  add_proto qw/void aom_ssim_parms_8x8/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
+  specialize qw/aom_ssim_parms_8x8/, "$sse2_x86_64";
+
+  add_proto qw/void aom_ssim_parms_16x16/, "const uint8_t *s, int sp, const uint8_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
+  specialize qw/aom_ssim_parms_16x16/, "$sse2_x86_64";
+
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    add_proto qw/void aom_highbd_ssim_parms_8x8/, "const uint16_t *s, int sp, const uint16_t *r, int rp, uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s, uint32_t *sum_sq_r, uint32_t *sum_sxr";
+  }
+}
+}  # CONFIG_ENCODERS
+
+if (aom_config("CONFIG_ENCODERS") eq "yes") {
+
+#
+# Specialty Variance
+#
+add_proto qw/void aom_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+
+add_proto qw/void aom_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+
+specialize qw/aom_get16x16var sse2 avx2 neon msa/;
+specialize qw/aom_get8x8var   sse2      neon msa/;
+
+
+add_proto qw/unsigned int aom_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+add_proto qw/unsigned int aom_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+add_proto qw/unsigned int aom_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+add_proto qw/unsigned int aom_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+
+specialize qw/aom_mse16x16          sse2 avx2 media neon msa/;
+specialize qw/aom_mse16x8           sse2                 msa/;
+specialize qw/aom_mse8x16           sse2                 msa/;
+specialize qw/aom_mse8x8            sse2                 msa/;
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  foreach $bd (8, 10, 12) {
+    add_proto qw/void/, "aom_highbd_${bd}_get16x16var", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+    add_proto qw/void/, "aom_highbd_${bd}_get8x8var", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+
+    add_proto qw/unsigned int/, "aom_highbd_${bd}_mse16x16", "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+    add_proto qw/unsigned int/, "aom_highbd_${bd}_mse16x8", "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+    add_proto qw/unsigned int/, "aom_highbd_${bd}_mse8x16", "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+    add_proto qw/unsigned int/, "aom_highbd_${bd}_mse8x8", "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+
+    specialize "aom_highbd_${bd}_mse16x16", qw/sse2/;
+    specialize "aom_highbd_${bd}_mse8x8", qw/sse2/;
+  }
+}
+
+#
+# ...
+#
+add_proto qw/void aom_upsampled_pred/, "uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride";
+specialize qw/aom_upsampled_pred sse2/;
+add_proto qw/void aom_comp_avg_upsampled_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride";
+specialize qw/aom_comp_avg_upsampled_pred sse2/;
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void aom_highbd_upsampled_pred/, "uint16_t *comp_pred, int width, int height, const uint8_t *ref8, int ref_stride";
+  specialize qw/aom_highbd_upsampled_pred sse2/;
+  add_proto qw/void aom_highbd_comp_avg_upsampled_pred/, "uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride";
+  specialize qw/aom_highbd_comp_avg_upsampled_pred sse2/;
+}
+
+#
+# ...
+#
+add_proto qw/unsigned int aom_get_mb_ss/, "const int16_t *";
+add_proto qw/unsigned int aom_get4x4sse_cs/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride";
+
+specialize qw/aom_get_mb_ss sse2 msa/;
+specialize qw/aom_get4x4sse_cs neon msa/;
+
+#
+# Variance / Subpixel Variance / Subpixel Avg Variance
+#
+  add_proto qw/unsigned int/, "aom_variance2x2", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+  add_proto qw/unsigned int/, "aom_variance2x4", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+  add_proto qw/unsigned int/, "aom_variance4x2", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+foreach (@block_sizes) {
+  ($w, $h) = @$_;
+  add_proto qw/unsigned int/, "aom_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/uint32_t/, "aom_sub_pixel_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  add_proto qw/uint32_t/, "aom_sub_pixel_avg_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+}
+
+specialize qw/aom_variance64x64     sse2 avx2       neon msa/;
+specialize qw/aom_variance64x32     sse2 avx2       neon msa/;
+specialize qw/aom_variance32x64     sse2            neon msa/;
+specialize qw/aom_variance32x32     sse2 avx2       neon msa/;
+specialize qw/aom_variance32x16     sse2 avx2            msa/;
+specialize qw/aom_variance16x32     sse2                 msa/;
+specialize qw/aom_variance16x16     sse2 avx2 media neon msa/;
+specialize qw/aom_variance16x8      sse2            neon msa/;
+specialize qw/aom_variance8x16      sse2            neon msa/;
+specialize qw/aom_variance8x8       sse2      media neon msa/;
+specialize qw/aom_variance8x4       sse2                 msa/;
+specialize qw/aom_variance4x8       sse2                 msa/;
+specialize qw/aom_variance4x4       sse2                 msa/;
+
+specialize qw/aom_sub_pixel_variance64x64     avx2       neon msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance64x32                     msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance32x64                     msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance32x32     avx2       neon msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance32x16                     msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance16x32                     msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance16x16          media neon msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance16x8                      msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance8x16                      msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance8x8            media neon msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance8x4                       msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance4x8                       msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_variance4x4                       msa sse2 ssse3/;
+
+specialize qw/aom_sub_pixel_avg_variance64x64 avx2 msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance64x32      msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance32x64      msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance32x32 avx2 msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance32x16      msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance16x32      msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance16x16      msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance16x8       msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance8x16       msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance8x8        msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance8x4        msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance4x8        msa sse2 ssse3/;
+specialize qw/aom_sub_pixel_avg_variance4x4        msa sse2 ssse3/;
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  foreach $bd (8, 10, 12) {
+    add_proto qw/unsigned int/, "aom_highbd_${bd}_variance2x2", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+    add_proto qw/unsigned int/, "aom_highbd_${bd}_variance2x4", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+    add_proto qw/unsigned int/, "aom_highbd_${bd}_variance4x2", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+    foreach (@block_sizes) {
+      ($w, $h) = @$_;
+      add_proto qw/unsigned int/, "aom_highbd_${bd}_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+      add_proto qw/uint32_t/, "aom_highbd_${bd}_sub_pixel_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+      add_proto qw/uint32_t/, "aom_highbd_${bd}_sub_pixel_avg_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+      if ($w != 128 && $h != 128 && $w != 4 && $h != 4) {
+        specialize "aom_highbd_${bd}_variance${w}x${h}", "sse2";
+      }
+      if ($w == 4 && $h == 4) {
+        specialize "aom_highbd_${bd}_variance${w}x${h}", "sse4_1";
+      }
+      if ($w != 128 && $h != 128 && $w != 4) {
+        specialize "aom_highbd_${bd}_sub_pixel_variance${w}x${h}", qw/sse2/;
+        specialize "aom_highbd_${bd}_sub_pixel_avg_variance${w}x${h}", qw/sse2/;
+      }
+      if ($w == 4 && $h == 4) {
+        specialize "aom_highbd_${bd}_sub_pixel_variance${w}x${h}", "sse4_1";
+        specialize "aom_highbd_${bd}_sub_pixel_avg_variance${w}x${h}", "sse4_1";
+      }
+    }
+  }
+}  # CONFIG_HIGHBITDEPTH
+
+if (aom_config("CONFIG_EXT_INTER") eq "yes") {
+#
+# Masked Variance / Masked Subpixel Variance
+#
+  foreach (@block_sizes) {
+    ($w, $h) = @$_;
+    add_proto qw/unsigned int/, "aom_masked_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *mask, int mask_stride, unsigned int *sse";
+    add_proto qw/unsigned int/, "aom_masked_sub_pixel_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, const uint8_t *mask, int mask_stride, unsigned int *sse";
+    specialize "aom_masked_variance${w}x${h}", qw/ssse3/;
+    specialize "aom_masked_sub_pixel_variance${w}x${h}", qw/ssse3/;
+  }
+
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    foreach $bd ("_", "_10_", "_12_") {
+      foreach (@block_sizes) {
+        ($w, $h) = @$_;
+        add_proto qw/unsigned int/, "aom_highbd${bd}masked_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *mask, int mask_stride, unsigned int *sse";
+        add_proto qw/unsigned int/, "aom_highbd${bd}masked_sub_pixel_variance${w}x${h}", "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, const uint8_t *m, int m_stride, unsigned int *sse";
+        specialize "aom_highbd${bd}masked_variance${w}x${h}", qw/ssse3/;
+        specialize "aom_highbd${bd}masked_sub_pixel_variance${w}x${h}", qw/ssse3/;
+      }
+    }
+  }
+}
+
+#
+# OBMC Variance / OBMC Subpixel Variance
+#
+if (aom_config("CONFIG_MOTION_VAR") eq "yes") {
+  foreach (@block_sizes) {
+    ($w, $h) = @$_;
+    add_proto qw/unsigned int/, "aom_obmc_variance${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse";
+    add_proto qw/unsigned int/, "aom_obmc_sub_pixel_variance${w}x${h}", "const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse";
+    specialize "aom_obmc_variance${w}x${h}", q/sse4_1/;
+  }
+
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    foreach $bd ("_", "_10_", "_12_") {
+      foreach (@block_sizes) {
+        ($w, $h) = @$_;
+        add_proto qw/unsigned int/, "aom_highbd${bd}obmc_variance${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse";
+        add_proto qw/unsigned int/, "aom_highbd${bd}obmc_sub_pixel_variance${w}x${h}", "const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse";
+        specialize "aom_highbd${bd}obmc_variance${w}x${h}", qw/sse4_1/;
+      }
+    }
+  }
+}
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance64x64 avx2 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance64x32 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance32x64 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance32x32 avx2 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance32x16 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance16x32 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance16x16 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance16x8 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance8x16 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance8x8 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance8x4 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance4x8 msa sse2 ssse3/;
+
+add_proto qw/uint32_t aom_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_sub_pixel_avg_variance4x4 msa sse2 ssse3/;
+
+#
+# Specialty Subpixel
+#
+add_proto qw/uint32_t aom_variance_halfpixvar16x16_h/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
+  specialize qw/aom_variance_halfpixvar16x16_h sse2 media/;
+
+add_proto qw/uint32_t aom_variance_halfpixvar16x16_v/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
+  specialize qw/aom_variance_halfpixvar16x16_v sse2 media/;
+
+add_proto qw/uint32_t aom_variance_halfpixvar16x16_hv/, "const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse";
+  specialize qw/aom_variance_halfpixvar16x16_hv sse2 media/;
+
+#
+# Comp Avg
+#
+add_proto qw/void aom_comp_avg_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride";
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/unsigned int aom_highbd_12_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_variance64x64 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_variance64x32 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_variance32x64 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_variance32x32 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_variance32x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_variance16x32 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_variance16x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_variance16x8 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_variance8x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_variance8x8 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_12_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_12_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+  add_proto qw/unsigned int aom_highbd_10_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_variance64x64 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_variance64x32 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_variance32x64 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_variance32x32 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_variance32x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_variance16x32 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_variance16x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_variance16x8 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_variance8x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_variance8x8 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_10_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_10_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+  add_proto qw/unsigned int aom_highbd_8_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_variance64x64 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_variance64x32 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_variance32x64 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_variance32x32 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_variance32x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_variance32x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_variance16x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_variance16x32 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_variance16x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_variance16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_variance16x8 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_variance8x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_variance8x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_variance8x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_variance8x8 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_variance8x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_8_variance4x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_8_variance4x4/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
+
+  add_proto qw/void aom_highbd_8_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+  add_proto qw/void aom_highbd_8_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+
+  add_proto qw/void aom_highbd_10_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+  add_proto qw/void aom_highbd_10_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+
+  add_proto qw/void aom_highbd_12_get16x16var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+  add_proto qw/void aom_highbd_12_get8x8var/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum";
+
+  add_proto qw/unsigned int aom_highbd_8_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_mse16x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_8_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_8_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_8_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/aom_highbd_8_mse8x8 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_mse16x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_10_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_10_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_10_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/aom_highbd_10_mse8x8 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_mse16x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_mse16x16 sse2/;
+
+  add_proto qw/unsigned int aom_highbd_12_mse16x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_12_mse8x16/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  add_proto qw/unsigned int aom_highbd_12_mse8x8/, "const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse";
+  specialize qw/aom_highbd_12_mse8x8 sse2/;
+
+  add_proto qw/void aom_highbd_comp_avg_pred/, "uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride";
+
+  #
+  # Subpixel Variance
+  #
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance64x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance64x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance32x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance32x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance32x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance16x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance16x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance16x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance8x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance8x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_12_sub_pixel_variance8x4 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance64x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance64x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance32x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance32x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance32x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance16x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance16x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance16x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance8x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance8x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_10_sub_pixel_variance8x4 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance64x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance64x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance32x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance32x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance32x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance16x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance16x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance16x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance8x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance8x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  specialize qw/aom_highbd_8_sub_pixel_variance8x4 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance64x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance64x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance32x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance32x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance32x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance16x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance16x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance16x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance8x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance8x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_12_sub_pixel_avg_variance8x4 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  add_proto qw/uint32_t aom_highbd_12_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance64x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance64x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance32x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance32x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance32x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance16x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance16x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance16x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance8x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance8x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_10_sub_pixel_avg_variance8x4 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  add_proto qw/uint32_t aom_highbd_10_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance64x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance64x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance32x64 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance32x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance32x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance16x32 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance16x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance16x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance8x16 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance8x8 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  specialize qw/aom_highbd_8_sub_pixel_avg_variance8x4 sse2/;
+
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+  add_proto qw/uint32_t aom_highbd_8_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
+
+}  # CONFIG_HIGHBITDEPTH
+
+}  # CONFIG_ENCODERS
+
+1;
diff --git a/third_party/aom/aom_dsp/aom_filter.h b/third_party/aom/aom_dsp/aom_filter.h
new file mode 100644
index 0000000000..04d113dd36
--- /dev/null
+++ b/third_party/aom/aom_dsp/aom_filter.h
@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_AOM_FILTER_H_
+#define AOM_DSP_AOM_FILTER_H_
+
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define FILTER_BITS 7
+
+#define SUBPEL_BITS 4
+#define SUBPEL_MASK ((1 << SUBPEL_BITS) - 1)
+#define SUBPEL_SHIFTS (1 << SUBPEL_BITS)
+#define SUBPEL_TAPS 8
+
+typedef int16_t InterpKernel[SUBPEL_TAPS];
+
+#define BIL_SUBPEL_BITS 3
+#define BIL_SUBPEL_SHIFTS (1 << BIL_SUBPEL_BITS)
+
+// 2 tap bilinear filters
+static const uint8_t bilinear_filters_2t[BIL_SUBPEL_SHIFTS][2] = {
+  { 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 },
+  { 64, 64 }, { 48, 80 },  { 32, 96 }, { 16, 112 },
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_AOM_FILTER_H_
diff --git a/third_party/aom/aom_dsp/aom_simd.h b/third_party/aom/aom_dsp/aom_simd.h
new file mode 100644
index 0000000000..469fd8ed23
--- /dev/null
+++ b/third_party/aom/aom_dsp/aom_simd.h
@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_AOM_AOM_SIMD_H_
+#define AOM_DSP_AOM_AOM_SIMD_H_
+
+#include <stdint.h>
+
+#if defined(_WIN32)
+#include <intrin.h>
+#endif
+
+#include "./aom_config.h"
+#include "./aom_simd_inline.h"
+
+#define SIMD_CHECK 1  // Sanity checks in C equivalents
+
+#if HAVE_NEON
+#include "simd/v256_intrinsics_arm.h"
+// VS compiling for 32 bit targets does not support vector types in
+// structs as arguments, which makes the v256 type of the intrinsics
+// hard to support, so optimizations for this target are disabled.
+#elif HAVE_SSE2 && (defined(_WIN64) || !defined(_MSC_VER) || defined(__clang__))
+#include "simd/v256_intrinsics_x86.h"
+#else
+#include "simd/v256_intrinsics.h"
+#endif
+
+#endif  // AOM_DSP_AOM_AOM_SIMD_H_
diff --git a/third_party/aom/aom_dsp/aom_simd_inline.h b/third_party/aom/aom_dsp/aom_simd_inline.h
new file mode 100644
index 0000000000..02a8b3a170
--- /dev/null
+++ b/third_party/aom/aom_dsp/aom_simd_inline.h
@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_AOM_SIMD_INLINE_H_
+#define AOM_DSP_AOM_SIMD_INLINE_H_
+
+#include "aom/aom_integer.h"
+
+#ifndef SIMD_INLINE
+#define SIMD_INLINE static AOM_FORCE_INLINE
+#endif
+
+#endif  // AOM_DSP_AOM_SIMD_INLINE_H_
diff --git a/third_party/aom/aom_dsp/arm/aom_convolve8_avg_neon.c b/third_party/aom/aom_dsp/arm/aom_convolve8_avg_neon.c
new file mode 100644
index 0000000000..09429d6d26
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/aom_convolve8_avg_neon.c
@@ -0,0 +1,364 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+static INLINE int32x4_t MULTIPLY_BY_Q0(int16x4_t dsrc0, int16x4_t dsrc1,
+                                       int16x4_t dsrc2, int16x4_t dsrc3,
+                                       int16x4_t dsrc4, int16x4_t dsrc5,
+                                       int16x4_t dsrc6, int16x4_t dsrc7,
+                                       int16x8_t q0s16) {
+  int32x4_t qdst;
+  int16x4_t d0s16, d1s16;
+
+  d0s16 = vget_low_s16(q0s16);
+  d1s16 = vget_high_s16(q0s16);
+
+  qdst = vmull_lane_s16(dsrc0, d0s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc1, d0s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc2, d0s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc3, d0s16, 3);
+  qdst = vmlal_lane_s16(qdst, dsrc4, d1s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc5, d1s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc6, d1s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc7, d1s16, 3);
+  return qdst;
+}
+
+void aom_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y,  // unused
+                                  int y_step_q4,            // unused
+                                  int w, int h) {
+  int width;
+  const uint8_t *s;
+  uint8_t *d;
+  uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
+  uint32x2_t d2u32, d3u32, d6u32, d7u32, d28u32, d29u32, d30u32, d31u32;
+  uint8x16_t q1u8, q3u8, q12u8, q13u8, q14u8, q15u8;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+  uint16x8x2_t q0x2u16;
+  uint8x8x2_t d0x2u8, d1x2u8;
+  uint32x2x2_t d0x2u32;
+  uint16x4x2_t d0x2u16, d1x2u16;
+  uint32x4x2_t q0x2u32;
+
+  assert(x_step_q4 == 16);
+
+  (void)x_step_q4;
+  (void)y_step_q4;
+  (void)filter_y;
+
+  q0s16 = vld1q_s16(filter_x);
+
+  src -= 3;                // adjust for taps
+  for (; h > 0; h -= 4) {  // loop_horiz_v
+    s = src;
+    d24u8 = vld1_u8(s);
+    s += src_stride;
+    d25u8 = vld1_u8(s);
+    s += src_stride;
+    d26u8 = vld1_u8(s);
+    s += src_stride;
+    d27u8 = vld1_u8(s);
+
+    q12u8 = vcombine_u8(d24u8, d25u8);
+    q13u8 = vcombine_u8(d26u8, d27u8);
+
+    q0x2u16 =
+        vtrnq_u16(vreinterpretq_u16_u8(q12u8), vreinterpretq_u16_u8(q13u8));
+    d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
+    d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
+    d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
+    d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
+    d0x2u8 = vtrn_u8(d24u8, d25u8);
+    d1x2u8 = vtrn_u8(d26u8, d27u8);
+
+    __builtin_prefetch(src + src_stride * 4);
+    __builtin_prefetch(src + src_stride * 5);
+
+    q8u16 = vmovl_u8(d0x2u8.val[0]);
+    q9u16 = vmovl_u8(d0x2u8.val[1]);
+    q10u16 = vmovl_u8(d1x2u8.val[0]);
+    q11u16 = vmovl_u8(d1x2u8.val[1]);
+
+    src += 7;
+    d16u16 = vget_low_u16(q8u16);
+    d17u16 = vget_high_u16(q8u16);
+    d18u16 = vget_low_u16(q9u16);
+    d19u16 = vget_high_u16(q9u16);
+    q8u16 = vcombine_u16(d16u16, d18u16);  // vswp 17 18
+    q9u16 = vcombine_u16(d17u16, d19u16);
+
+    d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));         // vmov 23 21
+    for (width = w; width > 0; width -= 4, src += 4, dst += 4) {  // loop_horiz
+      s = src;
+      d28u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d29u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d31u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d30u32 = vld1_dup_u32((const uint32_t *)s);
+
+      __builtin_prefetch(src + 64);
+
+      d0x2u16 =
+          vtrn_u16(vreinterpret_u16_u32(d28u32), vreinterpret_u16_u32(d31u32));
+      d1x2u16 =
+          vtrn_u16(vreinterpret_u16_u32(d29u32), vreinterpret_u16_u32(d30u32));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]),   // d28
+                       vreinterpret_u8_u16(d1x2u16.val[0]));  // d29
+      d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]),   // d31
+                       vreinterpret_u8_u16(d1x2u16.val[1]));  // d30
+
+      __builtin_prefetch(src + 64 + src_stride);
+
+      q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+      q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
+      q0x2u32 =
+          vtrnq_u32(vreinterpretq_u32_u8(q14u8), vreinterpretq_u32_u8(q15u8));
+
+      d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
+      d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
+      q12u16 = vmovl_u8(d28u8);
+      q13u16 = vmovl_u8(d29u8);
+
+      __builtin_prefetch(src + 64 + src_stride * 2);
+
+      d = dst;
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 0);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 0);
+      d += dst_stride;
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 1);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 1);
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+      d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+      d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16, d18s16, d19s16,
+                             d23s16, d24s16, q0s16);
+      q2s32 = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16, d19s16, d23s16,
+                             d24s16, d26s16, q0s16);
+      q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16, d23s16, d24s16,
+                              d26s16, d27s16, q0s16);
+      q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16, d24s16, d26s16,
+                              d27s16, d25s16, q0s16);
+
+      __builtin_prefetch(src + 64 + src_stride * 3);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u8 = vqmovn_u16(q1u16);
+      d3u8 = vqmovn_u16(q2u16);
+
+      d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8), vreinterpret_u16_u8(d3u8));
+      d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
+                         vreinterpret_u32_u16(d0x2u16.val[1]));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
+                       vreinterpret_u8_u32(d0x2u32.val[1]));
+
+      q1u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+      q3u8 = vreinterpretq_u8_u32(vcombine_u32(d6u32, d7u32));
+
+      q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+      d2u32 = vreinterpret_u32_u8(vget_low_u8(q1u8));
+      d3u32 = vreinterpret_u32_u8(vget_high_u8(q1u8));
+
+      d = dst;
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+
+      q8u16 = q9u16;
+      d20s16 = d23s16;
+      q11u16 = q12u16;
+      q9u16 = q13u16;
+      d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    }
+    src += src_stride * 4 - w - 7;
+    dst += dst_stride * 4 - w;
+  }
+  return;
+}
+
+void aom_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const int16_t *filter_x,  // unused
+                                 int x_step_q4,            // unused
+                                 const int16_t *filter_y, int y_step_q4, int w,
+                                 int h) {
+  int height;
+  const uint8_t *s;
+  uint8_t *d;
+  uint8x8_t d2u8, d3u8;
+  uint32x2_t d2u32, d3u32, d6u32, d7u32;
+  uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
+  uint8x16_t q1u8, q3u8;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+
+  assert(y_step_q4 == 16);
+
+  (void)x_step_q4;
+  (void)y_step_q4;
+  (void)filter_x;
+
+  src -= src_stride * 3;
+  q0s16 = vld1q_s16(filter_y);
+  for (; w > 0; w -= 4, src += 4, dst += 4) {  // loop_vert_h
+    s = src;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
+    s += src_stride;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
+    s += src_stride;
+    d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
+    s += src_stride;
+
+    q8u16 = vmovl_u8(vreinterpret_u8_u32(d16u32));
+    q9u16 = vmovl_u8(vreinterpret_u8_u32(d18u32));
+    q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
+    q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
+
+    d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+    d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d = dst;
+    for (height = h; height > 0; height -= 4) {  // loop_vert
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
+      s += src_stride;
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
+      s += src_stride;
+
+      q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
+      q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
+
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 0);
+      d += dst_stride;
+      d6u32 = vld1_lane_u32((const uint32_t *)d, d6u32, 1);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 0);
+      d += dst_stride;
+      d7u32 = vld1_lane_u32((const uint32_t *)d, d7u32, 1);
+      d -= dst_stride * 3;
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+      d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      __builtin_prefetch(s);
+      __builtin_prefetch(s + src_stride);
+      q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16, d20s16, d21s16,
+                             d22s16, d24s16, q0s16);
+      __builtin_prefetch(s + src_stride * 2);
+      __builtin_prefetch(s + src_stride * 3);
+      q2s32 = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16, d21s16, d22s16,
+                             d24s16, d26s16, q0s16);
+      __builtin_prefetch(d);
+      __builtin_prefetch(d + dst_stride);
+      q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16, d22s16, d24s16,
+                              d26s16, d27s16, q0s16);
+      __builtin_prefetch(d + dst_stride * 2);
+      __builtin_prefetch(d + dst_stride * 3);
+      q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16, d24s16, d26s16,
+                              d27s16, d25s16, q0s16);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u8 = vqmovn_u16(q1u16);
+      d3u8 = vqmovn_u16(q2u16);
+
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      q3u8 = vreinterpretq_u8_u32(vcombine_u32(d6u32, d7u32));
+
+      q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+      d2u32 = vreinterpret_u32_u8(vget_low_u8(q1u8));
+      d3u32 = vreinterpret_u32_u8(vget_high_u8(q1u8));
+
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+      d += dst_stride;
+
+      q8u16 = q10u16;
+      d18s16 = d22s16;
+      d19s16 = d24s16;
+      q10u16 = q13u16;
+      d22s16 = d25s16;
+    }
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/aom_convolve8_avg_neon_asm.asm b/third_party/aom/aom_dsp/arm/aom_convolve8_avg_neon_asm.asm
new file mode 100644
index 0000000000..80aef992dd
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/aom_convolve8_avg_neon_asm.asm
@@ -0,0 +1,295 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    ; These functions are only valid when:
+    ; x_step_q4 == 16
+    ; w%4 == 0
+    ; h%4 == 0
+    ; taps == 8
+    ; AV1_FILTER_WEIGHT == 128
+    ; AV1_FILTER_SHIFT == 7
+
+    EXPORT  |aom_convolve8_avg_horiz_neon|
+    EXPORT  |aom_convolve8_avg_vert_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Multiply and accumulate by q0
+    MACRO
+    MULTIPLY_BY_Q0 $dst, $src0, $src1, $src2, $src3, $src4, $src5, $src6, $src7
+    vmull.s16 $dst, $src0, d0[0]
+    vmlal.s16 $dst, $src1, d0[1]
+    vmlal.s16 $dst, $src2, d0[2]
+    vmlal.s16 $dst, $src3, d0[3]
+    vmlal.s16 $dst, $src4, d1[0]
+    vmlal.s16 $dst, $src5, d1[1]
+    vmlal.s16 $dst, $src6, d1[2]
+    vmlal.s16 $dst, $src7, d1[3]
+    MEND
+
+; r0    const uint8_t *src
+; r1    int src_stride
+; r2    uint8_t *dst
+; r3    int dst_stride
+; sp[]const int16_t *filter_x
+; sp[]int x_step_q4
+; sp[]const int16_t *filter_y ; unused
+; sp[]int y_step_q4           ; unused
+; sp[]int w
+; sp[]int h
+
+|aom_convolve8_avg_horiz_neon| PROC
+    push            {r4-r10, lr}
+
+    sub             r0, r0, #3              ; adjust for taps
+
+    ldr             r5, [sp, #32]           ; filter_x
+    ldr             r6, [sp, #48]           ; w
+    ldr             r7, [sp, #52]           ; h
+
+    vld1.s16        {q0}, [r5]              ; filter_x
+
+    sub             r8, r1, r1, lsl #2      ; -src_stride * 3
+    add             r8, r8, #4              ; -src_stride * 3 + 4
+
+    sub             r4, r3, r3, lsl #2      ; -dst_stride * 3
+    add             r4, r4, #4              ; -dst_stride * 3 + 4
+
+    rsb             r9, r6, r1, lsl #2      ; reset src for outer loop
+    sub             r9, r9, #7
+    rsb             r12, r6, r3, lsl #2     ; reset dst for outer loop
+
+    mov             r10, r6                 ; w loop counter
+
+aom_convolve8_avg_loop_horiz_v
+    vld1.8          {d24}, [r0], r1
+    vld1.8          {d25}, [r0], r1
+    vld1.8          {d26}, [r0], r1
+    vld1.8          {d27}, [r0], r8
+
+    vtrn.16         q12, q13
+    vtrn.8          d24, d25
+    vtrn.8          d26, d27
+
+    pld             [r0, r1, lsl #2]
+
+    vmovl.u8        q8, d24
+    vmovl.u8        q9, d25
+    vmovl.u8        q10, d26
+    vmovl.u8        q11, d27
+
+    ; save a few instructions in the inner loop
+    vswp            d17, d18
+    vmov            d23, d21
+
+    add             r0, r0, #3
+
+aom_convolve8_avg_loop_horiz
+    add             r5, r0, #64
+
+    vld1.32         {d28[]}, [r0], r1
+    vld1.32         {d29[]}, [r0], r1
+    vld1.32         {d31[]}, [r0], r1
+    vld1.32         {d30[]}, [r0], r8
+
+    pld             [r5]
+
+    vtrn.16         d28, d31
+    vtrn.16         d29, d30
+    vtrn.8          d28, d29
+    vtrn.8          d31, d30
+
+    pld             [r5, r1]
+
+    ; extract to s16
+    vtrn.32         q14, q15
+    vmovl.u8        q12, d28
+    vmovl.u8        q13, d29
+
+    pld             [r5, r1, lsl #1]
+
+    ; slightly out of order load to match the existing data
+    vld1.u32        {d6[0]}, [r2], r3
+    vld1.u32        {d7[0]}, [r2], r3
+    vld1.u32        {d6[1]}, [r2], r3
+    vld1.u32        {d7[1]}, [r2], r3
+
+    sub             r2, r2, r3, lsl #2      ; reset for store
+
+    ; src[] * filter_x
+    MULTIPLY_BY_Q0  q1,  d16, d17, d20, d22, d18, d19, d23, d24
+    MULTIPLY_BY_Q0  q2,  d17, d20, d22, d18, d19, d23, d24, d26
+    MULTIPLY_BY_Q0  q14, d20, d22, d18, d19, d23, d24, d26, d27
+    MULTIPLY_BY_Q0  q15, d22, d18, d19, d23, d24, d26, d27, d25
+
+    pld             [r5, -r8]
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; transpose
+    vtrn.16         d2, d3
+    vtrn.32         d2, d3
+    vtrn.8          d2, d3
+
+    ; average the new value and the dst value
+    vrhadd.u8       q1, q1, q3
+
+    vst1.u32        {d2[0]}, [r2@32], r3
+    vst1.u32        {d3[0]}, [r2@32], r3
+    vst1.u32        {d2[1]}, [r2@32], r3
+    vst1.u32        {d3[1]}, [r2@32], r4
+
+    vmov            q8,  q9
+    vmov            d20, d23
+    vmov            q11, q12
+    vmov            q9,  q13
+
+    subs            r6, r6, #4              ; w -= 4
+    bgt             aom_convolve8_avg_loop_horiz
+
+    ; outer loop
+    mov             r6, r10                 ; restore w counter
+    add             r0, r0, r9              ; src += src_stride * 4 - w
+    add             r2, r2, r12             ; dst += dst_stride * 4 - w
+    subs            r7, r7, #4              ; h -= 4
+    bgt aom_convolve8_avg_loop_horiz_v
+
+    pop             {r4-r10, pc}
+
+    ENDP
+
+|aom_convolve8_avg_vert_neon| PROC
+    push            {r4-r8, lr}
+
+    ; adjust for taps
+    sub             r0, r0, r1
+    sub             r0, r0, r1, lsl #1
+
+    ldr             r4, [sp, #32]           ; filter_y
+    ldr             r6, [sp, #40]           ; w
+    ldr             lr, [sp, #44]           ; h
+
+    vld1.s16        {q0}, [r4]              ; filter_y
+
+    lsl             r1, r1, #1
+    lsl             r3, r3, #1
+
+aom_convolve8_avg_loop_vert_h
+    mov             r4, r0
+    add             r7, r0, r1, asr #1
+    mov             r5, r2
+    add             r8, r2, r3, asr #1
+    mov             r12, lr                 ; h loop counter
+
+    vld1.u32        {d16[0]}, [r4], r1
+    vld1.u32        {d16[1]}, [r7], r1
+    vld1.u32        {d18[0]}, [r4], r1
+    vld1.u32        {d18[1]}, [r7], r1
+    vld1.u32        {d20[0]}, [r4], r1
+    vld1.u32        {d20[1]}, [r7], r1
+    vld1.u32        {d22[0]}, [r4], r1
+
+    vmovl.u8        q8, d16
+    vmovl.u8        q9, d18
+    vmovl.u8        q10, d20
+    vmovl.u8        q11, d22
+
+aom_convolve8_avg_loop_vert
+    ; always process a 4x4 block at a time
+    vld1.u32        {d24[0]}, [r7], r1
+    vld1.u32        {d26[0]}, [r4], r1
+    vld1.u32        {d26[1]}, [r7], r1
+    vld1.u32        {d24[1]}, [r4], r1
+
+    ; extract to s16
+    vmovl.u8        q12, d24
+    vmovl.u8        q13, d26
+
+    vld1.u32        {d6[0]}, [r5@32], r3
+    vld1.u32        {d6[1]}, [r8@32], r3
+    vld1.u32        {d7[0]}, [r5@32], r3
+    vld1.u32        {d7[1]}, [r8@32], r3
+
+    pld             [r7]
+    pld             [r4]
+
+    ; src[] * filter_y
+    MULTIPLY_BY_Q0  q1,  d16, d17, d18, d19, d20, d21, d22, d24
+
+    pld             [r7, r1]
+    pld             [r4, r1]
+
+    MULTIPLY_BY_Q0  q2,  d17, d18, d19, d20, d21, d22, d24, d26
+
+    pld             [r5]
+    pld             [r8]
+
+    MULTIPLY_BY_Q0  q14, d18, d19, d20, d21, d22, d24, d26, d27
+
+    pld             [r5, r3]
+    pld             [r8, r3]
+
+    MULTIPLY_BY_Q0  q15, d19, d20, d21, d22, d24, d26, d27, d25
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; average the new value and the dst value
+    vrhadd.u8       q1, q1, q3
+
+    sub             r5, r5, r3, lsl #1      ; reset for store
+    sub             r8, r8, r3, lsl #1
+
+    vst1.u32        {d2[0]}, [r5@32], r3
+    vst1.u32        {d2[1]}, [r8@32], r3
+    vst1.u32        {d3[0]}, [r5@32], r3
+    vst1.u32        {d3[1]}, [r8@32], r3
+
+    vmov            q8, q10
+    vmov            d18, d22
+    vmov            d19, d24
+    vmov            q10, q13
+    vmov            d22, d25
+
+    subs            r12, r12, #4            ; h -= 4
+    bgt             aom_convolve8_avg_loop_vert
+
+    ; outer loop
+    add             r0, r0, #4
+    add             r2, r2, #4
+    subs            r6, r6, #4              ; w -= 4
+    bgt             aom_convolve8_avg_loop_vert_h
+
+    pop             {r4-r8, pc}
+
+    ENDP
+    END
diff --git a/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c b/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c
new file mode 100644
index 0000000000..8ebffb5f9a
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/aom_convolve8_neon.c
@@ -0,0 +1,331 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+static INLINE int32x4_t MULTIPLY_BY_Q0(int16x4_t dsrc0, int16x4_t dsrc1,
+                                       int16x4_t dsrc2, int16x4_t dsrc3,
+                                       int16x4_t dsrc4, int16x4_t dsrc5,
+                                       int16x4_t dsrc6, int16x4_t dsrc7,
+                                       int16x8_t q0s16) {
+  int32x4_t qdst;
+  int16x4_t d0s16, d1s16;
+
+  d0s16 = vget_low_s16(q0s16);
+  d1s16 = vget_high_s16(q0s16);
+
+  qdst = vmull_lane_s16(dsrc0, d0s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc1, d0s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc2, d0s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc3, d0s16, 3);
+  qdst = vmlal_lane_s16(qdst, dsrc4, d1s16, 0);
+  qdst = vmlal_lane_s16(qdst, dsrc5, d1s16, 1);
+  qdst = vmlal_lane_s16(qdst, dsrc6, d1s16, 2);
+  qdst = vmlal_lane_s16(qdst, dsrc7, d1s16, 3);
+  return qdst;
+}
+
+void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y,  // unused
+                              int y_step_q4,            // unused
+                              int w, int h) {
+  int width;
+  const uint8_t *s, *psrc;
+  uint8_t *d, *pdst;
+  uint8x8_t d2u8, d3u8, d24u8, d25u8, d26u8, d27u8, d28u8, d29u8;
+  uint32x2_t d2u32, d3u32, d28u32, d29u32, d30u32, d31u32;
+  uint8x16_t q12u8, q13u8, q14u8, q15u8;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16, d16u16, d17u16, d18u16, d19u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+  uint16x8x2_t q0x2u16;
+  uint8x8x2_t d0x2u8, d1x2u8;
+  uint32x2x2_t d0x2u32;
+  uint16x4x2_t d0x2u16, d1x2u16;
+  uint32x4x2_t q0x2u32;
+
+  assert(x_step_q4 == 16);
+
+  (void)x_step_q4;
+  (void)y_step_q4;
+  (void)filter_y;
+
+  q0s16 = vld1q_s16(filter_x);
+
+  src -= 3;  // adjust for taps
+  for (; h > 0; h -= 4, src += src_stride * 4,
+                dst += dst_stride * 4) {  // loop_horiz_v
+    s = src;
+    d24u8 = vld1_u8(s);
+    s += src_stride;
+    d25u8 = vld1_u8(s);
+    s += src_stride;
+    d26u8 = vld1_u8(s);
+    s += src_stride;
+    d27u8 = vld1_u8(s);
+
+    q12u8 = vcombine_u8(d24u8, d25u8);
+    q13u8 = vcombine_u8(d26u8, d27u8);
+
+    q0x2u16 =
+        vtrnq_u16(vreinterpretq_u16_u8(q12u8), vreinterpretq_u16_u8(q13u8));
+    d24u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[0]));
+    d25u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[0]));
+    d26u8 = vreinterpret_u8_u16(vget_low_u16(q0x2u16.val[1]));
+    d27u8 = vreinterpret_u8_u16(vget_high_u16(q0x2u16.val[1]));
+    d0x2u8 = vtrn_u8(d24u8, d25u8);
+    d1x2u8 = vtrn_u8(d26u8, d27u8);
+
+    __builtin_prefetch(src + src_stride * 4);
+    __builtin_prefetch(src + src_stride * 5);
+    __builtin_prefetch(src + src_stride * 6);
+
+    q8u16 = vmovl_u8(d0x2u8.val[0]);
+    q9u16 = vmovl_u8(d0x2u8.val[1]);
+    q10u16 = vmovl_u8(d1x2u8.val[0]);
+    q11u16 = vmovl_u8(d1x2u8.val[1]);
+
+    d16u16 = vget_low_u16(q8u16);
+    d17u16 = vget_high_u16(q8u16);
+    d18u16 = vget_low_u16(q9u16);
+    d19u16 = vget_high_u16(q9u16);
+    q8u16 = vcombine_u16(d16u16, d18u16);  // vswp 17 18
+    q9u16 = vcombine_u16(d17u16, d19u16);
+
+    d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));  // vmov 23 21
+    for (width = w, psrc = src + 7, pdst = dst; width > 0;
+         width -= 4, psrc += 4, pdst += 4) {  // loop_horiz
+      s = psrc;
+      d28u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d29u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d31u32 = vld1_dup_u32((const uint32_t *)s);
+      s += src_stride;
+      d30u32 = vld1_dup_u32((const uint32_t *)s);
+
+      __builtin_prefetch(psrc + 64);
+
+      d0x2u16 =
+          vtrn_u16(vreinterpret_u16_u32(d28u32), vreinterpret_u16_u32(d31u32));
+      d1x2u16 =
+          vtrn_u16(vreinterpret_u16_u32(d29u32), vreinterpret_u16_u32(d30u32));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[0]),   // d28
+                       vreinterpret_u8_u16(d1x2u16.val[0]));  // d29
+      d1x2u8 = vtrn_u8(vreinterpret_u8_u16(d0x2u16.val[1]),   // d31
+                       vreinterpret_u8_u16(d1x2u16.val[1]));  // d30
+
+      __builtin_prefetch(psrc + 64 + src_stride);
+
+      q14u8 = vcombine_u8(d0x2u8.val[0], d0x2u8.val[1]);
+      q15u8 = vcombine_u8(d1x2u8.val[1], d1x2u8.val[0]);
+      q0x2u32 =
+          vtrnq_u32(vreinterpretq_u32_u8(q14u8), vreinterpretq_u32_u8(q15u8));
+
+      d28u8 = vreinterpret_u8_u32(vget_low_u32(q0x2u32.val[0]));
+      d29u8 = vreinterpret_u8_u32(vget_high_u32(q0x2u32.val[0]));
+      q12u16 = vmovl_u8(d28u8);
+      q13u16 = vmovl_u8(d29u8);
+
+      __builtin_prefetch(psrc + 64 + src_stride * 2);
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+      d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+      d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d20s16, d22s16, d18s16, d19s16,
+                             d23s16, d24s16, q0s16);
+      q2s32 = MULTIPLY_BY_Q0(d17s16, d20s16, d22s16, d18s16, d19s16, d23s16,
+                             d24s16, d26s16, q0s16);
+      q14s32 = MULTIPLY_BY_Q0(d20s16, d22s16, d18s16, d19s16, d23s16, d24s16,
+                              d26s16, d27s16, q0s16);
+      q15s32 = MULTIPLY_BY_Q0(d22s16, d18s16, d19s16, d23s16, d24s16, d26s16,
+                              d27s16, d25s16, q0s16);
+
+      __builtin_prefetch(psrc + 60 + src_stride * 3);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u8 = vqmovn_u16(q1u16);
+      d3u8 = vqmovn_u16(q2u16);
+
+      d0x2u16 = vtrn_u16(vreinterpret_u16_u8(d2u8), vreinterpret_u16_u8(d3u8));
+      d0x2u32 = vtrn_u32(vreinterpret_u32_u16(d0x2u16.val[0]),
+                         vreinterpret_u32_u16(d0x2u16.val[1]));
+      d0x2u8 = vtrn_u8(vreinterpret_u8_u32(d0x2u32.val[0]),
+                       vreinterpret_u8_u32(d0x2u32.val[1]));
+
+      d2u32 = vreinterpret_u32_u8(d0x2u8.val[0]);
+      d3u32 = vreinterpret_u32_u8(d0x2u8.val[1]);
+
+      d = pdst;
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+
+      q8u16 = q9u16;
+      d20s16 = d23s16;
+      q11u16 = q12u16;
+      q9u16 = q13u16;
+      d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    }
+  }
+  return;
+}
+
+void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x,  // unused
+                             int x_step_q4,            // unused
+                             const int16_t *filter_y, int y_step_q4, int w,
+                             int h) {
+  int height;
+  const uint8_t *s;
+  uint8_t *d;
+  uint32x2_t d2u32, d3u32;
+  uint32x2_t d16u32, d18u32, d20u32, d22u32, d24u32, d26u32;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16;
+  uint16x4_t d2u16, d3u16, d4u16, d5u16;
+  int16x8_t q0s16;
+  uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
+  int32x4_t q1s32, q2s32, q14s32, q15s32;
+
+  assert(y_step_q4 == 16);
+
+  (void)x_step_q4;
+  (void)y_step_q4;
+  (void)filter_x;
+
+  src -= src_stride * 3;
+  q0s16 = vld1q_s16(filter_y);
+  for (; w > 0; w -= 4, src += 4, dst += 4) {  // loop_vert_h
+    s = src;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 0);
+    s += src_stride;
+    d16u32 = vld1_lane_u32((const uint32_t *)s, d16u32, 1);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 0);
+    s += src_stride;
+    d18u32 = vld1_lane_u32((const uint32_t *)s, d18u32, 1);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 0);
+    s += src_stride;
+    d20u32 = vld1_lane_u32((const uint32_t *)s, d20u32, 1);
+    s += src_stride;
+    d22u32 = vld1_lane_u32((const uint32_t *)s, d22u32, 0);
+    s += src_stride;
+
+    q8u16 = vmovl_u8(vreinterpret_u8_u32(d16u32));
+    q9u16 = vmovl_u8(vreinterpret_u8_u32(d18u32));
+    q10u16 = vmovl_u8(vreinterpret_u8_u32(d20u32));
+    q11u16 = vmovl_u8(vreinterpret_u8_u32(d22u32));
+
+    d18s16 = vreinterpret_s16_u16(vget_low_u16(q9u16));
+    d19s16 = vreinterpret_s16_u16(vget_high_u16(q9u16));
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d = dst;
+    for (height = h; height > 0; height -= 4) {  // loop_vert
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 0);
+      s += src_stride;
+      d26u32 = vld1_lane_u32((const uint32_t *)s, d26u32, 1);
+      s += src_stride;
+      d24u32 = vld1_lane_u32((const uint32_t *)s, d24u32, 1);
+      s += src_stride;
+
+      q12u16 = vmovl_u8(vreinterpret_u8_u32(d24u32));
+      q13u16 = vmovl_u8(vreinterpret_u8_u32(d26u32));
+
+      d16s16 = vreinterpret_s16_u16(vget_low_u16(q8u16));
+      d17s16 = vreinterpret_s16_u16(vget_high_u16(q8u16));
+      d20s16 = vreinterpret_s16_u16(vget_low_u16(q10u16));
+      d21s16 = vreinterpret_s16_u16(vget_high_u16(q10u16));
+      d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+      d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+      d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+      d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+
+      __builtin_prefetch(d);
+      __builtin_prefetch(d + dst_stride);
+      q1s32 = MULTIPLY_BY_Q0(d16s16, d17s16, d18s16, d19s16, d20s16, d21s16,
+                             d22s16, d24s16, q0s16);
+      __builtin_prefetch(d + dst_stride * 2);
+      __builtin_prefetch(d + dst_stride * 3);
+      q2s32 = MULTIPLY_BY_Q0(d17s16, d18s16, d19s16, d20s16, d21s16, d22s16,
+                             d24s16, d26s16, q0s16);
+      __builtin_prefetch(s);
+      __builtin_prefetch(s + src_stride);
+      q14s32 = MULTIPLY_BY_Q0(d18s16, d19s16, d20s16, d21s16, d22s16, d24s16,
+                              d26s16, d27s16, q0s16);
+      __builtin_prefetch(s + src_stride * 2);
+      __builtin_prefetch(s + src_stride * 3);
+      q15s32 = MULTIPLY_BY_Q0(d19s16, d20s16, d21s16, d22s16, d24s16, d26s16,
+                              d27s16, d25s16, q0s16);
+
+      d2u16 = vqrshrun_n_s32(q1s32, 7);
+      d3u16 = vqrshrun_n_s32(q2s32, 7);
+      d4u16 = vqrshrun_n_s32(q14s32, 7);
+      d5u16 = vqrshrun_n_s32(q15s32, 7);
+
+      q1u16 = vcombine_u16(d2u16, d3u16);
+      q2u16 = vcombine_u16(d4u16, d5u16);
+
+      d2u32 = vreinterpret_u32_u8(vqmovn_u16(q1u16));
+      d3u32 = vreinterpret_u32_u8(vqmovn_u16(q2u16));
+
+      vst1_lane_u32((uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 0);
+      d += dst_stride;
+      vst1_lane_u32((uint32_t *)d, d3u32, 1);
+      d += dst_stride;
+
+      q8u16 = q10u16;
+      d18s16 = d22s16;
+      d19s16 = d24s16;
+      q10u16 = q13u16;
+      d22s16 = d25s16;
+    }
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/aom_convolve8_neon_asm.asm b/third_party/aom/aom_dsp/arm/aom_convolve8_neon_asm.asm
new file mode 100644
index 0000000000..38207d8649
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/aom_convolve8_neon_asm.asm
@@ -0,0 +1,273 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    ; These functions are only valid when:
+    ; x_step_q4 == 16
+    ; w%4 == 0
+    ; h%4 == 0
+    ; taps == 8
+    ; AV1_FILTER_WEIGHT == 128
+    ; AV1_FILTER_SHIFT == 7
+
+    EXPORT  |aom_convolve8_horiz_neon|
+    EXPORT  |aom_convolve8_vert_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Multiply and accumulate by q0
+    MACRO
+    MULTIPLY_BY_Q0 $dst, $src0, $src1, $src2, $src3, $src4, $src5, $src6, $src7
+    vmull.s16 $dst, $src0, d0[0]
+    vmlal.s16 $dst, $src1, d0[1]
+    vmlal.s16 $dst, $src2, d0[2]
+    vmlal.s16 $dst, $src3, d0[3]
+    vmlal.s16 $dst, $src4, d1[0]
+    vmlal.s16 $dst, $src5, d1[1]
+    vmlal.s16 $dst, $src6, d1[2]
+    vmlal.s16 $dst, $src7, d1[3]
+    MEND
+
+; r0    const uint8_t *src
+; r1    int src_stride
+; r2    uint8_t *dst
+; r3    int dst_stride
+; sp[]const int16_t *filter_x
+; sp[]int x_step_q4
+; sp[]const int16_t *filter_y ; unused
+; sp[]int y_step_q4           ; unused
+; sp[]int w
+; sp[]int h
+
+|aom_convolve8_horiz_neon| PROC
+    push            {r4-r10, lr}
+
+    sub             r0, r0, #3              ; adjust for taps
+
+    ldr             r5, [sp, #32]           ; filter_x
+    ldr             r6, [sp, #48]           ; w
+    ldr             r7, [sp, #52]           ; h
+
+    vld1.s16        {q0}, [r5]              ; filter_x
+
+    sub             r8, r1, r1, lsl #2      ; -src_stride * 3
+    add             r8, r8, #4              ; -src_stride * 3 + 4
+
+    sub             r4, r3, r3, lsl #2      ; -dst_stride * 3
+    add             r4, r4, #4              ; -dst_stride * 3 + 4
+
+    rsb             r9, r6, r1, lsl #2      ; reset src for outer loop
+    sub             r9, r9, #7
+    rsb             r12, r6, r3, lsl #2     ; reset dst for outer loop
+
+    mov             r10, r6                 ; w loop counter
+
+aom_convolve8_loop_horiz_v
+    vld1.8          {d24}, [r0], r1
+    vld1.8          {d25}, [r0], r1
+    vld1.8          {d26}, [r0], r1
+    vld1.8          {d27}, [r0], r8
+
+    vtrn.16         q12, q13
+    vtrn.8          d24, d25
+    vtrn.8          d26, d27
+
+    pld             [r0, r1, lsl #2]
+
+    vmovl.u8        q8, d24
+    vmovl.u8        q9, d25
+    vmovl.u8        q10, d26
+    vmovl.u8        q11, d27
+
+    ; save a few instructions in the inner loop
+    vswp            d17, d18
+    vmov            d23, d21
+
+    add             r0, r0, #3
+
+aom_convolve8_loop_horiz
+    add             r5, r0, #64
+
+    vld1.32         {d28[]}, [r0], r1
+    vld1.32         {d29[]}, [r0], r1
+    vld1.32         {d31[]}, [r0], r1
+    vld1.32         {d30[]}, [r0], r8
+
+    pld             [r5]
+
+    vtrn.16         d28, d31
+    vtrn.16         d29, d30
+    vtrn.8          d28, d29
+    vtrn.8          d31, d30
+
+    pld             [r5, r1]
+
+    ; extract to s16
+    vtrn.32         q14, q15
+    vmovl.u8        q12, d28
+    vmovl.u8        q13, d29
+
+    pld             [r5, r1, lsl #1]
+
+    ; src[] * filter_x
+    MULTIPLY_BY_Q0  q1,  d16, d17, d20, d22, d18, d19, d23, d24
+    MULTIPLY_BY_Q0  q2,  d17, d20, d22, d18, d19, d23, d24, d26
+    MULTIPLY_BY_Q0  q14, d20, d22, d18, d19, d23, d24, d26, d27
+    MULTIPLY_BY_Q0  q15, d22, d18, d19, d23, d24, d26, d27, d25
+
+    pld             [r5, -r8]
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    ; transpose
+    vtrn.16         d2, d3
+    vtrn.32         d2, d3
+    vtrn.8          d2, d3
+
+    vst1.u32        {d2[0]}, [r2@32], r3
+    vst1.u32        {d3[0]}, [r2@32], r3
+    vst1.u32        {d2[1]}, [r2@32], r3
+    vst1.u32        {d3[1]}, [r2@32], r4
+
+    vmov            q8,  q9
+    vmov            d20, d23
+    vmov            q11, q12
+    vmov            q9,  q13
+
+    subs            r6, r6, #4              ; w -= 4
+    bgt             aom_convolve8_loop_horiz
+
+    ; outer loop
+    mov             r6, r10                 ; restore w counter
+    add             r0, r0, r9              ; src += src_stride * 4 - w
+    add             r2, r2, r12             ; dst += dst_stride * 4 - w
+    subs            r7, r7, #4              ; h -= 4
+    bgt aom_convolve8_loop_horiz_v
+
+    pop             {r4-r10, pc}
+
+    ENDP
+
+|aom_convolve8_vert_neon| PROC
+    push            {r4-r8, lr}
+
+    ; adjust for taps
+    sub             r0, r0, r1
+    sub             r0, r0, r1, lsl #1
+
+    ldr             r4, [sp, #32]           ; filter_y
+    ldr             r6, [sp, #40]           ; w
+    ldr             lr, [sp, #44]           ; h
+
+    vld1.s16        {q0}, [r4]              ; filter_y
+
+    lsl             r1, r1, #1
+    lsl             r3, r3, #1
+
+aom_convolve8_loop_vert_h
+    mov             r4, r0
+    add             r7, r0, r1, asr #1
+    mov             r5, r2
+    add             r8, r2, r3, asr #1
+    mov             r12, lr                 ; h loop counter
+
+    vld1.u32        {d16[0]}, [r4], r1
+    vld1.u32        {d16[1]}, [r7], r1
+    vld1.u32        {d18[0]}, [r4], r1
+    vld1.u32        {d18[1]}, [r7], r1
+    vld1.u32        {d20[0]}, [r4], r1
+    vld1.u32        {d20[1]}, [r7], r1
+    vld1.u32        {d22[0]}, [r4], r1
+
+    vmovl.u8        q8, d16
+    vmovl.u8        q9, d18
+    vmovl.u8        q10, d20
+    vmovl.u8        q11, d22
+
+aom_convolve8_loop_vert
+    ; always process a 4x4 block at a time
+    vld1.u32        {d24[0]}, [r7], r1
+    vld1.u32        {d26[0]}, [r4], r1
+    vld1.u32        {d26[1]}, [r7], r1
+    vld1.u32        {d24[1]}, [r4], r1
+
+    ; extract to s16
+    vmovl.u8        q12, d24
+    vmovl.u8        q13, d26
+
+    pld             [r5]
+    pld             [r8]
+
+    ; src[] * filter_y
+    MULTIPLY_BY_Q0  q1,  d16, d17, d18, d19, d20, d21, d22, d24
+
+    pld             [r5, r3]
+    pld             [r8, r3]
+
+    MULTIPLY_BY_Q0  q2,  d17, d18, d19, d20, d21, d22, d24, d26
+
+    pld             [r7]
+    pld             [r4]
+
+    MULTIPLY_BY_Q0  q14, d18, d19, d20, d21, d22, d24, d26, d27
+
+    pld             [r7, r1]
+    pld             [r4, r1]
+
+    MULTIPLY_BY_Q0  q15, d19, d20, d21, d22, d24, d26, d27, d25
+
+    ; += 64 >> 7
+    vqrshrun.s32    d2, q1, #7
+    vqrshrun.s32    d3, q2, #7
+    vqrshrun.s32    d4, q14, #7
+    vqrshrun.s32    d5, q15, #7
+
+    ; saturate
+    vqmovn.u16      d2, q1
+    vqmovn.u16      d3, q2
+
+    vst1.u32        {d2[0]}, [r5@32], r3
+    vst1.u32        {d2[1]}, [r8@32], r3
+    vst1.u32        {d3[0]}, [r5@32], r3
+    vst1.u32        {d3[1]}, [r8@32], r3
+
+    vmov            q8, q10
+    vmov            d18, d22
+    vmov            d19, d24
+    vmov            q10, q13
+    vmov            d22, d25
+
+    subs            r12, r12, #4            ; h -= 4
+    bgt             aom_convolve8_loop_vert
+
+    ; outer loop
+    add             r0, r0, #4
+    add             r2, r2, #4
+    subs            r6, r6, #4              ; w -= 4
+    bgt             aom_convolve8_loop_vert_h
+
+    pop             {r4-r8, pc}
+
+    ENDP
+    END
diff --git a/third_party/aom/aom_dsp/arm/aom_convolve_avg_neon.c b/third_party/aom/aom_dsp/arm/aom_convolve_avg_neon.c
new file mode 100644
index 0000000000..f05d3ceae2
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/aom_convolve_avg_neon.c
@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+void aom_convolve_avg_neon(const uint8_t *src,    // r0
+                           ptrdiff_t src_stride,  // r1
+                           uint8_t *dst,          // r2
+                           ptrdiff_t dst_stride,  // r3
+                           const int16_t *filter_x, int filter_x_stride,
+                           const int16_t *filter_y, int filter_y_stride, int w,
+                           int h) {
+  uint8_t *d;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8;
+  uint32x2_t d0u32, d2u32;
+  uint8x16_t q0u8, q1u8, q2u8, q3u8, q8u8, q9u8, q10u8, q11u8;
+  (void)filter_x;
+  (void)filter_x_stride;
+  (void)filter_y;
+  (void)filter_y_stride;
+
+  d = dst;
+  if (w > 32) {  // avg64
+    for (; h > 0; h -= 1) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      q2u8 = vld1q_u8(src + 32);
+      q3u8 = vld1q_u8(src + 48);
+      src += src_stride;
+      q8u8 = vld1q_u8(d);
+      q9u8 = vld1q_u8(d + 16);
+      q10u8 = vld1q_u8(d + 32);
+      q11u8 = vld1q_u8(d + 48);
+      d += dst_stride;
+
+      q0u8 = vrhaddq_u8(q0u8, q8u8);
+      q1u8 = vrhaddq_u8(q1u8, q9u8);
+      q2u8 = vrhaddq_u8(q2u8, q10u8);
+      q3u8 = vrhaddq_u8(q3u8, q11u8);
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      vst1q_u8(dst + 32, q2u8);
+      vst1q_u8(dst + 48, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w == 32) {  // avg32
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      src += src_stride;
+      q2u8 = vld1q_u8(src);
+      q3u8 = vld1q_u8(src + 16);
+      src += src_stride;
+      q8u8 = vld1q_u8(d);
+      q9u8 = vld1q_u8(d + 16);
+      d += dst_stride;
+      q10u8 = vld1q_u8(d);
+      q11u8 = vld1q_u8(d + 16);
+      d += dst_stride;
+
+      q0u8 = vrhaddq_u8(q0u8, q8u8);
+      q1u8 = vrhaddq_u8(q1u8, q9u8);
+      q2u8 = vrhaddq_u8(q2u8, q10u8);
+      q3u8 = vrhaddq_u8(q3u8, q11u8);
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q2u8);
+      vst1q_u8(dst + 16, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w > 8) {  // avg16
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      src += src_stride;
+      q1u8 = vld1q_u8(src);
+      src += src_stride;
+      q2u8 = vld1q_u8(d);
+      d += dst_stride;
+      q3u8 = vld1q_u8(d);
+      d += dst_stride;
+
+      q0u8 = vrhaddq_u8(q0u8, q2u8);
+      q1u8 = vrhaddq_u8(q1u8, q3u8);
+
+      vst1q_u8(dst, q0u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q1u8);
+      dst += dst_stride;
+    }
+  } else if (w == 8) {  // avg8
+    for (; h > 0; h -= 2) {
+      d0u8 = vld1_u8(src);
+      src += src_stride;
+      d1u8 = vld1_u8(src);
+      src += src_stride;
+      d2u8 = vld1_u8(d);
+      d += dst_stride;
+      d3u8 = vld1_u8(d);
+      d += dst_stride;
+
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      q0u8 = vrhaddq_u8(q0u8, q1u8);
+
+      vst1_u8(dst, vget_low_u8(q0u8));
+      dst += dst_stride;
+      vst1_u8(dst, vget_high_u8(q0u8));
+      dst += dst_stride;
+    }
+  } else {  // avg4
+    for (; h > 0; h -= 2) {
+      d0u32 = vld1_lane_u32((const uint32_t *)src, d0u32, 0);
+      src += src_stride;
+      d0u32 = vld1_lane_u32((const uint32_t *)src, d0u32, 1);
+      src += src_stride;
+      d2u32 = vld1_lane_u32((const uint32_t *)d, d2u32, 0);
+      d += dst_stride;
+      d2u32 = vld1_lane_u32((const uint32_t *)d, d2u32, 1);
+      d += dst_stride;
+
+      d0u8 = vrhadd_u8(vreinterpret_u8_u32(d0u32), vreinterpret_u8_u32(d2u32));
+
+      d0u32 = vreinterpret_u32_u8(d0u8);
+      vst1_lane_u32((uint32_t *)dst, d0u32, 0);
+      dst += dst_stride;
+      vst1_lane_u32((uint32_t *)dst, d0u32, 1);
+      dst += dst_stride;
+    }
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/aom_convolve_avg_neon_asm.asm b/third_party/aom/aom_dsp/arm/aom_convolve_avg_neon_asm.asm
new file mode 100644
index 0000000000..43c300954f
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/aom_convolve_avg_neon_asm.asm
@@ -0,0 +1,119 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_convolve_avg_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|aom_convolve_avg_neon| PROC
+    push                {r4-r6, lr}
+    ldrd                r4, r5, [sp, #32]
+    mov                 r6, r2
+
+    cmp                 r4, #32
+    bgt                 avg64
+    beq                 avg32
+    cmp                 r4, #8
+    bgt                 avg16
+    beq                 avg8
+    b                   avg4
+
+avg64
+    sub                 lr, r1, #32
+    sub                 r4, r3, #32
+avg64_h
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0]!
+    vld1.8              {q2-q3}, [r0], lr
+    pld                 [r2, r3]
+    vld1.8              {q8-q9},   [r6@128]!
+    vld1.8              {q10-q11}, [r6@128], r4
+    vrhadd.u8           q0, q0, q8
+    vrhadd.u8           q1, q1, q9
+    vrhadd.u8           q2, q2, q10
+    vrhadd.u8           q3, q3, q11
+    vst1.8              {q0-q1}, [r2@128]!
+    vst1.8              {q2-q3}, [r2@128], r4
+    subs                r5, r5, #1
+    bgt                 avg64_h
+    pop                 {r4-r6, pc}
+
+avg32
+    vld1.8              {q0-q1}, [r0], r1
+    vld1.8              {q2-q3}, [r0], r1
+    vld1.8              {q8-q9},   [r6@128], r3
+    vld1.8              {q10-q11}, [r6@128], r3
+    pld                 [r0]
+    vrhadd.u8           q0, q0, q8
+    pld                 [r0, r1]
+    vrhadd.u8           q1, q1, q9
+    pld                 [r6]
+    vrhadd.u8           q2, q2, q10
+    pld                 [r6, r3]
+    vrhadd.u8           q3, q3, q11
+    vst1.8              {q0-q1}, [r2@128], r3
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 avg32
+    pop                 {r4-r6, pc}
+
+avg16
+    vld1.8              {q0}, [r0], r1
+    vld1.8              {q1}, [r0], r1
+    vld1.8              {q2}, [r6@128], r3
+    vld1.8              {q3}, [r6@128], r3
+    pld                 [r0]
+    pld                 [r0, r1]
+    vrhadd.u8           q0, q0, q2
+    pld                 [r6]
+    pld                 [r6, r3]
+    vrhadd.u8           q1, q1, q3
+    vst1.8              {q0}, [r2@128], r3
+    vst1.8              {q1}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 avg16
+    pop                 {r4-r6, pc}
+
+avg8
+    vld1.8              {d0}, [r0], r1
+    vld1.8              {d1}, [r0], r1
+    vld1.8              {d2}, [r6@64], r3
+    vld1.8              {d3}, [r6@64], r3
+    pld                 [r0]
+    pld                 [r0, r1]
+    vrhadd.u8           q0, q0, q1
+    pld                 [r6]
+    pld                 [r6, r3]
+    vst1.8              {d0}, [r2@64], r3
+    vst1.8              {d1}, [r2@64], r3
+    subs                r5, r5, #2
+    bgt                 avg8
+    pop                 {r4-r6, pc}
+
+avg4
+    vld1.32             {d0[0]}, [r0], r1
+    vld1.32             {d0[1]}, [r0], r1
+    vld1.32             {d2[0]}, [r6@32], r3
+    vld1.32             {d2[1]}, [r6@32], r3
+    vrhadd.u8           d0, d0, d2
+    vst1.32             {d0[0]}, [r2@32], r3
+    vst1.32             {d0[1]}, [r2@32], r3
+    subs                r5, r5, #2
+    bgt                 avg4
+    pop                 {r4-r6, pc}
+    ENDP
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/aom_convolve_copy_neon.c b/third_party/aom/aom_dsp/arm/aom_convolve_copy_neon.c
new file mode 100644
index 0000000000..9e57c7176f
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/aom_convolve_copy_neon.c
@@ -0,0 +1,93 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+void aom_convolve_copy_neon(const uint8_t *src,    // r0
+                            ptrdiff_t src_stride,  // r1
+                            uint8_t *dst,          // r2
+                            ptrdiff_t dst_stride,  // r3
+                            const int16_t *filter_x, int filter_x_stride,
+                            const int16_t *filter_y, int filter_y_stride, int w,
+                            int h) {
+  uint8x8_t d0u8, d2u8;
+  uint8x16_t q0u8, q1u8, q2u8, q3u8;
+  (void)filter_x;
+  (void)filter_x_stride;
+  (void)filter_y;
+  (void)filter_y_stride;
+
+  if (w > 32) {  // copy64
+    for (; h > 0; h--) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      q2u8 = vld1q_u8(src + 32);
+      q3u8 = vld1q_u8(src + 48);
+      src += src_stride;
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      vst1q_u8(dst + 32, q2u8);
+      vst1q_u8(dst + 48, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w == 32) {  // copy32
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      q1u8 = vld1q_u8(src + 16);
+      src += src_stride;
+      q2u8 = vld1q_u8(src);
+      q3u8 = vld1q_u8(src + 16);
+      src += src_stride;
+
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q1u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q2u8);
+      vst1q_u8(dst + 16, q3u8);
+      dst += dst_stride;
+    }
+  } else if (w > 8) {  // copy16
+    for (; h > 0; h -= 2) {
+      q0u8 = vld1q_u8(src);
+      src += src_stride;
+      q1u8 = vld1q_u8(src);
+      src += src_stride;
+
+      vst1q_u8(dst, q0u8);
+      dst += dst_stride;
+      vst1q_u8(dst, q1u8);
+      dst += dst_stride;
+    }
+  } else if (w == 8) {  // copy8
+    for (; h > 0; h -= 2) {
+      d0u8 = vld1_u8(src);
+      src += src_stride;
+      d2u8 = vld1_u8(src);
+      src += src_stride;
+
+      vst1_u8(dst, d0u8);
+      dst += dst_stride;
+      vst1_u8(dst, d2u8);
+      dst += dst_stride;
+    }
+  } else {  // copy4
+    for (; h > 0; h--) {
+      *(uint32_t *)dst = *(const uint32_t *)src;
+      src += src_stride;
+      dst += dst_stride;
+    }
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/aom_convolve_copy_neon_asm.asm b/third_party/aom/aom_dsp/arm/aom_convolve_copy_neon_asm.asm
new file mode 100644
index 0000000000..443d7178a6
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/aom_convolve_copy_neon_asm.asm
@@ -0,0 +1,87 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_convolve_copy_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|aom_convolve_copy_neon| PROC
+    push                {r4-r5, lr}
+    ldrd                r4, r5, [sp, #28]
+
+    cmp                 r4, #32
+    bgt                 copy64
+    beq                 copy32
+    cmp                 r4, #8
+    bgt                 copy16
+    beq                 copy8
+    b                   copy4
+
+copy64
+    sub                 lr, r1, #32
+    sub                 r3, r3, #32
+copy64_h
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0]!
+    vld1.8              {q2-q3}, [r0], lr
+    vst1.8              {q0-q1}, [r2@128]!
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #1
+    bgt                 copy64_h
+    pop                 {r4-r5, pc}
+
+copy32
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0-q1}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q2-q3}, [r0], r1
+    vst1.8              {q0-q1}, [r2@128], r3
+    vst1.8              {q2-q3}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 copy32
+    pop                 {r4-r5, pc}
+
+copy16
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q0}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {q1}, [r0], r1
+    vst1.8              {q0}, [r2@128], r3
+    vst1.8              {q1}, [r2@128], r3
+    subs                r5, r5, #2
+    bgt                 copy16
+    pop                 {r4-r5, pc}
+
+copy8
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {d0}, [r0], r1
+    pld                 [r0, r1, lsl #1]
+    vld1.8              {d2}, [r0], r1
+    vst1.8              {d0}, [r2@64], r3
+    vst1.8              {d2}, [r2@64], r3
+    subs                r5, r5, #2
+    bgt                 copy8
+    pop                 {r4-r5, pc}
+
+copy4
+    ldr                 r12, [r0], r1
+    str                 r12, [r2], r3
+    subs                r5, r5, #1
+    bgt                 copy4
+    pop                 {r4-r5, pc}
+    ENDP
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/aom_convolve_neon.c b/third_party/aom/aom_dsp/arm/aom_convolve_neon.c
new file mode 100644
index 0000000000..6c2997e040
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/aom_convolve_neon.c
@@ -0,0 +1,66 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+void aom_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                        ptrdiff_t dst_stride, const int16_t *filter_x,
+                        int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                        int w, int h) {
+  /* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the
+   * maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4).
+   */
+  DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
+
+  // Account for the vertical phase needing 3 lines prior and 4 lines post
+  int intermediate_height = h + 7;
+
+  assert(y_step_q4 == 16);
+  assert(x_step_q4 == 16);
+
+  /* Filter starting 3 lines back. The neon implementation will ignore the
+   * given height and filter a multiple of 4 lines. Since this goes in to
+   * the temp buffer which has lots of extra room and is subsequently discarded
+   * this is safe if somewhat less than ideal.
+   */
+  aom_convolve8_horiz_neon(src - src_stride * 3, src_stride, temp, 64, filter_x,
+                           x_step_q4, filter_y, y_step_q4, w,
+                           intermediate_height);
+
+  /* Step into the temp buffer 3 lines to get the actual frame data */
+  aom_convolve8_vert_neon(temp + 64 * 3, 64, dst, dst_stride, filter_x,
+                          x_step_q4, filter_y, y_step_q4, w, h);
+}
+
+void aom_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4, int w,
+                            int h) {
+  DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
+  int intermediate_height = h + 7;
+
+  assert(y_step_q4 == 16);
+  assert(x_step_q4 == 16);
+
+  /* This implementation has the same issues as above. In addition, we only want
+   * to average the values after both passes.
+   */
+  aom_convolve8_horiz_neon(src - src_stride * 3, src_stride, temp, 64, filter_x,
+                           x_step_q4, filter_y, y_step_q4, w,
+                           intermediate_height);
+  aom_convolve8_avg_vert_neon(temp + 64 * 3, 64, dst, dst_stride, filter_x,
+                              x_step_q4, filter_y, y_step_q4, w, h);
+}
diff --git a/third_party/aom/aom_dsp/arm/avg_neon.c b/third_party/aom/aom_dsp/arm/avg_neon.c
new file mode 100644
index 0000000000..e730ccbccc
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/avg_neon.c
@@ -0,0 +1,254 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+
+static INLINE unsigned int horizontal_add_u16x8(const uint16x8_t v_16x8) {
+  const uint32x4_t a = vpaddlq_u16(v_16x8);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+
+unsigned int aom_avg_4x4_neon(const uint8_t *s, int p) {
+  uint16x8_t v_sum;
+  uint32x2_t v_s0 = vdup_n_u32(0);
+  uint32x2_t v_s1 = vdup_n_u32(0);
+  v_s0 = vld1_lane_u32((const uint32_t *)s, v_s0, 0);
+  v_s0 = vld1_lane_u32((const uint32_t *)(s + p), v_s0, 1);
+  v_s1 = vld1_lane_u32((const uint32_t *)(s + 2 * p), v_s1, 0);
+  v_s1 = vld1_lane_u32((const uint32_t *)(s + 3 * p), v_s1, 1);
+  v_sum = vaddl_u8(vreinterpret_u8_u32(v_s0), vreinterpret_u8_u32(v_s1));
+  return (horizontal_add_u16x8(v_sum) + 8) >> 4;
+}
+
+unsigned int aom_avg_8x8_neon(const uint8_t *s, int p) {
+  uint8x8_t v_s0 = vld1_u8(s);
+  const uint8x8_t v_s1 = vld1_u8(s + p);
+  uint16x8_t v_sum = vaddl_u8(v_s0, v_s1);
+
+  v_s0 = vld1_u8(s + 2 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 3 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 4 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 5 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 6 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  v_s0 = vld1_u8(s + 7 * p);
+  v_sum = vaddw_u8(v_sum, v_s0);
+
+  return (horizontal_add_u16x8(v_sum) + 32) >> 6;
+}
+
+// coeff: 16 bits, dynamic range [-32640, 32640].
+// length: value range {16, 64, 256, 1024}.
+int aom_satd_neon(const int16_t *coeff, int length) {
+  const int16x4_t zero = vdup_n_s16(0);
+  int32x4_t accum = vdupq_n_s32(0);
+
+  do {
+    const int16x8_t src0 = vld1q_s16(coeff);
+    const int16x8_t src8 = vld1q_s16(coeff + 8);
+    accum = vabal_s16(accum, vget_low_s16(src0), zero);
+    accum = vabal_s16(accum, vget_high_s16(src0), zero);
+    accum = vabal_s16(accum, vget_low_s16(src8), zero);
+    accum = vabal_s16(accum, vget_high_s16(src8), zero);
+    length -= 16;
+    coeff += 16;
+  } while (length != 0);
+
+  {
+    // satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
+    const int64x2_t s0 = vpaddlq_s32(accum);  // cascading summation of 'accum'.
+    const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
+                                  vreinterpret_s32_s64(vget_high_s64(s0)));
+    const int satd = vget_lane_s32(s1, 0);
+    return satd;
+  }
+}
+
+void aom_int_pro_row_neon(int16_t hbuf[16], uint8_t const *ref, int ref_stride,
+                          int height) {
+  int i;
+  uint16x8_t vec_sum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_hi = vdupq_n_u16(0);
+  const int shift_factor = ((height >> 5) + 3) * -1;
+  const int16x8_t vec_shift = vdupq_n_s16(shift_factor);
+
+  for (i = 0; i < height; i += 8) {
+    const uint8x16_t vec_row1 = vld1q_u8(ref);
+    const uint8x16_t vec_row2 = vld1q_u8(ref + ref_stride);
+    const uint8x16_t vec_row3 = vld1q_u8(ref + ref_stride * 2);
+    const uint8x16_t vec_row4 = vld1q_u8(ref + ref_stride * 3);
+    const uint8x16_t vec_row5 = vld1q_u8(ref + ref_stride * 4);
+    const uint8x16_t vec_row6 = vld1q_u8(ref + ref_stride * 5);
+    const uint8x16_t vec_row7 = vld1q_u8(ref + ref_stride * 6);
+    const uint8x16_t vec_row8 = vld1q_u8(ref + ref_stride * 7);
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row1));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row1));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row2));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row2));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row3));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row3));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row4));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row4));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row5));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row5));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row6));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row6));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row7));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row7));
+
+    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row8));
+    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row8));
+
+    ref += ref_stride * 8;
+  }
+
+  vec_sum_lo = vshlq_u16(vec_sum_lo, vec_shift);
+  vec_sum_hi = vshlq_u16(vec_sum_hi, vec_shift);
+
+  vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_lo));
+  hbuf += 8;
+  vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_hi));
+}
+
+int16_t aom_int_pro_col_neon(uint8_t const *ref, const int width) {
+  int i;
+  uint16x8_t vec_sum = vdupq_n_u16(0);
+
+  for (i = 0; i < width; i += 16) {
+    const uint8x16_t vec_row = vld1q_u8(ref);
+    vec_sum = vaddw_u8(vec_sum, vget_low_u8(vec_row));
+    vec_sum = vaddw_u8(vec_sum, vget_high_u8(vec_row));
+    ref += 16;
+  }
+
+  return horizontal_add_u16x8(vec_sum);
+}
+
+// ref, src = [0, 510] - max diff = 16-bits
+// bwl = {2, 3, 4}, width = {16, 32, 64}
+int aom_vector_var_neon(int16_t const *ref, int16_t const *src, int bwl) {
+  int width = 4 << bwl;
+  int32x4_t sse = vdupq_n_s32(0);
+  int16x8_t total = vdupq_n_s16(0);
+
+  assert(width >= 8);
+  assert((width % 8) == 0);
+
+  do {
+    const int16x8_t r = vld1q_s16(ref);
+    const int16x8_t s = vld1q_s16(src);
+    const int16x8_t diff = vsubq_s16(r, s);  // [-510, 510], 10 bits.
+    const int16x4_t diff_lo = vget_low_s16(diff);
+    const int16x4_t diff_hi = vget_high_s16(diff);
+    sse = vmlal_s16(sse, diff_lo, diff_lo);  // dynamic range 26 bits.
+    sse = vmlal_s16(sse, diff_hi, diff_hi);
+    total = vaddq_s16(total, diff);  // dynamic range 16 bits.
+
+    ref += 8;
+    src += 8;
+    width -= 8;
+  } while (width != 0);
+
+  {
+    // Note: 'total''s pairwise addition could be implemented similarly to
+    // horizontal_add_u16x8(), but one less vpaddl with 'total' when paired
+    // with the summation of 'sse' performed better on a Cortex-A15.
+    const int32x4_t t0 = vpaddlq_s16(total);  // cascading summation of 'total'
+    const int32x2_t t1 = vadd_s32(vget_low_s32(t0), vget_high_s32(t0));
+    const int32x2_t t2 = vpadd_s32(t1, t1);
+    const int t = vget_lane_s32(t2, 0);
+    const int64x2_t s0 = vpaddlq_s32(sse);  // cascading summation of 'sse'.
+    const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
+                                  vreinterpret_s32_s64(vget_high_s64(s0)));
+    const int s = vget_lane_s32(s1, 0);
+    const int shift_factor = bwl + 2;
+    return s - ((t * t) >> shift_factor);
+  }
+}
+
+void aom_minmax_8x8_neon(const uint8_t *a, int a_stride, const uint8_t *b,
+                         int b_stride, int *min, int *max) {
+  // Load and concatenate.
+  const uint8x16_t a01 = vcombine_u8(vld1_u8(a), vld1_u8(a + a_stride));
+  const uint8x16_t a23 =
+      vcombine_u8(vld1_u8(a + 2 * a_stride), vld1_u8(a + 3 * a_stride));
+  const uint8x16_t a45 =
+      vcombine_u8(vld1_u8(a + 4 * a_stride), vld1_u8(a + 5 * a_stride));
+  const uint8x16_t a67 =
+      vcombine_u8(vld1_u8(a + 6 * a_stride), vld1_u8(a + 7 * a_stride));
+
+  const uint8x16_t b01 = vcombine_u8(vld1_u8(b), vld1_u8(b + b_stride));
+  const uint8x16_t b23 =
+      vcombine_u8(vld1_u8(b + 2 * b_stride), vld1_u8(b + 3 * b_stride));
+  const uint8x16_t b45 =
+      vcombine_u8(vld1_u8(b + 4 * b_stride), vld1_u8(b + 5 * b_stride));
+  const uint8x16_t b67 =
+      vcombine_u8(vld1_u8(b + 6 * b_stride), vld1_u8(b + 7 * b_stride));
+
+  // Absolute difference.
+  const uint8x16_t ab01_diff = vabdq_u8(a01, b01);
+  const uint8x16_t ab23_diff = vabdq_u8(a23, b23);
+  const uint8x16_t ab45_diff = vabdq_u8(a45, b45);
+  const uint8x16_t ab67_diff = vabdq_u8(a67, b67);
+
+  // Max values between the Q vectors.
+  const uint8x16_t ab0123_max = vmaxq_u8(ab01_diff, ab23_diff);
+  const uint8x16_t ab4567_max = vmaxq_u8(ab45_diff, ab67_diff);
+  const uint8x16_t ab0123_min = vminq_u8(ab01_diff, ab23_diff);
+  const uint8x16_t ab4567_min = vminq_u8(ab45_diff, ab67_diff);
+
+  const uint8x16_t ab07_max = vmaxq_u8(ab0123_max, ab4567_max);
+  const uint8x16_t ab07_min = vminq_u8(ab0123_min, ab4567_min);
+
+  // Split to D and start doing pairwise.
+  uint8x8_t ab_max = vmax_u8(vget_high_u8(ab07_max), vget_low_u8(ab07_max));
+  uint8x8_t ab_min = vmin_u8(vget_high_u8(ab07_min), vget_low_u8(ab07_min));
+
+  // Enough runs of vpmax/min propogate the max/min values to every position.
+  ab_max = vpmax_u8(ab_max, ab_max);
+  ab_min = vpmin_u8(ab_min, ab_min);
+
+  ab_max = vpmax_u8(ab_max, ab_max);
+  ab_min = vpmin_u8(ab_min, ab_min);
+
+  ab_max = vpmax_u8(ab_max, ab_max);
+  ab_min = vpmin_u8(ab_min, ab_min);
+
+  *min = *max = 0;  // Clear high bits
+  // Store directly to avoid costly neon->gpr transfer.
+  vst1_lane_u8((uint8_t *)max, ab_max, 0);
+  vst1_lane_u8((uint8_t *)min, ab_min, 0);
+}
diff --git a/third_party/aom/aom_dsp/arm/bilinear_filter_media.asm b/third_party/aom/aom_dsp/arm/bilinear_filter_media.asm
new file mode 100644
index 0000000000..17b7d25f97
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/bilinear_filter_media.asm
@@ -0,0 +1,240 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_filter_block2d_bil_first_pass_media|
+    EXPORT  |aom_filter_block2d_bil_second_pass_media|
+
+    AREA    |.text|, CODE, READONLY  ; name this block of code
+
+;-------------------------------------
+; r0    unsigned char  *src_ptr,
+; r1    unsigned short *dst_ptr,
+; r2    unsigned int    src_pitch,
+; r3    unsigned int    height,
+; stack unsigned int    width,
+; stack const short    *aom_filter
+;-------------------------------------
+; The output is transposed stroed in output array to make it easy for second pass filtering.
+|aom_filter_block2d_bil_first_pass_media| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; aom_filter address
+    ldr     r4, [sp, #36]                   ; width
+
+    mov     r12, r3                         ; outer-loop counter
+
+    add     r7, r2, r4                      ; preload next row
+    pld     [r0, r7]
+
+    sub     r2, r2, r4                      ; src increment for height loop
+
+    ldr     r5, [r11]                       ; load up filter coefficients
+
+    mov     r3, r3, lsl #1                  ; height*2
+    add     r3, r3, #2                      ; plus 2 to make output buffer 4-bit aligned since height is actually (height+1)
+
+    mov     r11, r1                         ; save dst_ptr for each row
+
+    cmp     r5, #128                        ; if filter coef = 128, then skip the filter
+    beq     bil_null_1st_filter
+
+|bil_height_loop_1st_v6|
+    ldrb    r6, [r0]                        ; load source data
+    ldrb    r7, [r0, #1]
+    ldrb    r8, [r0, #2]
+    mov     lr, r4, lsr #2                  ; 4-in-parellel loop counter
+
+|bil_width_loop_1st_v6|
+    ldrb    r9, [r0, #3]
+    ldrb    r10, [r0, #4]
+
+    pkhbt   r6, r6, r7, lsl #16             ; src[1] | src[0]
+    pkhbt   r7, r7, r8, lsl #16             ; src[2] | src[1]
+
+    smuad   r6, r6, r5                      ; apply the filter
+    pkhbt   r8, r8, r9, lsl #16             ; src[3] | src[2]
+    smuad   r7, r7, r5
+    pkhbt   r9, r9, r10, lsl #16            ; src[4] | src[3]
+
+    smuad   r8, r8, r5
+    smuad   r9, r9, r5
+
+    add     r0, r0, #4
+    subs    lr, lr, #1
+
+    add     r6, r6, #0x40                   ; round_shift_and_clamp
+    add     r7, r7, #0x40
+    usat    r6, #16, r6, asr #7
+    usat    r7, #16, r7, asr #7
+
+    strh    r6, [r1], r3                    ; result is transposed and stored
+
+    add     r8, r8, #0x40                   ; round_shift_and_clamp
+    strh    r7, [r1], r3
+    add     r9, r9, #0x40
+    usat    r8, #16, r8, asr #7
+    usat    r9, #16, r9, asr #7
+
+    strh    r8, [r1], r3                    ; result is transposed and stored
+
+    ldrneb  r6, [r0]                        ; load source data
+    strh    r9, [r1], r3
+
+    ldrneb  r7, [r0, #1]
+    ldrneb  r8, [r0, #2]
+
+    bne     bil_width_loop_1st_v6
+
+    add     r0, r0, r2                      ; move to next input row
+    subs    r12, r12, #1
+
+    add     r9, r2, r4, lsl #1              ; adding back block width
+    pld     [r0, r9]                        ; preload next row
+
+    add     r11, r11, #2                    ; move over to next column
+    mov     r1, r11
+
+    bne     bil_height_loop_1st_v6
+
+    ldmia   sp!, {r4 - r11, pc}
+
+|bil_null_1st_filter|
+|bil_height_loop_null_1st|
+    mov     lr, r4, lsr #2                  ; loop counter
+
+|bil_width_loop_null_1st|
+    ldrb    r6, [r0]                        ; load data
+    ldrb    r7, [r0, #1]
+    ldrb    r8, [r0, #2]
+    ldrb    r9, [r0, #3]
+
+    strh    r6, [r1], r3                    ; store it to immediate buffer
+    add     r0, r0, #4
+    strh    r7, [r1], r3
+    subs    lr, lr, #1
+    strh    r8, [r1], r3
+    strh    r9, [r1], r3
+
+    bne     bil_width_loop_null_1st
+
+    subs    r12, r12, #1
+    add     r0, r0, r2                      ; move to next input line
+    add     r11, r11, #2                    ; move over to next column
+    mov     r1, r11
+
+    bne     bil_height_loop_null_1st
+
+    ldmia   sp!, {r4 - r11, pc}
+
+    ENDP  ; |aom_filter_block2d_bil_first_pass_media|
+
+
+;---------------------------------
+; r0    unsigned short *src_ptr,
+; r1    unsigned char  *dst_ptr,
+; r2    int             dst_pitch,
+; r3    unsigned int    height,
+; stack unsigned int    width,
+; stack const short    *aom_filter
+;---------------------------------
+|aom_filter_block2d_bil_second_pass_media| PROC
+    stmdb   sp!, {r4 - r11, lr}
+
+    ldr     r11, [sp, #40]                  ; aom_filter address
+    ldr     r4, [sp, #36]                   ; width
+
+    ldr     r5, [r11]                       ; load up filter coefficients
+    mov     r12, r4                         ; outer-loop counter = width, since we work on transposed data matrix
+    mov     r11, r1
+
+    cmp     r5, #128                        ; if filter coef = 128, then skip the filter
+    beq     bil_null_2nd_filter
+
+|bil_height_loop_2nd|
+    ldr     r6, [r0]                        ; load the data
+    ldr     r8, [r0, #4]
+    ldrh    r10, [r0, #8]
+    mov     lr, r3, lsr #2                  ; loop counter
+
+|bil_width_loop_2nd|
+    pkhtb   r7, r6, r8                      ; src[1] | src[2]
+    pkhtb   r9, r8, r10                     ; src[3] | src[4]
+
+    smuad   r6, r6, r5                      ; apply filter
+    smuad   r8, r8, r5                      ; apply filter
+
+    subs    lr, lr, #1
+
+    smuadx  r7, r7, r5                      ; apply filter
+    smuadx  r9, r9, r5                      ; apply filter
+
+    add     r0, r0, #8
+
+    add     r6, r6, #0x40                   ; round_shift_and_clamp
+    add     r7, r7, #0x40
+    usat    r6, #8, r6, asr #7
+    usat    r7, #8, r7, asr #7
+    strb    r6, [r1], r2                    ; the result is transposed back and stored
+
+    add     r8, r8, #0x40                   ; round_shift_and_clamp
+    strb    r7, [r1], r2
+    add     r9, r9, #0x40
+    usat    r8, #8, r8, asr #7
+    usat    r9, #8, r9, asr #7
+    strb    r8, [r1], r2                    ; the result is transposed back and stored
+
+    ldrne   r6, [r0]                        ; load data
+    strb    r9, [r1], r2
+    ldrne   r8, [r0, #4]
+    ldrneh  r10, [r0, #8]
+
+    bne     bil_width_loop_2nd
+
+    subs    r12, r12, #1
+    add     r0, r0, #4                      ; update src for next row
+    add     r11, r11, #1
+    mov     r1, r11
+
+    bne     bil_height_loop_2nd
+    ldmia   sp!, {r4 - r11, pc}
+
+|bil_null_2nd_filter|
+|bil_height_loop_null_2nd|
+    mov     lr, r3, lsr #2
+
+|bil_width_loop_null_2nd|
+    ldr     r6, [r0], #4                    ; load data
+    subs    lr, lr, #1
+    ldr     r8, [r0], #4
+
+    strb    r6, [r1], r2                    ; store data
+    mov     r7, r6, lsr #16
+    strb    r7, [r1], r2
+    mov     r9, r8, lsr #16
+    strb    r8, [r1], r2
+    strb    r9, [r1], r2
+
+    bne     bil_width_loop_null_2nd
+
+    subs    r12, r12, #1
+    add     r0, r0, #4
+    add     r11, r11, #1
+    mov     r1, r11
+
+    bne     bil_height_loop_null_2nd
+
+    ldmia   sp!, {r4 - r11, pc}
+    ENDP  ; |aom_filter_block2d_second_pass_media|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/fwd_txfm_neon.c b/third_party/aom/aom_dsp/arm/fwd_txfm_neon.c
new file mode 100644
index 0000000000..1cf8a3a6ed
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/fwd_txfm_neon.c
@@ -0,0 +1,221 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "aom_dsp/txfm_common.h"
+
+void aom_fdct8x8_neon(const int16_t *input, int16_t *final_output, int stride) {
+  int i;
+  // stage 1
+  int16x8_t input_0 = vshlq_n_s16(vld1q_s16(&input[0 * stride]), 2);
+  int16x8_t input_1 = vshlq_n_s16(vld1q_s16(&input[1 * stride]), 2);
+  int16x8_t input_2 = vshlq_n_s16(vld1q_s16(&input[2 * stride]), 2);
+  int16x8_t input_3 = vshlq_n_s16(vld1q_s16(&input[3 * stride]), 2);
+  int16x8_t input_4 = vshlq_n_s16(vld1q_s16(&input[4 * stride]), 2);
+  int16x8_t input_5 = vshlq_n_s16(vld1q_s16(&input[5 * stride]), 2);
+  int16x8_t input_6 = vshlq_n_s16(vld1q_s16(&input[6 * stride]), 2);
+  int16x8_t input_7 = vshlq_n_s16(vld1q_s16(&input[7 * stride]), 2);
+  for (i = 0; i < 2; ++i) {
+    int16x8_t out_0, out_1, out_2, out_3, out_4, out_5, out_6, out_7;
+    const int16x8_t v_s0 = vaddq_s16(input_0, input_7);
+    const int16x8_t v_s1 = vaddq_s16(input_1, input_6);
+    const int16x8_t v_s2 = vaddq_s16(input_2, input_5);
+    const int16x8_t v_s3 = vaddq_s16(input_3, input_4);
+    const int16x8_t v_s4 = vsubq_s16(input_3, input_4);
+    const int16x8_t v_s5 = vsubq_s16(input_2, input_5);
+    const int16x8_t v_s6 = vsubq_s16(input_1, input_6);
+    const int16x8_t v_s7 = vsubq_s16(input_0, input_7);
+    // fdct4(step, step);
+    int16x8_t v_x0 = vaddq_s16(v_s0, v_s3);
+    int16x8_t v_x1 = vaddq_s16(v_s1, v_s2);
+    int16x8_t v_x2 = vsubq_s16(v_s1, v_s2);
+    int16x8_t v_x3 = vsubq_s16(v_s0, v_s3);
+    // fdct4(step, step);
+    int32x4_t v_t0_lo = vaddl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
+    int32x4_t v_t0_hi = vaddl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
+    int32x4_t v_t1_lo = vsubl_s16(vget_low_s16(v_x0), vget_low_s16(v_x1));
+    int32x4_t v_t1_hi = vsubl_s16(vget_high_s16(v_x0), vget_high_s16(v_x1));
+    int32x4_t v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_24_64);
+    int32x4_t v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_24_64);
+    int32x4_t v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_24_64);
+    int32x4_t v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_24_64);
+    v_t2_lo = vmlal_n_s16(v_t2_lo, vget_low_s16(v_x3), (int16_t)cospi_8_64);
+    v_t2_hi = vmlal_n_s16(v_t2_hi, vget_high_s16(v_x3), (int16_t)cospi_8_64);
+    v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x2), (int16_t)cospi_8_64);
+    v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x2), (int16_t)cospi_8_64);
+    v_t0_lo = vmulq_n_s32(v_t0_lo, (int32_t)cospi_16_64);
+    v_t0_hi = vmulq_n_s32(v_t0_hi, (int32_t)cospi_16_64);
+    v_t1_lo = vmulq_n_s32(v_t1_lo, (int32_t)cospi_16_64);
+    v_t1_hi = vmulq_n_s32(v_t1_hi, (int32_t)cospi_16_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
+      const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
+      const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
+      const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
+      out_0 = vcombine_s16(a, c);  // 00 01 02 03 40 41 42 43
+      out_2 = vcombine_s16(e, g);  // 20 21 22 23 60 61 62 63
+      out_4 = vcombine_s16(b, d);  // 04 05 06 07 44 45 46 47
+      out_6 = vcombine_s16(f, h);  // 24 25 26 27 64 65 66 67
+    }
+    // Stage 2
+    v_x0 = vsubq_s16(v_s6, v_s5);
+    v_x1 = vaddq_s16(v_s6, v_s5);
+    v_t0_lo = vmull_n_s16(vget_low_s16(v_x0), (int16_t)cospi_16_64);
+    v_t0_hi = vmull_n_s16(vget_high_s16(v_x0), (int16_t)cospi_16_64);
+    v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_16_64);
+    v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_16_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x8_t ab = vcombine_s16(a, b);
+      const int16x8_t cd = vcombine_s16(c, d);
+      // Stage 3
+      v_x0 = vaddq_s16(v_s4, ab);
+      v_x1 = vsubq_s16(v_s4, ab);
+      v_x2 = vsubq_s16(v_s7, cd);
+      v_x3 = vaddq_s16(v_s7, cd);
+    }
+    // Stage 4
+    v_t0_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_4_64);
+    v_t0_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_4_64);
+    v_t0_lo = vmlal_n_s16(v_t0_lo, vget_low_s16(v_x0), (int16_t)cospi_28_64);
+    v_t0_hi = vmlal_n_s16(v_t0_hi, vget_high_s16(v_x0), (int16_t)cospi_28_64);
+    v_t1_lo = vmull_n_s16(vget_low_s16(v_x1), (int16_t)cospi_12_64);
+    v_t1_hi = vmull_n_s16(vget_high_s16(v_x1), (int16_t)cospi_12_64);
+    v_t1_lo = vmlal_n_s16(v_t1_lo, vget_low_s16(v_x2), (int16_t)cospi_20_64);
+    v_t1_hi = vmlal_n_s16(v_t1_hi, vget_high_s16(v_x2), (int16_t)cospi_20_64);
+    v_t2_lo = vmull_n_s16(vget_low_s16(v_x2), (int16_t)cospi_12_64);
+    v_t2_hi = vmull_n_s16(vget_high_s16(v_x2), (int16_t)cospi_12_64);
+    v_t2_lo = vmlsl_n_s16(v_t2_lo, vget_low_s16(v_x1), (int16_t)cospi_20_64);
+    v_t2_hi = vmlsl_n_s16(v_t2_hi, vget_high_s16(v_x1), (int16_t)cospi_20_64);
+    v_t3_lo = vmull_n_s16(vget_low_s16(v_x3), (int16_t)cospi_28_64);
+    v_t3_hi = vmull_n_s16(vget_high_s16(v_x3), (int16_t)cospi_28_64);
+    v_t3_lo = vmlsl_n_s16(v_t3_lo, vget_low_s16(v_x0), (int16_t)cospi_4_64);
+    v_t3_hi = vmlsl_n_s16(v_t3_hi, vget_high_s16(v_x0), (int16_t)cospi_4_64);
+    {
+      const int16x4_t a = vrshrn_n_s32(v_t0_lo, DCT_CONST_BITS);
+      const int16x4_t b = vrshrn_n_s32(v_t0_hi, DCT_CONST_BITS);
+      const int16x4_t c = vrshrn_n_s32(v_t1_lo, DCT_CONST_BITS);
+      const int16x4_t d = vrshrn_n_s32(v_t1_hi, DCT_CONST_BITS);
+      const int16x4_t e = vrshrn_n_s32(v_t2_lo, DCT_CONST_BITS);
+      const int16x4_t f = vrshrn_n_s32(v_t2_hi, DCT_CONST_BITS);
+      const int16x4_t g = vrshrn_n_s32(v_t3_lo, DCT_CONST_BITS);
+      const int16x4_t h = vrshrn_n_s32(v_t3_hi, DCT_CONST_BITS);
+      out_1 = vcombine_s16(a, c);  // 10 11 12 13 50 51 52 53
+      out_3 = vcombine_s16(e, g);  // 30 31 32 33 70 71 72 73
+      out_5 = vcombine_s16(b, d);  // 14 15 16 17 54 55 56 57
+      out_7 = vcombine_s16(f, h);  // 34 35 36 37 74 75 76 77
+    }
+    // transpose 8x8
+    {
+      // 00 01 02 03 40 41 42 43
+      // 10 11 12 13 50 51 52 53
+      // 20 21 22 23 60 61 62 63
+      // 30 31 32 33 70 71 72 73
+      // 04 05 06 07 44 45 46 47
+      // 14 15 16 17 54 55 56 57
+      // 24 25 26 27 64 65 66 67
+      // 34 35 36 37 74 75 76 77
+      const int32x4x2_t r02_s32 =
+          vtrnq_s32(vreinterpretq_s32_s16(out_0), vreinterpretq_s32_s16(out_2));
+      const int32x4x2_t r13_s32 =
+          vtrnq_s32(vreinterpretq_s32_s16(out_1), vreinterpretq_s32_s16(out_3));
+      const int32x4x2_t r46_s32 =
+          vtrnq_s32(vreinterpretq_s32_s16(out_4), vreinterpretq_s32_s16(out_6));
+      const int32x4x2_t r57_s32 =
+          vtrnq_s32(vreinterpretq_s32_s16(out_5), vreinterpretq_s32_s16(out_7));
+      const int16x8x2_t r01_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[0]),
+                    vreinterpretq_s16_s32(r13_s32.val[0]));
+      const int16x8x2_t r23_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r02_s32.val[1]),
+                    vreinterpretq_s16_s32(r13_s32.val[1]));
+      const int16x8x2_t r45_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[0]),
+                    vreinterpretq_s16_s32(r57_s32.val[0]));
+      const int16x8x2_t r67_s16 =
+          vtrnq_s16(vreinterpretq_s16_s32(r46_s32.val[1]),
+                    vreinterpretq_s16_s32(r57_s32.val[1]));
+      input_0 = r01_s16.val[0];
+      input_1 = r01_s16.val[1];
+      input_2 = r23_s16.val[0];
+      input_3 = r23_s16.val[1];
+      input_4 = r45_s16.val[0];
+      input_5 = r45_s16.val[1];
+      input_6 = r67_s16.val[0];
+      input_7 = r67_s16.val[1];
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }  // for
+  {
+    // from aom_dct_sse2.c
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const int16x8_t sign_in0 = vshrq_n_s16(input_0, 15);
+    const int16x8_t sign_in1 = vshrq_n_s16(input_1, 15);
+    const int16x8_t sign_in2 = vshrq_n_s16(input_2, 15);
+    const int16x8_t sign_in3 = vshrq_n_s16(input_3, 15);
+    const int16x8_t sign_in4 = vshrq_n_s16(input_4, 15);
+    const int16x8_t sign_in5 = vshrq_n_s16(input_5, 15);
+    const int16x8_t sign_in6 = vshrq_n_s16(input_6, 15);
+    const int16x8_t sign_in7 = vshrq_n_s16(input_7, 15);
+    input_0 = vhsubq_s16(input_0, sign_in0);
+    input_1 = vhsubq_s16(input_1, sign_in1);
+    input_2 = vhsubq_s16(input_2, sign_in2);
+    input_3 = vhsubq_s16(input_3, sign_in3);
+    input_4 = vhsubq_s16(input_4, sign_in4);
+    input_5 = vhsubq_s16(input_5, sign_in5);
+    input_6 = vhsubq_s16(input_6, sign_in6);
+    input_7 = vhsubq_s16(input_7, sign_in7);
+    // store results
+    vst1q_s16(&final_output[0 * 8], input_0);
+    vst1q_s16(&final_output[1 * 8], input_1);
+    vst1q_s16(&final_output[2 * 8], input_2);
+    vst1q_s16(&final_output[3 * 8], input_3);
+    vst1q_s16(&final_output[4 * 8], input_4);
+    vst1q_s16(&final_output[5 * 8], input_5);
+    vst1q_s16(&final_output[6 * 8], input_6);
+    vst1q_s16(&final_output[7 * 8], input_7);
+  }
+}
+
+void aom_fdct8x8_1_neon(const int16_t *input, int16_t *output, int stride) {
+  int r;
+  int16x8_t sum = vld1q_s16(&input[0]);
+  for (r = 1; r < 8; ++r) {
+    const int16x8_t input_00 = vld1q_s16(&input[r * stride]);
+    sum = vaddq_s16(sum, input_00);
+  }
+  {
+    const int32x4_t a = vpaddlq_s16(sum);
+    const int64x2_t b = vpaddlq_s32(a);
+    const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                                 vreinterpret_s32_s64(vget_high_s64(b)));
+    output[0] = vget_lane_s16(vreinterpret_s16_s32(c), 0);
+    output[1] = 0;
+  }
+}
diff --git a/third_party/aom/aom_dsp/arm/hadamard_neon.c b/third_party/aom/aom_dsp/arm/hadamard_neon.c
new file mode 100644
index 0000000000..9baefae47f
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/hadamard_neon.c
@@ -0,0 +1,200 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+
+static void hadamard8x8_one_pass(int16x8_t *a0, int16x8_t *a1, int16x8_t *a2,
+                                 int16x8_t *a3, int16x8_t *a4, int16x8_t *a5,
+                                 int16x8_t *a6, int16x8_t *a7) {
+  const int16x8_t b0 = vaddq_s16(*a0, *a1);
+  const int16x8_t b1 = vsubq_s16(*a0, *a1);
+  const int16x8_t b2 = vaddq_s16(*a2, *a3);
+  const int16x8_t b3 = vsubq_s16(*a2, *a3);
+  const int16x8_t b4 = vaddq_s16(*a4, *a5);
+  const int16x8_t b5 = vsubq_s16(*a4, *a5);
+  const int16x8_t b6 = vaddq_s16(*a6, *a7);
+  const int16x8_t b7 = vsubq_s16(*a6, *a7);
+
+  const int16x8_t c0 = vaddq_s16(b0, b2);
+  const int16x8_t c1 = vaddq_s16(b1, b3);
+  const int16x8_t c2 = vsubq_s16(b0, b2);
+  const int16x8_t c3 = vsubq_s16(b1, b3);
+  const int16x8_t c4 = vaddq_s16(b4, b6);
+  const int16x8_t c5 = vaddq_s16(b5, b7);
+  const int16x8_t c6 = vsubq_s16(b4, b6);
+  const int16x8_t c7 = vsubq_s16(b5, b7);
+
+  *a0 = vaddq_s16(c0, c4);
+  *a1 = vsubq_s16(c2, c6);
+  *a2 = vsubq_s16(c0, c4);
+  *a3 = vaddq_s16(c2, c6);
+  *a4 = vaddq_s16(c3, c7);
+  *a5 = vsubq_s16(c3, c7);
+  *a6 = vsubq_s16(c1, c5);
+  *a7 = vaddq_s16(c1, c5);
+}
+
+// TODO(johannkoenig): Make a transpose library and dedup with idct. Consider
+// reversing transpose order which may make it easier for the compiler to
+// reconcile the vtrn.64 moves.
+static void transpose8x8(int16x8_t *a0, int16x8_t *a1, int16x8_t *a2,
+                         int16x8_t *a3, int16x8_t *a4, int16x8_t *a5,
+                         int16x8_t *a6, int16x8_t *a7) {
+  // Swap 64 bit elements. Goes from:
+  // a0: 00 01 02 03 04 05 06 07
+  // a1: 08 09 10 11 12 13 14 15
+  // a2: 16 17 18 19 20 21 22 23
+  // a3: 24 25 26 27 28 29 30 31
+  // a4: 32 33 34 35 36 37 38 39
+  // a5: 40 41 42 43 44 45 46 47
+  // a6: 48 49 50 51 52 53 54 55
+  // a7: 56 57 58 59 60 61 62 63
+  // to:
+  // a04_lo: 00 01 02 03 32 33 34 35
+  // a15_lo: 08 09 10 11 40 41 42 43
+  // a26_lo: 16 17 18 19 48 49 50 51
+  // a37_lo: 24 25 26 27 56 57 58 59
+  // a04_hi: 04 05 06 07 36 37 38 39
+  // a15_hi: 12 13 14 15 44 45 46 47
+  // a26_hi: 20 21 22 23 52 53 54 55
+  // a37_hi: 28 29 30 31 60 61 62 63
+  const int16x8_t a04_lo = vcombine_s16(vget_low_s16(*a0), vget_low_s16(*a4));
+  const int16x8_t a15_lo = vcombine_s16(vget_low_s16(*a1), vget_low_s16(*a5));
+  const int16x8_t a26_lo = vcombine_s16(vget_low_s16(*a2), vget_low_s16(*a6));
+  const int16x8_t a37_lo = vcombine_s16(vget_low_s16(*a3), vget_low_s16(*a7));
+  const int16x8_t a04_hi = vcombine_s16(vget_high_s16(*a0), vget_high_s16(*a4));
+  const int16x8_t a15_hi = vcombine_s16(vget_high_s16(*a1), vget_high_s16(*a5));
+  const int16x8_t a26_hi = vcombine_s16(vget_high_s16(*a2), vget_high_s16(*a6));
+  const int16x8_t a37_hi = vcombine_s16(vget_high_s16(*a3), vget_high_s16(*a7));
+
+  // Swap 32 bit elements resulting in:
+  // a0246_lo:
+  // 00 01 16 17 32 33 48 49
+  // 02 03 18 19 34 35 50 51
+  // a1357_lo:
+  // 08 09 24 25 40 41 56 57
+  // 10 11 26 27 42 43 58 59
+  // a0246_hi:
+  // 04 05 20 21 36 37 52 53
+  // 06 07 22 23 38 39 54 55
+  // a1657_hi:
+  // 12 13 28 29 44 45 60 61
+  // 14 15 30 31 46 47 62 63
+  const int32x4x2_t a0246_lo =
+      vtrnq_s32(vreinterpretq_s32_s16(a04_lo), vreinterpretq_s32_s16(a26_lo));
+  const int32x4x2_t a1357_lo =
+      vtrnq_s32(vreinterpretq_s32_s16(a15_lo), vreinterpretq_s32_s16(a37_lo));
+  const int32x4x2_t a0246_hi =
+      vtrnq_s32(vreinterpretq_s32_s16(a04_hi), vreinterpretq_s32_s16(a26_hi));
+  const int32x4x2_t a1357_hi =
+      vtrnq_s32(vreinterpretq_s32_s16(a15_hi), vreinterpretq_s32_s16(a37_hi));
+
+  // Swap 16 bit elements resulting in:
+  // b0:
+  // 00 08 16 24 32 40 48 56
+  // 01 09 17 25 33 41 49 57
+  // b1:
+  // 02 10 18 26 34 42 50 58
+  // 03 11 19 27 35 43 51 59
+  // b2:
+  // 04 12 20 28 36 44 52 60
+  // 05 13 21 29 37 45 53 61
+  // b3:
+  // 06 14 22 30 38 46 54 62
+  // 07 15 23 31 39 47 55 63
+  const int16x8x2_t b0 = vtrnq_s16(vreinterpretq_s16_s32(a0246_lo.val[0]),
+                                   vreinterpretq_s16_s32(a1357_lo.val[0]));
+  const int16x8x2_t b1 = vtrnq_s16(vreinterpretq_s16_s32(a0246_lo.val[1]),
+                                   vreinterpretq_s16_s32(a1357_lo.val[1]));
+  const int16x8x2_t b2 = vtrnq_s16(vreinterpretq_s16_s32(a0246_hi.val[0]),
+                                   vreinterpretq_s16_s32(a1357_hi.val[0]));
+  const int16x8x2_t b3 = vtrnq_s16(vreinterpretq_s16_s32(a0246_hi.val[1]),
+                                   vreinterpretq_s16_s32(a1357_hi.val[1]));
+
+  *a0 = b0.val[0];
+  *a1 = b0.val[1];
+  *a2 = b1.val[0];
+  *a3 = b1.val[1];
+  *a4 = b2.val[0];
+  *a5 = b2.val[1];
+  *a6 = b3.val[0];
+  *a7 = b3.val[1];
+}
+
+void aom_hadamard_8x8_neon(const int16_t *src_diff, int src_stride,
+                           int16_t *coeff) {
+  int16x8_t a0 = vld1q_s16(src_diff);
+  int16x8_t a1 = vld1q_s16(src_diff + src_stride);
+  int16x8_t a2 = vld1q_s16(src_diff + 2 * src_stride);
+  int16x8_t a3 = vld1q_s16(src_diff + 3 * src_stride);
+  int16x8_t a4 = vld1q_s16(src_diff + 4 * src_stride);
+  int16x8_t a5 = vld1q_s16(src_diff + 5 * src_stride);
+  int16x8_t a6 = vld1q_s16(src_diff + 6 * src_stride);
+  int16x8_t a7 = vld1q_s16(src_diff + 7 * src_stride);
+
+  hadamard8x8_one_pass(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+  transpose8x8(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+  hadamard8x8_one_pass(&a0, &a1, &a2, &a3, &a4, &a5, &a6, &a7);
+
+  // Skip the second transpose because it is not required.
+
+  vst1q_s16(coeff + 0, a0);
+  vst1q_s16(coeff + 8, a1);
+  vst1q_s16(coeff + 16, a2);
+  vst1q_s16(coeff + 24, a3);
+  vst1q_s16(coeff + 32, a4);
+  vst1q_s16(coeff + 40, a5);
+  vst1q_s16(coeff + 48, a6);
+  vst1q_s16(coeff + 56, a7);
+}
+
+void aom_hadamard_16x16_neon(const int16_t *src_diff, int src_stride,
+                             int16_t *coeff) {
+  int i;
+
+  /* Rearrange 16x16 to 8x32 and remove stride.
+   * Top left first. */
+  aom_hadamard_8x8_neon(src_diff + 0 + 0 * src_stride, src_stride, coeff + 0);
+  /* Top right. */
+  aom_hadamard_8x8_neon(src_diff + 8 + 0 * src_stride, src_stride, coeff + 64);
+  /* Bottom left. */
+  aom_hadamard_8x8_neon(src_diff + 0 + 8 * src_stride, src_stride, coeff + 128);
+  /* Bottom right. */
+  aom_hadamard_8x8_neon(src_diff + 8 + 8 * src_stride, src_stride, coeff + 192);
+
+  for (i = 0; i < 64; i += 8) {
+    const int16x8_t a0 = vld1q_s16(coeff + 0);
+    const int16x8_t a1 = vld1q_s16(coeff + 64);
+    const int16x8_t a2 = vld1q_s16(coeff + 128);
+    const int16x8_t a3 = vld1q_s16(coeff + 192);
+
+    const int16x8_t b0 = vhaddq_s16(a0, a1);
+    const int16x8_t b1 = vhsubq_s16(a0, a1);
+    const int16x8_t b2 = vhaddq_s16(a2, a3);
+    const int16x8_t b3 = vhsubq_s16(a2, a3);
+
+    const int16x8_t c0 = vaddq_s16(b0, b2);
+    const int16x8_t c1 = vaddq_s16(b1, b3);
+    const int16x8_t c2 = vsubq_s16(b0, b2);
+    const int16x8_t c3 = vsubq_s16(b1, b3);
+
+    vst1q_s16(coeff + 0, c0);
+    vst1q_s16(coeff + 64, c1);
+    vst1q_s16(coeff + 128, c2);
+    vst1q_s16(coeff + 192, c3);
+
+    coeff += 8;
+  }
+}
diff --git a/third_party/aom/aom_dsp/arm/idct16x16_1_add_neon.asm b/third_party/aom/aom_dsp/arm/idct16x16_1_add_neon.asm
new file mode 100644
index 0000000000..d01c4bc03e
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct16x16_1_add_neon.asm
@@ -0,0 +1,201 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+
+
+    EXPORT  |aom_idct16x16_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void aom_idct16x16_1_add_neon(int16_t *input, uint8_t *dest,
+;                                    int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct16x16_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 6)
+    add              r0, r0, #32               ; + (1 <<((6) - 1))
+    asr              r0, r0, #6                ; >> 6
+
+    vdup.s16         q0, r0                    ; duplicate a1
+    mov              r0, #8
+    sub              r2, #8
+
+    ; load destination data row0 - row3
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row4 - row7
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row8 - row11
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    ; load destination data row12 - row15
+    vld1.64          {d2}, [r1], r0
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r0
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r0
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r0
+    vld1.64          {d17}, [r1], r2
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r0
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r0
+    vst1.64          {d31}, [r12], r2
+
+    bx               lr
+    ENDP             ; |aom_idct16x16_1_add_neon|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/idct16x16_1_add_neon.c b/third_party/aom/aom_dsp/arm/idct16x16_1_add_neon.c
new file mode 100644
index 0000000000..196b2a890d
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct16x16_1_add_neon.c
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "aom_dsp/inv_txfm.h"
+#include "aom_ports/mem.h"
+
+void aom_idct16x16_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8x8_t d2u8, d3u8, d30u8, d31u8;
+  uint64x1_t d2u64, d3u64, d4u64, d5u64;
+  uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16;
+  int16x8_t q0s16;
+  uint8_t *d1, *d2;
+  int16_t i, j, a1;
+  int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+  out = dct_const_round_shift(out * cospi_16_64);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+
+  q0s16 = vdupq_n_s16(a1);
+  q0u16 = vreinterpretq_u16_s16(q0s16);
+
+  for (d1 = d2 = dest, i = 0; i < 4; i++) {
+    for (j = 0; j < 2; j++) {
+      d2u64 = vld1_u64((const uint64_t *)d1);
+      d3u64 = vld1_u64((const uint64_t *)(d1 + 8));
+      d1 += dest_stride;
+      d4u64 = vld1_u64((const uint64_t *)d1);
+      d5u64 = vld1_u64((const uint64_t *)(d1 + 8));
+      d1 += dest_stride;
+
+      q9u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64));
+      q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64));
+      q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64));
+      q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64));
+
+      d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+      d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+      d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+      d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+
+      vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+      vst1_u64((uint64_t *)(d2 + 8), vreinterpret_u64_u8(d3u8));
+      d2 += dest_stride;
+      vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8));
+      vst1_u64((uint64_t *)(d2 + 8), vreinterpret_u64_u8(d31u8));
+      d2 += dest_stride;
+    }
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/idct16x16_add_neon.asm b/third_party/aom/aom_dsp/arm/idct16x16_add_neon.asm
new file mode 100644
index 0000000000..4a8f8f183a
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct16x16_add_neon.asm
@@ -0,0 +1,1182 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_idct16x16_256_add_neon_pass1|
+    EXPORT  |aom_idct16x16_256_add_neon_pass2|
+    EXPORT  |aom_idct16x16_10_add_neon_pass1|
+    EXPORT  |aom_idct16x16_10_add_neon_pass2|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Transpose a 8x8 16bit data matrix. Datas are loaded in q8-q15.
+    MACRO
+    TRANSPOSE8X8
+    vswp            d17, d24
+    vswp            d23, d30
+    vswp            d21, d28
+    vswp            d19, d26
+    vtrn.32         q8, q10
+    vtrn.32         q9, q11
+    vtrn.32         q12, q14
+    vtrn.32         q13, q15
+    vtrn.16         q8, q9
+    vtrn.16         q10, q11
+    vtrn.16         q12, q13
+    vtrn.16         q14, q15
+    MEND
+
+    AREA    Block, CODE, READONLY ; name this block of code
+;void |aom_idct16x16_256_add_neon_pass1|(int16_t *input,
+;                                          int16_t *output, int output_stride)
+;
+; r0  int16_t input
+; r1  int16_t *output
+; r2  int  output_stride)
+
+; idct16 stage1 - stage6 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|aom_idct16x16_256_add_neon_pass1| PROC
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 0, 2, 4, 6, 8, 10, 12, 14 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q1,q2}, [r0]!
+    vmov.s16        q15, q1
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0xc00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r12, #0x3e00
+    add             r12, #0xc5
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         d0, r3                    ; duplicate cospi_28_64
+    vdup.16         d1, r12                   ; duplicate cospi_4_64
+
+    ; preloading to avoid stall
+    ; generate cospi_12_64 = 13623
+    mov             r3, #0x3500
+    add             r3, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r12, #0x2300
+    add             r12, #0x8e
+
+    ; step2[4] * cospi_28_64
+    vmull.s16       q2, d18, d0
+    vmull.s16       q3, d19, d0
+
+    ; step2[4] * cospi_4_64
+    vmull.s16       q5, d18, d1
+    vmull.s16       q6, d19, d1
+
+    ; temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64
+    vmlsl.s16       q2, d30, d1
+    vmlsl.s16       q3, d31, d1
+
+    ; temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64
+    vmlal.s16       q5, d30, d0
+    vmlal.s16       q6, d31, d0
+
+    vdup.16         d2, r3                    ; duplicate cospi_12_64
+    vdup.16         d3, r12                   ; duplicate cospi_20_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d8, q2, #14               ; >> 14
+    vqrshrn.s32     d9, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d14, q5, #14              ; >> 14
+    vqrshrn.s32     d15, q6, #14              ; >> 14
+
+    ; preloading to avoid stall
+    ; generate cospi_16_64 = 11585
+    mov             r3, #0x2d00
+    add             r3, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r12, #0x1800
+    add             r12, #0x7e
+
+    ; step2[5] * cospi_12_64
+    vmull.s16       q2, d26, d2
+    vmull.s16       q3, d27, d2
+
+    ; step2[5] * cospi_20_64
+    vmull.s16       q9, d26, d3
+    vmull.s16       q15, d27, d3
+
+    ; temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64
+    vmlsl.s16       q2, d22, d3
+    vmlsl.s16       q3, d23, d3
+
+    ; temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64
+    vmlal.s16       q9, d22, d2
+    vmlal.s16       q15, d23, d2
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d10, q2, #14              ; >> 14
+    vqrshrn.s32     d11, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q15, #14             ; >> 14
+
+    ; stage 4
+    vdup.16         d30, r3                   ; cospi_16_64
+
+    ; step1[0] * cospi_16_64
+    vmull.s16       q2, d16, d30
+    vmull.s16       q11, d17, d30
+
+    ; step1[1] * cospi_16_64
+    vmull.s16       q0, d24, d30
+    vmull.s16       q1, d25, d30
+
+    ; generate cospi_8_64 = 15137
+    mov             r3, #0x3b00
+    add             r3, #0x21
+
+    vdup.16         d30, r12                  ; duplicate cospi_24_64
+    vdup.16         d31, r3                   ; duplicate cospi_8_64
+
+    ; temp1 = (step1[0] + step1[1]) * cospi_16_64
+    vadd.s32        q3, q2, q0
+    vadd.s32        q12, q11, q1
+
+    ; temp2 = (step1[0] - step1[1]) * cospi_16_64
+    vsub.s32        q13, q2, q0
+    vsub.s32        q1, q11, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d16, q3, #14              ; >> 14
+    vqrshrn.s32     d17, q12, #14             ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d18, q13, #14             ; >> 14
+    vqrshrn.s32     d19, q1, #14              ; >> 14
+
+    ; step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+    ; step1[2] * cospi_8_64
+    vmull.s16       q0, d20, d31
+    vmull.s16       q1, d21, d31
+
+    ; step1[2] * cospi_24_64
+    vmull.s16       q12, d20, d30
+    vmull.s16       q13, d21, d30
+
+    ; temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64
+    vmlal.s16       q0, d28, d30
+    vmlal.s16       q1, d29, d30
+
+    ; temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64
+    vmlsl.s16       q12, d28, d31
+    vmlsl.s16       q13, d29, d31
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d22, q0, #14              ; >> 14
+    vqrshrn.s32     d23, q1, #14              ; >> 14
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d20, q12, #14             ; >> 14
+    vqrshrn.s32     d21, q13, #14             ; >> 14
+
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5];
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5];
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7];
+    vadd.s16        q15, q6, q7               ; step2[7] = step1[6] + step1[7];
+
+    ; generate cospi_16_64 = 11585
+    mov             r3, #0x2d00
+    add             r3, #0x41
+
+    ; stage 5
+    vadd.s16        q0, q8, q11               ; step1[0] = step2[0] + step2[3];
+    vadd.s16        q1, q9, q10               ; step1[1] = step2[1] + step2[2];
+    vsub.s16        q2, q9, q10               ; step1[2] = step2[1] - step2[2];
+    vsub.s16        q3, q8, q11               ; step1[3] = step2[0] - step2[3];
+
+    vdup.16         d16, r3;                  ; duplicate cospi_16_64
+
+    ; step2[5] * cospi_16_64
+    vmull.s16       q11, d26, d16
+    vmull.s16       q12, d27, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; temp1 = (step2[6] - step2[5]) * cospi_16_64
+    vsub.s32        q6, q9, q11
+    vsub.s32        q13, q10, q12
+
+    ; temp2 = (step2[5] + step2[6]) * cospi_16_64
+    vadd.s32        q9, q9, q11
+    vadd.s32        q10, q10, q12
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d10, q6, #14              ; >> 14
+    vqrshrn.s32     d11, q13, #14             ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q10, #14             ; >> 14
+
+    ; stage 6
+    vadd.s16        q8, q0, q15                ; step2[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; step2[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; step2[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; step2[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; step2[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; step2[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; step2[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q15              ; step2[7] = step1[0] - step1[7];
+
+    ; store the data
+    vst1.64         {d16}, [r1], r2
+    vst1.64         {d17}, [r1], r2
+    vst1.64         {d18}, [r1], r2
+    vst1.64         {d19}, [r1], r2
+    vst1.64         {d20}, [r1], r2
+    vst1.64         {d21}, [r1], r2
+    vst1.64         {d22}, [r1], r2
+    vst1.64         {d23}, [r1], r2
+    vst1.64         {d24}, [r1], r2
+    vst1.64         {d25}, [r1], r2
+    vst1.64         {d26}, [r1], r2
+    vst1.64         {d27}, [r1], r2
+    vst1.64         {d28}, [r1], r2
+    vst1.64         {d29}, [r1], r2
+    vst1.64         {d30}, [r1], r2
+    vst1.64         {d31}, [r1], r2
+
+    bx              lr
+    ENDP  ; |aom_idct16x16_256_add_neon_pass1|
+
+;void aom_idct16x16_256_add_neon_pass2(int16_t *src,
+;                                        int16_t *output,
+;                                        int16_t *pass1Output,
+;                                        int16_t skip_adding,
+;                                        uint8_t *dest,
+;                                        int dest_stride)
+;
+; r0  int16_t *src
+; r1  int16_t *output,
+; r2  int16_t *pass1Output,
+; r3  int16_t skip_adding,
+; r4  uint8_t *dest,
+; r5  int dest_stride)
+
+; idct16 stage1 - stage7 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|aom_idct16x16_256_add_neon_pass2| PROC
+    push            {r3-r9}
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 1, 3, 5, 7, 9, 11, 13, 15 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q0,q1}, [r0]!
+    vmov.s16        q15, q0;
+
+    ; generate  cospi_30_64 = 1606
+    mov             r3, #0x0600
+    add             r3, #0x46
+
+    ; generate cospi_2_64  = 16305
+    mov             r12, #0x3f00
+    add             r12, #0xb1
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         d12, r3                   ; duplicate cospi_30_64
+    vdup.16         d13, r12                  ; duplicate cospi_2_64
+
+    ; preloading to avoid stall
+    ; generate cospi_14_64 = 12665
+    mov             r3, #0x3100
+    add             r3, #0x79
+
+    ; generate cospi_18_64 = 10394
+    mov             r12, #0x2800
+    add             r12, #0x9a
+
+    ; step1[8] * cospi_30_64
+    vmull.s16       q2, d16, d12
+    vmull.s16       q3, d17, d12
+
+    ; step1[8] * cospi_2_64
+    vmull.s16       q1, d16, d13
+    vmull.s16       q4, d17, d13
+
+    ; temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64
+    vmlsl.s16       q2, d30, d13
+    vmlsl.s16       q3, d31, d13
+
+    ; temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64
+    vmlal.s16       q1, d30, d12
+    vmlal.s16       q4, d31, d12
+
+    vdup.16         d30, r3                   ; duplicate cospi_14_64
+    vdup.16         d31, r12                  ; duplicate cospi_18_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d0, q2, #14               ; >> 14
+    vqrshrn.s32     d1, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d14, q1, #14              ; >> 14
+    vqrshrn.s32     d15, q4, #14              ; >> 14
+
+    ; preloading to avoid stall
+    ; generate cospi_22_64 = 7723
+    mov             r3, #0x1e00
+    add             r3, #0x2b
+
+    ; generate cospi_10_64 = 14449
+    mov             r12, #0x3800
+    add             r12, #0x71
+
+    ; step1[9] * cospi_14_64
+    vmull.s16       q2, d24, d30
+    vmull.s16       q3, d25, d30
+
+    ; step1[9] * cospi_18_64
+    vmull.s16       q4, d24, d31
+    vmull.s16       q5, d25, d31
+
+    ; temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64
+    vmlsl.s16       q2, d22, d31
+    vmlsl.s16       q3, d23, d31
+
+    ; temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64
+    vmlal.s16       q4, d22, d30
+    vmlal.s16       q5, d23, d30
+
+    vdup.16         d30, r3                   ; duplicate cospi_22_64
+    vdup.16         d31, r12                  ; duplicate cospi_10_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d2, q2, #14               ; >> 14
+    vqrshrn.s32     d3, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q4, #14              ; >> 14
+    vqrshrn.s32     d13, q5, #14              ; >> 14
+
+    ; step1[10] * cospi_22_64
+    vmull.s16       q11, d20, d30
+    vmull.s16       q12, d21, d30
+
+    ; step1[10] * cospi_10_64
+    vmull.s16       q4, d20, d31
+    vmull.s16       q5, d21, d31
+
+    ; temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64
+    vmlsl.s16       q11, d26, d31
+    vmlsl.s16       q12, d27, d31
+
+    ; temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64
+    vmlal.s16       q4, d26, d30
+    vmlal.s16       q5, d27, d30
+
+    ; preloading to avoid stall
+    ; generate cospi_6_64 = 15679
+    mov             r3, #0x3d00
+    add             r3, #0x3f
+
+    ; generate cospi_26_64 = 4756
+    mov             r12, #0x1200
+    add             r12, #0x94
+
+    vdup.16         d30, r3                   ; duplicate cospi_6_64
+    vdup.16         d31, r12                  ; duplicate cospi_26_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q11, #14              ; >> 14
+    vqrshrn.s32     d5, q12, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d11, q5, #14              ; >> 14
+    vqrshrn.s32     d10, q4, #14              ; >> 14
+
+    ; step1[11] * cospi_6_64
+    vmull.s16       q10, d28, d30
+    vmull.s16       q11, d29, d30
+
+    ; step1[11] * cospi_26_64
+    vmull.s16       q12, d28, d31
+    vmull.s16       q13, d29, d31
+
+    ; temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64
+    vmlsl.s16       q10, d18, d31
+    vmlsl.s16       q11, d19, d31
+
+    ; temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64
+    vmlal.s16       q12, d18, d30
+    vmlal.s16       q13, d19, d30
+
+    vsub.s16        q9, q0, q1                ; step1[9]=step2[8]-step2[9]
+    vadd.s16        q0, q0, q1                ; step1[8]=step2[8]+step2[9]
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d6, q10, #14              ; >> 14
+    vqrshrn.s32     d7, q11, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d8, q12, #14              ; >> 14
+    vqrshrn.s32     d9, q13, #14              ; >> 14
+
+    ; stage 3
+    vsub.s16        q10, q3, q2               ; step1[10]=-step2[10]+step2[11]
+    vadd.s16        q11, q2, q3               ; step1[11]=step2[10]+step2[11]
+    vadd.s16        q12, q4, q5               ; step1[12]=step2[12]+step2[13]
+    vsub.s16        q13, q4, q5               ; step1[13]=step2[12]-step2[13]
+    vsub.s16        q14, q7, q6               ; step1[14]=-step2[14]+tep2[15]
+    vadd.s16        q7, q6, q7                ; step1[15]=step2[14]+step2[15]
+
+    ; stage 4
+    ; generate cospi_24_64 = 6270
+    mov             r3, #0x1800
+    add             r3, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r12, #0x3b00
+    add             r12, #0x21
+
+    ; -step1[9] * cospi_8_64 + step1[14] * cospi_24_64
+    vdup.16         d30, r12                  ; duplicate cospi_8_64
+    vdup.16         d31, r3                   ; duplicate cospi_24_64
+
+    ; step1[9] * cospi_24_64
+    vmull.s16       q2, d18, d31
+    vmull.s16       q3, d19, d31
+
+    ; step1[14] * cospi_24_64
+    vmull.s16       q4, d28, d31
+    vmull.s16       q5, d29, d31
+
+    ; temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64
+    vmlal.s16       q2, d28, d30
+    vmlal.s16       q3, d29, d30
+
+    ; temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64
+    vmlsl.s16       q4, d18, d30
+    vmlsl.s16       q5, d19, d30
+
+    rsb             r12, #0
+    vdup.16         d30, r12                  ; duplicate -cospi_8_64
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q2, #14              ; >> 14
+    vqrshrn.s32     d13, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d2, q4, #14               ; >> 14
+    vqrshrn.s32     d3, q5, #14               ; >> 14
+
+    vmov.s16        q3, q11
+    vmov.s16        q4, q12
+
+    ; - step1[13] * cospi_8_64
+    vmull.s16       q11, d26, d30
+    vmull.s16       q12, d27, d30
+
+    ; -step1[10] * cospi_8_64
+    vmull.s16       q8, d20, d30
+    vmull.s16       q9, d21, d30
+
+    ; temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64
+    vmlsl.s16       q11, d20, d31
+    vmlsl.s16       q12, d21, d31
+
+    ; temp1 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64
+    vmlal.s16       q8, d26, d31
+    vmlal.s16       q9, d27, d31
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d4, q11, #14              ; >> 14
+    vqrshrn.s32     d5, q12, #14              ; >> 14
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d10, q8, #14              ; >> 14
+    vqrshrn.s32     d11, q9, #14              ; >> 14
+
+    ; stage 5
+    vadd.s16        q8, q0, q3                ; step1[8] = step2[8]+step2[11];
+    vadd.s16        q9, q1, q2                ; step1[9] = step2[9]+step2[10];
+    vsub.s16        q10, q1, q2               ; step1[10] = step2[9]-step2[10];
+    vsub.s16        q11, q0, q3               ; step1[11] = step2[8]-step2[11];
+    vsub.s16        q12, q7, q4               ; step1[12] =-step2[12]+step2[15];
+    vsub.s16        q13, q6, q5               ; step1[13] =-step2[13]+step2[14];
+    vadd.s16        q14, q6, q5               ; step1[14] =step2[13]+step2[14];
+    vadd.s16        q15, q7, q4               ; step1[15] =step2[12]+step2[15];
+
+    ; stage 6.
+    ; generate cospi_16_64 = 11585
+    mov             r12, #0x2d00
+    add             r12, #0x41
+
+    vdup.16         d14, r12                  ; duplicate cospi_16_64
+
+    ; step1[13] * cospi_16_64
+    vmull.s16       q3, d26, d14
+    vmull.s16       q4, d27, d14
+
+    ; step1[10] * cospi_16_64
+    vmull.s16       q0, d20, d14
+    vmull.s16       q1, d21, d14
+
+    ; temp1 = (-step1[10] + step1[13]) * cospi_16_64
+    vsub.s32        q5, q3, q0
+    vsub.s32        q6, q4, q1
+
+    ; temp2 = (step1[10] + step1[13]) * cospi_16_64
+    vadd.s32        q10, q3, q0
+    vadd.s32        q4, q4, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q5, #14               ; >> 14
+    vqrshrn.s32     d5, q6, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d10, q10, #14             ; >> 14
+    vqrshrn.s32     d11, q4, #14              ; >> 14
+
+    ; step1[11] * cospi_16_64
+    vmull.s16       q0, d22, d14
+    vmull.s16       q1, d23, d14
+
+    ; step1[12] * cospi_16_64
+    vmull.s16       q13, d24, d14
+    vmull.s16       q6, d25, d14
+
+    ; temp1 = (-step1[11] + step1[12]) * cospi_16_64
+    vsub.s32        q10, q13, q0
+    vsub.s32        q4, q6, q1
+
+    ; temp2 = (step1[11] + step1[12]) * cospi_16_64
+    vadd.s32        q13, q13, q0
+    vadd.s32        q6, q6, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d6, q10, #14              ; >> 14
+    vqrshrn.s32     d7, q4, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d8, q13, #14              ; >> 14
+    vqrshrn.s32     d9, q6, #14               ; >> 14
+
+    mov              r4, #16                  ; pass1Output stride
+    ldr              r3, [sp]                 ; load skip_adding
+    cmp              r3, #0                   ; check if need adding dest data
+    beq              skip_adding_dest
+
+    ldr              r7, [sp, #28]            ; dest used to save element 0-7
+    mov              r9, r7                   ; save dest pointer for later use
+    ldr              r8, [sp, #32]            ; load dest_stride
+
+    ; stage 7
+    ; load the data in pass1
+    vld1.s16        {q0}, [r2], r4            ; load data step2[0]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[1]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[2]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[3]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q0, q15              ; step2[0] + step2[15]
+    vadd.s16        q13, q1, q14              ; step2[1] + step2[14]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vsub.s16        q14, q1, q14              ; step2[1] - step2[14]
+    vsub.s16        q15, q0, q15              ; step2[0] - step2[15]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q10, q5              ; step2[2] + step2[13]
+    vadd.s16        q13, q11, q4              ; step2[3] + step2[12]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vsub.s16        q4, q11, q4               ; step2[3] - step2[12]
+    vsub.s16        q5, q10, q5               ; step2[2] - step2[13]
+    vld1.s16        {q0}, [r2], r4            ; load data step2[4]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[5]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[6]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[7]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q0, q3               ; step2[4] + step2[11]
+    vadd.s16        q13, q1, q2               ; step2[5] + step2[10]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vsub.s16        q2, q1, q2                ; step2[5] - step2[10]
+    vsub.s16        q3, q0, q3                ; step2[4] - step2[11]
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vadd.s16        q12, q10, q9              ; step2[6] + step2[9]
+    vadd.s16        q13, q11, q8              ; step2[7] + step2[8]
+    vrshr.s16       q12, q12, #6              ; ROUND_POWER_OF_TWO
+    vrshr.s16       q13, q13, #6              ; ROUND_POWER_OF_TWO
+    vaddw.u8        q12, q12, d12             ; + dest[j * dest_stride + i]
+    vaddw.u8        q13, q13, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q12                  ; clip pixel
+    vqmovun.s16     d13, q13                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vsub.s16        q8, q11, q8               ; step2[7] - step2[8]
+    vsub.s16        q9, q10, q9               ; step2[6] - step2[9]
+
+    ; store the data  output 8,9,10,11,12,13,14,15
+    vrshr.s16       q8, q8, #6                ; ROUND_POWER_OF_TWO
+    vaddw.u8        q8, q8, d12               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q8                   ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q9, q9, #6
+    vaddw.u8        q9, q9, d13               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q9                   ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q2, q2, #6
+    vaddw.u8        q2, q2, d12               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q2                   ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q3, q3, #6
+    vaddw.u8        q3, q3, d13               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q3                   ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q4, q4, #6
+    vaddw.u8        q4, q4, d12               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q4                   ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q5, q5, #6
+    vaddw.u8        q5, q5, d13               ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q5                   ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    vld1.64         {d13}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q14, q14, #6
+    vaddw.u8        q14, q14, d12             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d12, q14                  ; clip pixel
+    vst1.64         {d12}, [r9], r8           ; store the data
+    vld1.64         {d12}, [r7], r8           ; load destinatoin data
+    vrshr.s16       q15, q15, #6
+    vaddw.u8        q15, q15, d13             ; + dest[j * dest_stride + i]
+    vqmovun.s16     d13, q15                  ; clip pixel
+    vst1.64         {d13}, [r9], r8           ; store the data
+    b               end_idct16x16_pass2
+
+skip_adding_dest
+    ; stage 7
+    ; load the data in pass1
+    mov              r5, #24
+    mov              r3, #8
+
+    vld1.s16        {q0}, [r2], r4            ; load data step2[0]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[1]
+    vadd.s16        q12, q0, q15              ; step2[0] + step2[15]
+    vadd.s16        q13, q1, q14              ; step2[1] + step2[14]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[2]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[3]
+    vst1.64         {d24}, [r1], r3           ; store output[0]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[1]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q5              ; step2[2] + step2[13]
+    vadd.s16        q13, q11, q4              ; step2[3] + step2[12]
+    vsub.s16        q14, q1, q14              ; step2[1] - step2[14]
+    vsub.s16        q15, q0, q15              ; step2[0] - step2[15]
+    vst1.64         {d24}, [r1], r3           ; store output[2]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[3]
+    vst1.64         {d27}, [r1], r5
+    vsub.s16        q4, q11, q4               ; step2[3] - step2[12]
+    vsub.s16        q5, q10, q5               ; step2[2] - step2[13]
+    vld1.s16        {q0}, [r2], r4            ; load data step2[4]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[5]
+    vadd.s16        q12, q0, q3               ; step2[4] + step2[11]
+    vadd.s16        q13, q1, q2               ; step2[5] + step2[10]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[6]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[7]
+    vst1.64         {d24}, [r1], r3           ; store output[4]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[5]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q9              ; step2[6] + step2[9]
+    vadd.s16        q13, q11, q8              ; step2[7] + step2[8]
+    vsub.s16        q2, q1, q2                ; step2[5] - step2[10]
+    vsub.s16        q3, q0, q3                ; step2[4] - step2[11]
+    vsub.s16        q8, q11, q8               ; step2[7] - step2[8]
+    vsub.s16        q9, q10, q9               ; step2[6] - step2[9]
+    vst1.64         {d24}, [r1], r3           ; store output[6]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[7]
+    vst1.64         {d27}, [r1], r5
+
+    ; store the data  output 8,9,10,11,12,13,14,15
+    vst1.64         {d16}, [r1], r3
+    vst1.64         {d17}, [r1], r5
+    vst1.64         {d18}, [r1], r3
+    vst1.64         {d19}, [r1], r5
+    vst1.64         {d4}, [r1], r3
+    vst1.64         {d5}, [r1], r5
+    vst1.64         {d6}, [r1], r3
+    vst1.64         {d7}, [r1], r5
+    vst1.64         {d8}, [r1], r3
+    vst1.64         {d9}, [r1], r5
+    vst1.64         {d10}, [r1], r3
+    vst1.64         {d11}, [r1], r5
+    vst1.64         {d28}, [r1], r3
+    vst1.64         {d29}, [r1], r5
+    vst1.64         {d30}, [r1], r3
+    vst1.64         {d31}, [r1], r5
+end_idct16x16_pass2
+    pop             {r3-r9}
+    bx              lr
+    ENDP  ; |aom_idct16x16_256_add_neon_pass2|
+
+;void |aom_idct16x16_10_add_neon_pass1|(int16_t *input,
+;                                             int16_t *output, int output_stride)
+;
+; r0  int16_t input
+; r1  int16_t *output
+; r2  int  output_stride)
+
+; idct16 stage1 - stage6 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|aom_idct16x16_10_add_neon_pass1| PROC
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 0, 2, 4, 6, 8, 10, 12, 14 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q1,q2}, [r0]!
+    vmov.s16        q15, q1
+
+    ; generate  cospi_28_64*2 = 6392
+    mov             r3, #0x1800
+    add             r3, #0xf8
+
+    ; generate cospi_4_64*2  = 32138
+    mov             r12, #0x7d00
+    add             r12, #0x8a
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         q0, r3                    ; duplicate cospi_28_64*2
+    vdup.16         q1, r12                   ; duplicate cospi_4_64*2
+
+    ; The following instructions use vqrdmulh to do the
+    ; dct_const_round_shift(step2[4] * cospi_28_64). vvqrdmulh will multiply,
+    ; double, and return the high 16 bits, effectively giving >> 15. Doubling
+    ; the constant will change this to >> 14.
+    ; dct_const_round_shift(step2[4] * cospi_28_64);
+    vqrdmulh.s16    q4, q9, q0
+
+    ; preloading to avoid stall
+    ; generate cospi_16_64*2 = 23170
+    mov             r3, #0x5a00
+    add             r3, #0x82
+
+    ; dct_const_round_shift(step2[4] * cospi_4_64);
+    vqrdmulh.s16    q7, q9, q1
+
+    ; stage 4
+    vdup.16         q1, r3                    ; cospi_16_64*2
+
+    ; generate cospi_16_64 = 11585
+    mov             r3, #0x2d00
+    add             r3, #0x41
+
+    vdup.16         d4, r3;                   ; duplicate cospi_16_64
+
+    ; dct_const_round_shift(step1[0] * cospi_16_64)
+    vqrdmulh.s16    q8, q8, q1
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d14, d4
+    vmull.s16       q10, d15, d4
+
+    ; step2[5] * cospi_16_64
+    vmull.s16       q12, d9, d4
+    vmull.s16       q11, d8, d4
+
+    ; temp1 = (step2[6] - step2[5]) * cospi_16_64
+    vsub.s32        q15, q10, q12
+    vsub.s32        q6, q9, q11
+
+    ; temp2 = (step2[5] + step2[6]) * cospi_16_64
+    vadd.s32        q9, q9, q11
+    vadd.s32        q10, q10, q12
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d11, q15, #14             ; >> 14
+    vqrshrn.s32     d10, q6, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q10, #14             ; >> 14
+
+    ; stage 6
+    vadd.s16        q2, q8, q7                ; step2[0] = step1[0] + step1[7];
+    vadd.s16        q10, q8, q5               ; step2[2] = step1[2] + step1[5];
+    vadd.s16        q11, q8, q4               ; step2[3] = step1[3] + step1[4];
+    vadd.s16        q9, q8, q6                ; step2[1] = step1[1] + step1[6];
+    vsub.s16        q12, q8, q4               ; step2[4] = step1[3] - step1[4];
+    vsub.s16        q13, q8, q5               ; step2[5] = step1[2] - step1[5];
+    vsub.s16        q14, q8, q6               ; step2[6] = step1[1] - step1[6];
+    vsub.s16        q15, q8, q7               ; step2[7] = step1[0] - step1[7];
+
+    ; store the data
+    vst1.64         {d4}, [r1], r2
+    vst1.64         {d5}, [r1], r2
+    vst1.64         {d18}, [r1], r2
+    vst1.64         {d19}, [r1], r2
+    vst1.64         {d20}, [r1], r2
+    vst1.64         {d21}, [r1], r2
+    vst1.64         {d22}, [r1], r2
+    vst1.64         {d23}, [r1], r2
+    vst1.64         {d24}, [r1], r2
+    vst1.64         {d25}, [r1], r2
+    vst1.64         {d26}, [r1], r2
+    vst1.64         {d27}, [r1], r2
+    vst1.64         {d28}, [r1], r2
+    vst1.64         {d29}, [r1], r2
+    vst1.64         {d30}, [r1], r2
+    vst1.64         {d31}, [r1], r2
+
+    bx              lr
+    ENDP  ; |aom_idct16x16_10_add_neon_pass1|
+
+;void aom_idct16x16_10_add_neon_pass2(int16_t *src,
+;                                           int16_t *output,
+;                                           int16_t *pass1Output,
+;                                           int16_t skip_adding,
+;                                           uint8_t *dest,
+;                                           int dest_stride)
+;
+; r0  int16_t *src
+; r1  int16_t *output,
+; r2  int16_t *pass1Output,
+; r3  int16_t skip_adding,
+; r4  uint8_t *dest,
+; r5  int dest_stride)
+
+; idct16 stage1 - stage7 on all the elements loaded in q8-q15. The output
+; will be stored back into q8-q15 registers. This function will touch q0-q7
+; registers and use them as buffer during calculation.
+|aom_idct16x16_10_add_neon_pass2| PROC
+    push            {r3-r9}
+
+    ; TODO(hkuang): Find a better way to load the elements.
+    ; load elements of 1, 3, 5, 7, 9, 11, 13, 15 into q8 - q15
+    vld2.s16        {q8,q9}, [r0]!
+    vld2.s16        {q9,q10}, [r0]!
+    vld2.s16        {q10,q11}, [r0]!
+    vld2.s16        {q11,q12}, [r0]!
+    vld2.s16        {q12,q13}, [r0]!
+    vld2.s16        {q13,q14}, [r0]!
+    vld2.s16        {q14,q15}, [r0]!
+    vld2.s16        {q0,q1}, [r0]!
+    vmov.s16        q15, q0;
+
+    ; generate 2*cospi_30_64 = 3212
+    mov             r3, #0xc00
+    add             r3, #0x8c
+
+    ; generate 2*cospi_2_64  = 32610
+    mov             r12, #0x7f00
+    add             r12, #0x62
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; stage 3
+    vdup.16         q6, r3                    ; duplicate 2*cospi_30_64
+
+    ; dct_const_round_shift(step1[8] * cospi_30_64)
+    vqrdmulh.s16    q0, q8, q6
+
+    vdup.16         q6, r12                   ; duplicate 2*cospi_2_64
+
+    ; dct_const_round_shift(step1[8] * cospi_2_64)
+    vqrdmulh.s16    q7, q8, q6
+
+    ; preloading to avoid stall
+    ; generate 2*cospi_26_64 = 9512
+    mov             r12, #0x2500
+    add             r12, #0x28
+    rsb             r12, #0
+    vdup.16         q15, r12                  ; duplicate -2*cospi_26_64
+
+    ; generate 2*cospi_6_64 = 31358
+    mov             r3, #0x7a00
+    add             r3, #0x7e
+    vdup.16         q14, r3                   ; duplicate 2*cospi_6_64
+
+    ; dct_const_round_shift(- step1[12] * cospi_26_64)
+    vqrdmulh.s16    q3, q9, q15
+
+    ; dct_const_round_shift(step1[12] * cospi_6_64)
+    vqrdmulh.s16    q4, q9, q14
+
+    ; stage 4
+    ; generate cospi_24_64 = 6270
+    mov             r3, #0x1800
+    add             r3, #0x7e
+    vdup.16         d31, r3                   ; duplicate cospi_24_64
+
+    ; generate cospi_8_64 = 15137
+    mov             r12, #0x3b00
+    add             r12, #0x21
+    vdup.16         d30, r12                  ; duplicate cospi_8_64
+
+    ; step1[14] * cospi_24_64
+    vmull.s16       q12, d14, d31
+    vmull.s16       q5, d15, d31
+
+    ; step1[9] * cospi_24_64
+    vmull.s16       q2, d0, d31
+    vmull.s16       q11, d1, d31
+
+    ; temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64
+    vmlsl.s16       q12, d0, d30
+    vmlsl.s16       q5, d1, d30
+
+    ; temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64
+    vmlal.s16       q2, d14, d30
+    vmlal.s16       q11, d15, d30
+
+    rsb              r12, #0
+    vdup.16          d30, r12                 ; duplicate -cospi_8_64
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d2, q12, #14              ; >> 14
+    vqrshrn.s32     d3, q5, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d12, q2, #14              ; >> 14
+    vqrshrn.s32     d13, q11, #14             ; >> 14
+
+    ; - step1[13] * cospi_8_64
+    vmull.s16       q10, d8, d30
+    vmull.s16       q13, d9, d30
+
+    ; -step1[10] * cospi_8_64
+    vmull.s16       q8, d6, d30
+    vmull.s16       q9, d7, d30
+
+    ; temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64
+    vmlsl.s16       q10, d6, d31
+    vmlsl.s16       q13, d7, d31
+
+    ; temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64
+    vmlal.s16       q8, d8, d31
+    vmlal.s16       q9, d9, d31
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q10, #14              ; >> 14
+    vqrshrn.s32     d5, q13, #14              ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d10, q8, #14              ; >> 14
+    vqrshrn.s32     d11, q9, #14              ; >> 14
+
+    ; stage 5
+    vadd.s16        q8, q0, q3                ; step1[8] = step2[8]+step2[11];
+    vadd.s16        q9, q1, q2                ; step1[9] = step2[9]+step2[10];
+    vsub.s16        q10, q1, q2               ; step1[10] = step2[9]-step2[10];
+    vsub.s16        q11, q0, q3               ; step1[11] = step2[8]-step2[11];
+    vsub.s16        q12, q7, q4               ; step1[12] =-step2[12]+step2[15];
+    vsub.s16        q13, q6, q5               ; step1[13] =-step2[13]+step2[14];
+    vadd.s16        q14, q6, q5               ; step1[14] =step2[13]+step2[14];
+    vadd.s16        q15, q7, q4               ; step1[15] =step2[12]+step2[15];
+
+    ; stage 6.
+    ; generate cospi_16_64 = 11585
+    mov             r12, #0x2d00
+    add             r12, #0x41
+
+    vdup.16         d14, r12                  ; duplicate cospi_16_64
+
+    ; step1[13] * cospi_16_64
+    vmull.s16       q3, d26, d14
+    vmull.s16       q4, d27, d14
+
+    ; step1[10] * cospi_16_64
+    vmull.s16       q0, d20, d14
+    vmull.s16       q1, d21, d14
+
+    ; temp1 = (-step1[10] + step1[13]) * cospi_16_64
+    vsub.s32        q5, q3, q0
+    vsub.s32        q6, q4, q1
+
+    ; temp2 = (step1[10] + step1[13]) * cospi_16_64
+    vadd.s32        q0, q3, q0
+    vadd.s32        q1, q4, q1
+
+    ; dct_const_round_shift(temp1)
+    vqrshrn.s32     d4, q5, #14               ; >> 14
+    vqrshrn.s32     d5, q6, #14               ; >> 14
+
+    ; dct_const_round_shift(temp2)
+    vqrshrn.s32     d10, q0, #14              ; >> 14
+    vqrshrn.s32     d11, q1, #14              ; >> 14
+
+    ; step1[11] * cospi_16_64
+    vmull.s16       q0, d22, d14
+    vmull.s16       q1, d23, d14
+
+    ; step1[12] * cospi_16_64
+    vmull.s16       q13, d24, d14
+    vmull.s16       q6, d25, d14
+
+    ; temp1 = (-step1[11] + step1[12]) * cospi_16_64
+    vsub.s32        q10, q13, q0
+    vsub.s32        q4, q6, q1
+
+    ; temp2 = (step1[11] + step1[12]) * cospi_16_64
+    vadd.s32        q13, q13, q0
+    vadd.s32        q6, q6, q1
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d6, q10, #14              ; >> 14
+    vqrshrn.s32     d7, q4, #14               ; >> 14
+
+    ; dct_const_round_shift((step1[11] + step1[12]) * cospi_16_64);
+    vqrshrn.s32     d8, q13, #14              ; >> 14
+    vqrshrn.s32     d9, q6, #14               ; >> 14
+
+    mov              r4, #16                  ; pass1Output stride
+    ldr              r3, [sp]                 ; load skip_adding
+
+    ; stage 7
+    ; load the data in pass1
+    mov              r5, #24
+    mov              r3, #8
+
+    vld1.s16        {q0}, [r2], r4            ; load data step2[0]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[1]
+    vadd.s16        q12, q0, q15              ; step2[0] + step2[15]
+    vadd.s16        q13, q1, q14              ; step2[1] + step2[14]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[2]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[3]
+    vst1.64         {d24}, [r1], r3           ; store output[0]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[1]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q5              ; step2[2] + step2[13]
+    vadd.s16        q13, q11, q4              ; step2[3] + step2[12]
+    vsub.s16        q14, q1, q14              ; step2[1] - step2[14]
+    vsub.s16        q15, q0, q15              ; step2[0] - step2[15]
+    vst1.64         {d24}, [r1], r3           ; store output[2]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[3]
+    vst1.64         {d27}, [r1], r5
+    vsub.s16        q4, q11, q4               ; step2[3] - step2[12]
+    vsub.s16        q5, q10, q5               ; step2[2] - step2[13]
+    vld1.s16        {q0}, [r2], r4            ; load data step2[4]
+    vld1.s16        {q1}, [r2], r4            ; load data step2[5]
+    vadd.s16        q12, q0, q3               ; step2[4] + step2[11]
+    vadd.s16        q13, q1, q2               ; step2[5] + step2[10]
+    vld1.s16        {q10}, [r2], r4           ; load data step2[6]
+    vld1.s16        {q11}, [r2], r4           ; load data step2[7]
+    vst1.64         {d24}, [r1], r3           ; store output[4]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[5]
+    vst1.64         {d27}, [r1], r5
+    vadd.s16        q12, q10, q9              ; step2[6] + step2[9]
+    vadd.s16        q13, q11, q8              ; step2[7] + step2[8]
+    vsub.s16        q2, q1, q2                ; step2[5] - step2[10]
+    vsub.s16        q3, q0, q3                ; step2[4] - step2[11]
+    vsub.s16        q8, q11, q8               ; step2[7] - step2[8]
+    vsub.s16        q9, q10, q9               ; step2[6] - step2[9]
+    vst1.64         {d24}, [r1], r3           ; store output[6]
+    vst1.64         {d25}, [r1], r5
+    vst1.64         {d26}, [r1], r3           ; store output[7]
+    vst1.64         {d27}, [r1], r5
+
+    ; store the data  output 8,9,10,11,12,13,14,15
+    vst1.64         {d16}, [r1], r3
+    vst1.64         {d17}, [r1], r5
+    vst1.64         {d18}, [r1], r3
+    vst1.64         {d19}, [r1], r5
+    vst1.64         {d4}, [r1], r3
+    vst1.64         {d5}, [r1], r5
+    vst1.64         {d6}, [r1], r3
+    vst1.64         {d7}, [r1], r5
+    vst1.64         {d8}, [r1], r3
+    vst1.64         {d9}, [r1], r5
+    vst1.64         {d10}, [r1], r3
+    vst1.64         {d11}, [r1], r5
+    vst1.64         {d28}, [r1], r3
+    vst1.64         {d29}, [r1], r5
+    vst1.64         {d30}, [r1], r3
+    vst1.64         {d31}, [r1], r5
+end_idct10_16x16_pass2
+    pop             {r3-r9}
+    bx              lr
+    ENDP  ; |aom_idct16x16_10_add_neon_pass2|
+    END
diff --git a/third_party/aom/aom_dsp/arm/idct16x16_add_neon.c b/third_party/aom/aom_dsp/arm/idct16x16_add_neon.c
new file mode 100644
index 0000000000..b4cb7a0cd8
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct16x16_add_neon.c
@@ -0,0 +1,1295 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "aom_dsp/txfm_common.h"
+
+static INLINE void TRANSPOSE8X8(int16x8_t *q8s16, int16x8_t *q9s16,
+                                int16x8_t *q10s16, int16x8_t *q11s16,
+                                int16x8_t *q12s16, int16x8_t *q13s16,
+                                int16x8_t *q14s16, int16x8_t *q15s16) {
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+  int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+  d16s16 = vget_low_s16(*q8s16);
+  d17s16 = vget_high_s16(*q8s16);
+  d18s16 = vget_low_s16(*q9s16);
+  d19s16 = vget_high_s16(*q9s16);
+  d20s16 = vget_low_s16(*q10s16);
+  d21s16 = vget_high_s16(*q10s16);
+  d22s16 = vget_low_s16(*q11s16);
+  d23s16 = vget_high_s16(*q11s16);
+  d24s16 = vget_low_s16(*q12s16);
+  d25s16 = vget_high_s16(*q12s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+  d30s16 = vget_low_s16(*q15s16);
+  d31s16 = vget_high_s16(*q15s16);
+
+  *q8s16 = vcombine_s16(d16s16, d24s16);   // vswp d17, d24
+  *q9s16 = vcombine_s16(d18s16, d26s16);   // vswp d19, d26
+  *q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+  *q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+  *q12s16 = vcombine_s16(d17s16, d25s16);
+  *q13s16 = vcombine_s16(d19s16, d27s16);
+  *q14s16 = vcombine_s16(d21s16, d29s16);
+  *q15s16 = vcombine_s16(d23s16, d31s16);
+
+  q0x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q8s16), vreinterpretq_s32_s16(*q10s16));
+  q1x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q9s16), vreinterpretq_s32_s16(*q11s16));
+  q2x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q12s16), vreinterpretq_s32_s16(*q14s16));
+  q3x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q13s16), vreinterpretq_s32_s16(*q15s16));
+
+  q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                      vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+  q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                      vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+  q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                      vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+  q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                      vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+  *q8s16 = q0x2s16.val[0];
+  *q9s16 = q0x2s16.val[1];
+  *q10s16 = q1x2s16.val[0];
+  *q11s16 = q1x2s16.val[1];
+  *q12s16 = q2x2s16.val[0];
+  *q13s16 = q2x2s16.val[1];
+  *q14s16 = q3x2s16.val[0];
+  *q15s16 = q3x2s16.val[1];
+  return;
+}
+
+void aom_idct16x16_256_add_neon_pass1(int16_t *in, int16_t *out,
+                                      int output_stride) {
+  int16x4_t d0s16, d1s16, d2s16, d3s16;
+  int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  uint64x1_t d16u64, d17u64, d18u64, d19u64, d20u64, d21u64, d22u64, d23u64;
+  uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64;
+  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  int32x4_t q0s32, q1s32, q2s32, q3s32, q5s32, q6s32, q9s32;
+  int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
+  int16x8x2_t q0x2s16;
+
+  q0x2s16 = vld2q_s16(in);
+  q8s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q9s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q10s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q11s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q12s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q13s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q14s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q15s16 = q0x2s16.val[0];
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  d16s16 = vget_low_s16(q8s16);
+  d17s16 = vget_high_s16(q8s16);
+  d18s16 = vget_low_s16(q9s16);
+  d19s16 = vget_high_s16(q9s16);
+  d20s16 = vget_low_s16(q10s16);
+  d21s16 = vget_high_s16(q10s16);
+  d22s16 = vget_low_s16(q11s16);
+  d23s16 = vget_high_s16(q11s16);
+  d24s16 = vget_low_s16(q12s16);
+  d25s16 = vget_high_s16(q12s16);
+  d26s16 = vget_low_s16(q13s16);
+  d27s16 = vget_high_s16(q13s16);
+  d28s16 = vget_low_s16(q14s16);
+  d29s16 = vget_high_s16(q14s16);
+  d30s16 = vget_low_s16(q15s16);
+  d31s16 = vget_high_s16(q15s16);
+
+  // stage 3
+  d0s16 = vdup_n_s16((int16_t)cospi_28_64);
+  d1s16 = vdup_n_s16((int16_t)cospi_4_64);
+
+  q2s32 = vmull_s16(d18s16, d0s16);
+  q3s32 = vmull_s16(d19s16, d0s16);
+  q5s32 = vmull_s16(d18s16, d1s16);
+  q6s32 = vmull_s16(d19s16, d1s16);
+
+  q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
+  q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
+  q5s32 = vmlal_s16(q5s32, d30s16, d0s16);
+  q6s32 = vmlal_s16(q6s32, d31s16, d0s16);
+
+  d2s16 = vdup_n_s16((int16_t)cospi_12_64);
+  d3s16 = vdup_n_s16((int16_t)cospi_20_64);
+
+  d8s16 = vqrshrn_n_s32(q2s32, 14);
+  d9s16 = vqrshrn_n_s32(q3s32, 14);
+  d14s16 = vqrshrn_n_s32(q5s32, 14);
+  d15s16 = vqrshrn_n_s32(q6s32, 14);
+  q4s16 = vcombine_s16(d8s16, d9s16);
+  q7s16 = vcombine_s16(d14s16, d15s16);
+
+  q2s32 = vmull_s16(d26s16, d2s16);
+  q3s32 = vmull_s16(d27s16, d2s16);
+  q9s32 = vmull_s16(d26s16, d3s16);
+  q15s32 = vmull_s16(d27s16, d3s16);
+
+  q2s32 = vmlsl_s16(q2s32, d22s16, d3s16);
+  q3s32 = vmlsl_s16(q3s32, d23s16, d3s16);
+  q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
+  q15s32 = vmlal_s16(q15s32, d23s16, d2s16);
+
+  d10s16 = vqrshrn_n_s32(q2s32, 14);
+  d11s16 = vqrshrn_n_s32(q3s32, 14);
+  d12s16 = vqrshrn_n_s32(q9s32, 14);
+  d13s16 = vqrshrn_n_s32(q15s32, 14);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  // stage 4
+  d30s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q2s32 = vmull_s16(d16s16, d30s16);
+  q11s32 = vmull_s16(d17s16, d30s16);
+  q0s32 = vmull_s16(d24s16, d30s16);
+  q1s32 = vmull_s16(d25s16, d30s16);
+
+  d30s16 = vdup_n_s16((int16_t)cospi_24_64);
+  d31s16 = vdup_n_s16((int16_t)cospi_8_64);
+
+  q3s32 = vaddq_s32(q2s32, q0s32);
+  q12s32 = vaddq_s32(q11s32, q1s32);
+  q13s32 = vsubq_s32(q2s32, q0s32);
+  q1s32 = vsubq_s32(q11s32, q1s32);
+
+  d16s16 = vqrshrn_n_s32(q3s32, 14);
+  d17s16 = vqrshrn_n_s32(q12s32, 14);
+  d18s16 = vqrshrn_n_s32(q13s32, 14);
+  d19s16 = vqrshrn_n_s32(q1s32, 14);
+  q8s16 = vcombine_s16(d16s16, d17s16);
+  q9s16 = vcombine_s16(d18s16, d19s16);
+
+  q0s32 = vmull_s16(d20s16, d31s16);
+  q1s32 = vmull_s16(d21s16, d31s16);
+  q12s32 = vmull_s16(d20s16, d30s16);
+  q13s32 = vmull_s16(d21s16, d30s16);
+
+  q0s32 = vmlal_s16(q0s32, d28s16, d30s16);
+  q1s32 = vmlal_s16(q1s32, d29s16, d30s16);
+  q12s32 = vmlsl_s16(q12s32, d28s16, d31s16);
+  q13s32 = vmlsl_s16(q13s32, d29s16, d31s16);
+
+  d22s16 = vqrshrn_n_s32(q0s32, 14);
+  d23s16 = vqrshrn_n_s32(q1s32, 14);
+  d20s16 = vqrshrn_n_s32(q12s32, 14);
+  d21s16 = vqrshrn_n_s32(q13s32, 14);
+  q10s16 = vcombine_s16(d20s16, d21s16);
+  q11s16 = vcombine_s16(d22s16, d23s16);
+
+  q13s16 = vsubq_s16(q4s16, q5s16);
+  q4s16 = vaddq_s16(q4s16, q5s16);
+  q14s16 = vsubq_s16(q7s16, q6s16);
+  q15s16 = vaddq_s16(q6s16, q7s16);
+  d26s16 = vget_low_s16(q13s16);
+  d27s16 = vget_high_s16(q13s16);
+  d28s16 = vget_low_s16(q14s16);
+  d29s16 = vget_high_s16(q14s16);
+
+  // stage 5
+  q0s16 = vaddq_s16(q8s16, q11s16);
+  q1s16 = vaddq_s16(q9s16, q10s16);
+  q2s16 = vsubq_s16(q9s16, q10s16);
+  q3s16 = vsubq_s16(q8s16, q11s16);
+
+  d16s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q11s32 = vmull_s16(d26s16, d16s16);
+  q12s32 = vmull_s16(d27s16, d16s16);
+  q9s32 = vmull_s16(d28s16, d16s16);
+  q10s32 = vmull_s16(d29s16, d16s16);
+
+  q6s32 = vsubq_s32(q9s32, q11s32);
+  q13s32 = vsubq_s32(q10s32, q12s32);
+  q9s32 = vaddq_s32(q9s32, q11s32);
+  q10s32 = vaddq_s32(q10s32, q12s32);
+
+  d10s16 = vqrshrn_n_s32(q6s32, 14);
+  d11s16 = vqrshrn_n_s32(q13s32, 14);
+  d12s16 = vqrshrn_n_s32(q9s32, 14);
+  d13s16 = vqrshrn_n_s32(q10s32, 14);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  // stage 6
+  q8s16 = vaddq_s16(q0s16, q15s16);
+  q9s16 = vaddq_s16(q1s16, q6s16);
+  q10s16 = vaddq_s16(q2s16, q5s16);
+  q11s16 = vaddq_s16(q3s16, q4s16);
+  q12s16 = vsubq_s16(q3s16, q4s16);
+  q13s16 = vsubq_s16(q2s16, q5s16);
+  q14s16 = vsubq_s16(q1s16, q6s16);
+  q15s16 = vsubq_s16(q0s16, q15s16);
+
+  d16u64 = vreinterpret_u64_s16(vget_low_s16(q8s16));
+  d17u64 = vreinterpret_u64_s16(vget_high_s16(q8s16));
+  d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16));
+  d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16));
+  d20u64 = vreinterpret_u64_s16(vget_low_s16(q10s16));
+  d21u64 = vreinterpret_u64_s16(vget_high_s16(q10s16));
+  d22u64 = vreinterpret_u64_s16(vget_low_s16(q11s16));
+  d23u64 = vreinterpret_u64_s16(vget_high_s16(q11s16));
+  d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+  d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+  d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+  d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+  d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16));
+  d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16));
+  d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16));
+  d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16));
+
+  // store the data
+  output_stride >>= 1;  // output_stride / 2, out is int16_t
+  vst1_u64((uint64_t *)out, d16u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d17u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d18u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d19u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d20u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d21u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d22u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d23u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d24u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d25u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d26u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d27u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d28u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d29u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d30u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d31u64);
+  return;
+}
+
+void aom_idct16x16_256_add_neon_pass2(int16_t *src, int16_t *out,
+                                      int16_t *pass1Output, int16_t skip_adding,
+                                      uint8_t *dest, int dest_stride) {
+  uint8_t *d;
+  uint8x8_t d12u8, d13u8;
+  int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+  int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  uint64x1_t d24u64, d25u64, d26u64, d27u64;
+  int64x1_t d12s64, d13s64;
+  uint16x8_t q2u16, q3u16, q4u16, q5u16, q8u16;
+  uint16x8_t q9u16, q12u16, q13u16, q14u16, q15u16;
+  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q8s32, q9s32;
+  int32x4_t q10s32, q11s32, q12s32, q13s32;
+  int16x8x2_t q0x2s16;
+
+  q0x2s16 = vld2q_s16(src);
+  q8s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q9s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q10s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q11s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q12s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q13s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q14s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q15s16 = q0x2s16.val[0];
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  d16s16 = vget_low_s16(q8s16);
+  d17s16 = vget_high_s16(q8s16);
+  d18s16 = vget_low_s16(q9s16);
+  d19s16 = vget_high_s16(q9s16);
+  d20s16 = vget_low_s16(q10s16);
+  d21s16 = vget_high_s16(q10s16);
+  d22s16 = vget_low_s16(q11s16);
+  d23s16 = vget_high_s16(q11s16);
+  d24s16 = vget_low_s16(q12s16);
+  d25s16 = vget_high_s16(q12s16);
+  d26s16 = vget_low_s16(q13s16);
+  d27s16 = vget_high_s16(q13s16);
+  d28s16 = vget_low_s16(q14s16);
+  d29s16 = vget_high_s16(q14s16);
+  d30s16 = vget_low_s16(q15s16);
+  d31s16 = vget_high_s16(q15s16);
+
+  // stage 3
+  d12s16 = vdup_n_s16((int16_t)cospi_30_64);
+  d13s16 = vdup_n_s16((int16_t)cospi_2_64);
+
+  q2s32 = vmull_s16(d16s16, d12s16);
+  q3s32 = vmull_s16(d17s16, d12s16);
+  q1s32 = vmull_s16(d16s16, d13s16);
+  q4s32 = vmull_s16(d17s16, d13s16);
+
+  q2s32 = vmlsl_s16(q2s32, d30s16, d13s16);
+  q3s32 = vmlsl_s16(q3s32, d31s16, d13s16);
+  q1s32 = vmlal_s16(q1s32, d30s16, d12s16);
+  q4s32 = vmlal_s16(q4s32, d31s16, d12s16);
+
+  d0s16 = vqrshrn_n_s32(q2s32, 14);
+  d1s16 = vqrshrn_n_s32(q3s32, 14);
+  d14s16 = vqrshrn_n_s32(q1s32, 14);
+  d15s16 = vqrshrn_n_s32(q4s32, 14);
+  q0s16 = vcombine_s16(d0s16, d1s16);
+  q7s16 = vcombine_s16(d14s16, d15s16);
+
+  d30s16 = vdup_n_s16((int16_t)cospi_14_64);
+  d31s16 = vdup_n_s16((int16_t)cospi_18_64);
+
+  q2s32 = vmull_s16(d24s16, d30s16);
+  q3s32 = vmull_s16(d25s16, d30s16);
+  q4s32 = vmull_s16(d24s16, d31s16);
+  q5s32 = vmull_s16(d25s16, d31s16);
+
+  q2s32 = vmlsl_s16(q2s32, d22s16, d31s16);
+  q3s32 = vmlsl_s16(q3s32, d23s16, d31s16);
+  q4s32 = vmlal_s16(q4s32, d22s16, d30s16);
+  q5s32 = vmlal_s16(q5s32, d23s16, d30s16);
+
+  d2s16 = vqrshrn_n_s32(q2s32, 14);
+  d3s16 = vqrshrn_n_s32(q3s32, 14);
+  d12s16 = vqrshrn_n_s32(q4s32, 14);
+  d13s16 = vqrshrn_n_s32(q5s32, 14);
+  q1s16 = vcombine_s16(d2s16, d3s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  d30s16 = vdup_n_s16((int16_t)cospi_22_64);
+  d31s16 = vdup_n_s16((int16_t)cospi_10_64);
+
+  q11s32 = vmull_s16(d20s16, d30s16);
+  q12s32 = vmull_s16(d21s16, d30s16);
+  q4s32 = vmull_s16(d20s16, d31s16);
+  q5s32 = vmull_s16(d21s16, d31s16);
+
+  q11s32 = vmlsl_s16(q11s32, d26s16, d31s16);
+  q12s32 = vmlsl_s16(q12s32, d27s16, d31s16);
+  q4s32 = vmlal_s16(q4s32, d26s16, d30s16);
+  q5s32 = vmlal_s16(q5s32, d27s16, d30s16);
+
+  d4s16 = vqrshrn_n_s32(q11s32, 14);
+  d5s16 = vqrshrn_n_s32(q12s32, 14);
+  d11s16 = vqrshrn_n_s32(q5s32, 14);
+  d10s16 = vqrshrn_n_s32(q4s32, 14);
+  q2s16 = vcombine_s16(d4s16, d5s16);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+
+  d30s16 = vdup_n_s16((int16_t)cospi_6_64);
+  d31s16 = vdup_n_s16((int16_t)cospi_26_64);
+
+  q10s32 = vmull_s16(d28s16, d30s16);
+  q11s32 = vmull_s16(d29s16, d30s16);
+  q12s32 = vmull_s16(d28s16, d31s16);
+  q13s32 = vmull_s16(d29s16, d31s16);
+
+  q10s32 = vmlsl_s16(q10s32, d18s16, d31s16);
+  q11s32 = vmlsl_s16(q11s32, d19s16, d31s16);
+  q12s32 = vmlal_s16(q12s32, d18s16, d30s16);
+  q13s32 = vmlal_s16(q13s32, d19s16, d30s16);
+
+  d6s16 = vqrshrn_n_s32(q10s32, 14);
+  d7s16 = vqrshrn_n_s32(q11s32, 14);
+  d8s16 = vqrshrn_n_s32(q12s32, 14);
+  d9s16 = vqrshrn_n_s32(q13s32, 14);
+  q3s16 = vcombine_s16(d6s16, d7s16);
+  q4s16 = vcombine_s16(d8s16, d9s16);
+
+  // stage 3
+  q9s16 = vsubq_s16(q0s16, q1s16);
+  q0s16 = vaddq_s16(q0s16, q1s16);
+  q10s16 = vsubq_s16(q3s16, q2s16);
+  q11s16 = vaddq_s16(q2s16, q3s16);
+  q12s16 = vaddq_s16(q4s16, q5s16);
+  q13s16 = vsubq_s16(q4s16, q5s16);
+  q14s16 = vsubq_s16(q7s16, q6s16);
+  q7s16 = vaddq_s16(q6s16, q7s16);
+
+  // stage 4
+  d18s16 = vget_low_s16(q9s16);
+  d19s16 = vget_high_s16(q9s16);
+  d20s16 = vget_low_s16(q10s16);
+  d21s16 = vget_high_s16(q10s16);
+  d26s16 = vget_low_s16(q13s16);
+  d27s16 = vget_high_s16(q13s16);
+  d28s16 = vget_low_s16(q14s16);
+  d29s16 = vget_high_s16(q14s16);
+
+  d30s16 = vdup_n_s16((int16_t)cospi_8_64);
+  d31s16 = vdup_n_s16((int16_t)cospi_24_64);
+
+  q2s32 = vmull_s16(d18s16, d31s16);
+  q3s32 = vmull_s16(d19s16, d31s16);
+  q4s32 = vmull_s16(d28s16, d31s16);
+  q5s32 = vmull_s16(d29s16, d31s16);
+
+  q2s32 = vmlal_s16(q2s32, d28s16, d30s16);
+  q3s32 = vmlal_s16(q3s32, d29s16, d30s16);
+  q4s32 = vmlsl_s16(q4s32, d18s16, d30s16);
+  q5s32 = vmlsl_s16(q5s32, d19s16, d30s16);
+
+  d12s16 = vqrshrn_n_s32(q2s32, 14);
+  d13s16 = vqrshrn_n_s32(q3s32, 14);
+  d2s16 = vqrshrn_n_s32(q4s32, 14);
+  d3s16 = vqrshrn_n_s32(q5s32, 14);
+  q1s16 = vcombine_s16(d2s16, d3s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  q3s16 = q11s16;
+  q4s16 = q12s16;
+
+  d30s16 = vdup_n_s16(-cospi_8_64);
+  q11s32 = vmull_s16(d26s16, d30s16);
+  q12s32 = vmull_s16(d27s16, d30s16);
+  q8s32 = vmull_s16(d20s16, d30s16);
+  q9s32 = vmull_s16(d21s16, d30s16);
+
+  q11s32 = vmlsl_s16(q11s32, d20s16, d31s16);
+  q12s32 = vmlsl_s16(q12s32, d21s16, d31s16);
+  q8s32 = vmlal_s16(q8s32, d26s16, d31s16);
+  q9s32 = vmlal_s16(q9s32, d27s16, d31s16);
+
+  d4s16 = vqrshrn_n_s32(q11s32, 14);
+  d5s16 = vqrshrn_n_s32(q12s32, 14);
+  d10s16 = vqrshrn_n_s32(q8s32, 14);
+  d11s16 = vqrshrn_n_s32(q9s32, 14);
+  q2s16 = vcombine_s16(d4s16, d5s16);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+
+  // stage 5
+  q8s16 = vaddq_s16(q0s16, q3s16);
+  q9s16 = vaddq_s16(q1s16, q2s16);
+  q10s16 = vsubq_s16(q1s16, q2s16);
+  q11s16 = vsubq_s16(q0s16, q3s16);
+  q12s16 = vsubq_s16(q7s16, q4s16);
+  q13s16 = vsubq_s16(q6s16, q5s16);
+  q14s16 = vaddq_s16(q6s16, q5s16);
+  q15s16 = vaddq_s16(q7s16, q4s16);
+
+  // stage 6
+  d20s16 = vget_low_s16(q10s16);
+  d21s16 = vget_high_s16(q10s16);
+  d22s16 = vget_low_s16(q11s16);
+  d23s16 = vget_high_s16(q11s16);
+  d24s16 = vget_low_s16(q12s16);
+  d25s16 = vget_high_s16(q12s16);
+  d26s16 = vget_low_s16(q13s16);
+  d27s16 = vget_high_s16(q13s16);
+
+  d14s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q3s32 = vmull_s16(d26s16, d14s16);
+  q4s32 = vmull_s16(d27s16, d14s16);
+  q0s32 = vmull_s16(d20s16, d14s16);
+  q1s32 = vmull_s16(d21s16, d14s16);
+
+  q5s32 = vsubq_s32(q3s32, q0s32);
+  q6s32 = vsubq_s32(q4s32, q1s32);
+  q10s32 = vaddq_s32(q3s32, q0s32);
+  q4s32 = vaddq_s32(q4s32, q1s32);
+
+  d4s16 = vqrshrn_n_s32(q5s32, 14);
+  d5s16 = vqrshrn_n_s32(q6s32, 14);
+  d10s16 = vqrshrn_n_s32(q10s32, 14);
+  d11s16 = vqrshrn_n_s32(q4s32, 14);
+  q2s16 = vcombine_s16(d4s16, d5s16);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+
+  q0s32 = vmull_s16(d22s16, d14s16);
+  q1s32 = vmull_s16(d23s16, d14s16);
+  q13s32 = vmull_s16(d24s16, d14s16);
+  q6s32 = vmull_s16(d25s16, d14s16);
+
+  q10s32 = vsubq_s32(q13s32, q0s32);
+  q4s32 = vsubq_s32(q6s32, q1s32);
+  q13s32 = vaddq_s32(q13s32, q0s32);
+  q6s32 = vaddq_s32(q6s32, q1s32);
+
+  d6s16 = vqrshrn_n_s32(q10s32, 14);
+  d7s16 = vqrshrn_n_s32(q4s32, 14);
+  d8s16 = vqrshrn_n_s32(q13s32, 14);
+  d9s16 = vqrshrn_n_s32(q6s32, 14);
+  q3s16 = vcombine_s16(d6s16, d7s16);
+  q4s16 = vcombine_s16(d8s16, d9s16);
+
+  // stage 7
+  if (skip_adding != 0) {
+    d = dest;
+    // load the data in pass1
+    q0s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q1s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    d13s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+
+    q12s16 = vaddq_s16(q0s16, q15s16);
+    q13s16 = vaddq_s16(q1s16, q14s16);
+    q12s16 = vrshrq_n_s16(q12s16, 6);
+    q13s16 = vrshrq_n_s16(q13s16, 6);
+    q12u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64));
+    q13u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+    d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+    d += dest_stride;
+    q14s16 = vsubq_s16(q1s16, q14s16);
+    q15s16 = vsubq_s16(q0s16, q15s16);
+
+    q10s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q11s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    d13s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    q12s16 = vaddq_s16(q10s16, q5s16);
+    q13s16 = vaddq_s16(q11s16, q4s16);
+    q12s16 = vrshrq_n_s16(q12s16, 6);
+    q13s16 = vrshrq_n_s16(q13s16, 6);
+    q12u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64));
+    q13u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+    d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+    d += dest_stride;
+    q4s16 = vsubq_s16(q11s16, q4s16);
+    q5s16 = vsubq_s16(q10s16, q5s16);
+
+    q0s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q1s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    d13s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    q12s16 = vaddq_s16(q0s16, q3s16);
+    q13s16 = vaddq_s16(q1s16, q2s16);
+    q12s16 = vrshrq_n_s16(q12s16, 6);
+    q13s16 = vrshrq_n_s16(q13s16, 6);
+    q12u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64));
+    q13u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+    d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+    d += dest_stride;
+    q2s16 = vsubq_s16(q1s16, q2s16);
+    q3s16 = vsubq_s16(q0s16, q3s16);
+
+    q10s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q11s16 = vld1q_s16(pass1Output);
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    d13s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    q12s16 = vaddq_s16(q10s16, q9s16);
+    q13s16 = vaddq_s16(q11s16, q8s16);
+    q12s16 = vrshrq_n_s16(q12s16, 6);
+    q13s16 = vrshrq_n_s16(q13s16, 6);
+    q12u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q12s16), vreinterpret_u8_s64(d12s64));
+    q13u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q13s16), vreinterpret_u8_s64(d13s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+    d13u8 = vqmovun_s16(vreinterpretq_s16_u16(q13u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d13u8));
+    d += dest_stride;
+    q8s16 = vsubq_s16(q11s16, q8s16);
+    q9s16 = vsubq_s16(q10s16, q9s16);
+
+    // store the data  out 8,9,10,11,12,13,14,15
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    q8s16 = vrshrq_n_s16(q8s16, 6);
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_s64(d12s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    q9s16 = vrshrq_n_s16(q9s16, 6);
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_s64(d12s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    q2s16 = vrshrq_n_s16(q2s16, 6);
+    q2u16 = vaddw_u8(vreinterpretq_u16_s16(q2s16), vreinterpret_u8_s64(d12s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q2u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    q3s16 = vrshrq_n_s16(q3s16, 6);
+    q3u16 = vaddw_u8(vreinterpretq_u16_s16(q3s16), vreinterpret_u8_s64(d12s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q3u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    q4s16 = vrshrq_n_s16(q4s16, 6);
+    q4u16 = vaddw_u8(vreinterpretq_u16_s16(q4s16), vreinterpret_u8_s64(d12s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q4u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    q5s16 = vrshrq_n_s16(q5s16, 6);
+    q5u16 = vaddw_u8(vreinterpretq_u16_s16(q5s16), vreinterpret_u8_s64(d12s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q5u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+
+    d12s64 = vld1_s64((int64_t *)dest);
+    dest += dest_stride;
+    q14s16 = vrshrq_n_s16(q14s16, 6);
+    q14u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q14s16), vreinterpret_u8_s64(d12s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q14u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+    d += dest_stride;
+
+    d12s64 = vld1_s64((int64_t *)dest);
+    q15s16 = vrshrq_n_s16(q15s16, 6);
+    q15u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q15s16), vreinterpret_u8_s64(d12s64));
+    d12u8 = vqmovun_s16(vreinterpretq_s16_u16(q15u16));
+    vst1_u64((uint64_t *)d, vreinterpret_u64_u8(d12u8));
+  } else {  // skip_adding_dest
+    q0s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q1s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q12s16 = vaddq_s16(q0s16, q15s16);
+    q13s16 = vaddq_s16(q1s16, q14s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q14s16 = vsubq_s16(q1s16, q14s16);
+    q15s16 = vsubq_s16(q0s16, q15s16);
+
+    q10s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q11s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q12s16 = vaddq_s16(q10s16, q5s16);
+    q13s16 = vaddq_s16(q11s16, q4s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q4s16 = vsubq_s16(q11s16, q4s16);
+    q5s16 = vsubq_s16(q10s16, q5s16);
+
+    q0s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q1s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q12s16 = vaddq_s16(q0s16, q3s16);
+    q13s16 = vaddq_s16(q1s16, q2s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q2s16 = vsubq_s16(q1s16, q2s16);
+    q3s16 = vsubq_s16(q0s16, q3s16);
+
+    q10s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q11s16 = vld1q_s16(pass1Output);
+    pass1Output += 8;
+    q12s16 = vaddq_s16(q10s16, q9s16);
+    q13s16 = vaddq_s16(q11s16, q8s16);
+    d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+    d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+    d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+    d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+    vst1_u64((uint64_t *)out, d24u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d25u64);
+    out += 12;
+    vst1_u64((uint64_t *)out, d26u64);
+    out += 4;
+    vst1_u64((uint64_t *)out, d27u64);
+    out += 12;
+    q8s16 = vsubq_s16(q11s16, q8s16);
+    q9s16 = vsubq_s16(q10s16, q9s16);
+
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q8s16)));
+    out += 4;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q8s16)));
+    out += 12;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q9s16)));
+    out += 4;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q9s16)));
+    out += 12;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q2s16)));
+    out += 4;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q2s16)));
+    out += 12;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q3s16)));
+    out += 4;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q3s16)));
+    out += 12;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q4s16)));
+    out += 4;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q4s16)));
+    out += 12;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q5s16)));
+    out += 4;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q5s16)));
+    out += 12;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q14s16)));
+    out += 4;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q14s16)));
+    out += 12;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_low_s16(q15s16)));
+    out += 4;
+    vst1_u64((uint64_t *)out, vreinterpret_u64_s16(vget_high_s16(q15s16)));
+  }
+  return;
+}
+
+void aom_idct16x16_10_add_neon_pass1(int16_t *in, int16_t *out,
+                                     int output_stride) {
+  int16x4_t d4s16;
+  int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+  uint64x1_t d4u64, d5u64, d18u64, d19u64, d20u64, d21u64, d22u64, d23u64;
+  uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64;
+  int16x8_t q0s16, q1s16, q2s16, q4s16, q5s16, q6s16, q7s16;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  int32x4_t q6s32, q9s32;
+  int32x4_t q10s32, q11s32, q12s32, q15s32;
+  int16x8x2_t q0x2s16;
+
+  q0x2s16 = vld2q_s16(in);
+  q8s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q9s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q10s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q11s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q12s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q13s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q14s16 = q0x2s16.val[0];
+  in += 16;
+  q0x2s16 = vld2q_s16(in);
+  q15s16 = q0x2s16.val[0];
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  // stage 3
+  q0s16 = vdupq_n_s16((int16_t)(cospi_28_64 * 2));
+  q1s16 = vdupq_n_s16((int16_t)(cospi_4_64 * 2));
+
+  q4s16 = vqrdmulhq_s16(q9s16, q0s16);
+  q7s16 = vqrdmulhq_s16(q9s16, q1s16);
+
+  // stage 4
+  q1s16 = vdupq_n_s16((int16_t)(cospi_16_64 * 2));
+  d4s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q8s16 = vqrdmulhq_s16(q8s16, q1s16);
+
+  d8s16 = vget_low_s16(q4s16);
+  d9s16 = vget_high_s16(q4s16);
+  d14s16 = vget_low_s16(q7s16);
+  d15s16 = vget_high_s16(q7s16);
+  q9s32 = vmull_s16(d14s16, d4s16);
+  q10s32 = vmull_s16(d15s16, d4s16);
+  q12s32 = vmull_s16(d9s16, d4s16);
+  q11s32 = vmull_s16(d8s16, d4s16);
+
+  q15s32 = vsubq_s32(q10s32, q12s32);
+  q6s32 = vsubq_s32(q9s32, q11s32);
+  q9s32 = vaddq_s32(q9s32, q11s32);
+  q10s32 = vaddq_s32(q10s32, q12s32);
+
+  d11s16 = vqrshrn_n_s32(q15s32, 14);
+  d10s16 = vqrshrn_n_s32(q6s32, 14);
+  d12s16 = vqrshrn_n_s32(q9s32, 14);
+  d13s16 = vqrshrn_n_s32(q10s32, 14);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  // stage 6
+  q2s16 = vaddq_s16(q8s16, q7s16);
+  q9s16 = vaddq_s16(q8s16, q6s16);
+  q10s16 = vaddq_s16(q8s16, q5s16);
+  q11s16 = vaddq_s16(q8s16, q4s16);
+  q12s16 = vsubq_s16(q8s16, q4s16);
+  q13s16 = vsubq_s16(q8s16, q5s16);
+  q14s16 = vsubq_s16(q8s16, q6s16);
+  q15s16 = vsubq_s16(q8s16, q7s16);
+
+  d4u64 = vreinterpret_u64_s16(vget_low_s16(q2s16));
+  d5u64 = vreinterpret_u64_s16(vget_high_s16(q2s16));
+  d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16));
+  d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16));
+  d20u64 = vreinterpret_u64_s16(vget_low_s16(q10s16));
+  d21u64 = vreinterpret_u64_s16(vget_high_s16(q10s16));
+  d22u64 = vreinterpret_u64_s16(vget_low_s16(q11s16));
+  d23u64 = vreinterpret_u64_s16(vget_high_s16(q11s16));
+  d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+  d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+  d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+  d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+  d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16));
+  d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16));
+  d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16));
+  d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16));
+
+  // store the data
+  output_stride >>= 1;  // output_stride / 2, out is int16_t
+  vst1_u64((uint64_t *)out, d4u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d5u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d18u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d19u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d20u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d21u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d22u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d23u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d24u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d25u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d26u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d27u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d28u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d29u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d30u64);
+  out += output_stride;
+  vst1_u64((uint64_t *)out, d31u64);
+  return;
+}
+
+void aom_idct16x16_10_add_neon_pass2(int16_t *src, int16_t *out,
+                                     int16_t *pass1Output, int16_t skip_adding,
+                                     uint8_t *dest, int dest_stride) {
+  int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+  int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+  int16x4_t d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d30s16, d31s16;
+  uint64x1_t d4u64, d5u64, d6u64, d7u64, d8u64, d9u64, d10u64, d11u64;
+  uint64x1_t d16u64, d17u64, d18u64, d19u64;
+  uint64x1_t d24u64, d25u64, d26u64, d27u64, d28u64, d29u64, d30u64, d31u64;
+  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q8s32, q9s32;
+  int32x4_t q10s32, q11s32, q12s32, q13s32;
+  int16x8x2_t q0x2s16;
+  (void)skip_adding;
+  (void)dest;
+  (void)dest_stride;
+
+  q0x2s16 = vld2q_s16(src);
+  q8s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q9s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q10s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q11s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q12s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q13s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q14s16 = q0x2s16.val[0];
+  src += 16;
+  q0x2s16 = vld2q_s16(src);
+  q15s16 = q0x2s16.val[0];
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  // stage 3
+  q6s16 = vdupq_n_s16((int16_t)(cospi_30_64 * 2));
+  q0s16 = vqrdmulhq_s16(q8s16, q6s16);
+  q6s16 = vdupq_n_s16((int16_t)(cospi_2_64 * 2));
+  q7s16 = vqrdmulhq_s16(q8s16, q6s16);
+
+  q15s16 = vdupq_n_s16(-cospi_26_64 * 2);
+  q14s16 = vdupq_n_s16((int16_t)(cospi_6_64 * 2));
+  q3s16 = vqrdmulhq_s16(q9s16, q15s16);
+  q4s16 = vqrdmulhq_s16(q9s16, q14s16);
+
+  // stage 4
+  d0s16 = vget_low_s16(q0s16);
+  d1s16 = vget_high_s16(q0s16);
+  d6s16 = vget_low_s16(q3s16);
+  d7s16 = vget_high_s16(q3s16);
+  d8s16 = vget_low_s16(q4s16);
+  d9s16 = vget_high_s16(q4s16);
+  d14s16 = vget_low_s16(q7s16);
+  d15s16 = vget_high_s16(q7s16);
+
+  d30s16 = vdup_n_s16((int16_t)cospi_8_64);
+  d31s16 = vdup_n_s16((int16_t)cospi_24_64);
+
+  q12s32 = vmull_s16(d14s16, d31s16);
+  q5s32 = vmull_s16(d15s16, d31s16);
+  q2s32 = vmull_s16(d0s16, d31s16);
+  q11s32 = vmull_s16(d1s16, d31s16);
+
+  q12s32 = vmlsl_s16(q12s32, d0s16, d30s16);
+  q5s32 = vmlsl_s16(q5s32, d1s16, d30s16);
+  q2s32 = vmlal_s16(q2s32, d14s16, d30s16);
+  q11s32 = vmlal_s16(q11s32, d15s16, d30s16);
+
+  d2s16 = vqrshrn_n_s32(q12s32, 14);
+  d3s16 = vqrshrn_n_s32(q5s32, 14);
+  d12s16 = vqrshrn_n_s32(q2s32, 14);
+  d13s16 = vqrshrn_n_s32(q11s32, 14);
+  q1s16 = vcombine_s16(d2s16, d3s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  d30s16 = vdup_n_s16(-cospi_8_64);
+  q10s32 = vmull_s16(d8s16, d30s16);
+  q13s32 = vmull_s16(d9s16, d30s16);
+  q8s32 = vmull_s16(d6s16, d30s16);
+  q9s32 = vmull_s16(d7s16, d30s16);
+
+  q10s32 = vmlsl_s16(q10s32, d6s16, d31s16);
+  q13s32 = vmlsl_s16(q13s32, d7s16, d31s16);
+  q8s32 = vmlal_s16(q8s32, d8s16, d31s16);
+  q9s32 = vmlal_s16(q9s32, d9s16, d31s16);
+
+  d4s16 = vqrshrn_n_s32(q10s32, 14);
+  d5s16 = vqrshrn_n_s32(q13s32, 14);
+  d10s16 = vqrshrn_n_s32(q8s32, 14);
+  d11s16 = vqrshrn_n_s32(q9s32, 14);
+  q2s16 = vcombine_s16(d4s16, d5s16);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+
+  // stage 5
+  q8s16 = vaddq_s16(q0s16, q3s16);
+  q9s16 = vaddq_s16(q1s16, q2s16);
+  q10s16 = vsubq_s16(q1s16, q2s16);
+  q11s16 = vsubq_s16(q0s16, q3s16);
+  q12s16 = vsubq_s16(q7s16, q4s16);
+  q13s16 = vsubq_s16(q6s16, q5s16);
+  q14s16 = vaddq_s16(q6s16, q5s16);
+  q15s16 = vaddq_s16(q7s16, q4s16);
+
+  // stage 6
+  d20s16 = vget_low_s16(q10s16);
+  d21s16 = vget_high_s16(q10s16);
+  d22s16 = vget_low_s16(q11s16);
+  d23s16 = vget_high_s16(q11s16);
+  d24s16 = vget_low_s16(q12s16);
+  d25s16 = vget_high_s16(q12s16);
+  d26s16 = vget_low_s16(q13s16);
+  d27s16 = vget_high_s16(q13s16);
+
+  d14s16 = vdup_n_s16((int16_t)cospi_16_64);
+  q3s32 = vmull_s16(d26s16, d14s16);
+  q4s32 = vmull_s16(d27s16, d14s16);
+  q0s32 = vmull_s16(d20s16, d14s16);
+  q1s32 = vmull_s16(d21s16, d14s16);
+
+  q5s32 = vsubq_s32(q3s32, q0s32);
+  q6s32 = vsubq_s32(q4s32, q1s32);
+  q0s32 = vaddq_s32(q3s32, q0s32);
+  q4s32 = vaddq_s32(q4s32, q1s32);
+
+  d4s16 = vqrshrn_n_s32(q5s32, 14);
+  d5s16 = vqrshrn_n_s32(q6s32, 14);
+  d10s16 = vqrshrn_n_s32(q0s32, 14);
+  d11s16 = vqrshrn_n_s32(q4s32, 14);
+  q2s16 = vcombine_s16(d4s16, d5s16);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+
+  q0s32 = vmull_s16(d22s16, d14s16);
+  q1s32 = vmull_s16(d23s16, d14s16);
+  q13s32 = vmull_s16(d24s16, d14s16);
+  q6s32 = vmull_s16(d25s16, d14s16);
+
+  q10s32 = vsubq_s32(q13s32, q0s32);
+  q4s32 = vsubq_s32(q6s32, q1s32);
+  q13s32 = vaddq_s32(q13s32, q0s32);
+  q6s32 = vaddq_s32(q6s32, q1s32);
+
+  d6s16 = vqrshrn_n_s32(q10s32, 14);
+  d7s16 = vqrshrn_n_s32(q4s32, 14);
+  d8s16 = vqrshrn_n_s32(q13s32, 14);
+  d9s16 = vqrshrn_n_s32(q6s32, 14);
+  q3s16 = vcombine_s16(d6s16, d7s16);
+  q4s16 = vcombine_s16(d8s16, d9s16);
+
+  // stage 7
+  q0s16 = vld1q_s16(pass1Output);
+  pass1Output += 8;
+  q1s16 = vld1q_s16(pass1Output);
+  pass1Output += 8;
+  q12s16 = vaddq_s16(q0s16, q15s16);
+  q13s16 = vaddq_s16(q1s16, q14s16);
+  d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+  d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+  d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+  d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+  vst1_u64((uint64_t *)out, d24u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d25u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d26u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d27u64);
+  out += 12;
+  q14s16 = vsubq_s16(q1s16, q14s16);
+  q15s16 = vsubq_s16(q0s16, q15s16);
+
+  q10s16 = vld1q_s16(pass1Output);
+  pass1Output += 8;
+  q11s16 = vld1q_s16(pass1Output);
+  pass1Output += 8;
+  q12s16 = vaddq_s16(q10s16, q5s16);
+  q13s16 = vaddq_s16(q11s16, q4s16);
+  d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+  d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+  d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+  d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+  vst1_u64((uint64_t *)out, d24u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d25u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d26u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d27u64);
+  out += 12;
+  q4s16 = vsubq_s16(q11s16, q4s16);
+  q5s16 = vsubq_s16(q10s16, q5s16);
+
+  q0s16 = vld1q_s16(pass1Output);
+  pass1Output += 8;
+  q1s16 = vld1q_s16(pass1Output);
+  pass1Output += 8;
+  q12s16 = vaddq_s16(q0s16, q3s16);
+  q13s16 = vaddq_s16(q1s16, q2s16);
+  d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+  d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+  d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+  d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+  vst1_u64((uint64_t *)out, d24u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d25u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d26u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d27u64);
+  out += 12;
+  q2s16 = vsubq_s16(q1s16, q2s16);
+  q3s16 = vsubq_s16(q0s16, q3s16);
+
+  q10s16 = vld1q_s16(pass1Output);
+  pass1Output += 8;
+  q11s16 = vld1q_s16(pass1Output);
+  q12s16 = vaddq_s16(q10s16, q9s16);
+  q13s16 = vaddq_s16(q11s16, q8s16);
+  d24u64 = vreinterpret_u64_s16(vget_low_s16(q12s16));
+  d25u64 = vreinterpret_u64_s16(vget_high_s16(q12s16));
+  d26u64 = vreinterpret_u64_s16(vget_low_s16(q13s16));
+  d27u64 = vreinterpret_u64_s16(vget_high_s16(q13s16));
+  vst1_u64((uint64_t *)out, d24u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d25u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d26u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d27u64);
+  out += 12;
+  q8s16 = vsubq_s16(q11s16, q8s16);
+  q9s16 = vsubq_s16(q10s16, q9s16);
+
+  d4u64 = vreinterpret_u64_s16(vget_low_s16(q2s16));
+  d5u64 = vreinterpret_u64_s16(vget_high_s16(q2s16));
+  d6u64 = vreinterpret_u64_s16(vget_low_s16(q3s16));
+  d7u64 = vreinterpret_u64_s16(vget_high_s16(q3s16));
+  d8u64 = vreinterpret_u64_s16(vget_low_s16(q4s16));
+  d9u64 = vreinterpret_u64_s16(vget_high_s16(q4s16));
+  d10u64 = vreinterpret_u64_s16(vget_low_s16(q5s16));
+  d11u64 = vreinterpret_u64_s16(vget_high_s16(q5s16));
+  d16u64 = vreinterpret_u64_s16(vget_low_s16(q8s16));
+  d17u64 = vreinterpret_u64_s16(vget_high_s16(q8s16));
+  d18u64 = vreinterpret_u64_s16(vget_low_s16(q9s16));
+  d19u64 = vreinterpret_u64_s16(vget_high_s16(q9s16));
+  d28u64 = vreinterpret_u64_s16(vget_low_s16(q14s16));
+  d29u64 = vreinterpret_u64_s16(vget_high_s16(q14s16));
+  d30u64 = vreinterpret_u64_s16(vget_low_s16(q15s16));
+  d31u64 = vreinterpret_u64_s16(vget_high_s16(q15s16));
+
+  vst1_u64((uint64_t *)out, d16u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d17u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d18u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d19u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d4u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d5u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d6u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d7u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d8u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d9u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d10u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d11u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d28u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d29u64);
+  out += 12;
+  vst1_u64((uint64_t *)out, d30u64);
+  out += 4;
+  vst1_u64((uint64_t *)out, d31u64);
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/idct16x16_neon.c b/third_party/aom/aom_dsp/arm/idct16x16_neon.c
new file mode 100644
index 0000000000..db0d4905b5
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct16x16_neon.c
@@ -0,0 +1,152 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_dsp_common.h"
+
+void aom_idct16x16_256_add_neon_pass1(const int16_t *input, int16_t *output,
+                                      int output_stride);
+void aom_idct16x16_256_add_neon_pass2(const int16_t *src, int16_t *output,
+                                      int16_t *pass1Output, int16_t skip_adding,
+                                      uint8_t *dest, int dest_stride);
+void aom_idct16x16_10_add_neon_pass1(const int16_t *input, int16_t *output,
+                                     int output_stride);
+void aom_idct16x16_10_add_neon_pass2(const int16_t *src, int16_t *output,
+                                     int16_t *pass1Output, int16_t skip_adding,
+                                     uint8_t *dest, int dest_stride);
+
+#if HAVE_NEON_ASM
+/* For ARM NEON, d8-d15 are callee-saved registers, and need to be saved. */
+extern void aom_push_neon(int64_t *store);
+extern void aom_pop_neon(int64_t *store);
+#endif  // HAVE_NEON_ASM
+
+void aom_idct16x16_256_add_neon(const int16_t *input, uint8_t *dest,
+                                int dest_stride) {
+#if HAVE_NEON_ASM
+  int64_t store_reg[8];
+#endif
+  int16_t pass1_output[16 * 16] = { 0 };
+  int16_t row_idct_output[16 * 16] = { 0 };
+
+#if HAVE_NEON_ASM
+  // save d8-d15 register values.
+  aom_push_neon(store_reg);
+#endif
+
+  /* Parallel idct on the upper 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(input, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  aom_idct16x16_256_add_neon_pass2(input + 1, row_idct_output, pass1_output, 0,
+                                   dest, dest_stride);
+
+  /* Parallel idct on the lower 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(input + 8 * 16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  aom_idct16x16_256_add_neon_pass2(input + 8 * 16 + 1, row_idct_output + 8,
+                                   pass1_output, 0, dest, dest_stride);
+
+  /* Parallel idct on the left 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(row_idct_output, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  aom_idct16x16_256_add_neon_pass2(row_idct_output + 1, row_idct_output,
+                                   pass1_output, 1, dest, dest_stride);
+
+  /* Parallel idct on the right 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(row_idct_output + 8 * 16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  aom_idct16x16_256_add_neon_pass2(row_idct_output + 8 * 16 + 1,
+                                   row_idct_output + 8, pass1_output, 1,
+                                   dest + 8, dest_stride);
+
+#if HAVE_NEON_ASM
+  // restore d8-d15 register values.
+  aom_pop_neon(store_reg);
+#endif
+
+  return;
+}
+
+void aom_idct16x16_10_add_neon(const int16_t *input, uint8_t *dest,
+                               int dest_stride) {
+#if HAVE_NEON_ASM
+  int64_t store_reg[8];
+#endif
+  int16_t pass1_output[16 * 16] = { 0 };
+  int16_t row_idct_output[16 * 16] = { 0 };
+
+#if HAVE_NEON_ASM
+  // save d8-d15 register values.
+  aom_push_neon(store_reg);
+#endif
+
+  /* Parallel idct on the upper 8 rows */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_10_add_neon_pass1(input, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7
+  // which will be saved into row_idct_output.
+  aom_idct16x16_10_add_neon_pass2(input + 1, row_idct_output, pass1_output, 0,
+                                  dest, dest_stride);
+
+  /* Skip Parallel idct on the lower 8 rows as they are all 0s */
+
+  /* Parallel idct on the left 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(row_idct_output, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  aom_idct16x16_256_add_neon_pass2(row_idct_output + 1, row_idct_output,
+                                   pass1_output, 1, dest, dest_stride);
+
+  /* Parallel idct on the right 8 columns */
+  // First pass processes even elements 0, 2, 4, 6, 8, 10, 12, 14 and save the
+  // stage 6 result in pass1_output.
+  aom_idct16x16_256_add_neon_pass1(row_idct_output + 8 * 16, pass1_output, 8);
+
+  // Second pass processes odd elements 1, 3, 5, 7, 9, 11, 13, 15 and combines
+  // with result in pass1(pass1_output) to calculate final result in stage 7.
+  // Then add the result to the destination data.
+  aom_idct16x16_256_add_neon_pass2(row_idct_output + 8 * 16 + 1,
+                                   row_idct_output + 8, pass1_output, 1,
+                                   dest + 8, dest_stride);
+
+#if HAVE_NEON_ASM
+  // restore d8-d15 register values.
+  aom_pop_neon(store_reg);
+#endif
+
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/idct32x32_1_add_neon.asm b/third_party/aom/aom_dsp/arm/idct32x32_1_add_neon.asm
new file mode 100644
index 0000000000..b04df2d0b8
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct32x32_1_add_neon.asm
@@ -0,0 +1,147 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+
+    EXPORT  |aom_idct32x32_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ;TODO(hkuang): put the following macros in a seperate
+    ;file so other idct function could also use them.
+    MACRO
+    LD_16x8          $src, $stride
+    vld1.8           {q8}, [$src], $stride
+    vld1.8           {q9}, [$src], $stride
+    vld1.8           {q10}, [$src], $stride
+    vld1.8           {q11}, [$src], $stride
+    vld1.8           {q12}, [$src], $stride
+    vld1.8           {q13}, [$src], $stride
+    vld1.8           {q14}, [$src], $stride
+    vld1.8           {q15}, [$src], $stride
+    MEND
+
+    MACRO
+    ADD_DIFF_16x8    $diff
+    vqadd.u8         q8, q8, $diff
+    vqadd.u8         q9, q9, $diff
+    vqadd.u8         q10, q10, $diff
+    vqadd.u8         q11, q11, $diff
+    vqadd.u8         q12, q12, $diff
+    vqadd.u8         q13, q13, $diff
+    vqadd.u8         q14, q14, $diff
+    vqadd.u8         q15, q15, $diff
+    MEND
+
+    MACRO
+    SUB_DIFF_16x8    $diff
+    vqsub.u8         q8, q8, $diff
+    vqsub.u8         q9, q9, $diff
+    vqsub.u8         q10, q10, $diff
+    vqsub.u8         q11, q11, $diff
+    vqsub.u8         q12, q12, $diff
+    vqsub.u8         q13, q13, $diff
+    vqsub.u8         q14, q14, $diff
+    vqsub.u8         q15, q15, $diff
+    MEND
+
+    MACRO
+    ST_16x8          $dst, $stride
+    vst1.8           {q8}, [$dst], $stride
+    vst1.8           {q9}, [$dst], $stride
+    vst1.8           {q10},[$dst], $stride
+    vst1.8           {q11},[$dst], $stride
+    vst1.8           {q12},[$dst], $stride
+    vst1.8           {q13},[$dst], $stride
+    vst1.8           {q14},[$dst], $stride
+    vst1.8           {q15},[$dst], $stride
+    MEND
+
+;void aom_idct32x32_1_add_neon(int16_t *input, uint8_t *dest,
+;                              int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride
+
+|aom_idct32x32_1_add_neon| PROC
+    push             {lr}
+    pld              [r1]
+    add              r3, r1, #16               ; r3 dest + 16 for second loop
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 6)
+    add              r0, r0, #32               ; + (1 <<((6) - 1))
+    asrs             r0, r0, #6                ; >> 6
+    bge              diff_positive_32_32
+
+diff_negative_32_32
+    neg              r0, r0
+    usat             r0, #8, r0
+    vdup.u8          q0, r0
+    mov              r0, #4
+
+diff_negative_32_32_loop
+    sub              r0, #1
+    LD_16x8          r1, r2
+    SUB_DIFF_16x8    q0
+    ST_16x8          r12, r2
+
+    LD_16x8          r1, r2
+    SUB_DIFF_16x8    q0
+    ST_16x8          r12, r2
+    cmp              r0, #2
+    moveq            r1, r3
+    moveq            r12, r3
+    cmp              r0, #0
+    bne              diff_negative_32_32_loop
+    pop              {pc}
+
+diff_positive_32_32
+    usat             r0, #8, r0
+    vdup.u8          q0, r0
+    mov              r0, #4
+
+diff_positive_32_32_loop
+    sub              r0, #1
+    LD_16x8          r1, r2
+    ADD_DIFF_16x8    q0
+    ST_16x8          r12, r2
+
+    LD_16x8          r1, r2
+    ADD_DIFF_16x8    q0
+    ST_16x8          r12, r2
+    cmp              r0, #2
+    moveq            r1, r3
+    moveq            r12, r3
+    cmp              r0, #0
+    bne              diff_positive_32_32_loop
+    pop              {pc}
+
+    ENDP             ; |aom_idct32x32_1_add_neon|
+    END
diff --git a/third_party/aom/aom_dsp/arm/idct32x32_1_add_neon.c b/third_party/aom/aom_dsp/arm/idct32x32_1_add_neon.c
new file mode 100644
index 0000000000..547567c5b4
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct32x32_1_add_neon.c
@@ -0,0 +1,141 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+
+#include "aom_dsp/inv_txfm.h"
+#include "aom_ports/mem.h"
+
+static INLINE void LD_16x8(uint8_t *d, int d_stride, uint8x16_t *q8u8,
+                           uint8x16_t *q9u8, uint8x16_t *q10u8,
+                           uint8x16_t *q11u8, uint8x16_t *q12u8,
+                           uint8x16_t *q13u8, uint8x16_t *q14u8,
+                           uint8x16_t *q15u8) {
+  *q8u8 = vld1q_u8(d);
+  d += d_stride;
+  *q9u8 = vld1q_u8(d);
+  d += d_stride;
+  *q10u8 = vld1q_u8(d);
+  d += d_stride;
+  *q11u8 = vld1q_u8(d);
+  d += d_stride;
+  *q12u8 = vld1q_u8(d);
+  d += d_stride;
+  *q13u8 = vld1q_u8(d);
+  d += d_stride;
+  *q14u8 = vld1q_u8(d);
+  d += d_stride;
+  *q15u8 = vld1q_u8(d);
+  return;
+}
+
+static INLINE void ADD_DIFF_16x8(uint8x16_t qdiffu8, uint8x16_t *q8u8,
+                                 uint8x16_t *q9u8, uint8x16_t *q10u8,
+                                 uint8x16_t *q11u8, uint8x16_t *q12u8,
+                                 uint8x16_t *q13u8, uint8x16_t *q14u8,
+                                 uint8x16_t *q15u8) {
+  *q8u8 = vqaddq_u8(*q8u8, qdiffu8);
+  *q9u8 = vqaddq_u8(*q9u8, qdiffu8);
+  *q10u8 = vqaddq_u8(*q10u8, qdiffu8);
+  *q11u8 = vqaddq_u8(*q11u8, qdiffu8);
+  *q12u8 = vqaddq_u8(*q12u8, qdiffu8);
+  *q13u8 = vqaddq_u8(*q13u8, qdiffu8);
+  *q14u8 = vqaddq_u8(*q14u8, qdiffu8);
+  *q15u8 = vqaddq_u8(*q15u8, qdiffu8);
+  return;
+}
+
+static INLINE void SUB_DIFF_16x8(uint8x16_t qdiffu8, uint8x16_t *q8u8,
+                                 uint8x16_t *q9u8, uint8x16_t *q10u8,
+                                 uint8x16_t *q11u8, uint8x16_t *q12u8,
+                                 uint8x16_t *q13u8, uint8x16_t *q14u8,
+                                 uint8x16_t *q15u8) {
+  *q8u8 = vqsubq_u8(*q8u8, qdiffu8);
+  *q9u8 = vqsubq_u8(*q9u8, qdiffu8);
+  *q10u8 = vqsubq_u8(*q10u8, qdiffu8);
+  *q11u8 = vqsubq_u8(*q11u8, qdiffu8);
+  *q12u8 = vqsubq_u8(*q12u8, qdiffu8);
+  *q13u8 = vqsubq_u8(*q13u8, qdiffu8);
+  *q14u8 = vqsubq_u8(*q14u8, qdiffu8);
+  *q15u8 = vqsubq_u8(*q15u8, qdiffu8);
+  return;
+}
+
+static INLINE void ST_16x8(uint8_t *d, int d_stride, uint8x16_t *q8u8,
+                           uint8x16_t *q9u8, uint8x16_t *q10u8,
+                           uint8x16_t *q11u8, uint8x16_t *q12u8,
+                           uint8x16_t *q13u8, uint8x16_t *q14u8,
+                           uint8x16_t *q15u8) {
+  vst1q_u8(d, *q8u8);
+  d += d_stride;
+  vst1q_u8(d, *q9u8);
+  d += d_stride;
+  vst1q_u8(d, *q10u8);
+  d += d_stride;
+  vst1q_u8(d, *q11u8);
+  d += d_stride;
+  vst1q_u8(d, *q12u8);
+  d += d_stride;
+  vst1q_u8(d, *q13u8);
+  d += d_stride;
+  vst1q_u8(d, *q14u8);
+  d += d_stride;
+  vst1q_u8(d, *q15u8);
+  return;
+}
+
+void aom_idct32x32_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8x16_t q0u8, q8u8, q9u8, q10u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+  int i, j, dest_stride8;
+  uint8_t *d;
+  int16_t a1;
+  int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+
+  out = dct_const_round_shift(out * cospi_16_64);
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+
+  dest_stride8 = dest_stride * 8;
+  if (a1 >= 0) {  // diff_positive_32_32
+    a1 = a1 < 0 ? 0 : a1 > 255 ? 255 : a1;
+    q0u8 = vdupq_n_u8(a1);
+    for (i = 0; i < 2; i++, dest += 16) {  // diff_positive_32_32_loop
+      d = dest;
+      for (j = 0; j < 4; j++) {
+        LD_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                &q14u8, &q15u8);
+        ADD_DIFF_16x8(q0u8, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                      &q14u8, &q15u8);
+        ST_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                &q14u8, &q15u8);
+        d += dest_stride8;
+      }
+    }
+  } else {  // diff_negative_32_32
+    a1 = -a1;
+    a1 = a1 < 0 ? 0 : a1 > 255 ? 255 : a1;
+    q0u8 = vdupq_n_u8(a1);
+    for (i = 0; i < 2; i++, dest += 16) {  // diff_negative_32_32_loop
+      d = dest;
+      for (j = 0; j < 4; j++) {
+        LD_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                &q14u8, &q15u8);
+        SUB_DIFF_16x8(q0u8, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                      &q14u8, &q15u8);
+        ST_16x8(d, dest_stride, &q8u8, &q9u8, &q10u8, &q11u8, &q12u8, &q13u8,
+                &q14u8, &q15u8);
+        d += dest_stride8;
+      }
+    }
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/idct32x32_add_neon.asm b/third_party/aom/aom_dsp/arm/idct32x32_add_neon.asm
new file mode 100644
index 0000000000..e7793fb16e
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct32x32_add_neon.asm
@@ -0,0 +1,1302 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+;TODO(cd): adjust these constant to be able to use vqdmulh for faster
+;          dct_const_round_shift(a * b) within butterfly calculations.
+cospi_1_64  EQU 16364
+cospi_2_64  EQU 16305
+cospi_3_64  EQU 16207
+cospi_4_64  EQU 16069
+cospi_5_64  EQU 15893
+cospi_6_64  EQU 15679
+cospi_7_64  EQU 15426
+cospi_8_64  EQU 15137
+cospi_9_64  EQU 14811
+cospi_10_64 EQU 14449
+cospi_11_64 EQU 14053
+cospi_12_64 EQU 13623
+cospi_13_64 EQU 13160
+cospi_14_64 EQU 12665
+cospi_15_64 EQU 12140
+cospi_16_64 EQU 11585
+cospi_17_64 EQU 11003
+cospi_18_64 EQU 10394
+cospi_19_64 EQU  9760
+cospi_20_64 EQU  9102
+cospi_21_64 EQU  8423
+cospi_22_64 EQU  7723
+cospi_23_64 EQU  7005
+cospi_24_64 EQU  6270
+cospi_25_64 EQU  5520
+cospi_26_64 EQU  4756
+cospi_27_64 EQU  3981
+cospi_28_64 EQU  3196
+cospi_29_64 EQU  2404
+cospi_30_64 EQU  1606
+cospi_31_64 EQU   804
+
+
+    EXPORT  |aom_idct32x32_1024_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    AREA     Block, CODE, READONLY
+
+    ; --------------------------------------------------------------------------
+    ; Load from transposed_buffer
+    ;   q13 = transposed_buffer[first_offset]
+    ;   q14 = transposed_buffer[second_offset]
+    ;   for proper address calculation, the last offset used when manipulating
+    ;   transposed_buffer must be passed in. use 0 for first use.
+    MACRO
+    LOAD_FROM_TRANSPOSED $prev_offset, $first_offset, $second_offset
+    ; address calculation with proper stride and loading
+    add r0, #($first_offset  - $prev_offset )*8*2
+    vld1.s16        {q14}, [r0]
+    add r0, #($second_offset - $first_offset)*8*2
+    vld1.s16        {q13}, [r0]
+    ; (used) two registers (q14, q13)
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Load from output (used as temporary storage)
+    ;   reg1 = output[first_offset]
+    ;   reg2 = output[second_offset]
+    ;   for proper address calculation, the last offset used when manipulating
+    ;   output, whether reading or storing) must be passed in. use 0 for first
+    ;   use.
+    MACRO
+    LOAD_FROM_OUTPUT $prev_offset, $first_offset, $second_offset, $reg1, $reg2
+    ; address calculation with proper stride and loading
+    add r1, #($first_offset  - $prev_offset )*32*2
+    vld1.s16        {$reg1}, [r1]
+    add r1, #($second_offset - $first_offset)*32*2
+    vld1.s16        {$reg2}, [r1]
+    ; (used) two registers ($reg1, $reg2)
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Store into output (sometimes as as temporary storage)
+    ;   output[first_offset] = reg1
+    ;   output[second_offset] = reg2
+    ;   for proper address calculation, the last offset used when manipulating
+    ;   output, whether reading or storing) must be passed in. use 0 for first
+    ;   use.
+    MACRO
+    STORE_IN_OUTPUT $prev_offset, $first_offset, $second_offset, $reg1, $reg2
+    ; address calculation with proper stride and storing
+    add r1, #($first_offset  - $prev_offset )*32*2
+    vst1.16 {$reg1}, [r1]
+    add r1, #($second_offset - $first_offset)*32*2
+    vst1.16 {$reg2}, [r1]
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q6-q9 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_CENTER_RESULTS
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d8}, [r10], r2
+    vld1.s16        {d11}, [r9], r11
+    vld1.s16        {d9}, [r10]
+    vld1.s16        {d10}, [r9]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q8, q8, #6
+    vrshr.s16       q9, q9, #6
+    vrshr.s16       q6, q6, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q7, q7, d9
+    vaddw.u8        q8, q8, d10
+    vaddw.u8        q9, q9, d11
+    vaddw.u8        q6, q6, d8
+    ; clip pixel
+    vqmovun.s16     d9,  q7
+    vqmovun.s16     d10, q8
+    vqmovun.s16     d11, q9
+    vqmovun.s16     d8,  q6
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d9}, [r10], r11
+    vst1.16         {d10}, [r9], r2
+    vst1.16         {d8}, [r10]
+    vst1.16         {d11}, [r9]
+    ; update pointers (by dest_stride * 2)
+    sub r9,  r9,  r2, lsl #1
+    add r10, r10, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q6-q9 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_CENTER_RESULTS_LAST
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d8}, [r10], r2
+    vld1.s16        {d11}, [r9], r11
+    vld1.s16        {d9}, [r10]
+    vld1.s16        {d10}, [r9]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q8, q8, #6
+    vrshr.s16       q9, q9, #6
+    vrshr.s16       q6, q6, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q7, q7, d9
+    vaddw.u8        q8, q8, d10
+    vaddw.u8        q9, q9, d11
+    vaddw.u8        q6, q6, d8
+    ; clip pixel
+    vqmovun.s16     d9,  q7
+    vqmovun.s16     d10, q8
+    vqmovun.s16     d11, q9
+    vqmovun.s16     d8,  q6
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d9}, [r10], r11
+    vst1.16         {d10}, [r9], r2
+    vst1.16         {d8}, [r10]!
+    vst1.16         {d11}, [r9]!
+    ; update pointers (by dest_stride * 2)
+    sub r9,  r9,  r2, lsl #1
+    add r10, r10, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q4-q7 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_EXTREME_RESULTS
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d4}, [r7], r2
+    vld1.s16        {d7}, [r6], r11
+    vld1.s16        {d5}, [r7]
+    vld1.s16        {d6}, [r6]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q5, q5, #6
+    vrshr.s16       q6, q6, #6
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q4, q4, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q5, q5, d5
+    vaddw.u8        q6, q6, d6
+    vaddw.u8        q7, q7, d7
+    vaddw.u8        q4, q4, d4
+    ; clip pixel
+    vqmovun.s16     d5, q5
+    vqmovun.s16     d6, q6
+    vqmovun.s16     d7, q7
+    vqmovun.s16     d4, q4
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d5}, [r7], r11
+    vst1.16         {d6}, [r6], r2
+    vst1.16         {d7}, [r6]
+    vst1.16         {d4}, [r7]
+    ; update pointers (by dest_stride * 2)
+    sub r6, r6, r2, lsl #1
+    add r7, r7, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Combine-add results with current destination content
+    ;   q4-q7 contain the results (out[j * 32 + 0-31])
+    MACRO
+    STORE_COMBINE_EXTREME_RESULTS_LAST
+    ; load dest[j * dest_stride + 0-31]
+    vld1.s16        {d4}, [r7], r2
+    vld1.s16        {d7}, [r6], r11
+    vld1.s16        {d5}, [r7]
+    vld1.s16        {d6}, [r6]
+    ; ROUND_POWER_OF_TWO
+    vrshr.s16       q5, q5, #6
+    vrshr.s16       q6, q6, #6
+    vrshr.s16       q7, q7, #6
+    vrshr.s16       q4, q4, #6
+    ; add to dest[j * dest_stride + 0-31]
+    vaddw.u8        q5, q5, d5
+    vaddw.u8        q6, q6, d6
+    vaddw.u8        q7, q7, d7
+    vaddw.u8        q4, q4, d4
+    ; clip pixel
+    vqmovun.s16     d5, q5
+    vqmovun.s16     d6, q6
+    vqmovun.s16     d7, q7
+    vqmovun.s16     d4, q4
+    ; store back into dest[j * dest_stride + 0-31]
+    vst1.16         {d5}, [r7], r11
+    vst1.16         {d6}, [r6], r2
+    vst1.16         {d7}, [r6]!
+    vst1.16         {d4}, [r7]!
+    ; update pointers (by dest_stride * 2)
+    sub r6, r6, r2, lsl #1
+    add r7, r7, r2, lsl #1
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Touches q8-q12, q15 (q13-q14 are preserved)
+    ; valid output registers are anything but q8-q11
+    MACRO
+    DO_BUTTERFLY $regC, $regD, $regA, $regB, $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4
+    ; TODO(cd): have special case to re-use constants when they are similar for
+    ;           consecutive butterflies
+    ; TODO(cd): have special case when both constants are the same, do the
+    ;           additions/subtractions before the multiplies.
+    ; generate the constants
+    ;   generate scalar constants
+    mov             r8,  #$first_constant  & 0xFF00
+    mov             r12, #$second_constant & 0xFF00
+    add             r8,  #$first_constant  & 0x00FF
+    add             r12, #$second_constant & 0x00FF
+    ;   generate vector constants
+    vdup.16         d30, r8
+    vdup.16         d31, r12
+    ; (used) two for inputs (regA-regD), one for constants (q15)
+    ; do some multiplications (ordered for maximum latency hiding)
+    vmull.s16 q8,  $regC, d30
+    vmull.s16 q10, $regA, d31
+    vmull.s16 q9,  $regD, d30
+    vmull.s16 q11, $regB, d31
+    vmull.s16 q12, $regC, d31
+    ; (used) five for intermediate (q8-q12), one for constants (q15)
+    ; do some addition/subtractions (to get back two register)
+    vsub.s32  q8, q8, q10
+    vsub.s32  q9, q9, q11
+    ; do more multiplications (ordered for maximum latency hiding)
+    vmull.s16 q10, $regD, d31
+    vmull.s16 q11, $regA, d30
+    vmull.s16 q15, $regB, d30
+    ; (used) six for intermediate (q8-q12, q15)
+    ; do more addition/subtractions
+    vadd.s32  q11, q12, q11
+    vadd.s32  q10, q10, q15
+    ; (used) four for intermediate (q8-q11)
+    ; dct_const_round_shift
+    vqrshrn.s32 $reg1, q8,  #14
+    vqrshrn.s32 $reg2, q9,  #14
+    vqrshrn.s32 $reg3, q11, #14
+    vqrshrn.s32 $reg4, q10, #14
+    ; (used) two for results, well four d registers
+    MEND
+    ; --------------------------------------------------------------------------
+    ; Touches q8-q12, q15 (q13-q14 are preserved)
+    ; valid output registers are anything but q8-q11
+    MACRO
+    DO_BUTTERFLY_STD $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4
+    DO_BUTTERFLY d28, d29, d26, d27, $first_constant, $second_constant, $reg1, $reg2, $reg3, $reg4
+    MEND
+    ; --------------------------------------------------------------------------
+
+;void aom_idct32x32_1024_add_neon(int16_t *input, uint8_t *dest, int dest_stride);
+;
+;   r0  int16_t *input,
+;   r1  uint8_t *dest,
+;   r2  int dest_stride)
+; loop counters
+;   r4  bands loop counter
+;   r5  pass loop counter
+;   r8  transpose loop counter
+; combine-add pointers
+;   r6  dest + 31 * dest_stride, descending (30, 29, 28, ...)
+;   r7  dest +  0 * dest_stride, ascending  (1, 2, 3, ...)
+;   r9  dest + 15 * dest_stride, descending (14, 13, 12, ...)
+;   r10 dest + 16 * dest_stride, ascending  (17, 18, 19, ...)
+
+|aom_idct32x32_1024_add_neon| PROC
+    ; This function does one pass of idct32x32 transform.
+    ;
+    ; This is done by transposing the input and then doing a 1d transform on
+    ; columns. In the first pass, the transposed columns are the original
+    ; rows. In the second pass, after the transposition, the colums are the
+    ; original columns.
+    ; The 1d transform is done by looping over bands of eight columns (the
+    ; idct32_bands loop). For each band, the transform input transposition
+    ; is done on demand, one band of four 8x8 matrices at a time. The four
+    ; matrices are transposed by pairs (the idct32_transpose_pair loop).
+    push  {r4-r11}
+    vpush {d8-d15}
+    ; stack operation
+    ; internal buffer used to transpose 8 lines into before transforming them
+    ;   int16_t transpose_buffer[32 * 8];
+    ;   at sp + [4096, 4607]
+    ; results of the first pass (transpose and transform rows)
+    ;   int16_t pass1[32 * 32];
+    ;   at sp + [0, 2047]
+    ; results of the second pass (transpose and transform columns)
+    ;   int16_t pass2[32 * 32];
+    ;   at sp + [2048, 4095]
+    sub sp, sp, #512+2048+2048
+
+    ; r6  = dest + 31 * dest_stride
+    ; r7  = dest +  0 * dest_stride
+    ; r9  = dest + 15 * dest_stride
+    ; r10 = dest + 16 * dest_stride
+    rsb r6,  r2, r2, lsl #5
+    rsb r9,  r2, r2, lsl #4
+    add r10, r1, r2, lsl #4
+    mov r7, r1
+    add r6, r6, r1
+    add r9, r9, r1
+    ; r11 = -dest_stride
+    neg r11, r2
+    ; r3 = input
+    mov r3, r0
+    ; parameters for first pass
+      ; r0 = transpose_buffer[32 * 8]
+    add r0, sp, #4096
+      ; r1 = pass1[32 * 32]
+    mov r1, sp
+
+    mov r5, #0          ; initialize pass loop counter
+idct32_pass_loop
+    mov r4, #4          ; initialize bands loop counter
+idct32_bands_loop
+    mov r8, #2          ; initialize transpose loop counter
+idct32_transpose_pair_loop
+    ; Load two horizontally consecutive 8x8 16bit data matrices. The first one
+    ; into q0-q7 and the second one into q8-q15. There is a stride of 64,
+    ; adjusted to 32 because of the two post-increments.
+    vld1.s16        {q8},  [r3]!
+    vld1.s16        {q0},  [r3]!
+    add r3, #32
+    vld1.s16        {q9},  [r3]!
+    vld1.s16        {q1},  [r3]!
+    add r3, #32
+    vld1.s16        {q10}, [r3]!
+    vld1.s16        {q2},  [r3]!
+    add r3, #32
+    vld1.s16        {q11}, [r3]!
+    vld1.s16        {q3},  [r3]!
+    add r3, #32
+    vld1.s16        {q12}, [r3]!
+    vld1.s16        {q4},  [r3]!
+    add r3, #32
+    vld1.s16        {q13}, [r3]!
+    vld1.s16        {q5},  [r3]!
+    add r3, #32
+    vld1.s16        {q14}, [r3]!
+    vld1.s16        {q6},  [r3]!
+    add r3, #32
+    vld1.s16        {q15}, [r3]!
+    vld1.s16        {q7},  [r3]!
+
+    ; Transpose the two 8x8 16bit data matrices.
+    vswp            d17, d24
+    vswp            d23, d30
+    vswp            d21, d28
+    vswp            d19, d26
+    vswp            d1,  d8
+    vswp            d7,  d14
+    vswp            d5,  d12
+    vswp            d3,  d10
+    vtrn.32         q8,  q10
+    vtrn.32         q9,  q11
+    vtrn.32         q12, q14
+    vtrn.32         q13, q15
+    vtrn.32         q0,  q2
+    vtrn.32         q1,  q3
+    vtrn.32         q4,  q6
+    vtrn.32         q5,  q7
+    vtrn.16         q8,  q9
+    vtrn.16         q10, q11
+    vtrn.16         q12, q13
+    vtrn.16         q14, q15
+    vtrn.16         q0,  q1
+    vtrn.16         q2,  q3
+    vtrn.16         q4,  q5
+    vtrn.16         q6,  q7
+
+    ; Store both matrices after each other. There is a stride of 32, which
+    ; adjusts to nothing because of the post-increments.
+    vst1.16        {q8},  [r0]!
+    vst1.16        {q9},  [r0]!
+    vst1.16        {q10}, [r0]!
+    vst1.16        {q11}, [r0]!
+    vst1.16        {q12}, [r0]!
+    vst1.16        {q13}, [r0]!
+    vst1.16        {q14}, [r0]!
+    vst1.16        {q15}, [r0]!
+    vst1.16        {q0},  [r0]!
+    vst1.16        {q1},  [r0]!
+    vst1.16        {q2},  [r0]!
+    vst1.16        {q3},  [r0]!
+    vst1.16        {q4},  [r0]!
+    vst1.16        {q5},  [r0]!
+    vst1.16        {q6},  [r0]!
+    vst1.16        {q7},  [r0]!
+
+    ; increment pointers by adjusted stride (not necessary for r0/out)
+    ;   go back by 7*32 for the seven lines moved fully by read and add
+    ;   go back by 32 for the eigth line only read
+    ;   advance by 16*2 to go the next pair
+    sub r3,  r3,  #7*32*2 + 32 - 16*2
+    ; transpose pair loop processing
+    subs r8, r8, #1
+    bne idct32_transpose_pair_loop
+
+    ; restore r0/input to its original value
+    sub r0, r0, #32*8*2
+
+    ; Instead of doing the transforms stage by stage, it is done by loading
+    ; some input values and doing as many stages as possible to minimize the
+    ; storing/loading of intermediate results. To fit within registers, the
+    ; final coefficients are cut into four blocks:
+    ; BLOCK A: 16-19,28-31
+    ; BLOCK B: 20-23,24-27
+    ; BLOCK C: 8-10,11-15
+    ; BLOCK D: 0-3,4-7
+    ; Blocks A and C are straight calculation through the various stages. In
+    ; block B, further calculations are performed using the results from
+    ; block A. In block D, further calculations are performed using the results
+    ; from block C and then the final calculations are done using results from
+    ; block A and B which have been combined at the end of block B.
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK A: 16-19,28-31
+    ; --------------------------------------------------------------------------
+    ; generate 16,17,30,31
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[1 * 32] * cospi_31_64 - input[31 * 32] *  cospi_1_64;
+    ;temp2 = input[1 * 32] *  cospi_1_64 + input[31 * 32] * cospi_31_64;
+    ;step1b[16][i] = dct_const_round_shift(temp1);
+    ;step1b[31][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 0, 1, 31
+    DO_BUTTERFLY_STD cospi_31_64, cospi_1_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[17 * 32] * cospi_15_64 - input[15 * 32] * cospi_17_64;
+    ;temp2 = input[17 * 32] * cospi_17_64 + input[15 * 32] * cospi_15_64;
+    ;step1b[17][i] = dct_const_round_shift(temp1);
+    ;step1b[30][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 31, 17, 15
+    DO_BUTTERFLY_STD cospi_15_64, cospi_17_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[16] =  step1b[16][i] + step1b[17][i];
+    ;step2[17] =  step1b[16][i] - step1b[17][i];
+    ;step2[30] = -step1b[30][i] + step1b[31][i];
+    ;step2[31] =  step1b[30][i] + step1b[31][i];
+    vadd.s16  q4, q0, q1
+    vsub.s16  q13, q0, q1
+    vadd.s16  q6, q2, q3
+    vsub.s16  q14, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[30][i] * cospi_28_64 - step1b[17][i] * cospi_4_64;
+    ;temp2 = step1b[30][i] * cospi_4_64  - step1b[17][i] * cospi_28_64;
+    ;step3[17] = dct_const_round_shift(temp1);
+    ;step3[30] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_28_64, cospi_4_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; generate 18,19,28,29
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[9 * 32] * cospi_23_64 - input[23 * 32] * cospi_9_64;
+    ;temp2 = input[9 * 32] *  cospi_9_64 + input[23 * 32] * cospi_23_64;
+    ;step1b[18][i] = dct_const_round_shift(temp1);
+    ;step1b[29][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 15, 9, 23
+    DO_BUTTERFLY_STD cospi_23_64, cospi_9_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[25 * 32] *  cospi_7_64 - input[7 * 32] * cospi_25_64;
+    ;temp2 = input[25 * 32] * cospi_25_64 + input[7 * 32] * cospi_7_64;
+    ;step1b[19][i] = dct_const_round_shift(temp1);
+    ;step1b[28][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 23, 25, 7
+    DO_BUTTERFLY_STD cospi_7_64, cospi_25_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[18] = -step1b[18][i] + step1b[19][i];
+    ;step2[19] =  step1b[18][i] + step1b[19][i];
+    ;step2[28] =  step1b[28][i] + step1b[29][i];
+    ;step2[29] =  step1b[28][i] - step1b[29][i];
+    vsub.s16  q13, q3, q2
+    vadd.s16  q3,  q3, q2
+    vsub.s16  q14, q1, q0
+    vadd.s16  q2,  q1, q0
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[18][i] * (-cospi_4_64)  - step1b[29][i] * (-cospi_28_64);
+    ;temp2 = step1b[18][i] * (-cospi_28_64) + step1b[29][i] * (-cospi_4_64);
+    ;step3[29] = dct_const_round_shift(temp1);
+    ;step3[18] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_4_64), (-cospi_28_64), d2, d3, d0, d1
+    ; --------------------------------------------------------------------------
+    ; combine 16-19,28-31
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[16] = step1b[16][i] + step1b[19][i];
+    ;step1[17] = step1b[17][i] + step1b[18][i];
+    ;step1[18] = step1b[17][i] - step1b[18][i];
+    ;step1[29] = step1b[30][i] - step1b[29][i];
+    ;step1[30] = step1b[30][i] + step1b[29][i];
+    ;step1[31] = step1b[31][i] + step1b[28][i];
+    vadd.s16  q8,  q4, q2
+    vadd.s16  q9,  q5, q0
+    vadd.s16  q10, q7, q1
+    vadd.s16  q15, q6, q3
+    vsub.s16  q13, q5, q0
+    vsub.s16  q14, q7, q1
+    STORE_IN_OUTPUT 0,  16, 31, q8,  q15
+    STORE_IN_OUTPUT 31, 17, 30, q9,  q10
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[29][i] * cospi_24_64 - step1b[18][i] * cospi_8_64;
+    ;temp2 = step1b[29][i] * cospi_8_64  + step1b[18][i] * cospi_24_64;
+    ;step2[18] = dct_const_round_shift(temp1);
+    ;step2[29] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d0, d1, d2, d3
+    STORE_IN_OUTPUT 30, 29, 18, q1, q0
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[19] = step1b[16][i] - step1b[19][i];
+    ;step1[28] = step1b[31][i] - step1b[28][i];
+    vsub.s16  q13, q4, q2
+    vsub.s16  q14, q6, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[28][i] * cospi_24_64 - step1b[19][i] * cospi_8_64;
+    ;temp2 = step1b[28][i] * cospi_8_64  + step1b[19][i] * cospi_24_64;
+    ;step2[19] = dct_const_round_shift(temp1);
+    ;step2[28] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d8, d9, d12, d13
+    STORE_IN_OUTPUT 18, 19, 28, q4, q6
+    ; --------------------------------------------------------------------------
+
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK B: 20-23,24-27
+    ; --------------------------------------------------------------------------
+    ; generate 20,21,26,27
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[5 * 32] * cospi_27_64 - input[27 * 32] * cospi_5_64;
+    ;temp2 = input[5 * 32] *  cospi_5_64 + input[27 * 32] * cospi_27_64;
+    ;step1b[20][i] = dct_const_round_shift(temp1);
+    ;step1b[27][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 7, 5, 27
+    DO_BUTTERFLY_STD cospi_27_64, cospi_5_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[21 * 32] * cospi_11_64 - input[11 * 32] * cospi_21_64;
+    ;temp2 = input[21 * 32] * cospi_21_64 + input[11 * 32] * cospi_11_64;
+    ;step1b[21][i] = dct_const_round_shift(temp1);
+    ;step1b[26][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 27, 21, 11
+    DO_BUTTERFLY_STD cospi_11_64, cospi_21_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[20] =  step1b[20][i] + step1b[21][i];
+    ;step2[21] =  step1b[20][i] - step1b[21][i];
+    ;step2[26] = -step1b[26][i] + step1b[27][i];
+    ;step2[27] =  step1b[26][i] + step1b[27][i];
+    vsub.s16  q13, q0, q1
+    vadd.s16  q0, q0, q1
+    vsub.s16  q14, q2, q3
+    vadd.s16  q2, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[26][i] * cospi_12_64 - step1b[21][i] * cospi_20_64;
+    ;temp2 = step1b[26][i] * cospi_20_64 + step1b[21][i] * cospi_12_64;
+    ;step3[21] = dct_const_round_shift(temp1);
+    ;step3[26] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_12_64, cospi_20_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; generate 22,23,24,25
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[13 * 32] * cospi_19_64 - input[19 * 32] * cospi_13_64;
+    ;temp2 = input[13 * 32] * cospi_13_64 + input[19 * 32] * cospi_19_64;
+    ;step1b[22][i] = dct_const_round_shift(temp1);
+    ;step1b[25][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 11, 13, 19
+    DO_BUTTERFLY_STD cospi_19_64, cospi_13_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; part of stage 1
+    ;temp1 = input[29 * 32] *  cospi_3_64 - input[3 * 32] * cospi_29_64;
+    ;temp2 = input[29 * 32] * cospi_29_64 + input[3 * 32] * cospi_3_64;
+    ;step1b[23][i] = dct_const_round_shift(temp1);
+    ;step1b[24][i] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 19, 29, 3
+    DO_BUTTERFLY_STD cospi_3_64, cospi_29_64, d8, d9, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;step2[22] = -step1b[22][i] + step1b[23][i];
+    ;step2[23] =  step1b[22][i] + step1b[23][i];
+    ;step2[24] =  step1b[24][i] + step1b[25][i];
+    ;step2[25] =  step1b[24][i] - step1b[25][i];
+    vsub.s16  q14, q4, q5
+    vadd.s16  q5, q4, q5
+    vsub.s16  q13, q6, q7
+    vadd.s16  q6, q6, q7
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = step1b[22][i] * (-cospi_20_64) - step1b[25][i] * (-cospi_12_64);
+    ;temp2 = step1b[22][i] * (-cospi_12_64) + step1b[25][i] * (-cospi_20_64);
+    ;step3[25] = dct_const_round_shift(temp1);
+    ;step3[22] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_20_64), (-cospi_12_64), d8, d9, d14, d15
+    ; --------------------------------------------------------------------------
+    ; combine 20-23,24-27
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[22] = step1b[22][i] + step1b[21][i];
+    ;step1[23] = step1b[23][i] + step1b[20][i];
+    vadd.s16  q10, q7, q1
+    vadd.s16  q11, q5, q0
+    ;step1[24] = step1b[24][i] + step1b[27][i];
+    ;step1[25] = step1b[25][i] + step1b[26][i];
+    vadd.s16  q12, q6, q2
+    vadd.s16  q15, q4, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[16] = step1b[16][i] + step1b[23][i];
+    ;step3[17] = step1b[17][i] + step1b[22][i];
+    ;step3[22] = step1b[17][i] - step1b[22][i];
+    ;step3[23] = step1b[16][i] - step1b[23][i];
+    LOAD_FROM_OUTPUT 28, 16, 17, q14, q13
+    vadd.s16  q8,  q14, q11
+    vadd.s16  q9,  q13, q10
+    vsub.s16  q13, q13, q10
+    vsub.s16  q11, q14, q11
+    STORE_IN_OUTPUT 17, 17, 16, q9, q8
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[24] = step1b[31][i] - step1b[24][i];
+    ;step3[25] = step1b[30][i] - step1b[25][i];
+    ;step3[30] = step1b[30][i] + step1b[25][i];
+    ;step3[31] = step1b[31][i] + step1b[24][i];
+    LOAD_FROM_OUTPUT 16, 30, 31, q14, q9
+    vsub.s16  q8,  q9,  q12
+    vadd.s16  q10, q14, q15
+    vsub.s16  q14, q14, q15
+    vadd.s16  q12, q9,  q12
+    STORE_IN_OUTPUT 31, 30, 31, q10, q12
+    ; --------------------------------------------------------------------------
+    ; TODO(cd) do some register allocation change to remove these push/pop
+    vpush {q8}  ; [24]
+    vpush {q11} ; [23]
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[25][i] - step1b[22][i]) * cospi_16_64;
+    ;temp2 = (step1b[25][i] + step1b[22][i]) * cospi_16_64;
+    ;step1[22] = dct_const_round_shift(temp1);
+    ;step1[25] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29
+    STORE_IN_OUTPUT 31, 25, 22, q14, q13
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[24][i] - step1b[23][i]) * cospi_16_64;
+    ;temp2 = (step1b[24][i] + step1b[23][i]) * cospi_16_64;
+    ;step1[23] = dct_const_round_shift(temp1);
+    ;step1[24] = dct_const_round_shift(temp2);
+    ; TODO(cd) do some register allocation change to remove these push/pop
+    vpop  {q13} ; [23]
+    vpop  {q14} ; [24]
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29
+    STORE_IN_OUTPUT 22, 24, 23, q14, q13
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[20] = step1b[23][i] - step1b[20][i];
+    ;step1[27] = step1b[24][i] - step1b[27][i];
+    vsub.s16  q14, q5, q0
+    vsub.s16  q13, q6, q2
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[20][i] * (-cospi_8_64)  - step1b[27][i] * (-cospi_24_64);
+    ;temp2 = step1b[20][i] * (-cospi_24_64) + step1b[27][i] * (-cospi_8_64);
+    ;step2[27] = dct_const_round_shift(temp1);
+    ;step2[20] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d10, d11, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[21] = step1b[22][i] - step1b[21][i];
+    ;step1[26] = step1b[25][i] - step1b[26][i];
+    vsub.s16  q14,  q7, q1
+    vsub.s16  q13,  q4, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = step1b[21][i] * (-cospi_8_64)  - step1b[26][i] * (-cospi_24_64);
+    ;temp2 = step1b[21][i] * (-cospi_24_64) + step1b[26][i] * (-cospi_8_64);
+    ;step2[26] = dct_const_round_shift(temp1);
+    ;step2[21] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d0, d1, d2, d3
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[18] = step1b[18][i] + step1b[21][i];
+    ;step3[19] = step1b[19][i] + step1b[20][i];
+    ;step3[20] = step1b[19][i] - step1b[20][i];
+    ;step3[21] = step1b[18][i] - step1b[21][i];
+    LOAD_FROM_OUTPUT 23, 18, 19, q14, q13
+    vadd.s16  q8,  q14, q1
+    vadd.s16  q9,  q13, q6
+    vsub.s16  q13, q13, q6
+    vsub.s16  q1,  q14, q1
+    STORE_IN_OUTPUT 19, 18, 19, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[27] = step1b[28][i] - step1b[27][i];
+    ;step3[28] = step1b[28][i] + step1b[27][i];
+    ;step3[29] = step1b[29][i] + step1b[26][i];
+    ;step3[26] = step1b[29][i] - step1b[26][i];
+    LOAD_FROM_OUTPUT 19, 28, 29, q8, q9
+    vsub.s16  q14, q8, q5
+    vadd.s16  q10, q8, q5
+    vadd.s16  q11, q9, q0
+    vsub.s16  q0, q9, q0
+    STORE_IN_OUTPUT 29, 28, 29, q10, q11
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[27][i] - step1b[20][i]) * cospi_16_64;
+    ;temp2 = (step1b[27][i] + step1b[20][i]) * cospi_16_64;
+    ;step1[20] = dct_const_round_shift(temp1);
+    ;step1[27] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d26, d27, d28, d29
+    STORE_IN_OUTPUT 29, 20, 27, q13, q14
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;temp1 = (step1b[26][i] - step1b[21][i]) * cospi_16_64;
+    ;temp2 = (step1b[26][i] + step1b[21][i]) * cospi_16_64;
+    ;step1[21] = dct_const_round_shift(temp1);
+    ;step1[26] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY d0, d1, d2, d3, cospi_16_64, cospi_16_64, d2, d3, d0, d1
+    STORE_IN_OUTPUT 27, 21, 26, q1, q0
+    ; --------------------------------------------------------------------------
+
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK C: 8-10,11-15
+    ; --------------------------------------------------------------------------
+    ; generate 8,9,14,15
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[2 * 32] * cospi_30_64 - input[30 * 32] * cospi_2_64;
+    ;temp2 = input[2 * 32] * cospi_2_64 + input[30 * 32] * cospi_30_64;
+    ;step2[8] = dct_const_round_shift(temp1);
+    ;step2[15] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 3, 2, 30
+    DO_BUTTERFLY_STD cospi_30_64, cospi_2_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[18 * 32] * cospi_14_64 - input[14 * 32] * cospi_18_64;
+    ;temp2 = input[18 * 32] * cospi_18_64 + input[14 * 32] * cospi_14_64;
+    ;step2[9] = dct_const_round_shift(temp1);
+    ;step2[14] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 30, 18, 14
+    DO_BUTTERFLY_STD cospi_14_64, cospi_18_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;step3[8] = step1b[8][i] + step1b[9][i];
+    ;step3[9] = step1b[8][i] - step1b[9][i];
+    ;step3[14] = step1b[15][i] - step1b[14][i];
+    ;step3[15] = step1b[15][i] + step1b[14][i];
+    vsub.s16  q13, q0, q1
+    vadd.s16  q0, q0, q1
+    vsub.s16  q14, q2, q3
+    vadd.s16  q2, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = step1b[14][i] * cospi_24_64 - step1b[9][i] * cospi_8_64;
+    ;temp2 = step1b[14][i] * cospi_8_64  + step1b[9][i] * cospi_24_64;
+    ;step1[9]  = dct_const_round_shift(temp1);
+    ;step1[14] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; generate 10,11,12,13
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[10 * 32] * cospi_22_64 - input[22 * 32] * cospi_10_64;
+    ;temp2 = input[10 * 32] * cospi_10_64 + input[22 * 32] * cospi_22_64;
+    ;step2[10] = dct_const_round_shift(temp1);
+    ;step2[13] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 14, 10, 22
+    DO_BUTTERFLY_STD cospi_22_64, cospi_10_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; part of stage 2
+    ;temp1 = input[26 * 32] * cospi_6_64 - input[6 * 32] * cospi_26_64;
+    ;temp2 = input[26 * 32] * cospi_26_64 + input[6 * 32] * cospi_6_64;
+    ;step2[11] = dct_const_round_shift(temp1);
+    ;step2[12] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 22, 26, 6
+    DO_BUTTERFLY_STD cospi_6_64, cospi_26_64, d8, d9, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;step3[10] = step1b[11][i] - step1b[10][i];
+    ;step3[11] = step1b[11][i] + step1b[10][i];
+    ;step3[12] = step1b[12][i] + step1b[13][i];
+    ;step3[13] = step1b[12][i] - step1b[13][i];
+    vsub.s16  q14, q4, q5
+    vadd.s16  q5, q4, q5
+    vsub.s16  q13, q6, q7
+    vadd.s16  q6, q6, q7
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = step1b[10][i] * (-cospi_8_64)  - step1b[13][i] * (-cospi_24_64);
+    ;temp2 = step1b[10][i] * (-cospi_24_64) + step1b[13][i] * (-cospi_8_64);
+    ;step1[13] = dct_const_round_shift(temp1);
+    ;step1[10] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD (-cospi_8_64), (-cospi_24_64), d8, d9, d14, d15
+    ; --------------------------------------------------------------------------
+    ; combine 8-10,11-15
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;step2[8]  = step1b[8][i] + step1b[11][i];
+    ;step2[9]  = step1b[9][i] + step1b[10][i];
+    ;step2[10] = step1b[9][i] - step1b[10][i];
+    vadd.s16  q8,  q0, q5
+    vadd.s16  q9,  q1, q7
+    vsub.s16  q13, q1, q7
+    ;step2[13] = step1b[14][i] - step1b[13][i];
+    ;step2[14] = step1b[14][i] + step1b[13][i];
+    ;step2[15] = step1b[15][i] + step1b[12][i];
+    vsub.s16  q14, q3, q4
+    vadd.s16  q10, q3, q4
+    vadd.s16  q15, q2, q6
+    STORE_IN_OUTPUT 26, 8, 15, q8, q15
+    STORE_IN_OUTPUT 15, 9, 14, q9, q10
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;temp1 = (step1b[13][i] - step1b[10][i]) * cospi_16_64;
+    ;temp2 = (step1b[13][i] + step1b[10][i]) * cospi_16_64;
+    ;step3[10] = dct_const_round_shift(temp1);
+    ;step3[13] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7
+    STORE_IN_OUTPUT 14, 13, 10, q3, q1
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;step2[11] = step1b[8][i] - step1b[11][i];
+    ;step2[12] = step1b[15][i] - step1b[12][i];
+    vsub.s16  q13, q0, q5
+    vsub.s16  q14,  q2, q6
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;temp1 = (step1b[12][i] - step1b[11][i]) * cospi_16_64;
+    ;temp2 = (step1b[12][i] + step1b[11][i]) * cospi_16_64;
+    ;step3[11] = dct_const_round_shift(temp1);
+    ;step3[12] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7
+    STORE_IN_OUTPUT 10, 11, 12, q1, q3
+    ; --------------------------------------------------------------------------
+
+
+    ; --------------------------------------------------------------------------
+    ; BLOCK D: 0-3,4-7
+    ; --------------------------------------------------------------------------
+    ; generate 4,5,6,7
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = input[4 * 32] * cospi_28_64 - input[28 * 32] * cospi_4_64;
+    ;temp2 = input[4 * 32] * cospi_4_64 + input[28 * 32] * cospi_28_64;
+    ;step3[4] = dct_const_round_shift(temp1);
+    ;step3[7] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 6, 4, 28
+    DO_BUTTERFLY_STD cospi_28_64, cospi_4_64, d0, d1, d4, d5
+    ; --------------------------------------------------------------------------
+    ; part of stage 3
+    ;temp1 = input[20 * 32] * cospi_12_64 - input[12 * 32] * cospi_20_64;
+    ;temp2 = input[20 * 32] * cospi_20_64 + input[12 * 32] * cospi_12_64;
+    ;step3[5] = dct_const_round_shift(temp1);
+    ;step3[6] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 28, 20, 12
+    DO_BUTTERFLY_STD cospi_12_64, cospi_20_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;step1[4] = step1b[4][i] + step1b[5][i];
+    ;step1[5] = step1b[4][i] - step1b[5][i];
+    ;step1[6] = step1b[7][i] - step1b[6][i];
+    ;step1[7] = step1b[7][i] + step1b[6][i];
+    vsub.s16  q13, q0, q1
+    vadd.s16  q0, q0, q1
+    vsub.s16  q14, q2, q3
+    vadd.s16  q2, q2, q3
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;temp1 = (step1b[6][i] - step1b[5][i]) * cospi_16_64;
+    ;temp2 = (step1b[5][i] + step1b[6][i]) * cospi_16_64;
+    ;step2[5] = dct_const_round_shift(temp1);
+    ;step2[6] = dct_const_round_shift(temp2);
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d2, d3, d6, d7
+    ; --------------------------------------------------------------------------
+    ; generate 0,1,2,3
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = (input[0 * 32] - input[16 * 32]) * cospi_16_64;
+    ;temp2 = (input[0 * 32] + input[16 * 32]) * cospi_16_64;
+    ;step1[1] = dct_const_round_shift(temp1);
+    ;step1[0] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 12, 0, 16
+    DO_BUTTERFLY_STD cospi_16_64, cospi_16_64, d10, d11, d14, d15
+    ; --------------------------------------------------------------------------
+    ; part of stage 4
+    ;temp1 = input[8 * 32] * cospi_24_64 - input[24 * 32] * cospi_8_64;
+    ;temp2 = input[8 * 32] * cospi_8_64 + input[24 * 32] * cospi_24_64;
+    ;step1[2] = dct_const_round_shift(temp1);
+    ;step1[3] = dct_const_round_shift(temp2);
+    LOAD_FROM_TRANSPOSED 16, 8, 24
+    DO_BUTTERFLY_STD cospi_24_64, cospi_8_64, d28, d29, d12, d13
+    ; --------------------------------------------------------------------------
+    ; part of stage 5
+    ;step2[0] = step1b[0][i] + step1b[3][i];
+    ;step2[1] = step1b[1][i] + step1b[2][i];
+    ;step2[2] = step1b[1][i] - step1b[2][i];
+    ;step2[3] = step1b[0][i] - step1b[3][i];
+    vadd.s16  q4, q7, q6
+    vsub.s16  q7, q7, q6
+    vsub.s16  q6, q5, q14
+    vadd.s16  q5, q5, q14
+    ; --------------------------------------------------------------------------
+    ; combine 0-3,4-7
+    ; --------------------------------------------------------------------------
+    ; part of stage 6
+    ;step3[0] = step1b[0][i] + step1b[7][i];
+    ;step3[1] = step1b[1][i] + step1b[6][i];
+    ;step3[2] = step1b[2][i] + step1b[5][i];
+    ;step3[3] = step1b[3][i] + step1b[4][i];
+    vadd.s16  q8,  q4, q2
+    vadd.s16  q9,  q5, q3
+    vadd.s16  q10, q6, q1
+    vadd.s16  q11, q7, q0
+    ;step3[4] = step1b[3][i] - step1b[4][i];
+    ;step3[5] = step1b[2][i] - step1b[5][i];
+    ;step3[6] = step1b[1][i] - step1b[6][i];
+    ;step3[7] = step1b[0][i] - step1b[7][i];
+    vsub.s16  q12, q7, q0
+    vsub.s16  q13, q6, q1
+    vsub.s16  q14, q5, q3
+    vsub.s16  q15, q4, q2
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[0] = step1b[0][i] + step1b[15][i];
+    ;step1[1] = step1b[1][i] + step1b[14][i];
+    ;step1[14] = step1b[1][i] - step1b[14][i];
+    ;step1[15] = step1b[0][i] - step1b[15][i];
+    LOAD_FROM_OUTPUT 12, 14, 15, q0, q1
+    vadd.s16  q2, q8, q1
+    vadd.s16  q3, q9, q0
+    vsub.s16  q4, q9, q0
+    vsub.s16  q5, q8, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[14 * 32] = step1b[14][i] + step1b[17][i];
+    ;output[15 * 32] = step1b[15][i] + step1b[16][i];
+    ;output[16 * 32] = step1b[15][i] - step1b[16][i];
+    ;output[17 * 32] = step1b[14][i] - step1b[17][i];
+    LOAD_FROM_OUTPUT 15, 16, 17, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+
+    cmp r5, #0
+    bgt idct32_bands_end_2nd_pass
+
+idct32_bands_end_1st_pass
+    STORE_IN_OUTPUT 17, 16, 17, q6, q7
+    STORE_IN_OUTPUT 17, 14, 15, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 0 * 32] = step1b[0][i] + step1b[31][i];
+    ;output[ 1 * 32] = step1b[1][i] + step1b[30][i];
+    ;output[30 * 32] = step1b[1][i] - step1b[30][i];
+    ;output[31 * 32] = step1b[0][i] - step1b[31][i];
+    LOAD_FROM_OUTPUT 15, 30, 31, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 31, 30, 31, q6, q7
+    STORE_IN_OUTPUT 31,  0,  1, q4, q5
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[2] = step1b[2][i] + step1b[13][i];
+    ;step1[3] = step1b[3][i] + step1b[12][i];
+    ;step1[12] = step1b[3][i] - step1b[12][i];
+    ;step1[13] = step1b[2][i] - step1b[13][i];
+    LOAD_FROM_OUTPUT 1, 12, 13, q0, q1
+    vadd.s16  q2, q10, q1
+    vadd.s16  q3, q11, q0
+    vsub.s16  q4, q11, q0
+    vsub.s16  q5, q10, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[12 * 32] = step1b[12][i] + step1b[19][i];
+    ;output[13 * 32] = step1b[13][i] + step1b[18][i];
+    ;output[18 * 32] = step1b[13][i] - step1b[18][i];
+    ;output[19 * 32] = step1b[12][i] - step1b[19][i];
+    LOAD_FROM_OUTPUT 13, 18, 19, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_IN_OUTPUT 19, 18, 19, q6, q7
+    STORE_IN_OUTPUT 19, 12, 13, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 2 * 32] = step1b[2][i] + step1b[29][i];
+    ;output[ 3 * 32] = step1b[3][i] + step1b[28][i];
+    ;output[28 * 32] = step1b[3][i] - step1b[28][i];
+    ;output[29 * 32] = step1b[2][i] - step1b[29][i];
+    LOAD_FROM_OUTPUT 13, 28, 29, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 29, 28, 29, q6, q7
+    STORE_IN_OUTPUT 29,  2,  3, q4, q5
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[4] = step1b[4][i] + step1b[11][i];
+    ;step1[5] = step1b[5][i] + step1b[10][i];
+    ;step1[10] = step1b[5][i] - step1b[10][i];
+    ;step1[11] = step1b[4][i] - step1b[11][i];
+    LOAD_FROM_OUTPUT 3, 10, 11, q0, q1
+    vadd.s16  q2, q12, q1
+    vadd.s16  q3, q13, q0
+    vsub.s16  q4, q13, q0
+    vsub.s16  q5, q12, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[10 * 32] = step1b[10][i] + step1b[21][i];
+    ;output[11 * 32] = step1b[11][i] + step1b[20][i];
+    ;output[20 * 32] = step1b[11][i] - step1b[20][i];
+    ;output[21 * 32] = step1b[10][i] - step1b[21][i];
+    LOAD_FROM_OUTPUT 11, 20, 21, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_IN_OUTPUT 21, 20, 21, q6, q7
+    STORE_IN_OUTPUT 21, 10, 11, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 4 * 32] = step1b[4][i] + step1b[27][i];
+    ;output[ 5 * 32] = step1b[5][i] + step1b[26][i];
+    ;output[26 * 32] = step1b[5][i] - step1b[26][i];
+    ;output[27 * 32] = step1b[4][i] - step1b[27][i];
+    LOAD_FROM_OUTPUT 11, 26, 27, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 27, 26, 27, q6, q7
+    STORE_IN_OUTPUT 27,  4,  5, q4, q5
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[6] = step1b[6][i] + step1b[9][i];
+    ;step1[7] = step1b[7][i] + step1b[8][i];
+    ;step1[8] = step1b[7][i] - step1b[8][i];
+    ;step1[9] = step1b[6][i] - step1b[9][i];
+    LOAD_FROM_OUTPUT 5, 8, 9, q0, q1
+    vadd.s16  q2, q14, q1
+    vadd.s16  q3, q15, q0
+    vsub.s16  q4, q15, q0
+    vsub.s16  q5, q14, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 8 * 32] = step1b[8][i] + step1b[23][i];
+    ;output[ 9 * 32] = step1b[9][i] + step1b[22][i];
+    ;output[22 * 32] = step1b[9][i] - step1b[22][i];
+    ;output[23 * 32] = step1b[8][i] - step1b[23][i];
+    LOAD_FROM_OUTPUT 9, 22, 23, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_IN_OUTPUT 23, 22, 23, q6, q7
+    STORE_IN_OUTPUT 23, 8, 9, q8, q9
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 6 * 32] = step1b[6][i] + step1b[25][i];
+    ;output[ 7 * 32] = step1b[7][i] + step1b[24][i];
+    ;output[24 * 32] = step1b[7][i] - step1b[24][i];
+    ;output[25 * 32] = step1b[6][i] - step1b[25][i];
+    LOAD_FROM_OUTPUT 9, 24, 25, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_IN_OUTPUT 25, 24, 25, q6, q7
+    STORE_IN_OUTPUT 25,  6,  7, q4, q5
+
+    ; restore r0 by removing the last offset from the last
+    ;     operation (LOAD_FROM_TRANSPOSED 16, 8, 24) => 24*8*2
+    sub r0, r0, #24*8*2
+    ; restore r1 by removing the last offset from the last
+    ;     operation (STORE_IN_OUTPUT 24,  6,  7) => 7*32*2
+    ; advance by 8 columns => 8*2
+    sub r1, r1, #7*32*2 - 8*2
+    ;   advance by 8 lines (8*32*2)
+    ;   go back by the two pairs from the loop (32*2)
+    add r3, r3, #8*32*2 - 32*2
+
+    ; bands loop processing
+    subs r4, r4, #1
+    bne idct32_bands_loop
+
+    ; parameters for second pass
+    ; the input of pass2 is the result of pass1. we have to remove the offset
+    ;   of 32 columns induced by the above idct32_bands_loop
+    sub r3, r1, #32*2
+      ; r1 = pass2[32 * 32]
+    add r1, sp, #2048
+
+    ; pass loop processing
+    add r5, r5, #1
+    b idct32_pass_loop
+
+idct32_bands_end_2nd_pass
+    STORE_COMBINE_CENTER_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 0 * 32] = step1b[0][i] + step1b[31][i];
+    ;output[ 1 * 32] = step1b[1][i] + step1b[30][i];
+    ;output[30 * 32] = step1b[1][i] - step1b[30][i];
+    ;output[31 * 32] = step1b[0][i] - step1b[31][i];
+    LOAD_FROM_OUTPUT 17, 30, 31, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[2] = step1b[2][i] + step1b[13][i];
+    ;step1[3] = step1b[3][i] + step1b[12][i];
+    ;step1[12] = step1b[3][i] - step1b[12][i];
+    ;step1[13] = step1b[2][i] - step1b[13][i];
+    LOAD_FROM_OUTPUT 31, 12, 13, q0, q1
+    vadd.s16  q2, q10, q1
+    vadd.s16  q3, q11, q0
+    vsub.s16  q4, q11, q0
+    vsub.s16  q5, q10, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[12 * 32] = step1b[12][i] + step1b[19][i];
+    ;output[13 * 32] = step1b[13][i] + step1b[18][i];
+    ;output[18 * 32] = step1b[13][i] - step1b[18][i];
+    ;output[19 * 32] = step1b[12][i] - step1b[19][i];
+    LOAD_FROM_OUTPUT 13, 18, 19, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_COMBINE_CENTER_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 2 * 32] = step1b[2][i] + step1b[29][i];
+    ;output[ 3 * 32] = step1b[3][i] + step1b[28][i];
+    ;output[28 * 32] = step1b[3][i] - step1b[28][i];
+    ;output[29 * 32] = step1b[2][i] - step1b[29][i];
+    LOAD_FROM_OUTPUT 19, 28, 29, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[4] = step1b[4][i] + step1b[11][i];
+    ;step1[5] = step1b[5][i] + step1b[10][i];
+    ;step1[10] = step1b[5][i] - step1b[10][i];
+    ;step1[11] = step1b[4][i] - step1b[11][i];
+    LOAD_FROM_OUTPUT 29, 10, 11, q0, q1
+    vadd.s16  q2, q12, q1
+    vadd.s16  q3, q13, q0
+    vsub.s16  q4, q13, q0
+    vsub.s16  q5, q12, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[10 * 32] = step1b[10][i] + step1b[21][i];
+    ;output[11 * 32] = step1b[11][i] + step1b[20][i];
+    ;output[20 * 32] = step1b[11][i] - step1b[20][i];
+    ;output[21 * 32] = step1b[10][i] - step1b[21][i];
+    LOAD_FROM_OUTPUT 11, 20, 21, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_COMBINE_CENTER_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 4 * 32] = step1b[4][i] + step1b[27][i];
+    ;output[ 5 * 32] = step1b[5][i] + step1b[26][i];
+    ;output[26 * 32] = step1b[5][i] - step1b[26][i];
+    ;output[27 * 32] = step1b[4][i] - step1b[27][i];
+    LOAD_FROM_OUTPUT 21, 26, 27, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS
+    ; --------------------------------------------------------------------------
+    ; part of stage 7
+    ;step1[6] = step1b[6][i] + step1b[9][i];
+    ;step1[7] = step1b[7][i] + step1b[8][i];
+    ;step1[8] = step1b[7][i] - step1b[8][i];
+    ;step1[9] = step1b[6][i] - step1b[9][i];
+    LOAD_FROM_OUTPUT 27, 8, 9, q0, q1
+    vadd.s16  q2, q14, q1
+    vadd.s16  q3, q15, q0
+    vsub.s16  q4, q15, q0
+    vsub.s16  q5, q14, q1
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 8 * 32] = step1b[8][i] + step1b[23][i];
+    ;output[ 9 * 32] = step1b[9][i] + step1b[22][i];
+    ;output[22 * 32] = step1b[9][i] - step1b[22][i];
+    ;output[23 * 32] = step1b[8][i] - step1b[23][i];
+    LOAD_FROM_OUTPUT 9, 22, 23, q0, q1
+    vadd.s16  q8, q4, q1
+    vadd.s16  q9, q5, q0
+    vsub.s16  q6, q5, q0
+    vsub.s16  q7, q4, q1
+    STORE_COMBINE_CENTER_RESULTS_LAST
+    ; --------------------------------------------------------------------------
+    ; part of final stage
+    ;output[ 6 * 32] = step1b[6][i] + step1b[25][i];
+    ;output[ 7 * 32] = step1b[7][i] + step1b[24][i];
+    ;output[24 * 32] = step1b[7][i] - step1b[24][i];
+    ;output[25 * 32] = step1b[6][i] - step1b[25][i];
+    LOAD_FROM_OUTPUT 23, 24, 25, q0, q1
+    vadd.s16  q4, q2, q1
+    vadd.s16  q5, q3, q0
+    vsub.s16  q6, q3, q0
+    vsub.s16  q7, q2, q1
+    STORE_COMBINE_EXTREME_RESULTS_LAST
+    ; --------------------------------------------------------------------------
+    ; restore pointers to their initial indices for next band pass by
+    ;     removing/adding dest_stride * 8. The actual increment by eight
+    ;     is taken care of within the _LAST macros.
+    add r6,  r6,  r2, lsl #3
+    add r9,  r9,  r2, lsl #3
+    sub r7,  r7,  r2, lsl #3
+    sub r10, r10, r2, lsl #3
+
+    ; restore r0 by removing the last offset from the last
+    ;     operation (LOAD_FROM_TRANSPOSED 16, 8, 24) => 24*8*2
+    sub r0, r0, #24*8*2
+    ; restore r1 by removing the last offset from the last
+    ;     operation (LOAD_FROM_OUTPUT 23, 24, 25) => 25*32*2
+    ; advance by 8 columns => 8*2
+    sub r1, r1, #25*32*2 - 8*2
+    ;   advance by 8 lines (8*32*2)
+    ;   go back by the two pairs from the loop (32*2)
+    add r3, r3, #8*32*2 - 32*2
+
+    ; bands loop processing
+    subs r4, r4, #1
+    bne idct32_bands_loop
+
+    ; stack operation
+    add sp, sp, #512+2048+2048
+    vpop {d8-d15}
+    pop  {r4-r11}
+    bx              lr
+    ENDP  ; |aom_idct32x32_1024_add_neon|
+    END
diff --git a/third_party/aom/aom_dsp/arm/idct32x32_add_neon.c b/third_party/aom/aom_dsp/arm/idct32x32_add_neon.c
new file mode 100644
index 0000000000..a7562c7d5d
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct32x32_add_neon.c
@@ -0,0 +1,686 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "aom_dsp/txfm_common.h"
+
+#define LOAD_FROM_TRANSPOSED(prev, first, second) \
+  q14s16 = vld1q_s16(trans_buf + first * 8);      \
+  q13s16 = vld1q_s16(trans_buf + second * 8);
+
+#define LOAD_FROM_OUTPUT(prev, first, second, qA, qB) \
+  qA = vld1q_s16(out + first * 32);                   \
+  qB = vld1q_s16(out + second * 32);
+
+#define STORE_IN_OUTPUT(prev, first, second, qA, qB) \
+  vst1q_s16(out + first * 32, qA);                   \
+  vst1q_s16(out + second * 32, qB);
+
+#define STORE_COMBINE_CENTER_RESULTS(r10, r9) \
+  __STORE_COMBINE_CENTER_RESULTS(r10, r9, stride, q6s16, q7s16, q8s16, q9s16);
+static INLINE void __STORE_COMBINE_CENTER_RESULTS(uint8_t *p1, uint8_t *p2,
+                                                  int stride, int16x8_t q6s16,
+                                                  int16x8_t q7s16,
+                                                  int16x8_t q8s16,
+                                                  int16x8_t q9s16) {
+  int16x4_t d8s16, d9s16, d10s16, d11s16;
+
+  d8s16 = vld1_s16((int16_t *)p1);
+  p1 += stride;
+  d11s16 = vld1_s16((int16_t *)p2);
+  p2 -= stride;
+  d9s16 = vld1_s16((int16_t *)p1);
+  d10s16 = vld1_s16((int16_t *)p2);
+
+  q7s16 = vrshrq_n_s16(q7s16, 6);
+  q8s16 = vrshrq_n_s16(q8s16, 6);
+  q9s16 = vrshrq_n_s16(q9s16, 6);
+  q6s16 = vrshrq_n_s16(q6s16, 6);
+
+  q7s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q7s16), vreinterpret_u8_s16(d9s16)));
+  q8s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_s16(d10s16)));
+  q9s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_s16(d11s16)));
+  q6s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q6s16), vreinterpret_u8_s16(d8s16)));
+
+  d9s16 = vreinterpret_s16_u8(vqmovun_s16(q7s16));
+  d10s16 = vreinterpret_s16_u8(vqmovun_s16(q8s16));
+  d11s16 = vreinterpret_s16_u8(vqmovun_s16(q9s16));
+  d8s16 = vreinterpret_s16_u8(vqmovun_s16(q6s16));
+
+  vst1_s16((int16_t *)p1, d9s16);
+  p1 -= stride;
+  vst1_s16((int16_t *)p2, d10s16);
+  p2 += stride;
+  vst1_s16((int16_t *)p1, d8s16);
+  vst1_s16((int16_t *)p2, d11s16);
+  return;
+}
+
+#define STORE_COMBINE_EXTREME_RESULTS(r7, r6) \
+  ;                                           \
+  __STORE_COMBINE_EXTREME_RESULTS(r7, r6, stride, q4s16, q5s16, q6s16, q7s16);
+static INLINE void __STORE_COMBINE_EXTREME_RESULTS(uint8_t *p1, uint8_t *p2,
+                                                   int stride, int16x8_t q4s16,
+                                                   int16x8_t q5s16,
+                                                   int16x8_t q6s16,
+                                                   int16x8_t q7s16) {
+  int16x4_t d4s16, d5s16, d6s16, d7s16;
+
+  d4s16 = vld1_s16((int16_t *)p1);
+  p1 += stride;
+  d7s16 = vld1_s16((int16_t *)p2);
+  p2 -= stride;
+  d5s16 = vld1_s16((int16_t *)p1);
+  d6s16 = vld1_s16((int16_t *)p2);
+
+  q5s16 = vrshrq_n_s16(q5s16, 6);
+  q6s16 = vrshrq_n_s16(q6s16, 6);
+  q7s16 = vrshrq_n_s16(q7s16, 6);
+  q4s16 = vrshrq_n_s16(q4s16, 6);
+
+  q5s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q5s16), vreinterpret_u8_s16(d5s16)));
+  q6s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q6s16), vreinterpret_u8_s16(d6s16)));
+  q7s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q7s16), vreinterpret_u8_s16(d7s16)));
+  q4s16 = vreinterpretq_s16_u16(
+      vaddw_u8(vreinterpretq_u16_s16(q4s16), vreinterpret_u8_s16(d4s16)));
+
+  d5s16 = vreinterpret_s16_u8(vqmovun_s16(q5s16));
+  d6s16 = vreinterpret_s16_u8(vqmovun_s16(q6s16));
+  d7s16 = vreinterpret_s16_u8(vqmovun_s16(q7s16));
+  d4s16 = vreinterpret_s16_u8(vqmovun_s16(q4s16));
+
+  vst1_s16((int16_t *)p1, d5s16);
+  p1 -= stride;
+  vst1_s16((int16_t *)p2, d6s16);
+  p2 += stride;
+  vst1_s16((int16_t *)p2, d7s16);
+  vst1_s16((int16_t *)p1, d4s16);
+  return;
+}
+
+#define DO_BUTTERFLY_STD(const_1, const_2, qA, qB) \
+  DO_BUTTERFLY(q14s16, q13s16, const_1, const_2, qA, qB);
+static INLINE void DO_BUTTERFLY(int16x8_t q14s16, int16x8_t q13s16,
+                                int16_t first_const, int16_t second_const,
+                                int16x8_t *qAs16, int16x8_t *qBs16) {
+  int16x4_t d30s16, d31s16;
+  int32x4_t q8s32, q9s32, q10s32, q11s32, q12s32, q15s32;
+  int16x4_t dCs16, dDs16, dAs16, dBs16;
+
+  dCs16 = vget_low_s16(q14s16);
+  dDs16 = vget_high_s16(q14s16);
+  dAs16 = vget_low_s16(q13s16);
+  dBs16 = vget_high_s16(q13s16);
+
+  d30s16 = vdup_n_s16(first_const);
+  d31s16 = vdup_n_s16(second_const);
+
+  q8s32 = vmull_s16(dCs16, d30s16);
+  q10s32 = vmull_s16(dAs16, d31s16);
+  q9s32 = vmull_s16(dDs16, d30s16);
+  q11s32 = vmull_s16(dBs16, d31s16);
+  q12s32 = vmull_s16(dCs16, d31s16);
+
+  q8s32 = vsubq_s32(q8s32, q10s32);
+  q9s32 = vsubq_s32(q9s32, q11s32);
+
+  q10s32 = vmull_s16(dDs16, d31s16);
+  q11s32 = vmull_s16(dAs16, d30s16);
+  q15s32 = vmull_s16(dBs16, d30s16);
+
+  q11s32 = vaddq_s32(q12s32, q11s32);
+  q10s32 = vaddq_s32(q10s32, q15s32);
+
+  *qAs16 = vcombine_s16(vqrshrn_n_s32(q8s32, 14), vqrshrn_n_s32(q9s32, 14));
+  *qBs16 = vcombine_s16(vqrshrn_n_s32(q11s32, 14), vqrshrn_n_s32(q10s32, 14));
+  return;
+}
+
+static INLINE void idct32_transpose_pair(int16_t *input, int16_t *t_buf) {
+  int16_t *in;
+  int i;
+  const int stride = 32;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+  int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+  for (i = 0; i < 4; i++, input += 8) {
+    in = input;
+    q8s16 = vld1q_s16(in);
+    in += stride;
+    q9s16 = vld1q_s16(in);
+    in += stride;
+    q10s16 = vld1q_s16(in);
+    in += stride;
+    q11s16 = vld1q_s16(in);
+    in += stride;
+    q12s16 = vld1q_s16(in);
+    in += stride;
+    q13s16 = vld1q_s16(in);
+    in += stride;
+    q14s16 = vld1q_s16(in);
+    in += stride;
+    q15s16 = vld1q_s16(in);
+
+    d16s16 = vget_low_s16(q8s16);
+    d17s16 = vget_high_s16(q8s16);
+    d18s16 = vget_low_s16(q9s16);
+    d19s16 = vget_high_s16(q9s16);
+    d20s16 = vget_low_s16(q10s16);
+    d21s16 = vget_high_s16(q10s16);
+    d22s16 = vget_low_s16(q11s16);
+    d23s16 = vget_high_s16(q11s16);
+    d24s16 = vget_low_s16(q12s16);
+    d25s16 = vget_high_s16(q12s16);
+    d26s16 = vget_low_s16(q13s16);
+    d27s16 = vget_high_s16(q13s16);
+    d28s16 = vget_low_s16(q14s16);
+    d29s16 = vget_high_s16(q14s16);
+    d30s16 = vget_low_s16(q15s16);
+    d31s16 = vget_high_s16(q15s16);
+
+    q8s16 = vcombine_s16(d16s16, d24s16);   // vswp d17, d24
+    q9s16 = vcombine_s16(d18s16, d26s16);   // vswp d19, d26
+    q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+    q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+    q12s16 = vcombine_s16(d17s16, d25s16);
+    q13s16 = vcombine_s16(d19s16, d27s16);
+    q14s16 = vcombine_s16(d21s16, d29s16);
+    q15s16 = vcombine_s16(d23s16, d31s16);
+
+    q0x2s32 =
+        vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q10s16));
+    q1x2s32 =
+        vtrnq_s32(vreinterpretq_s32_s16(q9s16), vreinterpretq_s32_s16(q11s16));
+    q2x2s32 =
+        vtrnq_s32(vreinterpretq_s32_s16(q12s16), vreinterpretq_s32_s16(q14s16));
+    q3x2s32 =
+        vtrnq_s32(vreinterpretq_s32_s16(q13s16), vreinterpretq_s32_s16(q15s16));
+
+    q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                        vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+    q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                        vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+    q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                        vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+    q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                        vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+    vst1q_s16(t_buf, q0x2s16.val[0]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q0x2s16.val[1]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q1x2s16.val[0]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q1x2s16.val[1]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q2x2s16.val[0]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q2x2s16.val[1]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q3x2s16.val[0]);
+    t_buf += 8;
+    vst1q_s16(t_buf, q3x2s16.val[1]);
+    t_buf += 8;
+  }
+  return;
+}
+
+static INLINE void idct32_bands_end_1st_pass(int16_t *out, int16x8_t q2s16,
+                                             int16x8_t q3s16, int16x8_t q6s16,
+                                             int16x8_t q7s16, int16x8_t q8s16,
+                                             int16x8_t q9s16, int16x8_t q10s16,
+                                             int16x8_t q11s16, int16x8_t q12s16,
+                                             int16x8_t q13s16, int16x8_t q14s16,
+                                             int16x8_t q15s16) {
+  int16x8_t q0s16, q1s16, q4s16, q5s16;
+
+  STORE_IN_OUTPUT(17, 16, 17, q6s16, q7s16);
+  STORE_IN_OUTPUT(17, 14, 15, q8s16, q9s16);
+
+  LOAD_FROM_OUTPUT(15, 30, 31, q0s16, q1s16);
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_IN_OUTPUT(31, 30, 31, q6s16, q7s16);
+  STORE_IN_OUTPUT(31, 0, 1, q4s16, q5s16);
+
+  LOAD_FROM_OUTPUT(1, 12, 13, q0s16, q1s16);
+  q2s16 = vaddq_s16(q10s16, q1s16);
+  q3s16 = vaddq_s16(q11s16, q0s16);
+  q4s16 = vsubq_s16(q11s16, q0s16);
+  q5s16 = vsubq_s16(q10s16, q1s16);
+
+  LOAD_FROM_OUTPUT(13, 18, 19, q0s16, q1s16);
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_IN_OUTPUT(19, 18, 19, q6s16, q7s16);
+  STORE_IN_OUTPUT(19, 12, 13, q8s16, q9s16);
+
+  LOAD_FROM_OUTPUT(13, 28, 29, q0s16, q1s16);
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_IN_OUTPUT(29, 28, 29, q6s16, q7s16);
+  STORE_IN_OUTPUT(29, 2, 3, q4s16, q5s16);
+
+  LOAD_FROM_OUTPUT(3, 10, 11, q0s16, q1s16);
+  q2s16 = vaddq_s16(q12s16, q1s16);
+  q3s16 = vaddq_s16(q13s16, q0s16);
+  q4s16 = vsubq_s16(q13s16, q0s16);
+  q5s16 = vsubq_s16(q12s16, q1s16);
+
+  LOAD_FROM_OUTPUT(11, 20, 21, q0s16, q1s16);
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_IN_OUTPUT(21, 20, 21, q6s16, q7s16);
+  STORE_IN_OUTPUT(21, 10, 11, q8s16, q9s16);
+
+  LOAD_FROM_OUTPUT(11, 26, 27, q0s16, q1s16);
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_IN_OUTPUT(27, 26, 27, q6s16, q7s16);
+  STORE_IN_OUTPUT(27, 4, 5, q4s16, q5s16);
+
+  LOAD_FROM_OUTPUT(5, 8, 9, q0s16, q1s16);
+  q2s16 = vaddq_s16(q14s16, q1s16);
+  q3s16 = vaddq_s16(q15s16, q0s16);
+  q4s16 = vsubq_s16(q15s16, q0s16);
+  q5s16 = vsubq_s16(q14s16, q1s16);
+
+  LOAD_FROM_OUTPUT(9, 22, 23, q0s16, q1s16);
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_IN_OUTPUT(23, 22, 23, q6s16, q7s16);
+  STORE_IN_OUTPUT(23, 8, 9, q8s16, q9s16);
+
+  LOAD_FROM_OUTPUT(9, 24, 25, q0s16, q1s16);
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_IN_OUTPUT(25, 24, 25, q6s16, q7s16);
+  STORE_IN_OUTPUT(25, 6, 7, q4s16, q5s16);
+  return;
+}
+
+static INLINE void idct32_bands_end_2nd_pass(
+    int16_t *out, uint8_t *dest, int stride, int16x8_t q2s16, int16x8_t q3s16,
+    int16x8_t q6s16, int16x8_t q7s16, int16x8_t q8s16, int16x8_t q9s16,
+    int16x8_t q10s16, int16x8_t q11s16, int16x8_t q12s16, int16x8_t q13s16,
+    int16x8_t q14s16, int16x8_t q15s16) {
+  uint8_t *r6 = dest + 31 * stride;
+  uint8_t *r7 = dest /* +  0 * stride*/;
+  uint8_t *r9 = dest + 15 * stride;
+  uint8_t *r10 = dest + 16 * stride;
+  int str2 = stride << 1;
+  int16x8_t q0s16, q1s16, q4s16, q5s16;
+
+  STORE_COMBINE_CENTER_RESULTS(r10, r9);
+  r10 += str2;
+  r9 -= str2;
+
+  LOAD_FROM_OUTPUT(17, 30, 31, q0s16, q1s16)
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+  r7 += str2;
+  r6 -= str2;
+
+  LOAD_FROM_OUTPUT(31, 12, 13, q0s16, q1s16)
+  q2s16 = vaddq_s16(q10s16, q1s16);
+  q3s16 = vaddq_s16(q11s16, q0s16);
+  q4s16 = vsubq_s16(q11s16, q0s16);
+  q5s16 = vsubq_s16(q10s16, q1s16);
+
+  LOAD_FROM_OUTPUT(13, 18, 19, q0s16, q1s16)
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_COMBINE_CENTER_RESULTS(r10, r9);
+  r10 += str2;
+  r9 -= str2;
+
+  LOAD_FROM_OUTPUT(19, 28, 29, q0s16, q1s16)
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+  r7 += str2;
+  r6 -= str2;
+
+  LOAD_FROM_OUTPUT(29, 10, 11, q0s16, q1s16)
+  q2s16 = vaddq_s16(q12s16, q1s16);
+  q3s16 = vaddq_s16(q13s16, q0s16);
+  q4s16 = vsubq_s16(q13s16, q0s16);
+  q5s16 = vsubq_s16(q12s16, q1s16);
+
+  LOAD_FROM_OUTPUT(11, 20, 21, q0s16, q1s16)
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_COMBINE_CENTER_RESULTS(r10, r9);
+  r10 += str2;
+  r9 -= str2;
+
+  LOAD_FROM_OUTPUT(21, 26, 27, q0s16, q1s16)
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+  r7 += str2;
+  r6 -= str2;
+
+  LOAD_FROM_OUTPUT(27, 8, 9, q0s16, q1s16)
+  q2s16 = vaddq_s16(q14s16, q1s16);
+  q3s16 = vaddq_s16(q15s16, q0s16);
+  q4s16 = vsubq_s16(q15s16, q0s16);
+  q5s16 = vsubq_s16(q14s16, q1s16);
+
+  LOAD_FROM_OUTPUT(9, 22, 23, q0s16, q1s16)
+  q8s16 = vaddq_s16(q4s16, q1s16);
+  q9s16 = vaddq_s16(q5s16, q0s16);
+  q6s16 = vsubq_s16(q5s16, q0s16);
+  q7s16 = vsubq_s16(q4s16, q1s16);
+  STORE_COMBINE_CENTER_RESULTS(r10, r9);
+
+  LOAD_FROM_OUTPUT(23, 24, 25, q0s16, q1s16)
+  q4s16 = vaddq_s16(q2s16, q1s16);
+  q5s16 = vaddq_s16(q3s16, q0s16);
+  q6s16 = vsubq_s16(q3s16, q0s16);
+  q7s16 = vsubq_s16(q2s16, q1s16);
+  STORE_COMBINE_EXTREME_RESULTS(r7, r6);
+  return;
+}
+
+void aom_idct32x32_1024_add_neon(int16_t *input, uint8_t *dest, int stride) {
+  int i, idct32_pass_loop;
+  int16_t trans_buf[32 * 8];
+  int16_t pass1[32 * 32];
+  int16_t pass2[32 * 32];
+  int16_t *out;
+  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+
+  for (idct32_pass_loop = 0, out = pass1; idct32_pass_loop < 2;
+       idct32_pass_loop++,
+      input = pass1,  // the input of pass2 is the result of pass1
+       out = pass2) {
+    for (i = 0; i < 4; i++, input += 32 * 8, out += 8) {  // idct32_bands_loop
+      idct32_transpose_pair(input, trans_buf);
+
+      // -----------------------------------------
+      // BLOCK A: 16-19,28-31
+      // -----------------------------------------
+      // generate 16,17,30,31
+      // part of stage 1
+      LOAD_FROM_TRANSPOSED(0, 1, 31)
+      DO_BUTTERFLY_STD(cospi_31_64, cospi_1_64, &q0s16, &q2s16)
+      LOAD_FROM_TRANSPOSED(31, 17, 15)
+      DO_BUTTERFLY_STD(cospi_15_64, cospi_17_64, &q1s16, &q3s16)
+      // part of stage 2
+      q4s16 = vaddq_s16(q0s16, q1s16);
+      q13s16 = vsubq_s16(q0s16, q1s16);
+      q6s16 = vaddq_s16(q2s16, q3s16);
+      q14s16 = vsubq_s16(q2s16, q3s16);
+      // part of stage 3
+      DO_BUTTERFLY_STD(cospi_28_64, cospi_4_64, &q5s16, &q7s16)
+
+      // generate 18,19,28,29
+      // part of stage 1
+      LOAD_FROM_TRANSPOSED(15, 9, 23)
+      DO_BUTTERFLY_STD(cospi_23_64, cospi_9_64, &q0s16, &q2s16)
+      LOAD_FROM_TRANSPOSED(23, 25, 7)
+      DO_BUTTERFLY_STD(cospi_7_64, cospi_25_64, &q1s16, &q3s16)
+      // part of stage 2
+      q13s16 = vsubq_s16(q3s16, q2s16);
+      q3s16 = vaddq_s16(q3s16, q2s16);
+      q14s16 = vsubq_s16(q1s16, q0s16);
+      q2s16 = vaddq_s16(q1s16, q0s16);
+      // part of stage 3
+      DO_BUTTERFLY_STD(-cospi_4_64, -cospi_28_64, &q1s16, &q0s16)
+      // part of stage 4
+      q8s16 = vaddq_s16(q4s16, q2s16);
+      q9s16 = vaddq_s16(q5s16, q0s16);
+      q10s16 = vaddq_s16(q7s16, q1s16);
+      q15s16 = vaddq_s16(q6s16, q3s16);
+      q13s16 = vsubq_s16(q5s16, q0s16);
+      q14s16 = vsubq_s16(q7s16, q1s16);
+      STORE_IN_OUTPUT(0, 16, 31, q8s16, q15s16)
+      STORE_IN_OUTPUT(31, 17, 30, q9s16, q10s16)
+      // part of stage 5
+      DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q0s16, &q1s16)
+      STORE_IN_OUTPUT(30, 29, 18, q1s16, q0s16)
+      // part of stage 4
+      q13s16 = vsubq_s16(q4s16, q2s16);
+      q14s16 = vsubq_s16(q6s16, q3s16);
+      // part of stage 5
+      DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q4s16, &q6s16)
+      STORE_IN_OUTPUT(18, 19, 28, q4s16, q6s16)
+
+      // -----------------------------------------
+      // BLOCK B: 20-23,24-27
+      // -----------------------------------------
+      // generate 20,21,26,27
+      // part of stage 1
+      LOAD_FROM_TRANSPOSED(7, 5, 27)
+      DO_BUTTERFLY_STD(cospi_27_64, cospi_5_64, &q0s16, &q2s16)
+      LOAD_FROM_TRANSPOSED(27, 21, 11)
+      DO_BUTTERFLY_STD(cospi_11_64, cospi_21_64, &q1s16, &q3s16)
+      // part of stage 2
+      q13s16 = vsubq_s16(q0s16, q1s16);
+      q0s16 = vaddq_s16(q0s16, q1s16);
+      q14s16 = vsubq_s16(q2s16, q3s16);
+      q2s16 = vaddq_s16(q2s16, q3s16);
+      // part of stage 3
+      DO_BUTTERFLY_STD(cospi_12_64, cospi_20_64, &q1s16, &q3s16)
+
+      // generate 22,23,24,25
+      // part of stage 1
+      LOAD_FROM_TRANSPOSED(11, 13, 19)
+      DO_BUTTERFLY_STD(cospi_19_64, cospi_13_64, &q5s16, &q7s16)
+      LOAD_FROM_TRANSPOSED(19, 29, 3)
+      DO_BUTTERFLY_STD(cospi_3_64, cospi_29_64, &q4s16, &q6s16)
+      // part of stage 2
+      q14s16 = vsubq_s16(q4s16, q5s16);
+      q5s16 = vaddq_s16(q4s16, q5s16);
+      q13s16 = vsubq_s16(q6s16, q7s16);
+      q6s16 = vaddq_s16(q6s16, q7s16);
+      // part of stage 3
+      DO_BUTTERFLY_STD(-cospi_20_64, -cospi_12_64, &q4s16, &q7s16)
+      // part of stage 4
+      q10s16 = vaddq_s16(q7s16, q1s16);
+      q11s16 = vaddq_s16(q5s16, q0s16);
+      q12s16 = vaddq_s16(q6s16, q2s16);
+      q15s16 = vaddq_s16(q4s16, q3s16);
+      // part of stage 6
+      LOAD_FROM_OUTPUT(28, 16, 17, q14s16, q13s16)
+      q8s16 = vaddq_s16(q14s16, q11s16);
+      q9s16 = vaddq_s16(q13s16, q10s16);
+      q13s16 = vsubq_s16(q13s16, q10s16);
+      q11s16 = vsubq_s16(q14s16, q11s16);
+      STORE_IN_OUTPUT(17, 17, 16, q9s16, q8s16)
+      LOAD_FROM_OUTPUT(16, 30, 31, q14s16, q9s16)
+      q8s16 = vsubq_s16(q9s16, q12s16);
+      q10s16 = vaddq_s16(q14s16, q15s16);
+      q14s16 = vsubq_s16(q14s16, q15s16);
+      q12s16 = vaddq_s16(q9s16, q12s16);
+      STORE_IN_OUTPUT(31, 30, 31, q10s16, q12s16)
+      // part of stage 7
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+      STORE_IN_OUTPUT(31, 25, 22, q14s16, q13s16)
+      q13s16 = q11s16;
+      q14s16 = q8s16;
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+      STORE_IN_OUTPUT(22, 24, 23, q14s16, q13s16)
+      // part of stage 4
+      q14s16 = vsubq_s16(q5s16, q0s16);
+      q13s16 = vsubq_s16(q6s16, q2s16);
+      DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q5s16, &q6s16);
+      q14s16 = vsubq_s16(q7s16, q1s16);
+      q13s16 = vsubq_s16(q4s16, q3s16);
+      DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q0s16, &q1s16);
+      // part of stage 6
+      LOAD_FROM_OUTPUT(23, 18, 19, q14s16, q13s16)
+      q8s16 = vaddq_s16(q14s16, q1s16);
+      q9s16 = vaddq_s16(q13s16, q6s16);
+      q13s16 = vsubq_s16(q13s16, q6s16);
+      q1s16 = vsubq_s16(q14s16, q1s16);
+      STORE_IN_OUTPUT(19, 18, 19, q8s16, q9s16)
+      LOAD_FROM_OUTPUT(19, 28, 29, q8s16, q9s16)
+      q14s16 = vsubq_s16(q8s16, q5s16);
+      q10s16 = vaddq_s16(q8s16, q5s16);
+      q11s16 = vaddq_s16(q9s16, q0s16);
+      q0s16 = vsubq_s16(q9s16, q0s16);
+      STORE_IN_OUTPUT(29, 28, 29, q10s16, q11s16)
+      // part of stage 7
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q13s16, &q14s16)
+      STORE_IN_OUTPUT(29, 20, 27, q13s16, q14s16)
+      DO_BUTTERFLY(q0s16, q1s16, cospi_16_64, cospi_16_64, &q1s16, &q0s16);
+      STORE_IN_OUTPUT(27, 21, 26, q1s16, q0s16)
+
+      // -----------------------------------------
+      // BLOCK C: 8-10,11-15
+      // -----------------------------------------
+      // generate 8,9,14,15
+      // part of stage 2
+      LOAD_FROM_TRANSPOSED(3, 2, 30)
+      DO_BUTTERFLY_STD(cospi_30_64, cospi_2_64, &q0s16, &q2s16)
+      LOAD_FROM_TRANSPOSED(30, 18, 14)
+      DO_BUTTERFLY_STD(cospi_14_64, cospi_18_64, &q1s16, &q3s16)
+      // part of stage 3
+      q13s16 = vsubq_s16(q0s16, q1s16);
+      q0s16 = vaddq_s16(q0s16, q1s16);
+      q14s16 = vsubq_s16(q2s16, q3s16);
+      q2s16 = vaddq_s16(q2s16, q3s16);
+      // part of stage 4
+      DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q1s16, &q3s16)
+
+      // generate 10,11,12,13
+      // part of stage 2
+      LOAD_FROM_TRANSPOSED(14, 10, 22)
+      DO_BUTTERFLY_STD(cospi_22_64, cospi_10_64, &q5s16, &q7s16)
+      LOAD_FROM_TRANSPOSED(22, 26, 6)
+      DO_BUTTERFLY_STD(cospi_6_64, cospi_26_64, &q4s16, &q6s16)
+      // part of stage 3
+      q14s16 = vsubq_s16(q4s16, q5s16);
+      q5s16 = vaddq_s16(q4s16, q5s16);
+      q13s16 = vsubq_s16(q6s16, q7s16);
+      q6s16 = vaddq_s16(q6s16, q7s16);
+      // part of stage 4
+      DO_BUTTERFLY_STD(-cospi_8_64, -cospi_24_64, &q4s16, &q7s16)
+      // part of stage 5
+      q8s16 = vaddq_s16(q0s16, q5s16);
+      q9s16 = vaddq_s16(q1s16, q7s16);
+      q13s16 = vsubq_s16(q1s16, q7s16);
+      q14s16 = vsubq_s16(q3s16, q4s16);
+      q10s16 = vaddq_s16(q3s16, q4s16);
+      q15s16 = vaddq_s16(q2s16, q6s16);
+      STORE_IN_OUTPUT(26, 8, 15, q8s16, q15s16)
+      STORE_IN_OUTPUT(15, 9, 14, q9s16, q10s16)
+      // part of stage 6
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+      STORE_IN_OUTPUT(14, 13, 10, q3s16, q1s16)
+      q13s16 = vsubq_s16(q0s16, q5s16);
+      q14s16 = vsubq_s16(q2s16, q6s16);
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+      STORE_IN_OUTPUT(10, 11, 12, q1s16, q3s16)
+
+      // -----------------------------------------
+      // BLOCK D: 0-3,4-7
+      // -----------------------------------------
+      // generate 4,5,6,7
+      // part of stage 3
+      LOAD_FROM_TRANSPOSED(6, 4, 28)
+      DO_BUTTERFLY_STD(cospi_28_64, cospi_4_64, &q0s16, &q2s16)
+      LOAD_FROM_TRANSPOSED(28, 20, 12)
+      DO_BUTTERFLY_STD(cospi_12_64, cospi_20_64, &q1s16, &q3s16)
+      // part of stage 4
+      q13s16 = vsubq_s16(q0s16, q1s16);
+      q0s16 = vaddq_s16(q0s16, q1s16);
+      q14s16 = vsubq_s16(q2s16, q3s16);
+      q2s16 = vaddq_s16(q2s16, q3s16);
+      // part of stage 5
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q1s16, &q3s16)
+
+      // generate 0,1,2,3
+      // part of stage 4
+      LOAD_FROM_TRANSPOSED(12, 0, 16)
+      DO_BUTTERFLY_STD(cospi_16_64, cospi_16_64, &q5s16, &q7s16)
+      LOAD_FROM_TRANSPOSED(16, 8, 24)
+      DO_BUTTERFLY_STD(cospi_24_64, cospi_8_64, &q14s16, &q6s16)
+      // part of stage 5
+      q4s16 = vaddq_s16(q7s16, q6s16);
+      q7s16 = vsubq_s16(q7s16, q6s16);
+      q6s16 = vsubq_s16(q5s16, q14s16);
+      q5s16 = vaddq_s16(q5s16, q14s16);
+      // part of stage 6
+      q8s16 = vaddq_s16(q4s16, q2s16);
+      q9s16 = vaddq_s16(q5s16, q3s16);
+      q10s16 = vaddq_s16(q6s16, q1s16);
+      q11s16 = vaddq_s16(q7s16, q0s16);
+      q12s16 = vsubq_s16(q7s16, q0s16);
+      q13s16 = vsubq_s16(q6s16, q1s16);
+      q14s16 = vsubq_s16(q5s16, q3s16);
+      q15s16 = vsubq_s16(q4s16, q2s16);
+      // part of stage 7
+      LOAD_FROM_OUTPUT(12, 14, 15, q0s16, q1s16)
+      q2s16 = vaddq_s16(q8s16, q1s16);
+      q3s16 = vaddq_s16(q9s16, q0s16);
+      q4s16 = vsubq_s16(q9s16, q0s16);
+      q5s16 = vsubq_s16(q8s16, q1s16);
+      LOAD_FROM_OUTPUT(15, 16, 17, q0s16, q1s16)
+      q8s16 = vaddq_s16(q4s16, q1s16);
+      q9s16 = vaddq_s16(q5s16, q0s16);
+      q6s16 = vsubq_s16(q5s16, q0s16);
+      q7s16 = vsubq_s16(q4s16, q1s16);
+
+      if (idct32_pass_loop == 0) {
+        idct32_bands_end_1st_pass(out, q2s16, q3s16, q6s16, q7s16, q8s16, q9s16,
+                                  q10s16, q11s16, q12s16, q13s16, q14s16,
+                                  q15s16);
+      } else {
+        idct32_bands_end_2nd_pass(out, dest, stride, q2s16, q3s16, q6s16, q7s16,
+                                  q8s16, q9s16, q10s16, q11s16, q12s16, q13s16,
+                                  q14s16, q15s16);
+        dest += 8;
+      }
+    }
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/idct4x4_1_add_neon.asm b/third_party/aom/aom_dsp/arm/idct4x4_1_add_neon.asm
new file mode 100644
index 0000000000..6bd733d5dd
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct4x4_1_add_neon.asm
@@ -0,0 +1,71 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+
+
+    EXPORT  |aom_idct4x4_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void aom_idct4x4_1_add_neon(int16_t *input, uint8_t *dest,
+;                                  int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct4x4_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 4)
+    add              r0, r0, #8                ; + (1 <<((4) - 1))
+    asr              r0, r0, #4                ; >> 4
+
+    vdup.s16         q0, r0                    ; duplicate a1
+
+    vld1.32          {d2[0]}, [r1], r2
+    vld1.32          {d2[1]}, [r1], r2
+    vld1.32          {d4[0]}, [r1], r2
+    vld1.32          {d4[1]}, [r1]
+
+    vaddw.u8         q8, q0, d2                ; dest[x] + a1
+    vaddw.u8         q9, q0, d4
+
+    vqmovun.s16      d6, q8                    ; clip_pixel
+    vqmovun.s16      d7, q9
+
+    vst1.32          {d6[0]}, [r12], r2
+    vst1.32          {d6[1]}, [r12], r2
+    vst1.32          {d7[0]}, [r12], r2
+    vst1.32          {d7[1]}, [r12]
+
+    bx               lr
+    ENDP             ; |aom_idct4x4_1_add_neon|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/idct4x4_1_add_neon.c b/third_party/aom/aom_dsp/arm/idct4x4_1_add_neon.c
new file mode 100644
index 0000000000..3df7a901b0
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct4x4_1_add_neon.c
@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "aom_dsp/inv_txfm.h"
+#include "aom_ports/mem.h"
+
+void aom_idct4x4_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8x8_t d6u8;
+  uint32x2_t d2u32 = vdup_n_u32(0);
+  uint16x8_t q8u16;
+  int16x8_t q0s16;
+  uint8_t *d1, *d2;
+  int16_t i, a1;
+  int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+  out = dct_const_round_shift(out * cospi_16_64);
+  a1 = ROUND_POWER_OF_TWO(out, 4);
+
+  q0s16 = vdupq_n_s16(a1);
+
+  // dc_only_idct_add
+  d1 = d2 = dest;
+  for (i = 0; i < 2; i++) {
+    d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 0);
+    d1 += dest_stride;
+    d2u32 = vld1_lane_u32((const uint32_t *)d1, d2u32, 1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q0s16), vreinterpret_u8_u32(d2u32));
+    d6u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+
+    vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 0);
+    d2 += dest_stride;
+    vst1_lane_u32((uint32_t *)d2, vreinterpret_u32_u8(d6u8), 1);
+    d2 += dest_stride;
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/idct4x4_add_neon.asm b/third_party/aom/aom_dsp/arm/idct4x4_add_neon.asm
new file mode 100644
index 0000000000..127acf6140
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct4x4_add_neon.asm
@@ -0,0 +1,193 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_idct4x4_16_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    AREA     Block, CODE, READONLY ; name this block of code
+;void aom_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct4x4_16_add_neon| PROC
+
+    ; The 2D transform is done with two passes which are actually pretty
+    ; similar. We first transform the rows. This is done by transposing
+    ; the inputs, doing an SIMD column transform (the columns are the
+    ; transposed rows) and then transpose the results (so that it goes back
+    ; in normal/row positions). Then, we transform the columns by doing
+    ; another SIMD column transform.
+    ; So, two passes of a transpose followed by a column transform.
+
+    ; load the inputs into q8-q9, d16-d19
+    vld1.s16        {q8,q9}, [r0]!
+
+    ; generate scalar constants
+    ; cospi_8_64 = 15137 = 0x3b21
+    mov             r0, #0x3b00
+    add             r0, #0x21
+    ; cospi_16_64 = 11585 = 0x2d41
+    mov             r3, #0x2d00
+    add             r3, #0x41
+    ; cospi_24_64 = 6270 = 0x 187e
+    mov             r12, #0x1800
+    add             r12, #0x7e
+
+    ; transpose the input data
+    ; 00 01 02 03   d16
+    ; 10 11 12 13   d17
+    ; 20 21 22 23   d18
+    ; 30 31 32 33   d19
+    vtrn.16         d16, d17
+    vtrn.16         d18, d19
+
+    ; generate constant vectors
+    vdup.16         d20, r0         ; replicate cospi_8_64
+    vdup.16         d21, r3         ; replicate cospi_16_64
+
+    ; 00 10 02 12   d16
+    ; 01 11 03 13   d17
+    ; 20 30 22 32   d18
+    ; 21 31 23 33   d19
+    vtrn.32         q8, q9
+    ; 00 10 20 30   d16
+    ; 01 11 21 31   d17
+    ; 02 12 22 32   d18
+    ; 03 13 23 33   d19
+
+    vdup.16         d22, r12        ; replicate cospi_24_64
+
+    ; do the transform on transposed rows
+
+    ; stage 1
+    vadd.s16  d23, d16, d18         ; (input[0] + input[2])
+    vsub.s16  d24, d16, d18         ; (input[0] - input[2])
+
+    vmull.s16 q15, d17, d22         ; input[1] * cospi_24_64
+    vmull.s16 q1,  d17, d20         ; input[1] * cospi_8_64
+
+    ; (input[0] + input[2]) * cospi_16_64;
+    ; (input[0] - input[2]) * cospi_16_64;
+    vmull.s16 q13, d23, d21
+    vmull.s16 q14, d24, d21
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64;
+    ; input[1] * cospi_8_64  + input[3] * cospi_24_64;
+    vmlsl.s16 q15, d19, d20
+    vmlal.s16 q1,  d19, d22
+
+    ; dct_const_round_shift
+    vqrshrn.s32 d26, q13, #14
+    vqrshrn.s32 d27, q14, #14
+    vqrshrn.s32 d29, q15, #14
+    vqrshrn.s32 d28, q1,  #14
+
+    ; stage 2
+    ; output[0] = step[0] + step[3];
+    ; output[1] = step[1] + step[2];
+    ; output[3] = step[0] - step[3];
+    ; output[2] = step[1] - step[2];
+    vadd.s16 q8,  q13, q14
+    vsub.s16 q9,  q13, q14
+    vswp     d18, d19
+
+    ; transpose the results
+    ; 00 01 02 03   d16
+    ; 10 11 12 13   d17
+    ; 20 21 22 23   d18
+    ; 30 31 32 33   d19
+    vtrn.16         d16, d17
+    vtrn.16         d18, d19
+    ; 00 10 02 12   d16
+    ; 01 11 03 13   d17
+    ; 20 30 22 32   d18
+    ; 21 31 23 33   d19
+    vtrn.32         q8, q9
+    ; 00 10 20 30   d16
+    ; 01 11 21 31   d17
+    ; 02 12 22 32   d18
+    ; 03 13 23 33   d19
+
+    ; do the transform on columns
+
+    ; stage 1
+    vadd.s16  d23, d16, d18         ; (input[0] + input[2])
+    vsub.s16  d24, d16, d18         ; (input[0] - input[2])
+
+    vmull.s16 q15, d17, d22         ; input[1] * cospi_24_64
+    vmull.s16 q1,  d17, d20         ; input[1] * cospi_8_64
+
+    ; (input[0] + input[2]) * cospi_16_64;
+    ; (input[0] - input[2]) * cospi_16_64;
+    vmull.s16 q13, d23, d21
+    vmull.s16 q14, d24, d21
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64;
+    ; input[1] * cospi_8_64  + input[3] * cospi_24_64;
+    vmlsl.s16 q15, d19, d20
+    vmlal.s16 q1,  d19, d22
+
+    ; dct_const_round_shift
+    vqrshrn.s32 d26, q13, #14
+    vqrshrn.s32 d27, q14, #14
+    vqrshrn.s32 d29, q15, #14
+    vqrshrn.s32 d28, q1,  #14
+
+    ; stage 2
+    ; output[0] = step[0] + step[3];
+    ; output[1] = step[1] + step[2];
+    ; output[3] = step[0] - step[3];
+    ; output[2] = step[1] - step[2];
+    vadd.s16 q8,  q13, q14
+    vsub.s16 q9,  q13, q14
+
+    ; The results are in two registers, one of them being swapped. This will
+    ; be taken care of by loading the 'dest' value in a swapped fashion and
+    ; also storing them in the same swapped fashion.
+    ; temp_out[0, 1] = d16, d17 = q8
+    ; temp_out[2, 3] = d19, d18 = q9 swapped
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 4)
+    vrshr.s16 q8, q8, #4
+    vrshr.s16 q9, q9, #4
+
+    vld1.32 {d26[0]}, [r1], r2
+    vld1.32 {d26[1]}, [r1], r2
+    vld1.32 {d27[1]}, [r1], r2
+    vld1.32 {d27[0]}, [r1]  ; no post-increment
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 4) + dest[j * dest_stride + i]
+    vaddw.u8 q8, q8, d26
+    vaddw.u8 q9, q9, d27
+
+    ; clip_pixel
+    vqmovun.s16 d26, q8
+    vqmovun.s16 d27, q9
+
+    ; do the stores in reverse order with negative post-increment, by changing
+    ; the sign of the stride
+    rsb r2, r2, #0
+    vst1.32 {d27[0]}, [r1], r2
+    vst1.32 {d27[1]}, [r1], r2
+    vst1.32 {d26[1]}, [r1], r2
+    vst1.32 {d26[0]}, [r1]  ; no post-increment
+    bx              lr
+    ENDP  ; |aom_idct4x4_16_add_neon|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/idct4x4_add_neon.c b/third_party/aom/aom_dsp/arm/idct4x4_add_neon.c
new file mode 100644
index 0000000000..763be1ab0e
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct4x4_add_neon.c
@@ -0,0 +1,146 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "aom_dsp/txfm_common.h"
+
+void aom_idct4x4_16_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8x8_t d26u8, d27u8;
+  uint32x2_t d26u32, d27u32;
+  uint16x8_t q8u16, q9u16;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16;
+  int16x4_t d22s16, d23s16, d24s16, d26s16, d27s16, d28s16, d29s16;
+  int16x8_t q8s16, q9s16, q13s16, q14s16;
+  int32x4_t q1s32, q13s32, q14s32, q15s32;
+  int16x4x2_t d0x2s16, d1x2s16;
+  int32x4x2_t q0x2s32;
+  uint8_t *d;
+
+  d26u32 = d27u32 = vdup_n_u32(0);
+
+  q8s16 = vld1q_s16(input);
+  q9s16 = vld1q_s16(input + 8);
+
+  d16s16 = vget_low_s16(q8s16);
+  d17s16 = vget_high_s16(q8s16);
+  d18s16 = vget_low_s16(q9s16);
+  d19s16 = vget_high_s16(q9s16);
+
+  d0x2s16 = vtrn_s16(d16s16, d17s16);
+  d1x2s16 = vtrn_s16(d18s16, d19s16);
+  q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
+  q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
+
+  d20s16 = vdup_n_s16((int16_t)cospi_8_64);
+  d21s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q0x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
+  d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+  d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+  d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+  d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+
+  d22s16 = vdup_n_s16((int16_t)cospi_24_64);
+
+  // stage 1
+  d23s16 = vadd_s16(d16s16, d18s16);
+  d24s16 = vsub_s16(d16s16, d18s16);
+
+  q15s32 = vmull_s16(d17s16, d22s16);
+  q1s32 = vmull_s16(d17s16, d20s16);
+  q13s32 = vmull_s16(d23s16, d21s16);
+  q14s32 = vmull_s16(d24s16, d21s16);
+
+  q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
+  q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
+
+  d26s16 = vqrshrn_n_s32(q13s32, 14);
+  d27s16 = vqrshrn_n_s32(q14s32, 14);
+  d29s16 = vqrshrn_n_s32(q15s32, 14);
+  d28s16 = vqrshrn_n_s32(q1s32, 14);
+  q13s16 = vcombine_s16(d26s16, d27s16);
+  q14s16 = vcombine_s16(d28s16, d29s16);
+
+  // stage 2
+  q8s16 = vaddq_s16(q13s16, q14s16);
+  q9s16 = vsubq_s16(q13s16, q14s16);
+
+  d16s16 = vget_low_s16(q8s16);
+  d17s16 = vget_high_s16(q8s16);
+  d18s16 = vget_high_s16(q9s16);  // vswp d18 d19
+  d19s16 = vget_low_s16(q9s16);
+
+  d0x2s16 = vtrn_s16(d16s16, d17s16);
+  d1x2s16 = vtrn_s16(d18s16, d19s16);
+  q8s16 = vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]);
+  q9s16 = vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]);
+
+  q0x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(q8s16), vreinterpretq_s32_s16(q9s16));
+  d16s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+  d17s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[0]));
+  d18s16 = vget_low_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+  d19s16 = vget_high_s16(vreinterpretq_s16_s32(q0x2s32.val[1]));
+
+  // do the transform on columns
+  // stage 1
+  d23s16 = vadd_s16(d16s16, d18s16);
+  d24s16 = vsub_s16(d16s16, d18s16);
+
+  q15s32 = vmull_s16(d17s16, d22s16);
+  q1s32 = vmull_s16(d17s16, d20s16);
+  q13s32 = vmull_s16(d23s16, d21s16);
+  q14s32 = vmull_s16(d24s16, d21s16);
+
+  q15s32 = vmlsl_s16(q15s32, d19s16, d20s16);
+  q1s32 = vmlal_s16(q1s32, d19s16, d22s16);
+
+  d26s16 = vqrshrn_n_s32(q13s32, 14);
+  d27s16 = vqrshrn_n_s32(q14s32, 14);
+  d29s16 = vqrshrn_n_s32(q15s32, 14);
+  d28s16 = vqrshrn_n_s32(q1s32, 14);
+  q13s16 = vcombine_s16(d26s16, d27s16);
+  q14s16 = vcombine_s16(d28s16, d29s16);
+
+  // stage 2
+  q8s16 = vaddq_s16(q13s16, q14s16);
+  q9s16 = vsubq_s16(q13s16, q14s16);
+
+  q8s16 = vrshrq_n_s16(q8s16, 4);
+  q9s16 = vrshrq_n_s16(q9s16, 4);
+
+  d = dest;
+  d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 0);
+  d += dest_stride;
+  d26u32 = vld1_lane_u32((const uint32_t *)d, d26u32, 1);
+  d += dest_stride;
+  d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 1);
+  d += dest_stride;
+  d27u32 = vld1_lane_u32((const uint32_t *)d, d27u32, 0);
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
+
+  d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+
+  d = dest;
+  vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 0);
+  d += dest_stride;
+  vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d26u8), 1);
+  d += dest_stride;
+  vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 1);
+  d += dest_stride;
+  vst1_lane_u32((uint32_t *)d, vreinterpret_u32_u8(d27u8), 0);
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/idct8x8_1_add_neon.asm b/third_party/aom/aom_dsp/arm/idct8x8_1_add_neon.asm
new file mode 100644
index 0000000000..ec07e2053b
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct8x8_1_add_neon.asm
@@ -0,0 +1,91 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+
+
+    EXPORT  |aom_idct8x8_1_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void aom_idct8x8_1_add_neon(int16_t *input, uint8_t *dest,
+;                                  int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct8x8_1_add_neon| PROC
+    ldrsh            r0, [r0]
+
+    ; generate cospi_16_64 = 11585
+    mov              r12, #0x2d00
+    add              r12, #0x41
+
+    ; out = dct_const_round_shift(input[0] * cospi_16_64)
+    mul              r0, r0, r12               ; input[0] * cospi_16_64
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; out = dct_const_round_shift(out * cospi_16_64)
+    mul              r0, r0, r12               ; out * cospi_16_64
+    mov              r12, r1                   ; save dest
+    add              r0, r0, #0x2000           ; +(1 << ((DCT_CONST_BITS) - 1))
+    asr              r0, r0, #14               ; >> DCT_CONST_BITS
+
+    ; a1 = ROUND_POWER_OF_TWO(out, 5)
+    add              r0, r0, #16               ; + (1 <<((5) - 1))
+    asr              r0, r0, #5                ; >> 5
+
+    vdup.s16         q0, r0                    ; duplicate a1
+
+    ; load destination data
+    vld1.64          {d2}, [r1], r2
+    vld1.64          {d3}, [r1], r2
+    vld1.64          {d4}, [r1], r2
+    vld1.64          {d5}, [r1], r2
+    vld1.64          {d6}, [r1], r2
+    vld1.64          {d7}, [r1], r2
+    vld1.64          {d16}, [r1], r2
+    vld1.64          {d17}, [r1]
+
+    vaddw.u8         q9, q0, d2                ; dest[x] + a1
+    vaddw.u8         q10, q0, d3               ; dest[x] + a1
+    vaddw.u8         q11, q0, d4               ; dest[x] + a1
+    vaddw.u8         q12, q0, d5               ; dest[x] + a1
+    vqmovun.s16      d2, q9                    ; clip_pixel
+    vqmovun.s16      d3, q10                   ; clip_pixel
+    vqmovun.s16      d30, q11                  ; clip_pixel
+    vqmovun.s16      d31, q12                  ; clip_pixel
+    vst1.64          {d2}, [r12], r2
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r2
+    vst1.64          {d31}, [r12], r2
+
+    vaddw.u8         q9, q0, d6                 ; dest[x] + a1
+    vaddw.u8         q10, q0, d7                ; dest[x] + a1
+    vaddw.u8         q11, q0, d16               ; dest[x] + a1
+    vaddw.u8         q12, q0, d17               ; dest[x] + a1
+    vqmovun.s16      d2, q9                     ; clip_pixel
+    vqmovun.s16      d3, q10                    ; clip_pixel
+    vqmovun.s16      d30, q11                   ; clip_pixel
+    vqmovun.s16      d31, q12                   ; clip_pixel
+    vst1.64          {d2}, [r12], r2
+    vst1.64          {d3}, [r12], r2
+    vst1.64          {d30}, [r12], r2
+    vst1.64          {d31}, [r12], r2
+
+    bx               lr
+    ENDP             ; |aom_idct8x8_1_add_neon|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/idct8x8_1_add_neon.c b/third_party/aom/aom_dsp/arm/idct8x8_1_add_neon.c
new file mode 100644
index 0000000000..c7926f9e4f
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct8x8_1_add_neon.c
@@ -0,0 +1,62 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "aom_dsp/inv_txfm.h"
+#include "aom_ports/mem.h"
+
+void aom_idct8x8_1_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8x8_t d2u8, d3u8, d30u8, d31u8;
+  uint64x1_t d2u64, d3u64, d4u64, d5u64;
+  uint16x8_t q0u16, q9u16, q10u16, q11u16, q12u16;
+  int16x8_t q0s16;
+  uint8_t *d1, *d2;
+  int16_t i, a1;
+  int16_t out = dct_const_round_shift(input[0] * cospi_16_64);
+  out = dct_const_round_shift(out * cospi_16_64);
+  a1 = ROUND_POWER_OF_TWO(out, 5);
+
+  q0s16 = vdupq_n_s16(a1);
+  q0u16 = vreinterpretq_u16_s16(q0s16);
+
+  d1 = d2 = dest;
+  for (i = 0; i < 2; i++) {
+    d2u64 = vld1_u64((const uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((const uint64_t *)d1);
+    d1 += dest_stride;
+    d4u64 = vld1_u64((const uint64_t *)d1);
+    d1 += dest_stride;
+    d5u64 = vld1_u64((const uint64_t *)d1);
+    d1 += dest_stride;
+
+    q9u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d2u64));
+    q10u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d3u64));
+    q11u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d4u64));
+    q12u16 = vaddw_u8(q0u16, vreinterpret_u8_u64(d5u64));
+
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d30u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+    d31u8 = vqmovun_s16(vreinterpretq_s16_u16(q12u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d30u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d31u8));
+    d2 += dest_stride;
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/idct8x8_add_neon.asm b/third_party/aom/aom_dsp/arm/idct8x8_add_neon.asm
new file mode 100644
index 0000000000..f3d5f246d0
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct8x8_add_neon.asm
@@ -0,0 +1,522 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_idct8x8_64_add_neon|
+    EXPORT  |aom_idct8x8_12_add_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+    ; Parallel 1D IDCT on all the columns of a 8x8 16bit data matrix which are
+    ; loaded in q8-q15. The output will be stored back into q8-q15 registers.
+    ; This macro will touch q0-q7 registers and use them as buffer during
+    ; calculation.
+    MACRO
+    IDCT8x8_1D
+    ; stage 1
+    vdup.16         d0, r3                    ; duplicate cospi_28_64
+    vdup.16         d1, r4                    ; duplicate cospi_4_64
+    vdup.16         d2, r5                    ; duplicate cospi_12_64
+    vdup.16         d3, r6                    ; duplicate cospi_20_64
+
+    ; input[1] * cospi_28_64
+    vmull.s16       q2, d18, d0
+    vmull.s16       q3, d19, d0
+
+    ; input[5] * cospi_12_64
+    vmull.s16       q5, d26, d2
+    vmull.s16       q6, d27, d2
+
+    ; input[1]*cospi_28_64-input[7]*cospi_4_64
+    vmlsl.s16       q2, d30, d1
+    vmlsl.s16       q3, d31, d1
+
+    ; input[5] * cospi_12_64 - input[3] * cospi_20_64
+    vmlsl.s16       q5, d22, d3
+    vmlsl.s16       q6, d23, d3
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d8, q2, #14               ; >> 14
+    vqrshrn.s32     d9, q3, #14               ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q5, #14              ; >> 14
+    vqrshrn.s32     d11, q6, #14              ; >> 14
+
+    ; input[1] * cospi_4_64
+    vmull.s16       q2, d18, d1
+    vmull.s16       q3, d19, d1
+
+    ; input[5] * cospi_20_64
+    vmull.s16       q9, d26, d3
+    vmull.s16       q13, d27, d3
+
+    ; input[1]*cospi_4_64+input[7]*cospi_28_64
+    vmlal.s16       q2, d30, d0
+    vmlal.s16       q3, d31, d0
+
+    ; input[5] * cospi_20_64 + input[3] * cospi_12_64
+    vmlal.s16       q9, d22, d2
+    vmlal.s16       q13, d23, d2
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d14, q2, #14              ; >> 14
+    vqrshrn.s32     d15, q3, #14              ; >> 14
+
+    ; stage 2 & stage 3 - even half
+    vdup.16         d0, r7                    ; duplicate cospi_16_64
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q9, #14              ; >> 14
+    vqrshrn.s32     d13, q13, #14              ; >> 14
+
+    ; input[0] * cospi_16_64
+    vmull.s16       q2, d16, d0
+    vmull.s16       q3, d17, d0
+
+    ; input[0] * cospi_16_64
+    vmull.s16       q13, d16, d0
+    vmull.s16       q15, d17, d0
+
+    ; (input[0] + input[2]) * cospi_16_64
+    vmlal.s16       q2,  d24, d0
+    vmlal.s16       q3, d25, d0
+
+    ; (input[0] - input[2]) * cospi_16_64
+    vmlsl.s16       q13, d24, d0
+    vmlsl.s16       q15, d25, d0
+
+    vdup.16         d0, r8                    ; duplicate cospi_24_64
+    vdup.16         d1, r9                    ; duplicate cospi_8_64
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d18, q2, #14              ; >> 14
+    vqrshrn.s32     d19, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d22, q13, #14              ; >> 14
+    vqrshrn.s32     d23, q15, #14              ; >> 14
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64
+    ; input[1] * cospi_24_64
+    vmull.s16       q2, d20, d0
+    vmull.s16       q3, d21, d0
+
+    ; input[1] * cospi_8_64
+    vmull.s16       q8, d20, d1
+    vmull.s16       q12, d21, d1
+
+    ; input[1] * cospi_24_64 - input[3] * cospi_8_64
+    vmlsl.s16       q2, d28, d1
+    vmlsl.s16       q3, d29, d1
+
+    ; input[1] * cospi_8_64 + input[3] * cospi_24_64
+    vmlal.s16       q8, d28, d0
+    vmlal.s16       q12, d29, d0
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d26, q2, #14              ; >> 14
+    vqrshrn.s32     d27, q3, #14              ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d30, q8, #14              ; >> 14
+    vqrshrn.s32     d31, q12, #14              ; >> 14
+
+    vadd.s16        q0, q9, q15               ; output[0] = step[0] + step[3]
+    vadd.s16        q1, q11, q13              ; output[1] = step[1] + step[2]
+    vsub.s16        q2, q11, q13              ; output[2] = step[1] - step[2]
+    vsub.s16        q3, q9, q15               ; output[3] = step[0] - step[3]
+
+    ; stage 3 -odd half
+    vdup.16         d16, r7                   ; duplicate cospi_16_64
+
+    ; stage 2 - odd half
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5]
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5]
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7]
+    vadd.s16        q7, q7, q6                ; step2[7] = step1[6] + step1[7]
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q11, d28, d16
+    vmull.s16       q12, d29, d16
+
+    ; (step2[6] - step2[5]) * cospi_16_64
+    vmlsl.s16       q9, d26, d16
+    vmlsl.s16       q10, d27, d16
+
+    ; (step2[5] + step2[6]) * cospi_16_64
+    vmlal.s16       q11, d26, d16
+    vmlal.s16       q12, d27, d16
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q9, #14              ; >> 14
+    vqrshrn.s32     d11, q10, #14             ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q11, #14              ; >> 14
+    vqrshrn.s32     d13, q12, #14             ; >> 14
+
+    ; stage 4
+    vadd.s16        q8, q0, q7                ; output[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; output[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; output[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; output[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; output[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; output[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; output[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q7               ; output[7] = step1[0] - step1[7];
+    MEND
+
+    ; Transpose a 8x8 16bit data matrix. Datas are loaded in q8-q15.
+    MACRO
+    TRANSPOSE8X8
+    vswp            d17, d24
+    vswp            d23, d30
+    vswp            d21, d28
+    vswp            d19, d26
+    vtrn.32         q8, q10
+    vtrn.32         q9, q11
+    vtrn.32         q12, q14
+    vtrn.32         q13, q15
+    vtrn.16         q8, q9
+    vtrn.16         q10, q11
+    vtrn.16         q12, q13
+    vtrn.16         q14, q15
+    MEND
+
+    AREA    Block, CODE, READONLY ; name this block of code
+;void aom_idct8x8_64_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct8x8_64_add_neon| PROC
+    push            {r4-r9}
+    vpush           {d8-d15}
+    vld1.s16        {q8,q9}, [r0]!
+    vld1.s16        {q10,q11}, [r0]!
+    vld1.s16        {q12,q13}, [r0]!
+    vld1.s16        {q14,q15}, [r0]!
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0x0c00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r4, #0x3e00
+    add             r4, #0xc5
+
+    ; generate cospi_12_64 = 13623
+    mov             r5, #0x3500
+    add             r5, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r6, #0x2300
+    add             r6, #0x8e
+
+    ; generate cospi_16_64 = 11585
+    mov             r7, #0x2d00
+    add             r7, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r8, #0x1800
+    add             r8, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r9, #0x3b00
+    add             r9, #0x21
+
+    ; First transform rows
+    IDCT8x8_1D
+
+    ; Transpose the matrix
+    TRANSPOSE8X8
+
+    ; Then transform columns
+    IDCT8x8_1D
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5)
+    vrshr.s16       q8, q8, #5
+    vrshr.s16       q9, q9, #5
+    vrshr.s16       q10, q10, #5
+    vrshr.s16       q11, q11, #5
+    vrshr.s16       q12, q12, #5
+    vrshr.s16       q13, q13, #5
+    vrshr.s16       q14, q14, #5
+    vrshr.s16       q15, q15, #5
+
+    ; save dest pointer
+    mov             r0, r1
+
+    ; load destination data
+    vld1.64         {d0}, [r1], r2
+    vld1.64         {d1}, [r1], r2
+    vld1.64         {d2}, [r1], r2
+    vld1.64         {d3}, [r1], r2
+    vld1.64         {d4}, [r1], r2
+    vld1.64         {d5}, [r1], r2
+    vld1.64         {d6}, [r1], r2
+    vld1.64         {d7}, [r1]
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i]
+    vaddw.u8        q8, q8, d0
+    vaddw.u8        q9, q9, d1
+    vaddw.u8        q10, q10, d2
+    vaddw.u8        q11, q11, d3
+    vaddw.u8        q12, q12, d4
+    vaddw.u8        q13, q13, d5
+    vaddw.u8        q14, q14, d6
+    vaddw.u8        q15, q15, d7
+
+    ; clip_pixel
+    vqmovun.s16     d0, q8
+    vqmovun.s16     d1, q9
+    vqmovun.s16     d2, q10
+    vqmovun.s16     d3, q11
+    vqmovun.s16     d4, q12
+    vqmovun.s16     d5, q13
+    vqmovun.s16     d6, q14
+    vqmovun.s16     d7, q15
+
+    ; store the data
+    vst1.64         {d0}, [r0], r2
+    vst1.64         {d1}, [r0], r2
+    vst1.64         {d2}, [r0], r2
+    vst1.64         {d3}, [r0], r2
+    vst1.64         {d4}, [r0], r2
+    vst1.64         {d5}, [r0], r2
+    vst1.64         {d6}, [r0], r2
+    vst1.64         {d7}, [r0], r2
+
+    vpop            {d8-d15}
+    pop             {r4-r9}
+    bx              lr
+    ENDP  ; |aom_idct8x8_64_add_neon|
+
+;void aom_idct8x8_12_add_neon(int16_t *input, uint8_t *dest, int dest_stride)
+;
+; r0  int16_t input
+; r1  uint8_t *dest
+; r2  int dest_stride)
+
+|aom_idct8x8_12_add_neon| PROC
+    push            {r4-r9}
+    vpush           {d8-d15}
+    vld1.s16        {q8,q9}, [r0]!
+    vld1.s16        {q10,q11}, [r0]!
+    vld1.s16        {q12,q13}, [r0]!
+    vld1.s16        {q14,q15}, [r0]!
+
+    ; transpose the input data
+    TRANSPOSE8X8
+
+    ; generate  cospi_28_64 = 3196
+    mov             r3, #0x0c00
+    add             r3, #0x7c
+
+    ; generate cospi_4_64  = 16069
+    mov             r4, #0x3e00
+    add             r4, #0xc5
+
+    ; generate cospi_12_64 = 13623
+    mov             r5, #0x3500
+    add             r5, #0x37
+
+    ; generate cospi_20_64 = 9102
+    mov             r6, #0x2300
+    add             r6, #0x8e
+
+    ; generate cospi_16_64 = 11585
+    mov             r7, #0x2d00
+    add             r7, #0x41
+
+    ; generate cospi_24_64 = 6270
+    mov             r8, #0x1800
+    add             r8, #0x7e
+
+    ; generate cospi_8_64 = 15137
+    mov             r9, #0x3b00
+    add             r9, #0x21
+
+    ; First transform rows
+    ; stage 1
+    ; The following instructions use vqrdmulh to do the
+    ; dct_const_round_shift(input[1] * cospi_28_64). vqrdmulh will do doubling
+    ; multiply and shift the result by 16 bits instead of 14 bits. So we need
+    ; to double the constants before multiplying to compensate this.
+    mov             r12, r3, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_28_64*2
+    mov             r12, r4, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_4_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_28_64)
+    vqrdmulh.s16    q4, q9, q0
+
+    mov             r12, r6, lsl #1
+    rsb             r12, #0
+    vdup.16         q0, r12                   ; duplicate -cospi_20_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_4_64)
+    vqrdmulh.s16    q7, q9, q1
+
+    mov             r12, r5, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_12_64*2
+
+    ; dct_const_round_shift(- input[3] * cospi_20_64)
+    vqrdmulh.s16    q5, q11, q0
+
+    mov             r12, r7, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_16_64*2
+
+    ; dct_const_round_shift(input[3] * cospi_12_64)
+    vqrdmulh.s16    q6, q11, q1
+
+    ; stage 2 & stage 3 - even half
+    mov             r12, r8, lsl #1
+    vdup.16         q1, r12                   ; duplicate cospi_24_64*2
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrdmulh.s16    q9, q8, q0
+
+    mov             r12, r9, lsl #1
+    vdup.16         q0, r12                   ; duplicate cospi_8_64*2
+
+    ; dct_const_round_shift(input[1] * cospi_24_64)
+    vqrdmulh.s16    q13, q10, q1
+
+    ; dct_const_round_shift(input[1] * cospi_8_64)
+    vqrdmulh.s16    q15, q10, q0
+
+    ; stage 3 -odd half
+    vdup.16         d16, r7                   ; duplicate cospi_16_64
+
+    vadd.s16        q0, q9, q15               ; output[0] = step[0] + step[3]
+    vadd.s16        q1, q9, q13               ; output[1] = step[1] + step[2]
+    vsub.s16        q2, q9, q13               ; output[2] = step[1] - step[2]
+    vsub.s16        q3, q9, q15               ; output[3] = step[0] - step[3]
+
+    ; stage 2 - odd half
+    vsub.s16        q13, q4, q5               ; step2[5] = step1[4] - step1[5]
+    vadd.s16        q4, q4, q5                ; step2[4] = step1[4] + step1[5]
+    vsub.s16        q14, q7, q6               ; step2[6] = -step1[6] + step1[7]
+    vadd.s16        q7, q7, q6                ; step2[7] = step1[6] + step1[7]
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q9, d28, d16
+    vmull.s16       q10, d29, d16
+
+    ; step2[6] * cospi_16_64
+    vmull.s16       q11, d28, d16
+    vmull.s16       q12, d29, d16
+
+    ; (step2[6] - step2[5]) * cospi_16_64
+    vmlsl.s16       q9, d26, d16
+    vmlsl.s16       q10, d27, d16
+
+    ; (step2[5] + step2[6]) * cospi_16_64
+    vmlal.s16       q11, d26, d16
+    vmlal.s16       q12, d27, d16
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d10, q9, #14              ; >> 14
+    vqrshrn.s32     d11, q10, #14             ; >> 14
+
+    ; dct_const_round_shift(input_dc * cospi_16_64)
+    vqrshrn.s32     d12, q11, #14              ; >> 14
+    vqrshrn.s32     d13, q12, #14             ; >> 14
+
+    ; stage 4
+    vadd.s16        q8, q0, q7                ; output[0] = step1[0] + step1[7];
+    vadd.s16        q9, q1, q6                ; output[1] = step1[1] + step1[6];
+    vadd.s16        q10, q2, q5               ; output[2] = step1[2] + step1[5];
+    vadd.s16        q11, q3, q4               ; output[3] = step1[3] + step1[4];
+    vsub.s16        q12, q3, q4               ; output[4] = step1[3] - step1[4];
+    vsub.s16        q13, q2, q5               ; output[5] = step1[2] - step1[5];
+    vsub.s16        q14, q1, q6               ; output[6] = step1[1] - step1[6];
+    vsub.s16        q15, q0, q7               ; output[7] = step1[0] - step1[7];
+
+    ; Transpose the matrix
+    TRANSPOSE8X8
+
+    ; Then transform columns
+    IDCT8x8_1D
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5)
+    vrshr.s16       q8, q8, #5
+    vrshr.s16       q9, q9, #5
+    vrshr.s16       q10, q10, #5
+    vrshr.s16       q11, q11, #5
+    vrshr.s16       q12, q12, #5
+    vrshr.s16       q13, q13, #5
+    vrshr.s16       q14, q14, #5
+    vrshr.s16       q15, q15, #5
+
+    ; save dest pointer
+    mov             r0, r1
+
+    ; load destination data
+    vld1.64         {d0}, [r1], r2
+    vld1.64         {d1}, [r1], r2
+    vld1.64         {d2}, [r1], r2
+    vld1.64         {d3}, [r1], r2
+    vld1.64         {d4}, [r1], r2
+    vld1.64         {d5}, [r1], r2
+    vld1.64         {d6}, [r1], r2
+    vld1.64         {d7}, [r1]
+
+    ; ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i]
+    vaddw.u8        q8, q8, d0
+    vaddw.u8        q9, q9, d1
+    vaddw.u8        q10, q10, d2
+    vaddw.u8        q11, q11, d3
+    vaddw.u8        q12, q12, d4
+    vaddw.u8        q13, q13, d5
+    vaddw.u8        q14, q14, d6
+    vaddw.u8        q15, q15, d7
+
+    ; clip_pixel
+    vqmovun.s16     d0, q8
+    vqmovun.s16     d1, q9
+    vqmovun.s16     d2, q10
+    vqmovun.s16     d3, q11
+    vqmovun.s16     d4, q12
+    vqmovun.s16     d5, q13
+    vqmovun.s16     d6, q14
+    vqmovun.s16     d7, q15
+
+    ; store the data
+    vst1.64         {d0}, [r0], r2
+    vst1.64         {d1}, [r0], r2
+    vst1.64         {d2}, [r0], r2
+    vst1.64         {d3}, [r0], r2
+    vst1.64         {d4}, [r0], r2
+    vst1.64         {d5}, [r0], r2
+    vst1.64         {d6}, [r0], r2
+    vst1.64         {d7}, [r0], r2
+
+    vpop            {d8-d15}
+    pop             {r4-r9}
+    bx              lr
+    ENDP  ; |aom_idct8x8_12_add_neon|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/idct8x8_add_neon.c b/third_party/aom/aom_dsp/arm/idct8x8_add_neon.c
new file mode 100644
index 0000000000..8ad70862d6
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/idct8x8_add_neon.c
@@ -0,0 +1,509 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "aom_dsp/txfm_common.h"
+
+static INLINE void TRANSPOSE8X8(int16x8_t *q8s16, int16x8_t *q9s16,
+                                int16x8_t *q10s16, int16x8_t *q11s16,
+                                int16x8_t *q12s16, int16x8_t *q13s16,
+                                int16x8_t *q14s16, int16x8_t *q15s16) {
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+  int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+  d16s16 = vget_low_s16(*q8s16);
+  d17s16 = vget_high_s16(*q8s16);
+  d18s16 = vget_low_s16(*q9s16);
+  d19s16 = vget_high_s16(*q9s16);
+  d20s16 = vget_low_s16(*q10s16);
+  d21s16 = vget_high_s16(*q10s16);
+  d22s16 = vget_low_s16(*q11s16);
+  d23s16 = vget_high_s16(*q11s16);
+  d24s16 = vget_low_s16(*q12s16);
+  d25s16 = vget_high_s16(*q12s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+  d30s16 = vget_low_s16(*q15s16);
+  d31s16 = vget_high_s16(*q15s16);
+
+  *q8s16 = vcombine_s16(d16s16, d24s16);   // vswp d17, d24
+  *q9s16 = vcombine_s16(d18s16, d26s16);   // vswp d19, d26
+  *q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+  *q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+  *q12s16 = vcombine_s16(d17s16, d25s16);
+  *q13s16 = vcombine_s16(d19s16, d27s16);
+  *q14s16 = vcombine_s16(d21s16, d29s16);
+  *q15s16 = vcombine_s16(d23s16, d31s16);
+
+  q0x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q8s16), vreinterpretq_s32_s16(*q10s16));
+  q1x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q9s16), vreinterpretq_s32_s16(*q11s16));
+  q2x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q12s16), vreinterpretq_s32_s16(*q14s16));
+  q3x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q13s16), vreinterpretq_s32_s16(*q15s16));
+
+  q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                      vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+  q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                      vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+  q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                      vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+  q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                      vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+  *q8s16 = q0x2s16.val[0];
+  *q9s16 = q0x2s16.val[1];
+  *q10s16 = q1x2s16.val[0];
+  *q11s16 = q1x2s16.val[1];
+  *q12s16 = q2x2s16.val[0];
+  *q13s16 = q2x2s16.val[1];
+  *q14s16 = q3x2s16.val[0];
+  *q15s16 = q3x2s16.val[1];
+  return;
+}
+
+static INLINE void IDCT8x8_1D(int16x8_t *q8s16, int16x8_t *q9s16,
+                              int16x8_t *q10s16, int16x8_t *q11s16,
+                              int16x8_t *q12s16, int16x8_t *q13s16,
+                              int16x8_t *q14s16, int16x8_t *q15s16) {
+  int16x4_t d0s16, d1s16, d2s16, d3s16;
+  int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+  int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
+  int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
+
+  d0s16 = vdup_n_s16((int16_t)cospi_28_64);
+  d1s16 = vdup_n_s16((int16_t)cospi_4_64);
+  d2s16 = vdup_n_s16((int16_t)cospi_12_64);
+  d3s16 = vdup_n_s16((int16_t)cospi_20_64);
+
+  d16s16 = vget_low_s16(*q8s16);
+  d17s16 = vget_high_s16(*q8s16);
+  d18s16 = vget_low_s16(*q9s16);
+  d19s16 = vget_high_s16(*q9s16);
+  d20s16 = vget_low_s16(*q10s16);
+  d21s16 = vget_high_s16(*q10s16);
+  d22s16 = vget_low_s16(*q11s16);
+  d23s16 = vget_high_s16(*q11s16);
+  d24s16 = vget_low_s16(*q12s16);
+  d25s16 = vget_high_s16(*q12s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+  d30s16 = vget_low_s16(*q15s16);
+  d31s16 = vget_high_s16(*q15s16);
+
+  q2s32 = vmull_s16(d18s16, d0s16);
+  q3s32 = vmull_s16(d19s16, d0s16);
+  q5s32 = vmull_s16(d26s16, d2s16);
+  q6s32 = vmull_s16(d27s16, d2s16);
+
+  q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
+  q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
+  q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
+  q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
+
+  d8s16 = vqrshrn_n_s32(q2s32, 14);
+  d9s16 = vqrshrn_n_s32(q3s32, 14);
+  d10s16 = vqrshrn_n_s32(q5s32, 14);
+  d11s16 = vqrshrn_n_s32(q6s32, 14);
+  q4s16 = vcombine_s16(d8s16, d9s16);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+
+  q2s32 = vmull_s16(d18s16, d1s16);
+  q3s32 = vmull_s16(d19s16, d1s16);
+  q9s32 = vmull_s16(d26s16, d3s16);
+  q13s32 = vmull_s16(d27s16, d3s16);
+
+  q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
+  q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
+  q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
+  q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
+
+  d14s16 = vqrshrn_n_s32(q2s32, 14);
+  d15s16 = vqrshrn_n_s32(q3s32, 14);
+  d12s16 = vqrshrn_n_s32(q9s32, 14);
+  d13s16 = vqrshrn_n_s32(q13s32, 14);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+  q7s16 = vcombine_s16(d14s16, d15s16);
+
+  d0s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q2s32 = vmull_s16(d16s16, d0s16);
+  q3s32 = vmull_s16(d17s16, d0s16);
+  q13s32 = vmull_s16(d16s16, d0s16);
+  q15s32 = vmull_s16(d17s16, d0s16);
+
+  q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
+  q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
+  q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
+  q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
+
+  d0s16 = vdup_n_s16((int16_t)cospi_24_64);
+  d1s16 = vdup_n_s16((int16_t)cospi_8_64);
+
+  d18s16 = vqrshrn_n_s32(q2s32, 14);
+  d19s16 = vqrshrn_n_s32(q3s32, 14);
+  d22s16 = vqrshrn_n_s32(q13s32, 14);
+  d23s16 = vqrshrn_n_s32(q15s32, 14);
+  *q9s16 = vcombine_s16(d18s16, d19s16);
+  *q11s16 = vcombine_s16(d22s16, d23s16);
+
+  q2s32 = vmull_s16(d20s16, d0s16);
+  q3s32 = vmull_s16(d21s16, d0s16);
+  q8s32 = vmull_s16(d20s16, d1s16);
+  q12s32 = vmull_s16(d21s16, d1s16);
+
+  q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
+  q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
+  q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
+  q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
+
+  d26s16 = vqrshrn_n_s32(q2s32, 14);
+  d27s16 = vqrshrn_n_s32(q3s32, 14);
+  d30s16 = vqrshrn_n_s32(q8s32, 14);
+  d31s16 = vqrshrn_n_s32(q12s32, 14);
+  *q13s16 = vcombine_s16(d26s16, d27s16);
+  *q15s16 = vcombine_s16(d30s16, d31s16);
+
+  q0s16 = vaddq_s16(*q9s16, *q15s16);
+  q1s16 = vaddq_s16(*q11s16, *q13s16);
+  q2s16 = vsubq_s16(*q11s16, *q13s16);
+  q3s16 = vsubq_s16(*q9s16, *q15s16);
+
+  *q13s16 = vsubq_s16(q4s16, q5s16);
+  q4s16 = vaddq_s16(q4s16, q5s16);
+  *q14s16 = vsubq_s16(q7s16, q6s16);
+  q7s16 = vaddq_s16(q7s16, q6s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+
+  d16s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q9s32 = vmull_s16(d28s16, d16s16);
+  q10s32 = vmull_s16(d29s16, d16s16);
+  q11s32 = vmull_s16(d28s16, d16s16);
+  q12s32 = vmull_s16(d29s16, d16s16);
+
+  q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
+  q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+  q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+  q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+  d10s16 = vqrshrn_n_s32(q9s32, 14);
+  d11s16 = vqrshrn_n_s32(q10s32, 14);
+  d12s16 = vqrshrn_n_s32(q11s32, 14);
+  d13s16 = vqrshrn_n_s32(q12s32, 14);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  *q8s16 = vaddq_s16(q0s16, q7s16);
+  *q9s16 = vaddq_s16(q1s16, q6s16);
+  *q10s16 = vaddq_s16(q2s16, q5s16);
+  *q11s16 = vaddq_s16(q3s16, q4s16);
+  *q12s16 = vsubq_s16(q3s16, q4s16);
+  *q13s16 = vsubq_s16(q2s16, q5s16);
+  *q14s16 = vsubq_s16(q1s16, q6s16);
+  *q15s16 = vsubq_s16(q0s16, q7s16);
+  return;
+}
+
+void aom_idct8x8_64_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8_t *d1, *d2;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8;
+  uint64x1_t d0u64, d1u64, d2u64, d3u64;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  uint16x8_t q8u16, q9u16, q10u16, q11u16;
+
+  q8s16 = vld1q_s16(input);
+  q9s16 = vld1q_s16(input + 8);
+  q10s16 = vld1q_s16(input + 16);
+  q11s16 = vld1q_s16(input + 24);
+  q12s16 = vld1q_s16(input + 32);
+  q13s16 = vld1q_s16(input + 40);
+  q14s16 = vld1q_s16(input + 48);
+  q15s16 = vld1q_s16(input + 56);
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+             &q15s16);
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+             &q15s16);
+
+  q8s16 = vrshrq_n_s16(q8s16, 5);
+  q9s16 = vrshrq_n_s16(q9s16, 5);
+  q10s16 = vrshrq_n_s16(q10s16, 5);
+  q11s16 = vrshrq_n_s16(q11s16, 5);
+  q12s16 = vrshrq_n_s16(q12s16, 5);
+  q13s16 = vrshrq_n_s16(q13s16, 5);
+  q14s16 = vrshrq_n_s16(q14s16, 5);
+  q15s16 = vrshrq_n_s16(q15s16, 5);
+
+  d1 = d2 = dest;
+
+  d0u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d1u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d2u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d3u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
+  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
+  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
+
+  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+  d2 += dest_stride;
+
+  q8s16 = q12s16;
+  q9s16 = q13s16;
+  q10s16 = q14s16;
+  q11s16 = q15s16;
+
+  d0u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d1u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d2u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d3u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
+  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
+  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
+
+  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+  d2 += dest_stride;
+  return;
+}
+
+void aom_idct8x8_12_add_neon(int16_t *input, uint8_t *dest, int dest_stride) {
+  uint8_t *d1, *d2;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8;
+  int16x4_t d10s16, d11s16, d12s16, d13s16, d16s16;
+  int16x4_t d26s16, d27s16, d28s16, d29s16;
+  uint64x1_t d0u64, d1u64, d2u64, d3u64;
+  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  uint16x8_t q8u16, q9u16, q10u16, q11u16;
+  int32x4_t q9s32, q10s32, q11s32, q12s32;
+
+  q8s16 = vld1q_s16(input);
+  q9s16 = vld1q_s16(input + 8);
+  q10s16 = vld1q_s16(input + 16);
+  q11s16 = vld1q_s16(input + 24);
+  q12s16 = vld1q_s16(input + 32);
+  q13s16 = vld1q_s16(input + 40);
+  q14s16 = vld1q_s16(input + 48);
+  q15s16 = vld1q_s16(input + 56);
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  // First transform rows
+  // stage 1
+  q0s16 = vdupq_n_s16((int16_t)cospi_28_64 * 2);
+  q1s16 = vdupq_n_s16((int16_t)cospi_4_64 * 2);
+
+  q4s16 = vqrdmulhq_s16(q9s16, q0s16);
+
+  q0s16 = vdupq_n_s16(-(int16_t)cospi_20_64 * 2);
+
+  q7s16 = vqrdmulhq_s16(q9s16, q1s16);
+
+  q1s16 = vdupq_n_s16((int16_t)cospi_12_64 * 2);
+
+  q5s16 = vqrdmulhq_s16(q11s16, q0s16);
+
+  q0s16 = vdupq_n_s16((int16_t)cospi_16_64 * 2);
+
+  q6s16 = vqrdmulhq_s16(q11s16, q1s16);
+
+  // stage 2 & stage 3 - even half
+  q1s16 = vdupq_n_s16((int16_t)cospi_24_64 * 2);
+
+  q9s16 = vqrdmulhq_s16(q8s16, q0s16);
+
+  q0s16 = vdupq_n_s16((int16_t)cospi_8_64 * 2);
+
+  q13s16 = vqrdmulhq_s16(q10s16, q1s16);
+
+  q15s16 = vqrdmulhq_s16(q10s16, q0s16);
+
+  // stage 3 -odd half
+  q0s16 = vaddq_s16(q9s16, q15s16);
+  q1s16 = vaddq_s16(q9s16, q13s16);
+  q2s16 = vsubq_s16(q9s16, q13s16);
+  q3s16 = vsubq_s16(q9s16, q15s16);
+
+  // stage 2 - odd half
+  q13s16 = vsubq_s16(q4s16, q5s16);
+  q4s16 = vaddq_s16(q4s16, q5s16);
+  q14s16 = vsubq_s16(q7s16, q6s16);
+  q7s16 = vaddq_s16(q7s16, q6s16);
+  d26s16 = vget_low_s16(q13s16);
+  d27s16 = vget_high_s16(q13s16);
+  d28s16 = vget_low_s16(q14s16);
+  d29s16 = vget_high_s16(q14s16);
+
+  d16s16 = vdup_n_s16((int16_t)cospi_16_64);
+  q9s32 = vmull_s16(d28s16, d16s16);
+  q10s32 = vmull_s16(d29s16, d16s16);
+  q11s32 = vmull_s16(d28s16, d16s16);
+  q12s32 = vmull_s16(d29s16, d16s16);
+
+  q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
+  q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+  q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+  q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+  d10s16 = vqrshrn_n_s32(q9s32, 14);
+  d11s16 = vqrshrn_n_s32(q10s32, 14);
+  d12s16 = vqrshrn_n_s32(q11s32, 14);
+  d13s16 = vqrshrn_n_s32(q12s32, 14);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  // stage 4
+  q8s16 = vaddq_s16(q0s16, q7s16);
+  q9s16 = vaddq_s16(q1s16, q6s16);
+  q10s16 = vaddq_s16(q2s16, q5s16);
+  q11s16 = vaddq_s16(q3s16, q4s16);
+  q12s16 = vsubq_s16(q3s16, q4s16);
+  q13s16 = vsubq_s16(q2s16, q5s16);
+  q14s16 = vsubq_s16(q1s16, q6s16);
+  q15s16 = vsubq_s16(q0s16, q7s16);
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+             &q15s16);
+
+  q8s16 = vrshrq_n_s16(q8s16, 5);
+  q9s16 = vrshrq_n_s16(q9s16, 5);
+  q10s16 = vrshrq_n_s16(q10s16, 5);
+  q11s16 = vrshrq_n_s16(q11s16, 5);
+  q12s16 = vrshrq_n_s16(q12s16, 5);
+  q13s16 = vrshrq_n_s16(q13s16, 5);
+  q14s16 = vrshrq_n_s16(q14s16, 5);
+  q15s16 = vrshrq_n_s16(q15s16, 5);
+
+  d1 = d2 = dest;
+
+  d0u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d1u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d2u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d3u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
+  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
+  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
+
+  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+  d2 += dest_stride;
+
+  q8s16 = q12s16;
+  q9s16 = q13s16;
+  q10s16 = q14s16;
+  q11s16 = q15s16;
+
+  d0u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d1u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d2u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+  d3u64 = vld1_u64((uint64_t *)d1);
+  d1 += dest_stride;
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
+  q10u16 = vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
+  q11u16 = vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
+
+  d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+  d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+  d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+  d2 += dest_stride;
+  vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+  d2 += dest_stride;
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/intrapred_neon.c b/third_party/aom/aom_dsp/arm/intrapred_neon.c
new file mode 100644
index 0000000000..2dc5b2e568
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/intrapred_neon.c
@@ -0,0 +1,757 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+//------------------------------------------------------------------------------
+// DC 4x4
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_4x4(uint8_t *dst, ptrdiff_t stride, const uint8_t *above,
+                          const uint8_t *left, int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x8_t A = vld1_u8(above);  // top row
+    const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    sum_top = vcombine_u16(p1, p1);
+  }
+
+  if (do_left) {
+    const uint8x8_t L = vld1_u8(left);   // left border
+    const uint16x4_t p0 = vpaddl_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    sum_left = vcombine_u16(p1, p1);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 3);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 2);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 2);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x8_t dc = vdup_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 4; ++i) {
+      vst1_lane_u32((uint32_t *)(dst + i * stride), vreinterpret_u32_u8(dc), 0);
+    }
+  }
+}
+
+void aom_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  dc_4x4(dst, stride, above, left, 1, 1);
+}
+
+void aom_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  dc_4x4(dst, stride, NULL, left, 0, 1);
+}
+
+void aom_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  dc_4x4(dst, stride, above, NULL, 1, 0);
+}
+
+void aom_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_4x4(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+// DC 8x8
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_8x8(uint8_t *dst, ptrdiff_t stride, const uint8_t *above,
+                          const uint8_t *left, int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x8_t A = vld1_u8(above);  // top row
+    const uint16x4_t p0 = vpaddl_u8(A);  // cascading summation of the top
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    sum_top = vcombine_u16(p2, p2);
+  }
+
+  if (do_left) {
+    const uint8x8_t L = vld1_u8(left);   // left border
+    const uint16x4_t p0 = vpaddl_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vpadd_u16(p0, p0);
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    sum_left = vcombine_u16(p2, p2);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 4);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 3);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 3);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x8_t dc = vdup_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 8; ++i) {
+      vst1_u32((uint32_t *)(dst + i * stride), vreinterpret_u32_u8(dc));
+    }
+  }
+}
+
+void aom_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  dc_8x8(dst, stride, above, left, 1, 1);
+}
+
+void aom_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  dc_8x8(dst, stride, NULL, left, 0, 1);
+}
+
+void aom_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  dc_8x8(dst, stride, above, NULL, 1, 0);
+}
+
+void aom_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_8x8(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+// DC 16x16
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_16x16(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left,
+                            int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x16_t A = vld1q_u8(above);  // top row
+    const uint16x8_t p0 = vpaddlq_u8(A);   // cascading summation of the top
+    const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    const uint16x4_t p3 = vpadd_u16(p2, p2);
+    sum_top = vcombine_u16(p3, p3);
+  }
+
+  if (do_left) {
+    const uint8x16_t L = vld1q_u8(left);  // left row
+    const uint16x8_t p0 = vpaddlq_u8(L);  // cascading summation of the left
+    const uint16x4_t p1 = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
+    const uint16x4_t p2 = vpadd_u16(p1, p1);
+    const uint16x4_t p3 = vpadd_u16(p2, p2);
+    sum_left = vcombine_u16(p3, p3);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 5);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 4);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 4);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 16; ++i) {
+      vst1q_u8(dst + i * stride, dc);
+    }
+  }
+}
+
+void aom_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  dc_16x16(dst, stride, above, left, 1, 1);
+}
+
+void aom_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  dc_16x16(dst, stride, NULL, left, 0, 1);
+}
+
+void aom_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  dc_16x16(dst, stride, above, NULL, 1, 0);
+}
+
+void aom_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_16x16(dst, stride, NULL, NULL, 0, 0);
+}
+
+//------------------------------------------------------------------------------
+// DC 32x32
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_32x32(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left,
+                            int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x16_t A0 = vld1q_u8(above);  // top row
+    const uint8x16_t A1 = vld1q_u8(above + 16);
+    const uint16x8_t p0 = vpaddlq_u8(A0);  // cascading summation of the top
+    const uint16x8_t p1 = vpaddlq_u8(A1);
+    const uint16x8_t p2 = vaddq_u16(p0, p1);
+    const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
+    const uint16x4_t p4 = vpadd_u16(p3, p3);
+    const uint16x4_t p5 = vpadd_u16(p4, p4);
+    sum_top = vcombine_u16(p5, p5);
+  }
+
+  if (do_left) {
+    const uint8x16_t L0 = vld1q_u8(left);  // left row
+    const uint8x16_t L1 = vld1q_u8(left + 16);
+    const uint16x8_t p0 = vpaddlq_u8(L0);  // cascading summation of the left
+    const uint16x8_t p1 = vpaddlq_u8(L1);
+    const uint16x8_t p2 = vaddq_u16(p0, p1);
+    const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
+    const uint16x4_t p4 = vpadd_u16(p3, p3);
+    const uint16x4_t p5 = vpadd_u16(p4, p4);
+    sum_left = vcombine_u16(p5, p5);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 6);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 5);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 5);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 32; ++i) {
+      vst1q_u8(dst + i * stride, dc);
+      vst1q_u8(dst + i * stride + 16, dc);
+    }
+  }
+}
+
+void aom_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  dc_32x32(dst, stride, above, left, 1, 1);
+}
+
+void aom_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  dc_32x32(dst, stride, NULL, left, 0, 1);
+}
+
+void aom_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  dc_32x32(dst, stride, above, NULL, 1, 0);
+}
+
+void aom_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_32x32(dst, stride, NULL, NULL, 0, 0);
+}
+
+// -----------------------------------------------------------------------------
+
+void aom_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const uint8x8_t XABCD_u8 = vld1_u8(above - 1);
+  const uint64x1_t XABCD = vreinterpret_u64_u8(XABCD_u8);
+  const uint64x1_t ____XABC = vshl_n_u64(XABCD, 32);
+  const uint32x2_t zero = vdup_n_u32(0);
+  const uint32x2_t IJKL = vld1_lane_u32((const uint32_t *)left, zero, 0);
+  const uint8x8_t IJKL_u8 = vreinterpret_u8_u32(IJKL);
+  const uint64x1_t LKJI____ = vreinterpret_u64_u8(vrev32_u8(IJKL_u8));
+  const uint64x1_t LKJIXABC = vorr_u64(LKJI____, ____XABC);
+  const uint8x8_t KJIXABC_ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 8));
+  const uint8x8_t JIXABC__ = vreinterpret_u8_u64(vshr_n_u64(LKJIXABC, 16));
+  const uint8_t D = vget_lane_u8(XABCD_u8, 4);
+  const uint8x8_t JIXABCD_ = vset_lane_u8(D, JIXABC__, 6);
+  const uint8x8_t LKJIXABC_u8 = vreinterpret_u8_u64(LKJIXABC);
+  const uint8x8_t avg1 = vhadd_u8(JIXABCD_, LKJIXABC_u8);
+  const uint8x8_t avg2 = vrhadd_u8(avg1, KJIXABC_);
+  const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
+  const uint32x2_t r3 = vreinterpret_u32_u8(avg2);
+  const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
+  const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
+  const uint32x2_t r0 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
+  vst1_lane_u32((uint32_t *)(dst + 0 * stride), r0, 0);
+  vst1_lane_u32((uint32_t *)(dst + 1 * stride), r1, 0);
+  vst1_lane_u32((uint32_t *)(dst + 2 * stride), r2, 0);
+  vst1_lane_u32((uint32_t *)(dst + 3 * stride), r3, 0);
+}
+
+#if !HAVE_NEON_ASM
+
+void aom_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint32x2_t d0u32 = vdup_n_u32(0);
+  (void)left;
+
+  d0u32 = vld1_lane_u32((const uint32_t *)above, d0u32, 0);
+  for (i = 0; i < 4; i++, dst += stride)
+    vst1_lane_u32((uint32_t *)dst, d0u32, 0);
+}
+
+void aom_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  (void)left;
+
+  d0u8 = vld1_u8(above);
+  for (i = 0; i < 8; i++, dst += stride) vst1_u8(dst, d0u8);
+}
+
+void aom_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  (void)left;
+
+  q0u8 = vld1q_u8(above);
+  for (i = 0; i < 16; i++, dst += stride) vst1q_u8(dst, q0u8);
+}
+
+void aom_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)left;
+
+  q0u8 = vld1q_u8(above);
+  q1u8 = vld1q_u8(above + 16);
+  for (i = 0; i < 32; i++, dst += stride) {
+    vst1q_u8(dst, q0u8);
+    vst1q_u8(dst + 16, q1u8);
+  }
+}
+
+void aom_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint32x2_t d1u32 = vdup_n_u32(0);
+  (void)above;
+
+  d1u32 = vld1_lane_u32((const uint32_t *)left, d1u32, 0);
+
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 0);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 1);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 2);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u32(d1u32), 3);
+  vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+}
+
+void aom_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint64x1_t d1u64 = vdup_n_u64(0);
+  (void)above;
+
+  d1u64 = vld1_u64((const uint64_t *)left);
+
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 0);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 1);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 2);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 3);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 4);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 5);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 6);
+  vst1_u8(dst, d0u8);
+  dst += stride;
+  d0u8 = vdup_lane_u8(vreinterpret_u8_u64(d1u64), 7);
+  vst1_u8(dst, d0u8);
+}
+
+void aom_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int j;
+  uint8x8_t d2u8 = vdup_n_u8(0);
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)above;
+
+  q1u8 = vld1q_u8(left);
+  d2u8 = vget_low_u8(q1u8);
+  for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) {
+    q0u8 = vdupq_lane_u8(d2u8, 0);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 1);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 2);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 3);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 4);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 5);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 6);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+    q0u8 = vdupq_lane_u8(d2u8, 7);
+    vst1q_u8(dst, q0u8);
+    dst += stride;
+  }
+}
+
+void aom_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint8x8_t d2u8 = vdup_n_u8(0);
+  uint8x16_t q0u8 = vdupq_n_u8(0);
+  uint8x16_t q1u8 = vdupq_n_u8(0);
+  (void)above;
+
+  for (k = 0; k < 2; k++, left += 16) {
+    q1u8 = vld1q_u8(left);
+    d2u8 = vget_low_u8(q1u8);
+    for (j = 0; j < 2; j++, d2u8 = vget_high_u8(q1u8)) {
+      q0u8 = vdupq_lane_u8(d2u8, 0);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 1);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 2);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 3);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 4);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 5);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 6);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+      q0u8 = vdupq_lane_u8(d2u8, 7);
+      vst1q_u8(dst, q0u8);
+      vst1q_u8(dst + 16, q0u8);
+      dst += stride;
+    }
+  }
+}
+
+void aom_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int i;
+  uint16x8_t q1u16, q3u16;
+  int16x8_t q1s16;
+  uint8x8_t d0u8 = vdup_n_u8(0);
+  uint32x2_t d2u32 = vdup_n_u32(0);
+
+  d0u8 = vld1_dup_u8(above - 1);
+  d2u32 = vld1_lane_u32((const uint32_t *)above, d2u32, 0);
+  q3u16 = vsubl_u8(vreinterpret_u8_u32(d2u32), d0u8);
+  for (i = 0; i < 4; i++, dst += stride) {
+    q1u16 = vdupq_n_u16((uint16_t)left[i]);
+    q1s16 =
+        vaddq_s16(vreinterpretq_s16_u16(q1u16), vreinterpretq_s16_u16(q3u16));
+    d0u8 = vqmovun_s16(q1s16);
+    vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d0u8), 0);
+  }
+}
+
+void aom_tm_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int j;
+  uint16x8_t q0u16, q3u16, q10u16;
+  int16x8_t q0s16;
+  uint16x4_t d20u16;
+  uint8x8_t d0u8, d2u8, d30u8;
+
+  d0u8 = vld1_dup_u8(above - 1);
+  d30u8 = vld1_u8(left);
+  d2u8 = vld1_u8(above);
+  q10u16 = vmovl_u8(d30u8);
+  q3u16 = vsubl_u8(d2u8, d0u8);
+  d20u16 = vget_low_u16(q10u16);
+  for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
+    q0u16 = vdupq_lane_u16(d20u16, 0);
+    q0s16 =
+        vaddq_s16(vreinterpretq_s16_u16(q3u16), vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 1);
+    q0s16 =
+        vaddq_s16(vreinterpretq_s16_u16(q3u16), vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 2);
+    q0s16 =
+        vaddq_s16(vreinterpretq_s16_u16(q3u16), vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+    q0u16 = vdupq_lane_u16(d20u16, 3);
+    q0s16 =
+        vaddq_s16(vreinterpretq_s16_u16(q3u16), vreinterpretq_s16_u16(q0u16));
+    d0u8 = vqmovun_s16(q0s16);
+    vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d0u8));
+    dst += stride;
+  }
+}
+
+void aom_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint16x8_t q0u16, q2u16, q3u16, q8u16, q10u16;
+  uint8x16_t q0u8, q1u8;
+  int16x8_t q0s16, q1s16, q8s16, q11s16;
+  uint16x4_t d20u16;
+  uint8x8_t d2u8, d3u8, d18u8, d22u8, d23u8;
+
+  q0u8 = vld1q_dup_u8(above - 1);
+  q1u8 = vld1q_u8(above);
+  q2u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
+  q3u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
+  for (k = 0; k < 2; k++, left += 8) {
+    d18u8 = vld1_u8(left);
+    q10u16 = vmovl_u8(d18u8);
+    d20u16 = vget_low_u16(q10u16);
+    for (j = 0; j < 2; j++, d20u16 = vget_high_u16(q10u16)) {
+      q0u16 = vdupq_lane_u16(d20u16, 0);
+      q8u16 = vdupq_lane_u16(d20u16, 1);
+      q1s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q2u16));
+      q0s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q3u16));
+      q11s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q8u16), vreinterpretq_s16_u16(q2u16));
+      q8s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q8u16), vreinterpretq_s16_u16(q3u16));
+      d2u8 = vqmovun_s16(q1s16);
+      d3u8 = vqmovun_s16(q0s16);
+      d22u8 = vqmovun_s16(q11s16);
+      d23u8 = vqmovun_s16(q8s16);
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
+      dst += stride;
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d20u16, 2);
+      q8u16 = vdupq_lane_u16(d20u16, 3);
+      q1s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q2u16));
+      q0s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q3u16));
+      q11s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q8u16), vreinterpretq_s16_u16(q2u16));
+      q8s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q8u16), vreinterpretq_s16_u16(q3u16));
+      d2u8 = vqmovun_s16(q1s16);
+      d3u8 = vqmovun_s16(q0s16);
+      d22u8 = vqmovun_s16(q11s16);
+      d23u8 = vqmovun_s16(q8s16);
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d2u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d3u8));
+      dst += stride;
+      vst1_u64((uint64_t *)dst, vreinterpret_u64_u8(d22u8));
+      vst1_u64((uint64_t *)(dst + 8), vreinterpret_u64_u8(d23u8));
+      dst += stride;
+    }
+  }
+}
+
+void aom_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  int j, k;
+  uint16x8_t q0u16, q3u16, q8u16, q9u16, q10u16, q11u16;
+  uint8x16_t q0u8, q1u8, q2u8;
+  int16x8_t q12s16, q13s16, q14s16, q15s16;
+  uint16x4_t d6u16;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8, d26u8;
+
+  q0u8 = vld1q_dup_u8(above - 1);
+  q1u8 = vld1q_u8(above);
+  q2u8 = vld1q_u8(above + 16);
+  q8u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q0u8));
+  q9u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q0u8));
+  q10u16 = vsubl_u8(vget_low_u8(q2u8), vget_low_u8(q0u8));
+  q11u16 = vsubl_u8(vget_high_u8(q2u8), vget_high_u8(q0u8));
+  for (k = 0; k < 4; k++, left += 8) {
+    d26u8 = vld1_u8(left);
+    q3u16 = vmovl_u8(d26u8);
+    d6u16 = vget_low_u16(q3u16);
+    for (j = 0; j < 2; j++, d6u16 = vget_high_u16(q3u16)) {
+      q0u16 = vdupq_lane_u16(d6u16, 0);
+      q12s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q8u16));
+      q13s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 1);
+      q12s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q8u16));
+      q13s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 2);
+      q12s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q8u16));
+      q13s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+
+      q0u16 = vdupq_lane_u16(d6u16, 3);
+      q12s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q8u16));
+      q13s16 =
+          vaddq_s16(vreinterpretq_s16_u16(q0u16), vreinterpretq_s16_u16(q9u16));
+      q14s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q10u16));
+      q15s16 = vaddq_s16(vreinterpretq_s16_u16(q0u16),
+                         vreinterpretq_s16_u16(q11u16));
+      d0u8 = vqmovun_s16(q12s16);
+      d1u8 = vqmovun_s16(q13s16);
+      d2u8 = vqmovun_s16(q14s16);
+      d3u8 = vqmovun_s16(q15s16);
+      q0u8 = vcombine_u8(d0u8, d1u8);
+      q1u8 = vcombine_u8(d2u8, d3u8);
+      vst1q_u64((uint64_t *)dst, vreinterpretq_u64_u8(q0u8));
+      vst1q_u64((uint64_t *)(dst + 16), vreinterpretq_u64_u8(q1u8));
+      dst += stride;
+    }
+  }
+}
+#endif  // !HAVE_NEON_ASM
diff --git a/third_party/aom/aom_dsp/arm/intrapred_neon_asm.asm b/third_party/aom/aom_dsp/arm/intrapred_neon_asm.asm
new file mode 100644
index 0000000000..7d04d35539
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/intrapred_neon_asm.asm
@@ -0,0 +1,633 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_v_predictor_4x4_neon|
+    EXPORT  |aom_v_predictor_8x8_neon|
+    EXPORT  |aom_v_predictor_16x16_neon|
+    EXPORT  |aom_v_predictor_32x32_neon|
+    EXPORT  |aom_h_predictor_4x4_neon|
+    EXPORT  |aom_h_predictor_8x8_neon|
+    EXPORT  |aom_h_predictor_16x16_neon|
+    EXPORT  |aom_h_predictor_32x32_neon|
+    EXPORT  |aom_tm_predictor_4x4_neon|
+    EXPORT  |aom_tm_predictor_8x8_neon|
+    EXPORT  |aom_tm_predictor_16x16_neon|
+    EXPORT  |aom_tm_predictor_32x32_neon|
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void aom_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_v_predictor_4x4_neon| PROC
+    vld1.32             {d0[0]}, [r2]
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d0[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_v_predictor_4x4_neon|
+
+;void aom_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_v_predictor_8x8_neon| PROC
+    vld1.8              {d0}, [r2]
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    vst1.8              {d0}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_v_predictor_8x8_neon|
+
+;void aom_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_v_predictor_16x16_neon| PROC
+    vld1.8              {q0}, [r2]
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    vst1.8              {q0}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_v_predictor_16x16_neon|
+
+;void aom_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_v_predictor_32x32_neon| PROC
+    vld1.8              {q0, q1}, [r2]
+    mov                 r2, #2
+loop_v
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    vst1.8              {q0, q1}, [r0], r1
+    subs                r2, r2, #1
+    bgt                 loop_v
+    bx                  lr
+    ENDP                ; |aom_v_predictor_32x32_neon|
+
+;void aom_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_h_predictor_4x4_neon| PROC
+    vld1.32             {d1[0]}, [r3]
+    vdup.8              d0, d1[0]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[1]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[2]
+    vst1.32             {d0[0]}, [r0], r1
+    vdup.8              d0, d1[3]
+    vst1.32             {d0[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_h_predictor_4x4_neon|
+
+;void aom_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                              const uint8_t *above,
+;                              const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_h_predictor_8x8_neon| PROC
+    vld1.64             {d1}, [r3]
+    vdup.8              d0, d1[0]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[1]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[2]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[3]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[4]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[5]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[6]
+    vst1.64             {d0}, [r0], r1
+    vdup.8              d0, d1[7]
+    vst1.64             {d0}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_h_predictor_8x8_neon|
+
+;void aom_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_h_predictor_16x16_neon| PROC
+    vld1.8              {q1}, [r3]
+    vdup.8              q0, d2[0]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[1]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[2]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[3]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[4]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[5]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[6]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[7]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[0]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[1]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[2]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[3]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[4]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[5]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[6]
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[7]
+    vst1.8              {q0}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_h_predictor_16x16_neon|
+
+;void aom_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_h_predictor_32x32_neon| PROC
+    sub                 r1, r1, #16
+    mov                 r2, #2
+loop_h
+    vld1.8              {q1}, [r3]!
+    vdup.8              q0, d2[0]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[1]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[2]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[3]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[4]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[5]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[6]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d2[7]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[0]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[1]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[2]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[3]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[4]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[5]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[6]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    vdup.8              q0, d3[7]
+    vst1.8              {q0}, [r0]!
+    vst1.8              {q0}, [r0], r1
+    subs                r2, r2, #1
+    bgt                 loop_h
+    bx                  lr
+    ENDP                ; |aom_h_predictor_32x32_neon|
+
+;void aom_tm_predictor_4x4_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_tm_predictor_4x4_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.u8             {d0[]}, [r12]
+
+    ; Load above 4 pixels
+    vld1.32             {d2[0]}, [r2]
+
+    ; Compute above - ytop_left
+    vsubl.u8            q3, d2, d0
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; 1st row and 2nd row
+    vld1.u8             {d2[]}, [r3]!
+    vld1.u8             {d4[]}, [r3]!
+    vmovl.u8            q1, d2
+    vmovl.u8            q2, d4
+    vadd.s16            q1, q1, q3
+    vadd.s16            q2, q2, q3
+    vqmovun.s16         d0, q1
+    vqmovun.s16         d1, q2
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d1[0]}, [r0], r1
+
+    ; 3rd row and 4th row
+    vld1.u8             {d2[]}, [r3]!
+    vld1.u8             {d4[]}, [r3]
+    vmovl.u8            q1, d2
+    vmovl.u8            q2, d4
+    vadd.s16            q1, q1, q3
+    vadd.s16            q2, q2, q3
+    vqmovun.s16         d0, q1
+    vqmovun.s16         d1, q2
+    vst1.32             {d0[0]}, [r0], r1
+    vst1.32             {d1[0]}, [r0], r1
+    bx                  lr
+    ENDP                ; |aom_tm_predictor_4x4_neon|
+
+;void aom_tm_predictor_8x8_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_tm_predictor_8x8_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; preload 8 left
+    vld1.8              {d30}, [r3]
+
+    ; Load above 8 pixels
+    vld1.64             {d2}, [r2]
+
+    vmovl.u8            q10, d30
+
+    ; Compute above - ytop_left
+    vsubl.u8            q3, d2, d0
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; 1st row and 2nd row
+    vdup.16             q0, d20[0]
+    vdup.16             q1, d20[1]
+    vadd.s16            q0, q3, q0
+    vadd.s16            q1, q3, q1
+
+    ; 3rd row and 4th row
+    vdup.16             q8, d20[2]
+    vdup.16             q9, d20[3]
+    vadd.s16            q8, q3, q8
+    vadd.s16            q9, q3, q9
+
+    vqmovun.s16         d0, q0
+    vqmovun.s16         d1, q1
+    vqmovun.s16         d2, q8
+    vqmovun.s16         d3, q9
+
+    vst1.64             {d0}, [r0], r1
+    vst1.64             {d1}, [r0], r1
+    vst1.64             {d2}, [r0], r1
+    vst1.64             {d3}, [r0], r1
+
+    ; 5th row and 6th row
+    vdup.16             q0, d21[0]
+    vdup.16             q1, d21[1]
+    vadd.s16            q0, q3, q0
+    vadd.s16            q1, q3, q1
+
+    ; 7th row and 8th row
+    vdup.16             q8, d21[2]
+    vdup.16             q9, d21[3]
+    vadd.s16            q8, q3, q8
+    vadd.s16            q9, q3, q9
+
+    vqmovun.s16         d0, q0
+    vqmovun.s16         d1, q1
+    vqmovun.s16         d2, q8
+    vqmovun.s16         d3, q9
+
+    vst1.64             {d0}, [r0], r1
+    vst1.64             {d1}, [r0], r1
+    vst1.64             {d2}, [r0], r1
+    vst1.64             {d3}, [r0], r1
+
+    bx                  lr
+    ENDP                ; |aom_tm_predictor_8x8_neon|
+
+;void aom_tm_predictor_16x16_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                const uint8_t *above,
+;                                const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_tm_predictor_16x16_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; Load above 8 pixels
+    vld1.8              {q1}, [r2]
+
+    ; preload 8 left into r12
+    vld1.8              {d18}, [r3]!
+
+    ; Compute above - ytop_left
+    vsubl.u8            q2, d2, d0
+    vsubl.u8            q3, d3, d0
+
+    vmovl.u8            q10, d18
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; Process 8 rows in each single loop and loop 2 times to process 16 rows.
+    mov                 r2, #2
+
+loop_16x16_neon
+    ; Process two rows.
+    vdup.16             q0, d20[0]
+    vdup.16             q8, d20[1]
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d20[2]                  ; proload next 2 rows data
+    vdup.16             q8, d20[3]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d21[0]                  ; proload next 2 rows data
+    vdup.16             q8, d21[1]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vdup.16             q0, d21[2]                  ; proload next 2 rows data
+    vdup.16             q8, d21[3]
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+
+    vadd.s16            q1, q0, q2
+    vadd.s16            q0, q0, q3
+    vadd.s16            q11, q8, q2
+    vadd.s16            q8, q8, q3
+    vqmovun.s16         d2, q1
+    vqmovun.s16         d3, q0
+    vqmovun.s16         d22, q11
+    vqmovun.s16         d23, q8
+    vld1.8              {d18}, [r3]!                  ; preload 8 left into r12
+    vmovl.u8            q10, d18
+    vst1.64             {d2,d3}, [r0], r1
+    vst1.64             {d22,d23}, [r0], r1
+
+    subs                r2, r2, #1
+    bgt                 loop_16x16_neon
+
+    bx                  lr
+    ENDP                ; |aom_tm_predictor_16x16_neon|
+
+;void aom_tm_predictor_32x32_neon (uint8_t *dst, ptrdiff_t y_stride,
+;                                  const uint8_t *above,
+;                                  const uint8_t *left)
+; r0  uint8_t *dst
+; r1  ptrdiff_t y_stride
+; r2  const uint8_t *above
+; r3  const uint8_t *left
+
+|aom_tm_predictor_32x32_neon| PROC
+    ; Load ytop_left = above[-1];
+    sub                 r12, r2, #1
+    vld1.8              {d0[]}, [r12]
+
+    ; Load above 32 pixels
+    vld1.8              {q1}, [r2]!
+    vld1.8              {q2}, [r2]
+
+    ; preload 8 left pixels
+    vld1.8              {d26}, [r3]!
+
+    ; Compute above - ytop_left
+    vsubl.u8            q8, d2, d0
+    vsubl.u8            q9, d3, d0
+    vsubl.u8            q10, d4, d0
+    vsubl.u8            q11, d5, d0
+
+    vmovl.u8            q3, d26
+
+    ; Load left row by row and compute left + (above - ytop_left)
+    ; Process 8 rows in each single loop and loop 4 times to process 32 rows.
+    mov                 r2, #4
+
+loop_32x32_neon
+    ; Process two rows.
+    vdup.16             q0, d6[0]
+    vdup.16             q2, d6[1]
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q1, d6[2]
+    vdup.16             q2, d6[3]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q1, q8
+    vadd.s16            q13, q1, q9
+    vadd.s16            q14, q1, q10
+    vadd.s16            q15, q1, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q0, d7[0]
+    vdup.16             q2, d7[1]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vdup.16             q0, d7[2]
+    vdup.16             q2, d7[3]
+    vst1.64             {d24-d27}, [r0], r1
+
+    ; Process two rows.
+    vadd.s16            q12, q0, q8
+    vadd.s16            q13, q0, q9
+    vadd.s16            q14, q0, q10
+    vadd.s16            q15, q0, q11
+    vqmovun.s16         d0, q12
+    vqmovun.s16         d1, q13
+    vadd.s16            q12, q2, q8
+    vadd.s16            q13, q2, q9
+    vqmovun.s16         d2, q14
+    vqmovun.s16         d3, q15
+    vadd.s16            q14, q2, q10
+    vadd.s16            q15, q2, q11
+    vst1.64             {d0-d3}, [r0], r1
+    vqmovun.s16         d24, q12
+    vqmovun.s16         d25, q13
+    vld1.8              {d0}, [r3]!                   ; preload 8 left pixels
+    vqmovun.s16         d26, q14
+    vqmovun.s16         d27, q15
+    vmovl.u8            q3, d0
+    vst1.64             {d24-d27}, [r0], r1
+
+    subs                r2, r2, #1
+    bgt                 loop_32x32_neon
+
+    bx                  lr
+    ENDP                ; |aom_tm_predictor_32x32_neon|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/loopfilter_16_neon.asm b/third_party/aom/aom_dsp/arm/loopfilter_16_neon.asm
new file mode 100644
index 0000000000..b6e2c9edb4
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/loopfilter_16_neon.asm
@@ -0,0 +1,202 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_lpf_horizontal_4_dual_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void aom_lpf_horizontal_4_dual_neon(uint8_t *s, int p,
+;                                    const uint8_t *blimit0,
+;                                    const uint8_t *limit0,
+;                                    const uint8_t *thresh0,
+;                                    const uint8_t *blimit1,
+;                                    const uint8_t *limit1,
+;                                    const uint8_t *thresh1)
+; r0    uint8_t *s,
+; r1    int p,
+; r2    const uint8_t *blimit0,
+; r3    const uint8_t *limit0,
+; sp    const uint8_t *thresh0,
+; sp+4  const uint8_t *blimit1,
+; sp+8  const uint8_t *limit1,
+; sp+12 const uint8_t *thresh1,
+
+|aom_lpf_horizontal_4_dual_neon| PROC
+    push        {lr}
+
+    ldr         r12, [sp, #4]              ; load thresh0
+    vld1.8      {d0}, [r2]                 ; load blimit0 to first half q
+    vld1.8      {d2}, [r3]                 ; load limit0 to first half q
+
+    add         r1, r1, r1                 ; double pitch
+    ldr         r2, [sp, #8]               ; load blimit1
+
+    vld1.8      {d4}, [r12]                ; load thresh0 to first half q
+
+    ldr         r3, [sp, #12]              ; load limit1
+    ldr         r12, [sp, #16]             ; load thresh1
+    vld1.8      {d1}, [r2]                 ; load blimit1 to 2nd half q
+
+    sub         r2, r0, r1, lsl #1         ; s[-4 * p]
+
+    vld1.8      {d3}, [r3]                 ; load limit1 to 2nd half q
+    vld1.8      {d5}, [r12]                ; load thresh1 to 2nd half q
+
+    vpush       {d8-d15}                   ; save neon registers
+
+    add         r3, r2, r1, lsr #1         ; s[-3 * p]
+
+    vld1.u8     {q3}, [r2@64], r1          ; p3
+    vld1.u8     {q4}, [r3@64], r1          ; p2
+    vld1.u8     {q5}, [r2@64], r1          ; p1
+    vld1.u8     {q6}, [r3@64], r1          ; p0
+    vld1.u8     {q7}, [r2@64], r1          ; q0
+    vld1.u8     {q8}, [r3@64], r1          ; q1
+    vld1.u8     {q9}, [r2@64]              ; q2
+    vld1.u8     {q10}, [r3@64]             ; q3
+
+    sub         r2, r2, r1, lsl #1
+    sub         r3, r3, r1, lsl #1
+
+    bl          aom_loop_filter_neon_16
+
+    vst1.u8     {q5}, [r2@64], r1          ; store op1
+    vst1.u8     {q6}, [r3@64], r1          ; store op0
+    vst1.u8     {q7}, [r2@64], r1          ; store oq0
+    vst1.u8     {q8}, [r3@64], r1          ; store oq1
+
+    vpop        {d8-d15}                   ; restore neon registers
+
+    pop         {pc}
+    ENDP        ; |aom_lpf_horizontal_4_dual_neon|
+
+; void aom_loop_filter_neon_16();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. This function uses
+; registers d8-d15, so the calling function must save those registers.
+;
+; r0-r3, r12 PRESERVE
+; q0    blimit
+; q1    limit
+; q2    thresh
+; q3    p3
+; q4    p2
+; q5    p1
+; q6    p0
+; q7    q0
+; q8    q1
+; q9    q2
+; q10   q3
+;
+; Outputs:
+; q5    op1
+; q6    op0
+; q7    oq0
+; q8    oq1
+|aom_loop_filter_neon_16| PROC
+
+    ; filter_mask
+    vabd.u8     q11, q3, q4                 ; m1 = abs(p3 - p2)
+    vabd.u8     q12, q4, q5                 ; m2 = abs(p2 - p1)
+    vabd.u8     q13, q5, q6                 ; m3 = abs(p1 - p0)
+    vabd.u8     q14, q8, q7                 ; m4 = abs(q1 - q0)
+    vabd.u8     q3, q9, q8                  ; m5 = abs(q2 - q1)
+    vabd.u8     q4, q10, q9                 ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     q11, q11, q12               ; m7 = max(m1, m2)
+    vmax.u8     q12, q13, q14               ; m8 = max(m3, m4)
+
+    vabd.u8     q9, q6, q7                  ; abs(p0 - q0)
+
+    vmax.u8     q3, q3, q4                  ; m9 = max(m5, m6)
+
+    vmov.u8     q10, #0x80
+
+    vmax.u8     q15, q11, q12               ; m10 = max(m7, m8)
+
+    vcgt.u8     q13, q13, q2                ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     q14, q14, q2                ; (abs(q1 - q0) > thresh)*-1
+    vmax.u8     q15, q15, q3                ; m11 = max(m10, m9)
+
+    vabd.u8     q2, q5, q8                  ; a = abs(p1 - q1)
+    vqadd.u8    q9, q9, q9                  ; b = abs(p0 - q0) * 2
+
+    veor        q7, q7, q10                 ; qs0
+
+    vcge.u8     q15, q1, q15                ; abs(m11) > limit
+
+    vshr.u8     q2, q2, #1                  ; a = a / 2
+    veor        q6, q6, q10                 ; ps0
+
+    veor        q5, q5, q10                 ; ps1
+    vqadd.u8    q9, q9, q2                  ; a = b + a
+
+    veor        q8, q8, q10                 ; qs1
+
+    vmov.u16    q4, #3
+
+    vsubl.s8    q2, d14, d12                ; ( qs0 - ps0)
+    vsubl.s8    q11, d15, d13
+
+    vcge.u8     q9, q0, q9                  ; a > blimit
+
+    vqsub.s8    q1, q5, q8                  ; filter = clamp(ps1-qs1)
+    vorr        q14, q13, q14               ; hev
+
+    vmul.i16    q2, q2, q4                  ; 3 * ( qs0 - ps0)
+    vmul.i16    q11, q11, q4
+
+    vand        q1, q1, q14                 ; filter &= hev
+    vand        q15, q15, q9                ; mask
+
+    vmov.u8     q4, #3
+
+    vaddw.s8    q2, q2, d2                  ; filter + 3 * (qs0 - ps0)
+    vaddw.s8    q11, q11, d3
+
+    vmov.u8     q9, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d2, q2
+    vqmovn.s16  d3, q11
+    vand        q1, q1, q15                 ; filter &= mask
+
+    vqadd.s8    q2, q1, q4                  ; filter2 = clamp(filter+3)
+    vqadd.s8    q1, q1, q9                  ; filter1 = clamp(filter+4)
+    vshr.s8     q2, q2, #3                  ; filter2 >>= 3
+    vshr.s8     q1, q1, #3                  ; filter1 >>= 3
+
+
+    vqadd.s8    q11, q6, q2                 ; u = clamp(ps0 + filter2)
+    vqsub.s8    q0, q7, q1                  ; u = clamp(qs0 - filter1)
+
+    ; outer tap adjustments
+    vrshr.s8    q1, q1, #1                  ; filter = ++filter1 >> 1
+
+    veor        q7, q0,  q10                ; *oq0 = u^0x80
+
+    vbic        q1, q1, q14                 ; filter &= ~hev
+
+    vqadd.s8    q13, q5, q1                 ; u = clamp(ps1 + filter)
+    vqsub.s8    q12, q8, q1                 ; u = clamp(qs1 - filter)
+
+    veor        q6, q11, q10                ; *op0 = u^0x80
+    veor        q5, q13, q10                ; *op1 = u^0x80
+    veor        q8, q12, q10                ; *oq1 = u^0x80
+
+    bx          lr
+    ENDP        ; |aom_loop_filter_neon_16|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/loopfilter_16_neon.c b/third_party/aom/aom_dsp/arm/loopfilter_16_neon.c
new file mode 100644
index 0000000000..c0562a6ea5
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/loopfilter_16_neon.c
@@ -0,0 +1,174 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+static INLINE void loop_filter_neon_16(uint8x16_t qblimit,  // blimit
+                                       uint8x16_t qlimit,   // limit
+                                       uint8x16_t qthresh,  // thresh
+                                       uint8x16_t q3,       // p3
+                                       uint8x16_t q4,       // p2
+                                       uint8x16_t q5,       // p1
+                                       uint8x16_t q6,       // p0
+                                       uint8x16_t q7,       // q0
+                                       uint8x16_t q8,       // q1
+                                       uint8x16_t q9,       // q2
+                                       uint8x16_t q10,      // q3
+                                       uint8x16_t *q5r,     // p1
+                                       uint8x16_t *q6r,     // p0
+                                       uint8x16_t *q7r,     // q0
+                                       uint8x16_t *q8r) {   // q1
+  uint8x16_t q1u8, q2u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+  int16x8_t q2s16, q11s16;
+  uint16x8_t q4u16;
+  int8x16_t q0s8, q1s8, q2s8, q11s8, q12s8, q13s8;
+  int8x8_t d2s8, d3s8;
+
+  q11u8 = vabdq_u8(q3, q4);
+  q12u8 = vabdq_u8(q4, q5);
+  q13u8 = vabdq_u8(q5, q6);
+  q14u8 = vabdq_u8(q8, q7);
+  q3 = vabdq_u8(q9, q8);
+  q4 = vabdq_u8(q10, q9);
+
+  q11u8 = vmaxq_u8(q11u8, q12u8);
+  q12u8 = vmaxq_u8(q13u8, q14u8);
+  q3 = vmaxq_u8(q3, q4);
+  q15u8 = vmaxq_u8(q11u8, q12u8);
+
+  q9 = vabdq_u8(q6, q7);
+
+  // aom_hevmask
+  q13u8 = vcgtq_u8(q13u8, qthresh);
+  q14u8 = vcgtq_u8(q14u8, qthresh);
+  q15u8 = vmaxq_u8(q15u8, q3);
+
+  q2u8 = vabdq_u8(q5, q8);
+  q9 = vqaddq_u8(q9, q9);
+
+  q15u8 = vcgeq_u8(qlimit, q15u8);
+
+  // aom_filter() function
+  // convert to signed
+  q10 = vdupq_n_u8(0x80);
+  q8 = veorq_u8(q8, q10);
+  q7 = veorq_u8(q7, q10);
+  q6 = veorq_u8(q6, q10);
+  q5 = veorq_u8(q5, q10);
+
+  q2u8 = vshrq_n_u8(q2u8, 1);
+  q9 = vqaddq_u8(q9, q2u8);
+
+  q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
+                   vget_low_s8(vreinterpretq_s8_u8(q6)));
+  q11s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
+                    vget_high_s8(vreinterpretq_s8_u8(q6)));
+
+  q9 = vcgeq_u8(qblimit, q9);
+
+  q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5), vreinterpretq_s8_u8(q8));
+
+  q14u8 = vorrq_u8(q13u8, q14u8);
+
+  q4u16 = vdupq_n_u16(3);
+  q2s16 = vmulq_s16(q2s16, vreinterpretq_s16_u16(q4u16));
+  q11s16 = vmulq_s16(q11s16, vreinterpretq_s16_u16(q4u16));
+
+  q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q14u8);
+  q15u8 = vandq_u8(q15u8, q9);
+
+  q1s8 = vreinterpretq_s8_u8(q1u8);
+  q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
+  q11s16 = vaddw_s8(q11s16, vget_high_s8(q1s8));
+
+  q4 = vdupq_n_u8(3);
+  q9 = vdupq_n_u8(4);
+  // aom_filter = clamp(aom_filter + 3 * ( qs0 - ps0))
+  d2s8 = vqmovn_s16(q2s16);
+  d3s8 = vqmovn_s16(q11s16);
+  q1s8 = vcombine_s8(d2s8, d3s8);
+  q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q15u8);
+  q1s8 = vreinterpretq_s8_u8(q1u8);
+
+  q2s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q4));
+  q1s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q9));
+  q2s8 = vshrq_n_s8(q2s8, 3);
+  q1s8 = vshrq_n_s8(q1s8, 3);
+
+  q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8);
+  q0s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q1s8);
+
+  q1s8 = vrshrq_n_s8(q1s8, 1);
+  q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+  q13s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q1s8);
+  q12s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q1s8);
+
+  *q8r = veorq_u8(vreinterpretq_u8_s8(q12s8), q10);
+  *q7r = veorq_u8(vreinterpretq_u8_s8(q0s8), q10);
+  *q6r = veorq_u8(vreinterpretq_u8_s8(q11s8), q10);
+  *q5r = veorq_u8(vreinterpretq_u8_s8(q13s8), q10);
+  return;
+}
+
+void aom_lpf_horizontal_4_dual_neon(
+    uint8_t *s, int p /* pitch */, const uint8_t *blimit0,
+    const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1,
+    const uint8_t *limit1, const uint8_t *thresh1) {
+  uint8x8_t dblimit0, dlimit0, dthresh0, dblimit1, dlimit1, dthresh1;
+  uint8x16_t qblimit, qlimit, qthresh;
+  uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8;
+
+  dblimit0 = vld1_u8(blimit0);
+  dlimit0 = vld1_u8(limit0);
+  dthresh0 = vld1_u8(thresh0);
+  dblimit1 = vld1_u8(blimit1);
+  dlimit1 = vld1_u8(limit1);
+  dthresh1 = vld1_u8(thresh1);
+  qblimit = vcombine_u8(dblimit0, dblimit1);
+  qlimit = vcombine_u8(dlimit0, dlimit1);
+  qthresh = vcombine_u8(dthresh0, dthresh1);
+
+  s -= (p << 2);
+
+  q3u8 = vld1q_u8(s);
+  s += p;
+  q4u8 = vld1q_u8(s);
+  s += p;
+  q5u8 = vld1q_u8(s);
+  s += p;
+  q6u8 = vld1q_u8(s);
+  s += p;
+  q7u8 = vld1q_u8(s);
+  s += p;
+  q8u8 = vld1q_u8(s);
+  s += p;
+  q9u8 = vld1q_u8(s);
+  s += p;
+  q10u8 = vld1q_u8(s);
+
+  loop_filter_neon_16(qblimit, qlimit, qthresh, q3u8, q4u8, q5u8, q6u8, q7u8,
+                      q8u8, q9u8, q10u8, &q5u8, &q6u8, &q7u8, &q8u8);
+
+  s -= (p * 5);
+  vst1q_u8(s, q5u8);
+  s += p;
+  vst1q_u8(s, q6u8);
+  s += p;
+  vst1q_u8(s, q7u8);
+  s += p;
+  vst1q_u8(s, q8u8);
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/loopfilter_4_neon.asm b/third_party/aom/aom_dsp/arm/loopfilter_4_neon.asm
new file mode 100644
index 0000000000..8b54984d56
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/loopfilter_4_neon.asm
@@ -0,0 +1,252 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_lpf_horizontal_4_neon|
+    EXPORT  |aom_lpf_vertical_4_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; Currently aom only works on iterations 8 at a time. The aom loop filter
+; works on 16 iterations at a time.
+;
+; void aom_lpf_horizontal_4_neon(uint8_t *s,
+;                                int p /* pitch */,
+;                                const uint8_t *blimit,
+;                                const uint8_t *limit,
+;                                const uint8_t *thresh)
+;
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|aom_lpf_horizontal_4_neon| PROC
+    push        {lr}
+
+    vld1.8      {d0[]}, [r2]               ; duplicate *blimit
+    ldr         r2, [sp, #4]               ; load thresh
+    add         r1, r1, r1                 ; double pitch
+
+    vld1.8      {d1[]}, [r3]               ; duplicate *limit
+    vld1.8      {d2[]}, [r2]               ; duplicate *thresh
+
+    sub         r2, r0, r1, lsl #1         ; move src pointer down by 4 lines
+    add         r3, r2, r1, lsr #1         ; set to 3 lines down
+
+    vld1.u8     {d3}, [r2@64], r1          ; p3
+    vld1.u8     {d4}, [r3@64], r1          ; p2
+    vld1.u8     {d5}, [r2@64], r1          ; p1
+    vld1.u8     {d6}, [r3@64], r1          ; p0
+    vld1.u8     {d7}, [r2@64], r1          ; q0
+    vld1.u8     {d16}, [r3@64], r1         ; q1
+    vld1.u8     {d17}, [r2@64]             ; q2
+    vld1.u8     {d18}, [r3@64]             ; q3
+
+    sub         r2, r2, r1, lsl #1
+    sub         r3, r3, r1, lsl #1
+
+    bl          aom_loop_filter_neon
+
+    vst1.u8     {d4}, [r2@64], r1          ; store op1
+    vst1.u8     {d5}, [r3@64], r1          ; store op0
+    vst1.u8     {d6}, [r2@64], r1          ; store oq0
+    vst1.u8     {d7}, [r3@64], r1          ; store oq1
+
+    pop         {pc}
+    ENDP        ; |aom_lpf_horizontal_4_neon|
+
+; Currently aom only works on iterations 8 at a time. The aom loop filter
+; works on 16 iterations at a time.
+;
+; void aom_lpf_vertical_4_neon(uint8_t *s,
+;                              int p /* pitch */,
+;                              const uint8_t *blimit,
+;                              const uint8_t *limit,
+;                              const uint8_t *thresh)
+;
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|aom_lpf_vertical_4_neon| PROC
+    push        {lr}
+
+    vld1.8      {d0[]}, [r2]              ; duplicate *blimit
+    vld1.8      {d1[]}, [r3]              ; duplicate *limit
+
+    ldr         r3, [sp, #4]              ; load thresh
+    sub         r2, r0, #4                ; move s pointer down by 4 columns
+
+    vld1.8      {d2[]}, [r3]              ; duplicate *thresh
+
+    vld1.u8     {d3}, [r2], r1             ; load s data
+    vld1.u8     {d4}, [r2], r1
+    vld1.u8     {d5}, [r2], r1
+    vld1.u8     {d6}, [r2], r1
+    vld1.u8     {d7}, [r2], r1
+    vld1.u8     {d16}, [r2], r1
+    vld1.u8     {d17}, [r2], r1
+    vld1.u8     {d18}, [r2]
+
+    ;transpose to 8x16 matrix
+    vtrn.32     d3, d7
+    vtrn.32     d4, d16
+    vtrn.32     d5, d17
+    vtrn.32     d6, d18
+
+    vtrn.16     d3, d5
+    vtrn.16     d4, d6
+    vtrn.16     d7, d17
+    vtrn.16     d16, d18
+
+    vtrn.8      d3, d4
+    vtrn.8      d5, d6
+    vtrn.8      d7, d16
+    vtrn.8      d17, d18
+
+    bl          aom_loop_filter_neon
+
+    sub         r0, r0, #2
+
+    ;store op1, op0, oq0, oq1
+    vst4.8      {d4[0], d5[0], d6[0], d7[0]}, [r0], r1
+    vst4.8      {d4[1], d5[1], d6[1], d7[1]}, [r0], r1
+    vst4.8      {d4[2], d5[2], d6[2], d7[2]}, [r0], r1
+    vst4.8      {d4[3], d5[3], d6[3], d7[3]}, [r0], r1
+    vst4.8      {d4[4], d5[4], d6[4], d7[4]}, [r0], r1
+    vst4.8      {d4[5], d5[5], d6[5], d7[5]}, [r0], r1
+    vst4.8      {d4[6], d5[6], d6[6], d7[6]}, [r0], r1
+    vst4.8      {d4[7], d5[7], d6[7], d7[7]}, [r0]
+
+    pop         {pc}
+    ENDP        ; |aom_lpf_vertical_4_neon|
+
+; void aom_loop_filter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. The function does not use
+; registers d8-d15.
+;
+; Inputs:
+; r0-r3, r12 PRESERVE
+; d0    blimit
+; d1    limit
+; d2    thresh
+; d3    p3
+; d4    p2
+; d5    p1
+; d6    p0
+; d7    q0
+; d16   q1
+; d17   q2
+; d18   q3
+;
+; Outputs:
+; d4    op1
+; d5    op0
+; d6    oq0
+; d7    oq1
+|aom_loop_filter_neon| PROC
+    ; filter_mask
+    vabd.u8     d19, d3, d4                 ; m1 = abs(p3 - p2)
+    vabd.u8     d20, d4, d5                 ; m2 = abs(p2 - p1)
+    vabd.u8     d21, d5, d6                 ; m3 = abs(p1 - p0)
+    vabd.u8     d22, d16, d7                ; m4 = abs(q1 - q0)
+    vabd.u8     d3, d17, d16                ; m5 = abs(q2 - q1)
+    vabd.u8     d4, d18, d17                ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20               ; m1 = max(m1, m2)
+    vmax.u8     d20, d21, d22               ; m2 = max(m3, m4)
+
+    vabd.u8     d17, d6, d7                 ; abs(p0 - q0)
+
+    vmax.u8     d3, d3, d4                  ; m3 = max(m5, m6)
+
+    vmov.u8     d18, #0x80
+
+    vmax.u8     d23, d19, d20               ; m1 = max(m1, m2)
+
+    ; hevmask
+    vcgt.u8     d21, d21, d2                ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     d22, d22, d2                ; (abs(q1 - q0) > thresh)*-1
+    vmax.u8     d23, d23, d3                ; m1 = max(m1, m3)
+
+    vabd.u8     d28, d5, d16                ; a = abs(p1 - q1)
+    vqadd.u8    d17, d17, d17               ; b = abs(p0 - q0) * 2
+
+    veor        d7, d7, d18                 ; qs0
+
+    vcge.u8     d23, d1, d23                ; abs(m1) > limit
+
+    ; filter() function
+    ; convert to signed
+
+    vshr.u8     d28, d28, #1                ; a = a / 2
+    veor        d6, d6, d18                 ; ps0
+
+    veor        d5, d5, d18                 ; ps1
+    vqadd.u8    d17, d17, d28               ; a = b + a
+
+    veor        d16, d16, d18               ; qs1
+
+    vmov.u8     d19, #3
+
+    vsub.s8     d28, d7, d6                 ; ( qs0 - ps0)
+
+    vcge.u8     d17, d0, d17                ; a > blimit
+
+    vqsub.s8    d27, d5, d16                ; filter = clamp(ps1-qs1)
+    vorr        d22, d21, d22               ; hevmask
+
+    vmull.s8    q12, d28, d19               ; 3 * ( qs0 - ps0)
+
+    vand        d27, d27, d22               ; filter &= hev
+    vand        d23, d23, d17               ; filter_mask
+
+    vaddw.s8    q12, q12, d27               ; filter + 3 * (qs0 - ps0)
+
+    vmov.u8     d17, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d27, q12
+
+    vand        d27, d27, d23               ; filter &= mask
+
+    vqadd.s8    d28, d27, d19               ; filter2 = clamp(filter+3)
+    vqadd.s8    d27, d27, d17               ; filter1 = clamp(filter+4)
+    vshr.s8     d28, d28, #3                ; filter2 >>= 3
+    vshr.s8     d27, d27, #3                ; filter1 >>= 3
+
+    vqadd.s8    d19, d6, d28                ; u = clamp(ps0 + filter2)
+    vqsub.s8    d26, d7, d27                ; u = clamp(qs0 - filter1)
+
+    ; outer tap adjustments
+    vrshr.s8    d27, d27, #1                ; filter = ++filter1 >> 1
+
+    veor        d6, d26, d18                ; *oq0 = u^0x80
+
+    vbic        d27, d27, d22               ; filter &= ~hev
+
+    vqadd.s8    d21, d5, d27                ; u = clamp(ps1 + filter)
+    vqsub.s8    d20, d16, d27               ; u = clamp(qs1 - filter)
+
+    veor        d5, d19, d18                ; *op0 = u^0x80
+    veor        d4, d21, d18                ; *op1 = u^0x80
+    veor        d7, d20, d18                ; *oq1 = u^0x80
+
+    bx          lr
+    ENDP        ; |aom_loop_filter_neon|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/loopfilter_4_neon.c b/third_party/aom/aom_dsp/arm/loopfilter_4_neon.c
new file mode 100644
index 0000000000..2b1f80b81e
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/loopfilter_4_neon.c
@@ -0,0 +1,250 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+
+static INLINE void loop_filter_neon(uint8x8_t dblimit,   // flimit
+                                    uint8x8_t dlimit,    // limit
+                                    uint8x8_t dthresh,   // thresh
+                                    uint8x8_t d3u8,      // p3
+                                    uint8x8_t d4u8,      // p2
+                                    uint8x8_t d5u8,      // p1
+                                    uint8x8_t d6u8,      // p0
+                                    uint8x8_t d7u8,      // q0
+                                    uint8x8_t d16u8,     // q1
+                                    uint8x8_t d17u8,     // q2
+                                    uint8x8_t d18u8,     // q3
+                                    uint8x8_t *d4ru8,    // p1
+                                    uint8x8_t *d5ru8,    // p0
+                                    uint8x8_t *d6ru8,    // q0
+                                    uint8x8_t *d7ru8) {  // q1
+  uint8x8_t d19u8, d20u8, d21u8, d22u8, d23u8, d27u8, d28u8;
+  int16x8_t q12s16;
+  int8x8_t d19s8, d20s8, d21s8, d26s8, d27s8, d28s8;
+
+  d19u8 = vabd_u8(d3u8, d4u8);
+  d20u8 = vabd_u8(d4u8, d5u8);
+  d21u8 = vabd_u8(d5u8, d6u8);
+  d22u8 = vabd_u8(d16u8, d7u8);
+  d3u8 = vabd_u8(d17u8, d16u8);
+  d4u8 = vabd_u8(d18u8, d17u8);
+
+  d19u8 = vmax_u8(d19u8, d20u8);
+  d20u8 = vmax_u8(d21u8, d22u8);
+  d3u8 = vmax_u8(d3u8, d4u8);
+  d23u8 = vmax_u8(d19u8, d20u8);
+
+  d17u8 = vabd_u8(d6u8, d7u8);
+
+  d21u8 = vcgt_u8(d21u8, dthresh);
+  d22u8 = vcgt_u8(d22u8, dthresh);
+  d23u8 = vmax_u8(d23u8, d3u8);
+
+  d28u8 = vabd_u8(d5u8, d16u8);
+  d17u8 = vqadd_u8(d17u8, d17u8);
+
+  d23u8 = vcge_u8(dlimit, d23u8);
+
+  d18u8 = vdup_n_u8(0x80);
+  d5u8 = veor_u8(d5u8, d18u8);
+  d6u8 = veor_u8(d6u8, d18u8);
+  d7u8 = veor_u8(d7u8, d18u8);
+  d16u8 = veor_u8(d16u8, d18u8);
+
+  d28u8 = vshr_n_u8(d28u8, 1);
+  d17u8 = vqadd_u8(d17u8, d28u8);
+
+  d19u8 = vdup_n_u8(3);
+
+  d28s8 = vsub_s8(vreinterpret_s8_u8(d7u8), vreinterpret_s8_u8(d6u8));
+
+  d17u8 = vcge_u8(dblimit, d17u8);
+
+  d27s8 = vqsub_s8(vreinterpret_s8_u8(d5u8), vreinterpret_s8_u8(d16u8));
+
+  d22u8 = vorr_u8(d21u8, d22u8);
+
+  q12s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d19u8));
+
+  d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d22u8);
+  d23u8 = vand_u8(d23u8, d17u8);
+
+  q12s16 = vaddw_s8(q12s16, vreinterpret_s8_u8(d27u8));
+
+  d17u8 = vdup_n_u8(4);
+
+  d27s8 = vqmovn_s16(q12s16);
+  d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d23u8);
+  d27s8 = vreinterpret_s8_u8(d27u8);
+
+  d28s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d19u8));
+  d27s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d17u8));
+  d28s8 = vshr_n_s8(d28s8, 3);
+  d27s8 = vshr_n_s8(d27s8, 3);
+
+  d19s8 = vqadd_s8(vreinterpret_s8_u8(d6u8), d28s8);
+  d26s8 = vqsub_s8(vreinterpret_s8_u8(d7u8), d27s8);
+
+  d27s8 = vrshr_n_s8(d27s8, 1);
+  d27s8 = vbic_s8(d27s8, vreinterpret_s8_u8(d22u8));
+
+  d21s8 = vqadd_s8(vreinterpret_s8_u8(d5u8), d27s8);
+  d20s8 = vqsub_s8(vreinterpret_s8_u8(d16u8), d27s8);
+
+  *d4ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d18u8);
+  *d5ru8 = veor_u8(vreinterpret_u8_s8(d19s8), d18u8);
+  *d6ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d18u8);
+  *d7ru8 = veor_u8(vreinterpret_u8_s8(d20s8), d18u8);
+  return;
+}
+
+void aom_lpf_horizontal_4_neon(uint8_t *src, int pitch, const uint8_t *blimit,
+                               const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+  uint8_t *s, *psrc;
+  uint8x8_t dblimit, dlimit, dthresh;
+  uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
+
+  dblimit = vld1_u8(blimit);
+  dlimit = vld1_u8(limit);
+  dthresh = vld1_u8(thresh);
+
+  psrc = src - (pitch << 2);
+  for (i = 0; i < 1; i++) {
+    s = psrc + i * 8;
+
+    d3u8 = vld1_u8(s);
+    s += pitch;
+    d4u8 = vld1_u8(s);
+    s += pitch;
+    d5u8 = vld1_u8(s);
+    s += pitch;
+    d6u8 = vld1_u8(s);
+    s += pitch;
+    d7u8 = vld1_u8(s);
+    s += pitch;
+    d16u8 = vld1_u8(s);
+    s += pitch;
+    d17u8 = vld1_u8(s);
+    s += pitch;
+    d18u8 = vld1_u8(s);
+
+    loop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
+                     d16u8, d17u8, d18u8, &d4u8, &d5u8, &d6u8, &d7u8);
+
+    s -= (pitch * 5);
+    vst1_u8(s, d4u8);
+    s += pitch;
+    vst1_u8(s, d5u8);
+    s += pitch;
+    vst1_u8(s, d6u8);
+    s += pitch;
+    vst1_u8(s, d7u8);
+  }
+  return;
+}
+
+void aom_lpf_vertical_4_neon(uint8_t *src, int pitch, const uint8_t *blimit,
+                             const uint8_t *limit, const uint8_t *thresh) {
+  int i, pitch8;
+  uint8_t *s;
+  uint8x8_t dblimit, dlimit, dthresh;
+  uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
+  uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
+  uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
+  uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
+  uint8x8x4_t d4Result;
+
+  dblimit = vld1_u8(blimit);
+  dlimit = vld1_u8(limit);
+  dthresh = vld1_u8(thresh);
+
+  pitch8 = pitch * 8;
+  for (i = 0; i < 1; i++, src += pitch8) {
+    s = src - (i + 1) * 4;
+
+    d3u8 = vld1_u8(s);
+    s += pitch;
+    d4u8 = vld1_u8(s);
+    s += pitch;
+    d5u8 = vld1_u8(s);
+    s += pitch;
+    d6u8 = vld1_u8(s);
+    s += pitch;
+    d7u8 = vld1_u8(s);
+    s += pitch;
+    d16u8 = vld1_u8(s);
+    s += pitch;
+    d17u8 = vld1_u8(s);
+    s += pitch;
+    d18u8 = vld1_u8(s);
+
+    d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8), vreinterpret_u32_u8(d7u8));
+    d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8), vreinterpret_u32_u8(d16u8));
+    d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8), vreinterpret_u32_u8(d17u8));
+    d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8), vreinterpret_u32_u8(d18u8));
+
+    d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
+                      vreinterpret_u16_u32(d2tmp2.val[0]));
+    d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
+                      vreinterpret_u16_u32(d2tmp3.val[0]));
+    d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
+                      vreinterpret_u16_u32(d2tmp2.val[1]));
+    d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
+                      vreinterpret_u16_u32(d2tmp3.val[1]));
+
+    d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
+                     vreinterpret_u8_u16(d2tmp5.val[0]));
+    d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
+                     vreinterpret_u8_u16(d2tmp5.val[1]));
+    d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
+                      vreinterpret_u8_u16(d2tmp7.val[0]));
+    d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
+                      vreinterpret_u8_u16(d2tmp7.val[1]));
+
+    d3u8 = d2tmp8.val[0];
+    d4u8 = d2tmp8.val[1];
+    d5u8 = d2tmp9.val[0];
+    d6u8 = d2tmp9.val[1];
+    d7u8 = d2tmp10.val[0];
+    d16u8 = d2tmp10.val[1];
+    d17u8 = d2tmp11.val[0];
+    d18u8 = d2tmp11.val[1];
+
+    loop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
+                     d16u8, d17u8, d18u8, &d4u8, &d5u8, &d6u8, &d7u8);
+
+    d4Result.val[0] = d4u8;
+    d4Result.val[1] = d5u8;
+    d4Result.val[2] = d6u8;
+    d4Result.val[3] = d7u8;
+
+    src -= 2;
+    vst4_lane_u8(src, d4Result, 0);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 1);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 2);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 3);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 4);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 5);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 6);
+    src += pitch;
+    vst4_lane_u8(src, d4Result, 7);
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/loopfilter_8_neon.asm b/third_party/aom/aom_dsp/arm/loopfilter_8_neon.asm
new file mode 100644
index 0000000000..9f3db66ee0
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/loopfilter_8_neon.asm
@@ -0,0 +1,428 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_lpf_horizontal_8_neon|
+    EXPORT  |aom_lpf_vertical_8_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; Currently aom only works on iterations 8 at a time. The aom loop filter
+; works on 16 iterations at a time.
+;
+; void aom_lpf_horizontal_8_neon(uint8_t *s, int p,
+;                                const uint8_t *blimit,
+;                                const uint8_t *limit,
+;                                const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|aom_lpf_horizontal_8_neon| PROC
+    push        {r4-r5, lr}
+
+    vld1.8      {d0[]}, [r2]               ; duplicate *blimit
+    ldr         r2, [sp, #12]              ; load thresh
+    add         r1, r1, r1                 ; double pitch
+
+    vld1.8      {d1[]}, [r3]               ; duplicate *limit
+    vld1.8      {d2[]}, [r2]               ; duplicate *thresh
+
+    sub         r3, r0, r1, lsl #1         ; move src pointer down by 4 lines
+    add         r2, r3, r1, lsr #1         ; set to 3 lines down
+
+    vld1.u8     {d3}, [r3@64], r1          ; p3
+    vld1.u8     {d4}, [r2@64], r1          ; p2
+    vld1.u8     {d5}, [r3@64], r1          ; p1
+    vld1.u8     {d6}, [r2@64], r1          ; p0
+    vld1.u8     {d7}, [r3@64], r1          ; q0
+    vld1.u8     {d16}, [r2@64], r1         ; q1
+    vld1.u8     {d17}, [r3@64]             ; q2
+    vld1.u8     {d18}, [r2@64], r1         ; q3
+
+    sub         r3, r3, r1, lsl #1
+    sub         r2, r2, r1, lsl #2
+
+    bl          aom_mbloop_filter_neon
+
+    vst1.u8     {d0}, [r2@64], r1          ; store op2
+    vst1.u8     {d1}, [r3@64], r1          ; store op1
+    vst1.u8     {d2}, [r2@64], r1          ; store op0
+    vst1.u8     {d3}, [r3@64], r1          ; store oq0
+    vst1.u8     {d4}, [r2@64], r1          ; store oq1
+    vst1.u8     {d5}, [r3@64], r1          ; store oq2
+
+    pop         {r4-r5, pc}
+
+    ENDP        ; |aom_lpf_horizontal_8_neon|
+
+; void aom_lpf_vertical_8_neon(uint8_t *s,
+;                              int pitch,
+;                              const uint8_t *blimit,
+;                              const uint8_t *limit,
+;                              const uint8_t *thresh)
+;
+; r0    uint8_t *s,
+; r1    int pitch,
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|aom_lpf_vertical_8_neon| PROC
+    push        {r4-r5, lr}
+
+    vld1.8      {d0[]}, [r2]              ; duplicate *blimit
+    vld1.8      {d1[]}, [r3]              ; duplicate *limit
+
+    ldr         r3, [sp, #12]             ; load thresh
+    sub         r2, r0, #4                ; move s pointer down by 4 columns
+
+    vld1.8      {d2[]}, [r3]              ; duplicate *thresh
+
+    vld1.u8     {d3}, [r2], r1             ; load s data
+    vld1.u8     {d4}, [r2], r1
+    vld1.u8     {d5}, [r2], r1
+    vld1.u8     {d6}, [r2], r1
+    vld1.u8     {d7}, [r2], r1
+    vld1.u8     {d16}, [r2], r1
+    vld1.u8     {d17}, [r2], r1
+    vld1.u8     {d18}, [r2]
+
+    ;transpose to 8x16 matrix
+    vtrn.32     d3, d7
+    vtrn.32     d4, d16
+    vtrn.32     d5, d17
+    vtrn.32     d6, d18
+
+    vtrn.16     d3, d5
+    vtrn.16     d4, d6
+    vtrn.16     d7, d17
+    vtrn.16     d16, d18
+
+    vtrn.8      d3, d4
+    vtrn.8      d5, d6
+    vtrn.8      d7, d16
+    vtrn.8      d17, d18
+
+    sub         r2, r0, #3
+    add         r3, r0, #1
+
+    bl          aom_mbloop_filter_neon
+
+    ;store op2, op1, op0, oq0
+    vst4.8      {d0[0], d1[0], d2[0], d3[0]}, [r2], r1
+    vst4.8      {d0[1], d1[1], d2[1], d3[1]}, [r2], r1
+    vst4.8      {d0[2], d1[2], d2[2], d3[2]}, [r2], r1
+    vst4.8      {d0[3], d1[3], d2[3], d3[3]}, [r2], r1
+    vst4.8      {d0[4], d1[4], d2[4], d3[4]}, [r2], r1
+    vst4.8      {d0[5], d1[5], d2[5], d3[5]}, [r2], r1
+    vst4.8      {d0[6], d1[6], d2[6], d3[6]}, [r2], r1
+    vst4.8      {d0[7], d1[7], d2[7], d3[7]}, [r2]
+
+    ;store oq1, oq2
+    vst2.8      {d4[0], d5[0]}, [r3], r1
+    vst2.8      {d4[1], d5[1]}, [r3], r1
+    vst2.8      {d4[2], d5[2]}, [r3], r1
+    vst2.8      {d4[3], d5[3]}, [r3], r1
+    vst2.8      {d4[4], d5[4]}, [r3], r1
+    vst2.8      {d4[5], d5[5]}, [r3], r1
+    vst2.8      {d4[6], d5[6]}, [r3], r1
+    vst2.8      {d4[7], d5[7]}, [r3]
+
+    pop         {r4-r5, pc}
+    ENDP        ; |aom_lpf_vertical_8_neon|
+
+; void aom_mbloop_filter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. The function does not use
+; registers d8-d15.
+;
+; Inputs:
+; r0-r3, r12 PRESERVE
+; d0    blimit
+; d1    limit
+; d2    thresh
+; d3    p3
+; d4    p2
+; d5    p1
+; d6    p0
+; d7    q0
+; d16   q1
+; d17   q2
+; d18   q3
+;
+; Outputs:
+; d0    op2
+; d1    op1
+; d2    op0
+; d3    oq0
+; d4    oq1
+; d5    oq2
+|aom_mbloop_filter_neon| PROC
+    ; filter_mask
+    vabd.u8     d19, d3, d4                ; m1 = abs(p3 - p2)
+    vabd.u8     d20, d4, d5                ; m2 = abs(p2 - p1)
+    vabd.u8     d21, d5, d6                ; m3 = abs(p1 - p0)
+    vabd.u8     d22, d16, d7               ; m4 = abs(q1 - q0)
+    vabd.u8     d23, d17, d16              ; m5 = abs(q2 - q1)
+    vabd.u8     d24, d18, d17              ; m6 = abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20              ; m1 = max(m1, m2)
+    vmax.u8     d20, d21, d22              ; m2 = max(m3, m4)
+
+    vabd.u8     d25, d6, d4                ; m7 = abs(p0 - p2)
+
+    vmax.u8     d23, d23, d24              ; m3 = max(m5, m6)
+
+    vabd.u8     d26, d7, d17               ; m8 = abs(q0 - q2)
+
+    vmax.u8     d19, d19, d20
+
+    vabd.u8     d24, d6, d7                ; m9 = abs(p0 - q0)
+    vabd.u8     d27, d3, d6                ; m10 = abs(p3 - p0)
+    vabd.u8     d28, d18, d7               ; m11 = abs(q3 - q0)
+
+    vmax.u8     d19, d19, d23
+
+    vabd.u8     d23, d5, d16               ; a = abs(p1 - q1)
+    vqadd.u8    d24, d24, d24              ; b = abs(p0 - q0) * 2
+
+    ; abs () > limit
+    vcge.u8     d19, d1, d19
+
+    ; only compare the largest value to thresh
+    vmax.u8     d25, d25, d26              ; m4 = max(m7, m8)
+    vmax.u8     d26, d27, d28              ; m5 = max(m10, m11)
+
+    vshr.u8     d23, d23, #1               ; a = a / 2
+
+    vmax.u8     d25, d25, d26              ; m4 = max(m4, m5)
+
+    vqadd.u8    d24, d24, d23              ; a = b + a
+
+    vmax.u8     d20, d20, d25              ; m2 = max(m2, m4)
+
+    vmov.u8     d23, #1
+    vcge.u8     d24, d0, d24               ; a > blimit
+
+    vcgt.u8     d21, d21, d2               ; (abs(p1 - p0) > thresh)*-1
+
+    vcge.u8     d20, d23, d20              ; flat
+
+    vand        d19, d19, d24              ; mask
+
+    vcgt.u8     d23, d22, d2               ; (abs(q1 - q0) > thresh)*-1
+
+    vand        d20, d20, d19              ; flat & mask
+
+    vmov.u8     d22, #0x80
+
+    vorr        d23, d21, d23              ; hev
+
+    ; This instruction will truncate the "flat & mask" masks down to 4 bits
+    ; each to fit into one 32 bit arm register. The values are stored in
+    ; q10.64[0].
+    vshrn.u16   d30, q10, #4
+    vmov.u32    r4, d30[0]                 ; flat & mask 4bits
+
+    adds        r5, r4, #1                 ; Check for all 1's
+
+    ; If mask and flat are 1's for all vectors, then we only need to execute
+    ; the power branch for all vectors.
+    beq         power_branch_only
+
+    cmp         r4, #0                     ; Check for 0, set flag for later
+
+    ; mbfilter() function
+    ; filter() function
+    ; convert to signed
+    veor        d21, d7, d22               ; qs0
+    veor        d24, d6, d22               ; ps0
+    veor        d25, d5, d22               ; ps1
+    veor        d26, d16, d22              ; qs1
+
+    vmov.u8     d27, #3
+
+    vsub.s8     d28, d21, d24              ; ( qs0 - ps0)
+
+    vqsub.s8    d29, d25, d26              ; filter = clamp(ps1-qs1)
+
+    vmull.s8    q15, d28, d27              ; 3 * ( qs0 - ps0)
+
+    vand        d29, d29, d23              ; filter &= hev
+
+    vaddw.s8    q15, q15, d29              ; filter + 3 * (qs0 - ps0)
+
+    vmov.u8     d29, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d28, q15
+
+    vand        d28, d28, d19              ; filter &= mask
+
+    vqadd.s8    d30, d28, d27              ; filter2 = clamp(filter+3)
+    vqadd.s8    d29, d28, d29              ; filter1 = clamp(filter+4)
+    vshr.s8     d30, d30, #3               ; filter2 >>= 3
+    vshr.s8     d29, d29, #3               ; filter1 >>= 3
+
+    vqadd.s8    d24, d24, d30              ; op0 = clamp(ps0 + filter2)
+    vqsub.s8    d21, d21, d29              ; oq0 = clamp(qs0 - filter1)
+
+    ; outer tap adjustments: ++filter1 >> 1
+    vrshr.s8    d29, d29, #1
+    vbic        d29, d29, d23              ; filter &= ~hev
+
+    vqadd.s8    d25, d25, d29              ; op1 = clamp(ps1 + filter)
+    vqsub.s8    d26, d26, d29              ; oq1 = clamp(qs1 - filter)
+
+    ; If mask and flat are 0's for all vectors, then we only need to execute
+    ; the filter branch for all vectors.
+    beq         filter_branch_only
+
+    ; If mask and flat are mixed then we must perform both branches and
+    ; combine the data.
+    veor        d24, d24, d22              ; *f_op0 = u^0x80
+    veor        d21, d21, d22              ; *f_oq0 = u^0x80
+    veor        d25, d25, d22              ; *f_op1 = u^0x80
+    veor        d26, d26, d22              ; *f_oq1 = u^0x80
+
+    ; At this point we have already executed the filter branch. The filter
+    ; branch does not set op2 or oq2, so use p2 and q2. Execute the power
+    ; branch and combine the data.
+    vmov.u8     d23, #2
+    vaddl.u8    q14, d6, d7                ; r_op2 = p0 + q0
+    vmlal.u8    q14, d3, d27               ; r_op2 += p3 * 3
+    vmlal.u8    q14, d4, d23               ; r_op2 += p2 * 2
+
+    vbif        d0, d4, d20                ; op2 |= p2 & ~(flat & mask)
+
+    vaddw.u8    q14, d5                    ; r_op2 += p1
+
+    vbif        d1, d25, d20               ; op1 |= f_op1 & ~(flat & mask)
+
+    vqrshrn.u16 d30, q14, #3               ; r_op2
+
+    vsubw.u8    q14, d3                    ; r_op1 = r_op2 - p3
+    vsubw.u8    q14, d4                    ; r_op1 -= p2
+    vaddw.u8    q14, d5                    ; r_op1 += p1
+    vaddw.u8    q14, d16                   ; r_op1 += q1
+
+    vbif        d2, d24, d20               ; op0 |= f_op0 & ~(flat & mask)
+
+    vqrshrn.u16 d31, q14, #3               ; r_op1
+
+    vsubw.u8    q14, d3                    ; r_op0 = r_op1 - p3
+    vsubw.u8    q14, d5                    ; r_op0 -= p1
+    vaddw.u8    q14, d6                    ; r_op0 += p0
+    vaddw.u8    q14, d17                   ; r_op0 += q2
+
+    vbit        d0, d30, d20               ; op2 |= r_op2 & (flat & mask)
+
+    vqrshrn.u16 d23, q14, #3               ; r_op0
+
+    vsubw.u8    q14, d3                    ; r_oq0 = r_op0 - p3
+    vsubw.u8    q14, d6                    ; r_oq0 -= p0
+    vaddw.u8    q14, d7                    ; r_oq0 += q0
+
+    vbit        d1, d31, d20               ; op1 |= r_op1 & (flat & mask)
+
+    vaddw.u8    q14, d18                   ; oq0 += q3
+
+    vbit        d2, d23, d20               ; op0 |= r_op0 & (flat & mask)
+
+    vqrshrn.u16 d22, q14, #3               ; r_oq0
+
+    vsubw.u8    q14, d4                    ; r_oq1 = r_oq0 - p2
+    vsubw.u8    q14, d7                    ; r_oq1 -= q0
+    vaddw.u8    q14, d16                   ; r_oq1 += q1
+
+    vbif        d3, d21, d20               ; oq0 |= f_oq0 & ~(flat & mask)
+
+    vaddw.u8    q14, d18                   ; r_oq1 += q3
+
+    vbif        d4, d26, d20               ; oq1 |= f_oq1 & ~(flat & mask)
+
+    vqrshrn.u16 d6, q14, #3                ; r_oq1
+
+    vsubw.u8    q14, d5                    ; r_oq2 = r_oq1 - p1
+    vsubw.u8    q14, d16                   ; r_oq2 -= q1
+    vaddw.u8    q14, d17                   ; r_oq2 += q2
+    vaddw.u8    q14, d18                   ; r_oq2 += q3
+
+    vbif        d5, d17, d20               ; oq2 |= q2 & ~(flat & mask)
+
+    vqrshrn.u16 d7, q14, #3                ; r_oq2
+
+    vbit        d3, d22, d20               ; oq0 |= r_oq0 & (flat & mask)
+    vbit        d4, d6, d20                ; oq1 |= r_oq1 & (flat & mask)
+    vbit        d5, d7, d20                ; oq2 |= r_oq2 & (flat & mask)
+
+    bx          lr
+
+power_branch_only
+    vmov.u8     d27, #3
+    vmov.u8     d21, #2
+    vaddl.u8    q14, d6, d7                ; op2 = p0 + q0
+    vmlal.u8    q14, d3, d27               ; op2 += p3 * 3
+    vmlal.u8    q14, d4, d21               ; op2 += p2 * 2
+    vaddw.u8    q14, d5                    ; op2 += p1
+    vqrshrn.u16 d0, q14, #3                ; op2
+
+    vsubw.u8    q14, d3                    ; op1 = op2 - p3
+    vsubw.u8    q14, d4                    ; op1 -= p2
+    vaddw.u8    q14, d5                    ; op1 += p1
+    vaddw.u8    q14, d16                   ; op1 += q1
+    vqrshrn.u16 d1, q14, #3                ; op1
+
+    vsubw.u8    q14, d3                    ; op0 = op1 - p3
+    vsubw.u8    q14, d5                    ; op0 -= p1
+    vaddw.u8    q14, d6                    ; op0 += p0
+    vaddw.u8    q14, d17                   ; op0 += q2
+    vqrshrn.u16 d2, q14, #3                ; op0
+
+    vsubw.u8    q14, d3                    ; oq0 = op0 - p3
+    vsubw.u8    q14, d6                    ; oq0 -= p0
+    vaddw.u8    q14, d7                    ; oq0 += q0
+    vaddw.u8    q14, d18                   ; oq0 += q3
+    vqrshrn.u16 d3, q14, #3                ; oq0
+
+    vsubw.u8    q14, d4                    ; oq1 = oq0 - p2
+    vsubw.u8    q14, d7                    ; oq1 -= q0
+    vaddw.u8    q14, d16                   ; oq1 += q1
+    vaddw.u8    q14, d18                   ; oq1 += q3
+    vqrshrn.u16 d4, q14, #3                ; oq1
+
+    vsubw.u8    q14, d5                    ; oq2 = oq1 - p1
+    vsubw.u8    q14, d16                   ; oq2 -= q1
+    vaddw.u8    q14, d17                   ; oq2 += q2
+    vaddw.u8    q14, d18                   ; oq2 += q3
+    vqrshrn.u16 d5, q14, #3                ; oq2
+
+    bx          lr
+
+filter_branch_only
+    ; TODO(fgalligan): See if we can rearange registers so we do not need to
+    ; do the 2 vswp.
+    vswp        d0, d4                      ; op2
+    vswp        d5, d17                     ; oq2
+    veor        d2, d24, d22                ; *op0 = u^0x80
+    veor        d3, d21, d22                ; *oq0 = u^0x80
+    veor        d1, d25, d22                ; *op1 = u^0x80
+    veor        d4, d26, d22                ; *oq1 = u^0x80
+
+    bx          lr
+
+    ENDP        ; |aom_mbloop_filter_neon|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/loopfilter_8_neon.c b/third_party/aom/aom_dsp/arm/loopfilter_8_neon.c
new file mode 100644
index 0000000000..c4502fdb5f
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/loopfilter_8_neon.c
@@ -0,0 +1,430 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+
+static INLINE void mbloop_filter_neon(uint8x8_t dblimit,   // mblimit
+                                      uint8x8_t dlimit,    // limit
+                                      uint8x8_t dthresh,   // thresh
+                                      uint8x8_t d3u8,      // p2
+                                      uint8x8_t d4u8,      // p2
+                                      uint8x8_t d5u8,      // p1
+                                      uint8x8_t d6u8,      // p0
+                                      uint8x8_t d7u8,      // q0
+                                      uint8x8_t d16u8,     // q1
+                                      uint8x8_t d17u8,     // q2
+                                      uint8x8_t d18u8,     // q3
+                                      uint8x8_t *d0ru8,    // p1
+                                      uint8x8_t *d1ru8,    // p1
+                                      uint8x8_t *d2ru8,    // p0
+                                      uint8x8_t *d3ru8,    // q0
+                                      uint8x8_t *d4ru8,    // q1
+                                      uint8x8_t *d5ru8) {  // q1
+  uint32_t flat;
+  uint8x8_t d0u8, d1u8, d2u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8;
+  uint8x8_t d25u8, d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
+  int16x8_t q15s16;
+  uint16x8_t q10u16, q14u16;
+  int8x8_t d21s8, d24s8, d25s8, d26s8, d28s8, d29s8, d30s8;
+
+  d19u8 = vabd_u8(d3u8, d4u8);
+  d20u8 = vabd_u8(d4u8, d5u8);
+  d21u8 = vabd_u8(d5u8, d6u8);
+  d22u8 = vabd_u8(d16u8, d7u8);
+  d23u8 = vabd_u8(d17u8, d16u8);
+  d24u8 = vabd_u8(d18u8, d17u8);
+
+  d19u8 = vmax_u8(d19u8, d20u8);
+  d20u8 = vmax_u8(d21u8, d22u8);
+
+  d25u8 = vabd_u8(d6u8, d4u8);
+
+  d23u8 = vmax_u8(d23u8, d24u8);
+
+  d26u8 = vabd_u8(d7u8, d17u8);
+
+  d19u8 = vmax_u8(d19u8, d20u8);
+
+  d24u8 = vabd_u8(d6u8, d7u8);
+  d27u8 = vabd_u8(d3u8, d6u8);
+  d28u8 = vabd_u8(d18u8, d7u8);
+
+  d19u8 = vmax_u8(d19u8, d23u8);
+
+  d23u8 = vabd_u8(d5u8, d16u8);
+  d24u8 = vqadd_u8(d24u8, d24u8);
+
+  d19u8 = vcge_u8(dlimit, d19u8);
+
+  d25u8 = vmax_u8(d25u8, d26u8);
+  d26u8 = vmax_u8(d27u8, d28u8);
+
+  d23u8 = vshr_n_u8(d23u8, 1);
+
+  d25u8 = vmax_u8(d25u8, d26u8);
+
+  d24u8 = vqadd_u8(d24u8, d23u8);
+
+  d20u8 = vmax_u8(d20u8, d25u8);
+
+  d23u8 = vdup_n_u8(1);
+  d24u8 = vcge_u8(dblimit, d24u8);
+
+  d21u8 = vcgt_u8(d21u8, dthresh);
+
+  d20u8 = vcge_u8(d23u8, d20u8);
+
+  d19u8 = vand_u8(d19u8, d24u8);
+
+  d23u8 = vcgt_u8(d22u8, dthresh);
+
+  d20u8 = vand_u8(d20u8, d19u8);
+
+  d22u8 = vdup_n_u8(0x80);
+
+  d23u8 = vorr_u8(d21u8, d23u8);
+
+  q10u16 = vcombine_u16(vreinterpret_u16_u8(d20u8), vreinterpret_u16_u8(d21u8));
+
+  d30u8 = vshrn_n_u16(q10u16, 4);
+  flat = vget_lane_u32(vreinterpret_u32_u8(d30u8), 0);
+
+  if (flat == 0xffffffff) {  // Check for all 1's, power_branch_only
+    d27u8 = vdup_n_u8(3);
+    d21u8 = vdup_n_u8(2);
+    q14u16 = vaddl_u8(d6u8, d7u8);
+    q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
+    q14u16 = vmlal_u8(q14u16, d4u8, d21u8);
+    q14u16 = vaddw_u8(q14u16, d5u8);
+    *d0ru8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d4u8);
+    q14u16 = vaddw_u8(q14u16, d5u8);
+    q14u16 = vaddw_u8(q14u16, d16u8);
+    *d1ru8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d5u8);
+    q14u16 = vaddw_u8(q14u16, d6u8);
+    q14u16 = vaddw_u8(q14u16, d17u8);
+    *d2ru8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d6u8);
+    q14u16 = vaddw_u8(q14u16, d7u8);
+    q14u16 = vaddw_u8(q14u16, d18u8);
+    *d3ru8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d4u8);
+    q14u16 = vsubw_u8(q14u16, d7u8);
+    q14u16 = vaddw_u8(q14u16, d16u8);
+    q14u16 = vaddw_u8(q14u16, d18u8);
+    *d4ru8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d5u8);
+    q14u16 = vsubw_u8(q14u16, d16u8);
+    q14u16 = vaddw_u8(q14u16, d17u8);
+    q14u16 = vaddw_u8(q14u16, d18u8);
+    *d5ru8 = vqrshrn_n_u16(q14u16, 3);
+  } else {
+    d21u8 = veor_u8(d7u8, d22u8);
+    d24u8 = veor_u8(d6u8, d22u8);
+    d25u8 = veor_u8(d5u8, d22u8);
+    d26u8 = veor_u8(d16u8, d22u8);
+
+    d27u8 = vdup_n_u8(3);
+
+    d28s8 = vsub_s8(vreinterpret_s8_u8(d21u8), vreinterpret_s8_u8(d24u8));
+    d29s8 = vqsub_s8(vreinterpret_s8_u8(d25u8), vreinterpret_s8_u8(d26u8));
+
+    q15s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d27u8));
+
+    d29s8 = vand_s8(d29s8, vreinterpret_s8_u8(d23u8));
+
+    q15s16 = vaddw_s8(q15s16, d29s8);
+
+    d29u8 = vdup_n_u8(4);
+
+    d28s8 = vqmovn_s16(q15s16);
+
+    d28s8 = vand_s8(d28s8, vreinterpret_s8_u8(d19u8));
+
+    d30s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d27u8));
+    d29s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d29u8));
+    d30s8 = vshr_n_s8(d30s8, 3);
+    d29s8 = vshr_n_s8(d29s8, 3);
+
+    d24s8 = vqadd_s8(vreinterpret_s8_u8(d24u8), d30s8);
+    d21s8 = vqsub_s8(vreinterpret_s8_u8(d21u8), d29s8);
+
+    d29s8 = vrshr_n_s8(d29s8, 1);
+    d29s8 = vbic_s8(d29s8, vreinterpret_s8_u8(d23u8));
+
+    d25s8 = vqadd_s8(vreinterpret_s8_u8(d25u8), d29s8);
+    d26s8 = vqsub_s8(vreinterpret_s8_u8(d26u8), d29s8);
+
+    if (flat == 0) {  // filter_branch_only
+      *d0ru8 = d4u8;
+      *d1ru8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
+      *d2ru8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
+      *d3ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
+      *d4ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
+      *d5ru8 = d17u8;
+      return;
+    }
+
+    d21u8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
+    d24u8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
+    d25u8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
+    d26u8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
+
+    d23u8 = vdup_n_u8(2);
+    q14u16 = vaddl_u8(d6u8, d7u8);
+    q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
+    q14u16 = vmlal_u8(q14u16, d4u8, d23u8);
+
+    d0u8 = vbsl_u8(d20u8, dblimit, d4u8);
+
+    q14u16 = vaddw_u8(q14u16, d5u8);
+
+    d1u8 = vbsl_u8(d20u8, dlimit, d25u8);
+
+    d30u8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d4u8);
+    q14u16 = vaddw_u8(q14u16, d5u8);
+    q14u16 = vaddw_u8(q14u16, d16u8);
+
+    d2u8 = vbsl_u8(d20u8, dthresh, d24u8);
+
+    d31u8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d5u8);
+    q14u16 = vaddw_u8(q14u16, d6u8);
+    q14u16 = vaddw_u8(q14u16, d17u8);
+
+    *d0ru8 = vbsl_u8(d20u8, d30u8, d0u8);
+
+    d23u8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d3u8);
+    q14u16 = vsubw_u8(q14u16, d6u8);
+    q14u16 = vaddw_u8(q14u16, d7u8);
+
+    *d1ru8 = vbsl_u8(d20u8, d31u8, d1u8);
+
+    q14u16 = vaddw_u8(q14u16, d18u8);
+
+    *d2ru8 = vbsl_u8(d20u8, d23u8, d2u8);
+
+    d22u8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d4u8);
+    q14u16 = vsubw_u8(q14u16, d7u8);
+    q14u16 = vaddw_u8(q14u16, d16u8);
+
+    d3u8 = vbsl_u8(d20u8, d3u8, d21u8);
+
+    q14u16 = vaddw_u8(q14u16, d18u8);
+
+    d4u8 = vbsl_u8(d20u8, d4u8, d26u8);
+
+    d6u8 = vqrshrn_n_u16(q14u16, 3);
+
+    q14u16 = vsubw_u8(q14u16, d5u8);
+    q14u16 = vsubw_u8(q14u16, d16u8);
+    q14u16 = vaddw_u8(q14u16, d17u8);
+    q14u16 = vaddw_u8(q14u16, d18u8);
+
+    d5u8 = vbsl_u8(d20u8, d5u8, d17u8);
+
+    d7u8 = vqrshrn_n_u16(q14u16, 3);
+
+    *d3ru8 = vbsl_u8(d20u8, d22u8, d3u8);
+    *d4ru8 = vbsl_u8(d20u8, d6u8, d4u8);
+    *d5ru8 = vbsl_u8(d20u8, d7u8, d5u8);
+  }
+  return;
+}
+
+void aom_lpf_horizontal_8_neon(uint8_t *src, int pitch, const uint8_t *blimit,
+                               const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+  uint8_t *s, *psrc;
+  uint8x8_t dblimit, dlimit, dthresh;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+  uint8x8_t d16u8, d17u8, d18u8;
+
+  dblimit = vld1_u8(blimit);
+  dlimit = vld1_u8(limit);
+  dthresh = vld1_u8(thresh);
+
+  psrc = src - (pitch << 2);
+  for (i = 0; i < 1; i++) {
+    s = psrc + i * 8;
+
+    d3u8 = vld1_u8(s);
+    s += pitch;
+    d4u8 = vld1_u8(s);
+    s += pitch;
+    d5u8 = vld1_u8(s);
+    s += pitch;
+    d6u8 = vld1_u8(s);
+    s += pitch;
+    d7u8 = vld1_u8(s);
+    s += pitch;
+    d16u8 = vld1_u8(s);
+    s += pitch;
+    d17u8 = vld1_u8(s);
+    s += pitch;
+    d18u8 = vld1_u8(s);
+
+    mbloop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
+                       d16u8, d17u8, d18u8, &d0u8, &d1u8, &d2u8, &d3u8, &d4u8,
+                       &d5u8);
+
+    s -= (pitch * 6);
+    vst1_u8(s, d0u8);
+    s += pitch;
+    vst1_u8(s, d1u8);
+    s += pitch;
+    vst1_u8(s, d2u8);
+    s += pitch;
+    vst1_u8(s, d3u8);
+    s += pitch;
+    vst1_u8(s, d4u8);
+    s += pitch;
+    vst1_u8(s, d5u8);
+  }
+  return;
+}
+
+void aom_lpf_vertical_8_neon(uint8_t *src, int pitch, const uint8_t *blimit,
+                             const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+  uint8_t *s;
+  uint8x8_t dblimit, dlimit, dthresh;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+  uint8x8_t d16u8, d17u8, d18u8;
+  uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
+  uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
+  uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
+  uint8x8x4_t d4Result;
+  uint8x8x2_t d2Result;
+
+  dblimit = vld1_u8(blimit);
+  dlimit = vld1_u8(limit);
+  dthresh = vld1_u8(thresh);
+
+  for (i = 0; i < 1; i++) {
+    s = src + (i * (pitch << 3)) - 4;
+
+    d3u8 = vld1_u8(s);
+    s += pitch;
+    d4u8 = vld1_u8(s);
+    s += pitch;
+    d5u8 = vld1_u8(s);
+    s += pitch;
+    d6u8 = vld1_u8(s);
+    s += pitch;
+    d7u8 = vld1_u8(s);
+    s += pitch;
+    d16u8 = vld1_u8(s);
+    s += pitch;
+    d17u8 = vld1_u8(s);
+    s += pitch;
+    d18u8 = vld1_u8(s);
+
+    d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8), vreinterpret_u32_u8(d7u8));
+    d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8), vreinterpret_u32_u8(d16u8));
+    d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8), vreinterpret_u32_u8(d17u8));
+    d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8), vreinterpret_u32_u8(d18u8));
+
+    d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
+                      vreinterpret_u16_u32(d2tmp2.val[0]));
+    d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
+                      vreinterpret_u16_u32(d2tmp3.val[0]));
+    d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
+                      vreinterpret_u16_u32(d2tmp2.val[1]));
+    d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
+                      vreinterpret_u16_u32(d2tmp3.val[1]));
+
+    d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
+                     vreinterpret_u8_u16(d2tmp5.val[0]));
+    d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
+                     vreinterpret_u8_u16(d2tmp5.val[1]));
+    d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
+                      vreinterpret_u8_u16(d2tmp7.val[0]));
+    d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
+                      vreinterpret_u8_u16(d2tmp7.val[1]));
+
+    d3u8 = d2tmp8.val[0];
+    d4u8 = d2tmp8.val[1];
+    d5u8 = d2tmp9.val[0];
+    d6u8 = d2tmp9.val[1];
+    d7u8 = d2tmp10.val[0];
+    d16u8 = d2tmp10.val[1];
+    d17u8 = d2tmp11.val[0];
+    d18u8 = d2tmp11.val[1];
+
+    mbloop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
+                       d16u8, d17u8, d18u8, &d0u8, &d1u8, &d2u8, &d3u8, &d4u8,
+                       &d5u8);
+
+    d4Result.val[0] = d0u8;
+    d4Result.val[1] = d1u8;
+    d4Result.val[2] = d2u8;
+    d4Result.val[3] = d3u8;
+
+    d2Result.val[0] = d4u8;
+    d2Result.val[1] = d5u8;
+
+    s = src - 3;
+    vst4_lane_u8(s, d4Result, 0);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 1);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 2);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 3);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 4);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 5);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 6);
+    s += pitch;
+    vst4_lane_u8(s, d4Result, 7);
+
+    s = src + 1;
+    vst2_lane_u8(s, d2Result, 0);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 1);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 2);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 3);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 4);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 5);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 6);
+    s += pitch;
+    vst2_lane_u8(s, d2Result, 7);
+  }
+  return;
+}
diff --git a/third_party/aom/aom_dsp/arm/loopfilter_mb_neon.asm b/third_party/aom/aom_dsp/arm/loopfilter_mb_neon.asm
new file mode 100644
index 0000000000..675928860f
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/loopfilter_mb_neon.asm
@@ -0,0 +1,638 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+    EXPORT  |aom_lpf_horizontal_edge_8_neon|
+    EXPORT  |aom_lpf_horizontal_edge_16_neon|
+    EXPORT  |aom_lpf_vertical_16_neon|
+    ARM
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; void mb_lpf_horizontal_edge(uint8_t *s, int p,
+;                             const uint8_t *blimit,
+;                             const uint8_t *limit,
+;                             const uint8_t *thresh,
+;                             int count)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+; r12   int count
+|mb_lpf_horizontal_edge| PROC
+    push        {r4-r8, lr}
+    vpush       {d8-d15}
+    ldr         r4, [sp, #88]              ; load thresh
+
+h_count
+    vld1.8      {d16[]}, [r2]              ; load *blimit
+    vld1.8      {d17[]}, [r3]              ; load *limit
+    vld1.8      {d18[]}, [r4]              ; load *thresh
+
+    sub         r8, r0, r1, lsl #3         ; move src pointer down by 8 lines
+
+    vld1.u8     {d0}, [r8@64], r1          ; p7
+    vld1.u8     {d1}, [r8@64], r1          ; p6
+    vld1.u8     {d2}, [r8@64], r1          ; p5
+    vld1.u8     {d3}, [r8@64], r1          ; p4
+    vld1.u8     {d4}, [r8@64], r1          ; p3
+    vld1.u8     {d5}, [r8@64], r1          ; p2
+    vld1.u8     {d6}, [r8@64], r1          ; p1
+    vld1.u8     {d7}, [r8@64], r1          ; p0
+    vld1.u8     {d8}, [r8@64], r1          ; q0
+    vld1.u8     {d9}, [r8@64], r1          ; q1
+    vld1.u8     {d10}, [r8@64], r1         ; q2
+    vld1.u8     {d11}, [r8@64], r1         ; q3
+    vld1.u8     {d12}, [r8@64], r1         ; q4
+    vld1.u8     {d13}, [r8@64], r1         ; q5
+    vld1.u8     {d14}, [r8@64], r1         ; q6
+    vld1.u8     {d15}, [r8@64], r1         ; q7
+
+    bl          aom_wide_mbfilter_neon
+
+    tst         r7, #1
+    beq         h_mbfilter
+
+    ; flat && mask were not set for any of the channels. Just store the values
+    ; from filter.
+    sub         r8, r0, r1, lsl #1
+
+    vst1.u8     {d25}, [r8@64], r1         ; store op1
+    vst1.u8     {d24}, [r8@64], r1         ; store op0
+    vst1.u8     {d23}, [r8@64], r1         ; store oq0
+    vst1.u8     {d26}, [r8@64], r1         ; store oq1
+
+    b           h_next
+
+h_mbfilter
+    tst         r7, #2
+    beq         h_wide_mbfilter
+
+    ; flat2 was not set for any of the channels. Just store the values from
+    ; mbfilter.
+    sub         r8, r0, r1, lsl #1
+    sub         r8, r8, r1
+
+    vst1.u8     {d18}, [r8@64], r1         ; store op2
+    vst1.u8     {d19}, [r8@64], r1         ; store op1
+    vst1.u8     {d20}, [r8@64], r1         ; store op0
+    vst1.u8     {d21}, [r8@64], r1         ; store oq0
+    vst1.u8     {d22}, [r8@64], r1         ; store oq1
+    vst1.u8     {d23}, [r8@64], r1         ; store oq2
+
+    b           h_next
+
+h_wide_mbfilter
+    sub         r8, r0, r1, lsl #3
+    add         r8, r8, r1
+
+    vst1.u8     {d16}, [r8@64], r1         ; store op6
+    vst1.u8     {d24}, [r8@64], r1         ; store op5
+    vst1.u8     {d25}, [r8@64], r1         ; store op4
+    vst1.u8     {d26}, [r8@64], r1         ; store op3
+    vst1.u8     {d27}, [r8@64], r1         ; store op2
+    vst1.u8     {d18}, [r8@64], r1         ; store op1
+    vst1.u8     {d19}, [r8@64], r1         ; store op0
+    vst1.u8     {d20}, [r8@64], r1         ; store oq0
+    vst1.u8     {d21}, [r8@64], r1         ; store oq1
+    vst1.u8     {d22}, [r8@64], r1         ; store oq2
+    vst1.u8     {d23}, [r8@64], r1         ; store oq3
+    vst1.u8     {d1}, [r8@64], r1          ; store oq4
+    vst1.u8     {d2}, [r8@64], r1          ; store oq5
+    vst1.u8     {d3}, [r8@64], r1          ; store oq6
+
+h_next
+    add         r0, r0, #8
+    subs        r12, r12, #1
+    bne         h_count
+
+    vpop        {d8-d15}
+    pop         {r4-r8, pc}
+
+    ENDP        ; |mb_lpf_horizontal_edge|
+
+; void aom_lpf_horizontal_edge_8_neon(uint8_t *s, int pitch,
+;                                     const uint8_t *blimit,
+;                                     const uint8_t *limit,
+;                                     const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int pitch,
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh
+|aom_lpf_horizontal_edge_8_neon| PROC
+    mov r12, #1
+    b mb_lpf_horizontal_edge
+    ENDP        ; |aom_lpf_horizontal_edge_8_neon|
+
+; void aom_lpf_horizontal_edge_16_neon(uint8_t *s, int pitch,
+;                                      const uint8_t *blimit,
+;                                      const uint8_t *limit,
+;                                      const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int pitch,
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh
+|aom_lpf_horizontal_edge_16_neon| PROC
+    mov r12, #2
+    b mb_lpf_horizontal_edge
+    ENDP        ; |aom_lpf_horizontal_edge_16_neon|
+
+; void aom_lpf_vertical_16_neon(uint8_t *s, int p,
+;                               const uint8_t *blimit,
+;                               const uint8_t *limit,
+;                               const uint8_t *thresh)
+; r0    uint8_t *s,
+; r1    int p, /* pitch */
+; r2    const uint8_t *blimit,
+; r3    const uint8_t *limit,
+; sp    const uint8_t *thresh,
+|aom_lpf_vertical_16_neon| PROC
+    push        {r4-r8, lr}
+    vpush       {d8-d15}
+    ldr         r4, [sp, #88]              ; load thresh
+
+    vld1.8      {d16[]}, [r2]              ; load *blimit
+    vld1.8      {d17[]}, [r3]              ; load *limit
+    vld1.8      {d18[]}, [r4]              ; load *thresh
+
+    sub         r8, r0, #8
+
+    vld1.8      {d0}, [r8@64], r1
+    vld1.8      {d8}, [r0@64], r1
+    vld1.8      {d1}, [r8@64], r1
+    vld1.8      {d9}, [r0@64], r1
+    vld1.8      {d2}, [r8@64], r1
+    vld1.8      {d10}, [r0@64], r1
+    vld1.8      {d3}, [r8@64], r1
+    vld1.8      {d11}, [r0@64], r1
+    vld1.8      {d4}, [r8@64], r1
+    vld1.8      {d12}, [r0@64], r1
+    vld1.8      {d5}, [r8@64], r1
+    vld1.8      {d13}, [r0@64], r1
+    vld1.8      {d6}, [r8@64], r1
+    vld1.8      {d14}, [r0@64], r1
+    vld1.8      {d7}, [r8@64], r1
+    vld1.8      {d15}, [r0@64], r1
+
+    sub         r0, r0, r1, lsl #3
+
+    vtrn.32     q0, q2
+    vtrn.32     q1, q3
+    vtrn.32     q4, q6
+    vtrn.32     q5, q7
+
+    vtrn.16     q0, q1
+    vtrn.16     q2, q3
+    vtrn.16     q4, q5
+    vtrn.16     q6, q7
+
+    vtrn.8      d0, d1
+    vtrn.8      d2, d3
+    vtrn.8      d4, d5
+    vtrn.8      d6, d7
+
+    vtrn.8      d8, d9
+    vtrn.8      d10, d11
+    vtrn.8      d12, d13
+    vtrn.8      d14, d15
+
+    bl          aom_wide_mbfilter_neon
+
+    tst         r7, #1
+    beq         v_mbfilter
+
+    ; flat && mask were not set for any of the channels. Just store the values
+    ; from filter.
+    sub         r8, r0, #2
+
+    vswp        d23, d25
+
+    vst4.8      {d23[0], d24[0], d25[0], d26[0]}, [r8], r1
+    vst4.8      {d23[1], d24[1], d25[1], d26[1]}, [r8], r1
+    vst4.8      {d23[2], d24[2], d25[2], d26[2]}, [r8], r1
+    vst4.8      {d23[3], d24[3], d25[3], d26[3]}, [r8], r1
+    vst4.8      {d23[4], d24[4], d25[4], d26[4]}, [r8], r1
+    vst4.8      {d23[5], d24[5], d25[5], d26[5]}, [r8], r1
+    vst4.8      {d23[6], d24[6], d25[6], d26[6]}, [r8], r1
+    vst4.8      {d23[7], d24[7], d25[7], d26[7]}, [r8], r1
+
+    b           v_end
+
+v_mbfilter
+    tst         r7, #2
+    beq         v_wide_mbfilter
+
+    ; flat2 was not set for any of the channels. Just store the values from
+    ; mbfilter.
+    sub         r8, r0, #3
+
+    vst3.8      {d18[0], d19[0], d20[0]}, [r8], r1
+    vst3.8      {d21[0], d22[0], d23[0]}, [r0], r1
+    vst3.8      {d18[1], d19[1], d20[1]}, [r8], r1
+    vst3.8      {d21[1], d22[1], d23[1]}, [r0], r1
+    vst3.8      {d18[2], d19[2], d20[2]}, [r8], r1
+    vst3.8      {d21[2], d22[2], d23[2]}, [r0], r1
+    vst3.8      {d18[3], d19[3], d20[3]}, [r8], r1
+    vst3.8      {d21[3], d22[3], d23[3]}, [r0], r1
+    vst3.8      {d18[4], d19[4], d20[4]}, [r8], r1
+    vst3.8      {d21[4], d22[4], d23[4]}, [r0], r1
+    vst3.8      {d18[5], d19[5], d20[5]}, [r8], r1
+    vst3.8      {d21[5], d22[5], d23[5]}, [r0], r1
+    vst3.8      {d18[6], d19[6], d20[6]}, [r8], r1
+    vst3.8      {d21[6], d22[6], d23[6]}, [r0], r1
+    vst3.8      {d18[7], d19[7], d20[7]}, [r8], r1
+    vst3.8      {d21[7], d22[7], d23[7]}, [r0], r1
+
+    b           v_end
+
+v_wide_mbfilter
+    sub         r8, r0, #8
+
+    vtrn.32     d0,  d26
+    vtrn.32     d16, d27
+    vtrn.32     d24, d18
+    vtrn.32     d25, d19
+
+    vtrn.16     d0,  d24
+    vtrn.16     d16, d25
+    vtrn.16     d26, d18
+    vtrn.16     d27, d19
+
+    vtrn.8      d0,  d16
+    vtrn.8      d24, d25
+    vtrn.8      d26, d27
+    vtrn.8      d18, d19
+
+    vtrn.32     d20, d1
+    vtrn.32     d21, d2
+    vtrn.32     d22, d3
+    vtrn.32     d23, d15
+
+    vtrn.16     d20, d22
+    vtrn.16     d21, d23
+    vtrn.16     d1,  d3
+    vtrn.16     d2,  d15
+
+    vtrn.8      d20, d21
+    vtrn.8      d22, d23
+    vtrn.8      d1,  d2
+    vtrn.8      d3,  d15
+
+    vst1.8      {d0}, [r8@64], r1
+    vst1.8      {d20}, [r0@64], r1
+    vst1.8      {d16}, [r8@64], r1
+    vst1.8      {d21}, [r0@64], r1
+    vst1.8      {d24}, [r8@64], r1
+    vst1.8      {d22}, [r0@64], r1
+    vst1.8      {d25}, [r8@64], r1
+    vst1.8      {d23}, [r0@64], r1
+    vst1.8      {d26}, [r8@64], r1
+    vst1.8      {d1}, [r0@64], r1
+    vst1.8      {d27}, [r8@64], r1
+    vst1.8      {d2}, [r0@64], r1
+    vst1.8      {d18}, [r8@64], r1
+    vst1.8      {d3}, [r0@64], r1
+    vst1.8      {d19}, [r8@64], r1
+    vst1.8      {d15}, [r0@64], r1
+
+v_end
+    vpop        {d8-d15}
+    pop         {r4-r8, pc}
+
+    ENDP        ; |aom_lpf_vertical_16_neon|
+
+; void aom_wide_mbfilter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store.
+;
+; r0-r3 PRESERVE
+; d16    blimit
+; d17    limit
+; d18    thresh
+; d0    p7
+; d1    p6
+; d2    p5
+; d3    p4
+; d4    p3
+; d5    p2
+; d6    p1
+; d7    p0
+; d8    q0
+; d9    q1
+; d10   q2
+; d11   q3
+; d12   q4
+; d13   q5
+; d14   q6
+; d15   q7
+|aom_wide_mbfilter_neon| PROC
+    mov         r7, #0
+
+    ; filter_mask
+    vabd.u8     d19, d4, d5                ; abs(p3 - p2)
+    vabd.u8     d20, d5, d6                ; abs(p2 - p1)
+    vabd.u8     d21, d6, d7                ; abs(p1 - p0)
+    vabd.u8     d22, d9, d8                ; abs(q1 - q0)
+    vabd.u8     d23, d10, d9               ; abs(q2 - q1)
+    vabd.u8     d24, d11, d10              ; abs(q3 - q2)
+
+    ; only compare the largest value to limit
+    vmax.u8     d19, d19, d20              ; max(abs(p3 - p2), abs(p2 - p1))
+    vmax.u8     d20, d21, d22              ; max(abs(p1 - p0), abs(q1 - q0))
+    vmax.u8     d23, d23, d24              ; max(abs(q2 - q1), abs(q3 - q2))
+    vmax.u8     d19, d19, d20
+
+    vabd.u8     d24, d7, d8                ; abs(p0 - q0)
+
+    vmax.u8     d19, d19, d23
+
+    vabd.u8     d23, d6, d9                ; a = abs(p1 - q1)
+    vqadd.u8    d24, d24, d24              ; b = abs(p0 - q0) * 2
+
+    ; abs () > limit
+    vcge.u8     d19, d17, d19
+
+    ; flatmask4
+    vabd.u8     d25, d7, d5                ; abs(p0 - p2)
+    vabd.u8     d26, d8, d10               ; abs(q0 - q2)
+    vabd.u8     d27, d4, d7                ; abs(p3 - p0)
+    vabd.u8     d28, d11, d8               ; abs(q3 - q0)
+
+    ; only compare the largest value to thresh
+    vmax.u8     d25, d25, d26              ; max(abs(p0 - p2), abs(q0 - q2))
+    vmax.u8     d26, d27, d28              ; max(abs(p3 - p0), abs(q3 - q0))
+    vmax.u8     d25, d25, d26
+    vmax.u8     d20, d20, d25
+
+    vshr.u8     d23, d23, #1               ; a = a / 2
+    vqadd.u8    d24, d24, d23              ; a = b + a
+
+    vmov.u8     d30, #1
+    vcge.u8     d24, d16, d24              ; (a > blimit * 2 + limit) * -1
+
+    vcge.u8     d20, d30, d20              ; flat
+
+    vand        d19, d19, d24              ; mask
+
+    ; hevmask
+    vcgt.u8     d21, d21, d18              ; (abs(p1 - p0) > thresh)*-1
+    vcgt.u8     d22, d22, d18              ; (abs(q1 - q0) > thresh)*-1
+    vorr        d21, d21, d22              ; hev
+
+    vand        d16, d20, d19              ; flat && mask
+    vmov        r5, r6, d16
+
+    ; flatmask5(1, p7, p6, p5, p4, p0, q0, q4, q5, q6, q7)
+    vabd.u8     d22, d3, d7                ; abs(p4 - p0)
+    vabd.u8     d23, d12, d8               ; abs(q4 - q0)
+    vabd.u8     d24, d7, d2                ; abs(p0 - p5)
+    vabd.u8     d25, d8, d13               ; abs(q0 - q5)
+    vabd.u8     d26, d1, d7                ; abs(p6 - p0)
+    vabd.u8     d27, d14, d8               ; abs(q6 - q0)
+    vabd.u8     d28, d0, d7                ; abs(p7 - p0)
+    vabd.u8     d29, d15, d8               ; abs(q7 - q0)
+
+    ; only compare the largest value to thresh
+    vmax.u8     d22, d22, d23              ; max(abs(p4 - p0), abs(q4 - q0))
+    vmax.u8     d23, d24, d25              ; max(abs(p0 - p5), abs(q0 - q5))
+    vmax.u8     d24, d26, d27              ; max(abs(p6 - p0), abs(q6 - q0))
+    vmax.u8     d25, d28, d29              ; max(abs(p7 - p0), abs(q7 - q0))
+
+    vmax.u8     d26, d22, d23
+    vmax.u8     d27, d24, d25
+    vmax.u8     d23, d26, d27
+
+    vcge.u8     d18, d30, d23              ; flat2
+
+    vmov.u8     d22, #0x80
+
+    orrs        r5, r5, r6                 ; Check for 0
+    orreq       r7, r7, #1                 ; Only do filter branch
+
+    vand        d17, d18, d16              ; flat2 && flat && mask
+    vmov        r5, r6, d17
+
+    ; mbfilter() function
+
+    ; filter() function
+    ; convert to signed
+    veor        d23, d8, d22               ; qs0
+    veor        d24, d7, d22               ; ps0
+    veor        d25, d6, d22               ; ps1
+    veor        d26, d9, d22               ; qs1
+
+    vmov.u8     d27, #3
+
+    vsub.s8     d28, d23, d24              ; ( qs0 - ps0)
+    vqsub.s8    d29, d25, d26              ; filter = clamp(ps1-qs1)
+    vmull.s8    q15, d28, d27              ; 3 * ( qs0 - ps0)
+    vand        d29, d29, d21              ; filter &= hev
+    vaddw.s8    q15, q15, d29              ; filter + 3 * (qs0 - ps0)
+    vmov.u8     d29, #4
+
+    ; filter = clamp(filter + 3 * ( qs0 - ps0))
+    vqmovn.s16  d28, q15
+
+    vand        d28, d28, d19              ; filter &= mask
+
+    vqadd.s8    d30, d28, d27              ; filter2 = clamp(filter+3)
+    vqadd.s8    d29, d28, d29              ; filter1 = clamp(filter+4)
+    vshr.s8     d30, d30, #3               ; filter2 >>= 3
+    vshr.s8     d29, d29, #3               ; filter1 >>= 3
+
+
+    vqadd.s8    d24, d24, d30              ; op0 = clamp(ps0 + filter2)
+    vqsub.s8    d23, d23, d29              ; oq0 = clamp(qs0 - filter1)
+
+    ; outer tap adjustments: ++filter1 >> 1
+    vrshr.s8    d29, d29, #1
+    vbic        d29, d29, d21              ; filter &= ~hev
+
+    vqadd.s8    d25, d25, d29              ; op1 = clamp(ps1 + filter)
+    vqsub.s8    d26, d26, d29              ; oq1 = clamp(qs1 - filter)
+
+    veor        d24, d24, d22              ; *f_op0 = u^0x80
+    veor        d23, d23, d22              ; *f_oq0 = u^0x80
+    veor        d25, d25, d22              ; *f_op1 = u^0x80
+    veor        d26, d26, d22              ; *f_oq1 = u^0x80
+
+    tst         r7, #1
+    bxne        lr
+
+    orrs        r5, r5, r6                 ; Check for 0
+    orreq       r7, r7, #2                 ; Only do mbfilter branch
+
+    ; mbfilter flat && mask branch
+    ; TODO(fgalligan): Can I decrease the cycles shifting to consective d's
+    ; and using vibt on the q's?
+    vmov.u8     d29, #2
+    vaddl.u8    q15, d7, d8                ; op2 = p0 + q0
+    vmlal.u8    q15, d4, d27               ; op2 = p0 + q0 + p3 * 3
+    vmlal.u8    q15, d5, d29               ; op2 = p0 + q0 + p3 * 3 + p2 * 2
+    vaddl.u8    q10, d4, d5
+    vaddw.u8    q15, d6                    ; op2=p1 + p0 + q0 + p3 * 3 + p2 *2
+    vaddl.u8    q14, d6, d9
+    vqrshrn.u16 d18, q15, #3               ; r_op2
+
+    vsub.i16    q15, q10
+    vaddl.u8    q10, d4, d6
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d7, d10
+    vqrshrn.u16 d19, q15, #3               ; r_op1
+
+    vsub.i16    q15, q10
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d8, d11
+    vqrshrn.u16 d20, q15, #3               ; r_op0
+
+    vsubw.u8    q15, d4                    ; oq0 = op0 - p3
+    vsubw.u8    q15, d7                    ; oq0 -= p0
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d9, d11
+    vqrshrn.u16 d21, q15, #3               ; r_oq0
+
+    vsubw.u8    q15, d5                    ; oq1 = oq0 - p2
+    vsubw.u8    q15, d8                    ; oq1 -= q0
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d10, d11
+    vqrshrn.u16 d22, q15, #3               ; r_oq1
+
+    vsubw.u8    q15, d6                    ; oq2 = oq0 - p1
+    vsubw.u8    q15, d9                    ; oq2 -= q1
+    vadd.i16    q15, q14
+    vqrshrn.u16 d27, q15, #3               ; r_oq2
+
+    ; Filter does not set op2 or oq2, so use p2 and q2.
+    vbif        d18, d5, d16               ; t_op2 |= p2 & ~(flat & mask)
+    vbif        d19, d25, d16              ; t_op1 |= f_op1 & ~(flat & mask)
+    vbif        d20, d24, d16              ; t_op0 |= f_op0 & ~(flat & mask)
+    vbif        d21, d23, d16              ; t_oq0 |= f_oq0 & ~(flat & mask)
+    vbif        d22, d26, d16              ; t_oq1 |= f_oq1 & ~(flat & mask)
+
+    vbit        d23, d27, d16              ; t_oq2 |= r_oq2 & (flat & mask)
+    vbif        d23, d10, d16              ; t_oq2 |= q2 & ~(flat & mask)
+
+    tst         r7, #2
+    bxne        lr
+
+    ; wide_mbfilter flat2 && flat && mask branch
+    vmov.u8     d16, #7
+    vaddl.u8    q15, d7, d8                ; op6 = p0 + q0
+    vaddl.u8    q12, d2, d3
+    vaddl.u8    q13, d4, d5
+    vaddl.u8    q14, d1, d6
+    vmlal.u8    q15, d0, d16               ; op6 += p7 * 3
+    vadd.i16    q12, q13
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d2, d9
+    vadd.i16    q15, q12
+    vaddl.u8    q12, d0, d1
+    vaddw.u8    q15, d1
+    vaddl.u8    q13, d0, d2
+    vadd.i16    q14, q15, q14
+    vqrshrn.u16 d16, q15, #4               ; w_op6
+
+    vsub.i16    q15, q14, q12
+    vaddl.u8    q14, d3, d10
+    vqrshrn.u16 d24, q15, #4               ; w_op5
+
+    vsub.i16    q15, q13
+    vaddl.u8    q13, d0, d3
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d4, d11
+    vqrshrn.u16 d25, q15, #4               ; w_op4
+
+    vadd.i16    q15, q14
+    vaddl.u8    q14, d0, d4
+    vsub.i16    q15, q13
+    vsub.i16    q14, q15, q14
+    vqrshrn.u16 d26, q15, #4               ; w_op3
+
+    vaddw.u8    q15, q14, d5               ; op2 += p2
+    vaddl.u8    q14, d0, d5
+    vaddw.u8    q15, d12                   ; op2 += q4
+    vbif        d26, d4, d17               ; op3 |= p3 & ~(f2 & f & m)
+    vqrshrn.u16 d27, q15, #4               ; w_op2
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d0, d6
+    vaddw.u8    q15, d6                    ; op1 += p1
+    vaddw.u8    q15, d13                   ; op1 += q5
+    vbif        d27, d18, d17              ; op2 |= t_op2 & ~(f2 & f & m)
+    vqrshrn.u16 d18, q15, #4               ; w_op1
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d0, d7
+    vaddw.u8    q15, d7                    ; op0 += p0
+    vaddw.u8    q15, d14                   ; op0 += q6
+    vbif        d18, d19, d17              ; op1 |= t_op1 & ~(f2 & f & m)
+    vqrshrn.u16 d19, q15, #4               ; w_op0
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d1, d8
+    vaddw.u8    q15, d8                    ; oq0 += q0
+    vaddw.u8    q15, d15                   ; oq0 += q7
+    vbif        d19, d20, d17              ; op0 |= t_op0 & ~(f2 & f & m)
+    vqrshrn.u16 d20, q15, #4               ; w_oq0
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d2, d9
+    vaddw.u8    q15, d9                    ; oq1 += q1
+    vaddl.u8    q4, d10, d15
+    vaddw.u8    q15, d15                   ; oq1 += q7
+    vbif        d20, d21, d17              ; oq0 |= t_oq0 & ~(f2 & f & m)
+    vqrshrn.u16 d21, q15, #4               ; w_oq1
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d3, d10
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d11, d15
+    vbif        d21, d22, d17              ; oq1 |= t_oq1 & ~(f2 & f & m)
+    vqrshrn.u16 d22, q15, #4               ; w_oq2
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d4, d11
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d12, d15
+    vbif        d22, d23, d17              ; oq2 |= t_oq2 & ~(f2 & f & m)
+    vqrshrn.u16 d23, q15, #4               ; w_oq3
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d5, d12
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d13, d15
+    vbif        d16, d1, d17               ; op6 |= p6 & ~(f2 & f & m)
+    vqrshrn.u16 d1, q15, #4                ; w_oq4
+
+    vsub.i16    q15, q14
+    vaddl.u8    q14, d6, d13
+    vadd.i16    q15, q4
+    vaddl.u8    q4, d14, d15
+    vbif        d24, d2, d17               ; op5 |= p5 & ~(f2 & f & m)
+    vqrshrn.u16 d2, q15, #4                ; w_oq5
+
+    vsub.i16    q15, q14
+    vbif        d25, d3, d17               ; op4 |= p4 & ~(f2 & f & m)
+    vadd.i16    q15, q4
+    vbif        d23, d11, d17              ; oq3 |= q3 & ~(f2 & f & m)
+    vqrshrn.u16 d3, q15, #4                ; w_oq6
+    vbif        d1, d12, d17               ; oq4 |= q4 & ~(f2 & f & m)
+    vbif        d2, d13, d17               ; oq5 |= q5 & ~(f2 & f & m)
+    vbif        d3, d14, d17               ; oq6 |= q6 & ~(f2 & f & m)
+
+    bx          lr
+    ENDP        ; |aom_wide_mbfilter_neon|
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/loopfilter_neon.c b/third_party/aom/aom_dsp/arm/loopfilter_neon.c
new file mode 100644
index 0000000000..c90d6bfde1
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/loopfilter_neon.c
@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+void aom_lpf_vertical_4_dual_neon(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0, const uint8_t *thresh0,
+                                  const uint8_t *blimit1, const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  aom_lpf_vertical_4_neon(s, p, blimit0, limit0, thresh0);
+  aom_lpf_vertical_4_neon(s + 8 * p, p, blimit1, limit1, thresh1);
+}
+
+#if HAVE_NEON_ASM
+void aom_lpf_horizontal_8_dual_neon(
+    uint8_t *s, int p /* pitch */, const uint8_t *blimit0,
+    const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1,
+    const uint8_t *limit1, const uint8_t *thresh1) {
+  aom_lpf_horizontal_8_neon(s, p, blimit0, limit0, thresh0);
+  aom_lpf_horizontal_8_neon(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void aom_lpf_vertical_8_dual_neon(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0, const uint8_t *thresh0,
+                                  const uint8_t *blimit1, const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  aom_lpf_vertical_8_neon(s, p, blimit0, limit0, thresh0);
+  aom_lpf_vertical_8_neon(s + 8 * p, p, blimit1, limit1, thresh1);
+}
+
+void aom_lpf_vertical_16_dual_neon(uint8_t *s, int p, const uint8_t *blimit,
+                                   const uint8_t *limit,
+                                   const uint8_t *thresh) {
+  aom_lpf_vertical_16_neon(s, p, blimit, limit, thresh);
+  aom_lpf_vertical_16_neon(s + 8 * p, p, blimit, limit, thresh);
+}
+#endif  // HAVE_NEON_ASM
diff --git a/third_party/aom/aom_dsp/arm/sad4d_neon.c b/third_party/aom/aom_dsp/arm/sad4d_neon.c
new file mode 100644
index 0000000000..a1eeaf4b77
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/sad4d_neon.c
@@ -0,0 +1,225 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
+                                                    const uint16x8_t vec_hi) {
+  const uint32x4_t vec_l_lo =
+      vaddl_u16(vget_low_u16(vec_lo), vget_high_u16(vec_lo));
+  const uint32x4_t vec_l_hi =
+      vaddl_u16(vget_low_u16(vec_hi), vget_high_u16(vec_hi));
+  const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+
+// Calculate the absolute difference of 64 bytes from vec_src_00, vec_src_16,
+// vec_src_32, vec_src_48 and ref. Accumulate partial sums in vec_sum_ref_lo
+// and vec_sum_ref_hi.
+static void sad_neon_64(const uint8x16_t vec_src_00,
+                        const uint8x16_t vec_src_16,
+                        const uint8x16_t vec_src_32,
+                        const uint8x16_t vec_src_48, const uint8_t *ref,
+                        uint16x8_t *vec_sum_ref_lo,
+                        uint16x8_t *vec_sum_ref_hi) {
+  const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+  const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+  const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32);
+  const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48);
+
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_00),
+                             vget_low_u8(vec_ref_00));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_00),
+                             vget_high_u8(vec_ref_00));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_16),
+                             vget_low_u8(vec_ref_16));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_16),
+                             vget_high_u8(vec_ref_16));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_32),
+                             vget_low_u8(vec_ref_32));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_32),
+                             vget_high_u8(vec_ref_32));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_48),
+                             vget_low_u8(vec_ref_48));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_48),
+                             vget_high_u8(vec_ref_48));
+}
+
+// Calculate the absolute difference of 32 bytes from vec_src_00, vec_src_16,
+// and ref. Accumulate partial sums in vec_sum_ref_lo and vec_sum_ref_hi.
+static void sad_neon_32(const uint8x16_t vec_src_00,
+                        const uint8x16_t vec_src_16, const uint8_t *ref,
+                        uint16x8_t *vec_sum_ref_lo,
+                        uint16x8_t *vec_sum_ref_hi) {
+  const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+  const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_00),
+                             vget_low_u8(vec_ref_00));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_00),
+                             vget_high_u8(vec_ref_00));
+  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_16),
+                             vget_low_u8(vec_ref_16));
+  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_16),
+                             vget_high_u8(vec_ref_16));
+}
+
+void aom_sad64x64x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  int i;
+  uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+
+  for (i = 0; i < 64; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_src_32 = vld1q_u8(src + 32);
+    const uint8x16_t vec_src_48 = vld1q_u8(src + 48);
+
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref0,
+                &vec_sum_ref0_lo, &vec_sum_ref0_hi);
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref1,
+                &vec_sum_ref1_lo, &vec_sum_ref1_hi);
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref2,
+                &vec_sum_ref2_lo, &vec_sum_ref2_hi);
+    sad_neon_64(vec_src_00, vec_src_16, vec_src_32, vec_src_48, ref3,
+                &vec_sum_ref3_lo, &vec_sum_ref3_hi);
+
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+
+  res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
+  res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
+  res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
+  res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
+}
+
+void aom_sad32x32x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  int i;
+  uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+
+  for (i = 0; i < 32; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+
+    sad_neon_32(vec_src_00, vec_src_16, ref0, &vec_sum_ref0_lo,
+                &vec_sum_ref0_hi);
+    sad_neon_32(vec_src_00, vec_src_16, ref1, &vec_sum_ref1_lo,
+                &vec_sum_ref1_hi);
+    sad_neon_32(vec_src_00, vec_src_16, ref2, &vec_sum_ref2_lo,
+                &vec_sum_ref2_hi);
+    sad_neon_32(vec_src_00, vec_src_16, ref3, &vec_sum_ref3_lo,
+                &vec_sum_ref3_hi);
+
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+
+  res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
+  res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
+  res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
+  res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
+}
+
+void aom_sad16x16x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  int i;
+  uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
+  uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+
+  for (i = 0; i < 16; ++i) {
+    const uint8x16_t vec_src = vld1q_u8(src);
+    const uint8x16_t vec_ref0 = vld1q_u8(ref0);
+    const uint8x16_t vec_ref1 = vld1q_u8(ref1);
+    const uint8x16_t vec_ref2 = vld1q_u8(ref2);
+    const uint8x16_t vec_ref3 = vld1q_u8(ref3);
+
+    vec_sum_ref0_lo =
+        vabal_u8(vec_sum_ref0_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref0));
+    vec_sum_ref0_hi = vabal_u8(vec_sum_ref0_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref0));
+    vec_sum_ref1_lo =
+        vabal_u8(vec_sum_ref1_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref1));
+    vec_sum_ref1_hi = vabal_u8(vec_sum_ref1_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref1));
+    vec_sum_ref2_lo =
+        vabal_u8(vec_sum_ref2_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref2));
+    vec_sum_ref2_hi = vabal_u8(vec_sum_ref2_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref2));
+    vec_sum_ref3_lo =
+        vabal_u8(vec_sum_ref3_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref3));
+    vec_sum_ref3_hi = vabal_u8(vec_sum_ref3_hi, vget_high_u8(vec_src),
+                               vget_high_u8(vec_ref3));
+
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+
+  res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
+  res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
+  res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
+  res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
+}
diff --git a/third_party/aom/aom_dsp/arm/sad_media.asm b/third_party/aom/aom_dsp/arm/sad_media.asm
new file mode 100644
index 0000000000..49ddb67642
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/sad_media.asm
@@ -0,0 +1,98 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_sad16x16_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    const unsigned char *src_ptr
+; r1    int  src_stride
+; r2    const unsigned char *ref_ptr
+; r3    int  ref_stride
+|aom_sad16x16_media| PROC
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+    pld     [r0, r1, lsl #1]
+    pld     [r2, r3, lsl #1]
+
+    mov     r4, #0              ; sad = 0;
+    mov     r5, #8              ; loop count
+
+loop
+    ; 1st row
+    ldr     r6, [r0, #0x0]      ; load 4 src pixels (1A)
+    ldr     r8, [r2, #0x0]      ; load 4 ref pixels (1A)
+    ldr     r7, [r0, #0x4]      ; load 4 src pixels (1A)
+    ldr     r9, [r2, #0x4]      ; load 4 ref pixels (1A)
+    ldr     r10, [r0, #0x8]     ; load 4 src pixels (1B)
+    ldr     r11, [r0, #0xC]     ; load 4 src pixels (1B)
+
+    usada8  r4, r8, r6, r4      ; calculate sad for 4 pixels
+    usad8   r8, r7, r9          ; calculate sad for 4 pixels
+
+    ldr     r12, [r2, #0x8]     ; load 4 ref pixels (1B)
+    ldr     lr, [r2, #0xC]      ; load 4 ref pixels (1B)
+
+    add     r0, r0, r1          ; set src pointer to next row
+    add     r2, r2, r3          ; set dst pointer to next row
+
+    pld     [r0, r1, lsl #1]
+    pld     [r2, r3, lsl #1]
+
+    usada8  r4, r10, r12, r4    ; calculate sad for 4 pixels
+    usada8  r8, r11, lr, r8     ; calculate sad for 4 pixels
+
+    ldr     r6, [r0, #0x0]      ; load 4 src pixels (2A)
+    ldr     r7, [r0, #0x4]      ; load 4 src pixels (2A)
+    add     r4, r4, r8          ; add partial sad values
+
+    ; 2nd row
+    ldr     r8, [r2, #0x0]      ; load 4 ref pixels (2A)
+    ldr     r9, [r2, #0x4]      ; load 4 ref pixels (2A)
+    ldr     r10, [r0, #0x8]     ; load 4 src pixels (2B)
+    ldr     r11, [r0, #0xC]     ; load 4 src pixels (2B)
+
+    usada8  r4, r6, r8, r4      ; calculate sad for 4 pixels
+    usad8   r8, r7, r9          ; calculate sad for 4 pixels
+
+    ldr     r12, [r2, #0x8]     ; load 4 ref pixels (2B)
+    ldr     lr, [r2, #0xC]      ; load 4 ref pixels (2B)
+
+    add     r0, r0, r1          ; set src pointer to next row
+    add     r2, r2, r3          ; set dst pointer to next row
+
+    usada8  r4, r10, r12, r4    ; calculate sad for 4 pixels
+    usada8  r8, r11, lr, r8     ; calculate sad for 4 pixels
+
+    pld     [r0, r1, lsl #1]
+    pld     [r2, r3, lsl #1]
+
+    subs    r5, r5, #1          ; decrement loop counter
+    add     r4, r4, r8          ; add partial sad values
+
+    bne     loop
+
+    mov     r0, r4              ; return sad
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+    END
+
diff --git a/third_party/aom/aom_dsp/arm/sad_neon.c b/third_party/aom/aom_dsp/arm/sad_neon.c
new file mode 100644
index 0000000000..2f452f55b5
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/sad_neon.c
@@ -0,0 +1,224 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+
+unsigned int aom_sad8x16_neon(unsigned char *src_ptr, int src_stride,
+                              unsigned char *ref_ptr, int ref_stride) {
+  uint8x8_t d0, d8;
+  uint16x8_t q12;
+  uint32x4_t q1;
+  uint64x2_t q3;
+  uint32x2_t d5;
+  int i;
+
+  d0 = vld1_u8(src_ptr);
+  src_ptr += src_stride;
+  d8 = vld1_u8(ref_ptr);
+  ref_ptr += ref_stride;
+  q12 = vabdl_u8(d0, d8);
+
+  for (i = 0; i < 15; i++) {
+    d0 = vld1_u8(src_ptr);
+    src_ptr += src_stride;
+    d8 = vld1_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabal_u8(q12, d0, d8);
+  }
+
+  q1 = vpaddlq_u16(q12);
+  q3 = vpaddlq_u32(q1);
+  d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
+                vreinterpret_u32_u64(vget_high_u64(q3)));
+
+  return vget_lane_u32(d5, 0);
+}
+
+unsigned int aom_sad4x4_neon(unsigned char *src_ptr, int src_stride,
+                             unsigned char *ref_ptr, int ref_stride) {
+  uint8x8_t d0, d8;
+  uint16x8_t q12;
+  uint32x2_t d1;
+  uint64x1_t d3;
+  int i;
+
+  d0 = vld1_u8(src_ptr);
+  src_ptr += src_stride;
+  d8 = vld1_u8(ref_ptr);
+  ref_ptr += ref_stride;
+  q12 = vabdl_u8(d0, d8);
+
+  for (i = 0; i < 3; i++) {
+    d0 = vld1_u8(src_ptr);
+    src_ptr += src_stride;
+    d8 = vld1_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabal_u8(q12, d0, d8);
+  }
+
+  d1 = vpaddl_u16(vget_low_u16(q12));
+  d3 = vpaddl_u32(d1);
+
+  return vget_lane_u32(vreinterpret_u32_u64(d3), 0);
+}
+
+unsigned int aom_sad16x8_neon(unsigned char *src_ptr, int src_stride,
+                              unsigned char *ref_ptr, int ref_stride) {
+  uint8x16_t q0, q4;
+  uint16x8_t q12, q13;
+  uint32x4_t q1;
+  uint64x2_t q3;
+  uint32x2_t d5;
+  int i;
+
+  q0 = vld1q_u8(src_ptr);
+  src_ptr += src_stride;
+  q4 = vld1q_u8(ref_ptr);
+  ref_ptr += ref_stride;
+  q12 = vabdl_u8(vget_low_u8(q0), vget_low_u8(q4));
+  q13 = vabdl_u8(vget_high_u8(q0), vget_high_u8(q4));
+
+  for (i = 0; i < 7; i++) {
+    q0 = vld1q_u8(src_ptr);
+    src_ptr += src_stride;
+    q4 = vld1q_u8(ref_ptr);
+    ref_ptr += ref_stride;
+    q12 = vabal_u8(q12, vget_low_u8(q0), vget_low_u8(q4));
+    q13 = vabal_u8(q13, vget_high_u8(q0), vget_high_u8(q4));
+  }
+
+  q12 = vaddq_u16(q12, q13);
+  q1 = vpaddlq_u16(q12);
+  q3 = vpaddlq_u32(q1);
+  d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
+                vreinterpret_u32_u64(vget_high_u64(q3)));
+
+  return vget_lane_u32(d5, 0);
+}
+
+static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
+                                                    const uint16x8_t vec_hi) {
+  const uint32x4_t vec_l_lo =
+      vaddl_u16(vget_low_u16(vec_lo), vget_high_u16(vec_lo));
+  const uint32x4_t vec_l_hi =
+      vaddl_u16(vget_low_u16(vec_hi), vget_high_u16(vec_hi));
+  const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+static INLINE unsigned int horizontal_add_16x8(const uint16x8_t vec_16x8) {
+  const uint32x4_t a = vpaddlq_u16(vec_16x8);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
+
+unsigned int aom_sad64x64_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+  for (i = 0; i < 64; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_src_32 = vld1q_u8(src + 32);
+    const uint8x16_t vec_src_48 = vld1q_u8(src + 48);
+    const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+    const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+    const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32);
+    const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
+                            vget_low_u8(vec_ref_00));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
+                            vget_high_u8(vec_ref_00));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
+                            vget_low_u8(vec_ref_16));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
+                            vget_high_u8(vec_ref_16));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_32),
+                            vget_low_u8(vec_ref_32));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_32),
+                            vget_high_u8(vec_ref_32));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_48),
+                            vget_low_u8(vec_ref_48));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_48),
+                            vget_high_u8(vec_ref_48));
+  }
+  return horizontal_long_add_16x8(vec_accum_lo, vec_accum_hi);
+}
+
+unsigned int aom_sad32x32_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+
+  for (i = 0; i < 32; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+    const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
+                            vget_low_u8(vec_ref_00));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
+                            vget_high_u8(vec_ref_00));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
+                            vget_low_u8(vec_ref_16));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
+                            vget_high_u8(vec_ref_16));
+  }
+  return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
+}
+
+unsigned int aom_sad16x16_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+
+  for (i = 0; i < 16; ++i) {
+    const uint8x16_t vec_src = vld1q_u8(src);
+    const uint8x16_t vec_ref = vld1q_u8(ref);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo =
+        vabal_u8(vec_accum_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref));
+    vec_accum_hi =
+        vabal_u8(vec_accum_hi, vget_high_u8(vec_src), vget_high_u8(vec_ref));
+  }
+  return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
+}
+
+unsigned int aom_sad8x8_neon(const uint8_t *src, int src_stride,
+                             const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum = vdupq_n_u16(0);
+
+  for (i = 0; i < 8; ++i) {
+    const uint8x8_t vec_src = vld1_u8(src);
+    const uint8x8_t vec_ref = vld1_u8(ref);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum = vabal_u8(vec_accum, vec_src, vec_ref);
+  }
+  return horizontal_add_16x8(vec_accum);
+}
diff --git a/third_party/aom/aom_dsp/arm/save_reg_neon.asm b/third_party/aom/aom_dsp/arm/save_reg_neon.asm
new file mode 100644
index 0000000000..e049698234
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/save_reg_neon.asm
@@ -0,0 +1,39 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_push_neon|
+    EXPORT  |aom_pop_neon|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+|aom_push_neon| PROC
+    vst1.i64            {d8, d9, d10, d11}, [r0]!
+    vst1.i64            {d12, d13, d14, d15}, [r0]!
+    bx              lr
+
+    ENDP
+
+|aom_pop_neon| PROC
+    vld1.i64            {d8, d9, d10, d11}, [r0]!
+    vld1.i64            {d12, d13, d14, d15}, [r0]!
+    bx              lr
+
+    ENDP
+
+    END
+
diff --git a/third_party/aom/aom_dsp/arm/subpel_variance_media.c b/third_party/aom/aom_dsp/arm/subpel_variance_media.c
new file mode 100644
index 0000000000..46ec028d37
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/subpel_variance_media.c
@@ -0,0 +1,81 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_MEDIA
+static const int16_t bilinear_filters_media[8][2] = { { 128, 0 }, { 112, 16 },
+                                                      { 96, 32 }, { 80, 48 },
+                                                      { 64, 64 }, { 48, 80 },
+                                                      { 32, 96 }, { 16, 112 } };
+
+extern void aom_filter_block2d_bil_first_pass_media(
+    const uint8_t *src_ptr, uint16_t *dst_ptr, uint32_t src_pitch,
+    uint32_t height, uint32_t width, const int16_t *filter);
+
+extern void aom_filter_block2d_bil_second_pass_media(
+    const uint16_t *src_ptr, uint8_t *dst_ptr, int32_t src_pitch,
+    uint32_t height, uint32_t width, const int16_t *filter);
+
+unsigned int aom_sub_pixel_variance8x8_media(
+    const uint8_t *src_ptr, int src_pixels_per_line, int xoffset, int yoffset,
+    const uint8_t *dst_ptr, int dst_pixels_per_line, unsigned int *sse) {
+  uint16_t first_pass[10 * 8];
+  uint8_t second_pass[8 * 8];
+  const int16_t *HFilter, *VFilter;
+
+  HFilter = bilinear_filters_media[xoffset];
+  VFilter = bilinear_filters_media[yoffset];
+
+  aom_filter_block2d_bil_first_pass_media(src_ptr, first_pass,
+                                          src_pixels_per_line, 9, 8, HFilter);
+  aom_filter_block2d_bil_second_pass_media(first_pass, second_pass, 8, 8, 8,
+                                           VFilter);
+
+  return aom_variance8x8_media(second_pass, 8, dst_ptr, dst_pixels_per_line,
+                               sse);
+}
+
+unsigned int aom_sub_pixel_variance16x16_media(
+    const uint8_t *src_ptr, int src_pixels_per_line, int xoffset, int yoffset,
+    const uint8_t *dst_ptr, int dst_pixels_per_line, unsigned int *sse) {
+  uint16_t first_pass[36 * 16];
+  uint8_t second_pass[20 * 16];
+  const int16_t *HFilter, *VFilter;
+  unsigned int var;
+
+  if (xoffset == 4 && yoffset == 0) {
+    var = aom_variance_halfpixvar16x16_h_media(
+        src_ptr, src_pixels_per_line, dst_ptr, dst_pixels_per_line, sse);
+  } else if (xoffset == 0 && yoffset == 4) {
+    var = aom_variance_halfpixvar16x16_v_media(
+        src_ptr, src_pixels_per_line, dst_ptr, dst_pixels_per_line, sse);
+  } else if (xoffset == 4 && yoffset == 4) {
+    var = aom_variance_halfpixvar16x16_hv_media(
+        src_ptr, src_pixels_per_line, dst_ptr, dst_pixels_per_line, sse);
+  } else {
+    HFilter = bilinear_filters_media[xoffset];
+    VFilter = bilinear_filters_media[yoffset];
+
+    aom_filter_block2d_bil_first_pass_media(
+        src_ptr, first_pass, src_pixels_per_line, 17, 16, HFilter);
+    aom_filter_block2d_bil_second_pass_media(first_pass, second_pass, 16, 16,
+                                             16, VFilter);
+
+    var = aom_variance16x16_media(second_pass, 16, dst_ptr, dst_pixels_per_line,
+                                  sse);
+  }
+  return var;
+}
+#endif  // HAVE_MEDIA
diff --git a/third_party/aom/aom_dsp/arm/subpel_variance_neon.c b/third_party/aom/aom_dsp/arm/subpel_variance_neon.c
new file mode 100644
index 0000000000..064b72d6fc
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/subpel_variance_neon.c
@@ -0,0 +1,134 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+
+#include "aom_ports/mem.h"
+#include "aom/aom_integer.h"
+
+#include "aom_dsp/variance.h"
+
+static const uint8_t bilinear_filters[8][2] = {
+  { 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 },
+  { 64, 64 }, { 48, 80 },  { 32, 96 }, { 16, 112 },
+};
+
+static void var_filter_block2d_bil_w8(const uint8_t *src_ptr,
+                                      uint8_t *output_ptr,
+                                      unsigned int src_pixels_per_line,
+                                      int pixel_step,
+                                      unsigned int output_height,
+                                      unsigned int output_width,
+                                      const uint8_t *filter) {
+  const uint8x8_t f0 = vmov_n_u8(filter[0]);
+  const uint8x8_t f1 = vmov_n_u8(filter[1]);
+  unsigned int i;
+  for (i = 0; i < output_height; ++i) {
+    const uint8x8_t src_0 = vld1_u8(&src_ptr[0]);
+    const uint8x8_t src_1 = vld1_u8(&src_ptr[pixel_step]);
+    const uint16x8_t a = vmull_u8(src_0, f0);
+    const uint16x8_t b = vmlal_u8(a, src_1, f1);
+    const uint8x8_t out = vrshrn_n_u16(b, FILTER_BITS);
+    vst1_u8(&output_ptr[0], out);
+    // Next row...
+    src_ptr += src_pixels_per_line;
+    output_ptr += output_width;
+  }
+}
+
+static void var_filter_block2d_bil_w16(const uint8_t *src_ptr,
+                                       uint8_t *output_ptr,
+                                       unsigned int src_pixels_per_line,
+                                       int pixel_step,
+                                       unsigned int output_height,
+                                       unsigned int output_width,
+                                       const uint8_t *filter) {
+  const uint8x8_t f0 = vmov_n_u8(filter[0]);
+  const uint8x8_t f1 = vmov_n_u8(filter[1]);
+  unsigned int i, j;
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; j += 16) {
+      const uint8x16_t src_0 = vld1q_u8(&src_ptr[j]);
+      const uint8x16_t src_1 = vld1q_u8(&src_ptr[j + pixel_step]);
+      const uint16x8_t a = vmull_u8(vget_low_u8(src_0), f0);
+      const uint16x8_t b = vmlal_u8(a, vget_low_u8(src_1), f1);
+      const uint8x8_t out_lo = vrshrn_n_u16(b, FILTER_BITS);
+      const uint16x8_t c = vmull_u8(vget_high_u8(src_0), f0);
+      const uint16x8_t d = vmlal_u8(c, vget_high_u8(src_1), f1);
+      const uint8x8_t out_hi = vrshrn_n_u16(d, FILTER_BITS);
+      vst1q_u8(&output_ptr[j], vcombine_u8(out_lo, out_hi));
+    }
+    // Next row...
+    src_ptr += src_pixels_per_line;
+    output_ptr += output_width;
+  }
+}
+
+unsigned int aom_sub_pixel_variance8x8_neon(const uint8_t *src, int src_stride,
+                                            int xoffset, int yoffset,
+                                            const uint8_t *dst, int dst_stride,
+                                            unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[8 * 8]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[9 * 8]);
+
+  var_filter_block2d_bil_w8(src, fdata3, src_stride, 1, 9, 8,
+                            bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w8(fdata3, temp2, 8, 8, 8, 8,
+                            bilinear_filters[yoffset]);
+  return aom_variance8x8_neon(temp2, 8, dst, dst_stride, sse);
+}
+
+unsigned int aom_sub_pixel_variance16x16_neon(const uint8_t *src,
+                                              int src_stride, int xoffset,
+                                              int yoffset, const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[16 * 16]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[17 * 16]);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 17, 16,
+                             bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w16(fdata3, temp2, 16, 16, 16, 16,
+                             bilinear_filters[yoffset]);
+  return aom_variance16x16_neon(temp2, 16, dst, dst_stride, sse);
+}
+
+unsigned int aom_sub_pixel_variance32x32_neon(const uint8_t *src,
+                                              int src_stride, int xoffset,
+                                              int yoffset, const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[32 * 32]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[33 * 32]);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 33, 32,
+                             bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w16(fdata3, temp2, 32, 32, 32, 32,
+                             bilinear_filters[yoffset]);
+  return aom_variance32x32_neon(temp2, 32, dst, dst_stride, sse);
+}
+
+unsigned int aom_sub_pixel_variance64x64_neon(const uint8_t *src,
+                                              int src_stride, int xoffset,
+                                              int yoffset, const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED(16, uint8_t, temp2[64 * 64]);
+  DECLARE_ALIGNED(16, uint8_t, fdata3[65 * 64]);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 65, 64,
+                             bilinear_filters[xoffset]);
+  var_filter_block2d_bil_w16(fdata3, temp2, 64, 64, 64, 64,
+                             bilinear_filters[yoffset]);
+  return aom_variance64x64_neon(temp2, 64, dst, dst_stride, sse);
+}
diff --git a/third_party/aom/aom_dsp/arm/subtract_neon.c b/third_party/aom/aom_dsp/arm/subtract_neon.c
new file mode 100644
index 0000000000..cb8a2daf8a
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/subtract_neon.c
@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+void aom_subtract_block_neon(int rows, int cols, int16_t *diff,
+                             ptrdiff_t diff_stride, const uint8_t *src,
+                             ptrdiff_t src_stride, const uint8_t *pred,
+                             ptrdiff_t pred_stride) {
+  int r, c;
+
+  if (cols > 16) {
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; c += 32) {
+        const uint8x16_t v_src_00 = vld1q_u8(&src[c + 0]);
+        const uint8x16_t v_src_16 = vld1q_u8(&src[c + 16]);
+        const uint8x16_t v_pred_00 = vld1q_u8(&pred[c + 0]);
+        const uint8x16_t v_pred_16 = vld1q_u8(&pred[c + 16]);
+        const uint16x8_t v_diff_lo_00 =
+            vsubl_u8(vget_low_u8(v_src_00), vget_low_u8(v_pred_00));
+        const uint16x8_t v_diff_hi_00 =
+            vsubl_u8(vget_high_u8(v_src_00), vget_high_u8(v_pred_00));
+        const uint16x8_t v_diff_lo_16 =
+            vsubl_u8(vget_low_u8(v_src_16), vget_low_u8(v_pred_16));
+        const uint16x8_t v_diff_hi_16 =
+            vsubl_u8(vget_high_u8(v_src_16), vget_high_u8(v_pred_16));
+        vst1q_s16(&diff[c + 0], vreinterpretq_s16_u16(v_diff_lo_00));
+        vst1q_s16(&diff[c + 8], vreinterpretq_s16_u16(v_diff_hi_00));
+        vst1q_s16(&diff[c + 16], vreinterpretq_s16_u16(v_diff_lo_16));
+        vst1q_s16(&diff[c + 24], vreinterpretq_s16_u16(v_diff_hi_16));
+      }
+      diff += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    }
+  } else if (cols > 8) {
+    for (r = 0; r < rows; ++r) {
+      const uint8x16_t v_src = vld1q_u8(&src[0]);
+      const uint8x16_t v_pred = vld1q_u8(&pred[0]);
+      const uint16x8_t v_diff_lo =
+          vsubl_u8(vget_low_u8(v_src), vget_low_u8(v_pred));
+      const uint16x8_t v_diff_hi =
+          vsubl_u8(vget_high_u8(v_src), vget_high_u8(v_pred));
+      vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff_lo));
+      vst1q_s16(&diff[8], vreinterpretq_s16_u16(v_diff_hi));
+      diff += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    }
+  } else if (cols > 4) {
+    for (r = 0; r < rows; ++r) {
+      const uint8x8_t v_src = vld1_u8(&src[0]);
+      const uint8x8_t v_pred = vld1_u8(&pred[0]);
+      const uint16x8_t v_diff = vsubl_u8(v_src, v_pred);
+      vst1q_s16(&diff[0], vreinterpretq_s16_u16(v_diff));
+      diff += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    }
+  } else {
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; ++c) diff[c] = src[c] - pred[c];
+
+      diff += diff_stride;
+      pred += pred_stride;
+      src += src_stride;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_h_media.asm b/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_h_media.asm
new file mode 100644
index 0000000000..1e5c9178e6
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_h_media.asm
@@ -0,0 +1,185 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_variance_halfpixvar16x16_h_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|aom_variance_halfpixvar16x16_h_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    ldr     r10, c80808080
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     lr, #0              ; constant zero
+loop
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load 4 src pixels
+    ldr     r6, [r0, #1]        ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load 4 src pixels
+    ldr     r6, [r0, #5]        ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load 4 src pixels
+    ldr     r6, [r0, #9]        ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load 4 src pixels
+    ldr     r6, [r0, #13]       ; load 4 src pixels with 1 byte offset
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    subs    r12, r12, #1
+
+    bne     loop
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+c80808080
+    DCD     0x80808080
+
+    END
+
diff --git a/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_hv_media.asm b/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_hv_media.asm
new file mode 100644
index 0000000000..9e0af830ee
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_hv_media.asm
@@ -0,0 +1,225 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_variance_halfpixvar16x16_hv_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|aom_variance_halfpixvar16x16_hv_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    ldr     r10, c80808080
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     lr, #0              ; constant zero
+loop
+    add     r9, r0, r1          ; pointer to pixels on the next row
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load source pixels a, row N
+    ldr     r6, [r0, #1]        ; load source pixels b, row N
+    ldr     r5, [r9, #0]        ; load source pixels c, row N+1
+    ldr     r7, [r9, #1]        ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load source pixels a, row N
+    ldr     r6, [r0, #5]        ; load source pixels b, row N
+    ldr     r5, [r9, #4]        ; load source pixels c, row N+1
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    ldr     r7, [r9, #5]        ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load source pixels a, row N
+    ldr     r6, [r0, #9]        ; load source pixels b, row N
+    ldr     r5, [r9, #8]        ; load source pixels c, row N+1
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    ldr     r7, [r9, #9]        ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load source pixels a, row N
+    ldr     r6, [r0, #13]       ; load source pixels b, row N
+    ldr     r5, [r9, #12]       ; load source pixels c, row N+1
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+    ldr     r7, [r9, #13]       ; load source pixels d, row N+1
+
+    ; x = (a + b + 1) >> 1, interpolate pixels horizontally on row N
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+    ; y = (c + d + 1) >> 1, interpolate pixels horizontally on row N+1
+    mvn     r7, r7
+    uhsub8  r5, r5, r7
+    eor     r5, r5, r10
+    ; z = (x + y + 1) >> 1, interpolate half pixel values vertically
+    mvn     r5, r5
+    uhsub8  r4, r4, r5
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    subs    r12, r12, #1
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    bne     loop
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+c80808080
+    DCD     0x80808080
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_v_media.asm b/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_v_media.asm
new file mode 100644
index 0000000000..545b681794
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/variance_halfpixvar16x16_v_media.asm
@@ -0,0 +1,187 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_variance_halfpixvar16x16_v_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|aom_variance_halfpixvar16x16_v_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    ldr     r10, c80808080
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     lr, #0              ; constant zero
+loop
+    add     r9, r0, r1          ; set src pointer to next row
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load 4 src pixels
+    ldr     r6, [r9, #0]        ; load 4 src pixels from next row
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load 4 src pixels
+    ldr     r6, [r9, #4]        ; load 4 src pixels from next row
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load 4 src pixels
+    ldr     r6, [r9, #8]        ; load 4 src pixels from next row
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load 4 src pixels
+    ldr     r6, [r9, #12]       ; load 4 src pixels from next row
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+
+    ; bilinear interpolation
+    mvn     r6, r6
+    uhsub8  r4, r4, r6
+    eor     r4, r4, r10
+
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r6, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r6, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r7, r6, ror #8      ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    smlad   r11, r7, r7, r11    ; dual signed multiply, add and accumulate (2)
+
+
+    subs    r12, r12, #1
+
+    bne     loop
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+c80808080
+    DCD     0x80808080
+
+    END
+
diff --git a/third_party/aom/aom_dsp/arm/variance_media.asm b/third_party/aom/aom_dsp/arm/variance_media.asm
new file mode 100644
index 0000000000..fdc311a81c
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/variance_media.asm
@@ -0,0 +1,361 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+    EXPORT  |aom_variance16x16_media|
+    EXPORT  |aom_variance8x8_media|
+    EXPORT  |aom_mse16x16_media|
+
+    ARM
+    REQUIRE8
+    PRESERVE8
+
+    AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|aom_variance16x16_media| PROC
+
+    stmfd   sp!, {r4-r12, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r8, #0              ; initialize sum = 0
+    mov     r11, #0             ; initialize sse = 0
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+
+loop16x16
+    ; 1st 4 pixels
+    ldr     r4, [r0, #0]        ; load 4 src pixels
+    ldr     r5, [r2, #0]        ; load 4 ref pixels
+
+    mov     lr, #0              ; constant zero
+
+    usub8   r6, r4, r5          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+    ; calculate total sum
+    adds    r8, r8, r4          ; add positive differences to sum
+    subs    r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r4, [r0, #4]        ; load 4 src pixels
+    ldr     r5, [r2, #4]        ; load 4 ref pixels
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r4, [r0, #8]        ; load 4 src pixels
+    ldr     r5, [r2, #8]        ; load 4 ref pixels
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r4, [r0, #12]       ; load 4 src pixels
+    ldr     r5, [r2, #12]       ; load 4 ref pixels
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+    usub8   r6, r4, r5          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r6, lr          ; select bytes with positive difference
+    usub8   r9, r5, r4          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r6, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r4, r7, lr          ; calculate sum of positive differences
+    usad8   r5, r6, lr          ; calculate sum of negative differences
+    orr     r6, r6, r7          ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r8, r8, r4          ; add positive differences to sum
+    sub     r8, r8, r5          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r5, r6              ; byte (two pixels) to halfwords
+    uxtb16  r10, r6, ror #8     ; another two pixels to halfwords
+    smlad   r11, r5, r5, r11    ; dual signed multiply, add and accumulate (1)
+    smlad   r11, r10, r10, r11  ; dual signed multiply, add and accumulate (2)
+
+
+    subs    r12, r12, #1
+
+    bne     loop16x16
+
+    ; return stuff
+    ldr     r6, [sp, #40]       ; get address of sse
+    mul     r0, r8, r8          ; sum * sum
+    str     r11, [r6]           ; store sse
+    sub     r0, r11, r0, lsr #8 ; return (sse - ((sum * sum) >> 8))
+
+    ldmfd   sp!, {r4-r12, pc}
+
+    ENDP
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+|aom_variance8x8_media| PROC
+
+    push    {r4-r10, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r12, #8             ; set loop counter to 8 (=block height)
+    mov     r4, #0              ; initialize sum = 0
+    mov     r5, #0              ; initialize sse = 0
+
+loop8x8
+    ; 1st 4 pixels
+    ldr     r6, [r0, #0x0]      ; load 4 src pixels
+    ldr     r7, [r2, #0x0]      ; load 4 ref pixels
+
+    mov     lr, #0              ; constant zero
+
+    usub8   r8, r6, r7          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r10, r8, lr         ; select bytes with positive difference
+    usub8   r9, r7, r6          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r6, r10, lr         ; calculate sum of positive differences
+    usad8   r7, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r10         ; differences of all 4 pixels
+    ; calculate total sum
+    add    r4, r4, r6           ; add positive differences to sum
+    sub    r4, r4, r7           ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r7, r8              ; byte (two pixels) to halfwords
+    uxtb16  r10, r8, ror #8     ; another two pixels to halfwords
+    smlad   r5, r7, r7, r5      ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r6, [r0, #0x4]      ; load 4 src pixels
+    ldr     r7, [r2, #0x4]      ; load 4 ref pixels
+    smlad   r5, r10, r10, r5    ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r6, r7          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r10, r8, lr         ; select bytes with positive difference
+    usub8   r9, r7, r6          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r6, r10, lr         ; calculate sum of positive differences
+    usad8   r7, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r10         ; differences of all 4 pixels
+
+    ; calculate total sum
+    add     r4, r4, r6          ; add positive differences to sum
+    sub     r4, r4, r7          ; subtract negative differences from sum
+
+    ; calculate sse
+    uxtb16  r7, r8              ; byte (two pixels) to halfwords
+    uxtb16  r10, r8, ror #8     ; another two pixels to halfwords
+    smlad   r5, r7, r7, r5      ; dual signed multiply, add and accumulate (1)
+    subs    r12, r12, #1        ; next row
+    smlad   r5, r10, r10, r5    ; dual signed multiply, add and accumulate (2)
+
+    bne     loop8x8
+
+    ; return stuff
+    ldr     r8, [sp, #32]       ; get address of sse
+    mul     r1, r4, r4          ; sum * sum
+    str     r5, [r8]            ; store sse
+    sub     r0, r5, r1, ASR #6  ; return (sse - ((sum * sum) >> 6))
+
+    pop     {r4-r10, pc}
+
+    ENDP
+
+; r0    unsigned char *src_ptr
+; r1    int source_stride
+; r2    unsigned char *ref_ptr
+; r3    int  recon_stride
+; stack unsigned int *sse
+;
+;note: Based on aom_variance16x16_media. In this function, sum is never used.
+;      So, we can remove this part of calculation.
+
+|aom_mse16x16_media| PROC
+
+    push    {r4-r9, lr}
+
+    pld     [r0, r1, lsl #0]
+    pld     [r2, r3, lsl #0]
+
+    mov     r12, #16            ; set loop counter to 16 (=block height)
+    mov     r4, #0              ; initialize sse = 0
+
+loopmse
+    ; 1st 4 pixels
+    ldr     r5, [r0, #0x0]      ; load 4 src pixels
+    ldr     r6, [r2, #0x0]      ; load 4 ref pixels
+
+    mov     lr, #0              ; constant zero
+
+    usub8   r8, r5, r6          ; calculate difference
+    pld     [r0, r1, lsl #1]
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    pld     [r2, r3, lsl #1]
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+
+    ldr     r5, [r0, #0x4]      ; load 4 src pixels
+
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+
+    ; 2nd 4 pixels
+    ldr     r6, [r2, #0x4]      ; load 4 ref pixels
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r5, r6          ; calculate difference
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+    ldr     r5, [r0, #0x8]      ; load 4 src pixels
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+
+    ; 3rd 4 pixels
+    ldr     r6, [r2, #0x8]      ; load 4 ref pixels
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r5, r6          ; calculate difference
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+
+    ldr     r5, [r0, #0xc]      ; load 4 src pixels
+
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+
+    ; 4th 4 pixels
+    ldr     r6, [r2, #0xc]      ; load 4 ref pixels
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    usub8   r8, r5, r6          ; calculate difference
+    add     r0, r0, r1          ; set src_ptr to next row
+    sel     r7, r8, lr          ; select bytes with positive difference
+    usub8   r9, r6, r5          ; calculate difference with reversed operands
+    add     r2, r2, r3          ; set dst_ptr to next row
+    sel     r8, r9, lr          ; select bytes with negative difference
+
+    ; calculate partial sums
+    usad8   r5, r7, lr          ; calculate sum of positive differences
+    usad8   r6, r8, lr          ; calculate sum of negative differences
+    orr     r8, r8, r7          ; differences of all 4 pixels
+
+    subs    r12, r12, #1        ; next row
+
+    ; calculate sse
+    uxtb16  r6, r8              ; byte (two pixels) to halfwords
+    uxtb16  r7, r8, ror #8      ; another two pixels to halfwords
+    smlad   r4, r6, r6, r4      ; dual signed multiply, add and accumulate (1)
+    smlad   r4, r7, r7, r4      ; dual signed multiply, add and accumulate (2)
+
+    bne     loopmse
+
+    ; return stuff
+    ldr     r1, [sp, #28]       ; get address of sse
+    mov     r0, r4              ; return sse
+    str     r4, [r1]            ; store sse
+
+    pop     {r4-r9, pc}
+
+    ENDP
+
+    END
diff --git a/third_party/aom/aom_dsp/arm/variance_neon.c b/third_party/aom/aom_dsp/arm/variance_neon.c
new file mode 100644
index 0000000000..dbab287e35
--- /dev/null
+++ b/third_party/aom/aom_dsp/arm/variance_neon.c
@@ -0,0 +1,400 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+static INLINE int horizontal_add_s16x8(const int16x8_t v_16x8) {
+  const int32x4_t a = vpaddlq_s16(v_16x8);
+  const int64x2_t b = vpaddlq_s32(a);
+  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                               vreinterpret_s32_s64(vget_high_s64(b)));
+  return vget_lane_s32(c, 0);
+}
+
+static INLINE int horizontal_add_s32x4(const int32x4_t v_32x4) {
+  const int64x2_t b = vpaddlq_s32(v_32x4);
+  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
+                               vreinterpret_s32_s64(vget_high_s64(b)));
+  return vget_lane_s32(c, 0);
+}
+
+// w * h must be less than 2048 or local variable v_sum may overflow.
+static void variance_neon_w8(const uint8_t *a, int a_stride, const uint8_t *b,
+                             int b_stride, int w, int h, uint32_t *sse,
+                             int *sum) {
+  int i, j;
+  int16x8_t v_sum = vdupq_n_s16(0);
+  int32x4_t v_sse_lo = vdupq_n_s32(0);
+  int32x4_t v_sse_hi = vdupq_n_s32(0);
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; j += 8) {
+      const uint8x8_t v_a = vld1_u8(&a[j]);
+      const uint8x8_t v_b = vld1_u8(&b[j]);
+      const uint16x8_t v_diff = vsubl_u8(v_a, v_b);
+      const int16x8_t sv_diff = vreinterpretq_s16_u16(v_diff);
+      v_sum = vaddq_s16(v_sum, sv_diff);
+      v_sse_lo =
+          vmlal_s16(v_sse_lo, vget_low_s16(sv_diff), vget_low_s16(sv_diff));
+      v_sse_hi =
+          vmlal_s16(v_sse_hi, vget_high_s16(sv_diff), vget_high_s16(sv_diff));
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+
+  *sum = horizontal_add_s16x8(v_sum);
+  *sse = (unsigned int)horizontal_add_s32x4(vaddq_s32(v_sse_lo, v_sse_hi));
+}
+
+void aom_get8x8var_neon(const uint8_t *a, int a_stride, const uint8_t *b,
+                        int b_stride, unsigned int *sse, int *sum) {
+  variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, sum);
+}
+
+void aom_get16x16var_neon(const uint8_t *a, int a_stride, const uint8_t *b,
+                          int b_stride, unsigned int *sse, int *sum) {
+  variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, sum);
+}
+
+unsigned int aom_variance8x8_neon(const uint8_t *a, int a_stride,
+                                  const uint8_t *b, int b_stride,
+                                  unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, &sum);
+  return *sse - ((sum * sum) >> 6);
+}
+
+unsigned int aom_variance16x16_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, &sum);
+  return *sse - (((unsigned int)((int64_t)sum * sum)) >> 8);
+}
+
+unsigned int aom_variance32x32_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, 32, 32, sse, &sum);
+  return *sse - (unsigned int)(((int64_t)sum * sum) >> 10);
+}
+
+unsigned int aom_variance32x64_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum1, sum2;
+  uint32_t sse1, sse2;
+  variance_neon_w8(a, a_stride, b, b_stride, 32, 32, &sse1, &sum1);
+  variance_neon_w8(a + (32 * a_stride), a_stride, b + (32 * b_stride), b_stride,
+                   32, 32, &sse2, &sum2);
+  *sse = sse1 + sse2;
+  sum1 += sum2;
+  return *sse - (unsigned int)(((int64_t)sum1 * sum1) >> 11);
+}
+
+unsigned int aom_variance64x32_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum1, sum2;
+  uint32_t sse1, sse2;
+  variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1);
+  variance_neon_w8(a + (16 * a_stride), a_stride, b + (16 * b_stride), b_stride,
+                   64, 16, &sse2, &sum2);
+  *sse = sse1 + sse2;
+  sum1 += sum2;
+  return *sse - (unsigned int)(((int64_t)sum1 * sum1) >> 11);
+}
+
+unsigned int aom_variance64x64_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum1, sum2;
+  uint32_t sse1, sse2;
+
+  variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1);
+  variance_neon_w8(a + (16 * a_stride), a_stride, b + (16 * b_stride), b_stride,
+                   64, 16, &sse2, &sum2);
+  sse1 += sse2;
+  sum1 += sum2;
+
+  variance_neon_w8(a + (16 * 2 * a_stride), a_stride, b + (16 * 2 * b_stride),
+                   b_stride, 64, 16, &sse2, &sum2);
+  sse1 += sse2;
+  sum1 += sum2;
+
+  variance_neon_w8(a + (16 * 3 * a_stride), a_stride, b + (16 * 3 * b_stride),
+                   b_stride, 64, 16, &sse2, &sum2);
+  *sse = sse1 + sse2;
+  sum1 += sum2;
+  return *sse - (unsigned int)(((int64_t)sum1 * sum1) >> 12);
+}
+
+unsigned int aom_variance16x8_neon(const unsigned char *src_ptr,
+                                   int source_stride,
+                                   const unsigned char *ref_ptr,
+                                   int recon_stride, unsigned int *sse) {
+  int i;
+  int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
+  uint32x2_t d0u32, d10u32;
+  int64x1_t d0s64, d1s64;
+  uint8x16_t q0u8, q1u8, q2u8, q3u8;
+  uint16x8_t q11u16, q12u16, q13u16, q14u16;
+  int32x4_t q8s32, q9s32, q10s32;
+  int64x2_t q0s64, q1s64, q5s64;
+
+  q8s32 = vdupq_n_s32(0);
+  q9s32 = vdupq_n_s32(0);
+  q10s32 = vdupq_n_s32(0);
+
+  for (i = 0; i < 4; i++) {
+    q0u8 = vld1q_u8(src_ptr);
+    src_ptr += source_stride;
+    q1u8 = vld1q_u8(src_ptr);
+    src_ptr += source_stride;
+    __builtin_prefetch(src_ptr);
+
+    q2u8 = vld1q_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    q3u8 = vld1q_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    __builtin_prefetch(ref_ptr);
+
+    q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
+    q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
+    q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
+    q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
+
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
+    q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
+    q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
+
+    d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+    d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
+    q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+    q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+
+    d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+    d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
+    q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
+    q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
+
+    d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
+    d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
+    q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
+    q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
+  }
+
+  q10s32 = vaddq_s32(q10s32, q9s32);
+  q0s64 = vpaddlq_s32(q8s32);
+  q1s64 = vpaddlq_s32(q10s32);
+
+  d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
+  d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+  q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64), vreinterpret_s32_s64(d0s64));
+  vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
+
+  d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
+  d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
+
+  return vget_lane_u32(d0u32, 0);
+}
+
+unsigned int aom_variance8x16_neon(const unsigned char *src_ptr,
+                                   int source_stride,
+                                   const unsigned char *ref_ptr,
+                                   int recon_stride, unsigned int *sse) {
+  int i;
+  uint8x8_t d0u8, d2u8, d4u8, d6u8;
+  int16x4_t d22s16, d23s16, d24s16, d25s16;
+  uint32x2_t d0u32, d10u32;
+  int64x1_t d0s64, d1s64;
+  uint16x8_t q11u16, q12u16;
+  int32x4_t q8s32, q9s32, q10s32;
+  int64x2_t q0s64, q1s64, q5s64;
+
+  q8s32 = vdupq_n_s32(0);
+  q9s32 = vdupq_n_s32(0);
+  q10s32 = vdupq_n_s32(0);
+
+  for (i = 0; i < 8; i++) {
+    d0u8 = vld1_u8(src_ptr);
+    src_ptr += source_stride;
+    d2u8 = vld1_u8(src_ptr);
+    src_ptr += source_stride;
+    __builtin_prefetch(src_ptr);
+
+    d4u8 = vld1_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    d6u8 = vld1_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    __builtin_prefetch(ref_ptr);
+
+    q11u16 = vsubl_u8(d0u8, d4u8);
+    q12u16 = vsubl_u8(d2u8, d6u8);
+
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
+    q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
+    q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
+
+    d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+    d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+    q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
+    q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+    q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+  }
+
+  q10s32 = vaddq_s32(q10s32, q9s32);
+  q0s64 = vpaddlq_s32(q8s32);
+  q1s64 = vpaddlq_s32(q10s32);
+
+  d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
+  d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+  q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64), vreinterpret_s32_s64(d0s64));
+  vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
+
+  d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
+  d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
+
+  return vget_lane_u32(d0u32, 0);
+}
+
+unsigned int aom_mse16x16_neon(const unsigned char *src_ptr, int source_stride,
+                               const unsigned char *ref_ptr, int recon_stride,
+                               unsigned int *sse) {
+  int i;
+  int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
+  int64x1_t d0s64;
+  uint8x16_t q0u8, q1u8, q2u8, q3u8;
+  int32x4_t q7s32, q8s32, q9s32, q10s32;
+  uint16x8_t q11u16, q12u16, q13u16, q14u16;
+  int64x2_t q1s64;
+
+  q7s32 = vdupq_n_s32(0);
+  q8s32 = vdupq_n_s32(0);
+  q9s32 = vdupq_n_s32(0);
+  q10s32 = vdupq_n_s32(0);
+
+  for (i = 0; i < 8; i++) {  // mse16x16_neon_loop
+    q0u8 = vld1q_u8(src_ptr);
+    src_ptr += source_stride;
+    q1u8 = vld1q_u8(src_ptr);
+    src_ptr += source_stride;
+    q2u8 = vld1q_u8(ref_ptr);
+    ref_ptr += recon_stride;
+    q3u8 = vld1q_u8(ref_ptr);
+    ref_ptr += recon_stride;
+
+    q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
+    q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
+    q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
+    q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
+
+    d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
+    d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
+    q7s32 = vmlal_s16(q7s32, d22s16, d22s16);
+    q8s32 = vmlal_s16(q8s32, d23s16, d23s16);
+
+    d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
+    d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
+    q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
+    q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
+
+    d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
+    d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
+    q7s32 = vmlal_s16(q7s32, d26s16, d26s16);
+    q8s32 = vmlal_s16(q8s32, d27s16, d27s16);
+
+    d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
+    d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
+    q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
+    q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
+  }
+
+  q7s32 = vaddq_s32(q7s32, q8s32);
+  q9s32 = vaddq_s32(q9s32, q10s32);
+  q10s32 = vaddq_s32(q7s32, q9s32);
+
+  q1s64 = vpaddlq_s32(q10s32);
+  d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+  vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d0s64), 0);
+  return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
+}
+
+unsigned int aom_get4x4sse_cs_neon(const unsigned char *src_ptr,
+                                   int source_stride,
+                                   const unsigned char *ref_ptr,
+                                   int recon_stride) {
+  int16x4_t d22s16, d24s16, d26s16, d28s16;
+  int64x1_t d0s64;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+  int32x4_t q7s32, q8s32, q9s32, q10s32;
+  uint16x8_t q11u16, q12u16, q13u16, q14u16;
+  int64x2_t q1s64;
+
+  d0u8 = vld1_u8(src_ptr);
+  src_ptr += source_stride;
+  d4u8 = vld1_u8(ref_ptr);
+  ref_ptr += recon_stride;
+  d1u8 = vld1_u8(src_ptr);
+  src_ptr += source_stride;
+  d5u8 = vld1_u8(ref_ptr);
+  ref_ptr += recon_stride;
+  d2u8 = vld1_u8(src_ptr);
+  src_ptr += source_stride;
+  d6u8 = vld1_u8(ref_ptr);
+  ref_ptr += recon_stride;
+  d3u8 = vld1_u8(src_ptr);
+  src_ptr += source_stride;
+  d7u8 = vld1_u8(ref_ptr);
+  ref_ptr += recon_stride;
+
+  q11u16 = vsubl_u8(d0u8, d4u8);
+  q12u16 = vsubl_u8(d1u8, d5u8);
+  q13u16 = vsubl_u8(d2u8, d6u8);
+  q14u16 = vsubl_u8(d3u8, d7u8);
+
+  d22s16 = vget_low_s16(vreinterpretq_s16_u16(q11u16));
+  d24s16 = vget_low_s16(vreinterpretq_s16_u16(q12u16));
+  d26s16 = vget_low_s16(vreinterpretq_s16_u16(q13u16));
+  d28s16 = vget_low_s16(vreinterpretq_s16_u16(q14u16));
+
+  q7s32 = vmull_s16(d22s16, d22s16);
+  q8s32 = vmull_s16(d24s16, d24s16);
+  q9s32 = vmull_s16(d26s16, d26s16);
+  q10s32 = vmull_s16(d28s16, d28s16);
+
+  q7s32 = vaddq_s32(q7s32, q8s32);
+  q9s32 = vaddq_s32(q9s32, q10s32);
+  q9s32 = vaddq_s32(q7s32, q9s32);
+
+  q1s64 = vpaddlq_s32(q9s32);
+  d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
+
+  return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
+}
diff --git a/third_party/aom/aom_dsp/avg.c b/third_party/aom/aom_dsp/avg.c
new file mode 100644
index 0000000000..eb60597052
--- /dev/null
+++ b/third_party/aom/aom_dsp/avg.c
@@ -0,0 +1,232 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <stdlib.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/mem.h"
+
+unsigned int aom_avg_8x8_c(const uint8_t *src, int stride) {
+  int i, j;
+  int sum = 0;
+  for (i = 0; i < 8; ++i, src += stride)
+    for (j = 0; j < 8; sum += src[j], ++j) {
+    }
+
+  return ROUND_POWER_OF_TWO(sum, 6);
+}
+
+unsigned int aom_avg_4x4_c(const uint8_t *src, int stride) {
+  int i, j;
+  int sum = 0;
+  for (i = 0; i < 4; ++i, src += stride)
+    for (j = 0; j < 4; sum += src[j], ++j) {
+    }
+
+  return ROUND_POWER_OF_TWO(sum, 4);
+}
+
+// src_diff: first pass, 9 bit, dynamic range [-255, 255]
+//           second pass, 12 bit, dynamic range [-2040, 2040]
+static void hadamard_col8(const int16_t *src_diff, int src_stride,
+                          int16_t *coeff) {
+  int16_t b0 = src_diff[0 * src_stride] + src_diff[1 * src_stride];
+  int16_t b1 = src_diff[0 * src_stride] - src_diff[1 * src_stride];
+  int16_t b2 = src_diff[2 * src_stride] + src_diff[3 * src_stride];
+  int16_t b3 = src_diff[2 * src_stride] - src_diff[3 * src_stride];
+  int16_t b4 = src_diff[4 * src_stride] + src_diff[5 * src_stride];
+  int16_t b5 = src_diff[4 * src_stride] - src_diff[5 * src_stride];
+  int16_t b6 = src_diff[6 * src_stride] + src_diff[7 * src_stride];
+  int16_t b7 = src_diff[6 * src_stride] - src_diff[7 * src_stride];
+
+  int16_t c0 = b0 + b2;
+  int16_t c1 = b1 + b3;
+  int16_t c2 = b0 - b2;
+  int16_t c3 = b1 - b3;
+  int16_t c4 = b4 + b6;
+  int16_t c5 = b5 + b7;
+  int16_t c6 = b4 - b6;
+  int16_t c7 = b5 - b7;
+
+  coeff[0] = c0 + c4;
+  coeff[7] = c1 + c5;
+  coeff[3] = c2 + c6;
+  coeff[4] = c3 + c7;
+  coeff[2] = c0 - c4;
+  coeff[6] = c1 - c5;
+  coeff[1] = c2 - c6;
+  coeff[5] = c3 - c7;
+}
+
+// The order of the output coeff of the hadamard is not important. For
+// optimization purposes the final transpose may be skipped.
+void aom_hadamard_8x8_c(const int16_t *src_diff, int src_stride,
+                        int16_t *coeff) {
+  int idx;
+  int16_t buffer[64];
+  int16_t *tmp_buf = &buffer[0];
+  for (idx = 0; idx < 8; ++idx) {
+    hadamard_col8(src_diff, src_stride, tmp_buf);  // src_diff: 9 bit
+                                                   // dynamic range [-255, 255]
+    tmp_buf += 8;
+    ++src_diff;
+  }
+
+  tmp_buf = &buffer[0];
+  for (idx = 0; idx < 8; ++idx) {
+    hadamard_col8(tmp_buf, 8, coeff);  // tmp_buf: 12 bit
+                                       // dynamic range [-2040, 2040]
+    coeff += 8;                        // coeff: 15 bit
+                                       // dynamic range [-16320, 16320]
+    ++tmp_buf;
+  }
+}
+
+// In place 16x16 2D Hadamard transform
+void aom_hadamard_16x16_c(const int16_t *src_diff, int src_stride,
+                          int16_t *coeff) {
+  int idx;
+  for (idx = 0; idx < 4; ++idx) {
+    // src_diff: 9 bit, dynamic range [-255, 255]
+    const int16_t *src_ptr =
+        src_diff + (idx >> 1) * 8 * src_stride + (idx & 0x01) * 8;
+    aom_hadamard_8x8_c(src_ptr, src_stride, coeff + idx * 64);
+  }
+
+  // coeff: 15 bit, dynamic range [-16320, 16320]
+  for (idx = 0; idx < 64; ++idx) {
+    int16_t a0 = coeff[0];
+    int16_t a1 = coeff[64];
+    int16_t a2 = coeff[128];
+    int16_t a3 = coeff[192];
+
+    int16_t b0 = (a0 + a1) >> 1;  // (a0 + a1): 16 bit, [-32640, 32640]
+    int16_t b1 = (a0 - a1) >> 1;  // b0-b3: 15 bit, dynamic range
+    int16_t b2 = (a2 + a3) >> 1;  // [-16320, 16320]
+    int16_t b3 = (a2 - a3) >> 1;
+
+    coeff[0] = b0 + b2;  // 16 bit, [-32640, 32640]
+    coeff[64] = b1 + b3;
+    coeff[128] = b0 - b2;
+    coeff[192] = b1 - b3;
+
+    ++coeff;
+  }
+}
+
+// coeff: 16 bits, dynamic range [-32640, 32640].
+// length: value range {16, 64, 256, 1024}.
+int aom_satd_c(const int16_t *coeff, int length) {
+  int i;
+  int satd = 0;
+  for (i = 0; i < length; ++i) satd += abs(coeff[i]);
+
+  // satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
+  return satd;
+}
+
+// Integer projection onto row vectors.
+// height: value range {16, 32, 64}.
+void aom_int_pro_row_c(int16_t hbuf[16], const uint8_t *ref, int ref_stride,
+                       int height) {
+  int idx;
+  const int norm_factor = height >> 1;
+  for (idx = 0; idx < 16; ++idx) {
+    int i;
+    hbuf[idx] = 0;
+    // hbuf[idx]: 14 bit, dynamic range [0, 16320].
+    for (i = 0; i < height; ++i) hbuf[idx] += ref[i * ref_stride];
+    // hbuf[idx]: 9 bit, dynamic range [0, 510].
+    hbuf[idx] /= norm_factor;
+    ++ref;
+  }
+}
+
+// width: value range {16, 32, 64}.
+int16_t aom_int_pro_col_c(const uint8_t *ref, int width) {
+  int idx;
+  int16_t sum = 0;
+  // sum: 14 bit, dynamic range [0, 16320]
+  for (idx = 0; idx < width; ++idx) sum += ref[idx];
+  return sum;
+}
+
+// ref: [0 - 510]
+// src: [0 - 510]
+// bwl: {2, 3, 4}
+int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl) {
+  int i;
+  int width = 4 << bwl;
+  int sse = 0, mean = 0, var;
+
+  for (i = 0; i < width; ++i) {
+    int diff = ref[i] - src[i];  // diff: dynamic range [-510, 510], 10 bits.
+    mean += diff;                // mean: dynamic range 16 bits.
+    sse += diff * diff;          // sse:  dynamic range 26 bits.
+  }
+
+  // (mean * mean): dynamic range 31 bits.
+  var = sse - ((mean * mean) >> (bwl + 2));
+  return var;
+}
+
+void aom_minmax_8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref,
+                      int ref_stride, int *min, int *max) {
+  int i, j;
+  *min = 255;
+  *max = 0;
+  for (i = 0; i < 8; ++i, src += src_stride, ref += ref_stride) {
+    for (j = 0; j < 8; ++j) {
+      int diff = abs(src[j] - ref[j]);
+      *min = diff < *min ? diff : *min;
+      *max = diff > *max ? diff : *max;
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+unsigned int aom_highbd_avg_8x8_c(const uint8_t *src, int stride) {
+  int i, j;
+  int sum = 0;
+  const uint16_t *s = CONVERT_TO_SHORTPTR(src);
+  for (i = 0; i < 8; ++i, s += stride)
+    for (j = 0; j < 8; sum += s[j], ++j) {
+    }
+
+  return ROUND_POWER_OF_TWO(sum, 6);
+}
+
+unsigned int aom_highbd_avg_4x4_c(const uint8_t *src, int stride) {
+  int i, j;
+  int sum = 0;
+  const uint16_t *s = CONVERT_TO_SHORTPTR(src);
+  for (i = 0; i < 4; ++i, s += stride)
+    for (j = 0; j < 4; sum += s[j], ++j) {
+    }
+
+  return ROUND_POWER_OF_TWO(sum, 4);
+}
+
+void aom_highbd_minmax_8x8_c(const uint8_t *s8, int p, const uint8_t *d8,
+                             int dp, int *min, int *max) {
+  int i, j;
+  const uint16_t *s = CONVERT_TO_SHORTPTR(s8);
+  const uint16_t *d = CONVERT_TO_SHORTPTR(d8);
+  *min = 255;
+  *max = 0;
+  for (i = 0; i < 8; ++i, s += p, d += dp) {
+    for (j = 0; j < 8; ++j) {
+      int diff = abs(s[j] - d[j]);
+      *min = diff < *min ? diff : *min;
+      *max = diff > *max ? diff : *max;
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/binary_codes_reader.c b/third_party/aom/aom_dsp/binary_codes_reader.c
new file mode 100644
index 0000000000..96c4cb436d
--- /dev/null
+++ b/third_party/aom/aom_dsp/binary_codes_reader.c
@@ -0,0 +1,117 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/bitreader.h"
+
+#include "av1/common/common.h"
+
+// Inverse recenters a non-negative literal v around a reference r
+static uint16_t inv_recenter_nonneg(uint16_t r, uint16_t v) {
+  if (v > (r << 1))
+    return v;
+  else if ((v & 1) == 0)
+    return (v >> 1) + r;
+  else
+    return r - ((v + 1) >> 1);
+}
+
+// Inverse recenters a non-negative literal v in [0, n-1] around a
+// reference r also in [0, n-1]
+static uint16_t inv_recenter_finite_nonneg(uint16_t n, uint16_t r, uint16_t v) {
+  if ((r << 1) <= n) {
+    return inv_recenter_nonneg(r, v);
+  } else {
+    return n - 1 - inv_recenter_nonneg(n - 1 - r, v);
+  }
+}
+
+int16_t aom_read_primitive_symmetric(aom_reader *r, unsigned int mag_bits) {
+  if (aom_read_bit(r, NULL)) {
+    int s = aom_read_bit(r, NULL);
+    int16_t x = aom_read_literal(r, mag_bits, NULL) + 1;
+    return (s > 0 ? -x : x);
+  } else {
+    return 0;
+  }
+}
+
+uint16_t aom_read_primitive_quniform(aom_reader *r, uint16_t n) {
+  if (n <= 1) return 0;
+  const int l = get_msb(n - 1) + 1;
+  const int m = (1 << l) - n;
+  const int v = aom_read_literal(r, l - 1, NULL);
+  return v < m ? v : (v << 1) - m + aom_read_bit(r, NULL);
+}
+
+uint16_t aom_read_primitive_refbilevel(aom_reader *r, uint16_t n, uint16_t p,
+                                       uint16_t ref) {
+  if (n <= 1) return 0;
+  assert(p > 0 && p <= n);
+  assert(ref < n);
+  int lolimit = ref - p / 2;
+  const int hilimit = lolimit + p - 1;
+  if (lolimit < 0) {
+    lolimit = 0;
+  } else if (hilimit >= n) {
+    lolimit = n - p;
+  }
+  int v;
+  if (aom_read_bit(r, NULL)) {
+    v = aom_read_primitive_quniform(r, p) + lolimit;
+  } else {
+    v = aom_read_primitive_quniform(r, n - p);
+    if (v >= lolimit) v += p;
+  }
+  return v;
+}
+
+// Decode finite subexponential code that for a symbol v in [0, n-1] with
+// parameter k
+uint16_t aom_read_primitive_subexpfin(aom_reader *r, uint16_t n, uint16_t k) {
+  int i = 0;
+  int mk = 0;
+  uint16_t v;
+  while (1) {
+    int b = (i ? k + i - 1 : k);
+    int a = (1 << b);
+    if (n <= mk + 3 * a) {
+      v = aom_read_primitive_quniform(r, n - mk) + mk;
+      break;
+    } else {
+      if (aom_read_bit(r, NULL)) {
+        i = i + 1;
+        mk += a;
+      } else {
+        v = aom_read_literal(r, b, NULL) + mk;
+        break;
+      }
+    }
+  }
+  return v;
+}
+
+// Decode finite subexponential code that for a symbol v in [0, n-1] with
+// parameter k
+// based on a reference ref also in [0, n-1].
+uint16_t aom_read_primitive_refsubexpfin(aom_reader *r, uint16_t n, uint16_t k,
+                                         uint16_t ref) {
+  return inv_recenter_finite_nonneg(n, ref,
+                                    aom_read_primitive_subexpfin(r, n, k));
+}
+
+// Decode finite subexponential code that for a symbol v in [-(n-1), n-1] with
+// parameter k based on a reference ref also in [-(n-1), n-1].
+int16_t aom_read_signed_primitive_refsubexpfin(aom_reader *r, uint16_t n,
+                                               uint16_t k, int16_t ref) {
+  ref += n - 1;
+  const uint16_t scaled_n = (n << 1) - 1;
+  return aom_read_primitive_refsubexpfin(r, scaled_n, k, ref) - n + 1;
+}
diff --git a/third_party/aom/aom_dsp/binary_codes_reader.h b/third_party/aom/aom_dsp/binary_codes_reader.h
new file mode 100644
index 0000000000..738d91da83
--- /dev/null
+++ b/third_party/aom/aom_dsp/binary_codes_reader.h
@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BINARY_CODES_READER_H_
+#define AOM_DSP_BINARY_CODES_READER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/bitreader.h"
+
+int16_t aom_read_primitive_symmetric(aom_reader *r, unsigned int mag_bits);
+
+uint16_t aom_read_primitive_quniform(aom_reader *r, uint16_t n);
+uint16_t aom_read_primitive_refbilevel(aom_reader *r, uint16_t n, uint16_t p,
+                                       uint16_t ref);
+uint16_t aom_read_primitive_subexpfin(aom_reader *r, uint16_t n, uint16_t k);
+uint16_t aom_read_primitive_refsubexpfin(aom_reader *r, uint16_t n, uint16_t k,
+                                         uint16_t ref);
+int16_t aom_read_signed_primitive_refsubexpfin(aom_reader *r, uint16_t n,
+                                               uint16_t k, int16_t ref);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BINARY_CODES_READER_H_
diff --git a/third_party/aom/aom_dsp/binary_codes_writer.c b/third_party/aom/aom_dsp/binary_codes_writer.c
new file mode 100644
index 0000000000..91e807b299
--- /dev/null
+++ b/third_party/aom/aom_dsp/binary_codes_writer.c
@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/bitwriter.h"
+
+#include "av1/common/common.h"
+
+// Recenters a non-negative literal v around a reference r
+static uint16_t recenter_nonneg(uint16_t r, uint16_t v) {
+  if (v > (r << 1))
+    return v;
+  else if (v >= r)
+    return ((v - r) << 1);
+  else
+    return ((r - v) << 1) - 1;
+}
+
+// Recenters a non-negative literal v in [0, n-1] around a
+// reference r also in [0, n-1]
+static uint16_t recenter_finite_nonneg(uint16_t n, uint16_t r, uint16_t v) {
+  if ((r << 1) <= n) {
+    return recenter_nonneg(r, v);
+  } else {
+    return recenter_nonneg(n - 1 - r, n - 1 - v);
+  }
+}
+
+// Codes a symbol v in [-2^mag_bits, 2^mag_bits].
+// mag_bits is number of bits for magnitude. The alphabet is of size
+// 2 * 2^mag_bits + 1, symmetric around 0, where one bit is used to
+// indicate 0 or non-zero, mag_bits bits are used to indicate magnitide
+// and 1 more bit for the sign if non-zero.
+void aom_write_primitive_symmetric(aom_writer *w, int16_t v,
+                                   unsigned int abs_bits) {
+  if (v == 0) {
+    aom_write_bit(w, 0);
+  } else {
+    const int x = abs(v);
+    const int s = v < 0;
+    aom_write_bit(w, 1);
+    aom_write_bit(w, s);
+    aom_write_literal(w, x - 1, abs_bits);
+  }
+}
+
+int aom_count_primitive_symmetric(int16_t v, unsigned int abs_bits) {
+  return (v == 0 ? 1 : abs_bits + 2);
+}
+
+// Encodes a value v in [0, n-1] quasi-uniformly
+void aom_write_primitive_quniform(aom_writer *w, uint16_t n, uint16_t v) {
+  if (n <= 1) return;
+  const int l = get_msb(n - 1) + 1;
+  const int m = (1 << l) - n;
+  if (v < m) {
+    aom_write_literal(w, v, l - 1);
+  } else {
+    aom_write_literal(w, m + ((v - m) >> 1), l - 1);
+    aom_write_bit(w, (v - m) & 1);
+  }
+}
+
+int aom_count_primitive_quniform(uint16_t n, uint16_t v) {
+  if (n <= 1) return 0;
+  const int l = get_msb(n - 1) + 1;
+  const int m = (1 << l) - n;
+  return v < m ? l - 1 : l;
+}
+
+// Encodes a value v in [0, n-1] based on a reference ref also in [0, n-1]
+// The closest p values of v from ref are coded using a p-ary quasi-unoform
+// short code while the remaining n-p values are coded with a longer code.
+void aom_write_primitive_refbilevel(aom_writer *w, uint16_t n, uint16_t p,
+                                    uint16_t ref, uint16_t v) {
+  if (n <= 1) return;
+  assert(p > 0 && p <= n);
+  assert(ref < n);
+  int lolimit = ref - p / 2;
+  int hilimit = lolimit + p - 1;
+  if (lolimit < 0) {
+    lolimit = 0;
+    hilimit = p - 1;
+  } else if (hilimit >= n) {
+    hilimit = n - 1;
+    lolimit = n - p;
+  }
+  if (v >= lolimit && v <= hilimit) {
+    aom_write_bit(w, 1);
+    v = v - lolimit;
+    aom_write_primitive_quniform(w, p, v);
+  } else {
+    aom_write_bit(w, 0);
+    if (v > hilimit) v -= p;
+    aom_write_primitive_quniform(w, n - p, v);
+  }
+}
+
+int aom_count_primitive_refbilevel(uint16_t n, uint16_t p, uint16_t ref,
+                                   uint16_t v) {
+  if (n <= 1) return 0;
+  assert(p > 0 && p <= n);
+  assert(ref < n);
+  int lolimit = ref - p / 2;
+  int hilimit = lolimit + p - 1;
+  if (lolimit < 0) {
+    lolimit = 0;
+    hilimit = p - 1;
+  } else if (hilimit >= n) {
+    hilimit = n - 1;
+    lolimit = n - p;
+  }
+  int count = 0;
+  if (v >= lolimit && v <= hilimit) {
+    count++;
+    v = v - lolimit;
+    count += aom_count_primitive_quniform(p, v);
+  } else {
+    count++;
+    if (v > hilimit) v -= p;
+    count += aom_count_primitive_quniform(n - p, v);
+  }
+  return count;
+}
+
+// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k
+void aom_write_primitive_subexpfin(aom_writer *w, uint16_t n, uint16_t k,
+                                   uint16_t v) {
+  int i = 0;
+  int mk = 0;
+  while (1) {
+    int b = (i ? k + i - 1 : k);
+    int a = (1 << b);
+    if (n <= mk + 3 * a) {
+      aom_write_primitive_quniform(w, n - mk, v - mk);
+      break;
+    } else {
+      int t = (v >= mk + a);
+      aom_write_bit(w, t);
+      if (t) {
+        i = i + 1;
+        mk += a;
+      } else {
+        aom_write_literal(w, v - mk, b);
+        break;
+      }
+    }
+  }
+}
+
+int aom_count_primitive_subexpfin(uint16_t n, uint16_t k, uint16_t v) {
+  int count = 0;
+  int i = 0;
+  int mk = 0;
+  while (1) {
+    int b = (i ? k + i - 1 : k);
+    int a = (1 << b);
+    if (n <= mk + 3 * a) {
+      count += aom_count_primitive_quniform(n - mk, v - mk);
+      break;
+    } else {
+      int t = (v >= mk + a);
+      count++;
+      if (t) {
+        i = i + 1;
+        mk += a;
+      } else {
+        count += b;
+        break;
+      }
+    }
+  }
+  return count;
+}
+
+// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k
+// based on a reference ref also in [0, n-1].
+// Recenters symbol around r first and then uses a finite subexponential code.
+void aom_write_primitive_refsubexpfin(aom_writer *w, uint16_t n, uint16_t k,
+                                      int16_t ref, int16_t v) {
+  aom_write_primitive_subexpfin(w, n, k, recenter_finite_nonneg(n, ref, v));
+}
+
+void aom_write_signed_primitive_refsubexpfin(aom_writer *w, uint16_t n,
+                                             uint16_t k, uint16_t ref,
+                                             uint16_t v) {
+  ref += n - 1;
+  v += n - 1;
+  const uint16_t scaled_n = (n << 1) - 1;
+  aom_write_primitive_refsubexpfin(w, scaled_n, k, ref, v);
+}
+
+int aom_count_primitive_refsubexpfin(uint16_t n, uint16_t k, uint16_t ref,
+                                     uint16_t v) {
+  return aom_count_primitive_subexpfin(n, k, recenter_finite_nonneg(n, ref, v));
+}
+
+int aom_count_signed_primitive_refsubexpfin(uint16_t n, uint16_t k, int16_t ref,
+                                            int16_t v) {
+  ref += n - 1;
+  v += n - 1;
+  const uint16_t scaled_n = (n << 1) - 1;
+  return aom_count_primitive_refsubexpfin(scaled_n, k, ref, v);
+}
diff --git a/third_party/aom/aom_dsp/binary_codes_writer.h b/third_party/aom/aom_dsp/binary_codes_writer.h
new file mode 100644
index 0000000000..ab5ccbf15c
--- /dev/null
+++ b/third_party/aom/aom_dsp/binary_codes_writer.h
@@ -0,0 +1,70 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BINARY_CODES_WRITER_H_
+#define AOM_DSP_BINARY_CODES_WRITER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/bitwriter.h"
+
+// Codes a symbol v in [-2^mag_bits, 2^mag_bits]
+// mag_bits is number of bits for magnitude. The alphabet is of size
+// 2 * 2^mag_bits + 1, symmetric around 0, where one bit is used to
+// indicate 0 or non-zero, mag_bits bits are used to indicate magnitide
+// and 1 more bit for the sign if non-zero.
+void aom_write_primitive_symmetric(aom_writer *w, int16_t v,
+                                   unsigned int mag_bits);
+
+// Encodes a value v in [0, n-1] quasi-uniformly
+void aom_write_primitive_quniform(aom_writer *w, uint16_t n, uint16_t v);
+
+// Encodes a value v in [0, n-1] based on a reference ref also in [0, n-1]
+// The closest p values of v from ref are coded using a p-ary quasi-unoform
+// short code while the remaining n-p values are coded with a longer code.
+void aom_write_primitive_refbilevel(aom_writer *w, uint16_t n, uint16_t p,
+                                    uint16_t ref, uint16_t v);
+
+// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k
+void aom_write_primitive_subexpfin(aom_writer *w, uint16_t n, uint16_t k,
+                                   uint16_t v);
+
+// Finite subexponential code that codes a symbol v in [0, n-1] with parameter k
+// based on a reference ref also in [0, n-1].
+void aom_write_primitive_refsubexpfin(aom_writer *w, uint16_t n, uint16_t k,
+                                      uint16_t ref, uint16_t v);
+
+// Finite subexponential code that codes a symbol v in [-(n-1), n-1] with
+// parameter k based on a reference ref also in [-(n-1), n-1].
+void aom_write_signed_primitive_refsubexpfin(aom_writer *w, uint16_t n,
+                                             uint16_t k, int16_t ref,
+                                             int16_t v);
+
+// Functions that counts bits for the above primitives
+int aom_count_primitive_symmetric(int16_t v, unsigned int mag_bits);
+int aom_count_primitive_quniform(uint16_t n, uint16_t v);
+int aom_count_primitive_refbilevel(uint16_t n, uint16_t p, uint16_t ref,
+                                   uint16_t v);
+int aom_count_primitive_subexpfin(uint16_t n, uint16_t k, uint16_t v);
+int aom_count_primitive_refsubexpfin(uint16_t n, uint16_t k, uint16_t ref,
+                                     uint16_t v);
+int aom_count_signed_primitive_refsubexpfin(uint16_t n, uint16_t k, int16_t ref,
+                                            int16_t v);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BINARY_CODES_WRITER_H_
diff --git a/third_party/aom/aom_dsp/bitreader.h b/third_party/aom/aom_dsp/bitreader.h
new file mode 100644
index 0000000000..9cd34dd483
--- /dev/null
+++ b/third_party/aom/aom_dsp/bitreader.h
@@ -0,0 +1,276 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BITREADER_H_
+#define AOM_DSP_BITREADER_H_
+
+#include <assert.h>
+#include <limits.h>
+
+#include "./aom_config.h"
+#if CONFIG_EC_ADAPT && !CONFIG_EC_MULTISYMBOL
+#error "CONFIG_EC_ADAPT is enabled without enabling CONFIG_EC_MULTISYMBOL."
+#endif
+
+#include "aom/aomdx.h"
+#include "aom/aom_integer.h"
+#if CONFIG_ANS
+#include "aom_dsp/ansreader.h"
+#elif CONFIG_DAALA_EC
+#include "aom_dsp/daalaboolreader.h"
+#else
+#include "aom_dsp/dkboolreader.h"
+#endif
+#include "aom_dsp/prob.h"
+#include "av1/common/odintrin.h"
+
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#define ACCT_STR_NAME acct_str
+#define ACCT_STR_PARAM , const char *ACCT_STR_NAME
+#define ACCT_STR_ARG(s) , s
+#else
+#define ACCT_STR_PARAM
+#define ACCT_STR_ARG(s)
+#endif
+
+#define aom_read(r, prob, ACCT_STR_NAME) \
+  aom_read_(r, prob ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_read_bit(r, ACCT_STR_NAME) \
+  aom_read_bit_(r ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_read_tree(r, tree, probs, ACCT_STR_NAME) \
+  aom_read_tree_(r, tree, probs ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_read_literal(r, bits, ACCT_STR_NAME) \
+  aom_read_literal_(r, bits ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_read_cdf(r, cdf, nsymbs, ACCT_STR_NAME) \
+  aom_read_cdf_(r, cdf, nsymbs ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_read_symbol(r, cdf, nsymbs, ACCT_STR_NAME) \
+  aom_read_symbol_(r, cdf, nsymbs ACCT_STR_ARG(ACCT_STR_NAME))
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_ANS
+typedef struct AnsDecoder aom_reader;
+#elif CONFIG_DAALA_EC
+typedef struct daala_reader aom_reader;
+#else
+typedef struct aom_dk_reader aom_reader;
+#endif
+
+static INLINE int aom_reader_init(aom_reader *r, const uint8_t *buffer,
+                                  size_t size, aom_decrypt_cb decrypt_cb,
+                                  void *decrypt_state) {
+#if CONFIG_ANS
+  (void)decrypt_cb;
+  (void)decrypt_state;
+  if (size > INT_MAX) return 1;
+  return ans_read_init(r, buffer, (int)size);
+#elif CONFIG_DAALA_EC
+  (void)decrypt_cb;
+  (void)decrypt_state;
+  return aom_daala_reader_init(r, buffer, (int)size);
+#else
+  return aom_dk_reader_init(r, buffer, size, decrypt_cb, decrypt_state);
+#endif
+}
+
+static INLINE const uint8_t *aom_reader_find_end(aom_reader *r) {
+#if CONFIG_ANS
+  (void)r;
+  assert(0 && "Use the raw buffer size with ANS");
+  return NULL;
+#elif CONFIG_DAALA_EC
+  return aom_daala_reader_find_end(r);
+#else
+  return aom_dk_reader_find_end(r);
+#endif
+}
+
+static INLINE int aom_reader_has_error(aom_reader *r) {
+#if CONFIG_ANS
+  return ans_reader_has_error(r);
+#elif CONFIG_DAALA_EC
+  return aom_daala_reader_has_error(r);
+#else
+  return aom_dk_reader_has_error(r);
+#endif
+}
+
+// Returns the position in the bit reader in bits.
+static INLINE uint32_t aom_reader_tell(const aom_reader *r) {
+#if CONFIG_ANS
+  (void)r;
+  assert(0 && "aom_reader_tell() is unimplemented for ANS");
+  return 0;
+#elif CONFIG_DAALA_EC
+  return aom_daala_reader_tell(r);
+#else
+  return aom_dk_reader_tell(r);
+#endif
+}
+
+// Returns the position in the bit reader in 1/8th bits.
+static INLINE uint32_t aom_reader_tell_frac(const aom_reader *r) {
+#if CONFIG_ANS
+  (void)r;
+  assert(0 && "aom_reader_tell_frac() is unimplemented for ANS");
+  return 0;
+#elif CONFIG_DAALA_EC
+  return aom_daala_reader_tell_frac(r);
+#else
+  return aom_dk_reader_tell_frac(r);
+#endif
+}
+
+#if CONFIG_ACCOUNTING
+static INLINE void aom_process_accounting(const aom_reader *r ACCT_STR_PARAM) {
+  if (r->accounting != NULL) {
+    uint32_t tell_frac;
+    tell_frac = aom_reader_tell_frac(r);
+    aom_accounting_record(r->accounting, ACCT_STR_NAME,
+                          tell_frac - r->accounting->last_tell_frac);
+    r->accounting->last_tell_frac = tell_frac;
+  }
+}
+
+static INLINE void aom_update_symb_counts(const aom_reader *r, int is_binary) {
+  if (r->accounting != NULL) {
+    r->accounting->syms.num_multi_syms += !is_binary;
+    r->accounting->syms.num_binary_syms += !!is_binary;
+  }
+}
+#endif
+
+static INLINE int aom_read_(aom_reader *r, int prob ACCT_STR_PARAM) {
+  int ret;
+#if CONFIG_ANS
+  ret = rabs_read(r, prob);
+#elif CONFIG_DAALA_EC
+  ret = aom_daala_read(r, prob);
+#else
+  ret = aom_dk_read(r, prob);
+#endif
+#if CONFIG_ACCOUNTING
+  if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME);
+  aom_update_symb_counts(r, 1);
+#endif
+  return ret;
+}
+
+static INLINE int aom_read_bit_(aom_reader *r ACCT_STR_PARAM) {
+  int ret;
+#if CONFIG_ANS
+  ret = rabs_read_bit(r);  // Non trivial optimization at half probability
+#elif CONFIG_DAALA_EC && CONFIG_RAWBITS
+  // Note this uses raw bits and is not the same as aom_daala_read(r, 128);
+  // Calls to this function are omitted from raw symbol accounting.
+  ret = aom_daala_read_bit(r);
+#else
+  ret = aom_read(r, 128, NULL);  // aom_prob_half
+#endif
+#if CONFIG_ACCOUNTING
+  if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME);
+#endif
+  return ret;
+}
+
+static INLINE int aom_read_literal_(aom_reader *r, int bits ACCT_STR_PARAM) {
+  int literal = 0, bit;
+
+  for (bit = bits - 1; bit >= 0; bit--) literal |= aom_read_bit(r, NULL) << bit;
+#if CONFIG_ACCOUNTING
+  if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME);
+#endif
+  return literal;
+}
+
+static INLINE int aom_read_tree_as_bits(aom_reader *r,
+                                        const aom_tree_index *tree,
+                                        const aom_prob *probs) {
+  aom_tree_index i = 0;
+
+  while ((i = tree[i + aom_read(r, probs[i >> 1], NULL)]) > 0) continue;
+  return -i;
+}
+
+#if CONFIG_EC_MULTISYMBOL
+static INLINE int aom_read_cdf_(aom_reader *r, const aom_cdf_prob *cdf,
+                                int nsymbs ACCT_STR_PARAM) {
+  int ret;
+#if CONFIG_ANS
+  (void)nsymbs;
+  ret = rans_read(r, cdf);
+#elif CONFIG_DAALA_EC
+  ret = daala_read_symbol(r, cdf, nsymbs);
+#else
+#error \
+    "CONFIG_EC_MULTISYMBOL is selected without a valid backing entropy " \
+  "coder. Enable daala_ec or ans for a valid configuration."
+#endif
+
+#if CONFIG_ACCOUNTING
+  if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME);
+  aom_update_symb_counts(r, (nsymbs == 2));
+#endif
+  return ret;
+}
+
+static INLINE int aom_read_symbol_(aom_reader *r, aom_cdf_prob *cdf,
+                                   int nsymbs ACCT_STR_PARAM) {
+  int ret;
+  ret = aom_read_cdf(r, cdf, nsymbs, ACCT_STR_NAME);
+#if CONFIG_EC_ADAPT
+  update_cdf(cdf, ret, nsymbs);
+#endif
+  return ret;
+}
+
+static INLINE int aom_read_tree_as_cdf(aom_reader *r,
+                                       const aom_tree_index *tree,
+                                       const aom_prob *probs) {
+  aom_tree_index i = 0;
+  do {
+    aom_cdf_prob cdf[16];
+    aom_tree_index index[16];
+    int path[16];
+    int dist[16];
+    int nsymbs;
+    int symb;
+    nsymbs = tree_to_cdf(tree, probs, i, cdf, index, path, dist);
+    symb = aom_read_cdf(r, cdf, nsymbs, NULL);
+    OD_ASSERT(symb >= 0 && symb < nsymbs);
+    i = index[symb];
+  } while (i > 0);
+  return -i;
+}
+#endif  // CONFIG_EC_MULTISYMBOL
+
+static INLINE int aom_read_tree_(aom_reader *r, const aom_tree_index *tree,
+                                 const aom_prob *probs ACCT_STR_PARAM) {
+  int ret;
+#if CONFIG_EC_MULTISYMBOL
+  ret = aom_read_tree_as_cdf(r, tree, probs);
+#else
+  ret = aom_read_tree_as_bits(r, tree, probs);
+#endif
+#if CONFIG_ACCOUNTING
+  if (ACCT_STR_NAME) aom_process_accounting(r, ACCT_STR_NAME);
+#endif
+  return ret;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BITREADER_H_
diff --git a/third_party/aom/aom_dsp/bitreader_buffer.c b/third_party/aom/aom_dsp/bitreader_buffer.c
new file mode 100644
index 0000000000..009682b4c8
--- /dev/null
+++ b/third_party/aom/aom_dsp/bitreader_buffer.c
@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "./aom_config.h"
+#include "./bitreader_buffer.h"
+
+size_t aom_rb_bytes_read(struct aom_read_bit_buffer *rb) {
+  return (rb->bit_offset + 7) >> 3;
+}
+
+int aom_rb_read_bit(struct aom_read_bit_buffer *rb) {
+  const uint32_t off = rb->bit_offset;
+  const uint32_t p = off >> 3;
+  const int q = 7 - (int)(off & 0x7);
+  if (rb->bit_buffer + p < rb->bit_buffer_end) {
+    const int bit = (rb->bit_buffer[p] >> q) & 1;
+    rb->bit_offset = off + 1;
+    return bit;
+  } else {
+    rb->error_handler(rb->error_handler_data);
+    return 0;
+  }
+}
+
+int aom_rb_read_literal(struct aom_read_bit_buffer *rb, int bits) {
+  int value = 0, bit;
+  for (bit = bits - 1; bit >= 0; bit--) value |= aom_rb_read_bit(rb) << bit;
+  return value;
+}
+
+int aom_rb_read_signed_literal(struct aom_read_bit_buffer *rb, int bits) {
+  const int value = aom_rb_read_literal(rb, bits);
+  return aom_rb_read_bit(rb) ? -value : value;
+}
+
+int aom_rb_read_inv_signed_literal(struct aom_read_bit_buffer *rb, int bits) {
+  const int nbits = sizeof(unsigned) * 8 - bits - 1;
+  const unsigned value = (unsigned)aom_rb_read_literal(rb, bits + 1) << nbits;
+  return ((int)value) >> nbits;
+}
diff --git a/third_party/aom/aom_dsp/bitreader_buffer.h b/third_party/aom/aom_dsp/bitreader_buffer.h
new file mode 100644
index 0000000000..22187357e0
--- /dev/null
+++ b/third_party/aom/aom_dsp/bitreader_buffer.h
@@ -0,0 +1,48 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BITREADER_BUFFER_H_
+#define AOM_DSP_BITREADER_BUFFER_H_
+
+#include <limits.h>
+
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef void (*aom_rb_error_handler)(void *data);
+
+struct aom_read_bit_buffer {
+  const uint8_t *bit_buffer;
+  const uint8_t *bit_buffer_end;
+  uint32_t bit_offset;
+
+  void *error_handler_data;
+  aom_rb_error_handler error_handler;
+};
+
+size_t aom_rb_bytes_read(struct aom_read_bit_buffer *rb);
+
+int aom_rb_read_bit(struct aom_read_bit_buffer *rb);
+
+int aom_rb_read_literal(struct aom_read_bit_buffer *rb, int bits);
+
+int aom_rb_read_signed_literal(struct aom_read_bit_buffer *rb, int bits);
+
+int aom_rb_read_inv_signed_literal(struct aom_read_bit_buffer *rb, int bits);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BITREADER_BUFFER_H_
diff --git a/third_party/aom/aom_dsp/bitwriter.h b/third_party/aom/aom_dsp/bitwriter.h
new file mode 100644
index 0000000000..6e3fac2607
--- /dev/null
+++ b/third_party/aom/aom_dsp/bitwriter.h
@@ -0,0 +1,255 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BITWRITER_H_
+#define AOM_DSP_BITWRITER_H_
+
+#include <assert.h>
+#include "./aom_config.h"
+#if CONFIG_EC_ADAPT && !CONFIG_EC_MULTISYMBOL
+#error "CONFIG_EC_ADAPT is enabled without enabling CONFIG_EC_MULTISYMBOL"
+#endif
+
+#if CONFIG_ANS
+#include "aom_dsp/buf_ans.h"
+#elif CONFIG_DAALA_EC
+#include "aom_dsp/daalaboolwriter.h"
+#else
+#include "aom_dsp/dkboolwriter.h"
+#endif
+#include "aom_dsp/prob.h"
+
+#if CONFIG_RD_DEBUG
+#include "av1/common/blockd.h"
+#include "av1/encoder/cost.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_ANS
+typedef struct BufAnsCoder aom_writer;
+#elif CONFIG_DAALA_EC
+typedef struct daala_writer aom_writer;
+#else
+typedef struct aom_dk_writer aom_writer;
+#endif
+
+typedef struct TOKEN_STATS {
+  int cost;
+#if CONFIG_VAR_TX
+#if CONFIG_RD_DEBUG
+  int txb_coeff_cost_map[TXB_COEFF_COST_MAP_SIZE][TXB_COEFF_COST_MAP_SIZE];
+#endif
+#endif
+} TOKEN_STATS;
+
+static INLINE void init_token_stats(TOKEN_STATS *token_stats) {
+#if CONFIG_VAR_TX
+#if CONFIG_RD_DEBUG
+  int r, c;
+  for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r) {
+    for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c) {
+      token_stats->txb_coeff_cost_map[r][c] = 0;
+    }
+  }
+#endif
+#endif
+  token_stats->cost = 0;
+}
+
+static INLINE void aom_start_encode(aom_writer *bc, uint8_t *buffer) {
+#if CONFIG_ANS
+  (void)bc;
+  (void)buffer;
+  assert(0 && "buf_ans requires a more complicated startup procedure");
+#elif CONFIG_DAALA_EC
+  aom_daala_start_encode(bc, buffer);
+#else
+  aom_dk_start_encode(bc, buffer);
+#endif
+}
+
+static INLINE void aom_stop_encode(aom_writer *bc) {
+#if CONFIG_ANS
+  (void)bc;
+  assert(0 && "buf_ans requires a more complicated shutdown procedure");
+#elif CONFIG_DAALA_EC
+  aom_daala_stop_encode(bc);
+#else
+  aom_dk_stop_encode(bc);
+#endif
+}
+
+static INLINE void aom_write(aom_writer *br, int bit, int probability) {
+#if CONFIG_ANS
+  buf_rabs_write(br, bit, probability);
+#elif CONFIG_DAALA_EC
+  aom_daala_write(br, bit, probability);
+#else
+  aom_dk_write(br, bit, probability);
+#endif
+}
+
+static INLINE void aom_write_record(aom_writer *br, int bit, int probability,
+                                    TOKEN_STATS *token_stats) {
+  aom_write(br, bit, probability);
+#if CONFIG_RD_DEBUG
+  token_stats->cost += av1_cost_bit(probability, bit);
+#else
+  (void)token_stats;
+#endif
+}
+
+static INLINE void aom_write_bit(aom_writer *w, int bit) {
+#if CONFIG_ANS
+  buf_rabs_write_bit(w, bit);
+#elif CONFIG_DAALA_EC && CONFIG_RAWBITS
+  // Note this uses raw bits and is not the same as aom_daala_write(r, 128);
+  aom_daala_write_bit(w, bit);
+#else
+  aom_write(w, bit, 128);  // aom_prob_half
+#endif
+}
+
+static INLINE void aom_write_bit_record(aom_writer *w, int bit,
+                                        TOKEN_STATS *token_stats) {
+  aom_write_bit(w, bit);
+#if CONFIG_RD_DEBUG
+  token_stats->cost += av1_cost_bit(128, bit);  // aom_prob_half
+#else
+  (void)token_stats;
+#endif
+}
+
+static INLINE void aom_write_literal(aom_writer *w, int data, int bits) {
+  int bit;
+
+  for (bit = bits - 1; bit >= 0; bit--) aom_write_bit(w, 1 & (data >> bit));
+}
+
+static INLINE void aom_write_tree_as_bits(aom_writer *w,
+                                          const aom_tree_index *tr,
+                                          const aom_prob *probs, int bits,
+                                          int len, aom_tree_index i) {
+  do {
+    const int bit = (bits >> --len) & 1;
+    aom_write(w, bit, probs[i >> 1]);
+    i = tr[i + bit];
+  } while (len);
+}
+
+static INLINE void aom_write_tree_as_bits_record(
+    aom_writer *w, const aom_tree_index *tr, const aom_prob *probs, int bits,
+    int len, aom_tree_index i, TOKEN_STATS *token_stats) {
+  do {
+    const int bit = (bits >> --len) & 1;
+    aom_write_record(w, bit, probs[i >> 1], token_stats);
+    i = tr[i + bit];
+  } while (len);
+}
+
+#if CONFIG_EC_MULTISYMBOL
+static INLINE void aom_write_cdf(aom_writer *w, int symb,
+                                 const aom_cdf_prob *cdf, int nsymbs) {
+#if CONFIG_ANS
+  (void)nsymbs;
+  assert(cdf);
+  const aom_cdf_prob cum_prob = symb > 0 ? cdf[symb - 1] : 0;
+  const aom_cdf_prob prob = cdf[symb] - cum_prob;
+  buf_rans_write(w, cum_prob, prob);
+#elif CONFIG_DAALA_EC
+  daala_write_symbol(w, symb, cdf, nsymbs);
+#else
+#error \
+    "CONFIG_EC_MULTISYMBOL is selected without a valid backing entropy " \
+  "coder. Enable daala_ec or ans for a valid configuration."
+#endif
+}
+
+static INLINE void aom_write_symbol(aom_writer *w, int symb, aom_cdf_prob *cdf,
+                                    int nsymbs) {
+  aom_write_cdf(w, symb, cdf, nsymbs);
+#if CONFIG_EC_ADAPT
+  update_cdf(cdf, symb, nsymbs);
+#endif
+}
+
+static INLINE void aom_write_tree_as_cdf(aom_writer *w,
+                                         const aom_tree_index *tree,
+                                         const aom_prob *probs, int bits,
+                                         int len, aom_tree_index i) {
+  aom_tree_index root;
+  root = i;
+  do {
+    aom_cdf_prob cdf[16];
+    aom_tree_index index[16];
+    int path[16];
+    int dist[16];
+    int nsymbs;
+    int symb;
+    int j;
+    /* Compute the CDF of the binary tree using the given probabilities. */
+    nsymbs = tree_to_cdf(tree, probs, root, cdf, index, path, dist);
+    /* Find the symbol to code. */
+    symb = -1;
+    for (j = 0; j < nsymbs; j++) {
+      /* If this symbol codes a leaf node,  */
+      if (index[j] <= 0) {
+        if (len == dist[j] && path[j] == bits) {
+          symb = j;
+          break;
+        }
+      } else {
+        if (len > dist[j] && path[j] == bits >> (len - dist[j])) {
+          symb = j;
+          break;
+        }
+      }
+    }
+    OD_ASSERT(symb != -1);
+    aom_write_cdf(w, symb, cdf, nsymbs);
+    bits &= (1 << (len - dist[symb])) - 1;
+    len -= dist[symb];
+  } while (len);
+}
+
+#endif  // CONFIG_EC_MULTISYMBOL
+
+static INLINE void aom_write_tree(aom_writer *w, const aom_tree_index *tree,
+                                  const aom_prob *probs, int bits, int len,
+                                  aom_tree_index i) {
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_tree_as_cdf(w, tree, probs, bits, len, i);
+#else
+  aom_write_tree_as_bits(w, tree, probs, bits, len, i);
+#endif
+}
+
+static INLINE void aom_write_tree_record(aom_writer *w,
+                                         const aom_tree_index *tree,
+                                         const aom_prob *probs, int bits,
+                                         int len, aom_tree_index i,
+                                         TOKEN_STATS *token_stats) {
+#if CONFIG_EC_MULTISYMBOL
+  (void)token_stats;
+  aom_write_tree_as_cdf(w, tree, probs, bits, len, i);
+#else
+  aom_write_tree_as_bits_record(w, tree, probs, bits, len, i, token_stats);
+#endif
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BITWRITER_H_
diff --git a/third_party/aom/aom_dsp/bitwriter_buffer.c b/third_party/aom/aom_dsp/bitwriter_buffer.c
new file mode 100644
index 0000000000..1b3dd2913e
--- /dev/null
+++ b/third_party/aom/aom_dsp/bitwriter_buffer.c
@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+#include <stdlib.h>
+
+#include "./aom_config.h"
+#include "./bitwriter_buffer.h"
+
+uint32_t aom_wb_bytes_written(const struct aom_write_bit_buffer *wb) {
+  return wb->bit_offset / CHAR_BIT + (wb->bit_offset % CHAR_BIT > 0);
+}
+
+void aom_wb_write_bit(struct aom_write_bit_buffer *wb, int bit) {
+  const int off = (int)wb->bit_offset;
+  const int p = off / CHAR_BIT;
+  const int q = CHAR_BIT - 1 - off % CHAR_BIT;
+  if (q == CHAR_BIT - 1) {
+    // Zero next char and write bit
+    wb->bit_buffer[p] = bit << q;
+  } else {
+    wb->bit_buffer[p] &= ~(1 << q);
+    wb->bit_buffer[p] |= bit << q;
+  }
+  wb->bit_offset = off + 1;
+}
+
+void aom_wb_overwrite_bit(struct aom_write_bit_buffer *wb, int bit) {
+  // Do not zero bytes but overwrite exisiting values
+  const int off = (int)wb->bit_offset;
+  const int p = off / CHAR_BIT;
+  const int q = CHAR_BIT - 1 - off % CHAR_BIT;
+  wb->bit_buffer[p] &= ~(1 << q);
+  wb->bit_buffer[p] |= bit << q;
+  wb->bit_offset = off + 1;
+}
+
+void aom_wb_write_literal(struct aom_write_bit_buffer *wb, int data, int bits) {
+  int bit;
+  for (bit = bits - 1; bit >= 0; bit--) aom_wb_write_bit(wb, (data >> bit) & 1);
+}
+
+void aom_wb_overwrite_literal(struct aom_write_bit_buffer *wb, int data,
+                              int bits) {
+  int bit;
+  for (bit = bits - 1; bit >= 0; bit--)
+    aom_wb_overwrite_bit(wb, (data >> bit) & 1);
+}
+
+void aom_wb_write_inv_signed_literal(struct aom_write_bit_buffer *wb, int data,
+                                     int bits) {
+  aom_wb_write_literal(wb, data, bits + 1);
+}
diff --git a/third_party/aom/aom_dsp/bitwriter_buffer.h b/third_party/aom/aom_dsp/bitwriter_buffer.h
new file mode 100644
index 0000000000..1f23dc857b
--- /dev/null
+++ b/third_party/aom/aom_dsp/bitwriter_buffer.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BITWRITER_BUFFER_H_
+#define AOM_DSP_BITWRITER_BUFFER_H_
+
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct aom_write_bit_buffer {
+  uint8_t *bit_buffer;
+  uint32_t bit_offset;
+};
+
+uint32_t aom_wb_bytes_written(const struct aom_write_bit_buffer *wb);
+
+void aom_wb_write_bit(struct aom_write_bit_buffer *wb, int bit);
+
+void aom_wb_overwrite_bit(struct aom_write_bit_buffer *wb, int bit);
+
+void aom_wb_write_literal(struct aom_write_bit_buffer *wb, int data, int bits);
+
+void aom_wb_overwrite_literal(struct aom_write_bit_buffer *wb, int data,
+                              int bits);
+
+void aom_wb_write_inv_signed_literal(struct aom_write_bit_buffer *wb, int data,
+                                     int bits);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_BITWRITER_BUFFER_H_
diff --git a/third_party/aom/aom_dsp/blend.h b/third_party/aom/aom_dsp/blend.h
new file mode 100644
index 0000000000..e5297ff83b
--- /dev/null
+++ b/third_party/aom/aom_dsp/blend.h
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BLEND_H_
+#define AOM_DSP_BLEND_H_
+
+#include "aom_ports/mem.h"
+
+// Various blending functions and macros.
+// See also the aom_blend_* functions in aom_dsp_rtcd.h
+
+// Alpha blending with alpha values from the range [0, 64], where 64
+// means use the first input and 0 means use the second input.
+
+#define AOM_BLEND_A64_ROUND_BITS 6
+#define AOM_BLEND_A64_MAX_ALPHA (1 << AOM_BLEND_A64_ROUND_BITS)  // 64
+
+#define AOM_BLEND_A64(a, v0, v1)                                          \
+  ROUND_POWER_OF_TWO((a) * (v0) + (AOM_BLEND_A64_MAX_ALPHA - (a)) * (v1), \
+                     AOM_BLEND_A64_ROUND_BITS)
+
+// Alpha blending with alpha values from the range [0, 256], where 256
+// means use the first input and 0 means use the second input.
+#define AOM_BLEND_A256_ROUND_BITS 8
+#define AOM_BLEND_A256_MAX_ALPHA (1 << AOM_BLEND_A256_ROUND_BITS)  // 256
+
+#define AOM_BLEND_A256(a, v0, v1)                                          \
+  ROUND_POWER_OF_TWO((a) * (v0) + (AOM_BLEND_A256_MAX_ALPHA - (a)) * (v1), \
+                     AOM_BLEND_A256_ROUND_BITS)
+
+// Blending by averaging.
+#define AOM_BLEND_AVG(v0, v1) ROUND_POWER_OF_TWO((v0) + (v1), 1)
+
+#endif  // AOM_DSP_BLEND_H_
diff --git a/third_party/aom/aom_dsp/blend_a64_hmask.c b/third_party/aom/aom_dsp/blend_a64_hmask.c
new file mode 100644
index 0000000000..99b4b8a599
--- /dev/null
+++ b/third_party/aom/aom_dsp/blend_a64_hmask.c
@@ -0,0 +1,71 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "./aom_dsp_rtcd.h"
+
+void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride,
+                           const uint8_t *src0, uint32_t src0_stride,
+                           const uint8_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int h, int w) {
+  int i, j;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(
+          mask[j], src0[i * src0_stride + j], src1[i * src1_stride + j]);
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_blend_a64_hmask_c(uint8_t *dst_8, uint32_t dst_stride,
+                                  const uint8_t *src0_8, uint32_t src0_stride,
+                                  const uint8_t *src1_8, uint32_t src1_stride,
+                                  const uint8_t *mask, int h, int w, int bd) {
+  int i, j;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8);
+  const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8);
+  const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8);
+  (void)bd;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(
+          mask[j], src0[i * src0_stride + j], src1[i * src1_stride + j]);
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/blend_a64_mask.c b/third_party/aom/aom_dsp/blend_a64_mask.c
new file mode 100644
index 0000000000..3e15542c93
--- /dev/null
+++ b/third_party/aom/aom_dsp/blend_a64_mask.c
@@ -0,0 +1,145 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/blend.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+#include "./aom_dsp_rtcd.h"
+
+// Blending with alpha mask. Mask values come from the range [0, 64],
+// as described for AOM_BLEND_A64 in aom_dsp/blend.h. src0 or src1 can
+// be the same as dst, or dst can be different from both sources.
+
+void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride,
+                          const uint8_t *src0, uint32_t src0_stride,
+                          const uint8_t *src1, uint32_t src1_stride,
+                          const uint8_t *mask, uint32_t mask_stride, int h,
+                          int w, int subh, int subw) {
+  int i, j;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  if (subw == 0 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = mask[i * mask_stride + j];
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 1) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = ROUND_POWER_OF_TWO(
+            mask[(2 * i) * mask_stride + (2 * j)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j)] +
+                mask[(2 * i) * mask_stride + (2 * j + 1)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j + 1)],
+            2);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[i * mask_stride + (2 * j)],
+                                    mask[i * mask_stride + (2 * j + 1)]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[(2 * i) * mask_stride + j],
+                                    mask[(2 * i + 1) * mask_stride + j]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_blend_a64_mask_c(uint8_t *dst_8, uint32_t dst_stride,
+                                 const uint8_t *src0_8, uint32_t src0_stride,
+                                 const uint8_t *src1_8, uint32_t src1_stride,
+                                 const uint8_t *mask, uint32_t mask_stride,
+                                 int h, int w, int subh, int subw, int bd) {
+  int i, j;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8);
+  const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8);
+  const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8);
+  (void)bd;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  if (subw == 0 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = mask[i * mask_stride + j];
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 1) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = ROUND_POWER_OF_TWO(
+            mask[(2 * i) * mask_stride + (2 * j)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j)] +
+                mask[(2 * i) * mask_stride + (2 * j + 1)] +
+                mask[(2 * i + 1) * mask_stride + (2 * j + 1)],
+            2);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else if (subw == 1 && subh == 0) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[i * mask_stride + (2 * j)],
+                                    mask[i * mask_stride + (2 * j + 1)]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  } else {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int m = AOM_BLEND_AVG(mask[(2 * i) * mask_stride + j],
+                                    mask[(2 * i + 1) * mask_stride + j]);
+        dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                                src1[i * src1_stride + j]);
+      }
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/blend_a64_vmask.c b/third_party/aom/aom_dsp/blend_a64_vmask.c
new file mode 100644
index 0000000000..1a5e30e315
--- /dev/null
+++ b/third_party/aom/aom_dsp/blend_a64_vmask.c
@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "./aom_dsp_rtcd.h"
+
+void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride,
+                           const uint8_t *src0, uint32_t src0_stride,
+                           const uint8_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int h, int w) {
+  int i, j;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  for (i = 0; i < h; ++i) {
+    const int m = mask[i];
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                              src1[i * src1_stride + j]);
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_blend_a64_vmask_c(uint8_t *dst_8, uint32_t dst_stride,
+                                  const uint8_t *src0_8, uint32_t src0_stride,
+                                  const uint8_t *src1_8, uint32_t src1_stride,
+                                  const uint8_t *mask, int h, int w, int bd) {
+  int i, j;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst_8);
+  const uint16_t *src0 = CONVERT_TO_SHORTPTR(src0_8);
+  const uint16_t *src1 = CONVERT_TO_SHORTPTR(src1_8);
+  (void)bd;
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  for (i = 0; i < h; ++i) {
+    const int m = mask[i];
+    for (j = 0; j < w; ++j) {
+      dst[i * dst_stride + j] = AOM_BLEND_A64(m, src0[i * src0_stride + j],
+                                              src1[i * src1_stride + j]);
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/buf_ans.c b/third_party/aom/aom_dsp/buf_ans.c
new file mode 100644
index 0000000000..8fe1ff7631
--- /dev/null
+++ b/third_party/aom/aom_dsp/buf_ans.c
@@ -0,0 +1,71 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <string.h>
+
+#include "aom_dsp/buf_ans.h"
+#include "aom_mem/aom_mem.h"
+#include "aom/internal/aom_codec_internal.h"
+
+void aom_buf_ans_alloc(struct BufAnsCoder *c,
+                       struct aom_internal_error_info *error, int size) {
+  c->error = error;
+  c->size = size;
+  assert(c->size > 1);
+  AOM_CHECK_MEM_ERROR(error, c->buf, aom_malloc(c->size * sizeof(*c->buf)));
+  // Initialize to overfull to trigger the assert in write.
+  c->offset = c->size + 1;
+}
+
+void aom_buf_ans_free(struct BufAnsCoder *c) {
+  aom_free(c->buf);
+  c->buf = NULL;
+  c->size = 0;
+}
+
+#if !ANS_MAX_SYMBOLS
+void aom_buf_ans_grow(struct BufAnsCoder *c) {
+  struct buffered_ans_symbol *new_buf = NULL;
+  int new_size = c->size * 2;
+  AOM_CHECK_MEM_ERROR(c->error, new_buf,
+                      aom_malloc(new_size * sizeof(*new_buf)));
+  memcpy(new_buf, c->buf, c->size * sizeof(*c->buf));
+  aom_free(c->buf);
+  c->buf = new_buf;
+  c->size = new_size;
+}
+#endif
+
+void aom_buf_ans_flush(struct BufAnsCoder *const c) {
+  int offset;
+#if ANS_MAX_SYMBOLS
+  if (c->offset == 0) return;
+#endif
+  assert(c->offset > 0);
+  offset = c->offset - 1;
+  // Code the first symbol such that it brings the state to the smallest normal
+  // state from an initial state that would have been a subnormal/refill state.
+  if (c->buf[offset].method == ANS_METHOD_RANS) {
+    c->ans.state += c->buf[offset].val_start;
+  } else {
+    c->ans.state += c->buf[offset].val_start ? c->buf[offset].prob : 0;
+  }
+  for (offset = offset - 1; offset >= 0; --offset) {
+    if (c->buf[offset].method == ANS_METHOD_RANS) {
+      rans_write(&c->ans, c->buf[offset].val_start, c->buf[offset].prob);
+    } else {
+      rabs_write(&c->ans, (uint8_t)c->buf[offset].val_start,
+                 (AnsP8)c->buf[offset].prob);
+    }
+  }
+  c->offset = 0;
+  c->output_bytes += ans_write_end(&c->ans);
+}
diff --git a/third_party/aom/aom_dsp/buf_ans.h b/third_party/aom/aom_dsp/buf_ans.h
new file mode 100644
index 0000000000..0768506b35
--- /dev/null
+++ b/third_party/aom/aom_dsp/buf_ans.h
@@ -0,0 +1,133 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_BUF_ANS_H_
+#define AOM_DSP_BUF_ANS_H_
+// Buffered forward ANS writer.
+// Symbols are written to the writer in forward (decode) order and serialized
+// backwards due to ANS's stack like behavior.
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/ans.h"
+#include "aom_dsp/answriter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+#define ANS_METHOD_RABS 0
+#define ANS_METHOD_RANS 1
+
+struct buffered_ans_symbol {
+  unsigned int method : 1;  // one of ANS_METHOD_RABS or ANS_METHOD_RANS
+  // TODO(aconverse): Should be possible to write this in terms of start for ABS
+  unsigned int val_start : RANS_PROB_BITS;  // Boolean value for ABS
+                                            // start in symbol cycle for Rans
+  unsigned int prob : RANS_PROB_BITS;       // Probability of this symbol
+};
+
+struct BufAnsCoder {
+  struct aom_internal_error_info *error;
+  struct buffered_ans_symbol *buf;
+  struct AnsCoder ans;
+  int size;
+  int offset;
+  int output_bytes;
+#if ANS_MAX_SYMBOLS
+  int window_size;
+#endif
+};
+
+// Allocate a buffered ANS coder to store size symbols.
+// When ANS_MAX_SYMBOLS is turned on, the size is the fixed size of each ANS
+// partition.
+// When ANS_MAX_SYMBOLS is turned off, size is merely an initial hint and the
+// buffer will grow on demand
+void aom_buf_ans_alloc(struct BufAnsCoder *c,
+                       struct aom_internal_error_info *error, int hint);
+
+void aom_buf_ans_free(struct BufAnsCoder *c);
+
+#if !ANS_MAX_SYMBOLS
+void aom_buf_ans_grow(struct BufAnsCoder *c);
+#endif
+
+void aom_buf_ans_flush(struct BufAnsCoder *const c);
+
+static INLINE void buf_ans_write_init(struct BufAnsCoder *const c,
+                                      uint8_t *const output_buffer) {
+  c->offset = 0;
+  c->output_bytes = 0;
+  ans_write_init(&c->ans, output_buffer);
+}
+
+static INLINE void buf_rabs_write(struct BufAnsCoder *const c, uint8_t val,
+                                  AnsP8 prob) {
+  assert(c->offset <= c->size);
+#if !ANS_MAX_SYMBOLS
+  if (c->offset == c->size) {
+    aom_buf_ans_grow(c);
+  }
+#endif
+  c->buf[c->offset].method = ANS_METHOD_RABS;
+  c->buf[c->offset].val_start = val;
+  c->buf[c->offset].prob = prob;
+  ++c->offset;
+#if ANS_MAX_SYMBOLS
+  if (c->offset == c->size) aom_buf_ans_flush(c);
+#endif
+}
+
+// Buffer one symbol for encoding using rANS.
+// cum_prob: The cumulative probability before this symbol (the offset of
+// the symbol in the symbol cycle)
+// prob: The probability of this symbol (l_s from the paper)
+// RANS_PRECISION takes the place of m from the paper.
+static INLINE void buf_rans_write(struct BufAnsCoder *const c,
+                                  aom_cdf_prob cum_prob, aom_cdf_prob prob) {
+  assert(c->offset <= c->size);
+#if !ANS_MAX_SYMBOLS
+  if (c->offset == c->size) {
+    aom_buf_ans_grow(c);
+  }
+#endif
+  c->buf[c->offset].method = ANS_METHOD_RANS;
+  c->buf[c->offset].val_start = cum_prob;
+  c->buf[c->offset].prob = prob;
+  ++c->offset;
+#if ANS_MAX_SYMBOLS
+  if (c->offset == c->size) aom_buf_ans_flush(c);
+#endif
+}
+
+static INLINE void buf_rabs_write_bit(struct BufAnsCoder *c, int bit) {
+  buf_rabs_write(c, bit, 128);
+}
+
+static INLINE void buf_rabs_write_literal(struct BufAnsCoder *c, int literal,
+                                          int bits) {
+  int bit;
+
+  assert(bits < 31);
+  for (bit = bits - 1; bit >= 0; bit--)
+    buf_rabs_write_bit(c, 1 & (literal >> bit));
+}
+
+static INLINE int buf_ans_write_end(struct BufAnsCoder *const c) {
+  assert(c->offset == 0);
+  return c->output_bytes;
+}
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_DSP_BUF_ANS_H_
diff --git a/third_party/aom/aom_dsp/daalaboolreader.c b/third_party/aom/aom_dsp/daalaboolreader.c
new file mode 100644
index 0000000000..0fc7b14a5a
--- /dev/null
+++ b/third_party/aom/aom_dsp/daalaboolreader.c
@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/daalaboolreader.h"
+
+int aom_daala_reader_init(daala_reader *r, const uint8_t *buffer, int size) {
+  if (size && !buffer) {
+    return 1;
+  }
+  r->buffer_end = buffer + size;
+  r->buffer = buffer;
+  od_ec_dec_init(&r->ec, buffer, size - 1);
+#if CONFIG_ACCOUNTING
+  r->accounting = NULL;
+#endif
+  return 0;
+}
+
+const uint8_t *aom_daala_reader_find_end(daala_reader *r) {
+  return r->buffer_end;
+}
+
+uint32_t aom_daala_reader_tell(const daala_reader *r) {
+  return od_ec_dec_tell(&r->ec);
+}
+
+uint32_t aom_daala_reader_tell_frac(const daala_reader *r) {
+  return od_ec_dec_tell_frac(&r->ec);
+}
diff --git a/third_party/aom/aom_dsp/daalaboolreader.h b/third_party/aom/aom_dsp/daalaboolreader.h
new file mode 100644
index 0000000000..428d74db09
--- /dev/null
+++ b/third_party/aom/aom_dsp/daalaboolreader.h
@@ -0,0 +1,164 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_DAALABOOLREADER_H_
+#define AOM_DSP_DAALABOOLREADER_H_
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/entdec.h"
+#include "aom_dsp/prob.h"
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#endif
+#if CONFIG_BITSTREAM_DEBUG
+#include <stdio.h>
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct daala_reader {
+  const uint8_t *buffer;
+  const uint8_t *buffer_end;
+  od_ec_dec ec;
+#if CONFIG_ACCOUNTING
+  Accounting *accounting;
+#endif
+};
+
+typedef struct daala_reader daala_reader;
+
+int aom_daala_reader_init(daala_reader *r, const uint8_t *buffer, int size);
+const uint8_t *aom_daala_reader_find_end(daala_reader *r);
+uint32_t aom_daala_reader_tell(const daala_reader *r);
+uint32_t aom_daala_reader_tell_frac(const daala_reader *r);
+
+static INLINE int aom_daala_read(daala_reader *r, int prob) {
+  int bit;
+#if CONFIG_EC_SMALLMUL
+  int p = (0x7FFFFF - (prob << 15) + prob) >> 8;
+#else
+  int p = ((prob << 15) + 256 - prob) >> 8;
+#endif
+#if CONFIG_BITSTREAM_DEBUG
+/*{
+  const int queue_r = bitstream_queue_get_read();
+  const int frame_idx = bitstream_queue_get_frame_read();
+  if (frame_idx == 0 && queue_r == 0) {
+    fprintf(stderr, "\n *** bitstream queue at frame_idx_r %d queue_r %d\n",
+            frame_idx, queue_r);
+  }
+}*/
+#endif
+
+  bit = od_ec_decode_bool_q15(&r->ec, p);
+
+#if CONFIG_BITSTREAM_DEBUG
+  {
+    int i;
+    int ref_bit, ref_nsymbs;
+    aom_cdf_prob ref_cdf[16];
+    const int queue_r = bitstream_queue_get_read();
+    const int frame_idx = bitstream_queue_get_frame_read();
+    bitstream_queue_pop(&ref_bit, ref_cdf, &ref_nsymbs);
+    if (ref_nsymbs != 2) {
+      fprintf(stderr,
+              "\n *** [bit] nsymbs error, frame_idx_r %d nsymbs %d ref_nsymbs "
+              "%d queue_r %d\n",
+              frame_idx, 2, ref_nsymbs, queue_r);
+      assert(0);
+    }
+    if ((ref_nsymbs != 2) || (ref_cdf[0] != (aom_cdf_prob)p) ||
+        (ref_cdf[1] != 32767)) {
+      fprintf(stderr,
+              "\n *** [bit] cdf error, frame_idx_r %d cdf {%d, %d} ref_cdf {%d",
+              frame_idx, p, 32767, ref_cdf[0]);
+      for (i = 1; i < ref_nsymbs; ++i) fprintf(stderr, ", %d", ref_cdf[i]);
+      fprintf(stderr, "} queue_r %d\n", queue_r);
+      assert(0);
+    }
+    if (bit != ref_bit) {
+      fprintf(stderr,
+              "\n *** [bit] symb error, frame_idx_r %d symb %d ref_symb %d "
+              "queue_r %d\n",
+              frame_idx, bit, ref_bit, queue_r);
+      assert(0);
+    }
+  }
+#endif
+
+  return bit;
+}
+
+#if CONFIG_RAWBITS
+static INLINE int aom_daala_read_bit(daala_reader *r) {
+  return od_ec_dec_bits(&r->ec, 1, "aom_bits");
+}
+#endif
+
+static INLINE int aom_daala_reader_has_error(daala_reader *r) {
+  return r->ec.error;
+}
+
+static INLINE int daala_read_symbol(daala_reader *r, const aom_cdf_prob *cdf,
+                                    int nsymbs) {
+  int symb;
+  symb = od_ec_decode_cdf_q15(&r->ec, cdf, nsymbs);
+
+#if CONFIG_BITSTREAM_DEBUG
+  {
+    int i;
+    int cdf_error = 0;
+    int ref_symb, ref_nsymbs;
+    aom_cdf_prob ref_cdf[16];
+    const int queue_r = bitstream_queue_get_read();
+    const int frame_idx = bitstream_queue_get_frame_read();
+    bitstream_queue_pop(&ref_symb, ref_cdf, &ref_nsymbs);
+    if (nsymbs != ref_nsymbs) {
+      fprintf(stderr,
+              "\n *** nsymbs error, frame_idx_r %d nsymbs %d ref_nsymbs %d "
+              "queue_r %d\n",
+              frame_idx, nsymbs, ref_nsymbs, queue_r);
+      cdf_error = 0;
+      assert(0);
+    } else {
+      for (i = 0; i < nsymbs; ++i)
+        if (cdf[i] != ref_cdf[i]) cdf_error = 1;
+    }
+    if (cdf_error) {
+      fprintf(stderr, "\n *** cdf error, frame_idx_r %d cdf {%d", frame_idx,
+              cdf[0]);
+      for (i = 1; i < nsymbs; ++i) fprintf(stderr, ", %d", cdf[i]);
+      fprintf(stderr, "} ref_cdf {%d", ref_cdf[0]);
+      for (i = 1; i < ref_nsymbs; ++i) fprintf(stderr, ", %d", ref_cdf[i]);
+      fprintf(stderr, "} queue_r %d\n", queue_r);
+      assert(0);
+    }
+    if (symb != ref_symb) {
+      fprintf(
+          stderr,
+          "\n *** symb error, frame_idx_r %d symb %d ref_symb %d queue_r %d\n",
+          frame_idx, symb, ref_symb, queue_r);
+      assert(0);
+    }
+  }
+#endif
+
+  return symb;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/third_party/aom/aom_dsp/daalaboolwriter.c b/third_party/aom/aom_dsp/daalaboolwriter.c
new file mode 100644
index 0000000000..0ba8f6ab8b
--- /dev/null
+++ b/third_party/aom/aom_dsp/daalaboolwriter.c
@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <string.h>
+#include "aom_dsp/daalaboolwriter.h"
+
+void aom_daala_start_encode(daala_writer *br, uint8_t *source) {
+  br->buffer = source;
+  br->pos = 0;
+  od_ec_enc_init(&br->ec, 62025);
+}
+
+void aom_daala_stop_encode(daala_writer *br) {
+  uint32_t daala_bytes;
+  unsigned char *daala_data;
+  daala_data = od_ec_enc_done(&br->ec, &daala_bytes);
+  memcpy(br->buffer, daala_data, daala_bytes);
+  br->pos = daala_bytes;
+  /* Prevent ec bitstream from being detected as a superframe marker.
+     Must always be added, so that rawbits knows the exact length of the
+      bitstream. */
+  br->buffer[br->pos++] = 0;
+  od_ec_enc_clear(&br->ec);
+}
diff --git a/third_party/aom/aom_dsp/daalaboolwriter.h b/third_party/aom/aom_dsp/daalaboolwriter.h
new file mode 100644
index 0000000000..bbaf53c69f
--- /dev/null
+++ b/third_party/aom/aom_dsp/daalaboolwriter.h
@@ -0,0 +1,87 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_DAALABOOLWRITER_H_
+#define AOM_DSP_DAALABOOLWRITER_H_
+
+#include <stdio.h>
+
+#include "aom_dsp/entenc.h"
+#include "aom_dsp/prob.h"
+#if CONFIG_BITSTREAM_DEBUG
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct daala_writer {
+  unsigned int pos;
+  uint8_t *buffer;
+  od_ec_enc ec;
+};
+
+typedef struct daala_writer daala_writer;
+
+void aom_daala_start_encode(daala_writer *w, uint8_t *buffer);
+void aom_daala_stop_encode(daala_writer *w);
+
+static INLINE void aom_daala_write(daala_writer *w, int bit, int prob) {
+#if CONFIG_EC_SMALLMUL
+  int p = (0x7FFFFF - (prob << 15) + prob) >> 8;
+#else
+  int p = ((prob << 15) + 256 - prob) >> 8;
+#endif
+#if CONFIG_BITSTREAM_DEBUG
+  aom_cdf_prob cdf[2] = { (aom_cdf_prob)p, 32767 };
+  /*int queue_r = 0;
+  int frame_idx_r = 0;
+  int queue_w = bitstream_queue_get_write();
+  int frame_idx_w = bitstream_queue_get_frame_write();
+  if (frame_idx_w == frame_idx_r && queue_w == queue_r) {
+    fprintf(stderr, "\n *** bitstream queue at frame_idx_w %d queue_w %d\n",
+    frame_idx_w, queue_w);
+  }*/
+  bitstream_queue_push(bit, cdf, 2);
+#endif
+
+  od_ec_encode_bool_q15(&w->ec, bit, p);
+}
+
+#if CONFIG_RAWBITS
+static INLINE void aom_daala_write_bit(daala_writer *w, int bit) {
+  od_ec_enc_bits(&w->ec, bit, 1);
+}
+#endif
+
+static INLINE void daala_write_symbol(daala_writer *w, int symb,
+                                      const aom_cdf_prob *cdf, int nsymbs) {
+#if CONFIG_BITSTREAM_DEBUG
+  /*int queue_r = 0;
+  int frame_idx_r = 0;
+  int queue_w = bitstream_queue_get_write();
+  int frame_idx_w = bitstream_queue_get_frame_write();
+  if (frame_idx_w == frame_idx_r && queue_w == queue_r) {
+    fprintf(stderr, "\n *** bitstream queue at frame_idx_w %d queue_w %d\n",
+    frame_idx_w, queue_w);
+  }*/
+  bitstream_queue_push(symb, cdf, nsymbs);
+#endif
+
+  od_ec_encode_cdf_q15(&w->ec, symb, cdf, nsymbs);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/third_party/aom/aom_dsp/dkboolreader.c b/third_party/aom/aom_dsp/dkboolreader.c
new file mode 100644
index 0000000000..288d5f1ce4
--- /dev/null
+++ b/third_party/aom/aom_dsp/dkboolreader.c
@@ -0,0 +1,110 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "./aom_config.h"
+
+#include "aom_dsp/dkboolreader.h"
+#include "aom_dsp/prob.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_util/endian_inl.h"
+
+static INLINE int aom_dk_read_bit(struct aom_dk_reader *r) {
+  return aom_dk_read(r, 128);  // aom_prob_half
+}
+
+int aom_dk_reader_init(struct aom_dk_reader *r, const uint8_t *buffer,
+                       size_t size, aom_decrypt_cb decrypt_cb,
+                       void *decrypt_state) {
+  if (size && !buffer) {
+    return 1;
+  } else {
+    r->buffer_end = buffer + size;
+    r->buffer_start = r->buffer = buffer;
+    r->value = 0;
+    r->count = -8;
+    r->range = 255;
+    r->decrypt_cb = decrypt_cb;
+    r->decrypt_state = decrypt_state;
+    aom_dk_reader_fill(r);
+#if CONFIG_ACCOUNTING
+    r->accounting = NULL;
+#endif
+    return aom_dk_read_bit(r) != 0;  // marker bit
+  }
+}
+
+void aom_dk_reader_fill(struct aom_dk_reader *r) {
+  const uint8_t *const buffer_end = r->buffer_end;
+  const uint8_t *buffer = r->buffer;
+  const uint8_t *buffer_start = buffer;
+  BD_VALUE value = r->value;
+  int count = r->count;
+  const size_t bytes_left = buffer_end - buffer;
+  const size_t bits_left = bytes_left * CHAR_BIT;
+  int shift = BD_VALUE_SIZE - CHAR_BIT - (count + CHAR_BIT);
+
+  if (r->decrypt_cb) {
+    size_t n = AOMMIN(sizeof(r->clear_buffer), bytes_left);
+    r->decrypt_cb(r->decrypt_state, buffer, r->clear_buffer, (int)n);
+    buffer = r->clear_buffer;
+    buffer_start = r->clear_buffer;
+  }
+  if (bits_left > BD_VALUE_SIZE) {
+    const int bits = (shift & 0xfffffff8) + CHAR_BIT;
+    BD_VALUE nv;
+    BD_VALUE big_endian_values;
+    memcpy(&big_endian_values, buffer, sizeof(BD_VALUE));
+#if SIZE_MAX == 0xffffffffffffffffULL
+    big_endian_values = HToBE64(big_endian_values);
+#else
+    big_endian_values = HToBE32(big_endian_values);
+#endif
+    nv = big_endian_values >> (BD_VALUE_SIZE - bits);
+    count += bits;
+    buffer += (bits >> 3);
+    value = r->value | (nv << (shift & 0x7));
+  } else {
+    const int bits_over = (int)(shift + CHAR_BIT - (int)bits_left);
+    int loop_end = 0;
+    if (bits_over >= 0) {
+      count += LOTS_OF_BITS;
+      loop_end = bits_over;
+    }
+
+    if (bits_over < 0 || bits_left) {
+      while (shift >= loop_end) {
+        count += CHAR_BIT;
+        value |= (BD_VALUE)*buffer++ << shift;
+        shift -= CHAR_BIT;
+      }
+    }
+  }
+
+  // NOTE: Variable 'buffer' may not relate to 'r->buffer' after decryption,
+  // so we increase 'r->buffer' by the amount that 'buffer' moved, rather than
+  // assign 'buffer' to 'r->buffer'.
+  r->buffer += buffer - buffer_start;
+  r->value = value;
+  r->count = count;
+}
+
+const uint8_t *aom_dk_reader_find_end(struct aom_dk_reader *r) {
+  // Find the end of the coded buffer
+  while (r->count > CHAR_BIT && r->count < BD_VALUE_SIZE) {
+    r->count -= CHAR_BIT;
+    r->buffer--;
+  }
+  return r->buffer;
+}
diff --git a/third_party/aom/aom_dsp/dkboolreader.h b/third_party/aom/aom_dsp/dkboolreader.h
new file mode 100644
index 0000000000..f0bc843813
--- /dev/null
+++ b/third_party/aom/aom_dsp/dkboolreader.h
@@ -0,0 +1,181 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_DKBOOLREADER_H_
+#define AOM_DSP_DKBOOLREADER_H_
+
+#include <assert.h>
+#include <stddef.h>
+#include <limits.h>
+
+#include "./aom_config.h"
+#if CONFIG_BITSTREAM_DEBUG
+#include <assert.h>
+#include <stdio.h>
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#include "aom_ports/mem.h"
+#include "aom/aomdx.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/prob.h"
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef size_t BD_VALUE;
+
+#define BD_VALUE_SIZE ((int)sizeof(BD_VALUE) * CHAR_BIT)
+
+// This is meant to be a large, positive constant that can still be efficiently
+// loaded as an immediate (on platforms like ARM, for example).
+// Even relatively modest values like 100 would work fine.
+#define LOTS_OF_BITS 0x40000000
+
+struct aom_dk_reader {
+  // Be careful when reordering this struct, it may impact the cache negatively.
+  BD_VALUE value;
+  unsigned int range;
+  int count;
+  const uint8_t *buffer_start;
+  const uint8_t *buffer_end;
+  const uint8_t *buffer;
+  aom_decrypt_cb decrypt_cb;
+  void *decrypt_state;
+  uint8_t clear_buffer[sizeof(BD_VALUE) + 1];
+#if CONFIG_ACCOUNTING
+  Accounting *accounting;
+#endif
+};
+
+int aom_dk_reader_init(struct aom_dk_reader *r, const uint8_t *buffer,
+                       size_t size, aom_decrypt_cb decrypt_cb,
+                       void *decrypt_state);
+
+void aom_dk_reader_fill(struct aom_dk_reader *r);
+
+const uint8_t *aom_dk_reader_find_end(struct aom_dk_reader *r);
+
+static INLINE uint32_t aom_dk_reader_tell(const struct aom_dk_reader *r) {
+  const uint32_t bits_read =
+      (uint32_t)((r->buffer - r->buffer_start) * CHAR_BIT);
+  const int count =
+      (r->count < LOTS_OF_BITS) ? r->count : r->count - LOTS_OF_BITS;
+  assert(r->buffer >= r->buffer_start);
+  return bits_read - (count + CHAR_BIT);
+}
+
+/*The resolution of fractional-precision bit usage measurements, i.e.,
+   3 => 1/8th bits.*/
+#define DK_BITRES (3)
+
+static INLINE uint32_t aom_dk_reader_tell_frac(const struct aom_dk_reader *r) {
+  uint32_t num_bits;
+  uint32_t range;
+  int l;
+  int i;
+  num_bits = aom_dk_reader_tell(r) << DK_BITRES;
+  range = r->range;
+  l = 0;
+  for (i = DK_BITRES; i-- > 0;) {
+    int b;
+    range = range * range >> 7;
+    b = (int)(range >> 8);
+    l = l << 1 | b;
+    range >>= b;
+  }
+  return num_bits - l;
+}
+
+static INLINE int aom_dk_reader_has_error(struct aom_dk_reader *r) {
+  // Check if we have reached the end of the buffer.
+  //
+  // Variable 'count' stores the number of bits in the 'value' buffer, minus
+  // 8. The top byte is part of the algorithm, and the remainder is buffered
+  // to be shifted into it. So if count == 8, the top 16 bits of 'value' are
+  // occupied, 8 for the algorithm and 8 in the buffer.
+  //
+  // When reading a byte from the user's buffer, count is filled with 8 and
+  // one byte is filled into the value buffer. When we reach the end of the
+  // data, count is additionally filled with LOTS_OF_BITS. So when
+  // count == LOTS_OF_BITS - 1, the user's data has been exhausted.
+  //
+  // 1 if we have tried to decode bits after the end of stream was encountered.
+  // 0 No error.
+  return r->count > BD_VALUE_SIZE && r->count < LOTS_OF_BITS;
+}
+
+static INLINE int aom_dk_read(struct aom_dk_reader *r, int prob) {
+  unsigned int bit = 0;
+  BD_VALUE value;
+  BD_VALUE bigsplit;
+  int count;
+  unsigned int range;
+  unsigned int split = (r->range * prob + (256 - prob)) >> CHAR_BIT;
+
+  if (r->count < 0) aom_dk_reader_fill(r);
+
+  value = r->value;
+  count = r->count;
+
+  bigsplit = (BD_VALUE)split << (BD_VALUE_SIZE - CHAR_BIT);
+
+  range = split;
+
+  if (value >= bigsplit) {
+    range = r->range - split;
+    value = value - bigsplit;
+    bit = 1;
+  }
+
+  {
+    register int shift = aom_norm[range];
+    range <<= shift;
+    value <<= shift;
+    count -= shift;
+  }
+  r->value = value;
+  r->count = count;
+  r->range = range;
+
+#if CONFIG_BITSTREAM_DEBUG
+  {
+    int ref_bit, ref_prob;
+    const int queue_r = bitstream_queue_get_read();
+    const int frame_idx = bitstream_queue_get_frame_read();
+    bitstream_queue_pop(&ref_bit, &ref_prob);
+    if (prob != ref_prob) {
+      fprintf(
+          stderr,
+          "\n *** prob error, frame_idx_r %d prob %d ref_prob %d queue_r %d\n",
+          frame_idx, prob, ref_prob, queue_r);
+      assert(0);
+    }
+    if ((int)bit != ref_bit) {
+      fprintf(stderr, "\n *** bit error, frame_idx_r %d bit %d ref_bit %d\n",
+              frame_idx, bit, ref_bit);
+      assert(0);
+    }
+  }
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+  return bit;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_DKBOOLREADER_H_
diff --git a/third_party/aom/aom_dsp/dkboolwriter.c b/third_party/aom/aom_dsp/dkboolwriter.c
new file mode 100644
index 0000000000..fc98e7c9be
--- /dev/null
+++ b/third_party/aom/aom_dsp/dkboolwriter.c
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "./dkboolwriter.h"
+
+static INLINE void aom_dk_write_bit(aom_dk_writer *w, int bit) {
+  aom_dk_write(w, bit, 128);  // aom_prob_half
+}
+
+void aom_dk_start_encode(aom_dk_writer *br, uint8_t *source) {
+  br->lowvalue = 0;
+  br->range = 255;
+  br->count = -24;
+  br->buffer = source;
+  br->pos = 0;
+  aom_dk_write_bit(br, 0);
+}
+
+void aom_dk_stop_encode(aom_dk_writer *br) {
+  int i;
+
+#if CONFIG_BITSTREAM_DEBUG
+  bitstream_queue_set_skip_write(1);
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+  for (i = 0; i < 32; i++) aom_dk_write_bit(br, 0);
+
+#if CONFIG_BITSTREAM_DEBUG
+  bitstream_queue_set_skip_write(0);
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+  // Ensure there's no ambigous collision with any index marker bytes
+  if ((br->buffer[br->pos - 1] & 0xe0) == 0xc0) br->buffer[br->pos++] = 0;
+}
diff --git a/third_party/aom/aom_dsp/dkboolwriter.h b/third_party/aom/aom_dsp/dkboolwriter.h
new file mode 100644
index 0000000000..835436885b
--- /dev/null
+++ b/third_party/aom/aom_dsp/dkboolwriter.h
@@ -0,0 +1,104 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_DKBOOLWRITER_H_
+#define AOM_DSP_DKBOOLWRITER_H_
+
+#include "./aom_config.h"
+
+#if CONFIG_BITSTREAM_DEBUG
+#include <stdio.h>
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#include "aom_dsp/prob.h"
+#include "aom_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct aom_dk_writer {
+  unsigned int lowvalue;
+  unsigned int range;
+  int count;
+  unsigned int pos;
+  uint8_t *buffer;
+} aom_dk_writer;
+
+void aom_dk_start_encode(aom_dk_writer *bc, uint8_t *buffer);
+void aom_dk_stop_encode(aom_dk_writer *bc);
+
+static INLINE void aom_dk_write(aom_dk_writer *br, int bit, int probability) {
+  unsigned int split;
+  int count = br->count;
+  unsigned int range = br->range;
+  unsigned int lowvalue = br->lowvalue;
+  register int shift;
+
+#if CONFIG_BITSTREAM_DEBUG
+  // int queue_r = 0;
+  // int frame_idx_r = 0;
+  // int queue_w = bitstream_queue_get_write();
+  // int frame_idx_w = bitstream_queue_get_frame_write();
+  // if (frame_idx_w == frame_idx_r && queue_w == queue_r) {
+  //   fprintf(stderr, "\n *** bitstream queue at frame_idx_w %d queue_w %d\n",
+  //   frame_idx_w, queue_w);
+  // }
+  bitstream_queue_push(bit, probability);
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+  split = 1 + (((range - 1) * probability) >> 8);
+
+  range = split;
+
+  if (bit) {
+    lowvalue += split;
+    range = br->range - split;
+  }
+
+  shift = aom_norm[range];
+
+  range <<= shift;
+  count += shift;
+
+  if (count >= 0) {
+    int offset = shift - count;
+
+    if ((lowvalue << (offset - 1)) & 0x80000000) {
+      int x = br->pos - 1;
+
+      while (x >= 0 && br->buffer[x] == 0xff) {
+        br->buffer[x] = 0;
+        x--;
+      }
+
+      br->buffer[x] += 1;
+    }
+
+    br->buffer[br->pos++] = (lowvalue >> (24 - offset));
+    lowvalue <<= offset;
+    shift = count;
+    lowvalue &= 0xffffff;
+    count -= 8;
+  }
+
+  lowvalue <<= shift;
+  br->count = count;
+  br->lowvalue = lowvalue;
+  br->range = range;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_DKBOOLWRITER_H_
diff --git a/third_party/aom/aom_dsp/entcode.c b/third_party/aom/aom_dsp/entcode.c
new file mode 100644
index 0000000000..ad76b7e3ea
--- /dev/null
+++ b/third_party/aom/aom_dsp/entcode.c
@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifdef HAVE_CONFIG_H
+#include "./config.h"
+#endif
+
+#include "aom_dsp/entcode.h"
+
+/*Given the current total integer number of bits used and the current value of
+   rng, computes the fraction number of bits used to OD_BITRES precision.
+  This is used by od_ec_enc_tell_frac() and od_ec_dec_tell_frac().
+  nbits_total: The number of whole bits currently used, i.e., the value
+                returned by od_ec_enc_tell() or od_ec_dec_tell().
+  rng: The current value of rng from either the encoder or decoder state.
+  Return: The number of bits scaled by 2**OD_BITRES.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+uint32_t od_ec_tell_frac(uint32_t nbits_total, uint32_t rng) {
+  uint32_t nbits;
+  int l;
+  int i;
+  /*To handle the non-integral number of bits still left in the encoder/decoder
+     state, we compute the worst-case number of bits of val that must be
+     encoded to ensure that the value is inside the range for any possible
+     subsequent bits.
+    The computation here is independent of val itself (the decoder does not
+     even track that value), even though the real number of bits used after
+     od_ec_enc_done() may be 1 smaller if rng is a power of two and the
+     corresponding trailing bits of val are all zeros.
+    If we did try to track that special case, then coding a value with a
+     probability of 1/(1 << n) might sometimes appear to use more than n bits.
+    This may help explain the surprising result that a newly initialized
+     encoder or decoder claims to have used 1 bit.*/
+  nbits = nbits_total << OD_BITRES;
+  l = 0;
+  for (i = OD_BITRES; i-- > 0;) {
+    int b;
+    rng = rng * rng >> 15;
+    b = (int)(rng >> 16);
+    l = l << 1 | b;
+    rng >>= b;
+  }
+  return nbits - l;
+}
diff --git a/third_party/aom/aom_dsp/entcode.h b/third_party/aom/aom_dsp/entcode.h
new file mode 100644
index 0000000000..534959e661
--- /dev/null
+++ b/third_party/aom/aom_dsp/entcode.h
@@ -0,0 +1,46 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if !defined(_entcode_H)
+#define _entcode_H (1)
+#include <limits.h>
+#include <stddef.h>
+#include "av1/common/odintrin.h"
+
+/*OPT: od_ec_window must be at least 32 bits, but if you have fast arithmetic
+   on a larger type, you can speed up the decoder by using it here.*/
+typedef uint32_t od_ec_window;
+
+#define OD_EC_WINDOW_SIZE ((int)sizeof(od_ec_window) * CHAR_BIT)
+
+/*The number of bits to use for the range-coded part of unsigned integers.*/
+#define OD_EC_UINT_BITS (4)
+
+/*The resolution of fractional-precision bit usage measurements, i.e.,
+   3 => 1/8th bits.*/
+#define OD_BITRES (3)
+
+/*With CONFIG_EC_SMALLMUL, the value stored in a CDF is 32768 minus the actual
+   Q15 cumulative probability (an "inverse" CDF).
+  This function converts from one representation to the other (and is its own
+   inverse).*/
+#if CONFIG_EC_SMALLMUL
+#define OD_ICDF(x) (32768U - (x))
+#else
+#define OD_ICDF(x) (x)
+#endif
+
+/*See entcode.c for further documentation.*/
+
+OD_WARN_UNUSED_RESULT uint32_t od_ec_tell_frac(uint32_t nbits_total,
+                                               uint32_t rng);
+
+#endif
diff --git a/third_party/aom/aom_dsp/entdec.c b/third_party/aom/aom_dsp/entdec.c
new file mode 100644
index 0000000000..49b176cd80
--- /dev/null
+++ b/third_party/aom/aom_dsp/entdec.c
@@ -0,0 +1,300 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifdef HAVE_CONFIG_H
+#include "./config.h"
+#endif
+
+#include "aom_dsp/entdec.h"
+
+/*A range decoder.
+  This is an entropy decoder based upon \cite{Mar79}, which is itself a
+   rediscovery of the FIFO arithmetic code introduced by \cite{Pas76}.
+  It is very similar to arithmetic encoding, except that encoding is done with
+   digits in any base, instead of with bits, and so it is faster when using
+   larger bases (i.e.: a byte).
+  The author claims an average waste of $\frac{1}{2}\log_b(2b)$ bits, where $b$
+   is the base, longer than the theoretical optimum, but to my knowledge there
+   is no published justification for this claim.
+  This only seems true when using near-infinite precision arithmetic so that
+   the process is carried out with no rounding errors.
+
+  An excellent description of implementation details is available at
+   http://www.arturocampos.com/ac_range.html
+  A recent work \cite{MNW98} which proposes several changes to arithmetic
+   encoding for efficiency actually re-discovers many of the principles
+   behind range encoding, and presents a good theoretical analysis of them.
+
+  End of stream is handled by writing out the smallest number of bits that
+   ensures that the stream will be correctly decoded regardless of the value of
+   any subsequent bits.
+  od_ec_dec_tell() can be used to determine how many bits were needed to decode
+   all the symbols thus far; other data can be packed in the remaining bits of
+   the input buffer.
+  @PHDTHESIS{Pas76,
+    author="Richard Clark Pasco",
+    title="Source coding algorithms for fast data compression",
+    school="Dept. of Electrical Engineering, Stanford University",
+    address="Stanford, CA",
+    month=May,
+    year=1976,
+    URL="http://www.richpasco.org/scaffdc.pdf"
+  }
+  @INPROCEEDINGS{Mar79,
+   author="Martin, G.N.N.",
+   title="Range encoding: an algorithm for removing redundancy from a digitised
+    message",
+   booktitle="Video & Data Recording Conference",
+   year=1979,
+   address="Southampton",
+   month=Jul,
+   URL="http://www.compressconsult.com/rangecoder/rngcod.pdf.gz"
+  }
+  @ARTICLE{MNW98,
+   author="Alistair Moffat and Radford Neal and Ian H. Witten",
+   title="Arithmetic Coding Revisited",
+   journal="{ACM} Transactions on Information Systems",
+   year=1998,
+   volume=16,
+   number=3,
+   pages="256--294",
+   month=Jul,
+   URL="http://researchcommons.waikato.ac.nz/bitstream/handle/10289/78/content.pdf"
+  }*/
+
+/*This is meant to be a large, positive constant that can still be efficiently
+   loaded as an immediate (on platforms like ARM, for example).
+  Even relatively modest values like 100 would work fine.*/
+#define OD_EC_LOTS_OF_BITS (0x4000)
+
+static void od_ec_dec_refill(od_ec_dec *dec) {
+  int s;
+  od_ec_window dif;
+  int16_t cnt;
+  const unsigned char *bptr;
+  const unsigned char *end;
+  dif = dec->dif;
+  cnt = dec->cnt;
+  bptr = dec->bptr;
+  end = dec->end;
+  s = OD_EC_WINDOW_SIZE - 9 - (cnt + 15);
+  for (; s >= 0 && bptr < end; s -= 8, bptr++) {
+    OD_ASSERT(s <= OD_EC_WINDOW_SIZE - 8);
+    dif ^= (od_ec_window)bptr[0] << s;
+    cnt += 8;
+  }
+  if (bptr >= end) {
+    dec->tell_offs += OD_EC_LOTS_OF_BITS - cnt;
+    cnt = OD_EC_LOTS_OF_BITS;
+  }
+  dec->dif = dif;
+  dec->cnt = cnt;
+  dec->bptr = bptr;
+}
+
+/*Takes updated dif and range values, renormalizes them so that
+   32768 <= rng < 65536 (reading more bytes from the stream into dif if
+   necessary), and stores them back in the decoder context.
+  dif: The new value of dif.
+  rng: The new value of the range.
+  ret: The value to return.
+  Return: ret.
+          This allows the compiler to jump to this function via a tail-call.*/
+static int od_ec_dec_normalize(od_ec_dec *dec, od_ec_window dif, unsigned rng,
+                               int ret) {
+  int d;
+  OD_ASSERT(rng <= 65535U);
+  d = 16 - OD_ILOG_NZ(rng);
+  dec->cnt -= d;
+#if CONFIG_EC_SMALLMUL
+  /*This is equivalent to shifting in 1's instead of 0's.*/
+  dec->dif = ((dif + 1) << d) - 1;
+#else
+  dec->dif = dif << d;
+#endif
+  dec->rng = rng << d;
+  if (dec->cnt < 0) od_ec_dec_refill(dec);
+  return ret;
+}
+
+/*Initializes the decoder.
+  buf: The input buffer to use.
+  Return: 0 on success, or a negative value on error.*/
+void od_ec_dec_init(od_ec_dec *dec, const unsigned char *buf,
+                    uint32_t storage) {
+  dec->buf = buf;
+  dec->eptr = buf + storage;
+  dec->end_window = 0;
+  dec->nend_bits = 0;
+  dec->tell_offs = 10 - (OD_EC_WINDOW_SIZE - 8);
+  dec->end = buf + storage;
+  dec->bptr = buf;
+#if CONFIG_EC_SMALLMUL
+  dec->dif = ((od_ec_window)1 << (OD_EC_WINDOW_SIZE - 1)) - 1;
+#else
+  dec->dif = 0;
+#endif
+  dec->rng = 0x8000;
+  dec->cnt = -15;
+  dec->error = 0;
+  od_ec_dec_refill(dec);
+}
+
+/*Decode a single binary value.
+  {EC_SMALLMUL} f: The probability that the bit is one, scaled by 32768.
+  {else} f: The probability that the bit is zero, scaled by 32768.
+  Return: The value decoded (0 or 1).*/
+int od_ec_decode_bool_q15(od_ec_dec *dec, unsigned f) {
+  od_ec_window dif;
+  od_ec_window vw;
+  unsigned r;
+  unsigned r_new;
+  unsigned v;
+  int ret;
+  OD_ASSERT(0 < f);
+  OD_ASSERT(f < 32768U);
+  dif = dec->dif;
+  r = dec->rng;
+  OD_ASSERT(dif >> (OD_EC_WINDOW_SIZE - 16) < r);
+  OD_ASSERT(32768U <= r);
+#if CONFIG_EC_SMALLMUL
+  v = (r >> 8) * (uint32_t)f >> 7;
+  vw = (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16);
+  ret = 1;
+  r_new = v;
+  if (dif >= vw) {
+    r_new = r - v;
+    dif -= vw;
+    ret = 0;
+  }
+#else
+  v = f * (uint32_t)r >> 15;
+  vw = (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16);
+  ret = 0;
+  r_new = v;
+  if (dif >= vw) {
+    r_new = r - v;
+    dif -= vw;
+    ret = 1;
+  }
+#endif
+  return od_ec_dec_normalize(dec, dif, r_new, ret);
+}
+
+/*Decodes a symbol given a cumulative distribution function (CDF) table in Q15.
+  cdf: The CDF, such that symbol s falls in the range
+        [s > 0 ? cdf[s - 1] : 0, cdf[s]).
+       The values must be monotonically non-increasing, and cdf[nsyms - 1]
+        must be 32768.
+       {EC_SMALLMUL}: The CDF contains 32768 minus those values.
+  nsyms: The number of symbols in the alphabet.
+         This should be at most 16.
+  Return: The decoded symbol s.*/
+int od_ec_decode_cdf_q15(od_ec_dec *dec, const uint16_t *cdf, int nsyms) {
+  od_ec_window dif;
+  unsigned r;
+  unsigned c;
+  unsigned u;
+  unsigned v;
+  int ret;
+  (void)nsyms;
+  dif = dec->dif;
+  r = dec->rng;
+  OD_ASSERT(dif >> (OD_EC_WINDOW_SIZE - 16) < r);
+  OD_ASSERT(cdf[nsyms - 1] == OD_ICDF(32768U));
+  OD_ASSERT(32768U <= r);
+#if CONFIG_EC_SMALLMUL
+  c = (unsigned)(dif >> (OD_EC_WINDOW_SIZE - 16));
+  v = r;
+  ret = -1;
+  do {
+    u = v;
+    v = (r >> 8) * (uint32_t)cdf[++ret] >> 7;
+  } while (c < v);
+  OD_ASSERT(v < u);
+  OD_ASSERT(u <= r);
+  r = u - v;
+  dif -= (od_ec_window)v << (OD_EC_WINDOW_SIZE - 16);
+#else
+  c = (unsigned)(dif >> (OD_EC_WINDOW_SIZE - 16));
+  v = 0;
+  ret = -1;
+  do {
+    u = v;
+    v = cdf[++ret] * (uint32_t)r >> 15;
+  } while (v <= c);
+  OD_ASSERT(u < v);
+  OD_ASSERT(v <= r);
+  r = v - u;
+  dif -= (od_ec_window)u << (OD_EC_WINDOW_SIZE - 16);
+#endif
+  return od_ec_dec_normalize(dec, dif, r, ret);
+}
+
+#if CONFIG_RAWBITS
+/*Extracts a sequence of raw bits from the stream.
+  The bits must have been encoded with od_ec_enc_bits().
+  ftb: The number of bits to extract.
+       This must be between 0 and 25, inclusive.
+  Return: The decoded bits.*/
+uint32_t od_ec_dec_bits_(od_ec_dec *dec, unsigned ftb) {
+  od_ec_window window;
+  int available;
+  uint32_t ret;
+  OD_ASSERT(ftb <= 25);
+  window = dec->end_window;
+  available = dec->nend_bits;
+  if ((unsigned)available < ftb) {
+    const unsigned char *buf;
+    const unsigned char *eptr;
+    buf = dec->buf;
+    eptr = dec->eptr;
+    OD_ASSERT(available <= OD_EC_WINDOW_SIZE - 8);
+    do {
+      if (eptr <= buf) {
+        dec->tell_offs += OD_EC_LOTS_OF_BITS - available;
+        available = OD_EC_LOTS_OF_BITS;
+        break;
+      }
+      window |= (od_ec_window) * --eptr << available;
+      available += 8;
+    } while (available <= OD_EC_WINDOW_SIZE - 8);
+    dec->eptr = eptr;
+  }
+  ret = (uint32_t)window & (((uint32_t)1 << ftb) - 1);
+  window >>= ftb;
+  available -= ftb;
+  dec->end_window = window;
+  dec->nend_bits = available;
+  return ret;
+}
+#endif
+
+/*Returns the number of bits "used" by the decoded symbols so far.
+  This same number can be computed in either the encoder or the decoder, and is
+   suitable for making coding decisions.
+  Return: The number of bits.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+int od_ec_dec_tell(const od_ec_dec *dec) {
+  return (int)(((dec->end - dec->eptr) + (dec->bptr - dec->buf)) * 8 -
+               dec->cnt - dec->nend_bits + dec->tell_offs);
+}
+
+/*Returns the number of bits "used" by the decoded symbols so far.
+  This same number can be computed in either the encoder or the decoder, and is
+   suitable for making coding decisions.
+  Return: The number of bits scaled by 2**OD_BITRES.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+uint32_t od_ec_dec_tell_frac(const od_ec_dec *dec) {
+  return od_ec_tell_frac(od_ec_dec_tell(dec), dec->rng);
+}
diff --git a/third_party/aom/aom_dsp/entdec.h b/third_party/aom/aom_dsp/entdec.h
new file mode 100644
index 0000000000..e1145e81d4
--- /dev/null
+++ b/third_party/aom/aom_dsp/entdec.h
@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if !defined(_entdec_H)
+#define _entdec_H (1)
+#include <limits.h>
+#include "aom_dsp/entcode.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct od_ec_dec od_ec_dec;
+
+#if defined(OD_ACCOUNTING) && OD_ACCOUNTING
+#define OD_ACC_STR , char *acc_str
+#define od_ec_dec_bits(dec, ftb, str) od_ec_dec_bits_(dec, ftb, str)
+#else
+#define OD_ACC_STR
+#define od_ec_dec_bits(dec, ftb, str) od_ec_dec_bits_(dec, ftb)
+#endif
+
+/*The entropy decoder context.*/
+struct od_ec_dec {
+  /*The start of the current input buffer.*/
+  const unsigned char *buf;
+  /*The read pointer for the raw bits.*/
+  const unsigned char *eptr;
+  /*Bits that will be read from/written at the end.*/
+  od_ec_window end_window;
+  /*Number of valid bits in end_window.*/
+  int nend_bits;
+  /*An offset used to keep track of tell after reaching the end of the stream.
+    This is constant throughout most of the decoding process, but becomes
+     important once we hit the end of the buffer and stop incrementing pointers
+     (and instead pretend cnt/nend_bits have lots of bits).*/
+  int32_t tell_offs;
+  /*The end of the current input buffer.*/
+  const unsigned char *end;
+  /*The read pointer for the entropy-coded bits.*/
+  const unsigned char *bptr;
+  /*The difference between the coded value and the low end of the current
+     range.
+    {EC_SMALLMUL} The difference between the high end of the current range,
+     (low + rng), and the coded value, minus 1.
+    This stores up to OD_EC_WINDOW_SIZE bits of that difference, but the
+     decoder only uses the top 16 bits of the window to decode the next symbol.
+    As we shift up during renormalization, if we don't have enough bits left in
+     the window to fill the top 16, we'll read in more bits of the coded
+     value.*/
+  od_ec_window dif;
+  /*The number of values in the current range.*/
+  uint16_t rng;
+  /*The number of bits of data in the current value.*/
+  int16_t cnt;
+  /*Nonzero if an error occurred.*/
+  int error;
+};
+
+/*See entdec.c for further documentation.*/
+
+void od_ec_dec_init(od_ec_dec *dec, const unsigned char *buf, uint32_t storage)
+    OD_ARG_NONNULL(1) OD_ARG_NONNULL(2);
+
+OD_WARN_UNUSED_RESULT int od_ec_decode_bool_q15(od_ec_dec *dec, unsigned f)
+    OD_ARG_NONNULL(1);
+OD_WARN_UNUSED_RESULT int od_ec_decode_cdf_q15(od_ec_dec *dec,
+                                               const uint16_t *cdf, int nsyms)
+    OD_ARG_NONNULL(1) OD_ARG_NONNULL(2);
+
+OD_WARN_UNUSED_RESULT uint32_t od_ec_dec_bits_(od_ec_dec *dec, unsigned ftb)
+    OD_ARG_NONNULL(1);
+
+OD_WARN_UNUSED_RESULT int od_ec_dec_tell(const od_ec_dec *dec)
+    OD_ARG_NONNULL(1);
+OD_WARN_UNUSED_RESULT uint32_t od_ec_dec_tell_frac(const od_ec_dec *dec)
+    OD_ARG_NONNULL(1);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/third_party/aom/aom_dsp/entenc.c b/third_party/aom/aom_dsp/entenc.c
new file mode 100644
index 0000000000..a350f27f4a
--- /dev/null
+++ b/third_party/aom/aom_dsp/entenc.c
@@ -0,0 +1,507 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifdef HAVE_CONFIG_H
+#include "./config.h"
+#endif
+
+#include <stdlib.h>
+#include <string.h>
+#include "aom_dsp/entenc.h"
+
+/*A range encoder.
+  See entdec.c and the references for implementation details \cite{Mar79,MNW98}.
+
+  @INPROCEEDINGS{Mar79,
+   author="Martin, G.N.N.",
+   title="Range encoding: an algorithm for removing redundancy from a digitised
+    message",
+   booktitle="Video \& Data Recording Conference",
+   year=1979,
+   address="Southampton",
+   month=Jul,
+   URL="http://www.compressconsult.com/rangecoder/rngcod.pdf.gz"
+  }
+  @ARTICLE{MNW98,
+   author="Alistair Moffat and Radford Neal and Ian H. Witten",
+   title="Arithmetic Coding Revisited",
+   journal="{ACM} Transactions on Information Systems",
+   year=1998,
+   volume=16,
+   number=3,
+   pages="256--294",
+   month=Jul,
+   URL="http://researchcommons.waikato.ac.nz/bitstream/handle/10289/78/content.pdf"
+  }*/
+
+/*Takes updated low and range values, renormalizes them so that
+   32768 <= rng < 65536 (flushing bytes from low to the pre-carry buffer if
+   necessary), and stores them back in the encoder context.
+  low: The new value of low.
+  rng: The new value of the range.*/
+static void od_ec_enc_normalize(od_ec_enc *enc, od_ec_window low,
+                                unsigned rng) {
+  int d;
+  int c;
+  int s;
+  c = enc->cnt;
+  OD_ASSERT(rng <= 65535U);
+  d = 16 - OD_ILOG_NZ(rng);
+  s = c + d;
+  /*TODO: Right now we flush every time we have at least one byte available.
+    Instead we should use an od_ec_window and flush right before we're about to
+     shift bits off the end of the window.
+    For a 32-bit window this is about the same amount of work, but for a 64-bit
+     window it should be a fair win.*/
+  if (s >= 0) {
+    uint16_t *buf;
+    uint32_t storage;
+    uint32_t offs;
+    unsigned m;
+    buf = enc->precarry_buf;
+    storage = enc->precarry_storage;
+    offs = enc->offs;
+    if (offs + 2 > storage) {
+      storage = 2 * storage + 2;
+      buf = (uint16_t *)realloc(buf, sizeof(*buf) * storage);
+      if (buf == NULL) {
+        enc->error = -1;
+        enc->offs = 0;
+        return;
+      }
+      enc->precarry_buf = buf;
+      enc->precarry_storage = storage;
+    }
+    c += 16;
+    m = (1 << c) - 1;
+    if (s >= 8) {
+      OD_ASSERT(offs < storage);
+      buf[offs++] = (uint16_t)(low >> c);
+      low &= m;
+      c -= 8;
+      m >>= 8;
+    }
+    OD_ASSERT(offs < storage);
+    buf[offs++] = (uint16_t)(low >> c);
+    s = c + d - 24;
+    low &= m;
+    enc->offs = offs;
+  }
+  enc->low = low << d;
+  enc->rng = rng << d;
+  enc->cnt = s;
+}
+
+/*Initializes the encoder.
+  size: The initial size of the buffer, in bytes.*/
+void od_ec_enc_init(od_ec_enc *enc, uint32_t size) {
+  od_ec_enc_reset(enc);
+  enc->buf = (unsigned char *)malloc(sizeof(*enc->buf) * size);
+  enc->storage = size;
+  if (size > 0 && enc->buf == NULL) {
+    enc->storage = 0;
+    enc->error = -1;
+  }
+  enc->precarry_buf = (uint16_t *)malloc(sizeof(*enc->precarry_buf) * size);
+  enc->precarry_storage = size;
+  if (size > 0 && enc->precarry_buf == NULL) {
+    enc->precarry_storage = 0;
+    enc->error = -1;
+  }
+}
+
+/*Reinitializes the encoder.*/
+void od_ec_enc_reset(od_ec_enc *enc) {
+  enc->end_offs = 0;
+  enc->end_window = 0;
+  enc->nend_bits = 0;
+  enc->offs = 0;
+  enc->low = 0;
+  enc->rng = 0x8000;
+  /*This is initialized to -9 so that it crosses zero after we've accumulated
+     one byte + one carry bit.*/
+  enc->cnt = -9;
+  enc->error = 0;
+#if OD_MEASURE_EC_OVERHEAD
+  enc->entropy = 0;
+  enc->nb_symbols = 0;
+#endif
+}
+
+/*Frees the buffers used by the encoder.*/
+void od_ec_enc_clear(od_ec_enc *enc) {
+  free(enc->precarry_buf);
+  free(enc->buf);
+}
+
+/*Encodes a symbol given its frequency in Q15.
+  fl: The cumulative frequency of all symbols that come before the one to be
+       encoded.
+  fh: The cumulative frequency of all symbols up to and including the one to
+       be encoded.
+  {EC_SMALLMUL} Both values are 32768 minus that.*/
+static void od_ec_encode_q15(od_ec_enc *enc, unsigned fl, unsigned fh) {
+  od_ec_window l;
+  unsigned r;
+  unsigned u;
+  unsigned v;
+  l = enc->low;
+  r = enc->rng;
+  OD_ASSERT(32768U <= r);
+#if CONFIG_EC_SMALLMUL
+  OD_ASSERT(fh < fl);
+  OD_ASSERT(fl <= 32768U);
+  if (fl < 32768U) {
+    u = (r >> 8) * (uint32_t)fl >> 7;
+    v = (r >> 8) * (uint32_t)fh >> 7;
+    l += r - u;
+    r = u - v;
+  } else {
+    r -= (r >> 8) * (uint32_t)fh >> 7;
+  }
+#else
+  OD_ASSERT(fl < fh);
+  OD_ASSERT(fh <= 32768U);
+  u = fl * (uint32_t)r >> 15;
+  v = fh * (uint32_t)r >> 15;
+  r = v - u;
+  l += u;
+#endif
+  od_ec_enc_normalize(enc, l, r);
+#if OD_MEASURE_EC_OVERHEAD
+  enc->entropy -= OD_LOG2((double)(OD_ICDF(fh) - OD_ICDF(fl)) / 32768.);
+  enc->nb_symbols++;
+#endif
+}
+
+/*Encode a single binary value.
+  val: The value to encode (0 or 1).
+  {EC_SMALLMUL} f: The probability that the val is one, scaled by 32768.
+  {else} f: The probability that val is zero, scaled by 32768.*/
+void od_ec_encode_bool_q15(od_ec_enc *enc, int val, unsigned f) {
+  od_ec_window l;
+  unsigned r;
+  unsigned v;
+  OD_ASSERT(0 < f);
+  OD_ASSERT(f < 32768U);
+  l = enc->low;
+  r = enc->rng;
+  OD_ASSERT(32768U <= r);
+#if CONFIG_EC_SMALLMUL
+  v = (r >> 8) * (uint32_t)f >> 7;
+  if (val) l += r - v;
+  r = val ? v : r - v;
+#else
+  v = f * (uint32_t)r >> 15;
+  if (val) l += v;
+  r = val ? r - v : v;
+#endif
+  od_ec_enc_normalize(enc, l, r);
+#if OD_MEASURE_EC_OVERHEAD
+  enc->entropy -=
+      OD_LOG2((double)(val ? 32768 - OD_ICDF(f) : OD_ICDF(f)) / 32768.);
+  enc->nb_symbols++;
+#endif
+}
+
+/*Encodes a symbol given a cumulative distribution function (CDF) table in Q15.
+  s: The index of the symbol to encode.
+  cdf: The CDF, such that symbol s falls in the range
+        [s > 0 ? cdf[s - 1] : 0, cdf[s]).
+       The values must be monotonically non-decreasing, and the last value
+        must be exactly 32768.
+  nsyms: The number of symbols in the alphabet.
+         This should be at most 16.*/
+void od_ec_encode_cdf_q15(od_ec_enc *enc, int s, const uint16_t *cdf,
+                          int nsyms) {
+  (void)nsyms;
+  OD_ASSERT(s >= 0);
+  OD_ASSERT(s < nsyms);
+  OD_ASSERT(cdf[nsyms - 1] == OD_ICDF(32768U));
+  od_ec_encode_q15(enc, s > 0 ? cdf[s - 1] : OD_ICDF(0), cdf[s]);
+}
+
+#if CONFIG_RAWBITS
+/*Encodes a sequence of raw bits in the stream.
+  fl: The bits to encode.
+  ftb: The number of bits to encode.
+       This must be between 0 and 25, inclusive.*/
+void od_ec_enc_bits(od_ec_enc *enc, uint32_t fl, unsigned ftb) {
+  od_ec_window end_window;
+  int nend_bits;
+  OD_ASSERT(ftb <= 25);
+  OD_ASSERT(fl < (uint32_t)1 << ftb);
+#if OD_MEASURE_EC_OVERHEAD
+  enc->entropy += ftb;
+#endif
+  end_window = enc->end_window;
+  nend_bits = enc->nend_bits;
+  if (nend_bits + ftb > OD_EC_WINDOW_SIZE) {
+    unsigned char *buf;
+    uint32_t storage;
+    uint32_t end_offs;
+    buf = enc->buf;
+    storage = enc->storage;
+    end_offs = enc->end_offs;
+    if (end_offs + (OD_EC_WINDOW_SIZE >> 3) >= storage) {
+      unsigned char *new_buf;
+      uint32_t new_storage;
+      new_storage = 2 * storage + (OD_EC_WINDOW_SIZE >> 3);
+      new_buf = (unsigned char *)malloc(sizeof(*new_buf) * new_storage);
+      if (new_buf == NULL) {
+        enc->error = -1;
+        enc->end_offs = 0;
+        return;
+      }
+      OD_COPY(new_buf + new_storage - end_offs, buf + storage - end_offs,
+              end_offs);
+      storage = new_storage;
+      free(buf);
+      enc->buf = buf = new_buf;
+      enc->storage = storage;
+    }
+    do {
+      OD_ASSERT(end_offs < storage);
+      buf[storage - ++end_offs] = (unsigned char)end_window;
+      end_window >>= 8;
+      nend_bits -= 8;
+    } while (nend_bits >= 8);
+    enc->end_offs = end_offs;
+  }
+  OD_ASSERT(nend_bits + ftb <= OD_EC_WINDOW_SIZE);
+  end_window |= (od_ec_window)fl << nend_bits;
+  nend_bits += ftb;
+  enc->end_window = end_window;
+  enc->nend_bits = nend_bits;
+}
+#endif
+
+/*Overwrites a few bits at the very start of an existing stream, after they
+   have already been encoded.
+  This makes it possible to have a few flags up front, where it is easy for
+   decoders to access them without parsing the whole stream, even if their
+   values are not determined until late in the encoding process, without having
+   to buffer all the intermediate symbols in the encoder.
+  In order for this to work, at least nbits bits must have already been encoded
+   using probabilities that are an exact power of two.
+  The encoder can verify the number of encoded bits is sufficient, but cannot
+   check this latter condition.
+  val: The bits to encode (in the least nbits significant bits).
+       They will be decoded in order from most-significant to least.
+  nbits: The number of bits to overwrite.
+         This must be no more than 8.*/
+void od_ec_enc_patch_initial_bits(od_ec_enc *enc, unsigned val, int nbits) {
+  int shift;
+  unsigned mask;
+  OD_ASSERT(nbits >= 0);
+  OD_ASSERT(nbits <= 8);
+  OD_ASSERT(val < 1U << nbits);
+  shift = 8 - nbits;
+  mask = ((1U << nbits) - 1) << shift;
+  if (enc->offs > 0) {
+    /*The first byte has been finalized.*/
+    enc->precarry_buf[0] =
+        (uint16_t)((enc->precarry_buf[0] & ~mask) | val << shift);
+  } else if (9 + enc->cnt + (enc->rng == 0x8000) > nbits) {
+    /*The first byte has yet to be output.*/
+    enc->low = (enc->low & ~((od_ec_window)mask << (16 + enc->cnt))) |
+               (od_ec_window)val << (16 + enc->cnt + shift);
+  } else {
+    /*The encoder hasn't even encoded _nbits of data yet.*/
+    enc->error = -1;
+  }
+}
+
+#if OD_MEASURE_EC_OVERHEAD
+#include <stdio.h>
+#endif
+
+/*Indicates that there are no more symbols to encode.
+  All remaining output bytes are flushed to the output buffer.
+  od_ec_enc_reset() should be called before using the encoder again.
+  bytes: Returns the size of the encoded data in the returned buffer.
+  Return: A pointer to the start of the final buffer, or NULL if there was an
+           encoding error.*/
+unsigned char *od_ec_enc_done(od_ec_enc *enc, uint32_t *nbytes) {
+  unsigned char *out;
+  uint32_t storage;
+  uint16_t *buf;
+  uint32_t offs;
+  uint32_t end_offs;
+  int nend_bits;
+  od_ec_window m;
+  od_ec_window e;
+  od_ec_window l;
+  unsigned r;
+  int c;
+  int s;
+  if (enc->error) return NULL;
+#if OD_MEASURE_EC_OVERHEAD
+  {
+    uint32_t tell;
+    /* Don't count the 1 bit we lose to raw bits as overhead. */
+    tell = od_ec_enc_tell(enc) - 1;
+    fprintf(stderr, "overhead: %f%%\n",
+            100 * (tell - enc->entropy) / enc->entropy);
+    fprintf(stderr, "efficiency: %f bits/symbol\n",
+            (double)tell / enc->nb_symbols);
+  }
+#endif
+  /*We output the minimum number of bits that ensures that the symbols encoded
+     thus far will be decoded correctly regardless of the bits that follow.*/
+  l = enc->low;
+  r = enc->rng;
+  c = enc->cnt;
+  s = 9;
+  m = 0x7FFF;
+  e = (l + m) & ~m;
+  while ((e | m) >= l + r) {
+    s++;
+    m >>= 1;
+    e = (l + m) & ~m;
+  }
+  s += c;
+  offs = enc->offs;
+  buf = enc->precarry_buf;
+  if (s > 0) {
+    unsigned n;
+    storage = enc->precarry_storage;
+    if (offs + ((s + 7) >> 3) > storage) {
+      storage = storage * 2 + ((s + 7) >> 3);
+      buf = (uint16_t *)realloc(buf, sizeof(*buf) * storage);
+      if (buf == NULL) {
+        enc->error = -1;
+        return NULL;
+      }
+      enc->precarry_buf = buf;
+      enc->precarry_storage = storage;
+    }
+    n = (1 << (c + 16)) - 1;
+    do {
+      OD_ASSERT(offs < storage);
+      buf[offs++] = (uint16_t)(e >> (c + 16));
+      e &= n;
+      s -= 8;
+      c -= 8;
+      n >>= 8;
+    } while (s > 0);
+  }
+  /*Make sure there's enough room for the entropy-coded bits and the raw
+     bits.*/
+  out = enc->buf;
+  storage = enc->storage;
+  end_offs = enc->end_offs;
+  e = enc->end_window;
+  nend_bits = enc->nend_bits;
+  s = -s;
+  c = OD_MAXI((nend_bits - s + 7) >> 3, 0);
+  if (offs + end_offs + c > storage) {
+    storage = offs + end_offs + c;
+    out = (unsigned char *)realloc(out, sizeof(*out) * storage);
+    if (out == NULL) {
+      enc->error = -1;
+      return NULL;
+    }
+    OD_MOVE(out + storage - end_offs, out + enc->storage - end_offs, end_offs);
+    enc->buf = out;
+    enc->storage = storage;
+  }
+  /*If we have buffered raw bits, flush them as well.*/
+  while (nend_bits > s) {
+    OD_ASSERT(end_offs < storage);
+    out[storage - ++end_offs] = (unsigned char)e;
+    e >>= 8;
+    nend_bits -= 8;
+  }
+  *nbytes = offs + end_offs;
+  /*Perform carry propagation.*/
+  OD_ASSERT(offs + end_offs <= storage);
+  out = out + storage - (offs + end_offs);
+  c = 0;
+  end_offs = offs;
+  while (offs > 0) {
+    offs--;
+    c = buf[offs] + c;
+    out[offs] = (unsigned char)c;
+    c >>= 8;
+  }
+  /*Add any remaining raw bits to the last byte.
+    There is guaranteed to be enough room, because nend_bits <= s.*/
+  OD_ASSERT(nend_bits <= 0 || end_offs > 0);
+  if (nend_bits > 0) out[end_offs - 1] |= (unsigned char)e;
+  /*Note: Unless there's an allocation error, if you keep encoding into the
+     current buffer and call this function again later, everything will work
+     just fine (you won't get a new packet out, but you will get a single
+     buffer with the new data appended to the old).
+    However, this function is O(N) where N is the amount of data coded so far,
+     so calling it more than once for a given packet is a bad idea.*/
+  return out;
+}
+
+/*Returns the number of bits "used" by the encoded symbols so far.
+  This same number can be computed in either the encoder or the decoder, and is
+   suitable for making coding decisions.
+  Warning: The value returned by this function can decrease compared to an
+   earlier call, even after encoding more data, if there is an encoding error
+   (i.e., a failure to allocate enough space for the output buffer).
+  Return: The number of bits.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+int od_ec_enc_tell(const od_ec_enc *enc) {
+  /*The 10 here counteracts the offset of -9 baked into cnt, and adds 1 extra
+     bit, which we reserve for terminating the stream.*/
+  return (enc->offs + enc->end_offs) * 8 + enc->cnt + enc->nend_bits + 10;
+}
+
+/*Returns the number of bits "used" by the encoded symbols so far.
+  This same number can be computed in either the encoder or the decoder, and is
+   suitable for making coding decisions.
+  Warning: The value returned by this function can decrease compared to an
+   earlier call, even after encoding more data, if there is an encoding error
+   (i.e., a failure to allocate enough space for the output buffer).
+  Return: The number of bits scaled by 2**OD_BITRES.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+uint32_t od_ec_enc_tell_frac(const od_ec_enc *enc) {
+  return od_ec_tell_frac(od_ec_enc_tell(enc), enc->rng);
+}
+
+/*Saves a entropy coder checkpoint to dst.
+  This allows an encoder to reverse a series of entropy coder
+   decisions if it decides that the information would have been
+   better coded some other way.*/
+void od_ec_enc_checkpoint(od_ec_enc *dst, const od_ec_enc *src) {
+  OD_COPY(dst, src, 1);
+}
+
+/*Restores an entropy coder checkpoint saved by od_ec_enc_checkpoint.
+  This can only be used to restore from checkpoints earlier in the target
+   state's history: you can not switch backwards and forwards or otherwise
+   switch to a state which isn't a casual ancestor of the current state.
+  Restore is also incompatible with patching the initial bits, as the
+   changes will remain in the restored version.*/
+void od_ec_enc_rollback(od_ec_enc *dst, const od_ec_enc *src) {
+  unsigned char *buf;
+  uint32_t storage;
+  uint16_t *precarry_buf;
+  uint32_t precarry_storage;
+  OD_ASSERT(dst->storage >= src->storage);
+  OD_ASSERT(dst->precarry_storage >= src->precarry_storage);
+  buf = dst->buf;
+  storage = dst->storage;
+  precarry_buf = dst->precarry_buf;
+  precarry_storage = dst->precarry_storage;
+  OD_COPY(dst, src, 1);
+  dst->buf = buf;
+  dst->storage = storage;
+  dst->precarry_buf = precarry_buf;
+  dst->precarry_storage = precarry_storage;
+}
diff --git a/third_party/aom/aom_dsp/entenc.h b/third_party/aom/aom_dsp/entenc.h
new file mode 100644
index 0000000000..314b363186
--- /dev/null
+++ b/third_party/aom/aom_dsp/entenc.h
@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if !defined(_entenc_H)
+#define _entenc_H (1)
+#include <stddef.h>
+#include "aom_dsp/entcode.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct od_ec_enc od_ec_enc;
+
+#define OD_MEASURE_EC_OVERHEAD (0)
+
+/*The entropy encoder context.*/
+struct od_ec_enc {
+  /*Buffered output.
+    This contains only the raw bits until the final call to od_ec_enc_done(),
+     where all the arithmetic-coded data gets prepended to it.*/
+  unsigned char *buf;
+  /*The size of the buffer.*/
+  uint32_t storage;
+  /*The offset at which the last byte containing raw bits was written.*/
+  uint32_t end_offs;
+  /*Bits that will be read from/written at the end.*/
+  od_ec_window end_window;
+  /*Number of valid bits in end_window.*/
+  int nend_bits;
+  /*A buffer for output bytes with their associated carry flags.*/
+  uint16_t *precarry_buf;
+  /*The size of the pre-carry buffer.*/
+  uint32_t precarry_storage;
+  /*The offset at which the next entropy-coded byte will be written.*/
+  uint32_t offs;
+  /*The low end of the current range.*/
+  od_ec_window low;
+  /*The number of values in the current range.*/
+  uint16_t rng;
+  /*The number of bits of data in the current value.*/
+  int16_t cnt;
+  /*Nonzero if an error occurred.*/
+  int error;
+#if OD_MEASURE_EC_OVERHEAD
+  double entropy;
+  int nb_symbols;
+#endif
+};
+
+/*See entenc.c for further documentation.*/
+
+void od_ec_enc_init(od_ec_enc *enc, uint32_t size) OD_ARG_NONNULL(1);
+void od_ec_enc_reset(od_ec_enc *enc) OD_ARG_NONNULL(1);
+void od_ec_enc_clear(od_ec_enc *enc) OD_ARG_NONNULL(1);
+
+void od_ec_encode_bool_q15(od_ec_enc *enc, int val, unsigned f_q15)
+    OD_ARG_NONNULL(1);
+void od_ec_encode_cdf_q15(od_ec_enc *enc, int s, const uint16_t *cdf, int nsyms)
+    OD_ARG_NONNULL(1) OD_ARG_NONNULL(3);
+
+void od_ec_enc_bits(od_ec_enc *enc, uint32_t fl, unsigned ftb)
+    OD_ARG_NONNULL(1);
+
+void od_ec_enc_patch_initial_bits(od_ec_enc *enc, unsigned val, int nbits)
+    OD_ARG_NONNULL(1);
+OD_WARN_UNUSED_RESULT unsigned char *od_ec_enc_done(od_ec_enc *enc,
+                                                    uint32_t *nbytes)
+    OD_ARG_NONNULL(1) OD_ARG_NONNULL(2);
+
+OD_WARN_UNUSED_RESULT int od_ec_enc_tell(const od_ec_enc *enc)
+    OD_ARG_NONNULL(1);
+OD_WARN_UNUSED_RESULT uint32_t od_ec_enc_tell_frac(const od_ec_enc *enc)
+    OD_ARG_NONNULL(1);
+
+void od_ec_enc_checkpoint(od_ec_enc *dst, const od_ec_enc *src);
+void od_ec_enc_rollback(od_ec_enc *dst, const od_ec_enc *src);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/third_party/aom/aom_dsp/fastssim.c b/third_party/aom/aom_dsp/fastssim.c
new file mode 100644
index 0000000000..09d945afc8
--- /dev/null
+++ b/third_party/aom/aom_dsp/fastssim.c
@@ -0,0 +1,493 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ *
+ *  This code was originally written by: Nathan E. Egge, at the Daala
+ *  project.
+ */
+#include <assert.h>
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/ssim.h"
+#include "aom_ports/system_state.h"
+
+typedef struct fs_level fs_level;
+typedef struct fs_ctx fs_ctx;
+
+#define SSIM_C1 (255 * 255 * 0.01 * 0.01)
+#define SSIM_C2 (255 * 255 * 0.03 * 0.03)
+#if CONFIG_HIGHBITDEPTH
+#define SSIM_C1_10 (1023 * 1023 * 0.01 * 0.01)
+#define SSIM_C1_12 (4095 * 4095 * 0.01 * 0.01)
+#define SSIM_C2_10 (1023 * 1023 * 0.03 * 0.03)
+#define SSIM_C2_12 (4095 * 4095 * 0.03 * 0.03)
+#endif
+#define FS_MINI(_a, _b) ((_a) < (_b) ? (_a) : (_b))
+#define FS_MAXI(_a, _b) ((_a) > (_b) ? (_a) : (_b))
+
+struct fs_level {
+  uint32_t *im1;
+  uint32_t *im2;
+  double *ssim;
+  int w;
+  int h;
+};
+
+struct fs_ctx {
+  fs_level *level;
+  int nlevels;
+  unsigned *col_buf;
+};
+
+static void fs_ctx_init(fs_ctx *_ctx, int _w, int _h, int _nlevels) {
+  unsigned char *data;
+  size_t data_size;
+  int lw;
+  int lh;
+  int l;
+  lw = (_w + 1) >> 1;
+  lh = (_h + 1) >> 1;
+  data_size =
+      _nlevels * sizeof(fs_level) + 2 * (lw + 8) * 8 * sizeof(*_ctx->col_buf);
+  for (l = 0; l < _nlevels; l++) {
+    size_t im_size;
+    size_t level_size;
+    im_size = lw * (size_t)lh;
+    level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
+    level_size += sizeof(*_ctx->level[l].ssim) - 1;
+    level_size /= sizeof(*_ctx->level[l].ssim);
+    level_size += im_size;
+    level_size *= sizeof(*_ctx->level[l].ssim);
+    data_size += level_size;
+    lw = (lw + 1) >> 1;
+    lh = (lh + 1) >> 1;
+  }
+  data = (unsigned char *)malloc(data_size);
+  _ctx->level = (fs_level *)data;
+  _ctx->nlevels = _nlevels;
+  data += _nlevels * sizeof(*_ctx->level);
+  lw = (_w + 1) >> 1;
+  lh = (_h + 1) >> 1;
+  for (l = 0; l < _nlevels; l++) {
+    size_t im_size;
+    size_t level_size;
+    _ctx->level[l].w = lw;
+    _ctx->level[l].h = lh;
+    im_size = lw * (size_t)lh;
+    level_size = 2 * im_size * sizeof(*_ctx->level[l].im1);
+    level_size += sizeof(*_ctx->level[l].ssim) - 1;
+    level_size /= sizeof(*_ctx->level[l].ssim);
+    level_size *= sizeof(*_ctx->level[l].ssim);
+    _ctx->level[l].im1 = (uint32_t *)data;
+    _ctx->level[l].im2 = _ctx->level[l].im1 + im_size;
+    data += level_size;
+    _ctx->level[l].ssim = (double *)data;
+    data += im_size * sizeof(*_ctx->level[l].ssim);
+    lw = (lw + 1) >> 1;
+    lh = (lh + 1) >> 1;
+  }
+  _ctx->col_buf = (unsigned *)data;
+}
+
+static void fs_ctx_clear(fs_ctx *_ctx) { free(_ctx->level); }
+
+static void fs_downsample_level(fs_ctx *_ctx, int _l) {
+  const uint32_t *src1;
+  const uint32_t *src2;
+  uint32_t *dst1;
+  uint32_t *dst2;
+  int w2;
+  int h2;
+  int w;
+  int h;
+  int i;
+  int j;
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  dst1 = _ctx->level[_l].im1;
+  dst2 = _ctx->level[_l].im2;
+  w2 = _ctx->level[_l - 1].w;
+  h2 = _ctx->level[_l - 1].h;
+  src1 = _ctx->level[_l - 1].im1;
+  src2 = _ctx->level[_l - 1].im2;
+  for (j = 0; j < h; j++) {
+    int j0offs;
+    int j1offs;
+    j0offs = 2 * j * w2;
+    j1offs = FS_MINI(2 * j + 1, h2) * w2;
+    for (i = 0; i < w; i++) {
+      int i0;
+      int i1;
+      i0 = 2 * i;
+      i1 = FS_MINI(i0 + 1, w2);
+      dst1[j * w + i] = src1[j0offs + i0] + src1[j0offs + i1] +
+                        src1[j1offs + i0] + src1[j1offs + i1];
+      dst2[j * w + i] = src2[j0offs + i0] + src2[j0offs + i1] +
+                        src2[j1offs + i0] + src2[j1offs + i1];
+    }
+  }
+}
+
+static void fs_downsample_level0(fs_ctx *_ctx, const uint8_t *_src1,
+                                 int _s1ystride, const uint8_t *_src2,
+                                 int _s2ystride, int _w, int _h, uint32_t bd,
+                                 uint32_t shift) {
+  uint32_t *dst1;
+  uint32_t *dst2;
+  int w;
+  int h;
+  int i;
+  int j;
+  w = _ctx->level[0].w;
+  h = _ctx->level[0].h;
+  dst1 = _ctx->level[0].im1;
+  dst2 = _ctx->level[0].im2;
+  for (j = 0; j < h; j++) {
+    int j0;
+    int j1;
+    j0 = 2 * j;
+    j1 = FS_MINI(j0 + 1, _h);
+    for (i = 0; i < w; i++) {
+      int i0;
+      int i1;
+      i0 = 2 * i;
+      i1 = FS_MINI(i0 + 1, _w);
+      if (bd == 8 && shift == 0) {
+        dst1[j * w + i] =
+            _src1[j0 * _s1ystride + i0] + _src1[j0 * _s1ystride + i1] +
+            _src1[j1 * _s1ystride + i0] + _src1[j1 * _s1ystride + i1];
+        dst2[j * w + i] =
+            _src2[j0 * _s2ystride + i0] + _src2[j0 * _s2ystride + i1] +
+            _src2[j1 * _s2ystride + i0] + _src2[j1 * _s2ystride + i1];
+      } else {
+        uint16_t *src1s = CONVERT_TO_SHORTPTR(_src1);
+        uint16_t *src2s = CONVERT_TO_SHORTPTR(_src2);
+        dst1[j * w + i] = (src1s[j0 * _s1ystride + i0] >> shift) +
+                          (src1s[j0 * _s1ystride + i1] >> shift) +
+                          (src1s[j1 * _s1ystride + i0] >> shift) +
+                          (src1s[j1 * _s1ystride + i1] >> shift);
+        dst2[j * w + i] = (src2s[j0 * _s2ystride + i0] >> shift) +
+                          (src2s[j0 * _s2ystride + i1] >> shift) +
+                          (src2s[j1 * _s2ystride + i0] >> shift) +
+                          (src2s[j1 * _s2ystride + i1] >> shift);
+      }
+    }
+  }
+}
+
+static void fs_apply_luminance(fs_ctx *_ctx, int _l, int bit_depth) {
+  unsigned *col_sums_x;
+  unsigned *col_sums_y;
+  uint32_t *im1;
+  uint32_t *im2;
+  double *ssim;
+  double c1;
+  int w;
+  int h;
+  int j0offs;
+  int j1offs;
+  int i;
+  int j;
+  double ssim_c1 = SSIM_C1;
+#if CONFIG_HIGHBITDEPTH
+  if (bit_depth == 10) ssim_c1 = SSIM_C1_10;
+  if (bit_depth == 12) ssim_c1 = SSIM_C1_12;
+#else
+  assert(bit_depth == 8);
+  (void)bit_depth;
+#endif
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  col_sums_x = _ctx->col_buf;
+  col_sums_y = col_sums_x + w;
+  im1 = _ctx->level[_l].im1;
+  im2 = _ctx->level[_l].im2;
+  for (i = 0; i < w; i++) col_sums_x[i] = 5 * im1[i];
+  for (i = 0; i < w; i++) col_sums_y[i] = 5 * im2[i];
+  for (j = 1; j < 4; j++) {
+    j1offs = FS_MINI(j, h - 1) * w;
+    for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i];
+    for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i];
+  }
+  ssim = _ctx->level[_l].ssim;
+  c1 = (double)(ssim_c1 * 4096 * (1 << 4 * _l));
+  for (j = 0; j < h; j++) {
+    unsigned mux;
+    unsigned muy;
+    int i0;
+    int i1;
+    mux = 5 * col_sums_x[0];
+    muy = 5 * col_sums_y[0];
+    for (i = 1; i < 4; i++) {
+      i1 = FS_MINI(i, w - 1);
+      mux += col_sums_x[i1];
+      muy += col_sums_y[i1];
+    }
+    for (i = 0; i < w; i++) {
+      ssim[j * w + i] *= (2 * mux * (double)muy + c1) /
+                         (mux * (double)mux + muy * (double)muy + c1);
+      if (i + 1 < w) {
+        i0 = FS_MAXI(0, i - 4);
+        i1 = FS_MINI(i + 4, w - 1);
+        mux += col_sums_x[i1] - col_sums_x[i0];
+        muy += col_sums_x[i1] - col_sums_x[i0];
+      }
+    }
+    if (j + 1 < h) {
+      j0offs = FS_MAXI(0, j - 4) * w;
+      for (i = 0; i < w; i++) col_sums_x[i] -= im1[j0offs + i];
+      for (i = 0; i < w; i++) col_sums_y[i] -= im2[j0offs + i];
+      j1offs = FS_MINI(j + 4, h - 1) * w;
+      for (i = 0; i < w; i++) col_sums_x[i] += im1[j1offs + i];
+      for (i = 0; i < w; i++) col_sums_y[i] += im2[j1offs + i];
+    }
+  }
+}
+
+#define FS_COL_SET(_col, _joffs, _ioffs)                       \
+  do {                                                         \
+    unsigned gx;                                               \
+    unsigned gy;                                               \
+    gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    col_sums_gx2[(_col)] = gx * (double)gx;                    \
+    col_sums_gy2[(_col)] = gy * (double)gy;                    \
+    col_sums_gxgy[(_col)] = gx * (double)gy;                   \
+  } while (0)
+
+#define FS_COL_ADD(_col, _joffs, _ioffs)                       \
+  do {                                                         \
+    unsigned gx;                                               \
+    unsigned gy;                                               \
+    gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    col_sums_gx2[(_col)] += gx * (double)gx;                   \
+    col_sums_gy2[(_col)] += gy * (double)gy;                   \
+    col_sums_gxgy[(_col)] += gx * (double)gy;                  \
+  } while (0)
+
+#define FS_COL_SUB(_col, _joffs, _ioffs)                       \
+  do {                                                         \
+    unsigned gx;                                               \
+    unsigned gy;                                               \
+    gx = gx_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    gy = gy_buf[((j + (_joffs)) & 7) * stride + i + (_ioffs)]; \
+    col_sums_gx2[(_col)] -= gx * (double)gx;                   \
+    col_sums_gy2[(_col)] -= gy * (double)gy;                   \
+    col_sums_gxgy[(_col)] -= gx * (double)gy;                  \
+  } while (0)
+
+#define FS_COL_COPY(_col1, _col2)                    \
+  do {                                               \
+    col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)];   \
+    col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)];   \
+    col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)]; \
+  } while (0)
+
+#define FS_COL_HALVE(_col1, _col2)                         \
+  do {                                                     \
+    col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 0.5;   \
+    col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 0.5;   \
+    col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 0.5; \
+  } while (0)
+
+#define FS_COL_DOUBLE(_col1, _col2)                      \
+  do {                                                   \
+    col_sums_gx2[(_col1)] = col_sums_gx2[(_col2)] * 2;   \
+    col_sums_gy2[(_col1)] = col_sums_gy2[(_col2)] * 2;   \
+    col_sums_gxgy[(_col1)] = col_sums_gxgy[(_col2)] * 2; \
+  } while (0)
+
+static void fs_calc_structure(fs_ctx *_ctx, int _l, int bit_depth) {
+  uint32_t *im1;
+  uint32_t *im2;
+  unsigned *gx_buf;
+  unsigned *gy_buf;
+  double *ssim;
+  double col_sums_gx2[8];
+  double col_sums_gy2[8];
+  double col_sums_gxgy[8];
+  double c2;
+  int stride;
+  int w;
+  int h;
+  int i;
+  int j;
+  double ssim_c2 = SSIM_C2;
+#if CONFIG_HIGHBITDEPTH
+  if (bit_depth == 10) ssim_c2 = SSIM_C2_10;
+  if (bit_depth == 12) ssim_c2 = SSIM_C2_12;
+#else
+  assert(bit_depth == 8);
+  (void)bit_depth;
+#endif
+
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  im1 = _ctx->level[_l].im1;
+  im2 = _ctx->level[_l].im2;
+  ssim = _ctx->level[_l].ssim;
+  gx_buf = _ctx->col_buf;
+  stride = w + 8;
+  gy_buf = gx_buf + 8 * stride;
+  memset(gx_buf, 0, 2 * 8 * stride * sizeof(*gx_buf));
+  c2 = ssim_c2 * (1 << 4 * _l) * 16 * 104;
+  for (j = 0; j < h + 4; j++) {
+    if (j < h - 1) {
+      for (i = 0; i < w - 1; i++) {
+        unsigned g1;
+        unsigned g2;
+        unsigned gx;
+        unsigned gy;
+        g1 = abs((int)im1[(j + 1) * w + i + 1] - (int)im1[j * w + i]);
+        g2 = abs((int)im1[(j + 1) * w + i] - (int)im1[j * w + i + 1]);
+        gx = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
+        g1 = abs((int)im2[(j + 1) * w + i + 1] - (int)im2[j * w + i]);
+        g2 = abs((int)im2[(j + 1) * w + i] - (int)im2[j * w + i + 1]);
+        gy = 4 * FS_MAXI(g1, g2) + FS_MINI(g1, g2);
+        gx_buf[(j & 7) * stride + i + 4] = gx;
+        gy_buf[(j & 7) * stride + i + 4] = gy;
+      }
+    } else {
+      memset(gx_buf + (j & 7) * stride, 0, stride * sizeof(*gx_buf));
+      memset(gy_buf + (j & 7) * stride, 0, stride * sizeof(*gy_buf));
+    }
+    if (j >= 4) {
+      int k;
+      col_sums_gx2[3] = col_sums_gx2[2] = col_sums_gx2[1] = col_sums_gx2[0] = 0;
+      col_sums_gy2[3] = col_sums_gy2[2] = col_sums_gy2[1] = col_sums_gy2[0] = 0;
+      col_sums_gxgy[3] = col_sums_gxgy[2] = col_sums_gxgy[1] =
+          col_sums_gxgy[0] = 0;
+      for (i = 4; i < 8; i++) {
+        FS_COL_SET(i, -1, 0);
+        FS_COL_ADD(i, 0, 0);
+        for (k = 1; k < 8 - i; k++) {
+          FS_COL_DOUBLE(i, i);
+          FS_COL_ADD(i, -k - 1, 0);
+          FS_COL_ADD(i, k, 0);
+        }
+      }
+      for (i = 0; i < w; i++) {
+        double mugx2;
+        double mugy2;
+        double mugxgy;
+        mugx2 = col_sums_gx2[0];
+        for (k = 1; k < 8; k++) mugx2 += col_sums_gx2[k];
+        mugy2 = col_sums_gy2[0];
+        for (k = 1; k < 8; k++) mugy2 += col_sums_gy2[k];
+        mugxgy = col_sums_gxgy[0];
+        for (k = 1; k < 8; k++) mugxgy += col_sums_gxgy[k];
+        ssim[(j - 4) * w + i] = (2 * mugxgy + c2) / (mugx2 + mugy2 + c2);
+        if (i + 1 < w) {
+          FS_COL_SET(0, -1, 1);
+          FS_COL_ADD(0, 0, 1);
+          FS_COL_SUB(2, -3, 2);
+          FS_COL_SUB(2, 2, 2);
+          FS_COL_HALVE(1, 2);
+          FS_COL_SUB(3, -4, 3);
+          FS_COL_SUB(3, 3, 3);
+          FS_COL_HALVE(2, 3);
+          FS_COL_COPY(3, 4);
+          FS_COL_DOUBLE(4, 5);
+          FS_COL_ADD(4, -4, 5);
+          FS_COL_ADD(4, 3, 5);
+          FS_COL_DOUBLE(5, 6);
+          FS_COL_ADD(5, -3, 6);
+          FS_COL_ADD(5, 2, 6);
+          FS_COL_DOUBLE(6, 7);
+          FS_COL_ADD(6, -2, 7);
+          FS_COL_ADD(6, 1, 7);
+          FS_COL_SET(7, -1, 8);
+          FS_COL_ADD(7, 0, 8);
+        }
+      }
+    }
+  }
+}
+
+#define FS_NLEVELS (4)
+
+/*These weights were derived from the default weights found in Wang's original
+ Matlab implementation: {0.0448, 0.2856, 0.2363, 0.1333}.
+ We drop the finest scale and renormalize the rest to sum to 1.*/
+
+static const double FS_WEIGHTS[FS_NLEVELS] = {
+  0.2989654541015625, 0.3141326904296875, 0.2473602294921875, 0.1395416259765625
+};
+
+static double fs_average(fs_ctx *_ctx, int _l) {
+  double *ssim;
+  double ret;
+  int w;
+  int h;
+  int i;
+  int j;
+  w = _ctx->level[_l].w;
+  h = _ctx->level[_l].h;
+  ssim = _ctx->level[_l].ssim;
+  ret = 0;
+  for (j = 0; j < h; j++)
+    for (i = 0; i < w; i++) ret += ssim[j * w + i];
+  return pow(ret / (w * h), FS_WEIGHTS[_l]);
+}
+
+static double convert_ssim_db(double _ssim, double _weight) {
+  assert(_weight >= _ssim);
+  if ((_weight - _ssim) < 1e-10) return MAX_SSIM_DB;
+  return 10 * (log10(_weight) - log10(_weight - _ssim));
+}
+
+static double calc_ssim(const uint8_t *_src, int _systride, const uint8_t *_dst,
+                        int _dystride, int _w, int _h, uint32_t _bd,
+                        uint32_t _shift) {
+  fs_ctx ctx;
+  double ret;
+  int l;
+  ret = 1;
+  fs_ctx_init(&ctx, _w, _h, FS_NLEVELS);
+  fs_downsample_level0(&ctx, _src, _systride, _dst, _dystride, _w, _h, _bd,
+                       _shift);
+  for (l = 0; l < FS_NLEVELS - 1; l++) {
+    fs_calc_structure(&ctx, l, _bd);
+    ret *= fs_average(&ctx, l);
+    fs_downsample_level(&ctx, l + 1);
+  }
+  fs_calc_structure(&ctx, l, _bd);
+  fs_apply_luminance(&ctx, l, _bd);
+  ret *= fs_average(&ctx, l);
+  fs_ctx_clear(&ctx);
+  return ret;
+}
+
+double aom_calc_fastssim(const YV12_BUFFER_CONFIG *source,
+                         const YV12_BUFFER_CONFIG *dest, double *ssim_y,
+                         double *ssim_u, double *ssim_v, uint32_t bd,
+                         uint32_t in_bd) {
+  double ssimv;
+  uint32_t bd_shift = 0;
+  aom_clear_system_state();
+  assert(bd >= in_bd);
+
+  bd_shift = bd - in_bd;
+
+  *ssim_y = calc_ssim(source->y_buffer, source->y_stride, dest->y_buffer,
+                      dest->y_stride, source->y_crop_width,
+                      source->y_crop_height, in_bd, bd_shift);
+  *ssim_u = calc_ssim(source->u_buffer, source->uv_stride, dest->u_buffer,
+                      dest->uv_stride, source->uv_crop_width,
+                      source->uv_crop_height, in_bd, bd_shift);
+  *ssim_v = calc_ssim(source->v_buffer, source->uv_stride, dest->v_buffer,
+                      dest->uv_stride, source->uv_crop_width,
+                      source->uv_crop_height, in_bd, bd_shift);
+  ssimv = (*ssim_y) * .8 + .1 * ((*ssim_u) + (*ssim_v));
+  return convert_ssim_db(ssimv, 1.0);
+}
diff --git a/third_party/aom/aom_dsp/fwd_txfm.c b/third_party/aom/aom_dsp/fwd_txfm.c
new file mode 100644
index 0000000000..12ee02ba1f
--- /dev/null
+++ b/third_party/aom/aom_dsp/fwd_txfm.c
@@ -0,0 +1,809 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/fwd_txfm.h"
+#include <assert.h>
+#include "./aom_dsp_rtcd.h"
+
+void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride) {
+  // The 2D transform is done with two passes which are actually pretty
+  // similar. In the first one, we transform the columns and transpose
+  // the results. In the second one, we transform the rows. To achieve that,
+  // as the first pass results are transposed, we transpose the columns (that
+  // is the transposed rows) and transpose the results (so that it goes back
+  // in normal/row positions).
+  int pass;
+  // We need an intermediate buffer between passes.
+  tran_low_t intermediate[4 * 4];
+  const tran_low_t *in_low = NULL;
+  tran_low_t *out = intermediate;
+  // Do the two transform/transpose passes
+  for (pass = 0; pass < 2; ++pass) {
+    tran_high_t in_high[4];    // canbe16
+    tran_high_t step[4];       // canbe16
+    tran_high_t temp1, temp2;  // needs32
+    int i;
+    for (i = 0; i < 4; ++i) {
+      // Load inputs.
+      if (pass == 0) {
+        in_high[0] = input[0 * stride] * 16;
+        in_high[1] = input[1 * stride] * 16;
+        in_high[2] = input[2 * stride] * 16;
+        in_high[3] = input[3 * stride] * 16;
+        if (i == 0 && in_high[0]) {
+          ++in_high[0];
+        }
+      } else {
+        assert(in_low != NULL);
+        in_high[0] = in_low[0 * 4];
+        in_high[1] = in_low[1 * 4];
+        in_high[2] = in_low[2 * 4];
+        in_high[3] = in_low[3 * 4];
+        ++in_low;
+      }
+      // Transform.
+      step[0] = in_high[0] + in_high[3];
+      step[1] = in_high[1] + in_high[2];
+      step[2] = in_high[1] - in_high[2];
+      step[3] = in_high[0] - in_high[3];
+      temp1 = (step[0] + step[1]) * cospi_16_64;
+      temp2 = (step[0] - step[1]) * cospi_16_64;
+      out[0] = (tran_low_t)fdct_round_shift(temp1);
+      out[2] = (tran_low_t)fdct_round_shift(temp2);
+      temp1 = step[2] * cospi_24_64 + step[3] * cospi_8_64;
+      temp2 = -step[2] * cospi_8_64 + step[3] * cospi_24_64;
+      out[1] = (tran_low_t)fdct_round_shift(temp1);
+      out[3] = (tran_low_t)fdct_round_shift(temp2);
+      // Do next column (which is a transposed row in second/horizontal pass)
+      ++input;
+      out += 4;
+    }
+    // Setup in/out for next pass.
+    in_low = intermediate;
+    out = output;
+  }
+
+  {
+    int i, j;
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j) output[j + i * 4] = (output[j + i * 4] + 1) >> 2;
+    }
+  }
+}
+
+void aom_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  tran_low_t sum = 0;
+  for (r = 0; r < 4; ++r)
+    for (c = 0; c < 4; ++c) sum += input[r * stride + c];
+
+  output[0] = sum << 1;
+}
+
+void aom_fdct8x8_c(const int16_t *input, tran_low_t *final_output, int stride) {
+  int i, j;
+  tran_low_t intermediate[64];
+  int pass;
+  tran_low_t *output = intermediate;
+  const tran_low_t *in = NULL;
+
+  // Transform columns
+  for (pass = 0; pass < 2; ++pass) {
+    tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
+    tran_high_t t0, t1, t2, t3;                  // needs32
+    tran_high_t x0, x1, x2, x3;                  // canbe16
+
+    for (i = 0; i < 8; i++) {
+      // stage 1
+      if (pass == 0) {
+        s0 = (input[0 * stride] + input[7 * stride]) * 4;
+        s1 = (input[1 * stride] + input[6 * stride]) * 4;
+        s2 = (input[2 * stride] + input[5 * stride]) * 4;
+        s3 = (input[3 * stride] + input[4 * stride]) * 4;
+        s4 = (input[3 * stride] - input[4 * stride]) * 4;
+        s5 = (input[2 * stride] - input[5 * stride]) * 4;
+        s6 = (input[1 * stride] - input[6 * stride]) * 4;
+        s7 = (input[0 * stride] - input[7 * stride]) * 4;
+        ++input;
+      } else {
+        s0 = in[0 * 8] + in[7 * 8];
+        s1 = in[1 * 8] + in[6 * 8];
+        s2 = in[2 * 8] + in[5 * 8];
+        s3 = in[3 * 8] + in[4 * 8];
+        s4 = in[3 * 8] - in[4 * 8];
+        s5 = in[2 * 8] - in[5 * 8];
+        s6 = in[1 * 8] - in[6 * 8];
+        s7 = in[0 * 8] - in[7 * 8];
+        ++in;
+      }
+
+      // fdct4(step, step);
+      x0 = s0 + s3;
+      x1 = s1 + s2;
+      x2 = s1 - s2;
+      x3 = s0 - s3;
+      t0 = (x0 + x1) * cospi_16_64;
+      t1 = (x0 - x1) * cospi_16_64;
+      t2 = x2 * cospi_24_64 + x3 * cospi_8_64;
+      t3 = -x2 * cospi_8_64 + x3 * cospi_24_64;
+      output[0] = (tran_low_t)fdct_round_shift(t0);
+      output[2] = (tran_low_t)fdct_round_shift(t2);
+      output[4] = (tran_low_t)fdct_round_shift(t1);
+      output[6] = (tran_low_t)fdct_round_shift(t3);
+
+      // Stage 2
+      t0 = (s6 - s5) * cospi_16_64;
+      t1 = (s6 + s5) * cospi_16_64;
+      t2 = fdct_round_shift(t0);
+      t3 = fdct_round_shift(t1);
+
+      // Stage 3
+      x0 = s4 + t2;
+      x1 = s4 - t2;
+      x2 = s7 - t3;
+      x3 = s7 + t3;
+
+      // Stage 4
+      t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
+      t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
+      t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
+      t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
+      output[1] = (tran_low_t)fdct_round_shift(t0);
+      output[3] = (tran_low_t)fdct_round_shift(t2);
+      output[5] = (tran_low_t)fdct_round_shift(t1);
+      output[7] = (tran_low_t)fdct_round_shift(t3);
+      output += 8;
+    }
+    in = intermediate;
+    output = final_output;
+  }
+
+  // Rows
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j) final_output[j + i * 8] /= 2;
+  }
+}
+
+void aom_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  tran_low_t sum = 0;
+  for (r = 0; r < 8; ++r)
+    for (c = 0; c < 8; ++c) sum += input[r * stride + c];
+
+  output[0] = sum;
+}
+
+void aom_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride) {
+  // The 2D transform is done with two passes which are actually pretty
+  // similar. In the first one, we transform the columns and transpose
+  // the results. In the second one, we transform the rows. To achieve that,
+  // as the first pass results are transposed, we transpose the columns (that
+  // is the transposed rows) and transpose the results (so that it goes back
+  // in normal/row positions).
+  int pass;
+  // We need an intermediate buffer between passes.
+  tran_low_t intermediate[256];
+  const tran_low_t *in_low = NULL;
+  tran_low_t *out = intermediate;
+  // Do the two transform/transpose passes
+  for (pass = 0; pass < 2; ++pass) {
+    tran_high_t step1[8];      // canbe16
+    tran_high_t step2[8];      // canbe16
+    tran_high_t step3[8];      // canbe16
+    tran_high_t in_high[8];    // canbe16
+    tran_high_t temp1, temp2;  // needs32
+    int i;
+    for (i = 0; i < 16; i++) {
+      if (0 == pass) {
+        // Calculate input for the first 8 results.
+        in_high[0] = (input[0 * stride] + input[15 * stride]) * 4;
+        in_high[1] = (input[1 * stride] + input[14 * stride]) * 4;
+        in_high[2] = (input[2 * stride] + input[13 * stride]) * 4;
+        in_high[3] = (input[3 * stride] + input[12 * stride]) * 4;
+        in_high[4] = (input[4 * stride] + input[11 * stride]) * 4;
+        in_high[5] = (input[5 * stride] + input[10 * stride]) * 4;
+        in_high[6] = (input[6 * stride] + input[9 * stride]) * 4;
+        in_high[7] = (input[7 * stride] + input[8 * stride]) * 4;
+        // Calculate input for the next 8 results.
+        step1[0] = (input[7 * stride] - input[8 * stride]) * 4;
+        step1[1] = (input[6 * stride] - input[9 * stride]) * 4;
+        step1[2] = (input[5 * stride] - input[10 * stride]) * 4;
+        step1[3] = (input[4 * stride] - input[11 * stride]) * 4;
+        step1[4] = (input[3 * stride] - input[12 * stride]) * 4;
+        step1[5] = (input[2 * stride] - input[13 * stride]) * 4;
+        step1[6] = (input[1 * stride] - input[14 * stride]) * 4;
+        step1[7] = (input[0 * stride] - input[15 * stride]) * 4;
+      } else {
+        // Calculate input for the first 8 results.
+        assert(in_low != NULL);
+        in_high[0] = ((in_low[0 * 16] + 1) >> 2) + ((in_low[15 * 16] + 1) >> 2);
+        in_high[1] = ((in_low[1 * 16] + 1) >> 2) + ((in_low[14 * 16] + 1) >> 2);
+        in_high[2] = ((in_low[2 * 16] + 1) >> 2) + ((in_low[13 * 16] + 1) >> 2);
+        in_high[3] = ((in_low[3 * 16] + 1) >> 2) + ((in_low[12 * 16] + 1) >> 2);
+        in_high[4] = ((in_low[4 * 16] + 1) >> 2) + ((in_low[11 * 16] + 1) >> 2);
+        in_high[5] = ((in_low[5 * 16] + 1) >> 2) + ((in_low[10 * 16] + 1) >> 2);
+        in_high[6] = ((in_low[6 * 16] + 1) >> 2) + ((in_low[9 * 16] + 1) >> 2);
+        in_high[7] = ((in_low[7 * 16] + 1) >> 2) + ((in_low[8 * 16] + 1) >> 2);
+        // Calculate input for the next 8 results.
+        step1[0] = ((in_low[7 * 16] + 1) >> 2) - ((in_low[8 * 16] + 1) >> 2);
+        step1[1] = ((in_low[6 * 16] + 1) >> 2) - ((in_low[9 * 16] + 1) >> 2);
+        step1[2] = ((in_low[5 * 16] + 1) >> 2) - ((in_low[10 * 16] + 1) >> 2);
+        step1[3] = ((in_low[4 * 16] + 1) >> 2) - ((in_low[11 * 16] + 1) >> 2);
+        step1[4] = ((in_low[3 * 16] + 1) >> 2) - ((in_low[12 * 16] + 1) >> 2);
+        step1[5] = ((in_low[2 * 16] + 1) >> 2) - ((in_low[13 * 16] + 1) >> 2);
+        step1[6] = ((in_low[1 * 16] + 1) >> 2) - ((in_low[14 * 16] + 1) >> 2);
+        step1[7] = ((in_low[0 * 16] + 1) >> 2) - ((in_low[15 * 16] + 1) >> 2);
+        in_low++;
+      }
+      // Work on the first eight values; fdct8(input, even_results);
+      {
+        tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
+        tran_high_t t0, t1, t2, t3;                  // needs32
+        tran_high_t x0, x1, x2, x3;                  // canbe16
+
+        // stage 1
+        s0 = in_high[0] + in_high[7];
+        s1 = in_high[1] + in_high[6];
+        s2 = in_high[2] + in_high[5];
+        s3 = in_high[3] + in_high[4];
+        s4 = in_high[3] - in_high[4];
+        s5 = in_high[2] - in_high[5];
+        s6 = in_high[1] - in_high[6];
+        s7 = in_high[0] - in_high[7];
+
+        // fdct4(step, step);
+        x0 = s0 + s3;
+        x1 = s1 + s2;
+        x2 = s1 - s2;
+        x3 = s0 - s3;
+        t0 = (x0 + x1) * cospi_16_64;
+        t1 = (x0 - x1) * cospi_16_64;
+        t2 = x3 * cospi_8_64 + x2 * cospi_24_64;
+        t3 = x3 * cospi_24_64 - x2 * cospi_8_64;
+        out[0] = (tran_low_t)fdct_round_shift(t0);
+        out[4] = (tran_low_t)fdct_round_shift(t2);
+        out[8] = (tran_low_t)fdct_round_shift(t1);
+        out[12] = (tran_low_t)fdct_round_shift(t3);
+
+        // Stage 2
+        t0 = (s6 - s5) * cospi_16_64;
+        t1 = (s6 + s5) * cospi_16_64;
+        t2 = fdct_round_shift(t0);
+        t3 = fdct_round_shift(t1);
+
+        // Stage 3
+        x0 = s4 + t2;
+        x1 = s4 - t2;
+        x2 = s7 - t3;
+        x3 = s7 + t3;
+
+        // Stage 4
+        t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
+        t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
+        t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
+        t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
+        out[2] = (tran_low_t)fdct_round_shift(t0);
+        out[6] = (tran_low_t)fdct_round_shift(t2);
+        out[10] = (tran_low_t)fdct_round_shift(t1);
+        out[14] = (tran_low_t)fdct_round_shift(t3);
+      }
+      // Work on the next eight values; step1 -> odd_results
+      {
+        // step 2
+        temp1 = (step1[5] - step1[2]) * cospi_16_64;
+        temp2 = (step1[4] - step1[3]) * cospi_16_64;
+        step2[2] = fdct_round_shift(temp1);
+        step2[3] = fdct_round_shift(temp2);
+        temp1 = (step1[4] + step1[3]) * cospi_16_64;
+        temp2 = (step1[5] + step1[2]) * cospi_16_64;
+        step2[4] = fdct_round_shift(temp1);
+        step2[5] = fdct_round_shift(temp2);
+        // step 3
+        step3[0] = step1[0] + step2[3];
+        step3[1] = step1[1] + step2[2];
+        step3[2] = step1[1] - step2[2];
+        step3[3] = step1[0] - step2[3];
+        step3[4] = step1[7] - step2[4];
+        step3[5] = step1[6] - step2[5];
+        step3[6] = step1[6] + step2[5];
+        step3[7] = step1[7] + step2[4];
+        // step 4
+        temp1 = step3[1] * -cospi_8_64 + step3[6] * cospi_24_64;
+        temp2 = step3[2] * cospi_24_64 + step3[5] * cospi_8_64;
+        step2[1] = fdct_round_shift(temp1);
+        step2[2] = fdct_round_shift(temp2);
+        temp1 = step3[2] * cospi_8_64 - step3[5] * cospi_24_64;
+        temp2 = step3[1] * cospi_24_64 + step3[6] * cospi_8_64;
+        step2[5] = fdct_round_shift(temp1);
+        step2[6] = fdct_round_shift(temp2);
+        // step 5
+        step1[0] = step3[0] + step2[1];
+        step1[1] = step3[0] - step2[1];
+        step1[2] = step3[3] + step2[2];
+        step1[3] = step3[3] - step2[2];
+        step1[4] = step3[4] - step2[5];
+        step1[5] = step3[4] + step2[5];
+        step1[6] = step3[7] - step2[6];
+        step1[7] = step3[7] + step2[6];
+        // step 6
+        temp1 = step1[0] * cospi_30_64 + step1[7] * cospi_2_64;
+        temp2 = step1[1] * cospi_14_64 + step1[6] * cospi_18_64;
+        out[1] = (tran_low_t)fdct_round_shift(temp1);
+        out[9] = (tran_low_t)fdct_round_shift(temp2);
+        temp1 = step1[2] * cospi_22_64 + step1[5] * cospi_10_64;
+        temp2 = step1[3] * cospi_6_64 + step1[4] * cospi_26_64;
+        out[5] = (tran_low_t)fdct_round_shift(temp1);
+        out[13] = (tran_low_t)fdct_round_shift(temp2);
+        temp1 = step1[3] * -cospi_26_64 + step1[4] * cospi_6_64;
+        temp2 = step1[2] * -cospi_10_64 + step1[5] * cospi_22_64;
+        out[3] = (tran_low_t)fdct_round_shift(temp1);
+        out[11] = (tran_low_t)fdct_round_shift(temp2);
+        temp1 = step1[1] * -cospi_18_64 + step1[6] * cospi_14_64;
+        temp2 = step1[0] * -cospi_2_64 + step1[7] * cospi_30_64;
+        out[7] = (tran_low_t)fdct_round_shift(temp1);
+        out[15] = (tran_low_t)fdct_round_shift(temp2);
+      }
+      // Do next column (which is a transposed row in second/horizontal pass)
+      input++;
+      out += 16;
+    }
+    // Setup in/out for next pass.
+    in_low = intermediate;
+    out = output;
+  }
+}
+
+void aom_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  int sum = 0;
+  for (r = 0; r < 16; ++r)
+    for (c = 0; c < 16; ++c) sum += input[r * stride + c];
+
+  output[0] = (tran_low_t)(sum >> 1);
+}
+
+static INLINE tran_high_t dct_32_round(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  // TODO(debargha, peter.derivaz): Find new bounds for this assert,
+  // and make the bounds consts.
+  // assert(-131072 <= rv && rv <= 131071);
+  return rv;
+}
+
+static INLINE tran_high_t half_round_shift(tran_high_t input) {
+  tran_high_t rv = (input + 1 + (input < 0)) >> 2;
+  return rv;
+}
+
+void aom_fdct32(const tran_high_t *input, tran_high_t *output, int round) {
+  tran_high_t step[32];
+  // Stage 1
+  step[0] = input[0] + input[(32 - 1)];
+  step[1] = input[1] + input[(32 - 2)];
+  step[2] = input[2] + input[(32 - 3)];
+  step[3] = input[3] + input[(32 - 4)];
+  step[4] = input[4] + input[(32 - 5)];
+  step[5] = input[5] + input[(32 - 6)];
+  step[6] = input[6] + input[(32 - 7)];
+  step[7] = input[7] + input[(32 - 8)];
+  step[8] = input[8] + input[(32 - 9)];
+  step[9] = input[9] + input[(32 - 10)];
+  step[10] = input[10] + input[(32 - 11)];
+  step[11] = input[11] + input[(32 - 12)];
+  step[12] = input[12] + input[(32 - 13)];
+  step[13] = input[13] + input[(32 - 14)];
+  step[14] = input[14] + input[(32 - 15)];
+  step[15] = input[15] + input[(32 - 16)];
+  step[16] = -input[16] + input[(32 - 17)];
+  step[17] = -input[17] + input[(32 - 18)];
+  step[18] = -input[18] + input[(32 - 19)];
+  step[19] = -input[19] + input[(32 - 20)];
+  step[20] = -input[20] + input[(32 - 21)];
+  step[21] = -input[21] + input[(32 - 22)];
+  step[22] = -input[22] + input[(32 - 23)];
+  step[23] = -input[23] + input[(32 - 24)];
+  step[24] = -input[24] + input[(32 - 25)];
+  step[25] = -input[25] + input[(32 - 26)];
+  step[26] = -input[26] + input[(32 - 27)];
+  step[27] = -input[27] + input[(32 - 28)];
+  step[28] = -input[28] + input[(32 - 29)];
+  step[29] = -input[29] + input[(32 - 30)];
+  step[30] = -input[30] + input[(32 - 31)];
+  step[31] = -input[31] + input[(32 - 32)];
+
+  // Stage 2
+  output[0] = step[0] + step[16 - 1];
+  output[1] = step[1] + step[16 - 2];
+  output[2] = step[2] + step[16 - 3];
+  output[3] = step[3] + step[16 - 4];
+  output[4] = step[4] + step[16 - 5];
+  output[5] = step[5] + step[16 - 6];
+  output[6] = step[6] + step[16 - 7];
+  output[7] = step[7] + step[16 - 8];
+  output[8] = -step[8] + step[16 - 9];
+  output[9] = -step[9] + step[16 - 10];
+  output[10] = -step[10] + step[16 - 11];
+  output[11] = -step[11] + step[16 - 12];
+  output[12] = -step[12] + step[16 - 13];
+  output[13] = -step[13] + step[16 - 14];
+  output[14] = -step[14] + step[16 - 15];
+  output[15] = -step[15] + step[16 - 16];
+
+  output[16] = step[16];
+  output[17] = step[17];
+  output[18] = step[18];
+  output[19] = step[19];
+
+  output[20] = dct_32_round((-step[20] + step[27]) * cospi_16_64);
+  output[21] = dct_32_round((-step[21] + step[26]) * cospi_16_64);
+  output[22] = dct_32_round((-step[22] + step[25]) * cospi_16_64);
+  output[23] = dct_32_round((-step[23] + step[24]) * cospi_16_64);
+
+  output[24] = dct_32_round((step[24] + step[23]) * cospi_16_64);
+  output[25] = dct_32_round((step[25] + step[22]) * cospi_16_64);
+  output[26] = dct_32_round((step[26] + step[21]) * cospi_16_64);
+  output[27] = dct_32_round((step[27] + step[20]) * cospi_16_64);
+
+  output[28] = step[28];
+  output[29] = step[29];
+  output[30] = step[30];
+  output[31] = step[31];
+
+  // dump the magnitude by 4, hence the intermediate values are within
+  // the range of 16 bits.
+  if (round) {
+    output[0] = half_round_shift(output[0]);
+    output[1] = half_round_shift(output[1]);
+    output[2] = half_round_shift(output[2]);
+    output[3] = half_round_shift(output[3]);
+    output[4] = half_round_shift(output[4]);
+    output[5] = half_round_shift(output[5]);
+    output[6] = half_round_shift(output[6]);
+    output[7] = half_round_shift(output[7]);
+    output[8] = half_round_shift(output[8]);
+    output[9] = half_round_shift(output[9]);
+    output[10] = half_round_shift(output[10]);
+    output[11] = half_round_shift(output[11]);
+    output[12] = half_round_shift(output[12]);
+    output[13] = half_round_shift(output[13]);
+    output[14] = half_round_shift(output[14]);
+    output[15] = half_round_shift(output[15]);
+
+    output[16] = half_round_shift(output[16]);
+    output[17] = half_round_shift(output[17]);
+    output[18] = half_round_shift(output[18]);
+    output[19] = half_round_shift(output[19]);
+    output[20] = half_round_shift(output[20]);
+    output[21] = half_round_shift(output[21]);
+    output[22] = half_round_shift(output[22]);
+    output[23] = half_round_shift(output[23]);
+    output[24] = half_round_shift(output[24]);
+    output[25] = half_round_shift(output[25]);
+    output[26] = half_round_shift(output[26]);
+    output[27] = half_round_shift(output[27]);
+    output[28] = half_round_shift(output[28]);
+    output[29] = half_round_shift(output[29]);
+    output[30] = half_round_shift(output[30]);
+    output[31] = half_round_shift(output[31]);
+  }
+
+  // Stage 3
+  step[0] = output[0] + output[(8 - 1)];
+  step[1] = output[1] + output[(8 - 2)];
+  step[2] = output[2] + output[(8 - 3)];
+  step[3] = output[3] + output[(8 - 4)];
+  step[4] = -output[4] + output[(8 - 5)];
+  step[5] = -output[5] + output[(8 - 6)];
+  step[6] = -output[6] + output[(8 - 7)];
+  step[7] = -output[7] + output[(8 - 8)];
+  step[8] = output[8];
+  step[9] = output[9];
+  step[10] = dct_32_round((-output[10] + output[13]) * cospi_16_64);
+  step[11] = dct_32_round((-output[11] + output[12]) * cospi_16_64);
+  step[12] = dct_32_round((output[12] + output[11]) * cospi_16_64);
+  step[13] = dct_32_round((output[13] + output[10]) * cospi_16_64);
+  step[14] = output[14];
+  step[15] = output[15];
+
+  step[16] = output[16] + output[23];
+  step[17] = output[17] + output[22];
+  step[18] = output[18] + output[21];
+  step[19] = output[19] + output[20];
+  step[20] = -output[20] + output[19];
+  step[21] = -output[21] + output[18];
+  step[22] = -output[22] + output[17];
+  step[23] = -output[23] + output[16];
+  step[24] = -output[24] + output[31];
+  step[25] = -output[25] + output[30];
+  step[26] = -output[26] + output[29];
+  step[27] = -output[27] + output[28];
+  step[28] = output[28] + output[27];
+  step[29] = output[29] + output[26];
+  step[30] = output[30] + output[25];
+  step[31] = output[31] + output[24];
+
+  // Stage 4
+  output[0] = step[0] + step[3];
+  output[1] = step[1] + step[2];
+  output[2] = -step[2] + step[1];
+  output[3] = -step[3] + step[0];
+  output[4] = step[4];
+  output[5] = dct_32_round((-step[5] + step[6]) * cospi_16_64);
+  output[6] = dct_32_round((step[6] + step[5]) * cospi_16_64);
+  output[7] = step[7];
+  output[8] = step[8] + step[11];
+  output[9] = step[9] + step[10];
+  output[10] = -step[10] + step[9];
+  output[11] = -step[11] + step[8];
+  output[12] = -step[12] + step[15];
+  output[13] = -step[13] + step[14];
+  output[14] = step[14] + step[13];
+  output[15] = step[15] + step[12];
+
+  output[16] = step[16];
+  output[17] = step[17];
+  output[18] = dct_32_round(step[18] * -cospi_8_64 + step[29] * cospi_24_64);
+  output[19] = dct_32_round(step[19] * -cospi_8_64 + step[28] * cospi_24_64);
+  output[20] = dct_32_round(step[20] * -cospi_24_64 + step[27] * -cospi_8_64);
+  output[21] = dct_32_round(step[21] * -cospi_24_64 + step[26] * -cospi_8_64);
+  output[22] = step[22];
+  output[23] = step[23];
+  output[24] = step[24];
+  output[25] = step[25];
+  output[26] = dct_32_round(step[26] * cospi_24_64 + step[21] * -cospi_8_64);
+  output[27] = dct_32_round(step[27] * cospi_24_64 + step[20] * -cospi_8_64);
+  output[28] = dct_32_round(step[28] * cospi_8_64 + step[19] * cospi_24_64);
+  output[29] = dct_32_round(step[29] * cospi_8_64 + step[18] * cospi_24_64);
+  output[30] = step[30];
+  output[31] = step[31];
+
+  // Stage 5
+  step[0] = dct_32_round((output[0] + output[1]) * cospi_16_64);
+  step[1] = dct_32_round((-output[1] + output[0]) * cospi_16_64);
+  step[2] = dct_32_round(output[2] * cospi_24_64 + output[3] * cospi_8_64);
+  step[3] = dct_32_round(output[3] * cospi_24_64 - output[2] * cospi_8_64);
+  step[4] = output[4] + output[5];
+  step[5] = -output[5] + output[4];
+  step[6] = -output[6] + output[7];
+  step[7] = output[7] + output[6];
+  step[8] = output[8];
+  step[9] = dct_32_round(output[9] * -cospi_8_64 + output[14] * cospi_24_64);
+  step[10] = dct_32_round(output[10] * -cospi_24_64 + output[13] * -cospi_8_64);
+  step[11] = output[11];
+  step[12] = output[12];
+  step[13] = dct_32_round(output[13] * cospi_24_64 + output[10] * -cospi_8_64);
+  step[14] = dct_32_round(output[14] * cospi_8_64 + output[9] * cospi_24_64);
+  step[15] = output[15];
+
+  step[16] = output[16] + output[19];
+  step[17] = output[17] + output[18];
+  step[18] = -output[18] + output[17];
+  step[19] = -output[19] + output[16];
+  step[20] = -output[20] + output[23];
+  step[21] = -output[21] + output[22];
+  step[22] = output[22] + output[21];
+  step[23] = output[23] + output[20];
+  step[24] = output[24] + output[27];
+  step[25] = output[25] + output[26];
+  step[26] = -output[26] + output[25];
+  step[27] = -output[27] + output[24];
+  step[28] = -output[28] + output[31];
+  step[29] = -output[29] + output[30];
+  step[30] = output[30] + output[29];
+  step[31] = output[31] + output[28];
+
+  // Stage 6
+  output[0] = step[0];
+  output[1] = step[1];
+  output[2] = step[2];
+  output[3] = step[3];
+  output[4] = dct_32_round(step[4] * cospi_28_64 + step[7] * cospi_4_64);
+  output[5] = dct_32_round(step[5] * cospi_12_64 + step[6] * cospi_20_64);
+  output[6] = dct_32_round(step[6] * cospi_12_64 + step[5] * -cospi_20_64);
+  output[7] = dct_32_round(step[7] * cospi_28_64 + step[4] * -cospi_4_64);
+  output[8] = step[8] + step[9];
+  output[9] = -step[9] + step[8];
+  output[10] = -step[10] + step[11];
+  output[11] = step[11] + step[10];
+  output[12] = step[12] + step[13];
+  output[13] = -step[13] + step[12];
+  output[14] = -step[14] + step[15];
+  output[15] = step[15] + step[14];
+
+  output[16] = step[16];
+  output[17] = dct_32_round(step[17] * -cospi_4_64 + step[30] * cospi_28_64);
+  output[18] = dct_32_round(step[18] * -cospi_28_64 + step[29] * -cospi_4_64);
+  output[19] = step[19];
+  output[20] = step[20];
+  output[21] = dct_32_round(step[21] * -cospi_20_64 + step[26] * cospi_12_64);
+  output[22] = dct_32_round(step[22] * -cospi_12_64 + step[25] * -cospi_20_64);
+  output[23] = step[23];
+  output[24] = step[24];
+  output[25] = dct_32_round(step[25] * cospi_12_64 + step[22] * -cospi_20_64);
+  output[26] = dct_32_round(step[26] * cospi_20_64 + step[21] * cospi_12_64);
+  output[27] = step[27];
+  output[28] = step[28];
+  output[29] = dct_32_round(step[29] * cospi_28_64 + step[18] * -cospi_4_64);
+  output[30] = dct_32_round(step[30] * cospi_4_64 + step[17] * cospi_28_64);
+  output[31] = step[31];
+
+  // Stage 7
+  step[0] = output[0];
+  step[1] = output[1];
+  step[2] = output[2];
+  step[3] = output[3];
+  step[4] = output[4];
+  step[5] = output[5];
+  step[6] = output[6];
+  step[7] = output[7];
+  step[8] = dct_32_round(output[8] * cospi_30_64 + output[15] * cospi_2_64);
+  step[9] = dct_32_round(output[9] * cospi_14_64 + output[14] * cospi_18_64);
+  step[10] = dct_32_round(output[10] * cospi_22_64 + output[13] * cospi_10_64);
+  step[11] = dct_32_round(output[11] * cospi_6_64 + output[12] * cospi_26_64);
+  step[12] = dct_32_round(output[12] * cospi_6_64 + output[11] * -cospi_26_64);
+  step[13] = dct_32_round(output[13] * cospi_22_64 + output[10] * -cospi_10_64);
+  step[14] = dct_32_round(output[14] * cospi_14_64 + output[9] * -cospi_18_64);
+  step[15] = dct_32_round(output[15] * cospi_30_64 + output[8] * -cospi_2_64);
+
+  step[16] = output[16] + output[17];
+  step[17] = -output[17] + output[16];
+  step[18] = -output[18] + output[19];
+  step[19] = output[19] + output[18];
+  step[20] = output[20] + output[21];
+  step[21] = -output[21] + output[20];
+  step[22] = -output[22] + output[23];
+  step[23] = output[23] + output[22];
+  step[24] = output[24] + output[25];
+  step[25] = -output[25] + output[24];
+  step[26] = -output[26] + output[27];
+  step[27] = output[27] + output[26];
+  step[28] = output[28] + output[29];
+  step[29] = -output[29] + output[28];
+  step[30] = -output[30] + output[31];
+  step[31] = output[31] + output[30];
+
+  // Final stage --- outputs indices are bit-reversed.
+  output[0] = step[0];
+  output[16] = step[1];
+  output[8] = step[2];
+  output[24] = step[3];
+  output[4] = step[4];
+  output[20] = step[5];
+  output[12] = step[6];
+  output[28] = step[7];
+  output[2] = step[8];
+  output[18] = step[9];
+  output[10] = step[10];
+  output[26] = step[11];
+  output[6] = step[12];
+  output[22] = step[13];
+  output[14] = step[14];
+  output[30] = step[15];
+
+  output[1] = dct_32_round(step[16] * cospi_31_64 + step[31] * cospi_1_64);
+  output[17] = dct_32_round(step[17] * cospi_15_64 + step[30] * cospi_17_64);
+  output[9] = dct_32_round(step[18] * cospi_23_64 + step[29] * cospi_9_64);
+  output[25] = dct_32_round(step[19] * cospi_7_64 + step[28] * cospi_25_64);
+  output[5] = dct_32_round(step[20] * cospi_27_64 + step[27] * cospi_5_64);
+  output[21] = dct_32_round(step[21] * cospi_11_64 + step[26] * cospi_21_64);
+  output[13] = dct_32_round(step[22] * cospi_19_64 + step[25] * cospi_13_64);
+  output[29] = dct_32_round(step[23] * cospi_3_64 + step[24] * cospi_29_64);
+  output[3] = dct_32_round(step[24] * cospi_3_64 + step[23] * -cospi_29_64);
+  output[19] = dct_32_round(step[25] * cospi_19_64 + step[22] * -cospi_13_64);
+  output[11] = dct_32_round(step[26] * cospi_11_64 + step[21] * -cospi_21_64);
+  output[27] = dct_32_round(step[27] * cospi_27_64 + step[20] * -cospi_5_64);
+  output[7] = dct_32_round(step[28] * cospi_7_64 + step[19] * -cospi_25_64);
+  output[23] = dct_32_round(step[29] * cospi_23_64 + step[18] * -cospi_9_64);
+  output[15] = dct_32_round(step[30] * cospi_15_64 + step[17] * -cospi_17_64);
+  output[31] = dct_32_round(step[31] * cospi_31_64 + step[16] * -cospi_1_64);
+}
+
+void aom_fdct32x32_c(const int16_t *input, tran_low_t *out, int stride) {
+  int i, j;
+  tran_high_t output[32 * 32];
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j) temp_in[j] = input[j * stride + i] * 4;
+    aom_fdct32(temp_in, temp_out, 0);
+    for (j = 0; j < 32; ++j)
+      output[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
+  }
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j) temp_in[j] = output[j + i * 32];
+    aom_fdct32(temp_in, temp_out, 0);
+    for (j = 0; j < 32; ++j)
+      out[j + i * 32] =
+          (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2);
+  }
+}
+
+// Note that although we use dct_32_round in dct32 computation flow,
+// this 2d fdct32x32 for rate-distortion optimization loop is operating
+// within 16 bits precision.
+void aom_fdct32x32_rd_c(const int16_t *input, tran_low_t *out, int stride) {
+  int i, j;
+  tran_high_t output[32 * 32];
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j) temp_in[j] = input[j * stride + i] * 4;
+    aom_fdct32(temp_in, temp_out, 0);
+    for (j = 0; j < 32; ++j)
+      // TODO(cd): see quality impact of only doing
+      //           output[j * 32 + i] = (temp_out[j] + 1) >> 2;
+      //           PS: also change code in aom_dsp/x86/aom_dct_sse2.c
+      output[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
+  }
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j) temp_in[j] = output[j + i * 32];
+    aom_fdct32(temp_in, temp_out, 1);
+    for (j = 0; j < 32; ++j) out[j + i * 32] = (tran_low_t)temp_out[j];
+  }
+}
+
+void aom_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride) {
+  int r, c;
+  int sum = 0;
+  for (r = 0; r < 32; ++r)
+    for (c = 0; c < 32; ++c) sum += input[r * stride + c];
+
+  output[0] = (tran_low_t)(sum >> 3);
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_fdct4x4_c(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  aom_fdct4x4_c(input, output, stride);
+}
+
+void aom_highbd_fdct8x8_c(const int16_t *input, tran_low_t *final_output,
+                          int stride) {
+  aom_fdct8x8_c(input, final_output, stride);
+}
+
+void aom_highbd_fdct8x8_1_c(const int16_t *input, tran_low_t *final_output,
+                            int stride) {
+  aom_fdct8x8_1_c(input, final_output, stride);
+}
+
+void aom_highbd_fdct16x16_c(const int16_t *input, tran_low_t *output,
+                            int stride) {
+  aom_fdct16x16_c(input, output, stride);
+}
+
+void aom_highbd_fdct16x16_1_c(const int16_t *input, tran_low_t *output,
+                              int stride) {
+  aom_fdct16x16_1_c(input, output, stride);
+}
+
+void aom_highbd_fdct32x32_c(const int16_t *input, tran_low_t *out, int stride) {
+  aom_fdct32x32_c(input, out, stride);
+}
+
+void aom_highbd_fdct32x32_rd_c(const int16_t *input, tran_low_t *out,
+                               int stride) {
+  aom_fdct32x32_rd_c(input, out, stride);
+}
+
+void aom_highbd_fdct32x32_1_c(const int16_t *input, tran_low_t *out,
+                              int stride) {
+  aom_fdct32x32_1_c(input, out, stride);
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/fwd_txfm.h b/third_party/aom/aom_dsp/fwd_txfm.h
new file mode 100644
index 0000000000..579dbd06e0
--- /dev/null
+++ b/third_party/aom/aom_dsp/fwd_txfm.h
@@ -0,0 +1,29 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_FWD_TXFM_H_
+#define AOM_DSP_FWD_TXFM_H_
+
+#include "aom_dsp/txfm_common.h"
+
+static INLINE tran_high_t saturate_int16(tran_high_t value) {
+  tran_high_t result;
+  result = value > INT16_MAX ? INT16_MAX : value;
+  return result < INT16_MIN ? INT16_MIN : result;
+}
+
+static INLINE tran_high_t fdct_round_shift(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  return rv;
+}
+
+void aom_fdct32(const tran_high_t *input, tran_high_t *output, int round);
+#endif  // AOM_DSP_FWD_TXFM_H_
diff --git a/third_party/aom/aom_dsp/intrapred.c b/third_party/aom/aom_dsp/intrapred.c
new file mode 100644
index 0000000000..1f0870b647
--- /dev/null
+++ b/third_party/aom/aom_dsp/intrapred.c
@@ -0,0 +1,971 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/bitops.h"
+
+#define DST(x, y) dst[(x) + (y)*stride]
+#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
+#define AVG2(a, b) (((a) + (b) + 1) >> 1)
+
+static INLINE void d207e_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                   const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  (void)above;
+
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = c & 1 ? AVG3(left[(c >> 1) + r], left[(c >> 1) + r + 1],
+                            left[(c >> 1) + r + 2])
+                     : AVG2(left[(c >> 1) + r], left[(c >> 1) + r + 1]);
+    }
+    dst += stride;
+  }
+}
+
+static INLINE void d63e_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  (void)left;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = r & 1 ? AVG3(above[(r >> 1) + c], above[(r >> 1) + c + 1],
+                            above[(r >> 1) + c + 2])
+                     : AVG2(above[(r >> 1) + c], above[(r >> 1) + c + 1]);
+    }
+    dst += stride;
+  }
+}
+
+static INLINE void d45e_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  (void)left;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = AVG3(above[r + c], above[r + c + 1],
+                    above[r + c + 1 + (r + c + 2 < bs * 2)]);
+    }
+    dst += stride;
+  }
+}
+
+static INLINE void d117_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+
+  // first row
+  for (c = 0; c < bs; c++) dst[c] = AVG2(above[c - 1], above[c]);
+  dst += stride;
+
+  // second row
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  for (c = 1; c < bs; c++) dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);
+  dst += stride;
+
+  // the rest of first col
+  dst[0] = AVG3(above[-1], left[0], left[1]);
+  for (r = 3; r < bs; ++r)
+    dst[(r - 2) * stride] = AVG3(left[r - 3], left[r - 2], left[r - 1]);
+
+  // the rest of the block
+  for (r = 2; r < bs; ++r) {
+    for (c = 1; c < bs; c++) dst[c] = dst[-2 * stride + c - 1];
+    dst += stride;
+  }
+}
+
+static INLINE void d135_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int i;
+#if CONFIG_TX64X64
+#if defined(__GNUC__) && __GNUC__ == 4 && __GNUC_MINOR__ > 7
+  // silence a spurious -Warray-bounds warning, possibly related to:
+  // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56273
+  uint8_t border[133];
+#else
+  uint8_t border[64 + 64 - 1];  // outer border from bottom-left to top-right
+#endif
+#else
+#if defined(__GNUC__) && __GNUC__ == 4 && __GNUC_MINOR__ > 7
+  // silence a spurious -Warray-bounds warning, possibly related to:
+  // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56273
+  uint8_t border[69];
+#else
+  uint8_t border[32 + 32 - 1];  // outer border from bottom-left to top-right
+#endif
+#endif  // CONFIG_TX64X64
+
+  // dst(bs, bs - 2)[0], i.e., border starting at bottom-left
+  for (i = 0; i < bs - 2; ++i) {
+    border[i] = AVG3(left[bs - 3 - i], left[bs - 2 - i], left[bs - 1 - i]);
+  }
+  border[bs - 2] = AVG3(above[-1], left[0], left[1]);
+  border[bs - 1] = AVG3(left[0], above[-1], above[0]);
+  border[bs - 0] = AVG3(above[-1], above[0], above[1]);
+  // dst[0][2, size), i.e., remaining top border ascending
+  for (i = 0; i < bs - 2; ++i) {
+    border[bs + 1 + i] = AVG3(above[i], above[i + 1], above[i + 2]);
+  }
+
+  for (i = 0; i < bs; ++i) {
+    memcpy(dst + i * stride, border + bs - 1 - i, bs);
+  }
+}
+
+static INLINE void d153_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                  const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  dst[0] = AVG2(above[-1], left[0]);
+  for (r = 1; r < bs; r++) dst[r * stride] = AVG2(left[r - 1], left[r]);
+  dst++;
+
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  dst[stride] = AVG3(above[-1], left[0], left[1]);
+  for (r = 2; r < bs; r++)
+    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);
+  dst++;
+
+  for (c = 0; c < bs - 2; c++)
+    dst[c] = AVG3(above[c - 1], above[c], above[c + 1]);
+  dst += stride;
+
+  for (r = 1; r < bs; ++r) {
+    for (c = 0; c < bs - 2; c++) dst[c] = dst[-stride + c - 2];
+    dst += stride;
+  }
+}
+
+static INLINE void v_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                               const uint8_t *above, const uint8_t *left) {
+  int r;
+  (void)left;
+
+  for (r = 0; r < bs; r++) {
+    memcpy(dst, above, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void h_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                               const uint8_t *above, const uint8_t *left) {
+  int r;
+  (void)above;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, left[r], bs);
+    dst += stride;
+  }
+}
+
+#if CONFIG_ALT_INTRA
+static INLINE int abs_diff(int a, int b) { return (a > b) ? a - b : b - a; }
+
+static INLINE uint16_t paeth_predictor_single(uint16_t left, uint16_t top,
+                                              uint16_t top_left) {
+  const int base = top + left - top_left;
+  const int p_left = abs_diff(base, left);
+  const int p_top = abs_diff(base, top);
+  const int p_top_left = abs_diff(base, top_left);
+
+  // Return nearest to base of left, top and top_left.
+  return (p_left <= p_top && p_left <= p_top_left)
+             ? left
+             : (p_top <= p_top_left) ? top : top_left;
+}
+
+static INLINE void paeth_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                   const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  const uint8_t ytop_left = above[-1];
+
+  for (r = 0; r < bs; r++) {
+    for (c = 0; c < bs; c++)
+      dst[c] = (uint8_t)paeth_predictor_single(left[r], above[c], ytop_left);
+    dst += stride;
+  }
+}
+
+// Weights are quadratic from '1' to '1 / block_size', scaled by
+// 2^sm_weight_log2_scale.
+static const int sm_weight_log2_scale = 8;
+
+#if CONFIG_TX64X64
+// max(block_size_wide[BLOCK_LARGEST], block_size_high[BLOCK_LARGEST])
+#define MAX_BLOCK_DIM 64
+#define NUM_BLOCK_DIMS 6  // log2(MAX_BLOCK_DIM)
+#else
+#define MAX_BLOCK_DIM 32
+#define NUM_BLOCK_DIMS 5
+#endif  // CONFIG_TX64X64
+
+static const uint8_t sm_weight_arrays[NUM_BLOCK_DIMS][MAX_BLOCK_DIM] = {
+  // bs = 2
+  { 255, 128 },
+  // bs = 4
+  { 255, 149, 85, 64 },
+  // bs = 8
+  { 255, 197, 146, 105, 73, 50, 37, 32 },
+  // bs = 16
+  { 255, 225, 196, 170, 145, 123, 102, 84, 68, 54, 43, 33, 26, 20, 17, 16 },
+  // bs = 32
+  {
+      255, 240, 225, 210, 196, 182, 169, 157, 145, 133, 122,
+      111, 101, 92,  83,  74,  66,  59,  52,  45,  39,  34,
+      29,  25,  21,  17,  14,  12,  10,  9,   8,   8 },
+#if CONFIG_TX64X64
+  // bs = 64
+  { 255, 248, 240, 233, 225, 218, 210, 203, 196, 189, 182, 176, 169,
+    163, 156, 150, 144, 138, 133, 127, 121, 116, 111, 106, 101, 96,
+    91,  86,  82,  77,  73,  69,  65,  61,  57,  54,  50,  47,  44,
+    41,  38,  35,  32,  29,  27,  25,  22,  20,  18,  16,  15,  13,
+    12,  10,  9,   8,   7,   6,   6,   5,   5,   4,   4,   4 },
+#endif  // CONFIG_TX64X64
+};
+
+// Some basic checks on weights for smooth predictor.
+#define sm_weights_sanity_checks(weights, weights_scale, pred_scale) \
+  assert(weights[0] < weights_scale);                                \
+  assert(weights_scale - weights[bs - 1] < weights_scale);           \
+  assert(pred_scale < 31)  // ensures no overflow when calculating predictor.
+
+#define divide_round(value, bits) (((value) + (1 << ((bits)-1))) >> (bits))
+
+static INLINE void smooth_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                    const uint8_t *above, const uint8_t *left) {
+  const uint8_t below_pred = left[bs - 1];   // estimated by bottom-left pixel
+  const uint8_t right_pred = above[bs - 1];  // estimated by top-right pixel
+  const int arr_index = get_msb(bs) - 1;
+  assert(arr_index >= 0);
+  assert(arr_index < NUM_BLOCK_DIMS);
+  const uint8_t *const sm_weights = sm_weight_arrays[arr_index];
+  // scale = 2 * 2^sm_weight_log2_scale
+  const int log2_scale = 1 + sm_weight_log2_scale;
+  const uint16_t scale = (1 << sm_weight_log2_scale);
+  sm_weights_sanity_checks(sm_weights, scale, log2_scale + sizeof(*dst));
+  int r;
+  for (r = 0; r < bs; ++r) {
+    int c;
+    for (c = 0; c < bs; ++c) {
+      const uint8_t pixels[] = { above[c], below_pred, left[r], right_pred };
+      const uint8_t weights[] = { sm_weights[r], scale - sm_weights[r],
+                                  sm_weights[c], scale - sm_weights[c] };
+      uint32_t this_pred = 0;
+      int i;
+      assert(scale >= sm_weights[r] && scale >= sm_weights[c]);
+      for (i = 0; i < 4; ++i) {
+        this_pred += weights[i] * pixels[i];
+      }
+      dst[c] = clip_pixel(divide_round(this_pred, log2_scale));
+    }
+    dst += stride;
+  }
+}
+
+#else
+
+static INLINE void tm_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                const uint8_t *above, const uint8_t *left) {
+  int r, c;
+  int ytop_left = above[-1];
+
+  for (r = 0; r < bs; r++) {
+    for (c = 0; c < bs; c++)
+      dst[c] = clip_pixel(left[r] + above[c] - ytop_left);
+    dst += stride;
+  }
+}
+#endif  // CONFIG_ALT_INTRA
+
+static INLINE void dc_128_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                    const uint8_t *above, const uint8_t *left) {
+  int r;
+  (void)above;
+  (void)left;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, 128, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void dc_left_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  int i, r, expected_dc, sum = 0;
+  (void)above;
+
+  for (i = 0; i < bs; i++) sum += left[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void dc_top_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                    const uint8_t *above, const uint8_t *left) {
+  int i, r, expected_dc, sum = 0;
+  (void)left;
+
+  for (i = 0; i < bs; i++) sum += above[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void dc_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
+                                const uint8_t *above, const uint8_t *left) {
+  int i, r, expected_dc, sum = 0;
+  const int count = 2 * bs;
+
+  for (i = 0; i < bs; i++) {
+    sum += above[i];
+    sum += left[i];
+  }
+
+  expected_dc = (sum + (count >> 1)) / count;
+
+  for (r = 0; r < bs; r++) {
+    memset(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+void aom_d45e_predictor_2x2_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  (void)stride;
+  (void)left;
+
+  DST(0, 0) = AVG3(A, B, C);
+  DST(1, 0) = DST(0, 1) = AVG3(B, C, D);
+  DST(1, 1) = AVG3(C, D, D);
+}
+
+void aom_d117_predictor_2x2_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  DST(0, 0) = AVG2(X, A);
+  DST(1, 0) = AVG2(A, B);
+  DST(0, 1) = AVG3(I, X, A);
+  DST(1, 1) = AVG3(X, A, B);
+}
+
+void aom_d135_predictor_2x2_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  (void)stride;
+  DST(0, 1) = AVG3(X, I, J);
+  DST(1, 1) = DST(0, 0) = AVG3(A, X, I);
+  DST(1, 0) = AVG3(B, A, X);
+}
+
+void aom_d153_predictor_2x2_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int X = above[-1];
+  const int A = above[0];
+
+  DST(0, 0) = AVG2(I, X);
+  DST(0, 1) = AVG2(J, I);
+  DST(1, 0) = AVG3(I, X, A);
+  DST(1, 1) = AVG3(J, I, X);
+}
+
+void aom_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  const int E = above[4];
+  const int F = above[5];
+  const int G = above[6];
+  const int H = above[7];
+  (void)stride;
+  (void)left;
+  DST(0, 0) = AVG3(A, B, C);
+  DST(1, 0) = DST(0, 1) = AVG3(B, C, D);
+  DST(2, 0) = DST(1, 1) = DST(0, 2) = AVG3(C, D, E);
+  DST(3, 0) = DST(2, 1) = DST(1, 2) = DST(0, 3) = AVG3(D, E, F);
+  DST(3, 1) = DST(2, 2) = DST(1, 3) = AVG3(E, F, G);
+  DST(3, 2) = DST(2, 3) = AVG3(F, G, H);
+  DST(3, 3) = AVG3(G, H, H);
+}
+
+void aom_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  DST(0, 0) = DST(1, 2) = AVG2(X, A);
+  DST(1, 0) = DST(2, 2) = AVG2(A, B);
+  DST(2, 0) = DST(3, 2) = AVG2(B, C);
+  DST(3, 0) = AVG2(C, D);
+
+  DST(0, 3) = AVG3(K, J, I);
+  DST(0, 2) = AVG3(J, I, X);
+  DST(0, 1) = DST(1, 3) = AVG3(I, X, A);
+  DST(1, 1) = DST(2, 3) = AVG3(X, A, B);
+  DST(2, 1) = DST(3, 3) = AVG3(A, B, C);
+  DST(3, 1) = AVG3(B, C, D);
+}
+
+void aom_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int L = left[3];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  (void)stride;
+  DST(0, 3) = AVG3(J, K, L);
+  DST(1, 3) = DST(0, 2) = AVG3(I, J, K);
+  DST(2, 3) = DST(1, 2) = DST(0, 1) = AVG3(X, I, J);
+  DST(3, 3) = DST(2, 2) = DST(1, 1) = DST(0, 0) = AVG3(A, X, I);
+  DST(3, 2) = DST(2, 1) = DST(1, 0) = AVG3(B, A, X);
+  DST(3, 1) = DST(2, 0) = AVG3(C, B, A);
+  DST(3, 0) = AVG3(D, C, B);
+}
+
+void aom_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int L = left[3];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+
+  DST(0, 0) = DST(2, 1) = AVG2(I, X);
+  DST(0, 1) = DST(2, 2) = AVG2(J, I);
+  DST(0, 2) = DST(2, 3) = AVG2(K, J);
+  DST(0, 3) = AVG2(L, K);
+
+  DST(3, 0) = AVG3(A, B, C);
+  DST(2, 0) = AVG3(X, A, B);
+  DST(1, 0) = DST(3, 1) = AVG3(I, X, A);
+  DST(1, 1) = DST(3, 2) = AVG3(J, I, X);
+  DST(1, 2) = DST(3, 3) = AVG3(K, J, I);
+  DST(1, 3) = AVG3(L, K, J);
+}
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE void highbd_d207e_predictor(uint16_t *dst, ptrdiff_t stride,
+                                          int bs, const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  int r, c;
+  (void)above;
+  (void)bd;
+
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = c & 1 ? AVG3(left[(c >> 1) + r], left[(c >> 1) + r + 1],
+                            left[(c >> 1) + r + 2])
+                     : AVG2(left[(c >> 1) + r], left[(c >> 1) + r + 1]);
+    }
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d63e_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void)left;
+  (void)bd;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = r & 1 ? AVG3(above[(r >> 1) + c], above[(r >> 1) + c + 1],
+                            above[(r >> 1) + c + 2])
+                     : AVG2(above[(r >> 1) + c], above[(r >> 1) + c + 1]);
+    }
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d45e_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void)left;
+  (void)bd;
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      dst[c] = AVG3(above[r + c], above[r + c + 1],
+                    above[r + c + 1 + (r + c + 2 < bs * 2)]);
+    }
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d117_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void)bd;
+
+  // first row
+  for (c = 0; c < bs; c++) dst[c] = AVG2(above[c - 1], above[c]);
+  dst += stride;
+
+  // second row
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  for (c = 1; c < bs; c++) dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);
+  dst += stride;
+
+  // the rest of first col
+  dst[0] = AVG3(above[-1], left[0], left[1]);
+  for (r = 3; r < bs; ++r)
+    dst[(r - 2) * stride] = AVG3(left[r - 3], left[r - 2], left[r - 1]);
+
+  // the rest of the block
+  for (r = 2; r < bs; ++r) {
+    for (c = 1; c < bs; c++) dst[c] = dst[-2 * stride + c - 1];
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d135_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void)bd;
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  for (c = 1; c < bs; c++) dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);
+
+  dst[stride] = AVG3(above[-1], left[0], left[1]);
+  for (r = 2; r < bs; ++r)
+    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);
+
+  dst += stride;
+  for (r = 1; r < bs; ++r) {
+    for (c = 1; c < bs; c++) dst[c] = dst[-stride + c - 1];
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_d153_predictor(uint16_t *dst, ptrdiff_t stride,
+                                         int bs, const uint16_t *above,
+                                         const uint16_t *left, int bd) {
+  int r, c;
+  (void)bd;
+  dst[0] = AVG2(above[-1], left[0]);
+  for (r = 1; r < bs; r++) dst[r * stride] = AVG2(left[r - 1], left[r]);
+  dst++;
+
+  dst[0] = AVG3(left[0], above[-1], above[0]);
+  dst[stride] = AVG3(above[-1], left[0], left[1]);
+  for (r = 2; r < bs; r++)
+    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);
+  dst++;
+
+  for (c = 0; c < bs - 2; c++)
+    dst[c] = AVG3(above[c - 1], above[c], above[c + 1]);
+  dst += stride;
+
+  for (r = 1; r < bs; ++r) {
+    for (c = 0; c < bs - 2; c++) dst[c] = dst[-stride + c - 2];
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_v_predictor(uint16_t *dst, ptrdiff_t stride, int bs,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  int r;
+  (void)left;
+  (void)bd;
+  for (r = 0; r < bs; r++) {
+    memcpy(dst, above, bs * sizeof(uint16_t));
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_h_predictor(uint16_t *dst, ptrdiff_t stride, int bs,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  int r;
+  (void)above;
+  (void)bd;
+  for (r = 0; r < bs; r++) {
+    aom_memset16(dst, left[r], bs);
+    dst += stride;
+  }
+}
+
+void aom_highbd_d207_predictor_2x2_c(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  const int I = left[0];
+  const int J = left[1];
+  const int K = left[2];
+  const int L = left[3];
+  (void)above;
+  (void)bd;
+  DST(0, 0) = AVG2(I, J);
+  DST(0, 1) = AVG2(J, K);
+  DST(1, 0) = AVG3(I, J, K);
+  DST(1, 1) = AVG3(J, K, L);
+}
+
+void aom_highbd_d63_predictor_2x2_c(uint16_t *dst, ptrdiff_t stride,
+                                    const uint16_t *above, const uint16_t *left,
+                                    int bd) {
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  (void)left;
+  (void)bd;
+  DST(0, 0) = AVG2(A, B);
+  DST(1, 0) = AVG2(B, C);
+  DST(0, 1) = AVG3(A, B, C);
+  DST(1, 1) = AVG3(B, C, D);
+}
+
+void aom_highbd_d45e_predictor_2x2_c(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  const int A = above[0];
+  const int B = above[1];
+  const int C = above[2];
+  const int D = above[3];
+  (void)stride;
+  (void)left;
+  (void)bd;
+  DST(0, 0) = AVG3(A, B, C);
+  DST(1, 0) = DST(0, 1) = AVG3(B, C, D);
+  DST(1, 1) = AVG3(C, D, D);
+}
+
+void aom_highbd_d117_predictor_2x2_c(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  const int I = left[0];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  (void)bd;
+  DST(0, 0) = AVG2(X, A);
+  DST(1, 0) = AVG2(A, B);
+  DST(0, 1) = AVG3(I, X, A);
+  DST(1, 1) = AVG3(X, A, B);
+}
+
+void aom_highbd_d135_predictor_2x2_c(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  const int I = left[0];
+  const int J = left[1];
+  const int X = above[-1];
+  const int A = above[0];
+  const int B = above[1];
+  (void)bd;
+  DST(0, 1) = AVG3(X, I, J);
+  DST(1, 1) = DST(0, 0) = AVG3(A, X, I);
+  DST(1, 0) = AVG3(B, A, X);
+}
+
+void aom_highbd_d153_predictor_2x2_c(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  const int I = left[0];
+  const int J = left[1];
+  const int X = above[-1];
+  const int A = above[0];
+  (void)bd;
+  DST(0, 0) = AVG2(I, X);
+  DST(0, 1) = AVG2(J, I);
+  DST(1, 0) = AVG3(I, X, A);
+  DST(1, 1) = AVG3(J, I, X);
+}
+
+#if CONFIG_ALT_INTRA
+static INLINE void highbd_paeth_predictor(uint16_t *dst, ptrdiff_t stride,
+                                          int bs, const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  int r, c;
+  const uint16_t ytop_left = above[-1];
+  (void)bd;
+
+  for (r = 0; r < bs; r++) {
+    for (c = 0; c < bs; c++)
+      dst[c] = paeth_predictor_single(left[r], above[c], ytop_left);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_smooth_predictor(uint16_t *dst, ptrdiff_t stride,
+                                           int bs, const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  const uint16_t below_pred = left[bs - 1];   // estimated by bottom-left pixel
+  const uint16_t right_pred = above[bs - 1];  // estimated by top-right pixel
+  const int arr_index = get_msb(bs) - 1;
+  assert(arr_index >= 0);
+  assert(arr_index < NUM_BLOCK_DIMS);
+  const uint8_t *const sm_weights = sm_weight_arrays[arr_index];
+  // scale = 2 * 2^sm_weight_log2_scale
+  const int log2_scale = 1 + sm_weight_log2_scale;
+  const uint16_t scale = (1 << sm_weight_log2_scale);
+  sm_weights_sanity_checks(sm_weights, scale, log2_scale + sizeof(*dst));
+  int r;
+  for (r = 0; r < bs; ++r) {
+    int c;
+    for (c = 0; c < bs; ++c) {
+      const uint16_t pixels[] = { above[c], below_pred, left[r], right_pred };
+      const uint8_t weights[] = { sm_weights[r], scale - sm_weights[r],
+                                  sm_weights[c], scale - sm_weights[c] };
+      uint32_t this_pred = 0;
+      int i;
+      assert(scale >= sm_weights[r] && scale >= sm_weights[c]);
+      for (i = 0; i < 4; ++i) {
+        this_pred += weights[i] * pixels[i];
+      }
+      dst[c] = clip_pixel_highbd(divide_round(this_pred, log2_scale), bd);
+    }
+    dst += stride;
+  }
+}
+
+#else
+static INLINE void highbd_tm_predictor(uint16_t *dst, ptrdiff_t stride, int bs,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int r, c;
+  int ytop_left = above[-1];
+  (void)bd;
+
+  for (r = 0; r < bs; r++) {
+    for (c = 0; c < bs; c++)
+      dst[c] = clip_pixel_highbd(left[r] + above[c] - ytop_left, bd);
+    dst += stride;
+  }
+}
+#endif  // CONFIG_ALT_INTRA
+
+static INLINE void highbd_dc_128_predictor(uint16_t *dst, ptrdiff_t stride,
+                                           int bs, const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  int r;
+  (void)above;
+  (void)left;
+
+  for (r = 0; r < bs; r++) {
+    aom_memset16(dst, 128 << (bd - 8), bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_left_predictor(uint16_t *dst, ptrdiff_t stride,
+                                            int bs, const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  int i, r, expected_dc, sum = 0;
+  (void)above;
+  (void)bd;
+
+  for (i = 0; i < bs; i++) sum += left[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    aom_memset16(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_top_predictor(uint16_t *dst, ptrdiff_t stride,
+                                           int bs, const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  int i, r, expected_dc, sum = 0;
+  (void)left;
+  (void)bd;
+
+  for (i = 0; i < bs; i++) sum += above[i];
+  expected_dc = (sum + (bs >> 1)) / bs;
+
+  for (r = 0; r < bs; r++) {
+    aom_memset16(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_dc_predictor(uint16_t *dst, ptrdiff_t stride, int bs,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int i, r, expected_dc, sum = 0;
+  const int count = 2 * bs;
+  (void)bd;
+
+  for (i = 0; i < bs; i++) {
+    sum += above[i];
+    sum += left[i];
+  }
+
+  expected_dc = (sum + (count >> 1)) / count;
+
+  for (r = 0; r < bs; r++) {
+    aom_memset16(dst, expected_dc, bs);
+    dst += stride;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+// This serves as a wrapper function, so that all the prediction functions
+// can be unified and accessed as a pointer array. Note that the boundary
+// above and left are not necessarily used all the time.
+#define intra_pred_sized(type, size)                        \
+  void aom_##type##_predictor_##size##x##size##_c(          \
+      uint8_t *dst, ptrdiff_t stride, const uint8_t *above, \
+      const uint8_t *left) {                                \
+    type##_predictor(dst, stride, size, above, left);       \
+  }
+
+#if CONFIG_HIGHBITDEPTH
+#define intra_pred_highbd_sized(type, size)                        \
+  void aom_highbd_##type##_predictor_##size##x##size##_c(          \
+      uint16_t *dst, ptrdiff_t stride, const uint16_t *above,      \
+      const uint16_t *left, int bd) {                              \
+    highbd_##type##_predictor(dst, stride, size, above, left, bd); \
+  }
+
+/* clang-format off */
+#if CONFIG_TX64X64
+#define intra_pred_allsizes(type) \
+  intra_pred_sized(type, 2) \
+  intra_pred_sized(type, 4) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32) \
+  intra_pred_sized(type, 64) \
+  intra_pred_highbd_sized(type, 4) \
+  intra_pred_highbd_sized(type, 8) \
+  intra_pred_highbd_sized(type, 16) \
+  intra_pred_highbd_sized(type, 32) \
+  intra_pred_highbd_sized(type, 64)
+
+#define intra_pred_above_4x4(type) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32) \
+  intra_pred_sized(type, 64) \
+  intra_pred_highbd_sized(type, 4) \
+  intra_pred_highbd_sized(type, 8) \
+  intra_pred_highbd_sized(type, 16) \
+  intra_pred_highbd_sized(type, 32) \
+  intra_pred_highbd_sized(type, 64)
+#else  // CONFIG_TX64X64
+#define intra_pred_allsizes(type) \
+  intra_pred_sized(type, 2) \
+  intra_pred_sized(type, 4) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32) \
+  intra_pred_highbd_sized(type, 2) \
+  intra_pred_highbd_sized(type, 4) \
+  intra_pred_highbd_sized(type, 8) \
+  intra_pred_highbd_sized(type, 16) \
+  intra_pred_highbd_sized(type, 32)
+
+#define intra_pred_above_4x4(type) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32) \
+  intra_pred_highbd_sized(type, 4) \
+  intra_pred_highbd_sized(type, 8) \
+  intra_pred_highbd_sized(type, 16) \
+  intra_pred_highbd_sized(type, 32)
+#endif  // CONFIG_TX64X64
+
+#else
+
+#if CONFIG_TX64X64
+#define intra_pred_allsizes(type) \
+  intra_pred_sized(type, 2) \
+  intra_pred_sized(type, 4) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32) \
+  intra_pred_sized(type, 64)
+
+#define intra_pred_above_4x4(type) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32) \
+  intra_pred_sized(type, 64)
+#else  // CONFIG_TX64X64
+#define intra_pred_allsizes(type) \
+  intra_pred_sized(type, 2) \
+  intra_pred_sized(type, 4) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32)
+
+#define intra_pred_above_4x4(type) \
+  intra_pred_sized(type, 8) \
+  intra_pred_sized(type, 16) \
+  intra_pred_sized(type, 32)
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_HIGHBITDEPTH
+
+intra_pred_allsizes(d207e)
+intra_pred_allsizes(d63e)
+intra_pred_above_4x4(d45e)
+intra_pred_above_4x4(d117)
+intra_pred_above_4x4(d135)
+intra_pred_above_4x4(d153)
+intra_pred_allsizes(v)
+intra_pred_allsizes(h)
+#if CONFIG_ALT_INTRA
+intra_pred_allsizes(paeth)
+intra_pred_allsizes(smooth)
+#else
+intra_pred_allsizes(tm)
+#endif  // CONFIG_ALT_INTRA
+intra_pred_allsizes(dc_128)
+intra_pred_allsizes(dc_left)
+intra_pred_allsizes(dc_top)
+intra_pred_allsizes(dc)
+/* clang-format on */
+#undef intra_pred_allsizes
diff --git a/third_party/aom/aom_dsp/inv_txfm.c b/third_party/aom/aom_dsp/inv_txfm.c
new file mode 100644
index 0000000000..bb995856ae
--- /dev/null
+++ b/third_party/aom/aom_dsp/inv_txfm.c
@@ -0,0 +1,1445 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <string.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/inv_txfm.h"
+
+void aom_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  /* 4-point reversible, orthonormal inverse Walsh-Hadamard in 3.5 adds,
+     0.5 shifts per pixel. */
+  int i;
+  tran_low_t output[16];
+  tran_high_t a1, b1, c1, d1, e1;
+  const tran_low_t *ip = input;
+  tran_low_t *op = output;
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip[0] >> UNIT_QUANT_SHIFT;
+    c1 = ip[1] >> UNIT_QUANT_SHIFT;
+    d1 = ip[2] >> UNIT_QUANT_SHIFT;
+    b1 = ip[3] >> UNIT_QUANT_SHIFT;
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    op[0] = WRAPLOW(a1);
+    op[1] = WRAPLOW(b1);
+    op[2] = WRAPLOW(c1);
+    op[3] = WRAPLOW(d1);
+    ip += 4;
+    op += 4;
+  }
+
+  ip = output;
+  for (i = 0; i < 4; i++) {
+    a1 = ip[4 * 0];
+    c1 = ip[4 * 1];
+    d1 = ip[4 * 2];
+    b1 = ip[4 * 3];
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    dest[stride * 0] = clip_pixel_add(dest[stride * 0], WRAPLOW(a1));
+    dest[stride * 1] = clip_pixel_add(dest[stride * 1], WRAPLOW(b1));
+    dest[stride * 2] = clip_pixel_add(dest[stride * 2], WRAPLOW(c1));
+    dest[stride * 3] = clip_pixel_add(dest[stride * 3], WRAPLOW(d1));
+
+    ip++;
+    dest++;
+  }
+}
+
+void aom_iwht4x4_1_add_c(const tran_low_t *in, uint8_t *dest, int dest_stride) {
+  int i;
+  tran_high_t a1, e1;
+  tran_low_t tmp[4];
+  const tran_low_t *ip = in;
+  tran_low_t *op = tmp;
+
+  a1 = ip[0] >> UNIT_QUANT_SHIFT;
+  e1 = a1 >> 1;
+  a1 -= e1;
+  op[0] = WRAPLOW(a1);
+  op[1] = op[2] = op[3] = WRAPLOW(e1);
+
+  ip = tmp;
+  for (i = 0; i < 4; i++) {
+    e1 = ip[0] >> 1;
+    a1 = ip[0] - e1;
+    dest[dest_stride * 0] = clip_pixel_add(dest[dest_stride * 0], a1);
+    dest[dest_stride * 1] = clip_pixel_add(dest[dest_stride * 1], e1);
+    dest[dest_stride * 2] = clip_pixel_add(dest[dest_stride * 2], e1);
+    dest[dest_stride * 3] = clip_pixel_add(dest[dest_stride * 3], e1);
+    ip++;
+    dest++;
+  }
+}
+
+void aom_idct4_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step[4];
+  tran_high_t temp1, temp2;
+  // stage 1
+  temp1 = (input[0] + input[2]) * cospi_16_64;
+  temp2 = (input[0] - input[2]) * cospi_16_64;
+  step[0] = WRAPLOW(dct_const_round_shift(temp1));
+  step[1] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+  temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+  step[2] = WRAPLOW(dct_const_round_shift(temp1));
+  step[3] = WRAPLOW(dct_const_round_shift(temp2));
+
+  // stage 2
+  output[0] = WRAPLOW(step[0] + step[3]);
+  output[1] = WRAPLOW(step[1] + step[2]);
+  output[2] = WRAPLOW(step[1] - step[2]);
+  output[3] = WRAPLOW(step[0] - step[3]);
+}
+
+void aom_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[4], temp_out[4];
+
+  // Rows
+  for (i = 0; i < 4; ++i) {
+    aom_idct4_c(input, outptr);
+    input += 4;
+    outptr += 4;
+  }
+
+  // Columns
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j) temp_in[j] = out[j * 4 + i];
+    aom_idct4_c(temp_in, temp_out);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 4));
+    }
+  }
+}
+
+void aom_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest,
+                         int dest_stride) {
+  int i;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64));
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
+  a1 = ROUND_POWER_OF_TWO(out, 4);
+
+  if (a1 == 0) return;
+
+  for (i = 0; i < 4; i++) {
+    dest[0] = clip_pixel_add(dest[0], a1);
+    dest[1] = clip_pixel_add(dest[1], a1);
+    dest[2] = clip_pixel_add(dest[2], a1);
+    dest[3] = clip_pixel_add(dest[3], a1);
+    dest += dest_stride;
+  }
+}
+
+void aom_idct8_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[8], step2[8];
+  tran_high_t temp1, temp2;
+  // stage 1
+  step1[0] = input[0];
+  step1[2] = input[4];
+  step1[1] = input[2];
+  step1[3] = input[6];
+  temp1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+  temp2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+  temp2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+
+  // stage 2
+  temp1 = (step1[0] + step1[2]) * cospi_16_64;
+  temp2 = (step1[0] - step1[2]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = step1[1] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[1] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[4] = WRAPLOW(step1[4] + step1[5]);
+  step2[5] = WRAPLOW(step1[4] - step1[5]);
+  step2[6] = WRAPLOW(-step1[6] + step1[7]);
+  step2[7] = WRAPLOW(step1[6] + step1[7]);
+
+  // stage 3
+  step1[0] = WRAPLOW(step2[0] + step2[3]);
+  step1[1] = WRAPLOW(step2[1] + step2[2]);
+  step1[2] = WRAPLOW(step2[1] - step2[2]);
+  step1[3] = WRAPLOW(step2[0] - step2[3]);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[7] = step2[7];
+
+  // stage 4
+  output[0] = WRAPLOW(step1[0] + step1[7]);
+  output[1] = WRAPLOW(step1[1] + step1[6]);
+  output[2] = WRAPLOW(step1[2] + step1[5]);
+  output[3] = WRAPLOW(step1[3] + step1[4]);
+  output[4] = WRAPLOW(step1[3] - step1[4]);
+  output[5] = WRAPLOW(step1[2] - step1[5]);
+  output[6] = WRAPLOW(step1[1] - step1[6]);
+  output[7] = WRAPLOW(step1[0] - step1[7]);
+}
+
+void aom_idct8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[8 * 8];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+
+  // First transform rows
+  for (i = 0; i < 8; ++i) {
+    aom_idct8_c(input, outptr);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i];
+    aom_idct8_c(temp_in, temp_out);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 5));
+    }
+  }
+}
+
+void aom_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64));
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
+  a1 = ROUND_POWER_OF_TWO(out, 5);
+  if (a1 == 0) return;
+  for (j = 0; j < 8; ++j) {
+    for (i = 0; i < 8; ++i) dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+void aom_iadst4_c(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_low_t x0 = input[0];
+  tran_low_t x1 = input[1];
+  tran_low_t x2 = input[2];
+  tran_low_t x3 = input[3];
+
+  if (!(x0 | x1 | x2 | x3)) {
+    output[0] = output[1] = output[2] = output[3] = 0;
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_2_9 * x0;
+  s2 = sinpi_3_9 * x1;
+  s3 = sinpi_4_9 * x2;
+  s4 = sinpi_1_9 * x2;
+  s5 = sinpi_2_9 * x3;
+  s6 = sinpi_4_9 * x3;
+  s7 = WRAPLOW(x0 - x2 + x3);
+
+  s0 = s0 + s3 + s5;
+  s1 = s1 - s4 - s6;
+  s3 = s2;
+  s2 = sinpi_3_9 * s7;
+
+  // 1-D transform scaling factor is sqrt(2).
+  // The overall dynamic range is 14b (input) + 14b (multiplication scaling)
+  // + 1b (addition) = 29b.
+  // Hence the output bit depth is 15b.
+  output[0] = WRAPLOW(dct_const_round_shift(s0 + s3));
+  output[1] = WRAPLOW(dct_const_round_shift(s1 + s3));
+  output[2] = WRAPLOW(dct_const_round_shift(s2));
+  output[3] = WRAPLOW(dct_const_round_shift(s0 + s1 - s3));
+}
+
+void aom_iadst8_c(const tran_low_t *input, tran_low_t *output) {
+  int s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_high_t x0 = input[7];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[5];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[3];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[1];
+  tran_high_t x7 = input[6];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7)) {
+    output[0] = output[1] = output[2] = output[3] = output[4] = output[5] =
+        output[6] = output[7] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = (int)(cospi_2_64 * x0 + cospi_30_64 * x1);
+  s1 = (int)(cospi_30_64 * x0 - cospi_2_64 * x1);
+  s2 = (int)(cospi_10_64 * x2 + cospi_22_64 * x3);
+  s3 = (int)(cospi_22_64 * x2 - cospi_10_64 * x3);
+  s4 = (int)(cospi_18_64 * x4 + cospi_14_64 * x5);
+  s5 = (int)(cospi_14_64 * x4 - cospi_18_64 * x5);
+  s6 = (int)(cospi_26_64 * x6 + cospi_6_64 * x7);
+  s7 = (int)(cospi_6_64 * x6 - cospi_26_64 * x7);
+
+  x0 = WRAPLOW(dct_const_round_shift(s0 + s4));
+  x1 = WRAPLOW(dct_const_round_shift(s1 + s5));
+  x2 = WRAPLOW(dct_const_round_shift(s2 + s6));
+  x3 = WRAPLOW(dct_const_round_shift(s3 + s7));
+  x4 = WRAPLOW(dct_const_round_shift(s0 - s4));
+  x5 = WRAPLOW(dct_const_round_shift(s1 - s5));
+  x6 = WRAPLOW(dct_const_round_shift(s2 - s6));
+  x7 = WRAPLOW(dct_const_round_shift(s3 - s7));
+
+  // stage 2
+  s0 = (int)x0;
+  s1 = (int)x1;
+  s2 = (int)x2;
+  s3 = (int)x3;
+  s4 = (int)(cospi_8_64 * x4 + cospi_24_64 * x5);
+  s5 = (int)(cospi_24_64 * x4 - cospi_8_64 * x5);
+  s6 = (int)(-cospi_24_64 * x6 + cospi_8_64 * x7);
+  s7 = (int)(cospi_8_64 * x6 + cospi_24_64 * x7);
+
+  x0 = WRAPLOW(s0 + s2);
+  x1 = WRAPLOW(s1 + s3);
+  x2 = WRAPLOW(s0 - s2);
+  x3 = WRAPLOW(s1 - s3);
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s6));
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s7));
+  x6 = WRAPLOW(dct_const_round_shift(s4 - s6));
+  x7 = WRAPLOW(dct_const_round_shift(s5 - s7));
+
+  // stage 3
+  s2 = (int)(cospi_16_64 * (x2 + x3));
+  s3 = (int)(cospi_16_64 * (x2 - x3));
+  s6 = (int)(cospi_16_64 * (x6 + x7));
+  s7 = (int)(cospi_16_64 * (x6 - x7));
+
+  x2 = WRAPLOW(dct_const_round_shift(s2));
+  x3 = WRAPLOW(dct_const_round_shift(s3));
+  x6 = WRAPLOW(dct_const_round_shift(s6));
+  x7 = WRAPLOW(dct_const_round_shift(s7));
+
+  output[0] = WRAPLOW(x0);
+  output[1] = WRAPLOW(-x4);
+  output[2] = WRAPLOW(x6);
+  output[3] = WRAPLOW(-x2);
+  output[4] = WRAPLOW(x3);
+  output[5] = WRAPLOW(-x7);
+  output[6] = WRAPLOW(x5);
+  output[7] = WRAPLOW(-x1);
+}
+
+void aom_idct8x8_12_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  tran_low_t out[8 * 8] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[8], temp_out[8];
+
+  // First transform rows
+  // only first 4 row has non-zero coefs
+  for (i = 0; i < 4; ++i) {
+    aom_idct8_c(input, outptr);
+    input += 8;
+    outptr += 8;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i];
+    aom_idct8_c(temp_in, temp_out);
+    for (j = 0; j < 8; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 5));
+    }
+  }
+}
+
+void aom_idct16_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[16], step2[16];
+  tran_high_t temp1, temp2;
+
+  // stage 1
+  step1[0] = input[0 / 2];
+  step1[1] = input[16 / 2];
+  step1[2] = input[8 / 2];
+  step1[3] = input[24 / 2];
+  step1[4] = input[4 / 2];
+  step1[5] = input[20 / 2];
+  step1[6] = input[12 / 2];
+  step1[7] = input[28 / 2];
+  step1[8] = input[2 / 2];
+  step1[9] = input[18 / 2];
+  step1[10] = input[10 / 2];
+  step1[11] = input[26 / 2];
+  step1[12] = input[6 / 2];
+  step1[13] = input[22 / 2];
+  step1[14] = input[14 / 2];
+  step1[15] = input[30 / 2];
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[15] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2));
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+
+  step1[8] = WRAPLOW(step2[8] + step2[9]);
+  step1[9] = WRAPLOW(step2[8] - step2[9]);
+  step1[10] = WRAPLOW(-step2[10] + step2[11]);
+  step1[11] = WRAPLOW(step2[10] + step2[11]);
+  step1[12] = WRAPLOW(step2[12] + step2[13]);
+  step1[13] = WRAPLOW(step2[12] - step2[13]);
+  step1[14] = WRAPLOW(-step2[14] + step2[15]);
+  step1[15] = WRAPLOW(step2[14] + step2[15]);
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[4] = WRAPLOW(step1[4] + step1[5]);
+  step2[5] = WRAPLOW(step1[4] - step1[5]);
+  step2[6] = WRAPLOW(-step1[6] + step1[7]);
+  step2[7] = WRAPLOW(step1[6] + step1[7]);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3]);
+  step1[1] = WRAPLOW(step2[1] + step2[2]);
+  step1[2] = WRAPLOW(step2[1] - step2[2]);
+  step1[3] = WRAPLOW(step2[0] - step2[3]);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11]);
+  step1[9] = WRAPLOW(step2[9] + step2[10]);
+  step1[10] = WRAPLOW(step2[9] - step2[10]);
+  step1[11] = WRAPLOW(step2[8] - step2[11]);
+  step1[12] = WRAPLOW(-step2[12] + step2[15]);
+  step1[13] = WRAPLOW(-step2[13] + step2[14]);
+  step1[14] = WRAPLOW(step2[13] + step2[14]);
+  step1[15] = WRAPLOW(step2[12] + step2[15]);
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7]);
+  step2[1] = WRAPLOW(step1[1] + step1[6]);
+  step2[2] = WRAPLOW(step1[2] + step1[5]);
+  step2[3] = WRAPLOW(step1[3] + step1[4]);
+  step2[4] = WRAPLOW(step1[3] - step1[4]);
+  step2[5] = WRAPLOW(step1[2] - step1[5]);
+  step2[6] = WRAPLOW(step1[1] - step1[6]);
+  step2[7] = WRAPLOW(step1[0] - step1[7]);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  // stage 7
+  output[0] = WRAPLOW(step2[0] + step2[15]);
+  output[1] = WRAPLOW(step2[1] + step2[14]);
+  output[2] = WRAPLOW(step2[2] + step2[13]);
+  output[3] = WRAPLOW(step2[3] + step2[12]);
+  output[4] = WRAPLOW(step2[4] + step2[11]);
+  output[5] = WRAPLOW(step2[5] + step2[10]);
+  output[6] = WRAPLOW(step2[6] + step2[9]);
+  output[7] = WRAPLOW(step2[7] + step2[8]);
+  output[8] = WRAPLOW(step2[7] - step2[8]);
+  output[9] = WRAPLOW(step2[6] - step2[9]);
+  output[10] = WRAPLOW(step2[5] - step2[10]);
+  output[11] = WRAPLOW(step2[4] - step2[11]);
+  output[12] = WRAPLOW(step2[3] - step2[12]);
+  output[13] = WRAPLOW(step2[2] - step2[13]);
+  output[14] = WRAPLOW(step2[1] - step2[14]);
+  output[15] = WRAPLOW(step2[0] - step2[15]);
+}
+
+void aom_idct16x16_256_add_c(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  tran_low_t out[16 * 16];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+
+  // First transform rows
+  for (i = 0; i < 16; ++i) {
+    aom_idct16_c(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j) temp_in[j] = out[j * 16 + i];
+    aom_idct16_c(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void aom_iadst16_c(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7, s8;
+  tran_high_t s9, s10, s11, s12, s13, s14, s15;
+
+  tran_high_t x0 = input[15];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[13];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[11];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[9];
+  tran_high_t x7 = input[6];
+  tran_high_t x8 = input[7];
+  tran_high_t x9 = input[8];
+  tran_high_t x10 = input[5];
+  tran_high_t x11 = input[10];
+  tran_high_t x12 = input[3];
+  tran_high_t x13 = input[12];
+  tran_high_t x14 = input[1];
+  tran_high_t x15 = input[14];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7 | x8 | x9 | x10 | x11 | x12 |
+        x13 | x14 | x15)) {
+    output[0] = output[1] = output[2] = output[3] = output[4] = output[5] =
+        output[6] = output[7] = output[8] = output[9] = output[10] =
+            output[11] = output[12] = output[13] = output[14] = output[15] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = x0 * cospi_1_64 + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64 + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64 + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64 - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64 - x15 * cospi_29_64;
+
+  x0 = WRAPLOW(dct_const_round_shift(s0 + s8));
+  x1 = WRAPLOW(dct_const_round_shift(s1 + s9));
+  x2 = WRAPLOW(dct_const_round_shift(s2 + s10));
+  x3 = WRAPLOW(dct_const_round_shift(s3 + s11));
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s12));
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s13));
+  x6 = WRAPLOW(dct_const_round_shift(s6 + s14));
+  x7 = WRAPLOW(dct_const_round_shift(s7 + s15));
+  x8 = WRAPLOW(dct_const_round_shift(s0 - s8));
+  x9 = WRAPLOW(dct_const_round_shift(s1 - s9));
+  x10 = WRAPLOW(dct_const_round_shift(s2 - s10));
+  x11 = WRAPLOW(dct_const_round_shift(s3 - s11));
+  x12 = WRAPLOW(dct_const_round_shift(s4 - s12));
+  x13 = WRAPLOW(dct_const_round_shift(s5 - s13));
+  x14 = WRAPLOW(dct_const_round_shift(s6 - s14));
+  x15 = WRAPLOW(dct_const_round_shift(s7 - s15));
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 = x8 * cospi_4_64 + x9 * cospi_28_64;
+  s9 = x8 * cospi_28_64 - x9 * cospi_4_64;
+  s10 = x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 = x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = -x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 = x12 * cospi_4_64 + x13 * cospi_28_64;
+  s14 = -x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 = x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = WRAPLOW(s0 + s4);
+  x1 = WRAPLOW(s1 + s5);
+  x2 = WRAPLOW(s2 + s6);
+  x3 = WRAPLOW(s3 + s7);
+  x4 = WRAPLOW(s0 - s4);
+  x5 = WRAPLOW(s1 - s5);
+  x6 = WRAPLOW(s2 - s6);
+  x7 = WRAPLOW(s3 - s7);
+  x8 = WRAPLOW(dct_const_round_shift(s8 + s12));
+  x9 = WRAPLOW(dct_const_round_shift(s9 + s13));
+  x10 = WRAPLOW(dct_const_round_shift(s10 + s14));
+  x11 = WRAPLOW(dct_const_round_shift(s11 + s15));
+  x12 = WRAPLOW(dct_const_round_shift(s8 - s12));
+  x13 = WRAPLOW(dct_const_round_shift(s9 - s13));
+  x14 = WRAPLOW(dct_const_round_shift(s10 - s14));
+  x15 = WRAPLOW(dct_const_round_shift(s11 - s15));
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64 + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = -x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 = x6 * cospi_8_64 + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64 + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = -x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 = x14 * cospi_8_64 + x15 * cospi_24_64;
+
+  x0 = WRAPLOW(s0 + s2);
+  x1 = WRAPLOW(s1 + s3);
+  x2 = WRAPLOW(s0 - s2);
+  x3 = WRAPLOW(s1 - s3);
+  x4 = WRAPLOW(dct_const_round_shift(s4 + s6));
+  x5 = WRAPLOW(dct_const_round_shift(s5 + s7));
+  x6 = WRAPLOW(dct_const_round_shift(s4 - s6));
+  x7 = WRAPLOW(dct_const_round_shift(s5 - s7));
+  x8 = WRAPLOW(s8 + s10);
+  x9 = WRAPLOW(s9 + s11);
+  x10 = WRAPLOW(s8 - s10);
+  x11 = WRAPLOW(s9 - s11);
+  x12 = WRAPLOW(dct_const_round_shift(s12 + s14));
+  x13 = WRAPLOW(dct_const_round_shift(s13 + s15));
+  x14 = WRAPLOW(dct_const_round_shift(s12 - s14));
+  x15 = WRAPLOW(dct_const_round_shift(s13 - s15));
+
+  // stage 4
+  s2 = (-cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (-x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (-x10 + x11);
+  s14 = (-cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = WRAPLOW(dct_const_round_shift(s2));
+  x3 = WRAPLOW(dct_const_round_shift(s3));
+  x6 = WRAPLOW(dct_const_round_shift(s6));
+  x7 = WRAPLOW(dct_const_round_shift(s7));
+  x10 = WRAPLOW(dct_const_round_shift(s10));
+  x11 = WRAPLOW(dct_const_round_shift(s11));
+  x14 = WRAPLOW(dct_const_round_shift(s14));
+  x15 = WRAPLOW(dct_const_round_shift(s15));
+
+  output[0] = WRAPLOW(x0);
+  output[1] = WRAPLOW(-x8);
+  output[2] = WRAPLOW(x12);
+  output[3] = WRAPLOW(-x4);
+  output[4] = WRAPLOW(x6);
+  output[5] = WRAPLOW(x14);
+  output[6] = WRAPLOW(x10);
+  output[7] = WRAPLOW(x2);
+  output[8] = WRAPLOW(x3);
+  output[9] = WRAPLOW(x11);
+  output[10] = WRAPLOW(x15);
+  output[11] = WRAPLOW(x7);
+  output[12] = WRAPLOW(x5);
+  output[13] = WRAPLOW(-x13);
+  output[14] = WRAPLOW(x9);
+  output[15] = WRAPLOW(-x1);
+}
+
+void aom_idct16x16_38_add_c(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  int i, j;
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  tran_low_t temp_in[16], temp_out[16];
+
+  // First transform rows. Since all non-zero dct coefficients are in
+  // upper-left 8x8 area, we only need to calculate first 8 rows here.
+  for (i = 0; i < 8; ++i) {
+    aom_idct16_c(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j) temp_in[j] = out[j * 16 + i];
+    aom_idct16_c(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void aom_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  tran_low_t out[16 * 16] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+
+  // First transform rows. Since all non-zero dct coefficients are in
+  // upper-left 4x4 area, we only need to calculate first 4 rows here.
+  for (i = 0; i < 4; ++i) {
+    aom_idct16_c(input, outptr);
+    input += 16;
+    outptr += 16;
+  }
+
+  // Then transform columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j) temp_in[j] = out[j * 16 + i];
+    aom_idct16_c(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void aom_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  int i, j;
+  tran_high_t a1;
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64));
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+  if (a1 == 0) return;
+  for (j = 0; j < 16; ++j) {
+    for (i = 0; i < 16; ++i) dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+void aom_idct32_c(const tran_low_t *input, tran_low_t *output) {
+  tran_low_t step1[32], step2[32];
+  tran_high_t temp1, temp2;
+
+  // stage 1
+  step1[0] = input[0];
+  step1[1] = input[16];
+  step1[2] = input[8];
+  step1[3] = input[24];
+  step1[4] = input[4];
+  step1[5] = input[20];
+  step1[6] = input[12];
+  step1[7] = input[28];
+  step1[8] = input[2];
+  step1[9] = input[18];
+  step1[10] = input[10];
+  step1[11] = input[26];
+  step1[12] = input[6];
+  step1[13] = input[22];
+  step1[14] = input[14];
+  step1[15] = input[30];
+
+  temp1 = input[1] * cospi_31_64 - input[31] * cospi_1_64;
+  temp2 = input[1] * cospi_1_64 + input[31] * cospi_31_64;
+  step1[16] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[31] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[17] * cospi_15_64 - input[15] * cospi_17_64;
+  temp2 = input[17] * cospi_17_64 + input[15] * cospi_15_64;
+  step1[17] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[30] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[9] * cospi_23_64 - input[23] * cospi_9_64;
+  temp2 = input[9] * cospi_9_64 + input[23] * cospi_23_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[25] * cospi_7_64 - input[7] * cospi_25_64;
+  temp2 = input[25] * cospi_25_64 + input[7] * cospi_7_64;
+  step1[19] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[28] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[5] * cospi_27_64 - input[27] * cospi_5_64;
+  temp2 = input[5] * cospi_5_64 + input[27] * cospi_27_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[21] * cospi_11_64 - input[11] * cospi_21_64;
+  temp2 = input[21] * cospi_21_64 + input[11] * cospi_11_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[13] * cospi_19_64 - input[19] * cospi_13_64;
+  temp2 = input[13] * cospi_13_64 + input[19] * cospi_19_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = input[29] * cospi_3_64 - input[3] * cospi_29_64;
+  temp2 = input[29] * cospi_29_64 + input[3] * cospi_3_64;
+  step1[23] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[24] = WRAPLOW(dct_const_round_shift(temp2));
+
+  // stage 2
+  step2[0] = step1[0];
+  step2[1] = step1[1];
+  step2[2] = step1[2];
+  step2[3] = step1[3];
+  step2[4] = step1[4];
+  step2[5] = step1[5];
+  step2[6] = step1[6];
+  step2[7] = step1[7];
+
+  temp1 = step1[8] * cospi_30_64 - step1[15] * cospi_2_64;
+  temp2 = step1[8] * cospi_2_64 + step1[15] * cospi_30_64;
+  step2[8] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[15] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[9] * cospi_14_64 - step1[14] * cospi_18_64;
+  temp2 = step1[9] * cospi_18_64 + step1[14] * cospi_14_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[10] * cospi_22_64 - step1[13] * cospi_10_64;
+  temp2 = step1[10] * cospi_10_64 + step1[13] * cospi_22_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+
+  temp1 = step1[11] * cospi_6_64 - step1[12] * cospi_26_64;
+  temp2 = step1[11] * cospi_26_64 + step1[12] * cospi_6_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2));
+
+  step2[16] = WRAPLOW(step1[16] + step1[17]);
+  step2[17] = WRAPLOW(step1[16] - step1[17]);
+  step2[18] = WRAPLOW(-step1[18] + step1[19]);
+  step2[19] = WRAPLOW(step1[18] + step1[19]);
+  step2[20] = WRAPLOW(step1[20] + step1[21]);
+  step2[21] = WRAPLOW(step1[20] - step1[21]);
+  step2[22] = WRAPLOW(-step1[22] + step1[23]);
+  step2[23] = WRAPLOW(step1[22] + step1[23]);
+  step2[24] = WRAPLOW(step1[24] + step1[25]);
+  step2[25] = WRAPLOW(step1[24] - step1[25]);
+  step2[26] = WRAPLOW(-step1[26] + step1[27]);
+  step2[27] = WRAPLOW(step1[26] + step1[27]);
+  step2[28] = WRAPLOW(step1[28] + step1[29]);
+  step2[29] = WRAPLOW(step1[28] - step1[29]);
+  step2[30] = WRAPLOW(-step1[30] + step1[31]);
+  step2[31] = WRAPLOW(step1[30] + step1[31]);
+
+  // stage 3
+  step1[0] = step2[0];
+  step1[1] = step2[1];
+  step1[2] = step2[2];
+  step1[3] = step2[3];
+
+  temp1 = step2[4] * cospi_28_64 - step2[7] * cospi_4_64;
+  temp2 = step2[4] * cospi_4_64 + step2[7] * cospi_28_64;
+  step1[4] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[7] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = step2[5] * cospi_12_64 - step2[6] * cospi_20_64;
+  temp2 = step2[5] * cospi_20_64 + step2[6] * cospi_12_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+
+  step1[8] = WRAPLOW(step2[8] + step2[9]);
+  step1[9] = WRAPLOW(step2[8] - step2[9]);
+  step1[10] = WRAPLOW(-step2[10] + step2[11]);
+  step1[11] = WRAPLOW(step2[10] + step2[11]);
+  step1[12] = WRAPLOW(step2[12] + step2[13]);
+  step1[13] = WRAPLOW(step2[12] - step2[13]);
+  step1[14] = WRAPLOW(-step2[14] + step2[15]);
+  step1[15] = WRAPLOW(step2[14] + step2[15]);
+
+  step1[16] = step2[16];
+  step1[31] = step2[31];
+  temp1 = -step2[17] * cospi_4_64 + step2[30] * cospi_28_64;
+  temp2 = step2[17] * cospi_28_64 + step2[30] * cospi_4_64;
+  step1[17] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[30] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step2[18] * cospi_28_64 - step2[29] * cospi_4_64;
+  temp2 = -step2[18] * cospi_4_64 + step2[29] * cospi_28_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[19] = step2[19];
+  step1[20] = step2[20];
+  temp1 = -step2[21] * cospi_20_64 + step2[26] * cospi_12_64;
+  temp2 = step2[21] * cospi_12_64 + step2[26] * cospi_20_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step2[22] * cospi_12_64 - step2[25] * cospi_20_64;
+  temp2 = -step2[22] * cospi_20_64 + step2[25] * cospi_12_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[27] = step2[27];
+  step1[28] = step2[28];
+
+  // stage 4
+  temp1 = (step1[0] + step1[1]) * cospi_16_64;
+  temp2 = (step1[0] - step1[1]) * cospi_16_64;
+  step2[0] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[1] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = step1[2] * cospi_24_64 - step1[3] * cospi_8_64;
+  temp2 = step1[2] * cospi_8_64 + step1[3] * cospi_24_64;
+  step2[2] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[3] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[4] = WRAPLOW(step1[4] + step1[5]);
+  step2[5] = WRAPLOW(step1[4] - step1[5]);
+  step2[6] = WRAPLOW(-step1[6] + step1[7]);
+  step2[7] = WRAPLOW(step1[6] + step1[7]);
+
+  step2[8] = step1[8];
+  step2[15] = step1[15];
+  temp1 = -step1[9] * cospi_8_64 + step1[14] * cospi_24_64;
+  temp2 = step1[9] * cospi_24_64 + step1[14] * cospi_8_64;
+  step2[9] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[14] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step1[10] * cospi_24_64 - step1[13] * cospi_8_64;
+  temp2 = -step1[10] * cospi_8_64 + step1[13] * cospi_24_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[11] = step1[11];
+  step2[12] = step1[12];
+
+  step2[16] = WRAPLOW(step1[16] + step1[19]);
+  step2[17] = WRAPLOW(step1[17] + step1[18]);
+  step2[18] = WRAPLOW(step1[17] - step1[18]);
+  step2[19] = WRAPLOW(step1[16] - step1[19]);
+  step2[20] = WRAPLOW(-step1[20] + step1[23]);
+  step2[21] = WRAPLOW(-step1[21] + step1[22]);
+  step2[22] = WRAPLOW(step1[21] + step1[22]);
+  step2[23] = WRAPLOW(step1[20] + step1[23]);
+
+  step2[24] = WRAPLOW(step1[24] + step1[27]);
+  step2[25] = WRAPLOW(step1[25] + step1[26]);
+  step2[26] = WRAPLOW(step1[25] - step1[26]);
+  step2[27] = WRAPLOW(step1[24] - step1[27]);
+  step2[28] = WRAPLOW(-step1[28] + step1[31]);
+  step2[29] = WRAPLOW(-step1[29] + step1[30]);
+  step2[30] = WRAPLOW(step1[29] + step1[30]);
+  step2[31] = WRAPLOW(step1[28] + step1[31]);
+
+  // stage 5
+  step1[0] = WRAPLOW(step2[0] + step2[3]);
+  step1[1] = WRAPLOW(step2[1] + step2[2]);
+  step1[2] = WRAPLOW(step2[1] - step2[2]);
+  step1[3] = WRAPLOW(step2[0] - step2[3]);
+  step1[4] = step2[4];
+  temp1 = (step2[6] - step2[5]) * cospi_16_64;
+  temp2 = (step2[5] + step2[6]) * cospi_16_64;
+  step1[5] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[6] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[7] = step2[7];
+
+  step1[8] = WRAPLOW(step2[8] + step2[11]);
+  step1[9] = WRAPLOW(step2[9] + step2[10]);
+  step1[10] = WRAPLOW(step2[9] - step2[10]);
+  step1[11] = WRAPLOW(step2[8] - step2[11]);
+  step1[12] = WRAPLOW(-step2[12] + step2[15]);
+  step1[13] = WRAPLOW(-step2[13] + step2[14]);
+  step1[14] = WRAPLOW(step2[13] + step2[14]);
+  step1[15] = WRAPLOW(step2[12] + step2[15]);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  temp1 = -step2[18] * cospi_8_64 + step2[29] * cospi_24_64;
+  temp2 = step2[18] * cospi_24_64 + step2[29] * cospi_8_64;
+  step1[18] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[29] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step2[19] * cospi_8_64 + step2[28] * cospi_24_64;
+  temp2 = step2[19] * cospi_24_64 + step2[28] * cospi_8_64;
+  step1[19] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[28] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step2[20] * cospi_24_64 - step2[27] * cospi_8_64;
+  temp2 = -step2[20] * cospi_8_64 + step2[27] * cospi_24_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = -step2[21] * cospi_24_64 - step2[26] * cospi_8_64;
+  temp2 = -step2[21] * cospi_8_64 + step2[26] * cospi_24_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[22] = step2[22];
+  step1[23] = step2[23];
+  step1[24] = step2[24];
+  step1[25] = step2[25];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // stage 6
+  step2[0] = WRAPLOW(step1[0] + step1[7]);
+  step2[1] = WRAPLOW(step1[1] + step1[6]);
+  step2[2] = WRAPLOW(step1[2] + step1[5]);
+  step2[3] = WRAPLOW(step1[3] + step1[4]);
+  step2[4] = WRAPLOW(step1[3] - step1[4]);
+  step2[5] = WRAPLOW(step1[2] - step1[5]);
+  step2[6] = WRAPLOW(step1[1] - step1[6]);
+  step2[7] = WRAPLOW(step1[0] - step1[7]);
+  step2[8] = step1[8];
+  step2[9] = step1[9];
+  temp1 = (-step1[10] + step1[13]) * cospi_16_64;
+  temp2 = (step1[10] + step1[13]) * cospi_16_64;
+  step2[10] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[13] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = (-step1[11] + step1[12]) * cospi_16_64;
+  temp2 = (step1[11] + step1[12]) * cospi_16_64;
+  step2[11] = WRAPLOW(dct_const_round_shift(temp1));
+  step2[12] = WRAPLOW(dct_const_round_shift(temp2));
+  step2[14] = step1[14];
+  step2[15] = step1[15];
+
+  step2[16] = WRAPLOW(step1[16] + step1[23]);
+  step2[17] = WRAPLOW(step1[17] + step1[22]);
+  step2[18] = WRAPLOW(step1[18] + step1[21]);
+  step2[19] = WRAPLOW(step1[19] + step1[20]);
+  step2[20] = WRAPLOW(step1[19] - step1[20]);
+  step2[21] = WRAPLOW(step1[18] - step1[21]);
+  step2[22] = WRAPLOW(step1[17] - step1[22]);
+  step2[23] = WRAPLOW(step1[16] - step1[23]);
+
+  step2[24] = WRAPLOW(-step1[24] + step1[31]);
+  step2[25] = WRAPLOW(-step1[25] + step1[30]);
+  step2[26] = WRAPLOW(-step1[26] + step1[29]);
+  step2[27] = WRAPLOW(-step1[27] + step1[28]);
+  step2[28] = WRAPLOW(step1[27] + step1[28]);
+  step2[29] = WRAPLOW(step1[26] + step1[29]);
+  step2[30] = WRAPLOW(step1[25] + step1[30]);
+  step2[31] = WRAPLOW(step1[24] + step1[31]);
+
+  // stage 7
+  step1[0] = WRAPLOW(step2[0] + step2[15]);
+  step1[1] = WRAPLOW(step2[1] + step2[14]);
+  step1[2] = WRAPLOW(step2[2] + step2[13]);
+  step1[3] = WRAPLOW(step2[3] + step2[12]);
+  step1[4] = WRAPLOW(step2[4] + step2[11]);
+  step1[5] = WRAPLOW(step2[5] + step2[10]);
+  step1[6] = WRAPLOW(step2[6] + step2[9]);
+  step1[7] = WRAPLOW(step2[7] + step2[8]);
+  step1[8] = WRAPLOW(step2[7] - step2[8]);
+  step1[9] = WRAPLOW(step2[6] - step2[9]);
+  step1[10] = WRAPLOW(step2[5] - step2[10]);
+  step1[11] = WRAPLOW(step2[4] - step2[11]);
+  step1[12] = WRAPLOW(step2[3] - step2[12]);
+  step1[13] = WRAPLOW(step2[2] - step2[13]);
+  step1[14] = WRAPLOW(step2[1] - step2[14]);
+  step1[15] = WRAPLOW(step2[0] - step2[15]);
+
+  step1[16] = step2[16];
+  step1[17] = step2[17];
+  step1[18] = step2[18];
+  step1[19] = step2[19];
+  temp1 = (-step2[20] + step2[27]) * cospi_16_64;
+  temp2 = (step2[20] + step2[27]) * cospi_16_64;
+  step1[20] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[27] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = (-step2[21] + step2[26]) * cospi_16_64;
+  temp2 = (step2[21] + step2[26]) * cospi_16_64;
+  step1[21] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[26] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = (-step2[22] + step2[25]) * cospi_16_64;
+  temp2 = (step2[22] + step2[25]) * cospi_16_64;
+  step1[22] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[25] = WRAPLOW(dct_const_round_shift(temp2));
+  temp1 = (-step2[23] + step2[24]) * cospi_16_64;
+  temp2 = (step2[23] + step2[24]) * cospi_16_64;
+  step1[23] = WRAPLOW(dct_const_round_shift(temp1));
+  step1[24] = WRAPLOW(dct_const_round_shift(temp2));
+  step1[28] = step2[28];
+  step1[29] = step2[29];
+  step1[30] = step2[30];
+  step1[31] = step2[31];
+
+  // final stage
+  output[0] = WRAPLOW(step1[0] + step1[31]);
+  output[1] = WRAPLOW(step1[1] + step1[30]);
+  output[2] = WRAPLOW(step1[2] + step1[29]);
+  output[3] = WRAPLOW(step1[3] + step1[28]);
+  output[4] = WRAPLOW(step1[4] + step1[27]);
+  output[5] = WRAPLOW(step1[5] + step1[26]);
+  output[6] = WRAPLOW(step1[6] + step1[25]);
+  output[7] = WRAPLOW(step1[7] + step1[24]);
+  output[8] = WRAPLOW(step1[8] + step1[23]);
+  output[9] = WRAPLOW(step1[9] + step1[22]);
+  output[10] = WRAPLOW(step1[10] + step1[21]);
+  output[11] = WRAPLOW(step1[11] + step1[20]);
+  output[12] = WRAPLOW(step1[12] + step1[19]);
+  output[13] = WRAPLOW(step1[13] + step1[18]);
+  output[14] = WRAPLOW(step1[14] + step1[17]);
+  output[15] = WRAPLOW(step1[15] + step1[16]);
+  output[16] = WRAPLOW(step1[15] - step1[16]);
+  output[17] = WRAPLOW(step1[14] - step1[17]);
+  output[18] = WRAPLOW(step1[13] - step1[18]);
+  output[19] = WRAPLOW(step1[12] - step1[19]);
+  output[20] = WRAPLOW(step1[11] - step1[20]);
+  output[21] = WRAPLOW(step1[10] - step1[21]);
+  output[22] = WRAPLOW(step1[9] - step1[22]);
+  output[23] = WRAPLOW(step1[8] - step1[23]);
+  output[24] = WRAPLOW(step1[7] - step1[24]);
+  output[25] = WRAPLOW(step1[6] - step1[25]);
+  output[26] = WRAPLOW(step1[5] - step1[26]);
+  output[27] = WRAPLOW(step1[4] - step1[27]);
+  output[28] = WRAPLOW(step1[3] - step1[28]);
+  output[29] = WRAPLOW(step1[2] - step1[29]);
+  output[30] = WRAPLOW(step1[1] - step1[30]);
+  output[31] = WRAPLOW(step1[0] - step1[31]);
+}
+
+void aom_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest,
+                              int stride) {
+  tran_low_t out[32 * 32];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    int16_t zero_coeff[16];
+    for (j = 0; j < 16; ++j) zero_coeff[j] = input[2 * j] | input[2 * j + 1];
+    for (j = 0; j < 8; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 4; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+    for (j = 0; j < 2; ++j)
+      zero_coeff[j] = zero_coeff[2 * j] | zero_coeff[2 * j + 1];
+
+    if (zero_coeff[0] | zero_coeff[1])
+      aom_idct32_c(input, outptr);
+    else
+      memset(outptr, 0, sizeof(tran_low_t) * 32);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j) temp_in[j] = out[j * 32 + i];
+    aom_idct32_c(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void aom_idct32x32_135_add_c(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  tran_low_t out[32 * 32] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+  // Rows
+  // only upper-left 16x16 has non-zero coeff
+  for (i = 0; i < 16; ++i) {
+    aom_idct32_c(input, outptr);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j) temp_in[j] = out[j * 32 + i];
+    aom_idct32_c(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void aom_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  tran_low_t out[32 * 32] = { 0 };
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+  // Rows
+  // only upper-left 8x8 has non-zero coeff
+  for (i = 0; i < 8; ++i) {
+    aom_idct32_c(input, outptr);
+    input += 32;
+    outptr += 32;
+  }
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j) temp_in[j] = out[j * 32 + i];
+    aom_idct32_c(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) {
+      dest[j * stride + i] = clip_pixel_add(dest[j * stride + i],
+                                            ROUND_POWER_OF_TWO(temp_out[j], 6));
+    }
+  }
+}
+
+void aom_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
+  int i, j;
+  tran_high_t a1;
+
+  tran_low_t out = WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64));
+  out = WRAPLOW(dct_const_round_shift(out * cospi_16_64));
+  a1 = ROUND_POWER_OF_TWO(out, 6);
+  if (a1 == 0) return;
+
+  for (j = 0; j < 32; ++j) {
+    for (i = 0; i < 32; ++i) dest[i] = clip_pixel_add(dest[i], a1);
+    dest += stride;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  /* 4-point reversible, orthonormal inverse Walsh-Hadamard in 3.5 adds,
+     0.5 shifts per pixel. */
+  int i;
+  tran_low_t output[16];
+  tran_high_t a1, b1, c1, d1, e1;
+  const tran_low_t *ip = input;
+  tran_low_t *op = output;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip[0] >> UNIT_QUANT_SHIFT;
+    c1 = ip[1] >> UNIT_QUANT_SHIFT;
+    d1 = ip[2] >> UNIT_QUANT_SHIFT;
+    b1 = ip[3] >> UNIT_QUANT_SHIFT;
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    op[0] = HIGHBD_WRAPLOW(a1, bd);
+    op[1] = HIGHBD_WRAPLOW(b1, bd);
+    op[2] = HIGHBD_WRAPLOW(c1, bd);
+    op[3] = HIGHBD_WRAPLOW(d1, bd);
+    ip += 4;
+    op += 4;
+  }
+
+  ip = output;
+  for (i = 0; i < 4; i++) {
+    a1 = ip[4 * 0];
+    c1 = ip[4 * 1];
+    d1 = ip[4 * 2];
+    b1 = ip[4 * 3];
+    a1 += c1;
+    d1 -= b1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= b1;
+    d1 += c1;
+    dest[stride * 0] =
+        highbd_clip_pixel_add(dest[stride * 0], HIGHBD_WRAPLOW(a1, bd), bd);
+    dest[stride * 1] =
+        highbd_clip_pixel_add(dest[stride * 1], HIGHBD_WRAPLOW(b1, bd), bd);
+    dest[stride * 2] =
+        highbd_clip_pixel_add(dest[stride * 2], HIGHBD_WRAPLOW(c1, bd), bd);
+    dest[stride * 3] =
+        highbd_clip_pixel_add(dest[stride * 3], HIGHBD_WRAPLOW(d1, bd), bd);
+
+    ip++;
+    dest++;
+  }
+}
+
+void aom_highbd_iwht4x4_1_add_c(const tran_low_t *in, uint8_t *dest8,
+                                int dest_stride, int bd) {
+  int i;
+  tran_high_t a1, e1;
+  tran_low_t tmp[4];
+  const tran_low_t *ip = in;
+  tran_low_t *op = tmp;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  (void)bd;
+
+  a1 = ip[0] >> UNIT_QUANT_SHIFT;
+  e1 = a1 >> 1;
+  a1 -= e1;
+  op[0] = HIGHBD_WRAPLOW(a1, bd);
+  op[1] = op[2] = op[3] = HIGHBD_WRAPLOW(e1, bd);
+
+  ip = tmp;
+  for (i = 0; i < 4; i++) {
+    e1 = ip[0] >> 1;
+    a1 = ip[0] - e1;
+    dest[dest_stride * 0] =
+        highbd_clip_pixel_add(dest[dest_stride * 0], a1, bd);
+    dest[dest_stride * 1] =
+        highbd_clip_pixel_add(dest[dest_stride * 1], e1, bd);
+    dest[dest_stride * 2] =
+        highbd_clip_pixel_add(dest[dest_stride * 2], e1, bd);
+    dest[dest_stride * 3] =
+        highbd_clip_pixel_add(dest[dest_stride * 3], e1, bd);
+    ip++;
+    dest++;
+  }
+}
+
+void aom_highbd_idct4_c(const tran_low_t *input, tran_low_t *output, int bd) {
+  tran_low_t step[4];
+  tran_high_t temp1, temp2;
+  (void)bd;
+  // stage 1
+  temp1 = (input[0] + input[2]) * cospi_16_64;
+  temp2 = (input[0] - input[2]) * cospi_16_64;
+  step[0] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd);
+  step[1] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd);
+  temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+  temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+  step[2] = HIGHBD_WRAPLOW(dct_const_round_shift(temp1), bd);
+  step[3] = HIGHBD_WRAPLOW(dct_const_round_shift(temp2), bd);
+
+  // stage 2
+  output[0] = HIGHBD_WRAPLOW(step[0] + step[3], bd);
+  output[1] = HIGHBD_WRAPLOW(step[1] + step[2], bd);
+  output[2] = HIGHBD_WRAPLOW(step[1] - step[2], bd);
+  output[3] = HIGHBD_WRAPLOW(step[0] - step[3], bd);
+}
+
+void aom_highbd_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int bd) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  tran_low_t temp_in[4], temp_out[4];
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  // Rows
+  for (i = 0; i < 4; ++i) {
+    aom_highbd_idct4_c(input, outptr, bd);
+    input += 4;
+    outptr += 4;
+  }
+
+  // Columns
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j) temp_in[j] = out[j * 4 + i];
+    aom_highbd_idct4_c(temp_in, temp_out, bd);
+    for (j = 0; j < 4; ++j) {
+      dest[j * stride + i] = highbd_clip_pixel_add(
+          dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
+    }
+  }
+}
+
+void aom_highbd_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int dest_stride, int bd) {
+  int i;
+  tran_high_t a1;
+  tran_low_t out =
+      HIGHBD_WRAPLOW(dct_const_round_shift(input[0] * cospi_16_64), bd);
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  out = HIGHBD_WRAPLOW(dct_const_round_shift(out * cospi_16_64), bd);
+  a1 = ROUND_POWER_OF_TWO(out, 4);
+
+  for (i = 0; i < 4; i++) {
+    dest[0] = highbd_clip_pixel_add(dest[0], a1, bd);
+    dest[1] = highbd_clip_pixel_add(dest[1], a1, bd);
+    dest[2] = highbd_clip_pixel_add(dest[2], a1, bd);
+    dest[3] = highbd_clip_pixel_add(dest[3], a1, bd);
+    dest += dest_stride;
+  }
+}
+
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/inv_txfm.h b/third_party/aom/aom_dsp/inv_txfm.h
new file mode 100644
index 0000000000..e64d463eae
--- /dev/null
+++ b/third_party/aom/aom_dsp/inv_txfm.h
@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_INV_TXFM_H_
+#define AOM_DSP_INV_TXFM_H_
+
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE tran_high_t dct_const_round_shift(tran_high_t input) {
+  tran_high_t rv = ROUND_POWER_OF_TWO(input, DCT_CONST_BITS);
+  return rv;
+}
+
+static INLINE tran_high_t check_range(tran_high_t input, int bd) {
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+  // For valid AV1 input streams, intermediate stage coefficients should always
+  // stay within the range of a signed 16 bit integer. Coefficients can go out
+  // of this range for invalid/corrupt AV1 streams. However, strictly checking
+  // this range for every intermediate coefficient can burdensome for a decoder,
+  // therefore the following assertion is only enabled when configured with
+  // --enable-coefficient-range-checking.
+  // For valid highbitdepth AV1 streams, intermediate stage coefficients will
+  // stay within the ranges:
+  // - 8 bit: signed 16 bit integer
+  // - 10 bit: signed 18 bit integer
+  // - 12 bit: signed 20 bit integer
+  const int32_t int_max = (1 << (7 + bd)) - 1;
+  const int32_t int_min = -int_max - 1;
+  assert(int_min <= input);
+  assert(input <= int_max);
+  (void)int_min;
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+  (void)bd;
+  return input;
+}
+
+#define WRAPLOW(x) ((int32_t)check_range(x, 8))
+#if CONFIG_HIGHBITDEPTH
+#define HIGHBD_WRAPLOW(x, bd) ((int32_t)check_range((x), bd))
+#endif  // CONFIG_HIGHBITDEPTH
+
+void aom_idct4_c(const tran_low_t *input, tran_low_t *output);
+void aom_idct8_c(const tran_low_t *input, tran_low_t *output);
+void aom_idct16_c(const tran_low_t *input, tran_low_t *output);
+void aom_idct32_c(const tran_low_t *input, tran_low_t *output);
+void aom_iadst4_c(const tran_low_t *input, tran_low_t *output);
+void aom_iadst8_c(const tran_low_t *input, tran_low_t *output);
+void aom_iadst16_c(const tran_low_t *input, tran_low_t *output);
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_idct4_c(const tran_low_t *input, tran_low_t *output, int bd);
+void aom_highbd_idct8_c(const tran_low_t *input, tran_low_t *output, int bd);
+void aom_highbd_idct16_c(const tran_low_t *input, tran_low_t *output, int bd);
+void aom_highbd_idct32_c(const tran_low_t *input, tran_low_t *output, int bd);
+
+void aom_highbd_iadst4_c(const tran_low_t *input, tran_low_t *output, int bd);
+void aom_highbd_iadst8_c(const tran_low_t *input, tran_low_t *output, int bd);
+void aom_highbd_iadst16_c(const tran_low_t *input, tran_low_t *output, int bd);
+
+static INLINE uint16_t highbd_clip_pixel_add(uint16_t dest, tran_high_t trans,
+                                             int bd) {
+  trans = HIGHBD_WRAPLOW(trans, bd);
+  return clip_pixel_highbd(dest + (int)trans, bd);
+}
+#endif
+
+static INLINE uint8_t clip_pixel_add(uint8_t dest, tran_high_t trans) {
+  trans = WRAPLOW(trans);
+  return clip_pixel(dest + (int)trans);
+}
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_INV_TXFM_H_
diff --git a/third_party/aom/aom_dsp/loopfilter.c b/third_party/aom/aom_dsp/loopfilter.c
new file mode 100644
index 0000000000..e2e8392194
--- /dev/null
+++ b/third_party/aom/aom_dsp/loopfilter.c
@@ -0,0 +1,900 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+static INLINE int8_t signed_char_clamp(int t) {
+  return (int8_t)clamp(t, -128, 127);
+}
+
+#define PARALLEL_DEBLOCKING_11_TAP 0
+#define PARALLEL_DEBLOCKING_9_TAP 0
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE int16_t signed_char_clamp_high(int t, int bd) {
+  switch (bd) {
+    case 10: return (int16_t)clamp(t, -128 * 4, 128 * 4 - 1);
+    case 12: return (int16_t)clamp(t, -128 * 16, 128 * 16 - 1);
+    case 8:
+    default: return (int16_t)clamp(t, -128, 128 - 1);
+  }
+}
+#endif
+#if CONFIG_PARALLEL_DEBLOCKING
+// should we apply any filter at all: 11111111 yes, 00000000 no
+static INLINE int8_t filter_mask2(uint8_t limit, uint8_t blimit, uint8_t p1,
+                                  uint8_t p0, uint8_t q0, uint8_t q1) {
+  int8_t mask = 0;
+  mask |= (abs(p1 - p0) > limit) * -1;
+  mask |= (abs(q1 - q0) > limit) * -1;
+  mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+  return ~mask;
+}
+#endif  // CONFIG_PARALLEL_DEBLOCKING
+static INLINE int8_t filter_mask(uint8_t limit, uint8_t blimit, uint8_t p3,
+                                 uint8_t p2, uint8_t p1, uint8_t p0, uint8_t q0,
+                                 uint8_t q1, uint8_t q2, uint8_t q3) {
+  int8_t mask = 0;
+  mask |= (abs(p3 - p2) > limit) * -1;
+  mask |= (abs(p2 - p1) > limit) * -1;
+  mask |= (abs(p1 - p0) > limit) * -1;
+  mask |= (abs(q1 - q0) > limit) * -1;
+  mask |= (abs(q2 - q1) > limit) * -1;
+  mask |= (abs(q3 - q2) > limit) * -1;
+  mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t flat_mask4(uint8_t thresh, uint8_t p3, uint8_t p2,
+                                uint8_t p1, uint8_t p0, uint8_t q0, uint8_t q1,
+                                uint8_t q2, uint8_t q3) {
+  int8_t mask = 0;
+  mask |= (abs(p1 - p0) > thresh) * -1;
+  mask |= (abs(q1 - q0) > thresh) * -1;
+  mask |= (abs(p2 - p0) > thresh) * -1;
+  mask |= (abs(q2 - q0) > thresh) * -1;
+  mask |= (abs(p3 - p0) > thresh) * -1;
+  mask |= (abs(q3 - q0) > thresh) * -1;
+  return ~mask;
+}
+
+#if PARALLEL_DEBLOCKING_9_TAP
+static INLINE int8_t flat_mask2(uint8_t thresh, uint8_t p4, uint8_t p0,
+                                uint8_t q0, uint8_t q4) {
+  int8_t mask = 0;
+  mask |= (abs(p4 - p0) > thresh) * -1;
+  mask |= (abs(q4 - q0) > thresh) * -1;
+  return ~mask;
+}
+#endif
+
+#if PARALLEL_DEBLOCKING_11_TAP
+static INLINE int8_t flat_mask3(uint8_t thresh, uint8_t p5, uint8_t p4,
+                                uint8_t p0, uint8_t q0, uint8_t q4,
+                                uint8_t q5) {
+  int8_t mask = 0;
+  mask |= (abs(p4 - p0) > thresh) * -1;
+  mask |= (abs(q4 - q0) > thresh) * -1;
+  mask |= (abs(p5 - p0) > thresh) * -1;
+  mask |= (abs(q5 - q0) > thresh) * -1;
+  return ~mask;
+}
+#endif
+
+static INLINE int8_t flat_mask5(uint8_t thresh, uint8_t p4, uint8_t p3,
+                                uint8_t p2, uint8_t p1, uint8_t p0, uint8_t q0,
+                                uint8_t q1, uint8_t q2, uint8_t q3,
+                                uint8_t q4) {
+  int8_t mask = ~flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3);
+  mask |= (abs(p4 - p0) > thresh) * -1;
+  mask |= (abs(q4 - q0) > thresh) * -1;
+  return ~mask;
+}
+
+// is there high edge variance internal edge: 11111111 yes, 00000000 no
+static INLINE int8_t hev_mask(uint8_t thresh, uint8_t p1, uint8_t p0,
+                              uint8_t q0, uint8_t q1) {
+  int8_t hev = 0;
+  hev |= (abs(p1 - p0) > thresh) * -1;
+  hev |= (abs(q1 - q0) > thresh) * -1;
+  return hev;
+}
+
+static INLINE void filter4(int8_t mask, uint8_t thresh, uint8_t *op1,
+                           uint8_t *op0, uint8_t *oq0, uint8_t *oq1) {
+  int8_t filter1, filter2;
+
+  const int8_t ps1 = (int8_t)*op1 ^ 0x80;
+  const int8_t ps0 = (int8_t)*op0 ^ 0x80;
+  const int8_t qs0 = (int8_t)*oq0 ^ 0x80;
+  const int8_t qs1 = (int8_t)*oq1 ^ 0x80;
+  const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1);
+
+  // add outer taps if we have high edge variance
+  int8_t filter = signed_char_clamp(ps1 - qs1) & hev;
+
+  // inner taps
+  filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask;
+
+  // save bottom 3 bits so that we round one side +4 and the other +3
+  // if it equals 4 we'll set to adjust by -1 to account for the fact
+  // we'd round 3 the other way
+  filter1 = signed_char_clamp(filter + 4) >> 3;
+  filter2 = signed_char_clamp(filter + 3) >> 3;
+
+  *oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80;
+  *op0 = signed_char_clamp(ps0 + filter2) ^ 0x80;
+
+  // outer tap adjustments
+  filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+
+  *oq1 = signed_char_clamp(qs1 - filter) ^ 0x80;
+  *op1 = signed_char_clamp(ps1 + filter) ^ 0x80;
+}
+
+void aom_lpf_horizontal_4_c(uint8_t *s, int p /* pitch */,
+                            const uint8_t *blimit, const uint8_t *limit,
+                            const uint8_t *thresh) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+#if !CONFIG_PARALLEL_DEBLOCKING
+    const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+    const int8_t mask =
+        filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3);
+#else   // CONFIG_PARALLEL_DEBLOCKING
+    const uint8_t p1 = s[-2 * p], p0 = s[-p];
+    const uint8_t q0 = s[0 * p], q1 = s[1 * p];
+    const int8_t mask = filter_mask2(*limit, *blimit, p1, p0, q0, q1);
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+    filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p);
+    ++s;
+  }
+}
+
+void aom_lpf_horizontal_4_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
+                                 const uint8_t *limit0, const uint8_t *thresh0,
+                                 const uint8_t *blimit1, const uint8_t *limit1,
+                                 const uint8_t *thresh1) {
+  aom_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0);
+  aom_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit,
+                          const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+#if !CONFIG_PARALLEL_DEBLOCKING
+    const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+    const int8_t mask =
+        filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3);
+#else   // CONFIG_PARALLEL_DEBLOCKING
+    const uint8_t p1 = s[-2], p0 = s[-1];
+    const uint8_t q0 = s[0], q1 = s[1];
+    const int8_t mask = filter_mask2(*limit, *blimit, p1, p0, q0, q1);
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+    filter4(mask, *thresh, s - 2, s - 1, s, s + 1);
+    s += pitch;
+  }
+}
+
+void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
+                               const uint8_t *limit0, const uint8_t *thresh0,
+                               const uint8_t *blimit1, const uint8_t *limit1,
+                               const uint8_t *thresh1) {
+  aom_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0);
+  aom_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1, thresh1);
+}
+
+static INLINE void filter8(int8_t mask, uint8_t thresh, uint8_t flat,
+                           uint8_t *op3, uint8_t *op2, uint8_t *op1,
+                           uint8_t *op0, uint8_t *oq0, uint8_t *oq1,
+                           uint8_t *oq2, uint8_t *oq3) {
+  if (flat && mask) {
+    const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+    const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+
+    // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
+    *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
+    *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
+    *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
+    *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
+    *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
+    *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
+  } else {
+    filter4(mask, thresh, op1, op0, oq0, oq1);
+  }
+}
+
+void aom_lpf_horizontal_8_c(uint8_t *s, int p, const uint8_t *blimit,
+                            const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+    const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+
+    const int8_t mask =
+        filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+    filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, s,
+            s + 1 * p, s + 2 * p, s + 3 * p);
+    ++s;
+  }
+}
+
+void aom_lpf_horizontal_8_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
+                                 const uint8_t *limit0, const uint8_t *thresh0,
+                                 const uint8_t *blimit1, const uint8_t *limit1,
+                                 const uint8_t *thresh1) {
+  aom_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0);
+  aom_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit,
+                          const uint8_t *limit, const uint8_t *thresh) {
+  int i;
+
+  for (i = 0; i < 8; ++i) {
+    const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+    const int8_t mask =
+        filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+    filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1, s, s + 1, s + 2,
+            s + 3);
+    s += pitch;
+  }
+}
+
+void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
+                               const uint8_t *limit0, const uint8_t *thresh0,
+                               const uint8_t *blimit1, const uint8_t *limit1,
+                               const uint8_t *thresh1) {
+  aom_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0);
+  aom_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1, thresh1);
+}
+
+#if PARALLEL_DEBLOCKING_11_TAP
+static INLINE void filter12(int8_t mask, uint8_t thresh, uint8_t flat,
+                            uint8_t flat2, uint8_t *op5, uint8_t *op4,
+                            uint8_t *op3, uint8_t *op2, uint8_t *op1,
+                            uint8_t *op0, uint8_t *oq0, uint8_t *oq1,
+                            uint8_t *oq2, uint8_t *oq3, uint8_t *oq4,
+                            uint8_t *oq5) {
+  if (flat2 && flat && mask) {
+    const uint8_t p5 = *op5, p4 = *op4, p3 = *op3, p2 = *op2, p1 = *op1,
+                  p0 = *op0;
+    const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3, q4 = *oq4,
+                  q5 = *oq5;
+
+    // 11-tap filter [1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1]
+    *op4 = (p5 * 5 + p4 * 2 + p3 + p2 + p1 + p0 + q0 + 6) / 12;
+    *op3 = (p5 * 4 + p4 + p3 * 2 + p2 + p1 + p0 + q0 + q1 + 6) / 12;
+    *op2 = (p5 * 3 + p4 + p3 + p2 * 2 + p1 + p0 + q0 + q1 + q2 + 6) / 12;
+    *op1 = (p5 * 2 + p4 + p3 + p2 + p1 * 2 + p0 + q0 + q1 + q2 + q3 + 6) / 12;
+    *op0 = (p5 + p4 + p3 + p2 + p1 + p0 * 2 + q0 + q1 + q2 + q3 + q4 + 6) / 12;
+    *oq0 = (p4 + p3 + p2 + p1 + p0 + q0 * 2 + q1 + q2 + q3 + q4 + q5 + 6) / 12;
+    *oq1 = (p3 + p2 + p1 + p0 + q0 + q1 * 2 + q2 + q3 + q4 + q5 * 2 + 6) / 12;
+    *oq2 = (p2 + p1 + p0 + q0 + q1 + q2 * 2 + q3 + q4 + q5 * 3 + 6) / 12;
+    *oq3 = (p1 + p0 + q0 + q1 + q2 + q3 * 2 + q4 + q5 * 4 + 6) / 12;
+    *oq4 = (p0 + q0 + q1 + q2 + q3 + q4 * 2 + q5 * 5 + 6) / 12;
+  } else {
+    filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
+  }
+}
+#endif
+
+#if PARALLEL_DEBLOCKING_9_TAP
+static INLINE void filter10(int8_t mask, uint8_t thresh, uint8_t flat,
+                            uint8_t flat2, uint8_t *op4, uint8_t *op3,
+                            uint8_t *op2, uint8_t *op1, uint8_t *op0,
+                            uint8_t *oq0, uint8_t *oq1, uint8_t *oq2,
+                            uint8_t *oq3, uint8_t *oq4) {
+  if (flat2 && flat && mask) {
+    const uint8_t p4 = *op4, p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+    const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3, q4 = *oq4;
+
+    // 9-tap filter [1, 1, 1, 1, 2, 1, 1, 1, 1]
+    *op3 = (p4 * 4 + p3 * 2 + p2 + p1 + p0 + q0 + 5) / 10;
+    *op2 = (p4 * 3 + p3 + p2 * 2 + p1 + p0 + q0 + q1 + 5) / 10;
+    *op1 = (p4 * 2 + p3 + p2 + p1 * 2 + p0 + q0 + q1 + q2 + 5) / 10;
+    *op0 = (p4 + p3 + p2 + p1 + p0 * 2 + q0 + q1 + q2 + q3 + 5) / 10;
+    *oq0 = (p3 + p2 + p1 + p0 + q0 * 2 + q1 + q2 + q3 + q4 + 5) / 10;
+    *oq1 = (p2 + p1 + p0 + q0 + q1 * 2 + q2 + q3 + q4 * 2 + 5) / 10;
+    *oq2 = (p1 + p0 + q0 + q1 + q2 * 2 + q3 + q4 * 3 + 5) / 10;
+    *oq3 = (p0 + q0 + q1 + q2 + q3 * 2 + q4 * 4 + 5) / 10;
+  } else {
+    filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
+  }
+}
+#endif
+
+static INLINE void filter16(int8_t mask, uint8_t thresh, uint8_t flat,
+                            uint8_t flat2, uint8_t *op7, uint8_t *op6,
+                            uint8_t *op5, uint8_t *op4, uint8_t *op3,
+                            uint8_t *op2, uint8_t *op1, uint8_t *op0,
+                            uint8_t *oq0, uint8_t *oq1, uint8_t *oq2,
+                            uint8_t *oq3, uint8_t *oq4, uint8_t *oq5,
+                            uint8_t *oq6, uint8_t *oq7) {
+  if (flat2 && flat && mask) {
+    const uint8_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4, p3 = *op3,
+                  p2 = *op2, p1 = *op1, p0 = *op0;
+
+    const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3, q4 = *oq4,
+                  q5 = *oq5, q6 = *oq6, q7 = *oq7;
+
+    // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
+    *op6 = ROUND_POWER_OF_TWO(
+        p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 + q0, 4);
+    *op5 = ROUND_POWER_OF_TWO(
+        p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 + q0 + q1, 4);
+    *op4 = ROUND_POWER_OF_TWO(
+        p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 + q0 + q1 + q2, 4);
+    *op3 = ROUND_POWER_OF_TWO(
+        p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 + q0 + q1 + q2 + q3, 4);
+    *op2 = ROUND_POWER_OF_TWO(
+        p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 + q0 + q1 + q2 + q3 + q4,
+        4);
+    *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
+                                  q0 + q1 + q2 + q3 + q4 + q5,
+                              4);
+    *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 + q0 +
+                                  q1 + q2 + q3 + q4 + q5 + q6,
+                              4);
+    *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 + q0 * 2 + q1 +
+                                  q2 + q3 + q4 + q5 + q6 + q7,
+                              4);
+    *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 + q0 + q1 * 2 + q2 +
+                                  q3 + q4 + q5 + q6 + q7 * 2,
+                              4);
+    *oq2 = ROUND_POWER_OF_TWO(
+        p4 + p3 + p2 + p1 + p0 + q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3,
+        4);
+    *oq3 = ROUND_POWER_OF_TWO(
+        p3 + p2 + p1 + p0 + q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
+    *oq4 = ROUND_POWER_OF_TWO(
+        p2 + p1 + p0 + q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
+    *oq5 = ROUND_POWER_OF_TWO(
+        p1 + p0 + q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
+    *oq6 = ROUND_POWER_OF_TWO(
+        p0 + q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
+  } else {
+    filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
+  }
+}
+
+static void mb_lpf_horizontal_edge_w(uint8_t *s, int p, const uint8_t *blimit,
+                                     const uint8_t *limit,
+                                     const uint8_t *thresh, int count) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint8_t p7 = s[-8 * p], p6 = s[-7 * p], p5 = s[-6 * p],
+                  p4 = s[-5 * p], p3 = s[-4 * p], p2 = s[-3 * p],
+                  p1 = s[-2 * p], p0 = s[-p];
+    const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p],
+                  q4 = s[4 * p], q5 = s[5 * p], q6 = s[6 * p], q7 = s[7 * p];
+    const int8_t mask =
+        filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+
+#if PARALLEL_DEBLOCKING_11_TAP
+    const int8_t flat2 = flat_mask3(1, p5, p4, p0, q0, q4, q5);
+
+    filter12(mask, *thresh, flat, flat2, s - 6 * p, s - 5 * p, s - 4 * p,
+             s - 3 * p, s - 2 * p, s - 1 * p, s, s + 1 * p, s + 2 * p,
+             s + 3 * p, s + 4 * p, s + 5 * p);
+
+#elif PARALLEL_DEBLOCKING_9_TAP
+    const int8_t flat2 = flat_mask2(1, p4, p0, q0, q4);
+
+    filter10(mask, *thresh, flat, flat2, s - 5 * p, s - 4 * p, s - 3 * p,
+             s - 2 * p, s - 1 * p, s, s + 1 * p, s + 2 * p, s + 3 * p,
+             s + 4 * p);
+#else
+    const int8_t flat2 = flat_mask5(1, p7, p6, p5, p4, p0, q0, q4, q5, q6, q7);
+
+    filter16(mask, *thresh, flat, flat2, s - 8 * p, s - 7 * p, s - 6 * p,
+             s - 5 * p, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, s,
+             s + 1 * p, s + 2 * p, s + 3 * p, s + 4 * p, s + 5 * p, s + 6 * p,
+             s + 7 * p);
+#endif
+
+    ++s;
+  }
+}
+
+void aom_lpf_horizontal_edge_8_c(uint8_t *s, int p, const uint8_t *blimit,
+                                 const uint8_t *limit, const uint8_t *thresh) {
+  mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 1);
+}
+
+void aom_lpf_horizontal_edge_16_c(uint8_t *s, int p, const uint8_t *blimit,
+                                  const uint8_t *limit, const uint8_t *thresh) {
+  mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 2);
+}
+
+static void mb_lpf_vertical_edge_w(uint8_t *s, int p, const uint8_t *blimit,
+                                   const uint8_t *limit, const uint8_t *thresh,
+                                   int count) {
+  int i;
+
+  for (i = 0; i < count; ++i) {
+    const uint8_t p7 = s[-8], p6 = s[-7], p5 = s[-6], p4 = s[-5], p3 = s[-4],
+                  p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3], q4 = s[4],
+                  q5 = s[5], q6 = s[6], q7 = s[7];
+    const int8_t mask =
+        filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3);
+    const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+
+#if PARALLEL_DEBLOCKING_11_TAP
+    const int8_t flat2 = flat_mask3(1, p5, p4, p0, q0, q4, q5);
+
+    filter12(mask, *thresh, flat, flat2, s - 6, s - 5, s - 4, s - 3, s - 2,
+             s - 1, s, s + 1, s + 2, s + 3, s + 4, s + 5);
+#elif PARALLEL_DEBLOCKING_9_TAP
+    const int8_t flat2 = flat_mask2(1, p4, p0, q0, q4);
+
+    filter10(mask, *thresh, flat, flat2, s - 5, s - 4, s - 3, s - 2, s - 1, s,
+             s + 1, s + 2, s + 3, s + 4);
+
+#else
+    const int8_t flat2 = flat_mask5(1, p7, p6, p5, p4, p0, q0, q4, q5, q6, q7);
+
+    filter16(mask, *thresh, flat, flat2, s - 8, s - 7, s - 6, s - 5, s - 4,
+             s - 3, s - 2, s - 1, s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6,
+             s + 7);
+#endif
+
+    s += p;
+  }
+}
+
+void aom_lpf_vertical_16_c(uint8_t *s, int p, const uint8_t *blimit,
+                           const uint8_t *limit, const uint8_t *thresh) {
+  mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8);
+}
+
+void aom_lpf_vertical_16_dual_c(uint8_t *s, int p, const uint8_t *blimit,
+                                const uint8_t *limit, const uint8_t *thresh) {
+  mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16);
+}
+
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_PARALLEL_DEBLOCKING
+// Should we apply any filter at all: 11111111 yes, 00000000 no ?
+static INLINE int8_t highbd_filter_mask2(uint8_t limit, uint8_t blimit,
+                                         uint16_t p1, uint16_t p0, uint16_t q0,
+                                         uint16_t q1, int bd) {
+  int8_t mask = 0;
+  int16_t limit16 = (uint16_t)limit << (bd - 8);
+  int16_t blimit16 = (uint16_t)blimit << (bd - 8);
+  mask |= (abs(p1 - p0) > limit16) * -1;
+  mask |= (abs(q1 - q0) > limit16) * -1;
+  mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1;
+  return ~mask;
+}
+#endif  // CONFIG_PARALLEL_DEBLOCKING
+
+// Should we apply any filter at all: 11111111 yes, 00000000 no ?
+static INLINE int8_t highbd_filter_mask(uint8_t limit, uint8_t blimit,
+                                        uint16_t p3, uint16_t p2, uint16_t p1,
+                                        uint16_t p0, uint16_t q0, uint16_t q1,
+                                        uint16_t q2, uint16_t q3, int bd) {
+  int8_t mask = 0;
+  int16_t limit16 = (uint16_t)limit << (bd - 8);
+  int16_t blimit16 = (uint16_t)blimit << (bd - 8);
+  mask |= (abs(p3 - p2) > limit16) * -1;
+  mask |= (abs(p2 - p1) > limit16) * -1;
+  mask |= (abs(p1 - p0) > limit16) * -1;
+  mask |= (abs(q1 - q0) > limit16) * -1;
+  mask |= (abs(q2 - q1) > limit16) * -1;
+  mask |= (abs(q3 - q2) > limit16) * -1;
+  mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t highbd_flat_mask4(uint8_t thresh, uint16_t p3, uint16_t p2,
+                                       uint16_t p1, uint16_t p0, uint16_t q0,
+                                       uint16_t q1, uint16_t q2, uint16_t q3,
+                                       int bd) {
+  int8_t mask = 0;
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  mask |= (abs(p1 - p0) > thresh16) * -1;
+  mask |= (abs(q1 - q0) > thresh16) * -1;
+  mask |= (abs(p2 - p0) > thresh16) * -1;
+  mask |= (abs(q2 - q0) > thresh16) * -1;
+  mask |= (abs(p3 - p0) > thresh16) * -1;
+  mask |= (abs(q3 - q0) > thresh16) * -1;
+  return ~mask;
+}
+
+static INLINE int8_t highbd_flat_mask5(uint8_t thresh, uint16_t p4, uint16_t p3,
+                                       uint16_t p2, uint16_t p1, uint16_t p0,
+                                       uint16_t q0, uint16_t q1, uint16_t q2,
+                                       uint16_t q3, uint16_t q4, int bd) {
+  int8_t mask = ~highbd_flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  mask |= (abs(p4 - p0) > thresh16) * -1;
+  mask |= (abs(q4 - q0) > thresh16) * -1;
+  return ~mask;
+}
+
+// Is there high edge variance internal edge:
+// 11111111_11111111 yes, 00000000_00000000 no ?
+static INLINE int16_t highbd_hev_mask(uint8_t thresh, uint16_t p1, uint16_t p0,
+                                      uint16_t q0, uint16_t q1, int bd) {
+  int16_t hev = 0;
+  int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+  hev |= (abs(p1 - p0) > thresh16) * -1;
+  hev |= (abs(q1 - q0) > thresh16) * -1;
+  return hev;
+}
+
+static INLINE void highbd_filter4(int8_t mask, uint8_t thresh, uint16_t *op1,
+                                  uint16_t *op0, uint16_t *oq0, uint16_t *oq1,
+                                  int bd) {
+  int16_t filter1, filter2;
+  // ^0x80 equivalent to subtracting 0x80 from the values to turn them
+  // into -128 to +127 instead of 0 to 255.
+  int shift = bd - 8;
+  const int16_t ps1 = (int16_t)*op1 - (0x80 << shift);
+  const int16_t ps0 = (int16_t)*op0 - (0x80 << shift);
+  const int16_t qs0 = (int16_t)*oq0 - (0x80 << shift);
+  const int16_t qs1 = (int16_t)*oq1 - (0x80 << shift);
+  const uint16_t hev = highbd_hev_mask(thresh, *op1, *op0, *oq0, *oq1, bd);
+
+  // Add outer taps if we have high edge variance.
+  int16_t filter = signed_char_clamp_high(ps1 - qs1, bd) & hev;
+
+  // Inner taps.
+  filter = signed_char_clamp_high(filter + 3 * (qs0 - ps0), bd) & mask;
+
+  // Save bottom 3 bits so that we round one side +4 and the other +3
+  // if it equals 4 we'll set to adjust by -1 to account for the fact
+  // we'd round 3 the other way.
+  filter1 = signed_char_clamp_high(filter + 4, bd) >> 3;
+  filter2 = signed_char_clamp_high(filter + 3, bd) >> 3;
+
+  *oq0 = signed_char_clamp_high(qs0 - filter1, bd) + (0x80 << shift);
+  *op0 = signed_char_clamp_high(ps0 + filter2, bd) + (0x80 << shift);
+
+  // Outer tap adjustments.
+  filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+
+  *oq1 = signed_char_clamp_high(qs1 - filter, bd) + (0x80 << shift);
+  *op1 = signed_char_clamp_high(ps1 + filter, bd) + (0x80 << shift);
+}
+
+void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int p /* pitch */,
+                                   const uint8_t *blimit, const uint8_t *limit,
+                                   const uint8_t *thresh, int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+#if !CONFIG_PARALLEL_DEBLOCKING
+    const uint16_t p3 = s[-4 * p];
+    const uint16_t p2 = s[-3 * p];
+    const uint16_t p1 = s[-2 * p];
+    const uint16_t p0 = s[-p];
+    const uint16_t q0 = s[0 * p];
+    const uint16_t q1 = s[1 * p];
+    const uint16_t q2 = s[2 * p];
+    const uint16_t q3 = s[3 * p];
+    const int8_t mask =
+        highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+#else   // CONFIG_PARALLEL_DEBLOCKING
+    const uint16_t p1 = s[-2 * p];
+    const uint16_t p0 = s[-p];
+    const uint16_t q0 = s[0 * p];
+    const uint16_t q1 = s[1 * p];
+    const int8_t mask =
+        highbd_filter_mask2(*limit, *blimit, p1, p0, q0, q1, bd);
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+    highbd_filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p, bd);
+    ++s;
+  }
+}
+
+void aom_highbd_lpf_horizontal_4_dual_c(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  aom_highbd_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, bd);
+  aom_highbd_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, bd);
+}
+
+void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit,
+                                 const uint8_t *limit, const uint8_t *thresh,
+                                 int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+#if !CONFIG_PARALLEL_DEBLOCKING
+    const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+    const int8_t mask =
+        highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+#else   // CONFIG_PARALLEL_DEBLOCKING
+    const uint16_t p1 = s[-2], p0 = s[-1];
+    const uint16_t q0 = s[0], q1 = s[1];
+    const int8_t mask =
+        highbd_filter_mask2(*limit, *blimit, p1, p0, q0, q1, bd);
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+    highbd_filter4(mask, *thresh, s - 2, s - 1, s, s + 1, bd);
+    s += pitch;
+  }
+}
+
+void aom_highbd_lpf_vertical_4_dual_c(
+    uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  aom_highbd_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, bd);
+  aom_highbd_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1, thresh1,
+                              bd);
+}
+
+static INLINE void highbd_filter8(int8_t mask, uint8_t thresh, uint8_t flat,
+                                  uint16_t *op3, uint16_t *op2, uint16_t *op1,
+                                  uint16_t *op0, uint16_t *oq0, uint16_t *oq1,
+                                  uint16_t *oq2, uint16_t *oq3, int bd) {
+  if (flat && mask) {
+    const uint16_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+    const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+
+    // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
+    *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
+    *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
+    *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
+    *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
+    *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
+    *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
+  } else {
+    highbd_filter4(mask, thresh, op1, op0, oq0, oq1, bd);
+  }
+}
+
+void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int p, const uint8_t *blimit,
+                                   const uint8_t *limit, const uint8_t *thresh,
+                                   int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8; ++i) {
+    const uint16_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+    const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+
+    const int8_t mask =
+        highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat =
+        highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    highbd_filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p,
+                   s - 1 * p, s, s + 1 * p, s + 2 * p, s + 3 * p, bd);
+    ++s;
+  }
+}
+
+void aom_highbd_lpf_horizontal_8_dual_c(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  aom_highbd_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, bd);
+  aom_highbd_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, bd);
+}
+
+void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit,
+                                 const uint8_t *limit, const uint8_t *thresh,
+                                 int bd) {
+  int i;
+
+  for (i = 0; i < 8; ++i) {
+    const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+    const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+    const int8_t mask =
+        highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat =
+        highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    highbd_filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1, s, s + 1,
+                   s + 2, s + 3, bd);
+    s += pitch;
+  }
+}
+
+void aom_highbd_lpf_vertical_8_dual_c(
+    uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  aom_highbd_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, bd);
+  aom_highbd_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1, thresh1,
+                              bd);
+}
+
+static INLINE void highbd_filter16(int8_t mask, uint8_t thresh, uint8_t flat,
+                                   uint8_t flat2, uint16_t *op7, uint16_t *op6,
+                                   uint16_t *op5, uint16_t *op4, uint16_t *op3,
+                                   uint16_t *op2, uint16_t *op1, uint16_t *op0,
+                                   uint16_t *oq0, uint16_t *oq1, uint16_t *oq2,
+                                   uint16_t *oq3, uint16_t *oq4, uint16_t *oq5,
+                                   uint16_t *oq6, uint16_t *oq7, int bd) {
+  if (flat2 && flat && mask) {
+    const uint16_t p7 = *op7;
+    const uint16_t p6 = *op6;
+    const uint16_t p5 = *op5;
+    const uint16_t p4 = *op4;
+    const uint16_t p3 = *op3;
+    const uint16_t p2 = *op2;
+    const uint16_t p1 = *op1;
+    const uint16_t p0 = *op0;
+    const uint16_t q0 = *oq0;
+    const uint16_t q1 = *oq1;
+    const uint16_t q2 = *oq2;
+    const uint16_t q3 = *oq3;
+    const uint16_t q4 = *oq4;
+    const uint16_t q5 = *oq5;
+    const uint16_t q6 = *oq6;
+    const uint16_t q7 = *oq7;
+
+    // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
+    *op6 = ROUND_POWER_OF_TWO(
+        p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 + q0, 4);
+    *op5 = ROUND_POWER_OF_TWO(
+        p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 + q0 + q1, 4);
+    *op4 = ROUND_POWER_OF_TWO(
+        p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 + q0 + q1 + q2, 4);
+    *op3 = ROUND_POWER_OF_TWO(
+        p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 + q0 + q1 + q2 + q3, 4);
+    *op2 = ROUND_POWER_OF_TWO(
+        p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 + q0 + q1 + q2 + q3 + q4,
+        4);
+    *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
+                                  q0 + q1 + q2 + q3 + q4 + q5,
+                              4);
+    *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 + q0 +
+                                  q1 + q2 + q3 + q4 + q5 + q6,
+                              4);
+    *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 + q0 * 2 + q1 +
+                                  q2 + q3 + q4 + q5 + q6 + q7,
+                              4);
+    *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 + q0 + q1 * 2 + q2 +
+                                  q3 + q4 + q5 + q6 + q7 * 2,
+                              4);
+    *oq2 = ROUND_POWER_OF_TWO(
+        p4 + p3 + p2 + p1 + p0 + q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3,
+        4);
+    *oq3 = ROUND_POWER_OF_TWO(
+        p3 + p2 + p1 + p0 + q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
+    *oq4 = ROUND_POWER_OF_TWO(
+        p2 + p1 + p0 + q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
+    *oq5 = ROUND_POWER_OF_TWO(
+        p1 + p0 + q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
+    *oq6 = ROUND_POWER_OF_TWO(
+        p0 + q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
+  } else {
+    highbd_filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3,
+                   bd);
+  }
+}
+
+static void highbd_mb_lpf_horizontal_edge_w(uint16_t *s, int p,
+                                            const uint8_t *blimit,
+                                            const uint8_t *limit,
+                                            const uint8_t *thresh, int count,
+                                            int bd) {
+  int i;
+
+  // loop filter designed to work using chars so that we can make maximum use
+  // of 8 bit simd instructions.
+  for (i = 0; i < 8 * count; ++i) {
+    const uint16_t p3 = s[-4 * p];
+    const uint16_t p2 = s[-3 * p];
+    const uint16_t p1 = s[-2 * p];
+    const uint16_t p0 = s[-p];
+    const uint16_t q0 = s[0 * p];
+    const uint16_t q1 = s[1 * p];
+    const uint16_t q2 = s[2 * p];
+    const uint16_t q3 = s[3 * p];
+    const int8_t mask =
+        highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat =
+        highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat2 =
+        highbd_flat_mask5(1, s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0, q0,
+                          s[4 * p], s[5 * p], s[6 * p], s[7 * p], bd);
+
+    highbd_filter16(mask, *thresh, flat, flat2, s - 8 * p, s - 7 * p, s - 6 * p,
+                    s - 5 * p, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, s,
+                    s + 1 * p, s + 2 * p, s + 3 * p, s + 4 * p, s + 5 * p,
+                    s + 6 * p, s + 7 * p, bd);
+    ++s;
+  }
+}
+
+void aom_highbd_lpf_horizontal_edge_8_c(uint16_t *s, int p,
+                                        const uint8_t *blimit,
+                                        const uint8_t *limit,
+                                        const uint8_t *thresh, int bd) {
+  highbd_mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 1, bd);
+}
+
+void aom_highbd_lpf_horizontal_edge_16_c(uint16_t *s, int p,
+                                         const uint8_t *blimit,
+                                         const uint8_t *limit,
+                                         const uint8_t *thresh, int bd) {
+  highbd_mb_lpf_horizontal_edge_w(s, p, blimit, limit, thresh, 2, bd);
+}
+
+static void highbd_mb_lpf_vertical_edge_w(uint16_t *s, int p,
+                                          const uint8_t *blimit,
+                                          const uint8_t *limit,
+                                          const uint8_t *thresh, int count,
+                                          int bd) {
+  int i;
+
+  for (i = 0; i < count; ++i) {
+    const uint16_t p3 = s[-4];
+    const uint16_t p2 = s[-3];
+    const uint16_t p1 = s[-2];
+    const uint16_t p0 = s[-1];
+    const uint16_t q0 = s[0];
+    const uint16_t q1 = s[1];
+    const uint16_t q2 = s[2];
+    const uint16_t q3 = s[3];
+    const int8_t mask =
+        highbd_filter_mask(*limit, *blimit, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat =
+        highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+    const int8_t flat2 = highbd_flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
+                                           q0, s[4], s[5], s[6], s[7], bd);
+
+    highbd_filter16(mask, *thresh, flat, flat2, s - 8, s - 7, s - 6, s - 5,
+                    s - 4, s - 3, s - 2, s - 1, s, s + 1, s + 2, s + 3, s + 4,
+                    s + 5, s + 6, s + 7, bd);
+    s += p;
+  }
+}
+
+void aom_highbd_lpf_vertical_16_c(uint16_t *s, int p, const uint8_t *blimit,
+                                  const uint8_t *limit, const uint8_t *thresh,
+                                  int bd) {
+  highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8, bd);
+}
+
+void aom_highbd_lpf_vertical_16_dual_c(uint16_t *s, int p,
+                                       const uint8_t *blimit,
+                                       const uint8_t *limit,
+                                       const uint8_t *thresh, int bd) {
+  highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16, bd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/mips/add_noise_msa.c b/third_party/aom/aom_dsp/mips/add_noise_msa.c
new file mode 100644
index 0000000000..4c6e201e11
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/add_noise_msa.c
@@ -0,0 +1,60 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include "./macros_msa.h"
+
+void aom_plane_add_noise_msa(uint8_t *start_ptr, char *noise,
+                             char blackclamp[16], char whiteclamp[16],
+                             char bothclamp[16], uint32_t width,
+                             uint32_t height, int32_t pitch) {
+  uint32_t i, j;
+
+  for (i = 0; i < height / 2; ++i) {
+    uint8_t *pos0_ptr = start_ptr + (2 * i) * pitch;
+    int8_t *ref0_ptr = (int8_t *)(noise + (rand() & 0xff));
+    uint8_t *pos1_ptr = start_ptr + (2 * i + 1) * pitch;
+    int8_t *ref1_ptr = (int8_t *)(noise + (rand() & 0xff));
+    for (j = width / 16; j--;) {
+      v16i8 temp00_s, temp01_s;
+      v16u8 temp00, temp01, black_clamp, white_clamp;
+      v16u8 pos0, ref0, pos1, ref1;
+      v16i8 const127 = __msa_ldi_b(127);
+
+      pos0 = LD_UB(pos0_ptr);
+      ref0 = LD_UB(ref0_ptr);
+      pos1 = LD_UB(pos1_ptr);
+      ref1 = LD_UB(ref1_ptr);
+      black_clamp = (v16u8)__msa_fill_b(blackclamp[0]);
+      white_clamp = (v16u8)__msa_fill_b(whiteclamp[0]);
+      temp00 = (pos0 < black_clamp);
+      pos0 = __msa_bmnz_v(pos0, black_clamp, temp00);
+      temp01 = (pos1 < black_clamp);
+      pos1 = __msa_bmnz_v(pos1, black_clamp, temp01);
+      XORI_B2_128_UB(pos0, pos1);
+      temp00_s = __msa_adds_s_b((v16i8)white_clamp, const127);
+      temp00 = (v16u8)(temp00_s < pos0);
+      pos0 = (v16u8)__msa_bmnz_v((v16u8)pos0, (v16u8)temp00_s, temp00);
+      temp01_s = __msa_adds_s_b((v16i8)white_clamp, const127);
+      temp01 = (temp01_s < pos1);
+      pos1 = (v16u8)__msa_bmnz_v((v16u8)pos1, (v16u8)temp01_s, temp01);
+      XORI_B2_128_UB(pos0, pos1);
+      pos0 += ref0;
+      ST_UB(pos0, pos0_ptr);
+      pos1 += ref1;
+      ST_UB(pos1, pos1_ptr);
+      pos0_ptr += 16;
+      pos1_ptr += 16;
+      ref0_ptr += 16;
+      ref1_ptr += 16;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/aom_convolve8_avg_horiz_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve8_avg_horiz_msa.c
new file mode 100644
index 0000000000..847394a3d2
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/aom_convolve8_avg_horiz_msa.c
@@ -0,0 +1,704 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/aom_convolve_msa.h"
+
+static void common_hz_8t_and_aver_dst_4x4_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 dst0, dst1, dst2, dst3, res2, res3;
+  v16u8 mask0, mask1, mask2, mask3;
+  v8i16 filt, res0, res1;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, res0, res1);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  SRARI_H2_SH(res0, res1, FILTER_BITS);
+  SAT_SH2_SH(res0, res1, 7);
+  PCKEV_B2_UB(res0, res0, res1, res1, res2, res3);
+  ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
+  XORI_B2_128_UB(res2, res3);
+  AVER_UB2_UB(res2, dst0, res3, dst2, res2, res3);
+  ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_8t_and_aver_dst_4x8_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, res0, res1, res2, res3;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v8i16 filt, vec0, vec1, vec2, vec3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  src += (4 * src_stride);
+  LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, vec0, vec1);
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, vec2, vec3);
+  SRARI_H4_SH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  SAT_SH4_SH(vec0, vec1, vec2, vec3, 7);
+  PCKEV_B4_UB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3, res0, res1, res2,
+              res3);
+  ILVR_D2_UB(res1, res0, res3, res2, res0, res2);
+  XORI_B2_128_UB(res0, res2);
+  ILVR_W4_UB(dst1, dst0, dst3, dst2, dst5, dst4, dst7, dst6, dst0, dst2, dst4,
+             dst6);
+  ILVR_D2_UB(dst2, dst0, dst6, dst4, dst0, dst4);
+  AVER_UB2_UB(res0, dst0, res2, dst4, res0, res2);
+  ST4x8_UB(res0, res2, dst, dst_stride);
+}
+
+static void common_hz_8t_and_aver_dst_4w_msa(const uint8_t *src,
+                                             int32_t src_stride, uint8_t *dst,
+                                             int32_t dst_stride, int8_t *filter,
+                                             int32_t height) {
+  if (4 == height) {
+    common_hz_8t_and_aver_dst_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_hz_8t_and_aver_dst_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_hz_8t_and_aver_dst_8w_msa(const uint8_t *src,
+                                             int32_t src_stride, uint8_t *dst,
+                                             int32_t dst_stride, int8_t *filter,
+                                             int32_t height) {
+  int32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, dst0, dst1, dst2, dst3;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    src += (4 * src_stride);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    CONVERT_UB_AVG_ST8x4_UB(out0, out1, out2, out3, dst0, dst1, dst2, dst3, dst,
+                            dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void common_hz_8t_and_aver_dst_16w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  int32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, dst0, dst1;
+  v8i16 filt, out0, out1, out2, out3;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8i16 vec8, vec9, vec10, vec11, vec12, vec13, vec14, vec15;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = height >> 1; loop_cnt--;) {
+    LD_SB2(src, src_stride, src0, src2);
+    LD_SB2(src + 8, src_stride, src1, src3);
+    src += (2 * src_stride);
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    VSHF_B4_SH(src0, src0, mask0, mask1, mask2, mask3, vec0, vec4, vec8, vec12);
+    VSHF_B4_SH(src1, src1, mask0, mask1, mask2, mask3, vec1, vec5, vec9, vec13);
+    VSHF_B4_SH(src2, src2, mask0, mask1, mask2, mask3, vec2, vec6, vec10,
+               vec14);
+    VSHF_B4_SH(src3, src3, mask0, mask1, mask2, mask3, vec3, vec7, vec11,
+               vec15);
+    DOTP_SB4_SH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    DOTP_SB4_SH(vec8, vec9, vec10, vec11, filt2, filt2, filt2, filt2, vec8,
+                vec9, vec10, vec11);
+    DPADD_SB4_SH(vec4, vec5, vec6, vec7, filt1, filt1, filt1, filt1, vec0, vec1,
+                 vec2, vec3);
+    DPADD_SB4_SH(vec12, vec13, vec14, vec15, filt3, filt3, filt3, filt3, vec8,
+                 vec9, vec10, vec11);
+    ADDS_SH4_SH(vec0, vec8, vec1, vec9, vec2, vec10, vec3, vec11, out0, out1,
+                out2, out3);
+    LD_UB2(dst, dst_stride, dst0, dst1);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    PCKEV_XORI128_AVG_ST_UB(out1, out0, dst0, dst);
+    dst += dst_stride;
+    PCKEV_XORI128_AVG_ST_UB(out3, out2, dst1, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_8t_and_aver_dst_32w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 dst1, dst2, mask0, mask1, mask2, mask3;
+  v8i16 filt, out0, out1, out2, out3;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8i16 vec8, vec9, vec10, vec11, vec12, vec13, vec14, vec15;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = height; loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    VSHF_B4_SH(src0, src0, mask0, mask1, mask2, mask3, vec0, vec4, vec8, vec12);
+    VSHF_B4_SH(src1, src1, mask0, mask1, mask2, mask3, vec1, vec5, vec9, vec13);
+    VSHF_B4_SH(src2, src2, mask0, mask1, mask2, mask3, vec2, vec6, vec10,
+               vec14);
+    VSHF_B4_SH(src3, src3, mask0, mask1, mask2, mask3, vec3, vec7, vec11,
+               vec15);
+    DOTP_SB4_SH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    DOTP_SB4_SH(vec8, vec9, vec10, vec11, filt2, filt2, filt2, filt2, vec8,
+                vec9, vec10, vec11);
+    DPADD_SB4_SH(vec4, vec5, vec6, vec7, filt1, filt1, filt1, filt1, vec0, vec1,
+                 vec2, vec3);
+    DPADD_SB4_SH(vec12, vec13, vec14, vec15, filt3, filt3, filt3, filt3, vec8,
+                 vec9, vec10, vec11);
+    ADDS_SH4_SH(vec0, vec8, vec1, vec9, vec2, vec10, vec3, vec11, out0, out1,
+                out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    LD_UB2(dst, 16, dst1, dst2);
+    PCKEV_XORI128_AVG_ST_UB(out1, out0, dst1, dst);
+    PCKEV_XORI128_AVG_ST_UB(out3, out2, dst2, dst + 16);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_8t_and_aver_dst_64w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  uint32_t loop_cnt, cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 dst1, dst2, mask0, mask1, mask2, mask3;
+  v8i16 filt, out0, out1, out2, out3;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8i16 vec8, vec9, vec10, vec11, vec12, vec13, vec14, vec15;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = height; loop_cnt--;) {
+    for (cnt = 0; cnt < 2; ++cnt) {
+      src0 = LD_SB(&src[cnt << 5]);
+      src2 = LD_SB(&src[16 + (cnt << 5)]);
+      src3 = LD_SB(&src[24 + (cnt << 5)]);
+      src1 = __msa_sldi_b(src2, src0, 8);
+
+      XORI_B4_128_SB(src0, src1, src2, src3);
+      VSHF_B4_SH(src0, src0, mask0, mask1, mask2, mask3, vec0, vec4, vec8,
+                 vec12);
+      VSHF_B4_SH(src1, src1, mask0, mask1, mask2, mask3, vec1, vec5, vec9,
+                 vec13);
+      VSHF_B4_SH(src2, src2, mask0, mask1, mask2, mask3, vec2, vec6, vec10,
+                 vec14);
+      VSHF_B4_SH(src3, src3, mask0, mask1, mask2, mask3, vec3, vec7, vec11,
+                 vec15);
+      DOTP_SB4_SH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0,
+                  vec1, vec2, vec3);
+      DOTP_SB4_SH(vec8, vec9, vec10, vec11, filt2, filt2, filt2, filt2, vec8,
+                  vec9, vec10, vec11);
+      DPADD_SB4_SH(vec4, vec5, vec6, vec7, filt1, filt1, filt1, filt1, vec0,
+                   vec1, vec2, vec3);
+      DPADD_SB4_SH(vec12, vec13, vec14, vec15, filt3, filt3, filt3, filt3, vec8,
+                   vec9, vec10, vec11);
+      ADDS_SH4_SH(vec0, vec8, vec1, vec9, vec2, vec10, vec3, vec11, out0, out1,
+                  out2, out3);
+      SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+      SAT_SH4_SH(out0, out1, out2, out3, 7);
+      LD_UB2(&dst[cnt << 5], 16, dst1, dst2);
+      PCKEV_XORI128_AVG_ST_UB(out1, out0, dst1, &dst[cnt << 5]);
+      PCKEV_XORI128_AVG_ST_UB(out3, out2, dst2, &dst[16 + (cnt << 5)]);
+    }
+
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_and_aver_dst_4x4_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3, vec0, vec1, res0, res1;
+  v8u16 vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
+  DOTP_UB2_UH(vec0, vec1, filt0, filt0, vec2, vec3);
+  SRARI_H2_UH(vec2, vec3, FILTER_BITS);
+  PCKEV_B2_UB(vec2, vec2, vec3, vec3, res0, res1);
+  ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
+  AVER_UB2_UB(res0, dst0, res1, dst2, res0, res1);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_2t_and_aver_dst_4x8_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, vec0, vec1, vec2, vec3, res0, res1, res2, res3;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v8u16 vec4, vec5, vec6, vec7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+  VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
+  VSHF_B2_UB(src4, src5, src6, src7, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec4, vec5,
+              vec6, vec7);
+  SRARI_H4_UH(vec4, vec5, vec6, vec7, FILTER_BITS);
+  PCKEV_B4_UB(vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7, res0, res1, res2,
+              res3);
+  ILVR_W4_UB(dst1, dst0, dst3, dst2, dst5, dst4, dst7, dst6, dst0, dst2, dst4,
+             dst6);
+  AVER_UB4_UB(res0, dst0, res1, dst2, res2, dst4, res3, dst6, res0, res1, res2,
+              res3);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+  dst += (4 * dst_stride);
+  ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_2t_and_aver_dst_4w_msa(const uint8_t *src,
+                                             int32_t src_stride, uint8_t *dst,
+                                             int32_t dst_stride, int8_t *filter,
+                                             int32_t height) {
+  if (4 == height) {
+    common_hz_2t_and_aver_dst_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_hz_2t_and_aver_dst_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_hz_2t_and_aver_dst_8x4_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3;
+  v8u16 vec0, vec1, vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  PCKEV_AVG_ST8x4_UB(vec0, dst0, vec1, dst1, vec2, dst2, vec3, dst3, dst,
+                     dst_stride);
+}
+
+static void common_hz_2t_and_aver_dst_8x8mult_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter, int32_t height) {
+  v16i8 src0, src1, src2, src3, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3;
+  v8u16 vec0, vec1, vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  src += (4 * src_stride);
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  src += (4 * src_stride);
+  PCKEV_AVG_ST8x4_UB(vec0, dst0, vec1, dst1, vec2, dst2, vec3, dst3, dst,
+                     dst_stride);
+  dst += (4 * dst_stride);
+
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  PCKEV_AVG_ST8x4_UB(vec0, dst0, vec1, dst1, vec2, dst2, vec3, dst3, dst,
+                     dst_stride);
+  dst += (4 * dst_stride);
+
+  if (16 == height) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    PCKEV_AVG_ST8x4_UB(vec0, dst0, vec1, dst1, vec2, dst2, vec3, dst3, dst,
+                       dst_stride);
+    dst += (4 * dst_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    PCKEV_AVG_ST8x4_UB(vec0, dst0, vec1, dst1, vec2, dst2, vec3, dst3, dst,
+                       dst_stride);
+  }
+}
+
+static void common_hz_2t_and_aver_dst_8w_msa(const uint8_t *src,
+                                             int32_t src_stride, uint8_t *dst,
+                                             int32_t dst_stride, int8_t *filter,
+                                             int32_t height) {
+  if (4 == height) {
+    common_hz_2t_and_aver_dst_8x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else {
+    common_hz_2t_and_aver_dst_8x8mult_msa(src, src_stride, dst, dst_stride,
+                                          filter, height);
+  }
+}
+
+static void common_hz_2t_and_aver_dst_16w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 res0, res1, res2, res3, res4, res5, res6, res7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src2, src4, src6);
+  LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+  src += (4 * src_stride);
+
+  VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+  VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, res0, res1,
+              res2, res3);
+  DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, res4, res5,
+              res6, res7);
+  SRARI_H4_UH(res0, res1, res2, res3, FILTER_BITS);
+  SRARI_H4_UH(res4, res5, res6, res7, FILTER_BITS);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  PCKEV_AVG_ST_UB(res1, res0, dst0, dst);
+  dst += dst_stride;
+  PCKEV_AVG_ST_UB(res3, res2, dst1, dst);
+  dst += dst_stride;
+  PCKEV_AVG_ST_UB(res5, res4, dst2, dst);
+  dst += dst_stride;
+  PCKEV_AVG_ST_UB(res7, res6, dst3, dst);
+  dst += dst_stride;
+
+  for (loop_cnt = (height >> 2) - 1; loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, res0, res1,
+                res2, res3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, res4, res5,
+                res6, res7);
+    SRARI_H4_UH(res0, res1, res2, res3, FILTER_BITS);
+    SRARI_H4_UH(res4, res5, res6, res7, FILTER_BITS);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    PCKEV_AVG_ST_UB(res1, res0, dst0, dst);
+    dst += dst_stride;
+    PCKEV_AVG_ST_UB(res3, res2, dst1, dst);
+    dst += dst_stride;
+    PCKEV_AVG_ST_UB(res5, res4, dst2, dst);
+    dst += dst_stride;
+    PCKEV_AVG_ST_UB(res7, res6, dst3, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_and_aver_dst_32w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 res0, res1, res2, res3, res4, res5, res6, res7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  for (loop_cnt = (height >> 1); loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+    src4 = LD_SB(src);
+    src6 = LD_SB(src + 16);
+    src7 = LD_SB(src + 24);
+    src5 = __msa_sldi_b(src6, src4, 8);
+    src += src_stride;
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, res0, res1,
+                res2, res3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, res4, res5,
+                res6, res7);
+    SRARI_H4_UH(res0, res1, res2, res3, FILTER_BITS);
+    SRARI_H4_UH(res4, res5, res6, res7, FILTER_BITS);
+    LD_UB2(dst, 16, dst0, dst1);
+    PCKEV_AVG_ST_UB(res1, res0, dst0, dst);
+    PCKEV_AVG_ST_UB(res3, res2, dst1, (dst + 16));
+    dst += dst_stride;
+    LD_UB2(dst, 16, dst2, dst3);
+    PCKEV_AVG_ST_UB(res5, res4, dst2, dst);
+    PCKEV_AVG_ST_UB(res7, res6, dst3, (dst + 16));
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_and_aver_dst_64w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, dst0, dst1, dst2, dst3;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  for (loop_cnt = height; loop_cnt--;) {
+    LD_SB4(src, 16, src0, src2, src4, src6);
+    src7 = LD_SB(src + 56);
+    SLDI_B3_SB(src2, src4, src6, src0, src2, src4, src1, src3, src5, 8);
+    src += src_stride;
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+                out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+                out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    LD_UB4(dst, 16, dst0, dst1, dst2, dst3);
+    PCKEV_AVG_ST_UB(out1, out0, dst0, dst);
+    PCKEV_AVG_ST_UB(out3, out2, dst1, dst + 16);
+    PCKEV_AVG_ST_UB(out5, out4, dst2, dst + 32);
+    PCKEV_AVG_ST_UB(out7, out6, dst3, dst + 48);
+    dst += dst_stride;
+  }
+}
+
+void aom_convolve8_avg_horiz_msa(const uint8_t *src, ptrdiff_t src_stride,
+                                 uint8_t *dst, ptrdiff_t dst_stride,
+                                 const int16_t *filter_x, int x_step_q4,
+                                 const int16_t *filter_y, int y_step_q4, int w,
+                                 int h) {
+  int8_t cnt, filt_hor[8];
+
+  assert(x_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+
+  for (cnt = 0; cnt < 8; ++cnt) {
+    filt_hor[cnt] = filter_x[cnt];
+  }
+
+  if (((const int32_t *)filter_x)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_hz_2t_and_aver_dst_4w_msa(src, (int32_t)src_stride, dst,
+                                         (int32_t)dst_stride, &filt_hor[3], h);
+        break;
+      case 8:
+        common_hz_2t_and_aver_dst_8w_msa(src, (int32_t)src_stride, dst,
+                                         (int32_t)dst_stride, &filt_hor[3], h);
+        break;
+      case 16:
+        common_hz_2t_and_aver_dst_16w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, &filt_hor[3], h);
+        break;
+      case 32:
+        common_hz_2t_and_aver_dst_32w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, &filt_hor[3], h);
+        break;
+      case 64:
+        common_hz_2t_and_aver_dst_64w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, &filt_hor[3], h);
+        break;
+      default:
+        aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                                  x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  } else {
+    switch (w) {
+      case 4:
+        common_hz_8t_and_aver_dst_4w_msa(src, (int32_t)src_stride, dst,
+                                         (int32_t)dst_stride, filt_hor, h);
+        break;
+      case 8:
+        common_hz_8t_and_aver_dst_8w_msa(src, (int32_t)src_stride, dst,
+                                         (int32_t)dst_stride, filt_hor, h);
+        break;
+      case 16:
+        common_hz_8t_and_aver_dst_16w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, filt_hor, h);
+        break;
+      case 32:
+        common_hz_8t_and_aver_dst_32w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, filt_hor, h);
+        break;
+      case 64:
+        common_hz_8t_and_aver_dst_64w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, filt_hor, h);
+        break;
+      default:
+        aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                                  x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/aom_convolve8_avg_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve8_avg_msa.c
new file mode 100644
index 0000000000..bed600d5b9
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/aom_convolve8_avg_msa.c
@@ -0,0 +1,605 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/aom_convolve_msa.h"
+
+static void common_hv_8ht_8vt_and_aver_dst_4w_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16u8 dst0, dst1, dst2, dst3, mask0, mask1, mask2, mask3, tmp0, tmp1;
+  v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3;
+  v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8i16 hz_out7, hz_out8, hz_out9, res0, res1, vec0, vec1, vec2, vec3, vec4;
+  v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= (3 + 3 * src_stride);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  hz_out0 = HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out2 = HORIZ_8TAP_FILT(src2, src3, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out4 = HORIZ_8TAP_FILT(src4, src5, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out5 = HORIZ_8TAP_FILT(src5, src6, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  SLDI_B2_SH(hz_out2, hz_out4, hz_out0, hz_out2, hz_out1, hz_out3, 8);
+
+  filt = LD_SH(filter_vert);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
+
+  ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+  vec2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    hz_out7 = HORIZ_8TAP_FILT(src7, src8, mask0, mask1, mask2, mask3, filt_hz0,
+                              filt_hz1, filt_hz2, filt_hz3);
+    hz_out6 = (v8i16)__msa_sldi_b((v16i8)hz_out7, (v16i8)hz_out5, 8);
+    vec3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+    res0 = FILT_8TAP_DPADD_S_H(vec0, vec1, vec2, vec3, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out9 = HORIZ_8TAP_FILT(src9, src10, mask0, mask1, mask2, mask3, filt_hz0,
+                              filt_hz1, filt_hz2, filt_hz3);
+    hz_out8 = (v8i16)__msa_sldi_b((v16i8)hz_out9, (v16i8)hz_out7, 8);
+    vec4 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8);
+    res1 = FILT_8TAP_DPADD_S_H(vec1, vec2, vec3, vec4, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+    ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
+
+    SRARI_H2_SH(res0, res1, FILTER_BITS);
+    SAT_SH2_SH(res0, res1, 7);
+    PCKEV_B2_UB(res0, res0, res1, res1, tmp0, tmp1);
+    XORI_B2_128_UB(tmp0, tmp1);
+    AVER_UB2_UB(tmp0, dst0, tmp1, dst2, tmp0, tmp1);
+    ST4x4_UB(tmp0, tmp1, 0, 1, 0, 1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    hz_out5 = hz_out9;
+    vec0 = vec2;
+    vec1 = vec3;
+    vec2 = vec4;
+  }
+}
+
+static void common_hv_8ht_8vt_and_aver_dst_8w_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3;
+  v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3;
+  v16u8 dst0, dst1, dst2, dst3, mask0, mask1, mask2, mask3;
+  v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8i16 hz_out7, hz_out8, hz_out9, hz_out10, tmp0, tmp1, tmp2, tmp3;
+  v8i16 out0, out1, out2, out3, out4, out5, out6, out7, out8, out9;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= (3 + 3 * src_stride);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  hz_out0 = HORIZ_8TAP_FILT(src0, src0, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out1 = HORIZ_8TAP_FILT(src1, src1, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out2 = HORIZ_8TAP_FILT(src2, src2, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out3 = HORIZ_8TAP_FILT(src3, src3, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out4 = HORIZ_8TAP_FILT(src4, src4, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out5 = HORIZ_8TAP_FILT(src5, src5, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out6 = HORIZ_8TAP_FILT(src6, src6, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+
+  filt = LD_SH(filter_vert);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
+
+  ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
+  ILVEV_B2_SH(hz_out4, hz_out5, hz_out1, hz_out2, out2, out4);
+  ILVEV_B2_SH(hz_out3, hz_out4, hz_out5, hz_out6, out5, out6);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+
+    hz_out7 = HORIZ_8TAP_FILT(src7, src7, mask0, mask1, mask2, mask3, filt_hz0,
+                              filt_hz1, filt_hz2, filt_hz3);
+    out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+    tmp0 = FILT_8TAP_DPADD_S_H(out0, out1, out2, out3, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out8 = HORIZ_8TAP_FILT(src8, src8, mask0, mask1, mask2, mask3, filt_hz0,
+                              filt_hz1, filt_hz2, filt_hz3);
+    out7 = (v8i16)__msa_ilvev_b((v16i8)hz_out8, (v16i8)hz_out7);
+    tmp1 = FILT_8TAP_DPADD_S_H(out4, out5, out6, out7, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out9 = HORIZ_8TAP_FILT(src9, src9, mask0, mask1, mask2, mask3, filt_hz0,
+                              filt_hz1, filt_hz2, filt_hz3);
+    out8 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8);
+    tmp2 = FILT_8TAP_DPADD_S_H(out1, out2, out3, out8, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out10 = HORIZ_8TAP_FILT(src10, src10, mask0, mask1, mask2, mask3,
+                               filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    out9 = (v8i16)__msa_ilvev_b((v16i8)hz_out10, (v16i8)hz_out9);
+    tmp3 = FILT_8TAP_DPADD_S_H(out5, out6, out7, out9, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    SRARI_H4_SH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    SAT_SH4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+    CONVERT_UB_AVG_ST8x4_UB(tmp0, tmp1, tmp2, tmp3, dst0, dst1, dst2, dst3, dst,
+                            dst_stride);
+    dst += (4 * dst_stride);
+
+    hz_out6 = hz_out10;
+    out0 = out2;
+    out1 = out3;
+    out2 = out8;
+    out4 = out6;
+    out5 = out7;
+    out6 = out9;
+  }
+}
+
+static void common_hv_8ht_8vt_and_aver_dst_16w_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 2; multiple8_cnt--;) {
+    common_hv_8ht_8vt_and_aver_dst_8w_msa(src, src_stride, dst, dst_stride,
+                                          filter_horiz, filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_8ht_8vt_and_aver_dst_32w_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 4; multiple8_cnt--;) {
+    common_hv_8ht_8vt_and_aver_dst_8w_msa(src, src_stride, dst, dst_stride,
+                                          filter_horiz, filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_8ht_8vt_and_aver_dst_64w_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 8; multiple8_cnt--;) {
+    common_hv_8ht_8vt_and_aver_dst_8w_msa(src, src_stride, dst, dst_stride,
+                                          filter_horiz, filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_4x4_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, mask;
+  v16u8 filt_hz, filt_vt, vec0, vec1;
+  v16u8 dst0, dst1, dst2, dst3, res0, res1;
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, tmp0, tmp1, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  filt = LD_UH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+  hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+  hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+  hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2);
+  ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
+  DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+  SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+  PCKEV_B2_UB(tmp0, tmp0, tmp1, tmp1, res0, res1);
+  AVER_UB2_UB(res0, dst0, res1, dst2, res0, res1);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_4x8_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, mask;
+  v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3, res0, res1, res2, res3;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8u16 hz_out7, hz_out8, tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  src += (8 * src_stride);
+  src8 = LD_SB(src);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+  hz_out4 = HORIZ_2TAP_FILT_UH(src4, src5, mask, filt_hz, FILTER_BITS);
+  hz_out6 = HORIZ_2TAP_FILT_UH(src6, src7, mask, filt_hz, FILTER_BITS);
+  hz_out8 = HORIZ_2TAP_FILT_UH(src8, src8, mask, filt_hz, FILTER_BITS);
+  SLDI_B3_UH(hz_out2, hz_out4, hz_out6, hz_out0, hz_out2, hz_out4, hz_out1,
+             hz_out3, hz_out5, 8);
+  hz_out7 = (v8u16)__msa_pckod_d((v2i64)hz_out8, (v2i64)hz_out6);
+
+  LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+  ILVR_W4_UB(dst1, dst0, dst3, dst2, dst5, dst4, dst7, dst6, dst0, dst2, dst4,
+             dst6);
+  ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+  ILVEV_B2_UB(hz_out4, hz_out5, hz_out6, hz_out7, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt_vt, filt_vt, filt_vt, filt_vt, tmp0,
+              tmp1, tmp2, tmp3);
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_B4_UB(tmp0, tmp0, tmp1, tmp1, tmp2, tmp2, tmp3, tmp3, res0, res1, res2,
+              res3);
+  AVER_UB4_UB(res0, dst0, res1, dst2, res2, dst4, res3, dst6, res0, res1, res2,
+              res3);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+  dst += (4 * dst_stride);
+  ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_4w_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  if (4 == height) {
+    common_hv_2ht_2vt_and_aver_dst_4x4_msa(src, src_stride, dst, dst_stride,
+                                           filter_horiz, filter_vert);
+  } else if (8 == height) {
+    common_hv_2ht_2vt_and_aver_dst_4x8_msa(src, src_stride, dst, dst_stride,
+                                           filter_horiz, filter_vert);
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_8x4_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, mask;
+  v16u8 filt_hz, filt_vt, dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3;
+  v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+  src += (5 * src_stride);
+
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+  vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+  tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+  vec1 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+  tmp1 = __msa_dotp_u_h(vec1, filt_vt);
+
+  hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+  vec2 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+  tmp2 = __msa_dotp_u_h(vec2, filt_vt);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+  vec3 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+  tmp3 = __msa_dotp_u_h(vec3, filt_vt);
+
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_AVG_ST8x4_UB(tmp0, dst0, tmp1, dst1, tmp2, dst2, tmp3, dst3, dst,
+                     dst_stride);
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_8x8mult_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, mask;
+  v16u8 filt_hz, filt_vt, vec0, dst0, dst1, dst2, dst3;
+  v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_SB(src);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp1 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp2 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp3 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    PCKEV_AVG_ST8x4_UB(tmp0, dst0, tmp1, dst1, tmp2, dst2, tmp3, dst3, dst,
+                       dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_8w_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  if (4 == height) {
+    common_hv_2ht_2vt_and_aver_dst_8x4_msa(src, src_stride, dst, dst_stride,
+                                           filter_horiz, filter_vert);
+  } else {
+    common_hv_2ht_2vt_and_aver_dst_8x8mult_msa(
+        src, src_stride, dst, dst_stride, filter_horiz, filter_vert, height);
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_16w_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt_hz, filt_vt, vec0, vec1, dst0, dst1, dst2, dst3;
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, tmp0, tmp1;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB2(src, 8, src0, src1);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst0, dst);
+    dst += dst_stride;
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst1, dst);
+    dst += dst_stride;
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst2, dst);
+    dst += dst_stride;
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst3, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_32w_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 2; multiple8_cnt--;) {
+    common_hv_2ht_2vt_and_aver_dst_16w_msa(src, src_stride, dst, dst_stride,
+                                           filter_horiz, filter_vert, height);
+    src += 16;
+    dst += 16;
+  }
+}
+
+static void common_hv_2ht_2vt_and_aver_dst_64w_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter_horiz, int8_t *filter_vert, int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 4; multiple8_cnt--;) {
+    common_hv_2ht_2vt_and_aver_dst_16w_msa(src, src_stride, dst, dst_stride,
+                                           filter_horiz, filter_vert, height);
+    src += 16;
+    dst += 16;
+  }
+}
+
+void aom_convolve8_avg_msa(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int x_step_q4,
+                           const int16_t *filter_y, int y_step_q4, int w,
+                           int h) {
+  int8_t cnt, filt_hor[8], filt_ver[8];
+
+  assert(x_step_q4 == 16);
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  for (cnt = 0; cnt < 8; ++cnt) {
+    filt_hor[cnt] = filter_x[cnt];
+    filt_ver[cnt] = filter_y[cnt];
+  }
+
+  if (((const int32_t *)filter_x)[0] == 0 &&
+      ((const int32_t *)filter_y)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_hv_2ht_2vt_and_aver_dst_4w_msa(src, (int32_t)src_stride, dst,
+                                              (int32_t)dst_stride, &filt_hor[3],
+                                              &filt_ver[3], h);
+        break;
+      case 8:
+        common_hv_2ht_2vt_and_aver_dst_8w_msa(src, (int32_t)src_stride, dst,
+                                              (int32_t)dst_stride, &filt_hor[3],
+                                              &filt_ver[3], h);
+        break;
+      case 16:
+        common_hv_2ht_2vt_and_aver_dst_16w_msa(src, (int32_t)src_stride, dst,
+                                               (int32_t)dst_stride,
+                                               &filt_hor[3], &filt_ver[3], h);
+        break;
+      case 32:
+        common_hv_2ht_2vt_and_aver_dst_32w_msa(src, (int32_t)src_stride, dst,
+                                               (int32_t)dst_stride,
+                                               &filt_hor[3], &filt_ver[3], h);
+        break;
+      case 64:
+        common_hv_2ht_2vt_and_aver_dst_64w_msa(src, (int32_t)src_stride, dst,
+                                               (int32_t)dst_stride,
+                                               &filt_hor[3], &filt_ver[3], h);
+        break;
+      default:
+        aom_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x,
+                            x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  } else if (((const int32_t *)filter_x)[0] == 0 ||
+             ((const int32_t *)filter_y)[0] == 0) {
+    aom_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                        filter_y, y_step_q4, w, h);
+  } else {
+    switch (w) {
+      case 4:
+        common_hv_8ht_8vt_and_aver_dst_4w_msa(src, (int32_t)src_stride, dst,
+                                              (int32_t)dst_stride, filt_hor,
+                                              filt_ver, h);
+        break;
+      case 8:
+        common_hv_8ht_8vt_and_aver_dst_8w_msa(src, (int32_t)src_stride, dst,
+                                              (int32_t)dst_stride, filt_hor,
+                                              filt_ver, h);
+        break;
+      case 16:
+        common_hv_8ht_8vt_and_aver_dst_16w_msa(src, (int32_t)src_stride, dst,
+                                               (int32_t)dst_stride, filt_hor,
+                                               filt_ver, h);
+        break;
+      case 32:
+        common_hv_8ht_8vt_and_aver_dst_32w_msa(src, (int32_t)src_stride, dst,
+                                               (int32_t)dst_stride, filt_hor,
+                                               filt_ver, h);
+        break;
+      case 64:
+        common_hv_8ht_8vt_and_aver_dst_64w_msa(src, (int32_t)src_stride, dst,
+                                               (int32_t)dst_stride, filt_hor,
+                                               filt_ver, h);
+        break;
+      default:
+        aom_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x,
+                            x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/aom_convolve8_avg_vert_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve8_avg_vert_msa.c
new file mode 100644
index 0000000000..dae771104f
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/aom_convolve8_avg_vert_msa.c
@@ -0,0 +1,677 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/aom_convolve_msa.h"
+
+static void common_vt_8t_and_aver_dst_4w_msa(const uint8_t *src,
+                                             int32_t src_stride, uint8_t *dst,
+                                             int32_t dst_stride, int8_t *filter,
+                                             int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16u8 dst0, dst1, dst2, dst3, out;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, src2110, src4332, src6554, src8776;
+  v16i8 src10998, filt0, filt1, filt2, filt3;
+  v8i16 filt, out10, out32;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+             src54_r, src21_r);
+  ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+  ILVR_D3_SB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r, src2110,
+             src4332, src6554);
+  XORI_B3_128_SB(src2110, src4332, src6554);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+               src87_r, src98_r, src109_r);
+    ILVR_D2_SB(src87_r, src76_r, src109_r, src98_r, src8776, src10998);
+    XORI_B2_128_SB(src8776, src10998);
+    out10 = FILT_8TAP_DPADD_S_H(src2110, src4332, src6554, src8776, filt0,
+                                filt1, filt2, filt3);
+    out32 = FILT_8TAP_DPADD_S_H(src4332, src6554, src8776, src10998, filt0,
+                                filt1, filt2, filt3);
+    SRARI_H2_SH(out10, out32, FILTER_BITS);
+    SAT_SH2_SH(out10, out32, 7);
+    out = PCKEV_XORI128_UB(out10, out32);
+    ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
+
+    dst0 = (v16u8)__msa_ilvr_d((v2i64)dst2, (v2i64)dst0);
+    out = __msa_aver_u_b(out, dst0);
+
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src2110 = src6554;
+    src4332 = src8776;
+    src6554 = src10998;
+    src6 = src10;
+  }
+}
+
+static void common_vt_8t_and_aver_dst_8w_msa(const uint8_t *src,
+                                             int32_t src_stride, uint8_t *dst,
+                                             int32_t dst_stride, int8_t *filter,
+                                             int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16u8 dst0, dst1, dst2, dst3;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, filt0, filt1, filt2, filt3;
+  v8i16 filt, out0, out1, out2, out3;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+             src54_r, src21_r);
+  ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+               src87_r, src98_r, src109_r);
+    out0 = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0, filt1,
+                               filt2, filt3);
+    out1 = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0, filt1,
+                               filt2, filt3);
+    out2 = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0, filt1,
+                               filt2, filt3);
+    out3 = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0,
+                               filt1, filt2, filt3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    CONVERT_UB_AVG_ST8x4_UB(out0, out1, out2, out3, dst0, dst1, dst2, dst3, dst,
+                            dst_stride);
+    dst += (4 * dst_stride);
+
+    src10_r = src54_r;
+    src32_r = src76_r;
+    src54_r = src98_r;
+    src21_r = src65_r;
+    src43_r = src87_r;
+    src65_r = src109_r;
+    src6 = src10;
+  }
+}
+
+static void common_vt_8t_and_aver_dst_16w_mult_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter, int32_t height, int32_t width) {
+  const uint8_t *src_tmp;
+  uint8_t *dst_tmp;
+  uint32_t loop_cnt, cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, src10_l, src32_l, src54_l, src76_l;
+  v16i8 src98_l, src21_l, src43_l, src65_l, src87_l, src109_l;
+  v16i8 filt0, filt1, filt2, filt3;
+  v16u8 dst0, dst1, dst2, dst3, tmp0, tmp1, tmp2, tmp3;
+  v8i16 out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l, filt;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  for (cnt = (width >> 4); cnt--;) {
+    src_tmp = src;
+    dst_tmp = dst;
+
+    LD_SB7(src_tmp, src_stride, src0, src1, src2, src3, src4, src5, src6);
+    XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+    src_tmp += (7 * src_stride);
+
+    ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+               src54_r, src21_r);
+    ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+    ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l, src32_l,
+               src54_l, src21_l);
+    ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;) {
+      LD_SB4(src_tmp, src_stride, src7, src8, src9, src10);
+      src_tmp += (4 * src_stride);
+
+      LD_UB4(dst_tmp, dst_stride, dst0, dst1, dst2, dst3);
+      XORI_B4_128_SB(src7, src8, src9, src10);
+      ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+                 src87_r, src98_r, src109_r);
+      ILVL_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_l,
+                 src87_l, src98_l, src109_l);
+      out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0,
+                                   filt1, filt2, filt3);
+      out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0,
+                                   filt1, filt2, filt3);
+      out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0,
+                                   filt1, filt2, filt3);
+      out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0,
+                                   filt1, filt2, filt3);
+      out0_l = FILT_8TAP_DPADD_S_H(src10_l, src32_l, src54_l, src76_l, filt0,
+                                   filt1, filt2, filt3);
+      out1_l = FILT_8TAP_DPADD_S_H(src21_l, src43_l, src65_l, src87_l, filt0,
+                                   filt1, filt2, filt3);
+      out2_l = FILT_8TAP_DPADD_S_H(src32_l, src54_l, src76_l, src98_l, filt0,
+                                   filt1, filt2, filt3);
+      out3_l = FILT_8TAP_DPADD_S_H(src43_l, src65_l, src87_l, src109_l, filt0,
+                                   filt1, filt2, filt3);
+      SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS);
+      SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, FILTER_BITS);
+      SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+      SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7);
+      PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l,
+                  out3_r, tmp0, tmp1, tmp2, tmp3);
+      XORI_B4_128_UB(tmp0, tmp1, tmp2, tmp3);
+      AVER_UB4_UB(tmp0, dst0, tmp1, dst1, tmp2, dst2, tmp3, dst3, dst0, dst1,
+                  dst2, dst3);
+      ST_UB4(dst0, dst1, dst2, dst3, dst_tmp, dst_stride);
+      dst_tmp += (4 * dst_stride);
+
+      src10_r = src54_r;
+      src32_r = src76_r;
+      src54_r = src98_r;
+      src21_r = src65_r;
+      src43_r = src87_r;
+      src65_r = src109_r;
+      src10_l = src54_l;
+      src32_l = src76_l;
+      src54_l = src98_l;
+      src21_l = src65_l;
+      src43_l = src87_l;
+      src65_l = src109_l;
+      src6 = src10;
+    }
+
+    src += 16;
+    dst += 16;
+  }
+}
+
+static void common_vt_8t_and_aver_dst_16w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  common_vt_8t_and_aver_dst_16w_mult_msa(src, src_stride, dst, dst_stride,
+                                         filter, height, 16);
+}
+
+static void common_vt_8t_and_aver_dst_32w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  common_vt_8t_and_aver_dst_16w_mult_msa(src, src_stride, dst, dst_stride,
+                                         filter, height, 32);
+}
+
+static void common_vt_8t_and_aver_dst_64w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  common_vt_8t_and_aver_dst_16w_mult_msa(src, src_stride, dst, dst_stride,
+                                         filter, height, 64);
+}
+
+static void common_vt_2t_and_aver_dst_4x4_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16i8 src0, src1, src2, src3, src4;
+  v16u8 dst0, dst1, dst2, dst3, out, filt0, src2110, src4332;
+  v16i8 src10_r, src32_r, src21_r, src43_r;
+  v8i16 filt;
+  v8u16 tmp0, tmp1;
+
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  src += (4 * src_stride);
+
+  src4 = LD_SB(src);
+  src += src_stride;
+
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst1);
+  dst0 = (v16u8)__msa_ilvr_d((v2i64)dst1, (v2i64)dst0);
+  ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+             src32_r, src43_r);
+  ILVR_D2_UB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+  DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1);
+  SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+
+  out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+  out = __msa_aver_u_b(out, dst0);
+
+  ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_vt_2t_and_aver_dst_4x8_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src87_r;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r, src65_r;
+  v16u8 src2110, src4332, src6554, src8776, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  src += (8 * src_stride);
+  src8 = LD_SB(src);
+
+  LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+  ILVR_W4_UB(dst1, dst0, dst3, dst2, dst5, dst4, dst7, dst6, dst0, dst1, dst2,
+             dst3);
+  ILVR_D2_UB(dst1, dst0, dst3, dst2, dst0, dst1);
+  ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+             src32_r, src43_r);
+  ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r, src65_r,
+             src76_r, src87_r);
+  ILVR_D4_UB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r, src87_r,
+             src76_r, src2110, src4332, src6554, src8776);
+  DOTP_UB4_UH(src2110, src4332, src6554, src8776, filt0, filt0, filt0, filt0,
+              tmp0, tmp1, tmp2, tmp3);
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, src2110, src4332);
+  AVER_UB2_UB(src2110, dst0, src4332, dst1, src2110, src4332);
+  ST4x4_UB(src2110, src2110, 0, 1, 2, 3, dst, dst_stride);
+  dst += (4 * dst_stride);
+  ST4x4_UB(src4332, src4332, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_vt_2t_and_aver_dst_4w_msa(const uint8_t *src,
+                                             int32_t src_stride, uint8_t *dst,
+                                             int32_t dst_stride, int8_t *filter,
+                                             int32_t height) {
+  if (4 == height) {
+    common_vt_2t_and_aver_dst_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_vt_2t_and_aver_dst_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_vt_2t_and_aver_dst_8x4_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter) {
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 dst0, dst1, dst2, dst3, vec0, vec1, vec2, vec3, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_UB5(src, src_stride, src0, src1, src2, src3, src4);
+  LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+  ILVR_B2_UB(src1, src0, src2, src1, vec0, vec1);
+  ILVR_B2_UB(src3, src2, src4, src3, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1,
+              tmp2, tmp3);
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_AVG_ST8x4_UB(tmp0, dst0, tmp1, dst1, tmp2, dst2, tmp3, dst3, dst,
+                     dst_stride);
+}
+
+static void common_vt_2t_and_aver_dst_8x8mult_msa(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+  v16u8 dst1, dst2, dst3, dst4, dst5, dst6, dst7, dst8;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 3); loop_cnt--;) {
+    LD_UB8(src, src_stride, src1, src2, src3, src4, src5, src6, src7, src8);
+    src += (8 * src_stride);
+    LD_UB8(dst, dst_stride, dst1, dst2, dst3, dst4, dst5, dst6, dst7, dst8);
+
+    ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, vec0, vec1, vec2,
+               vec3);
+    ILVR_B4_UB(src5, src4, src6, src5, src7, src6, src8, src7, vec4, vec5, vec6,
+               vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1,
+                tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST8x4_UB(tmp0, dst1, tmp1, dst2, tmp2, dst3, tmp3, dst4, dst,
+                       dst_stride);
+    dst += (4 * dst_stride);
+
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, tmp0, tmp1,
+                tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST8x4_UB(tmp0, dst5, tmp1, dst6, tmp2, dst7, tmp3, dst8, dst,
+                       dst_stride);
+    dst += (4 * dst_stride);
+
+    src0 = src8;
+  }
+}
+
+static void common_vt_2t_and_aver_dst_8w_msa(const uint8_t *src,
+                                             int32_t src_stride, uint8_t *dst,
+                                             int32_t dst_stride, int8_t *filter,
+                                             int32_t height) {
+  if (4 == height) {
+    common_vt_2t_and_aver_dst_8x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else {
+    common_vt_2t_and_aver_dst_8x8mult_msa(src, src_stride, dst, dst_stride,
+                                          filter, height);
+  }
+}
+
+static void common_vt_2t_and_aver_dst_16w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, dst0, dst1, dst2, dst3, filt0;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 tmp0, tmp1, tmp2, tmp3, filt;
+
+  /* rearranging filter_y */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst0, dst);
+    dst += dst_stride;
+
+    ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst1, dst);
+    dst += dst_stride;
+
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst2, dst);
+    dst += dst_stride;
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst3, dst);
+    dst += dst_stride;
+
+    src0 = src4;
+  }
+}
+
+static void common_vt_2t_and_aver_dst_32w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3, filt;
+
+  /* rearranging filter_y */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_UB2(src, 16, src0, src5);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+
+    LD_UB4(src + 16, src_stride, src6, src7, src8, src9);
+    LD_UB4(dst + 16, dst_stride, dst4, dst5, dst6, dst7);
+    src += (4 * src_stride);
+
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst0, dst);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst1, dst + dst_stride);
+
+    ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst2, dst + 2 * dst_stride);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst3, dst + 3 * dst_stride);
+
+    ILVR_B2_UB(src6, src5, src7, src6, vec0, vec2);
+    ILVL_B2_UB(src6, src5, src7, src6, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst4, dst + 16);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst5, dst + 16 + dst_stride);
+
+    ILVR_B2_UB(src8, src7, src9, src8, vec4, vec6);
+    ILVL_B2_UB(src8, src7, src9, src8, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst6, dst + 16 + 2 * dst_stride);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst7, dst + 16 + 3 * dst_stride);
+    dst += (4 * dst_stride);
+
+    src0 = src4;
+    src5 = src9;
+  }
+}
+
+static void common_vt_2t_and_aver_dst_64w_msa(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5;
+  v16u8 src6, src7, src8, src9, src10, src11, filt0;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8u16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_UB4(src, 16, src0, src3, src6, src9);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 1); loop_cnt--;) {
+    LD_UB2(src, src_stride, src1, src2);
+    LD_UB2(dst, dst_stride, dst0, dst1);
+    LD_UB2(src + 16, src_stride, src4, src5);
+    LD_UB2(dst + 16, dst_stride, dst2, dst3);
+    LD_UB2(src + 32, src_stride, src7, src8);
+    LD_UB2(dst + 32, dst_stride, dst4, dst5);
+    LD_UB2(src + 48, src_stride, src10, src11);
+    LD_UB2(dst + 48, dst_stride, dst6, dst7);
+    src += (2 * src_stride);
+
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst0, dst);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst1, dst + dst_stride);
+
+    ILVR_B2_UB(src4, src3, src5, src4, vec4, vec6);
+    ILVL_B2_UB(src4, src3, src5, src4, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp4, tmp5);
+    SRARI_H2_UH(tmp4, tmp5, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp5, tmp4, dst2, dst + 16);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp6, tmp7);
+    SRARI_H2_UH(tmp6, tmp7, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp7, tmp6, dst3, dst + 16 + dst_stride);
+
+    ILVR_B2_UB(src7, src6, src8, src7, vec0, vec2);
+    ILVL_B2_UB(src7, src6, src8, src7, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp1, tmp0, dst4, dst + 32);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp3, tmp2, dst5, dst + 32 + dst_stride);
+
+    ILVR_B2_UB(src10, src9, src11, src10, vec4, vec6);
+    ILVL_B2_UB(src10, src9, src11, src10, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp4, tmp5);
+    SRARI_H2_UH(tmp4, tmp5, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp5, tmp4, dst6, (dst + 48));
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp6, tmp7);
+    SRARI_H2_UH(tmp6, tmp7, FILTER_BITS);
+    PCKEV_AVG_ST_UB(tmp7, tmp6, dst7, dst + 48 + dst_stride);
+    dst += (2 * dst_stride);
+
+    src0 = src2;
+    src3 = src5;
+    src6 = src8;
+    src9 = src11;
+  }
+}
+
+void aom_convolve8_avg_vert_msa(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int x_step_q4,
+                                const int16_t *filter_y, int y_step_q4, int w,
+                                int h) {
+  int8_t cnt, filt_ver[8];
+
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  for (cnt = 0; cnt < 8; ++cnt) {
+    filt_ver[cnt] = filter_y[cnt];
+  }
+
+  if (((const int32_t *)filter_y)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_vt_2t_and_aver_dst_4w_msa(src, (int32_t)src_stride, dst,
+                                         (int32_t)dst_stride, &filt_ver[3], h);
+        break;
+      case 8:
+        common_vt_2t_and_aver_dst_8w_msa(src, (int32_t)src_stride, dst,
+                                         (int32_t)dst_stride, &filt_ver[3], h);
+        break;
+      case 16:
+        common_vt_2t_and_aver_dst_16w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, &filt_ver[3], h);
+        break;
+      case 32:
+        common_vt_2t_and_aver_dst_32w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, &filt_ver[3], h);
+        break;
+      case 64:
+        common_vt_2t_and_aver_dst_64w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, &filt_ver[3], h);
+        break;
+      default:
+        aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                                 x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  } else {
+    switch (w) {
+      case 4:
+        common_vt_8t_and_aver_dst_4w_msa(src, (int32_t)src_stride, dst,
+                                         (int32_t)dst_stride, filt_ver, h);
+        break;
+      case 8:
+        common_vt_8t_and_aver_dst_8w_msa(src, (int32_t)src_stride, dst,
+                                         (int32_t)dst_stride, filt_ver, h);
+        break;
+      case 16:
+        common_vt_8t_and_aver_dst_16w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, filt_ver, h);
+
+        break;
+      case 32:
+        common_vt_8t_and_aver_dst_32w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, filt_ver, h);
+        break;
+      case 64:
+        common_vt_8t_and_aver_dst_64w_msa(src, (int32_t)src_stride, dst,
+                                          (int32_t)dst_stride, filt_ver, h);
+        break;
+      default:
+        aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                                 x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/aom_convolve8_horiz_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve8_horiz_msa.c
new file mode 100644
index 0000000000..fc3a823c57
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/aom_convolve8_horiz_msa.c
@@ -0,0 +1,692 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/aom_convolve_msa.h"
+
+static void common_hz_8t_4x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v8i16 filt, out0, out1;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, out0, out1);
+  SRARI_H2_SH(out0, out1, FILTER_BITS);
+  SAT_SH2_SH(out0, out1, 7);
+  out = PCKEV_XORI128_UB(out0, out1);
+  ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_hz_8t_4x8_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16i8 filt0, filt1, filt2, filt3;
+  v16i8 src0, src1, src2, src3;
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  src += (4 * src_stride);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, out0, out1);
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, out2, out3);
+  SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+  SAT_SH4_SH(out0, out1, out2, out3, 7);
+  out = PCKEV_XORI128_UB(out0, out1);
+  ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+  dst += (4 * dst_stride);
+  out = PCKEV_XORI128_UB(out2, out3);
+  ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_hz_8t_4w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_hz_8t_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_hz_8t_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_hz_8t_8x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, tmp0, tmp1;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  XORI_B4_128_SB(src0, src1, src2, src3);
+  HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2, mask3,
+                             filt0, filt1, filt2, filt3, out0, out1, out2,
+                             out3);
+  SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+  SAT_SH4_SH(out0, out1, out2, out3, 7);
+  tmp0 = PCKEV_XORI128_UB(out0, out1);
+  tmp1 = PCKEV_XORI128_UB(out2, out3);
+  ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+}
+
+static void common_hz_8t_8x8mult_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, tmp0, tmp1;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    src += (4 * src_stride);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    tmp0 = PCKEV_XORI128_UB(out0, out1);
+    tmp1 = PCKEV_XORI128_UB(out2, out3);
+    ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void common_hz_8t_8w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_hz_8t_8x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else {
+    common_hz_8t_8x8mult_msa(src, src_stride, dst, dst_stride, filter, height);
+  }
+}
+
+static void common_hz_8t_16w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = (height >> 1); loop_cnt--;) {
+    LD_SB2(src, src_stride, src0, src2);
+    LD_SB2(src + 8, src_stride, src1, src3);
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    src += (2 * src_stride);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST_UB(out, dst);
+    dst += dst_stride;
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST_UB(out, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_8t_32w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = (height >> 1); loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST_UB(out, dst);
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST_UB(out, dst + 16);
+    dst += dst_stride;
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST_UB(out, dst);
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST_UB(out, dst + 16);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_8t_64w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  int32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, filt0, filt1, filt2, filt3;
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v8i16 filt, out0, out1, out2, out3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= 3;
+
+  /* rearranging filter */
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  for (loop_cnt = height; loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST_UB(out, dst);
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST_UB(out, dst + 16);
+
+    src0 = LD_SB(src + 32);
+    src2 = LD_SB(src + 48);
+    src3 = LD_SB(src + 56);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+
+    XORI_B4_128_SB(src0, src1, src2, src3);
+    HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1, mask2,
+                               mask3, filt0, filt1, filt2, filt3, out0, out1,
+                               out2, out3);
+    SRARI_H4_SH(out0, out1, out2, out3, FILTER_BITS);
+    SAT_SH4_SH(out0, out1, out2, out3, 7);
+    out = PCKEV_XORI128_UB(out0, out1);
+    ST_UB(out, dst + 32);
+    out = PCKEV_XORI128_UB(out2, out3);
+    ST_UB(out, dst + 48);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_4x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16i8 src0, src1, src2, src3, mask;
+  v16u8 filt0, vec0, vec1, res0, res1;
+  v8u16 vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
+  DOTP_UB2_UH(vec0, vec1, filt0, filt0, vec2, vec3);
+  SRARI_H2_UH(vec2, vec3, FILTER_BITS);
+  PCKEV_B2_UB(vec2, vec2, vec3, vec3, res0, res1);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_2t_4x8_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16u8 vec0, vec1, vec2, vec3, filt0;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16i8 res0, res1, res2, res3;
+  v8u16 vec4, vec5, vec6, vec7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  VSHF_B2_UB(src0, src1, src2, src3, mask, mask, vec0, vec1);
+  VSHF_B2_UB(src4, src5, src6, src7, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec4, vec5,
+              vec6, vec7);
+  SRARI_H4_UH(vec4, vec5, vec6, vec7, FILTER_BITS);
+  PCKEV_B4_SB(vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7, res0, res1, res2,
+              res3);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+  dst += (4 * dst_stride);
+  ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hz_2t_4w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_hz_2t_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_hz_2t_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_hz_2t_8x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16u8 filt0;
+  v16i8 src0, src1, src2, src3, mask;
+  v8u16 vec0, vec1, vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  PCKEV_B2_SB(vec1, vec0, vec3, vec2, src0, src1);
+  ST8x4_UB(src0, src1, dst, dst_stride);
+}
+
+static void common_hz_2t_8x8mult_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter, int32_t height) {
+  v16u8 filt0;
+  v16i8 src0, src1, src2, src3, mask, out0, out1;
+  v8u16 vec0, vec1, vec2, vec3, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  src += (4 * src_stride);
+
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+
+  LD_SB4(src, src_stride, src0, src1, src2, src3);
+  src += (4 * src_stride);
+
+  PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+  ST8x4_UB(out0, out1, dst, dst_stride);
+  dst += (4 * dst_stride);
+
+  VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+              vec2, vec3);
+  SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+  PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+  ST8x4_UB(out0, out1, dst, dst_stride);
+  dst += (4 * dst_stride);
+
+  if (16 == height) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+
+    PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    PCKEV_B2_SB(vec1, vec0, vec3, vec2, out0, out1);
+    ST8x4_UB(out0, out1, dst + 4 * dst_stride, dst_stride);
+  }
+}
+
+static void common_hz_2t_8w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_hz_2t_8x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else {
+    common_hz_2t_8x8mult_msa(src, src_stride, dst, dst_stride, filter, height);
+  }
+}
+
+static void common_hz_2t_16w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  loop_cnt = (height >> 2) - 1;
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB4(src, src_stride, src0, src2, src4, src6);
+  LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+  src += (4 * src_stride);
+
+  VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+  VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+  VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+  VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+              out2, out3);
+  DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+              out6, out7);
+  SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+  SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+  PCKEV_ST_SB(out0, out1, dst);
+  dst += dst_stride;
+  PCKEV_ST_SB(out2, out3, dst);
+  dst += dst_stride;
+  PCKEV_ST_SB(out4, out5, dst);
+  dst += dst_stride;
+  PCKEV_ST_SB(out6, out7, dst);
+  dst += dst_stride;
+
+  for (; loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+                out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+                out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    PCKEV_ST_SB(out0, out1, dst);
+    dst += dst_stride;
+    PCKEV_ST_SB(out2, out3, dst);
+    dst += dst_stride;
+    PCKEV_ST_SB(out4, out5, dst);
+    dst += dst_stride;
+    PCKEV_ST_SB(out6, out7, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_32w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  for (loop_cnt = height >> 1; loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src3 = LD_SB(src + 24);
+    src1 = __msa_sldi_b(src2, src0, 8);
+    src += src_stride;
+    src4 = LD_SB(src);
+    src6 = LD_SB(src + 16);
+    src7 = LD_SB(src + 24);
+    src5 = __msa_sldi_b(src6, src4, 8);
+    src += src_stride;
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+                out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+                out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    PCKEV_ST_SB(out0, out1, dst);
+    PCKEV_ST_SB(out2, out3, dst + 16);
+    dst += dst_stride;
+    PCKEV_ST_SB(out4, out5, dst);
+    PCKEV_ST_SB(out6, out7, dst + 16);
+    dst += dst_stride;
+  }
+}
+
+static void common_hz_2t_64w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt0, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter);
+  filt0 = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  for (loop_cnt = height; loop_cnt--;) {
+    src0 = LD_SB(src);
+    src2 = LD_SB(src + 16);
+    src4 = LD_SB(src + 32);
+    src6 = LD_SB(src + 48);
+    src7 = LD_SB(src + 56);
+    SLDI_B3_SB(src2, src4, src6, src0, src2, src4, src1, src3, src5, 8);
+    src += src_stride;
+
+    VSHF_B2_UB(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UB(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UB(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UB(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+                out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+                out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    PCKEV_ST_SB(out0, out1, dst);
+    PCKEV_ST_SB(out2, out3, dst + 16);
+    PCKEV_ST_SB(out4, out5, dst + 32);
+    PCKEV_ST_SB(out6, out7, dst + 48);
+    dst += dst_stride;
+  }
+}
+
+void aom_convolve8_horiz_msa(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int x_step_q4,
+                             const int16_t *filter_y, int y_step_q4, int w,
+                             int h) {
+  int8_t cnt, filt_hor[8];
+
+  assert(x_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+
+  for (cnt = 0; cnt < 8; ++cnt) {
+    filt_hor[cnt] = filter_x[cnt];
+  }
+
+  if (((const int32_t *)filter_x)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_hz_2t_4w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                            &filt_hor[3], h);
+        break;
+      case 8:
+        common_hz_2t_8w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                            &filt_hor[3], h);
+        break;
+      case 16:
+        common_hz_2t_16w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             &filt_hor[3], h);
+        break;
+      case 32:
+        common_hz_2t_32w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             &filt_hor[3], h);
+        break;
+      case 64:
+        common_hz_2t_64w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             &filt_hor[3], h);
+        break;
+      default:
+        aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                              x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  } else {
+    switch (w) {
+      case 4:
+        common_hz_8t_4w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                            filt_hor, h);
+        break;
+      case 8:
+        common_hz_8t_8w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                            filt_hor, h);
+        break;
+      case 16:
+        common_hz_8t_16w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             filt_hor, h);
+        break;
+      case 32:
+        common_hz_8t_32w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             filt_hor, h);
+        break;
+      case 64:
+        common_hz_8t_64w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             filt_hor, h);
+        break;
+      default:
+        aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                              x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/aom_convolve8_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve8_msa.c
new file mode 100644
index 0000000000..a4d594931a
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/aom_convolve8_msa.c
@@ -0,0 +1,630 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/aom_convolve_msa.h"
+
+const uint8_t mc_filt_mask_arr[16 * 3] = {
+  /* 8 width cases */
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+  /* 4 width cases */
+  0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20,
+  /* 4 width cases */
+  8, 9, 9, 10, 10, 11, 11, 12, 24, 25, 25, 26, 26, 27, 27, 28
+};
+
+static void common_hv_8ht_8vt_4w_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter_horiz, int8_t *filter_vert,
+                                     int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3;
+  v16u8 mask0, mask1, mask2, mask3, out;
+  v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8i16 hz_out7, hz_out8, hz_out9, tmp0, tmp1, out0, out1, out2, out3, out4;
+  v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[16]);
+  src -= (3 + 3 * src_stride);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  hz_out0 = HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out2 = HORIZ_8TAP_FILT(src2, src3, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out4 = HORIZ_8TAP_FILT(src4, src5, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out5 = HORIZ_8TAP_FILT(src5, src6, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  SLDI_B2_SH(hz_out2, hz_out4, hz_out0, hz_out2, hz_out1, hz_out3, 8);
+
+  filt = LD_SH(filter_vert);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
+
+  ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
+  out2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    hz_out7 = HORIZ_8TAP_FILT(src7, src8, mask0, mask1, mask2, mask3, filt_hz0,
+                              filt_hz1, filt_hz2, filt_hz3);
+    hz_out6 = (v8i16)__msa_sldi_b((v16i8)hz_out7, (v16i8)hz_out5, 8);
+    out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+    tmp0 = FILT_8TAP_DPADD_S_H(out0, out1, out2, out3, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out9 = HORIZ_8TAP_FILT(src9, src10, mask0, mask1, mask2, mask3, filt_hz0,
+                              filt_hz1, filt_hz2, filt_hz3);
+    hz_out8 = (v8i16)__msa_sldi_b((v16i8)hz_out9, (v16i8)hz_out7, 8);
+    out4 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8);
+    tmp1 = FILT_8TAP_DPADD_S_H(out1, out2, out3, out4, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+    SRARI_H2_SH(tmp0, tmp1, FILTER_BITS);
+    SAT_SH2_SH(tmp0, tmp1, 7);
+    out = PCKEV_XORI128_UB(tmp0, tmp1);
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    hz_out5 = hz_out9;
+    out0 = out2;
+    out1 = out3;
+    out2 = out4;
+  }
+}
+
+static void common_hv_8ht_8vt_8w_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter_horiz, int8_t *filter_vert,
+                                     int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3;
+  v16u8 mask0, mask1, mask2, mask3, vec0, vec1;
+  v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3;
+  v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8i16 hz_out7, hz_out8, hz_out9, hz_out10, tmp0, tmp1, tmp2, tmp3;
+  v8i16 out0, out1, out2, out3, out4, out5, out6, out7, out8, out9;
+
+  mask0 = LD_UB(&mc_filt_mask_arr[0]);
+  src -= (3 + 3 * src_stride);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+
+  mask1 = mask0 + 2;
+  mask2 = mask0 + 4;
+  mask3 = mask0 + 6;
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  hz_out0 = HORIZ_8TAP_FILT(src0, src0, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out1 = HORIZ_8TAP_FILT(src1, src1, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out2 = HORIZ_8TAP_FILT(src2, src2, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out3 = HORIZ_8TAP_FILT(src3, src3, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out4 = HORIZ_8TAP_FILT(src4, src4, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out5 = HORIZ_8TAP_FILT(src5, src5, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+  hz_out6 = HORIZ_8TAP_FILT(src6, src6, mask0, mask1, mask2, mask3, filt_hz0,
+                            filt_hz1, filt_hz2, filt_hz3);
+
+  filt = LD_SH(filter_vert);
+  SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
+
+  ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
+  ILVEV_B2_SH(hz_out4, hz_out5, hz_out1, hz_out2, out2, out4);
+  ILVEV_B2_SH(hz_out3, hz_out4, hz_out5, hz_out6, out5, out6);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    XORI_B4_128_SB(src7, src8, src9, src10);
+
+    hz_out7 = HORIZ_8TAP_FILT(src7, src7, mask0, mask1, mask2, mask3, filt_hz0,
+                              filt_hz1, filt_hz2, filt_hz3);
+    out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
+    tmp0 = FILT_8TAP_DPADD_S_H(out0, out1, out2, out3, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out8 = HORIZ_8TAP_FILT(src8, src8, mask0, mask1, mask2, mask3, filt_hz0,
+                              filt_hz1, filt_hz2, filt_hz3);
+    out7 = (v8i16)__msa_ilvev_b((v16i8)hz_out8, (v16i8)hz_out7);
+    tmp1 = FILT_8TAP_DPADD_S_H(out4, out5, out6, out7, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out9 = HORIZ_8TAP_FILT(src9, src9, mask0, mask1, mask2, mask3, filt_hz0,
+                              filt_hz1, filt_hz2, filt_hz3);
+    out8 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8);
+    tmp2 = FILT_8TAP_DPADD_S_H(out1, out2, out3, out8, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+
+    hz_out10 = HORIZ_8TAP_FILT(src10, src10, mask0, mask1, mask2, mask3,
+                               filt_hz0, filt_hz1, filt_hz2, filt_hz3);
+    out9 = (v8i16)__msa_ilvev_b((v16i8)hz_out10, (v16i8)hz_out9);
+    tmp3 = FILT_8TAP_DPADD_S_H(out5, out6, out7, out9, filt_vt0, filt_vt1,
+                               filt_vt2, filt_vt3);
+    SRARI_H4_SH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    SAT_SH4_SH(tmp0, tmp1, tmp2, tmp3, 7);
+    vec0 = PCKEV_XORI128_UB(tmp0, tmp1);
+    vec1 = PCKEV_XORI128_UB(tmp2, tmp3);
+    ST8x4_UB(vec0, vec1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    hz_out6 = hz_out10;
+    out0 = out2;
+    out1 = out3;
+    out2 = out8;
+    out4 = out6;
+    out5 = out7;
+    out6 = out9;
+  }
+}
+
+static void common_hv_8ht_8vt_16w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 2; multiple8_cnt--;) {
+    common_hv_8ht_8vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                             filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_8ht_8vt_32w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 4; multiple8_cnt--;) {
+    common_hv_8ht_8vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                             filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_8ht_8vt_64w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 8; multiple8_cnt--;) {
+    common_hv_8ht_8vt_8w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                             filter_vert, height);
+    src += 8;
+    dst += 8;
+  }
+}
+
+static void common_hv_2ht_2vt_4x4_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz,
+                                      int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, mask;
+  v16u8 filt_vt, filt_hz, vec0, vec1, res0, res1;
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, filt, tmp0, tmp1;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  filt = LD_UH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+  hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+  hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+  hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2);
+
+  ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+  DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+  SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+  PCKEV_B2_UB(tmp0, tmp0, tmp1, tmp1, res0, res1);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_4x8_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz,
+                                      int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, mask;
+  v16i8 res0, res1, res2, res3;
+  v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3;
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
+  v8u16 hz_out7, hz_out8, vec4, vec5, vec6, vec7, filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[16]);
+
+  /* rearranging filter */
+  filt = LD_UH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  filt = LD_UH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h((v8i16)filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  src += (8 * src_stride);
+  src8 = LD_SB(src);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+  hz_out4 = HORIZ_2TAP_FILT_UH(src4, src5, mask, filt_hz, FILTER_BITS);
+  hz_out6 = HORIZ_2TAP_FILT_UH(src6, src7, mask, filt_hz, FILTER_BITS);
+  hz_out8 = HORIZ_2TAP_FILT_UH(src8, src8, mask, filt_hz, FILTER_BITS);
+  SLDI_B3_UH(hz_out2, hz_out4, hz_out6, hz_out0, hz_out2, hz_out4, hz_out1,
+             hz_out3, hz_out5, 8);
+  hz_out7 = (v8u16)__msa_pckod_d((v2i64)hz_out8, (v2i64)hz_out6);
+
+  ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+  ILVEV_B2_UB(hz_out4, hz_out5, hz_out6, hz_out7, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt_vt, filt_vt, filt_vt, filt_vt, vec4,
+              vec5, vec6, vec7);
+  SRARI_H4_UH(vec4, vec5, vec6, vec7, FILTER_BITS);
+  PCKEV_B4_SB(vec4, vec4, vec5, vec5, vec6, vec6, vec7, vec7, res0, res1, res2,
+              res3);
+  ST4x4_UB(res0, res1, 0, 1, 0, 1, dst, dst_stride);
+  dst += (4 * dst_stride);
+  ST4x4_UB(res2, res3, 0, 1, 0, 1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_4w_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter_horiz, int8_t *filter_vert,
+                                     int32_t height) {
+  if (4 == height) {
+    common_hv_2ht_2vt_4x4_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                              filter_vert);
+  } else if (8 == height) {
+    common_hv_2ht_2vt_4x8_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                              filter_vert);
+  }
+}
+
+static void common_hv_2ht_2vt_8x4_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz,
+                                      int8_t *filter_vert) {
+  v16i8 src0, src1, src2, src3, src4, mask, out0, out1;
+  v16u8 filt_hz, filt_vt, vec0, vec1, vec2, vec3;
+  v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+  vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+  tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+  vec1 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+  tmp1 = __msa_dotp_u_h(vec1, filt_vt);
+
+  hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+  vec2 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+  tmp2 = __msa_dotp_u_h(vec2, filt_vt);
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+  vec3 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+  tmp3 = __msa_dotp_u_h(vec3, filt_vt);
+
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+  ST8x4_UB(out0, out1, dst, dst_stride);
+}
+
+static void common_hv_2ht_2vt_8x8mult_msa(const uint8_t *src,
+                                          int32_t src_stride, uint8_t *dst,
+                                          int32_t dst_stride,
+                                          int8_t *filter_horiz,
+                                          int8_t *filter_vert, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, mask, out0, out1;
+  v16u8 filt_hz, filt_vt, vec0;
+  v8u16 hz_out0, hz_out1, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_SB(src);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 3); loop_cnt--;) {
+    LD_SB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp1 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp2 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H2_UH(tmp1, tmp2, FILTER_BITS);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp3 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    LD_SB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp4 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H2_UH(tmp3, tmp4, FILTER_BITS);
+    PCKEV_B2_SB(tmp2, tmp1, tmp4, tmp3, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp5 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp6 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp7 = __msa_dotp_u_h(vec0, filt_vt);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp8 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H4_UH(tmp5, tmp6, tmp7, tmp8, FILTER_BITS);
+    PCKEV_B2_SB(tmp6, tmp5, tmp8, tmp7, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void common_hv_2ht_2vt_8w_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter_horiz, int8_t *filter_vert,
+                                     int32_t height) {
+  if (4 == height) {
+    common_hv_2ht_2vt_8x4_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                              filter_vert);
+  } else {
+    common_hv_2ht_2vt_8x8mult_msa(src, src_stride, dst, dst_stride,
+                                  filter_horiz, filter_vert, height);
+  }
+}
+
+static void common_hv_2ht_2vt_16w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, mask;
+  v16u8 filt_hz, filt_vt, vec0, vec1;
+  v8u16 tmp1, tmp2, hz_out0, hz_out1, hz_out2, hz_out3;
+  v8i16 filt;
+
+  mask = LD_SB(&mc_filt_mask_arr[0]);
+
+  /* rearranging filter */
+  filt = LD_SH(filter_horiz);
+  filt_hz = (v16u8)__msa_splati_h(filt, 0);
+
+  filt = LD_SH(filter_vert);
+  filt_vt = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB2(src, 8, src0, src1);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+    SRARI_H2_UH(tmp1, tmp2, FILTER_BITS);
+    PCKEV_ST_SB(tmp1, tmp2, dst);
+    dst += dst_stride;
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+    SRARI_H2_UH(tmp1, tmp2, FILTER_BITS);
+    PCKEV_ST_SB(tmp1, tmp2, dst);
+    dst += dst_stride;
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+    SRARI_H2_UH(tmp1, tmp2, FILTER_BITS);
+    PCKEV_ST_SB(tmp1, tmp2, dst);
+    dst += dst_stride;
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp1, tmp2);
+    SRARI_H2_UH(tmp1, tmp2, FILTER_BITS);
+    PCKEV_ST_SB(tmp1, tmp2, dst);
+    dst += dst_stride;
+  }
+}
+
+static void common_hv_2ht_2vt_32w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 2; multiple8_cnt--;) {
+    common_hv_2ht_2vt_16w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                              filter_vert, height);
+    src += 16;
+    dst += 16;
+  }
+}
+
+static void common_hv_2ht_2vt_64w_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter_horiz, int8_t *filter_vert,
+                                      int32_t height) {
+  int32_t multiple8_cnt;
+  for (multiple8_cnt = 4; multiple8_cnt--;) {
+    common_hv_2ht_2vt_16w_msa(src, src_stride, dst, dst_stride, filter_horiz,
+                              filter_vert, height);
+    src += 16;
+    dst += 16;
+  }
+}
+
+void aom_convolve8_msa(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                       ptrdiff_t dst_stride, const int16_t *filter_x,
+                       int32_t x_step_q4, const int16_t *filter_y,
+                       int32_t y_step_q4, int32_t w, int32_t h) {
+  int8_t cnt, filt_hor[8], filt_ver[8];
+
+  assert(x_step_q4 == 16);
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  for (cnt = 0; cnt < 8; ++cnt) {
+    filt_hor[cnt] = filter_x[cnt];
+    filt_ver[cnt] = filter_y[cnt];
+  }
+
+  if (((const int32_t *)filter_x)[0] == 0 &&
+      ((const int32_t *)filter_y)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_hv_2ht_2vt_4w_msa(src, (int32_t)src_stride, dst,
+                                 (int32_t)dst_stride, &filt_hor[3],
+                                 &filt_ver[3], (int32_t)h);
+        break;
+      case 8:
+        common_hv_2ht_2vt_8w_msa(src, (int32_t)src_stride, dst,
+                                 (int32_t)dst_stride, &filt_hor[3],
+                                 &filt_ver[3], (int32_t)h);
+        break;
+      case 16:
+        common_hv_2ht_2vt_16w_msa(src, (int32_t)src_stride, dst,
+                                  (int32_t)dst_stride, &filt_hor[3],
+                                  &filt_ver[3], (int32_t)h);
+        break;
+      case 32:
+        common_hv_2ht_2vt_32w_msa(src, (int32_t)src_stride, dst,
+                                  (int32_t)dst_stride, &filt_hor[3],
+                                  &filt_ver[3], (int32_t)h);
+        break;
+      case 64:
+        common_hv_2ht_2vt_64w_msa(src, (int32_t)src_stride, dst,
+                                  (int32_t)dst_stride, &filt_hor[3],
+                                  &filt_ver[3], (int32_t)h);
+        break;
+      default:
+        aom_convolve8_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                        filter_y, y_step_q4, w, h);
+        break;
+    }
+  } else if (((const int32_t *)filter_x)[0] == 0 ||
+             ((const int32_t *)filter_y)[0] == 0) {
+    aom_convolve8_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                    filter_y, y_step_q4, w, h);
+  } else {
+    switch (w) {
+      case 4:
+        common_hv_8ht_8vt_4w_msa(src, (int32_t)src_stride, dst,
+                                 (int32_t)dst_stride, filt_hor, filt_ver,
+                                 (int32_t)h);
+        break;
+      case 8:
+        common_hv_8ht_8vt_8w_msa(src, (int32_t)src_stride, dst,
+                                 (int32_t)dst_stride, filt_hor, filt_ver,
+                                 (int32_t)h);
+        break;
+      case 16:
+        common_hv_8ht_8vt_16w_msa(src, (int32_t)src_stride, dst,
+                                  (int32_t)dst_stride, filt_hor, filt_ver,
+                                  (int32_t)h);
+        break;
+      case 32:
+        common_hv_8ht_8vt_32w_msa(src, (int32_t)src_stride, dst,
+                                  (int32_t)dst_stride, filt_hor, filt_ver,
+                                  (int32_t)h);
+        break;
+      case 64:
+        common_hv_8ht_8vt_64w_msa(src, (int32_t)src_stride, dst,
+                                  (int32_t)dst_stride, filt_hor, filt_ver,
+                                  (int32_t)h);
+        break;
+      default:
+        aom_convolve8_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                        filter_y, y_step_q4, w, h);
+        break;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/aom_convolve8_vert_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve8_vert_msa.c
new file mode 100644
index 0000000000..f7bdfc2bd7
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/aom_convolve8_vert_msa.c
@@ -0,0 +1,699 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/aom_convolve_msa.h"
+
+static void common_vt_8t_4w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, src2110, src4332, src6554, src8776;
+  v16i8 src10998, filt0, filt1, filt2, filt3;
+  v16u8 out;
+  v8i16 filt, out10, out32;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+
+  ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+             src54_r, src21_r);
+  ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+  ILVR_D3_SB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r, src2110,
+             src4332, src6554);
+  XORI_B3_128_SB(src2110, src4332, src6554);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+               src87_r, src98_r, src109_r);
+    ILVR_D2_SB(src87_r, src76_r, src109_r, src98_r, src8776, src10998);
+    XORI_B2_128_SB(src8776, src10998);
+    out10 = FILT_8TAP_DPADD_S_H(src2110, src4332, src6554, src8776, filt0,
+                                filt1, filt2, filt3);
+    out32 = FILT_8TAP_DPADD_S_H(src4332, src6554, src8776, src10998, filt0,
+                                filt1, filt2, filt3);
+    SRARI_H2_SH(out10, out32, FILTER_BITS);
+    SAT_SH2_SH(out10, out32, 7);
+    out = PCKEV_XORI128_UB(out10, out32);
+    ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src2110 = src6554;
+    src4332 = src8776;
+    src6554 = src10998;
+    src6 = src10;
+  }
+}
+
+static void common_vt_8t_8w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, filt0, filt1, filt2, filt3;
+  v16u8 tmp0, tmp1;
+  v8i16 filt, out0_r, out1_r, out2_r, out3_r;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+  ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+             src54_r, src21_r);
+  ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+               src87_r, src98_r, src109_r);
+    out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0,
+                                 filt1, filt2, filt3);
+    out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0,
+                                 filt1, filt2, filt3);
+    out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0,
+                                 filt1, filt2, filt3);
+    out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0,
+                                 filt1, filt2, filt3);
+    SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS);
+    SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+    tmp0 = PCKEV_XORI128_UB(out0_r, out1_r);
+    tmp1 = PCKEV_XORI128_UB(out2_r, out3_r);
+    ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src10_r = src54_r;
+    src32_r = src76_r;
+    src54_r = src98_r;
+    src21_r = src65_r;
+    src43_r = src87_r;
+    src65_r = src109_r;
+    src6 = src10;
+  }
+}
+
+static void common_vt_8t_16w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 filt0, filt1, filt2, filt3;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, src10_l, src32_l, src54_l, src76_l;
+  v16i8 src98_l, src21_l, src43_l, src65_l, src87_l, src109_l;
+  v16u8 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt, out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
+  XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+  src += (7 * src_stride);
+  ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+             src54_r, src21_r);
+  ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+  ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l, src32_l,
+             src54_l, src21_l);
+  ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src7, src8, src9, src10);
+    XORI_B4_128_SB(src7, src8, src9, src10);
+    src += (4 * src_stride);
+
+    ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+               src87_r, src98_r, src109_r);
+    ILVL_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_l,
+               src87_l, src98_l, src109_l);
+    out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0,
+                                 filt1, filt2, filt3);
+    out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0,
+                                 filt1, filt2, filt3);
+    out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0,
+                                 filt1, filt2, filt3);
+    out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0,
+                                 filt1, filt2, filt3);
+    out0_l = FILT_8TAP_DPADD_S_H(src10_l, src32_l, src54_l, src76_l, filt0,
+                                 filt1, filt2, filt3);
+    out1_l = FILT_8TAP_DPADD_S_H(src21_l, src43_l, src65_l, src87_l, filt0,
+                                 filt1, filt2, filt3);
+    out2_l = FILT_8TAP_DPADD_S_H(src32_l, src54_l, src76_l, src98_l, filt0,
+                                 filt1, filt2, filt3);
+    out3_l = FILT_8TAP_DPADD_S_H(src43_l, src65_l, src87_l, src109_l, filt0,
+                                 filt1, filt2, filt3);
+    SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS);
+    SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, FILTER_BITS);
+    SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+    SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7);
+    PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l, out3_r,
+                tmp0, tmp1, tmp2, tmp3);
+    XORI_B4_128_UB(tmp0, tmp1, tmp2, tmp3);
+    ST_UB4(tmp0, tmp1, tmp2, tmp3, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src10_r = src54_r;
+    src32_r = src76_r;
+    src54_r = src98_r;
+    src21_r = src65_r;
+    src43_r = src87_r;
+    src65_r = src109_r;
+    src10_l = src54_l;
+    src32_l = src76_l;
+    src54_l = src98_l;
+    src21_l = src65_l;
+    src43_l = src87_l;
+    src65_l = src109_l;
+    src6 = src10;
+  }
+}
+
+static void common_vt_8t_16w_mult_msa(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      int8_t *filter, int32_t height,
+                                      int32_t width) {
+  const uint8_t *src_tmp;
+  uint8_t *dst_tmp;
+  uint32_t loop_cnt, cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16i8 filt0, filt1, filt2, filt3;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src98_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src109_r, src10_l, src32_l, src54_l, src76_l;
+  v16i8 src98_l, src21_l, src43_l, src65_l, src87_l, src109_l;
+  v16u8 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt, out0_r, out1_r, out2_r, out3_r, out0_l, out1_l, out2_l, out3_l;
+
+  src -= (3 * src_stride);
+
+  filt = LD_SH(filter);
+  SPLATI_H4_SB(filt, 0, 1, 2, 3, filt0, filt1, filt2, filt3);
+
+  for (cnt = (width >> 4); cnt--;) {
+    src_tmp = src;
+    dst_tmp = dst;
+
+    LD_SB7(src_tmp, src_stride, src0, src1, src2, src3, src4, src5, src6);
+    XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
+    src_tmp += (7 * src_stride);
+    ILVR_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_r, src32_r,
+               src54_r, src21_r);
+    ILVR_B2_SB(src4, src3, src6, src5, src43_r, src65_r);
+    ILVL_B4_SB(src1, src0, src3, src2, src5, src4, src2, src1, src10_l, src32_l,
+               src54_l, src21_l);
+    ILVL_B2_SB(src4, src3, src6, src5, src43_l, src65_l);
+
+    for (loop_cnt = (height >> 2); loop_cnt--;) {
+      LD_SB4(src_tmp, src_stride, src7, src8, src9, src10);
+      XORI_B4_128_SB(src7, src8, src9, src10);
+      src_tmp += (4 * src_stride);
+      ILVR_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_r,
+                 src87_r, src98_r, src109_r);
+      ILVL_B4_SB(src7, src6, src8, src7, src9, src8, src10, src9, src76_l,
+                 src87_l, src98_l, src109_l);
+      out0_r = FILT_8TAP_DPADD_S_H(src10_r, src32_r, src54_r, src76_r, filt0,
+                                   filt1, filt2, filt3);
+      out1_r = FILT_8TAP_DPADD_S_H(src21_r, src43_r, src65_r, src87_r, filt0,
+                                   filt1, filt2, filt3);
+      out2_r = FILT_8TAP_DPADD_S_H(src32_r, src54_r, src76_r, src98_r, filt0,
+                                   filt1, filt2, filt3);
+      out3_r = FILT_8TAP_DPADD_S_H(src43_r, src65_r, src87_r, src109_r, filt0,
+                                   filt1, filt2, filt3);
+      out0_l = FILT_8TAP_DPADD_S_H(src10_l, src32_l, src54_l, src76_l, filt0,
+                                   filt1, filt2, filt3);
+      out1_l = FILT_8TAP_DPADD_S_H(src21_l, src43_l, src65_l, src87_l, filt0,
+                                   filt1, filt2, filt3);
+      out2_l = FILT_8TAP_DPADD_S_H(src32_l, src54_l, src76_l, src98_l, filt0,
+                                   filt1, filt2, filt3);
+      out3_l = FILT_8TAP_DPADD_S_H(src43_l, src65_l, src87_l, src109_l, filt0,
+                                   filt1, filt2, filt3);
+      SRARI_H4_SH(out0_r, out1_r, out2_r, out3_r, FILTER_BITS);
+      SRARI_H4_SH(out0_l, out1_l, out2_l, out3_l, FILTER_BITS);
+      SAT_SH4_SH(out0_r, out1_r, out2_r, out3_r, 7);
+      SAT_SH4_SH(out0_l, out1_l, out2_l, out3_l, 7);
+      PCKEV_B4_UB(out0_l, out0_r, out1_l, out1_r, out2_l, out2_r, out3_l,
+                  out3_r, tmp0, tmp1, tmp2, tmp3);
+      XORI_B4_128_UB(tmp0, tmp1, tmp2, tmp3);
+      ST_UB4(tmp0, tmp1, tmp2, tmp3, dst_tmp, dst_stride);
+      dst_tmp += (4 * dst_stride);
+
+      src10_r = src54_r;
+      src32_r = src76_r;
+      src54_r = src98_r;
+      src21_r = src65_r;
+      src43_r = src87_r;
+      src65_r = src109_r;
+      src10_l = src54_l;
+      src32_l = src76_l;
+      src54_l = src98_l;
+      src21_l = src65_l;
+      src43_l = src87_l;
+      src65_l = src109_l;
+      src6 = src10;
+    }
+
+    src += 16;
+    dst += 16;
+  }
+}
+
+static void common_vt_8t_32w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  common_vt_8t_16w_mult_msa(src, src_stride, dst, dst_stride, filter, height,
+                            32);
+}
+
+static void common_vt_8t_64w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  common_vt_8t_16w_mult_msa(src, src_stride, dst, dst_stride, filter, height,
+                            64);
+}
+
+static void common_vt_2t_4x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16i8 src0, src1, src2, src3, src4;
+  v16i8 src10_r, src32_r, src21_r, src43_r, src2110, src4332;
+  v16u8 filt0;
+  v8i16 filt;
+  v8u16 tmp0, tmp1;
+
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB5(src, src_stride, src0, src1, src2, src3, src4);
+  src += (5 * src_stride);
+
+  ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+             src32_r, src43_r);
+  ILVR_D2_SB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+  DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1);
+  SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+  src2110 = __msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+  ST4x4_UB(src2110, src2110, 0, 1, 2, 3, dst, dst_stride);
+}
+
+static void common_vt_2t_4x8_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+  v16i8 src10_r, src32_r, src54_r, src76_r, src21_r, src43_r;
+  v16i8 src65_r, src87_r, src2110, src4332, src6554, src8776;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 filt0;
+  v8i16 filt;
+
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  src += (8 * src_stride);
+
+  src8 = LD_SB(src);
+  src += src_stride;
+
+  ILVR_B4_SB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+             src32_r, src43_r);
+  ILVR_B4_SB(src5, src4, src6, src5, src7, src6, src8, src7, src54_r, src65_r,
+             src76_r, src87_r);
+  ILVR_D4_SB(src21_r, src10_r, src43_r, src32_r, src65_r, src54_r, src87_r,
+             src76_r, src2110, src4332, src6554, src8776);
+  DOTP_UB4_UH(src2110, src4332, src6554, src8776, filt0, filt0, filt0, filt0,
+              tmp0, tmp1, tmp2, tmp3);
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, src2110, src4332);
+  ST4x4_UB(src2110, src2110, 0, 1, 2, 3, dst, dst_stride);
+  ST4x4_UB(src4332, src4332, 0, 1, 2, 3, dst + 4 * dst_stride, dst_stride);
+}
+
+static void common_vt_2t_4w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_vt_2t_4x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else if (8 == height) {
+    common_vt_2t_4x8_msa(src, src_stride, dst, dst_stride, filter);
+  }
+}
+
+static void common_vt_2t_8x4_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter) {
+  v16u8 src0, src1, src2, src3, src4, vec0, vec1, vec2, vec3, filt0;
+  v16i8 out0, out1;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_UB5(src, src_stride, src0, src1, src2, src3, src4);
+  ILVR_B2_UB(src1, src0, src2, src1, vec0, vec1);
+  ILVR_B2_UB(src3, src2, src4, src3, vec2, vec3);
+  DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1,
+              tmp2, tmp3);
+  SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+  PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+  ST8x4_UB(out0, out1, dst, dst_stride);
+}
+
+static void common_vt_2t_8x8mult_msa(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v16i8 out0, out1;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 3); loop_cnt--;) {
+    LD_UB8(src, src_stride, src1, src2, src3, src4, src5, src6, src7, src8);
+    src += (8 * src_stride);
+
+    ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, vec0, vec1, vec2,
+               vec3);
+    ILVR_B4_UB(src5, src4, src6, src5, src7, src6, src8, src7, vec4, vec5, vec6,
+               vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1,
+                tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, tmp0, tmp1,
+                tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_SB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+    ST8x4_UB(out0, out1, dst, dst_stride);
+    dst += (4 * dst_stride);
+
+    src0 = src8;
+  }
+}
+
+static void common_vt_2t_8w_msa(const uint8_t *src, int32_t src_stride,
+                                uint8_t *dst, int32_t dst_stride,
+                                int8_t *filter, int32_t height) {
+  if (4 == height) {
+    common_vt_2t_8x4_msa(src, src_stride, dst, dst_stride, filter);
+  } else {
+    common_vt_2t_8x8mult_msa(src, src_stride, dst, dst_stride, filter, height);
+  }
+}
+
+static void common_vt_2t_16w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst);
+    dst += dst_stride;
+
+    ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst);
+    dst += dst_stride;
+
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst);
+    dst += dst_stride;
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst);
+    dst += dst_stride;
+
+    src0 = src4;
+  }
+}
+
+static void common_vt_2t_32w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  src0 = LD_UB(src);
+  src5 = LD_UB(src + 16);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+
+    LD_UB4(src + 16, src_stride, src6, src7, src8, src9);
+    src += (4 * src_stride);
+
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst);
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + dst_stride);
+
+    ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst + 2 * dst_stride);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + 3 * dst_stride);
+
+    ILVR_B2_UB(src6, src5, src7, src6, vec0, vec2);
+    ILVL_B2_UB(src6, src5, src7, src6, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst + 16);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + 16 + dst_stride);
+
+    ILVR_B2_UB(src8, src7, src9, src8, vec4, vec6);
+    ILVL_B2_UB(src8, src7, src9, src8, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst + 16 + 2 * dst_stride);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + 16 + 3 * dst_stride);
+    dst += (4 * dst_stride);
+
+    src0 = src4;
+    src5 = src9;
+  }
+}
+
+static void common_vt_2t_64w_msa(const uint8_t *src, int32_t src_stride,
+                                 uint8_t *dst, int32_t dst_stride,
+                                 int8_t *filter, int32_t height) {
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
+  v16u8 src11, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, filt0;
+  v8u16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8i16 filt;
+
+  /* rearranging filter_y */
+  filt = LD_SH(filter);
+  filt0 = (v16u8)__msa_splati_h(filt, 0);
+
+  LD_UB4(src, 16, src0, src3, src6, src9);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 1); loop_cnt--;) {
+    LD_UB2(src, src_stride, src1, src2);
+    LD_UB2(src + 16, src_stride, src4, src5);
+    LD_UB2(src + 32, src_stride, src7, src8);
+    LD_UB2(src + 48, src_stride, src10, src11);
+    src += (2 * src_stride);
+
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + dst_stride);
+
+    ILVR_B2_UB(src4, src3, src5, src4, vec4, vec6);
+    ILVL_B2_UB(src4, src3, src5, src4, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp4, tmp5);
+    SRARI_H2_UH(tmp4, tmp5, FILTER_BITS);
+    PCKEV_ST_SB(tmp4, tmp5, dst + 16);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp6, tmp7);
+    SRARI_H2_UH(tmp6, tmp7, FILTER_BITS);
+    PCKEV_ST_SB(tmp6, tmp7, dst + 16 + dst_stride);
+
+    ILVR_B2_UB(src7, src6, src8, src7, vec0, vec2);
+    ILVL_B2_UB(src7, src6, src8, src7, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    PCKEV_ST_SB(tmp0, tmp1, dst + 32);
+
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_ST_SB(tmp2, tmp3, dst + 32 + dst_stride);
+
+    ILVR_B2_UB(src10, src9, src11, src10, vec4, vec6);
+    ILVL_B2_UB(src10, src9, src11, src10, vec5, vec7);
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp4, tmp5);
+    SRARI_H2_UH(tmp4, tmp5, FILTER_BITS);
+    PCKEV_ST_SB(tmp4, tmp5, dst + 48);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp6, tmp7);
+    SRARI_H2_UH(tmp6, tmp7, FILTER_BITS);
+    PCKEV_ST_SB(tmp6, tmp7, dst + 48 + dst_stride);
+    dst += (2 * dst_stride);
+
+    src0 = src2;
+    src3 = src5;
+    src6 = src8;
+    src9 = src11;
+  }
+}
+
+void aom_convolve8_vert_msa(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int x_step_q4,
+                            const int16_t *filter_y, int y_step_q4, int w,
+                            int h) {
+  int8_t cnt, filt_ver[8];
+
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  for (cnt = 8; cnt--;) {
+    filt_ver[cnt] = filter_y[cnt];
+  }
+
+  if (((const int32_t *)filter_y)[0] == 0) {
+    switch (w) {
+      case 4:
+        common_vt_2t_4w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                            &filt_ver[3], h);
+        break;
+      case 8:
+        common_vt_2t_8w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                            &filt_ver[3], h);
+        break;
+      case 16:
+        common_vt_2t_16w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             &filt_ver[3], h);
+        break;
+      case 32:
+        common_vt_2t_32w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             &filt_ver[3], h);
+        break;
+      case 64:
+        common_vt_2t_64w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             &filt_ver[3], h);
+        break;
+      default:
+        aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                             x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  } else {
+    switch (w) {
+      case 4:
+        common_vt_8t_4w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                            filt_ver, h);
+        break;
+      case 8:
+        common_vt_8t_8w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                            filt_ver, h);
+        break;
+      case 16:
+        common_vt_8t_16w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             filt_ver, h);
+        break;
+      case 32:
+        common_vt_8t_32w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             filt_ver, h);
+        break;
+      case 64:
+        common_vt_8t_64w_msa(src, (int32_t)src_stride, dst, (int32_t)dst_stride,
+                             filt_ver, h);
+        break;
+      default:
+        aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                             x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/aom_convolve_avg_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve_avg_msa.c
new file mode 100644
index 0000000000..75f8c7ea85
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/aom_convolve_avg_msa.c
@@ -0,0 +1,233 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/macros_msa.h"
+
+static void avg_width4_msa(const uint8_t *src, int32_t src_stride, uint8_t *dst,
+                           int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  uint32_t out0, out1, out2, out3;
+  v16u8 src0, src1, src2, src3;
+  v16u8 dst0, dst1, dst2, dst3;
+
+  if (0 == (height % 4)) {
+    for (cnt = (height / 4); cnt--;) {
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      src += (4 * src_stride);
+
+      LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+
+      AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3, dst0, dst1,
+                  dst2, dst3);
+
+      out0 = __msa_copy_u_w((v4i32)dst0, 0);
+      out1 = __msa_copy_u_w((v4i32)dst1, 0);
+      out2 = __msa_copy_u_w((v4i32)dst2, 0);
+      out3 = __msa_copy_u_w((v4i32)dst3, 0);
+      SW4(out0, out1, out2, out3, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == (height % 2)) {
+    for (cnt = (height / 2); cnt--;) {
+      LD_UB2(src, src_stride, src0, src1);
+      src += (2 * src_stride);
+
+      LD_UB2(dst, dst_stride, dst0, dst1);
+
+      AVER_UB2_UB(src0, dst0, src1, dst1, dst0, dst1);
+
+      out0 = __msa_copy_u_w((v4i32)dst0, 0);
+      out1 = __msa_copy_u_w((v4i32)dst1, 0);
+      SW(out0, dst);
+      dst += dst_stride;
+      SW(out1, dst);
+      dst += dst_stride;
+    }
+  }
+}
+
+static void avg_width8_msa(const uint8_t *src, int32_t src_stride, uint8_t *dst,
+                           int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  uint64_t out0, out1, out2, out3;
+  v16u8 src0, src1, src2, src3;
+  v16u8 dst0, dst1, dst2, dst3;
+
+  for (cnt = (height / 4); cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+
+    AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3, dst0, dst1,
+                dst2, dst3);
+
+    out0 = __msa_copy_u_d((v2i64)dst0, 0);
+    out1 = __msa_copy_u_d((v2i64)dst1, 0);
+    out2 = __msa_copy_u_d((v2i64)dst2, 0);
+    out3 = __msa_copy_u_d((v2i64)dst3, 0);
+    SD4(out0, out1, out2, out3, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void avg_width16_msa(const uint8_t *src, int32_t src_stride,
+                            uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+
+  for (cnt = (height / 8); cnt--;) {
+    LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    src += (8 * src_stride);
+    LD_UB8(dst, dst_stride, dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7);
+
+    AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3, dst0, dst1,
+                dst2, dst3);
+    AVER_UB4_UB(src4, dst4, src5, dst5, src6, dst6, src7, dst7, dst4, dst5,
+                dst6, dst7);
+    ST_UB8(dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7, dst, dst_stride);
+    dst += (8 * dst_stride);
+  }
+}
+
+static void avg_width32_msa(const uint8_t *src, int32_t src_stride,
+                            uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  uint8_t *dst_dup = dst;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 src8, src9, src10, src11, src12, src13, src14, src15;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16u8 dst8, dst9, dst10, dst11, dst12, dst13, dst14, dst15;
+
+  for (cnt = (height / 8); cnt--;) {
+    LD_UB4(src, src_stride, src0, src2, src4, src6);
+    LD_UB4(src + 16, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(dst_dup, dst_stride, dst0, dst2, dst4, dst6);
+    LD_UB4(dst_dup + 16, dst_stride, dst1, dst3, dst5, dst7);
+    dst_dup += (4 * dst_stride);
+    LD_UB4(src, src_stride, src8, src10, src12, src14);
+    LD_UB4(src + 16, src_stride, src9, src11, src13, src15);
+    src += (4 * src_stride);
+    LD_UB4(dst_dup, dst_stride, dst8, dst10, dst12, dst14);
+    LD_UB4(dst_dup + 16, dst_stride, dst9, dst11, dst13, dst15);
+    dst_dup += (4 * dst_stride);
+
+    AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3, dst0, dst1,
+                dst2, dst3);
+    AVER_UB4_UB(src4, dst4, src5, dst5, src6, dst6, src7, dst7, dst4, dst5,
+                dst6, dst7);
+    AVER_UB4_UB(src8, dst8, src9, dst9, src10, dst10, src11, dst11, dst8, dst9,
+                dst10, dst11);
+    AVER_UB4_UB(src12, dst12, src13, dst13, src14, dst14, src15, dst15, dst12,
+                dst13, dst14, dst15);
+
+    ST_UB4(dst0, dst2, dst4, dst6, dst, dst_stride);
+    ST_UB4(dst1, dst3, dst5, dst7, dst + 16, dst_stride);
+    dst += (4 * dst_stride);
+    ST_UB4(dst8, dst10, dst12, dst14, dst, dst_stride);
+    ST_UB4(dst9, dst11, dst13, dst15, dst + 16, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void avg_width64_msa(const uint8_t *src, int32_t src_stride,
+                            uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  uint8_t *dst_dup = dst;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 src8, src9, src10, src11, src12, src13, src14, src15;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16u8 dst8, dst9, dst10, dst11, dst12, dst13, dst14, dst15;
+
+  for (cnt = (height / 4); cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(src, 16, src4, src5, src6, src7);
+    src += src_stride;
+    LD_UB4(src, 16, src8, src9, src10, src11);
+    src += src_stride;
+    LD_UB4(src, 16, src12, src13, src14, src15);
+    src += src_stride;
+
+    LD_UB4(dst_dup, 16, dst0, dst1, dst2, dst3);
+    dst_dup += dst_stride;
+    LD_UB4(dst_dup, 16, dst4, dst5, dst6, dst7);
+    dst_dup += dst_stride;
+    LD_UB4(dst_dup, 16, dst8, dst9, dst10, dst11);
+    dst_dup += dst_stride;
+    LD_UB4(dst_dup, 16, dst12, dst13, dst14, dst15);
+    dst_dup += dst_stride;
+
+    AVER_UB4_UB(src0, dst0, src1, dst1, src2, dst2, src3, dst3, dst0, dst1,
+                dst2, dst3);
+    AVER_UB4_UB(src4, dst4, src5, dst5, src6, dst6, src7, dst7, dst4, dst5,
+                dst6, dst7);
+    AVER_UB4_UB(src8, dst8, src9, dst9, src10, dst10, src11, dst11, dst8, dst9,
+                dst10, dst11);
+    AVER_UB4_UB(src12, dst12, src13, dst13, src14, dst14, src15, dst15, dst12,
+                dst13, dst14, dst15);
+
+    ST_UB4(dst0, dst1, dst2, dst3, dst, 16);
+    dst += dst_stride;
+    ST_UB4(dst4, dst5, dst6, dst7, dst, 16);
+    dst += dst_stride;
+    ST_UB4(dst8, dst9, dst10, dst11, dst, 16);
+    dst += dst_stride;
+    ST_UB4(dst12, dst13, dst14, dst15, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+void aom_convolve_avg_msa(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const int16_t *filter_x, int32_t filter_x_stride,
+                          const int16_t *filter_y, int32_t filter_y_stride,
+                          int32_t w, int32_t h) {
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+
+  switch (w) {
+    case 4: {
+      avg_width4_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 8: {
+      avg_width8_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 16: {
+      avg_width16_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 32: {
+      avg_width32_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 64: {
+      avg_width64_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    default: {
+      int32_t lp, cnt;
+      for (cnt = h; cnt--;) {
+        for (lp = 0; lp < w; ++lp) {
+          dst[lp] = (((dst[lp] + src[lp]) + 1) >> 1);
+        }
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/aom_convolve_copy_msa.c b/third_party/aom/aom_dsp/mips/aom_convolve_copy_msa.c
new file mode 100644
index 0000000000..f7f116f4da
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/aom_convolve_copy_msa.c
@@ -0,0 +1,248 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <string.h>
+#include "aom_dsp/mips/macros_msa.h"
+
+static void copy_width8_msa(const uint8_t *src, int32_t src_stride,
+                            uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  uint64_t out0, out1, out2, out3, out4, out5, out6, out7;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+
+  if (0 == height % 12) {
+    for (cnt = (height / 12); cnt--;) {
+      LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+      src += (8 * src_stride);
+
+      out0 = __msa_copy_u_d((v2i64)src0, 0);
+      out1 = __msa_copy_u_d((v2i64)src1, 0);
+      out2 = __msa_copy_u_d((v2i64)src2, 0);
+      out3 = __msa_copy_u_d((v2i64)src3, 0);
+      out4 = __msa_copy_u_d((v2i64)src4, 0);
+      out5 = __msa_copy_u_d((v2i64)src5, 0);
+      out6 = __msa_copy_u_d((v2i64)src6, 0);
+      out7 = __msa_copy_u_d((v2i64)src7, 0);
+
+      SD4(out0, out1, out2, out3, dst, dst_stride);
+      dst += (4 * dst_stride);
+      SD4(out4, out5, out6, out7, dst, dst_stride);
+      dst += (4 * dst_stride);
+
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      src += (4 * src_stride);
+
+      out0 = __msa_copy_u_d((v2i64)src0, 0);
+      out1 = __msa_copy_u_d((v2i64)src1, 0);
+      out2 = __msa_copy_u_d((v2i64)src2, 0);
+      out3 = __msa_copy_u_d((v2i64)src3, 0);
+      SD4(out0, out1, out2, out3, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == height % 8) {
+    for (cnt = height >> 3; cnt--;) {
+      LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+      src += (8 * src_stride);
+
+      out0 = __msa_copy_u_d((v2i64)src0, 0);
+      out1 = __msa_copy_u_d((v2i64)src1, 0);
+      out2 = __msa_copy_u_d((v2i64)src2, 0);
+      out3 = __msa_copy_u_d((v2i64)src3, 0);
+      out4 = __msa_copy_u_d((v2i64)src4, 0);
+      out5 = __msa_copy_u_d((v2i64)src5, 0);
+      out6 = __msa_copy_u_d((v2i64)src6, 0);
+      out7 = __msa_copy_u_d((v2i64)src7, 0);
+
+      SD4(out0, out1, out2, out3, dst, dst_stride);
+      dst += (4 * dst_stride);
+      SD4(out4, out5, out6, out7, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == height % 4) {
+    for (cnt = (height / 4); cnt--;) {
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      src += (4 * src_stride);
+      out0 = __msa_copy_u_d((v2i64)src0, 0);
+      out1 = __msa_copy_u_d((v2i64)src1, 0);
+      out2 = __msa_copy_u_d((v2i64)src2, 0);
+      out3 = __msa_copy_u_d((v2i64)src3, 0);
+
+      SD4(out0, out1, out2, out3, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == height % 2) {
+    for (cnt = (height / 2); cnt--;) {
+      LD_UB2(src, src_stride, src0, src1);
+      src += (2 * src_stride);
+      out0 = __msa_copy_u_d((v2i64)src0, 0);
+      out1 = __msa_copy_u_d((v2i64)src1, 0);
+
+      SD(out0, dst);
+      dst += dst_stride;
+      SD(out1, dst);
+      dst += dst_stride;
+    }
+  }
+}
+
+static void copy_16multx8mult_msa(const uint8_t *src, int32_t src_stride,
+                                  uint8_t *dst, int32_t dst_stride,
+                                  int32_t height, int32_t width) {
+  int32_t cnt, loop_cnt;
+  const uint8_t *src_tmp;
+  uint8_t *dst_tmp;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+
+  for (cnt = (width >> 4); cnt--;) {
+    src_tmp = src;
+    dst_tmp = dst;
+
+    for (loop_cnt = (height >> 3); loop_cnt--;) {
+      LD_UB8(src_tmp, src_stride, src0, src1, src2, src3, src4, src5, src6,
+             src7);
+      src_tmp += (8 * src_stride);
+
+      ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst_tmp,
+             dst_stride);
+      dst_tmp += (8 * dst_stride);
+    }
+
+    src += 16;
+    dst += 16;
+  }
+}
+
+static void copy_width16_msa(const uint8_t *src, int32_t src_stride,
+                             uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+
+  if (0 == height % 12) {
+    for (cnt = (height / 12); cnt--;) {
+      LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+      src += (8 * src_stride);
+      ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, dst, dst_stride);
+      dst += (8 * dst_stride);
+
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      src += (4 * src_stride);
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == height % 8) {
+    copy_16multx8mult_msa(src, src_stride, dst, dst_stride, height, 16);
+  } else if (0 == height % 4) {
+    for (cnt = (height >> 2); cnt--;) {
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      src += (4 * src_stride);
+
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  }
+}
+
+static void copy_width32_msa(const uint8_t *src, int32_t src_stride,
+                             uint8_t *dst, int32_t dst_stride, int32_t height) {
+  int32_t cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+
+  if (0 == height % 12) {
+    for (cnt = (height / 12); cnt--;) {
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      LD_UB4(src + 16, src_stride, src4, src5, src6, src7);
+      src += (4 * src_stride);
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride);
+      dst += (4 * dst_stride);
+
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      LD_UB4(src + 16, src_stride, src4, src5, src6, src7);
+      src += (4 * src_stride);
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride);
+      dst += (4 * dst_stride);
+
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      LD_UB4(src + 16, src_stride, src4, src5, src6, src7);
+      src += (4 * src_stride);
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  } else if (0 == height % 8) {
+    copy_16multx8mult_msa(src, src_stride, dst, dst_stride, height, 32);
+  } else if (0 == height % 4) {
+    for (cnt = (height >> 2); cnt--;) {
+      LD_UB4(src, src_stride, src0, src1, src2, src3);
+      LD_UB4(src + 16, src_stride, src4, src5, src6, src7);
+      src += (4 * src_stride);
+      ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+      ST_UB4(src4, src5, src6, src7, dst + 16, dst_stride);
+      dst += (4 * dst_stride);
+    }
+  }
+}
+
+static void copy_width64_msa(const uint8_t *src, int32_t src_stride,
+                             uint8_t *dst, int32_t dst_stride, int32_t height) {
+  copy_16multx8mult_msa(src, src_stride, dst, dst_stride, height, 64);
+}
+
+void aom_convolve_copy_msa(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride,
+                           const int16_t *filter_x, int32_t filter_x_stride,
+                           const int16_t *filter_y, int32_t filter_y_stride,
+                           int32_t w, int32_t h) {
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+
+  switch (w) {
+    case 4: {
+      uint32_t cnt, tmp;
+      /* 1 word storage */
+      for (cnt = h; cnt--;) {
+        tmp = LW(src);
+        SW(tmp, dst);
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    }
+    case 8: {
+      copy_width8_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 16: {
+      copy_width16_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 32: {
+      copy_width32_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    case 64: {
+      copy_width64_msa(src, src_stride, dst, dst_stride, h);
+      break;
+    }
+    default: {
+      uint32_t cnt;
+      for (cnt = h; cnt--;) {
+        memcpy(dst, src, w);
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/aom_convolve_msa.h b/third_party/aom/aom_dsp/mips/aom_convolve_msa.h
new file mode 100644
index 0000000000..1a0ae4d8d4
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/aom_convolve_msa.h
@@ -0,0 +1,124 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_MIPS_AOM_CONVOLVE_MSA_H_
+#define AOM_DSP_MIPS_AOM_CONVOLVE_MSA_H_
+
+#include "aom_dsp/mips/macros_msa.h"
+#include "aom_dsp/aom_filter.h"
+
+extern const uint8_t mc_filt_mask_arr[16 * 3];
+
+#define FILT_8TAP_DPADD_S_H(vec0, vec1, vec2, vec3, filt0, filt1, filt2,   \
+                            filt3)                                         \
+  ({                                                                       \
+    v8i16 tmp_dpadd_0, tmp_dpadd_1;                                        \
+                                                                           \
+    tmp_dpadd_0 = __msa_dotp_s_h((v16i8)vec0, (v16i8)filt0);               \
+    tmp_dpadd_0 = __msa_dpadd_s_h(tmp_dpadd_0, (v16i8)vec1, (v16i8)filt1); \
+    tmp_dpadd_1 = __msa_dotp_s_h((v16i8)vec2, (v16i8)filt2);               \
+    tmp_dpadd_1 = __msa_dpadd_s_h(tmp_dpadd_1, (v16i8)vec3, (v16i8)filt3); \
+    tmp_dpadd_0 = __msa_adds_s_h(tmp_dpadd_0, tmp_dpadd_1);                \
+                                                                           \
+    tmp_dpadd_0;                                                           \
+  })
+
+#define HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_h0,       \
+                        filt_h1, filt_h2, filt_h3)                             \
+  ({                                                                           \
+    v16i8 vec0_m, vec1_m, vec2_m, vec3_m;                                      \
+    v8i16 hz_out_m;                                                            \
+                                                                               \
+    VSHF_B4_SB(src0, src1, mask0, mask1, mask2, mask3, vec0_m, vec1_m, vec2_m, \
+               vec3_m);                                                        \
+    hz_out_m = FILT_8TAP_DPADD_S_H(vec0_m, vec1_m, vec2_m, vec3_m, filt_h0,    \
+                                   filt_h1, filt_h2, filt_h3);                 \
+                                                                               \
+    hz_out_m = __msa_srari_h(hz_out_m, FILTER_BITS);                           \
+    hz_out_m = __msa_sat_s_h(hz_out_m, 7);                                     \
+                                                                               \
+    hz_out_m;                                                                  \
+  })
+
+#define HORIZ_8TAP_4WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,     \
+                                   mask2, mask3, filt0, filt1, filt2, filt3, \
+                                   out0, out1)                               \
+  {                                                                          \
+    v16i8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;    \
+    v8i16 res0_m, res1_m, res2_m, res3_m;                                    \
+                                                                             \
+    VSHF_B2_SB(src0, src1, src2, src3, mask0, mask0, vec0_m, vec1_m);        \
+    DOTP_SB2_SH(vec0_m, vec1_m, filt0, filt0, res0_m, res1_m);               \
+    VSHF_B2_SB(src0, src1, src2, src3, mask1, mask1, vec2_m, vec3_m);        \
+    DPADD_SB2_SH(vec2_m, vec3_m, filt1, filt1, res0_m, res1_m);              \
+    VSHF_B2_SB(src0, src1, src2, src3, mask2, mask2, vec4_m, vec5_m);        \
+    DOTP_SB2_SH(vec4_m, vec5_m, filt2, filt2, res2_m, res3_m);               \
+    VSHF_B2_SB(src0, src1, src2, src3, mask3, mask3, vec6_m, vec7_m);        \
+    DPADD_SB2_SH(vec6_m, vec7_m, filt3, filt3, res2_m, res3_m);              \
+    ADDS_SH2_SH(res0_m, res2_m, res1_m, res3_m, out0, out1);                 \
+  }
+
+#define HORIZ_8TAP_8WID_4VECS_FILT(src0, src1, src2, src3, mask0, mask1,     \
+                                   mask2, mask3, filt0, filt1, filt2, filt3, \
+                                   out0, out1, out2, out3)                   \
+  {                                                                          \
+    v16i8 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;    \
+    v8i16 res0_m, res1_m, res2_m, res3_m, res4_m, res5_m, res6_m, res7_m;    \
+                                                                             \
+    VSHF_B2_SB(src0, src0, src1, src1, mask0, mask0, vec0_m, vec1_m);        \
+    VSHF_B2_SB(src2, src2, src3, src3, mask0, mask0, vec2_m, vec3_m);        \
+    DOTP_SB4_SH(vec0_m, vec1_m, vec2_m, vec3_m, filt0, filt0, filt0, filt0,  \
+                res0_m, res1_m, res2_m, res3_m);                             \
+    VSHF_B2_SB(src0, src0, src1, src1, mask2, mask2, vec0_m, vec1_m);        \
+    VSHF_B2_SB(src2, src2, src3, src3, mask2, mask2, vec2_m, vec3_m);        \
+    DOTP_SB4_SH(vec0_m, vec1_m, vec2_m, vec3_m, filt2, filt2, filt2, filt2,  \
+                res4_m, res5_m, res6_m, res7_m);                             \
+    VSHF_B2_SB(src0, src0, src1, src1, mask1, mask1, vec4_m, vec5_m);        \
+    VSHF_B2_SB(src2, src2, src3, src3, mask1, mask1, vec6_m, vec7_m);        \
+    DPADD_SB4_SH(vec4_m, vec5_m, vec6_m, vec7_m, filt1, filt1, filt1, filt1, \
+                 res0_m, res1_m, res2_m, res3_m);                            \
+    VSHF_B2_SB(src0, src0, src1, src1, mask3, mask3, vec4_m, vec5_m);        \
+    VSHF_B2_SB(src2, src2, src3, src3, mask3, mask3, vec6_m, vec7_m);        \
+    DPADD_SB4_SH(vec4_m, vec5_m, vec6_m, vec7_m, filt3, filt3, filt3, filt3, \
+                 res4_m, res5_m, res6_m, res7_m);                            \
+    ADDS_SH4_SH(res0_m, res4_m, res1_m, res5_m, res2_m, res6_m, res3_m,      \
+                res7_m, out0, out1, out2, out3);                             \
+  }
+
+#define PCKEV_XORI128_AVG_ST_UB(in0, in1, dst, pdst) \
+  {                                                  \
+    v16u8 tmp_m;                                     \
+                                                     \
+    tmp_m = PCKEV_XORI128_UB(in1, in0);              \
+    tmp_m = __msa_aver_u_b(tmp_m, (v16u8)dst);       \
+    ST_UB(tmp_m, (pdst));                            \
+  }
+
+#define PCKEV_AVG_ST_UB(in0, in1, dst, pdst)              \
+  {                                                       \
+    v16u8 tmp_m;                                          \
+                                                          \
+    tmp_m = (v16u8)__msa_pckev_b((v16i8)in0, (v16i8)in1); \
+    tmp_m = __msa_aver_u_b(tmp_m, (v16u8)dst);            \
+    ST_UB(tmp_m, (pdst));                                 \
+  }
+
+#define PCKEV_AVG_ST8x4_UB(in1, dst0, in2, dst1, in3, dst2, in4, dst3, pdst, \
+                           stride)                                           \
+  {                                                                          \
+    v16u8 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                    \
+                                                                             \
+    PCKEV_B2_UB(in2, in1, in4, in3, tmp0_m, tmp1_m);                         \
+    PCKEV_D2_UB(dst1, dst0, dst3, dst2, tmp2_m, tmp3_m);                     \
+    AVER_UB2_UB(tmp0_m, tmp2_m, tmp1_m, tmp3_m, tmp0_m, tmp1_m);             \
+    ST8x4_UB(tmp0_m, tmp1_m, pdst, stride);                                  \
+  }
+#endif /* AOM_DSP_MIPS_AOM_CONVOLVE_MSA_H_ */
diff --git a/third_party/aom/aom_dsp/mips/avg_msa.c b/third_party/aom/aom_dsp/mips/avg_msa.c
new file mode 100644
index 0000000000..0e17281553
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/avg_msa.c
@@ -0,0 +1,57 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/macros_msa.h"
+
+uint32_t aom_avg_8x8_msa(const uint8_t *src, int32_t src_stride) {
+  uint32_t sum_out;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v8u16 sum0, sum1, sum2, sum3, sum4, sum5, sum6, sum7;
+  v4u32 sum = { 0 };
+
+  LD_UB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+  HADD_UB4_UH(src0, src1, src2, src3, sum0, sum1, sum2, sum3);
+  HADD_UB4_UH(src4, src5, src6, src7, sum4, sum5, sum6, sum7);
+  ADD4(sum0, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum0, sum2, sum4, sum6);
+  ADD2(sum0, sum2, sum4, sum6, sum0, sum4);
+  sum0 += sum4;
+
+  sum = __msa_hadd_u_w(sum0, sum0);
+  sum0 = (v8u16)__msa_pckev_h((v8i16)sum, (v8i16)sum);
+  sum = __msa_hadd_u_w(sum0, sum0);
+  sum = (v4u32)__msa_srari_w((v4i32)sum, 6);
+  sum_out = __msa_copy_u_w((v4i32)sum, 0);
+
+  return sum_out;
+}
+
+uint32_t aom_avg_4x4_msa(const uint8_t *src, int32_t src_stride) {
+  uint32_t sum_out;
+  uint32_t src0, src1, src2, src3;
+  v16u8 vec = { 0 };
+  v8u16 sum0;
+  v4u32 sum1;
+  v2u64 sum2;
+
+  LW4(src, src_stride, src0, src1, src2, src3);
+  INSERT_W4_UB(src0, src1, src2, src3, vec);
+
+  sum0 = __msa_hadd_u_h(vec, vec);
+  sum1 = __msa_hadd_u_w(sum0, sum0);
+  sum0 = (v8u16)__msa_pckev_h((v8i16)sum1, (v8i16)sum1);
+  sum1 = __msa_hadd_u_w(sum0, sum0);
+  sum2 = __msa_hadd_u_d(sum1, sum1);
+  sum1 = (v4u32)__msa_srari_w((v4i32)sum2, 4);
+  sum_out = __msa_copy_u_w((v4i32)sum1, 0);
+
+  return sum_out;
+}
diff --git a/third_party/aom/aom_dsp/mips/common_dspr2.c b/third_party/aom/aom_dsp/mips/common_dspr2.c
new file mode 100644
index 0000000000..00ab75dc31
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/common_dspr2.c
@@ -0,0 +1,31 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/common_dspr2.h"
+
+#if HAVE_DSPR2
+uint8_t aom_ff_cropTbl_a[256 + 2 * CROP_WIDTH];
+uint8_t *aom_ff_cropTbl;
+
+void aom_dsputil_static_init(void) {
+  int i;
+
+  for (i = 0; i < 256; i++) aom_ff_cropTbl_a[i + CROP_WIDTH] = i;
+
+  for (i = 0; i < CROP_WIDTH; i++) {
+    aom_ff_cropTbl_a[i] = 0;
+    aom_ff_cropTbl_a[i + CROP_WIDTH + 256] = 255;
+  }
+
+  aom_ff_cropTbl = &aom_ff_cropTbl_a[CROP_WIDTH];
+}
+
+#endif
diff --git a/third_party/aom/aom_dsp/mips/common_dspr2.h b/third_party/aom/aom_dsp/mips/common_dspr2.h
new file mode 100644
index 0000000000..31159fdcdc
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/common_dspr2.h
@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_COMMON_MIPS_DSPR2_H_
+#define AOM_COMMON_MIPS_DSPR2_H_
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+#if HAVE_DSPR2
+#define CROP_WIDTH 512
+
+extern uint8_t *aom_ff_cropTbl;  // From "aom_dsp/mips/intrapred4_dspr2.c"
+
+static INLINE void prefetch_load(const unsigned char *src) {
+  __asm__ __volatile__("pref   0,  0(%[src])   \n\t" : : [src] "r"(src));
+}
+
+/* prefetch data for store */
+static INLINE void prefetch_store(unsigned char *dst) {
+  __asm__ __volatile__("pref   1,  0(%[dst])   \n\t" : : [dst] "r"(dst));
+}
+
+static INLINE void prefetch_load_streamed(const unsigned char *src) {
+  __asm__ __volatile__("pref   4,  0(%[src])   \n\t" : : [src] "r"(src));
+}
+
+/* prefetch data for store */
+static INLINE void prefetch_store_streamed(unsigned char *dst) {
+  __asm__ __volatile__("pref   5,  0(%[dst])   \n\t" : : [dst] "r"(dst));
+}
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_COMMON_MIPS_DSPR2_H_
diff --git a/third_party/aom/aom_dsp/mips/convolve2_avg_dspr2.c b/third_party/aom/aom_dsp/mips/convolve2_avg_dspr2.c
new file mode 100644
index 0000000000..d557115b92
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve2_avg_dspr2.c
@@ -0,0 +1,256 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/convolve_common_dspr2.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_bi_avg_vert_4_dspr2(const uint8_t *src, int32_t src_stride,
+                                         uint8_t *dst, int32_t dst_stride,
+                                         const int16_t *filter_y, int32_t w,
+                                         int32_t h) {
+  int32_t x, y;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  uint32_t load1, load2;
+  uint32_t p1, p2;
+  uint32_t scratch1, scratch2;
+  uint32_t store1, store2;
+  int32_t Temp1, Temp2;
+  const int16_t *filter = &filter_y[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+
+    for (x = 0; x < w; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__(
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]     \n\t"
+
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "lbu              %[scratch1],  0(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 1 */
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 2 */
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+          "lbu              %[scratch1],  2(%[dst_ptr])                   \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  3(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 3 */
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 4 */
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r"(load1), [load2] "=&r"(load2), [p1] "=&r"(p1),
+            [p2] "=&r"(p2), [scratch1] "=&r"(scratch1),
+            [scratch2] "=&r"(scratch2), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [store1] "=&r"(store1),
+            [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr)
+          : [filter45] "r"(filter45), [vector4a] "r"(vector4a),
+            [src_stride] "r"(src_stride), [cm] "r"(cm), [dst_ptr] "r"(dst_ptr));
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_avg_vert_64_dspr2(const uint8_t *src,
+                                          int32_t src_stride, uint8_t *dst,
+                                          int32_t dst_stride,
+                                          const int16_t *filter_y, int32_t h) {
+  int32_t x, y;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  uint32_t load1, load2;
+  uint32_t p1, p2;
+  uint32_t scratch1, scratch2;
+  uint32_t store1, store2;
+  int32_t Temp1, Temp2;
+  const int16_t *filter = &filter_y[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+    prefetch_store(dst + dst_stride + 32);
+
+    for (x = 0; x < 64; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__(
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]     \n\t"
+
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "lbu              %[scratch1],  0(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 1 */
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 2 */
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+          "lbu              %[scratch1],  2(%[dst_ptr])                   \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  3(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 3 */
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 4 */
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r"(load1), [load2] "=&r"(load2), [p1] "=&r"(p1),
+            [p2] "=&r"(p2), [scratch1] "=&r"(scratch1),
+            [scratch2] "=&r"(scratch2), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [store1] "=&r"(store1),
+            [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr)
+          : [filter45] "r"(filter45), [vector4a] "r"(vector4a),
+            [src_stride] "r"(src_stride), [cm] "r"(cm), [dst_ptr] "r"(dst_ptr));
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void aom_convolve2_avg_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4, int w,
+                                  int h) {
+  uint32_t pos = 38;
+
+  assert(y_step_q4 == 16);
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                       :
+                       : [pos] "r"(pos));
+
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4:
+    case 8:
+    case 16:
+    case 32:
+      convolve_bi_avg_vert_4_dspr2(src, src_stride, dst, dst_stride, filter_y,
+                                   w, h);
+      break;
+    case 64:
+      prefetch_store(dst + 32);
+      convolve_bi_avg_vert_64_dspr2(src, src_stride, dst, dst_stride, filter_y,
+                                    h);
+      break;
+    default:
+      aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                               x_step_q4, filter_y, y_step_q4, w, h);
+      break;
+  }
+}
+#endif
diff --git a/third_party/aom/aom_dsp/mips/convolve2_avg_horiz_dspr2.c b/third_party/aom/aom_dsp/mips/convolve2_avg_horiz_dspr2.c
new file mode 100644
index 0000000000..efbdcf60fb
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve2_avg_horiz_dspr2.c
@@ -0,0 +1,802 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/convolve_common_dspr2.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_bi_avg_horiz_4_dspr2(const uint8_t *src,
+                                          int32_t src_stride, uint8_t *dst,
+                                          int32_t dst_stride,
+                                          const int16_t *filter_x0, int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  int32_t Temp1, Temp2, Temp3, Temp4;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3;
+  uint32_t tn1, tn2;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        "lbu              %[p2],          3(%[dst])                      \n\t" /* load odd 2 */
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],         %[Temp1](%[cm])                \n\t" /* even 1 */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "lbu              %[Temp1],       1(%[dst])                      \n\t" /* load odd 1 */
+        "preceu.ph.qbr    %[p1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p3],          %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tn2],         0(%[dst])                      \n\t" /* load even 1 */
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],         %[Temp3](%[cm])                \n\t" /* even 2 */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "lbux             %[tn1],         %[Temp2](%[cm])                \n\t" /* odd 1 */
+        "addqh_r.w        %[tn2],         %[tn2],         %[tp1]         \n\t" /* average even 1 */
+        "dpa.w.ph         $ac2,           %[p3],          %[filter45]    \n\t"
+        "extp             %[Temp4],       $ac2,           31             \n\t"
+
+        "lbu              %[tp1],         2(%[dst])                      \n\t" /* load even 2 */
+        "sb               %[tn2],         0(%[dst])                      \n\t" /* store even 1 */
+
+        /* clamp */
+        "addqh_r.w        %[Temp1],       %[Temp1],       %[tn1]         \n\t" /* average odd 1 */
+        "lbux             %[p3],          %[Temp4](%[cm])                \n\t" /* odd 2 */
+        "sb               %[Temp1],       1(%[dst])                      \n\t" /* store odd 1 */
+
+        "addqh_r.w        %[tp1],         %[tp1],         %[tp2]         \n\t" /* average even 2 */
+        "sb               %[tp1],         2(%[dst])                      \n\t" /* store even 2 */
+
+        "addqh_r.w        %[p2],          %[p2],          %[p3]          \n\t" /* average odd 2 */
+        "sb               %[p2],          3(%[dst])                      \n\t" /* store odd 2 */
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tn1] "=&r"(tn1),
+          [tn2] "=&r"(tn2), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3),
+          [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3),
+          [Temp4] "=&r"(Temp4)
+        : [filter45] "r"(filter45), [vector4a] "r"(vector4a), [cm] "r"(cm),
+          [dst] "r"(dst), [src] "r"(src));
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_avg_horiz_8_dspr2(const uint8_t *src,
+                                          int32_t src_stride, uint8_t *dst,
+                                          int32_t dst_stride,
+                                          const int16_t *filter_x0, int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2, tp3, tp4;
+  uint32_t p1, p2, p3, p4, n1;
+  uint32_t st0, st1;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp2]                         \n\t"
+        "ulw              %[tp3],         8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+        "lbu              %[Temp2],       0(%[dst])                      \n\t"
+        "lbu              %[tp4],         2(%[dst])                      \n\t"
+
+        /* even 2. pixel */
+        "dpa.w.ph         $ac2,           %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[st0],         %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac1                           \n\t"
+        "mthi             $zero,          $ac1                           \n\t"
+        "lbux             %[st1],         %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac1,           31             \n\t"
+
+        "addqh_r.w        %[Temp2],       %[Temp2],       %[st0]         \n\t"
+        "addqh_r.w        %[tp4],         %[tp4],         %[st1]         \n\t"
+        "sb               %[Temp2],       0(%[dst])                      \n\t"
+        "sb               %[tp4],         2(%[dst])                      \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+
+        "balign           %[tp3],         %[tp2],         3              \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+
+        "lbux             %[st0],         %[Temp1](%[cm])                \n\t"
+        "lbu              %[Temp2],       4(%[dst])                      \n\t"
+        "addqh_r.w        %[Temp2],       %[Temp2],       %[st0]         \n\t"
+
+        "dpa.w.ph         $ac2,           %[p4],          %[filter45]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a],    $ac1                           \n\t"
+        "mthi             $zero,          $ac1                           \n\t"
+        "sb               %[Temp2],       4(%[dst])                      \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp3]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp3]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tp1],         6(%[dst])                      \n\t"
+
+        /* odd 2. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "lbux             %[st0],         %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,           %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],       $ac1,           31             \n\t"
+
+        "lbu              %[tp2],         1(%[dst])                      \n\t"
+        "lbu              %[tp3],         3(%[dst])                      \n\t"
+        "addqh_r.w        %[tp1],         %[tp1],         %[st0]         \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[st1],         %[Temp2](%[cm])                \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[filter45]    \n\t"
+        "addqh_r.w        %[tp2],         %[tp2],         %[st1]         \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tp4],         5(%[dst])                      \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[tp2],         1(%[dst])                      \n\t"
+        "sb               %[tp1],         6(%[dst])                      \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac2,           31             \n\t"
+
+        "lbu              %[tp1],         7(%[dst])                      \n\t"
+
+        /* clamp */
+        "lbux             %[p4],          %[Temp3](%[cm])                \n\t"
+        "addqh_r.w        %[tp3],         %[tp3],         %[p4]          \n\t"
+
+        "lbux             %[p2],          %[Temp2](%[cm])                \n\t"
+        "addqh_r.w        %[tp4],         %[tp4],         %[p2]          \n\t"
+
+        "lbux             %[p1],          %[Temp1](%[cm])                \n\t"
+        "addqh_r.w        %[tp1],         %[tp1],         %[p1]          \n\t"
+
+        /* store bytes */
+        "sb               %[tp3],         3(%[dst])                      \n\t"
+        "sb               %[tp4],         5(%[dst])                      \n\t"
+        "sb               %[tp1],         7(%[dst])                      \n\t"
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3),
+          [tp4] "=&r"(tp4), [st0] "=&r"(st0), [st1] "=&r"(st1), [p1] "=&r"(p1),
+          [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4), [n1] "=&r"(n1),
+          [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3)
+        : [filter45] "r"(filter45), [vector4a] "r"(vector4a), [cm] "r"(cm),
+          [dst] "r"(dst), [src] "r"(src));
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_avg_horiz_16_dspr2(const uint8_t *src_ptr,
+                                           int32_t src_stride, uint8_t *dst_ptr,
+                                           int32_t dst_stride,
+                                           const int16_t *filter_x0, int32_t h,
+                                           int32_t count) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_store(dst_ptr + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+          "lbu              %[st2],       0(%[dst])                    \n\t" /* load even 1 from dst */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          "lbu              %[qload3],    2(%[dst])                    \n\t" /* load even 2 from dst */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st2],       0(%[dst])                    \n\t" /* store even 1 to dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter45]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st2]       \n\t" /* average even 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    2(%[dst])                    \n\t" /* store even 2 to dst */
+          "lbu              %[qload3],    4(%[dst])                    \n\t" /* load even 3 from dst */
+          "lbu              %[qload1],    6(%[dst])                    \n\t" /* load even 4 from dst */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter45]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 3 */
+          "sb               %[qload3],    4(%[dst])                    \n\t" /* store even 3 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average even 4 */
+          "sb               %[qload1],    6(%[dst])                    \n\t" /* store even 4 to dst */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter45]  \n\t" /* even 6 */
+          "lbu              %[qload2],    8(%[dst])                    \n\t" /* load even 5 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 5 */
+          "sb               %[qload2],    8(%[dst])                    \n\t" /* store even 5 to dst */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* even 7 */
+          "lbu              %[qload3],    10(%[dst])                   \n\t" /* load even 6 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          "lbu              %[st2],       12(%[dst])                   \n\t" /* load even 7 from dst */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* even 8 */
+          "sb               %[qload3],    10(%[dst])                   \n\t" /* store even 6 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                   \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 7 */
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       12(%[dst])                   \n\t" /* store even 7 to dst */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter45]  \n\t" /* odd 1 */
+          "lbu              %[qload2],    14(%[dst])                   \n\t" /* load even 8 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          "lbu              %[st1],       1(%[dst])                    \n\t" /* load odd 1 from dst */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 8 */
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    14(%[dst])                   \n\t" /* store even 8 to dst */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* odd 2 */
+          "lbu              %[qload3],    3(%[dst])                    \n\t" /* load odd 2 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st3],       %[st3],         %[st1]       \n\t" /* average odd 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* odd 3 */
+          "sb               %[st3],       1(%[dst])                    \n\t" /* store odd 1 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st1]       \n\t" /* average odd 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    3(%[dst])                    \n\t" /* store odd 2 to dst */
+          "lbu              %[qload1],    5(%[dst])                    \n\t" /* load odd 3 from dst */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter45]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          "lbu              %[st1],       7(%[dst])                    \n\t" /* load odd 4 from dst */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st2]       \n\t" /* average odd 3 */
+          "sb               %[qload1],    5(%[dst])                    \n\t" /* store odd 3 to dst */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          "lbu              %[qload1],    9(%[dst])                    \n\t" /* load odd 5 from dst */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st1],       %[st1],         %[st3]       \n\t" /* average odd 4 */
+          "sb               %[st1],       7(%[dst])                    \n\t" /* store odd 4 to dst */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter45]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 5 */
+          "sb               %[qload1],    9(%[dst])                    \n\t" /* store odd 5 to dst */
+          "lbu              %[qload2],    11(%[dst])                   \n\t" /* load odd 6 from dst */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          "lbu              %[qload3],    13(%[dst])                   \n\t" /* load odd 7 from dst */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter45]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbu              %[qload1],    15(%[dst])                   \n\t" /* load odd 8 from dst */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average odd 6 */
+
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average odd 7 */
+
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 8 */
+
+          "sb               %[qload2],    11(%[dst])                   \n\t" /* store odd 6 to dst */
+          "sb               %[qload3],    13(%[dst])                   \n\t" /* store odd 7 to dst */
+          "sb               %[qload1],    15(%[dst])                   \n\t" /* store odd 8 to dst */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), [st1] "=&r"(st1),
+            [st2] "=&r"(st2), [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2),
+            [p3] "=&r"(p3), [p4] "=&r"(p4), [qload3] "=&r"(qload3),
+            [p5] "=&r"(p5), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+            [Temp3] "=&r"(Temp3)
+          : [filter45] "r"(filter45), [vector_64] "r"(vector_64), [cm] "r"(cm),
+            [dst] "r"(dst), [src] "r"(src));
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void convolve_bi_avg_horiz_64_dspr2(const uint8_t *src_ptr,
+                                           int32_t src_stride, uint8_t *dst_ptr,
+                                           int32_t dst_stride,
+                                           const int16_t *filter_x0,
+                                           int32_t h) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+    prefetch_store(dst_ptr + dst_stride);
+    prefetch_store(dst_ptr + dst_stride + 32);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+          "lbu              %[st2],       0(%[dst])                    \n\t" /* load even 1 from dst */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          "lbu              %[qload3],    2(%[dst])                    \n\t" /* load even 2 from dst */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st2],       0(%[dst])                    \n\t" /* store even 1 to dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter45]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st2]       \n\t" /* average even 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    2(%[dst])                    \n\t" /* store even 2 to dst */
+          "lbu              %[qload3],    4(%[dst])                    \n\t" /* load even 3 from dst */
+          "lbu              %[qload1],    6(%[dst])                    \n\t" /* load even 4 from dst */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter45]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 3 */
+          "sb               %[qload3],    4(%[dst])                    \n\t" /* store even 3 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average even 4 */
+          "sb               %[qload1],    6(%[dst])                    \n\t" /* store even 4 to dst */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter45]  \n\t" /* even 6 */
+          "lbu              %[qload2],    8(%[dst])                    \n\t" /* load even 5 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 5 */
+          "sb               %[qload2],    8(%[dst])                    \n\t" /* store even 5 to dst */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* even 7 */
+          "lbu              %[qload3],    10(%[dst])                   \n\t" /* load even 6 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          "lbu              %[st2],       12(%[dst])                   \n\t" /* load even 7 from dst */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* even 8 */
+          "sb               %[qload3],    10(%[dst])                   \n\t" /* store even 6 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                   \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 7 */
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       12(%[dst])                   \n\t" /* store even 7 to dst */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter45]  \n\t" /* odd 1 */
+          "lbu              %[qload2],    14(%[dst])                   \n\t" /* load even 8 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          "lbu              %[st1],       1(%[dst])                    \n\t" /* load odd 1 from dst */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 8 */
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    14(%[dst])                   \n\t" /* store even 8 to dst */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* odd 2 */
+          "lbu              %[qload3],    3(%[dst])                    \n\t" /* load odd 2 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st3],       %[st3],         %[st1]       \n\t" /* average odd 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* odd 3 */
+          "sb               %[st3],       1(%[dst])                    \n\t" /* store odd 1 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st1]       \n\t" /* average odd 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    3(%[dst])                    \n\t" /* store odd 2 to dst */
+          "lbu              %[qload1],    5(%[dst])                    \n\t" /* load odd 3 from dst */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter45]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          "lbu              %[st1],       7(%[dst])                    \n\t" /* load odd 4 from dst */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st2]       \n\t" /* average odd 3 */
+          "sb               %[qload1],    5(%[dst])                    \n\t" /* store odd 3 to dst */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          "lbu              %[qload1],    9(%[dst])                    \n\t" /* load odd 5 from dst */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st1],       %[st1],         %[st3]       \n\t" /* average odd 4 */
+          "sb               %[st1],       7(%[dst])                    \n\t" /* store odd 4 to dst */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter45]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 5 */
+          "sb               %[qload1],    9(%[dst])                    \n\t" /* store odd 5 to dst */
+          "lbu              %[qload2],    11(%[dst])                   \n\t" /* load odd 6 from dst */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          "lbu              %[qload3],    13(%[dst])                   \n\t" /* load odd 7 from dst */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter45]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbu              %[qload1],    15(%[dst])                   \n\t" /* load odd 8 from dst */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average odd 6 */
+
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average odd 7 */
+
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 8 */
+
+          "sb               %[qload2],    11(%[dst])                   \n\t" /* store odd 6 to dst */
+          "sb               %[qload3],    13(%[dst])                   \n\t" /* store odd 7 to dst */
+          "sb               %[qload1],    15(%[dst])                   \n\t" /* store odd 8 to dst */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), [st1] "=&r"(st1),
+            [st2] "=&r"(st2), [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2),
+            [p3] "=&r"(p3), [p4] "=&r"(p4), [qload3] "=&r"(qload3),
+            [p5] "=&r"(p5), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+            [Temp3] "=&r"(Temp3)
+          : [filter45] "r"(filter45), [vector_64] "r"(vector_64), [cm] "r"(cm),
+            [dst] "r"(dst), [src] "r"(src));
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+void aom_convolve2_avg_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int x_step_q4,
+                                   const int16_t *filter_y, int y_step_q4,
+                                   int w, int h) {
+  uint32_t pos = 38;
+
+  assert(x_step_q4 == 16);
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                       :
+                       : [pos] "r"(pos));
+
+  /* prefetch data to cache memory */
+  prefetch_load(src);
+  prefetch_load(src + 32);
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4:
+      convolve_bi_avg_horiz_4_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                    h);
+      break;
+    case 8:
+      convolve_bi_avg_horiz_8_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                    h);
+      break;
+    case 16:
+      convolve_bi_avg_horiz_16_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                     h, 1);
+      break;
+    case 32:
+      convolve_bi_avg_horiz_16_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                     h, 2);
+      break;
+    case 64:
+      prefetch_load(src + 64);
+      prefetch_store(dst + 32);
+
+      convolve_bi_avg_horiz_64_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                     h);
+      break;
+    default:
+      aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                                x_step_q4, filter_y, y_step_q4, w, h);
+      break;
+  }
+}
+#endif
diff --git a/third_party/aom/aom_dsp/mips/convolve2_dspr2.c b/third_party/aom/aom_dsp/mips/convolve2_dspr2.c
new file mode 100644
index 0000000000..066308315d
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve2_dspr2.c
@@ -0,0 +1,1030 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/convolve_common_dspr2.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_bi_horiz_4_transposed_dspr2(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    const int16_t *filter_x0, int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint8_t *dst_ptr;
+  int32_t Temp1, Temp2;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    dst_ptr = dst;
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp2],       $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],         %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],         %[Temp2](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp2],       $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[p1],          %[Temp1](%[cm])                \n\t"
+        "lbux             %[p2],          %[Temp2](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[tp1],         0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[p1],          0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[tp2],         0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[p2],          0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [p1] "=&r"(p1), [p2] "=&r"(p2),
+          [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [dst_ptr] "+r"(dst_ptr)
+        : [filter45] "r"(filter45), [vector4a] "r"(vector4a), [cm] "r"(cm),
+          [src] "r"(src), [dst_stride] "r"(dst_stride));
+
+    /* Next row... */
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+static void convolve_bi_horiz_8_transposed_dspr2(
+    const uint8_t *src, int32_t src_stride, uint8_t *dst, int32_t dst_stride,
+    const int16_t *filter_x0, int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint8_t *dst_ptr;
+  uint32_t vector4a = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2, tp3;
+  uint32_t p1, p2, p3, p4;
+  uint8_t *odd_dst;
+  uint32_t dst_pitch_2 = (dst_stride << 1);
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+
+    dst_ptr = dst;
+    odd_dst = (dst_ptr + dst_stride);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],         0(%[src])                       \n\t"
+        "ulw              %[tp2],         4(%[src])                       \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                          \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                          \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp2]                          \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp2]                          \n\t"
+        "ulw              %[tp3],         8(%[src])                       \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]     \n\t"
+        "extp             %[Temp1],       $ac3,           31              \n\t"
+
+        /* even 2. pixel */
+        "dpa.w.ph         $ac2,           %[p2],          %[filter45]     \n\t"
+        "extp             %[Temp3],       $ac2,           31              \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[Temp2],       %[Temp1](%[cm])                 \n\t"
+        "mtlo             %[vector4a],    $ac1                            \n\t"
+        "mthi             $zero,          $ac1                            \n\t"
+        "balign           %[tp3],         %[tp2],         3              \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[filter45]     \n\t"
+        "lbux             %[tp1],         %[Temp3](%[cm])                 \n\t"
+        "extp             %[p3],          $ac1,           31              \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "sb               %[Temp2],       0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+        "sb               %[tp1],         0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+
+        "dpa.w.ph         $ac2,           %[p4],          %[filter45]     \n\t"
+        "extp             %[Temp3],       $ac2,           31              \n\t"
+
+        "lbux             %[Temp1],         %[p3](%[cm])                    "
+        "\n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a],    $ac1                            \n\t"
+        "mthi             $zero,          $ac1                            \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                          \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                          \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp3]                          \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp3]                          \n\t"
+        "sb               %[Temp1],       0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+
+        "dpa.w.ph         $ac3,           %[p1],          %[filter45]     \n\t"
+        "extp             %[Temp2],       $ac3,           31              \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp1],         %[Temp3](%[cm])                 \n\t"
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "dpa.w.ph         $ac1,           %[p2],          %[filter45]     \n\t"
+        "sb               %[tp1],         0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+        "extp             %[Temp3],       $ac1,           31              \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[tp3],         %[Temp2](%[cm])                 \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[filter45]     \n\t"
+        "extp             %[Temp2],       $ac3,           31              \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[tp3],         0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[filter45]     \n\t"
+        "extp             %[Temp1],       $ac2,           31              \n\t"
+
+        /* clamp */
+        "lbux             %[p4],          %[Temp3](%[cm])                 \n\t"
+        "lbux             %[p2],          %[Temp2](%[cm])                 \n\t"
+        "lbux             %[p1],          %[Temp1](%[cm])                 \n\t"
+
+        /* store bytes */
+        "sb               %[p4],          0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+        "sb               %[p2],          0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+        "sb               %[p1],          0(%[odd_dst])                   \n\t"
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3), [p1] "=&r"(p1),
+          [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4), [Temp1] "=&r"(Temp1),
+          [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3), [dst_ptr] "+r"(dst_ptr),
+          [odd_dst] "+r"(odd_dst)
+        : [filter45] "r"(filter45), [vector4a] "r"(vector4a), [cm] "r"(cm),
+          [src] "r"(src), [dst_pitch_2] "r"(dst_pitch_2));
+
+    /* Next row... */
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+static void convolve_bi_horiz_16_transposed_dspr2(
+    const uint8_t *src_ptr, int32_t src_stride, uint8_t *dst_ptr,
+    int32_t dst_stride, const int16_t *filter_x0, int32_t h, int32_t count) {
+  int32_t c, y;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  uint32_t dst_pitch_2 = (dst_stride << 1);
+  uint8_t *odd_dst;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+
+    src = src_ptr;
+    dst = dst_ptr;
+
+    odd_dst = (dst + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],        0(%[src])                       "
+          "\n\t"
+          "ulw              %[qload2],        4(%[src])                       "
+          "\n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 1 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 2 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       "
+          "\n\t"
+          "ulw              %[qload1],        8(%[src])                       "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter45]     "
+          "\n\t" /* even 1 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* even 3 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p5],            %[qload1]                       "
+          "\n\t"
+          "ulw              %[qload2],        12(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p2],          %[filter45]     "
+          "\n\t" /* even 1 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 1 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 4 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p2],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 1 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]   "
+          "          \n\t"
+          "dpa.w.ph         $ac3,             %[p3],          %[filter45]     "
+          "\n\t" /* even 3 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* even 3 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 5 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p3],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 2 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p4],          %[filter45]     "
+          "\n\t" /* even 4 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 4 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* even 6 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "sb               %[st3],           0(%[dst])                       "
+          "\n\t" /* even 3 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter45]     "
+          "\n\t" /* even 5 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 5 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 7 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 4 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload1],        20(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p5],          %[filter45]     "
+          "\n\t" /* even 6 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* even 6 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 8 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 5 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter45]     "
+          "\n\t" /* even 7 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 7 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 1 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter45]     "
+          "\n\t" /* even 8 */
+          "sb               %[st3],           0(%[dst])                       "
+          "\n\t" /* even 6 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 8 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],        1(%[src])                       "
+          "\n\t"
+          "ulw              %[qload2],        5(%[src])                       "
+          "\n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 2 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 7 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload2],        9(%[src])                       "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p1],          %[filter45]     "
+          "\n\t" /* odd 1 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 1 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* odd 3 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 8 */
+          "ulw              %[qload1],        13(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter45]     "
+          "\n\t" /* odd 2 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 2 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 4 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 1 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter45]     "
+          "\n\t" /* odd 3 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* odd 3 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 5 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 2 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p4],          %[filter45]     "
+          "\n\t" /* odd 4 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 4 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* odd 6 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "sb               %[st2],           0(%[odd_dst])                   "
+          "\n\t" /* odd 3 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter45]     "
+          "\n\t" /* odd 5 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 5 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 7 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 4 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload1],        21(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p5],          %[filter45]     "
+          "\n\t" /* odd 6 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* odd 6 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 8 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 5 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p2],          %[filter45]     "
+          "\n\t" /* odd 7 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter45]     "
+          "\n\t" /* odd 8 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 8 */
+
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* odd 6 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 7 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 8 */
+
+          "sb               %[st2],           0(%[odd_dst])                   "
+          "\n\t" /* odd 6 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 7 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 8 */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), [p5] "=&r"(p5),
+            [st1] "=&r"(st1), [st2] "=&r"(st2), [st3] "=&r"(st3),
+            [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4),
+            [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3),
+            [dst] "+r"(dst), [odd_dst] "+r"(odd_dst)
+          : [filter45] "r"(filter45), [vector_64] "r"(vector_64), [cm] "r"(cm),
+            [src] "r"(src), [dst_pitch_2] "r"(dst_pitch_2));
+
+      src += 16;
+      dst = (dst_ptr + ((c + 1) * 16 * dst_stride));
+      odd_dst = (dst + dst_stride);
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += 1;
+  }
+}
+
+static void convolve_bi_horiz_64_transposed_dspr2(
+    const uint8_t *src_ptr, int32_t src_stride, uint8_t *dst_ptr,
+    int32_t dst_stride, const int16_t *filter_x0, int32_t h) {
+  int32_t c, y;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  uint32_t dst_pitch_2 = (dst_stride << 1);
+  uint8_t *odd_dst;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+
+    src = src_ptr;
+    dst = dst_ptr;
+
+    odd_dst = (dst + dst_stride);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],        0(%[src])                       "
+          "\n\t"
+          "ulw              %[qload2],        4(%[src])                       "
+          "\n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 1 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 2 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       "
+          "\n\t"
+          "ulw              %[qload1],        8(%[src])                       "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter45]     "
+          "\n\t" /* even 1 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* even 3 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p5],            %[qload1]                       "
+          "\n\t"
+          "ulw              %[qload2],        12(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p2],          %[filter45]     "
+          "\n\t" /* even 1 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 1 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 4 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p2],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 1 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]   "
+          "          \n\t"
+          "dpa.w.ph         $ac3,             %[p3],          %[filter45]     "
+          "\n\t" /* even 3 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* even 3 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 5 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p3],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 2 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p4],          %[filter45]     "
+          "\n\t" /* even 4 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 4 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* even 6 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "sb               %[st3],           0(%[dst])                       "
+          "\n\t" /* even 3 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter45]     "
+          "\n\t" /* even 5 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 5 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 7 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 4 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload1],        20(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p5],          %[filter45]     "
+          "\n\t" /* even 6 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* even 6 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 8 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 5 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter45]     "
+          "\n\t" /* even 7 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 7 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 1 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter45]     "
+          "\n\t" /* even 8 */
+          "sb               %[st3],           0(%[dst])                       "
+          "\n\t" /* even 6 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 8 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],        1(%[src])                       "
+          "\n\t"
+          "ulw              %[qload2],        5(%[src])                       "
+          "\n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 2 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 7 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload2],        9(%[src])                       "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p1],          %[filter45]     "
+          "\n\t" /* odd 1 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 1 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* odd 3 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 8 */
+          "ulw              %[qload1],        13(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter45]     "
+          "\n\t" /* odd 2 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 2 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 4 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 1 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter45]     "
+          "\n\t" /* odd 3 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* odd 3 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 5 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 2 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p4],          %[filter45]     "
+          "\n\t" /* odd 4 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 4 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* odd 6 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "sb               %[st2],           0(%[odd_dst])                   "
+          "\n\t" /* odd 3 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter45]     "
+          "\n\t" /* odd 5 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 5 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 7 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 4 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload1],        21(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p5],          %[filter45]     "
+          "\n\t" /* odd 6 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* odd 6 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 8 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 5 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p2],          %[filter45]     "
+          "\n\t" /* odd 7 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter45]     "
+          "\n\t" /* odd 8 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 8 */
+
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* odd 6 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 7 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 8 */
+
+          "sb               %[st2],           0(%[odd_dst])                   "
+          "\n\t" /* odd 6 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 7 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 8 */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), [p5] "=&r"(p5),
+            [st1] "=&r"(st1), [st2] "=&r"(st2), [st3] "=&r"(st3),
+            [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4),
+            [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3),
+            [dst] "+r"(dst), [odd_dst] "+r"(odd_dst)
+          : [filter45] "r"(filter45), [vector_64] "r"(vector_64), [cm] "r"(cm),
+            [src] "r"(src), [dst_pitch_2] "r"(dst_pitch_2));
+
+      src += 16;
+      dst = (dst_ptr + ((c + 1) * 16 * dst_stride));
+      odd_dst = (dst + dst_stride);
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += 1;
+  }
+}
+
+void convolve_bi_horiz_transposed(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter, int w, int h) {
+  int x, y;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      int sum = 0;
+
+      sum += src[x] * filter[3];
+      sum += src[x + 1] * filter[4];
+
+      dst[x * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+void aom_convolve2_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const int16_t *filter, int w,
+                         int h) {
+  uint32_t pos = 38;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                       :
+                       : [pos] "r"(pos));
+
+  /* prefetch data to cache memory */
+  prefetch_load(src);
+  prefetch_load(src + 32);
+
+  switch (w) {
+    case 4:
+      convolve_bi_horiz_4_transposed_dspr2(src, src_stride, dst, dst_stride,
+                                           filter, h);
+      break;
+    case 8:
+      convolve_bi_horiz_8_transposed_dspr2(src, src_stride, dst, dst_stride,
+                                           filter, h);
+      break;
+    case 16:
+    case 32:
+      convolve_bi_horiz_16_transposed_dspr2(src, src_stride, dst, dst_stride,
+                                            filter, h, (w / 16));
+      break;
+    case 64:
+      prefetch_load(src + 32);
+      convolve_bi_horiz_64_transposed_dspr2(src, src_stride, dst, dst_stride,
+                                            filter, h);
+      break;
+    default:
+      convolve_bi_horiz_transposed(src, src_stride, dst, dst_stride, filter, w,
+                                   h);
+      break;
+  }
+}
+#endif
diff --git a/third_party/aom/aom_dsp/mips/convolve2_horiz_dspr2.c b/third_party/aom/aom_dsp/mips/convolve2_horiz_dspr2.c
new file mode 100644
index 0000000000..dc51ab1cb7
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve2_horiz_dspr2.c
@@ -0,0 +1,681 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/convolve_common_dspr2.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_bi_horiz_4_dspr2(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      const int16_t *filter_x0, int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  int32_t Temp1, Temp2, Temp3, Temp4;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],      0(%[src])                      \n\t"
+        "ulw              %[tp2],      4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],    $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "balign           %[tp2],      %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],      %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],      %[Temp3](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp4],    $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[p1],       %[Temp2](%[cm])                \n\t"
+        "lbux             %[p2],       %[Temp4](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[tp1],      0(%[dst])                      \n\t"
+        "sb               %[p1],       1(%[dst])                      \n\t"
+        "sb               %[tp2],      2(%[dst])                      \n\t"
+        "sb               %[p2],       3(%[dst])                      \n\t"
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [p1] "=&r"(p1), [p2] "=&r"(p2),
+          [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3),
+          [Temp4] "=&r"(Temp4)
+        : [filter45] "r"(filter45), [vector4a] "r"(vector4a), [cm] "r"(cm),
+          [dst] "r"(dst), [src] "r"(src));
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_horiz_8_dspr2(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      const int16_t *filter_x0, int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2, tp3;
+  uint32_t p1, p2, p3, p4;
+  uint32_t st0, st1;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],      0(%[src])                      \n\t"
+        "ulw              %[tp2],      4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],       %[tp2]                         \n\t"
+        "ulw              %[tp3],      8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp1],    $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "dpa.w.ph         $ac2,        %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[st0],      %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac1                           \n\t"
+        "mthi             $zero,       $ac1                           \n\t"
+        "dpa.w.ph         $ac1,        %[p3],          %[filter45]    \n\t"
+        "extp             %[Temp1],    $ac1,           31             \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "sb               %[st0],      0(%[dst])                      \n\t"
+        "lbux             %[st1],      %[Temp3](%[cm])                \n\t"
+
+        "balign           %[tp3],      %[tp2],         3              \n\t"
+        "balign           %[tp2],      %[tp1],         3              \n\t"
+
+        "dpa.w.ph         $ac2,        %[p4],          %[filter45]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        "lbux             %[st0],      %[Temp1](%[cm])                \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a], $ac1                           \n\t"
+        "mthi             $zero,       $ac1                           \n\t"
+        "sb               %[st1],      2(%[dst])                      \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[p3],       %[tp3]                         \n\t"
+        "preceu.ph.qbl    %[p4],       %[tp3]                         \n\t"
+        "sb               %[st0],      4(%[dst])                      \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[filter45]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "lbux             %[st0],      %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,        %[p2],          %[filter45]    \n\t"
+        "extp             %[Temp3],    $ac1,           31             \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[st1],      %[Temp2](%[cm])                \n\t"
+        "dpa.w.ph         $ac3,        %[p3],          %[filter45]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[st1],      1(%[dst])                      \n\t"
+        "sb               %[st0],      6(%[dst])                      \n\t"
+        "dpa.w.ph         $ac2,        %[p4],          %[filter45]    \n\t"
+        "extp             %[Temp1],    $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[p4],       %[Temp3](%[cm])                \n\t"
+        "lbux             %[p2],       %[Temp2](%[cm])                \n\t"
+        "lbux             %[p1],       %[Temp1](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[p4],       3(%[dst])                      \n\t"
+        "sb               %[p2],       5(%[dst])                      \n\t"
+        "sb               %[p1],       7(%[dst])                      \n\t"
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3),
+          [st0] "=&r"(st0), [st1] "=&r"(st1), [p1] "=&r"(p1), [p2] "=&r"(p2),
+          [p3] "=&r"(p3), [p4] "=&r"(p4), [Temp1] "=&r"(Temp1),
+          [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3)
+        : [filter45] "r"(filter45), [vector4a] "r"(vector4a), [cm] "r"(cm),
+          [dst] "r"(dst), [src] "r"(src));
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_horiz_16_dspr2(const uint8_t *src_ptr,
+                                       int32_t src_stride, uint8_t *dst_ptr,
+                                       int32_t dst_stride,
+                                       const int16_t *filter_x0, int32_t h,
+                                       int32_t count) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_store(dst_ptr + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st1],       0(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter45]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st2],       2(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter45]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "sb               %[st3],       4(%[dst])                    \n\t" /* even 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "sb               %[st1],       6(%[dst])                    \n\t" /* even 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter45]  \n\t" /* even 6 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "sb               %[st2],       8(%[dst])                    \n\t" /* even 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* even 7 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* even 8 */
+          "sb               %[st3],       10(%[dst])                   \n\t" /* even 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                    \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st1],       12(%[dst])                   \n\t" /* even 7 */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter45]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       14(%[dst])                   \n\t" /* even 8 */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* odd 2 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st3],       1(%[dst])                    \n\t" /* odd 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* odd 3 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st1],       3(%[dst])                    \n\t" /* odd 2 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter45]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "sb               %[st2],       5(%[dst])                    \n\t" /* odd 3 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "sb               %[st3],       7(%[dst])                    \n\t" /* odd 4 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter45]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "sb               %[st1],       9(%[dst])                    \n\t" /* odd 5 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter45]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+
+          "sb               %[st2],       11(%[dst])                   \n\t" /* odd 6 */
+          "sb               %[st3],       13(%[dst])                   \n\t" /* odd 7 */
+          "sb               %[st1],       15(%[dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2),
+            [qload3] "=&r"(qload3), [st1] "=&r"(st1), [st2] "=&r"(st2),
+            [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3),
+            [p4] "=&r"(p4), [p5] "=&r"(p5), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3)
+          : [filter45] "r"(filter45), [vector_64] "r"(vector_64), [cm] "r"(cm),
+            [dst] "r"(dst), [src] "r"(src));
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void convolve_bi_horiz_64_dspr2(const uint8_t *src_ptr,
+                                       int32_t src_stride, uint8_t *dst_ptr,
+                                       int32_t dst_stride,
+                                       const int16_t *filter_x0, int32_t h) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  const int16_t *filter = &filter_x0[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+    prefetch_store(dst_ptr + dst_stride);
+    prefetch_store(dst_ptr + dst_stride + 32);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter45]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st1],       0(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter45]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st2],       2(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter45]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "sb               %[st3],       4(%[dst])                    \n\t" /* even 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "sb               %[st1],       6(%[dst])                    \n\t" /* even 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter45]  \n\t" /* even 6 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "sb               %[st2],       8(%[dst])                    \n\t" /* even 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* even 7 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* even 8 */
+          "sb               %[st3],       10(%[dst])                   \n\t" /* even 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                    \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st1],       12(%[dst])                   \n\t" /* even 7 */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter45]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       14(%[dst])                   \n\t" /* even 8 */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]  \n\t" /* odd 2 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st3],       1(%[dst])                    \n\t" /* odd 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter45]  \n\t" /* odd 3 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st1],       3(%[dst])                    \n\t" /* odd 2 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter45]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "sb               %[st2],       5(%[dst])                    \n\t" /* odd 3 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter45]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "sb               %[st3],       7(%[dst])                    \n\t" /* odd 4 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter45]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "sb               %[st1],       9(%[dst])                    \n\t" /* odd 5 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter45]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+
+          "sb               %[st2],       11(%[dst])                   \n\t" /* odd 6 */
+          "sb               %[st3],       13(%[dst])                   \n\t" /* odd 7 */
+          "sb               %[st1],       15(%[dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2),
+            [qload3] "=&r"(qload3), [st1] "=&r"(st1), [st2] "=&r"(st2),
+            [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3),
+            [p4] "=&r"(p4), [p5] "=&r"(p5), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3)
+          : [filter45] "r"(filter45), [vector_64] "r"(vector_64), [cm] "r"(cm),
+            [dst] "r"(dst), [src] "r"(src));
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+void aom_convolve2_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4, int w,
+                               int h) {
+  uint32_t pos = 38;
+
+  assert(x_step_q4 == 16);
+
+  prefetch_load((const uint8_t *)filter_x);
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                       :
+                       : [pos] "r"(pos));
+
+  /* prefetch data to cache memory */
+  prefetch_load(src);
+  prefetch_load(src + 32);
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4:
+      convolve_bi_horiz_4_dspr2(src, (int32_t)src_stride, dst,
+                                (int32_t)dst_stride, filter_x, (int32_t)h);
+      break;
+    case 8:
+      convolve_bi_horiz_8_dspr2(src, (int32_t)src_stride, dst,
+                                (int32_t)dst_stride, filter_x, (int32_t)h);
+      break;
+    case 16:
+      convolve_bi_horiz_16_dspr2(src, (int32_t)src_stride, dst,
+                                 (int32_t)dst_stride, filter_x, (int32_t)h, 1);
+      break;
+    case 32:
+      convolve_bi_horiz_16_dspr2(src, (int32_t)src_stride, dst,
+                                 (int32_t)dst_stride, filter_x, (int32_t)h, 2);
+      break;
+    case 64:
+      prefetch_load(src + 64);
+      prefetch_store(dst + 32);
+
+      convolve_bi_horiz_64_dspr2(src, (int32_t)src_stride, dst,
+                                 (int32_t)dst_stride, filter_x, (int32_t)h);
+      break;
+    default:
+      aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                            x_step_q4, filter_y, y_step_q4, w, h);
+      break;
+  }
+}
+#endif
diff --git a/third_party/aom/aom_dsp/mips/convolve2_vert_dspr2.c b/third_party/aom/aom_dsp/mips/convolve2_vert_dspr2.c
new file mode 100644
index 0000000000..3367be01a2
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve2_vert_dspr2.c
@@ -0,0 +1,237 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/convolve_common_dspr2.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_bi_vert_4_dspr2(const uint8_t *src, int32_t src_stride,
+                                     uint8_t *dst, int32_t dst_stride,
+                                     const int16_t *filter_y, int32_t w,
+                                     int32_t h) {
+  int32_t x, y;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  uint32_t load1, load2;
+  uint32_t p1, p2;
+  uint32_t scratch1;
+  uint32_t store1, store2;
+  int32_t Temp1, Temp2;
+  const int16_t *filter = &filter_y[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+
+    for (x = 0; x < w; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__(
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]     \n\t"
+
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r"(load1), [load2] "=&r"(load2), [p1] "=&r"(p1),
+            [p2] "=&r"(p2), [scratch1] "=&r"(scratch1), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [store1] "=&r"(store1),
+            [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr)
+          : [filter45] "r"(filter45), [vector4a] "r"(vector4a),
+            [src_stride] "r"(src_stride), [cm] "r"(cm), [dst_ptr] "r"(dst_ptr));
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_bi_vert_64_dspr2(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      const int16_t *filter_y, int32_t h) {
+  int32_t x, y;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  uint32_t load1, load2;
+  uint32_t p1, p2;
+  uint32_t scratch1;
+  uint32_t store1, store2;
+  int32_t Temp1, Temp2;
+  const int16_t *filter = &filter_y[3];
+  uint32_t filter45;
+
+  filter45 = ((const int32_t *)filter)[0];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+
+    for (x = 0; x < 64; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__(
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter45]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+
+          "precrq.ph.w      %[p2],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[filter45]     \n\t"
+          "dpa.w.ph         $ac3,         %[p2],          %[filter45]     \n\t"
+
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r"(load1), [load2] "=&r"(load2), [p1] "=&r"(p1),
+            [p2] "=&r"(p2), [scratch1] "=&r"(scratch1), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [store1] "=&r"(store1),
+            [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr)
+          : [filter45] "r"(filter45), [vector4a] "r"(vector4a),
+            [src_stride] "r"(src_stride), [cm] "r"(cm), [dst_ptr] "r"(dst_ptr));
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void aom_convolve2_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4, int w,
+                              int h) {
+  uint32_t pos = 38;
+
+  assert(y_step_q4 == 16);
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                       :
+                       : [pos] "r"(pos));
+
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4:
+    case 8:
+    case 16:
+    case 32:
+      convolve_bi_vert_4_dspr2(src, src_stride, dst, dst_stride, filter_y, w,
+                               h);
+      break;
+    case 64:
+      prefetch_store(dst + 32);
+      convolve_bi_vert_64_dspr2(src, src_stride, dst, dst_stride, filter_y, h);
+      break;
+    default:
+      aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                           x_step_q4, filter_y, y_step_q4, w, h);
+      break;
+  }
+}
+#endif
diff --git a/third_party/aom/aom_dsp/mips/convolve8_avg_dspr2.c b/third_party/aom/aom_dsp/mips/convolve8_avg_dspr2.c
new file mode 100644
index 0000000000..298065adb0
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve8_avg_dspr2.c
@@ -0,0 +1,641 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/convolve_common_dspr2.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_avg_vert_4_dspr2(const uint8_t *src, int32_t src_stride,
+                                      uint8_t *dst, int32_t dst_stride,
+                                      const int16_t *filter_y, int32_t w,
+                                      int32_t h) {
+  int32_t x, y;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  uint32_t load1, load2, load3, load4;
+  uint32_t p1, p2;
+  uint32_t n1, n2;
+  uint32_t scratch1, scratch2;
+  uint32_t store1, store2;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2;
+
+  vector1b = ((const int32_t *)filter_y)[0];
+  vector2b = ((const int32_t *)filter_y)[1];
+  vector3b = ((const int32_t *)filter_y)[2];
+  vector4b = ((const int32_t *)filter_y)[3];
+
+  src -= 3 * src_stride;
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+
+    for (x = 0; x < w; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__(
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector2b]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector2b]     \n\t"
+
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "lbu              %[scratch1],  0(%[dst_ptr])                   \n\t"
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+          "lbu              %[scratch2],  1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "dpa.w.ph         $ac2,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector4b]     \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 1 */
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector4b]     \n\t"
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 2 */
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+          "lbu              %[scratch1],  2(%[dst_ptr])                   \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  3(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 3 */
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 4 */
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+            [load4] "=&r"(load4), [p1] "=&r"(p1), [p2] "=&r"(p2),
+            [n1] "=&r"(n1), [n2] "=&r"(n2), [scratch1] "=&r"(scratch1),
+            [scratch2] "=&r"(scratch2), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [store1] "=&r"(store1),
+            [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr)
+          : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
+            [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
+            [vector4a] "r"(vector4a), [src_stride] "r"(src_stride),
+            [cm] "r"(cm), [dst_ptr] "r"(dst_ptr));
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_avg_vert_64_dspr2(const uint8_t *src, int32_t src_stride,
+                                       uint8_t *dst, int32_t dst_stride,
+                                       const int16_t *filter_y, int32_t h) {
+  int32_t x, y;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  uint32_t load1, load2, load3, load4;
+  uint32_t p1, p2;
+  uint32_t n1, n2;
+  uint32_t scratch1, scratch2;
+  uint32_t store1, store2;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2;
+
+  vector1b = ((const int32_t *)filter_y)[0];
+  vector2b = ((const int32_t *)filter_y)[1];
+  vector3b = ((const int32_t *)filter_y)[2];
+  vector4b = ((const int32_t *)filter_y)[3];
+
+  src -= 3 * src_stride;
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+    prefetch_store(dst + dst_stride + 32);
+
+    for (x = 0; x < 64; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__(
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector2b]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector2b]     \n\t"
+
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "lbu              %[scratch1],  0(%[dst_ptr])                   \n\t"
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+          "lbu              %[scratch2],  1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "dpa.w.ph         $ac2,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector4b]     \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 1 */
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector4b]     \n\t"
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 2 */
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+          "lbu              %[scratch1],  2(%[dst_ptr])                   \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+          "lbu              %[scratch2],  3(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "addqh_r.w        %[store1],    %[store1],      %[scratch1]     \n\t" /* pixel 3 */
+          "addqh_r.w        %[store2],    %[store2],      %[scratch2]     \n\t" /* pixel 4 */
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+            [load4] "=&r"(load4), [p1] "=&r"(p1), [p2] "=&r"(p2),
+            [n1] "=&r"(n1), [n2] "=&r"(n2), [scratch1] "=&r"(scratch1),
+            [scratch2] "=&r"(scratch2), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [store1] "=&r"(store1),
+            [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr)
+          : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
+            [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
+            [vector4a] "r"(vector4a), [src_stride] "r"(src_stride),
+            [cm] "r"(cm), [dst_ptr] "r"(dst_ptr));
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void aom_convolve8_avg_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4, int w,
+                                  int h) {
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  if (((const int32_t *)filter_y)[0] == 0) {
+    aom_convolve2_avg_vert_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                 x_step_q4, filter_y, y_step_q4, w, h);
+  } else {
+    uint32_t pos = 38;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                         :
+                         : [pos] "r"(pos));
+
+    prefetch_store(dst);
+
+    switch (w) {
+      case 4:
+      case 8:
+      case 16:
+      case 32:
+        convolve_avg_vert_4_dspr2(src, src_stride, dst, dst_stride, filter_y, w,
+                                  h);
+        break;
+      case 64:
+        prefetch_store(dst + 32);
+        convolve_avg_vert_64_dspr2(src, src_stride, dst, dst_stride, filter_y,
+                                   h);
+        break;
+      default:
+        aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                                 x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  }
+}
+
+void aom_convolve8_avg_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int x_step_q4,
+                             const int16_t *filter_y, int y_step_q4, int w,
+                             int h) {
+  /* Fixed size intermediate buffer places limits on parameters. */
+  DECLARE_ALIGNED(32, uint8_t, temp[64 * 135]);
+  int32_t intermediate_height = ((h * y_step_q4) >> 4) + 7;
+
+  assert(w <= 64);
+  assert(h <= 64);
+  assert(x_step_q4 == 16);
+  assert(y_step_q4 == 16);
+
+  if (intermediate_height < h) intermediate_height = h;
+
+  aom_convolve8_horiz(src - (src_stride * 3), src_stride, temp, 64, filter_x,
+                      x_step_q4, filter_y, y_step_q4, w, intermediate_height);
+
+  aom_convolve8_avg_vert(temp + 64 * 3, 64, dst, dst_stride, filter_x,
+                         x_step_q4, filter_y, y_step_q4, w, h);
+}
+
+void aom_convolve_avg_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride,
+                            const int16_t *filter_x, int filter_x_stride,
+                            const int16_t *filter_y, int filter_y_stride, int w,
+                            int h) {
+  int x, y;
+  uint32_t tp1, tp2, tn1;
+  uint32_t tp3, tp4, tn2;
+
+  /* prefetch data to cache memory */
+  prefetch_load(src);
+  prefetch_load(src + 32);
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4:
+      /* 1 word storage */
+      for (y = h; y--;) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__(
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         0(%[dst])      \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "sw               %[tn1],         0(%[dst])      \n\t" /* store */
+
+            : [tn1] "=&r"(tn1), [tp1] "=&r"(tp1), [tp2] "=&r"(tp2)
+            : [src] "r"(src), [dst] "r"(dst));
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    case 8:
+      /* 2 word storage */
+      for (y = h; y--;) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__(
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         0(%[dst])      \n\t"
+            "ulw              %[tp3],         4(%[src])      \n\t"
+            "ulw              %[tp4],         4(%[dst])      \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "sw               %[tn1],         0(%[dst])      \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         4(%[dst])      \n\t" /* store */
+
+            : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3),
+              [tp4] "=&r"(tp4), [tn1] "=&r"(tn1), [tn2] "=&r"(tn2)
+            : [src] "r"(src), [dst] "r"(dst));
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    case 16:
+      /* 4 word storage */
+      for (y = h; y--;) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__(
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         0(%[dst])      \n\t"
+            "ulw              %[tp3],         4(%[src])      \n\t"
+            "ulw              %[tp4],         4(%[dst])      \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         8(%[src])      \n\t"
+            "ulw              %[tp2],         8(%[dst])      \n\t"
+            "sw               %[tn1],         0(%[dst])      \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         4(%[dst])      \n\t" /* store */
+            "ulw              %[tp3],         12(%[src])     \n\t"
+            "ulw              %[tp4],         12(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "sw               %[tn1],         8(%[dst])      \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         12(%[dst])     \n\t" /* store */
+
+            : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3),
+              [tp4] "=&r"(tp4), [tn1] "=&r"(tn1), [tn2] "=&r"(tn2)
+            : [src] "r"(src), [dst] "r"(dst));
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    case 32:
+      /* 8 word storage */
+      for (y = h; y--;) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__(
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         0(%[dst])      \n\t"
+            "ulw              %[tp3],         4(%[src])      \n\t"
+            "ulw              %[tp4],         4(%[dst])      \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         8(%[src])      \n\t"
+            "ulw              %[tp2],         8(%[dst])      \n\t"
+            "sw               %[tn1],         0(%[dst])      \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         4(%[dst])      \n\t" /* store */
+            "ulw              %[tp3],         12(%[src])     \n\t"
+            "ulw              %[tp4],         12(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         16(%[src])     \n\t"
+            "ulw              %[tp2],         16(%[dst])     \n\t"
+            "sw               %[tn1],         8(%[dst])      \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         12(%[dst])     \n\t" /* store */
+            "ulw              %[tp3],         20(%[src])     \n\t"
+            "ulw              %[tp4],         20(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         24(%[src])     \n\t"
+            "ulw              %[tp2],         24(%[dst])     \n\t"
+            "sw               %[tn1],         16(%[dst])     \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         20(%[dst])     \n\t" /* store */
+            "ulw              %[tp3],         28(%[src])     \n\t"
+            "ulw              %[tp4],         28(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "sw               %[tn1],         24(%[dst])     \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         28(%[dst])     \n\t" /* store */
+
+            : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3),
+              [tp4] "=&r"(tp4), [tn1] "=&r"(tn1), [tn2] "=&r"(tn2)
+            : [src] "r"(src), [dst] "r"(dst));
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    case 64:
+      prefetch_load(src + 64);
+      prefetch_store(dst + 32);
+
+      /* 16 word storage */
+      for (y = h; y--;) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_load(src + src_stride + 64);
+        prefetch_store(dst + dst_stride);
+        prefetch_store(dst + dst_stride + 32);
+
+        __asm__ __volatile__(
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         0(%[dst])      \n\t"
+            "ulw              %[tp3],         4(%[src])      \n\t"
+            "ulw              %[tp4],         4(%[dst])      \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         8(%[src])      \n\t"
+            "ulw              %[tp2],         8(%[dst])      \n\t"
+            "sw               %[tn1],         0(%[dst])      \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         4(%[dst])      \n\t" /* store */
+            "ulw              %[tp3],         12(%[src])     \n\t"
+            "ulw              %[tp4],         12(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         16(%[src])     \n\t"
+            "ulw              %[tp2],         16(%[dst])     \n\t"
+            "sw               %[tn1],         8(%[dst])      \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         12(%[dst])     \n\t" /* store */
+            "ulw              %[tp3],         20(%[src])     \n\t"
+            "ulw              %[tp4],         20(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         24(%[src])     \n\t"
+            "ulw              %[tp2],         24(%[dst])     \n\t"
+            "sw               %[tn1],         16(%[dst])     \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         20(%[dst])     \n\t" /* store */
+            "ulw              %[tp3],         28(%[src])     \n\t"
+            "ulw              %[tp4],         28(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         32(%[src])     \n\t"
+            "ulw              %[tp2],         32(%[dst])     \n\t"
+            "sw               %[tn1],         24(%[dst])     \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         28(%[dst])     \n\t" /* store */
+            "ulw              %[tp3],         36(%[src])     \n\t"
+            "ulw              %[tp4],         36(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         40(%[src])     \n\t"
+            "ulw              %[tp2],         40(%[dst])     \n\t"
+            "sw               %[tn1],         32(%[dst])     \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         36(%[dst])     \n\t" /* store */
+            "ulw              %[tp3],         44(%[src])     \n\t"
+            "ulw              %[tp4],         44(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         48(%[src])     \n\t"
+            "ulw              %[tp2],         48(%[dst])     \n\t"
+            "sw               %[tn1],         40(%[dst])     \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         44(%[dst])     \n\t" /* store */
+            "ulw              %[tp3],         52(%[src])     \n\t"
+            "ulw              %[tp4],         52(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "ulw              %[tp1],         56(%[src])     \n\t"
+            "ulw              %[tp2],         56(%[dst])     \n\t"
+            "sw               %[tn1],         48(%[dst])     \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         52(%[dst])     \n\t" /* store */
+            "ulw              %[tp3],         60(%[src])     \n\t"
+            "ulw              %[tp4],         60(%[dst])     \n\t"
+            "adduh_r.qb       %[tn1], %[tp2], %[tp1]         \n\t" /* average */
+            "sw               %[tn1],         56(%[dst])     \n\t" /* store */
+            "adduh_r.qb       %[tn2], %[tp3], %[tp4]         \n\t" /* average */
+            "sw               %[tn2],         60(%[dst])     \n\t" /* store */
+
+            : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3),
+              [tp4] "=&r"(tp4), [tn1] "=&r"(tn1), [tn2] "=&r"(tn2)
+            : [src] "r"(src), [dst] "r"(dst));
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+    default:
+      for (y = h; y > 0; --y) {
+        for (x = 0; x < w; ++x) {
+          dst[x] = (dst[x] + src[x] + 1) >> 1;
+        }
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+  }
+}
+#endif
diff --git a/third_party/aom/aom_dsp/mips/convolve8_avg_horiz_dspr2.c b/third_party/aom/aom_dsp/mips/convolve8_avg_horiz_dspr2.c
new file mode 100644
index 0000000000..f6534b4205
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve8_avg_horiz_dspr2.c
@@ -0,0 +1,998 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/convolve_common_dspr2.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_avg_horiz_4_dspr2(const uint8_t *src, int32_t src_stride,
+                                       uint8_t *dst, int32_t dst_stride,
+                                       const int16_t *filter_x0, int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3, Temp4;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3, p4;
+  uint32_t n1, n2, n3, n4;
+  uint32_t tn1, tn2;
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]    \n\t"
+        "ulw              %[tn2],         8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tn2]                         \n\t"
+        "balign           %[tn1],         %[tn2],         3              \n\t"
+        "balign           %[tn2],         %[tp2],         3              \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        "lbu              %[p2],          3(%[dst])                      \n\t" /* load odd 2 */
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],         %[Temp1](%[cm])                \n\t" /* even 1 */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "lbu              %[Temp1],       1(%[dst])                      \n\t" /* load odd 1 */
+        "preceu.ph.qbr    %[n1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[n2],          %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[n3],          %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[n4],          %[tn2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[n1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[n2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[n3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,           %[n4],          %[vector4b]    \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tn2],         0(%[dst])                      \n\t" /* load even 1 */
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],         %[Temp3](%[cm])                \n\t" /* even 2 */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[n1],          %[tn1]                         \n\t"
+        "lbux             %[tn1],         %[Temp2](%[cm])                \n\t" /* odd 1 */
+        "addqh_r.w        %[tn2],         %[tn2],         %[tp1]         \n\t" /* average even 1 */
+        "dpa.w.ph         $ac2,           %[n2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[n3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[n4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[n1],          %[vector4b]    \n\t"
+        "extp             %[Temp4],       $ac2,           31             \n\t"
+
+        "lbu              %[tp1],         2(%[dst])                      \n\t" /* load even 2 */
+        "sb               %[tn2],         0(%[dst])                      \n\t" /* store even 1 */
+
+        /* clamp */
+        "addqh_r.w        %[Temp1],       %[Temp1],       %[tn1]         \n\t" /* average odd 1 */
+        "lbux             %[n2],          %[Temp4](%[cm])                \n\t" /* odd 2 */
+        "sb               %[Temp1],       1(%[dst])                      \n\t" /* store odd 1 */
+
+        "addqh_r.w        %[tp1],         %[tp1],         %[tp2]         \n\t" /* average even 2 */
+        "sb               %[tp1],         2(%[dst])                      \n\t" /* store even 2 */
+
+        "addqh_r.w        %[p2],          %[p2],          %[n2]          \n\t" /* average odd 2 */
+        "sb               %[p2],          3(%[dst])                      \n\t" /* store odd 2 */
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tn1] "=&r"(tn1),
+          [tn2] "=&r"(tn2), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3),
+          [p4] "=&r"(p4), [n1] "=&r"(n1), [n2] "=&r"(n2), [n3] "=&r"(n3),
+          [n4] "=&r"(n4), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+          [Temp3] "=&r"(Temp3), [Temp4] "=&r"(Temp4)
+        : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
+          [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
+          [vector4a] "r"(vector4a), [cm] "r"(cm), [dst] "r"(dst),
+          [src] "r"(src));
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_avg_horiz_8_dspr2(const uint8_t *src, int32_t src_stride,
+                                       uint8_t *dst, int32_t dst_stride,
+                                       const int16_t *filter_x0, int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3, p4, n1;
+  uint32_t tn1, tn2, tn3;
+  uint32_t st0, st1;
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp2]                         \n\t"
+        "ulw              %[tn2],         8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+        "lbu              %[Temp2],       0(%[dst])                      \n\t"
+        "lbu              %[tn3],         2(%[dst])                      \n\t"
+
+        /* even 2. pixel */
+        "preceu.ph.qbr    %[p1],          %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[n1],          %[tn2]                         \n\t"
+        "ulw              %[tn1],         12(%[src])                     \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[st0],         %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac1                           \n\t"
+        "mthi             $zero,          $ac1                           \n\t"
+        "preceu.ph.qbr    %[p2],          %[tn1]                         \n\t"
+        "lbux             %[st1],         %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac1,           %[p4],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac1,           %[p1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac1,           %[n1],          %[vector4b]    \n\t"
+        "extp             %[Temp1],       $ac1,           31             \n\t"
+
+        "addqh_r.w        %[Temp2],       %[Temp2],       %[st0]         \n\t"
+        "addqh_r.w        %[tn3],         %[tn3],         %[st1]         \n\t"
+        "sb               %[Temp2],       0(%[dst])                      \n\t"
+        "sb               %[tn3],         2(%[dst])                      \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+
+        "balign           %[tn3],         %[tn1],         3              \n\t"
+        "balign           %[tn1],         %[tn2],         3              \n\t"
+        "balign           %[tn2],         %[tp2],         3              \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+
+        "lbux             %[st0],         %[Temp1](%[cm])                \n\t"
+        "lbu              %[Temp2],       4(%[dst])                      \n\t"
+        "addqh_r.w        %[Temp2],       %[Temp2],       %[st0]         \n\t"
+
+        "dpa.w.ph         $ac2,           %[p4],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[n1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector4b]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a],    $ac1                           \n\t"
+        "mthi             $zero,          $ac1                           \n\t"
+        "sb               %[Temp2],       4(%[dst])                      \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tn2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tp1],         6(%[dst])                      \n\t"
+
+        /* odd 2. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tn1]                         \n\t"
+        "preceu.ph.qbl    %[n1],          %[tn1]                         \n\t"
+        "lbux             %[st0],         %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,           %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac1,           %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac1,           %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],       $ac1,           31             \n\t"
+
+        "lbu              %[tp2],         1(%[dst])                      \n\t"
+        "lbu              %[tn2],         3(%[dst])                      \n\t"
+        "addqh_r.w        %[tp1],         %[tp1],         %[st0]         \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[st1],         %[Temp2](%[cm])                \n\t"
+        "preceu.ph.qbr    %[p2],          %[tn3]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,           %[n1],          %[vector4b]    \n\t"
+        "addqh_r.w        %[tp2],         %[tp2],         %[st1]         \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        "lbu              %[tn3],         5(%[dst])                      \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[tp2],         1(%[dst])                      \n\t"
+        "sb               %[tp1],         6(%[dst])                      \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[n1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector4b]    \n\t"
+        "extp             %[Temp1],       $ac2,           31             \n\t"
+
+        "lbu              %[tn1],         7(%[dst])                      \n\t"
+
+        /* clamp */
+        "lbux             %[p4],          %[Temp3](%[cm])                \n\t"
+        "addqh_r.w        %[tn2],         %[tn2],         %[p4]          \n\t"
+
+        "lbux             %[p2],          %[Temp2](%[cm])                \n\t"
+        "addqh_r.w        %[tn3],         %[tn3],         %[p2]          \n\t"
+
+        "lbux             %[n1],          %[Temp1](%[cm])                \n\t"
+        "addqh_r.w        %[tn1],         %[tn1],         %[n1]          \n\t"
+
+        /* store bytes */
+        "sb               %[tn2],         3(%[dst])                      \n\t"
+        "sb               %[tn3],         5(%[dst])                      \n\t"
+        "sb               %[tn1],         7(%[dst])                      \n\t"
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tn1] "=&r"(tn1),
+          [tn2] "=&r"(tn2), [tn3] "=&r"(tn3), [st0] "=&r"(st0),
+          [st1] "=&r"(st1), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3),
+          [p4] "=&r"(p4), [n1] "=&r"(n1), [Temp1] "=&r"(Temp1),
+          [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3)
+        : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
+          [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
+          [vector4a] "r"(vector4a), [cm] "r"(cm), [dst] "r"(dst),
+          [src] "r"(src));
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_avg_horiz_16_dspr2(const uint8_t *src_ptr,
+                                        int32_t src_stride, uint8_t *dst_ptr,
+                                        int32_t dst_stride,
+                                        const int16_t *filter_x0, int32_t h,
+                                        int32_t count) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t filter12, filter34, filter56, filter78;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_store(dst_ptr + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+          "lbu              %[st2],       0(%[dst])                    \n\t" /* load even 1 from dst */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          "lbu              %[qload3],    2(%[dst])                    \n\t" /* load even 2 from dst */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st2],       0(%[dst])                    \n\t" /* store even 1 to dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter12]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter34]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter56]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter78]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st2]       \n\t" /* average even 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    2(%[dst])                    \n\t" /* store even 2 to dst */
+          "ulw              %[qload2],    16(%[src])                   \n\t"
+          "lbu              %[qload3],    4(%[dst])                    \n\t" /* load even 3 from dst */
+          "lbu              %[qload1],    6(%[dst])                    \n\t" /* load even 4 from dst */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter12]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter34]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter56]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter78]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 3 */
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[qload3],    4(%[dst])                    \n\t" /* store even 3 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter12]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter34]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter56]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter78]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average even 4 */
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[qload1],    6(%[dst])                    \n\t" /* store even 4 to dst */
+          "ulw              %[qload3],    20(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p5],          %[filter12]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* even 6 */
+          "lbu              %[qload2],    8(%[dst])                    \n\t" /* load even 5 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 5 */
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    8(%[dst])                    \n\t" /* store even 5 to dst */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* even 7 */
+          "lbu              %[qload3],    10(%[dst])                   \n\t" /* load even 6 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          "lbu              %[st2],       12(%[dst])                   \n\t" /* load even 7 from dst */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* even 8 */
+          "sb               %[qload3],    10(%[dst])                   \n\t" /* store even 6 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* even 8 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                   \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 7 */
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       12(%[dst])                   \n\t" /* store even 7 to dst */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter12]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* odd 1 */
+          "lbu              %[qload2],    14(%[dst])                   \n\t" /* load even 8 from dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          "lbu              %[st1],       1(%[dst])                    \n\t" /* load odd 1 from dst */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 8 */
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    14(%[dst])                   \n\t" /* store even 8 to dst */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* odd 2 */
+          "lbu              %[qload3],    3(%[dst])                    \n\t" /* load odd 2 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st3],       %[st3],         %[st1]       \n\t" /* average odd 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* odd 3 */
+          "sb               %[st3],       1(%[dst])                    \n\t" /* store odd 1 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st1]       \n\t" /* average odd 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    3(%[dst])                    \n\t" /* store odd 2 to dst */
+          "lbu              %[qload1],    5(%[dst])                    \n\t" /* load odd 3 from dst */
+          "ulw              %[qload2],    17(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p4],          %[filter12]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter34]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter56]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter78]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          "lbu              %[st1],       7(%[dst])                    \n\t" /* load odd 4 from dst */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st2]       \n\t" /* average odd 3 */
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[qload1],    5(%[dst])                    \n\t" /* store odd 3 to dst */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter34]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter56]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter78]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          "lbu              %[qload1],    9(%[dst])                    \n\t" /* load odd 5 from dst */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st1],       %[st1],         %[st3]       \n\t" /* average odd 4 */
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st1],       7(%[dst])                    \n\t" /* store odd 4 to dst */
+          "ulw              %[qload3],    21(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p5],          %[filter12]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter34]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter56]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter78]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 5 */
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload1],    9(%[dst])                    \n\t" /* store odd 5 to dst */
+          "lbu              %[qload2],    11(%[dst])                   \n\t" /* load odd 6 from dst */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter12]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter34]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter56]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter78]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          "lbu              %[qload3],    13(%[dst])                   \n\t" /* load odd 7 from dst */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter12]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter34]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter56]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter78]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbu              %[qload1],    15(%[dst])                   \n\t" /* load odd 8 from dst */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average odd 6 */
+
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average odd 7 */
+
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 8 */
+
+          "sb               %[qload2],    11(%[dst])                   \n\t" /* store odd 6 to dst */
+          "sb               %[qload3],    13(%[dst])                   \n\t" /* store odd 7 to dst */
+          "sb               %[qload1],    15(%[dst])                   \n\t" /* store odd 8 to dst */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), [st1] "=&r"(st1),
+            [st2] "=&r"(st2), [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2),
+            [p3] "=&r"(p3), [p4] "=&r"(p4), [qload3] "=&r"(qload3),
+            [p5] "=&r"(p5), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+            [Temp3] "=&r"(Temp3)
+          : [filter12] "r"(filter12), [filter34] "r"(filter34),
+            [filter56] "r"(filter56), [filter78] "r"(filter78),
+            [vector_64] "r"(vector_64), [cm] "r"(cm), [dst] "r"(dst),
+            [src] "r"(src));
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void convolve_avg_horiz_64_dspr2(const uint8_t *src_ptr,
+                                        int32_t src_stride, uint8_t *dst_ptr,
+                                        int32_t dst_stride,
+                                        const int16_t *filter_x0, int32_t h) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t filter12, filter34, filter56, filter78;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+    prefetch_store(dst_ptr + dst_stride);
+    prefetch_store(dst_ptr + dst_stride + 32);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+          "lbu              %[st2],       0(%[dst])                    \n\t" /* load even 1 from dst */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          "lbu              %[qload3],    2(%[dst])                    \n\t" /* load even 2 from dst */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st2],       0(%[dst])                    \n\t" /* store even 1 to dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter12]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter34]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter56]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter78]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st2]       \n\t" /* average even 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    2(%[dst])                    \n\t" /* store even 2 to dst */
+          "ulw              %[qload2],    16(%[src])                   \n\t"
+          "lbu              %[qload3],    4(%[dst])                    \n\t" /* load even 3 from dst */
+          "lbu              %[qload1],    6(%[dst])                    \n\t" /* load even 4 from dst */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter12]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter34]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter56]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter78]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 3 */
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[qload3],    4(%[dst])                    \n\t" /* store even 3 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter12]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter34]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter56]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter78]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average even 4 */
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[qload1],    6(%[dst])                    \n\t" /* store even 4 to dst */
+          "ulw              %[qload3],    20(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p5],          %[filter12]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* even 6 */
+          "lbu              %[qload2],    8(%[dst])                    \n\t" /* load even 5 from dst */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 5 */
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    8(%[dst])                    \n\t" /* store even 5 to dst */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* even 7 */
+          "lbu              %[qload3],    10(%[dst])                   \n\t" /* load even 6 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          "lbu              %[st2],       12(%[dst])                   \n\t" /* load even 7 from dst */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average even 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* even 8 */
+          "sb               %[qload3],    10(%[dst])                   \n\t" /* store even 6 to dst */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* even 8 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                   \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          "addqh_r.w        %[st2],       %[st2],         %[st1]       \n\t" /* average even 7 */
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       12(%[dst])                   \n\t" /* store even 7 to dst */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter12]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* odd 1 */
+          "lbu              %[qload2],    14(%[dst])                   \n\t" /* load even 8 from dst */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          "lbu              %[st1],       1(%[dst])                    \n\t" /* load odd 1 from dst */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average even 8 */
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload2],    14(%[dst])                   \n\t" /* store even 8 to dst */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* odd 2 */
+          "lbu              %[qload3],    3(%[dst])                    \n\t" /* load odd 2 from dst */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st3],       %[st3],         %[st1]       \n\t" /* average odd 1 */
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* odd 3 */
+          "sb               %[st3],       1(%[dst])                    \n\t" /* store odd 1 to dst */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload3],    %[qload3],      %[st1]       \n\t" /* average odd 2 */
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[qload3],    3(%[dst])                    \n\t" /* store odd 2 to dst */
+          "lbu              %[qload1],    5(%[dst])                    \n\t" /* load odd 3 from dst */
+          "ulw              %[qload2],    17(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p4],          %[filter12]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter34]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter56]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter78]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          "lbu              %[st1],       7(%[dst])                    \n\t" /* load odd 4 from dst */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st2]       \n\t" /* average odd 3 */
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[qload1],    5(%[dst])                    \n\t" /* store odd 3 to dst */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter34]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter56]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter78]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          "lbu              %[qload1],    9(%[dst])                    \n\t" /* load odd 5 from dst */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "addqh_r.w        %[st1],       %[st1],         %[st3]       \n\t" /* average odd 4 */
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st1],       7(%[dst])                    \n\t" /* store odd 4 to dst */
+          "ulw              %[qload3],    21(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p5],          %[filter12]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter34]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter56]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter78]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 5 */
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[qload1],    9(%[dst])                    \n\t" /* store odd 5 to dst */
+          "lbu              %[qload2],    11(%[dst])                   \n\t" /* load odd 6 from dst */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter12]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter34]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter56]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter78]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          "lbu              %[qload3],    13(%[dst])                   \n\t" /* load odd 7 from dst */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter12]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter34]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter56]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter78]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbu              %[qload1],    15(%[dst])                   \n\t" /* load odd 8 from dst */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "addqh_r.w        %[qload2],    %[qload2],      %[st2]       \n\t" /* average odd 6 */
+
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "addqh_r.w        %[qload3],    %[qload3],      %[st3]       \n\t" /* average odd 7 */
+
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+          "addqh_r.w        %[qload1],    %[qload1],      %[st1]       \n\t" /* average odd 8 */
+
+          "sb               %[qload2],    11(%[dst])                   \n\t" /* store odd 6 to dst */
+          "sb               %[qload3],    13(%[dst])                   \n\t" /* store odd 7 to dst */
+          "sb               %[qload1],    15(%[dst])                   \n\t" /* store odd 8 to dst */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), [st1] "=&r"(st1),
+            [st2] "=&r"(st2), [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2),
+            [p3] "=&r"(p3), [p4] "=&r"(p4), [qload3] "=&r"(qload3),
+            [p5] "=&r"(p5), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+            [Temp3] "=&r"(Temp3)
+          : [filter12] "r"(filter12), [filter34] "r"(filter34),
+            [filter56] "r"(filter56), [filter78] "r"(filter78),
+            [vector_64] "r"(vector_64), [cm] "r"(cm), [dst] "r"(dst),
+            [src] "r"(src));
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+void aom_convolve8_avg_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int x_step_q4,
+                                   const int16_t *filter_y, int y_step_q4,
+                                   int w, int h) {
+  assert(x_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+
+  if (((const int32_t *)filter_x)[0] == 0) {
+    aom_convolve2_avg_horiz_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                  x_step_q4, filter_y, y_step_q4, w, h);
+  } else {
+    uint32_t pos = 38;
+
+    src -= 3;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                         :
+                         : [pos] "r"(pos));
+
+    /* prefetch data to cache memory */
+    prefetch_load(src);
+    prefetch_load(src + 32);
+    prefetch_store(dst);
+
+    switch (w) {
+      case 4:
+        convolve_avg_horiz_4_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                   h);
+        break;
+      case 8:
+        convolve_avg_horiz_8_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                   h);
+        break;
+      case 16:
+        convolve_avg_horiz_16_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                    h, 1);
+        break;
+      case 32:
+        convolve_avg_horiz_16_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                    h, 2);
+        break;
+      case 64:
+        prefetch_load(src + 64);
+        prefetch_store(dst + 32);
+
+        convolve_avg_horiz_64_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                                    h);
+        break;
+      default:
+        aom_convolve8_avg_horiz_c(src + 3, src_stride, dst, dst_stride,
+                                  filter_x, x_step_q4, filter_y, y_step_q4, w,
+                                  h);
+        break;
+    }
+  }
+}
+#endif
diff --git a/third_party/aom/aom_dsp/mips/convolve8_dspr2.c b/third_party/aom/aom_dsp/mips/convolve8_dspr2.c
new file mode 100644
index 0000000000..c871702f4b
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve8_dspr2.c
@@ -0,0 +1,1590 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/convolve_common_dspr2.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_horiz_4_transposed_dspr2(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              const int16_t *filter_x0,
+                                              int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint8_t *dst_ptr;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3, Temp4;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3, p4;
+  uint32_t tn1, tn2;
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    dst_ptr = dst;
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],         0(%[src])                      \n\t"
+        "ulw              %[tp2],         4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]    \n\t"
+        "ulw              %[tn2],         8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp1],       $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tn2]                         \n\t"
+        "balign           %[tn1],         %[tn2],         3              \n\t"
+        "balign           %[tn2],         %[tp2],         3              \n\t"
+        "balign           %[tp2],         %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],       $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],         %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac3                           \n\t"
+        "mthi             $zero,          $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[p3],          %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[p4],          %[tn2]                         \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp2],       $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],         %[Temp3](%[cm])                \n\t"
+        "mtlo             %[vector4a],    $ac2                           \n\t"
+        "mthi             $zero,          $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],          %[tn1]                         \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp4],       $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[tn1],         %[Temp2](%[cm])                \n\t"
+        "lbux             %[p2],          %[Temp4](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[tp1],         0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[tn1],         0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[tp2],         0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        "sb               %[p2],          0(%[dst_ptr])                  \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_stride]  \n\t"
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tn1] "=&r"(tn1),
+          [tn2] "=&r"(tn2), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3),
+          [p4] "=&r"(p4), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+          [Temp3] "=&r"(Temp3), [Temp4] "=&r"(Temp4), [dst_ptr] "+r"(dst_ptr)
+        : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
+          [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
+          [vector4a] "r"(vector4a), [cm] "r"(cm), [src] "r"(src),
+          [dst_stride] "r"(dst_stride));
+
+    /* Next row... */
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+static void convolve_horiz_8_transposed_dspr2(const uint8_t *src,
+                                              int32_t src_stride, uint8_t *dst,
+                                              int32_t dst_stride,
+                                              const int16_t *filter_x0,
+                                              int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint8_t *dst_ptr;
+  uint32_t vector4a = 64;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2, tp3;
+  uint32_t p1, p2, p3, p4, n1;
+  uint8_t *odd_dst;
+  uint32_t dst_pitch_2 = (dst_stride << 1);
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+
+    dst_ptr = dst;
+    odd_dst = (dst_ptr + dst_stride);
+
+    __asm__ __volatile__(
+        "ulw              %[tp2],         0(%[src])                       \n\t"
+        "ulw              %[tp1],         4(%[src])                       \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                          \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp2]                          \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp1]                          \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp1]                          \n\t"
+        "ulw              %[tp3],         8(%[src])                       \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]     \n\t"
+        "extp             %[Temp1],       $ac3,           31              \n\t"
+
+        /* even 2. pixel */
+        "preceu.ph.qbr    %[p1],          %[tp3]                          \n\t"
+        "preceu.ph.qbl    %[n1],          %[tp3]                          \n\t"
+        "ulw              %[tp2],         12(%[src])                      \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p3],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector4b]     \n\t"
+        "extp             %[Temp3],       $ac2,           31              \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[Temp2],       %[Temp1](%[cm])                 \n\t"
+        "mtlo             %[vector4a],    $ac1                            \n\t"
+        "mthi             $zero,          $ac1                            \n\t"
+        "preceu.ph.qbr    %[p2],          %[tp2]                          \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac1,           %[p4],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac1,           %[p1],          %[vector3b]     \n\t"
+        "lbux             %[tp3],         %[Temp3](%[cm])                 \n\t"
+        "dpa.w.ph         $ac1,           %[n1],          %[vector4b]     \n\t"
+        "extp             %[p3],          $ac1,           31              \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "sb               %[Temp2],       0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+        "sb               %[tp3],         0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+
+        "ulw              %[tp1],         1(%[src])                       \n\t"
+        "ulw              %[tp3],         5(%[src])                       \n\t"
+
+        "dpa.w.ph         $ac2,           %[p4],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac2,           %[n1],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector4b]     \n\t"
+        "extp             %[Temp3],       $ac2,           31              \n\t"
+
+        "lbux             %[tp2],         %[p3](%[cm])                    \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a],    $ac1                            \n\t"
+        "mthi             $zero,          $ac1                            \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp1]                          \n\t"
+        "preceu.ph.qbl    %[p2],          %[tp1]                          \n\t"
+        "preceu.ph.qbr    %[p3],          %[tp3]                          \n\t"
+        "preceu.ph.qbl    %[p4],          %[tp3]                          \n\t"
+        "sb               %[tp2],         0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+        "ulw              %[tp2],         9(%[src])                       \n\t"
+
+        "dpa.w.ph         $ac3,           %[p1],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p2],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector4b]     \n\t"
+        "extp             %[Temp2],       $ac3,           31              \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp1],         %[Temp3](%[cm])                 \n\t"
+        "mtlo             %[vector4a],    $ac3                            \n\t"
+        "mthi             $zero,          $ac3                            \n\t"
+        "mtlo             %[vector4a],    $ac2                            \n\t"
+        "mthi             $zero,          $ac2                            \n\t"
+        "preceu.ph.qbr    %[p1],          %[tp2]                          \n\t"
+        "preceu.ph.qbl    %[n1],          %[tp2]                          \n\t"
+        "ulw              %[Temp1],       13(%[src])                      \n\t"
+        "dpa.w.ph         $ac1,           %[p2],          %[vector1b]     \n\t"
+        "sb               %[tp1],         0(%[dst_ptr])                   \n\t"
+        "addu             %[dst_ptr],     %[dst_ptr],     %[dst_pitch_2]  \n\t"
+        "dpa.w.ph         $ac1,           %[p3],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac1,           %[p4],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac1,           %[p1],          %[vector4b]     \n\t"
+        "extp             %[Temp3],       $ac1,           31              \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[tp3],         %[Temp2](%[cm])                 \n\t"
+        "preceu.ph.qbr    %[p2],          %[Temp1]                        \n\t"
+        "dpa.w.ph         $ac3,           %[p3],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p4],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac3,           %[p1],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac3,           %[n1],          %[vector4b]     \n\t"
+        "extp             %[Temp2],       $ac3,           31              \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[tp3],         0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+        "dpa.w.ph         $ac2,           %[p4],          %[vector1b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p1],          %[vector2b]     \n\t"
+        "dpa.w.ph         $ac2,           %[n1],          %[vector3b]     \n\t"
+        "dpa.w.ph         $ac2,           %[p2],          %[vector4b]     \n\t"
+        "extp             %[Temp1],       $ac2,           31              \n\t"
+
+        /* clamp */
+        "lbux             %[p4],          %[Temp3](%[cm])                 \n\t"
+        "lbux             %[p2],          %[Temp2](%[cm])                 \n\t"
+        "lbux             %[n1],          %[Temp1](%[cm])                 \n\t"
+
+        /* store bytes */
+        "sb               %[p4],          0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+        "sb               %[p2],          0(%[odd_dst])                   \n\t"
+        "addu             %[odd_dst],     %[odd_dst],     %[dst_pitch_2]  \n\t"
+
+        "sb               %[n1],          0(%[odd_dst])                   \n\t"
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3), [p1] "=&r"(p1),
+          [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4), [n1] "=&r"(n1),
+          [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3),
+          [dst_ptr] "+r"(dst_ptr), [odd_dst] "+r"(odd_dst)
+        : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
+          [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
+          [vector4a] "r"(vector4a), [cm] "r"(cm), [src] "r"(src),
+          [dst_pitch_2] "r"(dst_pitch_2));
+
+    /* Next row... */
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+static void convolve_horiz_16_transposed_dspr2(
+    const uint8_t *src_ptr, int32_t src_stride, uint8_t *dst_ptr,
+    int32_t dst_stride, const int16_t *filter_x0, int32_t h, int32_t count) {
+  int32_t c, y;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t filter12, filter34, filter56, filter78;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  uint32_t dst_pitch_2 = (dst_stride << 1);
+  uint8_t *odd_dst;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+
+    src = src_ptr;
+    dst = dst_ptr;
+
+    odd_dst = (dst + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],        0(%[src])                       "
+          "\n\t"
+          "ulw              %[qload2],        4(%[src])                       "
+          "\n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 1 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 2 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       "
+          "\n\t"
+          "ulw              %[qload2],        8(%[src])                       "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter12]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter34]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter56]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter78]     "
+          "\n\t" /* even 1 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* even 3 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       "
+          "\n\t"
+          "ulw              %[qload1],        12(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p2],          %[filter12]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter34]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter56]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p1],          %[filter78]     "
+          "\n\t" /* even 1 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 1 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 4 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 1 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]   "
+          "          \n\t"
+          "dpa.w.ph         $ac3,             %[p3],          %[filter12]     "
+          "\n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter34]     "
+          "\n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter56]     "
+          "\n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p5],          %[filter78]     "
+          "\n\t" /* even 3 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* even 3 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 5 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 2 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload2],        16(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p4],          %[filter12]     "
+          "\n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter34]     "
+          "\n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter56]     "
+          "\n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter78]     "
+          "\n\t" /* even 4 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 4 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* even 6 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p4],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st3],           0(%[dst])                       "
+          "\n\t" /* even 3 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter12]     "
+          "\n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter34]     "
+          "\n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p2],          %[filter56]     "
+          "\n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter78]     "
+          "\n\t" /* even 5 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 5 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 7 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p1],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 4 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload1],        20(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p5],          %[filter12]     "
+          "\n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter34]     "
+          "\n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter56]     "
+          "\n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter78]     "
+          "\n\t" /* even 6 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* even 6 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 8 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 5 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter12]     "
+          "\n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter34]     "
+          "\n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter56]     "
+          "\n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter78]     "
+          "\n\t" /* even 7 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 7 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 1 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter12]     "
+          "\n\t" /* even 8 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter34]     "
+          "\n\t" /* even 8 */
+          "sb               %[st3],           0(%[dst])                       "
+          "\n\t" /* even 6 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter56]     "
+          "\n\t" /* even 8 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter78]     "
+          "\n\t" /* even 8 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 8 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],        1(%[src])                       "
+          "\n\t"
+          "ulw              %[qload2],        5(%[src])                       "
+          "\n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 2 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 7 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload2],        9(%[src])                       "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p1],          %[filter12]     "
+          "\n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter34]     "
+          "\n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter56]     "
+          "\n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter78]     "
+          "\n\t" /* odd 1 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 1 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* odd 3 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 8 */
+          "ulw              %[qload1],        13(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter12]     "
+          "\n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter34]     "
+          "\n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter56]     "
+          "\n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter78]     "
+          "\n\t" /* odd 2 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 2 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 4 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 1 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter12]     "
+          "\n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter34]     "
+          "\n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p1],          %[filter56]     "
+          "\n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter78]     "
+          "\n\t" /* odd 3 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* odd 3 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 5 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 2 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload2],        17(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p4],          %[filter12]     "
+          "\n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter34]     "
+          "\n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p5],          %[filter56]     "
+          "\n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter78]     "
+          "\n\t" /* odd 4 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 4 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* odd 6 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p4],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[odd_dst])                   "
+          "\n\t" /* odd 3 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter12]     "
+          "\n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter34]     "
+          "\n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter56]     "
+          "\n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter78]     "
+          "\n\t" /* odd 5 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 5 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 7 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p1],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 4 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload1],        21(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p5],          %[filter12]     "
+          "\n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p2],          %[filter34]     "
+          "\n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter56]     "
+          "\n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter78]     "
+          "\n\t" /* odd 6 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* odd 6 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 8 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 5 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p2],          %[filter12]     "
+          "\n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter34]     "
+          "\n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter56]     "
+          "\n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter78]     "
+          "\n\t" /* odd 7 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter12]     "
+          "\n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter34]     "
+          "\n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter56]     "
+          "\n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter78]     "
+          "\n\t" /* odd 8 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 8 */
+
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* odd 6 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 7 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 8 */
+
+          "sb               %[st2],           0(%[odd_dst])                   "
+          "\n\t" /* odd 6 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 7 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 8 */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), [p5] "=&r"(p5),
+            [st1] "=&r"(st1), [st2] "=&r"(st2), [st3] "=&r"(st3),
+            [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4),
+            [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3),
+            [dst] "+r"(dst), [odd_dst] "+r"(odd_dst)
+          : [filter12] "r"(filter12), [filter34] "r"(filter34),
+            [filter56] "r"(filter56), [filter78] "r"(filter78),
+            [vector_64] "r"(vector_64), [cm] "r"(cm), [src] "r"(src),
+            [dst_pitch_2] "r"(dst_pitch_2));
+
+      src += 16;
+      dst = (dst_ptr + ((c + 1) * 16 * dst_stride));
+      odd_dst = (dst + dst_stride);
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+
+    dst_ptr += 1;
+  }
+}
+
+static void convolve_horiz_64_transposed_dspr2(
+    const uint8_t *src_ptr, int32_t src_stride, uint8_t *dst_ptr,
+    int32_t dst_stride, const int16_t *filter_x0, int32_t h) {
+  int32_t c, y;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t filter12, filter34, filter56, filter78;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+  uint32_t dst_pitch_2 = (dst_stride << 1);
+  uint8_t *odd_dst;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+
+    src = src_ptr;
+    dst = dst_ptr;
+
+    odd_dst = (dst + dst_stride);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],        0(%[src])                       "
+          "\n\t"
+          "ulw              %[qload2],        4(%[src])                       "
+          "\n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 1 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 2 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       "
+          "\n\t"
+          "ulw              %[qload2],        8(%[src])                       "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter12]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter34]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter56]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter78]     "
+          "\n\t" /* even 1 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* even 3 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       "
+          "\n\t"
+          "ulw              %[qload1],        12(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p2],          %[filter12]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter34]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter56]     "
+          "\n\t" /* even 1 */
+          "dpa.w.ph         $ac2,             %[p1],          %[filter78]     "
+          "\n\t" /* even 1 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 1 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 4 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 1 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]   "
+          "          \n\t"
+          "dpa.w.ph         $ac3,             %[p3],          %[filter12]     "
+          "\n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter34]     "
+          "\n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter56]     "
+          "\n\t" /* even 3 */
+          "dpa.w.ph         $ac3,             %[p5],          %[filter78]     "
+          "\n\t" /* even 3 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* even 3 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 5 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 2 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload2],        16(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p4],          %[filter12]     "
+          "\n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter34]     "
+          "\n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter56]     "
+          "\n\t" /* even 4 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter78]     "
+          "\n\t" /* even 4 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 4 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* even 6 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p4],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st3],           0(%[dst])                       "
+          "\n\t" /* even 3 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter12]     "
+          "\n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter34]     "
+          "\n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p2],          %[filter56]     "
+          "\n\t" /* even 5 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter78]     "
+          "\n\t" /* even 5 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 5 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* even 7 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p1],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 4 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload1],        20(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p5],          %[filter12]     "
+          "\n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter34]     "
+          "\n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter56]     "
+          "\n\t" /* even 6 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter78]     "
+          "\n\t" /* even 6 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* even 6 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* even 8 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 5 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter12]     "
+          "\n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter34]     "
+          "\n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter56]     "
+          "\n\t" /* even 7 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter78]     "
+          "\n\t" /* even 7 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* even 7 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 1 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter12]     "
+          "\n\t" /* even 8 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter34]     "
+          "\n\t" /* even 8 */
+          "sb               %[st3],           0(%[dst])                       "
+          "\n\t" /* even 6 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p1],          %[filter56]     "
+          "\n\t" /* even 8 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter78]     "
+          "\n\t" /* even 8 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* even 8 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],        1(%[src])                       "
+          "\n\t"
+          "ulw              %[qload2],        5(%[src])                       "
+          "\n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 2 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p2],            %[qload1]                       "
+          "\n\t"
+          "preceu.ph.qbr    %[p3],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p4],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[dst])                       "
+          "\n\t" /* even 7 */
+          "addu             %[dst],           %[dst],         %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload2],        9(%[src])                       "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p1],          %[filter12]     "
+          "\n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter34]     "
+          "\n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter56]     "
+          "\n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter78]     "
+          "\n\t" /* odd 1 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 1 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* odd 3 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p1],            %[qload2]                       "
+          "\n\t"
+          "preceu.ph.qbl    %[p5],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[dst])                       "
+          "\n\t" /* even 8 */
+          "ulw              %[qload1],        13(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p2],          %[filter12]     "
+          "\n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter34]     "
+          "\n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter56]     "
+          "\n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter78]     "
+          "\n\t" /* odd 2 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 2 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 4 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p2],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 1 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p3],          %[filter12]     "
+          "\n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter34]     "
+          "\n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p1],          %[filter56]     "
+          "\n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,             %[p5],          %[filter78]     "
+          "\n\t" /* odd 3 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* odd 3 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 5 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p3],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 2 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload2],        17(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p4],          %[filter12]     "
+          "\n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter34]     "
+          "\n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p5],          %[filter56]     "
+          "\n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,             %[p2],          %[filter78]     "
+          "\n\t" /* odd 4 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 4 */
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64],     $ac2                            "
+          "\n\t" /* odd 6 */
+          "mthi             $zero,            $ac2                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p4],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st2],           0(%[odd_dst])                   "
+          "\n\t" /* odd 3 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac1,             %[p1],          %[filter12]     "
+          "\n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter34]     "
+          "\n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p2],          %[filter56]     "
+          "\n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter78]     "
+          "\n\t" /* odd 5 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 5 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64],     $ac3                            "
+          "\n\t" /* odd 7 */
+          "mthi             $zero,            $ac3                            "
+          "\n\t"
+          "preceu.ph.qbl    %[p1],            %[qload2]                       "
+          "\n\t"
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 4 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "ulw              %[qload1],        21(%[src])                      "
+          "\n\t"
+          "dpa.w.ph         $ac2,             %[p5],          %[filter12]     "
+          "\n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p2],          %[filter34]     "
+          "\n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p3],          %[filter56]     "
+          "\n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,             %[p4],          %[filter78]     "
+          "\n\t" /* odd 6 */
+          "extp             %[Temp2],         $ac2,           31              "
+          "\n\t" /* odd 6 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64],     $ac1                            "
+          "\n\t" /* odd 8 */
+          "mthi             $zero,            $ac1                            "
+          "\n\t"
+          "preceu.ph.qbr    %[p5],            %[qload1]                       "
+          "\n\t"
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 5 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+          "dpa.w.ph         $ac3,             %[p2],          %[filter12]     "
+          "\n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p3],          %[filter34]     "
+          "\n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p4],          %[filter56]     "
+          "\n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,             %[p1],          %[filter78]     "
+          "\n\t" /* odd 7 */
+          "extp             %[Temp3],         $ac3,           31              "
+          "\n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,             %[p3],          %[filter12]     "
+          "\n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p4],          %[filter34]     "
+          "\n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p1],          %[filter56]     "
+          "\n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,             %[p5],          %[filter78]     "
+          "\n\t" /* odd 8 */
+          "extp             %[Temp1],         $ac1,           31              "
+          "\n\t" /* odd 8 */
+
+          "lbux             %[st2],           %[Temp2](%[cm])                 "
+          "\n\t" /* odd 6 */
+          "lbux             %[st3],           %[Temp3](%[cm])                 "
+          "\n\t" /* odd 7 */
+          "lbux             %[st1],           %[Temp1](%[cm])                 "
+          "\n\t" /* odd 8 */
+
+          "sb               %[st2],           0(%[odd_dst])                   "
+          "\n\t" /* odd 6 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+
+          "sb               %[st3],           0(%[odd_dst])                   "
+          "\n\t" /* odd 7 */
+          "addu             %[odd_dst],       %[odd_dst],     %[dst_pitch_2]  "
+          "\n\t"
+
+          "sb               %[st1],           0(%[odd_dst])                   "
+          "\n\t" /* odd 8 */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2), [p5] "=&r"(p5),
+            [st1] "=&r"(st1), [st2] "=&r"(st2), [st3] "=&r"(st3),
+            [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4),
+            [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3),
+            [dst] "+r"(dst), [odd_dst] "+r"(odd_dst)
+          : [filter12] "r"(filter12), [filter34] "r"(filter34),
+            [filter56] "r"(filter56), [filter78] "r"(filter78),
+            [vector_64] "r"(vector_64), [cm] "r"(cm), [src] "r"(src),
+            [dst_pitch_2] "r"(dst_pitch_2));
+
+      src += 16;
+      dst = (dst_ptr + ((c + 1) * 16 * dst_stride));
+      odd_dst = (dst + dst_stride);
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+
+    dst_ptr += 1;
+  }
+}
+
+void convolve_horiz_transposed(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter, int w, int h) {
+  int x, y, k;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      int sum = 0;
+
+      for (k = 0; k < 8; ++k) sum += src[x + k] * filter[k];
+
+      dst[x * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+void copy_horiz_transposed(const uint8_t *src, ptrdiff_t src_stride,
+                           uint8_t *dst, ptrdiff_t dst_stride, int w, int h) {
+  int x, y;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      dst[x * dst_stride] = src[x];
+    }
+
+    src += src_stride;
+    dst += 1;
+  }
+}
+
+void aom_convolve8_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const int16_t *filter_x,
+                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                         int w, int h) {
+  DECLARE_ALIGNED(32, uint8_t, temp[64 * 135]);
+  int32_t intermediate_height = ((h * y_step_q4) >> 4) + 7;
+  uint32_t pos = 38;
+
+  assert(x_step_q4 == 16);
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                       :
+                       : [pos] "r"(pos));
+
+  if (intermediate_height < h) intermediate_height = h;
+
+  /* copy the src to dst */
+  if (filter_x[3] == 0x80) {
+    copy_horiz_transposed(src - src_stride * 3, src_stride, temp,
+                          intermediate_height, w, intermediate_height);
+  } else if (((const int32_t *)filter_x)[0] == 0) {
+    aom_convolve2_dspr2(src - src_stride * 3, src_stride, temp,
+                        intermediate_height, filter_x, w, intermediate_height);
+  } else {
+    src -= (src_stride * 3 + 3);
+
+    /* prefetch data to cache memory */
+    prefetch_load(src);
+    prefetch_load(src + 32);
+
+    switch (w) {
+      case 4:
+        convolve_horiz_4_transposed_dspr2(src, src_stride, temp,
+                                          intermediate_height, filter_x,
+                                          intermediate_height);
+        break;
+      case 8:
+        convolve_horiz_8_transposed_dspr2(src, src_stride, temp,
+                                          intermediate_height, filter_x,
+                                          intermediate_height);
+        break;
+      case 16:
+      case 32:
+        convolve_horiz_16_transposed_dspr2(src, src_stride, temp,
+                                           intermediate_height, filter_x,
+                                           intermediate_height, (w / 16));
+        break;
+      case 64:
+        prefetch_load(src + 32);
+        convolve_horiz_64_transposed_dspr2(src, src_stride, temp,
+                                           intermediate_height, filter_x,
+                                           intermediate_height);
+        break;
+      default:
+        convolve_horiz_transposed(src, src_stride, temp, intermediate_height,
+                                  filter_x, w, intermediate_height);
+        break;
+    }
+  }
+
+  /* copy the src to dst */
+  if (filter_y[3] == 0x80) {
+    copy_horiz_transposed(temp + 3, intermediate_height, dst, dst_stride, h, w);
+  } else if (((const int32_t *)filter_y)[0] == 0) {
+    aom_convolve2_dspr2(temp + 3, intermediate_height, dst, dst_stride,
+                        filter_y, h, w);
+  } else {
+    switch (h) {
+      case 4:
+        convolve_horiz_4_transposed_dspr2(temp, intermediate_height, dst,
+                                          dst_stride, filter_y, w);
+        break;
+      case 8:
+        convolve_horiz_8_transposed_dspr2(temp, intermediate_height, dst,
+                                          dst_stride, filter_y, w);
+        break;
+      case 16:
+      case 32:
+        convolve_horiz_16_transposed_dspr2(temp, intermediate_height, dst,
+                                           dst_stride, filter_y, w, (h / 16));
+        break;
+      case 64:
+        convolve_horiz_64_transposed_dspr2(temp, intermediate_height, dst,
+                                           dst_stride, filter_y, w);
+        break;
+      default:
+        convolve_horiz_transposed(temp, intermediate_height, dst, dst_stride,
+                                  filter_y, h, w);
+        break;
+    }
+  }
+}
+
+void aom_convolve_copy_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int filter_x_stride,
+                             const int16_t *filter_y, int filter_y_stride,
+                             int w, int h) {
+  int x, y;
+
+  /* prefetch data to cache memory */
+  prefetch_load(src);
+  prefetch_load(src + 32);
+  prefetch_store(dst);
+
+  switch (w) {
+    case 4: {
+      uint32_t tp1;
+
+      /* 1 word storage */
+      for (y = h; y--;) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__(
+            "ulw              %[tp1],         (%[src])      \n\t"
+            "sw               %[tp1],         (%[dst])      \n\t" /* store */
+
+            : [tp1] "=&r"(tp1)
+            : [src] "r"(src), [dst] "r"(dst));
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+    } break;
+    case 8: {
+      uint32_t tp1, tp2;
+
+      /* 2 word storage */
+      for (y = h; y--;) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__(
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         4(%[src])      \n\t"
+            "sw               %[tp1],         0(%[dst])      \n\t" /* store */
+            "sw               %[tp2],         4(%[dst])      \n\t" /* store */
+
+            : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2)
+            : [src] "r"(src), [dst] "r"(dst));
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+    } break;
+    case 16: {
+      uint32_t tp1, tp2, tp3, tp4;
+
+      /* 4 word storage */
+      for (y = h; y--;) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__(
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         4(%[src])      \n\t"
+            "ulw              %[tp3],         8(%[src])      \n\t"
+            "ulw              %[tp4],         12(%[src])     \n\t"
+
+            "sw               %[tp1],         0(%[dst])      \n\t" /* store */
+            "sw               %[tp2],         4(%[dst])      \n\t" /* store */
+            "sw               %[tp3],         8(%[dst])      \n\t" /* store */
+            "sw               %[tp4],         12(%[dst])     \n\t" /* store */
+
+            : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3),
+              [tp4] "=&r"(tp4)
+            : [src] "r"(src), [dst] "r"(dst));
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+    } break;
+    case 32: {
+      uint32_t tp1, tp2, tp3, tp4;
+      uint32_t tp5, tp6, tp7, tp8;
+
+      /* 8 word storage */
+      for (y = h; y--;) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_store(dst + dst_stride);
+
+        __asm__ __volatile__(
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         4(%[src])      \n\t"
+            "ulw              %[tp3],         8(%[src])      \n\t"
+            "ulw              %[tp4],         12(%[src])     \n\t"
+            "ulw              %[tp5],         16(%[src])     \n\t"
+            "ulw              %[tp6],         20(%[src])     \n\t"
+            "ulw              %[tp7],         24(%[src])     \n\t"
+            "ulw              %[tp8],         28(%[src])     \n\t"
+
+            "sw               %[tp1],         0(%[dst])      \n\t" /* store */
+            "sw               %[tp2],         4(%[dst])      \n\t" /* store */
+            "sw               %[tp3],         8(%[dst])      \n\t" /* store */
+            "sw               %[tp4],         12(%[dst])     \n\t" /* store */
+            "sw               %[tp5],         16(%[dst])     \n\t" /* store */
+            "sw               %[tp6],         20(%[dst])     \n\t" /* store */
+            "sw               %[tp7],         24(%[dst])     \n\t" /* store */
+            "sw               %[tp8],         28(%[dst])     \n\t" /* store */
+
+            : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3),
+              [tp4] "=&r"(tp4), [tp5] "=&r"(tp5), [tp6] "=&r"(tp6),
+              [tp7] "=&r"(tp7), [tp8] "=&r"(tp8)
+            : [src] "r"(src), [dst] "r"(dst));
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+    } break;
+    case 64: {
+      uint32_t tp1, tp2, tp3, tp4;
+      uint32_t tp5, tp6, tp7, tp8;
+
+      prefetch_load(src + 64);
+      prefetch_store(dst + 32);
+
+      /* 16 word storage */
+      for (y = h; y--;) {
+        prefetch_load(src + src_stride);
+        prefetch_load(src + src_stride + 32);
+        prefetch_load(src + src_stride + 64);
+        prefetch_store(dst + dst_stride);
+        prefetch_store(dst + dst_stride + 32);
+
+        __asm__ __volatile__(
+            "ulw              %[tp1],         0(%[src])      \n\t"
+            "ulw              %[tp2],         4(%[src])      \n\t"
+            "ulw              %[tp3],         8(%[src])      \n\t"
+            "ulw              %[tp4],         12(%[src])     \n\t"
+            "ulw              %[tp5],         16(%[src])     \n\t"
+            "ulw              %[tp6],         20(%[src])     \n\t"
+            "ulw              %[tp7],         24(%[src])     \n\t"
+            "ulw              %[tp8],         28(%[src])     \n\t"
+
+            "sw               %[tp1],         0(%[dst])      \n\t" /* store */
+            "sw               %[tp2],         4(%[dst])      \n\t" /* store */
+            "sw               %[tp3],         8(%[dst])      \n\t" /* store */
+            "sw               %[tp4],         12(%[dst])     \n\t" /* store */
+            "sw               %[tp5],         16(%[dst])     \n\t" /* store */
+            "sw               %[tp6],         20(%[dst])     \n\t" /* store */
+            "sw               %[tp7],         24(%[dst])     \n\t" /* store */
+            "sw               %[tp8],         28(%[dst])     \n\t" /* store */
+
+            "ulw              %[tp1],         32(%[src])     \n\t"
+            "ulw              %[tp2],         36(%[src])     \n\t"
+            "ulw              %[tp3],         40(%[src])     \n\t"
+            "ulw              %[tp4],         44(%[src])     \n\t"
+            "ulw              %[tp5],         48(%[src])     \n\t"
+            "ulw              %[tp6],         52(%[src])     \n\t"
+            "ulw              %[tp7],         56(%[src])     \n\t"
+            "ulw              %[tp8],         60(%[src])     \n\t"
+
+            "sw               %[tp1],         32(%[dst])     \n\t" /* store */
+            "sw               %[tp2],         36(%[dst])     \n\t" /* store */
+            "sw               %[tp3],         40(%[dst])     \n\t" /* store */
+            "sw               %[tp4],         44(%[dst])     \n\t" /* store */
+            "sw               %[tp5],         48(%[dst])     \n\t" /* store */
+            "sw               %[tp6],         52(%[dst])     \n\t" /* store */
+            "sw               %[tp7],         56(%[dst])     \n\t" /* store */
+            "sw               %[tp8],         60(%[dst])     \n\t" /* store */
+
+            : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tp3] "=&r"(tp3),
+              [tp4] "=&r"(tp4), [tp5] "=&r"(tp5), [tp6] "=&r"(tp6),
+              [tp7] "=&r"(tp7), [tp8] "=&r"(tp8)
+            : [src] "r"(src), [dst] "r"(dst));
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+    } break;
+    default:
+      for (y = h; y--;) {
+        for (x = 0; x < w; ++x) {
+          dst[x] = src[x];
+        }
+
+        src += src_stride;
+        dst += dst_stride;
+      }
+      break;
+  }
+}
+#endif
diff --git a/third_party/aom/aom_dsp/mips/convolve8_horiz_dspr2.c b/third_party/aom/aom_dsp/mips/convolve8_horiz_dspr2.c
new file mode 100644
index 0000000000..c605576179
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve8_horiz_dspr2.c
@@ -0,0 +1,878 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/convolve_common_dspr2.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_horiz_4_dspr2(const uint8_t *src, int32_t src_stride,
+                                   uint8_t *dst, int32_t dst_stride,
+                                   const int16_t *filter_x0, int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3, Temp4;
+  uint32_t vector4a = 64;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3, p4;
+  uint32_t n1, n2, n3, n4;
+  uint32_t tn1, tn2;
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],      0(%[src])                      \n\t"
+        "ulw              %[tp2],      4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],       %[tp2]                         \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p3],          %[vector3b]    \n\t"
+        "ulw              %[tn2],      8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,        %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp1],    $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tn2]                         \n\t"
+        "balign           %[tn1],      %[tn2],         3              \n\t"
+        "balign           %[tn2],      %[tp2],         3              \n\t"
+        "balign           %[tp2],      %[tp1],         3              \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        /* odd 1. pixel */
+        "lbux             %[tp1],      %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "preceu.ph.qbr    %[n1],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[n2],       %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[n3],       %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[n4],       %[tn2]                         \n\t"
+        "dpa.w.ph         $ac3,        %[n1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,        %[n2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,        %[n3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,        %[n4],          %[vector4b]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "lbux             %[tp2],      %[Temp3](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "preceu.ph.qbr    %[n1],       %[tn1]                         \n\t"
+        "dpa.w.ph         $ac2,        %[n2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,        %[n3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,        %[n4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,        %[n1],          %[vector4b]    \n\t"
+        "extp             %[Temp4],    $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[tn1],      %[Temp2](%[cm])                \n\t"
+        "lbux             %[n2],       %[Temp4](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[tp1],      0(%[dst])                      \n\t"
+        "sb               %[tn1],      1(%[dst])                      \n\t"
+        "sb               %[tp2],      2(%[dst])                      \n\t"
+        "sb               %[n2],       3(%[dst])                      \n\t"
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tn1] "=&r"(tn1),
+          [tn2] "=&r"(tn2), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3),
+          [p4] "=&r"(p4), [n1] "=&r"(n1), [n2] "=&r"(n2), [n3] "=&r"(n3),
+          [n4] "=&r"(n4), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+          [Temp3] "=&r"(Temp3), [Temp4] "=&r"(Temp4)
+        : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
+          [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
+          [vector4a] "r"(vector4a), [cm] "r"(cm), [dst] "r"(dst),
+          [src] "r"(src));
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_horiz_8_dspr2(const uint8_t *src, int32_t src_stride,
+                                   uint8_t *dst, int32_t dst_stride,
+                                   const int16_t *filter_x0, int32_t h) {
+  int32_t y;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t tp1, tp2;
+  uint32_t p1, p2, p3, p4, n1;
+  uint32_t tn1, tn2, tn3;
+  uint32_t st0, st1;
+
+  vector1b = ((const int32_t *)filter_x0)[0];
+  vector2b = ((const int32_t *)filter_x0)[1];
+  vector3b = ((const int32_t *)filter_x0)[2];
+  vector4b = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_load(src + src_stride);
+    prefetch_load(src + src_stride + 32);
+    prefetch_store(dst + dst_stride);
+
+    __asm__ __volatile__(
+        "ulw              %[tp1],      0(%[src])                      \n\t"
+        "ulw              %[tp2],      4(%[src])                      \n\t"
+
+        /* even 1. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp1]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp1]                         \n\t"
+        "preceu.ph.qbr    %[p3],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p4],       %[tp2]                         \n\t"
+        "ulw              %[tn2],      8(%[src])                      \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp1],    $ac3,           31             \n\t"
+
+        /* even 2. pixel */
+        "preceu.ph.qbr    %[p1],       %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[n1],       %[tn2]                         \n\t"
+        "ulw              %[tn1],      12(%[src])                     \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        /* even 3. pixel */
+        "lbux             %[st0],      %[Temp1](%[cm])                \n\t"
+        "mtlo             %[vector4a], $ac1                           \n\t"
+        "mthi             $zero,       $ac1                           \n\t"
+        "preceu.ph.qbr    %[p2],       %[tn1]                         \n\t"
+        "dpa.w.ph         $ac1,        %[p3],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac1,        %[p4],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac1,        %[p1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac1,        %[n1],          %[vector4b]    \n\t"
+        "extp             %[Temp1],    $ac1,           31             \n\t"
+
+        /* even 4. pixel */
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "sb               %[st0],      0(%[dst])                      \n\t"
+        "lbux             %[st1],      %[Temp3](%[cm])                \n\t"
+
+        "balign           %[tn3],      %[tn1],         3              \n\t"
+        "balign           %[tn1],      %[tn2],         3              \n\t"
+        "balign           %[tn2],      %[tp2],         3              \n\t"
+        "balign           %[tp2],      %[tp1],         3              \n\t"
+
+        "dpa.w.ph         $ac2,        %[p4],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p1],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,        %[n1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[vector4b]    \n\t"
+        "extp             %[Temp3],    $ac2,           31             \n\t"
+
+        "lbux             %[st0],      %[Temp1](%[cm])                \n\t"
+
+        /* odd 1. pixel */
+        "mtlo             %[vector4a], $ac1                           \n\t"
+        "mthi             $zero,       $ac1                           \n\t"
+        "sb               %[st1],      2(%[dst])                      \n\t"
+        "preceu.ph.qbr    %[p1],       %[tp2]                         \n\t"
+        "preceu.ph.qbl    %[p2],       %[tp2]                         \n\t"
+        "preceu.ph.qbr    %[p3],       %[tn2]                         \n\t"
+        "preceu.ph.qbl    %[p4],       %[tn2]                         \n\t"
+        "sb               %[st0],      4(%[dst])                      \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p2],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p3],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p4],          %[vector4b]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 2. pixel */
+        "mtlo             %[vector4a], $ac3                           \n\t"
+        "mthi             $zero,       $ac3                           \n\t"
+        "mtlo             %[vector4a], $ac2                           \n\t"
+        "mthi             $zero,       $ac2                           \n\t"
+        "preceu.ph.qbr    %[p1],       %[tn1]                         \n\t"
+        "preceu.ph.qbl    %[n1],       %[tn1]                         \n\t"
+        "lbux             %[st0],      %[Temp3](%[cm])                \n\t"
+        "dpa.w.ph         $ac1,        %[p2],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac1,        %[p3],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac1,        %[p4],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac1,        %[p1],          %[vector4b]    \n\t"
+        "extp             %[Temp3],    $ac1,           31             \n\t"
+
+        /* odd 3. pixel */
+        "lbux             %[st1],      %[Temp2](%[cm])                \n\t"
+        "preceu.ph.qbr    %[p2],       %[tn3]                         \n\t"
+        "dpa.w.ph         $ac3,        %[p3],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p4],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac3,        %[p1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac3,        %[n1],          %[vector4b]    \n\t"
+        "extp             %[Temp2],    $ac3,           31             \n\t"
+
+        /* odd 4. pixel */
+        "sb               %[st1],      1(%[dst])                      \n\t"
+        "sb               %[st0],      6(%[dst])                      \n\t"
+        "dpa.w.ph         $ac2,        %[p4],          %[vector1b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p1],          %[vector2b]    \n\t"
+        "dpa.w.ph         $ac2,        %[n1],          %[vector3b]    \n\t"
+        "dpa.w.ph         $ac2,        %[p2],          %[vector4b]    \n\t"
+        "extp             %[Temp1],    $ac2,           31             \n\t"
+
+        /* clamp */
+        "lbux             %[p4],       %[Temp3](%[cm])                \n\t"
+        "lbux             %[p2],       %[Temp2](%[cm])                \n\t"
+        "lbux             %[n1],       %[Temp1](%[cm])                \n\t"
+
+        /* store bytes */
+        "sb               %[p4],       3(%[dst])                      \n\t"
+        "sb               %[p2],       5(%[dst])                      \n\t"
+        "sb               %[n1],       7(%[dst])                      \n\t"
+
+        : [tp1] "=&r"(tp1), [tp2] "=&r"(tp2), [tn1] "=&r"(tn1),
+          [tn2] "=&r"(tn2), [tn3] "=&r"(tn3), [st0] "=&r"(st0),
+          [st1] "=&r"(st1), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3),
+          [p4] "=&r"(p4), [n1] "=&r"(n1), [Temp1] "=&r"(Temp1),
+          [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3)
+        : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
+          [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
+          [vector4a] "r"(vector4a), [cm] "r"(cm), [dst] "r"(dst),
+          [src] "r"(src));
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_horiz_16_dspr2(const uint8_t *src_ptr, int32_t src_stride,
+                                    uint8_t *dst_ptr, int32_t dst_stride,
+                                    const int16_t *filter_x0, int32_t h,
+                                    int32_t count) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t filter12, filter34, filter56, filter78;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_store(dst_ptr + dst_stride);
+
+    for (c = 0; c < count; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st1],       0(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter12]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter34]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter56]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter78]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st2],       2(%[dst])                    \n\t" /* even 1 */
+          "ulw              %[qload2],    16(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p4],          %[filter12]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter34]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter56]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter78]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st3],       4(%[dst])                    \n\t" /* even 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter12]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter34]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter56]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter78]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st1],       6(%[dst])                    \n\t" /* even 4 */
+          "ulw              %[qload3],    20(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p5],          %[filter12]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* even 6 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       8(%[dst])                    \n\t" /* even 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* even 7 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* even 8 */
+          "sb               %[st3],       10(%[dst])                   \n\t" /* even 6 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* even 8 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                    \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st1],       12(%[dst])                   \n\t" /* even 7 */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter12]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       14(%[dst])                   \n\t" /* even 8 */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* odd 2 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st3],       1(%[dst])                    \n\t" /* odd 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* odd 3 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st1],       3(%[dst])                    \n\t" /* odd 2 */
+          "ulw              %[qload2],    17(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p4],          %[filter12]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter34]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter56]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter78]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       5(%[dst])                    \n\t" /* odd 3 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter34]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter56]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter78]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st3],       7(%[dst])                    \n\t" /* odd 4 */
+          "ulw              %[qload3],    21(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p5],          %[filter12]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter34]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter56]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter78]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st1],       9(%[dst])                    \n\t" /* odd 5 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter12]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter34]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter56]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter78]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter12]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter34]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter56]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter78]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+
+          "sb               %[st2],       11(%[dst])                   \n\t" /* odd 6 */
+          "sb               %[st3],       13(%[dst])                   \n\t" /* odd 7 */
+          "sb               %[st1],       15(%[dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2),
+            [qload3] "=&r"(qload3), [st1] "=&r"(st1), [st2] "=&r"(st2),
+            [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3),
+            [p4] "=&r"(p4), [p5] "=&r"(p5), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3)
+          : [filter12] "r"(filter12), [filter34] "r"(filter34),
+            [filter56] "r"(filter56), [filter78] "r"(filter78),
+            [vector_64] "r"(vector_64), [cm] "r"(cm), [dst] "r"(dst),
+            [src] "r"(src));
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void convolve_horiz_64_dspr2(const uint8_t *src_ptr, int32_t src_stride,
+                                    uint8_t *dst_ptr, int32_t dst_stride,
+                                    const int16_t *filter_x0, int32_t h) {
+  int32_t y, c;
+  const uint8_t *src;
+  uint8_t *dst;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector_64 = 64;
+  int32_t filter12, filter34, filter56, filter78;
+  int32_t Temp1, Temp2, Temp3;
+  uint32_t qload1, qload2, qload3;
+  uint32_t p1, p2, p3, p4, p5;
+  uint32_t st1, st2, st3;
+
+  filter12 = ((const int32_t *)filter_x0)[0];
+  filter34 = ((const int32_t *)filter_x0)[1];
+  filter56 = ((const int32_t *)filter_x0)[2];
+  filter78 = ((const int32_t *)filter_x0)[3];
+
+  for (y = h; y--;) {
+    src = src_ptr;
+    dst = dst_ptr;
+
+    /* prefetch data to cache memory */
+    prefetch_load(src_ptr + src_stride);
+    prefetch_load(src_ptr + src_stride + 32);
+    prefetch_load(src_ptr + src_stride + 64);
+    prefetch_store(dst_ptr + dst_stride);
+    prefetch_store(dst_ptr + dst_stride + 32);
+
+    for (c = 0; c < 4; c++) {
+      __asm__ __volatile__(
+          "ulw              %[qload1],    0(%[src])                    \n\t"
+          "ulw              %[qload2],    4(%[src])                    \n\t"
+
+          /* even 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 1 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 2 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "ulw              %[qload3],    8(%[src])                    \n\t"
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 1 */
+
+          /* even 2. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 3 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "ulw              %[qload1],    12(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[filter12]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter34]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter56]  \n\t" /* even 1 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter78]  \n\t" /* even 1 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 1 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 1 */
+
+          /* even 3. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 4 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st1],       0(%[dst])                    \n\t" /* even 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter12]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter34]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter56]  \n\t" /* even 3 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter78]  \n\t" /* even 3 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 3 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 1 */
+
+          /* even 4. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 5 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st2],       2(%[dst])                    \n\t" /* even 1 */
+          "ulw              %[qload2],    16(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p4],          %[filter12]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter34]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter56]  \n\t" /* even 4 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter78]  \n\t" /* even 4 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 4 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 3 */
+
+          /* even 5. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* even 6 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st3],       4(%[dst])                    \n\t" /* even 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter12]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter34]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter56]  \n\t" /* even 5 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter78]  \n\t" /* even 5 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 5 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 4 */
+
+          /* even 6. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* even 7 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st1],       6(%[dst])                    \n\t" /* even 4 */
+          "ulw              %[qload3],    20(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p5],          %[filter12]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* even 6 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* even 6 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* even 6 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 5 */
+
+          /* even 7. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* even 8 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       8(%[dst])                    \n\t" /* even 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* even 7 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* even 7 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* even 7 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* even 6 */
+
+          /* even 8. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 1 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* even 8 */
+          "sb               %[st3],       10(%[dst])                   \n\t" /* even 6 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* even 8 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* even 8 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* even 8 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* even 7 */
+
+          /* ODD pixels */
+          "ulw              %[qload1],    1(%[src])                    \n\t"
+          "ulw              %[qload2],    5(%[src])                    \n\t"
+
+          /* odd 1. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 2 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload1]                    \n\t"
+          "preceu.ph.qbl    %[p2],        %[qload1]                    \n\t"
+          "preceu.ph.qbr    %[p3],        %[qload2]                    \n\t"
+          "preceu.ph.qbl    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st1],       12(%[dst])                   \n\t" /* even 7 */
+          "ulw              %[qload3],    9(%[src])                    \n\t"
+          "dpa.w.ph         $ac3,         %[p1],          %[filter12]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter34]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter56]  \n\t" /* odd 1 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter78]  \n\t" /* odd 1 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 1 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* even 8 */
+
+          /* odd 2. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 3 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p1],        %[qload3]                    \n\t"
+          "preceu.ph.qbl    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st2],       14(%[dst])                   \n\t" /* even 8 */
+          "ulw              %[qload1],    13(%[src])                   \n\t"
+          "dpa.w.ph         $ac1,         %[p2],          %[filter12]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter34]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter56]  \n\t" /* odd 2 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter78]  \n\t" /* odd 2 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 2 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 1 */
+
+          /* odd 3. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 4 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbr    %[p2],        %[qload1]                    \n\t"
+          "sb               %[st3],       1(%[dst])                    \n\t" /* odd 1 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter12]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter34]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p1],          %[filter56]  \n\t" /* odd 3 */
+          "dpa.w.ph         $ac2,         %[p5],          %[filter78]  \n\t" /* odd 3 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 3 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 2 */
+
+          /* odd 4. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 5 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbl    %[p3],        %[qload1]                    \n\t"
+          "sb               %[st1],       3(%[dst])                    \n\t" /* odd 2 */
+          "ulw              %[qload2],    17(%[src])                   \n\t"
+          "dpa.w.ph         $ac3,         %[p4],          %[filter12]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter34]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p5],          %[filter56]  \n\t" /* odd 4 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter78]  \n\t" /* odd 4 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 4 */
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 3 */
+
+          /* odd 5. pixel */
+          "mtlo             %[vector_64], $ac2                         \n\t" /* odd 6 */
+          "mthi             $zero,        $ac2                         \n\t"
+          "preceu.ph.qbr    %[p4],        %[qload2]                    \n\t"
+          "sb               %[st2],       5(%[dst])                    \n\t" /* odd 3 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter12]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter34]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p2],          %[filter56]  \n\t" /* odd 5 */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter78]  \n\t" /* odd 5 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 5 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 4 */
+
+          /* odd 6. pixel */
+          "mtlo             %[vector_64], $ac3                         \n\t" /* odd 7 */
+          "mthi             $zero,        $ac3                         \n\t"
+          "preceu.ph.qbl    %[p1],        %[qload2]                    \n\t"
+          "sb               %[st3],       7(%[dst])                    \n\t" /* odd 4 */
+          "ulw              %[qload3],    21(%[src])                   \n\t"
+          "dpa.w.ph         $ac2,         %[p5],          %[filter12]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p2],          %[filter34]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p3],          %[filter56]  \n\t" /* odd 6 */
+          "dpa.w.ph         $ac2,         %[p4],          %[filter78]  \n\t" /* odd 6 */
+          "extp             %[Temp2],     $ac2,           31           \n\t" /* odd 6 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 5 */
+
+          /* odd 7. pixel */
+          "mtlo             %[vector_64], $ac1                         \n\t" /* odd 8 */
+          "mthi             $zero,        $ac1                         \n\t"
+          "preceu.ph.qbr    %[p5],        %[qload3]                    \n\t"
+          "sb               %[st1],       9(%[dst])                    \n\t" /* odd 5 */
+          "dpa.w.ph         $ac3,         %[p2],          %[filter12]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p3],          %[filter34]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p4],          %[filter56]  \n\t" /* odd 7 */
+          "dpa.w.ph         $ac3,         %[p1],          %[filter78]  \n\t" /* odd 7 */
+          "extp             %[Temp3],     $ac3,           31           \n\t" /* odd 7 */
+
+          /* odd 8. pixel */
+          "dpa.w.ph         $ac1,         %[p3],          %[filter12]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p4],          %[filter34]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p1],          %[filter56]  \n\t" /* odd 8 */
+          "dpa.w.ph         $ac1,         %[p5],          %[filter78]  \n\t" /* odd 8 */
+          "extp             %[Temp1],     $ac1,           31           \n\t" /* odd 8 */
+
+          "lbux             %[st2],       %[Temp2](%[cm])              \n\t" /* odd 6 */
+          "lbux             %[st3],       %[Temp3](%[cm])              \n\t" /* odd 7 */
+          "lbux             %[st1],       %[Temp1](%[cm])              \n\t" /* odd 8 */
+
+          "sb               %[st2],       11(%[dst])                   \n\t" /* odd 6 */
+          "sb               %[st3],       13(%[dst])                   \n\t" /* odd 7 */
+          "sb               %[st1],       15(%[dst])                   \n\t" /* odd 8 */
+
+          : [qload1] "=&r"(qload1), [qload2] "=&r"(qload2),
+            [qload3] "=&r"(qload3), [st1] "=&r"(st1), [st2] "=&r"(st2),
+            [st3] "=&r"(st3), [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3),
+            [p4] "=&r"(p4), [p5] "=&r"(p5), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [Temp3] "=&r"(Temp3)
+          : [filter12] "r"(filter12), [filter34] "r"(filter34),
+            [filter56] "r"(filter56), [filter78] "r"(filter78),
+            [vector_64] "r"(vector_64), [cm] "r"(cm), [dst] "r"(dst),
+            [src] "r"(src));
+
+      src += 16;
+      dst += 16;
+    }
+
+    /* Next row... */
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+void aom_convolve8_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4, int w,
+                               int h) {
+  assert(x_step_q4 == 16);
+  assert(((const int32_t *)filter_x)[1] != 0x800000);
+
+  if (((const int32_t *)filter_x)[0] == 0) {
+    aom_convolve2_horiz_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                              x_step_q4, filter_y, y_step_q4, w, h);
+  } else {
+    uint32_t pos = 38;
+
+    prefetch_load((const uint8_t *)filter_x);
+    src -= 3;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                         :
+                         : [pos] "r"(pos));
+
+    /* prefetch data to cache memory */
+    prefetch_load(src);
+    prefetch_load(src + 32);
+    prefetch_store(dst);
+
+    switch (w) {
+      case 4:
+        convolve_horiz_4_dspr2(src, (int32_t)src_stride, dst,
+                               (int32_t)dst_stride, filter_x, (int32_t)h);
+        break;
+      case 8:
+        convolve_horiz_8_dspr2(src, (int32_t)src_stride, dst,
+                               (int32_t)dst_stride, filter_x, (int32_t)h);
+        break;
+      case 16:
+        convolve_horiz_16_dspr2(src, (int32_t)src_stride, dst,
+                                (int32_t)dst_stride, filter_x, (int32_t)h, 1);
+        break;
+      case 32:
+        convolve_horiz_16_dspr2(src, (int32_t)src_stride, dst,
+                                (int32_t)dst_stride, filter_x, (int32_t)h, 2);
+        break;
+      case 64:
+        prefetch_load(src + 64);
+        prefetch_store(dst + 32);
+
+        convolve_horiz_64_dspr2(src, (int32_t)src_stride, dst,
+                                (int32_t)dst_stride, filter_x, (int32_t)h);
+        break;
+      default:
+        aom_convolve8_horiz_c(src + 3, src_stride, dst, dst_stride, filter_x,
+                              x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  }
+}
+#endif
diff --git a/third_party/aom/aom_dsp/mips/convolve8_vert_dspr2.c b/third_party/aom/aom_dsp/mips/convolve8_vert_dspr2.c
new file mode 100644
index 0000000000..d8a90b6abd
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve8_vert_dspr2.c
@@ -0,0 +1,360 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/convolve_common_dspr2.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+static void convolve_vert_4_dspr2(const uint8_t *src, int32_t src_stride,
+                                  uint8_t *dst, int32_t dst_stride,
+                                  const int16_t *filter_y, int32_t w,
+                                  int32_t h) {
+  int32_t x, y;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  uint32_t load1, load2, load3, load4;
+  uint32_t p1, p2;
+  uint32_t n1, n2;
+  uint32_t scratch1, scratch2;
+  uint32_t store1, store2;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2;
+
+  vector1b = ((const int32_t *)filter_y)[0];
+  vector2b = ((const int32_t *)filter_y)[1];
+  vector3b = ((const int32_t *)filter_y)[2];
+  vector4b = ((const int32_t *)filter_y)[3];
+
+  src -= 3 * src_stride;
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+
+    for (x = 0; x < w; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__(
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector2b]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector2b]     \n\t"
+
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "dpa.w.ph         $ac2,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+            [load4] "=&r"(load4), [p1] "=&r"(p1), [p2] "=&r"(p2),
+            [n1] "=&r"(n1), [n2] "=&r"(n2), [scratch1] "=&r"(scratch1),
+            [scratch2] "=&r"(scratch2), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [store1] "=&r"(store1),
+            [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr)
+          : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
+            [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
+            [vector4a] "r"(vector4a), [src_stride] "r"(src_stride),
+            [cm] "r"(cm), [dst_ptr] "r"(dst_ptr));
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void convolve_vert_64_dspr2(const uint8_t *src, int32_t src_stride,
+                                   uint8_t *dst, int32_t dst_stride,
+                                   const int16_t *filter_y, int32_t h) {
+  int32_t x, y;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr;
+  uint8_t *cm = aom_ff_cropTbl;
+  uint32_t vector4a = 64;
+  uint32_t load1, load2, load3, load4;
+  uint32_t p1, p2;
+  uint32_t n1, n2;
+  uint32_t scratch1, scratch2;
+  uint32_t store1, store2;
+  int32_t vector1b, vector2b, vector3b, vector4b;
+  int32_t Temp1, Temp2;
+
+  vector1b = ((const int32_t *)filter_y)[0];
+  vector2b = ((const int32_t *)filter_y)[1];
+  vector3b = ((const int32_t *)filter_y)[2];
+  vector4b = ((const int32_t *)filter_y)[3];
+
+  src -= 3 * src_stride;
+
+  for (y = h; y--;) {
+    /* prefetch data to cache memory */
+    prefetch_store(dst + dst_stride);
+    prefetch_store(dst + dst_stride + 32);
+
+    for (x = 0; x < 64; x += 4) {
+      src_ptr = src + x;
+      dst_ptr = dst + x;
+
+      __asm__ __volatile__(
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "mtlo             %[vector4a],  $ac0                            \n\t"
+          "mtlo             %[vector4a],  $ac1                            \n\t"
+          "mtlo             %[vector4a],  $ac2                            \n\t"
+          "mtlo             %[vector4a],  $ac3                            \n\t"
+          "mthi             $zero,        $ac0                            \n\t"
+          "mthi             $zero,        $ac1                            \n\t"
+          "mthi             $zero,        $ac2                            \n\t"
+          "mthi             $zero,        $ac3                            \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector2b]     \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac2,         %[p1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector2b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector1b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector2b]     \n\t"
+
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load1],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load2],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load3],     0(%[src_ptr])                   \n\t"
+          "add              %[src_ptr],   %[src_ptr],     %[src_stride]   \n\t"
+          "ulw              %[load4],     0(%[src_ptr])                   \n\t"
+
+          "preceu.ph.qbr    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbr    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbr    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbr    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "dpa.w.ph         $ac0,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac0,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac0,           31              \n\t"
+          "dpa.w.ph         $ac1,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac1,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac1,           31              \n\t"
+
+          "preceu.ph.qbl    %[scratch1],  %[load1]                        \n\t"
+          "preceu.ph.qbl    %[p1],        %[load2]                        \n\t"
+          "precrq.ph.w      %[n1],        %[p1],          %[scratch1]     \n\t" /* pixel 2 */
+          "append           %[p1],        %[scratch1],    16              \n\t" /* pixel 1 */
+          "preceu.ph.qbl    %[scratch2],  %[load3]                        \n\t"
+          "preceu.ph.qbl    %[p2],        %[load4]                        \n\t"
+          "precrq.ph.w      %[n2],        %[p2],          %[scratch2]     \n\t" /* pixel 2 */
+          "append           %[p2],        %[scratch2],    16              \n\t" /* pixel 1 */
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "dpa.w.ph         $ac2,         %[p1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac2,         %[p2],          %[vector4b]     \n\t"
+          "extp             %[Temp1],     $ac2,           31              \n\t"
+
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+          "dpa.w.ph         $ac3,         %[n1],          %[vector3b]     \n\t"
+          "dpa.w.ph         $ac3,         %[n2],          %[vector4b]     \n\t"
+          "extp             %[Temp2],     $ac3,           31              \n\t"
+
+          "sb               %[store1],    0(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    1(%[dst_ptr])                   \n\t"
+
+          "lbux             %[store1],    %[Temp1](%[cm])                 \n\t"
+          "lbux             %[store2],    %[Temp2](%[cm])                 \n\t"
+
+          "sb               %[store1],    2(%[dst_ptr])                   \n\t"
+          "sb               %[store2],    3(%[dst_ptr])                   \n\t"
+
+          : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+            [load4] "=&r"(load4), [p1] "=&r"(p1), [p2] "=&r"(p2),
+            [n1] "=&r"(n1), [n2] "=&r"(n2), [scratch1] "=&r"(scratch1),
+            [scratch2] "=&r"(scratch2), [Temp1] "=&r"(Temp1),
+            [Temp2] "=&r"(Temp2), [store1] "=&r"(store1),
+            [store2] "=&r"(store2), [src_ptr] "+r"(src_ptr)
+          : [vector1b] "r"(vector1b), [vector2b] "r"(vector2b),
+            [vector3b] "r"(vector3b), [vector4b] "r"(vector4b),
+            [vector4a] "r"(vector4a), [src_stride] "r"(src_stride),
+            [cm] "r"(cm), [dst_ptr] "r"(dst_ptr));
+    }
+
+    /* Next row... */
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void aom_convolve8_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4, int w,
+                              int h) {
+  assert(y_step_q4 == 16);
+  assert(((const int32_t *)filter_y)[1] != 0x800000);
+
+  if (((const int32_t *)filter_y)[0] == 0) {
+    aom_convolve2_vert_dspr2(src, src_stride, dst, dst_stride, filter_x,
+                             x_step_q4, filter_y, y_step_q4, w, h);
+  } else {
+    uint32_t pos = 38;
+
+    /* bit positon for extract from acc */
+    __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                         :
+                         : [pos] "r"(pos));
+
+    prefetch_store(dst);
+
+    switch (w) {
+      case 4:
+      case 8:
+      case 16:
+      case 32:
+        convolve_vert_4_dspr2(src, src_stride, dst, dst_stride, filter_y, w, h);
+        break;
+      case 64:
+        prefetch_store(dst + 32);
+        convolve_vert_64_dspr2(src, src_stride, dst, dst_stride, filter_y, h);
+        break;
+      default:
+        aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x,
+                             x_step_q4, filter_y, y_step_q4, w, h);
+        break;
+    }
+  }
+}
+
+#endif
diff --git a/third_party/aom/aom_dsp/mips/convolve_common_dspr2.h b/third_party/aom/aom_dsp/mips/convolve_common_dspr2.h
new file mode 100644
index 0000000000..f8fd9e2b6c
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/convolve_common_dspr2.h
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_MIPS_AOM_COMMON_DSPR2_H_
+#define AOM_DSP_MIPS_AOM_COMMON_DSPR2_H_
+
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/mips/common_dspr2.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if HAVE_DSPR2
+void aom_convolve2_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4, int w,
+                               int h);
+
+void aom_convolve2_avg_horiz_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int x_step_q4,
+                                   const int16_t *filter_y, int y_step_q4,
+                                   int w, int h);
+
+void aom_convolve2_avg_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                                  uint8_t *dst, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int x_step_q4,
+                                  const int16_t *filter_y, int y_step_q4, int w,
+                                  int h);
+
+void aom_convolve2_dspr2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const int16_t *filter, int w,
+                         int h);
+
+void aom_convolve2_vert_dspr2(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4, int w,
+                              int h);
+
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_MIPS_AOM_COMMON_DSPR2_H_
diff --git a/third_party/aom/aom_dsp/mips/fwd_dct32x32_msa.c b/third_party/aom/aom_dsp/mips/fwd_dct32x32_msa.c
new file mode 100644
index 0000000000..dc9c63226a
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/fwd_dct32x32_msa.c
@@ -0,0 +1,948 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/fwd_txfm_msa.h"
+
+static void fdct8x32_1d_column_load_butterfly(const int16_t *input,
+                                              int32_t src_stride,
+                                              int16_t *temp_buff) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 step0, step1, step2, step3;
+  v8i16 in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1;
+  v8i16 step0_1, step1_1, step2_1, step3_1;
+
+  /* 1st and 2nd set */
+  LD_SH4(input, src_stride, in0, in1, in2, in3);
+  LD_SH4(input + (28 * src_stride), src_stride, in4, in5, in6, in7);
+  LD_SH4(input + (4 * src_stride), src_stride, in0_1, in1_1, in2_1, in3_1);
+  LD_SH4(input + (24 * src_stride), src_stride, in4_1, in5_1, in6_1, in7_1);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+  SLLI_4V(in0_1, in1_1, in2_1, in3_1, 2);
+  SLLI_4V(in4_1, in5_1, in6_1, in7_1, 2);
+  BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7, step0, step1, step2,
+              step3, in4, in5, in6, in7);
+  BUTTERFLY_8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1, step0_1,
+              step1_1, step2_1, step3_1, in4_1, in5_1, in6_1, in7_1);
+  ST_SH4(step0, step1, step2, step3, temp_buff, 8);
+  ST_SH4(in4, in5, in6, in7, temp_buff + (28 * 8), 8);
+  ST_SH4(step0_1, step1_1, step2_1, step3_1, temp_buff + (4 * 8), 8);
+  ST_SH4(in4_1, in5_1, in6_1, in7_1, temp_buff + (24 * 8), 8);
+
+  /* 3rd and 4th set */
+  LD_SH4(input + (8 * src_stride), src_stride, in0, in1, in2, in3);
+  LD_SH4(input + (20 * src_stride), src_stride, in4, in5, in6, in7);
+  LD_SH4(input + (12 * src_stride), src_stride, in0_1, in1_1, in2_1, in3_1);
+  LD_SH4(input + (16 * src_stride), src_stride, in4_1, in5_1, in6_1, in7_1);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+  SLLI_4V(in0_1, in1_1, in2_1, in3_1, 2);
+  SLLI_4V(in4_1, in5_1, in6_1, in7_1, 2);
+  BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7, step0, step1, step2,
+              step3, in4, in5, in6, in7);
+  BUTTERFLY_8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1, step0_1,
+              step1_1, step2_1, step3_1, in4_1, in5_1, in6_1, in7_1);
+  ST_SH4(step0, step1, step2, step3, temp_buff + (8 * 8), 8);
+  ST_SH4(in4, in5, in6, in7, temp_buff + (20 * 8), 8);
+  ST_SH4(step0_1, step1_1, step2_1, step3_1, temp_buff + (12 * 8), 8);
+  ST_SH4(in4_1, in5_1, in6_1, in7_1, temp_buff + (15 * 8) + 8, 8);
+}
+
+static void fdct8x32_1d_column_even_store(int16_t *input, int16_t *temp) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8i16 temp0, temp1;
+
+  /* fdct even */
+  LD_SH4(input, 8, in0, in1, in2, in3);
+  LD_SH4(input + 96, 8, in12, in13, in14, in15);
+  BUTTERFLY_8(in0, in1, in2, in3, in12, in13, in14, in15, vec0, vec1, vec2,
+              vec3, in12, in13, in14, in15);
+  LD_SH4(input + 32, 8, in4, in5, in6, in7);
+  LD_SH4(input + 64, 8, in8, in9, in10, in11);
+  BUTTERFLY_8(in4, in5, in6, in7, in8, in9, in10, in11, vec4, vec5, vec6, vec7,
+              in8, in9, in10, in11);
+
+  /* Stage 3 */
+  ADD4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, in0, in1, in2, in3);
+  BUTTERFLY_4(in0, in1, in2, in3, temp0, in4, in1, in0);
+  DOTP_CONST_PAIR(temp0, in4, cospi_16_64, cospi_16_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp);
+  ST_SH(temp1, temp + 512);
+
+  DOTP_CONST_PAIR(in0, in1, cospi_24_64, cospi_8_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 256);
+  ST_SH(temp1, temp + 768);
+
+  SUB4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, vec7, vec6, vec5, vec4);
+  DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+  ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 128);
+  ST_SH(temp1, temp + 896);
+
+  SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+  DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 640);
+  ST_SH(temp1, temp + 384);
+
+  DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+  DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+  ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+  DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+  ADD2(in0, in1, in2, in3, vec0, vec7);
+  DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 64);
+  ST_SH(temp1, temp + 960);
+
+  SUB2(in0, in1, in2, in3, in0, in2);
+  DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 576);
+  ST_SH(temp1, temp + 448);
+
+  SUB2(in9, vec2, in14, vec5, vec2, vec5);
+  DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+  SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5);
+  DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, temp1, temp0);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 320);
+  ST_SH(temp1, temp + 704);
+
+  ADD2(in3, in2, in0, in1, vec3, vec4);
+  DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, temp0, temp1);
+  FDCT32_POSTPROC_2V_POS_H(temp0, temp1);
+  ST_SH(temp0, temp + 192);
+  ST_SH(temp1, temp + 832);
+}
+
+static void fdct8x32_1d_column_odd_store(int16_t *input, int16_t *temp_ptr) {
+  v8i16 in16, in17, in18, in19, in20, in21, in22, in23;
+  v8i16 in24, in25, in26, in27, in28, in29, in30, in31, vec4, vec5;
+
+  in20 = LD_SH(input + 32);
+  in21 = LD_SH(input + 40);
+  in26 = LD_SH(input + 80);
+  in27 = LD_SH(input + 88);
+
+  DOTP_CONST_PAIR(in27, in20, cospi_16_64, cospi_16_64, in20, in27);
+  DOTP_CONST_PAIR(in26, in21, cospi_16_64, cospi_16_64, in21, in26);
+
+  in18 = LD_SH(input + 16);
+  in19 = LD_SH(input + 24);
+  in28 = LD_SH(input + 96);
+  in29 = LD_SH(input + 104);
+
+  vec4 = in19 - in20;
+  ST_SH(vec4, input + 32);
+  vec4 = in18 - in21;
+  ST_SH(vec4, input + 40);
+  vec4 = in29 - in26;
+  ST_SH(vec4, input + 80);
+  vec4 = in28 - in27;
+  ST_SH(vec4, input + 88);
+
+  in21 = in18 + in21;
+  in20 = in19 + in20;
+  in27 = in28 + in27;
+  in26 = in29 + in26;
+
+  LD_SH4(input + 48, 8, in22, in23, in24, in25);
+  DOTP_CONST_PAIR(in25, in22, cospi_16_64, cospi_16_64, in22, in25);
+  DOTP_CONST_PAIR(in24, in23, cospi_16_64, cospi_16_64, in23, in24);
+
+  in16 = LD_SH(input);
+  in17 = LD_SH(input + 8);
+  in30 = LD_SH(input + 112);
+  in31 = LD_SH(input + 120);
+
+  vec4 = in17 - in22;
+  ST_SH(vec4, input + 16);
+  vec4 = in16 - in23;
+  ST_SH(vec4, input + 24);
+  vec4 = in31 - in24;
+  ST_SH(vec4, input + 96);
+  vec4 = in30 - in25;
+  ST_SH(vec4, input + 104);
+
+  ADD4(in16, in23, in17, in22, in30, in25, in31, in24, in16, in17, in30, in31);
+  DOTP_CONST_PAIR(in26, in21, cospi_24_64, cospi_8_64, in18, in29);
+  DOTP_CONST_PAIR(in27, in20, cospi_24_64, cospi_8_64, in19, in28);
+  ADD4(in16, in19, in17, in18, in30, in29, in31, in28, in27, in22, in21, in25);
+  DOTP_CONST_PAIR(in21, in22, cospi_28_64, cospi_4_64, in26, in24);
+  ADD2(in27, in26, in25, in24, in23, in20);
+  DOTP_CONST_PAIR(in20, in23, cospi_31_64, cospi_1_64, vec4, vec5);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr);
+  ST_SH(vec4, temp_ptr + 960);
+
+  SUB2(in27, in26, in25, in24, in22, in21);
+  DOTP_CONST_PAIR(in21, in22, cospi_15_64, cospi_17_64, vec5, vec4);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr + 448);
+  ST_SH(vec4, temp_ptr + 512);
+
+  SUB4(in17, in18, in16, in19, in31, in28, in30, in29, in23, in26, in24, in20);
+  DOTP_CONST_PAIR((-in23), in20, cospi_28_64, cospi_4_64, in27, in25);
+  SUB2(in26, in27, in24, in25, in23, in20);
+  DOTP_CONST_PAIR(in20, in23, cospi_23_64, cospi_9_64, vec4, vec5);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec4, temp_ptr + 704);
+  ST_SH(vec5, temp_ptr + 256);
+
+  ADD2(in26, in27, in24, in25, in22, in21);
+  DOTP_CONST_PAIR(in21, in22, cospi_7_64, cospi_25_64, vec4, vec5);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec4, temp_ptr + 192);
+  ST_SH(vec5, temp_ptr + 768);
+
+  LD_SH4(input + 16, 8, in22, in23, in20, in21);
+  LD_SH4(input + 80, 8, in26, in27, in24, in25);
+  in16 = in20;
+  in17 = in21;
+  DOTP_CONST_PAIR(-in16, in27, cospi_24_64, cospi_8_64, in20, in27);
+  DOTP_CONST_PAIR(-in17, in26, cospi_24_64, cospi_8_64, in21, in26);
+  SUB4(in23, in20, in22, in21, in25, in26, in24, in27, in28, in17, in18, in31);
+  DOTP_CONST_PAIR(in18, in17, cospi_12_64, cospi_20_64, in29, in30);
+  ADD2(in28, in29, in31, in30, in16, in19);
+  DOTP_CONST_PAIR(in19, in16, cospi_27_64, cospi_5_64, vec5, vec4);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr + 832);
+  ST_SH(vec4, temp_ptr + 128);
+
+  SUB2(in28, in29, in31, in30, in17, in18);
+  DOTP_CONST_PAIR(in18, in17, cospi_11_64, cospi_21_64, vec5, vec4);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr + 320);
+  ST_SH(vec4, temp_ptr + 640);
+  ADD4(in22, in21, in23, in20, in24, in27, in25, in26, in16, in29, in30, in19);
+  DOTP_CONST_PAIR(-in16, in19, cospi_12_64, cospi_20_64, in28, in31);
+  SUB2(in29, in28, in30, in31, in16, in19);
+  DOTP_CONST_PAIR(in19, in16, cospi_19_64, cospi_13_64, vec5, vec4);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr + 576);
+  ST_SH(vec4, temp_ptr + 384);
+
+  ADD2(in29, in28, in30, in31, in17, in18);
+  DOTP_CONST_PAIR(in18, in17, cospi_3_64, cospi_29_64, vec5, vec4);
+  FDCT32_POSTPROC_2V_POS_H(vec5, vec4);
+  ST_SH(vec5, temp_ptr + 64);
+  ST_SH(vec4, temp_ptr + 896);
+}
+
+static void fdct8x32_1d_column(const int16_t *input, int32_t src_stride,
+                               int16_t *tmp_buf, int16_t *tmp_buf_big) {
+  fdct8x32_1d_column_load_butterfly(input, src_stride, tmp_buf);
+  fdct8x32_1d_column_even_store(tmp_buf, tmp_buf_big);
+  fdct8x32_1d_column_odd_store(tmp_buf + 128, (tmp_buf_big + 32));
+}
+
+static void fdct8x32_1d_row_load_butterfly(int16_t *temp_buff,
+                                           int16_t *output) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 step0, step1, step2, step3, step4, step5, step6, step7;
+
+  LD_SH8(temp_buff, 32, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8(temp_buff + 24, 32, in8, in9, in10, in11, in12, in13, in14, in15);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9,
+                     in10, in11, in12, in13, in14, in15);
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,
+               in12, in13, in14, in15, step0, step1, step2, step3, step4, step5,
+               step6, step7, in8, in9, in10, in11, in12, in13, in14, in15);
+  ST_SH8(step0, step1, step2, step3, step4, step5, step6, step7, output, 8);
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, (output + 24 * 8), 8);
+
+  /* 2nd set */
+  LD_SH8(temp_buff + 8, 32, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8(temp_buff + 16, 32, in8, in9, in10, in11, in12, in13, in14, in15);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9,
+                     in10, in11, in12, in13, in14, in15);
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,
+               in12, in13, in14, in15, step0, step1, step2, step3, step4, step5,
+               step6, step7, in8, in9, in10, in11, in12, in13, in14, in15);
+  ST_SH8(step0, step1, step2, step3, step4, step5, step6, step7,
+         (output + 8 * 8), 8);
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, (output + 16 * 8), 8);
+}
+
+static void fdct8x32_1d_row_even_4x(int16_t *input, int16_t *interm_ptr,
+                                    int16_t *out) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v4i32 vec0_l, vec1_l, vec2_l, vec3_l, vec4_l, vec5_l, vec6_l, vec7_l;
+  v4i32 vec0_r, vec1_r, vec2_r, vec3_r, vec4_r, vec5_r, vec6_r, vec7_r;
+  v4i32 tmp0_w, tmp1_w, tmp2_w, tmp3_w;
+
+  /* fdct32 even */
+  /* stage 2 */
+  LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8(input + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,
+               in12, in13, in14, in15, vec0, vec1, vec2, vec3, vec4, vec5, vec6,
+               vec7, in8, in9, in10, in11, in12, in13, in14, in15);
+  ST_SH8(vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, interm_ptr, 8);
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, interm_ptr + 64, 8);
+
+  /* Stage 3 */
+  UNPCK_SH_SW(vec0, vec0_l, vec0_r);
+  UNPCK_SH_SW(vec1, vec1_l, vec1_r);
+  UNPCK_SH_SW(vec2, vec2_l, vec2_r);
+  UNPCK_SH_SW(vec3, vec3_l, vec3_r);
+  UNPCK_SH_SW(vec4, vec4_l, vec4_r);
+  UNPCK_SH_SW(vec5, vec5_l, vec5_r);
+  UNPCK_SH_SW(vec6, vec6_l, vec6_r);
+  UNPCK_SH_SW(vec7, vec7_l, vec7_r);
+  ADD4(vec0_r, vec7_r, vec1_r, vec6_r, vec2_r, vec5_r, vec3_r, vec4_r, tmp0_w,
+       tmp1_w, tmp2_w, tmp3_w);
+  BUTTERFLY_4(tmp0_w, tmp1_w, tmp2_w, tmp3_w, vec4_r, vec6_r, vec7_r, vec5_r);
+  ADD4(vec0_l, vec7_l, vec1_l, vec6_l, vec2_l, vec5_l, vec3_l, vec4_l, vec0_r,
+       vec1_r, vec2_r, vec3_r);
+
+  tmp3_w = vec0_r + vec3_r;
+  vec0_r = vec0_r - vec3_r;
+  vec3_r = vec1_r + vec2_r;
+  vec1_r = vec1_r - vec2_r;
+
+  DOTP_CONST_PAIR_W(vec4_r, vec6_r, tmp3_w, vec3_r, cospi_16_64, cospi_16_64,
+                    vec4_r, tmp3_w, vec6_r, vec3_r);
+  FDCT32_POSTPROC_NEG_W(vec4_r);
+  FDCT32_POSTPROC_NEG_W(tmp3_w);
+  FDCT32_POSTPROC_NEG_W(vec6_r);
+  FDCT32_POSTPROC_NEG_W(vec3_r);
+  PCKEV_H2_SH(vec4_r, tmp3_w, vec6_r, vec3_r, vec4, vec5);
+  ST_SH2(vec5, vec4, out, 8);
+
+  DOTP_CONST_PAIR_W(vec5_r, vec7_r, vec0_r, vec1_r, cospi_24_64, cospi_8_64,
+                    vec4_r, tmp3_w, vec6_r, vec3_r);
+  FDCT32_POSTPROC_NEG_W(vec4_r);
+  FDCT32_POSTPROC_NEG_W(tmp3_w);
+  FDCT32_POSTPROC_NEG_W(vec6_r);
+  FDCT32_POSTPROC_NEG_W(vec3_r);
+  PCKEV_H2_SH(vec4_r, tmp3_w, vec6_r, vec3_r, vec4, vec5);
+  ST_SH2(vec5, vec4, out + 16, 8);
+
+  LD_SH8(interm_ptr, 8, vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7);
+  SUB4(vec3, vec4, vec2, vec5, vec1, vec6, vec0, vec7, vec4, vec5, vec6, vec7);
+  DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+  ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, in5, in4);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 32);
+  ST_SH(in5, out + 56);
+
+  SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+  DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, in5, in4);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 40);
+  ST_SH(in5, out + 48);
+
+  LD_SH8(interm_ptr + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+  DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+  DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+  ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+  DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+  ADD2(in0, in1, in2, in3, vec0, vec7);
+  DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, in5, in4);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 64);
+  ST_SH(in5, out + 120);
+
+  SUB2(in0, in1, in2, in3, in0, in2);
+  DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, in5, in4);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 72);
+  ST_SH(in5, out + 112);
+
+  SUB2(in9, vec2, in14, vec5, vec2, vec5);
+  DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+  SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5);
+  DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, in5, in4);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 80);
+  ST_SH(in5, out + 104);
+
+  ADD2(in3, in2, in0, in1, vec3, vec4);
+  DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, in4, in5);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  ST_SH(in4, out + 96);
+  ST_SH(in5, out + 88);
+}
+
+static void fdct8x32_1d_row_even(int16_t *temp, int16_t *out) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, temp0, temp1;
+
+  /* fdct32 even */
+  /* stage 2 */
+  LD_SH8(temp, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8(temp + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,
+               in12, in13, in14, in15, vec0, vec1, vec2, vec3, vec4, vec5, vec6,
+               vec7, in8, in9, in10, in11, in12, in13, in14, in15);
+
+  /* Stage 3 */
+  ADD4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, in0, in1, in2, in3);
+  BUTTERFLY_4(in0, in1, in2, in3, temp0, in4, in1, in0);
+  DOTP_CONST_PAIR(temp0, in4, cospi_16_64, cospi_16_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out);
+  ST_SH(temp1, out + 8);
+
+  DOTP_CONST_PAIR(in0, in1, cospi_24_64, cospi_8_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 16);
+  ST_SH(temp1, out + 24);
+
+  SUB4(vec3, vec4, vec2, vec5, vec1, vec6, vec0, vec7, vec4, vec5, vec6, vec7);
+  DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+  ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 32);
+  ST_SH(temp1, out + 56);
+
+  SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+  DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 40);
+  ST_SH(temp1, out + 48);
+
+  DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+  DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+  ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+  DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+  ADD2(in0, in1, in2, in3, vec0, vec7);
+  DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 64);
+  ST_SH(temp1, out + 120);
+
+  SUB2(in0, in1, in2, in3, in0, in2);
+  DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 72);
+  ST_SH(temp1, out + 112);
+
+  SUB2(in9, vec2, in14, vec5, vec2, vec5);
+  DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+  SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5)
+  DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, temp1, temp0);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 80);
+  ST_SH(temp1, out + 104);
+
+  ADD2(in3, in2, in0, in1, vec3, vec4);
+  DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, temp0, temp1);
+  FDCT_POSTPROC_2V_NEG_H(temp0, temp1);
+  ST_SH(temp0, out + 96);
+  ST_SH(temp1, out + 88);
+}
+
+static void fdct8x32_1d_row_odd(int16_t *temp, int16_t *interm_ptr,
+                                int16_t *out) {
+  v8i16 in16, in17, in18, in19, in20, in21, in22, in23;
+  v8i16 in24, in25, in26, in27, in28, in29, in30, in31, vec4, vec5;
+
+  in20 = LD_SH(temp + 32);
+  in21 = LD_SH(temp + 40);
+  in26 = LD_SH(temp + 80);
+  in27 = LD_SH(temp + 88);
+
+  DOTP_CONST_PAIR(in27, in20, cospi_16_64, cospi_16_64, in20, in27);
+  DOTP_CONST_PAIR(in26, in21, cospi_16_64, cospi_16_64, in21, in26);
+
+  in18 = LD_SH(temp + 16);
+  in19 = LD_SH(temp + 24);
+  in28 = LD_SH(temp + 96);
+  in29 = LD_SH(temp + 104);
+
+  vec4 = in19 - in20;
+  ST_SH(vec4, interm_ptr + 32);
+  vec4 = in18 - in21;
+  ST_SH(vec4, interm_ptr + 88);
+  vec4 = in28 - in27;
+  ST_SH(vec4, interm_ptr + 56);
+  vec4 = in29 - in26;
+  ST_SH(vec4, interm_ptr + 64);
+
+  ADD4(in18, in21, in19, in20, in28, in27, in29, in26, in21, in20, in27, in26);
+
+  in22 = LD_SH(temp + 48);
+  in23 = LD_SH(temp + 56);
+  in24 = LD_SH(temp + 64);
+  in25 = LD_SH(temp + 72);
+
+  DOTP_CONST_PAIR(in25, in22, cospi_16_64, cospi_16_64, in22, in25);
+  DOTP_CONST_PAIR(in24, in23, cospi_16_64, cospi_16_64, in23, in24);
+
+  in16 = LD_SH(temp);
+  in17 = LD_SH(temp + 8);
+  in30 = LD_SH(temp + 112);
+  in31 = LD_SH(temp + 120);
+
+  vec4 = in17 - in22;
+  ST_SH(vec4, interm_ptr + 40);
+  vec4 = in30 - in25;
+  ST_SH(vec4, interm_ptr + 48);
+  vec4 = in31 - in24;
+  ST_SH(vec4, interm_ptr + 72);
+  vec4 = in16 - in23;
+  ST_SH(vec4, interm_ptr + 80);
+
+  ADD4(in16, in23, in17, in22, in30, in25, in31, in24, in16, in17, in30, in31);
+  DOTP_CONST_PAIR(in26, in21, cospi_24_64, cospi_8_64, in18, in29);
+  DOTP_CONST_PAIR(in27, in20, cospi_24_64, cospi_8_64, in19, in28);
+
+  ADD4(in16, in19, in17, in18, in30, in29, in31, in28, in27, in22, in21, in25);
+  DOTP_CONST_PAIR(in21, in22, cospi_28_64, cospi_4_64, in26, in24);
+  ADD2(in27, in26, in25, in24, in23, in20);
+
+  DOTP_CONST_PAIR(in20, in23, cospi_31_64, cospi_1_64, vec4, vec5);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec5, out);
+  ST_SH(vec4, out + 120);
+
+  SUB2(in27, in26, in25, in24, in22, in21);
+
+  DOTP_CONST_PAIR(in21, in22, cospi_15_64, cospi_17_64, vec5, vec4);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec5, out + 112);
+  ST_SH(vec4, out + 8);
+
+  SUB4(in17, in18, in16, in19, in31, in28, in30, in29, in23, in26, in24, in20);
+  DOTP_CONST_PAIR((-in23), in20, cospi_28_64, cospi_4_64, in27, in25);
+  SUB2(in26, in27, in24, in25, in23, in20);
+
+  DOTP_CONST_PAIR(in20, in23, cospi_23_64, cospi_9_64, vec4, vec5);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec4, out + 16);
+  ST_SH(vec5, out + 104);
+
+  ADD2(in26, in27, in24, in25, in22, in21);
+  DOTP_CONST_PAIR(in21, in22, cospi_7_64, cospi_25_64, vec4, vec5);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec4, out + 24);
+  ST_SH(vec5, out + 96);
+
+  in20 = LD_SH(interm_ptr + 32);
+  in21 = LD_SH(interm_ptr + 88);
+  in27 = LD_SH(interm_ptr + 56);
+  in26 = LD_SH(interm_ptr + 64);
+
+  in16 = in20;
+  in17 = in21;
+  DOTP_CONST_PAIR(-in16, in27, cospi_24_64, cospi_8_64, in20, in27);
+  DOTP_CONST_PAIR(-in17, in26, cospi_24_64, cospi_8_64, in21, in26);
+
+  in22 = LD_SH(interm_ptr + 40);
+  in25 = LD_SH(interm_ptr + 48);
+  in24 = LD_SH(interm_ptr + 72);
+  in23 = LD_SH(interm_ptr + 80);
+
+  SUB4(in23, in20, in22, in21, in25, in26, in24, in27, in28, in17, in18, in31);
+  DOTP_CONST_PAIR(in18, in17, cospi_12_64, cospi_20_64, in29, in30);
+  ADD2(in28, in29, in31, in30, in16, in19);
+  DOTP_CONST_PAIR(in19, in16, cospi_27_64, cospi_5_64, vec5, vec4);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec5, out + 32);
+  ST_SH(vec4, out + 88);
+
+  SUB2(in28, in29, in31, in30, in17, in18);
+  DOTP_CONST_PAIR(in18, in17, cospi_11_64, cospi_21_64, vec5, vec4);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec5, out + 40);
+  ST_SH(vec4, out + 80);
+
+  ADD4(in22, in21, in23, in20, in24, in27, in25, in26, in16, in29, in30, in19);
+  DOTP_CONST_PAIR(-in16, in19, cospi_12_64, cospi_20_64, in28, in31);
+  SUB2(in29, in28, in30, in31, in16, in19);
+
+  DOTP_CONST_PAIR(in19, in16, cospi_19_64, cospi_13_64, vec5, vec4);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec5, out + 72);
+  ST_SH(vec4, out + 48);
+
+  ADD2(in29, in28, in30, in31, in17, in18);
+
+  DOTP_CONST_PAIR(in18, in17, cospi_3_64, cospi_29_64, vec5, vec4);
+  FDCT_POSTPROC_2V_NEG_H(vec5, vec4);
+  ST_SH(vec4, out + 56);
+  ST_SH(vec5, out + 64);
+}
+
+static void fdct8x32_1d_row_transpose_store(int16_t *temp, int16_t *output) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1;
+
+  /* 1st set */
+  in0 = LD_SH(temp);
+  in4 = LD_SH(temp + 32);
+  in2 = LD_SH(temp + 64);
+  in6 = LD_SH(temp + 96);
+  in1 = LD_SH(temp + 128);
+  in7 = LD_SH(temp + 152);
+  in3 = LD_SH(temp + 192);
+  in5 = LD_SH(temp + 216);
+
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+
+  /* 2nd set */
+  in0_1 = LD_SH(temp + 16);
+  in1_1 = LD_SH(temp + 232);
+  in2_1 = LD_SH(temp + 80);
+  in3_1 = LD_SH(temp + 168);
+  in4_1 = LD_SH(temp + 48);
+  in5_1 = LD_SH(temp + 176);
+  in6_1 = LD_SH(temp + 112);
+  in7_1 = LD_SH(temp + 240);
+
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output, 32);
+  TRANSPOSE8x8_SH_SH(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+                     in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1);
+
+  /* 3rd set */
+  in0 = LD_SH(temp + 8);
+  in1 = LD_SH(temp + 136);
+  in2 = LD_SH(temp + 72);
+  in3 = LD_SH(temp + 200);
+  in4 = LD_SH(temp + 40);
+  in5 = LD_SH(temp + 208);
+  in6 = LD_SH(temp + 104);
+  in7 = LD_SH(temp + 144);
+
+  ST_SH8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1, output + 8,
+         32);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output + 16, 32);
+
+  /* 4th set */
+  in0_1 = LD_SH(temp + 24);
+  in1_1 = LD_SH(temp + 224);
+  in2_1 = LD_SH(temp + 88);
+  in3_1 = LD_SH(temp + 160);
+  in4_1 = LD_SH(temp + 56);
+  in5_1 = LD_SH(temp + 184);
+  in6_1 = LD_SH(temp + 120);
+  in7_1 = LD_SH(temp + 248);
+
+  TRANSPOSE8x8_SH_SH(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1,
+                     in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1);
+  ST_SH8(in0_1, in1_1, in2_1, in3_1, in4_1, in5_1, in6_1, in7_1, output + 24,
+         32);
+}
+
+static void fdct32x8_1d_row(int16_t *temp, int16_t *temp_buf, int16_t *output) {
+  fdct8x32_1d_row_load_butterfly(temp, temp_buf);
+  fdct8x32_1d_row_even(temp_buf, temp_buf);
+  fdct8x32_1d_row_odd(temp_buf + 128, temp, temp_buf + 128);
+  fdct8x32_1d_row_transpose_store(temp_buf, output);
+}
+
+static void fdct32x8_1d_row_4x(int16_t *tmp_buf_big, int16_t *tmp_buf,
+                               int16_t *output) {
+  fdct8x32_1d_row_load_butterfly(tmp_buf_big, tmp_buf);
+  fdct8x32_1d_row_even_4x(tmp_buf, tmp_buf_big, tmp_buf);
+  fdct8x32_1d_row_odd(tmp_buf + 128, tmp_buf_big, tmp_buf + 128);
+  fdct8x32_1d_row_transpose_store(tmp_buf, output);
+}
+
+void aom_fdct32x32_msa(const int16_t *input, int16_t *output,
+                       int32_t src_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, tmp_buf_big[1024]);
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[256]);
+
+  /* column transform */
+  for (i = 0; i < 4; ++i) {
+    fdct8x32_1d_column(input + (8 * i), src_stride, tmp_buf,
+                       tmp_buf_big + (8 * i));
+  }
+
+  /* row transform */
+  fdct32x8_1d_row_4x(tmp_buf_big, tmp_buf, output);
+
+  /* row transform */
+  for (i = 1; i < 4; ++i) {
+    fdct32x8_1d_row(tmp_buf_big + (i * 256), tmp_buf, output + (i * 256));
+  }
+}
+
+static void fdct8x32_1d_row_even_rd(int16_t *temp, int16_t *out) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, temp0, temp1;
+
+  /* fdct32 even */
+  /* stage 2 */
+  LD_SH8(temp, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8(temp + 64, 8, in8, in9, in10, in11, in12, in13, in14, in15);
+
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,
+               in12, in13, in14, in15, vec0, vec1, vec2, vec3, vec4, vec5, vec6,
+               vec7, in8, in9, in10, in11, in12, in13, in14, in15);
+  FDCT_POSTPROC_2V_NEG_H(vec0, vec1);
+  FDCT_POSTPROC_2V_NEG_H(vec2, vec3);
+  FDCT_POSTPROC_2V_NEG_H(vec4, vec5);
+  FDCT_POSTPROC_2V_NEG_H(vec6, vec7);
+  FDCT_POSTPROC_2V_NEG_H(in8, in9);
+  FDCT_POSTPROC_2V_NEG_H(in10, in11);
+  FDCT_POSTPROC_2V_NEG_H(in12, in13);
+  FDCT_POSTPROC_2V_NEG_H(in14, in15);
+
+  /* Stage 3 */
+  ADD4(vec0, vec7, vec1, vec6, vec2, vec5, vec3, vec4, in0, in1, in2, in3);
+
+  temp0 = in0 + in3;
+  in0 = in0 - in3;
+  in3 = in1 + in2;
+  in1 = in1 - in2;
+
+  DOTP_CONST_PAIR(temp0, in3, cospi_16_64, cospi_16_64, temp1, temp0);
+  ST_SH(temp0, out);
+  ST_SH(temp1, out + 8);
+
+  DOTP_CONST_PAIR(in0, in1, cospi_24_64, cospi_8_64, temp1, temp0);
+  ST_SH(temp0, out + 16);
+  ST_SH(temp1, out + 24);
+
+  SUB4(vec3, vec4, vec2, vec5, vec1, vec6, vec0, vec7, vec4, vec5, vec6, vec7);
+  DOTP_CONST_PAIR(vec6, vec5, cospi_16_64, cospi_16_64, vec5, vec6);
+  ADD2(vec4, vec5, vec7, vec6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_28_64, cospi_4_64, temp1, temp0);
+  ST_SH(temp0, out + 32);
+  ST_SH(temp1, out + 56);
+
+  SUB2(vec4, vec5, vec7, vec6, vec4, vec7);
+  DOTP_CONST_PAIR(vec7, vec4, cospi_12_64, cospi_20_64, temp1, temp0);
+  ST_SH(temp0, out + 40);
+  ST_SH(temp1, out + 48);
+
+  DOTP_CONST_PAIR(in13, in10, cospi_16_64, cospi_16_64, vec2, vec5);
+  DOTP_CONST_PAIR(in12, in11, cospi_16_64, cospi_16_64, vec3, vec4);
+  ADD4(in8, vec3, in9, vec2, in14, vec5, in15, vec4, in0, vec1, vec6, in2);
+  DOTP_CONST_PAIR(vec6, vec1, cospi_24_64, cospi_8_64, in1, in3);
+  ADD2(in0, in1, in2, in3, vec0, vec7);
+  DOTP_CONST_PAIR(vec7, vec0, cospi_30_64, cospi_2_64, temp1, temp0);
+  ST_SH(temp0, out + 64);
+  ST_SH(temp1, out + 120);
+
+  SUB2(in0, in1, in2, in3, in0, in2);
+  DOTP_CONST_PAIR(in2, in0, cospi_14_64, cospi_18_64, temp1, temp0);
+  ST_SH(temp0, out + 72);
+  ST_SH(temp1, out + 112);
+
+  SUB2(in9, vec2, in14, vec5, vec2, vec5);
+  DOTP_CONST_PAIR((-vec2), vec5, cospi_24_64, cospi_8_64, in2, in1);
+  SUB4(in8, vec3, in15, vec4, in3, in2, in0, in1, in3, in0, vec2, vec5);
+  DOTP_CONST_PAIR(vec5, vec2, cospi_22_64, cospi_10_64, temp1, temp0);
+  ST_SH(temp0, out + 80);
+  ST_SH(temp1, out + 104);
+
+  ADD2(in3, in2, in0, in1, vec3, vec4);
+  DOTP_CONST_PAIR(vec4, vec3, cospi_6_64, cospi_26_64, temp0, temp1);
+  ST_SH(temp0, out + 96);
+  ST_SH(temp1, out + 88);
+}
+
+static void fdct8x32_1d_row_odd_rd(int16_t *temp, int16_t *interm_ptr,
+                                   int16_t *out) {
+  v8i16 in16, in17, in18, in19, in20, in21, in22, in23;
+  v8i16 in24, in25, in26, in27, in28, in29, in30, in31;
+  v8i16 vec4, vec5;
+
+  in20 = LD_SH(temp + 32);
+  in21 = LD_SH(temp + 40);
+  in26 = LD_SH(temp + 80);
+  in27 = LD_SH(temp + 88);
+
+  DOTP_CONST_PAIR(in27, in20, cospi_16_64, cospi_16_64, in20, in27);
+  DOTP_CONST_PAIR(in26, in21, cospi_16_64, cospi_16_64, in21, in26);
+
+  FDCT_POSTPROC_2V_NEG_H(in20, in21);
+  FDCT_POSTPROC_2V_NEG_H(in26, in27);
+
+  in18 = LD_SH(temp + 16);
+  in19 = LD_SH(temp + 24);
+  in28 = LD_SH(temp + 96);
+  in29 = LD_SH(temp + 104);
+
+  FDCT_POSTPROC_2V_NEG_H(in18, in19);
+  FDCT_POSTPROC_2V_NEG_H(in28, in29);
+
+  vec4 = in19 - in20;
+  ST_SH(vec4, interm_ptr + 32);
+  vec4 = in18 - in21;
+  ST_SH(vec4, interm_ptr + 88);
+  vec4 = in29 - in26;
+  ST_SH(vec4, interm_ptr + 64);
+  vec4 = in28 - in27;
+  ST_SH(vec4, interm_ptr + 56);
+
+  ADD4(in18, in21, in19, in20, in28, in27, in29, in26, in21, in20, in27, in26);
+
+  in22 = LD_SH(temp + 48);
+  in23 = LD_SH(temp + 56);
+  in24 = LD_SH(temp + 64);
+  in25 = LD_SH(temp + 72);
+
+  DOTP_CONST_PAIR(in25, in22, cospi_16_64, cospi_16_64, in22, in25);
+  DOTP_CONST_PAIR(in24, in23, cospi_16_64, cospi_16_64, in23, in24);
+  FDCT_POSTPROC_2V_NEG_H(in22, in23);
+  FDCT_POSTPROC_2V_NEG_H(in24, in25);
+
+  in16 = LD_SH(temp);
+  in17 = LD_SH(temp + 8);
+  in30 = LD_SH(temp + 112);
+  in31 = LD_SH(temp + 120);
+
+  FDCT_POSTPROC_2V_NEG_H(in16, in17);
+  FDCT_POSTPROC_2V_NEG_H(in30, in31);
+
+  vec4 = in17 - in22;
+  ST_SH(vec4, interm_ptr + 40);
+  vec4 = in30 - in25;
+  ST_SH(vec4, interm_ptr + 48);
+  vec4 = in31 - in24;
+  ST_SH(vec4, interm_ptr + 72);
+  vec4 = in16 - in23;
+  ST_SH(vec4, interm_ptr + 80);
+
+  ADD4(in16, in23, in17, in22, in30, in25, in31, in24, in16, in17, in30, in31);
+  DOTP_CONST_PAIR(in26, in21, cospi_24_64, cospi_8_64, in18, in29);
+  DOTP_CONST_PAIR(in27, in20, cospi_24_64, cospi_8_64, in19, in28);
+  ADD4(in16, in19, in17, in18, in30, in29, in31, in28, in27, in22, in21, in25);
+  DOTP_CONST_PAIR(in21, in22, cospi_28_64, cospi_4_64, in26, in24);
+  ADD2(in27, in26, in25, in24, in23, in20);
+  DOTP_CONST_PAIR(in20, in23, cospi_31_64, cospi_1_64, vec4, vec5);
+  ST_SH(vec5, out);
+  ST_SH(vec4, out + 120);
+
+  SUB2(in27, in26, in25, in24, in22, in21);
+  DOTP_CONST_PAIR(in21, in22, cospi_15_64, cospi_17_64, vec5, vec4);
+  ST_SH(vec5, out + 112);
+  ST_SH(vec4, out + 8);
+
+  SUB4(in17, in18, in16, in19, in31, in28, in30, in29, in23, in26, in24, in20);
+  DOTP_CONST_PAIR((-in23), in20, cospi_28_64, cospi_4_64, in27, in25);
+  SUB2(in26, in27, in24, in25, in23, in20);
+  DOTP_CONST_PAIR(in20, in23, cospi_23_64, cospi_9_64, vec4, vec5);
+  ST_SH(vec4, out + 16);
+  ST_SH(vec5, out + 104);
+
+  ADD2(in26, in27, in24, in25, in22, in21);
+  DOTP_CONST_PAIR(in21, in22, cospi_7_64, cospi_25_64, vec4, vec5);
+  ST_SH(vec4, out + 24);
+  ST_SH(vec5, out + 96);
+
+  in20 = LD_SH(interm_ptr + 32);
+  in21 = LD_SH(interm_ptr + 88);
+  in27 = LD_SH(interm_ptr + 56);
+  in26 = LD_SH(interm_ptr + 64);
+
+  in16 = in20;
+  in17 = in21;
+  DOTP_CONST_PAIR(-in16, in27, cospi_24_64, cospi_8_64, in20, in27);
+  DOTP_CONST_PAIR(-in17, in26, cospi_24_64, cospi_8_64, in21, in26);
+
+  in22 = LD_SH(interm_ptr + 40);
+  in25 = LD_SH(interm_ptr + 48);
+  in24 = LD_SH(interm_ptr + 72);
+  in23 = LD_SH(interm_ptr + 80);
+
+  SUB4(in23, in20, in22, in21, in25, in26, in24, in27, in28, in17, in18, in31);
+  DOTP_CONST_PAIR(in18, in17, cospi_12_64, cospi_20_64, in29, in30);
+  in16 = in28 + in29;
+  in19 = in31 + in30;
+  DOTP_CONST_PAIR(in19, in16, cospi_27_64, cospi_5_64, vec5, vec4);
+  ST_SH(vec5, out + 32);
+  ST_SH(vec4, out + 88);
+
+  SUB2(in28, in29, in31, in30, in17, in18);
+  DOTP_CONST_PAIR(in18, in17, cospi_11_64, cospi_21_64, vec5, vec4);
+  ST_SH(vec5, out + 40);
+  ST_SH(vec4, out + 80);
+
+  ADD4(in22, in21, in23, in20, in24, in27, in25, in26, in16, in29, in30, in19);
+  DOTP_CONST_PAIR(-in16, in19, cospi_12_64, cospi_20_64, in28, in31);
+  SUB2(in29, in28, in30, in31, in16, in19);
+  DOTP_CONST_PAIR(in19, in16, cospi_19_64, cospi_13_64, vec5, vec4);
+  ST_SH(vec5, out + 72);
+  ST_SH(vec4, out + 48);
+
+  ADD2(in29, in28, in30, in31, in17, in18);
+  DOTP_CONST_PAIR(in18, in17, cospi_3_64, cospi_29_64, vec5, vec4);
+  ST_SH(vec4, out + 56);
+  ST_SH(vec5, out + 64);
+}
+
+static void fdct32x8_1d_row_rd(int16_t *tmp_buf_big, int16_t *tmp_buf,
+                               int16_t *output) {
+  fdct8x32_1d_row_load_butterfly(tmp_buf_big, tmp_buf);
+  fdct8x32_1d_row_even_rd(tmp_buf, tmp_buf);
+  fdct8x32_1d_row_odd_rd((tmp_buf + 128), tmp_buf_big, (tmp_buf + 128));
+  fdct8x32_1d_row_transpose_store(tmp_buf, output);
+}
+
+void aom_fdct32x32_rd_msa(const int16_t *input, int16_t *out,
+                          int32_t src_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, tmp_buf_big[1024]);
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[256]);
+
+  /* column transform */
+  for (i = 0; i < 4; ++i) {
+    fdct8x32_1d_column(input + (8 * i), src_stride, &tmp_buf[0],
+                       &tmp_buf_big[0] + (8 * i));
+  }
+
+  /* row transform */
+  for (i = 0; i < 4; ++i) {
+    fdct32x8_1d_row_rd(&tmp_buf_big[0] + (8 * i * 32), &tmp_buf[0],
+                       out + (8 * i * 32));
+  }
+}
+
+void aom_fdct32x32_1_msa(const int16_t *input, int16_t *out, int32_t stride) {
+  int sum = LD_HADD(input, stride);
+  sum += LD_HADD(input + 8, stride);
+  sum += LD_HADD(input + 16, stride);
+  sum += LD_HADD(input + 24, stride);
+  sum += LD_HADD(input + 32 * 8, stride);
+  sum += LD_HADD(input + 32 * 8 + 8, stride);
+  sum += LD_HADD(input + 32 * 8 + 16, stride);
+  sum += LD_HADD(input + 32 * 8 + 24, stride);
+  sum += LD_HADD(input + 32 * 16, stride);
+  sum += LD_HADD(input + 32 * 16 + 8, stride);
+  sum += LD_HADD(input + 32 * 16 + 16, stride);
+  sum += LD_HADD(input + 32 * 16 + 24, stride);
+  sum += LD_HADD(input + 32 * 24, stride);
+  sum += LD_HADD(input + 32 * 24 + 8, stride);
+  sum += LD_HADD(input + 32 * 24 + 16, stride);
+  sum += LD_HADD(input + 32 * 24 + 24, stride);
+  out[0] = (int16_t)(sum >> 3);
+}
diff --git a/third_party/aom/aom_dsp/mips/fwd_txfm_msa.c b/third_party/aom/aom_dsp/mips/fwd_txfm_msa.c
new file mode 100644
index 0000000000..f16d290c82
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/fwd_txfm_msa.c
@@ -0,0 +1,246 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/fwd_txfm_msa.h"
+
+void fdct8x16_1d_column(const int16_t *input, int16_t *tmp_ptr,
+                        int32_t src_stride) {
+  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+  v8i16 stp21, stp22, stp23, stp24, stp25, stp26, stp30;
+  v8i16 stp31, stp32, stp33, stp34, stp35, stp36, stp37;
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, cnst0, cnst1, cnst4, cnst5;
+  v8i16 coeff = { cospi_16_64, -cospi_16_64, cospi_8_64,  cospi_24_64,
+                  -cospi_8_64, -cospi_24_64, cospi_12_64, cospi_20_64 };
+  v8i16 coeff1 = { cospi_2_64,  cospi_30_64, cospi_14_64, cospi_18_64,
+                   cospi_10_64, cospi_22_64, cospi_6_64,  cospi_26_64 };
+  v8i16 coeff2 = {
+    -cospi_2_64, -cospi_10_64, -cospi_18_64, -cospi_26_64, 0, 0, 0, 0
+  };
+
+  LD_SH16(input, src_stride, in0, in1, in2, in3, in4, in5, in6, in7, in8, in9,
+          in10, in11, in12, in13, in14, in15);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+  SLLI_4V(in8, in9, in10, in11, 2);
+  SLLI_4V(in12, in13, in14, in15, 2);
+  ADD4(in0, in15, in1, in14, in2, in13, in3, in12, tmp0, tmp1, tmp2, tmp3);
+  ADD4(in4, in11, in5, in10, in6, in9, in7, in8, tmp4, tmp5, tmp6, tmp7);
+  FDCT8x16_EVEN(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp0, tmp1,
+                tmp2, tmp3, tmp4, tmp5, tmp6, tmp7);
+  ST_SH8(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp_ptr, 32);
+  SUB4(in0, in15, in1, in14, in2, in13, in3, in12, in15, in14, in13, in12);
+  SUB4(in4, in11, in5, in10, in6, in9, in7, in8, in11, in10, in9, in8);
+
+  tmp_ptr += 16;
+
+  /* stp 1 */
+  ILVL_H2_SH(in10, in13, in11, in12, vec2, vec4);
+  ILVR_H2_SH(in10, in13, in11, in12, vec3, vec5);
+
+  cnst4 = __msa_splati_h(coeff, 0);
+  stp25 = DOT_SHIFT_RIGHT_PCK_H(vec2, vec3, cnst4);
+
+  cnst5 = __msa_splati_h(coeff, 1);
+  cnst5 = __msa_ilvev_h(cnst5, cnst4);
+  stp22 = DOT_SHIFT_RIGHT_PCK_H(vec2, vec3, cnst5);
+  stp24 = DOT_SHIFT_RIGHT_PCK_H(vec4, vec5, cnst4);
+  stp23 = DOT_SHIFT_RIGHT_PCK_H(vec4, vec5, cnst5);
+
+  /* stp2 */
+  BUTTERFLY_4(in8, in9, stp22, stp23, stp30, stp31, stp32, stp33);
+  BUTTERFLY_4(in15, in14, stp25, stp24, stp37, stp36, stp35, stp34);
+  ILVL_H2_SH(stp36, stp31, stp35, stp32, vec2, vec4);
+  ILVR_H2_SH(stp36, stp31, stp35, stp32, vec3, vec5);
+  SPLATI_H2_SH(coeff, 2, 3, cnst0, cnst1);
+  cnst0 = __msa_ilvev_h(cnst0, cnst1);
+  stp26 = DOT_SHIFT_RIGHT_PCK_H(vec2, vec3, cnst0);
+
+  cnst0 = __msa_splati_h(coeff, 4);
+  cnst1 = __msa_ilvev_h(cnst1, cnst0);
+  stp21 = DOT_SHIFT_RIGHT_PCK_H(vec2, vec3, cnst1);
+
+  BUTTERFLY_4(stp30, stp37, stp26, stp21, in8, in15, in14, in9);
+  ILVRL_H2_SH(in15, in8, vec1, vec0);
+  SPLATI_H2_SH(coeff1, 0, 1, cnst0, cnst1);
+  cnst0 = __msa_ilvev_h(cnst0, cnst1);
+
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr);
+
+  cnst0 = __msa_splati_h(coeff2, 0);
+  cnst0 = __msa_ilvev_h(cnst1, cnst0);
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr + 224);
+
+  ILVRL_H2_SH(in14, in9, vec1, vec0);
+  SPLATI_H2_SH(coeff1, 2, 3, cnst0, cnst1);
+  cnst1 = __msa_ilvev_h(cnst1, cnst0);
+
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst1);
+  ST_SH(in8, tmp_ptr + 128);
+
+  cnst1 = __msa_splati_h(coeff2, 2);
+  cnst0 = __msa_ilvev_h(cnst0, cnst1);
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr + 96);
+
+  SPLATI_H2_SH(coeff, 2, 5, cnst0, cnst1);
+  cnst1 = __msa_ilvev_h(cnst1, cnst0);
+
+  stp25 = DOT_SHIFT_RIGHT_PCK_H(vec4, vec5, cnst1);
+
+  cnst1 = __msa_splati_h(coeff, 3);
+  cnst1 = __msa_ilvev_h(cnst0, cnst1);
+  stp22 = DOT_SHIFT_RIGHT_PCK_H(vec4, vec5, cnst1);
+
+  /* stp4 */
+  ADD2(stp34, stp25, stp33, stp22, in13, in10);
+
+  ILVRL_H2_SH(in13, in10, vec1, vec0);
+  SPLATI_H2_SH(coeff1, 4, 5, cnst0, cnst1);
+  cnst0 = __msa_ilvev_h(cnst0, cnst1);
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr + 64);
+
+  cnst0 = __msa_splati_h(coeff2, 1);
+  cnst0 = __msa_ilvev_h(cnst1, cnst0);
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr + 160);
+
+  SUB2(stp34, stp25, stp33, stp22, in12, in11);
+  ILVRL_H2_SH(in12, in11, vec1, vec0);
+  SPLATI_H2_SH(coeff1, 6, 7, cnst0, cnst1);
+  cnst1 = __msa_ilvev_h(cnst1, cnst0);
+
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst1);
+  ST_SH(in8, tmp_ptr + 192);
+
+  cnst1 = __msa_splati_h(coeff2, 3);
+  cnst0 = __msa_ilvev_h(cnst0, cnst1);
+  in8 = DOT_SHIFT_RIGHT_PCK_H(vec0, vec1, cnst0);
+  ST_SH(in8, tmp_ptr + 32);
+}
+
+void fdct16x8_1d_row(int16_t *input, int16_t *output) {
+  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 in8, in9, in10, in11, in12, in13, in14, in15;
+
+  LD_SH8(input, 16, in0, in1, in2, in3, in4, in5, in6, in7);
+  LD_SH8((input + 8), 16, in8, in9, in10, in11, in12, in13, in14, in15);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9,
+                     in10, in11, in12, in13, in14, in15);
+  ADD4(in0, 1, in1, 1, in2, 1, in3, 1, in0, in1, in2, in3);
+  ADD4(in4, 1, in5, 1, in6, 1, in7, 1, in4, in5, in6, in7);
+  ADD4(in8, 1, in9, 1, in10, 1, in11, 1, in8, in9, in10, in11);
+  ADD4(in12, 1, in13, 1, in14, 1, in15, 1, in12, in13, in14, in15);
+  SRA_4V(in0, in1, in2, in3, 2);
+  SRA_4V(in4, in5, in6, in7, 2);
+  SRA_4V(in8, in9, in10, in11, 2);
+  SRA_4V(in12, in13, in14, in15, 2);
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,
+               in12, in13, in14, in15, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6,
+               tmp7, in8, in9, in10, in11, in12, in13, in14, in15);
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, input, 16);
+  FDCT8x16_EVEN(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp0, tmp1,
+                tmp2, tmp3, tmp4, tmp5, tmp6, tmp7);
+  LD_SH8(input, 16, in8, in9, in10, in11, in12, in13, in14, in15);
+  FDCT8x16_ODD(in8, in9, in10, in11, in12, in13, in14, in15, in0, in1, in2, in3,
+               in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3, tmp0, in0,
+                     tmp1, in1, tmp2, in2, tmp3, in3);
+  ST_SH8(tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3, output, 16);
+  TRANSPOSE8x8_SH_SH(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7, tmp4, in4,
+                     tmp5, in5, tmp6, in6, tmp7, in7);
+  ST_SH8(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7, output + 8, 16);
+}
+
+void aom_fdct4x4_msa(const int16_t *input, int16_t *output,
+                     int32_t src_stride) {
+  v8i16 in0, in1, in2, in3;
+
+  LD_SH4(input, src_stride, in0, in1, in2, in3);
+
+  /* fdct4 pre-process */
+  {
+    v8i16 vec, mask;
+    v16i8 zero = { 0 };
+    v16i8 one = __msa_ldi_b(1);
+
+    mask = (v8i16)__msa_sldi_b(zero, one, 15);
+    SLLI_4V(in0, in1, in2, in3, 4);
+    vec = __msa_ceqi_h(in0, 0);
+    vec = vec ^ 255;
+    vec = mask & vec;
+    in0 += vec;
+  }
+
+  AOM_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  AOM_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  ADD4(in0, 1, in1, 1, in2, 1, in3, 1, in0, in1, in2, in3);
+  SRA_4V(in0, in1, in2, in3, 2);
+  PCKEV_D2_SH(in1, in0, in3, in2, in0, in2);
+  ST_SH2(in0, in2, output, 8);
+}
+
+void aom_fdct8x8_msa(const int16_t *input, int16_t *output,
+                     int32_t src_stride) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  LD_SH8(input, src_stride, in0, in1, in2, in3, in4, in5, in6, in7);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+  AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+            in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+            in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7);
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output, 8);
+}
+
+void aom_fdct8x8_1_msa(const int16_t *input, int16_t *out, int32_t stride) {
+  out[0] = LD_HADD(input, stride);
+  out[1] = 0;
+}
+
+void aom_fdct16x16_msa(const int16_t *input, int16_t *output,
+                       int32_t src_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[16 * 16]);
+
+  /* column transform */
+  for (i = 0; i < 2; ++i) {
+    fdct8x16_1d_column((input + 8 * i), (&tmp_buf[0] + 8 * i), src_stride);
+  }
+
+  /* row transform */
+  for (i = 0; i < 2; ++i) {
+    fdct16x8_1d_row((&tmp_buf[0] + (128 * i)), (output + (128 * i)));
+  }
+}
+
+void aom_fdct16x16_1_msa(const int16_t *input, int16_t *out, int32_t stride) {
+  int sum = LD_HADD(input, stride);
+  sum += LD_HADD(input + 8, stride);
+  sum += LD_HADD(input + 16 * 8, stride);
+  sum += LD_HADD(input + 16 * 8 + 8, stride);
+  out[0] = (int16_t)(sum >> 1);
+}
diff --git a/third_party/aom/aom_dsp/mips/fwd_txfm_msa.h b/third_party/aom/aom_dsp/mips/fwd_txfm_msa.h
new file mode 100644
index 0000000000..ada25dffd9
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/fwd_txfm_msa.h
@@ -0,0 +1,381 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_MIPS_FWD_TXFM_MSA_H_
+#define AOM_DSP_MIPS_FWD_TXFM_MSA_H_
+
+#include "aom_dsp/mips/txfm_macros_msa.h"
+#include "aom_dsp/txfm_common.h"
+
+#define LD_HADD(psrc, stride)                                                  \
+  ({                                                                           \
+    v8i16 in0_m, in1_m, in2_m, in3_m, in4_m, in5_m, in6_m, in7_m;              \
+    v4i32 vec_w_m;                                                             \
+                                                                               \
+    LD_SH4((psrc), stride, in0_m, in1_m, in2_m, in3_m);                        \
+    ADD2(in0_m, in1_m, in2_m, in3_m, in0_m, in2_m);                            \
+    LD_SH4(((psrc) + 4 * stride), stride, in4_m, in5_m, in6_m, in7_m);         \
+    ADD4(in4_m, in5_m, in6_m, in7_m, in0_m, in2_m, in4_m, in6_m, in4_m, in6_m, \
+         in0_m, in4_m);                                                        \
+    in0_m += in4_m;                                                            \
+                                                                               \
+    vec_w_m = __msa_hadd_s_w(in0_m, in0_m);                                    \
+    HADD_SW_S32(vec_w_m);                                                      \
+  })
+
+#define AOM_FDCT4(in0, in1, in2, in3, out0, out1, out2, out3)                  \
+  {                                                                            \
+    v8i16 cnst0_m, cnst1_m, cnst2_m, cnst3_m;                                  \
+    v8i16 vec0_m, vec1_m, vec2_m, vec3_m;                                      \
+    v4i32 vec4_m, vec5_m, vec6_m, vec7_m;                                      \
+    v8i16 coeff_m = {                                                          \
+      cospi_16_64, -cospi_16_64, cospi_8_64, cospi_24_64, -cospi_8_64, 0, 0, 0 \
+    };                                                                         \
+                                                                               \
+    BUTTERFLY_4(in0, in1, in2, in3, vec0_m, vec1_m, vec2_m, vec3_m);           \
+    ILVR_H2_SH(vec1_m, vec0_m, vec3_m, vec2_m, vec0_m, vec2_m);                \
+    SPLATI_H2_SH(coeff_m, 0, 1, cnst0_m, cnst1_m);                             \
+    cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                 \
+    vec5_m = __msa_dotp_s_w(vec0_m, cnst1_m);                                  \
+                                                                               \
+    SPLATI_H2_SH(coeff_m, 4, 3, cnst2_m, cnst3_m);                             \
+    cnst2_m = __msa_ilvev_h(cnst3_m, cnst2_m);                                 \
+    vec7_m = __msa_dotp_s_w(vec2_m, cnst2_m);                                  \
+                                                                               \
+    vec4_m = __msa_dotp_s_w(vec0_m, cnst0_m);                                  \
+    cnst2_m = __msa_splati_h(coeff_m, 2);                                      \
+    cnst2_m = __msa_ilvev_h(cnst2_m, cnst3_m);                                 \
+    vec6_m = __msa_dotp_s_w(vec2_m, cnst2_m);                                  \
+                                                                               \
+    SRARI_W4_SW(vec4_m, vec5_m, vec6_m, vec7_m, DCT_CONST_BITS);               \
+    PCKEV_H4_SH(vec4_m, vec4_m, vec5_m, vec5_m, vec6_m, vec6_m, vec7_m,        \
+                vec7_m, out0, out2, out1, out3);                               \
+  }
+
+#define SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7)              \
+  {                                                                          \
+    v8i16 vec0_m, vec1_m, vec2_m, vec3_m, vec4_m, vec5_m, vec6_m, vec7_m;    \
+                                                                             \
+    SRLI_H4_SH(in0, in1, in2, in3, vec0_m, vec1_m, vec2_m, vec3_m, 15);      \
+    SRLI_H4_SH(in4, in5, in6, in7, vec4_m, vec5_m, vec6_m, vec7_m, 15);      \
+    AVE_SH4_SH(vec0_m, in0, vec1_m, in1, vec2_m, in2, vec3_m, in3, in0, in1, \
+               in2, in3);                                                    \
+    AVE_SH4_SH(vec4_m, in4, vec5_m, in5, vec6_m, in6, vec7_m, in7, in4, in5, \
+               in6, in7);                                                    \
+  }
+
+#define AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2,  \
+                  out3, out4, out5, out6, out7)                              \
+  {                                                                          \
+    v8i16 s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m;                          \
+    v8i16 s7_m, x0_m, x1_m, x2_m, x3_m;                                      \
+    v8i16 coeff_m = { cospi_16_64, -cospi_16_64, cospi_8_64,  cospi_24_64,   \
+                      cospi_4_64,  cospi_28_64,  cospi_12_64, cospi_20_64 }; \
+                                                                             \
+    /* FDCT stage1 */                                                        \
+    BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7, s0_m, s1_m, s2_m,    \
+                s3_m, s4_m, s5_m, s6_m, s7_m);                               \
+    BUTTERFLY_4(s0_m, s1_m, s2_m, s3_m, x0_m, x1_m, x2_m, x3_m);             \
+    ILVL_H2_SH(x1_m, x0_m, x3_m, x2_m, s0_m, s2_m);                          \
+    ILVR_H2_SH(x1_m, x0_m, x3_m, x2_m, s1_m, s3_m);                          \
+    SPLATI_H2_SH(coeff_m, 0, 1, x0_m, x1_m);                                 \
+    x1_m = __msa_ilvev_h(x1_m, x0_m);                                        \
+    out4 = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x1_m);                          \
+                                                                             \
+    SPLATI_H2_SH(coeff_m, 2, 3, x2_m, x3_m);                                 \
+    x2_m = -x2_m;                                                            \
+    x2_m = __msa_ilvev_h(x3_m, x2_m);                                        \
+    out6 = DOT_SHIFT_RIGHT_PCK_H(s2_m, s3_m, x2_m);                          \
+                                                                             \
+    out0 = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x0_m);                          \
+    x2_m = __msa_splati_h(coeff_m, 2);                                       \
+    x2_m = __msa_ilvev_h(x2_m, x3_m);                                        \
+    out2 = DOT_SHIFT_RIGHT_PCK_H(s2_m, s3_m, x2_m);                          \
+                                                                             \
+    /* stage2 */                                                             \
+    ILVRL_H2_SH(s5_m, s6_m, s1_m, s0_m);                                     \
+                                                                             \
+    s6_m = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x0_m);                          \
+    s5_m = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x1_m);                          \
+                                                                             \
+    /* stage3 */                                                             \
+    BUTTERFLY_4(s4_m, s7_m, s6_m, s5_m, x0_m, x3_m, x2_m, x1_m);             \
+                                                                             \
+    /* stage4 */                                                             \
+    ILVL_H2_SH(x3_m, x0_m, x2_m, x1_m, s4_m, s6_m);                          \
+    ILVR_H2_SH(x3_m, x0_m, x2_m, x1_m, s5_m, s7_m);                          \
+                                                                             \
+    SPLATI_H2_SH(coeff_m, 4, 5, x0_m, x1_m);                                 \
+    x1_m = __msa_ilvev_h(x0_m, x1_m);                                        \
+    out1 = DOT_SHIFT_RIGHT_PCK_H(s4_m, s5_m, x1_m);                          \
+                                                                             \
+    SPLATI_H2_SH(coeff_m, 6, 7, x2_m, x3_m);                                 \
+    x2_m = __msa_ilvev_h(x3_m, x2_m);                                        \
+    out5 = DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m);                          \
+                                                                             \
+    x1_m = __msa_splati_h(coeff_m, 5);                                       \
+    x0_m = -x0_m;                                                            \
+    x0_m = __msa_ilvev_h(x1_m, x0_m);                                        \
+    out7 = DOT_SHIFT_RIGHT_PCK_H(s4_m, s5_m, x0_m);                          \
+                                                                             \
+    x2_m = __msa_splati_h(coeff_m, 6);                                       \
+    x3_m = -x3_m;                                                            \
+    x2_m = __msa_ilvev_h(x2_m, x3_m);                                        \
+    out3 = DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m);                          \
+  }
+
+#define FDCT8x16_EVEN(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1,    \
+                      out2, out3, out4, out5, out6, out7)                    \
+  {                                                                          \
+    v8i16 s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m, s7_m;                    \
+    v8i16 x0_m, x1_m, x2_m, x3_m;                                            \
+    v8i16 coeff_m = { cospi_16_64, -cospi_16_64, cospi_8_64,  cospi_24_64,   \
+                      cospi_4_64,  cospi_28_64,  cospi_12_64, cospi_20_64 }; \
+                                                                             \
+    /* FDCT stage1 */                                                        \
+    BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7, s0_m, s1_m, s2_m,    \
+                s3_m, s4_m, s5_m, s6_m, s7_m);                               \
+    BUTTERFLY_4(s0_m, s1_m, s2_m, s3_m, x0_m, x1_m, x2_m, x3_m);             \
+    ILVL_H2_SH(x1_m, x0_m, x3_m, x2_m, s0_m, s2_m);                          \
+    ILVR_H2_SH(x1_m, x0_m, x3_m, x2_m, s1_m, s3_m);                          \
+    SPLATI_H2_SH(coeff_m, 0, 1, x0_m, x1_m);                                 \
+    x1_m = __msa_ilvev_h(x1_m, x0_m);                                        \
+    out4 = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x1_m);                          \
+                                                                             \
+    SPLATI_H2_SH(coeff_m, 2, 3, x2_m, x3_m);                                 \
+    x2_m = -x2_m;                                                            \
+    x2_m = __msa_ilvev_h(x3_m, x2_m);                                        \
+    out6 = DOT_SHIFT_RIGHT_PCK_H(s2_m, s3_m, x2_m);                          \
+                                                                             \
+    out0 = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x0_m);                          \
+    x2_m = __msa_splati_h(coeff_m, 2);                                       \
+    x2_m = __msa_ilvev_h(x2_m, x3_m);                                        \
+    out2 = DOT_SHIFT_RIGHT_PCK_H(s2_m, s3_m, x2_m);                          \
+                                                                             \
+    /* stage2 */                                                             \
+    ILVRL_H2_SH(s5_m, s6_m, s1_m, s0_m);                                     \
+                                                                             \
+    s6_m = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x0_m);                          \
+    s5_m = DOT_SHIFT_RIGHT_PCK_H(s0_m, s1_m, x1_m);                          \
+                                                                             \
+    /* stage3 */                                                             \
+    BUTTERFLY_4(s4_m, s7_m, s6_m, s5_m, x0_m, x3_m, x2_m, x1_m);             \
+                                                                             \
+    /* stage4 */                                                             \
+    ILVL_H2_SH(x3_m, x0_m, x2_m, x1_m, s4_m, s6_m);                          \
+    ILVR_H2_SH(x3_m, x0_m, x2_m, x1_m, s5_m, s7_m);                          \
+                                                                             \
+    SPLATI_H2_SH(coeff_m, 4, 5, x0_m, x1_m);                                 \
+    x1_m = __msa_ilvev_h(x0_m, x1_m);                                        \
+    out1 = DOT_SHIFT_RIGHT_PCK_H(s4_m, s5_m, x1_m);                          \
+                                                                             \
+    SPLATI_H2_SH(coeff_m, 6, 7, x2_m, x3_m);                                 \
+    x2_m = __msa_ilvev_h(x3_m, x2_m);                                        \
+    out5 = DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m);                          \
+                                                                             \
+    x1_m = __msa_splati_h(coeff_m, 5);                                       \
+    x0_m = -x0_m;                                                            \
+    x0_m = __msa_ilvev_h(x1_m, x0_m);                                        \
+    out7 = DOT_SHIFT_RIGHT_PCK_H(s4_m, s5_m, x0_m);                          \
+                                                                             \
+    x2_m = __msa_splati_h(coeff_m, 6);                                       \
+    x3_m = -x3_m;                                                            \
+    x2_m = __msa_ilvev_h(x2_m, x3_m);                                        \
+    out3 = DOT_SHIFT_RIGHT_PCK_H(s6_m, s7_m, x2_m);                          \
+  }
+
+#define FDCT8x16_ODD(input0, input1, input2, input3, input4, input5, input6,   \
+                     input7, out1, out3, out5, out7, out9, out11, out13,       \
+                     out15)                                                    \
+  {                                                                            \
+    v8i16 stp21_m, stp22_m, stp23_m, stp24_m, stp25_m, stp26_m;                \
+    v8i16 stp30_m, stp31_m, stp32_m, stp33_m, stp34_m, stp35_m;                \
+    v8i16 stp36_m, stp37_m, vec0_m, vec1_m;                                    \
+    v8i16 vec2_m, vec3_m, vec4_m, vec5_m, vec6_m;                              \
+    v8i16 cnst0_m, cnst1_m, cnst4_m, cnst5_m;                                  \
+    v8i16 coeff_m = { cospi_16_64, -cospi_16_64, cospi_8_64,  cospi_24_64,     \
+                      -cospi_8_64, -cospi_24_64, cospi_12_64, cospi_20_64 };   \
+    v8i16 coeff1_m = { cospi_2_64,  cospi_30_64, cospi_14_64, cospi_18_64,     \
+                       cospi_10_64, cospi_22_64, cospi_6_64,  cospi_26_64 };   \
+    v8i16 coeff2_m = {                                                         \
+      -cospi_2_64, -cospi_10_64, -cospi_18_64, -cospi_26_64, 0, 0, 0, 0        \
+    };                                                                         \
+                                                                               \
+    /* stp 1 */                                                                \
+    ILVL_H2_SH(input2, input5, input3, input4, vec2_m, vec4_m);                \
+    ILVR_H2_SH(input2, input5, input3, input4, vec3_m, vec5_m);                \
+                                                                               \
+    cnst4_m = __msa_splati_h(coeff_m, 0);                                      \
+    stp25_m = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst4_m);                  \
+                                                                               \
+    cnst5_m = __msa_splati_h(coeff_m, 1);                                      \
+    cnst5_m = __msa_ilvev_h(cnst5_m, cnst4_m);                                 \
+    stp22_m = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst5_m);                  \
+    stp24_m = DOT_SHIFT_RIGHT_PCK_H(vec4_m, vec5_m, cnst4_m);                  \
+    stp23_m = DOT_SHIFT_RIGHT_PCK_H(vec4_m, vec5_m, cnst5_m);                  \
+                                                                               \
+    /* stp2 */                                                                 \
+    BUTTERFLY_4(input0, input1, stp22_m, stp23_m, stp30_m, stp31_m, stp32_m,   \
+                stp33_m);                                                      \
+    BUTTERFLY_4(input7, input6, stp25_m, stp24_m, stp37_m, stp36_m, stp35_m,   \
+                stp34_m);                                                      \
+                                                                               \
+    ILVL_H2_SH(stp36_m, stp31_m, stp35_m, stp32_m, vec2_m, vec4_m);            \
+    ILVR_H2_SH(stp36_m, stp31_m, stp35_m, stp32_m, vec3_m, vec5_m);            \
+                                                                               \
+    SPLATI_H2_SH(coeff_m, 2, 3, cnst0_m, cnst1_m);                             \
+    cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m);                                 \
+    stp26_m = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst0_m);                  \
+                                                                               \
+    cnst0_m = __msa_splati_h(coeff_m, 4);                                      \
+    cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                 \
+    stp21_m = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst1_m);                  \
+                                                                               \
+    SPLATI_H2_SH(coeff_m, 5, 2, cnst0_m, cnst1_m);                             \
+    cnst1_m = __msa_ilvev_h(cnst0_m, cnst1_m);                                 \
+    stp25_m = DOT_SHIFT_RIGHT_PCK_H(vec4_m, vec5_m, cnst1_m);                  \
+                                                                               \
+    cnst0_m = __msa_splati_h(coeff_m, 3);                                      \
+    cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                 \
+    stp22_m = DOT_SHIFT_RIGHT_PCK_H(vec4_m, vec5_m, cnst1_m);                  \
+                                                                               \
+    /* stp4 */                                                                 \
+    BUTTERFLY_4(stp30_m, stp37_m, stp26_m, stp21_m, vec6_m, vec2_m, vec4_m,    \
+                vec5_m);                                                       \
+    BUTTERFLY_4(stp33_m, stp34_m, stp25_m, stp22_m, stp21_m, stp23_m, stp24_m, \
+                stp31_m);                                                      \
+                                                                               \
+    ILVRL_H2_SH(vec2_m, vec6_m, vec1_m, vec0_m);                               \
+    SPLATI_H2_SH(coeff1_m, 0, 1, cnst0_m, cnst1_m);                            \
+    cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m);                                 \
+                                                                               \
+    out1 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                     \
+                                                                               \
+    cnst0_m = __msa_splati_h(coeff2_m, 0);                                     \
+    cnst0_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                 \
+    out15 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                    \
+                                                                               \
+    ILVRL_H2_SH(vec4_m, vec5_m, vec1_m, vec0_m);                               \
+    SPLATI_H2_SH(coeff1_m, 2, 3, cnst0_m, cnst1_m);                            \
+    cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                 \
+                                                                               \
+    out9 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst1_m);                     \
+                                                                               \
+    cnst1_m = __msa_splati_h(coeff2_m, 2);                                     \
+    cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m);                                 \
+    out7 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                     \
+                                                                               \
+    ILVRL_H2_SH(stp23_m, stp21_m, vec1_m, vec0_m);                             \
+    SPLATI_H2_SH(coeff1_m, 4, 5, cnst0_m, cnst1_m);                            \
+    cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m);                                 \
+    out5 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                     \
+                                                                               \
+    cnst0_m = __msa_splati_h(coeff2_m, 1);                                     \
+    cnst0_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                 \
+    out11 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                    \
+                                                                               \
+    ILVRL_H2_SH(stp24_m, stp31_m, vec1_m, vec0_m);                             \
+    SPLATI_H2_SH(coeff1_m, 6, 7, cnst0_m, cnst1_m);                            \
+    cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                                 \
+                                                                               \
+    out13 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst1_m);                    \
+                                                                               \
+    cnst1_m = __msa_splati_h(coeff2_m, 3);                                     \
+    cnst0_m = __msa_ilvev_h(cnst0_m, cnst1_m);                                 \
+    out3 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                     \
+  }
+
+#define FDCT_POSTPROC_2V_NEG_H(vec0, vec1) \
+  {                                        \
+    v8i16 tp0_m, tp1_m;                    \
+    v8i16 one_m = __msa_ldi_h(1);          \
+                                           \
+    tp0_m = __msa_clti_s_h(vec0, 0);       \
+    tp1_m = __msa_clti_s_h(vec1, 0);       \
+    vec0 += 1;                             \
+    vec1 += 1;                             \
+    tp0_m = one_m & tp0_m;                 \
+    tp1_m = one_m & tp1_m;                 \
+    vec0 += tp0_m;                         \
+    vec1 += tp1_m;                         \
+    vec0 >>= 2;                            \
+    vec1 >>= 2;                            \
+  }
+
+#define FDCT32_POSTPROC_NEG_W(vec)   \
+  {                                  \
+    v4i32 temp_m;                    \
+    v4i32 one_m = __msa_ldi_w(1);    \
+                                     \
+    temp_m = __msa_clti_s_w(vec, 0); \
+    vec += 1;                        \
+    temp_m = one_m & temp_m;         \
+    vec += temp_m;                   \
+    vec >>= 2;                       \
+  }
+
+#define FDCT32_POSTPROC_2V_POS_H(vec0, vec1)        \
+  {                                                 \
+    v8i16 tp0_m, tp1_m;                             \
+    v8i16 one = __msa_ldi_h(1);                     \
+                                                    \
+    tp0_m = __msa_clei_s_h(vec0, 0);                \
+    tp1_m = __msa_clei_s_h(vec1, 0);                \
+    tp0_m = (v8i16)__msa_xori_b((v16u8)tp0_m, 255); \
+    tp1_m = (v8i16)__msa_xori_b((v16u8)tp1_m, 255); \
+    vec0 += 1;                                      \
+    vec1 += 1;                                      \
+    tp0_m = one & tp0_m;                            \
+    tp1_m = one & tp1_m;                            \
+    vec0 += tp0_m;                                  \
+    vec1 += tp1_m;                                  \
+    vec0 >>= 2;                                     \
+    vec1 >>= 2;                                     \
+  }
+
+#define DOTP_CONST_PAIR_W(reg0_left, reg1_left, reg0_right, reg1_right, \
+                          const0, const1, out0, out1, out2, out3)       \
+  {                                                                     \
+    v4i32 s0_m, s1_m, s2_m, s3_m, s4_m, s5_m, s6_m, s7_m;               \
+    v2i64 tp0_m, tp1_m, tp2_m, tp3_m;                                   \
+    v4i32 k0_m = __msa_fill_w((int32_t)const0);                         \
+                                                                        \
+    s0_m = __msa_fill_w((int32_t)const1);                               \
+    k0_m = __msa_ilvev_w(s0_m, k0_m);                                   \
+                                                                        \
+    ILVRL_W2_SW(-reg1_left, reg0_left, s1_m, s0_m);                     \
+    ILVRL_W2_SW(reg0_left, reg1_left, s3_m, s2_m);                      \
+    ILVRL_W2_SW(-reg1_right, reg0_right, s5_m, s4_m);                   \
+    ILVRL_W2_SW(reg0_right, reg1_right, s7_m, s6_m);                    \
+                                                                        \
+    DOTP_SW2_SD(s0_m, s1_m, k0_m, k0_m, tp0_m, tp1_m);                  \
+    DOTP_SW2_SD(s4_m, s5_m, k0_m, k0_m, tp2_m, tp3_m);                  \
+    tp0_m = __msa_srari_d(tp0_m, DCT_CONST_BITS);                       \
+    tp1_m = __msa_srari_d(tp1_m, DCT_CONST_BITS);                       \
+    tp2_m = __msa_srari_d(tp2_m, DCT_CONST_BITS);                       \
+    tp3_m = __msa_srari_d(tp3_m, DCT_CONST_BITS);                       \
+    out0 = __msa_pckev_w((v4i32)tp0_m, (v4i32)tp1_m);                   \
+    out1 = __msa_pckev_w((v4i32)tp2_m, (v4i32)tp3_m);                   \
+                                                                        \
+    DOTP_SW2_SD(s2_m, s3_m, k0_m, k0_m, tp0_m, tp1_m);                  \
+    DOTP_SW2_SD(s6_m, s7_m, k0_m, k0_m, tp2_m, tp3_m);                  \
+    tp0_m = __msa_srari_d(tp0_m, DCT_CONST_BITS);                       \
+    tp1_m = __msa_srari_d(tp1_m, DCT_CONST_BITS);                       \
+    tp2_m = __msa_srari_d(tp2_m, DCT_CONST_BITS);                       \
+    tp3_m = __msa_srari_d(tp3_m, DCT_CONST_BITS);                       \
+    out2 = __msa_pckev_w((v4i32)tp0_m, (v4i32)tp1_m);                   \
+    out3 = __msa_pckev_w((v4i32)tp2_m, (v4i32)tp3_m);                   \
+  }
+
+void fdct8x16_1d_column(const int16_t *input, int16_t *tmp_ptr,
+                        int32_t src_stride);
+void fdct16x8_1d_row(int16_t *input, int16_t *output);
+#endif  // AOM_DSP_MIPS_FWD_TXFM_MSA_H_
diff --git a/third_party/aom/aom_dsp/mips/idct16x16_msa.c b/third_party/aom/aom_dsp/mips/idct16x16_msa.c
new file mode 100644
index 0000000000..0ea127f524
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/idct16x16_msa.c
@@ -0,0 +1,486 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/inv_txfm_msa.h"
+
+void aom_idct16_1d_rows_msa(const int16_t *input, int16_t *output) {
+  v8i16 loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg2, reg4, reg6, reg8, reg10, reg12, reg14;
+  v8i16 reg3, reg13, reg11, reg5, reg7, reg9, reg1, reg15;
+  v8i16 tmp5, tmp6, tmp7;
+
+  LD_SH8(input, 16, reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+  input += 8;
+  LD_SH8(input, 16, reg8, reg9, reg10, reg11, reg12, reg13, reg14, reg15);
+
+  TRANSPOSE8x8_SH_SH(reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7, reg0, reg1,
+                     reg2, reg3, reg4, reg5, reg6, reg7);
+  TRANSPOSE8x8_SH_SH(reg8, reg9, reg10, reg11, reg12, reg13, reg14, reg15, reg8,
+                     reg9, reg10, reg11, reg12, reg13, reg14, reg15);
+  DOTP_CONST_PAIR(reg2, reg14, cospi_28_64, cospi_4_64, reg2, reg14);
+  DOTP_CONST_PAIR(reg10, reg6, cospi_12_64, cospi_20_64, reg10, reg6);
+  BUTTERFLY_4(reg2, reg14, reg6, reg10, loc0, loc1, reg14, reg2);
+  DOTP_CONST_PAIR(reg14, reg2, cospi_16_64, cospi_16_64, loc2, loc3);
+  DOTP_CONST_PAIR(reg0, reg8, cospi_16_64, cospi_16_64, reg0, reg8);
+  DOTP_CONST_PAIR(reg4, reg12, cospi_24_64, cospi_8_64, reg4, reg12);
+  BUTTERFLY_4(reg8, reg0, reg4, reg12, reg2, reg6, reg10, reg14);
+  SUB4(reg2, loc1, reg14, loc0, reg6, loc3, reg10, loc2, reg0, reg12, reg4,
+       reg8);
+  ADD4(reg2, loc1, reg14, loc0, reg6, loc3, reg10, loc2, reg2, reg14, reg6,
+       reg10);
+
+  /* stage 2 */
+  DOTP_CONST_PAIR(reg1, reg15, cospi_30_64, cospi_2_64, reg1, reg15);
+  DOTP_CONST_PAIR(reg9, reg7, cospi_14_64, cospi_18_64, loc2, loc3);
+
+  reg9 = reg1 - loc2;
+  reg1 = reg1 + loc2;
+  reg7 = reg15 - loc3;
+  reg15 = reg15 + loc3;
+
+  DOTP_CONST_PAIR(reg5, reg11, cospi_22_64, cospi_10_64, reg5, reg11);
+  DOTP_CONST_PAIR(reg13, reg3, cospi_6_64, cospi_26_64, loc0, loc1);
+  BUTTERFLY_4(loc0, loc1, reg11, reg5, reg13, reg3, reg11, reg5);
+
+  loc1 = reg15 + reg3;
+  reg3 = reg15 - reg3;
+  loc2 = reg2 + loc1;
+  reg15 = reg2 - loc1;
+
+  loc1 = reg1 + reg13;
+  reg13 = reg1 - reg13;
+  loc0 = reg0 + loc1;
+  loc1 = reg0 - loc1;
+  tmp6 = loc0;
+  tmp7 = loc1;
+  reg0 = loc2;
+
+  DOTP_CONST_PAIR(reg7, reg9, cospi_24_64, cospi_8_64, reg7, reg9);
+  DOTP_CONST_PAIR((-reg5), (-reg11), cospi_8_64, cospi_24_64, reg5, reg11);
+
+  loc0 = reg9 + reg5;
+  reg5 = reg9 - reg5;
+  reg2 = reg6 + loc0;
+  reg1 = reg6 - loc0;
+
+  loc0 = reg7 + reg11;
+  reg11 = reg7 - reg11;
+  loc1 = reg4 + loc0;
+  loc2 = reg4 - loc0;
+  tmp5 = loc1;
+
+  DOTP_CONST_PAIR(reg5, reg11, cospi_16_64, cospi_16_64, reg5, reg11);
+  BUTTERFLY_4(reg8, reg10, reg11, reg5, loc0, reg4, reg9, loc1);
+
+  reg10 = loc0;
+  reg11 = loc1;
+
+  DOTP_CONST_PAIR(reg3, reg13, cospi_16_64, cospi_16_64, reg3, reg13);
+  BUTTERFLY_4(reg12, reg14, reg13, reg3, reg8, reg6, reg7, reg5);
+
+  reg13 = loc2;
+
+  /* Transpose and store the output */
+  reg12 = tmp5;
+  reg14 = tmp6;
+  reg3 = tmp7;
+
+  /* transpose block */
+  TRANSPOSE8x8_SH_SH(reg0, reg2, reg4, reg6, reg8, reg10, reg12, reg14, reg0,
+                     reg2, reg4, reg6, reg8, reg10, reg12, reg14);
+  ST_SH8(reg0, reg2, reg4, reg6, reg8, reg10, reg12, reg14, output, 16);
+
+  /* transpose block */
+  TRANSPOSE8x8_SH_SH(reg3, reg13, reg11, reg5, reg7, reg9, reg1, reg15, reg3,
+                     reg13, reg11, reg5, reg7, reg9, reg1, reg15);
+  ST_SH8(reg3, reg13, reg11, reg5, reg7, reg9, reg1, reg15, (output + 8), 16);
+}
+
+void aom_idct16_1d_columns_addblk_msa(int16_t *input, uint8_t *dst,
+                                      int32_t dst_stride) {
+  v8i16 loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg2, reg4, reg6, reg8, reg10, reg12, reg14;
+  v8i16 reg3, reg13, reg11, reg5, reg7, reg9, reg1, reg15;
+  v8i16 tmp5, tmp6, tmp7;
+
+  /* load up 8x8 */
+  LD_SH8(input, 16, reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+  input += 8 * 16;
+  /* load bottom 8x8 */
+  LD_SH8(input, 16, reg8, reg9, reg10, reg11, reg12, reg13, reg14, reg15);
+
+  DOTP_CONST_PAIR(reg2, reg14, cospi_28_64, cospi_4_64, reg2, reg14);
+  DOTP_CONST_PAIR(reg10, reg6, cospi_12_64, cospi_20_64, reg10, reg6);
+  BUTTERFLY_4(reg2, reg14, reg6, reg10, loc0, loc1, reg14, reg2);
+  DOTP_CONST_PAIR(reg14, reg2, cospi_16_64, cospi_16_64, loc2, loc3);
+  DOTP_CONST_PAIR(reg0, reg8, cospi_16_64, cospi_16_64, reg0, reg8);
+  DOTP_CONST_PAIR(reg4, reg12, cospi_24_64, cospi_8_64, reg4, reg12);
+  BUTTERFLY_4(reg8, reg0, reg4, reg12, reg2, reg6, reg10, reg14);
+
+  reg0 = reg2 - loc1;
+  reg2 = reg2 + loc1;
+  reg12 = reg14 - loc0;
+  reg14 = reg14 + loc0;
+  reg4 = reg6 - loc3;
+  reg6 = reg6 + loc3;
+  reg8 = reg10 - loc2;
+  reg10 = reg10 + loc2;
+
+  /* stage 2 */
+  DOTP_CONST_PAIR(reg1, reg15, cospi_30_64, cospi_2_64, reg1, reg15);
+  DOTP_CONST_PAIR(reg9, reg7, cospi_14_64, cospi_18_64, loc2, loc3);
+
+  reg9 = reg1 - loc2;
+  reg1 = reg1 + loc2;
+  reg7 = reg15 - loc3;
+  reg15 = reg15 + loc3;
+
+  DOTP_CONST_PAIR(reg5, reg11, cospi_22_64, cospi_10_64, reg5, reg11);
+  DOTP_CONST_PAIR(reg13, reg3, cospi_6_64, cospi_26_64, loc0, loc1);
+  BUTTERFLY_4(loc0, loc1, reg11, reg5, reg13, reg3, reg11, reg5);
+
+  loc1 = reg15 + reg3;
+  reg3 = reg15 - reg3;
+  loc2 = reg2 + loc1;
+  reg15 = reg2 - loc1;
+
+  loc1 = reg1 + reg13;
+  reg13 = reg1 - reg13;
+  loc0 = reg0 + loc1;
+  loc1 = reg0 - loc1;
+  tmp6 = loc0;
+  tmp7 = loc1;
+  reg0 = loc2;
+
+  DOTP_CONST_PAIR(reg7, reg9, cospi_24_64, cospi_8_64, reg7, reg9);
+  DOTP_CONST_PAIR((-reg5), (-reg11), cospi_8_64, cospi_24_64, reg5, reg11);
+
+  loc0 = reg9 + reg5;
+  reg5 = reg9 - reg5;
+  reg2 = reg6 + loc0;
+  reg1 = reg6 - loc0;
+
+  loc0 = reg7 + reg11;
+  reg11 = reg7 - reg11;
+  loc1 = reg4 + loc0;
+  loc2 = reg4 - loc0;
+  tmp5 = loc1;
+
+  DOTP_CONST_PAIR(reg5, reg11, cospi_16_64, cospi_16_64, reg5, reg11);
+  BUTTERFLY_4(reg8, reg10, reg11, reg5, loc0, reg4, reg9, loc1);
+
+  reg10 = loc0;
+  reg11 = loc1;
+
+  DOTP_CONST_PAIR(reg3, reg13, cospi_16_64, cospi_16_64, reg3, reg13);
+  BUTTERFLY_4(reg12, reg14, reg13, reg3, reg8, reg6, reg7, reg5);
+  reg13 = loc2;
+
+  /* Transpose and store the output */
+  reg12 = tmp5;
+  reg14 = tmp6;
+  reg3 = tmp7;
+
+  SRARI_H4_SH(reg0, reg2, reg4, reg6, 6);
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, reg0, reg2, reg4, reg6);
+  dst += (4 * dst_stride);
+  SRARI_H4_SH(reg8, reg10, reg12, reg14, 6);
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, reg8, reg10, reg12, reg14);
+  dst += (4 * dst_stride);
+  SRARI_H4_SH(reg3, reg13, reg11, reg5, 6);
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, reg3, reg13, reg11, reg5);
+  dst += (4 * dst_stride);
+  SRARI_H4_SH(reg7, reg9, reg1, reg15, 6);
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, reg7, reg9, reg1, reg15);
+}
+
+void aom_idct16x16_256_add_msa(const int16_t *input, uint8_t *dst,
+                               int32_t dst_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, out_arr[16 * 16]);
+  int16_t *out = out_arr;
+
+  /* transform rows */
+  for (i = 0; i < 2; ++i) {
+    /* process 16 * 8 block */
+    aom_idct16_1d_rows_msa((input + (i << 7)), (out + (i << 7)));
+  }
+
+  /* transform columns */
+  for (i = 0; i < 2; ++i) {
+    /* process 8 * 16 block */
+    aom_idct16_1d_columns_addblk_msa((out + (i << 3)), (dst + (i << 3)),
+                                     dst_stride);
+  }
+}
+
+void aom_idct16x16_10_add_msa(const int16_t *input, uint8_t *dst,
+                              int32_t dst_stride) {
+  uint8_t i;
+  DECLARE_ALIGNED(32, int16_t, out_arr[16 * 16]);
+  int16_t *out = out_arr;
+
+  /* process 16 * 8 block */
+  aom_idct16_1d_rows_msa(input, out);
+
+  /* short case just considers top 4 rows as valid output */
+  out += 4 * 16;
+  for (i = 12; i--;) {
+    __asm__ __volatile__(
+        "sw     $zero,   0(%[out])     \n\t"
+        "sw     $zero,   4(%[out])     \n\t"
+        "sw     $zero,   8(%[out])     \n\t"
+        "sw     $zero,  12(%[out])     \n\t"
+        "sw     $zero,  16(%[out])     \n\t"
+        "sw     $zero,  20(%[out])     \n\t"
+        "sw     $zero,  24(%[out])     \n\t"
+        "sw     $zero,  28(%[out])     \n\t"
+
+        :
+        : [out] "r"(out));
+
+    out += 16;
+  }
+
+  out = out_arr;
+
+  /* transform columns */
+  for (i = 0; i < 2; ++i) {
+    /* process 8 * 16 block */
+    aom_idct16_1d_columns_addblk_msa((out + (i << 3)), (dst + (i << 3)),
+                                     dst_stride);
+  }
+}
+
+void aom_idct16x16_1_add_msa(const int16_t *input, uint8_t *dst,
+                             int32_t dst_stride) {
+  uint8_t i;
+  int16_t out;
+  v8i16 vec, res0, res1, res2, res3, res4, res5, res6, res7;
+  v16u8 dst0, dst1, dst2, dst3, tmp0, tmp1, tmp2, tmp3;
+
+  out = ROUND_POWER_OF_TWO((input[0] * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO((out * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO(out, 6);
+
+  vec = __msa_fill_h(out);
+
+  for (i = 4; i--;) {
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    UNPCK_UB_SH(dst0, res0, res4);
+    UNPCK_UB_SH(dst1, res1, res5);
+    UNPCK_UB_SH(dst2, res2, res6);
+    UNPCK_UB_SH(dst3, res3, res7);
+    ADD4(res0, vec, res1, vec, res2, vec, res3, vec, res0, res1, res2, res3);
+    ADD4(res4, vec, res5, vec, res6, vec, res7, vec, res4, res5, res6, res7);
+    CLIP_SH4_0_255(res0, res1, res2, res3);
+    CLIP_SH4_0_255(res4, res5, res6, res7);
+    PCKEV_B4_UB(res4, res0, res5, res1, res6, res2, res7, res3, tmp0, tmp1,
+                tmp2, tmp3);
+    ST_UB4(tmp0, tmp1, tmp2, tmp3, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+void aom_iadst16_1d_rows_msa(const int16_t *input, int16_t *output) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 l0, l1, l2, l3, l4, l5, l6, l7, l8, l9, l10, l11, l12, l13, l14, l15;
+
+  /* load input data */
+  LD_SH16(input, 8, l0, l8, l1, l9, l2, l10, l3, l11, l4, l12, l5, l13, l6, l14,
+          l7, l15);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7, l0, l1, l2, l3, l4, l5, l6,
+                     l7);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15, l8, l9, l10, l11,
+                     l12, l13, l14, l15);
+
+  /* ADST in horizontal */
+  AOM_IADST8x16_1D(l0, l1, l2, l3, l4, l5, l6, l7, l8, l9, l10, l11, l12, l13,
+                   l14, l15, r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11,
+                   r12, r13, r14, r15);
+
+  l1 = -r8;
+  l3 = -r4;
+  l13 = -r13;
+  l15 = -r1;
+
+  TRANSPOSE8x8_SH_SH(r0, l1, r12, l3, r6, r14, r10, r2, l0, l1, l2, l3, l4, l5,
+                     l6, l7);
+  ST_SH8(l0, l1, l2, l3, l4, l5, l6, l7, output, 16);
+  TRANSPOSE8x8_SH_SH(r3, r11, r15, r7, r5, l13, r9, l15, l8, l9, l10, l11, l12,
+                     l13, l14, l15);
+  ST_SH8(l8, l9, l10, l11, l12, l13, l14, l15, (output + 8), 16);
+}
+
+void aom_iadst16_1d_columns_addblk_msa(int16_t *input, uint8_t *dst,
+                                       int32_t dst_stride) {
+  v8i16 v0, v2, v4, v6, k0, k1, k2, k3;
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 out0, out1, out2, out3, out4, out5, out6, out7;
+  v8i16 out8, out9, out10, out11, out12, out13, out14, out15;
+  v8i16 g0, g1, g2, g3, g4, g5, g6, g7, g8, g9, g10, g11, g12, g13, g14, g15;
+  v8i16 h0, h1, h2, h3, h4, h5, h6, h7, h8, h9, h10, h11;
+  v8i16 res0, res1, res2, res3, res4, res5, res6, res7;
+  v8i16 res8, res9, res10, res11, res12, res13, res14, res15;
+  v16u8 dst0, dst1, dst2, dst3, dst4, dst5, dst6, dst7;
+  v16u8 dst8, dst9, dst10, dst11, dst12, dst13, dst14, dst15;
+  v16i8 zero = { 0 };
+
+  r0 = LD_SH(input + 0 * 16);
+  r3 = LD_SH(input + 3 * 16);
+  r4 = LD_SH(input + 4 * 16);
+  r7 = LD_SH(input + 7 * 16);
+  r8 = LD_SH(input + 8 * 16);
+  r11 = LD_SH(input + 11 * 16);
+  r12 = LD_SH(input + 12 * 16);
+  r15 = LD_SH(input + 15 * 16);
+
+  /* stage 1 */
+  k0 = AOM_SET_COSPI_PAIR(cospi_1_64, cospi_31_64);
+  k1 = AOM_SET_COSPI_PAIR(cospi_31_64, -cospi_1_64);
+  k2 = AOM_SET_COSPI_PAIR(cospi_17_64, cospi_15_64);
+  k3 = AOM_SET_COSPI_PAIR(cospi_15_64, -cospi_17_64);
+  MADD_BF(r15, r0, r7, r8, k0, k1, k2, k3, g0, g1, g2, g3);
+  k0 = AOM_SET_COSPI_PAIR(cospi_9_64, cospi_23_64);
+  k1 = AOM_SET_COSPI_PAIR(cospi_23_64, -cospi_9_64);
+  k2 = AOM_SET_COSPI_PAIR(cospi_25_64, cospi_7_64);
+  k3 = AOM_SET_COSPI_PAIR(cospi_7_64, -cospi_25_64);
+  MADD_BF(r11, r4, r3, r12, k0, k1, k2, k3, g8, g9, g10, g11);
+  BUTTERFLY_4(g0, g2, g10, g8, h8, h9, v2, v0);
+  k0 = AOM_SET_COSPI_PAIR(cospi_4_64, cospi_28_64);
+  k1 = AOM_SET_COSPI_PAIR(cospi_28_64, -cospi_4_64);
+  k2 = AOM_SET_COSPI_PAIR(-cospi_28_64, cospi_4_64);
+  MADD_BF(g1, g3, g9, g11, k0, k1, k2, k0, h0, h1, h2, h3);
+
+  r1 = LD_SH(input + 1 * 16);
+  r2 = LD_SH(input + 2 * 16);
+  r5 = LD_SH(input + 5 * 16);
+  r6 = LD_SH(input + 6 * 16);
+  r9 = LD_SH(input + 9 * 16);
+  r10 = LD_SH(input + 10 * 16);
+  r13 = LD_SH(input + 13 * 16);
+  r14 = LD_SH(input + 14 * 16);
+
+  k0 = AOM_SET_COSPI_PAIR(cospi_5_64, cospi_27_64);
+  k1 = AOM_SET_COSPI_PAIR(cospi_27_64, -cospi_5_64);
+  k2 = AOM_SET_COSPI_PAIR(cospi_21_64, cospi_11_64);
+  k3 = AOM_SET_COSPI_PAIR(cospi_11_64, -cospi_21_64);
+  MADD_BF(r13, r2, r5, r10, k0, k1, k2, k3, g4, g5, g6, g7);
+  k0 = AOM_SET_COSPI_PAIR(cospi_13_64, cospi_19_64);
+  k1 = AOM_SET_COSPI_PAIR(cospi_19_64, -cospi_13_64);
+  k2 = AOM_SET_COSPI_PAIR(cospi_29_64, cospi_3_64);
+  k3 = AOM_SET_COSPI_PAIR(cospi_3_64, -cospi_29_64);
+  MADD_BF(r9, r6, r1, r14, k0, k1, k2, k3, g12, g13, g14, g15);
+  BUTTERFLY_4(g4, g6, g14, g12, h10, h11, v6, v4);
+  BUTTERFLY_4(h8, h9, h11, h10, out0, out1, h11, h10);
+  out1 = -out1;
+  SRARI_H2_SH(out0, out1, 6);
+  dst0 = LD_UB(dst + 0 * dst_stride);
+  dst1 = LD_UB(dst + 15 * dst_stride);
+  ILVR_B2_SH(zero, dst0, zero, dst1, res0, res1);
+  ADD2(res0, out0, res1, out1, res0, res1);
+  CLIP_SH2_0_255(res0, res1);
+  PCKEV_B2_SH(res0, res0, res1, res1, res0, res1);
+  ST8x1_UB(res0, dst);
+  ST8x1_UB(res1, dst + 15 * dst_stride);
+
+  k0 = AOM_SET_COSPI_PAIR(cospi_12_64, cospi_20_64);
+  k1 = AOM_SET_COSPI_PAIR(-cospi_20_64, cospi_12_64);
+  k2 = AOM_SET_COSPI_PAIR(cospi_20_64, -cospi_12_64);
+  MADD_BF(g7, g5, g15, g13, k0, k1, k2, k0, h4, h5, h6, h7);
+  BUTTERFLY_4(h0, h2, h6, h4, out8, out9, out11, out10);
+  out8 = -out8;
+
+  SRARI_H2_SH(out8, out9, 6);
+  dst8 = LD_UB(dst + 1 * dst_stride);
+  dst9 = LD_UB(dst + 14 * dst_stride);
+  ILVR_B2_SH(zero, dst8, zero, dst9, res8, res9);
+  ADD2(res8, out8, res9, out9, res8, res9);
+  CLIP_SH2_0_255(res8, res9);
+  PCKEV_B2_SH(res8, res8, res9, res9, res8, res9);
+  ST8x1_UB(res8, dst + dst_stride);
+  ST8x1_UB(res9, dst + 14 * dst_stride);
+
+  k0 = AOM_SET_COSPI_PAIR(cospi_8_64, cospi_24_64);
+  k1 = AOM_SET_COSPI_PAIR(cospi_24_64, -cospi_8_64);
+  k2 = AOM_SET_COSPI_PAIR(-cospi_24_64, cospi_8_64);
+  MADD_BF(v0, v2, v4, v6, k0, k1, k2, k0, out4, out6, out5, out7);
+  out4 = -out4;
+  SRARI_H2_SH(out4, out5, 6);
+  dst4 = LD_UB(dst + 3 * dst_stride);
+  dst5 = LD_UB(dst + 12 * dst_stride);
+  ILVR_B2_SH(zero, dst4, zero, dst5, res4, res5);
+  ADD2(res4, out4, res5, out5, res4, res5);
+  CLIP_SH2_0_255(res4, res5);
+  PCKEV_B2_SH(res4, res4, res5, res5, res4, res5);
+  ST8x1_UB(res4, dst + 3 * dst_stride);
+  ST8x1_UB(res5, dst + 12 * dst_stride);
+
+  MADD_BF(h1, h3, h5, h7, k0, k1, k2, k0, out12, out14, out13, out15);
+  out13 = -out13;
+  SRARI_H2_SH(out12, out13, 6);
+  dst12 = LD_UB(dst + 2 * dst_stride);
+  dst13 = LD_UB(dst + 13 * dst_stride);
+  ILVR_B2_SH(zero, dst12, zero, dst13, res12, res13);
+  ADD2(res12, out12, res13, out13, res12, res13);
+  CLIP_SH2_0_255(res12, res13);
+  PCKEV_B2_SH(res12, res12, res13, res13, res12, res13);
+  ST8x1_UB(res12, dst + 2 * dst_stride);
+  ST8x1_UB(res13, dst + 13 * dst_stride);
+
+  k0 = AOM_SET_COSPI_PAIR(cospi_16_64, cospi_16_64);
+  k3 = AOM_SET_COSPI_PAIR(-cospi_16_64, cospi_16_64);
+  MADD_SHORT(out6, out7, k0, k3, out6, out7);
+  SRARI_H2_SH(out6, out7, 6);
+  dst6 = LD_UB(dst + 4 * dst_stride);
+  dst7 = LD_UB(dst + 11 * dst_stride);
+  ILVR_B2_SH(zero, dst6, zero, dst7, res6, res7);
+  ADD2(res6, out6, res7, out7, res6, res7);
+  CLIP_SH2_0_255(res6, res7);
+  PCKEV_B2_SH(res6, res6, res7, res7, res6, res7);
+  ST8x1_UB(res6, dst + 4 * dst_stride);
+  ST8x1_UB(res7, dst + 11 * dst_stride);
+
+  MADD_SHORT(out10, out11, k0, k3, out10, out11);
+  SRARI_H2_SH(out10, out11, 6);
+  dst10 = LD_UB(dst + 6 * dst_stride);
+  dst11 = LD_UB(dst + 9 * dst_stride);
+  ILVR_B2_SH(zero, dst10, zero, dst11, res10, res11);
+  ADD2(res10, out10, res11, out11, res10, res11);
+  CLIP_SH2_0_255(res10, res11);
+  PCKEV_B2_SH(res10, res10, res11, res11, res10, res11);
+  ST8x1_UB(res10, dst + 6 * dst_stride);
+  ST8x1_UB(res11, dst + 9 * dst_stride);
+
+  k1 = AOM_SET_COSPI_PAIR(-cospi_16_64, -cospi_16_64);
+  k2 = AOM_SET_COSPI_PAIR(cospi_16_64, -cospi_16_64);
+  MADD_SHORT(h10, h11, k1, k2, out2, out3);
+  SRARI_H2_SH(out2, out3, 6);
+  dst2 = LD_UB(dst + 7 * dst_stride);
+  dst3 = LD_UB(dst + 8 * dst_stride);
+  ILVR_B2_SH(zero, dst2, zero, dst3, res2, res3);
+  ADD2(res2, out2, res3, out3, res2, res3);
+  CLIP_SH2_0_255(res2, res3);
+  PCKEV_B2_SH(res2, res2, res3, res3, res2, res3);
+  ST8x1_UB(res2, dst + 7 * dst_stride);
+  ST8x1_UB(res3, dst + 8 * dst_stride);
+
+  MADD_SHORT(out14, out15, k1, k2, out14, out15);
+  SRARI_H2_SH(out14, out15, 6);
+  dst14 = LD_UB(dst + 5 * dst_stride);
+  dst15 = LD_UB(dst + 10 * dst_stride);
+  ILVR_B2_SH(zero, dst14, zero, dst15, res14, res15);
+  ADD2(res14, out14, res15, out15, res14, res15);
+  CLIP_SH2_0_255(res14, res15);
+  PCKEV_B2_SH(res14, res14, res15, res15, res14, res15);
+  ST8x1_UB(res14, dst + 5 * dst_stride);
+  ST8x1_UB(res15, dst + 10 * dst_stride);
+}
diff --git a/third_party/aom/aom_dsp/mips/idct32x32_msa.c b/third_party/aom/aom_dsp/mips/idct32x32_msa.c
new file mode 100644
index 0000000000..f1ca757a0d
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/idct32x32_msa.c
@@ -0,0 +1,730 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/inv_txfm_msa.h"
+
+static void idct32x8_row_transpose_store(const int16_t *input,
+                                         int16_t *tmp_buf) {
+  v8i16 m0, m1, m2, m3, m4, m5, m6, m7, n0, n1, n2, n3, n4, n5, n6, n7;
+
+  /* 1st & 2nd 8x8 */
+  LD_SH8(input, 32, m0, n0, m1, n1, m2, n2, m3, n3);
+  LD_SH8((input + 8), 32, m4, n4, m5, n5, m6, n6, m7, n7);
+  TRANSPOSE8x8_SH_SH(m0, n0, m1, n1, m2, n2, m3, n3, m0, n0, m1, n1, m2, n2, m3,
+                     n3);
+  TRANSPOSE8x8_SH_SH(m4, n4, m5, n5, m6, n6, m7, n7, m4, n4, m5, n5, m6, n6, m7,
+                     n7);
+  ST_SH8(m0, n0, m1, n1, m2, n2, m3, n3, (tmp_buf), 8);
+  ST_SH4(m4, n4, m5, n5, (tmp_buf + 8 * 8), 8);
+  ST_SH4(m6, n6, m7, n7, (tmp_buf + 12 * 8), 8);
+
+  /* 3rd & 4th 8x8 */
+  LD_SH8((input + 16), 32, m0, n0, m1, n1, m2, n2, m3, n3);
+  LD_SH8((input + 24), 32, m4, n4, m5, n5, m6, n6, m7, n7);
+  TRANSPOSE8x8_SH_SH(m0, n0, m1, n1, m2, n2, m3, n3, m0, n0, m1, n1, m2, n2, m3,
+                     n3);
+  TRANSPOSE8x8_SH_SH(m4, n4, m5, n5, m6, n6, m7, n7, m4, n4, m5, n5, m6, n6, m7,
+                     n7);
+  ST_SH4(m0, n0, m1, n1, (tmp_buf + 16 * 8), 8);
+  ST_SH4(m2, n2, m3, n3, (tmp_buf + 20 * 8), 8);
+  ST_SH4(m4, n4, m5, n5, (tmp_buf + 24 * 8), 8);
+  ST_SH4(m6, n6, m7, n7, (tmp_buf + 28 * 8), 8);
+}
+
+static void idct32x8_row_even_process_store(int16_t *tmp_buf,
+                                            int16_t *tmp_eve_buf) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v8i16 stp0, stp1, stp2, stp3, stp4, stp5, stp6, stp7;
+
+  /* Even stage 1 */
+  LD_SH8(tmp_buf, 32, reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+
+  DOTP_CONST_PAIR(reg1, reg7, cospi_28_64, cospi_4_64, reg1, reg7);
+  DOTP_CONST_PAIR(reg5, reg3, cospi_12_64, cospi_20_64, reg5, reg3);
+  BUTTERFLY_4(reg1, reg7, reg3, reg5, vec1, vec3, vec2, vec0);
+  DOTP_CONST_PAIR(vec2, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+
+  loc1 = vec3;
+  loc0 = vec1;
+
+  DOTP_CONST_PAIR(reg0, reg4, cospi_16_64, cospi_16_64, reg0, reg4);
+  DOTP_CONST_PAIR(reg2, reg6, cospi_24_64, cospi_8_64, reg2, reg6);
+  BUTTERFLY_4(reg4, reg0, reg2, reg6, vec1, vec3, vec2, vec0);
+  BUTTERFLY_4(vec0, vec1, loc1, loc0, stp3, stp0, stp7, stp4);
+  BUTTERFLY_4(vec2, vec3, loc3, loc2, stp2, stp1, stp6, stp5);
+
+  /* Even stage 2 */
+  LD_SH8((tmp_buf + 16), 32, reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+  DOTP_CONST_PAIR(reg0, reg7, cospi_30_64, cospi_2_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_14_64, cospi_18_64, reg4, reg3);
+  DOTP_CONST_PAIR(reg2, reg5, cospi_22_64, cospi_10_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg6, reg1, cospi_6_64, cospi_26_64, reg6, reg1);
+
+  vec0 = reg0 + reg4;
+  reg0 = reg0 - reg4;
+  reg4 = reg6 + reg2;
+  reg6 = reg6 - reg2;
+  reg2 = reg1 + reg5;
+  reg1 = reg1 - reg5;
+  reg5 = reg7 + reg3;
+  reg7 = reg7 - reg3;
+  reg3 = vec0;
+
+  vec1 = reg2;
+  reg2 = reg3 + reg4;
+  reg3 = reg3 - reg4;
+  reg4 = reg5 - vec1;
+  reg5 = reg5 + vec1;
+
+  DOTP_CONST_PAIR(reg7, reg0, cospi_24_64, cospi_8_64, reg0, reg7);
+  DOTP_CONST_PAIR((-reg6), reg1, cospi_24_64, cospi_8_64, reg6, reg1);
+
+  vec0 = reg0 - reg6;
+  reg0 = reg0 + reg6;
+  vec1 = reg7 - reg1;
+  reg7 = reg7 + reg1;
+
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, reg6, reg1);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_16_64, cospi_16_64, reg3, reg4);
+
+  /* Even stage 3 : Dependency on Even stage 1 & Even stage 2 */
+  BUTTERFLY_4(stp0, stp1, reg7, reg5, loc1, loc3, loc2, loc0);
+  ST_SH(loc0, (tmp_eve_buf + 15 * 8));
+  ST_SH(loc1, (tmp_eve_buf));
+  ST_SH(loc2, (tmp_eve_buf + 14 * 8));
+  ST_SH(loc3, (tmp_eve_buf + 8));
+
+  BUTTERFLY_4(stp2, stp3, reg4, reg1, loc1, loc3, loc2, loc0);
+  ST_SH(loc0, (tmp_eve_buf + 13 * 8));
+  ST_SH(loc1, (tmp_eve_buf + 2 * 8));
+  ST_SH(loc2, (tmp_eve_buf + 12 * 8));
+  ST_SH(loc3, (tmp_eve_buf + 3 * 8));
+
+  /* Store 8 */
+  BUTTERFLY_4(stp4, stp5, reg6, reg3, loc1, loc3, loc2, loc0);
+  ST_SH(loc0, (tmp_eve_buf + 11 * 8));
+  ST_SH(loc1, (tmp_eve_buf + 4 * 8));
+  ST_SH(loc2, (tmp_eve_buf + 10 * 8));
+  ST_SH(loc3, (tmp_eve_buf + 5 * 8));
+
+  BUTTERFLY_4(stp6, stp7, reg2, reg0, loc1, loc3, loc2, loc0);
+  ST_SH(loc0, (tmp_eve_buf + 9 * 8));
+  ST_SH(loc1, (tmp_eve_buf + 6 * 8));
+  ST_SH(loc2, (tmp_eve_buf + 8 * 8));
+  ST_SH(loc3, (tmp_eve_buf + 7 * 8));
+}
+
+static void idct32x8_row_odd_process_store(int16_t *tmp_buf,
+                                           int16_t *tmp_odd_buf) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+
+  /* Odd stage 1 */
+  reg0 = LD_SH(tmp_buf + 8);
+  reg1 = LD_SH(tmp_buf + 7 * 8);
+  reg2 = LD_SH(tmp_buf + 9 * 8);
+  reg3 = LD_SH(tmp_buf + 15 * 8);
+  reg4 = LD_SH(tmp_buf + 17 * 8);
+  reg5 = LD_SH(tmp_buf + 23 * 8);
+  reg6 = LD_SH(tmp_buf + 25 * 8);
+  reg7 = LD_SH(tmp_buf + 31 * 8);
+
+  DOTP_CONST_PAIR(reg0, reg7, cospi_31_64, cospi_1_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_15_64, cospi_17_64, reg3, reg4);
+  DOTP_CONST_PAIR(reg2, reg5, cospi_23_64, cospi_9_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg6, reg1, cospi_7_64, cospi_25_64, reg1, reg6);
+
+  vec0 = reg0 + reg3;
+  reg0 = reg0 - reg3;
+  reg3 = reg7 + reg4;
+  reg7 = reg7 - reg4;
+  reg4 = reg1 + reg2;
+  reg1 = reg1 - reg2;
+  reg2 = reg6 + reg5;
+  reg6 = reg6 - reg5;
+  reg5 = vec0;
+
+  /* 4 Stores */
+  ADD2(reg5, reg4, reg3, reg2, vec0, vec1);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 4 * 8), 8);
+
+  SUB2(reg5, reg4, reg3, reg2, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_24_64, cospi_8_64, vec0, vec1);
+  ST_SH2(vec0, vec1, (tmp_odd_buf), 8);
+
+  /* 4 Stores */
+  DOTP_CONST_PAIR(reg7, reg0, cospi_28_64, cospi_4_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg6, reg1, -cospi_4_64, cospi_28_64, reg1, reg6);
+  BUTTERFLY_4(reg0, reg7, reg6, reg1, vec0, vec1, vec2, vec3);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 6 * 8), 8);
+
+  DOTP_CONST_PAIR(vec2, vec3, cospi_24_64, cospi_8_64, vec2, vec3);
+  ST_SH2(vec2, vec3, (tmp_odd_buf + 2 * 8), 8);
+
+  /* Odd stage 2 */
+  /* 8 loads */
+  reg0 = LD_SH(tmp_buf + 3 * 8);
+  reg1 = LD_SH(tmp_buf + 5 * 8);
+  reg2 = LD_SH(tmp_buf + 11 * 8);
+  reg3 = LD_SH(tmp_buf + 13 * 8);
+  reg4 = LD_SH(tmp_buf + 19 * 8);
+  reg5 = LD_SH(tmp_buf + 21 * 8);
+  reg6 = LD_SH(tmp_buf + 27 * 8);
+  reg7 = LD_SH(tmp_buf + 29 * 8);
+
+  DOTP_CONST_PAIR(reg1, reg6, cospi_27_64, cospi_5_64, reg1, reg6);
+  DOTP_CONST_PAIR(reg5, reg2, cospi_11_64, cospi_21_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg3, reg4, cospi_19_64, cospi_13_64, reg3, reg4);
+  DOTP_CONST_PAIR(reg7, reg0, cospi_3_64, cospi_29_64, reg0, reg7);
+
+  /* 4 Stores */
+  SUB4(reg1, reg2, reg6, reg5, reg0, reg3, reg7, reg4, vec0, vec1, vec2, vec3);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_12_64, cospi_20_64, loc0, loc1);
+  DOTP_CONST_PAIR(vec3, vec2, -cospi_20_64, cospi_12_64, loc2, loc3);
+
+  BUTTERFLY_4(loc3, loc2, loc0, loc1, vec1, vec0, vec2, vec3);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 12 * 8), 3 * 8);
+
+  DOTP_CONST_PAIR(vec3, vec2, -cospi_8_64, cospi_24_64, vec0, vec1);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 10 * 8), 8);
+
+  /* 4 Stores */
+  ADD4(reg1, reg2, reg6, reg5, reg0, reg3, reg7, reg4, vec1, vec2, vec0, vec3);
+  BUTTERFLY_4(vec0, vec3, vec2, vec1, reg0, reg1, reg3, reg2);
+  ST_SH(reg0, (tmp_odd_buf + 13 * 8));
+  ST_SH(reg1, (tmp_odd_buf + 14 * 8));
+
+  DOTP_CONST_PAIR(reg3, reg2, -cospi_8_64, cospi_24_64, reg0, reg1);
+  ST_SH2(reg0, reg1, (tmp_odd_buf + 8 * 8), 8);
+
+  /* Odd stage 3 : Dependency on Odd stage 1 & Odd stage 2 */
+
+  /* Load 8 & Store 8 */
+  LD_SH4(tmp_odd_buf, 8, reg0, reg1, reg2, reg3);
+  LD_SH4((tmp_odd_buf + 8 * 8), 8, reg4, reg5, reg6, reg7);
+
+  ADD4(reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7, loc0, loc1, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, tmp_odd_buf, 8);
+
+  SUB2(reg0, reg4, reg1, reg5, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc0, loc1);
+
+  SUB2(reg2, reg6, reg3, reg7, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 8 * 8), 8);
+
+  /* Load 8 & Store 8 */
+  LD_SH4((tmp_odd_buf + 4 * 8), 8, reg1, reg2, reg0, reg3);
+  LD_SH4((tmp_odd_buf + 12 * 8), 8, reg4, reg5, reg6, reg7);
+
+  ADD4(reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7, loc0, loc1, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 4 * 8), 8);
+
+  SUB2(reg0, reg4, reg3, reg7, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc0, loc1);
+
+  SUB2(reg1, reg5, reg2, reg6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 12 * 8), 8);
+}
+
+static void idct_butterfly_transpose_store(int16_t *tmp_buf,
+                                           int16_t *tmp_eve_buf,
+                                           int16_t *tmp_odd_buf, int16_t *dst) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 m0, m1, m2, m3, m4, m5, m6, m7, n0, n1, n2, n3, n4, n5, n6, n7;
+
+  /* FINAL BUTTERFLY : Dependency on Even & Odd */
+  vec0 = LD_SH(tmp_odd_buf);
+  vec1 = LD_SH(tmp_odd_buf + 9 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 14 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 6 * 8);
+  loc0 = LD_SH(tmp_eve_buf);
+  loc1 = LD_SH(tmp_eve_buf + 8 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 4 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 12 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m0, m4, m2, m6);
+
+  ST_SH((loc0 - vec3), (tmp_buf + 31 * 8));
+  ST_SH((loc1 - vec2), (tmp_buf + 23 * 8));
+  ST_SH((loc2 - vec1), (tmp_buf + 27 * 8));
+  ST_SH((loc3 - vec0), (tmp_buf + 19 * 8));
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 4 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 13 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 10 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 3 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 2 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 10 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 6 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 14 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m1, m5, m3, m7);
+
+  ST_SH((loc0 - vec3), (tmp_buf + 29 * 8));
+  ST_SH((loc1 - vec2), (tmp_buf + 21 * 8));
+  ST_SH((loc2 - vec1), (tmp_buf + 25 * 8));
+  ST_SH((loc3 - vec0), (tmp_buf + 17 * 8));
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 2 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 11 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 12 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 7 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 1 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 9 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 5 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 13 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n0, n4, n2, n6);
+
+  ST_SH((loc0 - vec3), (tmp_buf + 30 * 8));
+  ST_SH((loc1 - vec2), (tmp_buf + 22 * 8));
+  ST_SH((loc2 - vec1), (tmp_buf + 26 * 8));
+  ST_SH((loc3 - vec0), (tmp_buf + 18 * 8));
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 5 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 15 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 8 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 1 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 3 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 11 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 7 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 15 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n1, n5, n3, n7);
+
+  ST_SH((loc0 - vec3), (tmp_buf + 28 * 8));
+  ST_SH((loc1 - vec2), (tmp_buf + 20 * 8));
+  ST_SH((loc2 - vec1), (tmp_buf + 24 * 8));
+  ST_SH((loc3 - vec0), (tmp_buf + 16 * 8));
+
+  /* Transpose : 16 vectors */
+  /* 1st & 2nd 8x8 */
+  TRANSPOSE8x8_SH_SH(m0, n0, m1, n1, m2, n2, m3, n3, m0, n0, m1, n1, m2, n2, m3,
+                     n3);
+  ST_SH4(m0, n0, m1, n1, (dst + 0), 32);
+  ST_SH4(m2, n2, m3, n3, (dst + 4 * 32), 32);
+
+  TRANSPOSE8x8_SH_SH(m4, n4, m5, n5, m6, n6, m7, n7, m4, n4, m5, n5, m6, n6, m7,
+                     n7);
+  ST_SH4(m4, n4, m5, n5, (dst + 8), 32);
+  ST_SH4(m6, n6, m7, n7, (dst + 8 + 4 * 32), 32);
+
+  /* 3rd & 4th 8x8 */
+  LD_SH8((tmp_buf + 8 * 16), 8, m0, n0, m1, n1, m2, n2, m3, n3);
+  LD_SH8((tmp_buf + 12 * 16), 8, m4, n4, m5, n5, m6, n6, m7, n7);
+  TRANSPOSE8x8_SH_SH(m0, n0, m1, n1, m2, n2, m3, n3, m0, n0, m1, n1, m2, n2, m3,
+                     n3);
+  ST_SH4(m0, n0, m1, n1, (dst + 16), 32);
+  ST_SH4(m2, n2, m3, n3, (dst + 16 + 4 * 32), 32);
+
+  TRANSPOSE8x8_SH_SH(m4, n4, m5, n5, m6, n6, m7, n7, m4, n4, m5, n5, m6, n6, m7,
+                     n7);
+  ST_SH4(m4, n4, m5, n5, (dst + 24), 32);
+  ST_SH4(m6, n6, m7, n7, (dst + 24 + 4 * 32), 32);
+}
+
+static void idct32x8_1d_rows_msa(const int16_t *input, int16_t *output) {
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[8 * 32]);
+  DECLARE_ALIGNED(32, int16_t, tmp_odd_buf[16 * 8]);
+  DECLARE_ALIGNED(32, int16_t, tmp_eve_buf[16 * 8]);
+
+  idct32x8_row_transpose_store(input, &tmp_buf[0]);
+  idct32x8_row_even_process_store(&tmp_buf[0], &tmp_eve_buf[0]);
+  idct32x8_row_odd_process_store(&tmp_buf[0], &tmp_odd_buf[0]);
+  idct_butterfly_transpose_store(&tmp_buf[0], &tmp_eve_buf[0], &tmp_odd_buf[0],
+                                 output);
+}
+
+static void idct8x32_column_even_process_store(int16_t *tmp_buf,
+                                               int16_t *tmp_eve_buf) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+  v8i16 stp0, stp1, stp2, stp3, stp4, stp5, stp6, stp7;
+
+  /* Even stage 1 */
+  LD_SH8(tmp_buf, (4 * 32), reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+  tmp_buf += (2 * 32);
+
+  DOTP_CONST_PAIR(reg1, reg7, cospi_28_64, cospi_4_64, reg1, reg7);
+  DOTP_CONST_PAIR(reg5, reg3, cospi_12_64, cospi_20_64, reg5, reg3);
+  BUTTERFLY_4(reg1, reg7, reg3, reg5, vec1, vec3, vec2, vec0);
+  DOTP_CONST_PAIR(vec2, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+
+  loc1 = vec3;
+  loc0 = vec1;
+
+  DOTP_CONST_PAIR(reg0, reg4, cospi_16_64, cospi_16_64, reg0, reg4);
+  DOTP_CONST_PAIR(reg2, reg6, cospi_24_64, cospi_8_64, reg2, reg6);
+  BUTTERFLY_4(reg4, reg0, reg2, reg6, vec1, vec3, vec2, vec0);
+  BUTTERFLY_4(vec0, vec1, loc1, loc0, stp3, stp0, stp7, stp4);
+  BUTTERFLY_4(vec2, vec3, loc3, loc2, stp2, stp1, stp6, stp5);
+
+  /* Even stage 2 */
+  /* Load 8 */
+  LD_SH8(tmp_buf, (4 * 32), reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7);
+
+  DOTP_CONST_PAIR(reg0, reg7, cospi_30_64, cospi_2_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_14_64, cospi_18_64, reg4, reg3);
+  DOTP_CONST_PAIR(reg2, reg5, cospi_22_64, cospi_10_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg6, reg1, cospi_6_64, cospi_26_64, reg6, reg1);
+
+  vec0 = reg0 + reg4;
+  reg0 = reg0 - reg4;
+  reg4 = reg6 + reg2;
+  reg6 = reg6 - reg2;
+  reg2 = reg1 + reg5;
+  reg1 = reg1 - reg5;
+  reg5 = reg7 + reg3;
+  reg7 = reg7 - reg3;
+  reg3 = vec0;
+
+  vec1 = reg2;
+  reg2 = reg3 + reg4;
+  reg3 = reg3 - reg4;
+  reg4 = reg5 - vec1;
+  reg5 = reg5 + vec1;
+
+  DOTP_CONST_PAIR(reg7, reg0, cospi_24_64, cospi_8_64, reg0, reg7);
+  DOTP_CONST_PAIR((-reg6), reg1, cospi_24_64, cospi_8_64, reg6, reg1);
+
+  vec0 = reg0 - reg6;
+  reg0 = reg0 + reg6;
+  vec1 = reg7 - reg1;
+  reg7 = reg7 + reg1;
+
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, reg6, reg1);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_16_64, cospi_16_64, reg3, reg4);
+
+  /* Even stage 3 : Dependency on Even stage 1 & Even stage 2 */
+  /* Store 8 */
+  BUTTERFLY_4(stp0, stp1, reg7, reg5, loc1, loc3, loc2, loc0);
+  ST_SH2(loc1, loc3, tmp_eve_buf, 8);
+  ST_SH2(loc2, loc0, (tmp_eve_buf + 14 * 8), 8);
+
+  BUTTERFLY_4(stp2, stp3, reg4, reg1, loc1, loc3, loc2, loc0);
+  ST_SH2(loc1, loc3, (tmp_eve_buf + 2 * 8), 8);
+  ST_SH2(loc2, loc0, (tmp_eve_buf + 12 * 8), 8);
+
+  /* Store 8 */
+  BUTTERFLY_4(stp4, stp5, reg6, reg3, loc1, loc3, loc2, loc0);
+  ST_SH2(loc1, loc3, (tmp_eve_buf + 4 * 8), 8);
+  ST_SH2(loc2, loc0, (tmp_eve_buf + 10 * 8), 8);
+
+  BUTTERFLY_4(stp6, stp7, reg2, reg0, loc1, loc3, loc2, loc0);
+  ST_SH2(loc1, loc3, (tmp_eve_buf + 6 * 8), 8);
+  ST_SH2(loc2, loc0, (tmp_eve_buf + 8 * 8), 8);
+}
+
+static void idct8x32_column_odd_process_store(int16_t *tmp_buf,
+                                              int16_t *tmp_odd_buf) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 reg0, reg1, reg2, reg3, reg4, reg5, reg6, reg7;
+
+  /* Odd stage 1 */
+  reg0 = LD_SH(tmp_buf + 32);
+  reg1 = LD_SH(tmp_buf + 7 * 32);
+  reg2 = LD_SH(tmp_buf + 9 * 32);
+  reg3 = LD_SH(tmp_buf + 15 * 32);
+  reg4 = LD_SH(tmp_buf + 17 * 32);
+  reg5 = LD_SH(tmp_buf + 23 * 32);
+  reg6 = LD_SH(tmp_buf + 25 * 32);
+  reg7 = LD_SH(tmp_buf + 31 * 32);
+
+  DOTP_CONST_PAIR(reg0, reg7, cospi_31_64, cospi_1_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg4, reg3, cospi_15_64, cospi_17_64, reg3, reg4);
+  DOTP_CONST_PAIR(reg2, reg5, cospi_23_64, cospi_9_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg6, reg1, cospi_7_64, cospi_25_64, reg1, reg6);
+
+  vec0 = reg0 + reg3;
+  reg0 = reg0 - reg3;
+  reg3 = reg7 + reg4;
+  reg7 = reg7 - reg4;
+  reg4 = reg1 + reg2;
+  reg1 = reg1 - reg2;
+  reg2 = reg6 + reg5;
+  reg6 = reg6 - reg5;
+  reg5 = vec0;
+
+  /* 4 Stores */
+  ADD2(reg5, reg4, reg3, reg2, vec0, vec1);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 4 * 8), 8);
+  SUB2(reg5, reg4, reg3, reg2, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_24_64, cospi_8_64, vec0, vec1);
+  ST_SH2(vec0, vec1, tmp_odd_buf, 8);
+
+  /* 4 Stores */
+  DOTP_CONST_PAIR(reg7, reg0, cospi_28_64, cospi_4_64, reg0, reg7);
+  DOTP_CONST_PAIR(reg6, reg1, -cospi_4_64, cospi_28_64, reg1, reg6);
+  BUTTERFLY_4(reg0, reg7, reg6, reg1, vec0, vec1, vec2, vec3);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 6 * 8), 8);
+  DOTP_CONST_PAIR(vec2, vec3, cospi_24_64, cospi_8_64, vec2, vec3);
+  ST_SH2(vec2, vec3, (tmp_odd_buf + 2 * 8), 8);
+
+  /* Odd stage 2 */
+  /* 8 loads */
+  reg0 = LD_SH(tmp_buf + 3 * 32);
+  reg1 = LD_SH(tmp_buf + 5 * 32);
+  reg2 = LD_SH(tmp_buf + 11 * 32);
+  reg3 = LD_SH(tmp_buf + 13 * 32);
+  reg4 = LD_SH(tmp_buf + 19 * 32);
+  reg5 = LD_SH(tmp_buf + 21 * 32);
+  reg6 = LD_SH(tmp_buf + 27 * 32);
+  reg7 = LD_SH(tmp_buf + 29 * 32);
+
+  DOTP_CONST_PAIR(reg1, reg6, cospi_27_64, cospi_5_64, reg1, reg6);
+  DOTP_CONST_PAIR(reg5, reg2, cospi_11_64, cospi_21_64, reg2, reg5);
+  DOTP_CONST_PAIR(reg3, reg4, cospi_19_64, cospi_13_64, reg3, reg4);
+  DOTP_CONST_PAIR(reg7, reg0, cospi_3_64, cospi_29_64, reg0, reg7);
+
+  /* 4 Stores */
+  SUB4(reg1, reg2, reg6, reg5, reg0, reg3, reg7, reg4, vec0, vec1, vec2, vec3);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_12_64, cospi_20_64, loc0, loc1);
+  DOTP_CONST_PAIR(vec3, vec2, -cospi_20_64, cospi_12_64, loc2, loc3);
+  BUTTERFLY_4(loc2, loc3, loc1, loc0, vec0, vec1, vec3, vec2);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 12 * 8), 3 * 8);
+  DOTP_CONST_PAIR(vec3, vec2, -cospi_8_64, cospi_24_64, vec0, vec1);
+  ST_SH2(vec0, vec1, (tmp_odd_buf + 10 * 8), 8);
+
+  /* 4 Stores */
+  ADD4(reg0, reg3, reg1, reg2, reg5, reg6, reg4, reg7, vec0, vec1, vec2, vec3);
+  BUTTERFLY_4(vec0, vec3, vec2, vec1, reg0, reg1, reg3, reg2);
+  ST_SH2(reg0, reg1, (tmp_odd_buf + 13 * 8), 8);
+  DOTP_CONST_PAIR(reg3, reg2, -cospi_8_64, cospi_24_64, reg0, reg1);
+  ST_SH2(reg0, reg1, (tmp_odd_buf + 8 * 8), 8);
+
+  /* Odd stage 3 : Dependency on Odd stage 1 & Odd stage 2 */
+  /* Load 8 & Store 8 */
+  LD_SH4(tmp_odd_buf, 8, reg0, reg1, reg2, reg3);
+  LD_SH4((tmp_odd_buf + 8 * 8), 8, reg4, reg5, reg6, reg7);
+
+  ADD4(reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7, loc0, loc1, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, tmp_odd_buf, 8);
+
+  SUB2(reg0, reg4, reg1, reg5, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc0, loc1);
+
+  SUB2(reg2, reg6, reg3, reg7, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 8 * 8), 8);
+
+  /* Load 8 & Store 8 */
+  LD_SH4((tmp_odd_buf + 4 * 8), 8, reg1, reg2, reg0, reg3);
+  LD_SH4((tmp_odd_buf + 12 * 8), 8, reg4, reg5, reg6, reg7);
+
+  ADD4(reg0, reg4, reg1, reg5, reg2, reg6, reg3, reg7, loc0, loc1, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 4 * 8), 8);
+
+  SUB2(reg0, reg4, reg3, reg7, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc0, loc1);
+
+  SUB2(reg1, reg5, reg2, reg6, vec0, vec1);
+  DOTP_CONST_PAIR(vec1, vec0, cospi_16_64, cospi_16_64, loc2, loc3);
+  ST_SH4(loc0, loc1, loc2, loc3, (tmp_odd_buf + 12 * 8), 8);
+}
+
+static void idct8x32_column_butterfly_addblk(int16_t *tmp_eve_buf,
+                                             int16_t *tmp_odd_buf, uint8_t *dst,
+                                             int32_t dst_stride) {
+  v8i16 vec0, vec1, vec2, vec3, loc0, loc1, loc2, loc3;
+  v8i16 m0, m1, m2, m3, m4, m5, m6, m7, n0, n1, n2, n3, n4, n5, n6, n7;
+
+  /* FINAL BUTTERFLY : Dependency on Even & Odd */
+  vec0 = LD_SH(tmp_odd_buf);
+  vec1 = LD_SH(tmp_odd_buf + 9 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 14 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 6 * 8);
+  loc0 = LD_SH(tmp_eve_buf);
+  loc1 = LD_SH(tmp_eve_buf + 8 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 4 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 12 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m0, m4, m2, m6);
+  SRARI_H4_SH(m0, m2, m4, m6, 6);
+  AOM_ADDBLK_ST8x4_UB(dst, (4 * dst_stride), m0, m2, m4, m6);
+
+  SUB4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m6, m2, m4, m0);
+  SRARI_H4_SH(m0, m2, m4, m6, 6);
+  AOM_ADDBLK_ST8x4_UB((dst + 19 * dst_stride), (4 * dst_stride), m0, m2, m4,
+                      m6);
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 4 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 13 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 10 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 3 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 2 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 10 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 6 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 14 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m1, m5, m3, m7);
+  SRARI_H4_SH(m1, m3, m5, m7, 6);
+  AOM_ADDBLK_ST8x4_UB((dst + 2 * dst_stride), (4 * dst_stride), m1, m3, m5, m7);
+
+  SUB4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, m7, m3, m5, m1);
+  SRARI_H4_SH(m1, m3, m5, m7, 6);
+  AOM_ADDBLK_ST8x4_UB((dst + 17 * dst_stride), (4 * dst_stride), m1, m3, m5,
+                      m7);
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 2 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 11 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 12 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 7 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 1 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 9 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 5 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 13 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n0, n4, n2, n6);
+  SRARI_H4_SH(n0, n2, n4, n6, 6);
+  AOM_ADDBLK_ST8x4_UB((dst + 1 * dst_stride), (4 * dst_stride), n0, n2, n4, n6);
+
+  SUB4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n6, n2, n4, n0);
+  SRARI_H4_SH(n0, n2, n4, n6, 6);
+  AOM_ADDBLK_ST8x4_UB((dst + 18 * dst_stride), (4 * dst_stride), n0, n2, n4,
+                      n6);
+
+  /* Load 8 & Store 8 */
+  vec0 = LD_SH(tmp_odd_buf + 5 * 8);
+  vec1 = LD_SH(tmp_odd_buf + 15 * 8);
+  vec2 = LD_SH(tmp_odd_buf + 8 * 8);
+  vec3 = LD_SH(tmp_odd_buf + 1 * 8);
+  loc0 = LD_SH(tmp_eve_buf + 3 * 8);
+  loc1 = LD_SH(tmp_eve_buf + 11 * 8);
+  loc2 = LD_SH(tmp_eve_buf + 7 * 8);
+  loc3 = LD_SH(tmp_eve_buf + 15 * 8);
+
+  ADD4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n1, n5, n3, n7);
+  SRARI_H4_SH(n1, n3, n5, n7, 6);
+  AOM_ADDBLK_ST8x4_UB((dst + 3 * dst_stride), (4 * dst_stride), n1, n3, n5, n7);
+
+  SUB4(loc0, vec3, loc1, vec2, loc2, vec1, loc3, vec0, n7, n3, n5, n1);
+  SRARI_H4_SH(n1, n3, n5, n7, 6);
+  AOM_ADDBLK_ST8x4_UB((dst + 16 * dst_stride), (4 * dst_stride), n1, n3, n5,
+                      n7);
+}
+
+static void idct8x32_1d_columns_addblk_msa(int16_t *input, uint8_t *dst,
+                                           int32_t dst_stride) {
+  DECLARE_ALIGNED(32, int16_t, tmp_odd_buf[16 * 8]);
+  DECLARE_ALIGNED(32, int16_t, tmp_eve_buf[16 * 8]);
+
+  idct8x32_column_even_process_store(input, &tmp_eve_buf[0]);
+  idct8x32_column_odd_process_store(input, &tmp_odd_buf[0]);
+  idct8x32_column_butterfly_addblk(&tmp_eve_buf[0], &tmp_odd_buf[0], dst,
+                                   dst_stride);
+}
+
+void aom_idct32x32_1024_add_msa(const int16_t *input, uint8_t *dst,
+                                int32_t dst_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, out_arr[32 * 32]);
+  int16_t *out_ptr = out_arr;
+
+  /* transform rows */
+  for (i = 0; i < 4; ++i) {
+    /* process 32 * 8 block */
+    idct32x8_1d_rows_msa((input + (i << 8)), (out_ptr + (i << 8)));
+  }
+
+  /* transform columns */
+  for (i = 0; i < 4; ++i) {
+    /* process 8 * 32 block */
+    idct8x32_1d_columns_addblk_msa((out_ptr + (i << 3)), (dst + (i << 3)),
+                                   dst_stride);
+  }
+}
+
+void aom_idct32x32_34_add_msa(const int16_t *input, uint8_t *dst,
+                              int32_t dst_stride) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, out_arr[32 * 32]);
+  int16_t *out_ptr = out_arr;
+
+  for (i = 32; i--;) {
+    __asm__ __volatile__(
+        "sw     $zero,      0(%[out_ptr])     \n\t"
+        "sw     $zero,      4(%[out_ptr])     \n\t"
+        "sw     $zero,      8(%[out_ptr])     \n\t"
+        "sw     $zero,     12(%[out_ptr])     \n\t"
+        "sw     $zero,     16(%[out_ptr])     \n\t"
+        "sw     $zero,     20(%[out_ptr])     \n\t"
+        "sw     $zero,     24(%[out_ptr])     \n\t"
+        "sw     $zero,     28(%[out_ptr])     \n\t"
+        "sw     $zero,     32(%[out_ptr])     \n\t"
+        "sw     $zero,     36(%[out_ptr])     \n\t"
+        "sw     $zero,     40(%[out_ptr])     \n\t"
+        "sw     $zero,     44(%[out_ptr])     \n\t"
+        "sw     $zero,     48(%[out_ptr])     \n\t"
+        "sw     $zero,     52(%[out_ptr])     \n\t"
+        "sw     $zero,     56(%[out_ptr])     \n\t"
+        "sw     $zero,     60(%[out_ptr])     \n\t"
+
+        :
+        : [out_ptr] "r"(out_ptr));
+
+    out_ptr += 32;
+  }
+
+  out_ptr = out_arr;
+
+  /* rows: only upper-left 8x8 has non-zero coeff */
+  idct32x8_1d_rows_msa(input, out_ptr);
+
+  /* transform columns */
+  for (i = 0; i < 4; ++i) {
+    /* process 8 * 32 block */
+    idct8x32_1d_columns_addblk_msa((out_ptr + (i << 3)), (dst + (i << 3)),
+                                   dst_stride);
+  }
+}
+
+void aom_idct32x32_1_add_msa(const int16_t *input, uint8_t *dst,
+                             int32_t dst_stride) {
+  int32_t i;
+  int16_t out;
+  v16u8 dst0, dst1, dst2, dst3, tmp0, tmp1, tmp2, tmp3;
+  v8i16 res0, res1, res2, res3, res4, res5, res6, res7, vec;
+
+  out = ROUND_POWER_OF_TWO((input[0] * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO((out * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO(out, 6);
+
+  vec = __msa_fill_h(out);
+
+  for (i = 16; i--;) {
+    LD_UB2(dst, 16, dst0, dst1);
+    LD_UB2(dst + dst_stride, 16, dst2, dst3);
+
+    UNPCK_UB_SH(dst0, res0, res4);
+    UNPCK_UB_SH(dst1, res1, res5);
+    UNPCK_UB_SH(dst2, res2, res6);
+    UNPCK_UB_SH(dst3, res3, res7);
+    ADD4(res0, vec, res1, vec, res2, vec, res3, vec, res0, res1, res2, res3);
+    ADD4(res4, vec, res5, vec, res6, vec, res7, vec, res4, res5, res6, res7);
+    CLIP_SH4_0_255(res0, res1, res2, res3);
+    CLIP_SH4_0_255(res4, res5, res6, res7);
+    PCKEV_B4_UB(res4, res0, res5, res1, res6, res2, res7, res3, tmp0, tmp1,
+                tmp2, tmp3);
+
+    ST_UB2(tmp0, tmp1, dst, 16);
+    dst += dst_stride;
+    ST_UB2(tmp2, tmp3, dst, 16);
+    dst += dst_stride;
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/idct4x4_msa.c b/third_party/aom/aom_dsp/mips/idct4x4_msa.c
new file mode 100644
index 0000000000..274818baae
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/idct4x4_msa.c
@@ -0,0 +1,99 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/inv_txfm_msa.h"
+
+void aom_iwht4x4_16_add_msa(const int16_t *input, uint8_t *dst,
+                            int32_t dst_stride) {
+  v8i16 in0, in1, in2, in3;
+  v4i32 in0_r, in1_r, in2_r, in3_r, in4_r;
+
+  /* load vector elements of 4x4 block */
+  LD4x4_SH(input, in0, in2, in3, in1);
+  TRANSPOSE4x4_SH_SH(in0, in2, in3, in1, in0, in2, in3, in1);
+  UNPCK_R_SH_SW(in0, in0_r);
+  UNPCK_R_SH_SW(in2, in2_r);
+  UNPCK_R_SH_SW(in3, in3_r);
+  UNPCK_R_SH_SW(in1, in1_r);
+  SRA_4V(in0_r, in1_r, in2_r, in3_r, UNIT_QUANT_SHIFT);
+
+  in0_r += in2_r;
+  in3_r -= in1_r;
+  in4_r = (in0_r - in3_r) >> 1;
+  in1_r = in4_r - in1_r;
+  in2_r = in4_r - in2_r;
+  in0_r -= in1_r;
+  in3_r += in2_r;
+
+  TRANSPOSE4x4_SW_SW(in0_r, in1_r, in2_r, in3_r, in0_r, in1_r, in2_r, in3_r);
+
+  in0_r += in1_r;
+  in2_r -= in3_r;
+  in4_r = (in0_r - in2_r) >> 1;
+  in3_r = in4_r - in3_r;
+  in1_r = in4_r - in1_r;
+  in0_r -= in3_r;
+  in2_r += in1_r;
+
+  PCKEV_H4_SH(in0_r, in0_r, in1_r, in1_r, in2_r, in2_r, in3_r, in3_r, in0, in1,
+              in2, in3);
+  ADDBLK_ST4x4_UB(in0, in3, in1, in2, dst, dst_stride);
+}
+
+void aom_iwht4x4_1_add_msa(const int16_t *input, uint8_t *dst,
+                           int32_t dst_stride) {
+  int16_t a1, e1;
+  v8i16 in1, in0 = { 0 };
+
+  a1 = input[0] >> UNIT_QUANT_SHIFT;
+  e1 = a1 >> 1;
+  a1 -= e1;
+
+  in0 = __msa_insert_h(in0, 0, a1);
+  in0 = __msa_insert_h(in0, 1, e1);
+  in0 = __msa_insert_h(in0, 2, e1);
+  in0 = __msa_insert_h(in0, 3, e1);
+
+  in1 = in0 >> 1;
+  in0 -= in1;
+
+  ADDBLK_ST4x4_UB(in0, in1, in1, in1, dst, dst_stride);
+}
+
+void aom_idct4x4_16_add_msa(const int16_t *input, uint8_t *dst,
+                            int32_t dst_stride) {
+  v8i16 in0, in1, in2, in3;
+
+  /* load vector elements of 4x4 block */
+  LD4x4_SH(input, in0, in1, in2, in3);
+  /* rows */
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  AOM_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+  /* columns */
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  AOM_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+  /* rounding (add 2^3, divide by 2^4) */
+  SRARI_H4_SH(in0, in1, in2, in3, 4);
+  ADDBLK_ST4x4_UB(in0, in1, in2, in3, dst, dst_stride);
+}
+
+void aom_idct4x4_1_add_msa(const int16_t *input, uint8_t *dst,
+                           int32_t dst_stride) {
+  int16_t out;
+  v8i16 vec;
+
+  out = ROUND_POWER_OF_TWO((input[0] * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO((out * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO(out, 4);
+  vec = __msa_fill_h(out);
+
+  ADDBLK_ST4x4_UB(vec, vec, vec, vec, dst, dst_stride);
+}
diff --git a/third_party/aom/aom_dsp/mips/idct8x8_msa.c b/third_party/aom/aom_dsp/mips/idct8x8_msa.c
new file mode 100644
index 0000000000..981c103cdc
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/idct8x8_msa.c
@@ -0,0 +1,117 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/inv_txfm_msa.h"
+
+void aom_idct8x8_64_add_msa(const int16_t *input, uint8_t *dst,
+                            int32_t dst_stride) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  /* load vector elements of 8x8 block */
+  LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+
+  /* rows transform */
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  /* 1D idct8x8 */
+  AOM_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                 in4, in5, in6, in7);
+  /* columns transform */
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  /* 1D idct8x8 */
+  AOM_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                 in4, in5, in6, in7);
+  /* final rounding (add 2^4, divide by 2^5) and shift */
+  SRARI_H4_SH(in0, in1, in2, in3, 5);
+  SRARI_H4_SH(in4, in5, in6, in7, 5);
+  /* add block and store 8x8 */
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, in0, in1, in2, in3);
+  dst += (4 * dst_stride);
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, in4, in5, in6, in7);
+}
+
+void aom_idct8x8_12_add_msa(const int16_t *input, uint8_t *dst,
+                            int32_t dst_stride) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+  v8i16 s0, s1, s2, s3, s4, s5, s6, s7, k0, k1, k2, k3, m0, m1, m2, m3;
+  v4i32 tmp0, tmp1, tmp2, tmp3;
+  v8i16 zero = { 0 };
+
+  /* load vector elements of 8x8 block */
+  LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+  TRANSPOSE8X4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+
+  /* stage1 */
+  ILVL_H2_SH(in3, in0, in2, in1, s0, s1);
+  k0 = AOM_SET_COSPI_PAIR(cospi_28_64, -cospi_4_64);
+  k1 = AOM_SET_COSPI_PAIR(cospi_4_64, cospi_28_64);
+  k2 = AOM_SET_COSPI_PAIR(-cospi_20_64, cospi_12_64);
+  k3 = AOM_SET_COSPI_PAIR(cospi_12_64, cospi_20_64);
+  DOTP_SH4_SW(s0, s0, s1, s1, k0, k1, k2, k3, tmp0, tmp1, tmp2, tmp3);
+  SRARI_W4_SW(tmp0, tmp1, tmp2, tmp3, DCT_CONST_BITS);
+  PCKEV_H2_SH(zero, tmp0, zero, tmp1, s0, s1);
+  PCKEV_H2_SH(zero, tmp2, zero, tmp3, s2, s3);
+  BUTTERFLY_4(s0, s1, s3, s2, s4, s7, s6, s5);
+
+  /* stage2 */
+  ILVR_H2_SH(in3, in1, in2, in0, s1, s0);
+  k0 = AOM_SET_COSPI_PAIR(cospi_16_64, cospi_16_64);
+  k1 = AOM_SET_COSPI_PAIR(cospi_16_64, -cospi_16_64);
+  k2 = AOM_SET_COSPI_PAIR(cospi_24_64, -cospi_8_64);
+  k3 = AOM_SET_COSPI_PAIR(cospi_8_64, cospi_24_64);
+  DOTP_SH4_SW(s0, s0, s1, s1, k0, k1, k2, k3, tmp0, tmp1, tmp2, tmp3);
+  SRARI_W4_SW(tmp0, tmp1, tmp2, tmp3, DCT_CONST_BITS);
+  PCKEV_H2_SH(zero, tmp0, zero, tmp1, s0, s1);
+  PCKEV_H2_SH(zero, tmp2, zero, tmp3, s2, s3);
+  BUTTERFLY_4(s0, s1, s2, s3, m0, m1, m2, m3);
+
+  /* stage3 */
+  s0 = __msa_ilvr_h(s6, s5);
+
+  k1 = AOM_SET_COSPI_PAIR(-cospi_16_64, cospi_16_64);
+  DOTP_SH2_SW(s0, s0, k1, k0, tmp0, tmp1);
+  SRARI_W2_SW(tmp0, tmp1, DCT_CONST_BITS);
+  PCKEV_H2_SH(zero, tmp0, zero, tmp1, s2, s3);
+
+  /* stage4 */
+  BUTTERFLY_8(m0, m1, m2, m3, s4, s2, s3, s7, in0, in1, in2, in3, in4, in5, in6,
+              in7);
+  TRANSPOSE4X8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  AOM_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                 in4, in5, in6, in7);
+
+  /* final rounding (add 2^4, divide by 2^5) and shift */
+  SRARI_H4_SH(in0, in1, in2, in3, 5);
+  SRARI_H4_SH(in4, in5, in6, in7, 5);
+
+  /* add block and store 8x8 */
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, in0, in1, in2, in3);
+  dst += (4 * dst_stride);
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, in4, in5, in6, in7);
+}
+
+void aom_idct8x8_1_add_msa(const int16_t *input, uint8_t *dst,
+                           int32_t dst_stride) {
+  int16_t out;
+  int32_t val;
+  v8i16 vec;
+
+  out = ROUND_POWER_OF_TWO((input[0] * cospi_16_64), DCT_CONST_BITS);
+  out = ROUND_POWER_OF_TWO((out * cospi_16_64), DCT_CONST_BITS);
+  val = ROUND_POWER_OF_TWO(out, 5);
+  vec = __msa_fill_h(val);
+
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, vec, vec, vec, vec);
+  dst += (4 * dst_stride);
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, vec, vec, vec, vec);
+}
diff --git a/third_party/aom/aom_dsp/mips/intrapred16_dspr2.c b/third_party/aom/aom_dsp/mips/intrapred16_dspr2.c
new file mode 100644
index 0000000000..dc8f20208a
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/intrapred16_dspr2.c
@@ -0,0 +1,325 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/common_dspr2.h"
+
+#if HAVE_DSPR2
+void aom_h_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  int32_t tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
+  int32_t tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
+
+  __asm__ __volatile__(
+      "lb         %[tmp1],      (%[left])                    \n\t"
+      "lb         %[tmp2],      1(%[left])                   \n\t"
+      "lb         %[tmp3],      2(%[left])                   \n\t"
+      "lb         %[tmp4],      3(%[left])                   \n\t"
+      "lb         %[tmp5],      4(%[left])                   \n\t"
+      "lb         %[tmp6],      5(%[left])                   \n\t"
+      "lb         %[tmp7],      6(%[left])                   \n\t"
+      "lb         %[tmp8],      7(%[left])                   \n\t"
+      "lb         %[tmp9],      8(%[left])                   \n\t"
+      "lb         %[tmp10],     9(%[left])                   \n\t"
+      "lb         %[tmp11],     10(%[left])                  \n\t"
+      "lb         %[tmp12],     11(%[left])                  \n\t"
+      "lb         %[tmp13],     12(%[left])                  \n\t"
+      "lb         %[tmp14],     13(%[left])                  \n\t"
+      "lb         %[tmp15],     14(%[left])                  \n\t"
+      "lb         %[tmp16],     15(%[left])                  \n\t"
+
+      "replv.qb   %[tmp1],      %[tmp1]                      \n\t"
+      "replv.qb   %[tmp2],      %[tmp2]                      \n\t"
+      "replv.qb   %[tmp3],      %[tmp3]                      \n\t"
+      "replv.qb   %[tmp4],      %[tmp4]                      \n\t"
+      "replv.qb   %[tmp5],      %[tmp5]                      \n\t"
+      "replv.qb   %[tmp6],      %[tmp6]                      \n\t"
+      "replv.qb   %[tmp7],      %[tmp7]                      \n\t"
+      "replv.qb   %[tmp8],      %[tmp8]                      \n\t"
+      "replv.qb   %[tmp9],      %[tmp9]                      \n\t"
+      "replv.qb   %[tmp10],     %[tmp10]                     \n\t"
+      "replv.qb   %[tmp11],     %[tmp11]                     \n\t"
+      "replv.qb   %[tmp12],     %[tmp12]                     \n\t"
+      "replv.qb   %[tmp13],     %[tmp13]                     \n\t"
+      "replv.qb   %[tmp14],     %[tmp14]                     \n\t"
+      "replv.qb   %[tmp15],     %[tmp15]                     \n\t"
+      "replv.qb   %[tmp16],     %[tmp16]                     \n\t"
+
+      "sw         %[tmp1],      (%[dst])                     \n\t"
+      "sw         %[tmp1],      4(%[dst])                    \n\t"
+      "sw         %[tmp1],      8(%[dst])                    \n\t"
+      "sw         %[tmp1],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp2],      (%[dst])                     \n\t"
+      "sw         %[tmp2],      4(%[dst])                    \n\t"
+      "sw         %[tmp2],      8(%[dst])                    \n\t"
+      "sw         %[tmp2],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp3],      (%[dst])                     \n\t"
+      "sw         %[tmp3],      4(%[dst])                    \n\t"
+      "sw         %[tmp3],      8(%[dst])                    \n\t"
+      "sw         %[tmp3],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp4],      (%[dst])                     \n\t"
+      "sw         %[tmp4],      4(%[dst])                    \n\t"
+      "sw         %[tmp4],      8(%[dst])                    \n\t"
+      "sw         %[tmp4],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp5],      (%[dst])                     \n\t"
+      "sw         %[tmp5],      4(%[dst])                    \n\t"
+      "sw         %[tmp5],      8(%[dst])                    \n\t"
+      "sw         %[tmp5],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp6],      (%[dst])                     \n\t"
+      "sw         %[tmp6],      4(%[dst])                    \n\t"
+      "sw         %[tmp6],      8(%[dst])                    \n\t"
+      "sw         %[tmp6],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp7],      (%[dst])                     \n\t"
+      "sw         %[tmp7],      4(%[dst])                    \n\t"
+      "sw         %[tmp7],      8(%[dst])                    \n\t"
+      "sw         %[tmp7],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp8],      (%[dst])                     \n\t"
+      "sw         %[tmp8],      4(%[dst])                    \n\t"
+      "sw         %[tmp8],      8(%[dst])                    \n\t"
+      "sw         %[tmp8],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp9],      (%[dst])                     \n\t"
+      "sw         %[tmp9],      4(%[dst])                    \n\t"
+      "sw         %[tmp9],      8(%[dst])                    \n\t"
+      "sw         %[tmp9],      12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp10],     (%[dst])                     \n\t"
+      "sw         %[tmp10],     4(%[dst])                    \n\t"
+      "sw         %[tmp10],     8(%[dst])                    \n\t"
+      "sw         %[tmp10],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp11],     (%[dst])                     \n\t"
+      "sw         %[tmp11],     4(%[dst])                    \n\t"
+      "sw         %[tmp11],     8(%[dst])                    \n\t"
+      "sw         %[tmp11],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp12],     (%[dst])                     \n\t"
+      "sw         %[tmp12],     4(%[dst])                    \n\t"
+      "sw         %[tmp12],     8(%[dst])                    \n\t"
+      "sw         %[tmp12],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp13],     (%[dst])                     \n\t"
+      "sw         %[tmp13],     4(%[dst])                    \n\t"
+      "sw         %[tmp13],     8(%[dst])                    \n\t"
+      "sw         %[tmp13],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp14],     (%[dst])                     \n\t"
+      "sw         %[tmp14],     4(%[dst])                    \n\t"
+      "sw         %[tmp14],     8(%[dst])                    \n\t"
+      "sw         %[tmp14],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp15],     (%[dst])                     \n\t"
+      "sw         %[tmp15],     4(%[dst])                    \n\t"
+      "sw         %[tmp15],     8(%[dst])                    \n\t"
+      "sw         %[tmp15],     12(%[dst])                   \n\t"
+
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp16],     (%[dst])                     \n\t"
+      "sw         %[tmp16],     4(%[dst])                    \n\t"
+      "sw         %[tmp16],     8(%[dst])                    \n\t"
+      "sw         %[tmp16],     12(%[dst])                   \n\t"
+
+      : [tmp1] "=&r"(tmp1), [tmp2] "=&r"(tmp2), [tmp3] "=&r"(tmp3),
+        [tmp4] "=&r"(tmp4), [tmp5] "=&r"(tmp5), [tmp7] "=&r"(tmp7),
+        [tmp6] "=&r"(tmp6), [tmp8] "=&r"(tmp8), [tmp9] "=&r"(tmp9),
+        [tmp10] "=&r"(tmp10), [tmp11] "=&r"(tmp11), [tmp12] "=&r"(tmp12),
+        [tmp13] "=&r"(tmp13), [tmp14] "=&r"(tmp14), [tmp15] "=&r"(tmp15),
+        [tmp16] "=&r"(tmp16)
+      : [left] "r"(left), [dst] "r"(dst), [stride] "r"(stride));
+}
+
+void aom_dc_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  int32_t expected_dc;
+  int32_t average;
+  int32_t tmp, above1, above_l1, above_r1, left1, left_r1, left_l1;
+  int32_t above2, left2;
+
+  __asm__ __volatile__(
+      "lw              %[above1],           (%[above])                    \n\t"
+      "lw              %[above2],           4(%[above])                   \n\t"
+      "lw              %[left1],            (%[left])                     \n\t"
+      "lw              %[left2],            4(%[left])                    \n\t"
+
+      "preceu.ph.qbl   %[above_l1],         %[above1]                     \n\t"
+      "preceu.ph.qbr   %[above_r1],         %[above1]                     \n\t"
+      "preceu.ph.qbl   %[left_l1],          %[left1]                      \n\t"
+      "preceu.ph.qbr   %[left_r1],          %[left1]                      \n\t"
+
+      "addu.ph         %[average],          %[above_r1],     %[above_l1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[left_l1]   \n\t"
+      "addu.ph         %[average],          %[average],      %[left_r1]   \n\t"
+
+      "preceu.ph.qbl   %[above_l1],         %[above2]                     \n\t"
+      "preceu.ph.qbr   %[above_r1],         %[above2]                     \n\t"
+      "preceu.ph.qbl   %[left_l1],          %[left2]                      \n\t"
+      "preceu.ph.qbr   %[left_r1],          %[left2]                      \n\t"
+
+      "addu.ph         %[average],          %[average],      %[above_l1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[above_r1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[left_l1]   \n\t"
+      "addu.ph         %[average],          %[average],      %[left_r1]   \n\t"
+
+      "lw              %[above1],           8(%[above])                   \n\t"
+      "lw              %[above2],           12(%[above])                  \n\t"
+      "lw              %[left1],            8(%[left])                    \n\t"
+      "lw              %[left2],            12(%[left])                   \n\t"
+
+      "preceu.ph.qbl   %[above_l1],         %[above1]                     \n\t"
+      "preceu.ph.qbr   %[above_r1],         %[above1]                     \n\t"
+      "preceu.ph.qbl   %[left_l1],          %[left1]                      \n\t"
+      "preceu.ph.qbr   %[left_r1],          %[left1]                      \n\t"
+
+      "addu.ph         %[average],          %[average],      %[above_l1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[above_r1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[left_l1]   \n\t"
+      "addu.ph         %[average],          %[average],      %[left_r1]   \n\t"
+
+      "preceu.ph.qbl   %[above_l1],         %[above2]                     \n\t"
+      "preceu.ph.qbr   %[above_r1],         %[above2]                     \n\t"
+      "preceu.ph.qbl   %[left_l1],          %[left2]                      \n\t"
+      "preceu.ph.qbr   %[left_r1],          %[left2]                      \n\t"
+
+      "addu.ph         %[average],          %[average],      %[above_l1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[above_r1]  \n\t"
+      "addu.ph         %[average],          %[average],      %[left_l1]   \n\t"
+      "addu.ph         %[average],          %[average],      %[left_r1]   \n\t"
+
+      "addiu           %[average],          %[average],      16           \n\t"
+      "srl             %[tmp],              %[average],      16           \n\t"
+      "addu.ph         %[average],          %[tmp],          %[average]   \n\t"
+      "srl             %[expected_dc],      %[average],      5            \n\t"
+      "replv.qb        %[expected_dc],      %[expected_dc]                \n\t"
+
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      "add             %[dst],              %[dst],          %[stride]    \n\t"
+      "sw              %[expected_dc],      (%[dst])                      \n\t"
+      "sw              %[expected_dc],      4(%[dst])                     \n\t"
+      "sw              %[expected_dc],      8(%[dst])                     \n\t"
+      "sw              %[expected_dc],      12(%[dst])                    \n\t"
+
+      : [left1] "=&r"(left1), [above1] "=&r"(above1), [left_l1] "=&r"(left_l1),
+        [above_l1] "=&r"(above_l1), [left_r1] "=&r"(left_r1),
+        [above_r1] "=&r"(above_r1), [above2] "=&r"(above2),
+        [left2] "=&r"(left2), [average] "=&r"(average), [tmp] "=&r"(tmp),
+        [expected_dc] "=&r"(expected_dc)
+      : [above] "r"(above), [left] "r"(left), [dst] "r"(dst),
+        [stride] "r"(stride));
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/intrapred4_dspr2.c b/third_party/aom/aom_dsp/mips/intrapred4_dspr2.c
new file mode 100644
index 0000000000..ea7c02810a
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/intrapred4_dspr2.c
@@ -0,0 +1,225 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/common_dspr2.h"
+
+#if HAVE_DSPR2
+void aom_h_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int32_t tmp1, tmp2, tmp3, tmp4;
+
+  __asm__ __volatile__(
+      "lb         %[tmp1],      (%[left])                    \n\t"
+      "lb         %[tmp2],      1(%[left])                   \n\t"
+      "lb         %[tmp3],      2(%[left])                   \n\t"
+      "lb         %[tmp4],      3(%[left])                   \n\t"
+      "replv.qb   %[tmp1],      %[tmp1]                      \n\t"
+      "replv.qb   %[tmp2],      %[tmp2]                      \n\t"
+      "replv.qb   %[tmp3],      %[tmp3]                      \n\t"
+      "replv.qb   %[tmp4],      %[tmp4]                      \n\t"
+      "sw         %[tmp1],      (%[dst])                     \n\t"
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp2],      (%[dst])                     \n\t"
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp3],      (%[dst])                     \n\t"
+      "add        %[dst],       %[dst],         %[stride]    \n\t"
+      "sw         %[tmp4],      (%[dst])                     \n\t"
+
+      : [tmp1] "=&r"(tmp1), [tmp2] "=&r"(tmp2), [tmp3] "=&r"(tmp3),
+        [tmp4] "=&r"(tmp4)
+      : [left] "r"(left), [dst] "r"(dst), [stride] "r"(stride));
+}
+
+void aom_dc_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int32_t expected_dc;
+  int32_t average;
+  int32_t tmp, above_c, above_l, above_r, left_c, left_r, left_l;
+
+  __asm__ __volatile__(
+      "lw              %[above_c],         (%[above])                    \n\t"
+      "lw              %[left_c],          (%[left])                     \n\t"
+
+      "preceu.ph.qbl   %[above_l],         %[above_c]                    \n\t"
+      "preceu.ph.qbr   %[above_r],         %[above_c]                    \n\t"
+      "preceu.ph.qbl   %[left_l],          %[left_c]                     \n\t"
+      "preceu.ph.qbr   %[left_r],          %[left_c]                     \n\t"
+
+      "addu.ph         %[average],         %[above_r],       %[above_l]  \n\t"
+      "addu.ph         %[average],         %[average],       %[left_l]   \n\t"
+      "addu.ph         %[average],         %[average],       %[left_r]   \n\t"
+      "addiu           %[average],         %[average],       4           \n\t"
+      "srl             %[tmp],             %[average],       16          \n\t"
+      "addu.ph         %[average],         %[tmp],           %[average]  \n\t"
+      "srl             %[expected_dc],     %[average],       3           \n\t"
+      "replv.qb        %[expected_dc],     %[expected_dc]                \n\t"
+
+      "sw              %[expected_dc],     (%[dst])                      \n\t"
+      "add             %[dst],              %[dst],          %[stride]   \n\t"
+      "sw              %[expected_dc],     (%[dst])                      \n\t"
+      "add             %[dst],              %[dst],          %[stride]   \n\t"
+      "sw              %[expected_dc],     (%[dst])                      \n\t"
+      "add             %[dst],              %[dst],          %[stride]   \n\t"
+      "sw              %[expected_dc],     (%[dst])                      \n\t"
+
+      : [above_c] "=&r"(above_c), [above_l] "=&r"(above_l),
+        [above_r] "=&r"(above_r), [left_c] "=&r"(left_c),
+        [left_l] "=&r"(left_l), [left_r] "=&r"(left_r),
+        [average] "=&r"(average), [tmp] "=&r"(tmp),
+        [expected_dc] "=&r"(expected_dc)
+      : [above] "r"(above), [left] "r"(left), [dst] "r"(dst),
+        [stride] "r"(stride));
+}
+
+void aom_tm_predictor_4x4_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int32_t abovel, abover;
+  int32_t left0, left1, left2, left3;
+  int32_t res0, res1;
+  int32_t resl;
+  int32_t resr;
+  int32_t top_left;
+  uint8_t *cm = aom_ff_cropTbl;
+
+  __asm__ __volatile__(
+      "ulw             %[resl],       (%[above])                         \n\t"
+
+      "lbu             %[left0],       (%[left])                         \n\t"
+      "lbu             %[left1],       1(%[left])                        \n\t"
+      "lbu             %[left2],       2(%[left])                        \n\t"
+      "lbu             %[left3],       3(%[left])                        \n\t"
+
+      "lbu             %[top_left],    -1(%[above])                      \n\t"
+
+      "preceu.ph.qbl   %[abovel],      %[resl]                           \n\t"
+      "preceu.ph.qbr   %[abover],      %[resl]                           \n\t"
+
+      "replv.ph        %[left0],       %[left0]                          \n\t"
+      "replv.ph        %[left1],       %[left1]                          \n\t"
+      "replv.ph        %[left2],       %[left2]                          \n\t"
+      "replv.ph        %[left3],       %[left3]                          \n\t"
+
+      "replv.ph        %[top_left],    %[top_left]                       \n\t"
+
+      "addu.ph         %[resl],        %[abovel],         %[left0]       \n\t"
+      "subu.ph         %[resl],        %[resl],           %[top_left]    \n\t"
+
+      "addu.ph         %[resr],        %[abover],         %[left0]       \n\t"
+      "subu.ph         %[resr],        %[resr],           %[top_left]    \n\t"
+
+      "sll             %[res0],        %[resr],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+      "sra             %[res1],        %[resr],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+      "sb              %[res0],        (%[dst])                          \n\t"
+
+      "sll             %[res0],        %[resl],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+      "sb              %[res1],        1(%[dst])                         \n\t"
+
+      "sra             %[res1],        %[resl],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+
+      "addu.ph         %[resl],        %[abovel],         %[left1]       \n\t"
+      "subu.ph         %[resl],        %[resl],           %[top_left]    \n\t"
+
+      "addu.ph         %[resr],        %[abover],         %[left1]       \n\t"
+      "subu.ph         %[resr],        %[resr],           %[top_left]    \n\t"
+
+      "sb              %[res0],        2(%[dst])                         \n\t"
+      "sb              %[res1],        3(%[dst])                         \n\t"
+
+      "add             %[dst],          %[dst],           %[stride]      \n\t"
+
+      "sll             %[res0],        %[resr],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+      "sra             %[res1],        %[resr],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+      "sb              %[res0],        (%[dst])                          \n\t"
+
+      "sll             %[res0],        %[resl],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+      "sb              %[res1],        1(%[dst])                         \n\t"
+      "sra             %[res1],        %[resl],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+
+      "addu.ph         %[resl],        %[abovel],         %[left2]       \n\t"
+      "subu.ph         %[resl],        %[resl],           %[top_left]    \n\t"
+
+      "addu.ph         %[resr],        %[abover],         %[left2]       \n\t"
+      "subu.ph         %[resr],        %[resr],           %[top_left]    \n\t"
+
+      "sb              %[res0],        2(%[dst])                         \n\t"
+      "sb              %[res1],        3(%[dst])                         \n\t"
+
+      "add             %[dst],          %[dst],           %[stride]      \n\t"
+
+      "sll             %[res0],        %[resr],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+      "sra             %[res1],        %[resr],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+      "sb              %[res0],        (%[dst])                          \n\t"
+
+      "sll             %[res0],        %[resl],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+      "sb              %[res1],        1(%[dst])                         \n\t"
+      "sra             %[res1],        %[resl],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+
+      "addu.ph         %[resl],        %[abovel],        %[left3]        \n\t"
+      "subu.ph         %[resl],        %[resl],          %[top_left]     \n\t"
+
+      "addu.ph         %[resr],        %[abover],        %[left3]        \n\t"
+      "subu.ph         %[resr],        %[resr],          %[top_left]     \n\t"
+
+      "sb              %[res0],        2(%[dst])                         \n\t"
+      "sb              %[res1],        3(%[dst])                         \n\t"
+
+      "add             %[dst],          %[dst],          %[stride]       \n\t"
+
+      "sll             %[res0],        %[resr],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+
+      "sra             %[res1],        %[resr],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+      "sb              %[res0],        (%[dst])                          \n\t"
+
+      "sll             %[res0],        %[resl],           16             \n\t"
+      "sra             %[res0],        %[res0],           16             \n\t"
+      "lbux            %[res0],        %[res0](%[cm])                    \n\t"
+      "sb              %[res1],        1(%[dst])                         \n\t"
+
+      "sra             %[res1],        %[resl],           16             \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                    \n\t"
+
+      "sb              %[res0],        2(%[dst])                         \n\t"
+      "sb              %[res1],        3(%[dst])                         \n\t"
+
+      : [abovel] "=&r"(abovel), [abover] "=&r"(abover), [left0] "=&r"(left0),
+        [left1] "=&r"(left1), [left2] "=&r"(left2), [res0] "=&r"(res0),
+        [res1] "=&r"(res1), [left3] "=&r"(left3), [resl] "=&r"(resl),
+        [resr] "=&r"(resr), [top_left] "=&r"(top_left)
+      : [above] "r"(above), [left] "r"(left), [dst] "r"(dst),
+        [stride] "r"(stride), [cm] "r"(cm));
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/intrapred8_dspr2.c b/third_party/aom/aom_dsp/mips/intrapred8_dspr2.c
new file mode 100644
index 0000000000..1114fbc006
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/intrapred8_dspr2.c
@@ -0,0 +1,603 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/common_dspr2.h"
+
+#if HAVE_DSPR2
+void aom_h_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  int32_t tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
+
+  __asm__ __volatile__(
+      "lb         %[tmp1],      (%[left])                   \n\t"
+      "lb         %[tmp2],      1(%[left])                  \n\t"
+      "lb         %[tmp3],      2(%[left])                  \n\t"
+      "lb         %[tmp4],      3(%[left])                  \n\t"
+      "lb         %[tmp5],      4(%[left])                  \n\t"
+      "lb         %[tmp6],      5(%[left])                  \n\t"
+      "lb         %[tmp7],      6(%[left])                  \n\t"
+      "lb         %[tmp8],      7(%[left])                  \n\t"
+
+      "replv.qb   %[tmp1],      %[tmp1]                     \n\t"
+      "replv.qb   %[tmp2],      %[tmp2]                     \n\t"
+      "replv.qb   %[tmp3],      %[tmp3]                     \n\t"
+      "replv.qb   %[tmp4],      %[tmp4]                     \n\t"
+      "replv.qb   %[tmp5],      %[tmp5]                     \n\t"
+      "replv.qb   %[tmp6],      %[tmp6]                     \n\t"
+      "replv.qb   %[tmp7],      %[tmp7]                     \n\t"
+      "replv.qb   %[tmp8],      %[tmp8]                     \n\t"
+
+      "sw         %[tmp1],      (%[dst])                    \n\t"
+      "sw         %[tmp1],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp2],      (%[dst])                    \n\t"
+      "sw         %[tmp2],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp3],      (%[dst])                    \n\t"
+      "sw         %[tmp3],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp4],      (%[dst])                    \n\t"
+      "sw         %[tmp4],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp5],      (%[dst])                    \n\t"
+      "sw         %[tmp5],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp6],      (%[dst])                    \n\t"
+      "sw         %[tmp6],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp7],      (%[dst])                    \n\t"
+      "sw         %[tmp7],      4(%[dst])                   \n\t"
+      "add        %[dst],       %[dst],         %[stride]   \n\t"
+      "sw         %[tmp8],      (%[dst])                    \n\t"
+      "sw         %[tmp8],      4(%[dst])                   \n\t"
+
+      : [tmp1] "=&r"(tmp1), [tmp2] "=&r"(tmp2), [tmp3] "=&r"(tmp3),
+        [tmp4] "=&r"(tmp4), [tmp5] "=&r"(tmp5), [tmp7] "=&r"(tmp7),
+        [tmp6] "=&r"(tmp6), [tmp8] "=&r"(tmp8)
+      : [left] "r"(left), [dst] "r"(dst), [stride] "r"(stride));
+}
+
+void aom_dc_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int32_t expected_dc;
+  int32_t average;
+  int32_t tmp, above1, above_l1, above_r1, left1, left_r1, left_l1;
+  int32_t above2, above_l2, above_r2, left2, left_r2, left_l2;
+
+  __asm__ __volatile__(
+      "lw              %[above1],         (%[above])                      \n\t"
+      "lw              %[above2],         4(%[above])                     \n\t"
+      "lw              %[left1],          (%[left])                       \n\t"
+      "lw              %[left2],          4(%[left])                      \n\t"
+
+      "preceu.ph.qbl   %[above_l1],       %[above1]                       \n\t"
+      "preceu.ph.qbr   %[above_r1],       %[above1]                       \n\t"
+      "preceu.ph.qbl   %[left_l1],        %[left1]                        \n\t"
+      "preceu.ph.qbr   %[left_r1],        %[left1]                        \n\t"
+
+      "preceu.ph.qbl   %[above_l2],       %[above2]                       \n\t"
+      "preceu.ph.qbr   %[above_r2],       %[above2]                       \n\t"
+      "preceu.ph.qbl   %[left_l2],        %[left2]                        \n\t"
+      "preceu.ph.qbr   %[left_r2],        %[left2]                        \n\t"
+
+      "addu.ph         %[average],        %[above_r1],      %[above_l1]   \n\t"
+      "addu.ph         %[average],        %[average],       %[left_l1]    \n\t"
+      "addu.ph         %[average],        %[average],       %[left_r1]    \n\t"
+
+      "addu.ph         %[average],        %[average],       %[above_l2]   \n\t"
+      "addu.ph         %[average],        %[average],       %[above_r2]   \n\t"
+      "addu.ph         %[average],        %[average],       %[left_l2]    \n\t"
+      "addu.ph         %[average],        %[average],       %[left_r2]    \n\t"
+
+      "addiu           %[average],        %[average],       8             \n\t"
+
+      "srl             %[tmp],            %[average],       16            \n\t"
+      "addu.ph         %[average],        %[tmp],           %[average]    \n\t"
+      "srl             %[expected_dc],    %[average],       4             \n\t"
+      "replv.qb        %[expected_dc],    %[expected_dc]                  \n\t"
+
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      "add             %[dst],             %[dst],          %[stride]     \n\t"
+      "sw              %[expected_dc],    (%[dst])                        \n\t"
+      "sw              %[expected_dc],    4(%[dst])                       \n\t"
+
+      : [above1] "=&r"(above1), [above_l1] "=&r"(above_l1),
+        [above_r1] "=&r"(above_r1), [left1] "=&r"(left1),
+        [left_l1] "=&r"(left_l1), [left_r1] "=&r"(left_r1),
+        [above2] "=&r"(above2), [above_l2] "=&r"(above_l2),
+        [above_r2] "=&r"(above_r2), [left2] "=&r"(left2),
+        [left_l2] "=&r"(left_l2), [left_r2] "=&r"(left_r2),
+        [average] "=&r"(average), [tmp] "=&r"(tmp),
+        [expected_dc] "=&r"(expected_dc)
+      : [above] "r"(above), [left] "r"(left), [dst] "r"(dst),
+        [stride] "r"(stride));
+}
+
+void aom_tm_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  int32_t abovel, abover;
+  int32_t abovel_1, abover_1;
+  int32_t left0;
+  int32_t res0, res1, res2, res3;
+  int32_t reshw;
+  int32_t top_left;
+  uint8_t *cm = aom_ff_cropTbl;
+
+  __asm__ __volatile__(
+      "ulw             %[reshw],       (%[above])                         \n\t"
+      "ulw             %[top_left],    4(%[above])                        \n\t"
+
+      "lbu             %[left0],       (%[left])                          \n\t"
+
+      "preceu.ph.qbl   %[abovel],      %[reshw]                           \n\t"
+      "preceu.ph.qbr   %[abover],      %[reshw]                           \n\t"
+      "preceu.ph.qbl   %[abovel_1],    %[top_left]                        \n\t"
+      "preceu.ph.qbr   %[abover_1],    %[top_left]                        \n\t"
+
+      "lbu             %[top_left],    -1(%[above])                       \n\t"
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+
+      "replv.ph        %[top_left],    %[top_left]                        \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       1(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       2(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       3(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       4(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       5(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       6(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbu             %[left0],       7(%[left])                         \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      "replv.ph        %[left0],       %[left0]                           \n\t"
+      "add             %[dst],          %[dst],             %[stride]     \n\t"
+
+      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        (%[dst])                           \n\t"
+      "sb              %[res1],        1(%[dst])                          \n\t"
+      "sb              %[res2],        2(%[dst])                          \n\t"
+      "sb              %[res3],        3(%[dst])                          \n\t"
+
+      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res2],        %[reshw],            16            \n\t"
+      "sra             %[res2],        %[res2],             16            \n\t"
+      "sra             %[res3],        %[reshw],            16            \n\t"
+
+      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
+      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
+
+      "sll             %[res0],        %[reshw],            16            \n\t"
+      "sra             %[res0],        %[res0],             16            \n\t"
+      "sra             %[res1],        %[reshw],            16            \n\t"
+
+      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
+      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
+      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
+      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
+
+      "sb              %[res0],        4(%[dst])                          \n\t"
+      "sb              %[res1],        5(%[dst])                          \n\t"
+      "sb              %[res2],        6(%[dst])                          \n\t"
+      "sb              %[res3],        7(%[dst])                          \n\t"
+
+      : [abovel] "=&r"(abovel), [abover] "=&r"(abover),
+        [abovel_1] "=&r"(abovel_1), [abover_1] "=&r"(abover_1),
+        [left0] "=&r"(left0), [res2] "=&r"(res2), [res3] "=&r"(res3),
+        [res0] "=&r"(res0), [res1] "=&r"(res1), [reshw] "=&r"(reshw),
+        [top_left] "=&r"(top_left)
+      : [above] "r"(above), [left] "r"(left), [dst] "r"(dst),
+        [stride] "r"(stride), [cm] "r"(cm));
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/intrapred_msa.c b/third_party/aom/aom_dsp/mips/intrapred_msa.c
new file mode 100644
index 0000000000..e8eaec7a94
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/intrapred_msa.c
@@ -0,0 +1,739 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/macros_msa.h"
+
+#define IPRED_SUBS_UH2_UH(in0, in1, out0, out1) \
+  {                                             \
+    out0 = __msa_subs_u_h(out0, in0);           \
+    out1 = __msa_subs_u_h(out1, in1);           \
+  }
+
+static void intra_predict_vert_4x4_msa(const uint8_t *src, uint8_t *dst,
+                                       int32_t dst_stride) {
+  uint32_t src_data;
+
+  src_data = LW(src);
+
+  SW4(src_data, src_data, src_data, src_data, dst, dst_stride);
+}
+
+static void intra_predict_vert_8x8_msa(const uint8_t *src, uint8_t *dst,
+                                       int32_t dst_stride) {
+  uint32_t row;
+  uint32_t src_data1, src_data2;
+
+  src_data1 = LW(src);
+  src_data2 = LW(src + 4);
+
+  for (row = 8; row--;) {
+    SW(src_data1, dst);
+    SW(src_data2, (dst + 4));
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_vert_16x16_msa(const uint8_t *src, uint8_t *dst,
+                                         int32_t dst_stride) {
+  uint32_t row;
+  v16u8 src0;
+
+  src0 = LD_UB(src);
+
+  for (row = 16; row--;) {
+    ST_UB(src0, dst);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_vert_32x32_msa(const uint8_t *src, uint8_t *dst,
+                                         int32_t dst_stride) {
+  uint32_t row;
+  v16u8 src1, src2;
+
+  src1 = LD_UB(src);
+  src2 = LD_UB(src + 16);
+
+  for (row = 32; row--;) {
+    ST_UB2(src1, src2, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_horiz_4x4_msa(const uint8_t *src, uint8_t *dst,
+                                        int32_t dst_stride) {
+  uint32_t out0, out1, out2, out3;
+
+  out0 = src[0] * 0x01010101;
+  out1 = src[1] * 0x01010101;
+  out2 = src[2] * 0x01010101;
+  out3 = src[3] * 0x01010101;
+
+  SW4(out0, out1, out2, out3, dst, dst_stride);
+}
+
+static void intra_predict_horiz_8x8_msa(const uint8_t *src, uint8_t *dst,
+                                        int32_t dst_stride) {
+  uint64_t out0, out1, out2, out3, out4, out5, out6, out7;
+
+  out0 = src[0] * 0x0101010101010101ull;
+  out1 = src[1] * 0x0101010101010101ull;
+  out2 = src[2] * 0x0101010101010101ull;
+  out3 = src[3] * 0x0101010101010101ull;
+  out4 = src[4] * 0x0101010101010101ull;
+  out5 = src[5] * 0x0101010101010101ull;
+  out6 = src[6] * 0x0101010101010101ull;
+  out7 = src[7] * 0x0101010101010101ull;
+
+  SD4(out0, out1, out2, out3, dst, dst_stride);
+  dst += (4 * dst_stride);
+  SD4(out4, out5, out6, out7, dst, dst_stride);
+}
+
+static void intra_predict_horiz_16x16_msa(const uint8_t *src, uint8_t *dst,
+                                          int32_t dst_stride) {
+  uint32_t row;
+  uint8_t inp0, inp1, inp2, inp3;
+  v16u8 src0, src1, src2, src3;
+
+  for (row = 4; row--;) {
+    inp0 = src[0];
+    inp1 = src[1];
+    inp2 = src[2];
+    inp3 = src[3];
+    src += 4;
+
+    src0 = (v16u8)__msa_fill_b(inp0);
+    src1 = (v16u8)__msa_fill_b(inp1);
+    src2 = (v16u8)__msa_fill_b(inp2);
+    src3 = (v16u8)__msa_fill_b(inp3);
+
+    ST_UB4(src0, src1, src2, src3, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void intra_predict_horiz_32x32_msa(const uint8_t *src, uint8_t *dst,
+                                          int32_t dst_stride) {
+  uint32_t row;
+  uint8_t inp0, inp1, inp2, inp3;
+  v16u8 src0, src1, src2, src3;
+
+  for (row = 8; row--;) {
+    inp0 = src[0];
+    inp1 = src[1];
+    inp2 = src[2];
+    inp3 = src[3];
+    src += 4;
+
+    src0 = (v16u8)__msa_fill_b(inp0);
+    src1 = (v16u8)__msa_fill_b(inp1);
+    src2 = (v16u8)__msa_fill_b(inp2);
+    src3 = (v16u8)__msa_fill_b(inp3);
+
+    ST_UB2(src0, src0, dst, 16);
+    dst += dst_stride;
+    ST_UB2(src1, src1, dst, 16);
+    dst += dst_stride;
+    ST_UB2(src2, src2, dst, 16);
+    dst += dst_stride;
+    ST_UB2(src3, src3, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_dc_4x4_msa(const uint8_t *src_top,
+                                     const uint8_t *src_left, uint8_t *dst,
+                                     int32_t dst_stride) {
+  uint32_t val0, val1;
+  v16i8 store, src = { 0 };
+  v8u16 sum_h;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  val0 = LW(src_top);
+  val1 = LW(src_left);
+  INSERT_W2_SB(val0, val1, src);
+  sum_h = __msa_hadd_u_h((v16u8)src, (v16u8)src);
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 3);
+  store = __msa_splati_b((v16i8)sum_w, 0);
+  val0 = __msa_copy_u_w((v4i32)store, 0);
+
+  SW4(val0, val0, val0, val0, dst, dst_stride);
+}
+
+static void intra_predict_dc_tl_4x4_msa(const uint8_t *src, uint8_t *dst,
+                                        int32_t dst_stride) {
+  uint32_t val0;
+  v16i8 store, data = { 0 };
+  v8u16 sum_h;
+  v4u32 sum_w;
+
+  val0 = LW(src);
+  data = (v16i8)__msa_insert_w((v4i32)data, 0, val0);
+  sum_h = __msa_hadd_u_h((v16u8)data, (v16u8)data);
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_w, 2);
+  store = __msa_splati_b((v16i8)sum_w, 0);
+  val0 = __msa_copy_u_w((v4i32)store, 0);
+
+  SW4(val0, val0, val0, val0, dst, dst_stride);
+}
+
+static void intra_predict_128dc_4x4_msa(uint8_t *dst, int32_t dst_stride) {
+  uint32_t out;
+  const v16i8 store = __msa_ldi_b(128);
+
+  out = __msa_copy_u_w((v4i32)store, 0);
+
+  SW4(out, out, out, out, dst, dst_stride);
+}
+
+static void intra_predict_dc_8x8_msa(const uint8_t *src_top,
+                                     const uint8_t *src_left, uint8_t *dst,
+                                     int32_t dst_stride) {
+  uint64_t val0, val1;
+  v16i8 store;
+  v16u8 src = { 0 };
+  v8u16 sum_h;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  val0 = LD(src_top);
+  val1 = LD(src_left);
+  INSERT_D2_UB(val0, val1, src);
+  sum_h = __msa_hadd_u_h(src, src);
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 4);
+  store = __msa_splati_b((v16i8)sum_w, 0);
+  val0 = __msa_copy_u_d((v2i64)store, 0);
+
+  SD4(val0, val0, val0, val0, dst, dst_stride);
+  dst += (4 * dst_stride);
+  SD4(val0, val0, val0, val0, dst, dst_stride);
+}
+
+static void intra_predict_dc_tl_8x8_msa(const uint8_t *src, uint8_t *dst,
+                                        int32_t dst_stride) {
+  uint64_t val0;
+  v16i8 store;
+  v16u8 data = { 0 };
+  v8u16 sum_h;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  val0 = LD(src);
+  data = (v16u8)__msa_insert_d((v2i64)data, 0, val0);
+  sum_h = __msa_hadd_u_h(data, data);
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 3);
+  store = __msa_splati_b((v16i8)sum_w, 0);
+  val0 = __msa_copy_u_d((v2i64)store, 0);
+
+  SD4(val0, val0, val0, val0, dst, dst_stride);
+  dst += (4 * dst_stride);
+  SD4(val0, val0, val0, val0, dst, dst_stride);
+}
+
+static void intra_predict_128dc_8x8_msa(uint8_t *dst, int32_t dst_stride) {
+  uint64_t out;
+  const v16i8 store = __msa_ldi_b(128);
+
+  out = __msa_copy_u_d((v2i64)store, 0);
+
+  SD4(out, out, out, out, dst, dst_stride);
+  dst += (4 * dst_stride);
+  SD4(out, out, out, out, dst, dst_stride);
+}
+
+static void intra_predict_dc_16x16_msa(const uint8_t *src_top,
+                                       const uint8_t *src_left, uint8_t *dst,
+                                       int32_t dst_stride) {
+  v16u8 top, left, out;
+  v8u16 sum_h, sum_top, sum_left;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  top = LD_UB(src_top);
+  left = LD_UB(src_left);
+  HADD_UB2_UH(top, left, sum_top, sum_left);
+  sum_h = sum_top + sum_left;
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 5);
+  out = (v16u8)__msa_splati_b((v16i8)sum_w, 0);
+
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+  dst += (8 * dst_stride);
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+}
+
+static void intra_predict_dc_tl_16x16_msa(const uint8_t *src, uint8_t *dst,
+                                          int32_t dst_stride) {
+  v16u8 data, out;
+  v8u16 sum_h;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  data = LD_UB(src);
+  sum_h = __msa_hadd_u_h(data, data);
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 4);
+  out = (v16u8)__msa_splati_b((v16i8)sum_w, 0);
+
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+  dst += (8 * dst_stride);
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+}
+
+static void intra_predict_128dc_16x16_msa(uint8_t *dst, int32_t dst_stride) {
+  const v16u8 out = (v16u8)__msa_ldi_b(128);
+
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+  dst += (8 * dst_stride);
+  ST_UB8(out, out, out, out, out, out, out, out, dst, dst_stride);
+}
+
+static void intra_predict_dc_32x32_msa(const uint8_t *src_top,
+                                       const uint8_t *src_left, uint8_t *dst,
+                                       int32_t dst_stride) {
+  uint32_t row;
+  v16u8 top0, top1, left0, left1, out;
+  v8u16 sum_h, sum_top0, sum_top1, sum_left0, sum_left1;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  LD_UB2(src_top, 16, top0, top1);
+  LD_UB2(src_left, 16, left0, left1);
+  HADD_UB2_UH(top0, top1, sum_top0, sum_top1);
+  HADD_UB2_UH(left0, left1, sum_left0, sum_left1);
+  sum_h = sum_top0 + sum_top1;
+  sum_h += sum_left0 + sum_left1;
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 6);
+  out = (v16u8)__msa_splati_b((v16i8)sum_w, 0);
+
+  for (row = 16; row--;) {
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_dc_tl_32x32_msa(const uint8_t *src, uint8_t *dst,
+                                          int32_t dst_stride) {
+  uint32_t row;
+  v16u8 data0, data1, out;
+  v8u16 sum_h, sum_data0, sum_data1;
+  v4u32 sum_w;
+  v2u64 sum_d;
+
+  LD_UB2(src, 16, data0, data1);
+  HADD_UB2_UH(data0, data1, sum_data0, sum_data1);
+  sum_h = sum_data0 + sum_data1;
+  sum_w = __msa_hadd_u_w(sum_h, sum_h);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_pckev_w((v4i32)sum_d, (v4i32)sum_d);
+  sum_d = __msa_hadd_u_d(sum_w, sum_w);
+  sum_w = (v4u32)__msa_srari_w((v4i32)sum_d, 5);
+  out = (v16u8)__msa_splati_b((v16i8)sum_w, 0);
+
+  for (row = 16; row--;) {
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_128dc_32x32_msa(uint8_t *dst, int32_t dst_stride) {
+  uint32_t row;
+  const v16u8 out = (v16u8)__msa_ldi_b(128);
+
+  for (row = 16; row--;) {
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+    ST_UB2(out, out, dst, 16);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_tm_4x4_msa(const uint8_t *src_top_ptr,
+                                     const uint8_t *src_left, uint8_t *dst,
+                                     int32_t dst_stride) {
+  uint32_t val;
+  uint8_t top_left = src_top_ptr[-1];
+  v16i8 src_left0, src_left1, src_left2, src_left3, tmp0, tmp1, src_top = { 0 };
+  v16u8 src0, src1, src2, src3;
+  v8u16 src_top_left, vec0, vec1, vec2, vec3;
+
+  src_top_left = (v8u16)__msa_fill_h(top_left);
+  val = LW(src_top_ptr);
+  src_top = (v16i8)__msa_insert_w((v4i32)src_top, 0, val);
+
+  src_left0 = __msa_fill_b(src_left[0]);
+  src_left1 = __msa_fill_b(src_left[1]);
+  src_left2 = __msa_fill_b(src_left[2]);
+  src_left3 = __msa_fill_b(src_left[3]);
+
+  ILVR_B4_UB(src_left0, src_top, src_left1, src_top, src_left2, src_top,
+             src_left3, src_top, src0, src1, src2, src3);
+  HADD_UB4_UH(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+  IPRED_SUBS_UH2_UH(src_top_left, src_top_left, vec0, vec1);
+  IPRED_SUBS_UH2_UH(src_top_left, src_top_left, vec2, vec3);
+  SAT_UH4_UH(vec0, vec1, vec2, vec3, 7);
+  PCKEV_B2_SB(vec1, vec0, vec3, vec2, tmp0, tmp1);
+  ST4x4_UB(tmp0, tmp1, 0, 2, 0, 2, dst, dst_stride);
+}
+
+static void intra_predict_tm_8x8_msa(const uint8_t *src_top_ptr,
+                                     const uint8_t *src_left, uint8_t *dst,
+                                     int32_t dst_stride) {
+  uint64_t val;
+  uint8_t top_left = src_top_ptr[-1];
+  uint32_t loop_cnt;
+  v16i8 src_left0, src_left1, src_left2, src_left3, tmp0, tmp1, src_top = { 0 };
+  v8u16 src_top_left, vec0, vec1, vec2, vec3;
+  v16u8 src0, src1, src2, src3;
+
+  val = LD(src_top_ptr);
+  src_top = (v16i8)__msa_insert_d((v2i64)src_top, 0, val);
+  src_top_left = (v8u16)__msa_fill_h(top_left);
+
+  for (loop_cnt = 2; loop_cnt--;) {
+    src_left0 = __msa_fill_b(src_left[0]);
+    src_left1 = __msa_fill_b(src_left[1]);
+    src_left2 = __msa_fill_b(src_left[2]);
+    src_left3 = __msa_fill_b(src_left[3]);
+    src_left += 4;
+
+    ILVR_B4_UB(src_left0, src_top, src_left1, src_top, src_left2, src_top,
+               src_left3, src_top, src0, src1, src2, src3);
+    HADD_UB4_UH(src0, src1, src2, src3, vec0, vec1, vec2, vec3);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, vec0, vec1);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, vec2, vec3);
+    SAT_UH4_UH(vec0, vec1, vec2, vec3, 7);
+    PCKEV_B2_SB(vec1, vec0, vec3, vec2, tmp0, tmp1);
+    ST8x4_UB(tmp0, tmp1, dst, dst_stride);
+    dst += (4 * dst_stride);
+  }
+}
+
+static void intra_predict_tm_16x16_msa(const uint8_t *src_top_ptr,
+                                       const uint8_t *src_left, uint8_t *dst,
+                                       int32_t dst_stride) {
+  uint8_t top_left = src_top_ptr[-1];
+  uint32_t loop_cnt;
+  v16i8 src_top, src_left0, src_left1, src_left2, src_left3;
+  v8u16 src_top_left, res_r, res_l;
+
+  src_top = LD_SB(src_top_ptr);
+  src_top_left = (v8u16)__msa_fill_h(top_left);
+
+  for (loop_cnt = 4; loop_cnt--;) {
+    src_left0 = __msa_fill_b(src_left[0]);
+    src_left1 = __msa_fill_b(src_left[1]);
+    src_left2 = __msa_fill_b(src_left[2]);
+    src_left3 = __msa_fill_b(src_left[3]);
+    src_left += 4;
+
+    ILVRL_B2_UH(src_left0, src_top, res_r, res_l);
+    HADD_UB2_UH(res_r, res_l, res_r, res_l);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r, res_l);
+
+    SAT_UH2_UH(res_r, res_l, 7);
+    PCKEV_ST_SB(res_r, res_l, dst);
+    dst += dst_stride;
+
+    ILVRL_B2_UH(src_left1, src_top, res_r, res_l);
+    HADD_UB2_UH(res_r, res_l, res_r, res_l);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r, res_l);
+    SAT_UH2_UH(res_r, res_l, 7);
+    PCKEV_ST_SB(res_r, res_l, dst);
+    dst += dst_stride;
+
+    ILVRL_B2_UH(src_left2, src_top, res_r, res_l);
+    HADD_UB2_UH(res_r, res_l, res_r, res_l);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r, res_l);
+    SAT_UH2_UH(res_r, res_l, 7);
+    PCKEV_ST_SB(res_r, res_l, dst);
+    dst += dst_stride;
+
+    ILVRL_B2_UH(src_left3, src_top, res_r, res_l);
+    HADD_UB2_UH(res_r, res_l, res_r, res_l);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r, res_l);
+    SAT_UH2_UH(res_r, res_l, 7);
+    PCKEV_ST_SB(res_r, res_l, dst);
+    dst += dst_stride;
+  }
+}
+
+static void intra_predict_tm_32x32_msa(const uint8_t *src_top,
+                                       const uint8_t *src_left, uint8_t *dst,
+                                       int32_t dst_stride) {
+  uint8_t top_left = src_top[-1];
+  uint32_t loop_cnt;
+  v16i8 src_top0, src_top1, src_left0, src_left1, src_left2, src_left3;
+  v8u16 src_top_left, res_r0, res_r1, res_l0, res_l1;
+
+  LD_SB2(src_top, 16, src_top0, src_top1);
+  src_top_left = (v8u16)__msa_fill_h(top_left);
+
+  for (loop_cnt = 8; loop_cnt--;) {
+    src_left0 = __msa_fill_b(src_left[0]);
+    src_left1 = __msa_fill_b(src_left[1]);
+    src_left2 = __msa_fill_b(src_left[2]);
+    src_left3 = __msa_fill_b(src_left[3]);
+    src_left += 4;
+
+    ILVR_B2_UH(src_left0, src_top0, src_left0, src_top1, res_r0, res_r1);
+    ILVL_B2_UH(src_left0, src_top0, src_left0, src_top1, res_l0, res_l1);
+    HADD_UB4_UH(res_r0, res_l0, res_r1, res_l1, res_r0, res_l0, res_r1, res_l1);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r0, res_l0);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r1, res_l1);
+    SAT_UH4_UH(res_r0, res_l0, res_r1, res_l1, 7);
+    PCKEV_ST_SB(res_r0, res_l0, dst);
+    PCKEV_ST_SB(res_r1, res_l1, dst + 16);
+    dst += dst_stride;
+
+    ILVR_B2_UH(src_left1, src_top0, src_left1, src_top1, res_r0, res_r1);
+    ILVL_B2_UH(src_left1, src_top0, src_left1, src_top1, res_l0, res_l1);
+    HADD_UB4_UH(res_r0, res_l0, res_r1, res_l1, res_r0, res_l0, res_r1, res_l1);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r0, res_l0);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r1, res_l1);
+    SAT_UH4_UH(res_r0, res_l0, res_r1, res_l1, 7);
+    PCKEV_ST_SB(res_r0, res_l0, dst);
+    PCKEV_ST_SB(res_r1, res_l1, dst + 16);
+    dst += dst_stride;
+
+    ILVR_B2_UH(src_left2, src_top0, src_left2, src_top1, res_r0, res_r1);
+    ILVL_B2_UH(src_left2, src_top0, src_left2, src_top1, res_l0, res_l1);
+    HADD_UB4_UH(res_r0, res_l0, res_r1, res_l1, res_r0, res_l0, res_r1, res_l1);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r0, res_l0);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r1, res_l1);
+    SAT_UH4_UH(res_r0, res_l0, res_r1, res_l1, 7);
+    PCKEV_ST_SB(res_r0, res_l0, dst);
+    PCKEV_ST_SB(res_r1, res_l1, dst + 16);
+    dst += dst_stride;
+
+    ILVR_B2_UH(src_left3, src_top0, src_left3, src_top1, res_r0, res_r1);
+    ILVL_B2_UH(src_left3, src_top0, src_left3, src_top1, res_l0, res_l1);
+    HADD_UB4_UH(res_r0, res_l0, res_r1, res_l1, res_r0, res_l0, res_r1, res_l1);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r0, res_l0);
+    IPRED_SUBS_UH2_UH(src_top_left, src_top_left, res_r1, res_l1);
+    SAT_UH4_UH(res_r0, res_l0, res_r1, res_l1, 7);
+    PCKEV_ST_SB(res_r0, res_l0, dst);
+    PCKEV_ST_SB(res_r1, res_l1, dst + 16);
+    dst += dst_stride;
+  }
+}
+
+void aom_v_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                             const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_vert_4x4_msa(above, dst, y_stride);
+}
+
+void aom_v_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                             const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_vert_8x8_msa(above, dst, y_stride);
+}
+
+void aom_v_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_vert_16x16_msa(above, dst, y_stride);
+}
+
+void aom_v_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_vert_32x32_msa(above, dst, y_stride);
+}
+
+void aom_h_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                             const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_horiz_4x4_msa(left, dst, y_stride);
+}
+
+void aom_h_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                             const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_horiz_8x8_msa(left, dst, y_stride);
+}
+
+void aom_h_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_horiz_16x16_msa(left, dst, y_stride);
+}
+
+void aom_h_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_horiz_32x32_msa(left, dst, y_stride);
+}
+
+void aom_dc_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                              const uint8_t *above, const uint8_t *left) {
+  intra_predict_dc_4x4_msa(above, left, dst, y_stride);
+}
+
+void aom_dc_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                              const uint8_t *above, const uint8_t *left) {
+  intra_predict_dc_8x8_msa(above, left, dst, y_stride);
+}
+
+void aom_dc_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                const uint8_t *above, const uint8_t *left) {
+  intra_predict_dc_16x16_msa(above, left, dst, y_stride);
+}
+
+void aom_dc_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                const uint8_t *above, const uint8_t *left) {
+  intra_predict_dc_32x32_msa(above, left, dst, y_stride);
+}
+
+void aom_dc_top_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_dc_tl_4x4_msa(above, dst, y_stride);
+}
+
+void aom_dc_top_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_dc_tl_8x8_msa(above, dst, y_stride);
+}
+
+void aom_dc_top_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_dc_tl_16x16_msa(above, dst, y_stride);
+}
+
+void aom_dc_top_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+
+  intra_predict_dc_tl_32x32_msa(above, dst, y_stride);
+}
+
+void aom_dc_left_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_dc_tl_4x4_msa(left, dst, y_stride);
+}
+
+void aom_dc_left_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+
+  intra_predict_dc_tl_8x8_msa(left, dst, y_stride);
+}
+
+void aom_dc_left_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+
+  intra_predict_dc_tl_16x16_msa(left, dst, y_stride);
+}
+
+void aom_dc_left_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+
+  intra_predict_dc_tl_32x32_msa(left, dst, y_stride);
+}
+
+void aom_dc_128_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+
+  intra_predict_128dc_4x4_msa(dst, y_stride);
+}
+
+void aom_dc_128_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+
+  intra_predict_128dc_8x8_msa(dst, y_stride);
+}
+
+void aom_dc_128_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+
+  intra_predict_128dc_16x16_msa(dst, y_stride);
+}
+
+void aom_dc_128_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+
+  intra_predict_128dc_32x32_msa(dst, y_stride);
+}
+
+void aom_tm_predictor_4x4_msa(uint8_t *dst, ptrdiff_t y_stride,
+                              const uint8_t *above, const uint8_t *left) {
+  intra_predict_tm_4x4_msa(above, left, dst, y_stride);
+}
+
+void aom_tm_predictor_8x8_msa(uint8_t *dst, ptrdiff_t y_stride,
+                              const uint8_t *above, const uint8_t *left) {
+  intra_predict_tm_8x8_msa(above, left, dst, y_stride);
+}
+
+void aom_tm_predictor_16x16_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                const uint8_t *above, const uint8_t *left) {
+  intra_predict_tm_16x16_msa(above, left, dst, y_stride);
+}
+
+void aom_tm_predictor_32x32_msa(uint8_t *dst, ptrdiff_t y_stride,
+                                const uint8_t *above, const uint8_t *left) {
+  intra_predict_tm_32x32_msa(above, left, dst, y_stride);
+}
diff --git a/third_party/aom/aom_dsp/mips/inv_txfm_dspr2.h b/third_party/aom/aom_dsp/mips/inv_txfm_dspr2.h
new file mode 100644
index 0000000000..8a85e26f3e
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/inv_txfm_dspr2.h
@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_MIPS_INV_TXFM_DSPR2_H_
+#define AOM_DSP_MIPS_INV_TXFM_DSPR2_H_
+
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/inv_txfm.h"
+#include "aom_dsp/mips/common_dspr2.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if HAVE_DSPR2
+#define DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input)                         \
+  ({                                                                           \
+                                                                               \
+    int32_t tmp, out;                                                          \
+    int dct_cost_rounding = DCT_CONST_ROUNDING;                                \
+    int in = input;                                                            \
+                                                                               \
+    __asm__ __volatile__(/* out = dct_const_round_shift(dc *  cospi_16_64); */ \
+                         "mtlo     %[dct_cost_rounding],   $ac1              " \
+                         "                \n\t"                                \
+                         "mthi     $zero,                  $ac1              " \
+                         "                \n\t"                                \
+                         "madd     $ac1,                   %[in],            " \
+                         "%[cospi_16_64]  \n\t"                                \
+                         "extp     %[tmp],                 $ac1,             " \
+                         "31              \n\t"                                \
+                                                                               \
+                         /* out = dct_const_round_shift(out * cospi_16_64); */ \
+                         "mtlo     %[dct_cost_rounding],   $ac2              " \
+                         "                \n\t"                                \
+                         "mthi     $zero,                  $ac2              " \
+                         "                \n\t"                                \
+                         "madd     $ac2,                   %[tmp],           " \
+                         "%[cospi_16_64]  \n\t"                                \
+                         "extp     %[out],                 $ac2,             " \
+                         "31              \n\t"                                \
+                                                                               \
+                         : [tmp] "=&r"(tmp), [out] "=r"(out)                   \
+                         : [in] "r"(in),                                       \
+                           [dct_cost_rounding] "r"(dct_cost_rounding),         \
+                           [cospi_16_64] "r"(cospi_16_64));                    \
+    out;                                                                       \
+  })
+
+void aom_idct32_cols_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                   int dest_stride);
+void aom_idct4_rows_dspr2(const int16_t *input, int16_t *output);
+void aom_idct4_columns_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                     int dest_stride);
+void iadst4_dspr2(const int16_t *input, int16_t *output);
+void idct8_rows_dspr2(const int16_t *input, int16_t *output, uint32_t no_rows);
+void idct8_columns_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                 int dest_stride);
+void iadst8_dspr2(const int16_t *input, int16_t *output);
+void idct16_rows_dspr2(const int16_t *input, int16_t *output, uint32_t no_rows);
+void idct16_cols_add_blk_dspr2(int16_t *input, uint8_t *dest, int dest_stride);
+void iadst16_dspr2(const int16_t *input, int16_t *output);
+
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_MIPS_INV_TXFM_DSPR2_H_
diff --git a/third_party/aom/aom_dsp/mips/inv_txfm_msa.h b/third_party/aom/aom_dsp/mips/inv_txfm_msa.h
new file mode 100644
index 0000000000..122667aa85
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/inv_txfm_msa.h
@@ -0,0 +1,412 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_MIPS_INV_TXFM_MSA_H_
+#define AOM_DSP_MIPS_INV_TXFM_MSA_H_
+
+#include "aom_dsp/mips/macros_msa.h"
+#include "aom_dsp/mips/txfm_macros_msa.h"
+#include "aom_dsp/txfm_common.h"
+
+#define AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2,  \
+                  out3, out4, out5, out6, out7)                              \
+  {                                                                          \
+    v8i16 cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst4_m;                       \
+    v8i16 vec0_m, vec1_m, vec2_m, vec3_m, s0_m, s1_m;                        \
+    v8i16 coeff0_m = { cospi_2_64,  cospi_6_64,  cospi_10_64, cospi_14_64,   \
+                       cospi_18_64, cospi_22_64, cospi_26_64, cospi_30_64 }; \
+    v8i16 coeff1_m = { cospi_8_64,  -cospi_8_64,  cospi_16_64, -cospi_16_64, \
+                       cospi_24_64, -cospi_24_64, 0,           0 };          \
+                                                                             \
+    SPLATI_H2_SH(coeff0_m, 0, 7, cnst0_m, cnst1_m);                          \
+    cnst2_m = -cnst0_m;                                                      \
+    ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);       \
+    SPLATI_H2_SH(coeff0_m, 4, 3, cnst2_m, cnst3_m);                          \
+    cnst4_m = -cnst2_m;                                                      \
+    ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);       \
+                                                                             \
+    ILVRL_H2_SH(in0, in7, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(in4, in3, vec3_m, vec2_m);                                   \
+    DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m, cnst1_m,  \
+                          cnst2_m, cnst3_m, in7, in0, in4, in3);             \
+                                                                             \
+    SPLATI_H2_SH(coeff0_m, 2, 5, cnst0_m, cnst1_m);                          \
+    cnst2_m = -cnst0_m;                                                      \
+    ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);       \
+    SPLATI_H2_SH(coeff0_m, 6, 1, cnst2_m, cnst3_m);                          \
+    cnst4_m = -cnst2_m;                                                      \
+    ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);       \
+                                                                             \
+    ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                   \
+                                                                             \
+    DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m, cnst1_m,  \
+                          cnst2_m, cnst3_m, in5, in2, in6, in1);             \
+    BUTTERFLY_4(in7, in0, in2, in5, s1_m, s0_m, in2, in5);                   \
+    out7 = -s0_m;                                                            \
+    out0 = s1_m;                                                             \
+                                                                             \
+    SPLATI_H4_SH(coeff1_m, 0, 4, 1, 5, cnst0_m, cnst1_m, cnst2_m, cnst3_m);  \
+                                                                             \
+    ILVEV_H2_SH(cnst3_m, cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst2_m);       \
+    cnst0_m = __msa_ilvev_h(cnst1_m, cnst0_m);                               \
+    cnst1_m = cnst0_m;                                                       \
+                                                                             \
+    ILVRL_H2_SH(in4, in3, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                   \
+    DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m, cnst2_m,  \
+                          cnst3_m, cnst1_m, out1, out6, s0_m, s1_m);         \
+                                                                             \
+    SPLATI_H2_SH(coeff1_m, 2, 3, cnst0_m, cnst1_m);                          \
+    cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                               \
+                                                                             \
+    ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(s0_m, s1_m, vec3_m, vec2_m);                                 \
+    out3 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                   \
+    out4 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst1_m);                   \
+    out2 = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst0_m);                   \
+    out5 = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst1_m);                   \
+                                                                             \
+    out1 = -out1;                                                            \
+    out3 = -out3;                                                            \
+    out5 = -out5;                                                            \
+  }
+
+#define AOM_SET_COSPI_PAIR(c0_h, c1_h)  \
+  ({                                    \
+    v8i16 out0_m, r0_m, r1_m;           \
+                                        \
+    r0_m = __msa_fill_h(c0_h);          \
+    r1_m = __msa_fill_h(c1_h);          \
+    out0_m = __msa_ilvev_h(r1_m, r0_m); \
+                                        \
+    out0_m;                             \
+  })
+
+#define AOM_ADDBLK_ST8x4_UB(dst, dst_stride, in0, in1, in2, in3)               \
+  {                                                                            \
+    uint8_t *dst_m = (uint8_t *)(dst);                                         \
+    v16u8 dst0_m, dst1_m, dst2_m, dst3_m;                                      \
+    v16i8 tmp0_m, tmp1_m;                                                      \
+    v16i8 zero_m = { 0 };                                                      \
+    v8i16 res0_m, res1_m, res2_m, res3_m;                                      \
+                                                                               \
+    LD_UB4(dst_m, dst_stride, dst0_m, dst1_m, dst2_m, dst3_m);                 \
+    ILVR_B4_SH(zero_m, dst0_m, zero_m, dst1_m, zero_m, dst2_m, zero_m, dst3_m, \
+               res0_m, res1_m, res2_m, res3_m);                                \
+    ADD4(res0_m, in0, res1_m, in1, res2_m, in2, res3_m, in3, res0_m, res1_m,   \
+         res2_m, res3_m);                                                      \
+    CLIP_SH4_0_255(res0_m, res1_m, res2_m, res3_m);                            \
+    PCKEV_B2_SB(res1_m, res0_m, res3_m, res2_m, tmp0_m, tmp1_m);               \
+    ST8x4_UB(tmp0_m, tmp1_m, dst_m, dst_stride);                               \
+  }
+
+#define AOM_IDCT4x4(in0, in1, in2, in3, out0, out1, out2, out3)             \
+  {                                                                         \
+    v8i16 c0_m, c1_m, c2_m, c3_m;                                           \
+    v8i16 step0_m, step1_m;                                                 \
+    v4i32 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                   \
+                                                                            \
+    c0_m = AOM_SET_COSPI_PAIR(cospi_16_64, cospi_16_64);                    \
+    c1_m = AOM_SET_COSPI_PAIR(cospi_16_64, -cospi_16_64);                   \
+    step0_m = __msa_ilvr_h(in2, in0);                                       \
+    DOTP_SH2_SW(step0_m, step0_m, c0_m, c1_m, tmp0_m, tmp1_m);              \
+                                                                            \
+    c2_m = AOM_SET_COSPI_PAIR(cospi_24_64, -cospi_8_64);                    \
+    c3_m = AOM_SET_COSPI_PAIR(cospi_8_64, cospi_24_64);                     \
+    step1_m = __msa_ilvr_h(in3, in1);                                       \
+    DOTP_SH2_SW(step1_m, step1_m, c2_m, c3_m, tmp2_m, tmp3_m);              \
+    SRARI_W4_SW(tmp0_m, tmp1_m, tmp2_m, tmp3_m, DCT_CONST_BITS);            \
+                                                                            \
+    PCKEV_H2_SW(tmp1_m, tmp0_m, tmp3_m, tmp2_m, tmp0_m, tmp2_m);            \
+    SLDI_B2_0_SW(tmp0_m, tmp2_m, tmp1_m, tmp3_m, 8);                        \
+    BUTTERFLY_4((v8i16)tmp0_m, (v8i16)tmp1_m, (v8i16)tmp2_m, (v8i16)tmp3_m, \
+                out0, out1, out2, out3);                                    \
+  }
+
+#define AOM_IADST4x4(in0, in1, in2, in3, out0, out1, out2, out3)       \
+  {                                                                    \
+    v8i16 res0_m, res1_m, c0_m, c1_m;                                  \
+    v8i16 k1_m, k2_m, k3_m, k4_m;                                      \
+    v8i16 zero_m = { 0 };                                              \
+    v4i32 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                              \
+    v4i32 int0_m, int1_m, int2_m, int3_m;                              \
+    v8i16 mask_m = { sinpi_1_9,  sinpi_2_9,  sinpi_3_9,  sinpi_4_9,    \
+                     -sinpi_1_9, -sinpi_2_9, -sinpi_3_9, -sinpi_4_9 }; \
+                                                                       \
+    SPLATI_H4_SH(mask_m, 3, 0, 1, 2, c0_m, c1_m, k1_m, k2_m);          \
+    ILVEV_H2_SH(c0_m, c1_m, k1_m, k2_m, c0_m, c1_m);                   \
+    ILVR_H2_SH(in0, in2, in1, in3, res0_m, res1_m);                    \
+    DOTP_SH2_SW(res0_m, res1_m, c0_m, c1_m, tmp2_m, tmp1_m);           \
+    int0_m = tmp2_m + tmp1_m;                                          \
+                                                                       \
+    SPLATI_H2_SH(mask_m, 4, 7, k4_m, k3_m);                            \
+    ILVEV_H2_SH(k4_m, k1_m, k3_m, k2_m, c0_m, c1_m);                   \
+    DOTP_SH2_SW(res0_m, res1_m, c0_m, c1_m, tmp0_m, tmp1_m);           \
+    int1_m = tmp0_m + tmp1_m;                                          \
+                                                                       \
+    c0_m = __msa_splati_h(mask_m, 6);                                  \
+    ILVL_H2_SH(k2_m, c0_m, zero_m, k2_m, c0_m, c1_m);                  \
+    ILVR_H2_SH(in0, in2, in1, in3, res0_m, res1_m);                    \
+    DOTP_SH2_SW(res0_m, res1_m, c0_m, c1_m, tmp0_m, tmp1_m);           \
+    int2_m = tmp0_m + tmp1_m;                                          \
+                                                                       \
+    c0_m = __msa_splati_h(mask_m, 6);                                  \
+    c0_m = __msa_ilvev_h(c0_m, k1_m);                                  \
+                                                                       \
+    res0_m = __msa_ilvr_h((in1), (in3));                               \
+    tmp0_m = __msa_dotp_s_w(res0_m, c0_m);                             \
+    int3_m = tmp2_m + tmp0_m;                                          \
+                                                                       \
+    res0_m = __msa_ilvr_h((in2), (in3));                               \
+    c1_m = __msa_ilvev_h(k4_m, k3_m);                                  \
+                                                                       \
+    tmp2_m = __msa_dotp_s_w(res0_m, c1_m);                             \
+    res1_m = __msa_ilvr_h((in0), (in2));                               \
+    c1_m = __msa_ilvev_h(k1_m, zero_m);                                \
+                                                                       \
+    tmp3_m = __msa_dotp_s_w(res1_m, c1_m);                             \
+    int3_m += tmp2_m;                                                  \
+    int3_m += tmp3_m;                                                  \
+                                                                       \
+    SRARI_W4_SW(int0_m, int1_m, int2_m, int3_m, DCT_CONST_BITS);       \
+    PCKEV_H2_SH(int0_m, int0_m, int1_m, int1_m, out0, out1);           \
+    PCKEV_H2_SH(int2_m, int2_m, int3_m, int3_m, out2, out3);           \
+  }
+
+#define AV1_SET_CONST_PAIR(mask_h, idx1_h, idx2_h)    \
+  ({                                                  \
+    v8i16 c0_m, c1_m;                                 \
+                                                      \
+    SPLATI_H2_SH(mask_h, idx1_h, idx2_h, c0_m, c1_m); \
+    c0_m = __msa_ilvev_h(c1_m, c0_m);                 \
+                                                      \
+    c0_m;                                             \
+  })
+
+/* multiply and add macro */
+#define AV1_MADD(inp0, inp1, inp2, inp3, cst0, cst1, cst2, cst3, out0, out1,  \
+                 out2, out3)                                                  \
+  {                                                                           \
+    v8i16 madd_s0_m, madd_s1_m, madd_s2_m, madd_s3_m;                         \
+    v4i32 tmp0_madd, tmp1_madd, tmp2_madd, tmp3_madd;                         \
+                                                                              \
+    ILVRL_H2_SH(inp1, inp0, madd_s1_m, madd_s0_m);                            \
+    ILVRL_H2_SH(inp3, inp2, madd_s3_m, madd_s2_m);                            \
+    DOTP_SH4_SW(madd_s1_m, madd_s0_m, madd_s1_m, madd_s0_m, cst0, cst0, cst1, \
+                cst1, tmp0_madd, tmp1_madd, tmp2_madd, tmp3_madd);            \
+    SRARI_W4_SW(tmp0_madd, tmp1_madd, tmp2_madd, tmp3_madd, DCT_CONST_BITS);  \
+    PCKEV_H2_SH(tmp1_madd, tmp0_madd, tmp3_madd, tmp2_madd, out0, out1);      \
+    DOTP_SH4_SW(madd_s3_m, madd_s2_m, madd_s3_m, madd_s2_m, cst2, cst2, cst3, \
+                cst3, tmp0_madd, tmp1_madd, tmp2_madd, tmp3_madd);            \
+    SRARI_W4_SW(tmp0_madd, tmp1_madd, tmp2_madd, tmp3_madd, DCT_CONST_BITS);  \
+    PCKEV_H2_SH(tmp1_madd, tmp0_madd, tmp3_madd, tmp2_madd, out2, out3);      \
+  }
+
+/* idct 8x8 macro */
+#define AOM_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1,    \
+                       out2, out3, out4, out5, out6, out7)                    \
+  {                                                                           \
+    v8i16 tp0_m, tp1_m, tp2_m, tp3_m, tp4_m, tp5_m, tp6_m, tp7_m;             \
+    v8i16 k0_m, k1_m, k2_m, k3_m, res0_m, res1_m, res2_m, res3_m;             \
+    v4i32 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                     \
+    v8i16 mask_m = { cospi_28_64, cospi_4_64,  cospi_20_64,  cospi_12_64,     \
+                     cospi_16_64, -cospi_4_64, -cospi_20_64, -cospi_16_64 };  \
+                                                                              \
+    k0_m = AV1_SET_CONST_PAIR(mask_m, 0, 5);                                  \
+    k1_m = AV1_SET_CONST_PAIR(mask_m, 1, 0);                                  \
+    k2_m = AV1_SET_CONST_PAIR(mask_m, 6, 3);                                  \
+    k3_m = AV1_SET_CONST_PAIR(mask_m, 3, 2);                                  \
+    AV1_MADD(in1, in7, in3, in5, k0_m, k1_m, k2_m, k3_m, in1, in7, in3, in5); \
+    SUB2(in1, in3, in7, in5, res0_m, res1_m);                                 \
+    k0_m = AV1_SET_CONST_PAIR(mask_m, 4, 7);                                  \
+    k1_m = __msa_splati_h(mask_m, 4);                                         \
+                                                                              \
+    ILVRL_H2_SH(res0_m, res1_m, res2_m, res3_m);                              \
+    DOTP_SH4_SW(res2_m, res3_m, res2_m, res3_m, k0_m, k0_m, k1_m, k1_m,       \
+                tmp0_m, tmp1_m, tmp2_m, tmp3_m);                              \
+    SRARI_W4_SW(tmp0_m, tmp1_m, tmp2_m, tmp3_m, DCT_CONST_BITS);              \
+    tp4_m = in1 + in3;                                                        \
+    PCKEV_H2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, tp5_m, tp6_m);                \
+    tp7_m = in7 + in5;                                                        \
+    k2_m = AOM_SET_COSPI_PAIR(cospi_24_64, -cospi_8_64);                      \
+    k3_m = AOM_SET_COSPI_PAIR(cospi_8_64, cospi_24_64);                       \
+    AV1_MADD(in0, in4, in2, in6, k1_m, k0_m, k2_m, k3_m, in0, in4, in2, in6); \
+    BUTTERFLY_4(in0, in4, in2, in6, tp0_m, tp1_m, tp2_m, tp3_m);              \
+    BUTTERFLY_8(tp0_m, tp1_m, tp2_m, tp3_m, tp4_m, tp5_m, tp6_m, tp7_m, out0, \
+                out1, out2, out3, out4, out5, out6, out7);                    \
+  }
+
+#define AV1_IADST8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1,   \
+                        out2, out3, out4, out5, out6, out7)                   \
+  {                                                                           \
+    v4i32 r0_m, r1_m, r2_m, r3_m, r4_m, r5_m, r6_m, r7_m;                     \
+    v4i32 m0_m, m1_m, m2_m, m3_m, t0_m, t1_m;                                 \
+    v8i16 res0_m, res1_m, res2_m, res3_m, k0_m, k1_m, in_s0, in_s1;           \
+    v8i16 mask1_m = { cospi_2_64,  cospi_30_64,  -cospi_2_64, cospi_10_64,    \
+                      cospi_22_64, -cospi_10_64, cospi_18_64, cospi_14_64 };  \
+    v8i16 mask2_m = { cospi_14_64,  -cospi_18_64, cospi_26_64, cospi_6_64,    \
+                      -cospi_26_64, cospi_8_64,   cospi_24_64, -cospi_8_64 }; \
+    v8i16 mask3_m = {                                                         \
+      -cospi_24_64, cospi_8_64, cospi_16_64, -cospi_16_64, 0, 0, 0, 0         \
+    };                                                                        \
+                                                                              \
+    k0_m = AV1_SET_CONST_PAIR(mask1_m, 0, 1);                                 \
+    k1_m = AV1_SET_CONST_PAIR(mask1_m, 1, 2);                                 \
+    ILVRL_H2_SH(in1, in0, in_s1, in_s0);                                      \
+    DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m, r0_m,     \
+                r1_m, r2_m, r3_m);                                            \
+    k0_m = AV1_SET_CONST_PAIR(mask1_m, 6, 7);                                 \
+    k1_m = AV1_SET_CONST_PAIR(mask2_m, 0, 1);                                 \
+    ILVRL_H2_SH(in5, in4, in_s1, in_s0);                                      \
+    DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m, r4_m,     \
+                r5_m, r6_m, r7_m);                                            \
+    ADD4(r0_m, r4_m, r1_m, r5_m, r2_m, r6_m, r3_m, r7_m, m0_m, m1_m, m2_m,    \
+         m3_m);                                                               \
+    SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                      \
+    PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, res0_m, res1_m);                      \
+    SUB4(r0_m, r4_m, r1_m, r5_m, r2_m, r6_m, r3_m, r7_m, m0_m, m1_m, m2_m,    \
+         m3_m);                                                               \
+    SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                      \
+    PCKEV_H2_SW(m1_m, m0_m, m3_m, m2_m, t0_m, t1_m);                          \
+    k0_m = AV1_SET_CONST_PAIR(mask1_m, 3, 4);                                 \
+    k1_m = AV1_SET_CONST_PAIR(mask1_m, 4, 5);                                 \
+    ILVRL_H2_SH(in3, in2, in_s1, in_s0);                                      \
+    DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m, r0_m,     \
+                r1_m, r2_m, r3_m);                                            \
+    k0_m = AV1_SET_CONST_PAIR(mask2_m, 2, 3);                                 \
+    k1_m = AV1_SET_CONST_PAIR(mask2_m, 3, 4);                                 \
+    ILVRL_H2_SH(in7, in6, in_s1, in_s0);                                      \
+    DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m, r4_m,     \
+                r5_m, r6_m, r7_m);                                            \
+    ADD4(r0_m, r4_m, r1_m, r5_m, r2_m, r6_m, r3_m, r7_m, m0_m, m1_m, m2_m,    \
+         m3_m);                                                               \
+    SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                      \
+    PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, res2_m, res3_m);                      \
+    SUB4(r0_m, r4_m, r1_m, r5_m, r2_m, r6_m, r3_m, r7_m, m0_m, m1_m, m2_m,    \
+         m3_m);                                                               \
+    SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                      \
+    PCKEV_H2_SW(m1_m, m0_m, m3_m, m2_m, r2_m, r3_m);                          \
+    ILVRL_H2_SW(r3_m, r2_m, m2_m, m3_m);                                      \
+    BUTTERFLY_4(res0_m, res1_m, res3_m, res2_m, out0, in7, in4, in3);         \
+    k0_m = AV1_SET_CONST_PAIR(mask2_m, 5, 6);                                 \
+    k1_m = AV1_SET_CONST_PAIR(mask2_m, 6, 7);                                 \
+    ILVRL_H2_SH(t1_m, t0_m, in_s1, in_s0);                                    \
+    DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m, r0_m,     \
+                r1_m, r2_m, r3_m);                                            \
+    k1_m = AV1_SET_CONST_PAIR(mask3_m, 0, 1);                                 \
+    DOTP_SH4_SW(m2_m, m3_m, m2_m, m3_m, k0_m, k0_m, k1_m, k1_m, r4_m, r5_m,   \
+                r6_m, r7_m);                                                  \
+    ADD4(r0_m, r6_m, r1_m, r7_m, r2_m, r4_m, r3_m, r5_m, m0_m, m1_m, m2_m,    \
+         m3_m);                                                               \
+    SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                      \
+    PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, in1, out6);                           \
+    SUB4(r0_m, r6_m, r1_m, r7_m, r2_m, r4_m, r3_m, r5_m, m0_m, m1_m, m2_m,    \
+         m3_m);                                                               \
+    SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                      \
+    PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, in2, in5);                            \
+    k0_m = AV1_SET_CONST_PAIR(mask3_m, 2, 2);                                 \
+    k1_m = AV1_SET_CONST_PAIR(mask3_m, 2, 3);                                 \
+    ILVRL_H2_SH(in4, in3, in_s1, in_s0);                                      \
+    DOTP_SH4_SW(in_s1, in_s0, in_s1, in_s0, k0_m, k0_m, k1_m, k1_m, m0_m,     \
+                m1_m, m2_m, m3_m);                                            \
+    SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                      \
+    PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, in3, out4);                           \
+    ILVRL_H2_SW(in5, in2, m2_m, m3_m);                                        \
+    DOTP_SH4_SW(m2_m, m3_m, m2_m, m3_m, k0_m, k0_m, k1_m, k1_m, m0_m, m1_m,   \
+                m2_m, m3_m);                                                  \
+    SRARI_W4_SW(m0_m, m1_m, m2_m, m3_m, DCT_CONST_BITS);                      \
+    PCKEV_H2_SH(m1_m, m0_m, m3_m, m2_m, out2, in5);                           \
+                                                                              \
+    out1 = -in1;                                                              \
+    out3 = -in3;                                                              \
+    out5 = -in5;                                                              \
+    out7 = -in7;                                                              \
+  }
+
+#define AOM_IADST8x16_1D(r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11,     \
+                         r12, r13, r14, r15, out0, out1, out2, out3, out4,     \
+                         out5, out6, out7, out8, out9, out10, out11, out12,    \
+                         out13, out14, out15)                                  \
+  {                                                                            \
+    v8i16 g0_m, g1_m, g2_m, g3_m, g4_m, g5_m, g6_m, g7_m;                      \
+    v8i16 g8_m, g9_m, g10_m, g11_m, g12_m, g13_m, g14_m, g15_m;                \
+    v8i16 h0_m, h1_m, h2_m, h3_m, h4_m, h5_m, h6_m, h7_m;                      \
+    v8i16 h8_m, h9_m, h10_m, h11_m;                                            \
+    v8i16 k0_m, k1_m, k2_m, k3_m;                                              \
+                                                                               \
+    /* stage 1 */                                                              \
+    k0_m = AOM_SET_COSPI_PAIR(cospi_1_64, cospi_31_64);                        \
+    k1_m = AOM_SET_COSPI_PAIR(cospi_31_64, -cospi_1_64);                       \
+    k2_m = AOM_SET_COSPI_PAIR(cospi_17_64, cospi_15_64);                       \
+    k3_m = AOM_SET_COSPI_PAIR(cospi_15_64, -cospi_17_64);                      \
+    MADD_BF(r15, r0, r7, r8, k0_m, k1_m, k2_m, k3_m, g0_m, g1_m, g2_m, g3_m);  \
+    k0_m = AOM_SET_COSPI_PAIR(cospi_5_64, cospi_27_64);                        \
+    k1_m = AOM_SET_COSPI_PAIR(cospi_27_64, -cospi_5_64);                       \
+    k2_m = AOM_SET_COSPI_PAIR(cospi_21_64, cospi_11_64);                       \
+    k3_m = AOM_SET_COSPI_PAIR(cospi_11_64, -cospi_21_64);                      \
+    MADD_BF(r13, r2, r5, r10, k0_m, k1_m, k2_m, k3_m, g4_m, g5_m, g6_m, g7_m); \
+    k0_m = AOM_SET_COSPI_PAIR(cospi_9_64, cospi_23_64);                        \
+    k1_m = AOM_SET_COSPI_PAIR(cospi_23_64, -cospi_9_64);                       \
+    k2_m = AOM_SET_COSPI_PAIR(cospi_25_64, cospi_7_64);                        \
+    k3_m = AOM_SET_COSPI_PAIR(cospi_7_64, -cospi_25_64);                       \
+    MADD_BF(r11, r4, r3, r12, k0_m, k1_m, k2_m, k3_m, g8_m, g9_m, g10_m,       \
+            g11_m);                                                            \
+    k0_m = AOM_SET_COSPI_PAIR(cospi_13_64, cospi_19_64);                       \
+    k1_m = AOM_SET_COSPI_PAIR(cospi_19_64, -cospi_13_64);                      \
+    k2_m = AOM_SET_COSPI_PAIR(cospi_29_64, cospi_3_64);                        \
+    k3_m = AOM_SET_COSPI_PAIR(cospi_3_64, -cospi_29_64);                       \
+    MADD_BF(r9, r6, r1, r14, k0_m, k1_m, k2_m, k3_m, g12_m, g13_m, g14_m,      \
+            g15_m);                                                            \
+                                                                               \
+    /* stage 2 */                                                              \
+    k0_m = AOM_SET_COSPI_PAIR(cospi_4_64, cospi_28_64);                        \
+    k1_m = AOM_SET_COSPI_PAIR(cospi_28_64, -cospi_4_64);                       \
+    k2_m = AOM_SET_COSPI_PAIR(-cospi_28_64, cospi_4_64);                       \
+    MADD_BF(g1_m, g3_m, g9_m, g11_m, k0_m, k1_m, k2_m, k0_m, h0_m, h1_m, h2_m, \
+            h3_m);                                                             \
+    k0_m = AOM_SET_COSPI_PAIR(cospi_12_64, cospi_20_64);                       \
+    k1_m = AOM_SET_COSPI_PAIR(-cospi_20_64, cospi_12_64);                      \
+    k2_m = AOM_SET_COSPI_PAIR(cospi_20_64, -cospi_12_64);                      \
+    MADD_BF(g7_m, g5_m, g15_m, g13_m, k0_m, k1_m, k2_m, k0_m, h4_m, h5_m,      \
+            h6_m, h7_m);                                                       \
+    BUTTERFLY_4(h0_m, h2_m, h6_m, h4_m, out8, out9, out11, out10);             \
+    BUTTERFLY_8(g0_m, g2_m, g4_m, g6_m, g14_m, g12_m, g10_m, g8_m, h8_m, h9_m, \
+                h10_m, h11_m, h6_m, h4_m, h2_m, h0_m);                         \
+                                                                               \
+    /* stage 3 */                                                              \
+    BUTTERFLY_4(h8_m, h9_m, h11_m, h10_m, out0, out1, h11_m, h10_m);           \
+    k0_m = AOM_SET_COSPI_PAIR(cospi_8_64, cospi_24_64);                        \
+    k1_m = AOM_SET_COSPI_PAIR(cospi_24_64, -cospi_8_64);                       \
+    k2_m = AOM_SET_COSPI_PAIR(-cospi_24_64, cospi_8_64);                       \
+    MADD_BF(h0_m, h2_m, h4_m, h6_m, k0_m, k1_m, k2_m, k0_m, out4, out6, out5,  \
+            out7);                                                             \
+    MADD_BF(h1_m, h3_m, h5_m, h7_m, k0_m, k1_m, k2_m, k0_m, out12, out14,      \
+            out13, out15);                                                     \
+                                                                               \
+    /* stage 4 */                                                              \
+    k0_m = AOM_SET_COSPI_PAIR(cospi_16_64, cospi_16_64);                       \
+    k1_m = AOM_SET_COSPI_PAIR(-cospi_16_64, -cospi_16_64);                     \
+    k2_m = AOM_SET_COSPI_PAIR(cospi_16_64, -cospi_16_64);                      \
+    k3_m = AOM_SET_COSPI_PAIR(-cospi_16_64, cospi_16_64);                      \
+    MADD_SHORT(h10_m, h11_m, k1_m, k2_m, out2, out3);                          \
+    MADD_SHORT(out6, out7, k0_m, k3_m, out6, out7);                            \
+    MADD_SHORT(out10, out11, k0_m, k3_m, out10, out11);                        \
+    MADD_SHORT(out14, out15, k1_m, k2_m, out14, out15);                        \
+  }
+
+void aom_idct16_1d_columns_addblk_msa(int16_t *input, uint8_t *dst,
+                                      int32_t dst_stride);
+void aom_idct16_1d_rows_msa(const int16_t *input, int16_t *output);
+void aom_iadst16_1d_columns_addblk_msa(int16_t *input, uint8_t *dst,
+                                       int32_t dst_stride);
+void aom_iadst16_1d_rows_msa(const int16_t *input, int16_t *output);
+#endif  // AOM_DSP_MIPS_INV_TXFM_MSA_H_
diff --git a/third_party/aom/aom_dsp/mips/itrans16_dspr2.c b/third_party/aom/aom_dsp/mips/itrans16_dspr2.c
new file mode 100644
index 0000000000..c63b1e8570
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/itrans16_dspr2.c
@@ -0,0 +1,1190 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/inv_txfm_dspr2.h"
+#include "aom_dsp/txfm_common.h"
+
+#if HAVE_DSPR2
+void idct16_rows_dspr2(const int16_t *input, int16_t *output,
+                       uint32_t no_rows) {
+  int i;
+  int step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6, step1_7;
+  int step1_10, step1_11, step1_12, step1_13;
+  int step2_0, step2_1, step2_2, step2_3;
+  int step2_8, step2_9, step2_10, step2_11;
+  int step2_12, step2_13, step2_14, step2_15;
+  int load1, load2, load3, load4, load5, load6, load7, load8;
+  int result1, result2, result3, result4;
+  const int const_2_power_13 = 8192;
+
+  for (i = no_rows; i--;) {
+    /* prefetch row */
+    prefetch_load((const uint8_t *)(input + 16));
+
+    __asm__ __volatile__(
+        "lh       %[load1],              0(%[input])                    \n\t"
+        "lh       %[load2],             16(%[input])                    \n\t"
+        "lh       %[load3],              8(%[input])                    \n\t"
+        "lh       %[load4],             24(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "add      %[result1],           %[load1],       %[load2]        \n\t"
+        "sub      %[result2],           %[load1],       %[load2]        \n\t"
+        "madd     $ac1,                 %[result1],     %[cospi_16_64]  \n\t"
+        "madd     $ac2,                 %[result2],     %[cospi_16_64]  \n\t"
+        "extp     %[step2_0],           $ac1,           31              \n\t"
+        "extp     %[step2_1],           $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "madd     $ac3,                 %[load3],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load4],       %[cospi_8_64]   \n\t"
+        "extp     %[step2_2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_24_64]  \n\t"
+        "extp     %[step2_3],           $ac1,           31              \n\t"
+
+        "add      %[step1_0],           %[step2_0],     %[step2_3]      \n\t"
+        "add      %[step1_1],           %[step2_1],     %[step2_2]      \n\t"
+        "sub      %[step1_2],           %[step2_1],     %[step2_2]      \n\t"
+        "sub      %[step1_3],           %[step2_0],     %[step2_3]      \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [result1] "=&r"(result1),
+          [result2] "=&r"(result2), [step2_0] "=&r"(step2_0),
+          [step2_1] "=&r"(step2_1), [step2_2] "=&r"(step2_2),
+          [step2_3] "=&r"(step2_3), [step1_0] "=r"(step1_0),
+          [step1_1] "=r"(step1_1), [step1_2] "=r"(step1_2),
+          [step1_3] "=r"(step1_3)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_24_64] "r"(cospi_24_64), [cospi_8_64] "r"(cospi_8_64),
+          [cospi_16_64] "r"(cospi_16_64));
+
+    __asm__ __volatile__(
+        "lh       %[load5],             2(%[input])                     \n\t"
+        "lh       %[load6],             30(%[input])                    \n\t"
+        "lh       %[load7],             18(%[input])                    \n\t"
+        "lh       %[load8],             14(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load5],       %[cospi_30_64]  \n\t"
+        "msub     $ac1,                 %[load6],       %[cospi_2_64]   \n\t"
+        "extp     %[result1],           $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load7],       %[cospi_14_64]  \n\t"
+        "msub     $ac3,                 %[load8],       %[cospi_18_64]  \n\t"
+        "extp     %[result2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac1,                 %[load7],       %[cospi_18_64]  \n\t"
+        "madd     $ac1,                 %[load8],       %[cospi_14_64]  \n\t"
+        "extp     %[result3],           $ac1,           31              \n\t"
+
+        "madd     $ac2,                 %[load5],       %[cospi_2_64]   \n\t"
+        "madd     $ac2,                 %[load6],       %[cospi_30_64]  \n\t"
+        "extp     %[result4],           $ac2,           31              \n\t"
+
+        "sub      %[load5],             %[result1],     %[result2]      \n\t"
+        "sub      %[load6],             %[result4],     %[result3]      \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load6],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load5],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load5],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load6],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_9],           $ac1,           31              \n\t"
+        "extp     %[step2_14],          $ac3,           31              \n\t"
+        "add      %[step2_8],           %[result1],     %[result2]      \n\t"
+        "add      %[step2_15],          %[result4],     %[result3]      \n\t"
+
+        : [load5] "=&r"(load5), [load6] "=&r"(load6), [load7] "=&r"(load7),
+          [load8] "=&r"(load8), [result1] "=&r"(result1),
+          [result2] "=&r"(result2), [result3] "=&r"(result3),
+          [result4] "=&r"(result4), [step2_8] "=r"(step2_8),
+          [step2_15] "=r"(step2_15), [step2_9] "=r"(step2_9),
+          [step2_14] "=r"(step2_14)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_30_64] "r"(cospi_30_64), [cospi_2_64] "r"(cospi_2_64),
+          [cospi_14_64] "r"(cospi_14_64), [cospi_18_64] "r"(cospi_18_64),
+          [cospi_24_64] "r"(cospi_24_64), [cospi_8_64] "r"(cospi_8_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             10(%[input])                    \n\t"
+        "lh       %[load2],             22(%[input])                    \n\t"
+        "lh       %[load3],             26(%[input])                    \n\t"
+        "lh       %[load4],             6(%[input])                     \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_22_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_10_64]  \n\t"
+        "extp     %[result1],           $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load3],       %[cospi_6_64]   \n\t"
+        "msub     $ac3,                 %[load4],       %[cospi_26_64]  \n\t"
+        "extp     %[result2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_10_64]  \n\t"
+        "madd     $ac1,                 %[load2],       %[cospi_22_64]  \n\t"
+        "extp     %[result3],           $ac1,           31              \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_26_64]  \n\t"
+        "madd     $ac2,                 %[load4],       %[cospi_6_64]   \n\t"
+        "extp     %[result4],           $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[result2],     %[result1]      \n\t"
+        "sub      %[load2],             %[result4],     %[result3]      \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_10],          $ac1,           31              \n\t"
+        "extp     %[step2_13],          $ac3,           31              \n\t"
+        "add      %[step2_11],          %[result1],     %[result2]      \n\t"
+        "add      %[step2_12],          %[result4],     %[result3]      \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [result1] "=&r"(result1),
+          [result2] "=&r"(result2), [result3] "=&r"(result3),
+          [result4] "=&r"(result4), [step2_10] "=r"(step2_10),
+          [step2_11] "=r"(step2_11), [step2_12] "=r"(step2_12),
+          [step2_13] "=r"(step2_13)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_22_64] "r"(cospi_22_64), [cospi_10_64] "r"(cospi_10_64),
+          [cospi_6_64] "r"(cospi_6_64), [cospi_26_64] "r"(cospi_26_64),
+          [cospi_24_64] "r"(cospi_24_64), [cospi_8_64] "r"(cospi_8_64));
+
+    __asm__ __volatile__(
+        "lh       %[load5],             4(%[input])                     \n\t"
+        "lh       %[load6],             28(%[input])                    \n\t"
+        "lh       %[load7],             20(%[input])                    \n\t"
+        "lh       %[load8],             12(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load5],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load6],       %[cospi_4_64]   \n\t"
+        "extp     %[result1],           $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load7],       %[cospi_12_64]  \n\t"
+        "msub     $ac3,                 %[load8],       %[cospi_20_64]  \n\t"
+        "extp     %[result2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac1,                 %[load7],       %[cospi_20_64]  \n\t"
+        "madd     $ac1,                 %[load8],       %[cospi_12_64]  \n\t"
+        "extp     %[result3],           $ac1,           31              \n\t"
+
+        "madd     $ac2,                 %[load5],       %[cospi_4_64]   \n\t"
+        "madd     $ac2,                 %[load6],       %[cospi_28_64]  \n\t"
+        "extp     %[result4],           $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load5],             %[result4],     %[result3]      \n\t"
+        "sub      %[load5],             %[load5],       %[result1]      \n\t"
+        "add      %[load5],             %[load5],       %[result2]      \n\t"
+
+        "sub      %[load6],             %[result1],     %[result2]      \n\t"
+        "sub      %[load6],             %[load6],       %[result3]      \n\t"
+        "add      %[load6],             %[load6],       %[result4]      \n\t"
+
+        "madd     $ac1,                 %[load5],       %[cospi_16_64]  \n\t"
+        "madd     $ac3,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_5],           $ac1,           31              \n\t"
+        "extp     %[step1_6],           $ac3,           31              \n\t"
+        "add      %[step1_4],           %[result1],     %[result2]      \n\t"
+        "add      %[step1_7],           %[result4],     %[result3]      \n\t"
+
+        : [load5] "=&r"(load5), [load6] "=&r"(load6), [load7] "=&r"(load7),
+          [load8] "=&r"(load8), [result1] "=&r"(result1),
+          [result2] "=&r"(result2), [result3] "=&r"(result3),
+          [result4] "=&r"(result4), [step1_4] "=r"(step1_4),
+          [step1_5] "=r"(step1_5), [step1_6] "=r"(step1_6),
+          [step1_7] "=r"(step1_7)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_20_64] "r"(cospi_20_64), [cospi_12_64] "r"(cospi_12_64),
+          [cospi_4_64] "r"(cospi_4_64), [cospi_28_64] "r"(cospi_28_64),
+          [cospi_16_64] "r"(cospi_16_64));
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        "sub      %[load5],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[load5],             %[load5],       %[step2_9]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_10]     \n\t"
+
+        "madd     $ac0,                 %[load5],       %[cospi_16_64]  \n\t"
+
+        "sub      %[load6],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[load6],             %[load6],       %[step2_10]     \n\t"
+        "add      %[load6],             %[load6],       %[step2_9]      \n\t"
+
+        "madd     $ac1,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load5],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[load5],             %[load5],       %[step2_8]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_11]     \n\t"
+
+        "madd     $ac2,                 %[load5],       %[cospi_16_64]  \n\t"
+
+        "sub      %[load6],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[load6],             %[load6],       %[step2_11]     \n\t"
+        "add      %[load6],             %[load6],       %[step2_8]      \n\t"
+
+        "madd     $ac3,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_10],          $ac0,           31              \n\t"
+        "extp     %[step1_13],          $ac1,           31              \n\t"
+        "extp     %[step1_11],          $ac2,           31              \n\t"
+        "extp     %[step1_12],          $ac3,           31              \n\t"
+
+        : [load5] "=&r"(load5), [load6] "=&r"(load6), [step1_10] "=r"(step1_10),
+          [step1_11] "=r"(step1_11), [step1_12] "=r"(step1_12),
+          [step1_13] "=r"(step1_13)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_14] "r"(step2_14),
+          [step2_13] "r"(step2_13), [step2_9] "r"(step2_9),
+          [step2_10] "r"(step2_10), [step2_15] "r"(step2_15),
+          [step2_12] "r"(step2_12), [step2_8] "r"(step2_8),
+          [step2_11] "r"(step2_11), [cospi_16_64] "r"(cospi_16_64));
+
+    __asm__ __volatile__(
+        "add      %[load5],             %[step1_0],     %[step1_7]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_12]     \n\t"
+        "add      %[load5],             %[load5],       %[step2_15]     \n\t"
+        "add      %[load6],             %[step1_1],     %[step1_6]      \n\t"
+        "add      %[load6],             %[load6],       %[step2_13]     \n\t"
+        "add      %[load6],             %[load6],       %[step2_14]     \n\t"
+        "sh       %[load5],             0(%[output])                    \n\t"
+        "sh       %[load6],             32(%[output])                   \n\t"
+        "sub      %[load5],             %[step1_1],     %[step1_6]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_9]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_10]     \n\t"
+        "sub      %[load6],             %[step1_0],     %[step1_7]      \n\t"
+        "add      %[load6],             %[load6],       %[step2_8]      \n\t"
+        "add      %[load6],             %[load6],       %[step2_11]     \n\t"
+        "sh       %[load5],             192(%[output])                  \n\t"
+        "sh       %[load6],             224(%[output])                  \n\t"
+        "sub      %[load5],             %[step1_0],     %[step1_7]      \n\t"
+        "sub      %[load5],             %[load5],       %[step2_8]      \n\t"
+        "sub      %[load5],             %[load5],       %[step2_11]     \n\t"
+        "sub      %[load6],             %[step1_1],     %[step1_6]      \n\t"
+        "sub      %[load6],             %[load6],       %[step2_9]      \n\t"
+        "sub      %[load6],             %[load6],       %[step2_10]     \n\t"
+        "sh       %[load5],             256(%[output])                  \n\t"
+        "sh       %[load6],             288(%[output])                  \n\t"
+        "add      %[load5],             %[step1_1],     %[step1_6]      \n\t"
+        "sub      %[load5],             %[load5],       %[step2_13]     \n\t"
+        "sub      %[load5],             %[load5],       %[step2_14]     \n\t"
+        "add      %[load6],             %[step1_0],     %[step1_7]      \n\t"
+        "sub      %[load6],             %[load6],       %[step2_12]     \n\t"
+        "sub      %[load6],             %[load6],       %[step2_15]     \n\t"
+        "sh       %[load5],             448(%[output])                  \n\t"
+        "sh       %[load6],             480(%[output])                  \n\t"
+
+        : [load5] "=&r"(load5), [load6] "=&r"(load6)
+        : [output] "r"(output), [step1_0] "r"(step1_0), [step1_1] "r"(step1_1),
+          [step1_6] "r"(step1_6), [step1_7] "r"(step1_7),
+          [step2_8] "r"(step2_8), [step2_9] "r"(step2_9),
+          [step2_10] "r"(step2_10), [step2_11] "r"(step2_11),
+          [step2_12] "r"(step2_12), [step2_13] "r"(step2_13),
+          [step2_14] "r"(step2_14), [step2_15] "r"(step2_15));
+
+    __asm__ __volatile__(
+        "add      %[load5],             %[step1_2],     %[step1_5]      \n\t"
+        "add      %[load5],             %[load5],       %[step1_13]     \n\t"
+        "add      %[load6],             %[step1_3],     %[step1_4]      \n\t"
+        "add      %[load6],             %[load6],       %[step1_12]     \n\t"
+        "sh       %[load5],             64(%[output])                   \n\t"
+        "sh       %[load6],             96(%[output])                   \n\t"
+        "sub      %[load5],             %[step1_3],     %[step1_4]      \n\t"
+        "add      %[load5],             %[load5],       %[step1_11]     \n\t"
+        "sub      %[load6],             %[step1_2],     %[step1_5]      \n\t"
+        "add      %[load6],             %[load6],       %[step1_10]     \n\t"
+        "sh       %[load5],             128(%[output])                  \n\t"
+        "sh       %[load6],             160(%[output])                  \n\t"
+        "sub      %[load5],             %[step1_2],     %[step1_5]      \n\t"
+        "sub      %[load5],             %[load5],       %[step1_10]     \n\t"
+        "sub      %[load6],             %[step1_3],     %[step1_4]      \n\t"
+        "sub      %[load6],             %[load6],       %[step1_11]     \n\t"
+        "sh       %[load5],             320(%[output])                  \n\t"
+        "sh       %[load6],             352(%[output])                  \n\t"
+        "add      %[load5],             %[step1_3],     %[step1_4]      \n\t"
+        "sub      %[load5],             %[load5],       %[step1_12]     \n\t"
+        "add      %[load6],             %[step1_2],     %[step1_5]      \n\t"
+        "sub      %[load6],             %[load6],       %[step1_13]     \n\t"
+        "sh       %[load5],             384(%[output])                  \n\t"
+        "sh       %[load6],             416(%[output])                  \n\t"
+
+        : [load5] "=&r"(load5), [load6] "=&r"(load6)
+        : [output] "r"(output), [step1_2] "r"(step1_2), [step1_3] "r"(step1_3),
+          [step1_4] "r"(step1_4), [step1_5] "r"(step1_5),
+          [step1_10] "r"(step1_10), [step1_11] "r"(step1_11),
+          [step1_12] "r"(step1_12), [step1_13] "r"(step1_13));
+
+    input += 16;
+    output += 1;
+  }
+}
+
+void idct16_cols_add_blk_dspr2(int16_t *input, uint8_t *dest, int dest_stride) {
+  int i;
+  int step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6, step1_7;
+  int step1_8, step1_9, step1_10, step1_11;
+  int step1_12, step1_13, step1_14, step1_15;
+  int step2_0, step2_1, step2_2, step2_3;
+  int step2_8, step2_9, step2_10, step2_11;
+  int step2_12, step2_13, step2_14, step2_15;
+  int load1, load2, load3, load4, load5, load6, load7, load8;
+  int result1, result2, result3, result4;
+  const int const_2_power_13 = 8192;
+  uint8_t *dest_pix;
+  uint8_t *cm = aom_ff_cropTbl;
+
+  /* prefetch aom_ff_cropTbl */
+  prefetch_load(aom_ff_cropTbl);
+  prefetch_load(aom_ff_cropTbl + 32);
+  prefetch_load(aom_ff_cropTbl + 64);
+  prefetch_load(aom_ff_cropTbl + 96);
+  prefetch_load(aom_ff_cropTbl + 128);
+  prefetch_load(aom_ff_cropTbl + 160);
+  prefetch_load(aom_ff_cropTbl + 192);
+  prefetch_load(aom_ff_cropTbl + 224);
+
+  for (i = 0; i < 16; ++i) {
+    dest_pix = (dest + i);
+    __asm__ __volatile__(
+        "lh       %[load1],              0(%[input])                    \n\t"
+        "lh       %[load2],             16(%[input])                    \n\t"
+        "lh       %[load3],              8(%[input])                    \n\t"
+        "lh       %[load4],             24(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "add      %[result1],           %[load1],       %[load2]        \n\t"
+        "sub      %[result2],           %[load1],       %[load2]        \n\t"
+        "madd     $ac1,                 %[result1],     %[cospi_16_64]  \n\t"
+        "madd     $ac2,                 %[result2],     %[cospi_16_64]  \n\t"
+        "extp     %[step2_0],           $ac1,           31              \n\t"
+        "extp     %[step2_1],           $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "madd     $ac3,                 %[load3],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load4],       %[cospi_8_64]   \n\t"
+        "extp     %[step2_2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_24_64]  \n\t"
+        "extp     %[step2_3],           $ac1,           31              \n\t"
+
+        "add      %[step1_0],           %[step2_0],     %[step2_3]      \n\t"
+        "add      %[step1_1],           %[step2_1],     %[step2_2]      \n\t"
+        "sub      %[step1_2],           %[step2_1],     %[step2_2]      \n\t"
+        "sub      %[step1_3],           %[step2_0],     %[step2_3]      \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [result1] "=&r"(result1),
+          [result2] "=&r"(result2), [step2_0] "=&r"(step2_0),
+          [step2_1] "=&r"(step2_1), [step2_2] "=&r"(step2_2),
+          [step2_3] "=&r"(step2_3), [step1_0] "=r"(step1_0),
+          [step1_1] "=r"(step1_1), [step1_2] "=r"(step1_2),
+          [step1_3] "=r"(step1_3)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_24_64] "r"(cospi_24_64), [cospi_8_64] "r"(cospi_8_64),
+          [cospi_16_64] "r"(cospi_16_64));
+
+    __asm__ __volatile__(
+        "lh       %[load5],             2(%[input])                     \n\t"
+        "lh       %[load6],             30(%[input])                    \n\t"
+        "lh       %[load7],             18(%[input])                    \n\t"
+        "lh       %[load8],             14(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load5],       %[cospi_30_64]  \n\t"
+        "msub     $ac1,                 %[load6],       %[cospi_2_64]   \n\t"
+        "extp     %[result1],           $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load7],       %[cospi_14_64]  \n\t"
+        "msub     $ac3,                 %[load8],       %[cospi_18_64]  \n\t"
+        "extp     %[result2],           $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac1,                 %[load7],       %[cospi_18_64]  \n\t"
+        "madd     $ac1,                 %[load8],       %[cospi_14_64]  \n\t"
+        "extp     %[result3],           $ac1,           31              \n\t"
+
+        "madd     $ac2,                 %[load5],        %[cospi_2_64]  \n\t"
+        "madd     $ac2,                 %[load6],        %[cospi_30_64] \n\t"
+        "extp     %[result4],           $ac2,            31             \n\t"
+
+        "sub      %[load5],             %[result1],     %[result2]      \n\t"
+        "sub      %[load6],             %[result4],     %[result3]      \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load6],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load5],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load5],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load6],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_9],           $ac1,           31              \n\t"
+        "extp     %[step2_14],          $ac3,           31              \n\t"
+        "add      %[step2_8],           %[result1],     %[result2]      \n\t"
+        "add      %[step2_15],          %[result4],     %[result3]      \n\t"
+
+        : [load5] "=&r"(load5), [load6] "=&r"(load6), [load7] "=&r"(load7),
+          [load8] "=&r"(load8), [result1] "=&r"(result1),
+          [result2] "=&r"(result2), [result3] "=&r"(result3),
+          [result4] "=&r"(result4), [step2_8] "=r"(step2_8),
+          [step2_15] "=r"(step2_15), [step2_9] "=r"(step2_9),
+          [step2_14] "=r"(step2_14)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_30_64] "r"(cospi_30_64), [cospi_2_64] "r"(cospi_2_64),
+          [cospi_14_64] "r"(cospi_14_64), [cospi_18_64] "r"(cospi_18_64),
+          [cospi_24_64] "r"(cospi_24_64), [cospi_8_64] "r"(cospi_8_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             10(%[input])                    \n\t"
+        "lh       %[load2],             22(%[input])                    \n\t"
+        "lh       %[load3],             26(%[input])                    \n\t"
+        "lh       %[load4],             6(%[input])                     \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],    %[cospi_22_64]     \n\t"
+        "msub     $ac1,                 %[load2],    %[cospi_10_64]     \n\t"
+        "extp     %[result1],           $ac1,        31                 \n\t"
+
+        "madd     $ac3,                 %[load3],    %[cospi_6_64]      \n\t"
+        "msub     $ac3,                 %[load4],    %[cospi_26_64]     \n\t"
+        "extp     %[result2],           $ac3,        31                 \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac1,                 %[load1],    %[cospi_10_64]     \n\t"
+        "madd     $ac1,                 %[load2],    %[cospi_22_64]     \n\t"
+        "extp     %[result3],           $ac1,        31                 \n\t"
+
+        "madd     $ac2,                 %[load3],    %[cospi_26_64]     \n\t"
+        "madd     $ac2,                 %[load4],    %[cospi_6_64]      \n\t"
+        "extp     %[result4],           $ac2,        31                 \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[result2],     %[result1]      \n\t"
+        "sub      %[load2],             %[result4],     %[result3]      \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_10],          $ac1,           31              \n\t"
+        "extp     %[step2_13],          $ac3,           31              \n\t"
+        "add      %[step2_11],          %[result1],     %[result2]      \n\t"
+        "add      %[step2_12],          %[result4],     %[result3]      \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [result1] "=&r"(result1),
+          [result2] "=&r"(result2), [result3] "=&r"(result3),
+          [result4] "=&r"(result4), [step2_10] "=r"(step2_10),
+          [step2_11] "=r"(step2_11), [step2_12] "=r"(step2_12),
+          [step2_13] "=r"(step2_13)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_22_64] "r"(cospi_22_64), [cospi_10_64] "r"(cospi_10_64),
+          [cospi_6_64] "r"(cospi_6_64), [cospi_26_64] "r"(cospi_26_64),
+          [cospi_24_64] "r"(cospi_24_64), [cospi_8_64] "r"(cospi_8_64));
+
+    __asm__ __volatile__(
+        "lh       %[load5],             4(%[input])                   \n\t"
+        "lh       %[load6],             28(%[input])                  \n\t"
+        "lh       %[load7],             20(%[input])                  \n\t"
+        "lh       %[load8],             12(%[input])                  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                          \n\t"
+        "mthi     $zero,                $ac1                          \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                          \n\t"
+        "mthi     $zero,                $ac3                          \n\t"
+
+        "madd     $ac1,                 %[load5],    %[cospi_28_64]   \n\t"
+        "msub     $ac1,                 %[load6],    %[cospi_4_64]    \n\t"
+        "extp     %[result1],           $ac1,        31               \n\t"
+
+        "madd     $ac3,                 %[load7],    %[cospi_12_64]   \n\t"
+        "msub     $ac3,                 %[load8],    %[cospi_20_64]   \n\t"
+        "extp     %[result2],           $ac3,        31               \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                          \n\t"
+        "mthi     $zero,                $ac1                          \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                          \n\t"
+        "mthi     $zero,                $ac2                          \n\t"
+
+        "madd     $ac1,                 %[load7],    %[cospi_20_64]   \n\t"
+        "madd     $ac1,                 %[load8],    %[cospi_12_64]   \n\t"
+        "extp     %[result3],           $ac1,        31               \n\t"
+
+        "madd     $ac2,                 %[load5],    %[cospi_4_64]    \n\t"
+        "madd     $ac2,                 %[load6],    %[cospi_28_64]   \n\t"
+        "extp     %[result4],           $ac2,        31               \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load5],             %[result4],     %[result3]      \n\t"
+        "sub      %[load5],             %[load5],       %[result1]      \n\t"
+        "add      %[load5],             %[load5],       %[result2]      \n\t"
+
+        "sub      %[load6],             %[result1],     %[result2]      \n\t"
+        "sub      %[load6],             %[load6],       %[result3]      \n\t"
+        "add      %[load6],             %[load6],       %[result4]      \n\t"
+
+        "madd     $ac1,                 %[load5],       %[cospi_16_64]  \n\t"
+        "madd     $ac3,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_5],           $ac1,           31              \n\t"
+        "extp     %[step1_6],           $ac3,           31              \n\t"
+
+        "add      %[step1_4],           %[result1],     %[result2]      \n\t"
+        "add      %[step1_7],           %[result4],     %[result3]      \n\t"
+
+        : [load5] "=&r"(load5), [load6] "=&r"(load6), [load7] "=&r"(load7),
+          [load8] "=&r"(load8), [result1] "=&r"(result1),
+          [result2] "=&r"(result2), [result3] "=&r"(result3),
+          [result4] "=&r"(result4), [step1_4] "=r"(step1_4),
+          [step1_5] "=r"(step1_5), [step1_6] "=r"(step1_6),
+          [step1_7] "=r"(step1_7)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_20_64] "r"(cospi_20_64), [cospi_12_64] "r"(cospi_12_64),
+          [cospi_4_64] "r"(cospi_4_64), [cospi_28_64] "r"(cospi_28_64),
+          [cospi_16_64] "r"(cospi_16_64));
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        "sub      %[load5],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[load5],             %[load5],       %[step2_9]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_10]     \n\t"
+
+        "madd     $ac0,                 %[load5],       %[cospi_16_64]  \n\t"
+
+        "sub      %[load6],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[load6],             %[load6],       %[step2_10]     \n\t"
+        "add      %[load6],             %[load6],       %[step2_9]      \n\t"
+
+        "madd     $ac1,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load5],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[load5],             %[load5],       %[step2_8]      \n\t"
+        "add      %[load5],             %[load5],       %[step2_11]     \n\t"
+
+        "madd     $ac2,                 %[load5],       %[cospi_16_64]  \n\t"
+
+        "sub      %[load6],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[load6],             %[load6],       %[step2_11]     \n\t"
+        "add      %[load6],             %[load6],       %[step2_8]      \n\t"
+
+        "madd     $ac3,                 %[load6],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_10],          $ac0,           31              \n\t"
+        "extp     %[step1_13],          $ac1,           31              \n\t"
+        "extp     %[step1_11],          $ac2,           31              \n\t"
+        "extp     %[step1_12],          $ac3,           31              \n\t"
+
+        : [load5] "=&r"(load5), [load6] "=&r"(load6), [step1_10] "=r"(step1_10),
+          [step1_11] "=r"(step1_11), [step1_12] "=r"(step1_12),
+          [step1_13] "=r"(step1_13)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_14] "r"(step2_14),
+          [step2_13] "r"(step2_13), [step2_9] "r"(step2_9),
+          [step2_10] "r"(step2_10), [step2_15] "r"(step2_15),
+          [step2_12] "r"(step2_12), [step2_8] "r"(step2_8),
+          [step2_11] "r"(step2_11), [cospi_16_64] "r"(cospi_16_64));
+
+    step1_8 = step2_8 + step2_11;
+    step1_9 = step2_9 + step2_10;
+    step1_14 = step2_13 + step2_14;
+    step1_15 = step2_12 + step2_15;
+
+    __asm__ __volatile__(
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "add      %[load5],         %[step1_0],         %[step1_7]      \n\t"
+        "add      %[load5],         %[load5],           %[step1_15]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "add      %[load6],         %[step1_1],         %[step1_6]      \n\t"
+        "add      %[load6],         %[load6],           %[step1_14]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "add      %[load5],         %[step1_2],         %[step1_5]      \n\t"
+        "add      %[load5],         %[load5],           %[step1_13]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "add      %[load6],         %[step1_3],         %[step1_4]      \n\t"
+        "add      %[load6],         %[load6],           %[step1_12]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "sub      %[load5],         %[step1_3],         %[step1_4]      \n\t"
+        "add      %[load5],         %[load5],           %[step1_11]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "sub      %[load6],         %[step1_2],         %[step1_5]      \n\t"
+        "add      %[load6],         %[load6],           %[step1_10]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "sub      %[load5],         %[step1_1],         %[step1_6]      \n\t"
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "add      %[load5],         %[load5],           %[step1_9]      \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "sub      %[load6],         %[step1_0],         %[step1_7]      \n\t"
+        "add      %[load6],         %[load6],           %[step1_8]      \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "sub      %[load5],         %[step1_0],         %[step1_7]      \n\t"
+        "sub      %[load5],         %[load5],           %[step1_8]      \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "sub      %[load6],         %[step1_1],         %[step1_6]      \n\t"
+        "sub      %[load6],         %[load6],           %[step1_9]      \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "sub      %[load5],         %[step1_2],         %[step1_5]      \n\t"
+        "sub      %[load5],         %[load5],           %[step1_10]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "sub      %[load6],         %[step1_3],         %[step1_4]      \n\t"
+        "sub      %[load6],         %[load6],           %[step1_11]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "add      %[load5],         %[step1_3],         %[step1_4]      \n\t"
+        "sub      %[load5],         %[load5],           %[step1_12]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "add      %[load6],         %[step1_2],         %[step1_5]      \n\t"
+        "sub      %[load6],         %[load6],           %[step1_13]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[load7],         0(%[dest_pix])                      \n\t"
+        "add      %[load5],         %[step1_1],         %[step1_6]      \n\t"
+        "sub      %[load5],         %[load5],           %[step1_14]     \n\t"
+        "addi     %[load5],         %[load5],           32              \n\t"
+        "sra      %[load5],         %[load5],           6               \n\t"
+        "add      %[load7],         %[load7],           %[load5]        \n\t"
+        "lbux     %[load5],         %[load7](%[cm])                     \n\t"
+        "add      %[load6],         %[step1_0],         %[step1_7]      \n\t"
+        "sub      %[load6],         %[load6],           %[step1_15]     \n\t"
+        "sb       %[load5],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[load8],         0(%[dest_pix])                      \n\t"
+        "addi     %[load6],         %[load6],           32              \n\t"
+        "sra      %[load6],         %[load6],           6               \n\t"
+        "add      %[load8],         %[load8],           %[load6]        \n\t"
+        "lbux     %[load6],         %[load8](%[cm])                     \n\t"
+        "sb       %[load6],         0(%[dest_pix])                      \n\t"
+
+        : [load5] "=&r"(load5), [load6] "=&r"(load6), [load7] "=&r"(load7),
+          [load8] "=&r"(load8), [dest_pix] "+r"(dest_pix)
+        :
+        [cm] "r"(cm), [dest_stride] "r"(dest_stride), [step1_0] "r"(step1_0),
+        [step1_1] "r"(step1_1), [step1_2] "r"(step1_2), [step1_3] "r"(step1_3),
+        [step1_4] "r"(step1_4), [step1_5] "r"(step1_5), [step1_6] "r"(step1_6),
+        [step1_7] "r"(step1_7), [step1_8] "r"(step1_8), [step1_9] "r"(step1_9),
+        [step1_10] "r"(step1_10), [step1_11] "r"(step1_11),
+        [step1_12] "r"(step1_12), [step1_13] "r"(step1_13),
+        [step1_14] "r"(step1_14), [step1_15] "r"(step1_15));
+
+    input += 16;
+  }
+}
+
+void aom_idct16x16_256_add_dspr2(const int16_t *input, uint8_t *dest,
+                                 int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t, out[16 * 16]);
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp    %[pos],    1    \n\t" : : [pos] "r"(pos));
+
+  // First transform rows
+  idct16_rows_dspr2(input, out, 16);
+
+  // Then transform columns and add to dest
+  idct16_cols_add_blk_dspr2(out, dest, dest_stride);
+}
+
+void aom_idct16x16_10_add_dspr2(const int16_t *input, uint8_t *dest,
+                                int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t, out[16 * 16]);
+  int16_t *outptr = out;
+  uint32_t i;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp    %[pos],    1    \n\t" : : [pos] "r"(pos));
+
+  // First transform rows. Since all non-zero dct coefficients are in
+  // upper-left 4x4 area, we only need to calculate first 4 rows here.
+  idct16_rows_dspr2(input, outptr, 4);
+
+  outptr += 4;
+  for (i = 0; i < 6; ++i) {
+    __asm__ __volatile__(
+        "sw     $zero,    0(%[outptr])     \n\t"
+        "sw     $zero,   32(%[outptr])     \n\t"
+        "sw     $zero,   64(%[outptr])     \n\t"
+        "sw     $zero,   96(%[outptr])     \n\t"
+        "sw     $zero,  128(%[outptr])     \n\t"
+        "sw     $zero,  160(%[outptr])     \n\t"
+        "sw     $zero,  192(%[outptr])     \n\t"
+        "sw     $zero,  224(%[outptr])     \n\t"
+        "sw     $zero,  256(%[outptr])     \n\t"
+        "sw     $zero,  288(%[outptr])     \n\t"
+        "sw     $zero,  320(%[outptr])     \n\t"
+        "sw     $zero,  352(%[outptr])     \n\t"
+        "sw     $zero,  384(%[outptr])     \n\t"
+        "sw     $zero,  416(%[outptr])     \n\t"
+        "sw     $zero,  448(%[outptr])     \n\t"
+        "sw     $zero,  480(%[outptr])     \n\t"
+
+        :
+        : [outptr] "r"(outptr));
+
+    outptr += 2;
+  }
+
+  // Then transform columns
+  idct16_cols_add_blk_dspr2(out, dest, dest_stride);
+}
+
+void aom_idct16x16_1_add_dspr2(const int16_t *input, uint8_t *dest,
+                               int dest_stride) {
+  uint32_t pos = 45;
+  int32_t out;
+  int32_t r;
+  int32_t a1, absa1;
+  int32_t vector_a1;
+  int32_t t1, t2, t3, t4;
+  int32_t vector_1, vector_2, vector_3, vector_4;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+
+                       :
+                       : [pos] "r"(pos));
+
+  out = DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input[0]);
+  __asm__ __volatile__(
+      "addi     %[out],     %[out],     32      \n\t"
+      "sra      %[a1],      %[out],     6       \n\t"
+
+      : [out] "+r"(out), [a1] "=r"(a1)
+      :);
+
+  if (a1 < 0) {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__(
+        "abs        %[absa1],       %[a1]       \n\t"
+        "replv.qb   %[vector_a1],   %[absa1]    \n\t"
+
+        : [absa1] "=r"(absa1), [vector_a1] "=r"(vector_a1)
+        : [a1] "r"(a1));
+
+    for (r = 16; r--;) {
+      __asm__ __volatile__(
+          "lw             %[t1],          0(%[dest])                      \n\t"
+          "lw             %[t2],          4(%[dest])                      \n\t"
+          "lw             %[t3],          8(%[dest])                      \n\t"
+          "lw             %[t4],          12(%[dest])                     \n\t"
+          "subu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    0(%[dest])                      \n\t"
+          "sw             %[vector_2],    4(%[dest])                      \n\t"
+          "sw             %[vector_3],    8(%[dest])                      \n\t"
+          "sw             %[vector_4],    12(%[dest])                     \n\t"
+          "add            %[dest],        %[dest],        %[dest_stride]  \n\t"
+
+          : [t1] "=&r"(t1), [t2] "=&r"(t2), [t3] "=&r"(t3), [t4] "=&r"(t4),
+            [vector_1] "=&r"(vector_1), [vector_2] "=&r"(vector_2),
+            [vector_3] "=&r"(vector_3), [vector_4] "=&r"(vector_4),
+            [dest] "+&r"(dest)
+          : [dest_stride] "r"(dest_stride), [vector_a1] "r"(vector_a1));
+    }
+  } else {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__("replv.qb   %[vector_a1],   %[a1]   \n\t"
+
+                         : [vector_a1] "=r"(vector_a1)
+                         : [a1] "r"(a1));
+
+    for (r = 16; r--;) {
+      __asm__ __volatile__(
+          "lw             %[t1],          0(%[dest])                      \n\t"
+          "lw             %[t2],          4(%[dest])                      \n\t"
+          "lw             %[t3],          8(%[dest])                      \n\t"
+          "lw             %[t4],          12(%[dest])                     \n\t"
+          "addu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    0(%[dest])                      \n\t"
+          "sw             %[vector_2],    4(%[dest])                      \n\t"
+          "sw             %[vector_3],    8(%[dest])                      \n\t"
+          "sw             %[vector_4],    12(%[dest])                     \n\t"
+          "add            %[dest],        %[dest],        %[dest_stride]  \n\t"
+
+          : [t1] "=&r"(t1), [t2] "=&r"(t2), [t3] "=&r"(t3), [t4] "=&r"(t4),
+            [vector_1] "=&r"(vector_1), [vector_2] "=&r"(vector_2),
+            [vector_3] "=&r"(vector_3), [vector_4] "=&r"(vector_4),
+            [dest] "+&r"(dest)
+          : [dest_stride] "r"(dest_stride), [vector_a1] "r"(vector_a1));
+    }
+  }
+}
+
+void iadst16_dspr2(const int16_t *input, int16_t *output) {
+  int s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14, s15;
+
+  int x0 = input[15];
+  int x1 = input[0];
+  int x2 = input[13];
+  int x3 = input[2];
+  int x4 = input[11];
+  int x5 = input[4];
+  int x6 = input[9];
+  int x7 = input[6];
+  int x8 = input[7];
+  int x9 = input[8];
+  int x10 = input[5];
+  int x11 = input[10];
+  int x12 = input[3];
+  int x13 = input[12];
+  int x14 = input[1];
+  int x15 = input[14];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7 | x8 | x9 | x10 | x11 | x12 |
+        x13 | x14 | x15)) {
+    output[0] = output[1] = output[2] = output[3] = output[4] = output[5] =
+        output[6] = output[7] = output[8] = output[9] = output[10] =
+            output[11] = output[12] = output[13] = output[14] = output[15] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = x0 * cospi_1_64 + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64 + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64 + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64 - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64 - x15 * cospi_29_64;
+
+  x0 = dct_const_round_shift(s0 + s8);
+  x1 = dct_const_round_shift(s1 + s9);
+  x2 = dct_const_round_shift(s2 + s10);
+  x3 = dct_const_round_shift(s3 + s11);
+  x4 = dct_const_round_shift(s4 + s12);
+  x5 = dct_const_round_shift(s5 + s13);
+  x6 = dct_const_round_shift(s6 + s14);
+  x7 = dct_const_round_shift(s7 + s15);
+  x8 = dct_const_round_shift(s0 - s8);
+  x9 = dct_const_round_shift(s1 - s9);
+  x10 = dct_const_round_shift(s2 - s10);
+  x11 = dct_const_round_shift(s3 - s11);
+  x12 = dct_const_round_shift(s4 - s12);
+  x13 = dct_const_round_shift(s5 - s13);
+  x14 = dct_const_round_shift(s6 - s14);
+  x15 = dct_const_round_shift(s7 - s15);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 = x8 * cospi_4_64 + x9 * cospi_28_64;
+  s9 = x8 * cospi_28_64 - x9 * cospi_4_64;
+  s10 = x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 = x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = -x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 = x12 * cospi_4_64 + x13 * cospi_28_64;
+  s14 = -x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 = x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = s0 + s4;
+  x1 = s1 + s5;
+  x2 = s2 + s6;
+  x3 = s3 + s7;
+  x4 = s0 - s4;
+  x5 = s1 - s5;
+  x6 = s2 - s6;
+  x7 = s3 - s7;
+  x8 = dct_const_round_shift(s8 + s12);
+  x9 = dct_const_round_shift(s9 + s13);
+  x10 = dct_const_round_shift(s10 + s14);
+  x11 = dct_const_round_shift(s11 + s15);
+  x12 = dct_const_round_shift(s8 - s12);
+  x13 = dct_const_round_shift(s9 - s13);
+  x14 = dct_const_round_shift(s10 - s14);
+  x15 = dct_const_round_shift(s11 - s15);
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64 + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = -x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 = x6 * cospi_8_64 + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64 + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = -x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 = x14 * cospi_8_64 + x15 * cospi_24_64;
+
+  x0 = s0 + s2;
+  x1 = s1 + s3;
+  x2 = s0 - s2;
+  x3 = s1 - s3;
+  x4 = dct_const_round_shift(s4 + s6);
+  x5 = dct_const_round_shift(s5 + s7);
+  x6 = dct_const_round_shift(s4 - s6);
+  x7 = dct_const_round_shift(s5 - s7);
+  x8 = s8 + s10;
+  x9 = s9 + s11;
+  x10 = s8 - s10;
+  x11 = s9 - s11;
+  x12 = dct_const_round_shift(s12 + s14);
+  x13 = dct_const_round_shift(s13 + s15);
+  x14 = dct_const_round_shift(s12 - s14);
+  x15 = dct_const_round_shift(s13 - s15);
+
+  // stage 4
+  s2 = (-cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (-x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (-x10 + x11);
+  s14 = (-cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = dct_const_round_shift(s2);
+  x3 = dct_const_round_shift(s3);
+  x6 = dct_const_round_shift(s6);
+  x7 = dct_const_round_shift(s7);
+  x10 = dct_const_round_shift(s10);
+  x11 = dct_const_round_shift(s11);
+  x14 = dct_const_round_shift(s14);
+  x15 = dct_const_round_shift(s15);
+
+  output[0] = x0;
+  output[1] = -x8;
+  output[2] = x12;
+  output[3] = -x4;
+  output[4] = x6;
+  output[5] = x14;
+  output[6] = x10;
+  output[7] = x2;
+  output[8] = x3;
+  output[9] = x11;
+  output[10] = x15;
+  output[11] = x7;
+  output[12] = x5;
+  output[13] = -x13;
+  output[14] = x9;
+  output[15] = -x1;
+}
+
+#endif  // HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/itrans32_cols_dspr2.c b/third_party/aom/aom_dsp/mips/itrans32_cols_dspr2.c
new file mode 100644
index 0000000000..d469d1ad0b
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/itrans32_cols_dspr2.c
@@ -0,0 +1,1042 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "aom_dsp/mips/inv_txfm_dspr2.h"
+#include "aom_dsp/txfm_common.h"
+
+#if HAVE_DSPR2
+void aom_idct32_cols_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                   int dest_stride) {
+  int16_t step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6;
+  int16_t step1_7, step1_8, step1_9, step1_10, step1_11, step1_12, step1_13;
+  int16_t step1_14, step1_15, step1_16, step1_17, step1_18, step1_19;
+  int16_t step1_20, step1_21, step1_22, step1_23, step1_24, step1_25, step1_26;
+  int16_t step1_27, step1_28, step1_29, step1_30, step1_31;
+  int16_t step2_0, step2_1, step2_2, step2_3, step2_4, step2_5, step2_6;
+  int16_t step2_7, step2_8, step2_9, step2_10, step2_11, step2_12, step2_13;
+  int16_t step2_14, step2_15, step2_16, step2_17, step2_18, step2_19, step2_20;
+  int16_t step2_21, step2_22, step2_23, step2_24, step2_25, step2_26, step2_27;
+  int16_t step2_28, step2_29, step2_30, step2_31;
+  int16_t step3_8, step3_9, step3_10, step3_11, step3_12, step3_13, step3_14;
+  int16_t step3_15, step3_16, step3_17, step3_18, step3_19, step3_20, step3_21;
+  int16_t step3_22, step3_23, step3_24, step3_25, step3_26, step3_27;
+  int16_t step3_28, step3_29, step3_30, step3_31;
+  int temp0, temp1, temp2, temp3;
+  int load1, load2, load3, load4;
+  int result1, result2;
+  int i, temp21;
+  uint8_t *dest_pix, *dest_pix1;
+  const int const_2_power_13 = 8192;
+  uint8_t *cm = aom_ff_cropTbl;
+
+  /* prefetch aom_ff_cropTbl */
+  prefetch_load(aom_ff_cropTbl);
+  prefetch_load(aom_ff_cropTbl + 32);
+  prefetch_load(aom_ff_cropTbl + 64);
+  prefetch_load(aom_ff_cropTbl + 96);
+  prefetch_load(aom_ff_cropTbl + 128);
+  prefetch_load(aom_ff_cropTbl + 160);
+  prefetch_load(aom_ff_cropTbl + 192);
+  prefetch_load(aom_ff_cropTbl + 224);
+
+  for (i = 0; i < 32; ++i) {
+    dest_pix = dest + i;
+    dest_pix1 = dest + i + 31 * dest_stride;
+
+    __asm__ __volatile__(
+        "lh       %[load1],             2(%[input])                     \n\t"
+        "lh       %[load2],             62(%[input])                    \n\t"
+        "lh       %[load3],             34(%[input])                    \n\t"
+        "lh       %[load4],             30(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_31_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_1_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_1_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_31_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_15_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_17_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_17_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_15_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp3],       %[temp2]        \n\t"
+        "sub      %[load2],             %[temp0],       %[temp1]        \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+
+        "extp     %[step1_17],          $ac1,           31              \n\t"
+        "extp     %[step1_30],          $ac3,           31              \n\t"
+        "add      %[step1_16],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_31],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_16] "=r"(step1_16),
+          [step1_17] "=r"(step1_17), [step1_30] "=r"(step1_30),
+          [step1_31] "=r"(step1_31)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_31_64] "r"(cospi_31_64), [cospi_1_64] "r"(cospi_1_64),
+          [cospi_4_64] "r"(cospi_4_64), [cospi_17_64] "r"(cospi_17_64),
+          [cospi_15_64] "r"(cospi_15_64), [cospi_28_64] "r"(cospi_28_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             18(%[input])                    \n\t"
+        "lh       %[load2],             46(%[input])                    \n\t"
+        "lh       %[load3],             50(%[input])                    \n\t"
+        "lh       %[load4],             14(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_23_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_9_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_9_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_23_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_7_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_25_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_25_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_7_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+
+        "extp     %[step1_18],          $ac1,           31              \n\t"
+        "extp     %[step1_29],          $ac3,           31              \n\t"
+        "add      %[step1_19],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_28],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_18] "=r"(step1_18),
+          [step1_19] "=r"(step1_19), [step1_28] "=r"(step1_28),
+          [step1_29] "=r"(step1_29)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_23_64] "r"(cospi_23_64), [cospi_9_64] "r"(cospi_9_64),
+          [cospi_4_64] "r"(cospi_4_64), [cospi_7_64] "r"(cospi_7_64),
+          [cospi_25_64] "r"(cospi_25_64), [cospi_28_64] "r"(cospi_28_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             10(%[input])                    \n\t"
+        "lh       %[load2],             54(%[input])                    \n\t"
+        "lh       %[load3],             42(%[input])                    \n\t"
+        "lh       %[load4],             22(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_27_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_5_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_5_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_27_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_11_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_21_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_21_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_11_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[temp3],       %[temp2]        \n\t"
+
+        "madd     $ac1,                 %[load2],       %[cospi_12_64]  \n\t"
+        "msub     $ac1,                 %[load1],       %[cospi_20_64]  \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_12_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_20_64]  \n\t"
+
+        "extp     %[step1_21],          $ac1,           31              \n\t"
+        "extp     %[step1_26],          $ac3,           31              \n\t"
+        "add      %[step1_20],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_27],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_20] "=r"(step1_20),
+          [step1_21] "=r"(step1_21), [step1_26] "=r"(step1_26),
+          [step1_27] "=r"(step1_27)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_27_64] "r"(cospi_27_64), [cospi_5_64] "r"(cospi_5_64),
+          [cospi_11_64] "r"(cospi_11_64), [cospi_21_64] "r"(cospi_21_64),
+          [cospi_12_64] "r"(cospi_12_64), [cospi_20_64] "r"(cospi_20_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             26(%[input])                    \n\t"
+        "lh       %[load2],             38(%[input])                    \n\t"
+        "lh       %[load3],             58(%[input])                    \n\t"
+        "lh       %[load4],              6(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_19_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_13_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_13_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_19_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_3_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_29_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_29_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_3_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+        "msub     $ac1,                 %[load1],       %[cospi_12_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_20_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_20_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_12_64]  \n\t"
+        "extp     %[step1_22],          $ac1,           31              \n\t"
+        "extp     %[step1_25],          $ac3,           31              \n\t"
+        "add      %[step1_23],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_24],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_22] "=r"(step1_22),
+          [step1_23] "=r"(step1_23), [step1_24] "=r"(step1_24),
+          [step1_25] "=r"(step1_25)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_19_64] "r"(cospi_19_64), [cospi_13_64] "r"(cospi_13_64),
+          [cospi_3_64] "r"(cospi_3_64), [cospi_29_64] "r"(cospi_29_64),
+          [cospi_12_64] "r"(cospi_12_64), [cospi_20_64] "r"(cospi_20_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],              4(%[input])                    \n\t"
+        "lh       %[load2],             60(%[input])                    \n\t"
+        "lh       %[load3],             36(%[input])                    \n\t"
+        "lh       %[load4],             28(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_30_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_2_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_2_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_30_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_14_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_18_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_18_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_14_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[temp3],       %[temp2]        \n\t"
+        "msub     $ac1,                 %[load1],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load2],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_8_64]   \n\t"
+        "extp     %[step2_9],           $ac1,           31              \n\t"
+        "extp     %[step2_14],          $ac3,           31              \n\t"
+        "add      %[step2_8],           %[temp0],       %[temp1]        \n\t"
+        "add      %[step2_15],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step2_8] "=r"(step2_8),
+          [step2_9] "=r"(step2_9), [step2_14] "=r"(step2_14),
+          [step2_15] "=r"(step2_15)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_30_64] "r"(cospi_30_64), [cospi_2_64] "r"(cospi_2_64),
+          [cospi_14_64] "r"(cospi_14_64), [cospi_18_64] "r"(cospi_18_64),
+          [cospi_8_64] "r"(cospi_8_64), [cospi_24_64] "r"(cospi_24_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             20(%[input])                    \n\t"
+        "lh       %[load2],             44(%[input])                    \n\t"
+        "lh       %[load3],             52(%[input])                    \n\t"
+        "lh       %[load4],             12(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_22_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_10_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_10_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_22_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_6_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_26_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_26_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_6_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+        "msub     $ac1,                 %[load1],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_8_64]   \n\t"
+        "extp     %[step2_10],          $ac1,           31              \n\t"
+        "extp     %[step2_13],          $ac3,           31              \n\t"
+        "add      %[step2_11],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step2_12],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step2_10] "=r"(step2_10),
+          [step2_11] "=r"(step2_11), [step2_12] "=r"(step2_12),
+          [step2_13] "=r"(step2_13)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_22_64] "r"(cospi_22_64), [cospi_10_64] "r"(cospi_10_64),
+          [cospi_6_64] "r"(cospi_6_64), [cospi_26_64] "r"(cospi_26_64),
+          [cospi_8_64] "r"(cospi_8_64), [cospi_24_64] "r"(cospi_24_64));
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "sub      %[temp0],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[temp0],             %[temp0],       %[step2_9]      \n\t"
+        "add      %[temp0],             %[temp0],       %[step2_10]     \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "sub      %[temp1],             %[step2_14],    %[step2_13]     \n\t"
+        "add      %[temp1],             %[temp1],       %[step2_9]      \n\t"
+        "sub      %[temp1],             %[temp1],       %[step2_10]     \n\t"
+        "madd     $ac1,                 %[temp1],       %[cospi_16_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "sub      %[temp0],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[temp0],             %[temp0],       %[step2_8]      \n\t"
+        "add      %[temp0],             %[temp0],       %[step2_11]     \n\t"
+        "madd     $ac2,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "sub      %[temp1],             %[step2_15],    %[step2_12]     \n\t"
+        "add      %[temp1],             %[temp1],       %[step2_8]      \n\t"
+        "sub      %[temp1],             %[temp1],       %[step2_11]     \n\t"
+        "madd     $ac3,                 %[temp1],       %[cospi_16_64]  \n\t"
+
+        "add      %[step3_8],           %[step2_8],     %[step2_11]     \n\t"
+        "add      %[step3_9],           %[step2_9],     %[step2_10]     \n\t"
+        "add      %[step3_14],          %[step2_13],    %[step2_14]     \n\t"
+        "add      %[step3_15],          %[step2_12],    %[step2_15]     \n\t"
+        "extp     %[step3_10],          $ac0,           31              \n\t"
+        "extp     %[step3_13],          $ac1,           31              \n\t"
+        "extp     %[step3_11],          $ac2,           31              \n\t"
+        "extp     %[step3_12],          $ac3,           31              \n\t"
+
+        : [temp0] "=&r"(temp0), [temp1] "=&r"(temp1), [step3_8] "=r"(step3_8),
+          [step3_9] "=r"(step3_9), [step3_10] "=r"(step3_10),
+          [step3_11] "=r"(step3_11), [step3_12] "=r"(step3_12),
+          [step3_13] "=r"(step3_13), [step3_14] "=r"(step3_14),
+          [step3_15] "=r"(step3_15)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_8] "r"(step2_8),
+          [step2_9] "r"(step2_9), [step2_10] "r"(step2_10),
+          [step2_11] "r"(step2_11), [step2_12] "r"(step2_12),
+          [step2_13] "r"(step2_13), [step2_14] "r"(step2_14),
+          [step2_15] "r"(step2_15), [cospi_16_64] "r"(cospi_16_64));
+
+    step2_18 = step1_17 - step1_18;
+    step2_29 = step1_30 - step1_29;
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_18],    %[cospi_8_64]   \n\t"
+        "madd     $ac0,                 %[step2_29],    %[cospi_24_64]  \n\t"
+        "extp     %[step3_18],          $ac0,           31              \n\t"
+
+        : [step3_18] "=r"(step3_18)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_18] "r"(step2_18),
+          [step2_29] "r"(step2_29), [cospi_24_64] "r"(cospi_24_64),
+          [cospi_8_64] "r"(cospi_8_64));
+
+    temp21 = step2_18 * cospi_24_64 + step2_29 * cospi_8_64;
+    step3_29 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step2_19 = step1_16 - step1_19;
+    step2_28 = step1_31 - step1_28;
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_19],    %[cospi_8_64]   \n\t"
+        "madd     $ac0,                 %[step2_28],    %[cospi_24_64]  \n\t"
+        "extp     %[step3_19],          $ac0,           31              \n\t"
+
+        : [step3_19] "=r"(step3_19)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_19] "r"(step2_19),
+          [step2_28] "r"(step2_28), [cospi_24_64] "r"(cospi_24_64),
+          [cospi_8_64] "r"(cospi_8_64));
+
+    temp21 = step2_19 * cospi_24_64 + step2_28 * cospi_8_64;
+    step3_28 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step3_16 = step1_16 + step1_19;
+    step3_17 = step1_17 + step1_18;
+    step3_30 = step1_29 + step1_30;
+    step3_31 = step1_28 + step1_31;
+
+    step2_20 = step1_23 - step1_20;
+    step2_27 = step1_24 - step1_27;
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_20],    %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[step2_27],    %[cospi_8_64]   \n\t"
+        "extp     %[step3_20],          $ac0,           31              \n\t"
+
+        : [step3_20] "=r"(step3_20)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_20] "r"(step2_20),
+          [step2_27] "r"(step2_27), [cospi_24_64] "r"(cospi_24_64),
+          [cospi_8_64] "r"(cospi_8_64));
+
+    temp21 = -step2_20 * cospi_8_64 + step2_27 * cospi_24_64;
+    step3_27 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step2_21 = step1_22 - step1_21;
+    step2_26 = step1_25 - step1_26;
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "msub     $ac1,                 %[step2_21],    %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[step2_26],    %[cospi_8_64]   \n\t"
+        "extp     %[step3_21],          $ac1,           31              \n\t"
+
+        : [step3_21] "=r"(step3_21)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_21] "r"(step2_21),
+          [step2_26] "r"(step2_26), [cospi_24_64] "r"(cospi_24_64),
+          [cospi_8_64] "r"(cospi_8_64));
+
+    temp21 = -step2_21 * cospi_8_64 + step2_26 * cospi_24_64;
+    step3_26 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step3_22 = step1_21 + step1_22;
+    step3_23 = step1_20 + step1_23;
+    step3_24 = step1_24 + step1_27;
+    step3_25 = step1_25 + step1_26;
+
+    step2_16 = step3_16 + step3_23;
+    step2_17 = step3_17 + step3_22;
+    step2_18 = step3_18 + step3_21;
+    step2_19 = step3_19 + step3_20;
+    step2_20 = step3_19 - step3_20;
+    step2_21 = step3_18 - step3_21;
+    step2_22 = step3_17 - step3_22;
+    step2_23 = step3_16 - step3_23;
+
+    step2_24 = step3_31 - step3_24;
+    step2_25 = step3_30 - step3_25;
+    step2_26 = step3_29 - step3_26;
+    step2_27 = step3_28 - step3_27;
+    step2_28 = step3_28 + step3_27;
+    step2_29 = step3_29 + step3_26;
+    step2_30 = step3_30 + step3_25;
+    step2_31 = step3_31 + step3_24;
+
+    __asm__ __volatile__(
+        "lh       %[load1],             0(%[input])                     \n\t"
+        "lh       %[load2],             32(%[input])                    \n\t"
+        "lh       %[load3],             16(%[input])                    \n\t"
+        "lh       %[load4],             48(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "add      %[result1],           %[load1],       %[load2]        \n\t"
+        "sub      %[result2],           %[load1],       %[load2]        \n\t"
+        "madd     $ac1,                 %[result1],     %[cospi_16_64]  \n\t"
+        "madd     $ac2,                 %[result2],     %[cospi_16_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "madd     $ac3,                 %[load3],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load4],       %[cospi_8_64]   \n\t"
+        "extp     %[temp2],             $ac3,           31              \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_24_64]  \n\t"
+        "extp     %[temp3],             $ac1,           31              \n\t"
+        "add      %[step1_0],           %[temp0],       %[temp3]        \n\t"
+        "add      %[step1_1],           %[temp1],       %[temp2]        \n\t"
+        "sub      %[step1_2],           %[temp1],       %[temp2]        \n\t"
+        "sub      %[step1_3],           %[temp0],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [result1] "=&r"(result1),
+          [result2] "=&r"(result2), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_0] "=r"(step1_0),
+          [step1_1] "=r"(step1_1), [step1_2] "=r"(step1_2),
+          [step1_3] "=r"(step1_3)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_24_64] "r"(cospi_24_64), [cospi_8_64] "r"(cospi_8_64),
+          [cospi_16_64] "r"(cospi_16_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             8(%[input])                     \n\t"
+        "lh       %[load2],             56(%[input])                    \n\t"
+        "lh       %[load3],             40(%[input])                    \n\t"
+        "lh       %[load4],             24(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_12_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_20_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_20_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_12_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp3],       %[temp2]        \n\t"
+        "sub      %[load1],             %[load1],       %[temp0]        \n\t"
+        "add      %[load1],             %[load1],       %[temp1]        \n\t"
+        "sub      %[load2],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[load2],       %[temp2]        \n\t"
+        "add      %[load2],             %[load2],       %[temp3]        \n\t"
+        "madd     $ac1,                 %[load1],       %[cospi_16_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_5],           $ac1,           31              \n\t"
+        "extp     %[step1_6],           $ac3,           31              \n\t"
+        "add      %[step1_4],           %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_7],           %[temp3],       %[temp2]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_4] "=r"(step1_4),
+          [step1_5] "=r"(step1_5), [step1_6] "=r"(step1_6),
+          [step1_7] "=r"(step1_7)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_20_64] "r"(cospi_20_64), [cospi_12_64] "r"(cospi_12_64),
+          [cospi_4_64] "r"(cospi_4_64), [cospi_28_64] "r"(cospi_28_64),
+          [cospi_16_64] "r"(cospi_16_64));
+
+    step2_0 = step1_0 + step1_7;
+    step2_1 = step1_1 + step1_6;
+    step2_2 = step1_2 + step1_5;
+    step2_3 = step1_3 + step1_4;
+    step2_4 = step1_3 - step1_4;
+    step2_5 = step1_2 - step1_5;
+    step2_6 = step1_1 - step1_6;
+    step2_7 = step1_0 - step1_7;
+
+    // stage 7
+    step1_0 = step2_0 + step3_15;
+    step1_1 = step2_1 + step3_14;
+    step1_2 = step2_2 + step3_13;
+    step1_3 = step2_3 + step3_12;
+    step1_4 = step2_4 + step3_11;
+    step1_5 = step2_5 + step3_10;
+    step1_6 = step2_6 + step3_9;
+    step1_7 = step2_7 + step3_8;
+    step1_8 = step2_7 - step3_8;
+    step1_9 = step2_6 - step3_9;
+    step1_10 = step2_5 - step3_10;
+    step1_11 = step2_4 - step3_11;
+    step1_12 = step2_3 - step3_12;
+    step1_13 = step2_2 - step3_13;
+    step1_14 = step2_1 - step3_14;
+    step1_15 = step2_0 - step3_15;
+
+    __asm__ __volatile__(
+        "sub      %[temp0],             %[step2_27],    %[step2_20]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_20],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r"(temp0), [step1_20] "=r"(step1_20)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_20] "r"(step2_20),
+          [step2_27] "r"(step2_27), [cospi_16_64] "r"(cospi_16_64));
+
+    temp21 = (step2_20 + step2_27) * cospi_16_64;
+    step1_27 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__(
+        "sub      %[temp0],             %[step2_26],    %[step2_21]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_21],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r"(temp0), [step1_21] "=r"(step1_21)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_26] "r"(step2_26),
+          [step2_21] "r"(step2_21), [cospi_16_64] "r"(cospi_16_64));
+
+    temp21 = (step2_21 + step2_26) * cospi_16_64;
+    step1_26 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__(
+        "sub      %[temp0],             %[step2_25],    %[step2_22]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_22],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r"(temp0), [step1_22] "=r"(step1_22)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_25] "r"(step2_25),
+          [step2_22] "r"(step2_22), [cospi_16_64] "r"(cospi_16_64));
+
+    temp21 = (step2_22 + step2_25) * cospi_16_64;
+    step1_25 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__(
+        "sub      %[temp0],             %[step2_24],    %[step2_23]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_23],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r"(temp0), [step1_23] "=r"(step1_23)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_24] "r"(step2_24),
+          [step2_23] "r"(step2_23), [cospi_16_64] "r"(cospi_16_64));
+
+    temp21 = (step2_23 + step2_24) * cospi_16_64;
+    step1_24 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__(
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_0],         %[step2_31]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_1],         %[step2_30]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_2],         %[step2_29]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_3],         %[step2_28]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        : [temp0] "=&r"(temp0), [temp1] "=&r"(temp1), [temp2] "=&r"(temp2),
+          [temp3] "=&r"(temp3), [dest_pix] "+r"(dest_pix)
+        : [cm] "r"(cm), [dest_stride] "r"(dest_stride), [step1_0] "r"(step1_0),
+          [step1_1] "r"(step1_1), [step1_2] "r"(step1_2),
+          [step1_3] "r"(step1_3), [step2_28] "r"(step2_28),
+          [step2_29] "r"(step2_29), [step2_30] "r"(step2_30),
+          [step2_31] "r"(step2_31));
+
+    step3_12 = ROUND_POWER_OF_TWO((step1_3 - step2_28), 6);
+    step3_13 = ROUND_POWER_OF_TWO((step1_2 - step2_29), 6);
+    step3_14 = ROUND_POWER_OF_TWO((step1_1 - step2_30), 6);
+    step3_15 = ROUND_POWER_OF_TWO((step1_0 - step2_31), 6);
+
+    __asm__ __volatile__(
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_15]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_14]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_13]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_12]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        : [temp0] "=&r"(temp0), [temp1] "=&r"(temp1), [temp2] "=&r"(temp2),
+          [temp3] "=&r"(temp3), [dest_pix1] "+r"(dest_pix1)
+        : [cm] "r"(cm), [dest_stride] "r"(dest_stride),
+          [step3_12] "r"(step3_12), [step3_13] "r"(step3_13),
+          [step3_14] "r"(step3_14), [step3_15] "r"(step3_15));
+
+    __asm__ __volatile__(
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_4],         %[step1_27]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_5],         %[step1_26]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_6],         %[step1_25]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_7],         %[step1_24]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        : [temp0] "=&r"(temp0), [temp1] "=&r"(temp1), [temp2] "=&r"(temp2),
+          [temp3] "=&r"(temp3), [dest_pix] "+r"(dest_pix)
+        : [cm] "r"(cm), [dest_stride] "r"(dest_stride), [step1_4] "r"(step1_4),
+          [step1_5] "r"(step1_5), [step1_6] "r"(step1_6),
+          [step1_7] "r"(step1_7), [step1_24] "r"(step1_24),
+          [step1_25] "r"(step1_25), [step1_26] "r"(step1_26),
+          [step1_27] "r"(step1_27));
+
+    step3_12 = ROUND_POWER_OF_TWO((step1_7 - step1_24), 6);
+    step3_13 = ROUND_POWER_OF_TWO((step1_6 - step1_25), 6);
+    step3_14 = ROUND_POWER_OF_TWO((step1_5 - step1_26), 6);
+    step3_15 = ROUND_POWER_OF_TWO((step1_4 - step1_27), 6);
+
+    __asm__ __volatile__(
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_15]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_14]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_13]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_12]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        : [temp0] "=&r"(temp0), [temp1] "=&r"(temp1), [temp2] "=&r"(temp2),
+          [temp3] "=&r"(temp3), [dest_pix1] "+r"(dest_pix1)
+        : [cm] "r"(cm), [dest_stride] "r"(dest_stride),
+          [step3_12] "r"(step3_12), [step3_13] "r"(step3_13),
+          [step3_14] "r"(step3_14), [step3_15] "r"(step3_15));
+
+    __asm__ __volatile__(
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_8],         %[step1_23]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_9],         %[step1_22]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_10],        %[step1_21]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_11],        %[step1_20]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        : [temp0] "=&r"(temp0), [temp1] "=&r"(temp1), [temp2] "=&r"(temp2),
+          [temp3] "=&r"(temp3), [dest_pix] "+r"(dest_pix)
+        : [cm] "r"(cm), [dest_stride] "r"(dest_stride), [step1_8] "r"(step1_8),
+          [step1_9] "r"(step1_9), [step1_10] "r"(step1_10),
+          [step1_11] "r"(step1_11), [step1_20] "r"(step1_20),
+          [step1_21] "r"(step1_21), [step1_22] "r"(step1_22),
+          [step1_23] "r"(step1_23));
+
+    step3_12 = ROUND_POWER_OF_TWO((step1_11 - step1_20), 6);
+    step3_13 = ROUND_POWER_OF_TWO((step1_10 - step1_21), 6);
+    step3_14 = ROUND_POWER_OF_TWO((step1_9 - step1_22), 6);
+    step3_15 = ROUND_POWER_OF_TWO((step1_8 - step1_23), 6);
+
+    __asm__ __volatile__(
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_15]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_14]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_13]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_12]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        : [temp0] "=&r"(temp0), [temp1] "=&r"(temp1), [temp2] "=&r"(temp2),
+          [temp3] "=&r"(temp3), [dest_pix1] "+r"(dest_pix1)
+        : [cm] "r"(cm), [dest_stride] "r"(dest_stride),
+          [step3_12] "r"(step3_12), [step3_13] "r"(step3_13),
+          [step3_14] "r"(step3_14), [step3_15] "r"(step3_15));
+
+    __asm__ __volatile__(
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_12],        %[step2_19]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_13],        %[step2_18]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix])                      \n\t"
+        "add      %[temp0],         %[step1_14],        %[step2_17]     \n\t"
+        "addi     %[temp0],         %[temp0],           32              \n\t"
+        "sra      %[temp0],         %[temp0],           6               \n\t"
+        "add      %[temp2],         %[temp2],           %[temp0]        \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "add      %[temp1],         %[step1_15],        %[step2_16]     \n\t"
+        "sb       %[temp0],         0(%[dest_pix])                      \n\t"
+        "addu     %[dest_pix],      %[dest_pix],        %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix])                      \n\t"
+        "addi     %[temp1],         %[temp1],           32              \n\t"
+        "sra      %[temp1],         %[temp1],           6               \n\t"
+        "add      %[temp3],         %[temp3],           %[temp1]        \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix])                      \n\t"
+
+        : [temp0] "=&r"(temp0), [temp1] "=&r"(temp1), [temp2] "=&r"(temp2),
+          [temp3] "=&r"(temp3), [dest_pix] "+r"(dest_pix)
+        : [cm] "r"(cm), [dest_stride] "r"(dest_stride),
+          [step1_12] "r"(step1_12), [step1_13] "r"(step1_13),
+          [step1_14] "r"(step1_14), [step1_15] "r"(step1_15),
+          [step2_16] "r"(step2_16), [step2_17] "r"(step2_17),
+          [step2_18] "r"(step2_18), [step2_19] "r"(step2_19));
+
+    step3_12 = ROUND_POWER_OF_TWO((step1_15 - step2_16), 6);
+    step3_13 = ROUND_POWER_OF_TWO((step1_14 - step2_17), 6);
+    step3_14 = ROUND_POWER_OF_TWO((step1_13 - step2_18), 6);
+    step3_15 = ROUND_POWER_OF_TWO((step1_12 - step2_19), 6);
+
+    __asm__ __volatile__(
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_15]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_14]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+
+        "lbu      %[temp2],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp2],         %[temp2],           %[step3_13]     \n\t"
+        "lbux     %[temp0],         %[temp2](%[cm])                     \n\t"
+        "sb       %[temp0],         0(%[dest_pix1])                     \n\t"
+        "subu     %[dest_pix1],     %[dest_pix1],       %[dest_stride]  \n\t"
+        "lbu      %[temp3],         0(%[dest_pix1])                     \n\t"
+        "add      %[temp3],         %[temp3],           %[step3_12]     \n\t"
+        "lbux     %[temp1],         %[temp3](%[cm])                     \n\t"
+        "sb       %[temp1],         0(%[dest_pix1])                     \n\t"
+
+        : [temp0] "=&r"(temp0), [temp1] "=&r"(temp1), [temp2] "=&r"(temp2),
+          [temp3] "=&r"(temp3), [dest_pix1] "+r"(dest_pix1)
+        : [cm] "r"(cm), [dest_stride] "r"(dest_stride),
+          [step3_12] "r"(step3_12), [step3_13] "r"(step3_13),
+          [step3_14] "r"(step3_14), [step3_15] "r"(step3_15));
+
+    input += 32;
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/itrans32_dspr2.c b/third_party/aom/aom_dsp/mips/itrans32_dspr2.c
new file mode 100644
index 0000000000..fa77032176
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/itrans32_dspr2.c
@@ -0,0 +1,1030 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_config.h"
+#include "aom_dsp/mips/inv_txfm_dspr2.h"
+#include "aom_dsp/txfm_common.h"
+
+#if HAVE_DSPR2
+static void idct32_rows_dspr2(const int16_t *input, int16_t *output,
+                              uint32_t no_rows) {
+  int16_t step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6;
+  int16_t step1_7, step1_8, step1_9, step1_10, step1_11, step1_12, step1_13;
+  int16_t step1_14, step1_15, step1_16, step1_17, step1_18, step1_19, step1_20;
+  int16_t step1_21, step1_22, step1_23, step1_24, step1_25, step1_26, step1_27;
+  int16_t step1_28, step1_29, step1_30, step1_31;
+  int16_t step2_0, step2_1, step2_2, step2_3, step2_4, step2_5, step2_6;
+  int16_t step2_7, step2_8, step2_9, step2_10, step2_11, step2_12, step2_13;
+  int16_t step2_14, step2_15, step2_16, step2_17, step2_18, step2_19, step2_20;
+  int16_t step2_21, step2_22, step2_23, step2_24, step2_25, step2_26, step2_27;
+  int16_t step2_28, step2_29, step2_30, step2_31;
+  int16_t step3_8, step3_9, step3_10, step3_11, step3_12, step3_13, step3_14;
+  int16_t step3_15, step3_16, step3_17, step3_18, step3_19, step3_20, step3_21;
+  int16_t step3_22, step3_23, step3_24, step3_25, step3_26, step3_27, step3_28;
+  int16_t step3_29, step3_30, step3_31;
+  int temp0, temp1, temp2, temp3;
+  int load1, load2, load3, load4;
+  int result1, result2;
+  int temp21;
+  int i;
+  const int const_2_power_13 = 8192;
+  const int32_t *input_int;
+
+  for (i = no_rows; i--;) {
+    input_int = (const int32_t *)input;
+
+    if (!(input_int[0] | input_int[1] | input_int[2] | input_int[3] |
+          input_int[4] | input_int[5] | input_int[6] | input_int[7] |
+          input_int[8] | input_int[9] | input_int[10] | input_int[11] |
+          input_int[12] | input_int[13] | input_int[14] | input_int[15])) {
+      input += 32;
+
+      __asm__ __volatile__(
+          "sh     $zero,     0(%[output])     \n\t"
+          "sh     $zero,    64(%[output])     \n\t"
+          "sh     $zero,   128(%[output])     \n\t"
+          "sh     $zero,   192(%[output])     \n\t"
+          "sh     $zero,   256(%[output])     \n\t"
+          "sh     $zero,   320(%[output])     \n\t"
+          "sh     $zero,   384(%[output])     \n\t"
+          "sh     $zero,   448(%[output])     \n\t"
+          "sh     $zero,   512(%[output])     \n\t"
+          "sh     $zero,   576(%[output])     \n\t"
+          "sh     $zero,   640(%[output])     \n\t"
+          "sh     $zero,   704(%[output])     \n\t"
+          "sh     $zero,   768(%[output])     \n\t"
+          "sh     $zero,   832(%[output])     \n\t"
+          "sh     $zero,   896(%[output])     \n\t"
+          "sh     $zero,   960(%[output])     \n\t"
+          "sh     $zero,  1024(%[output])     \n\t"
+          "sh     $zero,  1088(%[output])     \n\t"
+          "sh     $zero,  1152(%[output])     \n\t"
+          "sh     $zero,  1216(%[output])     \n\t"
+          "sh     $zero,  1280(%[output])     \n\t"
+          "sh     $zero,  1344(%[output])     \n\t"
+          "sh     $zero,  1408(%[output])     \n\t"
+          "sh     $zero,  1472(%[output])     \n\t"
+          "sh     $zero,  1536(%[output])     \n\t"
+          "sh     $zero,  1600(%[output])     \n\t"
+          "sh     $zero,  1664(%[output])     \n\t"
+          "sh     $zero,  1728(%[output])     \n\t"
+          "sh     $zero,  1792(%[output])     \n\t"
+          "sh     $zero,  1856(%[output])     \n\t"
+          "sh     $zero,  1920(%[output])     \n\t"
+          "sh     $zero,  1984(%[output])     \n\t"
+
+          :
+          : [output] "r"(output));
+
+      output += 1;
+
+      continue;
+    }
+
+    /* prefetch row */
+    prefetch_load((const uint8_t *)(input + 32));
+    prefetch_load((const uint8_t *)(input + 48));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             2(%[input])                     \n\t"
+        "lh       %[load2],             62(%[input])                    \n\t"
+        "lh       %[load3],             34(%[input])                    \n\t"
+        "lh       %[load4],             30(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_31_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_1_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_1_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_31_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_15_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_17_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_17_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_15_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp3],       %[temp2]        \n\t"
+        "sub      %[load2],             %[temp0],       %[temp1]        \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+
+        "extp     %[step1_17],          $ac1,           31              \n\t"
+        "extp     %[step1_30],          $ac3,           31              \n\t"
+        "add      %[step1_16],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_31],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_16] "=r"(step1_16),
+          [step1_17] "=r"(step1_17), [step1_30] "=r"(step1_30),
+          [step1_31] "=r"(step1_31)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_31_64] "r"(cospi_31_64), [cospi_1_64] "r"(cospi_1_64),
+          [cospi_4_64] "r"(cospi_4_64), [cospi_17_64] "r"(cospi_17_64),
+          [cospi_15_64] "r"(cospi_15_64), [cospi_28_64] "r"(cospi_28_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             18(%[input])                    \n\t"
+        "lh       %[load2],             46(%[input])                    \n\t"
+        "lh       %[load3],             50(%[input])                    \n\t"
+        "lh       %[load4],             14(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_23_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_9_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_9_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_23_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_7_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_25_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_25_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_7_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+
+        "extp     %[step1_18],          $ac1,           31              \n\t"
+        "extp     %[step1_29],          $ac3,           31              \n\t"
+        "add      %[step1_19],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_28],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_18] "=r"(step1_18),
+          [step1_19] "=r"(step1_19), [step1_28] "=r"(step1_28),
+          [step1_29] "=r"(step1_29)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_23_64] "r"(cospi_23_64), [cospi_9_64] "r"(cospi_9_64),
+          [cospi_4_64] "r"(cospi_4_64), [cospi_7_64] "r"(cospi_7_64),
+          [cospi_25_64] "r"(cospi_25_64), [cospi_28_64] "r"(cospi_28_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             10(%[input])                    \n\t"
+        "lh       %[load2],             54(%[input])                    \n\t"
+        "lh       %[load3],             42(%[input])                    \n\t"
+        "lh       %[load4],             22(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_27_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_5_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_5_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_27_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_11_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_21_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_21_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_11_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[temp3],       %[temp2]        \n\t"
+
+        "madd     $ac1,                 %[load2],       %[cospi_12_64]  \n\t"
+        "msub     $ac1,                 %[load1],       %[cospi_20_64]  \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_12_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_20_64]  \n\t"
+
+        "extp     %[step1_21],          $ac1,           31              \n\t"
+        "extp     %[step1_26],          $ac3,           31              \n\t"
+        "add      %[step1_20],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_27],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_20] "=r"(step1_20),
+          [step1_21] "=r"(step1_21), [step1_26] "=r"(step1_26),
+          [step1_27] "=r"(step1_27)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_27_64] "r"(cospi_27_64), [cospi_5_64] "r"(cospi_5_64),
+          [cospi_11_64] "r"(cospi_11_64), [cospi_21_64] "r"(cospi_21_64),
+          [cospi_12_64] "r"(cospi_12_64), [cospi_20_64] "r"(cospi_20_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             26(%[input])                    \n\t"
+        "lh       %[load2],             38(%[input])                    \n\t"
+        "lh       %[load3],             58(%[input])                    \n\t"
+        "lh       %[load4],              6(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_19_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_13_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_13_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_19_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_3_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_29_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_29_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_3_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_12_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_20_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_20_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_12_64]  \n\t"
+
+        "extp     %[step1_22],          $ac1,           31              \n\t"
+        "extp     %[step1_25],          $ac3,           31              \n\t"
+        "add      %[step1_23],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_24],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_22] "=r"(step1_22),
+          [step1_23] "=r"(step1_23), [step1_24] "=r"(step1_24),
+          [step1_25] "=r"(step1_25)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_19_64] "r"(cospi_19_64), [cospi_13_64] "r"(cospi_13_64),
+          [cospi_3_64] "r"(cospi_3_64), [cospi_29_64] "r"(cospi_29_64),
+          [cospi_12_64] "r"(cospi_12_64), [cospi_20_64] "r"(cospi_20_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],              4(%[input])                    \n\t"
+        "lh       %[load2],             60(%[input])                    \n\t"
+        "lh       %[load3],             36(%[input])                    \n\t"
+        "lh       %[load4],             28(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_30_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_2_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_2_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_30_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_14_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_18_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_18_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_14_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[temp3],       %[temp2]        \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load2],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load1],       %[cospi_24_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_9],           $ac1,           31              \n\t"
+        "extp     %[step2_14],          $ac3,           31              \n\t"
+        "add      %[step2_8],           %[temp0],       %[temp1]        \n\t"
+        "add      %[step2_15],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step2_8] "=r"(step2_8),
+          [step2_9] "=r"(step2_9), [step2_14] "=r"(step2_14),
+          [step2_15] "=r"(step2_15)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_30_64] "r"(cospi_30_64), [cospi_2_64] "r"(cospi_2_64),
+          [cospi_14_64] "r"(cospi_14_64), [cospi_18_64] "r"(cospi_18_64),
+          [cospi_8_64] "r"(cospi_8_64), [cospi_24_64] "r"(cospi_24_64));
+
+    __asm__ __volatile__(
+        "lh       %[load1],             20(%[input])                    \n\t"
+        "lh       %[load2],             44(%[input])                    \n\t"
+        "lh       %[load3],             52(%[input])                    \n\t"
+        "lh       %[load4],             12(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_22_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_10_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_10_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_22_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_6_64]   \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_26_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_26_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_6_64]   \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp1],       %[temp0]        \n\t"
+        "sub      %[load2],             %[temp2],       %[temp3]        \n\t"
+
+        "msub     $ac1,                 %[load1],       %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_8_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load1],       %[cospi_8_64]   \n\t"
+
+        "extp     %[step2_10],          $ac1,           31              \n\t"
+        "extp     %[step2_13],          $ac3,           31              \n\t"
+        "add      %[step2_11],          %[temp0],       %[temp1]        \n\t"
+        "add      %[step2_12],          %[temp2],       %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step2_10] "=r"(step2_10),
+          [step2_11] "=r"(step2_11), [step2_12] "=r"(step2_12),
+          [step2_13] "=r"(step2_13)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_22_64] "r"(cospi_22_64), [cospi_10_64] "r"(cospi_10_64),
+          [cospi_6_64] "r"(cospi_6_64), [cospi_26_64] "r"(cospi_26_64),
+          [cospi_8_64] "r"(cospi_8_64), [cospi_24_64] "r"(cospi_24_64));
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "sub      %[temp0],             %[step2_14],    %[step2_13]     \n\t"
+        "sub      %[temp0],             %[temp0],       %[step2_9]      \n\t"
+        "add      %[temp0],             %[temp0],       %[step2_10]     \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "sub      %[temp1],             %[step2_14],    %[step2_13]     \n\t"
+        "add      %[temp1],             %[temp1],       %[step2_9]      \n\t"
+        "sub      %[temp1],             %[temp1],       %[step2_10]     \n\t"
+        "madd     $ac1,                 %[temp1],       %[cospi_16_64]  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "sub      %[temp0],             %[step2_15],    %[step2_12]     \n\t"
+        "sub      %[temp0],             %[temp0],       %[step2_8]      \n\t"
+        "add      %[temp0],             %[temp0],       %[step2_11]     \n\t"
+        "madd     $ac2,                 %[temp0],       %[cospi_16_64]  \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "sub      %[temp1],             %[step2_15],    %[step2_12]     \n\t"
+        "add      %[temp1],             %[temp1],       %[step2_8]      \n\t"
+        "sub      %[temp1],             %[temp1],       %[step2_11]     \n\t"
+        "madd     $ac3,                 %[temp1],       %[cospi_16_64]  \n\t"
+
+        "add      %[step3_8],           %[step2_8],     %[step2_11]     \n\t"
+        "add      %[step3_9],           %[step2_9],     %[step2_10]     \n\t"
+        "add      %[step3_14],          %[step2_13],    %[step2_14]     \n\t"
+        "add      %[step3_15],          %[step2_12],    %[step2_15]     \n\t"
+
+        "extp     %[step3_10],          $ac0,           31              \n\t"
+        "extp     %[step3_13],          $ac1,           31              \n\t"
+        "extp     %[step3_11],          $ac2,           31              \n\t"
+        "extp     %[step3_12],          $ac3,           31              \n\t"
+
+        : [temp0] "=&r"(temp0), [temp1] "=&r"(temp1), [step3_8] "=r"(step3_8),
+          [step3_9] "=r"(step3_9), [step3_10] "=r"(step3_10),
+          [step3_11] "=r"(step3_11), [step3_12] "=r"(step3_12),
+          [step3_13] "=r"(step3_13), [step3_14] "=r"(step3_14),
+          [step3_15] "=r"(step3_15)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_8] "r"(step2_8),
+          [step2_9] "r"(step2_9), [step2_10] "r"(step2_10),
+          [step2_11] "r"(step2_11), [step2_12] "r"(step2_12),
+          [step2_13] "r"(step2_13), [step2_14] "r"(step2_14),
+          [step2_15] "r"(step2_15), [cospi_16_64] "r"(cospi_16_64));
+
+    step2_18 = step1_17 - step1_18;
+    step2_29 = step1_30 - step1_29;
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_18],    %[cospi_8_64]   \n\t"
+        "madd     $ac0,                 %[step2_29],    %[cospi_24_64]  \n\t"
+        "extp     %[step3_18],          $ac0,           31              \n\t"
+
+        : [step3_18] "=r"(step3_18)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_18] "r"(step2_18),
+          [step2_29] "r"(step2_29), [cospi_24_64] "r"(cospi_24_64),
+          [cospi_8_64] "r"(cospi_8_64));
+
+    temp21 = step2_18 * cospi_24_64 + step2_29 * cospi_8_64;
+    step3_29 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step2_19 = step1_16 - step1_19;
+    step2_28 = step1_31 - step1_28;
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_19],    %[cospi_8_64]   \n\t"
+        "madd     $ac0,                 %[step2_28],    %[cospi_24_64]  \n\t"
+        "extp     %[step3_19],          $ac0,           31              \n\t"
+
+        : [step3_19] "=r"(step3_19)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_19] "r"(step2_19),
+          [step2_28] "r"(step2_28), [cospi_24_64] "r"(cospi_24_64),
+          [cospi_8_64] "r"(cospi_8_64));
+
+    temp21 = step2_19 * cospi_24_64 + step2_28 * cospi_8_64;
+    step3_28 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step3_16 = step1_16 + step1_19;
+    step3_17 = step1_17 + step1_18;
+    step3_30 = step1_29 + step1_30;
+    step3_31 = step1_28 + step1_31;
+
+    step2_20 = step1_23 - step1_20;
+    step2_27 = step1_24 - step1_27;
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "msub     $ac0,                 %[step2_20],    %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[step2_27],    %[cospi_8_64]   \n\t"
+        "extp     %[step3_20],          $ac0,           31              \n\t"
+
+        : [step3_20] "=r"(step3_20)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_20] "r"(step2_20),
+          [step2_27] "r"(step2_27), [cospi_24_64] "r"(cospi_24_64),
+          [cospi_8_64] "r"(cospi_8_64));
+
+    temp21 = -step2_20 * cospi_8_64 + step2_27 * cospi_24_64;
+    step3_27 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step2_21 = step1_22 - step1_21;
+    step2_26 = step1_25 - step1_26;
+
+    __asm__ __volatile__(
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "msub     $ac1,                 %[step2_21],    %[cospi_24_64]  \n\t"
+        "msub     $ac1,                 %[step2_26],    %[cospi_8_64]   \n\t"
+        "extp     %[step3_21],          $ac1,           31              \n\t"
+
+        : [step3_21] "=r"(step3_21)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_21] "r"(step2_21),
+          [step2_26] "r"(step2_26), [cospi_24_64] "r"(cospi_24_64),
+          [cospi_8_64] "r"(cospi_8_64));
+
+    temp21 = -step2_21 * cospi_8_64 + step2_26 * cospi_24_64;
+    step3_26 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    step3_22 = step1_21 + step1_22;
+    step3_23 = step1_20 + step1_23;
+    step3_24 = step1_24 + step1_27;
+    step3_25 = step1_25 + step1_26;
+
+    step2_16 = step3_16 + step3_23;
+    step2_17 = step3_17 + step3_22;
+    step2_18 = step3_18 + step3_21;
+    step2_19 = step3_19 + step3_20;
+    step2_20 = step3_19 - step3_20;
+    step2_21 = step3_18 - step3_21;
+    step2_22 = step3_17 - step3_22;
+    step2_23 = step3_16 - step3_23;
+
+    step2_24 = step3_31 - step3_24;
+    step2_25 = step3_30 - step3_25;
+    step2_26 = step3_29 - step3_26;
+    step2_27 = step3_28 - step3_27;
+    step2_28 = step3_28 + step3_27;
+    step2_29 = step3_29 + step3_26;
+    step2_30 = step3_30 + step3_25;
+    step2_31 = step3_31 + step3_24;
+
+    __asm__ __volatile__(
+        "lh       %[load1],             0(%[input])                     \n\t"
+        "lh       %[load2],             32(%[input])                    \n\t"
+        "lh       %[load3],             16(%[input])                    \n\t"
+        "lh       %[load4],             48(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+        "add      %[result1],           %[load1],       %[load2]        \n\t"
+        "sub      %[result2],           %[load1],       %[load2]        \n\t"
+        "madd     $ac1,                 %[result1],     %[cospi_16_64]  \n\t"
+        "madd     $ac2,                 %[result2],     %[cospi_16_64]  \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+        "madd     $ac3,                 %[load3],       %[cospi_24_64]  \n\t"
+        "msub     $ac3,                 %[load4],       %[cospi_8_64]   \n\t"
+        "extp     %[temp2],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "madd     $ac1,                 %[load3],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_24_64]  \n\t"
+        "extp     %[temp3],             $ac1,           31              \n\t"
+
+        "add      %[step1_0],          %[temp0],        %[temp3]        \n\t"
+        "add      %[step1_1],          %[temp1],        %[temp2]        \n\t"
+        "sub      %[step1_2],          %[temp1],        %[temp2]        \n\t"
+        "sub      %[step1_3],          %[temp0],        %[temp3]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [result1] "=&r"(result1),
+          [result2] "=&r"(result2), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_0] "=r"(step1_0),
+          [step1_1] "=r"(step1_1), [step1_2] "=r"(step1_2),
+          [step1_3] "=r"(step1_3)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_16_64] "r"(cospi_16_64), [cospi_24_64] "r"(cospi_24_64),
+          [cospi_8_64] "r"(cospi_8_64)
+
+            );
+
+    __asm__ __volatile__(
+        "lh       %[load1],             8(%[input])                     \n\t"
+        "lh       %[load2],             56(%[input])                    \n\t"
+        "lh       %[load3],             40(%[input])                    \n\t"
+        "lh       %[load4],             24(%[input])                    \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_28_64]  \n\t"
+        "msub     $ac1,                 %[load2],       %[cospi_4_64]   \n\t"
+        "extp     %[temp0],             $ac1,           31              \n\t"
+
+        "madd     $ac3,                 %[load1],       %[cospi_4_64]   \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_28_64]  \n\t"
+        "extp     %[temp3],             $ac3,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac2                            \n\t"
+        "mthi     $zero,                $ac2                            \n\t"
+
+        "madd     $ac2,                 %[load3],       %[cospi_12_64]  \n\t"
+        "msub     $ac2,                 %[load4],       %[cospi_20_64]  \n\t"
+        "extp     %[temp1],             $ac2,           31              \n\t"
+
+        "madd     $ac1,                 %[load3],       %[cospi_20_64]  \n\t"
+        "madd     $ac1,                 %[load4],       %[cospi_12_64]  \n\t"
+        "extp     %[temp2],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac3                            \n\t"
+        "mthi     $zero,                $ac3                            \n\t"
+
+        "sub      %[load1],             %[temp3],       %[temp2]        \n\t"
+        "sub      %[load1],             %[load1],       %[temp0]        \n\t"
+        "add      %[load1],             %[load1],       %[temp1]        \n\t"
+
+        "sub      %[load2],             %[temp0],       %[temp1]        \n\t"
+        "sub      %[load2],             %[load2],       %[temp2]        \n\t"
+        "add      %[load2],             %[load2],       %[temp3]        \n\t"
+
+        "madd     $ac1,                 %[load1],       %[cospi_16_64]  \n\t"
+        "madd     $ac3,                 %[load2],       %[cospi_16_64]  \n\t"
+
+        "extp     %[step1_5],           $ac1,           31              \n\t"
+        "extp     %[step1_6],           $ac3,           31              \n\t"
+        "add      %[step1_4],           %[temp0],       %[temp1]        \n\t"
+        "add      %[step1_7],           %[temp3],       %[temp2]        \n\t"
+
+        : [load1] "=&r"(load1), [load2] "=&r"(load2), [load3] "=&r"(load3),
+          [load4] "=&r"(load4), [temp0] "=&r"(temp0), [temp1] "=&r"(temp1),
+          [temp2] "=&r"(temp2), [temp3] "=&r"(temp3), [step1_4] "=r"(step1_4),
+          [step1_5] "=r"(step1_5), [step1_6] "=r"(step1_6),
+          [step1_7] "=r"(step1_7)
+        : [const_2_power_13] "r"(const_2_power_13), [input] "r"(input),
+          [cospi_20_64] "r"(cospi_20_64), [cospi_12_64] "r"(cospi_12_64),
+          [cospi_4_64] "r"(cospi_4_64), [cospi_28_64] "r"(cospi_28_64),
+          [cospi_16_64] "r"(cospi_16_64));
+
+    step2_0 = step1_0 + step1_7;
+    step2_1 = step1_1 + step1_6;
+    step2_2 = step1_2 + step1_5;
+    step2_3 = step1_3 + step1_4;
+    step2_4 = step1_3 - step1_4;
+    step2_5 = step1_2 - step1_5;
+    step2_6 = step1_1 - step1_6;
+    step2_7 = step1_0 - step1_7;
+
+    step1_0 = step2_0 + step3_15;
+    step1_1 = step2_1 + step3_14;
+    step1_2 = step2_2 + step3_13;
+    step1_3 = step2_3 + step3_12;
+    step1_4 = step2_4 + step3_11;
+    step1_5 = step2_5 + step3_10;
+    step1_6 = step2_6 + step3_9;
+    step1_7 = step2_7 + step3_8;
+    step1_8 = step2_7 - step3_8;
+    step1_9 = step2_6 - step3_9;
+    step1_10 = step2_5 - step3_10;
+    step1_11 = step2_4 - step3_11;
+    step1_12 = step2_3 - step3_12;
+    step1_13 = step2_2 - step3_13;
+    step1_14 = step2_1 - step3_14;
+    step1_15 = step2_0 - step3_15;
+
+    __asm__ __volatile__(
+        "sub      %[temp0],             %[step2_27],    %[step2_20]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_20],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r"(temp0), [step1_20] "=r"(step1_20)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_20] "r"(step2_20),
+          [step2_27] "r"(step2_27), [cospi_16_64] "r"(cospi_16_64));
+
+    temp21 = (step2_20 + step2_27) * cospi_16_64;
+    step1_27 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__(
+        "sub      %[temp0],             %[step2_26],    %[step2_21]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_21],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r"(temp0), [step1_21] "=r"(step1_21)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_26] "r"(step2_26),
+          [step2_21] "r"(step2_21), [cospi_16_64] "r"(cospi_16_64));
+
+    temp21 = (step2_21 + step2_26) * cospi_16_64;
+    step1_26 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__(
+        "sub      %[temp0],             %[step2_25],    %[step2_22]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_22],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r"(temp0), [step1_22] "=r"(step1_22)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_25] "r"(step2_25),
+          [step2_22] "r"(step2_22), [cospi_16_64] "r"(cospi_16_64));
+
+    temp21 = (step2_22 + step2_25) * cospi_16_64;
+    step1_25 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    __asm__ __volatile__(
+        "sub      %[temp0],             %[step2_24],    %[step2_23]     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "madd     $ac0,                 %[temp0],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_23],          $ac0,           31              \n\t"
+
+        : [temp0] "=&r"(temp0), [step1_23] "=r"(step1_23)
+        : [const_2_power_13] "r"(const_2_power_13), [step2_24] "r"(step2_24),
+          [step2_23] "r"(step2_23), [cospi_16_64] "r"(cospi_16_64));
+
+    temp21 = (step2_23 + step2_24) * cospi_16_64;
+    step1_24 = (temp21 + DCT_CONST_ROUNDING) >> DCT_CONST_BITS;
+
+    // final stage
+    output[0 * 32] = step1_0 + step2_31;
+    output[1 * 32] = step1_1 + step2_30;
+    output[2 * 32] = step1_2 + step2_29;
+    output[3 * 32] = step1_3 + step2_28;
+    output[4 * 32] = step1_4 + step1_27;
+    output[5 * 32] = step1_5 + step1_26;
+    output[6 * 32] = step1_6 + step1_25;
+    output[7 * 32] = step1_7 + step1_24;
+    output[8 * 32] = step1_8 + step1_23;
+    output[9 * 32] = step1_9 + step1_22;
+    output[10 * 32] = step1_10 + step1_21;
+    output[11 * 32] = step1_11 + step1_20;
+    output[12 * 32] = step1_12 + step2_19;
+    output[13 * 32] = step1_13 + step2_18;
+    output[14 * 32] = step1_14 + step2_17;
+    output[15 * 32] = step1_15 + step2_16;
+    output[16 * 32] = step1_15 - step2_16;
+    output[17 * 32] = step1_14 - step2_17;
+    output[18 * 32] = step1_13 - step2_18;
+    output[19 * 32] = step1_12 - step2_19;
+    output[20 * 32] = step1_11 - step1_20;
+    output[21 * 32] = step1_10 - step1_21;
+    output[22 * 32] = step1_9 - step1_22;
+    output[23 * 32] = step1_8 - step1_23;
+    output[24 * 32] = step1_7 - step1_24;
+    output[25 * 32] = step1_6 - step1_25;
+    output[26 * 32] = step1_5 - step1_26;
+    output[27 * 32] = step1_4 - step1_27;
+    output[28 * 32] = step1_3 - step2_28;
+    output[29 * 32] = step1_2 - step2_29;
+    output[30 * 32] = step1_1 - step2_30;
+    output[31 * 32] = step1_0 - step2_31;
+
+    input += 32;
+    output += 1;
+  }
+}
+
+void aom_idct32x32_1024_add_dspr2(const int16_t *input, uint8_t *dest,
+                                  int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t, out[32 * 32]);
+  int16_t *outptr = out;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                       :
+                       : [pos] "r"(pos));
+
+  // Rows
+  idct32_rows_dspr2(input, outptr, 32);
+
+  // Columns
+  aom_idct32_cols_add_blk_dspr2(out, dest, dest_stride);
+}
+
+void aom_idct32x32_34_add_dspr2(const int16_t *input, uint8_t *dest,
+                                int stride) {
+  DECLARE_ALIGNED(32, int16_t, out[32 * 32]);
+  int16_t *outptr = out;
+  uint32_t i;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                       :
+                       : [pos] "r"(pos));
+
+  // Rows
+  idct32_rows_dspr2(input, outptr, 8);
+
+  outptr += 8;
+  __asm__ __volatile__(
+      "sw     $zero,      0(%[outptr])     \n\t"
+      "sw     $zero,      4(%[outptr])     \n\t"
+      "sw     $zero,      8(%[outptr])     \n\t"
+      "sw     $zero,     12(%[outptr])     \n\t"
+      "sw     $zero,     16(%[outptr])     \n\t"
+      "sw     $zero,     20(%[outptr])     \n\t"
+      "sw     $zero,     24(%[outptr])     \n\t"
+      "sw     $zero,     28(%[outptr])     \n\t"
+      "sw     $zero,     32(%[outptr])     \n\t"
+      "sw     $zero,     36(%[outptr])     \n\t"
+      "sw     $zero,     40(%[outptr])     \n\t"
+      "sw     $zero,     44(%[outptr])     \n\t"
+
+      :
+      : [outptr] "r"(outptr));
+
+  for (i = 0; i < 31; ++i) {
+    outptr += 32;
+
+    __asm__ __volatile__(
+        "sw     $zero,      0(%[outptr])     \n\t"
+        "sw     $zero,      4(%[outptr])     \n\t"
+        "sw     $zero,      8(%[outptr])     \n\t"
+        "sw     $zero,     12(%[outptr])     \n\t"
+        "sw     $zero,     16(%[outptr])     \n\t"
+        "sw     $zero,     20(%[outptr])     \n\t"
+        "sw     $zero,     24(%[outptr])     \n\t"
+        "sw     $zero,     28(%[outptr])     \n\t"
+        "sw     $zero,     32(%[outptr])     \n\t"
+        "sw     $zero,     36(%[outptr])     \n\t"
+        "sw     $zero,     40(%[outptr])     \n\t"
+        "sw     $zero,     44(%[outptr])     \n\t"
+
+        :
+        : [outptr] "r"(outptr));
+  }
+
+  // Columns
+  aom_idct32_cols_add_blk_dspr2(out, dest, stride);
+}
+
+void aom_idct32x32_1_add_dspr2(const int16_t *input, uint8_t *dest,
+                               int stride) {
+  int r, out;
+  int32_t a1, absa1;
+  int32_t vector_a1;
+  int32_t t1, t2, t3, t4;
+  int32_t vector_1, vector_2, vector_3, vector_4;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+
+                       :
+                       : [pos] "r"(pos));
+
+  out = DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input[0]);
+  __asm__ __volatile__(
+      "addi     %[out],    %[out],    32      \n\t"
+      "sra      %[a1],     %[out],    6       \n\t"
+
+      : [out] "+r"(out), [a1] "=r"(a1)
+      :);
+
+  if (a1 < 0) {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__(
+        "abs        %[absa1],     %[a1]         \n\t"
+        "replv.qb   %[vector_a1], %[absa1]      \n\t"
+
+        : [absa1] "=r"(absa1), [vector_a1] "=r"(vector_a1)
+        : [a1] "r"(a1));
+
+    for (r = 32; r--;) {
+      __asm__ __volatile__(
+          "lw             %[t1],          0(%[dest])                      \n\t"
+          "lw             %[t2],          4(%[dest])                      \n\t"
+          "lw             %[t3],          8(%[dest])                      \n\t"
+          "lw             %[t4],          12(%[dest])                     \n\t"
+          "subu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    0(%[dest])                      \n\t"
+          "sw             %[vector_2],    4(%[dest])                      \n\t"
+          "sw             %[vector_3],    8(%[dest])                      \n\t"
+          "sw             %[vector_4],    12(%[dest])                     \n\t"
+
+          "lw             %[t1],          16(%[dest])                     \n\t"
+          "lw             %[t2],          20(%[dest])                     \n\t"
+          "lw             %[t3],          24(%[dest])                     \n\t"
+          "lw             %[t4],          28(%[dest])                     \n\t"
+          "subu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "subu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    16(%[dest])                     \n\t"
+          "sw             %[vector_2],    20(%[dest])                     \n\t"
+          "sw             %[vector_3],    24(%[dest])                     \n\t"
+          "sw             %[vector_4],    28(%[dest])                     \n\t"
+
+          "add            %[dest],        %[dest],        %[stride]       \n\t"
+
+          : [t1] "=&r"(t1), [t2] "=&r"(t2), [t3] "=&r"(t3), [t4] "=&r"(t4),
+            [vector_1] "=&r"(vector_1), [vector_2] "=&r"(vector_2),
+            [vector_3] "=&r"(vector_3), [vector_4] "=&r"(vector_4),
+            [dest] "+&r"(dest)
+          : [stride] "r"(stride), [vector_a1] "r"(vector_a1));
+    }
+  } else {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__("replv.qb       %[vector_a1],   %[a1]     \n\t"
+
+                         : [vector_a1] "=r"(vector_a1)
+                         : [a1] "r"(a1));
+
+    for (r = 32; r--;) {
+      __asm__ __volatile__(
+          "lw             %[t1],          0(%[dest])                      \n\t"
+          "lw             %[t2],          4(%[dest])                      \n\t"
+          "lw             %[t3],          8(%[dest])                      \n\t"
+          "lw             %[t4],          12(%[dest])                     \n\t"
+          "addu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    0(%[dest])                      \n\t"
+          "sw             %[vector_2],    4(%[dest])                      \n\t"
+          "sw             %[vector_3],    8(%[dest])                      \n\t"
+          "sw             %[vector_4],    12(%[dest])                     \n\t"
+
+          "lw             %[t1],          16(%[dest])                     \n\t"
+          "lw             %[t2],          20(%[dest])                     \n\t"
+          "lw             %[t3],          24(%[dest])                     \n\t"
+          "lw             %[t4],          28(%[dest])                     \n\t"
+          "addu_s.qb      %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_3],    %[t3],          %[vector_a1]    \n\t"
+          "addu_s.qb      %[vector_4],    %[t4],          %[vector_a1]    \n\t"
+          "sw             %[vector_1],    16(%[dest])                     \n\t"
+          "sw             %[vector_2],    20(%[dest])                     \n\t"
+          "sw             %[vector_3],    24(%[dest])                     \n\t"
+          "sw             %[vector_4],    28(%[dest])                     \n\t"
+
+          "add            %[dest],        %[dest],        %[stride]       \n\t"
+
+          : [t1] "=&r"(t1), [t2] "=&r"(t2), [t3] "=&r"(t3), [t4] "=&r"(t4),
+            [vector_1] "=&r"(vector_1), [vector_2] "=&r"(vector_2),
+            [vector_3] "=&r"(vector_3), [vector_4] "=&r"(vector_4),
+            [dest] "+&r"(dest)
+          : [stride] "r"(stride), [vector_a1] "r"(vector_a1));
+    }
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/itrans4_dspr2.c b/third_party/aom/aom_dsp/mips/itrans4_dspr2.c
new file mode 100644
index 0000000000..e6d0367cd2
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/itrans4_dspr2.c
@@ -0,0 +1,342 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/inv_txfm_dspr2.h"
+#include "aom_dsp/txfm_common.h"
+
+#if HAVE_DSPR2
+void aom_idct4_rows_dspr2(const int16_t *input, int16_t *output) {
+  int16_t step_0, step_1, step_2, step_3;
+  int Temp0, Temp1, Temp2, Temp3;
+  const int const_2_power_13 = 8192;
+  int i;
+
+  for (i = 4; i--;) {
+    __asm__ __volatile__(
+        /*
+          temp_1 = (input[0] + input[2]) * cospi_16_64;
+          step_0 = dct_const_round_shift(temp_1);
+
+          temp_2 = (input[0] - input[2]) * cospi_16_64;
+          step_1 = dct_const_round_shift(temp_2);
+        */
+        "lh       %[Temp0],             0(%[input])                     \n\t"
+        "lh       %[Temp1],             4(%[input])                     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "add      %[Temp2],             %[Temp0],       %[Temp1]        \n\t"
+        "sub      %[Temp3],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac0,                 %[Temp2],       %[cospi_16_64]  \n\t"
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "extp     %[step_0],            $ac0,           31              \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+
+        "madd     $ac1,                 %[Temp3],       %[cospi_16_64]  \n\t"
+        "extp     %[step_1],            $ac1,           31              \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        /*
+          temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+          step_2 = dct_const_round_shift(temp1);
+        */
+        "madd     $ac0,                 %[Temp0],       %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_8_64]   \n\t"
+        "extp     %[step_2],            $ac0,           31              \n\t"
+
+        /*
+          temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+          step_3 = dct_const_round_shift(temp2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_24_64]  \n\t"
+        "extp     %[step_3],            $ac1,           31              \n\t"
+
+        /*
+          output[0]  = step_0 + step_3;
+          output[4]  = step_1 + step_2;
+          output[8]  = step_1 - step_2;
+          output[12] = step_0 - step_3;
+        */
+        "add      %[Temp0],             %[step_0],      %[step_3]       \n\t"
+        "sh       %[Temp0],             0(%[output])                    \n\t"
+
+        "add      %[Temp1],             %[step_1],      %[step_2]       \n\t"
+        "sh       %[Temp1],             8(%[output])                    \n\t"
+
+        "sub      %[Temp2],             %[step_1],      %[step_2]       \n\t"
+        "sh       %[Temp2],             16(%[output])                   \n\t"
+
+        "sub      %[Temp3],             %[step_0],      %[step_3]       \n\t"
+        "sh       %[Temp3],             24(%[output])                   \n\t"
+
+        : [Temp0] "=&r"(Temp0), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+          [Temp3] "=&r"(Temp3), [step_0] "=&r"(step_0), [step_1] "=&r"(step_1),
+          [step_2] "=&r"(step_2), [step_3] "=&r"(step_3), [output] "+r"(output)
+        : [const_2_power_13] "r"(const_2_power_13),
+          [cospi_8_64] "r"(cospi_8_64), [cospi_16_64] "r"(cospi_16_64),
+          [cospi_24_64] "r"(cospi_24_64), [input] "r"(input));
+
+    input += 4;
+    output += 1;
+  }
+}
+
+void aom_idct4_columns_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                     int dest_stride) {
+  int16_t step_0, step_1, step_2, step_3;
+  int Temp0, Temp1, Temp2, Temp3;
+  const int const_2_power_13 = 8192;
+  int i;
+  uint8_t *dest_pix;
+  uint8_t *cm = aom_ff_cropTbl;
+
+  /* prefetch aom_ff_cropTbl */
+  prefetch_load(aom_ff_cropTbl);
+  prefetch_load(aom_ff_cropTbl + 32);
+  prefetch_load(aom_ff_cropTbl + 64);
+  prefetch_load(aom_ff_cropTbl + 96);
+  prefetch_load(aom_ff_cropTbl + 128);
+  prefetch_load(aom_ff_cropTbl + 160);
+  prefetch_load(aom_ff_cropTbl + 192);
+  prefetch_load(aom_ff_cropTbl + 224);
+
+  for (i = 0; i < 4; ++i) {
+    dest_pix = (dest + i);
+
+    __asm__ __volatile__(
+        /*
+          temp_1 = (input[0] + input[2]) * cospi_16_64;
+          step_0 = dct_const_round_shift(temp_1);
+
+          temp_2 = (input[0] - input[2]) * cospi_16_64;
+          step_1 = dct_const_round_shift(temp_2);
+        */
+        "lh       %[Temp0],             0(%[input])                     \n\t"
+        "lh       %[Temp1],             4(%[input])                     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "add      %[Temp2],             %[Temp0],       %[Temp1]        \n\t"
+        "sub      %[Temp3],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac0,                 %[Temp2],       %[cospi_16_64]  \n\t"
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "extp     %[step_0],            $ac0,           31              \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+
+        "madd     $ac1,                 %[Temp3],       %[cospi_16_64]  \n\t"
+        "extp     %[step_1],            $ac1,           31              \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        /*
+          temp1 = input[1] * cospi_24_64 - input[3] * cospi_8_64;
+          step_2 = dct_const_round_shift(temp1);
+        */
+        "madd     $ac0,                 %[Temp0],       %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_8_64]   \n\t"
+        "extp     %[step_2],            $ac0,           31              \n\t"
+
+        /*
+          temp2 = input[1] * cospi_8_64 + input[3] * cospi_24_64;
+          step_3 = dct_const_round_shift(temp2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_24_64]  \n\t"
+        "extp     %[step_3],            $ac1,           31              \n\t"
+
+        /*
+          output[0]  = step_0 + step_3;
+          output[4]  = step_1 + step_2;
+          output[8]  = step_1 - step_2;
+          output[12] = step_0 - step_3;
+        */
+        "add      %[Temp0],             %[step_0],      %[step_3]       \n\t"
+        "addi     %[Temp0],             %[Temp0],       8               \n\t"
+        "sra      %[Temp0],             %[Temp0],       4               \n\t"
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "add      %[Temp0],             %[step_1],      %[step_2]       \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "addi     %[Temp0],             %[Temp0],       8               \n\t"
+        "sra      %[Temp0],             %[Temp0],       4               \n\t"
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step_1],      %[step_2]       \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "addi     %[Temp0],             %[Temp0],       8               \n\t"
+        "sra      %[Temp0],             %[Temp0],       4               \n\t"
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step_0],      %[step_3]       \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "addi     %[Temp0],             %[Temp0],       8               \n\t"
+        "sra      %[Temp0],             %[Temp0],       4               \n\t"
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+
+        : [Temp0] "=&r"(Temp0), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+          [Temp3] "=&r"(Temp3), [step_0] "=&r"(step_0), [step_1] "=&r"(step_1),
+          [step_2] "=&r"(step_2), [step_3] "=&r"(step_3),
+          [dest_pix] "+r"(dest_pix)
+        : [const_2_power_13] "r"(const_2_power_13),
+          [cospi_8_64] "r"(cospi_8_64), [cospi_16_64] "r"(cospi_16_64),
+          [cospi_24_64] "r"(cospi_24_64), [input] "r"(input), [cm] "r"(cm),
+          [dest_stride] "r"(dest_stride));
+
+    input += 4;
+  }
+}
+
+void aom_idct4x4_16_add_dspr2(const int16_t *input, uint8_t *dest,
+                              int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t, out[4 * 4]);
+  int16_t *outptr = out;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                       :
+                       : [pos] "r"(pos));
+
+  // Rows
+  aom_idct4_rows_dspr2(input, outptr);
+
+  // Columns
+  aom_idct4_columns_add_blk_dspr2(&out[0], dest, dest_stride);
+}
+
+void aom_idct4x4_1_add_dspr2(const int16_t *input, uint8_t *dest,
+                             int dest_stride) {
+  int a1, absa1;
+  int r;
+  int32_t out;
+  int t2, vector_a1, vector_a;
+  uint32_t pos = 45;
+  int16_t input_dc = input[0];
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+
+                       :
+                       : [pos] "r"(pos));
+
+  out = DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input_dc);
+  __asm__ __volatile__(
+      "addi     %[out],     %[out],    8       \n\t"
+      "sra      %[a1],      %[out],    4       \n\t"
+
+      : [out] "+r"(out), [a1] "=r"(a1)
+      :);
+
+  if (a1 < 0) {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__(
+        "abs        %[absa1],     %[a1]         \n\t"
+        "replv.qb   %[vector_a1], %[absa1]      \n\t"
+
+        : [absa1] "=r"(absa1), [vector_a1] "=r"(vector_a1)
+        : [a1] "r"(a1));
+
+    for (r = 4; r--;) {
+      __asm__ __volatile__(
+          "lw             %[t2],          0(%[dest])                      \n\t"
+          "subu_s.qb      %[vector_a],    %[t2],          %[vector_a1]    \n\t"
+          "sw             %[vector_a],    0(%[dest])                      \n\t"
+          "add            %[dest],        %[dest],        %[dest_stride]  \n\t"
+
+          : [t2] "=&r"(t2), [vector_a] "=&r"(vector_a), [dest] "+&r"(dest)
+          : [dest_stride] "r"(dest_stride), [vector_a1] "r"(vector_a1));
+    }
+  } else {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__("replv.qb       %[vector_a1],   %[a1]     \n\t"
+                         : [vector_a1] "=r"(vector_a1)
+                         : [a1] "r"(a1));
+
+    for (r = 4; r--;) {
+      __asm__ __volatile__(
+          "lw           %[t2],          0(%[dest])                        \n\t"
+          "addu_s.qb    %[vector_a],    %[t2],            %[vector_a1]    \n\t"
+          "sw           %[vector_a],    0(%[dest])                        \n\t"
+          "add          %[dest],        %[dest],          %[dest_stride]  \n\t"
+
+          : [t2] "=&r"(t2), [vector_a] "=&r"(vector_a), [dest] "+&r"(dest)
+          : [dest_stride] "r"(dest_stride), [vector_a1] "r"(vector_a1));
+    }
+  }
+}
+
+void iadst4_dspr2(const int16_t *input, int16_t *output) {
+  int s0, s1, s2, s3, s4, s5, s6, s7;
+  int x0, x1, x2, x3;
+
+  x0 = input[0];
+  x1 = input[1];
+  x2 = input[2];
+  x3 = input[3];
+
+  if (!(x0 | x1 | x2 | x3)) {
+    output[0] = output[1] = output[2] = output[3] = 0;
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_2_9 * x0;
+  s2 = sinpi_3_9 * x1;
+  s3 = sinpi_4_9 * x2;
+  s4 = sinpi_1_9 * x2;
+  s5 = sinpi_2_9 * x3;
+  s6 = sinpi_4_9 * x3;
+  s7 = x0 - x2 + x3;
+
+  x0 = s0 + s3 + s5;
+  x1 = s1 - s4 - s6;
+  x2 = sinpi_3_9 * s7;
+  x3 = s2;
+
+  s0 = x0 + x3;
+  s1 = x1 + x3;
+  s2 = x2;
+  s3 = x0 + x1 - x3;
+
+  // 1-D transform scaling factor is sqrt(2).
+  // The overall dynamic range is 14b (input) + 14b (multiplication scaling)
+  // + 1b (addition) = 29b.
+  // Hence the output bit depth is 15b.
+  output[0] = dct_const_round_shift(s0);
+  output[1] = dct_const_round_shift(s1);
+  output[2] = dct_const_round_shift(s2);
+  output[3] = dct_const_round_shift(s3);
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/itrans8_dspr2.c b/third_party/aom/aom_dsp/mips/itrans8_dspr2.c
new file mode 100644
index 0000000000..0a20f76f2d
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/itrans8_dspr2.c
@@ -0,0 +1,645 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/inv_txfm_dspr2.h"
+#include "aom_dsp/txfm_common.h"
+
+#if HAVE_DSPR2
+void idct8_rows_dspr2(const int16_t *input, int16_t *output, uint32_t no_rows) {
+  int step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6, step1_7;
+  const int const_2_power_13 = 8192;
+  int Temp0, Temp1, Temp2, Temp3, Temp4;
+  int i;
+
+  for (i = no_rows; i--;) {
+    __asm__ __volatile__(
+        /*
+          temp_1 = (input[0] + input[4]) * cospi_16_64;
+          step2_0 = dct_const_round_shift(temp_1);
+
+          temp_2 = (input[0] - input[4]) * cospi_16_64;
+          step2_1 = dct_const_round_shift(temp_2);
+        */
+        "lh       %[Temp0],             0(%[input])                     \n\t"
+        "lh       %[Temp1],             8(%[input])                     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "add      %[Temp2],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac0,                 %[Temp2],       %[cospi_16_64]  \n\t"
+        "extp     %[Temp4],             $ac0,           31              \n\t"
+
+        "sub      %[Temp3],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac1,                 %[Temp3],       %[cospi_16_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "extp     %[Temp2],             $ac1,           31              \n\t"
+
+        /*
+          temp_1 = input[2] * cospi_24_64 - input[6] * cospi_8_64;
+          step2_2 = dct_const_round_shift(temp_1);
+        */
+        "lh       %[Temp0],             4(%[input])                     \n\t"
+        "lh       %[Temp1],             12(%[input])                    \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_8_64]   \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "extp     %[Temp3],             $ac0,           31              \n\t"
+
+        /*
+          step1_1 = step2_1 + step2_2;
+          step1_2 = step2_1 - step2_2;
+        */
+        "add      %[step1_1],           %[Temp2],       %[Temp3]        \n\t"
+        "sub      %[step1_2],           %[Temp2],       %[Temp3]        \n\t"
+
+        /*
+          temp_2 = input[2] * cospi_8_64 + input[6] * cospi_24_64;
+          step2_3 = dct_const_round_shift(temp_2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_24_64]  \n\t"
+        "extp     %[Temp1],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+
+        /*
+          step1_0 = step2_0 + step2_3;
+          step1_3 = step2_0 - step2_3;
+        */
+        "add      %[step1_0],           %[Temp4],       %[Temp1]        \n\t"
+        "sub      %[step1_3],           %[Temp4],       %[Temp1]        \n\t"
+
+        /*
+          temp_1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+          step1_4 = dct_const_round_shift(temp_1);
+        */
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_28_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "lh       %[Temp1],             14(%[input])                    \n\t"
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_4_64]   \n\t"
+        "extp     %[step1_4],           $ac0,           31              \n\t"
+
+        /*
+          temp_2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+          step1_7 = dct_const_round_shift(temp_2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_4_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_28_64]  \n\t"
+        "extp     %[step1_7],           $ac1,           31              \n\t"
+
+        /*
+          temp_1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+          step1_5 = dct_const_round_shift(temp_1);
+        */
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "lh       %[Temp0],             10(%[input])                    \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_12_64]  \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_20_64]  \n\t"
+        "extp     %[step1_5],           $ac0,           31              \n\t"
+
+        /*
+          temp_2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+          step1_6 = dct_const_round_shift(temp_2);
+        */
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "lh       %[Temp0],             10(%[input])                    \n\t"
+        "madd     $ac1,                 %[Temp0],       %[cospi_20_64]  \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_12_64]  \n\t"
+        "extp     %[step1_6],           $ac1,           31              \n\t"
+
+        /*
+          temp_1 = (step1_7 - step1_6 - step1_4 + step1_5) * cospi_16_64;
+          temp_2 = (step1_4 - step1_5 - step1_6 + step1_7) * cospi_16_64;
+        */
+        "sub      %[Temp0],             %[step1_7],     %[step1_6]      \n\t"
+        "sub      %[Temp0],             %[Temp0],       %[step1_4]      \n\t"
+        "add      %[Temp0],             %[Temp0],       %[step1_5]      \n\t"
+        "sub      %[Temp1],             %[step1_4],     %[step1_5]      \n\t"
+        "sub      %[Temp1],             %[Temp1],       %[step1_6]      \n\t"
+        "add      %[Temp1],             %[Temp1],       %[step1_7]      \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        "madd     $ac0,                 %[Temp0],       %[cospi_16_64]  \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_16_64]  \n\t"
+
+        /*
+          step1_4 = step1_4 + step1_5;
+          step1_7 = step1_6 + step1_7;
+        */
+        "add      %[step1_4],           %[step1_4],     %[step1_5]      \n\t"
+        "add      %[step1_7],           %[step1_7],     %[step1_6]      \n\t"
+
+        "extp     %[step1_5],           $ac0,           31              \n\t"
+        "extp     %[step1_6],           $ac1,           31              \n\t"
+
+        "add      %[Temp0],             %[step1_0],     %[step1_7]      \n\t"
+        "sh       %[Temp0],             0(%[output])                    \n\t"
+        "add      %[Temp1],             %[step1_1],     %[step1_6]      \n\t"
+        "sh       %[Temp1],             16(%[output])                   \n\t"
+        "add      %[Temp0],             %[step1_2],     %[step1_5]      \n\t"
+        "sh       %[Temp0],             32(%[output])                   \n\t"
+        "add      %[Temp1],             %[step1_3],     %[step1_4]      \n\t"
+        "sh       %[Temp1],             48(%[output])                   \n\t"
+
+        "sub      %[Temp0],             %[step1_3],     %[step1_4]      \n\t"
+        "sh       %[Temp0],             64(%[output])                   \n\t"
+        "sub      %[Temp1],             %[step1_2],     %[step1_5]      \n\t"
+        "sh       %[Temp1],             80(%[output])                   \n\t"
+        "sub      %[Temp0],             %[step1_1],     %[step1_6]      \n\t"
+        "sh       %[Temp0],             96(%[output])                   \n\t"
+        "sub      %[Temp1],             %[step1_0],     %[step1_7]      \n\t"
+        "sh       %[Temp1],             112(%[output])                  \n\t"
+
+        : [step1_0] "=&r"(step1_0), [step1_1] "=&r"(step1_1),
+          [step1_2] "=&r"(step1_2), [step1_3] "=&r"(step1_3),
+          [step1_4] "=&r"(step1_4), [step1_5] "=&r"(step1_5),
+          [step1_6] "=&r"(step1_6), [step1_7] "=&r"(step1_7),
+          [Temp0] "=&r"(Temp0), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+          [Temp3] "=&r"(Temp3), [Temp4] "=&r"(Temp4)
+        : [const_2_power_13] "r"(const_2_power_13),
+          [cospi_16_64] "r"(cospi_16_64), [cospi_28_64] "r"(cospi_28_64),
+          [cospi_4_64] "r"(cospi_4_64), [cospi_12_64] "r"(cospi_12_64),
+          [cospi_20_64] "r"(cospi_20_64), [cospi_8_64] "r"(cospi_8_64),
+          [cospi_24_64] "r"(cospi_24_64), [output] "r"(output),
+          [input] "r"(input));
+
+    input += 8;
+    output += 1;
+  }
+}
+
+void idct8_columns_add_blk_dspr2(int16_t *input, uint8_t *dest,
+                                 int dest_stride) {
+  int step1_0, step1_1, step1_2, step1_3, step1_4, step1_5, step1_6, step1_7;
+  int Temp0, Temp1, Temp2, Temp3;
+  int i;
+  const int const_2_power_13 = 8192;
+  uint8_t *dest_pix;
+  uint8_t *cm = aom_ff_cropTbl;
+
+  /* prefetch aom_ff_cropTbl */
+  prefetch_load(aom_ff_cropTbl);
+  prefetch_load(aom_ff_cropTbl + 32);
+  prefetch_load(aom_ff_cropTbl + 64);
+  prefetch_load(aom_ff_cropTbl + 96);
+  prefetch_load(aom_ff_cropTbl + 128);
+  prefetch_load(aom_ff_cropTbl + 160);
+  prefetch_load(aom_ff_cropTbl + 192);
+  prefetch_load(aom_ff_cropTbl + 224);
+
+  for (i = 0; i < 8; ++i) {
+    dest_pix = (dest + i);
+
+    __asm__ __volatile__(
+        /*
+          temp_1 = (input[0] + input[4]) * cospi_16_64;
+          step2_0 = dct_const_round_shift(temp_1);
+
+          temp_2 = (input[0] - input[4]) * cospi_16_64;
+          step2_1 = dct_const_round_shift(temp_2);
+        */
+        "lh       %[Temp0],             0(%[input])                     \n\t"
+        "lh       %[Temp1],             8(%[input])                     \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "add      %[Temp2],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac0,                 %[Temp2],       %[cospi_16_64]  \n\t"
+        "extp     %[step1_6],           $ac0,           31              \n\t"
+
+        "sub      %[Temp3],             %[Temp0],       %[Temp1]        \n\t"
+        "madd     $ac1,                 %[Temp3],       %[cospi_16_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "extp     %[Temp2],             $ac1,           31              \n\t"
+
+        /*
+          temp_1 = input[2] * cospi_24_64 - input[6] * cospi_8_64;
+          step2_2 = dct_const_round_shift(temp_1);
+        */
+        "lh       %[Temp0],             4(%[input])                     \n\t"
+        "lh       %[Temp1],             12(%[input])                    \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_24_64]  \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_8_64]   \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "extp     %[Temp3],             $ac0,           31              \n\t"
+
+        /*
+          step1_1 = step2_1 + step2_2;
+          step1_2 = step2_1 - step2_2;
+        */
+        "add      %[step1_1],           %[Temp2],       %[Temp3]        \n\t"
+        "sub      %[step1_2],           %[Temp2],       %[Temp3]        \n\t"
+
+        /*
+          temp_2 = input[2] * cospi_8_64 + input[6] * cospi_24_64;
+          step2_3 = dct_const_round_shift(temp_2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_8_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_24_64]  \n\t"
+        "extp     %[Temp1],             $ac1,           31              \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+
+        /*
+          step1_0 = step2_0 + step2_3;
+          step1_3 = step2_0 - step2_3;
+        */
+        "add      %[step1_0],           %[step1_6],     %[Temp1]        \n\t"
+        "sub      %[step1_3],           %[step1_6],     %[Temp1]        \n\t"
+
+        /*
+          temp_1 = input[1] * cospi_28_64 - input[7] * cospi_4_64;
+          step1_4 = dct_const_round_shift(temp_1);
+        */
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_28_64]  \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "lh       %[Temp1],             14(%[input])                    \n\t"
+        "lh       %[Temp0],             2(%[input])                     \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_4_64]   \n\t"
+        "extp     %[step1_4],           $ac0,           31              \n\t"
+
+        /*
+          temp_2 = input[1] * cospi_4_64 + input[7] * cospi_28_64;
+          step1_7 = dct_const_round_shift(temp_2);
+        */
+        "madd     $ac1,                 %[Temp0],       %[cospi_4_64]   \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_28_64]  \n\t"
+        "extp     %[step1_7],           $ac1,           31              \n\t"
+
+        /*
+          temp_1 = input[5] * cospi_12_64 - input[3] * cospi_20_64;
+          step1_5 = dct_const_round_shift(temp_1);
+        */
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "lh       %[Temp0],             10(%[input])                    \n\t"
+        "madd     $ac0,                 %[Temp0],       %[cospi_12_64]  \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "msub     $ac0,                 %[Temp1],       %[cospi_20_64]  \n\t"
+        "extp     %[step1_5],           $ac0,           31              \n\t"
+
+        /*
+          temp_2 = input[5] * cospi_20_64 + input[3] * cospi_12_64;
+          step1_6 = dct_const_round_shift(temp_2);
+        */
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+        "lh       %[Temp0],             10(%[input])                    \n\t"
+        "madd     $ac1,                 %[Temp0],       %[cospi_20_64]  \n\t"
+        "lh       %[Temp1],             6(%[input])                     \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_12_64]  \n\t"
+        "extp     %[step1_6],           $ac1,           31              \n\t"
+
+        /*
+          temp_1 = (step1_7 - step1_6 - step1_4 + step1_5) * cospi_16_64;
+          temp_2 = (step1_4 - step1_5 - step1_6 + step1_7) * cospi_16_64;
+        */
+        "sub      %[Temp0],             %[step1_7],     %[step1_6]      \n\t"
+        "sub      %[Temp0],             %[Temp0],       %[step1_4]      \n\t"
+        "add      %[Temp0],             %[Temp0],       %[step1_5]      \n\t"
+        "sub      %[Temp1],             %[step1_4],     %[step1_5]      \n\t"
+        "sub      %[Temp1],             %[Temp1],       %[step1_6]      \n\t"
+        "add      %[Temp1],             %[Temp1],       %[step1_7]      \n\t"
+
+        "mtlo     %[const_2_power_13],  $ac0                            \n\t"
+        "mthi     $zero,                $ac0                            \n\t"
+        "mtlo     %[const_2_power_13],  $ac1                            \n\t"
+        "mthi     $zero,                $ac1                            \n\t"
+
+        "madd     $ac0,                 %[Temp0],       %[cospi_16_64]  \n\t"
+        "madd     $ac1,                 %[Temp1],       %[cospi_16_64]  \n\t"
+
+        /*
+          step1_4 = step1_4 + step1_5;
+          step1_7 = step1_6 + step1_7;
+        */
+        "add      %[step1_4],           %[step1_4],     %[step1_5]      \n\t"
+        "add      %[step1_7],           %[step1_7],     %[step1_6]      \n\t"
+
+        "extp     %[step1_5],           $ac0,           31              \n\t"
+        "extp     %[step1_6],           $ac1,           31              \n\t"
+
+        /* add block */
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "add      %[Temp0],             %[step1_0],     %[step1_7]      \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "add      %[Temp0],             %[step1_1],     %[step1_6]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "add      %[Temp0],             %[step1_2],     %[step1_5]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "add      %[Temp0],             %[step1_3],     %[step1_4]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step1_3],     %[step1_4]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step1_2],     %[step1_5]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step1_1],     %[step1_6]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "sub      %[Temp0],             %[step1_0],     %[step1_7]      \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+        "addu     %[dest_pix],          %[dest_pix],    %[dest_stride]  \n\t"
+
+        "lbu      %[Temp1],             0(%[dest_pix])                  \n\t"
+        "addi     %[Temp0],             %[Temp0],       16              \n\t"
+        "sra      %[Temp0],             %[Temp0],       5               \n\t"
+        "add      %[Temp1],             %[Temp1],       %[Temp0]        \n\t"
+        "lbux     %[Temp2],             %[Temp1](%[cm])                 \n\t"
+        "sb       %[Temp2],             0(%[dest_pix])                  \n\t"
+
+        : [step1_0] "=&r"(step1_0), [step1_1] "=&r"(step1_1),
+          [step1_2] "=&r"(step1_2), [step1_3] "=&r"(step1_3),
+          [step1_4] "=&r"(step1_4), [step1_5] "=&r"(step1_5),
+          [step1_6] "=&r"(step1_6), [step1_7] "=&r"(step1_7),
+          [Temp0] "=&r"(Temp0), [Temp1] "=&r"(Temp1), [Temp2] "=&r"(Temp2),
+          [Temp3] "=&r"(Temp3), [dest_pix] "+r"(dest_pix)
+        : [const_2_power_13] "r"(const_2_power_13),
+          [cospi_16_64] "r"(cospi_16_64), [cospi_28_64] "r"(cospi_28_64),
+          [cospi_4_64] "r"(cospi_4_64), [cospi_12_64] "r"(cospi_12_64),
+          [cospi_20_64] "r"(cospi_20_64), [cospi_8_64] "r"(cospi_8_64),
+          [cospi_24_64] "r"(cospi_24_64), [input] "r"(input), [cm] "r"(cm),
+          [dest_stride] "r"(dest_stride));
+
+    input += 8;
+  }
+}
+
+void aom_idct8x8_64_add_dspr2(const int16_t *input, uint8_t *dest,
+                              int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t, out[8 * 8]);
+  int16_t *outptr = out;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp    %[pos],    1    \n\t" : : [pos] "r"(pos));
+
+  // First transform rows
+  idct8_rows_dspr2(input, outptr, 8);
+
+  // Then transform columns and add to dest
+  idct8_columns_add_blk_dspr2(&out[0], dest, dest_stride);
+}
+
+void aom_idct8x8_12_add_dspr2(const int16_t *input, uint8_t *dest,
+                              int dest_stride) {
+  DECLARE_ALIGNED(32, int16_t, out[8 * 8]);
+  int16_t *outptr = out;
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp    %[pos],    1    \n\t" : : [pos] "r"(pos));
+
+  // First transform rows
+  idct8_rows_dspr2(input, outptr, 4);
+
+  outptr += 4;
+
+  __asm__ __volatile__(
+      "sw  $zero,   0(%[outptr])  \n\t"
+      "sw  $zero,   4(%[outptr])  \n\t"
+      "sw  $zero,  16(%[outptr])  \n\t"
+      "sw  $zero,  20(%[outptr])  \n\t"
+      "sw  $zero,  32(%[outptr])  \n\t"
+      "sw  $zero,  36(%[outptr])  \n\t"
+      "sw  $zero,  48(%[outptr])  \n\t"
+      "sw  $zero,  52(%[outptr])  \n\t"
+      "sw  $zero,  64(%[outptr])  \n\t"
+      "sw  $zero,  68(%[outptr])  \n\t"
+      "sw  $zero,  80(%[outptr])  \n\t"
+      "sw  $zero,  84(%[outptr])  \n\t"
+      "sw  $zero,  96(%[outptr])  \n\t"
+      "sw  $zero, 100(%[outptr])  \n\t"
+      "sw  $zero, 112(%[outptr])  \n\t"
+      "sw  $zero, 116(%[outptr])  \n\t"
+
+      :
+      : [outptr] "r"(outptr));
+
+  // Then transform columns and add to dest
+  idct8_columns_add_blk_dspr2(&out[0], dest, dest_stride);
+}
+
+void aom_idct8x8_1_add_dspr2(const int16_t *input, uint8_t *dest,
+                             int dest_stride) {
+  uint32_t pos = 45;
+  int32_t out;
+  int32_t r;
+  int32_t a1, absa1;
+  int32_t t1, t2, vector_a1, vector_1, vector_2;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+
+                       :
+                       : [pos] "r"(pos));
+
+  out = DCT_CONST_ROUND_SHIFT_TWICE_COSPI_16_64(input[0]);
+  __asm__ __volatile__(
+      "addi     %[out],     %[out],     16      \n\t"
+      "sra      %[a1],      %[out],     5       \n\t"
+
+      : [out] "+r"(out), [a1] "=r"(a1)
+      :);
+
+  if (a1 < 0) {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__(
+        "abs        %[absa1],       %[a1]       \n\t"
+        "replv.qb   %[vector_a1],   %[absa1]    \n\t"
+
+        : [absa1] "=r"(absa1), [vector_a1] "=r"(vector_a1)
+        : [a1] "r"(a1));
+
+    for (r = 8; r--;) {
+      __asm__ __volatile__(
+          "lw           %[t1],          0(%[dest])                      \n\t"
+          "lw           %[t2],          4(%[dest])                      \n\t"
+          "subu_s.qb    %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "subu_s.qb    %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "sw           %[vector_1],    0(%[dest])                      \n\t"
+          "sw           %[vector_2],    4(%[dest])                      \n\t"
+          "add          %[dest],        %[dest],        %[dest_stride]  \n\t"
+
+          : [t1] "=&r"(t1), [t2] "=&r"(t2), [vector_1] "=&r"(vector_1),
+            [vector_2] "=&r"(vector_2), [dest] "+&r"(dest)
+          : [dest_stride] "r"(dest_stride), [vector_a1] "r"(vector_a1));
+    }
+  } else {
+    /* use quad-byte
+     * input and output memory are four byte aligned */
+    __asm__ __volatile__("replv.qb   %[vector_a1],   %[a1]   \n\t"
+
+                         : [vector_a1] "=r"(vector_a1)
+                         : [a1] "r"(a1));
+
+    for (r = 8; r--;) {
+      __asm__ __volatile__(
+          "lw           %[t1],          0(%[dest])                      \n\t"
+          "lw           %[t2],          4(%[dest])                      \n\t"
+          "addu_s.qb    %[vector_1],    %[t1],          %[vector_a1]    \n\t"
+          "addu_s.qb    %[vector_2],    %[t2],          %[vector_a1]    \n\t"
+          "sw           %[vector_1],    0(%[dest])                      \n\t"
+          "sw           %[vector_2],    4(%[dest])                      \n\t"
+          "add          %[dest],        %[dest],        %[dest_stride]  \n\t"
+
+          : [t1] "=&r"(t1), [t2] "=&r"(t2), [vector_1] "=&r"(vector_1),
+            [vector_2] "=&r"(vector_2), [dest] "+r"(dest)
+          : [dest_stride] "r"(dest_stride), [vector_a1] "r"(vector_a1));
+    }
+  }
+}
+
+void iadst8_dspr2(const int16_t *input, int16_t *output) {
+  int s0, s1, s2, s3, s4, s5, s6, s7;
+  int x0, x1, x2, x3, x4, x5, x6, x7;
+
+  x0 = input[7];
+  x1 = input[0];
+  x2 = input[5];
+  x3 = input[2];
+  x4 = input[3];
+  x5 = input[4];
+  x6 = input[1];
+  x7 = input[6];
+
+  if (!(x0 | x1 | x2 | x3 | x4 | x5 | x6 | x7)) {
+    output[0] = output[1] = output[2] = output[3] = output[4] = output[5] =
+        output[6] = output[7] = 0;
+    return;
+  }
+
+  // stage 1
+  s0 = cospi_2_64 * x0 + cospi_30_64 * x1;
+  s1 = cospi_30_64 * x0 - cospi_2_64 * x1;
+  s2 = cospi_10_64 * x2 + cospi_22_64 * x3;
+  s3 = cospi_22_64 * x2 - cospi_10_64 * x3;
+  s4 = cospi_18_64 * x4 + cospi_14_64 * x5;
+  s5 = cospi_14_64 * x4 - cospi_18_64 * x5;
+  s6 = cospi_26_64 * x6 + cospi_6_64 * x7;
+  s7 = cospi_6_64 * x6 - cospi_26_64 * x7;
+
+  x0 = ROUND_POWER_OF_TWO((s0 + s4), DCT_CONST_BITS);
+  x1 = ROUND_POWER_OF_TWO((s1 + s5), DCT_CONST_BITS);
+  x2 = ROUND_POWER_OF_TWO((s2 + s6), DCT_CONST_BITS);
+  x3 = ROUND_POWER_OF_TWO((s3 + s7), DCT_CONST_BITS);
+  x4 = ROUND_POWER_OF_TWO((s0 - s4), DCT_CONST_BITS);
+  x5 = ROUND_POWER_OF_TWO((s1 - s5), DCT_CONST_BITS);
+  x6 = ROUND_POWER_OF_TWO((s2 - s6), DCT_CONST_BITS);
+  x7 = ROUND_POWER_OF_TWO((s3 - s7), DCT_CONST_BITS);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = cospi_8_64 * x4 + cospi_24_64 * x5;
+  s5 = cospi_24_64 * x4 - cospi_8_64 * x5;
+  s6 = -cospi_24_64 * x6 + cospi_8_64 * x7;
+  s7 = cospi_8_64 * x6 + cospi_24_64 * x7;
+
+  x0 = s0 + s2;
+  x1 = s1 + s3;
+  x2 = s0 - s2;
+  x3 = s1 - s3;
+  x4 = ROUND_POWER_OF_TWO((s4 + s6), DCT_CONST_BITS);
+  x5 = ROUND_POWER_OF_TWO((s5 + s7), DCT_CONST_BITS);
+  x6 = ROUND_POWER_OF_TWO((s4 - s6), DCT_CONST_BITS);
+  x7 = ROUND_POWER_OF_TWO((s5 - s7), DCT_CONST_BITS);
+
+  // stage 3
+  s2 = cospi_16_64 * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (x6 - x7);
+
+  x2 = ROUND_POWER_OF_TWO((s2), DCT_CONST_BITS);
+  x3 = ROUND_POWER_OF_TWO((s3), DCT_CONST_BITS);
+  x6 = ROUND_POWER_OF_TWO((s6), DCT_CONST_BITS);
+  x7 = ROUND_POWER_OF_TWO((s7), DCT_CONST_BITS);
+
+  output[0] = x0;
+  output[1] = -x4;
+  output[2] = x6;
+  output[3] = -x2;
+  output[4] = x3;
+  output[5] = -x7;
+  output[6] = x5;
+  output[7] = -x1;
+}
+#endif  // HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_16_msa.c b/third_party/aom/aom_dsp/mips/loopfilter_16_msa.c
new file mode 100644
index 0000000000..fc0c32ce39
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_16_msa.c
@@ -0,0 +1,1487 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_ports/mem.h"
+#include "aom_dsp/mips/loopfilter_msa.h"
+
+int32_t aom_hz_lpf_t4_and_t8_16w(uint8_t *src, int32_t pitch, uint8_t *filter48,
+                                 const uint8_t *b_limit_ptr,
+                                 const uint8_t *limit_ptr,
+                                 const uint8_t *thresh_ptr) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+  v16u8 zero = { 0 };
+
+  /* load vector elements */
+  LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev,
+               mask, flat);
+  AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  if (__msa_test_bz_v(flat)) {
+    ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch);
+
+    return 1;
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero,
+               q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r);
+    AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+    ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+    AOM_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+                p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+                p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
+    filter48 += (4 * 16);
+    ST_UB2(q1_out, q2_out, filter48, 16);
+    filter48 += (2 * 16);
+    ST_UB(flat, filter48);
+
+    return 0;
+  }
+}
+
+void aom_hz_lpf_t16_16w(uint8_t *src, int32_t pitch, uint8_t *filter48) {
+  v16u8 flat, flat2, filter8;
+  v16i8 zero = { 0 };
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
+  v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
+  v8u16 p7_l_in, p6_l_in, p5_l_in, p4_l_in, p3_l_in, p2_l_in, p1_l_in, p0_l_in;
+  v8u16 q7_l_in, q6_l_in, q5_l_in, q4_l_in, q3_l_in, q2_l_in, q1_l_in, q0_l_in;
+  v8u16 tmp0_r, tmp1_r, tmp0_l, tmp1_l;
+  v8i16 l_out, r_out;
+
+  flat = LD_UB(filter48 + 96);
+
+  LD_UB8((src - 8 * pitch), pitch, p7, p6, p5, p4, p3, p2, p1, p0);
+  LD_UB8(src, pitch, q0, q1, q2, q3, q4, q5, q6, q7);
+  AOM_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+  if (__msa_test_bz_v(flat2)) {
+    LD_UB4(filter48, 16, p2, p1, p0, q0);
+    LD_UB2(filter48 + 4 * 16, 16, q1, q2);
+
+    src -= 3 * pitch;
+    ST_UB4(p2, p1, p0, q0, src, pitch);
+    src += (4 * pitch);
+    ST_UB2(q1, q2, src, pitch);
+  } else {
+    src -= 7 * pitch;
+
+    ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2, zero,
+               p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in,
+               p2_r_in, p1_r_in, p0_r_in);
+
+    q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
+
+    tmp0_r = p7_r_in << 3;
+    tmp0_r -= p7_r_in;
+    tmp0_r += p6_r_in;
+    tmp0_r += q0_r_in;
+    tmp1_r = p6_r_in + p5_r_in;
+    tmp1_r += p4_r_in;
+    tmp1_r += p3_r_in;
+    tmp1_r += p2_r_in;
+    tmp1_r += p1_r_in;
+    tmp1_r += p0_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    ILVL_B4_UH(zero, p7, zero, p6, zero, p5, zero, p4, p7_l_in, p6_l_in,
+               p5_l_in, p4_l_in);
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l_in, p2_l_in,
+               p1_l_in, p0_l_in);
+    q0_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q0);
+
+    tmp0_l = p7_l_in << 3;
+    tmp0_l -= p7_l_in;
+    tmp0_l += p6_l_in;
+    tmp0_l += q0_l_in;
+    tmp1_l = p6_l_in + p5_l_in;
+    tmp1_l += p4_l_in;
+    tmp1_l += p3_l_in;
+    tmp1_l += p2_l_in;
+    tmp1_l += p1_l_in;
+    tmp1_l += p0_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
+    ST_UB(p6, src);
+    src += pitch;
+
+    /* p5 */
+    q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
+    tmp0_r = p5_r_in - p6_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q1_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q1);
+    tmp0_l = p5_l_in - p6_l_in;
+    tmp0_l += q1_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
+    ST_UB(p5, src);
+    src += pitch;
+
+    /* p4 */
+    q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
+    tmp0_r = p4_r_in - p5_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = (v8i16)__msa_srari_h((v8i16)tmp1_r, 4);
+
+    q2_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q2);
+    tmp0_l = p4_l_in - p5_l_in;
+    tmp0_l += q2_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
+    ST_UB(p4, src);
+    src += pitch;
+
+    /* p3 */
+    q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
+    tmp0_r = p3_r_in - p4_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q3_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q3);
+    tmp0_l = p3_l_in - p4_l_in;
+    tmp0_l += q3_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
+    ST_UB(p3, src);
+    src += pitch;
+
+    /* p2 */
+    q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
+    filter8 = LD_UB(filter48);
+    tmp0_r = p2_r_in - p3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q4_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q4);
+    tmp0_l = p2_l_in - p3_l_in;
+    tmp0_l += q4_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* p1 */
+    q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
+    filter8 = LD_UB(filter48 + 16);
+    tmp0_r = p1_r_in - p2_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q5_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q5);
+    tmp0_l = p1_l_in - p2_l_in;
+    tmp0_l += q5_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* p0 */
+    q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
+    filter8 = LD_UB(filter48 + 32);
+    tmp0_r = p0_r_in - p1_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q6_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q6);
+    tmp0_l = p0_l_in - p1_l_in;
+    tmp0_l += q6_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* q0 */
+    q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
+    filter8 = LD_UB(filter48 + 48);
+    tmp0_r = q7_r_in - p0_r_in;
+    tmp0_r += q0_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    q7_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q7);
+    tmp0_l = q7_l_in - p0_l_in;
+    tmp0_l += q0_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* q1 */
+    filter8 = LD_UB(filter48 + 64);
+    tmp0_r = q7_r_in - q0_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p6_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q0_l_in;
+    tmp0_l += q1_l_in;
+    tmp0_l -= p6_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* q2 */
+    filter8 = LD_UB(filter48 + 80);
+    tmp0_r = q7_r_in - q1_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p5_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q1_l_in;
+    tmp0_l += q2_l_in;
+    tmp0_l -= p5_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += pitch;
+
+    /* q3 */
+    tmp0_r = q7_r_in - q2_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p4_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q2_l_in;
+    tmp0_l += q3_l_in;
+    tmp0_l -= p4_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
+    ST_UB(q3, src);
+    src += pitch;
+
+    /* q4 */
+    tmp0_r = q7_r_in - q3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p3_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q3_l_in;
+    tmp0_l += q4_l_in;
+    tmp0_l -= p3_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
+    ST_UB(q4, src);
+    src += pitch;
+
+    /* q5 */
+    tmp0_r = q7_r_in - q4_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p2_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q4_l_in;
+    tmp0_l += q5_l_in;
+    tmp0_l -= p2_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
+    ST_UB(q5, src);
+    src += pitch;
+
+    /* q6 */
+    tmp0_r = q7_r_in - q5_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p1_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    tmp0_l = q7_l_in - q5_l_in;
+    tmp0_l += q6_l_in;
+    tmp0_l -= p1_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
+    ST_UB(q6, src);
+  }
+}
+
+void aom_lpf_horizontal_16_dual_msa(uint8_t *src, int32_t pitch,
+                                    const uint8_t *b_limit_ptr,
+                                    const uint8_t *limit_ptr,
+                                    const uint8_t *thresh_ptr, int32_t count) {
+  DECLARE_ALIGNED(32, uint8_t, filter48[16 * 8]);
+  uint8_t early_exit = 0;
+
+  (void)count;
+
+  early_exit = aom_hz_lpf_t4_and_t8_16w(src, pitch, &filter48[0], b_limit_ptr,
+                                        limit_ptr, thresh_ptr);
+
+  if (0 == early_exit) {
+    aom_hz_lpf_t16_16w(src, pitch, filter48);
+  }
+}
+
+static void mb_lpf_horizontal_edge(uint8_t *src, int32_t pitch,
+                                   const uint8_t *b_limit_ptr,
+                                   const uint8_t *limit_ptr,
+                                   const uint8_t *thresh_ptr, int32_t count) {
+  if (1 == count) {
+    uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
+    uint64_t dword0, dword1;
+    v16u8 flat2, mask, hev, flat, thresh, b_limit, limit;
+    v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p7, p6, p5, p4, q4, q5, q6, q7;
+    v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+    v16u8 p0_filter16, p1_filter16;
+    v8i16 p2_filter8, p1_filter8, p0_filter8;
+    v8i16 q0_filter8, q1_filter8, q2_filter8;
+    v8u16 p7_r, p6_r, p5_r, p4_r, q7_r, q6_r, q5_r, q4_r;
+    v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r;
+    v16i8 zero = { 0 };
+    v8u16 tmp0, tmp1, tmp2;
+
+    /* load vector elements */
+    LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+    thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+    b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+    limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+    LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev,
+                 mask, flat);
+    AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+    AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out,
+                       q1_out);
+
+    flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+    if (__msa_test_bz_v(flat)) {
+      p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+      p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+      q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+      q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+      SD4(p1_d, p0_d, q0_d, q1_d, src - 2 * pitch, pitch);
+    } else {
+      /* convert 8 bit input data into 16 bit */
+      ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+                 zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+                 q3_r);
+      AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8,
+                  p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8);
+
+      /* convert 16 bit output data into 8 bit */
+      PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8, zero,
+                  q0_filter8, p2_filter8, p1_filter8, p0_filter8, q0_filter8);
+      PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8);
+
+      /* store pixel values */
+      p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat);
+      p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat);
+      p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat);
+      q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat);
+      q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat);
+      q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat);
+
+      /* load 16 vector elements */
+      LD_UB4((src - 8 * pitch), pitch, p7, p6, p5, p4);
+      LD_UB4(src + (4 * pitch), pitch, q4, q5, q6, q7);
+
+      AOM_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+      if (__msa_test_bz_v(flat2)) {
+        p2_d = __msa_copy_u_d((v2i64)p2_out, 0);
+        p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+        p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+        q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+        q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+        q2_d = __msa_copy_u_d((v2i64)q2_out, 0);
+
+        SD4(p2_d, p1_d, p0_d, q0_d, src - 3 * pitch, pitch);
+        SD(q1_d, src + pitch);
+        SD(q2_d, src + 2 * pitch);
+      } else {
+        /* LSB(right) 8 pixel operation */
+        ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, q4, zero, q5,
+                   zero, q6, zero, q7, p7_r, p6_r, p5_r, p4_r, q4_r, q5_r, q6_r,
+                   q7_r);
+
+        tmp0 = p7_r << 3;
+        tmp0 -= p7_r;
+        tmp0 += p6_r;
+        tmp0 += q0_r;
+
+        src -= 7 * pitch;
+
+        /* calculation of p6 and p5 */
+        tmp1 = p6_r + p5_r + p4_r + p3_r;
+        tmp1 += (p2_r + p1_r + p0_r);
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp0 = p5_r - p6_r + q1_r - p7_r;
+        tmp1 += tmp0;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(p6, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(p5, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of p4 and p3 */
+        tmp0 = p4_r - p5_r + q2_r - p7_r;
+        tmp2 = p3_r - p4_r + q3_r - p7_r;
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(p4, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(p3, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of p2 and p1 */
+        tmp0 = p2_r - p3_r + q4_r - p7_r;
+        tmp2 = p1_r - p2_r + q5_r - p7_r;
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(p2_out, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(p1_out, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of p0 and q0 */
+        tmp0 = (p0_r - p1_r) + (q6_r - p7_r);
+        tmp2 = (q7_r - p0_r) + (q0_r - p7_r);
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(p0_out, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(q0_out, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of q1 and q2 */
+        tmp0 = q7_r - q0_r + q1_r - p6_r;
+        tmp2 = q7_r - q1_r + q2_r - p5_r;
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(q1_out, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(q2_out, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of q3 and q4 */
+        tmp0 = (q7_r - q2_r) + (q3_r - p4_r);
+        tmp2 = (q7_r - q3_r) + (q4_r - p3_r);
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(q3, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(q4, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+        src += pitch;
+
+        /* calculation of q5 and q6 */
+        tmp0 = (q7_r - q4_r) + (q5_r - p2_r);
+        tmp2 = (q7_r - q5_r) + (q6_r - p1_r);
+        tmp1 += tmp0;
+        p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        tmp1 += tmp2;
+        p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+        PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+                    p1_filter16);
+        p0_filter16 = __msa_bmnz_v(q5, p0_filter16, flat2);
+        p1_filter16 = __msa_bmnz_v(q6, p1_filter16, flat2);
+        dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+        dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+        SD(dword0, src);
+        src += pitch;
+        SD(dword1, src);
+      }
+    }
+  } else {
+    aom_lpf_horizontal_16_dual_msa(src, pitch, b_limit_ptr, limit_ptr,
+                                   thresh_ptr, count);
+  }
+}
+
+void aom_lpf_horizontal_edge_8_msa(uint8_t *src, int32_t pitch,
+                                   const uint8_t *b_limit_ptr,
+                                   const uint8_t *limit_ptr,
+                                   const uint8_t *thresh_ptr) {
+  mb_lpf_horizontal_edge(src, pitch, b_limit_ptr, limit_ptr, thresh_ptr, 1);
+}
+
+void aom_lpf_horizontal_edge_16_msa(uint8_t *src, int32_t pitch,
+                                    const uint8_t *b_limit_ptr,
+                                    const uint8_t *limit_ptr,
+                                    const uint8_t *thresh_ptr) {
+  mb_lpf_horizontal_edge(src, pitch, b_limit_ptr, limit_ptr, thresh_ptr, 2);
+}
+
+static void transpose_16x8_to_8x16(uint8_t *input, int32_t in_pitch,
+                                   uint8_t *output, int32_t out_pitch) {
+  v16u8 p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org, p0_org;
+  v16i8 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+
+  LD_UB8(input, in_pitch, p7_org, p6_org, p5_org, p4_org, p3_org, p2_org,
+         p1_org, p0_org);
+  /* 8x8 transpose */
+  TRANSPOSE8x8_UB_UB(p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org,
+                     p0_org, p7, p6, p5, p4, p3, p2, p1, p0);
+  /* 8x8 transpose */
+  ILVL_B4_SB(p5_org, p7_org, p4_org, p6_org, p1_org, p3_org, p0_org, p2_org,
+             tmp0, tmp1, tmp2, tmp3);
+  ILVR_B2_SB(tmp1, tmp0, tmp3, tmp2, tmp4, tmp6);
+  ILVL_B2_SB(tmp1, tmp0, tmp3, tmp2, tmp5, tmp7);
+  ILVR_W2_UB(tmp6, tmp4, tmp7, tmp5, q0, q4);
+  ILVL_W2_UB(tmp6, tmp4, tmp7, tmp5, q2, q6);
+  SLDI_B4_0_UB(q0, q2, q4, q6, q1, q3, q5, q7, 8);
+
+  ST_UB8(p7, p6, p5, p4, p3, p2, p1, p0, output, out_pitch);
+  output += (8 * out_pitch);
+  ST_UB8(q0, q1, q2, q3, q4, q5, q6, q7, output, out_pitch);
+}
+
+static void transpose_8x16_to_16x8(uint8_t *input, int32_t in_pitch,
+                                   uint8_t *output, int32_t out_pitch) {
+  v16u8 p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o;
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+
+  LD_UB8(input, in_pitch, p7, p6, p5, p4, p3, p2, p1, p0);
+  LD_UB8(input + (8 * in_pitch), in_pitch, q0, q1, q2, q3, q4, q5, q6, q7);
+  TRANSPOSE16x8_UB_UB(p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5,
+                      q6, q7, p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o);
+  ST_UB8(p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o, output, out_pitch);
+}
+
+static void transpose_16x16(uint8_t *input, int32_t in_pitch, uint8_t *output,
+                            int32_t out_pitch) {
+  v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
+  v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  v8i16 tmp0, tmp1, tmp4, tmp5, tmp6, tmp7;
+  v4i32 tmp2, tmp3;
+
+  LD_UB8(input, in_pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+  input += (8 * in_pitch);
+  LD_UB8(input, in_pitch, row8, row9, row10, row11, row12, row13, row14, row15);
+
+  TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, row8,
+                      row9, row10, row11, row12, row13, row14, row15, p7, p6,
+                      p5, p4, p3, p2, p1, p0);
+
+  /* transpose 16x8 matrix into 8x16 */
+  /* total 8 intermediate register and 32 instructions */
+  q7 = (v16u8)__msa_ilvod_d((v2i64)row8, (v2i64)row0);
+  q6 = (v16u8)__msa_ilvod_d((v2i64)row9, (v2i64)row1);
+  q5 = (v16u8)__msa_ilvod_d((v2i64)row10, (v2i64)row2);
+  q4 = (v16u8)__msa_ilvod_d((v2i64)row11, (v2i64)row3);
+  q3 = (v16u8)__msa_ilvod_d((v2i64)row12, (v2i64)row4);
+  q2 = (v16u8)__msa_ilvod_d((v2i64)row13, (v2i64)row5);
+  q1 = (v16u8)__msa_ilvod_d((v2i64)row14, (v2i64)row6);
+  q0 = (v16u8)__msa_ilvod_d((v2i64)row15, (v2i64)row7);
+
+  ILVEV_B2_SH(q7, q6, q5, q4, tmp0, tmp1);
+  tmp4 = (v8i16)__msa_ilvod_b((v16i8)q6, (v16i8)q7);
+  tmp5 = (v8i16)__msa_ilvod_b((v16i8)q4, (v16i8)q5);
+
+  ILVEV_B2_UB(q3, q2, q1, q0, q5, q7);
+  tmp6 = (v8i16)__msa_ilvod_b((v16i8)q2, (v16i8)q3);
+  tmp7 = (v8i16)__msa_ilvod_b((v16i8)q0, (v16i8)q1);
+
+  ILVEV_H2_SW(tmp0, tmp1, q5, q7, tmp2, tmp3);
+  q0 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+  q4 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+  tmp2 = (v4i32)__msa_ilvod_h(tmp1, tmp0);
+  tmp3 = (v4i32)__msa_ilvod_h((v8i16)q7, (v8i16)q5);
+  q2 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+  q6 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+  ILVEV_H2_SW(tmp4, tmp5, tmp6, tmp7, tmp2, tmp3);
+  q1 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+  q5 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+  tmp2 = (v4i32)__msa_ilvod_h(tmp5, tmp4);
+  tmp3 = (v4i32)__msa_ilvod_h(tmp7, tmp6);
+  q3 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+  q7 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+  ST_UB8(p7, p6, p5, p4, p3, p2, p1, p0, output, out_pitch);
+  output += (8 * out_pitch);
+  ST_UB8(q0, q1, q2, q3, q4, q5, q6, q7, output, out_pitch);
+}
+
+int32_t aom_vt_lpf_t4_and_t8_8w(uint8_t *src, uint8_t *filter48,
+                                uint8_t *src_org, int32_t pitch_org,
+                                const uint8_t *b_limit_ptr,
+                                const uint8_t *limit_ptr,
+                                const uint8_t *thresh_ptr) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v16i8 zero = { 0 };
+  v8i16 vec0, vec1, vec2, vec3;
+
+  /* load vector elements */
+  LD_UB8(src - (4 * 16), 16, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev,
+               mask, flat);
+  /* flat4 */
+  AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  /* filter4 */
+  AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+  if (__msa_test_bz_v(flat)) {
+    ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+    ST4x8_UB(vec2, vec3, (src_org - 2), pitch_org);
+    return 1;
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero,
+               q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r);
+    AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+    /* convert 16 bit output data into 8 bit */
+    p2_r = (v8u16)__msa_pckev_b((v16i8)p2_filt8_r, (v16i8)p2_filt8_r);
+    p1_r = (v8u16)__msa_pckev_b((v16i8)p1_filt8_r, (v16i8)p1_filt8_r);
+    p0_r = (v8u16)__msa_pckev_b((v16i8)p0_filt8_r, (v16i8)p0_filt8_r);
+    q0_r = (v8u16)__msa_pckev_b((v16i8)q0_filt8_r, (v16i8)q0_filt8_r);
+    q1_r = (v8u16)__msa_pckev_b((v16i8)q1_filt8_r, (v16i8)q1_filt8_r);
+    q2_r = (v8u16)__msa_pckev_b((v16i8)q2_filt8_r, (v16i8)q2_filt8_r);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_r, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_r, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_r, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_r, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_r, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_r, flat);
+
+    ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
+    filter48 += (4 * 16);
+    ST_UB2(q1_out, q2_out, filter48, 16);
+    filter48 += (2 * 16);
+    ST_UB(flat, filter48);
+
+    return 0;
+  }
+}
+
+int32_t aom_vt_lpf_t16_8w(uint8_t *src, uint8_t *src_org, int32_t pitch,
+                          uint8_t *filter48) {
+  v16i8 zero = { 0 };
+  v16u8 filter8, flat, flat2;
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
+  v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
+  v8u16 tmp0_r, tmp1_r;
+  v8i16 r_out;
+
+  flat = LD_UB(filter48 + 6 * 16);
+
+  LD_UB8((src - 8 * 16), 16, p7, p6, p5, p4, p3, p2, p1, p0);
+  LD_UB8(src, 16, q0, q1, q2, q3, q4, q5, q6, q7);
+
+  AOM_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+  if (__msa_test_bz_v(flat2)) {
+    v8i16 vec0, vec1, vec2, vec3, vec4;
+
+    LD_UB4(filter48, 16, p2, p1, p0, q0);
+    LD_UB2(filter48 + 4 * 16, 16, q1, q2);
+
+    ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec3, vec4);
+    vec2 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1);
+
+    src_org -= 3;
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec2, 0, (src_org + 4), pitch);
+    src_org += (4 * pitch);
+    ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec2, 4, (src_org + 4), pitch);
+
+    return 1;
+  } else {
+    src -= 7 * 16;
+
+    ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2, zero,
+               p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in,
+               p2_r_in, p1_r_in, p0_r_in);
+    q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
+
+    tmp0_r = p7_r_in << 3;
+    tmp0_r -= p7_r_in;
+    tmp0_r += p6_r_in;
+    tmp0_r += q0_r_in;
+    tmp1_r = p6_r_in + p5_r_in;
+    tmp1_r += p4_r_in;
+    tmp1_r += p3_r_in;
+    tmp1_r += p2_r_in;
+    tmp1_r += p1_r_in;
+    tmp1_r += p0_r_in;
+    tmp1_r += tmp0_r;
+
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
+    ST8x1_UB(p6, src);
+    src += 16;
+
+    /* p5 */
+    q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
+    tmp0_r = p5_r_in - p6_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
+    ST8x1_UB(p5, src);
+    src += 16;
+
+    /* p4 */
+    q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
+    tmp0_r = p4_r_in - p5_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
+    ST8x1_UB(p4, src);
+    src += 16;
+
+    /* p3 */
+    q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
+    tmp0_r = p3_r_in - p4_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
+    ST8x1_UB(p3, src);
+    src += 16;
+
+    /* p2 */
+    q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
+    filter8 = LD_UB(filter48);
+    tmp0_r = p2_r_in - p3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* p1 */
+    q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
+    filter8 = LD_UB(filter48 + 16);
+    tmp0_r = p1_r_in - p2_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* p0 */
+    q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
+    filter8 = LD_UB(filter48 + 32);
+    tmp0_r = p0_r_in - p1_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* q0 */
+    q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
+    filter8 = LD_UB(filter48 + 48);
+    tmp0_r = q7_r_in - p0_r_in;
+    tmp0_r += q0_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* q1 */
+    filter8 = LD_UB(filter48 + 64);
+    tmp0_r = q7_r_in - q0_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p6_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* q2 */
+    filter8 = LD_UB(filter48 + 80);
+    tmp0_r = q7_r_in - q1_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p5_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST8x1_UB(filter8, src);
+    src += 16;
+
+    /* q3 */
+    tmp0_r = q7_r_in - q2_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p4_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
+    ST8x1_UB(q3, src);
+    src += 16;
+
+    /* q4 */
+    tmp0_r = q7_r_in - q3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p3_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
+    ST8x1_UB(q4, src);
+    src += 16;
+
+    /* q5 */
+    tmp0_r = q7_r_in - q4_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p2_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
+    ST8x1_UB(q5, src);
+    src += 16;
+
+    /* q6 */
+    tmp0_r = q7_r_in - q5_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p1_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+    q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
+    ST8x1_UB(q6, src);
+
+    return 0;
+  }
+}
+
+void aom_lpf_vertical_16_msa(uint8_t *src, int32_t pitch,
+                             const uint8_t *b_limit_ptr,
+                             const uint8_t *limit_ptr,
+                             const uint8_t *thresh_ptr) {
+  uint8_t early_exit = 0;
+  DECLARE_ALIGNED(32, uint8_t, transposed_input[16 * 24]);
+  uint8_t *filter48 = &transposed_input[16 * 16];
+
+  transpose_16x8_to_8x16(src - 8, pitch, transposed_input, 16);
+
+  early_exit =
+      aom_vt_lpf_t4_and_t8_8w((transposed_input + 16 * 8), &filter48[0], src,
+                              pitch, b_limit_ptr, limit_ptr, thresh_ptr);
+
+  if (0 == early_exit) {
+    early_exit = aom_vt_lpf_t16_8w((transposed_input + 16 * 8), src, pitch,
+                                   &filter48[0]);
+
+    if (0 == early_exit) {
+      transpose_8x16_to_16x8(transposed_input, 16, src - 8, pitch);
+    }
+  }
+}
+
+int32_t aom_vt_lpf_t4_and_t8_16w(uint8_t *src, uint8_t *filter48,
+                                 uint8_t *src_org, int32_t pitch,
+                                 const uint8_t *b_limit_ptr,
+                                 const uint8_t *limit_ptr,
+                                 const uint8_t *thresh_ptr) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+  v16i8 zero = { 0 };
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5;
+
+  /* load vector elements */
+  LD_UB8(src - (4 * 16), 16, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev,
+               mask, flat);
+  /* flat4 */
+  AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  /* filter4 */
+  AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  if (__msa_test_bz_v(flat)) {
+    ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+    ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec4, vec5);
+
+    src_org -= 2;
+    ST4x8_UB(vec2, vec3, src_org, pitch);
+    src_org += 8 * pitch;
+    ST4x8_UB(vec4, vec5, src_org, pitch);
+
+    return 1;
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero,
+               q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r);
+    AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+    ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+    AOM_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+                p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+                p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
+    filter48 += (4 * 16);
+    ST_UB2(q1_out, q2_out, filter48, 16);
+    filter48 += (2 * 16);
+    ST_UB(flat, filter48);
+
+    return 0;
+  }
+}
+
+int32_t aom_vt_lpf_t16_16w(uint8_t *src, uint8_t *src_org, int32_t pitch,
+                           uint8_t *filter48) {
+  v16u8 flat, flat2, filter8;
+  v16i8 zero = { 0 };
+  v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
+  v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
+  v8u16 p7_l_in, p6_l_in, p5_l_in, p4_l_in, p3_l_in, p2_l_in, p1_l_in, p0_l_in;
+  v8u16 q7_l_in, q6_l_in, q5_l_in, q4_l_in, q3_l_in, q2_l_in, q1_l_in, q0_l_in;
+  v8u16 tmp0_r, tmp1_r, tmp0_l, tmp1_l;
+  v8i16 l_out, r_out;
+
+  flat = LD_UB(filter48 + 6 * 16);
+
+  LD_UB8((src - 8 * 16), 16, p7, p6, p5, p4, p3, p2, p1, p0);
+  LD_UB8(src, 16, q0, q1, q2, q3, q4, q5, q6, q7);
+
+  AOM_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+  if (__msa_test_bz_v(flat2)) {
+    v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+
+    LD_UB4(filter48, 16, p2, p1, p0, q0);
+    LD_UB2(filter48 + 4 * 16, 16, q1, q2);
+
+    ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec3, vec4);
+    ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec6, vec7);
+    ILVRL_B2_SH(q2, q1, vec2, vec5);
+
+    src_org -= 3;
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec2, 0, (src_org + 4), pitch);
+    src_org += (4 * pitch);
+    ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec2, 4, (src_org + 4), pitch);
+    src_org += (4 * pitch);
+    ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec5, 0, (src_org + 4), pitch);
+    src_org += (4 * pitch);
+    ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src_org, pitch);
+    ST2x4_UB(vec5, 4, (src_org + 4), pitch);
+
+    return 1;
+  } else {
+    src -= 7 * 16;
+
+    ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2, zero,
+               p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in,
+               p2_r_in, p1_r_in, p0_r_in);
+    q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
+
+    tmp0_r = p7_r_in << 3;
+    tmp0_r -= p7_r_in;
+    tmp0_r += p6_r_in;
+    tmp0_r += q0_r_in;
+    tmp1_r = p6_r_in + p5_r_in;
+    tmp1_r += p4_r_in;
+    tmp1_r += p3_r_in;
+    tmp1_r += p2_r_in;
+    tmp1_r += p1_r_in;
+    tmp1_r += p0_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+    ILVL_B4_UH(zero, p7, zero, p6, zero, p5, zero, p4, p7_l_in, p6_l_in,
+               p5_l_in, p4_l_in);
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l_in, p2_l_in,
+               p1_l_in, p0_l_in);
+    q0_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q0);
+
+    tmp0_l = p7_l_in << 3;
+    tmp0_l -= p7_l_in;
+    tmp0_l += p6_l_in;
+    tmp0_l += q0_l_in;
+    tmp1_l = p6_l_in + p5_l_in;
+    tmp1_l += p4_l_in;
+    tmp1_l += p3_l_in;
+    tmp1_l += p2_l_in;
+    tmp1_l += p1_l_in;
+    tmp1_l += p0_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
+    ST_UB(p6, src);
+    src += 16;
+
+    /* p5 */
+    q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
+    tmp0_r = p5_r_in - p6_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q1_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q1);
+    tmp0_l = p5_l_in - p6_l_in;
+    tmp0_l += q1_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
+    ST_UB(p5, src);
+    src += 16;
+
+    /* p4 */
+    q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
+    tmp0_r = p4_r_in - p5_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q2_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q2);
+    tmp0_l = p4_l_in - p5_l_in;
+    tmp0_l += q2_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
+    ST_UB(p4, src);
+    src += 16;
+
+    /* p3 */
+    q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
+    tmp0_r = p3_r_in - p4_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q3_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q3);
+    tmp0_l = p3_l_in - p4_l_in;
+    tmp0_l += q3_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
+    ST_UB(p3, src);
+    src += 16;
+
+    /* p2 */
+    q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
+    filter8 = LD_UB(filter48);
+    tmp0_r = p2_r_in - p3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q4_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q4);
+    tmp0_l = p2_l_in - p3_l_in;
+    tmp0_l += q4_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* p1 */
+    q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
+    filter8 = LD_UB(filter48 + 16);
+    tmp0_r = p1_r_in - p2_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q5_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q5);
+    tmp0_l = p1_l_in - p2_l_in;
+    tmp0_l += q5_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)(tmp1_l), 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* p0 */
+    q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
+    filter8 = LD_UB(filter48 + 32);
+    tmp0_r = p0_r_in - p1_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q6_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q6);
+    tmp0_l = p0_l_in - p1_l_in;
+    tmp0_l += q6_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* q0 */
+    q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
+    filter8 = LD_UB(filter48 + 48);
+    tmp0_r = q7_r_in - p0_r_in;
+    tmp0_r += q0_r_in;
+    tmp0_r -= p7_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    q7_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q7);
+    tmp0_l = q7_l_in - p0_l_in;
+    tmp0_l += q0_l_in;
+    tmp0_l -= p7_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* q1 */
+    filter8 = LD_UB(filter48 + 64);
+    tmp0_r = q7_r_in - q0_r_in;
+    tmp0_r += q1_r_in;
+    tmp0_r -= p6_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q0_l_in;
+    tmp0_l += q1_l_in;
+    tmp0_l -= p6_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* q2 */
+    filter8 = LD_UB(filter48 + 80);
+    tmp0_r = q7_r_in - q1_r_in;
+    tmp0_r += q2_r_in;
+    tmp0_r -= p5_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q1_l_in;
+    tmp0_l += q2_l_in;
+    tmp0_l -= p5_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+    ST_UB(filter8, src);
+    src += 16;
+
+    /* q3 */
+    tmp0_r = q7_r_in - q2_r_in;
+    tmp0_r += q3_r_in;
+    tmp0_r -= p4_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q2_l_in;
+    tmp0_l += q3_l_in;
+    tmp0_l -= p4_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
+    ST_UB(q3, src);
+    src += 16;
+
+    /* q4 */
+    tmp0_r = q7_r_in - q3_r_in;
+    tmp0_r += q4_r_in;
+    tmp0_r -= p3_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q3_l_in;
+    tmp0_l += q4_l_in;
+    tmp0_l -= p3_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
+    ST_UB(q4, src);
+    src += 16;
+
+    /* q5 */
+    tmp0_r = q7_r_in - q4_r_in;
+    tmp0_r += q5_r_in;
+    tmp0_r -= p2_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q4_l_in;
+    tmp0_l += q5_l_in;
+    tmp0_l -= p2_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
+    ST_UB(q5, src);
+    src += 16;
+
+    /* q6 */
+    tmp0_r = q7_r_in - q5_r_in;
+    tmp0_r += q6_r_in;
+    tmp0_r -= p1_r_in;
+    tmp1_r += tmp0_r;
+    r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+    tmp0_l = q7_l_in - q5_l_in;
+    tmp0_l += q6_l_in;
+    tmp0_l -= p1_l_in;
+    tmp1_l += tmp0_l;
+    l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+    r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+    q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
+    ST_UB(q6, src);
+
+    return 0;
+  }
+}
+
+void aom_lpf_vertical_16_dual_msa(uint8_t *src, int32_t pitch,
+                                  const uint8_t *b_limit_ptr,
+                                  const uint8_t *limit_ptr,
+                                  const uint8_t *thresh_ptr) {
+  uint8_t early_exit = 0;
+  DECLARE_ALIGNED(32, uint8_t, transposed_input[16 * 24]);
+  uint8_t *filter48 = &transposed_input[16 * 16];
+
+  transpose_16x16((src - 8), pitch, &transposed_input[0], 16);
+
+  early_exit =
+      aom_vt_lpf_t4_and_t8_16w((transposed_input + 16 * 8), &filter48[0], src,
+                               pitch, b_limit_ptr, limit_ptr, thresh_ptr);
+
+  if (0 == early_exit) {
+    early_exit = aom_vt_lpf_t16_16w((transposed_input + 16 * 8), src, pitch,
+                                    &filter48[0]);
+
+    if (0 == early_exit) {
+      transpose_16x16(transposed_input, 16, (src - 8), pitch);
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_4_msa.c b/third_party/aom/aom_dsp/mips/loopfilter_4_msa.c
new file mode 100644
index 0000000000..dc0a977645
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_4_msa.c
@@ -0,0 +1,147 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/loopfilter_msa.h"
+
+void aom_lpf_horizontal_4_msa(uint8_t *src, int32_t pitch,
+                              const uint8_t *b_limit_ptr,
+                              const uint8_t *limit_ptr,
+                              const uint8_t *thresh_ptr) {
+  uint64_t p1_d, p0_d, q0_d, q1_d;
+  v16u8 mask, hev, flat, thresh, b_limit, limit;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p1_out, p0_out, q0_out, q1_out;
+
+  /* load vector elements */
+  LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev,
+               mask, flat);
+  AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+  p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+  q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+  q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+  SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch);
+}
+
+void aom_lpf_horizontal_4_dual_msa(uint8_t *src, int32_t pitch,
+                                   const uint8_t *b_limit0_ptr,
+                                   const uint8_t *limit0_ptr,
+                                   const uint8_t *thresh0_ptr,
+                                   const uint8_t *b_limit1_ptr,
+                                   const uint8_t *limit1_ptr,
+                                   const uint8_t *thresh1_ptr) {
+  v16u8 mask, hev, flat, thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+
+  /* load vector elements */
+  LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
+  thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
+  thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
+
+  b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
+  b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
+  b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
+
+  limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
+  limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
+  limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0, hev,
+               mask, flat);
+  AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+
+  ST_UB4(p1, p0, q0, q1, (src - 2 * pitch), pitch);
+}
+
+void aom_lpf_vertical_4_msa(uint8_t *src, int32_t pitch,
+                            const uint8_t *b_limit_ptr,
+                            const uint8_t *limit_ptr,
+                            const uint8_t *thresh_ptr) {
+  v16u8 mask, hev, flat, limit, thresh, b_limit;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v8i16 vec0, vec1, vec2, vec3;
+
+  LD_UB8((src - 4), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3, p3, p2, p1, p0, q0, q1, q2,
+                     q3);
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev,
+               mask, flat);
+  AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+  ILVR_B2_SH(p0, p1, q1, q0, vec0, vec1);
+  ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+
+  src -= 2;
+  ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+  src += 4 * pitch;
+  ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+}
+
+void aom_lpf_vertical_4_dual_msa(uint8_t *src, int32_t pitch,
+                                 const uint8_t *b_limit0_ptr,
+                                 const uint8_t *limit0_ptr,
+                                 const uint8_t *thresh0_ptr,
+                                 const uint8_t *b_limit1_ptr,
+                                 const uint8_t *limit1_ptr,
+                                 const uint8_t *thresh1_ptr) {
+  v16u8 mask, hev, flat;
+  v16u8 thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
+  v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
+  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+
+  LD_UB8(src - 4, pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+  LD_UB8(src - 4 + (8 * pitch), pitch, row8, row9, row10, row11, row12, row13,
+         row14, row15);
+
+  TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, row8,
+                      row9, row10, row11, row12, row13, row14, row15, p3, p2,
+                      p1, p0, q0, q1, q2, q3);
+
+  thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
+  thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
+  thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
+
+  b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
+  b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
+  b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
+
+  limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
+  limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
+  limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0, hev,
+               mask, flat);
+  AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+  ILVR_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
+  ILVRL_H2_SH(tmp1, tmp0, tmp2, tmp3);
+  ILVL_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
+  ILVRL_H2_SH(tmp1, tmp0, tmp4, tmp5);
+
+  src -= 2;
+
+  ST4x8_UB(tmp2, tmp3, src, pitch);
+  src += (8 * pitch);
+  ST4x8_UB(tmp4, tmp5, src, pitch);
+}
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_8_msa.c b/third_party/aom/aom_dsp/mips/loopfilter_8_msa.c
new file mode 100644
index 0000000000..dc203e79cf
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_8_msa.c
@@ -0,0 +1,333 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/mips/loopfilter_msa.h"
+
+void aom_lpf_horizontal_8_msa(uint8_t *src, int32_t pitch,
+                              const uint8_t *b_limit_ptr,
+                              const uint8_t *limit_ptr,
+                              const uint8_t *thresh_ptr) {
+  uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
+  v16u8 mask, hev, flat, thresh, b_limit, limit;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v8i16 p2_filter8, p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r;
+  v16i8 zero = { 0 };
+
+  /* load vector elements */
+  LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev,
+               mask, flat);
+  AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+  if (__msa_test_bz_v(flat)) {
+    p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+    p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+    q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+    q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+    SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero,
+               q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r);
+    AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8,
+                p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8, zero,
+                q0_filter8, p2_filter8, p1_filter8, p0_filter8, q0_filter8);
+    PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat);
+
+    p2_d = __msa_copy_u_d((v2i64)p2_out, 0);
+    p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+    p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+    q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+    q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+    q2_d = __msa_copy_u_d((v2i64)q2_out, 0);
+
+    src -= 3 * pitch;
+
+    SD4(p2_d, p1_d, p0_d, q0_d, src, pitch);
+    src += (4 * pitch);
+    SD(q1_d, src);
+    src += pitch;
+    SD(q2_d, src);
+  }
+}
+
+void aom_lpf_horizontal_8_dual_msa(
+    uint8_t *src, int32_t pitch, const uint8_t *b_limit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *b_limit1, const uint8_t *limit1,
+    const uint8_t *thresh1) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+  v16u8 flat, mask, hev, tmp, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+  v16u8 zero = { 0 };
+
+  /* load vector elements */
+  LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh0);
+  tmp = (v16u8)__msa_fill_b(*thresh1);
+  thresh = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)thresh);
+
+  b_limit = (v16u8)__msa_fill_b(*b_limit0);
+  tmp = (v16u8)__msa_fill_b(*b_limit1);
+  b_limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)b_limit);
+
+  limit = (v16u8)__msa_fill_b(*limit0);
+  tmp = (v16u8)__msa_fill_b(*limit1);
+  limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)limit);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev,
+               mask, flat);
+  AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  if (__msa_test_bz_v(flat)) {
+    ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero,
+               q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r);
+    AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+    ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+    AOM_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+                p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+                p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    src -= 3 * pitch;
+
+    ST_UB4(p2_out, p1_out, p0_out, q0_out, src, pitch);
+    src += (4 * pitch);
+    ST_UB2(q1_out, q2_out, src, pitch);
+    src += (2 * pitch);
+  }
+}
+
+void aom_lpf_vertical_8_msa(uint8_t *src, int32_t pitch,
+                            const uint8_t *b_limit_ptr,
+                            const uint8_t *limit_ptr,
+                            const uint8_t *thresh_ptr) {
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p1_out, p0_out, q0_out, q1_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v16u8 zero = { 0 };
+  v8i16 vec0, vec1, vec2, vec3, vec4;
+
+  /* load vector elements */
+  LD_UB8(src - 4, pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+  TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3, p3, p2, p1, p0, q0, q1, q2,
+                     q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+  b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+  limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev,
+               mask, flat);
+  /* flat4 */
+  AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  /* filter4 */
+  AOM_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+  if (__msa_test_bz_v(flat)) {
+    /* Store 4 pixels p1-_q1 */
+    ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+
+    src -= 2;
+    ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+    src += 4 * pitch;
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero,
+               q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r);
+    AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_r, p2_filt8_r, p1_filt8_r, p1_filt8_r, p0_filt8_r,
+                p0_filt8_r, q0_filt8_r, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_r, q1_filt8_r, q2_filt8_r, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    /* Store 6 pixels p2-_q2 */
+    ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+    vec4 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1);
+
+    src -= 3;
+    ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec4, 0, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec4, 4, src + 4, pitch);
+  }
+}
+
+void aom_lpf_vertical_8_dual_msa(uint8_t *src, int32_t pitch,
+                                 const uint8_t *b_limit0, const uint8_t *limit0,
+                                 const uint8_t *thresh0,
+                                 const uint8_t *b_limit1, const uint8_t *limit1,
+                                 const uint8_t *thresh1) {
+  uint8_t *temp_src;
+  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+  v16u8 p1_out, p0_out, q0_out, q1_out;
+  v16u8 flat, mask, hev, thresh, b_limit, limit;
+  v16u8 row4, row5, row6, row7, row12, row13, row14, row15;
+  v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+  v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+  v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+  v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+  v16u8 zero = { 0 };
+  v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+
+  temp_src = src - 4;
+
+  LD_UB8(temp_src, pitch, p0, p1, p2, p3, row4, row5, row6, row7);
+  temp_src += (8 * pitch);
+  LD_UB8(temp_src, pitch, q3, q2, q1, q0, row12, row13, row14, row15);
+
+  /* transpose 16x8 matrix into 8x16 */
+  TRANSPOSE16x8_UB_UB(p0, p1, p2, p3, row4, row5, row6, row7, q3, q2, q1, q0,
+                      row12, row13, row14, row15, p3, p2, p1, p0, q0, q1, q2,
+                      q3);
+
+  thresh = (v16u8)__msa_fill_b(*thresh0);
+  vec0 = (v8i16)__msa_fill_b(*thresh1);
+  thresh = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)thresh);
+
+  b_limit = (v16u8)__msa_fill_b(*b_limit0);
+  vec0 = (v8i16)__msa_fill_b(*b_limit1);
+  b_limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)b_limit);
+
+  limit = (v16u8)__msa_fill_b(*limit0);
+  vec0 = (v8i16)__msa_fill_b(*limit1);
+  limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)limit);
+
+  /* mask and hev */
+  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh, hev,
+               mask, flat);
+  /* flat4 */
+  AOM_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+  /* filter4 */
+  AOM_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+  if (__msa_test_bz_v(flat)) {
+    ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+    ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec4, vec5);
+
+    src -= 2;
+    ST4x8_UB(vec2, vec3, src, pitch);
+    src += 8 * pitch;
+    ST4x8_UB(vec4, vec5, src, pitch);
+  } else {
+    ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1, zero,
+               q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r);
+    AOM_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+                p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+    ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+    ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+
+    /* filter8 */
+    AOM_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+                p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+    /* convert 16 bit output data into 8 bit */
+    PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+                p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+                p0_filt8_r, q0_filt8_r);
+    PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+                q2_filt8_r);
+
+    /* store pixel values */
+    p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+    p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+    p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+    q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+    q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+    q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+    ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec3, vec4);
+    ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1);
+    ILVRL_H2_SH(vec1, vec0, vec6, vec7);
+    ILVRL_B2_SH(q2, q1, vec2, vec5);
+
+    src -= 3;
+    ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec2, 0, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec2, 4, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec5, 0, src + 4, pitch);
+    src += (4 * pitch);
+    ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src, pitch);
+    ST2x4_UB(vec5, 4, src + 4, pitch);
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.c b/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.c
new file mode 100644
index 0000000000..883d0523d3
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.c
@@ -0,0 +1,327 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/mips/common_dspr2.h"
+#include "aom_dsp/mips/loopfilter_filters_dspr2.h"
+#include "aom_dsp/mips/loopfilter_macros_dspr2.h"
+#include "aom_dsp/mips/loopfilter_masks_dspr2.h"
+#include "aom_mem/aom_mem.h"
+
+#if HAVE_DSPR2
+void aom_lpf_horizontal_4_dspr2(unsigned char *s, int pitch,
+                                const uint8_t *blimit, const uint8_t *limit,
+                                const uint8_t *thresh) {
+  uint8_t i;
+  uint32_t mask;
+  uint32_t hev;
+  uint32_t pm1, p0, p1, p2, p3, p4, p5, p6;
+  uint8_t *sm1, *s0, *s1, *s2, *s3, *s4, *s5, *s6;
+  uint32_t thresh_vec, flimit_vec, limit_vec;
+  uint32_t uflimit, ulimit, uthresh;
+
+  uflimit = *blimit;
+  ulimit = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__(
+      "replv.qb       %[thresh_vec],    %[uthresh]    \n\t"
+      "replv.qb       %[flimit_vec],    %[uflimit]    \n\t"
+      "replv.qb       %[limit_vec],     %[ulimit]     \n\t"
+
+      : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec),
+        [limit_vec] "=r"(limit_vec)
+      : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit));
+
+  /* prefetch data for store */
+  prefetch_store(s);
+
+  /* loop filter designed to work using chars so that we can make maximum use
+     of 8 bit simd instructions. */
+  for (i = 0; i < 2; i++) {
+    sm1 = s - (pitch << 2);
+    s0 = sm1 + pitch;
+    s1 = s0 + pitch;
+    s2 = s - pitch;
+    s3 = s;
+    s4 = s + pitch;
+    s5 = s4 + pitch;
+    s6 = s5 + pitch;
+
+    __asm__ __volatile__(
+        "lw     %[p1],  (%[s1])    \n\t"
+        "lw     %[p2],  (%[s2])    \n\t"
+        "lw     %[p3],  (%[s3])    \n\t"
+        "lw     %[p4],  (%[s4])    \n\t"
+
+        : [p1] "=&r"(p1), [p2] "=&r"(p2), [p3] "=&r"(p3), [p4] "=&r"(p4)
+        : [s1] "r"(s1), [s2] "r"(s2), [s3] "r"(s3), [s4] "r"(s4));
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+       mask will be zero and filtering is not needed */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0))) {
+      __asm__ __volatile__(
+          "lw       %[pm1], (%[sm1])   \n\t"
+          "lw       %[p0],  (%[s0])    \n\t"
+          "lw       %[p5],  (%[s5])    \n\t"
+          "lw       %[p6],  (%[s6])    \n\t"
+
+          : [pm1] "=&r"(pm1), [p0] "=&r"(p0), [p5] "=&r"(p5), [p6] "=&r"(p6)
+          : [sm1] "r"(sm1), [s0] "r"(s0), [s5] "r"(s5), [s6] "r"(s6));
+
+      filter_hev_mask_dspr2(limit_vec, flimit_vec, p1, p2, pm1, p0, p3, p4, p5,
+                            p6, thresh_vec, &hev, &mask);
+
+      /* if mask == 0 do filtering is not needed */
+      if (mask) {
+        /* filtering */
+        filter_dspr2(mask, hev, &p1, &p2, &p3, &p4);
+
+        __asm__ __volatile__(
+            "sw     %[p1],  (%[s1])    \n\t"
+            "sw     %[p2],  (%[s2])    \n\t"
+            "sw     %[p3],  (%[s3])    \n\t"
+            "sw     %[p4],  (%[s4])    \n\t"
+
+            :
+            : [p1] "r"(p1), [p2] "r"(p2), [p3] "r"(p3), [p4] "r"(p4),
+              [s1] "r"(s1), [s2] "r"(s2), [s3] "r"(s3), [s4] "r"(s4));
+      }
+    }
+
+    s = s + 4;
+  }
+}
+
+void aom_lpf_vertical_4_dspr2(unsigned char *s, int pitch,
+                              const uint8_t *blimit, const uint8_t *limit,
+                              const uint8_t *thresh) {
+  uint8_t i;
+  uint32_t mask, hev;
+  uint32_t pm1, p0, p1, p2, p3, p4, p5, p6;
+  uint8_t *s1, *s2, *s3, *s4;
+  uint32_t prim1, prim2, sec3, sec4, prim3, prim4;
+  uint32_t thresh_vec, flimit_vec, limit_vec;
+  uint32_t uflimit, ulimit, uthresh;
+
+  uflimit = *blimit;
+  ulimit = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__(
+      "replv.qb       %[thresh_vec],    %[uthresh]    \n\t"
+      "replv.qb       %[flimit_vec],    %[uflimit]    \n\t"
+      "replv.qb       %[limit_vec],     %[ulimit]     \n\t"
+
+      : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec),
+        [limit_vec] "=r"(limit_vec)
+      : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit));
+
+  /* prefetch data for store */
+  prefetch_store(s + pitch);
+
+  for (i = 0; i < 2; i++) {
+    s1 = s;
+    s2 = s + pitch;
+    s3 = s2 + pitch;
+    s4 = s3 + pitch;
+    s = s4 + pitch;
+
+    /* load quad-byte vectors
+     * memory is 4 byte aligned
+     */
+    p2 = *((uint32_t *)(s1 - 4));
+    p6 = *((uint32_t *)(s1));
+    p1 = *((uint32_t *)(s2 - 4));
+    p5 = *((uint32_t *)(s2));
+    p0 = *((uint32_t *)(s3 - 4));
+    p4 = *((uint32_t *)(s3));
+    pm1 = *((uint32_t *)(s4 - 4));
+    p3 = *((uint32_t *)(s4));
+
+    /* transpose pm1, p0, p1, p2 */
+    __asm__ __volatile__(
+        "precrq.qb.ph   %[prim1],   %[p2],      %[p1]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p2],      %[p1]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p0],      %[pm1]      \n\t"
+        "precr.qb.ph    %[prim4],   %[p0],      %[pm1]      \n\t"
+
+        "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[pm1],     %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p2],      %[p1],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p0],      %[pm1],     %[sec4]     \n\t"
+        "append         %[p1],      %[sec3],    16          \n\t"
+        "append         %[pm1],     %[sec4],    16          \n\t"
+
+        : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3),
+          [prim4] "=&r"(prim4), [p2] "+r"(p2), [p1] "+r"(p1), [p0] "+r"(p0),
+          [pm1] "+r"(pm1), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4)
+        :);
+
+    /* transpose p3, p4, p5, p6 */
+    __asm__ __volatile__(
+        "precrq.qb.ph   %[prim1],   %[p6],      %[p5]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p6],      %[p5]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p4],      %[p3]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p4],      %[p3]       \n\t"
+
+        "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p6],      %[p5],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p4],      %[p3],      %[sec4]     \n\t"
+        "append         %[p5],      %[sec3],    16          \n\t"
+        "append         %[p3],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3),
+          [prim4] "=&r"(prim4), [p6] "+r"(p6), [p5] "+r"(p5), [p4] "+r"(p4),
+          [p3] "+r"(p3), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4)
+        :);
+
+    /* if (p1 - p4 == 0) and (p2 - p3 == 0)
+     * mask will be zero and filtering is not needed
+     */
+    if (!(((p1 - p4) == 0) && ((p2 - p3) == 0))) {
+      filter_hev_mask_dspr2(limit_vec, flimit_vec, p1, p2, pm1, p0, p3, p4, p5,
+                            p6, thresh_vec, &hev, &mask);
+
+      /* if mask == 0 do filtering is not needed */
+      if (mask) {
+        /* filtering */
+        filter_dspr2(mask, hev, &p1, &p2, &p3, &p4);
+
+        /* unpack processed 4x4 neighborhood
+         * don't use transpose on output data
+         * because memory isn't aligned
+         */
+        __asm__ __volatile__(
+            "sb     %[p4],   1(%[s4])    \n\t"
+            "sb     %[p3],   0(%[s4])    \n\t"
+            "sb     %[p2],  -1(%[s4])    \n\t"
+            "sb     %[p1],  -2(%[s4])    \n\t"
+
+            :
+            : [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1),
+              [s4] "r"(s4));
+
+        __asm__ __volatile__(
+            "srl    %[p4],  %[p4],  8     \n\t"
+            "srl    %[p3],  %[p3],  8     \n\t"
+            "srl    %[p2],  %[p2],  8     \n\t"
+            "srl    %[p1],  %[p1],  8     \n\t"
+
+            : [p4] "+r"(p4), [p3] "+r"(p3), [p2] "+r"(p2), [p1] "+r"(p1)
+            :);
+
+        __asm__ __volatile__(
+            "sb     %[p4],   1(%[s3])    \n\t"
+            "sb     %[p3],   0(%[s3])    \n\t"
+            "sb     %[p2],  -1(%[s3])    \n\t"
+            "sb     %[p1],  -2(%[s3])    \n\t"
+
+            : [p1] "+r"(p1)
+            : [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), [s3] "r"(s3));
+
+        __asm__ __volatile__(
+            "srl    %[p4],  %[p4],  8     \n\t"
+            "srl    %[p3],  %[p3],  8     \n\t"
+            "srl    %[p2],  %[p2],  8     \n\t"
+            "srl    %[p1],  %[p1],  8     \n\t"
+
+            : [p4] "+r"(p4), [p3] "+r"(p3), [p2] "+r"(p2), [p1] "+r"(p1)
+            :);
+
+        __asm__ __volatile__(
+            "sb     %[p4],   1(%[s2])    \n\t"
+            "sb     %[p3],   0(%[s2])    \n\t"
+            "sb     %[p2],  -1(%[s2])    \n\t"
+            "sb     %[p1],  -2(%[s2])    \n\t"
+
+            :
+            : [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1),
+              [s2] "r"(s2));
+
+        __asm__ __volatile__(
+            "srl    %[p4],  %[p4],  8     \n\t"
+            "srl    %[p3],  %[p3],  8     \n\t"
+            "srl    %[p2],  %[p2],  8     \n\t"
+            "srl    %[p1],  %[p1],  8     \n\t"
+
+            : [p4] "+r"(p4), [p3] "+r"(p3), [p2] "+r"(p2), [p1] "+r"(p1)
+            :);
+
+        __asm__ __volatile__(
+            "sb     %[p4],   1(%[s1])    \n\t"
+            "sb     %[p3],   0(%[s1])    \n\t"
+            "sb     %[p2],  -1(%[s1])    \n\t"
+            "sb     %[p1],  -2(%[s1])    \n\t"
+
+            :
+            : [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1),
+              [s1] "r"(s1));
+      }
+    }
+  }
+}
+
+void aom_lpf_horizontal_4_dual_dspr2(
+    uint8_t *s, int p /* pitch */, const uint8_t *blimit0,
+    const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1,
+    const uint8_t *limit1, const uint8_t *thresh1) {
+  aom_lpf_horizontal_4_dspr2(s, p, blimit0, limit0, thresh0);
+  aom_lpf_horizontal_4_dspr2(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void aom_lpf_horizontal_8_dual_dspr2(
+    uint8_t *s, int p /* pitch */, const uint8_t *blimit0,
+    const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1,
+    const uint8_t *limit1, const uint8_t *thresh1) {
+  aom_lpf_horizontal_8_dspr2(s, p, blimit0, limit0, thresh0);
+  aom_lpf_horizontal_8_dspr2(s + 8, p, blimit1, limit1, thresh1);
+}
+
+void aom_lpf_vertical_4_dual_dspr2(uint8_t *s, int p, const uint8_t *blimit0,
+                                   const uint8_t *limit0,
+                                   const uint8_t *thresh0,
+                                   const uint8_t *blimit1,
+                                   const uint8_t *limit1,
+                                   const uint8_t *thresh1) {
+  aom_lpf_vertical_4_dspr2(s, p, blimit0, limit0, thresh0);
+  aom_lpf_vertical_4_dspr2(s + 8 * p, p, blimit1, limit1, thresh1);
+}
+
+void aom_lpf_vertical_8_dual_dspr2(uint8_t *s, int p, const uint8_t *blimit0,
+                                   const uint8_t *limit0,
+                                   const uint8_t *thresh0,
+                                   const uint8_t *blimit1,
+                                   const uint8_t *limit1,
+                                   const uint8_t *thresh1) {
+  aom_lpf_vertical_8_dspr2(s, p, blimit0, limit0, thresh0);
+  aom_lpf_vertical_8_dspr2(s + 8 * p, p, blimit1, limit1, thresh1);
+}
+
+void aom_lpf_vertical_16_dual_dspr2(uint8_t *s, int p, const uint8_t *blimit,
+                                    const uint8_t *limit,
+                                    const uint8_t *thresh) {
+  aom_lpf_vertical_16_dspr2(s, p, blimit, limit, thresh);
+  aom_lpf_vertical_16_dspr2(s + 8 * p, p, blimit, limit, thresh);
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.h b/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.h
new file mode 100644
index 0000000000..72df098236
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_filters_dspr2.h
@@ -0,0 +1,735 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_MIPS_LOOPFILTER_FILTERS_DSPR2_H_
+#define AOM_DSP_MIPS_LOOPFILTER_FILTERS_DSPR2_H_
+
+#include <stdlib.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if HAVE_DSPR2
+/* inputs & outputs are quad-byte vectors */
+static INLINE void filter_dspr2(uint32_t mask, uint32_t hev, uint32_t *ps1,
+                                uint32_t *ps0, uint32_t *qs0, uint32_t *qs1) {
+  int32_t aom_filter_l, aom_filter_r;
+  int32_t Filter1_l, Filter1_r, Filter2_l, Filter2_r;
+  int32_t subr_r, subr_l;
+  uint32_t t1, t2, HWM, t3;
+  uint32_t hev_l, hev_r, mask_l, mask_r, invhev_l, invhev_r;
+  int32_t vps1, vps0, vqs0, vqs1;
+  int32_t vps1_l, vps1_r, vps0_l, vps0_r, vqs0_l, vqs0_r, vqs1_l, vqs1_r;
+  uint32_t N128;
+
+  N128 = 0x80808080;
+  t1 = 0x03000300;
+  t2 = 0x04000400;
+  t3 = 0x01000100;
+  HWM = 0xFF00FF00;
+
+  vps0 = (*ps0) ^ N128;
+  vps1 = (*ps1) ^ N128;
+  vqs0 = (*qs0) ^ N128;
+  vqs1 = (*qs1) ^ N128;
+
+  /* use halfword pairs instead quad-bytes because of accuracy */
+  vps0_l = vps0 & HWM;
+  vps0_r = vps0 << 8;
+  vps0_r = vps0_r & HWM;
+
+  vps1_l = vps1 & HWM;
+  vps1_r = vps1 << 8;
+  vps1_r = vps1_r & HWM;
+
+  vqs0_l = vqs0 & HWM;
+  vqs0_r = vqs0 << 8;
+  vqs0_r = vqs0_r & HWM;
+
+  vqs1_l = vqs1 & HWM;
+  vqs1_r = vqs1 << 8;
+  vqs1_r = vqs1_r & HWM;
+
+  mask_l = mask & HWM;
+  mask_r = mask << 8;
+  mask_r = mask_r & HWM;
+
+  hev_l = hev & HWM;
+  hev_r = hev << 8;
+  hev_r = hev_r & HWM;
+
+  __asm__ __volatile__(
+      /* aom_filter = aom_signed_char_clamp(ps1 - qs1); */
+      "subq_s.ph    %[aom_filter_l], %[vps1_l],       %[vqs1_l]       \n\t"
+      "subq_s.ph    %[aom_filter_r], %[vps1_r],       %[vqs1_r]       \n\t"
+
+      /* qs0 - ps0 */
+      "subq_s.ph    %[subr_l],       %[vqs0_l],       %[vps0_l]       \n\t"
+      "subq_s.ph    %[subr_r],       %[vqs0_r],       %[vps0_r]       \n\t"
+
+      /* aom_filter &= hev; */
+      "and          %[aom_filter_l], %[aom_filter_l], %[hev_l]        \n\t"
+      "and          %[aom_filter_r], %[aom_filter_r], %[hev_r]        \n\t"
+
+      /* aom_filter = aom_signed_char_clamp(aom_filter + 3 * (qs0 - ps0)); */
+      "addq_s.ph    %[aom_filter_l], %[aom_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[aom_filter_r], %[aom_filter_r], %[subr_r]       \n\t"
+      "xor          %[invhev_l],     %[hev_l],        %[HWM]          \n\t"
+      "addq_s.ph    %[aom_filter_l], %[aom_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[aom_filter_r], %[aom_filter_r], %[subr_r]       \n\t"
+      "xor          %[invhev_r],     %[hev_r],        %[HWM]          \n\t"
+      "addq_s.ph    %[aom_filter_l], %[aom_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[aom_filter_r], %[aom_filter_r], %[subr_r]       \n\t"
+
+      /* aom_filter &= mask; */
+      "and          %[aom_filter_l], %[aom_filter_l], %[mask_l]       \n\t"
+      "and          %[aom_filter_r], %[aom_filter_r], %[mask_r]       \n\t"
+
+      : [aom_filter_l] "=&r"(aom_filter_l), [aom_filter_r] "=&r"(aom_filter_r),
+        [subr_l] "=&r"(subr_l), [subr_r] "=&r"(subr_r),
+        [invhev_l] "=&r"(invhev_l), [invhev_r] "=&r"(invhev_r)
+      : [vps0_l] "r"(vps0_l), [vps0_r] "r"(vps0_r), [vps1_l] "r"(vps1_l),
+        [vps1_r] "r"(vps1_r), [vqs0_l] "r"(vqs0_l), [vqs0_r] "r"(vqs0_r),
+        [vqs1_l] "r"(vqs1_l), [vqs1_r] "r"(vqs1_r), [mask_l] "r"(mask_l),
+        [mask_r] "r"(mask_r), [hev_l] "r"(hev_l), [hev_r] "r"(hev_r),
+        [HWM] "r"(HWM));
+
+  /* save bottom 3 bits so that we round one side +4 and the other +3 */
+  __asm__ __volatile__(
+      /* Filter2 = aom_signed_char_clamp(aom_filter + 3) >>= 3; */
+      "addq_s.ph    %[Filter1_l],    %[aom_filter_l], %[t2]           \n\t"
+      "addq_s.ph    %[Filter1_r],    %[aom_filter_r], %[t2]           \n\t"
+
+      /* Filter1 = aom_signed_char_clamp(aom_filter + 4) >>= 3; */
+      "addq_s.ph    %[Filter2_l],    %[aom_filter_l], %[t1]           \n\t"
+      "addq_s.ph    %[Filter2_r],    %[aom_filter_r], %[t1]           \n\t"
+      "shra.ph      %[Filter1_r],    %[Filter1_r],    3               \n\t"
+      "shra.ph      %[Filter1_l],    %[Filter1_l],    3               \n\t"
+
+      "shra.ph      %[Filter2_l],    %[Filter2_l],    3               \n\t"
+      "shra.ph      %[Filter2_r],    %[Filter2_r],    3               \n\t"
+
+      "and          %[Filter1_l],    %[Filter1_l],    %[HWM]          \n\t"
+      "and          %[Filter1_r],    %[Filter1_r],    %[HWM]          \n\t"
+
+      /* vps0 = aom_signed_char_clamp(ps0 + Filter2); */
+      "addq_s.ph    %[vps0_l],       %[vps0_l],       %[Filter2_l]    \n\t"
+      "addq_s.ph    %[vps0_r],       %[vps0_r],       %[Filter2_r]    \n\t"
+
+      /* vqs0 = aom_signed_char_clamp(qs0 - Filter1); */
+      "subq_s.ph    %[vqs0_l],       %[vqs0_l],       %[Filter1_l]    \n\t"
+      "subq_s.ph    %[vqs0_r],       %[vqs0_r],       %[Filter1_r]    \n\t"
+
+      : [Filter1_l] "=&r"(Filter1_l), [Filter1_r] "=&r"(Filter1_r),
+        [Filter2_l] "=&r"(Filter2_l), [Filter2_r] "=&r"(Filter2_r),
+        [vps0_l] "+r"(vps0_l), [vps0_r] "+r"(vps0_r), [vqs0_l] "+r"(vqs0_l),
+        [vqs0_r] "+r"(vqs0_r)
+      : [t1] "r"(t1), [t2] "r"(t2), [HWM] "r"(HWM),
+        [aom_filter_l] "r"(aom_filter_l), [aom_filter_r] "r"(aom_filter_r));
+
+  __asm__ __volatile__(
+      /* (aom_filter += 1) >>= 1 */
+      "addqh.ph    %[Filter1_l],    %[Filter1_l],     %[t3]           \n\t"
+      "addqh.ph    %[Filter1_r],    %[Filter1_r],     %[t3]           \n\t"
+
+      /* aom_filter &= ~hev; */
+      "and          %[Filter1_l],    %[Filter1_l],    %[invhev_l]     \n\t"
+      "and          %[Filter1_r],    %[Filter1_r],    %[invhev_r]     \n\t"
+
+      /* vps1 = aom_signed_char_clamp(ps1 + aom_filter); */
+      "addq_s.ph    %[vps1_l],       %[vps1_l],       %[Filter1_l]    \n\t"
+      "addq_s.ph    %[vps1_r],       %[vps1_r],       %[Filter1_r]    \n\t"
+
+      /* vqs1 = aom_signed_char_clamp(qs1 - aom_filter); */
+      "subq_s.ph    %[vqs1_l],       %[vqs1_l],       %[Filter1_l]    \n\t"
+      "subq_s.ph    %[vqs1_r],       %[vqs1_r],       %[Filter1_r]    \n\t"
+
+      : [Filter1_l] "+r"(Filter1_l), [Filter1_r] "+r"(Filter1_r),
+        [vps1_l] "+r"(vps1_l), [vps1_r] "+r"(vps1_r), [vqs1_l] "+r"(vqs1_l),
+        [vqs1_r] "+r"(vqs1_r)
+      : [t3] "r"(t3), [invhev_l] "r"(invhev_l), [invhev_r] "r"(invhev_r));
+
+  /* Create quad-bytes from halfword pairs */
+  vqs0_l = vqs0_l & HWM;
+  vqs1_l = vqs1_l & HWM;
+  vps0_l = vps0_l & HWM;
+  vps1_l = vps1_l & HWM;
+
+  __asm__ __volatile__(
+      "shrl.ph      %[vqs0_r],       %[vqs0_r],       8   \n\t"
+      "shrl.ph      %[vps0_r],       %[vps0_r],       8   \n\t"
+      "shrl.ph      %[vqs1_r],       %[vqs1_r],       8   \n\t"
+      "shrl.ph      %[vps1_r],       %[vps1_r],       8   \n\t"
+
+      : [vps1_r] "+r"(vps1_r), [vqs1_r] "+r"(vqs1_r), [vps0_r] "+r"(vps0_r),
+        [vqs0_r] "+r"(vqs0_r)
+      :);
+
+  vqs0 = vqs0_l | vqs0_r;
+  vqs1 = vqs1_l | vqs1_r;
+  vps0 = vps0_l | vps0_r;
+  vps1 = vps1_l | vps1_r;
+
+  *ps0 = vps0 ^ N128;
+  *ps1 = vps1 ^ N128;
+  *qs0 = vqs0 ^ N128;
+  *qs1 = vqs1 ^ N128;
+}
+
+static INLINE void filter1_dspr2(uint32_t mask, uint32_t hev, uint32_t ps1,
+                                 uint32_t ps0, uint32_t qs0, uint32_t qs1,
+                                 uint32_t *p1_f0, uint32_t *p0_f0,
+                                 uint32_t *q0_f0, uint32_t *q1_f0) {
+  int32_t aom_filter_l, aom_filter_r;
+  int32_t Filter1_l, Filter1_r, Filter2_l, Filter2_r;
+  int32_t subr_r, subr_l;
+  uint32_t t1, t2, HWM, t3;
+  uint32_t hev_l, hev_r, mask_l, mask_r, invhev_l, invhev_r;
+  int32_t vps1, vps0, vqs0, vqs1;
+  int32_t vps1_l, vps1_r, vps0_l, vps0_r, vqs0_l, vqs0_r, vqs1_l, vqs1_r;
+  uint32_t N128;
+
+  N128 = 0x80808080;
+  t1 = 0x03000300;
+  t2 = 0x04000400;
+  t3 = 0x01000100;
+  HWM = 0xFF00FF00;
+
+  vps0 = (ps0) ^ N128;
+  vps1 = (ps1) ^ N128;
+  vqs0 = (qs0) ^ N128;
+  vqs1 = (qs1) ^ N128;
+
+  /* use halfword pairs instead quad-bytes because of accuracy */
+  vps0_l = vps0 & HWM;
+  vps0_r = vps0 << 8;
+  vps0_r = vps0_r & HWM;
+
+  vps1_l = vps1 & HWM;
+  vps1_r = vps1 << 8;
+  vps1_r = vps1_r & HWM;
+
+  vqs0_l = vqs0 & HWM;
+  vqs0_r = vqs0 << 8;
+  vqs0_r = vqs0_r & HWM;
+
+  vqs1_l = vqs1 & HWM;
+  vqs1_r = vqs1 << 8;
+  vqs1_r = vqs1_r & HWM;
+
+  mask_l = mask & HWM;
+  mask_r = mask << 8;
+  mask_r = mask_r & HWM;
+
+  hev_l = hev & HWM;
+  hev_r = hev << 8;
+  hev_r = hev_r & HWM;
+
+  __asm__ __volatile__(
+      /* aom_filter = aom_signed_char_clamp(ps1 - qs1); */
+      "subq_s.ph    %[aom_filter_l], %[vps1_l],       %[vqs1_l]       \n\t"
+      "subq_s.ph    %[aom_filter_r], %[vps1_r],       %[vqs1_r]       \n\t"
+
+      /* qs0 - ps0 */
+      "subq_s.ph    %[subr_l],       %[vqs0_l],       %[vps0_l]       \n\t"
+      "subq_s.ph    %[subr_r],       %[vqs0_r],       %[vps0_r]       \n\t"
+
+      /* aom_filter &= hev; */
+      "and          %[aom_filter_l], %[aom_filter_l], %[hev_l]        \n\t"
+      "and          %[aom_filter_r], %[aom_filter_r], %[hev_r]        \n\t"
+
+      /* aom_filter = aom_signed_char_clamp(aom_filter + 3 * (qs0 - ps0)); */
+      "addq_s.ph    %[aom_filter_l], %[aom_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[aom_filter_r], %[aom_filter_r], %[subr_r]       \n\t"
+      "xor          %[invhev_l],     %[hev_l],        %[HWM]          \n\t"
+      "addq_s.ph    %[aom_filter_l], %[aom_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[aom_filter_r], %[aom_filter_r], %[subr_r]       \n\t"
+      "xor          %[invhev_r],     %[hev_r],        %[HWM]          \n\t"
+      "addq_s.ph    %[aom_filter_l], %[aom_filter_l], %[subr_l]       \n\t"
+      "addq_s.ph    %[aom_filter_r], %[aom_filter_r], %[subr_r]       \n\t"
+
+      /* aom_filter &= mask; */
+      "and          %[aom_filter_l], %[aom_filter_l], %[mask_l]       \n\t"
+      "and          %[aom_filter_r], %[aom_filter_r], %[mask_r]       \n\t"
+
+      : [aom_filter_l] "=&r"(aom_filter_l), [aom_filter_r] "=&r"(aom_filter_r),
+        [subr_l] "=&r"(subr_l), [subr_r] "=&r"(subr_r),
+        [invhev_l] "=&r"(invhev_l), [invhev_r] "=&r"(invhev_r)
+      : [vps0_l] "r"(vps0_l), [vps0_r] "r"(vps0_r), [vps1_l] "r"(vps1_l),
+        [vps1_r] "r"(vps1_r), [vqs0_l] "r"(vqs0_l), [vqs0_r] "r"(vqs0_r),
+        [vqs1_l] "r"(vqs1_l), [vqs1_r] "r"(vqs1_r), [mask_l] "r"(mask_l),
+        [mask_r] "r"(mask_r), [hev_l] "r"(hev_l), [hev_r] "r"(hev_r),
+        [HWM] "r"(HWM));
+
+  /* save bottom 3 bits so that we round one side +4 and the other +3 */
+  __asm__ __volatile__(
+      /* Filter2 = aom_signed_char_clamp(aom_filter + 3) >>= 3; */
+      "addq_s.ph    %[Filter1_l],    %[aom_filter_l], %[t2]           \n\t"
+      "addq_s.ph    %[Filter1_r],    %[aom_filter_r], %[t2]           \n\t"
+
+      /* Filter1 = aom_signed_char_clamp(aom_filter + 4) >>= 3; */
+      "addq_s.ph    %[Filter2_l],    %[aom_filter_l], %[t1]           \n\t"
+      "addq_s.ph    %[Filter2_r],    %[aom_filter_r], %[t1]           \n\t"
+      "shra.ph      %[Filter1_r],    %[Filter1_r],    3               \n\t"
+      "shra.ph      %[Filter1_l],    %[Filter1_l],    3               \n\t"
+
+      "shra.ph      %[Filter2_l],    %[Filter2_l],    3               \n\t"
+      "shra.ph      %[Filter2_r],    %[Filter2_r],    3               \n\t"
+
+      "and          %[Filter1_l],    %[Filter1_l],    %[HWM]          \n\t"
+      "and          %[Filter1_r],    %[Filter1_r],    %[HWM]          \n\t"
+
+      /* vps0 = aom_signed_char_clamp(ps0 + Filter2); */
+      "addq_s.ph    %[vps0_l],       %[vps0_l],       %[Filter2_l]    \n\t"
+      "addq_s.ph    %[vps0_r],       %[vps0_r],       %[Filter2_r]    \n\t"
+
+      /* vqs0 = aom_signed_char_clamp(qs0 - Filter1); */
+      "subq_s.ph    %[vqs0_l],       %[vqs0_l],       %[Filter1_l]    \n\t"
+      "subq_s.ph    %[vqs0_r],       %[vqs0_r],       %[Filter1_r]    \n\t"
+
+      : [Filter1_l] "=&r"(Filter1_l), [Filter1_r] "=&r"(Filter1_r),
+        [Filter2_l] "=&r"(Filter2_l), [Filter2_r] "=&r"(Filter2_r),
+        [vps0_l] "+r"(vps0_l), [vps0_r] "+r"(vps0_r), [vqs0_l] "+r"(vqs0_l),
+        [vqs0_r] "+r"(vqs0_r)
+      : [t1] "r"(t1), [t2] "r"(t2), [HWM] "r"(HWM),
+        [aom_filter_l] "r"(aom_filter_l), [aom_filter_r] "r"(aom_filter_r));
+
+  __asm__ __volatile__(
+      /* (aom_filter += 1) >>= 1 */
+      "addqh.ph    %[Filter1_l],    %[Filter1_l],     %[t3]           \n\t"
+      "addqh.ph    %[Filter1_r],    %[Filter1_r],     %[t3]           \n\t"
+
+      /* aom_filter &= ~hev; */
+      "and          %[Filter1_l],    %[Filter1_l],    %[invhev_l]     \n\t"
+      "and          %[Filter1_r],    %[Filter1_r],    %[invhev_r]     \n\t"
+
+      /* vps1 = aom_signed_char_clamp(ps1 + aom_filter); */
+      "addq_s.ph    %[vps1_l],       %[vps1_l],       %[Filter1_l]    \n\t"
+      "addq_s.ph    %[vps1_r],       %[vps1_r],       %[Filter1_r]    \n\t"
+
+      /* vqs1 = aom_signed_char_clamp(qs1 - aom_filter); */
+      "subq_s.ph    %[vqs1_l],       %[vqs1_l],       %[Filter1_l]    \n\t"
+      "subq_s.ph    %[vqs1_r],       %[vqs1_r],       %[Filter1_r]    \n\t"
+
+      : [Filter1_l] "+r"(Filter1_l), [Filter1_r] "+r"(Filter1_r),
+        [vps1_l] "+r"(vps1_l), [vps1_r] "+r"(vps1_r), [vqs1_l] "+r"(vqs1_l),
+        [vqs1_r] "+r"(vqs1_r)
+      : [t3] "r"(t3), [invhev_l] "r"(invhev_l), [invhev_r] "r"(invhev_r));
+
+  /* Create quad-bytes from halfword pairs */
+  vqs0_l = vqs0_l & HWM;
+  vqs1_l = vqs1_l & HWM;
+  vps0_l = vps0_l & HWM;
+  vps1_l = vps1_l & HWM;
+
+  __asm__ __volatile__(
+      "shrl.ph      %[vqs0_r],       %[vqs0_r],       8   \n\t"
+      "shrl.ph      %[vps0_r],       %[vps0_r],       8   \n\t"
+      "shrl.ph      %[vqs1_r],       %[vqs1_r],       8   \n\t"
+      "shrl.ph      %[vps1_r],       %[vps1_r],       8   \n\t"
+
+      : [vps1_r] "+r"(vps1_r), [vqs1_r] "+r"(vqs1_r), [vps0_r] "+r"(vps0_r),
+        [vqs0_r] "+r"(vqs0_r)
+      :);
+
+  vqs0 = vqs0_l | vqs0_r;
+  vqs1 = vqs1_l | vqs1_r;
+  vps0 = vps0_l | vps0_r;
+  vps1 = vps1_l | vps1_r;
+
+  *p0_f0 = vps0 ^ N128;
+  *p1_f0 = vps1 ^ N128;
+  *q0_f0 = vqs0 ^ N128;
+  *q1_f0 = vqs1 ^ N128;
+}
+
+static INLINE void mbfilter_dspr2(uint32_t *op3, uint32_t *op2, uint32_t *op1,
+                                  uint32_t *op0, uint32_t *oq0, uint32_t *oq1,
+                                  uint32_t *oq2, uint32_t *oq3) {
+  /* use a 7 tap filter [1, 1, 1, 2, 1, 1, 1] for flat line */
+  const uint32_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+  const uint32_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+  uint32_t res_op2, res_op1, res_op0;
+  uint32_t res_oq0, res_oq1, res_oq2;
+  uint32_t tmp;
+  uint32_t add_p210_q012;
+  uint32_t u32Four = 0x00040004;
+
+  /* *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + p2 + p2 + p1 + p0 + q0, 3)  1 */
+  /* *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + p1 + p1 + p0 + q0 + q1, 3)  2 */
+  /* *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + p0 + q0 + q1 + q2, 3)  3 */
+  /* *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + q0 + q0 + q1 + q2 + q3, 3)  4 */
+  /* *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q1 + q2 + q3 + q3, 3)  5 */
+  /* *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q2 + q3 + q3 + q3, 3)  6 */
+
+  __asm__ __volatile__(
+      "addu.ph    %[add_p210_q012],  %[p2],             %[p1]            \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[p0]            \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q0]            \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q1]            \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q2]            \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[u32Four]       \n\t"
+
+      "shll.ph    %[tmp],            %[p3],             1                \n\t"
+      "addu.ph    %[res_op2],        %[tmp],            %[p3]            \n\t"
+      "addu.ph    %[res_op1],        %[p3],             %[p3]            \n\t"
+      "addu.ph    %[res_op2],        %[res_op2],        %[p2]            \n\t"
+      "addu.ph    %[res_op1],        %[res_op1],        %[p1]            \n\t"
+      "addu.ph    %[res_op2],        %[res_op2],        %[add_p210_q012] \n\t"
+      "addu.ph    %[res_op1],        %[res_op1],        %[add_p210_q012] \n\t"
+      "subu.ph    %[res_op2],        %[res_op2],        %[q1]            \n\t"
+      "subu.ph    %[res_op1],        %[res_op1],        %[q2]            \n\t"
+      "subu.ph    %[res_op2],        %[res_op2],        %[q2]            \n\t"
+      "shrl.ph    %[res_op1],        %[res_op1],        3                \n\t"
+      "shrl.ph    %[res_op2],        %[res_op2],        3                \n\t"
+      "addu.ph    %[res_op0],        %[p3],             %[p0]            \n\t"
+      "addu.ph    %[res_oq0],        %[q0],             %[q3]            \n\t"
+      "addu.ph    %[res_op0],        %[res_op0],        %[add_p210_q012] \n\t"
+      "addu.ph    %[res_oq0],        %[res_oq0],        %[add_p210_q012] \n\t"
+      "addu.ph    %[res_oq1],        %[q3],             %[q3]            \n\t"
+      "shll.ph    %[tmp],            %[q3],             1                \n\t"
+      "addu.ph    %[res_oq1],        %[res_oq1],        %[q1]            \n\t"
+      "addu.ph    %[res_oq2],        %[tmp],            %[q3]            \n\t"
+      "addu.ph    %[res_oq1],        %[res_oq1],        %[add_p210_q012] \n\t"
+      "addu.ph    %[res_oq2],        %[res_oq2],        %[add_p210_q012] \n\t"
+      "subu.ph    %[res_oq1],        %[res_oq1],        %[p2]            \n\t"
+      "addu.ph    %[res_oq2],        %[res_oq2],        %[q2]            \n\t"
+      "shrl.ph    %[res_oq1],        %[res_oq1],        3                \n\t"
+      "subu.ph    %[res_oq2],        %[res_oq2],        %[p2]            \n\t"
+      "shrl.ph    %[res_oq0],        %[res_oq0],        3                \n\t"
+      "subu.ph    %[res_oq2],        %[res_oq2],        %[p1]            \n\t"
+      "shrl.ph    %[res_op0],        %[res_op0],        3                \n\t"
+      "shrl.ph    %[res_oq2],        %[res_oq2],        3                \n\t"
+
+      : [add_p210_q012] "=&r"(add_p210_q012), [tmp] "=&r"(tmp),
+        [res_op2] "=&r"(res_op2), [res_op1] "=&r"(res_op1),
+        [res_op0] "=&r"(res_op0), [res_oq0] "=&r"(res_oq0),
+        [res_oq1] "=&r"(res_oq1), [res_oq2] "=&r"(res_oq2)
+      : [p0] "r"(p0), [q0] "r"(q0), [p1] "r"(p1), [q1] "r"(q1), [p2] "r"(p2),
+        [q2] "r"(q2), [p3] "r"(p3), [q3] "r"(q3), [u32Four] "r"(u32Four));
+
+  *op2 = res_op2;
+  *op1 = res_op1;
+  *op0 = res_op0;
+  *oq0 = res_oq0;
+  *oq1 = res_oq1;
+  *oq2 = res_oq2;
+}
+
+static INLINE void mbfilter1_dspr2(uint32_t p3, uint32_t p2, uint32_t p1,
+                                   uint32_t p0, uint32_t q0, uint32_t q1,
+                                   uint32_t q2, uint32_t q3, uint32_t *op2_f1,
+                                   uint32_t *op1_f1, uint32_t *op0_f1,
+                                   uint32_t *oq0_f1, uint32_t *oq1_f1,
+                                   uint32_t *oq2_f1) {
+  /* use a 7 tap filter [1, 1, 1, 2, 1, 1, 1] for flat line */
+  uint32_t res_op2, res_op1, res_op0;
+  uint32_t res_oq0, res_oq1, res_oq2;
+  uint32_t tmp;
+  uint32_t add_p210_q012;
+  uint32_t u32Four = 0x00040004;
+
+  /* *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + p2 + p2 + p1 + p0 + q0, 3)   1 */
+  /* *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + p1 + p1 + p0 + q0 + q1, 3)   2 */
+  /* *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + p0 + q0 + q1 + q2, 3)   3 */
+  /* *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + q0 + q0 + q1 + q2 + q3, 3)   4 */
+  /* *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q1 + q2 + q3 + q3, 3)   5 */
+  /* *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q2 + q3 + q3 + q3, 3)   6 */
+
+  __asm__ __volatile__(
+      "addu.ph    %[add_p210_q012],  %[p2],             %[p1]             \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[p0]             \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q0]             \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q1]             \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[q2]             \n\t"
+      "addu.ph    %[add_p210_q012],  %[add_p210_q012],  %[u32Four]        \n\t"
+
+      "shll.ph    %[tmp],            %[p3],             1                 \n\t"
+      "addu.ph    %[res_op2],        %[tmp],            %[p3]             \n\t"
+      "addu.ph    %[res_op1],        %[p3],             %[p3]             \n\t"
+      "addu.ph    %[res_op2],        %[res_op2],        %[p2]             \n\t"
+      "addu.ph    %[res_op1],        %[res_op1],        %[p1]             \n\t"
+      "addu.ph    %[res_op2],        %[res_op2],        %[add_p210_q012]  \n\t"
+      "addu.ph    %[res_op1],        %[res_op1],        %[add_p210_q012]  \n\t"
+      "subu.ph    %[res_op2],        %[res_op2],        %[q1]             \n\t"
+      "subu.ph    %[res_op1],        %[res_op1],        %[q2]             \n\t"
+      "subu.ph    %[res_op2],        %[res_op2],        %[q2]             \n\t"
+      "shrl.ph    %[res_op1],        %[res_op1],        3                 \n\t"
+      "shrl.ph    %[res_op2],        %[res_op2],        3                 \n\t"
+      "addu.ph    %[res_op0],        %[p3],             %[p0]             \n\t"
+      "addu.ph    %[res_oq0],        %[q0],             %[q3]             \n\t"
+      "addu.ph    %[res_op0],        %[res_op0],        %[add_p210_q012]  \n\t"
+      "addu.ph    %[res_oq0],        %[res_oq0],        %[add_p210_q012]  \n\t"
+      "addu.ph    %[res_oq1],        %[q3],             %[q3]             \n\t"
+      "shll.ph    %[tmp],            %[q3],             1                 \n\t"
+      "addu.ph    %[res_oq1],        %[res_oq1],        %[q1]             \n\t"
+      "addu.ph    %[res_oq2],        %[tmp],            %[q3]             \n\t"
+      "addu.ph    %[res_oq1],        %[res_oq1],        %[add_p210_q012]  \n\t"
+      "addu.ph    %[res_oq2],        %[res_oq2],        %[add_p210_q012]  \n\t"
+      "subu.ph    %[res_oq1],        %[res_oq1],        %[p2]             \n\t"
+      "addu.ph    %[res_oq2],        %[res_oq2],        %[q2]             \n\t"
+      "shrl.ph    %[res_oq1],        %[res_oq1],        3                 \n\t"
+      "subu.ph    %[res_oq2],        %[res_oq2],        %[p2]             \n\t"
+      "shrl.ph    %[res_oq0],        %[res_oq0],        3                 \n\t"
+      "subu.ph    %[res_oq2],        %[res_oq2],        %[p1]             \n\t"
+      "shrl.ph    %[res_op0],        %[res_op0],        3                 \n\t"
+      "shrl.ph    %[res_oq2],        %[res_oq2],        3                 \n\t"
+
+      : [add_p210_q012] "=&r"(add_p210_q012), [tmp] "=&r"(tmp),
+        [res_op2] "=&r"(res_op2), [res_op1] "=&r"(res_op1),
+        [res_op0] "=&r"(res_op0), [res_oq0] "=&r"(res_oq0),
+        [res_oq1] "=&r"(res_oq1), [res_oq2] "=&r"(res_oq2)
+      : [p0] "r"(p0), [q0] "r"(q0), [p1] "r"(p1), [q1] "r"(q1), [p2] "r"(p2),
+        [q2] "r"(q2), [p3] "r"(p3), [q3] "r"(q3), [u32Four] "r"(u32Four));
+
+  *op2_f1 = res_op2;
+  *op1_f1 = res_op1;
+  *op0_f1 = res_op0;
+  *oq0_f1 = res_oq0;
+  *oq1_f1 = res_oq1;
+  *oq2_f1 = res_oq2;
+}
+
+static INLINE void wide_mbfilter_dspr2(
+    uint32_t *op7, uint32_t *op6, uint32_t *op5, uint32_t *op4, uint32_t *op3,
+    uint32_t *op2, uint32_t *op1, uint32_t *op0, uint32_t *oq0, uint32_t *oq1,
+    uint32_t *oq2, uint32_t *oq3, uint32_t *oq4, uint32_t *oq5, uint32_t *oq6,
+    uint32_t *oq7) {
+  const uint32_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4;
+  const uint32_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+  const uint32_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+  const uint32_t q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7;
+  uint32_t res_op6, res_op5, res_op4, res_op3, res_op2, res_op1, res_op0;
+  uint32_t res_oq0, res_oq1, res_oq2, res_oq3, res_oq4, res_oq5, res_oq6;
+  uint32_t tmp;
+  uint32_t add_p6toq6;
+  uint32_t u32Eight = 0x00080008;
+
+  __asm__ __volatile__(
+      /* addition of p6,p5,p4,p3,p2,p1,p0,q0,q1,q2,q3,q4,q5,q6
+         which is used most of the time */
+      "addu.ph      %[add_p6toq6],     %[p6],              %[p5]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[p4]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[p3]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[p2]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[p1]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[p0]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q0]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q1]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q2]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q3]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q4]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q5]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[q6]         \n\t"
+      "addu.ph      %[add_p6toq6],     %[add_p6toq6],      %[u32Eight]   \n\t"
+
+      : [add_p6toq6] "=&r"(add_p6toq6)
+      : [p6] "r"(p6), [p5] "r"(p5), [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2),
+        [p1] "r"(p1), [p0] "r"(p0), [q0] "r"(q0), [q1] "r"(q1), [q2] "r"(q2),
+        [q3] "r"(q3), [q4] "r"(q4), [q5] "r"(q5), [q6] "r"(q6),
+        [u32Eight] "r"(u32Eight));
+
+  __asm__ __volatile__(
+      /* *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 +
+                                   p3 + p2 + p1 + p0 + q0, 4) */
+      "shll.ph       %[tmp],            %[p7],            3               \n\t"
+      "subu.ph       %[res_op6],        %[tmp],           %[p7]           \n\t"
+      "addu.ph       %[res_op6],        %[res_op6],       %[p6]           \n\t"
+      "addu.ph       %[res_op6],        %[res_op6],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q1]           \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q2]           \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q3]           \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q4]           \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q5]           \n\t"
+      "subu.ph       %[res_op6],        %[res_op6],       %[q6]           \n\t"
+      "shrl.ph       %[res_op6],        %[res_op6],       4               \n\t"
+
+      /* *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 +
+                                   p2 + p1 + p0 + q0 + q1, 4) */
+      "shll.ph       %[tmp],            %[p7],            2               \n\t"
+      "addu.ph       %[res_op5],        %[tmp],           %[p7]           \n\t"
+      "addu.ph       %[res_op5],        %[res_op5],       %[p7]           \n\t"
+      "addu.ph       %[res_op5],        %[res_op5],       %[p5]           \n\t"
+      "addu.ph       %[res_op5],        %[res_op5],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op5],        %[res_op5],       %[q2]           \n\t"
+      "subu.ph       %[res_op5],        %[res_op5],       %[q3]           \n\t"
+      "subu.ph       %[res_op5],        %[res_op5],       %[q4]           \n\t"
+      "subu.ph       %[res_op5],        %[res_op5],       %[q5]           \n\t"
+      "subu.ph       %[res_op5],        %[res_op5],       %[q6]           \n\t"
+      "shrl.ph       %[res_op5],        %[res_op5],       4               \n\t"
+
+      /* *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 +
+                                   p1 + p0 + q0 + q1 + q2, 4) */
+      "shll.ph       %[tmp],            %[p7],            2               \n\t"
+      "addu.ph       %[res_op4],        %[tmp],           %[p7]           \n\t"
+      "addu.ph       %[res_op4],        %[res_op4],       %[p4]           \n\t"
+      "addu.ph       %[res_op4],        %[res_op4],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op4],        %[res_op4],       %[q3]           \n\t"
+      "subu.ph       %[res_op4],        %[res_op4],       %[q4]           \n\t"
+      "subu.ph       %[res_op4],        %[res_op4],       %[q5]           \n\t"
+      "subu.ph       %[res_op4],        %[res_op4],       %[q6]           \n\t"
+      "shrl.ph       %[res_op4],        %[res_op4],       4               \n\t"
+
+      /* *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 +
+                                   p1 + p0 + q0 + q1 + q2 + q3, 4) */
+      "shll.ph       %[tmp],            %[p7],            2               \n\t"
+      "addu.ph       %[res_op3],        %[tmp],           %[p3]           \n\t"
+      "addu.ph       %[res_op3],        %[res_op3],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op3],        %[res_op3],       %[q4]           \n\t"
+      "subu.ph       %[res_op3],        %[res_op3],       %[q5]           \n\t"
+      "subu.ph       %[res_op3],        %[res_op3],       %[q6]           \n\t"
+      "shrl.ph       %[res_op3],        %[res_op3],       4               \n\t"
+
+      /* *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 +
+                                   p0 + q0 + q1 + q2 + q3 + q4, 4) */
+      "shll.ph       %[tmp],            %[p7],            1               \n\t"
+      "addu.ph       %[res_op2],        %[tmp],           %[p7]           \n\t"
+      "addu.ph       %[res_op2],        %[res_op2],       %[p2]           \n\t"
+      "addu.ph       %[res_op2],        %[res_op2],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op2],        %[res_op2],       %[q5]           \n\t"
+      "subu.ph       %[res_op2],        %[res_op2],       %[q6]           \n\t"
+      "shrl.ph       %[res_op2],        %[res_op2],       4               \n\t"
+
+      /* *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 +
+                                   p0 + q0 + q1 + q2 + q3 + q4 + q5, 4); */
+      "shll.ph       %[tmp],            %[p7],            1               \n\t"
+      "addu.ph       %[res_op1],        %[tmp],           %[p1]           \n\t"
+      "addu.ph       %[res_op1],        %[res_op1],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_op1],        %[res_op1],       %[q6]           \n\t"
+      "shrl.ph       %[res_op1],        %[res_op1],       4               \n\t"
+
+      /* *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
+                                  q0 + q1 + q2 + q3 + q4 + q5 + q6, 4) */
+      "addu.ph       %[res_op0],        %[p7],            %[p0]           \n\t"
+      "addu.ph       %[res_op0],        %[res_op0],       %[add_p6toq6]   \n\t"
+      "shrl.ph       %[res_op0],        %[res_op0],       4               \n\t"
+
+      : [res_op6] "=&r"(res_op6), [res_op5] "=&r"(res_op5),
+        [res_op4] "=&r"(res_op4), [res_op3] "=&r"(res_op3),
+        [res_op2] "=&r"(res_op2), [res_op1] "=&r"(res_op1),
+        [res_op0] "=&r"(res_op0), [tmp] "=&r"(tmp)
+      : [p7] "r"(p7), [p6] "r"(p6), [p5] "r"(p5), [p4] "r"(p4), [p3] "r"(p3),
+        [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), [q2] "r"(q2), [q1] "r"(q1),
+        [q3] "r"(q3), [q4] "r"(q4), [q5] "r"(q5), [q6] "r"(q6),
+        [add_p6toq6] "r"(add_p6toq6));
+
+  *op6 = res_op6;
+  *op5 = res_op5;
+  *op4 = res_op4;
+  *op3 = res_op3;
+  *op2 = res_op2;
+  *op1 = res_op1;
+  *op0 = res_op0;
+
+  __asm__ __volatile__(
+      /* *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 + q0 * 2 +
+                                   q1 + q2 + q3 + q4 + q5 + q6 + q7, 4); */
+      "addu.ph       %[res_oq0],        %[q7],            %[q0]           \n\t"
+      "addu.ph       %[res_oq0],        %[res_oq0],       %[add_p6toq6]   \n\t"
+      "shrl.ph       %[res_oq0],        %[res_oq0],       4               \n\t"
+
+      /* *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 + q0 + q1 * 2 +
+                                   q2 + q3 + q4 + q5 + q6 + q7 * 2, 4) */
+      "shll.ph       %[tmp],            %[q7],            1               \n\t"
+      "addu.ph       %[res_oq1],        %[tmp],           %[q1]           \n\t"
+      "addu.ph       %[res_oq1],        %[res_oq1],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq1],        %[res_oq1],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq1],        %[res_oq1],       4               \n\t"
+
+      /* *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 + q0 + q1 + q2 * 2 +
+                                   q3 + q4 + q5 + q6 + q7 * 3, 4) */
+      "shll.ph       %[tmp],            %[q7],            1               \n\t"
+      "addu.ph       %[res_oq2],        %[tmp],           %[q7]           \n\t"
+      "addu.ph       %[res_oq2],        %[res_oq2],       %[q2]           \n\t"
+      "addu.ph       %[res_oq2],        %[res_oq2],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq2],        %[res_oq2],       %[p5]           \n\t"
+      "subu.ph       %[res_oq2],        %[res_oq2],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq2],        %[res_oq2],       4               \n\t"
+
+      /* *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + q0 + q1 + q2 +
+                                   q3 * 2 + q4 + q5 + q6 + q7 * 4, 4) */
+      "shll.ph       %[tmp],            %[q7],            2               \n\t"
+      "addu.ph       %[res_oq3],        %[tmp],           %[q3]           \n\t"
+      "addu.ph       %[res_oq3],        %[res_oq3],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq3],        %[res_oq3],       %[p4]           \n\t"
+      "subu.ph       %[res_oq3],        %[res_oq3],       %[p5]           \n\t"
+      "subu.ph       %[res_oq3],        %[res_oq3],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq3],        %[res_oq3],       4               \n\t"
+
+      /* *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + q0 + q1 + q2 + q3 +
+                                   q4 * 2 + q5 + q6 + q7 * 5, 4) */
+      "shll.ph       %[tmp],            %[q7],            2               \n\t"
+      "addu.ph       %[res_oq4],        %[tmp],           %[q7]           \n\t"
+      "addu.ph       %[res_oq4],        %[res_oq4],       %[q4]           \n\t"
+      "addu.ph       %[res_oq4],        %[res_oq4],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq4],        %[res_oq4],       %[p3]           \n\t"
+      "subu.ph       %[res_oq4],        %[res_oq4],       %[p4]           \n\t"
+      "subu.ph       %[res_oq4],        %[res_oq4],       %[p5]           \n\t"
+      "subu.ph       %[res_oq4],        %[res_oq4],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq4],        %[res_oq4],       4               \n\t"
+
+      /* *oq5 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + q1 + q2 + q3 + q4 +
+                                   q5 * 2 + q6 + q7 * 6, 4) */
+      "shll.ph       %[tmp],            %[q7],            2               \n\t"
+      "addu.ph       %[res_oq5],        %[tmp],           %[q7]           \n\t"
+      "addu.ph       %[res_oq5],        %[res_oq5],       %[q7]           \n\t"
+      "addu.ph       %[res_oq5],        %[res_oq5],       %[q5]           \n\t"
+      "addu.ph       %[res_oq5],        %[res_oq5],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq5],        %[res_oq5],       %[p2]           \n\t"
+      "subu.ph       %[res_oq5],        %[res_oq5],       %[p3]           \n\t"
+      "subu.ph       %[res_oq5],        %[res_oq5],       %[p4]           \n\t"
+      "subu.ph       %[res_oq5],        %[res_oq5],       %[p5]           \n\t"
+      "subu.ph       %[res_oq5],        %[res_oq5],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq5],        %[res_oq5],       4               \n\t"
+
+      /* *oq6 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + q2 + q3 +
+                                   q4 + q5 + q6 * 2 + q7 * 7, 4) */
+      "shll.ph       %[tmp],            %[q7],            3               \n\t"
+      "subu.ph       %[res_oq6],        %[tmp],           %[q7]           \n\t"
+      "addu.ph       %[res_oq6],        %[res_oq6],       %[q6]           \n\t"
+      "addu.ph       %[res_oq6],        %[res_oq6],       %[add_p6toq6]   \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p1]           \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p2]           \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p3]           \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p4]           \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p5]           \n\t"
+      "subu.ph       %[res_oq6],        %[res_oq6],       %[p6]           \n\t"
+      "shrl.ph       %[res_oq6],        %[res_oq6],       4               \n\t"
+
+      : [res_oq6] "=&r"(res_oq6), [res_oq5] "=&r"(res_oq5),
+        [res_oq4] "=&r"(res_oq4), [res_oq3] "=&r"(res_oq3),
+        [res_oq2] "=&r"(res_oq2), [res_oq1] "=&r"(res_oq1),
+        [res_oq0] "=&r"(res_oq0), [tmp] "=&r"(tmp)
+      : [q7] "r"(q7), [q6] "r"(q6), [q5] "r"(q5), [q4] "r"(q4), [q3] "r"(q3),
+        [q2] "r"(q2), [q1] "r"(q1), [q0] "r"(q0), [p1] "r"(p1), [p2] "r"(p2),
+        [p3] "r"(p3), [p4] "r"(p4), [p5] "r"(p5), [p6] "r"(p6),
+        [add_p6toq6] "r"(add_p6toq6));
+
+  *oq0 = res_oq0;
+  *oq1 = res_oq1;
+  *oq2 = res_oq2;
+  *oq3 = res_oq3;
+  *oq4 = res_oq4;
+  *oq5 = res_oq5;
+  *oq6 = res_oq6;
+}
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_MIPS_LOOPFILTER_FILTERS_DSPR2_H_
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_macros_dspr2.h b/third_party/aom/aom_dsp/mips/loopfilter_macros_dspr2.h
new file mode 100644
index 0000000000..3e69947144
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_macros_dspr2.h
@@ -0,0 +1,436 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_MIPS_LOOPFILTER_MACROS_DSPR2_H_
+#define AOM_DSP_MIPS_LOOPFILTER_MACROS_DSPR2_H_
+
+#include <stdlib.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_mem/aom_mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if HAVE_DSPR2
+#define STORE_F0()                                                       \
+  {                                                                      \
+    __asm__ __volatile__(                                                \
+        "sb     %[q1_f0],    1(%[s4])           \n\t"                    \
+        "sb     %[q0_f0],    0(%[s4])           \n\t"                    \
+        "sb     %[p0_f0],   -1(%[s4])           \n\t"                    \
+        "sb     %[p1_f0],   -2(%[s4])           \n\t"                    \
+                                                                         \
+        :                                                                \
+        : [q1_f0] "r"(q1_f0), [q0_f0] "r"(q0_f0), [p0_f0] "r"(p0_f0),    \
+          [p1_f0] "r"(p1_f0), [s4] "r"(s4));                             \
+                                                                         \
+    __asm__ __volatile__(                                                \
+        "srl    %[q1_f0],   %[q1_f0],   8       \n\t"                    \
+        "srl    %[q0_f0],   %[q0_f0],   8       \n\t"                    \
+        "srl    %[p0_f0],   %[p0_f0],   8       \n\t"                    \
+        "srl    %[p1_f0],   %[p1_f0],   8       \n\t"                    \
+                                                                         \
+        : [q1_f0] "+r"(q1_f0), [q0_f0] "+r"(q0_f0), [p0_f0] "+r"(p0_f0), \
+          [p1_f0] "+r"(p1_f0)                                            \
+        :);                                                              \
+                                                                         \
+    __asm__ __volatile__(                                                \
+        "sb     %[q1_f0],    1(%[s3])           \n\t"                    \
+        "sb     %[q0_f0],    0(%[s3])           \n\t"                    \
+        "sb     %[p0_f0],   -1(%[s3])           \n\t"                    \
+        "sb     %[p1_f0],   -2(%[s3])           \n\t"                    \
+                                                                         \
+        : [p1_f0] "+r"(p1_f0)                                            \
+        : [q1_f0] "r"(q1_f0), [q0_f0] "r"(q0_f0), [s3] "r"(s3),          \
+          [p0_f0] "r"(p0_f0));                                           \
+                                                                         \
+    __asm__ __volatile__(                                                \
+        "srl    %[q1_f0],   %[q1_f0],   8       \n\t"                    \
+        "srl    %[q0_f0],   %[q0_f0],   8       \n\t"                    \
+        "srl    %[p0_f0],   %[p0_f0],   8       \n\t"                    \
+        "srl    %[p1_f0],   %[p1_f0],   8       \n\t"                    \
+                                                                         \
+        : [q1_f0] "+r"(q1_f0), [q0_f0] "+r"(q0_f0), [p0_f0] "+r"(p0_f0), \
+          [p1_f0] "+r"(p1_f0)                                            \
+        :);                                                              \
+                                                                         \
+    __asm__ __volatile__(                                                \
+        "sb     %[q1_f0],    1(%[s2])           \n\t"                    \
+        "sb     %[q0_f0],    0(%[s2])           \n\t"                    \
+        "sb     %[p0_f0],   -1(%[s2])           \n\t"                    \
+        "sb     %[p1_f0],   -2(%[s2])           \n\t"                    \
+                                                                         \
+        :                                                                \
+        : [q1_f0] "r"(q1_f0), [q0_f0] "r"(q0_f0), [p0_f0] "r"(p0_f0),    \
+          [p1_f0] "r"(p1_f0), [s2] "r"(s2));                             \
+                                                                         \
+    __asm__ __volatile__(                                                \
+        "srl    %[q1_f0],   %[q1_f0],   8       \n\t"                    \
+        "srl    %[q0_f0],   %[q0_f0],   8       \n\t"                    \
+        "srl    %[p0_f0],   %[p0_f0],   8       \n\t"                    \
+        "srl    %[p1_f0],   %[p1_f0],   8       \n\t"                    \
+                                                                         \
+        : [q1_f0] "+r"(q1_f0), [q0_f0] "+r"(q0_f0), [p0_f0] "+r"(p0_f0), \
+          [p1_f0] "+r"(p1_f0)                                            \
+        :);                                                              \
+                                                                         \
+    __asm__ __volatile__(                                                \
+        "sb     %[q1_f0],    1(%[s1])           \n\t"                    \
+        "sb     %[q0_f0],    0(%[s1])           \n\t"                    \
+        "sb     %[p0_f0],   -1(%[s1])           \n\t"                    \
+        "sb     %[p1_f0],   -2(%[s1])           \n\t"                    \
+                                                                         \
+        :                                                                \
+        : [q1_f0] "r"(q1_f0), [q0_f0] "r"(q0_f0), [p0_f0] "r"(p0_f0),    \
+          [p1_f0] "r"(p1_f0), [s1] "r"(s1));                             \
+  }
+
+#define STORE_F1()                                                             \
+  {                                                                            \
+    __asm__ __volatile__(                                                      \
+        "sb     %[q2_r],     2(%[s4])           \n\t"                          \
+        "sb     %[q1_r],     1(%[s4])           \n\t"                          \
+        "sb     %[q0_r],     0(%[s4])           \n\t"                          \
+        "sb     %[p0_r],    -1(%[s4])           \n\t"                          \
+        "sb     %[p1_r],    -2(%[s4])           \n\t"                          \
+        "sb     %[p2_r],    -3(%[s4])           \n\t"                          \
+                                                                               \
+        :                                                                      \
+        : [q2_r] "r"(q2_r), [q1_r] "r"(q1_r), [q0_r] "r"(q0_r),                \
+          [p0_r] "r"(p0_r), [p1_r] "r"(p1_r), [p2_r] "r"(p2_r), [s4] "r"(s4)); \
+                                                                               \
+    __asm__ __volatile__(                                                      \
+        "srl    %[q2_r],    %[q2_r],    16      \n\t"                          \
+        "srl    %[q1_r],    %[q1_r],    16      \n\t"                          \
+        "srl    %[q0_r],    %[q0_r],    16      \n\t"                          \
+        "srl    %[p0_r],    %[p0_r],    16      \n\t"                          \
+        "srl    %[p1_r],    %[p1_r],    16      \n\t"                          \
+        "srl    %[p2_r],    %[p2_r],    16      \n\t"                          \
+                                                                               \
+        : [q2_r] "+r"(q2_r), [q1_r] "+r"(q1_r), [q0_r] "+r"(q0_r),             \
+          [p0_r] "+r"(p0_r), [p1_r] "+r"(p1_r), [p2_r] "+r"(p2_r)              \
+        :);                                                                    \
+                                                                               \
+    __asm__ __volatile__(                                                      \
+        "sb     %[q2_r],     2(%[s3])           \n\t"                          \
+        "sb     %[q1_r],     1(%[s3])           \n\t"                          \
+        "sb     %[q0_r],     0(%[s3])           \n\t"                          \
+        "sb     %[p0_r],    -1(%[s3])           \n\t"                          \
+        "sb     %[p1_r],    -2(%[s3])           \n\t"                          \
+        "sb     %[p2_r],    -3(%[s3])           \n\t"                          \
+                                                                               \
+        :                                                                      \
+        : [q2_r] "r"(q2_r), [q1_r] "r"(q1_r), [q0_r] "r"(q0_r),                \
+          [p0_r] "r"(p0_r), [p1_r] "r"(p1_r), [p2_r] "r"(p2_r), [s3] "r"(s3)); \
+                                                                               \
+    __asm__ __volatile__(                                                      \
+        "sb     %[q2_l],     2(%[s2])           \n\t"                          \
+        "sb     %[q1_l],     1(%[s2])           \n\t"                          \
+        "sb     %[q0_l],     0(%[s2])           \n\t"                          \
+        "sb     %[p0_l],    -1(%[s2])           \n\t"                          \
+        "sb     %[p1_l],    -2(%[s2])           \n\t"                          \
+        "sb     %[p2_l],    -3(%[s2])           \n\t"                          \
+                                                                               \
+        :                                                                      \
+        : [q2_l] "r"(q2_l), [q1_l] "r"(q1_l), [q0_l] "r"(q0_l),                \
+          [p0_l] "r"(p0_l), [p1_l] "r"(p1_l), [p2_l] "r"(p2_l), [s2] "r"(s2)); \
+                                                                               \
+    __asm__ __volatile__(                                                      \
+        "srl    %[q2_l],    %[q2_l],    16      \n\t"                          \
+        "srl    %[q1_l],    %[q1_l],    16      \n\t"                          \
+        "srl    %[q0_l],    %[q0_l],    16      \n\t"                          \
+        "srl    %[p0_l],    %[p0_l],    16      \n\t"                          \
+        "srl    %[p1_l],    %[p1_l],    16      \n\t"                          \
+        "srl    %[p2_l],    %[p2_l],    16      \n\t"                          \
+                                                                               \
+        : [q2_l] "+r"(q2_l), [q1_l] "+r"(q1_l), [q0_l] "+r"(q0_l),             \
+          [p0_l] "+r"(p0_l), [p1_l] "+r"(p1_l), [p2_l] "+r"(p2_l)              \
+        :);                                                                    \
+                                                                               \
+    __asm__ __volatile__(                                                      \
+        "sb     %[q2_l],     2(%[s1])           \n\t"                          \
+        "sb     %[q1_l],     1(%[s1])           \n\t"                          \
+        "sb     %[q0_l],     0(%[s1])           \n\t"                          \
+        "sb     %[p0_l],    -1(%[s1])           \n\t"                          \
+        "sb     %[p1_l],    -2(%[s1])           \n\t"                          \
+        "sb     %[p2_l],    -3(%[s1])           \n\t"                          \
+                                                                               \
+        :                                                                      \
+        : [q2_l] "r"(q2_l), [q1_l] "r"(q1_l), [q0_l] "r"(q0_l),                \
+          [p0_l] "r"(p0_l), [p1_l] "r"(p1_l), [p2_l] "r"(p2_l), [s1] "r"(s1)); \
+  }
+
+#define STORE_F2()                                                 \
+  {                                                                \
+    __asm__ __volatile__(                                          \
+        "sb     %[q6_r],     6(%[s4])           \n\t"              \
+        "sb     %[q5_r],     5(%[s4])           \n\t"              \
+        "sb     %[q4_r],     4(%[s4])           \n\t"              \
+        "sb     %[q3_r],     3(%[s4])           \n\t"              \
+        "sb     %[q2_r],     2(%[s4])           \n\t"              \
+        "sb     %[q1_r],     1(%[s4])           \n\t"              \
+        "sb     %[q0_r],     0(%[s4])           \n\t"              \
+        "sb     %[p0_r],    -1(%[s4])           \n\t"              \
+        "sb     %[p1_r],    -2(%[s4])           \n\t"              \
+        "sb     %[p2_r],    -3(%[s4])           \n\t"              \
+        "sb     %[p3_r],    -4(%[s4])           \n\t"              \
+        "sb     %[p4_r],    -5(%[s4])           \n\t"              \
+        "sb     %[p5_r],    -6(%[s4])           \n\t"              \
+        "sb     %[p6_r],    -7(%[s4])           \n\t"              \
+                                                                   \
+        :                                                          \
+        : [q6_r] "r"(q6_r), [q5_r] "r"(q5_r), [q4_r] "r"(q4_r),    \
+          [q3_r] "r"(q3_r), [q2_r] "r"(q2_r), [q1_r] "r"(q1_r),    \
+          [q0_r] "r"(q0_r), [p0_r] "r"(p0_r), [p1_r] "r"(p1_r),    \
+          [p2_r] "r"(p2_r), [p3_r] "r"(p3_r), [p4_r] "r"(p4_r),    \
+          [p5_r] "r"(p5_r), [p6_r] "r"(p6_r), [s4] "r"(s4));       \
+                                                                   \
+    __asm__ __volatile__(                                          \
+        "srl    %[q6_r],    %[q6_r],    16      \n\t"              \
+        "srl    %[q5_r],    %[q5_r],    16      \n\t"              \
+        "srl    %[q4_r],    %[q4_r],    16      \n\t"              \
+        "srl    %[q3_r],    %[q3_r],    16      \n\t"              \
+        "srl    %[q2_r],    %[q2_r],    16      \n\t"              \
+        "srl    %[q1_r],    %[q1_r],    16      \n\t"              \
+        "srl    %[q0_r],    %[q0_r],    16      \n\t"              \
+        "srl    %[p0_r],    %[p0_r],    16      \n\t"              \
+        "srl    %[p1_r],    %[p1_r],    16      \n\t"              \
+        "srl    %[p2_r],    %[p2_r],    16      \n\t"              \
+        "srl    %[p3_r],    %[p3_r],    16      \n\t"              \
+        "srl    %[p4_r],    %[p4_r],    16      \n\t"              \
+        "srl    %[p5_r],    %[p5_r],    16      \n\t"              \
+        "srl    %[p6_r],    %[p6_r],    16      \n\t"              \
+                                                                   \
+        : [q6_r] "+r"(q6_r), [q5_r] "+r"(q5_r), [q4_r] "+r"(q4_r), \
+          [q3_r] "+r"(q3_r), [q2_r] "+r"(q2_r), [q1_r] "+r"(q1_r), \
+          [q0_r] "+r"(q0_r), [p0_r] "+r"(p0_r), [p1_r] "+r"(p1_r), \
+          [p2_r] "+r"(p2_r), [p3_r] "+r"(p3_r), [p4_r] "+r"(p4_r), \
+          [p5_r] "+r"(p5_r), [p6_r] "+r"(p6_r)                     \
+        :);                                                        \
+                                                                   \
+    __asm__ __volatile__(                                          \
+        "sb     %[q6_r],     6(%[s3])           \n\t"              \
+        "sb     %[q5_r],     5(%[s3])           \n\t"              \
+        "sb     %[q4_r],     4(%[s3])           \n\t"              \
+        "sb     %[q3_r],     3(%[s3])           \n\t"              \
+        "sb     %[q2_r],     2(%[s3])           \n\t"              \
+        "sb     %[q1_r],     1(%[s3])           \n\t"              \
+        "sb     %[q0_r],     0(%[s3])           \n\t"              \
+        "sb     %[p0_r],    -1(%[s3])           \n\t"              \
+        "sb     %[p1_r],    -2(%[s3])           \n\t"              \
+        "sb     %[p2_r],    -3(%[s3])           \n\t"              \
+        "sb     %[p3_r],    -4(%[s3])           \n\t"              \
+        "sb     %[p4_r],    -5(%[s3])           \n\t"              \
+        "sb     %[p5_r],    -6(%[s3])           \n\t"              \
+        "sb     %[p6_r],    -7(%[s3])           \n\t"              \
+                                                                   \
+        :                                                          \
+        : [q6_r] "r"(q6_r), [q5_r] "r"(q5_r), [q4_r] "r"(q4_r),    \
+          [q3_r] "r"(q3_r), [q2_r] "r"(q2_r), [q1_r] "r"(q1_r),    \
+          [q0_r] "r"(q0_r), [p0_r] "r"(p0_r), [p1_r] "r"(p1_r),    \
+          [p2_r] "r"(p2_r), [p3_r] "r"(p3_r), [p4_r] "r"(p4_r),    \
+          [p5_r] "r"(p5_r), [p6_r] "r"(p6_r), [s3] "r"(s3));       \
+                                                                   \
+    __asm__ __volatile__(                                          \
+        "sb     %[q6_l],     6(%[s2])           \n\t"              \
+        "sb     %[q5_l],     5(%[s2])           \n\t"              \
+        "sb     %[q4_l],     4(%[s2])           \n\t"              \
+        "sb     %[q3_l],     3(%[s2])           \n\t"              \
+        "sb     %[q2_l],     2(%[s2])           \n\t"              \
+        "sb     %[q1_l],     1(%[s2])           \n\t"              \
+        "sb     %[q0_l],     0(%[s2])           \n\t"              \
+        "sb     %[p0_l],    -1(%[s2])           \n\t"              \
+        "sb     %[p1_l],    -2(%[s2])           \n\t"              \
+        "sb     %[p2_l],    -3(%[s2])           \n\t"              \
+        "sb     %[p3_l],    -4(%[s2])           \n\t"              \
+        "sb     %[p4_l],    -5(%[s2])           \n\t"              \
+        "sb     %[p5_l],    -6(%[s2])           \n\t"              \
+        "sb     %[p6_l],    -7(%[s2])           \n\t"              \
+                                                                   \
+        :                                                          \
+        : [q6_l] "r"(q6_l), [q5_l] "r"(q5_l), [q4_l] "r"(q4_l),    \
+          [q3_l] "r"(q3_l), [q2_l] "r"(q2_l), [q1_l] "r"(q1_l),    \
+          [q0_l] "r"(q0_l), [p0_l] "r"(p0_l), [p1_l] "r"(p1_l),    \
+          [p2_l] "r"(p2_l), [p3_l] "r"(p3_l), [p4_l] "r"(p4_l),    \
+          [p5_l] "r"(p5_l), [p6_l] "r"(p6_l), [s2] "r"(s2));       \
+                                                                   \
+    __asm__ __volatile__(                                          \
+        "srl    %[q6_l],    %[q6_l],    16     \n\t"               \
+        "srl    %[q5_l],    %[q5_l],    16     \n\t"               \
+        "srl    %[q4_l],    %[q4_l],    16     \n\t"               \
+        "srl    %[q3_l],    %[q3_l],    16     \n\t"               \
+        "srl    %[q2_l],    %[q2_l],    16     \n\t"               \
+        "srl    %[q1_l],    %[q1_l],    16     \n\t"               \
+        "srl    %[q0_l],    %[q0_l],    16     \n\t"               \
+        "srl    %[p0_l],    %[p0_l],    16     \n\t"               \
+        "srl    %[p1_l],    %[p1_l],    16     \n\t"               \
+        "srl    %[p2_l],    %[p2_l],    16     \n\t"               \
+        "srl    %[p3_l],    %[p3_l],    16     \n\t"               \
+        "srl    %[p4_l],    %[p4_l],    16     \n\t"               \
+        "srl    %[p5_l],    %[p5_l],    16     \n\t"               \
+        "srl    %[p6_l],    %[p6_l],    16     \n\t"               \
+                                                                   \
+        : [q6_l] "+r"(q6_l), [q5_l] "+r"(q5_l), [q4_l] "+r"(q4_l), \
+          [q3_l] "+r"(q3_l), [q2_l] "+r"(q2_l), [q1_l] "+r"(q1_l), \
+          [q0_l] "+r"(q0_l), [p0_l] "+r"(p0_l), [p1_l] "+r"(p1_l), \
+          [p2_l] "+r"(p2_l), [p3_l] "+r"(p3_l), [p4_l] "+r"(p4_l), \
+          [p5_l] "+r"(p5_l), [p6_l] "+r"(p6_l)                     \
+        :);                                                        \
+                                                                   \
+    __asm__ __volatile__(                                          \
+        "sb     %[q6_l],     6(%[s1])           \n\t"              \
+        "sb     %[q5_l],     5(%[s1])           \n\t"              \
+        "sb     %[q4_l],     4(%[s1])           \n\t"              \
+        "sb     %[q3_l],     3(%[s1])           \n\t"              \
+        "sb     %[q2_l],     2(%[s1])           \n\t"              \
+        "sb     %[q1_l],     1(%[s1])           \n\t"              \
+        "sb     %[q0_l],     0(%[s1])           \n\t"              \
+        "sb     %[p0_l],    -1(%[s1])           \n\t"              \
+        "sb     %[p1_l],    -2(%[s1])           \n\t"              \
+        "sb     %[p2_l],    -3(%[s1])           \n\t"              \
+        "sb     %[p3_l],    -4(%[s1])           \n\t"              \
+        "sb     %[p4_l],    -5(%[s1])           \n\t"              \
+        "sb     %[p5_l],    -6(%[s1])           \n\t"              \
+        "sb     %[p6_l],    -7(%[s1])           \n\t"              \
+                                                                   \
+        :                                                          \
+        : [q6_l] "r"(q6_l), [q5_l] "r"(q5_l), [q4_l] "r"(q4_l),    \
+          [q3_l] "r"(q3_l), [q2_l] "r"(q2_l), [q1_l] "r"(q1_l),    \
+          [q0_l] "r"(q0_l), [p0_l] "r"(p0_l), [p1_l] "r"(p1_l),    \
+          [p2_l] "r"(p2_l), [p3_l] "r"(p3_l), [p4_l] "r"(p4_l),    \
+          [p5_l] "r"(p5_l), [p6_l] "r"(p6_l), [s1] "r"(s1));       \
+  }
+
+#define PACK_LEFT_0TO3()                                              \
+  {                                                                   \
+    __asm__ __volatile__(                                             \
+        "preceu.ph.qbl   %[p3_l],   %[p3]   \n\t"                     \
+        "preceu.ph.qbl   %[p2_l],   %[p2]   \n\t"                     \
+        "preceu.ph.qbl   %[p1_l],   %[p1]   \n\t"                     \
+        "preceu.ph.qbl   %[p0_l],   %[p0]   \n\t"                     \
+        "preceu.ph.qbl   %[q0_l],   %[q0]   \n\t"                     \
+        "preceu.ph.qbl   %[q1_l],   %[q1]   \n\t"                     \
+        "preceu.ph.qbl   %[q2_l],   %[q2]   \n\t"                     \
+        "preceu.ph.qbl   %[q3_l],   %[q3]   \n\t"                     \
+                                                                      \
+        : [p3_l] "=&r"(p3_l), [p2_l] "=&r"(p2_l), [p1_l] "=&r"(p1_l), \
+          [p0_l] "=&r"(p0_l), [q0_l] "=&r"(q0_l), [q1_l] "=&r"(q1_l), \
+          [q2_l] "=&r"(q2_l), [q3_l] "=&r"(q3_l)                      \
+        : [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0),     \
+          [q0] "r"(q0), [q1] "r"(q1), [q2] "r"(q2), [q3] "r"(q3));    \
+  }
+
+#define PACK_LEFT_4TO7()                                              \
+  {                                                                   \
+    __asm__ __volatile__(                                             \
+        "preceu.ph.qbl   %[p7_l],   %[p7]   \n\t"                     \
+        "preceu.ph.qbl   %[p6_l],   %[p6]   \n\t"                     \
+        "preceu.ph.qbl   %[p5_l],   %[p5]   \n\t"                     \
+        "preceu.ph.qbl   %[p4_l],   %[p4]   \n\t"                     \
+        "preceu.ph.qbl   %[q4_l],   %[q4]   \n\t"                     \
+        "preceu.ph.qbl   %[q5_l],   %[q5]   \n\t"                     \
+        "preceu.ph.qbl   %[q6_l],   %[q6]   \n\t"                     \
+        "preceu.ph.qbl   %[q7_l],   %[q7]   \n\t"                     \
+                                                                      \
+        : [p7_l] "=&r"(p7_l), [p6_l] "=&r"(p6_l), [p5_l] "=&r"(p5_l), \
+          [p4_l] "=&r"(p4_l), [q4_l] "=&r"(q4_l), [q5_l] "=&r"(q5_l), \
+          [q6_l] "=&r"(q6_l), [q7_l] "=&r"(q7_l)                      \
+        : [p7] "r"(p7), [p6] "r"(p6), [p5] "r"(p5), [p4] "r"(p4),     \
+          [q4] "r"(q4), [q5] "r"(q5), [q6] "r"(q6), [q7] "r"(q7));    \
+  }
+
+#define PACK_RIGHT_0TO3()                                             \
+  {                                                                   \
+    __asm__ __volatile__(                                             \
+        "preceu.ph.qbr   %[p3_r],   %[p3]  \n\t"                      \
+        "preceu.ph.qbr   %[p2_r],   %[p2]   \n\t"                     \
+        "preceu.ph.qbr   %[p1_r],   %[p1]   \n\t"                     \
+        "preceu.ph.qbr   %[p0_r],   %[p0]   \n\t"                     \
+        "preceu.ph.qbr   %[q0_r],   %[q0]   \n\t"                     \
+        "preceu.ph.qbr   %[q1_r],   %[q1]   \n\t"                     \
+        "preceu.ph.qbr   %[q2_r],   %[q2]   \n\t"                     \
+        "preceu.ph.qbr   %[q3_r],   %[q3]   \n\t"                     \
+                                                                      \
+        : [p3_r] "=&r"(p3_r), [p2_r] "=&r"(p2_r), [p1_r] "=&r"(p1_r), \
+          [p0_r] "=&r"(p0_r), [q0_r] "=&r"(q0_r), [q1_r] "=&r"(q1_r), \
+          [q2_r] "=&r"(q2_r), [q3_r] "=&r"(q3_r)                      \
+        : [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0),     \
+          [q0] "r"(q0), [q1] "r"(q1), [q2] "r"(q2), [q3] "r"(q3));    \
+  }
+
+#define PACK_RIGHT_4TO7()                                             \
+  {                                                                   \
+    __asm__ __volatile__(                                             \
+        "preceu.ph.qbr   %[p7_r],   %[p7]   \n\t"                     \
+        "preceu.ph.qbr   %[p6_r],   %[p6]   \n\t"                     \
+        "preceu.ph.qbr   %[p5_r],   %[p5]   \n\t"                     \
+        "preceu.ph.qbr   %[p4_r],   %[p4]   \n\t"                     \
+        "preceu.ph.qbr   %[q4_r],   %[q4]   \n\t"                     \
+        "preceu.ph.qbr   %[q5_r],   %[q5]   \n\t"                     \
+        "preceu.ph.qbr   %[q6_r],   %[q6]   \n\t"                     \
+        "preceu.ph.qbr   %[q7_r],   %[q7]   \n\t"                     \
+                                                                      \
+        : [p7_r] "=&r"(p7_r), [p6_r] "=&r"(p6_r), [p5_r] "=&r"(p5_r), \
+          [p4_r] "=&r"(p4_r), [q4_r] "=&r"(q4_r), [q5_r] "=&r"(q5_r), \
+          [q6_r] "=&r"(q6_r), [q7_r] "=&r"(q7_r)                      \
+        : [p7] "r"(p7), [p6] "r"(p6), [p5] "r"(p5), [p4] "r"(p4),     \
+          [q4] "r"(q4), [q5] "r"(q5), [q6] "r"(q6), [q7] "r"(q7));    \
+  }
+
+#define COMBINE_LEFT_RIGHT_0TO2()                                         \
+  {                                                                       \
+    __asm__ __volatile__(                                                 \
+        "precr.qb.ph    %[p2],  %[p2_l],    %[p2_r]    \n\t"              \
+        "precr.qb.ph    %[p1],  %[p1_l],    %[p1_r]    \n\t"              \
+        "precr.qb.ph    %[p0],  %[p0_l],    %[p0_r]    \n\t"              \
+        "precr.qb.ph    %[q0],  %[q0_l],    %[q0_r]    \n\t"              \
+        "precr.qb.ph    %[q1],  %[q1_l],    %[q1_r]    \n\t"              \
+        "precr.qb.ph    %[q2],  %[q2_l],    %[q2_r]    \n\t"              \
+                                                                          \
+        : [p2] "=&r"(p2), [p1] "=&r"(p1), [p0] "=&r"(p0), [q0] "=&r"(q0), \
+          [q1] "=&r"(q1), [q2] "=&r"(q2)                                  \
+        : [p2_l] "r"(p2_l), [p2_r] "r"(p2_r), [p1_l] "r"(p1_l),           \
+          [p1_r] "r"(p1_r), [p0_l] "r"(p0_l), [p0_r] "r"(p0_r),           \
+          [q0_l] "r"(q0_l), [q0_r] "r"(q0_r), [q1_l] "r"(q1_l),           \
+          [q1_r] "r"(q1_r), [q2_l] "r"(q2_l), [q2_r] "r"(q2_r));          \
+  }
+
+#define COMBINE_LEFT_RIGHT_3TO6()                                         \
+  {                                                                       \
+    __asm__ __volatile__(                                                 \
+        "precr.qb.ph    %[p6],  %[p6_l],    %[p6_r]    \n\t"              \
+        "precr.qb.ph    %[p5],  %[p5_l],    %[p5_r]    \n\t"              \
+        "precr.qb.ph    %[p4],  %[p4_l],    %[p4_r]    \n\t"              \
+        "precr.qb.ph    %[p3],  %[p3_l],    %[p3_r]    \n\t"              \
+        "precr.qb.ph    %[q3],  %[q3_l],    %[q3_r]    \n\t"              \
+        "precr.qb.ph    %[q4],  %[q4_l],    %[q4_r]    \n\t"              \
+        "precr.qb.ph    %[q5],  %[q5_l],    %[q5_r]    \n\t"              \
+        "precr.qb.ph    %[q6],  %[q6_l],    %[q6_r]    \n\t"              \
+                                                                          \
+        : [p6] "=&r"(p6), [p5] "=&r"(p5), [p4] "=&r"(p4), [p3] "=&r"(p3), \
+          [q3] "=&r"(q3), [q4] "=&r"(q4), [q5] "=&r"(q5), [q6] "=&r"(q6)  \
+        : [p6_l] "r"(p6_l), [p5_l] "r"(p5_l), [p4_l] "r"(p4_l),           \
+          [p3_l] "r"(p3_l), [p6_r] "r"(p6_r), [p5_r] "r"(p5_r),           \
+          [p4_r] "r"(p4_r), [p3_r] "r"(p3_r), [q3_l] "r"(q3_l),           \
+          [q4_l] "r"(q4_l), [q5_l] "r"(q5_l), [q6_l] "r"(q6_l),           \
+          [q3_r] "r"(q3_r), [q4_r] "r"(q4_r), [q5_r] "r"(q5_r),           \
+          [q6_r] "r"(q6_r));                                              \
+  }
+
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_MIPS_LOOPFILTER_MACROS_DSPR2_H_
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_masks_dspr2.h b/third_party/aom/aom_dsp/mips/loopfilter_masks_dspr2.h
new file mode 100644
index 0000000000..8db3e521f6
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_masks_dspr2.h
@@ -0,0 +1,356 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_MIPS_LOOPFILTER_MASKS_DSPR2_H_
+#define AOM_DSP_MIPS_LOOPFILTER_MASKS_DSPR2_H_
+
+#include <stdlib.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_mem/aom_mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if HAVE_DSPR2
+/* processing 4 pixels at the same time
+ * compute hev and mask in the same function */
+static INLINE void filter_hev_mask_dspr2(uint32_t limit, uint32_t flimit,
+                                         uint32_t p1, uint32_t p0, uint32_t p3,
+                                         uint32_t p2, uint32_t q0, uint32_t q1,
+                                         uint32_t q2, uint32_t q3,
+                                         uint32_t thresh, uint32_t *hev,
+                                         uint32_t *mask) {
+  uint32_t c, r, r3, r_k;
+  uint32_t s1, s2, s3;
+  uint32_t ones = 0xFFFFFFFF;
+  uint32_t hev1;
+
+  __asm__ __volatile__(
+      /* mask |= (abs(p3 - p2) > limit) */
+      "subu_s.qb      %[c],   %[p3],     %[p2]        \n\t"
+      "subu_s.qb      %[r_k], %[p2],     %[p3]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   $0,        %[c]         \n\t"
+
+      /* mask |= (abs(p2 - p1) > limit) */
+      "subu_s.qb      %[c],   %[p2],     %[p1]        \n\t"
+      "subu_s.qb      %[r_k], %[p1],     %[p2]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   %[r],      %[c]         \n\t"
+
+      /* mask |= (abs(p1 - p0) > limit)
+       * hev  |= (abs(p1 - p0) > thresh)
+       */
+      "subu_s.qb      %[c],   %[p1],     %[p0]        \n\t"
+      "subu_s.qb      %[r_k], %[p0],     %[p1]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[thresh], %[r_k]       \n\t"
+      "or             %[r3],  $0,        %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   %[r],      %[c]         \n\t"
+
+      /* mask |= (abs(q1 - q0) > limit)
+       * hev  |= (abs(q1 - q0) > thresh)
+       */
+      "subu_s.qb      %[c],   %[q1],     %[q0]        \n\t"
+      "subu_s.qb      %[r_k], %[q0],     %[q1]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[thresh], %[r_k]       \n\t"
+      "or             %[r3],  %[r3],     %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   %[r],      %[c]         \n\t"
+
+      /* mask |= (abs(q2 - q1) > limit) */
+      "subu_s.qb      %[c],   %[q2],     %[q1]        \n\t"
+      "subu_s.qb      %[r_k], %[q1],     %[q2]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   %[r],      %[c]         \n\t"
+      "sll            %[r3],    %[r3],    24          \n\t"
+
+      /* mask |= (abs(q3 - q2) > limit) */
+      "subu_s.qb      %[c],   %[q3],     %[q2]        \n\t"
+      "subu_s.qb      %[r_k], %[q2],     %[q3]        \n\t"
+      "or             %[r_k], %[r_k],    %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[limit],  %[r_k]       \n\t"
+      "or             %[r],   %[r],      %[c]         \n\t"
+
+      : [c] "=&r"(c), [r_k] "=&r"(r_k), [r] "=&r"(r), [r3] "=&r"(r3)
+      : [limit] "r"(limit), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1),
+        [p0] "r"(p0), [q1] "r"(q1), [q0] "r"(q0), [q2] "r"(q2), [q3] "r"(q3),
+        [thresh] "r"(thresh));
+
+  __asm__ __volatile__(
+      /* abs(p0 - q0) */
+      "subu_s.qb      %[c],   %[p0],     %[q0]        \n\t"
+      "subu_s.qb      %[r_k], %[q0],     %[p0]        \n\t"
+      "wrdsp          %[r3]                           \n\t"
+      "or             %[s1],  %[r_k],    %[c]         \n\t"
+
+      /* abs(p1 - q1) */
+      "subu_s.qb      %[c],    %[p1],    %[q1]        \n\t"
+      "addu_s.qb      %[s3],   %[s1],    %[s1]        \n\t"
+      "pick.qb        %[hev1], %[ones],  $0           \n\t"
+      "subu_s.qb      %[r_k],  %[q1],    %[p1]        \n\t"
+      "or             %[s2],   %[r_k],   %[c]         \n\t"
+
+      /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > flimit * 2 + limit */
+      "shrl.qb        %[s2],   %[s2],     1           \n\t"
+      "addu_s.qb      %[s1],   %[s2],     %[s3]       \n\t"
+      "cmpgu.lt.qb    %[c],    %[flimit], %[s1]       \n\t"
+      "or             %[r],    %[r],      %[c]        \n\t"
+      "sll            %[r],    %[r],      24          \n\t"
+
+      "wrdsp          %[r]                            \n\t"
+      "pick.qb        %[s2],  $0,         %[ones]     \n\t"
+
+      : [c] "=&r"(c), [r_k] "=&r"(r_k), [s1] "=&r"(s1), [hev1] "=&r"(hev1),
+        [s2] "=&r"(s2), [r] "+r"(r), [s3] "=&r"(s3)
+      : [p0] "r"(p0), [q0] "r"(q0), [p1] "r"(p1), [r3] "r"(r3), [q1] "r"(q1),
+        [ones] "r"(ones), [flimit] "r"(flimit));
+
+  *hev = hev1;
+  *mask = s2;
+}
+
+static INLINE void filter_hev_mask_flatmask4_dspr2(
+    uint32_t limit, uint32_t flimit, uint32_t thresh, uint32_t p1, uint32_t p0,
+    uint32_t p3, uint32_t p2, uint32_t q0, uint32_t q1, uint32_t q2,
+    uint32_t q3, uint32_t *hev, uint32_t *mask, uint32_t *flat) {
+  uint32_t c, r, r3, r_k, r_flat;
+  uint32_t s1, s2, s3;
+  uint32_t ones = 0xFFFFFFFF;
+  uint32_t flat_thresh = 0x01010101;
+  uint32_t hev1;
+  uint32_t flat1;
+
+  __asm__ __volatile__(
+      /* mask |= (abs(p3 - p2) > limit) */
+      "subu_s.qb      %[c],       %[p3],          %[p2]        \n\t"
+      "subu_s.qb      %[r_k],     %[p2],          %[p3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       $0,             %[c]         \n\t"
+
+      /* mask |= (abs(p2 - p1) > limit) */
+      "subu_s.qb      %[c],       %[p2],          %[p1]        \n\t"
+      "subu_s.qb      %[r_k],     %[p1],          %[p2]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       %[r],           %[c]         \n\t"
+
+      /* mask |= (abs(p1 - p0) > limit)
+       * hev  |= (abs(p1 - p0) > thresh)
+       * flat |= (abs(p1 - p0) > thresh)
+       */
+      "subu_s.qb      %[c],       %[p1],          %[p0]        \n\t"
+      "subu_s.qb      %[r_k],     %[p0],          %[p1]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[thresh],      %[r_k]       \n\t"
+      "or             %[r3],      $0,             %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       %[r],           %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  $0,             %[c]         \n\t"
+
+      /* mask |= (abs(q1 - q0) > limit)
+       * hev  |= (abs(q1 - q0) > thresh)
+       * flat |= (abs(q1 - q0) > thresh)
+       */
+      "subu_s.qb      %[c],       %[q1],          %[q0]        \n\t"
+      "subu_s.qb      %[r_k],     %[q0],          %[q1]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[thresh],      %[r_k]       \n\t"
+      "or             %[r3],      %[r3],          %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       %[r],           %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(p0 - p2) > thresh) */
+      "subu_s.qb      %[c],       %[p0],          %[p2]        \n\t"
+      "subu_s.qb      %[r_k],     %[p2],          %[p0]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(q0 - q2) > thresh) */
+      "subu_s.qb      %[c],       %[q0],          %[q2]        \n\t"
+      "subu_s.qb      %[r_k],     %[q2],          %[q0]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(p3 - p0) > thresh) */
+      "subu_s.qb      %[c],       %[p3],          %[p0]        \n\t"
+      "subu_s.qb      %[r_k],     %[p0],          %[p3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(q3 - q0) > thresh) */
+      "subu_s.qb      %[c],       %[q3],          %[q0]        \n\t"
+      "subu_s.qb      %[r_k],     %[q0],          %[q3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+      "sll            %[r_flat],  %[r_flat],      24           \n\t"
+      /* look at stall here */
+      "wrdsp          %[r_flat]                                \n\t"
+      "pick.qb        %[flat1],   $0,             %[ones]      \n\t"
+
+      /* mask |= (abs(q2 - q1) > limit) */
+      "subu_s.qb      %[c],       %[q2],          %[q1]        \n\t"
+      "subu_s.qb      %[r_k],     %[q1],          %[q2]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       %[r],           %[c]         \n\t"
+      "sll            %[r3],      %[r3],          24           \n\t"
+
+      /* mask |= (abs(q3 - q2) > limit) */
+      "subu_s.qb      %[c],       %[q3],          %[q2]        \n\t"
+      "subu_s.qb      %[r_k],     %[q2],          %[q3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[limit],       %[r_k]       \n\t"
+      "or             %[r],       %[r],           %[c]         \n\t"
+
+      : [c] "=&r"(c), [r_k] "=&r"(r_k), [r] "=&r"(r), [r3] "=&r"(r3),
+        [r_flat] "=&r"(r_flat), [flat1] "=&r"(flat1)
+      : [limit] "r"(limit), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1),
+        [p0] "r"(p0), [q1] "r"(q1), [q0] "r"(q0), [q2] "r"(q2), [q3] "r"(q3),
+        [thresh] "r"(thresh), [flat_thresh] "r"(flat_thresh), [ones] "r"(ones));
+
+  __asm__ __volatile__(
+      /* abs(p0 - q0) */
+      "subu_s.qb      %[c],   %[p0],     %[q0]        \n\t"
+      "subu_s.qb      %[r_k], %[q0],     %[p0]        \n\t"
+      "wrdsp          %[r3]                           \n\t"
+      "or             %[s1],  %[r_k],    %[c]         \n\t"
+
+      /* abs(p1 - q1) */
+      "subu_s.qb      %[c],    %[p1],    %[q1]        \n\t"
+      "addu_s.qb      %[s3],   %[s1],    %[s1]        \n\t"
+      "pick.qb        %[hev1], %[ones],  $0           \n\t"
+      "subu_s.qb      %[r_k],  %[q1],    %[p1]        \n\t"
+      "or             %[s2],   %[r_k],   %[c]         \n\t"
+
+      /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > flimit * 2 + limit */
+      "shrl.qb        %[s2],   %[s2],     1           \n\t"
+      "addu_s.qb      %[s1],   %[s2],     %[s3]       \n\t"
+      "cmpgu.lt.qb    %[c],    %[flimit], %[s1]       \n\t"
+      "or             %[r],    %[r],      %[c]        \n\t"
+      "sll            %[r],    %[r],      24          \n\t"
+
+      "wrdsp          %[r]                            \n\t"
+      "pick.qb        %[s2],   $0,        %[ones]     \n\t"
+
+      : [c] "=&r"(c), [r_k] "=&r"(r_k), [s1] "=&r"(s1), [hev1] "=&r"(hev1),
+        [s2] "=&r"(s2), [r] "+r"(r), [s3] "=&r"(s3)
+      : [p0] "r"(p0), [q0] "r"(q0), [p1] "r"(p1), [r3] "r"(r3), [q1] "r"(q1),
+        [ones] "r"(ones), [flimit] "r"(flimit));
+
+  *hev = hev1;
+  *mask = s2;
+  *flat = flat1;
+}
+
+static INLINE void flatmask5(uint32_t p4, uint32_t p3, uint32_t p2, uint32_t p1,
+                             uint32_t p0, uint32_t q0, uint32_t q1, uint32_t q2,
+                             uint32_t q3, uint32_t q4, uint32_t *flat2) {
+  uint32_t c, r, r_k, r_flat;
+  uint32_t ones = 0xFFFFFFFF;
+  uint32_t flat_thresh = 0x01010101;
+  uint32_t flat1, flat3;
+
+  __asm__ __volatile__(
+      /* flat |= (abs(p4 - p0) > thresh) */
+      "subu_s.qb      %[c],   %[p4],           %[p0]        \n\t"
+      "subu_s.qb      %[r_k], %[p0],           %[p4]        \n\t"
+      "or             %[r_k], %[r_k],          %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],   %[flat_thresh],  %[r_k]       \n\t"
+      "or             %[r],   $0,              %[c]         \n\t"
+
+      /* flat |= (abs(q4 - q0) > thresh) */
+      "subu_s.qb      %[c],     %[q4],           %[q0]     \n\t"
+      "subu_s.qb      %[r_k],   %[q0],           %[q4]     \n\t"
+      "or             %[r_k],   %[r_k],          %[c]      \n\t"
+      "cmpgu.lt.qb    %[c],     %[flat_thresh],  %[r_k]    \n\t"
+      "or             %[r],     %[r],            %[c]      \n\t"
+      "sll            %[r],     %[r],            24        \n\t"
+      "wrdsp          %[r]                                 \n\t"
+      "pick.qb        %[flat3], $0,           %[ones]      \n\t"
+
+      /* flat |= (abs(p1 - p0) > thresh) */
+      "subu_s.qb      %[c],       %[p1],          %[p0]        \n\t"
+      "subu_s.qb      %[r_k],     %[p0],          %[p1]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  $0,             %[c]         \n\t"
+
+      /* flat |= (abs(q1 - q0) > thresh) */
+      "subu_s.qb      %[c],      %[q1],           %[q0]        \n\t"
+      "subu_s.qb      %[r_k],    %[q0],           %[q1]        \n\t"
+      "or             %[r_k],    %[r_k],          %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],      %[flat_thresh],  %[r_k]       \n\t"
+      "or             %[r_flat], %[r_flat],       %[c]         \n\t"
+
+      /* flat |= (abs(p0 - p2) > thresh) */
+      "subu_s.qb      %[c],       %[p0],          %[p2]        \n\t"
+      "subu_s.qb      %[r_k],     %[p2],          %[p0]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(q0 - q2) > thresh) */
+      "subu_s.qb      %[c],       %[q0],          %[q2]        \n\t"
+      "subu_s.qb      %[r_k],     %[q2],          %[q0]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(p3 - p0) > thresh) */
+      "subu_s.qb      %[c],       %[p3],          %[p0]        \n\t"
+      "subu_s.qb      %[r_k],     %[p0],          %[p3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+
+      /* flat |= (abs(q3 - q0) > thresh) */
+      "subu_s.qb      %[c],       %[q3],          %[q0]        \n\t"
+      "subu_s.qb      %[r_k],     %[q0],          %[q3]        \n\t"
+      "or             %[r_k],     %[r_k],         %[c]         \n\t"
+      "cmpgu.lt.qb    %[c],       %[flat_thresh], %[r_k]       \n\t"
+      "or             %[r_flat],  %[r_flat],      %[c]         \n\t"
+      "sll            %[r_flat],  %[r_flat],      24           \n\t"
+      "wrdsp          %[r_flat]                                \n\t"
+      "pick.qb        %[flat1],   $0,             %[ones]      \n\t"
+      /* flat & flatmask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3) */
+      "and            %[flat1],  %[flat3],        %[flat1]     \n\t"
+
+      : [c] "=&r"(c), [r_k] "=&r"(r_k), [r] "=&r"(r), [r_flat] "=&r"(r_flat),
+        [flat1] "=&r"(flat1), [flat3] "=&r"(flat3)
+      : [p4] "r"(p4), [p3] "r"(p3), [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0),
+        [q0] "r"(q0), [q1] "r"(q1), [q2] "r"(q2), [q3] "r"(q3), [q4] "r"(q4),
+        [flat_thresh] "r"(flat_thresh), [ones] "r"(ones));
+
+  *flat2 = flat1;
+}
+#endif  // #if HAVE_DSPR2
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_MIPS_LOOPFILTER_MASKS_DSPR2_H_
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_mb_dspr2.c b/third_party/aom/aom_dsp/mips/loopfilter_mb_dspr2.c
new file mode 100644
index 0000000000..a3b5a9eb1c
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_mb_dspr2.c
@@ -0,0 +1,589 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/mips/common_dspr2.h"
+#include "aom_dsp/mips/loopfilter_filters_dspr2.h"
+#include "aom_dsp/mips/loopfilter_macros_dspr2.h"
+#include "aom_dsp/mips/loopfilter_masks_dspr2.h"
+#include "aom_mem/aom_mem.h"
+
+#if HAVE_DSPR2
+void aom_lpf_horizontal_8_dspr2(unsigned char *s, int pitch,
+                                const uint8_t *blimit, const uint8_t *limit,
+                                const uint8_t *thresh) {
+  uint32_t mask;
+  uint32_t hev, flat;
+  uint8_t i;
+  uint8_t *sp3, *sp2, *sp1, *sp0, *sq0, *sq1, *sq2, *sq3;
+  uint32_t thresh_vec, flimit_vec, limit_vec;
+  uint32_t uflimit, ulimit, uthresh;
+  uint32_t p1_f0, p0_f0, q0_f0, q1_f0;
+  uint32_t p3, p2, p1, p0, q0, q1, q2, q3;
+  uint32_t p0_l, p1_l, p2_l, p3_l, q0_l, q1_l, q2_l, q3_l;
+  uint32_t p0_r, p1_r, p2_r, p3_r, q0_r, q1_r, q2_r, q3_r;
+
+  uflimit = *blimit;
+  ulimit = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__(
+      "replv.qb       %[thresh_vec],    %[uthresh]    \n\t"
+      "replv.qb       %[flimit_vec],    %[uflimit]    \n\t"
+      "replv.qb       %[limit_vec],     %[ulimit]     \n\t"
+
+      : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec),
+        [limit_vec] "=r"(limit_vec)
+      : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit));
+
+  /* prefetch data for store */
+  prefetch_store(s);
+
+  for (i = 0; i < 2; i++) {
+    sp3 = s - (pitch << 2);
+    sp2 = sp3 + pitch;
+    sp1 = sp2 + pitch;
+    sp0 = sp1 + pitch;
+    sq0 = s;
+    sq1 = s + pitch;
+    sq2 = sq1 + pitch;
+    sq3 = sq2 + pitch;
+
+    __asm__ __volatile__(
+        "lw     %[p3],      (%[sp3])    \n\t"
+        "lw     %[p2],      (%[sp2])    \n\t"
+        "lw     %[p1],      (%[sp1])    \n\t"
+        "lw     %[p0],      (%[sp0])    \n\t"
+        "lw     %[q0],      (%[sq0])    \n\t"
+        "lw     %[q1],      (%[sq1])    \n\t"
+        "lw     %[q2],      (%[sq2])    \n\t"
+        "lw     %[q3],      (%[sq3])    \n\t"
+
+        : [p3] "=&r"(p3), [p2] "=&r"(p2), [p1] "=&r"(p1), [p0] "=&r"(p0),
+          [q3] "=&r"(q3), [q2] "=&r"(q2), [q1] "=&r"(q1), [q0] "=&r"(q0)
+        : [sp3] "r"(sp3), [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0),
+          [sq3] "r"(sq3), [sq2] "r"(sq2), [sq1] "r"(sq1), [sq0] "r"(sq0));
+
+    filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec, p1, p0,
+                                    p3, p2, q0, q1, q2, q3, &hev, &mask, &flat);
+
+    if ((flat == 0) && (mask != 0)) {
+      filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      __asm__ __volatile__(
+          "sw       %[p1_f0],   (%[sp1])    \n\t"
+          "sw       %[p0_f0],   (%[sp0])    \n\t"
+          "sw       %[q0_f0],   (%[sq0])    \n\t"
+          "sw       %[q1_f0],   (%[sq1])    \n\t"
+
+          :
+          : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+            [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0),
+            [sq1] "r"(sq1));
+    } else if ((mask & flat) == 0xFFFFFFFF) {
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r);
+
+      COMBINE_LEFT_RIGHT_0TO2()
+
+      __asm__ __volatile__(
+          "sw       %[p2],      (%[sp2])    \n\t"
+          "sw       %[p1],      (%[sp1])    \n\t"
+          "sw       %[p0],      (%[sp0])    \n\t"
+          "sw       %[q0],      (%[sq0])    \n\t"
+          "sw       %[q1],      (%[sq1])    \n\t"
+          "sw       %[q2],      (%[sq2])    \n\t"
+
+          :
+          : [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), [q0] "r"(q0),
+            [q1] "r"(q1), [q2] "r"(q2), [sp2] "r"(sp2), [sp1] "r"(sp1),
+            [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2));
+    } else if ((flat != 0) && (mask != 0)) {
+      /* filtering */
+      filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r);
+
+      if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb     %[p2_r],    (%[sp2])    \n\t"
+            "sb     %[p1_r],    (%[sp1])    \n\t"
+            "sb     %[p0_r],    (%[sp0])    \n\t"
+            "sb     %[q0_r],    (%[sq0])    \n\t"
+            "sb     %[q1_r],    (%[sq1])    \n\t"
+            "sb     %[q2_r],    (%[sq2])    \n\t"
+
+            :
+            : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r),
+              [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0),
+              [sq1] "r"(sq1), [sq2] "r"(sq2));
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  (%[sp1])    \n\t"
+            "sb         %[p0_f0],  (%[sp0])    \n\t"
+            "sb         %[q0_f0],  (%[sq0])    \n\t"
+            "sb         %[q1_f0],  (%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p2_r],    %[p2_r],    16      \n\t"
+          "srl      %[p1_r],    %[p1_r],    16      \n\t"
+          "srl      %[p0_r],    %[p0_r],    16      \n\t"
+          "srl      %[q0_r],    %[q0_r],    16      \n\t"
+          "srl      %[q1_r],    %[q1_r],    16      \n\t"
+          "srl      %[q2_r],    %[q2_r],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_r] "+r"(p2_r), [p1_r] "+r"(p1_r), [p0_r] "+r"(p0_r),
+            [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb     %[p2_r],    +1(%[sp2])    \n\t"
+            "sb     %[p1_r],    +1(%[sp1])    \n\t"
+            "sb     %[p0_r],    +1(%[sp0])    \n\t"
+            "sb     %[q0_r],    +1(%[sq0])    \n\t"
+            "sb     %[q1_r],    +1(%[sq1])    \n\t"
+            "sb     %[q2_r],    +1(%[sq2])    \n\t"
+
+            :
+            : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r),
+              [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0),
+              [sq1] "r"(sq1), [sq2] "r"(sq2));
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   +1(%[sp1])    \n\t"
+            "sb     %[p0_f0],   +1(%[sp0])    \n\t"
+            "sb     %[q0_f0],   +1(%[sq0])    \n\t"
+            "sb     %[q1_f0],   +1(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p1_f0],   %[p1_f0],   8     \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8     \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8     \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8     \n\t"
+
+          : [p2] "+r"(p2), [p1] "+r"(p1), [p0] "+r"(p0), [q0] "+r"(q0),
+            [q1] "+r"(q1), [q2] "+r"(q2), [p1_f0] "+r"(p1_f0),
+            [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb     %[p2_l],    +2(%[sp2])    \n\t"
+            "sb     %[p1_l],    +2(%[sp1])    \n\t"
+            "sb     %[p0_l],    +2(%[sp0])    \n\t"
+            "sb     %[q0_l],    +2(%[sq0])    \n\t"
+            "sb     %[q1_l],    +2(%[sq1])    \n\t"
+            "sb     %[q2_l],    +2(%[sq2])    \n\t"
+
+            :
+            : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l),
+              [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0),
+              [sq1] "r"(sq1), [sq2] "r"(sq2));
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   +2(%[sp1])    \n\t"
+            "sb     %[p0_f0],   +2(%[sp0])    \n\t"
+            "sb     %[q0_f0],   +2(%[sq0])    \n\t"
+            "sb     %[q1_f0],   +2(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p2_l],    %[p2_l],    16      \n\t"
+          "srl      %[p1_l],    %[p1_l],    16      \n\t"
+          "srl      %[p0_l],    %[p0_l],    16      \n\t"
+          "srl      %[q0_l],    %[q0_l],    16      \n\t"
+          "srl      %[q1_l],    %[q1_l],    16      \n\t"
+          "srl      %[q2_l],    %[q2_l],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_l] "+r"(p2_l), [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l),
+            [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb     %[p2_l],    +3(%[sp2])    \n\t"
+            "sb     %[p1_l],    +3(%[sp1])    \n\t"
+            "sb     %[p0_l],    +3(%[sp0])    \n\t"
+            "sb     %[q0_l],    +3(%[sq0])    \n\t"
+            "sb     %[q1_l],    +3(%[sq1])    \n\t"
+            "sb     %[q2_l],    +3(%[sq2])    \n\t"
+
+            :
+            : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l),
+              [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0),
+              [sq1] "r"(sq1), [sq2] "r"(sq2));
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   +3(%[sp1])    \n\t"
+            "sb     %[p0_f0],   +3(%[sp0])    \n\t"
+            "sb     %[q0_f0],   +3(%[sq0])    \n\t"
+            "sb     %[q1_f0],   +3(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+    }
+
+    s = s + 4;
+  }
+}
+
+void aom_lpf_vertical_8_dspr2(unsigned char *s, int pitch,
+                              const uint8_t *blimit, const uint8_t *limit,
+                              const uint8_t *thresh) {
+  uint8_t i;
+  uint32_t mask, hev, flat;
+  uint8_t *s1, *s2, *s3, *s4;
+  uint32_t prim1, prim2, sec3, sec4, prim3, prim4;
+  uint32_t thresh_vec, flimit_vec, limit_vec;
+  uint32_t uflimit, ulimit, uthresh;
+  uint32_t p3, p2, p1, p0, q3, q2, q1, q0;
+  uint32_t p1_f0, p0_f0, q0_f0, q1_f0;
+  uint32_t p0_l, p1_l, p2_l, p3_l, q0_l, q1_l, q2_l, q3_l;
+  uint32_t p0_r, p1_r, p2_r, p3_r, q0_r, q1_r, q2_r, q3_r;
+
+  uflimit = *blimit;
+  ulimit = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__(
+      "replv.qb     %[thresh_vec],  %[uthresh]    \n\t"
+      "replv.qb     %[flimit_vec],  %[uflimit]    \n\t"
+      "replv.qb     %[limit_vec],   %[ulimit]     \n\t"
+
+      : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec),
+        [limit_vec] "=r"(limit_vec)
+      : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit));
+
+  prefetch_store(s + pitch);
+
+  for (i = 0; i < 2; i++) {
+    s1 = s;
+    s2 = s + pitch;
+    s3 = s2 + pitch;
+    s4 = s3 + pitch;
+    s = s4 + pitch;
+
+    __asm__ __volatile__(
+        "lw     %[p0],  -4(%[s1])    \n\t"
+        "lw     %[p1],  -4(%[s2])    \n\t"
+        "lw     %[p2],  -4(%[s3])    \n\t"
+        "lw     %[p3],  -4(%[s4])    \n\t"
+        "lw     %[q3],    (%[s1])    \n\t"
+        "lw     %[q2],    (%[s2])    \n\t"
+        "lw     %[q1],    (%[s3])    \n\t"
+        "lw     %[q0],    (%[s4])    \n\t"
+
+        : [p3] "=&r"(p3), [p2] "=&r"(p2), [p1] "=&r"(p1), [p0] "=&r"(p0),
+          [q0] "=&r"(q0), [q1] "=&r"(q1), [q2] "=&r"(q2), [q3] "=&r"(q3)
+        : [s1] "r"(s1), [s2] "r"(s2), [s3] "r"(s3), [s4] "r"(s4));
+
+    /* transpose p3, p2, p1, p0
+       original (when loaded from memory)
+       register       -4    -3   -2     -1
+         p0         p0_0  p0_1  p0_2  p0_3
+         p1         p1_0  p1_1  p1_2  p1_3
+         p2         p2_0  p2_1  p2_2  p2_3
+         p3         p3_0  p3_1  p3_2  p3_3
+
+       after transpose
+       register
+         p0         p3_3  p2_3  p1_3  p0_3
+         p1         p3_2  p2_2  p1_2  p0_2
+         p2         p3_1  p2_1  p1_1  p0_1
+         p3         p3_0  p2_0  p1_0  p0_0
+    */
+    __asm__ __volatile__(
+        "precrq.qb.ph   %[prim1],   %[p0],      %[p1]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p0],      %[p1]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p2],      %[p3]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p2],      %[p3]       \n\t"
+
+        "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p0],      %[p1],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p2],      %[p3],      %[sec4]     \n\t"
+        "append         %[p1],      %[sec3],    16          \n\t"
+        "append         %[p3],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3),
+          [prim4] "=&r"(prim4), [p0] "+r"(p0), [p1] "+r"(p1), [p2] "+r"(p2),
+          [p3] "+r"(p3), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4)
+        :);
+
+    /* transpose q0, q1, q2, q3
+       original (when loaded from memory)
+       register       +1    +2    +3    +4
+         q3         q3_0  q3_1  q3_2  q3_3
+         q2         q2_0  q2_1  q2_2  q2_3
+         q1         q1_0  q1_1  q1_2  q1_3
+         q0         q0_0  q0_1  q0_2  q0_3
+
+       after transpose
+       register
+         q3         q0_3  q1_3  q2_3  q3_3
+         q2         q0_2  q1_2  q2_2  q3_2
+         q1         q0_1  q1_1  q2_1  q3_1
+         q0         q0_0  q1_0  q2_0  q3_0
+    */
+    __asm__ __volatile__(
+        "precrq.qb.ph   %[prim1],   %[q3],      %[q2]       \n\t"
+        "precr.qb.ph    %[prim2],   %[q3],      %[q2]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[q1],      %[q0]       \n\t"
+        "precr.qb.ph    %[prim4],   %[q1],      %[q0]       \n\t"
+
+        "precrq.qb.ph   %[q2],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[q0],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[q3],      %[q2],      %[sec3]     \n\t"
+        "precrq.ph.w    %[q1],      %[q0],      %[sec4]     \n\t"
+        "append         %[q2],      %[sec3],    16          \n\t"
+        "append         %[q0],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3),
+          [prim4] "=&r"(prim4), [q3] "+r"(q3), [q2] "+r"(q2), [q1] "+r"(q1),
+          [q0] "+r"(q0), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4)
+        :);
+
+    filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec, p1, p0,
+                                    p3, p2, q0, q1, q2, q3, &hev, &mask, &flat);
+
+    if ((flat == 0) && (mask != 0)) {
+      filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+      STORE_F0()
+    } else if ((mask & flat) == 0xFFFFFFFF) {
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r);
+
+      STORE_F1()
+    } else if ((flat != 0) && (mask != 0)) {
+      filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r);
+
+      if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb         %[p2_r],  -3(%[s4])    \n\t"
+            "sb         %[p1_r],  -2(%[s4])    \n\t"
+            "sb         %[p0_r],  -1(%[s4])    \n\t"
+            "sb         %[q0_r],    (%[s4])    \n\t"
+            "sb         %[q1_r],  +1(%[s4])    \n\t"
+            "sb         %[q2_r],  +2(%[s4])    \n\t"
+
+            :
+            : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r),
+              [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [s4] "r"(s4));
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  -2(%[s4])    \n\t"
+            "sb         %[p0_f0],  -1(%[s4])    \n\t"
+            "sb         %[q0_f0],    (%[s4])    \n\t"
+            "sb         %[q1_f0],  +1(%[s4])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s4] "r"(s4));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p2_r],    %[p2_r],    16      \n\t"
+          "srl      %[p1_r],    %[p1_r],    16      \n\t"
+          "srl      %[p0_r],    %[p0_r],    16      \n\t"
+          "srl      %[q0_r],    %[q0_r],    16      \n\t"
+          "srl      %[q1_r],    %[q1_r],    16      \n\t"
+          "srl      %[q2_r],    %[q2_r],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_r] "+r"(p2_r), [p1_r] "+r"(p1_r), [p0_r] "+r"(p0_r),
+            [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb         %[p2_r],  -3(%[s3])    \n\t"
+            "sb         %[p1_r],  -2(%[s3])    \n\t"
+            "sb         %[p0_r],  -1(%[s3])    \n\t"
+            "sb         %[q0_r],    (%[s3])    \n\t"
+            "sb         %[q1_r],  +1(%[s3])    \n\t"
+            "sb         %[q2_r],  +2(%[s3])    \n\t"
+
+            :
+            : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r),
+              [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [s3] "r"(s3));
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  -2(%[s3])    \n\t"
+            "sb         %[p0_f0],  -1(%[s3])    \n\t"
+            "sb         %[q0_f0],    (%[s3])    \n\t"
+            "sb         %[q1_f0],  +1(%[s3])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s3] "r"(s3));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p1_f0],   %[p1_f0],   8     \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8     \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8     \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8     \n\t"
+
+          : [p2] "+r"(p2), [p1] "+r"(p1), [p0] "+r"(p0), [q0] "+r"(q0),
+            [q1] "+r"(q1), [q2] "+r"(q2), [p1_f0] "+r"(p1_f0),
+            [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0), [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb         %[p2_l],  -3(%[s2])    \n\t"
+            "sb         %[p1_l],  -2(%[s2])    \n\t"
+            "sb         %[p0_l],  -1(%[s2])    \n\t"
+            "sb         %[q0_l],    (%[s2])    \n\t"
+            "sb         %[q1_l],  +1(%[s2])    \n\t"
+            "sb         %[q2_l],  +2(%[s2])    \n\t"
+
+            :
+            : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l),
+              [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [s2] "r"(s2));
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  -2(%[s2])    \n\t"
+            "sb         %[p0_f0],  -1(%[s2])    \n\t"
+            "sb         %[q0_f0],    (%[s2])    \n\t"
+            "sb         %[q1_f0],  +1(%[s2])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s2] "r"(s2));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p2_l],    %[p2_l],    16      \n\t"
+          "srl      %[p1_l],    %[p1_l],    16      \n\t"
+          "srl      %[p0_l],    %[p0_l],    16      \n\t"
+          "srl      %[q0_l],    %[q0_l],    16      \n\t"
+          "srl      %[q1_l],    %[q1_l],    16      \n\t"
+          "srl      %[q2_l],    %[q2_l],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_l] "+r"(p2_l), [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l),
+            [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb         %[p2_l],  -3(%[s1])    \n\t"
+            "sb         %[p1_l],  -2(%[s1])    \n\t"
+            "sb         %[p0_l],  -1(%[s1])    \n\t"
+            "sb         %[q0_l],    (%[s1])    \n\t"
+            "sb         %[q1_l],  +1(%[s1])    \n\t"
+            "sb         %[q2_l],  +2(%[s1])    \n\t"
+
+            :
+            : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l),
+              [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [s1] "r"(s1));
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  -2(%[s1])    \n\t"
+            "sb         %[p0_f0],  -1(%[s1])    \n\t"
+            "sb         %[q0_f0],    (%[s1])    \n\t"
+            "sb         %[q1_f0],  +1(%[s1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s1] "r"(s1));
+      }
+    }
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_mb_horiz_dspr2.c b/third_party/aom/aom_dsp/mips/loopfilter_mb_horiz_dspr2.c
new file mode 100644
index 0000000000..8d2fd69f7f
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_mb_horiz_dspr2.c
@@ -0,0 +1,734 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/mips/common_dspr2.h"
+#include "aom_dsp/mips/loopfilter_filters_dspr2.h"
+#include "aom_dsp/mips/loopfilter_macros_dspr2.h"
+#include "aom_dsp/mips/loopfilter_masks_dspr2.h"
+#include "aom_mem/aom_mem.h"
+
+#if HAVE_DSPR2
+static void mb_lpf_horizontal_edge(unsigned char *s, int pitch,
+                                   const uint8_t *blimit, const uint8_t *limit,
+                                   const uint8_t *thresh, int count) {
+  uint32_t mask;
+  uint32_t hev, flat, flat2;
+  uint8_t i;
+  uint8_t *sp7, *sp6, *sp5, *sp4, *sp3, *sp2, *sp1, *sp0;
+  uint8_t *sq0, *sq1, *sq2, *sq3, *sq4, *sq5, *sq6, *sq7;
+  uint32_t thresh_vec, flimit_vec, limit_vec;
+  uint32_t uflimit, ulimit, uthresh;
+  uint32_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  uint32_t p1_f0, p0_f0, q0_f0, q1_f0;
+  uint32_t p7_l, p6_l, p5_l, p4_l, p3_l, p2_l, p1_l, p0_l;
+  uint32_t q0_l, q1_l, q2_l, q3_l, q4_l, q5_l, q6_l, q7_l;
+  uint32_t p7_r, p6_r, p5_r, p4_r, p3_r, p2_r, p1_r, p0_r;
+  uint32_t q0_r, q1_r, q2_r, q3_r, q4_r, q5_r, q6_r, q7_r;
+  uint32_t p2_l_f1, p1_l_f1, p0_l_f1, p2_r_f1, p1_r_f1, p0_r_f1;
+  uint32_t q0_l_f1, q1_l_f1, q2_l_f1, q0_r_f1, q1_r_f1, q2_r_f1;
+
+  uflimit = *blimit;
+  ulimit = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__(
+      "replv.qb       %[thresh_vec],    %[uthresh]      \n\t"
+      "replv.qb       %[flimit_vec],    %[uflimit]      \n\t"
+      "replv.qb       %[limit_vec],     %[ulimit]       \n\t"
+
+      : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec),
+        [limit_vec] "=r"(limit_vec)
+      : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit));
+
+  /* prefetch data for store */
+  prefetch_store(s);
+
+  for (i = 0; i < (2 * count); i++) {
+    sp7 = s - (pitch << 3);
+    sp6 = sp7 + pitch;
+    sp5 = sp6 + pitch;
+    sp4 = sp5 + pitch;
+    sp3 = sp4 + pitch;
+    sp2 = sp3 + pitch;
+    sp1 = sp2 + pitch;
+    sp0 = sp1 + pitch;
+    sq0 = s;
+    sq1 = s + pitch;
+    sq2 = sq1 + pitch;
+    sq3 = sq2 + pitch;
+    sq4 = sq3 + pitch;
+    sq5 = sq4 + pitch;
+    sq6 = sq5 + pitch;
+    sq7 = sq6 + pitch;
+
+    __asm__ __volatile__(
+        "lw     %[p7],      (%[sp7])            \n\t"
+        "lw     %[p6],      (%[sp6])            \n\t"
+        "lw     %[p5],      (%[sp5])            \n\t"
+        "lw     %[p4],      (%[sp4])            \n\t"
+        "lw     %[p3],      (%[sp3])            \n\t"
+        "lw     %[p2],      (%[sp2])            \n\t"
+        "lw     %[p1],      (%[sp1])            \n\t"
+        "lw     %[p0],      (%[sp0])            \n\t"
+
+        : [p3] "=&r"(p3), [p2] "=&r"(p2), [p1] "=&r"(p1), [p0] "=&r"(p0),
+          [p7] "=&r"(p7), [p6] "=&r"(p6), [p5] "=&r"(p5), [p4] "=&r"(p4)
+        : [sp3] "r"(sp3), [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0),
+          [sp4] "r"(sp4), [sp5] "r"(sp5), [sp6] "r"(sp6), [sp7] "r"(sp7));
+
+    __asm__ __volatile__(
+        "lw     %[q0],      (%[sq0])            \n\t"
+        "lw     %[q1],      (%[sq1])            \n\t"
+        "lw     %[q2],      (%[sq2])            \n\t"
+        "lw     %[q3],      (%[sq3])            \n\t"
+        "lw     %[q4],      (%[sq4])            \n\t"
+        "lw     %[q5],      (%[sq5])            \n\t"
+        "lw     %[q6],      (%[sq6])            \n\t"
+        "lw     %[q7],      (%[sq7])            \n\t"
+
+        : [q3] "=&r"(q3), [q2] "=&r"(q2), [q1] "=&r"(q1), [q0] "=&r"(q0),
+          [q7] "=&r"(q7), [q6] "=&r"(q6), [q5] "=&r"(q5), [q4] "=&r"(q4)
+        : [sq3] "r"(sq3), [sq2] "r"(sq2), [sq1] "r"(sq1), [sq0] "r"(sq0),
+          [sq4] "r"(sq4), [sq5] "r"(sq5), [sq6] "r"(sq6), [sq7] "r"(sq7));
+
+    filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec, p1, p0,
+                                    p3, p2, q0, q1, q2, q3, &hev, &mask, &flat);
+
+    flatmask5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, &flat2);
+
+    /* f0 */
+    if (((flat2 == 0) && (flat == 0) && (mask != 0)) ||
+        ((flat2 != 0) && (flat == 0) && (mask != 0))) {
+      filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      __asm__ __volatile__(
+          "sw       %[p1_f0],   (%[sp1])            \n\t"
+          "sw       %[p0_f0],   (%[sp0])            \n\t"
+          "sw       %[q0_f0],   (%[sq0])            \n\t"
+          "sw       %[q1_f0],   (%[sq1])            \n\t"
+
+          :
+          : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+            [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0),
+            [sq1] "r"(sq1));
+    } else if ((flat2 == 0XFFFFFFFF) && (flat == 0xFFFFFFFF) &&
+               (mask == 0xFFFFFFFF)) {
+      /* f2 */
+      PACK_LEFT_0TO3()
+      PACK_LEFT_4TO7()
+      wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l, &p3_l, &p2_l, &p1_l,
+                          &p0_l, &q0_l, &q1_l, &q2_l, &q3_l, &q4_l, &q5_l,
+                          &q6_l, &q7_l);
+
+      PACK_RIGHT_0TO3()
+      PACK_RIGHT_4TO7()
+      wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r, &p3_r, &p2_r, &p1_r,
+                          &p0_r, &q0_r, &q1_r, &q2_r, &q3_r, &q4_r, &q5_r,
+                          &q6_r, &q7_r);
+
+      COMBINE_LEFT_RIGHT_0TO2()
+      COMBINE_LEFT_RIGHT_3TO6()
+
+      __asm__ __volatile__(
+          "sw         %[p6], (%[sp6])    \n\t"
+          "sw         %[p5], (%[sp5])    \n\t"
+          "sw         %[p4], (%[sp4])    \n\t"
+          "sw         %[p3], (%[sp3])    \n\t"
+          "sw         %[p2], (%[sp2])    \n\t"
+          "sw         %[p1], (%[sp1])    \n\t"
+          "sw         %[p0], (%[sp0])    \n\t"
+
+          :
+          : [p6] "r"(p6), [p5] "r"(p5), [p4] "r"(p4), [p3] "r"(p3),
+            [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), [sp6] "r"(sp6),
+            [sp5] "r"(sp5), [sp4] "r"(sp4), [sp3] "r"(sp3), [sp2] "r"(sp2),
+            [sp1] "r"(sp1), [sp0] "r"(sp0));
+
+      __asm__ __volatile__(
+          "sw         %[q6], (%[sq6])    \n\t"
+          "sw         %[q5], (%[sq5])    \n\t"
+          "sw         %[q4], (%[sq4])    \n\t"
+          "sw         %[q3], (%[sq3])    \n\t"
+          "sw         %[q2], (%[sq2])    \n\t"
+          "sw         %[q1], (%[sq1])    \n\t"
+          "sw         %[q0], (%[sq0])    \n\t"
+
+          :
+          : [q6] "r"(q6), [q5] "r"(q5), [q4] "r"(q4), [q3] "r"(q3),
+            [q2] "r"(q2), [q1] "r"(q1), [q0] "r"(q0), [sq6] "r"(sq6),
+            [sq5] "r"(sq5), [sq4] "r"(sq4), [sq3] "r"(sq3), [sq2] "r"(sq2),
+            [sq1] "r"(sq1), [sq0] "r"(sq0));
+    } else if ((flat2 == 0) && (flat == 0xFFFFFFFF) && (mask == 0xFFFFFFFF)) {
+      /* f1 */
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r);
+
+      COMBINE_LEFT_RIGHT_0TO2()
+
+      __asm__ __volatile__(
+          "sw         %[p2], (%[sp2])    \n\t"
+          "sw         %[p1], (%[sp1])    \n\t"
+          "sw         %[p0], (%[sp0])    \n\t"
+          "sw         %[q0], (%[sq0])    \n\t"
+          "sw         %[q1], (%[sq1])    \n\t"
+          "sw         %[q2], (%[sq2])    \n\t"
+
+          :
+          : [p2] "r"(p2), [p1] "r"(p1), [p0] "r"(p0), [q0] "r"(q0),
+            [q1] "r"(q1), [q2] "r"(q2), [sp2] "r"(sp2), [sp1] "r"(sp1),
+            [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2));
+    } else if ((flat2 == 0) && (flat != 0) && (mask != 0)) {
+      /* f0+f1 */
+      filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r);
+
+      if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb         %[p2_r],  (%[sp2])    \n\t"
+            "sb         %[p1_r],  (%[sp1])    \n\t"
+            "sb         %[p0_r],  (%[sp0])    \n\t"
+            "sb         %[q0_r],  (%[sq0])    \n\t"
+            "sb         %[q1_r],  (%[sq1])    \n\t"
+            "sb         %[q2_r],  (%[sq2])    \n\t"
+
+            :
+            : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r),
+              [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0),
+              [sq1] "r"(sq1), [sq2] "r"(sq2));
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  (%[sp1])    \n\t"
+            "sb         %[p0_f0],  (%[sp0])    \n\t"
+            "sb         %[q0_f0],  (%[sq0])    \n\t"
+            "sb         %[q1_f0],  (%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p2_r],    %[p2_r],    16      \n\t"
+          "srl      %[p1_r],    %[p1_r],    16      \n\t"
+          "srl      %[p0_r],    %[p0_r],    16      \n\t"
+          "srl      %[q0_r],    %[q0_r],    16      \n\t"
+          "srl      %[q1_r],    %[q1_r],    16      \n\t"
+          "srl      %[q2_r],    %[q2_r],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_r] "+r"(p2_r), [p1_r] "+r"(p1_r), [p0_r] "+r"(p0_r),
+            [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb         %[p2_r],  +1(%[sp2])    \n\t"
+            "sb         %[p1_r],  +1(%[sp1])    \n\t"
+            "sb         %[p0_r],  +1(%[sp0])    \n\t"
+            "sb         %[q0_r],  +1(%[sq0])    \n\t"
+            "sb         %[q1_r],  +1(%[sq1])    \n\t"
+            "sb         %[q2_r],  +1(%[sq2])    \n\t"
+
+            :
+            : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r),
+              [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0),
+              [sq1] "r"(sq1), [sq2] "r"(sq2));
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  +1(%[sp1])    \n\t"
+            "sb         %[p0_f0],  +1(%[sp0])    \n\t"
+            "sb         %[q0_f0],  +1(%[sq0])    \n\t"
+            "sb         %[q1_f0],  +1(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p1_f0],   %[p1_f0],   8     \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8     \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8     \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8     \n\t"
+
+          : [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb         %[p2_l],  +2(%[sp2])    \n\t"
+            "sb         %[p1_l],  +2(%[sp1])    \n\t"
+            "sb         %[p0_l],  +2(%[sp0])    \n\t"
+            "sb         %[q0_l],  +2(%[sq0])    \n\t"
+            "sb         %[q1_l],  +2(%[sq1])    \n\t"
+            "sb         %[q2_l],  +2(%[sq2])    \n\t"
+
+            :
+            : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l),
+              [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0),
+              [sq1] "r"(sq1), [sq2] "r"(sq2));
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  +2(%[sp1])    \n\t"
+            "sb         %[p0_f0],  +2(%[sp0])    \n\t"
+            "sb         %[q0_f0],  +2(%[sq0])    \n\t"
+            "sb         %[q1_f0],  +2(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p2_l],    %[p2_l],    16      \n\t"
+          "srl      %[p1_l],    %[p1_l],    16      \n\t"
+          "srl      %[p0_l],    %[p0_l],    16      \n\t"
+          "srl      %[q0_l],    %[q0_l],    16      \n\t"
+          "srl      %[q1_l],    %[q1_l],    16      \n\t"
+          "srl      %[q2_l],    %[q2_l],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_l] "+r"(p2_l), [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l),
+            [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb         %[p2_l],  +3(%[sp2])    \n\t"
+            "sb         %[p1_l],  +3(%[sp1])    \n\t"
+            "sb         %[p0_l],  +3(%[sp0])    \n\t"
+            "sb         %[q0_l],  +3(%[sq0])    \n\t"
+            "sb         %[q1_l],  +3(%[sq1])    \n\t"
+            "sb         %[q2_l],  +3(%[sq2])    \n\t"
+
+            :
+            : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l),
+              [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0),
+              [sq1] "r"(sq1), [sq2] "r"(sq2));
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  +3(%[sp1])    \n\t"
+            "sb         %[p0_f0],  +3(%[sp0])    \n\t"
+            "sb         %[q0_f0],  +3(%[sq0])    \n\t"
+            "sb         %[q1_f0],  +3(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+    } else if ((flat2 != 0) && (flat != 0) && (mask != 0)) {
+      /* f0 + f1 + f2 */
+      /* f0  function */
+      filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      /* f1  function */
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter1_dspr2(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, &p2_l_f1,
+                      &p1_l_f1, &p0_l_f1, &q0_l_f1, &q1_l_f1, &q2_l_f1);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter1_dspr2(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, &p2_r_f1,
+                      &p1_r_f1, &p0_r_f1, &q0_r_f1, &q1_r_f1, &q2_r_f1);
+
+      /* f2  function */
+      PACK_LEFT_4TO7()
+      wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l, &p3_l, &p2_l, &p1_l,
+                          &p0_l, &q0_l, &q1_l, &q2_l, &q3_l, &q4_l, &q5_l,
+                          &q6_l, &q7_l);
+
+      PACK_RIGHT_4TO7()
+      wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r, &p3_r, &p2_r, &p1_r,
+                          &p0_r, &q0_r, &q1_r, &q2_r, &q3_r, &q4_r, &q5_r,
+                          &q6_r, &q7_r);
+
+      if (mask & flat & flat2 & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb         %[p6_r],  (%[sp6])    \n\t"
+            "sb         %[p5_r],  (%[sp5])    \n\t"
+            "sb         %[p4_r],  (%[sp4])    \n\t"
+            "sb         %[p3_r],  (%[sp3])    \n\t"
+            "sb         %[p2_r],  (%[sp2])    \n\t"
+            "sb         %[p1_r],  (%[sp1])    \n\t"
+            "sb         %[p0_r],  (%[sp0])    \n\t"
+
+            :
+            : [p6_r] "r"(p6_r), [p5_r] "r"(p5_r), [p4_r] "r"(p4_r),
+              [p3_r] "r"(p3_r), [p2_r] "r"(p2_r), [p1_r] "r"(p1_r),
+              [sp6] "r"(sp6), [sp5] "r"(sp5), [sp4] "r"(sp4), [sp3] "r"(sp3),
+              [sp2] "r"(sp2), [sp1] "r"(sp1), [p0_r] "r"(p0_r), [sp0] "r"(sp0));
+
+        __asm__ __volatile__(
+            "sb         %[q0_r],  (%[sq0])    \n\t"
+            "sb         %[q1_r],  (%[sq1])    \n\t"
+            "sb         %[q2_r],  (%[sq2])    \n\t"
+            "sb         %[q3_r],  (%[sq3])    \n\t"
+            "sb         %[q4_r],  (%[sq4])    \n\t"
+            "sb         %[q5_r],  (%[sq5])    \n\t"
+            "sb         %[q6_r],  (%[sq6])    \n\t"
+
+            :
+            : [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [q3_r] "r"(q3_r), [q4_r] "r"(q4_r), [q5_r] "r"(q5_r),
+              [q6_r] "r"(q6_r), [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2),
+              [sq3] "r"(sq3), [sq4] "r"(sq4), [sq5] "r"(sq5), [sq6] "r"(sq6));
+      } else if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb         %[p2_r_f1],  (%[sp2])    \n\t"
+            "sb         %[p1_r_f1],  (%[sp1])    \n\t"
+            "sb         %[p0_r_f1],  (%[sp0])    \n\t"
+            "sb         %[q0_r_f1],  (%[sq0])    \n\t"
+            "sb         %[q1_r_f1],  (%[sq1])    \n\t"
+            "sb         %[q2_r_f1],  (%[sq2])    \n\t"
+
+            :
+            : [p2_r_f1] "r"(p2_r_f1), [p1_r_f1] "r"(p1_r_f1),
+              [p0_r_f1] "r"(p0_r_f1), [q0_r_f1] "r"(q0_r_f1),
+              [q1_r_f1] "r"(q1_r_f1), [q2_r_f1] "r"(q2_r_f1), [sp2] "r"(sp2),
+              [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1),
+              [sq2] "r"(sq2));
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  (%[sp1])    \n\t"
+            "sb         %[p0_f0],  (%[sp0])    \n\t"
+            "sb         %[q0_f0],  (%[sq0])    \n\t"
+            "sb         %[q1_f0],  (%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+
+      __asm__ __volatile__(
+          "srl        %[p6_r], %[p6_r], 16     \n\t"
+          "srl        %[p5_r], %[p5_r], 16     \n\t"
+          "srl        %[p4_r], %[p4_r], 16     \n\t"
+          "srl        %[p3_r], %[p3_r], 16     \n\t"
+          "srl        %[p2_r], %[p2_r], 16     \n\t"
+          "srl        %[p1_r], %[p1_r], 16     \n\t"
+          "srl        %[p0_r], %[p0_r], 16     \n\t"
+          "srl        %[q0_r], %[q0_r], 16     \n\t"
+          "srl        %[q1_r], %[q1_r], 16     \n\t"
+          "srl        %[q2_r], %[q2_r], 16     \n\t"
+          "srl        %[q3_r], %[q3_r], 16     \n\t"
+          "srl        %[q4_r], %[q4_r], 16     \n\t"
+          "srl        %[q5_r], %[q5_r], 16     \n\t"
+          "srl        %[q6_r], %[q6_r], 16     \n\t"
+
+          : [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r),
+            [q3_r] "+r"(q3_r), [q4_r] "+r"(q4_r), [q5_r] "+r"(q5_r),
+            [p6_r] "+r"(p6_r), [p5_r] "+r"(p5_r), [p4_r] "+r"(p4_r),
+            [p3_r] "+r"(p3_r), [p2_r] "+r"(p2_r), [p1_r] "+r"(p1_r),
+            [q6_r] "+r"(q6_r), [p0_r] "+r"(p0_r)
+          :);
+
+      __asm__ __volatile__(
+          "srl        %[p2_r_f1], %[p2_r_f1], 16     \n\t"
+          "srl        %[p1_r_f1], %[p1_r_f1], 16     \n\t"
+          "srl        %[p0_r_f1], %[p0_r_f1], 16     \n\t"
+          "srl        %[q0_r_f1], %[q0_r_f1], 16     \n\t"
+          "srl        %[q1_r_f1], %[q1_r_f1], 16     \n\t"
+          "srl        %[q2_r_f1], %[q2_r_f1], 16     \n\t"
+          "srl        %[p1_f0],   %[p1_f0],   8      \n\t"
+          "srl        %[p0_f0],   %[p0_f0],   8      \n\t"
+          "srl        %[q0_f0],   %[q0_f0],   8      \n\t"
+          "srl        %[q1_f0],   %[q1_f0],   8      \n\t"
+
+          : [p2_r_f1] "+r"(p2_r_f1), [p1_r_f1] "+r"(p1_r_f1),
+            [p0_r_f1] "+r"(p0_r_f1), [q0_r_f1] "+r"(q0_r_f1),
+            [q1_r_f1] "+r"(q1_r_f1), [q2_r_f1] "+r"(q2_r_f1),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & flat2 & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb         %[p6_r],  +1(%[sp6])    \n\t"
+            "sb         %[p5_r],  +1(%[sp5])    \n\t"
+            "sb         %[p4_r],  +1(%[sp4])    \n\t"
+            "sb         %[p3_r],  +1(%[sp3])    \n\t"
+            "sb         %[p2_r],  +1(%[sp2])    \n\t"
+            "sb         %[p1_r],  +1(%[sp1])    \n\t"
+            "sb         %[p0_r],  +1(%[sp0])    \n\t"
+
+            :
+            : [p6_r] "r"(p6_r), [p5_r] "r"(p5_r), [p4_r] "r"(p4_r),
+              [p3_r] "r"(p3_r), [p2_r] "r"(p2_r), [p1_r] "r"(p1_r),
+              [p0_r] "r"(p0_r), [sp6] "r"(sp6), [sp5] "r"(sp5), [sp4] "r"(sp4),
+              [sp3] "r"(sp3), [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0));
+
+        __asm__ __volatile__(
+            "sb         %[q0_r],  +1(%[sq0])    \n\t"
+            "sb         %[q1_r],  +1(%[sq1])    \n\t"
+            "sb         %[q2_r],  +1(%[sq2])    \n\t"
+            "sb         %[q3_r],  +1(%[sq3])    \n\t"
+            "sb         %[q4_r],  +1(%[sq4])    \n\t"
+            "sb         %[q5_r],  +1(%[sq5])    \n\t"
+            "sb         %[q6_r],  +1(%[sq6])    \n\t"
+
+            :
+            : [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [q3_r] "r"(q3_r), [q4_r] "r"(q4_r), [q5_r] "r"(q5_r),
+              [q6_r] "r"(q6_r), [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2),
+              [sq3] "r"(sq3), [sq4] "r"(sq4), [sq5] "r"(sq5), [sq6] "r"(sq6));
+      } else if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb         %[p2_r_f1],  +1(%[sp2])    \n\t"
+            "sb         %[p1_r_f1],  +1(%[sp1])    \n\t"
+            "sb         %[p0_r_f1],  +1(%[sp0])    \n\t"
+            "sb         %[q0_r_f1],  +1(%[sq0])    \n\t"
+            "sb         %[q1_r_f1],  +1(%[sq1])    \n\t"
+            "sb         %[q2_r_f1],  +1(%[sq2])    \n\t"
+
+            :
+            : [p2_r_f1] "r"(p2_r_f1), [p1_r_f1] "r"(p1_r_f1),
+              [p0_r_f1] "r"(p0_r_f1), [q0_r_f1] "r"(q0_r_f1),
+              [q1_r_f1] "r"(q1_r_f1), [q2_r_f1] "r"(q2_r_f1), [sp2] "r"(sp2),
+              [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1),
+              [sq2] "r"(sq2));
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  +1(%[sp1])    \n\t"
+            "sb         %[p0_f0],  +1(%[sp0])    \n\t"
+            "sb         %[q0_f0],  +1(%[sq0])    \n\t"
+            "sb         %[q1_f0],  +1(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+
+      __asm__ __volatile__(
+          "srl        %[p1_f0], %[p1_f0], 8     \n\t"
+          "srl        %[p0_f0], %[p0_f0], 8     \n\t"
+          "srl        %[q0_f0], %[q0_f0], 8     \n\t"
+          "srl        %[q1_f0], %[q1_f0], 8     \n\t"
+
+          : [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & flat2 & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb         %[p6_l],  +2(%[sp6])    \n\t"
+            "sb         %[p5_l],  +2(%[sp5])    \n\t"
+            "sb         %[p4_l],  +2(%[sp4])    \n\t"
+            "sb         %[p3_l],  +2(%[sp3])    \n\t"
+            "sb         %[p2_l],  +2(%[sp2])    \n\t"
+            "sb         %[p1_l],  +2(%[sp1])    \n\t"
+            "sb         %[p0_l],  +2(%[sp0])    \n\t"
+
+            :
+            : [p6_l] "r"(p6_l), [p5_l] "r"(p5_l), [p4_l] "r"(p4_l),
+              [p3_l] "r"(p3_l), [p2_l] "r"(p2_l), [p1_l] "r"(p1_l),
+              [p0_l] "r"(p0_l), [sp6] "r"(sp6), [sp5] "r"(sp5), [sp4] "r"(sp4),
+              [sp3] "r"(sp3), [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0));
+
+        __asm__ __volatile__(
+            "sb         %[q0_l],  +2(%[sq0])    \n\t"
+            "sb         %[q1_l],  +2(%[sq1])    \n\t"
+            "sb         %[q2_l],  +2(%[sq2])    \n\t"
+            "sb         %[q3_l],  +2(%[sq3])    \n\t"
+            "sb         %[q4_l],  +2(%[sq4])    \n\t"
+            "sb         %[q5_l],  +2(%[sq5])    \n\t"
+            "sb         %[q6_l],  +2(%[sq6])    \n\t"
+
+            :
+            : [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [q3_l] "r"(q3_l), [q4_l] "r"(q4_l), [q5_l] "r"(q5_l),
+              [q6_l] "r"(q6_l), [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2),
+              [sq3] "r"(sq3), [sq4] "r"(sq4), [sq5] "r"(sq5), [sq6] "r"(sq6));
+      } else if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb         %[p2_l_f1],  +2(%[sp2])    \n\t"
+            "sb         %[p1_l_f1],  +2(%[sp1])    \n\t"
+            "sb         %[p0_l_f1],  +2(%[sp0])    \n\t"
+            "sb         %[q0_l_f1],  +2(%[sq0])    \n\t"
+            "sb         %[q1_l_f1],  +2(%[sq1])    \n\t"
+            "sb         %[q2_l_f1],  +2(%[sq2])    \n\t"
+
+            :
+            : [p2_l_f1] "r"(p2_l_f1), [p1_l_f1] "r"(p1_l_f1),
+              [p0_l_f1] "r"(p0_l_f1), [q0_l_f1] "r"(q0_l_f1),
+              [q1_l_f1] "r"(q1_l_f1), [q2_l_f1] "r"(q2_l_f1), [sp2] "r"(sp2),
+              [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1),
+              [sq2] "r"(sq2));
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  +2(%[sp1])    \n\t"
+            "sb         %[p0_f0],  +2(%[sp0])    \n\t"
+            "sb         %[q0_f0],  +2(%[sq0])    \n\t"
+            "sb         %[q1_f0],  +2(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p6_l],    %[p6_l],    16   \n\t"
+          "srl      %[p5_l],    %[p5_l],    16   \n\t"
+          "srl      %[p4_l],    %[p4_l],    16   \n\t"
+          "srl      %[p3_l],    %[p3_l],    16   \n\t"
+          "srl      %[p2_l],    %[p2_l],    16   \n\t"
+          "srl      %[p1_l],    %[p1_l],    16   \n\t"
+          "srl      %[p0_l],    %[p0_l],    16   \n\t"
+          "srl      %[q0_l],    %[q0_l],    16   \n\t"
+          "srl      %[q1_l],    %[q1_l],    16   \n\t"
+          "srl      %[q2_l],    %[q2_l],    16   \n\t"
+          "srl      %[q3_l],    %[q3_l],    16   \n\t"
+          "srl      %[q4_l],    %[q4_l],    16   \n\t"
+          "srl      %[q5_l],    %[q5_l],    16   \n\t"
+          "srl      %[q6_l],    %[q6_l],    16   \n\t"
+
+          : [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l),
+            [q3_l] "+r"(q3_l), [q4_l] "+r"(q4_l), [q5_l] "+r"(q5_l),
+            [q6_l] "+r"(q6_l), [p6_l] "+r"(p6_l), [p5_l] "+r"(p5_l),
+            [p4_l] "+r"(p4_l), [p3_l] "+r"(p3_l), [p2_l] "+r"(p2_l),
+            [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l)
+          :);
+
+      __asm__ __volatile__(
+          "srl      %[p2_l_f1],   %[p2_l_f1],   16   \n\t"
+          "srl      %[p1_l_f1],   %[p1_l_f1],   16   \n\t"
+          "srl      %[p0_l_f1],   %[p0_l_f1],   16   \n\t"
+          "srl      %[q0_l_f1],   %[q0_l_f1],   16   \n\t"
+          "srl      %[q1_l_f1],   %[q1_l_f1],   16   \n\t"
+          "srl      %[q2_l_f1],   %[q2_l_f1],   16   \n\t"
+          "srl      %[p1_f0],     %[p1_f0],     8    \n\t"
+          "srl      %[p0_f0],     %[p0_f0],     8    \n\t"
+          "srl      %[q0_f0],     %[q0_f0],     8    \n\t"
+          "srl      %[q1_f0],     %[q1_f0],     8    \n\t"
+
+          : [p2_l_f1] "+r"(p2_l_f1), [p1_l_f1] "+r"(p1_l_f1),
+            [p0_l_f1] "+r"(p0_l_f1), [q0_l_f1] "+r"(q0_l_f1),
+            [q1_l_f1] "+r"(q1_l_f1), [q2_l_f1] "+r"(q2_l_f1),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & flat2 & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb     %[p6_l],    +3(%[sp6])    \n\t"
+            "sb     %[p5_l],    +3(%[sp5])    \n\t"
+            "sb     %[p4_l],    +3(%[sp4])    \n\t"
+            "sb     %[p3_l],    +3(%[sp3])    \n\t"
+            "sb     %[p2_l],    +3(%[sp2])    \n\t"
+            "sb     %[p1_l],    +3(%[sp1])    \n\t"
+            "sb     %[p0_l],    +3(%[sp0])    \n\t"
+
+            :
+            : [p6_l] "r"(p6_l), [p5_l] "r"(p5_l), [p4_l] "r"(p4_l),
+              [p3_l] "r"(p3_l), [p2_l] "r"(p2_l), [p1_l] "r"(p1_l),
+              [p0_l] "r"(p0_l), [sp6] "r"(sp6), [sp5] "r"(sp5), [sp4] "r"(sp4),
+              [sp3] "r"(sp3), [sp2] "r"(sp2), [sp1] "r"(sp1), [sp0] "r"(sp0));
+
+        __asm__ __volatile__(
+            "sb     %[q0_l],    +3(%[sq0])    \n\t"
+            "sb     %[q1_l],    +3(%[sq1])    \n\t"
+            "sb     %[q2_l],    +3(%[sq2])    \n\t"
+            "sb     %[q3_l],    +3(%[sq3])    \n\t"
+            "sb     %[q4_l],    +3(%[sq4])    \n\t"
+            "sb     %[q5_l],    +3(%[sq5])    \n\t"
+            "sb     %[q6_l],    +3(%[sq6])    \n\t"
+
+            :
+            : [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [q3_l] "r"(q3_l), [q4_l] "r"(q4_l), [q5_l] "r"(q5_l),
+              [sq0] "r"(sq0), [sq1] "r"(sq1), [sq2] "r"(sq2), [sq3] "r"(sq3),
+              [sq4] "r"(sq4), [sq5] "r"(sq5), [q6_l] "r"(q6_l), [sq6] "r"(sq6));
+      } else if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb     %[p2_l_f1],     +3(%[sp2])    \n\t"
+            "sb     %[p1_l_f1],     +3(%[sp1])    \n\t"
+            "sb     %[p0_l_f1],     +3(%[sp0])    \n\t"
+            "sb     %[q0_l_f1],     +3(%[sq0])    \n\t"
+            "sb     %[q1_l_f1],     +3(%[sq1])    \n\t"
+            "sb     %[q2_l_f1],     +3(%[sq2])    \n\t"
+
+            :
+            : [p2_l_f1] "r"(p2_l_f1), [p1_l_f1] "r"(p1_l_f1),
+              [p0_l_f1] "r"(p0_l_f1), [q0_l_f1] "r"(q0_l_f1),
+              [q1_l_f1] "r"(q1_l_f1), [q2_l_f1] "r"(q2_l_f1), [sp2] "r"(sp2),
+              [sp1] "r"(sp1), [sp0] "r"(sp0), [sq0] "r"(sq0), [sq1] "r"(sq1),
+              [sq2] "r"(sq2));
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   +3(%[sp1])    \n\t"
+            "sb     %[p0_f0],   +3(%[sp0])    \n\t"
+            "sb     %[q0_f0],   +3(%[sq0])    \n\t"
+            "sb     %[q1_f0],   +3(%[sq1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [sp1] "r"(sp1), [sp0] "r"(sp0),
+              [sq0] "r"(sq0), [sq1] "r"(sq1));
+      }
+    }
+
+    s = s + 4;
+  }
+}
+
+void aom_lpf_horizontal_edge_8_dspr2(unsigned char *s, int pitch,
+                                     const uint8_t *blimit,
+                                     const uint8_t *limit,
+                                     const uint8_t *thresh) {
+  mb_lpf_horizontal_edge(s, pitch, blimit, limit, thresh, 1);
+}
+
+void aom_lpf_horizontal_edge_16_dspr2(unsigned char *s, int pitch,
+                                      const uint8_t *blimit,
+                                      const uint8_t *limit,
+                                      const uint8_t *thresh) {
+  mb_lpf_horizontal_edge(s, pitch, blimit, limit, thresh, 2);
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_mb_vert_dspr2.c b/third_party/aom/aom_dsp/mips/loopfilter_mb_vert_dspr2.c
new file mode 100644
index 0000000000..28528869b4
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_mb_vert_dspr2.c
@@ -0,0 +1,757 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/mips/common_dspr2.h"
+#include "aom_dsp/mips/loopfilter_filters_dspr2.h"
+#include "aom_dsp/mips/loopfilter_macros_dspr2.h"
+#include "aom_dsp/mips/loopfilter_masks_dspr2.h"
+#include "aom_mem/aom_mem.h"
+
+#if HAVE_DSPR2
+void aom_lpf_vertical_16_dspr2(uint8_t *s, int pitch, const uint8_t *blimit,
+                               const uint8_t *limit, const uint8_t *thresh) {
+  uint8_t i;
+  uint32_t mask, hev, flat, flat2;
+  uint8_t *s1, *s2, *s3, *s4;
+  uint32_t prim1, prim2, sec3, sec4, prim3, prim4;
+  uint32_t thresh_vec, flimit_vec, limit_vec;
+  uint32_t uflimit, ulimit, uthresh;
+  uint32_t p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+  uint32_t p1_f0, p0_f0, q0_f0, q1_f0;
+  uint32_t p7_l, p6_l, p5_l, p4_l, p3_l, p2_l, p1_l, p0_l;
+  uint32_t q0_l, q1_l, q2_l, q3_l, q4_l, q5_l, q6_l, q7_l;
+  uint32_t p7_r, p6_r, p5_r, p4_r, p3_r, p2_r, p1_r, p0_r;
+  uint32_t q0_r, q1_r, q2_r, q3_r, q4_r, q5_r, q6_r, q7_r;
+  uint32_t p2_l_f1, p1_l_f1, p0_l_f1, p2_r_f1, p1_r_f1, p0_r_f1;
+  uint32_t q0_l_f1, q1_l_f1, q2_l_f1, q0_r_f1, q1_r_f1, q2_r_f1;
+
+  uflimit = *blimit;
+  ulimit = *limit;
+  uthresh = *thresh;
+
+  /* create quad-byte */
+  __asm__ __volatile__(
+      "replv.qb     %[thresh_vec],     %[uthresh]    \n\t"
+      "replv.qb     %[flimit_vec],     %[uflimit]    \n\t"
+      "replv.qb     %[limit_vec],      %[ulimit]     \n\t"
+
+      : [thresh_vec] "=&r"(thresh_vec), [flimit_vec] "=&r"(flimit_vec),
+        [limit_vec] "=r"(limit_vec)
+      : [uthresh] "r"(uthresh), [uflimit] "r"(uflimit), [ulimit] "r"(ulimit));
+
+  prefetch_store(s + pitch);
+
+  for (i = 0; i < 2; i++) {
+    s1 = s;
+    s2 = s + pitch;
+    s3 = s2 + pitch;
+    s4 = s3 + pitch;
+    s = s4 + pitch;
+
+    __asm__ __volatile__(
+        "lw     %[p0],  -4(%[s1])    \n\t"
+        "lw     %[p1],  -4(%[s2])    \n\t"
+        "lw     %[p2],  -4(%[s3])    \n\t"
+        "lw     %[p3],  -4(%[s4])    \n\t"
+        "lw     %[p4],  -8(%[s1])    \n\t"
+        "lw     %[p5],  -8(%[s2])    \n\t"
+        "lw     %[p6],  -8(%[s3])    \n\t"
+        "lw     %[p7],  -8(%[s4])    \n\t"
+
+        : [p3] "=&r"(p3), [p2] "=&r"(p2), [p1] "=&r"(p1), [p0] "=&r"(p0),
+          [p7] "=&r"(p7), [p6] "=&r"(p6), [p5] "=&r"(p5), [p4] "=&r"(p4)
+        : [s1] "r"(s1), [s2] "r"(s2), [s3] "r"(s3), [s4] "r"(s4));
+
+    __asm__ __volatile__(
+        "lw     %[q3],  (%[s1])     \n\t"
+        "lw     %[q2],  (%[s2])     \n\t"
+        "lw     %[q1],  (%[s3])     \n\t"
+        "lw     %[q0],  (%[s4])     \n\t"
+        "lw     %[q7],  +4(%[s1])   \n\t"
+        "lw     %[q6],  +4(%[s2])   \n\t"
+        "lw     %[q5],  +4(%[s3])   \n\t"
+        "lw     %[q4],  +4(%[s4])   \n\t"
+
+        : [q3] "=&r"(q3), [q2] "=&r"(q2), [q1] "=&r"(q1), [q0] "=&r"(q0),
+          [q7] "=&r"(q7), [q6] "=&r"(q6), [q5] "=&r"(q5), [q4] "=&r"(q4)
+        : [s1] "r"(s1), [s2] "r"(s2), [s3] "r"(s3), [s4] "r"(s4));
+
+    /* transpose p3, p2, p1, p0
+       original (when loaded from memory)
+       register       -4    -3   -2     -1
+         p0         p0_0  p0_1  p0_2  p0_3
+         p1         p1_0  p1_1  p1_2  p1_3
+         p2         p2_0  p2_1  p2_2  p2_3
+         p3         p3_0  p3_1  p3_2  p3_3
+
+       after transpose
+       register
+         p0         p3_3  p2_3  p1_3  p0_3
+         p1         p3_2  p2_2  p1_2  p0_2
+         p2         p3_1  p2_1  p1_1  p0_1
+         p3         p3_0  p2_0  p1_0  p0_0
+    */
+    __asm__ __volatile__(
+        "precrq.qb.ph   %[prim1],   %[p0],      %[p1]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p0],      %[p1]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p2],      %[p3]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p2],      %[p3]       \n\t"
+
+        "precrq.qb.ph   %[p1],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p3],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p0],      %[p1],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p2],      %[p3],      %[sec4]     \n\t"
+        "append         %[p1],      %[sec3],    16          \n\t"
+        "append         %[p3],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3),
+          [prim4] "=&r"(prim4), [p0] "+r"(p0), [p1] "+r"(p1), [p2] "+r"(p2),
+          [p3] "+r"(p3), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4)
+        :);
+
+    /* transpose q0, q1, q2, q3
+       original (when loaded from memory)
+       register       +1    +2    +3    +4
+         q3         q3_0  q3_1  q3_2  q3_3
+         q2         q2_0  q2_1  q2_2  q2_3
+         q1         q1_0  q1_1  q1_2  q1_3
+         q0         q0_0  q0_1  q0_2  q0_3
+
+       after transpose
+       register
+         q3         q0_3  q1_3  q2_3  q3_3
+         q2         q0_2  q1_2  q2_2  q3_2
+         q1         q0_1  q1_1  q2_1  q3_1
+         q0         q0_0  q1_0  q2_0  q3_0
+    */
+    __asm__ __volatile__(
+        "precrq.qb.ph   %[prim1],   %[q3],      %[q2]       \n\t"
+        "precr.qb.ph    %[prim2],   %[q3],      %[q2]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[q1],      %[q0]       \n\t"
+        "precr.qb.ph    %[prim4],   %[q1],      %[q0]       \n\t"
+
+        "precrq.qb.ph   %[q2],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[q0],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[q3],      %[q2],      %[sec3]     \n\t"
+        "precrq.ph.w    %[q1],      %[q0],      %[sec4]     \n\t"
+        "append         %[q2],      %[sec3],    16          \n\t"
+        "append         %[q0],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3),
+          [prim4] "=&r"(prim4), [q3] "+r"(q3), [q2] "+r"(q2), [q1] "+r"(q1),
+          [q0] "+r"(q0), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4)
+        :);
+
+    /* transpose p7, p6, p5, p4
+       original (when loaded from memory)
+       register      -8    -7   -6     -5
+         p4         p4_0  p4_1  p4_2  p4_3
+         p5         p5_0  p5_1  p5_2  p5_3
+         p6         p6_0  p6_1  p6_2  p6_3
+         p7         p7_0  p7_1  p7_2  p7_3
+
+       after transpose
+       register
+         p4         p7_3  p6_3  p5_3  p4_3
+         p5         p7_2  p6_2  p5_2  p4_2
+         p6         p7_1  p6_1  p5_1  p4_1
+         p7         p7_0  p6_0  p5_0  p4_0
+    */
+    __asm__ __volatile__(
+        "precrq.qb.ph   %[prim1],   %[p4],      %[p5]       \n\t"
+        "precr.qb.ph    %[prim2],   %[p4],      %[p5]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[p6],      %[p7]       \n\t"
+        "precr.qb.ph    %[prim4],   %[p6],      %[p7]       \n\t"
+
+        "precrq.qb.ph   %[p5],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[p7],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[p4],      %[p5],      %[sec3]     \n\t"
+        "precrq.ph.w    %[p6],      %[p7],      %[sec4]     \n\t"
+        "append         %[p5],      %[sec3],    16          \n\t"
+        "append         %[p7],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3),
+          [prim4] "=&r"(prim4), [p4] "+r"(p4), [p5] "+r"(p5), [p6] "+r"(p6),
+          [p7] "+r"(p7), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4)
+        :);
+
+    /* transpose q4, q5, q6, q7
+       original (when loaded from memory)
+       register      +5    +6    +7    +8
+         q7         q7_0  q7_1  q7_2  q7_3
+         q6         q6_0  q6_1  q6_2  q6_3
+         q5         q5_0  q5_1  q5_2  q5_3
+         q4         q4_0  q4_1  q4_2  q4_3
+
+       after transpose
+       register
+         q7         q4_3  q5_3  q26_3  q7_3
+         q6         q4_2  q5_2  q26_2  q7_2
+         q5         q4_1  q5_1  q26_1  q7_1
+         q4         q4_0  q5_0  q26_0  q7_0
+    */
+    __asm__ __volatile__(
+        "precrq.qb.ph   %[prim1],   %[q7],      %[q6]       \n\t"
+        "precr.qb.ph    %[prim2],   %[q7],      %[q6]       \n\t"
+        "precrq.qb.ph   %[prim3],   %[q5],      %[q4]       \n\t"
+        "precr.qb.ph    %[prim4],   %[q5],      %[q4]       \n\t"
+
+        "precrq.qb.ph   %[q6],      %[prim1],   %[prim2]    \n\t"
+        "precr.qb.ph    %[q4],      %[prim1],   %[prim2]    \n\t"
+        "precrq.qb.ph   %[sec3],    %[prim3],   %[prim4]    \n\t"
+        "precr.qb.ph    %[sec4],    %[prim3],   %[prim4]    \n\t"
+
+        "precrq.ph.w    %[q7],      %[q6],      %[sec3]     \n\t"
+        "precrq.ph.w    %[q5],      %[q4],      %[sec4]     \n\t"
+        "append         %[q6],      %[sec3],    16          \n\t"
+        "append         %[q4],      %[sec4],    16          \n\t"
+
+        : [prim1] "=&r"(prim1), [prim2] "=&r"(prim2), [prim3] "=&r"(prim3),
+          [prim4] "=&r"(prim4), [q7] "+r"(q7), [q6] "+r"(q6), [q5] "+r"(q5),
+          [q4] "+r"(q4), [sec3] "=&r"(sec3), [sec4] "=&r"(sec4)
+        :);
+
+    filter_hev_mask_flatmask4_dspr2(limit_vec, flimit_vec, thresh_vec, p1, p0,
+                                    p3, p2, q0, q1, q2, q3, &hev, &mask, &flat);
+
+    flatmask5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, &flat2);
+
+    /* f0 */
+    if (((flat2 == 0) && (flat == 0) && (mask != 0)) ||
+        ((flat2 != 0) && (flat == 0) && (mask != 0))) {
+      filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+      STORE_F0()
+    } else if ((flat2 == 0XFFFFFFFF) && (flat == 0xFFFFFFFF) &&
+               (mask == 0xFFFFFFFF)) {
+      /* f2 */
+      PACK_LEFT_0TO3()
+      PACK_LEFT_4TO7()
+      wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l, &p3_l, &p2_l, &p1_l,
+                          &p0_l, &q0_l, &q1_l, &q2_l, &q3_l, &q4_l, &q5_l,
+                          &q6_l, &q7_l);
+
+      PACK_RIGHT_0TO3()
+      PACK_RIGHT_4TO7()
+      wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r, &p3_r, &p2_r, &p1_r,
+                          &p0_r, &q0_r, &q1_r, &q2_r, &q3_r, &q4_r, &q5_r,
+                          &q6_r, &q7_r);
+
+      STORE_F2()
+    } else if ((flat2 == 0) && (flat == 0xFFFFFFFF) && (mask == 0xFFFFFFFF)) {
+      /* f1 */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l);
+
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r);
+
+      STORE_F1()
+    } else if ((flat2 == 0) && (flat != 0) && (mask != 0)) {
+      /* f0 + f1 */
+      filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      /* left 2 element operation */
+      PACK_LEFT_0TO3()
+      mbfilter_dspr2(&p3_l, &p2_l, &p1_l, &p0_l, &q0_l, &q1_l, &q2_l, &q3_l);
+
+      /* right 2 element operation */
+      PACK_RIGHT_0TO3()
+      mbfilter_dspr2(&p3_r, &p2_r, &p1_r, &p0_r, &q0_r, &q1_r, &q2_r, &q3_r);
+
+      if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb     %[p2_r],    -3(%[s4])    \n\t"
+            "sb     %[p1_r],    -2(%[s4])    \n\t"
+            "sb     %[p0_r],    -1(%[s4])    \n\t"
+            "sb     %[q0_r],      (%[s4])    \n\t"
+            "sb     %[q1_r],    +1(%[s4])    \n\t"
+            "sb     %[q2_r],    +2(%[s4])    \n\t"
+
+            :
+            : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r),
+              [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [s4] "r"(s4));
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb         %[p1_f0],  -2(%[s4])    \n\t"
+            "sb         %[p0_f0],  -1(%[s4])    \n\t"
+            "sb         %[q0_f0],    (%[s4])    \n\t"
+            "sb         %[q1_f0],  +1(%[s4])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s4] "r"(s4));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p2_r],    %[p2_r],    16      \n\t"
+          "srl      %[p1_r],    %[p1_r],    16      \n\t"
+          "srl      %[p0_r],    %[p0_r],    16      \n\t"
+          "srl      %[q0_r],    %[q0_r],    16      \n\t"
+          "srl      %[q1_r],    %[q1_r],    16      \n\t"
+          "srl      %[q2_r],    %[q2_r],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_r] "+r"(p2_r), [p1_r] "+r"(p1_r), [p0_r] "+r"(p0_r),
+            [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb     %[p2_r],    -3(%[s3])    \n\t"
+            "sb     %[p1_r],    -2(%[s3])    \n\t"
+            "sb     %[p0_r],    -1(%[s3])    \n\t"
+            "sb     %[q0_r],      (%[s3])    \n\t"
+            "sb     %[q1_r],    +1(%[s3])    \n\t"
+            "sb     %[q2_r],    +2(%[s3])    \n\t"
+
+            :
+            : [p2_r] "r"(p2_r), [p1_r] "r"(p1_r), [p0_r] "r"(p0_r),
+              [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [s3] "r"(s3));
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   -2(%[s3])    \n\t"
+            "sb     %[p0_f0],   -1(%[s3])    \n\t"
+            "sb     %[q0_f0],     (%[s3])    \n\t"
+            "sb     %[q1_f0],   +1(%[s3])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s3] "r"(s3));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p1_f0],   %[p1_f0],   8     \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8     \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8     \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8     \n\t"
+
+          : [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb       %[p2_l],    -3(%[s2])    \n\t"
+            "sb       %[p1_l],    -2(%[s2])    \n\t"
+            "sb       %[p0_l],    -1(%[s2])    \n\t"
+            "sb       %[q0_l],      (%[s2])    \n\t"
+            "sb       %[q1_l],    +1(%[s2])    \n\t"
+            "sb       %[q2_l],    +2(%[s2])    \n\t"
+
+            :
+            : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l),
+              [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [s2] "r"(s2));
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   -2(%[s2])    \n\t"
+            "sb     %[p0_f0],   -1(%[s2])    \n\t"
+            "sb     %[q0_f0],     (%[s2])    \n\t"
+            "sb     %[q1_f0],   +1(%[s2])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s2] "r"(s2));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p2_l],    %[p2_l],    16      \n\t"
+          "srl      %[p1_l],    %[p1_l],    16      \n\t"
+          "srl      %[p0_l],    %[p0_l],    16      \n\t"
+          "srl      %[q0_l],    %[q0_l],    16      \n\t"
+          "srl      %[q1_l],    %[q1_l],    16      \n\t"
+          "srl      %[q2_l],    %[q2_l],    16      \n\t"
+          "srl      %[p1_f0],   %[p1_f0],   8       \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8       \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8       \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8       \n\t"
+
+          : [p2_l] "+r"(p2_l), [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l),
+            [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb     %[p2_l],    -3(%[s1])    \n\t"
+            "sb     %[p1_l],    -2(%[s1])    \n\t"
+            "sb     %[p0_l],    -1(%[s1])    \n\t"
+            "sb     %[q0_l],      (%[s1])    \n\t"
+            "sb     %[q1_l],    +1(%[s1])    \n\t"
+            "sb     %[q2_l],    +2(%[s1])    \n\t"
+
+            :
+            : [p2_l] "r"(p2_l), [p1_l] "r"(p1_l), [p0_l] "r"(p0_l),
+              [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [s1] "r"(s1));
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   -2(%[s1])    \n\t"
+            "sb     %[p0_f0],   -1(%[s1])    \n\t"
+            "sb     %[q0_f0],     (%[s1])    \n\t"
+            "sb     %[q1_f0],   +1(%[s1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s1] "r"(s1));
+      }
+    } else if ((flat2 != 0) && (flat != 0) && (mask != 0)) {
+      /* f0+f1+f2 */
+      filter1_dspr2(mask, hev, p1, p0, q0, q1, &p1_f0, &p0_f0, &q0_f0, &q1_f0);
+
+      PACK_LEFT_0TO3()
+      mbfilter1_dspr2(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, &p2_l_f1,
+                      &p1_l_f1, &p0_l_f1, &q0_l_f1, &q1_l_f1, &q2_l_f1);
+
+      PACK_RIGHT_0TO3()
+      mbfilter1_dspr2(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, &p2_r_f1,
+                      &p1_r_f1, &p0_r_f1, &q0_r_f1, &q1_r_f1, &q2_r_f1);
+
+      PACK_LEFT_4TO7()
+      wide_mbfilter_dspr2(&p7_l, &p6_l, &p5_l, &p4_l, &p3_l, &p2_l, &p1_l,
+                          &p0_l, &q0_l, &q1_l, &q2_l, &q3_l, &q4_l, &q5_l,
+                          &q6_l, &q7_l);
+
+      PACK_RIGHT_4TO7()
+      wide_mbfilter_dspr2(&p7_r, &p6_r, &p5_r, &p4_r, &p3_r, &p2_r, &p1_r,
+                          &p0_r, &q0_r, &q1_r, &q2_r, &q3_r, &q4_r, &q5_r,
+                          &q6_r, &q7_r);
+
+      if (mask & flat & flat2 & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb     %[p6_r],    -7(%[s4])    \n\t"
+            "sb     %[p5_r],    -6(%[s4])    \n\t"
+            "sb     %[p4_r],    -5(%[s4])    \n\t"
+            "sb     %[p3_r],    -4(%[s4])    \n\t"
+            "sb     %[p2_r],    -3(%[s4])    \n\t"
+            "sb     %[p1_r],    -2(%[s4])    \n\t"
+            "sb     %[p0_r],    -1(%[s4])    \n\t"
+
+            :
+            : [p6_r] "r"(p6_r), [p5_r] "r"(p5_r), [p4_r] "r"(p4_r),
+              [p3_r] "r"(p3_r), [p2_r] "r"(p2_r), [p1_r] "r"(p1_r),
+              [p0_r] "r"(p0_r), [s4] "r"(s4));
+
+        __asm__ __volatile__(
+            "sb     %[q0_r],      (%[s4])    \n\t"
+            "sb     %[q1_r],    +1(%[s4])    \n\t"
+            "sb     %[q2_r],    +2(%[s4])    \n\t"
+            "sb     %[q3_r],    +3(%[s4])    \n\t"
+            "sb     %[q4_r],    +4(%[s4])    \n\t"
+            "sb     %[q5_r],    +5(%[s4])    \n\t"
+            "sb     %[q6_r],    +6(%[s4])    \n\t"
+
+            :
+            : [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [q3_r] "r"(q3_r), [q4_r] "r"(q4_r), [q5_r] "r"(q5_r),
+              [q6_r] "r"(q6_r), [s4] "r"(s4));
+      } else if (mask & flat & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb     %[p2_r_f1],     -3(%[s4])    \n\t"
+            "sb     %[p1_r_f1],     -2(%[s4])    \n\t"
+            "sb     %[p0_r_f1],     -1(%[s4])    \n\t"
+            "sb     %[q0_r_f1],       (%[s4])    \n\t"
+            "sb     %[q1_r_f1],     +1(%[s4])    \n\t"
+            "sb     %[q2_r_f1],     +2(%[s4])    \n\t"
+
+            :
+            : [p2_r_f1] "r"(p2_r_f1), [p1_r_f1] "r"(p1_r_f1),
+              [p0_r_f1] "r"(p0_r_f1), [q0_r_f1] "r"(q0_r_f1),
+              [q1_r_f1] "r"(q1_r_f1), [q2_r_f1] "r"(q2_r_f1), [s4] "r"(s4));
+      } else if (mask & 0x000000FF) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   -2(%[s4])    \n\t"
+            "sb     %[p0_f0],   -1(%[s4])    \n\t"
+            "sb     %[q0_f0],     (%[s4])    \n\t"
+            "sb     %[q1_f0],   +1(%[s4])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s4] "r"(s4));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p6_r],        %[p6_r],        16     \n\t"
+          "srl      %[p5_r],        %[p5_r],        16     \n\t"
+          "srl      %[p4_r],        %[p4_r],        16     \n\t"
+          "srl      %[p3_r],        %[p3_r],        16     \n\t"
+          "srl      %[p2_r],        %[p2_r],        16     \n\t"
+          "srl      %[p1_r],        %[p1_r],        16     \n\t"
+          "srl      %[p0_r],        %[p0_r],        16     \n\t"
+          "srl      %[q0_r],        %[q0_r],        16     \n\t"
+          "srl      %[q1_r],        %[q1_r],        16     \n\t"
+          "srl      %[q2_r],        %[q2_r],        16     \n\t"
+          "srl      %[q3_r],        %[q3_r],        16     \n\t"
+          "srl      %[q4_r],        %[q4_r],        16     \n\t"
+          "srl      %[q5_r],        %[q5_r],        16     \n\t"
+          "srl      %[q6_r],        %[q6_r],        16     \n\t"
+
+          : [q0_r] "+r"(q0_r), [q1_r] "+r"(q1_r), [q2_r] "+r"(q2_r),
+            [q3_r] "+r"(q3_r), [q4_r] "+r"(q4_r), [q5_r] "+r"(q5_r),
+            [q6_r] "+r"(q6_r), [p6_r] "+r"(p6_r), [p5_r] "+r"(p5_r),
+            [p4_r] "+r"(p4_r), [p3_r] "+r"(p3_r), [p2_r] "+r"(p2_r),
+            [p1_r] "+r"(p1_r), [p0_r] "+r"(p0_r)
+          :);
+
+      __asm__ __volatile__(
+          "srl      %[p2_r_f1],     %[p2_r_f1],     16      \n\t"
+          "srl      %[p1_r_f1],     %[p1_r_f1],     16      \n\t"
+          "srl      %[p0_r_f1],     %[p0_r_f1],     16      \n\t"
+          "srl      %[q0_r_f1],     %[q0_r_f1],     16      \n\t"
+          "srl      %[q1_r_f1],     %[q1_r_f1],     16      \n\t"
+          "srl      %[q2_r_f1],     %[q2_r_f1],     16      \n\t"
+          "srl      %[p1_f0],       %[p1_f0],       8       \n\t"
+          "srl      %[p0_f0],       %[p0_f0],       8       \n\t"
+          "srl      %[q0_f0],       %[q0_f0],       8       \n\t"
+          "srl      %[q1_f0],       %[q1_f0],       8       \n\t"
+
+          : [p2_r_f1] "+r"(p2_r_f1), [p1_r_f1] "+r"(p1_r_f1),
+            [p0_r_f1] "+r"(p0_r_f1), [q0_r_f1] "+r"(q0_r_f1),
+            [q1_r_f1] "+r"(q1_r_f1), [q2_r_f1] "+r"(q2_r_f1),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & flat2 & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb     %[p6_r],    -7(%[s3])    \n\t"
+            "sb     %[p5_r],    -6(%[s3])    \n\t"
+            "sb     %[p4_r],    -5(%[s3])    \n\t"
+            "sb     %[p3_r],    -4(%[s3])    \n\t"
+            "sb     %[p2_r],    -3(%[s3])    \n\t"
+            "sb     %[p1_r],    -2(%[s3])    \n\t"
+            "sb     %[p0_r],    -1(%[s3])    \n\t"
+
+            :
+            : [p6_r] "r"(p6_r), [p5_r] "r"(p5_r), [p4_r] "r"(p4_r),
+              [p3_r] "r"(p3_r), [p2_r] "r"(p2_r), [p1_r] "r"(p1_r),
+              [p0_r] "r"(p0_r), [s3] "r"(s3));
+
+        __asm__ __volatile__(
+            "sb     %[q0_r],      (%[s3])    \n\t"
+            "sb     %[q1_r],    +1(%[s3])    \n\t"
+            "sb     %[q2_r],    +2(%[s3])    \n\t"
+            "sb     %[q3_r],    +3(%[s3])    \n\t"
+            "sb     %[q4_r],    +4(%[s3])    \n\t"
+            "sb     %[q5_r],    +5(%[s3])    \n\t"
+            "sb     %[q6_r],    +6(%[s3])    \n\t"
+
+            :
+            : [q0_r] "r"(q0_r), [q1_r] "r"(q1_r), [q2_r] "r"(q2_r),
+              [q3_r] "r"(q3_r), [q4_r] "r"(q4_r), [q5_r] "r"(q5_r),
+              [q6_r] "r"(q6_r), [s3] "r"(s3));
+      } else if (mask & flat & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb     %[p2_r_f1],     -3(%[s3])    \n\t"
+            "sb     %[p1_r_f1],     -2(%[s3])    \n\t"
+            "sb     %[p0_r_f1],     -1(%[s3])    \n\t"
+            "sb     %[q0_r_f1],       (%[s3])    \n\t"
+            "sb     %[q1_r_f1],     +1(%[s3])    \n\t"
+            "sb     %[q2_r_f1],     +2(%[s3])    \n\t"
+
+            :
+            : [p2_r_f1] "r"(p2_r_f1), [p1_r_f1] "r"(p1_r_f1),
+              [p0_r_f1] "r"(p0_r_f1), [q0_r_f1] "r"(q0_r_f1),
+              [q1_r_f1] "r"(q1_r_f1), [q2_r_f1] "r"(q2_r_f1), [s3] "r"(s3));
+      } else if (mask & 0x0000FF00) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   -2(%[s3])    \n\t"
+            "sb     %[p0_f0],   -1(%[s3])    \n\t"
+            "sb     %[q0_f0],     (%[s3])    \n\t"
+            "sb     %[q1_f0],   +1(%[s3])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s3] "r"(s3));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p1_f0],   %[p1_f0],   8     \n\t"
+          "srl      %[p0_f0],   %[p0_f0],   8     \n\t"
+          "srl      %[q0_f0],   %[q0_f0],   8     \n\t"
+          "srl      %[q1_f0],   %[q1_f0],   8     \n\t"
+
+          : [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & flat2 & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb     %[p6_l],    -7(%[s2])    \n\t"
+            "sb     %[p5_l],    -6(%[s2])    \n\t"
+            "sb     %[p4_l],    -5(%[s2])    \n\t"
+            "sb     %[p3_l],    -4(%[s2])    \n\t"
+            "sb     %[p2_l],    -3(%[s2])    \n\t"
+            "sb     %[p1_l],    -2(%[s2])    \n\t"
+            "sb     %[p0_l],    -1(%[s2])    \n\t"
+
+            :
+            : [p6_l] "r"(p6_l), [p5_l] "r"(p5_l), [p4_l] "r"(p4_l),
+              [p3_l] "r"(p3_l), [p2_l] "r"(p2_l), [p1_l] "r"(p1_l),
+              [p0_l] "r"(p0_l), [s2] "r"(s2));
+
+        __asm__ __volatile__(
+            "sb     %[q0_l],      (%[s2])    \n\t"
+            "sb     %[q1_l],    +1(%[s2])    \n\t"
+            "sb     %[q2_l],    +2(%[s2])    \n\t"
+            "sb     %[q3_l],    +3(%[s2])    \n\t"
+            "sb     %[q4_l],    +4(%[s2])    \n\t"
+            "sb     %[q5_l],    +5(%[s2])    \n\t"
+            "sb     %[q6_l],    +6(%[s2])    \n\t"
+
+            :
+            : [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [q3_l] "r"(q3_l), [q4_l] "r"(q4_l), [q5_l] "r"(q5_l),
+              [q6_l] "r"(q6_l), [s2] "r"(s2));
+      } else if (mask & flat & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb     %[p2_l_f1],     -3(%[s2])    \n\t"
+            "sb     %[p1_l_f1],     -2(%[s2])    \n\t"
+            "sb     %[p0_l_f1],     -1(%[s2])    \n\t"
+            "sb     %[q0_l_f1],       (%[s2])    \n\t"
+            "sb     %[q1_l_f1],     +1(%[s2])    \n\t"
+            "sb     %[q2_l_f1],     +2(%[s2])    \n\t"
+
+            :
+            : [p2_l_f1] "r"(p2_l_f1), [p1_l_f1] "r"(p1_l_f1),
+              [p0_l_f1] "r"(p0_l_f1), [q0_l_f1] "r"(q0_l_f1),
+              [q1_l_f1] "r"(q1_l_f1), [q2_l_f1] "r"(q2_l_f1), [s2] "r"(s2));
+      } else if (mask & 0x00FF0000) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   -2(%[s2])    \n\t"
+            "sb     %[p0_f0],   -1(%[s2])    \n\t"
+            "sb     %[q0_f0],     (%[s2])    \n\t"
+            "sb     %[q1_f0],   +1(%[s2])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s2] "r"(s2));
+      }
+
+      __asm__ __volatile__(
+          "srl      %[p6_l],        %[p6_l],        16     \n\t"
+          "srl      %[p5_l],        %[p5_l],        16     \n\t"
+          "srl      %[p4_l],        %[p4_l],        16     \n\t"
+          "srl      %[p3_l],        %[p3_l],        16     \n\t"
+          "srl      %[p2_l],        %[p2_l],        16     \n\t"
+          "srl      %[p1_l],        %[p1_l],        16     \n\t"
+          "srl      %[p0_l],        %[p0_l],        16     \n\t"
+          "srl      %[q0_l],        %[q0_l],        16     \n\t"
+          "srl      %[q1_l],        %[q1_l],        16     \n\t"
+          "srl      %[q2_l],        %[q2_l],        16     \n\t"
+          "srl      %[q3_l],        %[q3_l],        16     \n\t"
+          "srl      %[q4_l],        %[q4_l],        16     \n\t"
+          "srl      %[q5_l],        %[q5_l],        16     \n\t"
+          "srl      %[q6_l],        %[q6_l],        16     \n\t"
+
+          : [q0_l] "+r"(q0_l), [q1_l] "+r"(q1_l), [q2_l] "+r"(q2_l),
+            [q3_l] "+r"(q3_l), [q4_l] "+r"(q4_l), [q5_l] "+r"(q5_l),
+            [q6_l] "+r"(q6_l), [p6_l] "+r"(p6_l), [p5_l] "+r"(p5_l),
+            [p4_l] "+r"(p4_l), [p3_l] "+r"(p3_l), [p2_l] "+r"(p2_l),
+            [p1_l] "+r"(p1_l), [p0_l] "+r"(p0_l)
+          :);
+
+      __asm__ __volatile__(
+          "srl      %[p2_l_f1],     %[p2_l_f1],     16      \n\t"
+          "srl      %[p1_l_f1],     %[p1_l_f1],     16      \n\t"
+          "srl      %[p0_l_f1],     %[p0_l_f1],     16      \n\t"
+          "srl      %[q0_l_f1],     %[q0_l_f1],     16      \n\t"
+          "srl      %[q1_l_f1],     %[q1_l_f1],     16      \n\t"
+          "srl      %[q2_l_f1],     %[q2_l_f1],     16      \n\t"
+          "srl      %[p1_f0],       %[p1_f0],       8       \n\t"
+          "srl      %[p0_f0],       %[p0_f0],       8       \n\t"
+          "srl      %[q0_f0],       %[q0_f0],       8       \n\t"
+          "srl      %[q1_f0],       %[q1_f0],       8       \n\t"
+
+          : [p2_l_f1] "+r"(p2_l_f1), [p1_l_f1] "+r"(p1_l_f1),
+            [p0_l_f1] "+r"(p0_l_f1), [q0_l_f1] "+r"(q0_l_f1),
+            [q1_l_f1] "+r"(q1_l_f1), [q2_l_f1] "+r"(q2_l_f1),
+            [p1_f0] "+r"(p1_f0), [p0_f0] "+r"(p0_f0), [q0_f0] "+r"(q0_f0),
+            [q1_f0] "+r"(q1_f0)
+          :);
+
+      if (mask & flat & flat2 & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb     %[p6_l],    -7(%[s1])    \n\t"
+            "sb     %[p5_l],    -6(%[s1])    \n\t"
+            "sb     %[p4_l],    -5(%[s1])    \n\t"
+            "sb     %[p3_l],    -4(%[s1])    \n\t"
+            "sb     %[p2_l],    -3(%[s1])    \n\t"
+            "sb     %[p1_l],    -2(%[s1])    \n\t"
+            "sb     %[p0_l],    -1(%[s1])    \n\t"
+
+            :
+            : [p6_l] "r"(p6_l), [p5_l] "r"(p5_l), [p4_l] "r"(p4_l),
+              [p3_l] "r"(p3_l), [p2_l] "r"(p2_l), [p1_l] "r"(p1_l),
+              [p0_l] "r"(p0_l), [s1] "r"(s1));
+
+        __asm__ __volatile__(
+            "sb     %[q0_l],     (%[s1])    \n\t"
+            "sb     %[q1_l],    1(%[s1])    \n\t"
+            "sb     %[q2_l],    2(%[s1])    \n\t"
+            "sb     %[q3_l],    3(%[s1])    \n\t"
+            "sb     %[q4_l],    4(%[s1])    \n\t"
+            "sb     %[q5_l],    5(%[s1])    \n\t"
+            "sb     %[q6_l],    6(%[s1])    \n\t"
+
+            :
+            : [q0_l] "r"(q0_l), [q1_l] "r"(q1_l), [q2_l] "r"(q2_l),
+              [q3_l] "r"(q3_l), [q4_l] "r"(q4_l), [q5_l] "r"(q5_l),
+              [q6_l] "r"(q6_l), [s1] "r"(s1));
+      } else if (mask & flat & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb     %[p2_l_f1],     -3(%[s1])    \n\t"
+            "sb     %[p1_l_f1],     -2(%[s1])    \n\t"
+            "sb     %[p0_l_f1],     -1(%[s1])    \n\t"
+            "sb     %[q0_l_f1],       (%[s1])    \n\t"
+            "sb     %[q1_l_f1],     +1(%[s1])    \n\t"
+            "sb     %[q2_l_f1],     +2(%[s1])    \n\t"
+
+            :
+            : [p2_l_f1] "r"(p2_l_f1), [p1_l_f1] "r"(p1_l_f1),
+              [p0_l_f1] "r"(p0_l_f1), [q0_l_f1] "r"(q0_l_f1),
+              [q1_l_f1] "r"(q1_l_f1), [q2_l_f1] "r"(q2_l_f1), [s1] "r"(s1));
+      } else if (mask & 0xFF000000) {
+        __asm__ __volatile__(
+            "sb     %[p1_f0],   -2(%[s1])    \n\t"
+            "sb     %[p0_f0],   -1(%[s1])    \n\t"
+            "sb     %[q0_f0],     (%[s1])    \n\t"
+            "sb     %[q1_f0],   +1(%[s1])    \n\t"
+
+            :
+            : [p1_f0] "r"(p1_f0), [p0_f0] "r"(p0_f0), [q0_f0] "r"(q0_f0),
+              [q1_f0] "r"(q1_f0), [s1] "r"(s1));
+      }
+    }
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/aom_dsp/mips/loopfilter_msa.h b/third_party/aom/aom_dsp/mips/loopfilter_msa.h
new file mode 100644
index 0000000000..450594262c
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/loopfilter_msa.h
@@ -0,0 +1,251 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_LOOPFILTER_MSA_H_
+#define AOM_DSP_LOOPFILTER_MSA_H_
+
+#include "aom_dsp/mips/macros_msa.h"
+
+#define AOM_LPF_FILTER4_8W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in, \
+                           p1_out, p0_out, q0_out, q1_out)              \
+  {                                                                     \
+    v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign;                 \
+    v16i8 filt, filt1, filt2, cnst4b, cnst3b;                           \
+    v8i16 q0_sub_p0_r, filt_r, cnst3h;                                  \
+                                                                        \
+    p1_m = (v16i8)__msa_xori_b(p1_in, 0x80);                            \
+    p0_m = (v16i8)__msa_xori_b(p0_in, 0x80);                            \
+    q0_m = (v16i8)__msa_xori_b(q0_in, 0x80);                            \
+    q1_m = (v16i8)__msa_xori_b(q1_in, 0x80);                            \
+                                                                        \
+    filt = __msa_subs_s_b(p1_m, q1_m);                                  \
+    filt = filt & (v16i8)hev_in;                                        \
+    q0_sub_p0 = q0_m - p0_m;                                            \
+    filt_sign = __msa_clti_s_b(filt, 0);                                \
+                                                                        \
+    cnst3h = __msa_ldi_h(3);                                            \
+    q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0);            \
+    q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h);    \
+    filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt);                      \
+    filt_r += q0_sub_p0_r;                                              \
+    filt_r = __msa_sat_s_h(filt_r, 7);                                  \
+                                                                        \
+    /* combine left and right part */                                   \
+    filt = __msa_pckev_b((v16i8)filt_r, (v16i8)filt_r);                 \
+                                                                        \
+    filt = filt & (v16i8)mask_in;                                       \
+    cnst4b = __msa_ldi_b(4);                                            \
+    filt1 = __msa_adds_s_b(filt, cnst4b);                               \
+    filt1 >>= 3;                                                        \
+                                                                        \
+    cnst3b = __msa_ldi_b(3);                                            \
+    filt2 = __msa_adds_s_b(filt, cnst3b);                               \
+    filt2 >>= 3;                                                        \
+                                                                        \
+    q0_m = __msa_subs_s_b(q0_m, filt1);                                 \
+    q0_out = __msa_xori_b((v16u8)q0_m, 0x80);                           \
+    p0_m = __msa_adds_s_b(p0_m, filt2);                                 \
+    p0_out = __msa_xori_b((v16u8)p0_m, 0x80);                           \
+                                                                        \
+    filt = __msa_srari_b(filt1, 1);                                     \
+    hev_in = __msa_xori_b((v16u8)hev_in, 0xff);                         \
+    filt = filt & (v16i8)hev_in;                                        \
+                                                                        \
+    q1_m = __msa_subs_s_b(q1_m, filt);                                  \
+    q1_out = __msa_xori_b((v16u8)q1_m, 0x80);                           \
+    p1_m = __msa_adds_s_b(p1_m, filt);                                  \
+    p1_out = __msa_xori_b((v16u8)p1_m, 0x80);                           \
+  }
+
+#define AOM_LPF_FILTER4_4W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in, \
+                           p1_out, p0_out, q0_out, q1_out)              \
+  {                                                                     \
+    v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign;                 \
+    v16i8 filt, filt1, filt2, cnst4b, cnst3b;                           \
+    v8i16 q0_sub_p0_r, q0_sub_p0_l, filt_l, filt_r, cnst3h;             \
+                                                                        \
+    p1_m = (v16i8)__msa_xori_b(p1_in, 0x80);                            \
+    p0_m = (v16i8)__msa_xori_b(p0_in, 0x80);                            \
+    q0_m = (v16i8)__msa_xori_b(q0_in, 0x80);                            \
+    q1_m = (v16i8)__msa_xori_b(q1_in, 0x80);                            \
+                                                                        \
+    filt = __msa_subs_s_b(p1_m, q1_m);                                  \
+                                                                        \
+    filt = filt & (v16i8)hev_in;                                        \
+                                                                        \
+    q0_sub_p0 = q0_m - p0_m;                                            \
+    filt_sign = __msa_clti_s_b(filt, 0);                                \
+                                                                        \
+    cnst3h = __msa_ldi_h(3);                                            \
+    q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0);            \
+    q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h);    \
+    filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt);                      \
+    filt_r += q0_sub_p0_r;                                              \
+    filt_r = __msa_sat_s_h(filt_r, 7);                                  \
+                                                                        \
+    q0_sub_p0_l = (v8i16)__msa_ilvl_b(q0_sub_p0, q0_sub_p0);            \
+    q0_sub_p0_l = __msa_dotp_s_h((v16i8)q0_sub_p0_l, (v16i8)cnst3h);    \
+    filt_l = (v8i16)__msa_ilvl_b(filt_sign, filt);                      \
+    filt_l += q0_sub_p0_l;                                              \
+    filt_l = __msa_sat_s_h(filt_l, 7);                                  \
+                                                                        \
+    filt = __msa_pckev_b((v16i8)filt_l, (v16i8)filt_r);                 \
+    filt = filt & (v16i8)mask_in;                                       \
+                                                                        \
+    cnst4b = __msa_ldi_b(4);                                            \
+    filt1 = __msa_adds_s_b(filt, cnst4b);                               \
+    filt1 >>= 3;                                                        \
+                                                                        \
+    cnst3b = __msa_ldi_b(3);                                            \
+    filt2 = __msa_adds_s_b(filt, cnst3b);                               \
+    filt2 >>= 3;                                                        \
+                                                                        \
+    q0_m = __msa_subs_s_b(q0_m, filt1);                                 \
+    q0_out = __msa_xori_b((v16u8)q0_m, 0x80);                           \
+    p0_m = __msa_adds_s_b(p0_m, filt2);                                 \
+    p0_out = __msa_xori_b((v16u8)p0_m, 0x80);                           \
+                                                                        \
+    filt = __msa_srari_b(filt1, 1);                                     \
+    hev_in = __msa_xori_b((v16u8)hev_in, 0xff);                         \
+    filt = filt & (v16i8)hev_in;                                        \
+                                                                        \
+    q1_m = __msa_subs_s_b(q1_m, filt);                                  \
+    q1_out = __msa_xori_b((v16u8)q1_m, 0x80);                           \
+    p1_m = __msa_adds_s_b(p1_m, filt);                                  \
+    p1_out = __msa_xori_b((v16u8)p1_m, 0x80);                           \
+  }
+
+#define AOM_FLAT4(p3_in, p2_in, p0_in, q0_in, q2_in, q3_in, flat_out)    \
+  {                                                                      \
+    v16u8 tmp_flat4, p2_a_sub_p0, q2_a_sub_q0, p3_a_sub_p0, q3_a_sub_q0; \
+    v16u8 zero_in = { 0 };                                               \
+                                                                         \
+    tmp_flat4 = __msa_ori_b(zero_in, 1);                                 \
+    p2_a_sub_p0 = __msa_asub_u_b(p2_in, p0_in);                          \
+    q2_a_sub_q0 = __msa_asub_u_b(q2_in, q0_in);                          \
+    p3_a_sub_p0 = __msa_asub_u_b(p3_in, p0_in);                          \
+    q3_a_sub_q0 = __msa_asub_u_b(q3_in, q0_in);                          \
+                                                                         \
+    p2_a_sub_p0 = __msa_max_u_b(p2_a_sub_p0, q2_a_sub_q0);               \
+    flat_out = __msa_max_u_b(p2_a_sub_p0, flat_out);                     \
+    p3_a_sub_p0 = __msa_max_u_b(p3_a_sub_p0, q3_a_sub_q0);               \
+    flat_out = __msa_max_u_b(p3_a_sub_p0, flat_out);                     \
+                                                                         \
+    flat_out = (tmp_flat4 < (v16u8)flat_out);                            \
+    flat_out = __msa_xori_b(flat_out, 0xff);                             \
+    flat_out = flat_out & (mask);                                        \
+  }
+
+#define AOM_FLAT5(p7_in, p6_in, p5_in, p4_in, p0_in, q0_in, q4_in, q5_in, \
+                  q6_in, q7_in, flat_in, flat2_out)                       \
+  {                                                                       \
+    v16u8 tmp_flat5, zero_in = { 0 };                                     \
+    v16u8 p4_a_sub_p0, q4_a_sub_q0, p5_a_sub_p0, q5_a_sub_q0;             \
+    v16u8 p6_a_sub_p0, q6_a_sub_q0, p7_a_sub_p0, q7_a_sub_q0;             \
+                                                                          \
+    tmp_flat5 = __msa_ori_b(zero_in, 1);                                  \
+    p4_a_sub_p0 = __msa_asub_u_b(p4_in, p0_in);                           \
+    q4_a_sub_q0 = __msa_asub_u_b(q4_in, q0_in);                           \
+    p5_a_sub_p0 = __msa_asub_u_b(p5_in, p0_in);                           \
+    q5_a_sub_q0 = __msa_asub_u_b(q5_in, q0_in);                           \
+    p6_a_sub_p0 = __msa_asub_u_b(p6_in, p0_in);                           \
+    q6_a_sub_q0 = __msa_asub_u_b(q6_in, q0_in);                           \
+    p7_a_sub_p0 = __msa_asub_u_b(p7_in, p0_in);                           \
+    q7_a_sub_q0 = __msa_asub_u_b(q7_in, q0_in);                           \
+                                                                          \
+    p4_a_sub_p0 = __msa_max_u_b(p4_a_sub_p0, q4_a_sub_q0);                \
+    flat2_out = __msa_max_u_b(p5_a_sub_p0, q5_a_sub_q0);                  \
+    flat2_out = __msa_max_u_b(p4_a_sub_p0, flat2_out);                    \
+    p6_a_sub_p0 = __msa_max_u_b(p6_a_sub_p0, q6_a_sub_q0);                \
+    flat2_out = __msa_max_u_b(p6_a_sub_p0, flat2_out);                    \
+    p7_a_sub_p0 = __msa_max_u_b(p7_a_sub_p0, q7_a_sub_q0);                \
+    flat2_out = __msa_max_u_b(p7_a_sub_p0, flat2_out);                    \
+                                                                          \
+    flat2_out = (tmp_flat5 < (v16u8)flat2_out);                           \
+    flat2_out = __msa_xori_b(flat2_out, 0xff);                            \
+    flat2_out = flat2_out & flat_in;                                      \
+  }
+
+#define AOM_FILTER8(p3_in, p2_in, p1_in, p0_in, q0_in, q1_in, q2_in, q3_in, \
+                    p2_filt8_out, p1_filt8_out, p0_filt8_out, q0_filt8_out, \
+                    q1_filt8_out, q2_filt8_out)                             \
+  {                                                                         \
+    v8u16 tmp_filt8_0, tmp_filt8_1, tmp_filt8_2;                            \
+                                                                            \
+    tmp_filt8_2 = p2_in + p1_in + p0_in;                                    \
+    tmp_filt8_0 = p3_in << 1;                                               \
+                                                                            \
+    tmp_filt8_0 = tmp_filt8_0 + tmp_filt8_2 + q0_in;                        \
+    tmp_filt8_1 = tmp_filt8_0 + p3_in + p2_in;                              \
+    p2_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_1, 3);             \
+                                                                            \
+    tmp_filt8_1 = tmp_filt8_0 + p1_in + q1_in;                              \
+    p1_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_1, 3);             \
+                                                                            \
+    tmp_filt8_1 = q2_in + q1_in + q0_in;                                    \
+    tmp_filt8_2 = tmp_filt8_2 + tmp_filt8_1;                                \
+    tmp_filt8_0 = tmp_filt8_2 + (p0_in);                                    \
+    tmp_filt8_0 = tmp_filt8_0 + (p3_in);                                    \
+    p0_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_0, 3);             \
+                                                                            \
+    tmp_filt8_0 = q2_in + q3_in;                                            \
+    tmp_filt8_0 = p0_in + tmp_filt8_1 + tmp_filt8_0;                        \
+    tmp_filt8_1 = q3_in + q3_in;                                            \
+    tmp_filt8_1 = tmp_filt8_1 + tmp_filt8_0;                                \
+    q2_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_1, 3);             \
+                                                                            \
+    tmp_filt8_0 = tmp_filt8_2 + q3_in;                                      \
+    tmp_filt8_1 = tmp_filt8_0 + q0_in;                                      \
+    q0_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_1, 3);             \
+                                                                            \
+    tmp_filt8_1 = tmp_filt8_0 - p2_in;                                      \
+    tmp_filt8_0 = q1_in + q3_in;                                            \
+    tmp_filt8_1 = tmp_filt8_0 + tmp_filt8_1;                                \
+    q1_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp_filt8_1, 3);             \
+  }
+
+#define LPF_MASK_HEV(p3_in, p2_in, p1_in, p0_in, q0_in, q1_in, q2_in, q3_in, \
+                     limit_in, b_limit_in, thresh_in, hev_out, mask_out,     \
+                     flat_out)                                               \
+  {                                                                          \
+    v16u8 p3_asub_p2_m, p2_asub_p1_m, p1_asub_p0_m, q1_asub_q0_m;            \
+    v16u8 p1_asub_q1_m, p0_asub_q0_m, q3_asub_q2_m, q2_asub_q1_m;            \
+                                                                             \
+    /* absolute subtraction of pixel values */                               \
+    p3_asub_p2_m = __msa_asub_u_b(p3_in, p2_in);                             \
+    p2_asub_p1_m = __msa_asub_u_b(p2_in, p1_in);                             \
+    p1_asub_p0_m = __msa_asub_u_b(p1_in, p0_in);                             \
+    q1_asub_q0_m = __msa_asub_u_b(q1_in, q0_in);                             \
+    q2_asub_q1_m = __msa_asub_u_b(q2_in, q1_in);                             \
+    q3_asub_q2_m = __msa_asub_u_b(q3_in, q2_in);                             \
+    p0_asub_q0_m = __msa_asub_u_b(p0_in, q0_in);                             \
+    p1_asub_q1_m = __msa_asub_u_b(p1_in, q1_in);                             \
+                                                                             \
+    /* calculation of hev */                                                 \
+    flat_out = __msa_max_u_b(p1_asub_p0_m, q1_asub_q0_m);                    \
+    hev_out = thresh_in < (v16u8)flat_out;                                   \
+                                                                             \
+    /* calculation of mask */                                                \
+    p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p0_asub_q0_m);               \
+    p1_asub_q1_m >>= 1;                                                      \
+    p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p1_asub_q1_m);               \
+                                                                             \
+    mask_out = b_limit_in < p0_asub_q0_m;                                    \
+    mask_out = __msa_max_u_b(flat_out, mask_out);                            \
+    p3_asub_p2_m = __msa_max_u_b(p3_asub_p2_m, p2_asub_p1_m);                \
+    mask_out = __msa_max_u_b(p3_asub_p2_m, mask_out);                        \
+    q2_asub_q1_m = __msa_max_u_b(q2_asub_q1_m, q3_asub_q2_m);                \
+    mask_out = __msa_max_u_b(q2_asub_q1_m, mask_out);                        \
+                                                                             \
+    mask_out = limit_in < (v16u8)mask_out;                                   \
+    mask_out = __msa_xori_b(mask_out, 0xff);                                 \
+  }
+#endif /* AOM_DSP_LOOPFILTER_MSA_H_ */
diff --git a/third_party/aom/aom_dsp/mips/macros_msa.h b/third_party/aom/aom_dsp/mips/macros_msa.h
new file mode 100644
index 0000000000..48fbcfd47f
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/macros_msa.h
@@ -0,0 +1,2057 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_MIPS_MACROS_MSA_H_
+#define AOM_DSP_MIPS_MACROS_MSA_H_
+
+#include <msa.h>
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+#define LD_B(RTYPE, psrc) *((const RTYPE *)(psrc))
+#define LD_UB(...) LD_B(v16u8, __VA_ARGS__)
+#define LD_SB(...) LD_B(v16i8, __VA_ARGS__)
+
+#define LD_H(RTYPE, psrc) *((const RTYPE *)(psrc))
+#define LD_UH(...) LD_H(v8u16, __VA_ARGS__)
+#define LD_SH(...) LD_H(v8i16, __VA_ARGS__)
+
+#define LD_W(RTYPE, psrc) *((const RTYPE *)(psrc))
+#define LD_SW(...) LD_W(v4i32, __VA_ARGS__)
+
+#define ST_B(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_UB(...) ST_B(v16u8, __VA_ARGS__)
+#define ST_SB(...) ST_B(v16i8, __VA_ARGS__)
+
+#define ST_H(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_SH(...) ST_H(v8i16, __VA_ARGS__)
+
+#define ST_W(RTYPE, in, pdst) *((RTYPE *)(pdst)) = (in)
+#define ST_SW(...) ST_W(v4i32, __VA_ARGS__)
+
+#if (__mips_isa_rev >= 6)
+#define LH(psrc)                                          \
+  ({                                                      \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);      \
+    uint16_t val_m;                                       \
+                                                          \
+    __asm__ __volatile__("lh  %[val_m],  %[psrc_m]  \n\t" \
+                                                          \
+                         : [val_m] "=r"(val_m)            \
+                         : [psrc_m] "m"(*psrc_m));        \
+                                                          \
+    val_m;                                                \
+  })
+
+#define LW(psrc)                                          \
+  ({                                                      \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);      \
+    uint32_t val_m;                                       \
+                                                          \
+    __asm__ __volatile__("lw  %[val_m],  %[psrc_m]  \n\t" \
+                                                          \
+                         : [val_m] "=r"(val_m)            \
+                         : [psrc_m] "m"(*psrc_m));        \
+                                                          \
+    val_m;                                                \
+  })
+
+#if (__mips == 64)
+#define LD(psrc)                                          \
+  ({                                                      \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);      \
+    uint64_t val_m = 0;                                   \
+                                                          \
+    __asm__ __volatile__("ld  %[val_m],  %[psrc_m]  \n\t" \
+                                                          \
+                         : [val_m] "=r"(val_m)            \
+                         : [psrc_m] "m"(*psrc_m));        \
+                                                          \
+    val_m;                                                \
+  })
+#else  // !(__mips == 64)
+#define LD(psrc)                                            \
+  ({                                                        \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);        \
+    uint32_t val0_m, val1_m;                                \
+    uint64_t val_m = 0;                                     \
+                                                            \
+    val0_m = LW(psrc_m);                                    \
+    val1_m = LW(psrc_m + 4);                                \
+                                                            \
+    val_m = (uint64_t)(val1_m);                             \
+    val_m = (uint64_t)((val_m << 32) & 0xFFFFFFFF00000000); \
+    val_m = (uint64_t)(val_m | (uint64_t)val0_m);           \
+                                                            \
+    val_m;                                                  \
+  })
+#endif  // (__mips == 64)
+
+#define SH(val, pdst)                                     \
+  {                                                       \
+    uint8_t *pdst_m = (uint8_t *)(pdst);                  \
+    const uint16_t val_m = (val);                         \
+                                                          \
+    __asm__ __volatile__("sh  %[val_m],  %[pdst_m]  \n\t" \
+                                                          \
+                         : [pdst_m] "=m"(*pdst_m)         \
+                         : [val_m] "r"(val_m));           \
+  }
+
+#define SW(val, pdst)                                     \
+  {                                                       \
+    uint8_t *pdst_m = (uint8_t *)(pdst);                  \
+    const uint32_t val_m = (val);                         \
+                                                          \
+    __asm__ __volatile__("sw  %[val_m],  %[pdst_m]  \n\t" \
+                                                          \
+                         : [pdst_m] "=m"(*pdst_m)         \
+                         : [val_m] "r"(val_m));           \
+  }
+
+#define SD(val, pdst)                                     \
+  {                                                       \
+    uint8_t *pdst_m = (uint8_t *)(pdst);                  \
+    const uint64_t val_m = (val);                         \
+                                                          \
+    __asm__ __volatile__("sd  %[val_m],  %[pdst_m]  \n\t" \
+                                                          \
+                         : [pdst_m] "=m"(*pdst_m)         \
+                         : [val_m] "r"(val_m));           \
+  }
+#else  // !(__mips_isa_rev >= 6)
+#define LH(psrc)                                           \
+  ({                                                       \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);       \
+    uint16_t val_m;                                        \
+                                                           \
+    __asm__ __volatile__("ulh  %[val_m],  %[psrc_m]  \n\t" \
+                                                           \
+                         : [val_m] "=r"(val_m)             \
+                         : [psrc_m] "m"(*psrc_m));         \
+                                                           \
+    val_m;                                                 \
+  })
+
+#define LW(psrc)                                           \
+  ({                                                       \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);       \
+    uint32_t val_m;                                        \
+                                                           \
+    __asm__ __volatile__("ulw  %[val_m],  %[psrc_m]  \n\t" \
+                                                           \
+                         : [val_m] "=r"(val_m)             \
+                         : [psrc_m] "m"(*psrc_m));         \
+                                                           \
+    val_m;                                                 \
+  })
+
+#if (__mips == 64)
+#define LD(psrc)                                           \
+  ({                                                       \
+    const uint8_t *psrc_m = (const uint8_t *)(psrc);       \
+    uint64_t val_m = 0;                                    \
+                                                           \
+    __asm__ __volatile__("uld  %[val_m],  %[psrc_m]  \n\t" \
+                                                           \
+                         : [val_m] "=r"(val_m)             \
+                         : [psrc_m] "m"(*psrc_m));         \
+                                                           \
+    val_m;                                                 \
+  })
+#else  // !(__mips == 64)
+#define LD(psrc)                                                              \
+  ({                                                                          \
+    const uint8_t *psrc_m1 = (const uint8_t *)(psrc);                         \
+    uint32_t val0_m, val1_m;                                                  \
+    uint64_t val_m_combined = 0;                                              \
+                                                                              \
+    val0_m = LW(psrc_m1);                                                     \
+    val1_m = LW(psrc_m1 + 4);                                                 \
+                                                                              \
+    val_m_combined = (uint64_t)(val1_m);                                      \
+    val_m_combined = (uint64_t)((val_m_combined << 32) & 0xFFFFFFFF00000000); \
+    val_m_combined = (uint64_t)(val_m_combined | (uint64_t)val0_m);           \
+                                                                              \
+    val_m_combined;                                                           \
+  })
+#endif  // (__mips == 64)
+
+#define SH(val, pdst)                                      \
+  {                                                        \
+    uint8_t *pdst_m = (uint8_t *)(pdst);                   \
+    const uint16_t val_m = (val);                          \
+                                                           \
+    __asm__ __volatile__("ush  %[val_m],  %[pdst_m]  \n\t" \
+                                                           \
+                         : [pdst_m] "=m"(*pdst_m)          \
+                         : [val_m] "r"(val_m));            \
+  }
+
+#define SW(val, pdst)                                      \
+  {                                                        \
+    uint8_t *pdst_m = (uint8_t *)(pdst);                   \
+    const uint32_t val_m = (val);                          \
+                                                           \
+    __asm__ __volatile__("usw  %[val_m],  %[pdst_m]  \n\t" \
+                                                           \
+                         : [pdst_m] "=m"(*pdst_m)          \
+                         : [val_m] "r"(val_m));            \
+  }
+
+#define SD(val, pdst)                                        \
+  {                                                          \
+    uint8_t *pdst_m1 = (uint8_t *)(pdst);                    \
+    uint32_t val0_m, val1_m;                                 \
+                                                             \
+    val0_m = (uint32_t)((val)&0x00000000FFFFFFFF);           \
+    val1_m = (uint32_t)(((val) >> 32) & 0x00000000FFFFFFFF); \
+                                                             \
+    SW(val0_m, pdst_m1);                                     \
+    SW(val1_m, pdst_m1 + 4);                                 \
+  }
+#endif  // (__mips_isa_rev >= 6)
+
+/* Description : Load 4 words with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1, out2, out3
+   Details     : Load word in 'out0' from (psrc)
+                 Load word in 'out1' from (psrc + stride)
+                 Load word in 'out2' from (psrc + 2 * stride)
+                 Load word in 'out3' from (psrc + 3 * stride)
+*/
+#define LW4(psrc, stride, out0, out1, out2, out3) \
+  {                                               \
+    out0 = LW((psrc));                            \
+    out1 = LW((psrc) + stride);                   \
+    out2 = LW((psrc) + 2 * stride);               \
+    out3 = LW((psrc) + 3 * stride);               \
+  }
+
+/* Description : Load double words with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+   Details     : Load double word in 'out0' from (psrc)
+                 Load double word in 'out1' from (psrc + stride)
+*/
+#define LD2(psrc, stride, out0, out1) \
+  {                                   \
+    out0 = LD((psrc));                \
+    out1 = LD((psrc) + stride);       \
+  }
+#define LD4(psrc, stride, out0, out1, out2, out3) \
+  {                                               \
+    LD2((psrc), stride, out0, out1);              \
+    LD2((psrc) + 2 * stride, stride, out2, out3); \
+  }
+
+/* Description : Store 4 words with stride
+   Arguments   : Inputs - in0, in1, in2, in3, pdst, stride
+   Details     : Store word from 'in0' to (pdst)
+                 Store word from 'in1' to (pdst + stride)
+                 Store word from 'in2' to (pdst + 2 * stride)
+                 Store word from 'in3' to (pdst + 3 * stride)
+*/
+#define SW4(in0, in1, in2, in3, pdst, stride) \
+  {                                           \
+    SW(in0, (pdst))                           \
+    SW(in1, (pdst) + stride);                 \
+    SW(in2, (pdst) + 2 * stride);             \
+    SW(in3, (pdst) + 3 * stride);             \
+  }
+
+/* Description : Store 4 double words with stride
+   Arguments   : Inputs - in0, in1, in2, in3, pdst, stride
+   Details     : Store double word from 'in0' to (pdst)
+                 Store double word from 'in1' to (pdst + stride)
+                 Store double word from 'in2' to (pdst + 2 * stride)
+                 Store double word from 'in3' to (pdst + 3 * stride)
+*/
+#define SD4(in0, in1, in2, in3, pdst, stride) \
+  {                                           \
+    SD(in0, (pdst))                           \
+    SD(in1, (pdst) + stride);                 \
+    SD(in2, (pdst) + 2 * stride);             \
+    SD(in3, (pdst) + 3 * stride);             \
+  }
+
+/* Description : Load vectors with 16 byte elements with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Load 16 byte elements in 'out0' from (psrc)
+                 Load 16 byte elements in 'out1' from (psrc + stride)
+*/
+#define LD_B2(RTYPE, psrc, stride, out0, out1) \
+  {                                            \
+    out0 = LD_B(RTYPE, (psrc));                \
+    out1 = LD_B(RTYPE, (psrc) + stride);       \
+  }
+#define LD_UB2(...) LD_B2(v16u8, __VA_ARGS__)
+#define LD_SB2(...) LD_B2(v16i8, __VA_ARGS__)
+
+#define LD_B3(RTYPE, psrc, stride, out0, out1, out2) \
+  {                                                  \
+    LD_B2(RTYPE, (psrc), stride, out0, out1);        \
+    out2 = LD_B(RTYPE, (psrc) + 2 * stride);         \
+  }
+#define LD_UB3(...) LD_B3(v16u8, __VA_ARGS__)
+
+#define LD_B4(RTYPE, psrc, stride, out0, out1, out2, out3) \
+  {                                                        \
+    LD_B2(RTYPE, (psrc), stride, out0, out1);              \
+    LD_B2(RTYPE, (psrc) + 2 * stride, stride, out2, out3); \
+  }
+#define LD_UB4(...) LD_B4(v16u8, __VA_ARGS__)
+#define LD_SB4(...) LD_B4(v16i8, __VA_ARGS__)
+
+#define LD_B5(RTYPE, psrc, stride, out0, out1, out2, out3, out4) \
+  {                                                              \
+    LD_B4(RTYPE, (psrc), stride, out0, out1, out2, out3);        \
+    out4 = LD_B(RTYPE, (psrc) + 4 * stride);                     \
+  }
+#define LD_UB5(...) LD_B5(v16u8, __VA_ARGS__)
+#define LD_SB5(...) LD_B5(v16i8, __VA_ARGS__)
+
+#define LD_B7(RTYPE, psrc, stride, out0, out1, out2, out3, out4, out5, out6) \
+  {                                                                          \
+    LD_B5(RTYPE, (psrc), stride, out0, out1, out2, out3, out4);              \
+    LD_B2(RTYPE, (psrc) + 5 * stride, stride, out5, out6);                   \
+  }
+#define LD_SB7(...) LD_B7(v16i8, __VA_ARGS__)
+
+#define LD_B8(RTYPE, psrc, stride, out0, out1, out2, out3, out4, out5, out6, \
+              out7)                                                          \
+  {                                                                          \
+    LD_B4(RTYPE, (psrc), stride, out0, out1, out2, out3);                    \
+    LD_B4(RTYPE, (psrc) + 4 * stride, stride, out4, out5, out6, out7);       \
+  }
+#define LD_UB8(...) LD_B8(v16u8, __VA_ARGS__)
+#define LD_SB8(...) LD_B8(v16i8, __VA_ARGS__)
+
+/* Description : Load vectors with 8 halfword elements with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+   Details     : Load 8 halfword elements in 'out0' from (psrc)
+                 Load 8 halfword elements in 'out1' from (psrc + stride)
+*/
+#define LD_H2(RTYPE, psrc, stride, out0, out1) \
+  {                                            \
+    out0 = LD_H(RTYPE, (psrc));                \
+    out1 = LD_H(RTYPE, (psrc) + (stride));     \
+  }
+#define LD_SH2(...) LD_H2(v8i16, __VA_ARGS__)
+
+#define LD_H4(RTYPE, psrc, stride, out0, out1, out2, out3) \
+  {                                                        \
+    LD_H2(RTYPE, (psrc), stride, out0, out1);              \
+    LD_H2(RTYPE, (psrc) + 2 * stride, stride, out2, out3); \
+  }
+#define LD_SH4(...) LD_H4(v8i16, __VA_ARGS__)
+
+#define LD_H8(RTYPE, psrc, stride, out0, out1, out2, out3, out4, out5, out6, \
+              out7)                                                          \
+  {                                                                          \
+    LD_H4(RTYPE, (psrc), stride, out0, out1, out2, out3);                    \
+    LD_H4(RTYPE, (psrc) + 4 * stride, stride, out4, out5, out6, out7);       \
+  }
+#define LD_SH8(...) LD_H8(v8i16, __VA_ARGS__)
+
+#define LD_H16(RTYPE, psrc, stride, out0, out1, out2, out3, out4, out5, out6,  \
+               out7, out8, out9, out10, out11, out12, out13, out14, out15)     \
+  {                                                                            \
+    LD_H8(RTYPE, (psrc), stride, out0, out1, out2, out3, out4, out5, out6,     \
+          out7);                                                               \
+    LD_H8(RTYPE, (psrc) + 8 * stride, stride, out8, out9, out10, out11, out12, \
+          out13, out14, out15);                                                \
+  }
+#define LD_SH16(...) LD_H16(v8i16, __VA_ARGS__)
+
+/* Description : Load 4x4 block of signed halfword elements from 1D source
+                 data into 4 vectors (Each vector with 4 signed halfwords)
+   Arguments   : Input   - psrc
+                 Outputs - out0, out1, out2, out3
+*/
+#define LD4x4_SH(psrc, out0, out1, out2, out3)            \
+  {                                                       \
+    out0 = LD_SH(psrc);                                   \
+    out2 = LD_SH(psrc + 8);                               \
+    out1 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out0); \
+    out3 = (v8i16)__msa_ilvl_d((v2i64)out2, (v2i64)out2); \
+  }
+
+/* Description : Load 2 vectors of signed word elements with stride
+   Arguments   : Inputs  - psrc, stride
+                 Outputs - out0, out1
+                 Return Type - signed word
+*/
+#define LD_SW2(psrc, stride, out0, out1) \
+  {                                      \
+    out0 = LD_SW((psrc));                \
+    out1 = LD_SW((psrc) + stride);       \
+  }
+
+/* Description : Store vectors of 16 byte elements with stride
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 16 byte elements from 'in0' to (pdst)
+                 Store 16 byte elements from 'in1' to (pdst + stride)
+*/
+#define ST_B2(RTYPE, in0, in1, pdst, stride) \
+  {                                          \
+    ST_B(RTYPE, in0, (pdst));                \
+    ST_B(RTYPE, in1, (pdst) + stride);       \
+  }
+#define ST_UB2(...) ST_B2(v16u8, __VA_ARGS__)
+
+#define ST_B4(RTYPE, in0, in1, in2, in3, pdst, stride)   \
+  {                                                      \
+    ST_B2(RTYPE, in0, in1, (pdst), stride);              \
+    ST_B2(RTYPE, in2, in3, (pdst) + 2 * stride, stride); \
+  }
+#define ST_UB4(...) ST_B4(v16u8, __VA_ARGS__)
+
+#define ST_B8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, pdst, stride) \
+  {                                                                        \
+    ST_B4(RTYPE, in0, in1, in2, in3, pdst, stride);                        \
+    ST_B4(RTYPE, in4, in5, in6, in7, (pdst) + 4 * stride, stride);         \
+  }
+#define ST_UB8(...) ST_B8(v16u8, __VA_ARGS__)
+
+/* Description : Store vectors of 8 halfword elements with stride
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 8 halfword elements from 'in0' to (pdst)
+                 Store 8 halfword elements from 'in1' to (pdst + stride)
+*/
+#define ST_H2(RTYPE, in0, in1, pdst, stride) \
+  {                                          \
+    ST_H(RTYPE, in0, (pdst));                \
+    ST_H(RTYPE, in1, (pdst) + stride);       \
+  }
+#define ST_SH2(...) ST_H2(v8i16, __VA_ARGS__)
+
+#define ST_H4(RTYPE, in0, in1, in2, in3, pdst, stride)   \
+  {                                                      \
+    ST_H2(RTYPE, in0, in1, (pdst), stride);              \
+    ST_H2(RTYPE, in2, in3, (pdst) + 2 * stride, stride); \
+  }
+#define ST_SH4(...) ST_H4(v8i16, __VA_ARGS__)
+
+#define ST_H8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, pdst, stride) \
+  {                                                                        \
+    ST_H4(RTYPE, in0, in1, in2, in3, (pdst), stride);                      \
+    ST_H4(RTYPE, in4, in5, in6, in7, (pdst) + 4 * stride, stride);         \
+  }
+#define ST_SH8(...) ST_H8(v8i16, __VA_ARGS__)
+
+/* Description : Store vectors of word elements with stride
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Store 4 word elements from 'in0' to (pdst)
+                 Store 4 word elements from 'in1' to (pdst + stride)
+*/
+#define ST_SW2(in0, in1, pdst, stride) \
+  {                                    \
+    ST_SW(in0, (pdst));                \
+    ST_SW(in1, (pdst) + stride);       \
+  }
+
+/* Description : Store 2x4 byte block to destination memory from input vector
+   Arguments   : Inputs - in, stidx, pdst, stride
+   Details     : Index 'stidx' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst)
+                 Index 'stidx+1' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst + stride)
+                 Index 'stidx+2' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst + 2 * stride)
+                 Index 'stidx+3' halfword element from 'in' vector is copied to
+                 the GP register and stored to (pdst + 3 * stride)
+*/
+#define ST2x4_UB(in, stidx, pdst, stride)            \
+  {                                                  \
+    uint16_t out0_m, out1_m, out2_m, out3_m;         \
+    uint8_t *pblk_2x4_m = (uint8_t *)(pdst);         \
+                                                     \
+    out0_m = __msa_copy_u_h((v8i16)in, (stidx));     \
+    out1_m = __msa_copy_u_h((v8i16)in, (stidx + 1)); \
+    out2_m = __msa_copy_u_h((v8i16)in, (stidx + 2)); \
+    out3_m = __msa_copy_u_h((v8i16)in, (stidx + 3)); \
+                                                     \
+    SH(out0_m, pblk_2x4_m);                          \
+    SH(out1_m, pblk_2x4_m + stride);                 \
+    SH(out2_m, pblk_2x4_m + 2 * stride);             \
+    SH(out3_m, pblk_2x4_m + 3 * stride);             \
+  }
+
+/* Description : Store 4x2 byte block to destination memory from input vector
+   Arguments   : Inputs - in, pdst, stride
+   Details     : Index 0 word element from 'in' vector is copied to the GP
+                 register and stored to (pdst)
+                 Index 1 word element from 'in' vector is copied to the GP
+                 register and stored to (pdst + stride)
+*/
+#define ST4x2_UB(in, pdst, stride)           \
+  {                                          \
+    uint32_t out0_m, out1_m;                 \
+    uint8_t *pblk_4x2_m = (uint8_t *)(pdst); \
+                                             \
+    out0_m = __msa_copy_u_w((v4i32)in, 0);   \
+    out1_m = __msa_copy_u_w((v4i32)in, 1);   \
+                                             \
+    SW(out0_m, pblk_4x2_m);                  \
+    SW(out1_m, pblk_4x2_m + stride);         \
+  }
+
+/* Description : Store 4x4 byte block to destination memory from input vector
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : 'Idx0' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst)
+                 'Idx1' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst + stride)
+                 'Idx2' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst + 2 * stride)
+                 'Idx3' word element from input vector 'in0' is copied to the
+                 GP register and stored to (pdst + 3 * stride)
+*/
+#define ST4x4_UB(in0, in1, idx0, idx1, idx2, idx3, pdst, stride) \
+  {                                                              \
+    uint32_t out0_m, out1_m, out2_m, out3_m;                     \
+    uint8_t *pblk_4x4_m = (uint8_t *)(pdst);                     \
+                                                                 \
+    out0_m = __msa_copy_u_w((v4i32)in0, idx0);                   \
+    out1_m = __msa_copy_u_w((v4i32)in0, idx1);                   \
+    out2_m = __msa_copy_u_w((v4i32)in1, idx2);                   \
+    out3_m = __msa_copy_u_w((v4i32)in1, idx3);                   \
+                                                                 \
+    SW4(out0_m, out1_m, out2_m, out3_m, pblk_4x4_m, stride);     \
+  }
+#define ST4x8_UB(in0, in1, pdst, stride)                           \
+  {                                                                \
+    uint8_t *pblk_4x8 = (uint8_t *)(pdst);                         \
+                                                                   \
+    ST4x4_UB(in0, in0, 0, 1, 2, 3, pblk_4x8, stride);              \
+    ST4x4_UB(in1, in1, 0, 1, 2, 3, pblk_4x8 + 4 * stride, stride); \
+  }
+
+/* Description : Store 8x1 byte block to destination memory from input vector
+   Arguments   : Inputs - in, pdst
+   Details     : Index 0 double word element from 'in' vector is copied to the
+                 GP register and stored to (pdst)
+*/
+#define ST8x1_UB(in, pdst)                 \
+  {                                        \
+    uint64_t out0_m;                       \
+                                           \
+    out0_m = __msa_copy_u_d((v2i64)in, 0); \
+    SD(out0_m, pdst);                      \
+  }
+
+/* Description : Store 8x2 byte block to destination memory from input vector
+   Arguments   : Inputs - in, pdst, stride
+   Details     : Index 0 double word element from 'in' vector is copied to the
+                 GP register and stored to (pdst)
+                 Index 1 double word element from 'in' vector is copied to the
+                 GP register and stored to (pdst + stride)
+*/
+#define ST8x2_UB(in, pdst, stride)           \
+  {                                          \
+    uint64_t out0_m, out1_m;                 \
+    uint8_t *pblk_8x2_m = (uint8_t *)(pdst); \
+                                             \
+    out0_m = __msa_copy_u_d((v2i64)in, 0);   \
+    out1_m = __msa_copy_u_d((v2i64)in, 1);   \
+                                             \
+    SD(out0_m, pblk_8x2_m);                  \
+    SD(out1_m, pblk_8x2_m + stride);         \
+  }
+
+/* Description : Store 8x4 byte block to destination memory from input
+                 vectors
+   Arguments   : Inputs - in0, in1, pdst, stride
+   Details     : Index 0 double word element from 'in0' vector is copied to the
+                 GP register and stored to (pdst)
+                 Index 1 double word element from 'in0' vector is copied to the
+                 GP register and stored to (pdst + stride)
+                 Index 0 double word element from 'in1' vector is copied to the
+                 GP register and stored to (pdst + 2 * stride)
+                 Index 1 double word element from 'in1' vector is copied to the
+                 GP register and stored to (pdst + 3 * stride)
+*/
+#define ST8x4_UB(in0, in1, pdst, stride)                     \
+  {                                                          \
+    uint64_t out0_m, out1_m, out2_m, out3_m;                 \
+    uint8_t *pblk_8x4_m = (uint8_t *)(pdst);                 \
+                                                             \
+    out0_m = __msa_copy_u_d((v2i64)in0, 0);                  \
+    out1_m = __msa_copy_u_d((v2i64)in0, 1);                  \
+    out2_m = __msa_copy_u_d((v2i64)in1, 0);                  \
+    out3_m = __msa_copy_u_d((v2i64)in1, 1);                  \
+                                                             \
+    SD4(out0_m, out1_m, out2_m, out3_m, pblk_8x4_m, stride); \
+  }
+
+/* Description : average with rounding (in0 + in1 + 1) / 2.
+   Arguments   : Inputs  - in0, in1, in2, in3,
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each unsigned byte element from 'in0' vector is added with
+                 each unsigned byte element from 'in1' vector. Then the average
+                 with rounding is calculated and written to 'out0'
+*/
+#define AVER_UB2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+  {                                                       \
+    out0 = (RTYPE)__msa_aver_u_b((v16u8)in0, (v16u8)in1); \
+    out1 = (RTYPE)__msa_aver_u_b((v16u8)in2, (v16u8)in3); \
+  }
+#define AVER_UB2_UB(...) AVER_UB2(v16u8, __VA_ARGS__)
+
+#define AVER_UB4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                 out2, out3)                                                \
+  {                                                                         \
+    AVER_UB2(RTYPE, in0, in1, in2, in3, out0, out1)                         \
+    AVER_UB2(RTYPE, in4, in5, in6, in7, out2, out3)                         \
+  }
+#define AVER_UB4_UB(...) AVER_UB4(v16u8, __VA_ARGS__)
+
+/* Description : Immediate number of elements to slide with zero
+   Arguments   : Inputs  - in0, in1, slide_val
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'zero_m' vector are slid into 'in0' by
+                 value specified in the 'slide_val'
+*/
+#define SLDI_B2_0(RTYPE, in0, in1, out0, out1, slide_val)             \
+  {                                                                   \
+    v16i8 zero_m = { 0 };                                             \
+    out0 = (RTYPE)__msa_sldi_b((v16i8)zero_m, (v16i8)in0, slide_val); \
+    out1 = (RTYPE)__msa_sldi_b((v16i8)zero_m, (v16i8)in1, slide_val); \
+  }
+#define SLDI_B2_0_SW(...) SLDI_B2_0(v4i32, __VA_ARGS__)
+
+#define SLDI_B4_0(RTYPE, in0, in1, in2, in3, out0, out1, out2, out3, \
+                  slide_val)                                         \
+  {                                                                  \
+    SLDI_B2_0(RTYPE, in0, in1, out0, out1, slide_val);               \
+    SLDI_B2_0(RTYPE, in2, in3, out2, out3, slide_val);               \
+  }
+#define SLDI_B4_0_UB(...) SLDI_B4_0(v16u8, __VA_ARGS__)
+
+/* Description : Immediate number of elements to slide
+   Arguments   : Inputs  - in0_0, in0_1, in1_0, in1_1, slide_val
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'in0_0' vector are slid into 'in1_0' by
+                 value specified in the 'slide_val'
+*/
+#define SLDI_B2(RTYPE, in0_0, in0_1, in1_0, in1_1, out0, out1, slide_val) \
+  {                                                                       \
+    out0 = (RTYPE)__msa_sldi_b((v16i8)in0_0, (v16i8)in1_0, slide_val);    \
+    out1 = (RTYPE)__msa_sldi_b((v16i8)in0_1, (v16i8)in1_1, slide_val);    \
+  }
+#define SLDI_B2_UB(...) SLDI_B2(v16u8, __VA_ARGS__)
+#define SLDI_B2_SH(...) SLDI_B2(v8i16, __VA_ARGS__)
+
+#define SLDI_B3(RTYPE, in0_0, in0_1, in0_2, in1_0, in1_1, in1_2, out0, out1, \
+                out2, slide_val)                                             \
+  {                                                                          \
+    SLDI_B2(RTYPE, in0_0, in0_1, in1_0, in1_1, out0, out1, slide_val)        \
+    out2 = (RTYPE)__msa_sldi_b((v16i8)in0_2, (v16i8)in1_2, slide_val);       \
+  }
+#define SLDI_B3_SB(...) SLDI_B3(v16i8, __VA_ARGS__)
+#define SLDI_B3_UH(...) SLDI_B3(v8u16, __VA_ARGS__)
+
+/* Description : Shuffle byte vector elements as per mask vector
+   Arguments   : Inputs  - in0, in1, in2, in3, mask0, mask1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Byte elements from 'in0' & 'in1' are copied selectively to
+                 'out0' as per control vector 'mask0'
+*/
+#define VSHF_B2(RTYPE, in0, in1, in2, in3, mask0, mask1, out0, out1)  \
+  {                                                                   \
+    out0 = (RTYPE)__msa_vshf_b((v16i8)mask0, (v16i8)in1, (v16i8)in0); \
+    out1 = (RTYPE)__msa_vshf_b((v16i8)mask1, (v16i8)in3, (v16i8)in2); \
+  }
+#define VSHF_B2_UB(...) VSHF_B2(v16u8, __VA_ARGS__)
+#define VSHF_B2_SB(...) VSHF_B2(v16i8, __VA_ARGS__)
+#define VSHF_B2_UH(...) VSHF_B2(v8u16, __VA_ARGS__)
+
+#define VSHF_B4(RTYPE, in0, in1, mask0, mask1, mask2, mask3, out0, out1, out2, \
+                out3)                                                          \
+  {                                                                            \
+    VSHF_B2(RTYPE, in0, in1, in0, in1, mask0, mask1, out0, out1);              \
+    VSHF_B2(RTYPE, in0, in1, in0, in1, mask2, mask3, out2, out3);              \
+  }
+#define VSHF_B4_SB(...) VSHF_B4(v16i8, __VA_ARGS__)
+#define VSHF_B4_SH(...) VSHF_B4(v8i16, __VA_ARGS__)
+
+/* Description : Dot product of byte vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Unsigned byte elements from 'mult0' are multiplied with
+                 unsigned byte elements from 'cnst0' producing a result
+                 twice the size of input i.e. unsigned halfword.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_UB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \
+  {                                                             \
+    out0 = (RTYPE)__msa_dotp_u_h((v16u8)mult0, (v16u8)cnst0);   \
+    out1 = (RTYPE)__msa_dotp_u_h((v16u8)mult1, (v16u8)cnst1);   \
+  }
+#define DOTP_UB2_UH(...) DOTP_UB2(v8u16, __VA_ARGS__)
+
+#define DOTP_UB4(RTYPE, mult0, mult1, mult2, mult3, cnst0, cnst1, cnst2, \
+                 cnst3, out0, out1, out2, out3)                          \
+  {                                                                      \
+    DOTP_UB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);             \
+    DOTP_UB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);             \
+  }
+#define DOTP_UB4_UH(...) DOTP_UB4(v8u16, __VA_ARGS__)
+
+/* Description : Dot product of byte vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed byte elements from 'mult0' are multiplied with
+                 signed byte elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed halfword.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \
+  {                                                             \
+    out0 = (RTYPE)__msa_dotp_s_h((v16i8)mult0, (v16i8)cnst0);   \
+    out1 = (RTYPE)__msa_dotp_s_h((v16i8)mult1, (v16i8)cnst1);   \
+  }
+#define DOTP_SB2_SH(...) DOTP_SB2(v8i16, __VA_ARGS__)
+
+#define DOTP_SB4(RTYPE, mult0, mult1, mult2, mult3, cnst0, cnst1, cnst2, \
+                 cnst3, out0, out1, out2, out3)                          \
+  {                                                                      \
+    DOTP_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);             \
+    DOTP_SB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);             \
+  }
+#define DOTP_SB4_SH(...) DOTP_SB4(v8i16, __VA_ARGS__)
+
+/* Description : Dot product of halfword vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed halfword elements from 'mult0' are multiplied with
+                 signed halfword elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed word.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \
+  {                                                             \
+    out0 = (RTYPE)__msa_dotp_s_w((v8i16)mult0, (v8i16)cnst0);   \
+    out1 = (RTYPE)__msa_dotp_s_w((v8i16)mult1, (v8i16)cnst1);   \
+  }
+#define DOTP_SH2_SW(...) DOTP_SH2(v4i32, __VA_ARGS__)
+
+#define DOTP_SH4(RTYPE, mult0, mult1, mult2, mult3, cnst0, cnst1, cnst2, \
+                 cnst3, out0, out1, out2, out3)                          \
+  {                                                                      \
+    DOTP_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);             \
+    DOTP_SH2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);             \
+  }
+#define DOTP_SH4_SW(...) DOTP_SH4(v4i32, __VA_ARGS__)
+
+/* Description : Dot product of word vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed word elements from 'mult0' are multiplied with
+                 signed word elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed double word.
+                 The multiplication result of adjacent odd-even elements
+                 are added together and written to the 'out0' vector
+*/
+#define DOTP_SW2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1) \
+  {                                                             \
+    out0 = (RTYPE)__msa_dotp_s_d((v4i32)mult0, (v4i32)cnst0);   \
+    out1 = (RTYPE)__msa_dotp_s_d((v4i32)mult1, (v4i32)cnst1);   \
+  }
+#define DOTP_SW2_SD(...) DOTP_SW2(v2i64, __VA_ARGS__)
+
+/* Description : Dot product & addition of byte vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed byte elements from 'mult0' are multiplied with
+                 signed byte elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed halfword.
+                 The multiplication result of adjacent odd-even elements
+                 are added to the 'out0' vector
+*/
+#define DPADD_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1)            \
+  {                                                                         \
+    out0 = (RTYPE)__msa_dpadd_s_h((v8i16)out0, (v16i8)mult0, (v16i8)cnst0); \
+    out1 = (RTYPE)__msa_dpadd_s_h((v8i16)out1, (v16i8)mult1, (v16i8)cnst1); \
+  }
+#define DPADD_SB2_SH(...) DPADD_SB2(v8i16, __VA_ARGS__)
+
+#define DPADD_SB4(RTYPE, mult0, mult1, mult2, mult3, cnst0, cnst1, cnst2, \
+                  cnst3, out0, out1, out2, out3)                          \
+  {                                                                       \
+    DPADD_SB2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1);             \
+    DPADD_SB2(RTYPE, mult2, mult3, cnst2, cnst3, out2, out3);             \
+  }
+#define DPADD_SB4_SH(...) DPADD_SB4(v8i16, __VA_ARGS__)
+
+/* Description : Dot product & addition of halfword vector elements
+   Arguments   : Inputs  - mult0, mult1, cnst0, cnst1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed halfword elements from 'mult0' are multiplied with
+                 signed halfword elements from 'cnst0' producing a result
+                 twice the size of input i.e. signed word.
+                 The multiplication result of adjacent odd-even elements
+                 are added to the 'out0' vector
+*/
+#define DPADD_SH2(RTYPE, mult0, mult1, cnst0, cnst1, out0, out1)            \
+  {                                                                         \
+    out0 = (RTYPE)__msa_dpadd_s_w((v4i32)out0, (v8i16)mult0, (v8i16)cnst0); \
+    out1 = (RTYPE)__msa_dpadd_s_w((v4i32)out1, (v8i16)mult1, (v8i16)cnst1); \
+  }
+#define DPADD_SH2_SW(...) DPADD_SH2(v4i32, __VA_ARGS__)
+
+/* Description : Dot product & addition of double word vector elements
+   Arguments   : Inputs  - mult0, mult1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each signed word element from 'mult0' is multiplied with itself
+                 producing an intermediate result twice the size of input
+                 i.e. signed double word
+                 The multiplication result of adjacent odd-even elements
+                 are added to the 'out0' vector
+*/
+#define DPADD_SD2(RTYPE, mult0, mult1, out0, out1)                          \
+  {                                                                         \
+    out0 = (RTYPE)__msa_dpadd_s_d((v2i64)out0, (v4i32)mult0, (v4i32)mult0); \
+    out1 = (RTYPE)__msa_dpadd_s_d((v2i64)out1, (v4i32)mult1, (v4i32)mult1); \
+  }
+#define DPADD_SD2_SD(...) DPADD_SD2(v2i64, __VA_ARGS__)
+
+/* Description : Minimum values between unsigned elements of
+                 either vector are copied to the output vector
+   Arguments   : Inputs  - in0, in1, min_vec
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Minimum of unsigned halfword element values from 'in0' and
+                 'min_vec' are written to output vector 'in0'
+*/
+#define MIN_UH2(RTYPE, in0, in1, min_vec)            \
+  {                                                  \
+    in0 = (RTYPE)__msa_min_u_h((v8u16)in0, min_vec); \
+    in1 = (RTYPE)__msa_min_u_h((v8u16)in1, min_vec); \
+  }
+#define MIN_UH2_UH(...) MIN_UH2(v8u16, __VA_ARGS__)
+
+#define MIN_UH4(RTYPE, in0, in1, in2, in3, min_vec) \
+  {                                                 \
+    MIN_UH2(RTYPE, in0, in1, min_vec);              \
+    MIN_UH2(RTYPE, in2, in3, min_vec);              \
+  }
+#define MIN_UH4_UH(...) MIN_UH4(v8u16, __VA_ARGS__)
+
+/* Description : Clips all signed halfword elements of input vector
+                 between 0 & 255
+   Arguments   : Input  - in
+                 Output - out_m
+                 Return Type - signed halfword
+*/
+#define CLIP_SH_0_255(in)                              \
+  ({                                                   \
+    v8i16 max_m = __msa_ldi_h(255);                    \
+    v8i16 out_m;                                       \
+                                                       \
+    out_m = __msa_maxi_s_h((v8i16)in, 0);              \
+    out_m = __msa_min_s_h((v8i16)max_m, (v8i16)out_m); \
+    out_m;                                             \
+  })
+#define CLIP_SH2_0_255(in0, in1) \
+  {                              \
+    in0 = CLIP_SH_0_255(in0);    \
+    in1 = CLIP_SH_0_255(in1);    \
+  }
+#define CLIP_SH4_0_255(in0, in1, in2, in3) \
+  {                                        \
+    CLIP_SH2_0_255(in0, in1);              \
+    CLIP_SH2_0_255(in2, in3);              \
+  }
+
+/* Description : Horizontal addition of 4 signed word elements of input vector
+   Arguments   : Input  - in       (signed word vector)
+                 Output - sum_m    (i32 sum)
+                 Return Type - signed word (GP)
+   Details     : 4 signed word elements of 'in' vector are added together and
+                 the resulting integer sum is returned
+*/
+#define HADD_SW_S32(in)                            \
+  ({                                               \
+    v2i64 res0_m, res1_m;                          \
+    int32_t sum_m;                                 \
+                                                   \
+    res0_m = __msa_hadd_s_d((v4i32)in, (v4i32)in); \
+    res1_m = __msa_splati_d(res0_m, 1);            \
+    res0_m = res0_m + res1_m;                      \
+    sum_m = __msa_copy_s_w((v4i32)res0_m, 0);      \
+    sum_m;                                         \
+  })
+
+/* Description : Horizontal addition of 8 unsigned halfword elements
+   Arguments   : Inputs  - in       (unsigned halfword vector)
+                 Outputs - sum_m    (u32 sum)
+                 Return Type - unsigned word
+   Details     : 8 unsigned halfword elements of input vector are added
+                 together and the resulting integer sum is returned
+*/
+#define HADD_UH_U32(in)                               \
+  ({                                                  \
+    v4u32 res_m;                                      \
+    v2u64 res0_m, res1_m;                             \
+    uint32_t sum_m;                                   \
+                                                      \
+    res_m = __msa_hadd_u_w((v8u16)in, (v8u16)in);     \
+    res0_m = __msa_hadd_u_d(res_m, res_m);            \
+    res1_m = (v2u64)__msa_splati_d((v2i64)res0_m, 1); \
+    res0_m = res0_m + res1_m;                         \
+    sum_m = __msa_copy_u_w((v4i32)res0_m, 0);         \
+    sum_m;                                            \
+  })
+
+/* Description : Horizontal addition of unsigned byte vector elements
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each unsigned odd byte element from 'in0' is added to
+                 even unsigned byte element from 'in0' (pairwise) and the
+                 halfword result is written to 'out0'
+*/
+#define HADD_UB2(RTYPE, in0, in1, out0, out1)             \
+  {                                                       \
+    out0 = (RTYPE)__msa_hadd_u_h((v16u8)in0, (v16u8)in0); \
+    out1 = (RTYPE)__msa_hadd_u_h((v16u8)in1, (v16u8)in1); \
+  }
+#define HADD_UB2_UH(...) HADD_UB2(v8u16, __VA_ARGS__)
+
+#define HADD_UB4(RTYPE, in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                                 \
+    HADD_UB2(RTYPE, in0, in1, out0, out1);                          \
+    HADD_UB2(RTYPE, in2, in3, out2, out3);                          \
+  }
+#define HADD_UB4_UH(...) HADD_UB4(v8u16, __VA_ARGS__)
+
+/* Description : Horizontal subtraction of unsigned byte vector elements
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each unsigned odd byte element from 'in0' is subtracted from
+                 even unsigned byte element from 'in0' (pairwise) and the
+                 halfword result is written to 'out0'
+*/
+#define HSUB_UB2(RTYPE, in0, in1, out0, out1)             \
+  {                                                       \
+    out0 = (RTYPE)__msa_hsub_u_h((v16u8)in0, (v16u8)in0); \
+    out1 = (RTYPE)__msa_hsub_u_h((v16u8)in1, (v16u8)in1); \
+  }
+#define HSUB_UB2_SH(...) HSUB_UB2(v8i16, __VA_ARGS__)
+
+/* Description : SAD (Sum of Absolute Difference)
+   Arguments   : Inputs  - in0, in1, ref0, ref1
+                 Outputs - sad_m                 (halfword vector)
+                 Return Type - unsigned halfword
+   Details     : Absolute difference of all the byte elements from 'in0' with
+                 'ref0' is calculated and preserved in 'diff0'. Then even-odd
+                 pairs are added together to generate 8 halfword results.
+*/
+#define SAD_UB2_UH(in0, in1, ref0, ref1)                     \
+  ({                                                         \
+    v16u8 diff0_m, diff1_m;                                  \
+    v8u16 sad_m = { 0 };                                     \
+                                                             \
+    diff0_m = __msa_asub_u_b((v16u8)in0, (v16u8)ref0);       \
+    diff1_m = __msa_asub_u_b((v16u8)in1, (v16u8)ref1);       \
+                                                             \
+    sad_m += __msa_hadd_u_h((v16u8)diff0_m, (v16u8)diff0_m); \
+    sad_m += __msa_hadd_u_h((v16u8)diff1_m, (v16u8)diff1_m); \
+                                                             \
+    sad_m;                                                   \
+  })
+
+/* Description : Horizontal subtraction of signed halfword vector elements
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Each signed odd halfword element from 'in0' is subtracted from
+                 even signed halfword element from 'in0' (pairwise) and the
+                 word result is written to 'out0'
+*/
+#define HSUB_UH2(RTYPE, in0, in1, out0, out1)             \
+  {                                                       \
+    out0 = (RTYPE)__msa_hsub_s_w((v8i16)in0, (v8i16)in0); \
+    out1 = (RTYPE)__msa_hsub_s_w((v8i16)in1, (v8i16)in1); \
+  }
+#define HSUB_UH2_SW(...) HSUB_UH2(v4i32, __VA_ARGS__)
+
+/* Description : Set element n input vector to GPR value
+   Arguments   : Inputs - in0, in1, in2, in3
+                 Output - out
+                 Return Type - as per RTYPE
+   Details     : Set element 0 in vector 'out' to value specified in 'in0'
+*/
+#define INSERT_W2(RTYPE, in0, in1, out)              \
+  {                                                  \
+    out = (RTYPE)__msa_insert_w((v4i32)out, 0, in0); \
+    out = (RTYPE)__msa_insert_w((v4i32)out, 1, in1); \
+  }
+#define INSERT_W2_SB(...) INSERT_W2(v16i8, __VA_ARGS__)
+
+#define INSERT_W4(RTYPE, in0, in1, in2, in3, out)    \
+  {                                                  \
+    out = (RTYPE)__msa_insert_w((v4i32)out, 0, in0); \
+    out = (RTYPE)__msa_insert_w((v4i32)out, 1, in1); \
+    out = (RTYPE)__msa_insert_w((v4i32)out, 2, in2); \
+    out = (RTYPE)__msa_insert_w((v4i32)out, 3, in3); \
+  }
+#define INSERT_W4_UB(...) INSERT_W4(v16u8, __VA_ARGS__)
+#define INSERT_W4_SB(...) INSERT_W4(v16i8, __VA_ARGS__)
+
+#define INSERT_D2(RTYPE, in0, in1, out)              \
+  {                                                  \
+    out = (RTYPE)__msa_insert_d((v2i64)out, 0, in0); \
+    out = (RTYPE)__msa_insert_d((v2i64)out, 1, in1); \
+  }
+#define INSERT_D2_UB(...) INSERT_D2(v16u8, __VA_ARGS__)
+#define INSERT_D2_SB(...) INSERT_D2(v16i8, __VA_ARGS__)
+
+/* Description : Interleave even byte elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even byte elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_B2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                      \
+    out0 = (RTYPE)__msa_ilvev_b((v16i8)in1, (v16i8)in0); \
+    out1 = (RTYPE)__msa_ilvev_b((v16i8)in3, (v16i8)in2); \
+  }
+#define ILVEV_B2_UB(...) ILVEV_B2(v16u8, __VA_ARGS__)
+#define ILVEV_B2_SH(...) ILVEV_B2(v8i16, __VA_ARGS__)
+
+/* Description : Interleave even halfword elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even halfword elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_H2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                      \
+    out0 = (RTYPE)__msa_ilvev_h((v8i16)in1, (v8i16)in0); \
+    out1 = (RTYPE)__msa_ilvev_h((v8i16)in3, (v8i16)in2); \
+  }
+#define ILVEV_H2_UB(...) ILVEV_H2(v16u8, __VA_ARGS__)
+#define ILVEV_H2_SH(...) ILVEV_H2(v8i16, __VA_ARGS__)
+#define ILVEV_H2_SW(...) ILVEV_H2(v4i32, __VA_ARGS__)
+
+/* Description : Interleave even word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even word elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_W2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                      \
+    out0 = (RTYPE)__msa_ilvev_w((v4i32)in1, (v4i32)in0); \
+    out1 = (RTYPE)__msa_ilvev_w((v4i32)in3, (v4i32)in2); \
+  }
+#define ILVEV_W2_SB(...) ILVEV_W2(v16i8, __VA_ARGS__)
+
+/* Description : Interleave even double word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even double word elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'
+*/
+#define ILVEV_D2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                      \
+    out0 = (RTYPE)__msa_ilvev_d((v2i64)in1, (v2i64)in0); \
+    out1 = (RTYPE)__msa_ilvev_d((v2i64)in3, (v2i64)in2); \
+  }
+#define ILVEV_D2_UB(...) ILVEV_D2(v16u8, __VA_ARGS__)
+
+/* Description : Interleave left half of byte elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Left half of byte elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'.
+*/
+#define ILVL_B2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                     \
+    out0 = (RTYPE)__msa_ilvl_b((v16i8)in0, (v16i8)in1); \
+    out1 = (RTYPE)__msa_ilvl_b((v16i8)in2, (v16i8)in3); \
+  }
+#define ILVL_B2_UB(...) ILVL_B2(v16u8, __VA_ARGS__)
+#define ILVL_B2_SB(...) ILVL_B2(v16i8, __VA_ARGS__)
+#define ILVL_B2_UH(...) ILVL_B2(v8u16, __VA_ARGS__)
+#define ILVL_B2_SH(...) ILVL_B2(v8i16, __VA_ARGS__)
+
+#define ILVL_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                out2, out3)                                                \
+  {                                                                        \
+    ILVL_B2(RTYPE, in0, in1, in2, in3, out0, out1);                        \
+    ILVL_B2(RTYPE, in4, in5, in6, in7, out2, out3);                        \
+  }
+#define ILVL_B4_SB(...) ILVL_B4(v16i8, __VA_ARGS__)
+#define ILVL_B4_SH(...) ILVL_B4(v8i16, __VA_ARGS__)
+#define ILVL_B4_UH(...) ILVL_B4(v8u16, __VA_ARGS__)
+
+/* Description : Interleave left half of halfword elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Left half of halfword elements of 'in0' and 'in1' are
+                 interleaved and written to 'out0'.
+*/
+#define ILVL_H2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                     \
+    out0 = (RTYPE)__msa_ilvl_h((v8i16)in0, (v8i16)in1); \
+    out1 = (RTYPE)__msa_ilvl_h((v8i16)in2, (v8i16)in3); \
+  }
+#define ILVL_H2_SH(...) ILVL_H2(v8i16, __VA_ARGS__)
+#define ILVL_H2_SW(...) ILVL_H2(v4i32, __VA_ARGS__)
+
+/* Description : Interleave left half of word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Left half of word elements of 'in0' and 'in1' are interleaved
+                 and written to 'out0'.
+*/
+#define ILVL_W2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                     \
+    out0 = (RTYPE)__msa_ilvl_w((v4i32)in0, (v4i32)in1); \
+    out1 = (RTYPE)__msa_ilvl_w((v4i32)in2, (v4i32)in3); \
+  }
+#define ILVL_W2_UB(...) ILVL_W2(v16u8, __VA_ARGS__)
+#define ILVL_W2_SH(...) ILVL_W2(v8i16, __VA_ARGS__)
+
+/* Description : Interleave right half of byte elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of byte elements of 'in0' and 'in1' are interleaved
+                 and written to out0.
+*/
+#define ILVR_B2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                     \
+    out0 = (RTYPE)__msa_ilvr_b((v16i8)in0, (v16i8)in1); \
+    out1 = (RTYPE)__msa_ilvr_b((v16i8)in2, (v16i8)in3); \
+  }
+#define ILVR_B2_UB(...) ILVR_B2(v16u8, __VA_ARGS__)
+#define ILVR_B2_SB(...) ILVR_B2(v16i8, __VA_ARGS__)
+#define ILVR_B2_UH(...) ILVR_B2(v8u16, __VA_ARGS__)
+#define ILVR_B2_SH(...) ILVR_B2(v8i16, __VA_ARGS__)
+
+#define ILVR_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                out2, out3)                                                \
+  {                                                                        \
+    ILVR_B2(RTYPE, in0, in1, in2, in3, out0, out1);                        \
+    ILVR_B2(RTYPE, in4, in5, in6, in7, out2, out3);                        \
+  }
+#define ILVR_B4_UB(...) ILVR_B4(v16u8, __VA_ARGS__)
+#define ILVR_B4_SB(...) ILVR_B4(v16i8, __VA_ARGS__)
+#define ILVR_B4_UH(...) ILVR_B4(v8u16, __VA_ARGS__)
+#define ILVR_B4_SH(...) ILVR_B4(v8i16, __VA_ARGS__)
+
+#define ILVR_B8(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, \
+                in11, in12, in13, in14, in15, out0, out1, out2, out3, out4,    \
+                out5, out6, out7)                                              \
+  {                                                                            \
+    ILVR_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2,   \
+            out3);                                                             \
+    ILVR_B4(RTYPE, in8, in9, in10, in11, in12, in13, in14, in15, out4, out5,   \
+            out6, out7);                                                       \
+  }
+#define ILVR_B8_UH(...) ILVR_B8(v8u16, __VA_ARGS__)
+
+/* Description : Interleave right half of halfword elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of halfword elements of 'in0' and 'in1' are
+                 interleaved and written to 'out0'.
+*/
+#define ILVR_H2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                     \
+    out0 = (RTYPE)__msa_ilvr_h((v8i16)in0, (v8i16)in1); \
+    out1 = (RTYPE)__msa_ilvr_h((v8i16)in2, (v8i16)in3); \
+  }
+#define ILVR_H2_SH(...) ILVR_H2(v8i16, __VA_ARGS__)
+#define ILVR_H2_SW(...) ILVR_H2(v4i32, __VA_ARGS__)
+
+#define ILVR_H4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                out2, out3)                                                \
+  {                                                                        \
+    ILVR_H2(RTYPE, in0, in1, in2, in3, out0, out1);                        \
+    ILVR_H2(RTYPE, in4, in5, in6, in7, out2, out3);                        \
+  }
+#define ILVR_H4_SH(...) ILVR_H4(v8i16, __VA_ARGS__)
+
+#define ILVR_W2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                     \
+    out0 = (RTYPE)__msa_ilvr_w((v4i32)in0, (v4i32)in1); \
+    out1 = (RTYPE)__msa_ilvr_w((v4i32)in2, (v4i32)in3); \
+  }
+#define ILVR_W2_UB(...) ILVR_W2(v16u8, __VA_ARGS__)
+#define ILVR_W2_SH(...) ILVR_W2(v8i16, __VA_ARGS__)
+
+#define ILVR_W4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                out2, out3)                                                \
+  {                                                                        \
+    ILVR_W2(RTYPE, in0, in1, in2, in3, out0, out1);                        \
+    ILVR_W2(RTYPE, in4, in5, in6, in7, out2, out3);                        \
+  }
+#define ILVR_W4_UB(...) ILVR_W4(v16u8, __VA_ARGS__)
+
+/* Description : Interleave right half of double word elements from vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of double word elements of 'in0' and 'in1' are
+                 interleaved and written to 'out0'.
+*/
+#define ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1)      \
+  {                                                         \
+    out0 = (RTYPE)__msa_ilvr_d((v2i64)(in0), (v2i64)(in1)); \
+    out1 = (RTYPE)__msa_ilvr_d((v2i64)(in2), (v2i64)(in3)); \
+  }
+#define ILVR_D2_UB(...) ILVR_D2(v16u8, __VA_ARGS__)
+#define ILVR_D2_SB(...) ILVR_D2(v16i8, __VA_ARGS__)
+#define ILVR_D2_SH(...) ILVR_D2(v8i16, __VA_ARGS__)
+
+#define ILVR_D3(RTYPE, in0, in1, in2, in3, in4, in5, out0, out1, out2) \
+  {                                                                    \
+    ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1);                    \
+    out2 = (RTYPE)__msa_ilvr_d((v2i64)(in4), (v2i64)(in5));            \
+  }
+#define ILVR_D3_SB(...) ILVR_D3(v16i8, __VA_ARGS__)
+
+#define ILVR_D4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                out2, out3)                                                \
+  {                                                                        \
+    ILVR_D2(RTYPE, in0, in1, in2, in3, out0, out1);                        \
+    ILVR_D2(RTYPE, in4, in5, in6, in7, out2, out3);                        \
+  }
+#define ILVR_D4_SB(...) ILVR_D4(v16i8, __VA_ARGS__)
+#define ILVR_D4_UB(...) ILVR_D4(v16u8, __VA_ARGS__)
+
+/* Description : Interleave both left and right half of input vectors
+   Arguments   : Inputs  - in0, in1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Right half of byte elements from 'in0' and 'in1' are
+                 interleaved and written to 'out0'
+*/
+#define ILVRL_B2(RTYPE, in0, in1, out0, out1)           \
+  {                                                     \
+    out0 = (RTYPE)__msa_ilvr_b((v16i8)in0, (v16i8)in1); \
+    out1 = (RTYPE)__msa_ilvl_b((v16i8)in0, (v16i8)in1); \
+  }
+#define ILVRL_B2_UB(...) ILVRL_B2(v16u8, __VA_ARGS__)
+#define ILVRL_B2_SB(...) ILVRL_B2(v16i8, __VA_ARGS__)
+#define ILVRL_B2_UH(...) ILVRL_B2(v8u16, __VA_ARGS__)
+#define ILVRL_B2_SH(...) ILVRL_B2(v8i16, __VA_ARGS__)
+
+#define ILVRL_H2(RTYPE, in0, in1, out0, out1)           \
+  {                                                     \
+    out0 = (RTYPE)__msa_ilvr_h((v8i16)in0, (v8i16)in1); \
+    out1 = (RTYPE)__msa_ilvl_h((v8i16)in0, (v8i16)in1); \
+  }
+#define ILVRL_H2_SH(...) ILVRL_H2(v8i16, __VA_ARGS__)
+#define ILVRL_H2_SW(...) ILVRL_H2(v4i32, __VA_ARGS__)
+
+#define ILVRL_W2(RTYPE, in0, in1, out0, out1)           \
+  {                                                     \
+    out0 = (RTYPE)__msa_ilvr_w((v4i32)in0, (v4i32)in1); \
+    out1 = (RTYPE)__msa_ilvl_w((v4i32)in0, (v4i32)in1); \
+  }
+#define ILVRL_W2_UB(...) ILVRL_W2(v16u8, __VA_ARGS__)
+#define ILVRL_W2_SH(...) ILVRL_W2(v8i16, __VA_ARGS__)
+#define ILVRL_W2_SW(...) ILVRL_W2(v4i32, __VA_ARGS__)
+
+/* Description : Saturate the halfword element values to the max
+                 unsigned value of (sat_val + 1) bits
+                 The element data width remains unchanged
+   Arguments   : Inputs  - in0, in1, sat_val
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each unsigned halfword element from 'in0' is saturated to the
+                 value generated with (sat_val + 1) bit range.
+                 The results are written in place
+*/
+#define SAT_UH2(RTYPE, in0, in1, sat_val)            \
+  {                                                  \
+    in0 = (RTYPE)__msa_sat_u_h((v8u16)in0, sat_val); \
+    in1 = (RTYPE)__msa_sat_u_h((v8u16)in1, sat_val); \
+  }
+#define SAT_UH2_UH(...) SAT_UH2(v8u16, __VA_ARGS__)
+
+#define SAT_UH4(RTYPE, in0, in1, in2, in3, sat_val) \
+  {                                                 \
+    SAT_UH2(RTYPE, in0, in1, sat_val);              \
+    SAT_UH2(RTYPE, in2, in3, sat_val)               \
+  }
+#define SAT_UH4_UH(...) SAT_UH4(v8u16, __VA_ARGS__)
+
+/* Description : Saturate the halfword element values to the max
+                 unsigned value of (sat_val + 1) bits
+                 The element data width remains unchanged
+   Arguments   : Inputs  - in0, in1, sat_val
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each unsigned halfword element from 'in0' is saturated to the
+                 value generated with (sat_val + 1) bit range
+                 The results are written in place
+*/
+#define SAT_SH2(RTYPE, in0, in1, sat_val)            \
+  {                                                  \
+    in0 = (RTYPE)__msa_sat_s_h((v8i16)in0, sat_val); \
+    in1 = (RTYPE)__msa_sat_s_h((v8i16)in1, sat_val); \
+  }
+#define SAT_SH2_SH(...) SAT_SH2(v8i16, __VA_ARGS__)
+
+#define SAT_SH4(RTYPE, in0, in1, in2, in3, sat_val) \
+  {                                                 \
+    SAT_SH2(RTYPE, in0, in1, sat_val);              \
+    SAT_SH2(RTYPE, in2, in3, sat_val);              \
+  }
+#define SAT_SH4_SH(...) SAT_SH4(v8i16, __VA_ARGS__)
+
+/* Description : Indexed halfword element values are replicated to all
+                 elements in output vector
+   Arguments   : Inputs  - in, idx0, idx1
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : 'idx0' element value from 'in' vector is replicated to all
+                  elements in 'out0' vector
+                  Valid index range for halfword operation is 0-7
+*/
+#define SPLATI_H2(RTYPE, in, idx0, idx1, out0, out1) \
+  {                                                  \
+    out0 = (RTYPE)__msa_splati_h((v8i16)in, idx0);   \
+    out1 = (RTYPE)__msa_splati_h((v8i16)in, idx1);   \
+  }
+#define SPLATI_H2_SH(...) SPLATI_H2(v8i16, __VA_ARGS__)
+
+#define SPLATI_H4(RTYPE, in, idx0, idx1, idx2, idx3, out0, out1, out2, out3) \
+  {                                                                          \
+    SPLATI_H2(RTYPE, in, idx0, idx1, out0, out1);                            \
+    SPLATI_H2(RTYPE, in, idx2, idx3, out2, out3);                            \
+  }
+#define SPLATI_H4_SB(...) SPLATI_H4(v16i8, __VA_ARGS__)
+#define SPLATI_H4_SH(...) SPLATI_H4(v8i16, __VA_ARGS__)
+
+/* Description : Pack even byte elements of vector pairs
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even byte elements of 'in0' are copied to the left half of
+                 'out0' & even byte elements of 'in1' are copied to the right
+                 half of 'out0'.
+*/
+#define PCKEV_B2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                      \
+    out0 = (RTYPE)__msa_pckev_b((v16i8)in0, (v16i8)in1); \
+    out1 = (RTYPE)__msa_pckev_b((v16i8)in2, (v16i8)in3); \
+  }
+#define PCKEV_B2_SB(...) PCKEV_B2(v16i8, __VA_ARGS__)
+#define PCKEV_B2_UB(...) PCKEV_B2(v16u8, __VA_ARGS__)
+#define PCKEV_B2_SH(...) PCKEV_B2(v8i16, __VA_ARGS__)
+
+#define PCKEV_B4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                 out2, out3)                                                \
+  {                                                                         \
+    PCKEV_B2(RTYPE, in0, in1, in2, in3, out0, out1);                        \
+    PCKEV_B2(RTYPE, in4, in5, in6, in7, out2, out3);                        \
+  }
+#define PCKEV_B4_SB(...) PCKEV_B4(v16i8, __VA_ARGS__)
+#define PCKEV_B4_UB(...) PCKEV_B4(v16u8, __VA_ARGS__)
+#define PCKEV_B4_SH(...) PCKEV_B4(v8i16, __VA_ARGS__)
+
+/* Description : Pack even halfword elements of vector pairs
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even halfword elements of 'in0' are copied to the left half of
+                 'out0' & even halfword elements of 'in1' are copied to the
+                 right half of 'out0'.
+*/
+#define PCKEV_H2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                      \
+    out0 = (RTYPE)__msa_pckev_h((v8i16)in0, (v8i16)in1); \
+    out1 = (RTYPE)__msa_pckev_h((v8i16)in2, (v8i16)in3); \
+  }
+#define PCKEV_H2_SH(...) PCKEV_H2(v8i16, __VA_ARGS__)
+#define PCKEV_H2_SW(...) PCKEV_H2(v4i32, __VA_ARGS__)
+
+#define PCKEV_H4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                 out2, out3)                                                \
+  {                                                                         \
+    PCKEV_H2(RTYPE, in0, in1, in2, in3, out0, out1);                        \
+    PCKEV_H2(RTYPE, in4, in5, in6, in7, out2, out3);                        \
+  }
+#define PCKEV_H4_SH(...) PCKEV_H4(v8i16, __VA_ARGS__)
+
+/* Description : Pack even double word elements of vector pairs
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Even double elements of 'in0' are copied to the left half of
+                 'out0' & even double elements of 'in1' are copied to the right
+                 half of 'out0'.
+*/
+#define PCKEV_D2(RTYPE, in0, in1, in2, in3, out0, out1)  \
+  {                                                      \
+    out0 = (RTYPE)__msa_pckev_d((v2i64)in0, (v2i64)in1); \
+    out1 = (RTYPE)__msa_pckev_d((v2i64)in2, (v2i64)in3); \
+  }
+#define PCKEV_D2_UB(...) PCKEV_D2(v16u8, __VA_ARGS__)
+#define PCKEV_D2_SH(...) PCKEV_D2(v8i16, __VA_ARGS__)
+
+#define PCKEV_D4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                 out2, out3)                                                \
+  {                                                                         \
+    PCKEV_D2(RTYPE, in0, in1, in2, in3, out0, out1);                        \
+    PCKEV_D2(RTYPE, in4, in5, in6, in7, out2, out3);                        \
+  }
+#define PCKEV_D4_UB(...) PCKEV_D4(v16u8, __VA_ARGS__)
+
+/* Description : Each byte element is logically xor'ed with immediate 128
+   Arguments   : Inputs  - in0, in1
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each unsigned byte element from input vector 'in0' is
+                 logically xor'ed with 128 and the result is stored in-place.
+*/
+#define XORI_B2_128(RTYPE, in0, in1)            \
+  {                                             \
+    in0 = (RTYPE)__msa_xori_b((v16u8)in0, 128); \
+    in1 = (RTYPE)__msa_xori_b((v16u8)in1, 128); \
+  }
+#define XORI_B2_128_UB(...) XORI_B2_128(v16u8, __VA_ARGS__)
+#define XORI_B2_128_SB(...) XORI_B2_128(v16i8, __VA_ARGS__)
+
+#define XORI_B3_128(RTYPE, in0, in1, in2)       \
+  {                                             \
+    XORI_B2_128(RTYPE, in0, in1);               \
+    in2 = (RTYPE)__msa_xori_b((v16u8)in2, 128); \
+  }
+#define XORI_B3_128_SB(...) XORI_B3_128(v16i8, __VA_ARGS__)
+
+#define XORI_B4_128(RTYPE, in0, in1, in2, in3) \
+  {                                            \
+    XORI_B2_128(RTYPE, in0, in1);              \
+    XORI_B2_128(RTYPE, in2, in3);              \
+  }
+#define XORI_B4_128_UB(...) XORI_B4_128(v16u8, __VA_ARGS__)
+#define XORI_B4_128_SB(...) XORI_B4_128(v16i8, __VA_ARGS__)
+
+#define XORI_B7_128(RTYPE, in0, in1, in2, in3, in4, in5, in6) \
+  {                                                           \
+    XORI_B4_128(RTYPE, in0, in1, in2, in3);                   \
+    XORI_B3_128(RTYPE, in4, in5, in6);                        \
+  }
+#define XORI_B7_128_SB(...) XORI_B7_128(v16i8, __VA_ARGS__)
+
+/* Description : Average of signed halfword elements -> (a + b) / 2
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3
+                 Return Type - as per RTYPE
+   Details     : Each signed halfword element from 'in0' is added to each
+                 signed halfword element of 'in1' with full precision resulting
+                 in one extra bit in the result. The result is then divided by
+                 2 and written to 'out0'
+*/
+#define AVE_SH4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                out2, out3)                                                \
+  {                                                                        \
+    out0 = (RTYPE)__msa_ave_s_h((v8i16)in0, (v8i16)in1);                   \
+    out1 = (RTYPE)__msa_ave_s_h((v8i16)in2, (v8i16)in3);                   \
+    out2 = (RTYPE)__msa_ave_s_h((v8i16)in4, (v8i16)in5);                   \
+    out3 = (RTYPE)__msa_ave_s_h((v8i16)in6, (v8i16)in7);                   \
+  }
+#define AVE_SH4_SH(...) AVE_SH4(v8i16, __VA_ARGS__)
+
+/* Description : Addition of signed halfword elements and signed saturation
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+                 Return Type - as per RTYPE
+   Details     : Signed halfword elements from 'in0' are added to signed
+                 halfword elements of 'in1'. The result is then signed saturated
+                 between halfword data type range
+*/
+#define ADDS_SH2(RTYPE, in0, in1, in2, in3, out0, out1)   \
+  {                                                       \
+    out0 = (RTYPE)__msa_adds_s_h((v8i16)in0, (v8i16)in1); \
+    out1 = (RTYPE)__msa_adds_s_h((v8i16)in2, (v8i16)in3); \
+  }
+#define ADDS_SH2_SH(...) ADDS_SH2(v8i16, __VA_ARGS__)
+
+#define ADDS_SH4(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                 out2, out3)                                                \
+  {                                                                         \
+    ADDS_SH2(RTYPE, in0, in1, in2, in3, out0, out1);                        \
+    ADDS_SH2(RTYPE, in4, in5, in6, in7, out2, out3);                        \
+  }
+#define ADDS_SH4_SH(...) ADDS_SH4(v8i16, __VA_ARGS__)
+
+/* Description : Shift left all elements of vector (generic for all data types)
+   Arguments   : Inputs  - in0, in1, in2, in3, shift
+                 Outputs - in place operation
+                 Return Type - as per input vector RTYPE
+   Details     : Each element of vector 'in0' is left shifted by 'shift' and
+                 the result is written in-place.
+*/
+#define SLLI_4V(in0, in1, in2, in3, shift) \
+  {                                        \
+    in0 = in0 << shift;                    \
+    in1 = in1 << shift;                    \
+    in2 = in2 << shift;                    \
+    in3 = in3 << shift;                    \
+  }
+
+/* Description : Arithmetic shift right all elements of vector
+                 (generic for all data types)
+   Arguments   : Inputs  - in0, in1, in2, in3, shift
+                 Outputs - in place operation
+                 Return Type - as per input vector RTYPE
+   Details     : Each element of vector 'in0' is right shifted by 'shift' and
+                 the result is written in-place. 'shift' is a GP variable.
+*/
+#define SRA_4V(in0, in1, in2, in3, shift) \
+  {                                       \
+    in0 = in0 >> shift;                   \
+    in1 = in1 >> shift;                   \
+    in2 = in2 >> shift;                   \
+    in3 = in3 >> shift;                   \
+  }
+
+/* Description : Shift right arithmetic rounded words
+   Arguments   : Inputs  - in0, in1, shift
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is shifted right arithmetically by
+                 the number of bits in the corresponding element in the vector
+                 'shift'. The last discarded bit is added to shifted value for
+                 rounding and the result is written in-place.
+                 'shift' is a vector.
+*/
+#define SRAR_W2(RTYPE, in0, in1, shift)                  \
+  {                                                      \
+    in0 = (RTYPE)__msa_srar_w((v4i32)in0, (v4i32)shift); \
+    in1 = (RTYPE)__msa_srar_w((v4i32)in1, (v4i32)shift); \
+  }
+
+#define SRAR_W4(RTYPE, in0, in1, in2, in3, shift) \
+  {                                               \
+    SRAR_W2(RTYPE, in0, in1, shift)               \
+    SRAR_W2(RTYPE, in2, in3, shift)               \
+  }
+#define SRAR_W4_SW(...) SRAR_W4(v4i32, __VA_ARGS__)
+
+/* Description : Shift right arithmetic rounded (immediate)
+   Arguments   : Inputs  - in0, in1, shift
+                 Outputs - in place operation
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is shifted right arithmetically by
+                 the value in 'shift'. The last discarded bit is added to the
+                 shifted value for rounding and the result is written in-place.
+                 'shift' is an immediate value.
+*/
+#define SRARI_H2(RTYPE, in0, in1, shift)           \
+  {                                                \
+    in0 = (RTYPE)__msa_srari_h((v8i16)in0, shift); \
+    in1 = (RTYPE)__msa_srari_h((v8i16)in1, shift); \
+  }
+#define SRARI_H2_UH(...) SRARI_H2(v8u16, __VA_ARGS__)
+#define SRARI_H2_SH(...) SRARI_H2(v8i16, __VA_ARGS__)
+
+#define SRARI_H4(RTYPE, in0, in1, in2, in3, shift) \
+  {                                                \
+    SRARI_H2(RTYPE, in0, in1, shift);              \
+    SRARI_H2(RTYPE, in2, in3, shift);              \
+  }
+#define SRARI_H4_UH(...) SRARI_H4(v8u16, __VA_ARGS__)
+#define SRARI_H4_SH(...) SRARI_H4(v8i16, __VA_ARGS__)
+
+#define SRARI_W2(RTYPE, in0, in1, shift)           \
+  {                                                \
+    in0 = (RTYPE)__msa_srari_w((v4i32)in0, shift); \
+    in1 = (RTYPE)__msa_srari_w((v4i32)in1, shift); \
+  }
+#define SRARI_W2_SW(...) SRARI_W2(v4i32, __VA_ARGS__)
+
+#define SRARI_W4(RTYPE, in0, in1, in2, in3, shift) \
+  {                                                \
+    SRARI_W2(RTYPE, in0, in1, shift);              \
+    SRARI_W2(RTYPE, in2, in3, shift);              \
+  }
+#define SRARI_W4_SW(...) SRARI_W4(v4i32, __VA_ARGS__)
+
+/* Description : Logical shift right all elements of vector (immediate)
+   Arguments   : Inputs  - in0, in1, in2, in3, shift
+                 Outputs - out0, out1, out2, out3
+                 Return Type - as per RTYPE
+   Details     : Each element of vector 'in0' is right shifted by 'shift' and
+                 the result is written in-place. 'shift' is an immediate value.
+*/
+#define SRLI_H4(RTYPE, in0, in1, in2, in3, out0, out1, out2, out3, shift) \
+  {                                                                       \
+    out0 = (RTYPE)__msa_srli_h((v8i16)in0, shift);                        \
+    out1 = (RTYPE)__msa_srli_h((v8i16)in1, shift);                        \
+    out2 = (RTYPE)__msa_srli_h((v8i16)in2, shift);                        \
+    out3 = (RTYPE)__msa_srli_h((v8i16)in3, shift);                        \
+  }
+#define SRLI_H4_SH(...) SRLI_H4(v8i16, __VA_ARGS__)
+
+/* Description : Multiplication of pairs of vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+   Details     : Each element from 'in0' is multiplied with elements from 'in1'
+                 and the result is written to 'out0'
+*/
+#define MUL2(in0, in1, in2, in3, out0, out1) \
+  {                                          \
+    out0 = in0 * in1;                        \
+    out1 = in2 * in3;                        \
+  }
+#define MUL4(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3) \
+  {                                                                          \
+    MUL2(in0, in1, in2, in3, out0, out1);                                    \
+    MUL2(in4, in5, in6, in7, out2, out3);                                    \
+  }
+
+/* Description : Addition of 2 pairs of vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+   Details     : Each element in 'in0' is added to 'in1' and result is written
+                 to 'out0'.
+*/
+#define ADD2(in0, in1, in2, in3, out0, out1) \
+  {                                          \
+    out0 = in0 + in1;                        \
+    out1 = in2 + in3;                        \
+  }
+#define ADD4(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3) \
+  {                                                                          \
+    ADD2(in0, in1, in2, in3, out0, out1);                                    \
+    ADD2(in4, in5, in6, in7, out2, out3);                                    \
+  }
+
+/* Description : Subtraction of 2 pairs of vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1
+   Details     : Each element in 'in1' is subtracted from 'in0' and result is
+                 written to 'out0'.
+*/
+#define SUB2(in0, in1, in2, in3, out0, out1) \
+  {                                          \
+    out0 = in0 - in1;                        \
+    out1 = in2 - in3;                        \
+  }
+#define SUB4(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3) \
+  {                                                                          \
+    out0 = in0 - in1;                                                        \
+    out1 = in2 - in3;                                                        \
+    out2 = in4 - in5;                                                        \
+    out3 = in6 - in7;                                                        \
+  }
+
+/* Description : Sign extend halfword elements from right half of the vector
+   Arguments   : Input  - in    (halfword vector)
+                 Output - out   (sign extended word vector)
+                 Return Type - signed word
+   Details     : Sign bit of halfword elements from input vector 'in' is
+                 extracted and interleaved with same vector 'in0' to generate
+                 4 word elements keeping sign intact
+*/
+#define UNPCK_R_SH_SW(in, out)                    \
+  {                                               \
+    v8i16 sign_m;                                 \
+                                                  \
+    sign_m = __msa_clti_s_h((v8i16)in, 0);        \
+    out = (v4i32)__msa_ilvr_h(sign_m, (v8i16)in); \
+  }
+
+/* Description : Zero extend unsigned byte elements to halfword elements
+   Arguments   : Input   - in          (unsigned byte vector)
+                 Outputs - out0, out1  (unsigned  halfword vectors)
+                 Return Type - signed halfword
+   Details     : Zero extended right half of vector is returned in 'out0'
+                 Zero extended left half of vector is returned in 'out1'
+*/
+#define UNPCK_UB_SH(in, out0, out1)      \
+  {                                      \
+    v16i8 zero_m = { 0 };                \
+                                         \
+    ILVRL_B2_SH(zero_m, in, out0, out1); \
+  }
+
+/* Description : Sign extend halfword elements from input vector and return
+                 the result in pair of vectors
+   Arguments   : Input   - in            (halfword vector)
+                 Outputs - out0, out1   (sign extended word vectors)
+                 Return Type - signed word
+   Details     : Sign bit of halfword elements from input vector 'in' is
+                 extracted and interleaved right with same vector 'in0' to
+                 generate 4 signed word elements in 'out0'
+                 Then interleaved left with same vector 'in0' to
+                 generate 4 signed word elements in 'out1'
+*/
+#define UNPCK_SH_SW(in, out0, out1)       \
+  {                                       \
+    v8i16 tmp_m;                          \
+                                          \
+    tmp_m = __msa_clti_s_h((v8i16)in, 0); \
+    ILVRL_H2_SW(tmp_m, in, out0, out1);   \
+  }
+
+/* Description : Butterfly of 4 input vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1, out2, out3
+   Details     : Butterfly operation
+*/
+#define BUTTERFLY_4(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                             \
+    out0 = in0 + in3;                                           \
+    out1 = in1 + in2;                                           \
+                                                                \
+    out2 = in1 - in2;                                           \
+    out3 = in0 - in3;                                           \
+  }
+
+/* Description : Butterfly of 8 input vectors
+   Arguments   : Inputs  - in0 ...  in7
+                 Outputs - out0 .. out7
+   Details     : Butterfly operation
+*/
+#define BUTTERFLY_8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, \
+                    out3, out4, out5, out6, out7)                             \
+  {                                                                           \
+    out0 = in0 + in7;                                                         \
+    out1 = in1 + in6;                                                         \
+    out2 = in2 + in5;                                                         \
+    out3 = in3 + in4;                                                         \
+                                                                              \
+    out4 = in3 - in4;                                                         \
+    out5 = in2 - in5;                                                         \
+    out6 = in1 - in6;                                                         \
+    out7 = in0 - in7;                                                         \
+  }
+
+/* Description : Butterfly of 16 input vectors
+   Arguments   : Inputs  - in0 ...  in15
+                 Outputs - out0 .. out15
+   Details     : Butterfly operation
+*/
+#define BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10,  \
+                     in11, in12, in13, in14, in15, out0, out1, out2, out3,    \
+                     out4, out5, out6, out7, out8, out9, out10, out11, out12, \
+                     out13, out14, out15)                                     \
+  {                                                                           \
+    out0 = in0 + in15;                                                        \
+    out1 = in1 + in14;                                                        \
+    out2 = in2 + in13;                                                        \
+    out3 = in3 + in12;                                                        \
+    out4 = in4 + in11;                                                        \
+    out5 = in5 + in10;                                                        \
+    out6 = in6 + in9;                                                         \
+    out7 = in7 + in8;                                                         \
+                                                                              \
+    out8 = in7 - in8;                                                         \
+    out9 = in6 - in9;                                                         \
+    out10 = in5 - in10;                                                       \
+    out11 = in4 - in11;                                                       \
+    out12 = in3 - in12;                                                       \
+    out13 = in2 - in13;                                                       \
+    out14 = in1 - in14;                                                       \
+    out15 = in0 - in15;                                                       \
+  }
+
+/* Description : Transpose input 8x8 byte block
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - as per RTYPE
+*/
+#define TRANSPOSE8x8_UB(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0,   \
+                        out1, out2, out3, out4, out5, out6, out7)              \
+  {                                                                            \
+    v16i8 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                      \
+    v16i8 tmp4_m, tmp5_m, tmp6_m, tmp7_m;                                      \
+                                                                               \
+    ILVR_B4_SB(in2, in0, in3, in1, in6, in4, in7, in5, tmp0_m, tmp1_m, tmp2_m, \
+               tmp3_m);                                                        \
+    ILVRL_B2_SB(tmp1_m, tmp0_m, tmp4_m, tmp5_m);                               \
+    ILVRL_B2_SB(tmp3_m, tmp2_m, tmp6_m, tmp7_m);                               \
+    ILVRL_W2(RTYPE, tmp6_m, tmp4_m, out0, out2);                               \
+    ILVRL_W2(RTYPE, tmp7_m, tmp5_m, out4, out6);                               \
+    SLDI_B2_0(RTYPE, out0, out2, out1, out3, 8);                               \
+    SLDI_B2_0(RTYPE, out4, out6, out5, out7, 8);                               \
+  }
+#define TRANSPOSE8x8_UB_UB(...) TRANSPOSE8x8_UB(v16u8, __VA_ARGS__)
+
+/* Description : Transpose 16x8 block into 8x16 with byte elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7,
+                           in8, in9, in10, in11, in12, in13, in14, in15
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - unsigned byte
+*/
+#define TRANSPOSE16x8_UB_UB(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, \
+                            in10, in11, in12, in13, in14, in15, out0, out1,   \
+                            out2, out3, out4, out5, out6, out7)               \
+  {                                                                           \
+    v16u8 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                     \
+    v16u8 tmp4_m, tmp5_m, tmp6_m, tmp7_m;                                     \
+                                                                              \
+    ILVEV_D2_UB(in0, in8, in1, in9, out7, out6);                              \
+    ILVEV_D2_UB(in2, in10, in3, in11, out5, out4);                            \
+    ILVEV_D2_UB(in4, in12, in5, in13, out3, out2);                            \
+    ILVEV_D2_UB(in6, in14, in7, in15, out1, out0);                            \
+                                                                              \
+    tmp0_m = (v16u8)__msa_ilvev_b((v16i8)out6, (v16i8)out7);                  \
+    tmp4_m = (v16u8)__msa_ilvod_b((v16i8)out6, (v16i8)out7);                  \
+    tmp1_m = (v16u8)__msa_ilvev_b((v16i8)out4, (v16i8)out5);                  \
+    tmp5_m = (v16u8)__msa_ilvod_b((v16i8)out4, (v16i8)out5);                  \
+    out5 = (v16u8)__msa_ilvev_b((v16i8)out2, (v16i8)out3);                    \
+    tmp6_m = (v16u8)__msa_ilvod_b((v16i8)out2, (v16i8)out3);                  \
+    out7 = (v16u8)__msa_ilvev_b((v16i8)out0, (v16i8)out1);                    \
+    tmp7_m = (v16u8)__msa_ilvod_b((v16i8)out0, (v16i8)out1);                  \
+                                                                              \
+    ILVEV_H2_UB(tmp0_m, tmp1_m, out5, out7, tmp2_m, tmp3_m);                  \
+    out0 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);                \
+    out4 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);                \
+                                                                              \
+    tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp1_m, (v8i16)tmp0_m);              \
+    tmp3_m = (v16u8)__msa_ilvod_h((v8i16)out7, (v8i16)out5);                  \
+    out2 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);                \
+    out6 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);                \
+                                                                              \
+    ILVEV_H2_UB(tmp4_m, tmp5_m, tmp6_m, tmp7_m, tmp2_m, tmp3_m);              \
+    out1 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);                \
+    out5 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);                \
+                                                                              \
+    tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp5_m, (v8i16)tmp4_m);              \
+    tmp2_m = (v16u8)__msa_ilvod_h((v8i16)tmp5_m, (v8i16)tmp4_m);              \
+    tmp3_m = (v16u8)__msa_ilvod_h((v8i16)tmp7_m, (v8i16)tmp6_m);              \
+    tmp3_m = (v16u8)__msa_ilvod_h((v8i16)tmp7_m, (v8i16)tmp6_m);              \
+    out3 = (v16u8)__msa_ilvev_w((v4i32)tmp3_m, (v4i32)tmp2_m);                \
+    out7 = (v16u8)__msa_ilvod_w((v4i32)tmp3_m, (v4i32)tmp2_m);                \
+  }
+
+/* Description : Transpose 4x4 block with half word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1, out2, out3
+                 Return Type - signed halfword
+*/
+#define TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                                    \
+    v8i16 s0_m, s1_m;                                                  \
+                                                                       \
+    ILVR_H2_SH(in1, in0, in3, in2, s0_m, s1_m);                        \
+    ILVRL_W2_SH(s1_m, s0_m, out0, out2);                               \
+    out1 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out0);              \
+    out3 = (v8i16)__msa_ilvl_d((v2i64)out0, (v2i64)out2);              \
+  }
+
+/* Description : Transpose 4x8 block with half word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - signed halfword
+*/
+#define TRANSPOSE4X8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                           out2, out3, out4, out5, out6, out7)                 \
+  {                                                                            \
+    v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                      \
+    v8i16 tmp0_n, tmp1_n, tmp2_n, tmp3_n;                                      \
+    v8i16 zero_m = { 0 };                                                      \
+                                                                               \
+    ILVR_H4_SH(in1, in0, in3, in2, in5, in4, in7, in6, tmp0_n, tmp1_n, tmp2_n, \
+               tmp3_n);                                                        \
+    ILVRL_W2_SH(tmp1_n, tmp0_n, tmp0_m, tmp2_m);                               \
+    ILVRL_W2_SH(tmp3_n, tmp2_n, tmp1_m, tmp3_m);                               \
+                                                                               \
+    out0 = (v8i16)__msa_ilvr_d((v2i64)tmp1_m, (v2i64)tmp0_m);                  \
+    out1 = (v8i16)__msa_ilvl_d((v2i64)tmp1_m, (v2i64)tmp0_m);                  \
+    out2 = (v8i16)__msa_ilvr_d((v2i64)tmp3_m, (v2i64)tmp2_m);                  \
+    out3 = (v8i16)__msa_ilvl_d((v2i64)tmp3_m, (v2i64)tmp2_m);                  \
+                                                                               \
+    out4 = zero_m;                                                             \
+    out5 = zero_m;                                                             \
+    out6 = zero_m;                                                             \
+    out7 = zero_m;                                                             \
+  }
+
+/* Description : Transpose 8x4 block with half word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - signed halfword
+*/
+#define TRANSPOSE8X4_SH_SH(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                                    \
+    v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                              \
+                                                                       \
+    ILVR_H2_SH(in1, in0, in3, in2, tmp0_m, tmp1_m);                    \
+    ILVL_H2_SH(in1, in0, in3, in2, tmp2_m, tmp3_m);                    \
+    ILVR_W2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out0, out2);            \
+    ILVL_W2_SH(tmp1_m, tmp0_m, tmp3_m, tmp2_m, out1, out3);            \
+  }
+
+/* Description : Transpose 8x8 block with half word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3, in4, in5, in6, in7
+                 Outputs - out0, out1, out2, out3, out4, out5, out6, out7
+                 Return Type - as per RTYPE
+*/
+#define TRANSPOSE8x8_H(RTYPE, in0, in1, in2, in3, in4, in5, in6, in7, out0, \
+                       out1, out2, out3, out4, out5, out6, out7)            \
+  {                                                                         \
+    v8i16 s0_m, s1_m;                                                       \
+    v8i16 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                   \
+    v8i16 tmp4_m, tmp5_m, tmp6_m, tmp7_m;                                   \
+                                                                            \
+    ILVR_H2_SH(in6, in4, in7, in5, s0_m, s1_m);                             \
+    ILVRL_H2_SH(s1_m, s0_m, tmp0_m, tmp1_m);                                \
+    ILVL_H2_SH(in6, in4, in7, in5, s0_m, s1_m);                             \
+    ILVRL_H2_SH(s1_m, s0_m, tmp2_m, tmp3_m);                                \
+    ILVR_H2_SH(in2, in0, in3, in1, s0_m, s1_m);                             \
+    ILVRL_H2_SH(s1_m, s0_m, tmp4_m, tmp5_m);                                \
+    ILVL_H2_SH(in2, in0, in3, in1, s0_m, s1_m);                             \
+    ILVRL_H2_SH(s1_m, s0_m, tmp6_m, tmp7_m);                                \
+    PCKEV_D4(RTYPE, tmp0_m, tmp4_m, tmp1_m, tmp5_m, tmp2_m, tmp6_m, tmp3_m, \
+             tmp7_m, out0, out2, out4, out6);                               \
+    out1 = (RTYPE)__msa_pckod_d((v2i64)tmp0_m, (v2i64)tmp4_m);              \
+    out3 = (RTYPE)__msa_pckod_d((v2i64)tmp1_m, (v2i64)tmp5_m);              \
+    out5 = (RTYPE)__msa_pckod_d((v2i64)tmp2_m, (v2i64)tmp6_m);              \
+    out7 = (RTYPE)__msa_pckod_d((v2i64)tmp3_m, (v2i64)tmp7_m);              \
+  }
+#define TRANSPOSE8x8_SH_SH(...) TRANSPOSE8x8_H(v8i16, __VA_ARGS__)
+
+/* Description : Transpose 4x4 block with word elements in vectors
+   Arguments   : Inputs  - in0, in1, in2, in3
+                 Outputs - out0, out1, out2, out3
+                 Return Type - signed word
+*/
+#define TRANSPOSE4x4_SW_SW(in0, in1, in2, in3, out0, out1, out2, out3) \
+  {                                                                    \
+    v4i32 s0_m, s1_m, s2_m, s3_m;                                      \
+                                                                       \
+    ILVRL_W2_SW(in1, in0, s0_m, s1_m);                                 \
+    ILVRL_W2_SW(in3, in2, s2_m, s3_m);                                 \
+                                                                       \
+    out0 = (v4i32)__msa_ilvr_d((v2i64)s2_m, (v2i64)s0_m);              \
+    out1 = (v4i32)__msa_ilvl_d((v2i64)s2_m, (v2i64)s0_m);              \
+    out2 = (v4i32)__msa_ilvr_d((v2i64)s3_m, (v2i64)s1_m);              \
+    out3 = (v4i32)__msa_ilvl_d((v2i64)s3_m, (v2i64)s1_m);              \
+  }
+
+/* Description : Add block 4x4
+   Arguments   : Inputs - in0, in1, in2, in3, pdst, stride
+   Details     : Least significant 4 bytes from each input vector are added to
+                 the destination bytes, clipped between 0-255 and stored.
+*/
+#define ADDBLK_ST4x4_UB(in0, in1, in2, in3, pdst, stride)        \
+  {                                                              \
+    uint32_t src0_m, src1_m, src2_m, src3_m;                     \
+    v8i16 inp0_m, inp1_m, res0_m, res1_m;                        \
+    v16i8 dst0_m = { 0 };                                        \
+    v16i8 dst1_m = { 0 };                                        \
+    v16i8 zero_m = { 0 };                                        \
+                                                                 \
+    ILVR_D2_SH(in1, in0, in3, in2, inp0_m, inp1_m)               \
+    LW4(pdst, stride, src0_m, src1_m, src2_m, src3_m);           \
+    INSERT_W2_SB(src0_m, src1_m, dst0_m);                        \
+    INSERT_W2_SB(src2_m, src3_m, dst1_m);                        \
+    ILVR_B2_SH(zero_m, dst0_m, zero_m, dst1_m, res0_m, res1_m);  \
+    ADD2(res0_m, inp0_m, res1_m, inp1_m, res0_m, res1_m);        \
+    CLIP_SH2_0_255(res0_m, res1_m);                              \
+    PCKEV_B2_SB(res0_m, res0_m, res1_m, res1_m, dst0_m, dst1_m); \
+    ST4x4_UB(dst0_m, dst1_m, 0, 1, 0, 1, pdst, stride);          \
+  }
+
+/* Description : Pack even elements of input vectors & xor with 128
+   Arguments   : Inputs - in0, in1
+                 Output - out_m
+                 Return Type - unsigned byte
+   Details     : Signed byte even elements from 'in0' and 'in1' are packed
+                 together in one vector and the resulting vector is xor'ed with
+                 128 to shift the range from signed to unsigned byte
+*/
+#define PCKEV_XORI128_UB(in0, in1)                        \
+  ({                                                      \
+    v16u8 out_m;                                          \
+                                                          \
+    out_m = (v16u8)__msa_pckev_b((v16i8)in1, (v16i8)in0); \
+    out_m = (v16u8)__msa_xori_b((v16u8)out_m, 128);       \
+    out_m;                                                \
+  })
+
+/* Description : Converts inputs to unsigned bytes, interleave, average & store
+                 as 8x4 unsigned byte block
+   Arguments   : Inputs - in0, in1, in2, in3, dst0, dst1, dst2, dst3,
+                          pdst, stride
+*/
+#define CONVERT_UB_AVG_ST8x4_UB(in0, in1, in2, in3, dst0, dst1, dst2, dst3, \
+                                pdst, stride)                               \
+  {                                                                         \
+    v16u8 tmp0_m, tmp1_m, tmp2_m, tmp3_m;                                   \
+                                                                            \
+    tmp0_m = PCKEV_XORI128_UB(in0, in1);                                    \
+    tmp1_m = PCKEV_XORI128_UB(in2, in3);                                    \
+    ILVR_D2_UB(dst1, dst0, dst3, dst2, tmp2_m, tmp3_m);                     \
+    AVER_UB2_UB(tmp0_m, tmp2_m, tmp1_m, tmp3_m, tmp0_m, tmp1_m);            \
+    ST8x4_UB(tmp0_m, tmp1_m, pdst, stride);                                 \
+  }
+
+/* Description : Pack even byte elements and store byte vector in destination
+                 memory
+   Arguments   : Inputs - in0, in1, pdst
+*/
+#define PCKEV_ST_SB(in0, in1, pdst)                \
+  {                                                \
+    v16i8 tmp_m;                                   \
+                                                   \
+    tmp_m = __msa_pckev_b((v16i8)in1, (v16i8)in0); \
+    ST_SB(tmp_m, (pdst));                          \
+  }
+
+/* Description : Horizontal 2 tap filter kernel code
+   Arguments   : Inputs - in0, in1, mask, coeff, shift
+*/
+#define HORIZ_2TAP_FILT_UH(in0, in1, mask, coeff, shift)        \
+  ({                                                            \
+    v16i8 tmp0_m;                                               \
+    v8u16 tmp1_m;                                               \
+                                                                \
+    tmp0_m = __msa_vshf_b((v16i8)mask, (v16i8)in1, (v16i8)in0); \
+    tmp1_m = __msa_dotp_u_h((v16u8)tmp0_m, (v16u8)coeff);       \
+    tmp1_m = (v8u16)__msa_srari_h((v8i16)tmp1_m, shift);        \
+                                                                \
+    tmp1_m;                                                     \
+  })
+#endif /* AOM_DSP_MIPS_MACROS_MSA_H_ */
diff --git a/third_party/aom/aom_dsp/mips/sad_msa.c b/third_party/aom/aom_dsp/mips/sad_msa.c
new file mode 100644
index 0000000000..258eb5c079
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/sad_msa.c
@@ -0,0 +1,1529 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/macros_msa.h"
+
+#define SAD_INSVE_W4(RTYPE, in0, in1, in2, in3, out)       \
+  {                                                        \
+    out = (RTYPE)__msa_insve_w((v4i32)out, 0, (v4i32)in0); \
+    out = (RTYPE)__msa_insve_w((v4i32)out, 1, (v4i32)in1); \
+    out = (RTYPE)__msa_insve_w((v4i32)out, 2, (v4i32)in2); \
+    out = (RTYPE)__msa_insve_w((v4i32)out, 3, (v4i32)in3); \
+  }
+#define SAD_INSVE_W4_UB(...) SAD_INSVE_W4(v16u8, __VA_ARGS__)
+
+static uint32_t sad_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 diff;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad += __msa_hadd_u_h(diff, diff);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_8width_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2, src0, src1,
+                ref0, ref1);
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_16width_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *ref, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(src, src_stride, src0, src1);
+    src += (2 * src_stride);
+    LD_UB2(ref, ref_stride, ref0, ref1);
+    ref += (2 * ref_stride);
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, src_stride, src0, src1);
+    src += (2 * src_stride);
+    LD_UB2(ref, ref_stride, ref0, ref1);
+    ref += (2 * ref_stride);
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_32width_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *ref, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB2(ref, 16, ref0, ref1);
+    ref += ref_stride;
+    sad += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t sad_64width_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *ref, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  uint32_t sad = 0;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+
+  for (ht_cnt = (height >> 1); ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad = HADD_UH_U32(sad0);
+  sad += HADD_UH_U32(sad1);
+
+  return sad;
+}
+
+static void sad_4width_x3_msa(const uint8_t *src_ptr, int32_t src_stride,
+                              const uint8_t *ref_ptr, int32_t ref_stride,
+                              int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 ref0, ref1, ref2, ref3, diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+
+    LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+}
+
+static void sad_8width_x3_msa(const uint8_t *src, int32_t src_stride,
+                              const uint8_t *ref, int32_t ref_stride,
+                              int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref00, ref11, ref22, ref33;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref00, ref11, ref22, ref33);
+    ref += (4 * ref_stride);
+    PCKEV_D4_UB(src1, src0, src3, src2, ref11, ref00, ref33, ref22, src0, src1,
+                ref0, ref1);
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+}
+
+static void sad_16width_x3_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src, ref, ref0, ref1, diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+
+  for (ht_cnt = (height >> 1); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+}
+
+static void sad_32width_x3_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0_0, ref0_1, ref0_2, ref0, ref1;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+
+  for (ht_cnt = height >> 1; ht_cnt--;) {
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB3(ref, 16, ref0_0, ref0_1, ref0_2);
+    ref += ref_stride;
+
+    sad0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB3(ref, 16, ref0_0, ref0_1, ref0_2);
+    ref += ref_stride;
+
+    sad0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+}
+
+static void sad_64width_x3_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0_0, ref0_1, ref0_2, ref0_3, ref0_4, ref0, ref1, ref2, ref3;
+  v8u16 sad0_0 = { 0 };
+  v8u16 sad0_1 = { 0 };
+  v8u16 sad1_0 = { 0 };
+  v8u16 sad1_1 = { 0 };
+  v8u16 sad2_0 = { 0 };
+  v8u16 sad2_1 = { 0 };
+  v4u32 sad;
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0_0, ref0_1, ref0_2, ref0_3);
+    ref0_4 = LD_UB(ref + 64);
+    ref += ref_stride;
+
+    sad0_0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
+    sad0_1 += SAD_UB2_UH(src2, src3, ref0_2, ref0_3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 1);
+    sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 2);
+    sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad = __msa_hadd_u_w(sad0_0, sad0_0);
+  sad += __msa_hadd_u_w(sad0_1, sad0_1);
+  sad_array[0] = HADD_SW_S32((v4i32)sad);
+
+  sad = __msa_hadd_u_w(sad1_0, sad1_0);
+  sad += __msa_hadd_u_w(sad1_1, sad1_1);
+  sad_array[1] = HADD_SW_S32((v4i32)sad);
+
+  sad = __msa_hadd_u_w(sad2_0, sad2_0);
+  sad += __msa_hadd_u_w(sad2_1, sad2_1);
+  sad_array[2] = HADD_SW_S32((v4i32)sad);
+}
+
+static void sad_4width_x8_msa(const uint8_t *src_ptr, int32_t src_stride,
+                              const uint8_t *ref_ptr, int32_t ref_stride,
+                              int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3, diff;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+  v8u16 sad4 = { 0 };
+  v8u16 sad5 = { 0 };
+  v8u16 sad6 = { 0 };
+  v8u16 sad7 = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad3 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad4 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad5 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad6 += __msa_hadd_u_h(diff, diff);
+
+    SLDI_B2_UB(ref0, ref1, ref0, ref1, ref0, ref1, 1);
+    SLDI_B2_UB(ref2, ref3, ref2, ref3, ref2, ref3, 1);
+    SAD_INSVE_W4_UB(ref0, ref1, ref2, ref3, ref);
+    diff = __msa_asub_u_b(src, ref);
+    sad7 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+  sad_array[4] = HADD_UH_U32(sad4);
+  sad_array[5] = HADD_UH_U32(sad5);
+  sad_array[6] = HADD_UH_U32(sad6);
+  sad_array[7] = HADD_UH_U32(sad7);
+}
+
+static void sad_8width_x8_msa(const uint8_t *src, int32_t src_stride,
+                              const uint8_t *ref, int32_t ref_stride,
+                              int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref00, ref11, ref22, ref33;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+  v8u16 sad4 = { 0 };
+  v8u16 sad5 = { 0 };
+  v8u16 sad6 = { 0 };
+  v8u16 sad7 = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref00, ref11, ref22, ref33);
+    ref += (4 * ref_stride);
+    PCKEV_D4_UB(src1, src0, src3, src2, ref11, ref00, ref33, ref22, src0, src1,
+                ref0, ref1);
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad4 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad5 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad6 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref00, ref11, ref00, ref11, ref00, ref11, 1);
+    SLDI_B2_UB(ref22, ref33, ref22, ref33, ref22, ref33, 1);
+    PCKEV_D2_UB(ref11, ref00, ref33, ref22, ref0, ref1);
+    sad7 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+  sad_array[4] = HADD_UH_U32(sad4);
+  sad_array[5] = HADD_UH_U32(sad5);
+  sad_array[6] = HADD_UH_U32(sad6);
+  sad_array[7] = HADD_UH_U32(sad7);
+}
+
+static void sad_16width_x8_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src, ref0, ref1, ref;
+  v16u8 diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+  v8u16 sad4 = { 0 };
+  v8u16 sad5 = { 0 };
+  v8u16 sad6 = { 0 };
+  v8u16 sad7 = { 0 };
+
+  for (ht_cnt = (height >> 1); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 3);
+    diff = __msa_asub_u_b(src, ref);
+    sad3 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 4);
+    diff = __msa_asub_u_b(src, ref);
+    sad4 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 5);
+    diff = __msa_asub_u_b(src, ref);
+    sad5 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 6);
+    diff = __msa_asub_u_b(src, ref);
+    sad6 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 7);
+    diff = __msa_asub_u_b(src, ref);
+    sad7 += __msa_hadd_u_h(diff, diff);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 1);
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 2);
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 3);
+    diff = __msa_asub_u_b(src, ref);
+    sad3 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 4);
+    diff = __msa_asub_u_b(src, ref);
+    sad4 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 5);
+    diff = __msa_asub_u_b(src, ref);
+    sad5 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 6);
+    diff = __msa_asub_u_b(src, ref);
+    sad6 += __msa_hadd_u_h(diff, diff);
+
+    ref = (v16u8)__msa_sldi_b((v16i8)ref1, (v16i8)ref0, 7);
+    diff = __msa_asub_u_b(src, ref);
+    sad7 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+  sad_array[4] = HADD_UH_U32(sad4);
+  sad_array[5] = HADD_UH_U32(sad5);
+  sad_array[6] = HADD_UH_U32(sad6);
+  sad_array[7] = HADD_UH_U32(sad7);
+}
+
+static void sad_32width_x8_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  int32_t ht_cnt;
+  v16u8 src0, src1;
+  v16u8 ref0, ref1, ref0_0, ref0_1, ref0_2;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+  v8u16 sad4 = { 0 };
+  v8u16 sad5 = { 0 };
+  v8u16 sad6 = { 0 };
+  v8u16 sad7 = { 0 };
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_UB3(ref, 16, ref0_0, ref0_1, ref0_2);
+    ref += ref_stride;
+
+    sad0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 3);
+    sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 4);
+    sad4 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 5);
+    sad5 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 6);
+    sad6 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 7);
+    sad7 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+  sad_array[4] = HADD_UH_U32(sad4);
+  sad_array[5] = HADD_UH_U32(sad5);
+  sad_array[6] = HADD_UH_U32(sad6);
+  sad_array[7] = HADD_UH_U32(sad7);
+}
+
+static void sad_64width_x8_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               int32_t height, uint32_t *sad_array) {
+  const uint8_t *src_dup, *ref_dup;
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0_0, ref0_1, ref0_2, ref0_3, ref0_4;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 sad0_0 = { 0 };
+  v8u16 sad0_1 = { 0 };
+  v8u16 sad1_0 = { 0 };
+  v8u16 sad1_1 = { 0 };
+  v8u16 sad2_0 = { 0 };
+  v8u16 sad2_1 = { 0 };
+  v8u16 sad3_0 = { 0 };
+  v8u16 sad3_1 = { 0 };
+  v4u32 sad;
+
+  src_dup = src;
+  ref_dup = ref;
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB5(ref, 16, ref0_0, ref0_1, ref0_2, ref0_3, ref0_4);
+    ref += ref_stride;
+
+    sad0_0 += SAD_UB2_UH(src0, src1, ref0_0, ref0_1);
+    sad0_1 += SAD_UB2_UH(src2, src3, ref0_2, ref0_3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 1);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 1);
+    sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 2);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 2);
+    sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 3);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 3);
+    sad3_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad3_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad = __msa_hadd_u_w(sad0_0, sad0_0);
+  sad += __msa_hadd_u_w(sad0_1, sad0_1);
+  sad_array[0] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad1_0, sad1_0);
+  sad += __msa_hadd_u_w(sad1_1, sad1_1);
+  sad_array[1] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad2_0, sad2_0);
+  sad += __msa_hadd_u_w(sad2_1, sad2_1);
+  sad_array[2] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad3_0, sad3_0);
+  sad += __msa_hadd_u_w(sad3_1, sad3_1);
+  sad_array[3] = HADD_SW_S32(sad);
+
+  sad0_0 = (v8u16)__msa_ldi_h(0);
+  sad0_1 = (v8u16)__msa_ldi_h(0);
+  sad1_0 = (v8u16)__msa_ldi_h(0);
+  sad1_1 = (v8u16)__msa_ldi_h(0);
+  sad2_0 = (v8u16)__msa_ldi_h(0);
+  sad2_1 = (v8u16)__msa_ldi_h(0);
+  sad3_0 = (v8u16)__msa_ldi_h(0);
+  sad3_1 = (v8u16)__msa_ldi_h(0);
+
+  for (ht_cnt = 64; ht_cnt--;) {
+    LD_UB4(src_dup, 16, src0, src1, src2, src3);
+    src_dup += src_stride;
+    LD_UB5(ref_dup, 16, ref0_0, ref0_1, ref0_2, ref0_3, ref0_4);
+    ref_dup += ref_stride;
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 4);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 4);
+    sad0_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad0_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 5);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 5);
+    sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 6);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 6);
+    sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    SLDI_B2_UB(ref0_1, ref0_2, ref0_0, ref0_1, ref0, ref1, 7);
+    SLDI_B2_UB(ref0_3, ref0_4, ref0_2, ref0_3, ref2, ref3, 7);
+    sad3_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad3_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad = __msa_hadd_u_w(sad0_0, sad0_0);
+  sad += __msa_hadd_u_w(sad0_1, sad0_1);
+  sad_array[4] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad1_0, sad1_0);
+  sad += __msa_hadd_u_w(sad1_1, sad1_1);
+  sad_array[5] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad2_0, sad2_0);
+  sad += __msa_hadd_u_w(sad2_1, sad2_1);
+  sad_array[6] = HADD_SW_S32(sad);
+
+  sad = __msa_hadd_u_w(sad3_0, sad3_0);
+  sad += __msa_hadd_u_w(sad3_1, sad3_1);
+  sad_array[7] = HADD_SW_S32(sad);
+}
+
+static void sad_4width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *const aref_ptr[],
+                               int32_t ref_stride, int32_t height,
+                               uint32_t *sad_array) {
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    src_ptr += (4 * src_stride);
+
+    LW4(ref0_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref0_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad0 += __msa_hadd_u_h(diff, diff);
+
+    LW4(ref1_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref1_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad1 += __msa_hadd_u_h(diff, diff);
+
+    LW4(ref2_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref2_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad2 += __msa_hadd_u_h(diff, diff);
+
+    LW4(ref3_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ref3_ptr += (4 * ref_stride);
+
+    diff = __msa_asub_u_b(src, ref);
+    sad3 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_8width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *const aref_ptr[],
+                               int32_t ref_stride, int32_t height,
+                               uint32_t *sad_array) {
+  int32_t ht_cnt;
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
+  v16u8 ref8, ref9, ref10, ref11, ref12, ref13, ref14, ref15;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref0_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref0_ptr += (4 * ref_stride);
+    LD_UB4(ref1_ptr, ref_stride, ref4, ref5, ref6, ref7);
+    ref1_ptr += (4 * ref_stride);
+    LD_UB4(ref2_ptr, ref_stride, ref8, ref9, ref10, ref11);
+    ref2_ptr += (4 * ref_stride);
+    LD_UB4(ref3_ptr, ref_stride, ref12, ref13, ref14, ref15);
+    ref3_ptr += (4 * ref_stride);
+
+    PCKEV_D2_UB(src1, src0, src3, src2, src0, src1);
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    PCKEV_D2_UB(ref5, ref4, ref7, ref6, ref0, ref1);
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    PCKEV_D2_UB(ref9, ref8, ref11, ref10, ref0, ref1);
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    PCKEV_D2_UB(ref13, ref12, ref15, ref14, ref0, ref1);
+    sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_16width_x4d_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                const uint8_t *const aref_ptr[],
+                                int32_t ref_stride, int32_t height,
+                                uint32_t *sad_array) {
+  int32_t ht_cnt;
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  v16u8 src, ref0, ref1, ref2, ref3, diff;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = (height >> 1); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref0 = LD_UB(ref0_ptr);
+    ref0_ptr += ref_stride;
+    ref1 = LD_UB(ref1_ptr);
+    ref1_ptr += ref_stride;
+    ref2 = LD_UB(ref2_ptr);
+    ref2_ptr += ref_stride;
+    ref3 = LD_UB(ref3_ptr);
+    ref3_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref1);
+    sad1 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref2);
+    sad2 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref3);
+    sad3 += __msa_hadd_u_h(diff, diff);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref0 = LD_UB(ref0_ptr);
+    ref0_ptr += ref_stride;
+    ref1 = LD_UB(ref1_ptr);
+    ref1_ptr += ref_stride;
+    ref2 = LD_UB(ref2_ptr);
+    ref2_ptr += ref_stride;
+    ref3 = LD_UB(ref3_ptr);
+    ref3_ptr += ref_stride;
+
+    diff = __msa_asub_u_b(src, ref0);
+    sad0 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref1);
+    sad1 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref2);
+    sad2 += __msa_hadd_u_h(diff, diff);
+    diff = __msa_asub_u_b(src, ref3);
+    sad3 += __msa_hadd_u_h(diff, diff);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_32width_x4d_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *const aref_ptr[],
+                                int32_t ref_stride, int32_t height,
+                                uint32_t *sad_array) {
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v8u16 sad2 = { 0 };
+  v8u16 sad3 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB2(src, 16, src0, src1);
+    src += src_stride;
+
+    LD_UB2(ref0_ptr, 16, ref0, ref1);
+    ref0_ptr += ref_stride;
+    sad0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(ref1_ptr, 16, ref0, ref1);
+    ref1_ptr += ref_stride;
+    sad1 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(ref2_ptr, 16, ref0, ref1);
+    ref2_ptr += ref_stride;
+    sad2 += SAD_UB2_UH(src0, src1, ref0, ref1);
+
+    LD_UB2(ref3_ptr, 16, ref0, ref1);
+    ref3_ptr += ref_stride;
+    sad3 += SAD_UB2_UH(src0, src1, ref0, ref1);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0);
+  sad_array[1] = HADD_UH_U32(sad1);
+  sad_array[2] = HADD_UH_U32(sad2);
+  sad_array[3] = HADD_UH_U32(sad3);
+}
+
+static void sad_64width_x4d_msa(const uint8_t *src, int32_t src_stride,
+                                const uint8_t *const aref_ptr[],
+                                int32_t ref_stride, int32_t height,
+                                uint32_t *sad_array) {
+  const uint8_t *ref0_ptr, *ref1_ptr, *ref2_ptr, *ref3_ptr;
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 sad0_0 = { 0 };
+  v8u16 sad0_1 = { 0 };
+  v8u16 sad1_0 = { 0 };
+  v8u16 sad1_1 = { 0 };
+  v8u16 sad2_0 = { 0 };
+  v8u16 sad2_1 = { 0 };
+  v8u16 sad3_0 = { 0 };
+  v8u16 sad3_1 = { 0 };
+
+  ref0_ptr = aref_ptr[0];
+  ref1_ptr = aref_ptr[1];
+  ref2_ptr = aref_ptr[2];
+  ref3_ptr = aref_ptr[3];
+
+  for (ht_cnt = height; ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+
+    LD_UB4(ref0_ptr, 16, ref0, ref1, ref2, ref3);
+    ref0_ptr += ref_stride;
+    sad0_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad0_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(ref1_ptr, 16, ref0, ref1, ref2, ref3);
+    ref1_ptr += ref_stride;
+    sad1_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad1_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(ref2_ptr, 16, ref0, ref1, ref2, ref3);
+    ref2_ptr += ref_stride;
+    sad2_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad2_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+
+    LD_UB4(ref3_ptr, 16, ref0, ref1, ref2, ref3);
+    ref3_ptr += ref_stride;
+    sad3_0 += SAD_UB2_UH(src0, src1, ref0, ref1);
+    sad3_1 += SAD_UB2_UH(src2, src3, ref2, ref3);
+  }
+
+  sad_array[0] = HADD_UH_U32(sad0_0);
+  sad_array[0] += HADD_UH_U32(sad0_1);
+  sad_array[1] = HADD_UH_U32(sad1_0);
+  sad_array[1] += HADD_UH_U32(sad1_1);
+  sad_array[2] = HADD_UH_U32(sad2_0);
+  sad_array[2] += HADD_UH_U32(sad2_1);
+  sad_array[3] = HADD_UH_U32(sad3_0);
+  sad_array[3] += HADD_UH_U32(sad3_1);
+}
+
+static uint32_t avgsad_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                  const uint8_t *ref_ptr, int32_t ref_stride,
+                                  int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v16u8 diff, pred, comp;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+
+    comp = __msa_aver_u_b(pred, ref);
+    diff = __msa_asub_u_b(src, comp);
+    sad += __msa_hadd_u_h(diff, diff);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_8width_msa(const uint8_t *src, int32_t src_stride,
+                                  const uint8_t *ref, int32_t ref_stride,
+                                  int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 diff0, diff1, pred0, pred1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2, src0, src1,
+                ref0, ref1);
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, diff0, diff1);
+    sad += SAD_UB2_UH(src0, src1, diff0, diff1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_16width_msa(const uint8_t *src, int32_t src_stride,
+                                   const uint8_t *ref, int32_t ref_stride,
+                                   int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3, comp0, comp1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 3); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * 16);
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1);
+    sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1);
+    sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+
+    LD_UB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(ref, ref_stride, ref0, ref1, ref2, ref3);
+    ref += (4 * ref_stride);
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * 16);
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1);
+    sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1);
+    sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_32width_msa(const uint8_t *src, int32_t src_stride,
+                                   const uint8_t *ref, int32_t ref_stride,
+                                   int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 ref0, ref1, ref2, ref3, ref4, ref5, ref6, ref7;
+  v16u8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7;
+  v16u8 comp0, comp1;
+  v8u16 sad = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, src_stride, src0, src2, src4, src6);
+    LD_UB4(src + 16, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+
+    LD_UB4(ref, ref_stride, ref0, ref2, ref4, ref6);
+    LD_UB4(ref + 16, ref_stride, ref1, ref3, ref5, ref7);
+    ref += (4 * ref_stride);
+
+    LD_UB4(sec_pred, 32, pred0, pred2, pred4, pred6);
+    LD_UB4(sec_pred + 16, 32, pred1, pred3, pred5, pred7);
+    sec_pred += (4 * 32);
+
+    AVER_UB2_UB(pred0, ref0, pred1, ref1, comp0, comp1);
+    sad += SAD_UB2_UH(src0, src1, comp0, comp1);
+    AVER_UB2_UB(pred2, ref2, pred3, ref3, comp0, comp1);
+    sad += SAD_UB2_UH(src2, src3, comp0, comp1);
+    AVER_UB2_UB(pred4, ref4, pred5, ref5, comp0, comp1);
+    sad += SAD_UB2_UH(src4, src5, comp0, comp1);
+    AVER_UB2_UB(pred6, ref6, pred7, ref7, comp0, comp1);
+    sad += SAD_UB2_UH(src6, src7, comp0, comp1);
+  }
+
+  return HADD_UH_U32(sad);
+}
+
+static uint32_t avgsad_64width_msa(const uint8_t *src, int32_t src_stride,
+                                   const uint8_t *ref, int32_t ref_stride,
+                                   int32_t height, const uint8_t *sec_pred) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 comp0, comp1, comp2, comp3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v8u16 sad0 = { 0 };
+  v8u16 sad1 = { 0 };
+  v4u32 sad;
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3, comp0,
+                comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3, comp0,
+                comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3, comp0,
+                comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+
+    LD_UB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_UB4(ref, 16, ref0, ref1, ref2, ref3);
+    ref += ref_stride;
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    AVER_UB4_UB(pred0, ref0, pred1, ref1, pred2, ref2, pred3, ref3, comp0,
+                comp1, comp2, comp3);
+    sad0 += SAD_UB2_UH(src0, src1, comp0, comp1);
+    sad1 += SAD_UB2_UH(src2, src3, comp2, comp3);
+  }
+
+  sad = __msa_hadd_u_w(sad0, sad0);
+  sad += __msa_hadd_u_w(sad1, sad1);
+
+  return HADD_SW_S32(sad);
+}
+
+#define AOM_SAD_4xHEIGHT_MSA(height)                                         \
+  uint32_t aom_sad4x##height##_msa(const uint8_t *src, int32_t src_stride,   \
+                                   const uint8_t *ref, int32_t ref_stride) { \
+    return sad_4width_msa(src, src_stride, ref, ref_stride, height);         \
+  }
+
+#define AOM_SAD_8xHEIGHT_MSA(height)                                         \
+  uint32_t aom_sad8x##height##_msa(const uint8_t *src, int32_t src_stride,   \
+                                   const uint8_t *ref, int32_t ref_stride) { \
+    return sad_8width_msa(src, src_stride, ref, ref_stride, height);         \
+  }
+
+#define AOM_SAD_16xHEIGHT_MSA(height)                                         \
+  uint32_t aom_sad16x##height##_msa(const uint8_t *src, int32_t src_stride,   \
+                                    const uint8_t *ref, int32_t ref_stride) { \
+    return sad_16width_msa(src, src_stride, ref, ref_stride, height);         \
+  }
+
+#define AOM_SAD_32xHEIGHT_MSA(height)                                         \
+  uint32_t aom_sad32x##height##_msa(const uint8_t *src, int32_t src_stride,   \
+                                    const uint8_t *ref, int32_t ref_stride) { \
+    return sad_32width_msa(src, src_stride, ref, ref_stride, height);         \
+  }
+
+#define AOM_SAD_64xHEIGHT_MSA(height)                                         \
+  uint32_t aom_sad64x##height##_msa(const uint8_t *src, int32_t src_stride,   \
+                                    const uint8_t *ref, int32_t ref_stride) { \
+    return sad_64width_msa(src, src_stride, ref, ref_stride, height);         \
+  }
+
+#define AOM_SAD_4xHEIGHTx3_MSA(height)                                   \
+  void aom_sad4x##height##x3_msa(const uint8_t *src, int32_t src_stride, \
+                                 const uint8_t *ref, int32_t ref_stride, \
+                                 uint32_t *sads) {                       \
+    sad_4width_x3_msa(src, src_stride, ref, ref_stride, height, sads);   \
+  }
+
+#define AOM_SAD_8xHEIGHTx3_MSA(height)                                   \
+  void aom_sad8x##height##x3_msa(const uint8_t *src, int32_t src_stride, \
+                                 const uint8_t *ref, int32_t ref_stride, \
+                                 uint32_t *sads) {                       \
+    sad_8width_x3_msa(src, src_stride, ref, ref_stride, height, sads);   \
+  }
+
+#define AOM_SAD_16xHEIGHTx3_MSA(height)                                   \
+  void aom_sad16x##height##x3_msa(const uint8_t *src, int32_t src_stride, \
+                                  const uint8_t *ref, int32_t ref_stride, \
+                                  uint32_t *sads) {                       \
+    sad_16width_x3_msa(src, src_stride, ref, ref_stride, height, sads);   \
+  }
+
+#define AOM_SAD_32xHEIGHTx3_MSA(height)                                   \
+  void aom_sad32x##height##x3_msa(const uint8_t *src, int32_t src_stride, \
+                                  const uint8_t *ref, int32_t ref_stride, \
+                                  uint32_t *sads) {                       \
+    sad_32width_x3_msa(src, src_stride, ref, ref_stride, height, sads);   \
+  }
+
+#define AOM_SAD_64xHEIGHTx3_MSA(height)                                   \
+  void aom_sad64x##height##x3_msa(const uint8_t *src, int32_t src_stride, \
+                                  const uint8_t *ref, int32_t ref_stride, \
+                                  uint32_t *sads) {                       \
+    sad_64width_x3_msa(src, src_stride, ref, ref_stride, height, sads);   \
+  }
+
+#define AOM_SAD_4xHEIGHTx8_MSA(height)                                   \
+  void aom_sad4x##height##x8_msa(const uint8_t *src, int32_t src_stride, \
+                                 const uint8_t *ref, int32_t ref_stride, \
+                                 uint32_t *sads) {                       \
+    sad_4width_x8_msa(src, src_stride, ref, ref_stride, height, sads);   \
+  }
+
+#define AOM_SAD_8xHEIGHTx8_MSA(height)                                   \
+  void aom_sad8x##height##x8_msa(const uint8_t *src, int32_t src_stride, \
+                                 const uint8_t *ref, int32_t ref_stride, \
+                                 uint32_t *sads) {                       \
+    sad_8width_x8_msa(src, src_stride, ref, ref_stride, height, sads);   \
+  }
+
+#define AOM_SAD_16xHEIGHTx8_MSA(height)                                   \
+  void aom_sad16x##height##x8_msa(const uint8_t *src, int32_t src_stride, \
+                                  const uint8_t *ref, int32_t ref_stride, \
+                                  uint32_t *sads) {                       \
+    sad_16width_x8_msa(src, src_stride, ref, ref_stride, height, sads);   \
+  }
+
+#define AOM_SAD_32xHEIGHTx8_MSA(height)                                   \
+  void aom_sad32x##height##x8_msa(const uint8_t *src, int32_t src_stride, \
+                                  const uint8_t *ref, int32_t ref_stride, \
+                                  uint32_t *sads) {                       \
+    sad_32width_x8_msa(src, src_stride, ref, ref_stride, height, sads);   \
+  }
+
+#define AOM_SAD_64xHEIGHTx8_MSA(height)                                   \
+  void aom_sad64x##height##x8_msa(const uint8_t *src, int32_t src_stride, \
+                                  const uint8_t *ref, int32_t ref_stride, \
+                                  uint32_t *sads) {                       \
+    sad_64width_x8_msa(src, src_stride, ref, ref_stride, height, sads);   \
+  }
+
+#define AOM_SAD_4xHEIGHTx4D_MSA(height)                                   \
+  void aom_sad4x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \
+                                  const uint8_t *const refs[],            \
+                                  int32_t ref_stride, uint32_t *sads) {   \
+    sad_4width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);  \
+  }
+
+#define AOM_SAD_8xHEIGHTx4D_MSA(height)                                   \
+  void aom_sad8x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \
+                                  const uint8_t *const refs[],            \
+                                  int32_t ref_stride, uint32_t *sads) {   \
+    sad_8width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);  \
+  }
+
+#define AOM_SAD_16xHEIGHTx4D_MSA(height)                                   \
+  void aom_sad16x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \
+                                   const uint8_t *const refs[],            \
+                                   int32_t ref_stride, uint32_t *sads) {   \
+    sad_16width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);  \
+  }
+
+#define AOM_SAD_32xHEIGHTx4D_MSA(height)                                   \
+  void aom_sad32x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \
+                                   const uint8_t *const refs[],            \
+                                   int32_t ref_stride, uint32_t *sads) {   \
+    sad_32width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);  \
+  }
+
+#define AOM_SAD_64xHEIGHTx4D_MSA(height)                                   \
+  void aom_sad64x##height##x4d_msa(const uint8_t *src, int32_t src_stride, \
+                                   const uint8_t *const refs[],            \
+                                   int32_t ref_stride, uint32_t *sads) {   \
+    sad_64width_x4d_msa(src, src_stride, refs, ref_stride, height, sads);  \
+  }
+
+#define AOM_AVGSAD_4xHEIGHT_MSA(height)                                        \
+  uint32_t aom_sad4x##height##_avg_msa(const uint8_t *src, int32_t src_stride, \
+                                       const uint8_t *ref, int32_t ref_stride, \
+                                       const uint8_t *second_pred) {           \
+    return avgsad_4width_msa(src, src_stride, ref, ref_stride, height,         \
+                             second_pred);                                     \
+  }
+
+#define AOM_AVGSAD_8xHEIGHT_MSA(height)                                        \
+  uint32_t aom_sad8x##height##_avg_msa(const uint8_t *src, int32_t src_stride, \
+                                       const uint8_t *ref, int32_t ref_stride, \
+                                       const uint8_t *second_pred) {           \
+    return avgsad_8width_msa(src, src_stride, ref, ref_stride, height,         \
+                             second_pred);                                     \
+  }
+
+#define AOM_AVGSAD_16xHEIGHT_MSA(height)                                \
+  uint32_t aom_sad16x##height##_avg_msa(                                \
+      const uint8_t *src, int32_t src_stride, const uint8_t *ref,       \
+      int32_t ref_stride, const uint8_t *second_pred) {                 \
+    return avgsad_16width_msa(src, src_stride, ref, ref_stride, height, \
+                              second_pred);                             \
+  }
+
+#define AOM_AVGSAD_32xHEIGHT_MSA(height)                                \
+  uint32_t aom_sad32x##height##_avg_msa(                                \
+      const uint8_t *src, int32_t src_stride, const uint8_t *ref,       \
+      int32_t ref_stride, const uint8_t *second_pred) {                 \
+    return avgsad_32width_msa(src, src_stride, ref, ref_stride, height, \
+                              second_pred);                             \
+  }
+
+#define AOM_AVGSAD_64xHEIGHT_MSA(height)                                \
+  uint32_t aom_sad64x##height##_avg_msa(                                \
+      const uint8_t *src, int32_t src_stride, const uint8_t *ref,       \
+      int32_t ref_stride, const uint8_t *second_pred) {                 \
+    return avgsad_64width_msa(src, src_stride, ref, ref_stride, height, \
+                              second_pred);                             \
+  }
+
+/* clang-format off */
+// 64x64
+AOM_SAD_64xHEIGHT_MSA(64)
+AOM_SAD_64xHEIGHTx3_MSA(64)
+AOM_SAD_64xHEIGHTx8_MSA(64)
+AOM_SAD_64xHEIGHTx4D_MSA(64)
+AOM_AVGSAD_64xHEIGHT_MSA(64)
+
+// 64x32
+AOM_SAD_64xHEIGHT_MSA(32)
+AOM_SAD_64xHEIGHTx3_MSA(32)
+AOM_SAD_64xHEIGHTx8_MSA(32)
+AOM_SAD_64xHEIGHTx4D_MSA(32)
+AOM_AVGSAD_64xHEIGHT_MSA(32)
+
+// 32x64
+AOM_SAD_32xHEIGHT_MSA(64)
+AOM_SAD_32xHEIGHTx3_MSA(64)
+AOM_SAD_32xHEIGHTx8_MSA(64)
+AOM_SAD_32xHEIGHTx4D_MSA(64)
+AOM_AVGSAD_32xHEIGHT_MSA(64)
+
+// 32x32
+AOM_SAD_32xHEIGHT_MSA(32)
+AOM_SAD_32xHEIGHTx3_MSA(32)
+AOM_SAD_32xHEIGHTx8_MSA(32)
+AOM_SAD_32xHEIGHTx4D_MSA(32)
+AOM_AVGSAD_32xHEIGHT_MSA(32)
+
+// 32x16
+AOM_SAD_32xHEIGHT_MSA(16)
+AOM_SAD_32xHEIGHTx3_MSA(16)
+AOM_SAD_32xHEIGHTx8_MSA(16)
+AOM_SAD_32xHEIGHTx4D_MSA(16)
+AOM_AVGSAD_32xHEIGHT_MSA(16)
+
+// 16x32
+AOM_SAD_16xHEIGHT_MSA(32)
+AOM_SAD_16xHEIGHTx3_MSA(32)
+AOM_SAD_16xHEIGHTx8_MSA(32)
+AOM_SAD_16xHEIGHTx4D_MSA(32)
+AOM_AVGSAD_16xHEIGHT_MSA(32)
+
+// 16x16
+AOM_SAD_16xHEIGHT_MSA(16)
+AOM_SAD_16xHEIGHTx3_MSA(16)
+AOM_SAD_16xHEIGHTx8_MSA(16)
+AOM_SAD_16xHEIGHTx4D_MSA(16)
+AOM_AVGSAD_16xHEIGHT_MSA(16)
+
+// 16x8
+AOM_SAD_16xHEIGHT_MSA(8)
+AOM_SAD_16xHEIGHTx3_MSA(8)
+AOM_SAD_16xHEIGHTx8_MSA(8)
+AOM_SAD_16xHEIGHTx4D_MSA(8)
+AOM_AVGSAD_16xHEIGHT_MSA(8)
+
+// 8x16
+AOM_SAD_8xHEIGHT_MSA(16)
+AOM_SAD_8xHEIGHTx3_MSA(16)
+AOM_SAD_8xHEIGHTx8_MSA(16)
+AOM_SAD_8xHEIGHTx4D_MSA(16)
+AOM_AVGSAD_8xHEIGHT_MSA(16)
+
+// 8x8
+AOM_SAD_8xHEIGHT_MSA(8)
+AOM_SAD_8xHEIGHTx3_MSA(8)
+AOM_SAD_8xHEIGHTx8_MSA(8)
+AOM_SAD_8xHEIGHTx4D_MSA(8)
+AOM_AVGSAD_8xHEIGHT_MSA(8)
+
+// 8x4
+AOM_SAD_8xHEIGHT_MSA(4)
+AOM_SAD_8xHEIGHTx3_MSA(4)
+AOM_SAD_8xHEIGHTx8_MSA(4)
+AOM_SAD_8xHEIGHTx4D_MSA(4)
+AOM_AVGSAD_8xHEIGHT_MSA(4)
+
+// 4x8
+AOM_SAD_4xHEIGHT_MSA(8)
+AOM_SAD_4xHEIGHTx3_MSA(8)
+AOM_SAD_4xHEIGHTx8_MSA(8)
+AOM_SAD_4xHEIGHTx4D_MSA(8)
+AOM_AVGSAD_4xHEIGHT_MSA(8)
+
+// 4x4
+AOM_SAD_4xHEIGHT_MSA(4)
+AOM_SAD_4xHEIGHTx3_MSA(4)
+AOM_SAD_4xHEIGHTx8_MSA(4)
+AOM_SAD_4xHEIGHTx4D_MSA(4)
+AOM_AVGSAD_4xHEIGHT_MSA(4)
+    /* clang-format on */
diff --git a/third_party/aom/aom_dsp/mips/sub_pixel_variance_msa.c b/third_party/aom/aom_dsp/mips/sub_pixel_variance_msa.c
new file mode 100644
index 0000000000..3eb85107d1
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/sub_pixel_variance_msa.c
@@ -0,0 +1,1795 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/mips/macros_msa.h"
+#include "aom_dsp/variance.h"
+
+static const uint8_t bilinear_filters_msa[8][2] = {
+  { 128, 0 }, { 112, 16 }, { 96, 32 }, { 80, 48 },
+  { 64, 64 }, { 48, 80 },  { 32, 96 }, { 16, 112 },
+};
+
+#define CALC_MSE_AVG_B(src, ref, var, sub)                          \
+  {                                                                 \
+    v16u8 src_l0_m, src_l1_m;                                       \
+    v8i16 res_l0_m, res_l1_m;                                       \
+                                                                    \
+    ILVRL_B2_UB(src, ref, src_l0_m, src_l1_m);                      \
+    HSUB_UB2_SH(src_l0_m, src_l1_m, res_l0_m, res_l1_m);            \
+    DPADD_SH2_SW(res_l0_m, res_l1_m, res_l0_m, res_l1_m, var, var); \
+                                                                    \
+    sub += res_l0_m + res_l1_m;                                     \
+  }
+
+#define VARIANCE_WxH(sse, diff, shift) sse - (((uint32_t)diff * diff) >> shift)
+
+#define VARIANCE_LARGE_WxH(sse, diff, shift) \
+  sse - (((int64_t)diff * diff) >> shift)
+
+static uint32_t avg_sse_diff_4width_msa(const uint8_t *src_ptr,
+                                        int32_t src_stride,
+                                        const uint8_t *ref_ptr,
+                                        int32_t ref_stride,
+                                        const uint8_t *sec_pred, int32_t height,
+                                        int32_t *diff) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 pred, src = { 0 };
+  v16u8 ref = { 0 };
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+
+    src = __msa_aver_u_b(src, pred);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_8width_msa(const uint8_t *src_ptr,
+                                        int32_t src_stride,
+                                        const uint8_t *ref_ptr,
+                                        int32_t ref_stride,
+                                        const uint8_t *sec_pred, int32_t height,
+                                        int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2, src0, src1,
+                ref0, ref1);
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_16width_msa(const uint8_t *src_ptr,
+                                         int32_t src_stride,
+                                         const uint8_t *ref_ptr,
+                                         int32_t ref_stride,
+                                         const uint8_t *sec_pred,
+                                         int32_t height, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src, ref, pred;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    src = __msa_aver_u_b(src, pred);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    src = __msa_aver_u_b(src, pred);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    src = __msa_aver_u_b(src, pred);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    src = __msa_aver_u_b(src, pred);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_32width_msa(const uint8_t *src_ptr,
+                                         int32_t src_stride,
+                                         const uint8_t *ref_ptr,
+                                         int32_t ref_stride,
+                                         const uint8_t *sec_pred,
+                                         int32_t height, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1, pred0, pred1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_32x64_msa(const uint8_t *src_ptr,
+                                       int32_t src_stride,
+                                       const uint8_t *ref_ptr,
+                                       int32_t ref_stride,
+                                       const uint8_t *sec_pred, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1, pred0, pred1;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 16; ht_cnt--;) {
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_64x32_msa(const uint8_t *src_ptr,
+                                       int32_t src_stride,
+                                       const uint8_t *ref_ptr,
+                                       int32_t ref_stride,
+                                       const uint8_t *sec_pred, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 16; ht_cnt--;) {
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3, src0, src1,
+                src2, src3);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src2, ref2, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src3, ref3, var, avg1);
+
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3, src0, src1,
+                src2, src3);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src2, ref2, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src3, ref3, var, avg1);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t avg_sse_diff_64x64_msa(const uint8_t *src_ptr,
+                                       int32_t src_stride,
+                                       const uint8_t *ref_ptr,
+                                       int32_t ref_stride,
+                                       const uint8_t *sec_pred, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v8i16 avg2 = { 0 };
+  v8i16 avg3 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 32; ht_cnt--;) {
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3, src0, src1,
+                src2, src3);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src2, ref2, var, avg2);
+    CALC_MSE_AVG_B(src3, ref3, var, avg3);
+
+    LD_UB4(sec_pred, 16, pred0, pred1, pred2, pred3);
+    sec_pred += 64;
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    AVER_UB4_UB(src0, pred0, src1, pred1, src2, pred2, src3, pred3, src0, src1,
+                src2, src3);
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src2, ref2, var, avg2);
+    CALC_MSE_AVG_B(src3, ref3, var, avg3);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+  vec += __msa_hadd_s_w(avg2, avg2);
+  vec += __msa_hadd_s_w(avg3, avg3);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_4width_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 filt0, ref = { 0 };
+  v16i8 src0, src1, src2, src3;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 vec0, vec1, vec2, vec3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3, src0, src1,
+                src2, src3);
+    ILVEV_W2_SB(src0, src1, src2, src3, src0, src2);
+    src0 = (v16i8)__msa_ilvev_d((v2i64)src2, (v2i64)src0);
+    CALC_MSE_AVG_B(src0, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_8width_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 filt0, out, ref0, ref1, ref2, ref3;
+  v16i8 src0, src1, src2, src3;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 vec0, vec1, vec2, vec3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3, src0, src1,
+                src2, src3);
+    out = (v16u8)__msa_ilvev_d((v2i64)src1, (v2i64)src0);
+    CALC_MSE_AVG_B(out, ref0, var, avg);
+    out = (v16u8)__msa_ilvev_d((v2i64)src3, (v2i64)src2);
+    CALC_MSE_AVG_B(out, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_16width_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v16u8 dst0, dst1, dst2, dst3, filt0;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    dst += (4 * dst_stride);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UH(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UH(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+                out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+                out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    PCKEV_B4_SB(out1, out0, out3, out2, out5, out4, out7, out6, src0, src1,
+                src2, src3);
+    CALC_MSE_AVG_B(src0, dst0, var, avg);
+    CALC_MSE_AVG_B(src1, dst1, var, avg);
+    CALC_MSE_AVG_B(src2, dst2, var, avg);
+    CALC_MSE_AVG_B(src3, dst3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_32width_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_h_msa(src, src_stride, dst, dst_stride,
+                                            filter, height, &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_sse_diff_64width_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_h_msa(src, src_stride, dst, dst_stride,
+                                            filter, height, &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_sse_diff_4width_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4, out;
+  v16u8 src10_r, src32_r, src21_r, src43_r;
+  v16u8 ref = { 0 };
+  v16u8 src2110, src4332;
+  v16u8 filt0;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+  v8u16 tmp0, tmp1;
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+               src32_r, src43_r);
+    ILVR_D2_UB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+    DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    CALC_MSE_AVG_B(out, ref, var, avg);
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_8width_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8u16 vec0, vec1, vec2, vec3;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 filt0;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    ILVR_B4_UH(src1, src0, src2, src1, src3, src2, src4, src3, vec0, vec1, vec2,
+               vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1,
+                tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_16width_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 out0, out1, out2, out3;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 filt0;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    ILVR_B2_UB(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UB(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    ILVR_B2_UB(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UB(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    out1 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    out3 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+
+    src0 = src4;
+
+    CALC_MSE_AVG_B(out0, ref0, var, avg);
+    CALC_MSE_AVG_B(out1, ref1, var, avg);
+    CALC_MSE_AVG_B(out2, ref2, var, avg);
+    CALC_MSE_AVG_B(out3, ref3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_32width_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_v_msa(src, src_stride, dst, dst_stride,
+                                            filter, height, &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_sse_diff_64width_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter, int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_v_msa(src, src_stride, dst, dst_stride,
+                                            filter, height, &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_sse_diff_4width_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter_horiz, const uint8_t *filter_vert,
+    int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 out, ref = { 0 };
+  v16u8 filt_vt, filt_hz, vec0, vec1;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20 };
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4;
+  v8u16 tmp0, tmp1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+    hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+    hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    CALC_MSE_AVG_B(out, ref, var, avg);
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_8width_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter_horiz, const uint8_t *filter_vert,
+    int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 out0, out1;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 hz_out0, hz_out1;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v16u8 filt_vt, filt_hz, vec0;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp1 = __msa_dotp_u_h(vec0, filt_vt);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp2 = __msa_dotp_u_h(vec0, filt_vt);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp3 = __msa_dotp_u_h(vec0, filt_vt);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+    CALC_MSE_AVG_B(out0, ref0, var, avg);
+    CALC_MSE_AVG_B(out1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_16width_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter_horiz, const uint8_t *filter_vert,
+    int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 filt_hz, filt_vt, vec0, vec1;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3;
+  v8u16 tmp0, tmp1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  LD_UB2(src, 8, src0, src1);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src0, src2, src4, src6);
+    LD_UB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    src0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    src1 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    src2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    src3 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+    CALC_MSE_AVG_B(src2, ref2, var, avg);
+    CALC_MSE_AVG_B(src3, ref3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_sse_diff_32width_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter_horiz, const uint8_t *filter_vert,
+    int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_hv_msa(src, src_stride, dst, dst_stride,
+                                             filter_horiz, filter_vert, height,
+                                             &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_sse_diff_64width_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *filter_horiz, const uint8_t *filter_vert,
+    int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse += sub_pixel_sse_diff_16width_hv_msa(src, src_stride, dst, dst_stride,
+                                             filter_horiz, filter_vert, height,
+                                             &diff0[loop_cnt]);
+    src += 16;
+    dst += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_4width_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 out, pred, filt0, ref = { 0 };
+  v16i8 src0, src1, src2, src3;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 vec0, vec1, vec2, vec3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3, src0, src1,
+                src2, src3);
+    ILVEV_W2_SB(src0, src1, src2, src3, src0, src2);
+    out = (v16u8)__msa_ilvev_d((v2i64)src2, (v2i64)src0);
+    out = __msa_aver_u_b(out, pred);
+    CALC_MSE_AVG_B(out, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_8width_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 out, pred, filt0;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16i8 src0, src1, src2, src3;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 vec0, vec1, vec2, vec3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src1, src2, src3);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, vec0, vec1,
+                vec2, vec3);
+    SRARI_H4_UH(vec0, vec1, vec2, vec3, FILTER_BITS);
+    PCKEV_B4_SB(vec0, vec0, vec1, vec1, vec2, vec2, vec3, vec3, src0, src1,
+                src2, src3);
+    out = (v16u8)__msa_ilvev_d((v2i64)src1, (v2i64)src0);
+
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    out = __msa_aver_u_b(out, pred);
+    CALC_MSE_AVG_B(out, ref0, var, avg);
+    out = (v16u8)__msa_ilvev_d((v2i64)src3, (v2i64)src2);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    out = __msa_aver_u_b(out, pred);
+    CALC_MSE_AVG_B(out, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t subpel_avg_ssediff_16w_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff, int32_t width) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16i8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v16u8 dst0, dst1, dst2, dst3;
+  v16u8 tmp0, tmp1, tmp2, tmp3;
+  v16u8 pred0, pred1, pred2, pred3, filt0;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 out0, out1, out2, out3, out4, out5, out6, out7;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_SB4(src, src_stride, src0, src2, src4, src6);
+    LD_SB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
+    dst += (4 * dst_stride);
+    LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * width);
+
+    VSHF_B2_UH(src0, src0, src1, src1, mask, mask, vec0, vec1);
+    VSHF_B2_UH(src2, src2, src3, src3, mask, mask, vec2, vec3);
+    VSHF_B2_UH(src4, src4, src5, src5, mask, mask, vec4, vec5);
+    VSHF_B2_UH(src6, src6, src7, src7, mask, mask, vec6, vec7);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, out0, out1,
+                out2, out3);
+    DOTP_UB4_UH(vec4, vec5, vec6, vec7, filt0, filt0, filt0, filt0, out4, out5,
+                out6, out7);
+    SRARI_H4_UH(out0, out1, out2, out3, FILTER_BITS);
+    SRARI_H4_UH(out4, out5, out6, out7, FILTER_BITS);
+    PCKEV_B4_UB(out1, out0, out3, out2, out5, out4, out7, out6, tmp0, tmp1,
+                tmp2, tmp3);
+    AVER_UB4_UB(tmp0, pred0, tmp1, pred1, tmp2, pred2, tmp3, pred3, tmp0, tmp1,
+                tmp2, tmp3);
+
+    CALC_MSE_AVG_B(tmp0, dst0, var, avg);
+    CALC_MSE_AVG_B(tmp1, dst1, var, avg);
+    CALC_MSE_AVG_B(tmp2, dst2, var, avg);
+    CALC_MSE_AVG_B(tmp3, dst3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_16width_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff) {
+  return subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride,
+                                      sec_pred, filter, height, diff, 16);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_32width_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse +=
+        subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride, sec_pred,
+                                     filter, height, &diff0[loop_cnt], 32);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_64width_h_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse +=
+        subpel_avg_ssediff_16w_h_msa(src, src_stride, dst, dst_stride, sec_pred,
+                                     filter, height, &diff0[loop_cnt], 64);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_4width_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 src10_r, src32_r, src21_r, src43_r;
+  v16u8 out, pred, ref = { 0 };
+  v16u8 src2110, src4332, filt0;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+  v8u16 tmp0, tmp1;
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    ILVR_B4_UB(src1, src0, src2, src1, src3, src2, src4, src3, src10_r, src21_r,
+               src32_r, src43_r);
+    ILVR_D2_UB(src21_r, src10_r, src43_r, src32_r, src2110, src4332);
+    DOTP_UB2_UH(src2110, src4332, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+
+    out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    out = __msa_aver_u_b(out, pred);
+    CALC_MSE_AVG_B(out, ref, var, avg);
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_8width_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, filt0;
+  v8u16 vec0, vec1, vec2, vec3;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    ILVR_B4_UH(src1, src0, src2, src1, src3, src2, src4, src3, vec0, vec1, vec2,
+               vec3);
+    DOTP_UB4_UH(vec0, vec1, vec2, vec3, filt0, filt0, filt0, filt0, tmp0, tmp1,
+                tmp2, tmp3);
+    SRARI_H4_UH(tmp0, tmp1, tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, src0, src1);
+    AVER_UB2_UB(src0, pred0, src1, pred1, src0, src1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t subpel_avg_ssediff_16w_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff, int32_t width) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 out0, out1, out2, out3, filt0;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 tmp0, tmp1, tmp2, tmp3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter);
+  filt0 = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * width);
+
+    ILVR_B2_UH(src1, src0, src2, src1, vec0, vec2);
+    ILVL_B2_UH(src1, src0, src2, src1, vec1, vec3);
+    DOTP_UB2_UH(vec0, vec1, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    ILVR_B2_UH(src3, src2, src4, src3, vec4, vec6);
+    ILVL_B2_UH(src3, src2, src4, src3, vec5, vec7);
+    DOTP_UB2_UH(vec2, vec3, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    out1 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+
+    DOTP_UB2_UH(vec4, vec5, filt0, filt0, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    DOTP_UB2_UH(vec6, vec7, filt0, filt0, tmp2, tmp3);
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    out3 = (v16u8)__msa_pckev_b((v16i8)tmp3, (v16i8)tmp2);
+
+    src0 = src4;
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    AVER_UB4_UB(out0, pred0, out1, pred1, out2, pred2, out3, pred3, out0, out1,
+                out2, out3);
+
+    CALC_MSE_AVG_B(out0, ref0, var, avg);
+    CALC_MSE_AVG_B(out1, ref1, var, avg);
+    CALC_MSE_AVG_B(out2, ref2, var, avg);
+    CALC_MSE_AVG_B(out3, ref3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_16width_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff) {
+  return subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride,
+                                      sec_pred, filter, height, diff, 16);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_32width_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse +=
+        subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride, sec_pred,
+                                     filter, height, &diff0[loop_cnt], 32);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_64width_v_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter,
+    int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse +=
+        subpel_avg_ssediff_16w_v_msa(src, src_stride, dst, dst_stride, sec_pred,
+                                     filter, height, &diff0[loop_cnt], 64);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_4width_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz,
+    const uint8_t *filter_vert, int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 16, 17, 17, 18, 18, 19, 19, 20 };
+  v16u8 filt_hz, filt_vt, vec0, vec1;
+  v16u8 out, pred, ref = { 0 };
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, tmp0, tmp1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    pred = LD_UB(sec_pred);
+    sec_pred += 16;
+    LW4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src0, src1, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src2, src3, mask, filt_hz, FILTER_BITS);
+    hz_out4 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out1 = (v8u16)__msa_sldi_b((v16i8)hz_out2, (v16i8)hz_out0, 8);
+    hz_out3 = (v8u16)__msa_pckod_d((v2i64)hz_out4, (v2i64)hz_out2);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+    out = __msa_aver_u_b(out, pred);
+    CALC_MSE_AVG_B(out, ref, var, avg);
+    src0 = src4;
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_8width_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz,
+    const uint8_t *filter_vert, int32_t height, int32_t *diff) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 src0, src1, src2, src3, src4;
+  v16u8 pred0, pred1, out0, out1;
+  v16u8 filt_hz, filt_vt, vec0;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 hz_out0, hz_out1, tmp0, tmp1, tmp2, tmp3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  src0 = LD_UB(src);
+  src += src_stride;
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src1, src2, src3, src4);
+    src += (4 * src_stride);
+    LD_UB2(sec_pred, 16, pred0, pred1);
+    sec_pred += 32;
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    PCKEV_D2_UB(ref1, ref0, ref3, ref2, ref0, ref1);
+    hz_out1 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp0 = __msa_dotp_u_h(vec0, filt_vt);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp1 = __msa_dotp_u_h(vec0, filt_vt);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    hz_out1 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out1, (v16i8)hz_out0);
+    tmp2 = __msa_dotp_u_h(vec0, filt_vt);
+    hz_out0 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)hz_out0, (v16i8)hz_out1);
+    tmp3 = __msa_dotp_u_h(vec0, filt_vt);
+
+    SRARI_H2_UH(tmp2, tmp3, FILTER_BITS);
+    PCKEV_B2_UB(tmp1, tmp0, tmp3, tmp2, out0, out1);
+    AVER_UB2_UB(out0, pred0, out1, pred1, out0, out1);
+
+    CALC_MSE_AVG_B(out0, ref0, var, avg);
+    CALC_MSE_AVG_B(out1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t subpel_avg_ssediff_16w_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz,
+    const uint8_t *filter_vert, int32_t height, int32_t *diff, int32_t width) {
+  int16_t filtval;
+  uint32_t loop_cnt;
+  v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16u8 ref0, ref1, ref2, ref3;
+  v16u8 pred0, pred1, pred2, pred3;
+  v16u8 out0, out1, out2, out3;
+  v16u8 filt_hz, filt_vt, vec0, vec1;
+  v16u8 mask = { 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8 };
+  v8u16 hz_out0, hz_out1, hz_out2, hz_out3, tmp0, tmp1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  filtval = LH(filter_horiz);
+  filt_hz = (v16u8)__msa_fill_h(filtval);
+  filtval = LH(filter_vert);
+  filt_vt = (v16u8)__msa_fill_h(filtval);
+
+  LD_UB2(src, 8, src0, src1);
+  src += src_stride;
+
+  hz_out0 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+  hz_out2 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+
+  for (loop_cnt = (height >> 2); loop_cnt--;) {
+    LD_UB4(src, src_stride, src0, src2, src4, src6);
+    LD_UB4(src + 8, src_stride, src1, src3, src5, src7);
+    src += (4 * src_stride);
+    LD_UB4(sec_pred, width, pred0, pred1, pred2, pred3);
+    sec_pred += (4 * width);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src0, src0, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src1, src1, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out0 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src2, src2, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src3, src3, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out1 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out1 = HORIZ_2TAP_FILT_UH(src4, src4, mask, filt_hz, FILTER_BITS);
+    hz_out3 = HORIZ_2TAP_FILT_UH(src5, src5, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out2 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    hz_out0 = HORIZ_2TAP_FILT_UH(src6, src6, mask, filt_hz, FILTER_BITS);
+    hz_out2 = HORIZ_2TAP_FILT_UH(src7, src7, mask, filt_hz, FILTER_BITS);
+    ILVEV_B2_UB(hz_out1, hz_out0, hz_out3, hz_out2, vec0, vec1);
+    DOTP_UB2_UH(vec0, vec1, filt_vt, filt_vt, tmp0, tmp1);
+    SRARI_H2_UH(tmp0, tmp1, FILTER_BITS);
+    out3 = (v16u8)__msa_pckev_b((v16i8)tmp1, (v16i8)tmp0);
+
+    LD_UB4(dst, dst_stride, ref0, ref1, ref2, ref3);
+    dst += (4 * dst_stride);
+
+    AVER_UB4_UB(out0, pred0, out1, pred1, out2, pred2, out3, pred3, out0, out1,
+                out2, out3);
+
+    CALC_MSE_AVG_B(out0, ref0, var, avg);
+    CALC_MSE_AVG_B(out1, ref1, var, avg);
+    CALC_MSE_AVG_B(out2, ref2, var, avg);
+    CALC_MSE_AVG_B(out3, ref3, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_16width_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz,
+    const uint8_t *filter_vert, int32_t height, int32_t *diff) {
+  return subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride,
+                                       sec_pred, filter_horiz, filter_vert,
+                                       height, diff, 16);
+}
+
+static uint32_t sub_pixel_avg_sse_diff_32width_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz,
+    const uint8_t *filter_vert, int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[2];
+
+  for (loop_cnt = 0; loop_cnt < 2; ++loop_cnt) {
+    sse += subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride,
+                                         sec_pred, filter_horiz, filter_vert,
+                                         height, &diff0[loop_cnt], 32);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1];
+
+  return sse;
+}
+
+static uint32_t sub_pixel_avg_sse_diff_64width_hv_msa(
+    const uint8_t *src, int32_t src_stride, const uint8_t *dst,
+    int32_t dst_stride, const uint8_t *sec_pred, const uint8_t *filter_horiz,
+    const uint8_t *filter_vert, int32_t height, int32_t *diff) {
+  uint32_t loop_cnt, sse = 0;
+  int32_t diff0[4];
+
+  for (loop_cnt = 0; loop_cnt < 4; ++loop_cnt) {
+    sse += subpel_avg_ssediff_16w_hv_msa(src, src_stride, dst, dst_stride,
+                                         sec_pred, filter_horiz, filter_vert,
+                                         height, &diff0[loop_cnt], 64);
+    src += 16;
+    dst += 16;
+    sec_pred += 16;
+  }
+
+  *diff = diff0[0] + diff0[1] + diff0[2] + diff0[3];
+
+  return sse;
+}
+
+#define VARIANCE_4Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 4);
+#define VARIANCE_4Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 5);
+#define VARIANCE_8Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 5);
+#define VARIANCE_8Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 6);
+#define VARIANCE_8Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 7);
+#define VARIANCE_16Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 7);
+#define VARIANCE_16Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 8);
+
+#define VARIANCE_16Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
+#define VARIANCE_32Wx16H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
+#define VARIANCE_32Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 10);
+#define VARIANCE_32Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
+#define VARIANCE_64Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
+#define VARIANCE_64Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 12);
+
+#define AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(wd, ht)                              \
+  uint32_t aom_sub_pixel_variance##wd##x##ht##_msa(                           \
+      const uint8_t *src, int32_t src_stride, int32_t xoffset,                \
+      int32_t yoffset, const uint8_t *ref, int32_t ref_stride,                \
+      uint32_t *sse) {                                                        \
+    int32_t diff;                                                             \
+    uint32_t var;                                                             \
+    const uint8_t *h_filter = bilinear_filters_msa[xoffset];                  \
+    const uint8_t *v_filter = bilinear_filters_msa[yoffset];                  \
+                                                                              \
+    if (yoffset) {                                                            \
+      if (xoffset) {                                                          \
+        *sse = sub_pixel_sse_diff_##wd##width_hv_msa(                         \
+            src, src_stride, ref, ref_stride, h_filter, v_filter, ht, &diff); \
+      } else {                                                                \
+        *sse = sub_pixel_sse_diff_##wd##width_v_msa(                          \
+            src, src_stride, ref, ref_stride, v_filter, ht, &diff);           \
+      }                                                                       \
+                                                                              \
+      var = VARIANCE_##wd##Wx##ht##H(*sse, diff);                             \
+    } else {                                                                  \
+      if (xoffset) {                                                          \
+        *sse = sub_pixel_sse_diff_##wd##width_h_msa(                          \
+            src, src_stride, ref, ref_stride, h_filter, ht, &diff);           \
+                                                                              \
+        var = VARIANCE_##wd##Wx##ht##H(*sse, diff);                           \
+      } else {                                                                \
+        var = aom_variance##wd##x##ht##_msa(src, src_stride, ref, ref_stride, \
+                                            sse);                             \
+      }                                                                       \
+    }                                                                         \
+                                                                              \
+    return var;                                                               \
+  }
+
+/* clang-format off */
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(4, 4)
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(4, 8)
+
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 4)
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 8)
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(8, 16)
+
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 8)
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 16)
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(16, 32)
+
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 16)
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 32)
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(32, 64)
+
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(64, 32)
+AOM_SUB_PIXEL_VARIANCE_WDXHT_MSA(64, 64)
+/* clang-format on */
+
+#define AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(wd, ht)                          \
+  uint32_t aom_sub_pixel_avg_variance##wd##x##ht##_msa(                       \
+      const uint8_t *src_ptr, int32_t src_stride, int32_t xoffset,            \
+      int32_t yoffset, const uint8_t *ref_ptr, int32_t ref_stride,            \
+      uint32_t *sse, const uint8_t *sec_pred) {                               \
+    int32_t diff;                                                             \
+    const uint8_t *h_filter = bilinear_filters_msa[xoffset];                  \
+    const uint8_t *v_filter = bilinear_filters_msa[yoffset];                  \
+                                                                              \
+    if (yoffset) {                                                            \
+      if (xoffset) {                                                          \
+        *sse = sub_pixel_avg_sse_diff_##wd##width_hv_msa(                     \
+            src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, h_filter,     \
+            v_filter, ht, &diff);                                             \
+      } else {                                                                \
+        *sse = sub_pixel_avg_sse_diff_##wd##width_v_msa(                      \
+            src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, v_filter, ht, \
+            &diff);                                                           \
+      }                                                                       \
+    } else {                                                                  \
+      if (xoffset) {                                                          \
+        *sse = sub_pixel_avg_sse_diff_##wd##width_h_msa(                      \
+            src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, h_filter, ht, \
+            &diff);                                                           \
+      } else {                                                                \
+        *sse = avg_sse_diff_##wd##width_msa(src_ptr, src_stride, ref_ptr,     \
+                                            ref_stride, sec_pred, ht, &diff); \
+      }                                                                       \
+    }                                                                         \
+                                                                              \
+    return VARIANCE_##wd##Wx##ht##H(*sse, diff);                              \
+  }
+
+/* clang-format off */
+AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(4, 4)
+AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(4, 8)
+
+AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 4)
+AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 8)
+AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(8, 16)
+
+AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 8)
+AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 16)
+AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(16, 32)
+
+AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(32, 16)
+AOM_SUB_PIXEL_AVG_VARIANCE_WDXHT_MSA(32, 32)
+/* clang-format on */
+
+uint32_t aom_sub_pixel_avg_variance32x64_msa(const uint8_t *src_ptr,
+                                             int32_t src_stride,
+                                             int32_t xoffset, int32_t yoffset,
+                                             const uint8_t *ref_ptr,
+                                             int32_t ref_stride, uint32_t *sse,
+                                             const uint8_t *sec_pred) {
+  int32_t diff;
+  const uint8_t *h_filter = bilinear_filters_msa[xoffset];
+  const uint8_t *v_filter = bilinear_filters_msa[yoffset];
+
+  if (yoffset) {
+    if (xoffset) {
+      *sse = sub_pixel_avg_sse_diff_32width_hv_msa(
+          src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, h_filter,
+          v_filter, 64, &diff);
+    } else {
+      *sse = sub_pixel_avg_sse_diff_32width_v_msa(src_ptr, src_stride, ref_ptr,
+                                                  ref_stride, sec_pred,
+                                                  v_filter, 64, &diff);
+    }
+  } else {
+    if (xoffset) {
+      *sse = sub_pixel_avg_sse_diff_32width_h_msa(src_ptr, src_stride, ref_ptr,
+                                                  ref_stride, sec_pred,
+                                                  h_filter, 64, &diff);
+    } else {
+      *sse = avg_sse_diff_32x64_msa(src_ptr, src_stride, ref_ptr, ref_stride,
+                                    sec_pred, &diff);
+    }
+  }
+
+  return VARIANCE_32Wx64H(*sse, diff);
+}
+
+#define AOM_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(ht)                           \
+  uint32_t aom_sub_pixel_avg_variance64x##ht##_msa(                           \
+      const uint8_t *src_ptr, int32_t src_stride, int32_t xoffset,            \
+      int32_t yoffset, const uint8_t *ref_ptr, int32_t ref_stride,            \
+      uint32_t *sse, const uint8_t *sec_pred) {                               \
+    int32_t diff;                                                             \
+    const uint8_t *h_filter = bilinear_filters_msa[xoffset];                  \
+    const uint8_t *v_filter = bilinear_filters_msa[yoffset];                  \
+                                                                              \
+    if (yoffset) {                                                            \
+      if (xoffset) {                                                          \
+        *sse = sub_pixel_avg_sse_diff_64width_hv_msa(                         \
+            src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, h_filter,     \
+            v_filter, ht, &diff);                                             \
+      } else {                                                                \
+        *sse = sub_pixel_avg_sse_diff_64width_v_msa(                          \
+            src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, v_filter, ht, \
+            &diff);                                                           \
+      }                                                                       \
+    } else {                                                                  \
+      if (xoffset) {                                                          \
+        *sse = sub_pixel_avg_sse_diff_64width_h_msa(                          \
+            src_ptr, src_stride, ref_ptr, ref_stride, sec_pred, h_filter, ht, \
+            &diff);                                                           \
+      } else {                                                                \
+        *sse = avg_sse_diff_64x##ht##_msa(src_ptr, src_stride, ref_ptr,       \
+                                          ref_stride, sec_pred, &diff);       \
+      }                                                                       \
+    }                                                                         \
+                                                                              \
+    return VARIANCE_64Wx##ht##H(*sse, diff);                                  \
+  }
+
+/* clang-format off */
+AOM_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(32)
+AOM_SUB_PIXEL_AVG_VARIANCE64XHEIGHT_MSA(64)
+/* clang-format on */
diff --git a/third_party/aom/aom_dsp/mips/subtract_msa.c b/third_party/aom/aom_dsp/mips/subtract_msa.c
new file mode 100644
index 0000000000..37b89765db
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/subtract_msa.c
@@ -0,0 +1,265 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/macros_msa.h"
+
+static void sub_blk_4x4_msa(const uint8_t *src_ptr, int32_t src_stride,
+                            const uint8_t *pred_ptr, int32_t pred_stride,
+                            int16_t *diff_ptr, int32_t diff_stride) {
+  uint32_t src0, src1, src2, src3;
+  uint32_t pred0, pred1, pred2, pred3;
+  v16i8 src = { 0 };
+  v16i8 pred = { 0 };
+  v16u8 src_l0, src_l1;
+  v8i16 diff0, diff1;
+
+  LW4(src_ptr, src_stride, src0, src1, src2, src3);
+  LW4(pred_ptr, pred_stride, pred0, pred1, pred2, pred3);
+  INSERT_W4_SB(src0, src1, src2, src3, src);
+  INSERT_W4_SB(pred0, pred1, pred2, pred3, pred);
+  ILVRL_B2_UB(src, pred, src_l0, src_l1);
+  HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+  ST8x4_UB(diff0, diff1, diff_ptr, (2 * diff_stride));
+}
+
+static void sub_blk_8x8_msa(const uint8_t *src_ptr, int32_t src_stride,
+                            const uint8_t *pred_ptr, int32_t pred_stride,
+                            int16_t *diff_ptr, int32_t diff_stride) {
+  uint32_t loop_cnt;
+  uint64_t src0, src1, pred0, pred1;
+  v16i8 src = { 0 };
+  v16i8 pred = { 0 };
+  v16u8 src_l0, src_l1;
+  v8i16 diff0, diff1;
+
+  for (loop_cnt = 4; loop_cnt--;) {
+    LD2(src_ptr, src_stride, src0, src1);
+    src_ptr += (2 * src_stride);
+    LD2(pred_ptr, pred_stride, pred0, pred1);
+    pred_ptr += (2 * pred_stride);
+
+    INSERT_D2_SB(src0, src1, src);
+    INSERT_D2_SB(pred0, pred1, pred);
+    ILVRL_B2_UB(src, pred, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff_ptr, diff_stride);
+    diff_ptr += (2 * diff_stride);
+  }
+}
+
+static void sub_blk_16x16_msa(const uint8_t *src, int32_t src_stride,
+                              const uint8_t *pred, int32_t pred_stride,
+                              int16_t *diff, int32_t diff_stride) {
+  int8_t count;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16i8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7;
+  v16u8 src_l0, src_l1;
+  v8i16 diff0, diff1;
+
+  for (count = 2; count--;) {
+    LD_SB8(src, src_stride, src0, src1, src2, src3, src4, src5, src6, src7);
+    src += (8 * src_stride);
+
+    LD_SB8(pred, pred_stride, pred0, pred1, pred2, pred3, pred4, pred5, pred6,
+           pred7);
+    pred += (8 * pred_stride);
+
+    ILVRL_B2_UB(src0, pred0, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src1, pred1, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src2, pred2, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src3, pred3, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src4, pred4, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src5, pred5, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src6, pred6, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src7, pred7, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    diff += diff_stride;
+  }
+}
+
+static void sub_blk_32x32_msa(const uint8_t *src, int32_t src_stride,
+                              const uint8_t *pred, int32_t pred_stride,
+                              int16_t *diff, int32_t diff_stride) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16i8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7;
+  v16u8 src_l0, src_l1;
+  v8i16 diff0, diff1;
+
+  for (loop_cnt = 8; loop_cnt--;) {
+    LD_SB2(src, 16, src0, src1);
+    src += src_stride;
+    LD_SB2(src, 16, src2, src3);
+    src += src_stride;
+    LD_SB2(src, 16, src4, src5);
+    src += src_stride;
+    LD_SB2(src, 16, src6, src7);
+    src += src_stride;
+
+    LD_SB2(pred, 16, pred0, pred1);
+    pred += pred_stride;
+    LD_SB2(pred, 16, pred2, pred3);
+    pred += pred_stride;
+    LD_SB2(pred, 16, pred4, pred5);
+    pred += pred_stride;
+    LD_SB2(pred, 16, pred6, pred7);
+    pred += pred_stride;
+
+    ILVRL_B2_UB(src0, pred0, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src1, pred1, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src2, pred2, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src3, pred3, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src4, pred4, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src5, pred5, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src6, pred6, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src7, pred7, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    diff += diff_stride;
+  }
+}
+
+static void sub_blk_64x64_msa(const uint8_t *src, int32_t src_stride,
+                              const uint8_t *pred, int32_t pred_stride,
+                              int16_t *diff, int32_t diff_stride) {
+  uint32_t loop_cnt;
+  v16i8 src0, src1, src2, src3, src4, src5, src6, src7;
+  v16i8 pred0, pred1, pred2, pred3, pred4, pred5, pred6, pred7;
+  v16u8 src_l0, src_l1;
+  v8i16 diff0, diff1;
+
+  for (loop_cnt = 32; loop_cnt--;) {
+    LD_SB4(src, 16, src0, src1, src2, src3);
+    src += src_stride;
+    LD_SB4(src, 16, src4, src5, src6, src7);
+    src += src_stride;
+
+    LD_SB4(pred, 16, pred0, pred1, pred2, pred3);
+    pred += pred_stride;
+    LD_SB4(pred, 16, pred4, pred5, pred6, pred7);
+    pred += pred_stride;
+
+    ILVRL_B2_UB(src0, pred0, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src1, pred1, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    ILVRL_B2_UB(src2, pred2, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 32, 8);
+    ILVRL_B2_UB(src3, pred3, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 48, 8);
+    diff += diff_stride;
+
+    ILVRL_B2_UB(src4, pred4, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff, 8);
+    ILVRL_B2_UB(src5, pred5, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 16, 8);
+    ILVRL_B2_UB(src6, pred6, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 32, 8);
+    ILVRL_B2_UB(src7, pred7, src_l0, src_l1);
+    HSUB_UB2_SH(src_l0, src_l1, diff0, diff1);
+    ST_SH2(diff0, diff1, diff + 48, 8);
+    diff += diff_stride;
+  }
+}
+
+void aom_subtract_block_msa(int32_t rows, int32_t cols, int16_t *diff_ptr,
+                            ptrdiff_t diff_stride, const uint8_t *src_ptr,
+                            ptrdiff_t src_stride, const uint8_t *pred_ptr,
+                            ptrdiff_t pred_stride) {
+  if (rows == cols) {
+    switch (rows) {
+      case 4:
+        sub_blk_4x4_msa(src_ptr, src_stride, pred_ptr, pred_stride, diff_ptr,
+                        diff_stride);
+        break;
+      case 8:
+        sub_blk_8x8_msa(src_ptr, src_stride, pred_ptr, pred_stride, diff_ptr,
+                        diff_stride);
+        break;
+      case 16:
+        sub_blk_16x16_msa(src_ptr, src_stride, pred_ptr, pred_stride, diff_ptr,
+                          diff_stride);
+        break;
+      case 32:
+        sub_blk_32x32_msa(src_ptr, src_stride, pred_ptr, pred_stride, diff_ptr,
+                          diff_stride);
+        break;
+      case 64:
+        sub_blk_64x64_msa(src_ptr, src_stride, pred_ptr, pred_stride, diff_ptr,
+                          diff_stride);
+        break;
+      default:
+        aom_subtract_block_c(rows, cols, diff_ptr, diff_stride, src_ptr,
+                             src_stride, pred_ptr, pred_stride);
+        break;
+    }
+  } else {
+    aom_subtract_block_c(rows, cols, diff_ptr, diff_stride, src_ptr, src_stride,
+                         pred_ptr, pred_stride);
+  }
+}
diff --git a/third_party/aom/aom_dsp/mips/txfm_macros_msa.h b/third_party/aom/aom_dsp/mips/txfm_macros_msa.h
new file mode 100644
index 0000000000..cba5d4445a
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/txfm_macros_msa.h
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_MIPS_TXFM_MACROS_MIPS_MSA_H_
+#define AOM_DSP_MIPS_TXFM_MACROS_MIPS_MSA_H_
+
+#include "aom_dsp/mips/macros_msa.h"
+
+#define DOTP_CONST_PAIR(reg0, reg1, cnst0, cnst1, out0, out1) \
+  {                                                           \
+    v8i16 k0_m = __msa_fill_h(cnst0);                         \
+    v4i32 s0_m, s1_m, s2_m, s3_m;                             \
+                                                              \
+    s0_m = (v4i32)__msa_fill_h(cnst1);                        \
+    k0_m = __msa_ilvev_h((v8i16)s0_m, k0_m);                  \
+                                                              \
+    ILVRL_H2_SW((-reg1), reg0, s1_m, s0_m);                   \
+    ILVRL_H2_SW(reg0, reg1, s3_m, s2_m);                      \
+    DOTP_SH2_SW(s1_m, s0_m, k0_m, k0_m, s1_m, s0_m);          \
+    SRARI_W2_SW(s1_m, s0_m, DCT_CONST_BITS);                  \
+    out0 = __msa_pckev_h((v8i16)s0_m, (v8i16)s1_m);           \
+                                                              \
+    DOTP_SH2_SW(s3_m, s2_m, k0_m, k0_m, s1_m, s0_m);          \
+    SRARI_W2_SW(s1_m, s0_m, DCT_CONST_BITS);                  \
+    out1 = __msa_pckev_h((v8i16)s0_m, (v8i16)s1_m);           \
+  }
+
+#define DOT_ADD_SUB_SRARI_PCK(in0, in1, in2, in3, in4, in5, in6, in7, dst0,   \
+                              dst1, dst2, dst3)                               \
+  {                                                                           \
+    v4i32 tp0_m, tp1_m, tp2_m, tp3_m, tp4_m;                                  \
+    v4i32 tp5_m, tp6_m, tp7_m, tp8_m, tp9_m;                                  \
+                                                                              \
+    DOTP_SH4_SW(in0, in1, in0, in1, in4, in4, in5, in5, tp0_m, tp2_m, tp3_m,  \
+                tp4_m);                                                       \
+    DOTP_SH4_SW(in2, in3, in2, in3, in6, in6, in7, in7, tp5_m, tp6_m, tp7_m,  \
+                tp8_m);                                                       \
+    BUTTERFLY_4(tp0_m, tp3_m, tp7_m, tp5_m, tp1_m, tp9_m, tp7_m, tp5_m);      \
+    BUTTERFLY_4(tp2_m, tp4_m, tp8_m, tp6_m, tp3_m, tp0_m, tp4_m, tp2_m);      \
+    SRARI_W4_SW(tp1_m, tp9_m, tp7_m, tp5_m, DCT_CONST_BITS);                  \
+    SRARI_W4_SW(tp3_m, tp0_m, tp4_m, tp2_m, DCT_CONST_BITS);                  \
+    PCKEV_H4_SH(tp1_m, tp3_m, tp9_m, tp0_m, tp7_m, tp4_m, tp5_m, tp2_m, dst0, \
+                dst1, dst2, dst3);                                            \
+  }
+
+#define DOT_SHIFT_RIGHT_PCK_H(in0, in1, in2)           \
+  ({                                                   \
+    v8i16 dst_m;                                       \
+    v4i32 tp0_m, tp1_m;                                \
+                                                       \
+    DOTP_SH2_SW(in0, in1, in2, in2, tp1_m, tp0_m);     \
+    SRARI_W2_SW(tp1_m, tp0_m, DCT_CONST_BITS);         \
+    dst_m = __msa_pckev_h((v8i16)tp1_m, (v8i16)tp0_m); \
+                                                       \
+    dst_m;                                             \
+  })
+
+#define MADD_SHORT(m0, m1, c0, c1, res0, res1)                              \
+  {                                                                         \
+    v4i32 madd0_m, madd1_m, madd2_m, madd3_m;                               \
+    v8i16 madd_s0_m, madd_s1_m;                                             \
+                                                                            \
+    ILVRL_H2_SH(m1, m0, madd_s0_m, madd_s1_m);                              \
+    DOTP_SH4_SW(madd_s0_m, madd_s1_m, madd_s0_m, madd_s1_m, c0, c0, c1, c1, \
+                madd0_m, madd1_m, madd2_m, madd3_m);                        \
+    SRARI_W4_SW(madd0_m, madd1_m, madd2_m, madd3_m, DCT_CONST_BITS);        \
+    PCKEV_H2_SH(madd1_m, madd0_m, madd3_m, madd2_m, res0, res1);            \
+  }
+
+#define MADD_BF(inp0, inp1, inp2, inp3, cst0, cst1, cst2, cst3, out0, out1,   \
+                out2, out3)                                                   \
+  {                                                                           \
+    v8i16 madd_s0_m, madd_s1_m, madd_s2_m, madd_s3_m;                         \
+    v4i32 tmp0_m, tmp1_m, tmp2_m, tmp3_m, m4_m, m5_m;                         \
+                                                                              \
+    ILVRL_H2_SH(inp1, inp0, madd_s0_m, madd_s1_m);                            \
+    ILVRL_H2_SH(inp3, inp2, madd_s2_m, madd_s3_m);                            \
+    DOTP_SH4_SW(madd_s0_m, madd_s1_m, madd_s2_m, madd_s3_m, cst0, cst0, cst2, \
+                cst2, tmp0_m, tmp1_m, tmp2_m, tmp3_m);                        \
+    BUTTERFLY_4(tmp0_m, tmp1_m, tmp3_m, tmp2_m, m4_m, m5_m, tmp3_m, tmp2_m);  \
+    SRARI_W4_SW(m4_m, m5_m, tmp2_m, tmp3_m, DCT_CONST_BITS);                  \
+    PCKEV_H2_SH(m5_m, m4_m, tmp3_m, tmp2_m, out0, out1);                      \
+    DOTP_SH4_SW(madd_s0_m, madd_s1_m, madd_s2_m, madd_s3_m, cst1, cst1, cst3, \
+                cst3, tmp0_m, tmp1_m, tmp2_m, tmp3_m);                        \
+    BUTTERFLY_4(tmp0_m, tmp1_m, tmp3_m, tmp2_m, m4_m, m5_m, tmp3_m, tmp2_m);  \
+    SRARI_W4_SW(m4_m, m5_m, tmp2_m, tmp3_m, DCT_CONST_BITS);                  \
+    PCKEV_H2_SH(m5_m, m4_m, tmp3_m, tmp2_m, out2, out3);                      \
+  }
+#endif  // AOM_DSP_MIPS_TXFM_MACROS_MIPS_MSA_H_
diff --git a/third_party/aom/aom_dsp/mips/variance_msa.c b/third_party/aom/aom_dsp/mips/variance_msa.c
new file mode 100644
index 0000000000..745fdfc9c5
--- /dev/null
+++ b/third_party/aom/aom_dsp/mips/variance_msa.c
@@ -0,0 +1,632 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/mips/macros_msa.h"
+
+#define CALC_MSE_B(src, ref, var)                                   \
+  {                                                                 \
+    v16u8 src_l0_m, src_l1_m;                                       \
+    v8i16 res_l0_m, res_l1_m;                                       \
+                                                                    \
+    ILVRL_B2_UB(src, ref, src_l0_m, src_l1_m);                      \
+    HSUB_UB2_SH(src_l0_m, src_l1_m, res_l0_m, res_l1_m);            \
+    DPADD_SH2_SW(res_l0_m, res_l1_m, res_l0_m, res_l1_m, var, var); \
+  }
+
+#define CALC_MSE_AVG_B(src, ref, var, sub)                          \
+  {                                                                 \
+    v16u8 src_l0_m, src_l1_m;                                       \
+    v8i16 res_l0_m, res_l1_m;                                       \
+                                                                    \
+    ILVRL_B2_UB(src, ref, src_l0_m, src_l1_m);                      \
+    HSUB_UB2_SH(src_l0_m, src_l1_m, res_l0_m, res_l1_m);            \
+    DPADD_SH2_SW(res_l0_m, res_l1_m, res_l0_m, res_l1_m, var, var); \
+                                                                    \
+    sub += res_l0_m + res_l1_m;                                     \
+  }
+
+#define VARIANCE_WxH(sse, diff, shift) sse - (((uint32_t)diff * diff) >> shift)
+
+#define VARIANCE_LARGE_WxH(sse, diff, shift) \
+  sse - (((int64_t)diff * diff) >> shift)
+
+static uint32_t sse_diff_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                    const uint8_t *ref_ptr, int32_t ref_stride,
+                                    int32_t height, int32_t *diff) {
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  int32_t ht_cnt;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    CALC_MSE_AVG_B(src, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_8width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                    const uint8_t *ref_ptr, int32_t ref_stride,
+                                    int32_t height, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2, src0, src1,
+                ref0, ref1);
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_16width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                     const uint8_t *ref_ptr, int32_t ref_stride,
+                                     int32_t height, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src, ref;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src, ref, var, avg);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src, ref, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_32width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                     const uint8_t *ref_ptr, int32_t ref_stride,
+                                     int32_t height, int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8i16 avg = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg);
+    CALC_MSE_AVG_B(src1, ref1, var, avg);
+  }
+
+  vec = __msa_hadd_s_w(avg, avg);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_32x64_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                   const uint8_t *ref_ptr, int32_t ref_stride,
+                                   int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 16; ht_cnt--;) {
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_64x32_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                   const uint8_t *ref_ptr, int32_t ref_stride,
+                                   int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 16; ht_cnt--;) {
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src2, ref2, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src3, ref3, var, avg1);
+
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src2, ref2, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src3, ref3, var, avg1);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_diff_64x64_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                   const uint8_t *ref_ptr, int32_t ref_stride,
+                                   int32_t *diff) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v8i16 avg0 = { 0 };
+  v8i16 avg1 = { 0 };
+  v8i16 avg2 = { 0 };
+  v8i16 avg3 = { 0 };
+  v4i32 vec, var = { 0 };
+
+  for (ht_cnt = 32; ht_cnt--;) {
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src2, ref2, var, avg2);
+    CALC_MSE_AVG_B(src3, ref3, var, avg3);
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    CALC_MSE_AVG_B(src0, ref0, var, avg0);
+    CALC_MSE_AVG_B(src1, ref1, var, avg1);
+    CALC_MSE_AVG_B(src2, ref2, var, avg2);
+    CALC_MSE_AVG_B(src3, ref3, var, avg3);
+  }
+
+  vec = __msa_hadd_s_w(avg0, avg0);
+  vec += __msa_hadd_s_w(avg1, avg1);
+  vec += __msa_hadd_s_w(avg2, avg2);
+  vec += __msa_hadd_s_w(avg3, avg3);
+  *diff = HADD_SW_S32(vec);
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t get_mb_ss_msa(const int16_t *src) {
+  uint32_t sum, cnt;
+  v8i16 src0, src1, src2, src3;
+  v4i32 src0_l, src1_l, src2_l, src3_l;
+  v4i32 src0_r, src1_r, src2_r, src3_r;
+  v2i64 sq_src_l = { 0 };
+  v2i64 sq_src_r = { 0 };
+
+  for (cnt = 8; cnt--;) {
+    LD_SH4(src, 8, src0, src1, src2, src3);
+    src += 4 * 8;
+
+    UNPCK_SH_SW(src0, src0_l, src0_r);
+    UNPCK_SH_SW(src1, src1_l, src1_r);
+    UNPCK_SH_SW(src2, src2_l, src2_r);
+    UNPCK_SH_SW(src3, src3_l, src3_r);
+
+    DPADD_SD2_SD(src0_l, src0_r, sq_src_l, sq_src_r);
+    DPADD_SD2_SD(src1_l, src1_r, sq_src_l, sq_src_r);
+    DPADD_SD2_SD(src2_l, src2_r, sq_src_l, sq_src_r);
+    DPADD_SD2_SD(src3_l, src3_r, sq_src_l, sq_src_r);
+  }
+
+  sq_src_l += __msa_splati_d(sq_src_l, 1);
+  sq_src_r += __msa_splati_d(sq_src_r, 1);
+
+  sum = __msa_copy_s_d(sq_src_l, 0);
+  sum += __msa_copy_s_d(sq_src_r, 0);
+
+  return sum;
+}
+
+static uint32_t sse_4width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height) {
+  int32_t ht_cnt;
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16u8 src = { 0 };
+  v16u8 ref = { 0 };
+  v4i32 var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LW4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    INSERT_W4_UB(src0, src1, src2, src3, src);
+    INSERT_W4_UB(ref0, ref1, ref2, ref3, ref);
+    CALC_MSE_B(src, ref, var);
+  }
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_8width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                               const uint8_t *ref_ptr, int32_t ref_stride,
+                               int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v4i32 var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB4(src_ptr, src_stride, src0, src1, src2, src3);
+    src_ptr += (4 * src_stride);
+    LD_UB4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+    ref_ptr += (4 * ref_stride);
+
+    PCKEV_D4_UB(src1, src0, src3, src2, ref1, ref0, ref3, ref2, src0, src1,
+                ref0, ref1);
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src1, ref1, var);
+  }
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_16width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                const uint8_t *ref_ptr, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src, ref;
+  v4i32 var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src, ref, var);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src, ref, var);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src, ref, var);
+
+    src = LD_UB(src_ptr);
+    src_ptr += src_stride;
+    ref = LD_UB(ref_ptr);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src, ref, var);
+  }
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_32width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                const uint8_t *ref_ptr, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, ref0, ref1;
+  v4i32 var = { 0 };
+
+  for (ht_cnt = (height >> 2); ht_cnt--;) {
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src1, ref1, var);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src1, ref1, var);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src1, ref1, var);
+
+    LD_UB2(src_ptr, 16, src0, src1);
+    src_ptr += src_stride;
+    LD_UB2(ref_ptr, 16, ref0, ref1);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src1, ref1, var);
+  }
+
+  return HADD_SW_S32(var);
+}
+
+static uint32_t sse_64width_msa(const uint8_t *src_ptr, int32_t src_stride,
+                                const uint8_t *ref_ptr, int32_t ref_stride,
+                                int32_t height) {
+  int32_t ht_cnt;
+  v16u8 src0, src1, src2, src3;
+  v16u8 ref0, ref1, ref2, ref3;
+  v4i32 var = { 0 };
+
+  for (ht_cnt = height >> 1; ht_cnt--;) {
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src2, ref2, var);
+    CALC_MSE_B(src1, ref1, var);
+    CALC_MSE_B(src3, ref3, var);
+
+    LD_UB4(src_ptr, 16, src0, src1, src2, src3);
+    src_ptr += src_stride;
+    LD_UB4(ref_ptr, 16, ref0, ref1, ref2, ref3);
+    ref_ptr += ref_stride;
+    CALC_MSE_B(src0, ref0, var);
+    CALC_MSE_B(src2, ref2, var);
+    CALC_MSE_B(src1, ref1, var);
+    CALC_MSE_B(src3, ref3, var);
+  }
+
+  return HADD_SW_S32(var);
+}
+
+uint32_t aom_get4x4sse_cs_msa(const uint8_t *src_ptr, int32_t src_stride,
+                              const uint8_t *ref_ptr, int32_t ref_stride) {
+  uint32_t err = 0;
+  uint32_t src0, src1, src2, src3;
+  uint32_t ref0, ref1, ref2, ref3;
+  v16i8 src = { 0 };
+  v16i8 ref = { 0 };
+  v16u8 src_vec0, src_vec1;
+  v8i16 diff0, diff1;
+  v4i32 err0 = { 0 };
+  v4i32 err1 = { 0 };
+
+  LW4(src_ptr, src_stride, src0, src1, src2, src3);
+  LW4(ref_ptr, ref_stride, ref0, ref1, ref2, ref3);
+  INSERT_W4_SB(src0, src1, src2, src3, src);
+  INSERT_W4_SB(ref0, ref1, ref2, ref3, ref);
+  ILVRL_B2_UB(src, ref, src_vec0, src_vec1);
+  HSUB_UB2_SH(src_vec0, src_vec1, diff0, diff1);
+  DPADD_SH2_SW(diff0, diff1, diff0, diff1, err0, err1);
+  err = HADD_SW_S32(err0);
+  err += HADD_SW_S32(err1);
+
+  return err;
+}
+
+#define VARIANCE_4Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 4);
+#define VARIANCE_4Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 5);
+#define VARIANCE_8Wx4H(sse, diff) VARIANCE_WxH(sse, diff, 5);
+#define VARIANCE_8Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 6);
+#define VARIANCE_8Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 7);
+#define VARIANCE_16Wx8H(sse, diff) VARIANCE_WxH(sse, diff, 7);
+#define VARIANCE_16Wx16H(sse, diff) VARIANCE_WxH(sse, diff, 8);
+
+#define VARIANCE_16Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
+#define VARIANCE_32Wx16H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 9);
+#define VARIANCE_32Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 10);
+#define VARIANCE_32Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
+#define VARIANCE_64Wx32H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 11);
+#define VARIANCE_64Wx64H(sse, diff) VARIANCE_LARGE_WxH(sse, diff, 12);
+
+#define AOM_VARIANCE_WDXHT_MSA(wd, ht)                                         \
+  uint32_t aom_variance##wd##x##ht##_msa(                                      \
+      const uint8_t *src, int32_t src_stride, const uint8_t *ref,              \
+      int32_t ref_stride, uint32_t *sse) {                                     \
+    int32_t diff;                                                              \
+                                                                               \
+    *sse =                                                                     \
+        sse_diff_##wd##width_msa(src, src_stride, ref, ref_stride, ht, &diff); \
+                                                                               \
+    return VARIANCE_##wd##Wx##ht##H(*sse, diff);                               \
+  }
+
+/* clang-format off */
+AOM_VARIANCE_WDXHT_MSA(4, 4)
+AOM_VARIANCE_WDXHT_MSA(4, 8)
+
+AOM_VARIANCE_WDXHT_MSA(8, 4)
+AOM_VARIANCE_WDXHT_MSA(8, 8)
+AOM_VARIANCE_WDXHT_MSA(8, 16)
+
+AOM_VARIANCE_WDXHT_MSA(16, 8)
+AOM_VARIANCE_WDXHT_MSA(16, 16)
+AOM_VARIANCE_WDXHT_MSA(16, 32)
+
+AOM_VARIANCE_WDXHT_MSA(32, 16)
+AOM_VARIANCE_WDXHT_MSA(32, 32)
+/* clang-format on */
+
+uint32_t aom_variance32x64_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               uint32_t *sse) {
+  int32_t diff;
+
+  *sse = sse_diff_32x64_msa(src, src_stride, ref, ref_stride, &diff);
+
+  return VARIANCE_32Wx64H(*sse, diff);
+}
+
+uint32_t aom_variance64x32_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               uint32_t *sse) {
+  int32_t diff;
+
+  *sse = sse_diff_64x32_msa(src, src_stride, ref, ref_stride, &diff);
+
+  return VARIANCE_64Wx32H(*sse, diff);
+}
+
+uint32_t aom_variance64x64_msa(const uint8_t *src, int32_t src_stride,
+                               const uint8_t *ref, int32_t ref_stride,
+                               uint32_t *sse) {
+  int32_t diff;
+
+  *sse = sse_diff_64x64_msa(src, src_stride, ref, ref_stride, &diff);
+
+  return VARIANCE_64Wx64H(*sse, diff);
+}
+
+uint32_t aom_mse8x8_msa(const uint8_t *src, int32_t src_stride,
+                        const uint8_t *ref, int32_t ref_stride, uint32_t *sse) {
+  *sse = sse_8width_msa(src, src_stride, ref, ref_stride, 8);
+
+  return *sse;
+}
+
+uint32_t aom_mse8x16_msa(const uint8_t *src, int32_t src_stride,
+                         const uint8_t *ref, int32_t ref_stride,
+                         uint32_t *sse) {
+  *sse = sse_8width_msa(src, src_stride, ref, ref_stride, 16);
+
+  return *sse;
+}
+
+uint32_t aom_mse16x8_msa(const uint8_t *src, int32_t src_stride,
+                         const uint8_t *ref, int32_t ref_stride,
+                         uint32_t *sse) {
+  *sse = sse_16width_msa(src, src_stride, ref, ref_stride, 8);
+
+  return *sse;
+}
+
+uint32_t aom_mse16x16_msa(const uint8_t *src, int32_t src_stride,
+                          const uint8_t *ref, int32_t ref_stride,
+                          uint32_t *sse) {
+  *sse = sse_16width_msa(src, src_stride, ref, ref_stride, 16);
+
+  return *sse;
+}
+
+void aom_get8x8var_msa(const uint8_t *src, int32_t src_stride,
+                       const uint8_t *ref, int32_t ref_stride, uint32_t *sse,
+                       int32_t *sum) {
+  *sse = sse_diff_8width_msa(src, src_stride, ref, ref_stride, 8, sum);
+}
+
+void aom_get16x16var_msa(const uint8_t *src, int32_t src_stride,
+                         const uint8_t *ref, int32_t ref_stride, uint32_t *sse,
+                         int32_t *sum) {
+  *sse = sse_diff_16width_msa(src, src_stride, ref, ref_stride, 16, sum);
+}
+
+uint32_t aom_get_mb_ss_msa(const int16_t *src) { return get_mb_ss_msa(src); }
diff --git a/third_party/aom/aom_dsp/postproc.h b/third_party/aom/aom_dsp/postproc.h
new file mode 100644
index 0000000000..11a8c5ad78
--- /dev/null
+++ b/third_party/aom/aom_dsp/postproc.h
@@ -0,0 +1,26 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_POSTPROC_H_
+#define AOM_DSP_POSTPROC_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Fills a noise buffer with gaussian noise strength determined by sigma.
+int aom_setup_noise(double sigma, int size, char *noise);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // AOM_DSP_POSTPROC_H_
diff --git a/third_party/aom/aom_dsp/prob.c b/third_party/aom/aom_dsp/prob.c
new file mode 100644
index 0000000000..c60bfdac5b
--- /dev/null
+++ b/third_party/aom/aom_dsp/prob.c
@@ -0,0 +1,236 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+
+#if CONFIG_EC_MULTISYMBOL
+#include <string.h>
+#endif
+
+#include "aom_dsp/prob.h"
+
+const uint8_t aom_norm[256] = {
+  0, 7, 6, 6, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
+  3, 3, 3, 3, 3, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+  2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
+};
+
+static unsigned int tree_merge_probs_impl(unsigned int i,
+                                          const aom_tree_index *tree,
+                                          const aom_prob *pre_probs,
+                                          const unsigned int *counts,
+                                          aom_prob *probs) {
+  const int l = tree[i];
+  const unsigned int left_count =
+      (l <= 0) ? counts[-l]
+               : tree_merge_probs_impl(l, tree, pre_probs, counts, probs);
+  const int r = tree[i + 1];
+  const unsigned int right_count =
+      (r <= 0) ? counts[-r]
+               : tree_merge_probs_impl(r, tree, pre_probs, counts, probs);
+  const unsigned int ct[2] = { left_count, right_count };
+  probs[i >> 1] = mode_mv_merge_probs(pre_probs[i >> 1], ct);
+  return left_count + right_count;
+}
+
+void aom_tree_merge_probs(const aom_tree_index *tree, const aom_prob *pre_probs,
+                          const unsigned int *counts, aom_prob *probs) {
+  tree_merge_probs_impl(0, tree, pre_probs, counts, probs);
+}
+
+#if CONFIG_EC_MULTISYMBOL
+typedef struct tree_node tree_node;
+
+struct tree_node {
+  aom_tree_index index;
+  uint8_t probs[16];
+  uint8_t prob;
+  int path;
+  int len;
+  int l;
+  int r;
+  aom_cdf_prob pdf;
+};
+
+/* Compute the probability of this node in Q23 */
+static uint32_t tree_node_prob(tree_node n, int i) {
+  uint32_t prob;
+  /* 1.0 in Q23 */
+  prob = 16777216;
+  for (; i < n.len; i++) {
+    prob = prob * n.probs[i] >> 8;
+  }
+  return prob;
+}
+
+static int tree_node_cmp(tree_node a, tree_node b) {
+  int i;
+  uint32_t pa;
+  uint32_t pb;
+  for (i = 0; i < AOMMIN(a.len, b.len) && a.probs[i] == b.probs[i]; i++) {
+  }
+  pa = tree_node_prob(a, i);
+  pb = tree_node_prob(b, i);
+  return pa > pb ? 1 : pa < pb ? -1 : 0;
+}
+
+/* Given a Q15 probability for symbol subtree rooted at tree[n], this function
+    computes the probability of each symbol (defined as a node that has no
+    children). */
+static aom_cdf_prob tree_node_compute_probs(tree_node *tree, int n,
+                                            aom_cdf_prob pdf) {
+  if (tree[n].l == 0) {
+    /* This prevents probability computations in Q15 that underflow from
+        producing a symbol that has zero probability. */
+    if (pdf == 0) pdf = 1;
+    tree[n].pdf = pdf;
+    return pdf;
+  } else {
+    /* We process the smaller probability first,  */
+    if (tree[n].prob < 128) {
+      aom_cdf_prob lp;
+      aom_cdf_prob rp;
+      lp = (((uint32_t)pdf) * tree[n].prob + 128) >> 8;
+      lp = tree_node_compute_probs(tree, tree[n].l, lp);
+      rp = tree_node_compute_probs(tree, tree[n].r, lp > pdf ? 0 : pdf - lp);
+      return lp + rp;
+    } else {
+      aom_cdf_prob rp;
+      aom_cdf_prob lp;
+      rp = (((uint32_t)pdf) * (256 - tree[n].prob) + 128) >> 8;
+      rp = tree_node_compute_probs(tree, tree[n].r, rp);
+      lp = tree_node_compute_probs(tree, tree[n].l, rp > pdf ? 0 : pdf - rp);
+      return lp + rp;
+    }
+  }
+}
+
+static int tree_node_extract(tree_node *tree, int n, int symb,
+                             aom_cdf_prob *pdf, aom_tree_index *index,
+                             int *path, int *len) {
+  if (tree[n].l == 0) {
+    pdf[symb] = tree[n].pdf;
+    if (index != NULL) index[symb] = tree[n].index;
+    if (path != NULL) path[symb] = tree[n].path;
+    if (len != NULL) len[symb] = tree[n].len;
+    return symb + 1;
+  } else {
+    symb = tree_node_extract(tree, tree[n].l, symb, pdf, index, path, len);
+    return tree_node_extract(tree, tree[n].r, symb, pdf, index, path, len);
+  }
+}
+
+int tree_to_cdf(const aom_tree_index *tree, const aom_prob *probs,
+                aom_tree_index root, aom_cdf_prob *cdf, aom_tree_index *index,
+                int *path, int *len) {
+  tree_node symb[2 * 16 - 1];
+  int nodes;
+  int next[16];
+  int size;
+  int nsymbs;
+  int i;
+  /* Create the root node with probability 1 in Q15. */
+  symb[0].index = root;
+  symb[0].path = 0;
+  symb[0].len = 0;
+  symb[0].l = symb[0].r = 0;
+  nodes = 1;
+  next[0] = 0;
+  size = 1;
+  nsymbs = 1;
+  while (size > 0 && nsymbs < 16) {
+    int m;
+    tree_node n;
+    aom_tree_index j;
+    uint8_t prob;
+    m = 0;
+    /* Find the internal node with the largest probability. */
+    for (i = 1; i < size; i++) {
+      if (tree_node_cmp(symb[next[i]], symb[next[m]]) > 0) m = i;
+    }
+    i = next[m];
+    memmove(&next[m], &next[m + 1], sizeof(*next) * (size - (m + 1)));
+    size--;
+    /* Split this symbol into two symbols */
+    n = symb[i];
+    j = n.index;
+    prob = probs[j >> 1];
+    /* Left */
+    n.index = tree[j];
+    n.path <<= 1;
+    n.len++;
+    n.probs[n.len - 1] = prob;
+    symb[nodes] = n;
+    if (n.index > 0) {
+      next[size++] = nodes;
+    }
+    /* Right */
+    n.index = tree[j + 1];
+    n.path += 1;
+    n.probs[n.len - 1] = 256 - prob;
+    symb[nodes + 1] = n;
+    if (n.index > 0) {
+      next[size++] = nodes + 1;
+    }
+    symb[i].prob = prob;
+    symb[i].l = nodes;
+    symb[i].r = nodes + 1;
+    nodes += 2;
+    nsymbs++;
+  }
+  /* Compute the probabilities of each symbol in Q15 */
+  tree_node_compute_probs(symb, 0, CDF_PROB_TOP);
+  /* Extract the cdf, index, path and length */
+  tree_node_extract(symb, 0, 0, cdf, index, path, len);
+  /* Convert to CDF */
+  cdf[0] = AOM_ICDF(cdf[0]);
+  for (i = 1; i < nsymbs; i++) {
+    cdf[i] = AOM_ICDF(AOM_ICDF(cdf[i - 1]) + cdf[i]);
+  }
+// Store symbol count at the end of the CDF
+#if CONFIG_EC_ADAPT
+  cdf[nsymbs] = 0;
+#endif
+  return nsymbs;
+}
+
+/* This code assumes that tree contains as unique leaf nodes the integer values
+    0 to len - 1 and produces the forward and inverse mapping tables in ind[]
+    and inv[] respectively. */
+static void tree_to_index(int *stack_index, int *ind, int *inv,
+                          const aom_tree_index *tree, int value, int index) {
+  value *= 2;
+
+  do {
+    const aom_tree_index content = tree[index];
+    ++index;
+    if (content <= 0) {
+      inv[*stack_index] = -content;
+      ind[-content] = *stack_index;
+      ++(*stack_index);
+    } else {
+      tree_to_index(stack_index, ind, inv, tree, value, content);
+    }
+  } while (++value & 1);
+}
+
+void av1_indices_from_tree(int *ind, int *inv, const aom_tree_index *tree) {
+  int stack_index = 0;
+  tree_to_index(&stack_index, ind, inv, tree, 0, 0);
+}
+#endif
diff --git a/third_party/aom/aom_dsp/prob.h b/third_party/aom/aom_dsp/prob.h
new file mode 100644
index 0000000000..8085929238
--- /dev/null
+++ b/third_party/aom/aom_dsp/prob.h
@@ -0,0 +1,198 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_PROB_H_
+#define AOM_DSP_PROB_H_
+
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_common.h"
+
+#include "aom_ports/bitops.h"
+#include "aom_ports/mem.h"
+
+#if CONFIG_DAALA_EC
+#include "aom_dsp/entcode.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef uint8_t aom_prob;
+
+// TODO(negge): Rename this aom_prob once we remove vpxbool.
+typedef uint16_t aom_cdf_prob;
+
+#if CONFIG_EC_MULTISYMBOL
+#define CDF_SIZE(x) ((x) + 1)
+#endif
+
+#define CDF_PROB_BITS 15
+#define CDF_PROB_TOP (1 << CDF_PROB_BITS)
+
+#if CONFIG_DAALA_EC
+#define AOM_ICDF OD_ICDF
+#else
+#define AOM_ICDF(x) (x)
+#endif
+
+#define MAX_PROB 255
+
+#define aom_prob_half ((aom_prob)128)
+
+typedef int8_t aom_tree_index;
+
+#define TREE_SIZE(leaf_count) (-2 + 2 * (leaf_count))
+
+#define MODE_MV_COUNT_SAT 20
+
+/* We build coding trees compactly in arrays.
+   Each node of the tree is a pair of aom_tree_indices.
+   Array index often references a corresponding probability table.
+   Index <= 0 means done encoding/decoding and value = -Index,
+   Index > 0 means need another bit, specification at index.
+   Nonnegative indices are always even;  processing begins at node 0. */
+
+typedef const aom_tree_index aom_tree[];
+
+static INLINE aom_prob get_prob(unsigned int num, unsigned int den) {
+  assert(den != 0);
+  {
+    const int p = (int)(((uint64_t)num * 256 + (den >> 1)) / den);
+    // (p > 255) ? 255 : (p < 1) ? 1 : p;
+    const int clipped_prob = p | ((255 - p) >> 23) | (p == 0);
+    return (aom_prob)clipped_prob;
+  }
+}
+
+static INLINE aom_prob get_binary_prob(unsigned int n0, unsigned int n1) {
+  const unsigned int den = n0 + n1;
+  if (den == 0) return 128u;
+  return get_prob(n0, den);
+}
+
+/* This function assumes prob1 and prob2 are already within [1,255] range. */
+static INLINE aom_prob weighted_prob(int prob1, int prob2, int factor) {
+  return ROUND_POWER_OF_TWO(prob1 * (256 - factor) + prob2 * factor, 8);
+}
+
+static INLINE aom_prob merge_probs(aom_prob pre_prob, const unsigned int ct[2],
+                                   unsigned int count_sat,
+                                   unsigned int max_update_factor) {
+  const aom_prob prob = get_binary_prob(ct[0], ct[1]);
+  const unsigned int count = AOMMIN(ct[0] + ct[1], count_sat);
+  const unsigned int factor = max_update_factor * count / count_sat;
+  return weighted_prob(pre_prob, prob, factor);
+}
+
+// MODE_MV_MAX_UPDATE_FACTOR (128) * count / MODE_MV_COUNT_SAT;
+static const int count_to_update_factor[MODE_MV_COUNT_SAT + 1] = {
+  0,  6,  12, 19, 25, 32,  38,  44,  51,  57, 64,
+  70, 76, 83, 89, 96, 102, 108, 115, 121, 128
+};
+
+static INLINE aom_prob mode_mv_merge_probs(aom_prob pre_prob,
+                                           const unsigned int ct[2]) {
+  const unsigned int den = ct[0] + ct[1];
+  if (den == 0) {
+    return pre_prob;
+  } else {
+    const unsigned int count = AOMMIN(den, MODE_MV_COUNT_SAT);
+    const unsigned int factor = count_to_update_factor[count];
+    const aom_prob prob = get_prob(ct[0], den);
+    return weighted_prob(pre_prob, prob, factor);
+  }
+}
+
+void aom_tree_merge_probs(const aom_tree_index *tree, const aom_prob *pre_probs,
+                          const unsigned int *counts, aom_prob *probs);
+
+#if CONFIG_EC_MULTISYMBOL
+int tree_to_cdf(const aom_tree_index *tree, const aom_prob *probs,
+                aom_tree_index root, aom_cdf_prob *cdf, aom_tree_index *ind,
+                int *pth, int *len);
+
+static INLINE void av1_tree_to_cdf(const aom_tree_index *tree,
+                                   const aom_prob *probs, aom_cdf_prob *cdf) {
+  aom_tree_index index[16];
+  int path[16];
+  int dist[16];
+  tree_to_cdf(tree, probs, 0, cdf, index, path, dist);
+}
+
+#define av1_tree_to_cdf_1D(tree, probs, cdf, u) \
+  do {                                          \
+    int i;                                      \
+    for (i = 0; i < u; i++) {                   \
+      av1_tree_to_cdf(tree, probs[i], cdf[i]);  \
+    }                                           \
+  } while (0)
+
+#define av1_tree_to_cdf_2D(tree, probs, cdf, v, u)     \
+  do {                                                 \
+    int j;                                             \
+    int i;                                             \
+    for (j = 0; j < v; j++) {                          \
+      for (i = 0; i < u; i++) {                        \
+        av1_tree_to_cdf(tree, probs[j][i], cdf[j][i]); \
+      }                                                \
+    }                                                  \
+  } while (0)
+
+void av1_indices_from_tree(int *ind, int *inv, const aom_tree_index *tree);
+#endif
+
+DECLARE_ALIGNED(16, extern const uint8_t, aom_norm[256]);
+
+#if CONFIG_EC_ADAPT
+static INLINE void update_cdf(aom_cdf_prob *cdf, int val, int nsymbs) {
+  const int rate = 4 + (cdf[nsymbs] > 31) + get_msb(nsymbs);
+  const int rate2 = 5;
+  int i, tmp;
+  int diff;
+#if 1
+  const int tmp0 = 1 << rate2;
+  tmp = AOM_ICDF(tmp0);
+  diff = ((CDF_PROB_TOP - (nsymbs << rate2)) >> rate) << rate;
+// Single loop (faster)
+#if CONFIG_DAALA_EC && CONFIG_EC_SMALLMUL
+  for (i = 0; i < nsymbs - 1; ++i, tmp -= tmp0) {
+    tmp -= (i == val ? diff : 0);
+    cdf[i] += ((tmp - cdf[i]) >> rate);
+  }
+#else
+  for (i = 0; i < nsymbs - 1; ++i, tmp += tmp0) {
+    tmp += (i == val ? diff : 0);
+    cdf[i] -= ((cdf[i] - tmp) >> rate);
+  }
+#endif
+#else
+  for (i = 0; i < nsymbs; ++i) {
+    tmp = (i + 1) << rate2;
+    cdf[i] -= ((cdf[i] - tmp) >> rate);
+  }
+  diff = CDF_PROB_TOP - cdf[nsymbs - 1];
+
+  for (i = val; i < nsymbs; ++i) {
+    cdf[i] += diff;
+  }
+#endif
+  cdf[nsymbs] += (cdf[nsymbs] < 32);
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_PROB_H_
diff --git a/third_party/aom/aom_dsp/psnr.c b/third_party/aom/aom_dsp/psnr.c
new file mode 100644
index 0000000000..461c13729c
--- /dev/null
+++ b/third_party/aom/aom_dsp/psnr.c
@@ -0,0 +1,373 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/psnr.h"
+#include "aom_scale/yv12config.h"
+
+double aom_sse_to_psnr(double samples, double peak, double sse) {
+  if (sse > 0.0) {
+    const double psnr = 10.0 * log10(samples * peak * peak / sse);
+    return psnr > MAX_PSNR ? MAX_PSNR : psnr;
+  } else {
+    return MAX_PSNR;
+  }
+}
+
+/* TODO(yaowu): The block_variance calls the unoptimized versions of variance()
+* and highbd_8_variance(). It should not.
+*/
+static void encoder_variance(const uint8_t *a, int a_stride, const uint8_t *b,
+                             int b_stride, int w, int h, unsigned int *sse,
+                             int *sum) {
+  int i, j;
+
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void encoder_highbd_variance64(const uint8_t *a8, int a_stride,
+                                      const uint8_t *b8, int b_stride, int w,
+                                      int h, uint64_t *sse, int64_t *sum) {
+  int i, j;
+
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+static void encoder_highbd_8_variance(const uint8_t *a8, int a_stride,
+                                      const uint8_t *b8, int b_stride, int w,
+                                      int h, unsigned int *sse, int *sum) {
+  uint64_t sse_long = 0;
+  int64_t sum_long = 0;
+  encoder_highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long,
+                            &sum_long);
+  *sse = (unsigned int)sse_long;
+  *sum = (int)sum_long;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static int64_t get_sse(const uint8_t *a, int a_stride, const uint8_t *b,
+                       int b_stride, int width, int height) {
+  const int dw = width % 16;
+  const int dh = height % 16;
+  int64_t total_sse = 0;
+  unsigned int sse = 0;
+  int sum = 0;
+  int x, y;
+
+  if (dw > 0) {
+    encoder_variance(&a[width - dw], a_stride, &b[width - dw], b_stride, dw,
+                     height, &sse, &sum);
+    total_sse += sse;
+  }
+
+  if (dh > 0) {
+    encoder_variance(&a[(height - dh) * a_stride], a_stride,
+                     &b[(height - dh) * b_stride], b_stride, width - dw, dh,
+                     &sse, &sum);
+    total_sse += sse;
+  }
+
+  for (y = 0; y < height / 16; ++y) {
+    const uint8_t *pa = a;
+    const uint8_t *pb = b;
+    for (x = 0; x < width / 16; ++x) {
+      aom_mse16x16(pa, a_stride, pb, b_stride, &sse);
+      total_sse += sse;
+
+      pa += 16;
+      pb += 16;
+    }
+
+    a += 16 * a_stride;
+    b += 16 * b_stride;
+  }
+
+  return total_sse;
+}
+
+#if CONFIG_HIGHBITDEPTH
+static int64_t highbd_get_sse_shift(const uint8_t *a8, int a_stride,
+                                    const uint8_t *b8, int b_stride, int width,
+                                    int height, unsigned int input_shift) {
+  const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  const uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  int64_t total_sse = 0;
+  int x, y;
+  for (y = 0; y < height; ++y) {
+    for (x = 0; x < width; ++x) {
+      int64_t diff;
+      diff = (a[x] >> input_shift) - (b[x] >> input_shift);
+      total_sse += diff * diff;
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+  return total_sse;
+}
+
+static int64_t highbd_get_sse(const uint8_t *a, int a_stride, const uint8_t *b,
+                              int b_stride, int width, int height) {
+  int64_t total_sse = 0;
+  int x, y;
+  const int dw = width % 16;
+  const int dh = height % 16;
+  unsigned int sse = 0;
+  int sum = 0;
+  if (dw > 0) {
+    encoder_highbd_8_variance(&a[width - dw], a_stride, &b[width - dw],
+                              b_stride, dw, height, &sse, &sum);
+    total_sse += sse;
+  }
+  if (dh > 0) {
+    encoder_highbd_8_variance(&a[(height - dh) * a_stride], a_stride,
+                              &b[(height - dh) * b_stride], b_stride,
+                              width - dw, dh, &sse, &sum);
+    total_sse += sse;
+  }
+  for (y = 0; y < height / 16; ++y) {
+    const uint8_t *pa = a;
+    const uint8_t *pb = b;
+    for (x = 0; x < width / 16; ++x) {
+      aom_highbd_8_mse16x16(pa, a_stride, pb, b_stride, &sse);
+      total_sse += sse;
+      pa += 16;
+      pb += 16;
+    }
+    a += 16 * a_stride;
+    b += 16 * b_stride;
+  }
+  return total_sse;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+int64_t aom_get_y_sse_part(const YV12_BUFFER_CONFIG *a,
+                           const YV12_BUFFER_CONFIG *b, int hstart, int width,
+                           int vstart, int height) {
+  return get_sse(a->y_buffer + vstart * a->y_stride + hstart, a->y_stride,
+                 b->y_buffer + vstart * b->y_stride + hstart, b->y_stride,
+                 width, height);
+}
+
+int64_t aom_get_y_sse(const YV12_BUFFER_CONFIG *a,
+                      const YV12_BUFFER_CONFIG *b) {
+  assert(a->y_crop_width == b->y_crop_width);
+  assert(a->y_crop_height == b->y_crop_height);
+
+  return get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
+                 a->y_crop_width, a->y_crop_height);
+}
+
+int64_t aom_get_u_sse_part(const YV12_BUFFER_CONFIG *a,
+                           const YV12_BUFFER_CONFIG *b, int hstart, int width,
+                           int vstart, int height) {
+  return get_sse(a->u_buffer + vstart * a->uv_stride + hstart, a->uv_stride,
+                 b->u_buffer + vstart * b->uv_stride + hstart, b->uv_stride,
+                 width, height);
+}
+
+int64_t aom_get_u_sse(const YV12_BUFFER_CONFIG *a,
+                      const YV12_BUFFER_CONFIG *b) {
+  assert(a->uv_crop_width == b->uv_crop_width);
+  assert(a->uv_crop_height == b->uv_crop_height);
+
+  return get_sse(a->u_buffer, a->uv_stride, b->u_buffer, b->uv_stride,
+                 a->uv_crop_width, a->uv_crop_height);
+}
+
+int64_t aom_get_v_sse_part(const YV12_BUFFER_CONFIG *a,
+                           const YV12_BUFFER_CONFIG *b, int hstart, int width,
+                           int vstart, int height) {
+  return get_sse(a->v_buffer + vstart * a->uv_stride + hstart, a->uv_stride,
+                 b->v_buffer + vstart * b->uv_stride + hstart, b->uv_stride,
+                 width, height);
+}
+
+int64_t aom_get_v_sse(const YV12_BUFFER_CONFIG *a,
+                      const YV12_BUFFER_CONFIG *b) {
+  assert(a->uv_crop_width == b->uv_crop_width);
+  assert(a->uv_crop_height == b->uv_crop_height);
+
+  return get_sse(a->v_buffer, a->uv_stride, b->v_buffer, b->uv_stride,
+                 a->uv_crop_width, a->uv_crop_height);
+}
+
+#if CONFIG_HIGHBITDEPTH
+int64_t aom_highbd_get_y_sse_part(const YV12_BUFFER_CONFIG *a,
+                                  const YV12_BUFFER_CONFIG *b, int hstart,
+                                  int width, int vstart, int height) {
+  return highbd_get_sse(
+      a->y_buffer + vstart * a->y_stride + hstart, a->y_stride,
+      b->y_buffer + vstart * b->y_stride + hstart, b->y_stride, width, height);
+}
+
+int64_t aom_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
+                             const YV12_BUFFER_CONFIG *b) {
+  assert(a->y_crop_width == b->y_crop_width);
+  assert(a->y_crop_height == b->y_crop_height);
+  assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
+  assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
+
+  return highbd_get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride,
+                        a->y_crop_width, a->y_crop_height);
+}
+
+int64_t aom_highbd_get_u_sse_part(const YV12_BUFFER_CONFIG *a,
+                                  const YV12_BUFFER_CONFIG *b, int hstart,
+                                  int width, int vstart, int height) {
+  return highbd_get_sse(a->u_buffer + vstart * a->uv_stride + hstart,
+                        a->uv_stride,
+                        b->u_buffer + vstart * b->uv_stride + hstart,
+                        b->uv_stride, width, height);
+}
+
+int64_t aom_highbd_get_u_sse(const YV12_BUFFER_CONFIG *a,
+                             const YV12_BUFFER_CONFIG *b) {
+  assert(a->uv_crop_width == b->uv_crop_width);
+  assert(a->uv_crop_height == b->uv_crop_height);
+  assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
+  assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
+
+  return highbd_get_sse(a->u_buffer, a->uv_stride, b->u_buffer, b->uv_stride,
+                        a->uv_crop_width, a->uv_crop_height);
+}
+
+int64_t aom_highbd_get_v_sse_part(const YV12_BUFFER_CONFIG *a,
+                                  const YV12_BUFFER_CONFIG *b, int hstart,
+                                  int width, int vstart, int height) {
+  return highbd_get_sse(a->v_buffer + vstart * a->uv_stride + hstart,
+                        a->uv_stride,
+                        b->v_buffer + vstart * b->uv_stride + hstart,
+                        b->uv_stride, width, height);
+}
+
+int64_t aom_highbd_get_v_sse(const YV12_BUFFER_CONFIG *a,
+                             const YV12_BUFFER_CONFIG *b) {
+  assert(a->uv_crop_width == b->uv_crop_width);
+  assert(a->uv_crop_height == b->uv_crop_height);
+  assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
+  assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0);
+
+  return highbd_get_sse(a->v_buffer, a->uv_stride, b->v_buffer, b->uv_stride,
+                        a->uv_crop_width, a->uv_crop_height);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_HIGHBITDEPTH
+void aom_calc_highbd_psnr(const YV12_BUFFER_CONFIG *a,
+                          const YV12_BUFFER_CONFIG *b, PSNR_STATS *psnr,
+                          uint32_t bit_depth, uint32_t in_bit_depth) {
+  const int widths[3] = { a->y_crop_width, a->uv_crop_width, a->uv_crop_width };
+  const int heights[3] = { a->y_crop_height, a->uv_crop_height,
+                           a->uv_crop_height };
+  const uint8_t *a_planes[3] = { a->y_buffer, a->u_buffer, a->v_buffer };
+  const int a_strides[3] = { a->y_stride, a->uv_stride, a->uv_stride };
+  const uint8_t *b_planes[3] = { b->y_buffer, b->u_buffer, b->v_buffer };
+  const int b_strides[3] = { b->y_stride, b->uv_stride, b->uv_stride };
+  int i;
+  uint64_t total_sse = 0;
+  uint32_t total_samples = 0;
+  const double peak = (double)((1 << in_bit_depth) - 1);
+  const unsigned int input_shift = bit_depth - in_bit_depth;
+
+  for (i = 0; i < 3; ++i) {
+    const int w = widths[i];
+    const int h = heights[i];
+    const uint32_t samples = w * h;
+    uint64_t sse;
+    if (a->flags & YV12_FLAG_HIGHBITDEPTH) {
+      if (input_shift) {
+        sse = highbd_get_sse_shift(a_planes[i], a_strides[i], b_planes[i],
+                                   b_strides[i], w, h, input_shift);
+      } else {
+        sse = highbd_get_sse(a_planes[i], a_strides[i], b_planes[i],
+                             b_strides[i], w, h);
+      }
+    } else {
+      sse = get_sse(a_planes[i], a_strides[i], b_planes[i], b_strides[i], w, h);
+    }
+    psnr->sse[1 + i] = sse;
+    psnr->samples[1 + i] = samples;
+    psnr->psnr[1 + i] = aom_sse_to_psnr(samples, peak, (double)sse);
+
+    total_sse += sse;
+    total_samples += samples;
+  }
+
+  psnr->sse[0] = total_sse;
+  psnr->samples[0] = total_samples;
+  psnr->psnr[0] =
+      aom_sse_to_psnr((double)total_samples, peak, (double)total_sse);
+}
+
+#endif  // !CONFIG_HIGHBITDEPTH
+
+void aom_calc_psnr(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b,
+                   PSNR_STATS *psnr) {
+  static const double peak = 255.0;
+  const int widths[3] = { a->y_crop_width, a->uv_crop_width, a->uv_crop_width };
+  const int heights[3] = { a->y_crop_height, a->uv_crop_height,
+                           a->uv_crop_height };
+  const uint8_t *a_planes[3] = { a->y_buffer, a->u_buffer, a->v_buffer };
+  const int a_strides[3] = { a->y_stride, a->uv_stride, a->uv_stride };
+  const uint8_t *b_planes[3] = { b->y_buffer, b->u_buffer, b->v_buffer };
+  const int b_strides[3] = { b->y_stride, b->uv_stride, b->uv_stride };
+  int i;
+  uint64_t total_sse = 0;
+  uint32_t total_samples = 0;
+
+  for (i = 0; i < 3; ++i) {
+    const int w = widths[i];
+    const int h = heights[i];
+    const uint32_t samples = w * h;
+    const uint64_t sse =
+        get_sse(a_planes[i], a_strides[i], b_planes[i], b_strides[i], w, h);
+    psnr->sse[1 + i] = sse;
+    psnr->samples[1 + i] = samples;
+    psnr->psnr[1 + i] = aom_sse_to_psnr(samples, peak, (double)sse);
+
+    total_sse += sse;
+    total_samples += samples;
+  }
+
+  psnr->sse[0] = total_sse;
+  psnr->samples[0] = total_samples;
+  psnr->psnr[0] =
+      aom_sse_to_psnr((double)total_samples, peak, (double)total_sse);
+}
diff --git a/third_party/aom/aom_dsp/psnr.h b/third_party/aom/aom_dsp/psnr.h
new file mode 100644
index 0000000000..480140e6f0
--- /dev/null
+++ b/third_party/aom/aom_dsp/psnr.h
@@ -0,0 +1,79 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_PSNR_H_
+#define AOM_DSP_PSNR_H_
+
+#include "aom_scale/yv12config.h"
+
+#define MAX_PSNR 100.0
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  double psnr[4];       // total/y/u/v
+  uint64_t sse[4];      // total/y/u/v
+  uint32_t samples[4];  // total/y/u/v
+} PSNR_STATS;
+
+/*!\brief Converts SSE to PSNR
+*
+* Converts sum of squared errros (SSE) to peak signal-to-noise ratio (PNSR).
+*
+* \param[in]    samples       Number of samples
+* \param[in]    peak          Max sample value
+* \param[in]    sse           Sum of squared errors
+*/
+double aom_sse_to_psnr(double samples, double peak, double sse);
+int64_t aom_get_y_sse_part(const YV12_BUFFER_CONFIG *a,
+                           const YV12_BUFFER_CONFIG *b, int hstart, int width,
+                           int vstart, int height);
+int64_t aom_get_y_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b);
+int64_t aom_get_u_sse_part(const YV12_BUFFER_CONFIG *a,
+                           const YV12_BUFFER_CONFIG *b, int hstart, int width,
+                           int vstart, int height);
+int64_t aom_get_u_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b);
+int64_t aom_get_v_sse_part(const YV12_BUFFER_CONFIG *a,
+                           const YV12_BUFFER_CONFIG *b, int hstart, int width,
+                           int vstart, int height);
+int64_t aom_get_v_sse(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b);
+#if CONFIG_HIGHBITDEPTH
+int64_t aom_highbd_get_y_sse_part(const YV12_BUFFER_CONFIG *a,
+                                  const YV12_BUFFER_CONFIG *b, int hstart,
+                                  int width, int vstart, int height);
+int64_t aom_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a,
+                             const YV12_BUFFER_CONFIG *b);
+int64_t aom_highbd_get_u_sse_part(const YV12_BUFFER_CONFIG *a,
+                                  const YV12_BUFFER_CONFIG *b, int hstart,
+                                  int width, int vstart, int height);
+int64_t aom_highbd_get_u_sse(const YV12_BUFFER_CONFIG *a,
+                             const YV12_BUFFER_CONFIG *b);
+int64_t aom_highbd_get_v_sse_part(const YV12_BUFFER_CONFIG *a,
+                                  const YV12_BUFFER_CONFIG *b, int hstart,
+                                  int width, int vstart, int height);
+int64_t aom_highbd_get_v_sse(const YV12_BUFFER_CONFIG *a,
+                             const YV12_BUFFER_CONFIG *b);
+void aom_calc_highbd_psnr(const YV12_BUFFER_CONFIG *a,
+                          const YV12_BUFFER_CONFIG *b, PSNR_STATS *psnr,
+                          unsigned int bit_depth, unsigned int in_bit_depth);
+#endif
+void aom_calc_psnr(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b,
+                   PSNR_STATS *psnr);
+
+double aom_psnrhvs(const YV12_BUFFER_CONFIG *source,
+                   const YV12_BUFFER_CONFIG *dest, double *phvs_y,
+                   double *phvs_u, double *phvs_v, uint32_t bd, uint32_t in_bd);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+#endif  // AOM_DSP_PSNR_H_
diff --git a/third_party/aom/aom_dsp/psnrhvs.c b/third_party/aom/aom_dsp/psnrhvs.c
new file mode 100644
index 0000000000..aeefd5908e
--- /dev/null
+++ b/third_party/aom/aom_dsp/psnrhvs.c
@@ -0,0 +1,276 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ *
+ *  This code was originally written by: Gregory Maxwell, at the Daala
+ *  project.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/psnr.h"
+#include "aom_dsp/ssim.h"
+#include "aom_ports/system_state.h"
+
+#if !defined(M_PI)
+#define M_PI (3.141592653589793238462643)
+#endif
+#include <string.h>
+
+static void od_bin_fdct8x8(tran_low_t *y, int ystride, const int16_t *x,
+                           int xstride) {
+  int i, j;
+  (void)xstride;
+  aom_fdct8x8(x, y, ystride);
+  for (i = 0; i < 8; i++)
+    for (j = 0; j < 8; j++)
+      *(y + ystride * i + j) = (*(y + ystride * i + j) + 4) >> 3;
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void hbd_od_bin_fdct8x8(tran_low_t *y, int ystride, const int16_t *x,
+                               int xstride) {
+  int i, j;
+  (void)xstride;
+  aom_highbd_fdct8x8(x, y, ystride);
+  for (i = 0; i < 8; i++)
+    for (j = 0; j < 8; j++)
+      *(y + ystride * i + j) = (*(y + ystride * i + j) + 4) >> 3;
+}
+#endif
+
+/* Normalized inverse quantization matrix for 8x8 DCT at the point of
+ * transparency. This is not the JPEG based matrix from the paper,
+ this one gives a slightly higher MOS agreement.*/
+static const double csf_y[8][8] = {
+  { 1.6193873005, 2.2901594831, 2.08509755623, 1.48366094411, 1.00227514334,
+    0.678296995242, 0.466224900598, 0.3265091542 },
+  { 2.2901594831, 1.94321815382, 2.04793073064, 1.68731108984, 1.2305666963,
+    0.868920337363, 0.61280991668, 0.436405793551 },
+  { 2.08509755623, 2.04793073064, 1.34329019223, 1.09205635862, 0.875748795257,
+    0.670882927016, 0.501731932449, 0.372504254596 },
+  { 1.48366094411, 1.68731108984, 1.09205635862, 0.772819797575, 0.605636379554,
+    0.48309405692, 0.380429446972, 0.295774038565 },
+  { 1.00227514334, 1.2305666963, 0.875748795257, 0.605636379554, 0.448996256676,
+    0.352889268808, 0.283006984131, 0.226951348204 },
+  { 0.678296995242, 0.868920337363, 0.670882927016, 0.48309405692,
+    0.352889268808, 0.27032073436, 0.215017739696, 0.17408067321 },
+  { 0.466224900598, 0.61280991668, 0.501731932449, 0.380429446972,
+    0.283006984131, 0.215017739696, 0.168869545842, 0.136153931001 },
+  { 0.3265091542, 0.436405793551, 0.372504254596, 0.295774038565,
+    0.226951348204, 0.17408067321, 0.136153931001, 0.109083846276 }
+};
+static const double csf_cb420[8][8] = {
+  { 1.91113096927, 2.46074210438, 1.18284184739, 1.14982565193, 1.05017074788,
+    0.898018824055, 0.74725392039, 0.615105596242 },
+  { 2.46074210438, 1.58529308355, 1.21363250036, 1.38190029285, 1.33100189972,
+    1.17428548929, 0.996404342439, 0.830890433625 },
+  { 1.18284184739, 1.21363250036, 0.978712413627, 1.02624506078, 1.03145147362,
+    0.960060382087, 0.849823426169, 0.731221236837 },
+  { 1.14982565193, 1.38190029285, 1.02624506078, 0.861317501629, 0.801821139099,
+    0.751437590932, 0.685398513368, 0.608694761374 },
+  { 1.05017074788, 1.33100189972, 1.03145147362, 0.801821139099, 0.676555426187,
+    0.605503172737, 0.55002013668, 0.495804539034 },
+  { 0.898018824055, 1.17428548929, 0.960060382087, 0.751437590932,
+    0.605503172737, 0.514674450957, 0.454353482512, 0.407050308965 },
+  { 0.74725392039, 0.996404342439, 0.849823426169, 0.685398513368,
+    0.55002013668, 0.454353482512, 0.389234902883, 0.342353999733 },
+  { 0.615105596242, 0.830890433625, 0.731221236837, 0.608694761374,
+    0.495804539034, 0.407050308965, 0.342353999733, 0.295530605237 }
+};
+static const double csf_cr420[8][8] = {
+  { 2.03871978502, 2.62502345193, 1.26180942886, 1.11019789803, 1.01397751469,
+    0.867069376285, 0.721500455585, 0.593906509971 },
+  { 2.62502345193, 1.69112867013, 1.17180569821, 1.3342742857, 1.28513006198,
+    1.13381474809, 0.962064122248, 0.802254508198 },
+  { 1.26180942886, 1.17180569821, 0.944981930573, 0.990876405848,
+    0.995903384143, 0.926972725286, 0.820534991409, 0.706020324706 },
+  { 1.11019789803, 1.3342742857, 0.990876405848, 0.831632933426, 0.77418706195,
+    0.725539939514, 0.661776842059, 0.587716619023 },
+  { 1.01397751469, 1.28513006198, 0.995903384143, 0.77418706195, 0.653238524286,
+    0.584635025748, 0.531064164893, 0.478717061273 },
+  { 0.867069376285, 1.13381474809, 0.926972725286, 0.725539939514,
+    0.584635025748, 0.496936637883, 0.438694579826, 0.393021669543 },
+  { 0.721500455585, 0.962064122248, 0.820534991409, 0.661776842059,
+    0.531064164893, 0.438694579826, 0.375820256136, 0.330555063063 },
+  { 0.593906509971, 0.802254508198, 0.706020324706, 0.587716619023,
+    0.478717061273, 0.393021669543, 0.330555063063, 0.285345396658 }
+};
+
+static double convert_score_db(double _score, double _weight, int bit_depth) {
+  int16_t pix_max = 255;
+  assert(_score * _weight >= 0.0);
+  if (bit_depth == 10)
+    pix_max = 1023;
+  else if (bit_depth == 12)
+    pix_max = 4095;
+
+  if (_weight * _score < pix_max * pix_max * 1e-10) return MAX_PSNR;
+  return 10 * (log10(pix_max * pix_max) - log10(_weight * _score));
+}
+
+static double calc_psnrhvs(const unsigned char *src, int _systride,
+                           const unsigned char *dst, int _dystride, double _par,
+                           int _w, int _h, int _step, const double _csf[8][8],
+                           uint32_t bit_depth, uint32_t _shift) {
+  double ret;
+  const uint8_t *_src8 = src;
+  const uint8_t *_dst8 = dst;
+  const uint16_t *_src16 = CONVERT_TO_SHORTPTR(src);
+  const uint16_t *_dst16 = CONVERT_TO_SHORTPTR(dst);
+  int16_t dct_s[8 * 8], dct_d[8 * 8];
+  tran_low_t dct_s_coef[8 * 8], dct_d_coef[8 * 8];
+  double mask[8][8];
+  int pixels;
+  int x;
+  int y;
+  (void)_par;
+  ret = pixels = 0;
+  /*In the PSNR-HVS-M paper[1] the authors describe the construction of
+   their masking table as "we have used the quantization table for the
+   color component Y of JPEG [6] that has been also obtained on the
+   basis of CSF. Note that the values in quantization table JPEG have
+   been normalized and then squared." Their CSF matrix (from PSNR-HVS)
+   was also constructed from the JPEG matrices. I can not find any obvious
+   scheme of normalizing to produce their table, but if I multiply their
+   CSF by 0.38857 and square the result I get their masking table.
+   I have no idea where this constant comes from, but deviating from it
+   too greatly hurts MOS agreement.
+
+   [1] Nikolay Ponomarenko, Flavia Silvestri, Karen Egiazarian, Marco Carli,
+   Jaakko Astola, Vladimir Lukin, "On between-coefficient contrast masking
+   of DCT basis functions", CD-ROM Proceedings of the Third
+   International Workshop on Video Processing and Quality Metrics for Consumer
+   Electronics VPQM-07, Scottsdale, Arizona, USA, 25-26 January, 2007, 4 p.*/
+  for (x = 0; x < 8; x++)
+    for (y = 0; y < 8; y++)
+      mask[x][y] =
+          (_csf[x][y] * 0.3885746225901003) * (_csf[x][y] * 0.3885746225901003);
+  for (y = 0; y < _h - 7; y += _step) {
+    for (x = 0; x < _w - 7; x += _step) {
+      int i;
+      int j;
+      double s_means[4];
+      double d_means[4];
+      double s_vars[4];
+      double d_vars[4];
+      double s_gmean = 0;
+      double d_gmean = 0;
+      double s_gvar = 0;
+      double d_gvar = 0;
+      double s_mask = 0;
+      double d_mask = 0;
+      for (i = 0; i < 4; i++)
+        s_means[i] = d_means[i] = s_vars[i] = d_vars[i] = 0;
+      for (i = 0; i < 8; i++) {
+        for (j = 0; j < 8; j++) {
+          int sub = ((i & 12) >> 2) + ((j & 12) >> 1);
+          if (bit_depth == 8 && _shift == 0) {
+            dct_s[i * 8 + j] = _src8[(y + i) * _systride + (j + x)];
+            dct_d[i * 8 + j] = _dst8[(y + i) * _dystride + (j + x)];
+          } else if (bit_depth == 10 || bit_depth == 12) {
+            dct_s[i * 8 + j] = _src16[(y + i) * _systride + (j + x)] >> _shift;
+            dct_d[i * 8 + j] = _dst16[(y + i) * _dystride + (j + x)] >> _shift;
+          }
+          s_gmean += dct_s[i * 8 + j];
+          d_gmean += dct_d[i * 8 + j];
+          s_means[sub] += dct_s[i * 8 + j];
+          d_means[sub] += dct_d[i * 8 + j];
+        }
+      }
+      s_gmean /= 64.f;
+      d_gmean /= 64.f;
+      for (i = 0; i < 4; i++) s_means[i] /= 16.f;
+      for (i = 0; i < 4; i++) d_means[i] /= 16.f;
+      for (i = 0; i < 8; i++) {
+        for (j = 0; j < 8; j++) {
+          int sub = ((i & 12) >> 2) + ((j & 12) >> 1);
+          s_gvar += (dct_s[i * 8 + j] - s_gmean) * (dct_s[i * 8 + j] - s_gmean);
+          d_gvar += (dct_d[i * 8 + j] - d_gmean) * (dct_d[i * 8 + j] - d_gmean);
+          s_vars[sub] += (dct_s[i * 8 + j] - s_means[sub]) *
+                         (dct_s[i * 8 + j] - s_means[sub]);
+          d_vars[sub] += (dct_d[i * 8 + j] - d_means[sub]) *
+                         (dct_d[i * 8 + j] - d_means[sub]);
+        }
+      }
+      s_gvar *= 1 / 63.f * 64;
+      d_gvar *= 1 / 63.f * 64;
+      for (i = 0; i < 4; i++) s_vars[i] *= 1 / 15.f * 16;
+      for (i = 0; i < 4; i++) d_vars[i] *= 1 / 15.f * 16;
+      if (s_gvar > 0)
+        s_gvar = (s_vars[0] + s_vars[1] + s_vars[2] + s_vars[3]) / s_gvar;
+      if (d_gvar > 0)
+        d_gvar = (d_vars[0] + d_vars[1] + d_vars[2] + d_vars[3]) / d_gvar;
+#if CONFIG_HIGHBITDEPTH
+      if (bit_depth == 10 || bit_depth == 12) {
+        hbd_od_bin_fdct8x8(dct_s_coef, 8, dct_s, 8);
+        hbd_od_bin_fdct8x8(dct_d_coef, 8, dct_d, 8);
+      }
+#endif
+      if (bit_depth == 8) {
+        od_bin_fdct8x8(dct_s_coef, 8, dct_s, 8);
+        od_bin_fdct8x8(dct_d_coef, 8, dct_d, 8);
+      }
+      for (i = 0; i < 8; i++)
+        for (j = (i == 0); j < 8; j++)
+          s_mask += dct_s_coef[i * 8 + j] * dct_s_coef[i * 8 + j] * mask[i][j];
+      for (i = 0; i < 8; i++)
+        for (j = (i == 0); j < 8; j++)
+          d_mask += dct_d_coef[i * 8 + j] * dct_d_coef[i * 8 + j] * mask[i][j];
+      s_mask = sqrt(s_mask * s_gvar) / 32.f;
+      d_mask = sqrt(d_mask * d_gvar) / 32.f;
+      if (d_mask > s_mask) s_mask = d_mask;
+      for (i = 0; i < 8; i++) {
+        for (j = 0; j < 8; j++) {
+          double err;
+          err = fabs((double)(dct_s_coef[i * 8 + j] - dct_d_coef[i * 8 + j]));
+          if (i != 0 || j != 0)
+            err = err < s_mask / mask[i][j] ? 0 : err - s_mask / mask[i][j];
+          ret += (err * _csf[i][j]) * (err * _csf[i][j]);
+          pixels++;
+        }
+      }
+    }
+  }
+  if (pixels <= 0) return 0;
+  ret /= pixels;
+  return ret;
+}
+
+double aom_psnrhvs(const YV12_BUFFER_CONFIG *src, const YV12_BUFFER_CONFIG *dst,
+                   double *y_psnrhvs, double *u_psnrhvs, double *v_psnrhvs,
+                   uint32_t bd, uint32_t in_bd) {
+  double psnrhvs;
+  const double par = 1.0;
+  const int step = 7;
+  uint32_t bd_shift = 0;
+  aom_clear_system_state();
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+  assert(bd >= in_bd);
+
+  bd_shift = bd - in_bd;
+
+  *y_psnrhvs = calc_psnrhvs(src->y_buffer, src->y_stride, dst->y_buffer,
+                            dst->y_stride, par, src->y_crop_width,
+                            src->y_crop_height, step, csf_y, bd, bd_shift);
+  *u_psnrhvs = calc_psnrhvs(src->u_buffer, src->uv_stride, dst->u_buffer,
+                            dst->uv_stride, par, src->uv_crop_width,
+                            src->uv_crop_height, step, csf_cb420, bd, bd_shift);
+  *v_psnrhvs = calc_psnrhvs(src->v_buffer, src->uv_stride, dst->v_buffer,
+                            dst->uv_stride, par, src->uv_crop_width,
+                            src->uv_crop_height, step, csf_cr420, bd, bd_shift);
+  psnrhvs = (*y_psnrhvs) * .8 + .1 * ((*u_psnrhvs) + (*v_psnrhvs));
+  return convert_score_db(psnrhvs, 1.0, in_bd);
+}
diff --git a/third_party/aom/aom_dsp/quantize.c b/third_party/aom/aom_dsp/quantize.c
new file mode 100644
index 0000000000..0759c22e31
--- /dev/null
+++ b/third_party/aom/aom_dsp/quantize.c
@@ -0,0 +1,832 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/quantize.h"
+#include "aom_mem/aom_mem.h"
+
+static void quantize_b_helper_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                int skip_block, const int16_t *zbin_ptr,
+                                const int16_t *round_ptr,
+                                const int16_t *quant_ptr,
+                                const int16_t *quant_shift_ptr,
+                                tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                                const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                                const int16_t *scan, const int16_t *iscan,
+#if CONFIG_AOM_QM
+                                const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr,
+#endif
+                                const int log_scale) {
+  const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], log_scale),
+                         ROUND_POWER_OF_TWO(zbin_ptr[1], log_scale) };
+  const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
+  int i, non_zero_count = (int)n_coeffs, eob = -1;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = (int)n_coeffs - 1; i >= 0; i--) {
+      const int rc = scan[i];
+#if CONFIG_AOM_QM
+      const qm_val_t wt = qm_ptr[rc];
+      const int coeff = coeff_ptr[rc] * wt;
+#else
+      const int coeff = coeff_ptr[rc];
+#endif  // CONFIG_AOM_QM
+
+#if CONFIG_AOM_QM
+      if (coeff < (zbins[rc != 0] << AOM_QM_BITS) &&
+          coeff > (nzbins[rc != 0] << AOM_QM_BITS))
+        non_zero_count--;
+#else
+      if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0]) non_zero_count--;
+#endif  // CONFIG_AOM_QM
+      else
+        break;
+    }
+
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < non_zero_count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      int tmp32;
+
+#if CONFIG_AOM_QM
+      const qm_val_t wt = qm_ptr[rc];
+      if (abs_coeff * wt >= (zbins[rc != 0] << AOM_QM_BITS)) {
+#else
+      if (abs_coeff >= zbins[rc != 0]) {
+#endif  // CONFIG_AOM_QM
+        int64_t tmp =
+            clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], log_scale),
+                  INT16_MIN, INT16_MAX);
+#if CONFIG_AOM_QM
+        tmp *= wt;
+        tmp32 = (int)(((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+                       quant_shift_ptr[rc != 0]) >>
+                      (16 - log_scale + AOM_QM_BITS));  // quantization
+#else
+        tmp32 = (int)(((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+                       quant_shift_ptr[rc != 0]) >>
+                      (16 - log_scale));  // quantization
+#endif  // CONFIG_AOM_QM
+        qcoeff_ptr[rc] = (tmp32 ^ coeff_sign) - coeff_sign;
+#if CONFIG_AOM_QM
+        const int dequant =
+            (dequant_ptr[rc != 0] * iqm_ptr[rc] + (1 << (AOM_QM_BITS - 1))) >>
+            AOM_QM_BITS;
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant / (1 << log_scale);
+#else
+        dqcoeff_ptr[rc] =
+            qcoeff_ptr[rc] * dequant_ptr[rc != 0] / (1 << log_scale);
+#endif  // CONFIG_AOM_QM
+
+        if (tmp32) eob = i;
+      }
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                      int skip_block, const int16_t *zbin_ptr,
+                      const int16_t *round_ptr, const int16_t *quant_ptr,
+                      const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+                      tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                      uint16_t *eob_ptr, const int16_t *scan,
+                      const int16_t *iscan
+#if CONFIG_AOM_QM
+                      ,
+                      const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr
+#endif
+                      ) {
+  quantize_b_helper_c(coeff_ptr, n_coeffs, skip_block, zbin_ptr, round_ptr,
+                      quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
+                      dequant_ptr, eob_ptr, scan, iscan,
+#if CONFIG_AOM_QM
+                      qm_ptr, iqm_ptr,
+#endif
+                      0);
+}
+
+void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            int skip_block, const int16_t *zbin_ptr,
+                            const int16_t *round_ptr, const int16_t *quant_ptr,
+                            const int16_t *quant_shift_ptr,
+                            tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                            const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                            const int16_t *scan, const int16_t *iscan
+#if CONFIG_AOM_QM
+                            ,
+                            const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr
+#endif
+                            ) {
+  quantize_b_helper_c(coeff_ptr, n_coeffs, skip_block, zbin_ptr, round_ptr,
+                      quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
+                      dequant_ptr, eob_ptr, scan, iscan,
+#if CONFIG_AOM_QM
+                      qm_ptr, iqm_ptr,
+#endif
+                      1);
+}
+
+#if CONFIG_TX64X64
+void aom_quantize_b_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            int skip_block, const int16_t *zbin_ptr,
+                            const int16_t *round_ptr, const int16_t *quant_ptr,
+                            const int16_t *quant_shift_ptr,
+                            tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                            const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                            const int16_t *scan, const int16_t *iscan
+#if CONFIG_AOM_QM
+                            ,
+                            const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr
+#endif
+                            ) {
+  quantize_b_helper_c(coeff_ptr, n_coeffs, skip_block, zbin_ptr, round_ptr,
+                      quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
+                      dequant_ptr, eob_ptr, scan, iscan,
+#if CONFIG_AOM_QM
+                      qm_ptr, iqm_ptr,
+#endif
+                      2);
+}
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_AOM_QM
+void aom_quantize_dc(const tran_low_t *coeff_ptr, int n_coeffs, int skip_block,
+                     const int16_t *round_ptr, const int16_t quant,
+                     tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                     const int16_t dequant_ptr, uint16_t *eob_ptr,
+                     const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr) {
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int64_t tmp, eob = -1;
+  int32_t tmp32;
+  int dequant =
+      (dequant_ptr * iqm_ptr[rc] + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+    tmp32 = (int32_t)((tmp * qm_ptr[rc] * quant) >> (16 + AOM_QM_BITS));
+    qcoeff_ptr[rc] = (tmp32 ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant;
+    if (tmp32) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void aom_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr,
+                           const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr) {
+  const int n_coeffs = 1024;
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int64_t tmp, eob = -1;
+  int32_t tmp32;
+  int dequant;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
+                INT16_MIN, INT16_MAX);
+    tmp32 = (int32_t)((tmp * qm_ptr[rc] * quant) >> (15 + AOM_QM_BITS));
+    qcoeff_ptr[rc] = (tmp32 ^ coeff_sign) - coeff_sign;
+    dequant =
+        (dequant_ptr * iqm_ptr[rc] + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
+    dqcoeff_ptr[rc] = (qcoeff_ptr[rc] * dequant) / 2;
+    if (tmp32) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_TX64X64
+void aom_quantize_dc_64x64(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr,
+                           const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr) {
+  const int n_coeffs = 1024;
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int64_t tmp, eob = -1;
+  int32_t tmp32;
+  int dequant;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 2),
+                INT16_MIN, INT16_MAX);
+    tmp32 = (int32_t)((tmp * qm_ptr[rc] * quant) >> (14 + AOM_QM_BITS));
+    qcoeff_ptr[rc] = (tmp32 ^ coeff_sign) - coeff_sign;
+    dequant =
+        (dequant_ptr * iqm_ptr[rc] + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
+    dqcoeff_ptr[rc] = (qcoeff_ptr[rc] * dequant) / 4;
+    if (tmp32) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_quantize_dc(const tran_low_t *coeff_ptr, int n_coeffs,
+                            int skip_block, const int16_t *round_ptr,
+                            const int16_t quant, tran_low_t *qcoeff_ptr,
+                            tran_low_t *dqcoeff_ptr, const int16_t dequant_ptr,
+                            uint16_t *eob_ptr, const qm_val_t *qm_ptr,
+                            const qm_val_t *iqm_ptr) {
+  int eob = -1;
+  int dequant =
+      (dequant_ptr * iqm_ptr[0] + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    const int coeff = coeff_ptr[0];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp = abs_coeff + round_ptr[0];
+    const uint32_t abs_qcoeff =
+        (uint32_t)((tmp * qm_ptr[0] * quant) >> (16 + AOM_QM_BITS));
+    qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+    dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant;
+    if (abs_qcoeff) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void aom_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                                  const int16_t *round_ptr, const int16_t quant,
+                                  tran_low_t *qcoeff_ptr,
+                                  tran_low_t *dqcoeff_ptr,
+                                  const int16_t dequant_ptr, uint16_t *eob_ptr,
+                                  const qm_val_t *qm_ptr,
+                                  const qm_val_t *iqm_ptr) {
+  const int n_coeffs = 1024;
+  int eob = -1;
+  int dequant;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    const int coeff = coeff_ptr[0];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[0], 1);
+    const uint32_t abs_qcoeff =
+        (uint32_t)((tmp * qm_ptr[0] * quant) >> (15 + AOM_QM_BITS));
+    qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+    dequant =
+        (dequant_ptr * iqm_ptr[0] + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
+    dqcoeff_ptr[0] = (qcoeff_ptr[0] * dequant) / 2;
+    if (abs_qcoeff) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_TX64X64
+void aom_highbd_quantize_dc_64x64(const tran_low_t *coeff_ptr, int skip_block,
+                                  const int16_t *round_ptr, const int16_t quant,
+                                  tran_low_t *qcoeff_ptr,
+                                  tran_low_t *dqcoeff_ptr,
+                                  const int16_t dequant_ptr, uint16_t *eob_ptr,
+                                  const qm_val_t *qm_ptr,
+                                  const qm_val_t *iqm_ptr) {
+  const int n_coeffs = 1024;
+  int eob = -1;
+  int dequant;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    const int coeff = coeff_ptr[0];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[0], 2);
+    const uint32_t abs_qcoeff =
+        (uint32_t)((tmp * qm_ptr[0] * quant) >> (14 + AOM_QM_BITS));
+    qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+    dequant =
+        (dequant_ptr * iqm_ptr[0] + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
+    dqcoeff_ptr[0] = (qcoeff_ptr[0] * dequant) / 4;
+    if (abs_qcoeff) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_TX64X64
+
+void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t *zbin_ptr,
+                             const int16_t *round_ptr, const int16_t *quant_ptr,
+                             const int16_t *quant_shift_ptr,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                             const int16_t *scan, const int16_t *iscan,
+                             const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr) {
+  int i, non_zero_count = (int)n_coeffs, eob = -1;
+  const int zbins[2] = { zbin_ptr[0], zbin_ptr[1] };
+  const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
+  int dequant;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = (int)n_coeffs - 1; i >= 0; i--) {
+      const int rc = scan[i];
+      const qm_val_t wt = qm_ptr[rc];
+      const int coeff = coeff_ptr[rc] * wt;
+
+      if (coeff < (zbins[rc != 0] << AOM_QM_BITS) &&
+          coeff > (nzbins[rc != 0] << AOM_QM_BITS))
+        non_zero_count--;
+      else
+        break;
+    }
+
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < non_zero_count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const qm_val_t wt = qm_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff * wt >= (zbins[rc != 0] << AOM_QM_BITS)) {
+        const int64_t tmp1 = abs_coeff + round_ptr[rc != 0];
+        const int64_t tmpw = tmp1 * wt;
+        const int64_t tmp2 = ((tmpw * quant_ptr[rc != 0]) >> 16) + tmpw;
+        const uint32_t abs_qcoeff =
+            (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> (16 + AOM_QM_BITS));
+        qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+        dequant =
+            (dequant_ptr[rc != 0] * iqm_ptr[rc] + (1 << (AOM_QM_BITS - 1))) >>
+            AOM_QM_BITS;
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant;
+        if (abs_qcoeff) eob = i;
+      }
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void aom_highbd_quantize_b_32x32_c(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr,
+    const qm_val_t *iqm_ptr) {
+  const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
+                         ROUND_POWER_OF_TWO(zbin_ptr[1], 1) };
+  const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
+
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  int dequant;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const qm_val_t wt = qm_ptr[rc];
+      const int coeff = coeff_ptr[rc] * wt;
+
+      // If the coefficient is out of the base ZBIN range, keep it for
+      // quantization.
+      if (coeff >= (zbins[rc != 0] << AOM_QM_BITS) ||
+          coeff <= (nzbins[rc != 0] << AOM_QM_BITS))
+        idx_arr[idx++] = i;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = scan[idx_arr[i]];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const qm_val_t wt = qm_ptr[rc];
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int64_t tmp1 =
+          abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+      const int64_t tmpw = tmp1 * wt;
+      const int64_t tmp2 = ((tmpw * quant_ptr[rc != 0]) >> 16) + tmpw;
+      const uint32_t abs_qcoeff =
+          (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> (15 + AOM_QM_BITS));
+      qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+      dequant =
+          (dequant_ptr[rc != 0] * iqm_ptr[rc] + (1 << (AOM_QM_BITS - 1))) >>
+          AOM_QM_BITS;
+      dqcoeff_ptr[rc] = (qcoeff_ptr[rc] * dequant) / 2;
+      if (abs_qcoeff) eob = idx_arr[i];
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_TX64X64
+void aom_highbd_quantize_b_64x64_c(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan, const qm_val_t *qm_ptr,
+    const qm_val_t *iqm_ptr) {
+  const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], 2),
+                         ROUND_POWER_OF_TWO(zbin_ptr[1], 2) };
+  const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
+
+  int idx = 0;
+  int idx_arr[4096];
+  int i, eob = -1;
+  int dequant;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const qm_val_t wt = qm_ptr[rc];
+      const int coeff = coeff_ptr[rc] * wt;
+
+      // If the coefficient is out of the base ZBIN range, keep it for
+      // quantization.
+      if (coeff >= (zbins[rc != 0] << AOM_QM_BITS) ||
+          coeff <= (nzbins[rc != 0] << AOM_QM_BITS))
+        idx_arr[idx++] = i;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = scan[idx_arr[i]];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const qm_val_t wt = qm_ptr[rc];
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int64_t tmp1 =
+          abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 2);
+      const int64_t tmpw = tmp1 * wt;
+      const int64_t tmp2 = ((tmpw * quant_ptr[rc != 0]) >> 16) + tmpw;
+      const uint32_t abs_qcoeff =
+          (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> (14 + AOM_QM_BITS));
+      qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+      dequant =
+          (dequant_ptr[rc != 0] * iqm_ptr[rc] + (1 << (AOM_QM_BITS - 1))) >>
+          AOM_QM_BITS;
+      dqcoeff_ptr[rc] = (qcoeff_ptr[rc] * dequant) / 4;
+      if (abs_qcoeff) eob = idx_arr[i];
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_HIGHBITDEPTH
+
+#else  // CONFIG_AOM_QM
+
+void aom_quantize_dc(const tran_low_t *coeff_ptr, int n_coeffs, int skip_block,
+                     const int16_t *round_ptr, const int16_t quant,
+                     tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                     const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int tmp, eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+    tmp = (tmp * quant) >> 16;
+    qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr;
+    if (tmp) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void aom_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  const int n_coeffs = 1024;
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int tmp, eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
+                INT16_MIN, INT16_MAX);
+    tmp = (tmp * quant) >> 15;
+    qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr / 2;
+    if (tmp) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_TX64X64
+void aom_quantize_dc_64x64(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr) {
+  const int n_coeffs = 4096;
+  const int rc = 0;
+  const int coeff = coeff_ptr[rc];
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int tmp, eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 2),
+                INT16_MIN, INT16_MAX);
+    tmp = (tmp * quant) >> 14;
+    qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr / 4;
+    if (tmp) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_quantize_dc(const tran_low_t *coeff_ptr, int n_coeffs,
+                            int skip_block, const int16_t *round_ptr,
+                            const int16_t quant, tran_low_t *qcoeff_ptr,
+                            tran_low_t *dqcoeff_ptr, const int16_t dequant_ptr,
+                            uint16_t *eob_ptr) {
+  int eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    const int coeff = coeff_ptr[0];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp = abs_coeff + round_ptr[0];
+    const uint32_t abs_qcoeff = (uint32_t)((tmp * quant) >> 16);
+    qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+    dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr;
+    if (abs_qcoeff) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void aom_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                                  const int16_t *round_ptr, const int16_t quant,
+                                  tran_low_t *qcoeff_ptr,
+                                  tran_low_t *dqcoeff_ptr,
+                                  const int16_t dequant_ptr,
+                                  uint16_t *eob_ptr) {
+  const int n_coeffs = 1024;
+  int eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    const int coeff = coeff_ptr[0];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[0], 1);
+    const uint32_t abs_qcoeff = (uint32_t)((tmp * quant) >> 15);
+    qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+    dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr / 2;
+    if (abs_qcoeff) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_TX64X64
+void aom_highbd_quantize_dc_64x64(const tran_low_t *coeff_ptr, int skip_block,
+                                  const int16_t *round_ptr, const int16_t quant,
+                                  tran_low_t *qcoeff_ptr,
+                                  tran_low_t *dqcoeff_ptr,
+                                  const int16_t dequant_ptr,
+                                  uint16_t *eob_ptr) {
+  const int n_coeffs = 4096;
+  int eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    const int coeff = coeff_ptr[0];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[0], 2);
+    const uint32_t abs_qcoeff = (uint32_t)((tmp * quant) >> 14);
+    qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+    dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr / 4;
+    if (abs_qcoeff) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_TX64X64
+
+void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t *zbin_ptr,
+                             const int16_t *round_ptr, const int16_t *quant_ptr,
+                             const int16_t *quant_shift_ptr,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                             const int16_t *scan, const int16_t *iscan) {
+  int i, non_zero_count = (int)n_coeffs, eob = -1;
+  const int zbins[2] = { zbin_ptr[0], zbin_ptr[1] };
+  const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = (int)n_coeffs - 1; i >= 0; i--) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
+        non_zero_count--;
+      else
+        break;
+    }
+
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < non_zero_count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      if (abs_coeff >= zbins[rc != 0]) {
+        const int64_t tmp1 = abs_coeff + round_ptr[rc != 0];
+        const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+        const uint32_t abs_qcoeff =
+            (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 16);
+        qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+        if (abs_qcoeff) eob = i;
+      }
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void aom_highbd_quantize_b_32x32_c(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
+                         ROUND_POWER_OF_TWO(zbin_ptr[1], 1) };
+  const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
+
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      // If the coefficient is out of the base ZBIN range, keep it for
+      // quantization.
+      if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
+        idx_arr[idx++] = i;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = scan[idx_arr[i]];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int64_t tmp1 =
+          abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+      const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+      const uint32_t abs_qcoeff =
+          (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
+      qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+      if (abs_qcoeff) eob = idx_arr[i];
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_TX64X64
+void aom_highbd_quantize_b_64x64_c(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], 2),
+                         ROUND_POWER_OF_TWO(zbin_ptr[1], 2) };
+  const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
+
+  int idx = 0;
+  int idx_arr[4096];
+  int i, eob = -1;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+
+      // If the coefficient is out of the base ZBIN range, keep it for
+      // quantization.
+      if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
+        idx_arr[idx++] = i;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = scan[idx_arr[i]];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int64_t tmp1 =
+          abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 2);
+      const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+      const uint32_t abs_qcoeff =
+          (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 14);
+      qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 4;
+      if (abs_qcoeff) eob = idx_arr[i];
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_AOM_QM
diff --git a/third_party/aom/aom_dsp/quantize.h b/third_party/aom/aom_dsp/quantize.h
new file mode 100644
index 0000000000..fe49b830f1
--- /dev/null
+++ b/third_party/aom/aom_dsp/quantize.h
@@ -0,0 +1,120 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_QUANTIZE_H_
+#define AOM_DSP_QUANTIZE_H_
+
+#include "./aom_config.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_AOM_QM
+void aom_quantize_dc(const tran_low_t *coeff_ptr, int n_coeffs, int skip_block,
+                     const int16_t *round_ptr, const int16_t quant_ptr,
+                     tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                     const int16_t dequant_ptr, uint16_t *eob_ptr,
+                     const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr);
+void aom_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant_ptr,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr,
+                           const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr);
+#if CONFIG_TX64X64
+void aom_quantize_dc_64x64(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant_ptr,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr,
+                           const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr);
+#endif  // CONFIG_TX64X64
+void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                      int skip_block, const int16_t *zbin_ptr,
+                      const int16_t *round_ptr, const int16_t *quant_ptr,
+                      const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+                      tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                      uint16_t *eob_ptr, const int16_t *scan,
+                      const int16_t *iscan, const qm_val_t *qm_ptr,
+                      const qm_val_t *iqm_ptr);
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_quantize_dc(const tran_low_t *coeff_ptr, int n_coeffs,
+                            int skip_block, const int16_t *round_ptr,
+                            const int16_t quant_ptr, tran_low_t *qcoeff_ptr,
+                            tran_low_t *dqcoeff_ptr, const int16_t dequant_ptr,
+                            uint16_t *eob_ptr, const qm_val_t *qm_ptr,
+                            const qm_val_t *iqm_ptr);
+void aom_highbd_quantize_dc_32x32(
+    const tran_low_t *coeff_ptr, int skip_block, const int16_t *round_ptr,
+    const int16_t quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+    const int16_t dequant_ptr, uint16_t *eob_ptr, const qm_val_t *qm_ptr,
+    const qm_val_t *iqm_ptr);
+#if CONFIG_TX64X64
+void aom_highbd_quantize_dc_64x64(
+    const tran_low_t *coeff_ptr, int skip_block, const int16_t *round_ptr,
+    const int16_t quant_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+    const int16_t dequant_ptr, uint16_t *eob_ptr, const qm_val_t *qm_ptr,
+    const qm_val_t *iqm_ptr);
+#endif  // CONFIG_TX64X64
+void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t *zbin_ptr,
+                             const int16_t *round_ptr, const int16_t *quant_ptr,
+                             const int16_t *quant_shift_ptr,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                             const int16_t *scan, const int16_t *iscan,
+                             const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#else  // CONFIG_AOM_QM
+
+void aom_quantize_dc(const tran_low_t *coeff_ptr, int n_coeffs, int skip_block,
+                     const int16_t *round_ptr, const int16_t quant_ptr,
+                     tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                     const int16_t dequant_ptr, uint16_t *eob_ptr);
+void aom_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant_ptr,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr);
+#if CONFIG_TX64X64
+void aom_quantize_dc_64x64(const tran_low_t *coeff_ptr, int skip_block,
+                           const int16_t *round_ptr, const int16_t quant_ptr,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           const int16_t dequant_ptr, uint16_t *eob_ptr);
+#endif  // CONFIG_TX64X64
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_quantize_dc(const tran_low_t *coeff_ptr, int n_coeffs,
+                            int skip_block, const int16_t *round_ptr,
+                            const int16_t quant_ptr, tran_low_t *qcoeff_ptr,
+                            tran_low_t *dqcoeff_ptr, const int16_t dequant_ptr,
+                            uint16_t *eob_ptr);
+void aom_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
+                                  const int16_t *round_ptr,
+                                  const int16_t quant_ptr,
+                                  tran_low_t *qcoeff_ptr,
+                                  tran_low_t *dqcoeff_ptr,
+                                  const int16_t dequant_ptr, uint16_t *eob_ptr);
+#if CONFIG_TX64X64
+void aom_highbd_quantize_dc_64x64(const tran_low_t *coeff_ptr, int skip_block,
+                                  const int16_t *round_ptr,
+                                  const int16_t quant_ptr,
+                                  tran_low_t *qcoeff_ptr,
+                                  tran_low_t *dqcoeff_ptr,
+                                  const int16_t dequant_ptr, uint16_t *eob_ptr);
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_AOM_QM
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_QUANTIZE_H_
diff --git a/third_party/aom/aom_dsp/sad.c b/third_party/aom/aom_dsp/sad.c
new file mode 100644
index 0000000000..3e10705195
--- /dev/null
+++ b/third_party/aom/aom_dsp/sad.c
@@ -0,0 +1,512 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+/* Sum the difference between every corresponding element of the buffers. */
+static INLINE unsigned int sad(const uint8_t *a, int a_stride, const uint8_t *b,
+                               int b_stride, int width, int height) {
+  int y, x;
+  unsigned int sad = 0;
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x++) sad += abs(a[x] - b[x]);
+
+    a += a_stride;
+    b += b_stride;
+  }
+  return sad;
+}
+
+#define sadMxN(m, n)                                                        \
+  unsigned int aom_sad##m##x##n##_c(const uint8_t *src, int src_stride,     \
+                                    const uint8_t *ref, int ref_stride) {   \
+    return sad(src, src_stride, ref, ref_stride, m, n);                     \
+  }                                                                         \
+  unsigned int aom_sad##m##x##n##_avg_c(const uint8_t *src, int src_stride, \
+                                        const uint8_t *ref, int ref_stride, \
+                                        const uint8_t *second_pred) {       \
+    uint8_t comp_pred[m * n];                                               \
+    aom_comp_avg_pred_c(comp_pred, second_pred, m, n, ref, ref_stride);     \
+    return sad(src, src_stride, comp_pred, m, m, n);                        \
+  }
+
+// depending on call sites, pass **ref_array to avoid & in subsequent call and
+// de-dup with 4D below.
+#define sadMxNxK(m, n, k)                                                   \
+  void aom_sad##m##x##n##x##k##_c(const uint8_t *src, int src_stride,       \
+                                  const uint8_t *ref_array, int ref_stride, \
+                                  uint32_t *sad_array) {                    \
+    int i;                                                                  \
+    for (i = 0; i < k; ++i)                                                 \
+      sad_array[i] =                                                        \
+          aom_sad##m##x##n##_c(src, src_stride, &ref_array[i], ref_stride); \
+  }
+
+// This appears to be equivalent to the above when k == 4 and refs is const
+#define sadMxNx4D(m, n)                                                    \
+  void aom_sad##m##x##n##x4d_c(const uint8_t *src, int src_stride,         \
+                               const uint8_t *const ref_array[],           \
+                               int ref_stride, uint32_t *sad_array) {      \
+    int i;                                                                 \
+    for (i = 0; i < 4; ++i)                                                \
+      sad_array[i] =                                                       \
+          aom_sad##m##x##n##_c(src, src_stride, ref_array[i], ref_stride); \
+  }
+
+/* clang-format off */
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+// 128x128
+sadMxN(128, 128)
+sadMxNxK(128, 128, 3)
+sadMxNxK(128, 128, 8)
+sadMxNx4D(128, 128)
+
+// 128x64
+sadMxN(128, 64)
+sadMxNx4D(128, 64)
+
+// 64x128
+sadMxN(64, 128)
+sadMxNx4D(64, 128)
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+
+// 64x64
+sadMxN(64, 64)
+sadMxNxK(64, 64, 3)
+sadMxNxK(64, 64, 8)
+sadMxNx4D(64, 64)
+
+// 64x32
+sadMxN(64, 32)
+sadMxNx4D(64, 32)
+
+// 32x64
+sadMxN(32, 64)
+sadMxNx4D(32, 64)
+
+// 32x32
+sadMxN(32, 32)
+sadMxNxK(32, 32, 3)
+sadMxNxK(32, 32, 8)
+sadMxNx4D(32, 32)
+
+// 32x16
+sadMxN(32, 16)
+sadMxNx4D(32, 16)
+
+// 16x32
+sadMxN(16, 32)
+sadMxNx4D(16, 32)
+
+// 16x16
+sadMxN(16, 16)
+sadMxNxK(16, 16, 3)
+sadMxNxK(16, 16, 8)
+sadMxNx4D(16, 16)
+
+// 16x8
+sadMxN(16, 8)
+sadMxNxK(16, 8, 3)
+sadMxNxK(16, 8, 8)
+sadMxNx4D(16, 8)
+
+// 8x16
+sadMxN(8, 16)
+sadMxNxK(8, 16, 3)
+sadMxNxK(8, 16, 8)
+sadMxNx4D(8, 16)
+
+// 8x8
+sadMxN(8, 8)
+sadMxNxK(8, 8, 3)
+sadMxNxK(8, 8, 8)
+sadMxNx4D(8, 8)
+
+// 8x4
+sadMxN(8, 4)
+sadMxNxK(8, 4, 8)
+sadMxNx4D(8, 4)
+
+// 4x8
+sadMxN(4, 8)
+sadMxNxK(4, 8, 8)
+sadMxNx4D(4, 8)
+
+// 4x4
+sadMxN(4, 4)
+sadMxNxK(4, 4, 3)
+sadMxNxK(4, 4, 8)
+sadMxNx4D(4, 4)
+/* clang-format on */
+
+#if CONFIG_HIGHBITDEPTH
+        static INLINE
+    unsigned int highbd_sad(const uint8_t *a8, int a_stride, const uint8_t *b8,
+                            int b_stride, int width, int height) {
+  int y, x;
+  unsigned int sad = 0;
+  const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  const uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x++) sad += abs(a[x] - b[x]);
+
+    a += a_stride;
+    b += b_stride;
+  }
+  return sad;
+}
+
+static INLINE unsigned int highbd_sadb(const uint8_t *a8, int a_stride,
+                                       const uint16_t *b, int b_stride,
+                                       int width, int height) {
+  int y, x;
+  unsigned int sad = 0;
+  const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x++) sad += abs(a[x] - b[x]);
+
+    a += a_stride;
+    b += b_stride;
+  }
+  return sad;
+}
+
+#define highbd_sadMxN(m, n)                                                    \
+  unsigned int aom_highbd_sad##m##x##n##_c(const uint8_t *src, int src_stride, \
+                                           const uint8_t *ref,                 \
+                                           int ref_stride) {                   \
+    return highbd_sad(src, src_stride, ref, ref_stride, m, n);                 \
+  }                                                                            \
+  unsigned int aom_highbd_sad##m##x##n##_avg_c(                                \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,  \
+      const uint8_t *second_pred) {                                            \
+    uint16_t comp_pred[m * n];                                                 \
+    aom_highbd_comp_avg_pred_c(comp_pred, second_pred, m, n, ref, ref_stride); \
+    return highbd_sadb(src, src_stride, comp_pred, m, m, n);                   \
+  }
+
+#define highbd_sadMxNxK(m, n, k)                                             \
+  void aom_highbd_sad##m##x##n##x##k##_c(                                    \
+      const uint8_t *src, int src_stride, const uint8_t *ref_array,          \
+      int ref_stride, uint32_t *sad_array) {                                 \
+    int i;                                                                   \
+    for (i = 0; i < k; ++i) {                                                \
+      sad_array[i] = aom_highbd_sad##m##x##n##_c(src, src_stride,            \
+                                                 &ref_array[i], ref_stride); \
+    }                                                                        \
+  }
+
+#define highbd_sadMxNx4D(m, n)                                               \
+  void aom_highbd_sad##m##x##n##x4d_c(const uint8_t *src, int src_stride,    \
+                                      const uint8_t *const ref_array[],      \
+                                      int ref_stride, uint32_t *sad_array) { \
+    int i;                                                                   \
+    for (i = 0; i < 4; ++i) {                                                \
+      sad_array[i] = aom_highbd_sad##m##x##n##_c(src, src_stride,            \
+                                                 ref_array[i], ref_stride);  \
+    }                                                                        \
+  }
+
+/* clang-format off */
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+// 128x128
+highbd_sadMxN(128, 128)
+highbd_sadMxNxK(128, 128, 3)
+highbd_sadMxNxK(128, 128, 8)
+highbd_sadMxNx4D(128, 128)
+
+// 128x64
+highbd_sadMxN(128, 64)
+highbd_sadMxNx4D(128, 64)
+
+// 64x128
+highbd_sadMxN(64, 128)
+highbd_sadMxNx4D(64, 128)
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+
+// 64x64
+highbd_sadMxN(64, 64)
+highbd_sadMxNxK(64, 64, 3)
+highbd_sadMxNxK(64, 64, 8)
+highbd_sadMxNx4D(64, 64)
+
+// 64x32
+highbd_sadMxN(64, 32)
+highbd_sadMxNx4D(64, 32)
+
+// 32x64
+highbd_sadMxN(32, 64)
+highbd_sadMxNx4D(32, 64)
+
+// 32x32
+highbd_sadMxN(32, 32)
+highbd_sadMxNxK(32, 32, 3)
+highbd_sadMxNxK(32, 32, 8)
+highbd_sadMxNx4D(32, 32)
+
+// 32x16
+highbd_sadMxN(32, 16)
+highbd_sadMxNx4D(32, 16)
+
+// 16x32
+highbd_sadMxN(16, 32)
+highbd_sadMxNx4D(16, 32)
+
+// 16x16
+highbd_sadMxN(16, 16)
+highbd_sadMxNxK(16, 16, 3)
+highbd_sadMxNxK(16, 16, 8)
+highbd_sadMxNx4D(16, 16)
+
+// 16x8
+highbd_sadMxN(16, 8)
+highbd_sadMxNxK(16, 8, 3)
+highbd_sadMxNxK(16, 8, 8)
+highbd_sadMxNx4D(16, 8)
+
+// 8x16
+highbd_sadMxN(8, 16)
+highbd_sadMxNxK(8, 16, 3)
+highbd_sadMxNxK(8, 16, 8)
+highbd_sadMxNx4D(8, 16)
+
+// 8x8
+highbd_sadMxN(8, 8)
+highbd_sadMxNxK(8, 8, 3)
+highbd_sadMxNxK(8, 8, 8)
+highbd_sadMxNx4D(8, 8)
+
+// 8x4
+highbd_sadMxN(8, 4)
+highbd_sadMxNxK(8, 4, 8)
+highbd_sadMxNx4D(8, 4)
+
+// 4x8
+highbd_sadMxN(4, 8)
+highbd_sadMxNxK(4, 8, 8)
+highbd_sadMxNx4D(4, 8)
+
+// 4x4
+highbd_sadMxN(4, 4)
+highbd_sadMxNxK(4, 4, 3)
+highbd_sadMxNxK(4, 4, 8)
+highbd_sadMxNx4D(4, 4)
+/* clang-format on */
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_AV1 && CONFIG_EXT_INTER
+            static INLINE
+    unsigned int masked_sad(const uint8_t *a, int a_stride, const uint8_t *b,
+                            int b_stride, const uint8_t *m, int m_stride,
+                            int width, int height) {
+  int y, x;
+  unsigned int sad = 0;
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x++) sad += m[x] * abs(a[x] - b[x]);
+
+    a += a_stride;
+    b += b_stride;
+    m += m_stride;
+  }
+  sad = (sad + 31) >> 6;
+
+  return sad;
+}
+
+#define MASKSADMxN(m, n)                                                      \
+  unsigned int aom_masked_sad##m##x##n##_c(                                   \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      const uint8_t *msk, int msk_stride) {                                   \
+    return masked_sad(src, src_stride, ref, ref_stride, msk, msk_stride, m,   \
+                      n);                                                     \
+  }
+
+/* clang-format off */
+#if CONFIG_EXT_PARTITION
+MASKSADMxN(128, 128)
+MASKSADMxN(128, 64)
+MASKSADMxN(64, 128)
+#endif  // CONFIG_EXT_PARTITION
+MASKSADMxN(64, 64)
+MASKSADMxN(64, 32)
+MASKSADMxN(32, 64)
+MASKSADMxN(32, 32)
+MASKSADMxN(32, 16)
+MASKSADMxN(16, 32)
+MASKSADMxN(16, 16)
+MASKSADMxN(16, 8)
+MASKSADMxN(8, 16)
+MASKSADMxN(8, 8)
+MASKSADMxN(8, 4)
+MASKSADMxN(4, 8)
+MASKSADMxN(4, 4)
+/* clang-format on */
+
+#if CONFIG_HIGHBITDEPTH
+                    static INLINE
+    unsigned int highbd_masked_sad(const uint8_t *a8, int a_stride,
+                                   const uint8_t *b8, int b_stride,
+                                   const uint8_t *m, int m_stride, int width,
+                                   int height) {
+  int y, x;
+  unsigned int sad = 0;
+  const uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  const uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x++) sad += m[x] * abs(a[x] - b[x]);
+
+    a += a_stride;
+    b += b_stride;
+    m += m_stride;
+  }
+  sad = (sad + 31) >> 6;
+
+  return sad;
+}
+
+#define HIGHBD_MASKSADMXN(m, n)                                               \
+  unsigned int aom_highbd_masked_sad##m##x##n##_c(                            \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      const uint8_t *msk, int msk_stride) {                                   \
+    return highbd_masked_sad(src, src_stride, ref, ref_stride, msk,           \
+                             msk_stride, m, n);                               \
+  }
+
+#if CONFIG_EXT_PARTITION
+HIGHBD_MASKSADMXN(128, 128)
+HIGHBD_MASKSADMXN(128, 64)
+HIGHBD_MASKSADMXN(64, 128)
+#endif  // CONFIG_EXT_PARTITION
+HIGHBD_MASKSADMXN(64, 64)
+HIGHBD_MASKSADMXN(64, 32)
+HIGHBD_MASKSADMXN(32, 64)
+HIGHBD_MASKSADMXN(32, 32)
+HIGHBD_MASKSADMXN(32, 16)
+HIGHBD_MASKSADMXN(16, 32)
+HIGHBD_MASKSADMXN(16, 16)
+HIGHBD_MASKSADMXN(16, 8)
+HIGHBD_MASKSADMXN(8, 16)
+HIGHBD_MASKSADMXN(8, 8)
+HIGHBD_MASKSADMXN(8, 4)
+HIGHBD_MASKSADMXN(4, 8)
+HIGHBD_MASKSADMXN(4, 4)
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_AV1 && CONFIG_EXT_INTER
+
+#if CONFIG_AV1 && CONFIG_MOTION_VAR
+// pre: predictor being evaluated
+// wsrc: target weighted prediction (has been *4096 to keep precision)
+// mask: 2d weights (scaled by 4096)
+static INLINE unsigned int obmc_sad(const uint8_t *pre, int pre_stride,
+                                    const int32_t *wsrc, const int32_t *mask,
+                                    int width, int height) {
+  int y, x;
+  unsigned int sad = 0;
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x++)
+      sad += ROUND_POWER_OF_TWO(abs(wsrc[x] - pre[x] * mask[x]), 12);
+
+    pre += pre_stride;
+    wsrc += width;
+    mask += width;
+  }
+
+  return sad;
+}
+
+#define OBMCSADMxN(m, n)                                                     \
+  unsigned int aom_obmc_sad##m##x##n##_c(const uint8_t *ref, int ref_stride, \
+                                         const int32_t *wsrc,                \
+                                         const int32_t *mask) {              \
+    return obmc_sad(ref, ref_stride, wsrc, mask, m, n);                      \
+  }
+
+/* clang-format off */
+#if CONFIG_EXT_PARTITION
+OBMCSADMxN(128, 128)
+OBMCSADMxN(128, 64)
+OBMCSADMxN(64, 128)
+#endif  // CONFIG_EXT_PARTITION
+OBMCSADMxN(64, 64)
+OBMCSADMxN(64, 32)
+OBMCSADMxN(32, 64)
+OBMCSADMxN(32, 32)
+OBMCSADMxN(32, 16)
+OBMCSADMxN(16, 32)
+OBMCSADMxN(16, 16)
+OBMCSADMxN(16, 8)
+OBMCSADMxN(8, 16)
+OBMCSADMxN(8, 8)
+OBMCSADMxN(8, 4)
+OBMCSADMxN(4, 8)
+OBMCSADMxN(4, 4)
+/* clang-format on */
+
+#if CONFIG_HIGHBITDEPTH
+                    static INLINE
+    unsigned int highbd_obmc_sad(const uint8_t *pre8, int pre_stride,
+                                 const int32_t *wsrc, const int32_t *mask,
+                                 int width, int height) {
+  int y, x;
+  unsigned int sad = 0;
+  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x++)
+      sad += ROUND_POWER_OF_TWO(abs(wsrc[x] - pre[x] * mask[x]), 12);
+
+    pre += pre_stride;
+    wsrc += width;
+    mask += width;
+  }
+
+  return sad;
+}
+
+#define HIGHBD_OBMCSADMXN(m, n)                                \
+  unsigned int aom_highbd_obmc_sad##m##x##n##_c(               \
+      const uint8_t *ref, int ref_stride, const int32_t *wsrc, \
+      const int32_t *mask) {                                   \
+    return highbd_obmc_sad(ref, ref_stride, wsrc, mask, m, n); \
+  }
+
+/* clang-format off */
+#if CONFIG_EXT_PARTITION
+HIGHBD_OBMCSADMXN(128, 128)
+HIGHBD_OBMCSADMXN(128, 64)
+HIGHBD_OBMCSADMXN(64, 128)
+#endif  // CONFIG_EXT_PARTITION
+HIGHBD_OBMCSADMXN(64, 64)
+HIGHBD_OBMCSADMXN(64, 32)
+HIGHBD_OBMCSADMXN(32, 64)
+HIGHBD_OBMCSADMXN(32, 32)
+HIGHBD_OBMCSADMXN(32, 16)
+HIGHBD_OBMCSADMXN(16, 32)
+HIGHBD_OBMCSADMXN(16, 16)
+HIGHBD_OBMCSADMXN(16, 8)
+HIGHBD_OBMCSADMXN(8, 16)
+HIGHBD_OBMCSADMXN(8, 8)
+HIGHBD_OBMCSADMXN(8, 4)
+HIGHBD_OBMCSADMXN(4, 8)
+HIGHBD_OBMCSADMXN(4, 4)
+/* clang-format on */
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_AV1 && CONFIG_MOTION_VAR
diff --git a/third_party/aom/aom_dsp/simd/v128_intrinsics.h b/third_party/aom/aom_dsp/simd/v128_intrinsics.h
new file mode 100644
index 0000000000..8f6509383a
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v128_intrinsics.h
@@ -0,0 +1,268 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V128_INTRINSICS_H
+#define _V128_INTRINSICS_H
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "./v128_intrinsics_c.h"
+#include "./v64_intrinsics.h"
+
+/* Fallback to plain, unoptimised C. */
+
+typedef c_v128 v128;
+
+SIMD_INLINE uint32_t v128_low_u32(v128 a) { return c_v128_low_u32(a); }
+SIMD_INLINE v64 v128_low_v64(v128 a) { return c_v128_low_v64(a); }
+SIMD_INLINE v64 v128_high_v64(v128 a) { return c_v128_high_v64(a); }
+SIMD_INLINE v128 v128_from_64(uint64_t hi, uint64_t lo) {
+  return c_v128_from_64(hi, lo);
+}
+SIMD_INLINE v128 v128_from_v64(v64 hi, v64 lo) {
+  return c_v128_from_v64(hi, lo);
+}
+SIMD_INLINE v128 v128_from_32(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
+  return c_v128_from_32(a, b, c, d);
+}
+
+SIMD_INLINE v128 v128_load_unaligned(const void *p) {
+  return c_v128_load_unaligned(p);
+}
+SIMD_INLINE v128 v128_load_aligned(const void *p) {
+  return c_v128_load_aligned(p);
+}
+
+SIMD_INLINE void v128_store_unaligned(void *p, v128 a) {
+  c_v128_store_unaligned(p, a);
+}
+SIMD_INLINE void v128_store_aligned(void *p, v128 a) {
+  c_v128_store_aligned(p, a);
+}
+
+SIMD_INLINE v128 v128_align(v128 a, v128 b, unsigned int c) {
+  return c_v128_align(a, b, c);
+}
+
+SIMD_INLINE v128 v128_zero() { return c_v128_zero(); }
+SIMD_INLINE v128 v128_dup_8(uint8_t x) { return c_v128_dup_8(x); }
+SIMD_INLINE v128 v128_dup_16(uint16_t x) { return c_v128_dup_16(x); }
+SIMD_INLINE v128 v128_dup_32(uint32_t x) { return c_v128_dup_32(x); }
+
+typedef uint32_t sad128_internal;
+SIMD_INLINE sad128_internal v128_sad_u8_init() { return c_v128_sad_u8_init(); }
+SIMD_INLINE sad128_internal v128_sad_u8(sad128_internal s, v128 a, v128 b) {
+  return c_v128_sad_u8(s, a, b);
+}
+SIMD_INLINE uint32_t v128_sad_u8_sum(sad128_internal s) {
+  return c_v128_sad_u8_sum(s);
+}
+typedef uint32_t ssd128_internal;
+SIMD_INLINE ssd128_internal v128_ssd_u8_init() { return c_v128_ssd_u8_init(); }
+SIMD_INLINE ssd128_internal v128_ssd_u8(ssd128_internal s, v128 a, v128 b) {
+  return c_v128_ssd_u8(s, a, b);
+}
+SIMD_INLINE uint32_t v128_ssd_u8_sum(ssd128_internal s) {
+  return c_v128_ssd_u8_sum(s);
+}
+SIMD_INLINE int64_t v128_dotp_s16(v128 a, v128 b) {
+  return c_v128_dotp_s16(a, b);
+}
+SIMD_INLINE uint64_t v128_hadd_u8(v128 a) { return c_v128_hadd_u8(a); }
+
+SIMD_INLINE v128 v128_or(v128 a, v128 b) { return c_v128_or(a, b); }
+SIMD_INLINE v128 v128_xor(v128 a, v128 b) { return c_v128_xor(a, b); }
+SIMD_INLINE v128 v128_and(v128 a, v128 b) { return c_v128_and(a, b); }
+SIMD_INLINE v128 v128_andn(v128 a, v128 b) { return c_v128_andn(a, b); }
+
+SIMD_INLINE v128 v128_add_8(v128 a, v128 b) { return c_v128_add_8(a, b); }
+SIMD_INLINE v128 v128_add_16(v128 a, v128 b) { return c_v128_add_16(a, b); }
+SIMD_INLINE v128 v128_sadd_s16(v128 a, v128 b) { return c_v128_sadd_s16(a, b); }
+SIMD_INLINE v128 v128_add_32(v128 a, v128 b) { return c_v128_add_32(a, b); }
+SIMD_INLINE v128 v128_padd_s16(v128 a) { return c_v128_padd_s16(a); }
+SIMD_INLINE v128 v128_sub_8(v128 a, v128 b) { return c_v128_sub_8(a, b); }
+SIMD_INLINE v128 v128_ssub_u8(v128 a, v128 b) { return c_v128_ssub_u8(a, b); }
+SIMD_INLINE v128 v128_ssub_s8(v128 a, v128 b) { return c_v128_ssub_s8(a, b); }
+SIMD_INLINE v128 v128_sub_16(v128 a, v128 b) { return c_v128_sub_16(a, b); }
+SIMD_INLINE v128 v128_ssub_s16(v128 a, v128 b) { return c_v128_ssub_s16(a, b); }
+SIMD_INLINE v128 v128_ssub_u16(v128 a, v128 b) { return c_v128_ssub_u16(a, b); }
+SIMD_INLINE v128 v128_sub_32(v128 a, v128 b) { return c_v128_sub_32(a, b); }
+SIMD_INLINE v128 v128_abs_s16(v128 a) { return c_v128_abs_s16(a); }
+SIMD_INLINE v128 v128_abs_s8(v128 a) { return c_v128_abs_s8(a); }
+
+SIMD_INLINE v128 v128_mul_s16(v64 a, v64 b) { return c_v128_mul_s16(a, b); }
+SIMD_INLINE v128 v128_mullo_s16(v128 a, v128 b) {
+  return c_v128_mullo_s16(a, b);
+}
+SIMD_INLINE v128 v128_mulhi_s16(v128 a, v128 b) {
+  return c_v128_mulhi_s16(a, b);
+}
+SIMD_INLINE v128 v128_mullo_s32(v128 a, v128 b) {
+  return c_v128_mullo_s32(a, b);
+}
+SIMD_INLINE v128 v128_madd_s16(v128 a, v128 b) { return c_v128_madd_s16(a, b); }
+SIMD_INLINE v128 v128_madd_us8(v128 a, v128 b) { return c_v128_madd_us8(a, b); }
+
+SIMD_INLINE v128 v128_avg_u8(v128 a, v128 b) { return c_v128_avg_u8(a, b); }
+SIMD_INLINE v128 v128_rdavg_u8(v128 a, v128 b) { return c_v128_rdavg_u8(a, b); }
+SIMD_INLINE v128 v128_avg_u16(v128 a, v128 b) { return c_v128_avg_u16(a, b); }
+SIMD_INLINE v128 v128_min_u8(v128 a, v128 b) { return c_v128_min_u8(a, b); }
+SIMD_INLINE v128 v128_max_u8(v128 a, v128 b) { return c_v128_max_u8(a, b); }
+SIMD_INLINE v128 v128_min_s8(v128 a, v128 b) { return c_v128_min_s8(a, b); }
+SIMD_INLINE v128 v128_max_s8(v128 a, v128 b) { return c_v128_max_s8(a, b); }
+SIMD_INLINE v128 v128_min_s16(v128 a, v128 b) { return c_v128_min_s16(a, b); }
+SIMD_INLINE v128 v128_max_s16(v128 a, v128 b) { return c_v128_max_s16(a, b); }
+
+SIMD_INLINE v128 v128_ziplo_8(v128 a, v128 b) { return c_v128_ziplo_8(a, b); }
+SIMD_INLINE v128 v128_ziphi_8(v128 a, v128 b) { return c_v128_ziphi_8(a, b); }
+SIMD_INLINE v128 v128_ziplo_16(v128 a, v128 b) { return c_v128_ziplo_16(a, b); }
+SIMD_INLINE v128 v128_ziphi_16(v128 a, v128 b) { return c_v128_ziphi_16(a, b); }
+SIMD_INLINE v128 v128_ziplo_32(v128 a, v128 b) { return c_v128_ziplo_32(a, b); }
+SIMD_INLINE v128 v128_ziphi_32(v128 a, v128 b) { return c_v128_ziphi_32(a, b); }
+SIMD_INLINE v128 v128_ziplo_64(v128 a, v128 b) { return c_v128_ziplo_64(a, b); }
+SIMD_INLINE v128 v128_ziphi_64(v128 a, v128 b) { return c_v128_ziphi_64(a, b); }
+SIMD_INLINE v128 v128_zip_8(v64 a, v64 b) { return c_v128_zip_8(a, b); }
+SIMD_INLINE v128 v128_zip_16(v64 a, v64 b) { return c_v128_zip_16(a, b); }
+SIMD_INLINE v128 v128_zip_32(v64 a, v64 b) { return c_v128_zip_32(a, b); }
+SIMD_INLINE v128 v128_unziplo_8(v128 a, v128 b) {
+  return c_v128_unziplo_8(a, b);
+}
+SIMD_INLINE v128 v128_unziphi_8(v128 a, v128 b) {
+  return c_v128_unziphi_8(a, b);
+}
+SIMD_INLINE v128 v128_unziplo_16(v128 a, v128 b) {
+  return c_v128_unziplo_16(a, b);
+}
+SIMD_INLINE v128 v128_unziphi_16(v128 a, v128 b) {
+  return c_v128_unziphi_16(a, b);
+}
+SIMD_INLINE v128 v128_unziplo_32(v128 a, v128 b) {
+  return c_v128_unziplo_32(a, b);
+}
+SIMD_INLINE v128 v128_unziphi_32(v128 a, v128 b) {
+  return c_v128_unziphi_32(a, b);
+}
+SIMD_INLINE v128 v128_unpack_u8_s16(v64 a) { return c_v128_unpack_u8_s16(a); }
+SIMD_INLINE v128 v128_unpacklo_u8_s16(v128 a) {
+  return c_v128_unpacklo_u8_s16(a);
+}
+SIMD_INLINE v128 v128_unpackhi_u8_s16(v128 a) {
+  return c_v128_unpackhi_u8_s16(a);
+}
+SIMD_INLINE v128 v128_unpack_s8_s16(v64 a) { return c_v128_unpack_s8_s16(a); }
+SIMD_INLINE v128 v128_unpacklo_s8_s16(v128 a) {
+  return c_v128_unpacklo_s8_s16(a);
+}
+SIMD_INLINE v128 v128_unpackhi_s8_s16(v128 a) {
+  return c_v128_unpackhi_s8_s16(a);
+}
+SIMD_INLINE v128 v128_pack_s32_s16(v128 a, v128 b) {
+  return c_v128_pack_s32_s16(a, b);
+}
+SIMD_INLINE v128 v128_pack_s16_u8(v128 a, v128 b) {
+  return c_v128_pack_s16_u8(a, b);
+}
+SIMD_INLINE v128 v128_pack_s16_s8(v128 a, v128 b) {
+  return c_v128_pack_s16_s8(a, b);
+}
+SIMD_INLINE v128 v128_unpack_u16_s32(v64 a) { return c_v128_unpack_u16_s32(a); }
+SIMD_INLINE v128 v128_unpack_s16_s32(v64 a) { return c_v128_unpack_s16_s32(a); }
+SIMD_INLINE v128 v128_unpacklo_u16_s32(v128 a) {
+  return c_v128_unpacklo_u16_s32(a);
+}
+SIMD_INLINE v128 v128_unpacklo_s16_s32(v128 a) {
+  return c_v128_unpacklo_s16_s32(a);
+}
+SIMD_INLINE v128 v128_unpackhi_u16_s32(v128 a) {
+  return c_v128_unpackhi_u16_s32(a);
+}
+SIMD_INLINE v128 v128_unpackhi_s16_s32(v128 a) {
+  return c_v128_unpackhi_s16_s32(a);
+}
+SIMD_INLINE v128 v128_shuffle_8(v128 a, v128 pattern) {
+  return c_v128_shuffle_8(a, pattern);
+}
+
+SIMD_INLINE v128 v128_cmpgt_s8(v128 a, v128 b) { return c_v128_cmpgt_s8(a, b); }
+SIMD_INLINE v128 v128_cmplt_s8(v128 a, v128 b) { return c_v128_cmplt_s8(a, b); }
+SIMD_INLINE v128 v128_cmpeq_8(v128 a, v128 b) { return c_v128_cmpeq_8(a, b); }
+SIMD_INLINE v128 v128_cmpgt_s16(v128 a, v128 b) {
+  return c_v128_cmpgt_s16(a, b);
+}
+SIMD_INLINE v128 v128_cmplt_s16(v128 a, v128 b) {
+  return c_v128_cmplt_s16(a, b);
+}
+SIMD_INLINE v128 v128_cmpeq_16(v128 a, v128 b) { return c_v128_cmpeq_16(a, b); }
+
+SIMD_INLINE v128 v128_shl_8(v128 a, unsigned int c) {
+  return c_v128_shl_8(a, c);
+}
+SIMD_INLINE v128 v128_shr_u8(v128 a, unsigned int c) {
+  return c_v128_shr_u8(a, c);
+}
+SIMD_INLINE v128 v128_shr_s8(v128 a, unsigned int c) {
+  return c_v128_shr_s8(a, c);
+}
+SIMD_INLINE v128 v128_shl_16(v128 a, unsigned int c) {
+  return c_v128_shl_16(a, c);
+}
+SIMD_INLINE v128 v128_shr_u16(v128 a, unsigned int c) {
+  return c_v128_shr_u16(a, c);
+}
+SIMD_INLINE v128 v128_shr_s16(v128 a, unsigned int c) {
+  return c_v128_shr_s16(a, c);
+}
+SIMD_INLINE v128 v128_shl_32(v128 a, unsigned int c) {
+  return c_v128_shl_32(a, c);
+}
+SIMD_INLINE v128 v128_shr_u32(v128 a, unsigned int c) {
+  return c_v128_shr_u32(a, c);
+}
+SIMD_INLINE v128 v128_shr_s32(v128 a, unsigned int c) {
+  return c_v128_shr_s32(a, c);
+}
+
+SIMD_INLINE v128 v128_shr_n_byte(v128 a, unsigned int n) {
+  return c_v128_shr_n_byte(a, n);
+}
+SIMD_INLINE v128 v128_shl_n_byte(v128 a, unsigned int n) {
+  return c_v128_shl_n_byte(a, n);
+}
+SIMD_INLINE v128 v128_shl_n_8(v128 a, unsigned int n) {
+  return c_v128_shl_n_8(a, n);
+}
+SIMD_INLINE v128 v128_shl_n_16(v128 a, unsigned int n) {
+  return c_v128_shl_n_16(a, n);
+}
+SIMD_INLINE v128 v128_shl_n_32(v128 a, unsigned int n) {
+  return c_v128_shl_n_32(a, n);
+}
+SIMD_INLINE v128 v128_shr_n_u8(v128 a, unsigned int n) {
+  return c_v128_shr_n_u8(a, n);
+}
+SIMD_INLINE v128 v128_shr_n_u16(v128 a, unsigned int n) {
+  return c_v128_shr_n_u16(a, n);
+}
+SIMD_INLINE v128 v128_shr_n_u32(v128 a, unsigned int n) {
+  return c_v128_shr_n_u32(a, n);
+}
+SIMD_INLINE v128 v128_shr_n_s8(v128 a, unsigned int n) {
+  return c_v128_shr_n_s8(a, n);
+}
+SIMD_INLINE v128 v128_shr_n_s16(v128 a, unsigned int n) {
+  return c_v128_shr_n_s16(a, n);
+}
+SIMD_INLINE v128 v128_shr_n_s32(v128 a, unsigned int n) {
+  return c_v128_shr_n_s32(a, n);
+}
+
+#endif /* _V128_INTRINSICS_H */
diff --git a/third_party/aom/aom_dsp/simd/v128_intrinsics_arm.h b/third_party/aom/aom_dsp/simd/v128_intrinsics_arm.h
new file mode 100644
index 0000000000..0377d4ce10
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v128_intrinsics_arm.h
@@ -0,0 +1,671 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V128_INTRINSICS_H
+#define _V128_INTRINSICS_H
+
+#include <arm_neon.h>
+#include "./v64_intrinsics_arm.h"
+
+typedef int64x2_t v128;
+
+SIMD_INLINE uint32_t v128_low_u32(v128 a) {
+  return v64_low_u32(vget_low_s64(a));
+}
+
+SIMD_INLINE v64 v128_low_v64(v128 a) { return vget_low_s64(a); }
+
+SIMD_INLINE v64 v128_high_v64(v128 a) { return vget_high_s64(a); }
+
+SIMD_INLINE v128 v128_from_v64(v64 a, v64 b) { return vcombine_s64(b, a); }
+
+SIMD_INLINE v128 v128_from_64(uint64_t a, uint64_t b) {
+  return vcombine_s64((uint64x1_t)b, (uint64x1_t)a);
+}
+
+SIMD_INLINE v128 v128_from_32(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
+  return vcombine_s64(v64_from_32(c, d), v64_from_32(a, b));
+}
+
+SIMD_INLINE v128 v128_load_aligned(const void *p) {
+  return vreinterpretq_s64_u8(vld1q_u8((const uint8_t *)p));
+}
+
+SIMD_INLINE v128 v128_load_unaligned(const void *p) {
+  return v128_load_aligned(p);
+}
+
+SIMD_INLINE void v128_store_aligned(void *p, v128 r) {
+  vst1q_u8((uint8_t *)p, vreinterpretq_u8_s64(r));
+}
+
+SIMD_INLINE void v128_store_unaligned(void *p, v128 r) {
+  vst1q_u8((uint8_t *)p, vreinterpretq_u8_s64(r));
+}
+
+SIMD_INLINE v128 v128_align(v128 a, v128 b, unsigned int c) {
+// The following functions require an immediate.
+// Some compilers will check this during optimisation, others wont.
+#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__)
+  return c ? vreinterpretq_s64_s8(
+                 vextq_s8(vreinterpretq_s8_s64(b), vreinterpretq_s8_s64(a), c))
+           : b;
+#else
+  return c < 8 ? v128_from_v64(v64_align(v128_low_v64(a), v128_high_v64(b), c),
+                               v64_align(v128_high_v64(b), v128_low_v64(b), c))
+               : v128_from_v64(
+                     v64_align(v128_high_v64(a), v128_low_v64(a), c - 8),
+                     v64_align(v128_low_v64(a), v128_high_v64(b), c - 8));
+#endif
+}
+
+SIMD_INLINE v128 v128_zero() { return vreinterpretq_s64_u8(vdupq_n_u8(0)); }
+
+SIMD_INLINE v128 v128_ones() { return vreinterpretq_s64_u8(vdupq_n_u8(-1)); }
+
+SIMD_INLINE v128 v128_dup_8(uint8_t x) {
+  return vreinterpretq_s64_u8(vdupq_n_u8(x));
+}
+
+SIMD_INLINE v128 v128_dup_16(uint16_t x) {
+  return vreinterpretq_s64_u16(vdupq_n_u16(x));
+}
+
+SIMD_INLINE v128 v128_dup_32(uint32_t x) {
+  return vreinterpretq_s64_u32(vdupq_n_u32(x));
+}
+
+SIMD_INLINE int64_t v128_dotp_s16(v128 a, v128 b) {
+  return v64_dotp_s16(vget_high_s64(a), vget_high_s64(b)) +
+         v64_dotp_s16(vget_low_s64(a), vget_low_s64(b));
+}
+
+SIMD_INLINE uint64_t v128_hadd_u8(v128 x) {
+  uint64x2_t t = vpaddlq_u32(vpaddlq_u16(vpaddlq_u8(vreinterpretq_u8_s64(x))));
+  return vget_lane_s32(
+      vreinterpret_s32_u64(vadd_u64(vget_high_u64(t), vget_low_u64(t))), 0);
+}
+
+SIMD_INLINE v128 v128_padd_s16(v128 a) {
+  return vreinterpretq_s64_s32(vpaddlq_s16(vreinterpretq_s16_s64(a)));
+}
+
+typedef struct { sad64_internal hi, lo; } sad128_internal;
+
+SIMD_INLINE sad128_internal v128_sad_u8_init() {
+  sad128_internal s;
+  s.hi = s.lo = vdupq_n_u16(0);
+  return s;
+}
+
+/* Implementation dependent return value.  Result must be finalised with
+   v128_sad_u8_sum().
+   The result for more than 32 v128_sad_u8() calls is undefined. */
+SIMD_INLINE sad128_internal v128_sad_u8(sad128_internal s, v128 a, v128 b) {
+  sad128_internal r;
+  r.hi = v64_sad_u8(s.hi, vget_high_s64(a), vget_high_s64(b));
+  r.lo = v64_sad_u8(s.lo, vget_low_s64(a), vget_low_s64(b));
+  return r;
+}
+
+SIMD_INLINE uint32_t v128_sad_u8_sum(sad128_internal s) {
+  return (uint32_t)(v64_sad_u8_sum(s.hi) + v64_sad_u8_sum(s.lo));
+}
+
+typedef struct { ssd64_internal hi, lo; } ssd128_internal;
+
+SIMD_INLINE ssd128_internal v128_ssd_u8_init() {
+  ssd128_internal s;
+  s.hi = s.lo = (ssd64_internal)(uint64_t)0;
+  return s;
+}
+
+/* Implementation dependent return value.  Result must be finalised with
+ * v128_ssd_u8_sum(). */
+SIMD_INLINE ssd128_internal v128_ssd_u8(ssd128_internal s, v128 a, v128 b) {
+  ssd128_internal r;
+  r.hi = v64_ssd_u8(s.hi, vget_high_s64(a), vget_high_s64(b));
+  r.lo = v64_ssd_u8(s.lo, vget_low_s64(a), vget_low_s64(b));
+  return r;
+}
+
+SIMD_INLINE uint32_t v128_ssd_u8_sum(ssd128_internal s) {
+  return (uint32_t)(v64_ssd_u8_sum(s.hi) + v64_ssd_u8_sum(s.lo));
+}
+
+SIMD_INLINE v128 v128_or(v128 x, v128 y) { return vorrq_s64(x, y); }
+
+SIMD_INLINE v128 v128_xor(v128 x, v128 y) { return veorq_s64(x, y); }
+
+SIMD_INLINE v128 v128_and(v128 x, v128 y) { return vandq_s64(x, y); }
+
+SIMD_INLINE v128 v128_andn(v128 x, v128 y) { return vbicq_s64(x, y); }
+
+SIMD_INLINE v128 v128_add_8(v128 x, v128 y) {
+  return vreinterpretq_s64_u8(
+      vaddq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_add_16(v128 x, v128 y) {
+  return vreinterpretq_s64_s16(
+      vaddq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_sadd_s16(v128 x, v128 y) {
+  return vreinterpretq_s64_s16(
+      vqaddq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_add_32(v128 x, v128 y) {
+  return vreinterpretq_s64_u32(
+      vaddq_u32(vreinterpretq_u32_s64(x), vreinterpretq_u32_s64(y)));
+}
+
+SIMD_INLINE v128 v128_sub_8(v128 x, v128 y) {
+  return vreinterpretq_s64_u8(
+      vsubq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_sub_16(v128 x, v128 y) {
+  return vreinterpretq_s64_s16(
+      vsubq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_ssub_s16(v128 x, v128 y) {
+  return vreinterpretq_s64_s16(
+      vqsubq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_ssub_u16(v128 x, v128 y) {
+  return vreinterpretq_s64_u16(
+      vqsubq_u16(vreinterpretq_u16_s64(x), vreinterpretq_u16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_ssub_u8(v128 x, v128 y) {
+  return vreinterpretq_s64_u8(
+      vqsubq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_ssub_s8(v128 x, v128 y) {
+  return vreinterpretq_s64_s8(
+      vqsubq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_sub_32(v128 x, v128 y) {
+  return vreinterpretq_s64_s32(
+      vsubq_s32(vreinterpretq_s32_s64(x), vreinterpretq_s32_s64(y)));
+}
+
+SIMD_INLINE v128 v128_abs_s16(v128 x) {
+  return vreinterpretq_s64_s16(vabsq_s16(vreinterpretq_s16_s64(x)));
+}
+
+SIMD_INLINE v128 v128_abs_s8(v128 x) {
+  return vreinterpretq_s64_s8(vabsq_s8(vreinterpretq_s8_s64(x)));
+}
+
+SIMD_INLINE v128 v128_mul_s16(v64 a, v64 b) {
+  return vreinterpretq_s64_s32(
+      vmull_s16(vreinterpret_s16_s64(a), vreinterpret_s16_s64(b)));
+}
+
+SIMD_INLINE v128 v128_mullo_s16(v128 a, v128 b) {
+  return vreinterpretq_s64_s16(
+      vmulq_s16(vreinterpretq_s16_s64(a), vreinterpretq_s16_s64(b)));
+}
+
+SIMD_INLINE v128 v128_mulhi_s16(v128 a, v128 b) {
+  return v128_from_v64(v64_mulhi_s16(vget_high_s64(a), vget_high_s64(b)),
+                       v64_mulhi_s16(vget_low_s64(a), vget_low_s64(b)));
+}
+
+SIMD_INLINE v128 v128_mullo_s32(v128 a, v128 b) {
+  return vreinterpretq_s64_s32(
+      vmulq_s32(vreinterpretq_s32_s64(a), vreinterpretq_s32_s64(b)));
+}
+
+SIMD_INLINE v128 v128_madd_s16(v128 a, v128 b) {
+  return v128_from_v64(v64_madd_s16(vget_high_s64(a), vget_high_s64(b)),
+                       v64_madd_s16(vget_low_s64(a), vget_low_s64(b)));
+}
+
+SIMD_INLINE v128 v128_madd_us8(v128 a, v128 b) {
+  return v128_from_v64(v64_madd_us8(vget_high_s64(a), vget_high_s64(b)),
+                       v64_madd_us8(vget_low_s64(a), vget_low_s64(b)));
+}
+
+SIMD_INLINE v128 v128_avg_u8(v128 x, v128 y) {
+  return vreinterpretq_s64_u8(
+      vrhaddq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_rdavg_u8(v128 x, v128 y) {
+  return vreinterpretq_s64_u8(
+      vhaddq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_avg_u16(v128 x, v128 y) {
+  return vreinterpretq_s64_u16(
+      vrhaddq_u16(vreinterpretq_u16_s64(x), vreinterpretq_u16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_min_u8(v128 x, v128 y) {
+  return vreinterpretq_s64_u8(
+      vminq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_max_u8(v128 x, v128 y) {
+  return vreinterpretq_s64_u8(
+      vmaxq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_min_s8(v128 x, v128 y) {
+  return vreinterpretq_s64_s8(
+      vminq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_max_s8(v128 x, v128 y) {
+  return vreinterpretq_s64_s8(
+      vmaxq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_min_s16(v128 x, v128 y) {
+  return vreinterpretq_s64_s16(
+      vminq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_max_s16(v128 x, v128 y) {
+  return vreinterpretq_s64_s16(
+      vmaxq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_ziplo_8(v128 x, v128 y) {
+  uint8x16x2_t r = vzipq_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x));
+  return vreinterpretq_s64_u8(r.val[0]);
+}
+
+SIMD_INLINE v128 v128_ziphi_8(v128 x, v128 y) {
+  uint8x16x2_t r = vzipq_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x));
+  return vreinterpretq_s64_u8(r.val[1]);
+}
+
+SIMD_INLINE v128 v128_zip_8(v64 x, v64 y) {
+  uint8x8x2_t r = vzip_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x));
+  return vreinterpretq_s64_u8(vcombine_u8(r.val[0], r.val[1]));
+}
+
+SIMD_INLINE v128 v128_ziplo_16(v128 x, v128 y) {
+  int16x8x2_t r = vzipq_s16(vreinterpretq_s16_s64(y), vreinterpretq_s16_s64(x));
+  return vreinterpretq_s64_s16(r.val[0]);
+}
+
+SIMD_INLINE v128 v128_ziphi_16(v128 x, v128 y) {
+  int16x8x2_t r = vzipq_s16(vreinterpretq_s16_s64(y), vreinterpretq_s16_s64(x));
+  return vreinterpretq_s64_s16(r.val[1]);
+}
+
+SIMD_INLINE v128 v128_zip_16(v64 x, v64 y) {
+  uint16x4x2_t r = vzip_u16(vreinterpret_u16_s64(y), vreinterpret_u16_s64(x));
+  return vreinterpretq_s64_u16(vcombine_u16(r.val[0], r.val[1]));
+}
+
+SIMD_INLINE v128 v128_ziplo_32(v128 x, v128 y) {
+  int32x4x2_t r = vzipq_s32(vreinterpretq_s32_s64(y), vreinterpretq_s32_s64(x));
+  return vreinterpretq_s64_s32(r.val[0]);
+}
+
+SIMD_INLINE v128 v128_ziphi_32(v128 x, v128 y) {
+  int32x4x2_t r = vzipq_s32(vreinterpretq_s32_s64(y), vreinterpretq_s32_s64(x));
+  return vreinterpretq_s64_s32(r.val[1]);
+}
+
+SIMD_INLINE v128 v128_zip_32(v64 x, v64 y) {
+  uint32x2x2_t r = vzip_u32(vreinterpret_u32_s64(y), vreinterpret_u32_s64(x));
+  return vreinterpretq_s64_u32(vcombine_u32(r.val[0], r.val[1]));
+}
+
+SIMD_INLINE v128 v128_ziplo_64(v128 a, v128 b) {
+  return v128_from_v64(vget_low_u64((uint64x2_t)a),
+                       vget_low_u64((uint64x2_t)b));
+}
+
+SIMD_INLINE v128 v128_ziphi_64(v128 a, v128 b) {
+  return v128_from_v64(vget_high_u64((uint64x2_t)a),
+                       vget_high_u64((uint64x2_t)b));
+}
+
+SIMD_INLINE v128 v128_unziplo_8(v128 x, v128 y) {
+  uint8x16x2_t r = vuzpq_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x));
+  return vreinterpretq_s64_u8(r.val[0]);
+}
+
+SIMD_INLINE v128 v128_unziphi_8(v128 x, v128 y) {
+  uint8x16x2_t r = vuzpq_u8(vreinterpretq_u8_s64(y), vreinterpretq_u8_s64(x));
+  return vreinterpretq_s64_u8(r.val[1]);
+}
+
+SIMD_INLINE v128 v128_unziplo_16(v128 x, v128 y) {
+  uint16x8x2_t r =
+      vuzpq_u16(vreinterpretq_u16_s64(y), vreinterpretq_u16_s64(x));
+  return vreinterpretq_s64_u16(r.val[0]);
+}
+
+SIMD_INLINE v128 v128_unziphi_16(v128 x, v128 y) {
+  uint16x8x2_t r =
+      vuzpq_u16(vreinterpretq_u16_s64(y), vreinterpretq_u16_s64(x));
+  return vreinterpretq_s64_u16(r.val[1]);
+}
+
+SIMD_INLINE v128 v128_unziplo_32(v128 x, v128 y) {
+  uint32x4x2_t r =
+      vuzpq_u32(vreinterpretq_u32_s64(y), vreinterpretq_u32_s64(x));
+  return vreinterpretq_s64_u32(r.val[0]);
+}
+
+SIMD_INLINE v128 v128_unziphi_32(v128 x, v128 y) {
+  uint32x4x2_t r =
+      vuzpq_u32(vreinterpretq_u32_s64(y), vreinterpretq_u32_s64(x));
+  return vreinterpretq_s64_u32(r.val[1]);
+}
+
+SIMD_INLINE v128 v128_unpack_u8_s16(v64 a) {
+  return vreinterpretq_s64_u16(vmovl_u8(vreinterpret_u8_s64(a)));
+}
+
+SIMD_INLINE v128 v128_unpacklo_u8_s16(v128 a) {
+  return vreinterpretq_s64_u16(vmovl_u8(vreinterpret_u8_s64(vget_low_s64(a))));
+}
+
+SIMD_INLINE v128 v128_unpackhi_u8_s16(v128 a) {
+  return vreinterpretq_s64_u16(vmovl_u8(vreinterpret_u8_s64(vget_high_s64(a))));
+}
+
+SIMD_INLINE v128 v128_unpack_s8_s16(v64 a) {
+  return vreinterpretq_s64_s16(vmovl_s8(vreinterpret_s8_s64(a)));
+}
+
+SIMD_INLINE v128 v128_unpacklo_s8_s16(v128 a) {
+  return vreinterpretq_s64_s16(vmovl_s8(vreinterpret_s8_s64(vget_low_s64(a))));
+}
+
+SIMD_INLINE v128 v128_unpackhi_s8_s16(v128 a) {
+  return vreinterpretq_s64_s16(vmovl_s8(vreinterpret_s8_s64(vget_high_s64(a))));
+}
+
+SIMD_INLINE v128 v128_pack_s32_s16(v128 a, v128 b) {
+  return v128_from_v64(
+      vreinterpret_s64_s16(vqmovn_s32(vreinterpretq_s32_s64(a))),
+      vreinterpret_s64_s16(vqmovn_s32(vreinterpretq_s32_s64(b))));
+}
+
+SIMD_INLINE v128 v128_pack_s16_u8(v128 a, v128 b) {
+  return v128_from_v64(
+      vreinterpret_s64_u8(vqmovun_s16(vreinterpretq_s16_s64(a))),
+      vreinterpret_s64_u8(vqmovun_s16(vreinterpretq_s16_s64(b))));
+}
+
+SIMD_INLINE v128 v128_pack_s16_s8(v128 a, v128 b) {
+  return v128_from_v64(
+      vreinterpret_s64_s8(vqmovn_s16(vreinterpretq_s16_s64(a))),
+      vreinterpret_s64_s8(vqmovn_s16(vreinterpretq_s16_s64(b))));
+}
+
+SIMD_INLINE v128 v128_unpack_u16_s32(v64 a) {
+  return vreinterpretq_s64_u32(vmovl_u16(vreinterpret_u16_s64(a)));
+}
+
+SIMD_INLINE v128 v128_unpack_s16_s32(v64 a) {
+  return vreinterpretq_s64_s32(vmovl_s16(vreinterpret_s16_s64(a)));
+}
+
+SIMD_INLINE v128 v128_unpacklo_u16_s32(v128 a) {
+  return vreinterpretq_s64_u32(
+      vmovl_u16(vreinterpret_u16_s64(vget_low_s64(a))));
+}
+
+SIMD_INLINE v128 v128_unpacklo_s16_s32(v128 a) {
+  return vreinterpretq_s64_s32(
+      vmovl_s16(vreinterpret_s16_s64(vget_low_s64(a))));
+}
+
+SIMD_INLINE v128 v128_unpackhi_u16_s32(v128 a) {
+  return vreinterpretq_s64_u32(
+      vmovl_u16(vreinterpret_u16_s64(vget_high_s64(a))));
+}
+
+SIMD_INLINE v128 v128_unpackhi_s16_s32(v128 a) {
+  return vreinterpretq_s64_s32(
+      vmovl_s16(vreinterpret_s16_s64(vget_high_s64(a))));
+}
+
+SIMD_INLINE v128 v128_shuffle_8(v128 x, v128 pattern) {
+  return v128_from_64(
+      (uint64_t)vreinterpret_s64_u8(
+          vtbl2_u8((uint8x8x2_t){ { vget_low_u8(vreinterpretq_u8_s64(x)),
+                                    vget_high_u8(vreinterpretq_u8_s64(x)) } },
+                   vreinterpret_u8_s64(vget_high_s64(pattern)))),
+      (uint64_t)vreinterpret_s64_u8(
+          vtbl2_u8((uint8x8x2_t){ { vget_low_u8(vreinterpretq_u8_s64(x)),
+                                    vget_high_u8(vreinterpretq_u8_s64(x)) } },
+                   vreinterpret_u8_s64(vget_low_s64(pattern)))));
+}
+
+SIMD_INLINE v128 v128_cmpgt_s8(v128 x, v128 y) {
+  return vreinterpretq_s64_u8(
+      vcgtq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_cmplt_s8(v128 x, v128 y) {
+  return vreinterpretq_s64_u8(
+      vcltq_s8(vreinterpretq_s8_s64(x), vreinterpretq_s8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_cmpeq_8(v128 x, v128 y) {
+  return vreinterpretq_s64_u8(
+      vceqq_u8(vreinterpretq_u8_s64(x), vreinterpretq_u8_s64(y)));
+}
+
+SIMD_INLINE v128 v128_cmpgt_s16(v128 x, v128 y) {
+  return vreinterpretq_s64_u16(
+      vcgtq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_cmplt_s16(v128 x, v128 y) {
+  return vreinterpretq_s64_u16(
+      vcltq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_cmpeq_16(v128 x, v128 y) {
+  return vreinterpretq_s64_u16(
+      vceqq_s16(vreinterpretq_s16_s64(x), vreinterpretq_s16_s64(y)));
+}
+
+SIMD_INLINE v128 v128_shl_8(v128 a, unsigned int c) {
+  return (c > 7) ? v128_zero() : vreinterpretq_s64_u8(vshlq_u8(
+                                     vreinterpretq_u8_s64(a), vdupq_n_s8(c)));
+}
+
+SIMD_INLINE v128 v128_shr_u8(v128 a, unsigned int c) {
+  return (c > 7) ? v128_zero() : vreinterpretq_s64_u8(vshlq_u8(
+                                     vreinterpretq_u8_s64(a), vdupq_n_s8(-c)));
+}
+
+SIMD_INLINE v128 v128_shr_s8(v128 a, unsigned int c) {
+  return (c > 7) ? v128_ones() : vreinterpretq_s64_s8(vshlq_s8(
+                                     vreinterpretq_s8_s64(a), vdupq_n_s8(-c)));
+}
+
+SIMD_INLINE v128 v128_shl_16(v128 a, unsigned int c) {
+  return (c > 15) ? v128_zero()
+                  : vreinterpretq_s64_u16(
+                        vshlq_u16(vreinterpretq_u16_s64(a), vdupq_n_s16(c)));
+}
+
+SIMD_INLINE v128 v128_shr_u16(v128 a, unsigned int c) {
+  return (c > 15) ? v128_zero()
+                  : vreinterpretq_s64_u16(
+                        vshlq_u16(vreinterpretq_u16_s64(a), vdupq_n_s16(-c)));
+}
+
+SIMD_INLINE v128 v128_shr_s16(v128 a, unsigned int c) {
+  return (c > 15) ? v128_ones()
+                  : vreinterpretq_s64_s16(
+                        vshlq_s16(vreinterpretq_s16_s64(a), vdupq_n_s16(-c)));
+}
+
+SIMD_INLINE v128 v128_shl_32(v128 a, unsigned int c) {
+  return (c > 31) ? v128_zero()
+                  : vreinterpretq_s64_u32(
+                        vshlq_u32(vreinterpretq_u32_s64(a), vdupq_n_s32(c)));
+}
+
+SIMD_INLINE v128 v128_shr_u32(v128 a, unsigned int c) {
+  return (c > 31) ? v128_zero()
+                  : vreinterpretq_s64_u32(
+                        vshlq_u32(vreinterpretq_u32_s64(a), vdupq_n_s32(-c)));
+}
+
+SIMD_INLINE v128 v128_shr_s32(v128 a, unsigned int c) {
+  return (c > 31) ? v128_ones()
+                  : vreinterpretq_s64_s32(
+                        vshlq_s32(vreinterpretq_s32_s64(a), vdupq_n_s32(-c)));
+}
+
+#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__)
+
+SIMD_INLINE v128 v128_shl_n_byte(v128 a, unsigned int n) {
+  return n < 8
+             ? v128_from_64(
+                   (uint64_t)vorr_u64(
+                       vshl_n_u64(vreinterpret_u64_s64(vget_high_s64(a)),
+                                  n * 8),
+                       vshr_n_u64(vreinterpret_u64_s64(vget_low_s64(a)),
+                                  (8 - n) * 8)),
+                   (uint64_t)vshl_n_u64(vreinterpret_u64_s64(vget_low_s64(a)),
+                                        n * 8))
+             : (n == 8 ? v128_from_64(
+                             (uint64_t)vreinterpret_u64_s64(vget_low_s64(a)), 0)
+                       : v128_from_64((uint64_t)vshl_n_u64(
+                                          vreinterpret_u64_s64(vget_low_s64(a)),
+                                          (n - 8) * 8),
+                                      0));
+}
+
+SIMD_INLINE v128 v128_shr_n_byte(v128 a, unsigned int n) {
+  return n < 8
+             ? v128_from_64(
+                   vshr_n_u64(vreinterpret_u64_s64(vget_high_s64(a)), n * 8),
+                   vorr_u64(
+                       vshr_n_u64(vreinterpret_u64_s64(vget_low_s64(a)), n * 8),
+                       vshl_n_u64(vreinterpret_u64_s64(vget_high_s64(a)),
+                                  (8 - n) * 8)))
+             : (n == 8
+                    ? v128_from_64(0, vreinterpret_u64_s64(vget_high_s64(a)))
+                    : v128_from_64(
+                          0, vshr_n_u64(vreinterpret_u64_s64(vget_high_s64(a)),
+                                        (n - 8) * 8)));
+}
+
+SIMD_INLINE v128 v128_shl_n_8(v128 a, unsigned int c) {
+  return vreinterpretq_s64_u8(vshlq_n_u8(vreinterpretq_u8_s64(a), c));
+}
+
+SIMD_INLINE v128 v128_shr_n_u8(v128 a, unsigned int c) {
+  return vreinterpretq_s64_u8(vshrq_n_u8(vreinterpretq_u8_s64(a), c));
+}
+
+SIMD_INLINE v128 v128_shr_n_s8(v128 a, unsigned int c) {
+  return vreinterpretq_s64_s8(vshrq_n_s8(vreinterpretq_s8_s64(a), c));
+}
+
+SIMD_INLINE v128 v128_shl_n_16(v128 a, unsigned int c) {
+  return vreinterpretq_s64_u16(vshlq_n_u16(vreinterpretq_u16_s64(a), c));
+}
+
+SIMD_INLINE v128 v128_shr_n_u16(v128 a, unsigned int c) {
+  return vreinterpretq_s64_u16(vshrq_n_u16(vreinterpretq_u16_s64(a), c));
+}
+
+SIMD_INLINE v128 v128_shr_n_s16(v128 a, unsigned int c) {
+  return vreinterpretq_s64_s16(vshrq_n_s16(vreinterpretq_s16_s64(a), c));
+}
+
+SIMD_INLINE v128 v128_shl_n_32(v128 a, unsigned int c) {
+  return vreinterpretq_s64_u32(vshlq_n_u32(vreinterpretq_u32_s64(a), c));
+}
+
+SIMD_INLINE v128 v128_shr_n_u32(v128 a, unsigned int c) {
+  return vreinterpretq_s64_u32(vshrq_n_u32(vreinterpretq_u32_s64(a), c));
+}
+
+SIMD_INLINE v128 v128_shr_n_s32(v128 a, unsigned int c) {
+  return vreinterpretq_s64_s32(vshrq_n_s32(vreinterpretq_s32_s64(a), c));
+}
+
+#else
+
+SIMD_INLINE v128 v128_shl_n_byte(v128 a, unsigned int n) {
+  if (n < 8)
+    return v128_from_v64(v64_or(v64_shl_n_byte(v128_high_v64(a), n),
+                                v64_shr_n_byte(v128_low_v64(a), 8 - n)),
+                         v64_shl_n_byte(v128_low_v64(a), n));
+  else
+    return v128_from_v64(v64_shl_n_byte(v128_low_v64(a), n - 8), v64_zero());
+}
+
+SIMD_INLINE v128 v128_shr_n_byte(v128 a, unsigned int n) {
+  if (n < 8)
+    return v128_from_v64(v64_shr_n_byte(v128_high_v64(a), n),
+                         v64_or(v64_shr_n_byte(v128_low_v64(a), n),
+                                v64_shl_n_byte(v128_high_v64(a), 8 - n)));
+  else
+    return v128_from_v64(v64_zero(), v64_shr_n_byte(v128_high_v64(a), n - 8));
+}
+
+SIMD_INLINE v128 v128_shl_n_8(v128 a, unsigned int c) {
+  return v128_shl_8(a, c);
+}
+
+SIMD_INLINE v128 v128_shr_n_u8(v128 a, unsigned int c) {
+  return v128_shr_u8(a, c);
+}
+
+SIMD_INLINE v128 v128_shr_n_s8(v128 a, unsigned int c) {
+  return v128_shr_s8(a, c);
+}
+
+SIMD_INLINE v128 v128_shl_n_16(v128 a, unsigned int c) {
+  return v128_shl_16(a, c);
+}
+
+SIMD_INLINE v128 v128_shr_n_u16(v128 a, unsigned int c) {
+  return v128_shr_u16(a, c);
+}
+
+SIMD_INLINE v128 v128_shr_n_s16(v128 a, unsigned int c) {
+  return v128_shr_s16(a, c);
+}
+
+SIMD_INLINE v128 v128_shl_n_32(v128 a, unsigned int c) {
+  return v128_shl_32(a, c);
+}
+
+SIMD_INLINE v128 v128_shr_n_u32(v128 a, unsigned int c) {
+  return v128_shr_u32(a, c);
+}
+
+SIMD_INLINE v128 v128_shr_n_s32(v128 a, unsigned int c) {
+  return v128_shr_s32(a, c);
+}
+
+#endif
+
+#endif /* _V128_INTRINSICS_H */
diff --git a/third_party/aom/aom_dsp/simd/v128_intrinsics_c.h b/third_party/aom/aom_dsp/simd/v128_intrinsics_c.h
new file mode 100644
index 0000000000..32e7c32de5
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v128_intrinsics_c.h
@@ -0,0 +1,707 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V128_INTRINSICS_C_H
+#define _V128_INTRINSICS_C_H
+
+#include <stdio.h>
+#include <stdlib.h>
+#include "./v64_intrinsics_c.h"
+#include "./aom_config.h"
+
+typedef union {
+  uint8_t u8[16];
+  uint16_t u16[8];
+  uint32_t u32[4];
+  uint64_t u64[2];
+  int8_t s8[16];
+  int16_t s16[8];
+  int32_t s32[4];
+  int64_t s64[2];
+  c_v64 v64[2];
+} c_v128;
+
+SIMD_INLINE uint32_t c_v128_low_u32(c_v128 a) { return a.u32[0]; }
+
+SIMD_INLINE c_v64 c_v128_low_v64(c_v128 a) { return a.v64[0]; }
+
+SIMD_INLINE c_v64 c_v128_high_v64(c_v128 a) { return a.v64[1]; }
+
+SIMD_INLINE c_v128 c_v128_from_64(uint64_t hi, uint64_t lo) {
+  c_v128 t;
+  t.u64[1] = hi;
+  t.u64[0] = lo;
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_from_v64(c_v64 hi, c_v64 lo) {
+  c_v128 t;
+  t.v64[1] = hi;
+  t.v64[0] = lo;
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_from_32(uint32_t a, uint32_t b, uint32_t c,
+                                  uint32_t d) {
+  c_v128 t;
+  t.u32[3] = a;
+  t.u32[2] = b;
+  t.u32[1] = c;
+  t.u32[0] = d;
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_load_unaligned(const void *p) {
+  c_v128 t;
+  uint8_t *pp = (uint8_t *)p;
+  uint8_t *q = (uint8_t *)&t;
+  int c;
+  for (c = 0; c < 16; c++) q[c] = pp[c];
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_load_aligned(const void *p) {
+  if (SIMD_CHECK && (uintptr_t)p & 15) {
+    fprintf(stderr, "Error: unaligned v128 load at %p\n", p);
+    abort();
+  }
+  return c_v128_load_unaligned(p);
+}
+
+SIMD_INLINE void c_v128_store_unaligned(void *p, c_v128 a) {
+  uint8_t *pp = (uint8_t *)p;
+  uint8_t *q = (uint8_t *)&a;
+  int c;
+  for (c = 0; c < 16; c++) pp[c] = q[c];
+}
+
+SIMD_INLINE void c_v128_store_aligned(void *p, c_v128 a) {
+  if (SIMD_CHECK && (uintptr_t)p & 15) {
+    fprintf(stderr, "Error: unaligned v128 store at %p\n", p);
+    abort();
+  }
+  c_v128_store_unaligned(p, a);
+}
+
+SIMD_INLINE c_v128 c_v128_zero() {
+  c_v128 t;
+  t.u64[1] = t.u64[0] = 0;
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_dup_8(uint8_t x) {
+  c_v128 t;
+  t.v64[1] = t.v64[0] = c_v64_dup_8(x);
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_dup_16(uint16_t x) {
+  c_v128 t;
+  t.v64[1] = t.v64[0] = c_v64_dup_16(x);
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_dup_32(uint32_t x) {
+  c_v128 t;
+  t.v64[1] = t.v64[0] = c_v64_dup_32(x);
+  return t;
+}
+
+SIMD_INLINE int64_t c_v128_dotp_s16(c_v128 a, c_v128 b) {
+  return c_v64_dotp_s16(a.v64[1], b.v64[1]) +
+         c_v64_dotp_s16(a.v64[0], b.v64[0]);
+}
+
+SIMD_INLINE uint64_t c_v128_hadd_u8(c_v128 a) {
+  return c_v64_hadd_u8(a.v64[1]) + c_v64_hadd_u8(a.v64[0]);
+}
+
+typedef uint32_t c_sad128_internal;
+
+SIMD_INLINE c_sad128_internal c_v128_sad_u8_init() { return 0; }
+
+/* Implementation dependent return value.  Result must be finalised with
+   v128_sad_u8_sum().
+   The result for more than 32 v128_sad_u8() calls is undefined. */
+SIMD_INLINE c_sad128_internal c_v128_sad_u8(c_sad128_internal s, c_v128 a,
+                                            c_v128 b) {
+  int c;
+  for (c = 0; c < 16; c++)
+    s += a.u8[c] > b.u8[c] ? a.u8[c] - b.u8[c] : b.u8[c] - a.u8[c];
+  return s;
+}
+
+SIMD_INLINE uint32_t c_v128_sad_u8_sum(c_sad128_internal s) { return s; }
+
+typedef uint32_t c_ssd128_internal;
+
+SIMD_INLINE c_ssd128_internal c_v128_ssd_u8_init() { return 0; }
+
+/* Implementation dependent return value.  Result must be finalised with
+ * v128_ssd_u8_sum(). */
+SIMD_INLINE c_ssd128_internal c_v128_ssd_u8(c_ssd128_internal s, c_v128 a,
+                                            c_v128 b) {
+  int c;
+  for (c = 0; c < 16; c++) s += (a.u8[c] - b.u8[c]) * (a.u8[c] - b.u8[c]);
+  return s;
+}
+
+SIMD_INLINE uint32_t c_v128_ssd_u8_sum(c_ssd128_internal s) { return s; }
+
+SIMD_INLINE c_v128 c_v128_or(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_or(a.v64[1], b.v64[1]),
+                         c_v64_or(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_xor(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_xor(a.v64[1], b.v64[1]),
+                         c_v64_xor(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_and(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_and(a.v64[1], b.v64[1]),
+                         c_v64_and(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_andn(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_andn(a.v64[1], b.v64[1]),
+                         c_v64_andn(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_add_8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_add_8(a.v64[1], b.v64[1]),
+                         c_v64_add_8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_add_16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_add_16(a.v64[1], b.v64[1]),
+                         c_v64_add_16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_sadd_s16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_sadd_s16(a.v64[1], b.v64[1]),
+                         c_v64_sadd_s16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_add_32(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_add_32(a.v64[1], b.v64[1]),
+                         c_v64_add_32(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_padd_s16(c_v128 a) {
+  c_v128 t;
+  t.s32[0] = (int32_t)a.s16[0] + (int32_t)a.s16[1];
+  t.s32[1] = (int32_t)a.s16[2] + (int32_t)a.s16[3];
+  t.s32[2] = (int32_t)a.s16[4] + (int32_t)a.s16[5];
+  t.s32[3] = (int32_t)a.s16[6] + (int32_t)a.s16[7];
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_sub_8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_sub_8(a.v64[1], b.v64[1]),
+                         c_v64_sub_8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_ssub_u8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_ssub_u8(a.v64[1], b.v64[1]),
+                         c_v64_ssub_u8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_ssub_s8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_ssub_s8(a.v64[1], b.v64[1]),
+                         c_v64_ssub_s8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_sub_16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_sub_16(a.v64[1], b.v64[1]),
+                         c_v64_sub_16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_ssub_s16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_ssub_s16(a.v64[1], b.v64[1]),
+                         c_v64_ssub_s16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_ssub_u16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_ssub_u16(a.v64[1], b.v64[1]),
+                         c_v64_ssub_u16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_sub_32(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_sub_32(a.v64[1], b.v64[1]),
+                         c_v64_sub_32(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_abs_s16(c_v128 a) {
+  return c_v128_from_v64(c_v64_abs_s16(a.v64[1]), c_v64_abs_s16(a.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_abs_s8(c_v128 a) {
+  return c_v128_from_v64(c_v64_abs_s8(a.v64[1]), c_v64_abs_s8(a.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_mul_s16(c_v64 a, c_v64 b) {
+  c_v64 lo_bits = c_v64_mullo_s16(a, b);
+  c_v64 hi_bits = c_v64_mulhi_s16(a, b);
+  return c_v128_from_v64(c_v64_ziphi_16(hi_bits, lo_bits),
+                         c_v64_ziplo_16(hi_bits, lo_bits));
+}
+
+SIMD_INLINE c_v128 c_v128_mullo_s16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_mullo_s16(a.v64[1], b.v64[1]),
+                         c_v64_mullo_s16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_mulhi_s16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_mulhi_s16(a.v64[1], b.v64[1]),
+                         c_v64_mulhi_s16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_mullo_s32(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_mullo_s32(a.v64[1], b.v64[1]),
+                         c_v64_mullo_s32(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_madd_s16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_madd_s16(a.v64[1], b.v64[1]),
+                         c_v64_madd_s16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_madd_us8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_madd_us8(a.v64[1], b.v64[1]),
+                         c_v64_madd_us8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_avg_u8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_avg_u8(a.v64[1], b.v64[1]),
+                         c_v64_avg_u8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_rdavg_u8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_rdavg_u8(a.v64[1], b.v64[1]),
+                         c_v64_rdavg_u8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_avg_u16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_avg_u16(a.v64[1], b.v64[1]),
+                         c_v64_avg_u16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_min_u8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_min_u8(a.v64[1], b.v64[1]),
+                         c_v64_min_u8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_max_u8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_max_u8(a.v64[1], b.v64[1]),
+                         c_v64_max_u8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_min_s8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_min_s8(a.v64[1], b.v64[1]),
+                         c_v64_min_s8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_max_s8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_max_s8(a.v64[1], b.v64[1]),
+                         c_v64_max_s8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_min_s16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_min_s16(a.v64[1], b.v64[1]),
+                         c_v64_min_s16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_max_s16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_max_s16(a.v64[1], b.v64[1]),
+                         c_v64_max_s16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_ziplo_8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_ziphi_8(a.v64[0], b.v64[0]),
+                         c_v64_ziplo_8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_ziphi_8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_ziphi_8(a.v64[1], b.v64[1]),
+                         c_v64_ziplo_8(a.v64[1], b.v64[1]));
+}
+
+SIMD_INLINE c_v128 c_v128_ziplo_16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_ziphi_16(a.v64[0], b.v64[0]),
+                         c_v64_ziplo_16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_ziphi_16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_ziphi_16(a.v64[1], b.v64[1]),
+                         c_v64_ziplo_16(a.v64[1], b.v64[1]));
+}
+
+SIMD_INLINE c_v128 c_v128_ziplo_32(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_ziphi_32(a.v64[0], b.v64[0]),
+                         c_v64_ziplo_32(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_ziphi_32(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_ziphi_32(a.v64[1], b.v64[1]),
+                         c_v64_ziplo_32(a.v64[1], b.v64[1]));
+}
+
+SIMD_INLINE c_v128 c_v128_ziplo_64(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(a.v64[0], b.v64[0]);
+}
+
+SIMD_INLINE c_v128 c_v128_ziphi_64(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(a.v64[1], b.v64[1]);
+}
+
+SIMD_INLINE c_v128 c_v128_zip_8(c_v64 a, c_v64 b) {
+  return c_v128_from_v64(c_v64_ziphi_8(a, b), c_v64_ziplo_8(a, b));
+}
+
+SIMD_INLINE c_v128 c_v128_zip_16(c_v64 a, c_v64 b) {
+  return c_v128_from_v64(c_v64_ziphi_16(a, b), c_v64_ziplo_16(a, b));
+}
+
+SIMD_INLINE c_v128 c_v128_zip_32(c_v64 a, c_v64 b) {
+  return c_v128_from_v64(c_v64_ziphi_32(a, b), c_v64_ziplo_32(a, b));
+}
+
+SIMD_INLINE c_v128 _c_v128_unzip_8(c_v128 a, c_v128 b, int mode) {
+  c_v128 t;
+  if (mode) {
+    t.u8[15] = b.u8[15];
+    t.u8[14] = b.u8[13];
+    t.u8[13] = b.u8[11];
+    t.u8[12] = b.u8[9];
+    t.u8[11] = b.u8[7];
+    t.u8[10] = b.u8[5];
+    t.u8[9] = b.u8[3];
+    t.u8[8] = b.u8[1];
+    t.u8[7] = a.u8[15];
+    t.u8[6] = a.u8[13];
+    t.u8[5] = a.u8[11];
+    t.u8[4] = a.u8[9];
+    t.u8[3] = a.u8[7];
+    t.u8[2] = a.u8[5];
+    t.u8[1] = a.u8[3];
+    t.u8[0] = a.u8[1];
+  } else {
+    t.u8[15] = a.u8[14];
+    t.u8[14] = a.u8[12];
+    t.u8[13] = a.u8[10];
+    t.u8[12] = a.u8[8];
+    t.u8[11] = a.u8[6];
+    t.u8[10] = a.u8[4];
+    t.u8[9] = a.u8[2];
+    t.u8[8] = a.u8[0];
+    t.u8[7] = b.u8[14];
+    t.u8[6] = b.u8[12];
+    t.u8[5] = b.u8[10];
+    t.u8[4] = b.u8[8];
+    t.u8[3] = b.u8[6];
+    t.u8[2] = b.u8[4];
+    t.u8[1] = b.u8[2];
+    t.u8[0] = b.u8[0];
+  }
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_unziplo_8(c_v128 a, c_v128 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v128_unzip_8(a, b, 1)
+                           : _c_v128_unzip_8(a, b, 0);
+}
+
+SIMD_INLINE c_v128 c_v128_unziphi_8(c_v128 a, c_v128 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v128_unzip_8(b, a, 0)
+                           : _c_v128_unzip_8(b, a, 1);
+}
+
+SIMD_INLINE c_v128 _c_v128_unzip_16(c_v128 a, c_v128 b, int mode) {
+  c_v128 t;
+  if (mode) {
+    t.u16[7] = b.u16[7];
+    t.u16[6] = b.u16[5];
+    t.u16[5] = b.u16[3];
+    t.u16[4] = b.u16[1];
+    t.u16[3] = a.u16[7];
+    t.u16[2] = a.u16[5];
+    t.u16[1] = a.u16[3];
+    t.u16[0] = a.u16[1];
+  } else {
+    t.u16[7] = a.u16[6];
+    t.u16[6] = a.u16[4];
+    t.u16[5] = a.u16[2];
+    t.u16[4] = a.u16[0];
+    t.u16[3] = b.u16[6];
+    t.u16[2] = b.u16[4];
+    t.u16[1] = b.u16[2];
+    t.u16[0] = b.u16[0];
+  }
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_unziplo_16(c_v128 a, c_v128 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v128_unzip_16(a, b, 1)
+                           : _c_v128_unzip_16(a, b, 0);
+}
+
+SIMD_INLINE c_v128 c_v128_unziphi_16(c_v128 a, c_v128 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v128_unzip_16(b, a, 0)
+                           : _c_v128_unzip_16(b, a, 1);
+}
+
+SIMD_INLINE c_v128 _c_v128_unzip_32(c_v128 a, c_v128 b, int mode) {
+  c_v128 t;
+  if (mode) {
+    t.u32[3] = b.u32[3];
+    t.u32[2] = b.u32[1];
+    t.u32[1] = a.u32[3];
+    t.u32[0] = a.u32[1];
+  } else {
+    t.u32[3] = a.u32[2];
+    t.u32[2] = a.u32[0];
+    t.u32[1] = b.u32[2];
+    t.u32[0] = b.u32[0];
+  }
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_unziplo_32(c_v128 a, c_v128 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v128_unzip_32(a, b, 1)
+                           : _c_v128_unzip_32(a, b, 0);
+}
+
+SIMD_INLINE c_v128 c_v128_unziphi_32(c_v128 a, c_v128 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v128_unzip_32(b, a, 0)
+                           : _c_v128_unzip_32(b, a, 1);
+}
+
+SIMD_INLINE c_v128 c_v128_unpack_u8_s16(c_v64 a) {
+  return c_v128_from_v64(c_v64_unpackhi_u8_s16(a), c_v64_unpacklo_u8_s16(a));
+}
+
+SIMD_INLINE c_v128 c_v128_unpacklo_u8_s16(c_v128 a) {
+  return c_v128_from_v64(c_v64_unpackhi_u8_s16(a.v64[0]),
+                         c_v64_unpacklo_u8_s16(a.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_unpackhi_u8_s16(c_v128 a) {
+  return c_v128_from_v64(c_v64_unpackhi_u8_s16(a.v64[1]),
+                         c_v64_unpacklo_u8_s16(a.v64[1]));
+}
+
+SIMD_INLINE c_v128 c_v128_unpack_s8_s16(c_v64 a) {
+  return c_v128_from_v64(c_v64_unpackhi_s8_s16(a), c_v64_unpacklo_s8_s16(a));
+}
+
+SIMD_INLINE c_v128 c_v128_unpacklo_s8_s16(c_v128 a) {
+  return c_v128_from_v64(c_v64_unpackhi_s8_s16(a.v64[0]),
+                         c_v64_unpacklo_s8_s16(a.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_unpackhi_s8_s16(c_v128 a) {
+  return c_v128_from_v64(c_v64_unpackhi_s8_s16(a.v64[1]),
+                         c_v64_unpacklo_s8_s16(a.v64[1]));
+}
+
+SIMD_INLINE c_v128 c_v128_pack_s32_s16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_pack_s32_s16(a.v64[1], a.v64[0]),
+                         c_v64_pack_s32_s16(b.v64[1], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_pack_s16_u8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_pack_s16_u8(a.v64[1], a.v64[0]),
+                         c_v64_pack_s16_u8(b.v64[1], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_pack_s16_s8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_pack_s16_s8(a.v64[1], a.v64[0]),
+                         c_v64_pack_s16_s8(b.v64[1], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_unpack_u16_s32(c_v64 a) {
+  return c_v128_from_v64(c_v64_unpackhi_u16_s32(a), c_v64_unpacklo_u16_s32(a));
+}
+
+SIMD_INLINE c_v128 c_v128_unpack_s16_s32(c_v64 a) {
+  return c_v128_from_v64(c_v64_unpackhi_s16_s32(a), c_v64_unpacklo_s16_s32(a));
+}
+
+SIMD_INLINE c_v128 c_v128_unpacklo_u16_s32(c_v128 a) {
+  return c_v128_from_v64(c_v64_unpackhi_u16_s32(a.v64[0]),
+                         c_v64_unpacklo_u16_s32(a.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_unpacklo_s16_s32(c_v128 a) {
+  return c_v128_from_v64(c_v64_unpackhi_s16_s32(a.v64[0]),
+                         c_v64_unpacklo_s16_s32(a.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_unpackhi_u16_s32(c_v128 a) {
+  return c_v128_from_v64(c_v64_unpackhi_u16_s32(a.v64[1]),
+                         c_v64_unpacklo_u16_s32(a.v64[1]));
+}
+
+SIMD_INLINE c_v128 c_v128_unpackhi_s16_s32(c_v128 a) {
+  return c_v128_from_v64(c_v64_unpackhi_s16_s32(a.v64[1]),
+                         c_v64_unpacklo_s16_s32(a.v64[1]));
+}
+
+SIMD_INLINE c_v128 c_v128_shuffle_8(c_v128 a, c_v128 pattern) {
+  c_v128 t;
+  int c;
+  for (c = 0; c < 16; c++) {
+    if (pattern.u8[c] & ~15) {
+      fprintf(stderr, "Undefined v128_shuffle_8 index %d/%d\n", pattern.u8[c],
+              c);
+      abort();
+    }
+    t.u8[c] = a.u8[CONFIG_BIG_ENDIAN ? 15 - (pattern.u8[c] & 15)
+                                     : pattern.u8[c] & 15];
+  }
+  return t;
+}
+
+SIMD_INLINE c_v128 c_v128_cmpgt_s8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_cmpgt_s8(a.v64[1], b.v64[1]),
+                         c_v64_cmpgt_s8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_cmplt_s8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_cmplt_s8(a.v64[1], b.v64[1]),
+                         c_v64_cmplt_s8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_cmpeq_8(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_cmpeq_8(a.v64[1], b.v64[1]),
+                         c_v64_cmpeq_8(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_cmpgt_s16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_cmpgt_s16(a.v64[1], b.v64[1]),
+                         c_v64_cmpgt_s16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_cmplt_s16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_cmplt_s16(a.v64[1], b.v64[1]),
+                         c_v64_cmplt_s16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_cmpeq_16(c_v128 a, c_v128 b) {
+  return c_v128_from_v64(c_v64_cmpeq_16(a.v64[1], b.v64[1]),
+                         c_v64_cmpeq_16(a.v64[0], b.v64[0]));
+}
+
+SIMD_INLINE c_v128 c_v128_shl_n_byte(c_v128 a, unsigned int n) {
+  if (n < 8)
+    return c_v128_from_v64(c_v64_or(c_v64_shl_n_byte(a.v64[1], n),
+                                    c_v64_shr_n_byte(a.v64[0], 8 - n)),
+                           c_v64_shl_n_byte(a.v64[0], n));
+  else
+    return c_v128_from_v64(c_v64_shl_n_byte(a.v64[0], n - 8), c_v64_zero());
+}
+
+SIMD_INLINE c_v128 c_v128_shr_n_byte(c_v128 a, unsigned int n) {
+  if (n < 8)
+    return c_v128_from_v64(c_v64_shr_n_byte(a.v64[1], n),
+                           c_v64_or(c_v64_shr_n_byte(a.v64[0], n),
+                                    c_v64_shl_n_byte(a.v64[1], 8 - n)));
+  else
+    return c_v128_from_v64(c_v64_zero(), c_v64_shr_n_byte(a.v64[1], n - 8));
+}
+
+SIMD_INLINE c_v128 c_v128_align(c_v128 a, c_v128 b, unsigned int c) {
+  if (SIMD_CHECK && c > 15) {
+    fprintf(stderr, "Error: undefined alignment %d\n", c);
+    abort();
+  }
+  return c ? c_v128_or(c_v128_shr_n_byte(b, c), c_v128_shl_n_byte(a, 16 - c))
+           : b;
+}
+
+SIMD_INLINE c_v128 c_v128_shl_8(c_v128 a, unsigned int c) {
+  return c_v128_from_v64(c_v64_shl_8(a.v64[1], c), c_v64_shl_8(a.v64[0], c));
+}
+
+SIMD_INLINE c_v128 c_v128_shr_u8(c_v128 a, unsigned int c) {
+  return c_v128_from_v64(c_v64_shr_u8(a.v64[1], c), c_v64_shr_u8(a.v64[0], c));
+}
+
+SIMD_INLINE c_v128 c_v128_shr_s8(c_v128 a, unsigned int c) {
+  return c_v128_from_v64(c_v64_shr_s8(a.v64[1], c), c_v64_shr_s8(a.v64[0], c));
+}
+
+SIMD_INLINE c_v128 c_v128_shl_16(c_v128 a, unsigned int c) {
+  return c_v128_from_v64(c_v64_shl_16(a.v64[1], c), c_v64_shl_16(a.v64[0], c));
+}
+
+SIMD_INLINE c_v128 c_v128_shr_u16(c_v128 a, unsigned int c) {
+  return c_v128_from_v64(c_v64_shr_u16(a.v64[1], c),
+                         c_v64_shr_u16(a.v64[0], c));
+}
+
+SIMD_INLINE c_v128 c_v128_shr_s16(c_v128 a, unsigned int c) {
+  return c_v128_from_v64(c_v64_shr_s16(a.v64[1], c),
+                         c_v64_shr_s16(a.v64[0], c));
+}
+
+SIMD_INLINE c_v128 c_v128_shl_32(c_v128 a, unsigned int c) {
+  return c_v128_from_v64(c_v64_shl_32(a.v64[1], c), c_v64_shl_32(a.v64[0], c));
+}
+
+SIMD_INLINE c_v128 c_v128_shr_u32(c_v128 a, unsigned int c) {
+  return c_v128_from_v64(c_v64_shr_u32(a.v64[1], c),
+                         c_v64_shr_u32(a.v64[0], c));
+}
+
+SIMD_INLINE c_v128 c_v128_shr_s32(c_v128 a, unsigned int c) {
+  return c_v128_from_v64(c_v64_shr_s32(a.v64[1], c),
+                         c_v64_shr_s32(a.v64[0], c));
+}
+
+SIMD_INLINE c_v128 c_v128_shl_n_8(c_v128 a, unsigned int n) {
+  return c_v128_shl_8(a, n);
+}
+
+SIMD_INLINE c_v128 c_v128_shl_n_16(c_v128 a, unsigned int n) {
+  return c_v128_shl_16(a, n);
+}
+
+SIMD_INLINE c_v128 c_v128_shl_n_32(c_v128 a, unsigned int n) {
+  return c_v128_shl_32(a, n);
+}
+
+SIMD_INLINE c_v128 c_v128_shr_n_u8(c_v128 a, unsigned int n) {
+  return c_v128_shr_u8(a, n);
+}
+
+SIMD_INLINE c_v128 c_v128_shr_n_u16(c_v128 a, unsigned int n) {
+  return c_v128_shr_u16(a, n);
+}
+
+SIMD_INLINE c_v128 c_v128_shr_n_u32(c_v128 a, unsigned int n) {
+  return c_v128_shr_u32(a, n);
+}
+
+SIMD_INLINE c_v128 c_v128_shr_n_s8(c_v128 a, unsigned int n) {
+  return c_v128_shr_s8(a, n);
+}
+
+SIMD_INLINE c_v128 c_v128_shr_n_s16(c_v128 a, unsigned int n) {
+  return c_v128_shr_s16(a, n);
+}
+
+SIMD_INLINE c_v128 c_v128_shr_n_s32(c_v128 a, unsigned int n) {
+  return c_v128_shr_s32(a, n);
+}
+
+#endif /* _V128_INTRINSICS_C_H */
diff --git a/third_party/aom/aom_dsp/simd/v128_intrinsics_x86.h b/third_party/aom/aom_dsp/simd/v128_intrinsics_x86.h
new file mode 100644
index 0000000000..cca1788d54
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v128_intrinsics_x86.h
@@ -0,0 +1,511 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V128_INTRINSICS_H
+#define _V128_INTRINSICS_H
+
+#include "./v64_intrinsics_x86.h"
+
+typedef __m128i v128;
+
+SIMD_INLINE uint32_t v128_low_u32(v128 a) {
+  return (uint32_t)_mm_cvtsi128_si32(a);
+}
+
+SIMD_INLINE v64 v128_low_v64(v128 a) {
+  return _mm_unpacklo_epi64(a, v64_zero());
+}
+
+SIMD_INLINE v64 v128_high_v64(v128 a) { return _mm_srli_si128(a, 8); }
+
+SIMD_INLINE v128 v128_from_v64(v64 a, v64 b) {
+  return _mm_unpacklo_epi64(b, a);
+}
+
+SIMD_INLINE v128 v128_from_64(uint64_t a, uint64_t b) {
+  return v128_from_v64(v64_from_64(a), v64_from_64(b));
+}
+
+SIMD_INLINE v128 v128_from_32(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
+  return _mm_set_epi32(a, b, c, d);
+}
+
+SIMD_INLINE v128 v128_load_aligned(const void *p) {
+  return _mm_load_si128((__m128i *)p);
+}
+
+SIMD_INLINE v128 v128_load_unaligned(const void *p) {
+#if defined(__SSSE3__)
+  return (__m128i)_mm_lddqu_si128((__m128i *)p);
+#else
+  return _mm_loadu_si128((__m128i *)p);
+#endif
+}
+
+SIMD_INLINE void v128_store_aligned(void *p, v128 a) {
+  _mm_store_si128((__m128i *)p, a);
+}
+
+SIMD_INLINE void v128_store_unaligned(void *p, v128 a) {
+  _mm_storeu_si128((__m128i *)p, a);
+}
+
+// The following function requires an immediate.
+// Some compilers will check this during optimisation, others wont.
+#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__)
+#if defined(__SSSE3__)
+SIMD_INLINE v128 v128_align(v128 a, v128 b, unsigned int c) {
+  return c ? _mm_alignr_epi8(a, b, c) : b;
+}
+#else
+#define v128_align(a, b, c) \
+  ((c) ? _mm_or_si128(_mm_srli_si128(b, c), _mm_slli_si128(a, 16 - (c))) : (b))
+#endif
+#else
+#if defined(__SSSE3__)
+#define v128_align(a, b, c) ((c) ? _mm_alignr_epi8(a, b, c) : (b))
+#else
+#define v128_align(a, b, c) \
+  ((c) ? _mm_or_si128(_mm_srli_si128(b, c), _mm_slli_si128(a, 16 - (c))) : (b))
+#endif
+#endif
+
+SIMD_INLINE v128 v128_zero() { return _mm_setzero_si128(); }
+
+SIMD_INLINE v128 v128_dup_8(uint8_t x) { return _mm_set1_epi8(x); }
+
+SIMD_INLINE v128 v128_dup_16(uint16_t x) { return _mm_set1_epi16(x); }
+
+SIMD_INLINE v128 v128_dup_32(uint32_t x) { return _mm_set1_epi32(x); }
+
+SIMD_INLINE v128 v128_add_8(v128 a, v128 b) { return _mm_add_epi8(a, b); }
+
+SIMD_INLINE v128 v128_add_16(v128 a, v128 b) { return _mm_add_epi16(a, b); }
+
+SIMD_INLINE v128 v128_sadd_s16(v128 a, v128 b) { return _mm_adds_epi16(a, b); }
+
+SIMD_INLINE v128 v128_add_32(v128 a, v128 b) { return _mm_add_epi32(a, b); }
+
+SIMD_INLINE v128 v128_padd_s16(v128 a) {
+  return _mm_madd_epi16(a, _mm_set1_epi16(1));
+}
+
+SIMD_INLINE v128 v128_sub_8(v128 a, v128 b) { return _mm_sub_epi8(a, b); }
+
+SIMD_INLINE v128 v128_ssub_u8(v128 a, v128 b) { return _mm_subs_epu8(a, b); }
+
+SIMD_INLINE v128 v128_ssub_s8(v128 a, v128 b) { return _mm_subs_epi8(a, b); }
+
+SIMD_INLINE v128 v128_sub_16(v128 a, v128 b) { return _mm_sub_epi16(a, b); }
+
+SIMD_INLINE v128 v128_ssub_s16(v128 a, v128 b) { return _mm_subs_epi16(a, b); }
+
+SIMD_INLINE v128 v128_ssub_u16(v128 a, v128 b) { return _mm_subs_epu16(a, b); }
+
+SIMD_INLINE v128 v128_sub_32(v128 a, v128 b) { return _mm_sub_epi32(a, b); }
+
+SIMD_INLINE v128 v128_abs_s16(v128 a) {
+#if defined(__SSSE3__)
+  return _mm_abs_epi16(a);
+#else
+  return _mm_max_epi16(a, _mm_sub_epi16(_mm_setzero_si128(), a));
+#endif
+}
+
+SIMD_INLINE v128 v128_abs_s8(v128 a) {
+#if defined(__SSSE3__)
+  return _mm_abs_epi8(a);
+#else
+  v128 sign = _mm_cmplt_epi8(a, _mm_setzero_si128());
+  return _mm_xor_si128(sign, _mm_add_epi8(a, sign));
+#endif
+}
+
+SIMD_INLINE v128 v128_ziplo_8(v128 a, v128 b) {
+  return _mm_unpacklo_epi8(b, a);
+}
+
+SIMD_INLINE v128 v128_ziphi_8(v128 a, v128 b) {
+  return _mm_unpackhi_epi8(b, a);
+}
+
+SIMD_INLINE v128 v128_ziplo_16(v128 a, v128 b) {
+  return _mm_unpacklo_epi16(b, a);
+}
+
+SIMD_INLINE v128 v128_ziphi_16(v128 a, v128 b) {
+  return _mm_unpackhi_epi16(b, a);
+}
+
+SIMD_INLINE v128 v128_ziplo_32(v128 a, v128 b) {
+  return _mm_unpacklo_epi32(b, a);
+}
+
+SIMD_INLINE v128 v128_ziphi_32(v128 a, v128 b) {
+  return _mm_unpackhi_epi32(b, a);
+}
+
+SIMD_INLINE v128 v128_ziplo_64(v128 a, v128 b) {
+  return _mm_unpacklo_epi64(b, a);
+}
+
+SIMD_INLINE v128 v128_ziphi_64(v128 a, v128 b) {
+  return _mm_unpackhi_epi64(b, a);
+}
+
+SIMD_INLINE v128 v128_zip_8(v64 a, v64 b) { return _mm_unpacklo_epi8(b, a); }
+
+SIMD_INLINE v128 v128_zip_16(v64 a, v64 b) { return _mm_unpacklo_epi16(b, a); }
+
+SIMD_INLINE v128 v128_zip_32(v64 a, v64 b) { return _mm_unpacklo_epi32(b, a); }
+
+SIMD_INLINE v128 v128_unziphi_8(v128 a, v128 b) {
+  return _mm_packs_epi16(_mm_srai_epi16(b, 8), _mm_srai_epi16(a, 8));
+}
+
+SIMD_INLINE v128 v128_unziplo_8(v128 a, v128 b) {
+#if defined(__SSSE3__)
+#ifdef __x86_64__
+  v128 order = _mm_cvtsi64_si128(0x0e0c0a0806040200LL);
+#else
+  v128 order = _mm_set_epi32(0, 0, 0x0e0c0a08, 0x06040200);
+#endif
+  return _mm_unpacklo_epi64(_mm_shuffle_epi8(b, order),
+                            _mm_shuffle_epi8(a, order));
+#else
+  return v128_unziphi_8(_mm_slli_si128(a, 1), _mm_slli_si128(b, 1));
+#endif
+}
+
+SIMD_INLINE v128 v128_unziphi_16(v128 a, v128 b) {
+  return _mm_packs_epi32(_mm_srai_epi32(b, 16), _mm_srai_epi32(a, 16));
+}
+
+SIMD_INLINE v128 v128_unziplo_16(v128 a, v128 b) {
+#if defined(__SSSE3__)
+#ifdef __x86_64__
+  v128 order = _mm_cvtsi64_si128(0x0d0c090805040100LL);
+#else
+  v128 order = _mm_set_epi32(0, 0, 0x0d0c0908, 0x05040100);
+#endif
+  return _mm_unpacklo_epi64(_mm_shuffle_epi8(b, order),
+                            _mm_shuffle_epi8(a, order));
+#else
+  return v128_unziphi_16(_mm_slli_si128(a, 2), _mm_slli_si128(b, 2));
+#endif
+}
+
+SIMD_INLINE v128 v128_unziphi_32(v128 a, v128 b) {
+  return _mm_castps_si128(_mm_shuffle_ps(
+      _mm_castsi128_ps(b), _mm_castsi128_ps(a), _MM_SHUFFLE(3, 1, 3, 1)));
+}
+
+SIMD_INLINE v128 v128_unziplo_32(v128 a, v128 b) {
+  return _mm_castps_si128(_mm_shuffle_ps(
+      _mm_castsi128_ps(b), _mm_castsi128_ps(a), _MM_SHUFFLE(2, 0, 2, 0)));
+}
+
+SIMD_INLINE v128 v128_unpack_u8_s16(v64 a) {
+  return _mm_unpacklo_epi8(a, _mm_setzero_si128());
+}
+
+SIMD_INLINE v128 v128_unpacklo_u8_s16(v128 a) {
+  return _mm_unpacklo_epi8(a, _mm_setzero_si128());
+}
+
+SIMD_INLINE v128 v128_unpackhi_u8_s16(v128 a) {
+  return _mm_unpackhi_epi8(a, _mm_setzero_si128());
+}
+
+SIMD_INLINE v128 v128_unpack_s8_s16(v64 a) {
+  return _mm_srai_epi16(_mm_unpacklo_epi8(a, a), 8);
+}
+
+SIMD_INLINE v128 v128_unpacklo_s8_s16(v128 a) {
+  return _mm_srai_epi16(_mm_unpacklo_epi8(a, a), 8);
+}
+
+SIMD_INLINE v128 v128_unpackhi_s8_s16(v128 a) {
+  return _mm_srai_epi16(_mm_unpackhi_epi8(a, a), 8);
+}
+
+SIMD_INLINE v128 v128_pack_s32_s16(v128 a, v128 b) {
+  return _mm_packs_epi32(b, a);
+}
+
+SIMD_INLINE v128 v128_pack_s16_u8(v128 a, v128 b) {
+  return _mm_packus_epi16(b, a);
+}
+
+SIMD_INLINE v128 v128_pack_s16_s8(v128 a, v128 b) {
+  return _mm_packs_epi16(b, a);
+}
+
+SIMD_INLINE v128 v128_unpack_u16_s32(v64 a) {
+  return _mm_unpacklo_epi16(a, _mm_setzero_si128());
+}
+
+SIMD_INLINE v128 v128_unpack_s16_s32(v64 a) {
+  return _mm_srai_epi32(_mm_unpacklo_epi16(a, a), 16);
+}
+
+SIMD_INLINE v128 v128_unpacklo_u16_s32(v128 a) {
+  return _mm_unpacklo_epi16(a, _mm_setzero_si128());
+}
+
+SIMD_INLINE v128 v128_unpacklo_s16_s32(v128 a) {
+  return _mm_srai_epi32(_mm_unpacklo_epi16(a, a), 16);
+}
+
+SIMD_INLINE v128 v128_unpackhi_u16_s32(v128 a) {
+  return _mm_unpackhi_epi16(a, _mm_setzero_si128());
+}
+
+SIMD_INLINE v128 v128_unpackhi_s16_s32(v128 a) {
+  return _mm_srai_epi32(_mm_unpackhi_epi16(a, a), 16);
+}
+
+SIMD_INLINE v128 v128_shuffle_8(v128 x, v128 pattern) {
+#if defined(__SSSE3__)
+  return _mm_shuffle_epi8(x, pattern);
+#else
+  v128 output;
+  unsigned char *input = (unsigned char *)&x;
+  unsigned char *index = (unsigned char *)&pattern;
+  char *selected = (char *)&output;
+  int counter;
+
+  for (counter = 0; counter < 16; counter++) {
+    selected[counter] = input[index[counter] & 15];
+  }
+
+  return output;
+#endif
+}
+
+SIMD_INLINE int64_t v128_dotp_s16(v128 a, v128 b) {
+  v128 r = _mm_madd_epi16(a, b);
+#if defined(__SSE4_1__) && defined(__x86_64__)
+  v128 c = _mm_add_epi64(_mm_cvtepi32_epi64(r),
+                         _mm_cvtepi32_epi64(_mm_srli_si128(r, 8)));
+  return _mm_cvtsi128_si64(_mm_add_epi64(c, _mm_srli_si128(c, 8)));
+#else
+  return (int64_t)_mm_cvtsi128_si32(r) +
+         (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 4)) +
+         (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 8)) +
+         (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 12));
+#endif
+}
+
+SIMD_INLINE uint64_t v128_hadd_u8(v128 a) {
+  v128 t = _mm_sad_epu8(a, _mm_setzero_si128());
+  return v64_low_u32(v128_low_v64(t)) + v64_low_u32(v128_high_v64(t));
+}
+
+typedef v128 sad128_internal;
+
+SIMD_INLINE sad128_internal v128_sad_u8_init() { return _mm_setzero_si128(); }
+
+/* Implementation dependent return value.  Result must be finalised with
+   v128_sad_sum().
+   The result for more than 32 v128_sad_u8() calls is undefined. */
+SIMD_INLINE sad128_internal v128_sad_u8(sad128_internal s, v128 a, v128 b) {
+  return _mm_add_epi64(s, _mm_sad_epu8(a, b));
+}
+
+SIMD_INLINE uint32_t v128_sad_u8_sum(sad128_internal s) {
+  return v128_low_u32(_mm_add_epi32(s, _mm_unpackhi_epi64(s, s)));
+}
+
+typedef v128 ssd128_internal;
+
+SIMD_INLINE ssd128_internal v128_ssd_u8_init() { return _mm_setzero_si128(); }
+
+/* Implementation dependent return value.  Result must be finalised with
+ * v128_ssd_sum(). */
+SIMD_INLINE ssd128_internal v128_ssd_u8(ssd128_internal s, v128 a, v128 b) {
+  v128 l = _mm_sub_epi16(_mm_unpacklo_epi8(a, _mm_setzero_si128()),
+                         _mm_unpacklo_epi8(b, _mm_setzero_si128()));
+  v128 h = _mm_sub_epi16(_mm_unpackhi_epi8(a, _mm_setzero_si128()),
+                         _mm_unpackhi_epi8(b, _mm_setzero_si128()));
+  v128 rl = _mm_madd_epi16(l, l);
+  v128 rh = _mm_madd_epi16(h, h);
+  v128 c = _mm_cvtsi32_si128(32);
+  rl = _mm_add_epi32(rl, _mm_srli_si128(rl, 8));
+  rl = _mm_add_epi32(rl, _mm_srli_si128(rl, 4));
+  rh = _mm_add_epi32(rh, _mm_srli_si128(rh, 8));
+  rh = _mm_add_epi32(rh, _mm_srli_si128(rh, 4));
+  return _mm_add_epi64(
+      s, _mm_srl_epi64(_mm_sll_epi64(_mm_unpacklo_epi64(rl, rh), c), c));
+}
+
+SIMD_INLINE uint32_t v128_ssd_u8_sum(ssd128_internal s) {
+  return v128_low_u32(_mm_add_epi32(s, _mm_unpackhi_epi64(s, s)));
+}
+
+SIMD_INLINE v128 v128_or(v128 a, v128 b) { return _mm_or_si128(a, b); }
+
+SIMD_INLINE v128 v128_xor(v128 a, v128 b) { return _mm_xor_si128(a, b); }
+
+SIMD_INLINE v128 v128_and(v128 a, v128 b) { return _mm_and_si128(a, b); }
+
+SIMD_INLINE v128 v128_andn(v128 a, v128 b) { return _mm_andnot_si128(b, a); }
+
+SIMD_INLINE v128 v128_mul_s16(v64 a, v64 b) {
+  v64 lo_bits = v64_mullo_s16(a, b);
+  v64 hi_bits = v64_mulhi_s16(a, b);
+  return v128_from_v64(v64_ziphi_16(hi_bits, lo_bits),
+                       v64_ziplo_16(hi_bits, lo_bits));
+}
+
+SIMD_INLINE v128 v128_mullo_s16(v128 a, v128 b) {
+  return _mm_mullo_epi16(a, b);
+}
+
+SIMD_INLINE v128 v128_mulhi_s16(v128 a, v128 b) {
+  return _mm_mulhi_epi16(a, b);
+}
+
+SIMD_INLINE v128 v128_mullo_s32(v128 a, v128 b) {
+#if defined(__SSE4_1__)
+  return _mm_mullo_epi32(a, b);
+#else
+  return _mm_unpacklo_epi32(
+      _mm_shuffle_epi32(_mm_mul_epu32(a, b), 8),
+      _mm_shuffle_epi32(
+          _mm_mul_epu32(_mm_srli_si128(a, 4), _mm_srli_si128(b, 4)), 8));
+#endif
+}
+
+SIMD_INLINE v128 v128_madd_s16(v128 a, v128 b) { return _mm_madd_epi16(a, b); }
+
+SIMD_INLINE v128 v128_madd_us8(v128 a, v128 b) {
+#if defined(__SSSE3__)
+  return _mm_maddubs_epi16(a, b);
+#else
+  return _mm_packs_epi32(
+      _mm_madd_epi16(_mm_unpacklo_epi8(a, _mm_setzero_si128()),
+                     _mm_srai_epi16(_mm_unpacklo_epi8(b, b), 8)),
+      _mm_madd_epi16(_mm_unpackhi_epi8(a, _mm_setzero_si128()),
+                     _mm_srai_epi16(_mm_unpackhi_epi8(b, b), 8)));
+#endif
+}
+
+SIMD_INLINE v128 v128_avg_u8(v128 a, v128 b) { return _mm_avg_epu8(a, b); }
+
+SIMD_INLINE v128 v128_rdavg_u8(v128 a, v128 b) {
+  return _mm_sub_epi8(_mm_avg_epu8(a, b),
+                      _mm_and_si128(_mm_xor_si128(a, b), v128_dup_8(1)));
+}
+
+SIMD_INLINE v128 v128_avg_u16(v128 a, v128 b) { return _mm_avg_epu16(a, b); }
+
+SIMD_INLINE v128 v128_min_u8(v128 a, v128 b) { return _mm_min_epu8(a, b); }
+
+SIMD_INLINE v128 v128_max_u8(v128 a, v128 b) { return _mm_max_epu8(a, b); }
+
+SIMD_INLINE v128 v128_min_s8(v128 a, v128 b) {
+#if defined(__SSE4_1__)
+  return _mm_min_epi8(a, b);
+#else
+  v128 mask = _mm_cmplt_epi8(a, b);
+  return _mm_or_si128(_mm_andnot_si128(mask, b), _mm_and_si128(mask, a));
+#endif
+}
+
+SIMD_INLINE v128 v128_max_s8(v128 a, v128 b) {
+#if defined(__SSE4_1__)
+  return _mm_max_epi8(a, b);
+#else
+  v128 mask = _mm_cmplt_epi8(b, a);
+  return _mm_or_si128(_mm_andnot_si128(mask, b), _mm_and_si128(mask, a));
+#endif
+}
+
+SIMD_INLINE v128 v128_min_s16(v128 a, v128 b) { return _mm_min_epi16(a, b); }
+
+SIMD_INLINE v128 v128_max_s16(v128 a, v128 b) { return _mm_max_epi16(a, b); }
+
+SIMD_INLINE v128 v128_cmpgt_s8(v128 a, v128 b) { return _mm_cmpgt_epi8(a, b); }
+
+SIMD_INLINE v128 v128_cmplt_s8(v128 a, v128 b) { return _mm_cmplt_epi8(a, b); }
+
+SIMD_INLINE v128 v128_cmpeq_8(v128 a, v128 b) { return _mm_cmpeq_epi8(a, b); }
+
+SIMD_INLINE v128 v128_cmpgt_s16(v128 a, v128 b) {
+  return _mm_cmpgt_epi16(a, b);
+}
+
+SIMD_INLINE v128 v128_cmplt_s16(v128 a, v128 b) {
+  return _mm_cmplt_epi16(a, b);
+}
+
+SIMD_INLINE v128 v128_cmpeq_16(v128 a, v128 b) { return _mm_cmpeq_epi16(a, b); }
+
+SIMD_INLINE v128 v128_shl_8(v128 a, unsigned int c) {
+  return _mm_and_si128(_mm_set1_epi8((uint8_t)(0xff << c)),
+                       _mm_sll_epi16(a, _mm_cvtsi32_si128(c)));
+}
+
+SIMD_INLINE v128 v128_shr_u8(v128 a, unsigned int c) {
+  return _mm_and_si128(_mm_set1_epi8(0xff >> c),
+                       _mm_srl_epi16(a, _mm_cvtsi32_si128(c)));
+}
+
+SIMD_INLINE v128 v128_shr_s8(v128 a, unsigned int c) {
+  __m128i x = _mm_cvtsi32_si128(c + 8);
+  return _mm_packs_epi16(_mm_sra_epi16(_mm_unpacklo_epi8(a, a), x),
+                         _mm_sra_epi16(_mm_unpackhi_epi8(a, a), x));
+}
+
+SIMD_INLINE v128 v128_shl_16(v128 a, unsigned int c) {
+  return _mm_sll_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v128 v128_shr_u16(v128 a, unsigned int c) {
+  return _mm_srl_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v128 v128_shr_s16(v128 a, unsigned int c) {
+  return _mm_sra_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v128 v128_shl_32(v128 a, unsigned int c) {
+  return _mm_sll_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v128 v128_shr_u32(v128 a, unsigned int c) {
+  return _mm_srl_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v128 v128_shr_s32(v128 a, unsigned int c) {
+  return _mm_sra_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+/* These intrinsics require immediate values, so we must use #defines
+   to enforce that. */
+#define v128_shl_n_byte(a, c) _mm_slli_si128(a, c)
+#define v128_shr_n_byte(a, c) _mm_srli_si128(a, c)
+#define v128_shl_n_8(a, c) \
+  _mm_and_si128(_mm_set1_epi8((uint8_t)(0xff << (c))), _mm_slli_epi16(a, c))
+#define v128_shr_n_u8(a, c) \
+  _mm_and_si128(_mm_set1_epi8(0xff >> (c)), _mm_srli_epi16(a, c))
+#define v128_shr_n_s8(a, c)                                         \
+  _mm_packs_epi16(_mm_srai_epi16(_mm_unpacklo_epi8(a, a), (c) + 8), \
+                  _mm_srai_epi16(_mm_unpackhi_epi8(a, a), (c) + 8))
+#define v128_shl_n_16(a, c) _mm_slli_epi16(a, c)
+#define v128_shr_n_u16(a, c) _mm_srli_epi16(a, c)
+#define v128_shr_n_s16(a, c) _mm_srai_epi16(a, c)
+#define v128_shl_n_32(a, c) _mm_slli_epi32(a, c)
+#define v128_shr_n_u32(a, c) _mm_srli_epi32(a, c)
+#define v128_shr_n_s32(a, c) _mm_srai_epi32(a, c)
+
+#endif /* _V128_INTRINSICS_H */
diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics.h b/third_party/aom/aom_dsp/simd/v256_intrinsics.h
new file mode 100644
index 0000000000..1896374ee3
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v256_intrinsics.h
@@ -0,0 +1,283 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V256_INTRINSICS_H
+#define _V256_INTRINSICS_H
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "./v256_intrinsics_c.h"
+#include "./v128_intrinsics.h"
+#include "./v64_intrinsics.h"
+
+/* Fallback to plain, unoptimised C. */
+
+typedef c_v256 v256;
+
+SIMD_INLINE uint32_t v256_low_u32(v256 a) { return c_v256_low_u32(a); }
+SIMD_INLINE v64 v256_low_v64(v256 a) { return c_v256_low_v64(a); }
+SIMD_INLINE v128 v256_low_v128(v256 a) { return c_v256_low_v128(a); }
+SIMD_INLINE v128 v256_high_v128(v256 a) { return c_v256_high_v128(a); }
+SIMD_INLINE v256 v256_from_v128(v128 hi, v128 lo) {
+  return c_v256_from_v128(hi, lo);
+}
+SIMD_INLINE v256 v256_from_64(uint64_t a, uint64_t b, uint64_t c, uint64_t d) {
+  return c_v256_from_64(a, b, c, d);
+}
+SIMD_INLINE v256 v256_from_v64(v64 a, v64 b, v64 c, v64 d) {
+  return c_v256_from_v64(a, b, c, d);
+}
+
+SIMD_INLINE v256 v256_load_unaligned(const void *p) {
+  return c_v256_load_unaligned(p);
+}
+SIMD_INLINE v256 v256_load_aligned(const void *p) {
+  return c_v256_load_aligned(p);
+}
+
+SIMD_INLINE void v256_store_unaligned(void *p, v256 a) {
+  c_v256_store_unaligned(p, a);
+}
+SIMD_INLINE void v256_store_aligned(void *p, v256 a) {
+  c_v256_store_aligned(p, a);
+}
+
+SIMD_INLINE v256 v256_align(v256 a, v256 b, unsigned int c) {
+  return c_v256_align(a, b, c);
+}
+
+SIMD_INLINE v256 v256_zero() { return c_v256_zero(); }
+SIMD_INLINE v256 v256_dup_8(uint8_t x) { return c_v256_dup_8(x); }
+SIMD_INLINE v256 v256_dup_16(uint16_t x) { return c_v256_dup_16(x); }
+SIMD_INLINE v256 v256_dup_32(uint32_t x) { return c_v256_dup_32(x); }
+
+typedef uint32_t sad256_internal;
+SIMD_INLINE sad256_internal v256_sad_u8_init() { return c_v256_sad_u8_init(); }
+SIMD_INLINE sad256_internal v256_sad_u8(sad256_internal s, v256 a, v256 b) {
+  return c_v256_sad_u8(s, a, b);
+}
+SIMD_INLINE uint32_t v256_sad_u8_sum(sad256_internal s) {
+  return c_v256_sad_u8_sum(s);
+}
+typedef uint32_t ssd256_internal;
+SIMD_INLINE ssd256_internal v256_ssd_u8_init() { return c_v256_ssd_u8_init(); }
+SIMD_INLINE ssd256_internal v256_ssd_u8(ssd256_internal s, v256 a, v256 b) {
+  return c_v256_ssd_u8(s, a, b);
+}
+SIMD_INLINE uint32_t v256_ssd_u8_sum(ssd256_internal s) {
+  return c_v256_ssd_u8_sum(s);
+}
+SIMD_INLINE int64_t v256_dotp_s16(v256 a, v256 b) {
+  return c_v256_dotp_s16(a, b);
+}
+SIMD_INLINE uint64_t v256_hadd_u8(v256 a) { return c_v256_hadd_u8(a); }
+
+SIMD_INLINE v256 v256_or(v256 a, v256 b) { return c_v256_or(a, b); }
+SIMD_INLINE v256 v256_xor(v256 a, v256 b) { return c_v256_xor(a, b); }
+SIMD_INLINE v256 v256_and(v256 a, v256 b) { return c_v256_and(a, b); }
+SIMD_INLINE v256 v256_andn(v256 a, v256 b) { return c_v256_andn(a, b); }
+
+SIMD_INLINE v256 v256_add_8(v256 a, v256 b) { return c_v256_add_8(a, b); }
+SIMD_INLINE v256 v256_add_16(v256 a, v256 b) { return c_v256_add_16(a, b); }
+SIMD_INLINE v256 v256_sadd_s16(v256 a, v256 b) { return c_v256_sadd_s16(a, b); }
+SIMD_INLINE v256 v256_add_32(v256 a, v256 b) { return c_v256_add_32(a, b); }
+SIMD_INLINE v256 v256_padd_s16(v256 a) { return c_v256_padd_s16(a); }
+SIMD_INLINE v256 v256_sub_8(v256 a, v256 b) { return c_v256_sub_8(a, b); }
+SIMD_INLINE v256 v256_ssub_u8(v256 a, v256 b) { return c_v256_ssub_u8(a, b); }
+SIMD_INLINE v256 v256_ssub_s8(v256 a, v256 b) { return c_v256_ssub_s8(a, b); }
+SIMD_INLINE v256 v256_sub_16(v256 a, v256 b) { return c_v256_sub_16(a, b); }
+SIMD_INLINE v256 v256_ssub_s16(v256 a, v256 b) { return c_v256_ssub_s16(a, b); }
+SIMD_INLINE v256 v256_ssub_u16(v256 a, v256 b) { return c_v256_ssub_u16(a, b); }
+SIMD_INLINE v256 v256_sub_32(v256 a, v256 b) { return c_v256_sub_32(a, b); }
+SIMD_INLINE v256 v256_abs_s16(v256 a) { return c_v256_abs_s16(a); }
+SIMD_INLINE v256 v256_abs_s8(v256 a) { return c_v256_abs_s8(a); }
+
+SIMD_INLINE v256 v256_mul_s16(v128 a, v128 b) { return c_v256_mul_s16(a, b); }
+SIMD_INLINE v256 v256_mullo_s16(v256 a, v256 b) {
+  return c_v256_mullo_s16(a, b);
+}
+SIMD_INLINE v256 v256_mulhi_s16(v256 a, v256 b) {
+  return c_v256_mulhi_s16(a, b);
+}
+SIMD_INLINE v256 v256_mullo_s32(v256 a, v256 b) {
+  return c_v256_mullo_s32(a, b);
+}
+SIMD_INLINE v256 v256_madd_s16(v256 a, v256 b) { return c_v256_madd_s16(a, b); }
+SIMD_INLINE v256 v256_madd_us8(v256 a, v256 b) { return c_v256_madd_us8(a, b); }
+
+SIMD_INLINE v256 v256_avg_u8(v256 a, v256 b) { return c_v256_avg_u8(a, b); }
+SIMD_INLINE v256 v256_rdavg_u8(v256 a, v256 b) { return c_v256_rdavg_u8(a, b); }
+SIMD_INLINE v256 v256_avg_u16(v256 a, v256 b) { return c_v256_avg_u16(a, b); }
+SIMD_INLINE v256 v256_min_u8(v256 a, v256 b) { return c_v256_min_u8(a, b); }
+SIMD_INLINE v256 v256_max_u8(v256 a, v256 b) { return c_v256_max_u8(a, b); }
+SIMD_INLINE v256 v256_min_s8(v256 a, v256 b) { return c_v256_min_s8(a, b); }
+SIMD_INLINE v256 v256_max_s8(v256 a, v256 b) { return c_v256_max_s8(a, b); }
+SIMD_INLINE v256 v256_min_s16(v256 a, v256 b) { return c_v256_min_s16(a, b); }
+SIMD_INLINE v256 v256_max_s16(v256 a, v256 b) { return c_v256_max_s16(a, b); }
+
+SIMD_INLINE v256 v256_ziplo_8(v256 a, v256 b) { return c_v256_ziplo_8(a, b); }
+SIMD_INLINE v256 v256_ziphi_8(v256 a, v256 b) { return c_v256_ziphi_8(a, b); }
+SIMD_INLINE v256 v256_ziplo_16(v256 a, v256 b) { return c_v256_ziplo_16(a, b); }
+SIMD_INLINE v256 v256_ziphi_16(v256 a, v256 b) { return c_v256_ziphi_16(a, b); }
+SIMD_INLINE v256 v256_ziplo_32(v256 a, v256 b) { return c_v256_ziplo_32(a, b); }
+SIMD_INLINE v256 v256_ziphi_32(v256 a, v256 b) { return c_v256_ziphi_32(a, b); }
+SIMD_INLINE v256 v256_ziplo_64(v256 a, v256 b) { return c_v256_ziplo_64(a, b); }
+SIMD_INLINE v256 v256_ziphi_64(v256 a, v256 b) { return c_v256_ziphi_64(a, b); }
+SIMD_INLINE v256 v256_ziplo_128(v256 a, v256 b) {
+  return c_v256_ziplo_128(a, b);
+}
+SIMD_INLINE v256 v256_ziphi_128(v256 a, v256 b) {
+  return c_v256_ziphi_128(a, b);
+}
+SIMD_INLINE v256 v256_zip_8(v128 a, v128 b) { return c_v256_zip_8(a, b); }
+SIMD_INLINE v256 v256_zip_16(v128 a, v128 b) { return c_v256_zip_16(a, b); }
+SIMD_INLINE v256 v256_zip_32(v128 a, v128 b) { return c_v256_zip_32(a, b); }
+SIMD_INLINE v256 v256_unziplo_8(v256 a, v256 b) {
+  return c_v256_unziplo_8(a, b);
+}
+SIMD_INLINE v256 v256_unziphi_8(v256 a, v256 b) {
+  return c_v256_unziphi_8(a, b);
+}
+SIMD_INLINE v256 v256_unziplo_16(v256 a, v256 b) {
+  return c_v256_unziplo_16(a, b);
+}
+SIMD_INLINE v256 v256_unziphi_16(v256 a, v256 b) {
+  return c_v256_unziphi_16(a, b);
+}
+SIMD_INLINE v256 v256_unziplo_32(v256 a, v256 b) {
+  return c_v256_unziplo_32(a, b);
+}
+SIMD_INLINE v256 v256_unziphi_32(v256 a, v256 b) {
+  return c_v256_unziphi_32(a, b);
+}
+SIMD_INLINE v256 v256_unpack_u8_s16(v128 a) { return c_v256_unpack_u8_s16(a); }
+SIMD_INLINE v256 v256_unpacklo_u8_s16(v256 a) {
+  return c_v256_unpacklo_u8_s16(a);
+}
+SIMD_INLINE v256 v256_unpackhi_u8_s16(v256 a) {
+  return c_v256_unpackhi_u8_s16(a);
+}
+SIMD_INLINE v256 v256_unpack_s8_s16(v128 a) { return c_v256_unpack_s8_s16(a); }
+SIMD_INLINE v256 v256_unpacklo_s8_s16(v256 a) {
+  return c_v256_unpacklo_s8_s16(a);
+}
+SIMD_INLINE v256 v256_unpackhi_s8_s16(v256 a) {
+  return c_v256_unpackhi_s8_s16(a);
+}
+SIMD_INLINE v256 v256_pack_s32_s16(v256 a, v256 b) {
+  return c_v256_pack_s32_s16(a, b);
+}
+SIMD_INLINE v256 v256_pack_s16_u8(v256 a, v256 b) {
+  return c_v256_pack_s16_u8(a, b);
+}
+SIMD_INLINE v256 v256_pack_s16_s8(v256 a, v256 b) {
+  return c_v256_pack_s16_s8(a, b);
+}
+SIMD_INLINE v256 v256_unpack_u16_s32(v128 a) {
+  return c_v256_unpack_u16_s32(a);
+}
+SIMD_INLINE v256 v256_unpack_s16_s32(v128 a) {
+  return c_v256_unpack_s16_s32(a);
+}
+SIMD_INLINE v256 v256_unpacklo_u16_s32(v256 a) {
+  return c_v256_unpacklo_u16_s32(a);
+}
+SIMD_INLINE v256 v256_unpacklo_s16_s32(v256 a) {
+  return c_v256_unpacklo_s16_s32(a);
+}
+SIMD_INLINE v256 v256_unpackhi_u16_s32(v256 a) {
+  return c_v256_unpackhi_u16_s32(a);
+}
+SIMD_INLINE v256 v256_unpackhi_s16_s32(v256 a) {
+  return c_v256_unpackhi_s16_s32(a);
+}
+SIMD_INLINE v256 v256_shuffle_8(v256 a, v256 pattern) {
+  return c_v256_shuffle_8(a, pattern);
+}
+SIMD_INLINE v256 v256_pshuffle_8(v256 a, v256 pattern) {
+  return c_v256_pshuffle_8(a, pattern);
+}
+
+SIMD_INLINE v256 v256_cmpgt_s8(v256 a, v256 b) { return c_v256_cmpgt_s8(a, b); }
+SIMD_INLINE v256 v256_cmplt_s8(v256 a, v256 b) { return c_v256_cmplt_s8(a, b); }
+SIMD_INLINE v256 v256_cmpeq_8(v256 a, v256 b) { return c_v256_cmpeq_8(a, b); }
+SIMD_INLINE v256 v256_cmpgt_s16(v256 a, v256 b) {
+  return c_v256_cmpgt_s16(a, b);
+}
+SIMD_INLINE v256 v256_cmplt_s16(v256 a, v256 b) {
+  return c_v256_cmplt_s16(a, b);
+}
+SIMD_INLINE v256 v256_cmpeq_16(v256 a, v256 b) { return c_v256_cmpeq_16(a, b); }
+
+SIMD_INLINE v256 v256_shl_8(v256 a, unsigned int c) {
+  return c_v256_shl_8(a, c);
+}
+SIMD_INLINE v256 v256_shr_u8(v256 a, unsigned int c) {
+  return c_v256_shr_u8(a, c);
+}
+SIMD_INLINE v256 v256_shr_s8(v256 a, unsigned int c) {
+  return c_v256_shr_s8(a, c);
+}
+SIMD_INLINE v256 v256_shl_16(v256 a, unsigned int c) {
+  return c_v256_shl_16(a, c);
+}
+SIMD_INLINE v256 v256_shr_u16(v256 a, unsigned int c) {
+  return c_v256_shr_u16(a, c);
+}
+SIMD_INLINE v256 v256_shr_s16(v256 a, unsigned int c) {
+  return c_v256_shr_s16(a, c);
+}
+SIMD_INLINE v256 v256_shl_32(v256 a, unsigned int c) {
+  return c_v256_shl_32(a, c);
+}
+SIMD_INLINE v256 v256_shr_u32(v256 a, unsigned int c) {
+  return c_v256_shr_u32(a, c);
+}
+SIMD_INLINE v256 v256_shr_s32(v256 a, unsigned int c) {
+  return c_v256_shr_s32(a, c);
+}
+
+SIMD_INLINE v256 v256_shr_n_byte(v256 a, unsigned int n) {
+  return c_v256_shr_n_byte(a, n);
+}
+SIMD_INLINE v256 v256_shl_n_byte(v256 a, unsigned int n) {
+  return c_v256_shl_n_byte(a, n);
+}
+SIMD_INLINE v256 v256_shl_n_8(v256 a, unsigned int n) {
+  return c_v256_shl_n_8(a, n);
+}
+SIMD_INLINE v256 v256_shl_n_16(v256 a, unsigned int n) {
+  return c_v256_shl_n_16(a, n);
+}
+SIMD_INLINE v256 v256_shl_n_32(v256 a, unsigned int n) {
+  return c_v256_shl_n_32(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_u8(v256 a, unsigned int n) {
+  return c_v256_shr_n_u8(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_u16(v256 a, unsigned int n) {
+  return c_v256_shr_n_u16(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_u32(v256 a, unsigned int n) {
+  return c_v256_shr_n_u32(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_s8(v256 a, unsigned int n) {
+  return c_v256_shr_n_s8(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_s16(v256 a, unsigned int n) {
+  return c_v256_shr_n_s16(a, n);
+}
+SIMD_INLINE v256 v256_shr_n_s32(v256 a, unsigned int n) {
+  return c_v256_shr_n_s32(a, n);
+}
+
+#endif /* _V256_INTRINSICS_H */
diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics_arm.h b/third_party/aom/aom_dsp/simd/v256_intrinsics_arm.h
new file mode 100644
index 0000000000..ba4ed719df
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v256_intrinsics_arm.h
@@ -0,0 +1,17 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V256_INTRINSICS_H
+#define _V256_INTRINSICS_H
+
+#include "./v256_intrinsics_v128.h"
+
+#endif /* _V256_INTRINSICS_H */
diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics_c.h b/third_party/aom/aom_dsp/simd/v256_intrinsics_c.h
new file mode 100644
index 0000000000..f96ca7fa6a
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v256_intrinsics_c.h
@@ -0,0 +1,724 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V256_INTRINSICS_C_H
+#define _V256_INTRINSICS_C_H
+
+#include <stdio.h>
+#include <stdlib.h>
+#include "./v128_intrinsics_c.h"
+#include "./aom_config.h"
+
+typedef union {
+  uint8_t u8[32];
+  uint16_t u16[16];
+  uint32_t u32[8];
+  uint64_t u64[4];
+  int8_t s8[32];
+  int16_t s16[16];
+  int32_t s32[8];
+  int64_t s64[4];
+  c_v64 v64[4];
+  c_v128 v128[2];
+} c_v256;
+
+SIMD_INLINE uint32_t c_v256_low_u32(c_v256 a) { return a.u32[0]; }
+
+SIMD_INLINE c_v64 c_v256_low_v64(c_v256 a) { return a.v64[0]; }
+
+SIMD_INLINE c_v128 c_v256_low_v128(c_v256 a) { return a.v128[0]; }
+
+SIMD_INLINE c_v128 c_v256_high_v128(c_v256 a) { return a.v128[1]; }
+
+SIMD_INLINE c_v256 c_v256_from_v128(c_v128 hi, c_v128 lo) {
+  c_v256 t;
+  t.v128[1] = hi;
+  t.v128[0] = lo;
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_from_64(uint64_t a, uint64_t b, uint64_t c,
+                                  uint64_t d) {
+  c_v256 t;
+  t.u64[3] = a;
+  t.u64[2] = b;
+  t.u64[1] = c;
+  t.u64[0] = d;
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_from_v64(c_v64 a, c_v64 b, c_v64 c, c_v64 d) {
+  c_v256 t;
+  t.u64[3] = a.u64;
+  t.u64[2] = b.u64;
+  t.u64[1] = c.u64;
+  t.u64[0] = d.u64;
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_load_unaligned(const void *p) {
+  c_v256 t;
+  uint8_t *pp = (uint8_t *)p;
+  uint8_t *q = (uint8_t *)&t;
+  int c;
+  for (c = 0; c < 32; c++) q[c] = pp[c];
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_load_aligned(const void *p) {
+  if (SIMD_CHECK && (uintptr_t)p & 31) {
+    fprintf(stderr, "Error: unaligned v256 load at %p\n", p);
+    abort();
+  }
+  return c_v256_load_unaligned(p);
+}
+
+SIMD_INLINE void c_v256_store_unaligned(void *p, c_v256 a) {
+  uint8_t *pp = (uint8_t *)p;
+  uint8_t *q = (uint8_t *)&a;
+  int c;
+  for (c = 0; c < 32; c++) pp[c] = q[c];
+}
+
+SIMD_INLINE void c_v256_store_aligned(void *p, c_v256 a) {
+  if (SIMD_CHECK && (uintptr_t)p & 31) {
+    fprintf(stderr, "Error: unaligned v256 store at %p\n", p);
+    abort();
+  }
+  c_v256_store_unaligned(p, a);
+}
+
+SIMD_INLINE c_v256 c_v256_zero() {
+  c_v256 t;
+  t.u64[3] = t.u64[2] = t.u64[1] = t.u64[0] = 0;
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_dup_8(uint8_t x) {
+  c_v256 t;
+  t.v64[3] = t.v64[2] = t.v64[1] = t.v64[0] = c_v64_dup_8(x);
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_dup_16(uint16_t x) {
+  c_v256 t;
+  t.v64[3] = t.v64[2] = t.v64[1] = t.v64[0] = c_v64_dup_16(x);
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_dup_32(uint32_t x) {
+  c_v256 t;
+  t.v64[3] = t.v64[2] = t.v64[1] = t.v64[0] = c_v64_dup_32(x);
+  return t;
+}
+
+SIMD_INLINE int64_t c_v256_dotp_s16(c_v256 a, c_v256 b) {
+  return c_v128_dotp_s16(a.v128[1], b.v128[1]) +
+         c_v128_dotp_s16(a.v128[0], b.v128[0]);
+}
+
+SIMD_INLINE uint64_t c_v256_hadd_u8(c_v256 a) {
+  return c_v128_hadd_u8(a.v128[1]) + c_v128_hadd_u8(a.v128[0]);
+}
+
+typedef uint32_t c_sad256_internal;
+
+SIMD_INLINE c_sad128_internal c_v256_sad_u8_init() { return 0; }
+
+/* Implementation dependent return value.  Result must be finalised with
+   v256_sad_u8_sum().
+   The result for more than 16 v256_sad_u8() calls is undefined. */
+SIMD_INLINE c_sad128_internal c_v256_sad_u8(c_sad256_internal s, c_v256 a,
+                                            c_v256 b) {
+  int c;
+  for (c = 0; c < 32; c++)
+    s += a.u8[c] > b.u8[c] ? a.u8[c] - b.u8[c] : b.u8[c] - a.u8[c];
+  return s;
+}
+
+SIMD_INLINE uint32_t c_v256_sad_u8_sum(c_sad256_internal s) { return s; }
+
+typedef uint32_t c_ssd256_internal;
+
+SIMD_INLINE c_ssd256_internal c_v256_ssd_u8_init() { return 0; }
+
+/* Implementation dependent return value.  Result must be finalised with
+ * v256_ssd_u8_sum(). */
+SIMD_INLINE c_ssd256_internal c_v256_ssd_u8(c_ssd256_internal s, c_v256 a,
+                                            c_v256 b) {
+  int c;
+  for (c = 0; c < 32; c++) s += (a.u8[c] - b.u8[c]) * (a.u8[c] - b.u8[c]);
+  return s;
+}
+
+SIMD_INLINE uint32_t c_v256_ssd_u8_sum(c_ssd256_internal s) { return s; }
+
+SIMD_INLINE c_v256 c_v256_or(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_or(a.v128[1], b.v128[1]),
+                          c_v128_or(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_xor(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_xor(a.v128[1], b.v128[1]),
+                          c_v128_xor(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_and(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_and(a.v128[1], b.v128[1]),
+                          c_v128_and(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_andn(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_andn(a.v128[1], b.v128[1]),
+                          c_v128_andn(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_add_8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_add_8(a.v128[1], b.v128[1]),
+                          c_v128_add_8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_add_16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_add_16(a.v128[1], b.v128[1]),
+                          c_v128_add_16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_sadd_s16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_sadd_s16(a.v128[1], b.v128[1]),
+                          c_v128_sadd_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_add_32(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_add_32(a.v128[1], b.v128[1]),
+                          c_v128_add_32(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_padd_s16(c_v256 a) {
+  c_v256 t;
+  t.s32[0] = (int32_t)a.s16[0] + (int32_t)a.s16[1];
+  t.s32[1] = (int32_t)a.s16[2] + (int32_t)a.s16[3];
+  t.s32[2] = (int32_t)a.s16[4] + (int32_t)a.s16[5];
+  t.s32[3] = (int32_t)a.s16[6] + (int32_t)a.s16[7];
+  t.s32[4] = (int32_t)a.s16[8] + (int32_t)a.s16[9];
+  t.s32[5] = (int32_t)a.s16[10] + (int32_t)a.s16[11];
+  t.s32[6] = (int32_t)a.s16[12] + (int32_t)a.s16[13];
+  t.s32[7] = (int32_t)a.s16[14] + (int32_t)a.s16[15];
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_sub_8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_sub_8(a.v128[1], b.v128[1]),
+                          c_v128_sub_8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ssub_u8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ssub_u8(a.v128[1], b.v128[1]),
+                          c_v128_ssub_u8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ssub_s8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ssub_s8(a.v128[1], b.v128[1]),
+                          c_v128_ssub_s8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_sub_16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_sub_16(a.v128[1], b.v128[1]),
+                          c_v128_sub_16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ssub_s16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ssub_s16(a.v128[1], b.v128[1]),
+                          c_v128_ssub_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ssub_u16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ssub_u16(a.v128[1], b.v128[1]),
+                          c_v128_ssub_u16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_sub_32(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_sub_32(a.v128[1], b.v128[1]),
+                          c_v128_sub_32(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_abs_s16(c_v256 a) {
+  return c_v256_from_v128(c_v128_abs_s16(a.v128[1]), c_v128_abs_s16(a.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_abs_s8(c_v256 a) {
+  return c_v256_from_v128(c_v128_abs_s8(a.v128[1]), c_v128_abs_s8(a.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_mul_s16(c_v128 a, c_v128 b) {
+  c_v128 lo_bits = c_v128_mullo_s16(a, b);
+  c_v128 hi_bits = c_v128_mulhi_s16(a, b);
+  return c_v256_from_v128(c_v128_ziphi_16(hi_bits, lo_bits),
+                          c_v128_ziplo_16(hi_bits, lo_bits));
+}
+
+SIMD_INLINE c_v256 c_v256_mullo_s16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_mullo_s16(a.v128[1], b.v128[1]),
+                          c_v128_mullo_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_mulhi_s16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_mulhi_s16(a.v128[1], b.v128[1]),
+                          c_v128_mulhi_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_mullo_s32(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_mullo_s32(a.v128[1], b.v128[1]),
+                          c_v128_mullo_s32(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_madd_s16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_madd_s16(a.v128[1], b.v128[1]),
+                          c_v128_madd_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_madd_us8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_madd_us8(a.v128[1], b.v128[1]),
+                          c_v128_madd_us8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_avg_u8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_avg_u8(a.v128[1], b.v128[1]),
+                          c_v128_avg_u8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_rdavg_u8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_rdavg_u8(a.v128[1], b.v128[1]),
+                          c_v128_rdavg_u8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_avg_u16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_avg_u16(a.v128[1], b.v128[1]),
+                          c_v128_avg_u16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_min_u8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_min_u8(a.v128[1], b.v128[1]),
+                          c_v128_min_u8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_max_u8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_max_u8(a.v128[1], b.v128[1]),
+                          c_v128_max_u8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_min_s8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_min_s8(a.v128[1], b.v128[1]),
+                          c_v128_min_s8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_max_s8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_max_s8(a.v128[1], b.v128[1]),
+                          c_v128_max_s8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_min_s16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_min_s16(a.v128[1], b.v128[1]),
+                          c_v128_min_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_max_s16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_max_s16(a.v128[1], b.v128[1]),
+                          c_v128_max_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziplo_8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ziphi_8(a.v128[0], b.v128[0]),
+                          c_v128_ziplo_8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziphi_8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ziphi_8(a.v128[1], b.v128[1]),
+                          c_v128_ziplo_8(a.v128[1], b.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziplo_16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ziphi_16(a.v128[0], b.v128[0]),
+                          c_v128_ziplo_16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziphi_16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ziphi_16(a.v128[1], b.v128[1]),
+                          c_v128_ziplo_16(a.v128[1], b.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziplo_32(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ziphi_32(a.v128[0], b.v128[0]),
+                          c_v128_ziplo_32(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziphi_32(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ziphi_32(a.v128[1], b.v128[1]),
+                          c_v128_ziplo_32(a.v128[1], b.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziplo_64(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ziphi_64(a.v128[0], b.v128[0]),
+                          c_v128_ziplo_64(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziphi_64(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_ziphi_64(a.v128[1], b.v128[1]),
+                          c_v128_ziplo_64(a.v128[1], b.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_ziplo_128(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(a.v128[0], b.v128[0]);
+}
+
+SIMD_INLINE c_v256 c_v256_ziphi_128(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(a.v128[1], b.v128[1]);
+}
+
+SIMD_INLINE c_v256 c_v256_zip_8(c_v128 a, c_v128 b) {
+  return c_v256_from_v128(c_v128_ziphi_8(a, b), c_v128_ziplo_8(a, b));
+}
+
+SIMD_INLINE c_v256 c_v256_zip_16(c_v128 a, c_v128 b) {
+  return c_v256_from_v128(c_v128_ziphi_16(a, b), c_v128_ziplo_16(a, b));
+}
+
+SIMD_INLINE c_v256 c_v256_zip_32(c_v128 a, c_v128 b) {
+  return c_v256_from_v128(c_v128_ziphi_32(a, b), c_v128_ziplo_32(a, b));
+}
+
+SIMD_INLINE c_v256 _c_v256_unzip_8(c_v256 a, c_v256 b, int mode) {
+  c_v256 t;
+  int i;
+  if (mode) {
+    for (i = 0; i < 16; i++) {
+      t.u8[i] = a.u8[i * 2 + 1];
+      t.u8[i + 16] = b.u8[i * 2 + 1];
+    }
+  } else {
+    for (i = 0; i < 16; i++) {
+      t.u8[i] = b.u8[i * 2];
+      t.u8[i + 16] = a.u8[i * 2];
+    }
+  }
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_unziplo_8(c_v256 a, c_v256 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v256_unzip_8(a, b, 1)
+                           : _c_v256_unzip_8(a, b, 0);
+}
+
+SIMD_INLINE c_v256 c_v256_unziphi_8(c_v256 a, c_v256 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v256_unzip_8(b, a, 0)
+                           : _c_v256_unzip_8(b, a, 1);
+}
+
+SIMD_INLINE c_v256 _c_v256_unzip_16(c_v256 a, c_v256 b, int mode) {
+  c_v256 t;
+  int i;
+  if (mode) {
+    for (i = 0; i < 8; i++) {
+      t.u16[i] = a.u16[i * 2 + 1];
+      t.u16[i + 8] = b.u16[i * 2 + 1];
+    }
+  } else {
+    for (i = 0; i < 8; i++) {
+      t.u16[i] = b.u16[i * 2];
+      t.u16[i + 8] = a.u16[i * 2];
+    }
+  }
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_unziplo_16(c_v256 a, c_v256 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v256_unzip_16(a, b, 1)
+                           : _c_v256_unzip_16(a, b, 0);
+}
+
+SIMD_INLINE c_v256 c_v256_unziphi_16(c_v256 a, c_v256 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v256_unzip_16(b, a, 0)
+                           : _c_v256_unzip_16(b, a, 1);
+}
+
+SIMD_INLINE c_v256 _c_v256_unzip_32(c_v256 a, c_v256 b, int mode) {
+  c_v256 t;
+  if (mode) {
+    t.u32[7] = b.u32[7];
+    t.u32[6] = b.u32[5];
+    t.u32[5] = b.u32[3];
+    t.u32[4] = b.u32[1];
+    t.u32[3] = a.u32[7];
+    t.u32[2] = a.u32[5];
+    t.u32[1] = a.u32[3];
+    t.u32[0] = a.u32[1];
+  } else {
+    t.u32[7] = a.u32[6];
+    t.u32[6] = a.u32[4];
+    t.u32[5] = a.u32[2];
+    t.u32[4] = a.u32[0];
+    t.u32[3] = b.u32[6];
+    t.u32[2] = b.u32[4];
+    t.u32[1] = b.u32[2];
+    t.u32[0] = b.u32[0];
+  }
+  return t;
+}
+
+SIMD_INLINE c_v256 c_v256_unziplo_32(c_v256 a, c_v256 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v256_unzip_32(a, b, 1)
+                           : _c_v256_unzip_32(a, b, 0);
+}
+
+SIMD_INLINE c_v256 c_v256_unziphi_32(c_v256 a, c_v256 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v256_unzip_32(b, a, 0)
+                           : _c_v256_unzip_32(b, a, 1);
+}
+
+SIMD_INLINE c_v256 c_v256_unpack_u8_s16(c_v128 a) {
+  return c_v256_from_v128(c_v128_unpackhi_u8_s16(a), c_v128_unpacklo_u8_s16(a));
+}
+
+SIMD_INLINE c_v256 c_v256_unpacklo_u8_s16(c_v256 a) {
+  return c_v256_from_v128(c_v128_unpackhi_u8_s16(a.v128[0]),
+                          c_v128_unpacklo_u8_s16(a.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpackhi_u8_s16(c_v256 a) {
+  return c_v256_from_v128(c_v128_unpackhi_u8_s16(a.v128[1]),
+                          c_v128_unpacklo_u8_s16(a.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpack_s8_s16(c_v128 a) {
+  return c_v256_from_v128(c_v128_unpackhi_s8_s16(a), c_v128_unpacklo_s8_s16(a));
+}
+
+SIMD_INLINE c_v256 c_v256_unpacklo_s8_s16(c_v256 a) {
+  return c_v256_from_v128(c_v128_unpackhi_s8_s16(a.v128[0]),
+                          c_v128_unpacklo_s8_s16(a.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpackhi_s8_s16(c_v256 a) {
+  return c_v256_from_v128(c_v128_unpackhi_s8_s16(a.v128[1]),
+                          c_v128_unpacklo_s8_s16(a.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_pack_s32_s16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_pack_s32_s16(a.v128[1], a.v128[0]),
+                          c_v128_pack_s32_s16(b.v128[1], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_pack_s16_u8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_pack_s16_u8(a.v128[1], a.v128[0]),
+                          c_v128_pack_s16_u8(b.v128[1], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_pack_s16_s8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_pack_s16_s8(a.v128[1], a.v128[0]),
+                          c_v128_pack_s16_s8(b.v128[1], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpack_u16_s32(c_v128 a) {
+  return c_v256_from_v128(c_v128_unpackhi_u16_s32(a),
+                          c_v128_unpacklo_u16_s32(a));
+}
+
+SIMD_INLINE c_v256 c_v256_unpack_s16_s32(c_v128 a) {
+  return c_v256_from_v128(c_v128_unpackhi_s16_s32(a),
+                          c_v128_unpacklo_s16_s32(a));
+}
+
+SIMD_INLINE c_v256 c_v256_unpacklo_u16_s32(c_v256 a) {
+  return c_v256_from_v128(c_v128_unpackhi_u16_s32(a.v128[0]),
+                          c_v128_unpacklo_u16_s32(a.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpacklo_s16_s32(c_v256 a) {
+  return c_v256_from_v128(c_v128_unpackhi_s16_s32(a.v128[0]),
+                          c_v128_unpacklo_s16_s32(a.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpackhi_u16_s32(c_v256 a) {
+  return c_v256_from_v128(c_v128_unpackhi_u16_s32(a.v128[1]),
+                          c_v128_unpacklo_u16_s32(a.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_unpackhi_s16_s32(c_v256 a) {
+  return c_v256_from_v128(c_v128_unpackhi_s16_s32(a.v128[1]),
+                          c_v128_unpacklo_s16_s32(a.v128[1]));
+}
+
+SIMD_INLINE c_v256 c_v256_shuffle_8(c_v256 a, c_v256 pattern) {
+  c_v256 t;
+  int c;
+  for (c = 0; c < 32; c++) {
+    if (pattern.u8[c] & ~31) {
+      fprintf(stderr, "Undefined v256_shuffle_8 index %d/%d\n", pattern.u8[c],
+              c);
+      abort();
+    }
+    t.u8[c] = a.u8[CONFIG_BIG_ENDIAN ? 31 - (pattern.u8[c] & 31)
+                                     : pattern.u8[c] & 31];
+  }
+  return t;
+}
+
+// Pairwise / dual-lane shuffle: shuffle two 128 bit lates.
+SIMD_INLINE c_v256 c_v256_pshuffle_8(c_v256 a, c_v256 pattern) {
+  return c_v256_from_v128(
+      c_v128_shuffle_8(c_v256_high_v128(a), c_v256_high_v128(pattern)),
+      c_v128_shuffle_8(c_v256_low_v128(a), c_v256_low_v128(pattern)));
+}
+
+SIMD_INLINE c_v256 c_v256_cmpgt_s8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_cmpgt_s8(a.v128[1], b.v128[1]),
+                          c_v128_cmpgt_s8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_cmplt_s8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_cmplt_s8(a.v128[1], b.v128[1]),
+                          c_v128_cmplt_s8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_cmpeq_8(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_cmpeq_8(a.v128[1], b.v128[1]),
+                          c_v128_cmpeq_8(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_cmpgt_s16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_cmpgt_s16(a.v128[1], b.v128[1]),
+                          c_v128_cmpgt_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_cmplt_s16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_cmplt_s16(a.v128[1], b.v128[1]),
+                          c_v128_cmplt_s16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_cmpeq_16(c_v256 a, c_v256 b) {
+  return c_v256_from_v128(c_v128_cmpeq_16(a.v128[1], b.v128[1]),
+                          c_v128_cmpeq_16(a.v128[0], b.v128[0]));
+}
+
+SIMD_INLINE c_v256 c_v256_shl_n_byte(c_v256 a, unsigned int n) {
+  if (n < 16)
+    return c_v256_from_v128(c_v128_or(c_v128_shl_n_byte(a.v128[1], n),
+                                      c_v128_shr_n_byte(a.v128[0], 16 - n)),
+                            c_v128_shl_n_byte(a.v128[0], n));
+  else if (n > 16)
+    return c_v256_from_v128(c_v128_shl_n_byte(a.v128[0], n - 16),
+                            c_v128_zero());
+  else
+    return c_v256_from_v128(c_v256_low_v128(a), c_v128_zero());
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_byte(c_v256 a, unsigned int n) {
+  if (n < 16)
+    return c_v256_from_v128(c_v128_shr_n_byte(a.v128[1], n),
+                            c_v128_or(c_v128_shr_n_byte(a.v128[0], n),
+                                      c_v128_shl_n_byte(a.v128[1], 16 - n)));
+  else if (n > 16)
+    return c_v256_from_v128(c_v128_zero(),
+                            c_v128_shr_n_byte(a.v128[1], n - 16));
+  else
+    return c_v256_from_v128(c_v128_zero(), c_v256_high_v128(a));
+}
+
+SIMD_INLINE c_v256 c_v256_align(c_v256 a, c_v256 b, unsigned int c) {
+  if (SIMD_CHECK && c > 31) {
+    fprintf(stderr, "Error: undefined alignment %d\n", c);
+    abort();
+  }
+  return c ? c_v256_or(c_v256_shr_n_byte(b, c), c_v256_shl_n_byte(a, 32 - c))
+           : b;
+}
+
+SIMD_INLINE c_v256 c_v256_shl_8(c_v256 a, unsigned int c) {
+  return c_v256_from_v128(c_v128_shl_8(a.v128[1], c),
+                          c_v128_shl_8(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_u8(c_v256 a, unsigned int c) {
+  return c_v256_from_v128(c_v128_shr_u8(a.v128[1], c),
+                          c_v128_shr_u8(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_s8(c_v256 a, unsigned int c) {
+  return c_v256_from_v128(c_v128_shr_s8(a.v128[1], c),
+                          c_v128_shr_s8(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shl_16(c_v256 a, unsigned int c) {
+  return c_v256_from_v128(c_v128_shl_16(a.v128[1], c),
+                          c_v128_shl_16(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_u16(c_v256 a, unsigned int c) {
+  return c_v256_from_v128(c_v128_shr_u16(a.v128[1], c),
+                          c_v128_shr_u16(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_s16(c_v256 a, unsigned int c) {
+  return c_v256_from_v128(c_v128_shr_s16(a.v128[1], c),
+                          c_v128_shr_s16(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shl_32(c_v256 a, unsigned int c) {
+  return c_v256_from_v128(c_v128_shl_32(a.v128[1], c),
+                          c_v128_shl_32(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_u32(c_v256 a, unsigned int c) {
+  return c_v256_from_v128(c_v128_shr_u32(a.v128[1], c),
+                          c_v128_shr_u32(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shr_s32(c_v256 a, unsigned int c) {
+  return c_v256_from_v128(c_v128_shr_s32(a.v128[1], c),
+                          c_v128_shr_s32(a.v128[0], c));
+}
+
+SIMD_INLINE c_v256 c_v256_shl_n_8(c_v256 a, unsigned int n) {
+  return c_v256_shl_8(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shl_n_16(c_v256 a, unsigned int n) {
+  return c_v256_shl_16(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shl_n_32(c_v256 a, unsigned int n) {
+  return c_v256_shl_32(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_u8(c_v256 a, unsigned int n) {
+  return c_v256_shr_u8(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_u16(c_v256 a, unsigned int n) {
+  return c_v256_shr_u16(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_u32(c_v256 a, unsigned int n) {
+  return c_v256_shr_u32(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_s8(c_v256 a, unsigned int n) {
+  return c_v256_shr_s8(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_s16(c_v256 a, unsigned int n) {
+  return c_v256_shr_s16(a, n);
+}
+
+SIMD_INLINE c_v256 c_v256_shr_n_s32(c_v256 a, unsigned int n) {
+  return c_v256_shr_s32(a, n);
+}
+
+#endif /* _V256_INTRINSICS_C_H */
diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics_v128.h b/third_party/aom/aom_dsp/simd/v256_intrinsics_v128.h
new file mode 100644
index 0000000000..a4b334ea65
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v256_intrinsics_v128.h
@@ -0,0 +1,545 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V256_INTRINSICS_V128_H
+#define _V256_INTRINSICS_V128_H
+
+#if HAVE_NEON
+#include "./v128_intrinsics_arm.h"
+#elif HAVE_SSE2
+#include "./v128_intrinsics_x86.h"
+#else
+#include "./v128_intrinsics.h"
+#endif
+
+typedef struct { v128 lo, hi; } v256;
+
+SIMD_INLINE uint32_t v256_low_u32(v256 a) { return v128_low_u32(a.lo); }
+
+SIMD_INLINE v64 v256_low_v64(v256 a) { return v128_low_v64(a.lo); }
+
+SIMD_INLINE v128 v256_low_v128(v256 a) { return a.lo; }
+
+SIMD_INLINE v128 v256_high_v128(v256 a) { return a.hi; }
+
+SIMD_INLINE v256 v256_from_v128(v128 hi, v128 lo) {
+  v256 t;
+  t.hi = hi;
+  t.lo = lo;
+  return t;
+}
+
+SIMD_INLINE v256 v256_from_64(uint64_t a, uint64_t b, uint64_t c, uint64_t d) {
+  return v256_from_v128(v128_from_64(a, b), v128_from_64(c, d));
+}
+
+SIMD_INLINE v256 v256_from_v64(v64 a, v64 b, v64 c, v64 d) {
+  return v256_from_v128(v128_from_v64(a, b), v128_from_v64(c, d));
+}
+
+SIMD_INLINE v256 v256_load_unaligned(const void *p) {
+  return v256_from_v128(v128_load_unaligned((uint8_t *)p + 16),
+                        v128_load_unaligned(p));
+}
+
+SIMD_INLINE v256 v256_load_aligned(const void *p) {
+  return v256_from_v128(v128_load_aligned((uint8_t *)p + 16),
+                        v128_load_aligned(p));
+}
+
+SIMD_INLINE void v256_store_unaligned(void *p, v256 a) {
+  v128_store_unaligned(p, a.lo);
+  v128_store_unaligned((uint8_t *)p + 16, a.hi);
+}
+
+SIMD_INLINE void v256_store_aligned(void *p, v256 a) {
+  v128_store_aligned(p, a.lo);
+  v128_store_aligned((uint8_t *)p + 16, a.hi);
+}
+
+SIMD_INLINE v256 v256_zero() {
+  return v256_from_v128(v128_zero(), v128_zero());
+}
+
+SIMD_INLINE v256 v256_dup_8(uint8_t x) {
+  v128 t = v128_dup_8(x);
+  return v256_from_v128(t, t);
+}
+
+SIMD_INLINE v256 v256_dup_16(uint16_t x) {
+  v128 t = v128_dup_16(x);
+  return v256_from_v128(t, t);
+}
+
+SIMD_INLINE v256 v256_dup_32(uint32_t x) {
+  v128 t = v128_dup_32(x);
+  return v256_from_v128(t, t);
+}
+
+SIMD_INLINE int64_t v256_dotp_s16(v256 a, v256 b) {
+  return v128_dotp_s16(a.hi, b.hi) + v128_dotp_s16(a.lo, b.lo);
+}
+
+SIMD_INLINE uint64_t v256_hadd_u8(v256 a) {
+  return v128_hadd_u8(a.hi) + v128_hadd_u8(a.lo);
+}
+
+typedef struct {
+  sad128_internal hi;
+  sad128_internal lo;
+} sad256_internal;
+
+SIMD_INLINE sad256_internal v256_sad_u8_init() {
+  sad256_internal t;
+  t.hi = v128_sad_u8_init();
+  t.lo = v128_sad_u8_init();
+  return t;
+}
+
+/* Implementation dependent return value.  Result must be finalised with
+   v256_sad_u8_sum().
+   The result for more than 16 v256_sad_u8() calls is undefined. */
+SIMD_INLINE sad256_internal v256_sad_u8(sad256_internal s, v256 a, v256 b) {
+  sad256_internal t;
+  t.hi = v128_sad_u8(s.hi, a.hi, b.hi);
+  t.lo = v128_sad_u8(s.lo, a.lo, b.lo);
+  return t;
+}
+
+SIMD_INLINE uint32_t v256_sad_u8_sum(sad256_internal s) {
+  return v128_sad_u8_sum(s.hi) + v128_sad_u8_sum(s.lo);
+}
+
+typedef struct {
+  ssd128_internal hi;
+  ssd128_internal lo;
+} ssd256_internal;
+
+SIMD_INLINE ssd256_internal v256_ssd_u8_init() {
+  ssd256_internal t;
+  t.hi = v128_ssd_u8_init();
+  t.lo = v128_ssd_u8_init();
+  return t;
+}
+
+/* Implementation dependent return value.  Result must be finalised with
+ * v256_ssd_u8_sum(). */
+SIMD_INLINE ssd256_internal v256_ssd_u8(ssd256_internal s, v256 a, v256 b) {
+  ssd256_internal t;
+  t.hi = v128_ssd_u8(s.hi, a.hi, b.hi);
+  t.lo = v128_ssd_u8(s.lo, a.lo, b.lo);
+  return t;
+}
+
+SIMD_INLINE uint32_t v256_ssd_u8_sum(ssd256_internal s) {
+  return v128_ssd_u8_sum(s.hi) + v128_ssd_u8_sum(s.lo);
+}
+
+SIMD_INLINE v256 v256_or(v256 a, v256 b) {
+  return v256_from_v128(v128_or(a.hi, b.hi), v128_or(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_xor(v256 a, v256 b) {
+  return v256_from_v128(v128_xor(a.hi, b.hi), v128_xor(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_and(v256 a, v256 b) {
+  return v256_from_v128(v128_and(a.hi, b.hi), v128_and(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_andn(v256 a, v256 b) {
+  return v256_from_v128(v128_andn(a.hi, b.hi), v128_andn(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_add_8(v256 a, v256 b) {
+  return v256_from_v128(v128_add_8(a.hi, b.hi), v128_add_8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_add_16(v256 a, v256 b) {
+  return v256_from_v128(v128_add_16(a.hi, b.hi), v128_add_16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_sadd_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_sadd_s16(a.hi, b.hi), v128_sadd_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_add_32(v256 a, v256 b) {
+  return v256_from_v128(v128_add_32(a.hi, b.hi), v128_add_32(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_padd_s16(v256 a) {
+  return v256_from_v128(v128_padd_s16(a.hi), v128_padd_s16(a.lo));
+}
+
+SIMD_INLINE v256 v256_sub_8(v256 a, v256 b) {
+  return v256_from_v128(v128_sub_8(a.hi, b.hi), v128_sub_8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ssub_u8(v256 a, v256 b) {
+  return v256_from_v128(v128_ssub_u8(a.hi, b.hi), v128_ssub_u8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ssub_s8(v256 a, v256 b) {
+  return v256_from_v128(v128_ssub_s8(a.hi, b.hi), v128_ssub_s8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_sub_16(v256 a, v256 b) {
+  return v256_from_v128(v128_sub_16(a.hi, b.hi), v128_sub_16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ssub_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_ssub_s16(a.hi, b.hi), v128_ssub_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ssub_u16(v256 a, v256 b) {
+  return v256_from_v128(v128_ssub_u16(a.hi, b.hi), v128_ssub_u16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_sub_32(v256 a, v256 b) {
+  return v256_from_v128(v128_sub_32(a.hi, b.hi), v128_sub_32(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_abs_s16(v256 a) {
+  return v256_from_v128(v128_abs_s16(a.hi), v128_abs_s16(a.lo));
+}
+
+SIMD_INLINE v256 v256_abs_s8(v256 a) {
+  return v256_from_v128(v128_abs_s8(a.hi), v128_abs_s8(a.lo));
+}
+
+SIMD_INLINE v256 v256_mul_s16(v128 a, v128 b) {
+  v128 lo_bits = v128_mullo_s16(a, b);
+  v128 hi_bits = v128_mulhi_s16(a, b);
+  return v256_from_v128(v128_ziphi_16(hi_bits, lo_bits),
+                        v128_ziplo_16(hi_bits, lo_bits));
+}
+
+SIMD_INLINE v256 v256_mullo_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_mullo_s16(a.hi, b.hi), v128_mullo_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_mulhi_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_mulhi_s16(a.hi, b.hi), v128_mulhi_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_mullo_s32(v256 a, v256 b) {
+  return v256_from_v128(v128_mullo_s32(a.hi, b.hi), v128_mullo_s32(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_madd_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_madd_s16(a.hi, b.hi), v128_madd_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_madd_us8(v256 a, v256 b) {
+  return v256_from_v128(v128_madd_us8(a.hi, b.hi), v128_madd_us8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_avg_u8(v256 a, v256 b) {
+  return v256_from_v128(v128_avg_u8(a.hi, b.hi), v128_avg_u8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_rdavg_u8(v256 a, v256 b) {
+  return v256_from_v128(v128_rdavg_u8(a.hi, b.hi), v128_rdavg_u8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_avg_u16(v256 a, v256 b) {
+  return v256_from_v128(v128_avg_u16(a.hi, b.hi), v128_avg_u16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_min_u8(v256 a, v256 b) {
+  return v256_from_v128(v128_min_u8(a.hi, b.hi), v128_min_u8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_max_u8(v256 a, v256 b) {
+  return v256_from_v128(v128_max_u8(a.hi, b.hi), v128_max_u8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_min_s8(v256 a, v256 b) {
+  return v256_from_v128(v128_min_s8(a.hi, b.hi), v128_min_s8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_max_s8(v256 a, v256 b) {
+  return v256_from_v128(v128_max_s8(a.hi, b.hi), v128_max_s8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_min_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_min_s16(a.hi, b.hi), v128_min_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_max_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_max_s16(a.hi, b.hi), v128_max_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ziplo_8(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_8(a.lo, b.lo), v128_ziplo_8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ziphi_8(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_8(a.hi, b.hi), v128_ziplo_8(a.hi, b.hi));
+}
+
+SIMD_INLINE v256 v256_ziplo_16(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_16(a.lo, b.lo), v128_ziplo_16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ziphi_16(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_16(a.hi, b.hi), v128_ziplo_16(a.hi, b.hi));
+}
+
+SIMD_INLINE v256 v256_ziplo_32(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_32(a.lo, b.lo), v128_ziplo_32(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ziphi_32(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_32(a.hi, b.hi), v128_ziplo_32(a.hi, b.hi));
+}
+
+SIMD_INLINE v256 v256_ziplo_64(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_64(a.lo, b.lo), v128_ziplo_64(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_ziphi_64(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_64(a.hi, b.hi), v128_ziplo_64(a.hi, b.hi));
+}
+
+SIMD_INLINE v256 v256_ziplo_128(v256 a, v256 b) {
+  return v256_from_v128(a.lo, b.lo);
+}
+
+SIMD_INLINE v256 v256_ziphi_128(v256 a, v256 b) {
+  return v256_from_v128(a.hi, b.hi);
+}
+
+SIMD_INLINE v256 v256_zip_8(v128 a, v128 b) {
+  return v256_from_v128(v128_ziphi_8(a, b), v128_ziplo_8(a, b));
+}
+
+SIMD_INLINE v256 v256_zip_16(v128 a, v128 b) {
+  return v256_from_v128(v128_ziphi_16(a, b), v128_ziplo_16(a, b));
+}
+
+SIMD_INLINE v256 v256_zip_32(v128 a, v128 b) {
+  return v256_from_v128(v128_ziphi_32(a, b), v128_ziplo_32(a, b));
+}
+
+SIMD_INLINE v256 v256_unziplo_8(v256 a, v256 b) {
+  return v256_from_v128(v128_unziplo_8(a.hi, a.lo), v128_unziplo_8(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unziphi_8(v256 a, v256 b) {
+  return v256_from_v128(v128_unziphi_8(a.hi, a.lo), v128_unziphi_8(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unziplo_16(v256 a, v256 b) {
+  return v256_from_v128(v128_unziplo_16(a.hi, a.lo),
+                        v128_unziplo_16(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unziphi_16(v256 a, v256 b) {
+  return v256_from_v128(v128_unziphi_16(a.hi, a.lo),
+                        v128_unziphi_16(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unziplo_32(v256 a, v256 b) {
+  return v256_from_v128(v128_unziplo_32(a.hi, a.lo),
+                        v128_unziplo_32(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unziphi_32(v256 a, v256 b) {
+  return v256_from_v128(v128_unziphi_32(a.hi, a.lo),
+                        v128_unziphi_32(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unpack_u8_s16(v128 a) {
+  return v256_from_v128(v128_unpackhi_u8_s16(a), v128_unpacklo_u8_s16(a));
+}
+
+SIMD_INLINE v256 v256_unpacklo_u8_s16(v256 a) {
+  return v256_from_v128(v128_unpackhi_u8_s16(a.lo), v128_unpacklo_u8_s16(a.lo));
+}
+
+SIMD_INLINE v256 v256_unpackhi_u8_s16(v256 a) {
+  return v256_from_v128(v128_unpackhi_u8_s16(a.hi), v128_unpacklo_u8_s16(a.hi));
+}
+
+SIMD_INLINE v256 v256_unpack_s8_s16(v128 a) {
+  return v256_from_v128(v128_unpackhi_s8_s16(a), v128_unpacklo_s8_s16(a));
+}
+
+SIMD_INLINE v256 v256_unpacklo_s8_s16(v256 a) {
+  return v256_from_v128(v128_unpackhi_s8_s16(a.lo), v128_unpacklo_s8_s16(a.lo));
+}
+
+SIMD_INLINE v256 v256_unpackhi_s8_s16(v256 a) {
+  return v256_from_v128(v128_unpackhi_s8_s16(a.hi), v128_unpacklo_s8_s16(a.hi));
+}
+
+SIMD_INLINE v256 v256_pack_s32_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_pack_s32_s16(a.hi, a.lo),
+                        v128_pack_s32_s16(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_pack_s16_u8(v256 a, v256 b) {
+  return v256_from_v128(v128_pack_s16_u8(a.hi, a.lo),
+                        v128_pack_s16_u8(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_pack_s16_s8(v256 a, v256 b) {
+  return v256_from_v128(v128_pack_s16_s8(a.hi, a.lo),
+                        v128_pack_s16_s8(b.hi, b.lo));
+}
+
+SIMD_INLINE v256 v256_unpack_u16_s32(v128 a) {
+  return v256_from_v128(v128_unpackhi_u16_s32(a), v128_unpacklo_u16_s32(a));
+}
+
+SIMD_INLINE v256 v256_unpack_s16_s32(v128 a) {
+  return v256_from_v128(v128_unpackhi_s16_s32(a), v128_unpacklo_s16_s32(a));
+}
+
+SIMD_INLINE v256 v256_unpacklo_u16_s32(v256 a) {
+  return v256_from_v128(v128_unpackhi_u16_s32(a.lo),
+                        v128_unpacklo_u16_s32(a.lo));
+}
+
+SIMD_INLINE v256 v256_unpacklo_s16_s32(v256 a) {
+  return v256_from_v128(v128_unpackhi_s16_s32(a.lo),
+                        v128_unpacklo_s16_s32(a.lo));
+}
+
+SIMD_INLINE v256 v256_unpackhi_u16_s32(v256 a) {
+  return v256_from_v128(v128_unpackhi_u16_s32(a.hi),
+                        v128_unpacklo_u16_s32(a.hi));
+}
+
+SIMD_INLINE v256 v256_unpackhi_s16_s32(v256 a) {
+  return v256_from_v128(v128_unpackhi_s16_s32(a.hi),
+                        v128_unpacklo_s16_s32(a.hi));
+}
+
+SIMD_INLINE v256 v256_shuffle_8(v256 a, v256 pattern) {
+  v128 c16 = v128_dup_8(16);
+  v128 maskhi = v128_cmplt_s8(pattern.hi, c16);
+  v128 masklo = v128_cmplt_s8(pattern.lo, c16);
+  return v256_from_v128(
+      v128_or(
+          v128_and(v128_shuffle_8(a.lo, pattern.hi), maskhi),
+          v128_andn(v128_shuffle_8(a.hi, v128_sub_8(pattern.hi, c16)), maskhi)),
+      v128_or(v128_and(v128_shuffle_8(a.lo, pattern.lo), masklo),
+              v128_andn(v128_shuffle_8(a.hi, v128_sub_8(pattern.lo, c16)),
+                        masklo)));
+}
+
+SIMD_INLINE v256 v256_pshuffle_8(v256 a, v256 pattern) {
+  return v256_from_v128(
+      v128_shuffle_8(v256_high_v128(a), v256_high_v128(pattern)),
+      v128_shuffle_8(v256_low_v128(a), v256_low_v128(pattern)));
+}
+
+SIMD_INLINE v256 v256_cmpgt_s8(v256 a, v256 b) {
+  return v256_from_v128(v128_cmpgt_s8(a.hi, b.hi), v128_cmpgt_s8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_cmplt_s8(v256 a, v256 b) {
+  return v256_from_v128(v128_cmplt_s8(a.hi, b.hi), v128_cmplt_s8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_cmpeq_8(v256 a, v256 b) {
+  return v256_from_v128(v128_cmpeq_8(a.hi, b.hi), v128_cmpeq_8(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_cmpgt_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_cmpgt_s16(a.hi, b.hi), v128_cmpgt_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_cmplt_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_cmplt_s16(a.hi, b.hi), v128_cmplt_s16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_cmpeq_16(v256 a, v256 b) {
+  return v256_from_v128(v128_cmpeq_16(a.hi, b.hi), v128_cmpeq_16(a.lo, b.lo));
+}
+
+SIMD_INLINE v256 v256_shl_8(v256 a, unsigned int c) {
+  return v256_from_v128(v128_shl_8(a.hi, c), v128_shl_8(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_u8(v256 a, unsigned int c) {
+  return v256_from_v128(v128_shr_u8(a.hi, c), v128_shr_u8(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_s8(v256 a, unsigned int c) {
+  return v256_from_v128(v128_shr_s8(a.hi, c), v128_shr_s8(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shl_16(v256 a, unsigned int c) {
+  return v256_from_v128(v128_shl_16(a.hi, c), v128_shl_16(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_u16(v256 a, unsigned int c) {
+  return v256_from_v128(v128_shr_u16(a.hi, c), v128_shr_u16(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_s16(v256 a, unsigned int c) {
+  return v256_from_v128(v128_shr_s16(a.hi, c), v128_shr_s16(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shl_32(v256 a, unsigned int c) {
+  return v256_from_v128(v128_shl_32(a.hi, c), v128_shl_32(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_u32(v256 a, unsigned int c) {
+  return v256_from_v128(v128_shr_u32(a.hi, c), v128_shr_u32(a.lo, c));
+}
+
+SIMD_INLINE v256 v256_shr_s32(v256 a, unsigned int c) {
+  return v256_from_v128(v128_shr_s32(a.hi, c), v128_shr_s32(a.lo, c));
+}
+
+/* These intrinsics require immediate values, so we must use #defines
+   to enforce that. */
+#define v256_shl_n_byte(a, n)                                                 \
+  ((n) < 16 ? v256_from_v128(v128_or(v128_shl_n_byte(a.hi, n),                \
+                                     v128_shr_n_byte(a.lo, 16 - (n))),        \
+                             v128_shl_n_byte(a.lo, (n)))                      \
+            : v256_from_v128((n) > 16 ? v128_shl_n_byte(a.lo, (n)-16) : a.lo, \
+                             v128_zero()))
+
+#define v256_shr_n_byte(a, n)                                          \
+  ((n) < 16 ? v256_from_v128(v128_shr_n_byte(a.hi, n),                 \
+                             v128_or(v128_shr_n_byte(a.lo, n),         \
+                                     v128_shl_n_byte(a.hi, 16 - (n)))) \
+            : v256_from_v128(v128_zero(),                              \
+                             (n) > 16 ? v128_shr_n_byte(a.hi, (n)-16) : a.hi))
+
+#define v256_align(a, b, c) \
+  ((c) ? v256_or(v256_shr_n_byte(b, c), v256_shl_n_byte(a, 32 - (c))) : b)
+
+#define v256_shl_n_8(a, n) \
+  v256_from_v128(v128_shl_n_8(a.hi, n), v128_shl_n_8(a.lo, n))
+#define v256_shl_n_16(a, n) \
+  v256_from_v128(v128_shl_n_16(a.hi, n), v128_shl_n_16(a.lo, n))
+#define v256_shl_n_32(a, n) \
+  v256_from_v128(v128_shl_n_32(a.hi, n), v128_shl_n_32(a.lo, n))
+#define v256_shr_n_u8(a, n) \
+  v256_from_v128(v128_shr_n_u8(a.hi, n), v128_shr_n_u8(a.lo, n))
+#define v256_shr_n_u16(a, n) \
+  v256_from_v128(v128_shr_n_u16(a.hi, n), v128_shr_n_u16(a.lo, n))
+#define v256_shr_n_u32(a, n) \
+  v256_from_v128(v128_shr_n_u32(a.hi, n), v128_shr_n_u32(a.lo, n))
+#define v256_shr_n_s8(a, n) \
+  v256_from_v128(v128_shr_n_s8(a.hi, n), v128_shr_n_s8(a.lo, n))
+#define v256_shr_n_s16(a, n) \
+  v256_from_v128(v128_shr_n_s16(a.hi, n), v128_shr_n_s16(a.lo, n))
+#define v256_shr_n_s32(a, n) \
+  v256_from_v128(v128_shr_n_s32(a.hi, n), v128_shr_n_s32(a.lo, n))
+
+#endif /* _V256_INTRINSICS_V128_H */
diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics_x86.h b/third_party/aom/aom_dsp/simd/v256_intrinsics_x86.h
new file mode 100644
index 0000000000..b82daab686
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v256_intrinsics_x86.h
@@ -0,0 +1,548 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V256_INTRINSICS_H
+#define _V256_INTRINSICS_H
+
+#if !defined(__AVX2__)
+
+#include "./v256_intrinsics_v128.h"
+
+#else
+
+// The _m256i type seems to cause problems for g++'s mangling prior to
+// version 5, but adding -fabi-version=0 fixes this.
+#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ < 5 && \
+    defined(__AVX2__) && defined(__cplusplus)
+#pragma GCC optimize "-fabi-version=0"
+#endif
+
+#include <immintrin.h>
+#include "./v128_intrinsics_x86.h"
+
+typedef __m256i v256;
+
+SIMD_INLINE uint32_t v256_low_u32(v256 a) {
+  return (uint32_t)_mm_cvtsi128_si32(_mm256_extracti128_si256(a, 0));
+}
+
+SIMD_INLINE v64 v256_low_v64(v256 a) {
+  return _mm_unpacklo_epi64(_mm256_extracti128_si256(a, 0), v64_zero());
+}
+
+SIMD_INLINE v128 v256_low_v128(v256 a) {
+  return _mm256_extracti128_si256(a, 0);
+}
+
+SIMD_INLINE v128 v256_high_v128(v256 a) {
+  return _mm256_extracti128_si256(a, 1);
+}
+
+SIMD_INLINE v256 v256_from_v128(v128 a, v128 b) {
+  // gcc seems to be missing _mm256_set_m128i()
+  return _mm256_insertf128_si256(
+      _mm256_insertf128_si256(_mm256_setzero_si256(), b, 0), a, 1);
+}
+
+SIMD_INLINE v256 v256_from_v64(v64 a, v64 b, v64 c, v64 d) {
+  return v256_from_v128(v128_from_v64(a, b), v128_from_v64(c, d));
+}
+
+SIMD_INLINE v256 v256_from_64(uint64_t a, uint64_t b, uint64_t c, uint64_t d) {
+  return v256_from_v128(v128_from_64(a, b), v128_from_64(c, d));
+}
+
+SIMD_INLINE v256 v256_load_aligned(const void *p) {
+  return _mm256_load_si256((const __m256i *)p);
+}
+
+SIMD_INLINE v256 v256_load_unaligned(const void *p) {
+  return _mm256_loadu_si256((const __m256i *)p);
+}
+
+SIMD_INLINE void v256_store_aligned(void *p, v256 a) {
+  _mm256_store_si256((__m256i *)p, a);
+}
+
+SIMD_INLINE void v256_store_unaligned(void *p, v256 a) {
+  _mm256_storeu_si256((__m256i *)p, a);
+}
+
+SIMD_INLINE v256 v256_zero() { return _mm256_setzero_si256(); }
+
+SIMD_INLINE v256 v256_dup_8(uint8_t x) { return _mm256_set1_epi8(x); }
+
+SIMD_INLINE v256 v256_dup_16(uint16_t x) { return _mm256_set1_epi16(x); }
+
+SIMD_INLINE v256 v256_dup_32(uint32_t x) { return _mm256_set1_epi32(x); }
+
+SIMD_INLINE v256 v256_add_8(v256 a, v256 b) { return _mm256_add_epi8(a, b); }
+
+SIMD_INLINE v256 v256_add_16(v256 a, v256 b) { return _mm256_add_epi16(a, b); }
+
+SIMD_INLINE v256 v256_sadd_s16(v256 a, v256 b) {
+  return _mm256_adds_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_add_32(v256 a, v256 b) { return _mm256_add_epi32(a, b); }
+
+SIMD_INLINE v256 v256_padd_s16(v256 a) {
+  return _mm256_madd_epi16(a, _mm256_set1_epi16(1));
+}
+
+SIMD_INLINE v256 v256_sub_8(v256 a, v256 b) { return _mm256_sub_epi8(a, b); }
+
+SIMD_INLINE v256 v256_ssub_u8(v256 a, v256 b) { return _mm256_subs_epu8(a, b); }
+
+SIMD_INLINE v256 v256_ssub_s8(v256 a, v256 b) { return _mm256_subs_epi8(a, b); }
+
+SIMD_INLINE v256 v256_sub_16(v256 a, v256 b) { return _mm256_sub_epi16(a, b); }
+
+SIMD_INLINE v256 v256_ssub_s16(v256 a, v256 b) {
+  return _mm256_subs_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_ssub_u16(v256 a, v256 b) {
+  return _mm256_subs_epu16(a, b);
+}
+
+SIMD_INLINE v256 v256_sub_32(v256 a, v256 b) { return _mm256_sub_epi32(a, b); }
+
+SIMD_INLINE v256 v256_abs_s16(v256 a) { return _mm256_abs_epi16(a); }
+
+SIMD_INLINE v256 v256_abs_s8(v256 a) { return _mm256_abs_epi8(a); }
+
+// AVX doesn't have the direct intrinsics to zip/unzip 8, 16, 32 bit
+// lanes of lower or upper halves of a 256bit vector because the
+// unpack/pack intrinsics operate on the 256 bit input vector as 2
+// independent 128 bit vectors.
+SIMD_INLINE v256 v256_ziplo_8(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_8(v256_low_v128(a), v256_low_v128(b)),
+                        v128_ziplo_8(v256_low_v128(a), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziphi_8(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_8(v256_high_v128(a), v256_high_v128(b)),
+                        v128_ziplo_8(v256_high_v128(a), v256_high_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziplo_16(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_16(v256_low_v128(a), v256_low_v128(b)),
+                        v128_ziplo_16(v256_low_v128(a), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziphi_16(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_16(v256_high_v128(a), v256_high_v128(b)),
+                        v128_ziplo_16(v256_high_v128(a), v256_high_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziplo_32(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_32(v256_low_v128(a), v256_low_v128(b)),
+                        v128_ziplo_32(v256_low_v128(a), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziphi_32(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_32(v256_high_v128(a), v256_high_v128(b)),
+                        v128_ziplo_32(v256_high_v128(a), v256_high_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziplo_64(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_64(v256_low_v128(a), v256_low_v128(b)),
+                        v128_ziplo_64(v256_low_v128(a), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziphi_64(v256 a, v256 b) {
+  return v256_from_v128(v128_ziphi_64(v256_high_v128(a), v256_high_v128(b)),
+                        v128_ziplo_64(v256_high_v128(a), v256_high_v128(b)));
+}
+
+SIMD_INLINE v256 v256_ziplo_128(v256 a, v256 b) {
+  return v256_from_v128(v256_low_v128(a), v256_low_v128(b));
+}
+
+SIMD_INLINE v256 v256_ziphi_128(v256 a, v256 b) {
+  return v256_from_v128(v256_high_v128(a), v256_high_v128(b));
+}
+
+SIMD_INLINE v256 v256_zip_8(v128 a, v128 b) {
+  return v256_from_v128(v128_ziphi_8(a, b), v128_ziplo_8(a, b));
+}
+
+SIMD_INLINE v256 v256_zip_16(v128 a, v128 b) {
+  return v256_from_v128(v128_ziphi_16(a, b), v128_ziplo_16(a, b));
+}
+
+SIMD_INLINE v256 v256_zip_32(v128 a, v128 b) {
+  return v256_from_v128(v128_ziphi_32(a, b), v128_ziplo_32(a, b));
+}
+
+SIMD_INLINE v256 v256_unziplo_8(v256 a, v256 b) {
+  return v256_from_v128(v128_unziplo_8(v256_high_v128(a), v256_low_v128(a)),
+                        v128_unziplo_8(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unziphi_8(v256 a, v256 b) {
+  return v256_from_v128(v128_unziphi_8(v256_high_v128(a), v256_low_v128(a)),
+                        v128_unziphi_8(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unziplo_16(v256 a, v256 b) {
+  return v256_from_v128(v128_unziplo_16(v256_high_v128(a), v256_low_v128(a)),
+                        v128_unziplo_16(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unziphi_16(v256 a, v256 b) {
+  return v256_from_v128(v128_unziphi_16(v256_high_v128(a), v256_low_v128(a)),
+                        v128_unziphi_16(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unziplo_32(v256 a, v256 b) {
+  return v256_from_v128(v128_unziplo_32(v256_high_v128(a), v256_low_v128(a)),
+                        v128_unziplo_32(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unziphi_32(v256 a, v256 b) {
+  return v256_from_v128(v128_unziphi_32(v256_high_v128(a), v256_low_v128(a)),
+                        v128_unziphi_32(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unpack_u8_s16(v128 a) {
+  return v256_from_v128(v128_unpackhi_u8_s16(a), v128_unpacklo_u8_s16(a));
+}
+
+SIMD_INLINE v256 v256_unpacklo_u8_s16(v256 a) {
+  return v256_from_v128(v128_unpackhi_u8_s16(v256_low_v128(a)),
+                        v128_unpacklo_u8_s16(v256_low_v128(a)));
+}
+
+SIMD_INLINE v256 v256_unpackhi_u8_s16(v256 a) {
+  return v256_from_v128(v128_unpackhi_u8_s16(v256_high_v128(a)),
+                        v128_unpacklo_u8_s16(v256_high_v128(a)));
+}
+
+SIMD_INLINE v256 v256_unpack_s8_s16(v128 a) {
+  return v256_from_v128(v128_unpackhi_s8_s16(a), v128_unpacklo_s8_s16(a));
+}
+
+SIMD_INLINE v256 v256_unpacklo_s8_s16(v256 a) {
+  return v256_from_v128(v128_unpackhi_s8_s16(v256_low_v128(a)),
+                        v128_unpacklo_s8_s16(v256_low_v128(a)));
+}
+
+SIMD_INLINE v256 v256_unpackhi_s8_s16(v256 a) {
+  return v256_from_v128(v128_unpackhi_s8_s16(v256_high_v128(a)),
+                        v128_unpacklo_s8_s16(v256_high_v128(a)));
+}
+
+SIMD_INLINE v256 v256_pack_s32_s16(v256 a, v256 b) {
+  return v256_from_v128(v128_pack_s32_s16(v256_high_v128(a), v256_low_v128(a)),
+                        v128_pack_s32_s16(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_pack_s16_u8(v256 a, v256 b) {
+  return v256_from_v128(v128_pack_s16_u8(v256_high_v128(a), v256_low_v128(a)),
+                        v128_pack_s16_u8(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_pack_s16_s8(v256 a, v256 b) {
+  return v256_from_v128(v128_pack_s16_s8(v256_high_v128(a), v256_low_v128(a)),
+                        v128_pack_s16_s8(v256_high_v128(b), v256_low_v128(b)));
+}
+
+SIMD_INLINE v256 v256_unpack_u16_s32(v128 a) {
+  return v256_from_v128(v128_unpackhi_u16_s32(a), v128_unpacklo_u16_s32(a));
+}
+
+SIMD_INLINE v256 v256_unpack_s16_s32(v128 a) {
+  return v256_from_v128(v128_unpackhi_s16_s32(a), v128_unpacklo_s16_s32(a));
+}
+
+SIMD_INLINE v256 v256_unpacklo_u16_s32(v256 a) {
+  return v256_from_v128(v128_unpackhi_u16_s32(v256_low_v128(a)),
+                        v128_unpacklo_u16_s32(v256_low_v128(a)));
+}
+
+SIMD_INLINE v256 v256_unpacklo_s16_s32(v256 a) {
+  return v256_from_v128(v128_unpackhi_s16_s32(v256_low_v128(a)),
+                        v128_unpacklo_s16_s32(v256_low_v128(a)));
+}
+
+SIMD_INLINE v256 v256_unpackhi_u16_s32(v256 a) {
+  return v256_from_v128(v128_unpackhi_u16_s32(v256_high_v128(a)),
+                        v128_unpacklo_u16_s32(v256_high_v128(a)));
+}
+
+SIMD_INLINE v256 v256_unpackhi_s16_s32(v256 a) {
+  return v256_from_v128(v128_unpackhi_s16_s32(v256_high_v128(a)),
+                        v128_unpacklo_s16_s32(v256_high_v128(a)));
+}
+SIMD_INLINE v256 v256_shuffle_8(v256 a, v256 pattern) {
+  v128 c16 = v128_dup_8(16);
+  v128 hi = v256_high_v128(pattern);
+  v128 lo = v256_low_v128(pattern);
+  v128 maskhi = v128_cmplt_s8(hi, c16);
+  v128 masklo = v128_cmplt_s8(lo, c16);
+  return v256_from_v128(
+      v128_or(v128_and(v128_shuffle_8(v256_low_v128(a), hi), maskhi),
+              v128_andn(v128_shuffle_8(v256_high_v128(a), v128_sub_8(hi, c16)),
+                        maskhi)),
+      v128_or(v128_and(v128_shuffle_8(v256_low_v128(a), lo), masklo),
+              v128_andn(v128_shuffle_8(v256_high_v128(a), v128_sub_8(lo, c16)),
+                        masklo)));
+}
+
+SIMD_INLINE v256 v256_pshuffle_8(v256 a, v256 pattern) {
+  return _mm256_shuffle_epi8(a, pattern);
+}
+
+SIMD_INLINE int64_t v256_dotp_s16(v256 a, v256 b) {
+  v256 r = _mm256_madd_epi16(a, b);
+#if defined(__x86_64__)
+  v128 t;
+  r = _mm256_add_epi64(_mm256_cvtepi32_epi64(v256_high_v128(r)),
+                       _mm256_cvtepi32_epi64(v256_low_v128(r)));
+  t = v256_low_v128(_mm256_add_epi64(
+      r, _mm256_permute2x128_si256(r, r, _MM_SHUFFLE(2, 0, 0, 1))));
+  return _mm_cvtsi128_si64(_mm_add_epi64(t, _mm_srli_si128(t, 8)));
+#else
+  v128 l = v256_low_v128(r);
+  v128 h = v256_high_v128(r);
+  return (int64_t)_mm_cvtsi128_si32(l) +
+         (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 4)) +
+         (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 8)) +
+         (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(l, 12)) +
+         (int64_t)_mm_cvtsi128_si32(h) +
+         (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 4)) +
+         (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 8)) +
+         (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(h, 12));
+#endif
+}
+
+SIMD_INLINE uint64_t v256_hadd_u8(v256 a) {
+  v256 t = _mm256_sad_epu8(a, _mm256_setzero_si256());
+  v128 lo = v256_low_v128(t);
+  v128 hi = v256_high_v128(t);
+  lo = v128_add_32(lo, hi);
+  return v64_low_u32(v128_low_v64(lo)) + v128_low_u32(v128_high_v64(lo));
+}
+
+typedef v256 sad256_internal;
+
+SIMD_INLINE sad256_internal v256_sad_u8_init() {
+  return _mm256_setzero_si256();
+}
+
+/* Implementation dependent return value.  Result must be finalised with
+   v256_sad_sum().
+   The result for more than 32 v256_sad_u8() calls is undefined. */
+SIMD_INLINE sad256_internal v256_sad_u8(sad256_internal s, v256 a, v256 b) {
+  return _mm256_add_epi64(s, _mm256_sad_epu8(a, b));
+}
+
+SIMD_INLINE uint32_t v256_sad_u8_sum(sad256_internal s) {
+  v256 t = _mm256_add_epi32(s, _mm256_unpackhi_epi64(s, s));
+  return v128_low_u32(_mm_add_epi32(v256_high_v128(t), v256_low_v128(t)));
+}
+
+typedef v256 ssd256_internal;
+
+SIMD_INLINE ssd256_internal v256_ssd_u8_init() {
+  return _mm256_setzero_si256();
+}
+
+/* Implementation dependent return value.  Result must be finalised with
+ * v256_ssd_sum(). */
+SIMD_INLINE ssd256_internal v256_ssd_u8(ssd256_internal s, v256 a, v256 b) {
+  v256 l = _mm256_sub_epi16(_mm256_unpacklo_epi8(a, _mm256_setzero_si256()),
+                            _mm256_unpacklo_epi8(b, _mm256_setzero_si256()));
+  v256 h = _mm256_sub_epi16(_mm256_unpackhi_epi8(a, _mm256_setzero_si256()),
+                            _mm256_unpackhi_epi8(b, _mm256_setzero_si256()));
+  v256 rl = _mm256_madd_epi16(l, l);
+  v256 rh = _mm256_madd_epi16(h, h);
+  v128 c = _mm_cvtsi32_si128(32);
+  rl = _mm256_add_epi32(rl, _mm256_srli_si256(rl, 8));
+  rl = _mm256_add_epi32(rl, _mm256_srli_si256(rl, 4));
+  rh = _mm256_add_epi32(rh, _mm256_srli_si256(rh, 8));
+  rh = _mm256_add_epi32(rh, _mm256_srli_si256(rh, 4));
+  return _mm256_add_epi64(
+      s,
+      _mm256_srl_epi64(_mm256_sll_epi64(_mm256_unpacklo_epi64(rl, rh), c), c));
+}
+
+SIMD_INLINE uint32_t v256_ssd_u8_sum(ssd256_internal s) {
+  v256 t = _mm256_add_epi32(s, _mm256_unpackhi_epi64(s, s));
+  return v128_low_u32(_mm_add_epi32(v256_high_v128(t), v256_low_v128(t)));
+}
+
+SIMD_INLINE v256 v256_or(v256 a, v256 b) { return _mm256_or_si256(a, b); }
+
+SIMD_INLINE v256 v256_xor(v256 a, v256 b) { return _mm256_xor_si256(a, b); }
+
+SIMD_INLINE v256 v256_and(v256 a, v256 b) { return _mm256_and_si256(a, b); }
+
+SIMD_INLINE v256 v256_andn(v256 a, v256 b) { return _mm256_andnot_si256(b, a); }
+
+SIMD_INLINE v256 v256_mul_s16(v64 a, v64 b) {
+  v128 lo_bits = v128_mullo_s16(a, b);
+  v128 hi_bits = v128_mulhi_s16(a, b);
+  return v256_from_v128(v128_ziphi_16(hi_bits, lo_bits),
+                        v128_ziplo_16(hi_bits, lo_bits));
+}
+
+SIMD_INLINE v256 v256_mullo_s16(v256 a, v256 b) {
+  return _mm256_mullo_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_mulhi_s16(v256 a, v256 b) {
+  return _mm256_mulhi_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_mullo_s32(v256 a, v256 b) {
+  return _mm256_mullo_epi32(a, b);
+}
+
+SIMD_INLINE v256 v256_madd_s16(v256 a, v256 b) {
+  return _mm256_madd_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_madd_us8(v256 a, v256 b) {
+  return _mm256_maddubs_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_avg_u8(v256 a, v256 b) { return _mm256_avg_epu8(a, b); }
+
+SIMD_INLINE v256 v256_rdavg_u8(v256 a, v256 b) {
+  return _mm256_sub_epi8(
+      _mm256_avg_epu8(a, b),
+      _mm256_and_si256(_mm256_xor_si256(a, b), v256_dup_8(1)));
+}
+
+SIMD_INLINE v256 v256_avg_u16(v256 a, v256 b) { return _mm256_avg_epu16(a, b); }
+
+SIMD_INLINE v256 v256_min_u8(v256 a, v256 b) { return _mm256_min_epu8(a, b); }
+
+SIMD_INLINE v256 v256_max_u8(v256 a, v256 b) { return _mm256_max_epu8(a, b); }
+
+SIMD_INLINE v256 v256_min_s8(v256 a, v256 b) { return _mm256_min_epi8(a, b); }
+
+SIMD_INLINE v256 v256_max_s8(v256 a, v256 b) { return _mm256_max_epi8(a, b); }
+
+SIMD_INLINE v256 v256_min_s16(v256 a, v256 b) { return _mm256_min_epi16(a, b); }
+
+SIMD_INLINE v256 v256_max_s16(v256 a, v256 b) { return _mm256_max_epi16(a, b); }
+
+SIMD_INLINE v256 v256_cmpgt_s8(v256 a, v256 b) {
+  return _mm256_cmpgt_epi8(a, b);
+}
+
+SIMD_INLINE v256 v256_cmplt_s8(v256 a, v256 b) {
+  return v256_andn(_mm256_cmpgt_epi8(b, a), _mm256_cmpeq_epi8(b, a));
+}
+
+SIMD_INLINE v256 v256_cmpeq_8(v256 a, v256 b) {
+  return _mm256_cmpeq_epi8(a, b);
+}
+
+SIMD_INLINE v256 v256_cmpgt_s16(v256 a, v256 b) {
+  return _mm256_cmpgt_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_cmplt_s16(v256 a, v256 b) {
+  return v256_andn(_mm256_cmpgt_epi16(b, a), _mm256_cmpeq_epi16(b, a));
+}
+
+SIMD_INLINE v256 v256_cmpeq_16(v256 a, v256 b) {
+  return _mm256_cmpeq_epi16(a, b);
+}
+
+SIMD_INLINE v256 v256_shl_8(v256 a, unsigned int c) {
+  return _mm256_and_si256(_mm256_set1_epi8((uint8_t)(0xff << c)),
+                          _mm256_sll_epi16(a, _mm_cvtsi32_si128(c)));
+}
+
+SIMD_INLINE v256 v256_shr_u8(v256 a, unsigned int c) {
+  return _mm256_and_si256(_mm256_set1_epi8(0xff >> c),
+                          _mm256_srl_epi16(a, _mm_cvtsi32_si128(c)));
+}
+
+SIMD_INLINE v256 v256_shr_s8(v256 a, unsigned int c) {
+  __m128i x = _mm_cvtsi32_si128(c + 8);
+  return _mm256_packs_epi16(_mm256_sra_epi16(_mm256_unpacklo_epi8(a, a), x),
+                            _mm256_sra_epi16(_mm256_unpackhi_epi8(a, a), x));
+}
+
+SIMD_INLINE v256 v256_shl_16(v256 a, unsigned int c) {
+  return _mm256_sll_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v256 v256_shr_u16(v256 a, unsigned int c) {
+  return _mm256_srl_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v256 v256_shr_s16(v256 a, unsigned int c) {
+  return _mm256_sra_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v256 v256_shl_32(v256 a, unsigned int c) {
+  return _mm256_sll_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v256 v256_shr_u32(v256 a, unsigned int c) {
+  return _mm256_srl_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v256 v256_shr_s32(v256 a, unsigned int c) {
+  return _mm256_sra_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+/* These intrinsics require immediate values, so we must use #defines
+   to enforce that. */
+// _mm256_slli_si256 works on 128 bit lanes and can't be used
+#define v256_shl_n_byte(a, n)                                                 \
+  ((n) < 16                                                                   \
+       ? v256_from_v128(v128_or(v128_shl_n_byte(v256_high_v128(a), n),        \
+                                v128_shr_n_byte(v256_low_v128(a), 16 - (n))), \
+                        v128_shl_n_byte(v256_low_v128(a), n))                 \
+       : v256_from_v128(v128_shl_n_byte(v256_low_v128(a), (n)-16),            \
+                        v128_zero()))
+
+// _mm256_srli_si256 works on 128 bit lanes and can't be used
+#define v256_shr_n_byte(a, n)                                                 \
+  ((n) < 16                                                                   \
+       ? _mm256_alignr_epi8(                                                  \
+             _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(2, 0, 0, 1)), a, n)  \
+       : ((n) > 16                                                            \
+              ? _mm256_srli_si256(                                            \
+                    _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(2, 0, 0, 1)), \
+                    (n)-16)                                                   \
+              : _mm256_permute2x128_si256(a, a, _MM_SHUFFLE(2, 0, 0, 1))))
+
+// _mm256_alignr_epi8 works on two 128 bit lanes and can't be used
+#define v256_align(a, b, c) \
+  ((c) ? v256_or(v256_shr_n_byte(b, c), v256_shl_n_byte(a, 32 - c)) : b)
+
+#define v256_shl_n_8(a, c)                                   \
+  _mm256_and_si256(_mm256_set1_epi8((uint8_t)(0xff << (c))), \
+                   _mm256_slli_epi16(a, c))
+#define v256_shr_n_u8(a, c) \
+  _mm256_and_si256(_mm256_set1_epi8(0xff >> (c)), _mm256_srli_epi16(a, c))
+#define v256_shr_n_s8(a, c)                                                  \
+  _mm256_packs_epi16(_mm256_srai_epi16(_mm256_unpacklo_epi8(a, a), (c) + 8), \
+                     _mm256_srai_epi16(_mm256_unpackhi_epi8(a, a), (c) + 8))
+#define v256_shl_n_16(a, c) _mm256_slli_epi16(a, c)
+#define v256_shr_n_u16(a, c) _mm256_srli_epi16(a, c)
+#define v256_shr_n_s16(a, c) _mm256_srai_epi16(a, c)
+#define v256_shl_n_32(a, c) _mm256_slli_epi32(a, c)
+#define v256_shr_n_u32(a, c) _mm256_srli_epi32(a, c)
+#define v256_shr_n_s32(a, c) _mm256_srai_epi32(a, c)
+#endif
+
+#endif /* _V256_INTRINSICS_H */
diff --git a/third_party/aom/aom_dsp/simd/v64_intrinsics.h b/third_party/aom/aom_dsp/simd/v64_intrinsics.h
new file mode 100644
index 0000000000..ee2b683a44
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v64_intrinsics.h
@@ -0,0 +1,223 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V64_INTRINSICS_H
+#define _V64_INTRINSICS_H
+
+#include <stdio.h>
+#include <stdlib.h>
+#include "./v64_intrinsics_c.h"
+
+/* Fallback to plain, unoptimised C. */
+
+typedef c_v64 v64;
+
+SIMD_INLINE uint32_t v64_low_u32(v64 a) { return c_v64_low_u32(a); }
+SIMD_INLINE uint32_t v64_high_u32(v64 a) { return c_v64_high_u32(a); }
+SIMD_INLINE int32_t v64_low_s32(v64 a) { return c_v64_low_s32(a); }
+SIMD_INLINE int32_t v64_high_s32(v64 a) { return c_v64_high_s32(a); }
+SIMD_INLINE v64 v64_from_32(uint32_t x, uint32_t y) {
+  return c_v64_from_32(x, y);
+}
+SIMD_INLINE v64 v64_from_64(uint64_t x) { return c_v64_from_64(x); }
+SIMD_INLINE uint64_t v64_u64(v64 x) { return c_v64_u64(x); }
+SIMD_INLINE v64 v64_from_16(uint16_t a, uint16_t b, uint16_t c, uint16_t d) {
+  return c_v64_from_16(a, b, c, d);
+}
+
+SIMD_INLINE uint32_t u32_load_unaligned(const void *p) {
+  return c_u32_load_unaligned(p);
+}
+SIMD_INLINE uint32_t u32_load_aligned(const void *p) {
+  return c_u32_load_aligned(p);
+}
+SIMD_INLINE void u32_store_unaligned(void *p, uint32_t a) {
+  c_u32_store_unaligned(p, a);
+}
+SIMD_INLINE void u32_store_aligned(void *p, uint32_t a) {
+  c_u32_store_aligned(p, a);
+}
+
+SIMD_INLINE v64 v64_load_unaligned(const void *p) {
+  return c_v64_load_unaligned(p);
+}
+SIMD_INLINE v64 v64_load_aligned(const void *p) {
+  return c_v64_load_aligned(p);
+}
+
+SIMD_INLINE void v64_store_unaligned(void *p, v64 a) {
+  c_v64_store_unaligned(p, a);
+}
+SIMD_INLINE void v64_store_aligned(void *p, v64 a) {
+  c_v64_store_aligned(p, a);
+}
+
+SIMD_INLINE v64 v64_align(v64 a, v64 b, c) { return c_v64_align(a, b, c); }
+
+SIMD_INLINE v64 v64_zero() { return c_v64_zero(); }
+SIMD_INLINE v64 v64_dup_8(uint8_t x) { return c_v64_dup_8(x); }
+SIMD_INLINE v64 v64_dup_16(uint16_t x) { return c_v64_dup_16(x); }
+SIMD_INLINE v64 v64_dup_32(uint32_t x) { return c_v64_dup_32(x); }
+
+SIMD_INLINE v64 v64_add_8(v64 a, v64 b) { return c_v64_add_8(a, b); }
+SIMD_INLINE v64 v64_add_16(v64 a, v64 b) { return c_v64_add_16(a, b); }
+SIMD_INLINE v64 v64_sadd_s16(v64 a, v64 b) { return c_v64_sadd_s16(a, b); }
+SIMD_INLINE v64 v64_add_32(v64 a, v64 b) { return c_v64_add_32(a, b); }
+SIMD_INLINE v64 v64_sub_8(v64 a, v64 b) { return c_v64_sub_8(a, b); }
+SIMD_INLINE v64 v64_ssub_u8(v64 a, v64 b) { return c_v64_ssub_u8(a, b); }
+SIMD_INLINE v64 v64_ssub_s8(v64 a, v64 b) { return c_v64_ssub_s8(a, b); }
+SIMD_INLINE v64 v64_sub_16(v64 a, v64 b) { return c_v64_sub_16(a, b); }
+SIMD_INLINE v64 v64_ssub_s16(v64 a, v64 b) { return c_v64_ssub_s16(a, b); }
+SIMD_INLINE v64 v64_ssub_u16(v64 a, v64 b) { return c_v64_ssub_u16(a, b); }
+SIMD_INLINE v64 v64_sub_32(v64 a, v64 b) { return c_v64_sub_32(a, b); }
+SIMD_INLINE v64 v64_abs_s16(v64 a) { return c_v64_abs_s16(a); }
+SIMD_INLINE v64 v64_abs_s8(v64 a) { return c_v64_abs_s8(a); }
+
+SIMD_INLINE v64 v64_ziplo_8(v64 a, v64 b) { return c_v64_ziplo_8(a, b); }
+SIMD_INLINE v64 v64_ziphi_8(v64 a, v64 b) { return c_v64_ziphi_8(a, b); }
+SIMD_INLINE v64 v64_ziplo_16(v64 a, v64 b) { return c_v64_ziplo_16(a, b); }
+SIMD_INLINE v64 v64_ziphi_16(v64 a, v64 b) { return c_v64_ziphi_16(a, b); }
+SIMD_INLINE v64 v64_ziplo_32(v64 a, v64 b) { return c_v64_ziplo_32(a, b); }
+SIMD_INLINE v64 v64_ziphi_32(v64 a, v64 b) { return c_v64_ziphi_32(a, b); }
+SIMD_INLINE v64 v64_unziplo_8(v64 a, v64 b) { return c_v64_unziplo_8(a, b); }
+SIMD_INLINE v64 v64_unziphi_8(v64 a, v64 b) { return c_v64_unziphi_8(a, b); }
+SIMD_INLINE v64 v64_unziplo_16(v64 a, v64 b) { return c_v64_unziplo_16(a, b); }
+SIMD_INLINE v64 v64_unziphi_16(v64 a, v64 b) { return c_v64_unziphi_16(a, b); }
+SIMD_INLINE v64 v64_unpacklo_u8_s16(v64 a) { return c_v64_unpacklo_u8_s16(a); }
+SIMD_INLINE v64 v64_unpackhi_u8_s16(v64 a) { return c_v64_unpackhi_u8_s16(a); }
+SIMD_INLINE v64 v64_unpacklo_s8_s16(v64 a) { return c_v64_unpacklo_s8_s16(a); }
+SIMD_INLINE v64 v64_unpackhi_s8_s16(v64 a) { return c_v64_unpackhi_s8_s16(a); }
+SIMD_INLINE v64 v64_pack_s32_s16(v64 a, v64 b) {
+  return c_v64_pack_s32_s16(a, b);
+}
+SIMD_INLINE v64 v64_pack_s16_u8(v64 a, v64 b) {
+  return c_v64_pack_s16_u8(a, b);
+}
+SIMD_INLINE v64 v64_pack_s16_s8(v64 a, v64 b) {
+  return c_v64_pack_s16_s8(a, b);
+}
+SIMD_INLINE v64 v64_unpacklo_u16_s32(v64 a) {
+  return c_v64_unpacklo_u16_s32(a);
+}
+SIMD_INLINE v64 v64_unpacklo_s16_s32(v64 a) {
+  return c_v64_unpacklo_s16_s32(a);
+}
+SIMD_INLINE v64 v64_unpackhi_u16_s32(v64 a) {
+  return c_v64_unpackhi_u16_s32(a);
+}
+SIMD_INLINE v64 v64_unpackhi_s16_s32(v64 a) {
+  return c_v64_unpackhi_s16_s32(a);
+}
+SIMD_INLINE v64 v64_shuffle_8(v64 a, v64 pattern) {
+  return c_v64_shuffle_8(a, pattern);
+}
+
+typedef uint32_t sad64_internal;
+SIMD_INLINE sad64_internal v64_sad_u8_init() { return c_v64_sad_u8_init(); }
+SIMD_INLINE sad64_internal v64_sad_u8(sad64_internal s, v64 a, v64 b) {
+  return c_v64_sad_u8(s, a, b);
+}
+SIMD_INLINE uint32_t v64_sad_u8_sum(sad64_internal s) {
+  return c_v64_sad_u8_sum(s);
+}
+typedef uint32_t ssd64_internal;
+SIMD_INLINE ssd64_internal v64_ssd_u8_init() { return c_v64_ssd_u8_init(); }
+SIMD_INLINE ssd64_internal v64_ssd_u8(ssd64_internal s, v64 a, v64 b) {
+  return c_v64_ssd_u8(s, a, b);
+}
+SIMD_INLINE uint32_t v64_ssd_u8_sum(ssd64_internal s) {
+  return c_v64_ssd_u8_sum(s);
+}
+SIMD_INLINE int64_t v64_dotp_su8(v64 a, v64 b) { return c_v64_dotp_su8(a, b); }
+SIMD_INLINE int64_t v64_dotp_s16(v64 a, v64 b) { return c_v64_dotp_s16(a, b); }
+SIMD_INLINE uint64_t v64_hadd_u8(v64 a) { return c_v64_hadd_u8(a); }
+SIMD_INLINE int64_t v64_hadd_s16(v64 a) { return c_v64_hadd_s16(a); }
+
+SIMD_INLINE v64 v64_or(v64 a, v64 b) { return c_v64_or(a, b); }
+SIMD_INLINE v64 v64_xor(v64 a, v64 b) { return c_v64_xor(a, b); }
+SIMD_INLINE v64 v64_and(v64 a, v64 b) { return c_v64_and(a, b); }
+SIMD_INLINE v64 v64_andn(v64 a, v64 b) { return c_v64_andn(a, b); }
+
+SIMD_INLINE v64 v64_mullo_s16(v64 a, v64 b) { return c_v64_mullo_s16(a, b); }
+SIMD_INLINE v64 v64_mulhi_s16(v64 a, v64 b) { return c_v64_mulhi_s16(a, b); }
+SIMD_INLINE v64 v64_mullo_s32(v64 a, v64 b) { return c_v64_mullo_s32(a, b); }
+SIMD_INLINE v64 v64_madd_s16(v64 a, v64 b) { return c_v64_madd_s16(a, b); }
+SIMD_INLINE v64 v64_madd_us8(v64 a, v64 b) { return c_v64_madd_us8(a, b); }
+
+SIMD_INLINE v64 v64_avg_u8(v64 a, v64 b) { return c_v64_avg_u8(a, b); }
+SIMD_INLINE v64 v64_rdavg_u8(v64 a, v64 b) { return c_v64_rdavg_u8(a, b); }
+SIMD_INLINE v64 v64_avg_u16(v64 a, v64 b) { return c_v64_avg_u16(a, b); }
+SIMD_INLINE v64 v64_min_u8(v64 a, v64 b) { return c_v64_min_u8(a, b); }
+SIMD_INLINE v64 v64_max_u8(v64 a, v64 b) { return c_v64_max_u8(a, b); }
+SIMD_INLINE v64 v64_min_s8(v64 a, v64 b) { return c_v64_min_s8(a, b); }
+SIMD_INLINE v64 v64_max_s8(v64 a, v64 b) { return c_v64_max_s8(a, b); }
+SIMD_INLINE v64 v64_min_s16(v64 a, v64 b) { return c_v64_min_s16(a, b); }
+SIMD_INLINE v64 v64_max_s16(v64 a, v64 b) { return c_v64_max_s16(a, b); }
+
+SIMD_INLINE v64 v64_cmpgt_s8(v64 a, v64 b) { return c_v64_cmpgt_s8(a, b); }
+SIMD_INLINE v64 v64_cmplt_s8(v64 a, v64 b) { return c_v64_cmplt_s8(a, b); }
+SIMD_INLINE v64 v64_cmpeq_8(v64 a, v64 b) { return c_v64_cmpeq_8(a, b); }
+SIMD_INLINE v64 v64_cmpgt_s16(v64 a, v64 b) { return c_v64_cmpgt_s16(a, b); }
+SIMD_INLINE v64 v64_cmplt_s16(v64 a, v64 b) { return c_v64_cmplt_s16(a, b); }
+SIMD_INLINE v64 v64_cmpeq_16(v64 a, v64 b) { return c_v64_cmpeq_16(a, b); }
+
+SIMD_INLINE v64 v64_shl_8(v64 a, unsigned int n) { return c_v64_shl_8(a, n); }
+SIMD_INLINE v64 v64_shr_u8(v64 a, unsigned int n) { return c_v64_shr_u8(a, n); }
+SIMD_INLINE v64 v64_shr_s8(v64 a, unsigned int n) { return c_v64_shr_s8(a, n); }
+SIMD_INLINE v64 v64_shl_16(v64 a, unsigned int n) { return c_v64_shl_16(a, n); }
+SIMD_INLINE v64 v64_shr_u16(v64 a, unsigned int n) {
+  return c_v64_shr_u16(a, n);
+}
+SIMD_INLINE v64 v64_shr_s16(v64 a, unsigned int n) {
+  return c_v64_shr_s16(a, n);
+}
+SIMD_INLINE v64 v64_shl_32(v64 a, unsigned int n) { return c_v64_shl_32(a, n); }
+SIMD_INLINE v64 v64_shr_u32(v64 a, unsigned int n) {
+  return c_v64_shr_u32(a, n);
+}
+SIMD_INLINE v64 v64_shr_s32(v64 a, unsigned int n) {
+  return c_v64_shr_s32(a, n);
+}
+SIMD_INLINE v64 v64_shr_n_byte(v64 a, unsigned int n) {
+  return c_v64_shr_n_byte(a, n);
+}
+SIMD_INLINE v64 v64_shl_n_byte(v64 a, unsigned int n) {
+  return c_v64_shl_n_byte(a, n);
+}
+SIMD_INLINE v64 v64_shl_n_8(v64 a, unsigned int c) {
+  return c_v64_shl_n_8(a, c);
+}
+SIMD_INLINE v64 v64_shr_n_u8(v64 a, unsigned int c) {
+  return c_v64_shr_n_u8(a, c);
+}
+SIMD_INLINE v64 v64_shr_n_s8(v64 a, unsigned int c) {
+  return c_v64_shr_n_s8(a, c);
+}
+SIMD_INLINE v64 v64_shl_n_16(v64 a, unsigned int c) {
+  return c_v64_shl_n_16(a, c);
+}
+SIMD_INLINE v64 v64_shr_n_u16(v64 a, unsigned int c) {
+  return c_v64_shr_n_u16(a, c);
+}
+SIMD_INLINE v64 v64_shr_n_s16(v64 a, unsigned int c) {
+  return c_v64_shr_n_s16(a, c);
+}
+SIMD_INLINE v64 v64_shl_n_32(v64 a, unsigned int c) {
+  return c_v64_shl_n_32(a, c);
+}
+SIMD_INLINE v64 v64_shr_n_u32(v64 a, unsigned int c) {
+  return c_v64_shr_n_u32(a, c);
+}
+SIMD_INLINE v64 v64_shr_n_s32(v64 a, unsigned int c) {
+  return c_v64_shr_n_s32(a, c);
+}
+
+#endif /* _V64_INTRINSICS_H */
diff --git a/third_party/aom/aom_dsp/simd/v64_intrinsics_arm.h b/third_party/aom/aom_dsp/simd/v64_intrinsics_arm.h
new file mode 100644
index 0000000000..c7574eef50
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v64_intrinsics_arm.h
@@ -0,0 +1,583 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V64_INTRINSICS_H
+#define _V64_INTRINSICS_H
+
+#include <arm_neon.h>
+#include "./v64_intrinsics_arm.h"
+#include "aom_ports/arm.h"
+
+#ifdef AOM_INCOMPATIBLE_GCC
+#error Incompatible gcc
+#endif
+
+typedef int64x1_t v64;
+
+SIMD_INLINE uint32_t v64_low_u32(v64 a) {
+  return vget_lane_u32(vreinterpret_u32_s64(a), 0);
+}
+
+SIMD_INLINE uint32_t v64_high_u32(v64 a) {
+  return vget_lane_u32(vreinterpret_u32_s64(a), 1);
+}
+
+SIMD_INLINE int32_t v64_low_s32(v64 a) {
+  return vget_lane_s32(vreinterpret_s32_s64(a), 0);
+}
+
+SIMD_INLINE int32_t v64_high_s32(v64 a) {
+  return vget_lane_s32(vreinterpret_s32_s64(a), 1);
+}
+
+SIMD_INLINE v64 v64_from_16(uint16_t a, uint16_t b, uint16_t c, uint16_t d) {
+  return vcreate_s64((uint64_t)a << 48 | (uint64_t)b << 32 | (uint64_t)c << 16 |
+                     d);
+}
+
+SIMD_INLINE v64 v64_from_32(uint32_t x, uint32_t y) {
+  return vcreate_s64((uint64_t)x << 32 | y);
+}
+
+SIMD_INLINE v64 v64_from_64(uint64_t x) { return vcreate_s64(x); }
+
+SIMD_INLINE uint64_t v64_u64(v64 x) { return (uint64_t)x; }
+
+SIMD_INLINE uint32_t u32_load_aligned(const void *p) {
+  return *((uint32_t *)p);
+}
+
+SIMD_INLINE uint32_t u32_load_unaligned(const void *p) {
+  return vget_lane_u32(vreinterpret_u32_u8(vld1_u8((const uint8_t *)p)), 0);
+}
+
+SIMD_INLINE void u32_store_aligned(void *p, uint32_t a) {
+  *((uint32_t *)p) = a;
+}
+
+SIMD_INLINE void u32_store_unaligned(void *p, uint32_t a) {
+#if defined(__clang__)
+  vst1_lane_u32((uint32_t *)p, vreinterpret_u32_s64((uint64x1_t)(uint64_t)a),
+                0);
+#elif defined(__CC_ARM)
+  *(__packed uint32_t *)p) = a;
+#elif defined(__GNUC__)
+  *((__attribute((packed)) uint32_t *)p) = a;
+#else
+  vst1_lane_u32((uint32_t *)p, vreinterpret_u32_s64((uint64x1_t)(uint64_t)a),
+                0);
+#endif
+}
+
+SIMD_INLINE v64 v64_load_aligned(const void *p) {
+  return vreinterpret_s64_u8(vld1_u8((const uint8_t *)p));
+}
+
+SIMD_INLINE v64 v64_load_unaligned(const void *p) {
+  return v64_load_aligned(p);
+}
+
+SIMD_INLINE void v64_store_aligned(void *p, v64 r) {
+  vst1_u8((uint8_t *)p, vreinterpret_u8_s64(r));
+}
+
+SIMD_INLINE void v64_store_unaligned(void *p, v64 r) {
+  vst1_u8((uint8_t *)p, vreinterpret_u8_s64(r));
+}
+
+// The following function requires an immediate.
+// Some compilers will check this if it's optimising, others wont.
+SIMD_INLINE v64 v64_align(v64 a, v64 b, unsigned int c) {
+#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__)
+  return c ? vreinterpret_s64_s8(
+                 vext_s8(vreinterpret_s8_s64(b), vreinterpret_s8_s64(a), c))
+           : b;
+#else
+  return c ? v64_from_64(((uint64_t)b >> c * 8) | ((uint64_t)a << (8 - c) * 8))
+           : b;
+#endif
+}
+
+SIMD_INLINE v64 v64_zero() { return vreinterpret_s64_u8(vdup_n_u8(0)); }
+
+SIMD_INLINE v64 v64_dup_8(uint8_t x) {
+  return vreinterpret_s64_u8(vdup_n_u8(x));
+}
+
+SIMD_INLINE v64 v64_dup_16(uint16_t x) {
+  return vreinterpret_s64_u16(vdup_n_u16(x));
+}
+
+SIMD_INLINE v64 v64_dup_32(uint32_t x) {
+  return vreinterpret_s64_u32(vdup_n_u32(x));
+}
+
+SIMD_INLINE int64_t v64_dotp_su8(v64 x, v64 y) {
+  int64x2_t r = vpaddlq_s32(vpaddlq_s16(
+      vmulq_s16(vmovl_s8(vreinterpret_s8_s64(x)),
+                vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_s64(y))))));
+  return (int64_t)vadd_s64(vget_high_s64(r), vget_low_s64(r));
+}
+
+SIMD_INLINE int64_t v64_dotp_s16(v64 x, v64 y) {
+  int64x2_t r =
+      vpaddlq_s32(vmull_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+  return (int64_t)(vget_high_s64(r) + vget_low_s64(r));
+}
+
+SIMD_INLINE uint64_t v64_hadd_u8(v64 x) {
+  return (uint64_t)vpaddl_u32(vpaddl_u16(vpaddl_u8(vreinterpret_u8_s64(x))));
+}
+
+SIMD_INLINE int64_t v64_hadd_s16(v64 a) {
+  return (int64_t)vpaddl_s32(vpaddl_s16(vreinterpret_s16_s64(a)));
+}
+
+typedef uint16x8_t sad64_internal;
+
+SIMD_INLINE sad64_internal v64_sad_u8_init() { return vdupq_n_u16(0); }
+
+/* Implementation dependent return value.  Result must be finalised with
+   v64_sad_u8_sum().
+   The result for more than 32 v64_sad_u8() calls is undefined. */
+SIMD_INLINE sad64_internal v64_sad_u8(sad64_internal s, v64 a, v64 b) {
+  return vabal_u8(s, vreinterpret_u8_s64(a), vreinterpret_u8_s64(b));
+}
+
+SIMD_INLINE uint32_t v64_sad_u8_sum(sad64_internal s) {
+  uint64x2_t r = vpaddlq_u32(vpaddlq_u16(s));
+  return (uint32_t)(uint64_t)(vget_high_u64(r) + vget_low_u64(r));
+}
+
+typedef int64x1_t ssd64_internal;
+
+SIMD_INLINE ssd64_internal v64_ssd_u8_init() {
+  return (ssd64_internal)(uint64_t)0;
+}
+
+/* Implementation dependent return value.  Result must be finalised with
+ * v64_ssd_u8_sum(). */
+SIMD_INLINE ssd64_internal v64_ssd_u8(ssd64_internal s, v64 a, v64 b) {
+  uint8x8_t t = vabd_u8(vreinterpret_u8_s64(a), vreinterpret_u8_s64(b));
+  uint64x2_t r = vpaddlq_u32(vpaddlq_u16(vmull_u8(t, t)));
+  return vadd_u64(s, vadd_u64(vget_high_u64(r), vget_low_u64(r)));
+}
+
+SIMD_INLINE uint32_t v64_ssd_u8_sum(ssd64_internal s) {
+  return (uint32_t)(uint64_t)s;
+}
+
+SIMD_INLINE v64 v64_or(v64 x, v64 y) { return vorr_s64(x, y); }
+
+SIMD_INLINE v64 v64_xor(v64 x, v64 y) { return veor_s64(x, y); }
+
+SIMD_INLINE v64 v64_and(v64 x, v64 y) { return vand_s64(x, y); }
+
+SIMD_INLINE v64 v64_andn(v64 x, v64 y) { return vbic_s64(x, y); }
+
+SIMD_INLINE v64 v64_add_8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(
+      vadd_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_add_16(v64 x, v64 y) {
+  return vreinterpret_s64_s16(
+      vadd_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_sadd_s16(v64 x, v64 y) {
+  return vreinterpret_s64_s16(
+      vqadd_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_add_32(v64 x, v64 y) {
+  return vreinterpret_s64_u32(
+      vadd_u32(vreinterpret_u32_s64(x), vreinterpret_u32_s64(y)));
+}
+
+SIMD_INLINE v64 v64_sub_8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(
+      vsub_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_sub_16(v64 x, v64 y) {
+  return vreinterpret_s64_s16(
+      vsub_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_ssub_s16(v64 x, v64 y) {
+  return vreinterpret_s64_s16(
+      vqsub_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_ssub_u16(v64 x, v64 y) {
+  return vreinterpret_s64_u16(
+      vqsub_u16(vreinterpret_u16_s64(x), vreinterpret_u16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_ssub_u8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(
+      vqsub_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_ssub_s8(v64 x, v64 y) {
+  return vreinterpret_s64_s8(
+      vqsub_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_sub_32(v64 x, v64 y) {
+  return vreinterpret_s64_s32(
+      vsub_s32(vreinterpret_s32_s64(x), vreinterpret_s32_s64(y)));
+}
+
+SIMD_INLINE v64 v64_abs_s16(v64 x) {
+  return vreinterpret_s64_s16(vabs_s16(vreinterpret_s16_s64(x)));
+}
+
+SIMD_INLINE v64 v64_abs_s8(v64 x) {
+  return vreinterpret_s64_s8(vabs_s8(vreinterpret_s8_s64(x)));
+}
+
+SIMD_INLINE v64 v64_mullo_s16(v64 x, v64 y) {
+  return vreinterpret_s64_s16(
+      vmul_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_mulhi_s16(v64 x, v64 y) {
+  return vreinterpret_s64_s16(vmovn_s32(vshrq_n_s32(
+      vmull_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)), 16)));
+}
+
+SIMD_INLINE v64 v64_mullo_s32(v64 x, v64 y) {
+  return vreinterpret_s64_s32(
+      vmul_s32(vreinterpret_s32_s64(x), vreinterpret_s32_s64(y)));
+}
+
+SIMD_INLINE v64 v64_madd_s16(v64 x, v64 y) {
+  int32x4_t t = vmull_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y));
+  return vreinterpret_s64_s32(
+      vpadd_s32(vreinterpret_s32_s64(vget_low_s64(vreinterpretq_s64_s32(t))),
+                vreinterpret_s32_s64(vget_high_s64(vreinterpretq_s64_s32(t)))));
+}
+
+SIMD_INLINE v64 v64_madd_us8(v64 x, v64 y) {
+  return vreinterpret_s64_s16(vqmovn_s32(vpaddlq_s16(
+      vaddq_s16(vmull_s8(vadd_s8(vreinterpret_s8_s64(x), vdup_n_s8(-128)),
+                         vreinterpret_s8_s64(y)),
+                vshlq_n_s16(vmovl_s8(vreinterpret_s8_s64(y)), 7)))));
+}
+
+SIMD_INLINE v64 v64_avg_u8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(
+      vrhadd_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_rdavg_u8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(
+      vhadd_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_avg_u16(v64 x, v64 y) {
+  return vreinterpret_s64_u16(
+      vrhadd_u16(vreinterpret_u16_s64(x), vreinterpret_u16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_max_u8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(
+      vmax_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_min_u8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(
+      vmin_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_max_s8(v64 x, v64 y) {
+  return vreinterpret_s64_s8(
+      vmax_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_min_s8(v64 x, v64 y) {
+  return vreinterpret_s64_s8(
+      vmin_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_max_s16(v64 x, v64 y) {
+  return vreinterpret_s64_s16(
+      vmax_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_min_s16(v64 x, v64 y) {
+  return vreinterpret_s64_s16(
+      vmin_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_ziplo_8(v64 x, v64 y) {
+  uint8x8x2_t r = vzip_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x));
+  return vreinterpret_s64_u8(r.val[0]);
+}
+
+SIMD_INLINE v64 v64_ziphi_8(v64 x, v64 y) {
+  uint8x8x2_t r = vzip_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x));
+  return vreinterpret_s64_u8(r.val[1]);
+}
+
+SIMD_INLINE v64 v64_ziplo_16(v64 x, v64 y) {
+  int16x4x2_t r = vzip_s16(vreinterpret_s16_s64(y), vreinterpret_s16_s64(x));
+  return vreinterpret_s64_s16(r.val[0]);
+}
+
+SIMD_INLINE v64 v64_ziphi_16(v64 x, v64 y) {
+  int16x4x2_t r = vzip_s16(vreinterpret_s16_s64(y), vreinterpret_s16_s64(x));
+  return vreinterpret_s64_s16(r.val[1]);
+}
+
+SIMD_INLINE v64 v64_ziplo_32(v64 x, v64 y) {
+  int32x2x2_t r = vzip_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x));
+  return vreinterpret_s64_s32(r.val[0]);
+}
+
+SIMD_INLINE v64 v64_ziphi_32(v64 x, v64 y) {
+  int32x2x2_t r = vzip_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x));
+  return vreinterpret_s64_s32(r.val[1]);
+}
+
+SIMD_INLINE v64 v64_unpacklo_u8_s16(v64 a) {
+  return vreinterpret_s64_u16(vget_low_u16(vmovl_u8(vreinterpret_u8_s64(a))));
+}
+
+SIMD_INLINE v64 v64_unpackhi_u8_s16(v64 a) {
+  return vreinterpret_s64_u16(vget_high_u16(vmovl_u8(vreinterpret_u8_s64(a))));
+}
+
+SIMD_INLINE v64 v64_unpacklo_s8_s16(v64 a) {
+  return vreinterpret_s64_s16(vget_low_s16(vmovl_s8(vreinterpret_s8_s64(a))));
+}
+
+SIMD_INLINE v64 v64_unpackhi_s8_s16(v64 a) {
+  return vreinterpret_s64_s16(vget_high_s16(vmovl_s8(vreinterpret_s8_s64(a))));
+}
+
+SIMD_INLINE v64 v64_pack_s32_s16(v64 x, v64 y) {
+  return vreinterpret_s64_s16(vqmovn_s32(
+      vcombine_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x))));
+}
+
+SIMD_INLINE v64 v64_pack_s16_u8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(vqmovun_s16(vreinterpretq_s16_s32(
+      vcombine_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x)))));
+}
+
+SIMD_INLINE v64 v64_pack_s16_s8(v64 x, v64 y) {
+  return vreinterpret_s64_s8(vqmovn_s16(vreinterpretq_s16_s32(
+      vcombine_s32(vreinterpret_s32_s64(y), vreinterpret_s32_s64(x)))));
+}
+
+SIMD_INLINE v64 v64_unziplo_8(v64 x, v64 y) {
+  uint8x8x2_t r = vuzp_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x));
+  return vreinterpret_s64_u8(r.val[0]);
+}
+
+SIMD_INLINE v64 v64_unziphi_8(v64 x, v64 y) {
+  uint8x8x2_t r = vuzp_u8(vreinterpret_u8_s64(y), vreinterpret_u8_s64(x));
+  return vreinterpret_s64_u8(r.val[1]);
+}
+
+SIMD_INLINE v64 v64_unziplo_16(v64 x, v64 y) {
+  uint16x4x2_t r = vuzp_u16(vreinterpret_u16_s64(y), vreinterpret_u16_s64(x));
+  return vreinterpret_s64_u16(r.val[0]);
+}
+
+SIMD_INLINE v64 v64_unziphi_16(v64 x, v64 y) {
+  uint16x4x2_t r = vuzp_u16(vreinterpret_u16_s64(y), vreinterpret_u16_s64(x));
+  return vreinterpret_s64_u16(r.val[1]);
+}
+
+SIMD_INLINE v64 v64_unpacklo_s16_s32(v64 x) {
+  return vreinterpret_s64_s32(vget_low_s32(vmovl_s16(vreinterpret_s16_s64(x))));
+}
+
+SIMD_INLINE v64 v64_unpacklo_u16_s32(v64 x) {
+  return vreinterpret_s64_u32(vget_low_u32(vmovl_u16(vreinterpret_u16_s64(x))));
+}
+
+SIMD_INLINE v64 v64_unpackhi_s16_s32(v64 x) {
+  return vreinterpret_s64_s32(
+      vget_high_s32(vmovl_s16(vreinterpret_s16_s64(x))));
+}
+
+SIMD_INLINE v64 v64_unpackhi_u16_s32(v64 x) {
+  return vreinterpret_s64_u32(
+      vget_high_u32(vmovl_u16(vreinterpret_u16_s64(x))));
+}
+
+SIMD_INLINE v64 v64_shuffle_8(v64 x, v64 pattern) {
+  return vreinterpret_s64_u8(
+      vtbl1_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(pattern)));
+}
+
+SIMD_INLINE v64 v64_cmpgt_s8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(
+      vcgt_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_cmplt_s8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(
+      vclt_s8(vreinterpret_s8_s64(x), vreinterpret_s8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_cmpeq_8(v64 x, v64 y) {
+  return vreinterpret_s64_u8(
+      vceq_u8(vreinterpret_u8_s64(x), vreinterpret_u8_s64(y)));
+}
+
+SIMD_INLINE v64 v64_cmpgt_s16(v64 x, v64 y) {
+  return vreinterpret_s64_u16(
+      vcgt_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_cmplt_s16(v64 x, v64 y) {
+  return vreinterpret_s64_u16(
+      vclt_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_cmpeq_16(v64 x, v64 y) {
+  return vreinterpret_s64_u16(
+      vceq_s16(vreinterpret_s16_s64(x), vreinterpret_s16_s64(y)));
+}
+
+SIMD_INLINE v64 v64_shl_8(v64 a, unsigned int c) {
+  return vreinterpret_s64_u8(vshl_u8(vreinterpret_u8_s64(a), vdup_n_s8(c)));
+}
+
+SIMD_INLINE v64 v64_shr_u8(v64 a, unsigned int c) {
+  return vreinterpret_s64_u8(vshl_u8(vreinterpret_u8_s64(a), vdup_n_s8(-c)));
+}
+
+SIMD_INLINE v64 v64_shr_s8(v64 a, unsigned int c) {
+  return vreinterpret_s64_s8(vshl_s8(vreinterpret_s8_s64(a), vdup_n_s8(-c)));
+}
+
+SIMD_INLINE v64 v64_shl_16(v64 a, unsigned int c) {
+  return vreinterpret_s64_u16(vshl_u16(vreinterpret_u16_s64(a), vdup_n_s16(c)));
+}
+
+SIMD_INLINE v64 v64_shr_u16(v64 a, unsigned int c) {
+  return vreinterpret_s64_u16(
+      vshl_u16(vreinterpret_u16_s64(a), vdup_n_s16(-(int)c)));
+}
+
+SIMD_INLINE v64 v64_shr_s16(v64 a, unsigned int c) {
+  return vreinterpret_s64_s16(
+      vshl_s16(vreinterpret_s16_s64(a), vdup_n_s16(-(int)c)));
+}
+
+SIMD_INLINE v64 v64_shl_32(v64 a, unsigned int c) {
+  return vreinterpret_s64_u32(vshl_u32(vreinterpret_u32_s64(a), vdup_n_s32(c)));
+}
+
+SIMD_INLINE v64 v64_shr_u32(v64 a, unsigned int c) {
+  return vreinterpret_s64_u32(
+      vshl_u32(vreinterpret_u32_s64(a), vdup_n_s32(-(int)c)));
+}
+
+SIMD_INLINE v64 v64_shr_s32(v64 a, unsigned int c) {
+  return vreinterpret_s64_s32(
+      vshl_s32(vreinterpret_s32_s64(a), vdup_n_s32(-(int)c)));
+}
+
+// The following functions require an immediate.
+// Some compilers will check this during optimisation, others wont.
+#if defined(__OPTIMIZE__) && __OPTIMIZE__ && !defined(__clang__)
+
+SIMD_INLINE v64 v64_shl_n_byte(v64 a, unsigned int c) {
+  return vshl_n_s64(a, c * 8);
+}
+
+SIMD_INLINE v64 v64_shr_n_byte(v64 a, unsigned int c) {
+  return c ? (v64)vshr_n_u64(vreinterpret_u64_s64(a), c * 8) : a;
+}
+
+SIMD_INLINE v64 v64_shl_n_8(v64 a, unsigned int c) {
+  return vreinterpret_s64_u8(vshl_n_u8(vreinterpret_u8_s64(a), c));
+}
+
+SIMD_INLINE v64 v64_shr_n_u8(v64 a, unsigned int c) {
+  return vreinterpret_s64_u8(vshr_n_u8(vreinterpret_u8_s64(a), c));
+}
+
+SIMD_INLINE v64 v64_shr_n_s8(v64 a, unsigned int c) {
+  return vreinterpret_s64_s8(vshr_n_s8(vreinterpret_s8_s64(a), c));
+}
+
+SIMD_INLINE v64 v64_shl_n_16(v64 a, unsigned int c) {
+  return vreinterpret_s64_u16(vshl_n_u16(vreinterpret_u16_s64(a), c));
+}
+
+SIMD_INLINE v64 v64_shr_n_u16(v64 a, unsigned int c) {
+  return vreinterpret_s64_u16(vshr_n_u16(vreinterpret_u16_s64(a), c));
+}
+
+SIMD_INLINE v64 v64_shr_n_s16(v64 a, unsigned int c) {
+  return vreinterpret_s64_s16(vshr_n_s16(vreinterpret_s16_s64(a), c));
+}
+
+SIMD_INLINE v64 v64_shl_n_32(v64 a, unsigned int c) {
+  return vreinterpret_s64_u32(vshl_n_u32(vreinterpret_u32_s64(a), c));
+}
+
+SIMD_INLINE v64 v64_shr_n_u32(v64 a, unsigned int c) {
+  return vreinterpret_s64_u32(vshr_n_u32(vreinterpret_u32_s64(a), c));
+}
+
+SIMD_INLINE v64 v64_shr_n_s32(v64 a, unsigned int c) {
+  return vreinterpret_s64_s32(vshr_n_s32(vreinterpret_s32_s64(a), c));
+}
+
+#else
+
+SIMD_INLINE v64 v64_shl_n_byte(v64 a, unsigned int c) {
+  return v64_from_64(v64_u64(a) << c * 8);
+}
+
+SIMD_INLINE v64 v64_shr_n_byte(v64 a, unsigned int c) {
+  return v64_from_64(v64_u64(a) >> c * 8);
+}
+
+SIMD_INLINE v64 v64_shl_n_8(v64 a, unsigned int c) { return v64_shl_8(a, c); }
+
+SIMD_INLINE v64 v64_shr_n_u8(v64 a, unsigned int c) { return v64_shr_u8(a, c); }
+
+SIMD_INLINE v64 v64_shr_n_s8(v64 a, unsigned int c) { return v64_shr_s8(a, c); }
+
+SIMD_INLINE v64 v64_shl_n_16(v64 a, unsigned int c) { return v64_shl_16(a, c); }
+
+SIMD_INLINE v64 v64_shr_n_u16(v64 a, unsigned int c) {
+  return v64_shr_u16(a, c);
+}
+
+SIMD_INLINE v64 v64_shr_n_s16(v64 a, unsigned int c) {
+  return v64_shr_s16(a, c);
+}
+
+SIMD_INLINE v64 v64_shl_n_32(v64 a, unsigned int c) { return v64_shl_32(a, c); }
+
+SIMD_INLINE v64 v64_shr_n_u32(v64 a, unsigned int c) {
+  return v64_shr_u32(a, c);
+}
+
+SIMD_INLINE v64 v64_shr_n_s32(v64 a, unsigned int c) {
+  return v64_shr_s32(a, c);
+}
+
+#endif
+
+#endif /* _V64_INTRINSICS_H */
diff --git a/third_party/aom/aom_dsp/simd/v64_intrinsics_c.h b/third_party/aom/aom_dsp/simd/v64_intrinsics_c.h
new file mode 100644
index 0000000000..5032238b66
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v64_intrinsics_c.h
@@ -0,0 +1,919 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V64_INTRINSICS_C_H
+#define _V64_INTRINSICS_C_H
+
+/* Note: This implements the intrinsics in plain, unoptimised C.
+   Intended for reference, porting or debugging. */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include "./aom_config.h"
+
+typedef union {
+  uint8_t u8[8];
+  uint16_t u16[4];
+  uint32_t u32[2];
+  uint64_t u64;
+  int8_t s8[8];
+  int16_t s16[4];
+  int32_t s32[2];
+  int64_t s64;
+} c_v64;
+
+SIMD_INLINE uint32_t c_v64_low_u32(c_v64 a) { return a.u32[CONFIG_BIG_ENDIAN]; }
+
+SIMD_INLINE uint32_t c_v64_high_u32(c_v64 a) {
+  return a.u32[!CONFIG_BIG_ENDIAN];
+}
+
+SIMD_INLINE int32_t c_v64_low_s32(c_v64 a) { return a.s32[CONFIG_BIG_ENDIAN]; }
+
+SIMD_INLINE int32_t c_v64_high_s32(c_v64 a) {
+  return a.s32[!CONFIG_BIG_ENDIAN];
+}
+
+SIMD_INLINE c_v64 c_v64_from_32(uint32_t x, uint32_t y) {
+  c_v64 t;
+  t.u32[!CONFIG_BIG_ENDIAN] = x;
+  t.u32[CONFIG_BIG_ENDIAN] = y;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_from_64(uint64_t x) {
+  c_v64 t;
+  t.u64 = x;
+  return t;
+}
+
+SIMD_INLINE uint64_t c_v64_u64(c_v64 x) { return x.u64; }
+
+SIMD_INLINE c_v64 c_v64_from_16(uint16_t a, uint16_t b, uint16_t c,
+                                uint16_t d) {
+  c_v64 t;
+  if (CONFIG_BIG_ENDIAN) {
+    t.u16[0] = a;
+    t.u16[1] = b;
+    t.u16[2] = c;
+    t.u16[3] = d;
+  } else {
+    t.u16[3] = a;
+    t.u16[2] = b;
+    t.u16[1] = c;
+    t.u16[0] = d;
+  }
+  return t;
+}
+
+SIMD_INLINE uint32_t c_u32_load_unaligned(const void *p) {
+  uint32_t t;
+  uint8_t *pp = (uint8_t *)p;
+  uint8_t *q = (uint8_t *)&t;
+  int c;
+  for (c = 0; c < 4; c++) q[c] = pp[c];
+  return t;
+}
+
+SIMD_INLINE void c_u32_store_unaligned(void *p, uint32_t a) {
+  uint8_t *pp = (uint8_t *)p;
+  uint8_t *q = (uint8_t *)&a;
+  int c;
+  for (c = 0; c < 4; c++) pp[c] = q[c];
+}
+
+SIMD_INLINE uint32_t c_u32_load_aligned(const void *p) {
+  if (SIMD_CHECK && (uintptr_t)p & 3) {
+    fprintf(stderr, "Error: Unaligned u32 load at %p\n", p);
+    abort();
+  }
+  return c_u32_load_unaligned(p);
+}
+
+SIMD_INLINE void c_u32_store_aligned(void *p, uint32_t a) {
+  if (SIMD_CHECK && (uintptr_t)p & 3) {
+    fprintf(stderr, "Error: Unaligned u32 store at %p\n", p);
+    abort();
+  }
+  c_u32_store_unaligned(p, a);
+}
+
+SIMD_INLINE c_v64 c_v64_load_unaligned(const void *p) {
+  c_v64 t;
+  uint8_t *pp = (uint8_t *)p;
+  uint8_t *q = (uint8_t *)&t;
+  int c;
+  for (c = 0; c < 8; c++) q[c] = pp[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_load_aligned(const void *p) {
+  if (SIMD_CHECK && (uintptr_t)p & 7) {
+    fprintf(stderr, "Error: Unaligned c_v64 load at %p\n", p);
+    abort();
+  }
+  return c_v64_load_unaligned(p);
+}
+
+SIMD_INLINE void c_v64_store_unaligned(void *p, c_v64 a) {
+  uint8_t *q = (uint8_t *)p;
+  uint8_t *r = (uint8_t *)&a;
+  int c;
+  for (c = 0; c < 8; c++) q[c] = r[c];
+}
+
+SIMD_INLINE void c_v64_store_aligned(void *p, c_v64 a) {
+  if (SIMD_CHECK && (uintptr_t)p & 7) {
+    fprintf(stderr, "Error: Unaligned c_v64 store at %p\n", p);
+    abort();
+  }
+  c_v64_store_unaligned(p, a);
+}
+
+SIMD_INLINE c_v64 c_v64_zero() {
+  c_v64 t;
+  t.u64 = 0;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_dup_8(uint8_t x) {
+  c_v64 t;
+  t.u8[0] = t.u8[1] = t.u8[2] = t.u8[3] = t.u8[4] = t.u8[5] = t.u8[6] =
+      t.u8[7] = x;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_dup_16(uint16_t x) {
+  c_v64 t;
+  t.u16[0] = t.u16[1] = t.u16[2] = t.u16[3] = x;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_dup_32(uint32_t x) {
+  c_v64 t;
+  t.u32[0] = t.u32[1] = x;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_add_8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] + b.u8[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_add_16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++) t.u16[c] = a.u16[c] + b.u16[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_sadd_s16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++)
+    t.s16[c] = (int32_t)a.s16[c] + (int32_t)b.s16[c] > 32767
+                   ? 32767
+                   : (int32_t)a.s16[c] + (int32_t)b.s16[c] < -32768
+                         ? -32768
+                         : (int32_t)a.s16[c] + (int32_t)b.s16[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_add_32(c_v64 a, c_v64 b) {
+  c_v64 t;
+  t.u32[0] = (uint32_t)((uint64_t)a.u32[0] + b.u32[0]);
+  t.u32[1] = (uint32_t)((uint64_t)a.u32[1] + b.u32[1]);
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_sub_8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] - b.u8[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_ssub_u8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++)
+    t.u8[c] = (int32_t)a.u8[c] - (int32_t)b.u8[c] < 0 ? 0 : a.u8[c] - b.u8[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_ssub_s8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) {
+    int16_t d = (int16_t)a.s8[c] - (int16_t)b.s8[c];
+    t.s8[c] = d > 127 ? 127 : (d < -128 ? -128 : d);
+  }
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_sub_16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++) t.u16[c] = a.u16[c] - b.u16[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_ssub_s16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++)
+    t.s16[c] = (int32_t)a.s16[c] - (int32_t)b.s16[c] < -32768
+                   ? -32768
+                   : (int32_t)a.s16[c] - (int32_t)b.s16[c] > 32767
+                         ? 32767
+                         : (int32_t)a.s16[c] - (int32_t)b.s16[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_ssub_u16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++)
+    t.u16[c] =
+        (int32_t)a.u16[c] - (int32_t)b.u16[c] < 0 ? 0 : a.u16[c] - b.u16[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_sub_32(c_v64 a, c_v64 b) {
+  c_v64 t;
+  t.u32[0] = (uint32_t)((int64_t)a.u32[0] - b.u32[0]);
+  t.u32[1] = (uint32_t)((int64_t)a.u32[1] - b.u32[1]);
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_abs_s16(c_v64 a) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++)
+    t.u16[c] = (int16_t)a.u16[c] > 0 ? a.u16[c] : -a.u16[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_abs_s8(c_v64 a) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.u8[c] = (int8_t)a.u8[c] > 0 ? a.u8[c] : -a.u8[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 _c_v64_zip_8(c_v64 a, c_v64 b, int mode) {
+  c_v64 t;
+  if (mode) {
+    t.u8[7] = a.u8[7];
+    t.u8[6] = b.u8[7];
+    t.u8[5] = a.u8[6];
+    t.u8[4] = b.u8[6];
+    t.u8[3] = a.u8[5];
+    t.u8[2] = b.u8[5];
+    t.u8[1] = a.u8[4];
+    t.u8[0] = b.u8[4];
+  } else {
+    t.u8[7] = a.u8[3];
+    t.u8[6] = b.u8[3];
+    t.u8[5] = a.u8[2];
+    t.u8[4] = b.u8[2];
+    t.u8[3] = a.u8[1];
+    t.u8[2] = b.u8[1];
+    t.u8[1] = a.u8[0];
+    t.u8[0] = b.u8[0];
+  }
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_ziplo_8(c_v64 a, c_v64 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v64_zip_8(b, a, 1) : _c_v64_zip_8(a, b, 0);
+}
+
+SIMD_INLINE c_v64 c_v64_ziphi_8(c_v64 a, c_v64 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v64_zip_8(b, a, 0) : _c_v64_zip_8(a, b, 1);
+}
+
+SIMD_INLINE c_v64 _c_v64_zip_16(c_v64 a, c_v64 b, int mode) {
+  c_v64 t;
+  if (mode) {
+    t.u16[3] = a.u16[3];
+    t.u16[2] = b.u16[3];
+    t.u16[1] = a.u16[2];
+    t.u16[0] = b.u16[2];
+  } else {
+    t.u16[3] = a.u16[1];
+    t.u16[2] = b.u16[1];
+    t.u16[1] = a.u16[0];
+    t.u16[0] = b.u16[0];
+  }
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_ziplo_16(c_v64 a, c_v64 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v64_zip_16(b, a, 1) : _c_v64_zip_16(a, b, 0);
+}
+
+SIMD_INLINE c_v64 c_v64_ziphi_16(c_v64 a, c_v64 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v64_zip_16(b, a, 0) : _c_v64_zip_16(a, b, 1);
+}
+
+SIMD_INLINE c_v64 _c_v64_zip_32(c_v64 a, c_v64 b, int mode) {
+  c_v64 t;
+  if (mode) {
+    t.u32[1] = a.u32[1];
+    t.u32[0] = b.u32[1];
+  } else {
+    t.u32[1] = a.u32[0];
+    t.u32[0] = b.u32[0];
+  }
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_ziplo_32(c_v64 a, c_v64 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v64_zip_32(b, a, 1) : _c_v64_zip_32(a, b, 0);
+}
+
+SIMD_INLINE c_v64 c_v64_ziphi_32(c_v64 a, c_v64 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v64_zip_32(b, a, 0) : _c_v64_zip_32(a, b, 1);
+}
+
+SIMD_INLINE c_v64 _c_v64_unzip_8(c_v64 a, c_v64 b, int mode) {
+  c_v64 t;
+  if (mode) {
+    t.u8[7] = b.u8[7];
+    t.u8[6] = b.u8[5];
+    t.u8[5] = b.u8[3];
+    t.u8[4] = b.u8[1];
+    t.u8[3] = a.u8[7];
+    t.u8[2] = a.u8[5];
+    t.u8[1] = a.u8[3];
+    t.u8[0] = a.u8[1];
+  } else {
+    t.u8[7] = a.u8[6];
+    t.u8[6] = a.u8[4];
+    t.u8[5] = a.u8[2];
+    t.u8[4] = a.u8[0];
+    t.u8[3] = b.u8[6];
+    t.u8[2] = b.u8[4];
+    t.u8[1] = b.u8[2];
+    t.u8[0] = b.u8[0];
+  }
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_unziplo_8(c_v64 a, c_v64 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v64_unzip_8(a, b, 1) : _c_v64_unzip_8(a, b, 0);
+}
+
+SIMD_INLINE c_v64 c_v64_unziphi_8(c_v64 a, c_v64 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v64_unzip_8(b, a, 0) : _c_v64_unzip_8(b, a, 1);
+}
+
+SIMD_INLINE c_v64 _c_v64_unzip_16(c_v64 a, c_v64 b, int mode) {
+  c_v64 t;
+  if (mode) {
+    t.u16[3] = b.u16[3];
+    t.u16[2] = b.u16[1];
+    t.u16[1] = a.u16[3];
+    t.u16[0] = a.u16[1];
+  } else {
+    t.u16[3] = a.u16[2];
+    t.u16[2] = a.u16[0];
+    t.u16[1] = b.u16[2];
+    t.u16[0] = b.u16[0];
+  }
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_unziplo_16(c_v64 a, c_v64 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v64_unzip_16(a, b, 1)
+                           : _c_v64_unzip_16(a, b, 0);
+}
+
+SIMD_INLINE c_v64 c_v64_unziphi_16(c_v64 a, c_v64 b) {
+  return CONFIG_BIG_ENDIAN ? _c_v64_unzip_16(b, a, 0)
+                           : _c_v64_unzip_16(b, a, 1);
+}
+
+SIMD_INLINE c_v64 c_v64_unpacklo_u8_s16(c_v64 a) {
+  c_v64 t;
+  int endian = !!CONFIG_BIG_ENDIAN * 4;
+  t.s16[3] = (int16_t)a.u8[3 + endian];
+  t.s16[2] = (int16_t)a.u8[2 + endian];
+  t.s16[1] = (int16_t)a.u8[1 + endian];
+  t.s16[0] = (int16_t)a.u8[0 + endian];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_unpackhi_u8_s16(c_v64 a) {
+  c_v64 t;
+  int endian = !!CONFIG_BIG_ENDIAN * 4;
+  t.s16[3] = (int16_t)a.u8[7 - endian];
+  t.s16[2] = (int16_t)a.u8[6 - endian];
+  t.s16[1] = (int16_t)a.u8[5 - endian];
+  t.s16[0] = (int16_t)a.u8[4 - endian];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_unpacklo_s8_s16(c_v64 a) {
+  c_v64 t;
+  int endian = !!CONFIG_BIG_ENDIAN * 4;
+  t.s16[3] = (int16_t)a.s8[3 + endian];
+  t.s16[2] = (int16_t)a.s8[2 + endian];
+  t.s16[1] = (int16_t)a.s8[1 + endian];
+  t.s16[0] = (int16_t)a.s8[0 + endian];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_unpackhi_s8_s16(c_v64 a) {
+  c_v64 t;
+  int endian = !!CONFIG_BIG_ENDIAN * 4;
+  t.s16[3] = (int16_t)a.s8[7 - endian];
+  t.s16[2] = (int16_t)a.s8[6 - endian];
+  t.s16[1] = (int16_t)a.s8[5 - endian];
+  t.s16[0] = (int16_t)a.s8[4 - endian];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_pack_s32_s16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  if (CONFIG_BIG_ENDIAN) {
+    c_v64 u = a;
+    a = b;
+    b = u;
+  }
+  t.s16[3] = a.s32[1] > 32767 ? 32767 : a.s32[1] < -32768 ? -32768 : a.s32[1];
+  t.s16[2] = a.s32[0] > 32767 ? 32767 : a.s32[0] < -32768 ? -32768 : a.s32[0];
+  t.s16[1] = b.s32[1] > 32767 ? 32767 : b.s32[1] < -32768 ? -32768 : b.s32[1];
+  t.s16[0] = b.s32[0] > 32767 ? 32767 : b.s32[0] < -32768 ? -32768 : b.s32[0];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_pack_s16_u8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  if (CONFIG_BIG_ENDIAN) {
+    c_v64 u = a;
+    a = b;
+    b = u;
+  }
+  t.u8[7] = a.s16[3] > 255 ? 255 : a.s16[3] < 0 ? 0 : a.s16[3];
+  t.u8[6] = a.s16[2] > 255 ? 255 : a.s16[2] < 0 ? 0 : a.s16[2];
+  t.u8[5] = a.s16[1] > 255 ? 255 : a.s16[1] < 0 ? 0 : a.s16[1];
+  t.u8[4] = a.s16[0] > 255 ? 255 : a.s16[0] < 0 ? 0 : a.s16[0];
+  t.u8[3] = b.s16[3] > 255 ? 255 : b.s16[3] < 0 ? 0 : b.s16[3];
+  t.u8[2] = b.s16[2] > 255 ? 255 : b.s16[2] < 0 ? 0 : b.s16[2];
+  t.u8[1] = b.s16[1] > 255 ? 255 : b.s16[1] < 0 ? 0 : b.s16[1];
+  t.u8[0] = b.s16[0] > 255 ? 255 : b.s16[0] < 0 ? 0 : b.s16[0];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_pack_s16_s8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  if (CONFIG_BIG_ENDIAN) {
+    c_v64 u = a;
+    a = b;
+    b = u;
+  }
+  t.u8[7] = a.s16[3] > 127 ? 127 : a.s16[3] < -128 ? 128 : a.s16[3];
+  t.u8[6] = a.s16[2] > 127 ? 127 : a.s16[2] < -128 ? 128 : a.s16[2];
+  t.u8[5] = a.s16[1] > 127 ? 127 : a.s16[1] < -128 ? 128 : a.s16[1];
+  t.u8[4] = a.s16[0] > 127 ? 127 : a.s16[0] < -128 ? 128 : a.s16[0];
+  t.u8[3] = b.s16[3] > 127 ? 127 : b.s16[3] < -128 ? 128 : b.s16[3];
+  t.u8[2] = b.s16[2] > 127 ? 127 : b.s16[2] < -128 ? 128 : b.s16[2];
+  t.u8[1] = b.s16[1] > 127 ? 127 : b.s16[1] < -128 ? 128 : b.s16[1];
+  t.u8[0] = b.s16[0] > 127 ? 127 : b.s16[0] < -128 ? 128 : b.s16[0];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_unpacklo_u16_s32(c_v64 a) {
+  c_v64 t;
+  t.s32[1] = a.u16[1 + !!CONFIG_BIG_ENDIAN * 2];
+  t.s32[0] = a.u16[0 + !!CONFIG_BIG_ENDIAN * 2];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_unpacklo_s16_s32(c_v64 a) {
+  c_v64 t;
+  t.s32[1] = a.s16[1 + !!CONFIG_BIG_ENDIAN * 2];
+  t.s32[0] = a.s16[0 + !!CONFIG_BIG_ENDIAN * 2];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_unpackhi_u16_s32(c_v64 a) {
+  c_v64 t;
+  t.s32[1] = a.u16[3 - !!CONFIG_BIG_ENDIAN * 2];
+  t.s32[0] = a.u16[2 - !!CONFIG_BIG_ENDIAN * 2];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_unpackhi_s16_s32(c_v64 a) {
+  c_v64 t;
+  t.s32[1] = a.s16[3 - !!CONFIG_BIG_ENDIAN * 2];
+  t.s32[0] = a.s16[2 - !!CONFIG_BIG_ENDIAN * 2];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shuffle_8(c_v64 a, c_v64 pattern) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) {
+    if (SIMD_CHECK && (pattern.u8[c] & ~7)) {
+      fprintf(stderr, "Error: Undefined v64_shuffle_8 index %d/%d\n",
+              pattern.u8[c], c);
+      abort();
+    }
+    t.u8[c] =
+        a.u8[CONFIG_BIG_ENDIAN ? 7 - (pattern.u8[c] & 7) : pattern.u8[c] & 7];
+  }
+  return t;
+}
+
+SIMD_INLINE int64_t c_v64_dotp_su8(c_v64 a, c_v64 b) {
+  return a.s8[7] * b.u8[7] + a.s8[6] * b.u8[6] + a.s8[5] * b.u8[5] +
+         a.s8[4] * b.u8[4] + a.s8[3] * b.u8[3] + a.s8[2] * b.u8[2] +
+         a.s8[1] * b.u8[1] + a.s8[0] * b.u8[0];
+}
+
+SIMD_INLINE int64_t c_v64_dotp_s16(c_v64 a, c_v64 b) {
+  return (int64_t)(a.s16[3] * b.s16[3] + a.s16[2] * b.s16[2]) +
+         (int64_t)(a.s16[1] * b.s16[1] + a.s16[0] * b.s16[0]);
+}
+
+SIMD_INLINE uint64_t c_v64_hadd_u8(c_v64 a) {
+  return a.u8[7] + a.u8[6] + a.u8[5] + a.u8[4] + a.u8[3] + a.u8[2] + a.u8[1] +
+         a.u8[0];
+}
+
+SIMD_INLINE int64_t c_v64_hadd_s16(c_v64 a) {
+  return a.s16[3] + a.s16[2] + a.s16[1] + a.s16[0];
+}
+
+typedef uint32_t c_sad64_internal;
+
+/* Implementation dependent return value.  Result must be finalised with
+   v64_sad_u8_sum().
+   The result for more than 32 v64_sad_u8() calls is undefined. */
+SIMD_INLINE c_sad64_internal c_v64_sad_u8_init() { return 0; }
+
+SIMD_INLINE c_sad64_internal c_v64_sad_u8(c_sad64_internal s, c_v64 a,
+                                          c_v64 b) {
+  int c;
+  for (c = 0; c < 8; c++)
+    s += a.u8[c] > b.u8[c] ? a.u8[c] - b.u8[c] : b.u8[c] - a.u8[c];
+  return s;
+}
+
+SIMD_INLINE uint32_t c_v64_sad_u8_sum(c_sad64_internal s) { return s; }
+
+typedef uint32_t c_ssd64_internal;
+
+/* Implementation dependent return value.  Result must be finalised with
+ * v64_ssd_u8_sum(). */
+SIMD_INLINE c_ssd64_internal c_v64_ssd_u8_init() { return 0; }
+
+SIMD_INLINE c_ssd64_internal c_v64_ssd_u8(c_ssd64_internal s, c_v64 a,
+                                          c_v64 b) {
+  int c;
+  for (c = 0; c < 8; c++) s += (a.u8[c] - b.u8[c]) * (a.u8[c] - b.u8[c]);
+  return s;
+}
+
+SIMD_INLINE uint32_t c_v64_ssd_u8_sum(c_ssd64_internal s) { return s; }
+
+SIMD_INLINE c_v64 c_v64_or(c_v64 a, c_v64 b) {
+  c_v64 t;
+  t.u64 = a.u64 | b.u64;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_xor(c_v64 a, c_v64 b) {
+  c_v64 t;
+  t.u64 = a.u64 ^ b.u64;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_and(c_v64 a, c_v64 b) {
+  c_v64 t;
+  t.u64 = a.u64 & b.u64;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_andn(c_v64 a, c_v64 b) {
+  c_v64 t;
+  t.u64 = a.u64 & ~b.u64;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_mullo_s16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++) t.s16[c] = (int16_t)(a.s16[c] * b.s16[c]);
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_mulhi_s16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++) t.s16[c] = (a.s16[c] * b.s16[c]) >> 16;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_mullo_s32(c_v64 a, c_v64 b) {
+  c_v64 t;
+  t.s32[0] = (int32_t)((int64_t)a.s32[0] * b.s32[0]);
+  t.s32[1] = (int32_t)((int64_t)a.s32[1] * b.s32[1]);
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_madd_s16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  t.s32[0] = a.s16[0] * b.s16[0] + a.s16[1] * b.s16[1];
+  t.s32[1] = a.s16[2] * b.s16[2] + a.s16[3] * b.s16[3];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_madd_us8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int32_t u;
+  u = a.u8[0] * b.s8[0] + a.u8[1] * b.s8[1];
+  t.s16[0] = u > 32767 ? 32767 : u < -32768 ? -32768 : u;
+  u = a.u8[2] * b.s8[2] + a.u8[3] * b.s8[3];
+  t.s16[1] = u > 32767 ? 32767 : u < -32768 ? -32768 : u;
+  u = a.u8[4] * b.s8[4] + a.u8[5] * b.s8[5];
+  t.s16[2] = u > 32767 ? 32767 : u < -32768 ? -32768 : u;
+  u = a.u8[6] * b.s8[6] + a.u8[7] * b.s8[7];
+  t.s16[3] = u > 32767 ? 32767 : u < -32768 ? -32768 : u;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_avg_u8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.u8[c] = (a.u8[c] + b.u8[c] + 1) >> 1;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_rdavg_u8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.u8[c] = (a.u8[c] + b.u8[c]) >> 1;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_avg_u16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++) t.u16[c] = (a.u16[c] + b.u16[c] + 1) >> 1;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_min_u8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] > b.u8[c] ? b.u8[c] : a.u8[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_max_u8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] > b.u8[c] ? a.u8[c] : b.u8[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_min_s8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.s8[c] = a.s8[c] > b.s8[c] ? b.s8[c] : a.s8[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_max_s8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.s8[c] = a.s8[c] > b.s8[c] ? a.s8[c] : b.s8[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_min_s16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++) t.s16[c] = a.s16[c] > b.s16[c] ? b.s16[c] : a.s16[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_max_s16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++) t.s16[c] = a.s16[c] > b.s16[c] ? a.s16[c] : b.s16[c];
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_cmpgt_s8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.s8[c] = -(a.s8[c] > b.s8[c]);
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_cmplt_s8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.s8[c] = -(a.s8[c] < b.s8[c]);
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_cmpeq_8(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 8; c++) t.s8[c] = -(a.u8[c] == b.u8[c]);
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_cmpgt_s16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++) t.s16[c] = -(a.s16[c] > b.s16[c]);
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_cmplt_s16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++) t.s16[c] = -(a.s16[c] < b.s16[c]);
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_cmpeq_16(c_v64 a, c_v64 b) {
+  c_v64 t;
+  int c;
+  for (c = 0; c < 4; c++) t.s16[c] = -(a.u16[c] == b.u16[c]);
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shl_8(c_v64 a, unsigned int n) {
+  c_v64 t;
+  int c;
+  if (SIMD_CHECK && n > 7) {
+    fprintf(stderr, "Error: Undefined u8 shift left %d\n", n);
+    abort();
+  }
+  for (c = 0; c < 8; c++) t.s8[c] = a.u8[c] << n;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shr_u8(c_v64 a, unsigned int n) {
+  c_v64 t;
+  int c;
+  if (SIMD_CHECK && n > 7) {
+    fprintf(stderr, "Error: Undefined u8 shift right %d\n", n);
+    abort();
+  }
+  for (c = 0; c < 8; c++) t.u8[c] = a.u8[c] >> n;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shr_s8(c_v64 a, unsigned int n) {
+  c_v64 t;
+  int c;
+  if (SIMD_CHECK && n > 7) {
+    fprintf(stderr, "Error: Undefined s8 shift right %d\n", n);
+    abort();
+  }
+  for (c = 0; c < 8; c++) t.s8[c] = a.s8[c] >> n;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shl_16(c_v64 a, unsigned int n) {
+  c_v64 t;
+  int c;
+  if (SIMD_CHECK && n > 15) {
+    fprintf(stderr, "Error: Undefined u16 shift left %d\n", n);
+    abort();
+  }
+  for (c = 0; c < 4; c++) t.u16[c] = a.u16[c] << n;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shr_u16(c_v64 a, unsigned int n) {
+  c_v64 t;
+  int c;
+  if (SIMD_CHECK && n > 15) {
+    fprintf(stderr, "Error: Undefined u16 shift right %d\n", n);
+    abort();
+  }
+  for (c = 0; c < 4; c++) t.u16[c] = a.u16[c] >> n;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shr_s16(c_v64 a, unsigned int n) {
+  c_v64 t;
+  int c;
+  if (SIMD_CHECK && n > 15) {
+    fprintf(stderr, "Error: undefined s16 shift right %d\n", n);
+    abort();
+  }
+  for (c = 0; c < 4; c++) t.s16[c] = a.s16[c] >> n;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shl_32(c_v64 a, unsigned int n) {
+  c_v64 t;
+  if (SIMD_CHECK && n > 31) {
+    fprintf(stderr, "Error: undefined u32 shift left %d\n", n);
+    abort();
+  }
+  t.u32[1] = a.u32[1] << n;
+  t.u32[0] = a.u32[0] << n;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shr_u32(c_v64 a, unsigned int n) {
+  c_v64 t;
+  if (SIMD_CHECK && n > 31) {
+    fprintf(stderr, "Error: undefined u32 shift right %d\n", n);
+    abort();
+  }
+  t.u32[1] = a.u32[1] >> n;
+  t.u32[0] = a.u32[0] >> n;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shr_s32(c_v64 a, unsigned int n) {
+  c_v64 t;
+  if (SIMD_CHECK && n > 31) {
+    fprintf(stderr, "Error: undefined s32 shift right %d\n", n);
+    abort();
+  }
+  t.s32[1] = a.s32[1] >> n;
+  t.s32[0] = a.s32[0] >> n;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shr_n_byte(c_v64 x, unsigned int i) {
+  c_v64 t;
+  t.u64 = x.u64 >> i * 8;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_shl_n_byte(c_v64 x, unsigned int i) {
+  c_v64 t;
+  t.u64 = x.u64 << i * 8;
+  return t;
+}
+
+SIMD_INLINE c_v64 c_v64_align(c_v64 a, c_v64 b, unsigned int c) {
+  if (SIMD_CHECK && c > 7) {
+    fprintf(stderr, "Error: undefined alignment %d\n", c);
+    abort();
+  }
+  return c ? c_v64_or(c_v64_shr_n_byte(b, c), c_v64_shl_n_byte(a, 8 - c)) : b;
+}
+
+SIMD_INLINE c_v64 c_v64_shl_n_8(c_v64 a, unsigned int c) {
+  return c_v64_shl_8(a, c);
+}
+
+SIMD_INLINE c_v64 c_v64_shr_n_u8(c_v64 a, unsigned int c) {
+  return c_v64_shr_u8(a, c);
+}
+
+SIMD_INLINE c_v64 c_v64_shr_n_s8(c_v64 a, unsigned int c) {
+  return c_v64_shr_s8(a, c);
+}
+
+SIMD_INLINE c_v64 c_v64_shl_n_16(c_v64 a, unsigned int c) {
+  return c_v64_shl_16(a, c);
+}
+
+SIMD_INLINE c_v64 c_v64_shr_n_u16(c_v64 a, unsigned int c) {
+  return c_v64_shr_u16(a, c);
+}
+
+SIMD_INLINE c_v64 c_v64_shr_n_s16(c_v64 a, unsigned int c) {
+  return c_v64_shr_s16(a, c);
+}
+
+SIMD_INLINE c_v64 c_v64_shl_n_32(c_v64 a, unsigned int c) {
+  return c_v64_shl_32(a, c);
+}
+
+SIMD_INLINE c_v64 c_v64_shr_n_u32(c_v64 a, unsigned int c) {
+  return c_v64_shr_u32(a, c);
+}
+
+SIMD_INLINE c_v64 c_v64_shr_n_s32(c_v64 a, unsigned int c) {
+  return c_v64_shr_s32(a, c);
+}
+
+#endif /* _V64_INTRINSICS_C_H */
diff --git a/third_party/aom/aom_dsp/simd/v64_intrinsics_x86.h b/third_party/aom/aom_dsp/simd/v64_intrinsics_x86.h
new file mode 100644
index 0000000000..8dcc9f6fc7
--- /dev/null
+++ b/third_party/aom/aom_dsp/simd/v64_intrinsics_x86.h
@@ -0,0 +1,470 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _V64_INTRINSICS_H
+#define _V64_INTRINSICS_H
+
+#include <emmintrin.h>
+#if defined(__SSSE3__)
+#include <tmmintrin.h>
+#endif
+#if defined(__SSE4_1__)
+#include <smmintrin.h>
+#endif
+
+typedef __m128i v64;
+
+SIMD_INLINE uint32_t v64_low_u32(v64 a) {
+  return (uint32_t)_mm_cvtsi128_si32(a);
+}
+
+SIMD_INLINE uint32_t v64_high_u32(v64 a) {
+  return (uint32_t)_mm_cvtsi128_si32(_mm_srli_si128(a, 4));
+}
+
+SIMD_INLINE int32_t v64_low_s32(v64 a) { return (int32_t)_mm_cvtsi128_si32(a); }
+
+SIMD_INLINE int32_t v64_high_s32(v64 a) {
+  return (int32_t)_mm_cvtsi128_si32(_mm_srli_si128(a, 4));
+}
+
+SIMD_INLINE v64 v64_from_16(uint16_t a, uint16_t b, uint16_t c, uint16_t d) {
+  return _mm_packs_epi32(
+      _mm_set_epi32((int16_t)a, (int16_t)b, (int16_t)c, (int16_t)d),
+      _mm_setzero_si128());
+}
+
+SIMD_INLINE v64 v64_from_32(uint32_t x, uint32_t y) {
+  return _mm_set_epi32(0, 0, x, y);
+}
+
+SIMD_INLINE v64 v64_from_64(uint64_t x) {
+#ifdef __x86_64__
+  return _mm_cvtsi64_si128(x);
+#else
+  return _mm_set_epi32(0, 0, x >> 32, (uint32_t)x);
+#endif
+}
+
+SIMD_INLINE uint64_t v64_u64(v64 x) {
+  return (uint64_t)v64_low_u32(x) | ((uint64_t)v64_high_u32(x) << 32);
+}
+
+SIMD_INLINE uint32_t u32_load_aligned(const void *p) {
+  return *((uint32_t *)p);
+}
+
+SIMD_INLINE uint32_t u32_load_unaligned(const void *p) {
+  return *((uint32_t *)p);
+}
+
+SIMD_INLINE void u32_store_aligned(void *p, uint32_t a) {
+  *((uint32_t *)p) = a;
+}
+
+SIMD_INLINE void u32_store_unaligned(void *p, uint32_t a) {
+  *((uint32_t *)p) = a;
+}
+
+SIMD_INLINE v64 v64_load_aligned(const void *p) {
+  return _mm_loadl_epi64((__m128i *)p);
+}
+
+SIMD_INLINE v64 v64_load_unaligned(const void *p) {
+  return _mm_loadl_epi64((__m128i *)p);
+}
+
+SIMD_INLINE void v64_store_aligned(void *p, v64 a) {
+  _mm_storel_epi64((__m128i *)p, a);
+}
+
+SIMD_INLINE void v64_store_unaligned(void *p, v64 a) {
+  _mm_storel_epi64((__m128i *)p, a);
+}
+
+// The following function requires an immediate.
+#if defined(__OPTIMIZE__) && __OPTIMIZE__
+#define v64_align(a, b, c) \
+  ((c) ? _mm_srli_si128(_mm_unpacklo_epi64(b, a), (c)) : b)
+#else
+#define v64_align(a, b, c)                                                  \
+  ((c) ? v64_from_64((v64_u64(b) >> (c)*8) | (v64_u64(a) << (8 - (c)) * 8)) \
+       : (b))
+#endif
+
+SIMD_INLINE v64 v64_zero() { return _mm_setzero_si128(); }
+
+SIMD_INLINE v64 v64_dup_8(uint8_t x) { return _mm_set1_epi8(x); }
+
+SIMD_INLINE v64 v64_dup_16(uint16_t x) { return _mm_set1_epi16(x); }
+
+SIMD_INLINE v64 v64_dup_32(uint32_t x) { return _mm_set1_epi32(x); }
+
+SIMD_INLINE v64 v64_add_8(v64 a, v64 b) { return _mm_add_epi8(a, b); }
+
+SIMD_INLINE v64 v64_add_16(v64 a, v64 b) { return _mm_add_epi16(a, b); }
+
+SIMD_INLINE v64 v64_sadd_s16(v64 a, v64 b) { return _mm_adds_epi16(a, b); }
+
+SIMD_INLINE v64 v64_add_32(v64 a, v64 b) { return _mm_add_epi32(a, b); }
+
+SIMD_INLINE v64 v64_sub_8(v64 a, v64 b) { return _mm_sub_epi8(a, b); }
+
+SIMD_INLINE v64 v64_ssub_u8(v64 a, v64 b) { return _mm_subs_epu8(a, b); }
+
+SIMD_INLINE v64 v64_ssub_s8(v64 a, v64 b) { return _mm_subs_epi8(a, b); }
+
+SIMD_INLINE v64 v64_sub_16(v64 a, v64 b) { return _mm_sub_epi16(a, b); }
+
+SIMD_INLINE v64 v64_ssub_s16(v64 a, v64 b) { return _mm_subs_epi16(a, b); }
+
+SIMD_INLINE v64 v64_ssub_u16(v64 a, v64 b) { return _mm_subs_epu16(a, b); }
+
+SIMD_INLINE v64 v64_sub_32(v64 a, v64 b) { return _mm_sub_epi32(a, b); }
+
+SIMD_INLINE v64 v64_abs_s16(v64 a) {
+#if defined(__SSSE3__)
+  return _mm_abs_epi16(a);
+#else
+  return _mm_max_epi16(a, _mm_sub_epi16(_mm_setzero_si128(), a));
+#endif
+}
+
+SIMD_INLINE v64 v64_abs_s8(v64 a) {
+#if defined(__SSSE3__)
+  return _mm_abs_epi8(a);
+#else
+  v64 sign = _mm_cmplt_epi8(a, _mm_setzero_si128());
+  return _mm_xor_si128(sign, _mm_add_epi8(a, sign));
+#endif
+}
+
+SIMD_INLINE v64 v64_ziplo_8(v64 a, v64 b) { return _mm_unpacklo_epi8(b, a); }
+
+SIMD_INLINE v64 v64_ziphi_8(v64 a, v64 b) {
+  return _mm_srli_si128(_mm_unpacklo_epi8(b, a), 8);
+}
+
+SIMD_INLINE v64 v64_ziplo_16(v64 a, v64 b) { return _mm_unpacklo_epi16(b, a); }
+
+SIMD_INLINE v64 v64_ziphi_16(v64 a, v64 b) {
+  return _mm_srli_si128(_mm_unpacklo_epi16(b, a), 8);
+}
+
+SIMD_INLINE v64 v64_ziplo_32(v64 a, v64 b) { return _mm_unpacklo_epi32(b, a); }
+
+SIMD_INLINE v64 v64_ziphi_32(v64 a, v64 b) {
+  return _mm_srli_si128(_mm_unpacklo_epi32(b, a), 8);
+}
+
+SIMD_INLINE v64 v64_pack_s32_s16(v64 a, v64 b) {
+  __m128i t = _mm_unpacklo_epi64(b, a);
+  return _mm_packs_epi32(t, t);
+}
+
+SIMD_INLINE v64 v64_pack_s16_u8(v64 a, v64 b) {
+  __m128i t = _mm_unpacklo_epi64(b, a);
+  return _mm_packus_epi16(t, t);
+}
+
+SIMD_INLINE v64 v64_pack_s16_s8(v64 a, v64 b) {
+  __m128i t = _mm_unpacklo_epi64(b, a);
+  return _mm_packs_epi16(t, t);
+}
+
+SIMD_INLINE v64 v64_unziphi_8(v64 a, v64 b) {
+#if defined(__SSSE3__)
+  return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a),
+                          v64_from_64(0x0f0d0b0907050301LL));
+#else
+  return _mm_packus_epi16(
+      _mm_unpacklo_epi64(_mm_srli_epi16(b, 8), _mm_srli_epi16(a, 8)),
+      _mm_setzero_si128());
+#endif
+}
+
+SIMD_INLINE v64 v64_unziplo_8(v64 a, v64 b) {
+#if defined(__SSSE3__)
+  return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a),
+                          v64_from_64(0x0e0c0a0806040200LL));
+#else
+  return v64_unziphi_8(_mm_slli_si128(a, 1), _mm_slli_si128(b, 1));
+#endif
+}
+
+SIMD_INLINE v64 v64_unziphi_16(v64 a, v64 b) {
+#if defined(__SSSE3__)
+  return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a),
+                          v64_from_64(0x0f0e0b0a07060302LL));
+#else
+  return _mm_packs_epi32(
+      _mm_unpacklo_epi64(_mm_srai_epi32(b, 16), _mm_srai_epi32(a, 16)),
+      _mm_setzero_si128());
+#endif
+}
+
+SIMD_INLINE v64 v64_unziplo_16(v64 a, v64 b) {
+#if defined(__SSSE3__)
+  return _mm_shuffle_epi8(_mm_unpacklo_epi64(b, a),
+                          v64_from_64(0x0d0c090805040100LL));
+#else
+  return v64_unziphi_16(_mm_slli_si128(a, 2), _mm_slli_si128(b, 2));
+#endif
+}
+
+SIMD_INLINE v64 v64_unpacklo_u8_s16(v64 a) {
+  return _mm_unpacklo_epi8(a, _mm_setzero_si128());
+}
+
+SIMD_INLINE v64 v64_unpackhi_u8_s16(v64 a) {
+  return _mm_srli_si128(_mm_unpacklo_epi8(a, _mm_setzero_si128()), 8);
+}
+
+SIMD_INLINE v64 v64_unpacklo_s8_s16(v64 a) {
+  return _mm_srai_epi16(_mm_unpacklo_epi8(a, a), 8);
+}
+
+SIMD_INLINE v64 v64_unpackhi_s8_s16(v64 a) {
+  return _mm_srli_si128(_mm_srai_epi16(_mm_unpacklo_epi8(a, a), 8), 8);
+}
+
+SIMD_INLINE v64 v64_unpacklo_u16_s32(v64 a) {
+  return _mm_unpacklo_epi16(a, _mm_setzero_si128());
+}
+
+SIMD_INLINE v64 v64_unpacklo_s16_s32(v64 a) {
+  return _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), a), 16);
+}
+
+SIMD_INLINE v64 v64_unpackhi_u16_s32(v64 a) {
+  return _mm_srli_si128(_mm_unpacklo_epi16(a, _mm_setzero_si128()), 8);
+}
+
+SIMD_INLINE v64 v64_unpackhi_s16_s32(v64 a) {
+  return _mm_srli_si128(
+      _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), a), 16), 8);
+}
+
+SIMD_INLINE v64 v64_shuffle_8(v64 x, v64 pattern) {
+#if defined(__SSSE3__)
+  return _mm_shuffle_epi8(x, pattern);
+#else
+  v64 output;
+  unsigned char *input = (unsigned char *)&x;
+  unsigned char *index = (unsigned char *)&pattern;
+  char *selected = (char *)&output;
+  int counter;
+
+  for (counter = 0; counter < 8; counter++) {
+    selected[counter] = input[index[counter]];
+  }
+
+  return output;
+#endif
+}
+
+SIMD_INLINE int64_t v64_dotp_su8(v64 a, v64 b) {
+  __m128i r, r1, r2, z;
+  z = _mm_setzero_si128();
+  r1 = _mm_madd_epi16(_mm_slli_epi16(_mm_unpacklo_epi8(a, z), 8),
+                      _mm_unpacklo_epi8(b, z));
+  r2 = _mm_srli_si128(r1, 8);
+  r = _mm_add_epi32(r1, r2);
+  r = _mm_add_epi32(r, _mm_srli_si128(r, 4));
+  return ((int32_t)v64_low_u32(r)) >> 8;
+}
+
+SIMD_INLINE int64_t v64_dotp_s16(v64 a, v64 b) {
+  __m128i r = _mm_madd_epi16(a, b);
+#if defined(__SSE4_1__) && defined(__x86_64__)
+  __m128i x = _mm_cvtepi32_epi64(r);
+  return _mm_cvtsi128_si64(_mm_add_epi64(x, _mm_srli_si128(x, 8)));
+#else
+  return (int64_t)_mm_cvtsi128_si32(_mm_srli_si128(r, 4)) +
+         (int64_t)_mm_cvtsi128_si32(r);
+#endif
+}
+
+SIMD_INLINE uint64_t v64_hadd_u8(v64 a) {
+  return v64_low_u32(_mm_sad_epu8(a, _mm_setzero_si128()));
+}
+
+SIMD_INLINE int64_t v64_hadd_s16(v64 a) {
+  return v64_dotp_s16(a, v64_dup_16(1));
+}
+
+typedef v64 sad64_internal;
+
+SIMD_INLINE sad64_internal v64_sad_u8_init() { return _mm_setzero_si128(); }
+
+/* Implementation dependent return value.  Result must be finalised with
+   v64_sad_u8_sum().
+   The result for more than 32 v64_sad_u8() calls is undefined. */
+SIMD_INLINE sad64_internal v64_sad_u8(sad64_internal s, v64 a, v64 b) {
+  return _mm_add_epi64(s, _mm_sad_epu8(a, b));
+}
+
+SIMD_INLINE uint32_t v64_sad_u8_sum(sad64_internal s) { return v64_low_u32(s); }
+
+typedef v64 ssd64_internal;
+
+SIMD_INLINE ssd64_internal v64_ssd_u8_init() { return _mm_setzero_si128(); }
+
+/* Implementation dependent return value.  Result must be finalised with
+ * v64_ssd_u8_sum(). */
+SIMD_INLINE ssd64_internal v64_ssd_u8(ssd64_internal s, v64 a, v64 b) {
+  v64 l = v64_sub_16(v64_ziplo_8(v64_zero(), a), v64_ziplo_8(v64_zero(), b));
+  v64 h = v64_sub_16(v64_ziphi_8(v64_zero(), a), v64_ziphi_8(v64_zero(), b));
+  v64 r = v64_add_32(_mm_madd_epi16(l, l), _mm_madd_epi16(h, h));
+  return _mm_add_epi64(
+      s, v64_ziplo_32(v64_zero(), _mm_add_epi32(r, _mm_srli_si128(r, 4))));
+}
+
+SIMD_INLINE uint32_t v64_ssd_u8_sum(sad64_internal s) { return v64_low_u32(s); }
+
+SIMD_INLINE v64 v64_or(v64 a, v64 b) { return _mm_or_si128(a, b); }
+
+SIMD_INLINE v64 v64_xor(v64 a, v64 b) { return _mm_xor_si128(a, b); }
+
+SIMD_INLINE v64 v64_and(v64 a, v64 b) { return _mm_and_si128(a, b); }
+
+SIMD_INLINE v64 v64_andn(v64 a, v64 b) { return _mm_andnot_si128(b, a); }
+
+SIMD_INLINE v64 v64_mullo_s16(v64 a, v64 b) { return _mm_mullo_epi16(a, b); }
+
+SIMD_INLINE v64 v64_mulhi_s16(v64 a, v64 b) { return _mm_mulhi_epi16(a, b); }
+
+SIMD_INLINE v64 v64_mullo_s32(v64 a, v64 b) {
+#if defined(__SSE4_1__)
+  return _mm_mullo_epi32(a, b);
+#else
+  return _mm_unpacklo_epi32(
+      _mm_mul_epu32(a, b),
+      _mm_mul_epu32(_mm_srli_si128(a, 4), _mm_srli_si128(b, 4)));
+#endif
+}
+
+SIMD_INLINE v64 v64_madd_s16(v64 a, v64 b) { return _mm_madd_epi16(a, b); }
+
+SIMD_INLINE v64 v64_madd_us8(v64 a, v64 b) {
+#if defined(__SSSE3__)
+  return _mm_maddubs_epi16(a, b);
+#else
+  __m128i t = _mm_madd_epi16(_mm_unpacklo_epi8(a, _mm_setzero_si128()),
+                             _mm_srai_epi16(_mm_unpacklo_epi8(b, b), 8));
+  return _mm_packs_epi32(t, t);
+#endif
+}
+
+SIMD_INLINE v64 v64_avg_u8(v64 a, v64 b) { return _mm_avg_epu8(a, b); }
+
+SIMD_INLINE v64 v64_rdavg_u8(v64 a, v64 b) {
+  return _mm_sub_epi8(_mm_avg_epu8(a, b),
+                      _mm_and_si128(_mm_xor_si128(a, b), v64_dup_8(1)));
+}
+
+SIMD_INLINE v64 v64_avg_u16(v64 a, v64 b) { return _mm_avg_epu16(a, b); }
+
+SIMD_INLINE v64 v64_min_u8(v64 a, v64 b) { return _mm_min_epu8(a, b); }
+
+SIMD_INLINE v64 v64_max_u8(v64 a, v64 b) { return _mm_max_epu8(a, b); }
+
+SIMD_INLINE v64 v64_min_s8(v64 a, v64 b) {
+#if defined(__SSE4_1__)
+  return _mm_min_epi8(a, b);
+#else
+  v64 mask = _mm_cmplt_epi8(a, b);
+  return _mm_or_si128(_mm_andnot_si128(mask, b), _mm_and_si128(mask, a));
+#endif
+}
+
+SIMD_INLINE v64 v64_max_s8(v64 a, v64 b) {
+#if defined(__SSE4_1__)
+  return _mm_max_epi8(a, b);
+#else
+  v64 mask = _mm_cmplt_epi8(b, a);
+  return _mm_or_si128(_mm_andnot_si128(mask, b), _mm_and_si128(mask, a));
+#endif
+}
+
+SIMD_INLINE v64 v64_min_s16(v64 a, v64 b) { return _mm_min_epi16(a, b); }
+
+SIMD_INLINE v64 v64_max_s16(v64 a, v64 b) { return _mm_max_epi16(a, b); }
+
+SIMD_INLINE v64 v64_cmpgt_s8(v64 a, v64 b) { return _mm_cmpgt_epi8(a, b); }
+
+SIMD_INLINE v64 v64_cmplt_s8(v64 a, v64 b) { return _mm_cmplt_epi8(a, b); }
+
+SIMD_INLINE v64 v64_cmpeq_8(v64 a, v64 b) { return _mm_cmpeq_epi8(a, b); }
+
+SIMD_INLINE v64 v64_cmpgt_s16(v64 a, v64 b) { return _mm_cmpgt_epi16(a, b); }
+
+SIMD_INLINE v64 v64_cmplt_s16(v64 a, v64 b) { return _mm_cmplt_epi16(a, b); }
+
+SIMD_INLINE v64 v64_cmpeq_16(v64 a, v64 b) { return _mm_cmpeq_epi16(a, b); }
+
+SIMD_INLINE v64 v64_shl_8(v64 a, unsigned int c) {
+  return _mm_and_si128(_mm_set1_epi8((uint8_t)(0xff << c)),
+                       _mm_sll_epi16(a, _mm_cvtsi32_si128(c)));
+}
+
+SIMD_INLINE v64 v64_shr_u8(v64 a, unsigned int c) {
+  return _mm_and_si128(_mm_set1_epi8(0xff >> c),
+                       _mm_srl_epi16(a, _mm_cvtsi32_si128(c)));
+}
+
+SIMD_INLINE v64 v64_shr_s8(v64 a, unsigned int c) {
+  return _mm_packs_epi16(
+      _mm_sra_epi16(_mm_unpacklo_epi8(a, a), _mm_cvtsi32_si128(c + 8)), a);
+}
+
+SIMD_INLINE v64 v64_shl_16(v64 a, unsigned int c) {
+  return _mm_sll_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v64 v64_shr_u16(v64 a, unsigned int c) {
+  return _mm_srl_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v64 v64_shr_s16(v64 a, unsigned int c) {
+  return _mm_sra_epi16(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v64 v64_shl_32(v64 a, unsigned int c) {
+  return _mm_sll_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v64 v64_shr_u32(v64 a, unsigned int c) {
+  return _mm_srl_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+SIMD_INLINE v64 v64_shr_s32(v64 a, unsigned int c) {
+  return _mm_sra_epi32(a, _mm_cvtsi32_si128(c));
+}
+
+/* These intrinsics require immediate values, so we must use #defines
+   to enforce that. */
+#define v64_shl_n_byte(a, c) _mm_slli_si128(a, c)
+#define v64_shr_n_byte(a, c) _mm_srli_si128(_mm_unpacklo_epi64(a, a), c + 8)
+#define v64_shl_n_8(a, c) \
+  _mm_and_si128(_mm_set1_epi8((uint8_t)(0xff << (c))), _mm_slli_epi16(a, c))
+#define v64_shr_n_u8(a, c) \
+  _mm_and_si128(_mm_set1_epi8(0xff >> (c)), _mm_srli_epi16(a, c))
+#define v64_shr_n_s8(a, c) \
+  _mm_packs_epi16(_mm_srai_epi16(_mm_unpacklo_epi8(a, a), (c) + 8), a)
+#define v64_shl_n_16(a, c) _mm_slli_epi16(a, c)
+#define v64_shr_n_u16(a, c) _mm_srli_epi16(a, c)
+#define v64_shr_n_s16(a, c) _mm_srai_epi16(a, c)
+#define v64_shl_n_32(a, c) _mm_slli_epi32(a, c)
+#define v64_shr_n_u32(a, c) _mm_srli_epi32(a, c)
+#define v64_shr_n_s32(a, c) _mm_srai_epi32(a, c)
+
+#endif /* _V64_INTRINSICS_H */
diff --git a/third_party/aom/aom_dsp/ssim.c b/third_party/aom/aom_dsp/ssim.c
new file mode 100644
index 0000000000..141bf01c74
--- /dev/null
+++ b/third_party/aom/aom_dsp/ssim.c
@@ -0,0 +1,462 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/ssim.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/system_state.h"
+
+void aom_ssim_parms_16x16_c(const uint8_t *s, int sp, const uint8_t *r, int rp,
+                            uint32_t *sum_s, uint32_t *sum_r,
+                            uint32_t *sum_sq_s, uint32_t *sum_sq_r,
+                            uint32_t *sum_sxr) {
+  int i, j;
+  for (i = 0; i < 16; i++, s += sp, r += rp) {
+    for (j = 0; j < 16; j++) {
+      *sum_s += s[j];
+      *sum_r += r[j];
+      *sum_sq_s += s[j] * s[j];
+      *sum_sq_r += r[j] * r[j];
+      *sum_sxr += s[j] * r[j];
+    }
+  }
+}
+void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp,
+                          uint32_t *sum_s, uint32_t *sum_r, uint32_t *sum_sq_s,
+                          uint32_t *sum_sq_r, uint32_t *sum_sxr) {
+  int i, j;
+  for (i = 0; i < 8; i++, s += sp, r += rp) {
+    for (j = 0; j < 8; j++) {
+      *sum_s += s[j];
+      *sum_r += r[j];
+      *sum_sq_s += s[j] * s[j];
+      *sum_sq_r += r[j] * r[j];
+      *sum_sxr += s[j] * r[j];
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_ssim_parms_8x8_c(const uint16_t *s, int sp, const uint16_t *r,
+                                 int rp, uint32_t *sum_s, uint32_t *sum_r,
+                                 uint32_t *sum_sq_s, uint32_t *sum_sq_r,
+                                 uint32_t *sum_sxr) {
+  int i, j;
+  for (i = 0; i < 8; i++, s += sp, r += rp) {
+    for (j = 0; j < 8; j++) {
+      *sum_s += s[j];
+      *sum_r += r[j];
+      *sum_sq_s += s[j] * s[j];
+      *sum_sq_r += r[j] * r[j];
+      *sum_sxr += s[j] * r[j];
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static const int64_t cc1 = 26634;        // (64^2*(.01*255)^2
+static const int64_t cc2 = 239708;       // (64^2*(.03*255)^2
+static const int64_t cc1_10 = 428658;    // (64^2*(.01*1023)^2
+static const int64_t cc2_10 = 3857925;   // (64^2*(.03*1023)^2
+static const int64_t cc1_12 = 6868593;   // (64^2*(.01*4095)^2
+static const int64_t cc2_12 = 61817334;  // (64^2*(.03*4095)^2
+
+static double similarity(uint32_t sum_s, uint32_t sum_r, uint32_t sum_sq_s,
+                         uint32_t sum_sq_r, uint32_t sum_sxr, int count,
+                         uint32_t bd) {
+  int64_t ssim_n, ssim_d;
+  int64_t c1, c2;
+  if (bd == 8) {
+    // scale the constants by number of pixels
+    c1 = (cc1 * count * count) >> 12;
+    c2 = (cc2 * count * count) >> 12;
+  } else if (bd == 10) {
+    c1 = (cc1_10 * count * count) >> 12;
+    c2 = (cc2_10 * count * count) >> 12;
+  } else if (bd == 12) {
+    c1 = (cc1_12 * count * count) >> 12;
+    c2 = (cc2_12 * count * count) >> 12;
+  } else {
+    c1 = c2 = 0;
+    assert(0);
+  }
+
+  ssim_n = (2 * sum_s * sum_r + c1) *
+           ((int64_t)2 * count * sum_sxr - (int64_t)2 * sum_s * sum_r + c2);
+
+  ssim_d = (sum_s * sum_s + sum_r * sum_r + c1) *
+           ((int64_t)count * sum_sq_s - (int64_t)sum_s * sum_s +
+            (int64_t)count * sum_sq_r - (int64_t)sum_r * sum_r + c2);
+
+  return ssim_n * 1.0 / ssim_d;
+}
+
+static double ssim_8x8(const uint8_t *s, int sp, const uint8_t *r, int rp) {
+  uint32_t sum_s = 0, sum_r = 0, sum_sq_s = 0, sum_sq_r = 0, sum_sxr = 0;
+  aom_ssim_parms_8x8(s, sp, r, rp, &sum_s, &sum_r, &sum_sq_s, &sum_sq_r,
+                     &sum_sxr);
+  return similarity(sum_s, sum_r, sum_sq_s, sum_sq_r, sum_sxr, 64, 8);
+}
+
+#if CONFIG_HIGHBITDEPTH
+static double highbd_ssim_8x8(const uint16_t *s, int sp, const uint16_t *r,
+                              int rp, uint32_t bd, uint32_t shift) {
+  uint32_t sum_s = 0, sum_r = 0, sum_sq_s = 0, sum_sq_r = 0, sum_sxr = 0;
+  aom_highbd_ssim_parms_8x8(s, sp, r, rp, &sum_s, &sum_r, &sum_sq_s, &sum_sq_r,
+                            &sum_sxr);
+  return similarity(sum_s >> shift, sum_r >> shift, sum_sq_s >> (2 * shift),
+                    sum_sq_r >> (2 * shift), sum_sxr >> (2 * shift), 64, bd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+// We are using a 8x8 moving window with starting location of each 8x8 window
+// on the 4x4 pixel grid. Such arrangement allows the windows to overlap
+// block boundaries to penalize blocking artifacts.
+static double aom_ssim2(const uint8_t *img1, const uint8_t *img2,
+                        int stride_img1, int stride_img2, int width,
+                        int height) {
+  int i, j;
+  int samples = 0;
+  double ssim_total = 0;
+
+  // sample point start with each 4x4 location
+  for (i = 0; i <= height - 8;
+       i += 4, img1 += stride_img1 * 4, img2 += stride_img2 * 4) {
+    for (j = 0; j <= width - 8; j += 4) {
+      double v = ssim_8x8(img1 + j, stride_img1, img2 + j, stride_img2);
+      ssim_total += v;
+      samples++;
+    }
+  }
+  ssim_total /= samples;
+  return ssim_total;
+}
+
+#if CONFIG_HIGHBITDEPTH
+static double aom_highbd_ssim2(const uint8_t *img1, const uint8_t *img2,
+                               int stride_img1, int stride_img2, int width,
+                               int height, uint32_t bd, uint32_t shift) {
+  int i, j;
+  int samples = 0;
+  double ssim_total = 0;
+
+  // sample point start with each 4x4 location
+  for (i = 0; i <= height - 8;
+       i += 4, img1 += stride_img1 * 4, img2 += stride_img2 * 4) {
+    for (j = 0; j <= width - 8; j += 4) {
+      double v = highbd_ssim_8x8(CONVERT_TO_SHORTPTR(img1 + j), stride_img1,
+                                 CONVERT_TO_SHORTPTR(img2 + j), stride_img2, bd,
+                                 shift);
+      ssim_total += v;
+      samples++;
+    }
+  }
+  ssim_total /= samples;
+  return ssim_total;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+double aom_calc_ssim(const YV12_BUFFER_CONFIG *source,
+                     const YV12_BUFFER_CONFIG *dest, double *weight) {
+  double a, b, c;
+  double ssimv;
+
+  a = aom_ssim2(source->y_buffer, dest->y_buffer, source->y_stride,
+                dest->y_stride, source->y_crop_width, source->y_crop_height);
+
+  b = aom_ssim2(source->u_buffer, dest->u_buffer, source->uv_stride,
+                dest->uv_stride, source->uv_crop_width, source->uv_crop_height);
+
+  c = aom_ssim2(source->v_buffer, dest->v_buffer, source->uv_stride,
+                dest->uv_stride, source->uv_crop_width, source->uv_crop_height);
+
+  ssimv = a * .8 + .1 * (b + c);
+
+  *weight = 1;
+
+  return ssimv;
+}
+
+// traditional ssim as per: http://en.wikipedia.org/wiki/Structural_similarity
+//
+// Re working out the math ->
+//
+// ssim(x,y) =  (2*mean(x)*mean(y) + c1)*(2*cov(x,y)+c2) /
+//   ((mean(x)^2+mean(y)^2+c1)*(var(x)+var(y)+c2))
+//
+// mean(x) = sum(x) / n
+//
+// cov(x,y) = (n*sum(xi*yi)-sum(x)*sum(y))/(n*n)
+//
+// var(x) = (n*sum(xi*xi)-sum(xi)*sum(xi))/(n*n)
+//
+// ssim(x,y) =
+//   (2*sum(x)*sum(y)/(n*n) + c1)*(2*(n*sum(xi*yi)-sum(x)*sum(y))/(n*n)+c2) /
+//   (((sum(x)*sum(x)+sum(y)*sum(y))/(n*n) +c1) *
+//    ((n*sum(xi*xi) - sum(xi)*sum(xi))/(n*n)+
+//     (n*sum(yi*yi) - sum(yi)*sum(yi))/(n*n)+c2)))
+//
+// factoring out n*n
+//
+// ssim(x,y) =
+//   (2*sum(x)*sum(y) + n*n*c1)*(2*(n*sum(xi*yi)-sum(x)*sum(y))+n*n*c2) /
+//   (((sum(x)*sum(x)+sum(y)*sum(y)) + n*n*c1) *
+//    (n*sum(xi*xi)-sum(xi)*sum(xi)+n*sum(yi*yi)-sum(yi)*sum(yi)+n*n*c2))
+//
+// Replace c1 with n*n * c1 for the final step that leads to this code:
+// The final step scales by 12 bits so we don't lose precision in the constants.
+
+static double ssimv_similarity(const Ssimv *sv, int64_t n) {
+  // Scale the constants by number of pixels.
+  const int64_t c1 = (cc1 * n * n) >> 12;
+  const int64_t c2 = (cc2 * n * n) >> 12;
+
+  const double l = 1.0 * (2 * sv->sum_s * sv->sum_r + c1) /
+                   (sv->sum_s * sv->sum_s + sv->sum_r * sv->sum_r + c1);
+
+  // Since these variables are unsigned sums, convert to double so
+  // math is done in double arithmetic.
+  const double v = (2.0 * n * sv->sum_sxr - 2 * sv->sum_s * sv->sum_r + c2) /
+                   (n * sv->sum_sq_s - sv->sum_s * sv->sum_s +
+                    n * sv->sum_sq_r - sv->sum_r * sv->sum_r + c2);
+
+  return l * v;
+}
+
+// The first term of the ssim metric is a luminance factor.
+//
+// (2*mean(x)*mean(y) + c1)/ (mean(x)^2+mean(y)^2+c1)
+//
+// This luminance factor is super sensitive to the dark side of luminance
+// values and completely insensitive on the white side.  check out 2 sets
+// (1,3) and (250,252) the term gives ( 2*1*3/(1+9) = .60
+// 2*250*252/ (250^2+252^2) => .99999997
+//
+// As a result in this tweaked version of the calculation in which the
+// luminance is taken as percentage off from peak possible.
+//
+// 255 * 255 - (sum_s - sum_r) / count * (sum_s - sum_r) / count
+//
+static double ssimv_similarity2(const Ssimv *sv, int64_t n) {
+  // Scale the constants by number of pixels.
+  const int64_t c1 = (cc1 * n * n) >> 12;
+  const int64_t c2 = (cc2 * n * n) >> 12;
+
+  const double mean_diff = (1.0 * sv->sum_s - sv->sum_r) / n;
+  const double l = (255 * 255 - mean_diff * mean_diff + c1) / (255 * 255 + c1);
+
+  // Since these variables are unsigned, sums convert to double so
+  // math is done in double arithmetic.
+  const double v = (2.0 * n * sv->sum_sxr - 2 * sv->sum_s * sv->sum_r + c2) /
+                   (n * sv->sum_sq_s - sv->sum_s * sv->sum_s +
+                    n * sv->sum_sq_r - sv->sum_r * sv->sum_r + c2);
+
+  return l * v;
+}
+static void ssimv_parms(uint8_t *img1, int img1_pitch, uint8_t *img2,
+                        int img2_pitch, Ssimv *sv) {
+  aom_ssim_parms_8x8(img1, img1_pitch, img2, img2_pitch, &sv->sum_s, &sv->sum_r,
+                     &sv->sum_sq_s, &sv->sum_sq_r, &sv->sum_sxr);
+}
+
+double aom_get_ssim_metrics(uint8_t *img1, int img1_pitch, uint8_t *img2,
+                            int img2_pitch, int width, int height, Ssimv *sv2,
+                            Metrics *m, int do_inconsistency) {
+  double dssim_total = 0;
+  double ssim_total = 0;
+  double ssim2_total = 0;
+  double inconsistency_total = 0;
+  int i, j;
+  int c = 0;
+  double norm;
+  double old_ssim_total = 0;
+  aom_clear_system_state();
+  // We can sample points as frequently as we like start with 1 per 4x4.
+  for (i = 0; i < height;
+       i += 4, img1 += img1_pitch * 4, img2 += img2_pitch * 4) {
+    for (j = 0; j < width; j += 4, ++c) {
+      Ssimv sv = { 0 };
+      double ssim;
+      double ssim2;
+      double dssim;
+      uint32_t var_new;
+      uint32_t var_old;
+      uint32_t mean_new;
+      uint32_t mean_old;
+      double ssim_new;
+      double ssim_old;
+
+      // Not sure there's a great way to handle the edge pixels
+      // in ssim when using a window. Seems biased against edge pixels
+      // however you handle this. This uses only samples that are
+      // fully in the frame.
+      if (j + 8 <= width && i + 8 <= height) {
+        ssimv_parms(img1 + j, img1_pitch, img2 + j, img2_pitch, &sv);
+      }
+
+      ssim = ssimv_similarity(&sv, 64);
+      ssim2 = ssimv_similarity2(&sv, 64);
+
+      sv.ssim = ssim2;
+
+      // dssim is calculated to use as an actual error metric and
+      // is scaled up to the same range as sum square error.
+      // Since we are subsampling every 16th point maybe this should be
+      // *16 ?
+      dssim = 255 * 255 * (1 - ssim2) / 2;
+
+      // Here I introduce a new error metric: consistency-weighted
+      // SSIM-inconsistency.  This metric isolates frames where the
+      // SSIM 'suddenly' changes, e.g. if one frame in every 8 is much
+      // sharper or blurrier than the others. Higher values indicate a
+      // temporally inconsistent SSIM. There are two ideas at work:
+      //
+      // 1) 'SSIM-inconsistency': the total inconsistency value
+      // reflects how much SSIM values are changing between this
+      // source / reference frame pair and the previous pair.
+      //
+      // 2) 'consistency-weighted': weights de-emphasize areas in the
+      // frame where the scene content has changed. Changes in scene
+      // content are detected via changes in local variance and local
+      // mean.
+      //
+      // Thus the overall measure reflects how inconsistent the SSIM
+      // values are, over consistent regions of the frame.
+      //
+      // The metric has three terms:
+      //
+      // term 1 -> uses change in scene Variance to weight error score
+      //  2 * var(Fi)*var(Fi-1) / (var(Fi)^2+var(Fi-1)^2)
+      //  larger changes from one frame to the next mean we care
+      //  less about consistency.
+      //
+      // term 2 -> uses change in local scene luminance to weight error
+      //  2 * avg(Fi)*avg(Fi-1) / (avg(Fi)^2+avg(Fi-1)^2)
+      //  larger changes from one frame to the next mean we care
+      //  less about consistency.
+      //
+      // term3 -> measures inconsistency in ssim scores between frames
+      //   1 - ( 2 * ssim(Fi)*ssim(Fi-1)/(ssim(Fi)^2+sssim(Fi-1)^2).
+      //
+      // This term compares the ssim score for the same location in 2
+      // subsequent frames.
+      var_new = sv.sum_sq_s - sv.sum_s * sv.sum_s / 64;
+      var_old = sv2[c].sum_sq_s - sv2[c].sum_s * sv2[c].sum_s / 64;
+      mean_new = sv.sum_s;
+      mean_old = sv2[c].sum_s;
+      ssim_new = sv.ssim;
+      ssim_old = sv2[c].ssim;
+
+      if (do_inconsistency) {
+        // We do the metric once for every 4x4 block in the image. Since
+        // we are scaling the error to SSE for use in a psnr calculation
+        // 1.0 = 4x4x255x255 the worst error we can possibly have.
+        static const double kScaling = 4. * 4 * 255 * 255;
+
+        // The constants have to be non 0 to avoid potential divide by 0
+        // issues other than that they affect kind of a weighting between
+        // the terms.  No testing of what the right terms should be has been
+        // done.
+        static const double c1 = 1, c2 = 1, c3 = 1;
+
+        // This measures how much consistent variance is in two consecutive
+        // source frames. 1.0 means they have exactly the same variance.
+        const double variance_term =
+            (2.0 * var_old * var_new + c1) /
+            (1.0 * var_old * var_old + 1.0 * var_new * var_new + c1);
+
+        // This measures how consistent the local mean are between two
+        // consecutive frames. 1.0 means they have exactly the same mean.
+        const double mean_term =
+            (2.0 * mean_old * mean_new + c2) /
+            (1.0 * mean_old * mean_old + 1.0 * mean_new * mean_new + c2);
+
+        // This measures how consistent the ssims of two
+        // consecutive frames is. 1.0 means they are exactly the same.
+        double ssim_term =
+            pow((2.0 * ssim_old * ssim_new + c3) /
+                    (ssim_old * ssim_old + ssim_new * ssim_new + c3),
+                5);
+
+        double this_inconsistency;
+
+        // Floating point math sometimes makes this > 1 by a tiny bit.
+        // We want the metric to scale between 0 and 1.0 so we can convert
+        // it to an snr scaled value.
+        if (ssim_term > 1) ssim_term = 1;
+
+        // This converts the consistency metric to an inconsistency metric
+        // ( so we can scale it like psnr to something like sum square error.
+        // The reason for the variance and mean terms is the assumption that
+        // if there are big changes in the source we shouldn't penalize
+        // inconsistency in ssim scores a bit less as it will be less visible
+        // to the user.
+        this_inconsistency = (1 - ssim_term) * variance_term * mean_term;
+
+        this_inconsistency *= kScaling;
+        inconsistency_total += this_inconsistency;
+      }
+      sv2[c] = sv;
+      ssim_total += ssim;
+      ssim2_total += ssim2;
+      dssim_total += dssim;
+
+      old_ssim_total += ssim_old;
+    }
+    old_ssim_total += 0;
+  }
+
+  norm = 1. / (width / 4) / (height / 4);
+  ssim_total *= norm;
+  ssim2_total *= norm;
+  m->ssim2 = ssim2_total;
+  m->ssim = ssim_total;
+  if (old_ssim_total == 0) inconsistency_total = 0;
+
+  m->ssimc = inconsistency_total;
+
+  m->dssim = dssim_total;
+  return inconsistency_total;
+}
+
+#if CONFIG_HIGHBITDEPTH
+double aom_highbd_calc_ssim(const YV12_BUFFER_CONFIG *source,
+                            const YV12_BUFFER_CONFIG *dest, double *weight,
+                            uint32_t bd, uint32_t in_bd) {
+  double a, b, c;
+  double ssimv;
+  uint32_t shift = 0;
+
+  assert(bd >= in_bd);
+  shift = bd - in_bd;
+
+  a = aom_highbd_ssim2(source->y_buffer, dest->y_buffer, source->y_stride,
+                       dest->y_stride, source->y_crop_width,
+                       source->y_crop_height, in_bd, shift);
+
+  b = aom_highbd_ssim2(source->u_buffer, dest->u_buffer, source->uv_stride,
+                       dest->uv_stride, source->uv_crop_width,
+                       source->uv_crop_height, in_bd, shift);
+
+  c = aom_highbd_ssim2(source->v_buffer, dest->v_buffer, source->uv_stride,
+                       dest->uv_stride, source->uv_crop_width,
+                       source->uv_crop_height, in_bd, shift);
+
+  ssimv = a * .8 + .1 * (b + c);
+
+  *weight = 1;
+
+  return ssimv;
+}
+
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/ssim.h b/third_party/aom/aom_dsp/ssim.h
new file mode 100644
index 0000000000..902735e50e
--- /dev/null
+++ b/third_party/aom/aom_dsp/ssim.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_SSIM_H_
+#define AOM_DSP_SSIM_H_
+
+#define MAX_SSIM_DB 100.0;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./aom_config.h"
+#include "aom_scale/yv12config.h"
+
+// metrics used for calculating ssim, ssim2, dssim, and ssimc
+typedef struct {
+  // source sum ( over 8x8 region )
+  uint32_t sum_s;
+
+  // reference sum (over 8x8 region )
+  uint32_t sum_r;
+
+  // source sum squared ( over 8x8 region )
+  uint32_t sum_sq_s;
+
+  // reference sum squared (over 8x8 region )
+  uint32_t sum_sq_r;
+
+  // sum of source times reference (over 8x8 region)
+  uint32_t sum_sxr;
+
+  // calculated ssim score between source and reference
+  double ssim;
+} Ssimv;
+
+// metrics collected on a frame basis
+typedef struct {
+  // ssim consistency error metric ( see code for explanation )
+  double ssimc;
+
+  // standard ssim
+  double ssim;
+
+  // revised ssim ( see code for explanation)
+  double ssim2;
+
+  // ssim restated as an error metric like sse
+  double dssim;
+
+  // dssim converted to decibels
+  double dssimd;
+
+  // ssimc converted to decibels
+  double ssimcd;
+} Metrics;
+
+double aom_get_ssim_metrics(uint8_t *img1, int img1_pitch, uint8_t *img2,
+                            int img2_pitch, int width, int height, Ssimv *sv2,
+                            Metrics *m, int do_inconsistency);
+
+double aom_calc_ssim(const YV12_BUFFER_CONFIG *source,
+                     const YV12_BUFFER_CONFIG *dest, double *weight);
+
+double aom_calc_fastssim(const YV12_BUFFER_CONFIG *source,
+                         const YV12_BUFFER_CONFIG *dest, double *ssim_y,
+                         double *ssim_u, double *ssim_v, uint32_t bd,
+                         uint32_t in_bd);
+
+#if CONFIG_HIGHBITDEPTH
+double aom_highbd_calc_ssim(const YV12_BUFFER_CONFIG *source,
+                            const YV12_BUFFER_CONFIG *dest, double *weight,
+                            uint32_t bd, uint32_t in_bd);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_SSIM_H_
diff --git a/third_party/aom/aom_dsp/subtract.c b/third_party/aom/aom_dsp/subtract.c
new file mode 100644
index 0000000000..8dda96efbc
--- /dev/null
+++ b/third_party/aom/aom_dsp/subtract.c
@@ -0,0 +1,55 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+void aom_subtract_block_c(int rows, int cols, int16_t *diff,
+                          ptrdiff_t diff_stride, const uint8_t *src,
+                          ptrdiff_t src_stride, const uint8_t *pred,
+                          ptrdiff_t pred_stride) {
+  int r, c;
+
+  for (r = 0; r < rows; r++) {
+    for (c = 0; c < cols; c++) diff[c] = src[c] - pred[c];
+
+    diff += diff_stride;
+    pred += pred_stride;
+    src += src_stride;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_subtract_block_c(int rows, int cols, int16_t *diff,
+                                 ptrdiff_t diff_stride, const uint8_t *src8,
+                                 ptrdiff_t src_stride, const uint8_t *pred8,
+                                 ptrdiff_t pred_stride, int bd) {
+  int r, c;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  (void)bd;
+
+  for (r = 0; r < rows; r++) {
+    for (c = 0; c < cols; c++) {
+      diff[c] = src[c] - pred[c];
+    }
+
+    diff += diff_stride;
+    pred += pred_stride;
+    src += src_stride;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/sum_squares.c b/third_party/aom/aom_dsp/sum_squares.c
new file mode 100644
index 0000000000..b9155fdc05
--- /dev/null
+++ b/third_party/aom/aom_dsp/sum_squares.c
@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "./aom_dsp_rtcd.h"
+
+uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int src_stride, int width,
+                                  int height) {
+  int r, c;
+  uint64_t ss = 0;
+
+  for (r = 0; r < height; r++) {
+    for (c = 0; c < width; c++) {
+      const int16_t v = src[c];
+      ss += v * v;
+    }
+    src += src_stride;
+  }
+
+  return ss;
+}
+
+uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t n) {
+  uint64_t ss = 0;
+  do {
+    const int16_t v = *src++;
+    ss += v * v;
+  } while (--n);
+
+  return ss;
+}
diff --git a/third_party/aom/aom_dsp/txfm_common.h b/third_party/aom/aom_dsp/txfm_common.h
new file mode 100644
index 0000000000..a5e964aadb
--- /dev/null
+++ b/third_party/aom/aom_dsp/txfm_common.h
@@ -0,0 +1,70 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_TXFM_COMMON_H_
+#define AOM_DSP_TXFM_COMMON_H_
+
+#include "aom_dsp/aom_dsp_common.h"
+
+// Constants and Macros used by all idct/dct functions
+#define DCT_CONST_BITS 14
+#define DCT_CONST_ROUNDING (1 << (DCT_CONST_BITS - 1))
+
+#define UNIT_QUANT_SHIFT 2
+#define UNIT_QUANT_FACTOR (1 << UNIT_QUANT_SHIFT)
+
+// Constants:
+//  for (int i = 1; i< 32; ++i)
+//    printf("static const int cospi_%d_64 = %.0f;\n", i,
+//           round(16384 * cos(i*M_PI/64)));
+// Note: sin(k*Pi/64) = cos((32-k)*Pi/64)
+static const tran_high_t cospi_1_64 = 16364;
+static const tran_high_t cospi_2_64 = 16305;
+static const tran_high_t cospi_3_64 = 16207;
+static const tran_high_t cospi_4_64 = 16069;
+static const tran_high_t cospi_5_64 = 15893;
+static const tran_high_t cospi_6_64 = 15679;
+static const tran_high_t cospi_7_64 = 15426;
+static const tran_high_t cospi_8_64 = 15137;
+static const tran_high_t cospi_9_64 = 14811;
+static const tran_high_t cospi_10_64 = 14449;
+static const tran_high_t cospi_11_64 = 14053;
+static const tran_high_t cospi_12_64 = 13623;
+static const tran_high_t cospi_13_64 = 13160;
+static const tran_high_t cospi_14_64 = 12665;
+static const tran_high_t cospi_15_64 = 12140;
+static const tran_high_t cospi_16_64 = 11585;
+static const tran_high_t cospi_17_64 = 11003;
+static const tran_high_t cospi_18_64 = 10394;
+static const tran_high_t cospi_19_64 = 9760;
+static const tran_high_t cospi_20_64 = 9102;
+static const tran_high_t cospi_21_64 = 8423;
+static const tran_high_t cospi_22_64 = 7723;
+static const tran_high_t cospi_23_64 = 7005;
+static const tran_high_t cospi_24_64 = 6270;
+static const tran_high_t cospi_25_64 = 5520;
+static const tran_high_t cospi_26_64 = 4756;
+static const tran_high_t cospi_27_64 = 3981;
+static const tran_high_t cospi_28_64 = 3196;
+static const tran_high_t cospi_29_64 = 2404;
+static const tran_high_t cospi_30_64 = 1606;
+static const tran_high_t cospi_31_64 = 804;
+
+//  16384 * sqrt(2) * sin(kPi/9) * 2 / 3
+static const tran_high_t sinpi_1_9 = 5283;
+static const tran_high_t sinpi_2_9 = 9929;
+static const tran_high_t sinpi_3_9 = 13377;
+static const tran_high_t sinpi_4_9 = 15212;
+
+// 16384 * sqrt(2)
+static const tran_high_t Sqrt2 = 23170;
+
+#endif  // AOM_DSP_TXFM_COMMON_H_
diff --git a/third_party/aom/aom_dsp/variance.c b/third_party/aom/aom_dsp/variance.c
new file mode 100644
index 0000000000..9fc0db783f
--- /dev/null
+++ b/third_party/aom/aom_dsp/variance.c
@@ -0,0 +1,1249 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <stdlib.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "aom/aom_integer.h"
+
+#include "aom_dsp/variance.h"
+#include "aom_dsp/aom_filter.h"
+
+uint32_t aom_get4x4sse_cs_c(const uint8_t *a, int a_stride, const uint8_t *b,
+                            int b_stride) {
+  int distortion = 0;
+  int r, c;
+
+  for (r = 0; r < 4; ++r) {
+    for (c = 0; c < 4; ++c) {
+      int diff = a[c] - b[c];
+      distortion += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+
+  return distortion;
+}
+
+uint32_t aom_get_mb_ss_c(const int16_t *a) {
+  unsigned int i, sum = 0;
+
+  for (i = 0; i < 256; ++i) {
+    sum += a[i] * a[i];
+  }
+
+  return sum;
+}
+
+uint32_t aom_variance_halfpixvar16x16_h_c(const uint8_t *a, int a_stride,
+                                          const uint8_t *b, int b_stride,
+                                          uint32_t *sse) {
+  return aom_sub_pixel_variance16x16_c(a, a_stride, 4, 0, b, b_stride, sse);
+}
+
+uint32_t aom_variance_halfpixvar16x16_v_c(const uint8_t *a, int a_stride,
+                                          const uint8_t *b, int b_stride,
+                                          uint32_t *sse) {
+  return aom_sub_pixel_variance16x16_c(a, a_stride, 0, 4, b, b_stride, sse);
+}
+
+uint32_t aom_variance_halfpixvar16x16_hv_c(const uint8_t *a, int a_stride,
+                                           const uint8_t *b, int b_stride,
+                                           uint32_t *sse) {
+  return aom_sub_pixel_variance16x16_c(a, a_stride, 4, 4, b, b_stride, sse);
+}
+
+static void variance(const uint8_t *a, int a_stride, const uint8_t *b,
+                     int b_stride, int w, int h, uint32_t *sse, int *sum) {
+  int i, j;
+
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+uint32_t aom_sse_odd_size(const uint8_t *a, int a_stride, const uint8_t *b,
+                          int b_stride, int w, int h) {
+  uint32_t sse;
+  int sum;
+  variance(a, a_stride, b, b_stride, w, h, &sse, &sum);
+  return sse;
+}
+
+// Applies a 1-D 2-tap bilinear filter to the source block in either horizontal
+// or vertical direction to produce the filtered output block. Used to implement
+// the first-pass of 2-D separable filter.
+//
+// Produces int16_t output to retain precision for the next pass. Two filter
+// taps should sum to FILTER_WEIGHT. pixel_step defines whether the filter is
+// applied horizontally (pixel_step = 1) or vertically (pixel_step = stride).
+// It defines the offset required to move from one input to the next.
+static void var_filter_block2d_bil_first_pass(const uint8_t *a, uint16_t *b,
+                                              unsigned int src_pixels_per_line,
+                                              int pixel_step,
+                                              unsigned int output_height,
+                                              unsigned int output_width,
+                                              const uint8_t *filter) {
+  unsigned int i, j;
+
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      b[j] = ROUND_POWER_OF_TWO(
+          (int)a[0] * filter[0] + (int)a[pixel_step] * filter[1], FILTER_BITS);
+
+      ++a;
+    }
+
+    a += src_pixels_per_line - output_width;
+    b += output_width;
+  }
+}
+
+// Applies a 1-D 2-tap bilinear filter to the source block in either horizontal
+// or vertical direction to produce the filtered output block. Used to implement
+// the second-pass of 2-D separable filter.
+//
+// Requires 16-bit input as produced by filter_block2d_bil_first_pass. Two
+// filter taps should sum to FILTER_WEIGHT. pixel_step defines whether the
+// filter is applied horizontally (pixel_step = 1) or vertically
+// (pixel_step = stride). It defines the offset required to move from one input
+// to the next. Output is 8-bit.
+static void var_filter_block2d_bil_second_pass(const uint16_t *a, uint8_t *b,
+                                               unsigned int src_pixels_per_line,
+                                               unsigned int pixel_step,
+                                               unsigned int output_height,
+                                               unsigned int output_width,
+                                               const uint8_t *filter) {
+  unsigned int i, j;
+
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      b[j] = ROUND_POWER_OF_TWO(
+          (int)a[0] * filter[0] + (int)a[pixel_step] * filter[1], FILTER_BITS);
+      ++a;
+    }
+
+    a += src_pixels_per_line - output_width;
+    b += output_width;
+  }
+}
+
+#define VAR(W, H)                                                    \
+  uint32_t aom_variance##W##x##H##_c(const uint8_t *a, int a_stride, \
+                                     const uint8_t *b, int b_stride, \
+                                     uint32_t *sse) {                \
+    int sum;                                                         \
+    variance(a, a_stride, b, b_stride, W, H, sse, &sum);             \
+    return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H));        \
+  }
+
+#define SUBPIX_VAR(W, H)                                                \
+  uint32_t aom_sub_pixel_variance##W##x##H##_c(                         \
+      const uint8_t *a, int a_stride, int xoffset, int yoffset,         \
+      const uint8_t *b, int b_stride, uint32_t *sse) {                  \
+    uint16_t fdata3[(H + 1) * W];                                       \
+    uint8_t temp2[H * W];                                               \
+                                                                        \
+    var_filter_block2d_bil_first_pass(a, fdata3, a_stride, 1, H + 1, W, \
+                                      bilinear_filters_2t[xoffset]);    \
+    var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W,       \
+                                       bilinear_filters_2t[yoffset]);   \
+                                                                        \
+    return aom_variance##W##x##H##_c(temp2, W, b, b_stride, sse);       \
+  }
+
+#define SUBPIX_AVG_VAR(W, H)                                            \
+  uint32_t aom_sub_pixel_avg_variance##W##x##H##_c(                     \
+      const uint8_t *a, int a_stride, int xoffset, int yoffset,         \
+      const uint8_t *b, int b_stride, uint32_t *sse,                    \
+      const uint8_t *second_pred) {                                     \
+    uint16_t fdata3[(H + 1) * W];                                       \
+    uint8_t temp2[H * W];                                               \
+    DECLARE_ALIGNED(16, uint8_t, temp3[H * W]);                         \
+                                                                        \
+    var_filter_block2d_bil_first_pass(a, fdata3, a_stride, 1, H + 1, W, \
+                                      bilinear_filters_2t[xoffset]);    \
+    var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W,       \
+                                       bilinear_filters_2t[yoffset]);   \
+                                                                        \
+    aom_comp_avg_pred(temp3, second_pred, W, H, temp2, W);              \
+                                                                        \
+    return aom_variance##W##x##H##_c(temp3, W, b, b_stride, sse);       \
+  }
+
+/* Identical to the variance call except it takes an additional parameter, sum,
+ * and returns that value using pass-by-reference instead of returning
+ * sse - sum^2 / w*h
+ */
+#define GET_VAR(W, H)                                                         \
+  void aom_get##W##x##H##var_c(const uint8_t *a, int a_stride,                \
+                               const uint8_t *b, int b_stride, uint32_t *sse, \
+                               int *sum) {                                    \
+    variance(a, a_stride, b, b_stride, W, H, sse, sum);                       \
+  }
+
+/* Identical to the variance call except it does not calculate the
+ * sse - sum^2 / w*h and returns sse in addtion to modifying the passed in
+ * variable.
+ */
+#define MSE(W, H)                                               \
+  uint32_t aom_mse##W##x##H##_c(const uint8_t *a, int a_stride, \
+                                const uint8_t *b, int b_stride, \
+                                uint32_t *sse) {                \
+    int sum;                                                    \
+    variance(a, a_stride, b, b_stride, W, H, sse, &sum);        \
+    return *sse;                                                \
+  }
+
+/* All three forms of the variance are available in the same sizes. */
+#define VARIANCES(W, H) \
+  VAR(W, H)             \
+  SUBPIX_VAR(W, H)      \
+  SUBPIX_AVG_VAR(W, H)
+
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+VARIANCES(128, 128)
+VARIANCES(128, 64)
+VARIANCES(64, 128)
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+VARIANCES(64, 64)
+VARIANCES(64, 32)
+VARIANCES(32, 64)
+VARIANCES(32, 32)
+VARIANCES(32, 16)
+VARIANCES(16, 32)
+VARIANCES(16, 16)
+VARIANCES(16, 8)
+VARIANCES(8, 16)
+VARIANCES(8, 8)
+VARIANCES(8, 4)
+VARIANCES(4, 8)
+VARIANCES(4, 4)
+VARIANCES(4, 2)
+VARIANCES(2, 4)
+VARIANCES(2, 2)
+
+GET_VAR(16, 16)
+GET_VAR(8, 8)
+
+MSE(16, 16)
+MSE(16, 8)
+MSE(8, 16)
+MSE(8, 8)
+
+void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width,
+                         int height, const uint8_t *ref, int ref_stride) {
+  int i, j;
+
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      const int tmp = pred[j] + ref[j];
+      comp_pred[j] = ROUND_POWER_OF_TWO(tmp, 1);
+    }
+    comp_pred += width;
+    pred += width;
+    ref += ref_stride;
+  }
+}
+
+// Get pred block from up-sampled reference.
+void aom_upsampled_pred_c(uint8_t *comp_pred, int width, int height,
+                          const uint8_t *ref, int ref_stride) {
+  int i, j, k;
+  int stride = ref_stride << 3;
+
+  for (i = 0; i < height; i++) {
+    for (j = 0, k = 0; j < width; j++, k += 8) {
+      comp_pred[j] = ref[k];
+    }
+    comp_pred += width;
+    ref += stride;
+  }
+}
+
+void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred,
+                                   int width, int height, const uint8_t *ref,
+                                   int ref_stride) {
+  int i, j;
+  int stride = ref_stride << 3;
+
+  for (i = 0; i < height; i++) {
+    for (j = 0; j < width; j++) {
+      const int tmp = ref[(j << 3)] + pred[j];
+      comp_pred[j] = ROUND_POWER_OF_TWO(tmp, 1);
+    }
+    comp_pred += width;
+    pred += width;
+    ref += stride;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_variance64(const uint8_t *a8, int a_stride,
+                              const uint8_t *b8, int b_stride, int w, int h,
+                              uint64_t *sse, int64_t *sum) {
+  int i, j;
+
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+uint64_t aom_highbd_sse_odd_size(const uint8_t *a, int a_stride,
+                                 const uint8_t *b, int b_stride, int w, int h) {
+  uint64_t sse;
+  int64_t sum;
+  highbd_variance64(a, a_stride, b, b_stride, w, h, &sse, &sum);
+  return sse;
+}
+
+static void highbd_8_variance(const uint8_t *a8, int a_stride,
+                              const uint8_t *b8, int b_stride, int w, int h,
+                              uint32_t *sse, int *sum) {
+  uint64_t sse_long = 0;
+  int64_t sum_long = 0;
+  highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
+  *sse = (uint32_t)sse_long;
+  *sum = (int)sum_long;
+}
+
+static void highbd_10_variance(const uint8_t *a8, int a_stride,
+                               const uint8_t *b8, int b_stride, int w, int h,
+                               uint32_t *sse, int *sum) {
+  uint64_t sse_long = 0;
+  int64_t sum_long = 0;
+  highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4);
+  *sum = (int)ROUND_POWER_OF_TWO(sum_long, 2);
+}
+
+static void highbd_12_variance(const uint8_t *a8, int a_stride,
+                               const uint8_t *b8, int b_stride, int w, int h,
+                               uint32_t *sse, int *sum) {
+  uint64_t sse_long = 0;
+  int64_t sum_long = 0;
+  highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8);
+  *sum = (int)ROUND_POWER_OF_TWO(sum_long, 4);
+}
+
+#define HIGHBD_VAR(W, H)                                                       \
+  uint32_t aom_highbd_8_variance##W##x##H##_c(const uint8_t *a, int a_stride,  \
+                                              const uint8_t *b, int b_stride,  \
+                                              uint32_t *sse) {                 \
+    int sum;                                                                   \
+    highbd_8_variance(a, a_stride, b, b_stride, W, H, sse, &sum);              \
+    return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H));                  \
+  }                                                                            \
+                                                                               \
+  uint32_t aom_highbd_10_variance##W##x##H##_c(const uint8_t *a, int a_stride, \
+                                               const uint8_t *b, int b_stride, \
+                                               uint32_t *sse) {                \
+    int sum;                                                                   \
+    int64_t var;                                                               \
+    highbd_10_variance(a, a_stride, b, b_stride, W, H, sse, &sum);             \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));                  \
+    return (var >= 0) ? (uint32_t)var : 0;                                     \
+  }                                                                            \
+                                                                               \
+  uint32_t aom_highbd_12_variance##W##x##H##_c(const uint8_t *a, int a_stride, \
+                                               const uint8_t *b, int b_stride, \
+                                               uint32_t *sse) {                \
+    int sum;                                                                   \
+    int64_t var;                                                               \
+    highbd_12_variance(a, a_stride, b, b_stride, W, H, sse, &sum);             \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));                  \
+    return (var >= 0) ? (uint32_t)var : 0;                                     \
+  }
+
+#define HIGHBD_GET_VAR(S)                                                    \
+  void aom_highbd_8_get##S##x##S##var_c(const uint8_t *src, int src_stride,  \
+                                        const uint8_t *ref, int ref_stride,  \
+                                        uint32_t *sse, int *sum) {           \
+    highbd_8_variance(src, src_stride, ref, ref_stride, S, S, sse, sum);     \
+  }                                                                          \
+                                                                             \
+  void aom_highbd_10_get##S##x##S##var_c(const uint8_t *src, int src_stride, \
+                                         const uint8_t *ref, int ref_stride, \
+                                         uint32_t *sse, int *sum) {          \
+    highbd_10_variance(src, src_stride, ref, ref_stride, S, S, sse, sum);    \
+  }                                                                          \
+                                                                             \
+  void aom_highbd_12_get##S##x##S##var_c(const uint8_t *src, int src_stride, \
+                                         const uint8_t *ref, int ref_stride, \
+                                         uint32_t *sse, int *sum) {          \
+    highbd_12_variance(src, src_stride, ref, ref_stride, S, S, sse, sum);    \
+  }
+
+#define HIGHBD_MSE(W, H)                                                      \
+  uint32_t aom_highbd_8_mse##W##x##H##_c(const uint8_t *src, int src_stride,  \
+                                         const uint8_t *ref, int ref_stride,  \
+                                         uint32_t *sse) {                     \
+    int sum;                                                                  \
+    highbd_8_variance(src, src_stride, ref, ref_stride, W, H, sse, &sum);     \
+    return *sse;                                                              \
+  }                                                                           \
+                                                                              \
+  uint32_t aom_highbd_10_mse##W##x##H##_c(const uint8_t *src, int src_stride, \
+                                          const uint8_t *ref, int ref_stride, \
+                                          uint32_t *sse) {                    \
+    int sum;                                                                  \
+    highbd_10_variance(src, src_stride, ref, ref_stride, W, H, sse, &sum);    \
+    return *sse;                                                              \
+  }                                                                           \
+                                                                              \
+  uint32_t aom_highbd_12_mse##W##x##H##_c(const uint8_t *src, int src_stride, \
+                                          const uint8_t *ref, int ref_stride, \
+                                          uint32_t *sse) {                    \
+    int sum;                                                                  \
+    highbd_12_variance(src, src_stride, ref, ref_stride, W, H, sse, &sum);    \
+    return *sse;                                                              \
+  }
+
+void aom_highbd_var_filter_block2d_bil_first_pass(
+    const uint8_t *src_ptr8, uint16_t *output_ptr,
+    unsigned int src_pixels_per_line, int pixel_step,
+    unsigned int output_height, unsigned int output_width,
+    const uint8_t *filter) {
+  unsigned int i, j;
+  uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src_ptr8);
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      output_ptr[j] = ROUND_POWER_OF_TWO(
+          (int)src_ptr[0] * filter[0] + (int)src_ptr[pixel_step] * filter[1],
+          FILTER_BITS);
+
+      ++src_ptr;
+    }
+
+    // Next row...
+    src_ptr += src_pixels_per_line - output_width;
+    output_ptr += output_width;
+  }
+}
+
+void aom_highbd_var_filter_block2d_bil_second_pass(
+    const uint16_t *src_ptr, uint16_t *output_ptr,
+    unsigned int src_pixels_per_line, unsigned int pixel_step,
+    unsigned int output_height, unsigned int output_width,
+    const uint8_t *filter) {
+  unsigned int i, j;
+
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      output_ptr[j] = ROUND_POWER_OF_TWO(
+          (int)src_ptr[0] * filter[0] + (int)src_ptr[pixel_step] * filter[1],
+          FILTER_BITS);
+      ++src_ptr;
+    }
+
+    src_ptr += src_pixels_per_line - output_width;
+    output_ptr += output_width;
+  }
+}
+
+#define HIGHBD_SUBPIX_VAR(W, H)                                              \
+  uint32_t aom_highbd_8_sub_pixel_variance##W##x##H##_c(                     \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,          \
+      const uint8_t *dst, int dst_stride, uint32_t *sse) {                   \
+    uint16_t fdata3[(H + 1) * W];                                            \
+    uint16_t temp2[H * W];                                                   \
+                                                                             \
+    aom_highbd_var_filter_block2d_bil_first_pass(                            \
+        src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \
+    aom_highbd_var_filter_block2d_bil_second_pass(                           \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);            \
+                                                                             \
+    return aom_highbd_8_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), W,  \
+                                              dst, dst_stride, sse);         \
+  }                                                                          \
+                                                                             \
+  uint32_t aom_highbd_10_sub_pixel_variance##W##x##H##_c(                    \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,          \
+      const uint8_t *dst, int dst_stride, uint32_t *sse) {                   \
+    uint16_t fdata3[(H + 1) * W];                                            \
+    uint16_t temp2[H * W];                                                   \
+                                                                             \
+    aom_highbd_var_filter_block2d_bil_first_pass(                            \
+        src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \
+    aom_highbd_var_filter_block2d_bil_second_pass(                           \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);            \
+                                                                             \
+    return aom_highbd_10_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), W, \
+                                               dst, dst_stride, sse);        \
+  }                                                                          \
+                                                                             \
+  uint32_t aom_highbd_12_sub_pixel_variance##W##x##H##_c(                    \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,          \
+      const uint8_t *dst, int dst_stride, uint32_t *sse) {                   \
+    uint16_t fdata3[(H + 1) * W];                                            \
+    uint16_t temp2[H * W];                                                   \
+                                                                             \
+    aom_highbd_var_filter_block2d_bil_first_pass(                            \
+        src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \
+    aom_highbd_var_filter_block2d_bil_second_pass(                           \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);            \
+                                                                             \
+    return aom_highbd_12_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), W, \
+                                               dst, dst_stride, sse);        \
+  }
+
+#define HIGHBD_SUBPIX_AVG_VAR(W, H)                                          \
+  uint32_t aom_highbd_8_sub_pixel_avg_variance##W##x##H##_c(                 \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,          \
+      const uint8_t *dst, int dst_stride, uint32_t *sse,                     \
+      const uint8_t *second_pred) {                                          \
+    uint16_t fdata3[(H + 1) * W];                                            \
+    uint16_t temp2[H * W];                                                   \
+    DECLARE_ALIGNED(16, uint16_t, temp3[H * W]);                             \
+                                                                             \
+    aom_highbd_var_filter_block2d_bil_first_pass(                            \
+        src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \
+    aom_highbd_var_filter_block2d_bil_second_pass(                           \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);            \
+                                                                             \
+    aom_highbd_comp_avg_pred_c(temp3, second_pred, W, H,                     \
+                               CONVERT_TO_BYTEPTR(temp2), W);                \
+                                                                             \
+    return aom_highbd_8_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W,  \
+                                              dst, dst_stride, sse);         \
+  }                                                                          \
+                                                                             \
+  uint32_t aom_highbd_10_sub_pixel_avg_variance##W##x##H##_c(                \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,          \
+      const uint8_t *dst, int dst_stride, uint32_t *sse,                     \
+      const uint8_t *second_pred) {                                          \
+    uint16_t fdata3[(H + 1) * W];                                            \
+    uint16_t temp2[H * W];                                                   \
+    DECLARE_ALIGNED(16, uint16_t, temp3[H * W]);                             \
+                                                                             \
+    aom_highbd_var_filter_block2d_bil_first_pass(                            \
+        src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \
+    aom_highbd_var_filter_block2d_bil_second_pass(                           \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);            \
+                                                                             \
+    aom_highbd_comp_avg_pred_c(temp3, second_pred, W, H,                     \
+                               CONVERT_TO_BYTEPTR(temp2), W);                \
+                                                                             \
+    return aom_highbd_10_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \
+                                               dst, dst_stride, sse);        \
+  }                                                                          \
+                                                                             \
+  uint32_t aom_highbd_12_sub_pixel_avg_variance##W##x##H##_c(                \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,          \
+      const uint8_t *dst, int dst_stride, uint32_t *sse,                     \
+      const uint8_t *second_pred) {                                          \
+    uint16_t fdata3[(H + 1) * W];                                            \
+    uint16_t temp2[H * W];                                                   \
+    DECLARE_ALIGNED(16, uint16_t, temp3[H * W]);                             \
+                                                                             \
+    aom_highbd_var_filter_block2d_bil_first_pass(                            \
+        src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \
+    aom_highbd_var_filter_block2d_bil_second_pass(                           \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);            \
+                                                                             \
+    aom_highbd_comp_avg_pred_c(temp3, second_pred, W, H,                     \
+                               CONVERT_TO_BYTEPTR(temp2), W);                \
+                                                                             \
+    return aom_highbd_12_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp3), W, \
+                                               dst, dst_stride, sse);        \
+  }
+
+/* All three forms of the variance are available in the same sizes. */
+#define HIGHBD_VARIANCES(W, H) \
+  HIGHBD_VAR(W, H)             \
+  HIGHBD_SUBPIX_VAR(W, H)      \
+  HIGHBD_SUBPIX_AVG_VAR(W, H)
+
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+HIGHBD_VARIANCES(128, 128)
+HIGHBD_VARIANCES(128, 64)
+HIGHBD_VARIANCES(64, 128)
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+HIGHBD_VARIANCES(64, 64)
+HIGHBD_VARIANCES(64, 32)
+HIGHBD_VARIANCES(32, 64)
+HIGHBD_VARIANCES(32, 32)
+HIGHBD_VARIANCES(32, 16)
+HIGHBD_VARIANCES(16, 32)
+HIGHBD_VARIANCES(16, 16)
+HIGHBD_VARIANCES(16, 8)
+HIGHBD_VARIANCES(8, 16)
+HIGHBD_VARIANCES(8, 8)
+HIGHBD_VARIANCES(8, 4)
+HIGHBD_VARIANCES(4, 8)
+HIGHBD_VARIANCES(4, 4)
+HIGHBD_VARIANCES(4, 2)
+HIGHBD_VARIANCES(2, 4)
+HIGHBD_VARIANCES(2, 2)
+
+HIGHBD_GET_VAR(8)
+HIGHBD_GET_VAR(16)
+
+HIGHBD_MSE(16, 16)
+HIGHBD_MSE(16, 8)
+HIGHBD_MSE(8, 16)
+HIGHBD_MSE(8, 8)
+
+void aom_highbd_comp_avg_pred_c(uint16_t *comp_pred, const uint8_t *pred8,
+                                int width, int height, const uint8_t *ref8,
+                                int ref_stride) {
+  int i, j;
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      const int tmp = pred[j] + ref[j];
+      comp_pred[j] = ROUND_POWER_OF_TWO(tmp, 1);
+    }
+    comp_pred += width;
+    pred += width;
+    ref += ref_stride;
+  }
+}
+
+void aom_highbd_upsampled_pred_c(uint16_t *comp_pred, int width, int height,
+                                 const uint8_t *ref8, int ref_stride) {
+  int i, j;
+  int stride = ref_stride << 3;
+
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      comp_pred[j] = ref[(j << 3)];
+    }
+    comp_pred += width;
+    ref += stride;
+  }
+}
+
+void aom_highbd_comp_avg_upsampled_pred_c(uint16_t *comp_pred,
+                                          const uint8_t *pred8, int width,
+                                          int height, const uint8_t *ref8,
+                                          int ref_stride) {
+  int i, j;
+  int stride = ref_stride << 3;
+
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      const int tmp = pred[j] + ref[(j << 3)];
+      comp_pred[j] = ROUND_POWER_OF_TWO(tmp, 1);
+    }
+    comp_pred += width;
+    pred += width;
+    ref += stride;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_AV1 && CONFIG_EXT_INTER
+void masked_variance(const uint8_t *a, int a_stride, const uint8_t *b,
+                     int b_stride, const uint8_t *m, int m_stride, int w, int h,
+                     unsigned int *sse, int *sum) {
+  int i, j;
+
+  int64_t sum64 = 0;
+  uint64_t sse64 = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = (a[j] - b[j]) * (m[j]);
+      sum64 += diff;
+      sse64 += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+    m += m_stride;
+  }
+  sum64 = (sum64 >= 0) ? sum64 : -sum64;
+  *sum = (int)ROUND_POWER_OF_TWO(sum64, 6);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse64, 12);
+}
+
+#define MASK_VAR(W, H)                                                       \
+  unsigned int aom_masked_variance##W##x##H##_c(                             \
+      const uint8_t *a, int a_stride, const uint8_t *b, int b_stride,        \
+      const uint8_t *m, int m_stride, unsigned int *sse) {                   \
+    int sum;                                                                 \
+    masked_variance(a, a_stride, b, b_stride, m, m_stride, W, H, sse, &sum); \
+    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H));            \
+  }
+
+#define MASK_SUBPIX_VAR(W, H)                                                 \
+  unsigned int aom_masked_sub_pixel_variance##W##x##H##_c(                    \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, \
+      unsigned int *sse) {                                                    \
+    uint16_t fdata3[(H + 1) * W];                                             \
+    uint8_t temp2[H * W];                                                     \
+                                                                              \
+    var_filter_block2d_bil_first_pass(src, fdata3, src_stride, 1, H + 1, W,   \
+                                      bilinear_filters_2t[xoffset]);          \
+    var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W,             \
+                                       bilinear_filters_2t[yoffset]);         \
+                                                                              \
+    return aom_masked_variance##W##x##H##_c(temp2, W, dst, dst_stride, msk,   \
+                                            msk_stride, sse);                 \
+  }
+
+MASK_VAR(4, 4)
+MASK_SUBPIX_VAR(4, 4)
+
+MASK_VAR(4, 8)
+MASK_SUBPIX_VAR(4, 8)
+
+MASK_VAR(8, 4)
+MASK_SUBPIX_VAR(8, 4)
+
+MASK_VAR(8, 8)
+MASK_SUBPIX_VAR(8, 8)
+
+MASK_VAR(8, 16)
+MASK_SUBPIX_VAR(8, 16)
+
+MASK_VAR(16, 8)
+MASK_SUBPIX_VAR(16, 8)
+
+MASK_VAR(16, 16)
+MASK_SUBPIX_VAR(16, 16)
+
+MASK_VAR(16, 32)
+MASK_SUBPIX_VAR(16, 32)
+
+MASK_VAR(32, 16)
+MASK_SUBPIX_VAR(32, 16)
+
+MASK_VAR(32, 32)
+MASK_SUBPIX_VAR(32, 32)
+
+MASK_VAR(32, 64)
+MASK_SUBPIX_VAR(32, 64)
+
+MASK_VAR(64, 32)
+MASK_SUBPIX_VAR(64, 32)
+
+MASK_VAR(64, 64)
+MASK_SUBPIX_VAR(64, 64)
+
+#if CONFIG_EXT_PARTITION
+MASK_VAR(64, 128)
+MASK_SUBPIX_VAR(64, 128)
+
+MASK_VAR(128, 64)
+MASK_SUBPIX_VAR(128, 64)
+
+MASK_VAR(128, 128)
+MASK_SUBPIX_VAR(128, 128)
+#endif  // CONFIG_EXT_PARTITION
+
+#if CONFIG_HIGHBITDEPTH
+void highbd_masked_variance64(const uint8_t *a8, int a_stride,
+                              const uint8_t *b8, int b_stride, const uint8_t *m,
+                              int m_stride, int w, int h, uint64_t *sse,
+                              int64_t *sum) {
+  int i, j;
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = (a[j] - b[j]) * (m[j]);
+      *sum += (int64_t)diff;
+      *sse += (int64_t)diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+    m += m_stride;
+  }
+  *sum = (*sum >= 0) ? *sum : -*sum;
+  *sum = ROUND_POWER_OF_TWO(*sum, 6);
+  *sse = ROUND_POWER_OF_TWO(*sse, 12);
+}
+
+void highbd_masked_variance(const uint8_t *a8, int a_stride, const uint8_t *b8,
+                            int b_stride, const uint8_t *m, int m_stride, int w,
+                            int h, unsigned int *sse, int *sum) {
+  int64_t sum64;
+  uint64_t sse64;
+  highbd_masked_variance64(a8, a_stride, b8, b_stride, m, m_stride, w, h,
+                           &sse64, &sum64);
+  *sum = (int)sum64;
+  *sse = (unsigned int)sse64;
+}
+
+void highbd_10_masked_variance(const uint8_t *a8, int a_stride,
+                               const uint8_t *b8, int b_stride,
+                               const uint8_t *m, int m_stride, int w, int h,
+                               unsigned int *sse, int *sum) {
+  int64_t sum64;
+  uint64_t sse64;
+  highbd_masked_variance64(a8, a_stride, b8, b_stride, m, m_stride, w, h,
+                           &sse64, &sum64);
+  *sum = (int)ROUND_POWER_OF_TWO(sum64, 2);
+  *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 4);
+}
+
+void highbd_12_masked_variance(const uint8_t *a8, int a_stride,
+                               const uint8_t *b8, int b_stride,
+                               const uint8_t *m, int m_stride, int w, int h,
+                               unsigned int *sse, int *sum) {
+  int64_t sum64;
+  uint64_t sse64;
+  highbd_masked_variance64(a8, a_stride, b8, b_stride, m, m_stride, w, h,
+                           &sse64, &sum64);
+  *sum = (int)ROUND_POWER_OF_TWO(sum64, 4);
+  *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 8);
+}
+
+#define HIGHBD_MASK_VAR(W, H)                                                \
+  unsigned int aom_highbd_masked_variance##W##x##H##_c(                      \
+      const uint8_t *a, int a_stride, const uint8_t *b, int b_stride,        \
+      const uint8_t *m, int m_stride, unsigned int *sse) {                   \
+    int sum;                                                                 \
+    highbd_masked_variance(a, a_stride, b, b_stride, m, m_stride, W, H, sse, \
+                           &sum);                                            \
+    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H));            \
+  }                                                                          \
+                                                                             \
+  unsigned int aom_highbd_10_masked_variance##W##x##H##_c(                   \
+      const uint8_t *a, int a_stride, const uint8_t *b, int b_stride,        \
+      const uint8_t *m, int m_stride, unsigned int *sse) {                   \
+    int sum;                                                                 \
+    int64_t var;                                                             \
+    highbd_10_masked_variance(a, a_stride, b, b_stride, m, m_stride, W, H,   \
+                              sse, &sum);                                    \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));                \
+    return (var >= 0) ? (uint32_t)var : 0;                                   \
+  }                                                                          \
+                                                                             \
+  unsigned int aom_highbd_12_masked_variance##W##x##H##_c(                   \
+      const uint8_t *a, int a_stride, const uint8_t *b, int b_stride,        \
+      const uint8_t *m, int m_stride, unsigned int *sse) {                   \
+    int sum;                                                                 \
+    int64_t var;                                                             \
+    highbd_12_masked_variance(a, a_stride, b, b_stride, m, m_stride, W, H,   \
+                              sse, &sum);                                    \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));                \
+    return (var >= 0) ? (uint32_t)var : 0;                                   \
+  }
+
+#define HIGHBD_MASK_SUBPIX_VAR(W, H)                                          \
+  unsigned int aom_highbd_masked_sub_pixel_variance##W##x##H##_c(             \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, \
+      unsigned int *sse) {                                                    \
+    uint16_t fdata3[(H + 1) * W];                                             \
+    uint16_t temp2[H * W];                                                    \
+                                                                              \
+    aom_highbd_var_filter_block2d_bil_first_pass(                             \
+        src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]);  \
+    aom_highbd_var_filter_block2d_bil_second_pass(                            \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);             \
+                                                                              \
+    return aom_highbd_masked_variance##W##x##H##_c(                           \
+        CONVERT_TO_BYTEPTR(temp2), W, dst, dst_stride, msk, msk_stride, sse); \
+  }                                                                           \
+                                                                              \
+  unsigned int aom_highbd_10_masked_sub_pixel_variance##W##x##H##_c(          \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, \
+      unsigned int *sse) {                                                    \
+    uint16_t fdata3[(H + 1) * W];                                             \
+    uint16_t temp2[H * W];                                                    \
+                                                                              \
+    aom_highbd_var_filter_block2d_bil_first_pass(                             \
+        src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]);  \
+    aom_highbd_var_filter_block2d_bil_second_pass(                            \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);             \
+                                                                              \
+    return aom_highbd_10_masked_variance##W##x##H##_c(                        \
+        CONVERT_TO_BYTEPTR(temp2), W, dst, dst_stride, msk, msk_stride, sse); \
+  }                                                                           \
+                                                                              \
+  unsigned int aom_highbd_12_masked_sub_pixel_variance##W##x##H##_c(          \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride, \
+      unsigned int *sse) {                                                    \
+    uint16_t fdata3[(H + 1) * W];                                             \
+    uint16_t temp2[H * W];                                                    \
+                                                                              \
+    aom_highbd_var_filter_block2d_bil_first_pass(                             \
+        src, fdata3, src_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]);  \
+    aom_highbd_var_filter_block2d_bil_second_pass(                            \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);             \
+                                                                              \
+    return aom_highbd_12_masked_variance##W##x##H##_c(                        \
+        CONVERT_TO_BYTEPTR(temp2), W, dst, dst_stride, msk, msk_stride, sse); \
+  }
+
+HIGHBD_MASK_VAR(4, 4)
+HIGHBD_MASK_SUBPIX_VAR(4, 4)
+
+HIGHBD_MASK_VAR(4, 8)
+HIGHBD_MASK_SUBPIX_VAR(4, 8)
+
+HIGHBD_MASK_VAR(8, 4)
+HIGHBD_MASK_SUBPIX_VAR(8, 4)
+
+HIGHBD_MASK_VAR(8, 8)
+HIGHBD_MASK_SUBPIX_VAR(8, 8)
+
+HIGHBD_MASK_VAR(8, 16)
+HIGHBD_MASK_SUBPIX_VAR(8, 16)
+
+HIGHBD_MASK_VAR(16, 8)
+HIGHBD_MASK_SUBPIX_VAR(16, 8)
+
+HIGHBD_MASK_VAR(16, 16)
+HIGHBD_MASK_SUBPIX_VAR(16, 16)
+
+HIGHBD_MASK_VAR(16, 32)
+HIGHBD_MASK_SUBPIX_VAR(16, 32)
+
+HIGHBD_MASK_VAR(32, 16)
+HIGHBD_MASK_SUBPIX_VAR(32, 16)
+
+HIGHBD_MASK_VAR(32, 32)
+HIGHBD_MASK_SUBPIX_VAR(32, 32)
+
+HIGHBD_MASK_VAR(32, 64)
+HIGHBD_MASK_SUBPIX_VAR(32, 64)
+
+HIGHBD_MASK_VAR(64, 32)
+HIGHBD_MASK_SUBPIX_VAR(64, 32)
+
+HIGHBD_MASK_VAR(64, 64)
+HIGHBD_MASK_SUBPIX_VAR(64, 64)
+
+#if CONFIG_EXT_PARTITION
+HIGHBD_MASK_VAR(64, 128)
+HIGHBD_MASK_SUBPIX_VAR(64, 128)
+
+HIGHBD_MASK_VAR(128, 64)
+HIGHBD_MASK_SUBPIX_VAR(128, 64)
+
+HIGHBD_MASK_VAR(128, 128)
+HIGHBD_MASK_SUBPIX_VAR(128, 128)
+#endif  // CONFIG_EXT_PARTITION
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_AV1 && CONFIG_EXT_INTER
+
+#if CONFIG_AV1 && CONFIG_MOTION_VAR
+static INLINE void obmc_variance(const uint8_t *pre, int pre_stride,
+                                 const int32_t *wsrc, const int32_t *mask,
+                                 int w, int h, unsigned int *sse, int *sum) {
+  int i, j;
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      int diff = ROUND_POWER_OF_TWO_SIGNED(wsrc[j] - pre[j] * mask[j], 12);
+      *sum += diff;
+      *sse += diff * diff;
+    }
+
+    pre += pre_stride;
+    wsrc += w;
+    mask += w;
+  }
+}
+
+#define OBMC_VAR(W, H)                                            \
+  unsigned int aom_obmc_variance##W##x##H##_c(                    \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,    \
+      const int32_t *mask, unsigned int *sse) {                   \
+    int sum;                                                      \
+    obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum);  \
+    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H)); \
+  }
+
+#define OBMC_SUBPIX_VAR(W, H)                                               \
+  unsigned int aom_obmc_sub_pixel_variance##W##x##H##_c(                    \
+      const uint8_t *pre, int pre_stride, int xoffset, int yoffset,         \
+      const int32_t *wsrc, const int32_t *mask, unsigned int *sse) {        \
+    uint16_t fdata3[(H + 1) * W];                                           \
+    uint8_t temp2[H * W];                                                   \
+                                                                            \
+    var_filter_block2d_bil_first_pass(pre, fdata3, pre_stride, 1, H + 1, W, \
+                                      bilinear_filters_2t[xoffset]);        \
+    var_filter_block2d_bil_second_pass(fdata3, temp2, W, W, H, W,           \
+                                       bilinear_filters_2t[yoffset]);       \
+                                                                            \
+    return aom_obmc_variance##W##x##H##_c(temp2, W, wsrc, mask, sse);       \
+  }
+
+OBMC_VAR(4, 4)
+OBMC_SUBPIX_VAR(4, 4)
+
+OBMC_VAR(4, 8)
+OBMC_SUBPIX_VAR(4, 8)
+
+OBMC_VAR(8, 4)
+OBMC_SUBPIX_VAR(8, 4)
+
+OBMC_VAR(8, 8)
+OBMC_SUBPIX_VAR(8, 8)
+
+OBMC_VAR(8, 16)
+OBMC_SUBPIX_VAR(8, 16)
+
+OBMC_VAR(16, 8)
+OBMC_SUBPIX_VAR(16, 8)
+
+OBMC_VAR(16, 16)
+OBMC_SUBPIX_VAR(16, 16)
+
+OBMC_VAR(16, 32)
+OBMC_SUBPIX_VAR(16, 32)
+
+OBMC_VAR(32, 16)
+OBMC_SUBPIX_VAR(32, 16)
+
+OBMC_VAR(32, 32)
+OBMC_SUBPIX_VAR(32, 32)
+
+OBMC_VAR(32, 64)
+OBMC_SUBPIX_VAR(32, 64)
+
+OBMC_VAR(64, 32)
+OBMC_SUBPIX_VAR(64, 32)
+
+OBMC_VAR(64, 64)
+OBMC_SUBPIX_VAR(64, 64)
+
+#if CONFIG_EXT_PARTITION
+OBMC_VAR(64, 128)
+OBMC_SUBPIX_VAR(64, 128)
+
+OBMC_VAR(128, 64)
+OBMC_SUBPIX_VAR(128, 64)
+
+OBMC_VAR(128, 128)
+OBMC_SUBPIX_VAR(128, 128)
+#endif  // CONFIG_EXT_PARTITION
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE void highbd_obmc_variance64(const uint8_t *pre8, int pre_stride,
+                                          const int32_t *wsrc,
+                                          const int32_t *mask, int w, int h,
+                                          uint64_t *sse, int64_t *sum) {
+  int i, j;
+  uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      int diff = ROUND_POWER_OF_TWO_SIGNED(wsrc[j] - pre[j] * mask[j], 12);
+      *sum += diff;
+      *sse += diff * diff;
+    }
+
+    pre += pre_stride;
+    wsrc += w;
+    mask += w;
+  }
+}
+
+static INLINE void highbd_obmc_variance(const uint8_t *pre8, int pre_stride,
+                                        const int32_t *wsrc,
+                                        const int32_t *mask, int w, int h,
+                                        unsigned int *sse, int *sum) {
+  int64_t sum64;
+  uint64_t sse64;
+  highbd_obmc_variance64(pre8, pre_stride, wsrc, mask, w, h, &sse64, &sum64);
+  *sum = (int)sum64;
+  *sse = (unsigned int)sse64;
+}
+
+static INLINE void highbd_10_obmc_variance(const uint8_t *pre8, int pre_stride,
+                                           const int32_t *wsrc,
+                                           const int32_t *mask, int w, int h,
+                                           unsigned int *sse, int *sum) {
+  int64_t sum64;
+  uint64_t sse64;
+  highbd_obmc_variance64(pre8, pre_stride, wsrc, mask, w, h, &sse64, &sum64);
+  *sum = (int)ROUND_POWER_OF_TWO(sum64, 2);
+  *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 4);
+}
+
+static INLINE void highbd_12_obmc_variance(const uint8_t *pre8, int pre_stride,
+                                           const int32_t *wsrc,
+                                           const int32_t *mask, int w, int h,
+                                           unsigned int *sse, int *sum) {
+  int64_t sum64;
+  uint64_t sse64;
+  highbd_obmc_variance64(pre8, pre_stride, wsrc, mask, w, h, &sse64, &sum64);
+  *sum = (int)ROUND_POWER_OF_TWO(sum64, 4);
+  *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 8);
+}
+
+#define HIGHBD_OBMC_VAR(W, H)                                              \
+  unsigned int aom_highbd_obmc_variance##W##x##H##_c(                      \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
+      const int32_t *mask, unsigned int *sse) {                            \
+    int sum;                                                               \
+    highbd_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum);    \
+    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H));          \
+  }                                                                        \
+                                                                           \
+  unsigned int aom_highbd_10_obmc_variance##W##x##H##_c(                   \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
+      const int32_t *mask, unsigned int *sse) {                            \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    highbd_10_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));              \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }                                                                        \
+                                                                           \
+  unsigned int aom_highbd_12_obmc_variance##W##x##H##_c(                   \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
+      const int32_t *mask, unsigned int *sse) {                            \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    highbd_12_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));              \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }
+
+#define HIGHBD_OBMC_SUBPIX_VAR(W, H)                                           \
+  unsigned int aom_highbd_obmc_sub_pixel_variance##W##x##H##_c(                \
+      const uint8_t *pre, int pre_stride, int xoffset, int yoffset,            \
+      const int32_t *wsrc, const int32_t *mask, unsigned int *sse) {           \
+    uint16_t fdata3[(H + 1) * W];                                              \
+    uint16_t temp2[H * W];                                                     \
+                                                                               \
+    aom_highbd_var_filter_block2d_bil_first_pass(                              \
+        pre, fdata3, pre_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]);   \
+    aom_highbd_var_filter_block2d_bil_second_pass(                             \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);              \
+                                                                               \
+    return aom_highbd_obmc_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), W, \
+                                                 wsrc, mask, sse);             \
+  }                                                                            \
+                                                                               \
+  unsigned int aom_highbd_10_obmc_sub_pixel_variance##W##x##H##_c(             \
+      const uint8_t *pre, int pre_stride, int xoffset, int yoffset,            \
+      const int32_t *wsrc, const int32_t *mask, unsigned int *sse) {           \
+    uint16_t fdata3[(H + 1) * W];                                              \
+    uint16_t temp2[H * W];                                                     \
+                                                                               \
+    aom_highbd_var_filter_block2d_bil_first_pass(                              \
+        pre, fdata3, pre_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]);   \
+    aom_highbd_var_filter_block2d_bil_second_pass(                             \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);              \
+                                                                               \
+    return aom_highbd_10_obmc_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), \
+                                                    W, wsrc, mask, sse);       \
+  }                                                                            \
+                                                                               \
+  unsigned int aom_highbd_12_obmc_sub_pixel_variance##W##x##H##_c(             \
+      const uint8_t *pre, int pre_stride, int xoffset, int yoffset,            \
+      const int32_t *wsrc, const int32_t *mask, unsigned int *sse) {           \
+    uint16_t fdata3[(H + 1) * W];                                              \
+    uint16_t temp2[H * W];                                                     \
+                                                                               \
+    aom_highbd_var_filter_block2d_bil_first_pass(                              \
+        pre, fdata3, pre_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]);   \
+    aom_highbd_var_filter_block2d_bil_second_pass(                             \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);              \
+                                                                               \
+    return aom_highbd_12_obmc_variance##W##x##H##_c(CONVERT_TO_BYTEPTR(temp2), \
+                                                    W, wsrc, mask, sse);       \
+  }
+
+HIGHBD_OBMC_VAR(4, 4)
+HIGHBD_OBMC_SUBPIX_VAR(4, 4)
+
+HIGHBD_OBMC_VAR(4, 8)
+HIGHBD_OBMC_SUBPIX_VAR(4, 8)
+
+HIGHBD_OBMC_VAR(8, 4)
+HIGHBD_OBMC_SUBPIX_VAR(8, 4)
+
+HIGHBD_OBMC_VAR(8, 8)
+HIGHBD_OBMC_SUBPIX_VAR(8, 8)
+
+HIGHBD_OBMC_VAR(8, 16)
+HIGHBD_OBMC_SUBPIX_VAR(8, 16)
+
+HIGHBD_OBMC_VAR(16, 8)
+HIGHBD_OBMC_SUBPIX_VAR(16, 8)
+
+HIGHBD_OBMC_VAR(16, 16)
+HIGHBD_OBMC_SUBPIX_VAR(16, 16)
+
+HIGHBD_OBMC_VAR(16, 32)
+HIGHBD_OBMC_SUBPIX_VAR(16, 32)
+
+HIGHBD_OBMC_VAR(32, 16)
+HIGHBD_OBMC_SUBPIX_VAR(32, 16)
+
+HIGHBD_OBMC_VAR(32, 32)
+HIGHBD_OBMC_SUBPIX_VAR(32, 32)
+
+HIGHBD_OBMC_VAR(32, 64)
+HIGHBD_OBMC_SUBPIX_VAR(32, 64)
+
+HIGHBD_OBMC_VAR(64, 32)
+HIGHBD_OBMC_SUBPIX_VAR(64, 32)
+
+HIGHBD_OBMC_VAR(64, 64)
+HIGHBD_OBMC_SUBPIX_VAR(64, 64)
+
+#if CONFIG_EXT_PARTITION
+HIGHBD_OBMC_VAR(64, 128)
+HIGHBD_OBMC_SUBPIX_VAR(64, 128)
+
+HIGHBD_OBMC_VAR(128, 64)
+HIGHBD_OBMC_SUBPIX_VAR(128, 64)
+
+HIGHBD_OBMC_VAR(128, 128)
+HIGHBD_OBMC_SUBPIX_VAR(128, 128)
+#endif  // CONFIG_EXT_PARTITION
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_AV1 && CONFIG_MOTION_VAR
diff --git a/third_party/aom/aom_dsp/variance.h b/third_party/aom/aom_dsp/variance.h
new file mode 100644
index 0000000000..7c925cfac0
--- /dev/null
+++ b/third_party/aom/aom_dsp/variance.h
@@ -0,0 +1,132 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_VARIANCE_H_
+#define AOM_DSP_VARIANCE_H_
+
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define FILTER_BITS 7
+#define FILTER_WEIGHT 128
+
+typedef unsigned int (*aom_sad_fn_t)(const uint8_t *a, int a_stride,
+                                     const uint8_t *b, int b_stride);
+
+typedef unsigned int (*aom_sad_avg_fn_t)(const uint8_t *a, int a_stride,
+                                         const uint8_t *b, int b_stride,
+                                         const uint8_t *second_pred);
+
+typedef void (*aom_copy32xn_fn_t)(const uint8_t *a, int a_stride, uint8_t *b,
+                                  int b_stride, int n);
+
+typedef void (*aom_sad_multi_fn_t)(const uint8_t *a, int a_stride,
+                                   const uint8_t *b, int b_stride,
+                                   unsigned int *sad_array);
+
+typedef void (*aom_sad_multi_d_fn_t)(const uint8_t *a, int a_stride,
+                                     const uint8_t *const b_array[],
+                                     int b_stride, unsigned int *sad_array);
+
+typedef unsigned int (*aom_variance_fn_t)(const uint8_t *a, int a_stride,
+                                          const uint8_t *b, int b_stride,
+                                          unsigned int *sse);
+
+typedef unsigned int (*aom_subpixvariance_fn_t)(const uint8_t *a, int a_stride,
+                                                int xoffset, int yoffset,
+                                                const uint8_t *b, int b_stride,
+                                                unsigned int *sse);
+
+typedef unsigned int (*aom_subp_avg_variance_fn_t)(
+    const uint8_t *a, int a_stride, int xoffset, int yoffset, const uint8_t *b,
+    int b_stride, unsigned int *sse, const uint8_t *second_pred);
+
+#if CONFIG_AV1 && CONFIG_EXT_INTER
+typedef unsigned int (*aom_masked_sad_fn_t)(const uint8_t *src, int src_stride,
+                                            const uint8_t *ref, int ref_stride,
+                                            const uint8_t *msk_ptr,
+                                            int msk_stride);
+typedef unsigned int (*aom_masked_variance_fn_t)(
+    const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,
+    const uint8_t *msk, int msk_stride, unsigned int *sse);
+typedef unsigned int (*aom_masked_subpixvariance_fn_t)(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *ref, int ref_stride, const uint8_t *msk, int msk_stride,
+    unsigned int *sse);
+#endif  // CONFIG_AV1 && CONFIG_EXT_INTER
+
+#if CONFIG_AV1 && CONFIG_MOTION_VAR
+typedef unsigned int (*aom_obmc_sad_fn_t)(const uint8_t *pred, int pred_stride,
+                                          const int32_t *wsrc,
+                                          const int32_t *msk);
+typedef unsigned int (*aom_obmc_variance_fn_t)(const uint8_t *pred,
+                                               int pred_stride,
+                                               const int32_t *wsrc,
+                                               const int32_t *msk,
+                                               unsigned int *sse);
+typedef unsigned int (*aom_obmc_subpixvariance_fn_t)(
+    const uint8_t *pred, int pred_stride, int xoffset, int yoffset,
+    const int32_t *wsrc, const int32_t *msk, unsigned int *sse);
+#endif  // CONFIG_AV1 && CONFIG_MOTION_VAR
+
+#if CONFIG_AV1
+typedef struct aom_variance_vtable {
+  aom_sad_fn_t sdf;
+  aom_sad_avg_fn_t sdaf;
+  aom_variance_fn_t vf;
+  aom_subpixvariance_fn_t svf;
+  aom_subp_avg_variance_fn_t svaf;
+  aom_sad_multi_fn_t sdx3f;
+  aom_sad_multi_fn_t sdx8f;
+  aom_sad_multi_d_fn_t sdx4df;
+#if CONFIG_EXT_INTER
+  aom_masked_sad_fn_t msdf;
+  aom_masked_variance_fn_t mvf;
+  aom_masked_subpixvariance_fn_t msvf;
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR
+  aom_obmc_sad_fn_t osdf;
+  aom_obmc_variance_fn_t ovf;
+  aom_obmc_subpixvariance_fn_t osvf;
+#endif  // CONFIG_MOTION_VAR
+} aom_variance_fn_ptr_t;
+#endif  // CONFIG_AV1
+
+void aom_highbd_var_filter_block2d_bil_first_pass(
+    const uint8_t *src_ptr8, uint16_t *output_ptr,
+    unsigned int src_pixels_per_line, int pixel_step,
+    unsigned int output_height, unsigned int output_width,
+    const uint8_t *filter);
+
+void aom_highbd_var_filter_block2d_bil_second_pass(
+    const uint16_t *src_ptr, uint16_t *output_ptr,
+    unsigned int src_pixels_per_line, unsigned int pixel_step,
+    unsigned int output_height, unsigned int output_width,
+    const uint8_t *filter);
+
+uint32_t aom_sse_odd_size(const uint8_t *a, int a_stride, const uint8_t *b,
+                          int b_stride, int w, int h);
+
+#if CONFIG_HIGHBITDEPTH
+uint64_t aom_highbd_sse_odd_size(const uint8_t *a, int a_stride,
+                                 const uint8_t *b, int b_stride, int w, int h);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_VARIANCE_H_
diff --git a/third_party/aom/aom_dsp/x86/aom_asm_stubs.c b/third_party/aom/aom_dsp/x86/aom_asm_stubs.c
new file mode 100644
index 0000000000..4067b0b531
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_asm_stubs.c
@@ -0,0 +1,182 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/x86/convolve.h"
+
+#if HAVE_SSE2
+filter8_1dfunction aom_filter_block1d16_v8_sse2;
+filter8_1dfunction aom_filter_block1d16_h8_sse2;
+filter8_1dfunction aom_filter_block1d8_v8_sse2;
+filter8_1dfunction aom_filter_block1d8_h8_sse2;
+filter8_1dfunction aom_filter_block1d4_v8_sse2;
+filter8_1dfunction aom_filter_block1d4_h8_sse2;
+filter8_1dfunction aom_filter_block1d16_v8_avg_sse2;
+filter8_1dfunction aom_filter_block1d16_h8_avg_sse2;
+filter8_1dfunction aom_filter_block1d8_v8_avg_sse2;
+filter8_1dfunction aom_filter_block1d8_h8_avg_sse2;
+filter8_1dfunction aom_filter_block1d4_v8_avg_sse2;
+filter8_1dfunction aom_filter_block1d4_h8_avg_sse2;
+
+filter8_1dfunction aom_filter_block1d16_v2_sse2;
+filter8_1dfunction aom_filter_block1d16_h2_sse2;
+filter8_1dfunction aom_filter_block1d8_v2_sse2;
+filter8_1dfunction aom_filter_block1d8_h2_sse2;
+filter8_1dfunction aom_filter_block1d4_v2_sse2;
+filter8_1dfunction aom_filter_block1d4_h2_sse2;
+filter8_1dfunction aom_filter_block1d16_v2_avg_sse2;
+filter8_1dfunction aom_filter_block1d16_h2_avg_sse2;
+filter8_1dfunction aom_filter_block1d8_v2_avg_sse2;
+filter8_1dfunction aom_filter_block1d8_h2_avg_sse2;
+filter8_1dfunction aom_filter_block1d4_v2_avg_sse2;
+filter8_1dfunction aom_filter_block1d4_h2_avg_sse2;
+
+// void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+// void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                              uint8_t *dst, ptrdiff_t dst_stride,
+//                              const int16_t *filter_x, int x_step_q4,
+//                              const int16_t *filter_y, int y_step_q4,
+//                              int w, int h);
+// void aom_convolve8_avg_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                   uint8_t *dst, ptrdiff_t dst_stride,
+//                                   const int16_t *filter_x, int x_step_q4,
+//                                   const int16_t *filter_y, int y_step_q4,
+//                                   int w, int h);
+// void aom_convolve8_avg_vert_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                  uint8_t *dst, ptrdiff_t dst_stride,
+//                                  const int16_t *filter_x, int x_step_q4,
+//                                  const int16_t *filter_y, int y_step_q4,
+//                                  int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2);
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2);
+FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, sse2);
+FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_, sse2);
+
+// void aom_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                         uint8_t *dst, ptrdiff_t dst_stride,
+//                         const int16_t *filter_x, int x_step_q4,
+//                         const int16_t *filter_y, int y_step_q4,
+//                         int w, int h);
+// void aom_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                             uint8_t *dst, ptrdiff_t dst_stride,
+//                             const int16_t *filter_x, int x_step_q4,
+//                             const int16_t *filter_y, int y_step_q4,
+//                             int w, int h);
+FUN_CONV_2D(, sse2);
+FUN_CONV_2D(avg_, sse2);
+
+#if CONFIG_HIGHBITDEPTH && ARCH_X86_64
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_v8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_h8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_v8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_h8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_v8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_h8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_v8_avg_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_h8_avg_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_v8_avg_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_h8_avg_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_v8_avg_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_h8_avg_sse2;
+
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_v2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_h2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_v2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_h2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_v2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_h2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_v2_avg_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_h2_avg_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_v2_avg_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_h2_avg_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_v2_avg_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_h2_avg_sse2;
+
+// void aom_highbd_convolve8_horiz_sse2(const uint8_t *src,
+//                                      ptrdiff_t src_stride,
+//                                      uint8_t *dst,
+//                                      ptrdiff_t dst_stride,
+//                                      const int16_t *filter_x,
+//                                      int x_step_q4,
+//                                      const int16_t *filter_y,
+//                                      int y_step_q4,
+//                                      int w, int h, int bd);
+// void aom_highbd_convolve8_vert_sse2(const uint8_t *src,
+//                                     ptrdiff_t src_stride,
+//                                     uint8_t *dst,
+//                                     ptrdiff_t dst_stride,
+//                                     const int16_t *filter_x,
+//                                     int x_step_q4,
+//                                     const int16_t *filter_y,
+//                                     int y_step_q4,
+//                                     int w, int h, int bd);
+// void aom_highbd_convolve8_avg_horiz_sse2(const uint8_t *src,
+//                                          ptrdiff_t src_stride,
+//                                          uint8_t *dst,
+//                                          ptrdiff_t dst_stride,
+//                                          const int16_t *filter_x,
+//                                          int x_step_q4,
+//                                          const int16_t *filter_y,
+//                                          int y_step_q4,
+//                                          int w, int h, int bd);
+// void aom_highbd_convolve8_avg_vert_sse2(const uint8_t *src,
+//                                         ptrdiff_t src_stride,
+//                                         uint8_t *dst,
+//                                         ptrdiff_t dst_stride,
+//                                         const int16_t *filter_x,
+//                                         int x_step_q4,
+//                                         const int16_t *filter_y,
+//                                         int y_step_q4,
+//                                         int w, int h, int bd);
+HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2);
+HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2);
+HIGH_FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, sse2);
+HIGH_FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_,
+                 sse2);
+
+// void aom_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h, int bd);
+// void aom_highbd_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                                    uint8_t *dst, ptrdiff_t dst_stride,
+//                                    const int16_t *filter_x, int x_step_q4,
+//                                    const int16_t *filter_y, int y_step_q4,
+//                                    int w, int h, int bd);
+HIGH_FUN_CONV_2D(, sse2);
+HIGH_FUN_CONV_2D(avg_, sse2);
+
+#if CONFIG_LOOP_RESTORATION
+// The SSE2 highbd convolve functions can deal with coefficients up to 32767.
+// So redirect highbd_convolve8_add_src to regular highbd_convolve8.
+void aom_highbd_convolve8_add_src_sse2(const uint8_t *src, ptrdiff_t src_stride,
+                                       uint8_t *dst, ptrdiff_t dst_stride,
+                                       const int16_t *filter_x, int x_step_q4,
+                                       const int16_t *filter_y, int y_step_q4,
+                                       int w, int h, int bd) {
+  assert(x_step_q4 == 16);
+  assert(y_step_q4 == 16);
+  ((int16_t *)filter_x)[3] += 128;
+  ((int16_t *)filter_y)[3] += 128;
+  aom_highbd_convolve8_sse2(src, src_stride, dst, dst_stride, filter_x,
+                            x_step_q4, filter_y, y_step_q4, w, h, bd);
+  ((int16_t *)filter_x)[3] -= 128;
+  ((int16_t *)filter_y)[3] -= 128;
+}
+#endif  // CONFIG_LOOP_RESTORATION
+#endif  // CONFIG_HIGHBITDEPTH && ARCH_X86_64
+#endif  // HAVE_SSE2
diff --git a/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.asm b/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.asm
new file mode 100644
index 0000000000..4d31428671
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.asm
@@ -0,0 +1,345 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro convolve_fn 1-2
+%ifidn %1, avg
+%define AUX_XMM_REGS 4
+%else
+%define AUX_XMM_REGS 0
+%endif
+%ifidn %2, highbd
+%define pavg pavgw
+cglobal %2_convolve_%1, 4, 7, 4+AUX_XMM_REGS, src, src_stride, \
+                                              dst, dst_stride, \
+                                              fx, fxs, fy, fys, w, h, bd
+%else
+%define pavg pavgb
+cglobal convolve_%1, 4, 7, 4+AUX_XMM_REGS, src, src_stride, \
+                                           dst, dst_stride, \
+                                           fx, fxs, fy, fys, w, h
+%endif
+  mov r4d, dword wm
+%ifidn %2, highbd
+  shl r4d, 1
+  shl srcq, 1
+  shl src_strideq, 1
+  shl dstq, 1
+  shl dst_strideq, 1
+%else
+  cmp r4d, 4
+  je .w4
+%endif
+  cmp r4d, 8
+  je .w8
+  cmp r4d, 16
+  je .w16
+  cmp r4d, 32
+  je .w32
+
+%if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  cmp r4d, 64
+  je .w64
+%ifidn %2, highbd
+  cmp r4d, 128
+  je .w128
+
+.w256:
+  mov                    r4d, dword hm
+.loop256:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+16]
+  movu                    m2, [srcq+32]
+  movu                    m3, [srcq+48]
+%ifidn %1, avg
+  pavg                    m0, [dstq]
+  pavg                    m1, [dstq+16]
+  pavg                    m2, [dstq+32]
+  pavg                    m3, [dstq+48]
+%endif
+  mova             [dstq   ], m0
+  mova             [dstq+16], m1
+  mova             [dstq+32], m2
+  mova             [dstq+48], m3
+  movu                    m0, [srcq+64]
+  movu                    m1, [srcq+80]
+  movu                    m2, [srcq+96]
+  movu                    m3, [srcq+112]
+%ifidn %1, avg
+  pavg                    m0, [dstq+64]
+  pavg                    m1, [dstq+80]
+  pavg                    m2, [dstq+96]
+  pavg                    m3, [dstq+112]
+%endif
+  mova             [dstq+64], m0
+  mova             [dstq+80], m1
+  mova             [dstq+96], m2
+  mova            [dstq+112], m3
+  movu                    m0, [srcq+128]
+  movu                    m1, [srcq+128+16]
+  movu                    m2, [srcq+128+32]
+  movu                    m3, [srcq+128+48]
+%ifidn %1, avg
+  pavg                    m0, [dstq+128]
+  pavg                    m1, [dstq+128+16]
+  pavg                    m2, [dstq+128+32]
+  pavg                    m3, [dstq+128+48]
+%endif
+  mova         [dstq+128   ], m0
+  mova         [dstq+128+16], m1
+  mova         [dstq+128+32], m2
+  mova         [dstq+128+48], m3
+  movu                    m0, [srcq+128+64]
+  movu                    m1, [srcq+128+80]
+  movu                    m2, [srcq+128+96]
+  movu                    m3, [srcq+128+112]
+  add                   srcq, src_strideq
+%ifidn %1, avg
+  pavg                    m0, [dstq+128+64]
+  pavg                    m1, [dstq+128+80]
+  pavg                    m2, [dstq+128+96]
+  pavg                    m3, [dstq+128+112]
+%endif
+  mova         [dstq+128+64], m0
+  mova         [dstq+128+80], m1
+  mova         [dstq+128+96], m2
+  mova        [dstq+128+112], m3
+  add                   dstq, dst_strideq
+  sub                    r4d, 1
+  jnz .loop256
+  RET
+%endif
+
+.w128:
+  mov                    r4d, dword hm
+.loop128:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+16]
+  movu                    m2, [srcq+32]
+  movu                    m3, [srcq+48]
+%ifidn %1, avg
+  pavg                    m0, [dstq]
+  pavg                    m1, [dstq+16]
+  pavg                    m2, [dstq+32]
+  pavg                    m3, [dstq+48]
+%endif
+  mova             [dstq   ], m0
+  mova             [dstq+16], m1
+  mova             [dstq+32], m2
+  mova             [dstq+48], m3
+  movu                    m0, [srcq+64]
+  movu                    m1, [srcq+80]
+  movu                    m2, [srcq+96]
+  movu                    m3, [srcq+112]
+  add                   srcq, src_strideq
+%ifidn %1, avg
+  pavg                    m0, [dstq+64]
+  pavg                    m1, [dstq+80]
+  pavg                    m2, [dstq+96]
+  pavg                    m3, [dstq+112]
+%endif
+  mova             [dstq+64], m0
+  mova             [dstq+80], m1
+  mova             [dstq+96], m2
+  mova            [dstq+112], m3
+  add                   dstq, dst_strideq
+  sub                    r4d, 1
+  jnz .loop128
+  RET
+
+%else  ; CONFIG_AV1 && CONFIG_EXT_PARTITION
+
+%ifidn %2, highbd
+  cmp r4d, 64
+  je .w64
+
+  mov                    r4d, dword hm
+.loop128:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+16]
+  movu                    m2, [srcq+32]
+  movu                    m3, [srcq+48]
+%ifidn %1, avg
+  pavg                    m0, [dstq]
+  pavg                    m1, [dstq+16]
+  pavg                    m2, [dstq+32]
+  pavg                    m3, [dstq+48]
+%endif
+  mova             [dstq   ], m0
+  mova             [dstq+16], m1
+  mova             [dstq+32], m2
+  mova             [dstq+48], m3
+  movu                    m0, [srcq+64]
+  movu                    m1, [srcq+80]
+  movu                    m2, [srcq+96]
+  movu                    m3, [srcq+112]
+  add                   srcq, src_strideq
+%ifidn %1, avg
+  pavg                    m0, [dstq+64]
+  pavg                    m1, [dstq+80]
+  pavg                    m2, [dstq+96]
+  pavg                    m3, [dstq+112]
+%endif
+  mova             [dstq+64], m0
+  mova             [dstq+80], m1
+  mova             [dstq+96], m2
+  mova            [dstq+112], m3
+  add                   dstq, dst_strideq
+  sub                    r4d, 1
+  jnz .loop128
+  RET
+%endif
+%endif  ; CONFIG_AV1 && CONFIG_EXT_PARTITION
+
+.w64:
+  mov                    r4d, dword hm
+.loop64:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+16]
+  movu                    m2, [srcq+32]
+  movu                    m3, [srcq+48]
+  add                   srcq, src_strideq
+%ifidn %1, avg
+  pavg                    m0, [dstq]
+  pavg                    m1, [dstq+16]
+  pavg                    m2, [dstq+32]
+  pavg                    m3, [dstq+48]
+%endif
+  mova             [dstq   ], m0
+  mova             [dstq+16], m1
+  mova             [dstq+32], m2
+  mova             [dstq+48], m3
+  add                   dstq, dst_strideq
+  sub                    r4d, 1
+  jnz .loop64
+  RET
+
+.w32:
+  mov                    r4d, dword hm
+.loop32:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+16]
+  movu                    m2, [srcq+src_strideq]
+  movu                    m3, [srcq+src_strideq+16]
+  lea                   srcq, [srcq+src_strideq*2]
+%ifidn %1, avg
+  pavg                    m0, [dstq]
+  pavg                    m1, [dstq            +16]
+  pavg                    m2, [dstq+dst_strideq]
+  pavg                    m3, [dstq+dst_strideq+16]
+%endif
+  mova [dstq               ], m0
+  mova [dstq            +16], m1
+  mova [dstq+dst_strideq   ], m2
+  mova [dstq+dst_strideq+16], m3
+  lea                   dstq, [dstq+dst_strideq*2]
+  sub                    r4d, 2
+  jnz .loop32
+  RET
+
+.w16:
+  mov                    r4d, dword hm
+  lea                    r5q, [src_strideq*3]
+  lea                    r6q, [dst_strideq*3]
+.loop16:
+  movu                    m0, [srcq]
+  movu                    m1, [srcq+src_strideq]
+  movu                    m2, [srcq+src_strideq*2]
+  movu                    m3, [srcq+r5q]
+  lea                   srcq, [srcq+src_strideq*4]
+%ifidn %1, avg
+  pavg                    m0, [dstq]
+  pavg                    m1, [dstq+dst_strideq]
+  pavg                    m2, [dstq+dst_strideq*2]
+  pavg                    m3, [dstq+r6q]
+%endif
+  mova  [dstq              ], m0
+  mova  [dstq+dst_strideq  ], m1
+  mova  [dstq+dst_strideq*2], m2
+  mova  [dstq+r6q          ], m3
+  lea                   dstq, [dstq+dst_strideq*4]
+  sub                    r4d, 4
+  jnz .loop16
+  RET
+
+.w8:
+  mov                    r4d, dword hm
+  lea                    r5q, [src_strideq*3]
+  lea                    r6q, [dst_strideq*3]
+.loop8:
+  movh                    m0, [srcq]
+  movh                    m1, [srcq+src_strideq]
+  movh                    m2, [srcq+src_strideq*2]
+  movh                    m3, [srcq+r5q]
+  lea                   srcq, [srcq+src_strideq*4]
+%ifidn %1, avg
+  movh                    m4, [dstq]
+  movh                    m5, [dstq+dst_strideq]
+  movh                    m6, [dstq+dst_strideq*2]
+  movh                    m7, [dstq+r6q]
+  pavg                    m0, m4
+  pavg                    m1, m5
+  pavg                    m2, m6
+  pavg                    m3, m7
+%endif
+  movh  [dstq              ], m0
+  movh  [dstq+dst_strideq  ], m1
+  movh  [dstq+dst_strideq*2], m2
+  movh  [dstq+r6q          ], m3
+  lea                   dstq, [dstq+dst_strideq*4]
+  sub                    r4d, 4
+  jnz .loop8
+  RET
+
+%ifnidn %2, highbd
+.w4:
+  mov                    r4d, dword hm
+  lea                    r5q, [src_strideq*3]
+  lea                    r6q, [dst_strideq*3]
+.loop4:
+  movd                    m0, [srcq]
+  movd                    m1, [srcq+src_strideq]
+  movd                    m2, [srcq+src_strideq*2]
+  movd                    m3, [srcq+r5q]
+  lea                   srcq, [srcq+src_strideq*4]
+%ifidn %1, avg
+  movd                    m4, [dstq]
+  movd                    m5, [dstq+dst_strideq]
+  movd                    m6, [dstq+dst_strideq*2]
+  movd                    m7, [dstq+r6q]
+  pavg                    m0, m4
+  pavg                    m1, m5
+  pavg                    m2, m6
+  pavg                    m3, m7
+%endif
+  movd  [dstq              ], m0
+  movd  [dstq+dst_strideq  ], m1
+  movd  [dstq+dst_strideq*2], m2
+  movd  [dstq+r6q          ], m3
+  lea                   dstq, [dstq+dst_strideq*4]
+  sub                    r4d, 4
+  jnz .loop4
+  RET
+%endif
+%endmacro
+
+INIT_XMM sse2
+convolve_fn copy
+convolve_fn avg
+%if CONFIG_HIGHBITDEPTH
+convolve_fn copy, highbd
+convolve_fn avg, highbd
+%endif
diff --git a/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm b/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm
new file mode 100644
index 0000000000..e6d357ba3f
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm
@@ -0,0 +1,965 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+%include "aom_ports/x86_abi_support.asm"
+
+;Note: tap3 and tap4 have to be applied and added after other taps to avoid
+;overflow.
+
+%macro HIGH_GET_FILTERS_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    psrldq      xmm7, 8
+    pshuflw     xmm4, xmm7, 0b              ;k4
+    pshuflw     xmm5, xmm7, 01010101b       ;k5
+    pshuflw     xmm6, xmm7, 10101010b       ;k6
+    pshuflw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklwd   xmm0, xmm6
+    punpcklwd   xmm2, xmm5
+    punpcklwd   xmm3, xmm4
+    punpcklwd   xmm1, xmm7
+
+    movdqa      k0k6, xmm0
+    movdqa      k2k5, xmm2
+    movdqa      k3k4, xmm3
+    movdqa      k1k7, xmm1
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6
+
+    ;Compute max and min values of a pixel
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)      ;bps
+    movq        xmm0, rdx
+    movq        xmm1, rcx
+    pshufd      xmm0, xmm0, 0b
+    movdqa      xmm2, xmm0
+    psllw       xmm0, xmm1
+    psubw       xmm0, xmm2
+    pxor        xmm1, xmm1
+    movdqa      max, xmm0                  ;max value (for clamping)
+    movdqa      min, xmm1                  ;min value (for clamping)
+
+%endm
+
+%macro HIGH_APPLY_FILTER_4 1
+    punpcklwd   xmm0, xmm6                  ;two row in one register
+    punpcklwd   xmm1, xmm7
+    punpcklwd   xmm2, xmm5
+    punpcklwd   xmm3, xmm4
+
+    pmaddwd     xmm0, k0k6                  ;multiply the filter factors
+    pmaddwd     xmm1, k1k7
+    pmaddwd     xmm2, k2k5
+    pmaddwd     xmm3, k3k4
+
+    paddd       xmm0, xmm1                  ;sum
+    paddd       xmm0, xmm2
+    paddd       xmm0, xmm3
+
+    paddd       xmm0, krd                   ;rounding
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm0                  ;pack to word
+
+    ;clamp the values
+    pminsw      xmm0, max
+    pmaxsw      xmm0, min
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0
+%endm
+
+%macro HIGH_GET_FILTERS 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    pshufhw     xmm4, xmm7, 0b              ;k4
+    pshufhw     xmm5, xmm7, 01010101b       ;k5
+    pshufhw     xmm6, xmm7, 10101010b       ;k6
+    pshufhw     xmm7, xmm7, 11111111b       ;k7
+    punpcklqdq  xmm2, xmm2
+    punpcklqdq  xmm3, xmm3
+    punpcklwd   xmm0, xmm1
+    punpckhwd   xmm6, xmm7
+    punpckhwd   xmm2, xmm5
+    punpckhwd   xmm3, xmm4
+
+    movdqa      k0k1, xmm0                  ;store filter factors on stack
+    movdqa      k6k7, xmm6
+    movdqa      k2k5, xmm2
+    movdqa      k3k4, xmm3
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6                   ;rounding
+
+    ;Compute max and min values of a pixel
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm0, rdx
+    movq        xmm1, rcx
+    pshufd      xmm0, xmm0, 0b
+    movdqa      xmm2, xmm0
+    psllw       xmm0, xmm1
+    psubw       xmm0, xmm2
+    pxor        xmm1, xmm1
+    movdqa      max, xmm0                  ;max value (for clamping)
+    movdqa      min, xmm1                  ;min value (for clamping)
+%endm
+
+%macro LOAD_VERT_8 1
+    movdqu      xmm0, [rsi + %1]            ;0
+    movdqu      xmm1, [rsi + rax + %1]      ;1
+    movdqu      xmm6, [rsi + rdx * 2 + %1]  ;6
+    lea         rsi,  [rsi + rax]
+    movdqu      xmm7, [rsi + rdx * 2 + %1]  ;7
+    movdqu      xmm2, [rsi + rax + %1]      ;2
+    movdqu      xmm3, [rsi + rax * 2 + %1]  ;3
+    movdqu      xmm4, [rsi + rdx + %1]      ;4
+    movdqu      xmm5, [rsi + rax * 4 + %1]  ;5
+%endm
+
+%macro HIGH_APPLY_FILTER_8 2
+    movdqu      temp, xmm4
+    movdqa      xmm4, xmm0
+    punpcklwd   xmm0, xmm1
+    punpckhwd   xmm4, xmm1
+    movdqa      xmm1, xmm6
+    punpcklwd   xmm6, xmm7
+    punpckhwd   xmm1, xmm7
+    movdqa      xmm7, xmm2
+    punpcklwd   xmm2, xmm5
+    punpckhwd   xmm7, xmm5
+
+    movdqu      xmm5, temp
+    movdqu      temp, xmm4
+    movdqa      xmm4, xmm3
+    punpcklwd   xmm3, xmm5
+    punpckhwd   xmm4, xmm5
+    movdqu      xmm5, temp
+
+    pmaddwd     xmm0, k0k1
+    pmaddwd     xmm5, k0k1
+    pmaddwd     xmm6, k6k7
+    pmaddwd     xmm1, k6k7
+    pmaddwd     xmm2, k2k5
+    pmaddwd     xmm7, k2k5
+    pmaddwd     xmm3, k3k4
+    pmaddwd     xmm4, k3k4
+
+    paddd       xmm0, xmm6
+    paddd       xmm0, xmm2
+    paddd       xmm0, xmm3
+    paddd       xmm5, xmm1
+    paddd       xmm5, xmm7
+    paddd       xmm5, xmm4
+
+    paddd       xmm0, krd                   ;rounding
+    paddd       xmm5, krd
+    psrad       xmm0, 7                     ;shift
+    psrad       xmm5, 7
+    packssdw    xmm0, xmm5                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, max
+    pmaxsw      xmm0, min
+
+%if %1
+    movdqu      xmm1, [rdi + %2]
+    pavgw       xmm0, xmm1
+%endif
+    movdqu      [rdi + %2], xmm0
+%endm
+
+;void aom_filter_block1d4_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(aom_highbd_filter_block1d4_v8_sse2) PRIVATE
+sym(aom_highbd_filter_block1d4_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movq        xmm0, [rsi]                 ;load src: row 0
+    movq        xmm1, [rsi + rax]           ;1
+    movq        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2]       ;7
+    movq        xmm2, [rsi + rax]           ;2
+    movq        xmm3, [rsi + rax * 2]       ;3
+    movq        xmm4, [rsi + rdx]           ;4
+    movq        xmm5, [rsi + rax * 4]       ;5
+
+    HIGH_APPLY_FILTER_4 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d8_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(aom_highbd_filter_block1d8_v8_sse2) PRIVATE
+sym(aom_highbd_filter_block1d8_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 0, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d16_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(aom_highbd_filter_block1d16_v8_sse2) PRIVATE
+sym(aom_highbd_filter_block1d16_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 0, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 16
+    HIGH_APPLY_FILTER_8 0, 16
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_highbd_filter_block1d4_v8_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d4_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movq        xmm0, [rsi]                 ;load src: row 0
+    movq        xmm1, [rsi + rax]           ;1
+    movq        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2]       ;7
+    movq        xmm2, [rsi + rax]           ;2
+    movq        xmm3, [rsi + rax * 2]       ;3
+    movq        xmm4, [rsi + rdx]           ;4
+    movq        xmm5, [rsi + rax * 4]       ;5
+
+    HIGH_APPLY_FILTER_4 1
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_highbd_filter_block1d8_v8_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d8_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 1, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_highbd_filter_block1d16_v8_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d16_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 1, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 16
+    HIGH_APPLY_FILTER_8 1, 16
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d4_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(aom_highbd_filter_block1d4_h8_sse2) PRIVATE
+sym(aom_highbd_filter_block1d4_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm4,   [rsi + 2]
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm4
+    movdqa      xmm7, xmm4
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm4
+
+    psrldq      xmm1, 2
+    psrldq      xmm6, 4
+    psrldq      xmm7, 6
+    psrldq      xmm2, 4
+    psrldq      xmm3, 6
+    psrldq      xmm5, 2
+
+    HIGH_APPLY_FILTER_4 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d8_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(aom_highbd_filter_block1d8_h8_sse2) PRIVATE
+sym(aom_highbd_filter_block1d8_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 0, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d16_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(aom_highbd_filter_block1d16_h8_sse2) PRIVATE
+sym(aom_highbd_filter_block1d16_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 0, 0
+
+    movdqu      xmm0,   [rsi + 10]           ;load src
+    movdqu      xmm1,   [rsi + 12]
+    movdqu      xmm2,   [rsi + 14]
+    movdqu      xmm3,   [rsi + 16]
+    movdqu      xmm4,   [rsi + 18]
+    movdqu      xmm5,   [rsi + 20]
+    movdqu      xmm6,   [rsi + 22]
+    movdqu      xmm7,   [rsi + 24]
+
+    HIGH_APPLY_FILTER_8 0, 16
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_highbd_filter_block1d4_h8_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d4_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm4,   [rsi + 2]
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm4
+    movdqa      xmm7, xmm4
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm4
+
+    psrldq      xmm1, 2
+    psrldq      xmm6, 4
+    psrldq      xmm7, 6
+    psrldq      xmm2, 4
+    psrldq      xmm3, 6
+    psrldq      xmm5, 2
+
+    HIGH_APPLY_FILTER_4 1
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_highbd_filter_block1d8_h8_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d8_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 1, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_highbd_filter_block1d16_h8_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d16_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 1, 0
+
+    movdqu      xmm0,   [rsi + 10]           ;load src
+    movdqu      xmm1,   [rsi + 12]
+    movdqu      xmm2,   [rsi + 14]
+    movdqu      xmm3,   [rsi + 16]
+    movdqu      xmm4,   [rsi + 18]
+    movdqu      xmm5,   [rsi + 20]
+    movdqu      xmm6,   [rsi + 22]
+    movdqu      xmm7,   [rsi + 24]
+
+    HIGH_APPLY_FILTER_8 1, 16
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm b/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm
new file mode 100644
index 0000000000..9e2ec748c7
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm
@@ -0,0 +1,497 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "aom_ports/x86_abi_support.asm"
+
+%macro HIGH_GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    pshuflw     xmm4, xmm3, 11111111b       ;k3
+    psrldq      xmm3, 8
+    pshuflw     xmm3, xmm3, 0b              ;k4
+    punpcklwd   xmm4, xmm3                  ;k3k4
+
+    movq        xmm3, rcx                   ;rounding
+    pshufd      xmm3, xmm3, 0
+
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm5, rdx
+    movq        xmm2, rcx
+    pshufd      xmm5, xmm5, 0b
+    movdqa      xmm1, xmm5
+    psllw       xmm5, xmm2
+    psubw       xmm5, xmm1                  ;max value (for clamping)
+    pxor        xmm2, xmm2                  ;min value (for clamping)
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro HIGH_APPLY_FILTER_4 1
+
+    punpcklwd   xmm0, xmm1                  ;two row in one register
+    pmaddwd     xmm0, xmm4                  ;multiply the filter factors
+
+    paddd       xmm0, xmm3                  ;rounding
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm0                  ;pack to word
+
+    ;clamp the values
+    pminsw      xmm0, xmm5
+    pmaxsw      xmm0, xmm2
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+
+    movq        [rdi], xmm0
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+
+%if ARCH_X86_64
+%macro HIGH_GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm6, [rdx]                 ;load filters
+
+    pshuflw     xmm7, xmm6, 11111111b       ;k3
+    pshufhw     xmm6, xmm6, 0b              ;k4
+    psrldq      xmm6, 8
+    punpcklwd   xmm7, xmm6                  ;k3k4k3k4k3k4k3k4
+
+    movq        xmm4, rcx                   ;rounding
+    pshufd      xmm4, xmm4, 0
+
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm8, rdx
+    movq        xmm5, rcx
+    pshufd      xmm8, xmm8, 0b
+    movdqa      xmm1, xmm8
+    psllw       xmm8, xmm5
+    psubw       xmm8, xmm1                  ;max value (for clamping)
+    pxor        xmm5, xmm5                  ;min value (for clamping)
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro HIGH_APPLY_FILTER_8 1
+    movdqa      xmm6, xmm0
+    punpckhwd   xmm6, xmm1
+    punpcklwd   xmm0, xmm1
+    pmaddwd     xmm6, xmm7
+    pmaddwd     xmm0, xmm7
+
+    paddd       xmm6, xmm4                  ;rounding
+    paddd       xmm0, xmm4                  ;rounding
+    psrad       xmm6, 7                     ;shift
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm6                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, xmm8
+    pmaxsw      xmm0, xmm5
+
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+
+%macro HIGH_APPLY_FILTER_16 1
+    movdqa      xmm9, xmm0
+    movdqa      xmm6, xmm2
+    punpckhwd   xmm9, xmm1
+    punpckhwd   xmm6, xmm3
+    punpcklwd   xmm0, xmm1
+    punpcklwd   xmm2, xmm3
+
+    pmaddwd     xmm9, xmm7
+    pmaddwd     xmm6, xmm7
+    pmaddwd     xmm0, xmm7
+    pmaddwd     xmm2, xmm7
+
+    paddd       xmm9, xmm4                  ;rounding
+    paddd       xmm6, xmm4
+    paddd       xmm0, xmm4
+    paddd       xmm2, xmm4
+
+    psrad       xmm9, 7                     ;shift
+    psrad       xmm6, 7
+    psrad       xmm0, 7
+    psrad       xmm2, 7
+
+    packssdw    xmm0, xmm9                  ;pack back to word
+    packssdw    xmm2, xmm6                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, xmm8
+    pmaxsw      xmm0, xmm5
+    pminsw      xmm2, xmm8
+    pmaxsw      xmm2, xmm5
+
+%if %1
+    movdqu      xmm1, [rdi]
+    movdqu      xmm3, [rdi + 16]
+    pavgw       xmm0, xmm1
+    pavgw       xmm2, xmm3
+%endif
+    movdqu      [rdi], xmm0               ;store the result
+    movdqu      [rdi + 16], xmm2          ;store the result
+
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+%endif
+
+global sym(aom_highbd_filter_block1d4_v2_sse2) PRIVATE
+sym(aom_highbd_filter_block1d4_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movq        xmm0, [rsi]                 ;load src
+    movq        xmm1, [rsi + 2*rax]
+
+    HIGH_APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(aom_highbd_filter_block1d8_v2_sse2) PRIVATE
+sym(aom_highbd_filter_block1d8_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;0
+    movdqu      xmm1, [rsi + 2*rax]         ;1
+
+    HIGH_APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_highbd_filter_block1d16_v2_sse2) PRIVATE
+sym(aom_highbd_filter_block1d16_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm2, [rsi + 16]
+    movdqu        xmm1, [rsi + 2*rax]       ;1
+    movdqu        xmm3, [rsi + 2*rax + 16]
+
+    HIGH_APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
+
+global sym(aom_highbd_filter_block1d4_v2_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d4_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movq        xmm0, [rsi]                 ;load src
+    movq        xmm1, [rsi + 2*rax]
+
+    HIGH_APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(aom_highbd_filter_block1d8_v2_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d8_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;0
+    movdqu      xmm1, [rsi + 2*rax]         ;1
+
+    HIGH_APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_highbd_filter_block1d16_v2_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d16_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + 2*rax]       ;1
+    movdqu        xmm2, [rsi + 16]
+    movdqu        xmm3, [rsi + 2*rax + 16]
+
+    HIGH_APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
+
+global sym(aom_highbd_filter_block1d4_h2_sse2) PRIVATE
+sym(aom_highbd_filter_block1d4_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 2
+
+    HIGH_APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(aom_highbd_filter_block1d8_h2_sse2) PRIVATE
+sym(aom_highbd_filter_block1d8_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqu      xmm1, [rsi + 2]
+
+    HIGH_APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_highbd_filter_block1d16_h2_sse2) PRIVATE
+sym(aom_highbd_filter_block1d16_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 2]
+    movdqu      xmm2,   [rsi + 16]
+    movdqu      xmm3,   [rsi + 18]
+
+    HIGH_APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
+
+global sym(aom_highbd_filter_block1d4_h2_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d4_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 2
+
+    HIGH_APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+%if ARCH_X86_64
+global sym(aom_highbd_filter_block1d8_h2_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d8_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqu      xmm1, [rsi + 2]
+
+    HIGH_APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_highbd_filter_block1d16_h2_avg_sse2) PRIVATE
+sym(aom_highbd_filter_block1d16_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 2]
+    movdqu      xmm2,   [rsi + 16]
+    movdqu      xmm3,   [rsi + 18]
+
+    HIGH_APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+%endif
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c
new file mode 100644
index 0000000000..61476b8be6
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c
@@ -0,0 +1,575 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/x86/convolve.h"
+#include "aom_ports/mem.h"
+
+// filters for 16_h8 and 16_v8
+DECLARE_ALIGNED(32, static const uint8_t, filt1_global_avx2[32]) = {
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt2_global_avx2[32]) = {
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10,
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt3_global_avx2[32]) = {
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12,
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt4_global_avx2[32]) = {
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14,
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
+};
+
+#if defined(__clang__)
+#if (__clang_major__ > 0 && __clang_major__ < 3) ||            \
+    (__clang_major__ == 3 && __clang_minor__ <= 3) ||          \
+    (defined(__APPLE__) && defined(__apple_build_version__) && \
+     ((__clang_major__ == 4 && __clang_minor__ <= 2) ||        \
+      (__clang_major__ == 5 && __clang_minor__ == 0)))
+#define MM256_BROADCASTSI128_SI256(x) \
+  _mm_broadcastsi128_si256((__m128i const *)&(x))
+#else  // clang > 3.3, and not 5.0 on macosx.
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif  // clang <= 3.3
+#elif defined(__GNUC__)
+#if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 6)
+#define MM256_BROADCASTSI128_SI256(x) \
+  _mm_broadcastsi128_si256((__m128i const *)&(x))
+#elif __GNUC__ == 4 && __GNUC_MINOR__ == 7
+#define MM256_BROADCASTSI128_SI256(x) _mm_broadcastsi128_si256(x)
+#else  // gcc > 4.7
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif  // gcc <= 4.6
+#else   // !(gcc || clang)
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif  // __clang__
+
+static void aom_filter_block1d16_h8_avx2(
+    const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr,
+    ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) {
+  __m128i filtersReg;
+  __m256i addFilterReg64, filt1Reg, filt2Reg, filt3Reg, filt4Reg;
+  __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3;
+  __m256i srcReg32b1, srcReg32b2, filtersReg32;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to 8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
+  // have the same data in both lanes of a 256 bit register
+  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 256 bit register
+  firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 256 bit register
+  secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 256 bit register
+  thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 256 bit register
+  forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u));
+
+  filt1Reg = _mm256_load_si256((__m256i const *)filt1_global_avx2);
+  filt2Reg = _mm256_load_si256((__m256i const *)filt2_global_avx2);
+  filt3Reg = _mm256_load_si256((__m256i const *)filt3_global_avx2);
+  filt4Reg = _mm256_load_si256((__m256i const *)filt4_global_avx2);
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pixels_per_line << 1;
+  dst_stride = output_pitch << 1;
+  for (i = output_height; i > 1; i -= 2) {
+    // load the 2 strides of source
+    srcReg32b1 =
+        _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr - 3)));
+    srcReg32b1 = _mm256_inserti128_si256(
+        srcReg32b1,
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line - 3)),
+        1);
+
+    // filter the source buffer
+    srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2);
+
+    // filter the source buffer
+    srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b1_1 = _mm256_adds_epi16(
+        srcRegFilt32b1_1, _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+    // reading 2 strides of the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg32b2 =
+        _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr + 5)));
+    srcReg32b2 = _mm256_inserti128_si256(
+        srcReg32b2,
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line + 5)),
+        1);
+
+    // add and saturate the results together
+    srcRegFilt32b1_1 = _mm256_adds_epi16(
+        srcRegFilt32b1_1, _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+    // filter the source buffer
+    srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b2_1 = _mm256_maddubs_epi16(srcRegFilt32b2_1, firstFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, srcRegFilt32b2);
+
+    // filter the source buffer
+    srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b2, filt2Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b2_1 = _mm256_adds_epi16(
+        srcRegFilt32b2_1, _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+    srcRegFilt32b2_1 = _mm256_adds_epi16(
+        srcRegFilt32b2_1, _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg64);
+
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg64);
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 7);
+    srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1);
+
+    src_ptr += src_stride;
+
+    // save 16 bytes
+    _mm_store_si128((__m128i *)output_ptr,
+                    _mm256_castsi256_si128(srcRegFilt32b1_1));
+
+    // save the next 16 bits
+    _mm_store_si128((__m128i *)(output_ptr + output_pitch),
+                    _mm256_extractf128_si256(srcRegFilt32b1_1, 1));
+    output_ptr += dst_stride;
+  }
+
+  // if the number of strides is odd.
+  // process only 16 bytes
+  if (i > 0) {
+    __m128i srcReg1, srcReg2, srcRegFilt1_1, srcRegFilt2_1;
+    __m128i srcRegFilt2, srcRegFilt3;
+
+    srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+
+    // filter the source buffer
+    srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt4Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1_1 =
+        _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters));
+    srcRegFilt2 =
+        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
+
+    // filter the source buffer
+    srcRegFilt3 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt3Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 =
+        _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters));
+    srcRegFilt2 =
+        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt1_1 =
+        _mm_adds_epi16(srcRegFilt1_1, _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+
+    // reading the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 5));
+
+    // add and saturate the results together
+    srcRegFilt1_1 =
+        _mm_adds_epi16(srcRegFilt1_1, _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+
+    // filter the source buffer
+    srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt1Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt4Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt2_1 =
+        _mm_maddubs_epi16(srcRegFilt2_1, _mm256_castsi256_si128(firstFilters));
+    srcRegFilt2 =
+        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
+
+    // filter the source buffer
+    srcRegFilt3 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt2Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt3Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 =
+        _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters));
+    srcRegFilt2 =
+        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt2_1 =
+        _mm_adds_epi16(srcRegFilt2_1, _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+    srcRegFilt2_1 =
+        _mm_adds_epi16(srcRegFilt2_1, _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+
+    srcRegFilt1_1 =
+        _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg64));
+
+    srcRegFilt2_1 =
+        _mm_adds_epi16(srcRegFilt2_1, _mm256_castsi256_si128(addFilterReg64));
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 7);
+    srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, srcRegFilt2_1);
+
+    // save 16 bytes
+    _mm_store_si128((__m128i *)output_ptr, srcRegFilt1_1);
+  }
+}
+
+static void aom_filter_block1d16_v8_avx2(
+    const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr,
+    ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) {
+  __m128i filtersReg;
+  __m256i addFilterReg64;
+  __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5;
+  __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10;
+  __m256i srcReg32b11, srcReg32b12, filtersReg32;
+  __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the
+  // same data in both lanes of 128 bit register.
+  filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
+  // have the same data in both lanes of a 256 bit register
+  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 256 bit register
+  firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 256 bit register
+  secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 256 bit register
+  thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 256 bit register
+  forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  // load 16 bytes 7 times in stride of src_pitch
+  srcReg32b1 =
+      _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr)));
+  srcReg32b2 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch)));
+  srcReg32b3 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2)));
+  srcReg32b4 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3)));
+  srcReg32b5 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4)));
+  srcReg32b6 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5)));
+  srcReg32b7 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6)));
+
+  // have each consecutive loads on the same 256 register
+  srcReg32b1 = _mm256_inserti128_si256(srcReg32b1,
+                                       _mm256_castsi256_si128(srcReg32b2), 1);
+  srcReg32b2 = _mm256_inserti128_si256(srcReg32b2,
+                                       _mm256_castsi256_si128(srcReg32b3), 1);
+  srcReg32b3 = _mm256_inserti128_si256(srcReg32b3,
+                                       _mm256_castsi256_si128(srcReg32b4), 1);
+  srcReg32b4 = _mm256_inserti128_si256(srcReg32b4,
+                                       _mm256_castsi256_si128(srcReg32b5), 1);
+  srcReg32b5 = _mm256_inserti128_si256(srcReg32b5,
+                                       _mm256_castsi256_si128(srcReg32b6), 1);
+  srcReg32b6 = _mm256_inserti128_si256(srcReg32b6,
+                                       _mm256_castsi256_si128(srcReg32b7), 1);
+
+  // merge every two consecutive registers except the last one
+  srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2);
+  srcReg32b1 = _mm256_unpackhi_epi8(srcReg32b1, srcReg32b2);
+
+  // save
+  srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4);
+
+  // save
+  srcReg32b3 = _mm256_unpackhi_epi8(srcReg32b3, srcReg32b4);
+
+  // save
+  srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6);
+
+  // save
+  srcReg32b5 = _mm256_unpackhi_epi8(srcReg32b5, srcReg32b6);
+
+  for (i = output_height; i > 1; i -= 2) {
+    // load the last 2 loads of 16 bytes and have every two
+    // consecutive loads in the same 256 bit register
+    srcReg32b8 = _mm256_castsi128_si256(
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7)));
+    srcReg32b7 = _mm256_inserti128_si256(srcReg32b7,
+                                         _mm256_castsi256_si128(srcReg32b8), 1);
+    srcReg32b9 = _mm256_castsi128_si256(
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 8)));
+    srcReg32b8 = _mm256_inserti128_si256(srcReg32b8,
+                                         _mm256_castsi256_si128(srcReg32b9), 1);
+
+    // merge every two consecutive registers
+    // save
+    srcReg32b4 = _mm256_unpacklo_epi8(srcReg32b7, srcReg32b8);
+    srcReg32b7 = _mm256_unpackhi_epi8(srcReg32b7, srcReg32b8);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcReg32b10 = _mm256_maddubs_epi16(srcReg32b10, firstFilters);
+    srcReg32b6 = _mm256_maddubs_epi16(srcReg32b4, forthFilters);
+
+    // add and saturate the results together
+    srcReg32b10 = _mm256_adds_epi16(srcReg32b10, srcReg32b6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcReg32b8 = _mm256_maddubs_epi16(srcReg32b11, secondFilters);
+    srcReg32b12 = _mm256_maddubs_epi16(srcReg32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
+                                    _mm256_min_epi16(srcReg32b8, srcReg32b12));
+    srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
+                                    _mm256_max_epi16(srcReg32b8, srcReg32b12));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcReg32b1 = _mm256_maddubs_epi16(srcReg32b1, firstFilters);
+    srcReg32b6 = _mm256_maddubs_epi16(srcReg32b7, forthFilters);
+
+    srcReg32b1 = _mm256_adds_epi16(srcReg32b1, srcReg32b6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcReg32b8 = _mm256_maddubs_epi16(srcReg32b3, secondFilters);
+    srcReg32b12 = _mm256_maddubs_epi16(srcReg32b5, thirdFilters);
+
+    // add and saturate the results together
+    srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
+                                   _mm256_min_epi16(srcReg32b8, srcReg32b12));
+    srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
+                                   _mm256_max_epi16(srcReg32b8, srcReg32b12));
+
+    srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg64);
+    srcReg32b1 = _mm256_adds_epi16(srcReg32b1, addFilterReg64);
+
+    // shift by 7 bit each 16 bit
+    srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 7);
+    srcReg32b1 = _mm256_srai_epi16(srcReg32b1, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcReg32b1 = _mm256_packus_epi16(srcReg32b10, srcReg32b1);
+
+    src_ptr += src_stride;
+
+    // save 16 bytes
+    _mm_store_si128((__m128i *)output_ptr, _mm256_castsi256_si128(srcReg32b1));
+
+    // save the next 16 bits
+    _mm_store_si128((__m128i *)(output_ptr + out_pitch),
+                    _mm256_extractf128_si256(srcReg32b1, 1));
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg32b10 = srcReg32b11;
+    srcReg32b1 = srcReg32b3;
+    srcReg32b11 = srcReg32b2;
+    srcReg32b3 = srcReg32b5;
+    srcReg32b2 = srcReg32b4;
+    srcReg32b5 = srcReg32b7;
+    srcReg32b7 = srcReg32b9;
+  }
+  if (i > 0) {
+    __m128i srcRegFilt1, srcRegFilt3, srcRegFilt4, srcRegFilt5;
+    __m128i srcRegFilt6, srcRegFilt7, srcRegFilt8;
+    // load the last 16 bytes
+    srcRegFilt8 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+
+    // merge the last 2 results together
+    srcRegFilt4 =
+        _mm_unpacklo_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
+    srcRegFilt7 =
+        _mm_unpackhi_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b10),
+                                    _mm256_castsi256_si128(firstFilters));
+    srcRegFilt4 =
+        _mm_maddubs_epi16(srcRegFilt4, _mm256_castsi256_si128(forthFilters));
+    srcRegFilt3 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b1),
+                                    _mm256_castsi256_si128(firstFilters));
+    srcRegFilt7 =
+        _mm_maddubs_epi16(srcRegFilt7, _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3, srcRegFilt7);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11),
+                                    _mm256_castsi256_si128(secondFilters));
+    srcRegFilt5 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b3),
+                                    _mm256_castsi256_si128(secondFilters));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt6 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b2),
+                                    _mm256_castsi256_si128(thirdFilters));
+    srcRegFilt7 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b5),
+                                    _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt1 =
+        _mm_adds_epi16(srcRegFilt1, _mm_min_epi16(srcRegFilt4, srcRegFilt6));
+    srcRegFilt3 =
+        _mm_adds_epi16(srcRegFilt3, _mm_min_epi16(srcRegFilt5, srcRegFilt7));
+
+    // add and saturate the results together
+    srcRegFilt1 =
+        _mm_adds_epi16(srcRegFilt1, _mm_max_epi16(srcRegFilt4, srcRegFilt6));
+    srcRegFilt3 =
+        _mm_adds_epi16(srcRegFilt3, _mm_max_epi16(srcRegFilt5, srcRegFilt7));
+
+    srcRegFilt1 =
+        _mm_adds_epi16(srcRegFilt1, _mm256_castsi256_si128(addFilterReg64));
+    srcRegFilt3 =
+        _mm_adds_epi16(srcRegFilt3, _mm256_castsi256_si128(addFilterReg64));
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+    srcRegFilt3 = _mm_srai_epi16(srcRegFilt3, 7);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt3);
+
+    // save 16 bytes
+    _mm_store_si128((__m128i *)output_ptr, srcRegFilt1);
+  }
+}
+
+#if HAVE_AVX2 && HAVE_SSSE3
+filter8_1dfunction aom_filter_block1d4_v8_ssse3;
+#if ARCH_X86_64
+filter8_1dfunction aom_filter_block1d8_v8_intrin_ssse3;
+filter8_1dfunction aom_filter_block1d8_h8_intrin_ssse3;
+filter8_1dfunction aom_filter_block1d4_h8_intrin_ssse3;
+#define aom_filter_block1d8_v8_avx2 aom_filter_block1d8_v8_intrin_ssse3
+#define aom_filter_block1d8_h8_avx2 aom_filter_block1d8_h8_intrin_ssse3
+#define aom_filter_block1d4_h8_avx2 aom_filter_block1d4_h8_intrin_ssse3
+#else  // ARCH_X86
+filter8_1dfunction aom_filter_block1d8_v8_ssse3;
+filter8_1dfunction aom_filter_block1d8_h8_ssse3;
+filter8_1dfunction aom_filter_block1d4_h8_ssse3;
+#define aom_filter_block1d8_v8_avx2 aom_filter_block1d8_v8_ssse3
+#define aom_filter_block1d8_h8_avx2 aom_filter_block1d8_h8_ssse3
+#define aom_filter_block1d4_h8_avx2 aom_filter_block1d4_h8_ssse3
+#endif  // ARCH_X86_64
+filter8_1dfunction aom_filter_block1d16_v2_ssse3;
+filter8_1dfunction aom_filter_block1d16_h2_ssse3;
+filter8_1dfunction aom_filter_block1d8_v2_ssse3;
+filter8_1dfunction aom_filter_block1d8_h2_ssse3;
+filter8_1dfunction aom_filter_block1d4_v2_ssse3;
+filter8_1dfunction aom_filter_block1d4_h2_ssse3;
+#define aom_filter_block1d4_v8_avx2 aom_filter_block1d4_v8_ssse3
+#define aom_filter_block1d16_v2_avx2 aom_filter_block1d16_v2_ssse3
+#define aom_filter_block1d16_h2_avx2 aom_filter_block1d16_h2_ssse3
+#define aom_filter_block1d8_v2_avx2 aom_filter_block1d8_v2_ssse3
+#define aom_filter_block1d8_h2_avx2 aom_filter_block1d8_h2_ssse3
+#define aom_filter_block1d4_v2_avx2 aom_filter_block1d4_v2_ssse3
+#define aom_filter_block1d4_h2_avx2 aom_filter_block1d4_h2_ssse3
+// void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h);
+// void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2);
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2);
+
+// void aom_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                          uint8_t *dst, ptrdiff_t dst_stride,
+//                          const int16_t *filter_x, int x_step_q4,
+//                          const int16_t *filter_y, int y_step_q4,
+//                          int w, int h);
+FUN_CONV_2D(, avx2);
+#endif  // HAVE_AX2 && HAVE_SSSE3
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c
new file mode 100644
index 0000000000..be37738df6
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c
@@ -0,0 +1,920 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <tmmintrin.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/convolve.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/emmintrin_compat.h"
+
+// filters only for the 4_h8 convolution
+DECLARE_ALIGNED(16, static const uint8_t, filt1_4_h8[16]) = {
+  0, 1, 1, 2, 2, 3, 3, 4, 2, 3, 3, 4, 4, 5, 5, 6
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt2_4_h8[16]) = {
+  4, 5, 5, 6, 6, 7, 7, 8, 6, 7, 7, 8, 8, 9, 9, 10
+};
+
+// filters for 8_h8 and 16_h8
+DECLARE_ALIGNED(16, static const uint8_t, filt1_global[16]) = {
+  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt2_global[16]) = {
+  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt3_global[16]) = {
+  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
+};
+
+DECLARE_ALIGNED(16, static const uint8_t, filt4_global[16]) = {
+  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
+};
+
+// These are reused by the avx2 intrinsics.
+filter8_1dfunction aom_filter_block1d8_v8_intrin_ssse3;
+filter8_1dfunction aom_filter_block1d8_h8_intrin_ssse3;
+filter8_1dfunction aom_filter_block1d4_h8_intrin_ssse3;
+
+void aom_filter_block1d4_h8_intrin_ssse3(
+    const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr,
+    ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) {
+  __m128i firstFilters, secondFilters, shuffle1, shuffle2;
+  __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4;
+  __m128i addFilterReg64, filtersReg, srcReg, minReg;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits in the filter into the first lane
+  firstFilters = _mm_shufflelo_epi16(filtersReg, 0);
+  // duplicate only the third 16 bit in the filter into the first lane
+  secondFilters = _mm_shufflelo_epi16(filtersReg, 0xAAu);
+  // duplicate only the seconds 16 bits in the filter into the second lane
+  // firstFilters: k0 k1 k0 k1 k0 k1 k0 k1 k2 k3 k2 k3 k2 k3 k2 k3
+  firstFilters = _mm_shufflehi_epi16(firstFilters, 0x55u);
+  // duplicate only the forth 16 bits in the filter into the second lane
+  // secondFilters: k4 k5 k4 k5 k4 k5 k4 k5 k6 k7 k6 k7 k6 k7 k6 k7
+  secondFilters = _mm_shufflehi_epi16(secondFilters, 0xFFu);
+
+  // loading the local filters
+  shuffle1 = _mm_load_si128((__m128i const *)filt1_4_h8);
+  shuffle2 = _mm_load_si128((__m128i const *)filt2_4_h8);
+
+  for (i = 0; i < output_height; i++) {
+    srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+
+    // filter the source buffer
+    srcRegFilt1 = _mm_shuffle_epi8(srcReg, shuffle1);
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg, shuffle2);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
+
+    // extract the higher half of the lane
+    srcRegFilt3 = _mm_srli_si128(srcRegFilt1, 8);
+    srcRegFilt4 = _mm_srli_si128(srcRegFilt2, 8);
+
+    minReg = _mm_min_epi16(srcRegFilt3, srcRegFilt2);
+
+    // add and saturate all the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+    srcRegFilt3 = _mm_max_epi16(srcRegFilt3, srcRegFilt2);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
+
+    // shift by 7 bit each 16 bits
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+
+    // shrink to 8 bit each 16 bits
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1);
+    src_ptr += src_pixels_per_line;
+
+    // save only 4 bytes
+    *((int *)&output_ptr[0]) = _mm_cvtsi128_si32(srcRegFilt1);
+
+    output_ptr += output_pitch;
+  }
+}
+
+void aom_filter_block1d8_h8_intrin_ssse3(
+    const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr,
+    ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) {
+  __m128i firstFilters, secondFilters, thirdFilters, forthFilters, srcReg;
+  __m128i filt1Reg, filt2Reg, filt3Reg, filt4Reg;
+  __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt4;
+  __m128i addFilterReg64, filtersReg, minReg;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 128 bit register
+  firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 128 bit register
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 128 bit register
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 128 bit register
+  forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u));
+
+  filt1Reg = _mm_load_si128((__m128i const *)filt1_global);
+  filt2Reg = _mm_load_si128((__m128i const *)filt2_global);
+  filt3Reg = _mm_load_si128((__m128i const *)filt3_global);
+  filt4Reg = _mm_load_si128((__m128i const *)filt4_global);
+
+  for (i = 0; i < output_height; i++) {
+    srcReg = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+
+    // filter the source buffer
+    srcRegFilt1 = _mm_shuffle_epi8(srcReg, filt1Reg);
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg, filt2Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, secondFilters);
+
+    // filter the source buffer
+    srcRegFilt3 = _mm_shuffle_epi8(srcReg, filt3Reg);
+    srcRegFilt4 = _mm_shuffle_epi8(srcReg, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, thirdFilters);
+    srcRegFilt4 = _mm_maddubs_epi16(srcRegFilt4, forthFilters);
+
+    // add and saturate all the results together
+    minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+
+    srcRegFilt2 = _mm_max_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
+
+    // shift by 7 bit each 16 bits
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+
+    // shrink to 8 bit each 16 bits
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1);
+
+    src_ptr += src_pixels_per_line;
+
+    // save only 8 bytes
+    _mm_storel_epi64((__m128i *)&output_ptr[0], srcRegFilt1);
+
+    output_ptr += output_pitch;
+  }
+}
+
+void aom_filter_block1d8_v8_intrin_ssse3(
+    const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr,
+    ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) {
+  __m128i addFilterReg64, filtersReg, minReg;
+  __m128i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m128i srcRegFilt1, srcRegFilt2, srcRegFilt3, srcRegFilt5;
+  __m128i srcReg1, srcReg2, srcReg3, srcReg4, srcReg5, srcReg6, srcReg7;
+  __m128i srcReg8;
+  unsigned int i;
+
+  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
+  addFilterReg64 = _mm_set1_epi32((int)0x0400040u);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
+
+  // duplicate only the first 16 bits in the filter
+  firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u));
+  // duplicate only the second 16 bits in the filter
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits in the filter
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits in the filter
+  forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u));
+
+  // load the first 7 rows of 8 bytes
+  srcReg1 = _mm_loadl_epi64((const __m128i *)src_ptr);
+  srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch));
+  srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
+  srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
+  srcReg7 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
+
+  for (i = 0; i < output_height; i++) {
+    // load the last 8 bytes
+    srcReg8 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7));
+
+    // merge the result together
+    srcRegFilt1 = _mm_unpacklo_epi8(srcReg1, srcReg2);
+    srcRegFilt3 = _mm_unpacklo_epi8(srcReg3, srcReg4);
+
+    // merge the result together
+    srcRegFilt2 = _mm_unpacklo_epi8(srcReg5, srcReg6);
+    srcRegFilt5 = _mm_unpacklo_epi8(srcReg7, srcReg8);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
+    srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
+    srcRegFilt2 = _mm_maddubs_epi16(srcRegFilt2, thirdFilters);
+    srcRegFilt5 = _mm_maddubs_epi16(srcRegFilt5, forthFilters);
+
+    // add and saturate the results together
+    minReg = _mm_min_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt5);
+    srcRegFilt2 = _mm_max_epi16(srcRegFilt2, srcRegFilt3);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, minReg);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt2);
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
+
+    // shift by 7 bit each 16 bit
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
+
+    // shrink to 8 bit each 16 bits
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt1);
+
+    src_ptr += src_pitch;
+
+    // shift down a row
+    srcReg1 = srcReg2;
+    srcReg2 = srcReg3;
+    srcReg3 = srcReg4;
+    srcReg4 = srcReg5;
+    srcReg5 = srcReg6;
+    srcReg6 = srcReg7;
+    srcReg7 = srcReg8;
+
+    // save only 8 bytes convolve result
+    _mm_storel_epi64((__m128i *)&output_ptr[0], srcRegFilt1);
+
+    output_ptr += out_pitch;
+  }
+}
+
+filter8_1dfunction aom_filter_block1d16_v8_ssse3;
+filter8_1dfunction aom_filter_block1d16_h8_ssse3;
+filter8_1dfunction aom_filter_block1d8_v8_ssse3;
+filter8_1dfunction aom_filter_block1d8_h8_ssse3;
+filter8_1dfunction aom_filter_block1d4_v8_ssse3;
+filter8_1dfunction aom_filter_block1d4_h8_ssse3;
+filter8_1dfunction aom_filter_block1d16_v8_avg_ssse3;
+filter8_1dfunction aom_filter_block1d16_h8_avg_ssse3;
+filter8_1dfunction aom_filter_block1d8_v8_avg_ssse3;
+filter8_1dfunction aom_filter_block1d8_h8_avg_ssse3;
+filter8_1dfunction aom_filter_block1d4_v8_avg_ssse3;
+filter8_1dfunction aom_filter_block1d4_h8_avg_ssse3;
+#if CONFIG_LOOP_RESTORATION
+filter8_1dfunction aom_filter_block1d16_v8_add_src_ssse3;
+filter8_1dfunction aom_filter_block1d16_h8_add_src_ssse3;
+filter8_1dfunction aom_filter_block1d8_v8_add_src_ssse3;
+filter8_1dfunction aom_filter_block1d8_h8_add_src_ssse3;
+filter8_1dfunction aom_filter_block1d4_v8_add_src_ssse3;
+filter8_1dfunction aom_filter_block1d4_h8_add_src_ssse3;
+#endif
+
+filter8_1dfunction aom_filter_block1d16_v2_ssse3;
+filter8_1dfunction aom_filter_block1d16_h2_ssse3;
+filter8_1dfunction aom_filter_block1d8_v2_ssse3;
+filter8_1dfunction aom_filter_block1d8_h2_ssse3;
+filter8_1dfunction aom_filter_block1d4_v2_ssse3;
+filter8_1dfunction aom_filter_block1d4_h2_ssse3;
+filter8_1dfunction aom_filter_block1d16_v2_avg_ssse3;
+filter8_1dfunction aom_filter_block1d16_h2_avg_ssse3;
+filter8_1dfunction aom_filter_block1d8_v2_avg_ssse3;
+filter8_1dfunction aom_filter_block1d8_h2_avg_ssse3;
+filter8_1dfunction aom_filter_block1d4_v2_avg_ssse3;
+filter8_1dfunction aom_filter_block1d4_h2_avg_ssse3;
+
+// void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h);
+// void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+// void aom_convolve8_avg_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                                    uint8_t *dst, ptrdiff_t dst_stride,
+//                                    const int16_t *filter_x, int x_step_q4,
+//                                    const int16_t *filter_y, int y_step_q4,
+//                                    int w, int h);
+// void aom_convolve8_avg_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                                   uint8_t *dst, ptrdiff_t dst_stride,
+//                                   const int16_t *filter_x, int x_step_q4,
+//                                   const int16_t *filter_y, int y_step_q4,
+//                                   int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , ssse3);
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , ssse3);
+FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, ssse3);
+FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_,
+            ssse3);
+
+#if CONFIG_LOOP_RESTORATION
+FUN_CONV_1D_NO_BILINEAR(add_src_horiz, x_step_q4, filter_x, h, src, add_src_,
+                        ssse3);
+FUN_CONV_1D_NO_BILINEAR(add_src_vert, y_step_q4, filter_y, v,
+                        src - src_stride * 3, add_src_, ssse3);
+#endif
+
+#define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                      out2, out3, out4, out5, out6, out7)                 \
+  {                                                                       \
+    const __m128i tr0_0 = _mm_unpacklo_epi8(in0, in1);                    \
+    const __m128i tr0_1 = _mm_unpacklo_epi8(in2, in3);                    \
+    const __m128i tr0_2 = _mm_unpacklo_epi8(in4, in5);                    \
+    const __m128i tr0_3 = _mm_unpacklo_epi8(in6, in7);                    \
+                                                                          \
+    const __m128i tr1_0 = _mm_unpacklo_epi16(tr0_0, tr0_1);               \
+    const __m128i tr1_1 = _mm_unpackhi_epi16(tr0_0, tr0_1);               \
+    const __m128i tr1_2 = _mm_unpacklo_epi16(tr0_2, tr0_3);               \
+    const __m128i tr1_3 = _mm_unpackhi_epi16(tr0_2, tr0_3);               \
+                                                                          \
+    const __m128i tr2_0 = _mm_unpacklo_epi32(tr1_0, tr1_2);               \
+    const __m128i tr2_1 = _mm_unpackhi_epi32(tr1_0, tr1_2);               \
+    const __m128i tr2_2 = _mm_unpacklo_epi32(tr1_1, tr1_3);               \
+    const __m128i tr2_3 = _mm_unpackhi_epi32(tr1_1, tr1_3);               \
+                                                                          \
+    out0 = _mm_unpacklo_epi64(tr2_0, tr2_0);                              \
+    out1 = _mm_unpackhi_epi64(tr2_0, tr2_0);                              \
+    out2 = _mm_unpacklo_epi64(tr2_1, tr2_1);                              \
+    out3 = _mm_unpackhi_epi64(tr2_1, tr2_1);                              \
+    out4 = _mm_unpacklo_epi64(tr2_2, tr2_2);                              \
+    out5 = _mm_unpackhi_epi64(tr2_2, tr2_2);                              \
+    out6 = _mm_unpacklo_epi64(tr2_3, tr2_3);                              \
+    out7 = _mm_unpackhi_epi64(tr2_3, tr2_3);                              \
+  }
+
+static void filter_horiz_w8_ssse3(const uint8_t *src_x, ptrdiff_t src_pitch,
+                                  uint8_t *dst, const int16_t *x_filter) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i f_values = _mm_load_si128((const __m128i *)x_filter);
+  // pack and duplicate the filter values
+  const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+  const __m128i A = _mm_loadl_epi64((const __m128i *)src_x);
+  const __m128i B = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch));
+  const __m128i C = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 2));
+  const __m128i D = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 3));
+  const __m128i E = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 4));
+  const __m128i F = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 5));
+  const __m128i G = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 6));
+  const __m128i H = _mm_loadl_epi64((const __m128i *)(src_x + src_pitch * 7));
+  // 00 01 10 11 02 03 12 13 04 05 14 15 06 07 16 17
+  const __m128i tr0_0 = _mm_unpacklo_epi16(A, B);
+  // 20 21 30 31 22 23 32 33 24 25 34 35 26 27 36 37
+  const __m128i tr0_1 = _mm_unpacklo_epi16(C, D);
+  // 40 41 50 51 42 43 52 53 44 45 54 55 46 47 56 57
+  const __m128i tr0_2 = _mm_unpacklo_epi16(E, F);
+  // 60 61 70 71 62 63 72 73 64 65 74 75 66 67 76 77
+  const __m128i tr0_3 = _mm_unpacklo_epi16(G, H);
+  // 00 01 10 11 20 21 30 31 02 03 12 13 22 23 32 33
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  // 04 05 14 15 24 25 34 35 06 07 16 17 26 27 36 37
+  const __m128i tr1_1 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  // 40 41 50 51 60 61 70 71 42 43 52 53 62 63 72 73
+  const __m128i tr1_2 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  // 44 45 54 55 64 65 74 75 46 47 56 57 66 67 76 77
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  // 00 01 10 11 20 21 30 31 40 41 50 51 60 61 70 71
+  const __m128i s1s0 = _mm_unpacklo_epi64(tr1_0, tr1_2);
+  const __m128i s3s2 = _mm_unpackhi_epi64(tr1_0, tr1_2);
+  const __m128i s5s4 = _mm_unpacklo_epi64(tr1_1, tr1_3);
+  const __m128i s7s6 = _mm_unpackhi_epi64(tr1_1, tr1_3);
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0);
+  const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2);
+  const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4);
+  const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6);
+  // add and saturate the results together
+  const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+  const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+  __m128i temp = _mm_adds_epi16(x0, x3);
+  temp = _mm_adds_epi16(temp, min_x2x1);
+  temp = _mm_adds_epi16(temp, max_x2x1);
+  // round and shift by 7 bit each 16 bit
+  temp = _mm_mulhrs_epi16(temp, k_256);
+  // shrink to 8 bit each 16 bits
+  temp = _mm_packus_epi16(temp, temp);
+  // save only 8 bytes convolve result
+  _mm_storel_epi64((__m128i *)dst, temp);
+}
+
+static void transpose8x8_to_dst(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride) {
+  __m128i A, B, C, D, E, F, G, H;
+
+  A = _mm_loadl_epi64((const __m128i *)src);
+  B = _mm_loadl_epi64((const __m128i *)(src + src_stride));
+  C = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2));
+  D = _mm_loadl_epi64((const __m128i *)(src + src_stride * 3));
+  E = _mm_loadl_epi64((const __m128i *)(src + src_stride * 4));
+  F = _mm_loadl_epi64((const __m128i *)(src + src_stride * 5));
+  G = _mm_loadl_epi64((const __m128i *)(src + src_stride * 6));
+  H = _mm_loadl_epi64((const __m128i *)(src + src_stride * 7));
+
+  TRANSPOSE_8X8(A, B, C, D, E, F, G, H, A, B, C, D, E, F, G, H);
+
+  _mm_storel_epi64((__m128i *)dst, A);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 1), B);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 2), C);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 3), D);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 4), E);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 5), F);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 6), G);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 7), H);
+}
+
+static void scaledconvolve_horiz_w8(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const InterpKernel *x_filters, int x0_q4,
+                                    int x_step_q4, int w, int h) {
+  DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]);
+  int x, y, z;
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  // This function processes 8x8 areas.  The intermediate height is not always
+  // a multiple of 8, so force it to be a multiple of 8 here.
+  y = h + (8 - (h & 0x7));
+
+  do {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; x += 8) {
+      // process 8 src_x steps
+      for (z = 0; z < 8; ++z) {
+        const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+        const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+        if (x_q4 & SUBPEL_MASK) {
+          filter_horiz_w8_ssse3(src_x, src_stride, temp + (z * 8), x_filter);
+        } else {
+          int i;
+          for (i = 0; i < 8; ++i) {
+            temp[z * 8 + i] = src_x[i * src_stride + 3];
+          }
+        }
+        x_q4 += x_step_q4;
+      }
+
+      // transpose the 8x8 filters values back to dst
+      transpose8x8_to_dst(temp, 8, dst + x, dst_stride);
+    }
+
+    src += src_stride * 8;
+    dst += dst_stride * 8;
+  } while (y -= 8);
+}
+
+static void filter_horiz_w4_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                  uint8_t *dst, const int16_t *filter) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+  // pack and duplicate the filter values
+  const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+  const __m128i A = _mm_loadl_epi64((const __m128i *)src_ptr);
+  const __m128i B = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch));
+  const __m128i C = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  const __m128i D = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  // TRANSPOSE...
+  // 00 01 02 03 04 05 06 07
+  // 10 11 12 13 14 15 16 17
+  // 20 21 22 23 24 25 26 27
+  // 30 31 32 33 34 35 36 37
+  //
+  // TO
+  //
+  // 00 10 20 30
+  // 01 11 21 31
+  // 02 12 22 32
+  // 03 13 23 33
+  // 04 14 24 34
+  // 05 15 25 35
+  // 06 16 26 36
+  // 07 17 27 37
+  //
+  // 00 01 10 11 02 03 12 13 04 05 14 15 06 07 16 17
+  const __m128i tr0_0 = _mm_unpacklo_epi16(A, B);
+  // 20 21 30 31 22 23 32 33 24 25 34 35 26 27 36 37
+  const __m128i tr0_1 = _mm_unpacklo_epi16(C, D);
+  // 00 01 10 11 20 21 30 31 02 03 12 13 22 23 32 33
+  const __m128i s1s0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  // 04 05 14 15 24 25 34 35 06 07 16 17 26 27 36 37
+  const __m128i s5s4 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  // 02 03 12 13 22 23 32 33
+  const __m128i s3s2 = _mm_srli_si128(s1s0, 8);
+  // 06 07 16 17 26 27 36 37
+  const __m128i s7s6 = _mm_srli_si128(s5s4, 8);
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0);
+  const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2);
+  const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4);
+  const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6);
+  // add and saturate the results together
+  const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+  const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+  __m128i temp = _mm_adds_epi16(x0, x3);
+  temp = _mm_adds_epi16(temp, min_x2x1);
+  temp = _mm_adds_epi16(temp, max_x2x1);
+  // round and shift by 7 bit each 16 bit
+  temp = _mm_mulhrs_epi16(temp, k_256);
+  // shrink to 8 bit each 16 bits
+  temp = _mm_packus_epi16(temp, temp);
+  // save only 4 bytes
+  *(int *)dst = _mm_cvtsi128_si32(temp);
+}
+
+static void transpose4x4_to_dst(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride) {
+  __m128i A = _mm_cvtsi32_si128(*(const int *)src);
+  __m128i B = _mm_cvtsi32_si128(*(const int *)(src + src_stride));
+  __m128i C = _mm_cvtsi32_si128(*(const int *)(src + src_stride * 2));
+  __m128i D = _mm_cvtsi32_si128(*(const int *)(src + src_stride * 3));
+  // 00 10 01 11 02 12 03 13
+  const __m128i tr0_0 = _mm_unpacklo_epi8(A, B);
+  // 20 30 21 31 22 32 23 33
+  const __m128i tr0_1 = _mm_unpacklo_epi8(C, D);
+  // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+  A = _mm_unpacklo_epi16(tr0_0, tr0_1);
+  B = _mm_srli_si128(A, 4);
+  C = _mm_srli_si128(A, 8);
+  D = _mm_srli_si128(A, 12);
+
+  *(int *)(dst) = _mm_cvtsi128_si32(A);
+  *(int *)(dst + dst_stride) = _mm_cvtsi128_si32(B);
+  *(int *)(dst + dst_stride * 2) = _mm_cvtsi128_si32(C);
+  *(int *)(dst + dst_stride * 3) = _mm_cvtsi128_si32(D);
+}
+
+static void scaledconvolve_horiz_w4(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const InterpKernel *x_filters, int x0_q4,
+                                    int x_step_q4, int w, int h) {
+  DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]);
+  int x, y, z;
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  for (y = 0; y < h; y += 4) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; x += 4) {
+      // process 4 src_x steps
+      for (z = 0; z < 4; ++z) {
+        const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+        const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+        if (x_q4 & SUBPEL_MASK) {
+          filter_horiz_w4_ssse3(src_x, src_stride, temp + (z * 4), x_filter);
+        } else {
+          int i;
+          for (i = 0; i < 4; ++i) {
+            temp[z * 4 + i] = src_x[i * src_stride + 3];
+          }
+        }
+        x_q4 += x_step_q4;
+      }
+
+      // transpose the 4x4 filters values back to dst
+      transpose4x4_to_dst(temp, 4, dst + x, dst_stride);
+    }
+
+    src += src_stride * 4;
+    dst += dst_stride * 4;
+  }
+}
+
+static void filter_vert_w4_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                 uint8_t *dst, const int16_t *filter) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+  // pack and duplicate the filter values
+  const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+  const __m128i A = _mm_cvtsi32_si128(*(const int *)src_ptr);
+  const __m128i B = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch));
+  const __m128i C = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 2));
+  const __m128i D = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 3));
+  const __m128i E = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 4));
+  const __m128i F = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 5));
+  const __m128i G = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 6));
+  const __m128i H = _mm_cvtsi32_si128(*(const int *)(src_ptr + src_pitch * 7));
+  const __m128i s1s0 = _mm_unpacklo_epi8(A, B);
+  const __m128i s3s2 = _mm_unpacklo_epi8(C, D);
+  const __m128i s5s4 = _mm_unpacklo_epi8(E, F);
+  const __m128i s7s6 = _mm_unpacklo_epi8(G, H);
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0);
+  const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2);
+  const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4);
+  const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6);
+  // add and saturate the results together
+  const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+  const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+  __m128i temp = _mm_adds_epi16(x0, x3);
+  temp = _mm_adds_epi16(temp, min_x2x1);
+  temp = _mm_adds_epi16(temp, max_x2x1);
+  // round and shift by 7 bit each 16 bit
+  temp = _mm_mulhrs_epi16(temp, k_256);
+  // shrink to 8 bit each 16 bits
+  temp = _mm_packus_epi16(temp, temp);
+  // save only 4 bytes
+  *(int *)dst = _mm_cvtsi128_si32(temp);
+}
+
+static void scaledconvolve_vert_w4(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const InterpKernel *y_filters, int y0_q4,
+                                   int y_step_q4, int w, int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (y = 0; y < h; ++y) {
+    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w4_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter);
+    } else {
+      memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+    }
+
+    y_q4 += y_step_q4;
+  }
+}
+
+static void filter_vert_w8_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                 uint8_t *dst, const int16_t *filter) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+  // pack and duplicate the filter values
+  const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+  const __m128i A = _mm_loadl_epi64((const __m128i *)src_ptr);
+  const __m128i B = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch));
+  const __m128i C = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  const __m128i D = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  const __m128i E = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
+  const __m128i F = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
+  const __m128i G = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
+  const __m128i H = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7));
+  const __m128i s1s0 = _mm_unpacklo_epi8(A, B);
+  const __m128i s3s2 = _mm_unpacklo_epi8(C, D);
+  const __m128i s5s4 = _mm_unpacklo_epi8(E, F);
+  const __m128i s7s6 = _mm_unpacklo_epi8(G, H);
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m128i x0 = _mm_maddubs_epi16(s1s0, f1f0);
+  const __m128i x1 = _mm_maddubs_epi16(s3s2, f3f2);
+  const __m128i x2 = _mm_maddubs_epi16(s5s4, f5f4);
+  const __m128i x3 = _mm_maddubs_epi16(s7s6, f7f6);
+  // add and saturate the results together
+  const __m128i min_x2x1 = _mm_min_epi16(x2, x1);
+  const __m128i max_x2x1 = _mm_max_epi16(x2, x1);
+  __m128i temp = _mm_adds_epi16(x0, x3);
+  temp = _mm_adds_epi16(temp, min_x2x1);
+  temp = _mm_adds_epi16(temp, max_x2x1);
+  // round and shift by 7 bit each 16 bit
+  temp = _mm_mulhrs_epi16(temp, k_256);
+  // shrink to 8 bit each 16 bits
+  temp = _mm_packus_epi16(temp, temp);
+  // save only 8 bytes convolve result
+  _mm_storel_epi64((__m128i *)dst, temp);
+}
+
+static void scaledconvolve_vert_w8(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const InterpKernel *y_filters, int y0_q4,
+                                   int y_step_q4, int w, int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (y = 0; y < h; ++y) {
+    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w8_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter);
+    } else {
+      memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+    }
+    y_q4 += y_step_q4;
+  }
+}
+
+static void filter_vert_w16_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                  uint8_t *dst, const int16_t *filter, int w) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+  // pack and duplicate the filter values
+  const __m128i f1f0 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  const __m128i f3f2 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  const __m128i f5f4 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  const __m128i f7f6 = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+  int i;
+
+  for (i = 0; i < w; i += 16) {
+    const __m128i A = _mm_loadu_si128((const __m128i *)src_ptr);
+    const __m128i B = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch));
+    const __m128i C =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
+    const __m128i D =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
+    const __m128i E =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4));
+    const __m128i F =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5));
+    const __m128i G =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6));
+    const __m128i H =
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+    // merge the result together
+    const __m128i s1s0_lo = _mm_unpacklo_epi8(A, B);
+    const __m128i s7s6_lo = _mm_unpacklo_epi8(G, H);
+    const __m128i s1s0_hi = _mm_unpackhi_epi8(A, B);
+    const __m128i s7s6_hi = _mm_unpackhi_epi8(G, H);
+    // multiply 2 adjacent elements with the filter and add the result
+    const __m128i x0_lo = _mm_maddubs_epi16(s1s0_lo, f1f0);
+    const __m128i x3_lo = _mm_maddubs_epi16(s7s6_lo, f7f6);
+    const __m128i x0_hi = _mm_maddubs_epi16(s1s0_hi, f1f0);
+    const __m128i x3_hi = _mm_maddubs_epi16(s7s6_hi, f7f6);
+    // add and saturate the results together
+    const __m128i x3x0_lo = _mm_adds_epi16(x0_lo, x3_lo);
+    const __m128i x3x0_hi = _mm_adds_epi16(x0_hi, x3_hi);
+    // merge the result together
+    const __m128i s3s2_lo = _mm_unpacklo_epi8(C, D);
+    const __m128i s3s2_hi = _mm_unpackhi_epi8(C, D);
+    // multiply 2 adjacent elements with the filter and add the result
+    const __m128i x1_lo = _mm_maddubs_epi16(s3s2_lo, f3f2);
+    const __m128i x1_hi = _mm_maddubs_epi16(s3s2_hi, f3f2);
+    // merge the result together
+    const __m128i s5s4_lo = _mm_unpacklo_epi8(E, F);
+    const __m128i s5s4_hi = _mm_unpackhi_epi8(E, F);
+    // multiply 2 adjacent elements with the filter and add the result
+    const __m128i x2_lo = _mm_maddubs_epi16(s5s4_lo, f5f4);
+    const __m128i x2_hi = _mm_maddubs_epi16(s5s4_hi, f5f4);
+    // add and saturate the results together
+    __m128i temp_lo = _mm_adds_epi16(x3x0_lo, _mm_min_epi16(x1_lo, x2_lo));
+    __m128i temp_hi = _mm_adds_epi16(x3x0_hi, _mm_min_epi16(x1_hi, x2_hi));
+
+    // add and saturate the results together
+    temp_lo = _mm_adds_epi16(temp_lo, _mm_max_epi16(x1_lo, x2_lo));
+    temp_hi = _mm_adds_epi16(temp_hi, _mm_max_epi16(x1_hi, x2_hi));
+    // round and shift by 7 bit each 16 bit
+    temp_lo = _mm_mulhrs_epi16(temp_lo, k_256);
+    temp_hi = _mm_mulhrs_epi16(temp_hi, k_256);
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    temp_hi = _mm_packus_epi16(temp_lo, temp_hi);
+    src_ptr += 16;
+    // save 16 bytes convolve result
+    _mm_store_si128((__m128i *)&dst[i], temp_hi);
+  }
+}
+
+static void scaledconvolve_vert_w16(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const InterpKernel *y_filters, int y0_q4,
+                                    int y_step_q4, int w, int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (y = 0; y < h; ++y) {
+    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w16_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter,
+                            w);
+    } else {
+      memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+    }
+    y_q4 += y_step_q4;
+  }
+}
+
+static void scaledconvolve2d(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const InterpKernel *const x_filters, int x0_q4,
+                             int x_step_q4, const InterpKernel *const y_filters,
+                             int y0_q4, int y_step_q4, int w, int h) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  // --Require an additional 8 rows for the horiz_w8 transpose tail.
+  DECLARE_ALIGNED(16, uint8_t, temp[(MAX_EXT_SIZE + 8) * MAX_SB_SIZE]);
+  const int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  if (w >= 8) {
+    scaledconvolve_horiz_w8(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                            src_stride, temp, MAX_SB_SIZE, x_filters, x0_q4,
+                            x_step_q4, w, intermediate_height);
+  } else {
+    scaledconvolve_horiz_w4(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                            src_stride, temp, MAX_SB_SIZE, x_filters, x0_q4,
+                            x_step_q4, w, intermediate_height);
+  }
+
+  if (w >= 16) {
+    scaledconvolve_vert_w16(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
+                            MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4,
+                            y_step_q4, w, h);
+  } else if (w == 8) {
+    scaledconvolve_vert_w8(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
+                           MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4,
+                           y_step_q4, w, h);
+  } else {
+    scaledconvolve_vert_w4(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
+                           MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4,
+                           y_step_q4, w, h);
+  }
+}
+
+static const InterpKernel *get_filter_base(const int16_t *filter) {
+  // NOTE: This assumes that the filter table is 256-byte aligned.
+  // TODO(agrange) Modify to make independent of table alignment.
+  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
+}
+
+static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
+  return (int)((const InterpKernel *)(intptr_t)f - base);
+}
+
+void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const int16_t *filter_x,
+                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
+                         int w, int h) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  scaledconvolve2d(src, src_stride, dst, dst_stride, filters_x, x0_q4,
+                   x_step_q4, filters_y, y0_q4, y_step_q4, w, h);
+}
+
+// void aom_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                          uint8_t *dst, ptrdiff_t dst_stride,
+//                          const int16_t *filter_x, int x_step_q4,
+//                          const int16_t *filter_y, int y_step_q4,
+//                          int w, int h);
+// void aom_convolve8_avg_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                              uint8_t *dst, ptrdiff_t dst_stride,
+//                              const int16_t *filter_x, int x_step_q4,
+//                              const int16_t *filter_y, int y_step_q4,
+//                              int w, int h);
+FUN_CONV_2D(, ssse3);
+FUN_CONV_2D(avg_, ssse3);
+#if CONFIG_LOOP_RESTORATION
+FUN_CONV_2D_NO_BILINEAR(add_src_, add_src_, ssse3);
+#endif
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm
new file mode 100644
index 0000000000..b946010d32
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm
@@ -0,0 +1,990 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+%include "aom_ports/x86_abi_support.asm"
+
+;Note: tap3 and tap4 have to be applied and added after other taps to avoid
+;overflow.
+
+%macro GET_FILTERS_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    psrldq      xmm7, 8
+    pshuflw     xmm4, xmm7, 0b              ;k4
+    pshuflw     xmm5, xmm7, 01010101b       ;k5
+    pshuflw     xmm6, xmm7, 10101010b       ;k6
+    pshuflw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklqdq  xmm0, xmm1
+    punpcklqdq  xmm2, xmm3
+    punpcklqdq  xmm5, xmm4
+    punpcklqdq  xmm6, xmm7
+
+    movdqa      k0k1, xmm0
+    movdqa      k2k3, xmm2
+    movdqa      k5k4, xmm5
+    movdqa      k6k7, xmm6
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6
+
+    pxor        xmm7, xmm7
+    movdqa      zero, xmm7
+%endm
+
+%macro APPLY_FILTER_4 1
+    punpckldq   xmm0, xmm1                  ;two row in one register
+    punpckldq   xmm6, xmm7
+    punpckldq   xmm2, xmm3
+    punpckldq   xmm5, xmm4
+
+    punpcklbw   xmm0, zero                  ;unpack to word
+    punpcklbw   xmm6, zero
+    punpcklbw   xmm2, zero
+    punpcklbw   xmm5, zero
+
+    pmullw      xmm0, k0k1                  ;multiply the filter factors
+    pmullw      xmm6, k6k7
+    pmullw      xmm2, k2k3
+    pmullw      xmm5, k5k4
+
+    paddsw      xmm0, xmm6                  ;sum
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 8
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm2
+    psrldq      xmm2, 8
+    paddsw      xmm0, xmm5
+    psrldq      xmm5, 8
+    paddsw      xmm0, xmm2
+    paddsw      xmm0, xmm5
+
+    paddsw      xmm0, krd                   ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movd        [rdi], xmm0
+%endm
+
+%macro GET_FILTERS 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    pshufhw     xmm4, xmm7, 0b              ;k4
+    pshufhw     xmm5, xmm7, 01010101b       ;k5
+    pshufhw     xmm6, xmm7, 10101010b       ;k6
+    pshufhw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklwd   xmm0, xmm0
+    punpcklwd   xmm1, xmm1
+    punpcklwd   xmm2, xmm2
+    punpcklwd   xmm3, xmm3
+    punpckhwd   xmm4, xmm4
+    punpckhwd   xmm5, xmm5
+    punpckhwd   xmm6, xmm6
+    punpckhwd   xmm7, xmm7
+
+    movdqa      k0,   xmm0                  ;store filter factors on stack
+    movdqa      k1,   xmm1
+    movdqa      k2,   xmm2
+    movdqa      k3,   xmm3
+    movdqa      k4,   xmm4
+    movdqa      k5,   xmm5
+    movdqa      k6,   xmm6
+    movdqa      k7,   xmm7
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6                   ;rounding
+
+    pxor        xmm7, xmm7
+    movdqa      zero, xmm7
+%endm
+
+%macro LOAD_VERT_8 1
+    movq        xmm0, [rsi + %1]            ;0
+    movq        xmm1, [rsi + rax + %1]      ;1
+    movq        xmm6, [rsi + rdx * 2 + %1]  ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2 + %1]  ;7
+    movq        xmm2, [rsi + rax + %1]      ;2
+    movq        xmm3, [rsi + rax * 2 + %1]  ;3
+    movq        xmm4, [rsi + rdx + %1]      ;4
+    movq        xmm5, [rsi + rax * 4 + %1]  ;5
+%endm
+
+%macro APPLY_FILTER_8 2
+    punpcklbw   xmm0, zero
+    punpcklbw   xmm1, zero
+    punpcklbw   xmm6, zero
+    punpcklbw   xmm7, zero
+    punpcklbw   xmm2, zero
+    punpcklbw   xmm5, zero
+    punpcklbw   xmm3, zero
+    punpcklbw   xmm4, zero
+
+    pmullw      xmm0, k0
+    pmullw      xmm1, k1
+    pmullw      xmm6, k6
+    pmullw      xmm7, k7
+    pmullw      xmm2, k2
+    pmullw      xmm5, k5
+    pmullw      xmm3, k3
+    pmullw      xmm4, k4
+
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm6
+    paddsw      xmm0, xmm7
+    paddsw      xmm0, xmm2
+    paddsw      xmm0, xmm5
+    paddsw      xmm0, xmm3
+    paddsw      xmm0, xmm4
+
+    paddsw      xmm0, krd                   ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack back to byte
+%if %1
+    movq        xmm1, [rdi + %2]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi + %2], xmm0
+%endm
+
+;void aom_filter_block1d4_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(aom_filter_block1d4_v8_sse2) PRIVATE
+sym(aom_filter_block1d4_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movd        xmm0, [rsi]                 ;load src: row 0
+    movd        xmm1, [rsi + rax]           ;1
+    movd        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movd        xmm7, [rsi + rdx * 2]       ;7
+    movd        xmm2, [rsi + rax]           ;2
+    movd        xmm3, [rsi + rax * 2]       ;3
+    movd        xmm4, [rsi + rdx]           ;4
+    movd        xmm5, [rsi + rax * 4]       ;5
+
+    APPLY_FILTER_4 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d8_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(aom_filter_block1d8_v8_sse2) PRIVATE
+sym(aom_filter_block1d8_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 0, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d16_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+global sym(aom_filter_block1d16_v8_sse2) PRIVATE
+sym(aom_filter_block1d16_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 0, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 8
+    APPLY_FILTER_8 0, 8
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d4_v8_avg_sse2) PRIVATE
+sym(aom_filter_block1d4_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movd        xmm0, [rsi]                 ;load src: row 0
+    movd        xmm1, [rsi + rax]           ;1
+    movd        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movd        xmm7, [rsi + rdx * 2]       ;7
+    movd        xmm2, [rsi + rax]           ;2
+    movd        xmm3, [rsi + rax * 2]       ;3
+    movd        xmm4, [rsi + rdx]           ;4
+    movd        xmm5, [rsi + rax * 4]       ;5
+
+    APPLY_FILTER_4 1
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d8_v8_avg_sse2) PRIVATE
+sym(aom_filter_block1d8_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 1, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d16_v8_avg_sse2) PRIVATE
+sym(aom_filter_block1d16_v8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 1, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 8
+    APPLY_FILTER_8 1, 8
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d4_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(aom_filter_block1d4_h8_sse2) PRIVATE
+sym(aom_filter_block1d4_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm3, 3
+    psrldq      xmm5, 5
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_4 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d8_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(aom_filter_block1d8_h8_sse2) PRIVATE
+sym(aom_filter_block1d8_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d16_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+global sym(aom_filter_block1d16_h8_sse2) PRIVATE
+sym(aom_filter_block1d16_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 0
+
+    movdqu      xmm0,   [rsi + 5]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 8
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d4_h8_avg_sse2) PRIVATE
+sym(aom_filter_block1d4_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm3, 3
+    psrldq      xmm5, 5
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_4 1
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d8_h8_avg_sse2) PRIVATE
+sym(aom_filter_block1d8_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 1, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d16_h8_avg_sse2) PRIVATE
+sym(aom_filter_block1d16_h8_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 1, 0
+
+    movdqu      xmm0,   [rsi + 5]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 1, 8
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm
new file mode 100644
index 0000000000..357f374016
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm
@@ -0,0 +1,883 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_64:    times 8 dw 64
+even_byte_mask: times 8 dw 0x00ff
+
+; %define USE_PMULHRSW
+; NOTE: pmulhrsw has a latency of 5 cycles.  Tests showed a performance loss
+; when using this instruction.
+;
+; The add order below (based on ffav1) must be followed to prevent outranges.
+; x = k0k1 + k4k5
+; y = k2k3 + k6k7
+; z = signed SAT(x + y)
+
+SECTION .text
+%define LOCAL_VARS_SIZE 16*6
+
+%macro SETUP_LOCAL_VARS 0
+    ; TODO(slavarnway): using xmm registers for these on ARCH_X86_64 +
+    ; pmaddubsw has a higher latency on some platforms, this might be eased by
+    ; interleaving the instructions.
+    %define    k0k1  [rsp + 16*0]
+    %define    k2k3  [rsp + 16*1]
+    %define    k4k5  [rsp + 16*2]
+    %define    k6k7  [rsp + 16*3]
+    packsswb     m4, m4
+    ; TODO(slavarnway): multiple pshufb instructions had a higher latency on
+    ; some platforms.
+    pshuflw      m0, m4, 0b              ;k0_k1
+    pshuflw      m1, m4, 01010101b       ;k2_k3
+    pshuflw      m2, m4, 10101010b       ;k4_k5
+    pshuflw      m3, m4, 11111111b       ;k6_k7
+    punpcklqdq   m0, m0
+    punpcklqdq   m1, m1
+    punpcklqdq   m2, m2
+    punpcklqdq   m3, m3
+    mova       k0k1, m0
+    mova       k2k3, m1
+    mova       k4k5, m2
+    mova       k6k7, m3
+%if ARCH_X86_64
+    %define     krd  m12
+    %define    tmp0  [rsp + 16*4]
+    %define    tmp1  [rsp + 16*5]
+    mova        krd, [GLOBAL(pw_64)]
+%else
+    %define     krd  [rsp + 16*4]
+%if CONFIG_PIC=0
+    mova         m6, [GLOBAL(pw_64)]
+%else
+    ; build constants without accessing global memory
+    pcmpeqb      m6, m6                  ;all ones
+    psrlw        m6, 15
+    psllw        m6, 6                   ;aka pw_64
+%endif
+    mova        krd, m6
+%endif
+%endm
+
+;-------------------------------------------------------------------------------
+%if ARCH_X86_64
+  %define LOCAL_VARS_SIZE_H4 0
+%else
+  %define LOCAL_VARS_SIZE_H4 16*4
+%endif
+
+%macro SUBPIX_HFILTER4 1
+cglobal filter_block1d4_%1, 6, 6, 11, LOCAL_VARS_SIZE_H4, \
+                            src, sstride, dst, dstride, height, filter
+    mova                m4, [filterq]
+    packsswb            m4, m4
+%if ARCH_X86_64
+    %define       k0k1k4k5  m8
+    %define       k2k3k6k7  m9
+    %define            krd  m10
+    mova               krd, [GLOBAL(pw_64)]
+    pshuflw       k0k1k4k5, m4, 0b              ;k0_k1
+    pshufhw       k0k1k4k5, k0k1k4k5, 10101010b ;k0_k1_k4_k5
+    pshuflw       k2k3k6k7, m4, 01010101b       ;k2_k3
+    pshufhw       k2k3k6k7, k2k3k6k7, 11111111b ;k2_k3_k6_k7
+%else
+    %define       k0k1k4k5  [rsp + 16*0]
+    %define       k2k3k6k7  [rsp + 16*1]
+    %define            krd  [rsp + 16*2]
+    pshuflw             m6, m4, 0b              ;k0_k1
+    pshufhw             m6, m6, 10101010b       ;k0_k1_k4_k5
+    pshuflw             m7, m4, 01010101b       ;k2_k3
+    pshufhw             m7, m7, 11111111b       ;k2_k3_k6_k7
+%if CONFIG_PIC=0
+    mova                m1, [GLOBAL(pw_64)]
+%else
+    ; build constants without accessing global memory
+    pcmpeqb             m1, m1                  ;all ones
+    psrlw               m1, 15
+    psllw               m1, 6                   ;aka pw_64
+%endif
+    mova          k0k1k4k5, m6
+    mova          k2k3k6k7, m7
+    mova               krd, m1
+%endif
+    dec            heightd
+
+.loop:
+    ;Do two rows at once
+    movu                m4, [srcq - 3]
+    movu                m5, [srcq + sstrideq - 3]
+    punpckhbw           m1, m4, m4
+    punpcklbw           m4, m4
+    punpckhbw           m3, m5, m5
+    punpcklbw           m5, m5
+    palignr             m0, m1, m4, 1
+    pmaddubsw           m0, k0k1k4k5
+    palignr             m1, m4, 5
+    pmaddubsw           m1, k2k3k6k7
+    palignr             m2, m3, m5, 1
+    pmaddubsw           m2, k0k1k4k5
+    palignr             m3, m5, 5
+    pmaddubsw           m3, k2k3k6k7
+    punpckhqdq          m4, m0, m2
+    punpcklqdq          m0, m2
+    punpckhqdq          m5, m1, m3
+    punpcklqdq          m1, m3
+    paddsw              m0, m4
+    paddsw              m1, m5
+%ifidn %1, h8_avg
+    movd                m4, [dstq]
+    movd                m5, [dstq + dstrideq]
+%endif
+    paddsw              m0, m1
+    paddsw              m0, krd
+    psraw               m0, 7
+%ifidn %1, h8_add_src
+    pxor                 m3, m3
+    movu                 m4, [srcq]
+    movu                 m5, [srcq + sstrideq]
+    punpckldq            m4, m5 ; Bytes 0,1,2,3 from row 0, then 0,1,2,3 from row 2
+    punpcklbw            m4, m3
+    paddsw               m0, m4
+%endif
+    packuswb            m0, m0
+    psrldq              m1, m0, 4
+
+%ifidn %1, h8_avg
+    pavgb               m0, m4
+    pavgb               m1, m5
+%endif
+    movd            [dstq], m0
+    movd [dstq + dstrideq], m1
+
+    lea               srcq, [srcq + sstrideq        ]
+    prefetcht0              [srcq + 4 * sstrideq - 3]
+    lea               srcq, [srcq + sstrideq        ]
+    lea               dstq, [dstq + 2 * dstrideq    ]
+    prefetcht0              [srcq + 2 * sstrideq - 3]
+
+    sub            heightd, 2
+    jg               .loop
+
+    ; Do last row if output_height is odd
+    jne              .done
+
+    movu                m4, [srcq - 3]
+    punpckhbw           m1, m4, m4
+    punpcklbw           m4, m4
+    palignr             m0, m1, m4, 1
+    palignr             m1, m4, 5
+    pmaddubsw           m0, k0k1k4k5
+    pmaddubsw           m1, k2k3k6k7
+    psrldq              m2, m0, 8
+    psrldq              m3, m1, 8
+    paddsw              m0, m2
+    paddsw              m1, m3
+    paddsw              m0, m1
+    paddsw              m0, krd
+    psraw               m0, 7
+%ifidn %1, h8_add_src
+    pxor                m3, m3
+    movu                m4, [srcq]
+    punpcklbw           m4, m3
+    paddsw              m0, m4
+%endif
+    packuswb            m0, m0
+%ifidn %1, h8_avg
+    movd                m4, [dstq]
+    pavgb               m0, m4
+%endif
+    movd            [dstq], m0
+.done:
+    REP_RET
+%endm
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_HFILTER8 1
+cglobal filter_block1d8_%1, 6, 6, 14, LOCAL_VARS_SIZE, \
+                            src, sstride, dst, dstride, height, filter
+    mova                 m4, [filterq]
+    SETUP_LOCAL_VARS
+    dec             heightd
+
+.loop:
+    ;Do two rows at once
+    movu                 m0, [srcq - 3]
+    movu                 m4, [srcq + sstrideq - 3]
+    punpckhbw            m1, m0, m0
+    punpcklbw            m0, m0
+    palignr              m5, m1, m0, 13
+    pmaddubsw            m5, k6k7
+    palignr              m2, m1, m0, 5
+    palignr              m3, m1, m0, 9
+    palignr              m1, m0, 1
+    pmaddubsw            m1, k0k1
+    punpckhbw            m6, m4, m4
+    punpcklbw            m4, m4
+    pmaddubsw            m2, k2k3
+    pmaddubsw            m3, k4k5
+
+    palignr              m7, m6, m4, 13
+    palignr              m0, m6, m4, 5
+    pmaddubsw            m7, k6k7
+    paddsw               m1, m3
+    paddsw               m2, m5
+    paddsw               m1, m2
+%ifidn %1, h8_avg
+    movh                 m2, [dstq]
+    movhps               m2, [dstq + dstrideq]
+%endif
+    palignr              m5, m6, m4, 9
+    palignr              m6, m4, 1
+    pmaddubsw            m0, k2k3
+    pmaddubsw            m6, k0k1
+    paddsw               m1, krd
+    pmaddubsw            m5, k4k5
+    psraw                m1, 7
+    paddsw               m0, m7
+    paddsw               m6, m5
+    paddsw               m6, m0
+    paddsw               m6, krd
+    psraw                m6, 7
+%ifidn %1, h8_add_src
+    pxor                 m3, m3
+    movu                 m4, [srcq]
+    movu                 m5, [srcq + sstrideq]
+    punpcklbw            m4, m3
+    punpcklbw            m5, m3
+    paddsw               m1, m4
+    paddsw               m6, m5
+%endif
+    packuswb             m1, m6
+%ifidn %1, h8_avg
+    pavgb                m1, m2
+%endif
+    movh              [dstq], m1
+    movhps [dstq + dstrideq], m1
+
+    lea                srcq, [srcq + sstrideq        ]
+    prefetcht0               [srcq + 4 * sstrideq - 3]
+    lea                srcq, [srcq + sstrideq        ]
+    lea                dstq, [dstq + 2 * dstrideq    ]
+    prefetcht0               [srcq + 2 * sstrideq - 3]
+    sub             heightd, 2
+    jg                .loop
+
+    ; Do last row if output_height is odd
+    jne               .done
+
+    movu                 m0, [srcq - 3]
+    punpckhbw            m3, m0, m0
+    punpcklbw            m0, m0
+    palignr              m1, m3, m0, 1
+    palignr              m2, m3, m0, 5
+    palignr              m4, m3, m0, 13
+    palignr              m3, m0, 9
+    pmaddubsw            m1, k0k1
+    pmaddubsw            m2, k2k3
+    pmaddubsw            m3, k4k5
+    pmaddubsw            m4, k6k7
+    paddsw               m1, m3
+    paddsw               m4, m2
+    paddsw               m1, m4
+    paddsw               m1, krd
+    psraw                m1, 7
+%ifidn %1, h8_add_src
+    pxor                 m6, m6
+    movu                 m5, [srcq]
+    punpcklbw            m5, m6
+    paddsw               m1, m5
+%endif
+    packuswb             m1, m1
+%ifidn %1, h8_avg
+    movh                 m0, [dstq]
+    pavgb                m1, m0
+%endif
+    movh             [dstq], m1
+.done:
+    REP_RET
+%endm
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_HFILTER16 1
+cglobal filter_block1d16_%1, 6, 6, 14, LOCAL_VARS_SIZE, \
+                             src, sstride, dst, dstride, height, filter
+    mova          m4, [filterq]
+    SETUP_LOCAL_VARS
+
+.loop:
+    prefetcht0        [srcq + 2 * sstrideq -3]
+
+    movu          m0, [srcq - 3]
+    movu          m4, [srcq - 2]
+    pmaddubsw     m0, k0k1
+    pmaddubsw     m4, k0k1
+    movu          m1, [srcq - 1]
+    movu          m5, [srcq + 0]
+    pmaddubsw     m1, k2k3
+    pmaddubsw     m5, k2k3
+    movu          m2, [srcq + 1]
+    movu          m6, [srcq + 2]
+    pmaddubsw     m2, k4k5
+    pmaddubsw     m6, k4k5
+    movu          m3, [srcq + 3]
+    movu          m7, [srcq + 4]
+    pmaddubsw     m3, k6k7
+    pmaddubsw     m7, k6k7
+    paddsw        m0, m2
+    paddsw        m1, m3
+    paddsw        m0, m1
+    paddsw        m4, m6
+    paddsw        m5, m7
+    paddsw        m4, m5
+    paddsw        m0, krd
+    paddsw        m4, krd
+    psraw         m0, 7
+    psraw         m4, 7
+%ifidn %1, h8_add_src
+    movu          m5, [srcq]
+    mova          m7, m5
+    pand          m5, [even_byte_mask]
+    psrlw         m7, 8
+    paddsw        m0, m5
+    paddsw        m4, m7
+%endif
+    packuswb      m0, m0
+    packuswb      m4, m4
+    punpcklbw     m0, m4
+%ifidn %1, h8_avg
+    pavgb         m0, [dstq]
+%endif
+    lea         srcq, [srcq + sstrideq]
+    mova      [dstq], m0
+    lea         dstq, [dstq + dstrideq]
+    dec      heightd
+    jnz        .loop
+    REP_RET
+%endm
+
+INIT_XMM ssse3
+SUBPIX_HFILTER16 h8
+SUBPIX_HFILTER16 h8_avg
+SUBPIX_HFILTER8  h8
+SUBPIX_HFILTER8  h8_avg
+SUBPIX_HFILTER4  h8
+SUBPIX_HFILTER4  h8_avg
+
+%if CONFIG_LOOP_RESTORATION
+SUBPIX_HFILTER16 h8_add_src
+SUBPIX_HFILTER8  h8_add_src
+SUBPIX_HFILTER4  h8_add_src
+%endif
+
+;-------------------------------------------------------------------------------
+
+; TODO(Linfeng): Detect cpu type and choose the code with better performance.
+%define X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON 1
+
+%if ARCH_X86_64 && X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON
+    %define NUM_GENERAL_REG_USED 9
+%else
+    %define NUM_GENERAL_REG_USED 6
+%endif
+
+%macro SUBPIX_VFILTER 2
+cglobal filter_block1d%2_%1, 6, NUM_GENERAL_REG_USED, 15, LOCAL_VARS_SIZE, \
+                             src, sstride, dst, dstride, height, filter
+    mova          m4, [filterq]
+    SETUP_LOCAL_VARS
+
+%ifidn %2, 8
+    %define                movx  movh
+%else
+    %define                movx  movd
+%endif
+
+    dec                 heightd
+
+%if ARCH_X86 || X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON
+
+%if ARCH_X86_64
+    %define               src1q  r7
+    %define           sstride6q  r8
+    %define          dst_stride  dstrideq
+%else
+    %define               src1q  filterq
+    %define           sstride6q  dstrideq
+    %define          dst_stride  dstridemp
+%endif
+    mov                   src1q, srcq
+    add                   src1q, sstrideq
+    lea               sstride6q, [sstrideq + sstrideq * 4]
+    add               sstride6q, sstrideq                   ;pitch * 6
+
+.loop:
+    ;Do two rows at once
+    movx                     m0, [srcq                ]     ;A
+    movx                     m1, [src1q               ]     ;B
+    punpcklbw                m0, m1                         ;A B
+    movx                     m2, [srcq + sstrideq * 2 ]     ;C
+    pmaddubsw                m0, k0k1
+    mova                     m6, m2
+    movx                     m3, [src1q + sstrideq * 2]     ;D
+    punpcklbw                m2, m3                         ;C D
+    pmaddubsw                m2, k2k3
+    movx                     m4, [srcq + sstrideq * 4 ]     ;E
+    mova                     m7, m4
+    movx                     m5, [src1q + sstrideq * 4]     ;F
+    punpcklbw                m4, m5                         ;E F
+    pmaddubsw                m4, k4k5
+    punpcklbw                m1, m6                         ;A B next iter
+    movx                     m6, [srcq + sstride6q    ]     ;G
+    punpcklbw                m5, m6                         ;E F next iter
+    punpcklbw                m3, m7                         ;C D next iter
+    pmaddubsw                m5, k4k5
+    movx                     m7, [src1q + sstride6q   ]     ;H
+    punpcklbw                m6, m7                         ;G H
+    pmaddubsw                m6, k6k7
+    pmaddubsw                m3, k2k3
+    pmaddubsw                m1, k0k1
+    paddsw                   m0, m4
+    paddsw                   m2, m6
+    movx                     m6, [srcq + sstrideq * 8 ]     ;H next iter
+    punpcklbw                m7, m6
+    pmaddubsw                m7, k6k7
+    paddsw                   m0, m2
+    paddsw                   m0, krd
+    psraw                    m0, 7
+    paddsw                   m1, m5
+%ifidn %1, v8_add_src
+    pxor                     m6, m6
+    movu                     m4, [srcq]
+    punpcklbw                m4, m6
+    paddsw                   m0, m4
+%endif
+    packuswb                 m0, m0
+
+    paddsw                   m3, m7
+    paddsw                   m1, m3
+    paddsw                   m1, krd
+    psraw                    m1, 7
+%ifidn %1, v8_add_src
+    movu                     m4, [src1q]
+    punpcklbw                m4, m6
+    paddsw                   m1, m4
+%endif
+    lea                    srcq, [srcq + sstrideq * 2 ]
+    lea                   src1q, [src1q + sstrideq * 2]
+    packuswb                 m1, m1
+
+%ifidn %1, v8_avg
+    movx                     m2, [dstq]
+    pavgb                    m0, m2
+%endif
+    movx                 [dstq], m0
+    add                    dstq, dst_stride
+%ifidn %1, v8_avg
+    movx                     m3, [dstq]
+    pavgb                    m1, m3
+%endif
+    movx                 [dstq], m1
+    add                    dstq, dst_stride
+    sub                 heightd, 2
+    jg                    .loop
+
+    ; Do last row if output_height is odd
+    jne                   .done
+
+    movx                     m0, [srcq                ]     ;A
+    movx                     m1, [srcq + sstrideq     ]     ;B
+    movx                     m6, [srcq + sstride6q    ]     ;G
+    punpcklbw                m0, m1                         ;A B
+    movx                     m7, [src1q + sstride6q   ]     ;H
+    pmaddubsw                m0, k0k1
+    movx                     m2, [srcq + sstrideq * 2 ]     ;C
+    punpcklbw                m6, m7                         ;G H
+    movx                     m3, [src1q + sstrideq * 2]     ;D
+    pmaddubsw                m6, k6k7
+    movx                     m4, [srcq + sstrideq * 4 ]     ;E
+    punpcklbw                m2, m3                         ;C D
+    movx                     m5, [src1q + sstrideq * 4]     ;F
+    punpcklbw                m4, m5                         ;E F
+    pmaddubsw                m2, k2k3
+    pmaddubsw                m4, k4k5
+    paddsw                   m2, m6
+    paddsw                   m0, m4
+    paddsw                   m0, m2
+    paddsw                   m0, krd
+    psraw                    m0, 7
+%ifidn %1, v8_add_src
+    pxor                     m6, m6
+    movu                     m4, [srcq]
+    punpcklbw                m4, m6
+    paddsw                   m0, m4
+%endif
+    packuswb                 m0, m0
+%ifidn %1, v8_avg
+    movx                     m1, [dstq]
+    pavgb                    m0, m1
+%endif
+    movx                 [dstq], m0
+
+%else
+    ; ARCH_X86_64
+
+    movx                     m0, [srcq                ]     ;A
+    movx                     m1, [srcq + sstrideq     ]     ;B
+    lea                    srcq, [srcq + sstrideq * 2 ]
+    movx                     m2, [srcq]                     ;C
+    movx                     m3, [srcq + sstrideq]          ;D
+    lea                    srcq, [srcq + sstrideq * 2 ]
+    movx                     m4, [srcq]                     ;E
+    movx                     m5, [srcq + sstrideq]          ;F
+    lea                    srcq, [srcq + sstrideq * 2 ]
+    movx                     m6, [srcq]                     ;G
+    punpcklbw                m0, m1                         ;A B
+    punpcklbw                m1, m2                         ;A B next iter
+    punpcklbw                m2, m3                         ;C D
+    punpcklbw                m3, m4                         ;C D next iter
+    punpcklbw                m4, m5                         ;E F
+    punpcklbw                m5, m6                         ;E F next iter
+
+.loop:
+    ;Do two rows at once
+    movx                     m7, [srcq + sstrideq]          ;H
+    lea                    srcq, [srcq + sstrideq * 2 ]
+    movx                    m14, [srcq]                     ;H next iter
+    punpcklbw                m6, m7                         ;G H
+    punpcklbw                m7, m14                        ;G H next iter
+    pmaddubsw                m8, m0, k0k1
+    pmaddubsw                m9, m1, k0k1
+    mova                     m0, m2
+    mova                     m1, m3
+    pmaddubsw               m10, m2, k2k3
+    pmaddubsw               m11, m3, k2k3
+    mova                     m2, m4
+    mova                     m3, m5
+    pmaddubsw                m4, k4k5
+    pmaddubsw                m5, k4k5
+    paddsw                   m8, m4
+    paddsw                   m9, m5
+    mova                     m4, m6
+    mova                     m5, m7
+    pmaddubsw                m6, k6k7
+    pmaddubsw                m7, k6k7
+    paddsw                  m10, m6
+    paddsw                  m11, m7
+    paddsw                   m8, m10
+    paddsw                   m9, m11
+    mova                     m6, m14
+    paddsw                   m8, krd
+    paddsw                   m9, krd
+    psraw                    m8, 7
+    psraw                    m9, 7
+%ifidn %2, 4
+    packuswb                 m8, m8
+    packuswb                 m9, m9
+%else
+    packuswb                 m8, m9
+%endif
+
+%ifidn %1, v8_avg
+    movx                     m7, [dstq]
+%ifidn %2, 4
+    movx                    m10, [dstq + dstrideq]
+    pavgb                    m9, m10
+%else
+    movhpd                   m7, [dstq + dstrideq]
+%endif
+    pavgb                    m8, m7
+%endif
+    movx                 [dstq], m8
+%ifidn %2, 4
+    movx      [dstq + dstrideq], m9
+%else
+    movhpd    [dstq + dstrideq], m8
+%endif
+
+    lea                    dstq, [dstq + dstrideq * 2 ]
+    sub                 heightd, 2
+    jg                    .loop
+
+    ; Do last row if output_height is odd
+    jne                   .done
+
+    movx                     m7, [srcq + sstrideq]          ;H
+    punpcklbw                m6, m7                         ;G H
+    pmaddubsw                m0, k0k1
+    pmaddubsw                m2, k2k3
+    pmaddubsw                m4, k4k5
+    pmaddubsw                m6, k6k7
+    paddsw                   m0, m4
+    paddsw                   m2, m6
+    paddsw                   m0, m2
+    paddsw                   m0, krd
+    psraw                    m0, 7
+    packuswb                 m0, m0
+%ifidn %1, v8_avg
+    movx                     m1, [dstq]
+    pavgb                    m0, m1
+%endif
+    movx                 [dstq], m0
+
+%endif ; ARCH_X86_64
+
+.done:
+    REP_RET
+
+%endm
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_VFILTER16 1
+cglobal filter_block1d16_%1, 6, NUM_GENERAL_REG_USED, 16, LOCAL_VARS_SIZE, \
+                             src, sstride, dst, dstride, height, filter
+    mova                     m4, [filterq]
+    SETUP_LOCAL_VARS
+
+%if ARCH_X86 || X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON
+
+%if ARCH_X86_64
+    %define               src1q  r7
+    %define           sstride6q  r8
+    %define          dst_stride  dstrideq
+%else
+    %define               src1q  filterq
+    %define           sstride6q  dstrideq
+    %define          dst_stride  dstridemp
+%endif
+    lea                   src1q, [srcq + sstrideq]
+    lea               sstride6q, [sstrideq + sstrideq * 4]
+    add               sstride6q, sstrideq                   ;pitch * 6
+
+.loop:
+    movh                     m0, [srcq                ]     ;A
+    movh                     m1, [src1q               ]     ;B
+    movh                     m2, [srcq + sstrideq * 2 ]     ;C
+    movh                     m3, [src1q + sstrideq * 2]     ;D
+    movh                     m4, [srcq + sstrideq * 4 ]     ;E
+    movh                     m5, [src1q + sstrideq * 4]     ;F
+
+    punpcklbw                m0, m1                         ;A B
+    movh                     m6, [srcq + sstride6q]         ;G
+    punpcklbw                m2, m3                         ;C D
+    movh                     m7, [src1q + sstride6q]        ;H
+    punpcklbw                m4, m5                         ;E F
+    pmaddubsw                m0, k0k1
+    movh                     m3, [srcq + 8]                 ;A
+    pmaddubsw                m2, k2k3
+    punpcklbw                m6, m7                         ;G H
+    movh                     m5, [srcq + sstrideq + 8]      ;B
+    pmaddubsw                m4, k4k5
+    punpcklbw                m3, m5                         ;A B
+    movh                     m7, [srcq + sstrideq * 2 + 8]  ;C
+    pmaddubsw                m6, k6k7
+    movh                     m5, [src1q + sstrideq * 2 + 8] ;D
+    punpcklbw                m7, m5                         ;C D
+    paddsw                   m2, m6
+    pmaddubsw                m3, k0k1
+    movh                     m1, [srcq + sstrideq * 4 + 8]  ;E
+    paddsw                   m0, m4
+    pmaddubsw                m7, k2k3
+    movh                     m6, [src1q + sstrideq * 4 + 8] ;F
+    punpcklbw                m1, m6                         ;E F
+    paddsw                   m0, m2
+    paddsw                   m0, krd
+    movh                     m2, [srcq + sstride6q + 8]     ;G
+    pmaddubsw                m1, k4k5
+    movh                     m5, [src1q + sstride6q + 8]    ;H
+    psraw                    m0, 7
+    punpcklbw                m2, m5                         ;G H
+    pmaddubsw                m2, k6k7
+    paddsw                   m7, m2
+    paddsw                   m3, m1
+    paddsw                   m3, m7
+    paddsw                   m3, krd
+    psraw                    m3, 7
+%ifidn %1, v8_add_src
+    pxor                     m6, m6
+    movu                     m4, [src1q + 2 * sstrideq] ; Fetch from 3 rows down
+    mova                     m5, m4
+    punpcklbw                m4, m6
+    punpckhbw                m5, m6
+    paddsw                   m0, m4
+    paddsw                   m3, m5
+%endif
+    packuswb                 m0, m3
+
+    add                    srcq, sstrideq
+    add                   src1q, sstrideq
+%ifidn %1, v8_avg
+    pavgb                    m0, [dstq]
+%endif
+    mova                 [dstq], m0
+    add                    dstq, dst_stride
+    dec                 heightd
+    jnz                   .loop
+    REP_RET
+
+%else
+    ; ARCH_X86_64
+    dec                 heightd
+
+    movu                     m1, [srcq                ]     ;A
+    movu                     m3, [srcq + sstrideq     ]     ;B
+    lea                    srcq, [srcq + sstrideq * 2]
+    punpcklbw                m0, m1, m3                     ;A B
+    punpckhbw                m1, m3                         ;A B
+    movu                     m5, [srcq]                     ;C
+    punpcklbw                m2, m3, m5                     ;A B next iter
+    punpckhbw                m3, m5                         ;A B next iter
+    mova                   tmp0, m2                         ;store to stack
+    mova                   tmp1, m3                         ;store to stack
+    movu                     m7, [srcq + sstrideq]          ;D
+    lea                    srcq, [srcq + sstrideq * 2]
+    punpcklbw                m4, m5, m7                     ;C D
+    punpckhbw                m5, m7                         ;C D
+    movu                     m9, [srcq]                     ;E
+    punpcklbw                m6, m7, m9                     ;C D next iter
+    punpckhbw                m7, m9                         ;C D next iter
+    movu                    m11, [srcq + sstrideq]          ;F
+    lea                    srcq, [srcq + sstrideq * 2]
+    punpcklbw                m8, m9, m11                    ;E F
+    punpckhbw                m9, m11                        ;E F
+    movu                     m2, [srcq]                     ;G
+    punpcklbw               m10, m11, m2                    ;E F next iter
+    punpckhbw               m11, m2                         ;E F next iter
+
+.loop:
+    ;Do two rows at once
+    pmaddubsw               m13, m0, k0k1
+    mova                     m0, m4
+    pmaddubsw               m14, m8, k4k5
+    pmaddubsw               m15, m4, k2k3
+    mova                     m4, m8
+    paddsw                  m13, m14
+    movu                     m3, [srcq + sstrideq]          ;H
+    lea                    srcq, [srcq + sstrideq * 2]
+    punpcklbw               m14, m2, m3                     ;G H
+    mova                     m8, m14
+    pmaddubsw               m14, k6k7
+    paddsw                  m15, m14
+    paddsw                  m13, m15
+    paddsw                  m13, krd
+    psraw                   m13, 7
+
+    pmaddubsw               m14, m1, k0k1
+    pmaddubsw                m1, m9, k4k5
+    pmaddubsw               m15, m5, k2k3
+    paddsw                  m14, m1
+    mova                     m1, m5
+    mova                     m5, m9
+    punpckhbw                m2, m3                         ;G H
+    mova                     m9, m2
+    pmaddubsw                m2, k6k7
+    paddsw                  m15, m2
+    paddsw                  m14, m15
+    paddsw                  m14, krd
+    psraw                   m14, 7
+    packuswb                m13, m14
+%ifidn %1, v8_avg
+    pavgb                   m13, [dstq]
+%endif
+    mova                 [dstq], m13
+
+    ; next iter
+    pmaddubsw               m15, tmp0, k0k1
+    pmaddubsw               m14, m10, k4k5
+    pmaddubsw               m13, m6, k2k3
+    paddsw                  m15, m14
+    mova                   tmp0, m6
+    mova                     m6, m10
+    movu                     m2, [srcq]                     ;G next iter
+    punpcklbw               m14, m3, m2                     ;G H next iter
+    mova                    m10, m14
+    pmaddubsw               m14, k6k7
+    paddsw                  m13, m14
+    paddsw                  m15, m13
+    paddsw                  m15, krd
+    psraw                   m15, 7
+
+    pmaddubsw               m14, tmp1, k0k1
+    mova                   tmp1, m7
+    pmaddubsw               m13, m7, k2k3
+    mova                     m7, m11
+    pmaddubsw               m11, k4k5
+    paddsw                  m14, m11
+    punpckhbw                m3, m2                         ;G H next iter
+    mova                    m11, m3
+    pmaddubsw                m3, k6k7
+    paddsw                  m13, m3
+    paddsw                  m14, m13
+    paddsw                  m14, krd
+    psraw                   m14, 7
+    packuswb                m15, m14
+%ifidn %1, v8_avg
+    pavgb                   m15, [dstq + dstrideq]
+%endif
+    mova      [dstq + dstrideq], m15
+    lea                    dstq, [dstq + dstrideq * 2]
+    sub                 heightd, 2
+    jg                    .loop
+
+    ; Do last row if output_height is odd
+    jne                   .done
+
+    movu                     m3, [srcq + sstrideq]          ;H
+    punpcklbw                m6, m2, m3                     ;G H
+    punpckhbw                m2, m3                         ;G H
+    pmaddubsw                m0, k0k1
+    pmaddubsw                m1, k0k1
+    pmaddubsw                m4, k2k3
+    pmaddubsw                m5, k2k3
+    pmaddubsw                m8, k4k5
+    pmaddubsw                m9, k4k5
+    pmaddubsw                m6, k6k7
+    pmaddubsw                m2, k6k7
+    paddsw                   m0, m8
+    paddsw                   m1, m9
+    paddsw                   m4, m6
+    paddsw                   m5, m2
+    paddsw                   m0, m4
+    paddsw                   m1, m5
+    paddsw                   m0, krd
+    paddsw                   m1, krd
+    psraw                    m0, 7
+    psraw                    m1, 7
+    packuswb                 m0, m1
+%ifidn %1, v8_avg
+    pavgb                    m0, [dstq]
+%endif
+    mova                 [dstq], m0
+
+.done:
+    REP_RET
+
+%endif ; ARCH_X86_64
+
+%endm
+
+INIT_XMM ssse3
+SUBPIX_VFILTER16     v8
+SUBPIX_VFILTER16 v8_avg
+SUBPIX_VFILTER       v8, 8
+SUBPIX_VFILTER   v8_avg, 8
+SUBPIX_VFILTER       v8, 4
+SUBPIX_VFILTER   v8_avg, 4
+
+%if (ARCH_X86 || X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON) && \
+    CONFIG_LOOP_RESTORATION
+SUBPIX_VFILTER16 v8_add_src
+SUBPIX_VFILTER   v8_add_src, 8
+SUBPIX_VFILTER   v8_add_src, 4
+%endif
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm
new file mode 100644
index 0000000000..8f025a8be4
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm
@@ -0,0 +1,451 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "aom_ports/x86_abi_support.asm"
+
+%macro GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    pshuflw     xmm4, xmm3, 11111111b       ;k3
+    psrldq      xmm3, 8
+    pshuflw     xmm3, xmm3, 0b              ;k4
+    punpcklqdq  xmm4, xmm3                  ;k3k4
+
+    movq        xmm3, rcx                   ;rounding
+    pshufd      xmm3, xmm3, 0
+
+    pxor        xmm2, xmm2
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_4 1
+
+    punpckldq   xmm0, xmm1                  ;two row in one register
+    punpcklbw   xmm0, xmm2                  ;unpack to word
+    pmullw      xmm0, xmm4                  ;multiply the filter factors
+
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 8
+    paddsw      xmm0, xmm1
+
+    paddsw      xmm0, xmm3                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+
+    movd        [rdi], xmm0
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+
+    pshuflw     xmm6, xmm7, 11111111b       ;k3
+    pshufhw     xmm7, xmm7, 0b              ;k4
+    punpcklwd   xmm6, xmm6
+    punpckhwd   xmm7, xmm7
+
+    movq        xmm4, rcx                   ;rounding
+    pshufd      xmm4, xmm4, 0
+
+    pxor        xmm5, xmm5
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_8 1
+    punpcklbw   xmm0, xmm5
+    punpcklbw   xmm1, xmm5
+
+    pmullw      xmm0, xmm6
+    pmullw      xmm1, xmm7
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm4                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack back to byte
+%if %1
+    movq        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro APPLY_FILTER_16 1
+    punpcklbw   xmm0, xmm5
+    punpcklbw   xmm1, xmm5
+    punpckhbw   xmm2, xmm5
+    punpckhbw   xmm3, xmm5
+
+    pmullw      xmm0, xmm6
+    pmullw      xmm1, xmm7
+    pmullw      xmm2, xmm6
+    pmullw      xmm3, xmm7
+
+    paddsw      xmm0, xmm1
+    paddsw      xmm2, xmm3
+
+    paddsw      xmm0, xmm4                  ;rounding
+    paddsw      xmm2, xmm4
+    psraw       xmm0, 7                     ;shift
+    psraw       xmm2, 7
+    packuswb    xmm0, xmm2                  ;pack back to byte
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+global sym(aom_filter_block1d4_v2_sse2) PRIVATE
+sym(aom_filter_block1d4_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d8_v2_sse2) PRIVATE
+sym(aom_filter_block1d8_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d16_v2_sse2) PRIVATE
+sym(aom_filter_block1d16_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+    movdqa        xmm3, xmm1
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d4_v2_avg_sse2) PRIVATE
+sym(aom_filter_block1d4_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d8_v2_avg_sse2) PRIVATE
+sym(aom_filter_block1d8_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d16_v2_avg_sse2) PRIVATE
+sym(aom_filter_block1d16_v2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+    movdqa        xmm3, xmm1
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d4_h2_sse2) PRIVATE
+sym(aom_filter_block1d4_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d8_h2_sse2) PRIVATE
+sym(aom_filter_block1d8_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d16_h2_sse2) PRIVATE
+sym(aom_filter_block1d16_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm1
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d4_h2_avg_sse2) PRIVATE
+sym(aom_filter_block1d4_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d8_h2_avg_sse2) PRIVATE
+sym(aom_filter_block1d8_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d16_h2_avg_sse2) PRIVATE
+sym(aom_filter_block1d16_h2_avg_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm1
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm
new file mode 100644
index 0000000000..b9b2da0bea
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm
@@ -0,0 +1,421 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "aom_ports/x86_abi_support.asm"
+
+%macro GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         ecx, 0x01000100
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    psrldq      xmm3, 6
+    packsswb    xmm3, xmm3
+    pshuflw     xmm3, xmm3, 0b              ;k3_k4
+
+    movd        xmm2, ecx                   ;rounding_shift
+    pshufd      xmm2, xmm2, 0
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_4 1
+    punpcklbw   xmm0, xmm1
+    pmaddubsw   xmm0, xmm3
+
+    pmulhrsw    xmm0, xmm2                  ;rounding(+64)+shift(>>7)
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movd        [rdi], xmm0
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         ecx, 0x01000100
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    psrldq      xmm7, 6
+    packsswb    xmm7, xmm7
+    pshuflw     xmm7, xmm7, 0b              ;k3_k4
+    punpcklwd   xmm7, xmm7
+
+    movd        xmm6, ecx                   ;rounding_shift
+    pshufd      xmm6, xmm6, 0
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_8 1
+    punpcklbw   xmm0, xmm1
+    pmaddubsw   xmm0, xmm7
+
+    pmulhrsw    xmm0, xmm6                  ;rounding(+64)+shift(>>7)
+    packuswb    xmm0, xmm0                  ;pack back to byte
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro APPLY_FILTER_16 1
+    punpcklbw   xmm0, xmm1
+    punpckhbw   xmm2, xmm1
+    pmaddubsw   xmm0, xmm7
+    pmaddubsw   xmm2, xmm7
+
+    pmulhrsw    xmm0, xmm6                  ;rounding(+64)+shift(>>7)
+    pmulhrsw    xmm2, xmm6
+    packuswb    xmm0, xmm2                  ;pack back to byte
+
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+global sym(aom_filter_block1d4_v2_ssse3) PRIVATE
+sym(aom_filter_block1d4_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d8_v2_ssse3) PRIVATE
+sym(aom_filter_block1d8_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d16_v2_ssse3) PRIVATE
+sym(aom_filter_block1d16_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d4_v2_avg_ssse3) PRIVATE
+sym(aom_filter_block1d4_v2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d8_v2_avg_ssse3) PRIVATE
+sym(aom_filter_block1d8_v2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d16_v2_avg_ssse3) PRIVATE
+sym(aom_filter_block1d16_v2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d4_h2_ssse3) PRIVATE
+sym(aom_filter_block1d4_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d8_h2_ssse3) PRIVATE
+sym(aom_filter_block1d8_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d16_h2_ssse3) PRIVATE
+sym(aom_filter_block1d16_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d4_h2_avg_ssse3) PRIVATE
+sym(aom_filter_block1d4_h2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d8_h2_avg_ssse3) PRIVATE
+sym(aom_filter_block1d8_h2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+global sym(aom_filter_block1d16_h2_avg_ssse3) PRIVATE
+sym(aom_filter_block1d16_h2_avg_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+
+    APPLY_FILTER_16 1
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c b/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c
new file mode 100644
index 0000000000..bcdc20f638
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c
@@ -0,0 +1,426 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>
+
+#include "aom_dsp/x86/synonyms.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/mem.h"
+
+void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp,
+                         int *min, int *max) {
+  __m128i u0, s0, d0, diff, maxabsdiff, minabsdiff, negdiff, absdiff0, absdiff;
+  u0 = _mm_setzero_si128();
+  // Row 0
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff0 = _mm_max_epi16(diff, negdiff);
+  // Row 1
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(absdiff0, absdiff);
+  minabsdiff = _mm_min_epi16(absdiff0, absdiff);
+  // Row 2
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 2 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 3
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 3 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 4
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 4 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 5
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 5 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 6
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 6 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 7
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 7 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_si128(maxabsdiff, 8));
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 32));
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 16));
+  *max = _mm_extract_epi16(maxabsdiff, 0);
+
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_si128(minabsdiff, 8));
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 32));
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 16));
+  *min = _mm_extract_epi16(minabsdiff, 0);
+}
+
+unsigned int aom_avg_8x8_sse2(const uint8_t *s, int p) {
+  __m128i s0, s1, u0;
+  unsigned int avg = 0;
+  u0 = _mm_setzero_si128();
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  s0 = _mm_adds_epu16(s0, _mm_srli_si128(s0, 8));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 32));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 16));
+  avg = _mm_extract_epi16(s0, 0);
+  return (avg + 32) >> 6;
+}
+
+unsigned int aom_avg_4x4_sse2(const uint8_t *s, int p) {
+  __m128i s0, s1, u0;
+  unsigned int avg = 0;
+
+  u0 = _mm_setzero_si128();
+  s0 = _mm_unpacklo_epi8(xx_loadl_32(s), u0);
+  s1 = _mm_unpacklo_epi8(xx_loadl_32(s + p), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(xx_loadl_32(s + 2 * p), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(xx_loadl_32(s + 3 * p), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  s0 = _mm_adds_epu16(s0, _mm_srli_si128(s0, 4));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 16));
+  avg = _mm_extract_epi16(s0, 0);
+  return (avg + 8) >> 4;
+}
+
+static void hadamard_col8_sse2(__m128i *in, int iter) {
+  __m128i a0 = in[0];
+  __m128i a1 = in[1];
+  __m128i a2 = in[2];
+  __m128i a3 = in[3];
+  __m128i a4 = in[4];
+  __m128i a5 = in[5];
+  __m128i a6 = in[6];
+  __m128i a7 = in[7];
+
+  __m128i b0 = _mm_add_epi16(a0, a1);
+  __m128i b1 = _mm_sub_epi16(a0, a1);
+  __m128i b2 = _mm_add_epi16(a2, a3);
+  __m128i b3 = _mm_sub_epi16(a2, a3);
+  __m128i b4 = _mm_add_epi16(a4, a5);
+  __m128i b5 = _mm_sub_epi16(a4, a5);
+  __m128i b6 = _mm_add_epi16(a6, a7);
+  __m128i b7 = _mm_sub_epi16(a6, a7);
+
+  a0 = _mm_add_epi16(b0, b2);
+  a1 = _mm_add_epi16(b1, b3);
+  a2 = _mm_sub_epi16(b0, b2);
+  a3 = _mm_sub_epi16(b1, b3);
+  a4 = _mm_add_epi16(b4, b6);
+  a5 = _mm_add_epi16(b5, b7);
+  a6 = _mm_sub_epi16(b4, b6);
+  a7 = _mm_sub_epi16(b5, b7);
+
+  if (iter == 0) {
+    b0 = _mm_add_epi16(a0, a4);
+    b7 = _mm_add_epi16(a1, a5);
+    b3 = _mm_add_epi16(a2, a6);
+    b4 = _mm_add_epi16(a3, a7);
+    b2 = _mm_sub_epi16(a0, a4);
+    b6 = _mm_sub_epi16(a1, a5);
+    b1 = _mm_sub_epi16(a2, a6);
+    b5 = _mm_sub_epi16(a3, a7);
+
+    a0 = _mm_unpacklo_epi16(b0, b1);
+    a1 = _mm_unpacklo_epi16(b2, b3);
+    a2 = _mm_unpackhi_epi16(b0, b1);
+    a3 = _mm_unpackhi_epi16(b2, b3);
+    a4 = _mm_unpacklo_epi16(b4, b5);
+    a5 = _mm_unpacklo_epi16(b6, b7);
+    a6 = _mm_unpackhi_epi16(b4, b5);
+    a7 = _mm_unpackhi_epi16(b6, b7);
+
+    b0 = _mm_unpacklo_epi32(a0, a1);
+    b1 = _mm_unpacklo_epi32(a4, a5);
+    b2 = _mm_unpackhi_epi32(a0, a1);
+    b3 = _mm_unpackhi_epi32(a4, a5);
+    b4 = _mm_unpacklo_epi32(a2, a3);
+    b5 = _mm_unpacklo_epi32(a6, a7);
+    b6 = _mm_unpackhi_epi32(a2, a3);
+    b7 = _mm_unpackhi_epi32(a6, a7);
+
+    in[0] = _mm_unpacklo_epi64(b0, b1);
+    in[1] = _mm_unpackhi_epi64(b0, b1);
+    in[2] = _mm_unpacklo_epi64(b2, b3);
+    in[3] = _mm_unpackhi_epi64(b2, b3);
+    in[4] = _mm_unpacklo_epi64(b4, b5);
+    in[5] = _mm_unpackhi_epi64(b4, b5);
+    in[6] = _mm_unpacklo_epi64(b6, b7);
+    in[7] = _mm_unpackhi_epi64(b6, b7);
+  } else {
+    in[0] = _mm_add_epi16(a0, a4);
+    in[7] = _mm_add_epi16(a1, a5);
+    in[3] = _mm_add_epi16(a2, a6);
+    in[4] = _mm_add_epi16(a3, a7);
+    in[2] = _mm_sub_epi16(a0, a4);
+    in[6] = _mm_sub_epi16(a1, a5);
+    in[1] = _mm_sub_epi16(a2, a6);
+    in[5] = _mm_sub_epi16(a3, a7);
+  }
+}
+
+void aom_hadamard_8x8_sse2(int16_t const *src_diff, int src_stride,
+                           int16_t *coeff) {
+  __m128i src[8];
+  src[0] = _mm_load_si128((const __m128i *)src_diff);
+  src[1] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[2] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[3] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[4] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[5] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[6] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[7] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+
+  hadamard_col8_sse2(src, 0);
+  hadamard_col8_sse2(src, 1);
+
+  _mm_store_si128((__m128i *)coeff, src[0]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[1]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[2]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[3]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[4]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[5]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[6]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[7]);
+}
+
+void aom_hadamard_16x16_sse2(int16_t const *src_diff, int src_stride,
+                             int16_t *coeff) {
+  int idx;
+  for (idx = 0; idx < 4; ++idx) {
+    int16_t const *src_ptr =
+        src_diff + (idx >> 1) * 8 * src_stride + (idx & 0x01) * 8;
+    aom_hadamard_8x8_sse2(src_ptr, src_stride, coeff + idx * 64);
+  }
+
+  for (idx = 0; idx < 64; idx += 8) {
+    __m128i coeff0 = _mm_load_si128((const __m128i *)coeff);
+    __m128i coeff1 = _mm_load_si128((const __m128i *)(coeff + 64));
+    __m128i coeff2 = _mm_load_si128((const __m128i *)(coeff + 128));
+    __m128i coeff3 = _mm_load_si128((const __m128i *)(coeff + 192));
+
+    __m128i b0 = _mm_add_epi16(coeff0, coeff1);
+    __m128i b1 = _mm_sub_epi16(coeff0, coeff1);
+    __m128i b2 = _mm_add_epi16(coeff2, coeff3);
+    __m128i b3 = _mm_sub_epi16(coeff2, coeff3);
+
+    b0 = _mm_srai_epi16(b0, 1);
+    b1 = _mm_srai_epi16(b1, 1);
+    b2 = _mm_srai_epi16(b2, 1);
+    b3 = _mm_srai_epi16(b3, 1);
+
+    coeff0 = _mm_add_epi16(b0, b2);
+    coeff1 = _mm_add_epi16(b1, b3);
+    _mm_store_si128((__m128i *)coeff, coeff0);
+    _mm_store_si128((__m128i *)(coeff + 64), coeff1);
+
+    coeff2 = _mm_sub_epi16(b0, b2);
+    coeff3 = _mm_sub_epi16(b1, b3);
+    _mm_store_si128((__m128i *)(coeff + 128), coeff2);
+    _mm_store_si128((__m128i *)(coeff + 192), coeff3);
+
+    coeff += 8;
+  }
+}
+
+int aom_satd_sse2(const int16_t *coeff, int length) {
+  int i;
+  const __m128i zero = _mm_setzero_si128();
+  __m128i accum = zero;
+
+  for (i = 0; i < length; i += 8) {
+    const __m128i src_line = _mm_load_si128((const __m128i *)coeff);
+    const __m128i inv = _mm_sub_epi16(zero, src_line);
+    const __m128i abs = _mm_max_epi16(src_line, inv);  // abs(src_line)
+    const __m128i abs_lo = _mm_unpacklo_epi16(abs, zero);
+    const __m128i abs_hi = _mm_unpackhi_epi16(abs, zero);
+    const __m128i sum = _mm_add_epi32(abs_lo, abs_hi);
+    accum = _mm_add_epi32(accum, sum);
+    coeff += 8;
+  }
+
+  {  // cascading summation of accum
+    __m128i hi = _mm_srli_si128(accum, 8);
+    accum = _mm_add_epi32(accum, hi);
+    hi = _mm_srli_epi64(accum, 32);
+    accum = _mm_add_epi32(accum, hi);
+  }
+
+  return _mm_cvtsi128_si32(accum);
+}
+
+void aom_int_pro_row_sse2(int16_t *hbuf, uint8_t const *ref, int ref_stride,
+                          int height) {
+  int idx;
+  __m128i zero = _mm_setzero_si128();
+  __m128i src_line = _mm_loadu_si128((const __m128i *)ref);
+  __m128i s0 = _mm_unpacklo_epi8(src_line, zero);
+  __m128i s1 = _mm_unpackhi_epi8(src_line, zero);
+  __m128i t0, t1;
+  int height_1 = height - 1;
+  ref += ref_stride;
+
+  for (idx = 1; idx < height_1; idx += 2) {
+    src_line = _mm_loadu_si128((const __m128i *)ref);
+    t0 = _mm_unpacklo_epi8(src_line, zero);
+    t1 = _mm_unpackhi_epi8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, t0);
+    s1 = _mm_adds_epu16(s1, t1);
+    ref += ref_stride;
+
+    src_line = _mm_loadu_si128((const __m128i *)ref);
+    t0 = _mm_unpacklo_epi8(src_line, zero);
+    t1 = _mm_unpackhi_epi8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, t0);
+    s1 = _mm_adds_epu16(s1, t1);
+    ref += ref_stride;
+  }
+
+  src_line = _mm_loadu_si128((const __m128i *)ref);
+  t0 = _mm_unpacklo_epi8(src_line, zero);
+  t1 = _mm_unpackhi_epi8(src_line, zero);
+  s0 = _mm_adds_epu16(s0, t0);
+  s1 = _mm_adds_epu16(s1, t1);
+
+  if (height == 64) {
+    s0 = _mm_srai_epi16(s0, 5);
+    s1 = _mm_srai_epi16(s1, 5);
+  } else if (height == 32) {
+    s0 = _mm_srai_epi16(s0, 4);
+    s1 = _mm_srai_epi16(s1, 4);
+  } else {
+    s0 = _mm_srai_epi16(s0, 3);
+    s1 = _mm_srai_epi16(s1, 3);
+  }
+
+  _mm_storeu_si128((__m128i *)hbuf, s0);
+  hbuf += 8;
+  _mm_storeu_si128((__m128i *)hbuf, s1);
+}
+
+int16_t aom_int_pro_col_sse2(uint8_t const *ref, int width) {
+  __m128i zero = _mm_setzero_si128();
+  __m128i src_line = _mm_load_si128((const __m128i *)ref);
+  __m128i s0 = _mm_sad_epu8(src_line, zero);
+  __m128i s1;
+  int i;
+
+  for (i = 16; i < width; i += 16) {
+    ref += 16;
+    src_line = _mm_load_si128((const __m128i *)ref);
+    s1 = _mm_sad_epu8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, s1);
+  }
+
+  s1 = _mm_srli_si128(s0, 8);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  return _mm_extract_epi16(s0, 0);
+}
+
+int aom_vector_var_sse2(int16_t const *ref, int16_t const *src, int bwl) {
+  int idx;
+  int width = 4 << bwl;
+  int16_t mean;
+  __m128i v0 = _mm_loadu_si128((const __m128i *)ref);
+  __m128i v1 = _mm_load_si128((const __m128i *)src);
+  __m128i diff = _mm_subs_epi16(v0, v1);
+  __m128i sum = diff;
+  __m128i sse = _mm_madd_epi16(diff, diff);
+
+  ref += 8;
+  src += 8;
+
+  for (idx = 8; idx < width; idx += 8) {
+    v0 = _mm_loadu_si128((const __m128i *)ref);
+    v1 = _mm_load_si128((const __m128i *)src);
+    diff = _mm_subs_epi16(v0, v1);
+
+    sum = _mm_add_epi16(sum, diff);
+    v0 = _mm_madd_epi16(diff, diff);
+    sse = _mm_add_epi32(sse, v0);
+
+    ref += 8;
+    src += 8;
+  }
+
+  v0 = _mm_srli_si128(sum, 8);
+  sum = _mm_add_epi16(sum, v0);
+  v0 = _mm_srli_epi64(sum, 32);
+  sum = _mm_add_epi16(sum, v0);
+  v0 = _mm_srli_epi32(sum, 16);
+  sum = _mm_add_epi16(sum, v0);
+
+  v1 = _mm_srli_si128(sse, 8);
+  sse = _mm_add_epi32(sse, v1);
+  v1 = _mm_srli_epi64(sse, 32);
+  sse = _mm_add_epi32(sse, v1);
+
+  mean = _mm_extract_epi16(sum, 0);
+
+  return _mm_cvtsi128_si32(sse) - ((mean * mean) >> (bwl + 2));
+}
diff --git a/third_party/aom/aom_dsp/x86/avg_ssse3_x86_64.asm b/third_party/aom/aom_dsp/x86/avg_ssse3_x86_64.asm
new file mode 100644
index 0000000000..b2d150296b
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/avg_ssse3_x86_64.asm
@@ -0,0 +1,124 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%define private_prefix aom
+
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the hadamard transformation. Part
+; of the macro definitions are originally derived from the ffmpeg project.
+; The current version applies to x86 64-bit only.
+
+SECTION .text
+
+%if ARCH_X86_64
+; matrix transpose
+%macro INTERLEAVE_2X 4
+  punpckh%1          m%4, m%2, m%3
+  punpckl%1          m%2, m%3
+  SWAP               %3,  %4
+%endmacro
+
+%macro TRANSPOSE8X8 9
+  INTERLEAVE_2X  wd, %1, %2, %9
+  INTERLEAVE_2X  wd, %3, %4, %9
+  INTERLEAVE_2X  wd, %5, %6, %9
+  INTERLEAVE_2X  wd, %7, %8, %9
+
+  INTERLEAVE_2X  dq, %1, %3, %9
+  INTERLEAVE_2X  dq, %2, %4, %9
+  INTERLEAVE_2X  dq, %5, %7, %9
+  INTERLEAVE_2X  dq, %6, %8, %9
+
+  INTERLEAVE_2X  qdq, %1, %5, %9
+  INTERLEAVE_2X  qdq, %3, %7, %9
+  INTERLEAVE_2X  qdq, %2, %6, %9
+  INTERLEAVE_2X  qdq, %4, %8, %9
+
+  SWAP  %2, %5
+  SWAP  %4, %7
+%endmacro
+
+%macro HMD8_1D 0
+  psubw              m8, m0, m1
+  psubw              m9, m2, m3
+  paddw              m0, m1
+  paddw              m2, m3
+  SWAP               1, 8
+  SWAP               3, 9
+  psubw              m8, m4, m5
+  psubw              m9, m6, m7
+  paddw              m4, m5
+  paddw              m6, m7
+  SWAP               5, 8
+  SWAP               7, 9
+
+  psubw              m8, m0, m2
+  psubw              m9, m1, m3
+  paddw              m0, m2
+  paddw              m1, m3
+  SWAP               2, 8
+  SWAP               3, 9
+  psubw              m8, m4, m6
+  psubw              m9, m5, m7
+  paddw              m4, m6
+  paddw              m5, m7
+  SWAP               6, 8
+  SWAP               7, 9
+
+  psubw              m8, m0, m4
+  psubw              m9, m1, m5
+  paddw              m0, m4
+  paddw              m1, m5
+  SWAP               4, 8
+  SWAP               5, 9
+  psubw              m8, m2, m6
+  psubw              m9, m3, m7
+  paddw              m2, m6
+  paddw              m3, m7
+  SWAP               6, 8
+  SWAP               7, 9
+%endmacro
+
+INIT_XMM ssse3
+cglobal hadamard_8x8, 3, 5, 10, input, stride, output
+  lea                r3, [2 * strideq]
+  lea                r4, [4 * strideq]
+
+  mova               m0, [inputq]
+  mova               m1, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m2, [inputq]
+  mova               m3, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m4, [inputq]
+  mova               m5, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m6, [inputq]
+  mova               m7, [inputq + r3]
+
+  HMD8_1D
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+  HMD8_1D
+
+  mova              [outputq +   0], m0
+  mova              [outputq +  16], m1
+  mova              [outputq +  32], m2
+  mova              [outputq +  48], m3
+  mova              [outputq +  64], m4
+  mova              [outputq +  80], m5
+  mova              [outputq +  96], m6
+  mova              [outputq + 112], m7
+
+  RET
+%endif
diff --git a/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c
new file mode 100644
index 0000000000..e916e4ff9e
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c
@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom/aom_integer.h"
+
+#include "./aom_dsp_rtcd.h"
+
+// To start out, just dispatch to the function using the 2D mask and
+// pass mask stride as 0. This can be improved upon if necessary.
+
+void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                const uint8_t *src0, uint32_t src0_stride,
+                                const uint8_t *src1, uint32_t src1_stride,
+                                const uint8_t *mask, int h, int w) {
+  aom_blend_a64_mask_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                            src1_stride, mask, 0, h, w, 0, 0);
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_blend_a64_hmask_sse4_1(
+    uint8_t *dst_8, uint32_t dst_stride, const uint8_t *src0_8,
+    uint32_t src0_stride, const uint8_t *src1_8, uint32_t src1_stride,
+    const uint8_t *mask, int h, int w, int bd) {
+  aom_highbd_blend_a64_mask_sse4_1(dst_8, dst_stride, src0_8, src0_stride,
+                                   src1_8, src1_stride, mask, 0, h, w, 0, 0,
+                                   bd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c
new file mode 100644
index 0000000000..68d74e5176
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c
@@ -0,0 +1,924 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <smmintrin.h>  // SSE4.1
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/blend_sse4.h"
+
+#include "./aom_dsp_rtcd.h"
+
+//////////////////////////////////////////////////////////////////////////////
+// No sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_mask_w4_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                     const uint8_t *src0, uint32_t src0_stride,
+                                     const uint8_t *src1, uint32_t src1_stride,
+                                     const uint8_t *mask, uint32_t mask_stride,
+                                     int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_m0_b = xx_loadl_32(mask);
+    const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_w8_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                     const uint8_t *src0, uint32_t src0_stride,
+                                     const uint8_t *src1, uint32_t src1_stride,
+                                     const uint8_t *mask, uint32_t mask_stride,
+                                     int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_m0_b = xx_loadl_64(mask);
+    const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_w16n_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_m0l_b = xx_loadl_64(mask + c);
+      const __m128i v_m0h_b = xx_loadl_64(mask + c + 8);
+      const __m128i v_m0l_w = _mm_cvtepu8_epi16(v_m0l_b);
+      const __m128i v_m0h_w = _mm_cvtepu8_epi16(v_m0h_b);
+      const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w);
+      const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w);
+
+      const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w);
+      const __m128i v_resh_w =
+          blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w);
+
+      const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Horizontal sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_mask_sx_w4_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0,
+                                         0xff, 0, 0xff, 0, 0xff, 0, 0xff);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_r_b = xx_loadl_64(mask);
+    const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1));
+
+    const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sx_w8_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0,
+                                         0xff, 0, 0xff, 0, 0xff, 0, 0xff);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_r_b = xx_loadu_128(mask);
+    const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1));
+
+    const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sx_w16n_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0,
+                                         0xff, 0, 0xff, 0, 0xff, 0, 0xff);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_rl_b = xx_loadu_128(mask + 2 * c);
+      const __m128i v_rh_b = xx_loadu_128(mask + 2 * c + 16);
+      const __m128i v_al_b = _mm_avg_epu8(v_rl_b, _mm_srli_si128(v_rl_b, 1));
+      const __m128i v_ah_b = _mm_avg_epu8(v_rh_b, _mm_srli_si128(v_rh_b, 1));
+
+      const __m128i v_m0l_w = _mm_and_si128(v_al_b, v_zmask_b);
+      const __m128i v_m0h_w = _mm_and_si128(v_ah_b, v_zmask_b);
+      const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w);
+      const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w);
+
+      const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w);
+      const __m128i v_resh_w =
+          blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w);
+
+      const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Vertical sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_mask_sy_w4_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_ra_b = xx_loadl_32(mask);
+    const __m128i v_rb_b = xx_loadl_32(mask + mask_stride);
+    const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+
+    const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sy_w8_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_ra_b = xx_loadl_64(mask);
+    const __m128i v_rb_b = xx_loadl_64(mask + mask_stride);
+    const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+
+    const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sy_w16n_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  const __m128i v_zero = _mm_setzero_si128();
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_ra_b = xx_loadu_128(mask + c);
+      const __m128i v_rb_b = xx_loadu_128(mask + c + mask_stride);
+      const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+
+      const __m128i v_m0l_w = _mm_cvtepu8_epi16(v_a_b);
+      const __m128i v_m0h_w = _mm_unpackhi_epi8(v_a_b, v_zero);
+      const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w);
+      const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w);
+
+      const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w);
+      const __m128i v_resh_w =
+          blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w);
+
+      const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Horizontal and Vertical sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_mask_sx_sy_w4_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0,
+                                         0xff, 0, 0xff, 0, 0xff, 0, 0xff);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_ra_b = xx_loadl_64(mask);
+    const __m128i v_rb_b = xx_loadl_64(mask + mask_stride);
+    const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b);
+    const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b);
+    const __m128i v_rvsb_w =
+        _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b);
+    const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w);
+
+    const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sx_sy_w8_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0,
+                                         0xff, 0, 0xff, 0, 0xff, 0, 0xff);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_ra_b = xx_loadu_128(mask);
+    const __m128i v_rb_b = xx_loadu_128(mask + mask_stride);
+    const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b);
+    const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b);
+    const __m128i v_rvsb_w =
+        _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b);
+    const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w);
+
+    const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sx_sy_w16n_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0,
+                                         0xff, 0, 0xff, 0, 0xff, 0, 0xff);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_ral_b = xx_loadu_128(mask + 2 * c);
+      const __m128i v_rah_b = xx_loadu_128(mask + 2 * c + 16);
+      const __m128i v_rbl_b = xx_loadu_128(mask + mask_stride + 2 * c);
+      const __m128i v_rbh_b = xx_loadu_128(mask + mask_stride + 2 * c + 16);
+      const __m128i v_rvsl_b = _mm_add_epi8(v_ral_b, v_rbl_b);
+      const __m128i v_rvsh_b = _mm_add_epi8(v_rah_b, v_rbh_b);
+      const __m128i v_rvsal_w = _mm_and_si128(v_rvsl_b, v_zmask_b);
+      const __m128i v_rvsah_w = _mm_and_si128(v_rvsh_b, v_zmask_b);
+      const __m128i v_rvsbl_w =
+          _mm_and_si128(_mm_srli_si128(v_rvsl_b, 1), v_zmask_b);
+      const __m128i v_rvsbh_w =
+          _mm_and_si128(_mm_srli_si128(v_rvsh_b, 1), v_zmask_b);
+      const __m128i v_rsl_w = _mm_add_epi16(v_rvsal_w, v_rvsbl_w);
+      const __m128i v_rsh_w = _mm_add_epi16(v_rvsah_w, v_rvsbh_w);
+
+      const __m128i v_m0l_w = xx_roundn_epu16(v_rsl_w, 2);
+      const __m128i v_m0h_w = xx_roundn_epu16(v_rsh_w, 2);
+      const __m128i v_m1l_w = _mm_sub_epi16(v_maxval_w, v_m0l_w);
+      const __m128i v_m1h_w = _mm_sub_epi16(v_maxval_w, v_m0h_w);
+
+      const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0l_w, v_m1l_w);
+      const __m128i v_resh_w =
+          blend_8(src0 + c + 8, src1 + c + 8, v_m0h_w, v_m1h_w);
+
+      const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Dispatch
+//////////////////////////////////////////////////////////////////////////////
+
+void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                               const uint8_t *src0, uint32_t src0_stride,
+                               const uint8_t *src1, uint32_t src1_stride,
+                               const uint8_t *mask, uint32_t mask_stride, int h,
+                               int w, int suby, int subx) {
+  typedef void (*blend_fn)(
+      uint8_t * dst, uint32_t dst_stride, const uint8_t *src0,
+      uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+      const uint8_t *mask, uint32_t mask_stride, int h, int w);
+
+  // Dimensions are: width_index X subx X suby
+  static const blend_fn blend[3][2][2] = {
+    { // w % 16 == 0
+      { blend_a64_mask_w16n_sse4_1, blend_a64_mask_sy_w16n_sse4_1 },
+      { blend_a64_mask_sx_w16n_sse4_1, blend_a64_mask_sx_sy_w16n_sse4_1 } },
+    { // w == 4
+      { blend_a64_mask_w4_sse4_1, blend_a64_mask_sy_w4_sse4_1 },
+      { blend_a64_mask_sx_w4_sse4_1, blend_a64_mask_sx_sy_w4_sse4_1 } },
+    { // w == 8
+      { blend_a64_mask_w8_sse4_1, blend_a64_mask_sy_w8_sse4_1 },
+      { blend_a64_mask_sx_w8_sse4_1, blend_a64_mask_sx_sy_w8_sse4_1 } }
+  };
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  if (UNLIKELY((h | w) & 3)) {  // if (w <= 2 || h <= 2)
+    aom_blend_a64_mask_c(dst, dst_stride, src0, src0_stride, src1, src1_stride,
+                         mask, mask_stride, h, w, suby, subx);
+  } else {
+    blend[(w >> 2) & 3][subx != 0][suby != 0](dst, dst_stride, src0,
+                                              src0_stride, src1, src1_stride,
+                                              mask, mask_stride, h, w);
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+//////////////////////////////////////////////////////////////////////////////
+// No sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_mask_bn_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    const __m128i v_m0_b = xx_loadl_32(mask);
+    const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w);
+
+    xx_storel_64(dst, v_res_w);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  (void)w;
+  blend_a64_mask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                              src1_stride, mask, mask_stride, h, blend_4_b10);
+}
+
+static void blend_a64_mask_b12_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  (void)w;
+  blend_a64_mask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                              src1_stride, mask, mask_stride, h, blend_4_b12);
+}
+
+static INLINE void blend_a64_mask_bn_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 8) {
+      const __m128i v_m0_b = xx_loadl_64(mask + c);
+      const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b);
+      const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+      const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w);
+
+      xx_storeu_128(dst + c, v_res_w);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  blend_a64_mask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                               src1_stride, mask, mask_stride, h, w,
+                               blend_8_b10);
+}
+
+static void blend_a64_mask_b12_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  blend_a64_mask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                               src1_stride, mask, mask_stride, h, w,
+                               blend_8_b12);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Horizontal sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_mask_bn_sx_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) {
+  const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0,
+                                         0xff, 0, 0xff, 0, 0xff, 0, 0xff);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    const __m128i v_r_b = xx_loadl_64(mask);
+    const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1));
+
+    const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w);
+
+    xx_storel_64(dst, v_res_w);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sx_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  (void)w;
+  blend_a64_mask_bn_sx_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                 src1_stride, mask, mask_stride, h,
+                                 blend_4_b10);
+}
+
+static void blend_a64_mask_b12_sx_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  (void)w;
+  blend_a64_mask_bn_sx_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                 src1_stride, mask, mask_stride, h,
+                                 blend_4_b12);
+}
+
+static INLINE void blend_a64_mask_bn_sx_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    blend_unit_fn blend) {
+  const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0,
+                                         0xff, 0, 0xff, 0, 0xff, 0, 0xff);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 8) {
+      const __m128i v_r_b = xx_loadu_128(mask + 2 * c);
+      const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1));
+
+      const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b);
+      const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+      const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w);
+
+      xx_storeu_128(dst + c, v_res_w);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sx_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  blend_a64_mask_bn_sx_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, mask_stride, h, w,
+                                  blend_8_b10);
+}
+
+static void blend_a64_mask_b12_sx_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  blend_a64_mask_bn_sx_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, mask_stride, h, w,
+                                  blend_8_b12);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Vertical sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_mask_bn_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    const __m128i v_ra_b = xx_loadl_32(mask);
+    const __m128i v_rb_b = xx_loadl_32(mask + mask_stride);
+    const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+
+    const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w);
+
+    xx_storel_64(dst, v_res_w);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  (void)w;
+  blend_a64_mask_bn_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                 src1_stride, mask, mask_stride, h,
+                                 blend_4_b10);
+}
+
+static void blend_a64_mask_b12_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  (void)w;
+  blend_a64_mask_bn_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                 src1_stride, mask, mask_stride, h,
+                                 blend_4_b12);
+}
+
+static INLINE void blend_a64_mask_bn_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 8) {
+      const __m128i v_ra_b = xx_loadl_64(mask + c);
+      const __m128i v_rb_b = xx_loadl_64(mask + c + mask_stride);
+      const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+
+      const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b);
+      const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+      const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w);
+
+      xx_storeu_128(dst + c, v_res_w);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  blend_a64_mask_bn_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, mask_stride, h, w,
+                                  blend_8_b10);
+}
+
+static void blend_a64_mask_b12_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  blend_a64_mask_bn_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, mask_stride, h, w,
+                                  blend_8_b12);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Horizontal and Vertical sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_mask_bn_sx_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) {
+  const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0,
+                                         0xff, 0, 0xff, 0, 0xff, 0, 0xff);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    const __m128i v_ra_b = xx_loadl_64(mask);
+    const __m128i v_rb_b = xx_loadl_64(mask + mask_stride);
+    const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b);
+    const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b);
+    const __m128i v_rvsb_w =
+        _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b);
+    const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w);
+
+    const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w);
+
+    xx_storel_64(dst, v_res_w);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sx_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  (void)w;
+  blend_a64_mask_bn_sx_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                    src1_stride, mask, mask_stride, h,
+                                    blend_4_b10);
+}
+
+static void blend_a64_mask_b12_sx_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  (void)w;
+  blend_a64_mask_bn_sx_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                    src1_stride, mask, mask_stride, h,
+                                    blend_4_b12);
+}
+
+static INLINE void blend_a64_mask_bn_sx_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    blend_unit_fn blend) {
+  const __m128i v_zmask_b = _mm_set_epi8(0, 0xff, 0, 0xff, 0, 0xff, 0, 0xff, 0,
+                                         0xff, 0, 0xff, 0, 0xff, 0, 0xff);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 8) {
+      const __m128i v_ra_b = xx_loadu_128(mask + 2 * c);
+      const __m128i v_rb_b = xx_loadu_128(mask + 2 * c + mask_stride);
+      const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b);
+      const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b);
+      const __m128i v_rvsb_w =
+          _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b);
+      const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w);
+
+      const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2);
+      const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+      const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w);
+
+      xx_storeu_128(dst + c, v_res_w);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sx_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  blend_a64_mask_bn_sx_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                     src1_stride, mask, mask_stride, h, w,
+                                     blend_8_b10);
+}
+
+static void blend_a64_mask_b12_sx_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w) {
+  blend_a64_mask_bn_sx_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                     src1_stride, mask, mask_stride, h, w,
+                                     blend_8_b12);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Dispatch
+//////////////////////////////////////////////////////////////////////////////
+
+void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst_8, uint32_t dst_stride,
+                                      const uint8_t *src0_8,
+                                      uint32_t src0_stride,
+                                      const uint8_t *src1_8,
+                                      uint32_t src1_stride, const uint8_t *mask,
+                                      uint32_t mask_stride, int h, int w,
+                                      int suby, int subx, int bd) {
+  typedef void (*blend_fn)(
+      uint16_t * dst, uint32_t dst_stride, const uint16_t *src0,
+      uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+      const uint8_t *mask, uint32_t mask_stride, int h, int w);
+
+  // Dimensions are: bd_index X width_index X subx X suby
+  static const blend_fn blend[2][2][2][2] = {
+    {   // bd == 8 or 10
+      { // w % 8 == 0
+        { blend_a64_mask_b10_w8n_sse4_1, blend_a64_mask_b10_sy_w8n_sse4_1 },
+        { blend_a64_mask_b10_sx_w8n_sse4_1,
+          blend_a64_mask_b10_sx_sy_w8n_sse4_1 } },
+      { // w == 4
+        { blend_a64_mask_b10_w4_sse4_1, blend_a64_mask_b10_sy_w4_sse4_1 },
+        { blend_a64_mask_b10_sx_w4_sse4_1,
+          blend_a64_mask_b10_sx_sy_w4_sse4_1 } } },
+    {   // bd == 12
+      { // w % 8 == 0
+        { blend_a64_mask_b12_w8n_sse4_1, blend_a64_mask_b12_sy_w8n_sse4_1 },
+        { blend_a64_mask_b12_sx_w8n_sse4_1,
+          blend_a64_mask_b12_sx_sy_w8n_sse4_1 } },
+      { // w == 4
+        { blend_a64_mask_b12_w4_sse4_1, blend_a64_mask_b12_sy_w4_sse4_1 },
+        { blend_a64_mask_b12_sx_w4_sse4_1,
+          blend_a64_mask_b12_sx_sy_w4_sse4_1 } } }
+  };
+
+  assert(IMPLIES(src0_8 == dst_8, src0_stride == dst_stride));
+  assert(IMPLIES(src1_8 == dst_8, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+  if (UNLIKELY((h | w) & 3)) {  // if (w <= 2 || h <= 2)
+    aom_highbd_blend_a64_mask_c(dst_8, dst_stride, src0_8, src0_stride, src1_8,
+                                src1_stride, mask, mask_stride, h, w, suby,
+                                subx, bd);
+  } else {
+    uint16_t *const dst = CONVERT_TO_SHORTPTR(dst_8);
+    const uint16_t *const src0 = CONVERT_TO_SHORTPTR(src0_8);
+    const uint16_t *const src1 = CONVERT_TO_SHORTPTR(src1_8);
+
+    blend[bd == 12][(w >> 2) & 1][subx != 0][suby != 0](
+        dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+        mask_stride, h, w);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c
new file mode 100644
index 0000000000..9dabe5b798
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c
@@ -0,0 +1,285 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <smmintrin.h>  // SSE4.1
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/blend_sse4.h"
+
+#include "./aom_dsp_rtcd.h"
+
+//////////////////////////////////////////////////////////////////////////////
+// Implementation - No sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_vmask_w4_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                      const uint8_t *src0, uint32_t src0_stride,
+                                      const uint8_t *src1, uint32_t src1_stride,
+                                      const uint8_t *mask, int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_4(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+static void blend_a64_vmask_w8_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                      const uint8_t *src0, uint32_t src0_stride,
+                                      const uint8_t *src1, uint32_t src1_stride,
+                                      const uint8_t *mask, int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_8(src0, src1, v_m0_w, v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+static void blend_a64_vmask_w16n_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                        const uint8_t *src0,
+                                        uint32_t src0_stride,
+                                        const uint8_t *src1,
+                                        uint32_t src1_stride,
+                                        const uint8_t *mask, int h, int w) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_resl_w = blend_8(src0 + c, src1 + c, v_m0_w, v_m1_w);
+      const __m128i v_resh_w =
+          blend_8(src0 + c + 8, src1 + c + 8, v_m0_w, v_m1_w);
+
+      const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Dispatch
+//////////////////////////////////////////////////////////////////////////////
+
+void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                const uint8_t *src0, uint32_t src0_stride,
+                                const uint8_t *src1, uint32_t src1_stride,
+                                const uint8_t *mask, int h, int w) {
+  typedef void (*blend_fn)(uint8_t * dst, uint32_t dst_stride,
+                           const uint8_t *src0, uint32_t src0_stride,
+                           const uint8_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int h, int w);
+
+  // Dimension: width_index
+  static const blend_fn blend[9] = {
+    blend_a64_vmask_w16n_sse4_1,  // w % 16 == 0
+    aom_blend_a64_vmask_c,        // w == 1
+    aom_blend_a64_vmask_c,        // w == 2
+    NULL,                         // INVALID
+    blend_a64_vmask_w4_sse4_1,    // w == 4
+    NULL,                         // INVALID
+    NULL,                         // INVALID
+    NULL,                         // INVALID
+    blend_a64_vmask_w8_sse4_1,    // w == 8
+  };
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  blend[w & 0xf](dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, h,
+                 w);
+}
+
+#if CONFIG_HIGHBITDEPTH
+//////////////////////////////////////////////////////////////////////////////
+// Implementation - No sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_vmask_bn_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, int h, blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w);
+
+    xx_storel_64(dst, v_res_w);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+static void blend_a64_vmask_b10_w4_sse4_1(uint16_t *dst, uint32_t dst_stride,
+                                          const uint16_t *src0,
+                                          uint32_t src0_stride,
+                                          const uint16_t *src1,
+                                          uint32_t src1_stride,
+                                          const uint8_t *mask, int h, int w) {
+  (void)w;
+  blend_a64_vmask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                               src1_stride, mask, h, blend_4_b10);
+}
+
+static void blend_a64_vmask_b12_w4_sse4_1(uint16_t *dst, uint32_t dst_stride,
+                                          const uint16_t *src0,
+                                          uint32_t src0_stride,
+                                          const uint16_t *src1,
+                                          uint32_t src1_stride,
+                                          const uint8_t *mask, int h, int w) {
+  (void)w;
+  blend_a64_vmask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                               src1_stride, mask, h, blend_4_b12);
+}
+
+static INLINE void blend_a64_vmask_bn_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, int h, int w, blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+    for (c = 0; c < w; c += 8) {
+      const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w);
+
+      xx_storeu_128(dst + c, v_res_w);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+static void blend_a64_vmask_b10_w8n_sse4_1(uint16_t *dst, uint32_t dst_stride,
+                                           const uint16_t *src0,
+                                           uint32_t src0_stride,
+                                           const uint16_t *src1,
+                                           uint32_t src1_stride,
+                                           const uint8_t *mask, int h, int w) {
+  blend_a64_vmask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                src1_stride, mask, h, w, blend_8_b10);
+}
+
+static void blend_a64_vmask_b12_w8n_sse4_1(uint16_t *dst, uint32_t dst_stride,
+                                           const uint16_t *src0,
+                                           uint32_t src0_stride,
+                                           const uint16_t *src1,
+                                           uint32_t src1_stride,
+                                           const uint8_t *mask, int h, int w) {
+  blend_a64_vmask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                src1_stride, mask, h, w, blend_8_b12);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Dispatch
+//////////////////////////////////////////////////////////////////////////////
+
+void aom_highbd_blend_a64_vmask_sse4_1(
+    uint8_t *dst_8, uint32_t dst_stride, const uint8_t *src0_8,
+    uint32_t src0_stride, const uint8_t *src1_8, uint32_t src1_stride,
+    const uint8_t *mask, int h, int w, int bd) {
+  typedef void (*blend_fn)(uint16_t * dst, uint32_t dst_stride,
+                           const uint16_t *src0, uint32_t src0_stride,
+                           const uint16_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int h, int w);
+
+  // Dimensions are: bd_index X width_index
+  static const blend_fn blend[2][2] = {
+    {
+        // bd == 8 or 10
+        blend_a64_vmask_b10_w8n_sse4_1,  // w % 8 == 0
+        blend_a64_vmask_b10_w4_sse4_1,   // w == 4
+    },
+    {
+        // bd == 12
+        blend_a64_vmask_b12_w8n_sse4_1,  // w % 8 == 0
+        blend_a64_vmask_b12_w4_sse4_1,   // w == 4
+    }
+  };
+
+  assert(IMPLIES(src0_8 == dst_8, src0_stride == dst_stride));
+  assert(IMPLIES(src1_8 == dst_8, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  if (UNLIKELY((h | w) & 3)) {  // if (w <= 2 || h <= 2)
+    aom_highbd_blend_a64_vmask_c(dst_8, dst_stride, src0_8, src0_stride, src1_8,
+                                 src1_stride, mask, h, w, bd);
+  } else {
+    uint16_t *const dst = CONVERT_TO_SHORTPTR(dst_8);
+    const uint16_t *const src0 = CONVERT_TO_SHORTPTR(src0_8);
+    const uint16_t *const src1 = CONVERT_TO_SHORTPTR(src1_8);
+
+    blend[bd == 12][(w >> 2) & 1](dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, h, w);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/blend_sse4.h b/third_party/aom/aom_dsp/x86/blend_sse4.h
new file mode 100644
index 0000000000..daa2b2b3ae
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blend_sse4.h
@@ -0,0 +1,146 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_X86_BLEND_SSE4_H_
+#define AOM_DSP_X86_BLEND_SSE4_H_
+
+#include "aom_dsp/blend.h"
+#include "aom_dsp/x86/synonyms.h"
+
+//////////////////////////////////////////////////////////////////////////////
+// Common kernels
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE __m128i blend_4(const uint8_t *src0, const uint8_t *src1,
+                              const __m128i v_m0_w, const __m128i v_m1_w) {
+  const __m128i v_s0_b = xx_loadl_32(src0);
+  const __m128i v_s1_b = xx_loadl_32(src1);
+  const __m128i v_s0_w = _mm_cvtepu8_epi16(v_s0_b);
+  const __m128i v_s1_w = _mm_cvtepu8_epi16(v_s1_b);
+
+  const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w);
+  const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w);
+
+  const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w);
+
+  const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS);
+
+  return v_res_w;
+}
+
+static INLINE __m128i blend_8(const uint8_t *src0, const uint8_t *src1,
+                              const __m128i v_m0_w, const __m128i v_m1_w) {
+  const __m128i v_s0_b = xx_loadl_64(src0);
+  const __m128i v_s1_b = xx_loadl_64(src1);
+  const __m128i v_s0_w = _mm_cvtepu8_epi16(v_s0_b);
+  const __m128i v_s1_w = _mm_cvtepu8_epi16(v_s1_b);
+
+  const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w);
+  const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w);
+
+  const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w);
+
+  const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS);
+
+  return v_res_w;
+}
+
+#if CONFIG_HIGHBITDEPTH
+typedef __m128i (*blend_unit_fn)(const uint16_t *src0, const uint16_t *src1,
+                                 const __m128i v_m0_w, const __m128i v_m1_w);
+
+static INLINE __m128i blend_4_b10(const uint16_t *src0, const uint16_t *src1,
+                                  const __m128i v_m0_w, const __m128i v_m1_w) {
+  const __m128i v_s0_w = xx_loadl_64(src0);
+  const __m128i v_s1_w = xx_loadl_64(src1);
+
+  const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w);
+  const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w);
+
+  const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w);
+
+  const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS);
+
+  return v_res_w;
+}
+
+static INLINE __m128i blend_8_b10(const uint16_t *src0, const uint16_t *src1,
+                                  const __m128i v_m0_w, const __m128i v_m1_w) {
+  const __m128i v_s0_w = xx_loadu_128(src0);
+  const __m128i v_s1_w = xx_loadu_128(src1);
+
+  const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w);
+  const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w);
+
+  const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w);
+
+  const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS);
+
+  return v_res_w;
+}
+
+static INLINE __m128i blend_4_b12(const uint16_t *src0, const uint16_t *src1,
+                                  const __m128i v_m0_w, const __m128i v_m1_w) {
+  const __m128i v_s0_w = xx_loadl_64(src0);
+  const __m128i v_s1_w = xx_loadl_64(src1);
+
+  // Interleave
+  const __m128i v_m01_w = _mm_unpacklo_epi16(v_m0_w, v_m1_w);
+  const __m128i v_s01_w = _mm_unpacklo_epi16(v_s0_w, v_s1_w);
+
+  // Multiply-Add
+  const __m128i v_sum_d = _mm_madd_epi16(v_s01_w, v_m01_w);
+
+  // Scale
+  const __m128i v_ssum_d =
+      _mm_srli_epi32(v_sum_d, AOM_BLEND_A64_ROUND_BITS - 1);
+
+  // Pack
+  const __m128i v_pssum_d = _mm_packs_epi32(v_ssum_d, v_ssum_d);
+
+  // Round
+  const __m128i v_res_w = xx_round_epu16(v_pssum_d);
+
+  return v_res_w;
+}
+
+static INLINE __m128i blend_8_b12(const uint16_t *src0, const uint16_t *src1,
+                                  const __m128i v_m0_w, const __m128i v_m1_w) {
+  const __m128i v_s0_w = xx_loadu_128(src0);
+  const __m128i v_s1_w = xx_loadu_128(src1);
+
+  // Interleave
+  const __m128i v_m01l_w = _mm_unpacklo_epi16(v_m0_w, v_m1_w);
+  const __m128i v_m01h_w = _mm_unpackhi_epi16(v_m0_w, v_m1_w);
+  const __m128i v_s01l_w = _mm_unpacklo_epi16(v_s0_w, v_s1_w);
+  const __m128i v_s01h_w = _mm_unpackhi_epi16(v_s0_w, v_s1_w);
+
+  // Multiply-Add
+  const __m128i v_suml_d = _mm_madd_epi16(v_s01l_w, v_m01l_w);
+  const __m128i v_sumh_d = _mm_madd_epi16(v_s01h_w, v_m01h_w);
+
+  // Scale
+  const __m128i v_ssuml_d =
+      _mm_srli_epi32(v_suml_d, AOM_BLEND_A64_ROUND_BITS - 1);
+  const __m128i v_ssumh_d =
+      _mm_srli_epi32(v_sumh_d, AOM_BLEND_A64_ROUND_BITS - 1);
+
+  // Pack
+  const __m128i v_pssum_d = _mm_packs_epi32(v_ssuml_d, v_ssumh_d);
+
+  // Round
+  const __m128i v_res_w = xx_round_epu16(v_pssum_d);
+
+  return v_res_w;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#endif  // AOM_DSP_X86_BLEND_SSE4_H_
diff --git a/third_party/aom/aom_dsp/x86/convolve.h b/third_party/aom/aom_dsp/x86/convolve.h
new file mode 100644
index 0000000000..8641164db9
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/convolve.h
@@ -0,0 +1,288 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_DSP_X86_CONVOLVE_H_
+#define AOM_DSP_X86_CONVOLVE_H_
+
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_convolve.h"
+
+typedef void filter8_1dfunction(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                uint8_t *output_ptr, ptrdiff_t out_pitch,
+                                uint32_t output_height, const int16_t *filter);
+
+#define FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt)         \
+  void aom_convolve8_##name##_##opt(                                         \
+      const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,                \
+      ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,          \
+      const int16_t *filter_y, int y_step_q4, int w, int h) {                \
+    (void)filter_x;                                                          \
+    (void)x_step_q4;                                                         \
+    (void)filter_y;                                                          \
+    (void)y_step_q4;                                                         \
+    assert((-128 <= filter[3]) && (filter[3] <= 127));                       \
+    assert(step_q4 == 16);                                                   \
+    if (filter[0] | filter[1] | filter[2]) {                                 \
+      while (w >= 16) {                                                      \
+        aom_filter_block1d16_##dir##8_##avg##opt(src_start, src_stride, dst, \
+                                                 dst_stride, h, filter);     \
+        src += 16;                                                           \
+        dst += 16;                                                           \
+        w -= 16;                                                             \
+      }                                                                      \
+      while (w >= 8) {                                                       \
+        aom_filter_block1d8_##dir##8_##avg##opt(src_start, src_stride, dst,  \
+                                                dst_stride, h, filter);      \
+        src += 8;                                                            \
+        dst += 8;                                                            \
+        w -= 8;                                                              \
+      }                                                                      \
+      while (w >= 4) {                                                       \
+        aom_filter_block1d4_##dir##8_##avg##opt(src_start, src_stride, dst,  \
+                                                dst_stride, h, filter);      \
+        src += 4;                                                            \
+        dst += 4;                                                            \
+        w -= 4;                                                              \
+      }                                                                      \
+    } else {                                                                 \
+      while (w >= 16) {                                                      \
+        aom_filter_block1d16_##dir##2_##avg##opt(src, src_stride, dst,       \
+                                                 dst_stride, h, filter);     \
+        src += 16;                                                           \
+        dst += 16;                                                           \
+        w -= 16;                                                             \
+      }                                                                      \
+      while (w >= 8) {                                                       \
+        aom_filter_block1d8_##dir##2_##avg##opt(src, src_stride, dst,        \
+                                                dst_stride, h, filter);      \
+        src += 8;                                                            \
+        dst += 8;                                                            \
+        w -= 8;                                                              \
+      }                                                                      \
+      while (w >= 4) {                                                       \
+        aom_filter_block1d4_##dir##2_##avg##opt(src, src_stride, dst,        \
+                                                dst_stride, h, filter);      \
+        src += 4;                                                            \
+        dst += 4;                                                            \
+        w -= 4;                                                              \
+      }                                                                      \
+    }                                                                        \
+    if (w) {                                                                 \
+      aom_convolve8_##name##_c(src, src_stride, dst, dst_stride, filter_x,   \
+                               x_step_q4, filter_y, y_step_q4, w, h);        \
+    }                                                                        \
+  }
+
+#define FUN_CONV_2D(avg, opt)                                                \
+  void aom_convolve8_##avg##opt(                                             \
+      const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,                \
+      ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,          \
+      const int16_t *filter_y, int y_step_q4, int w, int h) {                \
+    assert((-128 <= filter_x[3]) && (filter_x[3] <= 127));                   \
+    assert((-128 <= filter_y[3]) && (filter_y[3] <= 127));                   \
+    assert(w <= MAX_SB_SIZE);                                                \
+    assert(h <= MAX_SB_SIZE);                                                \
+    assert(x_step_q4 == 16);                                                 \
+    assert(y_step_q4 == 16);                                                 \
+    if (filter_x[0] || filter_x[1] || filter_x[2] || filter_y[0] ||          \
+        filter_y[1] || filter_y[2]) {                                        \
+      DECLARE_ALIGNED(16, uint8_t, fdata2[MAX_SB_SIZE * (MAX_SB_SIZE + 7)]); \
+      aom_convolve8_horiz_##opt(src - 3 * src_stride, src_stride, fdata2,    \
+                                MAX_SB_SIZE, filter_x, x_step_q4, filter_y,  \
+                                y_step_q4, w, h + 7);                        \
+      aom_convolve8_##avg##vert_##opt(fdata2 + 3 * MAX_SB_SIZE, MAX_SB_SIZE, \
+                                      dst, dst_stride, filter_x, x_step_q4,  \
+                                      filter_y, y_step_q4, w, h);            \
+    } else {                                                                 \
+      DECLARE_ALIGNED(16, uint8_t, fdata2[MAX_SB_SIZE * (MAX_SB_SIZE + 1)]); \
+      aom_convolve8_horiz_##opt(src, src_stride, fdata2, MAX_SB_SIZE,        \
+                                filter_x, x_step_q4, filter_y, y_step_q4, w, \
+                                h + 1);                                      \
+      aom_convolve8_##avg##vert_##opt(fdata2, MAX_SB_SIZE, dst, dst_stride,  \
+                                      filter_x, x_step_q4, filter_y,         \
+                                      y_step_q4, w, h);                      \
+    }                                                                        \
+  }
+
+#if CONFIG_LOOP_RESTORATION
+// convolve_add_src is only used by the Wiener filter, which will never
+// end up calling the bilinear functions (it uses a symmetric filter, so
+// the possible numbers of taps are 1,3,5,7)
+#define FUN_CONV_1D_NO_BILINEAR(name, step_q4, filter, dir, src_start, avg, \
+                                opt)                                        \
+  void aom_convolve8_##name##_##opt(                                        \
+      const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,               \
+      ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,         \
+      const int16_t *filter_y, int y_step_q4, int w, int h) {               \
+    (void)filter_x;                                                         \
+    (void)x_step_q4;                                                        \
+    (void)filter_y;                                                         \
+    (void)y_step_q4;                                                        \
+    assert((-128 <= filter[3]) && (filter[3] <= 127));                      \
+    assert(step_q4 == 16);                                                  \
+    while (w >= 16) {                                                       \
+      aom_filter_block1d16_##dir##8_##avg##opt(src_start, src_stride, dst,  \
+                                               dst_stride, h, filter);      \
+      src += 16;                                                            \
+      dst += 16;                                                            \
+      w -= 16;                                                              \
+    }                                                                       \
+    while (w >= 8) {                                                        \
+      aom_filter_block1d8_##dir##8_##avg##opt(src_start, src_stride, dst,   \
+                                              dst_stride, h, filter);       \
+      src += 8;                                                             \
+      dst += 8;                                                             \
+      w -= 8;                                                               \
+    }                                                                       \
+    while (w >= 4) {                                                        \
+      aom_filter_block1d4_##dir##8_##avg##opt(src_start, src_stride, dst,   \
+                                              dst_stride, h, filter);       \
+      src += 4;                                                             \
+      dst += 4;                                                             \
+      w -= 4;                                                               \
+    }                                                                       \
+    if (w) {                                                                \
+      aom_convolve8_##name##_c(src, src_stride, dst, dst_stride, filter_x,  \
+                               x_step_q4, filter_y, y_step_q4, w, h);       \
+    }                                                                       \
+  }
+
+#define FUN_CONV_2D_NO_BILINEAR(type, htype, opt)                           \
+  void aom_convolve8_##type##opt(                                           \
+      const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,               \
+      ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,         \
+      const int16_t *filter_y, int y_step_q4, int w, int h) {               \
+    DECLARE_ALIGNED(16, uint8_t, fdata2[MAX_SB_SIZE * (MAX_SB_SIZE + 7)]);  \
+    assert((-128 <= filter_x[3]) && (filter_x[3] <= 127));                  \
+    assert((-128 <= filter_y[3]) && (filter_y[3] <= 127));                  \
+    assert(w <= MAX_SB_SIZE);                                               \
+    assert(h <= MAX_SB_SIZE);                                               \
+    assert(x_step_q4 == 16);                                                \
+    assert(y_step_q4 == 16);                                                \
+    aom_convolve8_##htype##horiz_##opt(                                     \
+        src - 3 * src_stride, src_stride, fdata2, MAX_SB_SIZE, filter_x,    \
+        x_step_q4, filter_y, y_step_q4, w, h + 7);                          \
+    aom_convolve8_##type##vert_##opt(fdata2 + 3 * MAX_SB_SIZE, MAX_SB_SIZE, \
+                                     dst, dst_stride, filter_x, x_step_q4,  \
+                                     filter_y, y_step_q4, w, h);            \
+  }
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+typedef void highbd_filter8_1dfunction(const uint16_t *src_ptr,
+                                       const ptrdiff_t src_pitch,
+                                       uint16_t *output_ptr,
+                                       ptrdiff_t out_pitch,
+                                       unsigned int output_height,
+                                       const int16_t *filter, int bd);
+
+#define HIGH_FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt)  \
+  void aom_highbd_convolve8_##name##_##opt(                                \
+      const uint8_t *src8, ptrdiff_t src_stride, uint8_t *dst8,            \
+      ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,        \
+      const int16_t *filter_y, int y_step_q4, int w, int h, int bd) {      \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                             \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                             \
+    if (step_q4 == 16 && filter[3] != 128) {                               \
+      if (filter[0] | filter[1] | filter[2]) {                             \
+        while (w >= 16) {                                                  \
+          aom_highbd_filter_block1d16_##dir##8_##avg##opt(                 \
+              src_start, src_stride, dst, dst_stride, h, filter, bd);      \
+          src += 16;                                                       \
+          dst += 16;                                                       \
+          w -= 16;                                                         \
+        }                                                                  \
+        while (w >= 8) {                                                   \
+          aom_highbd_filter_block1d8_##dir##8_##avg##opt(                  \
+              src_start, src_stride, dst, dst_stride, h, filter, bd);      \
+          src += 8;                                                        \
+          dst += 8;                                                        \
+          w -= 8;                                                          \
+        }                                                                  \
+        while (w >= 4) {                                                   \
+          aom_highbd_filter_block1d4_##dir##8_##avg##opt(                  \
+              src_start, src_stride, dst, dst_stride, h, filter, bd);      \
+          src += 4;                                                        \
+          dst += 4;                                                        \
+          w -= 4;                                                          \
+        }                                                                  \
+      } else {                                                             \
+        while (w >= 16) {                                                  \
+          aom_highbd_filter_block1d16_##dir##2_##avg##opt(                 \
+              src, src_stride, dst, dst_stride, h, filter, bd);            \
+          src += 16;                                                       \
+          dst += 16;                                                       \
+          w -= 16;                                                         \
+        }                                                                  \
+        while (w >= 8) {                                                   \
+          aom_highbd_filter_block1d8_##dir##2_##avg##opt(                  \
+              src, src_stride, dst, dst_stride, h, filter, bd);            \
+          src += 8;                                                        \
+          dst += 8;                                                        \
+          w -= 8;                                                          \
+        }                                                                  \
+        while (w >= 4) {                                                   \
+          aom_highbd_filter_block1d4_##dir##2_##avg##opt(                  \
+              src, src_stride, dst, dst_stride, h, filter, bd);            \
+          src += 4;                                                        \
+          dst += 4;                                                        \
+          w -= 4;                                                          \
+        }                                                                  \
+      }                                                                    \
+    }                                                                      \
+    if (w) {                                                               \
+      aom_highbd_convolve8_##name##_c(                                     \
+          CONVERT_TO_BYTEPTR(src), src_stride, CONVERT_TO_BYTEPTR(dst),    \
+          dst_stride, filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd); \
+    }                                                                      \
+  }
+
+#define HIGH_FUN_CONV_2D(avg, opt)                                            \
+  void aom_highbd_convolve8_##avg##opt(                                       \
+      const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,                 \
+      ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,           \
+      const int16_t *filter_y, int y_step_q4, int w, int h, int bd) {         \
+    assert(w <= MAX_SB_SIZE);                                                 \
+    assert(h <= MAX_SB_SIZE);                                                 \
+    if (x_step_q4 == 16 && y_step_q4 == 16) {                                 \
+      if (filter_x[0] || filter_x[1] || filter_x[2] || filter_x[3] == 128 ||  \
+          filter_y[0] || filter_y[1] || filter_y[2] || filter_y[3] == 128) {  \
+        DECLARE_ALIGNED(16, uint16_t,                                         \
+                        fdata2[MAX_SB_SIZE * (MAX_SB_SIZE + 7)]);             \
+        aom_highbd_convolve8_horiz_##opt(src - 3 * src_stride, src_stride,    \
+                                         CONVERT_TO_BYTEPTR(fdata2),          \
+                                         MAX_SB_SIZE, filter_x, x_step_q4,    \
+                                         filter_y, y_step_q4, w, h + 7, bd);  \
+        aom_highbd_convolve8_##avg##vert_##opt(                               \
+            CONVERT_TO_BYTEPTR(fdata2) + 3 * MAX_SB_SIZE, MAX_SB_SIZE, dst,   \
+            dst_stride, filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);  \
+      } else {                                                                \
+        DECLARE_ALIGNED(16, uint16_t,                                         \
+                        fdata2[MAX_SB_SIZE * (MAX_SB_SIZE + 1)]);             \
+        aom_highbd_convolve8_horiz_##opt(                                     \
+            src, src_stride, CONVERT_TO_BYTEPTR(fdata2), MAX_SB_SIZE,         \
+            filter_x, x_step_q4, filter_y, y_step_q4, w, h + 1, bd);          \
+        aom_highbd_convolve8_##avg##vert_##opt(                               \
+            CONVERT_TO_BYTEPTR(fdata2), MAX_SB_SIZE, dst, dst_stride,         \
+            filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);              \
+      }                                                                       \
+    } else {                                                                  \
+      aom_highbd_convolve8_##avg##c(src, src_stride, dst, dst_stride,         \
+                                    filter_x, x_step_q4, filter_y, y_step_q4, \
+                                    w, h, bd);                                \
+    }                                                                         \
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+#endif  // AOM_DSP_X86_CONVOLVE_H_
diff --git a/third_party/aom/aom_dsp/x86/fwd_dct32_8cols_sse2.c b/third_party/aom/aom_dsp/x86/fwd_dct32_8cols_sse2.c
new file mode 100644
index 0000000000..b8ec08de79
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_dct32_8cols_sse2.c
@@ -0,0 +1,862 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "aom_dsp/fwd_txfm.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+
+// Apply a 32-element IDCT to 8 columns. This does not do any transposition
+// of its output - the caller is expected to do that.
+// The input buffers are the top and bottom halves of an 8x32 block.
+void fdct32_8col(__m128i *in0, __m128i *in1) {
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(+cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(+cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(+cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(+cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m28_m04 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_m12_m20 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_p30_p02 = pair_set_epi16(+cospi_30_64, cospi_2_64);
+  const __m128i k__cospi_p14_p18 = pair_set_epi16(+cospi_14_64, cospi_18_64);
+  const __m128i k__cospi_p22_p10 = pair_set_epi16(+cospi_22_64, cospi_10_64);
+  const __m128i k__cospi_p06_p26 = pair_set_epi16(+cospi_6_64, cospi_26_64);
+  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p31_p01 = pair_set_epi16(+cospi_31_64, cospi_1_64);
+  const __m128i k__cospi_p15_p17 = pair_set_epi16(+cospi_15_64, cospi_17_64);
+  const __m128i k__cospi_p23_p09 = pair_set_epi16(+cospi_23_64, cospi_9_64);
+  const __m128i k__cospi_p07_p25 = pair_set_epi16(+cospi_7_64, cospi_25_64);
+  const __m128i k__cospi_m25_p07 = pair_set_epi16(-cospi_25_64, cospi_7_64);
+  const __m128i k__cospi_m09_p23 = pair_set_epi16(-cospi_9_64, cospi_23_64);
+  const __m128i k__cospi_m17_p15 = pair_set_epi16(-cospi_17_64, cospi_15_64);
+  const __m128i k__cospi_m01_p31 = pair_set_epi16(-cospi_1_64, cospi_31_64);
+  const __m128i k__cospi_p27_p05 = pair_set_epi16(+cospi_27_64, cospi_5_64);
+  const __m128i k__cospi_p11_p21 = pair_set_epi16(+cospi_11_64, cospi_21_64);
+  const __m128i k__cospi_p19_p13 = pair_set_epi16(+cospi_19_64, cospi_13_64);
+  const __m128i k__cospi_p03_p29 = pair_set_epi16(+cospi_3_64, cospi_29_64);
+  const __m128i k__cospi_m29_p03 = pair_set_epi16(-cospi_29_64, cospi_3_64);
+  const __m128i k__cospi_m13_p19 = pair_set_epi16(-cospi_13_64, cospi_19_64);
+  const __m128i k__cospi_m21_p11 = pair_set_epi16(-cospi_21_64, cospi_11_64);
+  const __m128i k__cospi_m05_p27 = pair_set_epi16(-cospi_5_64, cospi_27_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i step1[32];
+  __m128i step2[32];
+  __m128i step3[32];
+  __m128i out[32];
+  // Stage 1
+  {
+    const __m128i *ina = in0;
+    const __m128i *inb = in1 + 15;
+    __m128i *step1a = &step1[0];
+    __m128i *step1b = &step1[31];
+    const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
+    const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + 1));
+    const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + 2));
+    const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + 3));
+    const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - 3));
+    const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - 2));
+    const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - 1));
+    const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
+    step1a[0] = _mm_add_epi16(ina0, inb0);
+    step1a[1] = _mm_add_epi16(ina1, inb1);
+    step1a[2] = _mm_add_epi16(ina2, inb2);
+    step1a[3] = _mm_add_epi16(ina3, inb3);
+    step1b[-3] = _mm_sub_epi16(ina3, inb3);
+    step1b[-2] = _mm_sub_epi16(ina2, inb2);
+    step1b[-1] = _mm_sub_epi16(ina1, inb1);
+    step1b[-0] = _mm_sub_epi16(ina0, inb0);
+  }
+  {
+    const __m128i *ina = in0 + 4;
+    const __m128i *inb = in1 + 11;
+    __m128i *step1a = &step1[4];
+    __m128i *step1b = &step1[27];
+    const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
+    const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + 1));
+    const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + 2));
+    const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + 3));
+    const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - 3));
+    const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - 2));
+    const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - 1));
+    const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
+    step1a[0] = _mm_add_epi16(ina0, inb0);
+    step1a[1] = _mm_add_epi16(ina1, inb1);
+    step1a[2] = _mm_add_epi16(ina2, inb2);
+    step1a[3] = _mm_add_epi16(ina3, inb3);
+    step1b[-3] = _mm_sub_epi16(ina3, inb3);
+    step1b[-2] = _mm_sub_epi16(ina2, inb2);
+    step1b[-1] = _mm_sub_epi16(ina1, inb1);
+    step1b[-0] = _mm_sub_epi16(ina0, inb0);
+  }
+  {
+    const __m128i *ina = in0 + 8;
+    const __m128i *inb = in1 + 7;
+    __m128i *step1a = &step1[8];
+    __m128i *step1b = &step1[23];
+    const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
+    const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + 1));
+    const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + 2));
+    const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + 3));
+    const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - 3));
+    const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - 2));
+    const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - 1));
+    const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
+    step1a[0] = _mm_add_epi16(ina0, inb0);
+    step1a[1] = _mm_add_epi16(ina1, inb1);
+    step1a[2] = _mm_add_epi16(ina2, inb2);
+    step1a[3] = _mm_add_epi16(ina3, inb3);
+    step1b[-3] = _mm_sub_epi16(ina3, inb3);
+    step1b[-2] = _mm_sub_epi16(ina2, inb2);
+    step1b[-1] = _mm_sub_epi16(ina1, inb1);
+    step1b[-0] = _mm_sub_epi16(ina0, inb0);
+  }
+  {
+    const __m128i *ina = in0 + 12;
+    const __m128i *inb = in1 + 3;
+    __m128i *step1a = &step1[12];
+    __m128i *step1b = &step1[19];
+    const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
+    const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + 1));
+    const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + 2));
+    const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + 3));
+    const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - 3));
+    const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - 2));
+    const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - 1));
+    const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
+    step1a[0] = _mm_add_epi16(ina0, inb0);
+    step1a[1] = _mm_add_epi16(ina1, inb1);
+    step1a[2] = _mm_add_epi16(ina2, inb2);
+    step1a[3] = _mm_add_epi16(ina3, inb3);
+    step1b[-3] = _mm_sub_epi16(ina3, inb3);
+    step1b[-2] = _mm_sub_epi16(ina2, inb2);
+    step1b[-1] = _mm_sub_epi16(ina1, inb1);
+    step1b[-0] = _mm_sub_epi16(ina0, inb0);
+  }
+  // Stage 2
+  {
+    step2[0] = _mm_add_epi16(step1[0], step1[15]);
+    step2[1] = _mm_add_epi16(step1[1], step1[14]);
+    step2[2] = _mm_add_epi16(step1[2], step1[13]);
+    step2[3] = _mm_add_epi16(step1[3], step1[12]);
+    step2[4] = _mm_add_epi16(step1[4], step1[11]);
+    step2[5] = _mm_add_epi16(step1[5], step1[10]);
+    step2[6] = _mm_add_epi16(step1[6], step1[9]);
+    step2[7] = _mm_add_epi16(step1[7], step1[8]);
+    step2[8] = _mm_sub_epi16(step1[7], step1[8]);
+    step2[9] = _mm_sub_epi16(step1[6], step1[9]);
+    step2[10] = _mm_sub_epi16(step1[5], step1[10]);
+    step2[11] = _mm_sub_epi16(step1[4], step1[11]);
+    step2[12] = _mm_sub_epi16(step1[3], step1[12]);
+    step2[13] = _mm_sub_epi16(step1[2], step1[13]);
+    step2[14] = _mm_sub_epi16(step1[1], step1[14]);
+    step2[15] = _mm_sub_epi16(step1[0], step1[15]);
+  }
+  {
+    const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]);
+    const __m128i s2_20_1 = _mm_unpackhi_epi16(step1[27], step1[20]);
+    const __m128i s2_21_0 = _mm_unpacklo_epi16(step1[26], step1[21]);
+    const __m128i s2_21_1 = _mm_unpackhi_epi16(step1[26], step1[21]);
+    const __m128i s2_22_0 = _mm_unpacklo_epi16(step1[25], step1[22]);
+    const __m128i s2_22_1 = _mm_unpackhi_epi16(step1[25], step1[22]);
+    const __m128i s2_23_0 = _mm_unpacklo_epi16(step1[24], step1[23]);
+    const __m128i s2_23_1 = _mm_unpackhi_epi16(step1[24], step1[23]);
+    const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
+    const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
+    const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
+    const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
+    const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
+    const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
+    const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
+    const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
+    const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
+    const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16);
+    const __m128i s2_25_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_p16);
+    const __m128i s2_25_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_p16);
+    const __m128i s2_26_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_p16);
+    const __m128i s2_26_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_p16);
+    const __m128i s2_27_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_p16);
+    const __m128i s2_27_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_p16);
+    // dct_const_round_shift
+    const __m128i s2_20_4 = _mm_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_20_5 = _mm_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_21_4 = _mm_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_21_5 = _mm_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_22_4 = _mm_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_22_5 = _mm_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_23_4 = _mm_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_23_5 = _mm_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_24_4 = _mm_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_24_5 = _mm_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_25_4 = _mm_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_25_5 = _mm_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_26_4 = _mm_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_26_5 = _mm_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_27_4 = _mm_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_27_5 = _mm_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_20_6 = _mm_srai_epi32(s2_20_4, DCT_CONST_BITS);
+    const __m128i s2_20_7 = _mm_srai_epi32(s2_20_5, DCT_CONST_BITS);
+    const __m128i s2_21_6 = _mm_srai_epi32(s2_21_4, DCT_CONST_BITS);
+    const __m128i s2_21_7 = _mm_srai_epi32(s2_21_5, DCT_CONST_BITS);
+    const __m128i s2_22_6 = _mm_srai_epi32(s2_22_4, DCT_CONST_BITS);
+    const __m128i s2_22_7 = _mm_srai_epi32(s2_22_5, DCT_CONST_BITS);
+    const __m128i s2_23_6 = _mm_srai_epi32(s2_23_4, DCT_CONST_BITS);
+    const __m128i s2_23_7 = _mm_srai_epi32(s2_23_5, DCT_CONST_BITS);
+    const __m128i s2_24_6 = _mm_srai_epi32(s2_24_4, DCT_CONST_BITS);
+    const __m128i s2_24_7 = _mm_srai_epi32(s2_24_5, DCT_CONST_BITS);
+    const __m128i s2_25_6 = _mm_srai_epi32(s2_25_4, DCT_CONST_BITS);
+    const __m128i s2_25_7 = _mm_srai_epi32(s2_25_5, DCT_CONST_BITS);
+    const __m128i s2_26_6 = _mm_srai_epi32(s2_26_4, DCT_CONST_BITS);
+    const __m128i s2_26_7 = _mm_srai_epi32(s2_26_5, DCT_CONST_BITS);
+    const __m128i s2_27_6 = _mm_srai_epi32(s2_27_4, DCT_CONST_BITS);
+    const __m128i s2_27_7 = _mm_srai_epi32(s2_27_5, DCT_CONST_BITS);
+    // Combine
+    step2[20] = _mm_packs_epi32(s2_20_6, s2_20_7);
+    step2[21] = _mm_packs_epi32(s2_21_6, s2_21_7);
+    step2[22] = _mm_packs_epi32(s2_22_6, s2_22_7);
+    step2[23] = _mm_packs_epi32(s2_23_6, s2_23_7);
+    step2[24] = _mm_packs_epi32(s2_24_6, s2_24_7);
+    step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7);
+    step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7);
+    step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7);
+  }
+  // Stage 3
+  {
+    step3[0] = _mm_add_epi16(step2[(8 - 1)], step2[0]);
+    step3[1] = _mm_add_epi16(step2[(8 - 2)], step2[1]);
+    step3[2] = _mm_add_epi16(step2[(8 - 3)], step2[2]);
+    step3[3] = _mm_add_epi16(step2[(8 - 4)], step2[3]);
+    step3[4] = _mm_sub_epi16(step2[(8 - 5)], step2[4]);
+    step3[5] = _mm_sub_epi16(step2[(8 - 6)], step2[5]);
+    step3[6] = _mm_sub_epi16(step2[(8 - 7)], step2[6]);
+    step3[7] = _mm_sub_epi16(step2[(8 - 8)], step2[7]);
+  }
+  {
+    const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
+    const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
+    const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
+    const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
+    const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
+    const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
+    const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
+    const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
+    const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
+    const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
+    const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
+    const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
+    // dct_const_round_shift
+    const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_10_6 = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
+    const __m128i s3_10_7 = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
+    const __m128i s3_11_6 = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
+    const __m128i s3_11_7 = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
+    const __m128i s3_12_6 = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
+    const __m128i s3_12_7 = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
+    const __m128i s3_13_6 = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
+    const __m128i s3_13_7 = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
+    // Combine
+    step3[10] = _mm_packs_epi32(s3_10_6, s3_10_7);
+    step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7);
+    step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7);
+    step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7);
+  }
+  {
+    step3[16] = _mm_add_epi16(step2[23], step1[16]);
+    step3[17] = _mm_add_epi16(step2[22], step1[17]);
+    step3[18] = _mm_add_epi16(step2[21], step1[18]);
+    step3[19] = _mm_add_epi16(step2[20], step1[19]);
+    step3[20] = _mm_sub_epi16(step1[19], step2[20]);
+    step3[21] = _mm_sub_epi16(step1[18], step2[21]);
+    step3[22] = _mm_sub_epi16(step1[17], step2[22]);
+    step3[23] = _mm_sub_epi16(step1[16], step2[23]);
+    step3[24] = _mm_sub_epi16(step1[31], step2[24]);
+    step3[25] = _mm_sub_epi16(step1[30], step2[25]);
+    step3[26] = _mm_sub_epi16(step1[29], step2[26]);
+    step3[27] = _mm_sub_epi16(step1[28], step2[27]);
+    step3[28] = _mm_add_epi16(step2[27], step1[28]);
+    step3[29] = _mm_add_epi16(step2[26], step1[29]);
+    step3[30] = _mm_add_epi16(step2[25], step1[30]);
+    step3[31] = _mm_add_epi16(step2[24], step1[31]);
+  }
+
+  // Stage 4
+  {
+    step1[0] = _mm_add_epi16(step3[3], step3[0]);
+    step1[1] = _mm_add_epi16(step3[2], step3[1]);
+    step1[2] = _mm_sub_epi16(step3[1], step3[2]);
+    step1[3] = _mm_sub_epi16(step3[0], step3[3]);
+    step1[8] = _mm_add_epi16(step3[11], step2[8]);
+    step1[9] = _mm_add_epi16(step3[10], step2[9]);
+    step1[10] = _mm_sub_epi16(step2[9], step3[10]);
+    step1[11] = _mm_sub_epi16(step2[8], step3[11]);
+    step1[12] = _mm_sub_epi16(step2[15], step3[12]);
+    step1[13] = _mm_sub_epi16(step2[14], step3[13]);
+    step1[14] = _mm_add_epi16(step3[13], step2[14]);
+    step1[15] = _mm_add_epi16(step3[12], step2[15]);
+  }
+  {
+    const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]);
+    const __m128i s1_05_1 = _mm_unpackhi_epi16(step3[6], step3[5]);
+    const __m128i s1_05_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_m16);
+    const __m128i s1_05_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_m16);
+    const __m128i s1_06_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_p16);
+    const __m128i s1_06_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_p16);
+    // dct_const_round_shift
+    const __m128i s1_05_4 = _mm_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
+    const __m128i s1_05_5 = _mm_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
+    const __m128i s1_06_4 = _mm_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
+    const __m128i s1_06_5 = _mm_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
+    const __m128i s1_05_6 = _mm_srai_epi32(s1_05_4, DCT_CONST_BITS);
+    const __m128i s1_05_7 = _mm_srai_epi32(s1_05_5, DCT_CONST_BITS);
+    const __m128i s1_06_6 = _mm_srai_epi32(s1_06_4, DCT_CONST_BITS);
+    const __m128i s1_06_7 = _mm_srai_epi32(s1_06_5, DCT_CONST_BITS);
+    // Combine
+    step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7);
+    step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7);
+  }
+  {
+    const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]);
+    const __m128i s1_18_1 = _mm_unpackhi_epi16(step3[18], step3[29]);
+    const __m128i s1_19_0 = _mm_unpacklo_epi16(step3[19], step3[28]);
+    const __m128i s1_19_1 = _mm_unpackhi_epi16(step3[19], step3[28]);
+    const __m128i s1_20_0 = _mm_unpacklo_epi16(step3[20], step3[27]);
+    const __m128i s1_20_1 = _mm_unpackhi_epi16(step3[20], step3[27]);
+    const __m128i s1_21_0 = _mm_unpacklo_epi16(step3[21], step3[26]);
+    const __m128i s1_21_1 = _mm_unpackhi_epi16(step3[21], step3[26]);
+    const __m128i s1_18_2 = _mm_madd_epi16(s1_18_0, k__cospi_m08_p24);
+    const __m128i s1_18_3 = _mm_madd_epi16(s1_18_1, k__cospi_m08_p24);
+    const __m128i s1_19_2 = _mm_madd_epi16(s1_19_0, k__cospi_m08_p24);
+    const __m128i s1_19_3 = _mm_madd_epi16(s1_19_1, k__cospi_m08_p24);
+    const __m128i s1_20_2 = _mm_madd_epi16(s1_20_0, k__cospi_m24_m08);
+    const __m128i s1_20_3 = _mm_madd_epi16(s1_20_1, k__cospi_m24_m08);
+    const __m128i s1_21_2 = _mm_madd_epi16(s1_21_0, k__cospi_m24_m08);
+    const __m128i s1_21_3 = _mm_madd_epi16(s1_21_1, k__cospi_m24_m08);
+    const __m128i s1_26_2 = _mm_madd_epi16(s1_21_0, k__cospi_m08_p24);
+    const __m128i s1_26_3 = _mm_madd_epi16(s1_21_1, k__cospi_m08_p24);
+    const __m128i s1_27_2 = _mm_madd_epi16(s1_20_0, k__cospi_m08_p24);
+    const __m128i s1_27_3 = _mm_madd_epi16(s1_20_1, k__cospi_m08_p24);
+    const __m128i s1_28_2 = _mm_madd_epi16(s1_19_0, k__cospi_p24_p08);
+    const __m128i s1_28_3 = _mm_madd_epi16(s1_19_1, k__cospi_p24_p08);
+    const __m128i s1_29_2 = _mm_madd_epi16(s1_18_0, k__cospi_p24_p08);
+    const __m128i s1_29_3 = _mm_madd_epi16(s1_18_1, k__cospi_p24_p08);
+    // dct_const_round_shift
+    const __m128i s1_18_4 = _mm_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
+    const __m128i s1_18_5 = _mm_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
+    const __m128i s1_19_4 = _mm_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
+    const __m128i s1_19_5 = _mm_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
+    const __m128i s1_20_4 = _mm_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
+    const __m128i s1_20_5 = _mm_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
+    const __m128i s1_21_4 = _mm_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
+    const __m128i s1_21_5 = _mm_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
+    const __m128i s1_26_4 = _mm_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
+    const __m128i s1_26_5 = _mm_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
+    const __m128i s1_27_4 = _mm_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
+    const __m128i s1_27_5 = _mm_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
+    const __m128i s1_28_4 = _mm_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
+    const __m128i s1_28_5 = _mm_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
+    const __m128i s1_29_4 = _mm_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
+    const __m128i s1_29_5 = _mm_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
+    const __m128i s1_18_6 = _mm_srai_epi32(s1_18_4, DCT_CONST_BITS);
+    const __m128i s1_18_7 = _mm_srai_epi32(s1_18_5, DCT_CONST_BITS);
+    const __m128i s1_19_6 = _mm_srai_epi32(s1_19_4, DCT_CONST_BITS);
+    const __m128i s1_19_7 = _mm_srai_epi32(s1_19_5, DCT_CONST_BITS);
+    const __m128i s1_20_6 = _mm_srai_epi32(s1_20_4, DCT_CONST_BITS);
+    const __m128i s1_20_7 = _mm_srai_epi32(s1_20_5, DCT_CONST_BITS);
+    const __m128i s1_21_6 = _mm_srai_epi32(s1_21_4, DCT_CONST_BITS);
+    const __m128i s1_21_7 = _mm_srai_epi32(s1_21_5, DCT_CONST_BITS);
+    const __m128i s1_26_6 = _mm_srai_epi32(s1_26_4, DCT_CONST_BITS);
+    const __m128i s1_26_7 = _mm_srai_epi32(s1_26_5, DCT_CONST_BITS);
+    const __m128i s1_27_6 = _mm_srai_epi32(s1_27_4, DCT_CONST_BITS);
+    const __m128i s1_27_7 = _mm_srai_epi32(s1_27_5, DCT_CONST_BITS);
+    const __m128i s1_28_6 = _mm_srai_epi32(s1_28_4, DCT_CONST_BITS);
+    const __m128i s1_28_7 = _mm_srai_epi32(s1_28_5, DCT_CONST_BITS);
+    const __m128i s1_29_6 = _mm_srai_epi32(s1_29_4, DCT_CONST_BITS);
+    const __m128i s1_29_7 = _mm_srai_epi32(s1_29_5, DCT_CONST_BITS);
+    // Combine
+    step1[18] = _mm_packs_epi32(s1_18_6, s1_18_7);
+    step1[19] = _mm_packs_epi32(s1_19_6, s1_19_7);
+    step1[20] = _mm_packs_epi32(s1_20_6, s1_20_7);
+    step1[21] = _mm_packs_epi32(s1_21_6, s1_21_7);
+    step1[26] = _mm_packs_epi32(s1_26_6, s1_26_7);
+    step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7);
+    step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7);
+    step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7);
+  }
+  // Stage 5
+  {
+    step2[4] = _mm_add_epi16(step1[5], step3[4]);
+    step2[5] = _mm_sub_epi16(step3[4], step1[5]);
+    step2[6] = _mm_sub_epi16(step3[7], step1[6]);
+    step2[7] = _mm_add_epi16(step1[6], step3[7]);
+  }
+  {
+    const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]);
+    const __m128i out_00_1 = _mm_unpackhi_epi16(step1[0], step1[1]);
+    const __m128i out_08_0 = _mm_unpacklo_epi16(step1[2], step1[3]);
+    const __m128i out_08_1 = _mm_unpackhi_epi16(step1[2], step1[3]);
+    const __m128i out_00_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_p16);
+    const __m128i out_00_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_p16);
+    const __m128i out_16_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_m16);
+    const __m128i out_16_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_m16);
+    const __m128i out_08_2 = _mm_madd_epi16(out_08_0, k__cospi_p24_p08);
+    const __m128i out_08_3 = _mm_madd_epi16(out_08_1, k__cospi_p24_p08);
+    const __m128i out_24_2 = _mm_madd_epi16(out_08_0, k__cospi_m08_p24);
+    const __m128i out_24_3 = _mm_madd_epi16(out_08_1, k__cospi_m08_p24);
+    // dct_const_round_shift
+    const __m128i out_00_4 = _mm_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_00_5 = _mm_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_16_4 = _mm_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_16_5 = _mm_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_08_4 = _mm_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_08_5 = _mm_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_24_4 = _mm_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_24_5 = _mm_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_00_6 = _mm_srai_epi32(out_00_4, DCT_CONST_BITS);
+    const __m128i out_00_7 = _mm_srai_epi32(out_00_5, DCT_CONST_BITS);
+    const __m128i out_16_6 = _mm_srai_epi32(out_16_4, DCT_CONST_BITS);
+    const __m128i out_16_7 = _mm_srai_epi32(out_16_5, DCT_CONST_BITS);
+    const __m128i out_08_6 = _mm_srai_epi32(out_08_4, DCT_CONST_BITS);
+    const __m128i out_08_7 = _mm_srai_epi32(out_08_5, DCT_CONST_BITS);
+    const __m128i out_24_6 = _mm_srai_epi32(out_24_4, DCT_CONST_BITS);
+    const __m128i out_24_7 = _mm_srai_epi32(out_24_5, DCT_CONST_BITS);
+    // Combine
+    out[0] = _mm_packs_epi32(out_00_6, out_00_7);
+    out[16] = _mm_packs_epi32(out_16_6, out_16_7);
+    out[8] = _mm_packs_epi32(out_08_6, out_08_7);
+    out[24] = _mm_packs_epi32(out_24_6, out_24_7);
+  }
+  {
+    const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[9], step1[14]);
+    const __m128i s2_09_1 = _mm_unpackhi_epi16(step1[9], step1[14]);
+    const __m128i s2_10_0 = _mm_unpacklo_epi16(step1[10], step1[13]);
+    const __m128i s2_10_1 = _mm_unpackhi_epi16(step1[10], step1[13]);
+    const __m128i s2_09_2 = _mm_madd_epi16(s2_09_0, k__cospi_m08_p24);
+    const __m128i s2_09_3 = _mm_madd_epi16(s2_09_1, k__cospi_m08_p24);
+    const __m128i s2_10_2 = _mm_madd_epi16(s2_10_0, k__cospi_m24_m08);
+    const __m128i s2_10_3 = _mm_madd_epi16(s2_10_1, k__cospi_m24_m08);
+    const __m128i s2_13_2 = _mm_madd_epi16(s2_10_0, k__cospi_m08_p24);
+    const __m128i s2_13_3 = _mm_madd_epi16(s2_10_1, k__cospi_m08_p24);
+    const __m128i s2_14_2 = _mm_madd_epi16(s2_09_0, k__cospi_p24_p08);
+    const __m128i s2_14_3 = _mm_madd_epi16(s2_09_1, k__cospi_p24_p08);
+    // dct_const_round_shift
+    const __m128i s2_09_4 = _mm_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_09_5 = _mm_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_10_4 = _mm_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_10_5 = _mm_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_13_4 = _mm_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_13_5 = _mm_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_14_4 = _mm_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
+    const __m128i s2_14_5 = _mm_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
+    const __m128i s2_09_6 = _mm_srai_epi32(s2_09_4, DCT_CONST_BITS);
+    const __m128i s2_09_7 = _mm_srai_epi32(s2_09_5, DCT_CONST_BITS);
+    const __m128i s2_10_6 = _mm_srai_epi32(s2_10_4, DCT_CONST_BITS);
+    const __m128i s2_10_7 = _mm_srai_epi32(s2_10_5, DCT_CONST_BITS);
+    const __m128i s2_13_6 = _mm_srai_epi32(s2_13_4, DCT_CONST_BITS);
+    const __m128i s2_13_7 = _mm_srai_epi32(s2_13_5, DCT_CONST_BITS);
+    const __m128i s2_14_6 = _mm_srai_epi32(s2_14_4, DCT_CONST_BITS);
+    const __m128i s2_14_7 = _mm_srai_epi32(s2_14_5, DCT_CONST_BITS);
+    // Combine
+    step2[9] = _mm_packs_epi32(s2_09_6, s2_09_7);
+    step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7);
+    step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7);
+    step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7);
+  }
+  {
+    step2[16] = _mm_add_epi16(step1[19], step3[16]);
+    step2[17] = _mm_add_epi16(step1[18], step3[17]);
+    step2[18] = _mm_sub_epi16(step3[17], step1[18]);
+    step2[19] = _mm_sub_epi16(step3[16], step1[19]);
+    step2[20] = _mm_sub_epi16(step3[23], step1[20]);
+    step2[21] = _mm_sub_epi16(step3[22], step1[21]);
+    step2[22] = _mm_add_epi16(step1[21], step3[22]);
+    step2[23] = _mm_add_epi16(step1[20], step3[23]);
+    step2[24] = _mm_add_epi16(step1[27], step3[24]);
+    step2[25] = _mm_add_epi16(step1[26], step3[25]);
+    step2[26] = _mm_sub_epi16(step3[25], step1[26]);
+    step2[27] = _mm_sub_epi16(step3[24], step1[27]);
+    step2[28] = _mm_sub_epi16(step3[31], step1[28]);
+    step2[29] = _mm_sub_epi16(step3[30], step1[29]);
+    step2[30] = _mm_add_epi16(step1[29], step3[30]);
+    step2[31] = _mm_add_epi16(step1[28], step3[31]);
+  }
+  // Stage 6
+  {
+    const __m128i out_04_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
+    const __m128i out_04_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
+    const __m128i out_20_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
+    const __m128i out_20_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
+    const __m128i out_12_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
+    const __m128i out_12_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
+    const __m128i out_28_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
+    const __m128i out_28_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
+    const __m128i out_04_2 = _mm_madd_epi16(out_04_0, k__cospi_p28_p04);
+    const __m128i out_04_3 = _mm_madd_epi16(out_04_1, k__cospi_p28_p04);
+    const __m128i out_20_2 = _mm_madd_epi16(out_20_0, k__cospi_p12_p20);
+    const __m128i out_20_3 = _mm_madd_epi16(out_20_1, k__cospi_p12_p20);
+    const __m128i out_12_2 = _mm_madd_epi16(out_12_0, k__cospi_m20_p12);
+    const __m128i out_12_3 = _mm_madd_epi16(out_12_1, k__cospi_m20_p12);
+    const __m128i out_28_2 = _mm_madd_epi16(out_28_0, k__cospi_m04_p28);
+    const __m128i out_28_3 = _mm_madd_epi16(out_28_1, k__cospi_m04_p28);
+    // dct_const_round_shift
+    const __m128i out_04_4 = _mm_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_04_5 = _mm_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_20_4 = _mm_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_20_5 = _mm_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_12_4 = _mm_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_12_5 = _mm_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_28_4 = _mm_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_28_5 = _mm_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_04_6 = _mm_srai_epi32(out_04_4, DCT_CONST_BITS);
+    const __m128i out_04_7 = _mm_srai_epi32(out_04_5, DCT_CONST_BITS);
+    const __m128i out_20_6 = _mm_srai_epi32(out_20_4, DCT_CONST_BITS);
+    const __m128i out_20_7 = _mm_srai_epi32(out_20_5, DCT_CONST_BITS);
+    const __m128i out_12_6 = _mm_srai_epi32(out_12_4, DCT_CONST_BITS);
+    const __m128i out_12_7 = _mm_srai_epi32(out_12_5, DCT_CONST_BITS);
+    const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS);
+    const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS);
+    // Combine
+    out[4] = _mm_packs_epi32(out_04_6, out_04_7);
+    out[20] = _mm_packs_epi32(out_20_6, out_20_7);
+    out[12] = _mm_packs_epi32(out_12_6, out_12_7);
+    out[28] = _mm_packs_epi32(out_28_6, out_28_7);
+  }
+  {
+    step3[8] = _mm_add_epi16(step2[9], step1[8]);
+    step3[9] = _mm_sub_epi16(step1[8], step2[9]);
+    step3[10] = _mm_sub_epi16(step1[11], step2[10]);
+    step3[11] = _mm_add_epi16(step2[10], step1[11]);
+    step3[12] = _mm_add_epi16(step2[13], step1[12]);
+    step3[13] = _mm_sub_epi16(step1[12], step2[13]);
+    step3[14] = _mm_sub_epi16(step1[15], step2[14]);
+    step3[15] = _mm_add_epi16(step2[14], step1[15]);
+  }
+  {
+    const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]);
+    const __m128i s3_17_1 = _mm_unpackhi_epi16(step2[17], step2[30]);
+    const __m128i s3_18_0 = _mm_unpacklo_epi16(step2[18], step2[29]);
+    const __m128i s3_18_1 = _mm_unpackhi_epi16(step2[18], step2[29]);
+    const __m128i s3_21_0 = _mm_unpacklo_epi16(step2[21], step2[26]);
+    const __m128i s3_21_1 = _mm_unpackhi_epi16(step2[21], step2[26]);
+    const __m128i s3_22_0 = _mm_unpacklo_epi16(step2[22], step2[25]);
+    const __m128i s3_22_1 = _mm_unpackhi_epi16(step2[22], step2[25]);
+    const __m128i s3_17_2 = _mm_madd_epi16(s3_17_0, k__cospi_m04_p28);
+    const __m128i s3_17_3 = _mm_madd_epi16(s3_17_1, k__cospi_m04_p28);
+    const __m128i s3_18_2 = _mm_madd_epi16(s3_18_0, k__cospi_m28_m04);
+    const __m128i s3_18_3 = _mm_madd_epi16(s3_18_1, k__cospi_m28_m04);
+    const __m128i s3_21_2 = _mm_madd_epi16(s3_21_0, k__cospi_m20_p12);
+    const __m128i s3_21_3 = _mm_madd_epi16(s3_21_1, k__cospi_m20_p12);
+    const __m128i s3_22_2 = _mm_madd_epi16(s3_22_0, k__cospi_m12_m20);
+    const __m128i s3_22_3 = _mm_madd_epi16(s3_22_1, k__cospi_m12_m20);
+    const __m128i s3_25_2 = _mm_madd_epi16(s3_22_0, k__cospi_m20_p12);
+    const __m128i s3_25_3 = _mm_madd_epi16(s3_22_1, k__cospi_m20_p12);
+    const __m128i s3_26_2 = _mm_madd_epi16(s3_21_0, k__cospi_p12_p20);
+    const __m128i s3_26_3 = _mm_madd_epi16(s3_21_1, k__cospi_p12_p20);
+    const __m128i s3_29_2 = _mm_madd_epi16(s3_18_0, k__cospi_m04_p28);
+    const __m128i s3_29_3 = _mm_madd_epi16(s3_18_1, k__cospi_m04_p28);
+    const __m128i s3_30_2 = _mm_madd_epi16(s3_17_0, k__cospi_p28_p04);
+    const __m128i s3_30_3 = _mm_madd_epi16(s3_17_1, k__cospi_p28_p04);
+    // dct_const_round_shift
+    const __m128i s3_17_4 = _mm_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_17_5 = _mm_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_18_4 = _mm_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_18_5 = _mm_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_21_4 = _mm_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_21_5 = _mm_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_22_4 = _mm_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_22_5 = _mm_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_17_6 = _mm_srai_epi32(s3_17_4, DCT_CONST_BITS);
+    const __m128i s3_17_7 = _mm_srai_epi32(s3_17_5, DCT_CONST_BITS);
+    const __m128i s3_18_6 = _mm_srai_epi32(s3_18_4, DCT_CONST_BITS);
+    const __m128i s3_18_7 = _mm_srai_epi32(s3_18_5, DCT_CONST_BITS);
+    const __m128i s3_21_6 = _mm_srai_epi32(s3_21_4, DCT_CONST_BITS);
+    const __m128i s3_21_7 = _mm_srai_epi32(s3_21_5, DCT_CONST_BITS);
+    const __m128i s3_22_6 = _mm_srai_epi32(s3_22_4, DCT_CONST_BITS);
+    const __m128i s3_22_7 = _mm_srai_epi32(s3_22_5, DCT_CONST_BITS);
+    const __m128i s3_25_4 = _mm_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_25_5 = _mm_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_26_4 = _mm_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_26_5 = _mm_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_29_4 = _mm_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_29_5 = _mm_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_30_4 = _mm_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
+    const __m128i s3_30_5 = _mm_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
+    const __m128i s3_25_6 = _mm_srai_epi32(s3_25_4, DCT_CONST_BITS);
+    const __m128i s3_25_7 = _mm_srai_epi32(s3_25_5, DCT_CONST_BITS);
+    const __m128i s3_26_6 = _mm_srai_epi32(s3_26_4, DCT_CONST_BITS);
+    const __m128i s3_26_7 = _mm_srai_epi32(s3_26_5, DCT_CONST_BITS);
+    const __m128i s3_29_6 = _mm_srai_epi32(s3_29_4, DCT_CONST_BITS);
+    const __m128i s3_29_7 = _mm_srai_epi32(s3_29_5, DCT_CONST_BITS);
+    const __m128i s3_30_6 = _mm_srai_epi32(s3_30_4, DCT_CONST_BITS);
+    const __m128i s3_30_7 = _mm_srai_epi32(s3_30_5, DCT_CONST_BITS);
+    // Combine
+    step3[17] = _mm_packs_epi32(s3_17_6, s3_17_7);
+    step3[18] = _mm_packs_epi32(s3_18_6, s3_18_7);
+    step3[21] = _mm_packs_epi32(s3_21_6, s3_21_7);
+    step3[22] = _mm_packs_epi32(s3_22_6, s3_22_7);
+    // Combine
+    step3[25] = _mm_packs_epi32(s3_25_6, s3_25_7);
+    step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7);
+    step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7);
+    step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7);
+  }
+  // Stage 7
+  {
+    const __m128i out_02_0 = _mm_unpacklo_epi16(step3[8], step3[15]);
+    const __m128i out_02_1 = _mm_unpackhi_epi16(step3[8], step3[15]);
+    const __m128i out_18_0 = _mm_unpacklo_epi16(step3[9], step3[14]);
+    const __m128i out_18_1 = _mm_unpackhi_epi16(step3[9], step3[14]);
+    const __m128i out_10_0 = _mm_unpacklo_epi16(step3[10], step3[13]);
+    const __m128i out_10_1 = _mm_unpackhi_epi16(step3[10], step3[13]);
+    const __m128i out_26_0 = _mm_unpacklo_epi16(step3[11], step3[12]);
+    const __m128i out_26_1 = _mm_unpackhi_epi16(step3[11], step3[12]);
+    const __m128i out_02_2 = _mm_madd_epi16(out_02_0, k__cospi_p30_p02);
+    const __m128i out_02_3 = _mm_madd_epi16(out_02_1, k__cospi_p30_p02);
+    const __m128i out_18_2 = _mm_madd_epi16(out_18_0, k__cospi_p14_p18);
+    const __m128i out_18_3 = _mm_madd_epi16(out_18_1, k__cospi_p14_p18);
+    const __m128i out_10_2 = _mm_madd_epi16(out_10_0, k__cospi_p22_p10);
+    const __m128i out_10_3 = _mm_madd_epi16(out_10_1, k__cospi_p22_p10);
+    const __m128i out_26_2 = _mm_madd_epi16(out_26_0, k__cospi_p06_p26);
+    const __m128i out_26_3 = _mm_madd_epi16(out_26_1, k__cospi_p06_p26);
+    const __m128i out_06_2 = _mm_madd_epi16(out_26_0, k__cospi_m26_p06);
+    const __m128i out_06_3 = _mm_madd_epi16(out_26_1, k__cospi_m26_p06);
+    const __m128i out_22_2 = _mm_madd_epi16(out_10_0, k__cospi_m10_p22);
+    const __m128i out_22_3 = _mm_madd_epi16(out_10_1, k__cospi_m10_p22);
+    const __m128i out_14_2 = _mm_madd_epi16(out_18_0, k__cospi_m18_p14);
+    const __m128i out_14_3 = _mm_madd_epi16(out_18_1, k__cospi_m18_p14);
+    const __m128i out_30_2 = _mm_madd_epi16(out_02_0, k__cospi_m02_p30);
+    const __m128i out_30_3 = _mm_madd_epi16(out_02_1, k__cospi_m02_p30);
+    // dct_const_round_shift
+    const __m128i out_02_4 = _mm_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_02_5 = _mm_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_18_4 = _mm_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_18_5 = _mm_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_10_4 = _mm_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_10_5 = _mm_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_26_4 = _mm_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_26_5 = _mm_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_06_4 = _mm_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_06_5 = _mm_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_22_4 = _mm_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_22_5 = _mm_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_14_4 = _mm_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_14_5 = _mm_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_30_4 = _mm_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_30_5 = _mm_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_02_6 = _mm_srai_epi32(out_02_4, DCT_CONST_BITS);
+    const __m128i out_02_7 = _mm_srai_epi32(out_02_5, DCT_CONST_BITS);
+    const __m128i out_18_6 = _mm_srai_epi32(out_18_4, DCT_CONST_BITS);
+    const __m128i out_18_7 = _mm_srai_epi32(out_18_5, DCT_CONST_BITS);
+    const __m128i out_10_6 = _mm_srai_epi32(out_10_4, DCT_CONST_BITS);
+    const __m128i out_10_7 = _mm_srai_epi32(out_10_5, DCT_CONST_BITS);
+    const __m128i out_26_6 = _mm_srai_epi32(out_26_4, DCT_CONST_BITS);
+    const __m128i out_26_7 = _mm_srai_epi32(out_26_5, DCT_CONST_BITS);
+    const __m128i out_06_6 = _mm_srai_epi32(out_06_4, DCT_CONST_BITS);
+    const __m128i out_06_7 = _mm_srai_epi32(out_06_5, DCT_CONST_BITS);
+    const __m128i out_22_6 = _mm_srai_epi32(out_22_4, DCT_CONST_BITS);
+    const __m128i out_22_7 = _mm_srai_epi32(out_22_5, DCT_CONST_BITS);
+    const __m128i out_14_6 = _mm_srai_epi32(out_14_4, DCT_CONST_BITS);
+    const __m128i out_14_7 = _mm_srai_epi32(out_14_5, DCT_CONST_BITS);
+    const __m128i out_30_6 = _mm_srai_epi32(out_30_4, DCT_CONST_BITS);
+    const __m128i out_30_7 = _mm_srai_epi32(out_30_5, DCT_CONST_BITS);
+    // Combine
+    out[2] = _mm_packs_epi32(out_02_6, out_02_7);
+    out[18] = _mm_packs_epi32(out_18_6, out_18_7);
+    out[10] = _mm_packs_epi32(out_10_6, out_10_7);
+    out[26] = _mm_packs_epi32(out_26_6, out_26_7);
+    out[6] = _mm_packs_epi32(out_06_6, out_06_7);
+    out[22] = _mm_packs_epi32(out_22_6, out_22_7);
+    out[14] = _mm_packs_epi32(out_14_6, out_14_7);
+    out[30] = _mm_packs_epi32(out_30_6, out_30_7);
+  }
+  {
+    step1[16] = _mm_add_epi16(step3[17], step2[16]);
+    step1[17] = _mm_sub_epi16(step2[16], step3[17]);
+    step1[18] = _mm_sub_epi16(step2[19], step3[18]);
+    step1[19] = _mm_add_epi16(step3[18], step2[19]);
+    step1[20] = _mm_add_epi16(step3[21], step2[20]);
+    step1[21] = _mm_sub_epi16(step2[20], step3[21]);
+    step1[22] = _mm_sub_epi16(step2[23], step3[22]);
+    step1[23] = _mm_add_epi16(step3[22], step2[23]);
+    step1[24] = _mm_add_epi16(step3[25], step2[24]);
+    step1[25] = _mm_sub_epi16(step2[24], step3[25]);
+    step1[26] = _mm_sub_epi16(step2[27], step3[26]);
+    step1[27] = _mm_add_epi16(step3[26], step2[27]);
+    step1[28] = _mm_add_epi16(step3[29], step2[28]);
+    step1[29] = _mm_sub_epi16(step2[28], step3[29]);
+    step1[30] = _mm_sub_epi16(step2[31], step3[30]);
+    step1[31] = _mm_add_epi16(step3[30], step2[31]);
+  }
+  // Final stage --- outputs indices are bit-reversed.
+  {
+    const __m128i out_01_0 = _mm_unpacklo_epi16(step1[16], step1[31]);
+    const __m128i out_01_1 = _mm_unpackhi_epi16(step1[16], step1[31]);
+    const __m128i out_17_0 = _mm_unpacklo_epi16(step1[17], step1[30]);
+    const __m128i out_17_1 = _mm_unpackhi_epi16(step1[17], step1[30]);
+    const __m128i out_09_0 = _mm_unpacklo_epi16(step1[18], step1[29]);
+    const __m128i out_09_1 = _mm_unpackhi_epi16(step1[18], step1[29]);
+    const __m128i out_25_0 = _mm_unpacklo_epi16(step1[19], step1[28]);
+    const __m128i out_25_1 = _mm_unpackhi_epi16(step1[19], step1[28]);
+    const __m128i out_01_2 = _mm_madd_epi16(out_01_0, k__cospi_p31_p01);
+    const __m128i out_01_3 = _mm_madd_epi16(out_01_1, k__cospi_p31_p01);
+    const __m128i out_17_2 = _mm_madd_epi16(out_17_0, k__cospi_p15_p17);
+    const __m128i out_17_3 = _mm_madd_epi16(out_17_1, k__cospi_p15_p17);
+    const __m128i out_09_2 = _mm_madd_epi16(out_09_0, k__cospi_p23_p09);
+    const __m128i out_09_3 = _mm_madd_epi16(out_09_1, k__cospi_p23_p09);
+    const __m128i out_25_2 = _mm_madd_epi16(out_25_0, k__cospi_p07_p25);
+    const __m128i out_25_3 = _mm_madd_epi16(out_25_1, k__cospi_p07_p25);
+    const __m128i out_07_2 = _mm_madd_epi16(out_25_0, k__cospi_m25_p07);
+    const __m128i out_07_3 = _mm_madd_epi16(out_25_1, k__cospi_m25_p07);
+    const __m128i out_23_2 = _mm_madd_epi16(out_09_0, k__cospi_m09_p23);
+    const __m128i out_23_3 = _mm_madd_epi16(out_09_1, k__cospi_m09_p23);
+    const __m128i out_15_2 = _mm_madd_epi16(out_17_0, k__cospi_m17_p15);
+    const __m128i out_15_3 = _mm_madd_epi16(out_17_1, k__cospi_m17_p15);
+    const __m128i out_31_2 = _mm_madd_epi16(out_01_0, k__cospi_m01_p31);
+    const __m128i out_31_3 = _mm_madd_epi16(out_01_1, k__cospi_m01_p31);
+    // dct_const_round_shift
+    const __m128i out_01_4 = _mm_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_01_5 = _mm_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_17_4 = _mm_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_17_5 = _mm_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_09_4 = _mm_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_09_5 = _mm_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_25_4 = _mm_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_25_5 = _mm_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_07_4 = _mm_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_07_5 = _mm_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_23_4 = _mm_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_23_5 = _mm_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_15_4 = _mm_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_15_5 = _mm_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_31_4 = _mm_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_31_5 = _mm_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_01_6 = _mm_srai_epi32(out_01_4, DCT_CONST_BITS);
+    const __m128i out_01_7 = _mm_srai_epi32(out_01_5, DCT_CONST_BITS);
+    const __m128i out_17_6 = _mm_srai_epi32(out_17_4, DCT_CONST_BITS);
+    const __m128i out_17_7 = _mm_srai_epi32(out_17_5, DCT_CONST_BITS);
+    const __m128i out_09_6 = _mm_srai_epi32(out_09_4, DCT_CONST_BITS);
+    const __m128i out_09_7 = _mm_srai_epi32(out_09_5, DCT_CONST_BITS);
+    const __m128i out_25_6 = _mm_srai_epi32(out_25_4, DCT_CONST_BITS);
+    const __m128i out_25_7 = _mm_srai_epi32(out_25_5, DCT_CONST_BITS);
+    const __m128i out_07_6 = _mm_srai_epi32(out_07_4, DCT_CONST_BITS);
+    const __m128i out_07_7 = _mm_srai_epi32(out_07_5, DCT_CONST_BITS);
+    const __m128i out_23_6 = _mm_srai_epi32(out_23_4, DCT_CONST_BITS);
+    const __m128i out_23_7 = _mm_srai_epi32(out_23_5, DCT_CONST_BITS);
+    const __m128i out_15_6 = _mm_srai_epi32(out_15_4, DCT_CONST_BITS);
+    const __m128i out_15_7 = _mm_srai_epi32(out_15_5, DCT_CONST_BITS);
+    const __m128i out_31_6 = _mm_srai_epi32(out_31_4, DCT_CONST_BITS);
+    const __m128i out_31_7 = _mm_srai_epi32(out_31_5, DCT_CONST_BITS);
+    // Combine
+    out[1] = _mm_packs_epi32(out_01_6, out_01_7);
+    out[17] = _mm_packs_epi32(out_17_6, out_17_7);
+    out[9] = _mm_packs_epi32(out_09_6, out_09_7);
+    out[25] = _mm_packs_epi32(out_25_6, out_25_7);
+    out[7] = _mm_packs_epi32(out_07_6, out_07_7);
+    out[23] = _mm_packs_epi32(out_23_6, out_23_7);
+    out[15] = _mm_packs_epi32(out_15_6, out_15_7);
+    out[31] = _mm_packs_epi32(out_31_6, out_31_7);
+  }
+  {
+    const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]);
+    const __m128i out_05_1 = _mm_unpackhi_epi16(step1[20], step1[27]);
+    const __m128i out_21_0 = _mm_unpacklo_epi16(step1[21], step1[26]);
+    const __m128i out_21_1 = _mm_unpackhi_epi16(step1[21], step1[26]);
+    const __m128i out_13_0 = _mm_unpacklo_epi16(step1[22], step1[25]);
+    const __m128i out_13_1 = _mm_unpackhi_epi16(step1[22], step1[25]);
+    const __m128i out_29_0 = _mm_unpacklo_epi16(step1[23], step1[24]);
+    const __m128i out_29_1 = _mm_unpackhi_epi16(step1[23], step1[24]);
+    const __m128i out_05_2 = _mm_madd_epi16(out_05_0, k__cospi_p27_p05);
+    const __m128i out_05_3 = _mm_madd_epi16(out_05_1, k__cospi_p27_p05);
+    const __m128i out_21_2 = _mm_madd_epi16(out_21_0, k__cospi_p11_p21);
+    const __m128i out_21_3 = _mm_madd_epi16(out_21_1, k__cospi_p11_p21);
+    const __m128i out_13_2 = _mm_madd_epi16(out_13_0, k__cospi_p19_p13);
+    const __m128i out_13_3 = _mm_madd_epi16(out_13_1, k__cospi_p19_p13);
+    const __m128i out_29_2 = _mm_madd_epi16(out_29_0, k__cospi_p03_p29);
+    const __m128i out_29_3 = _mm_madd_epi16(out_29_1, k__cospi_p03_p29);
+    const __m128i out_03_2 = _mm_madd_epi16(out_29_0, k__cospi_m29_p03);
+    const __m128i out_03_3 = _mm_madd_epi16(out_29_1, k__cospi_m29_p03);
+    const __m128i out_19_2 = _mm_madd_epi16(out_13_0, k__cospi_m13_p19);
+    const __m128i out_19_3 = _mm_madd_epi16(out_13_1, k__cospi_m13_p19);
+    const __m128i out_11_2 = _mm_madd_epi16(out_21_0, k__cospi_m21_p11);
+    const __m128i out_11_3 = _mm_madd_epi16(out_21_1, k__cospi_m21_p11);
+    const __m128i out_27_2 = _mm_madd_epi16(out_05_0, k__cospi_m05_p27);
+    const __m128i out_27_3 = _mm_madd_epi16(out_05_1, k__cospi_m05_p27);
+    // dct_const_round_shift
+    const __m128i out_05_4 = _mm_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_05_5 = _mm_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_21_4 = _mm_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_21_5 = _mm_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_13_4 = _mm_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_13_5 = _mm_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_29_4 = _mm_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_29_5 = _mm_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_03_4 = _mm_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_03_5 = _mm_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_19_4 = _mm_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_19_5 = _mm_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_11_4 = _mm_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_11_5 = _mm_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_27_4 = _mm_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
+    const __m128i out_27_5 = _mm_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
+    const __m128i out_05_6 = _mm_srai_epi32(out_05_4, DCT_CONST_BITS);
+    const __m128i out_05_7 = _mm_srai_epi32(out_05_5, DCT_CONST_BITS);
+    const __m128i out_21_6 = _mm_srai_epi32(out_21_4, DCT_CONST_BITS);
+    const __m128i out_21_7 = _mm_srai_epi32(out_21_5, DCT_CONST_BITS);
+    const __m128i out_13_6 = _mm_srai_epi32(out_13_4, DCT_CONST_BITS);
+    const __m128i out_13_7 = _mm_srai_epi32(out_13_5, DCT_CONST_BITS);
+    const __m128i out_29_6 = _mm_srai_epi32(out_29_4, DCT_CONST_BITS);
+    const __m128i out_29_7 = _mm_srai_epi32(out_29_5, DCT_CONST_BITS);
+    const __m128i out_03_6 = _mm_srai_epi32(out_03_4, DCT_CONST_BITS);
+    const __m128i out_03_7 = _mm_srai_epi32(out_03_5, DCT_CONST_BITS);
+    const __m128i out_19_6 = _mm_srai_epi32(out_19_4, DCT_CONST_BITS);
+    const __m128i out_19_7 = _mm_srai_epi32(out_19_5, DCT_CONST_BITS);
+    const __m128i out_11_6 = _mm_srai_epi32(out_11_4, DCT_CONST_BITS);
+    const __m128i out_11_7 = _mm_srai_epi32(out_11_5, DCT_CONST_BITS);
+    const __m128i out_27_6 = _mm_srai_epi32(out_27_4, DCT_CONST_BITS);
+    const __m128i out_27_7 = _mm_srai_epi32(out_27_5, DCT_CONST_BITS);
+    // Combine
+    out[5] = _mm_packs_epi32(out_05_6, out_05_7);
+    out[21] = _mm_packs_epi32(out_21_6, out_21_7);
+    out[13] = _mm_packs_epi32(out_13_6, out_13_7);
+    out[29] = _mm_packs_epi32(out_29_6, out_29_7);
+    out[3] = _mm_packs_epi32(out_03_6, out_03_7);
+    out[19] = _mm_packs_epi32(out_19_6, out_19_7);
+    out[11] = _mm_packs_epi32(out_11_6, out_11_7);
+    out[27] = _mm_packs_epi32(out_27_6, out_27_7);
+  }
+
+  // Output results
+  {
+    int j;
+    for (j = 0; j < 16; ++j) {
+      _mm_storeu_si128((__m128i *)(in0 + j), out[j]);
+      _mm_storeu_si128((__m128i *)(in1 + j), out[j + 16]);
+    }
+  }
+}  // NOLINT
diff --git a/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_avx2.h b/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_avx2.h
new file mode 100644
index 0000000000..216739581b
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_avx2.h
@@ -0,0 +1,3022 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "aom_dsp/txfm_common.h"
+#include "aom_dsp/x86/txfm_common_intrin.h"
+#include "aom_dsp/x86/txfm_common_avx2.h"
+
+#if FDCT32x32_HIGH_PRECISION
+static INLINE __m256i k_madd_epi32_avx2(__m256i a, __m256i b) {
+  __m256i buf0, buf1;
+  buf0 = _mm256_mul_epu32(a, b);
+  a = _mm256_srli_epi64(a, 32);
+  b = _mm256_srli_epi64(b, 32);
+  buf1 = _mm256_mul_epu32(a, b);
+  return _mm256_add_epi64(buf0, buf1);
+}
+
+static INLINE __m256i k_packs_epi64_avx2(__m256i a, __m256i b) {
+  __m256i buf0 = _mm256_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
+  __m256i buf1 = _mm256_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
+  return _mm256_unpacklo_epi64(buf0, buf1);
+}
+#endif
+
+#ifndef STORE_COEFF_FUNC
+#define STORE_COEFF_FUNC
+static void store_coeff(const __m256i *coeff, tran_low_t *curr,
+                        tran_low_t *next) {
+  __m128i u = _mm256_castsi256_si128(*coeff);
+  storeu_output(&u, curr);
+  u = _mm256_extractf128_si256(*coeff, 1);
+  storeu_output(&u, next);
+}
+#endif
+
+void FDCT32x32_2D_AVX2(const int16_t *input, tran_low_t *output_org,
+                       int stride) {
+  // Calculate pre-multiplied strides
+  const int str1 = stride;
+  const int str2 = 2 * stride;
+  const int str3 = 2 * stride + str1;
+  // We need an intermediate buffer between passes.
+  DECLARE_ALIGNED(32, int16_t, intermediate[32 * 32]);
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m256i k__cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64);
+  const __m256i k__cospi_p16_m16 =
+      pair256_set_epi16(+cospi_16_64, -cospi_16_64);
+  const __m256i k__cospi_m08_p24 = pair256_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m256i k__cospi_m24_m08 = pair256_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m256i k__cospi_p24_p08 = pair256_set_epi16(+cospi_24_64, cospi_8_64);
+  const __m256i k__cospi_p12_p20 = pair256_set_epi16(+cospi_12_64, cospi_20_64);
+  const __m256i k__cospi_m20_p12 = pair256_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m256i k__cospi_m04_p28 = pair256_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m256i k__cospi_p28_p04 = pair256_set_epi16(+cospi_28_64, cospi_4_64);
+  const __m256i k__cospi_m28_m04 = pair256_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m256i k__cospi_m12_m20 =
+      pair256_set_epi16(-cospi_12_64, -cospi_20_64);
+  const __m256i k__cospi_p30_p02 = pair256_set_epi16(+cospi_30_64, cospi_2_64);
+  const __m256i k__cospi_p14_p18 = pair256_set_epi16(+cospi_14_64, cospi_18_64);
+  const __m256i k__cospi_p22_p10 = pair256_set_epi16(+cospi_22_64, cospi_10_64);
+  const __m256i k__cospi_p06_p26 = pair256_set_epi16(+cospi_6_64, cospi_26_64);
+  const __m256i k__cospi_m26_p06 = pair256_set_epi16(-cospi_26_64, cospi_6_64);
+  const __m256i k__cospi_m10_p22 = pair256_set_epi16(-cospi_10_64, cospi_22_64);
+  const __m256i k__cospi_m18_p14 = pair256_set_epi16(-cospi_18_64, cospi_14_64);
+  const __m256i k__cospi_m02_p30 = pair256_set_epi16(-cospi_2_64, cospi_30_64);
+  const __m256i k__cospi_p31_p01 = pair256_set_epi16(+cospi_31_64, cospi_1_64);
+  const __m256i k__cospi_p15_p17 = pair256_set_epi16(+cospi_15_64, cospi_17_64);
+  const __m256i k__cospi_p23_p09 = pair256_set_epi16(+cospi_23_64, cospi_9_64);
+  const __m256i k__cospi_p07_p25 = pair256_set_epi16(+cospi_7_64, cospi_25_64);
+  const __m256i k__cospi_m25_p07 = pair256_set_epi16(-cospi_25_64, cospi_7_64);
+  const __m256i k__cospi_m09_p23 = pair256_set_epi16(-cospi_9_64, cospi_23_64);
+  const __m256i k__cospi_m17_p15 = pair256_set_epi16(-cospi_17_64, cospi_15_64);
+  const __m256i k__cospi_m01_p31 = pair256_set_epi16(-cospi_1_64, cospi_31_64);
+  const __m256i k__cospi_p27_p05 = pair256_set_epi16(+cospi_27_64, cospi_5_64);
+  const __m256i k__cospi_p11_p21 = pair256_set_epi16(+cospi_11_64, cospi_21_64);
+  const __m256i k__cospi_p19_p13 = pair256_set_epi16(+cospi_19_64, cospi_13_64);
+  const __m256i k__cospi_p03_p29 = pair256_set_epi16(+cospi_3_64, cospi_29_64);
+  const __m256i k__cospi_m29_p03 = pair256_set_epi16(-cospi_29_64, cospi_3_64);
+  const __m256i k__cospi_m13_p19 = pair256_set_epi16(-cospi_13_64, cospi_19_64);
+  const __m256i k__cospi_m21_p11 = pair256_set_epi16(-cospi_21_64, cospi_11_64);
+  const __m256i k__cospi_m05_p27 = pair256_set_epi16(-cospi_5_64, cospi_27_64);
+  const __m256i k__DCT_CONST_ROUNDING = _mm256_set1_epi32(DCT_CONST_ROUNDING);
+  const __m256i kZero = _mm256_set1_epi16(0);
+  const __m256i kOne = _mm256_set1_epi16(1);
+  // Do the two transform/transpose passes
+  int pass;
+  for (pass = 0; pass < 2; ++pass) {
+    // We process sixteen columns (transposed rows in second pass) at a time.
+    int column_start;
+    for (column_start = 0; column_start < 32; column_start += 16) {
+      __m256i step1[32];
+      __m256i step2[32];
+      __m256i step3[32];
+      __m256i out[32];
+      // Stage 1
+      // Note: even though all the loads below are aligned, using the aligned
+      //       intrinsic make the code slightly slower.
+      if (0 == pass) {
+        const int16_t *in = &input[column_start];
+        // step1[i] =  (in[ 0 * stride] + in[(32 -  1) * stride]) << 2;
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          const int16_t *ina = in + 0 * str1;
+          const int16_t *inb = in + 31 * str1;
+          __m256i *step1a = &step1[0];
+          __m256i *step1b = &step1[31];
+          const __m256i ina0 = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1 =
+              _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2 =
+              _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3 =
+              _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3 =
+              _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2 =
+              _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1 =
+              _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0 = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[0] = _mm256_add_epi16(ina0, inb0);
+          step1a[1] = _mm256_add_epi16(ina1, inb1);
+          step1a[2] = _mm256_add_epi16(ina2, inb2);
+          step1a[3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[0] = _mm256_slli_epi16(step1a[0], 2);
+          step1a[1] = _mm256_slli_epi16(step1a[1], 2);
+          step1a[2] = _mm256_slli_epi16(step1a[2], 2);
+          step1a[3] = _mm256_slli_epi16(step1a[3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina = in + 4 * str1;
+          const int16_t *inb = in + 27 * str1;
+          __m256i *step1a = &step1[4];
+          __m256i *step1b = &step1[27];
+          const __m256i ina0 = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1 =
+              _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2 =
+              _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3 =
+              _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3 =
+              _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2 =
+              _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1 =
+              _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0 = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[0] = _mm256_add_epi16(ina0, inb0);
+          step1a[1] = _mm256_add_epi16(ina1, inb1);
+          step1a[2] = _mm256_add_epi16(ina2, inb2);
+          step1a[3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[0] = _mm256_slli_epi16(step1a[0], 2);
+          step1a[1] = _mm256_slli_epi16(step1a[1], 2);
+          step1a[2] = _mm256_slli_epi16(step1a[2], 2);
+          step1a[3] = _mm256_slli_epi16(step1a[3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina = in + 8 * str1;
+          const int16_t *inb = in + 23 * str1;
+          __m256i *step1a = &step1[8];
+          __m256i *step1b = &step1[23];
+          const __m256i ina0 = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1 =
+              _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2 =
+              _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3 =
+              _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3 =
+              _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2 =
+              _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1 =
+              _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0 = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[0] = _mm256_add_epi16(ina0, inb0);
+          step1a[1] = _mm256_add_epi16(ina1, inb1);
+          step1a[2] = _mm256_add_epi16(ina2, inb2);
+          step1a[3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[0] = _mm256_slli_epi16(step1a[0], 2);
+          step1a[1] = _mm256_slli_epi16(step1a[1], 2);
+          step1a[2] = _mm256_slli_epi16(step1a[2], 2);
+          step1a[3] = _mm256_slli_epi16(step1a[3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina = in + 12 * str1;
+          const int16_t *inb = in + 19 * str1;
+          __m256i *step1a = &step1[12];
+          __m256i *step1b = &step1[19];
+          const __m256i ina0 = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1 =
+              _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2 =
+              _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3 =
+              _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3 =
+              _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2 =
+              _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1 =
+              _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0 = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[0] = _mm256_add_epi16(ina0, inb0);
+          step1a[1] = _mm256_add_epi16(ina1, inb1);
+          step1a[2] = _mm256_add_epi16(ina2, inb2);
+          step1a[3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[0] = _mm256_slli_epi16(step1a[0], 2);
+          step1a[1] = _mm256_slli_epi16(step1a[1], 2);
+          step1a[2] = _mm256_slli_epi16(step1a[2], 2);
+          step1a[3] = _mm256_slli_epi16(step1a[3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+      } else {
+        int16_t *in = &intermediate[column_start];
+        // step1[i] =  in[ 0 * 32] + in[(32 -  1) * 32];
+        // Note: using the same approach as above to have common offset is
+        //       counter-productive as all offsets can be calculated at compile
+        //       time.
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          __m256i in00 = _mm256_loadu_si256((const __m256i *)(in + 0 * 32));
+          __m256i in01 = _mm256_loadu_si256((const __m256i *)(in + 1 * 32));
+          __m256i in02 = _mm256_loadu_si256((const __m256i *)(in + 2 * 32));
+          __m256i in03 = _mm256_loadu_si256((const __m256i *)(in + 3 * 32));
+          __m256i in28 = _mm256_loadu_si256((const __m256i *)(in + 28 * 32));
+          __m256i in29 = _mm256_loadu_si256((const __m256i *)(in + 29 * 32));
+          __m256i in30 = _mm256_loadu_si256((const __m256i *)(in + 30 * 32));
+          __m256i in31 = _mm256_loadu_si256((const __m256i *)(in + 31 * 32));
+          step1[0] = _mm256_add_epi16(in00, in31);
+          step1[1] = _mm256_add_epi16(in01, in30);
+          step1[2] = _mm256_add_epi16(in02, in29);
+          step1[3] = _mm256_add_epi16(in03, in28);
+          step1[28] = _mm256_sub_epi16(in03, in28);
+          step1[29] = _mm256_sub_epi16(in02, in29);
+          step1[30] = _mm256_sub_epi16(in01, in30);
+          step1[31] = _mm256_sub_epi16(in00, in31);
+        }
+        {
+          __m256i in04 = _mm256_loadu_si256((const __m256i *)(in + 4 * 32));
+          __m256i in05 = _mm256_loadu_si256((const __m256i *)(in + 5 * 32));
+          __m256i in06 = _mm256_loadu_si256((const __m256i *)(in + 6 * 32));
+          __m256i in07 = _mm256_loadu_si256((const __m256i *)(in + 7 * 32));
+          __m256i in24 = _mm256_loadu_si256((const __m256i *)(in + 24 * 32));
+          __m256i in25 = _mm256_loadu_si256((const __m256i *)(in + 25 * 32));
+          __m256i in26 = _mm256_loadu_si256((const __m256i *)(in + 26 * 32));
+          __m256i in27 = _mm256_loadu_si256((const __m256i *)(in + 27 * 32));
+          step1[4] = _mm256_add_epi16(in04, in27);
+          step1[5] = _mm256_add_epi16(in05, in26);
+          step1[6] = _mm256_add_epi16(in06, in25);
+          step1[7] = _mm256_add_epi16(in07, in24);
+          step1[24] = _mm256_sub_epi16(in07, in24);
+          step1[25] = _mm256_sub_epi16(in06, in25);
+          step1[26] = _mm256_sub_epi16(in05, in26);
+          step1[27] = _mm256_sub_epi16(in04, in27);
+        }
+        {
+          __m256i in08 = _mm256_loadu_si256((const __m256i *)(in + 8 * 32));
+          __m256i in09 = _mm256_loadu_si256((const __m256i *)(in + 9 * 32));
+          __m256i in10 = _mm256_loadu_si256((const __m256i *)(in + 10 * 32));
+          __m256i in11 = _mm256_loadu_si256((const __m256i *)(in + 11 * 32));
+          __m256i in20 = _mm256_loadu_si256((const __m256i *)(in + 20 * 32));
+          __m256i in21 = _mm256_loadu_si256((const __m256i *)(in + 21 * 32));
+          __m256i in22 = _mm256_loadu_si256((const __m256i *)(in + 22 * 32));
+          __m256i in23 = _mm256_loadu_si256((const __m256i *)(in + 23 * 32));
+          step1[8] = _mm256_add_epi16(in08, in23);
+          step1[9] = _mm256_add_epi16(in09, in22);
+          step1[10] = _mm256_add_epi16(in10, in21);
+          step1[11] = _mm256_add_epi16(in11, in20);
+          step1[20] = _mm256_sub_epi16(in11, in20);
+          step1[21] = _mm256_sub_epi16(in10, in21);
+          step1[22] = _mm256_sub_epi16(in09, in22);
+          step1[23] = _mm256_sub_epi16(in08, in23);
+        }
+        {
+          __m256i in12 = _mm256_loadu_si256((const __m256i *)(in + 12 * 32));
+          __m256i in13 = _mm256_loadu_si256((const __m256i *)(in + 13 * 32));
+          __m256i in14 = _mm256_loadu_si256((const __m256i *)(in + 14 * 32));
+          __m256i in15 = _mm256_loadu_si256((const __m256i *)(in + 15 * 32));
+          __m256i in16 = _mm256_loadu_si256((const __m256i *)(in + 16 * 32));
+          __m256i in17 = _mm256_loadu_si256((const __m256i *)(in + 17 * 32));
+          __m256i in18 = _mm256_loadu_si256((const __m256i *)(in + 18 * 32));
+          __m256i in19 = _mm256_loadu_si256((const __m256i *)(in + 19 * 32));
+          step1[12] = _mm256_add_epi16(in12, in19);
+          step1[13] = _mm256_add_epi16(in13, in18);
+          step1[14] = _mm256_add_epi16(in14, in17);
+          step1[15] = _mm256_add_epi16(in15, in16);
+          step1[16] = _mm256_sub_epi16(in15, in16);
+          step1[17] = _mm256_sub_epi16(in14, in17);
+          step1[18] = _mm256_sub_epi16(in13, in18);
+          step1[19] = _mm256_sub_epi16(in12, in19);
+        }
+      }
+      // Stage 2
+      {
+        step2[0] = _mm256_add_epi16(step1[0], step1[15]);
+        step2[1] = _mm256_add_epi16(step1[1], step1[14]);
+        step2[2] = _mm256_add_epi16(step1[2], step1[13]);
+        step2[3] = _mm256_add_epi16(step1[3], step1[12]);
+        step2[4] = _mm256_add_epi16(step1[4], step1[11]);
+        step2[5] = _mm256_add_epi16(step1[5], step1[10]);
+        step2[6] = _mm256_add_epi16(step1[6], step1[9]);
+        step2[7] = _mm256_add_epi16(step1[7], step1[8]);
+        step2[8] = _mm256_sub_epi16(step1[7], step1[8]);
+        step2[9] = _mm256_sub_epi16(step1[6], step1[9]);
+        step2[10] = _mm256_sub_epi16(step1[5], step1[10]);
+        step2[11] = _mm256_sub_epi16(step1[4], step1[11]);
+        step2[12] = _mm256_sub_epi16(step1[3], step1[12]);
+        step2[13] = _mm256_sub_epi16(step1[2], step1[13]);
+        step2[14] = _mm256_sub_epi16(step1[1], step1[14]);
+        step2[15] = _mm256_sub_epi16(step1[0], step1[15]);
+      }
+      {
+        const __m256i s2_20_0 = _mm256_unpacklo_epi16(step1[27], step1[20]);
+        const __m256i s2_20_1 = _mm256_unpackhi_epi16(step1[27], step1[20]);
+        const __m256i s2_21_0 = _mm256_unpacklo_epi16(step1[26], step1[21]);
+        const __m256i s2_21_1 = _mm256_unpackhi_epi16(step1[26], step1[21]);
+        const __m256i s2_22_0 = _mm256_unpacklo_epi16(step1[25], step1[22]);
+        const __m256i s2_22_1 = _mm256_unpackhi_epi16(step1[25], step1[22]);
+        const __m256i s2_23_0 = _mm256_unpacklo_epi16(step1[24], step1[23]);
+        const __m256i s2_23_1 = _mm256_unpackhi_epi16(step1[24], step1[23]);
+        const __m256i s2_20_2 = _mm256_madd_epi16(s2_20_0, k__cospi_p16_m16);
+        const __m256i s2_20_3 = _mm256_madd_epi16(s2_20_1, k__cospi_p16_m16);
+        const __m256i s2_21_2 = _mm256_madd_epi16(s2_21_0, k__cospi_p16_m16);
+        const __m256i s2_21_3 = _mm256_madd_epi16(s2_21_1, k__cospi_p16_m16);
+        const __m256i s2_22_2 = _mm256_madd_epi16(s2_22_0, k__cospi_p16_m16);
+        const __m256i s2_22_3 = _mm256_madd_epi16(s2_22_1, k__cospi_p16_m16);
+        const __m256i s2_23_2 = _mm256_madd_epi16(s2_23_0, k__cospi_p16_m16);
+        const __m256i s2_23_3 = _mm256_madd_epi16(s2_23_1, k__cospi_p16_m16);
+        const __m256i s2_24_2 = _mm256_madd_epi16(s2_23_0, k__cospi_p16_p16);
+        const __m256i s2_24_3 = _mm256_madd_epi16(s2_23_1, k__cospi_p16_p16);
+        const __m256i s2_25_2 = _mm256_madd_epi16(s2_22_0, k__cospi_p16_p16);
+        const __m256i s2_25_3 = _mm256_madd_epi16(s2_22_1, k__cospi_p16_p16);
+        const __m256i s2_26_2 = _mm256_madd_epi16(s2_21_0, k__cospi_p16_p16);
+        const __m256i s2_26_3 = _mm256_madd_epi16(s2_21_1, k__cospi_p16_p16);
+        const __m256i s2_27_2 = _mm256_madd_epi16(s2_20_0, k__cospi_p16_p16);
+        const __m256i s2_27_3 = _mm256_madd_epi16(s2_20_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m256i s2_20_4 =
+            _mm256_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_20_5 =
+            _mm256_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_21_4 =
+            _mm256_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_21_5 =
+            _mm256_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_22_4 =
+            _mm256_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_22_5 =
+            _mm256_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_23_4 =
+            _mm256_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_23_5 =
+            _mm256_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_24_4 =
+            _mm256_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_24_5 =
+            _mm256_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_25_4 =
+            _mm256_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_25_5 =
+            _mm256_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_26_4 =
+            _mm256_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_26_5 =
+            _mm256_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_27_4 =
+            _mm256_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_27_5 =
+            _mm256_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_20_6 = _mm256_srai_epi32(s2_20_4, DCT_CONST_BITS);
+        const __m256i s2_20_7 = _mm256_srai_epi32(s2_20_5, DCT_CONST_BITS);
+        const __m256i s2_21_6 = _mm256_srai_epi32(s2_21_4, DCT_CONST_BITS);
+        const __m256i s2_21_7 = _mm256_srai_epi32(s2_21_5, DCT_CONST_BITS);
+        const __m256i s2_22_6 = _mm256_srai_epi32(s2_22_4, DCT_CONST_BITS);
+        const __m256i s2_22_7 = _mm256_srai_epi32(s2_22_5, DCT_CONST_BITS);
+        const __m256i s2_23_6 = _mm256_srai_epi32(s2_23_4, DCT_CONST_BITS);
+        const __m256i s2_23_7 = _mm256_srai_epi32(s2_23_5, DCT_CONST_BITS);
+        const __m256i s2_24_6 = _mm256_srai_epi32(s2_24_4, DCT_CONST_BITS);
+        const __m256i s2_24_7 = _mm256_srai_epi32(s2_24_5, DCT_CONST_BITS);
+        const __m256i s2_25_6 = _mm256_srai_epi32(s2_25_4, DCT_CONST_BITS);
+        const __m256i s2_25_7 = _mm256_srai_epi32(s2_25_5, DCT_CONST_BITS);
+        const __m256i s2_26_6 = _mm256_srai_epi32(s2_26_4, DCT_CONST_BITS);
+        const __m256i s2_26_7 = _mm256_srai_epi32(s2_26_5, DCT_CONST_BITS);
+        const __m256i s2_27_6 = _mm256_srai_epi32(s2_27_4, DCT_CONST_BITS);
+        const __m256i s2_27_7 = _mm256_srai_epi32(s2_27_5, DCT_CONST_BITS);
+        // Combine
+        step2[20] = _mm256_packs_epi32(s2_20_6, s2_20_7);
+        step2[21] = _mm256_packs_epi32(s2_21_6, s2_21_7);
+        step2[22] = _mm256_packs_epi32(s2_22_6, s2_22_7);
+        step2[23] = _mm256_packs_epi32(s2_23_6, s2_23_7);
+        step2[24] = _mm256_packs_epi32(s2_24_6, s2_24_7);
+        step2[25] = _mm256_packs_epi32(s2_25_6, s2_25_7);
+        step2[26] = _mm256_packs_epi32(s2_26_6, s2_26_7);
+        step2[27] = _mm256_packs_epi32(s2_27_6, s2_27_7);
+      }
+
+#if !FDCT32x32_HIGH_PRECISION
+      // dump the magnitude by half, hence the intermediate values are within
+      // the range of 16 bits.
+      if (1 == pass) {
+        __m256i s3_00_0 = _mm256_cmpgt_epi16(kZero, step2[0]);
+        __m256i s3_01_0 = _mm256_cmpgt_epi16(kZero, step2[1]);
+        __m256i s3_02_0 = _mm256_cmpgt_epi16(kZero, step2[2]);
+        __m256i s3_03_0 = _mm256_cmpgt_epi16(kZero, step2[3]);
+        __m256i s3_04_0 = _mm256_cmpgt_epi16(kZero, step2[4]);
+        __m256i s3_05_0 = _mm256_cmpgt_epi16(kZero, step2[5]);
+        __m256i s3_06_0 = _mm256_cmpgt_epi16(kZero, step2[6]);
+        __m256i s3_07_0 = _mm256_cmpgt_epi16(kZero, step2[7]);
+        __m256i s2_08_0 = _mm256_cmpgt_epi16(kZero, step2[8]);
+        __m256i s2_09_0 = _mm256_cmpgt_epi16(kZero, step2[9]);
+        __m256i s3_10_0 = _mm256_cmpgt_epi16(kZero, step2[10]);
+        __m256i s3_11_0 = _mm256_cmpgt_epi16(kZero, step2[11]);
+        __m256i s3_12_0 = _mm256_cmpgt_epi16(kZero, step2[12]);
+        __m256i s3_13_0 = _mm256_cmpgt_epi16(kZero, step2[13]);
+        __m256i s2_14_0 = _mm256_cmpgt_epi16(kZero, step2[14]);
+        __m256i s2_15_0 = _mm256_cmpgt_epi16(kZero, step2[15]);
+        __m256i s3_16_0 = _mm256_cmpgt_epi16(kZero, step1[16]);
+        __m256i s3_17_0 = _mm256_cmpgt_epi16(kZero, step1[17]);
+        __m256i s3_18_0 = _mm256_cmpgt_epi16(kZero, step1[18]);
+        __m256i s3_19_0 = _mm256_cmpgt_epi16(kZero, step1[19]);
+        __m256i s3_20_0 = _mm256_cmpgt_epi16(kZero, step2[20]);
+        __m256i s3_21_0 = _mm256_cmpgt_epi16(kZero, step2[21]);
+        __m256i s3_22_0 = _mm256_cmpgt_epi16(kZero, step2[22]);
+        __m256i s3_23_0 = _mm256_cmpgt_epi16(kZero, step2[23]);
+        __m256i s3_24_0 = _mm256_cmpgt_epi16(kZero, step2[24]);
+        __m256i s3_25_0 = _mm256_cmpgt_epi16(kZero, step2[25]);
+        __m256i s3_26_0 = _mm256_cmpgt_epi16(kZero, step2[26]);
+        __m256i s3_27_0 = _mm256_cmpgt_epi16(kZero, step2[27]);
+        __m256i s3_28_0 = _mm256_cmpgt_epi16(kZero, step1[28]);
+        __m256i s3_29_0 = _mm256_cmpgt_epi16(kZero, step1[29]);
+        __m256i s3_30_0 = _mm256_cmpgt_epi16(kZero, step1[30]);
+        __m256i s3_31_0 = _mm256_cmpgt_epi16(kZero, step1[31]);
+
+        step2[0] = _mm256_sub_epi16(step2[0], s3_00_0);
+        step2[1] = _mm256_sub_epi16(step2[1], s3_01_0);
+        step2[2] = _mm256_sub_epi16(step2[2], s3_02_0);
+        step2[3] = _mm256_sub_epi16(step2[3], s3_03_0);
+        step2[4] = _mm256_sub_epi16(step2[4], s3_04_0);
+        step2[5] = _mm256_sub_epi16(step2[5], s3_05_0);
+        step2[6] = _mm256_sub_epi16(step2[6], s3_06_0);
+        step2[7] = _mm256_sub_epi16(step2[7], s3_07_0);
+        step2[8] = _mm256_sub_epi16(step2[8], s2_08_0);
+        step2[9] = _mm256_sub_epi16(step2[9], s2_09_0);
+        step2[10] = _mm256_sub_epi16(step2[10], s3_10_0);
+        step2[11] = _mm256_sub_epi16(step2[11], s3_11_0);
+        step2[12] = _mm256_sub_epi16(step2[12], s3_12_0);
+        step2[13] = _mm256_sub_epi16(step2[13], s3_13_0);
+        step2[14] = _mm256_sub_epi16(step2[14], s2_14_0);
+        step2[15] = _mm256_sub_epi16(step2[15], s2_15_0);
+        step1[16] = _mm256_sub_epi16(step1[16], s3_16_0);
+        step1[17] = _mm256_sub_epi16(step1[17], s3_17_0);
+        step1[18] = _mm256_sub_epi16(step1[18], s3_18_0);
+        step1[19] = _mm256_sub_epi16(step1[19], s3_19_0);
+        step2[20] = _mm256_sub_epi16(step2[20], s3_20_0);
+        step2[21] = _mm256_sub_epi16(step2[21], s3_21_0);
+        step2[22] = _mm256_sub_epi16(step2[22], s3_22_0);
+        step2[23] = _mm256_sub_epi16(step2[23], s3_23_0);
+        step2[24] = _mm256_sub_epi16(step2[24], s3_24_0);
+        step2[25] = _mm256_sub_epi16(step2[25], s3_25_0);
+        step2[26] = _mm256_sub_epi16(step2[26], s3_26_0);
+        step2[27] = _mm256_sub_epi16(step2[27], s3_27_0);
+        step1[28] = _mm256_sub_epi16(step1[28], s3_28_0);
+        step1[29] = _mm256_sub_epi16(step1[29], s3_29_0);
+        step1[30] = _mm256_sub_epi16(step1[30], s3_30_0);
+        step1[31] = _mm256_sub_epi16(step1[31], s3_31_0);
+
+        step2[0] = _mm256_add_epi16(step2[0], kOne);
+        step2[1] = _mm256_add_epi16(step2[1], kOne);
+        step2[2] = _mm256_add_epi16(step2[2], kOne);
+        step2[3] = _mm256_add_epi16(step2[3], kOne);
+        step2[4] = _mm256_add_epi16(step2[4], kOne);
+        step2[5] = _mm256_add_epi16(step2[5], kOne);
+        step2[6] = _mm256_add_epi16(step2[6], kOne);
+        step2[7] = _mm256_add_epi16(step2[7], kOne);
+        step2[8] = _mm256_add_epi16(step2[8], kOne);
+        step2[9] = _mm256_add_epi16(step2[9], kOne);
+        step2[10] = _mm256_add_epi16(step2[10], kOne);
+        step2[11] = _mm256_add_epi16(step2[11], kOne);
+        step2[12] = _mm256_add_epi16(step2[12], kOne);
+        step2[13] = _mm256_add_epi16(step2[13], kOne);
+        step2[14] = _mm256_add_epi16(step2[14], kOne);
+        step2[15] = _mm256_add_epi16(step2[15], kOne);
+        step1[16] = _mm256_add_epi16(step1[16], kOne);
+        step1[17] = _mm256_add_epi16(step1[17], kOne);
+        step1[18] = _mm256_add_epi16(step1[18], kOne);
+        step1[19] = _mm256_add_epi16(step1[19], kOne);
+        step2[20] = _mm256_add_epi16(step2[20], kOne);
+        step2[21] = _mm256_add_epi16(step2[21], kOne);
+        step2[22] = _mm256_add_epi16(step2[22], kOne);
+        step2[23] = _mm256_add_epi16(step2[23], kOne);
+        step2[24] = _mm256_add_epi16(step2[24], kOne);
+        step2[25] = _mm256_add_epi16(step2[25], kOne);
+        step2[26] = _mm256_add_epi16(step2[26], kOne);
+        step2[27] = _mm256_add_epi16(step2[27], kOne);
+        step1[28] = _mm256_add_epi16(step1[28], kOne);
+        step1[29] = _mm256_add_epi16(step1[29], kOne);
+        step1[30] = _mm256_add_epi16(step1[30], kOne);
+        step1[31] = _mm256_add_epi16(step1[31], kOne);
+
+        step2[0] = _mm256_srai_epi16(step2[0], 2);
+        step2[1] = _mm256_srai_epi16(step2[1], 2);
+        step2[2] = _mm256_srai_epi16(step2[2], 2);
+        step2[3] = _mm256_srai_epi16(step2[3], 2);
+        step2[4] = _mm256_srai_epi16(step2[4], 2);
+        step2[5] = _mm256_srai_epi16(step2[5], 2);
+        step2[6] = _mm256_srai_epi16(step2[6], 2);
+        step2[7] = _mm256_srai_epi16(step2[7], 2);
+        step2[8] = _mm256_srai_epi16(step2[8], 2);
+        step2[9] = _mm256_srai_epi16(step2[9], 2);
+        step2[10] = _mm256_srai_epi16(step2[10], 2);
+        step2[11] = _mm256_srai_epi16(step2[11], 2);
+        step2[12] = _mm256_srai_epi16(step2[12], 2);
+        step2[13] = _mm256_srai_epi16(step2[13], 2);
+        step2[14] = _mm256_srai_epi16(step2[14], 2);
+        step2[15] = _mm256_srai_epi16(step2[15], 2);
+        step1[16] = _mm256_srai_epi16(step1[16], 2);
+        step1[17] = _mm256_srai_epi16(step1[17], 2);
+        step1[18] = _mm256_srai_epi16(step1[18], 2);
+        step1[19] = _mm256_srai_epi16(step1[19], 2);
+        step2[20] = _mm256_srai_epi16(step2[20], 2);
+        step2[21] = _mm256_srai_epi16(step2[21], 2);
+        step2[22] = _mm256_srai_epi16(step2[22], 2);
+        step2[23] = _mm256_srai_epi16(step2[23], 2);
+        step2[24] = _mm256_srai_epi16(step2[24], 2);
+        step2[25] = _mm256_srai_epi16(step2[25], 2);
+        step2[26] = _mm256_srai_epi16(step2[26], 2);
+        step2[27] = _mm256_srai_epi16(step2[27], 2);
+        step1[28] = _mm256_srai_epi16(step1[28], 2);
+        step1[29] = _mm256_srai_epi16(step1[29], 2);
+        step1[30] = _mm256_srai_epi16(step1[30], 2);
+        step1[31] = _mm256_srai_epi16(step1[31], 2);
+      }
+#endif
+
+#if FDCT32x32_HIGH_PRECISION
+      if (pass == 0) {
+#endif
+        // Stage 3
+        {
+          step3[0] = _mm256_add_epi16(step2[(8 - 1)], step2[0]);
+          step3[1] = _mm256_add_epi16(step2[(8 - 2)], step2[1]);
+          step3[2] = _mm256_add_epi16(step2[(8 - 3)], step2[2]);
+          step3[3] = _mm256_add_epi16(step2[(8 - 4)], step2[3]);
+          step3[4] = _mm256_sub_epi16(step2[(8 - 5)], step2[4]);
+          step3[5] = _mm256_sub_epi16(step2[(8 - 6)], step2[5]);
+          step3[6] = _mm256_sub_epi16(step2[(8 - 7)], step2[6]);
+          step3[7] = _mm256_sub_epi16(step2[(8 - 8)], step2[7]);
+        }
+        {
+          const __m256i s3_10_0 = _mm256_unpacklo_epi16(step2[13], step2[10]);
+          const __m256i s3_10_1 = _mm256_unpackhi_epi16(step2[13], step2[10]);
+          const __m256i s3_11_0 = _mm256_unpacklo_epi16(step2[12], step2[11]);
+          const __m256i s3_11_1 = _mm256_unpackhi_epi16(step2[12], step2[11]);
+          const __m256i s3_10_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_m16);
+          const __m256i s3_10_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_m16);
+          const __m256i s3_11_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_m16);
+          const __m256i s3_11_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_m16);
+          const __m256i s3_12_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_p16);
+          const __m256i s3_12_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_p16);
+          const __m256i s3_13_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_p16);
+          const __m256i s3_13_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_p16);
+          // dct_const_round_shift
+          const __m256i s3_10_4 =
+              _mm256_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_10_5 =
+              _mm256_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_11_4 =
+              _mm256_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_11_5 =
+              _mm256_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_12_4 =
+              _mm256_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_12_5 =
+              _mm256_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_13_4 =
+              _mm256_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_13_5 =
+              _mm256_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_10_6 = _mm256_srai_epi32(s3_10_4, DCT_CONST_BITS);
+          const __m256i s3_10_7 = _mm256_srai_epi32(s3_10_5, DCT_CONST_BITS);
+          const __m256i s3_11_6 = _mm256_srai_epi32(s3_11_4, DCT_CONST_BITS);
+          const __m256i s3_11_7 = _mm256_srai_epi32(s3_11_5, DCT_CONST_BITS);
+          const __m256i s3_12_6 = _mm256_srai_epi32(s3_12_4, DCT_CONST_BITS);
+          const __m256i s3_12_7 = _mm256_srai_epi32(s3_12_5, DCT_CONST_BITS);
+          const __m256i s3_13_6 = _mm256_srai_epi32(s3_13_4, DCT_CONST_BITS);
+          const __m256i s3_13_7 = _mm256_srai_epi32(s3_13_5, DCT_CONST_BITS);
+          // Combine
+          step3[10] = _mm256_packs_epi32(s3_10_6, s3_10_7);
+          step3[11] = _mm256_packs_epi32(s3_11_6, s3_11_7);
+          step3[12] = _mm256_packs_epi32(s3_12_6, s3_12_7);
+          step3[13] = _mm256_packs_epi32(s3_13_6, s3_13_7);
+        }
+        {
+          step3[16] = _mm256_add_epi16(step2[23], step1[16]);
+          step3[17] = _mm256_add_epi16(step2[22], step1[17]);
+          step3[18] = _mm256_add_epi16(step2[21], step1[18]);
+          step3[19] = _mm256_add_epi16(step2[20], step1[19]);
+          step3[20] = _mm256_sub_epi16(step1[19], step2[20]);
+          step3[21] = _mm256_sub_epi16(step1[18], step2[21]);
+          step3[22] = _mm256_sub_epi16(step1[17], step2[22]);
+          step3[23] = _mm256_sub_epi16(step1[16], step2[23]);
+          step3[24] = _mm256_sub_epi16(step1[31], step2[24]);
+          step3[25] = _mm256_sub_epi16(step1[30], step2[25]);
+          step3[26] = _mm256_sub_epi16(step1[29], step2[26]);
+          step3[27] = _mm256_sub_epi16(step1[28], step2[27]);
+          step3[28] = _mm256_add_epi16(step2[27], step1[28]);
+          step3[29] = _mm256_add_epi16(step2[26], step1[29]);
+          step3[30] = _mm256_add_epi16(step2[25], step1[30]);
+          step3[31] = _mm256_add_epi16(step2[24], step1[31]);
+        }
+
+        // Stage 4
+        {
+          step1[0] = _mm256_add_epi16(step3[3], step3[0]);
+          step1[1] = _mm256_add_epi16(step3[2], step3[1]);
+          step1[2] = _mm256_sub_epi16(step3[1], step3[2]);
+          step1[3] = _mm256_sub_epi16(step3[0], step3[3]);
+          step1[8] = _mm256_add_epi16(step3[11], step2[8]);
+          step1[9] = _mm256_add_epi16(step3[10], step2[9]);
+          step1[10] = _mm256_sub_epi16(step2[9], step3[10]);
+          step1[11] = _mm256_sub_epi16(step2[8], step3[11]);
+          step1[12] = _mm256_sub_epi16(step2[15], step3[12]);
+          step1[13] = _mm256_sub_epi16(step2[14], step3[13]);
+          step1[14] = _mm256_add_epi16(step3[13], step2[14]);
+          step1[15] = _mm256_add_epi16(step3[12], step2[15]);
+        }
+        {
+          const __m256i s1_05_0 = _mm256_unpacklo_epi16(step3[6], step3[5]);
+          const __m256i s1_05_1 = _mm256_unpackhi_epi16(step3[6], step3[5]);
+          const __m256i s1_05_2 = _mm256_madd_epi16(s1_05_0, k__cospi_p16_m16);
+          const __m256i s1_05_3 = _mm256_madd_epi16(s1_05_1, k__cospi_p16_m16);
+          const __m256i s1_06_2 = _mm256_madd_epi16(s1_05_0, k__cospi_p16_p16);
+          const __m256i s1_06_3 = _mm256_madd_epi16(s1_05_1, k__cospi_p16_p16);
+          // dct_const_round_shift
+          const __m256i s1_05_4 =
+              _mm256_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
+          const __m256i s1_05_5 =
+              _mm256_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
+          const __m256i s1_06_4 =
+              _mm256_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
+          const __m256i s1_06_5 =
+              _mm256_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
+          const __m256i s1_05_6 = _mm256_srai_epi32(s1_05_4, DCT_CONST_BITS);
+          const __m256i s1_05_7 = _mm256_srai_epi32(s1_05_5, DCT_CONST_BITS);
+          const __m256i s1_06_6 = _mm256_srai_epi32(s1_06_4, DCT_CONST_BITS);
+          const __m256i s1_06_7 = _mm256_srai_epi32(s1_06_5, DCT_CONST_BITS);
+          // Combine
+          step1[5] = _mm256_packs_epi32(s1_05_6, s1_05_7);
+          step1[6] = _mm256_packs_epi32(s1_06_6, s1_06_7);
+        }
+        {
+          const __m256i s1_18_0 = _mm256_unpacklo_epi16(step3[18], step3[29]);
+          const __m256i s1_18_1 = _mm256_unpackhi_epi16(step3[18], step3[29]);
+          const __m256i s1_19_0 = _mm256_unpacklo_epi16(step3[19], step3[28]);
+          const __m256i s1_19_1 = _mm256_unpackhi_epi16(step3[19], step3[28]);
+          const __m256i s1_20_0 = _mm256_unpacklo_epi16(step3[20], step3[27]);
+          const __m256i s1_20_1 = _mm256_unpackhi_epi16(step3[20], step3[27]);
+          const __m256i s1_21_0 = _mm256_unpacklo_epi16(step3[21], step3[26]);
+          const __m256i s1_21_1 = _mm256_unpackhi_epi16(step3[21], step3[26]);
+          const __m256i s1_18_2 = _mm256_madd_epi16(s1_18_0, k__cospi_m08_p24);
+          const __m256i s1_18_3 = _mm256_madd_epi16(s1_18_1, k__cospi_m08_p24);
+          const __m256i s1_19_2 = _mm256_madd_epi16(s1_19_0, k__cospi_m08_p24);
+          const __m256i s1_19_3 = _mm256_madd_epi16(s1_19_1, k__cospi_m08_p24);
+          const __m256i s1_20_2 = _mm256_madd_epi16(s1_20_0, k__cospi_m24_m08);
+          const __m256i s1_20_3 = _mm256_madd_epi16(s1_20_1, k__cospi_m24_m08);
+          const __m256i s1_21_2 = _mm256_madd_epi16(s1_21_0, k__cospi_m24_m08);
+          const __m256i s1_21_3 = _mm256_madd_epi16(s1_21_1, k__cospi_m24_m08);
+          const __m256i s1_26_2 = _mm256_madd_epi16(s1_21_0, k__cospi_m08_p24);
+          const __m256i s1_26_3 = _mm256_madd_epi16(s1_21_1, k__cospi_m08_p24);
+          const __m256i s1_27_2 = _mm256_madd_epi16(s1_20_0, k__cospi_m08_p24);
+          const __m256i s1_27_3 = _mm256_madd_epi16(s1_20_1, k__cospi_m08_p24);
+          const __m256i s1_28_2 = _mm256_madd_epi16(s1_19_0, k__cospi_p24_p08);
+          const __m256i s1_28_3 = _mm256_madd_epi16(s1_19_1, k__cospi_p24_p08);
+          const __m256i s1_29_2 = _mm256_madd_epi16(s1_18_0, k__cospi_p24_p08);
+          const __m256i s1_29_3 = _mm256_madd_epi16(s1_18_1, k__cospi_p24_p08);
+          // dct_const_round_shift
+          const __m256i s1_18_4 =
+              _mm256_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
+          const __m256i s1_18_5 =
+              _mm256_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
+          const __m256i s1_19_4 =
+              _mm256_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
+          const __m256i s1_19_5 =
+              _mm256_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
+          const __m256i s1_20_4 =
+              _mm256_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
+          const __m256i s1_20_5 =
+              _mm256_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
+          const __m256i s1_21_4 =
+              _mm256_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
+          const __m256i s1_21_5 =
+              _mm256_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
+          const __m256i s1_26_4 =
+              _mm256_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
+          const __m256i s1_26_5 =
+              _mm256_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
+          const __m256i s1_27_4 =
+              _mm256_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
+          const __m256i s1_27_5 =
+              _mm256_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
+          const __m256i s1_28_4 =
+              _mm256_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
+          const __m256i s1_28_5 =
+              _mm256_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
+          const __m256i s1_29_4 =
+              _mm256_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
+          const __m256i s1_29_5 =
+              _mm256_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
+          const __m256i s1_18_6 = _mm256_srai_epi32(s1_18_4, DCT_CONST_BITS);
+          const __m256i s1_18_7 = _mm256_srai_epi32(s1_18_5, DCT_CONST_BITS);
+          const __m256i s1_19_6 = _mm256_srai_epi32(s1_19_4, DCT_CONST_BITS);
+          const __m256i s1_19_7 = _mm256_srai_epi32(s1_19_5, DCT_CONST_BITS);
+          const __m256i s1_20_6 = _mm256_srai_epi32(s1_20_4, DCT_CONST_BITS);
+          const __m256i s1_20_7 = _mm256_srai_epi32(s1_20_5, DCT_CONST_BITS);
+          const __m256i s1_21_6 = _mm256_srai_epi32(s1_21_4, DCT_CONST_BITS);
+          const __m256i s1_21_7 = _mm256_srai_epi32(s1_21_5, DCT_CONST_BITS);
+          const __m256i s1_26_6 = _mm256_srai_epi32(s1_26_4, DCT_CONST_BITS);
+          const __m256i s1_26_7 = _mm256_srai_epi32(s1_26_5, DCT_CONST_BITS);
+          const __m256i s1_27_6 = _mm256_srai_epi32(s1_27_4, DCT_CONST_BITS);
+          const __m256i s1_27_7 = _mm256_srai_epi32(s1_27_5, DCT_CONST_BITS);
+          const __m256i s1_28_6 = _mm256_srai_epi32(s1_28_4, DCT_CONST_BITS);
+          const __m256i s1_28_7 = _mm256_srai_epi32(s1_28_5, DCT_CONST_BITS);
+          const __m256i s1_29_6 = _mm256_srai_epi32(s1_29_4, DCT_CONST_BITS);
+          const __m256i s1_29_7 = _mm256_srai_epi32(s1_29_5, DCT_CONST_BITS);
+          // Combine
+          step1[18] = _mm256_packs_epi32(s1_18_6, s1_18_7);
+          step1[19] = _mm256_packs_epi32(s1_19_6, s1_19_7);
+          step1[20] = _mm256_packs_epi32(s1_20_6, s1_20_7);
+          step1[21] = _mm256_packs_epi32(s1_21_6, s1_21_7);
+          step1[26] = _mm256_packs_epi32(s1_26_6, s1_26_7);
+          step1[27] = _mm256_packs_epi32(s1_27_6, s1_27_7);
+          step1[28] = _mm256_packs_epi32(s1_28_6, s1_28_7);
+          step1[29] = _mm256_packs_epi32(s1_29_6, s1_29_7);
+        }
+        // Stage 5
+        {
+          step2[4] = _mm256_add_epi16(step1[5], step3[4]);
+          step2[5] = _mm256_sub_epi16(step3[4], step1[5]);
+          step2[6] = _mm256_sub_epi16(step3[7], step1[6]);
+          step2[7] = _mm256_add_epi16(step1[6], step3[7]);
+        }
+        {
+          const __m256i out_00_0 = _mm256_unpacklo_epi16(step1[0], step1[1]);
+          const __m256i out_00_1 = _mm256_unpackhi_epi16(step1[0], step1[1]);
+          const __m256i out_08_0 = _mm256_unpacklo_epi16(step1[2], step1[3]);
+          const __m256i out_08_1 = _mm256_unpackhi_epi16(step1[2], step1[3]);
+          const __m256i out_00_2 =
+              _mm256_madd_epi16(out_00_0, k__cospi_p16_p16);
+          const __m256i out_00_3 =
+              _mm256_madd_epi16(out_00_1, k__cospi_p16_p16);
+          const __m256i out_16_2 =
+              _mm256_madd_epi16(out_00_0, k__cospi_p16_m16);
+          const __m256i out_16_3 =
+              _mm256_madd_epi16(out_00_1, k__cospi_p16_m16);
+          const __m256i out_08_2 =
+              _mm256_madd_epi16(out_08_0, k__cospi_p24_p08);
+          const __m256i out_08_3 =
+              _mm256_madd_epi16(out_08_1, k__cospi_p24_p08);
+          const __m256i out_24_2 =
+              _mm256_madd_epi16(out_08_0, k__cospi_m08_p24);
+          const __m256i out_24_3 =
+              _mm256_madd_epi16(out_08_1, k__cospi_m08_p24);
+          // dct_const_round_shift
+          const __m256i out_00_4 =
+              _mm256_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_00_5 =
+              _mm256_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_16_4 =
+              _mm256_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_16_5 =
+              _mm256_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_08_4 =
+              _mm256_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_08_5 =
+              _mm256_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_24_4 =
+              _mm256_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_24_5 =
+              _mm256_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_00_6 = _mm256_srai_epi32(out_00_4, DCT_CONST_BITS);
+          const __m256i out_00_7 = _mm256_srai_epi32(out_00_5, DCT_CONST_BITS);
+          const __m256i out_16_6 = _mm256_srai_epi32(out_16_4, DCT_CONST_BITS);
+          const __m256i out_16_7 = _mm256_srai_epi32(out_16_5, DCT_CONST_BITS);
+          const __m256i out_08_6 = _mm256_srai_epi32(out_08_4, DCT_CONST_BITS);
+          const __m256i out_08_7 = _mm256_srai_epi32(out_08_5, DCT_CONST_BITS);
+          const __m256i out_24_6 = _mm256_srai_epi32(out_24_4, DCT_CONST_BITS);
+          const __m256i out_24_7 = _mm256_srai_epi32(out_24_5, DCT_CONST_BITS);
+          // Combine
+          out[0] = _mm256_packs_epi32(out_00_6, out_00_7);
+          out[16] = _mm256_packs_epi32(out_16_6, out_16_7);
+          out[8] = _mm256_packs_epi32(out_08_6, out_08_7);
+          out[24] = _mm256_packs_epi32(out_24_6, out_24_7);
+        }
+        {
+          const __m256i s2_09_0 = _mm256_unpacklo_epi16(step1[9], step1[14]);
+          const __m256i s2_09_1 = _mm256_unpackhi_epi16(step1[9], step1[14]);
+          const __m256i s2_10_0 = _mm256_unpacklo_epi16(step1[10], step1[13]);
+          const __m256i s2_10_1 = _mm256_unpackhi_epi16(step1[10], step1[13]);
+          const __m256i s2_09_2 = _mm256_madd_epi16(s2_09_0, k__cospi_m08_p24);
+          const __m256i s2_09_3 = _mm256_madd_epi16(s2_09_1, k__cospi_m08_p24);
+          const __m256i s2_10_2 = _mm256_madd_epi16(s2_10_0, k__cospi_m24_m08);
+          const __m256i s2_10_3 = _mm256_madd_epi16(s2_10_1, k__cospi_m24_m08);
+          const __m256i s2_13_2 = _mm256_madd_epi16(s2_10_0, k__cospi_m08_p24);
+          const __m256i s2_13_3 = _mm256_madd_epi16(s2_10_1, k__cospi_m08_p24);
+          const __m256i s2_14_2 = _mm256_madd_epi16(s2_09_0, k__cospi_p24_p08);
+          const __m256i s2_14_3 = _mm256_madd_epi16(s2_09_1, k__cospi_p24_p08);
+          // dct_const_round_shift
+          const __m256i s2_09_4 =
+              _mm256_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
+          const __m256i s2_09_5 =
+              _mm256_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
+          const __m256i s2_10_4 =
+              _mm256_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
+          const __m256i s2_10_5 =
+              _mm256_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
+          const __m256i s2_13_4 =
+              _mm256_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
+          const __m256i s2_13_5 =
+              _mm256_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
+          const __m256i s2_14_4 =
+              _mm256_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
+          const __m256i s2_14_5 =
+              _mm256_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
+          const __m256i s2_09_6 = _mm256_srai_epi32(s2_09_4, DCT_CONST_BITS);
+          const __m256i s2_09_7 = _mm256_srai_epi32(s2_09_5, DCT_CONST_BITS);
+          const __m256i s2_10_6 = _mm256_srai_epi32(s2_10_4, DCT_CONST_BITS);
+          const __m256i s2_10_7 = _mm256_srai_epi32(s2_10_5, DCT_CONST_BITS);
+          const __m256i s2_13_6 = _mm256_srai_epi32(s2_13_4, DCT_CONST_BITS);
+          const __m256i s2_13_7 = _mm256_srai_epi32(s2_13_5, DCT_CONST_BITS);
+          const __m256i s2_14_6 = _mm256_srai_epi32(s2_14_4, DCT_CONST_BITS);
+          const __m256i s2_14_7 = _mm256_srai_epi32(s2_14_5, DCT_CONST_BITS);
+          // Combine
+          step2[9] = _mm256_packs_epi32(s2_09_6, s2_09_7);
+          step2[10] = _mm256_packs_epi32(s2_10_6, s2_10_7);
+          step2[13] = _mm256_packs_epi32(s2_13_6, s2_13_7);
+          step2[14] = _mm256_packs_epi32(s2_14_6, s2_14_7);
+        }
+        {
+          step2[16] = _mm256_add_epi16(step1[19], step3[16]);
+          step2[17] = _mm256_add_epi16(step1[18], step3[17]);
+          step2[18] = _mm256_sub_epi16(step3[17], step1[18]);
+          step2[19] = _mm256_sub_epi16(step3[16], step1[19]);
+          step2[20] = _mm256_sub_epi16(step3[23], step1[20]);
+          step2[21] = _mm256_sub_epi16(step3[22], step1[21]);
+          step2[22] = _mm256_add_epi16(step1[21], step3[22]);
+          step2[23] = _mm256_add_epi16(step1[20], step3[23]);
+          step2[24] = _mm256_add_epi16(step1[27], step3[24]);
+          step2[25] = _mm256_add_epi16(step1[26], step3[25]);
+          step2[26] = _mm256_sub_epi16(step3[25], step1[26]);
+          step2[27] = _mm256_sub_epi16(step3[24], step1[27]);
+          step2[28] = _mm256_sub_epi16(step3[31], step1[28]);
+          step2[29] = _mm256_sub_epi16(step3[30], step1[29]);
+          step2[30] = _mm256_add_epi16(step1[29], step3[30]);
+          step2[31] = _mm256_add_epi16(step1[28], step3[31]);
+        }
+        // Stage 6
+        {
+          const __m256i out_04_0 = _mm256_unpacklo_epi16(step2[4], step2[7]);
+          const __m256i out_04_1 = _mm256_unpackhi_epi16(step2[4], step2[7]);
+          const __m256i out_20_0 = _mm256_unpacklo_epi16(step2[5], step2[6]);
+          const __m256i out_20_1 = _mm256_unpackhi_epi16(step2[5], step2[6]);
+          const __m256i out_12_0 = _mm256_unpacklo_epi16(step2[5], step2[6]);
+          const __m256i out_12_1 = _mm256_unpackhi_epi16(step2[5], step2[6]);
+          const __m256i out_28_0 = _mm256_unpacklo_epi16(step2[4], step2[7]);
+          const __m256i out_28_1 = _mm256_unpackhi_epi16(step2[4], step2[7]);
+          const __m256i out_04_2 =
+              _mm256_madd_epi16(out_04_0, k__cospi_p28_p04);
+          const __m256i out_04_3 =
+              _mm256_madd_epi16(out_04_1, k__cospi_p28_p04);
+          const __m256i out_20_2 =
+              _mm256_madd_epi16(out_20_0, k__cospi_p12_p20);
+          const __m256i out_20_3 =
+              _mm256_madd_epi16(out_20_1, k__cospi_p12_p20);
+          const __m256i out_12_2 =
+              _mm256_madd_epi16(out_12_0, k__cospi_m20_p12);
+          const __m256i out_12_3 =
+              _mm256_madd_epi16(out_12_1, k__cospi_m20_p12);
+          const __m256i out_28_2 =
+              _mm256_madd_epi16(out_28_0, k__cospi_m04_p28);
+          const __m256i out_28_3 =
+              _mm256_madd_epi16(out_28_1, k__cospi_m04_p28);
+          // dct_const_round_shift
+          const __m256i out_04_4 =
+              _mm256_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_04_5 =
+              _mm256_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_20_4 =
+              _mm256_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_20_5 =
+              _mm256_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_12_4 =
+              _mm256_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_12_5 =
+              _mm256_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_28_4 =
+              _mm256_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_28_5 =
+              _mm256_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_04_6 = _mm256_srai_epi32(out_04_4, DCT_CONST_BITS);
+          const __m256i out_04_7 = _mm256_srai_epi32(out_04_5, DCT_CONST_BITS);
+          const __m256i out_20_6 = _mm256_srai_epi32(out_20_4, DCT_CONST_BITS);
+          const __m256i out_20_7 = _mm256_srai_epi32(out_20_5, DCT_CONST_BITS);
+          const __m256i out_12_6 = _mm256_srai_epi32(out_12_4, DCT_CONST_BITS);
+          const __m256i out_12_7 = _mm256_srai_epi32(out_12_5, DCT_CONST_BITS);
+          const __m256i out_28_6 = _mm256_srai_epi32(out_28_4, DCT_CONST_BITS);
+          const __m256i out_28_7 = _mm256_srai_epi32(out_28_5, DCT_CONST_BITS);
+          // Combine
+          out[4] = _mm256_packs_epi32(out_04_6, out_04_7);
+          out[20] = _mm256_packs_epi32(out_20_6, out_20_7);
+          out[12] = _mm256_packs_epi32(out_12_6, out_12_7);
+          out[28] = _mm256_packs_epi32(out_28_6, out_28_7);
+        }
+        {
+          step3[8] = _mm256_add_epi16(step2[9], step1[8]);
+          step3[9] = _mm256_sub_epi16(step1[8], step2[9]);
+          step3[10] = _mm256_sub_epi16(step1[11], step2[10]);
+          step3[11] = _mm256_add_epi16(step2[10], step1[11]);
+          step3[12] = _mm256_add_epi16(step2[13], step1[12]);
+          step3[13] = _mm256_sub_epi16(step1[12], step2[13]);
+          step3[14] = _mm256_sub_epi16(step1[15], step2[14]);
+          step3[15] = _mm256_add_epi16(step2[14], step1[15]);
+        }
+        {
+          const __m256i s3_17_0 = _mm256_unpacklo_epi16(step2[17], step2[30]);
+          const __m256i s3_17_1 = _mm256_unpackhi_epi16(step2[17], step2[30]);
+          const __m256i s3_18_0 = _mm256_unpacklo_epi16(step2[18], step2[29]);
+          const __m256i s3_18_1 = _mm256_unpackhi_epi16(step2[18], step2[29]);
+          const __m256i s3_21_0 = _mm256_unpacklo_epi16(step2[21], step2[26]);
+          const __m256i s3_21_1 = _mm256_unpackhi_epi16(step2[21], step2[26]);
+          const __m256i s3_22_0 = _mm256_unpacklo_epi16(step2[22], step2[25]);
+          const __m256i s3_22_1 = _mm256_unpackhi_epi16(step2[22], step2[25]);
+          const __m256i s3_17_2 = _mm256_madd_epi16(s3_17_0, k__cospi_m04_p28);
+          const __m256i s3_17_3 = _mm256_madd_epi16(s3_17_1, k__cospi_m04_p28);
+          const __m256i s3_18_2 = _mm256_madd_epi16(s3_18_0, k__cospi_m28_m04);
+          const __m256i s3_18_3 = _mm256_madd_epi16(s3_18_1, k__cospi_m28_m04);
+          const __m256i s3_21_2 = _mm256_madd_epi16(s3_21_0, k__cospi_m20_p12);
+          const __m256i s3_21_3 = _mm256_madd_epi16(s3_21_1, k__cospi_m20_p12);
+          const __m256i s3_22_2 = _mm256_madd_epi16(s3_22_0, k__cospi_m12_m20);
+          const __m256i s3_22_3 = _mm256_madd_epi16(s3_22_1, k__cospi_m12_m20);
+          const __m256i s3_25_2 = _mm256_madd_epi16(s3_22_0, k__cospi_m20_p12);
+          const __m256i s3_25_3 = _mm256_madd_epi16(s3_22_1, k__cospi_m20_p12);
+          const __m256i s3_26_2 = _mm256_madd_epi16(s3_21_0, k__cospi_p12_p20);
+          const __m256i s3_26_3 = _mm256_madd_epi16(s3_21_1, k__cospi_p12_p20);
+          const __m256i s3_29_2 = _mm256_madd_epi16(s3_18_0, k__cospi_m04_p28);
+          const __m256i s3_29_3 = _mm256_madd_epi16(s3_18_1, k__cospi_m04_p28);
+          const __m256i s3_30_2 = _mm256_madd_epi16(s3_17_0, k__cospi_p28_p04);
+          const __m256i s3_30_3 = _mm256_madd_epi16(s3_17_1, k__cospi_p28_p04);
+          // dct_const_round_shift
+          const __m256i s3_17_4 =
+              _mm256_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_17_5 =
+              _mm256_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_18_4 =
+              _mm256_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_18_5 =
+              _mm256_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_21_4 =
+              _mm256_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_21_5 =
+              _mm256_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_22_4 =
+              _mm256_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_22_5 =
+              _mm256_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_17_6 = _mm256_srai_epi32(s3_17_4, DCT_CONST_BITS);
+          const __m256i s3_17_7 = _mm256_srai_epi32(s3_17_5, DCT_CONST_BITS);
+          const __m256i s3_18_6 = _mm256_srai_epi32(s3_18_4, DCT_CONST_BITS);
+          const __m256i s3_18_7 = _mm256_srai_epi32(s3_18_5, DCT_CONST_BITS);
+          const __m256i s3_21_6 = _mm256_srai_epi32(s3_21_4, DCT_CONST_BITS);
+          const __m256i s3_21_7 = _mm256_srai_epi32(s3_21_5, DCT_CONST_BITS);
+          const __m256i s3_22_6 = _mm256_srai_epi32(s3_22_4, DCT_CONST_BITS);
+          const __m256i s3_22_7 = _mm256_srai_epi32(s3_22_5, DCT_CONST_BITS);
+          const __m256i s3_25_4 =
+              _mm256_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_25_5 =
+              _mm256_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_26_4 =
+              _mm256_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_26_5 =
+              _mm256_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_29_4 =
+              _mm256_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_29_5 =
+              _mm256_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_30_4 =
+              _mm256_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_30_5 =
+              _mm256_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_25_6 = _mm256_srai_epi32(s3_25_4, DCT_CONST_BITS);
+          const __m256i s3_25_7 = _mm256_srai_epi32(s3_25_5, DCT_CONST_BITS);
+          const __m256i s3_26_6 = _mm256_srai_epi32(s3_26_4, DCT_CONST_BITS);
+          const __m256i s3_26_7 = _mm256_srai_epi32(s3_26_5, DCT_CONST_BITS);
+          const __m256i s3_29_6 = _mm256_srai_epi32(s3_29_4, DCT_CONST_BITS);
+          const __m256i s3_29_7 = _mm256_srai_epi32(s3_29_5, DCT_CONST_BITS);
+          const __m256i s3_30_6 = _mm256_srai_epi32(s3_30_4, DCT_CONST_BITS);
+          const __m256i s3_30_7 = _mm256_srai_epi32(s3_30_5, DCT_CONST_BITS);
+          // Combine
+          step3[17] = _mm256_packs_epi32(s3_17_6, s3_17_7);
+          step3[18] = _mm256_packs_epi32(s3_18_6, s3_18_7);
+          step3[21] = _mm256_packs_epi32(s3_21_6, s3_21_7);
+          step3[22] = _mm256_packs_epi32(s3_22_6, s3_22_7);
+          // Combine
+          step3[25] = _mm256_packs_epi32(s3_25_6, s3_25_7);
+          step3[26] = _mm256_packs_epi32(s3_26_6, s3_26_7);
+          step3[29] = _mm256_packs_epi32(s3_29_6, s3_29_7);
+          step3[30] = _mm256_packs_epi32(s3_30_6, s3_30_7);
+        }
+        // Stage 7
+        {
+          const __m256i out_02_0 = _mm256_unpacklo_epi16(step3[8], step3[15]);
+          const __m256i out_02_1 = _mm256_unpackhi_epi16(step3[8], step3[15]);
+          const __m256i out_18_0 = _mm256_unpacklo_epi16(step3[9], step3[14]);
+          const __m256i out_18_1 = _mm256_unpackhi_epi16(step3[9], step3[14]);
+          const __m256i out_10_0 = _mm256_unpacklo_epi16(step3[10], step3[13]);
+          const __m256i out_10_1 = _mm256_unpackhi_epi16(step3[10], step3[13]);
+          const __m256i out_26_0 = _mm256_unpacklo_epi16(step3[11], step3[12]);
+          const __m256i out_26_1 = _mm256_unpackhi_epi16(step3[11], step3[12]);
+          const __m256i out_02_2 =
+              _mm256_madd_epi16(out_02_0, k__cospi_p30_p02);
+          const __m256i out_02_3 =
+              _mm256_madd_epi16(out_02_1, k__cospi_p30_p02);
+          const __m256i out_18_2 =
+              _mm256_madd_epi16(out_18_0, k__cospi_p14_p18);
+          const __m256i out_18_3 =
+              _mm256_madd_epi16(out_18_1, k__cospi_p14_p18);
+          const __m256i out_10_2 =
+              _mm256_madd_epi16(out_10_0, k__cospi_p22_p10);
+          const __m256i out_10_3 =
+              _mm256_madd_epi16(out_10_1, k__cospi_p22_p10);
+          const __m256i out_26_2 =
+              _mm256_madd_epi16(out_26_0, k__cospi_p06_p26);
+          const __m256i out_26_3 =
+              _mm256_madd_epi16(out_26_1, k__cospi_p06_p26);
+          const __m256i out_06_2 =
+              _mm256_madd_epi16(out_26_0, k__cospi_m26_p06);
+          const __m256i out_06_3 =
+              _mm256_madd_epi16(out_26_1, k__cospi_m26_p06);
+          const __m256i out_22_2 =
+              _mm256_madd_epi16(out_10_0, k__cospi_m10_p22);
+          const __m256i out_22_3 =
+              _mm256_madd_epi16(out_10_1, k__cospi_m10_p22);
+          const __m256i out_14_2 =
+              _mm256_madd_epi16(out_18_0, k__cospi_m18_p14);
+          const __m256i out_14_3 =
+              _mm256_madd_epi16(out_18_1, k__cospi_m18_p14);
+          const __m256i out_30_2 =
+              _mm256_madd_epi16(out_02_0, k__cospi_m02_p30);
+          const __m256i out_30_3 =
+              _mm256_madd_epi16(out_02_1, k__cospi_m02_p30);
+          // dct_const_round_shift
+          const __m256i out_02_4 =
+              _mm256_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_02_5 =
+              _mm256_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_18_4 =
+              _mm256_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_18_5 =
+              _mm256_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_10_4 =
+              _mm256_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_10_5 =
+              _mm256_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_26_4 =
+              _mm256_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_26_5 =
+              _mm256_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_06_4 =
+              _mm256_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_06_5 =
+              _mm256_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_22_4 =
+              _mm256_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_22_5 =
+              _mm256_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_14_4 =
+              _mm256_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_14_5 =
+              _mm256_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_30_4 =
+              _mm256_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_30_5 =
+              _mm256_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_02_6 = _mm256_srai_epi32(out_02_4, DCT_CONST_BITS);
+          const __m256i out_02_7 = _mm256_srai_epi32(out_02_5, DCT_CONST_BITS);
+          const __m256i out_18_6 = _mm256_srai_epi32(out_18_4, DCT_CONST_BITS);
+          const __m256i out_18_7 = _mm256_srai_epi32(out_18_5, DCT_CONST_BITS);
+          const __m256i out_10_6 = _mm256_srai_epi32(out_10_4, DCT_CONST_BITS);
+          const __m256i out_10_7 = _mm256_srai_epi32(out_10_5, DCT_CONST_BITS);
+          const __m256i out_26_6 = _mm256_srai_epi32(out_26_4, DCT_CONST_BITS);
+          const __m256i out_26_7 = _mm256_srai_epi32(out_26_5, DCT_CONST_BITS);
+          const __m256i out_06_6 = _mm256_srai_epi32(out_06_4, DCT_CONST_BITS);
+          const __m256i out_06_7 = _mm256_srai_epi32(out_06_5, DCT_CONST_BITS);
+          const __m256i out_22_6 = _mm256_srai_epi32(out_22_4, DCT_CONST_BITS);
+          const __m256i out_22_7 = _mm256_srai_epi32(out_22_5, DCT_CONST_BITS);
+          const __m256i out_14_6 = _mm256_srai_epi32(out_14_4, DCT_CONST_BITS);
+          const __m256i out_14_7 = _mm256_srai_epi32(out_14_5, DCT_CONST_BITS);
+          const __m256i out_30_6 = _mm256_srai_epi32(out_30_4, DCT_CONST_BITS);
+          const __m256i out_30_7 = _mm256_srai_epi32(out_30_5, DCT_CONST_BITS);
+          // Combine
+          out[2] = _mm256_packs_epi32(out_02_6, out_02_7);
+          out[18] = _mm256_packs_epi32(out_18_6, out_18_7);
+          out[10] = _mm256_packs_epi32(out_10_6, out_10_7);
+          out[26] = _mm256_packs_epi32(out_26_6, out_26_7);
+          out[6] = _mm256_packs_epi32(out_06_6, out_06_7);
+          out[22] = _mm256_packs_epi32(out_22_6, out_22_7);
+          out[14] = _mm256_packs_epi32(out_14_6, out_14_7);
+          out[30] = _mm256_packs_epi32(out_30_6, out_30_7);
+        }
+        {
+          step1[16] = _mm256_add_epi16(step3[17], step2[16]);
+          step1[17] = _mm256_sub_epi16(step2[16], step3[17]);
+          step1[18] = _mm256_sub_epi16(step2[19], step3[18]);
+          step1[19] = _mm256_add_epi16(step3[18], step2[19]);
+          step1[20] = _mm256_add_epi16(step3[21], step2[20]);
+          step1[21] = _mm256_sub_epi16(step2[20], step3[21]);
+          step1[22] = _mm256_sub_epi16(step2[23], step3[22]);
+          step1[23] = _mm256_add_epi16(step3[22], step2[23]);
+          step1[24] = _mm256_add_epi16(step3[25], step2[24]);
+          step1[25] = _mm256_sub_epi16(step2[24], step3[25]);
+          step1[26] = _mm256_sub_epi16(step2[27], step3[26]);
+          step1[27] = _mm256_add_epi16(step3[26], step2[27]);
+          step1[28] = _mm256_add_epi16(step3[29], step2[28]);
+          step1[29] = _mm256_sub_epi16(step2[28], step3[29]);
+          step1[30] = _mm256_sub_epi16(step2[31], step3[30]);
+          step1[31] = _mm256_add_epi16(step3[30], step2[31]);
+        }
+        // Final stage --- outputs indices are bit-reversed.
+        {
+          const __m256i out_01_0 = _mm256_unpacklo_epi16(step1[16], step1[31]);
+          const __m256i out_01_1 = _mm256_unpackhi_epi16(step1[16], step1[31]);
+          const __m256i out_17_0 = _mm256_unpacklo_epi16(step1[17], step1[30]);
+          const __m256i out_17_1 = _mm256_unpackhi_epi16(step1[17], step1[30]);
+          const __m256i out_09_0 = _mm256_unpacklo_epi16(step1[18], step1[29]);
+          const __m256i out_09_1 = _mm256_unpackhi_epi16(step1[18], step1[29]);
+          const __m256i out_25_0 = _mm256_unpacklo_epi16(step1[19], step1[28]);
+          const __m256i out_25_1 = _mm256_unpackhi_epi16(step1[19], step1[28]);
+          const __m256i out_01_2 =
+              _mm256_madd_epi16(out_01_0, k__cospi_p31_p01);
+          const __m256i out_01_3 =
+              _mm256_madd_epi16(out_01_1, k__cospi_p31_p01);
+          const __m256i out_17_2 =
+              _mm256_madd_epi16(out_17_0, k__cospi_p15_p17);
+          const __m256i out_17_3 =
+              _mm256_madd_epi16(out_17_1, k__cospi_p15_p17);
+          const __m256i out_09_2 =
+              _mm256_madd_epi16(out_09_0, k__cospi_p23_p09);
+          const __m256i out_09_3 =
+              _mm256_madd_epi16(out_09_1, k__cospi_p23_p09);
+          const __m256i out_25_2 =
+              _mm256_madd_epi16(out_25_0, k__cospi_p07_p25);
+          const __m256i out_25_3 =
+              _mm256_madd_epi16(out_25_1, k__cospi_p07_p25);
+          const __m256i out_07_2 =
+              _mm256_madd_epi16(out_25_0, k__cospi_m25_p07);
+          const __m256i out_07_3 =
+              _mm256_madd_epi16(out_25_1, k__cospi_m25_p07);
+          const __m256i out_23_2 =
+              _mm256_madd_epi16(out_09_0, k__cospi_m09_p23);
+          const __m256i out_23_3 =
+              _mm256_madd_epi16(out_09_1, k__cospi_m09_p23);
+          const __m256i out_15_2 =
+              _mm256_madd_epi16(out_17_0, k__cospi_m17_p15);
+          const __m256i out_15_3 =
+              _mm256_madd_epi16(out_17_1, k__cospi_m17_p15);
+          const __m256i out_31_2 =
+              _mm256_madd_epi16(out_01_0, k__cospi_m01_p31);
+          const __m256i out_31_3 =
+              _mm256_madd_epi16(out_01_1, k__cospi_m01_p31);
+          // dct_const_round_shift
+          const __m256i out_01_4 =
+              _mm256_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_01_5 =
+              _mm256_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_17_4 =
+              _mm256_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_17_5 =
+              _mm256_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_09_4 =
+              _mm256_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_09_5 =
+              _mm256_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_25_4 =
+              _mm256_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_25_5 =
+              _mm256_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_07_4 =
+              _mm256_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_07_5 =
+              _mm256_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_23_4 =
+              _mm256_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_23_5 =
+              _mm256_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_15_4 =
+              _mm256_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_15_5 =
+              _mm256_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_31_4 =
+              _mm256_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_31_5 =
+              _mm256_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_01_6 = _mm256_srai_epi32(out_01_4, DCT_CONST_BITS);
+          const __m256i out_01_7 = _mm256_srai_epi32(out_01_5, DCT_CONST_BITS);
+          const __m256i out_17_6 = _mm256_srai_epi32(out_17_4, DCT_CONST_BITS);
+          const __m256i out_17_7 = _mm256_srai_epi32(out_17_5, DCT_CONST_BITS);
+          const __m256i out_09_6 = _mm256_srai_epi32(out_09_4, DCT_CONST_BITS);
+          const __m256i out_09_7 = _mm256_srai_epi32(out_09_5, DCT_CONST_BITS);
+          const __m256i out_25_6 = _mm256_srai_epi32(out_25_4, DCT_CONST_BITS);
+          const __m256i out_25_7 = _mm256_srai_epi32(out_25_5, DCT_CONST_BITS);
+          const __m256i out_07_6 = _mm256_srai_epi32(out_07_4, DCT_CONST_BITS);
+          const __m256i out_07_7 = _mm256_srai_epi32(out_07_5, DCT_CONST_BITS);
+          const __m256i out_23_6 = _mm256_srai_epi32(out_23_4, DCT_CONST_BITS);
+          const __m256i out_23_7 = _mm256_srai_epi32(out_23_5, DCT_CONST_BITS);
+          const __m256i out_15_6 = _mm256_srai_epi32(out_15_4, DCT_CONST_BITS);
+          const __m256i out_15_7 = _mm256_srai_epi32(out_15_5, DCT_CONST_BITS);
+          const __m256i out_31_6 = _mm256_srai_epi32(out_31_4, DCT_CONST_BITS);
+          const __m256i out_31_7 = _mm256_srai_epi32(out_31_5, DCT_CONST_BITS);
+          // Combine
+          out[1] = _mm256_packs_epi32(out_01_6, out_01_7);
+          out[17] = _mm256_packs_epi32(out_17_6, out_17_7);
+          out[9] = _mm256_packs_epi32(out_09_6, out_09_7);
+          out[25] = _mm256_packs_epi32(out_25_6, out_25_7);
+          out[7] = _mm256_packs_epi32(out_07_6, out_07_7);
+          out[23] = _mm256_packs_epi32(out_23_6, out_23_7);
+          out[15] = _mm256_packs_epi32(out_15_6, out_15_7);
+          out[31] = _mm256_packs_epi32(out_31_6, out_31_7);
+        }
+        {
+          const __m256i out_05_0 = _mm256_unpacklo_epi16(step1[20], step1[27]);
+          const __m256i out_05_1 = _mm256_unpackhi_epi16(step1[20], step1[27]);
+          const __m256i out_21_0 = _mm256_unpacklo_epi16(step1[21], step1[26]);
+          const __m256i out_21_1 = _mm256_unpackhi_epi16(step1[21], step1[26]);
+          const __m256i out_13_0 = _mm256_unpacklo_epi16(step1[22], step1[25]);
+          const __m256i out_13_1 = _mm256_unpackhi_epi16(step1[22], step1[25]);
+          const __m256i out_29_0 = _mm256_unpacklo_epi16(step1[23], step1[24]);
+          const __m256i out_29_1 = _mm256_unpackhi_epi16(step1[23], step1[24]);
+          const __m256i out_05_2 =
+              _mm256_madd_epi16(out_05_0, k__cospi_p27_p05);
+          const __m256i out_05_3 =
+              _mm256_madd_epi16(out_05_1, k__cospi_p27_p05);
+          const __m256i out_21_2 =
+              _mm256_madd_epi16(out_21_0, k__cospi_p11_p21);
+          const __m256i out_21_3 =
+              _mm256_madd_epi16(out_21_1, k__cospi_p11_p21);
+          const __m256i out_13_2 =
+              _mm256_madd_epi16(out_13_0, k__cospi_p19_p13);
+          const __m256i out_13_3 =
+              _mm256_madd_epi16(out_13_1, k__cospi_p19_p13);
+          const __m256i out_29_2 =
+              _mm256_madd_epi16(out_29_0, k__cospi_p03_p29);
+          const __m256i out_29_3 =
+              _mm256_madd_epi16(out_29_1, k__cospi_p03_p29);
+          const __m256i out_03_2 =
+              _mm256_madd_epi16(out_29_0, k__cospi_m29_p03);
+          const __m256i out_03_3 =
+              _mm256_madd_epi16(out_29_1, k__cospi_m29_p03);
+          const __m256i out_19_2 =
+              _mm256_madd_epi16(out_13_0, k__cospi_m13_p19);
+          const __m256i out_19_3 =
+              _mm256_madd_epi16(out_13_1, k__cospi_m13_p19);
+          const __m256i out_11_2 =
+              _mm256_madd_epi16(out_21_0, k__cospi_m21_p11);
+          const __m256i out_11_3 =
+              _mm256_madd_epi16(out_21_1, k__cospi_m21_p11);
+          const __m256i out_27_2 =
+              _mm256_madd_epi16(out_05_0, k__cospi_m05_p27);
+          const __m256i out_27_3 =
+              _mm256_madd_epi16(out_05_1, k__cospi_m05_p27);
+          // dct_const_round_shift
+          const __m256i out_05_4 =
+              _mm256_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_05_5 =
+              _mm256_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_21_4 =
+              _mm256_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_21_5 =
+              _mm256_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_13_4 =
+              _mm256_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_13_5 =
+              _mm256_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_29_4 =
+              _mm256_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_29_5 =
+              _mm256_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_03_4 =
+              _mm256_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_03_5 =
+              _mm256_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_19_4 =
+              _mm256_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_19_5 =
+              _mm256_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_11_4 =
+              _mm256_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_11_5 =
+              _mm256_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_27_4 =
+              _mm256_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
+          const __m256i out_27_5 =
+              _mm256_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
+          const __m256i out_05_6 = _mm256_srai_epi32(out_05_4, DCT_CONST_BITS);
+          const __m256i out_05_7 = _mm256_srai_epi32(out_05_5, DCT_CONST_BITS);
+          const __m256i out_21_6 = _mm256_srai_epi32(out_21_4, DCT_CONST_BITS);
+          const __m256i out_21_7 = _mm256_srai_epi32(out_21_5, DCT_CONST_BITS);
+          const __m256i out_13_6 = _mm256_srai_epi32(out_13_4, DCT_CONST_BITS);
+          const __m256i out_13_7 = _mm256_srai_epi32(out_13_5, DCT_CONST_BITS);
+          const __m256i out_29_6 = _mm256_srai_epi32(out_29_4, DCT_CONST_BITS);
+          const __m256i out_29_7 = _mm256_srai_epi32(out_29_5, DCT_CONST_BITS);
+          const __m256i out_03_6 = _mm256_srai_epi32(out_03_4, DCT_CONST_BITS);
+          const __m256i out_03_7 = _mm256_srai_epi32(out_03_5, DCT_CONST_BITS);
+          const __m256i out_19_6 = _mm256_srai_epi32(out_19_4, DCT_CONST_BITS);
+          const __m256i out_19_7 = _mm256_srai_epi32(out_19_5, DCT_CONST_BITS);
+          const __m256i out_11_6 = _mm256_srai_epi32(out_11_4, DCT_CONST_BITS);
+          const __m256i out_11_7 = _mm256_srai_epi32(out_11_5, DCT_CONST_BITS);
+          const __m256i out_27_6 = _mm256_srai_epi32(out_27_4, DCT_CONST_BITS);
+          const __m256i out_27_7 = _mm256_srai_epi32(out_27_5, DCT_CONST_BITS);
+          // Combine
+          out[5] = _mm256_packs_epi32(out_05_6, out_05_7);
+          out[21] = _mm256_packs_epi32(out_21_6, out_21_7);
+          out[13] = _mm256_packs_epi32(out_13_6, out_13_7);
+          out[29] = _mm256_packs_epi32(out_29_6, out_29_7);
+          out[3] = _mm256_packs_epi32(out_03_6, out_03_7);
+          out[19] = _mm256_packs_epi32(out_19_6, out_19_7);
+          out[11] = _mm256_packs_epi32(out_11_6, out_11_7);
+          out[27] = _mm256_packs_epi32(out_27_6, out_27_7);
+        }
+#if FDCT32x32_HIGH_PRECISION
+      } else {
+        __m256i lstep1[64], lstep2[64], lstep3[64];
+        __m256i u[32], v[32], sign[16];
+        const __m256i K32One = _mm256_set_epi32(1, 1, 1, 1, 1, 1, 1, 1);
+        // start using 32-bit operations
+        // stage 3
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[0] = _mm256_unpacklo_epi16(step2[0], kZero);
+          lstep2[1] = _mm256_unpackhi_epi16(step2[0], kZero);
+          lstep2[2] = _mm256_unpacklo_epi16(step2[1], kZero);
+          lstep2[3] = _mm256_unpackhi_epi16(step2[1], kZero);
+          lstep2[4] = _mm256_unpacklo_epi16(step2[2], kZero);
+          lstep2[5] = _mm256_unpackhi_epi16(step2[2], kZero);
+          lstep2[6] = _mm256_unpacklo_epi16(step2[3], kZero);
+          lstep2[7] = _mm256_unpackhi_epi16(step2[3], kZero);
+          lstep2[8] = _mm256_unpacklo_epi16(step2[4], kZero);
+          lstep2[9] = _mm256_unpackhi_epi16(step2[4], kZero);
+          lstep2[10] = _mm256_unpacklo_epi16(step2[5], kZero);
+          lstep2[11] = _mm256_unpackhi_epi16(step2[5], kZero);
+          lstep2[12] = _mm256_unpacklo_epi16(step2[6], kZero);
+          lstep2[13] = _mm256_unpackhi_epi16(step2[6], kZero);
+          lstep2[14] = _mm256_unpacklo_epi16(step2[7], kZero);
+          lstep2[15] = _mm256_unpackhi_epi16(step2[7], kZero);
+          lstep2[0] = _mm256_madd_epi16(lstep2[0], kOne);
+          lstep2[1] = _mm256_madd_epi16(lstep2[1], kOne);
+          lstep2[2] = _mm256_madd_epi16(lstep2[2], kOne);
+          lstep2[3] = _mm256_madd_epi16(lstep2[3], kOne);
+          lstep2[4] = _mm256_madd_epi16(lstep2[4], kOne);
+          lstep2[5] = _mm256_madd_epi16(lstep2[5], kOne);
+          lstep2[6] = _mm256_madd_epi16(lstep2[6], kOne);
+          lstep2[7] = _mm256_madd_epi16(lstep2[7], kOne);
+          lstep2[8] = _mm256_madd_epi16(lstep2[8], kOne);
+          lstep2[9] = _mm256_madd_epi16(lstep2[9], kOne);
+          lstep2[10] = _mm256_madd_epi16(lstep2[10], kOne);
+          lstep2[11] = _mm256_madd_epi16(lstep2[11], kOne);
+          lstep2[12] = _mm256_madd_epi16(lstep2[12], kOne);
+          lstep2[13] = _mm256_madd_epi16(lstep2[13], kOne);
+          lstep2[14] = _mm256_madd_epi16(lstep2[14], kOne);
+          lstep2[15] = _mm256_madd_epi16(lstep2[15], kOne);
+
+          lstep3[0] = _mm256_add_epi32(lstep2[14], lstep2[0]);
+          lstep3[1] = _mm256_add_epi32(lstep2[15], lstep2[1]);
+          lstep3[2] = _mm256_add_epi32(lstep2[12], lstep2[2]);
+          lstep3[3] = _mm256_add_epi32(lstep2[13], lstep2[3]);
+          lstep3[4] = _mm256_add_epi32(lstep2[10], lstep2[4]);
+          lstep3[5] = _mm256_add_epi32(lstep2[11], lstep2[5]);
+          lstep3[6] = _mm256_add_epi32(lstep2[8], lstep2[6]);
+          lstep3[7] = _mm256_add_epi32(lstep2[9], lstep2[7]);
+          lstep3[8] = _mm256_sub_epi32(lstep2[6], lstep2[8]);
+          lstep3[9] = _mm256_sub_epi32(lstep2[7], lstep2[9]);
+          lstep3[10] = _mm256_sub_epi32(lstep2[4], lstep2[10]);
+          lstep3[11] = _mm256_sub_epi32(lstep2[5], lstep2[11]);
+          lstep3[12] = _mm256_sub_epi32(lstep2[2], lstep2[12]);
+          lstep3[13] = _mm256_sub_epi32(lstep2[3], lstep2[13]);
+          lstep3[14] = _mm256_sub_epi32(lstep2[0], lstep2[14]);
+          lstep3[15] = _mm256_sub_epi32(lstep2[1], lstep2[15]);
+        }
+        {
+          const __m256i s3_10_0 = _mm256_unpacklo_epi16(step2[13], step2[10]);
+          const __m256i s3_10_1 = _mm256_unpackhi_epi16(step2[13], step2[10]);
+          const __m256i s3_11_0 = _mm256_unpacklo_epi16(step2[12], step2[11]);
+          const __m256i s3_11_1 = _mm256_unpackhi_epi16(step2[12], step2[11]);
+          const __m256i s3_10_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_m16);
+          const __m256i s3_10_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_m16);
+          const __m256i s3_11_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_m16);
+          const __m256i s3_11_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_m16);
+          const __m256i s3_12_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_p16);
+          const __m256i s3_12_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_p16);
+          const __m256i s3_13_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_p16);
+          const __m256i s3_13_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_p16);
+          // dct_const_round_shift
+          const __m256i s3_10_4 =
+              _mm256_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_10_5 =
+              _mm256_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_11_4 =
+              _mm256_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_11_5 =
+              _mm256_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_12_4 =
+              _mm256_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_12_5 =
+              _mm256_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_13_4 =
+              _mm256_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_13_5 =
+              _mm256_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+          lstep3[20] = _mm256_srai_epi32(s3_10_4, DCT_CONST_BITS);
+          lstep3[21] = _mm256_srai_epi32(s3_10_5, DCT_CONST_BITS);
+          lstep3[22] = _mm256_srai_epi32(s3_11_4, DCT_CONST_BITS);
+          lstep3[23] = _mm256_srai_epi32(s3_11_5, DCT_CONST_BITS);
+          lstep3[24] = _mm256_srai_epi32(s3_12_4, DCT_CONST_BITS);
+          lstep3[25] = _mm256_srai_epi32(s3_12_5, DCT_CONST_BITS);
+          lstep3[26] = _mm256_srai_epi32(s3_13_4, DCT_CONST_BITS);
+          lstep3[27] = _mm256_srai_epi32(s3_13_5, DCT_CONST_BITS);
+        }
+        {
+          lstep2[40] = _mm256_unpacklo_epi16(step2[20], kZero);
+          lstep2[41] = _mm256_unpackhi_epi16(step2[20], kZero);
+          lstep2[42] = _mm256_unpacklo_epi16(step2[21], kZero);
+          lstep2[43] = _mm256_unpackhi_epi16(step2[21], kZero);
+          lstep2[44] = _mm256_unpacklo_epi16(step2[22], kZero);
+          lstep2[45] = _mm256_unpackhi_epi16(step2[22], kZero);
+          lstep2[46] = _mm256_unpacklo_epi16(step2[23], kZero);
+          lstep2[47] = _mm256_unpackhi_epi16(step2[23], kZero);
+          lstep2[48] = _mm256_unpacklo_epi16(step2[24], kZero);
+          lstep2[49] = _mm256_unpackhi_epi16(step2[24], kZero);
+          lstep2[50] = _mm256_unpacklo_epi16(step2[25], kZero);
+          lstep2[51] = _mm256_unpackhi_epi16(step2[25], kZero);
+          lstep2[52] = _mm256_unpacklo_epi16(step2[26], kZero);
+          lstep2[53] = _mm256_unpackhi_epi16(step2[26], kZero);
+          lstep2[54] = _mm256_unpacklo_epi16(step2[27], kZero);
+          lstep2[55] = _mm256_unpackhi_epi16(step2[27], kZero);
+          lstep2[40] = _mm256_madd_epi16(lstep2[40], kOne);
+          lstep2[41] = _mm256_madd_epi16(lstep2[41], kOne);
+          lstep2[42] = _mm256_madd_epi16(lstep2[42], kOne);
+          lstep2[43] = _mm256_madd_epi16(lstep2[43], kOne);
+          lstep2[44] = _mm256_madd_epi16(lstep2[44], kOne);
+          lstep2[45] = _mm256_madd_epi16(lstep2[45], kOne);
+          lstep2[46] = _mm256_madd_epi16(lstep2[46], kOne);
+          lstep2[47] = _mm256_madd_epi16(lstep2[47], kOne);
+          lstep2[48] = _mm256_madd_epi16(lstep2[48], kOne);
+          lstep2[49] = _mm256_madd_epi16(lstep2[49], kOne);
+          lstep2[50] = _mm256_madd_epi16(lstep2[50], kOne);
+          lstep2[51] = _mm256_madd_epi16(lstep2[51], kOne);
+          lstep2[52] = _mm256_madd_epi16(lstep2[52], kOne);
+          lstep2[53] = _mm256_madd_epi16(lstep2[53], kOne);
+          lstep2[54] = _mm256_madd_epi16(lstep2[54], kOne);
+          lstep2[55] = _mm256_madd_epi16(lstep2[55], kOne);
+
+          lstep1[32] = _mm256_unpacklo_epi16(step1[16], kZero);
+          lstep1[33] = _mm256_unpackhi_epi16(step1[16], kZero);
+          lstep1[34] = _mm256_unpacklo_epi16(step1[17], kZero);
+          lstep1[35] = _mm256_unpackhi_epi16(step1[17], kZero);
+          lstep1[36] = _mm256_unpacklo_epi16(step1[18], kZero);
+          lstep1[37] = _mm256_unpackhi_epi16(step1[18], kZero);
+          lstep1[38] = _mm256_unpacklo_epi16(step1[19], kZero);
+          lstep1[39] = _mm256_unpackhi_epi16(step1[19], kZero);
+          lstep1[56] = _mm256_unpacklo_epi16(step1[28], kZero);
+          lstep1[57] = _mm256_unpackhi_epi16(step1[28], kZero);
+          lstep1[58] = _mm256_unpacklo_epi16(step1[29], kZero);
+          lstep1[59] = _mm256_unpackhi_epi16(step1[29], kZero);
+          lstep1[60] = _mm256_unpacklo_epi16(step1[30], kZero);
+          lstep1[61] = _mm256_unpackhi_epi16(step1[30], kZero);
+          lstep1[62] = _mm256_unpacklo_epi16(step1[31], kZero);
+          lstep1[63] = _mm256_unpackhi_epi16(step1[31], kZero);
+          lstep1[32] = _mm256_madd_epi16(lstep1[32], kOne);
+          lstep1[33] = _mm256_madd_epi16(lstep1[33], kOne);
+          lstep1[34] = _mm256_madd_epi16(lstep1[34], kOne);
+          lstep1[35] = _mm256_madd_epi16(lstep1[35], kOne);
+          lstep1[36] = _mm256_madd_epi16(lstep1[36], kOne);
+          lstep1[37] = _mm256_madd_epi16(lstep1[37], kOne);
+          lstep1[38] = _mm256_madd_epi16(lstep1[38], kOne);
+          lstep1[39] = _mm256_madd_epi16(lstep1[39], kOne);
+          lstep1[56] = _mm256_madd_epi16(lstep1[56], kOne);
+          lstep1[57] = _mm256_madd_epi16(lstep1[57], kOne);
+          lstep1[58] = _mm256_madd_epi16(lstep1[58], kOne);
+          lstep1[59] = _mm256_madd_epi16(lstep1[59], kOne);
+          lstep1[60] = _mm256_madd_epi16(lstep1[60], kOne);
+          lstep1[61] = _mm256_madd_epi16(lstep1[61], kOne);
+          lstep1[62] = _mm256_madd_epi16(lstep1[62], kOne);
+          lstep1[63] = _mm256_madd_epi16(lstep1[63], kOne);
+
+          lstep3[32] = _mm256_add_epi32(lstep2[46], lstep1[32]);
+          lstep3[33] = _mm256_add_epi32(lstep2[47], lstep1[33]);
+
+          lstep3[34] = _mm256_add_epi32(lstep2[44], lstep1[34]);
+          lstep3[35] = _mm256_add_epi32(lstep2[45], lstep1[35]);
+          lstep3[36] = _mm256_add_epi32(lstep2[42], lstep1[36]);
+          lstep3[37] = _mm256_add_epi32(lstep2[43], lstep1[37]);
+          lstep3[38] = _mm256_add_epi32(lstep2[40], lstep1[38]);
+          lstep3[39] = _mm256_add_epi32(lstep2[41], lstep1[39]);
+          lstep3[40] = _mm256_sub_epi32(lstep1[38], lstep2[40]);
+          lstep3[41] = _mm256_sub_epi32(lstep1[39], lstep2[41]);
+          lstep3[42] = _mm256_sub_epi32(lstep1[36], lstep2[42]);
+          lstep3[43] = _mm256_sub_epi32(lstep1[37], lstep2[43]);
+          lstep3[44] = _mm256_sub_epi32(lstep1[34], lstep2[44]);
+          lstep3[45] = _mm256_sub_epi32(lstep1[35], lstep2[45]);
+          lstep3[46] = _mm256_sub_epi32(lstep1[32], lstep2[46]);
+          lstep3[47] = _mm256_sub_epi32(lstep1[33], lstep2[47]);
+          lstep3[48] = _mm256_sub_epi32(lstep1[62], lstep2[48]);
+          lstep3[49] = _mm256_sub_epi32(lstep1[63], lstep2[49]);
+          lstep3[50] = _mm256_sub_epi32(lstep1[60], lstep2[50]);
+          lstep3[51] = _mm256_sub_epi32(lstep1[61], lstep2[51]);
+          lstep3[52] = _mm256_sub_epi32(lstep1[58], lstep2[52]);
+          lstep3[53] = _mm256_sub_epi32(lstep1[59], lstep2[53]);
+          lstep3[54] = _mm256_sub_epi32(lstep1[56], lstep2[54]);
+          lstep3[55] = _mm256_sub_epi32(lstep1[57], lstep2[55]);
+          lstep3[56] = _mm256_add_epi32(lstep2[54], lstep1[56]);
+          lstep3[57] = _mm256_add_epi32(lstep2[55], lstep1[57]);
+          lstep3[58] = _mm256_add_epi32(lstep2[52], lstep1[58]);
+          lstep3[59] = _mm256_add_epi32(lstep2[53], lstep1[59]);
+          lstep3[60] = _mm256_add_epi32(lstep2[50], lstep1[60]);
+          lstep3[61] = _mm256_add_epi32(lstep2[51], lstep1[61]);
+          lstep3[62] = _mm256_add_epi32(lstep2[48], lstep1[62]);
+          lstep3[63] = _mm256_add_epi32(lstep2[49], lstep1[63]);
+        }
+
+        // stage 4
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[16] = _mm256_unpacklo_epi16(step2[8], kZero);
+          lstep2[17] = _mm256_unpackhi_epi16(step2[8], kZero);
+          lstep2[18] = _mm256_unpacklo_epi16(step2[9], kZero);
+          lstep2[19] = _mm256_unpackhi_epi16(step2[9], kZero);
+          lstep2[28] = _mm256_unpacklo_epi16(step2[14], kZero);
+          lstep2[29] = _mm256_unpackhi_epi16(step2[14], kZero);
+          lstep2[30] = _mm256_unpacklo_epi16(step2[15], kZero);
+          lstep2[31] = _mm256_unpackhi_epi16(step2[15], kZero);
+          lstep2[16] = _mm256_madd_epi16(lstep2[16], kOne);
+          lstep2[17] = _mm256_madd_epi16(lstep2[17], kOne);
+          lstep2[18] = _mm256_madd_epi16(lstep2[18], kOne);
+          lstep2[19] = _mm256_madd_epi16(lstep2[19], kOne);
+          lstep2[28] = _mm256_madd_epi16(lstep2[28], kOne);
+          lstep2[29] = _mm256_madd_epi16(lstep2[29], kOne);
+          lstep2[30] = _mm256_madd_epi16(lstep2[30], kOne);
+          lstep2[31] = _mm256_madd_epi16(lstep2[31], kOne);
+
+          lstep1[0] = _mm256_add_epi32(lstep3[6], lstep3[0]);
+          lstep1[1] = _mm256_add_epi32(lstep3[7], lstep3[1]);
+          lstep1[2] = _mm256_add_epi32(lstep3[4], lstep3[2]);
+          lstep1[3] = _mm256_add_epi32(lstep3[5], lstep3[3]);
+          lstep1[4] = _mm256_sub_epi32(lstep3[2], lstep3[4]);
+          lstep1[5] = _mm256_sub_epi32(lstep3[3], lstep3[5]);
+          lstep1[6] = _mm256_sub_epi32(lstep3[0], lstep3[6]);
+          lstep1[7] = _mm256_sub_epi32(lstep3[1], lstep3[7]);
+          lstep1[16] = _mm256_add_epi32(lstep3[22], lstep2[16]);
+          lstep1[17] = _mm256_add_epi32(lstep3[23], lstep2[17]);
+          lstep1[18] = _mm256_add_epi32(lstep3[20], lstep2[18]);
+          lstep1[19] = _mm256_add_epi32(lstep3[21], lstep2[19]);
+          lstep1[20] = _mm256_sub_epi32(lstep2[18], lstep3[20]);
+          lstep1[21] = _mm256_sub_epi32(lstep2[19], lstep3[21]);
+          lstep1[22] = _mm256_sub_epi32(lstep2[16], lstep3[22]);
+          lstep1[23] = _mm256_sub_epi32(lstep2[17], lstep3[23]);
+          lstep1[24] = _mm256_sub_epi32(lstep2[30], lstep3[24]);
+          lstep1[25] = _mm256_sub_epi32(lstep2[31], lstep3[25]);
+          lstep1[26] = _mm256_sub_epi32(lstep2[28], lstep3[26]);
+          lstep1[27] = _mm256_sub_epi32(lstep2[29], lstep3[27]);
+          lstep1[28] = _mm256_add_epi32(lstep3[26], lstep2[28]);
+          lstep1[29] = _mm256_add_epi32(lstep3[27], lstep2[29]);
+          lstep1[30] = _mm256_add_epi32(lstep3[24], lstep2[30]);
+          lstep1[31] = _mm256_add_epi32(lstep3[25], lstep2[31]);
+        }
+        {
+          // to be continued...
+          //
+          const __m256i k32_p16_p16 =
+              pair256_set_epi32(cospi_16_64, cospi_16_64);
+          const __m256i k32_p16_m16 =
+              pair256_set_epi32(cospi_16_64, -cospi_16_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep3[12], lstep3[10]);
+          u[1] = _mm256_unpackhi_epi32(lstep3[12], lstep3[10]);
+          u[2] = _mm256_unpacklo_epi32(lstep3[13], lstep3[11]);
+          u[3] = _mm256_unpackhi_epi32(lstep3[13], lstep3[11]);
+
+          // TODO(jingning): manually inline k_madd_epi32_avx2_ to further hide
+          // instruction latency.
+          v[0] = k_madd_epi32_avx2(u[0], k32_p16_m16);
+          v[1] = k_madd_epi32_avx2(u[1], k32_p16_m16);
+          v[2] = k_madd_epi32_avx2(u[2], k32_p16_m16);
+          v[3] = k_madd_epi32_avx2(u[3], k32_p16_m16);
+          v[4] = k_madd_epi32_avx2(u[0], k32_p16_p16);
+          v[5] = k_madd_epi32_avx2(u[1], k32_p16_p16);
+          v[6] = k_madd_epi32_avx2(u[2], k32_p16_p16);
+          v[7] = k_madd_epi32_avx2(u[3], k32_p16_p16);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+
+          lstep1[10] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          lstep1[11] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          lstep1[12] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          lstep1[13] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+        }
+        {
+          const __m256i k32_m08_p24 =
+              pair256_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m256i k32_m24_m08 =
+              pair256_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m256i k32_p24_p08 =
+              pair256_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep3[36], lstep3[58]);
+          u[1] = _mm256_unpackhi_epi32(lstep3[36], lstep3[58]);
+          u[2] = _mm256_unpacklo_epi32(lstep3[37], lstep3[59]);
+          u[3] = _mm256_unpackhi_epi32(lstep3[37], lstep3[59]);
+          u[4] = _mm256_unpacklo_epi32(lstep3[38], lstep3[56]);
+          u[5] = _mm256_unpackhi_epi32(lstep3[38], lstep3[56]);
+          u[6] = _mm256_unpacklo_epi32(lstep3[39], lstep3[57]);
+          u[7] = _mm256_unpackhi_epi32(lstep3[39], lstep3[57]);
+          u[8] = _mm256_unpacklo_epi32(lstep3[40], lstep3[54]);
+          u[9] = _mm256_unpackhi_epi32(lstep3[40], lstep3[54]);
+          u[10] = _mm256_unpacklo_epi32(lstep3[41], lstep3[55]);
+          u[11] = _mm256_unpackhi_epi32(lstep3[41], lstep3[55]);
+          u[12] = _mm256_unpacklo_epi32(lstep3[42], lstep3[52]);
+          u[13] = _mm256_unpackhi_epi32(lstep3[42], lstep3[52]);
+          u[14] = _mm256_unpacklo_epi32(lstep3[43], lstep3[53]);
+          u[15] = _mm256_unpackhi_epi32(lstep3[43], lstep3[53]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_m08_p24);
+          v[1] = k_madd_epi32_avx2(u[1], k32_m08_p24);
+          v[2] = k_madd_epi32_avx2(u[2], k32_m08_p24);
+          v[3] = k_madd_epi32_avx2(u[3], k32_m08_p24);
+          v[4] = k_madd_epi32_avx2(u[4], k32_m08_p24);
+          v[5] = k_madd_epi32_avx2(u[5], k32_m08_p24);
+          v[6] = k_madd_epi32_avx2(u[6], k32_m08_p24);
+          v[7] = k_madd_epi32_avx2(u[7], k32_m08_p24);
+          v[8] = k_madd_epi32_avx2(u[8], k32_m24_m08);
+          v[9] = k_madd_epi32_avx2(u[9], k32_m24_m08);
+          v[10] = k_madd_epi32_avx2(u[10], k32_m24_m08);
+          v[11] = k_madd_epi32_avx2(u[11], k32_m24_m08);
+          v[12] = k_madd_epi32_avx2(u[12], k32_m24_m08);
+          v[13] = k_madd_epi32_avx2(u[13], k32_m24_m08);
+          v[14] = k_madd_epi32_avx2(u[14], k32_m24_m08);
+          v[15] = k_madd_epi32_avx2(u[15], k32_m24_m08);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m08_p24);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m08_p24);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m08_p24);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m08_p24);
+          v[20] = k_madd_epi32_avx2(u[8], k32_m08_p24);
+          v[21] = k_madd_epi32_avx2(u[9], k32_m08_p24);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m08_p24);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m08_p24);
+          v[24] = k_madd_epi32_avx2(u[4], k32_p24_p08);
+          v[25] = k_madd_epi32_avx2(u[5], k32_p24_p08);
+          v[26] = k_madd_epi32_avx2(u[6], k32_p24_p08);
+          v[27] = k_madd_epi32_avx2(u[7], k32_p24_p08);
+          v[28] = k_madd_epi32_avx2(u[0], k32_p24_p08);
+          v[29] = k_madd_epi32_avx2(u[1], k32_p24_p08);
+          v[30] = k_madd_epi32_avx2(u[2], k32_p24_p08);
+          v[31] = k_madd_epi32_avx2(u[3], k32_p24_p08);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+          v[8] = _mm256_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+          v[9] = _mm256_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep1[36] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          lstep1[37] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          lstep1[38] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          lstep1[39] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          lstep1[40] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          lstep1[41] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          lstep1[42] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          lstep1[43] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+          lstep1[52] = _mm256_srai_epi32(v[8], DCT_CONST_BITS);
+          lstep1[53] = _mm256_srai_epi32(v[9], DCT_CONST_BITS);
+          lstep1[54] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep1[55] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep1[56] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep1[57] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep1[58] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep1[59] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 5
+        {
+          lstep2[8] = _mm256_add_epi32(lstep1[10], lstep3[8]);
+          lstep2[9] = _mm256_add_epi32(lstep1[11], lstep3[9]);
+          lstep2[10] = _mm256_sub_epi32(lstep3[8], lstep1[10]);
+          lstep2[11] = _mm256_sub_epi32(lstep3[9], lstep1[11]);
+          lstep2[12] = _mm256_sub_epi32(lstep3[14], lstep1[12]);
+          lstep2[13] = _mm256_sub_epi32(lstep3[15], lstep1[13]);
+          lstep2[14] = _mm256_add_epi32(lstep1[12], lstep3[14]);
+          lstep2[15] = _mm256_add_epi32(lstep1[13], lstep3[15]);
+        }
+        {
+          const __m256i k32_p16_p16 =
+              pair256_set_epi32(cospi_16_64, cospi_16_64);
+          const __m256i k32_p16_m16 =
+              pair256_set_epi32(cospi_16_64, -cospi_16_64);
+          const __m256i k32_p24_p08 =
+              pair256_set_epi32(cospi_24_64, cospi_8_64);
+          const __m256i k32_m08_p24 =
+              pair256_set_epi32(-cospi_8_64, cospi_24_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep1[0], lstep1[2]);
+          u[1] = _mm256_unpackhi_epi32(lstep1[0], lstep1[2]);
+          u[2] = _mm256_unpacklo_epi32(lstep1[1], lstep1[3]);
+          u[3] = _mm256_unpackhi_epi32(lstep1[1], lstep1[3]);
+          u[4] = _mm256_unpacklo_epi32(lstep1[4], lstep1[6]);
+          u[5] = _mm256_unpackhi_epi32(lstep1[4], lstep1[6]);
+          u[6] = _mm256_unpacklo_epi32(lstep1[5], lstep1[7]);
+          u[7] = _mm256_unpackhi_epi32(lstep1[5], lstep1[7]);
+
+          // TODO(jingning): manually inline k_madd_epi32_avx2_ to further hide
+          // instruction latency.
+          v[0] = k_madd_epi32_avx2(u[0], k32_p16_p16);
+          v[1] = k_madd_epi32_avx2(u[1], k32_p16_p16);
+          v[2] = k_madd_epi32_avx2(u[2], k32_p16_p16);
+          v[3] = k_madd_epi32_avx2(u[3], k32_p16_p16);
+          v[4] = k_madd_epi32_avx2(u[0], k32_p16_m16);
+          v[5] = k_madd_epi32_avx2(u[1], k32_p16_m16);
+          v[6] = k_madd_epi32_avx2(u[2], k32_p16_m16);
+          v[7] = k_madd_epi32_avx2(u[3], k32_p16_m16);
+          v[8] = k_madd_epi32_avx2(u[4], k32_p24_p08);
+          v[9] = k_madd_epi32_avx2(u[5], k32_p24_p08);
+          v[10] = k_madd_epi32_avx2(u[6], k32_p24_p08);
+          v[11] = k_madd_epi32_avx2(u[7], k32_p24_p08);
+          v[12] = k_madd_epi32_avx2(u[4], k32_m08_p24);
+          v[13] = k_madd_epi32_avx2(u[5], k32_m08_p24);
+          v[14] = k_madd_epi32_avx2(u[6], k32_m08_p24);
+          v[15] = k_madd_epi32_avx2(u[7], k32_m08_p24);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm256_cmpgt_epi32(kZero, u[0]);
+          sign[1] = _mm256_cmpgt_epi32(kZero, u[1]);
+          sign[2] = _mm256_cmpgt_epi32(kZero, u[2]);
+          sign[3] = _mm256_cmpgt_epi32(kZero, u[3]);
+          sign[4] = _mm256_cmpgt_epi32(kZero, u[4]);
+          sign[5] = _mm256_cmpgt_epi32(kZero, u[5]);
+          sign[6] = _mm256_cmpgt_epi32(kZero, u[6]);
+          sign[7] = _mm256_cmpgt_epi32(kZero, u[7]);
+
+          u[0] = _mm256_sub_epi32(u[0], sign[0]);
+          u[1] = _mm256_sub_epi32(u[1], sign[1]);
+          u[2] = _mm256_sub_epi32(u[2], sign[2]);
+          u[3] = _mm256_sub_epi32(u[3], sign[3]);
+          u[4] = _mm256_sub_epi32(u[4], sign[4]);
+          u[5] = _mm256_sub_epi32(u[5], sign[5]);
+          u[6] = _mm256_sub_epi32(u[6], sign[6]);
+          u[7] = _mm256_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm256_add_epi32(u[0], K32One);
+          u[1] = _mm256_add_epi32(u[1], K32One);
+          u[2] = _mm256_add_epi32(u[2], K32One);
+          u[3] = _mm256_add_epi32(u[3], K32One);
+          u[4] = _mm256_add_epi32(u[4], K32One);
+          u[5] = _mm256_add_epi32(u[5], K32One);
+          u[6] = _mm256_add_epi32(u[6], K32One);
+          u[7] = _mm256_add_epi32(u[7], K32One);
+
+          u[0] = _mm256_srai_epi32(u[0], 2);
+          u[1] = _mm256_srai_epi32(u[1], 2);
+          u[2] = _mm256_srai_epi32(u[2], 2);
+          u[3] = _mm256_srai_epi32(u[3], 2);
+          u[4] = _mm256_srai_epi32(u[4], 2);
+          u[5] = _mm256_srai_epi32(u[5], 2);
+          u[6] = _mm256_srai_epi32(u[6], 2);
+          u[7] = _mm256_srai_epi32(u[7], 2);
+
+          // Combine
+          out[0] = _mm256_packs_epi32(u[0], u[1]);
+          out[16] = _mm256_packs_epi32(u[2], u[3]);
+          out[8] = _mm256_packs_epi32(u[4], u[5]);
+          out[24] = _mm256_packs_epi32(u[6], u[7]);
+        }
+        {
+          const __m256i k32_m08_p24 =
+              pair256_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m256i k32_m24_m08 =
+              pair256_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m256i k32_p24_p08 =
+              pair256_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep1[18], lstep1[28]);
+          u[1] = _mm256_unpackhi_epi32(lstep1[18], lstep1[28]);
+          u[2] = _mm256_unpacklo_epi32(lstep1[19], lstep1[29]);
+          u[3] = _mm256_unpackhi_epi32(lstep1[19], lstep1[29]);
+          u[4] = _mm256_unpacklo_epi32(lstep1[20], lstep1[26]);
+          u[5] = _mm256_unpackhi_epi32(lstep1[20], lstep1[26]);
+          u[6] = _mm256_unpacklo_epi32(lstep1[21], lstep1[27]);
+          u[7] = _mm256_unpackhi_epi32(lstep1[21], lstep1[27]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_m08_p24);
+          v[1] = k_madd_epi32_avx2(u[1], k32_m08_p24);
+          v[2] = k_madd_epi32_avx2(u[2], k32_m08_p24);
+          v[3] = k_madd_epi32_avx2(u[3], k32_m08_p24);
+          v[4] = k_madd_epi32_avx2(u[4], k32_m24_m08);
+          v[5] = k_madd_epi32_avx2(u[5], k32_m24_m08);
+          v[6] = k_madd_epi32_avx2(u[6], k32_m24_m08);
+          v[7] = k_madd_epi32_avx2(u[7], k32_m24_m08);
+          v[8] = k_madd_epi32_avx2(u[4], k32_m08_p24);
+          v[9] = k_madd_epi32_avx2(u[5], k32_m08_p24);
+          v[10] = k_madd_epi32_avx2(u[6], k32_m08_p24);
+          v[11] = k_madd_epi32_avx2(u[7], k32_m08_p24);
+          v[12] = k_madd_epi32_avx2(u[0], k32_p24_p08);
+          v[13] = k_madd_epi32_avx2(u[1], k32_p24_p08);
+          v[14] = k_madd_epi32_avx2(u[2], k32_p24_p08);
+          v[15] = k_madd_epi32_avx2(u[3], k32_p24_p08);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+
+          u[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          u[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          u[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          u[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          u[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          u[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          u[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          u[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          lstep2[18] = _mm256_srai_epi32(u[0], DCT_CONST_BITS);
+          lstep2[19] = _mm256_srai_epi32(u[1], DCT_CONST_BITS);
+          lstep2[20] = _mm256_srai_epi32(u[2], DCT_CONST_BITS);
+          lstep2[21] = _mm256_srai_epi32(u[3], DCT_CONST_BITS);
+          lstep2[26] = _mm256_srai_epi32(u[4], DCT_CONST_BITS);
+          lstep2[27] = _mm256_srai_epi32(u[5], DCT_CONST_BITS);
+          lstep2[28] = _mm256_srai_epi32(u[6], DCT_CONST_BITS);
+          lstep2[29] = _mm256_srai_epi32(u[7], DCT_CONST_BITS);
+        }
+        {
+          lstep2[32] = _mm256_add_epi32(lstep1[38], lstep3[32]);
+          lstep2[33] = _mm256_add_epi32(lstep1[39], lstep3[33]);
+          lstep2[34] = _mm256_add_epi32(lstep1[36], lstep3[34]);
+          lstep2[35] = _mm256_add_epi32(lstep1[37], lstep3[35]);
+          lstep2[36] = _mm256_sub_epi32(lstep3[34], lstep1[36]);
+          lstep2[37] = _mm256_sub_epi32(lstep3[35], lstep1[37]);
+          lstep2[38] = _mm256_sub_epi32(lstep3[32], lstep1[38]);
+          lstep2[39] = _mm256_sub_epi32(lstep3[33], lstep1[39]);
+          lstep2[40] = _mm256_sub_epi32(lstep3[46], lstep1[40]);
+          lstep2[41] = _mm256_sub_epi32(lstep3[47], lstep1[41]);
+          lstep2[42] = _mm256_sub_epi32(lstep3[44], lstep1[42]);
+          lstep2[43] = _mm256_sub_epi32(lstep3[45], lstep1[43]);
+          lstep2[44] = _mm256_add_epi32(lstep1[42], lstep3[44]);
+          lstep2[45] = _mm256_add_epi32(lstep1[43], lstep3[45]);
+          lstep2[46] = _mm256_add_epi32(lstep1[40], lstep3[46]);
+          lstep2[47] = _mm256_add_epi32(lstep1[41], lstep3[47]);
+          lstep2[48] = _mm256_add_epi32(lstep1[54], lstep3[48]);
+          lstep2[49] = _mm256_add_epi32(lstep1[55], lstep3[49]);
+          lstep2[50] = _mm256_add_epi32(lstep1[52], lstep3[50]);
+          lstep2[51] = _mm256_add_epi32(lstep1[53], lstep3[51]);
+          lstep2[52] = _mm256_sub_epi32(lstep3[50], lstep1[52]);
+          lstep2[53] = _mm256_sub_epi32(lstep3[51], lstep1[53]);
+          lstep2[54] = _mm256_sub_epi32(lstep3[48], lstep1[54]);
+          lstep2[55] = _mm256_sub_epi32(lstep3[49], lstep1[55]);
+          lstep2[56] = _mm256_sub_epi32(lstep3[62], lstep1[56]);
+          lstep2[57] = _mm256_sub_epi32(lstep3[63], lstep1[57]);
+          lstep2[58] = _mm256_sub_epi32(lstep3[60], lstep1[58]);
+          lstep2[59] = _mm256_sub_epi32(lstep3[61], lstep1[59]);
+          lstep2[60] = _mm256_add_epi32(lstep1[58], lstep3[60]);
+          lstep2[61] = _mm256_add_epi32(lstep1[59], lstep3[61]);
+          lstep2[62] = _mm256_add_epi32(lstep1[56], lstep3[62]);
+          lstep2[63] = _mm256_add_epi32(lstep1[57], lstep3[63]);
+        }
+        // stage 6
+        {
+          const __m256i k32_p28_p04 =
+              pair256_set_epi32(cospi_28_64, cospi_4_64);
+          const __m256i k32_p12_p20 =
+              pair256_set_epi32(cospi_12_64, cospi_20_64);
+          const __m256i k32_m20_p12 =
+              pair256_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m256i k32_m04_p28 =
+              pair256_set_epi32(-cospi_4_64, cospi_28_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep2[8], lstep2[14]);
+          u[1] = _mm256_unpackhi_epi32(lstep2[8], lstep2[14]);
+          u[2] = _mm256_unpacklo_epi32(lstep2[9], lstep2[15]);
+          u[3] = _mm256_unpackhi_epi32(lstep2[9], lstep2[15]);
+          u[4] = _mm256_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[5] = _mm256_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[6] = _mm256_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[7] = _mm256_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[8] = _mm256_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[9] = _mm256_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[10] = _mm256_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[11] = _mm256_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[12] = _mm256_unpacklo_epi32(lstep2[8], lstep2[14]);
+          u[13] = _mm256_unpackhi_epi32(lstep2[8], lstep2[14]);
+          u[14] = _mm256_unpacklo_epi32(lstep2[9], lstep2[15]);
+          u[15] = _mm256_unpackhi_epi32(lstep2[9], lstep2[15]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_p28_p04);
+          v[1] = k_madd_epi32_avx2(u[1], k32_p28_p04);
+          v[2] = k_madd_epi32_avx2(u[2], k32_p28_p04);
+          v[3] = k_madd_epi32_avx2(u[3], k32_p28_p04);
+          v[4] = k_madd_epi32_avx2(u[4], k32_p12_p20);
+          v[5] = k_madd_epi32_avx2(u[5], k32_p12_p20);
+          v[6] = k_madd_epi32_avx2(u[6], k32_p12_p20);
+          v[7] = k_madd_epi32_avx2(u[7], k32_p12_p20);
+          v[8] = k_madd_epi32_avx2(u[8], k32_m20_p12);
+          v[9] = k_madd_epi32_avx2(u[9], k32_m20_p12);
+          v[10] = k_madd_epi32_avx2(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32_avx2(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32_avx2(u[12], k32_m04_p28);
+          v[13] = k_madd_epi32_avx2(u[13], k32_m04_p28);
+          v[14] = k_madd_epi32_avx2(u[14], k32_m04_p28);
+          v[15] = k_madd_epi32_avx2(u[15], k32_m04_p28);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm256_cmpgt_epi32(kZero, u[0]);
+          sign[1] = _mm256_cmpgt_epi32(kZero, u[1]);
+          sign[2] = _mm256_cmpgt_epi32(kZero, u[2]);
+          sign[3] = _mm256_cmpgt_epi32(kZero, u[3]);
+          sign[4] = _mm256_cmpgt_epi32(kZero, u[4]);
+          sign[5] = _mm256_cmpgt_epi32(kZero, u[5]);
+          sign[6] = _mm256_cmpgt_epi32(kZero, u[6]);
+          sign[7] = _mm256_cmpgt_epi32(kZero, u[7]);
+
+          u[0] = _mm256_sub_epi32(u[0], sign[0]);
+          u[1] = _mm256_sub_epi32(u[1], sign[1]);
+          u[2] = _mm256_sub_epi32(u[2], sign[2]);
+          u[3] = _mm256_sub_epi32(u[3], sign[3]);
+          u[4] = _mm256_sub_epi32(u[4], sign[4]);
+          u[5] = _mm256_sub_epi32(u[5], sign[5]);
+          u[6] = _mm256_sub_epi32(u[6], sign[6]);
+          u[7] = _mm256_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm256_add_epi32(u[0], K32One);
+          u[1] = _mm256_add_epi32(u[1], K32One);
+          u[2] = _mm256_add_epi32(u[2], K32One);
+          u[3] = _mm256_add_epi32(u[3], K32One);
+          u[4] = _mm256_add_epi32(u[4], K32One);
+          u[5] = _mm256_add_epi32(u[5], K32One);
+          u[6] = _mm256_add_epi32(u[6], K32One);
+          u[7] = _mm256_add_epi32(u[7], K32One);
+
+          u[0] = _mm256_srai_epi32(u[0], 2);
+          u[1] = _mm256_srai_epi32(u[1], 2);
+          u[2] = _mm256_srai_epi32(u[2], 2);
+          u[3] = _mm256_srai_epi32(u[3], 2);
+          u[4] = _mm256_srai_epi32(u[4], 2);
+          u[5] = _mm256_srai_epi32(u[5], 2);
+          u[6] = _mm256_srai_epi32(u[6], 2);
+          u[7] = _mm256_srai_epi32(u[7], 2);
+
+          out[4] = _mm256_packs_epi32(u[0], u[1]);
+          out[20] = _mm256_packs_epi32(u[2], u[3]);
+          out[12] = _mm256_packs_epi32(u[4], u[5]);
+          out[28] = _mm256_packs_epi32(u[6], u[7]);
+        }
+        {
+          lstep3[16] = _mm256_add_epi32(lstep2[18], lstep1[16]);
+          lstep3[17] = _mm256_add_epi32(lstep2[19], lstep1[17]);
+          lstep3[18] = _mm256_sub_epi32(lstep1[16], lstep2[18]);
+          lstep3[19] = _mm256_sub_epi32(lstep1[17], lstep2[19]);
+          lstep3[20] = _mm256_sub_epi32(lstep1[22], lstep2[20]);
+          lstep3[21] = _mm256_sub_epi32(lstep1[23], lstep2[21]);
+          lstep3[22] = _mm256_add_epi32(lstep2[20], lstep1[22]);
+          lstep3[23] = _mm256_add_epi32(lstep2[21], lstep1[23]);
+          lstep3[24] = _mm256_add_epi32(lstep2[26], lstep1[24]);
+          lstep3[25] = _mm256_add_epi32(lstep2[27], lstep1[25]);
+          lstep3[26] = _mm256_sub_epi32(lstep1[24], lstep2[26]);
+          lstep3[27] = _mm256_sub_epi32(lstep1[25], lstep2[27]);
+          lstep3[28] = _mm256_sub_epi32(lstep1[30], lstep2[28]);
+          lstep3[29] = _mm256_sub_epi32(lstep1[31], lstep2[29]);
+          lstep3[30] = _mm256_add_epi32(lstep2[28], lstep1[30]);
+          lstep3[31] = _mm256_add_epi32(lstep2[29], lstep1[31]);
+        }
+        {
+          const __m256i k32_m04_p28 =
+              pair256_set_epi32(-cospi_4_64, cospi_28_64);
+          const __m256i k32_m28_m04 =
+              pair256_set_epi32(-cospi_28_64, -cospi_4_64);
+          const __m256i k32_m20_p12 =
+              pair256_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m256i k32_m12_m20 =
+              pair256_set_epi32(-cospi_12_64, -cospi_20_64);
+          const __m256i k32_p12_p20 =
+              pair256_set_epi32(cospi_12_64, cospi_20_64);
+          const __m256i k32_p28_p04 =
+              pair256_set_epi32(cospi_28_64, cospi_4_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep2[34], lstep2[60]);
+          u[1] = _mm256_unpackhi_epi32(lstep2[34], lstep2[60]);
+          u[2] = _mm256_unpacklo_epi32(lstep2[35], lstep2[61]);
+          u[3] = _mm256_unpackhi_epi32(lstep2[35], lstep2[61]);
+          u[4] = _mm256_unpacklo_epi32(lstep2[36], lstep2[58]);
+          u[5] = _mm256_unpackhi_epi32(lstep2[36], lstep2[58]);
+          u[6] = _mm256_unpacklo_epi32(lstep2[37], lstep2[59]);
+          u[7] = _mm256_unpackhi_epi32(lstep2[37], lstep2[59]);
+          u[8] = _mm256_unpacklo_epi32(lstep2[42], lstep2[52]);
+          u[9] = _mm256_unpackhi_epi32(lstep2[42], lstep2[52]);
+          u[10] = _mm256_unpacklo_epi32(lstep2[43], lstep2[53]);
+          u[11] = _mm256_unpackhi_epi32(lstep2[43], lstep2[53]);
+          u[12] = _mm256_unpacklo_epi32(lstep2[44], lstep2[50]);
+          u[13] = _mm256_unpackhi_epi32(lstep2[44], lstep2[50]);
+          u[14] = _mm256_unpacklo_epi32(lstep2[45], lstep2[51]);
+          u[15] = _mm256_unpackhi_epi32(lstep2[45], lstep2[51]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_m04_p28);
+          v[1] = k_madd_epi32_avx2(u[1], k32_m04_p28);
+          v[2] = k_madd_epi32_avx2(u[2], k32_m04_p28);
+          v[3] = k_madd_epi32_avx2(u[3], k32_m04_p28);
+          v[4] = k_madd_epi32_avx2(u[4], k32_m28_m04);
+          v[5] = k_madd_epi32_avx2(u[5], k32_m28_m04);
+          v[6] = k_madd_epi32_avx2(u[6], k32_m28_m04);
+          v[7] = k_madd_epi32_avx2(u[7], k32_m28_m04);
+          v[8] = k_madd_epi32_avx2(u[8], k32_m20_p12);
+          v[9] = k_madd_epi32_avx2(u[9], k32_m20_p12);
+          v[10] = k_madd_epi32_avx2(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32_avx2(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32_avx2(u[12], k32_m12_m20);
+          v[13] = k_madd_epi32_avx2(u[13], k32_m12_m20);
+          v[14] = k_madd_epi32_avx2(u[14], k32_m12_m20);
+          v[15] = k_madd_epi32_avx2(u[15], k32_m12_m20);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m20_p12);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m20_p12);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m20_p12);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m20_p12);
+          v[20] = k_madd_epi32_avx2(u[8], k32_p12_p20);
+          v[21] = k_madd_epi32_avx2(u[9], k32_p12_p20);
+          v[22] = k_madd_epi32_avx2(u[10], k32_p12_p20);
+          v[23] = k_madd_epi32_avx2(u[11], k32_p12_p20);
+          v[24] = k_madd_epi32_avx2(u[4], k32_m04_p28);
+          v[25] = k_madd_epi32_avx2(u[5], k32_m04_p28);
+          v[26] = k_madd_epi32_avx2(u[6], k32_m04_p28);
+          v[27] = k_madd_epi32_avx2(u[7], k32_m04_p28);
+          v[28] = k_madd_epi32_avx2(u[0], k32_p28_p04);
+          v[29] = k_madd_epi32_avx2(u[1], k32_p28_p04);
+          v[30] = k_madd_epi32_avx2(u[2], k32_p28_p04);
+          v[31] = k_madd_epi32_avx2(u[3], k32_p28_p04);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+          v[8] = _mm256_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+          v[9] = _mm256_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep3[34] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          lstep3[35] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          lstep3[36] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          lstep3[37] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          lstep3[42] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          lstep3[43] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          lstep3[44] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          lstep3[45] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+          lstep3[50] = _mm256_srai_epi32(v[8], DCT_CONST_BITS);
+          lstep3[51] = _mm256_srai_epi32(v[9], DCT_CONST_BITS);
+          lstep3[52] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep3[53] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep3[58] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep3[59] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep3[60] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep3[61] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 7
+        {
+          const __m256i k32_p30_p02 =
+              pair256_set_epi32(cospi_30_64, cospi_2_64);
+          const __m256i k32_p14_p18 =
+              pair256_set_epi32(cospi_14_64, cospi_18_64);
+          const __m256i k32_p22_p10 =
+              pair256_set_epi32(cospi_22_64, cospi_10_64);
+          const __m256i k32_p06_p26 =
+              pair256_set_epi32(cospi_6_64, cospi_26_64);
+          const __m256i k32_m26_p06 =
+              pair256_set_epi32(-cospi_26_64, cospi_6_64);
+          const __m256i k32_m10_p22 =
+              pair256_set_epi32(-cospi_10_64, cospi_22_64);
+          const __m256i k32_m18_p14 =
+              pair256_set_epi32(-cospi_18_64, cospi_14_64);
+          const __m256i k32_m02_p30 =
+              pair256_set_epi32(-cospi_2_64, cospi_30_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep3[16], lstep3[30]);
+          u[1] = _mm256_unpackhi_epi32(lstep3[16], lstep3[30]);
+          u[2] = _mm256_unpacklo_epi32(lstep3[17], lstep3[31]);
+          u[3] = _mm256_unpackhi_epi32(lstep3[17], lstep3[31]);
+          u[4] = _mm256_unpacklo_epi32(lstep3[18], lstep3[28]);
+          u[5] = _mm256_unpackhi_epi32(lstep3[18], lstep3[28]);
+          u[6] = _mm256_unpacklo_epi32(lstep3[19], lstep3[29]);
+          u[7] = _mm256_unpackhi_epi32(lstep3[19], lstep3[29]);
+          u[8] = _mm256_unpacklo_epi32(lstep3[20], lstep3[26]);
+          u[9] = _mm256_unpackhi_epi32(lstep3[20], lstep3[26]);
+          u[10] = _mm256_unpacklo_epi32(lstep3[21], lstep3[27]);
+          u[11] = _mm256_unpackhi_epi32(lstep3[21], lstep3[27]);
+          u[12] = _mm256_unpacklo_epi32(lstep3[22], lstep3[24]);
+          u[13] = _mm256_unpackhi_epi32(lstep3[22], lstep3[24]);
+          u[14] = _mm256_unpacklo_epi32(lstep3[23], lstep3[25]);
+          u[15] = _mm256_unpackhi_epi32(lstep3[23], lstep3[25]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_p30_p02);
+          v[1] = k_madd_epi32_avx2(u[1], k32_p30_p02);
+          v[2] = k_madd_epi32_avx2(u[2], k32_p30_p02);
+          v[3] = k_madd_epi32_avx2(u[3], k32_p30_p02);
+          v[4] = k_madd_epi32_avx2(u[4], k32_p14_p18);
+          v[5] = k_madd_epi32_avx2(u[5], k32_p14_p18);
+          v[6] = k_madd_epi32_avx2(u[6], k32_p14_p18);
+          v[7] = k_madd_epi32_avx2(u[7], k32_p14_p18);
+          v[8] = k_madd_epi32_avx2(u[8], k32_p22_p10);
+          v[9] = k_madd_epi32_avx2(u[9], k32_p22_p10);
+          v[10] = k_madd_epi32_avx2(u[10], k32_p22_p10);
+          v[11] = k_madd_epi32_avx2(u[11], k32_p22_p10);
+          v[12] = k_madd_epi32_avx2(u[12], k32_p06_p26);
+          v[13] = k_madd_epi32_avx2(u[13], k32_p06_p26);
+          v[14] = k_madd_epi32_avx2(u[14], k32_p06_p26);
+          v[15] = k_madd_epi32_avx2(u[15], k32_p06_p26);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m26_p06);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m26_p06);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m26_p06);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m26_p06);
+          v[20] = k_madd_epi32_avx2(u[8], k32_m10_p22);
+          v[21] = k_madd_epi32_avx2(u[9], k32_m10_p22);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m10_p22);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m10_p22);
+          v[24] = k_madd_epi32_avx2(u[4], k32_m18_p14);
+          v[25] = k_madd_epi32_avx2(u[5], k32_m18_p14);
+          v[26] = k_madd_epi32_avx2(u[6], k32_m18_p14);
+          v[27] = k_madd_epi32_avx2(u[7], k32_m18_p14);
+          v[28] = k_madd_epi32_avx2(u[0], k32_m02_p30);
+          v[29] = k_madd_epi32_avx2(u[1], k32_m02_p30);
+          v[30] = k_madd_epi32_avx2(u[2], k32_m02_p30);
+          v[31] = k_madd_epi32_avx2(u[3], k32_m02_p30);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+          v[8] = _mm256_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+          v[9] = _mm256_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+          u[8] = _mm256_srai_epi32(v[8], DCT_CONST_BITS);
+          u[9] = _mm256_srai_epi32(v[9], DCT_CONST_BITS);
+          u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[0] = _mm256_cmpgt_epi32(kZero, u[0]);
+          v[1] = _mm256_cmpgt_epi32(kZero, u[1]);
+          v[2] = _mm256_cmpgt_epi32(kZero, u[2]);
+          v[3] = _mm256_cmpgt_epi32(kZero, u[3]);
+          v[4] = _mm256_cmpgt_epi32(kZero, u[4]);
+          v[5] = _mm256_cmpgt_epi32(kZero, u[5]);
+          v[6] = _mm256_cmpgt_epi32(kZero, u[6]);
+          v[7] = _mm256_cmpgt_epi32(kZero, u[7]);
+          v[8] = _mm256_cmpgt_epi32(kZero, u[8]);
+          v[9] = _mm256_cmpgt_epi32(kZero, u[9]);
+          v[10] = _mm256_cmpgt_epi32(kZero, u[10]);
+          v[11] = _mm256_cmpgt_epi32(kZero, u[11]);
+          v[12] = _mm256_cmpgt_epi32(kZero, u[12]);
+          v[13] = _mm256_cmpgt_epi32(kZero, u[13]);
+          v[14] = _mm256_cmpgt_epi32(kZero, u[14]);
+          v[15] = _mm256_cmpgt_epi32(kZero, u[15]);
+
+          u[0] = _mm256_sub_epi32(u[0], v[0]);
+          u[1] = _mm256_sub_epi32(u[1], v[1]);
+          u[2] = _mm256_sub_epi32(u[2], v[2]);
+          u[3] = _mm256_sub_epi32(u[3], v[3]);
+          u[4] = _mm256_sub_epi32(u[4], v[4]);
+          u[5] = _mm256_sub_epi32(u[5], v[5]);
+          u[6] = _mm256_sub_epi32(u[6], v[6]);
+          u[7] = _mm256_sub_epi32(u[7], v[7]);
+          u[8] = _mm256_sub_epi32(u[8], v[8]);
+          u[9] = _mm256_sub_epi32(u[9], v[9]);
+          u[10] = _mm256_sub_epi32(u[10], v[10]);
+          u[11] = _mm256_sub_epi32(u[11], v[11]);
+          u[12] = _mm256_sub_epi32(u[12], v[12]);
+          u[13] = _mm256_sub_epi32(u[13], v[13]);
+          u[14] = _mm256_sub_epi32(u[14], v[14]);
+          u[15] = _mm256_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], K32One);
+          v[1] = _mm256_add_epi32(u[1], K32One);
+          v[2] = _mm256_add_epi32(u[2], K32One);
+          v[3] = _mm256_add_epi32(u[3], K32One);
+          v[4] = _mm256_add_epi32(u[4], K32One);
+          v[5] = _mm256_add_epi32(u[5], K32One);
+          v[6] = _mm256_add_epi32(u[6], K32One);
+          v[7] = _mm256_add_epi32(u[7], K32One);
+          v[8] = _mm256_add_epi32(u[8], K32One);
+          v[9] = _mm256_add_epi32(u[9], K32One);
+          v[10] = _mm256_add_epi32(u[10], K32One);
+          v[11] = _mm256_add_epi32(u[11], K32One);
+          v[12] = _mm256_add_epi32(u[12], K32One);
+          v[13] = _mm256_add_epi32(u[13], K32One);
+          v[14] = _mm256_add_epi32(u[14], K32One);
+          v[15] = _mm256_add_epi32(u[15], K32One);
+
+          u[0] = _mm256_srai_epi32(v[0], 2);
+          u[1] = _mm256_srai_epi32(v[1], 2);
+          u[2] = _mm256_srai_epi32(v[2], 2);
+          u[3] = _mm256_srai_epi32(v[3], 2);
+          u[4] = _mm256_srai_epi32(v[4], 2);
+          u[5] = _mm256_srai_epi32(v[5], 2);
+          u[6] = _mm256_srai_epi32(v[6], 2);
+          u[7] = _mm256_srai_epi32(v[7], 2);
+          u[8] = _mm256_srai_epi32(v[8], 2);
+          u[9] = _mm256_srai_epi32(v[9], 2);
+          u[10] = _mm256_srai_epi32(v[10], 2);
+          u[11] = _mm256_srai_epi32(v[11], 2);
+          u[12] = _mm256_srai_epi32(v[12], 2);
+          u[13] = _mm256_srai_epi32(v[13], 2);
+          u[14] = _mm256_srai_epi32(v[14], 2);
+          u[15] = _mm256_srai_epi32(v[15], 2);
+
+          out[2] = _mm256_packs_epi32(u[0], u[1]);
+          out[18] = _mm256_packs_epi32(u[2], u[3]);
+          out[10] = _mm256_packs_epi32(u[4], u[5]);
+          out[26] = _mm256_packs_epi32(u[6], u[7]);
+          out[6] = _mm256_packs_epi32(u[8], u[9]);
+          out[22] = _mm256_packs_epi32(u[10], u[11]);
+          out[14] = _mm256_packs_epi32(u[12], u[13]);
+          out[30] = _mm256_packs_epi32(u[14], u[15]);
+        }
+        {
+          lstep1[32] = _mm256_add_epi32(lstep3[34], lstep2[32]);
+          lstep1[33] = _mm256_add_epi32(lstep3[35], lstep2[33]);
+          lstep1[34] = _mm256_sub_epi32(lstep2[32], lstep3[34]);
+          lstep1[35] = _mm256_sub_epi32(lstep2[33], lstep3[35]);
+          lstep1[36] = _mm256_sub_epi32(lstep2[38], lstep3[36]);
+          lstep1[37] = _mm256_sub_epi32(lstep2[39], lstep3[37]);
+          lstep1[38] = _mm256_add_epi32(lstep3[36], lstep2[38]);
+          lstep1[39] = _mm256_add_epi32(lstep3[37], lstep2[39]);
+          lstep1[40] = _mm256_add_epi32(lstep3[42], lstep2[40]);
+          lstep1[41] = _mm256_add_epi32(lstep3[43], lstep2[41]);
+          lstep1[42] = _mm256_sub_epi32(lstep2[40], lstep3[42]);
+          lstep1[43] = _mm256_sub_epi32(lstep2[41], lstep3[43]);
+          lstep1[44] = _mm256_sub_epi32(lstep2[46], lstep3[44]);
+          lstep1[45] = _mm256_sub_epi32(lstep2[47], lstep3[45]);
+          lstep1[46] = _mm256_add_epi32(lstep3[44], lstep2[46]);
+          lstep1[47] = _mm256_add_epi32(lstep3[45], lstep2[47]);
+          lstep1[48] = _mm256_add_epi32(lstep3[50], lstep2[48]);
+          lstep1[49] = _mm256_add_epi32(lstep3[51], lstep2[49]);
+          lstep1[50] = _mm256_sub_epi32(lstep2[48], lstep3[50]);
+          lstep1[51] = _mm256_sub_epi32(lstep2[49], lstep3[51]);
+          lstep1[52] = _mm256_sub_epi32(lstep2[54], lstep3[52]);
+          lstep1[53] = _mm256_sub_epi32(lstep2[55], lstep3[53]);
+          lstep1[54] = _mm256_add_epi32(lstep3[52], lstep2[54]);
+          lstep1[55] = _mm256_add_epi32(lstep3[53], lstep2[55]);
+          lstep1[56] = _mm256_add_epi32(lstep3[58], lstep2[56]);
+          lstep1[57] = _mm256_add_epi32(lstep3[59], lstep2[57]);
+          lstep1[58] = _mm256_sub_epi32(lstep2[56], lstep3[58]);
+          lstep1[59] = _mm256_sub_epi32(lstep2[57], lstep3[59]);
+          lstep1[60] = _mm256_sub_epi32(lstep2[62], lstep3[60]);
+          lstep1[61] = _mm256_sub_epi32(lstep2[63], lstep3[61]);
+          lstep1[62] = _mm256_add_epi32(lstep3[60], lstep2[62]);
+          lstep1[63] = _mm256_add_epi32(lstep3[61], lstep2[63]);
+        }
+        // stage 8
+        {
+          const __m256i k32_p31_p01 =
+              pair256_set_epi32(cospi_31_64, cospi_1_64);
+          const __m256i k32_p15_p17 =
+              pair256_set_epi32(cospi_15_64, cospi_17_64);
+          const __m256i k32_p23_p09 =
+              pair256_set_epi32(cospi_23_64, cospi_9_64);
+          const __m256i k32_p07_p25 =
+              pair256_set_epi32(cospi_7_64, cospi_25_64);
+          const __m256i k32_m25_p07 =
+              pair256_set_epi32(-cospi_25_64, cospi_7_64);
+          const __m256i k32_m09_p23 =
+              pair256_set_epi32(-cospi_9_64, cospi_23_64);
+          const __m256i k32_m17_p15 =
+              pair256_set_epi32(-cospi_17_64, cospi_15_64);
+          const __m256i k32_m01_p31 =
+              pair256_set_epi32(-cospi_1_64, cospi_31_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep1[32], lstep1[62]);
+          u[1] = _mm256_unpackhi_epi32(lstep1[32], lstep1[62]);
+          u[2] = _mm256_unpacklo_epi32(lstep1[33], lstep1[63]);
+          u[3] = _mm256_unpackhi_epi32(lstep1[33], lstep1[63]);
+          u[4] = _mm256_unpacklo_epi32(lstep1[34], lstep1[60]);
+          u[5] = _mm256_unpackhi_epi32(lstep1[34], lstep1[60]);
+          u[6] = _mm256_unpacklo_epi32(lstep1[35], lstep1[61]);
+          u[7] = _mm256_unpackhi_epi32(lstep1[35], lstep1[61]);
+          u[8] = _mm256_unpacklo_epi32(lstep1[36], lstep1[58]);
+          u[9] = _mm256_unpackhi_epi32(lstep1[36], lstep1[58]);
+          u[10] = _mm256_unpacklo_epi32(lstep1[37], lstep1[59]);
+          u[11] = _mm256_unpackhi_epi32(lstep1[37], lstep1[59]);
+          u[12] = _mm256_unpacklo_epi32(lstep1[38], lstep1[56]);
+          u[13] = _mm256_unpackhi_epi32(lstep1[38], lstep1[56]);
+          u[14] = _mm256_unpacklo_epi32(lstep1[39], lstep1[57]);
+          u[15] = _mm256_unpackhi_epi32(lstep1[39], lstep1[57]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_p31_p01);
+          v[1] = k_madd_epi32_avx2(u[1], k32_p31_p01);
+          v[2] = k_madd_epi32_avx2(u[2], k32_p31_p01);
+          v[3] = k_madd_epi32_avx2(u[3], k32_p31_p01);
+          v[4] = k_madd_epi32_avx2(u[4], k32_p15_p17);
+          v[5] = k_madd_epi32_avx2(u[5], k32_p15_p17);
+          v[6] = k_madd_epi32_avx2(u[6], k32_p15_p17);
+          v[7] = k_madd_epi32_avx2(u[7], k32_p15_p17);
+          v[8] = k_madd_epi32_avx2(u[8], k32_p23_p09);
+          v[9] = k_madd_epi32_avx2(u[9], k32_p23_p09);
+          v[10] = k_madd_epi32_avx2(u[10], k32_p23_p09);
+          v[11] = k_madd_epi32_avx2(u[11], k32_p23_p09);
+          v[12] = k_madd_epi32_avx2(u[12], k32_p07_p25);
+          v[13] = k_madd_epi32_avx2(u[13], k32_p07_p25);
+          v[14] = k_madd_epi32_avx2(u[14], k32_p07_p25);
+          v[15] = k_madd_epi32_avx2(u[15], k32_p07_p25);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m25_p07);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m25_p07);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m25_p07);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m25_p07);
+          v[20] = k_madd_epi32_avx2(u[8], k32_m09_p23);
+          v[21] = k_madd_epi32_avx2(u[9], k32_m09_p23);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m09_p23);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m09_p23);
+          v[24] = k_madd_epi32_avx2(u[4], k32_m17_p15);
+          v[25] = k_madd_epi32_avx2(u[5], k32_m17_p15);
+          v[26] = k_madd_epi32_avx2(u[6], k32_m17_p15);
+          v[27] = k_madd_epi32_avx2(u[7], k32_m17_p15);
+          v[28] = k_madd_epi32_avx2(u[0], k32_m01_p31);
+          v[29] = k_madd_epi32_avx2(u[1], k32_m01_p31);
+          v[30] = k_madd_epi32_avx2(u[2], k32_m01_p31);
+          v[31] = k_madd_epi32_avx2(u[3], k32_m01_p31);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+          v[8] = _mm256_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+          v[9] = _mm256_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+          u[8] = _mm256_srai_epi32(v[8], DCT_CONST_BITS);
+          u[9] = _mm256_srai_epi32(v[9], DCT_CONST_BITS);
+          u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[0] = _mm256_cmpgt_epi32(kZero, u[0]);
+          v[1] = _mm256_cmpgt_epi32(kZero, u[1]);
+          v[2] = _mm256_cmpgt_epi32(kZero, u[2]);
+          v[3] = _mm256_cmpgt_epi32(kZero, u[3]);
+          v[4] = _mm256_cmpgt_epi32(kZero, u[4]);
+          v[5] = _mm256_cmpgt_epi32(kZero, u[5]);
+          v[6] = _mm256_cmpgt_epi32(kZero, u[6]);
+          v[7] = _mm256_cmpgt_epi32(kZero, u[7]);
+          v[8] = _mm256_cmpgt_epi32(kZero, u[8]);
+          v[9] = _mm256_cmpgt_epi32(kZero, u[9]);
+          v[10] = _mm256_cmpgt_epi32(kZero, u[10]);
+          v[11] = _mm256_cmpgt_epi32(kZero, u[11]);
+          v[12] = _mm256_cmpgt_epi32(kZero, u[12]);
+          v[13] = _mm256_cmpgt_epi32(kZero, u[13]);
+          v[14] = _mm256_cmpgt_epi32(kZero, u[14]);
+          v[15] = _mm256_cmpgt_epi32(kZero, u[15]);
+
+          u[0] = _mm256_sub_epi32(u[0], v[0]);
+          u[1] = _mm256_sub_epi32(u[1], v[1]);
+          u[2] = _mm256_sub_epi32(u[2], v[2]);
+          u[3] = _mm256_sub_epi32(u[3], v[3]);
+          u[4] = _mm256_sub_epi32(u[4], v[4]);
+          u[5] = _mm256_sub_epi32(u[5], v[5]);
+          u[6] = _mm256_sub_epi32(u[6], v[6]);
+          u[7] = _mm256_sub_epi32(u[7], v[7]);
+          u[8] = _mm256_sub_epi32(u[8], v[8]);
+          u[9] = _mm256_sub_epi32(u[9], v[9]);
+          u[10] = _mm256_sub_epi32(u[10], v[10]);
+          u[11] = _mm256_sub_epi32(u[11], v[11]);
+          u[12] = _mm256_sub_epi32(u[12], v[12]);
+          u[13] = _mm256_sub_epi32(u[13], v[13]);
+          u[14] = _mm256_sub_epi32(u[14], v[14]);
+          u[15] = _mm256_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], K32One);
+          v[1] = _mm256_add_epi32(u[1], K32One);
+          v[2] = _mm256_add_epi32(u[2], K32One);
+          v[3] = _mm256_add_epi32(u[3], K32One);
+          v[4] = _mm256_add_epi32(u[4], K32One);
+          v[5] = _mm256_add_epi32(u[5], K32One);
+          v[6] = _mm256_add_epi32(u[6], K32One);
+          v[7] = _mm256_add_epi32(u[7], K32One);
+          v[8] = _mm256_add_epi32(u[8], K32One);
+          v[9] = _mm256_add_epi32(u[9], K32One);
+          v[10] = _mm256_add_epi32(u[10], K32One);
+          v[11] = _mm256_add_epi32(u[11], K32One);
+          v[12] = _mm256_add_epi32(u[12], K32One);
+          v[13] = _mm256_add_epi32(u[13], K32One);
+          v[14] = _mm256_add_epi32(u[14], K32One);
+          v[15] = _mm256_add_epi32(u[15], K32One);
+
+          u[0] = _mm256_srai_epi32(v[0], 2);
+          u[1] = _mm256_srai_epi32(v[1], 2);
+          u[2] = _mm256_srai_epi32(v[2], 2);
+          u[3] = _mm256_srai_epi32(v[3], 2);
+          u[4] = _mm256_srai_epi32(v[4], 2);
+          u[5] = _mm256_srai_epi32(v[5], 2);
+          u[6] = _mm256_srai_epi32(v[6], 2);
+          u[7] = _mm256_srai_epi32(v[7], 2);
+          u[8] = _mm256_srai_epi32(v[8], 2);
+          u[9] = _mm256_srai_epi32(v[9], 2);
+          u[10] = _mm256_srai_epi32(v[10], 2);
+          u[11] = _mm256_srai_epi32(v[11], 2);
+          u[12] = _mm256_srai_epi32(v[12], 2);
+          u[13] = _mm256_srai_epi32(v[13], 2);
+          u[14] = _mm256_srai_epi32(v[14], 2);
+          u[15] = _mm256_srai_epi32(v[15], 2);
+
+          out[1] = _mm256_packs_epi32(u[0], u[1]);
+          out[17] = _mm256_packs_epi32(u[2], u[3]);
+          out[9] = _mm256_packs_epi32(u[4], u[5]);
+          out[25] = _mm256_packs_epi32(u[6], u[7]);
+          out[7] = _mm256_packs_epi32(u[8], u[9]);
+          out[23] = _mm256_packs_epi32(u[10], u[11]);
+          out[15] = _mm256_packs_epi32(u[12], u[13]);
+          out[31] = _mm256_packs_epi32(u[14], u[15]);
+        }
+        {
+          const __m256i k32_p27_p05 =
+              pair256_set_epi32(cospi_27_64, cospi_5_64);
+          const __m256i k32_p11_p21 =
+              pair256_set_epi32(cospi_11_64, cospi_21_64);
+          const __m256i k32_p19_p13 =
+              pair256_set_epi32(cospi_19_64, cospi_13_64);
+          const __m256i k32_p03_p29 =
+              pair256_set_epi32(cospi_3_64, cospi_29_64);
+          const __m256i k32_m29_p03 =
+              pair256_set_epi32(-cospi_29_64, cospi_3_64);
+          const __m256i k32_m13_p19 =
+              pair256_set_epi32(-cospi_13_64, cospi_19_64);
+          const __m256i k32_m21_p11 =
+              pair256_set_epi32(-cospi_21_64, cospi_11_64);
+          const __m256i k32_m05_p27 =
+              pair256_set_epi32(-cospi_5_64, cospi_27_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep1[40], lstep1[54]);
+          u[1] = _mm256_unpackhi_epi32(lstep1[40], lstep1[54]);
+          u[2] = _mm256_unpacklo_epi32(lstep1[41], lstep1[55]);
+          u[3] = _mm256_unpackhi_epi32(lstep1[41], lstep1[55]);
+          u[4] = _mm256_unpacklo_epi32(lstep1[42], lstep1[52]);
+          u[5] = _mm256_unpackhi_epi32(lstep1[42], lstep1[52]);
+          u[6] = _mm256_unpacklo_epi32(lstep1[43], lstep1[53]);
+          u[7] = _mm256_unpackhi_epi32(lstep1[43], lstep1[53]);
+          u[8] = _mm256_unpacklo_epi32(lstep1[44], lstep1[50]);
+          u[9] = _mm256_unpackhi_epi32(lstep1[44], lstep1[50]);
+          u[10] = _mm256_unpacklo_epi32(lstep1[45], lstep1[51]);
+          u[11] = _mm256_unpackhi_epi32(lstep1[45], lstep1[51]);
+          u[12] = _mm256_unpacklo_epi32(lstep1[46], lstep1[48]);
+          u[13] = _mm256_unpackhi_epi32(lstep1[46], lstep1[48]);
+          u[14] = _mm256_unpacklo_epi32(lstep1[47], lstep1[49]);
+          u[15] = _mm256_unpackhi_epi32(lstep1[47], lstep1[49]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_p27_p05);
+          v[1] = k_madd_epi32_avx2(u[1], k32_p27_p05);
+          v[2] = k_madd_epi32_avx2(u[2], k32_p27_p05);
+          v[3] = k_madd_epi32_avx2(u[3], k32_p27_p05);
+          v[4] = k_madd_epi32_avx2(u[4], k32_p11_p21);
+          v[5] = k_madd_epi32_avx2(u[5], k32_p11_p21);
+          v[6] = k_madd_epi32_avx2(u[6], k32_p11_p21);
+          v[7] = k_madd_epi32_avx2(u[7], k32_p11_p21);
+          v[8] = k_madd_epi32_avx2(u[8], k32_p19_p13);
+          v[9] = k_madd_epi32_avx2(u[9], k32_p19_p13);
+          v[10] = k_madd_epi32_avx2(u[10], k32_p19_p13);
+          v[11] = k_madd_epi32_avx2(u[11], k32_p19_p13);
+          v[12] = k_madd_epi32_avx2(u[12], k32_p03_p29);
+          v[13] = k_madd_epi32_avx2(u[13], k32_p03_p29);
+          v[14] = k_madd_epi32_avx2(u[14], k32_p03_p29);
+          v[15] = k_madd_epi32_avx2(u[15], k32_p03_p29);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m29_p03);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m29_p03);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m29_p03);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m29_p03);
+          v[20] = k_madd_epi32_avx2(u[8], k32_m13_p19);
+          v[21] = k_madd_epi32_avx2(u[9], k32_m13_p19);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m13_p19);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m13_p19);
+          v[24] = k_madd_epi32_avx2(u[4], k32_m21_p11);
+          v[25] = k_madd_epi32_avx2(u[5], k32_m21_p11);
+          v[26] = k_madd_epi32_avx2(u[6], k32_m21_p11);
+          v[27] = k_madd_epi32_avx2(u[7], k32_m21_p11);
+          v[28] = k_madd_epi32_avx2(u[0], k32_m05_p27);
+          v[29] = k_madd_epi32_avx2(u[1], k32_m05_p27);
+          v[30] = k_madd_epi32_avx2(u[2], k32_m05_p27);
+          v[31] = k_madd_epi32_avx2(u[3], k32_m05_p27);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+          v[8] = _mm256_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+          v[9] = _mm256_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+          u[8] = _mm256_srai_epi32(v[8], DCT_CONST_BITS);
+          u[9] = _mm256_srai_epi32(v[9], DCT_CONST_BITS);
+          u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[0] = _mm256_cmpgt_epi32(kZero, u[0]);
+          v[1] = _mm256_cmpgt_epi32(kZero, u[1]);
+          v[2] = _mm256_cmpgt_epi32(kZero, u[2]);
+          v[3] = _mm256_cmpgt_epi32(kZero, u[3]);
+          v[4] = _mm256_cmpgt_epi32(kZero, u[4]);
+          v[5] = _mm256_cmpgt_epi32(kZero, u[5]);
+          v[6] = _mm256_cmpgt_epi32(kZero, u[6]);
+          v[7] = _mm256_cmpgt_epi32(kZero, u[7]);
+          v[8] = _mm256_cmpgt_epi32(kZero, u[8]);
+          v[9] = _mm256_cmpgt_epi32(kZero, u[9]);
+          v[10] = _mm256_cmpgt_epi32(kZero, u[10]);
+          v[11] = _mm256_cmpgt_epi32(kZero, u[11]);
+          v[12] = _mm256_cmpgt_epi32(kZero, u[12]);
+          v[13] = _mm256_cmpgt_epi32(kZero, u[13]);
+          v[14] = _mm256_cmpgt_epi32(kZero, u[14]);
+          v[15] = _mm256_cmpgt_epi32(kZero, u[15]);
+
+          u[0] = _mm256_sub_epi32(u[0], v[0]);
+          u[1] = _mm256_sub_epi32(u[1], v[1]);
+          u[2] = _mm256_sub_epi32(u[2], v[2]);
+          u[3] = _mm256_sub_epi32(u[3], v[3]);
+          u[4] = _mm256_sub_epi32(u[4], v[4]);
+          u[5] = _mm256_sub_epi32(u[5], v[5]);
+          u[6] = _mm256_sub_epi32(u[6], v[6]);
+          u[7] = _mm256_sub_epi32(u[7], v[7]);
+          u[8] = _mm256_sub_epi32(u[8], v[8]);
+          u[9] = _mm256_sub_epi32(u[9], v[9]);
+          u[10] = _mm256_sub_epi32(u[10], v[10]);
+          u[11] = _mm256_sub_epi32(u[11], v[11]);
+          u[12] = _mm256_sub_epi32(u[12], v[12]);
+          u[13] = _mm256_sub_epi32(u[13], v[13]);
+          u[14] = _mm256_sub_epi32(u[14], v[14]);
+          u[15] = _mm256_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], K32One);
+          v[1] = _mm256_add_epi32(u[1], K32One);
+          v[2] = _mm256_add_epi32(u[2], K32One);
+          v[3] = _mm256_add_epi32(u[3], K32One);
+          v[4] = _mm256_add_epi32(u[4], K32One);
+          v[5] = _mm256_add_epi32(u[5], K32One);
+          v[6] = _mm256_add_epi32(u[6], K32One);
+          v[7] = _mm256_add_epi32(u[7], K32One);
+          v[8] = _mm256_add_epi32(u[8], K32One);
+          v[9] = _mm256_add_epi32(u[9], K32One);
+          v[10] = _mm256_add_epi32(u[10], K32One);
+          v[11] = _mm256_add_epi32(u[11], K32One);
+          v[12] = _mm256_add_epi32(u[12], K32One);
+          v[13] = _mm256_add_epi32(u[13], K32One);
+          v[14] = _mm256_add_epi32(u[14], K32One);
+          v[15] = _mm256_add_epi32(u[15], K32One);
+
+          u[0] = _mm256_srai_epi32(v[0], 2);
+          u[1] = _mm256_srai_epi32(v[1], 2);
+          u[2] = _mm256_srai_epi32(v[2], 2);
+          u[3] = _mm256_srai_epi32(v[3], 2);
+          u[4] = _mm256_srai_epi32(v[4], 2);
+          u[5] = _mm256_srai_epi32(v[5], 2);
+          u[6] = _mm256_srai_epi32(v[6], 2);
+          u[7] = _mm256_srai_epi32(v[7], 2);
+          u[8] = _mm256_srai_epi32(v[8], 2);
+          u[9] = _mm256_srai_epi32(v[9], 2);
+          u[10] = _mm256_srai_epi32(v[10], 2);
+          u[11] = _mm256_srai_epi32(v[11], 2);
+          u[12] = _mm256_srai_epi32(v[12], 2);
+          u[13] = _mm256_srai_epi32(v[13], 2);
+          u[14] = _mm256_srai_epi32(v[14], 2);
+          u[15] = _mm256_srai_epi32(v[15], 2);
+
+          out[5] = _mm256_packs_epi32(u[0], u[1]);
+          out[21] = _mm256_packs_epi32(u[2], u[3]);
+          out[13] = _mm256_packs_epi32(u[4], u[5]);
+          out[29] = _mm256_packs_epi32(u[6], u[7]);
+          out[3] = _mm256_packs_epi32(u[8], u[9]);
+          out[19] = _mm256_packs_epi32(u[10], u[11]);
+          out[11] = _mm256_packs_epi32(u[12], u[13]);
+          out[27] = _mm256_packs_epi32(u[14], u[15]);
+        }
+      }
+#endif
+      // Transpose the results, do it as four 8x8 transposes.
+      {
+        int transpose_block;
+        int16_t *output_currStep, *output_nextStep;
+        tran_low_t *curr_out, *next_out;
+        // Pass 0
+        output_currStep = &intermediate[column_start * 32];
+        output_nextStep = &intermediate[(column_start + 8) * 32];
+        // Pass 1
+        curr_out = &output_org[column_start * 32];
+        next_out = &output_org[(column_start + 8) * 32];
+
+        for (transpose_block = 0; transpose_block < 4; ++transpose_block) {
+          __m256i *this_out = &out[8 * transpose_block];
+          // 00  01  02  03  04  05  06  07  08  09  10  11  12  13  14  15
+          // 20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35
+          // 40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55
+          // 60  61  62  63  64  65  66  67  68  69  70  71  72  73  74  75
+          // 80  81  82  83  84  85  86  87  88  89  90  91  92  93  94  95
+          // 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115
+          // 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
+          // 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155
+          const __m256i tr0_0 = _mm256_unpacklo_epi16(this_out[0], this_out[1]);
+          const __m256i tr0_1 = _mm256_unpacklo_epi16(this_out[2], this_out[3]);
+          const __m256i tr0_2 = _mm256_unpackhi_epi16(this_out[0], this_out[1]);
+          const __m256i tr0_3 = _mm256_unpackhi_epi16(this_out[2], this_out[3]);
+          const __m256i tr0_4 = _mm256_unpacklo_epi16(this_out[4], this_out[5]);
+          const __m256i tr0_5 = _mm256_unpacklo_epi16(this_out[6], this_out[7]);
+          const __m256i tr0_6 = _mm256_unpackhi_epi16(this_out[4], this_out[5]);
+          const __m256i tr0_7 = _mm256_unpackhi_epi16(this_out[6], this_out[7]);
+          // 00  20  01  21  02  22  03  23  08  28  09  29  10  30  11  31
+          // 40  60  41  61  42  62  43  63  48  68  49  69  50  70  51  71
+          // 04  24  05  25  06  26  07  27  12  32  13  33  14  34  15  35
+          // 44  64  45  65  46  66  47  67  52  72  53  73  54  74  55  75
+          // 80  100 81  101 82  102 83  103 88  108 89  109 90  110 91  101
+          // 120 140 121 141 122 142 123 143 128 148 129 149 130 150 131 151
+          // 84  104 85  105 86  106 87  107 92  112 93  113 94  114 95  115
+          // 124 144 125 145 126 146 127 147 132 152 133 153 134 154 135 155
+
+          const __m256i tr1_0 = _mm256_unpacklo_epi32(tr0_0, tr0_1);
+          const __m256i tr1_1 = _mm256_unpacklo_epi32(tr0_2, tr0_3);
+          const __m256i tr1_2 = _mm256_unpackhi_epi32(tr0_0, tr0_1);
+          const __m256i tr1_3 = _mm256_unpackhi_epi32(tr0_2, tr0_3);
+          const __m256i tr1_4 = _mm256_unpacklo_epi32(tr0_4, tr0_5);
+          const __m256i tr1_5 = _mm256_unpacklo_epi32(tr0_6, tr0_7);
+          const __m256i tr1_6 = _mm256_unpackhi_epi32(tr0_4, tr0_5);
+          const __m256i tr1_7 = _mm256_unpackhi_epi32(tr0_6, tr0_7);
+          // 00 20  40  60  01 21  41  61  08 28  48  68  09 29  49  69
+          // 04 24  44  64  05 25  45  65  12 32  52  72  13 33  53  73
+          // 02 22  42  62  03 23  43  63  10 30  50  70  11 31  51  71
+          // 06 26  46  66  07 27  47  67  14 34  54  74  15 35  55  75
+          // 80 100 120 140 81 101 121 141 88 108 128 148 89 109 129 149
+          // 84 104 124 144 85 105 125 145 92 112 132 152 93 113 133 153
+          // 82 102 122 142 83 103 123 143 90 110 130 150 91 101 131 151
+          // 86 106 126 146 87 107 127 147 94 114 134 154 95 115 135 155
+          __m256i tr2_0 = _mm256_unpacklo_epi64(tr1_0, tr1_4);
+          __m256i tr2_1 = _mm256_unpackhi_epi64(tr1_0, tr1_4);
+          __m256i tr2_2 = _mm256_unpacklo_epi64(tr1_2, tr1_6);
+          __m256i tr2_3 = _mm256_unpackhi_epi64(tr1_2, tr1_6);
+          __m256i tr2_4 = _mm256_unpacklo_epi64(tr1_1, tr1_5);
+          __m256i tr2_5 = _mm256_unpackhi_epi64(tr1_1, tr1_5);
+          __m256i tr2_6 = _mm256_unpacklo_epi64(tr1_3, tr1_7);
+          __m256i tr2_7 = _mm256_unpackhi_epi64(tr1_3, tr1_7);
+          // 00 20 40 60 80 100 120 140 08 28 48 68 88 108 128 148
+          // 01 21 41 61 81 101 121 141 09 29 49 69 89 109 129 149
+          // 02 22 42 62 82 102 122 142 10 30 50 70 90 110 130 150
+          // 03 23 43 63 83 103 123 143 11 31 51 71 91 101 131 151
+          // 04 24 44 64 84 104 124 144 12 32 52 72 92 112 132 152
+          // 05 25 45 65 85 105 125 145 13 33 53 73 93 113 133 153
+          // 06 26 46 66 86 106 126 146 14 34 54 74 94 114 134 154
+          // 07 27 47 67 87 107 127 147 15 35 55 75 95 115 135 155
+          if (0 == pass) {
+            // output[j] = (output[j] + 1 + (output[j] > 0)) >> 2;
+            // TODO(cd): see quality impact of only doing
+            //           output[j] = (output[j] + 1) >> 2;
+            //           which would remove the code between here ...
+            __m256i tr2_0_0 = _mm256_cmpgt_epi16(tr2_0, kZero);
+            __m256i tr2_1_0 = _mm256_cmpgt_epi16(tr2_1, kZero);
+            __m256i tr2_2_0 = _mm256_cmpgt_epi16(tr2_2, kZero);
+            __m256i tr2_3_0 = _mm256_cmpgt_epi16(tr2_3, kZero);
+            __m256i tr2_4_0 = _mm256_cmpgt_epi16(tr2_4, kZero);
+            __m256i tr2_5_0 = _mm256_cmpgt_epi16(tr2_5, kZero);
+            __m256i tr2_6_0 = _mm256_cmpgt_epi16(tr2_6, kZero);
+            __m256i tr2_7_0 = _mm256_cmpgt_epi16(tr2_7, kZero);
+            tr2_0 = _mm256_sub_epi16(tr2_0, tr2_0_0);
+            tr2_1 = _mm256_sub_epi16(tr2_1, tr2_1_0);
+            tr2_2 = _mm256_sub_epi16(tr2_2, tr2_2_0);
+            tr2_3 = _mm256_sub_epi16(tr2_3, tr2_3_0);
+            tr2_4 = _mm256_sub_epi16(tr2_4, tr2_4_0);
+            tr2_5 = _mm256_sub_epi16(tr2_5, tr2_5_0);
+            tr2_6 = _mm256_sub_epi16(tr2_6, tr2_6_0);
+            tr2_7 = _mm256_sub_epi16(tr2_7, tr2_7_0);
+            //           ... and here.
+            //           PS: also change code in av1/encoder/av1_dct.c
+            tr2_0 = _mm256_add_epi16(tr2_0, kOne);
+            tr2_1 = _mm256_add_epi16(tr2_1, kOne);
+            tr2_2 = _mm256_add_epi16(tr2_2, kOne);
+            tr2_3 = _mm256_add_epi16(tr2_3, kOne);
+            tr2_4 = _mm256_add_epi16(tr2_4, kOne);
+            tr2_5 = _mm256_add_epi16(tr2_5, kOne);
+            tr2_6 = _mm256_add_epi16(tr2_6, kOne);
+            tr2_7 = _mm256_add_epi16(tr2_7, kOne);
+            tr2_0 = _mm256_srai_epi16(tr2_0, 2);
+            tr2_1 = _mm256_srai_epi16(tr2_1, 2);
+            tr2_2 = _mm256_srai_epi16(tr2_2, 2);
+            tr2_3 = _mm256_srai_epi16(tr2_3, 2);
+            tr2_4 = _mm256_srai_epi16(tr2_4, 2);
+            tr2_5 = _mm256_srai_epi16(tr2_5, 2);
+            tr2_6 = _mm256_srai_epi16(tr2_6, 2);
+            tr2_7 = _mm256_srai_epi16(tr2_7, 2);
+          }
+          if (0 == pass) {
+            // Note: even though all these stores are aligned, using the aligned
+            //       intrinsic make the code slightly slower.
+            _mm_storeu_si128((__m128i *)(output_currStep + 0 * 32),
+                             _mm256_castsi256_si128(tr2_0));
+            _mm_storeu_si128((__m128i *)(output_currStep + 1 * 32),
+                             _mm256_castsi256_si128(tr2_1));
+            _mm_storeu_si128((__m128i *)(output_currStep + 2 * 32),
+                             _mm256_castsi256_si128(tr2_2));
+            _mm_storeu_si128((__m128i *)(output_currStep + 3 * 32),
+                             _mm256_castsi256_si128(tr2_3));
+            _mm_storeu_si128((__m128i *)(output_currStep + 4 * 32),
+                             _mm256_castsi256_si128(tr2_4));
+            _mm_storeu_si128((__m128i *)(output_currStep + 5 * 32),
+                             _mm256_castsi256_si128(tr2_5));
+            _mm_storeu_si128((__m128i *)(output_currStep + 6 * 32),
+                             _mm256_castsi256_si128(tr2_6));
+            _mm_storeu_si128((__m128i *)(output_currStep + 7 * 32),
+                             _mm256_castsi256_si128(tr2_7));
+
+            _mm_storeu_si128((__m128i *)(output_nextStep + 0 * 32),
+                             _mm256_extractf128_si256(tr2_0, 1));
+            _mm_storeu_si128((__m128i *)(output_nextStep + 1 * 32),
+                             _mm256_extractf128_si256(tr2_1, 1));
+            _mm_storeu_si128((__m128i *)(output_nextStep + 2 * 32),
+                             _mm256_extractf128_si256(tr2_2, 1));
+            _mm_storeu_si128((__m128i *)(output_nextStep + 3 * 32),
+                             _mm256_extractf128_si256(tr2_3, 1));
+            _mm_storeu_si128((__m128i *)(output_nextStep + 4 * 32),
+                             _mm256_extractf128_si256(tr2_4, 1));
+            _mm_storeu_si128((__m128i *)(output_nextStep + 5 * 32),
+                             _mm256_extractf128_si256(tr2_5, 1));
+            _mm_storeu_si128((__m128i *)(output_nextStep + 6 * 32),
+                             _mm256_extractf128_si256(tr2_6, 1));
+            _mm_storeu_si128((__m128i *)(output_nextStep + 7 * 32),
+                             _mm256_extractf128_si256(tr2_7, 1));
+            // Process next 8x8
+            output_currStep += 8;
+            output_nextStep += 8;
+          }
+          if (1 == pass) {
+            store_coeff(&tr2_0, curr_out + 0 * 32, next_out + 0 * 32);
+            store_coeff(&tr2_1, curr_out + 1 * 32, next_out + 1 * 32);
+            store_coeff(&tr2_2, curr_out + 2 * 32, next_out + 2 * 32);
+            store_coeff(&tr2_3, curr_out + 3 * 32, next_out + 3 * 32);
+            store_coeff(&tr2_4, curr_out + 4 * 32, next_out + 4 * 32);
+            store_coeff(&tr2_5, curr_out + 5 * 32, next_out + 5 * 32);
+            store_coeff(&tr2_6, curr_out + 6 * 32, next_out + 6 * 32);
+            store_coeff(&tr2_7, curr_out + 7 * 32, next_out + 7 * 32);
+            curr_out += 8;
+            next_out += 8;
+          }
+        }
+      }
+    }
+  }
+  _mm256_zeroupper();
+}  // NOLINT
diff --git a/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_sse2.h b/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_sse2.h
new file mode 100644
index 0000000000..69dd6af119
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_dct32x32_impl_sse2.h
@@ -0,0 +1,3201 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "aom_dsp/fwd_txfm.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+
+// TODO(jingning) The high bit-depth version needs re-work for performance.
+// The current SSE2 implementation also causes cross reference to the static
+// functions in the C implementation file.
+#if DCT_HIGH_BIT_DEPTH
+#define ADD_EPI16 _mm_adds_epi16
+#define SUB_EPI16 _mm_subs_epi16
+#if FDCT32x32_HIGH_PRECISION
+void aom_fdct32x32_rows_c(const int16_t *intermediate, tran_low_t *out) {
+  int i, j;
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j) temp_in[j] = intermediate[j * 32 + i];
+    aom_fdct32(temp_in, temp_out, 0);
+    for (j = 0; j < 32; ++j)
+      out[j + i * 32] =
+          (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2);
+  }
+}
+#define HIGH_FDCT32x32_2D_C aom_highbd_fdct32x32_c
+#define HIGH_FDCT32x32_2D_ROWS_C aom_fdct32x32_rows_c
+#else
+void aom_fdct32x32_rd_rows_c(const int16_t *intermediate, tran_low_t *out) {
+  int i, j;
+  for (i = 0; i < 32; ++i) {
+    tran_high_t temp_in[32], temp_out[32];
+    for (j = 0; j < 32; ++j) temp_in[j] = intermediate[j * 32 + i];
+    aom_fdct32(temp_in, temp_out, 1);
+    for (j = 0; j < 32; ++j) out[j + i * 32] = (tran_low_t)temp_out[j];
+  }
+}
+#define HIGH_FDCT32x32_2D_C aom_highbd_fdct32x32_rd_c
+#define HIGH_FDCT32x32_2D_ROWS_C aom_fdct32x32_rd_rows_c
+#endif  // FDCT32x32_HIGH_PRECISION
+#else
+#define ADD_EPI16 _mm_add_epi16
+#define SUB_EPI16 _mm_sub_epi16
+#endif  // DCT_HIGH_BIT_DEPTH
+
+void FDCT32x32_2D(const int16_t *input, tran_low_t *output_org, int stride) {
+  // Calculate pre-multiplied strides
+  const int str1 = stride;
+  const int str2 = 2 * stride;
+  const int str3 = 2 * stride + str1;
+  // We need an intermediate buffer between passes.
+  DECLARE_ALIGNED(16, int16_t, intermediate[32 * 32]);
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(+cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(+cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(+cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(+cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m28_m04 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_m12_m20 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_p30_p02 = pair_set_epi16(+cospi_30_64, cospi_2_64);
+  const __m128i k__cospi_p14_p18 = pair_set_epi16(+cospi_14_64, cospi_18_64);
+  const __m128i k__cospi_p22_p10 = pair_set_epi16(+cospi_22_64, cospi_10_64);
+  const __m128i k__cospi_p06_p26 = pair_set_epi16(+cospi_6_64, cospi_26_64);
+  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p31_p01 = pair_set_epi16(+cospi_31_64, cospi_1_64);
+  const __m128i k__cospi_p15_p17 = pair_set_epi16(+cospi_15_64, cospi_17_64);
+  const __m128i k__cospi_p23_p09 = pair_set_epi16(+cospi_23_64, cospi_9_64);
+  const __m128i k__cospi_p07_p25 = pair_set_epi16(+cospi_7_64, cospi_25_64);
+  const __m128i k__cospi_m25_p07 = pair_set_epi16(-cospi_25_64, cospi_7_64);
+  const __m128i k__cospi_m09_p23 = pair_set_epi16(-cospi_9_64, cospi_23_64);
+  const __m128i k__cospi_m17_p15 = pair_set_epi16(-cospi_17_64, cospi_15_64);
+  const __m128i k__cospi_m01_p31 = pair_set_epi16(-cospi_1_64, cospi_31_64);
+  const __m128i k__cospi_p27_p05 = pair_set_epi16(+cospi_27_64, cospi_5_64);
+  const __m128i k__cospi_p11_p21 = pair_set_epi16(+cospi_11_64, cospi_21_64);
+  const __m128i k__cospi_p19_p13 = pair_set_epi16(+cospi_19_64, cospi_13_64);
+  const __m128i k__cospi_p03_p29 = pair_set_epi16(+cospi_3_64, cospi_29_64);
+  const __m128i k__cospi_m29_p03 = pair_set_epi16(-cospi_29_64, cospi_3_64);
+  const __m128i k__cospi_m13_p19 = pair_set_epi16(-cospi_13_64, cospi_19_64);
+  const __m128i k__cospi_m21_p11 = pair_set_epi16(-cospi_21_64, cospi_11_64);
+  const __m128i k__cospi_m05_p27 = pair_set_epi16(-cospi_5_64, cospi_27_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i kOne = _mm_set1_epi16(1);
+  // Do the two transform/transpose passes
+  int pass;
+#if DCT_HIGH_BIT_DEPTH
+  int overflow;
+#endif
+  for (pass = 0; pass < 2; ++pass) {
+    // We process eight columns (transposed rows in second pass) at a time.
+    int column_start;
+    for (column_start = 0; column_start < 32; column_start += 8) {
+      __m128i step1[32];
+      __m128i step2[32];
+      __m128i step3[32];
+      __m128i out[32];
+      // Stage 1
+      // Note: even though all the loads below are aligned, using the aligned
+      //       intrinsic make the code slightly slower.
+      if (0 == pass) {
+        const int16_t *in = &input[column_start];
+        // step1[i] =  (in[ 0 * stride] + in[(32 -  1) * stride]) << 2;
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          const int16_t *ina = in + 0 * str1;
+          const int16_t *inb = in + 31 * str1;
+          __m128i *step1a = &step1[0];
+          __m128i *step1b = &step1[31];
+          const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[0] = _mm_add_epi16(ina0, inb0);
+          step1a[1] = _mm_add_epi16(ina1, inb1);
+          step1a[2] = _mm_add_epi16(ina2, inb2);
+          step1a[3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[0] = _mm_slli_epi16(step1a[0], 2);
+          step1a[1] = _mm_slli_epi16(step1a[1], 2);
+          step1a[2] = _mm_slli_epi16(step1a[2], 2);
+          step1a[3] = _mm_slli_epi16(step1a[3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina = in + 4 * str1;
+          const int16_t *inb = in + 27 * str1;
+          __m128i *step1a = &step1[4];
+          __m128i *step1b = &step1[27];
+          const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[0] = _mm_add_epi16(ina0, inb0);
+          step1a[1] = _mm_add_epi16(ina1, inb1);
+          step1a[2] = _mm_add_epi16(ina2, inb2);
+          step1a[3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[0] = _mm_slli_epi16(step1a[0], 2);
+          step1a[1] = _mm_slli_epi16(step1a[1], 2);
+          step1a[2] = _mm_slli_epi16(step1a[2], 2);
+          step1a[3] = _mm_slli_epi16(step1a[3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina = in + 8 * str1;
+          const int16_t *inb = in + 23 * str1;
+          __m128i *step1a = &step1[8];
+          __m128i *step1b = &step1[23];
+          const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[0] = _mm_add_epi16(ina0, inb0);
+          step1a[1] = _mm_add_epi16(ina1, inb1);
+          step1a[2] = _mm_add_epi16(ina2, inb2);
+          step1a[3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[0] = _mm_slli_epi16(step1a[0], 2);
+          step1a[1] = _mm_slli_epi16(step1a[1], 2);
+          step1a[2] = _mm_slli_epi16(step1a[2], 2);
+          step1a[3] = _mm_slli_epi16(step1a[3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina = in + 12 * str1;
+          const int16_t *inb = in + 19 * str1;
+          __m128i *step1a = &step1[12];
+          __m128i *step1b = &step1[19];
+          const __m128i ina0 = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1 = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2 = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3 = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3 = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2 = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1 = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0 = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[0] = _mm_add_epi16(ina0, inb0);
+          step1a[1] = _mm_add_epi16(ina1, inb1);
+          step1a[2] = _mm_add_epi16(ina2, inb2);
+          step1a[3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[0] = _mm_slli_epi16(step1a[0], 2);
+          step1a[1] = _mm_slli_epi16(step1a[1], 2);
+          step1a[2] = _mm_slli_epi16(step1a[2], 2);
+          step1a[3] = _mm_slli_epi16(step1a[3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+      } else {
+        int16_t *in = &intermediate[column_start];
+        // step1[i] =  in[ 0 * 32] + in[(32 -  1) * 32];
+        // Note: using the same approach as above to have common offset is
+        //       counter-productive as all offsets can be calculated at compile
+        //       time.
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          __m128i in00 = _mm_loadu_si128((const __m128i *)(in + 0 * 32));
+          __m128i in01 = _mm_loadu_si128((const __m128i *)(in + 1 * 32));
+          __m128i in02 = _mm_loadu_si128((const __m128i *)(in + 2 * 32));
+          __m128i in03 = _mm_loadu_si128((const __m128i *)(in + 3 * 32));
+          __m128i in28 = _mm_loadu_si128((const __m128i *)(in + 28 * 32));
+          __m128i in29 = _mm_loadu_si128((const __m128i *)(in + 29 * 32));
+          __m128i in30 = _mm_loadu_si128((const __m128i *)(in + 30 * 32));
+          __m128i in31 = _mm_loadu_si128((const __m128i *)(in + 31 * 32));
+          step1[0] = ADD_EPI16(in00, in31);
+          step1[1] = ADD_EPI16(in01, in30);
+          step1[2] = ADD_EPI16(in02, in29);
+          step1[3] = ADD_EPI16(in03, in28);
+          step1[28] = SUB_EPI16(in03, in28);
+          step1[29] = SUB_EPI16(in02, in29);
+          step1[30] = SUB_EPI16(in01, in30);
+          step1[31] = SUB_EPI16(in00, in31);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[0], &step1[1], &step1[2],
+                                             &step1[3], &step1[28], &step1[29],
+                                             &step1[30], &step1[31]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          __m128i in04 = _mm_loadu_si128((const __m128i *)(in + 4 * 32));
+          __m128i in05 = _mm_loadu_si128((const __m128i *)(in + 5 * 32));
+          __m128i in06 = _mm_loadu_si128((const __m128i *)(in + 6 * 32));
+          __m128i in07 = _mm_loadu_si128((const __m128i *)(in + 7 * 32));
+          __m128i in24 = _mm_loadu_si128((const __m128i *)(in + 24 * 32));
+          __m128i in25 = _mm_loadu_si128((const __m128i *)(in + 25 * 32));
+          __m128i in26 = _mm_loadu_si128((const __m128i *)(in + 26 * 32));
+          __m128i in27 = _mm_loadu_si128((const __m128i *)(in + 27 * 32));
+          step1[4] = ADD_EPI16(in04, in27);
+          step1[5] = ADD_EPI16(in05, in26);
+          step1[6] = ADD_EPI16(in06, in25);
+          step1[7] = ADD_EPI16(in07, in24);
+          step1[24] = SUB_EPI16(in07, in24);
+          step1[25] = SUB_EPI16(in06, in25);
+          step1[26] = SUB_EPI16(in05, in26);
+          step1[27] = SUB_EPI16(in04, in27);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[4], &step1[5], &step1[6],
+                                             &step1[7], &step1[24], &step1[25],
+                                             &step1[26], &step1[27]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          __m128i in08 = _mm_loadu_si128((const __m128i *)(in + 8 * 32));
+          __m128i in09 = _mm_loadu_si128((const __m128i *)(in + 9 * 32));
+          __m128i in10 = _mm_loadu_si128((const __m128i *)(in + 10 * 32));
+          __m128i in11 = _mm_loadu_si128((const __m128i *)(in + 11 * 32));
+          __m128i in20 = _mm_loadu_si128((const __m128i *)(in + 20 * 32));
+          __m128i in21 = _mm_loadu_si128((const __m128i *)(in + 21 * 32));
+          __m128i in22 = _mm_loadu_si128((const __m128i *)(in + 22 * 32));
+          __m128i in23 = _mm_loadu_si128((const __m128i *)(in + 23 * 32));
+          step1[8] = ADD_EPI16(in08, in23);
+          step1[9] = ADD_EPI16(in09, in22);
+          step1[10] = ADD_EPI16(in10, in21);
+          step1[11] = ADD_EPI16(in11, in20);
+          step1[20] = SUB_EPI16(in11, in20);
+          step1[21] = SUB_EPI16(in10, in21);
+          step1[22] = SUB_EPI16(in09, in22);
+          step1[23] = SUB_EPI16(in08, in23);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[8], &step1[9], &step1[10],
+                                             &step1[11], &step1[20], &step1[21],
+                                             &step1[22], &step1[23]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          __m128i in12 = _mm_loadu_si128((const __m128i *)(in + 12 * 32));
+          __m128i in13 = _mm_loadu_si128((const __m128i *)(in + 13 * 32));
+          __m128i in14 = _mm_loadu_si128((const __m128i *)(in + 14 * 32));
+          __m128i in15 = _mm_loadu_si128((const __m128i *)(in + 15 * 32));
+          __m128i in16 = _mm_loadu_si128((const __m128i *)(in + 16 * 32));
+          __m128i in17 = _mm_loadu_si128((const __m128i *)(in + 17 * 32));
+          __m128i in18 = _mm_loadu_si128((const __m128i *)(in + 18 * 32));
+          __m128i in19 = _mm_loadu_si128((const __m128i *)(in + 19 * 32));
+          step1[12] = ADD_EPI16(in12, in19);
+          step1[13] = ADD_EPI16(in13, in18);
+          step1[14] = ADD_EPI16(in14, in17);
+          step1[15] = ADD_EPI16(in15, in16);
+          step1[16] = SUB_EPI16(in15, in16);
+          step1[17] = SUB_EPI16(in14, in17);
+          step1[18] = SUB_EPI16(in13, in18);
+          step1[19] = SUB_EPI16(in12, in19);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[12], &step1[13], &step1[14],
+                                             &step1[15], &step1[16], &step1[17],
+                                             &step1[18], &step1[19]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+      // Stage 2
+      {
+        step2[0] = ADD_EPI16(step1[0], step1[15]);
+        step2[1] = ADD_EPI16(step1[1], step1[14]);
+        step2[2] = ADD_EPI16(step1[2], step1[13]);
+        step2[3] = ADD_EPI16(step1[3], step1[12]);
+        step2[4] = ADD_EPI16(step1[4], step1[11]);
+        step2[5] = ADD_EPI16(step1[5], step1[10]);
+        step2[6] = ADD_EPI16(step1[6], step1[9]);
+        step2[7] = ADD_EPI16(step1[7], step1[8]);
+        step2[8] = SUB_EPI16(step1[7], step1[8]);
+        step2[9] = SUB_EPI16(step1[6], step1[9]);
+        step2[10] = SUB_EPI16(step1[5], step1[10]);
+        step2[11] = SUB_EPI16(step1[4], step1[11]);
+        step2[12] = SUB_EPI16(step1[3], step1[12]);
+        step2[13] = SUB_EPI16(step1[2], step1[13]);
+        step2[14] = SUB_EPI16(step1[1], step1[14]);
+        step2[15] = SUB_EPI16(step1[0], step1[15]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step2[0], &step2[1], &step2[2], &step2[3], &step2[4], &step2[5],
+            &step2[6], &step2[7], &step2[8], &step2[9], &step2[10], &step2[11],
+            &step2[12], &step2[13], &step2[14], &step2[15]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]);
+        const __m128i s2_20_1 = _mm_unpackhi_epi16(step1[27], step1[20]);
+        const __m128i s2_21_0 = _mm_unpacklo_epi16(step1[26], step1[21]);
+        const __m128i s2_21_1 = _mm_unpackhi_epi16(step1[26], step1[21]);
+        const __m128i s2_22_0 = _mm_unpacklo_epi16(step1[25], step1[22]);
+        const __m128i s2_22_1 = _mm_unpackhi_epi16(step1[25], step1[22]);
+        const __m128i s2_23_0 = _mm_unpacklo_epi16(step1[24], step1[23]);
+        const __m128i s2_23_1 = _mm_unpackhi_epi16(step1[24], step1[23]);
+        const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
+        const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
+        const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
+        const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
+        const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
+        const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
+        const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
+        const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
+        const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
+        const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16);
+        const __m128i s2_25_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_p16);
+        const __m128i s2_25_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_p16);
+        const __m128i s2_26_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_p16);
+        const __m128i s2_26_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_p16);
+        const __m128i s2_27_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_p16);
+        const __m128i s2_27_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m128i s2_20_4 = _mm_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_20_5 = _mm_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_21_4 = _mm_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_21_5 = _mm_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_22_4 = _mm_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_22_5 = _mm_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_23_4 = _mm_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_23_5 = _mm_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_24_4 = _mm_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_24_5 = _mm_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_25_4 = _mm_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_25_5 = _mm_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_26_4 = _mm_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_26_5 = _mm_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_27_4 = _mm_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_27_5 = _mm_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_20_6 = _mm_srai_epi32(s2_20_4, DCT_CONST_BITS);
+        const __m128i s2_20_7 = _mm_srai_epi32(s2_20_5, DCT_CONST_BITS);
+        const __m128i s2_21_6 = _mm_srai_epi32(s2_21_4, DCT_CONST_BITS);
+        const __m128i s2_21_7 = _mm_srai_epi32(s2_21_5, DCT_CONST_BITS);
+        const __m128i s2_22_6 = _mm_srai_epi32(s2_22_4, DCT_CONST_BITS);
+        const __m128i s2_22_7 = _mm_srai_epi32(s2_22_5, DCT_CONST_BITS);
+        const __m128i s2_23_6 = _mm_srai_epi32(s2_23_4, DCT_CONST_BITS);
+        const __m128i s2_23_7 = _mm_srai_epi32(s2_23_5, DCT_CONST_BITS);
+        const __m128i s2_24_6 = _mm_srai_epi32(s2_24_4, DCT_CONST_BITS);
+        const __m128i s2_24_7 = _mm_srai_epi32(s2_24_5, DCT_CONST_BITS);
+        const __m128i s2_25_6 = _mm_srai_epi32(s2_25_4, DCT_CONST_BITS);
+        const __m128i s2_25_7 = _mm_srai_epi32(s2_25_5, DCT_CONST_BITS);
+        const __m128i s2_26_6 = _mm_srai_epi32(s2_26_4, DCT_CONST_BITS);
+        const __m128i s2_26_7 = _mm_srai_epi32(s2_26_5, DCT_CONST_BITS);
+        const __m128i s2_27_6 = _mm_srai_epi32(s2_27_4, DCT_CONST_BITS);
+        const __m128i s2_27_7 = _mm_srai_epi32(s2_27_5, DCT_CONST_BITS);
+        // Combine
+        step2[20] = _mm_packs_epi32(s2_20_6, s2_20_7);
+        step2[21] = _mm_packs_epi32(s2_21_6, s2_21_7);
+        step2[22] = _mm_packs_epi32(s2_22_6, s2_22_7);
+        step2[23] = _mm_packs_epi32(s2_23_6, s2_23_7);
+        step2[24] = _mm_packs_epi32(s2_24_6, s2_24_7);
+        step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7);
+        step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7);
+        step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step2[20], &step2[21], &step2[22],
+                                           &step2[23], &step2[24], &step2[25],
+                                           &step2[26], &step2[27]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+
+#if !FDCT32x32_HIGH_PRECISION
+      // dump the magnitude by half, hence the intermediate values are within
+      // the range of 16 bits.
+      if (1 == pass) {
+        __m128i s3_00_0 = _mm_cmplt_epi16(step2[0], kZero);
+        __m128i s3_01_0 = _mm_cmplt_epi16(step2[1], kZero);
+        __m128i s3_02_0 = _mm_cmplt_epi16(step2[2], kZero);
+        __m128i s3_03_0 = _mm_cmplt_epi16(step2[3], kZero);
+        __m128i s3_04_0 = _mm_cmplt_epi16(step2[4], kZero);
+        __m128i s3_05_0 = _mm_cmplt_epi16(step2[5], kZero);
+        __m128i s3_06_0 = _mm_cmplt_epi16(step2[6], kZero);
+        __m128i s3_07_0 = _mm_cmplt_epi16(step2[7], kZero);
+        __m128i s2_08_0 = _mm_cmplt_epi16(step2[8], kZero);
+        __m128i s2_09_0 = _mm_cmplt_epi16(step2[9], kZero);
+        __m128i s3_10_0 = _mm_cmplt_epi16(step2[10], kZero);
+        __m128i s3_11_0 = _mm_cmplt_epi16(step2[11], kZero);
+        __m128i s3_12_0 = _mm_cmplt_epi16(step2[12], kZero);
+        __m128i s3_13_0 = _mm_cmplt_epi16(step2[13], kZero);
+        __m128i s2_14_0 = _mm_cmplt_epi16(step2[14], kZero);
+        __m128i s2_15_0 = _mm_cmplt_epi16(step2[15], kZero);
+        __m128i s3_16_0 = _mm_cmplt_epi16(step1[16], kZero);
+        __m128i s3_17_0 = _mm_cmplt_epi16(step1[17], kZero);
+        __m128i s3_18_0 = _mm_cmplt_epi16(step1[18], kZero);
+        __m128i s3_19_0 = _mm_cmplt_epi16(step1[19], kZero);
+        __m128i s3_20_0 = _mm_cmplt_epi16(step2[20], kZero);
+        __m128i s3_21_0 = _mm_cmplt_epi16(step2[21], kZero);
+        __m128i s3_22_0 = _mm_cmplt_epi16(step2[22], kZero);
+        __m128i s3_23_0 = _mm_cmplt_epi16(step2[23], kZero);
+        __m128i s3_24_0 = _mm_cmplt_epi16(step2[24], kZero);
+        __m128i s3_25_0 = _mm_cmplt_epi16(step2[25], kZero);
+        __m128i s3_26_0 = _mm_cmplt_epi16(step2[26], kZero);
+        __m128i s3_27_0 = _mm_cmplt_epi16(step2[27], kZero);
+        __m128i s3_28_0 = _mm_cmplt_epi16(step1[28], kZero);
+        __m128i s3_29_0 = _mm_cmplt_epi16(step1[29], kZero);
+        __m128i s3_30_0 = _mm_cmplt_epi16(step1[30], kZero);
+        __m128i s3_31_0 = _mm_cmplt_epi16(step1[31], kZero);
+
+        step2[0] = SUB_EPI16(step2[0], s3_00_0);
+        step2[1] = SUB_EPI16(step2[1], s3_01_0);
+        step2[2] = SUB_EPI16(step2[2], s3_02_0);
+        step2[3] = SUB_EPI16(step2[3], s3_03_0);
+        step2[4] = SUB_EPI16(step2[4], s3_04_0);
+        step2[5] = SUB_EPI16(step2[5], s3_05_0);
+        step2[6] = SUB_EPI16(step2[6], s3_06_0);
+        step2[7] = SUB_EPI16(step2[7], s3_07_0);
+        step2[8] = SUB_EPI16(step2[8], s2_08_0);
+        step2[9] = SUB_EPI16(step2[9], s2_09_0);
+        step2[10] = SUB_EPI16(step2[10], s3_10_0);
+        step2[11] = SUB_EPI16(step2[11], s3_11_0);
+        step2[12] = SUB_EPI16(step2[12], s3_12_0);
+        step2[13] = SUB_EPI16(step2[13], s3_13_0);
+        step2[14] = SUB_EPI16(step2[14], s2_14_0);
+        step2[15] = SUB_EPI16(step2[15], s2_15_0);
+        step1[16] = SUB_EPI16(step1[16], s3_16_0);
+        step1[17] = SUB_EPI16(step1[17], s3_17_0);
+        step1[18] = SUB_EPI16(step1[18], s3_18_0);
+        step1[19] = SUB_EPI16(step1[19], s3_19_0);
+        step2[20] = SUB_EPI16(step2[20], s3_20_0);
+        step2[21] = SUB_EPI16(step2[21], s3_21_0);
+        step2[22] = SUB_EPI16(step2[22], s3_22_0);
+        step2[23] = SUB_EPI16(step2[23], s3_23_0);
+        step2[24] = SUB_EPI16(step2[24], s3_24_0);
+        step2[25] = SUB_EPI16(step2[25], s3_25_0);
+        step2[26] = SUB_EPI16(step2[26], s3_26_0);
+        step2[27] = SUB_EPI16(step2[27], s3_27_0);
+        step1[28] = SUB_EPI16(step1[28], s3_28_0);
+        step1[29] = SUB_EPI16(step1[29], s3_29_0);
+        step1[30] = SUB_EPI16(step1[30], s3_30_0);
+        step1[31] = SUB_EPI16(step1[31], s3_31_0);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x32(
+            &step2[0], &step2[1], &step2[2], &step2[3], &step2[4], &step2[5],
+            &step2[6], &step2[7], &step2[8], &step2[9], &step2[10], &step2[11],
+            &step2[12], &step2[13], &step2[14], &step2[15], &step1[16],
+            &step1[17], &step1[18], &step1[19], &step2[20], &step2[21],
+            &step2[22], &step2[23], &step2[24], &step2[25], &step2[26],
+            &step2[27], &step1[28], &step1[29], &step1[30], &step1[31]);
+        if (overflow) {
+          HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        step2[0] = _mm_add_epi16(step2[0], kOne);
+        step2[1] = _mm_add_epi16(step2[1], kOne);
+        step2[2] = _mm_add_epi16(step2[2], kOne);
+        step2[3] = _mm_add_epi16(step2[3], kOne);
+        step2[4] = _mm_add_epi16(step2[4], kOne);
+        step2[5] = _mm_add_epi16(step2[5], kOne);
+        step2[6] = _mm_add_epi16(step2[6], kOne);
+        step2[7] = _mm_add_epi16(step2[7], kOne);
+        step2[8] = _mm_add_epi16(step2[8], kOne);
+        step2[9] = _mm_add_epi16(step2[9], kOne);
+        step2[10] = _mm_add_epi16(step2[10], kOne);
+        step2[11] = _mm_add_epi16(step2[11], kOne);
+        step2[12] = _mm_add_epi16(step2[12], kOne);
+        step2[13] = _mm_add_epi16(step2[13], kOne);
+        step2[14] = _mm_add_epi16(step2[14], kOne);
+        step2[15] = _mm_add_epi16(step2[15], kOne);
+        step1[16] = _mm_add_epi16(step1[16], kOne);
+        step1[17] = _mm_add_epi16(step1[17], kOne);
+        step1[18] = _mm_add_epi16(step1[18], kOne);
+        step1[19] = _mm_add_epi16(step1[19], kOne);
+        step2[20] = _mm_add_epi16(step2[20], kOne);
+        step2[21] = _mm_add_epi16(step2[21], kOne);
+        step2[22] = _mm_add_epi16(step2[22], kOne);
+        step2[23] = _mm_add_epi16(step2[23], kOne);
+        step2[24] = _mm_add_epi16(step2[24], kOne);
+        step2[25] = _mm_add_epi16(step2[25], kOne);
+        step2[26] = _mm_add_epi16(step2[26], kOne);
+        step2[27] = _mm_add_epi16(step2[27], kOne);
+        step1[28] = _mm_add_epi16(step1[28], kOne);
+        step1[29] = _mm_add_epi16(step1[29], kOne);
+        step1[30] = _mm_add_epi16(step1[30], kOne);
+        step1[31] = _mm_add_epi16(step1[31], kOne);
+
+        step2[0] = _mm_srai_epi16(step2[0], 2);
+        step2[1] = _mm_srai_epi16(step2[1], 2);
+        step2[2] = _mm_srai_epi16(step2[2], 2);
+        step2[3] = _mm_srai_epi16(step2[3], 2);
+        step2[4] = _mm_srai_epi16(step2[4], 2);
+        step2[5] = _mm_srai_epi16(step2[5], 2);
+        step2[6] = _mm_srai_epi16(step2[6], 2);
+        step2[7] = _mm_srai_epi16(step2[7], 2);
+        step2[8] = _mm_srai_epi16(step2[8], 2);
+        step2[9] = _mm_srai_epi16(step2[9], 2);
+        step2[10] = _mm_srai_epi16(step2[10], 2);
+        step2[11] = _mm_srai_epi16(step2[11], 2);
+        step2[12] = _mm_srai_epi16(step2[12], 2);
+        step2[13] = _mm_srai_epi16(step2[13], 2);
+        step2[14] = _mm_srai_epi16(step2[14], 2);
+        step2[15] = _mm_srai_epi16(step2[15], 2);
+        step1[16] = _mm_srai_epi16(step1[16], 2);
+        step1[17] = _mm_srai_epi16(step1[17], 2);
+        step1[18] = _mm_srai_epi16(step1[18], 2);
+        step1[19] = _mm_srai_epi16(step1[19], 2);
+        step2[20] = _mm_srai_epi16(step2[20], 2);
+        step2[21] = _mm_srai_epi16(step2[21], 2);
+        step2[22] = _mm_srai_epi16(step2[22], 2);
+        step2[23] = _mm_srai_epi16(step2[23], 2);
+        step2[24] = _mm_srai_epi16(step2[24], 2);
+        step2[25] = _mm_srai_epi16(step2[25], 2);
+        step2[26] = _mm_srai_epi16(step2[26], 2);
+        step2[27] = _mm_srai_epi16(step2[27], 2);
+        step1[28] = _mm_srai_epi16(step1[28], 2);
+        step1[29] = _mm_srai_epi16(step1[29], 2);
+        step1[30] = _mm_srai_epi16(step1[30], 2);
+        step1[31] = _mm_srai_epi16(step1[31], 2);
+      }
+#endif  // !FDCT32x32_HIGH_PRECISION
+
+#if FDCT32x32_HIGH_PRECISION
+      if (pass == 0) {
+#endif
+        // Stage 3
+        {
+          step3[0] = ADD_EPI16(step2[(8 - 1)], step2[0]);
+          step3[1] = ADD_EPI16(step2[(8 - 2)], step2[1]);
+          step3[2] = ADD_EPI16(step2[(8 - 3)], step2[2]);
+          step3[3] = ADD_EPI16(step2[(8 - 4)], step2[3]);
+          step3[4] = SUB_EPI16(step2[(8 - 5)], step2[4]);
+          step3[5] = SUB_EPI16(step2[(8 - 6)], step2[5]);
+          step3[6] = SUB_EPI16(step2[(8 - 7)], step2[6]);
+          step3[7] = SUB_EPI16(step2[(8 - 8)], step2[7]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step3[0], &step3[1], &step3[2],
+                                             &step3[3], &step3[4], &step3[5],
+                                             &step3[6], &step3[7]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
+          const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
+          const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
+          const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
+          const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
+          const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
+          const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
+          const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
+          const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
+          const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
+          const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
+          const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
+          // dct_const_round_shift
+          const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_10_6 = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
+          const __m128i s3_10_7 = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
+          const __m128i s3_11_6 = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
+          const __m128i s3_11_7 = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
+          const __m128i s3_12_6 = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
+          const __m128i s3_12_7 = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
+          const __m128i s3_13_6 = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
+          const __m128i s3_13_7 = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
+          // Combine
+          step3[10] = _mm_packs_epi32(s3_10_6, s3_10_7);
+          step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7);
+          step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7);
+          step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&step3[10], &step3[11], &step3[12],
+                                             &step3[13]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          step3[16] = ADD_EPI16(step2[23], step1[16]);
+          step3[17] = ADD_EPI16(step2[22], step1[17]);
+          step3[18] = ADD_EPI16(step2[21], step1[18]);
+          step3[19] = ADD_EPI16(step2[20], step1[19]);
+          step3[20] = SUB_EPI16(step1[19], step2[20]);
+          step3[21] = SUB_EPI16(step1[18], step2[21]);
+          step3[22] = SUB_EPI16(step1[17], step2[22]);
+          step3[23] = SUB_EPI16(step1[16], step2[23]);
+          step3[24] = SUB_EPI16(step1[31], step2[24]);
+          step3[25] = SUB_EPI16(step1[30], step2[25]);
+          step3[26] = SUB_EPI16(step1[29], step2[26]);
+          step3[27] = SUB_EPI16(step1[28], step2[27]);
+          step3[28] = ADD_EPI16(step2[27], step1[28]);
+          step3[29] = ADD_EPI16(step2[26], step1[29]);
+          step3[30] = ADD_EPI16(step2[25], step1[30]);
+          step3[31] = ADD_EPI16(step2[24], step1[31]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x16(
+              &step3[16], &step3[17], &step3[18], &step3[19], &step3[20],
+              &step3[21], &step3[22], &step3[23], &step3[24], &step3[25],
+              &step3[26], &step3[27], &step3[28], &step3[29], &step3[30],
+              &step3[31]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+
+        // Stage 4
+        {
+          step1[0] = ADD_EPI16(step3[3], step3[0]);
+          step1[1] = ADD_EPI16(step3[2], step3[1]);
+          step1[2] = SUB_EPI16(step3[1], step3[2]);
+          step1[3] = SUB_EPI16(step3[0], step3[3]);
+          step1[8] = ADD_EPI16(step3[11], step2[8]);
+          step1[9] = ADD_EPI16(step3[10], step2[9]);
+          step1[10] = SUB_EPI16(step2[9], step3[10]);
+          step1[11] = SUB_EPI16(step2[8], step3[11]);
+          step1[12] = SUB_EPI16(step2[15], step3[12]);
+          step1[13] = SUB_EPI16(step2[14], step3[13]);
+          step1[14] = ADD_EPI16(step3[13], step2[14]);
+          step1[15] = ADD_EPI16(step3[12], step2[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x16(
+              &step1[0], &step1[1], &step1[2], &step1[3], &step1[4], &step1[5],
+              &step1[6], &step1[7], &step1[8], &step1[9], &step1[10],
+              &step1[11], &step1[12], &step1[13], &step1[14], &step1[15]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]);
+          const __m128i s1_05_1 = _mm_unpackhi_epi16(step3[6], step3[5]);
+          const __m128i s1_05_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_m16);
+          const __m128i s1_05_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_m16);
+          const __m128i s1_06_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_p16);
+          const __m128i s1_06_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_p16);
+          // dct_const_round_shift
+          const __m128i s1_05_4 = _mm_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
+          const __m128i s1_05_5 = _mm_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
+          const __m128i s1_06_4 = _mm_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
+          const __m128i s1_06_5 = _mm_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
+          const __m128i s1_05_6 = _mm_srai_epi32(s1_05_4, DCT_CONST_BITS);
+          const __m128i s1_05_7 = _mm_srai_epi32(s1_05_5, DCT_CONST_BITS);
+          const __m128i s1_06_6 = _mm_srai_epi32(s1_06_4, DCT_CONST_BITS);
+          const __m128i s1_06_7 = _mm_srai_epi32(s1_06_5, DCT_CONST_BITS);
+          // Combine
+          step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7);
+          step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x2(&step1[5], &step1[6]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]);
+          const __m128i s1_18_1 = _mm_unpackhi_epi16(step3[18], step3[29]);
+          const __m128i s1_19_0 = _mm_unpacklo_epi16(step3[19], step3[28]);
+          const __m128i s1_19_1 = _mm_unpackhi_epi16(step3[19], step3[28]);
+          const __m128i s1_20_0 = _mm_unpacklo_epi16(step3[20], step3[27]);
+          const __m128i s1_20_1 = _mm_unpackhi_epi16(step3[20], step3[27]);
+          const __m128i s1_21_0 = _mm_unpacklo_epi16(step3[21], step3[26]);
+          const __m128i s1_21_1 = _mm_unpackhi_epi16(step3[21], step3[26]);
+          const __m128i s1_18_2 = _mm_madd_epi16(s1_18_0, k__cospi_m08_p24);
+          const __m128i s1_18_3 = _mm_madd_epi16(s1_18_1, k__cospi_m08_p24);
+          const __m128i s1_19_2 = _mm_madd_epi16(s1_19_0, k__cospi_m08_p24);
+          const __m128i s1_19_3 = _mm_madd_epi16(s1_19_1, k__cospi_m08_p24);
+          const __m128i s1_20_2 = _mm_madd_epi16(s1_20_0, k__cospi_m24_m08);
+          const __m128i s1_20_3 = _mm_madd_epi16(s1_20_1, k__cospi_m24_m08);
+          const __m128i s1_21_2 = _mm_madd_epi16(s1_21_0, k__cospi_m24_m08);
+          const __m128i s1_21_3 = _mm_madd_epi16(s1_21_1, k__cospi_m24_m08);
+          const __m128i s1_26_2 = _mm_madd_epi16(s1_21_0, k__cospi_m08_p24);
+          const __m128i s1_26_3 = _mm_madd_epi16(s1_21_1, k__cospi_m08_p24);
+          const __m128i s1_27_2 = _mm_madd_epi16(s1_20_0, k__cospi_m08_p24);
+          const __m128i s1_27_3 = _mm_madd_epi16(s1_20_1, k__cospi_m08_p24);
+          const __m128i s1_28_2 = _mm_madd_epi16(s1_19_0, k__cospi_p24_p08);
+          const __m128i s1_28_3 = _mm_madd_epi16(s1_19_1, k__cospi_p24_p08);
+          const __m128i s1_29_2 = _mm_madd_epi16(s1_18_0, k__cospi_p24_p08);
+          const __m128i s1_29_3 = _mm_madd_epi16(s1_18_1, k__cospi_p24_p08);
+          // dct_const_round_shift
+          const __m128i s1_18_4 = _mm_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
+          const __m128i s1_18_5 = _mm_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
+          const __m128i s1_19_4 = _mm_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
+          const __m128i s1_19_5 = _mm_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
+          const __m128i s1_20_4 = _mm_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
+          const __m128i s1_20_5 = _mm_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
+          const __m128i s1_21_4 = _mm_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
+          const __m128i s1_21_5 = _mm_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
+          const __m128i s1_26_4 = _mm_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
+          const __m128i s1_26_5 = _mm_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
+          const __m128i s1_27_4 = _mm_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
+          const __m128i s1_27_5 = _mm_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
+          const __m128i s1_28_4 = _mm_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
+          const __m128i s1_28_5 = _mm_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
+          const __m128i s1_29_4 = _mm_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
+          const __m128i s1_29_5 = _mm_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
+          const __m128i s1_18_6 = _mm_srai_epi32(s1_18_4, DCT_CONST_BITS);
+          const __m128i s1_18_7 = _mm_srai_epi32(s1_18_5, DCT_CONST_BITS);
+          const __m128i s1_19_6 = _mm_srai_epi32(s1_19_4, DCT_CONST_BITS);
+          const __m128i s1_19_7 = _mm_srai_epi32(s1_19_5, DCT_CONST_BITS);
+          const __m128i s1_20_6 = _mm_srai_epi32(s1_20_4, DCT_CONST_BITS);
+          const __m128i s1_20_7 = _mm_srai_epi32(s1_20_5, DCT_CONST_BITS);
+          const __m128i s1_21_6 = _mm_srai_epi32(s1_21_4, DCT_CONST_BITS);
+          const __m128i s1_21_7 = _mm_srai_epi32(s1_21_5, DCT_CONST_BITS);
+          const __m128i s1_26_6 = _mm_srai_epi32(s1_26_4, DCT_CONST_BITS);
+          const __m128i s1_26_7 = _mm_srai_epi32(s1_26_5, DCT_CONST_BITS);
+          const __m128i s1_27_6 = _mm_srai_epi32(s1_27_4, DCT_CONST_BITS);
+          const __m128i s1_27_7 = _mm_srai_epi32(s1_27_5, DCT_CONST_BITS);
+          const __m128i s1_28_6 = _mm_srai_epi32(s1_28_4, DCT_CONST_BITS);
+          const __m128i s1_28_7 = _mm_srai_epi32(s1_28_5, DCT_CONST_BITS);
+          const __m128i s1_29_6 = _mm_srai_epi32(s1_29_4, DCT_CONST_BITS);
+          const __m128i s1_29_7 = _mm_srai_epi32(s1_29_5, DCT_CONST_BITS);
+          // Combine
+          step1[18] = _mm_packs_epi32(s1_18_6, s1_18_7);
+          step1[19] = _mm_packs_epi32(s1_19_6, s1_19_7);
+          step1[20] = _mm_packs_epi32(s1_20_6, s1_20_7);
+          step1[21] = _mm_packs_epi32(s1_21_6, s1_21_7);
+          step1[26] = _mm_packs_epi32(s1_26_6, s1_26_7);
+          step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7);
+          step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7);
+          step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[18], &step1[19], &step1[20],
+                                             &step1[21], &step1[26], &step1[27],
+                                             &step1[28], &step1[29]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // Stage 5
+        {
+          step2[4] = ADD_EPI16(step1[5], step3[4]);
+          step2[5] = SUB_EPI16(step3[4], step1[5]);
+          step2[6] = SUB_EPI16(step3[7], step1[6]);
+          step2[7] = ADD_EPI16(step1[6], step3[7]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&step2[4], &step2[5], &step2[6],
+                                             &step2[7]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]);
+          const __m128i out_00_1 = _mm_unpackhi_epi16(step1[0], step1[1]);
+          const __m128i out_08_0 = _mm_unpacklo_epi16(step1[2], step1[3]);
+          const __m128i out_08_1 = _mm_unpackhi_epi16(step1[2], step1[3]);
+          const __m128i out_00_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_p16);
+          const __m128i out_00_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_p16);
+          const __m128i out_16_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_m16);
+          const __m128i out_16_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_m16);
+          const __m128i out_08_2 = _mm_madd_epi16(out_08_0, k__cospi_p24_p08);
+          const __m128i out_08_3 = _mm_madd_epi16(out_08_1, k__cospi_p24_p08);
+          const __m128i out_24_2 = _mm_madd_epi16(out_08_0, k__cospi_m08_p24);
+          const __m128i out_24_3 = _mm_madd_epi16(out_08_1, k__cospi_m08_p24);
+          // dct_const_round_shift
+          const __m128i out_00_4 =
+              _mm_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_00_5 =
+              _mm_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_16_4 =
+              _mm_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_16_5 =
+              _mm_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_08_4 =
+              _mm_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_08_5 =
+              _mm_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_24_4 =
+              _mm_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_24_5 =
+              _mm_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_00_6 = _mm_srai_epi32(out_00_4, DCT_CONST_BITS);
+          const __m128i out_00_7 = _mm_srai_epi32(out_00_5, DCT_CONST_BITS);
+          const __m128i out_16_6 = _mm_srai_epi32(out_16_4, DCT_CONST_BITS);
+          const __m128i out_16_7 = _mm_srai_epi32(out_16_5, DCT_CONST_BITS);
+          const __m128i out_08_6 = _mm_srai_epi32(out_08_4, DCT_CONST_BITS);
+          const __m128i out_08_7 = _mm_srai_epi32(out_08_5, DCT_CONST_BITS);
+          const __m128i out_24_6 = _mm_srai_epi32(out_24_4, DCT_CONST_BITS);
+          const __m128i out_24_7 = _mm_srai_epi32(out_24_5, DCT_CONST_BITS);
+          // Combine
+          out[0] = _mm_packs_epi32(out_00_6, out_00_7);
+          out[16] = _mm_packs_epi32(out_16_6, out_16_7);
+          out[8] = _mm_packs_epi32(out_08_6, out_08_7);
+          out[24] = _mm_packs_epi32(out_24_6, out_24_7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x4(&out[0], &out[16], &out[8], &out[24]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[9], step1[14]);
+          const __m128i s2_09_1 = _mm_unpackhi_epi16(step1[9], step1[14]);
+          const __m128i s2_10_0 = _mm_unpacklo_epi16(step1[10], step1[13]);
+          const __m128i s2_10_1 = _mm_unpackhi_epi16(step1[10], step1[13]);
+          const __m128i s2_09_2 = _mm_madd_epi16(s2_09_0, k__cospi_m08_p24);
+          const __m128i s2_09_3 = _mm_madd_epi16(s2_09_1, k__cospi_m08_p24);
+          const __m128i s2_10_2 = _mm_madd_epi16(s2_10_0, k__cospi_m24_m08);
+          const __m128i s2_10_3 = _mm_madd_epi16(s2_10_1, k__cospi_m24_m08);
+          const __m128i s2_13_2 = _mm_madd_epi16(s2_10_0, k__cospi_m08_p24);
+          const __m128i s2_13_3 = _mm_madd_epi16(s2_10_1, k__cospi_m08_p24);
+          const __m128i s2_14_2 = _mm_madd_epi16(s2_09_0, k__cospi_p24_p08);
+          const __m128i s2_14_3 = _mm_madd_epi16(s2_09_1, k__cospi_p24_p08);
+          // dct_const_round_shift
+          const __m128i s2_09_4 = _mm_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
+          const __m128i s2_09_5 = _mm_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
+          const __m128i s2_10_4 = _mm_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
+          const __m128i s2_10_5 = _mm_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
+          const __m128i s2_13_4 = _mm_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
+          const __m128i s2_13_5 = _mm_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
+          const __m128i s2_14_4 = _mm_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
+          const __m128i s2_14_5 = _mm_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
+          const __m128i s2_09_6 = _mm_srai_epi32(s2_09_4, DCT_CONST_BITS);
+          const __m128i s2_09_7 = _mm_srai_epi32(s2_09_5, DCT_CONST_BITS);
+          const __m128i s2_10_6 = _mm_srai_epi32(s2_10_4, DCT_CONST_BITS);
+          const __m128i s2_10_7 = _mm_srai_epi32(s2_10_5, DCT_CONST_BITS);
+          const __m128i s2_13_6 = _mm_srai_epi32(s2_13_4, DCT_CONST_BITS);
+          const __m128i s2_13_7 = _mm_srai_epi32(s2_13_5, DCT_CONST_BITS);
+          const __m128i s2_14_6 = _mm_srai_epi32(s2_14_4, DCT_CONST_BITS);
+          const __m128i s2_14_7 = _mm_srai_epi32(s2_14_5, DCT_CONST_BITS);
+          // Combine
+          step2[9] = _mm_packs_epi32(s2_09_6, s2_09_7);
+          step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7);
+          step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7);
+          step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&step2[9], &step2[10], &step2[13],
+                                             &step2[14]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          step2[16] = ADD_EPI16(step1[19], step3[16]);
+          step2[17] = ADD_EPI16(step1[18], step3[17]);
+          step2[18] = SUB_EPI16(step3[17], step1[18]);
+          step2[19] = SUB_EPI16(step3[16], step1[19]);
+          step2[20] = SUB_EPI16(step3[23], step1[20]);
+          step2[21] = SUB_EPI16(step3[22], step1[21]);
+          step2[22] = ADD_EPI16(step1[21], step3[22]);
+          step2[23] = ADD_EPI16(step1[20], step3[23]);
+          step2[24] = ADD_EPI16(step1[27], step3[24]);
+          step2[25] = ADD_EPI16(step1[26], step3[25]);
+          step2[26] = SUB_EPI16(step3[25], step1[26]);
+          step2[27] = SUB_EPI16(step3[24], step1[27]);
+          step2[28] = SUB_EPI16(step3[31], step1[28]);
+          step2[29] = SUB_EPI16(step3[30], step1[29]);
+          step2[30] = ADD_EPI16(step1[29], step3[30]);
+          step2[31] = ADD_EPI16(step1[28], step3[31]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x16(
+              &step2[16], &step2[17], &step2[18], &step2[19], &step2[20],
+              &step2[21], &step2[22], &step2[23], &step2[24], &step2[25],
+              &step2[26], &step2[27], &step2[28], &step2[29], &step2[30],
+              &step2[31]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // Stage 6
+        {
+          const __m128i out_04_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
+          const __m128i out_04_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
+          const __m128i out_20_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
+          const __m128i out_20_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
+          const __m128i out_12_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
+          const __m128i out_12_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
+          const __m128i out_28_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
+          const __m128i out_28_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
+          const __m128i out_04_2 = _mm_madd_epi16(out_04_0, k__cospi_p28_p04);
+          const __m128i out_04_3 = _mm_madd_epi16(out_04_1, k__cospi_p28_p04);
+          const __m128i out_20_2 = _mm_madd_epi16(out_20_0, k__cospi_p12_p20);
+          const __m128i out_20_3 = _mm_madd_epi16(out_20_1, k__cospi_p12_p20);
+          const __m128i out_12_2 = _mm_madd_epi16(out_12_0, k__cospi_m20_p12);
+          const __m128i out_12_3 = _mm_madd_epi16(out_12_1, k__cospi_m20_p12);
+          const __m128i out_28_2 = _mm_madd_epi16(out_28_0, k__cospi_m04_p28);
+          const __m128i out_28_3 = _mm_madd_epi16(out_28_1, k__cospi_m04_p28);
+          // dct_const_round_shift
+          const __m128i out_04_4 =
+              _mm_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_04_5 =
+              _mm_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_20_4 =
+              _mm_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_20_5 =
+              _mm_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_12_4 =
+              _mm_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_12_5 =
+              _mm_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_28_4 =
+              _mm_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_28_5 =
+              _mm_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_04_6 = _mm_srai_epi32(out_04_4, DCT_CONST_BITS);
+          const __m128i out_04_7 = _mm_srai_epi32(out_04_5, DCT_CONST_BITS);
+          const __m128i out_20_6 = _mm_srai_epi32(out_20_4, DCT_CONST_BITS);
+          const __m128i out_20_7 = _mm_srai_epi32(out_20_5, DCT_CONST_BITS);
+          const __m128i out_12_6 = _mm_srai_epi32(out_12_4, DCT_CONST_BITS);
+          const __m128i out_12_7 = _mm_srai_epi32(out_12_5, DCT_CONST_BITS);
+          const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS);
+          const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS);
+          // Combine
+          out[4] = _mm_packs_epi32(out_04_6, out_04_7);
+          out[20] = _mm_packs_epi32(out_20_6, out_20_7);
+          out[12] = _mm_packs_epi32(out_12_6, out_12_7);
+          out[28] = _mm_packs_epi32(out_28_6, out_28_7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x4(&out[4], &out[20], &out[12], &out[28]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          step3[8] = ADD_EPI16(step2[9], step1[8]);
+          step3[9] = SUB_EPI16(step1[8], step2[9]);
+          step3[10] = SUB_EPI16(step1[11], step2[10]);
+          step3[11] = ADD_EPI16(step2[10], step1[11]);
+          step3[12] = ADD_EPI16(step2[13], step1[12]);
+          step3[13] = SUB_EPI16(step1[12], step2[13]);
+          step3[14] = SUB_EPI16(step1[15], step2[14]);
+          step3[15] = ADD_EPI16(step2[14], step1[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step3[8], &step3[9], &step3[10],
+                                             &step3[11], &step3[12], &step3[13],
+                                             &step3[14], &step3[15]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]);
+          const __m128i s3_17_1 = _mm_unpackhi_epi16(step2[17], step2[30]);
+          const __m128i s3_18_0 = _mm_unpacklo_epi16(step2[18], step2[29]);
+          const __m128i s3_18_1 = _mm_unpackhi_epi16(step2[18], step2[29]);
+          const __m128i s3_21_0 = _mm_unpacklo_epi16(step2[21], step2[26]);
+          const __m128i s3_21_1 = _mm_unpackhi_epi16(step2[21], step2[26]);
+          const __m128i s3_22_0 = _mm_unpacklo_epi16(step2[22], step2[25]);
+          const __m128i s3_22_1 = _mm_unpackhi_epi16(step2[22], step2[25]);
+          const __m128i s3_17_2 = _mm_madd_epi16(s3_17_0, k__cospi_m04_p28);
+          const __m128i s3_17_3 = _mm_madd_epi16(s3_17_1, k__cospi_m04_p28);
+          const __m128i s3_18_2 = _mm_madd_epi16(s3_18_0, k__cospi_m28_m04);
+          const __m128i s3_18_3 = _mm_madd_epi16(s3_18_1, k__cospi_m28_m04);
+          const __m128i s3_21_2 = _mm_madd_epi16(s3_21_0, k__cospi_m20_p12);
+          const __m128i s3_21_3 = _mm_madd_epi16(s3_21_1, k__cospi_m20_p12);
+          const __m128i s3_22_2 = _mm_madd_epi16(s3_22_0, k__cospi_m12_m20);
+          const __m128i s3_22_3 = _mm_madd_epi16(s3_22_1, k__cospi_m12_m20);
+          const __m128i s3_25_2 = _mm_madd_epi16(s3_22_0, k__cospi_m20_p12);
+          const __m128i s3_25_3 = _mm_madd_epi16(s3_22_1, k__cospi_m20_p12);
+          const __m128i s3_26_2 = _mm_madd_epi16(s3_21_0, k__cospi_p12_p20);
+          const __m128i s3_26_3 = _mm_madd_epi16(s3_21_1, k__cospi_p12_p20);
+          const __m128i s3_29_2 = _mm_madd_epi16(s3_18_0, k__cospi_m04_p28);
+          const __m128i s3_29_3 = _mm_madd_epi16(s3_18_1, k__cospi_m04_p28);
+          const __m128i s3_30_2 = _mm_madd_epi16(s3_17_0, k__cospi_p28_p04);
+          const __m128i s3_30_3 = _mm_madd_epi16(s3_17_1, k__cospi_p28_p04);
+          // dct_const_round_shift
+          const __m128i s3_17_4 = _mm_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_17_5 = _mm_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_18_4 = _mm_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_18_5 = _mm_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_21_4 = _mm_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_21_5 = _mm_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_22_4 = _mm_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_22_5 = _mm_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_17_6 = _mm_srai_epi32(s3_17_4, DCT_CONST_BITS);
+          const __m128i s3_17_7 = _mm_srai_epi32(s3_17_5, DCT_CONST_BITS);
+          const __m128i s3_18_6 = _mm_srai_epi32(s3_18_4, DCT_CONST_BITS);
+          const __m128i s3_18_7 = _mm_srai_epi32(s3_18_5, DCT_CONST_BITS);
+          const __m128i s3_21_6 = _mm_srai_epi32(s3_21_4, DCT_CONST_BITS);
+          const __m128i s3_21_7 = _mm_srai_epi32(s3_21_5, DCT_CONST_BITS);
+          const __m128i s3_22_6 = _mm_srai_epi32(s3_22_4, DCT_CONST_BITS);
+          const __m128i s3_22_7 = _mm_srai_epi32(s3_22_5, DCT_CONST_BITS);
+          const __m128i s3_25_4 = _mm_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_25_5 = _mm_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_26_4 = _mm_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_26_5 = _mm_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_29_4 = _mm_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_29_5 = _mm_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_30_4 = _mm_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_30_5 = _mm_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_25_6 = _mm_srai_epi32(s3_25_4, DCT_CONST_BITS);
+          const __m128i s3_25_7 = _mm_srai_epi32(s3_25_5, DCT_CONST_BITS);
+          const __m128i s3_26_6 = _mm_srai_epi32(s3_26_4, DCT_CONST_BITS);
+          const __m128i s3_26_7 = _mm_srai_epi32(s3_26_5, DCT_CONST_BITS);
+          const __m128i s3_29_6 = _mm_srai_epi32(s3_29_4, DCT_CONST_BITS);
+          const __m128i s3_29_7 = _mm_srai_epi32(s3_29_5, DCT_CONST_BITS);
+          const __m128i s3_30_6 = _mm_srai_epi32(s3_30_4, DCT_CONST_BITS);
+          const __m128i s3_30_7 = _mm_srai_epi32(s3_30_5, DCT_CONST_BITS);
+          // Combine
+          step3[17] = _mm_packs_epi32(s3_17_6, s3_17_7);
+          step3[18] = _mm_packs_epi32(s3_18_6, s3_18_7);
+          step3[21] = _mm_packs_epi32(s3_21_6, s3_21_7);
+          step3[22] = _mm_packs_epi32(s3_22_6, s3_22_7);
+          // Combine
+          step3[25] = _mm_packs_epi32(s3_25_6, s3_25_7);
+          step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7);
+          step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7);
+          step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step3[17], &step3[18], &step3[21],
+                                             &step3[22], &step3[25], &step3[26],
+                                             &step3[29], &step3[30]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // Stage 7
+        {
+          const __m128i out_02_0 = _mm_unpacklo_epi16(step3[8], step3[15]);
+          const __m128i out_02_1 = _mm_unpackhi_epi16(step3[8], step3[15]);
+          const __m128i out_18_0 = _mm_unpacklo_epi16(step3[9], step3[14]);
+          const __m128i out_18_1 = _mm_unpackhi_epi16(step3[9], step3[14]);
+          const __m128i out_10_0 = _mm_unpacklo_epi16(step3[10], step3[13]);
+          const __m128i out_10_1 = _mm_unpackhi_epi16(step3[10], step3[13]);
+          const __m128i out_26_0 = _mm_unpacklo_epi16(step3[11], step3[12]);
+          const __m128i out_26_1 = _mm_unpackhi_epi16(step3[11], step3[12]);
+          const __m128i out_02_2 = _mm_madd_epi16(out_02_0, k__cospi_p30_p02);
+          const __m128i out_02_3 = _mm_madd_epi16(out_02_1, k__cospi_p30_p02);
+          const __m128i out_18_2 = _mm_madd_epi16(out_18_0, k__cospi_p14_p18);
+          const __m128i out_18_3 = _mm_madd_epi16(out_18_1, k__cospi_p14_p18);
+          const __m128i out_10_2 = _mm_madd_epi16(out_10_0, k__cospi_p22_p10);
+          const __m128i out_10_3 = _mm_madd_epi16(out_10_1, k__cospi_p22_p10);
+          const __m128i out_26_2 = _mm_madd_epi16(out_26_0, k__cospi_p06_p26);
+          const __m128i out_26_3 = _mm_madd_epi16(out_26_1, k__cospi_p06_p26);
+          const __m128i out_06_2 = _mm_madd_epi16(out_26_0, k__cospi_m26_p06);
+          const __m128i out_06_3 = _mm_madd_epi16(out_26_1, k__cospi_m26_p06);
+          const __m128i out_22_2 = _mm_madd_epi16(out_10_0, k__cospi_m10_p22);
+          const __m128i out_22_3 = _mm_madd_epi16(out_10_1, k__cospi_m10_p22);
+          const __m128i out_14_2 = _mm_madd_epi16(out_18_0, k__cospi_m18_p14);
+          const __m128i out_14_3 = _mm_madd_epi16(out_18_1, k__cospi_m18_p14);
+          const __m128i out_30_2 = _mm_madd_epi16(out_02_0, k__cospi_m02_p30);
+          const __m128i out_30_3 = _mm_madd_epi16(out_02_1, k__cospi_m02_p30);
+          // dct_const_round_shift
+          const __m128i out_02_4 =
+              _mm_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_02_5 =
+              _mm_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_18_4 =
+              _mm_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_18_5 =
+              _mm_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_10_4 =
+              _mm_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_10_5 =
+              _mm_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_26_4 =
+              _mm_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_26_5 =
+              _mm_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_06_4 =
+              _mm_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_06_5 =
+              _mm_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_22_4 =
+              _mm_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_22_5 =
+              _mm_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_14_4 =
+              _mm_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_14_5 =
+              _mm_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_30_4 =
+              _mm_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_30_5 =
+              _mm_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_02_6 = _mm_srai_epi32(out_02_4, DCT_CONST_BITS);
+          const __m128i out_02_7 = _mm_srai_epi32(out_02_5, DCT_CONST_BITS);
+          const __m128i out_18_6 = _mm_srai_epi32(out_18_4, DCT_CONST_BITS);
+          const __m128i out_18_7 = _mm_srai_epi32(out_18_5, DCT_CONST_BITS);
+          const __m128i out_10_6 = _mm_srai_epi32(out_10_4, DCT_CONST_BITS);
+          const __m128i out_10_7 = _mm_srai_epi32(out_10_5, DCT_CONST_BITS);
+          const __m128i out_26_6 = _mm_srai_epi32(out_26_4, DCT_CONST_BITS);
+          const __m128i out_26_7 = _mm_srai_epi32(out_26_5, DCT_CONST_BITS);
+          const __m128i out_06_6 = _mm_srai_epi32(out_06_4, DCT_CONST_BITS);
+          const __m128i out_06_7 = _mm_srai_epi32(out_06_5, DCT_CONST_BITS);
+          const __m128i out_22_6 = _mm_srai_epi32(out_22_4, DCT_CONST_BITS);
+          const __m128i out_22_7 = _mm_srai_epi32(out_22_5, DCT_CONST_BITS);
+          const __m128i out_14_6 = _mm_srai_epi32(out_14_4, DCT_CONST_BITS);
+          const __m128i out_14_7 = _mm_srai_epi32(out_14_5, DCT_CONST_BITS);
+          const __m128i out_30_6 = _mm_srai_epi32(out_30_4, DCT_CONST_BITS);
+          const __m128i out_30_7 = _mm_srai_epi32(out_30_5, DCT_CONST_BITS);
+          // Combine
+          out[2] = _mm_packs_epi32(out_02_6, out_02_7);
+          out[18] = _mm_packs_epi32(out_18_6, out_18_7);
+          out[10] = _mm_packs_epi32(out_10_6, out_10_7);
+          out[26] = _mm_packs_epi32(out_26_6, out_26_7);
+          out[6] = _mm_packs_epi32(out_06_6, out_06_7);
+          out[22] = _mm_packs_epi32(out_22_6, out_22_7);
+          out[14] = _mm_packs_epi32(out_14_6, out_14_7);
+          out[30] = _mm_packs_epi32(out_30_6, out_30_7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x8(&out[2], &out[18], &out[10], &out[26],
+                                      &out[6], &out[22], &out[14], &out[30]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          step1[16] = ADD_EPI16(step3[17], step2[16]);
+          step1[17] = SUB_EPI16(step2[16], step3[17]);
+          step1[18] = SUB_EPI16(step2[19], step3[18]);
+          step1[19] = ADD_EPI16(step3[18], step2[19]);
+          step1[20] = ADD_EPI16(step3[21], step2[20]);
+          step1[21] = SUB_EPI16(step2[20], step3[21]);
+          step1[22] = SUB_EPI16(step2[23], step3[22]);
+          step1[23] = ADD_EPI16(step3[22], step2[23]);
+          step1[24] = ADD_EPI16(step3[25], step2[24]);
+          step1[25] = SUB_EPI16(step2[24], step3[25]);
+          step1[26] = SUB_EPI16(step2[27], step3[26]);
+          step1[27] = ADD_EPI16(step3[26], step2[27]);
+          step1[28] = ADD_EPI16(step3[29], step2[28]);
+          step1[29] = SUB_EPI16(step2[28], step3[29]);
+          step1[30] = SUB_EPI16(step2[31], step3[30]);
+          step1[31] = ADD_EPI16(step3[30], step2[31]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x16(
+              &step1[16], &step1[17], &step1[18], &step1[19], &step1[20],
+              &step1[21], &step1[22], &step1[23], &step1[24], &step1[25],
+              &step1[26], &step1[27], &step1[28], &step1[29], &step1[30],
+              &step1[31]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // Final stage --- outputs indices are bit-reversed.
+        {
+          const __m128i out_01_0 = _mm_unpacklo_epi16(step1[16], step1[31]);
+          const __m128i out_01_1 = _mm_unpackhi_epi16(step1[16], step1[31]);
+          const __m128i out_17_0 = _mm_unpacklo_epi16(step1[17], step1[30]);
+          const __m128i out_17_1 = _mm_unpackhi_epi16(step1[17], step1[30]);
+          const __m128i out_09_0 = _mm_unpacklo_epi16(step1[18], step1[29]);
+          const __m128i out_09_1 = _mm_unpackhi_epi16(step1[18], step1[29]);
+          const __m128i out_25_0 = _mm_unpacklo_epi16(step1[19], step1[28]);
+          const __m128i out_25_1 = _mm_unpackhi_epi16(step1[19], step1[28]);
+          const __m128i out_01_2 = _mm_madd_epi16(out_01_0, k__cospi_p31_p01);
+          const __m128i out_01_3 = _mm_madd_epi16(out_01_1, k__cospi_p31_p01);
+          const __m128i out_17_2 = _mm_madd_epi16(out_17_0, k__cospi_p15_p17);
+          const __m128i out_17_3 = _mm_madd_epi16(out_17_1, k__cospi_p15_p17);
+          const __m128i out_09_2 = _mm_madd_epi16(out_09_0, k__cospi_p23_p09);
+          const __m128i out_09_3 = _mm_madd_epi16(out_09_1, k__cospi_p23_p09);
+          const __m128i out_25_2 = _mm_madd_epi16(out_25_0, k__cospi_p07_p25);
+          const __m128i out_25_3 = _mm_madd_epi16(out_25_1, k__cospi_p07_p25);
+          const __m128i out_07_2 = _mm_madd_epi16(out_25_0, k__cospi_m25_p07);
+          const __m128i out_07_3 = _mm_madd_epi16(out_25_1, k__cospi_m25_p07);
+          const __m128i out_23_2 = _mm_madd_epi16(out_09_0, k__cospi_m09_p23);
+          const __m128i out_23_3 = _mm_madd_epi16(out_09_1, k__cospi_m09_p23);
+          const __m128i out_15_2 = _mm_madd_epi16(out_17_0, k__cospi_m17_p15);
+          const __m128i out_15_3 = _mm_madd_epi16(out_17_1, k__cospi_m17_p15);
+          const __m128i out_31_2 = _mm_madd_epi16(out_01_0, k__cospi_m01_p31);
+          const __m128i out_31_3 = _mm_madd_epi16(out_01_1, k__cospi_m01_p31);
+          // dct_const_round_shift
+          const __m128i out_01_4 =
+              _mm_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_01_5 =
+              _mm_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_17_4 =
+              _mm_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_17_5 =
+              _mm_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_09_4 =
+              _mm_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_09_5 =
+              _mm_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_25_4 =
+              _mm_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_25_5 =
+              _mm_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_07_4 =
+              _mm_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_07_5 =
+              _mm_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_23_4 =
+              _mm_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_23_5 =
+              _mm_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_15_4 =
+              _mm_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_15_5 =
+              _mm_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_31_4 =
+              _mm_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_31_5 =
+              _mm_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_01_6 = _mm_srai_epi32(out_01_4, DCT_CONST_BITS);
+          const __m128i out_01_7 = _mm_srai_epi32(out_01_5, DCT_CONST_BITS);
+          const __m128i out_17_6 = _mm_srai_epi32(out_17_4, DCT_CONST_BITS);
+          const __m128i out_17_7 = _mm_srai_epi32(out_17_5, DCT_CONST_BITS);
+          const __m128i out_09_6 = _mm_srai_epi32(out_09_4, DCT_CONST_BITS);
+          const __m128i out_09_7 = _mm_srai_epi32(out_09_5, DCT_CONST_BITS);
+          const __m128i out_25_6 = _mm_srai_epi32(out_25_4, DCT_CONST_BITS);
+          const __m128i out_25_7 = _mm_srai_epi32(out_25_5, DCT_CONST_BITS);
+          const __m128i out_07_6 = _mm_srai_epi32(out_07_4, DCT_CONST_BITS);
+          const __m128i out_07_7 = _mm_srai_epi32(out_07_5, DCT_CONST_BITS);
+          const __m128i out_23_6 = _mm_srai_epi32(out_23_4, DCT_CONST_BITS);
+          const __m128i out_23_7 = _mm_srai_epi32(out_23_5, DCT_CONST_BITS);
+          const __m128i out_15_6 = _mm_srai_epi32(out_15_4, DCT_CONST_BITS);
+          const __m128i out_15_7 = _mm_srai_epi32(out_15_5, DCT_CONST_BITS);
+          const __m128i out_31_6 = _mm_srai_epi32(out_31_4, DCT_CONST_BITS);
+          const __m128i out_31_7 = _mm_srai_epi32(out_31_5, DCT_CONST_BITS);
+          // Combine
+          out[1] = _mm_packs_epi32(out_01_6, out_01_7);
+          out[17] = _mm_packs_epi32(out_17_6, out_17_7);
+          out[9] = _mm_packs_epi32(out_09_6, out_09_7);
+          out[25] = _mm_packs_epi32(out_25_6, out_25_7);
+          out[7] = _mm_packs_epi32(out_07_6, out_07_7);
+          out[23] = _mm_packs_epi32(out_23_6, out_23_7);
+          out[15] = _mm_packs_epi32(out_15_6, out_15_7);
+          out[31] = _mm_packs_epi32(out_31_6, out_31_7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x8(&out[1], &out[17], &out[9], &out[25],
+                                      &out[7], &out[23], &out[15], &out[31]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]);
+          const __m128i out_05_1 = _mm_unpackhi_epi16(step1[20], step1[27]);
+          const __m128i out_21_0 = _mm_unpacklo_epi16(step1[21], step1[26]);
+          const __m128i out_21_1 = _mm_unpackhi_epi16(step1[21], step1[26]);
+          const __m128i out_13_0 = _mm_unpacklo_epi16(step1[22], step1[25]);
+          const __m128i out_13_1 = _mm_unpackhi_epi16(step1[22], step1[25]);
+          const __m128i out_29_0 = _mm_unpacklo_epi16(step1[23], step1[24]);
+          const __m128i out_29_1 = _mm_unpackhi_epi16(step1[23], step1[24]);
+          const __m128i out_05_2 = _mm_madd_epi16(out_05_0, k__cospi_p27_p05);
+          const __m128i out_05_3 = _mm_madd_epi16(out_05_1, k__cospi_p27_p05);
+          const __m128i out_21_2 = _mm_madd_epi16(out_21_0, k__cospi_p11_p21);
+          const __m128i out_21_3 = _mm_madd_epi16(out_21_1, k__cospi_p11_p21);
+          const __m128i out_13_2 = _mm_madd_epi16(out_13_0, k__cospi_p19_p13);
+          const __m128i out_13_3 = _mm_madd_epi16(out_13_1, k__cospi_p19_p13);
+          const __m128i out_29_2 = _mm_madd_epi16(out_29_0, k__cospi_p03_p29);
+          const __m128i out_29_3 = _mm_madd_epi16(out_29_1, k__cospi_p03_p29);
+          const __m128i out_03_2 = _mm_madd_epi16(out_29_0, k__cospi_m29_p03);
+          const __m128i out_03_3 = _mm_madd_epi16(out_29_1, k__cospi_m29_p03);
+          const __m128i out_19_2 = _mm_madd_epi16(out_13_0, k__cospi_m13_p19);
+          const __m128i out_19_3 = _mm_madd_epi16(out_13_1, k__cospi_m13_p19);
+          const __m128i out_11_2 = _mm_madd_epi16(out_21_0, k__cospi_m21_p11);
+          const __m128i out_11_3 = _mm_madd_epi16(out_21_1, k__cospi_m21_p11);
+          const __m128i out_27_2 = _mm_madd_epi16(out_05_0, k__cospi_m05_p27);
+          const __m128i out_27_3 = _mm_madd_epi16(out_05_1, k__cospi_m05_p27);
+          // dct_const_round_shift
+          const __m128i out_05_4 =
+              _mm_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_05_5 =
+              _mm_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_21_4 =
+              _mm_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_21_5 =
+              _mm_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_13_4 =
+              _mm_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_13_5 =
+              _mm_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_29_4 =
+              _mm_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_29_5 =
+              _mm_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_03_4 =
+              _mm_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_03_5 =
+              _mm_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_19_4 =
+              _mm_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_19_5 =
+              _mm_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_11_4 =
+              _mm_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_11_5 =
+              _mm_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_27_4 =
+              _mm_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
+          const __m128i out_27_5 =
+              _mm_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
+          const __m128i out_05_6 = _mm_srai_epi32(out_05_4, DCT_CONST_BITS);
+          const __m128i out_05_7 = _mm_srai_epi32(out_05_5, DCT_CONST_BITS);
+          const __m128i out_21_6 = _mm_srai_epi32(out_21_4, DCT_CONST_BITS);
+          const __m128i out_21_7 = _mm_srai_epi32(out_21_5, DCT_CONST_BITS);
+          const __m128i out_13_6 = _mm_srai_epi32(out_13_4, DCT_CONST_BITS);
+          const __m128i out_13_7 = _mm_srai_epi32(out_13_5, DCT_CONST_BITS);
+          const __m128i out_29_6 = _mm_srai_epi32(out_29_4, DCT_CONST_BITS);
+          const __m128i out_29_7 = _mm_srai_epi32(out_29_5, DCT_CONST_BITS);
+          const __m128i out_03_6 = _mm_srai_epi32(out_03_4, DCT_CONST_BITS);
+          const __m128i out_03_7 = _mm_srai_epi32(out_03_5, DCT_CONST_BITS);
+          const __m128i out_19_6 = _mm_srai_epi32(out_19_4, DCT_CONST_BITS);
+          const __m128i out_19_7 = _mm_srai_epi32(out_19_5, DCT_CONST_BITS);
+          const __m128i out_11_6 = _mm_srai_epi32(out_11_4, DCT_CONST_BITS);
+          const __m128i out_11_7 = _mm_srai_epi32(out_11_5, DCT_CONST_BITS);
+          const __m128i out_27_6 = _mm_srai_epi32(out_27_4, DCT_CONST_BITS);
+          const __m128i out_27_7 = _mm_srai_epi32(out_27_5, DCT_CONST_BITS);
+          // Combine
+          out[5] = _mm_packs_epi32(out_05_6, out_05_7);
+          out[21] = _mm_packs_epi32(out_21_6, out_21_7);
+          out[13] = _mm_packs_epi32(out_13_6, out_13_7);
+          out[29] = _mm_packs_epi32(out_29_6, out_29_7);
+          out[3] = _mm_packs_epi32(out_03_6, out_03_7);
+          out[19] = _mm_packs_epi32(out_19_6, out_19_7);
+          out[11] = _mm_packs_epi32(out_11_6, out_11_7);
+          out[27] = _mm_packs_epi32(out_27_6, out_27_7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x8(&out[5], &out[21], &out[13], &out[29],
+                                      &out[3], &out[19], &out[11], &out[27]);
+          if (overflow) {
+            if (pass == 0)
+              HIGH_FDCT32x32_2D_C(input, output_org, stride);
+            else
+              HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+#if FDCT32x32_HIGH_PRECISION
+      } else {
+        __m128i lstep1[64], lstep2[64], lstep3[64];
+        __m128i u[32], v[32], sign[16];
+        const __m128i K32One = _mm_set_epi32(1, 1, 1, 1);
+        // start using 32-bit operations
+        // stage 3
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[0] = _mm_unpacklo_epi16(step2[0], kZero);
+          lstep2[1] = _mm_unpackhi_epi16(step2[0], kZero);
+          lstep2[2] = _mm_unpacklo_epi16(step2[1], kZero);
+          lstep2[3] = _mm_unpackhi_epi16(step2[1], kZero);
+          lstep2[4] = _mm_unpacklo_epi16(step2[2], kZero);
+          lstep2[5] = _mm_unpackhi_epi16(step2[2], kZero);
+          lstep2[6] = _mm_unpacklo_epi16(step2[3], kZero);
+          lstep2[7] = _mm_unpackhi_epi16(step2[3], kZero);
+          lstep2[8] = _mm_unpacklo_epi16(step2[4], kZero);
+          lstep2[9] = _mm_unpackhi_epi16(step2[4], kZero);
+          lstep2[10] = _mm_unpacklo_epi16(step2[5], kZero);
+          lstep2[11] = _mm_unpackhi_epi16(step2[5], kZero);
+          lstep2[12] = _mm_unpacklo_epi16(step2[6], kZero);
+          lstep2[13] = _mm_unpackhi_epi16(step2[6], kZero);
+          lstep2[14] = _mm_unpacklo_epi16(step2[7], kZero);
+          lstep2[15] = _mm_unpackhi_epi16(step2[7], kZero);
+          lstep2[0] = _mm_madd_epi16(lstep2[0], kOne);
+          lstep2[1] = _mm_madd_epi16(lstep2[1], kOne);
+          lstep2[2] = _mm_madd_epi16(lstep2[2], kOne);
+          lstep2[3] = _mm_madd_epi16(lstep2[3], kOne);
+          lstep2[4] = _mm_madd_epi16(lstep2[4], kOne);
+          lstep2[5] = _mm_madd_epi16(lstep2[5], kOne);
+          lstep2[6] = _mm_madd_epi16(lstep2[6], kOne);
+          lstep2[7] = _mm_madd_epi16(lstep2[7], kOne);
+          lstep2[8] = _mm_madd_epi16(lstep2[8], kOne);
+          lstep2[9] = _mm_madd_epi16(lstep2[9], kOne);
+          lstep2[10] = _mm_madd_epi16(lstep2[10], kOne);
+          lstep2[11] = _mm_madd_epi16(lstep2[11], kOne);
+          lstep2[12] = _mm_madd_epi16(lstep2[12], kOne);
+          lstep2[13] = _mm_madd_epi16(lstep2[13], kOne);
+          lstep2[14] = _mm_madd_epi16(lstep2[14], kOne);
+          lstep2[15] = _mm_madd_epi16(lstep2[15], kOne);
+
+          lstep3[0] = _mm_add_epi32(lstep2[14], lstep2[0]);
+          lstep3[1] = _mm_add_epi32(lstep2[15], lstep2[1]);
+          lstep3[2] = _mm_add_epi32(lstep2[12], lstep2[2]);
+          lstep3[3] = _mm_add_epi32(lstep2[13], lstep2[3]);
+          lstep3[4] = _mm_add_epi32(lstep2[10], lstep2[4]);
+          lstep3[5] = _mm_add_epi32(lstep2[11], lstep2[5]);
+          lstep3[6] = _mm_add_epi32(lstep2[8], lstep2[6]);
+          lstep3[7] = _mm_add_epi32(lstep2[9], lstep2[7]);
+          lstep3[8] = _mm_sub_epi32(lstep2[6], lstep2[8]);
+          lstep3[9] = _mm_sub_epi32(lstep2[7], lstep2[9]);
+          lstep3[10] = _mm_sub_epi32(lstep2[4], lstep2[10]);
+          lstep3[11] = _mm_sub_epi32(lstep2[5], lstep2[11]);
+          lstep3[12] = _mm_sub_epi32(lstep2[2], lstep2[12]);
+          lstep3[13] = _mm_sub_epi32(lstep2[3], lstep2[13]);
+          lstep3[14] = _mm_sub_epi32(lstep2[0], lstep2[14]);
+          lstep3[15] = _mm_sub_epi32(lstep2[1], lstep2[15]);
+        }
+        {
+          const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
+          const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
+          const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
+          const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
+          const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
+          const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
+          const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
+          const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
+          const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
+          const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
+          const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
+          const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
+          // dct_const_round_shift
+          const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+          lstep3[20] = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
+          lstep3[21] = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
+          lstep3[22] = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
+          lstep3[23] = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
+          lstep3[24] = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
+          lstep3[25] = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
+          lstep3[26] = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
+          lstep3[27] = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
+        }
+        {
+          lstep2[40] = _mm_unpacklo_epi16(step2[20], kZero);
+          lstep2[41] = _mm_unpackhi_epi16(step2[20], kZero);
+          lstep2[42] = _mm_unpacklo_epi16(step2[21], kZero);
+          lstep2[43] = _mm_unpackhi_epi16(step2[21], kZero);
+          lstep2[44] = _mm_unpacklo_epi16(step2[22], kZero);
+          lstep2[45] = _mm_unpackhi_epi16(step2[22], kZero);
+          lstep2[46] = _mm_unpacklo_epi16(step2[23], kZero);
+          lstep2[47] = _mm_unpackhi_epi16(step2[23], kZero);
+          lstep2[48] = _mm_unpacklo_epi16(step2[24], kZero);
+          lstep2[49] = _mm_unpackhi_epi16(step2[24], kZero);
+          lstep2[50] = _mm_unpacklo_epi16(step2[25], kZero);
+          lstep2[51] = _mm_unpackhi_epi16(step2[25], kZero);
+          lstep2[52] = _mm_unpacklo_epi16(step2[26], kZero);
+          lstep2[53] = _mm_unpackhi_epi16(step2[26], kZero);
+          lstep2[54] = _mm_unpacklo_epi16(step2[27], kZero);
+          lstep2[55] = _mm_unpackhi_epi16(step2[27], kZero);
+          lstep2[40] = _mm_madd_epi16(lstep2[40], kOne);
+          lstep2[41] = _mm_madd_epi16(lstep2[41], kOne);
+          lstep2[42] = _mm_madd_epi16(lstep2[42], kOne);
+          lstep2[43] = _mm_madd_epi16(lstep2[43], kOne);
+          lstep2[44] = _mm_madd_epi16(lstep2[44], kOne);
+          lstep2[45] = _mm_madd_epi16(lstep2[45], kOne);
+          lstep2[46] = _mm_madd_epi16(lstep2[46], kOne);
+          lstep2[47] = _mm_madd_epi16(lstep2[47], kOne);
+          lstep2[48] = _mm_madd_epi16(lstep2[48], kOne);
+          lstep2[49] = _mm_madd_epi16(lstep2[49], kOne);
+          lstep2[50] = _mm_madd_epi16(lstep2[50], kOne);
+          lstep2[51] = _mm_madd_epi16(lstep2[51], kOne);
+          lstep2[52] = _mm_madd_epi16(lstep2[52], kOne);
+          lstep2[53] = _mm_madd_epi16(lstep2[53], kOne);
+          lstep2[54] = _mm_madd_epi16(lstep2[54], kOne);
+          lstep2[55] = _mm_madd_epi16(lstep2[55], kOne);
+
+          lstep1[32] = _mm_unpacklo_epi16(step1[16], kZero);
+          lstep1[33] = _mm_unpackhi_epi16(step1[16], kZero);
+          lstep1[34] = _mm_unpacklo_epi16(step1[17], kZero);
+          lstep1[35] = _mm_unpackhi_epi16(step1[17], kZero);
+          lstep1[36] = _mm_unpacklo_epi16(step1[18], kZero);
+          lstep1[37] = _mm_unpackhi_epi16(step1[18], kZero);
+          lstep1[38] = _mm_unpacklo_epi16(step1[19], kZero);
+          lstep1[39] = _mm_unpackhi_epi16(step1[19], kZero);
+          lstep1[56] = _mm_unpacklo_epi16(step1[28], kZero);
+          lstep1[57] = _mm_unpackhi_epi16(step1[28], kZero);
+          lstep1[58] = _mm_unpacklo_epi16(step1[29], kZero);
+          lstep1[59] = _mm_unpackhi_epi16(step1[29], kZero);
+          lstep1[60] = _mm_unpacklo_epi16(step1[30], kZero);
+          lstep1[61] = _mm_unpackhi_epi16(step1[30], kZero);
+          lstep1[62] = _mm_unpacklo_epi16(step1[31], kZero);
+          lstep1[63] = _mm_unpackhi_epi16(step1[31], kZero);
+          lstep1[32] = _mm_madd_epi16(lstep1[32], kOne);
+          lstep1[33] = _mm_madd_epi16(lstep1[33], kOne);
+          lstep1[34] = _mm_madd_epi16(lstep1[34], kOne);
+          lstep1[35] = _mm_madd_epi16(lstep1[35], kOne);
+          lstep1[36] = _mm_madd_epi16(lstep1[36], kOne);
+          lstep1[37] = _mm_madd_epi16(lstep1[37], kOne);
+          lstep1[38] = _mm_madd_epi16(lstep1[38], kOne);
+          lstep1[39] = _mm_madd_epi16(lstep1[39], kOne);
+          lstep1[56] = _mm_madd_epi16(lstep1[56], kOne);
+          lstep1[57] = _mm_madd_epi16(lstep1[57], kOne);
+          lstep1[58] = _mm_madd_epi16(lstep1[58], kOne);
+          lstep1[59] = _mm_madd_epi16(lstep1[59], kOne);
+          lstep1[60] = _mm_madd_epi16(lstep1[60], kOne);
+          lstep1[61] = _mm_madd_epi16(lstep1[61], kOne);
+          lstep1[62] = _mm_madd_epi16(lstep1[62], kOne);
+          lstep1[63] = _mm_madd_epi16(lstep1[63], kOne);
+
+          lstep3[32] = _mm_add_epi32(lstep2[46], lstep1[32]);
+          lstep3[33] = _mm_add_epi32(lstep2[47], lstep1[33]);
+
+          lstep3[34] = _mm_add_epi32(lstep2[44], lstep1[34]);
+          lstep3[35] = _mm_add_epi32(lstep2[45], lstep1[35]);
+          lstep3[36] = _mm_add_epi32(lstep2[42], lstep1[36]);
+          lstep3[37] = _mm_add_epi32(lstep2[43], lstep1[37]);
+          lstep3[38] = _mm_add_epi32(lstep2[40], lstep1[38]);
+          lstep3[39] = _mm_add_epi32(lstep2[41], lstep1[39]);
+          lstep3[40] = _mm_sub_epi32(lstep1[38], lstep2[40]);
+          lstep3[41] = _mm_sub_epi32(lstep1[39], lstep2[41]);
+          lstep3[42] = _mm_sub_epi32(lstep1[36], lstep2[42]);
+          lstep3[43] = _mm_sub_epi32(lstep1[37], lstep2[43]);
+          lstep3[44] = _mm_sub_epi32(lstep1[34], lstep2[44]);
+          lstep3[45] = _mm_sub_epi32(lstep1[35], lstep2[45]);
+          lstep3[46] = _mm_sub_epi32(lstep1[32], lstep2[46]);
+          lstep3[47] = _mm_sub_epi32(lstep1[33], lstep2[47]);
+          lstep3[48] = _mm_sub_epi32(lstep1[62], lstep2[48]);
+          lstep3[49] = _mm_sub_epi32(lstep1[63], lstep2[49]);
+          lstep3[50] = _mm_sub_epi32(lstep1[60], lstep2[50]);
+          lstep3[51] = _mm_sub_epi32(lstep1[61], lstep2[51]);
+          lstep3[52] = _mm_sub_epi32(lstep1[58], lstep2[52]);
+          lstep3[53] = _mm_sub_epi32(lstep1[59], lstep2[53]);
+          lstep3[54] = _mm_sub_epi32(lstep1[56], lstep2[54]);
+          lstep3[55] = _mm_sub_epi32(lstep1[57], lstep2[55]);
+          lstep3[56] = _mm_add_epi32(lstep2[54], lstep1[56]);
+          lstep3[57] = _mm_add_epi32(lstep2[55], lstep1[57]);
+          lstep3[58] = _mm_add_epi32(lstep2[52], lstep1[58]);
+          lstep3[59] = _mm_add_epi32(lstep2[53], lstep1[59]);
+          lstep3[60] = _mm_add_epi32(lstep2[50], lstep1[60]);
+          lstep3[61] = _mm_add_epi32(lstep2[51], lstep1[61]);
+          lstep3[62] = _mm_add_epi32(lstep2[48], lstep1[62]);
+          lstep3[63] = _mm_add_epi32(lstep2[49], lstep1[63]);
+        }
+
+        // stage 4
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[16] = _mm_unpacklo_epi16(step2[8], kZero);
+          lstep2[17] = _mm_unpackhi_epi16(step2[8], kZero);
+          lstep2[18] = _mm_unpacklo_epi16(step2[9], kZero);
+          lstep2[19] = _mm_unpackhi_epi16(step2[9], kZero);
+          lstep2[28] = _mm_unpacklo_epi16(step2[14], kZero);
+          lstep2[29] = _mm_unpackhi_epi16(step2[14], kZero);
+          lstep2[30] = _mm_unpacklo_epi16(step2[15], kZero);
+          lstep2[31] = _mm_unpackhi_epi16(step2[15], kZero);
+          lstep2[16] = _mm_madd_epi16(lstep2[16], kOne);
+          lstep2[17] = _mm_madd_epi16(lstep2[17], kOne);
+          lstep2[18] = _mm_madd_epi16(lstep2[18], kOne);
+          lstep2[19] = _mm_madd_epi16(lstep2[19], kOne);
+          lstep2[28] = _mm_madd_epi16(lstep2[28], kOne);
+          lstep2[29] = _mm_madd_epi16(lstep2[29], kOne);
+          lstep2[30] = _mm_madd_epi16(lstep2[30], kOne);
+          lstep2[31] = _mm_madd_epi16(lstep2[31], kOne);
+
+          lstep1[0] = _mm_add_epi32(lstep3[6], lstep3[0]);
+          lstep1[1] = _mm_add_epi32(lstep3[7], lstep3[1]);
+          lstep1[2] = _mm_add_epi32(lstep3[4], lstep3[2]);
+          lstep1[3] = _mm_add_epi32(lstep3[5], lstep3[3]);
+          lstep1[4] = _mm_sub_epi32(lstep3[2], lstep3[4]);
+          lstep1[5] = _mm_sub_epi32(lstep3[3], lstep3[5]);
+          lstep1[6] = _mm_sub_epi32(lstep3[0], lstep3[6]);
+          lstep1[7] = _mm_sub_epi32(lstep3[1], lstep3[7]);
+          lstep1[16] = _mm_add_epi32(lstep3[22], lstep2[16]);
+          lstep1[17] = _mm_add_epi32(lstep3[23], lstep2[17]);
+          lstep1[18] = _mm_add_epi32(lstep3[20], lstep2[18]);
+          lstep1[19] = _mm_add_epi32(lstep3[21], lstep2[19]);
+          lstep1[20] = _mm_sub_epi32(lstep2[18], lstep3[20]);
+          lstep1[21] = _mm_sub_epi32(lstep2[19], lstep3[21]);
+          lstep1[22] = _mm_sub_epi32(lstep2[16], lstep3[22]);
+          lstep1[23] = _mm_sub_epi32(lstep2[17], lstep3[23]);
+          lstep1[24] = _mm_sub_epi32(lstep2[30], lstep3[24]);
+          lstep1[25] = _mm_sub_epi32(lstep2[31], lstep3[25]);
+          lstep1[26] = _mm_sub_epi32(lstep2[28], lstep3[26]);
+          lstep1[27] = _mm_sub_epi32(lstep2[29], lstep3[27]);
+          lstep1[28] = _mm_add_epi32(lstep3[26], lstep2[28]);
+          lstep1[29] = _mm_add_epi32(lstep3[27], lstep2[29]);
+          lstep1[30] = _mm_add_epi32(lstep3[24], lstep2[30]);
+          lstep1[31] = _mm_add_epi32(lstep3[25], lstep2[31]);
+        }
+        {
+          // to be continued...
+          //
+          const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
+          const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep3[12], lstep3[10]);
+          u[1] = _mm_unpackhi_epi32(lstep3[12], lstep3[10]);
+          u[2] = _mm_unpacklo_epi32(lstep3[13], lstep3[11]);
+          u[3] = _mm_unpackhi_epi32(lstep3[13], lstep3[11]);
+
+          // TODO(jingning): manually inline k_madd_epi32_ to further hide
+          // instruction latency.
+          v[0] = k_madd_epi32(u[0], k32_p16_m16);
+          v[1] = k_madd_epi32(u[1], k32_p16_m16);
+          v[2] = k_madd_epi32(u[2], k32_p16_m16);
+          v[3] = k_madd_epi32(u[3], k32_p16_m16);
+          v[4] = k_madd_epi32(u[0], k32_p16_p16);
+          v[5] = k_madd_epi32(u[1], k32_p16_p16);
+          v[6] = k_madd_epi32(u[2], k32_p16_p16);
+          v[7] = k_madd_epi32(u[3], k32_p16_p16);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_8(&v[0], &v[1], &v[2], &v[3], &v[4],
+                                              &v[5], &v[6], &v[7], &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+
+          lstep1[10] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          lstep1[11] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          lstep1[12] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          lstep1[13] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+        }
+        {
+          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep3[36], lstep3[58]);
+          u[1] = _mm_unpackhi_epi32(lstep3[36], lstep3[58]);
+          u[2] = _mm_unpacklo_epi32(lstep3[37], lstep3[59]);
+          u[3] = _mm_unpackhi_epi32(lstep3[37], lstep3[59]);
+          u[4] = _mm_unpacklo_epi32(lstep3[38], lstep3[56]);
+          u[5] = _mm_unpackhi_epi32(lstep3[38], lstep3[56]);
+          u[6] = _mm_unpacklo_epi32(lstep3[39], lstep3[57]);
+          u[7] = _mm_unpackhi_epi32(lstep3[39], lstep3[57]);
+          u[8] = _mm_unpacklo_epi32(lstep3[40], lstep3[54]);
+          u[9] = _mm_unpackhi_epi32(lstep3[40], lstep3[54]);
+          u[10] = _mm_unpacklo_epi32(lstep3[41], lstep3[55]);
+          u[11] = _mm_unpackhi_epi32(lstep3[41], lstep3[55]);
+          u[12] = _mm_unpacklo_epi32(lstep3[42], lstep3[52]);
+          u[13] = _mm_unpackhi_epi32(lstep3[42], lstep3[52]);
+          u[14] = _mm_unpacklo_epi32(lstep3[43], lstep3[53]);
+          u[15] = _mm_unpackhi_epi32(lstep3[43], lstep3[53]);
+
+          v[0] = k_madd_epi32(u[0], k32_m08_p24);
+          v[1] = k_madd_epi32(u[1], k32_m08_p24);
+          v[2] = k_madd_epi32(u[2], k32_m08_p24);
+          v[3] = k_madd_epi32(u[3], k32_m08_p24);
+          v[4] = k_madd_epi32(u[4], k32_m08_p24);
+          v[5] = k_madd_epi32(u[5], k32_m08_p24);
+          v[6] = k_madd_epi32(u[6], k32_m08_p24);
+          v[7] = k_madd_epi32(u[7], k32_m08_p24);
+          v[8] = k_madd_epi32(u[8], k32_m24_m08);
+          v[9] = k_madd_epi32(u[9], k32_m24_m08);
+          v[10] = k_madd_epi32(u[10], k32_m24_m08);
+          v[11] = k_madd_epi32(u[11], k32_m24_m08);
+          v[12] = k_madd_epi32(u[12], k32_m24_m08);
+          v[13] = k_madd_epi32(u[13], k32_m24_m08);
+          v[14] = k_madd_epi32(u[14], k32_m24_m08);
+          v[15] = k_madd_epi32(u[15], k32_m24_m08);
+          v[16] = k_madd_epi32(u[12], k32_m08_p24);
+          v[17] = k_madd_epi32(u[13], k32_m08_p24);
+          v[18] = k_madd_epi32(u[14], k32_m08_p24);
+          v[19] = k_madd_epi32(u[15], k32_m08_p24);
+          v[20] = k_madd_epi32(u[8], k32_m08_p24);
+          v[21] = k_madd_epi32(u[9], k32_m08_p24);
+          v[22] = k_madd_epi32(u[10], k32_m08_p24);
+          v[23] = k_madd_epi32(u[11], k32_m08_p24);
+          v[24] = k_madd_epi32(u[4], k32_p24_p08);
+          v[25] = k_madd_epi32(u[5], k32_p24_p08);
+          v[26] = k_madd_epi32(u[6], k32_p24_p08);
+          v[27] = k_madd_epi32(u[7], k32_p24_p08);
+          v[28] = k_madd_epi32(u[0], k32_p24_p08);
+          v[29] = k_madd_epi32(u[1], k32_p24_p08);
+          v[30] = k_madd_epi32(u[2], k32_p24_p08);
+          v[31] = k_madd_epi32(u[3], k32_p24_p08);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
+              &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16],
+              &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24],
+              &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+          u[8] = k_packs_epi64(v[16], v[17]);
+          u[9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+          v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+          v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep1[36] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          lstep1[37] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          lstep1[38] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          lstep1[39] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          lstep1[40] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          lstep1[41] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          lstep1[42] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          lstep1[43] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+          lstep1[52] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+          lstep1[53] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+          lstep1[54] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep1[55] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep1[56] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep1[57] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep1[58] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep1[59] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 5
+        {
+          lstep2[8] = _mm_add_epi32(lstep1[10], lstep3[8]);
+          lstep2[9] = _mm_add_epi32(lstep1[11], lstep3[9]);
+          lstep2[10] = _mm_sub_epi32(lstep3[8], lstep1[10]);
+          lstep2[11] = _mm_sub_epi32(lstep3[9], lstep1[11]);
+          lstep2[12] = _mm_sub_epi32(lstep3[14], lstep1[12]);
+          lstep2[13] = _mm_sub_epi32(lstep3[15], lstep1[13]);
+          lstep2[14] = _mm_add_epi32(lstep1[12], lstep3[14]);
+          lstep2[15] = _mm_add_epi32(lstep1[13], lstep3[15]);
+        }
+        {
+          const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
+          const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
+          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
+          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep1[0], lstep1[2]);
+          u[1] = _mm_unpackhi_epi32(lstep1[0], lstep1[2]);
+          u[2] = _mm_unpacklo_epi32(lstep1[1], lstep1[3]);
+          u[3] = _mm_unpackhi_epi32(lstep1[1], lstep1[3]);
+          u[4] = _mm_unpacklo_epi32(lstep1[4], lstep1[6]);
+          u[5] = _mm_unpackhi_epi32(lstep1[4], lstep1[6]);
+          u[6] = _mm_unpacklo_epi32(lstep1[5], lstep1[7]);
+          u[7] = _mm_unpackhi_epi32(lstep1[5], lstep1[7]);
+
+          // TODO(jingning): manually inline k_madd_epi32_ to further hide
+          // instruction latency.
+          v[0] = k_madd_epi32(u[0], k32_p16_p16);
+          v[1] = k_madd_epi32(u[1], k32_p16_p16);
+          v[2] = k_madd_epi32(u[2], k32_p16_p16);
+          v[3] = k_madd_epi32(u[3], k32_p16_p16);
+          v[4] = k_madd_epi32(u[0], k32_p16_m16);
+          v[5] = k_madd_epi32(u[1], k32_p16_m16);
+          v[6] = k_madd_epi32(u[2], k32_p16_m16);
+          v[7] = k_madd_epi32(u[3], k32_p16_m16);
+          v[8] = k_madd_epi32(u[4], k32_p24_p08);
+          v[9] = k_madd_epi32(u[5], k32_p24_p08);
+          v[10] = k_madd_epi32(u[6], k32_p24_p08);
+          v[11] = k_madd_epi32(u[7], k32_p24_p08);
+          v[12] = k_madd_epi32(u[4], k32_m08_p24);
+          v[13] = k_madd_epi32(u[5], k32_m08_p24);
+          v[14] = k_madd_epi32(u[6], k32_m08_p24);
+          v[15] = k_madd_epi32(u[7], k32_m08_p24);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_16(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
+              &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm_cmplt_epi32(u[0], kZero);
+          sign[1] = _mm_cmplt_epi32(u[1], kZero);
+          sign[2] = _mm_cmplt_epi32(u[2], kZero);
+          sign[3] = _mm_cmplt_epi32(u[3], kZero);
+          sign[4] = _mm_cmplt_epi32(u[4], kZero);
+          sign[5] = _mm_cmplt_epi32(u[5], kZero);
+          sign[6] = _mm_cmplt_epi32(u[6], kZero);
+          sign[7] = _mm_cmplt_epi32(u[7], kZero);
+
+          u[0] = _mm_sub_epi32(u[0], sign[0]);
+          u[1] = _mm_sub_epi32(u[1], sign[1]);
+          u[2] = _mm_sub_epi32(u[2], sign[2]);
+          u[3] = _mm_sub_epi32(u[3], sign[3]);
+          u[4] = _mm_sub_epi32(u[4], sign[4]);
+          u[5] = _mm_sub_epi32(u[5], sign[5]);
+          u[6] = _mm_sub_epi32(u[6], sign[6]);
+          u[7] = _mm_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm_add_epi32(u[0], K32One);
+          u[1] = _mm_add_epi32(u[1], K32One);
+          u[2] = _mm_add_epi32(u[2], K32One);
+          u[3] = _mm_add_epi32(u[3], K32One);
+          u[4] = _mm_add_epi32(u[4], K32One);
+          u[5] = _mm_add_epi32(u[5], K32One);
+          u[6] = _mm_add_epi32(u[6], K32One);
+          u[7] = _mm_add_epi32(u[7], K32One);
+
+          u[0] = _mm_srai_epi32(u[0], 2);
+          u[1] = _mm_srai_epi32(u[1], 2);
+          u[2] = _mm_srai_epi32(u[2], 2);
+          u[3] = _mm_srai_epi32(u[3], 2);
+          u[4] = _mm_srai_epi32(u[4], 2);
+          u[5] = _mm_srai_epi32(u[5], 2);
+          u[6] = _mm_srai_epi32(u[6], 2);
+          u[7] = _mm_srai_epi32(u[7], 2);
+
+          // Combine
+          out[0] = _mm_packs_epi32(u[0], u[1]);
+          out[16] = _mm_packs_epi32(u[2], u[3]);
+          out[8] = _mm_packs_epi32(u[4], u[5]);
+          out[24] = _mm_packs_epi32(u[6], u[7]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x4(&out[0], &out[16], &out[8], &out[24]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep1[18], lstep1[28]);
+          u[1] = _mm_unpackhi_epi32(lstep1[18], lstep1[28]);
+          u[2] = _mm_unpacklo_epi32(lstep1[19], lstep1[29]);
+          u[3] = _mm_unpackhi_epi32(lstep1[19], lstep1[29]);
+          u[4] = _mm_unpacklo_epi32(lstep1[20], lstep1[26]);
+          u[5] = _mm_unpackhi_epi32(lstep1[20], lstep1[26]);
+          u[6] = _mm_unpacklo_epi32(lstep1[21], lstep1[27]);
+          u[7] = _mm_unpackhi_epi32(lstep1[21], lstep1[27]);
+
+          v[0] = k_madd_epi32(u[0], k32_m08_p24);
+          v[1] = k_madd_epi32(u[1], k32_m08_p24);
+          v[2] = k_madd_epi32(u[2], k32_m08_p24);
+          v[3] = k_madd_epi32(u[3], k32_m08_p24);
+          v[4] = k_madd_epi32(u[4], k32_m24_m08);
+          v[5] = k_madd_epi32(u[5], k32_m24_m08);
+          v[6] = k_madd_epi32(u[6], k32_m24_m08);
+          v[7] = k_madd_epi32(u[7], k32_m24_m08);
+          v[8] = k_madd_epi32(u[4], k32_m08_p24);
+          v[9] = k_madd_epi32(u[5], k32_m08_p24);
+          v[10] = k_madd_epi32(u[6], k32_m08_p24);
+          v[11] = k_madd_epi32(u[7], k32_m08_p24);
+          v[12] = k_madd_epi32(u[0], k32_p24_p08);
+          v[13] = k_madd_epi32(u[1], k32_p24_p08);
+          v[14] = k_madd_epi32(u[2], k32_p24_p08);
+          v[15] = k_madd_epi32(u[3], k32_p24_p08);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_16(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
+              &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+
+          u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          lstep2[18] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+          lstep2[19] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+          lstep2[20] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+          lstep2[21] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+          lstep2[26] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+          lstep2[27] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+          lstep2[28] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+          lstep2[29] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+        }
+        {
+          lstep2[32] = _mm_add_epi32(lstep1[38], lstep3[32]);
+          lstep2[33] = _mm_add_epi32(lstep1[39], lstep3[33]);
+          lstep2[34] = _mm_add_epi32(lstep1[36], lstep3[34]);
+          lstep2[35] = _mm_add_epi32(lstep1[37], lstep3[35]);
+          lstep2[36] = _mm_sub_epi32(lstep3[34], lstep1[36]);
+          lstep2[37] = _mm_sub_epi32(lstep3[35], lstep1[37]);
+          lstep2[38] = _mm_sub_epi32(lstep3[32], lstep1[38]);
+          lstep2[39] = _mm_sub_epi32(lstep3[33], lstep1[39]);
+          lstep2[40] = _mm_sub_epi32(lstep3[46], lstep1[40]);
+          lstep2[41] = _mm_sub_epi32(lstep3[47], lstep1[41]);
+          lstep2[42] = _mm_sub_epi32(lstep3[44], lstep1[42]);
+          lstep2[43] = _mm_sub_epi32(lstep3[45], lstep1[43]);
+          lstep2[44] = _mm_add_epi32(lstep1[42], lstep3[44]);
+          lstep2[45] = _mm_add_epi32(lstep1[43], lstep3[45]);
+          lstep2[46] = _mm_add_epi32(lstep1[40], lstep3[46]);
+          lstep2[47] = _mm_add_epi32(lstep1[41], lstep3[47]);
+          lstep2[48] = _mm_add_epi32(lstep1[54], lstep3[48]);
+          lstep2[49] = _mm_add_epi32(lstep1[55], lstep3[49]);
+          lstep2[50] = _mm_add_epi32(lstep1[52], lstep3[50]);
+          lstep2[51] = _mm_add_epi32(lstep1[53], lstep3[51]);
+          lstep2[52] = _mm_sub_epi32(lstep3[50], lstep1[52]);
+          lstep2[53] = _mm_sub_epi32(lstep3[51], lstep1[53]);
+          lstep2[54] = _mm_sub_epi32(lstep3[48], lstep1[54]);
+          lstep2[55] = _mm_sub_epi32(lstep3[49], lstep1[55]);
+          lstep2[56] = _mm_sub_epi32(lstep3[62], lstep1[56]);
+          lstep2[57] = _mm_sub_epi32(lstep3[63], lstep1[57]);
+          lstep2[58] = _mm_sub_epi32(lstep3[60], lstep1[58]);
+          lstep2[59] = _mm_sub_epi32(lstep3[61], lstep1[59]);
+          lstep2[60] = _mm_add_epi32(lstep1[58], lstep3[60]);
+          lstep2[61] = _mm_add_epi32(lstep1[59], lstep3[61]);
+          lstep2[62] = _mm_add_epi32(lstep1[56], lstep3[62]);
+          lstep2[63] = _mm_add_epi32(lstep1[57], lstep3[63]);
+        }
+        // stage 6
+        {
+          const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
+          const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
+          const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep2[8], lstep2[14]);
+          u[1] = _mm_unpackhi_epi32(lstep2[8], lstep2[14]);
+          u[2] = _mm_unpacklo_epi32(lstep2[9], lstep2[15]);
+          u[3] = _mm_unpackhi_epi32(lstep2[9], lstep2[15]);
+          u[4] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[5] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[6] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[7] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[8] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[9] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[10] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[11] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[12] = _mm_unpacklo_epi32(lstep2[8], lstep2[14]);
+          u[13] = _mm_unpackhi_epi32(lstep2[8], lstep2[14]);
+          u[14] = _mm_unpacklo_epi32(lstep2[9], lstep2[15]);
+          u[15] = _mm_unpackhi_epi32(lstep2[9], lstep2[15]);
+
+          v[0] = k_madd_epi32(u[0], k32_p28_p04);
+          v[1] = k_madd_epi32(u[1], k32_p28_p04);
+          v[2] = k_madd_epi32(u[2], k32_p28_p04);
+          v[3] = k_madd_epi32(u[3], k32_p28_p04);
+          v[4] = k_madd_epi32(u[4], k32_p12_p20);
+          v[5] = k_madd_epi32(u[5], k32_p12_p20);
+          v[6] = k_madd_epi32(u[6], k32_p12_p20);
+          v[7] = k_madd_epi32(u[7], k32_p12_p20);
+          v[8] = k_madd_epi32(u[8], k32_m20_p12);
+          v[9] = k_madd_epi32(u[9], k32_m20_p12);
+          v[10] = k_madd_epi32(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32(u[12], k32_m04_p28);
+          v[13] = k_madd_epi32(u[13], k32_m04_p28);
+          v[14] = k_madd_epi32(u[14], k32_m04_p28);
+          v[15] = k_madd_epi32(u[15], k32_m04_p28);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_16(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
+              &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm_cmplt_epi32(u[0], kZero);
+          sign[1] = _mm_cmplt_epi32(u[1], kZero);
+          sign[2] = _mm_cmplt_epi32(u[2], kZero);
+          sign[3] = _mm_cmplt_epi32(u[3], kZero);
+          sign[4] = _mm_cmplt_epi32(u[4], kZero);
+          sign[5] = _mm_cmplt_epi32(u[5], kZero);
+          sign[6] = _mm_cmplt_epi32(u[6], kZero);
+          sign[7] = _mm_cmplt_epi32(u[7], kZero);
+
+          u[0] = _mm_sub_epi32(u[0], sign[0]);
+          u[1] = _mm_sub_epi32(u[1], sign[1]);
+          u[2] = _mm_sub_epi32(u[2], sign[2]);
+          u[3] = _mm_sub_epi32(u[3], sign[3]);
+          u[4] = _mm_sub_epi32(u[4], sign[4]);
+          u[5] = _mm_sub_epi32(u[5], sign[5]);
+          u[6] = _mm_sub_epi32(u[6], sign[6]);
+          u[7] = _mm_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm_add_epi32(u[0], K32One);
+          u[1] = _mm_add_epi32(u[1], K32One);
+          u[2] = _mm_add_epi32(u[2], K32One);
+          u[3] = _mm_add_epi32(u[3], K32One);
+          u[4] = _mm_add_epi32(u[4], K32One);
+          u[5] = _mm_add_epi32(u[5], K32One);
+          u[6] = _mm_add_epi32(u[6], K32One);
+          u[7] = _mm_add_epi32(u[7], K32One);
+
+          u[0] = _mm_srai_epi32(u[0], 2);
+          u[1] = _mm_srai_epi32(u[1], 2);
+          u[2] = _mm_srai_epi32(u[2], 2);
+          u[3] = _mm_srai_epi32(u[3], 2);
+          u[4] = _mm_srai_epi32(u[4], 2);
+          u[5] = _mm_srai_epi32(u[5], 2);
+          u[6] = _mm_srai_epi32(u[6], 2);
+          u[7] = _mm_srai_epi32(u[7], 2);
+
+          out[4] = _mm_packs_epi32(u[0], u[1]);
+          out[20] = _mm_packs_epi32(u[2], u[3]);
+          out[12] = _mm_packs_epi32(u[4], u[5]);
+          out[28] = _mm_packs_epi32(u[6], u[7]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x4(&out[4], &out[20], &out[12], &out[28]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          lstep3[16] = _mm_add_epi32(lstep2[18], lstep1[16]);
+          lstep3[17] = _mm_add_epi32(lstep2[19], lstep1[17]);
+          lstep3[18] = _mm_sub_epi32(lstep1[16], lstep2[18]);
+          lstep3[19] = _mm_sub_epi32(lstep1[17], lstep2[19]);
+          lstep3[20] = _mm_sub_epi32(lstep1[22], lstep2[20]);
+          lstep3[21] = _mm_sub_epi32(lstep1[23], lstep2[21]);
+          lstep3[22] = _mm_add_epi32(lstep2[20], lstep1[22]);
+          lstep3[23] = _mm_add_epi32(lstep2[21], lstep1[23]);
+          lstep3[24] = _mm_add_epi32(lstep2[26], lstep1[24]);
+          lstep3[25] = _mm_add_epi32(lstep2[27], lstep1[25]);
+          lstep3[26] = _mm_sub_epi32(lstep1[24], lstep2[26]);
+          lstep3[27] = _mm_sub_epi32(lstep1[25], lstep2[27]);
+          lstep3[28] = _mm_sub_epi32(lstep1[30], lstep2[28]);
+          lstep3[29] = _mm_sub_epi32(lstep1[31], lstep2[29]);
+          lstep3[30] = _mm_add_epi32(lstep2[28], lstep1[30]);
+          lstep3[31] = _mm_add_epi32(lstep2[29], lstep1[31]);
+        }
+        {
+          const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
+          const __m128i k32_m28_m04 = pair_set_epi32(-cospi_28_64, -cospi_4_64);
+          const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m128i k32_m12_m20 =
+              pair_set_epi32(-cospi_12_64, -cospi_20_64);
+          const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
+          const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep2[34], lstep2[60]);
+          u[1] = _mm_unpackhi_epi32(lstep2[34], lstep2[60]);
+          u[2] = _mm_unpacklo_epi32(lstep2[35], lstep2[61]);
+          u[3] = _mm_unpackhi_epi32(lstep2[35], lstep2[61]);
+          u[4] = _mm_unpacklo_epi32(lstep2[36], lstep2[58]);
+          u[5] = _mm_unpackhi_epi32(lstep2[36], lstep2[58]);
+          u[6] = _mm_unpacklo_epi32(lstep2[37], lstep2[59]);
+          u[7] = _mm_unpackhi_epi32(lstep2[37], lstep2[59]);
+          u[8] = _mm_unpacklo_epi32(lstep2[42], lstep2[52]);
+          u[9] = _mm_unpackhi_epi32(lstep2[42], lstep2[52]);
+          u[10] = _mm_unpacklo_epi32(lstep2[43], lstep2[53]);
+          u[11] = _mm_unpackhi_epi32(lstep2[43], lstep2[53]);
+          u[12] = _mm_unpacklo_epi32(lstep2[44], lstep2[50]);
+          u[13] = _mm_unpackhi_epi32(lstep2[44], lstep2[50]);
+          u[14] = _mm_unpacklo_epi32(lstep2[45], lstep2[51]);
+          u[15] = _mm_unpackhi_epi32(lstep2[45], lstep2[51]);
+
+          v[0] = k_madd_epi32(u[0], k32_m04_p28);
+          v[1] = k_madd_epi32(u[1], k32_m04_p28);
+          v[2] = k_madd_epi32(u[2], k32_m04_p28);
+          v[3] = k_madd_epi32(u[3], k32_m04_p28);
+          v[4] = k_madd_epi32(u[4], k32_m28_m04);
+          v[5] = k_madd_epi32(u[5], k32_m28_m04);
+          v[6] = k_madd_epi32(u[6], k32_m28_m04);
+          v[7] = k_madd_epi32(u[7], k32_m28_m04);
+          v[8] = k_madd_epi32(u[8], k32_m20_p12);
+          v[9] = k_madd_epi32(u[9], k32_m20_p12);
+          v[10] = k_madd_epi32(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32(u[12], k32_m12_m20);
+          v[13] = k_madd_epi32(u[13], k32_m12_m20);
+          v[14] = k_madd_epi32(u[14], k32_m12_m20);
+          v[15] = k_madd_epi32(u[15], k32_m12_m20);
+          v[16] = k_madd_epi32(u[12], k32_m20_p12);
+          v[17] = k_madd_epi32(u[13], k32_m20_p12);
+          v[18] = k_madd_epi32(u[14], k32_m20_p12);
+          v[19] = k_madd_epi32(u[15], k32_m20_p12);
+          v[20] = k_madd_epi32(u[8], k32_p12_p20);
+          v[21] = k_madd_epi32(u[9], k32_p12_p20);
+          v[22] = k_madd_epi32(u[10], k32_p12_p20);
+          v[23] = k_madd_epi32(u[11], k32_p12_p20);
+          v[24] = k_madd_epi32(u[4], k32_m04_p28);
+          v[25] = k_madd_epi32(u[5], k32_m04_p28);
+          v[26] = k_madd_epi32(u[6], k32_m04_p28);
+          v[27] = k_madd_epi32(u[7], k32_m04_p28);
+          v[28] = k_madd_epi32(u[0], k32_p28_p04);
+          v[29] = k_madd_epi32(u[1], k32_p28_p04);
+          v[30] = k_madd_epi32(u[2], k32_p28_p04);
+          v[31] = k_madd_epi32(u[3], k32_p28_p04);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
+              &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16],
+              &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24],
+              &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+          u[8] = k_packs_epi64(v[16], v[17]);
+          u[9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+          v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+          v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep3[34] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          lstep3[35] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          lstep3[36] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          lstep3[37] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          lstep3[42] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          lstep3[43] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          lstep3[44] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          lstep3[45] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+          lstep3[50] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+          lstep3[51] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+          lstep3[52] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep3[53] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep3[58] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep3[59] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep3[60] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep3[61] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 7
+        {
+          const __m128i k32_p30_p02 = pair_set_epi32(cospi_30_64, cospi_2_64);
+          const __m128i k32_p14_p18 = pair_set_epi32(cospi_14_64, cospi_18_64);
+          const __m128i k32_p22_p10 = pair_set_epi32(cospi_22_64, cospi_10_64);
+          const __m128i k32_p06_p26 = pair_set_epi32(cospi_6_64, cospi_26_64);
+          const __m128i k32_m26_p06 = pair_set_epi32(-cospi_26_64, cospi_6_64);
+          const __m128i k32_m10_p22 = pair_set_epi32(-cospi_10_64, cospi_22_64);
+          const __m128i k32_m18_p14 = pair_set_epi32(-cospi_18_64, cospi_14_64);
+          const __m128i k32_m02_p30 = pair_set_epi32(-cospi_2_64, cospi_30_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep3[16], lstep3[30]);
+          u[1] = _mm_unpackhi_epi32(lstep3[16], lstep3[30]);
+          u[2] = _mm_unpacklo_epi32(lstep3[17], lstep3[31]);
+          u[3] = _mm_unpackhi_epi32(lstep3[17], lstep3[31]);
+          u[4] = _mm_unpacklo_epi32(lstep3[18], lstep3[28]);
+          u[5] = _mm_unpackhi_epi32(lstep3[18], lstep3[28]);
+          u[6] = _mm_unpacklo_epi32(lstep3[19], lstep3[29]);
+          u[7] = _mm_unpackhi_epi32(lstep3[19], lstep3[29]);
+          u[8] = _mm_unpacklo_epi32(lstep3[20], lstep3[26]);
+          u[9] = _mm_unpackhi_epi32(lstep3[20], lstep3[26]);
+          u[10] = _mm_unpacklo_epi32(lstep3[21], lstep3[27]);
+          u[11] = _mm_unpackhi_epi32(lstep3[21], lstep3[27]);
+          u[12] = _mm_unpacklo_epi32(lstep3[22], lstep3[24]);
+          u[13] = _mm_unpackhi_epi32(lstep3[22], lstep3[24]);
+          u[14] = _mm_unpacklo_epi32(lstep3[23], lstep3[25]);
+          u[15] = _mm_unpackhi_epi32(lstep3[23], lstep3[25]);
+
+          v[0] = k_madd_epi32(u[0], k32_p30_p02);
+          v[1] = k_madd_epi32(u[1], k32_p30_p02);
+          v[2] = k_madd_epi32(u[2], k32_p30_p02);
+          v[3] = k_madd_epi32(u[3], k32_p30_p02);
+          v[4] = k_madd_epi32(u[4], k32_p14_p18);
+          v[5] = k_madd_epi32(u[5], k32_p14_p18);
+          v[6] = k_madd_epi32(u[6], k32_p14_p18);
+          v[7] = k_madd_epi32(u[7], k32_p14_p18);
+          v[8] = k_madd_epi32(u[8], k32_p22_p10);
+          v[9] = k_madd_epi32(u[9], k32_p22_p10);
+          v[10] = k_madd_epi32(u[10], k32_p22_p10);
+          v[11] = k_madd_epi32(u[11], k32_p22_p10);
+          v[12] = k_madd_epi32(u[12], k32_p06_p26);
+          v[13] = k_madd_epi32(u[13], k32_p06_p26);
+          v[14] = k_madd_epi32(u[14], k32_p06_p26);
+          v[15] = k_madd_epi32(u[15], k32_p06_p26);
+          v[16] = k_madd_epi32(u[12], k32_m26_p06);
+          v[17] = k_madd_epi32(u[13], k32_m26_p06);
+          v[18] = k_madd_epi32(u[14], k32_m26_p06);
+          v[19] = k_madd_epi32(u[15], k32_m26_p06);
+          v[20] = k_madd_epi32(u[8], k32_m10_p22);
+          v[21] = k_madd_epi32(u[9], k32_m10_p22);
+          v[22] = k_madd_epi32(u[10], k32_m10_p22);
+          v[23] = k_madd_epi32(u[11], k32_m10_p22);
+          v[24] = k_madd_epi32(u[4], k32_m18_p14);
+          v[25] = k_madd_epi32(u[5], k32_m18_p14);
+          v[26] = k_madd_epi32(u[6], k32_m18_p14);
+          v[27] = k_madd_epi32(u[7], k32_m18_p14);
+          v[28] = k_madd_epi32(u[0], k32_m02_p30);
+          v[29] = k_madd_epi32(u[1], k32_m02_p30);
+          v[30] = k_madd_epi32(u[2], k32_m02_p30);
+          v[31] = k_madd_epi32(u[3], k32_m02_p30);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
+              &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16],
+              &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24],
+              &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+          u[8] = k_packs_epi64(v[16], v[17]);
+          u[9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+          v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+          v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+          u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+          u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[0] = _mm_cmplt_epi32(u[0], kZero);
+          v[1] = _mm_cmplt_epi32(u[1], kZero);
+          v[2] = _mm_cmplt_epi32(u[2], kZero);
+          v[3] = _mm_cmplt_epi32(u[3], kZero);
+          v[4] = _mm_cmplt_epi32(u[4], kZero);
+          v[5] = _mm_cmplt_epi32(u[5], kZero);
+          v[6] = _mm_cmplt_epi32(u[6], kZero);
+          v[7] = _mm_cmplt_epi32(u[7], kZero);
+          v[8] = _mm_cmplt_epi32(u[8], kZero);
+          v[9] = _mm_cmplt_epi32(u[9], kZero);
+          v[10] = _mm_cmplt_epi32(u[10], kZero);
+          v[11] = _mm_cmplt_epi32(u[11], kZero);
+          v[12] = _mm_cmplt_epi32(u[12], kZero);
+          v[13] = _mm_cmplt_epi32(u[13], kZero);
+          v[14] = _mm_cmplt_epi32(u[14], kZero);
+          v[15] = _mm_cmplt_epi32(u[15], kZero);
+
+          u[0] = _mm_sub_epi32(u[0], v[0]);
+          u[1] = _mm_sub_epi32(u[1], v[1]);
+          u[2] = _mm_sub_epi32(u[2], v[2]);
+          u[3] = _mm_sub_epi32(u[3], v[3]);
+          u[4] = _mm_sub_epi32(u[4], v[4]);
+          u[5] = _mm_sub_epi32(u[5], v[5]);
+          u[6] = _mm_sub_epi32(u[6], v[6]);
+          u[7] = _mm_sub_epi32(u[7], v[7]);
+          u[8] = _mm_sub_epi32(u[8], v[8]);
+          u[9] = _mm_sub_epi32(u[9], v[9]);
+          u[10] = _mm_sub_epi32(u[10], v[10]);
+          u[11] = _mm_sub_epi32(u[11], v[11]);
+          u[12] = _mm_sub_epi32(u[12], v[12]);
+          u[13] = _mm_sub_epi32(u[13], v[13]);
+          u[14] = _mm_sub_epi32(u[14], v[14]);
+          u[15] = _mm_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], K32One);
+          v[1] = _mm_add_epi32(u[1], K32One);
+          v[2] = _mm_add_epi32(u[2], K32One);
+          v[3] = _mm_add_epi32(u[3], K32One);
+          v[4] = _mm_add_epi32(u[4], K32One);
+          v[5] = _mm_add_epi32(u[5], K32One);
+          v[6] = _mm_add_epi32(u[6], K32One);
+          v[7] = _mm_add_epi32(u[7], K32One);
+          v[8] = _mm_add_epi32(u[8], K32One);
+          v[9] = _mm_add_epi32(u[9], K32One);
+          v[10] = _mm_add_epi32(u[10], K32One);
+          v[11] = _mm_add_epi32(u[11], K32One);
+          v[12] = _mm_add_epi32(u[12], K32One);
+          v[13] = _mm_add_epi32(u[13], K32One);
+          v[14] = _mm_add_epi32(u[14], K32One);
+          v[15] = _mm_add_epi32(u[15], K32One);
+
+          u[0] = _mm_srai_epi32(v[0], 2);
+          u[1] = _mm_srai_epi32(v[1], 2);
+          u[2] = _mm_srai_epi32(v[2], 2);
+          u[3] = _mm_srai_epi32(v[3], 2);
+          u[4] = _mm_srai_epi32(v[4], 2);
+          u[5] = _mm_srai_epi32(v[5], 2);
+          u[6] = _mm_srai_epi32(v[6], 2);
+          u[7] = _mm_srai_epi32(v[7], 2);
+          u[8] = _mm_srai_epi32(v[8], 2);
+          u[9] = _mm_srai_epi32(v[9], 2);
+          u[10] = _mm_srai_epi32(v[10], 2);
+          u[11] = _mm_srai_epi32(v[11], 2);
+          u[12] = _mm_srai_epi32(v[12], 2);
+          u[13] = _mm_srai_epi32(v[13], 2);
+          u[14] = _mm_srai_epi32(v[14], 2);
+          u[15] = _mm_srai_epi32(v[15], 2);
+
+          out[2] = _mm_packs_epi32(u[0], u[1]);
+          out[18] = _mm_packs_epi32(u[2], u[3]);
+          out[10] = _mm_packs_epi32(u[4], u[5]);
+          out[26] = _mm_packs_epi32(u[6], u[7]);
+          out[6] = _mm_packs_epi32(u[8], u[9]);
+          out[22] = _mm_packs_epi32(u[10], u[11]);
+          out[14] = _mm_packs_epi32(u[12], u[13]);
+          out[30] = _mm_packs_epi32(u[14], u[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x8(&out[2], &out[18], &out[10], &out[26],
+                                      &out[6], &out[22], &out[14], &out[30]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          lstep1[32] = _mm_add_epi32(lstep3[34], lstep2[32]);
+          lstep1[33] = _mm_add_epi32(lstep3[35], lstep2[33]);
+          lstep1[34] = _mm_sub_epi32(lstep2[32], lstep3[34]);
+          lstep1[35] = _mm_sub_epi32(lstep2[33], lstep3[35]);
+          lstep1[36] = _mm_sub_epi32(lstep2[38], lstep3[36]);
+          lstep1[37] = _mm_sub_epi32(lstep2[39], lstep3[37]);
+          lstep1[38] = _mm_add_epi32(lstep3[36], lstep2[38]);
+          lstep1[39] = _mm_add_epi32(lstep3[37], lstep2[39]);
+          lstep1[40] = _mm_add_epi32(lstep3[42], lstep2[40]);
+          lstep1[41] = _mm_add_epi32(lstep3[43], lstep2[41]);
+          lstep1[42] = _mm_sub_epi32(lstep2[40], lstep3[42]);
+          lstep1[43] = _mm_sub_epi32(lstep2[41], lstep3[43]);
+          lstep1[44] = _mm_sub_epi32(lstep2[46], lstep3[44]);
+          lstep1[45] = _mm_sub_epi32(lstep2[47], lstep3[45]);
+          lstep1[46] = _mm_add_epi32(lstep3[44], lstep2[46]);
+          lstep1[47] = _mm_add_epi32(lstep3[45], lstep2[47]);
+          lstep1[48] = _mm_add_epi32(lstep3[50], lstep2[48]);
+          lstep1[49] = _mm_add_epi32(lstep3[51], lstep2[49]);
+          lstep1[50] = _mm_sub_epi32(lstep2[48], lstep3[50]);
+          lstep1[51] = _mm_sub_epi32(lstep2[49], lstep3[51]);
+          lstep1[52] = _mm_sub_epi32(lstep2[54], lstep3[52]);
+          lstep1[53] = _mm_sub_epi32(lstep2[55], lstep3[53]);
+          lstep1[54] = _mm_add_epi32(lstep3[52], lstep2[54]);
+          lstep1[55] = _mm_add_epi32(lstep3[53], lstep2[55]);
+          lstep1[56] = _mm_add_epi32(lstep3[58], lstep2[56]);
+          lstep1[57] = _mm_add_epi32(lstep3[59], lstep2[57]);
+          lstep1[58] = _mm_sub_epi32(lstep2[56], lstep3[58]);
+          lstep1[59] = _mm_sub_epi32(lstep2[57], lstep3[59]);
+          lstep1[60] = _mm_sub_epi32(lstep2[62], lstep3[60]);
+          lstep1[61] = _mm_sub_epi32(lstep2[63], lstep3[61]);
+          lstep1[62] = _mm_add_epi32(lstep3[60], lstep2[62]);
+          lstep1[63] = _mm_add_epi32(lstep3[61], lstep2[63]);
+        }
+        // stage 8
+        {
+          const __m128i k32_p31_p01 = pair_set_epi32(cospi_31_64, cospi_1_64);
+          const __m128i k32_p15_p17 = pair_set_epi32(cospi_15_64, cospi_17_64);
+          const __m128i k32_p23_p09 = pair_set_epi32(cospi_23_64, cospi_9_64);
+          const __m128i k32_p07_p25 = pair_set_epi32(cospi_7_64, cospi_25_64);
+          const __m128i k32_m25_p07 = pair_set_epi32(-cospi_25_64, cospi_7_64);
+          const __m128i k32_m09_p23 = pair_set_epi32(-cospi_9_64, cospi_23_64);
+          const __m128i k32_m17_p15 = pair_set_epi32(-cospi_17_64, cospi_15_64);
+          const __m128i k32_m01_p31 = pair_set_epi32(-cospi_1_64, cospi_31_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep1[32], lstep1[62]);
+          u[1] = _mm_unpackhi_epi32(lstep1[32], lstep1[62]);
+          u[2] = _mm_unpacklo_epi32(lstep1[33], lstep1[63]);
+          u[3] = _mm_unpackhi_epi32(lstep1[33], lstep1[63]);
+          u[4] = _mm_unpacklo_epi32(lstep1[34], lstep1[60]);
+          u[5] = _mm_unpackhi_epi32(lstep1[34], lstep1[60]);
+          u[6] = _mm_unpacklo_epi32(lstep1[35], lstep1[61]);
+          u[7] = _mm_unpackhi_epi32(lstep1[35], lstep1[61]);
+          u[8] = _mm_unpacklo_epi32(lstep1[36], lstep1[58]);
+          u[9] = _mm_unpackhi_epi32(lstep1[36], lstep1[58]);
+          u[10] = _mm_unpacklo_epi32(lstep1[37], lstep1[59]);
+          u[11] = _mm_unpackhi_epi32(lstep1[37], lstep1[59]);
+          u[12] = _mm_unpacklo_epi32(lstep1[38], lstep1[56]);
+          u[13] = _mm_unpackhi_epi32(lstep1[38], lstep1[56]);
+          u[14] = _mm_unpacklo_epi32(lstep1[39], lstep1[57]);
+          u[15] = _mm_unpackhi_epi32(lstep1[39], lstep1[57]);
+
+          v[0] = k_madd_epi32(u[0], k32_p31_p01);
+          v[1] = k_madd_epi32(u[1], k32_p31_p01);
+          v[2] = k_madd_epi32(u[2], k32_p31_p01);
+          v[3] = k_madd_epi32(u[3], k32_p31_p01);
+          v[4] = k_madd_epi32(u[4], k32_p15_p17);
+          v[5] = k_madd_epi32(u[5], k32_p15_p17);
+          v[6] = k_madd_epi32(u[6], k32_p15_p17);
+          v[7] = k_madd_epi32(u[7], k32_p15_p17);
+          v[8] = k_madd_epi32(u[8], k32_p23_p09);
+          v[9] = k_madd_epi32(u[9], k32_p23_p09);
+          v[10] = k_madd_epi32(u[10], k32_p23_p09);
+          v[11] = k_madd_epi32(u[11], k32_p23_p09);
+          v[12] = k_madd_epi32(u[12], k32_p07_p25);
+          v[13] = k_madd_epi32(u[13], k32_p07_p25);
+          v[14] = k_madd_epi32(u[14], k32_p07_p25);
+          v[15] = k_madd_epi32(u[15], k32_p07_p25);
+          v[16] = k_madd_epi32(u[12], k32_m25_p07);
+          v[17] = k_madd_epi32(u[13], k32_m25_p07);
+          v[18] = k_madd_epi32(u[14], k32_m25_p07);
+          v[19] = k_madd_epi32(u[15], k32_m25_p07);
+          v[20] = k_madd_epi32(u[8], k32_m09_p23);
+          v[21] = k_madd_epi32(u[9], k32_m09_p23);
+          v[22] = k_madd_epi32(u[10], k32_m09_p23);
+          v[23] = k_madd_epi32(u[11], k32_m09_p23);
+          v[24] = k_madd_epi32(u[4], k32_m17_p15);
+          v[25] = k_madd_epi32(u[5], k32_m17_p15);
+          v[26] = k_madd_epi32(u[6], k32_m17_p15);
+          v[27] = k_madd_epi32(u[7], k32_m17_p15);
+          v[28] = k_madd_epi32(u[0], k32_m01_p31);
+          v[29] = k_madd_epi32(u[1], k32_m01_p31);
+          v[30] = k_madd_epi32(u[2], k32_m01_p31);
+          v[31] = k_madd_epi32(u[3], k32_m01_p31);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
+              &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16],
+              &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24],
+              &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+          u[8] = k_packs_epi64(v[16], v[17]);
+          u[9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+          v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+          v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+          u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+          u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[0] = _mm_cmplt_epi32(u[0], kZero);
+          v[1] = _mm_cmplt_epi32(u[1], kZero);
+          v[2] = _mm_cmplt_epi32(u[2], kZero);
+          v[3] = _mm_cmplt_epi32(u[3], kZero);
+          v[4] = _mm_cmplt_epi32(u[4], kZero);
+          v[5] = _mm_cmplt_epi32(u[5], kZero);
+          v[6] = _mm_cmplt_epi32(u[6], kZero);
+          v[7] = _mm_cmplt_epi32(u[7], kZero);
+          v[8] = _mm_cmplt_epi32(u[8], kZero);
+          v[9] = _mm_cmplt_epi32(u[9], kZero);
+          v[10] = _mm_cmplt_epi32(u[10], kZero);
+          v[11] = _mm_cmplt_epi32(u[11], kZero);
+          v[12] = _mm_cmplt_epi32(u[12], kZero);
+          v[13] = _mm_cmplt_epi32(u[13], kZero);
+          v[14] = _mm_cmplt_epi32(u[14], kZero);
+          v[15] = _mm_cmplt_epi32(u[15], kZero);
+
+          u[0] = _mm_sub_epi32(u[0], v[0]);
+          u[1] = _mm_sub_epi32(u[1], v[1]);
+          u[2] = _mm_sub_epi32(u[2], v[2]);
+          u[3] = _mm_sub_epi32(u[3], v[3]);
+          u[4] = _mm_sub_epi32(u[4], v[4]);
+          u[5] = _mm_sub_epi32(u[5], v[5]);
+          u[6] = _mm_sub_epi32(u[6], v[6]);
+          u[7] = _mm_sub_epi32(u[7], v[7]);
+          u[8] = _mm_sub_epi32(u[8], v[8]);
+          u[9] = _mm_sub_epi32(u[9], v[9]);
+          u[10] = _mm_sub_epi32(u[10], v[10]);
+          u[11] = _mm_sub_epi32(u[11], v[11]);
+          u[12] = _mm_sub_epi32(u[12], v[12]);
+          u[13] = _mm_sub_epi32(u[13], v[13]);
+          u[14] = _mm_sub_epi32(u[14], v[14]);
+          u[15] = _mm_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], K32One);
+          v[1] = _mm_add_epi32(u[1], K32One);
+          v[2] = _mm_add_epi32(u[2], K32One);
+          v[3] = _mm_add_epi32(u[3], K32One);
+          v[4] = _mm_add_epi32(u[4], K32One);
+          v[5] = _mm_add_epi32(u[5], K32One);
+          v[6] = _mm_add_epi32(u[6], K32One);
+          v[7] = _mm_add_epi32(u[7], K32One);
+          v[8] = _mm_add_epi32(u[8], K32One);
+          v[9] = _mm_add_epi32(u[9], K32One);
+          v[10] = _mm_add_epi32(u[10], K32One);
+          v[11] = _mm_add_epi32(u[11], K32One);
+          v[12] = _mm_add_epi32(u[12], K32One);
+          v[13] = _mm_add_epi32(u[13], K32One);
+          v[14] = _mm_add_epi32(u[14], K32One);
+          v[15] = _mm_add_epi32(u[15], K32One);
+
+          u[0] = _mm_srai_epi32(v[0], 2);
+          u[1] = _mm_srai_epi32(v[1], 2);
+          u[2] = _mm_srai_epi32(v[2], 2);
+          u[3] = _mm_srai_epi32(v[3], 2);
+          u[4] = _mm_srai_epi32(v[4], 2);
+          u[5] = _mm_srai_epi32(v[5], 2);
+          u[6] = _mm_srai_epi32(v[6], 2);
+          u[7] = _mm_srai_epi32(v[7], 2);
+          u[8] = _mm_srai_epi32(v[8], 2);
+          u[9] = _mm_srai_epi32(v[9], 2);
+          u[10] = _mm_srai_epi32(v[10], 2);
+          u[11] = _mm_srai_epi32(v[11], 2);
+          u[12] = _mm_srai_epi32(v[12], 2);
+          u[13] = _mm_srai_epi32(v[13], 2);
+          u[14] = _mm_srai_epi32(v[14], 2);
+          u[15] = _mm_srai_epi32(v[15], 2);
+
+          out[1] = _mm_packs_epi32(u[0], u[1]);
+          out[17] = _mm_packs_epi32(u[2], u[3]);
+          out[9] = _mm_packs_epi32(u[4], u[5]);
+          out[25] = _mm_packs_epi32(u[6], u[7]);
+          out[7] = _mm_packs_epi32(u[8], u[9]);
+          out[23] = _mm_packs_epi32(u[10], u[11]);
+          out[15] = _mm_packs_epi32(u[12], u[13]);
+          out[31] = _mm_packs_epi32(u[14], u[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x8(&out[1], &out[17], &out[9], &out[25],
+                                      &out[7], &out[23], &out[15], &out[31]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i k32_p27_p05 = pair_set_epi32(cospi_27_64, cospi_5_64);
+          const __m128i k32_p11_p21 = pair_set_epi32(cospi_11_64, cospi_21_64);
+          const __m128i k32_p19_p13 = pair_set_epi32(cospi_19_64, cospi_13_64);
+          const __m128i k32_p03_p29 = pair_set_epi32(cospi_3_64, cospi_29_64);
+          const __m128i k32_m29_p03 = pair_set_epi32(-cospi_29_64, cospi_3_64);
+          const __m128i k32_m13_p19 = pair_set_epi32(-cospi_13_64, cospi_19_64);
+          const __m128i k32_m21_p11 = pair_set_epi32(-cospi_21_64, cospi_11_64);
+          const __m128i k32_m05_p27 = pair_set_epi32(-cospi_5_64, cospi_27_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep1[40], lstep1[54]);
+          u[1] = _mm_unpackhi_epi32(lstep1[40], lstep1[54]);
+          u[2] = _mm_unpacklo_epi32(lstep1[41], lstep1[55]);
+          u[3] = _mm_unpackhi_epi32(lstep1[41], lstep1[55]);
+          u[4] = _mm_unpacklo_epi32(lstep1[42], lstep1[52]);
+          u[5] = _mm_unpackhi_epi32(lstep1[42], lstep1[52]);
+          u[6] = _mm_unpacklo_epi32(lstep1[43], lstep1[53]);
+          u[7] = _mm_unpackhi_epi32(lstep1[43], lstep1[53]);
+          u[8] = _mm_unpacklo_epi32(lstep1[44], lstep1[50]);
+          u[9] = _mm_unpackhi_epi32(lstep1[44], lstep1[50]);
+          u[10] = _mm_unpacklo_epi32(lstep1[45], lstep1[51]);
+          u[11] = _mm_unpackhi_epi32(lstep1[45], lstep1[51]);
+          u[12] = _mm_unpacklo_epi32(lstep1[46], lstep1[48]);
+          u[13] = _mm_unpackhi_epi32(lstep1[46], lstep1[48]);
+          u[14] = _mm_unpacklo_epi32(lstep1[47], lstep1[49]);
+          u[15] = _mm_unpackhi_epi32(lstep1[47], lstep1[49]);
+
+          v[0] = k_madd_epi32(u[0], k32_p27_p05);
+          v[1] = k_madd_epi32(u[1], k32_p27_p05);
+          v[2] = k_madd_epi32(u[2], k32_p27_p05);
+          v[3] = k_madd_epi32(u[3], k32_p27_p05);
+          v[4] = k_madd_epi32(u[4], k32_p11_p21);
+          v[5] = k_madd_epi32(u[5], k32_p11_p21);
+          v[6] = k_madd_epi32(u[6], k32_p11_p21);
+          v[7] = k_madd_epi32(u[7], k32_p11_p21);
+          v[8] = k_madd_epi32(u[8], k32_p19_p13);
+          v[9] = k_madd_epi32(u[9], k32_p19_p13);
+          v[10] = k_madd_epi32(u[10], k32_p19_p13);
+          v[11] = k_madd_epi32(u[11], k32_p19_p13);
+          v[12] = k_madd_epi32(u[12], k32_p03_p29);
+          v[13] = k_madd_epi32(u[13], k32_p03_p29);
+          v[14] = k_madd_epi32(u[14], k32_p03_p29);
+          v[15] = k_madd_epi32(u[15], k32_p03_p29);
+          v[16] = k_madd_epi32(u[12], k32_m29_p03);
+          v[17] = k_madd_epi32(u[13], k32_m29_p03);
+          v[18] = k_madd_epi32(u[14], k32_m29_p03);
+          v[19] = k_madd_epi32(u[15], k32_m29_p03);
+          v[20] = k_madd_epi32(u[8], k32_m13_p19);
+          v[21] = k_madd_epi32(u[9], k32_m13_p19);
+          v[22] = k_madd_epi32(u[10], k32_m13_p19);
+          v[23] = k_madd_epi32(u[11], k32_m13_p19);
+          v[24] = k_madd_epi32(u[4], k32_m21_p11);
+          v[25] = k_madd_epi32(u[5], k32_m21_p11);
+          v[26] = k_madd_epi32(u[6], k32_m21_p11);
+          v[27] = k_madd_epi32(u[7], k32_m21_p11);
+          v[28] = k_madd_epi32(u[0], k32_m05_p27);
+          v[29] = k_madd_epi32(u[1], k32_m05_p27);
+          v[30] = k_madd_epi32(u[2], k32_m05_p27);
+          v[31] = k_madd_epi32(u[3], k32_m05_p27);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7], &v[8],
+              &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15], &v[16],
+              &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23], &v[24],
+              &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31], &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+          u[8] = k_packs_epi64(v[16], v[17]);
+          u[9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+          v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+          v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+          u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+          u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[0] = _mm_cmplt_epi32(u[0], kZero);
+          v[1] = _mm_cmplt_epi32(u[1], kZero);
+          v[2] = _mm_cmplt_epi32(u[2], kZero);
+          v[3] = _mm_cmplt_epi32(u[3], kZero);
+          v[4] = _mm_cmplt_epi32(u[4], kZero);
+          v[5] = _mm_cmplt_epi32(u[5], kZero);
+          v[6] = _mm_cmplt_epi32(u[6], kZero);
+          v[7] = _mm_cmplt_epi32(u[7], kZero);
+          v[8] = _mm_cmplt_epi32(u[8], kZero);
+          v[9] = _mm_cmplt_epi32(u[9], kZero);
+          v[10] = _mm_cmplt_epi32(u[10], kZero);
+          v[11] = _mm_cmplt_epi32(u[11], kZero);
+          v[12] = _mm_cmplt_epi32(u[12], kZero);
+          v[13] = _mm_cmplt_epi32(u[13], kZero);
+          v[14] = _mm_cmplt_epi32(u[14], kZero);
+          v[15] = _mm_cmplt_epi32(u[15], kZero);
+
+          u[0] = _mm_sub_epi32(u[0], v[0]);
+          u[1] = _mm_sub_epi32(u[1], v[1]);
+          u[2] = _mm_sub_epi32(u[2], v[2]);
+          u[3] = _mm_sub_epi32(u[3], v[3]);
+          u[4] = _mm_sub_epi32(u[4], v[4]);
+          u[5] = _mm_sub_epi32(u[5], v[5]);
+          u[6] = _mm_sub_epi32(u[6], v[6]);
+          u[7] = _mm_sub_epi32(u[7], v[7]);
+          u[8] = _mm_sub_epi32(u[8], v[8]);
+          u[9] = _mm_sub_epi32(u[9], v[9]);
+          u[10] = _mm_sub_epi32(u[10], v[10]);
+          u[11] = _mm_sub_epi32(u[11], v[11]);
+          u[12] = _mm_sub_epi32(u[12], v[12]);
+          u[13] = _mm_sub_epi32(u[13], v[13]);
+          u[14] = _mm_sub_epi32(u[14], v[14]);
+          u[15] = _mm_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], K32One);
+          v[1] = _mm_add_epi32(u[1], K32One);
+          v[2] = _mm_add_epi32(u[2], K32One);
+          v[3] = _mm_add_epi32(u[3], K32One);
+          v[4] = _mm_add_epi32(u[4], K32One);
+          v[5] = _mm_add_epi32(u[5], K32One);
+          v[6] = _mm_add_epi32(u[6], K32One);
+          v[7] = _mm_add_epi32(u[7], K32One);
+          v[8] = _mm_add_epi32(u[8], K32One);
+          v[9] = _mm_add_epi32(u[9], K32One);
+          v[10] = _mm_add_epi32(u[10], K32One);
+          v[11] = _mm_add_epi32(u[11], K32One);
+          v[12] = _mm_add_epi32(u[12], K32One);
+          v[13] = _mm_add_epi32(u[13], K32One);
+          v[14] = _mm_add_epi32(u[14], K32One);
+          v[15] = _mm_add_epi32(u[15], K32One);
+
+          u[0] = _mm_srai_epi32(v[0], 2);
+          u[1] = _mm_srai_epi32(v[1], 2);
+          u[2] = _mm_srai_epi32(v[2], 2);
+          u[3] = _mm_srai_epi32(v[3], 2);
+          u[4] = _mm_srai_epi32(v[4], 2);
+          u[5] = _mm_srai_epi32(v[5], 2);
+          u[6] = _mm_srai_epi32(v[6], 2);
+          u[7] = _mm_srai_epi32(v[7], 2);
+          u[8] = _mm_srai_epi32(v[8], 2);
+          u[9] = _mm_srai_epi32(v[9], 2);
+          u[10] = _mm_srai_epi32(v[10], 2);
+          u[11] = _mm_srai_epi32(v[11], 2);
+          u[12] = _mm_srai_epi32(v[12], 2);
+          u[13] = _mm_srai_epi32(v[13], 2);
+          u[14] = _mm_srai_epi32(v[14], 2);
+          u[15] = _mm_srai_epi32(v[15], 2);
+
+          out[5] = _mm_packs_epi32(u[0], u[1]);
+          out[21] = _mm_packs_epi32(u[2], u[3]);
+          out[13] = _mm_packs_epi32(u[4], u[5]);
+          out[29] = _mm_packs_epi32(u[6], u[7]);
+          out[3] = _mm_packs_epi32(u[8], u[9]);
+          out[19] = _mm_packs_epi32(u[10], u[11]);
+          out[11] = _mm_packs_epi32(u[12], u[13]);
+          out[27] = _mm_packs_epi32(u[14], u[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x8(&out[5], &out[21], &out[13], &out[29],
+                                      &out[3], &out[19], &out[11], &out[27]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+#endif  // FDCT32x32_HIGH_PRECISION
+      // Transpose the results, do it as four 8x8 transposes.
+      {
+        int transpose_block;
+        int16_t *output0 = &intermediate[column_start * 32];
+        tran_low_t *output1 = &output_org[column_start * 32];
+        for (transpose_block = 0; transpose_block < 4; ++transpose_block) {
+          __m128i *this_out = &out[8 * transpose_block];
+          // 00 01 02 03 04 05 06 07
+          // 10 11 12 13 14 15 16 17
+          // 20 21 22 23 24 25 26 27
+          // 30 31 32 33 34 35 36 37
+          // 40 41 42 43 44 45 46 47
+          // 50 51 52 53 54 55 56 57
+          // 60 61 62 63 64 65 66 67
+          // 70 71 72 73 74 75 76 77
+          const __m128i tr0_0 = _mm_unpacklo_epi16(this_out[0], this_out[1]);
+          const __m128i tr0_1 = _mm_unpacklo_epi16(this_out[2], this_out[3]);
+          const __m128i tr0_2 = _mm_unpackhi_epi16(this_out[0], this_out[1]);
+          const __m128i tr0_3 = _mm_unpackhi_epi16(this_out[2], this_out[3]);
+          const __m128i tr0_4 = _mm_unpacklo_epi16(this_out[4], this_out[5]);
+          const __m128i tr0_5 = _mm_unpacklo_epi16(this_out[6], this_out[7]);
+          const __m128i tr0_6 = _mm_unpackhi_epi16(this_out[4], this_out[5]);
+          const __m128i tr0_7 = _mm_unpackhi_epi16(this_out[6], this_out[7]);
+          // 00 10 01 11 02 12 03 13
+          // 20 30 21 31 22 32 23 33
+          // 04 14 05 15 06 16 07 17
+          // 24 34 25 35 26 36 27 37
+          // 40 50 41 51 42 52 43 53
+          // 60 70 61 71 62 72 63 73
+          // 54 54 55 55 56 56 57 57
+          // 64 74 65 75 66 76 67 77
+          const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+          const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+          const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+          const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+          const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+          const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+          const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+          const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+          // 00 10 20 30 01 11 21 31
+          // 40 50 60 70 41 51 61 71
+          // 02 12 22 32 03 13 23 33
+          // 42 52 62 72 43 53 63 73
+          // 04 14 24 34 05 15 21 36
+          // 44 54 64 74 45 55 61 76
+          // 06 16 26 36 07 17 27 37
+          // 46 56 66 76 47 57 67 77
+          __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+          __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+          __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+          __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+          __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+          __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+          __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+          __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+          // 00 10 20 30 40 50 60 70
+          // 01 11 21 31 41 51 61 71
+          // 02 12 22 32 42 52 62 72
+          // 03 13 23 33 43 53 63 73
+          // 04 14 24 34 44 54 64 74
+          // 05 15 25 35 45 55 65 75
+          // 06 16 26 36 46 56 66 76
+          // 07 17 27 37 47 57 67 77
+          if (0 == pass) {
+            // output[j] = (output[j] + 1 + (output[j] > 0)) >> 2;
+            // TODO(cd): see quality impact of only doing
+            //           output[j] = (output[j] + 1) >> 2;
+            //           which would remove the code between here ...
+            __m128i tr2_0_0 = _mm_cmpgt_epi16(tr2_0, kZero);
+            __m128i tr2_1_0 = _mm_cmpgt_epi16(tr2_1, kZero);
+            __m128i tr2_2_0 = _mm_cmpgt_epi16(tr2_2, kZero);
+            __m128i tr2_3_0 = _mm_cmpgt_epi16(tr2_3, kZero);
+            __m128i tr2_4_0 = _mm_cmpgt_epi16(tr2_4, kZero);
+            __m128i tr2_5_0 = _mm_cmpgt_epi16(tr2_5, kZero);
+            __m128i tr2_6_0 = _mm_cmpgt_epi16(tr2_6, kZero);
+            __m128i tr2_7_0 = _mm_cmpgt_epi16(tr2_7, kZero);
+            tr2_0 = _mm_sub_epi16(tr2_0, tr2_0_0);
+            tr2_1 = _mm_sub_epi16(tr2_1, tr2_1_0);
+            tr2_2 = _mm_sub_epi16(tr2_2, tr2_2_0);
+            tr2_3 = _mm_sub_epi16(tr2_3, tr2_3_0);
+            tr2_4 = _mm_sub_epi16(tr2_4, tr2_4_0);
+            tr2_5 = _mm_sub_epi16(tr2_5, tr2_5_0);
+            tr2_6 = _mm_sub_epi16(tr2_6, tr2_6_0);
+            tr2_7 = _mm_sub_epi16(tr2_7, tr2_7_0);
+            //           ... and here.
+            //           PS: also change code in av1/encoder/av1_dct.c
+            tr2_0 = _mm_add_epi16(tr2_0, kOne);
+            tr2_1 = _mm_add_epi16(tr2_1, kOne);
+            tr2_2 = _mm_add_epi16(tr2_2, kOne);
+            tr2_3 = _mm_add_epi16(tr2_3, kOne);
+            tr2_4 = _mm_add_epi16(tr2_4, kOne);
+            tr2_5 = _mm_add_epi16(tr2_5, kOne);
+            tr2_6 = _mm_add_epi16(tr2_6, kOne);
+            tr2_7 = _mm_add_epi16(tr2_7, kOne);
+            tr2_0 = _mm_srai_epi16(tr2_0, 2);
+            tr2_1 = _mm_srai_epi16(tr2_1, 2);
+            tr2_2 = _mm_srai_epi16(tr2_2, 2);
+            tr2_3 = _mm_srai_epi16(tr2_3, 2);
+            tr2_4 = _mm_srai_epi16(tr2_4, 2);
+            tr2_5 = _mm_srai_epi16(tr2_5, 2);
+            tr2_6 = _mm_srai_epi16(tr2_6, 2);
+            tr2_7 = _mm_srai_epi16(tr2_7, 2);
+          }
+          // Note: even though all these stores are aligned, using the aligned
+          //       intrinsic make the code slightly slower.
+          if (pass == 0) {
+            _mm_storeu_si128((__m128i *)(output0 + 0 * 32), tr2_0);
+            _mm_storeu_si128((__m128i *)(output0 + 1 * 32), tr2_1);
+            _mm_storeu_si128((__m128i *)(output0 + 2 * 32), tr2_2);
+            _mm_storeu_si128((__m128i *)(output0 + 3 * 32), tr2_3);
+            _mm_storeu_si128((__m128i *)(output0 + 4 * 32), tr2_4);
+            _mm_storeu_si128((__m128i *)(output0 + 5 * 32), tr2_5);
+            _mm_storeu_si128((__m128i *)(output0 + 6 * 32), tr2_6);
+            _mm_storeu_si128((__m128i *)(output0 + 7 * 32), tr2_7);
+            // Process next 8x8
+            output0 += 8;
+          } else {
+            storeu_output(&tr2_0, (output1 + 0 * 32));
+            storeu_output(&tr2_1, (output1 + 1 * 32));
+            storeu_output(&tr2_2, (output1 + 2 * 32));
+            storeu_output(&tr2_3, (output1 + 3 * 32));
+            storeu_output(&tr2_4, (output1 + 4 * 32));
+            storeu_output(&tr2_5, (output1 + 5 * 32));
+            storeu_output(&tr2_6, (output1 + 6 * 32));
+            storeu_output(&tr2_7, (output1 + 7 * 32));
+            // Process next 8x8
+            output1 += 8;
+          }
+        }
+      }
+    }
+  }
+}  // NOLINT
+
+#undef ADD_EPI16
+#undef SUB_EPI16
+#undef HIGH_FDCT32x32_2D_C
+#undef HIGH_FDCT32x32_2D_ROWS_C
diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.c b/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.c
new file mode 100644
index 0000000000..670f864d07
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.c
@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+
+#define FDCT32x32_2D_AVX2 aom_fdct32x32_rd_avx2
+#define FDCT32x32_HIGH_PRECISION 0
+#include "aom_dsp/x86/fwd_dct32x32_impl_avx2.h"
+#undef FDCT32x32_2D_AVX2
+#undef FDCT32x32_HIGH_PRECISION
+
+#define FDCT32x32_2D_AVX2 aom_fdct32x32_avx2
+#define FDCT32x32_HIGH_PRECISION 1
+#include "aom_dsp/x86/fwd_dct32x32_impl_avx2.h"  // NOLINT
+#undef FDCT32x32_2D_AVX2
+#undef FDCT32x32_HIGH_PRECISION
diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.h b/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.h
new file mode 100644
index 0000000000..d3aceae009
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_txfm_avx2.h
@@ -0,0 +1,35 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_X86_FWD_TXFM_AVX2_H
+#define AOM_DSP_X86_FWD_TXFM_AVX2_H
+
+#include "./aom_config.h"
+
+static INLINE void storeu_output_avx2(const __m256i *coeff, tran_low_t *out) {
+#if CONFIG_HIGHBITDEPTH
+  const __m256i zero = _mm256_setzero_si256();
+  const __m256i sign = _mm256_cmpgt_epi16(zero, *coeff);
+
+  __m256i x0 = _mm256_unpacklo_epi16(*coeff, sign);
+  __m256i x1 = _mm256_unpackhi_epi16(*coeff, sign);
+
+  __m256i y0 = _mm256_permute2x128_si256(x0, x1, 0x20);
+  __m256i y1 = _mm256_permute2x128_si256(x0, x1, 0x31);
+
+  _mm256_storeu_si256((__m256i *)out, y0);
+  _mm256_storeu_si256((__m256i *)(out + 8), y1);
+#else
+  _mm256_storeu_si256((__m256i *)out, *coeff);
+#endif
+}
+
+#endif  // AOM_DSP_X86_FWD_TXFM_AVX2_H
diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h b/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h
new file mode 100644
index 0000000000..7bb1db70af
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h
@@ -0,0 +1,1014 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_dsp/x86/fwd_txfm_sse2.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+#include "aom_ports/mem.h"
+
+// TODO(jingning) The high bit-depth functions need rework for performance.
+// After we properly fix the high bit-depth function implementations, this
+// file's dependency should be substantially simplified.
+#if DCT_HIGH_BIT_DEPTH
+#define ADD_EPI16 _mm_adds_epi16
+#define SUB_EPI16 _mm_subs_epi16
+
+#else
+#define ADD_EPI16 _mm_add_epi16
+#define SUB_EPI16 _mm_sub_epi16
+#endif
+
+void FDCT4x4_2D(const int16_t *input, tran_low_t *output, int stride) {
+  // This 2D transform implements 4 vertical 1D transforms followed
+  // by 4 horizontal 1D transforms.  The multiplies and adds are as given
+  // by Chen, Smith and Fralick ('77).  The commands for moving the data
+  // around have been minimized by hand.
+  // For the purposes of the comments, the 16 inputs are referred to at i0
+  // through iF (in raster order), intermediate variables are a0, b0, c0
+  // through f, and correspond to the in-place computations mapped to input
+  // locations.  The outputs, o0 through oF are labeled according to the
+  // output locations.
+
+  // Constants
+  // These are the coefficients used for the multiplies.
+  // In the comments, pN means cos(N pi /64) and mN is -cos(N pi /64),
+  // where cospi_N_64 = cos(N pi /64)
+  const __m128i k__cospi_A =
+      octa_set_epi16(cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64,
+                     cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_B =
+      octa_set_epi16(cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64,
+                     cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_C =
+      octa_set_epi16(cospi_8_64, cospi_24_64, cospi_8_64, cospi_24_64,
+                     cospi_24_64, -cospi_8_64, cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_D =
+      octa_set_epi16(cospi_24_64, -cospi_8_64, cospi_24_64, -cospi_8_64,
+                     cospi_8_64, cospi_24_64, cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_E =
+      octa_set_epi16(cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64,
+                     cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_F =
+      octa_set_epi16(cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64,
+                     cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_G =
+      octa_set_epi16(cospi_8_64, cospi_24_64, cospi_8_64, cospi_24_64,
+                     -cospi_8_64, -cospi_24_64, -cospi_8_64, -cospi_24_64);
+  const __m128i k__cospi_H =
+      octa_set_epi16(cospi_24_64, -cospi_8_64, cospi_24_64, -cospi_8_64,
+                     -cospi_24_64, cospi_8_64, -cospi_24_64, cospi_8_64);
+
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // This second rounding constant saves doing some extra adds at the end
+  const __m128i k__DCT_CONST_ROUNDING2 =
+      _mm_set1_epi32(DCT_CONST_ROUNDING + (DCT_CONST_ROUNDING << 1));
+  const int DCT_CONST_BITS2 = DCT_CONST_BITS + 2;
+  const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
+  const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
+  __m128i in0, in1;
+#if DCT_HIGH_BIT_DEPTH
+  __m128i cmp0, cmp1;
+  int test, overflow;
+#endif
+
+  // Load inputs.
+  in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+  in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+  // in0 = [i0 i1 i2 i3 iC iD iE iF]
+  // in1 = [i4 i5 i6 i7 i8 i9 iA iB]
+  in1 = _mm_unpacklo_epi64(
+      in1, _mm_loadl_epi64((const __m128i *)(input + 2 * stride)));
+  in0 = _mm_unpacklo_epi64(
+      in0, _mm_loadl_epi64((const __m128i *)(input + 3 * stride)));
+#if DCT_HIGH_BIT_DEPTH
+  // Check inputs small enough to use optimised code
+  cmp0 = _mm_xor_si128(_mm_cmpgt_epi16(in0, _mm_set1_epi16(0x3ff)),
+                       _mm_cmplt_epi16(in0, _mm_set1_epi16(0xfc00)));
+  cmp1 = _mm_xor_si128(_mm_cmpgt_epi16(in1, _mm_set1_epi16(0x3ff)),
+                       _mm_cmplt_epi16(in1, _mm_set1_epi16(0xfc00)));
+  test = _mm_movemask_epi8(_mm_or_si128(cmp0, cmp1));
+  if (test) {
+    aom_highbd_fdct4x4_c(input, output, stride);
+    return;
+  }
+#endif  // DCT_HIGH_BIT_DEPTH
+
+  // multiply by 16 to give some extra precision
+  in0 = _mm_slli_epi16(in0, 4);
+  in1 = _mm_slli_epi16(in1, 4);
+  // if (i == 0 && input[0]) input[0] += 1;
+  // add 1 to the upper left pixel if it is non-zero, which helps reduce
+  // the round-trip error
+  {
+    // The mask will only contain whether the first value is zero, all
+    // other comparison will fail as something shifted by 4 (above << 4)
+    // can never be equal to one. To increment in the non-zero case, we
+    // add the mask and one for the first element:
+    //   - if zero, mask = -1, v = v - 1 + 1 = v
+    //   - if non-zero, mask = 0, v = v + 0 + 1 = v + 1
+    __m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a);
+    in0 = _mm_add_epi16(in0, mask);
+    in0 = _mm_add_epi16(in0, k__nonzero_bias_b);
+  }
+  // There are 4 total stages, alternating between an add/subtract stage
+  // followed by an multiply-and-add stage.
+  {
+    // Stage 1: Add/subtract
+
+    // in0 = [i0 i1 i2 i3 iC iD iE iF]
+    // in1 = [i4 i5 i6 i7 i8 i9 iA iB]
+    const __m128i r0 = _mm_unpacklo_epi16(in0, in1);
+    const __m128i r1 = _mm_unpackhi_epi16(in0, in1);
+    // r0 = [i0 i4 i1 i5 i2 i6 i3 i7]
+    // r1 = [iC i8 iD i9 iE iA iF iB]
+    const __m128i r2 = _mm_shuffle_epi32(r0, 0xB4);
+    const __m128i r3 = _mm_shuffle_epi32(r1, 0xB4);
+    // r2 = [i0 i4 i1 i5 i3 i7 i2 i6]
+    // r3 = [iC i8 iD i9 iF iB iE iA]
+
+    const __m128i t0 = _mm_add_epi16(r2, r3);
+    const __m128i t1 = _mm_sub_epi16(r2, r3);
+    // t0 = [a0 a4 a1 a5 a3 a7 a2 a6]
+    // t1 = [aC a8 aD a9 aF aB aE aA]
+
+    // Stage 2: multiply by constants (which gets us into 32 bits).
+    // The constants needed here are:
+    // k__cospi_A = [p16 p16 p16 p16 p16 m16 p16 m16]
+    // k__cospi_B = [p16 m16 p16 m16 p16 p16 p16 p16]
+    // k__cospi_C = [p08 p24 p08 p24 p24 m08 p24 m08]
+    // k__cospi_D = [p24 m08 p24 m08 p08 p24 p08 p24]
+    const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A);
+    const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B);
+    const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C);
+    const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D);
+    // Then add and right-shift to get back to 16-bit range
+    const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+    const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+    const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+    const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+    const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+    const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+    const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+    const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+    // w0 = [b0 b1 b7 b6]
+    // w1 = [b8 b9 bF bE]
+    // w2 = [b4 b5 b3 b2]
+    // w3 = [bC bD bB bA]
+    const __m128i x0 = _mm_packs_epi32(w0, w1);
+    const __m128i x1 = _mm_packs_epi32(w2, w3);
+#if DCT_HIGH_BIT_DEPTH
+    overflow = check_epi16_overflow_x2(&x0, &x1);
+    if (overflow) {
+      aom_highbd_fdct4x4_c(input, output, stride);
+      return;
+    }
+#endif  // DCT_HIGH_BIT_DEPTH
+    // x0 = [b0 b1 b7 b6 b8 b9 bF bE]
+    // x1 = [b4 b5 b3 b2 bC bD bB bA]
+    in0 = _mm_shuffle_epi32(x0, 0xD8);
+    in1 = _mm_shuffle_epi32(x1, 0x8D);
+    // in0 = [b0 b1 b8 b9 b7 b6 bF bE]
+    // in1 = [b3 b2 bB bA b4 b5 bC bD]
+  }
+  {
+    // vertical DCTs finished. Now we do the horizontal DCTs.
+    // Stage 3: Add/subtract
+
+    // t0 = [c0 c1 c8 c9  c4  c5  cC  cD]
+    // t1 = [c3 c2 cB cA -c7 -c6 -cF -cE]
+    const __m128i t0 = ADD_EPI16(in0, in1);
+    const __m128i t1 = SUB_EPI16(in0, in1);
+#if DCT_HIGH_BIT_DEPTH
+    overflow = check_epi16_overflow_x2(&t0, &t1);
+    if (overflow) {
+      aom_highbd_fdct4x4_c(input, output, stride);
+      return;
+    }
+#endif  // DCT_HIGH_BIT_DEPTH
+
+    // Stage 4: multiply by constants (which gets us into 32 bits).
+    {
+      // The constants needed here are:
+      // k__cospi_E = [p16 p16 p16 p16 p16 p16 p16 p16]
+      // k__cospi_F = [p16 m16 p16 m16 p16 m16 p16 m16]
+      // k__cospi_G = [p08 p24 p08 p24 m08 m24 m08 m24]
+      // k__cospi_H = [p24 m08 p24 m08 m24 p08 m24 p08]
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E);
+      const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F);
+      const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H);
+      // Then add and right-shift to get back to 16-bit range
+      // but this combines the final right-shift as well to save operations
+      // This unusual rounding operations is to maintain bit-accurate
+      // compatibility with the c version of this function which has two
+      // rounding steps in a row.
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING2);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING2);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING2);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING2);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS2);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS2);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS2);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS2);
+      // w0 = [o0 o4 o8 oC]
+      // w1 = [o2 o6 oA oE]
+      // w2 = [o1 o5 o9 oD]
+      // w3 = [o3 o7 oB oF]
+      // remember the o's are numbered according to the correct output location
+      const __m128i x0 = _mm_packs_epi32(w0, w1);
+      const __m128i x1 = _mm_packs_epi32(w2, w3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x2(&x0, &x1);
+      if (overflow) {
+        aom_highbd_fdct4x4_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      {
+        // x0 = [o0 o4 o8 oC o2 o6 oA oE]
+        // x1 = [o1 o5 o9 oD o3 o7 oB oF]
+        const __m128i y0 = _mm_unpacklo_epi16(x0, x1);
+        const __m128i y1 = _mm_unpackhi_epi16(x0, x1);
+        // y0 = [o0 o1 o4 o5 o8 o9 oC oD]
+        // y1 = [o2 o3 o6 o7 oA oB oE oF]
+        in0 = _mm_unpacklo_epi32(y0, y1);
+        // in0 = [o0 o1 o2 o3 o4 o5 o6 o7]
+        in1 = _mm_unpackhi_epi32(y0, y1);
+        // in1 = [o8 o9 oA oB oC oD oE oF]
+      }
+    }
+  }
+  // Post-condition (v + 1) >> 2 is now incorporated into previous
+  // add and right-shift commands.  Only 2 store instructions needed
+  // because we are using the fact that 1/3 are stored just after 0/2.
+  storeu_output(&in0, output + 0 * 4);
+  storeu_output(&in1, output + 2 * 4);
+}
+
+void FDCT8x8_2D(const int16_t *input, tran_low_t *output, int stride) {
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+#if DCT_HIGH_BIT_DEPTH
+  int overflow;
+#endif
+  // Load input
+  __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = ADD_EPI16(in0, in7);
+    const __m128i q1 = ADD_EPI16(in1, in6);
+    const __m128i q2 = ADD_EPI16(in2, in5);
+    const __m128i q3 = ADD_EPI16(in3, in4);
+    const __m128i q4 = SUB_EPI16(in3, in4);
+    const __m128i q5 = SUB_EPI16(in2, in5);
+    const __m128i q6 = SUB_EPI16(in1, in6);
+    const __m128i q7 = SUB_EPI16(in0, in7);
+#if DCT_HIGH_BIT_DEPTH
+    if (pass == 1) {
+      overflow =
+          check_epi16_overflow_x8(&q0, &q1, &q2, &q3, &q4, &q5, &q6, &q7);
+      if (overflow) {
+        aom_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+    }
+#endif  // DCT_HIGH_BIT_DEPTH
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = ADD_EPI16(q0, q3);
+      const __m128i r1 = ADD_EPI16(q1, q2);
+      const __m128i r2 = SUB_EPI16(q1, q2);
+      const __m128i r3 = SUB_EPI16(q0, q3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3);
+      if (overflow) {
+        aom_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      // Interleave to do the multiply by constants which gets us into 32bits
+      {
+        const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+        const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+        const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+        const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+        const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+        const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+        const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+        const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+        const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+        const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+        const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+        const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+        // dct_const_round_shift
+        const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+        const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+        const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+        const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+        const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+        const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+        const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+        const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+        const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+        const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+        const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+        const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+        const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+        const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+        const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+        const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+        // Combine
+        res0 = _mm_packs_epi32(w0, w1);
+        res4 = _mm_packs_epi32(w2, w3);
+        res2 = _mm_packs_epi32(w4, w5);
+        res6 = _mm_packs_epi32(w6, w7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&res0, &res4, &res2, &res6);
+        if (overflow) {
+          aom_highbd_fdct8x8_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
+      const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
+      const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
+      const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
+      const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
+      const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
+      // dct_const_round_shift
+      const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
+      const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
+      const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
+      const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
+      const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
+      const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
+      const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
+      const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
+      // Combine
+      const __m128i r0 = _mm_packs_epi32(s0, s1);
+      const __m128i r1 = _mm_packs_epi32(s2, s3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x2(&r0, &r1);
+      if (overflow) {
+        aom_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      {
+        // Add/subtract
+        const __m128i x0 = ADD_EPI16(q4, r0);
+        const __m128i x1 = SUB_EPI16(q4, r0);
+        const __m128i x2 = SUB_EPI16(q7, r1);
+        const __m128i x3 = ADD_EPI16(q7, r1);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3);
+        if (overflow) {
+          aom_highbd_fdct8x8_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        // Interleave to do the multiply by constants which gets us into 32bits
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+          const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+          const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+          const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+          const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+          const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+          const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+          const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+          const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+          const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+          // dct_const_round_shift
+          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+          const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+          const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+          const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+          const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+          const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+          const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+          const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+          const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+          // Combine
+          res1 = _mm_packs_epi32(w0, w1);
+          res7 = _mm_packs_epi32(w2, w3);
+          res5 = _mm_packs_epi32(w4, w5);
+          res3 = _mm_packs_epi32(w6, w7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&res1, &res7, &res5, &res3);
+          if (overflow) {
+            aom_highbd_fdct8x8_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+    // store results
+    store_output(&in0, (output + 0 * 8));
+    store_output(&in1, (output + 1 * 8));
+    store_output(&in2, (output + 2 * 8));
+    store_output(&in3, (output + 3 * 8));
+    store_output(&in4, (output + 4 * 8));
+    store_output(&in5, (output + 5 * 8));
+    store_output(&in6, (output + 6 * 8));
+    store_output(&in7, (output + 7 * 8));
+  }
+}
+
+void FDCT16x16_2D(const int16_t *input, tran_low_t *output, int stride) {
+  // The 2D transform is done with two passes which are actually pretty
+  // similar. In the first one, we transform the columns and transpose
+  // the results. In the second one, we transform the rows. To achieve that,
+  // as the first pass results are transposed, we transpose the columns (that
+  // is the transposed rows) and transpose the results (so that it goes back
+  // in normal/row positions).
+  int pass;
+  // We need an intermediate buffer between passes.
+  DECLARE_ALIGNED(16, int16_t, intermediate[256]);
+  const int16_t *in = input;
+  int16_t *out0 = intermediate;
+  tran_low_t *out1 = output;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
+  const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
+  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
+  const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
+  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kOne = _mm_set1_epi16(1);
+  // Do the two transform/transpose passes
+  for (pass = 0; pass < 2; ++pass) {
+    // We process eight columns (transposed rows in second pass) at a time.
+    int column_start;
+#if DCT_HIGH_BIT_DEPTH
+    int overflow;
+#endif
+    for (column_start = 0; column_start < 16; column_start += 8) {
+      __m128i in00, in01, in02, in03, in04, in05, in06, in07;
+      __m128i in08, in09, in10, in11, in12, in13, in14, in15;
+      __m128i input0, input1, input2, input3, input4, input5, input6, input7;
+      __m128i step1_0, step1_1, step1_2, step1_3;
+      __m128i step1_4, step1_5, step1_6, step1_7;
+      __m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6;
+      __m128i step3_0, step3_1, step3_2, step3_3;
+      __m128i step3_4, step3_5, step3_6, step3_7;
+      __m128i res00, res01, res02, res03, res04, res05, res06, res07;
+      __m128i res08, res09, res10, res11, res12, res13, res14, res15;
+      // Load and pre-condition input.
+      if (0 == pass) {
+        in00 = _mm_load_si128((const __m128i *)(in + 0 * stride));
+        in01 = _mm_load_si128((const __m128i *)(in + 1 * stride));
+        in02 = _mm_load_si128((const __m128i *)(in + 2 * stride));
+        in03 = _mm_load_si128((const __m128i *)(in + 3 * stride));
+        in04 = _mm_load_si128((const __m128i *)(in + 4 * stride));
+        in05 = _mm_load_si128((const __m128i *)(in + 5 * stride));
+        in06 = _mm_load_si128((const __m128i *)(in + 6 * stride));
+        in07 = _mm_load_si128((const __m128i *)(in + 7 * stride));
+        in08 = _mm_load_si128((const __m128i *)(in + 8 * stride));
+        in09 = _mm_load_si128((const __m128i *)(in + 9 * stride));
+        in10 = _mm_load_si128((const __m128i *)(in + 10 * stride));
+        in11 = _mm_load_si128((const __m128i *)(in + 11 * stride));
+        in12 = _mm_load_si128((const __m128i *)(in + 12 * stride));
+        in13 = _mm_load_si128((const __m128i *)(in + 13 * stride));
+        in14 = _mm_load_si128((const __m128i *)(in + 14 * stride));
+        in15 = _mm_load_si128((const __m128i *)(in + 15 * stride));
+        // x = x << 2
+        in00 = _mm_slli_epi16(in00, 2);
+        in01 = _mm_slli_epi16(in01, 2);
+        in02 = _mm_slli_epi16(in02, 2);
+        in03 = _mm_slli_epi16(in03, 2);
+        in04 = _mm_slli_epi16(in04, 2);
+        in05 = _mm_slli_epi16(in05, 2);
+        in06 = _mm_slli_epi16(in06, 2);
+        in07 = _mm_slli_epi16(in07, 2);
+        in08 = _mm_slli_epi16(in08, 2);
+        in09 = _mm_slli_epi16(in09, 2);
+        in10 = _mm_slli_epi16(in10, 2);
+        in11 = _mm_slli_epi16(in11, 2);
+        in12 = _mm_slli_epi16(in12, 2);
+        in13 = _mm_slli_epi16(in13, 2);
+        in14 = _mm_slli_epi16(in14, 2);
+        in15 = _mm_slli_epi16(in15, 2);
+      } else {
+        in00 = _mm_load_si128((const __m128i *)(in + 0 * 16));
+        in01 = _mm_load_si128((const __m128i *)(in + 1 * 16));
+        in02 = _mm_load_si128((const __m128i *)(in + 2 * 16));
+        in03 = _mm_load_si128((const __m128i *)(in + 3 * 16));
+        in04 = _mm_load_si128((const __m128i *)(in + 4 * 16));
+        in05 = _mm_load_si128((const __m128i *)(in + 5 * 16));
+        in06 = _mm_load_si128((const __m128i *)(in + 6 * 16));
+        in07 = _mm_load_si128((const __m128i *)(in + 7 * 16));
+        in08 = _mm_load_si128((const __m128i *)(in + 8 * 16));
+        in09 = _mm_load_si128((const __m128i *)(in + 9 * 16));
+        in10 = _mm_load_si128((const __m128i *)(in + 10 * 16));
+        in11 = _mm_load_si128((const __m128i *)(in + 11 * 16));
+        in12 = _mm_load_si128((const __m128i *)(in + 12 * 16));
+        in13 = _mm_load_si128((const __m128i *)(in + 13 * 16));
+        in14 = _mm_load_si128((const __m128i *)(in + 14 * 16));
+        in15 = _mm_load_si128((const __m128i *)(in + 15 * 16));
+        // x = (x + 1) >> 2
+        in00 = _mm_add_epi16(in00, kOne);
+        in01 = _mm_add_epi16(in01, kOne);
+        in02 = _mm_add_epi16(in02, kOne);
+        in03 = _mm_add_epi16(in03, kOne);
+        in04 = _mm_add_epi16(in04, kOne);
+        in05 = _mm_add_epi16(in05, kOne);
+        in06 = _mm_add_epi16(in06, kOne);
+        in07 = _mm_add_epi16(in07, kOne);
+        in08 = _mm_add_epi16(in08, kOne);
+        in09 = _mm_add_epi16(in09, kOne);
+        in10 = _mm_add_epi16(in10, kOne);
+        in11 = _mm_add_epi16(in11, kOne);
+        in12 = _mm_add_epi16(in12, kOne);
+        in13 = _mm_add_epi16(in13, kOne);
+        in14 = _mm_add_epi16(in14, kOne);
+        in15 = _mm_add_epi16(in15, kOne);
+        in00 = _mm_srai_epi16(in00, 2);
+        in01 = _mm_srai_epi16(in01, 2);
+        in02 = _mm_srai_epi16(in02, 2);
+        in03 = _mm_srai_epi16(in03, 2);
+        in04 = _mm_srai_epi16(in04, 2);
+        in05 = _mm_srai_epi16(in05, 2);
+        in06 = _mm_srai_epi16(in06, 2);
+        in07 = _mm_srai_epi16(in07, 2);
+        in08 = _mm_srai_epi16(in08, 2);
+        in09 = _mm_srai_epi16(in09, 2);
+        in10 = _mm_srai_epi16(in10, 2);
+        in11 = _mm_srai_epi16(in11, 2);
+        in12 = _mm_srai_epi16(in12, 2);
+        in13 = _mm_srai_epi16(in13, 2);
+        in14 = _mm_srai_epi16(in14, 2);
+        in15 = _mm_srai_epi16(in15, 2);
+      }
+      in += 8;
+      // Calculate input for the first 8 results.
+      {
+        input0 = ADD_EPI16(in00, in15);
+        input1 = ADD_EPI16(in01, in14);
+        input2 = ADD_EPI16(in02, in13);
+        input3 = ADD_EPI16(in03, in12);
+        input4 = ADD_EPI16(in04, in11);
+        input5 = ADD_EPI16(in05, in10);
+        input6 = ADD_EPI16(in06, in09);
+        input7 = ADD_EPI16(in07, in08);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&input0, &input1, &input2, &input3,
+                                           &input4, &input5, &input6, &input7);
+        if (overflow) {
+          aom_highbd_fdct16x16_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Calculate input for the next 8 results.
+      {
+        step1_0 = SUB_EPI16(in07, in08);
+        step1_1 = SUB_EPI16(in06, in09);
+        step1_2 = SUB_EPI16(in05, in10);
+        step1_3 = SUB_EPI16(in04, in11);
+        step1_4 = SUB_EPI16(in03, in12);
+        step1_5 = SUB_EPI16(in02, in13);
+        step1_6 = SUB_EPI16(in01, in14);
+        step1_7 = SUB_EPI16(in00, in15);
+#if DCT_HIGH_BIT_DEPTH
+        overflow =
+            check_epi16_overflow_x8(&step1_0, &step1_1, &step1_2, &step1_3,
+                                    &step1_4, &step1_5, &step1_6, &step1_7);
+        if (overflow) {
+          aom_highbd_fdct16x16_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Work on the first eight values; fdct8(input, even_results);
+      {
+        // Add/subtract
+        const __m128i q0 = ADD_EPI16(input0, input7);
+        const __m128i q1 = ADD_EPI16(input1, input6);
+        const __m128i q2 = ADD_EPI16(input2, input5);
+        const __m128i q3 = ADD_EPI16(input3, input4);
+        const __m128i q4 = SUB_EPI16(input3, input4);
+        const __m128i q5 = SUB_EPI16(input2, input5);
+        const __m128i q6 = SUB_EPI16(input1, input6);
+        const __m128i q7 = SUB_EPI16(input0, input7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow =
+            check_epi16_overflow_x8(&q0, &q1, &q2, &q3, &q4, &q5, &q6, &q7);
+        if (overflow) {
+          aom_highbd_fdct16x16_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        // Work on first four results
+        {
+          // Add/subtract
+          const __m128i r0 = ADD_EPI16(q0, q3);
+          const __m128i r1 = ADD_EPI16(q1, q2);
+          const __m128i r2 = SUB_EPI16(q1, q2);
+          const __m128i r3 = SUB_EPI16(q0, q3);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3);
+          if (overflow) {
+            aom_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          // Interleave to do the multiply by constants which gets us
+          // into 32 bits.
+          {
+            const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+            const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+            const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+            const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+            res00 = mult_round_shift(&t0, &t1, &k__cospi_p16_p16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+            res08 = mult_round_shift(&t0, &t1, &k__cospi_p16_m16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+            res04 = mult_round_shift(&t2, &t3, &k__cospi_p24_p08,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+            res12 = mult_round_shift(&t2, &t3, &k__cospi_m08_p24,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+            overflow = check_epi16_overflow_x4(&res00, &res08, &res04, &res12);
+            if (overflow) {
+              aom_highbd_fdct16x16_c(input, output, stride);
+              return;
+            }
+#endif  // DCT_HIGH_BIT_DEPTH
+          }
+        }
+        // Work on next four results
+        {
+          // Interleave to do the multiply by constants which gets us
+          // into 32 bits.
+          const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
+          const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
+          const __m128i r0 =
+              mult_round_shift(&d0, &d1, &k__cospi_p16_m16,
+                               &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          const __m128i r1 =
+              mult_round_shift(&d0, &d1, &k__cospi_p16_p16,
+                               &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x2(&r0, &r1);
+          if (overflow) {
+            aom_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          {
+            // Add/subtract
+            const __m128i x0 = ADD_EPI16(q4, r0);
+            const __m128i x1 = SUB_EPI16(q4, r0);
+            const __m128i x2 = SUB_EPI16(q7, r1);
+            const __m128i x3 = ADD_EPI16(q7, r1);
+#if DCT_HIGH_BIT_DEPTH
+            overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3);
+            if (overflow) {
+              aom_highbd_fdct16x16_c(input, output, stride);
+              return;
+            }
+#endif  // DCT_HIGH_BIT_DEPTH
+            // Interleave to do the multiply by constants which gets us
+            // into 32 bits.
+            {
+              const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+              const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+              const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+              const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+              res02 = mult_round_shift(&t0, &t1, &k__cospi_p28_p04,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+              res14 = mult_round_shift(&t0, &t1, &k__cospi_m04_p28,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+              res10 = mult_round_shift(&t2, &t3, &k__cospi_p12_p20,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+              res06 = mult_round_shift(&t2, &t3, &k__cospi_m20_p12,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+              overflow =
+                  check_epi16_overflow_x4(&res02, &res14, &res10, &res06);
+              if (overflow) {
+                aom_highbd_fdct16x16_c(input, output, stride);
+                return;
+              }
+#endif  // DCT_HIGH_BIT_DEPTH
+            }
+          }
+        }
+      }
+      // Work on the next eight values; step1 -> odd_results
+      {
+        // step 2
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2);
+          const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2);
+          const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3);
+          const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3);
+          step2_2 = mult_round_shift(&t0, &t1, &k__cospi_p16_m16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_3 = mult_round_shift(&t2, &t3, &k__cospi_p16_m16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_5 = mult_round_shift(&t0, &t1, &k__cospi_p16_p16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_4 = mult_round_shift(&t2, &t3, &k__cospi_p16_p16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x4(&step2_2, &step2_3, &step2_5, &step2_4);
+          if (overflow) {
+            aom_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 3
+        {
+          step3_0 = ADD_EPI16(step1_0, step2_3);
+          step3_1 = ADD_EPI16(step1_1, step2_2);
+          step3_2 = SUB_EPI16(step1_1, step2_2);
+          step3_3 = SUB_EPI16(step1_0, step2_3);
+          step3_4 = SUB_EPI16(step1_7, step2_4);
+          step3_5 = SUB_EPI16(step1_6, step2_5);
+          step3_6 = ADD_EPI16(step1_6, step2_5);
+          step3_7 = ADD_EPI16(step1_7, step2_4);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x8(&step3_0, &step3_1, &step3_2, &step3_3,
+                                      &step3_4, &step3_5, &step3_6, &step3_7);
+          if (overflow) {
+            aom_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 4
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6);
+          const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6);
+          const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5);
+          const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
+          step2_1 = mult_round_shift(&t0, &t1, &k__cospi_m08_p24,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_2 = mult_round_shift(&t2, &t3, &k__cospi_p24_p08,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_6 = mult_round_shift(&t0, &t1, &k__cospi_p24_p08,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_5 = mult_round_shift(&t2, &t3, &k__cospi_p08_m24,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x4(&step2_1, &step2_2, &step2_6, &step2_5);
+          if (overflow) {
+            aom_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 5
+        {
+          step1_0 = ADD_EPI16(step3_0, step2_1);
+          step1_1 = SUB_EPI16(step3_0, step2_1);
+          step1_2 = ADD_EPI16(step3_3, step2_2);
+          step1_3 = SUB_EPI16(step3_3, step2_2);
+          step1_4 = SUB_EPI16(step3_4, step2_5);
+          step1_5 = ADD_EPI16(step3_4, step2_5);
+          step1_6 = SUB_EPI16(step3_7, step2_6);
+          step1_7 = ADD_EPI16(step3_7, step2_6);
+#if DCT_HIGH_BIT_DEPTH
+          overflow =
+              check_epi16_overflow_x8(&step1_0, &step1_1, &step1_2, &step1_3,
+                                      &step1_4, &step1_5, &step1_6, &step1_7);
+          if (overflow) {
+            aom_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 6
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7);
+          const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7);
+          const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6);
+          const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6);
+          res01 = mult_round_shift(&t0, &t1, &k__cospi_p30_p02,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res09 = mult_round_shift(&t2, &t3, &k__cospi_p14_p18,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res15 = mult_round_shift(&t0, &t1, &k__cospi_m02_p30,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res07 = mult_round_shift(&t2, &t3, &k__cospi_m18_p14,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&res01, &res09, &res15, &res07);
+          if (overflow) {
+            aom_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5);
+          const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5);
+          const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4);
+          const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4);
+          res05 = mult_round_shift(&t0, &t1, &k__cospi_p22_p10,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res13 = mult_round_shift(&t2, &t3, &k__cospi_p06_p26,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res11 = mult_round_shift(&t0, &t1, &k__cospi_m10_p22,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res03 = mult_round_shift(&t2, &t3, &k__cospi_m26_p06,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&res05, &res13, &res11, &res03);
+          if (overflow) {
+            aom_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+      // Transpose the results, do it as two 8x8 transposes.
+      transpose_and_output8x8(&res00, &res01, &res02, &res03, &res04, &res05,
+                              &res06, &res07, pass, out0, out1);
+      transpose_and_output8x8(&res08, &res09, &res10, &res11, &res12, &res13,
+                              &res14, &res15, pass, out0 + 8, out1 + 8);
+      if (pass == 0) {
+        out0 += 8 * 16;
+      } else {
+        out1 += 8 * 16;
+      }
+    }
+    // Setup in/out for next pass.
+    in = intermediate;
+  }
+}
+
+#undef ADD_EPI16
+#undef SUB_EPI16
diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c
new file mode 100644
index 0000000000..a337e618de
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c
@@ -0,0 +1,273 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/x86/fwd_txfm_sse2.h"
+
+void aom_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride) {
+  __m128i in0, in1;
+  __m128i tmp;
+  const __m128i zero = _mm_setzero_si128();
+  in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+  in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+  in1 = _mm_unpacklo_epi64(
+      in1, _mm_loadl_epi64((const __m128i *)(input + 2 * stride)));
+  in0 = _mm_unpacklo_epi64(
+      in0, _mm_loadl_epi64((const __m128i *)(input + 3 * stride)));
+
+  tmp = _mm_add_epi16(in0, in1);
+  in0 = _mm_unpacklo_epi16(zero, tmp);
+  in1 = _mm_unpackhi_epi16(zero, tmp);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  tmp = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(tmp, zero);
+  in1 = _mm_unpackhi_epi32(tmp, zero);
+
+  tmp = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(tmp, 8);
+
+  in1 = _mm_add_epi32(tmp, in0);
+  in0 = _mm_slli_epi32(in1, 1);
+  output[0] = (tran_low_t)_mm_cvtsi128_si32(in0);
+}
+
+void aom_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride) {
+  __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i u0, u1, sum;
+
+  u0 = _mm_add_epi16(in0, in1);
+  u1 = _mm_add_epi16(in2, in3);
+
+  in0 = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  in1 = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  in2 = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  in3 = _mm_load_si128((const __m128i *)(input + 7 * stride));
+
+  sum = _mm_add_epi16(u0, u1);
+
+  in0 = _mm_add_epi16(in0, in1);
+  in2 = _mm_add_epi16(in2, in3);
+  sum = _mm_add_epi16(sum, in0);
+
+  u0 = _mm_setzero_si128();
+  sum = _mm_add_epi16(sum, in2);
+
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  output[0] = (tran_low_t)_mm_cvtsi128_si32(in1);
+}
+
+void aom_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  __m128i in0, in1, in2, in3;
+  __m128i u0, u1;
+  __m128i sum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 2; ++i) {
+    in0 = _mm_load_si128((const __m128i *)(input + 0 * stride + 0));
+    in1 = _mm_load_si128((const __m128i *)(input + 0 * stride + 8));
+    in2 = _mm_load_si128((const __m128i *)(input + 1 * stride + 0));
+    in3 = _mm_load_si128((const __m128i *)(input + 1 * stride + 8));
+
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0 = _mm_load_si128((const __m128i *)(input + 2 * stride + 0));
+    in1 = _mm_load_si128((const __m128i *)(input + 2 * stride + 8));
+    in2 = _mm_load_si128((const __m128i *)(input + 3 * stride + 0));
+    in3 = _mm_load_si128((const __m128i *)(input + 3 * stride + 8));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0 = _mm_load_si128((const __m128i *)(input + 4 * stride + 0));
+    in1 = _mm_load_si128((const __m128i *)(input + 4 * stride + 8));
+    in2 = _mm_load_si128((const __m128i *)(input + 5 * stride + 0));
+    in3 = _mm_load_si128((const __m128i *)(input + 5 * stride + 8));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0 = _mm_load_si128((const __m128i *)(input + 6 * stride + 0));
+    in1 = _mm_load_si128((const __m128i *)(input + 6 * stride + 8));
+    in2 = _mm_load_si128((const __m128i *)(input + 7 * stride + 0));
+    in3 = _mm_load_si128((const __m128i *)(input + 7 * stride + 8));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    sum = _mm_add_epi16(sum, u1);
+    input += 8 * stride;
+  }
+
+  u0 = _mm_setzero_si128();
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  in1 = _mm_srai_epi32(in1, 1);
+  output[0] = (tran_low_t)_mm_cvtsi128_si32(in1);
+}
+
+void aom_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  __m128i in0, in1, in2, in3;
+  __m128i u0, u1;
+  __m128i sum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 8; ++i) {
+    in0 = _mm_load_si128((const __m128i *)(input + 0));
+    in1 = _mm_load_si128((const __m128i *)(input + 8));
+    in2 = _mm_load_si128((const __m128i *)(input + 16));
+    in3 = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0 = _mm_load_si128((const __m128i *)(input + 0));
+    in1 = _mm_load_si128((const __m128i *)(input + 8));
+    in2 = _mm_load_si128((const __m128i *)(input + 16));
+    in3 = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0 = _mm_load_si128((const __m128i *)(input + 0));
+    in1 = _mm_load_si128((const __m128i *)(input + 8));
+    in2 = _mm_load_si128((const __m128i *)(input + 16));
+    in3 = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0 = _mm_load_si128((const __m128i *)(input + 0));
+    in1 = _mm_load_si128((const __m128i *)(input + 8));
+    in2 = _mm_load_si128((const __m128i *)(input + 16));
+    in3 = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    sum = _mm_add_epi16(sum, u1);
+  }
+
+  u0 = _mm_setzero_si128();
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  in1 = _mm_srai_epi32(in1, 3);
+  output[0] = (tran_low_t)_mm_cvtsi128_si32(in1);
+}
+
+#define DCT_HIGH_BIT_DEPTH 0
+#define FDCT4x4_2D aom_fdct4x4_sse2
+#define FDCT8x8_2D aom_fdct8x8_sse2
+#define FDCT16x16_2D aom_fdct16x16_sse2
+#include "aom_dsp/x86/fwd_txfm_impl_sse2.h"
+#undef FDCT4x4_2D
+#undef FDCT8x8_2D
+#undef FDCT16x16_2D
+
+#define FDCT32x32_2D aom_fdct32x32_rd_sse2
+#define FDCT32x32_HIGH_PRECISION 0
+#include "aom_dsp/x86/fwd_dct32x32_impl_sse2.h"
+#undef FDCT32x32_2D
+#undef FDCT32x32_HIGH_PRECISION
+
+#define FDCT32x32_2D aom_fdct32x32_sse2
+#define FDCT32x32_HIGH_PRECISION 1
+#include "aom_dsp/x86/fwd_dct32x32_impl_sse2.h"  // NOLINT
+#undef FDCT32x32_2D
+#undef FDCT32x32_HIGH_PRECISION
+#undef DCT_HIGH_BIT_DEPTH
+
+#if CONFIG_HIGHBITDEPTH
+#define DCT_HIGH_BIT_DEPTH 1
+#define FDCT4x4_2D aom_highbd_fdct4x4_sse2
+#define FDCT8x8_2D aom_highbd_fdct8x8_sse2
+#define FDCT16x16_2D aom_highbd_fdct16x16_sse2
+#include "aom_dsp/x86/fwd_txfm_impl_sse2.h"  // NOLINT
+#undef FDCT4x4_2D
+#undef FDCT8x8_2D
+#undef FDCT16x16_2D
+
+#define FDCT32x32_2D aom_highbd_fdct32x32_rd_sse2
+#define FDCT32x32_HIGH_PRECISION 0
+#include "aom_dsp/x86/fwd_dct32x32_impl_sse2.h"  // NOLINT
+#undef FDCT32x32_2D
+#undef FDCT32x32_HIGH_PRECISION
+
+#define FDCT32x32_2D aom_highbd_fdct32x32_sse2
+#define FDCT32x32_HIGH_PRECISION 1
+#include "aom_dsp/x86/fwd_dct32x32_impl_sse2.h"  // NOLINT
+#undef FDCT32x32_2D
+#undef FDCT32x32_HIGH_PRECISION
+#undef DCT_HIGH_BIT_DEPTH
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h
new file mode 100644
index 0000000000..26b2db2e02
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h
@@ -0,0 +1,362 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_X86_FWD_TXFM_SSE2_H_
+#define AOM_DSP_X86_FWD_TXFM_SSE2_H_
+
+#include "aom_dsp/x86/txfm_common_intrin.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define pair_set_epi32(a, b) \
+  _mm_set_epi32((int)(b), (int)(a), (int)(b), (int)(a))
+
+static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) {
+  __m128i buf0, buf1;
+  buf0 = _mm_mul_epu32(a, b);
+  a = _mm_srli_epi64(a, 32);
+  b = _mm_srli_epi64(b, 32);
+  buf1 = _mm_mul_epu32(a, b);
+  return _mm_add_epi64(buf0, buf1);
+}
+
+static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) {
+  __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
+  __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
+  return _mm_unpacklo_epi64(buf0, buf1);
+}
+
+static INLINE int check_epi16_overflow_x2(const __m128i *preg0,
+                                          const __m128i *preg1) {
+  const __m128i max_overflow = _mm_set1_epi16(0x7fff);
+  const __m128i min_overflow = _mm_set1_epi16(0x8000);
+  __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
+                              _mm_cmpeq_epi16(*preg0, min_overflow));
+  __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
+                              _mm_cmpeq_epi16(*preg1, min_overflow));
+  cmp0 = _mm_or_si128(cmp0, cmp1);
+  return _mm_movemask_epi8(cmp0);
+}
+
+static INLINE int check_epi16_overflow_x4(const __m128i *preg0,
+                                          const __m128i *preg1,
+                                          const __m128i *preg2,
+                                          const __m128i *preg3) {
+  const __m128i max_overflow = _mm_set1_epi16(0x7fff);
+  const __m128i min_overflow = _mm_set1_epi16(0x8000);
+  __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
+                              _mm_cmpeq_epi16(*preg0, min_overflow));
+  __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
+                              _mm_cmpeq_epi16(*preg1, min_overflow));
+  __m128i cmp2 = _mm_or_si128(_mm_cmpeq_epi16(*preg2, max_overflow),
+                              _mm_cmpeq_epi16(*preg2, min_overflow));
+  __m128i cmp3 = _mm_or_si128(_mm_cmpeq_epi16(*preg3, max_overflow),
+                              _mm_cmpeq_epi16(*preg3, min_overflow));
+  cmp0 = _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3));
+  return _mm_movemask_epi8(cmp0);
+}
+
+static INLINE int check_epi16_overflow_x8(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x12(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
+    const __m128i *preg9, const __m128i *preg10, const __m128i *preg11) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x16(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
+    const __m128i *preg9, const __m128i *preg10, const __m128i *preg11,
+    const __m128i *preg12, const __m128i *preg13, const __m128i *preg14,
+    const __m128i *preg15) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) {
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+    if (!res1) res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
+  }
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x32(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
+    const __m128i *preg9, const __m128i *preg10, const __m128i *preg11,
+    const __m128i *preg12, const __m128i *preg13, const __m128i *preg14,
+    const __m128i *preg15, const __m128i *preg16, const __m128i *preg17,
+    const __m128i *preg18, const __m128i *preg19, const __m128i *preg20,
+    const __m128i *preg21, const __m128i *preg22, const __m128i *preg23,
+    const __m128i *preg24, const __m128i *preg25, const __m128i *preg26,
+    const __m128i *preg27, const __m128i *preg28, const __m128i *preg29,
+    const __m128i *preg30, const __m128i *preg31) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) {
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+    if (!res1) {
+      res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
+      if (!res0) {
+        res0 = check_epi16_overflow_x4(preg16, preg17, preg18, preg19);
+        if (!res1) {
+          res1 = check_epi16_overflow_x4(preg20, preg21, preg22, preg23);
+          if (!res0) {
+            res0 = check_epi16_overflow_x4(preg24, preg25, preg26, preg27);
+            if (!res1)
+              res1 = check_epi16_overflow_x4(preg28, preg29, preg30, preg31);
+          }
+        }
+      }
+    }
+  }
+  return res0 + res1;
+}
+
+static INLINE int k_check_epi32_overflow_4(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *zero) {
+  __m128i minus_one = _mm_set1_epi32(-1);
+  // Check for overflows
+  __m128i reg0_shifted = _mm_slli_epi64(*preg0, 1);
+  __m128i reg1_shifted = _mm_slli_epi64(*preg1, 1);
+  __m128i reg2_shifted = _mm_slli_epi64(*preg2, 1);
+  __m128i reg3_shifted = _mm_slli_epi64(*preg3, 1);
+  __m128i reg0_top_dwords =
+      _mm_shuffle_epi32(reg0_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg1_top_dwords =
+      _mm_shuffle_epi32(reg1_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg2_top_dwords =
+      _mm_shuffle_epi32(reg2_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg3_top_dwords =
+      _mm_shuffle_epi32(reg3_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i top_dwords_01 = _mm_unpacklo_epi64(reg0_top_dwords, reg1_top_dwords);
+  __m128i top_dwords_23 = _mm_unpacklo_epi64(reg2_top_dwords, reg3_top_dwords);
+  __m128i valid_positve_01 = _mm_cmpeq_epi32(top_dwords_01, *zero);
+  __m128i valid_positve_23 = _mm_cmpeq_epi32(top_dwords_23, *zero);
+  __m128i valid_negative_01 = _mm_cmpeq_epi32(top_dwords_01, minus_one);
+  __m128i valid_negative_23 = _mm_cmpeq_epi32(top_dwords_23, minus_one);
+  int overflow_01 =
+      _mm_movemask_epi8(_mm_cmpeq_epi32(valid_positve_01, valid_negative_01));
+  int overflow_23 =
+      _mm_movemask_epi8(_mm_cmpeq_epi32(valid_positve_23, valid_negative_23));
+  return (overflow_01 + overflow_23);
+}
+
+static INLINE int k_check_epi32_overflow_8(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7, const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+  }
+  return overflow;
+}
+
+static INLINE int k_check_epi32_overflow_16(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
+    const __m128i *preg9, const __m128i *preg10, const __m128i *preg11,
+    const __m128i *preg12, const __m128i *preg13, const __m128i *preg14,
+    const __m128i *preg15, const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+    if (!overflow) {
+      overflow = k_check_epi32_overflow_4(preg8, preg9, preg10, preg11, zero);
+      if (!overflow) {
+        overflow =
+            k_check_epi32_overflow_4(preg12, preg13, preg14, preg15, zero);
+      }
+    }
+  }
+  return overflow;
+}
+
+static INLINE int k_check_epi32_overflow_32(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
+    const __m128i *preg9, const __m128i *preg10, const __m128i *preg11,
+    const __m128i *preg12, const __m128i *preg13, const __m128i *preg14,
+    const __m128i *preg15, const __m128i *preg16, const __m128i *preg17,
+    const __m128i *preg18, const __m128i *preg19, const __m128i *preg20,
+    const __m128i *preg21, const __m128i *preg22, const __m128i *preg23,
+    const __m128i *preg24, const __m128i *preg25, const __m128i *preg26,
+    const __m128i *preg27, const __m128i *preg28, const __m128i *preg29,
+    const __m128i *preg30, const __m128i *preg31, const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+    if (!overflow) {
+      overflow = k_check_epi32_overflow_4(preg8, preg9, preg10, preg11, zero);
+      if (!overflow) {
+        overflow =
+            k_check_epi32_overflow_4(preg12, preg13, preg14, preg15, zero);
+        if (!overflow) {
+          overflow =
+              k_check_epi32_overflow_4(preg16, preg17, preg18, preg19, zero);
+          if (!overflow) {
+            overflow =
+                k_check_epi32_overflow_4(preg20, preg21, preg22, preg23, zero);
+            if (!overflow) {
+              overflow = k_check_epi32_overflow_4(preg24, preg25, preg26,
+                                                  preg27, zero);
+              if (!overflow) {
+                overflow = k_check_epi32_overflow_4(preg28, preg29, preg30,
+                                                    preg31, zero);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  return overflow;
+}
+
+static INLINE void store_output(const __m128i *poutput, tran_low_t *dst_ptr) {
+#if CONFIG_HIGHBITDEPTH
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  _mm_store_si128((__m128i *)(dst_ptr), out0);
+  _mm_store_si128((__m128i *)(dst_ptr + 4), out1);
+#else
+  _mm_store_si128((__m128i *)(dst_ptr), *poutput);
+#endif  // CONFIG_HIGHBITDEPTH
+}
+
+static INLINE __m128i mult_round_shift(const __m128i *pin0, const __m128i *pin1,
+                                       const __m128i *pmultiplier,
+                                       const __m128i *prounding, int shift) {
+  const __m128i u0 = _mm_madd_epi16(*pin0, *pmultiplier);
+  const __m128i u1 = _mm_madd_epi16(*pin1, *pmultiplier);
+  const __m128i v0 = _mm_add_epi32(u0, *prounding);
+  const __m128i v1 = _mm_add_epi32(u1, *prounding);
+  const __m128i w0 = _mm_srai_epi32(v0, shift);
+  const __m128i w1 = _mm_srai_epi32(v1, shift);
+  return _mm_packs_epi32(w0, w1);
+}
+
+static INLINE void transpose_and_output8x8(
+    const __m128i *pin00, const __m128i *pin01, const __m128i *pin02,
+    const __m128i *pin03, const __m128i *pin04, const __m128i *pin05,
+    const __m128i *pin06, const __m128i *pin07, int pass, int16_t *out0_ptr,
+    tran_low_t *out1_ptr) {
+  // 00 01 02 03 04 05 06 07
+  // 10 11 12 13 14 15 16 17
+  // 20 21 22 23 24 25 26 27
+  // 30 31 32 33 34 35 36 37
+  // 40 41 42 43 44 45 46 47
+  // 50 51 52 53 54 55 56 57
+  // 60 61 62 63 64 65 66 67
+  // 70 71 72 73 74 75 76 77
+  const __m128i tr0_0 = _mm_unpacklo_epi16(*pin00, *pin01);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(*pin02, *pin03);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(*pin00, *pin01);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(*pin02, *pin03);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(*pin04, *pin05);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(*pin06, *pin07);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(*pin04, *pin05);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(*pin06, *pin07);
+  // 00 10 01 11 02 12 03 13
+  // 20 30 21 31 22 32 23 33
+  // 04 14 05 15 06 16 07 17
+  // 24 34 25 35 26 36 27 37
+  // 40 50 41 51 42 52 43 53
+  // 60 70 61 71 62 72 63 73
+  // 54 54 55 55 56 56 57 57
+  // 64 74 65 75 66 76 67 77
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+  // 00 10 20 30 01 11 21 31
+  // 40 50 60 70 41 51 61 71
+  // 02 12 22 32 03 13 23 33
+  // 42 52 62 72 43 53 63 73
+  // 04 14 24 34 05 15 21 36
+  // 44 54 64 74 45 55 61 76
+  // 06 16 26 36 07 17 27 37
+  // 46 56 66 76 47 57 67 77
+  const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+  const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+  const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+  const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+  const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+  const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+  const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+  const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+  // 00 10 20 30 40 50 60 70
+  // 01 11 21 31 41 51 61 71
+  // 02 12 22 32 42 52 62 72
+  // 03 13 23 33 43 53 63 73
+  // 04 14 24 34 44 54 64 74
+  // 05 15 25 35 45 55 65 75
+  // 06 16 26 36 46 56 66 76
+  // 07 17 27 37 47 57 67 77
+  if (pass == 0) {
+    _mm_storeu_si128((__m128i *)(out0_ptr + 0 * 16), tr2_0);
+    _mm_storeu_si128((__m128i *)(out0_ptr + 1 * 16), tr2_1);
+    _mm_storeu_si128((__m128i *)(out0_ptr + 2 * 16), tr2_2);
+    _mm_storeu_si128((__m128i *)(out0_ptr + 3 * 16), tr2_3);
+    _mm_storeu_si128((__m128i *)(out0_ptr + 4 * 16), tr2_4);
+    _mm_storeu_si128((__m128i *)(out0_ptr + 5 * 16), tr2_5);
+    _mm_storeu_si128((__m128i *)(out0_ptr + 6 * 16), tr2_6);
+    _mm_storeu_si128((__m128i *)(out0_ptr + 7 * 16), tr2_7);
+  } else {
+    storeu_output(&tr2_0, (out1_ptr + 0 * 16));
+    storeu_output(&tr2_1, (out1_ptr + 1 * 16));
+    storeu_output(&tr2_2, (out1_ptr + 2 * 16));
+    storeu_output(&tr2_3, (out1_ptr + 3 * 16));
+    storeu_output(&tr2_4, (out1_ptr + 4 * 16));
+    storeu_output(&tr2_5, (out1_ptr + 5 * 16));
+    storeu_output(&tr2_6, (out1_ptr + 6 * 16));
+    storeu_output(&tr2_7, (out1_ptr + 7 * 16));
+  }
+}
+
+void fdct32_8col(__m128i *in0, __m128i *in1);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_DSP_X86_FWD_TXFM_SSE2_H_
diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm b/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm
new file mode 100644
index 0000000000..8fa1c04d0c
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm
@@ -0,0 +1,204 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the forward transformation. Part
+; of the macro definitions are originally derived from the ffmpeg project.
+; The current version applies to x86 64-bit only.
+
+SECTION_RODATA
+
+pw_11585x2: times 8 dw 23170
+pd_8192:    times 4 dd 8192
+
+%macro TRANSFORM_COEFFS 2
+pw_%1_%2:   dw  %1,  %2,  %1,  %2,  %1,  %2,  %1,  %2
+pw_%2_m%1:  dw  %2, -%1,  %2, -%1,  %2, -%1,  %2, -%1
+%endmacro
+
+TRANSFORM_COEFFS 11585,  11585
+TRANSFORM_COEFFS 15137,   6270
+TRANSFORM_COEFFS 16069,   3196
+TRANSFORM_COEFFS  9102,  13623
+
+SECTION .text
+
+%if ARCH_X86_64
+%macro SUM_SUB 3
+  psubw  m%3, m%1, m%2
+  paddw  m%1, m%2
+  SWAP    %2, %3
+%endmacro
+
+; butterfly operation
+%macro MUL_ADD_2X 6 ; dst1, dst2, src, round, coefs1, coefs2
+  pmaddwd            m%1, m%3, %5
+  pmaddwd            m%2, m%3, %6
+  paddd              m%1,  %4
+  paddd              m%2,  %4
+  psrad              m%1,  14
+  psrad              m%2,  14
+%endmacro
+
+%macro BUTTERFLY_4X 7 ; dst1, dst2, coef1, coef2, round, tmp1, tmp2
+  punpckhwd          m%6, m%2, m%1
+  MUL_ADD_2X         %7,  %6,  %6,  %5, [pw_%4_%3], [pw_%3_m%4]
+  punpcklwd          m%2, m%1
+  MUL_ADD_2X         %1,  %2,  %2,  %5, [pw_%4_%3], [pw_%3_m%4]
+  packssdw           m%1, m%7
+  packssdw           m%2, m%6
+%endmacro
+
+; matrix transpose
+%macro INTERLEAVE_2X 4
+  punpckh%1          m%4, m%2, m%3
+  punpckl%1          m%2, m%3
+  SWAP               %3,  %4
+%endmacro
+
+%macro TRANSPOSE8X8 9
+  INTERLEAVE_2X  wd, %1, %2, %9
+  INTERLEAVE_2X  wd, %3, %4, %9
+  INTERLEAVE_2X  wd, %5, %6, %9
+  INTERLEAVE_2X  wd, %7, %8, %9
+
+  INTERLEAVE_2X  dq, %1, %3, %9
+  INTERLEAVE_2X  dq, %2, %4, %9
+  INTERLEAVE_2X  dq, %5, %7, %9
+  INTERLEAVE_2X  dq, %6, %8, %9
+
+  INTERLEAVE_2X  qdq, %1, %5, %9
+  INTERLEAVE_2X  qdq, %3, %7, %9
+  INTERLEAVE_2X  qdq, %2, %6, %9
+  INTERLEAVE_2X  qdq, %4, %8, %9
+
+  SWAP  %2, %5
+  SWAP  %4, %7
+%endmacro
+
+; 1D forward 8x8 DCT transform
+%macro FDCT8_1D 1
+  SUM_SUB            0,  7,  9
+  SUM_SUB            1,  6,  9
+  SUM_SUB            2,  5,  9
+  SUM_SUB            3,  4,  9
+
+  SUM_SUB            0,  3,  9
+  SUM_SUB            1,  2,  9
+  SUM_SUB            6,  5,  9
+%if %1 == 0
+  SUM_SUB            0,  1,  9
+%endif
+
+  BUTTERFLY_4X       2,  3,  6270,  15137,  m8,  9,  10
+
+  pmulhrsw           m6, m12
+  pmulhrsw           m5, m12
+%if %1 == 0
+  pmulhrsw           m0, m12
+  pmulhrsw           m1, m12
+%else
+  BUTTERFLY_4X       1,  0,  11585, 11585,  m8,  9,  10
+  SWAP               0,  1
+%endif
+
+  SUM_SUB            4,  5,  9
+  SUM_SUB            7,  6,  9
+  BUTTERFLY_4X       4,  7,  3196,  16069,  m8,  9,  10
+  BUTTERFLY_4X       5,  6,  13623,  9102,  m8,  9,  10
+  SWAP               1,  4
+  SWAP               3,  6
+%endmacro
+
+%macro DIVIDE_ROUND_2X 4 ; dst1, dst2, tmp1, tmp2
+  psraw              m%3, m%1, 15
+  psraw              m%4, m%2, 15
+  psubw              m%1, m%3
+  psubw              m%2, m%4
+  psraw              m%1, 1
+  psraw              m%2, 1
+%endmacro
+
+%macro STORE_OUTPUT 2 ; index, result
+%if CONFIG_HIGHBITDEPTH
+  ; const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  ; __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  ; __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  ; _mm_store_si128((__m128i *)(dst_ptr), out0);
+  ; _mm_store_si128((__m128i *)(dst_ptr + 4), out1);
+  pxor               m11, m11
+  pcmpgtw            m11, m%2
+  movdqa             m12, m%2
+  punpcklwd          m%2, m11
+  punpckhwd          m12, m11
+  mova               [outputq + 4*%1 +  0], m%2
+  mova               [outputq + 4*%1 + 16], m12
+%else
+  mova               [outputq + 2*%1], m%2
+%endif
+%endmacro
+
+INIT_XMM ssse3
+cglobal fdct8x8, 3, 5, 13, input, output, stride
+
+  mova               m8, [pd_8192]
+  mova              m12, [pw_11585x2]
+
+  lea                r3, [2 * strideq]
+  lea                r4, [4 * strideq]
+  mova               m0, [inputq]
+  mova               m1, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m2, [inputq]
+  mova               m3, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m4, [inputq]
+  mova               m5, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m6, [inputq]
+  mova               m7, [inputq + r3]
+
+  ; left shift by 2 to increase forward transformation precision
+  psllw              m0, 2
+  psllw              m1, 2
+  psllw              m2, 2
+  psllw              m3, 2
+  psllw              m4, 2
+  psllw              m5, 2
+  psllw              m6, 2
+  psllw              m7, 2
+
+  ; column transform
+  FDCT8_1D  0
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  FDCT8_1D  1
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  DIVIDE_ROUND_2X   0, 1, 9, 10
+  DIVIDE_ROUND_2X   2, 3, 9, 10
+  DIVIDE_ROUND_2X   4, 5, 9, 10
+  DIVIDE_ROUND_2X   6, 7, 9, 10
+
+  STORE_OUTPUT       0, 0
+  STORE_OUTPUT       8, 1
+  STORE_OUTPUT      16, 2
+  STORE_OUTPUT      24, 3
+  STORE_OUTPUT      32, 4
+  STORE_OUTPUT      40, 5
+  STORE_OUTPUT      48, 6
+  STORE_OUTPUT      56, 7
+
+  RET
+%endif
diff --git a/third_party/aom/aom_dsp/x86/halfpix_variance_impl_sse2.asm b/third_party/aom/aom_dsp/x86/halfpix_variance_impl_sse2.asm
new file mode 100644
index 0000000000..60446b0869
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/halfpix_variance_impl_sse2.asm
@@ -0,0 +1,349 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "aom_ports/x86_abi_support.asm"
+
+;void aom_half_horiz_vert_variance16x_h_sse2(unsigned char *ref,
+;                                            int ref_stride,
+;                                            unsigned char *src,
+;                                            int src_stride,
+;                                            unsigned int height,
+;                                            int *sum,
+;                                            unsigned int *sumsquared)
+global sym(aom_half_horiz_vert_variance16x_h_sse2) PRIVATE
+sym(aom_half_horiz_vert_variance16x_h_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        pxor            xmm6,           xmm6                ;  error accumulator
+        pxor            xmm7,           xmm7                ;  sse eaccumulator
+        mov             rsi,            arg(0) ;ref
+
+        mov             rdi,            arg(2) ;src
+        movsxd          rcx,            dword ptr arg(4) ;height
+        movsxd          rax,            dword ptr arg(1) ;ref_stride
+        movsxd          rdx,            dword ptr arg(3)    ;src_stride
+
+        pxor            xmm0,           xmm0                ;
+
+        movdqu          xmm5,           XMMWORD PTR [rsi]
+        movdqu          xmm3,           XMMWORD PTR [rsi+1]
+        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3) horizontal line 1
+
+        lea             rsi,            [rsi + rax]
+
+aom_half_horiz_vert_variance16x_h_1:
+        movdqu          xmm1,           XMMWORD PTR [rsi]     ;
+        movdqu          xmm2,           XMMWORD PTR [rsi+1]   ;
+        pavgb           xmm1,           xmm2                ;  xmm1 = avg(xmm1,xmm3) horizontal line i+1
+
+        pavgb           xmm5,           xmm1                ;  xmm = vertical average of the above
+
+        movdqa          xmm4,           xmm5
+        punpcklbw       xmm5,           xmm0                ;  xmm5 = words of above
+        punpckhbw       xmm4,           xmm0
+
+        movq            xmm3,           QWORD PTR [rdi]     ;  xmm3 = d0,d1,d2..d7
+        punpcklbw       xmm3,           xmm0                ;  xmm3 = words of above
+        psubw           xmm5,           xmm3                ;  xmm5 -= xmm3
+
+        movq            xmm3,           QWORD PTR [rdi+8]
+        punpcklbw       xmm3,           xmm0
+        psubw           xmm4,           xmm3
+
+        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
+        paddw           xmm6,           xmm4
+        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
+        pmaddwd         xmm4,           xmm4
+        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
+        paddd           xmm7,           xmm4
+
+        movdqa          xmm5,           xmm1                ;  save xmm1 for use on the next row
+
+        lea             rsi,            [rsi + rax]
+        lea             rdi,            [rdi + rdx]
+
+        sub             rcx,            1                   ;
+        jnz             aom_half_horiz_vert_variance16x_h_1     ;
+
+        pxor        xmm1,           xmm1
+        pxor        xmm5,           xmm5
+
+        punpcklwd   xmm0,           xmm6
+        punpckhwd   xmm1,           xmm6
+        psrad       xmm0,           16
+        psrad       xmm1,           16
+        paddd       xmm0,           xmm1
+        movdqa      xmm1,           xmm0
+
+        movdqa      xmm6,           xmm7
+        punpckldq   xmm6,           xmm5
+        punpckhdq   xmm7,           xmm5
+        paddd       xmm6,           xmm7
+
+        punpckldq   xmm0,           xmm5
+        punpckhdq   xmm1,           xmm5
+        paddd       xmm0,           xmm1
+
+        movdqa      xmm7,           xmm6
+        movdqa      xmm1,           xmm0
+
+        psrldq      xmm7,           8
+        psrldq      xmm1,           8
+
+        paddd       xmm6,           xmm7
+        paddd       xmm0,           xmm1
+
+        mov         rsi,            arg(5) ;[Sum]
+        mov         rdi,            arg(6) ;[SSE]
+
+        movd        [rsi],       xmm0
+        movd        [rdi],       xmm6
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void aom_half_vert_variance16x_h_sse2(unsigned char *ref,
+;                                      int ref_stride,
+;                                      unsigned char *src,
+;                                      int src_stride,
+;                                      unsigned int height,
+;                                      int *sum,
+;                                      unsigned int *sumsquared)
+global sym(aom_half_vert_variance16x_h_sse2) PRIVATE
+sym(aom_half_vert_variance16x_h_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        pxor            xmm6,           xmm6                ;  error accumulator
+        pxor            xmm7,           xmm7                ;  sse eaccumulator
+        mov             rsi,            arg(0)              ;ref
+
+        mov             rdi,            arg(2)              ;src
+        movsxd          rcx,            dword ptr arg(4)    ;height
+        movsxd          rax,            dword ptr arg(1)    ;ref_stride
+        movsxd          rdx,            dword ptr arg(3)    ;src_stride
+
+        movdqu          xmm5,           XMMWORD PTR [rsi]
+        lea             rsi,            [rsi + rax          ]
+        pxor            xmm0,           xmm0
+
+aom_half_vert_variance16x_h_1:
+        movdqu          xmm3,           XMMWORD PTR [rsi]
+
+        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3)
+        movdqa          xmm4,           xmm5
+        punpcklbw       xmm5,           xmm0
+        punpckhbw       xmm4,           xmm0
+
+        movq            xmm2,           QWORD PTR [rdi]
+        punpcklbw       xmm2,           xmm0
+        psubw           xmm5,           xmm2
+        movq            xmm2,           QWORD PTR [rdi+8]
+        punpcklbw       xmm2,           xmm0
+        psubw           xmm4,           xmm2
+
+        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
+        paddw           xmm6,           xmm4
+        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
+        pmaddwd         xmm4,           xmm4
+        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
+        paddd           xmm7,           xmm4
+
+        movdqa          xmm5,           xmm3
+
+        lea             rsi,            [rsi + rax]
+        lea             rdi,            [rdi + rdx]
+
+        sub             rcx,            1
+        jnz             aom_half_vert_variance16x_h_1
+
+        pxor        xmm1,           xmm1
+        pxor        xmm5,           xmm5
+
+        punpcklwd   xmm0,           xmm6
+        punpckhwd   xmm1,           xmm6
+        psrad       xmm0,           16
+        psrad       xmm1,           16
+        paddd       xmm0,           xmm1
+        movdqa      xmm1,           xmm0
+
+        movdqa      xmm6,           xmm7
+        punpckldq   xmm6,           xmm5
+        punpckhdq   xmm7,           xmm5
+        paddd       xmm6,           xmm7
+
+        punpckldq   xmm0,           xmm5
+        punpckhdq   xmm1,           xmm5
+        paddd       xmm0,           xmm1
+
+        movdqa      xmm7,           xmm6
+        movdqa      xmm1,           xmm0
+
+        psrldq      xmm7,           8
+        psrldq      xmm1,           8
+
+        paddd       xmm6,           xmm7
+        paddd       xmm0,           xmm1
+
+        mov         rsi,            arg(5) ;[Sum]
+        mov         rdi,            arg(6) ;[SSE]
+
+        movd        [rsi],       xmm0
+        movd        [rdi],       xmm6
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void aom_half_horiz_variance16x_h_sse2(unsigned char *ref,
+;                                       int ref_stride
+;                                       unsigned char *src,
+;                                       int src_stride,
+;                                       unsigned int height,
+;                                       int *sum,
+;                                       unsigned int *sumsquared)
+global sym(aom_half_horiz_variance16x_h_sse2) PRIVATE
+sym(aom_half_horiz_variance16x_h_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        pxor            xmm6,           xmm6                ;  error accumulator
+        pxor            xmm7,           xmm7                ;  sse eaccumulator
+        mov             rsi,            arg(0) ;ref
+
+        mov             rdi,            arg(2) ;src
+        movsxd          rcx,            dword ptr arg(4) ;height
+        movsxd          rax,            dword ptr arg(1) ;ref_stride
+        movsxd          rdx,            dword ptr arg(3)    ;src_stride
+
+        pxor            xmm0,           xmm0                ;
+
+aom_half_horiz_variance16x_h_1:
+        movdqu          xmm5,           XMMWORD PTR [rsi]     ;  xmm5 = s0,s1,s2..s15
+        movdqu          xmm3,           XMMWORD PTR [rsi+1]   ;  xmm3 = s1,s2,s3..s16
+
+        pavgb           xmm5,           xmm3                ;  xmm5 = avg(xmm1,xmm3)
+        movdqa          xmm1,           xmm5
+        punpcklbw       xmm5,           xmm0                ;  xmm5 = words of above
+        punpckhbw       xmm1,           xmm0
+
+        movq            xmm3,           QWORD PTR [rdi]     ;  xmm3 = d0,d1,d2..d7
+        punpcklbw       xmm3,           xmm0                ;  xmm3 = words of above
+        movq            xmm2,           QWORD PTR [rdi+8]
+        punpcklbw       xmm2,           xmm0
+
+        psubw           xmm5,           xmm3                ;  xmm5 -= xmm3
+        psubw           xmm1,           xmm2
+        paddw           xmm6,           xmm5                ;  xmm6 += accumulated column differences
+        paddw           xmm6,           xmm1
+        pmaddwd         xmm5,           xmm5                ;  xmm5 *= xmm5
+        pmaddwd         xmm1,           xmm1
+        paddd           xmm7,           xmm5                ;  xmm7 += accumulated square column differences
+        paddd           xmm7,           xmm1
+
+        lea             rsi,            [rsi + rax]
+        lea             rdi,            [rdi + rdx]
+
+        sub             rcx,            1                   ;
+        jnz             aom_half_horiz_variance16x_h_1        ;
+
+        pxor        xmm1,           xmm1
+        pxor        xmm5,           xmm5
+
+        punpcklwd   xmm0,           xmm6
+        punpckhwd   xmm1,           xmm6
+        psrad       xmm0,           16
+        psrad       xmm1,           16
+        paddd       xmm0,           xmm1
+        movdqa      xmm1,           xmm0
+
+        movdqa      xmm6,           xmm7
+        punpckldq   xmm6,           xmm5
+        punpckhdq   xmm7,           xmm5
+        paddd       xmm6,           xmm7
+
+        punpckldq   xmm0,           xmm5
+        punpckhdq   xmm1,           xmm5
+        paddd       xmm0,           xmm1
+
+        movdqa      xmm7,           xmm6
+        movdqa      xmm1,           xmm0
+
+        psrldq      xmm7,           8
+        psrldq      xmm1,           8
+
+        paddd       xmm6,           xmm7
+        paddd       xmm0,           xmm1
+
+        mov         rsi,            arg(5) ;[Sum]
+        mov         rdi,            arg(6) ;[SSE]
+
+        movd        [rsi],       xmm0
+        movd        [rdi],       xmm6
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+;    short xmm_bi_rd[8] = { 64, 64, 64, 64,64, 64, 64, 64};
+align 16
+xmm_bi_rd:
+    times 8 dw 64
+align 16
+aom_bilinear_filters_sse2:
+    dw 128, 128, 128, 128, 128, 128, 128, 128,  0,  0,  0,  0,  0,  0,  0,  0
+    dw 112, 112, 112, 112, 112, 112, 112, 112, 16, 16, 16, 16, 16, 16, 16, 16
+    dw 96, 96, 96, 96, 96, 96, 96, 96, 32, 32, 32, 32, 32, 32, 32, 32
+    dw 80, 80, 80, 80, 80, 80, 80, 80, 48, 48, 48, 48, 48, 48, 48, 48
+    dw 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64
+    dw 48, 48, 48, 48, 48, 48, 48, 48, 80, 80, 80, 80, 80, 80, 80, 80
+    dw 32, 32, 32, 32, 32, 32, 32, 32, 96, 96, 96, 96, 96, 96, 96, 96
+    dw 16, 16, 16, 16, 16, 16, 16, 16, 112, 112, 112, 112, 112, 112, 112, 112
diff --git a/third_party/aom/aom_dsp/x86/halfpix_variance_sse2.c b/third_party/aom/aom_dsp/x86/halfpix_variance_sse2.c
new file mode 100644
index 0000000000..a99c0b40e0
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/halfpix_variance_sse2.c
@@ -0,0 +1,77 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+void aom_half_horiz_vert_variance16x_h_sse2(const unsigned char *ref,
+                                            int ref_stride,
+                                            const unsigned char *src,
+                                            int src_stride, unsigned int height,
+                                            int *sum, unsigned int *sumsquared);
+void aom_half_horiz_variance16x_h_sse2(const unsigned char *ref, int ref_stride,
+                                       const unsigned char *src, int src_stride,
+                                       unsigned int height, int *sum,
+                                       unsigned int *sumsquared);
+void aom_half_vert_variance16x_h_sse2(const unsigned char *ref, int ref_stride,
+                                      const unsigned char *src, int src_stride,
+                                      unsigned int height, int *sum,
+                                      unsigned int *sumsquared);
+
+uint32_t aom_variance_halfpixvar16x16_h_sse2(const unsigned char *src,
+                                             int src_stride,
+                                             const unsigned char *dst,
+                                             int dst_stride, uint32_t *sse) {
+  int xsum0;
+  unsigned int xxsum0;
+
+  aom_half_horiz_variance16x_h_sse2(src, src_stride, dst, dst_stride, 16,
+                                    &xsum0, &xxsum0);
+
+  *sse = xxsum0;
+  assert(xsum0 <= 255 * 16 * 16);
+  assert(xsum0 >= -255 * 16 * 16);
+  return (xxsum0 - ((uint32_t)((int64_t)xsum0 * xsum0) >> 8));
+}
+
+uint32_t aom_variance_halfpixvar16x16_v_sse2(const unsigned char *src,
+                                             int src_stride,
+                                             const unsigned char *dst,
+                                             int dst_stride, uint32_t *sse) {
+  int xsum0;
+  unsigned int xxsum0;
+  aom_half_vert_variance16x_h_sse2(src, src_stride, dst, dst_stride, 16, &xsum0,
+                                   &xxsum0);
+
+  *sse = xxsum0;
+  assert(xsum0 <= 255 * 16 * 16);
+  assert(xsum0 >= -255 * 16 * 16);
+  return (xxsum0 - ((uint32_t)((int64_t)xsum0 * xsum0) >> 8));
+}
+
+uint32_t aom_variance_halfpixvar16x16_hv_sse2(const unsigned char *src,
+                                              int src_stride,
+                                              const unsigned char *dst,
+                                              int dst_stride, uint32_t *sse) {
+  int xsum0;
+  unsigned int xxsum0;
+
+  aom_half_horiz_vert_variance16x_h_sse2(src, src_stride, dst, dst_stride, 16,
+                                         &xsum0, &xxsum0);
+
+  *sse = xxsum0;
+  assert(xsum0 <= 255 * 16 * 16);
+  assert(xsum0 >= -255 * 16 * 16);
+  return (xxsum0 - ((uint32_t)((int64_t)xsum0 * xsum0) >> 8));
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c
new file mode 100644
index 0000000000..7d96e26ae4
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c
@@ -0,0 +1,1151 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <immintrin.h>
+#include <string.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/x86/convolve.h"
+
+#define CONV8_ROUNDING_BITS (7)
+
+static const uint8_t signal_pattern_0[32] = { 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6,
+                                              7, 6, 7, 8, 9, 0, 1, 2, 3, 2, 3,
+                                              4, 5, 4, 5, 6, 7, 6, 7, 8, 9 };
+
+static const uint8_t signal_pattern_1[32] = { 4, 5, 6,  7,  6,  7,  8,  9,
+                                              8, 9, 10, 11, 10, 11, 12, 13,
+                                              4, 5, 6,  7,  6,  7,  8,  9,
+                                              8, 9, 10, 11, 10, 11, 12, 13 };
+
+static const uint8_t signal_pattern_2[32] = { 6,  7,  8,  9,  8,  9,  10, 11,
+                                              10, 11, 12, 13, 12, 13, 14, 15,
+                                              6,  7,  8,  9,  8,  9,  10, 11,
+                                              10, 11, 12, 13, 12, 13, 14, 15 };
+
+static const uint32_t signal_index[8] = { 2, 3, 4, 5, 2, 3, 4, 5 };
+
+typedef enum { PACK_8x1, PACK_8x2, PACK_16x1 } PixelPackFormat;
+
+typedef void (*WritePixels)(const __m256i *y0, const __m256i *y1,
+                            const __m256i *mask, uint16_t *dst,
+                            ptrdiff_t pitch);
+
+// -----------------------------------------------------------------------------
+// Copy and average
+
+void aom_highbd_convolve_copy_avx2(const uint8_t *src8, ptrdiff_t src_stride,
+                                   uint8_t *dst8, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int filter_x_stride,
+                                   const int16_t *filter_y, int filter_y_stride,
+                                   int width, int h, int bd) {
+  const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  assert(width % 4 == 0);
+  if (width > 32) {  // width = 64
+    do {
+      const __m256i p0 = _mm256_loadu_si256((const __m256i *)src);
+      const __m256i p1 = _mm256_loadu_si256((const __m256i *)(src + 16));
+      const __m256i p2 = _mm256_loadu_si256((const __m256i *)(src + 32));
+      const __m256i p3 = _mm256_loadu_si256((const __m256i *)(src + 48));
+      src += src_stride;
+      _mm256_storeu_si256((__m256i *)dst, p0);
+      _mm256_storeu_si256((__m256i *)(dst + 16), p1);
+      _mm256_storeu_si256((__m256i *)(dst + 32), p2);
+      _mm256_storeu_si256((__m256i *)(dst + 48), p3);
+      dst += dst_stride;
+      h--;
+    } while (h > 0);
+  } else if (width > 16) {  // width = 32
+    do {
+      const __m256i p0 = _mm256_loadu_si256((const __m256i *)src);
+      const __m256i p1 = _mm256_loadu_si256((const __m256i *)(src + 16));
+      src += src_stride;
+      _mm256_storeu_si256((__m256i *)dst, p0);
+      _mm256_storeu_si256((__m256i *)(dst + 16), p1);
+      dst += dst_stride;
+      h--;
+    } while (h > 0);
+  } else if (width > 8) {  // width = 16
+    __m256i p0, p1;
+    do {
+      p0 = _mm256_loadu_si256((const __m256i *)src);
+      src += src_stride;
+      p1 = _mm256_loadu_si256((const __m256i *)src);
+      src += src_stride;
+
+      _mm256_storeu_si256((__m256i *)dst, p0);
+      dst += dst_stride;
+      _mm256_storeu_si256((__m256i *)dst, p1);
+      dst += dst_stride;
+      h -= 2;
+    } while (h > 0);
+  } else if (width > 4) {  // width = 8
+    __m128i p0, p1;
+    do {
+      p0 = _mm_loadu_si128((const __m128i *)src);
+      src += src_stride;
+      p1 = _mm_loadu_si128((const __m128i *)src);
+      src += src_stride;
+
+      _mm_storeu_si128((__m128i *)dst, p0);
+      dst += dst_stride;
+      _mm_storeu_si128((__m128i *)dst, p1);
+      dst += dst_stride;
+      h -= 2;
+    } while (h > 0);
+  } else {  // width = 4
+    __m128i p0, p1;
+    do {
+      p0 = _mm_loadl_epi64((const __m128i *)src);
+      src += src_stride;
+      p1 = _mm_loadl_epi64((const __m128i *)src);
+      src += src_stride;
+
+      _mm_storel_epi64((__m128i *)dst, p0);
+      dst += dst_stride;
+      _mm_storel_epi64((__m128i *)dst, p1);
+      dst += dst_stride;
+      h -= 2;
+    } while (h > 0);
+  }
+}
+
+void aom_highbd_convolve_avg_avx2(const uint8_t *src8, ptrdiff_t src_stride,
+                                  uint8_t *dst8, ptrdiff_t dst_stride,
+                                  const int16_t *filter_x, int filter_x_stride,
+                                  const int16_t *filter_y, int filter_y_stride,
+                                  int width, int h, int bd) {
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  assert(width % 4 == 0);
+  if (width > 32) {  // width = 64
+    __m256i p0, p1, p2, p3, u0, u1, u2, u3;
+    do {
+      p0 = _mm256_loadu_si256((const __m256i *)src);
+      p1 = _mm256_loadu_si256((const __m256i *)(src + 16));
+      p2 = _mm256_loadu_si256((const __m256i *)(src + 32));
+      p3 = _mm256_loadu_si256((const __m256i *)(src + 48));
+      src += src_stride;
+      u0 = _mm256_loadu_si256((const __m256i *)dst);
+      u1 = _mm256_loadu_si256((const __m256i *)(dst + 16));
+      u2 = _mm256_loadu_si256((const __m256i *)(dst + 32));
+      u3 = _mm256_loadu_si256((const __m256i *)(dst + 48));
+      _mm256_storeu_si256((__m256i *)dst, _mm256_avg_epu16(p0, u0));
+      _mm256_storeu_si256((__m256i *)(dst + 16), _mm256_avg_epu16(p1, u1));
+      _mm256_storeu_si256((__m256i *)(dst + 32), _mm256_avg_epu16(p2, u2));
+      _mm256_storeu_si256((__m256i *)(dst + 48), _mm256_avg_epu16(p3, u3));
+      dst += dst_stride;
+      h--;
+    } while (h > 0);
+  } else if (width > 16) {  // width = 32
+    __m256i p0, p1, u0, u1;
+    do {
+      p0 = _mm256_loadu_si256((const __m256i *)src);
+      p1 = _mm256_loadu_si256((const __m256i *)(src + 16));
+      src += src_stride;
+      u0 = _mm256_loadu_si256((const __m256i *)dst);
+      u1 = _mm256_loadu_si256((const __m256i *)(dst + 16));
+      _mm256_storeu_si256((__m256i *)dst, _mm256_avg_epu16(p0, u0));
+      _mm256_storeu_si256((__m256i *)(dst + 16), _mm256_avg_epu16(p1, u1));
+      dst += dst_stride;
+      h--;
+    } while (h > 0);
+  } else if (width > 8) {  // width = 16
+    __m256i p0, p1, u0, u1;
+    do {
+      p0 = _mm256_loadu_si256((const __m256i *)src);
+      p1 = _mm256_loadu_si256((const __m256i *)(src + src_stride));
+      src += src_stride << 1;
+      u0 = _mm256_loadu_si256((const __m256i *)dst);
+      u1 = _mm256_loadu_si256((const __m256i *)(dst + dst_stride));
+
+      _mm256_storeu_si256((__m256i *)dst, _mm256_avg_epu16(p0, u0));
+      _mm256_storeu_si256((__m256i *)(dst + dst_stride),
+                          _mm256_avg_epu16(p1, u1));
+      dst += dst_stride << 1;
+      h -= 2;
+    } while (h > 0);
+  } else if (width > 4) {  // width = 8
+    __m128i p0, p1, u0, u1;
+    do {
+      p0 = _mm_loadu_si128((const __m128i *)src);
+      p1 = _mm_loadu_si128((const __m128i *)(src + src_stride));
+      src += src_stride << 1;
+      u0 = _mm_loadu_si128((const __m128i *)dst);
+      u1 = _mm_loadu_si128((const __m128i *)(dst + dst_stride));
+
+      _mm_storeu_si128((__m128i *)dst, _mm_avg_epu16(p0, u0));
+      _mm_storeu_si128((__m128i *)(dst + dst_stride), _mm_avg_epu16(p1, u1));
+      dst += dst_stride << 1;
+      h -= 2;
+    } while (h > 0);
+  } else {  // width = 4
+    __m128i p0, p1, u0, u1;
+    do {
+      p0 = _mm_loadl_epi64((const __m128i *)src);
+      p1 = _mm_loadl_epi64((const __m128i *)(src + src_stride));
+      src += src_stride << 1;
+      u0 = _mm_loadl_epi64((const __m128i *)dst);
+      u1 = _mm_loadl_epi64((const __m128i *)(dst + dst_stride));
+
+      _mm_storel_epi64((__m128i *)dst, _mm_avg_epu16(u0, p0));
+      _mm_storel_epi64((__m128i *)(dst + dst_stride), _mm_avg_epu16(u1, p1));
+      dst += dst_stride << 1;
+      h -= 2;
+    } while (h > 0);
+  }
+}
+
+// -----------------------------------------------------------------------------
+// Horizontal Filtering
+
+static INLINE void pack_pixels(const __m256i *s, __m256i *p /*p[4]*/) {
+  const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index);
+  const __m256i sf0 = _mm256_loadu_si256((const __m256i *)signal_pattern_0);
+  const __m256i sf1 = _mm256_loadu_si256((const __m256i *)signal_pattern_1);
+  const __m256i c = _mm256_permutevar8x32_epi32(*s, idx);
+
+  p[0] = _mm256_shuffle_epi8(*s, sf0);  // x0x6
+  p[1] = _mm256_shuffle_epi8(*s, sf1);  // x1x7
+  p[2] = _mm256_shuffle_epi8(c, sf0);   // x2x4
+  p[3] = _mm256_shuffle_epi8(c, sf1);   // x3x5
+}
+
+// Note:
+//  Shared by 8x2 and 16x1 block
+static INLINE void pack_16_pixels(const __m256i *s0, const __m256i *s1,
+                                  __m256i *x /*x[8]*/) {
+  __m256i pp[8];
+  pack_pixels(s0, pp);
+  pack_pixels(s1, &pp[4]);
+  x[0] = _mm256_permute2x128_si256(pp[0], pp[4], 0x20);
+  x[1] = _mm256_permute2x128_si256(pp[1], pp[5], 0x20);
+  x[2] = _mm256_permute2x128_si256(pp[2], pp[6], 0x20);
+  x[3] = _mm256_permute2x128_si256(pp[3], pp[7], 0x20);
+  x[4] = x[2];
+  x[5] = x[3];
+  x[6] = _mm256_permute2x128_si256(pp[0], pp[4], 0x31);
+  x[7] = _mm256_permute2x128_si256(pp[1], pp[5], 0x31);
+}
+
+static INLINE void pack_pixels_with_format(const uint16_t *src,
+                                           PixelPackFormat fmt,
+                                           ptrdiff_t stride, __m256i *x) {
+  switch (fmt) {
+    case PACK_8x1: {
+      __m256i pp[8];
+      __m256i s0;
+      s0 = _mm256_loadu_si256((const __m256i *)src);
+      pack_pixels(&s0, pp);
+      x[0] = _mm256_permute2x128_si256(pp[0], pp[2], 0x30);
+      x[1] = _mm256_permute2x128_si256(pp[1], pp[3], 0x30);
+      x[2] = _mm256_permute2x128_si256(pp[2], pp[0], 0x30);
+      x[3] = _mm256_permute2x128_si256(pp[3], pp[1], 0x30);
+      break;
+    }
+    case PACK_8x2: {
+      __m256i s0, s1;
+      s0 = _mm256_loadu_si256((const __m256i *)src);
+      s1 = _mm256_loadu_si256((const __m256i *)(src + stride));
+      pack_16_pixels(&s0, &s1, x);
+      break;
+    }
+    case PACK_16x1: {
+      __m256i s0, s1;
+      s0 = _mm256_loadu_si256((const __m256i *)src);
+      s1 = _mm256_loadu_si256((const __m256i *)(src + 8));
+      pack_16_pixels(&s0, &s1, x);
+      break;
+    }
+    default: { assert(0); }
+  }
+}
+
+static INLINE void pack_8x1_pixels(const uint16_t *src, const ptrdiff_t pitch,
+                                   __m256i *x /*x[4]*/) {
+  pack_pixels_with_format(src, PACK_8x1, pitch, x);
+}
+
+static INLINE void pack_8x2_pixels(const uint16_t *src, const ptrdiff_t pitch,
+                                   __m256i *x /*x[8]*/) {
+  pack_pixels_with_format(src, PACK_8x2, pitch, x);
+}
+
+static INLINE void pack_16x1_pixels(const uint16_t *src, const ptrdiff_t pitch,
+                                    __m256i *x /*x[8]*/) {
+  pack_pixels_with_format(src, PACK_16x1, pitch, x);
+}
+
+// Note:
+//  Shared by horizontal and vertical filtering
+static INLINE void pack_filters(const int16_t *filter, __m256i *f /*f[4]*/) {
+  const __m128i h = _mm_loadu_si128((const __m128i *)filter);
+  const __m256i hh = _mm256_insertf128_si256(_mm256_castsi128_si256(h), h, 1);
+  const __m256i p0 = _mm256_set1_epi32(0x03020100);
+  const __m256i p1 = _mm256_set1_epi32(0x07060504);
+  const __m256i p2 = _mm256_set1_epi32(0x0b0a0908);
+  const __m256i p3 = _mm256_set1_epi32(0x0f0e0d0c);
+  f[0] = _mm256_shuffle_epi8(hh, p0);
+  f[1] = _mm256_shuffle_epi8(hh, p1);
+  f[2] = _mm256_shuffle_epi8(hh, p2);
+  f[3] = _mm256_shuffle_epi8(hh, p3);
+}
+
+static INLINE void filter_8x1_pixels(const __m256i *sig /*sig[4]*/,
+                                     const __m256i *fil /*fil[4]*/,
+                                     __m256i *y) {
+  __m256i a, a0, a1;
+
+  a0 = _mm256_madd_epi16(fil[0], sig[0]);
+  a1 = _mm256_madd_epi16(fil[3], sig[3]);
+  a = _mm256_add_epi32(a0, a1);
+
+  a0 = _mm256_madd_epi16(fil[1], sig[1]);
+  a1 = _mm256_madd_epi16(fil[2], sig[2]);
+
+  const __m256i min = _mm256_min_epi32(a0, a1);
+  a = _mm256_add_epi32(a, min);
+
+  const __m256i max = _mm256_max_epi32(a0, a1);
+  a = _mm256_add_epi32(a, max);
+
+  const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1));
+  a = _mm256_add_epi32(a, rounding);
+  *y = _mm256_srai_epi32(a, CONV8_ROUNDING_BITS);
+}
+
+static void write_8x1_pixels(const __m256i *y, const __m256i *z,
+                             const __m256i *mask, uint16_t *dst,
+                             ptrdiff_t pitch) {
+  const __m128i a0 = _mm256_castsi256_si128(*y);
+  const __m128i a1 = _mm256_extractf128_si256(*y, 1);
+  __m128i res = _mm_packus_epi32(a0, a1);
+  (void)z;
+  (void)pitch;
+  res = _mm_min_epi16(res, _mm256_castsi256_si128(*mask));
+  _mm_storeu_si128((__m128i *)dst, res);
+}
+
+static void write_8x2_pixels(const __m256i *y0, const __m256i *y1,
+                             const __m256i *mask, uint16_t *dst,
+                             ptrdiff_t pitch) {
+  __m256i a = _mm256_packus_epi32(*y0, *y1);
+  a = _mm256_min_epi16(a, *mask);
+  _mm_storeu_si128((__m128i *)dst, _mm256_castsi256_si128(a));
+  _mm_storeu_si128((__m128i *)(dst + pitch), _mm256_extractf128_si256(a, 1));
+}
+
+static void write_16x1_pixels(const __m256i *y0, const __m256i *y1,
+                              const __m256i *mask, uint16_t *dst,
+                              ptrdiff_t dst_pitch) {
+  (void)dst_pitch;
+  __m256i a = _mm256_packus_epi32(*y0, *y1);
+  a = _mm256_min_epi16(a, *mask);
+  _mm256_storeu_si256((__m256i *)dst, a);
+}
+
+static void filter_block_width8_horiz(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, const WritePixels write_8x1,
+    const WritePixels write_8x2, uint16_t *dst_ptr, ptrdiff_t dst_pitch,
+    uint32_t height, const int16_t *filter, int bd) {
+  __m256i signal[8], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff[4];
+  pack_filters(filter, ff);
+
+  src_ptr -= 3;
+  do {
+    pack_8x2_pixels(src_ptr, src_pitch, signal);
+    filter_8x1_pixels(signal, ff, &res0);
+    filter_8x1_pixels(&signal[4], ff, &res1);
+    write_8x2(&res0, &res1, &max, dst_ptr, dst_pitch);
+    height -= 2;
+    src_ptr += src_pitch << 1;
+    dst_ptr += dst_pitch << 1;
+  } while (height > 1);
+
+  if (height > 0) {
+    pack_8x1_pixels(src_ptr, src_pitch, signal);
+    filter_8x1_pixels(signal, ff, &res0);
+    write_8x1(&res0, &res1, &max, dst_ptr, dst_pitch);
+  }
+}
+
+static void aom_highbd_filter_block1d8_h8_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width8_horiz(src, src_pitch, write_8x1_pixels, write_8x2_pixels,
+                            dst, dst_pitch, height, filter, bd);
+}
+
+static void filter_block_width16_horiz(const uint16_t *src_ptr,
+                                       ptrdiff_t src_pitch,
+                                       const WritePixels write_16x1,
+                                       uint16_t *dst_ptr, ptrdiff_t dst_pitch,
+                                       uint32_t height, const int16_t *filter,
+                                       int bd) {
+  __m256i signal[8], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff[4];
+  pack_filters(filter, ff);
+
+  src_ptr -= 3;
+  do {
+    pack_16x1_pixels(src_ptr, src_pitch, signal);
+    filter_8x1_pixels(signal, ff, &res0);
+    filter_8x1_pixels(&signal[4], ff, &res1);
+    write_16x1(&res0, &res1, &max, dst_ptr, dst_pitch);
+    height -= 1;
+    src_ptr += src_pitch;
+    dst_ptr += dst_pitch;
+  } while (height > 0);
+}
+
+static void aom_highbd_filter_block1d16_h8_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width16_horiz(src, src_pitch, write_16x1_pixels, dst, dst_pitch,
+                             height, filter, bd);
+}
+
+// 2-tap horizontal filtering
+
+static INLINE void pack_2t_filter(const int16_t *filter, __m256i *f) {
+  const __m128i h = _mm_loadu_si128((const __m128i *)filter);
+  const __m256i hh = _mm256_insertf128_si256(_mm256_castsi128_si256(h), h, 1);
+  const __m256i p = _mm256_set1_epi32(0x09080706);
+  f[0] = _mm256_shuffle_epi8(hh, p);
+}
+
+// can be used by pack_8x2_2t_pixels() and pack_16x1_2t_pixels()
+// the difference is s0/s1 specifies first and second rows or,
+// first 16 samples and 8-sample shifted 16 samples
+static INLINE void pack_16_2t_pixels(const __m256i *s0, const __m256i *s1,
+                                     __m256i *sig) {
+  const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index);
+  const __m256i sf2 = _mm256_loadu_si256((const __m256i *)signal_pattern_2);
+  __m256i x0 = _mm256_shuffle_epi8(*s0, sf2);
+  __m256i x1 = _mm256_shuffle_epi8(*s1, sf2);
+  __m256i r0 = _mm256_permutevar8x32_epi32(*s0, idx);
+  __m256i r1 = _mm256_permutevar8x32_epi32(*s1, idx);
+  r0 = _mm256_shuffle_epi8(r0, sf2);
+  r1 = _mm256_shuffle_epi8(r1, sf2);
+  sig[0] = _mm256_permute2x128_si256(x0, x1, 0x20);
+  sig[1] = _mm256_permute2x128_si256(r0, r1, 0x20);
+}
+
+static INLINE void pack_8x2_2t_pixels(const uint16_t *src,
+                                      const ptrdiff_t pitch, __m256i *sig) {
+  const __m256i r0 = _mm256_loadu_si256((const __m256i *)src);
+  const __m256i r1 = _mm256_loadu_si256((const __m256i *)(src + pitch));
+  pack_16_2t_pixels(&r0, &r1, sig);
+}
+
+static INLINE void pack_16x1_2t_pixels(const uint16_t *src,
+                                       __m256i *sig /*sig[2]*/) {
+  const __m256i r0 = _mm256_loadu_si256((const __m256i *)src);
+  const __m256i r1 = _mm256_loadu_si256((const __m256i *)(src + 8));
+  pack_16_2t_pixels(&r0, &r1, sig);
+}
+
+static INLINE void pack_8x1_2t_pixels(const uint16_t *src,
+                                      __m256i *sig /*sig[2]*/) {
+  const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index);
+  const __m256i sf2 = _mm256_loadu_si256((const __m256i *)signal_pattern_2);
+  __m256i r0 = _mm256_loadu_si256((const __m256i *)src);
+  __m256i x0 = _mm256_shuffle_epi8(r0, sf2);
+  r0 = _mm256_permutevar8x32_epi32(r0, idx);
+  r0 = _mm256_shuffle_epi8(r0, sf2);
+  sig[0] = _mm256_permute2x128_si256(x0, r0, 0x20);
+}
+
+// can be used by filter_8x2_2t_pixels() and filter_16x1_2t_pixels()
+static INLINE void filter_16_2t_pixels(const __m256i *sig, const __m256i *f,
+                                       __m256i *y0, __m256i *y1) {
+  const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1));
+  __m256i x0 = _mm256_madd_epi16(sig[0], *f);
+  __m256i x1 = _mm256_madd_epi16(sig[1], *f);
+  x0 = _mm256_add_epi32(x0, rounding);
+  x1 = _mm256_add_epi32(x1, rounding);
+  *y0 = _mm256_srai_epi32(x0, CONV8_ROUNDING_BITS);
+  *y1 = _mm256_srai_epi32(x1, CONV8_ROUNDING_BITS);
+}
+
+static INLINE void filter_8x2_2t_pixels(const __m256i *sig, const __m256i *f,
+                                        __m256i *y0, __m256i *y1) {
+  filter_16_2t_pixels(sig, f, y0, y1);
+}
+
+static INLINE void filter_16x1_2t_pixels(const __m256i *sig, const __m256i *f,
+                                         __m256i *y0, __m256i *y1) {
+  filter_16_2t_pixels(sig, f, y0, y1);
+}
+
+static INLINE void filter_8x1_2t_pixels(const __m256i *sig, const __m256i *f,
+                                        __m256i *y0) {
+  const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1));
+  __m256i x0 = _mm256_madd_epi16(sig[0], *f);
+  x0 = _mm256_add_epi32(x0, rounding);
+  *y0 = _mm256_srai_epi32(x0, CONV8_ROUNDING_BITS);
+}
+
+static void filter_block_width8_2t_horiz(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, const WritePixels write_8x1,
+    const WritePixels write_8x2, uint16_t *dst_ptr, ptrdiff_t dst_pitch,
+    uint32_t height, const int16_t *filter, int bd) {
+  __m256i signal[2], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff;
+  pack_2t_filter(filter, &ff);
+
+  src_ptr -= 3;
+  do {
+    pack_8x2_2t_pixels(src_ptr, src_pitch, signal);
+    filter_8x2_2t_pixels(signal, &ff, &res0, &res1);
+    write_8x2(&res0, &res1, &max, dst_ptr, dst_pitch);
+    height -= 2;
+    src_ptr += src_pitch << 1;
+    dst_ptr += dst_pitch << 1;
+  } while (height > 1);
+
+  if (height > 0) {
+    pack_8x1_2t_pixels(src_ptr, signal);
+    filter_8x1_2t_pixels(signal, &ff, &res0);
+    write_8x1(&res0, &res1, &max, dst_ptr, dst_pitch);
+  }
+}
+
+static void aom_highbd_filter_block1d8_h2_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width8_2t_horiz(src, src_pitch, write_8x1_pixels,
+                               write_8x2_pixels, dst, dst_pitch, height, filter,
+                               bd);
+}
+
+static void filter_block_width16_2t_horiz(const uint16_t *src_ptr,
+                                          ptrdiff_t src_pitch,
+                                          const WritePixels write_16x1,
+                                          uint16_t *dst_ptr,
+                                          ptrdiff_t dst_pitch, uint32_t height,
+                                          const int16_t *filter, int bd) {
+  __m256i signal[2], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff;
+  pack_2t_filter(filter, &ff);
+
+  src_ptr -= 3;
+  do {
+    pack_16x1_2t_pixels(src_ptr, signal);
+    filter_16x1_2t_pixels(signal, &ff, &res0, &res1);
+    write_16x1(&res0, &res1, &max, dst_ptr, dst_pitch);
+    height -= 1;
+    src_ptr += src_pitch;
+    dst_ptr += dst_pitch;
+  } while (height > 0);
+}
+
+static void aom_highbd_filter_block1d16_h2_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width16_2t_horiz(src, src_pitch, write_16x1_pixels, dst,
+                                dst_pitch, height, filter, bd);
+}
+
+// Vertical Filtering
+
+static void pack_8x9_init(const uint16_t *src, ptrdiff_t pitch, __m256i *sig) {
+  __m256i s0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)src));
+  __m256i s1 =
+      _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src + pitch)));
+  __m256i s2 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 2 * pitch)));
+  __m256i s3 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 3 * pitch)));
+  __m256i s4 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 4 * pitch)));
+  __m256i s5 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 5 * pitch)));
+  __m256i s6 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 6 * pitch)));
+
+  s0 = _mm256_inserti128_si256(s0, _mm256_castsi256_si128(s1), 1);
+  s1 = _mm256_inserti128_si256(s1, _mm256_castsi256_si128(s2), 1);
+  s2 = _mm256_inserti128_si256(s2, _mm256_castsi256_si128(s3), 1);
+  s3 = _mm256_inserti128_si256(s3, _mm256_castsi256_si128(s4), 1);
+  s4 = _mm256_inserti128_si256(s4, _mm256_castsi256_si128(s5), 1);
+  s5 = _mm256_inserti128_si256(s5, _mm256_castsi256_si128(s6), 1);
+
+  sig[0] = _mm256_unpacklo_epi16(s0, s1);
+  sig[4] = _mm256_unpackhi_epi16(s0, s1);
+  sig[1] = _mm256_unpacklo_epi16(s2, s3);
+  sig[5] = _mm256_unpackhi_epi16(s2, s3);
+  sig[2] = _mm256_unpacklo_epi16(s4, s5);
+  sig[6] = _mm256_unpackhi_epi16(s4, s5);
+  sig[8] = s6;
+}
+
+static INLINE void pack_8x9_pixels(const uint16_t *src, ptrdiff_t pitch,
+                                   __m256i *sig) {
+  // base + 7th row
+  __m256i s0 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 7 * pitch)));
+  // base + 8th row
+  __m256i s1 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 8 * pitch)));
+  __m256i s2 = _mm256_inserti128_si256(sig[8], _mm256_castsi256_si128(s0), 1);
+  __m256i s3 = _mm256_inserti128_si256(s0, _mm256_castsi256_si128(s1), 1);
+  sig[3] = _mm256_unpacklo_epi16(s2, s3);
+  sig[7] = _mm256_unpackhi_epi16(s2, s3);
+  sig[8] = s1;
+}
+
+static INLINE void filter_8x9_pixels(const __m256i *sig, const __m256i *f,
+                                     __m256i *y0, __m256i *y1) {
+  filter_8x1_pixels(sig, f, y0);
+  filter_8x1_pixels(&sig[4], f, y1);
+}
+
+static INLINE void update_pixels(__m256i *sig) {
+  int i;
+  for (i = 0; i < 3; ++i) {
+    sig[i] = sig[i + 1];
+    sig[i + 4] = sig[i + 5];
+  }
+}
+
+static INLINE void write_8x1_pixels_ver(const __m256i *y0, const __m256i *y1,
+                                        const __m256i *mask, uint16_t *dst,
+                                        ptrdiff_t pitch) {
+  (void)pitch;
+  const __m128i v0 = _mm256_castsi256_si128(*y0);
+  const __m128i v1 = _mm256_castsi256_si128(*y1);
+  __m128i p = _mm_packus_epi32(v0, v1);
+  p = _mm_min_epi16(p, _mm256_castsi256_si128(*mask));
+  _mm_storeu_si128((__m128i *)dst, p);
+}
+
+static void filter_block_width8_vert(const uint16_t *src_ptr,
+                                     ptrdiff_t src_pitch, WritePixels write_8x1,
+                                     WritePixels write_8x2, uint16_t *dst_ptr,
+                                     ptrdiff_t dst_pitch, uint32_t height,
+                                     const int16_t *filter, int bd) {
+  __m256i signal[9], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff[4];
+  pack_filters(filter, ff);
+
+  pack_8x9_init(src_ptr, src_pitch, signal);
+
+  do {
+    pack_8x9_pixels(src_ptr, src_pitch, signal);
+
+    filter_8x9_pixels(signal, ff, &res0, &res1);
+    write_8x2(&res0, &res1, &max, dst_ptr, dst_pitch);
+    update_pixels(signal);
+
+    src_ptr += src_pitch << 1;
+    dst_ptr += dst_pitch << 1;
+    height -= 2;
+  } while (height > 1);
+
+  if (height > 0) {
+    pack_8x9_pixels(src_ptr, src_pitch, signal);
+    filter_8x9_pixels(signal, ff, &res0, &res1);
+    write_8x1(&res0, &res1, &max, dst_ptr, dst_pitch);
+  }
+}
+
+static void aom_highbd_filter_block1d8_v8_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width8_vert(src, src_pitch, write_8x1_pixels_ver,
+                           write_8x2_pixels, dst, dst_pitch, height, filter,
+                           bd);
+}
+
+static void pack_16x9_init(const uint16_t *src, ptrdiff_t pitch, __m256i *sig) {
+  __m256i u0, u1, u2, u3;
+  // load 0-6 rows
+  const __m256i s0 = _mm256_loadu_si256((const __m256i *)src);
+  const __m256i s1 = _mm256_loadu_si256((const __m256i *)(src + pitch));
+  const __m256i s2 = _mm256_loadu_si256((const __m256i *)(src + 2 * pitch));
+  const __m256i s3 = _mm256_loadu_si256((const __m256i *)(src + 3 * pitch));
+  const __m256i s4 = _mm256_loadu_si256((const __m256i *)(src + 4 * pitch));
+  const __m256i s5 = _mm256_loadu_si256((const __m256i *)(src + 5 * pitch));
+  const __m256i s6 = _mm256_loadu_si256((const __m256i *)(src + 6 * pitch));
+
+  u0 = _mm256_permute2x128_si256(s0, s1, 0x20);  // 0, 1 low
+  u1 = _mm256_permute2x128_si256(s0, s1, 0x31);  // 0, 1 high
+
+  u2 = _mm256_permute2x128_si256(s1, s2, 0x20);  // 1, 2 low
+  u3 = _mm256_permute2x128_si256(s1, s2, 0x31);  // 1, 2 high
+
+  sig[0] = _mm256_unpacklo_epi16(u0, u2);
+  sig[4] = _mm256_unpackhi_epi16(u0, u2);
+
+  sig[8] = _mm256_unpacklo_epi16(u1, u3);
+  sig[12] = _mm256_unpackhi_epi16(u1, u3);
+
+  u0 = _mm256_permute2x128_si256(s2, s3, 0x20);
+  u1 = _mm256_permute2x128_si256(s2, s3, 0x31);
+
+  u2 = _mm256_permute2x128_si256(s3, s4, 0x20);
+  u3 = _mm256_permute2x128_si256(s3, s4, 0x31);
+
+  sig[1] = _mm256_unpacklo_epi16(u0, u2);
+  sig[5] = _mm256_unpackhi_epi16(u0, u2);
+
+  sig[9] = _mm256_unpacklo_epi16(u1, u3);
+  sig[13] = _mm256_unpackhi_epi16(u1, u3);
+
+  u0 = _mm256_permute2x128_si256(s4, s5, 0x20);
+  u1 = _mm256_permute2x128_si256(s4, s5, 0x31);
+
+  u2 = _mm256_permute2x128_si256(s5, s6, 0x20);
+  u3 = _mm256_permute2x128_si256(s5, s6, 0x31);
+
+  sig[2] = _mm256_unpacklo_epi16(u0, u2);
+  sig[6] = _mm256_unpackhi_epi16(u0, u2);
+
+  sig[10] = _mm256_unpacklo_epi16(u1, u3);
+  sig[14] = _mm256_unpackhi_epi16(u1, u3);
+
+  sig[16] = s6;
+}
+
+static void pack_16x9_pixels(const uint16_t *src, ptrdiff_t pitch,
+                             __m256i *sig) {
+  // base + 7th row
+  const __m256i s7 = _mm256_loadu_si256((const __m256i *)(src + 7 * pitch));
+  // base + 8th row
+  const __m256i s8 = _mm256_loadu_si256((const __m256i *)(src + 8 * pitch));
+
+  __m256i u0, u1, u2, u3;
+  u0 = _mm256_permute2x128_si256(sig[16], s7, 0x20);
+  u1 = _mm256_permute2x128_si256(sig[16], s7, 0x31);
+
+  u2 = _mm256_permute2x128_si256(s7, s8, 0x20);
+  u3 = _mm256_permute2x128_si256(s7, s8, 0x31);
+
+  sig[3] = _mm256_unpacklo_epi16(u0, u2);
+  sig[7] = _mm256_unpackhi_epi16(u0, u2);
+
+  sig[11] = _mm256_unpacklo_epi16(u1, u3);
+  sig[15] = _mm256_unpackhi_epi16(u1, u3);
+
+  sig[16] = s8;
+}
+
+static INLINE void filter_16x9_pixels(const __m256i *sig, const __m256i *f,
+                                      __m256i *y0, __m256i *y1) {
+  __m256i res[4];
+  int i;
+  for (i = 0; i < 4; ++i) {
+    filter_8x1_pixels(&sig[i << 2], f, &res[i]);
+  }
+
+  const __m256i l0l1 = _mm256_packus_epi32(res[0], res[1]);
+  const __m256i h0h1 = _mm256_packus_epi32(res[2], res[3]);
+  *y0 = _mm256_permute2x128_si256(l0l1, h0h1, 0x20);
+  *y1 = _mm256_permute2x128_si256(l0l1, h0h1, 0x31);
+}
+
+static INLINE void write_16x2_pixels(const __m256i *y0, const __m256i *y1,
+                                     const __m256i *mask, uint16_t *dst,
+                                     ptrdiff_t pitch) {
+  __m256i p = _mm256_min_epi16(*y0, *mask);
+  _mm256_storeu_si256((__m256i *)dst, p);
+  p = _mm256_min_epi16(*y1, *mask);
+  _mm256_storeu_si256((__m256i *)(dst + pitch), p);
+}
+
+static INLINE void write_16x1_pixels_ver(const __m256i *y0, const __m256i *y1,
+                                         const __m256i *mask, uint16_t *dst,
+                                         ptrdiff_t pitch) {
+  (void)y1;
+  (void)pitch;
+  const __m256i p = _mm256_min_epi16(*y0, *mask);
+  _mm256_storeu_si256((__m256i *)dst, p);
+}
+
+static void update_16x9_pixels(__m256i *sig) {
+  update_pixels(&sig[0]);
+  update_pixels(&sig[8]);
+}
+
+static void filter_block_width16_vert(const uint16_t *src_ptr,
+                                      ptrdiff_t src_pitch,
+                                      WritePixels write_16x1,
+                                      WritePixels write_16x2, uint16_t *dst_ptr,
+                                      ptrdiff_t dst_pitch, uint32_t height,
+                                      const int16_t *filter, int bd) {
+  __m256i signal[17], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff[4];
+  pack_filters(filter, ff);
+
+  pack_16x9_init(src_ptr, src_pitch, signal);
+
+  do {
+    pack_16x9_pixels(src_ptr, src_pitch, signal);
+    filter_16x9_pixels(signal, ff, &res0, &res1);
+    write_16x2(&res0, &res1, &max, dst_ptr, dst_pitch);
+    update_16x9_pixels(signal);
+
+    src_ptr += src_pitch << 1;
+    dst_ptr += dst_pitch << 1;
+    height -= 2;
+  } while (height > 1);
+
+  if (height > 0) {
+    pack_16x9_pixels(src_ptr, src_pitch, signal);
+    filter_16x9_pixels(signal, ff, &res0, &res1);
+    write_16x1(&res0, &res1, &max, dst_ptr, dst_pitch);
+  }
+}
+
+static void aom_highbd_filter_block1d16_v8_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width16_vert(src, src_pitch, write_16x1_pixels_ver,
+                            write_16x2_pixels, dst, dst_pitch, height, filter,
+                            bd);
+}
+
+// 2-tap vertical filtering
+
+static void pack_16x2_init(const uint16_t *src, __m256i *sig) {
+  sig[2] = _mm256_loadu_si256((const __m256i *)src);
+}
+
+static INLINE void pack_16x2_2t_pixels(const uint16_t *src, ptrdiff_t pitch,
+                                       __m256i *sig) {
+  // load the next row
+  const __m256i u = _mm256_loadu_si256((const __m256i *)(src + pitch));
+  sig[0] = _mm256_unpacklo_epi16(sig[2], u);
+  sig[1] = _mm256_unpackhi_epi16(sig[2], u);
+  sig[2] = u;
+}
+
+static INLINE void filter_16x2_2t_pixels(const __m256i *sig, const __m256i *f,
+                                         __m256i *y0, __m256i *y1) {
+  filter_16_2t_pixels(sig, f, y0, y1);
+}
+
+static void filter_block_width16_2t_vert(const uint16_t *src_ptr,
+                                         ptrdiff_t src_pitch,
+                                         WritePixels write_16x1,
+                                         uint16_t *dst_ptr, ptrdiff_t dst_pitch,
+                                         uint32_t height, const int16_t *filter,
+                                         int bd) {
+  __m256i signal[3], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+  __m256i ff;
+
+  pack_2t_filter(filter, &ff);
+  pack_16x2_init(src_ptr, signal);
+
+  do {
+    pack_16x2_2t_pixels(src_ptr, src_pitch, signal);
+    filter_16x2_2t_pixels(signal, &ff, &res0, &res1);
+    write_16x1(&res0, &res1, &max, dst_ptr, dst_pitch);
+
+    src_ptr += src_pitch;
+    dst_ptr += dst_pitch;
+    height -= 1;
+  } while (height > 0);
+}
+
+static void aom_highbd_filter_block1d16_v2_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width16_2t_vert(src, src_pitch, write_16x1_pixels, dst,
+                               dst_pitch, height, filter, bd);
+}
+
+static INLINE void pack_8x1_2t_filter(const int16_t *filter, __m128i *f) {
+  const __m128i h = _mm_loadu_si128((const __m128i *)filter);
+  const __m128i p = _mm_set1_epi32(0x09080706);
+  f[0] = _mm_shuffle_epi8(h, p);
+}
+
+static void pack_8x2_init(const uint16_t *src, __m128i *sig) {
+  sig[2] = _mm_loadu_si128((const __m128i *)src);
+}
+
+static INLINE void pack_8x2_2t_pixels_ver(const uint16_t *src, ptrdiff_t pitch,
+                                          __m128i *sig) {
+  // load the next row
+  const __m128i u = _mm_loadu_si128((const __m128i *)(src + pitch));
+  sig[0] = _mm_unpacklo_epi16(sig[2], u);
+  sig[1] = _mm_unpackhi_epi16(sig[2], u);
+  sig[2] = u;
+}
+
+static INLINE void filter_8_2t_pixels(const __m128i *sig, const __m128i *f,
+                                      __m128i *y0, __m128i *y1) {
+  const __m128i rounding = _mm_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1));
+  __m128i x0 = _mm_madd_epi16(sig[0], *f);
+  __m128i x1 = _mm_madd_epi16(sig[1], *f);
+  x0 = _mm_add_epi32(x0, rounding);
+  x1 = _mm_add_epi32(x1, rounding);
+  *y0 = _mm_srai_epi32(x0, CONV8_ROUNDING_BITS);
+  *y1 = _mm_srai_epi32(x1, CONV8_ROUNDING_BITS);
+}
+
+static void write_8x1_2t_pixels_ver(const __m128i *y0, const __m128i *y1,
+                                    const __m128i *mask, uint16_t *dst) {
+  __m128i res = _mm_packus_epi32(*y0, *y1);
+  res = _mm_min_epi16(res, *mask);
+  _mm_storeu_si128((__m128i *)dst, res);
+}
+
+typedef void (*Write8Pixels)(const __m128i *y0, const __m128i *y1,
+                             const __m128i *mask, uint16_t *dst);
+
+static void filter_block_width8_2t_vert(const uint16_t *src_ptr,
+                                        ptrdiff_t src_pitch,
+                                        Write8Pixels write_8x1,
+                                        uint16_t *dst_ptr, ptrdiff_t dst_pitch,
+                                        uint32_t height, const int16_t *filter,
+                                        int bd) {
+  __m128i signal[3], res0, res1;
+  const __m128i max = _mm_set1_epi16((1 << bd) - 1);
+  __m128i ff;
+
+  pack_8x1_2t_filter(filter, &ff);
+  pack_8x2_init(src_ptr, signal);
+
+  do {
+    pack_8x2_2t_pixels_ver(src_ptr, src_pitch, signal);
+    filter_8_2t_pixels(signal, &ff, &res0, &res1);
+    write_8x1(&res0, &res1, &max, dst_ptr);
+
+    src_ptr += src_pitch;
+    dst_ptr += dst_pitch;
+    height -= 1;
+  } while (height > 0);
+}
+
+static void aom_highbd_filter_block1d8_v2_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width8_2t_vert(src, src_pitch, write_8x1_2t_pixels_ver, dst,
+                              dst_pitch, height, filter, bd);
+}
+
+// Calculation with averaging the input pixels
+
+static void write_8x1_avg_pixels(const __m256i *y0, const __m256i *y1,
+                                 const __m256i *mask, uint16_t *dst,
+                                 ptrdiff_t pitch) {
+  (void)y1;
+  (void)pitch;
+  const __m128i a0 = _mm256_castsi256_si128(*y0);
+  const __m128i a1 = _mm256_extractf128_si256(*y0, 1);
+  __m128i res = _mm_packus_epi32(a0, a1);
+  const __m128i pix = _mm_loadu_si128((const __m128i *)dst);
+  res = _mm_min_epi16(res, _mm256_castsi256_si128(*mask));
+  res = _mm_avg_epu16(res, pix);
+  _mm_storeu_si128((__m128i *)dst, res);
+}
+
+static void write_8x2_avg_pixels(const __m256i *y0, const __m256i *y1,
+                                 const __m256i *mask, uint16_t *dst,
+                                 ptrdiff_t pitch) {
+  __m256i a = _mm256_packus_epi32(*y0, *y1);
+  const __m128i pix0 = _mm_loadu_si128((const __m128i *)dst);
+  const __m128i pix1 = _mm_loadu_si128((const __m128i *)(dst + pitch));
+  const __m256i pix =
+      _mm256_insertf128_si256(_mm256_castsi128_si256(pix0), pix1, 1);
+  a = _mm256_min_epi16(a, *mask);
+  a = _mm256_avg_epu16(a, pix);
+  _mm_storeu_si128((__m128i *)dst, _mm256_castsi256_si128(a));
+  _mm_storeu_si128((__m128i *)(dst + pitch), _mm256_extractf128_si256(a, 1));
+}
+
+static void write_16x1_avg_pixels(const __m256i *y0, const __m256i *y1,
+                                  const __m256i *mask, uint16_t *dst,
+                                  ptrdiff_t pitch) {
+  (void)pitch;
+  __m256i a = _mm256_packus_epi32(*y0, *y1);
+  const __m256i pix = _mm256_loadu_si256((const __m256i *)dst);
+  a = _mm256_min_epi16(a, *mask);
+  a = _mm256_avg_epu16(a, pix);
+  _mm256_storeu_si256((__m256i *)dst, a);
+}
+
+static INLINE void write_8x1_avg_pixels_ver(const __m256i *y0,
+                                            const __m256i *y1,
+                                            const __m256i *mask, uint16_t *dst,
+                                            ptrdiff_t pitch) {
+  (void)pitch;
+  const __m128i v0 = _mm256_castsi256_si128(*y0);
+  const __m128i v1 = _mm256_castsi256_si128(*y1);
+  __m128i p = _mm_packus_epi32(v0, v1);
+  const __m128i pix = _mm_loadu_si128((const __m128i *)dst);
+  p = _mm_min_epi16(p, _mm256_castsi256_si128(*mask));
+  p = _mm_avg_epu16(p, pix);
+  _mm_storeu_si128((__m128i *)dst, p);
+}
+
+static INLINE void write_16x2_avg_pixels(const __m256i *y0, const __m256i *y1,
+                                         const __m256i *mask, uint16_t *dst,
+                                         ptrdiff_t pitch) {
+  const __m256i pix0 = _mm256_loadu_si256((const __m256i *)dst);
+  const __m256i pix1 = _mm256_loadu_si256((const __m256i *)(dst + pitch));
+  __m256i p = _mm256_min_epi16(*y0, *mask);
+  p = _mm256_avg_epu16(p, pix0);
+  _mm256_storeu_si256((__m256i *)dst, p);
+
+  p = _mm256_min_epi16(*y1, *mask);
+  p = _mm256_avg_epu16(p, pix1);
+  _mm256_storeu_si256((__m256i *)(dst + pitch), p);
+}
+
+static INLINE void write_16x1_avg_pixels_ver(const __m256i *y0,
+                                             const __m256i *y1,
+                                             const __m256i *mask, uint16_t *dst,
+                                             ptrdiff_t pitch) {
+  (void)y1;
+  (void)pitch;
+  __m256i p = _mm256_min_epi16(*y0, *mask);
+  const __m256i pix = _mm256_loadu_si256((const __m256i *)dst);
+  p = _mm256_avg_epu16(p, pix);
+  _mm256_storeu_si256((__m256i *)dst, p);
+}
+
+static void write_8x1_2t_avg_pixels_ver(const __m128i *y0, const __m128i *y1,
+                                        const __m128i *mask, uint16_t *dst) {
+  __m128i res = _mm_packus_epi32(*y0, *y1);
+  const __m128i pix = _mm_loadu_si128((const __m128i *)dst);
+  res = _mm_min_epi16(res, *mask);
+  res = _mm_avg_epu16(res, pix);
+  _mm_storeu_si128((__m128i *)dst, res);
+}
+
+static void aom_highbd_filter_block1d8_h8_avg_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width8_horiz(src, src_pitch, write_8x1_avg_pixels,
+                            write_8x2_avg_pixels, dst, dst_pitch, height,
+                            filter, bd);
+}
+
+static void aom_highbd_filter_block1d16_h8_avg_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width16_horiz(src, src_pitch, write_16x1_avg_pixels, dst,
+                             dst_pitch, height, filter, bd);
+}
+
+static void aom_highbd_filter_block1d8_v8_avg_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width8_vert(src, src_pitch, write_8x1_avg_pixels_ver,
+                           write_8x2_avg_pixels, dst, dst_pitch, height, filter,
+                           bd);
+}
+
+static void aom_highbd_filter_block1d16_v8_avg_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width16_vert(src, src_pitch, write_16x1_avg_pixels_ver,
+                            write_16x2_avg_pixels, dst, dst_pitch, height,
+                            filter, bd);
+}
+
+// 2-tap averaging
+
+static void aom_highbd_filter_block1d8_h2_avg_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width8_2t_horiz(src, src_pitch, write_8x1_avg_pixels,
+                               write_8x2_avg_pixels, dst, dst_pitch, height,
+                               filter, bd);
+}
+
+static void aom_highbd_filter_block1d16_h2_avg_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width16_2t_horiz(src, src_pitch, write_16x1_avg_pixels, dst,
+                                dst_pitch, height, filter, bd);
+}
+
+static void aom_highbd_filter_block1d16_v2_avg_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width16_2t_vert(src, src_pitch, write_16x1_avg_pixels, dst,
+                               dst_pitch, height, filter, bd);
+}
+
+static void aom_highbd_filter_block1d8_v2_avg_avx2(
+    const uint16_t *src, ptrdiff_t src_pitch, uint16_t *dst,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  filter_block_width8_2t_vert(src, src_pitch, write_8x1_2t_avg_pixels_ver, dst,
+                              dst_pitch, height, filter, bd);
+}
+
+typedef void HbdFilter1dFunc(const uint16_t *, ptrdiff_t, uint16_t *, ptrdiff_t,
+                             uint32_t, const int16_t *, int);
+
+#define HIGHBD_FUNC(width, dir, avg, opt) \
+  aom_highbd_filter_block1d##width##_##dir##_##avg##opt
+
+HbdFilter1dFunc HIGHBD_FUNC(4, h8, , sse2);
+HbdFilter1dFunc HIGHBD_FUNC(4, h2, , sse2);
+HbdFilter1dFunc HIGHBD_FUNC(4, v8, , sse2);
+HbdFilter1dFunc HIGHBD_FUNC(4, v2, , sse2);
+
+#define aom_highbd_filter_block1d4_h8_avx2 HIGHBD_FUNC(4, h8, , sse2)
+#define aom_highbd_filter_block1d4_h2_avx2 HIGHBD_FUNC(4, h2, , sse2)
+#define aom_highbd_filter_block1d4_v8_avx2 HIGHBD_FUNC(4, v8, , sse2)
+#define aom_highbd_filter_block1d4_v2_avx2 HIGHBD_FUNC(4, v2, , sse2)
+
+HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2);
+HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2);
+HIGH_FUN_CONV_2D(, avx2);
+
+HbdFilter1dFunc HIGHBD_FUNC(4, h8, avg_, sse2);
+HbdFilter1dFunc HIGHBD_FUNC(4, h2, avg_, sse2);
+HbdFilter1dFunc HIGHBD_FUNC(4, v8, avg_, sse2);
+HbdFilter1dFunc HIGHBD_FUNC(4, v2, avg_, sse2);
+
+#define aom_highbd_filter_block1d4_h8_avg_avx2 HIGHBD_FUNC(4, h8, avg_, sse2)
+#define aom_highbd_filter_block1d4_h2_avg_avx2 HIGHBD_FUNC(4, h2, avg_, sse2)
+#define aom_highbd_filter_block1d4_v8_avg_avx2 HIGHBD_FUNC(4, v8, avg_, sse2)
+#define aom_highbd_filter_block1d4_v2_avg_avx2 HIGHBD_FUNC(4, v2, avg_, sse2)
+
+HIGH_FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, avx2);
+HIGH_FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_,
+                 avx2);
+HIGH_FUN_CONV_2D(avg_, avx2);
+
+#undef HIGHBD_FUNC
diff --git a/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.asm
new file mode 100644
index 0000000000..5d84ef8a75
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.asm
@@ -0,0 +1,456 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_4:  times 8 dw 4
+pw_8:  times 8 dw 8
+pw_16: times 4 dd 16
+pw_32: times 4 dd 32
+
+SECTION .text
+INIT_XMM sse2
+cglobal highbd_dc_predictor_4x4, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  movq                  m0, [aboveq]
+  movq                  m2, [leftq]
+  paddw                 m0, m2
+  pshuflw               m1, m0, 0xe
+  paddw                 m0, m1
+  pshuflw               m1, m0, 0x1
+  paddw                 m0, m1
+  paddw                 m0, [GLOBAL(pw_4)]
+  psraw                 m0, 3
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_8x8, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3, one
+  mov                 oned, 0x00010001
+  lea             stride3q, [strideq*3]
+  movd                  m3, oned
+  pshufd                m3, m3, 0x0
+  paddw                 m0, m2
+  pmaddwd               m0, m3
+  packssdw              m0, m1
+  pmaddwd               m0, m3
+  packssdw              m0, m1
+  pmaddwd               m0, m3
+  paddw                 m0, [GLOBAL(pw_8)]
+  psrlw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  lea                 dstq, [dstq+strideq*8]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_16x16, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m3, [aboveq+16]
+  mova                  m2, [leftq]
+  mova                  m4, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  paddw                 m0, m2
+  paddw                 m0, m3
+  paddw                 m0, m4
+  movhlps               m2, m0
+  paddw                 m0, m2
+  punpcklwd             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  punpckldq             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  paddd                 m0, [GLOBAL(pw_16)]
+  psrad                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+.loop:
+  mova   [dstq              ], m0
+  mova   [dstq           +16], m0
+  mova   [dstq+strideq*2    ], m0
+  mova   [dstq+strideq*2 +16], m0
+  mova   [dstq+strideq*4    ], m0
+  mova   [dstq+strideq*4 +16], m0
+  mova   [dstq+stride3q*2   ], m0
+  mova   [dstq+stride3q*2+16], m0
+  lea                 dstq, [dstq+strideq*8]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_32x32, 4, 5, 7, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [aboveq+32]
+  mova                  m4, [aboveq+48]
+  paddw                 m0, m2
+  paddw                 m3, m4
+  mova                  m2, [leftq]
+  mova                  m4, [leftq+16]
+  mova                  m5, [leftq+32]
+  mova                  m6, [leftq+48]
+  paddw                 m2, m4
+  paddw                 m5, m6
+  paddw                 m0, m3
+  paddw                 m2, m5
+  pxor                  m1, m1
+  paddw                 m0, m2
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  movhlps               m2, m0
+  paddw                 m0, m2
+  punpcklwd             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  punpckldq             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  paddd                 m0, [GLOBAL(pw_32)]
+  psrad                 m0, 6
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+.loop:
+  mova [dstq               ], m0
+  mova [dstq          +16  ], m0
+  mova [dstq          +32  ], m0
+  mova [dstq          +48  ], m0
+  mova [dstq+strideq*2     ], m0
+  mova [dstq+strideq*2+16  ], m0
+  mova [dstq+strideq*2+32  ], m0
+  mova [dstq+strideq*2+48  ], m0
+  mova [dstq+strideq*4     ], m0
+  mova [dstq+strideq*4+16  ], m0
+  mova [dstq+strideq*4+32  ], m0
+  mova [dstq+strideq*4+48  ], m0
+  mova [dstq+stride3q*2    ], m0
+  mova [dstq+stride3q*2 +16], m0
+  mova [dstq+stride3q*2 +32], m0
+  mova [dstq+stride3q*2 +48], m0
+  lea                 dstq, [dstq+strideq*8]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_4x4, 3, 3, 1, dst, stride, above
+  movq                  m0, [aboveq]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_8x8, 3, 3, 1, dst, stride, above
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  lea                 dstq, [dstq+strideq*8]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_16x16, 3, 4, 2, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 4
+.loop:
+  mova    [dstq              ], m0
+  mova    [dstq           +16], m1
+  mova    [dstq+strideq*2    ], m0
+  mova    [dstq+strideq*2 +16], m1
+  mova    [dstq+strideq*4    ], m0
+  mova    [dstq+strideq*4 +16], m1
+  mova    [dstq+stride3q*2   ], m0
+  mova    [dstq+stride3q*2+16], m1
+  lea                 dstq, [dstq+strideq*8]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_32x32, 3, 4, 4, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  mova                  m2, [aboveq+32]
+  mova                  m3, [aboveq+48]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 8
+.loop:
+  mova [dstq               ], m0
+  mova [dstq            +16], m1
+  mova [dstq            +32], m2
+  mova [dstq            +48], m3
+  mova [dstq+strideq*2     ], m0
+  mova [dstq+strideq*2  +16], m1
+  mova [dstq+strideq*2  +32], m2
+  mova [dstq+strideq*2  +48], m3
+  mova [dstq+strideq*4     ], m0
+  mova [dstq+strideq*4  +16], m1
+  mova [dstq+strideq*4  +32], m2
+  mova [dstq+strideq*4  +48], m3
+  mova [dstq+stride3q*2    ], m0
+  mova [dstq+stride3q*2 +16], m1
+  mova [dstq+stride3q*2 +32], m2
+  mova [dstq+stride3q*2 +48], m3
+  lea                 dstq, [dstq+strideq*8]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_tm_predictor_4x4, 5, 5, 6, dst, stride, above, left, bps
+  movd                  m1, [aboveq-2]
+  movq                  m0, [aboveq]
+  pshuflw               m1, m1, 0x0
+  movlhps               m0, m0         ; t1 t2 t3 t4 t1 t2 t3 t4
+  movlhps               m1, m1         ; tl tl tl tl tl tl tl tl
+  ; Get the values to compute the maximum value at this bit depth
+  pcmpeqw               m3, m3
+  movd                  m4, bpsd
+  psubw                 m0, m1         ; t1-tl t2-tl t3-tl t4-tl
+  psllw                 m3, m4
+  pcmpeqw               m2, m2
+  pxor                  m4, m4         ; min possible value
+  pxor                  m3, m2         ; max possible value
+  mova                  m1, [leftq]
+  pshuflw               m2, m1, 0x0
+  pshuflw               m5, m1, 0x55
+  movlhps               m2, m5         ; l1 l1 l1 l1 l2 l2 l2 l2
+  paddw                 m2, m0
+  ;Clamp to the bit-depth
+  pminsw                m2, m3
+  pmaxsw                m2, m4
+  ;Store the values
+  movq    [dstq          ], m2
+  movhpd  [dstq+strideq*2], m2
+  lea                 dstq, [dstq+strideq*4]
+  pshuflw               m2, m1, 0xaa
+  pshuflw               m5, m1, 0xff
+  movlhps               m2, m5
+  paddw                 m2, m0
+  ;Clamp to the bit-depth
+  pminsw                m2, m3
+  pmaxsw                m2, m4
+  ;Store the values
+  movq    [dstq          ], m2
+  movhpd  [dstq+strideq*2], m2
+  RET
+
+INIT_XMM sse2
+cglobal highbd_tm_predictor_8x8, 5, 6, 5, dst, stride, above, left, bps, one
+  movd                  m1, [aboveq-2]
+  mova                  m0, [aboveq]
+  pshuflw               m1, m1, 0x0
+  ; Get the values to compute the maximum value at this bit depth
+  mov                 oned, 1
+  pxor                  m3, m3
+  pxor                  m4, m4
+  pinsrw                m3, oned, 0
+  pinsrw                m4, bpsd, 0
+  pshuflw               m3, m3, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  punpcklqdq            m3, m3
+  mov                lineq, -4
+  mova                  m2, m3
+  punpcklqdq            m1, m1
+  psllw                 m3, m4
+  add                leftq, 16
+  psubw                 m3, m2 ; max possible value
+  pxor                  m4, m4 ; min possible value
+  psubw                 m0, m1
+.loop:
+  movd                  m1, [leftq+lineq*4]
+  movd                  m2, [leftq+lineq*4+2]
+  pshuflw               m1, m1, 0x0
+  pshuflw               m2, m2, 0x0
+  punpcklqdq            m1, m1
+  punpcklqdq            m2, m2
+  paddw                 m1, m0
+  paddw                 m2, m0
+  ;Clamp to the bit-depth
+  pminsw                m1, m3
+  pminsw                m2, m3
+  pmaxsw                m1, m4
+  pmaxsw                m2, m4
+  ;Store the values
+  mova      [dstq          ], m1
+  mova      [dstq+strideq*2], m2
+  lea                 dstq, [dstq+strideq*4]
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_tm_predictor_16x16, 5, 5, 8, dst, stride, above, left, bps
+  movd                  m2, [aboveq-2]
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  pshuflw               m2, m2, 0x0
+  ; Get the values to compute the maximum value at this bit depth
+  pcmpeqw               m3, m3
+  movd                  m4, bpsd
+  punpcklqdq            m2, m2
+  psllw                 m3, m4
+  pcmpeqw               m5, m5
+  pxor                  m4, m4         ; min possible value
+  pxor                  m3, m5         ; max possible value
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -8
+  psubw                 m0, m2
+  psubw                 m1, m2
+.loop:
+  movd                  m7, [leftq]
+  pshuflw               m5, m7, 0x0
+  pshuflw               m2, m7, 0x55
+  punpcklqdq            m5, m5         ; l1 l1 l1 l1 l1 l1 l1 l1
+  punpcklqdq            m2, m2         ; l2 l2 l2 l2 l2 l2 l2 l2
+  paddw                 m6, m5, m0     ; t1-tl+l1 to t4-tl+l1
+  paddw                 m5, m1         ; t5-tl+l1 to t8-tl+l1
+  pminsw                m6, m3
+  pminsw                m5, m3
+  pmaxsw                m6, m4         ; Clamp to the bit-depth
+  pmaxsw                m5, m4
+  mova   [dstq           ], m6
+  mova   [dstq        +16], m5
+  paddw                 m6, m2, m0
+  paddw                 m2, m1
+  pminsw                m6, m3
+  pminsw                m2, m3
+  pmaxsw                m6, m4
+  pmaxsw                m2, m4
+  mova   [dstq+strideq*2 ], m6
+  mova [dstq+strideq*2+16], m2
+  lea                 dstq, [dstq+strideq*4]
+  inc                lineq
+  lea                leftq, [leftq+4]
+
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_tm_predictor_32x32, 5, 5, 8, dst, stride, above, left, bps
+  movd                  m0, [aboveq-2]
+  mova                  m1, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [aboveq+32]
+  mova                  m4, [aboveq+48]
+  pshuflw               m0, m0, 0x0
+  ; Get the values to compute the maximum value at this bit depth
+  pcmpeqw               m5, m5
+  movd                  m6, bpsd
+  psllw                 m5, m6
+  pcmpeqw               m7, m7
+  pxor                  m6, m6         ; min possible value
+  pxor                  m5, m7         ; max possible value
+  punpcklqdq            m0, m0
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -16
+  psubw                 m1, m0
+  psubw                 m2, m0
+  psubw                 m3, m0
+  psubw                 m4, m0
+.loop:
+  movd                  m7, [leftq]
+  pshuflw               m7, m7, 0x0
+  punpcklqdq            m7, m7         ; l1 l1 l1 l1 l1 l1 l1 l1
+  paddw                 m0, m7, m1
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq           ], m0
+  paddw                 m0, m7, m2
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq        +16], m0
+  paddw                 m0, m7, m3
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq        +32], m0
+  paddw                 m0, m7, m4
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq        +48], m0
+  movd                  m7, [leftq+2]
+  pshuflw               m7, m7, 0x0
+  punpcklqdq            m7, m7         ; l2 l2 l2 l2 l2 l2 l2 l2
+  paddw                 m0, m7, m1
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq+strideq*2 ], m0
+  paddw                 m0, m7, m2
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq+strideq*2+16], m0
+  paddw                 m0, m7, m3
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq+strideq*2+32], m0
+  paddw                 m0, m7, m4
+  pminsw                m0, m5
+  pmaxsw                m0, m6
+  mova   [dstq+strideq*2+48], m0
+  lea                 dstq, [dstq+strideq*4]
+  lea                leftq, [leftq+4]
+  inc                lineq
+  jnz .loop
+  REP_RET
diff --git a/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c b/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c
new file mode 100644
index 0000000000..76369871ba
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c
@@ -0,0 +1,1140 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/emmintrin_compat.h"
+
+static INLINE __m128i signed_char_clamp_bd_sse2(__m128i value, int bd) {
+  __m128i ubounded;
+  __m128i lbounded;
+  __m128i retval;
+
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i t80, max, min;
+
+  if (bd == 8) {
+    t80 = _mm_set1_epi16(0x80);
+    max = _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, 8), one), t80);
+  } else if (bd == 10) {
+    t80 = _mm_set1_epi16(0x200);
+    max = _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, 10), one), t80);
+  } else {  // bd == 12
+    t80 = _mm_set1_epi16(0x800);
+    max = _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, 12), one), t80);
+  }
+
+  min = _mm_subs_epi16(zero, t80);
+
+  ubounded = _mm_cmpgt_epi16(value, max);
+  lbounded = _mm_cmplt_epi16(value, min);
+  retval = _mm_andnot_si128(_mm_or_si128(ubounded, lbounded), value);
+  ubounded = _mm_and_si128(ubounded, max);
+  lbounded = _mm_and_si128(lbounded, min);
+  retval = _mm_or_si128(retval, ubounded);
+  retval = _mm_or_si128(retval, lbounded);
+  return retval;
+}
+
+// TODO(debargha, peter): Break up large functions into smaller ones
+// in this file.
+void aom_highbd_lpf_horizontal_edge_8_sse2(uint16_t *s, int p,
+                                           const uint8_t *_blimit,
+                                           const uint8_t *_limit,
+                                           const uint8_t *_thresh, int bd) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i blimit, limit, thresh;
+  __m128i q7, p7, q6, p6, q5, p5, q4, p4, q3, p3, q2, p2, q1, p1, q0, p0;
+  __m128i mask, hev, flat, flat2, abs_p1p0, abs_q1q0;
+  __m128i ps1, qs1, ps0, qs0;
+  __m128i abs_p0q0, abs_p1q1, ffff, work;
+  __m128i filt, work_a, filter1, filter2;
+  __m128i flat2_q6, flat2_p6, flat2_q5, flat2_p5, flat2_q4, flat2_p4;
+  __m128i flat2_q3, flat2_p3, flat2_q2, flat2_p2, flat2_q1, flat2_p1;
+  __m128i flat2_q0, flat2_p0;
+  __m128i flat_q2, flat_p2, flat_q1, flat_p1, flat_q0, flat_p0;
+  __m128i pixelFilter_p, pixelFilter_q;
+  __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+  __m128i sum_p7, sum_q7, sum_p3, sum_q3;
+  __m128i t4, t3, t80, t1;
+  __m128i eight, four;
+
+  if (bd == 8) {
+    blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+    limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+    thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+  } else if (bd == 10) {
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+    limit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+    thresh = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+  } else {  // bd == 12
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+    limit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+    thresh = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+  }
+
+  q4 = _mm_load_si128((__m128i *)(s + 4 * p));
+  p4 = _mm_load_si128((__m128i *)(s - 5 * p));
+  q3 = _mm_load_si128((__m128i *)(s + 3 * p));
+  p3 = _mm_load_si128((__m128i *)(s - 4 * p));
+  q2 = _mm_load_si128((__m128i *)(s + 2 * p));
+  p2 = _mm_load_si128((__m128i *)(s - 3 * p));
+  q1 = _mm_load_si128((__m128i *)(s + 1 * p));
+  p1 = _mm_load_si128((__m128i *)(s - 2 * p));
+  q0 = _mm_load_si128((__m128i *)(s + 0 * p));
+  p0 = _mm_load_si128((__m128i *)(s - 1 * p));
+
+  //  highbd_filter_mask
+  abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1));
+  abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1));
+
+  ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
+
+  abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0));
+  abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1));
+
+  //  highbd_hev_mask (in C code this is actually called from highbd_filter4)
+  flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  hev = _mm_subs_epu16(flat, thresh);
+  hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+  abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0);  // abs(p0 - q0) * 2
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);         // abs(p1 - q1) / 2
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1)),
+      _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p2, p1), _mm_subs_epu16(p1, p2)),
+      _mm_or_si128(_mm_subs_epu16(q2, q1), _mm_subs_epu16(q1, q2)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p3, p2), _mm_subs_epu16(p2, p3)),
+      _mm_or_si128(_mm_subs_epu16(q3, q2), _mm_subs_epu16(q2, q3)));
+  mask = _mm_max_epi16(work, mask);
+
+  mask = _mm_subs_epu16(mask, limit);
+  mask = _mm_cmpeq_epi16(mask, zero);  // return ~mask
+
+  // lp filter
+  // highbd_filter4
+  t4 = _mm_set1_epi16(4);
+  t3 = _mm_set1_epi16(3);
+  if (bd == 8)
+    t80 = _mm_set1_epi16(0x80);
+  else if (bd == 10)
+    t80 = _mm_set1_epi16(0x200);
+  else  // bd == 12
+    t80 = _mm_set1_epi16(0x800);
+
+  t1 = _mm_set1_epi16(0x1);
+
+  ps1 = _mm_subs_epi16(p1, t80);
+  qs1 = _mm_subs_epi16(q1, t80);
+  ps0 = _mm_subs_epi16(p0, t80);
+  qs0 = _mm_subs_epi16(q0, t80);
+
+  filt = _mm_and_si128(signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd),
+                       hev);
+  work_a = _mm_subs_epi16(qs0, ps0);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
+  filt = _mm_and_si128(filt, mask);
+  filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
+  filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
+
+  // Filter1 >> 3
+  filter1 = _mm_srai_epi16(filter1, 0x3);
+  filter2 = _mm_srai_epi16(filter2, 0x3);
+
+  qs0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd), t80);
+  ps0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80);
+  filt = _mm_adds_epi16(filter1, t1);
+  filt = _mm_srai_epi16(filt, 1);
+  filt = _mm_andnot_si128(hev, filt);
+  qs1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd),
+                       t80);
+  ps1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd),
+                       t80);
+
+  // end highbd_filter4
+  // loopfilter done
+
+  // highbd_flat_mask4
+  flat = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p2, p0), _mm_subs_epu16(p0, p2)),
+      _mm_or_si128(_mm_subs_epu16(p3, p0), _mm_subs_epu16(p0, p3)));
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(q2, q0), _mm_subs_epu16(q0, q2)),
+      _mm_or_si128(_mm_subs_epu16(q3, q0), _mm_subs_epu16(q0, q3)));
+  flat = _mm_max_epi16(work, flat);
+  work = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  flat = _mm_max_epi16(work, flat);
+
+  if (bd == 8)
+    flat = _mm_subs_epu16(flat, one);
+  else if (bd == 10)
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2));
+  else  // bd == 12
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4));
+
+  flat = _mm_cmpeq_epi16(flat, zero);
+  // end flat_mask4
+
+  // flat & mask = flat && mask (as used in filter8)
+  // (because, in both vars, each block of 16 either all 1s or all 0s)
+  flat = _mm_and_si128(flat, mask);
+
+  p5 = _mm_load_si128((__m128i *)(s - 6 * p));
+  q5 = _mm_load_si128((__m128i *)(s + 5 * p));
+  p6 = _mm_load_si128((__m128i *)(s - 7 * p));
+  q6 = _mm_load_si128((__m128i *)(s + 6 * p));
+  p7 = _mm_load_si128((__m128i *)(s - 8 * p));
+  q7 = _mm_load_si128((__m128i *)(s + 7 * p));
+
+  // highbd_flat_mask5 (arguments passed in are p0, q0, p4-p7, q4-q7
+  // but referred to as p0-p4 & q0-q4 in fn)
+  flat2 = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p4, p0), _mm_subs_epu16(p0, p4)),
+      _mm_or_si128(_mm_subs_epu16(q4, q0), _mm_subs_epu16(q0, q4)));
+
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p5, p0), _mm_subs_epu16(p0, p5)),
+      _mm_or_si128(_mm_subs_epu16(q5, q0), _mm_subs_epu16(q0, q5)));
+  flat2 = _mm_max_epi16(work, flat2);
+
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p6, p0), _mm_subs_epu16(p0, p6)),
+      _mm_or_si128(_mm_subs_epu16(q6, q0), _mm_subs_epu16(q0, q6)));
+  flat2 = _mm_max_epi16(work, flat2);
+
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p7, p0), _mm_subs_epu16(p0, p7)),
+      _mm_or_si128(_mm_subs_epu16(q7, q0), _mm_subs_epu16(q0, q7)));
+  flat2 = _mm_max_epi16(work, flat2);
+
+  if (bd == 8)
+    flat2 = _mm_subs_epu16(flat2, one);
+  else if (bd == 10)
+    flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 2));
+  else  // bd == 12
+    flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 4));
+
+  flat2 = _mm_cmpeq_epi16(flat2, zero);
+  flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+  // end highbd_flat_mask5
+
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // flat and wide flat calculations
+  eight = _mm_set1_epi16(8);
+  four = _mm_set1_epi16(4);
+
+  pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6, p5), _mm_add_epi16(p4, p3));
+  pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6, q5), _mm_add_epi16(q4, q3));
+
+  pixetFilter_p2p1p0 = _mm_add_epi16(p0, _mm_add_epi16(p2, p1));
+  pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+  pixetFilter_q2q1q0 = _mm_add_epi16(q0, _mm_add_epi16(q2, q1));
+  pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+  pixelFilter_p =
+      _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q));
+  pixetFilter_p2p1p0 = _mm_add_epi16(
+      four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+  flat2_p0 =
+      _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7, p0)), 4);
+  flat2_q0 =
+      _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7, q0)), 4);
+  flat_p0 = _mm_srli_epi16(
+      _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3, p0)), 3);
+  flat_q0 = _mm_srli_epi16(
+      _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3, q0)), 3);
+
+  sum_p7 = _mm_add_epi16(p7, p7);
+  sum_q7 = _mm_add_epi16(q7, q7);
+  sum_p3 = _mm_add_epi16(p3, p3);
+  sum_q3 = _mm_add_epi16(q3, q3);
+
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6);
+  flat2_p1 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1)), 4);
+  flat2_q1 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1)), 4);
+
+  pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2);
+  pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2);
+  flat_p1 = _mm_srli_epi16(
+      _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1)), 3);
+  flat_q1 = _mm_srli_epi16(
+      _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1)), 3);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  sum_p3 = _mm_add_epi16(sum_p3, p3);
+  sum_q3 = _mm_add_epi16(sum_q3, q3);
+
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5);
+  flat2_p2 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2)), 4);
+  flat2_q2 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2)), 4);
+
+  pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1);
+  pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1);
+  flat_p2 = _mm_srli_epi16(
+      _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2)), 3);
+  flat_q2 = _mm_srli_epi16(
+      _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2)), 3);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4);
+  flat2_p3 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3)), 4);
+  flat2_q3 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3)), 4);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3);
+  flat2_p4 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4)), 4);
+  flat2_q4 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4)), 4);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2);
+  flat2_p5 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5)), 4);
+  flat2_q5 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5)), 4);
+
+  sum_p7 = _mm_add_epi16(sum_p7, p7);
+  sum_q7 = _mm_add_epi16(sum_q7, q7);
+  pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1);
+  pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1);
+  flat2_p6 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6)), 4);
+  flat2_q6 = _mm_srli_epi16(
+      _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6)), 4);
+
+  //  wide flat
+  //  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+  //  highbd_filter8
+  p2 = _mm_andnot_si128(flat, p2);
+  //  p2 remains unchanged if !(flat && mask)
+  flat_p2 = _mm_and_si128(flat, flat_p2);
+  //  when (flat && mask)
+  p2 = _mm_or_si128(p2, flat_p2);  // full list of p2 values
+  q2 = _mm_andnot_si128(flat, q2);
+  flat_q2 = _mm_and_si128(flat, flat_q2);
+  q2 = _mm_or_si128(q2, flat_q2);  // full list of q2 values
+
+  ps1 = _mm_andnot_si128(flat, ps1);
+  //  p1 takes the value assigned to in in filter4 if !(flat && mask)
+  flat_p1 = _mm_and_si128(flat, flat_p1);
+  //  when (flat && mask)
+  p1 = _mm_or_si128(ps1, flat_p1);  // full list of p1 values
+  qs1 = _mm_andnot_si128(flat, qs1);
+  flat_q1 = _mm_and_si128(flat, flat_q1);
+  q1 = _mm_or_si128(qs1, flat_q1);  // full list of q1 values
+
+  ps0 = _mm_andnot_si128(flat, ps0);
+  //  p0 takes the value assigned to in in filter4 if !(flat && mask)
+  flat_p0 = _mm_and_si128(flat, flat_p0);
+  //  when (flat && mask)
+  p0 = _mm_or_si128(ps0, flat_p0);  // full list of p0 values
+  qs0 = _mm_andnot_si128(flat, qs0);
+  flat_q0 = _mm_and_si128(flat, flat_q0);
+  q0 = _mm_or_si128(qs0, flat_q0);  // full list of q0 values
+  // end highbd_filter8
+
+  // highbd_filter16
+  p6 = _mm_andnot_si128(flat2, p6);
+  //  p6 remains unchanged if !(flat2 && flat && mask)
+  flat2_p6 = _mm_and_si128(flat2, flat2_p6);
+  //  get values for when (flat2 && flat && mask)
+  p6 = _mm_or_si128(p6, flat2_p6);  // full list of p6 values
+  q6 = _mm_andnot_si128(flat2, q6);
+  //  q6 remains unchanged if !(flat2 && flat && mask)
+  flat2_q6 = _mm_and_si128(flat2, flat2_q6);
+  //  get values for when (flat2 && flat && mask)
+  q6 = _mm_or_si128(q6, flat2_q6);  // full list of q6 values
+  _mm_store_si128((__m128i *)(s - 7 * p), p6);
+  _mm_store_si128((__m128i *)(s + 6 * p), q6);
+
+  p5 = _mm_andnot_si128(flat2, p5);
+  //  p5 remains unchanged if !(flat2 && flat && mask)
+  flat2_p5 = _mm_and_si128(flat2, flat2_p5);
+  //  get values for when (flat2 && flat && mask)
+  p5 = _mm_or_si128(p5, flat2_p5);
+  //  full list of p5 values
+  q5 = _mm_andnot_si128(flat2, q5);
+  //  q5 remains unchanged if !(flat2 && flat && mask)
+  flat2_q5 = _mm_and_si128(flat2, flat2_q5);
+  //  get values for when (flat2 && flat && mask)
+  q5 = _mm_or_si128(q5, flat2_q5);
+  //  full list of q5 values
+  _mm_store_si128((__m128i *)(s - 6 * p), p5);
+  _mm_store_si128((__m128i *)(s + 5 * p), q5);
+
+  p4 = _mm_andnot_si128(flat2, p4);
+  //  p4 remains unchanged if !(flat2 && flat && mask)
+  flat2_p4 = _mm_and_si128(flat2, flat2_p4);
+  //  get values for when (flat2 && flat && mask)
+  p4 = _mm_or_si128(p4, flat2_p4);  // full list of p4 values
+  q4 = _mm_andnot_si128(flat2, q4);
+  //  q4 remains unchanged if !(flat2 && flat && mask)
+  flat2_q4 = _mm_and_si128(flat2, flat2_q4);
+  //  get values for when (flat2 && flat && mask)
+  q4 = _mm_or_si128(q4, flat2_q4);  // full list of q4 values
+  _mm_store_si128((__m128i *)(s - 5 * p), p4);
+  _mm_store_si128((__m128i *)(s + 4 * p), q4);
+
+  p3 = _mm_andnot_si128(flat2, p3);
+  //  p3 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p3 = _mm_and_si128(flat2, flat2_p3);
+  //  get values for when (flat2 && flat && mask)
+  p3 = _mm_or_si128(p3, flat2_p3);  // full list of p3 values
+  q3 = _mm_andnot_si128(flat2, q3);
+  //  q3 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q3 = _mm_and_si128(flat2, flat2_q3);
+  //  get values for when (flat2 && flat && mask)
+  q3 = _mm_or_si128(q3, flat2_q3);  // full list of q3 values
+  _mm_store_si128((__m128i *)(s - 4 * p), p3);
+  _mm_store_si128((__m128i *)(s + 3 * p), q3);
+
+  p2 = _mm_andnot_si128(flat2, p2);
+  //  p2 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p2 = _mm_and_si128(flat2, flat2_p2);
+  //  get values for when (flat2 && flat && mask)
+  p2 = _mm_or_si128(p2, flat2_p2);
+  //  full list of p2 values
+  q2 = _mm_andnot_si128(flat2, q2);
+  //  q2 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q2 = _mm_and_si128(flat2, flat2_q2);
+  //  get values for when (flat2 && flat && mask)
+  q2 = _mm_or_si128(q2, flat2_q2);  // full list of q2 values
+  _mm_store_si128((__m128i *)(s - 3 * p), p2);
+  _mm_store_si128((__m128i *)(s + 2 * p), q2);
+
+  p1 = _mm_andnot_si128(flat2, p1);
+  //  p1 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p1 = _mm_and_si128(flat2, flat2_p1);
+  //  get values for when (flat2 && flat && mask)
+  p1 = _mm_or_si128(p1, flat2_p1);  // full list of p1 values
+  q1 = _mm_andnot_si128(flat2, q1);
+  //  q1 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q1 = _mm_and_si128(flat2, flat2_q1);
+  //  get values for when (flat2 && flat && mask)
+  q1 = _mm_or_si128(q1, flat2_q1);  // full list of q1 values
+  _mm_store_si128((__m128i *)(s - 2 * p), p1);
+  _mm_store_si128((__m128i *)(s + 1 * p), q1);
+
+  p0 = _mm_andnot_si128(flat2, p0);
+  //  p0 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_p0 = _mm_and_si128(flat2, flat2_p0);
+  //  get values for when (flat2 && flat && mask)
+  p0 = _mm_or_si128(p0, flat2_p0);  // full list of p0 values
+  q0 = _mm_andnot_si128(flat2, q0);
+  //  q0 takes value from highbd_filter8 if !(flat2 && flat && mask)
+  flat2_q0 = _mm_and_si128(flat2, flat2_q0);
+  //  get values for when (flat2 && flat && mask)
+  q0 = _mm_or_si128(q0, flat2_q0);  // full list of q0 values
+  _mm_store_si128((__m128i *)(s - 1 * p), p0);
+  _mm_store_si128((__m128i *)(s - 0 * p), q0);
+}
+
+void aom_highbd_lpf_horizontal_edge_16_sse2(uint16_t *s, int p,
+                                            const uint8_t *_blimit,
+                                            const uint8_t *_limit,
+                                            const uint8_t *_thresh, int bd) {
+  aom_highbd_lpf_horizontal_edge_8_sse2(s, p, _blimit, _limit, _thresh, bd);
+  aom_highbd_lpf_horizontal_edge_8_sse2(s + 8, p, _blimit, _limit, _thresh, bd);
+}
+
+void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p,
+                                      const uint8_t *_blimit,
+                                      const uint8_t *_limit,
+                                      const uint8_t *_thresh, int bd) {
+  DECLARE_ALIGNED(16, uint16_t, flat_op2[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_op1[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_op0[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_oq2[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_oq1[16]);
+  DECLARE_ALIGNED(16, uint16_t, flat_oq0[16]);
+  const __m128i zero = _mm_set1_epi16(0);
+  __m128i blimit, limit, thresh;
+  __m128i mask, hev, flat;
+  __m128i p3 = _mm_load_si128((__m128i *)(s - 4 * p));
+  __m128i q3 = _mm_load_si128((__m128i *)(s + 3 * p));
+  __m128i p2 = _mm_load_si128((__m128i *)(s - 3 * p));
+  __m128i q2 = _mm_load_si128((__m128i *)(s + 2 * p));
+  __m128i p1 = _mm_load_si128((__m128i *)(s - 2 * p));
+  __m128i q1 = _mm_load_si128((__m128i *)(s + 1 * p));
+  __m128i p0 = _mm_load_si128((__m128i *)(s - 1 * p));
+  __m128i q0 = _mm_load_si128((__m128i *)(s + 0 * p));
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i ffff = _mm_cmpeq_epi16(one, one);
+  __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+  const __m128i four = _mm_set1_epi16(4);
+  __m128i workp_a, workp_b, workp_shft;
+
+  const __m128i t4 = _mm_set1_epi16(4);
+  const __m128i t3 = _mm_set1_epi16(3);
+  __m128i t80;
+  const __m128i t1 = _mm_set1_epi16(0x1);
+  __m128i ps1, ps0, qs0, qs1;
+  __m128i filt;
+  __m128i work_a;
+  __m128i filter1, filter2;
+
+  if (bd == 8) {
+    blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+    limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+    thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+    t80 = _mm_set1_epi16(0x80);
+  } else if (bd == 10) {
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+    limit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+    thresh = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+    t80 = _mm_set1_epi16(0x200);
+  } else {  // bd == 12
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+    limit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+    thresh = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+    t80 = _mm_set1_epi16(0x800);
+  }
+
+  ps1 = _mm_subs_epi16(p1, t80);
+  ps0 = _mm_subs_epi16(p0, t80);
+  qs0 = _mm_subs_epi16(q0, t80);
+  qs1 = _mm_subs_epi16(q1, t80);
+
+  // filter_mask and hev_mask
+  abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1));
+  abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1));
+
+  abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0));
+  abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1));
+  flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  hev = _mm_subs_epu16(flat, thresh);
+  hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+  abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+  // So taking maximums continues to work:
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+  mask = _mm_max_epi16(abs_p1p0, mask);
+  // mask |= (abs(p1 - p0) > limit) * -1;
+  mask = _mm_max_epi16(abs_q1q0, mask);
+  // mask |= (abs(q1 - q0) > limit) * -1;
+
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p2, p1), _mm_subs_epu16(p1, p2)),
+      _mm_or_si128(_mm_subs_epu16(q2, q1), _mm_subs_epu16(q1, q2)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p3, p2), _mm_subs_epu16(p2, p3)),
+      _mm_or_si128(_mm_subs_epu16(q3, q2), _mm_subs_epu16(q2, q3)));
+  mask = _mm_max_epi16(work, mask);
+  mask = _mm_subs_epu16(mask, limit);
+  mask = _mm_cmpeq_epi16(mask, zero);
+
+  // flat_mask4
+  flat = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p2, p0), _mm_subs_epu16(p0, p2)),
+      _mm_or_si128(_mm_subs_epu16(q2, q0), _mm_subs_epu16(q0, q2)));
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p3, p0), _mm_subs_epu16(p0, p3)),
+      _mm_or_si128(_mm_subs_epu16(q3, q0), _mm_subs_epu16(q0, q3)));
+  flat = _mm_max_epi16(work, flat);
+  flat = _mm_max_epi16(abs_p1p0, flat);
+  flat = _mm_max_epi16(abs_q1q0, flat);
+
+  if (bd == 8)
+    flat = _mm_subs_epu16(flat, one);
+  else if (bd == 10)
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2));
+  else  // bd == 12
+    flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4));
+
+  flat = _mm_cmpeq_epi16(flat, zero);
+  flat = _mm_and_si128(flat, mask);  // flat & mask
+
+  // Added before shift for rounding part of ROUND_POWER_OF_TWO
+
+  workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+  workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+  workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_op2[0], workp_shft);
+
+  workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_op1[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_op0[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_oq0[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_oq1[0], workp_shft);
+
+  workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+  workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+  workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+  _mm_store_si128((__m128i *)&flat_oq2[0], workp_shft);
+
+  // lp filter
+  filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd);
+  filt = _mm_and_si128(filt, hev);
+  work_a = _mm_subs_epi16(qs0, ps0);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  // (aom_filter + 3 * (qs0 - ps0)) & mask
+  filt = signed_char_clamp_bd_sse2(filt, bd);
+  filt = _mm_and_si128(filt, mask);
+
+  filter1 = _mm_adds_epi16(filt, t4);
+  filter2 = _mm_adds_epi16(filt, t3);
+
+  // Filter1 >> 3
+  filter1 = signed_char_clamp_bd_sse2(filter1, bd);
+  filter1 = _mm_srai_epi16(filter1, 3);
+
+  // Filter2 >> 3
+  filter2 = signed_char_clamp_bd_sse2(filter2, bd);
+  filter2 = _mm_srai_epi16(filter2, 3);
+
+  // filt >> 1
+  filt = _mm_adds_epi16(filter1, t1);
+  filt = _mm_srai_epi16(filt, 1);
+  // filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+  filt = _mm_andnot_si128(hev, filt);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  q0 = _mm_load_si128((__m128i *)flat_oq0);
+  work_a = _mm_andnot_si128(flat, work_a);
+  q0 = _mm_and_si128(flat, q0);
+  q0 = _mm_or_si128(work_a, q0);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  q1 = _mm_load_si128((__m128i *)flat_oq1);
+  work_a = _mm_andnot_si128(flat, work_a);
+  q1 = _mm_and_si128(flat, q1);
+  q1 = _mm_or_si128(work_a, q1);
+
+  work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q2 = _mm_load_si128((__m128i *)flat_oq2);
+  work_a = _mm_andnot_si128(flat, work_a);
+  q2 = _mm_and_si128(flat, q2);
+  q2 = _mm_or_si128(work_a, q2);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  p0 = _mm_load_si128((__m128i *)flat_op0);
+  work_a = _mm_andnot_si128(flat, work_a);
+  p0 = _mm_and_si128(flat, p0);
+  p0 = _mm_or_si128(work_a, p0);
+
+  work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd);
+  work_a = _mm_adds_epi16(work_a, t80);
+  p1 = _mm_load_si128((__m128i *)flat_op1);
+  work_a = _mm_andnot_si128(flat, work_a);
+  p1 = _mm_and_si128(flat, p1);
+  p1 = _mm_or_si128(work_a, p1);
+
+  work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p2 = _mm_load_si128((__m128i *)flat_op2);
+  work_a = _mm_andnot_si128(flat, work_a);
+  p2 = _mm_and_si128(flat, p2);
+  p2 = _mm_or_si128(work_a, p2);
+
+  _mm_store_si128((__m128i *)(s - 3 * p), p2);
+  _mm_store_si128((__m128i *)(s - 2 * p), p1);
+  _mm_store_si128((__m128i *)(s - 1 * p), p0);
+  _mm_store_si128((__m128i *)(s + 0 * p), q0);
+  _mm_store_si128((__m128i *)(s + 1 * p), q1);
+  _mm_store_si128((__m128i *)(s + 2 * p), q2);
+}
+
+void aom_highbd_lpf_horizontal_8_dual_sse2(
+    uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0,
+    const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1,
+    const uint8_t *_thresh1, int bd) {
+  aom_highbd_lpf_horizontal_8_sse2(s, p, _blimit0, _limit0, _thresh0, bd);
+  aom_highbd_lpf_horizontal_8_sse2(s + 8, p, _blimit1, _limit1, _thresh1, bd);
+}
+
+void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p,
+                                      const uint8_t *_blimit,
+                                      const uint8_t *_limit,
+                                      const uint8_t *_thresh, int bd) {
+  const __m128i zero = _mm_set1_epi16(0);
+  __m128i blimit, limit, thresh;
+  __m128i mask, hev, flat;
+  __m128i p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  __m128i p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  __m128i p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  __m128i p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  __m128i q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  __m128i q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  __m128i q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  __m128i q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  const __m128i abs_p1p0 =
+      _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1));
+  const __m128i abs_q1q0 =
+      _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1));
+  const __m128i ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i abs_p0q0 =
+      _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0));
+  __m128i abs_p1q1 =
+      _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1));
+  __m128i work;
+  const __m128i t4 = _mm_set1_epi16(4);
+  const __m128i t3 = _mm_set1_epi16(3);
+  __m128i t80;
+  __m128i tff80;
+  __m128i tffe0;
+  __m128i t1f;
+  // equivalent to shifting 0x1f left by bitdepth - 8
+  // and setting new bits to 1
+  const __m128i t1 = _mm_set1_epi16(0x1);
+  __m128i t7f;
+  // equivalent to shifting 0x7f left by bitdepth - 8
+  // and setting new bits to 1
+  __m128i ps1, ps0, qs0, qs1;
+  __m128i filt;
+  __m128i work_a;
+  __m128i filter1, filter2;
+
+  if (bd == 8) {
+    blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+    limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+    thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+    t80 = _mm_set1_epi16(0x80);
+    tff80 = _mm_set1_epi16(0xff80);
+    tffe0 = _mm_set1_epi16(0xffe0);
+    t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 8);
+    t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 8);
+  } else if (bd == 10) {
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+    limit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+    thresh = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+    t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 2);
+    tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 2);
+    tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 2);
+    t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 6);
+    t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 6);
+  } else {  // bd == 12
+    blimit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+    limit = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+    thresh = _mm_slli_epi16(
+        _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+    t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 4);
+    tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 4);
+    tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 4);
+    t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 4);
+    t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 4);
+  }
+
+  ps1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80);
+  ps0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80);
+  qs0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80);
+  qs1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80);
+
+  // filter_mask and hev_mask
+  flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  hev = _mm_subs_epu16(flat, thresh);
+  hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+  abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+  // So taking maximums continues to work:
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+  mask = _mm_max_epi16(flat, mask);
+  // mask |= (abs(p1 - p0) > limit) * -1;
+  // mask |= (abs(q1 - q0) > limit) * -1;
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(p2, p1), _mm_subs_epu16(p1, p2)),
+      _mm_or_si128(_mm_subs_epu16(p3, p2), _mm_subs_epu16(p2, p3)));
+  mask = _mm_max_epi16(work, mask);
+  work = _mm_max_epi16(
+      _mm_or_si128(_mm_subs_epu16(q2, q1), _mm_subs_epu16(q1, q2)),
+      _mm_or_si128(_mm_subs_epu16(q3, q2), _mm_subs_epu16(q2, q3)));
+  mask = _mm_max_epi16(work, mask);
+  mask = _mm_subs_epu16(mask, limit);
+  mask = _mm_cmpeq_epi16(mask, zero);
+
+  // filter4
+  filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd);
+  filt = _mm_and_si128(filt, hev);
+  work_a = _mm_subs_epi16(qs0, ps0);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = _mm_adds_epi16(filt, work_a);
+  filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
+
+  // (aom_filter + 3 * (qs0 - ps0)) & mask
+  filt = _mm_and_si128(filt, mask);
+
+  filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
+  filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
+
+  // Filter1 >> 3
+  work_a = _mm_cmpgt_epi16(zero, filter1);  // get the values that are <0
+  filter1 = _mm_srli_epi16(filter1, 3);
+  work_a = _mm_and_si128(work_a, tffe0);    // sign bits for the values < 0
+  filter1 = _mm_and_si128(filter1, t1f);    // clamp the range
+  filter1 = _mm_or_si128(filter1, work_a);  // reinsert the sign bits
+
+  // Filter2 >> 3
+  work_a = _mm_cmpgt_epi16(zero, filter2);
+  filter2 = _mm_srli_epi16(filter2, 3);
+  work_a = _mm_and_si128(work_a, tffe0);
+  filter2 = _mm_and_si128(filter2, t1f);
+  filter2 = _mm_or_si128(filter2, work_a);
+
+  // filt >> 1
+  filt = _mm_adds_epi16(filter1, t1);
+  work_a = _mm_cmpgt_epi16(zero, filt);
+  filt = _mm_srli_epi16(filt, 1);
+  work_a = _mm_and_si128(work_a, tff80);
+  filt = _mm_and_si128(filt, t7f);
+  filt = _mm_or_si128(filt, work_a);
+
+  filt = _mm_andnot_si128(hev, filt);
+
+  q0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd), t80);
+  q1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd),
+                      t80);
+  p0 = _mm_adds_epi16(
+      signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80);
+  p1 = _mm_adds_epi16(signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd),
+                      t80);
+
+  _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+  _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+  _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+  _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+}
+
+void aom_highbd_lpf_horizontal_4_dual_sse2(
+    uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0,
+    const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1,
+    const uint8_t *_thresh1, int bd) {
+  aom_highbd_lpf_horizontal_4_sse2(s, p, _blimit0, _limit0, _thresh0, bd);
+  aom_highbd_lpf_horizontal_4_sse2(s + 8, p, _blimit1, _limit1, _thresh1, bd);
+}
+
+static INLINE void highbd_transpose(uint16_t *src[], int in_p, uint16_t *dst[],
+                                    int out_p, int num_8x8_to_transpose) {
+  int idx8x8 = 0;
+  __m128i p0, p1, p2, p3, p4, p5, p6, p7, x0, x1, x2, x3, x4, x5, x6, x7;
+  do {
+    uint16_t *in = src[idx8x8];
+    uint16_t *out = dst[idx8x8];
+
+    p0 =
+        _mm_loadu_si128((__m128i *)(in + 0 * in_p));  // 00 01 02 03 04 05 06 07
+    p1 =
+        _mm_loadu_si128((__m128i *)(in + 1 * in_p));  // 10 11 12 13 14 15 16 17
+    p2 =
+        _mm_loadu_si128((__m128i *)(in + 2 * in_p));  // 20 21 22 23 24 25 26 27
+    p3 =
+        _mm_loadu_si128((__m128i *)(in + 3 * in_p));  // 30 31 32 33 34 35 36 37
+    p4 =
+        _mm_loadu_si128((__m128i *)(in + 4 * in_p));  // 40 41 42 43 44 45 46 47
+    p5 =
+        _mm_loadu_si128((__m128i *)(in + 5 * in_p));  // 50 51 52 53 54 55 56 57
+    p6 =
+        _mm_loadu_si128((__m128i *)(in + 6 * in_p));  // 60 61 62 63 64 65 66 67
+    p7 =
+        _mm_loadu_si128((__m128i *)(in + 7 * in_p));  // 70 71 72 73 74 75 76 77
+    // 00 10 01 11 02 12 03 13
+    x0 = _mm_unpacklo_epi16(p0, p1);
+    // 20 30 21 31 22 32 23 33
+    x1 = _mm_unpacklo_epi16(p2, p3);
+    // 40 50 41 51 42 52 43 53
+    x2 = _mm_unpacklo_epi16(p4, p5);
+    // 60 70 61 71 62 72 63 73
+    x3 = _mm_unpacklo_epi16(p6, p7);
+    // 00 10 20 30 01 11 21 31
+    x4 = _mm_unpacklo_epi32(x0, x1);
+    // 40 50 60 70 41 51 61 71
+    x5 = _mm_unpacklo_epi32(x2, x3);
+    // 00 10 20 30 40 50 60 70
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 01 11 21 31 41 51 61 71
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 0 * out_p), x6);
+    // 00 10 20 30 40 50 60 70
+    _mm_storeu_si128((__m128i *)(out + 1 * out_p), x7);
+    // 01 11 21 31 41 51 61 71
+
+    // 02 12 22 32 03 13 23 33
+    x4 = _mm_unpackhi_epi32(x0, x1);
+    // 42 52 62 72 43 53 63 73
+    x5 = _mm_unpackhi_epi32(x2, x3);
+    // 02 12 22 32 42 52 62 72
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 03 13 23 33 43 53 63 73
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 2 * out_p), x6);
+    // 02 12 22 32 42 52 62 72
+    _mm_storeu_si128((__m128i *)(out + 3 * out_p), x7);
+    // 03 13 23 33 43 53 63 73
+
+    // 04 14 05 15 06 16 07 17
+    x0 = _mm_unpackhi_epi16(p0, p1);
+    // 24 34 25 35 26 36 27 37
+    x1 = _mm_unpackhi_epi16(p2, p3);
+    // 44 54 45 55 46 56 47 57
+    x2 = _mm_unpackhi_epi16(p4, p5);
+    // 64 74 65 75 66 76 67 77
+    x3 = _mm_unpackhi_epi16(p6, p7);
+    // 04 14 24 34 05 15 25 35
+    x4 = _mm_unpacklo_epi32(x0, x1);
+    // 44 54 64 74 45 55 65 75
+    x5 = _mm_unpacklo_epi32(x2, x3);
+    // 04 14 24 34 44 54 64 74
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 05 15 25 35 45 55 65 75
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 4 * out_p), x6);
+    // 04 14 24 34 44 54 64 74
+    _mm_storeu_si128((__m128i *)(out + 5 * out_p), x7);
+    // 05 15 25 35 45 55 65 75
+
+    // 06 16 26 36 07 17 27 37
+    x4 = _mm_unpackhi_epi32(x0, x1);
+    // 46 56 66 76 47 57 67 77
+    x5 = _mm_unpackhi_epi32(x2, x3);
+    // 06 16 26 36 46 56 66 76
+    x6 = _mm_unpacklo_epi64(x4, x5);
+    // 07 17 27 37 47 57 67 77
+    x7 = _mm_unpackhi_epi64(x4, x5);
+
+    _mm_storeu_si128((__m128i *)(out + 6 * out_p), x6);
+    // 06 16 26 36 46 56 66 76
+    _mm_storeu_si128((__m128i *)(out + 7 * out_p), x7);
+    // 07 17 27 37 47 57 67 77
+  } while (++idx8x8 < num_8x8_to_transpose);
+}
+
+static INLINE void highbd_transpose8x16(uint16_t *in0, uint16_t *in1, int in_p,
+                                        uint16_t *out, int out_p) {
+  uint16_t *src0[1];
+  uint16_t *src1[1];
+  uint16_t *dest0[1];
+  uint16_t *dest1[1];
+  src0[0] = in0;
+  src1[0] = in1;
+  dest0[0] = out;
+  dest1[0] = out + 8;
+  highbd_transpose(src0, in_p, dest0, out_p, 1);
+  highbd_transpose(src1, in_p, dest1, out_p, 1);
+}
+
+void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int p, const uint8_t *blimit,
+                                    const uint8_t *limit, const uint8_t *thresh,
+                                    int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]);
+  uint16_t *src[1];
+  uint16_t *dst[1];
+
+  // Transpose 8x8
+  src[0] = s - 4;
+  dst[0] = t_dst;
+
+  highbd_transpose(src, p, dst, 8, 1);
+
+  // Loop filtering
+  aom_highbd_lpf_horizontal_4_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, bd);
+
+  src[0] = t_dst;
+  dst[0] = s - 4;
+
+  // Transpose back
+  highbd_transpose(src, 8, dst, p, 1);
+}
+
+void aom_highbd_lpf_vertical_4_dual_sse2(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]);
+  uint16_t *src[2];
+  uint16_t *dst[2];
+
+  // Transpose 8x16
+  highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  aom_highbd_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0,
+                                        thresh0, blimit1, limit1, thresh1, bd);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  highbd_transpose(src, 16, dst, p, 2);
+}
+
+void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int p, const uint8_t *blimit,
+                                    const uint8_t *limit, const uint8_t *thresh,
+                                    int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]);
+  uint16_t *src[1];
+  uint16_t *dst[1];
+
+  // Transpose 8x8
+  src[0] = s - 4;
+  dst[0] = t_dst;
+
+  highbd_transpose(src, p, dst, 8, 1);
+
+  // Loop filtering
+  aom_highbd_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, bd);
+
+  src[0] = t_dst;
+  dst[0] = s - 4;
+
+  // Transpose back
+  highbd_transpose(src, 8, dst, p, 1);
+}
+
+void aom_highbd_lpf_vertical_8_dual_sse2(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]);
+  uint16_t *src[2];
+  uint16_t *dst[2];
+
+  // Transpose 8x16
+  highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  aom_highbd_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0,
+                                        thresh0, blimit1, limit1, thresh1, bd);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  highbd_transpose(src, 16, dst, p, 2);
+}
+
+void aom_highbd_lpf_vertical_16_sse2(uint16_t *s, int p, const uint8_t *blimit,
+                                     const uint8_t *limit,
+                                     const uint8_t *thresh, int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 16]);
+  uint16_t *src[2];
+  uint16_t *dst[2];
+
+  src[0] = s - 8;
+  src[1] = s;
+  dst[0] = t_dst;
+  dst[1] = t_dst + 8 * 8;
+
+  // Transpose 16x8
+  highbd_transpose(src, p, dst, 8, 2);
+
+  // Loop filtering
+  aom_highbd_lpf_horizontal_edge_8_sse2(t_dst + 8 * 8, 8, blimit, limit, thresh,
+                                        bd);
+  src[0] = t_dst;
+  src[1] = t_dst + 8 * 8;
+  dst[0] = s - 8;
+  dst[1] = s;
+
+  // Transpose back
+  highbd_transpose(src, 8, dst, p, 2);
+}
+
+void aom_highbd_lpf_vertical_16_dual_sse2(uint16_t *s, int p,
+                                          const uint8_t *blimit,
+                                          const uint8_t *limit,
+                                          const uint8_t *thresh, int bd) {
+  DECLARE_ALIGNED(16, uint16_t, t_dst[256]);
+
+  //  Transpose 16x16
+  highbd_transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16);
+  highbd_transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16);
+
+  //  Loop filtering
+  aom_highbd_lpf_horizontal_edge_16_sse2(t_dst + 8 * 16, 16, blimit, limit,
+                                         thresh, bd);
+
+  //  Transpose back
+  highbd_transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p);
+  highbd_transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p);
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c
new file mode 100644
index 0000000000..3ee24ab161
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c
@@ -0,0 +1,155 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#if CONFIG_HIGHBITDEPTH
+void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t count,
+                                int skip_block, const int16_t *zbin_ptr,
+                                const int16_t *round_ptr,
+                                const int16_t *quant_ptr,
+                                const int16_t *quant_shift_ptr,
+                                tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                                const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                                const int16_t *scan, const int16_t *iscan) {
+  int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
+  __m128i zbins[2];
+  __m128i nzbins[2];
+
+  zbins[0] = _mm_set_epi32((int)zbin_ptr[1], (int)zbin_ptr[1], (int)zbin_ptr[1],
+                           (int)zbin_ptr[0]);
+  zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  (void)scan;
+
+  memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = ((int)count / 4) - 1; i >= 0; i--) {
+      __m128i coeffs, cmp1, cmp2;
+      int test;
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+      cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+      cmp1 = _mm_and_si128(cmp1, cmp2);
+      test = _mm_movemask_epi8(cmp1);
+      if (test == 0xffff)
+        non_zero_regs--;
+      else
+        break;
+    }
+
+    // Quantization pass:
+    for (i = 0; i < non_zero_regs; i++) {
+      __m128i coeffs, coeffs_sign, tmp1, tmp2;
+      int test;
+      int abs_coeff[4];
+      int coeff_sign[4];
+
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      coeffs_sign = _mm_srai_epi32(coeffs, 31);
+      coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
+      tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
+      tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
+      tmp1 = _mm_or_si128(tmp1, tmp2);
+      test = _mm_movemask_epi8(tmp1);
+      _mm_storeu_si128((__m128i *)abs_coeff, coeffs);
+      _mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign);
+
+      for (j = 0; j < 4; j++) {
+        if (test & (1 << (4 * j))) {
+          int k = 4 * i + j;
+          const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0];
+          const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3;
+          const uint32_t abs_qcoeff =
+              (uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16);
+          qcoeff_ptr[k] = (int)(abs_qcoeff ^ coeff_sign[j]) - coeff_sign[j];
+          dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
+          if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
+        }
+      }
+    }
+  }
+  *eob_ptr = eob_i + 1;
+}
+
+void aom_highbd_quantize_b_32x32_sse2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  __m128i zbins[2];
+  __m128i nzbins[2];
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
+  const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
+  (void)scan;
+  zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp);
+  zbins[1] = _mm_set1_epi32(zbin1_tmp);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs / 4; i++) {
+      __m128i coeffs, cmp1, cmp2;
+      int test;
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+      cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+      cmp1 = _mm_and_si128(cmp1, cmp2);
+      test = _mm_movemask_epi8(cmp1);
+      if (!(test & 0xf)) idx_arr[idx++] = i * 4;
+      if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
+      if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
+      if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = idx_arr[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int64_t tmp1 =
+          abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+      const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+      const uint32_t abs_qcoeff =
+          (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
+      qcoeff_ptr[rc] = (int)(abs_qcoeff ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+      if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
diff --git a/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm
new file mode 100644
index 0000000000..0c7cb3998b
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm
@@ -0,0 +1,290 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; HIGH_PROCESS_4x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_4x2x4 5-6 0
+  movh                  m0, [srcq +%2*2]
+%if %1 == 1
+  movu                  m4, [ref1q+%3*2]
+  movu                  m5, [ref2q+%3*2]
+  movu                  m6, [ref3q+%3*2]
+  movu                  m7, [ref4q+%3*2]
+  movhps                m0, [srcq +%4*2]
+  movhps                m4, [ref1q+%5*2]
+  movhps                m5, [ref2q+%5*2]
+  movhps                m6, [ref3q+%5*2]
+  movhps                m7, [ref4q+%5*2]
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m4
+  psubusw               m2, m5
+  psubusw               m4, m0
+  psubusw               m5, m0
+  por                   m4, m3
+  por                   m5, m2
+  pmaddwd               m4, m1
+  pmaddwd               m5, m1
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m6
+  psubusw               m2, m7
+  psubusw               m6, m0
+  psubusw               m7, m0
+  por                   m6, m3
+  por                   m7, m2
+  pmaddwd               m6, m1
+  pmaddwd               m7, m1
+%else
+  movu                  m2, [ref1q+%3*2]
+  movhps                m0, [srcq +%4*2]
+  movhps                m2, [ref1q+%5*2]
+  mova                  m3, m0
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m4, m2
+
+  movu                  m2, [ref2q+%3*2]
+  mova                  m3, m0
+  movhps                m2, [ref2q+%5*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m5, m2
+
+  movu                  m2, [ref3q+%3*2]
+  mova                  m3, m0
+  movhps                m2, [ref3q+%5*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m6, m2
+
+  movu                  m2, [ref4q+%3*2]
+  mova                  m3, m0
+  movhps                m2, [ref4q+%5*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m7, m2
+%endif
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*4]
+  lea                ref1q, [ref1q+ref_strideq*4]
+  lea                ref2q, [ref2q+ref_strideq*4]
+  lea                ref3q, [ref3q+ref_strideq*4]
+  lea                ref4q, [ref4q+ref_strideq*4]
+%endif
+%endmacro
+
+; PROCESS_8x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_8x2x4 5-6 0
+  ; 1st 8 px
+  mova                  m0, [srcq +%2*2]
+%if %1 == 1
+  movu                  m4, [ref1q+%3*2]
+  movu                  m5, [ref2q+%3*2]
+  movu                  m6, [ref3q+%3*2]
+  movu                  m7, [ref4q+%3*2]
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m4
+  psubusw               m2, m5
+  psubusw               m4, m0
+  psubusw               m5, m0
+  por                   m4, m3
+  por                   m5, m2
+  pmaddwd               m4, m1
+  pmaddwd               m5, m1
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m6
+  psubusw               m2, m7
+  psubusw               m6, m0
+  psubusw               m7, m0
+  por                   m6, m3
+  por                   m7, m2
+  pmaddwd               m6, m1
+  pmaddwd               m7, m1
+%else
+  mova                  m3, m0
+  movu                  m2, [ref1q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m4, m2
+  movu                  m2, [ref2q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m5, m2
+  movu                  m2, [ref3q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m6, m2
+  movu                  m2, [ref4q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m7, m2
+%endif
+
+  ; 2nd 8 px
+  mova                  m0, [srcq +(%4)*2]
+  mova                  m3, m0
+  movu                  m2, [ref1q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m4, m2
+  movu                  m2, [ref2q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m5, m2
+  movu                  m2, [ref3q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m6, m2
+  movu                  m2, [ref4q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*4]
+  lea                ref1q, [ref1q+ref_strideq*4]
+  lea                ref2q, [ref2q+ref_strideq*4]
+  lea                ref3q, [ref3q+ref_strideq*4]
+  lea                ref4q, [ref4q+ref_strideq*4]
+%endif
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m7, m2
+%endmacro
+
+; HIGH_PROCESS_16x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_16x2x4 5-6 0
+  HIGH_PROCESS_8x2x4 %1, %2, %3, (%2 + 8), (%3 + 8)
+  HIGH_PROCESS_8x2x4  0, %4, %5, (%4 + 8), (%5 + 8), %6
+%endmacro
+
+; HIGH_PROCESS_32x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_32x2x4 5-6 0
+  HIGH_PROCESS_16x2x4 %1, %2, %3, (%2 + 16), (%3 + 16)
+  HIGH_PROCESS_16x2x4  0, %4, %5, (%4 + 16), (%5 + 16), %6
+%endmacro
+
+; HIGH_PROCESS_64x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_64x2x4 5-6 0
+  HIGH_PROCESS_32x2x4 %1, %2, %3, (%2 + 32), (%3 + 32)
+  HIGH_PROCESS_32x2x4  0, %4, %5, (%4 + 32), (%5 + 32), %6
+%endmacro
+
+; void aom_highbd_sadNxNx4d_sse2(uint8_t *src,    int src_stride,
+;                         uint8_t *ref[4], int ref_stride,
+;                         uint32_t res[4]);
+; where NxN = 64x64, 32x32, 16x16, 16x8, 8x16 or 8x8
+%macro HIGH_SADNXN4D 2
+%if UNIX64
+cglobal highbd_sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \
+                              res, ref2, ref3, ref4
+%else
+cglobal highbd_sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \
+                              ref2, ref3, ref4
+%endif
+
+; set m1
+  push                srcq
+  mov                 srcd, 0x00010001
+  movd                  m1, srcd
+  pshufd                m1, m1, 0x0
+  pop                 srcq
+
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+  mov                ref2q, [ref1q+gprsize*1]
+  mov                ref3q, [ref1q+gprsize*2]
+  mov                ref4q, [ref1q+gprsize*3]
+  mov                ref1q, [ref1q+gprsize*0]
+
+; convert byte pointers to short pointers
+  shl                 srcq, 1
+  shl                ref2q, 1
+  shl                ref3q, 1
+  shl                ref4q, 1
+  shl                ref1q, 1
+
+  HIGH_PROCESS_%1x2x4 1, 0, 0, src_strideq, ref_strideq, 1
+%rep (%2-4)/2
+  HIGH_PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 1
+%endrep
+  HIGH_PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 0
+  ; N.B. HIGH_PROCESS outputs dwords (32 bits)
+  ; so in high bit depth even the smallest width (4) needs 128bits i.e. XMM
+  movhlps               m0, m4
+  movhlps               m1, m5
+  movhlps               m2, m6
+  movhlps               m3, m7
+  paddd                 m4, m0
+  paddd                 m5, m1
+  paddd                 m6, m2
+  paddd                 m7, m3
+  punpckldq             m4, m5
+  punpckldq             m6, m7
+  movhlps               m0, m4
+  movhlps               m1, m6
+  paddd                 m4, m0
+  paddd                 m6, m1
+  punpcklqdq            m4, m6
+  movifnidn             r4, r4mp
+  movu                [r4], m4
+  RET
+%endmacro
+
+
+INIT_XMM sse2
+HIGH_SADNXN4D 64, 64
+HIGH_SADNXN4D 64, 32
+HIGH_SADNXN4D 32, 64
+HIGH_SADNXN4D 32, 32
+HIGH_SADNXN4D 32, 16
+HIGH_SADNXN4D 16, 32
+HIGH_SADNXN4D 16, 16
+HIGH_SADNXN4D 16,  8
+HIGH_SADNXN4D  8, 16
+HIGH_SADNXN4D  8,  8
+HIGH_SADNXN4D  8,  4
+HIGH_SADNXN4D  4,  8
+HIGH_SADNXN4D  4,  4
diff --git a/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm
new file mode 100644
index 0000000000..8427b891cc
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm
@@ -0,0 +1,366 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro HIGH_SAD_FN 4
+%if %4 == 0
+%if %3 == 5
+cglobal highbd_sad%1x%2, 4, %3, 7, src, src_stride, ref, ref_stride, n_rows
+%else ; %3 == 7
+cglobal highbd_sad%1x%2, 4, %3, 7, src, src_stride, ref, ref_stride, \
+                            src_stride3, ref_stride3, n_rows
+%endif ; %3 == 5/7
+%else ; avg
+%if %3 == 5
+cglobal highbd_sad%1x%2_avg, 5, 1 + %3, 7, src, src_stride, ref, ref_stride, \
+                                    second_pred, n_rows
+%else ; %3 == 7
+cglobal highbd_sad%1x%2_avg, 5, ARCH_X86_64 + %3, 7, src, src_stride, \
+                                              ref, ref_stride, \
+                                              second_pred, \
+                                              src_stride3, ref_stride3
+%if ARCH_X86_64
+%define n_rowsd r7d
+%else ; x86-32
+%define n_rowsd dword r0m
+%endif ; x86-32/64
+%endif ; %3 == 5/7
+%endif ; avg/sad
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+%if %3 == 7
+  lea         src_stride3q, [src_strideq*3]
+  lea         ref_stride3q, [ref_strideq*3]
+%endif ; %3 == 7
+; convert src, ref & second_pred to short ptrs (from byte ptrs)
+  shl                 srcq, 1
+  shl                 refq, 1
+%if %4 == 1
+  shl         second_predq, 1
+%endif
+%endmacro
+
+; unsigned int aom_highbd_sad64x{16,32,64}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD64XN 1-2 0
+  HIGH_SAD_FN 64, %1, 5, %2
+  mov              n_rowsd, %1
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  ; first half of each row
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+16]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+16]
+  por                   m2, m5
+  mova                  m5, [srcq+32]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+32]
+  por                   m3, m5
+  mova                  m5, [srcq+48]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+48]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  paddd                 m0, m1
+  paddd                 m0, m3
+  ; second half of each row
+  movu                  m1, [refq+64]
+  movu                  m2, [refq+80]
+  movu                  m3, [refq+96]
+  movu                  m4, [refq+112]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq+64]
+  psubusw               m5, m1
+  psubusw               m1, [srcq+64]
+  por                   m1, m5
+  mova                  m5, [srcq+80]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+80]
+  por                   m2, m5
+  mova                  m5, [srcq+96]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+96]
+  por                   m3, m5
+  mova                  m5, [srcq+112]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+112]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*2]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*2]
+  paddd                 m0, m3
+
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD64XN 64 ; highbd_sad64x64_sse2
+HIGH_SAD64XN 32 ; highbd_sad64x32_sse2
+HIGH_SAD64XN 64, 1 ; highbd_sad64x64_avg_sse2
+HIGH_SAD64XN 32, 1 ; highbd_sad64x32_avg_sse2
+
+
+; unsigned int aom_highbd_sad32x{16,32,64}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD32XN 1-2 0
+  HIGH_SAD_FN 32, %1, 5, %2
+  mov              n_rowsd, %1
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+16]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+16]
+  por                   m2, m5
+  mova                  m5, [srcq+32]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+32]
+  por                   m3, m5
+  mova                  m5, [srcq+48]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+48]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*2]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*2]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD32XN 64 ; highbd_sad32x64_sse2
+HIGH_SAD32XN 32 ; highbd_sad32x32_sse2
+HIGH_SAD32XN 16 ; highbd_sad32x16_sse2
+HIGH_SAD32XN 64, 1 ; highbd_sad32x64_avg_sse2
+HIGH_SAD32XN 32, 1 ; highbd_sad32x32_avg_sse2
+HIGH_SAD32XN 16, 1 ; highbd_sad32x16_avg_sse2
+
+; unsigned int aom_highbd_sad16x{8,16,32}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD16XN 1-2 0
+  HIGH_SAD_FN 16, %1, 5, %2
+  mov              n_rowsd, %1/2
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+ref_strideq*2]
+  movu                  m4, [refq+ref_strideq*2+16]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+16]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*2+16]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+16]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+16]
+  por                   m2, m5
+  mova                  m5, [srcq+src_strideq*2]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+src_strideq*2]
+  por                   m3, m5
+  mova                  m5, [srcq+src_strideq*2+16]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+src_strideq*2+16]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD16XN 32 ; highbd_sad16x32_sse2
+HIGH_SAD16XN 16 ; highbd_sad16x16_sse2
+HIGH_SAD16XN  8 ; highbd_sad16x8_sse2
+HIGH_SAD16XN 32, 1 ; highbd_sad16x32_avg_sse2
+HIGH_SAD16XN 16, 1 ; highbd_sad16x16_avg_sse2
+HIGH_SAD16XN  8, 1 ; highbd_sad16x8_avg_sse2
+
+
+; unsigned int aom_highbd_sad8x{4,8,16}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD8XN 1-2 0
+  HIGH_SAD_FN 8, %1, 7, %2
+  mov              n_rowsd, %1/4
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+ref_strideq*2]
+  movu                  m3, [refq+ref_strideq*4]
+  movu                  m4, [refq+ref_stride3q*2]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+src_strideq*2]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+src_strideq*2]
+  por                   m2, m5
+  mova                  m5, [srcq+src_strideq*4]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+src_strideq*4]
+  por                   m3, m5
+  mova                  m5, [srcq+src_stride3q*2]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+src_stride3q*2]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*8]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*8]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD8XN 16 ; highbd_sad8x16_sse2
+HIGH_SAD8XN  8 ; highbd_sad8x8_sse2
+HIGH_SAD8XN  4 ; highbd_sad8x4_sse2
+HIGH_SAD8XN 16, 1 ; highbd_sad8x16_avg_sse2
+HIGH_SAD8XN  8, 1 ; highbd_sad8x8_avg_sse2
+HIGH_SAD8XN  4, 1 ; highbd_sad8x4_avg_sse2
diff --git a/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm
new file mode 100644
index 0000000000..797e9c1d44
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm
@@ -0,0 +1,1040 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_8: times  8 dw  8
+bilin_filter_m_sse2: times  8 dw 16
+                     times  8 dw  0
+                     times  8 dw 14
+                     times  8 dw  2
+                     times  8 dw 12
+                     times  8 dw  4
+                     times  8 dw 10
+                     times  8 dw  6
+                     times 16 dw  8
+                     times  8 dw  6
+                     times  8 dw 10
+                     times  8 dw  4
+                     times  8 dw 12
+                     times  8 dw  2
+                     times  8 dw 14
+
+SECTION .text
+
+; int aom_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride,
+;                               int x_offset, int y_offset,
+;                               const uint8_t *dst, ptrdiff_t dst_stride,
+;                               int height, unsigned int *sse);
+;
+; This function returns the SE and stores SSE in the given pointer.
+
+%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse
+  psubw                %3, %4
+  psubw                %1, %2
+  mova                 %4, %3       ; make copies to manipulate to calc sum
+  mova                 %2, %1       ; use originals for calc sse
+  pmaddwd              %3, %3
+  paddw                %4, %2
+  pmaddwd              %1, %1
+  movhlps              %2, %4
+  paddd                %6, %3
+  paddw                %4, %2
+  pxor                 %2, %2
+  pcmpgtw              %2, %4       ; mask for 0 > %4 (sum)
+  punpcklwd            %4, %2       ; sign-extend word to dword
+  paddd                %6, %1
+  paddd                %5, %4
+
+%endmacro
+
+%macro STORE_AND_RET 0
+%if mmsize == 16
+  ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit
+  ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg.
+  ; We have to sign-extend it before adding the words within the register
+  ; and outputing to a dword.
+  movhlps              m3, m7
+  movhlps              m4, m6
+  paddd                m7, m3
+  paddd                m6, m4
+  pshufd               m3, m7, 0x1
+  pshufd               m4, m6, 0x1
+  paddd                m7, m3
+  paddd                m6, m4
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  movd               [r1], m7           ; store sse
+  movd                rax, m6           ; store sum as return value
+%endif
+  RET
+%endmacro
+
+%macro INC_SRC_BY_SRC_STRIDE  0
+%if ARCH_X86=1 && CONFIG_PIC=1
+  add                srcq, src_stridemp
+  add                srcq, src_stridemp
+%else
+  lea                srcq, [srcq + src_strideq*2]
+%endif
+%endmacro
+
+%macro SUBPEL_VARIANCE 1-2 0 ; W
+%define bilin_filter_m bilin_filter_m_sse2
+%define filter_idx_shift 5
+
+
+%ifdef PIC    ; 64bit PIC
+  %if %2 == 1 ; avg
+    cglobal highbd_sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \
+                                      x_offset, y_offset, \
+                                      dst, dst_stride, \
+                                      sec, sec_stride, height, sse
+    %define sec_str sec_strideq
+  %else
+    cglobal highbd_sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, x_offset, \
+                                  y_offset, dst, dst_stride, height, sse
+  %endif
+  %define block_height heightd
+  %define bilin_filter sseq
+%else
+  %if ARCH_X86=1 && CONFIG_PIC=1
+    %if %2 == 1 ; avg
+      cglobal highbd_sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \
+                                  x_offset, y_offset, \
+                                  dst, dst_stride, \
+                                  sec, sec_stride, \
+                                  height, sse, g_bilin_filter, g_pw_8
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+
+      ; Store bilin_filter and pw_8 location in stack
+      %if GET_GOT_DEFINED == 1
+        GET_GOT eax
+        add esp, 4                ; restore esp
+      %endif
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %else
+      cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \
+                                x_offset, y_offset, dst, dst_stride, height, \
+                                sse, g_bilin_filter, g_pw_8
+      %define block_height heightd
+
+      ; Store bilin_filter and pw_8 location in stack
+      %if GET_GOT_DEFINED == 1
+        GET_GOT eax
+        add esp, 4                ; restore esp
+      %endif
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %endif
+  %else
+    %if %2 == 1 ; avg
+      cglobal highbd_sub_pixel_avg_variance%1xh, 7 + 2 * ARCH_X86_64, \
+                        7 + 2 * ARCH_X86_64, 13, src, src_stride, \
+                                             x_offset, y_offset, \
+                                             dst, dst_stride, \
+                                             sec, sec_stride, \
+                                             height, sse
+      %if ARCH_X86_64
+      %define block_height heightd
+      %define sec_str sec_strideq
+      %else
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+      %endif
+    %else
+      cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \
+                              x_offset, y_offset, dst, dst_stride, height, sse
+      %define block_height heightd
+    %endif
+
+    %define bilin_filter bilin_filter_m
+  %endif
+%endif
+
+  ASSERT               %1 <= 16         ; m6 overflows if w > 16
+  pxor                 m6, m6           ; sum
+  pxor                 m7, m7           ; sse
+
+%if %1 < 16
+  sar                   block_height, 1
+%endif
+%if %2 == 1 ; avg
+  shl             sec_str, 1
+%endif
+
+  ; FIXME(rbultje) replace by jumptable?
+  test          x_offsetd, x_offsetd
+  jnz .x_nonzero
+  ; x_offset == 0
+  test          y_offsetd, y_offsetd
+  jnz .x_zero_y_nonzero
+
+  ; x_offset == 0 && y_offset == 0
+.x_zero_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq + 16]
+  mova                 m1, [dstq]
+  mova                 m3, [dstq + 16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m2, [secq+16]
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq + src_strideq*2]
+  mova                 m1, [dstq]
+  mova                 m3, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_zero_y_zero_loop
+  STORE_AND_RET
+
+.x_zero_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_zero_y_nonhalf
+
+  ; x_offset == 0 && y_offset == 0.5
+.x_zero_y_half_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m4, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*2+16]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+16]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*4]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+dst_strideq*2]
+  pavgw                m0, m1
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_zero_y_half_loop
+  STORE_AND_RET
+
+.x_zero_y_nonhalf:
+  ; x_offset == 0 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+y_offsetq]
+  mova                 m9, [bilin_filter+y_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else ; x86-32 or mmx
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0, reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_zero_y_other_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + 16]
+  movu                 m4, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*2+16]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+16]
+  ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can
+  ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of
+  ; instructions is the same (5), but it is 1 mul instead of 2, so might be
+  ; slightly faster because of pmullw latency. It would also cut our rodata
+  ; tables in half for this function, and save 1-2 registers on x86-64.
+  pmullw               m1, filter_y_a
+  pmullw               m5, filter_y_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m1, m5
+  paddw                m0, m4
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*4]
+  mova                 m4, m1
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+dst_strideq*2]
+  pmullw               m1, filter_y_a
+  pmullw               m5, filter_y_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m1, m5
+  paddw                m0, m4
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_zero_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonzero:
+  cmp           x_offsetd, 8
+  jne .x_nonhalf
+  ; x_offset == 0.5
+  test          y_offsetd, y_offsetd
+  jnz .x_half_y_nonzero
+
+  ; x_offset == 0.5 && y_offset == 0
+.x_half_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + 16]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + 18]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq + 16]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + src_strideq*2]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + src_strideq*2 + 2]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq + dst_strideq*2]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_half_y_zero_loop
+  STORE_AND_RET
+
+.x_half_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_half_y_nonhalf
+
+  ; x_offset == 0.5 && y_offset == 0.5
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+  pavgw                m1, m3
+.x_half_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq + 16]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + 18]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  pavgw                m0, m2
+  pavgw                m1, m3
+  mova                 m4, [dstq]
+  mova                 m5, [dstq + 16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+  mova                 m0, m2
+  mova                 m1, m3
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+.x_half_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq + src_strideq*2]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + src_strideq*2 + 2]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  pavgw                m0, m2
+  pavgw                m2, m3
+  mova                 m4, [dstq]
+  mova                 m5, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m4, m2, m5, m6, m7
+  mova                 m0, m3
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_half_y_half_loop
+  STORE_AND_RET
+
+.x_half_y_nonhalf:
+  ; x_offset == 0.5 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+y_offsetq]
+  mova                 m9, [bilin_filter+y_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else  ; x86_32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0.5. We can reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+  pavgw                m1, m3
+.x_half_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+16]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+18]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m1, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m1, filter_rnd
+  paddw                m1, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  psrlw                m1, 4
+  paddw                m0, m2
+  mova                 m2, [dstq]
+  psrlw                m0, 4
+  mova                 m3, [dstq+16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+  mova                 m0, m4
+  mova                 m1, m5
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+.x_half_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+src_strideq*2]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+src_strideq*2+2]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m4, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m4, filter_rnd
+  paddw                m4, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  psrlw                m4, 4
+  paddw                m0, m2
+  mova                 m2, [dstq]
+  psrlw                m0, 4
+  mova                 m3, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m4, [secq]
+%endif
+  SUM_SSE              m0, m2, m4, m3, m6, m7
+  mova                 m0, m5
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_half_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf:
+  test          y_offsetd, y_offsetd
+  jnz .x_nonhalf_y_nonzero
+
+  ; x_offset == bilin interpolation && y_offset == 0
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_other_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+16]
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m1, m3
+  paddw                m0, m2
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+src_strideq*2+2]
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+dst_strideq*2]
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m1, m3
+  paddw                m0, m2
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+
+  lea                srcq, [srcq+src_strideq*4]
+  lea                dstq, [dstq+dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_other_y_zero_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_nonhalf_y_nonhalf
+
+  ; x_offset == bilin interpolation && y_offset == 0.5
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0.5. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m0, m2
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+  lea                srcq, [srcq+src_strideq*2]
+.x_other_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+16]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+18]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+16]
+  psrlw                m2, 4
+  psrlw                m3, 4
+  pavgw                m0, m2
+  pavgw                m1, m3
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+  mova                 m0, m2
+  mova                 m1, m3
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m2
+  psrlw                m0, 4
+  lea                srcq, [srcq+src_strideq*2]
+.x_other_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+src_strideq*2]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+src_strideq*2+2]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+dst_strideq*2]
+  psrlw                m2, 4
+  psrlw                m3, 4
+  pavgw                m0, m2
+  pavgw                m2, m3
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m4, m2, m5, m6, m7
+  mova                 m0, m3
+
+  lea                srcq, [srcq+src_strideq*4]
+  lea                dstq, [dstq+dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_other_y_half_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf_y_nonhalf:
+; loading filter - this is same as in 8-bit depth
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift ; filter_idx_shift = 5
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [bilin_filter+y_offsetq]
+  mova                m11, [bilin_filter+y_offsetq+16]
+  mova                m12, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_y_a m10
+%define filter_y_b m11
+%define filter_rnd m12
+%else   ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; In this case, there is NO unused register. Used src_stride register. Later,
+; src_stride has to be loaded from stack when it is needed.
+%define tempq src_strideq
+  mov tempq, g_bilin_filterm
+  add           x_offsetq, tempq
+  add           y_offsetq, tempq
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+  add           y_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+; end of load filter
+
+  ; x_offset == bilin interpolation && y_offset == bilin interpolation
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  movu                 m1, [srcq+16]
+  movu                 m3, [srcq+18]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m0, m2
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m4, [srcq+2]
+  movu                 m3, [srcq+16]
+  movu                 m5, [srcq+18]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  psrlw                m2, 4
+  psrlw                m3, 4
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, m2
+  paddw                m1, filter_rnd
+  mova                 m2, [dstq]
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+  mova                 m3, [dstq+16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+  mova                 m0, m4
+  mova                 m1, m5
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq + dst_strideq * 2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m2
+  psrlw                m0, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m4, [srcq+2]
+  INC_SRC_BY_SRC_STRIDE
+  movu                 m3, [srcq]
+  movu                 m5, [srcq+2]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  psrlw                m2, 4
+  psrlw                m3, 4
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m4, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, m2
+  paddw                m4, filter_rnd
+  mova                 m2, [dstq]
+  paddw                m4, m3
+  psrlw                m0, 4
+  psrlw                m4, 4
+  mova                 m3, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m4, [secq]
+%endif
+  SUM_SSE              m0, m2, m4, m3, m6, m7
+  mova                 m0, m5
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq + dst_strideq * 4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_other_y_other_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+%endmacro
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
diff --git a/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c b/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c
new file mode 100644
index 0000000000..7bc8a0df32
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c
@@ -0,0 +1,364 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>
+#include <stddef.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+typedef void (*SubtractWxHFuncType)(int16_t *diff, ptrdiff_t diff_stride,
+                                    const uint16_t *src, ptrdiff_t src_stride,
+                                    const uint16_t *pred,
+                                    ptrdiff_t pred_stride);
+
+static void subtract_4x4(int16_t *diff, ptrdiff_t diff_stride,
+                         const uint16_t *src, ptrdiff_t src_stride,
+                         const uint16_t *pred, ptrdiff_t pred_stride) {
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3;
+  __m128i x0, x1, x2, x3;
+  int64_t *store_diff = (int64_t *)(diff + 0 * diff_stride);
+
+  u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride));
+  u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride));
+  u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+
+  v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride));
+  v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride));
+  v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride));
+  v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride));
+
+  x0 = _mm_sub_epi16(u0, v0);
+  x1 = _mm_sub_epi16(u1, v1);
+  x2 = _mm_sub_epi16(u2, v2);
+  x3 = _mm_sub_epi16(u3, v3);
+
+  _mm_storel_epi64((__m128i *)store_diff, x0);
+  store_diff = (int64_t *)(diff + 1 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x1);
+  store_diff = (int64_t *)(diff + 2 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x2);
+  store_diff = (int64_t *)(diff + 3 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x3);
+}
+
+static void subtract_4x8(int16_t *diff, ptrdiff_t diff_stride,
+                         const uint16_t *src, ptrdiff_t src_stride,
+                         const uint16_t *pred, ptrdiff_t pred_stride) {
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  int64_t *store_diff = (int64_t *)(diff + 0 * diff_stride);
+
+  u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride));
+  u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride));
+  u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+  u4 = _mm_loadu_si128((__m128i const *)(src + 4 * src_stride));
+  u5 = _mm_loadu_si128((__m128i const *)(src + 5 * src_stride));
+  u6 = _mm_loadu_si128((__m128i const *)(src + 6 * src_stride));
+  u7 = _mm_loadu_si128((__m128i const *)(src + 7 * src_stride));
+
+  v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride));
+  v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride));
+  v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride));
+  v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride));
+  v4 = _mm_loadu_si128((__m128i const *)(pred + 4 * pred_stride));
+  v5 = _mm_loadu_si128((__m128i const *)(pred + 5 * pred_stride));
+  v6 = _mm_loadu_si128((__m128i const *)(pred + 6 * pred_stride));
+  v7 = _mm_loadu_si128((__m128i const *)(pred + 7 * pred_stride));
+
+  x0 = _mm_sub_epi16(u0, v0);
+  x1 = _mm_sub_epi16(u1, v1);
+  x2 = _mm_sub_epi16(u2, v2);
+  x3 = _mm_sub_epi16(u3, v3);
+  x4 = _mm_sub_epi16(u4, v4);
+  x5 = _mm_sub_epi16(u5, v5);
+  x6 = _mm_sub_epi16(u6, v6);
+  x7 = _mm_sub_epi16(u7, v7);
+
+  _mm_storel_epi64((__m128i *)store_diff, x0);
+  store_diff = (int64_t *)(diff + 1 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x1);
+  store_diff = (int64_t *)(diff + 2 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x2);
+  store_diff = (int64_t *)(diff + 3 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x3);
+  store_diff = (int64_t *)(diff + 4 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x4);
+  store_diff = (int64_t *)(diff + 5 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x5);
+  store_diff = (int64_t *)(diff + 6 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x6);
+  store_diff = (int64_t *)(diff + 7 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x7);
+}
+
+static void subtract_8x4(int16_t *diff, ptrdiff_t diff_stride,
+                         const uint16_t *src, ptrdiff_t src_stride,
+                         const uint16_t *pred, ptrdiff_t pred_stride) {
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3;
+  __m128i x0, x1, x2, x3;
+
+  u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride));
+  u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride));
+  u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+
+  v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride));
+  v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride));
+  v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride));
+  v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride));
+
+  x0 = _mm_sub_epi16(u0, v0);
+  x1 = _mm_sub_epi16(u1, v1);
+  x2 = _mm_sub_epi16(u2, v2);
+  x3 = _mm_sub_epi16(u3, v3);
+
+  _mm_storeu_si128((__m128i *)(diff + 0 * diff_stride), x0);
+  _mm_storeu_si128((__m128i *)(diff + 1 * diff_stride), x1);
+  _mm_storeu_si128((__m128i *)(diff + 2 * diff_stride), x2);
+  _mm_storeu_si128((__m128i *)(diff + 3 * diff_stride), x3);
+}
+
+static void subtract_8x8(int16_t *diff, ptrdiff_t diff_stride,
+                         const uint16_t *src, ptrdiff_t src_stride,
+                         const uint16_t *pred, ptrdiff_t pred_stride) {
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+
+  u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride));
+  u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride));
+  u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+  u4 = _mm_loadu_si128((__m128i const *)(src + 4 * src_stride));
+  u5 = _mm_loadu_si128((__m128i const *)(src + 5 * src_stride));
+  u6 = _mm_loadu_si128((__m128i const *)(src + 6 * src_stride));
+  u7 = _mm_loadu_si128((__m128i const *)(src + 7 * src_stride));
+
+  v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride));
+  v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride));
+  v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride));
+  v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride));
+  v4 = _mm_loadu_si128((__m128i const *)(pred + 4 * pred_stride));
+  v5 = _mm_loadu_si128((__m128i const *)(pred + 5 * pred_stride));
+  v6 = _mm_loadu_si128((__m128i const *)(pred + 6 * pred_stride));
+  v7 = _mm_loadu_si128((__m128i const *)(pred + 7 * pred_stride));
+
+  x0 = _mm_sub_epi16(u0, v0);
+  x1 = _mm_sub_epi16(u1, v1);
+  x2 = _mm_sub_epi16(u2, v2);
+  x3 = _mm_sub_epi16(u3, v3);
+  x4 = _mm_sub_epi16(u4, v4);
+  x5 = _mm_sub_epi16(u5, v5);
+  x6 = _mm_sub_epi16(u6, v6);
+  x7 = _mm_sub_epi16(u7, v7);
+
+  _mm_storeu_si128((__m128i *)(diff + 0 * diff_stride), x0);
+  _mm_storeu_si128((__m128i *)(diff + 1 * diff_stride), x1);
+  _mm_storeu_si128((__m128i *)(diff + 2 * diff_stride), x2);
+  _mm_storeu_si128((__m128i *)(diff + 3 * diff_stride), x3);
+  _mm_storeu_si128((__m128i *)(diff + 4 * diff_stride), x4);
+  _mm_storeu_si128((__m128i *)(diff + 5 * diff_stride), x5);
+  _mm_storeu_si128((__m128i *)(diff + 6 * diff_stride), x6);
+  _mm_storeu_si128((__m128i *)(diff + 7 * diff_stride), x7);
+}
+
+static void subtract_8x16(int16_t *diff, ptrdiff_t diff_stride,
+                          const uint16_t *src, ptrdiff_t src_stride,
+                          const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_8x8(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += diff_stride << 3;
+  src += src_stride << 3;
+  pred += pred_stride << 3;
+  subtract_8x8(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_16x8(int16_t *diff, ptrdiff_t diff_stride,
+                          const uint16_t *src, ptrdiff_t src_stride,
+                          const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_8x8(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += 8;
+  src += 8;
+  pred += 8;
+  subtract_8x8(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_16x16(int16_t *diff, ptrdiff_t diff_stride,
+                           const uint16_t *src, ptrdiff_t src_stride,
+                           const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_16x8(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += diff_stride << 3;
+  src += src_stride << 3;
+  pred += pred_stride << 3;
+  subtract_16x8(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_16x32(int16_t *diff, ptrdiff_t diff_stride,
+                           const uint16_t *src, ptrdiff_t src_stride,
+                           const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_16x16(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += diff_stride << 4;
+  src += src_stride << 4;
+  pred += pred_stride << 4;
+  subtract_16x16(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_32x16(int16_t *diff, ptrdiff_t diff_stride,
+                           const uint16_t *src, ptrdiff_t src_stride,
+                           const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_16x16(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += 16;
+  src += 16;
+  pred += 16;
+  subtract_16x16(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_32x32(int16_t *diff, ptrdiff_t diff_stride,
+                           const uint16_t *src, ptrdiff_t src_stride,
+                           const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_32x16(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += diff_stride << 4;
+  src += src_stride << 4;
+  pred += pred_stride << 4;
+  subtract_32x16(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_32x64(int16_t *diff, ptrdiff_t diff_stride,
+                           const uint16_t *src, ptrdiff_t src_stride,
+                           const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_32x32(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += diff_stride << 5;
+  src += src_stride << 5;
+  pred += pred_stride << 5;
+  subtract_32x32(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_64x32(int16_t *diff, ptrdiff_t diff_stride,
+                           const uint16_t *src, ptrdiff_t src_stride,
+                           const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_32x32(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += 32;
+  src += 32;
+  pred += 32;
+  subtract_32x32(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_64x64(int16_t *diff, ptrdiff_t diff_stride,
+                           const uint16_t *src, ptrdiff_t src_stride,
+                           const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_64x32(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += diff_stride << 5;
+  src += src_stride << 5;
+  pred += pred_stride << 5;
+  subtract_64x32(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_64x128(int16_t *diff, ptrdiff_t diff_stride,
+                            const uint16_t *src, ptrdiff_t src_stride,
+                            const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_64x64(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += diff_stride << 6;
+  src += src_stride << 6;
+  pred += pred_stride << 6;
+  subtract_64x64(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_128x64(int16_t *diff, ptrdiff_t diff_stride,
+                            const uint16_t *src, ptrdiff_t src_stride,
+                            const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_64x64(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += 64;
+  src += 64;
+  pred += 64;
+  subtract_64x64(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static void subtract_128x128(int16_t *diff, ptrdiff_t diff_stride,
+                             const uint16_t *src, ptrdiff_t src_stride,
+                             const uint16_t *pred, ptrdiff_t pred_stride) {
+  subtract_128x64(diff, diff_stride, src, src_stride, pred, pred_stride);
+  diff += diff_stride << 6;
+  src += src_stride << 6;
+  pred += pred_stride << 6;
+  subtract_128x64(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
+
+static SubtractWxHFuncType getSubtractFunc(int rows, int cols) {
+  SubtractWxHFuncType ret_func_ptr = NULL;
+  if (rows == 4) {
+    if (cols == 4) {
+      ret_func_ptr = subtract_4x4;
+    } else if (cols == 8) {
+      ret_func_ptr = subtract_8x4;
+    }
+  } else if (rows == 8) {
+    if (cols == 4) {
+      ret_func_ptr = subtract_4x8;
+    } else if (cols == 8) {
+      ret_func_ptr = subtract_8x8;
+    } else if (cols == 16) {
+      ret_func_ptr = subtract_16x8;
+    }
+  } else if (rows == 16) {
+    if (cols == 8) {
+      ret_func_ptr = subtract_8x16;
+    } else if (cols == 16) {
+      ret_func_ptr = subtract_16x16;
+    } else if (cols == 32) {
+      ret_func_ptr = subtract_32x16;
+    }
+  } else if (rows == 32) {
+    if (cols == 16) {
+      ret_func_ptr = subtract_16x32;
+    } else if (cols == 32) {
+      ret_func_ptr = subtract_32x32;
+    } else if (cols == 64) {
+      ret_func_ptr = subtract_64x32;
+    }
+  } else if (rows == 64) {
+    if (cols == 32) {
+      ret_func_ptr = subtract_32x64;
+    } else if (cols == 64) {
+      ret_func_ptr = subtract_64x64;
+    } else if (cols == 128) {
+      ret_func_ptr = subtract_128x64;
+    }
+  } else if (rows == 128) {
+    if (cols == 64) {
+      ret_func_ptr = subtract_64x128;
+    } else if (cols == 128) {
+      ret_func_ptr = subtract_128x128;
+    }
+  }
+  if (!ret_func_ptr) {
+    assert(0);
+  }
+  return ret_func_ptr;
+}
+
+void aom_highbd_subtract_block_sse2(int rows, int cols, int16_t *diff,
+                                    ptrdiff_t diff_stride, const uint8_t *src8,
+                                    ptrdiff_t src_stride, const uint8_t *pred8,
+                                    ptrdiff_t pred_stride, int bd) {
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  SubtractWxHFuncType func;
+  (void)bd;
+
+  func = getSubtractFunc(rows, cols);
+  func(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm
new file mode 100644
index 0000000000..cf8ea498c7
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm
@@ -0,0 +1,316 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+%include "aom_ports/x86_abi_support.asm"
+
+;unsigned int aom_highbd_calc16x16var_sse2
+;(
+;    unsigned char   *  src_ptr,
+;    int             source_stride,
+;    unsigned char   *  ref_ptr,
+;    int             recon_stride,
+;    unsigned int    *  SSE,
+;    int             *  Sum
+;)
+global sym(aom_highbd_calc16x16var_sse2) PRIVATE
+sym(aom_highbd_calc16x16var_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        mov         rsi,            arg(0) ;[src_ptr]
+        mov         rdi,            arg(2) ;[ref_ptr]
+
+        movsxd      rax,            DWORD PTR arg(1) ;[source_stride]
+        movsxd      rdx,            DWORD PTR arg(3) ;[recon_stride]
+        add         rax,            rax ; source stride in bytes
+        add         rdx,            rdx ; recon stride in bytes
+
+        ; Prefetch data
+        prefetcht0      [rsi]
+        prefetcht0      [rsi+16]
+        prefetcht0      [rsi+rax]
+        prefetcht0      [rsi+rax+16]
+        lea             rbx,    [rsi+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rax]
+        prefetcht0      [rbx+rax+16]
+
+        prefetcht0      [rdi]
+        prefetcht0      [rdi+16]
+        prefetcht0      [rdi+rdx]
+        prefetcht0      [rdi+rdx+16]
+        lea             rbx,    [rdi+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rdx]
+        prefetcht0      [rbx+rdx+16]
+
+        pxor        xmm0,           xmm0     ; clear xmm0 for unpack
+        pxor        xmm7,           xmm7     ; clear xmm7 for accumulating diffs
+
+        pxor        xmm6,           xmm6     ; clear xmm6 for accumulating sse
+        mov         rcx,            16
+
+.var16loop:
+        movdqu      xmm1,           XMMWORD PTR [rsi]
+        movdqu      xmm2,           XMMWORD PTR [rdi]
+
+        lea             rbx,    [rsi+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rax]
+        prefetcht0      [rbx+rax+16]
+        lea             rbx,    [rdi+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rdx]
+        prefetcht0      [rbx+rdx+16]
+
+        pxor        xmm5,           xmm5
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+16]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+16]
+        paddd       xmm6,           xmm1
+
+        psubw       xmm3,           xmm2
+        movdqu      xmm1,           XMMWORD PTR [rsi+rax]
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx]
+        paddd       xmm6,           xmm3
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+rax+16]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx+16]
+        paddd       xmm6,           xmm1
+
+        psubw       xmm3,           xmm2
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        paddd       xmm6,           xmm3
+
+        movdqa      xmm1,           xmm5
+        movdqa      xmm2,           xmm5
+        pcmpgtw     xmm1,           xmm0
+        pcmpeqw     xmm2,           xmm0
+        por         xmm1,           xmm2
+        pcmpeqw     xmm1,           xmm0
+        movdqa      xmm2,           xmm5
+        punpcklwd   xmm5,           xmm1
+        punpckhwd   xmm2,           xmm1
+        paddd       xmm7,           xmm5
+        paddd       xmm7,           xmm2
+
+        lea         rsi,            [rsi + 2*rax]
+        lea         rdi,            [rdi + 2*rdx]
+        sub         rcx,            2
+        jnz         .var16loop
+
+        movdqa      xmm4,           xmm6
+        punpckldq   xmm6,           xmm0
+
+        punpckhdq   xmm4,           xmm0
+        movdqa      xmm5,           xmm7
+
+        paddd       xmm6,           xmm4
+        punpckldq   xmm7,           xmm0
+
+        punpckhdq   xmm5,           xmm0
+        paddd       xmm7,           xmm5
+
+        movdqa      xmm4,           xmm6
+        movdqa      xmm5,           xmm7
+
+        psrldq      xmm4,           8
+        psrldq      xmm5,           8
+
+        paddd       xmm6,           xmm4
+        paddd       xmm7,           xmm5
+
+        mov         rdi,            arg(4)   ; [SSE]
+        mov         rax,            arg(5)   ; [Sum]
+
+        movd DWORD PTR [rdi],       xmm6
+        movd DWORD PTR [rax],       xmm7
+
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    pop rbx
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;unsigned int aom_highbd_calc8x8var_sse2
+;(
+;    unsigned char   *  src_ptr,
+;    int             source_stride,
+;    unsigned char   *  ref_ptr,
+;    int             recon_stride,
+;    unsigned int    *  SSE,
+;    int             *  Sum
+;)
+global sym(aom_highbd_calc8x8var_sse2) PRIVATE
+sym(aom_highbd_calc8x8var_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        mov         rsi,            arg(0) ;[src_ptr]
+        mov         rdi,            arg(2) ;[ref_ptr]
+
+        movsxd      rax,            DWORD PTR arg(1) ;[source_stride]
+        movsxd      rdx,            DWORD PTR arg(3) ;[recon_stride]
+        add         rax,            rax ; source stride in bytes
+        add         rdx,            rdx ; recon stride in bytes
+
+        ; Prefetch data
+        prefetcht0      [rsi]
+        prefetcht0      [rsi+rax]
+        lea             rbx,    [rsi+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rax]
+
+        prefetcht0      [rdi]
+        prefetcht0      [rdi+rdx]
+        lea             rbx,    [rdi+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rdx]
+
+        pxor        xmm0,           xmm0     ; clear xmm0 for unpack
+        pxor        xmm7,           xmm7     ; clear xmm7 for accumulating diffs
+
+        pxor        xmm6,           xmm6     ; clear xmm6 for accumulating sse
+        mov         rcx,            8
+
+.var8loop:
+        movdqu      xmm1,           XMMWORD PTR [rsi]
+        movdqu      xmm2,           XMMWORD PTR [rdi]
+
+        lea             rbx,    [rsi+rax*4]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rax]
+        lea             rbx,    [rbx+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rax]
+        lea             rbx,    [rdi+rdx*4]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rdx]
+        lea             rbx,    [rbx+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rdx]
+
+        pxor        xmm5,           xmm5
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+rax]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx]
+        paddd       xmm6,           xmm1
+
+        lea         rsi,            [rsi + 2*rax]
+        lea         rdi,            [rdi + 2*rdx]
+
+        psubw       xmm3,           xmm2
+        movdqu      xmm1,           XMMWORD PTR [rsi]
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        movdqu      xmm2,           XMMWORD PTR [rdi]
+        paddd       xmm6,           xmm3
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+rax]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx]
+        paddd       xmm6,           xmm1
+
+        psubw       xmm3,           xmm2
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        paddd       xmm6,           xmm3
+
+        movdqa      xmm1,           xmm5
+        movdqa      xmm2,           xmm5
+        pcmpgtw     xmm1,           xmm0
+        pcmpeqw     xmm2,           xmm0
+        por         xmm1,           xmm2
+        pcmpeqw     xmm1,           xmm0
+        movdqa      xmm2,           xmm5
+        punpcklwd   xmm5,           xmm1
+        punpckhwd   xmm2,           xmm1
+        paddd       xmm7,           xmm5
+        paddd       xmm7,           xmm2
+
+        lea         rsi,            [rsi + 2*rax]
+        lea         rdi,            [rdi + 2*rdx]
+        sub         rcx,            4
+        jnz         .var8loop
+
+        movdqa      xmm4,           xmm6
+        punpckldq   xmm6,           xmm0
+
+        punpckhdq   xmm4,           xmm0
+        movdqa      xmm5,           xmm7
+
+        paddd       xmm6,           xmm4
+        punpckldq   xmm7,           xmm0
+
+        punpckhdq   xmm5,           xmm0
+        paddd       xmm7,           xmm5
+
+        movdqa      xmm4,           xmm6
+        movdqa      xmm5,           xmm7
+
+        psrldq      xmm4,           8
+        psrldq      xmm5,           8
+
+        paddd       xmm6,           xmm4
+        paddd       xmm7,           xmm5
+
+        mov         rdi,            arg(4)   ; [SSE]
+        mov         rax,            arg(5)   ; [Sum]
+
+        movd DWORD PTR [rdi],       xmm6
+        movd DWORD PTR [rax],       xmm7
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    pop rbx
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c b/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c
new file mode 100644
index 0000000000..29f96ce248
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c
@@ -0,0 +1,695 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom_ports/mem.h"
+
+typedef uint32_t (*high_variance_fn_t)(const uint16_t *src, int src_stride,
+                                       const uint16_t *ref, int ref_stride,
+                                       uint32_t *sse, int *sum);
+
+uint32_t aom_highbd_calc8x8var_sse2(const uint16_t *src, int src_stride,
+                                    const uint16_t *ref, int ref_stride,
+                                    uint32_t *sse, int *sum);
+
+uint32_t aom_highbd_calc16x16var_sse2(const uint16_t *src, int src_stride,
+                                      const uint16_t *ref, int ref_stride,
+                                      uint32_t *sse, int *sum);
+
+static void highbd_8_variance_sse2(const uint16_t *src, int src_stride,
+                                   const uint16_t *ref, int ref_stride, int w,
+                                   int h, uint32_t *sse, int *sum,
+                                   high_variance_fn_t var_fn, int block_size) {
+  int i, j;
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
+             ref_stride, &sse0, &sum0);
+      *sse += sse0;
+      *sum += sum0;
+    }
+  }
+}
+
+static void highbd_10_variance_sse2(const uint16_t *src, int src_stride,
+                                    const uint16_t *ref, int ref_stride, int w,
+                                    int h, uint32_t *sse, int *sum,
+                                    high_variance_fn_t var_fn, int block_size) {
+  int i, j;
+  uint64_t sse_long = 0;
+  int32_t sum_long = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
+             ref_stride, &sse0, &sum0);
+      sse_long += sse0;
+      sum_long += sum0;
+    }
+  }
+  *sum = ROUND_POWER_OF_TWO(sum_long, 2);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4);
+}
+
+static void highbd_12_variance_sse2(const uint16_t *src, int src_stride,
+                                    const uint16_t *ref, int ref_stride, int w,
+                                    int h, uint32_t *sse, int *sum,
+                                    high_variance_fn_t var_fn, int block_size) {
+  int i, j;
+  uint64_t sse_long = 0;
+  int32_t sum_long = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
+             ref_stride, &sse0, &sum0);
+      sse_long += sse0;
+      sum_long += sum0;
+    }
+  }
+  *sum = ROUND_POWER_OF_TWO(sum_long, 4);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8);
+}
+
+#define HIGH_GET_VAR(S)                                                       \
+  void aom_highbd_get##S##x##S##var_sse2(const uint8_t *src8, int src_stride, \
+                                         const uint8_t *ref8, int ref_stride, \
+                                         uint32_t *sse, int *sum) {           \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                \
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                                \
+    aom_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \
+                                       sum);                                  \
+  }                                                                           \
+                                                                              \
+  void aom_highbd_10_get##S##x##S##var_sse2(                                  \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,               \
+      int ref_stride, uint32_t *sse, int *sum) {                              \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                \
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                                \
+    aom_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \
+                                       sum);                                  \
+    *sum = ROUND_POWER_OF_TWO(*sum, 2);                                       \
+    *sse = ROUND_POWER_OF_TWO(*sse, 4);                                       \
+  }                                                                           \
+                                                                              \
+  void aom_highbd_12_get##S##x##S##var_sse2(                                  \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,               \
+      int ref_stride, uint32_t *sse, int *sum) {                              \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                \
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                                \
+    aom_highbd_calc##S##x##S##var_sse2(src, src_stride, ref, ref_stride, sse, \
+                                       sum);                                  \
+    *sum = ROUND_POWER_OF_TWO(*sum, 4);                                       \
+    *sse = ROUND_POWER_OF_TWO(*sse, 8);                                       \
+  }
+
+HIGH_GET_VAR(16);
+HIGH_GET_VAR(8);
+
+#undef HIGH_GET_VAR
+
+#define VAR_FN(w, h, block_size, shift)                                    \
+  uint32_t aom_highbd_8_variance##w##x##h##_sse2(                          \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,            \
+      int ref_stride, uint32_t *sse) {                                     \
+    int sum;                                                               \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                             \
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                             \
+    highbd_8_variance_sse2(                                                \
+        src, src_stride, ref, ref_stride, w, h, sse, &sum,                 \
+        aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
+    return *sse - (uint32_t)(((int64_t)sum * sum) >> shift);               \
+  }                                                                        \
+                                                                           \
+  uint32_t aom_highbd_10_variance##w##x##h##_sse2(                         \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,            \
+      int ref_stride, uint32_t *sse) {                                     \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                             \
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                             \
+    highbd_10_variance_sse2(                                               \
+        src, src_stride, ref, ref_stride, w, h, sse, &sum,                 \
+        aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) >> shift);               \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }                                                                        \
+                                                                           \
+  uint32_t aom_highbd_12_variance##w##x##h##_sse2(                         \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,            \
+      int ref_stride, uint32_t *sse) {                                     \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                             \
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                             \
+    highbd_12_variance_sse2(                                               \
+        src, src_stride, ref, ref_stride, w, h, sse, &sum,                 \
+        aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) >> shift);               \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }
+
+VAR_FN(64, 64, 16, 12);
+VAR_FN(64, 32, 16, 11);
+VAR_FN(32, 64, 16, 11);
+VAR_FN(32, 32, 16, 10);
+VAR_FN(32, 16, 16, 9);
+VAR_FN(16, 32, 16, 9);
+VAR_FN(16, 16, 16, 8);
+VAR_FN(16, 8, 8, 7);
+VAR_FN(8, 16, 8, 7);
+VAR_FN(8, 8, 8, 6);
+
+#undef VAR_FN
+
+unsigned int aom_highbd_8_mse16x16_sse2(const uint8_t *src8, int src_stride,
+                                        const uint8_t *ref8, int ref_stride,
+                                        unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum,
+                         aom_highbd_calc16x16var_sse2, 16);
+  return *sse;
+}
+
+unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src8, int src_stride,
+                                         const uint8_t *ref8, int ref_stride,
+                                         unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum,
+                          aom_highbd_calc16x16var_sse2, 16);
+  return *sse;
+}
+
+unsigned int aom_highbd_12_mse16x16_sse2(const uint8_t *src8, int src_stride,
+                                         const uint8_t *ref8, int ref_stride,
+                                         unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum,
+                          aom_highbd_calc16x16var_sse2, 16);
+  return *sse;
+}
+
+unsigned int aom_highbd_8_mse8x8_sse2(const uint8_t *src8, int src_stride,
+                                      const uint8_t *ref8, int ref_stride,
+                                      unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum,
+                         aom_highbd_calc8x8var_sse2, 8);
+  return *sse;
+}
+
+unsigned int aom_highbd_10_mse8x8_sse2(const uint8_t *src8, int src_stride,
+                                       const uint8_t *ref8, int ref_stride,
+                                       unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum,
+                          aom_highbd_calc8x8var_sse2, 8);
+  return *sse;
+}
+
+unsigned int aom_highbd_12_mse8x8_sse2(const uint8_t *src8, int src_stride,
+                                       const uint8_t *ref8, int ref_stride,
+                                       unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum,
+                          aom_highbd_calc8x8var_sse2, 8);
+  return *sse;
+}
+
+// The 2 unused parameters are place holders for PIC enabled build.
+// These definitions are for functions defined in
+// highbd_subpel_variance_impl_sse2.asm
+#define DECL(w, opt)                                                         \
+  int aom_highbd_sub_pixel_variance##w##xh_##opt(                            \
+      const uint16_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
+      const uint16_t *dst, ptrdiff_t dst_stride, int height,                 \
+      unsigned int *sse, void *unused0, void *unused);
+#define DECLS(opt) \
+  DECL(8, opt);    \
+  DECL(16, opt)
+
+DECLS(sse2);
+
+#undef DECLS
+#undef DECL
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast)                                  \
+  uint32_t aom_highbd_8_sub_pixel_variance##w##x##h##_##opt(                   \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) {                \
+    uint32_t sse;                                                              \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    int se = aom_highbd_sub_pixel_variance##wf##xh_##opt(                      \
+        src, src_stride, x_offset, y_offset, dst, dst_stride, h, &sse, NULL,   \
+        NULL);                                                                 \
+    if (w > wf) {                                                              \
+      unsigned int sse2;                                                       \
+      int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+          src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, h,   \
+          &sse2, NULL, NULL);                                                  \
+      se += se2;                                                               \
+      sse += sse2;                                                             \
+      if (w > wf * 2) {                                                        \
+        se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                     \
+            src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, h, \
+            &sse2, NULL, NULL);                                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+        se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                     \
+            src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, h, \
+            &sse2, NULL, NULL);                                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+      }                                                                        \
+    }                                                                          \
+    *sse_ptr = sse;                                                            \
+    return sse - (uint32_t)((cast se * se) >> (wlog2 + hlog2));                \
+  }                                                                            \
+                                                                               \
+  uint32_t aom_highbd_10_sub_pixel_variance##w##x##h##_##opt(                  \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) {                \
+    int64_t var;                                                               \
+    uint32_t sse;                                                              \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    int se = aom_highbd_sub_pixel_variance##wf##xh_##opt(                      \
+        src, src_stride, x_offset, y_offset, dst, dst_stride, h, &sse, NULL,   \
+        NULL);                                                                 \
+    if (w > wf) {                                                              \
+      uint32_t sse2;                                                           \
+      int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+          src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, h,   \
+          &sse2, NULL, NULL);                                                  \
+      se += se2;                                                               \
+      sse += sse2;                                                             \
+      if (w > wf * 2) {                                                        \
+        se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                     \
+            src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, h, \
+            &sse2, NULL, NULL);                                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+        se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                     \
+            src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, h, \
+            &sse2, NULL, NULL);                                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+      }                                                                        \
+    }                                                                          \
+    se = ROUND_POWER_OF_TWO(se, 2);                                            \
+    sse = ROUND_POWER_OF_TWO(sse, 4);                                          \
+    *sse_ptr = sse;                                                            \
+    var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2));                \
+    return (var >= 0) ? (uint32_t)var : 0;                                     \
+  }                                                                            \
+                                                                               \
+  uint32_t aom_highbd_12_sub_pixel_variance##w##x##h##_##opt(                  \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) {                \
+    int start_row;                                                             \
+    uint32_t sse;                                                              \
+    int se = 0;                                                                \
+    int64_t var;                                                               \
+    uint64_t long_sse = 0;                                                     \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    for (start_row = 0; start_row < h; start_row += 16) {                      \
+      uint32_t sse2;                                                           \
+      int height = h - start_row < 16 ? h - start_row : 16;                    \
+      int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+          src + (start_row * src_stride), src_stride, x_offset, y_offset,      \
+          dst + (start_row * dst_stride), dst_stride, height, &sse2, NULL,     \
+          NULL);                                                               \
+      se += se2;                                                               \
+      long_sse += sse2;                                                        \
+      if (w > wf) {                                                            \
+        se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                     \
+            src + 16 + (start_row * src_stride), src_stride, x_offset,         \
+            y_offset, dst + 16 + (start_row * dst_stride), dst_stride, height, \
+            &sse2, NULL, NULL);                                                \
+        se += se2;                                                             \
+        long_sse += sse2;                                                      \
+        if (w > wf * 2) {                                                      \
+          se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+              src + 32 + (start_row * src_stride), src_stride, x_offset,       \
+              y_offset, dst + 32 + (start_row * dst_stride), dst_stride,       \
+              height, &sse2, NULL, NULL);                                      \
+          se += se2;                                                           \
+          long_sse += sse2;                                                    \
+          se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+              src + 48 + (start_row * src_stride), src_stride, x_offset,       \
+              y_offset, dst + 48 + (start_row * dst_stride), dst_stride,       \
+              height, &sse2, NULL, NULL);                                      \
+          se += se2;                                                           \
+          long_sse += sse2;                                                    \
+        }                                                                      \
+      }                                                                        \
+    }                                                                          \
+    se = ROUND_POWER_OF_TWO(se, 4);                                            \
+    sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8);                           \
+    *sse_ptr = sse;                                                            \
+    var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2));                \
+    return (var >= 0) ? (uint32_t)var : 0;                                     \
+  }
+
+#define FNS(opt)                        \
+  FN(64, 64, 16, 6, 6, opt, (int64_t)); \
+  FN(64, 32, 16, 6, 5, opt, (int64_t)); \
+  FN(32, 64, 16, 5, 6, opt, (int64_t)); \
+  FN(32, 32, 16, 5, 5, opt, (int64_t)); \
+  FN(32, 16, 16, 5, 4, opt, (int64_t)); \
+  FN(16, 32, 16, 4, 5, opt, (int64_t)); \
+  FN(16, 16, 16, 4, 4, opt, (int64_t)); \
+  FN(16, 8, 16, 4, 3, opt, (int64_t));  \
+  FN(8, 16, 8, 3, 4, opt, (int64_t));   \
+  FN(8, 8, 8, 3, 3, opt, (int64_t));    \
+  FN(8, 4, 8, 3, 2, opt, (int64_t));
+
+FNS(sse2);
+
+#undef FNS
+#undef FN
+
+// The 2 unused parameters are place holders for PIC enabled build.
+#define DECL(w, opt)                                                         \
+  int aom_highbd_sub_pixel_avg_variance##w##xh_##opt(                        \
+      const uint16_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
+      const uint16_t *dst, ptrdiff_t dst_stride, const uint16_t *sec,        \
+      ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0,    \
+      void *unused);
+#define DECLS(opt1) \
+  DECL(16, opt1)    \
+  DECL(8, opt1)
+
+DECLS(sse2);
+#undef DECL
+#undef DECLS
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast)                                  \
+  uint32_t aom_highbd_8_sub_pixel_avg_variance##w##x##h##_##opt(               \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr,                  \
+      const uint8_t *sec8) {                                                   \
+    uint32_t sse;                                                              \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    uint16_t *sec = CONVERT_TO_SHORTPTR(sec8);                                 \
+    int se = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                  \
+        src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \
+        NULL, NULL);                                                           \
+    if (w > wf) {                                                              \
+      uint32_t sse2;                                                           \
+      int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(               \
+          src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride,      \
+          sec + 16, w, h, &sse2, NULL, NULL);                                  \
+      se += se2;                                                               \
+      sse += sse2;                                                             \
+      if (w > wf * 2) {                                                        \
+        se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                 \
+            src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride,    \
+            sec + 32, w, h, &sse2, NULL, NULL);                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+        se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                 \
+            src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride,    \
+            sec + 48, w, h, &sse2, NULL, NULL);                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+      }                                                                        \
+    }                                                                          \
+    *sse_ptr = sse;                                                            \
+    return sse - (uint32_t)((cast se * se) >> (wlog2 + hlog2));                \
+  }                                                                            \
+                                                                               \
+  uint32_t aom_highbd_10_sub_pixel_avg_variance##w##x##h##_##opt(              \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr,                  \
+      const uint8_t *sec8) {                                                   \
+    int64_t var;                                                               \
+    uint32_t sse;                                                              \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    uint16_t *sec = CONVERT_TO_SHORTPTR(sec8);                                 \
+    int se = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                  \
+        src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \
+        NULL, NULL);                                                           \
+    if (w > wf) {                                                              \
+      uint32_t sse2;                                                           \
+      int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(               \
+          src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride,      \
+          sec + 16, w, h, &sse2, NULL, NULL);                                  \
+      se += se2;                                                               \
+      sse += sse2;                                                             \
+      if (w > wf * 2) {                                                        \
+        se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                 \
+            src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride,    \
+            sec + 32, w, h, &sse2, NULL, NULL);                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+        se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                 \
+            src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride,    \
+            sec + 48, w, h, &sse2, NULL, NULL);                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+      }                                                                        \
+    }                                                                          \
+    se = ROUND_POWER_OF_TWO(se, 2);                                            \
+    sse = ROUND_POWER_OF_TWO(sse, 4);                                          \
+    *sse_ptr = sse;                                                            \
+    var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2));                \
+    return (var >= 0) ? (uint32_t)var : 0;                                     \
+  }                                                                            \
+                                                                               \
+  uint32_t aom_highbd_12_sub_pixel_avg_variance##w##x##h##_##opt(              \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr,                  \
+      const uint8_t *sec8) {                                                   \
+    int start_row;                                                             \
+    int64_t var;                                                               \
+    uint32_t sse;                                                              \
+    int se = 0;                                                                \
+    uint64_t long_sse = 0;                                                     \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    uint16_t *sec = CONVERT_TO_SHORTPTR(sec8);                                 \
+    for (start_row = 0; start_row < h; start_row += 16) {                      \
+      uint32_t sse2;                                                           \
+      int height = h - start_row < 16 ? h - start_row : 16;                    \
+      int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(               \
+          src + (start_row * src_stride), src_stride, x_offset, y_offset,      \
+          dst + (start_row * dst_stride), dst_stride, sec + (start_row * w),   \
+          w, height, &sse2, NULL, NULL);                                       \
+      se += se2;                                                               \
+      long_sse += sse2;                                                        \
+      if (w > wf) {                                                            \
+        se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                 \
+            src + 16 + (start_row * src_stride), src_stride, x_offset,         \
+            y_offset, dst + 16 + (start_row * dst_stride), dst_stride,         \
+            sec + 16 + (start_row * w), w, height, &sse2, NULL, NULL);         \
+        se += se2;                                                             \
+        long_sse += sse2;                                                      \
+        if (w > wf * 2) {                                                      \
+          se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(               \
+              src + 32 + (start_row * src_stride), src_stride, x_offset,       \
+              y_offset, dst + 32 + (start_row * dst_stride), dst_stride,       \
+              sec + 32 + (start_row * w), w, height, &sse2, NULL, NULL);       \
+          se += se2;                                                           \
+          long_sse += sse2;                                                    \
+          se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(               \
+              src + 48 + (start_row * src_stride), src_stride, x_offset,       \
+              y_offset, dst + 48 + (start_row * dst_stride), dst_stride,       \
+              sec + 48 + (start_row * w), w, height, &sse2, NULL, NULL);       \
+          se += se2;                                                           \
+          long_sse += sse2;                                                    \
+        }                                                                      \
+      }                                                                        \
+    }                                                                          \
+    se = ROUND_POWER_OF_TWO(se, 4);                                            \
+    sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8);                           \
+    *sse_ptr = sse;                                                            \
+    var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2));                \
+    return (var >= 0) ? (uint32_t)var : 0;                                     \
+  }
+
+#define FNS(opt1)                        \
+  FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
+  FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
+  FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
+  FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
+  FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
+  FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
+  FN(16, 16, 16, 4, 4, opt1, (int64_t)); \
+  FN(16, 8, 16, 4, 3, opt1, (int64_t));  \
+  FN(8, 16, 8, 4, 3, opt1, (int64_t));   \
+  FN(8, 8, 8, 3, 3, opt1, (int64_t));    \
+  FN(8, 4, 8, 3, 2, opt1, (int64_t));
+
+FNS(sse2);
+
+#undef FNS
+#undef FN
+
+void aom_highbd_upsampled_pred_sse2(uint16_t *comp_pred, int width, int height,
+                                    const uint8_t *ref8, int ref_stride) {
+  int i, j;
+  int stride = ref_stride << 3;
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+
+  if (width >= 8) {
+    // read 8 points at one time
+    for (i = 0; i < height; i++) {
+      for (j = 0; j < width; j += 8) {
+        __m128i s0 = _mm_cvtsi32_si128(*(const uint32_t *)ref);
+        __m128i s1 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 8));
+        __m128i s2 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 16));
+        __m128i s3 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 24));
+        __m128i s4 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 32));
+        __m128i s5 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 40));
+        __m128i s6 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 48));
+        __m128i s7 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 56));
+        __m128i t0, t1, t2, t3;
+
+        t0 = _mm_unpacklo_epi16(s0, s1);
+        t1 = _mm_unpacklo_epi16(s2, s3);
+        t2 = _mm_unpacklo_epi16(s4, s5);
+        t3 = _mm_unpacklo_epi16(s6, s7);
+        t0 = _mm_unpacklo_epi32(t0, t1);
+        t2 = _mm_unpacklo_epi32(t2, t3);
+        t0 = _mm_unpacklo_epi64(t0, t2);
+
+        _mm_storeu_si128((__m128i *)(comp_pred), t0);
+        comp_pred += 8;
+        ref += 64;  // 8 * 8;
+      }
+      ref += stride - (width << 3);
+    }
+  } else {
+    // read 4 points at one time
+    for (i = 0; i < height; i++) {
+      for (j = 0; j < width; j += 4) {
+        __m128i s0 = _mm_cvtsi32_si128(*(const uint32_t *)ref);
+        __m128i s1 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 8));
+        __m128i s2 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 16));
+        __m128i s3 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 24));
+        __m128i t0, t1;
+
+        t0 = _mm_unpacklo_epi16(s0, s1);
+        t1 = _mm_unpacklo_epi16(s2, s3);
+        t0 = _mm_unpacklo_epi32(t0, t1);
+
+        _mm_storel_epi64((__m128i *)(comp_pred), t0);
+        comp_pred += 4;
+        ref += 4 * 8;
+      }
+      ref += stride - (width << 3);
+    }
+  }
+}
+
+void aom_highbd_comp_avg_upsampled_pred_sse2(uint16_t *comp_pred,
+                                             const uint8_t *pred8, int width,
+                                             int height, const uint8_t *ref8,
+                                             int ref_stride) {
+  const __m128i one = _mm_set1_epi16(1);
+  int i, j;
+  int stride = ref_stride << 3;
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+
+  if (width >= 8) {
+    // read 8 points at one time
+    for (i = 0; i < height; i++) {
+      for (j = 0; j < width; j += 8) {
+        __m128i s0 = _mm_cvtsi32_si128(*(const uint32_t *)ref);
+        __m128i s1 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 8));
+        __m128i s2 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 16));
+        __m128i s3 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 24));
+        __m128i s4 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 32));
+        __m128i s5 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 40));
+        __m128i s6 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 48));
+        __m128i s7 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 56));
+        __m128i p0 = _mm_loadu_si128((const __m128i *)pred);
+        __m128i t0, t1, t2, t3;
+
+        t0 = _mm_unpacklo_epi16(s0, s1);
+        t1 = _mm_unpacklo_epi16(s2, s3);
+        t2 = _mm_unpacklo_epi16(s4, s5);
+        t3 = _mm_unpacklo_epi16(s6, s7);
+        t0 = _mm_unpacklo_epi32(t0, t1);
+        t2 = _mm_unpacklo_epi32(t2, t3);
+        t0 = _mm_unpacklo_epi64(t0, t2);
+
+        p0 = _mm_adds_epu16(t0, p0);
+        p0 = _mm_adds_epu16(p0, one);
+        p0 = _mm_srli_epi16(p0, 1);
+
+        _mm_storeu_si128((__m128i *)(comp_pred), p0);
+        comp_pred += 8;
+        pred += 8;
+        ref += 8 * 8;
+      }
+      ref += stride - (width << 3);
+    }
+  } else {
+    // read 4 points at one time
+    for (i = 0; i < height; i++) {
+      for (j = 0; j < width; j += 4) {
+        __m128i s0 = _mm_cvtsi32_si128(*(const uint32_t *)ref);
+        __m128i s1 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 8));
+        __m128i s2 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 16));
+        __m128i s3 = _mm_cvtsi32_si128(*(const uint32_t *)(ref + 24));
+        __m128i p0 = _mm_loadl_epi64((const __m128i *)pred);
+        __m128i t0, t1;
+
+        t0 = _mm_unpacklo_epi16(s0, s1);
+        t1 = _mm_unpacklo_epi16(s2, s3);
+        t0 = _mm_unpacklo_epi32(t0, t1);
+
+        p0 = _mm_adds_epu16(t0, p0);
+        p0 = _mm_adds_epu16(p0, one);
+        p0 = _mm_srli_epi16(p0, 1);
+
+        _mm_storel_epi64((__m128i *)(comp_pred), p0);
+        comp_pred += 4;
+        pred += 4;
+        ref += 4 * 8;
+      }
+      ref += stride - (width << 3);
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c b/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c
new file mode 100644
index 0000000000..cc7f528113
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c
@@ -0,0 +1,216 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <smmintrin.h> /* SSE4.1 */
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom_dsp/variance.h"
+#include "aom_dsp/aom_filter.h"
+
+static INLINE void variance4x4_64_sse4_1(const uint8_t *a8, int a_stride,
+                                         const uint8_t *b8, int b_stride,
+                                         uint64_t *sse, int64_t *sum) {
+  __m128i u0, u1, u2, u3;
+  __m128i s0, s1, s2, s3;
+  __m128i t0, t1, x0, y0;
+  __m128i a0, a1, a2, a3;
+  __m128i b0, b1, b2, b3;
+  __m128i k_one_epi16 = _mm_set1_epi16((int16_t)1);
+
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+
+  a0 = _mm_loadl_epi64((__m128i const *)(a + 0 * a_stride));
+  a1 = _mm_loadl_epi64((__m128i const *)(a + 1 * a_stride));
+  a2 = _mm_loadl_epi64((__m128i const *)(a + 2 * a_stride));
+  a3 = _mm_loadl_epi64((__m128i const *)(a + 3 * a_stride));
+
+  b0 = _mm_loadl_epi64((__m128i const *)(b + 0 * b_stride));
+  b1 = _mm_loadl_epi64((__m128i const *)(b + 1 * b_stride));
+  b2 = _mm_loadl_epi64((__m128i const *)(b + 2 * b_stride));
+  b3 = _mm_loadl_epi64((__m128i const *)(b + 3 * b_stride));
+
+  u0 = _mm_unpacklo_epi16(a0, a1);
+  u1 = _mm_unpacklo_epi16(a2, a3);
+  u2 = _mm_unpacklo_epi16(b0, b1);
+  u3 = _mm_unpacklo_epi16(b2, b3);
+
+  s0 = _mm_sub_epi16(u0, u2);
+  s1 = _mm_sub_epi16(u1, u3);
+
+  t0 = _mm_madd_epi16(s0, k_one_epi16);
+  t1 = _mm_madd_epi16(s1, k_one_epi16);
+
+  s2 = _mm_hadd_epi32(t0, t1);
+  s3 = _mm_hadd_epi32(s2, s2);
+  y0 = _mm_hadd_epi32(s3, s3);
+
+  t0 = _mm_madd_epi16(s0, s0);
+  t1 = _mm_madd_epi16(s1, s1);
+
+  s2 = _mm_hadd_epi32(t0, t1);
+  s3 = _mm_hadd_epi32(s2, s2);
+  x0 = _mm_hadd_epi32(s3, s3);
+
+  *sse = (uint64_t)_mm_extract_epi32(x0, 0);
+  *sum = (int64_t)_mm_extract_epi32(y0, 0);
+}
+
+uint32_t aom_highbd_8_variance4x4_sse4_1(const uint8_t *a, int a_stride,
+                                         const uint8_t *b, int b_stride,
+                                         uint32_t *sse) {
+  int64_t sum, diff;
+  uint64_t local_sse;
+
+  variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum);
+  *sse = (uint32_t)local_sse;
+
+  diff = (int64_t)*sse - ((sum * sum) >> 4);
+  return (diff >= 0) ? (uint32_t)diff : 0;
+}
+
+uint32_t aom_highbd_10_variance4x4_sse4_1(const uint8_t *a, int a_stride,
+                                          const uint8_t *b, int b_stride,
+                                          uint32_t *sse) {
+  int64_t sum, diff;
+  uint64_t local_sse;
+
+  variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(local_sse, 4);
+  sum = ROUND_POWER_OF_TWO(sum, 2);
+
+  diff = (int64_t)*sse - ((sum * sum) >> 4);
+  return (diff >= 0) ? (uint32_t)diff : 0;
+}
+
+uint32_t aom_highbd_12_variance4x4_sse4_1(const uint8_t *a, int a_stride,
+                                          const uint8_t *b, int b_stride,
+                                          uint32_t *sse) {
+  int64_t sum, diff;
+  uint64_t local_sse;
+
+  variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(local_sse, 8);
+  sum = ROUND_POWER_OF_TWO(sum, 4);
+
+  diff = (int64_t)*sse - ((sum * sum) >> 4);
+  return diff >= 0 ? (uint32_t)diff : 0;
+}
+
+// Sub-pixel
+uint32_t aom_highbd_8_sub_pixel_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  return aom_highbd_8_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst, dst_stride,
+                                  sse);
+}
+
+uint32_t aom_highbd_10_sub_pixel_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  return aom_highbd_10_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst,
+                                   dst_stride, sse);
+}
+
+uint32_t aom_highbd_12_sub_pixel_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  return aom_highbd_12_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst,
+                                   dst_stride, sse);
+}
+
+// Sub-pixel average
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse,
+    const uint8_t *second_pred) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+  DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]);
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  aom_highbd_comp_avg_pred(temp3, second_pred, 4, 4, CONVERT_TO_BYTEPTR(temp2),
+                           4);
+
+  return aom_highbd_8_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst, dst_stride,
+                                  sse);
+}
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse,
+    const uint8_t *second_pred) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+  DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]);
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  aom_highbd_comp_avg_pred(temp3, second_pred, 4, 4, CONVERT_TO_BYTEPTR(temp2),
+                           4);
+
+  return aom_highbd_10_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst,
+                                   dst_stride, sse);
+}
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse,
+    const uint8_t *second_pred) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+  DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]);
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  aom_highbd_comp_avg_pred(temp3, second_pred, 4, 4, CONVERT_TO_BYTEPTR(temp2),
+                           4);
+
+  return aom_highbd_12_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst,
+                                   dst_stride, sse);
+}
diff --git a/third_party/aom/aom_dsp/x86/intrapred_sse2.asm b/third_party/aom/aom_dsp/x86/intrapred_sse2.asm
new file mode 100644
index 0000000000..02567db492
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/intrapred_sse2.asm
@@ -0,0 +1,771 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pb_1: times 16 db 1
+pw_4:  times 8 dw 4
+pw_8:  times 8 dw 8
+pw_16: times 8 dw 16
+pw_32: times 8 dw 32
+dc_128: times 16 db 128
+pw2_4:  times 8 dw 2
+pw2_8:  times 8 dw 4
+pw2_16:  times 8 dw 8
+pw2_32:  times 8 dw 16
+
+SECTION .text
+
+; ------------------------------------------
+; input: x, y, z, result
+;
+; trick from pascal
+; (x+2y+z+2)>>2 can be calculated as:
+; result = avg(x,z)
+; result -= xor(x,z) & 1
+; result = avg(result,y)
+; ------------------------------------------
+%macro X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 4
+  pavgb               %4, %1, %3
+  pxor                %3, %1
+  pand                %3, [GLOBAL(pb_1)]
+  psubb               %4, %3
+  pavgb               %4, %2
+%endmacro
+
+INIT_XMM sse2
+cglobal dc_predictor_4x4, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  movd                  m2, [leftq]
+  movd                  m0, [aboveq]
+  pxor                  m1, m1
+  punpckldq             m0, m2
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw_4)]
+  psraw                 m0, 3
+  pshuflw               m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_4x4, 2, 5, 2, dst, stride, above, left, goffset
+  movifnidn          leftq, leftmp
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movd                  m0, [leftq]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_4)]
+  psraw                 m0, 2
+  pshuflw               m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_top_predictor_4x4, 3, 5, 2, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movd                  m0, [aboveq]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_4)]
+  psraw                 m0, 2
+  pshuflw               m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [aboveq]
+  movq                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_8)]
+  psraw                 m0, 4
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_top_predictor_8x8, 3, 5, 2, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_8)]
+  psraw                 m0, 3
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_8x8, 2, 5, 2, dst, stride, above, left, goffset
+  movifnidn          leftq, leftmp
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [leftq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_8)]
+  psraw                 m0, 3
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_4x4, 2, 5, 1, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movd     m0,        [GLOBAL(dc_128)]
+  movd    [dstq          ], m0
+  movd    [dstq+strideq  ], m0
+  movd    [dstq+strideq*2], m0
+  movd    [dstq+stride3q ], m0
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_8x8, 2, 5, 1, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movq    m0,        [GLOBAL(dc_128)]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_16)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+
+INIT_XMM sse2
+cglobal dc_top_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_16)]
+  psraw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [leftq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_16)]
+  psraw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  mova    m0,        [GLOBAL(dc_128)]
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+  RESTORE_GOT
+  RET
+
+
+INIT_XMM sse2
+cglobal dc_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [leftq]
+  mova                  m4, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  psadbw                m3, m1
+  psadbw                m4, m1
+  paddw                 m0, m2
+  paddw                 m0, m3
+  paddw                 m0, m4
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_32)]
+  psraw                 m0, 6
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_top_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_32)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [leftq]
+  mova                  m2, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_32)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_32x32, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  mova    m0,        [GLOBAL(dc_128)]
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal v_predictor_4x4, 3, 3, 1, dst, stride, above
+  movd                  m0, [aboveq]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  RET
+
+INIT_XMM sse2
+cglobal v_predictor_8x8, 3, 3, 1, dst, stride, above
+  movq                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  RET
+
+INIT_XMM sse2
+cglobal v_predictor_16x16, 3, 4, 1, dst, stride, above
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 4
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal v_predictor_32x32, 3, 4, 2, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 8
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m1
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m1
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m1
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m1
+  lea                 dstq, [dstq+strideq*4]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal h_predictor_4x4, 2, 4, 4, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  movd                  m0, [leftq]
+  punpcklbw             m0, m0
+  punpcklbw             m0, m0
+  pshufd                m1, m0, 0x1
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m1
+  pshufd                m2, m0, 0x2
+  lea                 dstq, [dstq+strideq*2]
+  pshufd                m3, m0, 0x3
+  movd      [dstq        ], m2
+  movd      [dstq+strideq], m3
+  RET
+
+INIT_XMM sse2
+cglobal h_predictor_8x8, 2, 5, 3, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  mov                lineq, -2
+  DEFINE_ARGS  dst, stride, line, left, stride3
+  lea             stride3q, [strideq*3]
+  movq                  m0, [leftq    ]
+  punpcklbw             m0, m0              ; l1 l1 l2 l2 ... l8 l8
+.loop:
+  pshuflw               m1, m0, 0x0         ; l1 l1 l1 l1 l1 l1 l1 l1
+  pshuflw               m2, m0, 0x55        ; l2 l2 l2 l2 l2 l2 l2 l2
+  movq      [dstq        ], m1
+  movq      [dstq+strideq], m2
+  pshuflw               m1, m0, 0xaa
+  pshuflw               m2, m0, 0xff
+  movq    [dstq+strideq*2], m1
+  movq    [dstq+stride3q ], m2
+  pshufd                m0, m0, 0xe         ; [63:0] l5 l5 l6 l6 l7 l7 l8 l8
+  inc                lineq
+  lea                 dstq, [dstq+strideq*4]
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal h_predictor_16x16, 2, 5, 3, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  mov                lineq, -4
+  DEFINE_ARGS dst, stride, line, left, stride3
+  lea             stride3q, [strideq*3]
+.loop:
+  movd                  m0, [leftq]
+  punpcklbw             m0, m0
+  punpcklbw             m0, m0              ; l1 to l4 each repeated 4 times
+  pshufd            m1, m0, 0x0             ; l1 repeated 16 times
+  pshufd            m2, m0, 0x55            ; l2 repeated 16 times
+  mova    [dstq          ], m1
+  mova    [dstq+strideq  ], m2
+  pshufd            m1, m0, 0xaa
+  pshufd            m2, m0, 0xff
+  mova    [dstq+strideq*2], m1
+  mova    [dstq+stride3q ], m2
+  inc                lineq
+  lea                leftq, [leftq+4       ]
+  lea                 dstq, [dstq+strideq*4]
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal h_predictor_32x32, 2, 5, 3, dst, stride, line, left
+  movifnidn              leftq, leftmp
+  mov                    lineq, -8
+  DEFINE_ARGS dst, stride, line, left, stride3
+  lea                 stride3q, [strideq*3]
+.loop:
+  movd                      m0, [leftq]
+  punpcklbw                 m0, m0
+  punpcklbw                 m0, m0              ; l1 to l4 each repeated 4 times
+  pshufd                m1, m0, 0x0             ; l1 repeated 16 times
+  pshufd                m2, m0, 0x55            ; l2 repeated 16 times
+  mova     [dstq             ], m1
+  mova     [dstq+16          ], m1
+  mova     [dstq+strideq     ], m2
+  mova     [dstq+strideq+16  ], m2
+  pshufd                m1, m0, 0xaa
+  pshufd                m2, m0, 0xff
+  mova     [dstq+strideq*2   ], m1
+  mova     [dstq+strideq*2+16], m1
+  mova     [dstq+stride3q    ], m2
+  mova     [dstq+stride3q+16 ], m2
+  inc                    lineq
+  lea                    leftq, [leftq+4       ]
+  lea                     dstq, [dstq+strideq*4]
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal tm_predictor_4x4, 4, 4, 5, dst, stride, above, left
+  pxor                  m1, m1
+  movq                  m0, [aboveq-1]; [63:0] tl t1 t2 t3 t4 x x x
+  punpcklbw             m0, m1
+  pshuflw               m2, m0, 0x0   ; [63:0] tl tl tl tl [word]
+  psrldq                m0, 2
+  psubw                 m0, m2        ; [63:0] t1-tl t2-tl t3-tl t4-tl [word]
+  movd                  m2, [leftq]
+  punpcklbw             m2, m1
+  pshuflw               m4, m2, 0x0   ; [63:0] l1 l1 l1 l1 [word]
+  pshuflw               m3, m2, 0x55  ; [63:0] l2 l2 l2 l2 [word]
+  paddw                 m4, m0
+  paddw                 m3, m0
+  packuswb              m4, m4
+  packuswb              m3, m3
+  movd      [dstq        ], m4
+  movd      [dstq+strideq], m3
+  lea                 dstq, [dstq+strideq*2]
+  pshuflw               m4, m2, 0xaa
+  pshuflw               m3, m2, 0xff
+  paddw                 m4, m0
+  paddw                 m3, m0
+  packuswb              m4, m4
+  packuswb              m3, m3
+  movd      [dstq        ], m4
+  movd      [dstq+strideq], m3
+  RET
+
+INIT_XMM sse2
+cglobal tm_predictor_8x8, 4, 4, 5, dst, stride, above, left
+  pxor                  m1, m1
+  movd                  m2, [aboveq-1]
+  movq                  m0, [aboveq]
+  punpcklbw             m2, m1
+  punpcklbw             m0, m1        ; t1 t2 t3 t4 t5 t6 t7 t8 [word]
+  pshuflw               m2, m2, 0x0   ; [63:0] tl tl tl tl [word]
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -4
+  punpcklqdq            m2, m2        ; tl tl tl tl tl tl tl tl [word]
+  psubw                 m0, m2        ; t1-tl t2-tl ... t8-tl [word]
+  movq                  m2, [leftq]
+  punpcklbw             m2, m1        ; l1 l2 l3 l4 l5 l6 l7 l8 [word]
+.loop:
+  pshuflw               m4, m2, 0x0   ; [63:0] l1 l1 l1 l1 [word]
+  pshuflw               m3, m2, 0x55  ; [63:0] l2 l2 l2 l2 [word]
+  punpcklqdq            m4, m4        ; l1 l1 l1 l1 l1 l1 l1 l1 [word]
+  punpcklqdq            m3, m3        ; l2 l2 l2 l2 l2 l2 l2 l2 [word]
+  paddw                 m4, m0
+  paddw                 m3, m0
+  packuswb              m4, m3
+  movq      [dstq        ], m4
+  movhps    [dstq+strideq], m4
+  lea                 dstq, [dstq+strideq*2]
+  psrldq                m2, 4
+  inc                lineq
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal tm_predictor_16x16, 4, 5, 8, dst, stride, above, left
+  pxor                  m1, m1
+  mova                  m2, [aboveq-16];
+  mova                  m0, [aboveq]   ; t1 t2 ... t16 [byte]
+  punpckhbw             m2, m1         ; [127:112] tl [word]
+  punpckhbw             m4, m0, m1
+  punpcklbw             m0, m1         ; m0:m4 t1 t2 ... t16 [word]
+  DEFINE_ARGS dst, stride, line, left, stride8
+  mov                lineq, -8
+  pshufhw               m2, m2, 0xff
+  mova                  m3, [leftq]    ; l1 l2 ... l16 [byte]
+  punpckhqdq            m2, m2         ; tl repeated 8 times [word]
+  psubw                 m0, m2
+  psubw                 m4, m2         ; m0:m4 t1-tl t2-tl ... t16-tl [word]
+  punpckhbw             m5, m3, m1
+  punpcklbw             m3, m1         ; m3:m5 l1 l2 ... l16 [word]
+  lea             stride8q, [strideq*8]
+.loop:
+  pshuflw               m6, m3, 0x0
+  pshuflw               m7, m5, 0x0
+  punpcklqdq            m6, m6         ; l1 repeated 8 times [word]
+  punpcklqdq            m7, m7         ; l8 repeated 8 times [word]
+  paddw                 m1, m6, m0
+  paddw                 m6, m4         ; m1:m6 ti-tl+l1 [i=1,15] [word]
+  psrldq                m5, 2
+  packuswb              m1, m6
+  mova     [dstq         ], m1
+  paddw                 m1, m7, m0
+  paddw                 m7, m4         ; m1:m7 ti-tl+l8 [i=1,15] [word]
+  psrldq                m3, 2
+  packuswb              m1, m7
+  mova     [dstq+stride8q], m1
+  inc                lineq
+  lea                 dstq, [dstq+strideq]
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal tm_predictor_32x32, 4, 4, 8, dst, stride, above, left
+  pxor                  m1, m1
+  movd                  m2, [aboveq-1]
+  mova                  m0, [aboveq]
+  mova                  m4, [aboveq+16]
+  punpcklbw             m2, m1
+  punpckhbw             m3, m0, m1
+  punpckhbw             m5, m4, m1
+  punpcklbw             m0, m1
+  punpcklbw             m4, m1
+  pshuflw               m2, m2, 0x0
+  DEFINE_ARGS dst, stride, line, left
+  mov                lineq, -16
+  punpcklqdq            m2, m2
+  add                leftq, 32
+  psubw                 m0, m2
+  psubw                 m3, m2
+  psubw                 m4, m2
+  psubw                 m5, m2
+.loop:
+  movd                  m2, [leftq+lineq*2]
+  pxor                  m1, m1
+  punpcklbw             m2, m1
+  pshuflw               m7, m2, 0x55
+  pshuflw               m2, m2, 0x0
+  punpcklqdq            m2, m2
+  punpcklqdq            m7, m7
+  paddw                 m6, m2, m3
+  paddw                 m1, m2, m0
+  packuswb              m1, m6
+  mova   [dstq           ], m1
+  paddw                 m6, m2, m5
+  paddw                 m1, m2, m4
+  packuswb              m1, m6
+  mova   [dstq+16        ], m1
+  paddw                 m6, m7, m3
+  paddw                 m1, m7, m0
+  packuswb              m1, m6
+  mova   [dstq+strideq   ], m1
+  paddw                 m6, m7, m5
+  paddw                 m1, m7, m4
+  packuswb              m1, m6
+  mova   [dstq+strideq+16], m1
+  lea                 dstq, [dstq+strideq*2]
+  inc                lineq
+  jnz .loop
+  REP_RET
diff --git a/third_party/aom/aom_dsp/x86/intrapred_ssse3.asm b/third_party/aom/aom_dsp/x86/intrapred_ssse3.asm
new file mode 100644
index 0000000000..bc1bb2ff3d
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/intrapred_ssse3.asm
@@ -0,0 +1,410 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+
+pb_1: times 16 db 1
+sh_b12345677: db 1, 2, 3, 4, 5, 6, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b23456777: db 2, 3, 4, 5, 6, 7, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b0123456777777777: db 0, 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7
+sh_b1234567777777777: db 1, 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7
+sh_b2345677777777777: db 2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7
+sh_b123456789abcdeff: db 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 15
+sh_b23456789abcdefff: db 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 15, 15
+sh_b32104567: db 3, 2, 1, 0, 4, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b8091a2b345: db 8, 0, 9, 1, 10, 2, 11, 3, 4, 5, 0, 0, 0, 0, 0, 0
+sh_b76543210: db 7, 6, 5, 4, 3, 2, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b65432108: db 6, 5, 4, 3, 2, 1, 0, 8, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b54321089: db 5, 4, 3, 2, 1, 0, 8, 9, 0, 0, 0, 0, 0, 0, 0, 0
+sh_b89abcdef: db 8, 9, 10, 11, 12, 13, 14, 15, 0, 0, 0, 0, 0, 0, 0, 0
+sh_bfedcba9876543210: db 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0
+
+SECTION .text
+
+; ------------------------------------------
+; input: x, y, z, result
+;
+; trick from pascal
+; (x+2y+z+2)>>2 can be calculated as:
+; result = avg(x,z)
+; result -= xor(x,z) & 1
+; result = avg(result,y)
+; ------------------------------------------
+%macro X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 4
+  pavgb               %4, %1, %3
+  pxor                %3, %1
+  pand                %3, [GLOBAL(pb_1)]
+  psubb               %4, %3
+  pavgb               %4, %2
+%endmacro
+
+INIT_XMM ssse3
+cglobal d63e_predictor_4x4, 3, 4, 5, dst, stride, above, goffset
+  GET_GOT     goffsetq
+
+  movq                m3, [aboveq]
+  pshufb              m1, m3, [GLOBAL(sh_b23456777)]
+  pshufb              m2, m3, [GLOBAL(sh_b12345677)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m3, m2, m1, m4
+  pavgb               m3, m2
+
+  ; store 4 lines
+  movd    [dstq        ], m3
+  movd    [dstq+strideq], m4
+  lea               dstq, [dstq+strideq*2]
+  psrldq              m3, 1
+  psrldq              m4, 1
+  movd    [dstq        ], m3
+  movd    [dstq+strideq], m4
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_4x4, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  movd                m0, [leftq]               ; l1, l2, l3, l4
+  movd                m1, [aboveq-1]            ; tl, t1, t2, t3
+  punpckldq           m0, m1                    ; l1, l2, l3, l4, tl, t1, t2, t3
+  pshufb              m0, [GLOBAL(sh_b32104567)]; l4, l3, l2, l1, tl, t1, t2, t3
+  psrldq              m1, m0, 1                 ; l3, l2, l1, tl, t1, t2, t3
+  psrldq              m2, m0, 2                 ; l2, l1, tl, t1, t2, t3
+  ; comments below are for a predictor like this
+  ; A1 B1 C1 D1
+  ; A2 B2 A1 B1
+  ; A3 B3 A2 B2
+  ; A4 B4 A3 B3
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m1, m2, m3  ; 3-tap avg B4 B3 B2 B1 C1 D1
+  pavgb               m1, m0                    ; 2-tap avg A4 A3 A2 A1
+
+  punpcklqdq          m3, m1                    ; B4 B3 B2 B1 C1 D1 x x A4 A3 A2 A1 ..
+
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+  pshufb              m3, [GLOBAL(sh_b8091a2b345)] ; A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 ..
+  movd  [dstq+stride3q ], m3
+  psrldq              m3, 2                     ; A3 B3 A2 B2 A1 B1 C1 D1 ..
+  movd  [dstq+strideq*2], m3
+  psrldq              m3, 2                     ; A2 B2 A1 B1 C1 D1 ..
+  movd  [dstq+strideq  ], m3
+  psrldq              m3, 2                     ; A1 B1 C1 D1 ..
+  movd  [dstq          ], m3
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_8x8, 4, 5, 8, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  movq                m0, [leftq]                     ; [0- 7] l1-8 [byte]
+  movhps              m0, [aboveq-1]                  ; [8-15] tl, t1-7 [byte]
+  pshufb              m1, m0, [GLOBAL(sh_b76543210)]  ; l8-1 [word]
+  pshufb              m2, m0, [GLOBAL(sh_b65432108)]  ; l7-1,tl [word]
+  pshufb              m3, m0, [GLOBAL(sh_b54321089)]  ; l6-1,tl,t1 [word]
+  pshufb              m0, [GLOBAL(sh_b89abcdef)]      ; tl,t1-7 [word]
+  psrldq              m4, m0, 1                       ; t1-7 [word]
+  psrldq              m5, m0, 2                       ; t2-7 [word]
+  ; comments below are for a predictor like this
+  ; A1 B1 C1 D1 E1 F1 G1 H1
+  ; A2 B2 A1 B1 C1 D1 E1 F1
+  ; A3 B3 A2 B2 A1 B1 C1 D1
+  ; A4 B4 A3 B3 A2 B2 A1 B1
+  ; A5 B5 A4 B4 A3 B3 A2 B2
+  ; A6 B6 A5 B5 A4 B4 A3 B3
+  ; A7 B7 A6 B6 A5 B5 A4 B4
+  ; A8 B8 A7 B7 A6 B6 A5 B5
+  pavgb               m6, m1, m2                ; 2-tap avg A8-A1
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m4, m5, m7  ; 3-tap avg C-H1
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m1, m2, m3, m0  ; 3-tap avg B8-1
+
+  punpcklbw           m6, m0                    ; A-B8, A-B7 ... A-B2, A-B1
+
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+
+  movhps [dstq+stride3q], m6                    ; A-B4, A-B3, A-B2, A-B1
+  palignr             m0, m7, m6, 10            ; A-B3, A-B2, A-B1, C-H1
+  movq  [dstq+strideq*2], m0
+  psrldq              m0, 2                     ; A-B2, A-B1, C-H1
+  movq  [dstq+strideq  ], m0
+  psrldq              m0, 2                     ; A-H1
+  movq  [dstq          ], m0
+  lea               dstq, [dstq+strideq*4]
+  movq  [dstq+stride3q ], m6                    ; A-B8, A-B7, A-B6, A-B5
+  psrldq              m6, 2                     ; A-B7, A-B6, A-B5, A-B4
+  movq  [dstq+strideq*2], m6
+  psrldq              m6, 2                     ; A-B6, A-B5, A-B4, A-B3
+  movq  [dstq+strideq  ], m6
+  psrldq              m6, 2                     ; A-B5, A-B4, A-B3, A-B2
+  movq  [dstq          ], m6
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_16x16, 4, 5, 8, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  mova                m0, [leftq]
+  movu                m7, [aboveq-1]
+  ; comments below are for a predictor like this
+  ; A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1 M1 N1 O1 P1
+  ; A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1 M1 N1
+  ; A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1 K1 L1
+  ; A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1 I1 J1
+  ; A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1 G1 H1
+  ; A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1 E1 F1
+  ; A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1 C1 D1
+  ; A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2 A1 B1
+  ; A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3 A2 B2
+  ; Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4 A3 B3
+  ; Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5 A4 B4
+  ; Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6 A5 B5
+  ; Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7 A6 B6
+  ; Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8 A7 B7
+  ; Af Bf Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9 A8 B8
+  ; Ag Bg Af Bf Ae Be Ad Bd Ac Bc Ab Bb Aa Ba A9 B9
+  pshufb              m6, m7, [GLOBAL(sh_bfedcba9876543210)]
+  palignr             m5, m0, m6, 15
+  palignr             m3, m0, m6, 14
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m5, m3, m4          ; 3-tap avg B3-Bg
+  pshufb              m1, m0, [GLOBAL(sh_b123456789abcdeff)]
+  pavgb               m5, m0                            ; A1 - Ag
+
+  punpcklbw           m0, m4, m5                        ; A-B8 ... A-B1
+  punpckhbw           m4, m5                            ; A-B9 ... A-Bg
+
+  pshufb              m3, m7, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb              m5, m7, [GLOBAL(sh_b23456789abcdefff)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m5, m1          ; 3-tap avg C1-P1
+
+  pshufb              m6, m0, [GLOBAL(sh_bfedcba9876543210)]
+  DEFINE_ARGS dst, stride, stride3
+  lea           stride3q, [strideq*3]
+  palignr             m2, m1, m6, 14
+  mova  [dstq          ], m2
+  palignr             m2, m1, m6, 12
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m1, m6, 10
+  mova  [dstq+strideq*2], m2
+  palignr             m2, m1, m6, 8
+  mova  [dstq+stride3q ], m2
+  lea               dstq, [dstq+strideq*4]
+  palignr             m2, m1, m6, 6
+  mova  [dstq          ], m2
+  palignr             m2, m1, m6, 4
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m1, m6, 2
+  mova  [dstq+strideq*2], m2
+  pshufb              m4, [GLOBAL(sh_bfedcba9876543210)]
+  mova  [dstq+stride3q ], m6
+  lea               dstq, [dstq+strideq*4]
+
+  palignr             m2, m6, m4, 14
+  mova  [dstq          ], m2
+  palignr             m2, m6, m4, 12
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m6, m4, 10
+  mova  [dstq+strideq*2], m2
+  palignr             m2, m6, m4, 8
+  mova  [dstq+stride3q ], m2
+  lea               dstq, [dstq+strideq*4]
+  palignr             m2, m6, m4, 6
+  mova  [dstq          ], m2
+  palignr             m2, m6, m4, 4
+  mova  [dstq+strideq  ], m2
+  palignr             m2, m6, m4, 2
+  mova  [dstq+strideq*2], m2
+  mova  [dstq+stride3q ], m4
+  RESTORE_GOT
+  RET
+
+INIT_XMM ssse3
+cglobal d153_predictor_32x32, 4, 5, 8, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+  mova                  m0, [leftq]
+  movu                  m7, [aboveq-1]
+  movu                  m1, [aboveq+15]
+
+  pshufb                m4, m1, [GLOBAL(sh_b123456789abcdeff)]
+  pshufb                m6, m1, [GLOBAL(sh_b23456789abcdefff)]
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m1, m4, m6, m2          ; 3-tap avg above [high]
+
+  palignr               m3, m1, m7, 1
+  palignr               m5, m1, m7, 2
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m7, m3, m5, m1          ; 3-tap avg above [low]
+
+  pshufb                m7, [GLOBAL(sh_bfedcba9876543210)]
+  palignr               m5, m0, m7, 15
+  palignr               m3, m0, m7, 14
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m0, m5, m3, m4          ; 3-tap avg B3-Bg
+  pavgb                 m5, m0                            ; A1 - Ag
+  punpcklbw             m6, m4, m5                        ; A-B8 ... A-B1
+  punpckhbw             m4, m5                            ; A-B9 ... A-Bg
+  pshufb                m6, [GLOBAL(sh_bfedcba9876543210)]
+  pshufb                m4, [GLOBAL(sh_bfedcba9876543210)]
+
+  DEFINE_ARGS dst, stride, stride3, left, line
+  lea             stride3q, [strideq*3]
+
+  palignr               m5, m2, m1, 14
+  palignr               m7, m1, m6, 14
+  mova  [dstq            ], m7
+  mova  [dstq+16         ], m5
+  palignr               m5, m2, m1, 12
+  palignr               m7, m1, m6, 12
+  mova  [dstq+strideq    ], m7
+  mova  [dstq+strideq+16 ], m5
+  palignr                m5, m2, m1, 10
+  palignr                m7, m1, m6, 10
+  mova  [dstq+strideq*2   ], m7
+  mova  [dstq+strideq*2+16], m5
+  palignr                m5, m2, m1, 8
+  palignr                m7, m1, m6, 8
+  mova  [dstq+stride3q    ], m7
+  mova  [dstq+stride3q+16 ], m5
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m5, m2, m1, 6
+  palignr                m7, m1, m6, 6
+  mova  [dstq             ], m7
+  mova  [dstq+16          ], m5
+  palignr                m5, m2, m1, 4
+  palignr                m7, m1, m6, 4
+  mova  [dstq+strideq     ], m7
+  mova  [dstq+strideq+16  ], m5
+  palignr                m5, m2, m1, 2
+  palignr                m7, m1, m6, 2
+  mova  [dstq+strideq*2   ], m7
+  mova  [dstq+strideq*2+16], m5
+  mova  [dstq+stride3q    ], m6
+  mova  [dstq+stride3q+16 ], m1
+  lea                  dstq, [dstq+strideq*4]
+
+  palignr                m5, m1, m6, 14
+  palignr                m3, m6, m4, 14
+  mova  [dstq             ], m3
+  mova  [dstq+16          ], m5
+  palignr                m5, m1, m6, 12
+  palignr                m3, m6, m4, 12
+  mova  [dstq+strideq     ], m3
+  mova  [dstq+strideq+16  ], m5
+  palignr                m5, m1, m6, 10
+  palignr                m3, m6, m4, 10
+  mova  [dstq+strideq*2   ], m3
+  mova  [dstq+strideq*2+16], m5
+  palignr                m5, m1, m6, 8
+  palignr                m3, m6, m4, 8
+  mova  [dstq+stride3q    ], m3
+  mova  [dstq+stride3q+16 ], m5
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m5, m1, m6, 6
+  palignr                m3, m6, m4, 6
+  mova  [dstq             ], m3
+  mova  [dstq+16          ], m5
+  palignr                m5, m1, m6, 4
+  palignr                m3, m6, m4, 4
+  mova  [dstq+strideq     ], m3
+  mova  [dstq+strideq+16  ], m5
+  palignr                m5, m1, m6, 2
+  palignr                m3, m6, m4, 2
+  mova  [dstq+strideq*2   ], m3
+  mova  [dstq+strideq*2+16], m5
+  mova  [dstq+stride3q    ], m4
+  mova  [dstq+stride3q+16 ], m6
+  lea               dstq, [dstq+strideq*4]
+
+  mova                   m7, [leftq]
+  mova                   m3, [leftq+16]
+  palignr                m5, m3, m7, 15
+  palignr                m0, m3, m7, 14
+
+  X_PLUS_2Y_PLUS_Z_PLUS_2_RSH_2 m3, m5, m0, m2          ; 3-tap avg Bh -
+  pavgb                  m5, m3                            ; Ah -
+  punpcklbw              m3, m2, m5                        ; A-B8 ... A-B1
+  punpckhbw              m2, m5                            ; A-B9 ... A-Bg
+  pshufb                 m3, [GLOBAL(sh_bfedcba9876543210)]
+  pshufb                 m2, [GLOBAL(sh_bfedcba9876543210)]
+
+  palignr                m7, m6, m4, 14
+  palignr                m0, m4, m3, 14
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m6, m4, 12
+  palignr                m0, m4, m3, 12
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m6, m4, 10
+  palignr                m0, m4, m3, 10
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  palignr                m7, m6, m4, 8
+  palignr                m0, m4, m3, 8
+  mova  [dstq+stride3q    ], m0
+  mova  [dstq+stride3q+16 ], m7
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m7, m6, m4, 6
+  palignr                m0, m4, m3, 6
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m6, m4, 4
+  palignr                m0, m4, m3, 4
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m6, m4, 2
+  palignr                m0, m4, m3, 2
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  mova  [dstq+stride3q    ], m3
+  mova  [dstq+stride3q+16 ], m4
+  lea                  dstq, [dstq+strideq*4]
+
+  palignr                m7, m4, m3, 14
+  palignr                m0, m3, m2, 14
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m4, m3, 12
+  palignr                m0, m3, m2, 12
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m4, m3, 10
+  palignr                m0, m3, m2, 10
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  palignr                m7, m4, m3, 8
+  palignr                m0, m3, m2, 8
+  mova  [dstq+stride3q    ], m0
+  mova  [dstq+stride3q+16 ], m7
+  lea                  dstq, [dstq+strideq*4]
+  palignr                m7, m4, m3, 6
+  palignr                m0, m3, m2, 6
+  mova  [dstq             ], m0
+  mova  [dstq+16          ], m7
+  palignr                m7, m4, m3, 4
+  palignr                m0, m3, m2, 4
+  mova  [dstq+strideq     ], m0
+  mova  [dstq+strideq+16  ], m7
+  palignr                m7, m4, m3, 2
+  palignr                m0, m3, m2, 2
+  mova  [dstq+strideq*2   ], m0
+  mova  [dstq+strideq*2+16], m7
+  mova  [dstq+stride3q    ], m2
+  mova  [dstq+stride3q+16 ], m3
+
+  RESTORE_GOT
+  RET
diff --git a/third_party/aom/aom_dsp/x86/inv_txfm_sse2.c b/third_party/aom/aom_dsp/x86/inv_txfm_sse2.c
new file mode 100644
index 0000000000..5795a1845c
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/inv_txfm_sse2.c
@@ -0,0 +1,3631 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/x86/inv_txfm_sse2.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+
+#define RECON_AND_STORE4X4(dest, in_x)                    \
+  {                                                       \
+    __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest)); \
+    d0 = _mm_unpacklo_epi8(d0, zero);                     \
+    d0 = _mm_add_epi16(in_x, d0);                         \
+    d0 = _mm_packus_epi16(d0, d0);                        \
+    *(int *)(dest) = _mm_cvtsi128_si32(d0);               \
+  }
+
+void aom_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i eight = _mm_set1_epi16(8);
+  const __m128i cst = _mm_setr_epi16(
+      (int16_t)cospi_16_64, (int16_t)cospi_16_64, (int16_t)cospi_16_64,
+      (int16_t)-cospi_16_64, (int16_t)cospi_24_64, (int16_t)-cospi_8_64,
+      (int16_t)cospi_8_64, (int16_t)cospi_24_64);
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i input0, input1, input2, input3;
+
+  // Rows
+  input0 = load_input_data(input);
+  input2 = load_input_data(input + 8);
+
+  // Construct i3, i1, i3, i1, i2, i0, i2, i0
+  input0 = _mm_shufflelo_epi16(input0, 0xd8);
+  input0 = _mm_shufflehi_epi16(input0, 0xd8);
+  input2 = _mm_shufflelo_epi16(input2, 0xd8);
+  input2 = _mm_shufflehi_epi16(input2, 0xd8);
+
+  input1 = _mm_unpackhi_epi32(input0, input0);
+  input0 = _mm_unpacklo_epi32(input0, input0);
+  input3 = _mm_unpackhi_epi32(input2, input2);
+  input2 = _mm_unpacklo_epi32(input2, input2);
+
+  // Stage 1
+  input0 = _mm_madd_epi16(input0, cst);
+  input1 = _mm_madd_epi16(input1, cst);
+  input2 = _mm_madd_epi16(input2, cst);
+  input3 = _mm_madd_epi16(input3, cst);
+
+  input0 = _mm_add_epi32(input0, rounding);
+  input1 = _mm_add_epi32(input1, rounding);
+  input2 = _mm_add_epi32(input2, rounding);
+  input3 = _mm_add_epi32(input3, rounding);
+
+  input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
+  input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
+  input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
+  input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
+
+  // Stage 2
+  input0 = _mm_packs_epi32(input0, input1);
+  input1 = _mm_packs_epi32(input2, input3);
+
+  // Transpose
+  input2 = _mm_unpacklo_epi16(input0, input1);
+  input3 = _mm_unpackhi_epi16(input0, input1);
+  input0 = _mm_unpacklo_epi32(input2, input3);
+  input1 = _mm_unpackhi_epi32(input2, input3);
+
+  // Switch column2, column 3, and then, we got:
+  // input2: column1, column 0;  input3: column2, column 3.
+  input1 = _mm_shuffle_epi32(input1, 0x4e);
+  input2 = _mm_add_epi16(input0, input1);
+  input3 = _mm_sub_epi16(input0, input1);
+
+  // Columns
+  // Construct i3, i1, i3, i1, i2, i0, i2, i0
+  input0 = _mm_unpacklo_epi32(input2, input2);
+  input1 = _mm_unpackhi_epi32(input2, input2);
+  input2 = _mm_unpackhi_epi32(input3, input3);
+  input3 = _mm_unpacklo_epi32(input3, input3);
+
+  // Stage 1
+  input0 = _mm_madd_epi16(input0, cst);
+  input1 = _mm_madd_epi16(input1, cst);
+  input2 = _mm_madd_epi16(input2, cst);
+  input3 = _mm_madd_epi16(input3, cst);
+
+  input0 = _mm_add_epi32(input0, rounding);
+  input1 = _mm_add_epi32(input1, rounding);
+  input2 = _mm_add_epi32(input2, rounding);
+  input3 = _mm_add_epi32(input3, rounding);
+
+  input0 = _mm_srai_epi32(input0, DCT_CONST_BITS);
+  input1 = _mm_srai_epi32(input1, DCT_CONST_BITS);
+  input2 = _mm_srai_epi32(input2, DCT_CONST_BITS);
+  input3 = _mm_srai_epi32(input3, DCT_CONST_BITS);
+
+  // Stage 2
+  input0 = _mm_packs_epi32(input0, input2);
+  input1 = _mm_packs_epi32(input1, input3);
+
+  // Transpose
+  input2 = _mm_unpacklo_epi16(input0, input1);
+  input3 = _mm_unpackhi_epi16(input0, input1);
+  input0 = _mm_unpacklo_epi32(input2, input3);
+  input1 = _mm_unpackhi_epi32(input2, input3);
+
+  // Switch column2, column 3, and then, we got:
+  // input2: column1, column 0;  input3: column2, column 3.
+  input1 = _mm_shuffle_epi32(input1, 0x4e);
+  input2 = _mm_add_epi16(input0, input1);
+  input3 = _mm_sub_epi16(input0, input1);
+
+  // Final round and shift
+  input2 = _mm_add_epi16(input2, eight);
+  input3 = _mm_add_epi16(input3, eight);
+
+  input2 = _mm_srai_epi16(input2, 4);
+  input3 = _mm_srai_epi16(input3, 4);
+
+  // Reconstruction and Store
+  {
+    __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest));
+    __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2));
+    d0 = _mm_unpacklo_epi32(d0,
+                            _mm_cvtsi32_si128(*(const int *)(dest + stride)));
+    d2 = _mm_unpacklo_epi32(
+        _mm_cvtsi32_si128(*(const int *)(dest + stride * 3)), d2);
+    d0 = _mm_unpacklo_epi8(d0, zero);
+    d2 = _mm_unpacklo_epi8(d2, zero);
+    d0 = _mm_add_epi16(d0, input2);
+    d2 = _mm_add_epi16(d2, input3);
+    d0 = _mm_packus_epi16(d0, d2);
+    // store input0
+    *(int *)dest = _mm_cvtsi128_si32(d0);
+    // store input1
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride) = _mm_cvtsi128_si32(d0);
+    // store input2
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0);
+    // store input3
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0);
+  }
+}
+
+void aom_idct4x4_1_add_sse2(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a;
+
+  a = (int)dct_const_round_shift(input[0] * cospi_16_64);
+  a = (int)dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 4);
+
+  if (a == 0) return;
+
+  dc_value = _mm_set1_epi16(a);
+
+  RECON_AND_STORE4X4(dest + 0 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 1 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 2 * stride, dc_value);
+  RECON_AND_STORE4X4(dest + 3 * stride, dc_value);
+}
+
+void aom_idct4_sse2(__m128i *in) {
+  const __m128i k__cospi_p16_p16 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8];
+
+  array_transpose_4x4(in);
+  // stage 1
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpackhi_epi16(in[0], in[1]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[1] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[2] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+
+  u[0] = _mm_packs_epi32(v[0], v[1]);
+  u[1] = _mm_packs_epi32(v[3], v[2]);
+
+  // stage 2
+  in[0] = _mm_add_epi16(u[0], u[1]);
+  in[1] = _mm_sub_epi16(u[0], u[1]);
+  in[1] = _mm_shuffle_epi32(in[1], 0x4E);
+}
+
+void aom_iadst4_sse2(__m128i *in) {
+  const __m128i k__sinpi_p01_p04 = pair_set_epi16(sinpi_1_9, sinpi_4_9);
+  const __m128i k__sinpi_p03_p02 = pair_set_epi16(sinpi_3_9, sinpi_2_9);
+  const __m128i k__sinpi_p02_m01 = pair_set_epi16(sinpi_2_9, -sinpi_1_9);
+  const __m128i k__sinpi_p03_m04 = pair_set_epi16(sinpi_3_9, -sinpi_4_9);
+  const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8], in7;
+
+  array_transpose_4x4(in);
+  in7 = _mm_srli_si128(in[1], 8);
+  in7 = _mm_add_epi16(in7, in[0]);
+  in7 = _mm_sub_epi16(in7, in[1]);
+
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpackhi_epi16(in[0], in[1]);
+  u[2] = _mm_unpacklo_epi16(in7, kZero);
+  u[3] = _mm_unpackhi_epi16(in[0], kZero);
+
+  v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p04);  // s0 + s3
+  v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p02);  // s2 + s5
+  v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x2
+  v[3] = _mm_madd_epi16(u[0], k__sinpi_p02_m01);  // s1 - s4
+  v[4] = _mm_madd_epi16(u[1], k__sinpi_p03_m04);  // s2 - s6
+  v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s2
+
+  u[0] = _mm_add_epi32(v[0], v[1]);
+  u[1] = _mm_add_epi32(v[3], v[4]);
+  u[2] = v[2];
+  u[3] = _mm_add_epi32(u[0], u[1]);
+  u[4] = _mm_slli_epi32(v[5], 2);
+  u[5] = _mm_add_epi32(u[3], v[5]);
+  u[6] = _mm_sub_epi32(u[5], u[4]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[1]);
+  in[1] = _mm_packs_epi32(u[2], u[3]);
+}
+
+// Define Macro for multiplying elements by constants and adding them together.
+#define MULTIPLICATION_AND_ADD(lo_0, hi_0, lo_1, hi_1, cst0, cst1, cst2, cst3, \
+                               res0, res1, res2, res3)                         \
+  {                                                                            \
+    tmp0 = _mm_madd_epi16(lo_0, cst0);                                         \
+    tmp1 = _mm_madd_epi16(hi_0, cst0);                                         \
+    tmp2 = _mm_madd_epi16(lo_0, cst1);                                         \
+    tmp3 = _mm_madd_epi16(hi_0, cst1);                                         \
+    tmp4 = _mm_madd_epi16(lo_1, cst2);                                         \
+    tmp5 = _mm_madd_epi16(hi_1, cst2);                                         \
+    tmp6 = _mm_madd_epi16(lo_1, cst3);                                         \
+    tmp7 = _mm_madd_epi16(hi_1, cst3);                                         \
+                                                                               \
+    tmp0 = _mm_add_epi32(tmp0, rounding);                                      \
+    tmp1 = _mm_add_epi32(tmp1, rounding);                                      \
+    tmp2 = _mm_add_epi32(tmp2, rounding);                                      \
+    tmp3 = _mm_add_epi32(tmp3, rounding);                                      \
+    tmp4 = _mm_add_epi32(tmp4, rounding);                                      \
+    tmp5 = _mm_add_epi32(tmp5, rounding);                                      \
+    tmp6 = _mm_add_epi32(tmp6, rounding);                                      \
+    tmp7 = _mm_add_epi32(tmp7, rounding);                                      \
+                                                                               \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);                               \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);                               \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);                               \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);                               \
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);                               \
+    tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);                               \
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);                               \
+    tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);                               \
+                                                                               \
+    res0 = _mm_packs_epi32(tmp0, tmp1);                                        \
+    res1 = _mm_packs_epi32(tmp2, tmp3);                                        \
+    res2 = _mm_packs_epi32(tmp4, tmp5);                                        \
+    res3 = _mm_packs_epi32(tmp6, tmp7);                                        \
+  }
+
+#define MULTIPLICATION_AND_ADD_2(lo_0, hi_0, cst0, cst1, res0, res1) \
+  {                                                                  \
+    tmp0 = _mm_madd_epi16(lo_0, cst0);                               \
+    tmp1 = _mm_madd_epi16(hi_0, cst0);                               \
+    tmp2 = _mm_madd_epi16(lo_0, cst1);                               \
+    tmp3 = _mm_madd_epi16(hi_0, cst1);                               \
+                                                                     \
+    tmp0 = _mm_add_epi32(tmp0, rounding);                            \
+    tmp1 = _mm_add_epi32(tmp1, rounding);                            \
+    tmp2 = _mm_add_epi32(tmp2, rounding);                            \
+    tmp3 = _mm_add_epi32(tmp3, rounding);                            \
+                                                                     \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);                     \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);                     \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);                     \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);                     \
+                                                                     \
+    res0 = _mm_packs_epi32(tmp0, tmp1);                              \
+    res1 = _mm_packs_epi32(tmp2, tmp3);                              \
+  }
+
+#define IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2, out3, \
+              out4, out5, out6, out7)                                         \
+  {                                                                           \
+    /* Stage1 */                                                              \
+    {                                                                         \
+      const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7);                     \
+      const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7);                     \
+      const __m128i lo_35 = _mm_unpacklo_epi16(in3, in5);                     \
+      const __m128i hi_35 = _mm_unpackhi_epi16(in3, in5);                     \
+                                                                              \
+      MULTIPLICATION_AND_ADD(lo_17, hi_17, lo_35, hi_35, stg1_0, stg1_1,      \
+                             stg1_2, stg1_3, stp1_4, stp1_7, stp1_5, stp1_6)  \
+    }                                                                         \
+                                                                              \
+    /* Stage2 */                                                              \
+    {                                                                         \
+      const __m128i lo_04 = _mm_unpacklo_epi16(in0, in4);                     \
+      const __m128i hi_04 = _mm_unpackhi_epi16(in0, in4);                     \
+      const __m128i lo_26 = _mm_unpacklo_epi16(in2, in6);                     \
+      const __m128i hi_26 = _mm_unpackhi_epi16(in2, in6);                     \
+                                                                              \
+      MULTIPLICATION_AND_ADD(lo_04, hi_04, lo_26, hi_26, stg2_0, stg2_1,      \
+                             stg2_2, stg2_3, stp2_0, stp2_1, stp2_2, stp2_3)  \
+                                                                              \
+      stp2_4 = _mm_adds_epi16(stp1_4, stp1_5);                                \
+      stp2_5 = _mm_subs_epi16(stp1_4, stp1_5);                                \
+      stp2_6 = _mm_subs_epi16(stp1_7, stp1_6);                                \
+      stp2_7 = _mm_adds_epi16(stp1_7, stp1_6);                                \
+    }                                                                         \
+                                                                              \
+    /* Stage3 */                                                              \
+    {                                                                         \
+      const __m128i lo_56 = _mm_unpacklo_epi16(stp2_6, stp2_5);               \
+      const __m128i hi_56 = _mm_unpackhi_epi16(stp2_6, stp2_5);               \
+                                                                              \
+      stp1_0 = _mm_adds_epi16(stp2_0, stp2_3);                                \
+      stp1_1 = _mm_adds_epi16(stp2_1, stp2_2);                                \
+      stp1_2 = _mm_subs_epi16(stp2_1, stp2_2);                                \
+      stp1_3 = _mm_subs_epi16(stp2_0, stp2_3);                                \
+                                                                              \
+      tmp0 = _mm_madd_epi16(lo_56, stg2_1);                                   \
+      tmp1 = _mm_madd_epi16(hi_56, stg2_1);                                   \
+      tmp2 = _mm_madd_epi16(lo_56, stg2_0);                                   \
+      tmp3 = _mm_madd_epi16(hi_56, stg2_0);                                   \
+                                                                              \
+      tmp0 = _mm_add_epi32(tmp0, rounding);                                   \
+      tmp1 = _mm_add_epi32(tmp1, rounding);                                   \
+      tmp2 = _mm_add_epi32(tmp2, rounding);                                   \
+      tmp3 = _mm_add_epi32(tmp3, rounding);                                   \
+                                                                              \
+      tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);                            \
+      tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);                            \
+      tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);                            \
+      tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);                            \
+                                                                              \
+      stp1_5 = _mm_packs_epi32(tmp0, tmp1);                                   \
+      stp1_6 = _mm_packs_epi32(tmp2, tmp3);                                   \
+    }                                                                         \
+                                                                              \
+    /* Stage4  */                                                             \
+    out0 = _mm_adds_epi16(stp1_0, stp2_7);                                    \
+    out1 = _mm_adds_epi16(stp1_1, stp1_6);                                    \
+    out2 = _mm_adds_epi16(stp1_2, stp1_5);                                    \
+    out3 = _mm_adds_epi16(stp1_3, stp2_4);                                    \
+    out4 = _mm_subs_epi16(stp1_3, stp2_4);                                    \
+    out5 = _mm_subs_epi16(stp1_2, stp1_5);                                    \
+    out6 = _mm_subs_epi16(stp1_1, stp1_6);                                    \
+    out7 = _mm_subs_epi16(stp1_0, stp2_7);                                    \
+  }
+
+void aom_idct8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  // Load input data.
+  in0 = load_input_data(input);
+  in1 = load_input_data(input + 8 * 1);
+  in2 = load_input_data(input + 8 * 2);
+  in3 = load_input_data(input + 8 * 3);
+  in4 = load_input_data(input + 8 * 4);
+  in5 = load_input_data(input + 8 * 5);
+  in6 = load_input_data(input + 8 * 6);
+  in7 = load_input_data(input + 8 * 7);
+
+  // 2-D
+  for (i = 0; i < 2; i++) {
+    // 8x8 Transpose is copied from aom_fdct8x8_sse2()
+    TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                  in4, in5, in6, in7);
+
+    // 4-stage 1D idct8x8
+    IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4, in5,
+          in6, in7);
+  }
+
+  // Final rounding and shift
+  in0 = _mm_adds_epi16(in0, final_rounding);
+  in1 = _mm_adds_epi16(in1, final_rounding);
+  in2 = _mm_adds_epi16(in2, final_rounding);
+  in3 = _mm_adds_epi16(in3, final_rounding);
+  in4 = _mm_adds_epi16(in4, final_rounding);
+  in5 = _mm_adds_epi16(in5, final_rounding);
+  in6 = _mm_adds_epi16(in6, final_rounding);
+  in7 = _mm_adds_epi16(in7, final_rounding);
+
+  in0 = _mm_srai_epi16(in0, 5);
+  in1 = _mm_srai_epi16(in1, 5);
+  in2 = _mm_srai_epi16(in2, 5);
+  in3 = _mm_srai_epi16(in3, 5);
+  in4 = _mm_srai_epi16(in4, 5);
+  in5 = _mm_srai_epi16(in5, 5);
+  in6 = _mm_srai_epi16(in6, 5);
+  in7 = _mm_srai_epi16(in7, 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in0);
+  RECON_AND_STORE(dest + 1 * stride, in1);
+  RECON_AND_STORE(dest + 2 * stride, in2);
+  RECON_AND_STORE(dest + 3 * stride, in3);
+  RECON_AND_STORE(dest + 4 * stride, in4);
+  RECON_AND_STORE(dest + 5 * stride, in5);
+  RECON_AND_STORE(dest + 6 * stride, in6);
+  RECON_AND_STORE(dest + 7 * stride, in7);
+}
+
+void aom_idct8x8_1_add_sse2(const tran_low_t *input, uint8_t *dest,
+                            int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a;
+
+  a = (int)dct_const_round_shift(input[0] * cospi_16_64);
+  a = (int)dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 5);
+
+  if (a == 0) return;
+
+  dc_value = _mm_set1_epi16(a);
+
+  RECON_AND_STORE(dest + 0 * stride, dc_value);
+  RECON_AND_STORE(dest + 1 * stride, dc_value);
+  RECON_AND_STORE(dest + 2 * stride, dc_value);
+  RECON_AND_STORE(dest + 3 * stride, dc_value);
+  RECON_AND_STORE(dest + 4 * stride, dc_value);
+  RECON_AND_STORE(dest + 5 * stride, dc_value);
+  RECON_AND_STORE(dest + 6 * stride, dc_value);
+  RECON_AND_STORE(dest + 7 * stride, dc_value);
+}
+
+void aom_idct8_sse2(__m128i *in) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+  // 8x8 Transpose is copied from aom_fdct8x8_sse2()
+  TRANSPOSE_8X8(in[0], in[1], in[2], in[3], in[4], in[5], in[6], in[7], in0,
+                in1, in2, in3, in4, in5, in6, in7);
+
+  // 4-stage 1D idct8x8
+  IDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in[0], in[1], in[2], in[3],
+        in[4], in[5], in[6], in[7]);
+}
+
+void aom_iadst8_sse2(__m128i *in) {
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__const_0 = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+
+  // transpose
+  array_transpose_8x8(in, in);
+
+  // properly aligned for butterfly input
+  in0 = in[7];
+  in1 = in[0];
+  in2 = in[5];
+  in3 = in[2];
+  in4 = in[3];
+  in5 = in[4];
+  in6 = in[1];
+  in7 = in[6];
+
+  // column transformation
+  // stage 1
+  // interleave and multiply/add into 32-bit integer
+  s0 = _mm_unpacklo_epi16(in0, in1);
+  s1 = _mm_unpackhi_epi16(in0, in1);
+  s2 = _mm_unpacklo_epi16(in2, in3);
+  s3 = _mm_unpackhi_epi16(in2, in3);
+  s4 = _mm_unpacklo_epi16(in4, in5);
+  s5 = _mm_unpackhi_epi16(in4, in5);
+  s6 = _mm_unpacklo_epi16(in6, in7);
+  s7 = _mm_unpackhi_epi16(in6, in7);
+
+  u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
+  u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
+  u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
+  u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
+  u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
+  u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
+  u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
+  u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
+  u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
+  u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
+  u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
+  u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
+  u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
+  u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
+  u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
+  u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
+
+  // addition
+  w0 = _mm_add_epi32(u0, u8);
+  w1 = _mm_add_epi32(u1, u9);
+  w2 = _mm_add_epi32(u2, u10);
+  w3 = _mm_add_epi32(u3, u11);
+  w4 = _mm_add_epi32(u4, u12);
+  w5 = _mm_add_epi32(u5, u13);
+  w6 = _mm_add_epi32(u6, u14);
+  w7 = _mm_add_epi32(u7, u15);
+  w8 = _mm_sub_epi32(u0, u8);
+  w9 = _mm_sub_epi32(u1, u9);
+  w10 = _mm_sub_epi32(u2, u10);
+  w11 = _mm_sub_epi32(u3, u11);
+  w12 = _mm_sub_epi32(u4, u12);
+  w13 = _mm_sub_epi32(u5, u13);
+  w14 = _mm_sub_epi32(u6, u14);
+  w15 = _mm_sub_epi32(u7, u15);
+
+  // shift and rounding
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+  v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
+  v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
+  v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
+  v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
+  v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
+  v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
+  v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
+  v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+  u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
+  u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
+  u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
+  u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
+  u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
+  u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
+  u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
+  u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
+
+  // back to 16-bit and pack 8 integers into __m128i
+  in[0] = _mm_packs_epi32(u0, u1);
+  in[1] = _mm_packs_epi32(u2, u3);
+  in[2] = _mm_packs_epi32(u4, u5);
+  in[3] = _mm_packs_epi32(u6, u7);
+  in[4] = _mm_packs_epi32(u8, u9);
+  in[5] = _mm_packs_epi32(u10, u11);
+  in[6] = _mm_packs_epi32(u12, u13);
+  in[7] = _mm_packs_epi32(u14, u15);
+
+  // stage 2
+  s0 = _mm_add_epi16(in[0], in[2]);
+  s1 = _mm_add_epi16(in[1], in[3]);
+  s2 = _mm_sub_epi16(in[0], in[2]);
+  s3 = _mm_sub_epi16(in[1], in[3]);
+  u0 = _mm_unpacklo_epi16(in[4], in[5]);
+  u1 = _mm_unpackhi_epi16(in[4], in[5]);
+  u2 = _mm_unpacklo_epi16(in[6], in[7]);
+  u3 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
+  v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
+  v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
+  v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
+  v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
+  v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
+  v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
+  v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
+
+  w0 = _mm_add_epi32(v0, v4);
+  w1 = _mm_add_epi32(v1, v5);
+  w2 = _mm_add_epi32(v2, v6);
+  w3 = _mm_add_epi32(v3, v7);
+  w4 = _mm_sub_epi32(v0, v4);
+  w5 = _mm_sub_epi32(v1, v5);
+  w6 = _mm_sub_epi32(v2, v6);
+  w7 = _mm_sub_epi32(v3, v7);
+
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  // back to 16-bit intergers
+  s4 = _mm_packs_epi32(u0, u1);
+  s5 = _mm_packs_epi32(u2, u3);
+  s6 = _mm_packs_epi32(u4, u5);
+  s7 = _mm_packs_epi32(u6, u7);
+
+  // stage 3
+  u0 = _mm_unpacklo_epi16(s2, s3);
+  u1 = _mm_unpackhi_epi16(s2, s3);
+  u2 = _mm_unpacklo_epi16(s6, s7);
+  u3 = _mm_unpackhi_epi16(s6, s7);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
+  v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
+  v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
+  v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
+  v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
+  v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
+  v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
+  v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
+
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  s2 = _mm_packs_epi32(v0, v1);
+  s3 = _mm_packs_epi32(v2, v3);
+  s6 = _mm_packs_epi32(v4, v5);
+  s7 = _mm_packs_epi32(v6, v7);
+
+  in[0] = s0;
+  in[1] = _mm_sub_epi16(k__const_0, s4);
+  in[2] = s6;
+  in[3] = _mm_sub_epi16(k__const_0, s2);
+  in[4] = s3;
+  in[5] = _mm_sub_epi16(k__const_0, s7);
+  in[6] = s5;
+  in[7] = _mm_sub_epi16(k__const_0, s1);
+}
+
+void aom_idct8x8_12_add_sse2(const tran_low_t *input, uint8_t *dest,
+                             int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg3_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+  // Rows. Load 4-row input data.
+  in0 = load_input_data(input);
+  in1 = load_input_data(input + 8 * 1);
+  in2 = load_input_data(input + 8 * 2);
+  in3 = load_input_data(input + 8 * 3);
+
+  // 8x4 Transpose
+  TRANSPOSE_8X8_10(in0, in1, in2, in3, in0, in1);
+  // Stage1
+  {
+    const __m128i lo_17 = _mm_unpackhi_epi16(in0, zero);
+    const __m128i lo_35 = _mm_unpackhi_epi16(in1, zero);
+
+    tmp0 = _mm_madd_epi16(lo_17, stg1_0);
+    tmp2 = _mm_madd_epi16(lo_17, stg1_1);
+    tmp4 = _mm_madd_epi16(lo_35, stg1_2);
+    tmp6 = _mm_madd_epi16(lo_35, stg1_3);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp1_4 = _mm_packs_epi32(tmp0, tmp2);
+    stp1_5 = _mm_packs_epi32(tmp4, tmp6);
+  }
+
+  // Stage2
+  {
+    const __m128i lo_04 = _mm_unpacklo_epi16(in0, zero);
+    const __m128i lo_26 = _mm_unpacklo_epi16(in1, zero);
+
+    tmp0 = _mm_madd_epi16(lo_04, stg2_0);
+    tmp2 = _mm_madd_epi16(lo_04, stg2_1);
+    tmp4 = _mm_madd_epi16(lo_26, stg2_2);
+    tmp6 = _mm_madd_epi16(lo_26, stg2_3);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp2_0 = _mm_packs_epi32(tmp0, tmp2);
+    stp2_2 = _mm_packs_epi32(tmp6, tmp4);
+
+    tmp0 = _mm_adds_epi16(stp1_4, stp1_5);
+    tmp1 = _mm_subs_epi16(stp1_4, stp1_5);
+
+    stp2_4 = tmp0;
+    stp2_5 = _mm_unpacklo_epi64(tmp1, zero);
+    stp2_6 = _mm_unpackhi_epi64(tmp1, zero);
+  }
+
+  // Stage3
+  {
+    const __m128i lo_56 = _mm_unpacklo_epi16(stp2_5, stp2_6);
+
+    tmp4 = _mm_adds_epi16(stp2_0, stp2_2);
+    tmp6 = _mm_subs_epi16(stp2_0, stp2_2);
+
+    stp1_2 = _mm_unpackhi_epi64(tmp6, tmp4);
+    stp1_3 = _mm_unpacklo_epi64(tmp6, tmp4);
+
+    tmp0 = _mm_madd_epi16(lo_56, stg3_0);
+    tmp2 = _mm_madd_epi16(lo_56, stg2_0);  // stg3_1 = stg2_0
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+
+    stp1_5 = _mm_packs_epi32(tmp0, tmp2);
+  }
+
+  // Stage4
+  tmp0 = _mm_adds_epi16(stp1_3, stp2_4);
+  tmp1 = _mm_adds_epi16(stp1_2, stp1_5);
+  tmp2 = _mm_subs_epi16(stp1_3, stp2_4);
+  tmp3 = _mm_subs_epi16(stp1_2, stp1_5);
+
+  TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, in0, in1, in2, in3)
+
+  IDCT8(in0, in1, in2, in3, zero, zero, zero, zero, in0, in1, in2, in3, in4,
+        in5, in6, in7);
+  // Final rounding and shift
+  in0 = _mm_adds_epi16(in0, final_rounding);
+  in1 = _mm_adds_epi16(in1, final_rounding);
+  in2 = _mm_adds_epi16(in2, final_rounding);
+  in3 = _mm_adds_epi16(in3, final_rounding);
+  in4 = _mm_adds_epi16(in4, final_rounding);
+  in5 = _mm_adds_epi16(in5, final_rounding);
+  in6 = _mm_adds_epi16(in6, final_rounding);
+  in7 = _mm_adds_epi16(in7, final_rounding);
+
+  in0 = _mm_srai_epi16(in0, 5);
+  in1 = _mm_srai_epi16(in1, 5);
+  in2 = _mm_srai_epi16(in2, 5);
+  in3 = _mm_srai_epi16(in3, 5);
+  in4 = _mm_srai_epi16(in4, 5);
+  in5 = _mm_srai_epi16(in5, 5);
+  in6 = _mm_srai_epi16(in6, 5);
+  in7 = _mm_srai_epi16(in7, 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in0);
+  RECON_AND_STORE(dest + 1 * stride, in1);
+  RECON_AND_STORE(dest + 2 * stride, in2);
+  RECON_AND_STORE(dest + 3 * stride, in3);
+  RECON_AND_STORE(dest + 4 * stride, in4);
+  RECON_AND_STORE(dest + 5 * stride, in5);
+  RECON_AND_STORE(dest + 6 * stride, in6);
+  RECON_AND_STORE(dest + 7 * stride, in7);
+}
+
+#define IDCT16                                                                 \
+  /* Stage2 */                                                                 \
+  {                                                                            \
+    const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], in[15]);                 \
+    const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], in[15]);                 \
+    const __m128i lo_9_7 = _mm_unpacklo_epi16(in[9], in[7]);                   \
+    const __m128i hi_9_7 = _mm_unpackhi_epi16(in[9], in[7]);                   \
+    const __m128i lo_5_11 = _mm_unpacklo_epi16(in[5], in[11]);                 \
+    const __m128i hi_5_11 = _mm_unpackhi_epi16(in[5], in[11]);                 \
+    const __m128i lo_13_3 = _mm_unpacklo_epi16(in[13], in[3]);                 \
+    const __m128i hi_13_3 = _mm_unpackhi_epi16(in[13], in[3]);                 \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_9_7, hi_9_7, stg2_0, stg2_1,   \
+                           stg2_2, stg2_3, stp2_8, stp2_15, stp2_9, stp2_14)   \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_5_11, hi_5_11, lo_13_3, hi_13_3, stg2_4, stg2_5, \
+                           stg2_6, stg2_7, stp2_10, stp2_13, stp2_11, stp2_12) \
+  }                                                                            \
+                                                                               \
+  /* Stage3 */                                                                 \
+  {                                                                            \
+    const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], in[14]);                 \
+    const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], in[14]);                 \
+    const __m128i lo_10_6 = _mm_unpacklo_epi16(in[10], in[6]);                 \
+    const __m128i hi_10_6 = _mm_unpackhi_epi16(in[10], in[6]);                 \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_2_14, hi_2_14, lo_10_6, hi_10_6, stg3_0, stg3_1, \
+                           stg3_2, stg3_3, stp1_4, stp1_7, stp1_5, stp1_6)     \
+                                                                               \
+    stp1_8_0 = _mm_add_epi16(stp2_8, stp2_9);                                  \
+    stp1_9 = _mm_sub_epi16(stp2_8, stp2_9);                                    \
+    stp1_10 = _mm_sub_epi16(stp2_11, stp2_10);                                 \
+    stp1_11 = _mm_add_epi16(stp2_11, stp2_10);                                 \
+                                                                               \
+    stp1_12_0 = _mm_add_epi16(stp2_12, stp2_13);                               \
+    stp1_13 = _mm_sub_epi16(stp2_12, stp2_13);                                 \
+    stp1_14 = _mm_sub_epi16(stp2_15, stp2_14);                                 \
+    stp1_15 = _mm_add_epi16(stp2_15, stp2_14);                                 \
+  }                                                                            \
+                                                                               \
+  /* Stage4 */                                                                 \
+  {                                                                            \
+    const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], in[8]);                   \
+    const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], in[8]);                   \
+    const __m128i lo_4_12 = _mm_unpacklo_epi16(in[4], in[12]);                 \
+    const __m128i hi_4_12 = _mm_unpackhi_epi16(in[4], in[12]);                 \
+                                                                               \
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14);               \
+    const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14);               \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);             \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13);             \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_0_8, hi_0_8, lo_4_12, hi_4_12, stg4_0, stg4_1,   \
+                           stg4_2, stg4_3, stp2_0, stp2_1, stp2_2, stp2_3)     \
+                                                                               \
+    stp2_4 = _mm_add_epi16(stp1_4, stp1_5);                                    \
+    stp2_5 = _mm_sub_epi16(stp1_4, stp1_5);                                    \
+    stp2_6 = _mm_sub_epi16(stp1_7, stp1_6);                                    \
+    stp2_7 = _mm_add_epi16(stp1_7, stp1_6);                                    \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4,       \
+                           stg4_5, stg4_6, stg4_7, stp2_9, stp2_14, stp2_10,   \
+                           stp2_13)                                            \
+  }                                                                            \
+                                                                               \
+  /* Stage5 */                                                                 \
+  {                                                                            \
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5);                 \
+    const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5);                 \
+                                                                               \
+    stp1_0 = _mm_add_epi16(stp2_0, stp2_3);                                    \
+    stp1_1 = _mm_add_epi16(stp2_1, stp2_2);                                    \
+    stp1_2 = _mm_sub_epi16(stp2_1, stp2_2);                                    \
+    stp1_3 = _mm_sub_epi16(stp2_0, stp2_3);                                    \
+                                                                               \
+    tmp0 = _mm_madd_epi16(lo_6_5, stg4_1);                                     \
+    tmp1 = _mm_madd_epi16(hi_6_5, stg4_1);                                     \
+    tmp2 = _mm_madd_epi16(lo_6_5, stg4_0);                                     \
+    tmp3 = _mm_madd_epi16(hi_6_5, stg4_0);                                     \
+                                                                               \
+    tmp0 = _mm_add_epi32(tmp0, rounding);                                      \
+    tmp1 = _mm_add_epi32(tmp1, rounding);                                      \
+    tmp2 = _mm_add_epi32(tmp2, rounding);                                      \
+    tmp3 = _mm_add_epi32(tmp3, rounding);                                      \
+                                                                               \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);                               \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);                               \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);                               \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);                               \
+                                                                               \
+    stp1_5 = _mm_packs_epi32(tmp0, tmp1);                                      \
+    stp1_6 = _mm_packs_epi32(tmp2, tmp3);                                      \
+                                                                               \
+    stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11);                                 \
+    stp1_9 = _mm_add_epi16(stp2_9, stp2_10);                                   \
+    stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);                                  \
+    stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11);                                \
+                                                                               \
+    stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0);                               \
+    stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);                                 \
+    stp1_14 = _mm_add_epi16(stp2_14, stp2_13);                                 \
+    stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0);                               \
+  }                                                                            \
+                                                                               \
+  /* Stage6 */                                                                 \
+  {                                                                            \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);             \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13);             \
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12);             \
+    const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12);             \
+                                                                               \
+    stp2_0 = _mm_add_epi16(stp1_0, stp2_7);                                    \
+    stp2_1 = _mm_add_epi16(stp1_1, stp1_6);                                    \
+    stp2_2 = _mm_add_epi16(stp1_2, stp1_5);                                    \
+    stp2_3 = _mm_add_epi16(stp1_3, stp2_4);                                    \
+    stp2_4 = _mm_sub_epi16(stp1_3, stp2_4);                                    \
+    stp2_5 = _mm_sub_epi16(stp1_2, stp1_5);                                    \
+    stp2_6 = _mm_sub_epi16(stp1_1, stp1_6);                                    \
+    stp2_7 = _mm_sub_epi16(stp1_0, stp2_7);                                    \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0,     \
+                           stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11,  \
+                           stp2_12)                                            \
+  }
+
+#define IDCT16_10                                                              \
+  /* Stage2 */                                                                 \
+  {                                                                            \
+    const __m128i lo_1_15 = _mm_unpacklo_epi16(in[1], zero);                   \
+    const __m128i hi_1_15 = _mm_unpackhi_epi16(in[1], zero);                   \
+    const __m128i lo_13_3 = _mm_unpacklo_epi16(zero, in[3]);                   \
+    const __m128i hi_13_3 = _mm_unpackhi_epi16(zero, in[3]);                   \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_1_15, hi_1_15, lo_13_3, hi_13_3, stg2_0, stg2_1, \
+                           stg2_6, stg2_7, stp1_8_0, stp1_15, stp1_11,         \
+                           stp1_12_0)                                          \
+  }                                                                            \
+                                                                               \
+  /* Stage3 */                                                                 \
+  {                                                                            \
+    const __m128i lo_2_14 = _mm_unpacklo_epi16(in[2], zero);                   \
+    const __m128i hi_2_14 = _mm_unpackhi_epi16(in[2], zero);                   \
+                                                                               \
+    MULTIPLICATION_AND_ADD_2(lo_2_14, hi_2_14, stg3_0, stg3_1, stp2_4, stp2_7) \
+                                                                               \
+    stp1_9 = stp1_8_0;                                                         \
+    stp1_10 = stp1_11;                                                         \
+                                                                               \
+    stp1_13 = stp1_12_0;                                                       \
+    stp1_14 = stp1_15;                                                         \
+  }                                                                            \
+                                                                               \
+  /* Stage4 */                                                                 \
+  {                                                                            \
+    const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero);                    \
+    const __m128i hi_0_8 = _mm_unpackhi_epi16(in[0], zero);                    \
+                                                                               \
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14);               \
+    const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14);               \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);             \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13);             \
+                                                                               \
+    MULTIPLICATION_AND_ADD_2(lo_0_8, hi_0_8, stg4_0, stg4_1, stp1_0, stp1_1)   \
+    stp2_5 = stp2_4;                                                           \
+    stp2_6 = stp2_7;                                                           \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4,       \
+                           stg4_5, stg4_6, stg4_7, stp2_9, stp2_14, stp2_10,   \
+                           stp2_13)                                            \
+  }                                                                            \
+                                                                               \
+  /* Stage5 */                                                                 \
+  {                                                                            \
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5);                 \
+    const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5);                 \
+                                                                               \
+    stp1_2 = stp1_1;                                                           \
+    stp1_3 = stp1_0;                                                           \
+                                                                               \
+    tmp0 = _mm_madd_epi16(lo_6_5, stg4_1);                                     \
+    tmp1 = _mm_madd_epi16(hi_6_5, stg4_1);                                     \
+    tmp2 = _mm_madd_epi16(lo_6_5, stg4_0);                                     \
+    tmp3 = _mm_madd_epi16(hi_6_5, stg4_0);                                     \
+                                                                               \
+    tmp0 = _mm_add_epi32(tmp0, rounding);                                      \
+    tmp1 = _mm_add_epi32(tmp1, rounding);                                      \
+    tmp2 = _mm_add_epi32(tmp2, rounding);                                      \
+    tmp3 = _mm_add_epi32(tmp3, rounding);                                      \
+                                                                               \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);                               \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);                               \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);                               \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);                               \
+                                                                               \
+    stp1_5 = _mm_packs_epi32(tmp0, tmp1);                                      \
+    stp1_6 = _mm_packs_epi32(tmp2, tmp3);                                      \
+                                                                               \
+    stp1_8 = _mm_add_epi16(stp1_8_0, stp1_11);                                 \
+    stp1_9 = _mm_add_epi16(stp2_9, stp2_10);                                   \
+    stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);                                  \
+    stp1_11 = _mm_sub_epi16(stp1_8_0, stp1_11);                                \
+                                                                               \
+    stp1_12 = _mm_sub_epi16(stp1_15, stp1_12_0);                               \
+    stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);                                 \
+    stp1_14 = _mm_add_epi16(stp2_14, stp2_13);                                 \
+    stp1_15 = _mm_add_epi16(stp1_15, stp1_12_0);                               \
+  }                                                                            \
+                                                                               \
+  /* Stage6 */                                                                 \
+  {                                                                            \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);             \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13);             \
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12);             \
+    const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12);             \
+                                                                               \
+    stp2_0 = _mm_add_epi16(stp1_0, stp2_7);                                    \
+    stp2_1 = _mm_add_epi16(stp1_1, stp1_6);                                    \
+    stp2_2 = _mm_add_epi16(stp1_2, stp1_5);                                    \
+    stp2_3 = _mm_add_epi16(stp1_3, stp2_4);                                    \
+    stp2_4 = _mm_sub_epi16(stp1_3, stp2_4);                                    \
+    stp2_5 = _mm_sub_epi16(stp1_2, stp1_5);                                    \
+    stp2_6 = _mm_sub_epi16(stp1_1, stp1_6);                                    \
+    stp2_7 = _mm_sub_epi16(stp1_0, stp2_7);                                    \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0,     \
+                           stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11,  \
+                           stp2_12)                                            \
+  }
+
+void aom_idct16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest,
+                                int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[16], l[16], r[16], *curr1;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+      stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+      stp1_8_0, stp1_12_0;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+      stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  curr1 = l;
+  for (i = 0; i < 2; i++) {
+    // 1-D idct
+
+    // Load input data.
+    in[0] = load_input_data(input);
+    in[8] = load_input_data(input + 8 * 1);
+    in[1] = load_input_data(input + 8 * 2);
+    in[9] = load_input_data(input + 8 * 3);
+    in[2] = load_input_data(input + 8 * 4);
+    in[10] = load_input_data(input + 8 * 5);
+    in[3] = load_input_data(input + 8 * 6);
+    in[11] = load_input_data(input + 8 * 7);
+    in[4] = load_input_data(input + 8 * 8);
+    in[12] = load_input_data(input + 8 * 9);
+    in[5] = load_input_data(input + 8 * 10);
+    in[13] = load_input_data(input + 8 * 11);
+    in[6] = load_input_data(input + 8 * 12);
+    in[14] = load_input_data(input + 8 * 13);
+    in[7] = load_input_data(input + 8 * 14);
+    in[15] = load_input_data(input + 8 * 15);
+
+    array_transpose_8x8(in, in);
+    array_transpose_8x8(in + 8, in + 8);
+
+    IDCT16
+
+    // Stage7
+    curr1[0] = _mm_add_epi16(stp2_0, stp1_15);
+    curr1[1] = _mm_add_epi16(stp2_1, stp1_14);
+    curr1[2] = _mm_add_epi16(stp2_2, stp2_13);
+    curr1[3] = _mm_add_epi16(stp2_3, stp2_12);
+    curr1[4] = _mm_add_epi16(stp2_4, stp2_11);
+    curr1[5] = _mm_add_epi16(stp2_5, stp2_10);
+    curr1[6] = _mm_add_epi16(stp2_6, stp1_9);
+    curr1[7] = _mm_add_epi16(stp2_7, stp1_8);
+    curr1[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    curr1[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    curr1[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    curr1[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    curr1[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    curr1[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    curr1[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    curr1[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    curr1 = r;
+    input += 128;
+  }
+  for (i = 0; i < 2; i++) {
+    int j;
+    // 1-D idct
+    array_transpose_8x8(l + i * 8, in);
+    array_transpose_8x8(r + i * 8, in + 8);
+
+    IDCT16
+
+    // 2-D
+    in[0] = _mm_add_epi16(stp2_0, stp1_15);
+    in[1] = _mm_add_epi16(stp2_1, stp1_14);
+    in[2] = _mm_add_epi16(stp2_2, stp2_13);
+    in[3] = _mm_add_epi16(stp2_3, stp2_12);
+    in[4] = _mm_add_epi16(stp2_4, stp2_11);
+    in[5] = _mm_add_epi16(stp2_5, stp2_10);
+    in[6] = _mm_add_epi16(stp2_6, stp1_9);
+    in[7] = _mm_add_epi16(stp2_7, stp1_8);
+    in[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    in[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    in[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    in[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    in[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    in[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    in[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    in[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    for (j = 0; j < 16; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void aom_idct16x16_1_add_sse2(const tran_low_t *input, uint8_t *dest,
+                              int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a, i;
+
+  a = (int)dct_const_round_shift(input[0] * cospi_16_64);
+  a = (int)dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 6);
+
+  if (a == 0) return;
+
+  dc_value = _mm_set1_epi16(a);
+
+  for (i = 0; i < 16; ++i) {
+    RECON_AND_STORE(dest + 0, dc_value);
+    RECON_AND_STORE(dest + 8, dc_value);
+    dest += stride;
+  }
+}
+
+void iadst16_8col(__m128i *in) {
+  // perform 16x16 1-D ADST for 8 columns
+  __m128i s[16], x[16], u[32], v[32];
+  const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
+  const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kZero = _mm_set1_epi16(0);
+
+  u[0] = _mm_unpacklo_epi16(in[15], in[0]);
+  u[1] = _mm_unpackhi_epi16(in[15], in[0]);
+  u[2] = _mm_unpacklo_epi16(in[13], in[2]);
+  u[3] = _mm_unpackhi_epi16(in[13], in[2]);
+  u[4] = _mm_unpacklo_epi16(in[11], in[4]);
+  u[5] = _mm_unpackhi_epi16(in[11], in[4]);
+  u[6] = _mm_unpacklo_epi16(in[9], in[6]);
+  u[7] = _mm_unpackhi_epi16(in[9], in[6]);
+  u[8] = _mm_unpacklo_epi16(in[7], in[8]);
+  u[9] = _mm_unpackhi_epi16(in[7], in[8]);
+  u[10] = _mm_unpacklo_epi16(in[5], in[10]);
+  u[11] = _mm_unpackhi_epi16(in[5], in[10]);
+  u[12] = _mm_unpacklo_epi16(in[3], in[12]);
+  u[13] = _mm_unpackhi_epi16(in[3], in[12]);
+  u[14] = _mm_unpacklo_epi16(in[1], in[14]);
+  u[15] = _mm_unpackhi_epi16(in[1], in[14]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
+  v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
+  v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
+  v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
+  v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
+  v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
+  v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
+  v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
+  v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
+  v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
+  v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
+  v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
+  v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
+  v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
+  v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
+  v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
+  v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
+
+  u[0] = _mm_add_epi32(v[0], v[16]);
+  u[1] = _mm_add_epi32(v[1], v[17]);
+  u[2] = _mm_add_epi32(v[2], v[18]);
+  u[3] = _mm_add_epi32(v[3], v[19]);
+  u[4] = _mm_add_epi32(v[4], v[20]);
+  u[5] = _mm_add_epi32(v[5], v[21]);
+  u[6] = _mm_add_epi32(v[6], v[22]);
+  u[7] = _mm_add_epi32(v[7], v[23]);
+  u[8] = _mm_add_epi32(v[8], v[24]);
+  u[9] = _mm_add_epi32(v[9], v[25]);
+  u[10] = _mm_add_epi32(v[10], v[26]);
+  u[11] = _mm_add_epi32(v[11], v[27]);
+  u[12] = _mm_add_epi32(v[12], v[28]);
+  u[13] = _mm_add_epi32(v[13], v[29]);
+  u[14] = _mm_add_epi32(v[14], v[30]);
+  u[15] = _mm_add_epi32(v[15], v[31]);
+  u[16] = _mm_sub_epi32(v[0], v[16]);
+  u[17] = _mm_sub_epi32(v[1], v[17]);
+  u[18] = _mm_sub_epi32(v[2], v[18]);
+  u[19] = _mm_sub_epi32(v[3], v[19]);
+  u[20] = _mm_sub_epi32(v[4], v[20]);
+  u[21] = _mm_sub_epi32(v[5], v[21]);
+  u[22] = _mm_sub_epi32(v[6], v[22]);
+  u[23] = _mm_sub_epi32(v[7], v[23]);
+  u[24] = _mm_sub_epi32(v[8], v[24]);
+  u[25] = _mm_sub_epi32(v[9], v[25]);
+  u[26] = _mm_sub_epi32(v[10], v[26]);
+  u[27] = _mm_sub_epi32(v[11], v[27]);
+  u[28] = _mm_sub_epi32(v[12], v[28]);
+  u[29] = _mm_sub_epi32(v[13], v[29]);
+  u[30] = _mm_sub_epi32(v[14], v[30]);
+  u[31] = _mm_sub_epi32(v[15], v[31]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+  v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
+  v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
+  v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
+  v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
+  v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
+  v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
+  v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
+  v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
+  v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
+  v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
+  v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
+  v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
+  v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
+  v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
+  v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
+  v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+  u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
+  u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
+  u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
+  u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
+  u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
+  u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
+  u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
+  u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
+  u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
+  u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
+  u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
+  u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
+  u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
+  u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
+  u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
+  u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
+
+  s[0] = _mm_packs_epi32(u[0], u[1]);
+  s[1] = _mm_packs_epi32(u[2], u[3]);
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_packs_epi32(u[8], u[9]);
+  s[5] = _mm_packs_epi32(u[10], u[11]);
+  s[6] = _mm_packs_epi32(u[12], u[13]);
+  s[7] = _mm_packs_epi32(u[14], u[15]);
+  s[8] = _mm_packs_epi32(u[16], u[17]);
+  s[9] = _mm_packs_epi32(u[18], u[19]);
+  s[10] = _mm_packs_epi32(u[20], u[21]);
+  s[11] = _mm_packs_epi32(u[22], u[23]);
+  s[12] = _mm_packs_epi32(u[24], u[25]);
+  s[13] = _mm_packs_epi32(u[26], u[27]);
+  s[14] = _mm_packs_epi32(u[28], u[29]);
+  s[15] = _mm_packs_epi32(u[30], u[31]);
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[9]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[9]);
+  u[2] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[3] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[4] = _mm_unpacklo_epi16(s[12], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[12], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], v[8]);
+  u[1] = _mm_add_epi32(v[1], v[9]);
+  u[2] = _mm_add_epi32(v[2], v[10]);
+  u[3] = _mm_add_epi32(v[3], v[11]);
+  u[4] = _mm_add_epi32(v[4], v[12]);
+  u[5] = _mm_add_epi32(v[5], v[13]);
+  u[6] = _mm_add_epi32(v[6], v[14]);
+  u[7] = _mm_add_epi32(v[7], v[15]);
+  u[8] = _mm_sub_epi32(v[0], v[8]);
+  u[9] = _mm_sub_epi32(v[1], v[9]);
+  u[10] = _mm_sub_epi32(v[2], v[10]);
+  u[11] = _mm_sub_epi32(v[3], v[11]);
+  u[12] = _mm_sub_epi32(v[4], v[12]);
+  u[13] = _mm_sub_epi32(v[5], v[13]);
+  u[14] = _mm_sub_epi32(v[6], v[14]);
+  u[15] = _mm_sub_epi32(v[7], v[15]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+  x[0] = _mm_add_epi16(s[0], s[4]);
+  x[1] = _mm_add_epi16(s[1], s[5]);
+  x[2] = _mm_add_epi16(s[2], s[6]);
+  x[3] = _mm_add_epi16(s[3], s[7]);
+  x[4] = _mm_sub_epi16(s[0], s[4]);
+  x[5] = _mm_sub_epi16(s[1], s[5]);
+  x[6] = _mm_sub_epi16(s[2], s[6]);
+  x[7] = _mm_sub_epi16(s[3], s[7]);
+  x[8] = _mm_packs_epi32(u[0], u[1]);
+  x[9] = _mm_packs_epi32(u[2], u[3]);
+  x[10] = _mm_packs_epi32(u[4], u[5]);
+  x[11] = _mm_packs_epi32(u[6], u[7]);
+  x[12] = _mm_packs_epi32(u[8], u[9]);
+  x[13] = _mm_packs_epi32(u[10], u[11]);
+  x[14] = _mm_packs_epi32(u[12], u[13]);
+  x[15] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  u[0] = _mm_unpacklo_epi16(x[4], x[5]);
+  u[1] = _mm_unpackhi_epi16(x[4], x[5]);
+  u[2] = _mm_unpacklo_epi16(x[6], x[7]);
+  u[3] = _mm_unpackhi_epi16(x[6], x[7]);
+  u[4] = _mm_unpacklo_epi16(x[12], x[13]);
+  u[5] = _mm_unpackhi_epi16(x[12], x[13]);
+  u[6] = _mm_unpacklo_epi16(x[14], x[15]);
+  u[7] = _mm_unpackhi_epi16(x[14], x[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], v[4]);
+  u[1] = _mm_add_epi32(v[1], v[5]);
+  u[2] = _mm_add_epi32(v[2], v[6]);
+  u[3] = _mm_add_epi32(v[3], v[7]);
+  u[4] = _mm_sub_epi32(v[0], v[4]);
+  u[5] = _mm_sub_epi32(v[1], v[5]);
+  u[6] = _mm_sub_epi32(v[2], v[6]);
+  u[7] = _mm_sub_epi32(v[3], v[7]);
+  u[8] = _mm_add_epi32(v[8], v[12]);
+  u[9] = _mm_add_epi32(v[9], v[13]);
+  u[10] = _mm_add_epi32(v[10], v[14]);
+  u[11] = _mm_add_epi32(v[11], v[15]);
+  u[12] = _mm_sub_epi32(v[8], v[12]);
+  u[13] = _mm_sub_epi32(v[9], v[13]);
+  u[14] = _mm_sub_epi32(v[10], v[14]);
+  u[15] = _mm_sub_epi32(v[11], v[15]);
+
+  u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[0] = _mm_add_epi16(x[0], x[2]);
+  s[1] = _mm_add_epi16(x[1], x[3]);
+  s[2] = _mm_sub_epi16(x[0], x[2]);
+  s[3] = _mm_sub_epi16(x[1], x[3]);
+  s[4] = _mm_packs_epi32(v[0], v[1]);
+  s[5] = _mm_packs_epi32(v[2], v[3]);
+  s[6] = _mm_packs_epi32(v[4], v[5]);
+  s[7] = _mm_packs_epi32(v[6], v[7]);
+  s[8] = _mm_add_epi16(x[8], x[10]);
+  s[9] = _mm_add_epi16(x[9], x[11]);
+  s[10] = _mm_sub_epi16(x[8], x[10]);
+  s[11] = _mm_sub_epi16(x[9], x[11]);
+  s[12] = _mm_packs_epi32(v[8], v[9]);
+  s[13] = _mm_packs_epi32(v[10], v[11]);
+  s[14] = _mm_packs_epi32(v[12], v[13]);
+  s[15] = _mm_packs_epi32(v[14], v[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(s[2], s[3]);
+  u[1] = _mm_unpackhi_epi16(s[2], s[3]);
+  u[2] = _mm_unpacklo_epi16(s[6], s[7]);
+  u[3] = _mm_unpackhi_epi16(s[6], s[7]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  in[0] = s[0];
+  in[1] = _mm_sub_epi16(kZero, s[8]);
+  in[2] = s[12];
+  in[3] = _mm_sub_epi16(kZero, s[4]);
+  in[4] = _mm_packs_epi32(v[4], v[5]);
+  in[5] = _mm_packs_epi32(v[12], v[13]);
+  in[6] = _mm_packs_epi32(v[8], v[9]);
+  in[7] = _mm_packs_epi32(v[0], v[1]);
+  in[8] = _mm_packs_epi32(v[2], v[3]);
+  in[9] = _mm_packs_epi32(v[10], v[11]);
+  in[10] = _mm_packs_epi32(v[14], v[15]);
+  in[11] = _mm_packs_epi32(v[6], v[7]);
+  in[12] = s[5];
+  in[13] = _mm_sub_epi16(kZero, s[13]);
+  in[14] = s[9];
+  in[15] = _mm_sub_epi16(kZero, s[1]);
+}
+
+void idct16_8col(__m128i *in) {
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i v[16], u[16], s[16], t[16];
+
+  // stage 1
+  s[0] = in[0];
+  s[1] = in[8];
+  s[2] = in[4];
+  s[3] = in[12];
+  s[4] = in[2];
+  s[5] = in[10];
+  s[6] = in[6];
+  s[7] = in[14];
+  s[8] = in[1];
+  s[9] = in[9];
+  s[10] = in[5];
+  s[11] = in[13];
+  s[12] = in[3];
+  s[13] = in[11];
+  s[14] = in[7];
+  s[15] = in[15];
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[15]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[15]);
+  u[2] = _mm_unpacklo_epi16(s[9], s[14]);
+  u[3] = _mm_unpackhi_epi16(s[9], s[14]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[11], s[12]);
+  u[7] = _mm_unpackhi_epi16(s[11], s[12]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p30_m02);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p30_m02);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p02_p30);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p02_p30);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p14_m18);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p14_m18);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p18_p14);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p18_p14);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p22_m10);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p22_m10);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p10_p22);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p10_p22);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p06_m26);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p06_m26);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p26_p06);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p26_p06);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[8] = _mm_packs_epi32(u[0], u[1]);
+  s[15] = _mm_packs_epi32(u[2], u[3]);
+  s[9] = _mm_packs_epi32(u[4], u[5]);
+  s[14] = _mm_packs_epi32(u[6], u[7]);
+  s[10] = _mm_packs_epi32(u[8], u[9]);
+  s[13] = _mm_packs_epi32(u[10], u[11]);
+  s[11] = _mm_packs_epi32(u[12], u[13]);
+  s[12] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  t[0] = s[0];
+  t[1] = s[1];
+  t[2] = s[2];
+  t[3] = s[3];
+  u[0] = _mm_unpacklo_epi16(s[4], s[7]);
+  u[1] = _mm_unpackhi_epi16(s[4], s[7]);
+  u[2] = _mm_unpacklo_epi16(s[5], s[6]);
+  u[3] = _mm_unpackhi_epi16(s[5], s[6]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[4] = _mm_packs_epi32(u[0], u[1]);
+  t[7] = _mm_packs_epi32(u[2], u[3]);
+  t[5] = _mm_packs_epi32(u[4], u[5]);
+  t[6] = _mm_packs_epi32(u[6], u[7]);
+  t[8] = _mm_add_epi16(s[8], s[9]);
+  t[9] = _mm_sub_epi16(s[8], s[9]);
+  t[10] = _mm_sub_epi16(s[11], s[10]);
+  t[11] = _mm_add_epi16(s[10], s[11]);
+  t[12] = _mm_add_epi16(s[12], s[13]);
+  t[13] = _mm_sub_epi16(s[12], s[13]);
+  t[14] = _mm_sub_epi16(s[15], s[14]);
+  t[15] = _mm_add_epi16(s[14], s[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(t[0], t[1]);
+  u[1] = _mm_unpackhi_epi16(t[0], t[1]);
+  u[2] = _mm_unpacklo_epi16(t[2], t[3]);
+  u[3] = _mm_unpackhi_epi16(t[2], t[3]);
+  u[4] = _mm_unpacklo_epi16(t[9], t[14]);
+  u[5] = _mm_unpackhi_epi16(t[9], t[14]);
+  u[6] = _mm_unpacklo_epi16(t[10], t[13]);
+  u[7] = _mm_unpackhi_epi16(t[10], t[13]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p24_m08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p24_m08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_p08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_p08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_m08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_m08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_m08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_m08_p24);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[0] = _mm_packs_epi32(u[0], u[1]);
+  s[1] = _mm_packs_epi32(u[2], u[3]);
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_add_epi16(t[4], t[5]);
+  s[5] = _mm_sub_epi16(t[4], t[5]);
+  s[6] = _mm_sub_epi16(t[7], t[6]);
+  s[7] = _mm_add_epi16(t[6], t[7]);
+  s[8] = t[8];
+  s[15] = t[15];
+  s[9] = _mm_packs_epi32(u[8], u[9]);
+  s[14] = _mm_packs_epi32(u[10], u[11]);
+  s[10] = _mm_packs_epi32(u[12], u[13]);
+  s[13] = _mm_packs_epi32(u[14], u[15]);
+  s[11] = t[11];
+  s[12] = t[12];
+
+  // stage 5
+  t[0] = _mm_add_epi16(s[0], s[3]);
+  t[1] = _mm_add_epi16(s[1], s[2]);
+  t[2] = _mm_sub_epi16(s[1], s[2]);
+  t[3] = _mm_sub_epi16(s[0], s[3]);
+  t[4] = s[4];
+  t[7] = s[7];
+
+  u[0] = _mm_unpacklo_epi16(s[5], s[6]);
+  u[1] = _mm_unpackhi_epi16(s[5], s[6]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  t[5] = _mm_packs_epi32(u[0], u[1]);
+  t[6] = _mm_packs_epi32(u[2], u[3]);
+
+  t[8] = _mm_add_epi16(s[8], s[11]);
+  t[9] = _mm_add_epi16(s[9], s[10]);
+  t[10] = _mm_sub_epi16(s[9], s[10]);
+  t[11] = _mm_sub_epi16(s[8], s[11]);
+  t[12] = _mm_sub_epi16(s[15], s[12]);
+  t[13] = _mm_sub_epi16(s[14], s[13]);
+  t[14] = _mm_add_epi16(s[13], s[14]);
+  t[15] = _mm_add_epi16(s[12], s[15]);
+
+  // stage 6
+  s[0] = _mm_add_epi16(t[0], t[7]);
+  s[1] = _mm_add_epi16(t[1], t[6]);
+  s[2] = _mm_add_epi16(t[2], t[5]);
+  s[3] = _mm_add_epi16(t[3], t[4]);
+  s[4] = _mm_sub_epi16(t[3], t[4]);
+  s[5] = _mm_sub_epi16(t[2], t[5]);
+  s[6] = _mm_sub_epi16(t[1], t[6]);
+  s[7] = _mm_sub_epi16(t[0], t[7]);
+  s[8] = t[8];
+  s[9] = t[9];
+
+  u[0] = _mm_unpacklo_epi16(t[10], t[13]);
+  u[1] = _mm_unpackhi_epi16(t[10], t[13]);
+  u[2] = _mm_unpacklo_epi16(t[11], t[12]);
+  u[3] = _mm_unpackhi_epi16(t[11], t[12]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  s[10] = _mm_packs_epi32(u[0], u[1]);
+  s[13] = _mm_packs_epi32(u[2], u[3]);
+  s[11] = _mm_packs_epi32(u[4], u[5]);
+  s[12] = _mm_packs_epi32(u[6], u[7]);
+  s[14] = t[14];
+  s[15] = t[15];
+
+  // stage 7
+  in[0] = _mm_add_epi16(s[0], s[15]);
+  in[1] = _mm_add_epi16(s[1], s[14]);
+  in[2] = _mm_add_epi16(s[2], s[13]);
+  in[3] = _mm_add_epi16(s[3], s[12]);
+  in[4] = _mm_add_epi16(s[4], s[11]);
+  in[5] = _mm_add_epi16(s[5], s[10]);
+  in[6] = _mm_add_epi16(s[6], s[9]);
+  in[7] = _mm_add_epi16(s[7], s[8]);
+  in[8] = _mm_sub_epi16(s[7], s[8]);
+  in[9] = _mm_sub_epi16(s[6], s[9]);
+  in[10] = _mm_sub_epi16(s[5], s[10]);
+  in[11] = _mm_sub_epi16(s[4], s[11]);
+  in[12] = _mm_sub_epi16(s[3], s[12]);
+  in[13] = _mm_sub_epi16(s[2], s[13]);
+  in[14] = _mm_sub_epi16(s[1], s[14]);
+  in[15] = _mm_sub_epi16(s[0], s[15]);
+}
+
+void aom_idct16_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  idct16_8col(in0);
+  idct16_8col(in1);
+}
+
+void aom_iadst16_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  iadst16_8col(in0);
+  iadst16_8col(in1);
+}
+
+void aom_idct16x16_10_add_sse2(const tran_low_t *input, uint8_t *dest,
+                               int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i stg4_7 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  __m128i in[16], l[16];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_8,
+      stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15, stp1_8_0,
+      stp1_12_0;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+      stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+  // First 1-D inverse DCT
+  // Load input data.
+  in[0] = load_input_data(input);
+  in[1] = load_input_data(input + 8 * 2);
+  in[2] = load_input_data(input + 8 * 4);
+  in[3] = load_input_data(input + 8 * 6);
+
+  TRANSPOSE_8X4(in[0], in[1], in[2], in[3], in[0], in[1]);
+
+  // Stage2
+  {
+    const __m128i lo_1_15 = _mm_unpackhi_epi16(in[0], zero);
+    const __m128i lo_13_3 = _mm_unpackhi_epi16(zero, in[1]);
+
+    tmp0 = _mm_madd_epi16(lo_1_15, stg2_0);
+    tmp2 = _mm_madd_epi16(lo_1_15, stg2_1);
+    tmp5 = _mm_madd_epi16(lo_13_3, stg2_6);
+    tmp7 = _mm_madd_epi16(lo_13_3, stg2_7);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp5 = _mm_add_epi32(tmp5, rounding);
+    tmp7 = _mm_add_epi32(tmp7, rounding);
+
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
+    tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
+
+    stp2_8 = _mm_packs_epi32(tmp0, tmp2);
+    stp2_11 = _mm_packs_epi32(tmp5, tmp7);
+  }
+
+  // Stage3
+  {
+    const __m128i lo_2_14 = _mm_unpacklo_epi16(in[1], zero);
+
+    tmp0 = _mm_madd_epi16(lo_2_14, stg3_0);
+    tmp2 = _mm_madd_epi16(lo_2_14, stg3_1);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+
+    stp1_13 = _mm_unpackhi_epi64(stp2_11, zero);
+    stp1_14 = _mm_unpackhi_epi64(stp2_8, zero);
+
+    stp1_4 = _mm_packs_epi32(tmp0, tmp2);
+  }
+
+  // Stage4
+  {
+    const __m128i lo_0_8 = _mm_unpacklo_epi16(in[0], zero);
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp1_14);
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp1_13);
+
+    tmp0 = _mm_madd_epi16(lo_0_8, stg4_0);
+    tmp2 = _mm_madd_epi16(lo_0_8, stg4_1);
+    tmp1 = _mm_madd_epi16(lo_9_14, stg4_4);
+    tmp3 = _mm_madd_epi16(lo_9_14, stg4_5);
+    tmp5 = _mm_madd_epi16(lo_10_13, stg4_6);
+    tmp7 = _mm_madd_epi16(lo_10_13, stg4_7);
+
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp1 = _mm_add_epi32(tmp1, rounding);
+    tmp3 = _mm_add_epi32(tmp3, rounding);
+    tmp5 = _mm_add_epi32(tmp5, rounding);
+    tmp7 = _mm_add_epi32(tmp7, rounding);
+
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+    tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
+    tmp7 = _mm_srai_epi32(tmp7, DCT_CONST_BITS);
+
+    stp1_0 = _mm_packs_epi32(tmp0, tmp0);
+    stp1_1 = _mm_packs_epi32(tmp2, tmp2);
+    stp2_9 = _mm_packs_epi32(tmp1, tmp3);
+    stp2_10 = _mm_packs_epi32(tmp5, tmp7);
+
+    stp2_6 = _mm_unpackhi_epi64(stp1_4, zero);
+  }
+
+  // Stage5 and Stage6
+  {
+    tmp0 = _mm_add_epi16(stp2_8, stp2_11);
+    tmp1 = _mm_sub_epi16(stp2_8, stp2_11);
+    tmp2 = _mm_add_epi16(stp2_9, stp2_10);
+    tmp3 = _mm_sub_epi16(stp2_9, stp2_10);
+
+    stp1_9 = _mm_unpacklo_epi64(tmp2, zero);
+    stp1_10 = _mm_unpacklo_epi64(tmp3, zero);
+    stp1_8 = _mm_unpacklo_epi64(tmp0, zero);
+    stp1_11 = _mm_unpacklo_epi64(tmp1, zero);
+
+    stp1_13 = _mm_unpackhi_epi64(tmp3, zero);
+    stp1_14 = _mm_unpackhi_epi64(tmp2, zero);
+    stp1_12 = _mm_unpackhi_epi64(tmp1, zero);
+    stp1_15 = _mm_unpackhi_epi64(tmp0, zero);
+  }
+
+  // Stage6
+  {
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp1_4);
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12);
+
+    tmp1 = _mm_madd_epi16(lo_6_5, stg4_1);
+    tmp3 = _mm_madd_epi16(lo_6_5, stg4_0);
+    tmp0 = _mm_madd_epi16(lo_10_13, stg6_0);
+    tmp2 = _mm_madd_epi16(lo_10_13, stg4_0);
+    tmp4 = _mm_madd_epi16(lo_11_12, stg6_0);
+    tmp6 = _mm_madd_epi16(lo_11_12, stg4_0);
+
+    tmp1 = _mm_add_epi32(tmp1, rounding);
+    tmp3 = _mm_add_epi32(tmp3, rounding);
+    tmp0 = _mm_add_epi32(tmp0, rounding);
+    tmp2 = _mm_add_epi32(tmp2, rounding);
+    tmp4 = _mm_add_epi32(tmp4, rounding);
+    tmp6 = _mm_add_epi32(tmp6, rounding);
+
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+    tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+    tmp6 = _mm_srai_epi32(tmp6, DCT_CONST_BITS);
+
+    stp1_6 = _mm_packs_epi32(tmp3, tmp1);
+
+    stp2_10 = _mm_packs_epi32(tmp0, zero);
+    stp2_13 = _mm_packs_epi32(tmp2, zero);
+    stp2_11 = _mm_packs_epi32(tmp4, zero);
+    stp2_12 = _mm_packs_epi32(tmp6, zero);
+
+    tmp0 = _mm_add_epi16(stp1_0, stp1_4);
+    tmp1 = _mm_sub_epi16(stp1_0, stp1_4);
+    tmp2 = _mm_add_epi16(stp1_1, stp1_6);
+    tmp3 = _mm_sub_epi16(stp1_1, stp1_6);
+
+    stp2_0 = _mm_unpackhi_epi64(tmp0, zero);
+    stp2_1 = _mm_unpacklo_epi64(tmp2, zero);
+    stp2_2 = _mm_unpackhi_epi64(tmp2, zero);
+    stp2_3 = _mm_unpacklo_epi64(tmp0, zero);
+    stp2_4 = _mm_unpacklo_epi64(tmp1, zero);
+    stp2_5 = _mm_unpackhi_epi64(tmp3, zero);
+    stp2_6 = _mm_unpacklo_epi64(tmp3, zero);
+    stp2_7 = _mm_unpackhi_epi64(tmp1, zero);
+  }
+
+  // Stage7. Left 8x16 only.
+  l[0] = _mm_add_epi16(stp2_0, stp1_15);
+  l[1] = _mm_add_epi16(stp2_1, stp1_14);
+  l[2] = _mm_add_epi16(stp2_2, stp2_13);
+  l[3] = _mm_add_epi16(stp2_3, stp2_12);
+  l[4] = _mm_add_epi16(stp2_4, stp2_11);
+  l[5] = _mm_add_epi16(stp2_5, stp2_10);
+  l[6] = _mm_add_epi16(stp2_6, stp1_9);
+  l[7] = _mm_add_epi16(stp2_7, stp1_8);
+  l[8] = _mm_sub_epi16(stp2_7, stp1_8);
+  l[9] = _mm_sub_epi16(stp2_6, stp1_9);
+  l[10] = _mm_sub_epi16(stp2_5, stp2_10);
+  l[11] = _mm_sub_epi16(stp2_4, stp2_11);
+  l[12] = _mm_sub_epi16(stp2_3, stp2_12);
+  l[13] = _mm_sub_epi16(stp2_2, stp2_13);
+  l[14] = _mm_sub_epi16(stp2_1, stp1_14);
+  l[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+  // Second 1-D inverse transform, performed per 8x16 block
+  for (i = 0; i < 2; i++) {
+    int j;
+    array_transpose_4X8(l + 8 * i, in);
+
+    IDCT16_10
+
+    // Stage7
+    in[0] = _mm_add_epi16(stp2_0, stp1_15);
+    in[1] = _mm_add_epi16(stp2_1, stp1_14);
+    in[2] = _mm_add_epi16(stp2_2, stp2_13);
+    in[3] = _mm_add_epi16(stp2_3, stp2_12);
+    in[4] = _mm_add_epi16(stp2_4, stp2_11);
+    in[5] = _mm_add_epi16(stp2_5, stp2_10);
+    in[6] = _mm_add_epi16(stp2_6, stp1_9);
+    in[7] = _mm_add_epi16(stp2_7, stp1_8);
+    in[8] = _mm_sub_epi16(stp2_7, stp1_8);
+    in[9] = _mm_sub_epi16(stp2_6, stp1_9);
+    in[10] = _mm_sub_epi16(stp2_5, stp2_10);
+    in[11] = _mm_sub_epi16(stp2_4, stp2_11);
+    in[12] = _mm_sub_epi16(stp2_3, stp2_12);
+    in[13] = _mm_sub_epi16(stp2_2, stp2_13);
+    in[14] = _mm_sub_epi16(stp2_1, stp1_14);
+    in[15] = _mm_sub_epi16(stp2_0, stp1_15);
+
+    for (j = 0; j < 16; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+#define LOAD_DQCOEFF(reg, input)  \
+  {                               \
+    reg = load_input_data(input); \
+    input += 8;                   \
+  }
+
+#define IDCT32_34                                                              \
+  /* Stage1 */                                                                 \
+  {                                                                            \
+    const __m128i lo_1_31 = _mm_unpacklo_epi16(in[1], zero);                   \
+    const __m128i hi_1_31 = _mm_unpackhi_epi16(in[1], zero);                   \
+                                                                               \
+    const __m128i lo_25_7 = _mm_unpacklo_epi16(zero, in[7]);                   \
+    const __m128i hi_25_7 = _mm_unpackhi_epi16(zero, in[7]);                   \
+                                                                               \
+    const __m128i lo_5_27 = _mm_unpacklo_epi16(in[5], zero);                   \
+    const __m128i hi_5_27 = _mm_unpackhi_epi16(in[5], zero);                   \
+                                                                               \
+    const __m128i lo_29_3 = _mm_unpacklo_epi16(zero, in[3]);                   \
+    const __m128i hi_29_3 = _mm_unpackhi_epi16(zero, in[3]);                   \
+                                                                               \
+    MULTIPLICATION_AND_ADD_2(lo_1_31, hi_1_31, stg1_0, stg1_1, stp1_16,        \
+                             stp1_31);                                         \
+    MULTIPLICATION_AND_ADD_2(lo_25_7, hi_25_7, stg1_6, stg1_7, stp1_19,        \
+                             stp1_28);                                         \
+    MULTIPLICATION_AND_ADD_2(lo_5_27, hi_5_27, stg1_8, stg1_9, stp1_20,        \
+                             stp1_27);                                         \
+    MULTIPLICATION_AND_ADD_2(lo_29_3, hi_29_3, stg1_14, stg1_15, stp1_23,      \
+                             stp1_24);                                         \
+  }                                                                            \
+                                                                               \
+  /* Stage2 */                                                                 \
+  {                                                                            \
+    const __m128i lo_2_30 = _mm_unpacklo_epi16(in[2], zero);                   \
+    const __m128i hi_2_30 = _mm_unpackhi_epi16(in[2], zero);                   \
+                                                                               \
+    const __m128i lo_26_6 = _mm_unpacklo_epi16(zero, in[6]);                   \
+    const __m128i hi_26_6 = _mm_unpackhi_epi16(zero, in[6]);                   \
+                                                                               \
+    MULTIPLICATION_AND_ADD_2(lo_2_30, hi_2_30, stg2_0, stg2_1, stp2_8,         \
+                             stp2_15);                                         \
+    MULTIPLICATION_AND_ADD_2(lo_26_6, hi_26_6, stg2_6, stg2_7, stp2_11,        \
+                             stp2_12);                                         \
+                                                                               \
+    stp2_16 = stp1_16;                                                         \
+    stp2_19 = stp1_19;                                                         \
+                                                                               \
+    stp2_20 = stp1_20;                                                         \
+    stp2_23 = stp1_23;                                                         \
+                                                                               \
+    stp2_24 = stp1_24;                                                         \
+    stp2_27 = stp1_27;                                                         \
+                                                                               \
+    stp2_28 = stp1_28;                                                         \
+    stp2_31 = stp1_31;                                                         \
+  }                                                                            \
+                                                                               \
+  /* Stage3 */                                                                 \
+  {                                                                            \
+    const __m128i lo_4_28 = _mm_unpacklo_epi16(in[4], zero);                   \
+    const __m128i hi_4_28 = _mm_unpackhi_epi16(in[4], zero);                   \
+                                                                               \
+    const __m128i lo_17_30 = _mm_unpacklo_epi16(stp1_16, stp1_31);             \
+    const __m128i hi_17_30 = _mm_unpackhi_epi16(stp1_16, stp1_31);             \
+    const __m128i lo_18_29 = _mm_unpacklo_epi16(stp1_19, stp1_28);             \
+    const __m128i hi_18_29 = _mm_unpackhi_epi16(stp1_19, stp1_28);             \
+                                                                               \
+    const __m128i lo_21_26 = _mm_unpacklo_epi16(stp1_20, stp1_27);             \
+    const __m128i hi_21_26 = _mm_unpackhi_epi16(stp1_20, stp1_27);             \
+    const __m128i lo_22_25 = _mm_unpacklo_epi16(stp1_23, stp1_24);             \
+    const __m128i hi_22_25 = _mm_unpackhi_epi16(stp1_23, stp2_24);             \
+                                                                               \
+    MULTIPLICATION_AND_ADD_2(lo_4_28, hi_4_28, stg3_0, stg3_1, stp1_4,         \
+                             stp1_7);                                          \
+                                                                               \
+    stp1_8 = stp2_8;                                                           \
+    stp1_11 = stp2_11;                                                         \
+    stp1_12 = stp2_12;                                                         \
+    stp1_15 = stp2_15;                                                         \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4,     \
+                           stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, stp1_18,  \
+                           stp1_29)                                            \
+    MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8,     \
+                           stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, stp1_22, \
+                           stp1_25)                                            \
+                                                                               \
+    stp1_16 = stp2_16;                                                         \
+    stp1_31 = stp2_31;                                                         \
+    stp1_19 = stp2_19;                                                         \
+    stp1_20 = stp2_20;                                                         \
+    stp1_23 = stp2_23;                                                         \
+    stp1_24 = stp2_24;                                                         \
+    stp1_27 = stp2_27;                                                         \
+    stp1_28 = stp2_28;                                                         \
+  }                                                                            \
+                                                                               \
+  /* Stage4 */                                                                 \
+  {                                                                            \
+    const __m128i lo_0_16 = _mm_unpacklo_epi16(in[0], zero);                   \
+    const __m128i hi_0_16 = _mm_unpackhi_epi16(in[0], zero);                   \
+                                                                               \
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp2_8, stp2_15);               \
+    const __m128i hi_9_14 = _mm_unpackhi_epi16(stp2_8, stp2_15);               \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp2_11, stp2_12);             \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp2_11, stp2_12);             \
+                                                                               \
+    MULTIPLICATION_AND_ADD_2(lo_0_16, hi_0_16, stg4_0, stg4_1, stp2_0,         \
+                             stp2_1);                                          \
+                                                                               \
+    stp2_4 = stp1_4;                                                           \
+    stp2_5 = stp1_4;                                                           \
+    stp2_6 = stp1_7;                                                           \
+    stp2_7 = stp1_7;                                                           \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4,       \
+                           stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, stp2_10,   \
+                           stp2_13)                                            \
+                                                                               \
+    stp2_8 = stp1_8;                                                           \
+    stp2_15 = stp1_15;                                                         \
+    stp2_11 = stp1_11;                                                         \
+    stp2_12 = stp1_12;                                                         \
+                                                                               \
+    stp2_16 = _mm_add_epi16(stp1_16, stp1_19);                                 \
+    stp2_17 = _mm_add_epi16(stp1_17, stp1_18);                                 \
+    stp2_18 = _mm_sub_epi16(stp1_17, stp1_18);                                 \
+    stp2_19 = _mm_sub_epi16(stp1_16, stp1_19);                                 \
+    stp2_20 = _mm_sub_epi16(stp1_23, stp1_20);                                 \
+    stp2_21 = _mm_sub_epi16(stp1_22, stp1_21);                                 \
+    stp2_22 = _mm_add_epi16(stp1_22, stp1_21);                                 \
+    stp2_23 = _mm_add_epi16(stp1_23, stp1_20);                                 \
+                                                                               \
+    stp2_24 = _mm_add_epi16(stp1_24, stp1_27);                                 \
+    stp2_25 = _mm_add_epi16(stp1_25, stp1_26);                                 \
+    stp2_26 = _mm_sub_epi16(stp1_25, stp1_26);                                 \
+    stp2_27 = _mm_sub_epi16(stp1_24, stp1_27);                                 \
+    stp2_28 = _mm_sub_epi16(stp1_31, stp1_28);                                 \
+    stp2_29 = _mm_sub_epi16(stp1_30, stp1_29);                                 \
+    stp2_30 = _mm_add_epi16(stp1_29, stp1_30);                                 \
+    stp2_31 = _mm_add_epi16(stp1_28, stp1_31);                                 \
+  }                                                                            \
+                                                                               \
+  /* Stage5 */                                                                 \
+  {                                                                            \
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5);                 \
+    const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5);                 \
+    const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29);             \
+    const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29);             \
+                                                                               \
+    const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28);             \
+    const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28);             \
+    const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27);             \
+    const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27);             \
+                                                                               \
+    const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26);             \
+    const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26);             \
+                                                                               \
+    stp1_0 = stp2_0;                                                           \
+    stp1_1 = stp2_1;                                                           \
+    stp1_2 = stp2_1;                                                           \
+    stp1_3 = stp2_0;                                                           \
+                                                                               \
+    tmp0 = _mm_madd_epi16(lo_6_5, stg4_1);                                     \
+    tmp1 = _mm_madd_epi16(hi_6_5, stg4_1);                                     \
+    tmp2 = _mm_madd_epi16(lo_6_5, stg4_0);                                     \
+    tmp3 = _mm_madd_epi16(hi_6_5, stg4_0);                                     \
+                                                                               \
+    tmp0 = _mm_add_epi32(tmp0, rounding);                                      \
+    tmp1 = _mm_add_epi32(tmp1, rounding);                                      \
+    tmp2 = _mm_add_epi32(tmp2, rounding);                                      \
+    tmp3 = _mm_add_epi32(tmp3, rounding);                                      \
+                                                                               \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);                               \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);                               \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);                               \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);                               \
+                                                                               \
+    stp1_5 = _mm_packs_epi32(tmp0, tmp1);                                      \
+    stp1_6 = _mm_packs_epi32(tmp2, tmp3);                                      \
+                                                                               \
+    stp1_4 = stp2_4;                                                           \
+    stp1_7 = stp2_7;                                                           \
+                                                                               \
+    stp1_8 = _mm_add_epi16(stp2_8, stp2_11);                                   \
+    stp1_9 = _mm_add_epi16(stp2_9, stp2_10);                                   \
+    stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);                                  \
+    stp1_11 = _mm_sub_epi16(stp2_8, stp2_11);                                  \
+    stp1_12 = _mm_sub_epi16(stp2_15, stp2_12);                                 \
+    stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);                                 \
+    stp1_14 = _mm_add_epi16(stp2_14, stp2_13);                                 \
+    stp1_15 = _mm_add_epi16(stp2_15, stp2_12);                                 \
+                                                                               \
+    stp1_16 = stp2_16;                                                         \
+    stp1_17 = stp2_17;                                                         \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4,     \
+                           stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, stp1_19,  \
+                           stp1_28)                                            \
+    MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6,     \
+                           stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, stp1_21,  \
+                           stp1_26)                                            \
+                                                                               \
+    stp1_22 = stp2_22;                                                         \
+    stp1_23 = stp2_23;                                                         \
+    stp1_24 = stp2_24;                                                         \
+    stp1_25 = stp2_25;                                                         \
+    stp1_30 = stp2_30;                                                         \
+    stp1_31 = stp2_31;                                                         \
+  }                                                                            \
+                                                                               \
+  /* Stage6 */                                                                 \
+  {                                                                            \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);             \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13);             \
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12);             \
+    const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12);             \
+                                                                               \
+    stp2_0 = _mm_add_epi16(stp1_0, stp1_7);                                    \
+    stp2_1 = _mm_add_epi16(stp1_1, stp1_6);                                    \
+    stp2_2 = _mm_add_epi16(stp1_2, stp1_5);                                    \
+    stp2_3 = _mm_add_epi16(stp1_3, stp1_4);                                    \
+    stp2_4 = _mm_sub_epi16(stp1_3, stp1_4);                                    \
+    stp2_5 = _mm_sub_epi16(stp1_2, stp1_5);                                    \
+    stp2_6 = _mm_sub_epi16(stp1_1, stp1_6);                                    \
+    stp2_7 = _mm_sub_epi16(stp1_0, stp1_7);                                    \
+                                                                               \
+    stp2_8 = stp1_8;                                                           \
+    stp2_9 = stp1_9;                                                           \
+    stp2_14 = stp1_14;                                                         \
+    stp2_15 = stp1_15;                                                         \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0,     \
+                           stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11,  \
+                           stp2_12)                                            \
+                                                                               \
+    stp2_16 = _mm_add_epi16(stp1_16, stp1_23);                                 \
+    stp2_17 = _mm_add_epi16(stp1_17, stp1_22);                                 \
+    stp2_18 = _mm_add_epi16(stp1_18, stp1_21);                                 \
+    stp2_19 = _mm_add_epi16(stp1_19, stp1_20);                                 \
+    stp2_20 = _mm_sub_epi16(stp1_19, stp1_20);                                 \
+    stp2_21 = _mm_sub_epi16(stp1_18, stp1_21);                                 \
+    stp2_22 = _mm_sub_epi16(stp1_17, stp1_22);                                 \
+    stp2_23 = _mm_sub_epi16(stp1_16, stp1_23);                                 \
+                                                                               \
+    stp2_24 = _mm_sub_epi16(stp1_31, stp1_24);                                 \
+    stp2_25 = _mm_sub_epi16(stp1_30, stp1_25);                                 \
+    stp2_26 = _mm_sub_epi16(stp1_29, stp1_26);                                 \
+    stp2_27 = _mm_sub_epi16(stp1_28, stp1_27);                                 \
+    stp2_28 = _mm_add_epi16(stp1_27, stp1_28);                                 \
+    stp2_29 = _mm_add_epi16(stp1_26, stp1_29);                                 \
+    stp2_30 = _mm_add_epi16(stp1_25, stp1_30);                                 \
+    stp2_31 = _mm_add_epi16(stp1_24, stp1_31);                                 \
+  }                                                                            \
+                                                                               \
+  /* Stage7 */                                                                 \
+  {                                                                            \
+    const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27);             \
+    const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27);             \
+    const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26);             \
+    const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26);             \
+                                                                               \
+    const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25);             \
+    const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25);             \
+    const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24);             \
+    const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24);             \
+                                                                               \
+    stp1_0 = _mm_add_epi16(stp2_0, stp2_15);                                   \
+    stp1_1 = _mm_add_epi16(stp2_1, stp2_14);                                   \
+    stp1_2 = _mm_add_epi16(stp2_2, stp2_13);                                   \
+    stp1_3 = _mm_add_epi16(stp2_3, stp2_12);                                   \
+    stp1_4 = _mm_add_epi16(stp2_4, stp2_11);                                   \
+    stp1_5 = _mm_add_epi16(stp2_5, stp2_10);                                   \
+    stp1_6 = _mm_add_epi16(stp2_6, stp2_9);                                    \
+    stp1_7 = _mm_add_epi16(stp2_7, stp2_8);                                    \
+    stp1_8 = _mm_sub_epi16(stp2_7, stp2_8);                                    \
+    stp1_9 = _mm_sub_epi16(stp2_6, stp2_9);                                    \
+    stp1_10 = _mm_sub_epi16(stp2_5, stp2_10);                                  \
+    stp1_11 = _mm_sub_epi16(stp2_4, stp2_11);                                  \
+    stp1_12 = _mm_sub_epi16(stp2_3, stp2_12);                                  \
+    stp1_13 = _mm_sub_epi16(stp2_2, stp2_13);                                  \
+    stp1_14 = _mm_sub_epi16(stp2_1, stp2_14);                                  \
+    stp1_15 = _mm_sub_epi16(stp2_0, stp2_15);                                  \
+                                                                               \
+    stp1_16 = stp2_16;                                                         \
+    stp1_17 = stp2_17;                                                         \
+    stp1_18 = stp2_18;                                                         \
+    stp1_19 = stp2_19;                                                         \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0,     \
+                           stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, stp1_21,  \
+                           stp1_26)                                            \
+    MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0,     \
+                           stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, stp1_23,  \
+                           stp1_24)                                            \
+                                                                               \
+    stp1_28 = stp2_28;                                                         \
+    stp1_29 = stp2_29;                                                         \
+    stp1_30 = stp2_30;                                                         \
+    stp1_31 = stp2_31;                                                         \
+  }
+
+#define IDCT32(in0, in1)                                                       \
+  /* Stage1 */                                                                 \
+  {                                                                            \
+    const __m128i lo_1_31 = _mm_unpacklo_epi16((in0)[1], (in1)[15]);           \
+    const __m128i hi_1_31 = _mm_unpackhi_epi16((in0)[1], (in1)[15]);           \
+    const __m128i lo_17_15 = _mm_unpacklo_epi16((in1)[1], (in0)[15]);          \
+    const __m128i hi_17_15 = _mm_unpackhi_epi16((in1)[1], (in0)[15]);          \
+                                                                               \
+    const __m128i lo_9_23 = _mm_unpacklo_epi16((in0)[9], (in1)[7]);            \
+    const __m128i hi_9_23 = _mm_unpackhi_epi16((in0)[9], (in1)[7]);            \
+    const __m128i lo_25_7 = _mm_unpacklo_epi16((in1)[9], (in0)[7]);            \
+    const __m128i hi_25_7 = _mm_unpackhi_epi16((in1)[9], (in0)[7]);            \
+                                                                               \
+    const __m128i lo_5_27 = _mm_unpacklo_epi16((in0)[5], (in1)[11]);           \
+    const __m128i hi_5_27 = _mm_unpackhi_epi16((in0)[5], (in1)[11]);           \
+    const __m128i lo_21_11 = _mm_unpacklo_epi16((in1)[5], (in0)[11]);          \
+    const __m128i hi_21_11 = _mm_unpackhi_epi16((in1)[5], (in0)[11]);          \
+                                                                               \
+    const __m128i lo_13_19 = _mm_unpacklo_epi16((in0)[13], (in1)[3]);          \
+    const __m128i hi_13_19 = _mm_unpackhi_epi16((in0)[13], (in1)[3]);          \
+    const __m128i lo_29_3 = _mm_unpacklo_epi16((in1)[13], (in0)[3]);           \
+    const __m128i hi_29_3 = _mm_unpackhi_epi16((in1)[13], (in0)[3]);           \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_1_31, hi_1_31, lo_17_15, hi_17_15, stg1_0,       \
+                           stg1_1, stg1_2, stg1_3, stp1_16, stp1_31, stp1_17,  \
+                           stp1_30)                                            \
+    MULTIPLICATION_AND_ADD(lo_9_23, hi_9_23, lo_25_7, hi_25_7, stg1_4, stg1_5, \
+                           stg1_6, stg1_7, stp1_18, stp1_29, stp1_19, stp1_28) \
+    MULTIPLICATION_AND_ADD(lo_5_27, hi_5_27, lo_21_11, hi_21_11, stg1_8,       \
+                           stg1_9, stg1_10, stg1_11, stp1_20, stp1_27,         \
+                           stp1_21, stp1_26)                                   \
+    MULTIPLICATION_AND_ADD(lo_13_19, hi_13_19, lo_29_3, hi_29_3, stg1_12,      \
+                           stg1_13, stg1_14, stg1_15, stp1_22, stp1_25,        \
+                           stp1_23, stp1_24)                                   \
+  }                                                                            \
+                                                                               \
+  /* Stage2 */                                                                 \
+  {                                                                            \
+    const __m128i lo_2_30 = _mm_unpacklo_epi16((in0)[2], (in1)[14]);           \
+    const __m128i hi_2_30 = _mm_unpackhi_epi16((in0)[2], (in1)[14]);           \
+    const __m128i lo_18_14 = _mm_unpacklo_epi16((in1)[2], (in0)[14]);          \
+    const __m128i hi_18_14 = _mm_unpackhi_epi16((in1)[2], (in0)[14]);          \
+                                                                               \
+    const __m128i lo_10_22 = _mm_unpacklo_epi16((in0)[10], (in1)[6]);          \
+    const __m128i hi_10_22 = _mm_unpackhi_epi16((in0)[10], (in1)[6]);          \
+    const __m128i lo_26_6 = _mm_unpacklo_epi16((in1)[10], (in0)[6]);           \
+    const __m128i hi_26_6 = _mm_unpackhi_epi16((in1)[10], (in0)[6]);           \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_2_30, hi_2_30, lo_18_14, hi_18_14, stg2_0,       \
+                           stg2_1, stg2_2, stg2_3, stp2_8, stp2_15, stp2_9,    \
+                           stp2_14)                                            \
+    MULTIPLICATION_AND_ADD(lo_10_22, hi_10_22, lo_26_6, hi_26_6, stg2_4,       \
+                           stg2_5, stg2_6, stg2_7, stp2_10, stp2_13, stp2_11,  \
+                           stp2_12)                                            \
+                                                                               \
+    stp2_16 = _mm_add_epi16(stp1_16, stp1_17);                                 \
+    stp2_17 = _mm_sub_epi16(stp1_16, stp1_17);                                 \
+    stp2_18 = _mm_sub_epi16(stp1_19, stp1_18);                                 \
+    stp2_19 = _mm_add_epi16(stp1_19, stp1_18);                                 \
+                                                                               \
+    stp2_20 = _mm_add_epi16(stp1_20, stp1_21);                                 \
+    stp2_21 = _mm_sub_epi16(stp1_20, stp1_21);                                 \
+    stp2_22 = _mm_sub_epi16(stp1_23, stp1_22);                                 \
+    stp2_23 = _mm_add_epi16(stp1_23, stp1_22);                                 \
+                                                                               \
+    stp2_24 = _mm_add_epi16(stp1_24, stp1_25);                                 \
+    stp2_25 = _mm_sub_epi16(stp1_24, stp1_25);                                 \
+    stp2_26 = _mm_sub_epi16(stp1_27, stp1_26);                                 \
+    stp2_27 = _mm_add_epi16(stp1_27, stp1_26);                                 \
+                                                                               \
+    stp2_28 = _mm_add_epi16(stp1_28, stp1_29);                                 \
+    stp2_29 = _mm_sub_epi16(stp1_28, stp1_29);                                 \
+    stp2_30 = _mm_sub_epi16(stp1_31, stp1_30);                                 \
+    stp2_31 = _mm_add_epi16(stp1_31, stp1_30);                                 \
+  }                                                                            \
+                                                                               \
+  /* Stage3 */                                                                 \
+  {                                                                            \
+    const __m128i lo_4_28 = _mm_unpacklo_epi16((in0)[4], (in1)[12]);           \
+    const __m128i hi_4_28 = _mm_unpackhi_epi16((in0)[4], (in1)[12]);           \
+    const __m128i lo_20_12 = _mm_unpacklo_epi16((in1)[4], (in0)[12]);          \
+    const __m128i hi_20_12 = _mm_unpackhi_epi16((in1)[4], (in0)[12]);          \
+                                                                               \
+    const __m128i lo_17_30 = _mm_unpacklo_epi16(stp2_17, stp2_30);             \
+    const __m128i hi_17_30 = _mm_unpackhi_epi16(stp2_17, stp2_30);             \
+    const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29);             \
+    const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29);             \
+                                                                               \
+    const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26);             \
+    const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26);             \
+    const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25);             \
+    const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25);             \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_4_28, hi_4_28, lo_20_12, hi_20_12, stg3_0,       \
+                           stg3_1, stg3_2, stg3_3, stp1_4, stp1_7, stp1_5,     \
+                           stp1_6)                                             \
+                                                                               \
+    stp1_8 = _mm_add_epi16(stp2_8, stp2_9);                                    \
+    stp1_9 = _mm_sub_epi16(stp2_8, stp2_9);                                    \
+    stp1_10 = _mm_sub_epi16(stp2_11, stp2_10);                                 \
+    stp1_11 = _mm_add_epi16(stp2_11, stp2_10);                                 \
+    stp1_12 = _mm_add_epi16(stp2_12, stp2_13);                                 \
+    stp1_13 = _mm_sub_epi16(stp2_12, stp2_13);                                 \
+    stp1_14 = _mm_sub_epi16(stp2_15, stp2_14);                                 \
+    stp1_15 = _mm_add_epi16(stp2_15, stp2_14);                                 \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_17_30, hi_17_30, lo_18_29, hi_18_29, stg3_4,     \
+                           stg3_5, stg3_6, stg3_4, stp1_17, stp1_30, stp1_18,  \
+                           stp1_29)                                            \
+    MULTIPLICATION_AND_ADD(lo_21_26, hi_21_26, lo_22_25, hi_22_25, stg3_8,     \
+                           stg3_9, stg3_10, stg3_8, stp1_21, stp1_26, stp1_22, \
+                           stp1_25)                                            \
+                                                                               \
+    stp1_16 = stp2_16;                                                         \
+    stp1_31 = stp2_31;                                                         \
+    stp1_19 = stp2_19;                                                         \
+    stp1_20 = stp2_20;                                                         \
+    stp1_23 = stp2_23;                                                         \
+    stp1_24 = stp2_24;                                                         \
+    stp1_27 = stp2_27;                                                         \
+    stp1_28 = stp2_28;                                                         \
+  }                                                                            \
+                                                                               \
+  /* Stage4 */                                                                 \
+  {                                                                            \
+    const __m128i lo_0_16 = _mm_unpacklo_epi16((in0)[0], (in1)[0]);            \
+    const __m128i hi_0_16 = _mm_unpackhi_epi16((in0)[0], (in1)[0]);            \
+    const __m128i lo_8_24 = _mm_unpacklo_epi16((in0)[8], (in1)[8]);            \
+    const __m128i hi_8_24 = _mm_unpackhi_epi16((in0)[8], (in1)[8]);            \
+                                                                               \
+    const __m128i lo_9_14 = _mm_unpacklo_epi16(stp1_9, stp1_14);               \
+    const __m128i hi_9_14 = _mm_unpackhi_epi16(stp1_9, stp1_14);               \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);             \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13);             \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_0_16, hi_0_16, lo_8_24, hi_8_24, stg4_0, stg4_1, \
+                           stg4_2, stg4_3, stp2_0, stp2_1, stp2_2, stp2_3)     \
+                                                                               \
+    stp2_4 = _mm_add_epi16(stp1_4, stp1_5);                                    \
+    stp2_5 = _mm_sub_epi16(stp1_4, stp1_5);                                    \
+    stp2_6 = _mm_sub_epi16(stp1_7, stp1_6);                                    \
+    stp2_7 = _mm_add_epi16(stp1_7, stp1_6);                                    \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_9_14, hi_9_14, lo_10_13, hi_10_13, stg4_4,       \
+                           stg4_5, stg4_6, stg4_4, stp2_9, stp2_14, stp2_10,   \
+                           stp2_13)                                            \
+                                                                               \
+    stp2_8 = stp1_8;                                                           \
+    stp2_15 = stp1_15;                                                         \
+    stp2_11 = stp1_11;                                                         \
+    stp2_12 = stp1_12;                                                         \
+                                                                               \
+    stp2_16 = _mm_add_epi16(stp1_16, stp1_19);                                 \
+    stp2_17 = _mm_add_epi16(stp1_17, stp1_18);                                 \
+    stp2_18 = _mm_sub_epi16(stp1_17, stp1_18);                                 \
+    stp2_19 = _mm_sub_epi16(stp1_16, stp1_19);                                 \
+    stp2_20 = _mm_sub_epi16(stp1_23, stp1_20);                                 \
+    stp2_21 = _mm_sub_epi16(stp1_22, stp1_21);                                 \
+    stp2_22 = _mm_add_epi16(stp1_22, stp1_21);                                 \
+    stp2_23 = _mm_add_epi16(stp1_23, stp1_20);                                 \
+                                                                               \
+    stp2_24 = _mm_add_epi16(stp1_24, stp1_27);                                 \
+    stp2_25 = _mm_add_epi16(stp1_25, stp1_26);                                 \
+    stp2_26 = _mm_sub_epi16(stp1_25, stp1_26);                                 \
+    stp2_27 = _mm_sub_epi16(stp1_24, stp1_27);                                 \
+    stp2_28 = _mm_sub_epi16(stp1_31, stp1_28);                                 \
+    stp2_29 = _mm_sub_epi16(stp1_30, stp1_29);                                 \
+    stp2_30 = _mm_add_epi16(stp1_29, stp1_30);                                 \
+    stp2_31 = _mm_add_epi16(stp1_28, stp1_31);                                 \
+  }                                                                            \
+                                                                               \
+  /* Stage5 */                                                                 \
+  {                                                                            \
+    const __m128i lo_6_5 = _mm_unpacklo_epi16(stp2_6, stp2_5);                 \
+    const __m128i hi_6_5 = _mm_unpackhi_epi16(stp2_6, stp2_5);                 \
+    const __m128i lo_18_29 = _mm_unpacklo_epi16(stp2_18, stp2_29);             \
+    const __m128i hi_18_29 = _mm_unpackhi_epi16(stp2_18, stp2_29);             \
+                                                                               \
+    const __m128i lo_19_28 = _mm_unpacklo_epi16(stp2_19, stp2_28);             \
+    const __m128i hi_19_28 = _mm_unpackhi_epi16(stp2_19, stp2_28);             \
+    const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27);             \
+    const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27);             \
+                                                                               \
+    const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26);             \
+    const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26);             \
+                                                                               \
+    stp1_0 = _mm_add_epi16(stp2_0, stp2_3);                                    \
+    stp1_1 = _mm_add_epi16(stp2_1, stp2_2);                                    \
+    stp1_2 = _mm_sub_epi16(stp2_1, stp2_2);                                    \
+    stp1_3 = _mm_sub_epi16(stp2_0, stp2_3);                                    \
+                                                                               \
+    tmp0 = _mm_madd_epi16(lo_6_5, stg4_1);                                     \
+    tmp1 = _mm_madd_epi16(hi_6_5, stg4_1);                                     \
+    tmp2 = _mm_madd_epi16(lo_6_5, stg4_0);                                     \
+    tmp3 = _mm_madd_epi16(hi_6_5, stg4_0);                                     \
+                                                                               \
+    tmp0 = _mm_add_epi32(tmp0, rounding);                                      \
+    tmp1 = _mm_add_epi32(tmp1, rounding);                                      \
+    tmp2 = _mm_add_epi32(tmp2, rounding);                                      \
+    tmp3 = _mm_add_epi32(tmp3, rounding);                                      \
+                                                                               \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS);                               \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);                               \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);                               \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);                               \
+                                                                               \
+    stp1_5 = _mm_packs_epi32(tmp0, tmp1);                                      \
+    stp1_6 = _mm_packs_epi32(tmp2, tmp3);                                      \
+                                                                               \
+    stp1_4 = stp2_4;                                                           \
+    stp1_7 = stp2_7;                                                           \
+                                                                               \
+    stp1_8 = _mm_add_epi16(stp2_8, stp2_11);                                   \
+    stp1_9 = _mm_add_epi16(stp2_9, stp2_10);                                   \
+    stp1_10 = _mm_sub_epi16(stp2_9, stp2_10);                                  \
+    stp1_11 = _mm_sub_epi16(stp2_8, stp2_11);                                  \
+    stp1_12 = _mm_sub_epi16(stp2_15, stp2_12);                                 \
+    stp1_13 = _mm_sub_epi16(stp2_14, stp2_13);                                 \
+    stp1_14 = _mm_add_epi16(stp2_14, stp2_13);                                 \
+    stp1_15 = _mm_add_epi16(stp2_15, stp2_12);                                 \
+                                                                               \
+    stp1_16 = stp2_16;                                                         \
+    stp1_17 = stp2_17;                                                         \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_18_29, hi_18_29, lo_19_28, hi_19_28, stg4_4,     \
+                           stg4_5, stg4_4, stg4_5, stp1_18, stp1_29, stp1_19,  \
+                           stp1_28)                                            \
+    MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg4_6,     \
+                           stg4_4, stg4_6, stg4_4, stp1_20, stp1_27, stp1_21,  \
+                           stp1_26)                                            \
+                                                                               \
+    stp1_22 = stp2_22;                                                         \
+    stp1_23 = stp2_23;                                                         \
+    stp1_24 = stp2_24;                                                         \
+    stp1_25 = stp2_25;                                                         \
+    stp1_30 = stp2_30;                                                         \
+    stp1_31 = stp2_31;                                                         \
+  }                                                                            \
+                                                                               \
+  /* Stage6 */                                                                 \
+  {                                                                            \
+    const __m128i lo_10_13 = _mm_unpacklo_epi16(stp1_10, stp1_13);             \
+    const __m128i hi_10_13 = _mm_unpackhi_epi16(stp1_10, stp1_13);             \
+    const __m128i lo_11_12 = _mm_unpacklo_epi16(stp1_11, stp1_12);             \
+    const __m128i hi_11_12 = _mm_unpackhi_epi16(stp1_11, stp1_12);             \
+                                                                               \
+    stp2_0 = _mm_add_epi16(stp1_0, stp1_7);                                    \
+    stp2_1 = _mm_add_epi16(stp1_1, stp1_6);                                    \
+    stp2_2 = _mm_add_epi16(stp1_2, stp1_5);                                    \
+    stp2_3 = _mm_add_epi16(stp1_3, stp1_4);                                    \
+    stp2_4 = _mm_sub_epi16(stp1_3, stp1_4);                                    \
+    stp2_5 = _mm_sub_epi16(stp1_2, stp1_5);                                    \
+    stp2_6 = _mm_sub_epi16(stp1_1, stp1_6);                                    \
+    stp2_7 = _mm_sub_epi16(stp1_0, stp1_7);                                    \
+                                                                               \
+    stp2_8 = stp1_8;                                                           \
+    stp2_9 = stp1_9;                                                           \
+    stp2_14 = stp1_14;                                                         \
+    stp2_15 = stp1_15;                                                         \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_10_13, hi_10_13, lo_11_12, hi_11_12, stg6_0,     \
+                           stg4_0, stg6_0, stg4_0, stp2_10, stp2_13, stp2_11,  \
+                           stp2_12)                                            \
+                                                                               \
+    stp2_16 = _mm_add_epi16(stp1_16, stp1_23);                                 \
+    stp2_17 = _mm_add_epi16(stp1_17, stp1_22);                                 \
+    stp2_18 = _mm_add_epi16(stp1_18, stp1_21);                                 \
+    stp2_19 = _mm_add_epi16(stp1_19, stp1_20);                                 \
+    stp2_20 = _mm_sub_epi16(stp1_19, stp1_20);                                 \
+    stp2_21 = _mm_sub_epi16(stp1_18, stp1_21);                                 \
+    stp2_22 = _mm_sub_epi16(stp1_17, stp1_22);                                 \
+    stp2_23 = _mm_sub_epi16(stp1_16, stp1_23);                                 \
+                                                                               \
+    stp2_24 = _mm_sub_epi16(stp1_31, stp1_24);                                 \
+    stp2_25 = _mm_sub_epi16(stp1_30, stp1_25);                                 \
+    stp2_26 = _mm_sub_epi16(stp1_29, stp1_26);                                 \
+    stp2_27 = _mm_sub_epi16(stp1_28, stp1_27);                                 \
+    stp2_28 = _mm_add_epi16(stp1_27, stp1_28);                                 \
+    stp2_29 = _mm_add_epi16(stp1_26, stp1_29);                                 \
+    stp2_30 = _mm_add_epi16(stp1_25, stp1_30);                                 \
+    stp2_31 = _mm_add_epi16(stp1_24, stp1_31);                                 \
+  }                                                                            \
+                                                                               \
+  /* Stage7 */                                                                 \
+  {                                                                            \
+    const __m128i lo_20_27 = _mm_unpacklo_epi16(stp2_20, stp2_27);             \
+    const __m128i hi_20_27 = _mm_unpackhi_epi16(stp2_20, stp2_27);             \
+    const __m128i lo_21_26 = _mm_unpacklo_epi16(stp2_21, stp2_26);             \
+    const __m128i hi_21_26 = _mm_unpackhi_epi16(stp2_21, stp2_26);             \
+                                                                               \
+    const __m128i lo_22_25 = _mm_unpacklo_epi16(stp2_22, stp2_25);             \
+    const __m128i hi_22_25 = _mm_unpackhi_epi16(stp2_22, stp2_25);             \
+    const __m128i lo_23_24 = _mm_unpacklo_epi16(stp2_23, stp2_24);             \
+    const __m128i hi_23_24 = _mm_unpackhi_epi16(stp2_23, stp2_24);             \
+                                                                               \
+    stp1_0 = _mm_add_epi16(stp2_0, stp2_15);                                   \
+    stp1_1 = _mm_add_epi16(stp2_1, stp2_14);                                   \
+    stp1_2 = _mm_add_epi16(stp2_2, stp2_13);                                   \
+    stp1_3 = _mm_add_epi16(stp2_3, stp2_12);                                   \
+    stp1_4 = _mm_add_epi16(stp2_4, stp2_11);                                   \
+    stp1_5 = _mm_add_epi16(stp2_5, stp2_10);                                   \
+    stp1_6 = _mm_add_epi16(stp2_6, stp2_9);                                    \
+    stp1_7 = _mm_add_epi16(stp2_7, stp2_8);                                    \
+    stp1_8 = _mm_sub_epi16(stp2_7, stp2_8);                                    \
+    stp1_9 = _mm_sub_epi16(stp2_6, stp2_9);                                    \
+    stp1_10 = _mm_sub_epi16(stp2_5, stp2_10);                                  \
+    stp1_11 = _mm_sub_epi16(stp2_4, stp2_11);                                  \
+    stp1_12 = _mm_sub_epi16(stp2_3, stp2_12);                                  \
+    stp1_13 = _mm_sub_epi16(stp2_2, stp2_13);                                  \
+    stp1_14 = _mm_sub_epi16(stp2_1, stp2_14);                                  \
+    stp1_15 = _mm_sub_epi16(stp2_0, stp2_15);                                  \
+                                                                               \
+    stp1_16 = stp2_16;                                                         \
+    stp1_17 = stp2_17;                                                         \
+    stp1_18 = stp2_18;                                                         \
+    stp1_19 = stp2_19;                                                         \
+                                                                               \
+    MULTIPLICATION_AND_ADD(lo_20_27, hi_20_27, lo_21_26, hi_21_26, stg6_0,     \
+                           stg4_0, stg6_0, stg4_0, stp1_20, stp1_27, stp1_21,  \
+                           stp1_26)                                            \
+    MULTIPLICATION_AND_ADD(lo_22_25, hi_22_25, lo_23_24, hi_23_24, stg6_0,     \
+                           stg4_0, stg6_0, stg4_0, stp1_22, stp1_25, stp1_23,  \
+                           stp1_24)                                            \
+                                                                               \
+    stp1_28 = stp2_28;                                                         \
+    stp1_29 = stp2_29;                                                         \
+    stp1_30 = stp2_30;                                                         \
+    stp1_31 = stp2_31;                                                         \
+  }
+
+// Only upper-left 8x8 has non-zero coeff
+void aom_idct32x32_34_add_sse2(const tran_low_t *input, uint8_t *dest,
+                               int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+
+  // idct constants for each stage
+  const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[32], col[32];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+      stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+      stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23,
+      stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+      stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
+      stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23,
+      stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  // Load input data. Only need to load the top left 8x8 block.
+  in[0] = load_input_data(input);
+  in[1] = load_input_data(input + 32);
+  in[2] = load_input_data(input + 64);
+  in[3] = load_input_data(input + 96);
+  in[4] = load_input_data(input + 128);
+  in[5] = load_input_data(input + 160);
+  in[6] = load_input_data(input + 192);
+  in[7] = load_input_data(input + 224);
+
+  for (i = 8; i < 32; ++i) {
+    in[i] = _mm_setzero_si128();
+  }
+
+  array_transpose_8x8(in, in);
+  // TODO(hkuang): Following transposes are unnecessary. But remove them will
+  // lead to performance drop on some devices.
+  array_transpose_8x8(in + 8, in + 8);
+  array_transpose_8x8(in + 16, in + 16);
+  array_transpose_8x8(in + 24, in + 24);
+
+  IDCT32_34
+
+  // 1_D: Store 32 intermediate results for each 8x32 block.
+  col[0] = _mm_add_epi16(stp1_0, stp1_31);
+  col[1] = _mm_add_epi16(stp1_1, stp1_30);
+  col[2] = _mm_add_epi16(stp1_2, stp1_29);
+  col[3] = _mm_add_epi16(stp1_3, stp1_28);
+  col[4] = _mm_add_epi16(stp1_4, stp1_27);
+  col[5] = _mm_add_epi16(stp1_5, stp1_26);
+  col[6] = _mm_add_epi16(stp1_6, stp1_25);
+  col[7] = _mm_add_epi16(stp1_7, stp1_24);
+  col[8] = _mm_add_epi16(stp1_8, stp1_23);
+  col[9] = _mm_add_epi16(stp1_9, stp1_22);
+  col[10] = _mm_add_epi16(stp1_10, stp1_21);
+  col[11] = _mm_add_epi16(stp1_11, stp1_20);
+  col[12] = _mm_add_epi16(stp1_12, stp1_19);
+  col[13] = _mm_add_epi16(stp1_13, stp1_18);
+  col[14] = _mm_add_epi16(stp1_14, stp1_17);
+  col[15] = _mm_add_epi16(stp1_15, stp1_16);
+  col[16] = _mm_sub_epi16(stp1_15, stp1_16);
+  col[17] = _mm_sub_epi16(stp1_14, stp1_17);
+  col[18] = _mm_sub_epi16(stp1_13, stp1_18);
+  col[19] = _mm_sub_epi16(stp1_12, stp1_19);
+  col[20] = _mm_sub_epi16(stp1_11, stp1_20);
+  col[21] = _mm_sub_epi16(stp1_10, stp1_21);
+  col[22] = _mm_sub_epi16(stp1_9, stp1_22);
+  col[23] = _mm_sub_epi16(stp1_8, stp1_23);
+  col[24] = _mm_sub_epi16(stp1_7, stp1_24);
+  col[25] = _mm_sub_epi16(stp1_6, stp1_25);
+  col[26] = _mm_sub_epi16(stp1_5, stp1_26);
+  col[27] = _mm_sub_epi16(stp1_4, stp1_27);
+  col[28] = _mm_sub_epi16(stp1_3, stp1_28);
+  col[29] = _mm_sub_epi16(stp1_2, stp1_29);
+  col[30] = _mm_sub_epi16(stp1_1, stp1_30);
+  col[31] = _mm_sub_epi16(stp1_0, stp1_31);
+  for (i = 0; i < 4; i++) {
+    int j;
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(col + i * 8, in);
+    IDCT32_34
+
+    // 2_D: Calculate the results and store them to destination.
+    in[0] = _mm_add_epi16(stp1_0, stp1_31);
+    in[1] = _mm_add_epi16(stp1_1, stp1_30);
+    in[2] = _mm_add_epi16(stp1_2, stp1_29);
+    in[3] = _mm_add_epi16(stp1_3, stp1_28);
+    in[4] = _mm_add_epi16(stp1_4, stp1_27);
+    in[5] = _mm_add_epi16(stp1_5, stp1_26);
+    in[6] = _mm_add_epi16(stp1_6, stp1_25);
+    in[7] = _mm_add_epi16(stp1_7, stp1_24);
+    in[8] = _mm_add_epi16(stp1_8, stp1_23);
+    in[9] = _mm_add_epi16(stp1_9, stp1_22);
+    in[10] = _mm_add_epi16(stp1_10, stp1_21);
+    in[11] = _mm_add_epi16(stp1_11, stp1_20);
+    in[12] = _mm_add_epi16(stp1_12, stp1_19);
+    in[13] = _mm_add_epi16(stp1_13, stp1_18);
+    in[14] = _mm_add_epi16(stp1_14, stp1_17);
+    in[15] = _mm_add_epi16(stp1_15, stp1_16);
+    in[16] = _mm_sub_epi16(stp1_15, stp1_16);
+    in[17] = _mm_sub_epi16(stp1_14, stp1_17);
+    in[18] = _mm_sub_epi16(stp1_13, stp1_18);
+    in[19] = _mm_sub_epi16(stp1_12, stp1_19);
+    in[20] = _mm_sub_epi16(stp1_11, stp1_20);
+    in[21] = _mm_sub_epi16(stp1_10, stp1_21);
+    in[22] = _mm_sub_epi16(stp1_9, stp1_22);
+    in[23] = _mm_sub_epi16(stp1_8, stp1_23);
+    in[24] = _mm_sub_epi16(stp1_7, stp1_24);
+    in[25] = _mm_sub_epi16(stp1_6, stp1_25);
+    in[26] = _mm_sub_epi16(stp1_5, stp1_26);
+    in[27] = _mm_sub_epi16(stp1_4, stp1_27);
+    in[28] = _mm_sub_epi16(stp1_3, stp1_28);
+    in[29] = _mm_sub_epi16(stp1_2, stp1_29);
+    in[30] = _mm_sub_epi16(stp1_1, stp1_30);
+    in[31] = _mm_sub_epi16(stp1_0, stp1_31);
+
+    for (j = 0; j < 32; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void aom_idct32x32_1024_add_sse2(const tran_low_t *input, uint8_t *dest,
+                                 int stride) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+
+  // idct constants for each stage
+  const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in[32], col[128], zero_idx[16];
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+      stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+      stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23,
+      stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+      stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
+      stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23,
+      stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i, j, i32;
+
+  for (i = 0; i < 4; i++) {
+    i32 = (i << 5);
+    // First 1-D idct
+    // Load input data.
+    LOAD_DQCOEFF(in[0], input);
+    LOAD_DQCOEFF(in[8], input);
+    LOAD_DQCOEFF(in[16], input);
+    LOAD_DQCOEFF(in[24], input);
+    LOAD_DQCOEFF(in[1], input);
+    LOAD_DQCOEFF(in[9], input);
+    LOAD_DQCOEFF(in[17], input);
+    LOAD_DQCOEFF(in[25], input);
+    LOAD_DQCOEFF(in[2], input);
+    LOAD_DQCOEFF(in[10], input);
+    LOAD_DQCOEFF(in[18], input);
+    LOAD_DQCOEFF(in[26], input);
+    LOAD_DQCOEFF(in[3], input);
+    LOAD_DQCOEFF(in[11], input);
+    LOAD_DQCOEFF(in[19], input);
+    LOAD_DQCOEFF(in[27], input);
+
+    LOAD_DQCOEFF(in[4], input);
+    LOAD_DQCOEFF(in[12], input);
+    LOAD_DQCOEFF(in[20], input);
+    LOAD_DQCOEFF(in[28], input);
+    LOAD_DQCOEFF(in[5], input);
+    LOAD_DQCOEFF(in[13], input);
+    LOAD_DQCOEFF(in[21], input);
+    LOAD_DQCOEFF(in[29], input);
+    LOAD_DQCOEFF(in[6], input);
+    LOAD_DQCOEFF(in[14], input);
+    LOAD_DQCOEFF(in[22], input);
+    LOAD_DQCOEFF(in[30], input);
+    LOAD_DQCOEFF(in[7], input);
+    LOAD_DQCOEFF(in[15], input);
+    LOAD_DQCOEFF(in[23], input);
+    LOAD_DQCOEFF(in[31], input);
+
+    // checking if all entries are zero
+    zero_idx[0] = _mm_or_si128(in[0], in[1]);
+    zero_idx[1] = _mm_or_si128(in[2], in[3]);
+    zero_idx[2] = _mm_or_si128(in[4], in[5]);
+    zero_idx[3] = _mm_or_si128(in[6], in[7]);
+    zero_idx[4] = _mm_or_si128(in[8], in[9]);
+    zero_idx[5] = _mm_or_si128(in[10], in[11]);
+    zero_idx[6] = _mm_or_si128(in[12], in[13]);
+    zero_idx[7] = _mm_or_si128(in[14], in[15]);
+    zero_idx[8] = _mm_or_si128(in[16], in[17]);
+    zero_idx[9] = _mm_or_si128(in[18], in[19]);
+    zero_idx[10] = _mm_or_si128(in[20], in[21]);
+    zero_idx[11] = _mm_or_si128(in[22], in[23]);
+    zero_idx[12] = _mm_or_si128(in[24], in[25]);
+    zero_idx[13] = _mm_or_si128(in[26], in[27]);
+    zero_idx[14] = _mm_or_si128(in[28], in[29]);
+    zero_idx[15] = _mm_or_si128(in[30], in[31]);
+
+    zero_idx[0] = _mm_or_si128(zero_idx[0], zero_idx[1]);
+    zero_idx[1] = _mm_or_si128(zero_idx[2], zero_idx[3]);
+    zero_idx[2] = _mm_or_si128(zero_idx[4], zero_idx[5]);
+    zero_idx[3] = _mm_or_si128(zero_idx[6], zero_idx[7]);
+    zero_idx[4] = _mm_or_si128(zero_idx[8], zero_idx[9]);
+    zero_idx[5] = _mm_or_si128(zero_idx[10], zero_idx[11]);
+    zero_idx[6] = _mm_or_si128(zero_idx[12], zero_idx[13]);
+    zero_idx[7] = _mm_or_si128(zero_idx[14], zero_idx[15]);
+
+    zero_idx[8] = _mm_or_si128(zero_idx[0], zero_idx[1]);
+    zero_idx[9] = _mm_or_si128(zero_idx[2], zero_idx[3]);
+    zero_idx[10] = _mm_or_si128(zero_idx[4], zero_idx[5]);
+    zero_idx[11] = _mm_or_si128(zero_idx[6], zero_idx[7]);
+    zero_idx[12] = _mm_or_si128(zero_idx[8], zero_idx[9]);
+    zero_idx[13] = _mm_or_si128(zero_idx[10], zero_idx[11]);
+    zero_idx[14] = _mm_or_si128(zero_idx[12], zero_idx[13]);
+
+    if (_mm_movemask_epi8(_mm_cmpeq_epi32(zero_idx[14], zero)) == 0xFFFF) {
+      col[i32 + 0] = _mm_setzero_si128();
+      col[i32 + 1] = _mm_setzero_si128();
+      col[i32 + 2] = _mm_setzero_si128();
+      col[i32 + 3] = _mm_setzero_si128();
+      col[i32 + 4] = _mm_setzero_si128();
+      col[i32 + 5] = _mm_setzero_si128();
+      col[i32 + 6] = _mm_setzero_si128();
+      col[i32 + 7] = _mm_setzero_si128();
+      col[i32 + 8] = _mm_setzero_si128();
+      col[i32 + 9] = _mm_setzero_si128();
+      col[i32 + 10] = _mm_setzero_si128();
+      col[i32 + 11] = _mm_setzero_si128();
+      col[i32 + 12] = _mm_setzero_si128();
+      col[i32 + 13] = _mm_setzero_si128();
+      col[i32 + 14] = _mm_setzero_si128();
+      col[i32 + 15] = _mm_setzero_si128();
+      col[i32 + 16] = _mm_setzero_si128();
+      col[i32 + 17] = _mm_setzero_si128();
+      col[i32 + 18] = _mm_setzero_si128();
+      col[i32 + 19] = _mm_setzero_si128();
+      col[i32 + 20] = _mm_setzero_si128();
+      col[i32 + 21] = _mm_setzero_si128();
+      col[i32 + 22] = _mm_setzero_si128();
+      col[i32 + 23] = _mm_setzero_si128();
+      col[i32 + 24] = _mm_setzero_si128();
+      col[i32 + 25] = _mm_setzero_si128();
+      col[i32 + 26] = _mm_setzero_si128();
+      col[i32 + 27] = _mm_setzero_si128();
+      col[i32 + 28] = _mm_setzero_si128();
+      col[i32 + 29] = _mm_setzero_si128();
+      col[i32 + 30] = _mm_setzero_si128();
+      col[i32 + 31] = _mm_setzero_si128();
+      continue;
+    }
+
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(in, in);
+    array_transpose_8x8(in + 8, in + 8);
+    array_transpose_8x8(in + 16, in + 16);
+    array_transpose_8x8(in + 24, in + 24);
+
+    IDCT32(in, in + 16)
+
+    // 1_D: Store 32 intermediate results for each 8x32 block.
+    col[i32 + 0] = _mm_add_epi16(stp1_0, stp1_31);
+    col[i32 + 1] = _mm_add_epi16(stp1_1, stp1_30);
+    col[i32 + 2] = _mm_add_epi16(stp1_2, stp1_29);
+    col[i32 + 3] = _mm_add_epi16(stp1_3, stp1_28);
+    col[i32 + 4] = _mm_add_epi16(stp1_4, stp1_27);
+    col[i32 + 5] = _mm_add_epi16(stp1_5, stp1_26);
+    col[i32 + 6] = _mm_add_epi16(stp1_6, stp1_25);
+    col[i32 + 7] = _mm_add_epi16(stp1_7, stp1_24);
+    col[i32 + 8] = _mm_add_epi16(stp1_8, stp1_23);
+    col[i32 + 9] = _mm_add_epi16(stp1_9, stp1_22);
+    col[i32 + 10] = _mm_add_epi16(stp1_10, stp1_21);
+    col[i32 + 11] = _mm_add_epi16(stp1_11, stp1_20);
+    col[i32 + 12] = _mm_add_epi16(stp1_12, stp1_19);
+    col[i32 + 13] = _mm_add_epi16(stp1_13, stp1_18);
+    col[i32 + 14] = _mm_add_epi16(stp1_14, stp1_17);
+    col[i32 + 15] = _mm_add_epi16(stp1_15, stp1_16);
+    col[i32 + 16] = _mm_sub_epi16(stp1_15, stp1_16);
+    col[i32 + 17] = _mm_sub_epi16(stp1_14, stp1_17);
+    col[i32 + 18] = _mm_sub_epi16(stp1_13, stp1_18);
+    col[i32 + 19] = _mm_sub_epi16(stp1_12, stp1_19);
+    col[i32 + 20] = _mm_sub_epi16(stp1_11, stp1_20);
+    col[i32 + 21] = _mm_sub_epi16(stp1_10, stp1_21);
+    col[i32 + 22] = _mm_sub_epi16(stp1_9, stp1_22);
+    col[i32 + 23] = _mm_sub_epi16(stp1_8, stp1_23);
+    col[i32 + 24] = _mm_sub_epi16(stp1_7, stp1_24);
+    col[i32 + 25] = _mm_sub_epi16(stp1_6, stp1_25);
+    col[i32 + 26] = _mm_sub_epi16(stp1_5, stp1_26);
+    col[i32 + 27] = _mm_sub_epi16(stp1_4, stp1_27);
+    col[i32 + 28] = _mm_sub_epi16(stp1_3, stp1_28);
+    col[i32 + 29] = _mm_sub_epi16(stp1_2, stp1_29);
+    col[i32 + 30] = _mm_sub_epi16(stp1_1, stp1_30);
+    col[i32 + 31] = _mm_sub_epi16(stp1_0, stp1_31);
+  }
+  for (i = 0; i < 4; i++) {
+    // Second 1-D idct
+    j = i << 3;
+
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(col + j, in);
+    array_transpose_8x8(col + j + 32, in + 8);
+    array_transpose_8x8(col + j + 64, in + 16);
+    array_transpose_8x8(col + j + 96, in + 24);
+
+    IDCT32(in, in + 16)
+
+    // 2_D: Calculate the results and store them to destination.
+    in[0] = _mm_add_epi16(stp1_0, stp1_31);
+    in[1] = _mm_add_epi16(stp1_1, stp1_30);
+    in[2] = _mm_add_epi16(stp1_2, stp1_29);
+    in[3] = _mm_add_epi16(stp1_3, stp1_28);
+    in[4] = _mm_add_epi16(stp1_4, stp1_27);
+    in[5] = _mm_add_epi16(stp1_5, stp1_26);
+    in[6] = _mm_add_epi16(stp1_6, stp1_25);
+    in[7] = _mm_add_epi16(stp1_7, stp1_24);
+    in[8] = _mm_add_epi16(stp1_8, stp1_23);
+    in[9] = _mm_add_epi16(stp1_9, stp1_22);
+    in[10] = _mm_add_epi16(stp1_10, stp1_21);
+    in[11] = _mm_add_epi16(stp1_11, stp1_20);
+    in[12] = _mm_add_epi16(stp1_12, stp1_19);
+    in[13] = _mm_add_epi16(stp1_13, stp1_18);
+    in[14] = _mm_add_epi16(stp1_14, stp1_17);
+    in[15] = _mm_add_epi16(stp1_15, stp1_16);
+    in[16] = _mm_sub_epi16(stp1_15, stp1_16);
+    in[17] = _mm_sub_epi16(stp1_14, stp1_17);
+    in[18] = _mm_sub_epi16(stp1_13, stp1_18);
+    in[19] = _mm_sub_epi16(stp1_12, stp1_19);
+    in[20] = _mm_sub_epi16(stp1_11, stp1_20);
+    in[21] = _mm_sub_epi16(stp1_10, stp1_21);
+    in[22] = _mm_sub_epi16(stp1_9, stp1_22);
+    in[23] = _mm_sub_epi16(stp1_8, stp1_23);
+    in[24] = _mm_sub_epi16(stp1_7, stp1_24);
+    in[25] = _mm_sub_epi16(stp1_6, stp1_25);
+    in[26] = _mm_sub_epi16(stp1_5, stp1_26);
+    in[27] = _mm_sub_epi16(stp1_4, stp1_27);
+    in[28] = _mm_sub_epi16(stp1_3, stp1_28);
+    in[29] = _mm_sub_epi16(stp1_2, stp1_29);
+    in[30] = _mm_sub_epi16(stp1_1, stp1_30);
+    in[31] = _mm_sub_epi16(stp1_0, stp1_31);
+
+    for (j = 0; j < 32; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+void aom_idct32x32_1_add_sse2(const tran_low_t *input, uint8_t *dest,
+                              int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a, j;
+
+  a = (int)dct_const_round_shift(input[0] * cospi_16_64);
+  a = (int)dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 6);
+
+  if (a == 0) return;
+
+  dc_value = _mm_set1_epi16(a);
+
+  for (j = 0; j < 32; ++j) {
+    RECON_AND_STORE(dest + 0 + j * stride, dc_value);
+    RECON_AND_STORE(dest + 8 + j * stride, dc_value);
+    RECON_AND_STORE(dest + 16 + j * stride, dc_value);
+    RECON_AND_STORE(dest + 24 + j * stride, dc_value);
+  }
+}
+
+// Apply a 32-element IDCT to 8 columns. This does not do any transposition
+// of its input - the caller is expected to have done that.
+// The input buffers are the top and bottom halves of an 8x32 block.
+void idct32_8col(__m128i *in0, __m128i *in1) {
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  // idct constants for each stage
+  const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
+
+  const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+
+  const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7,
+      stp1_8, stp1_9, stp1_10, stp1_11, stp1_12, stp1_13, stp1_14, stp1_15,
+      stp1_16, stp1_17, stp1_18, stp1_19, stp1_20, stp1_21, stp1_22, stp1_23,
+      stp1_24, stp1_25, stp1_26, stp1_27, stp1_28, stp1_29, stp1_30, stp1_31;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7,
+      stp2_8, stp2_9, stp2_10, stp2_11, stp2_12, stp2_13, stp2_14, stp2_15,
+      stp2_16, stp2_17, stp2_18, stp2_19, stp2_20, stp2_21, stp2_22, stp2_23,
+      stp2_24, stp2_25, stp2_26, stp2_27, stp2_28, stp2_29, stp2_30, stp2_31;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+
+  IDCT32(in0, in1)
+
+  // 2_D: Calculate the results and store them to destination.
+  in0[0] = _mm_add_epi16(stp1_0, stp1_31);
+  in0[1] = _mm_add_epi16(stp1_1, stp1_30);
+  in0[2] = _mm_add_epi16(stp1_2, stp1_29);
+  in0[3] = _mm_add_epi16(stp1_3, stp1_28);
+  in0[4] = _mm_add_epi16(stp1_4, stp1_27);
+  in0[5] = _mm_add_epi16(stp1_5, stp1_26);
+  in0[6] = _mm_add_epi16(stp1_6, stp1_25);
+  in0[7] = _mm_add_epi16(stp1_7, stp1_24);
+  in0[8] = _mm_add_epi16(stp1_8, stp1_23);
+  in0[9] = _mm_add_epi16(stp1_9, stp1_22);
+  in0[10] = _mm_add_epi16(stp1_10, stp1_21);
+  in0[11] = _mm_add_epi16(stp1_11, stp1_20);
+  in0[12] = _mm_add_epi16(stp1_12, stp1_19);
+  in0[13] = _mm_add_epi16(stp1_13, stp1_18);
+  in0[14] = _mm_add_epi16(stp1_14, stp1_17);
+  in0[15] = _mm_add_epi16(stp1_15, stp1_16);
+  in1[0] = _mm_sub_epi16(stp1_15, stp1_16);
+  in1[1] = _mm_sub_epi16(stp1_14, stp1_17);
+  in1[2] = _mm_sub_epi16(stp1_13, stp1_18);
+  in1[3] = _mm_sub_epi16(stp1_12, stp1_19);
+  in1[4] = _mm_sub_epi16(stp1_11, stp1_20);
+  in1[5] = _mm_sub_epi16(stp1_10, stp1_21);
+  in1[6] = _mm_sub_epi16(stp1_9, stp1_22);
+  in1[7] = _mm_sub_epi16(stp1_8, stp1_23);
+  in1[8] = _mm_sub_epi16(stp1_7, stp1_24);
+  in1[9] = _mm_sub_epi16(stp1_6, stp1_25);
+  in1[10] = _mm_sub_epi16(stp1_5, stp1_26);
+  in1[11] = _mm_sub_epi16(stp1_4, stp1_27);
+  in1[12] = _mm_sub_epi16(stp1_3, stp1_28);
+  in1[13] = _mm_sub_epi16(stp1_2, stp1_29);
+  in1[14] = _mm_sub_epi16(stp1_1, stp1_30);
+  in1[15] = _mm_sub_epi16(stp1_0, stp1_31);
+}
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE __m128i clamp_high_sse2(__m128i value, int bd) {
+  __m128i ubounded, retval;
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i max = _mm_subs_epi16(_mm_slli_epi16(one, bd), one);
+  ubounded = _mm_cmpgt_epi16(value, max);
+  retval = _mm_andnot_si128(ubounded, value);
+  ubounded = _mm_and_si128(ubounded, max);
+  retval = _mm_or_si128(retval, ubounded);
+  retval = _mm_and_si128(retval, _mm_cmpgt_epi16(retval, zero));
+  return retval;
+}
+
+void aom_highbd_idct4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest8,
+                                    int stride, int bd) {
+  tran_low_t out[4 * 4];
+  tran_low_t *outptr = out;
+  int i, j;
+  __m128i inptr[4];
+  __m128i sign_bits[2];
+  __m128i temp_mm, min_input, max_input;
+  int test;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+  int optimised_cols = 0;
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i eight = _mm_set1_epi16(8);
+  const __m128i max = _mm_set1_epi16(12043);
+  const __m128i min = _mm_set1_epi16(-12043);
+  // Load input into __m128i
+  inptr[0] = _mm_loadu_si128((const __m128i *)input);
+  inptr[1] = _mm_loadu_si128((const __m128i *)(input + 4));
+  inptr[2] = _mm_loadu_si128((const __m128i *)(input + 8));
+  inptr[3] = _mm_loadu_si128((const __m128i *)(input + 12));
+
+  // Pack to 16 bits
+  inptr[0] = _mm_packs_epi32(inptr[0], inptr[1]);
+  inptr[1] = _mm_packs_epi32(inptr[2], inptr[3]);
+
+  max_input = _mm_max_epi16(inptr[0], inptr[1]);
+  min_input = _mm_min_epi16(inptr[0], inptr[1]);
+  max_input = _mm_cmpgt_epi16(max_input, max);
+  min_input = _mm_cmplt_epi16(min_input, min);
+  temp_mm = _mm_or_si128(max_input, min_input);
+  test = _mm_movemask_epi8(temp_mm);
+
+  if (!test) {
+    // Do the row transform
+    aom_idct4_sse2(inptr);
+
+    // Check the min & max values
+    max_input = _mm_max_epi16(inptr[0], inptr[1]);
+    min_input = _mm_min_epi16(inptr[0], inptr[1]);
+    max_input = _mm_cmpgt_epi16(max_input, max);
+    min_input = _mm_cmplt_epi16(min_input, min);
+    temp_mm = _mm_or_si128(max_input, min_input);
+    test = _mm_movemask_epi8(temp_mm);
+
+    if (test) {
+      array_transpose_4x4(inptr);
+      sign_bits[0] = _mm_cmplt_epi16(inptr[0], zero);
+      sign_bits[1] = _mm_cmplt_epi16(inptr[1], zero);
+      inptr[3] = _mm_unpackhi_epi16(inptr[1], sign_bits[1]);
+      inptr[2] = _mm_unpacklo_epi16(inptr[1], sign_bits[1]);
+      inptr[1] = _mm_unpackhi_epi16(inptr[0], sign_bits[0]);
+      inptr[0] = _mm_unpacklo_epi16(inptr[0], sign_bits[0]);
+      _mm_storeu_si128((__m128i *)outptr, inptr[0]);
+      _mm_storeu_si128((__m128i *)(outptr + 4), inptr[1]);
+      _mm_storeu_si128((__m128i *)(outptr + 8), inptr[2]);
+      _mm_storeu_si128((__m128i *)(outptr + 12), inptr[3]);
+    } else {
+      // Set to use the optimised transform for the column
+      optimised_cols = 1;
+    }
+  } else {
+    // Run the un-optimised row transform
+    for (i = 0; i < 4; ++i) {
+      aom_highbd_idct4_c(input, outptr, bd);
+      input += 4;
+      outptr += 4;
+    }
+  }
+
+  if (optimised_cols) {
+    aom_idct4_sse2(inptr);
+
+    // Final round and shift
+    inptr[0] = _mm_add_epi16(inptr[0], eight);
+    inptr[1] = _mm_add_epi16(inptr[1], eight);
+
+    inptr[0] = _mm_srai_epi16(inptr[0], 4);
+    inptr[1] = _mm_srai_epi16(inptr[1], 4);
+
+    // Reconstruction and Store
+    {
+      __m128i d0 = _mm_loadl_epi64((const __m128i *)dest);
+      __m128i d2 = _mm_loadl_epi64((const __m128i *)(dest + stride * 2));
+      d0 = _mm_unpacklo_epi64(
+          d0, _mm_loadl_epi64((const __m128i *)(dest + stride)));
+      d2 = _mm_unpacklo_epi64(
+          d2, _mm_loadl_epi64((const __m128i *)(dest + stride * 3)));
+      d0 = clamp_high_sse2(_mm_adds_epi16(d0, inptr[0]), bd);
+      d2 = clamp_high_sse2(_mm_adds_epi16(d2, inptr[1]), bd);
+      // store input0
+      _mm_storel_epi64((__m128i *)dest, d0);
+      // store input1
+      d0 = _mm_srli_si128(d0, 8);
+      _mm_storel_epi64((__m128i *)(dest + stride), d0);
+      // store input2
+      _mm_storel_epi64((__m128i *)(dest + stride * 2), d2);
+      // store input3
+      d2 = _mm_srli_si128(d2, 8);
+      _mm_storel_epi64((__m128i *)(dest + stride * 3), d2);
+    }
+  } else {
+    // Run the un-optimised column transform
+    tran_low_t temp_in[4], temp_out[4];
+    // Columns
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j) temp_in[j] = out[j * 4 + i];
+      aom_highbd_idct4_c(temp_in, temp_out, bd);
+      for (j = 0; j < 4; ++j) {
+        dest[j * stride + i] = highbd_clip_pixel_add(
+            dest[j * stride + i], ROUND_POWER_OF_TWO(temp_out[j], 4), bd);
+      }
+    }
+  }
+}
+
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/inv_txfm_sse2.h b/third_party/aom/aom_dsp/x86/inv_txfm_sse2.h
new file mode 100644
index 0000000000..95d246c3c5
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/inv_txfm_sse2.h
@@ -0,0 +1,265 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_X86_INV_TXFM_SSE2_H_
+#define AOM_DSP_X86_INV_TXFM_SSE2_H_
+
+#include <emmintrin.h>  // SSE2
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/inv_txfm.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+
+// perform 8x8 transpose
+static INLINE void array_transpose_4x4(__m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
+  const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
+
+  res[0] = _mm_unpacklo_epi16(tr0_0, tr0_1);
+  res[1] = _mm_unpackhi_epi16(tr0_0, tr0_1);
+}
+
+static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+
+  res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
+  res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
+  res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
+  res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
+  res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
+  res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
+  res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
+  res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
+}
+
+#define TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, \
+                      out2, out3, out4, out5, out6, out7)                 \
+  {                                                                       \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1);                   \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3);                   \
+    const __m128i tr0_2 = _mm_unpackhi_epi16(in0, in1);                   \
+    const __m128i tr0_3 = _mm_unpackhi_epi16(in2, in3);                   \
+    const __m128i tr0_4 = _mm_unpacklo_epi16(in4, in5);                   \
+    const __m128i tr0_5 = _mm_unpacklo_epi16(in6, in7);                   \
+    const __m128i tr0_6 = _mm_unpackhi_epi16(in4, in5);                   \
+    const __m128i tr0_7 = _mm_unpackhi_epi16(in6, in7);                   \
+                                                                          \
+    const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);               \
+    const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);               \
+    const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);               \
+    const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);               \
+    const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);               \
+    const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);               \
+    const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);               \
+    const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);               \
+                                                                          \
+    out0 = _mm_unpacklo_epi64(tr1_0, tr1_4);                              \
+    out1 = _mm_unpackhi_epi64(tr1_0, tr1_4);                              \
+    out2 = _mm_unpacklo_epi64(tr1_2, tr1_6);                              \
+    out3 = _mm_unpackhi_epi64(tr1_2, tr1_6);                              \
+    out4 = _mm_unpacklo_epi64(tr1_1, tr1_5);                              \
+    out5 = _mm_unpackhi_epi64(tr1_1, tr1_5);                              \
+    out6 = _mm_unpacklo_epi64(tr1_3, tr1_7);                              \
+    out7 = _mm_unpackhi_epi64(tr1_3, tr1_7);                              \
+  }
+
+#define TRANSPOSE_8X4(in0, in1, in2, in3, out0, out1)   \
+  {                                                     \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1); \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3); \
+                                                        \
+    in0 = _mm_unpacklo_epi32(tr0_0, tr0_1); /* i1 i0 */ \
+    in1 = _mm_unpackhi_epi32(tr0_0, tr0_1); /* i3 i2 */ \
+  }
+
+static INLINE void array_transpose_4X8(__m128i *in, __m128i *out) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+
+  out[0] = _mm_unpacklo_epi64(tr1_0, tr1_4);
+  out[1] = _mm_unpackhi_epi64(tr1_0, tr1_4);
+  out[2] = _mm_unpacklo_epi64(tr1_2, tr1_6);
+  out[3] = _mm_unpackhi_epi64(tr1_2, tr1_6);
+}
+
+static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
+  __m128i tbuf[8];
+  array_transpose_8x8(res0, res0);
+  array_transpose_8x8(res1, tbuf);
+  array_transpose_8x8(res0 + 8, res1);
+  array_transpose_8x8(res1 + 8, res1 + 8);
+
+  res0[8] = tbuf[0];
+  res0[9] = tbuf[1];
+  res0[10] = tbuf[2];
+  res0[11] = tbuf[3];
+  res0[12] = tbuf[4];
+  res0[13] = tbuf[5];
+  res0[14] = tbuf[6];
+  res0[15] = tbuf[7];
+}
+
+// Function to allow 8 bit optimisations to be used when profile 0 is used with
+// highbitdepth enabled
+static INLINE __m128i load_input_data(const tran_low_t *data) {
+#if CONFIG_HIGHBITDEPTH
+  return octa_set_epi16(data[0], data[1], data[2], data[3], data[4], data[5],
+                        data[6], data[7]);
+#else
+  return _mm_load_si128((const __m128i *)data);
+#endif
+}
+
+static INLINE void load_buffer_8x16(const tran_low_t *input, __m128i *in) {
+  in[0] = load_input_data(input + 0 * 16);
+  in[1] = load_input_data(input + 1 * 16);
+  in[2] = load_input_data(input + 2 * 16);
+  in[3] = load_input_data(input + 3 * 16);
+  in[4] = load_input_data(input + 4 * 16);
+  in[5] = load_input_data(input + 5 * 16);
+  in[6] = load_input_data(input + 6 * 16);
+  in[7] = load_input_data(input + 7 * 16);
+
+  in[8] = load_input_data(input + 8 * 16);
+  in[9] = load_input_data(input + 9 * 16);
+  in[10] = load_input_data(input + 10 * 16);
+  in[11] = load_input_data(input + 11 * 16);
+  in[12] = load_input_data(input + 12 * 16);
+  in[13] = load_input_data(input + 13 * 16);
+  in[14] = load_input_data(input + 14 * 16);
+  in[15] = load_input_data(input + 15 * 16);
+}
+
+#define RECON_AND_STORE(dest, in_x)                  \
+  {                                                  \
+    __m128i d0 = _mm_loadl_epi64((__m128i *)(dest)); \
+    d0 = _mm_unpacklo_epi8(d0, zero);                \
+    d0 = _mm_add_epi16(in_x, d0);                    \
+    d0 = _mm_packus_epi16(d0, d0);                   \
+    _mm_storel_epi64((__m128i *)(dest), d0);         \
+  }
+
+static INLINE void write_buffer_8x16(uint8_t *dest, __m128i *in, int stride) {
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+  in[4] = _mm_adds_epi16(in[4], final_rounding);
+  in[5] = _mm_adds_epi16(in[5], final_rounding);
+  in[6] = _mm_adds_epi16(in[6], final_rounding);
+  in[7] = _mm_adds_epi16(in[7], final_rounding);
+  in[8] = _mm_adds_epi16(in[8], final_rounding);
+  in[9] = _mm_adds_epi16(in[9], final_rounding);
+  in[10] = _mm_adds_epi16(in[10], final_rounding);
+  in[11] = _mm_adds_epi16(in[11], final_rounding);
+  in[12] = _mm_adds_epi16(in[12], final_rounding);
+  in[13] = _mm_adds_epi16(in[13], final_rounding);
+  in[14] = _mm_adds_epi16(in[14], final_rounding);
+  in[15] = _mm_adds_epi16(in[15], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 6);
+  in[1] = _mm_srai_epi16(in[1], 6);
+  in[2] = _mm_srai_epi16(in[2], 6);
+  in[3] = _mm_srai_epi16(in[3], 6);
+  in[4] = _mm_srai_epi16(in[4], 6);
+  in[5] = _mm_srai_epi16(in[5], 6);
+  in[6] = _mm_srai_epi16(in[6], 6);
+  in[7] = _mm_srai_epi16(in[7], 6);
+  in[8] = _mm_srai_epi16(in[8], 6);
+  in[9] = _mm_srai_epi16(in[9], 6);
+  in[10] = _mm_srai_epi16(in[10], 6);
+  in[11] = _mm_srai_epi16(in[11], 6);
+  in[12] = _mm_srai_epi16(in[12], 6);
+  in[13] = _mm_srai_epi16(in[13], 6);
+  in[14] = _mm_srai_epi16(in[14], 6);
+  in[15] = _mm_srai_epi16(in[15], 6);
+
+  RECON_AND_STORE(dest + 0 * stride, in[0]);
+  RECON_AND_STORE(dest + 1 * stride, in[1]);
+  RECON_AND_STORE(dest + 2 * stride, in[2]);
+  RECON_AND_STORE(dest + 3 * stride, in[3]);
+  RECON_AND_STORE(dest + 4 * stride, in[4]);
+  RECON_AND_STORE(dest + 5 * stride, in[5]);
+  RECON_AND_STORE(dest + 6 * stride, in[6]);
+  RECON_AND_STORE(dest + 7 * stride, in[7]);
+  RECON_AND_STORE(dest + 8 * stride, in[8]);
+  RECON_AND_STORE(dest + 9 * stride, in[9]);
+  RECON_AND_STORE(dest + 10 * stride, in[10]);
+  RECON_AND_STORE(dest + 11 * stride, in[11]);
+  RECON_AND_STORE(dest + 12 * stride, in[12]);
+  RECON_AND_STORE(dest + 13 * stride, in[13]);
+  RECON_AND_STORE(dest + 14 * stride, in[14]);
+  RECON_AND_STORE(dest + 15 * stride, in[15]);
+}
+
+#define TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, out0, out1, out2, out3) \
+  {                                                                      \
+    const __m128i tr0_0 = _mm_unpackhi_epi16(tmp0, tmp1);                \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(tmp1, tmp0);                \
+    const __m128i tr0_4 = _mm_unpacklo_epi16(tmp2, tmp3);                \
+    const __m128i tr0_5 = _mm_unpackhi_epi16(tmp3, tmp2);                \
+                                                                         \
+    const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);              \
+    const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);              \
+    const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);              \
+    const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);              \
+                                                                         \
+    out0 = _mm_unpacklo_epi64(tr1_0, tr1_4);                             \
+    out1 = _mm_unpackhi_epi64(tr1_0, tr1_4);                             \
+    out2 = _mm_unpacklo_epi64(tr1_2, tr1_6);                             \
+    out3 = _mm_unpackhi_epi64(tr1_2, tr1_6);                             \
+  }
+
+#define TRANSPOSE_8X8_10(in0, in1, in2, in3, out0, out1) \
+  {                                                      \
+    const __m128i tr0_0 = _mm_unpacklo_epi16(in0, in1);  \
+    const __m128i tr0_1 = _mm_unpacklo_epi16(in2, in3);  \
+    out0 = _mm_unpacklo_epi32(tr0_0, tr0_1);             \
+    out1 = _mm_unpackhi_epi32(tr0_0, tr0_1);             \
+  }
+
+void iadst16_8col(__m128i *in);
+void idct16_8col(__m128i *in);
+void aom_idct4_sse2(__m128i *in);
+void aom_idct8_sse2(__m128i *in);
+void aom_idct16_sse2(__m128i *in0, __m128i *in1);
+void aom_iadst4_sse2(__m128i *in);
+void aom_iadst8_sse2(__m128i *in);
+void aom_iadst16_sse2(__m128i *in0, __m128i *in1);
+void idct32_8col(__m128i *in0, __m128i *in1);
+
+#endif  // AOM_DSP_X86_INV_TXFM_SSE2_H_
diff --git a/third_party/aom/aom_dsp/x86/inv_txfm_ssse3.c b/third_party/aom/aom_dsp/x86/inv_txfm_ssse3.c
new file mode 100644
index 0000000000..9d006797be
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/inv_txfm_ssse3.c
@@ -0,0 +1,1333 @@
+/*
+ *  Copyright (c) 2017 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <tmmintrin.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/x86/inv_txfm_sse2.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+
+void aom_idct8x8_64_add_ssse3(const tran_low_t *input, uint8_t *dest,
+                              int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i stg1_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i stg1_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i stg1_2 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg1_3 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stk2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stk2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i stg2_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  int i;
+
+  // Load input data.
+  in0 = load_input_data(input);
+  in1 = load_input_data(input + 8 * 1);
+  in2 = load_input_data(input + 8 * 2);
+  in3 = load_input_data(input + 8 * 3);
+  in4 = load_input_data(input + 8 * 4);
+  in5 = load_input_data(input + 8 * 5);
+  in6 = load_input_data(input + 8 * 6);
+  in7 = load_input_data(input + 8 * 7);
+
+  // 2-D
+  for (i = 0; i < 2; i++) {
+    // 8x8 Transpose is copied from vpx_fdct8x8_sse2()
+    TRANSPOSE_8X8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                  in4, in5, in6, in7);
+
+    // 4-stage 1D idct8x8
+    {
+      /* Stage1 */
+      {
+        const __m128i lo_17 = _mm_unpacklo_epi16(in1, in7);
+        const __m128i hi_17 = _mm_unpackhi_epi16(in1, in7);
+        const __m128i lo_35 = _mm_unpacklo_epi16(in3, in5);
+        const __m128i hi_35 = _mm_unpackhi_epi16(in3, in5);
+
+        {
+          tmp0 = _mm_madd_epi16(lo_17, stg1_0);
+          tmp1 = _mm_madd_epi16(hi_17, stg1_0);
+          tmp2 = _mm_madd_epi16(lo_17, stg1_1);
+          tmp3 = _mm_madd_epi16(hi_17, stg1_1);
+          tmp4 = _mm_madd_epi16(lo_35, stg1_2);
+          tmp5 = _mm_madd_epi16(hi_35, stg1_2);
+          tmp6 = _mm_madd_epi16(lo_35, stg1_3);
+          tmp7 = _mm_madd_epi16(hi_35, stg1_3);
+
+          tmp0 = _mm_add_epi32(tmp0, rounding);
+          tmp1 = _mm_add_epi32(tmp1, rounding);
+          tmp2 = _mm_add_epi32(tmp2, rounding);
+          tmp3 = _mm_add_epi32(tmp3, rounding);
+          tmp4 = _mm_add_epi32(tmp4, rounding);
+          tmp5 = _mm_add_epi32(tmp5, rounding);
+          tmp6 = _mm_add_epi32(tmp6, rounding);
+          tmp7 = _mm_add_epi32(tmp7, rounding);
+
+          tmp0 = _mm_srai_epi32(tmp0, 14);
+          tmp1 = _mm_srai_epi32(tmp1, 14);
+          tmp2 = _mm_srai_epi32(tmp2, 14);
+          tmp3 = _mm_srai_epi32(tmp3, 14);
+          tmp4 = _mm_srai_epi32(tmp4, 14);
+          tmp5 = _mm_srai_epi32(tmp5, 14);
+          tmp6 = _mm_srai_epi32(tmp6, 14);
+          tmp7 = _mm_srai_epi32(tmp7, 14);
+
+          stp1_4 = _mm_packs_epi32(tmp0, tmp1);
+          stp1_7 = _mm_packs_epi32(tmp2, tmp3);
+          stp1_5 = _mm_packs_epi32(tmp4, tmp5);
+          stp1_6 = _mm_packs_epi32(tmp6, tmp7);
+        }
+      }
+
+      /* Stage2 */
+      {
+        const __m128i lo_26 = _mm_unpacklo_epi16(in2, in6);
+        const __m128i hi_26 = _mm_unpackhi_epi16(in2, in6);
+
+        {
+          tmp0 = _mm_unpacklo_epi16(in0, in4);
+          tmp1 = _mm_unpackhi_epi16(in0, in4);
+
+          tmp2 = _mm_madd_epi16(tmp0, stk2_0);
+          tmp3 = _mm_madd_epi16(tmp1, stk2_0);
+          tmp4 = _mm_madd_epi16(tmp0, stk2_1);
+          tmp5 = _mm_madd_epi16(tmp1, stk2_1);
+
+          tmp2 = _mm_add_epi32(tmp2, rounding);
+          tmp3 = _mm_add_epi32(tmp3, rounding);
+          tmp4 = _mm_add_epi32(tmp4, rounding);
+          tmp5 = _mm_add_epi32(tmp5, rounding);
+
+          tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+          tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+          tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+          tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
+
+          stp2_0 = _mm_packs_epi32(tmp2, tmp3);
+          stp2_1 = _mm_packs_epi32(tmp4, tmp5);
+
+          tmp0 = _mm_madd_epi16(lo_26, stg2_2);
+          tmp1 = _mm_madd_epi16(hi_26, stg2_2);
+          tmp2 = _mm_madd_epi16(lo_26, stg2_3);
+          tmp3 = _mm_madd_epi16(hi_26, stg2_3);
+
+          tmp0 = _mm_add_epi32(tmp0, rounding);
+          tmp1 = _mm_add_epi32(tmp1, rounding);
+          tmp2 = _mm_add_epi32(tmp2, rounding);
+          tmp3 = _mm_add_epi32(tmp3, rounding);
+
+          tmp0 = _mm_srai_epi32(tmp0, 14);
+          tmp1 = _mm_srai_epi32(tmp1, 14);
+          tmp2 = _mm_srai_epi32(tmp2, 14);
+          tmp3 = _mm_srai_epi32(tmp3, 14);
+
+          stp2_2 = _mm_packs_epi32(tmp0, tmp1);
+          stp2_3 = _mm_packs_epi32(tmp2, tmp3);
+        }
+
+        stp2_4 = _mm_add_epi16(stp1_4, stp1_5);
+        stp2_5 = _mm_sub_epi16(stp1_4, stp1_5);
+        stp2_6 = _mm_sub_epi16(stp1_7, stp1_6);
+        stp2_7 = _mm_add_epi16(stp1_7, stp1_6);
+      }
+
+      /* Stage3 */
+      {
+        stp1_0 = _mm_add_epi16(stp2_0, stp2_3);
+        stp1_1 = _mm_add_epi16(stp2_1, stp2_2);
+        stp1_2 = _mm_sub_epi16(stp2_1, stp2_2);
+        stp1_3 = _mm_sub_epi16(stp2_0, stp2_3);
+
+        tmp0 = _mm_unpacklo_epi16(stp2_6, stp2_5);
+        tmp1 = _mm_unpackhi_epi16(stp2_6, stp2_5);
+
+        tmp2 = _mm_madd_epi16(tmp0, stk2_1);
+        tmp3 = _mm_madd_epi16(tmp1, stk2_1);
+        tmp4 = _mm_madd_epi16(tmp0, stk2_0);
+        tmp5 = _mm_madd_epi16(tmp1, stk2_0);
+
+        tmp2 = _mm_add_epi32(tmp2, rounding);
+        tmp3 = _mm_add_epi32(tmp3, rounding);
+        tmp4 = _mm_add_epi32(tmp4, rounding);
+        tmp5 = _mm_add_epi32(tmp5, rounding);
+
+        tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+        tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+        tmp4 = _mm_srai_epi32(tmp4, DCT_CONST_BITS);
+        tmp5 = _mm_srai_epi32(tmp5, DCT_CONST_BITS);
+
+        stp1_5 = _mm_packs_epi32(tmp2, tmp3);
+        stp1_6 = _mm_packs_epi32(tmp4, tmp5);
+      }
+
+      /* Stage4  */
+      in0 = _mm_add_epi16(stp1_0, stp2_7);
+      in1 = _mm_add_epi16(stp1_1, stp1_6);
+      in2 = _mm_add_epi16(stp1_2, stp1_5);
+      in3 = _mm_add_epi16(stp1_3, stp2_4);
+      in4 = _mm_sub_epi16(stp1_3, stp2_4);
+      in5 = _mm_sub_epi16(stp1_2, stp1_5);
+      in6 = _mm_sub_epi16(stp1_1, stp1_6);
+      in7 = _mm_sub_epi16(stp1_0, stp2_7);
+    }
+  }
+
+  // Final rounding and shift
+  in0 = _mm_adds_epi16(in0, final_rounding);
+  in1 = _mm_adds_epi16(in1, final_rounding);
+  in2 = _mm_adds_epi16(in2, final_rounding);
+  in3 = _mm_adds_epi16(in3, final_rounding);
+  in4 = _mm_adds_epi16(in4, final_rounding);
+  in5 = _mm_adds_epi16(in5, final_rounding);
+  in6 = _mm_adds_epi16(in6, final_rounding);
+  in7 = _mm_adds_epi16(in7, final_rounding);
+
+  in0 = _mm_srai_epi16(in0, 5);
+  in1 = _mm_srai_epi16(in1, 5);
+  in2 = _mm_srai_epi16(in2, 5);
+  in3 = _mm_srai_epi16(in3, 5);
+  in4 = _mm_srai_epi16(in4, 5);
+  in5 = _mm_srai_epi16(in5, 5);
+  in6 = _mm_srai_epi16(in6, 5);
+  in7 = _mm_srai_epi16(in7, 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in0);
+  RECON_AND_STORE(dest + 1 * stride, in1);
+  RECON_AND_STORE(dest + 2 * stride, in2);
+  RECON_AND_STORE(dest + 3 * stride, in3);
+  RECON_AND_STORE(dest + 4 * stride, in4);
+  RECON_AND_STORE(dest + 5 * stride, in5);
+  RECON_AND_STORE(dest + 6 * stride, in6);
+  RECON_AND_STORE(dest + 7 * stride, in7);
+}
+
+void aom_idct8x8_12_add_ssse3(const tran_low_t *input, uint8_t *dest,
+                              int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i stg1_0 = pair_set_epi16(2 * cospi_28_64, 2 * cospi_28_64);
+  const __m128i stg1_1 = pair_set_epi16(2 * cospi_4_64, 2 * cospi_4_64);
+  const __m128i stg1_2 = pair_set_epi16(-2 * cospi_20_64, -2 * cospi_20_64);
+  const __m128i stg1_3 = pair_set_epi16(2 * cospi_12_64, 2 * cospi_12_64);
+  const __m128i stg2_0 = pair_set_epi16(2 * cospi_16_64, 2 * cospi_16_64);
+  const __m128i stk2_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stk2_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg2_2 = pair_set_epi16(2 * cospi_24_64, 2 * cospi_24_64);
+  const __m128i stg2_3 = pair_set_epi16(2 * cospi_8_64, 2 * cospi_8_64);
+  const __m128i stg3_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+  __m128i stp1_0, stp1_1, stp1_2, stp1_3, stp1_4, stp1_5, stp1_6, stp1_7;
+  __m128i stp2_0, stp2_1, stp2_2, stp2_3, stp2_4, stp2_5, stp2_6, stp2_7;
+  __m128i tmp0, tmp1, tmp2, tmp3;
+
+  // Rows. Load 4-row input data.
+  in0 = load_input_data(input);
+  in1 = load_input_data(input + 8 * 1);
+  in2 = load_input_data(input + 8 * 2);
+  in3 = load_input_data(input + 8 * 3);
+
+  // 8x4 Transpose
+  TRANSPOSE_8X8_10(in0, in1, in2, in3, in0, in1);
+
+  // Stage1
+  tmp0 = _mm_mulhrs_epi16(in0, stg1_0);
+  tmp1 = _mm_mulhrs_epi16(in0, stg1_1);
+  tmp2 = _mm_mulhrs_epi16(in1, stg1_2);
+  tmp3 = _mm_mulhrs_epi16(in1, stg1_3);
+
+  stp1_4 = _mm_unpackhi_epi64(tmp0, tmp1);
+  stp1_5 = _mm_unpackhi_epi64(tmp2, tmp3);
+
+  // Stage2
+  tmp0 = _mm_mulhrs_epi16(in0, stg2_0);
+  stp2_0 = _mm_unpacklo_epi64(tmp0, tmp0);
+
+  tmp1 = _mm_mulhrs_epi16(in1, stg2_2);
+  tmp2 = _mm_mulhrs_epi16(in1, stg2_3);
+  stp2_2 = _mm_unpacklo_epi64(tmp2, tmp1);
+
+  tmp0 = _mm_add_epi16(stp1_4, stp1_5);
+  tmp1 = _mm_sub_epi16(stp1_4, stp1_5);
+
+  stp2_4 = tmp0;
+  stp2_5 = _mm_unpacklo_epi64(tmp1, zero);
+  stp2_6 = _mm_unpackhi_epi64(tmp1, zero);
+
+  tmp0 = _mm_unpacklo_epi16(stp2_5, stp2_6);
+  tmp1 = _mm_madd_epi16(tmp0, stg3_0);
+  tmp2 = _mm_madd_epi16(tmp0, stk2_0);  // stg3_1 = stk2_0
+
+  tmp1 = _mm_add_epi32(tmp1, rounding);
+  tmp2 = _mm_add_epi32(tmp2, rounding);
+  tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS);
+  tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+
+  stp1_5 = _mm_packs_epi32(tmp1, tmp2);
+
+  // Stage3
+  tmp2 = _mm_add_epi16(stp2_0, stp2_2);
+  tmp3 = _mm_sub_epi16(stp2_0, stp2_2);
+
+  stp1_2 = _mm_unpackhi_epi64(tmp3, tmp2);
+  stp1_3 = _mm_unpacklo_epi64(tmp3, tmp2);
+
+  // Stage4
+  tmp0 = _mm_add_epi16(stp1_3, stp2_4);
+  tmp1 = _mm_add_epi16(stp1_2, stp1_5);
+  tmp2 = _mm_sub_epi16(stp1_3, stp2_4);
+  tmp3 = _mm_sub_epi16(stp1_2, stp1_5);
+
+  TRANSPOSE_4X8_10(tmp0, tmp1, tmp2, tmp3, in0, in1, in2, in3)
+
+  /* Stage1 */
+  stp1_4 = _mm_mulhrs_epi16(in1, stg1_0);
+  stp1_7 = _mm_mulhrs_epi16(in1, stg1_1);
+  stp1_5 = _mm_mulhrs_epi16(in3, stg1_2);
+  stp1_6 = _mm_mulhrs_epi16(in3, stg1_3);
+
+  /* Stage2 */
+  stp2_0 = _mm_mulhrs_epi16(in0, stg2_0);
+  stp2_1 = _mm_mulhrs_epi16(in0, stg2_0);
+
+  stp2_2 = _mm_mulhrs_epi16(in2, stg2_2);
+  stp2_3 = _mm_mulhrs_epi16(in2, stg2_3);
+
+  stp2_4 = _mm_add_epi16(stp1_4, stp1_5);
+  stp2_5 = _mm_sub_epi16(stp1_4, stp1_5);
+  stp2_6 = _mm_sub_epi16(stp1_7, stp1_6);
+  stp2_7 = _mm_add_epi16(stp1_7, stp1_6);
+
+  /* Stage3 */
+  stp1_0 = _mm_add_epi16(stp2_0, stp2_3);
+  stp1_1 = _mm_add_epi16(stp2_1, stp2_2);
+  stp1_2 = _mm_sub_epi16(stp2_1, stp2_2);
+  stp1_3 = _mm_sub_epi16(stp2_0, stp2_3);
+
+  tmp0 = _mm_unpacklo_epi16(stp2_6, stp2_5);
+  tmp1 = _mm_unpackhi_epi16(stp2_6, stp2_5);
+
+  tmp2 = _mm_madd_epi16(tmp0, stk2_0);
+  tmp3 = _mm_madd_epi16(tmp1, stk2_0);
+  tmp2 = _mm_add_epi32(tmp2, rounding);
+  tmp3 = _mm_add_epi32(tmp3, rounding);
+  tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+  tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+  stp1_6 = _mm_packs_epi32(tmp2, tmp3);
+
+  tmp2 = _mm_madd_epi16(tmp0, stk2_1);
+  tmp3 = _mm_madd_epi16(tmp1, stk2_1);
+  tmp2 = _mm_add_epi32(tmp2, rounding);
+  tmp3 = _mm_add_epi32(tmp3, rounding);
+  tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS);
+  tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS);
+  stp1_5 = _mm_packs_epi32(tmp2, tmp3);
+
+  /* Stage4  */
+  in0 = _mm_add_epi16(stp1_0, stp2_7);
+  in1 = _mm_add_epi16(stp1_1, stp1_6);
+  in2 = _mm_add_epi16(stp1_2, stp1_5);
+  in3 = _mm_add_epi16(stp1_3, stp2_4);
+  in4 = _mm_sub_epi16(stp1_3, stp2_4);
+  in5 = _mm_sub_epi16(stp1_2, stp1_5);
+  in6 = _mm_sub_epi16(stp1_1, stp1_6);
+  in7 = _mm_sub_epi16(stp1_0, stp2_7);
+
+  // Final rounding and shift
+  in0 = _mm_adds_epi16(in0, final_rounding);
+  in1 = _mm_adds_epi16(in1, final_rounding);
+  in2 = _mm_adds_epi16(in2, final_rounding);
+  in3 = _mm_adds_epi16(in3, final_rounding);
+  in4 = _mm_adds_epi16(in4, final_rounding);
+  in5 = _mm_adds_epi16(in5, final_rounding);
+  in6 = _mm_adds_epi16(in6, final_rounding);
+  in7 = _mm_adds_epi16(in7, final_rounding);
+
+  in0 = _mm_srai_epi16(in0, 5);
+  in1 = _mm_srai_epi16(in1, 5);
+  in2 = _mm_srai_epi16(in2, 5);
+  in3 = _mm_srai_epi16(in3, 5);
+  in4 = _mm_srai_epi16(in4, 5);
+  in5 = _mm_srai_epi16(in5, 5);
+  in6 = _mm_srai_epi16(in6, 5);
+  in7 = _mm_srai_epi16(in7, 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in0);
+  RECON_AND_STORE(dest + 1 * stride, in1);
+  RECON_AND_STORE(dest + 2 * stride, in2);
+  RECON_AND_STORE(dest + 3 * stride, in3);
+  RECON_AND_STORE(dest + 4 * stride, in4);
+  RECON_AND_STORE(dest + 5 * stride, in5);
+  RECON_AND_STORE(dest + 6 * stride, in6);
+  RECON_AND_STORE(dest + 7 * stride, in7);
+}
+
+// Only do addition and subtraction butterfly, size = 16, 32
+static INLINE void add_sub_butterfly(const __m128i *in, __m128i *out,
+                                     int size) {
+  int i = 0;
+  const int num = size >> 1;
+  const int bound = size - 1;
+  while (i < num) {
+    out[i] = _mm_add_epi16(in[i], in[bound - i]);
+    out[bound - i] = _mm_sub_epi16(in[i], in[bound - i]);
+    i++;
+  }
+}
+
+#define BUTTERFLY_PAIR(x0, x1, co0, co1)         \
+  do {                                           \
+    tmp0 = _mm_madd_epi16(x0, co0);              \
+    tmp1 = _mm_madd_epi16(x1, co0);              \
+    tmp2 = _mm_madd_epi16(x0, co1);              \
+    tmp3 = _mm_madd_epi16(x1, co1);              \
+    tmp0 = _mm_add_epi32(tmp0, rounding);        \
+    tmp1 = _mm_add_epi32(tmp1, rounding);        \
+    tmp2 = _mm_add_epi32(tmp2, rounding);        \
+    tmp3 = _mm_add_epi32(tmp3, rounding);        \
+    tmp0 = _mm_srai_epi32(tmp0, DCT_CONST_BITS); \
+    tmp1 = _mm_srai_epi32(tmp1, DCT_CONST_BITS); \
+    tmp2 = _mm_srai_epi32(tmp2, DCT_CONST_BITS); \
+    tmp3 = _mm_srai_epi32(tmp3, DCT_CONST_BITS); \
+  } while (0)
+
+static INLINE void butterfly(const __m128i *x0, const __m128i *x1,
+                             const __m128i *c0, const __m128i *c1, __m128i *y0,
+                             __m128i *y1) {
+  __m128i tmp0, tmp1, tmp2, tmp3, u0, u1;
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  u0 = _mm_unpacklo_epi16(*x0, *x1);
+  u1 = _mm_unpackhi_epi16(*x0, *x1);
+  BUTTERFLY_PAIR(u0, u1, *c0, *c1);
+  *y0 = _mm_packs_epi32(tmp0, tmp1);
+  *y1 = _mm_packs_epi32(tmp2, tmp3);
+}
+
+static INLINE void butterfly_self(__m128i *x0, __m128i *x1, const __m128i *c0,
+                                  const __m128i *c1) {
+  __m128i tmp0, tmp1, tmp2, tmp3, u0, u1;
+  const __m128i rounding = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  u0 = _mm_unpacklo_epi16(*x0, *x1);
+  u1 = _mm_unpackhi_epi16(*x0, *x1);
+  BUTTERFLY_PAIR(u0, u1, *c0, *c1);
+  *x0 = _mm_packs_epi32(tmp0, tmp1);
+  *x1 = _mm_packs_epi32(tmp2, tmp3);
+}
+
+static void idct32_34_first_half(const __m128i *in, __m128i *stp1) {
+  const __m128i stk2_0 = pair_set_epi16(2 * cospi_30_64, 2 * cospi_30_64);
+  const __m128i stk2_1 = pair_set_epi16(2 * cospi_2_64, 2 * cospi_2_64);
+  const __m128i stk2_6 = pair_set_epi16(-2 * cospi_26_64, -2 * cospi_26_64);
+  const __m128i stk2_7 = pair_set_epi16(2 * cospi_6_64, 2 * cospi_6_64);
+
+  const __m128i stk3_0 = pair_set_epi16(2 * cospi_28_64, 2 * cospi_28_64);
+  const __m128i stk3_1 = pair_set_epi16(2 * cospi_4_64, 2 * cospi_4_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stk4_0 = pair_set_epi16(2 * cospi_16_64, 2 * cospi_16_64);
+  const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i x0, x1, x4, x5, x6, x7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
+
+  // phase 1
+
+  // 0, 15
+  u2 = _mm_mulhrs_epi16(in[2], stk2_1);  // stp2_15
+  u3 = _mm_mulhrs_epi16(in[6], stk2_7);  // stp2_12
+  v15 = _mm_add_epi16(u2, u3);
+  // in[0], in[4]
+  x0 = _mm_mulhrs_epi16(in[0], stk4_0);  // stp1[0]
+  x7 = _mm_mulhrs_epi16(in[4], stk3_1);  // stp1[7]
+  v0 = _mm_add_epi16(x0, x7);            // stp2_0
+  stp1[0] = _mm_add_epi16(v0, v15);
+  stp1[15] = _mm_sub_epi16(v0, v15);
+
+  // in[2], in[6]
+  u0 = _mm_mulhrs_epi16(in[2], stk2_0);             // stp2_8
+  u1 = _mm_mulhrs_epi16(in[6], stk2_6);             // stp2_11
+  butterfly(&u0, &u2, &stg4_4, &stg4_5, &u4, &u5);  // stp2_9, stp2_14
+  butterfly(&u1, &u3, &stg4_6, &stg4_4, &u6, &u7);  // stp2_10, stp2_13
+
+  v8 = _mm_add_epi16(u0, u1);
+  v9 = _mm_add_epi16(u4, u6);
+  v10 = _mm_sub_epi16(u4, u6);
+  v11 = _mm_sub_epi16(u0, u1);
+  v12 = _mm_sub_epi16(u2, u3);
+  v13 = _mm_sub_epi16(u5, u7);
+  v14 = _mm_add_epi16(u5, u7);
+
+  butterfly_self(&v10, &v13, &stg6_0, &stg4_0);
+  butterfly_self(&v11, &v12, &stg6_0, &stg4_0);
+
+  // 1, 14
+  x1 = _mm_mulhrs_epi16(in[0], stk4_0);  // stp1[1], stk4_1 = stk4_0
+  // stp1[2] = stp1[0], stp1[3] = stp1[1]
+  x4 = _mm_mulhrs_epi16(in[4], stk3_0);  // stp1[4]
+  butterfly(&x7, &x4, &stg4_1, &stg4_0, &x5, &x6);
+  v1 = _mm_add_epi16(x1, x6);  // stp2_1
+  v2 = _mm_add_epi16(x0, x5);  // stp2_2
+  stp1[1] = _mm_add_epi16(v1, v14);
+  stp1[14] = _mm_sub_epi16(v1, v14);
+
+  stp1[2] = _mm_add_epi16(v2, v13);
+  stp1[13] = _mm_sub_epi16(v2, v13);
+
+  v3 = _mm_add_epi16(x1, x4);  // stp2_3
+  v4 = _mm_sub_epi16(x1, x4);  // stp2_4
+
+  v5 = _mm_sub_epi16(x0, x5);  // stp2_5
+
+  v6 = _mm_sub_epi16(x1, x6);  // stp2_6
+  v7 = _mm_sub_epi16(x0, x7);  // stp2_7
+  stp1[3] = _mm_add_epi16(v3, v12);
+  stp1[12] = _mm_sub_epi16(v3, v12);
+
+  stp1[6] = _mm_add_epi16(v6, v9);
+  stp1[9] = _mm_sub_epi16(v6, v9);
+
+  stp1[7] = _mm_add_epi16(v7, v8);
+  stp1[8] = _mm_sub_epi16(v7, v8);
+
+  stp1[4] = _mm_add_epi16(v4, v11);
+  stp1[11] = _mm_sub_epi16(v4, v11);
+
+  stp1[5] = _mm_add_epi16(v5, v10);
+  stp1[10] = _mm_sub_epi16(v5, v10);
+}
+
+static void idct32_34_second_half(const __m128i *in, __m128i *stp1) {
+  const __m128i stk1_0 = pair_set_epi16(2 * cospi_31_64, 2 * cospi_31_64);
+  const __m128i stk1_1 = pair_set_epi16(2 * cospi_1_64, 2 * cospi_1_64);
+  const __m128i stk1_6 = pair_set_epi16(-2 * cospi_25_64, -2 * cospi_25_64);
+  const __m128i stk1_7 = pair_set_epi16(2 * cospi_7_64, 2 * cospi_7_64);
+  const __m128i stk1_8 = pair_set_epi16(2 * cospi_27_64, 2 * cospi_27_64);
+  const __m128i stk1_9 = pair_set_epi16(2 * cospi_5_64, 2 * cospi_5_64);
+  const __m128i stk1_14 = pair_set_epi16(-2 * cospi_29_64, -2 * cospi_29_64);
+  const __m128i stk1_15 = pair_set_epi16(2 * cospi_3_64, 2 * cospi_3_64);
+  const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+  const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  __m128i v16, v17, v18, v19, v20, v21, v22, v23;
+  __m128i v24, v25, v26, v27, v28, v29, v30, v31;
+  __m128i u16, u17, u18, u19, u20, u21, u22, u23;
+  __m128i u24, u25, u26, u27, u28, u29, u30, u31;
+
+  v16 = _mm_mulhrs_epi16(in[1], stk1_0);
+  v31 = _mm_mulhrs_epi16(in[1], stk1_1);
+
+  v19 = _mm_mulhrs_epi16(in[7], stk1_6);
+  v28 = _mm_mulhrs_epi16(in[7], stk1_7);
+
+  v20 = _mm_mulhrs_epi16(in[5], stk1_8);
+  v27 = _mm_mulhrs_epi16(in[5], stk1_9);
+
+  v23 = _mm_mulhrs_epi16(in[3], stk1_14);
+  v24 = _mm_mulhrs_epi16(in[3], stk1_15);
+
+  butterfly(&v16, &v31, &stg3_4, &stg3_5, &v17, &v30);
+  butterfly(&v19, &v28, &stg3_6, &stg3_4, &v18, &v29);
+  butterfly(&v20, &v27, &stg3_8, &stg3_9, &v21, &v26);
+  butterfly(&v23, &v24, &stg3_10, &stg3_8, &v22, &v25);
+
+  u16 = _mm_add_epi16(v16, v19);
+  u17 = _mm_add_epi16(v17, v18);
+  u18 = _mm_sub_epi16(v17, v18);
+  u19 = _mm_sub_epi16(v16, v19);
+  u20 = _mm_sub_epi16(v23, v20);
+  u21 = _mm_sub_epi16(v22, v21);
+  u22 = _mm_add_epi16(v22, v21);
+  u23 = _mm_add_epi16(v23, v20);
+  u24 = _mm_add_epi16(v24, v27);
+  u27 = _mm_sub_epi16(v24, v27);
+  u25 = _mm_add_epi16(v25, v26);
+  u26 = _mm_sub_epi16(v25, v26);
+  u28 = _mm_sub_epi16(v31, v28);
+  u31 = _mm_add_epi16(v28, v31);
+  u29 = _mm_sub_epi16(v30, v29);
+  u30 = _mm_add_epi16(v29, v30);
+
+  butterfly_self(&u18, &u29, &stg4_4, &stg4_5);
+  butterfly_self(&u19, &u28, &stg4_4, &stg4_5);
+  butterfly_self(&u20, &u27, &stg4_6, &stg4_4);
+  butterfly_self(&u21, &u26, &stg4_6, &stg4_4);
+
+  stp1[16] = _mm_add_epi16(u16, u23);
+  stp1[23] = _mm_sub_epi16(u16, u23);
+
+  stp1[17] = _mm_add_epi16(u17, u22);
+  stp1[22] = _mm_sub_epi16(u17, u22);
+
+  stp1[18] = _mm_add_epi16(u18, u21);
+  stp1[21] = _mm_sub_epi16(u18, u21);
+
+  stp1[19] = _mm_add_epi16(u19, u20);
+  stp1[20] = _mm_sub_epi16(u19, u20);
+
+  stp1[24] = _mm_sub_epi16(u31, u24);
+  stp1[31] = _mm_add_epi16(u24, u31);
+
+  stp1[25] = _mm_sub_epi16(u30, u25);
+  stp1[30] = _mm_add_epi16(u25, u30);
+
+  stp1[26] = _mm_sub_epi16(u29, u26);
+  stp1[29] = _mm_add_epi16(u26, u29);
+
+  stp1[27] = _mm_sub_epi16(u28, u27);
+  stp1[28] = _mm_add_epi16(u27, u28);
+
+  butterfly_self(&stp1[20], &stp1[27], &stg6_0, &stg4_0);
+  butterfly_self(&stp1[21], &stp1[26], &stg6_0, &stg4_0);
+  butterfly_self(&stp1[22], &stp1[25], &stg6_0, &stg4_0);
+  butterfly_self(&stp1[23], &stp1[24], &stg6_0, &stg4_0);
+}
+
+// Only upper-left 8x8 has non-zero coeff
+void aom_idct32x32_34_add_ssse3(const tran_low_t *input, uint8_t *dest,
+                                int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  __m128i in[32], col[32];
+  __m128i stp1[32];
+  int i;
+
+  // Load input data. Only need to load the top left 8x8 block.
+  in[0] = load_input_data(input);
+  in[1] = load_input_data(input + 32);
+  in[2] = load_input_data(input + 64);
+  in[3] = load_input_data(input + 96);
+  in[4] = load_input_data(input + 128);
+  in[5] = load_input_data(input + 160);
+  in[6] = load_input_data(input + 192);
+  in[7] = load_input_data(input + 224);
+
+  array_transpose_8x8(in, in);
+  idct32_34_first_half(in, stp1);
+  idct32_34_second_half(in, stp1);
+
+  // 1_D: Store 32 intermediate results for each 8x32 block.
+  add_sub_butterfly(stp1, col, 32);
+  for (i = 0; i < 4; i++) {
+    int j;
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(col + i * 8, in);
+    idct32_34_first_half(in, stp1);
+    idct32_34_second_half(in, stp1);
+
+    // 2_D: Calculate the results and store them to destination.
+    add_sub_butterfly(stp1, in, 32);
+    for (j = 0; j < 32; ++j) {
+      // Final rounding and shift
+      in[j] = _mm_adds_epi16(in[j], final_rounding);
+      in[j] = _mm_srai_epi16(in[j], 6);
+      RECON_AND_STORE(dest + j * stride, in[j]);
+    }
+
+    dest += 8;
+  }
+}
+
+// in0[16] represents the left 8x16 block
+// in1[16] represents the right 8x16 block
+static void load_buffer_16x16(const tran_low_t *input, __m128i *in0,
+                              __m128i *in1) {
+  int i;
+  for (i = 0; i < 16; i++) {
+    in0[i] = load_input_data(input);
+    in1[i] = load_input_data(input + 8);
+    input += 32;
+  }
+}
+
+static void array_transpose_16x16_2(__m128i *in0, __m128i *in1, __m128i *out0,
+                                    __m128i *out1) {
+  array_transpose_8x8(in0, out0);
+  array_transpose_8x8(&in0[8], out1);
+  array_transpose_8x8(in1, &out0[8]);
+  array_transpose_8x8(&in1[8], &out1[8]);
+}
+
+// Group the coefficient calculation into smaller functions
+// to prevent stack spillover:
+// quarter_1: 0-7
+// quarter_2: 8-15
+// quarter_3_4: 16-23, 24-31
+static void idct32_8x32_135_quarter_1(const __m128i *in /*in[16]*/,
+                                      __m128i *out /*out[8]*/) {
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+
+  {
+    const __m128i stk4_0 = pair_set_epi16(2 * cospi_16_64, 2 * cospi_16_64);
+    const __m128i stk4_2 = pair_set_epi16(2 * cospi_24_64, 2 * cospi_24_64);
+    const __m128i stk4_3 = pair_set_epi16(2 * cospi_8_64, 2 * cospi_8_64);
+    u0 = _mm_mulhrs_epi16(in[0], stk4_0);
+    u2 = _mm_mulhrs_epi16(in[8], stk4_2);
+    u3 = _mm_mulhrs_epi16(in[8], stk4_3);
+    u1 = u0;
+  }
+
+  v0 = _mm_add_epi16(u0, u3);
+  v1 = _mm_add_epi16(u1, u2);
+  v2 = _mm_sub_epi16(u1, u2);
+  v3 = _mm_sub_epi16(u0, u3);
+
+  {
+    const __m128i stk3_0 = pair_set_epi16(2 * cospi_28_64, 2 * cospi_28_64);
+    const __m128i stk3_1 = pair_set_epi16(2 * cospi_4_64, 2 * cospi_4_64);
+    const __m128i stk3_2 = pair_set_epi16(-2 * cospi_20_64, -2 * cospi_20_64);
+    const __m128i stk3_3 = pair_set_epi16(2 * cospi_12_64, 2 * cospi_12_64);
+    u4 = _mm_mulhrs_epi16(in[4], stk3_0);
+    u7 = _mm_mulhrs_epi16(in[4], stk3_1);
+    u5 = _mm_mulhrs_epi16(in[12], stk3_2);
+    u6 = _mm_mulhrs_epi16(in[12], stk3_3);
+  }
+
+  v4 = _mm_add_epi16(u4, u5);
+  v5 = _mm_sub_epi16(u4, u5);
+  v6 = _mm_sub_epi16(u7, u6);
+  v7 = _mm_add_epi16(u7, u6);
+
+  {
+    const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+    const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+    butterfly(&v6, &v5, &stg4_1, &stg4_0, &v5, &v6);
+  }
+
+  out[0] = _mm_add_epi16(v0, v7);
+  out[1] = _mm_add_epi16(v1, v6);
+  out[2] = _mm_add_epi16(v2, v5);
+  out[3] = _mm_add_epi16(v3, v4);
+  out[4] = _mm_sub_epi16(v3, v4);
+  out[5] = _mm_sub_epi16(v2, v5);
+  out[6] = _mm_sub_epi16(v1, v6);
+  out[7] = _mm_sub_epi16(v0, v7);
+}
+
+static void idct32_8x32_135_quarter_2(const __m128i *in /*in[16]*/,
+                                      __m128i *out /*out[8]*/) {
+  __m128i u8, u9, u10, u11, u12, u13, u14, u15;
+  __m128i v8, v9, v10, v11, v12, v13, v14, v15;
+
+  {
+    const __m128i stk2_0 = pair_set_epi16(2 * cospi_30_64, 2 * cospi_30_64);
+    const __m128i stk2_1 = pair_set_epi16(2 * cospi_2_64, 2 * cospi_2_64);
+    const __m128i stk2_2 = pair_set_epi16(-2 * cospi_18_64, -2 * cospi_18_64);
+    const __m128i stk2_3 = pair_set_epi16(2 * cospi_14_64, 2 * cospi_14_64);
+    const __m128i stk2_4 = pair_set_epi16(2 * cospi_22_64, 2 * cospi_22_64);
+    const __m128i stk2_5 = pair_set_epi16(2 * cospi_10_64, 2 * cospi_10_64);
+    const __m128i stk2_6 = pair_set_epi16(-2 * cospi_26_64, -2 * cospi_26_64);
+    const __m128i stk2_7 = pair_set_epi16(2 * cospi_6_64, 2 * cospi_6_64);
+    u8 = _mm_mulhrs_epi16(in[2], stk2_0);
+    u15 = _mm_mulhrs_epi16(in[2], stk2_1);
+    u9 = _mm_mulhrs_epi16(in[14], stk2_2);
+    u14 = _mm_mulhrs_epi16(in[14], stk2_3);
+    u10 = _mm_mulhrs_epi16(in[10], stk2_4);
+    u13 = _mm_mulhrs_epi16(in[10], stk2_5);
+    u11 = _mm_mulhrs_epi16(in[6], stk2_6);
+    u12 = _mm_mulhrs_epi16(in[6], stk2_7);
+  }
+
+  v8 = _mm_add_epi16(u8, u9);
+  v9 = _mm_sub_epi16(u8, u9);
+  v10 = _mm_sub_epi16(u11, u10);
+  v11 = _mm_add_epi16(u11, u10);
+  v12 = _mm_add_epi16(u12, u13);
+  v13 = _mm_sub_epi16(u12, u13);
+  v14 = _mm_sub_epi16(u15, u14);
+  v15 = _mm_add_epi16(u15, u14);
+
+  {
+    const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+    const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+    const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+    butterfly_self(&v9, &v14, &stg4_4, &stg4_5);
+    butterfly_self(&v10, &v13, &stg4_6, &stg4_4);
+  }
+
+  out[0] = _mm_add_epi16(v8, v11);
+  out[1] = _mm_add_epi16(v9, v10);
+  out[2] = _mm_sub_epi16(v9, v10);
+  out[3] = _mm_sub_epi16(v8, v11);
+  out[4] = _mm_sub_epi16(v15, v12);
+  out[5] = _mm_sub_epi16(v14, v13);
+  out[6] = _mm_add_epi16(v14, v13);
+  out[7] = _mm_add_epi16(v15, v12);
+
+  {
+    const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+    const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+    butterfly_self(&out[2], &out[5], &stg6_0, &stg4_0);
+    butterfly_self(&out[3], &out[4], &stg6_0, &stg4_0);
+  }
+}
+
+// 8x32 block even indexed 8 inputs of in[16],
+// output first half 16 to out[32]
+static void idct32_8x32_quarter_1_2(const __m128i *in /*in[16]*/,
+                                    __m128i *out /*out[32]*/) {
+  __m128i temp[16];
+  idct32_8x32_135_quarter_1(in, temp);
+  idct32_8x32_135_quarter_2(in, &temp[8]);
+  add_sub_butterfly(temp, out, 16);
+}
+
+// 8x32 block odd indexed 8 inputs of in[16],
+// output second half 16 to out[32]
+static void idct32_8x32_quarter_3_4(const __m128i *in /*in[16]*/,
+                                    __m128i *out /*out[32]*/) {
+  __m128i v16, v17, v18, v19, v20, v21, v22, v23;
+  __m128i v24, v25, v26, v27, v28, v29, v30, v31;
+  __m128i u16, u17, u18, u19, u20, u21, u22, u23;
+  __m128i u24, u25, u26, u27, u28, u29, u30, u31;
+
+  {
+    const __m128i stk1_0 = pair_set_epi16(2 * cospi_31_64, 2 * cospi_31_64);
+    const __m128i stk1_1 = pair_set_epi16(2 * cospi_1_64, 2 * cospi_1_64);
+    const __m128i stk1_2 = pair_set_epi16(-2 * cospi_17_64, -2 * cospi_17_64);
+    const __m128i stk1_3 = pair_set_epi16(2 * cospi_15_64, 2 * cospi_15_64);
+
+    const __m128i stk1_4 = pair_set_epi16(2 * cospi_23_64, 2 * cospi_23_64);
+    const __m128i stk1_5 = pair_set_epi16(2 * cospi_9_64, 2 * cospi_9_64);
+    const __m128i stk1_6 = pair_set_epi16(-2 * cospi_25_64, -2 * cospi_25_64);
+    const __m128i stk1_7 = pair_set_epi16(2 * cospi_7_64, 2 * cospi_7_64);
+    const __m128i stk1_8 = pair_set_epi16(2 * cospi_27_64, 2 * cospi_27_64);
+    const __m128i stk1_9 = pair_set_epi16(2 * cospi_5_64, 2 * cospi_5_64);
+    const __m128i stk1_10 = pair_set_epi16(-2 * cospi_21_64, -2 * cospi_21_64);
+    const __m128i stk1_11 = pair_set_epi16(2 * cospi_11_64, 2 * cospi_11_64);
+
+    const __m128i stk1_12 = pair_set_epi16(2 * cospi_19_64, 2 * cospi_19_64);
+    const __m128i stk1_13 = pair_set_epi16(2 * cospi_13_64, 2 * cospi_13_64);
+    const __m128i stk1_14 = pair_set_epi16(-2 * cospi_29_64, -2 * cospi_29_64);
+    const __m128i stk1_15 = pair_set_epi16(2 * cospi_3_64, 2 * cospi_3_64);
+    u16 = _mm_mulhrs_epi16(in[1], stk1_0);
+    u31 = _mm_mulhrs_epi16(in[1], stk1_1);
+    u17 = _mm_mulhrs_epi16(in[15], stk1_2);
+    u30 = _mm_mulhrs_epi16(in[15], stk1_3);
+
+    u18 = _mm_mulhrs_epi16(in[9], stk1_4);
+    u29 = _mm_mulhrs_epi16(in[9], stk1_5);
+    u19 = _mm_mulhrs_epi16(in[7], stk1_6);
+    u28 = _mm_mulhrs_epi16(in[7], stk1_7);
+
+    u20 = _mm_mulhrs_epi16(in[5], stk1_8);
+    u27 = _mm_mulhrs_epi16(in[5], stk1_9);
+    u21 = _mm_mulhrs_epi16(in[11], stk1_10);
+    u26 = _mm_mulhrs_epi16(in[11], stk1_11);
+
+    u22 = _mm_mulhrs_epi16(in[13], stk1_12);
+    u25 = _mm_mulhrs_epi16(in[13], stk1_13);
+    u23 = _mm_mulhrs_epi16(in[3], stk1_14);
+    u24 = _mm_mulhrs_epi16(in[3], stk1_15);
+  }
+
+  v16 = _mm_add_epi16(u16, u17);
+  v17 = _mm_sub_epi16(u16, u17);
+  v18 = _mm_sub_epi16(u19, u18);
+  v19 = _mm_add_epi16(u19, u18);
+
+  v20 = _mm_add_epi16(u20, u21);
+  v21 = _mm_sub_epi16(u20, u21);
+  v22 = _mm_sub_epi16(u23, u22);
+  v23 = _mm_add_epi16(u23, u22);
+
+  v24 = _mm_add_epi16(u24, u25);
+  v25 = _mm_sub_epi16(u24, u25);
+  v26 = _mm_sub_epi16(u27, u26);
+  v27 = _mm_add_epi16(u27, u26);
+
+  v28 = _mm_add_epi16(u28, u29);
+  v29 = _mm_sub_epi16(u28, u29);
+  v30 = _mm_sub_epi16(u31, u30);
+  v31 = _mm_add_epi16(u31, u30);
+
+  {
+    const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+    const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+    const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+    const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+    const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+    const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+
+    butterfly_self(&v17, &v30, &stg3_4, &stg3_5);
+    butterfly_self(&v18, &v29, &stg3_6, &stg3_4);
+    butterfly_self(&v21, &v26, &stg3_8, &stg3_9);
+    butterfly_self(&v22, &v25, &stg3_10, &stg3_8);
+  }
+
+  u16 = _mm_add_epi16(v16, v19);
+  u17 = _mm_add_epi16(v17, v18);
+  u18 = _mm_sub_epi16(v17, v18);
+  u19 = _mm_sub_epi16(v16, v19);
+  u20 = _mm_sub_epi16(v23, v20);
+  u21 = _mm_sub_epi16(v22, v21);
+  u22 = _mm_add_epi16(v22, v21);
+  u23 = _mm_add_epi16(v23, v20);
+
+  u24 = _mm_add_epi16(v24, v27);
+  u25 = _mm_add_epi16(v25, v26);
+  u26 = _mm_sub_epi16(v25, v26);
+  u27 = _mm_sub_epi16(v24, v27);
+  u28 = _mm_sub_epi16(v31, v28);
+  u29 = _mm_sub_epi16(v30, v29);
+  u30 = _mm_add_epi16(v29, v30);
+  u31 = _mm_add_epi16(v28, v31);
+
+  {
+    const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+    const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+    const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+    butterfly_self(&u18, &u29, &stg4_4, &stg4_5);
+    butterfly_self(&u19, &u28, &stg4_4, &stg4_5);
+    butterfly_self(&u20, &u27, &stg4_6, &stg4_4);
+    butterfly_self(&u21, &u26, &stg4_6, &stg4_4);
+  }
+
+  out[0] = _mm_add_epi16(u16, u23);
+  out[1] = _mm_add_epi16(u17, u22);
+  out[2] = _mm_add_epi16(u18, u21);
+  out[3] = _mm_add_epi16(u19, u20);
+  v20 = _mm_sub_epi16(u19, u20);
+  v21 = _mm_sub_epi16(u18, u21);
+  v22 = _mm_sub_epi16(u17, u22);
+  v23 = _mm_sub_epi16(u16, u23);
+
+  v24 = _mm_sub_epi16(u31, u24);
+  v25 = _mm_sub_epi16(u30, u25);
+  v26 = _mm_sub_epi16(u29, u26);
+  v27 = _mm_sub_epi16(u28, u27);
+  out[12] = _mm_add_epi16(u27, u28);
+  out[13] = _mm_add_epi16(u26, u29);
+  out[14] = _mm_add_epi16(u25, u30);
+  out[15] = _mm_add_epi16(u24, u31);
+
+  {
+    const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+    const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+    butterfly(&v20, &v27, &stg6_0, &stg4_0, &out[4], &out[11]);
+    butterfly(&v21, &v26, &stg6_0, &stg4_0, &out[5], &out[10]);
+    butterfly(&v22, &v25, &stg6_0, &stg4_0, &out[6], &out[9]);
+    butterfly(&v23, &v24, &stg6_0, &stg4_0, &out[7], &out[8]);
+  }
+}
+
+// 8x16 block, input __m128i in[16], output __m128i in[32]
+static void idct32_8x32_135(__m128i *in /*in[32]*/) {
+  __m128i out[32];
+  idct32_8x32_quarter_1_2(in, out);
+  idct32_8x32_quarter_3_4(in, &out[16]);
+  add_sub_butterfly(out, in, 32);
+}
+
+static INLINE void store_buffer_8x32(__m128i *in, uint8_t *dst, int stride) {
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+  int j = 0;
+  while (j < 32) {
+    in[j] = _mm_adds_epi16(in[j], final_rounding);
+    in[j + 1] = _mm_adds_epi16(in[j + 1], final_rounding);
+
+    in[j] = _mm_srai_epi16(in[j], 6);
+    in[j + 1] = _mm_srai_epi16(in[j + 1], 6);
+
+    RECON_AND_STORE(dst, in[j]);
+    dst += stride;
+    RECON_AND_STORE(dst, in[j + 1]);
+    dst += stride;
+    j += 2;
+  }
+}
+
+static INLINE void recon_and_store(__m128i *in0, __m128i *in1, uint8_t *dest,
+                                   int stride) {
+  store_buffer_8x32(in0, dest, stride);
+  store_buffer_8x32(in1, dest + 8, stride);
+}
+
+static INLINE void idct32_135(__m128i *col0, __m128i *col1) {
+  idct32_8x32_135(col0);
+  idct32_8x32_135(col1);
+}
+
+typedef enum { left_16, right_16 } ColsIndicator;
+
+static void transpose_and_copy_16x16(__m128i *in0, __m128i *in1, __m128i *store,
+                                     ColsIndicator cols) {
+  switch (cols) {
+    case left_16: {
+      int i;
+      array_transpose_16x16(in0, in1);
+      for (i = 0; i < 16; ++i) {
+        store[i] = in0[16 + i];
+        store[16 + i] = in1[16 + i];
+      }
+      break;
+    }
+    case right_16: {
+      array_transpose_16x16_2(store, &store[16], in0, in1);
+      break;
+    }
+    default: { assert(0); }
+  }
+}
+
+// Only upper-left 16x16 has non-zero coeff
+void aom_idct32x32_135_add_ssse3(const tran_low_t *input, uint8_t *dest,
+                                 int stride) {
+  // Each array represents an 8x32 block
+  __m128i col0[32], col1[32];
+  // This array represents a 16x16 block
+  __m128i temp[32];
+
+  // Load input data. Only need to load the top left 16x16 block.
+  load_buffer_16x16(input, col0, col1);
+
+  // columns
+  array_transpose_16x16(col0, col1);
+  idct32_135(col0, col1);
+
+  // rows
+  transpose_and_copy_16x16(col0, col1, temp, left_16);
+  idct32_135(col0, col1);
+  recon_and_store(col0, col1, dest, stride);
+
+  transpose_and_copy_16x16(col0, col1, temp, right_16);
+  idct32_135(col0, col1);
+  recon_and_store(col0, col1, dest + 16, stride);
+}
+
+// For each 8x32 block __m128i in[32],
+// Input with index, 2, 6, 10, 14, 18, 22, 26, 30
+// output pixels: 8-15 in __m128i in[32]
+static void idct32_full_8x32_quarter_2(const __m128i *in /*in[32]*/,
+                                       __m128i *out /*out[16]*/) {
+  __m128i u8, u9, u10, u11, u12, u13, u14, u15;  // stp2_
+  __m128i v8, v9, v10, v11, v12, v13, v14, v15;  // stp1_
+
+  {
+    const __m128i stg2_0 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+    const __m128i stg2_1 = pair_set_epi16(cospi_2_64, cospi_30_64);
+    const __m128i stg2_2 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+    const __m128i stg2_3 = pair_set_epi16(cospi_18_64, cospi_14_64);
+    butterfly(&in[2], &in[30], &stg2_0, &stg2_1, &u8, &u15);
+    butterfly(&in[18], &in[14], &stg2_2, &stg2_3, &u9, &u14);
+  }
+
+  v8 = _mm_add_epi16(u8, u9);
+  v9 = _mm_sub_epi16(u8, u9);
+  v14 = _mm_sub_epi16(u15, u14);
+  v15 = _mm_add_epi16(u15, u14);
+
+  {
+    const __m128i stg2_4 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+    const __m128i stg2_5 = pair_set_epi16(cospi_10_64, cospi_22_64);
+    const __m128i stg2_6 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+    const __m128i stg2_7 = pair_set_epi16(cospi_26_64, cospi_6_64);
+    butterfly(&in[10], &in[22], &stg2_4, &stg2_5, &u10, &u13);
+    butterfly(&in[26], &in[6], &stg2_6, &stg2_7, &u11, &u12);
+  }
+
+  v10 = _mm_sub_epi16(u11, u10);
+  v11 = _mm_add_epi16(u11, u10);
+  v12 = _mm_add_epi16(u12, u13);
+  v13 = _mm_sub_epi16(u12, u13);
+
+  {
+    const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+    const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+    const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+    butterfly_self(&v9, &v14, &stg4_4, &stg4_5);
+    butterfly_self(&v10, &v13, &stg4_6, &stg4_4);
+  }
+
+  out[0] = _mm_add_epi16(v8, v11);
+  out[1] = _mm_add_epi16(v9, v10);
+  out[6] = _mm_add_epi16(v14, v13);
+  out[7] = _mm_add_epi16(v15, v12);
+
+  out[2] = _mm_sub_epi16(v9, v10);
+  out[3] = _mm_sub_epi16(v8, v11);
+  out[4] = _mm_sub_epi16(v15, v12);
+  out[5] = _mm_sub_epi16(v14, v13);
+
+  {
+    const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+    const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+    butterfly_self(&out[2], &out[5], &stg6_0, &stg4_0);
+    butterfly_self(&out[3], &out[4], &stg6_0, &stg4_0);
+  }
+}
+
+// For each 8x32 block __m128i in[32],
+// Input with index, 0, 4, 8, 12, 16, 20, 24, 28
+// output pixels: 0-7 in __m128i in[32]
+static void idct32_full_8x32_quarter_1(const __m128i *in /*in[32]*/,
+                                       __m128i *out /*out[8]*/) {
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;  // stp1_
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;  // stp2_
+
+  {
+    const __m128i stg3_0 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+    const __m128i stg3_1 = pair_set_epi16(cospi_4_64, cospi_28_64);
+    const __m128i stg3_2 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+    const __m128i stg3_3 = pair_set_epi16(cospi_20_64, cospi_12_64);
+    butterfly(&in[4], &in[28], &stg3_0, &stg3_1, &u4, &u7);
+    butterfly(&in[20], &in[12], &stg3_2, &stg3_3, &u5, &u6);
+  }
+
+  v4 = _mm_add_epi16(u4, u5);
+  v5 = _mm_sub_epi16(u4, u5);
+  v6 = _mm_sub_epi16(u7, u6);
+  v7 = _mm_add_epi16(u7, u6);
+
+  {
+    const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+    const __m128i stg4_1 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+    const __m128i stg4_2 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+    const __m128i stg4_3 = pair_set_epi16(cospi_8_64, cospi_24_64);
+    butterfly(&v6, &v5, &stg4_1, &stg4_0, &v5, &v6);
+
+    butterfly(&in[0], &in[16], &stg4_0, &stg4_1, &u0, &u1);
+    butterfly(&in[8], &in[24], &stg4_2, &stg4_3, &u2, &u3);
+  }
+
+  v0 = _mm_add_epi16(u0, u3);
+  v1 = _mm_add_epi16(u1, u2);
+  v2 = _mm_sub_epi16(u1, u2);
+  v3 = _mm_sub_epi16(u0, u3);
+
+  out[0] = _mm_add_epi16(v0, v7);
+  out[1] = _mm_add_epi16(v1, v6);
+  out[2] = _mm_add_epi16(v2, v5);
+  out[3] = _mm_add_epi16(v3, v4);
+  out[4] = _mm_sub_epi16(v3, v4);
+  out[5] = _mm_sub_epi16(v2, v5);
+  out[6] = _mm_sub_epi16(v1, v6);
+  out[7] = _mm_sub_epi16(v0, v7);
+}
+
+// For each 8x32 block __m128i in[32],
+// Input with odd index,
+// 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31
+// output pixels: 16-23, 24-31 in __m128i in[32]
+// We avoid hide an offset, 16, inside this function. So we output 0-15 into
+// array out[16]
+static void idct32_full_8x32_quarter_3_4(const __m128i *in /*in[32]*/,
+                                         __m128i *out /*out[16]*/) {
+  __m128i v16, v17, v18, v19, v20, v21, v22, v23;
+  __m128i v24, v25, v26, v27, v28, v29, v30, v31;
+  __m128i u16, u17, u18, u19, u20, u21, u22, u23;
+  __m128i u24, u25, u26, u27, u28, u29, u30, u31;
+
+  {
+    const __m128i stg1_0 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+    const __m128i stg1_1 = pair_set_epi16(cospi_1_64, cospi_31_64);
+    const __m128i stg1_2 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+    const __m128i stg1_3 = pair_set_epi16(cospi_17_64, cospi_15_64);
+    const __m128i stg1_4 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+    const __m128i stg1_5 = pair_set_epi16(cospi_9_64, cospi_23_64);
+    const __m128i stg1_6 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+    const __m128i stg1_7 = pair_set_epi16(cospi_25_64, cospi_7_64);
+    const __m128i stg1_8 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+    const __m128i stg1_9 = pair_set_epi16(cospi_5_64, cospi_27_64);
+    const __m128i stg1_10 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+    const __m128i stg1_11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+    const __m128i stg1_12 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+    const __m128i stg1_13 = pair_set_epi16(cospi_13_64, cospi_19_64);
+    const __m128i stg1_14 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+    const __m128i stg1_15 = pair_set_epi16(cospi_29_64, cospi_3_64);
+    butterfly(&in[1], &in[31], &stg1_0, &stg1_1, &u16, &u31);
+    butterfly(&in[17], &in[15], &stg1_2, &stg1_3, &u17, &u30);
+    butterfly(&in[9], &in[23], &stg1_4, &stg1_5, &u18, &u29);
+    butterfly(&in[25], &in[7], &stg1_6, &stg1_7, &u19, &u28);
+
+    butterfly(&in[5], &in[27], &stg1_8, &stg1_9, &u20, &u27);
+    butterfly(&in[21], &in[11], &stg1_10, &stg1_11, &u21, &u26);
+
+    butterfly(&in[13], &in[19], &stg1_12, &stg1_13, &u22, &u25);
+    butterfly(&in[29], &in[3], &stg1_14, &stg1_15, &u23, &u24);
+  }
+
+  v16 = _mm_add_epi16(u16, u17);
+  v17 = _mm_sub_epi16(u16, u17);
+  v18 = _mm_sub_epi16(u19, u18);
+  v19 = _mm_add_epi16(u19, u18);
+
+  v20 = _mm_add_epi16(u20, u21);
+  v21 = _mm_sub_epi16(u20, u21);
+  v22 = _mm_sub_epi16(u23, u22);
+  v23 = _mm_add_epi16(u23, u22);
+
+  v24 = _mm_add_epi16(u24, u25);
+  v25 = _mm_sub_epi16(u24, u25);
+  v26 = _mm_sub_epi16(u27, u26);
+  v27 = _mm_add_epi16(u27, u26);
+
+  v28 = _mm_add_epi16(u28, u29);
+  v29 = _mm_sub_epi16(u28, u29);
+  v30 = _mm_sub_epi16(u31, u30);
+  v31 = _mm_add_epi16(u31, u30);
+
+  {
+    const __m128i stg3_4 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+    const __m128i stg3_5 = pair_set_epi16(cospi_28_64, cospi_4_64);
+    const __m128i stg3_6 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+    const __m128i stg3_8 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+    const __m128i stg3_9 = pair_set_epi16(cospi_12_64, cospi_20_64);
+    const __m128i stg3_10 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+    butterfly_self(&v17, &v30, &stg3_4, &stg3_5);
+    butterfly_self(&v18, &v29, &stg3_6, &stg3_4);
+    butterfly_self(&v21, &v26, &stg3_8, &stg3_9);
+    butterfly_self(&v22, &v25, &stg3_10, &stg3_8);
+  }
+
+  u16 = _mm_add_epi16(v16, v19);
+  u17 = _mm_add_epi16(v17, v18);
+  u18 = _mm_sub_epi16(v17, v18);
+  u19 = _mm_sub_epi16(v16, v19);
+  u20 = _mm_sub_epi16(v23, v20);
+  u21 = _mm_sub_epi16(v22, v21);
+  u22 = _mm_add_epi16(v22, v21);
+  u23 = _mm_add_epi16(v23, v20);
+
+  u24 = _mm_add_epi16(v24, v27);
+  u25 = _mm_add_epi16(v25, v26);
+  u26 = _mm_sub_epi16(v25, v26);
+  u27 = _mm_sub_epi16(v24, v27);
+
+  u28 = _mm_sub_epi16(v31, v28);
+  u29 = _mm_sub_epi16(v30, v29);
+  u30 = _mm_add_epi16(v29, v30);
+  u31 = _mm_add_epi16(v28, v31);
+
+  {
+    const __m128i stg4_4 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+    const __m128i stg4_5 = pair_set_epi16(cospi_24_64, cospi_8_64);
+    const __m128i stg4_6 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+    butterfly_self(&u18, &u29, &stg4_4, &stg4_5);
+    butterfly_self(&u19, &u28, &stg4_4, &stg4_5);
+    butterfly_self(&u20, &u27, &stg4_6, &stg4_4);
+    butterfly_self(&u21, &u26, &stg4_6, &stg4_4);
+  }
+
+  out[0] = _mm_add_epi16(u16, u23);
+  out[1] = _mm_add_epi16(u17, u22);
+  out[2] = _mm_add_epi16(u18, u21);
+  out[3] = _mm_add_epi16(u19, u20);
+  out[4] = _mm_sub_epi16(u19, u20);
+  out[5] = _mm_sub_epi16(u18, u21);
+  out[6] = _mm_sub_epi16(u17, u22);
+  out[7] = _mm_sub_epi16(u16, u23);
+
+  out[8] = _mm_sub_epi16(u31, u24);
+  out[9] = _mm_sub_epi16(u30, u25);
+  out[10] = _mm_sub_epi16(u29, u26);
+  out[11] = _mm_sub_epi16(u28, u27);
+  out[12] = _mm_add_epi16(u27, u28);
+  out[13] = _mm_add_epi16(u26, u29);
+  out[14] = _mm_add_epi16(u25, u30);
+  out[15] = _mm_add_epi16(u24, u31);
+
+  {
+    const __m128i stg4_0 = pair_set_epi16(cospi_16_64, cospi_16_64);
+    const __m128i stg6_0 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+    butterfly_self(&out[4], &out[11], &stg6_0, &stg4_0);
+    butterfly_self(&out[5], &out[10], &stg6_0, &stg4_0);
+    butterfly_self(&out[6], &out[9], &stg6_0, &stg4_0);
+    butterfly_self(&out[7], &out[8], &stg6_0, &stg4_0);
+  }
+}
+
+static void idct32_full_8x32_quarter_1_2(const __m128i *in /*in[32]*/,
+                                         __m128i *out /*out[32]*/) {
+  __m128i temp[16];
+  idct32_full_8x32_quarter_1(in, temp);
+  idct32_full_8x32_quarter_2(in, &temp[8]);
+  add_sub_butterfly(temp, out, 16);
+}
+
+static void idct32_full_8x32(const __m128i *in /*in[32]*/,
+                             __m128i *out /*out[32]*/) {
+  __m128i temp[32];
+  idct32_full_8x32_quarter_1_2(in, temp);
+  idct32_full_8x32_quarter_3_4(in, &temp[16]);
+  add_sub_butterfly(temp, out, 32);
+}
+
+static void load_buffer_8x32(const tran_low_t *input, __m128i *in) {
+  int i;
+  for (i = 0; i < 8; ++i) {
+    in[i] = load_input_data(input);
+    in[i + 8] = load_input_data(input + 8);
+    in[i + 16] = load_input_data(input + 16);
+    in[i + 24] = load_input_data(input + 24);
+    input += 32;
+  }
+}
+
+void aom_idct32x32_1024_add_ssse3(const tran_low_t *input, uint8_t *dest,
+                                  int stride) {
+  __m128i col[128], in[32];
+  int i, j;
+
+  // rows
+  for (i = 0; i < 4; ++i) {
+    load_buffer_8x32(input, in);
+    input += 32 << 3;
+
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(in, in);
+    array_transpose_8x8(in + 8, in + 8);
+    array_transpose_8x8(in + 16, in + 16);
+    array_transpose_8x8(in + 24, in + 24);
+
+    idct32_full_8x32(in, col + (i << 5));
+  }
+
+  // columns
+  for (i = 0; i < 4; ++i) {
+    j = i << 3;
+    // Transpose 32x8 block to 8x32 block
+    array_transpose_8x8(col + j, in);
+    array_transpose_8x8(col + j + 32, in + 8);
+    array_transpose_8x8(col + j + 64, in + 16);
+    array_transpose_8x8(col + j + 96, in + 24);
+
+    idct32_full_8x32(in, in);
+    store_buffer_8x32(in, dest, stride);
+    dest += 8;
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm b/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm
new file mode 100644
index 0000000000..f0668e6f34
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm
@@ -0,0 +1,112 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro REORDER_INPUTS 0
+  ; a c d b  to  a b c d
+  SWAP 1, 3, 2
+%endmacro
+
+%macro TRANSFORM_COLS 0
+  ; input:
+  ; m0 a
+  ; m1 b
+  ; m2 c
+  ; m3 d
+  paddw           m0,        m2
+  psubw           m3,        m1
+
+  ; wide subtract
+  punpcklwd       m4,        m0
+  punpcklwd       m5,        m3
+  psrad           m4,        16
+  psrad           m5,        16
+  psubd           m4,        m5
+  psrad           m4,        1
+  packssdw        m4,        m4             ; e
+
+  psubw           m5,        m4,        m1  ; b
+  psubw           m4,        m2             ; c
+  psubw           m0,        m5
+  paddw           m3,        m4
+                                ; m0 a
+  SWAP            1,         5  ; m1 b
+  SWAP            2,         4  ; m2 c
+                                ; m3 d
+%endmacro
+
+%macro TRANSPOSE_4X4 0
+  punpcklwd       m0,        m2
+  punpcklwd       m1,        m3
+  mova            m2,        m0
+  punpcklwd       m0,        m1
+  punpckhwd       m2,        m1
+  pshufd          m1,        m0, 0x0e
+  pshufd          m3,        m2, 0x0e
+%endmacro
+
+; transpose a 4x4 int16 matrix in xmm0 and xmm1 to the bottom half of xmm0-xmm3
+%macro TRANSPOSE_4X4_WIDE 0
+  mova            m3, m0
+  punpcklwd       m0, m1
+  punpckhwd       m3, m1
+  mova            m2, m0
+  punpcklwd       m0, m3
+  punpckhwd       m2, m3
+  pshufd          m1, m0, 0x0e
+  pshufd          m3, m2, 0x0e
+%endmacro
+
+%macro ADD_STORE_4P_2X 5  ; src1, src2, tmp1, tmp2, zero
+  movd            m%3,       [outputq]
+  movd            m%4,       [outputq + strideq]
+  punpcklbw       m%3,       m%5
+  punpcklbw       m%4,       m%5
+  paddw           m%1,       m%3
+  paddw           m%2,       m%4
+  packuswb        m%1,       m%5
+  packuswb        m%2,       m%5
+  movd            [outputq], m%1
+  movd            [outputq + strideq], m%2
+%endmacro
+
+INIT_XMM sse2
+cglobal iwht4x4_16_add, 3, 3, 7, input, output, stride
+%if CONFIG_HIGHBITDEPTH
+  mova            m0,        [inputq +  0]
+  packssdw        m0,        [inputq + 16]
+  mova            m1,        [inputq + 32]
+  packssdw        m1,        [inputq + 48]
+%else
+  mova            m0,        [inputq +  0]
+  mova            m1,        [inputq + 16]
+%endif
+  psraw           m0,        2
+  psraw           m1,        2
+
+  TRANSPOSE_4X4_WIDE
+  REORDER_INPUTS
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+  REORDER_INPUTS
+  TRANSFORM_COLS
+
+  pxor            m4, m4
+  ADD_STORE_4P_2X  0, 1, 5, 6, 4
+  lea             outputq, [outputq + 2 * strideq]
+  ADD_STORE_4P_2X  2, 3, 5, 6, 4
+
+  RET
diff --git a/third_party/aom/aom_dsp/x86/loopfilter_avx2.c b/third_party/aom/aom_dsp/x86/loopfilter_avx2.c
new file mode 100644
index 0000000000..bf8150e2a3
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/loopfilter_avx2.c
@@ -0,0 +1,915 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h> /* AVX2 */
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/mem.h"
+
+void aom_lpf_horizontal_edge_8_avx2(unsigned char *s, int p,
+                                    const unsigned char *_blimit,
+                                    const unsigned char *_limit,
+                                    const unsigned char *_thresh) {
+  __m128i mask, hev, flat, flat2;
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi8(1);
+  __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
+  __m128i abs_p1p0;
+
+  const __m128i thresh =
+      _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh[0]));
+  const __m128i limit = _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit[0]));
+  const __m128i blimit =
+      _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit[0]));
+
+  q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
+  q4p4 = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(q4p4), (__m64 *)(s + 4 * p)));
+  q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
+  q3p3 = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(q3p3), (__m64 *)(s + 3 * p)));
+  q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+  q2p2 = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(q2p2), (__m64 *)(s + 2 * p)));
+  q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  q1p1 = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(q1p1), (__m64 *)(s + 1 * p)));
+  p1q1 = _mm_shuffle_epi32(q1p1, 78);
+  q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  q0p0 = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(q0p0), (__m64 *)(s - 0 * p)));
+  p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+  {
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
+    abs_p1p0 =
+        _mm_or_si128(_mm_subs_epu8(q1p1, q0p0), _mm_subs_epu8(q0p0, q1p1));
+    abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+    fe = _mm_set1_epi8(0xfe);
+    ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    abs_p0q0 =
+        _mm_or_si128(_mm_subs_epu8(q0p0, p0q0), _mm_subs_epu8(p0q0, q0p0));
+    abs_p1q1 =
+        _mm_or_si128(_mm_subs_epu8(q1p1, p1q1), _mm_subs_epu8(p1q1, q1p1));
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(q2p2, q1p1), _mm_subs_epu8(q1p1, q2p2)),
+        _mm_or_si128(_mm_subs_epu8(q3p3, q2p2), _mm_subs_epu8(q2p2, q3p3)));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i t1 = _mm_set1_epi16(0x1);
+    __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
+    __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
+    __m128i qs0 = _mm_xor_si128(p0q0, t80);
+    __m128i qs1 = _mm_xor_si128(p1q1, t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+    __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+    __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
+
+    filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, qs0ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    /* (aom_filter + 3 * (qs0 - ps0)) & mask */
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    filter1 = _mm_unpacklo_epi8(zero, filter1);
+    filter1 = _mm_srai_epi16(filter1, 0xB);
+    filter2 = _mm_unpacklo_epi8(zero, filter2);
+    filter2 = _mm_srai_epi16(filter2, 0xB);
+
+    /* Filter1 >> 3 */
+    filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
+    qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
+
+    /* filt >> 1 */
+    filt = _mm_adds_epi16(filter1, t1);
+    filt = _mm_srai_epi16(filt, 1);
+    filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8),
+                            filt);
+    filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
+    qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
+    // loopfilter done
+
+    {
+      __m128i work;
+      flat = _mm_max_epu8(
+          _mm_or_si128(_mm_subs_epu8(q2p2, q0p0), _mm_subs_epu8(q0p0, q2p2)),
+          _mm_or_si128(_mm_subs_epu8(q3p3, q0p0), _mm_subs_epu8(q0p0, q3p3)));
+      flat = _mm_max_epu8(abs_p1p0, flat);
+      flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+      flat = _mm_subs_epu8(flat, one);
+      flat = _mm_cmpeq_epi8(flat, zero);
+      flat = _mm_and_si128(flat, mask);
+
+      q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
+      q5p5 = _mm_castps_si128(
+          _mm_loadh_pi(_mm_castsi128_ps(q5p5), (__m64 *)(s + 5 * p)));
+
+      q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
+      q6p6 = _mm_castps_si128(
+          _mm_loadh_pi(_mm_castsi128_ps(q6p6), (__m64 *)(s + 6 * p)));
+
+      flat2 = _mm_max_epu8(
+          _mm_or_si128(_mm_subs_epu8(q4p4, q0p0), _mm_subs_epu8(q0p0, q4p4)),
+          _mm_or_si128(_mm_subs_epu8(q5p5, q0p0), _mm_subs_epu8(q0p0, q5p5)));
+
+      q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
+      q7p7 = _mm_castps_si128(
+          _mm_loadh_pi(_mm_castsi128_ps(q7p7), (__m64 *)(s + 7 * p)));
+
+      work = _mm_max_epu8(
+          _mm_or_si128(_mm_subs_epu8(q6p6, q0p0), _mm_subs_epu8(q0p0, q6p6)),
+          _mm_or_si128(_mm_subs_epu8(q7p7, q0p0), _mm_subs_epu8(q0p0, q7p7)));
+
+      flat2 = _mm_max_epu8(work, flat2);
+      flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
+      flat2 = _mm_subs_epu8(flat2, one);
+      flat2 = _mm_cmpeq_epi8(flat2, zero);
+      flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+    }
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // flat and wide flat calculations
+    {
+      const __m128i eight = _mm_set1_epi16(8);
+      const __m128i four = _mm_set1_epi16(4);
+      __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+      __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+      __m128i pixelFilter_p, pixelFilter_q;
+      __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+      __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
+
+      p7_16 = _mm_unpacklo_epi8(q7p7, zero);
+      p6_16 = _mm_unpacklo_epi8(q6p6, zero);
+      p5_16 = _mm_unpacklo_epi8(q5p5, zero);
+      p4_16 = _mm_unpacklo_epi8(q4p4, zero);
+      p3_16 = _mm_unpacklo_epi8(q3p3, zero);
+      p2_16 = _mm_unpacklo_epi8(q2p2, zero);
+      p1_16 = _mm_unpacklo_epi8(q1p1, zero);
+      p0_16 = _mm_unpacklo_epi8(q0p0, zero);
+      q0_16 = _mm_unpackhi_epi8(q0p0, zero);
+      q1_16 = _mm_unpackhi_epi8(q1p1, zero);
+      q2_16 = _mm_unpackhi_epi8(q2p2, zero);
+      q3_16 = _mm_unpackhi_epi8(q3p3, zero);
+      q4_16 = _mm_unpackhi_epi8(q4p4, zero);
+      q5_16 = _mm_unpackhi_epi8(q5p5, zero);
+      q6_16 = _mm_unpackhi_epi8(q6p6, zero);
+      q7_16 = _mm_unpackhi_epi8(q7p7, zero);
+
+      pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
+                                    _mm_add_epi16(p4_16, p3_16));
+      pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
+                                    _mm_add_epi16(q4_16, q3_16));
+
+      pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
+      pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+      pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16));
+      pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+      pixelFilter_p =
+          _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q));
+      pixetFilter_p2p1p0 = _mm_add_epi16(
+          four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)), 4);
+      flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3_16, p0_16)), 3);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3_16, q0_16)), 3);
+
+      flat_q0p0 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(p7_16, p7_16);
+      sum_q7 = _mm_add_epi16(q7_16, q7_16);
+      sum_p3 = _mm_add_epi16(p3_16, p3_16);
+      sum_q3 = _mm_add_epi16(q3_16, q3_16);
+
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)), 4);
+      flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+      pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
+      pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1_16)), 3);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1_16)), 3);
+      flat_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      sum_p3 = _mm_add_epi16(sum_p3, p3_16);
+      sum_q3 = _mm_add_epi16(sum_q3, q3_16);
+
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)), 4);
+      flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+      pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
+      pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
+
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2_16)), 3);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2_16)), 3);
+      flat_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)), 4);
+      flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)), 4);
+      flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)), 4);
+      flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)), 4);
+      flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
+    }
+    // wide flat
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    flat = _mm_shuffle_epi32(flat, 68);
+    flat2 = _mm_shuffle_epi32(flat2, 68);
+
+    q2p2 = _mm_andnot_si128(flat, q2p2);
+    flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
+    q2p2 = _mm_or_si128(q2p2, flat_q2p2);
+
+    qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+    flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
+    q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
+
+    qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+    flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
+    q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
+
+    q6p6 = _mm_andnot_si128(flat2, q6p6);
+    flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
+    q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
+    _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6);
+    _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6));
+
+    q5p5 = _mm_andnot_si128(flat2, q5p5);
+    flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
+    q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
+    _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5);
+    _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5));
+
+    q4p4 = _mm_andnot_si128(flat2, q4p4);
+    flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
+    q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
+    _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4);
+    _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4));
+
+    q3p3 = _mm_andnot_si128(flat2, q3p3);
+    flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
+    q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
+    _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3);
+    _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3));
+
+    q2p2 = _mm_andnot_si128(flat2, q2p2);
+    flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
+    q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
+    _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2);
+    _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2));
+
+    q1p1 = _mm_andnot_si128(flat2, q1p1);
+    flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
+    q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
+    _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1);
+    _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1));
+
+    q0p0 = _mm_andnot_si128(flat2, q0p0);
+    flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
+    q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
+    _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0);
+    _mm_storeh_pi((__m64 *)(s - 0 * p), _mm_castsi128_ps(q0p0));
+  }
+}
+
+DECLARE_ALIGNED(32, static const uint8_t, filt_loopfilter_avx2[32]) = {
+  0, 128, 1, 128, 2,  128, 3,  128, 4,  128, 5,  128, 6,  128, 7,  128,
+  8, 128, 9, 128, 10, 128, 11, 128, 12, 128, 13, 128, 14, 128, 15, 128
+};
+
+void aom_lpf_horizontal_edge_16_avx2(unsigned char *s, int p,
+                                     const unsigned char *_blimit,
+                                     const unsigned char *_limit,
+                                     const unsigned char *_thresh) {
+  __m128i mask, hev, flat, flat2;
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi8(1);
+  __m128i p7, p6, p5;
+  __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
+  __m128i q5, q6, q7;
+  __m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4, q256_4,
+      p256_3, q256_3, p256_2, q256_2, p256_1, q256_1, p256_0, q256_0;
+
+  const __m128i thresh =
+      _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh[0]));
+  const __m128i limit = _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit[0]));
+  const __m128i blimit =
+      _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit[0]));
+
+  p256_4 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 5 * p)));
+  p256_3 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 4 * p)));
+  p256_2 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 3 * p)));
+  p256_1 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 2 * p)));
+  p256_0 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 1 * p)));
+  q256_0 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 0 * p)));
+  q256_1 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 1 * p)));
+  q256_2 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 2 * p)));
+  q256_3 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 3 * p)));
+  q256_4 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 4 * p)));
+
+  p4 = _mm256_castsi256_si128(p256_4);
+  p3 = _mm256_castsi256_si128(p256_3);
+  p2 = _mm256_castsi256_si128(p256_2);
+  p1 = _mm256_castsi256_si128(p256_1);
+  p0 = _mm256_castsi256_si128(p256_0);
+  q0 = _mm256_castsi256_si128(q256_0);
+  q1 = _mm256_castsi256_si128(q256_1);
+  q2 = _mm256_castsi256_si128(q256_2);
+  q3 = _mm256_castsi256_si128(q256_3);
+  q4 = _mm256_castsi256_si128(q256_4);
+
+  {
+    const __m128i abs_p1p0 =
+        _mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1));
+    const __m128i abs_q1q0 =
+        _mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1));
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    __m128i abs_p0q0 =
+        _mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0));
+    __m128i abs_p1q1 =
+        _mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1));
+    __m128i work;
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(flat, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
+        _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
+        _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i te0 = _mm_set1_epi8(0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+
+    __m128i ps1 = _mm_xor_si128(p1, t80);
+    __m128i ps0 = _mm_xor_si128(p0, t80);
+    __m128i qs0 = _mm_xor_si128(q0, t80);
+    __m128i qs1 = _mm_xor_si128(q1, t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+    __m128i flat2_p6, flat2_p5, flat2_p4, flat2_p3, flat2_p2, flat2_p1,
+        flat2_p0, flat2_q0, flat2_q1, flat2_q2, flat2_q3, flat2_q4, flat2_q5,
+        flat2_q6, flat_p2, flat_p1, flat_p0, flat_q0, flat_q1, flat_q2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    /* (aom_filter + 3 * (qs0 - ps0)) & mask */
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    /* Filter1 >> 3 */
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+
+    /* Filter2 >> 3 */
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+
+    /* filt >> 1 */
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+    qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    // loopfilter done
+
+    {
+      __m128i work;
+      work = _mm_max_epu8(
+          _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
+          _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
+      flat = _mm_max_epu8(work, flat);
+      work = _mm_max_epu8(
+          _mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)),
+          _mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3)));
+      flat = _mm_max_epu8(work, flat);
+      work = _mm_max_epu8(
+          _mm_or_si128(_mm_subs_epu8(p4, p0), _mm_subs_epu8(p0, p4)),
+          _mm_or_si128(_mm_subs_epu8(q4, q0), _mm_subs_epu8(q0, q4)));
+      flat = _mm_subs_epu8(flat, one);
+      flat = _mm_cmpeq_epi8(flat, zero);
+      flat = _mm_and_si128(flat, mask);
+
+      p256_5 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s - 6 * p)));
+      q256_5 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s + 5 * p)));
+      p5 = _mm256_castsi256_si128(p256_5);
+      q5 = _mm256_castsi256_si128(q256_5);
+      flat2 = _mm_max_epu8(
+          _mm_or_si128(_mm_subs_epu8(p5, p0), _mm_subs_epu8(p0, p5)),
+          _mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5)));
+
+      flat2 = _mm_max_epu8(work, flat2);
+      p256_6 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s - 7 * p)));
+      q256_6 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s + 6 * p)));
+      p6 = _mm256_castsi256_si128(p256_6);
+      q6 = _mm256_castsi256_si128(q256_6);
+      work = _mm_max_epu8(
+          _mm_or_si128(_mm_subs_epu8(p6, p0), _mm_subs_epu8(p0, p6)),
+          _mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6)));
+
+      flat2 = _mm_max_epu8(work, flat2);
+
+      p256_7 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s - 8 * p)));
+      q256_7 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s + 7 * p)));
+      p7 = _mm256_castsi256_si128(p256_7);
+      q7 = _mm256_castsi256_si128(q256_7);
+      work = _mm_max_epu8(
+          _mm_or_si128(_mm_subs_epu8(p7, p0), _mm_subs_epu8(p0, p7)),
+          _mm_or_si128(_mm_subs_epu8(q7, q0), _mm_subs_epu8(q0, q7)));
+
+      flat2 = _mm_max_epu8(work, flat2);
+      flat2 = _mm_subs_epu8(flat2, one);
+      flat2 = _mm_cmpeq_epi8(flat2, zero);
+      flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+    }
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // flat and wide flat calculations
+    {
+      const __m256i eight = _mm256_set1_epi16(8);
+      const __m256i four = _mm256_set1_epi16(4);
+      __m256i pixelFilter_p, pixelFilter_q, pixetFilter_p2p1p0,
+          pixetFilter_q2q1q0, sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
+
+      const __m256i filter =
+          _mm256_load_si256((__m256i const *)filt_loopfilter_avx2);
+      p256_7 = _mm256_shuffle_epi8(p256_7, filter);
+      p256_6 = _mm256_shuffle_epi8(p256_6, filter);
+      p256_5 = _mm256_shuffle_epi8(p256_5, filter);
+      p256_4 = _mm256_shuffle_epi8(p256_4, filter);
+      p256_3 = _mm256_shuffle_epi8(p256_3, filter);
+      p256_2 = _mm256_shuffle_epi8(p256_2, filter);
+      p256_1 = _mm256_shuffle_epi8(p256_1, filter);
+      p256_0 = _mm256_shuffle_epi8(p256_0, filter);
+      q256_0 = _mm256_shuffle_epi8(q256_0, filter);
+      q256_1 = _mm256_shuffle_epi8(q256_1, filter);
+      q256_2 = _mm256_shuffle_epi8(q256_2, filter);
+      q256_3 = _mm256_shuffle_epi8(q256_3, filter);
+      q256_4 = _mm256_shuffle_epi8(q256_4, filter);
+      q256_5 = _mm256_shuffle_epi8(q256_5, filter);
+      q256_6 = _mm256_shuffle_epi8(q256_6, filter);
+      q256_7 = _mm256_shuffle_epi8(q256_7, filter);
+
+      pixelFilter_p = _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_5),
+                                       _mm256_add_epi16(p256_4, p256_3));
+      pixelFilter_q = _mm256_add_epi16(_mm256_add_epi16(q256_6, q256_5),
+                                       _mm256_add_epi16(q256_4, q256_3));
+
+      pixetFilter_p2p1p0 =
+          _mm256_add_epi16(p256_0, _mm256_add_epi16(p256_2, p256_1));
+      pixelFilter_p = _mm256_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+      pixetFilter_q2q1q0 =
+          _mm256_add_epi16(q256_0, _mm256_add_epi16(q256_2, q256_1));
+      pixelFilter_q = _mm256_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+
+      pixelFilter_p = _mm256_add_epi16(
+          eight, _mm256_add_epi16(pixelFilter_p, pixelFilter_q));
+
+      pixetFilter_p2p1p0 = _mm256_add_epi16(
+          four, _mm256_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+
+      res_p = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(p256_7, p256_0)), 4);
+
+      flat2_p0 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+
+      res_q = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(q256_7, q256_0)), 4);
+
+      flat2_q0 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+
+      res_p =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
+                                             _mm256_add_epi16(p256_3, p256_0)),
+                            3);
+
+      flat_p0 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+
+      res_q =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
+                                             _mm256_add_epi16(q256_3, q256_0)),
+                            3);
+
+      flat_q0 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+
+      sum_p7 = _mm256_add_epi16(p256_7, p256_7);
+
+      sum_q7 = _mm256_add_epi16(q256_7, q256_7);
+
+      sum_p3 = _mm256_add_epi16(p256_3, p256_3);
+
+      sum_q3 = _mm256_add_epi16(q256_3, q256_3);
+
+      pixelFilter_q = _mm256_sub_epi16(pixelFilter_p, p256_6);
+
+      pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_6);
+
+      res_p = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_1)), 4);
+
+      flat2_p1 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+
+      res_q = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_1)), 4);
+
+      flat2_q1 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+
+      pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_p2p1p0, p256_2);
+
+      pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_2);
+
+      res_p =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
+                                             _mm256_add_epi16(sum_p3, p256_1)),
+                            3);
+
+      flat_p1 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+
+      res_q =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0,
+                                             _mm256_add_epi16(sum_q3, q256_1)),
+                            3);
+
+      flat_q1 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+
+      sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+      sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+      sum_p3 = _mm256_add_epi16(sum_p3, p256_3);
+
+      sum_q3 = _mm256_add_epi16(sum_q3, q256_3);
+
+      pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_5);
+
+      pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_5);
+
+      res_p = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_2)), 4);
+
+      flat2_p2 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+
+      res_q = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_2)), 4);
+
+      flat2_q2 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+
+      pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_1);
+
+      pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_q2q1q0, p256_1);
+
+      res_p =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
+                                             _mm256_add_epi16(sum_p3, p256_2)),
+                            3);
+
+      flat_p2 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+
+      res_q =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0,
+                                             _mm256_add_epi16(sum_q3, q256_2)),
+                            3);
+
+      flat_q2 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+
+      sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+      sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+      pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_4);
+
+      pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_4);
+
+      res_p = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_3)), 4);
+
+      flat2_p3 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+
+      res_q = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_3)), 4);
+
+      flat2_q3 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+
+      sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+      sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+      pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_3);
+
+      pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_3);
+
+      res_p = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_4)), 4);
+
+      flat2_p4 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+
+      res_q = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_4)), 4);
+
+      flat2_q4 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+
+      sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+      sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+      pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_2);
+
+      pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_2);
+
+      res_p = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_5)), 4);
+
+      flat2_p5 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+
+      res_q = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_5)), 4);
+
+      flat2_q5 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+
+      sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+      sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+      pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_1);
+
+      pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_1);
+
+      res_p = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_6)), 4);
+
+      flat2_p6 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+
+      res_q = _mm256_srli_epi16(
+          _mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_6)), 4);
+
+      flat2_q6 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+    }
+
+    // wide flat
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    p2 = _mm_andnot_si128(flat, p2);
+    flat_p2 = _mm_and_si128(flat, flat_p2);
+    p2 = _mm_or_si128(flat_p2, p2);
+
+    p1 = _mm_andnot_si128(flat, ps1);
+    flat_p1 = _mm_and_si128(flat, flat_p1);
+    p1 = _mm_or_si128(flat_p1, p1);
+
+    p0 = _mm_andnot_si128(flat, ps0);
+    flat_p0 = _mm_and_si128(flat, flat_p0);
+    p0 = _mm_or_si128(flat_p0, p0);
+
+    q0 = _mm_andnot_si128(flat, qs0);
+    flat_q0 = _mm_and_si128(flat, flat_q0);
+    q0 = _mm_or_si128(flat_q0, q0);
+
+    q1 = _mm_andnot_si128(flat, qs1);
+    flat_q1 = _mm_and_si128(flat, flat_q1);
+    q1 = _mm_or_si128(flat_q1, q1);
+
+    q2 = _mm_andnot_si128(flat, q2);
+    flat_q2 = _mm_and_si128(flat, flat_q2);
+    q2 = _mm_or_si128(flat_q2, q2);
+
+    p6 = _mm_andnot_si128(flat2, p6);
+    flat2_p6 = _mm_and_si128(flat2, flat2_p6);
+    p6 = _mm_or_si128(flat2_p6, p6);
+    _mm_storeu_si128((__m128i *)(s - 7 * p), p6);
+
+    p5 = _mm_andnot_si128(flat2, p5);
+    flat2_p5 = _mm_and_si128(flat2, flat2_p5);
+    p5 = _mm_or_si128(flat2_p5, p5);
+    _mm_storeu_si128((__m128i *)(s - 6 * p), p5);
+
+    p4 = _mm_andnot_si128(flat2, p4);
+    flat2_p4 = _mm_and_si128(flat2, flat2_p4);
+    p4 = _mm_or_si128(flat2_p4, p4);
+    _mm_storeu_si128((__m128i *)(s - 5 * p), p4);
+
+    p3 = _mm_andnot_si128(flat2, p3);
+    flat2_p3 = _mm_and_si128(flat2, flat2_p3);
+    p3 = _mm_or_si128(flat2_p3, p3);
+    _mm_storeu_si128((__m128i *)(s - 4 * p), p3);
+
+    p2 = _mm_andnot_si128(flat2, p2);
+    flat2_p2 = _mm_and_si128(flat2, flat2_p2);
+    p2 = _mm_or_si128(flat2_p2, p2);
+    _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+
+    p1 = _mm_andnot_si128(flat2, p1);
+    flat2_p1 = _mm_and_si128(flat2, flat2_p1);
+    p1 = _mm_or_si128(flat2_p1, p1);
+    _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+
+    p0 = _mm_andnot_si128(flat2, p0);
+    flat2_p0 = _mm_and_si128(flat2, flat2_p0);
+    p0 = _mm_or_si128(flat2_p0, p0);
+    _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+
+    q0 = _mm_andnot_si128(flat2, q0);
+    flat2_q0 = _mm_and_si128(flat2, flat2_q0);
+    q0 = _mm_or_si128(flat2_q0, q0);
+    _mm_storeu_si128((__m128i *)(s - 0 * p), q0);
+
+    q1 = _mm_andnot_si128(flat2, q1);
+    flat2_q1 = _mm_and_si128(flat2, flat2_q1);
+    q1 = _mm_or_si128(flat2_q1, q1);
+    _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+
+    q2 = _mm_andnot_si128(flat2, q2);
+    flat2_q2 = _mm_and_si128(flat2, flat2_q2);
+    q2 = _mm_or_si128(flat2_q2, q2);
+    _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+
+    q3 = _mm_andnot_si128(flat2, q3);
+    flat2_q3 = _mm_and_si128(flat2, flat2_q3);
+    q3 = _mm_or_si128(flat2_q3, q3);
+    _mm_storeu_si128((__m128i *)(s + 3 * p), q3);
+
+    q4 = _mm_andnot_si128(flat2, q4);
+    flat2_q4 = _mm_and_si128(flat2, flat2_q4);
+    q4 = _mm_or_si128(flat2_q4, q4);
+    _mm_storeu_si128((__m128i *)(s + 4 * p), q4);
+
+    q5 = _mm_andnot_si128(flat2, q5);
+    flat2_q5 = _mm_and_si128(flat2, flat2_q5);
+    q5 = _mm_or_si128(flat2_q5, q5);
+    _mm_storeu_si128((__m128i *)(s + 5 * p), q5);
+
+    q6 = _mm_andnot_si128(flat2, q6);
+    flat2_q6 = _mm_and_si128(flat2, flat2_q6);
+    q6 = _mm_or_si128(flat2_q6, q6);
+    _mm_storeu_si128((__m128i *)(s + 6 * p), q6);
+  }
+  _mm256_zeroupper();
+}
diff --git a/third_party/aom/aom_dsp/x86/loopfilter_sse2.c b/third_party/aom/aom_dsp/x86/loopfilter_sse2.c
new file mode 100644
index 0000000000..7e134dc63f
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/loopfilter_sse2.c
@@ -0,0 +1,1892 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/emmintrin_compat.h"
+
+static INLINE __m128i abs_diff(__m128i a, __m128i b) {
+  return _mm_or_si128(_mm_subs_epu8(a, b), _mm_subs_epu8(b, a));
+}
+
+#if CONFIG_PARALLEL_DEBLOCKING
+// filter_mask and hev_mask
+#define FILTER_HEV_MASK4                                                      \
+  do {                                                                        \
+    /* (abs(q1 - q0), abs(p1 - p0) */                                         \
+    __m128i flat = abs_diff(q1p1, q0p0);                                      \
+    /* abs(p1 - q1), abs(p0 - q0) */                                          \
+    const __m128i abs_p1q1p0q0 = abs_diff(p1p0, q1q0);                        \
+    __m128i abs_p0q0, abs_p1q1;                                               \
+                                                                              \
+    /* const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); */       \
+    hev =                                                                     \
+        _mm_unpacklo_epi8(_mm_max_epu8(flat, _mm_srli_si128(flat, 8)), zero); \
+    hev = _mm_cmpgt_epi16(hev, thresh);                                       \
+    hev = _mm_packs_epi16(hev, hev);                                          \
+                                                                              \
+    /* const int8_t mask = filter_mask2(*limit, *blimit, */                   \
+    /*                                  p1, p0, q0, q1); */                   \
+    abs_p0q0 =                                                                \
+        _mm_adds_epu8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p0 - q0) * 2 */     \
+    abs_p1q1 =                                                                \
+        _mm_unpackhi_epi8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p1 - q1) */     \
+    abs_p1q1 = _mm_srli_epi16(abs_p1q1, 9);                                   \
+    abs_p1q1 = _mm_packs_epi16(abs_p1q1, abs_p1q1); /* abs(p1 - q1) / 2 */    \
+    /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 */                                 \
+    mask = _mm_adds_epu8(abs_p0q0, abs_p1q1);                                 \
+    flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));                       \
+    mask = _mm_unpacklo_epi64(mask, flat);                                    \
+    mask = _mm_subs_epu8(mask, limit);                                        \
+    mask = _mm_cmpeq_epi8(mask, zero);                                        \
+    mask = _mm_and_si128(mask, _mm_srli_si128(mask, 8));                      \
+  } while (0)
+#endif  // CONFIG_PARALLEL_DEBLOCKING
+
+// filter_mask and hev_mask
+#define FILTER_HEV_MASK                                                       \
+  do {                                                                        \
+    /* (abs(q1 - q0), abs(p1 - p0) */                                         \
+    __m128i flat = abs_diff(q1p1, q0p0);                                      \
+    /* abs(p1 - q1), abs(p0 - q0) */                                          \
+    const __m128i abs_p1q1p0q0 = abs_diff(p1p0, q1q0);                        \
+    __m128i abs_p0q0, abs_p1q1, work;                                         \
+                                                                              \
+    /* const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); */       \
+    hev =                                                                     \
+        _mm_unpacklo_epi8(_mm_max_epu8(flat, _mm_srli_si128(flat, 8)), zero); \
+    hev = _mm_cmpgt_epi16(hev, thresh);                                       \
+    hev = _mm_packs_epi16(hev, hev);                                          \
+                                                                              \
+    /* const int8_t mask = filter_mask(*limit, *blimit, */                    \
+    /*                                 p3, p2, p1, p0, q0, q1, q2, q3); */    \
+    abs_p0q0 =                                                                \
+        _mm_adds_epu8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p0 - q0) * 2 */     \
+    abs_p1q1 =                                                                \
+        _mm_unpackhi_epi8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p1 - q1) */     \
+    abs_p1q1 = _mm_srli_epi16(abs_p1q1, 9);                                   \
+    abs_p1q1 = _mm_packs_epi16(abs_p1q1, abs_p1q1); /* abs(p1 - q1) / 2 */    \
+    /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 */                                 \
+    mask = _mm_adds_epu8(abs_p0q0, abs_p1q1);                                 \
+    /* abs(p3 - p2), abs(p2 - p1) */                                          \
+    work = abs_diff(p3p2, p2p1);                                              \
+    flat = _mm_max_epu8(work, flat);                                          \
+    /* abs(q3 - q2), abs(q2 - q1) */                                          \
+    work = abs_diff(q3q2, q2q1);                                              \
+    flat = _mm_max_epu8(work, flat);                                          \
+    flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));                       \
+    mask = _mm_unpacklo_epi64(mask, flat);                                    \
+    mask = _mm_subs_epu8(mask, limit);                                        \
+    mask = _mm_cmpeq_epi8(mask, zero);                                        \
+    mask = _mm_and_si128(mask, _mm_srli_si128(mask, 8));                      \
+  } while (0)
+
+#define FILTER4                                                             \
+  do {                                                                      \
+    const __m128i t3t4 =                                                    \
+        _mm_set_epi8(3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4);       \
+    const __m128i t80 = _mm_set1_epi8(0x80);                                \
+    __m128i filter, filter2filter1, work;                                   \
+                                                                            \
+    ps1ps0 = _mm_xor_si128(p1p0, t80); /* ^ 0x80 */                         \
+    qs1qs0 = _mm_xor_si128(q1q0, t80);                                      \
+                                                                            \
+    /* int8_t filter = signed_char_clamp(ps1 - qs1) & hev; */               \
+    work = _mm_subs_epi8(ps1ps0, qs1qs0);                                   \
+    filter = _mm_and_si128(_mm_srli_si128(work, 8), hev);                   \
+    /* filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; */      \
+    filter = _mm_subs_epi8(filter, work);                                   \
+    filter = _mm_subs_epi8(filter, work);                                   \
+    filter = _mm_subs_epi8(filter, work); /* + 3 * (qs0 - ps0) */           \
+    filter = _mm_and_si128(filter, mask); /* & mask */                      \
+    filter = _mm_unpacklo_epi64(filter, filter);                            \
+                                                                            \
+    /* filter1 = signed_char_clamp(filter + 4) >> 3; */                     \
+    /* filter2 = signed_char_clamp(filter + 3) >> 3; */                     \
+    filter2filter1 = _mm_adds_epi8(filter, t3t4); /* signed_char_clamp */   \
+    filter = _mm_unpackhi_epi8(filter2filter1, filter2filter1);             \
+    filter2filter1 = _mm_unpacklo_epi8(filter2filter1, filter2filter1);     \
+    filter2filter1 = _mm_srai_epi16(filter2filter1, 11); /* >> 3 */         \
+    filter = _mm_srai_epi16(filter, 11);                 /* >> 3 */         \
+    filter2filter1 = _mm_packs_epi16(filter2filter1, filter);               \
+                                                                            \
+    /* filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; */                   \
+    filter = _mm_subs_epi8(filter2filter1, ff); /* + 1 */                   \
+    filter = _mm_unpacklo_epi8(filter, filter);                             \
+    filter = _mm_srai_epi16(filter, 9); /* round */                         \
+    filter = _mm_packs_epi16(filter, filter);                               \
+    filter = _mm_andnot_si128(hev, filter);                                 \
+                                                                            \
+    hev = _mm_unpackhi_epi64(filter2filter1, filter);                       \
+    filter2filter1 = _mm_unpacklo_epi64(filter2filter1, filter);            \
+                                                                            \
+    /* signed_char_clamp(qs1 - filter), signed_char_clamp(qs0 - filter1) */ \
+    qs1qs0 = _mm_subs_epi8(qs1qs0, filter2filter1);                         \
+    /* signed_char_clamp(ps1 + filter), signed_char_clamp(ps0 + filter2) */ \
+    ps1ps0 = _mm_adds_epi8(ps1ps0, hev);                                    \
+    qs1qs0 = _mm_xor_si128(qs1qs0, t80); /* ^ 0x80 */                       \
+    ps1ps0 = _mm_xor_si128(ps1ps0, t80); /* ^ 0x80 */                       \
+  } while (0)
+
+void aom_lpf_horizontal_4_sse2(uint8_t *s, int p /* pitch */,
+                               const uint8_t *_blimit, const uint8_t *_limit,
+                               const uint8_t *_thresh) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i limit =
+      _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)_blimit),
+                         _mm_loadl_epi64((const __m128i *)_limit));
+  const __m128i thresh =
+      _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero);
+  const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+#if !CONFIG_PARALLEL_DEBLOCKING
+  __m128i p3p2, p2p1, q3q2, q2q1;
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  __m128i q1p1, q0p0, p1p0, q1q0, ps1ps0, qs1qs0;
+  __m128i mask, hev;
+#if !CONFIG_PARALLEL_DEBLOCKING
+  p3p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)),
+                            _mm_loadl_epi64((__m128i *)(s - 4 * p)));
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 1 * p)));
+  q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 0 * p)));
+#if !CONFIG_PARALLEL_DEBLOCKING
+  q3q2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s + 2 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 3 * p)));
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  p1p0 = _mm_unpacklo_epi64(q0p0, q1p1);
+  q1q0 = _mm_unpackhi_epi64(q0p0, q1p1);
+#if !CONFIG_PARALLEL_DEBLOCKING
+  p2p1 = _mm_unpacklo_epi64(q1p1, p3p2);
+  q2q1 = _mm_unpacklo_epi64(_mm_srli_si128(q1p1, 8), q3q2);
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+#if !CONFIG_PARALLEL_DEBLOCKING
+  FILTER_HEV_MASK;
+#else   // CONFIG_PARALLEL_DEBLOCKING
+  FILTER_HEV_MASK4;
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  FILTER4;
+
+  _mm_storeh_pi((__m64 *)(s - 2 * p), _mm_castsi128_ps(ps1ps0));  // *op1
+  _mm_storel_epi64((__m128i *)(s - 1 * p), ps1ps0);               // *op0
+  _mm_storel_epi64((__m128i *)(s + 0 * p), qs1qs0);               // *oq0
+  _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(qs1qs0));  // *oq1
+}
+
+void aom_lpf_vertical_4_sse2(uint8_t *s, int p /* pitch */,
+                             const uint8_t *_blimit, const uint8_t *_limit,
+                             const uint8_t *_thresh) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i limit =
+      _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)_blimit),
+                         _mm_loadl_epi64((const __m128i *)_limit));
+  const __m128i thresh =
+      _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero);
+  const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+  __m128i x0, x1, x2, x3;
+#if !CONFIG_PARALLEL_DEBLOCKING
+  __m128i p3p2, p2p1, q3q2, q2q1;
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  __m128i q1p1, q0p0, p1p0, q1q0, ps1ps0, qs1qs0;
+  __m128i mask, hev;
+
+  // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+  q1q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 0 * p - 4)),
+                           _mm_loadl_epi64((__m128i *)(s + 1 * p - 4)));
+
+  // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+  x1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 2 * p - 4)),
+                         _mm_loadl_epi64((__m128i *)(s + 3 * p - 4)));
+
+  // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+  x2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 4 * p - 4)),
+                         _mm_loadl_epi64((__m128i *)(s + 5 * p - 4)));
+
+  // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+  x3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(s + 6 * p - 4)),
+                         _mm_loadl_epi64((__m128i *)(s + 7 * p - 4)));
+
+  // Transpose 8x8
+  // 00 10 20 30 01 11 21 31  02 12 22 32 03 13 23 33
+  p1p0 = _mm_unpacklo_epi16(q1q0, x1);
+  // 40 50 60 70 41 51 61 71  42 52 62 72 43 53 63 73
+  x0 = _mm_unpacklo_epi16(x2, x3);
+#if !CONFIG_PARALLEL_DEBLOCKING
+  // 00 10 20 30 40 50 60 70  01 11 21 31 41 51 61 71
+  p3p2 = _mm_unpacklo_epi32(p1p0, x0);
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  // 02 12 22 32 42 52 62 72  03 13 23 33 43 53 63 73
+  p1p0 = _mm_unpackhi_epi32(p1p0, x0);
+#if !CONFIG_PARALLEL_DEBLOCKING
+  p3p2 = _mm_unpackhi_epi64(p3p2, _mm_slli_si128(p3p2, 8));  // swap lo and high
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  p1p0 = _mm_unpackhi_epi64(p1p0, _mm_slli_si128(p1p0, 8));  // swap lo and high
+
+  // 04 14 24 34 05 15 25 35  06 16 26 36 07 17 27 37
+  q1q0 = _mm_unpackhi_epi16(q1q0, x1);
+  // 44 54 64 74 45 55 65 75  46 56 66 76 47 57 67 77
+  x2 = _mm_unpackhi_epi16(x2, x3);
+#if !CONFIG_PARALLEL_DEBLOCKING
+  // 06 16 26 36 46 56 66 76  07 17 27 37 47 57 67 77
+  q3q2 = _mm_unpackhi_epi32(q1q0, x2);
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  // 04 14 24 34 44 54 64 74  05 15 25 35 45 55 65 75
+  q1q0 = _mm_unpacklo_epi32(q1q0, x2);
+
+  q0p0 = _mm_unpacklo_epi64(p1p0, q1q0);
+  q1p1 = _mm_unpackhi_epi64(p1p0, q1q0);
+  p1p0 = _mm_unpacklo_epi64(q0p0, q1p1);
+#if !CONFIG_PARALLEL_DEBLOCKING
+  p2p1 = _mm_unpacklo_epi64(q1p1, p3p2);
+  q2q1 = _mm_unpacklo_epi64(_mm_srli_si128(q1p1, 8), q3q2);
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+#if !CONFIG_PARALLEL_DEBLOCKING
+  FILTER_HEV_MASK;
+#else   // CONFIG_PARALLEL_DEBLOCKING
+  FILTER_HEV_MASK4;
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  FILTER4;
+
+  // Transpose 8x4 to 4x8
+  // qs1qs0: 20 21 22 23 24 25 26 27  30 31 32 33 34 34 36 37
+  // ps1ps0: 10 11 12 13 14 15 16 17  00 01 02 03 04 05 06 07
+  // 00 01 02 03 04 05 06 07  10 11 12 13 14 15 16 17
+  ps1ps0 = _mm_unpackhi_epi64(ps1ps0, _mm_slli_si128(ps1ps0, 8));
+  // 10 30 11 31 12 32 13 33  14 34 15 35 16 36 17 37
+  x0 = _mm_unpackhi_epi8(ps1ps0, qs1qs0);
+  // 00 20 01 21 02 22 03 23  04 24 05 25 06 26 07 27
+  ps1ps0 = _mm_unpacklo_epi8(ps1ps0, qs1qs0);
+  // 04 14 24 34 05 15 25 35  06 16 26 36 07 17 27 37
+  qs1qs0 = _mm_unpackhi_epi8(ps1ps0, x0);
+  // 00 10 20 30 01 11 21 31  02 12 22 32 03 13 23 33
+  ps1ps0 = _mm_unpacklo_epi8(ps1ps0, x0);
+
+  *(int *)(s + 0 * p - 2) = _mm_cvtsi128_si32(ps1ps0);
+  ps1ps0 = _mm_srli_si128(ps1ps0, 4);
+  *(int *)(s + 1 * p - 2) = _mm_cvtsi128_si32(ps1ps0);
+  ps1ps0 = _mm_srli_si128(ps1ps0, 4);
+  *(int *)(s + 2 * p - 2) = _mm_cvtsi128_si32(ps1ps0);
+  ps1ps0 = _mm_srli_si128(ps1ps0, 4);
+  *(int *)(s + 3 * p - 2) = _mm_cvtsi128_si32(ps1ps0);
+
+  *(int *)(s + 4 * p - 2) = _mm_cvtsi128_si32(qs1qs0);
+  qs1qs0 = _mm_srli_si128(qs1qs0, 4);
+  *(int *)(s + 5 * p - 2) = _mm_cvtsi128_si32(qs1qs0);
+  qs1qs0 = _mm_srli_si128(qs1qs0, 4);
+  *(int *)(s + 6 * p - 2) = _mm_cvtsi128_si32(qs1qs0);
+  qs1qs0 = _mm_srli_si128(qs1qs0, 4);
+  *(int *)(s + 7 * p - 2) = _mm_cvtsi128_si32(qs1qs0);
+}
+
+void aom_lpf_horizontal_edge_8_sse2(unsigned char *s, int p,
+                                    const unsigned char *_blimit,
+                                    const unsigned char *_limit,
+                                    const unsigned char *_thresh) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+  __m128i mask, hev, flat, flat2;
+  __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
+  __m128i abs_p1p0;
+
+  q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
+  q4p4 = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(q4p4), (__m64 *)(s + 4 * p)));
+  q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
+  q3p3 = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(q3p3), (__m64 *)(s + 3 * p)));
+  q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+  q2p2 = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(q2p2), (__m64 *)(s + 2 * p)));
+  q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  q1p1 = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(q1p1), (__m64 *)(s + 1 * p)));
+  p1q1 = _mm_shuffle_epi32(q1p1, 78);
+  q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  q0p0 = _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(q0p0), (__m64 *)(s - 0 * p)));
+  p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+  {
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
+    abs_p1p0 = abs_diff(q1p1, q0p0);
+    abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+    fe = _mm_set1_epi8(0xfe);
+    ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    abs_p0q0 = abs_diff(q0p0, p0q0);
+    abs_p1q1 = abs_diff(q1p1, p1q1);
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = _mm_max_epu8(abs_diff(q2p2, q1p1), abs_diff(q3p3, q2p2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i t1 = _mm_set1_epi16(0x1);
+    __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
+    __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
+    __m128i qs0 = _mm_xor_si128(p0q0, t80);
+    __m128i qs1 = _mm_xor_si128(p1q1, t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+    __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+    __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
+
+    filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, qs0ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (aom_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    filter1 = _mm_unpacklo_epi8(zero, filter1);
+    filter1 = _mm_srai_epi16(filter1, 0xB);
+    filter2 = _mm_unpacklo_epi8(zero, filter2);
+    filter2 = _mm_srai_epi16(filter2, 0xB);
+
+    // Filter1 >> 3
+    filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
+    qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
+
+    // filt >> 1
+    filt = _mm_adds_epi16(filter1, t1);
+    filt = _mm_srai_epi16(filt, 1);
+    filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8),
+                            filt);
+    filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
+    qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
+    // loopfilter done
+
+    {
+      __m128i work;
+      flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0));
+      flat = _mm_max_epu8(abs_p1p0, flat);
+      flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+      flat = _mm_subs_epu8(flat, one);
+      flat = _mm_cmpeq_epi8(flat, zero);
+      flat = _mm_and_si128(flat, mask);
+
+      q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
+      q5p5 = _mm_castps_si128(
+          _mm_loadh_pi(_mm_castsi128_ps(q5p5), (__m64 *)(s + 5 * p)));
+
+      q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
+      q6p6 = _mm_castps_si128(
+          _mm_loadh_pi(_mm_castsi128_ps(q6p6), (__m64 *)(s + 6 * p)));
+      flat2 = _mm_max_epu8(abs_diff(q4p4, q0p0), abs_diff(q5p5, q0p0));
+
+      q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
+      q7p7 = _mm_castps_si128(
+          _mm_loadh_pi(_mm_castsi128_ps(q7p7), (__m64 *)(s + 7 * p)));
+      work = _mm_max_epu8(abs_diff(q6p6, q0p0), abs_diff(q7p7, q0p0));
+      flat2 = _mm_max_epu8(work, flat2);
+      flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
+      flat2 = _mm_subs_epu8(flat2, one);
+      flat2 = _mm_cmpeq_epi8(flat2, zero);
+      flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+    }
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // flat and wide flat calculations
+    {
+      const __m128i eight = _mm_set1_epi16(8);
+      const __m128i four = _mm_set1_epi16(4);
+      __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+      __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+      __m128i pixelFilter_p, pixelFilter_q;
+      __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+      __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
+
+      p7_16 = _mm_unpacklo_epi8(q7p7, zero);
+      p6_16 = _mm_unpacklo_epi8(q6p6, zero);
+      p5_16 = _mm_unpacklo_epi8(q5p5, zero);
+      p4_16 = _mm_unpacklo_epi8(q4p4, zero);
+      p3_16 = _mm_unpacklo_epi8(q3p3, zero);
+      p2_16 = _mm_unpacklo_epi8(q2p2, zero);
+      p1_16 = _mm_unpacklo_epi8(q1p1, zero);
+      p0_16 = _mm_unpacklo_epi8(q0p0, zero);
+      q0_16 = _mm_unpackhi_epi8(q0p0, zero);
+      q1_16 = _mm_unpackhi_epi8(q1p1, zero);
+      q2_16 = _mm_unpackhi_epi8(q2p2, zero);
+      q3_16 = _mm_unpackhi_epi8(q3p3, zero);
+      q4_16 = _mm_unpackhi_epi8(q4p4, zero);
+      q5_16 = _mm_unpackhi_epi8(q5p5, zero);
+      q6_16 = _mm_unpackhi_epi8(q6p6, zero);
+      q7_16 = _mm_unpackhi_epi8(q7p7, zero);
+
+      pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
+                                    _mm_add_epi16(p4_16, p3_16));
+      pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
+                                    _mm_add_epi16(q4_16, q3_16));
+
+      pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
+      pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+      pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16));
+      pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+      pixelFilter_p =
+          _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q));
+      pixetFilter_p2p1p0 = _mm_add_epi16(
+          four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)), 4);
+      flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3_16, p0_16)), 3);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3_16, q0_16)), 3);
+
+      flat_q0p0 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(p7_16, p7_16);
+      sum_q7 = _mm_add_epi16(q7_16, q7_16);
+      sum_p3 = _mm_add_epi16(p3_16, p3_16);
+      sum_q3 = _mm_add_epi16(q3_16, q3_16);
+
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)), 4);
+      flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+      pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
+      pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1_16)), 3);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1_16)), 3);
+      flat_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      sum_p3 = _mm_add_epi16(sum_p3, p3_16);
+      sum_q3 = _mm_add_epi16(sum_q3, q3_16);
+
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)), 4);
+      flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+      pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
+      pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
+
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2_16)), 3);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2_16)), 3);
+      flat_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)), 4);
+      flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)), 4);
+      flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)), 4);
+      flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+      sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6_16)), 4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)), 4);
+      flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
+    }
+    // wide flat
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+    flat = _mm_shuffle_epi32(flat, 68);
+    flat2 = _mm_shuffle_epi32(flat2, 68);
+
+    q2p2 = _mm_andnot_si128(flat, q2p2);
+    flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
+    q2p2 = _mm_or_si128(q2p2, flat_q2p2);
+
+    qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+    flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
+    q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
+
+    qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+    flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
+    q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
+
+    q6p6 = _mm_andnot_si128(flat2, q6p6);
+    flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
+    q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
+    _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6);
+    _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6));
+
+    q5p5 = _mm_andnot_si128(flat2, q5p5);
+    flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
+    q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
+    _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5);
+    _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5));
+
+    q4p4 = _mm_andnot_si128(flat2, q4p4);
+    flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
+    q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
+    _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4);
+    _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4));
+
+    q3p3 = _mm_andnot_si128(flat2, q3p3);
+    flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
+    q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
+    _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3);
+    _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3));
+
+    q2p2 = _mm_andnot_si128(flat2, q2p2);
+    flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
+    q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
+    _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2);
+    _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2));
+
+    q1p1 = _mm_andnot_si128(flat2, q1p1);
+    flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
+    q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
+    _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1);
+    _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1));
+
+    q0p0 = _mm_andnot_si128(flat2, q0p0);
+    flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
+    q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
+    _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0);
+    _mm_storeh_pi((__m64 *)(s - 0 * p), _mm_castsi128_ps(q0p0));
+  }
+}
+
+static INLINE __m128i filter_add2_sub2(const __m128i *const total,
+                                       const __m128i *const a1,
+                                       const __m128i *const a2,
+                                       const __m128i *const s1,
+                                       const __m128i *const s2) {
+  __m128i x = _mm_add_epi16(*a1, *total);
+  x = _mm_add_epi16(_mm_sub_epi16(x, _mm_add_epi16(*s1, *s2)), *a2);
+  return x;
+}
+
+static INLINE __m128i filter8_mask(const __m128i *const flat,
+                                   const __m128i *const other_filt,
+                                   const __m128i *const f8_lo,
+                                   const __m128i *const f8_hi) {
+  const __m128i f8 =
+      _mm_packus_epi16(_mm_srli_epi16(*f8_lo, 3), _mm_srli_epi16(*f8_hi, 3));
+  const __m128i result = _mm_and_si128(*flat, f8);
+  return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
+}
+
+static INLINE __m128i filter16_mask(const __m128i *const flat,
+                                    const __m128i *const other_filt,
+                                    const __m128i *const f_lo,
+                                    const __m128i *const f_hi) {
+  const __m128i f =
+      _mm_packus_epi16(_mm_srli_epi16(*f_lo, 4), _mm_srli_epi16(*f_hi, 4));
+  const __m128i result = _mm_and_si128(*flat, f);
+  return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
+}
+
+void aom_lpf_horizontal_edge_16_sse2(unsigned char *s, int p,
+                                     const unsigned char *_blimit,
+                                     const unsigned char *_limit,
+                                     const unsigned char *_thresh) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+  __m128i mask, hev, flat, flat2;
+  __m128i p7, p6, p5;
+  __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
+  __m128i q5, q6, q7;
+
+  __m128i op2, op1, op0, oq0, oq1, oq2;
+
+  __m128i max_abs_p1p0q1q0;
+
+  p7 = _mm_loadu_si128((__m128i *)(s - 8 * p));
+  p6 = _mm_loadu_si128((__m128i *)(s - 7 * p));
+  p5 = _mm_loadu_si128((__m128i *)(s - 6 * p));
+  p4 = _mm_loadu_si128((__m128i *)(s - 5 * p));
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  q4 = _mm_loadu_si128((__m128i *)(s + 4 * p));
+  q5 = _mm_loadu_si128((__m128i *)(s + 5 * p));
+  q6 = _mm_loadu_si128((__m128i *)(s + 6 * p));
+  q7 = _mm_loadu_si128((__m128i *)(s + 7 * p));
+
+  {
+    const __m128i abs_p1p0 = abs_diff(p1, p0);
+    const __m128i abs_q1q0 = abs_diff(q1, q0);
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+    __m128i abs_p0q0 = abs_diff(p0, q0);
+    __m128i abs_p1q1 = abs_diff(p1, q1);
+    __m128i work;
+    max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(max_abs_p1p0q1q0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(p3, p2));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(abs_diff(q2, q1), abs_diff(q3, q2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  {
+    __m128i work;
+    work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0));
+    flat = _mm_max_epu8(work, max_abs_p1p0q1q0);
+    work = _mm_max_epu8(abs_diff(p3, p0), abs_diff(q3, q0));
+    flat = _mm_max_epu8(work, flat);
+    work = _mm_max_epu8(abs_diff(p4, p0), abs_diff(q4, q0));
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+    flat2 = _mm_max_epu8(abs_diff(p5, p0), abs_diff(q5, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    work = _mm_max_epu8(abs_diff(p6, p0), abs_diff(q6, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    work = _mm_max_epu8(abs_diff(p7, p0), abs_diff(q7, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    flat2 = _mm_subs_epu8(flat2, one);
+    flat2 = _mm_cmpeq_epi8(flat2, zero);
+    flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+  }
+
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // filter4
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i te0 = _mm_set1_epi8(0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+    const __m128i ff = _mm_cmpeq_epi8(t4, t4);
+
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    op1 = _mm_xor_si128(p1, t80);
+    op0 = _mm_xor_si128(p0, t80);
+    oq0 = _mm_xor_si128(q0, t80);
+    oq1 = _mm_xor_si128(q1, t80);
+
+    hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev);
+
+    work_a = _mm_subs_epi8(oq0, op0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (aom_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80);
+
+    // Filter2 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80);
+    oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80);
+    // loopfilter done
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // filter8
+    {
+      const __m128i four = _mm_set1_epi16(4);
+      const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+      const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+      const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+      const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+      const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+      const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+      const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+      const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+
+      const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+      const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+      const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+      const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+      const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+      const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+      const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+      const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+      __m128i f8_lo, f8_hi;
+
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, four),
+                            _mm_add_epi16(p3_lo, p2_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f8_lo),
+                            _mm_add_epi16(p2_lo, p1_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo);
+
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, four),
+                            _mm_add_epi16(p3_hi, p2_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f8_hi),
+                            _mm_add_epi16(p2_hi, p1_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi);
+
+      op2 = filter8_mask(&flat, &p2, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q1_lo, &p1_lo, &p2_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q1_hi, &p1_hi, &p2_hi, &p3_hi);
+      op1 = filter8_mask(&flat, &op1, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q2_lo, &p0_lo, &p1_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q2_hi, &p0_hi, &p1_hi, &p3_hi);
+      op0 = filter8_mask(&flat, &op0, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q0_lo, &p0_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q0_hi, &p0_hi, &p3_hi);
+      oq0 = filter8_mask(&flat, &oq0, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q1_lo, &q0_lo, &p2_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q1_hi, &q0_hi, &p2_hi);
+      oq1 = filter8_mask(&flat, &oq1, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q2_lo, &q1_lo, &p1_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q2_hi, &q1_hi, &p1_hi);
+      oq2 = filter8_mask(&flat, &q2, &f8_lo, &f8_hi);
+    }
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // wide flat calculations
+    {
+      const __m128i eight = _mm_set1_epi16(8);
+      const __m128i p7_lo = _mm_unpacklo_epi8(p7, zero);
+      const __m128i p6_lo = _mm_unpacklo_epi8(p6, zero);
+      const __m128i p5_lo = _mm_unpacklo_epi8(p5, zero);
+      const __m128i p4_lo = _mm_unpacklo_epi8(p4, zero);
+      const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+      const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+      const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+      const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+      const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+      const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+      const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+      const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+      const __m128i q4_lo = _mm_unpacklo_epi8(q4, zero);
+      const __m128i q5_lo = _mm_unpacklo_epi8(q5, zero);
+      const __m128i q6_lo = _mm_unpacklo_epi8(q6, zero);
+      const __m128i q7_lo = _mm_unpacklo_epi8(q7, zero);
+
+      const __m128i p7_hi = _mm_unpackhi_epi8(p7, zero);
+      const __m128i p6_hi = _mm_unpackhi_epi8(p6, zero);
+      const __m128i p5_hi = _mm_unpackhi_epi8(p5, zero);
+      const __m128i p4_hi = _mm_unpackhi_epi8(p4, zero);
+      const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+      const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+      const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+      const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+      const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+      const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+      const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+      const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+      const __m128i q4_hi = _mm_unpackhi_epi8(q4, zero);
+      const __m128i q5_hi = _mm_unpackhi_epi8(q5, zero);
+      const __m128i q6_hi = _mm_unpackhi_epi8(q6, zero);
+      const __m128i q7_hi = _mm_unpackhi_epi8(q7, zero);
+
+      __m128i f_lo;
+      __m128i f_hi;
+
+      f_lo = _mm_sub_epi16(_mm_slli_epi16(p7_lo, 3), p7_lo);  // p7 * 7
+      f_lo =
+          _mm_add_epi16(_mm_slli_epi16(p6_lo, 1), _mm_add_epi16(p4_lo, f_lo));
+      f_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f_lo),
+                           _mm_add_epi16(p2_lo, p1_lo));
+      f_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f_lo);
+      f_lo = _mm_add_epi16(_mm_add_epi16(p5_lo, eight), f_lo);
+
+      f_hi = _mm_sub_epi16(_mm_slli_epi16(p7_hi, 3), p7_hi);  // p7 * 7
+      f_hi =
+          _mm_add_epi16(_mm_slli_epi16(p6_hi, 1), _mm_add_epi16(p4_hi, f_hi));
+      f_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f_hi),
+                           _mm_add_epi16(p2_hi, p1_hi));
+      f_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f_hi);
+      f_hi = _mm_add_epi16(_mm_add_epi16(p5_hi, eight), f_hi);
+
+      p6 = filter16_mask(&flat2, &p6, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 7 * p), p6);
+
+      f_lo = filter_add2_sub2(&f_lo, &q1_lo, &p5_lo, &p6_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q1_hi, &p5_hi, &p6_hi, &p7_hi);
+      p5 = filter16_mask(&flat2, &p5, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 6 * p), p5);
+
+      f_lo = filter_add2_sub2(&f_lo, &q2_lo, &p4_lo, &p5_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q2_hi, &p4_hi, &p5_hi, &p7_hi);
+      p4 = filter16_mask(&flat2, &p4, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 5 * p), p4);
+
+      f_lo = filter_add2_sub2(&f_lo, &q3_lo, &p3_lo, &p4_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q3_hi, &p3_hi, &p4_hi, &p7_hi);
+      p3 = filter16_mask(&flat2, &p3, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 4 * p), p3);
+
+      f_lo = filter_add2_sub2(&f_lo, &q4_lo, &p2_lo, &p3_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q4_hi, &p2_hi, &p3_hi, &p7_hi);
+      op2 = filter16_mask(&flat2, &op2, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 3 * p), op2);
+
+      f_lo = filter_add2_sub2(&f_lo, &q5_lo, &p1_lo, &p2_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q5_hi, &p1_hi, &p2_hi, &p7_hi);
+      op1 = filter16_mask(&flat2, &op1, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+
+      f_lo = filter_add2_sub2(&f_lo, &q6_lo, &p0_lo, &p1_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q6_hi, &p0_hi, &p1_hi, &p7_hi);
+      op0 = filter16_mask(&flat2, &op0, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q0_lo, &p0_lo, &p7_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q0_hi, &p0_hi, &p7_hi);
+      oq0 = filter16_mask(&flat2, &oq0, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q1_lo, &p6_lo, &q0_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q1_hi, &p6_hi, &q0_hi);
+      oq1 = filter16_mask(&flat2, &oq1, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q2_lo, &p5_lo, &q1_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q2_hi, &p5_hi, &q1_hi);
+      oq2 = filter16_mask(&flat2, &oq2, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 2 * p), oq2);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q3_lo, &p4_lo, &q2_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q3_hi, &p4_hi, &q2_hi);
+      q3 = filter16_mask(&flat2, &q3, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 3 * p), q3);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q4_lo, &p3_lo, &q3_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q4_hi, &p3_hi, &q3_hi);
+      q4 = filter16_mask(&flat2, &q4, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 4 * p), q4);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q5_lo, &p2_lo, &q4_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q5_hi, &p2_hi, &q4_hi);
+      q5 = filter16_mask(&flat2, &q5, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 5 * p), q5);
+
+      f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q6_lo, &p1_lo, &q5_lo);
+      f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q6_hi, &p1_hi, &q5_hi);
+      q6 = filter16_mask(&flat2, &q6, &f_lo, &f_hi);
+      _mm_storeu_si128((__m128i *)(s + 6 * p), q6);
+    }
+    // wide flat
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  }
+}
+
+void aom_lpf_horizontal_8_sse2(unsigned char *s, int p,
+                               const unsigned char *_blimit,
+                               const unsigned char *_limit,
+                               const unsigned char *_thresh) {
+  DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+  __m128i mask, hev, flat;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+  __m128i q3p3, q2p2, q1p1, q0p0, p1q1, p0q0;
+
+  q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 3 * p)));
+  q2p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 2 * p)));
+  q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 1 * p)));
+  q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+                            _mm_loadl_epi64((__m128i *)(s - 0 * p)));
+  p1q1 = _mm_shuffle_epi32(q1p1, 78);
+  p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+  {
+    // filter_mask and hev_mask
+    const __m128i one = _mm_set1_epi8(1);
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(fe, fe);
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+    abs_p1p0 = abs_diff(q1p1, q0p0);
+    abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+
+    abs_p0q0 = abs_diff(q0p0, p0q0);
+    abs_p1q1 = abs_diff(q1p1, p1q1);
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = _mm_max_epu8(abs_diff(q2p2, q1p1), abs_diff(q3p3, q2p2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+
+    // flat_mask4
+
+    flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0));
+    flat = _mm_max_epu8(abs_p1p0, flat);
+    flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+  }
+
+  {
+    const __m128i four = _mm_set1_epi16(4);
+    unsigned char *src = s;
+    {
+      __m128i workp_a, workp_b, workp_shft;
+      p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
+      p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
+      p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
+      p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
+      q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
+      q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
+      q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
+      q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
+
+      workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+      workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op2[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op1[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op0[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq0[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq1[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq2[0],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+    }
+  }
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i ps1 =
+        _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 2 * p)), t80);
+    const __m128i ps0 =
+        _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 1 * p)), t80);
+    const __m128i qs0 =
+        _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 0 * p)), t80);
+    const __m128i qs1 =
+        _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 1 * p)), t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (aom_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    filter1 = _mm_unpacklo_epi8(zero, filter1);
+    filter1 = _mm_srai_epi16(filter1, 11);
+    filter1 = _mm_packs_epi16(filter1, filter1);
+
+    // Filter2 >> 3
+    filter2 = _mm_unpacklo_epi8(zero, filter2);
+    filter2 = _mm_srai_epi16(filter2, 11);
+    filter2 = _mm_packs_epi16(filter2, zero);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    filt = _mm_unpacklo_epi8(zero, filt);
+    filt = _mm_srai_epi16(filt, 9);
+    filt = _mm_packs_epi16(filt, zero);
+
+    filt = _mm_andnot_si128(hev, filt);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+    q0 = _mm_loadl_epi64((__m128i *)flat_oq0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q0 = _mm_and_si128(flat, q0);
+    q0 = _mm_or_si128(work_a, q0);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    q1 = _mm_loadl_epi64((__m128i *)flat_oq1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q1 = _mm_and_si128(flat, q1);
+    q1 = _mm_or_si128(work_a, q1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+    q2 = _mm_loadl_epi64((__m128i *)flat_oq2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q2 = _mm_and_si128(flat, q2);
+    q2 = _mm_or_si128(work_a, q2);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+    p0 = _mm_loadl_epi64((__m128i *)flat_op0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p0 = _mm_and_si128(flat, p0);
+    p0 = _mm_or_si128(work_a, p0);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+    p1 = _mm_loadl_epi64((__m128i *)flat_op1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p1 = _mm_and_si128(flat, p1);
+    p1 = _mm_or_si128(work_a, p1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+    p2 = _mm_loadl_epi64((__m128i *)flat_op2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p2 = _mm_and_si128(flat, p2);
+    p2 = _mm_or_si128(work_a, p2);
+
+    _mm_storel_epi64((__m128i *)(s - 3 * p), p2);
+    _mm_storel_epi64((__m128i *)(s - 2 * p), p1);
+    _mm_storel_epi64((__m128i *)(s - 1 * p), p0);
+    _mm_storel_epi64((__m128i *)(s + 0 * p), q0);
+    _mm_storel_epi64((__m128i *)(s + 1 * p), q1);
+    _mm_storel_epi64((__m128i *)(s + 2 * p), q2);
+  }
+}
+
+void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int p, const uint8_t *_blimit0,
+                                    const uint8_t *_limit0,
+                                    const uint8_t *_thresh0,
+                                    const uint8_t *_blimit1,
+                                    const uint8_t *_limit1,
+                                    const uint8_t *_thresh1) {
+  DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i blimit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0),
+                         _mm_load_si128((const __m128i *)_blimit1));
+  const __m128i limit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0),
+                         _mm_load_si128((const __m128i *)_limit1));
+  const __m128i thresh =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0),
+                         _mm_load_si128((const __m128i *)_thresh1));
+
+  __m128i mask, hev, flat;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  {
+    const __m128i abs_p1p0 =
+        _mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1));
+    const __m128i abs_q1q0 =
+        _mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1));
+    const __m128i one = _mm_set1_epi8(1);
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    __m128i abs_p0q0 =
+        _mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0));
+    __m128i abs_p1q1 =
+        _mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1));
+    __m128i work;
+
+    // filter_mask and hev_mask
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(flat, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
+        _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
+        _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+
+    // flat_mask4
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
+        _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
+    flat = _mm_max_epu8(work, flat);
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)),
+        _mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3)));
+    flat = _mm_max_epu8(work, flat);
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+  }
+  {
+    const __m128i four = _mm_set1_epi16(4);
+    unsigned char *src = s;
+    int i = 0;
+
+    do {
+      __m128i workp_a, workp_b, workp_shft;
+      p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
+      p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
+      p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
+      p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
+      q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
+      q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
+      q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
+      q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
+
+      workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+      workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op2[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op1[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_op0[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq0[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq1[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+      _mm_storel_epi64((__m128i *)&flat_oq2[i * 8],
+                       _mm_packus_epi16(workp_shft, workp_shft));
+
+      src += 8;
+    } while (++i < 2);
+  }
+  // lp filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i te0 = _mm_set1_epi8(0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+
+    const __m128i ps1 =
+        _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80);
+    const __m128i ps0 =
+        _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80);
+    const __m128i qs0 =
+        _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80);
+    const __m128i qs1 =
+        _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (aom_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+
+    // Filter2 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+
+    filt = _mm_andnot_si128(hev, filt);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+    q0 = _mm_load_si128((__m128i *)flat_oq0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q0 = _mm_and_si128(flat, q0);
+    q0 = _mm_or_si128(work_a, q0);
+
+    work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    q1 = _mm_load_si128((__m128i *)flat_oq1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q1 = _mm_and_si128(flat, q1);
+    q1 = _mm_or_si128(work_a, q1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+    q2 = _mm_load_si128((__m128i *)flat_oq2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    q2 = _mm_and_si128(flat, q2);
+    q2 = _mm_or_si128(work_a, q2);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+    p0 = _mm_load_si128((__m128i *)flat_op0);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p0 = _mm_and_si128(flat, p0);
+    p0 = _mm_or_si128(work_a, p0);
+
+    work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+    p1 = _mm_load_si128((__m128i *)flat_op1);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p1 = _mm_and_si128(flat, p1);
+    p1 = _mm_or_si128(work_a, p1);
+
+    work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+    p2 = _mm_load_si128((__m128i *)flat_op2);
+    work_a = _mm_andnot_si128(flat, work_a);
+    p2 = _mm_and_si128(flat, p2);
+    p2 = _mm_or_si128(work_a, p2);
+
+    _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+    _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+    _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+    _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+    _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+    _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+  }
+}
+
+void aom_lpf_horizontal_4_dual_sse2(unsigned char *s, int p,
+                                    const unsigned char *_blimit0,
+                                    const unsigned char *_limit0,
+                                    const unsigned char *_thresh0,
+                                    const unsigned char *_blimit1,
+                                    const unsigned char *_limit1,
+                                    const unsigned char *_thresh1) {
+  const __m128i blimit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0),
+                         _mm_load_si128((const __m128i *)_blimit1));
+  const __m128i limit =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0),
+                         _mm_load_si128((const __m128i *)_limit1));
+  const __m128i thresh =
+      _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0),
+                         _mm_load_si128((const __m128i *)_thresh1));
+  const __m128i zero = _mm_set1_epi16(0);
+#if !CONFIG_PARALLEL_DEBLOCKING
+  __m128i p3, p2, q2, q3;
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  __m128i p1, p0, q0, q1;
+  __m128i mask, hev, flat;
+#if !CONFIG_PARALLEL_DEBLOCKING
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+#if !CONFIG_PARALLEL_DEBLOCKING
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  // filter_mask and hev_mask
+  {
+    const __m128i abs_p1p0 =
+        _mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1));
+    const __m128i abs_q1q0 =
+        _mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1));
+    const __m128i fe = _mm_set1_epi8(0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    __m128i abs_p0q0 =
+        _mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0));
+    __m128i abs_p1q1 =
+        _mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1));
+#if !CONFIG_PARALLEL_DEBLOCKING
+    __m128i work;
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(flat, mask);
+#if !CONFIG_PARALLEL_DEBLOCKING
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
+        _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
+        _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
+    mask = _mm_max_epu8(work, mask);
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+    mask = _mm_subs_epu8(mask, limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  // filter4
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8(0x80);
+    const __m128i te0 = _mm_set1_epi8(0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+
+    const __m128i ps1 =
+        _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80);
+    const __m128i ps0 =
+        _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80);
+    const __m128i qs0 =
+        _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80);
+    const __m128i qs1 =
+        _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    // (aom_filter + 3 * (qs0 - ps0)) & mask
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    // Filter1 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+
+    // Filter2 >> 3
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+
+    // filt >> 1
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+
+    filt = _mm_andnot_si128(hev, filt);
+
+    q0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+    q1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    p0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+    p1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+
+    _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+    _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+    _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+    _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+  }
+}
+
+static INLINE void transpose8x16(unsigned char *in0, unsigned char *in1,
+                                 int in_p, unsigned char *out, int out_p) {
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i x8, x9, x10, x11, x12, x13, x14, x15;
+
+  // 2-way interleave w/hoisting of unpacks
+  x0 = _mm_loadl_epi64((__m128i *)in0);           // 1
+  x1 = _mm_loadl_epi64((__m128i *)(in0 + in_p));  // 3
+  x0 = _mm_unpacklo_epi8(x0, x1);                 // 1
+
+  x2 = _mm_loadl_epi64((__m128i *)(in0 + 2 * in_p));  // 5
+  x3 = _mm_loadl_epi64((__m128i *)(in0 + 3 * in_p));  // 7
+  x1 = _mm_unpacklo_epi8(x2, x3);                     // 2
+
+  x4 = _mm_loadl_epi64((__m128i *)(in0 + 4 * in_p));  // 9
+  x5 = _mm_loadl_epi64((__m128i *)(in0 + 5 * in_p));  // 11
+  x2 = _mm_unpacklo_epi8(x4, x5);                     // 3
+
+  x6 = _mm_loadl_epi64((__m128i *)(in0 + 6 * in_p));  // 13
+  x7 = _mm_loadl_epi64((__m128i *)(in0 + 7 * in_p));  // 15
+  x3 = _mm_unpacklo_epi8(x6, x7);                     // 4
+  x4 = _mm_unpacklo_epi16(x0, x1);                    // 9
+
+  x8 = _mm_loadl_epi64((__m128i *)in1);           // 2
+  x9 = _mm_loadl_epi64((__m128i *)(in1 + in_p));  // 4
+  x8 = _mm_unpacklo_epi8(x8, x9);                 // 5
+  x5 = _mm_unpacklo_epi16(x2, x3);                // 10
+
+  x10 = _mm_loadl_epi64((__m128i *)(in1 + 2 * in_p));  // 6
+  x11 = _mm_loadl_epi64((__m128i *)(in1 + 3 * in_p));  // 8
+  x9 = _mm_unpacklo_epi8(x10, x11);                    // 6
+
+  x12 = _mm_loadl_epi64((__m128i *)(in1 + 4 * in_p));  // 10
+  x13 = _mm_loadl_epi64((__m128i *)(in1 + 5 * in_p));  // 12
+  x10 = _mm_unpacklo_epi8(x12, x13);                   // 7
+  x12 = _mm_unpacklo_epi16(x8, x9);                    // 11
+
+  x14 = _mm_loadl_epi64((__m128i *)(in1 + 6 * in_p));  // 14
+  x15 = _mm_loadl_epi64((__m128i *)(in1 + 7 * in_p));  // 16
+  x11 = _mm_unpacklo_epi8(x14, x15);                   // 8
+  x13 = _mm_unpacklo_epi16(x10, x11);                  // 12
+
+  x6 = _mm_unpacklo_epi32(x4, x5);     // 13
+  x7 = _mm_unpackhi_epi32(x4, x5);     // 14
+  x14 = _mm_unpacklo_epi32(x12, x13);  // 15
+  x15 = _mm_unpackhi_epi32(x12, x13);  // 16
+
+  // Store first 4-line result
+  _mm_storeu_si128((__m128i *)out, _mm_unpacklo_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + out_p), _mm_unpackhi_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 2 * out_p), _mm_unpacklo_epi64(x7, x15));
+  _mm_storeu_si128((__m128i *)(out + 3 * out_p), _mm_unpackhi_epi64(x7, x15));
+
+  x4 = _mm_unpackhi_epi16(x0, x1);
+  x5 = _mm_unpackhi_epi16(x2, x3);
+  x12 = _mm_unpackhi_epi16(x8, x9);
+  x13 = _mm_unpackhi_epi16(x10, x11);
+
+  x6 = _mm_unpacklo_epi32(x4, x5);
+  x7 = _mm_unpackhi_epi32(x4, x5);
+  x14 = _mm_unpacklo_epi32(x12, x13);
+  x15 = _mm_unpackhi_epi32(x12, x13);
+
+  // Store second 4-line result
+  _mm_storeu_si128((__m128i *)(out + 4 * out_p), _mm_unpacklo_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 5 * out_p), _mm_unpackhi_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 6 * out_p), _mm_unpacklo_epi64(x7, x15));
+  _mm_storeu_si128((__m128i *)(out + 7 * out_p), _mm_unpackhi_epi64(x7, x15));
+}
+
+#if CONFIG_PARALLEL_DEBLOCKING
+#define movq(p) _mm_loadl_epi64((const __m128i *)(p))
+#define punpcklbw(r0, r1) _mm_unpacklo_epi8(r0, r1)
+#define punpcklwd(r0, r1) _mm_unpacklo_epi16(r0, r1)
+#define punpckhwd(r0, r1) _mm_unpackhi_epi16(r0, r1)
+#define movd(p, r) *((uint32_t *)(p)) = _mm_cvtsi128_si32(r)
+#define pshufd(r, imm) _mm_shuffle_epi32(r, imm)
+enum { ROTATE_DWORD_RIGHT = 0x39 };
+static INLINE void transpose16x4(uint8_t *pDst, const ptrdiff_t dstStride,
+                                 const uint8_t *pSrc,
+                                 const ptrdiff_t srcStride) {
+  for (uint32_t idx = 0; idx < 2; idx += 1) {
+    __m128i r0, r1, r2, r3;
+    // load data
+    r0 = movq(pSrc);
+    r1 = movq(pSrc + srcStride);
+    r2 = movq(pSrc + srcStride * 2);
+    r3 = movq(pSrc + srcStride * 3);
+    // transpose
+    r0 = punpcklbw(r0, r1);
+    r2 = punpcklbw(r2, r3);
+    r1 = punpckhwd(r0, r2);
+    r0 = punpcklwd(r0, r2);
+    // store data
+    movd(pDst, r0);
+    r0 = pshufd(r0, ROTATE_DWORD_RIGHT);
+    movd(pDst + dstStride, r0);
+    r0 = pshufd(r0, ROTATE_DWORD_RIGHT);
+    movd(pDst + dstStride * 2, r0);
+    r0 = pshufd(r0, ROTATE_DWORD_RIGHT);
+    movd(pDst + dstStride * 3, r0);
+    movd(pDst + dstStride * 4, r1);
+    r1 = pshufd(r1, ROTATE_DWORD_RIGHT);
+    movd(pDst + dstStride * 5, r1);
+    r1 = pshufd(r1, ROTATE_DWORD_RIGHT);
+    movd(pDst + dstStride * 6, r1);
+    r1 = pshufd(r1, ROTATE_DWORD_RIGHT);
+    movd(pDst + dstStride * 7, r1);
+    // advance the pointers
+    pDst += dstStride * 8;
+    pSrc += 8;
+  }
+}
+
+#endif  // CONFIG_PARALLEL_DEBLOCKING
+static INLINE void transpose(unsigned char *src[], int in_p,
+                             unsigned char *dst[], int out_p,
+                             int num_8x8_to_transpose) {
+  int idx8x8 = 0;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  do {
+    unsigned char *in = src[idx8x8];
+    unsigned char *out = dst[idx8x8];
+
+    x0 =
+        _mm_loadl_epi64((__m128i *)(in + 0 * in_p));  // 00 01 02 03 04 05 06 07
+    x1 =
+        _mm_loadl_epi64((__m128i *)(in + 1 * in_p));  // 10 11 12 13 14 15 16 17
+    // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+    x0 = _mm_unpacklo_epi8(x0, x1);
+
+    x2 =
+        _mm_loadl_epi64((__m128i *)(in + 2 * in_p));  // 20 21 22 23 24 25 26 27
+    x3 =
+        _mm_loadl_epi64((__m128i *)(in + 3 * in_p));  // 30 31 32 33 34 35 36 37
+    // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+    x1 = _mm_unpacklo_epi8(x2, x3);
+
+    x4 =
+        _mm_loadl_epi64((__m128i *)(in + 4 * in_p));  // 40 41 42 43 44 45 46 47
+    x5 =
+        _mm_loadl_epi64((__m128i *)(in + 5 * in_p));  // 50 51 52 53 54 55 56 57
+    // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+    x2 = _mm_unpacklo_epi8(x4, x5);
+
+    x6 =
+        _mm_loadl_epi64((__m128i *)(in + 6 * in_p));  // 60 61 62 63 64 65 66 67
+    x7 =
+        _mm_loadl_epi64((__m128i *)(in + 7 * in_p));  // 70 71 72 73 74 75 76 77
+    // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+    x3 = _mm_unpacklo_epi8(x6, x7);
+
+    // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+    x4 = _mm_unpacklo_epi16(x0, x1);
+    // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73
+    x5 = _mm_unpacklo_epi16(x2, x3);
+    // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
+    x6 = _mm_unpacklo_epi32(x4, x5);
+    _mm_storel_pd((double *)(out + 0 * out_p),
+                  _mm_castsi128_pd(x6));  // 00 10 20 30 40 50 60 70
+    _mm_storeh_pd((double *)(out + 1 * out_p),
+                  _mm_castsi128_pd(x6));  // 01 11 21 31 41 51 61 71
+    // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
+    x7 = _mm_unpackhi_epi32(x4, x5);
+    _mm_storel_pd((double *)(out + 2 * out_p),
+                  _mm_castsi128_pd(x7));  // 02 12 22 32 42 52 62 72
+    _mm_storeh_pd((double *)(out + 3 * out_p),
+                  _mm_castsi128_pd(x7));  // 03 13 23 33 43 53 63 73
+
+    // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+    x4 = _mm_unpackhi_epi16(x0, x1);
+    // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77
+    x5 = _mm_unpackhi_epi16(x2, x3);
+    // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75
+    x6 = _mm_unpacklo_epi32(x4, x5);
+    _mm_storel_pd((double *)(out + 4 * out_p),
+                  _mm_castsi128_pd(x6));  // 04 14 24 34 44 54 64 74
+    _mm_storeh_pd((double *)(out + 5 * out_p),
+                  _mm_castsi128_pd(x6));  // 05 15 25 35 45 55 65 75
+    // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77
+    x7 = _mm_unpackhi_epi32(x4, x5);
+
+    _mm_storel_pd((double *)(out + 6 * out_p),
+                  _mm_castsi128_pd(x7));  // 06 16 26 36 46 56 66 76
+    _mm_storeh_pd((double *)(out + 7 * out_p),
+                  _mm_castsi128_pd(x7));  // 07 17 27 37 47 57 67 77
+  } while (++idx8x8 < num_8x8_to_transpose);
+}
+
+void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0, const uint8_t *thresh0,
+                                  const uint8_t *blimit1, const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+#if !CONFIG_PARALLEL_DEBLOCKING
+  unsigned char *src[2];
+  unsigned char *dst[2];
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+  // Transpose 8x16
+  transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  aom_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,
+                                 blimit1, limit1, thresh1);
+#if !CONFIG_PARALLEL_DEBLOCKING
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  transpose(src, 16, dst, p, 2);
+#else  // CONFIG_PARALLEL_DEBLOCKING
+  transpose16x4(s - 2, p, t_dst + 16 * 2, 16);
+#endif  // !CONFIG_PARALLEL_DEBLOCKING
+}
+
+void aom_lpf_vertical_8_sse2(unsigned char *s, int p,
+                             const unsigned char *blimit,
+                             const unsigned char *limit,
+                             const unsigned char *thresh) {
+  DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 8]);
+  unsigned char *src[1];
+  unsigned char *dst[1];
+
+  // Transpose 8x8
+  src[0] = s - 4;
+  dst[0] = t_dst;
+
+  transpose(src, p, dst, 8, 1);
+
+  // Loop filtering
+  aom_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh);
+
+  src[0] = t_dst;
+  dst[0] = s - 4;
+
+  // Transpose back
+  transpose(src, 8, dst, p, 1);
+}
+
+void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0, const uint8_t *thresh0,
+                                  const uint8_t *blimit1, const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+  unsigned char *src[2];
+  unsigned char *dst[2];
+
+  // Transpose 8x16
+  transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+  // Loop filtering
+  aom_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,
+                                 blimit1, limit1, thresh1);
+  src[0] = t_dst;
+  src[1] = t_dst + 8;
+
+  dst[0] = s - 4;
+  dst[1] = s - 4 + p * 8;
+
+  // Transpose back
+  transpose(src, 16, dst, p, 2);
+}
+
+void aom_lpf_vertical_16_sse2(unsigned char *s, int p,
+                              const unsigned char *blimit,
+                              const unsigned char *limit,
+                              const unsigned char *thresh) {
+  DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 16]);
+  unsigned char *src[2];
+  unsigned char *dst[2];
+
+  src[0] = s - 8;
+  src[1] = s;
+  dst[0] = t_dst;
+  dst[1] = t_dst + 8 * 8;
+
+  // Transpose 16x8
+  transpose(src, p, dst, 8, 2);
+
+  // Loop filtering
+  aom_lpf_horizontal_edge_8_sse2(t_dst + 8 * 8, 8, blimit, limit, thresh);
+
+  src[0] = t_dst;
+  src[1] = t_dst + 8 * 8;
+  dst[0] = s - 8;
+  dst[1] = s;
+
+  // Transpose back
+  transpose(src, 8, dst, p, 2);
+}
+
+void aom_lpf_vertical_16_dual_sse2(unsigned char *s, int p,
+                                   const uint8_t *blimit, const uint8_t *limit,
+                                   const uint8_t *thresh) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[256]);
+
+  // Transpose 16x16
+  transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16);
+  transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16);
+
+  // Loop filtering
+  aom_lpf_horizontal_edge_16_sse2(t_dst + 8 * 16, 16, blimit, limit, thresh);
+
+  // Transpose back
+  transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p);
+  transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p);
+}
diff --git a/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c
new file mode 100644
index 0000000000..5166e9e0af
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c
@@ -0,0 +1,334 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include <emmintrin.h>
+#include <tmmintrin.h>
+
+#include "aom_ports/mem.h"
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+static INLINE __m128i width8_load_2rows(const uint8_t *ptr, int stride) {
+  __m128i temp1 = _mm_loadl_epi64((const __m128i *)ptr);
+  __m128i temp2 = _mm_loadl_epi64((const __m128i *)(ptr + stride));
+  return _mm_unpacklo_epi64(temp1, temp2);
+}
+
+static INLINE __m128i width4_load_4rows(const uint8_t *ptr, int stride) {
+  __m128i temp1 = _mm_cvtsi32_si128(*(const uint32_t *)ptr);
+  __m128i temp2 = _mm_cvtsi32_si128(*(const uint32_t *)(ptr + stride));
+  __m128i temp3 = _mm_unpacklo_epi32(temp1, temp2);
+  temp1 = _mm_cvtsi32_si128(*(const uint32_t *)(ptr + stride * 2));
+  temp2 = _mm_cvtsi32_si128(*(const uint32_t *)(ptr + stride * 3));
+  temp1 = _mm_unpacklo_epi32(temp1, temp2);
+  return _mm_unpacklo_epi64(temp3, temp1);
+}
+
+static INLINE unsigned int masked_sad_ssse3(const uint8_t *a_ptr, int a_stride,
+                                            const uint8_t *b_ptr, int b_stride,
+                                            const uint8_t *m_ptr, int m_stride,
+                                            int width, int height);
+
+static INLINE unsigned int masked_sad8xh_ssse3(
+    const uint8_t *a_ptr, int a_stride, const uint8_t *b_ptr, int b_stride,
+    const uint8_t *m_ptr, int m_stride, int height);
+
+static INLINE unsigned int masked_sad4xh_ssse3(
+    const uint8_t *a_ptr, int a_stride, const uint8_t *b_ptr, int b_stride,
+    const uint8_t *m_ptr, int m_stride, int height);
+
+#define MASKSADMXN_SSSE3(m, n)                                                 \
+  unsigned int aom_masked_sad##m##x##n##_ssse3(                                \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,  \
+      const uint8_t *msk, int msk_stride) {                                    \
+    return masked_sad_ssse3(src, src_stride, ref, ref_stride, msk, msk_stride, \
+                            m, n);                                             \
+  }
+
+#if CONFIG_EXT_PARTITION
+MASKSADMXN_SSSE3(128, 128)
+MASKSADMXN_SSSE3(128, 64)
+MASKSADMXN_SSSE3(64, 128)
+#endif  // CONFIG_EXT_PARTITION
+MASKSADMXN_SSSE3(64, 64)
+MASKSADMXN_SSSE3(64, 32)
+MASKSADMXN_SSSE3(32, 64)
+MASKSADMXN_SSSE3(32, 32)
+MASKSADMXN_SSSE3(32, 16)
+MASKSADMXN_SSSE3(16, 32)
+MASKSADMXN_SSSE3(16, 16)
+MASKSADMXN_SSSE3(16, 8)
+
+#define MASKSAD8XN_SSSE3(n)                                                   \
+  unsigned int aom_masked_sad8x##n##_ssse3(                                   \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      const uint8_t *msk, int msk_stride) {                                   \
+    return masked_sad8xh_ssse3(src, src_stride, ref, ref_stride, msk,         \
+                               msk_stride, n);                                \
+  }
+
+MASKSAD8XN_SSSE3(16)
+MASKSAD8XN_SSSE3(8)
+MASKSAD8XN_SSSE3(4)
+
+#define MASKSAD4XN_SSSE3(n)                                                   \
+  unsigned int aom_masked_sad4x##n##_ssse3(                                   \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      const uint8_t *msk, int msk_stride) {                                   \
+    return masked_sad4xh_ssse3(src, src_stride, ref, ref_stride, msk,         \
+                               msk_stride, n);                                \
+  }
+
+MASKSAD4XN_SSSE3(8)
+MASKSAD4XN_SSSE3(4)
+
+// For width a multiple of 16
+// Assumes values in m are <=64
+static INLINE unsigned int masked_sad_ssse3(const uint8_t *a_ptr, int a_stride,
+                                            const uint8_t *b_ptr, int b_stride,
+                                            const uint8_t *m_ptr, int m_stride,
+                                            int width, int height) {
+  int y, x;
+  __m128i a, b, m, temp1, temp2;
+  __m128i res = _mm_setzero_si128();
+  __m128i one = _mm_set1_epi16(1);
+  // For each row
+  for (y = 0; y < height; y++) {
+    // Covering the full width
+    for (x = 0; x < width; x += 16) {
+      // Load a, b, m in xmm registers
+      a = _mm_loadu_si128((const __m128i *)(a_ptr + x));
+      b = _mm_loadu_si128((const __m128i *)(b_ptr + x));
+      m = _mm_loadu_si128((const __m128i *)(m_ptr + x));
+
+      // Calculate the difference between a & b
+      temp1 = _mm_subs_epu8(a, b);
+      temp2 = _mm_subs_epu8(b, a);
+      temp1 = _mm_or_si128(temp1, temp2);
+
+      // Multiply by m and add together
+      temp2 = _mm_maddubs_epi16(temp1, m);
+      // Pad out row result to 32 bit integers & add to running total
+      res = _mm_add_epi32(res, _mm_madd_epi16(temp2, one));
+    }
+    // Move onto the next row
+    a_ptr += a_stride;
+    b_ptr += b_stride;
+    m_ptr += m_stride;
+  }
+  res = _mm_hadd_epi32(res, _mm_setzero_si128());
+  res = _mm_hadd_epi32(res, _mm_setzero_si128());
+  // sad = (sad + 31) >> 6;
+  return (_mm_cvtsi128_si32(res) + 31) >> 6;
+}
+
+static INLINE unsigned int masked_sad8xh_ssse3(
+    const uint8_t *a_ptr, int a_stride, const uint8_t *b_ptr, int b_stride,
+    const uint8_t *m_ptr, int m_stride, int height) {
+  int y;
+  __m128i a, b, m, temp1, temp2, row_res;
+  __m128i res = _mm_setzero_si128();
+  __m128i one = _mm_set1_epi16(1);
+  // Add the masked SAD for 2 rows at a time
+  for (y = 0; y < height; y += 2) {
+    // Load a, b, m in xmm registers
+    a = width8_load_2rows(a_ptr, a_stride);
+    b = width8_load_2rows(b_ptr, b_stride);
+    m = width8_load_2rows(m_ptr, m_stride);
+
+    // Calculate the difference between a & b
+    temp1 = _mm_subs_epu8(a, b);
+    temp2 = _mm_subs_epu8(b, a);
+    temp1 = _mm_or_si128(temp1, temp2);
+
+    // Multiply by m and add together
+    row_res = _mm_maddubs_epi16(temp1, m);
+
+    // Pad out row result to 32 bit integers & add to running total
+    res = _mm_add_epi32(res, _mm_madd_epi16(row_res, one));
+
+    // Move onto the next rows
+    a_ptr += a_stride * 2;
+    b_ptr += b_stride * 2;
+    m_ptr += m_stride * 2;
+  }
+  res = _mm_hadd_epi32(res, _mm_setzero_si128());
+  res = _mm_hadd_epi32(res, _mm_setzero_si128());
+  // sad = (sad + 31) >> 6;
+  return (_mm_cvtsi128_si32(res) + 31) >> 6;
+}
+
+static INLINE unsigned int masked_sad4xh_ssse3(
+    const uint8_t *a_ptr, int a_stride, const uint8_t *b_ptr, int b_stride,
+    const uint8_t *m_ptr, int m_stride, int height) {
+  int y;
+  __m128i a, b, m, temp1, temp2, row_res;
+  __m128i res = _mm_setzero_si128();
+  __m128i one = _mm_set1_epi16(1);
+  // Add the masked SAD for 4 rows at a time
+  for (y = 0; y < height; y += 4) {
+    // Load a, b, m in xmm registers
+    a = width4_load_4rows(a_ptr, a_stride);
+    b = width4_load_4rows(b_ptr, b_stride);
+    m = width4_load_4rows(m_ptr, m_stride);
+
+    // Calculate the difference between a & b
+    temp1 = _mm_subs_epu8(a, b);
+    temp2 = _mm_subs_epu8(b, a);
+    temp1 = _mm_or_si128(temp1, temp2);
+
+    // Multiply by m and add together
+    row_res = _mm_maddubs_epi16(temp1, m);
+
+    // Pad out row result to 32 bit integers & add to running total
+    res = _mm_add_epi32(res, _mm_madd_epi16(row_res, one));
+
+    // Move onto the next rows
+    a_ptr += a_stride * 4;
+    b_ptr += b_stride * 4;
+    m_ptr += m_stride * 4;
+  }
+  // Pad out row result to 32 bit integers & add to running total
+  res = _mm_hadd_epi32(res, _mm_setzero_si128());
+  res = _mm_hadd_epi32(res, _mm_setzero_si128());
+  // sad = (sad + 31) >> 6;
+  return (_mm_cvtsi128_si32(res) + 31) >> 6;
+}
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE __m128i highbd_width4_load_2rows(const uint16_t *ptr,
+                                               int stride) {
+  __m128i temp1 = _mm_loadl_epi64((const __m128i *)ptr);
+  __m128i temp2 = _mm_loadl_epi64((const __m128i *)(ptr + stride));
+  return _mm_unpacklo_epi64(temp1, temp2);
+}
+
+static INLINE unsigned int highbd_masked_sad_ssse3(
+    const uint8_t *a8_ptr, int a_stride, const uint8_t *b8_ptr, int b_stride,
+    const uint8_t *m_ptr, int m_stride, int width, int height);
+
+static INLINE unsigned int highbd_masked_sad4xh_ssse3(
+    const uint8_t *a8_ptr, int a_stride, const uint8_t *b8_ptr, int b_stride,
+    const uint8_t *m_ptr, int m_stride, int height);
+
+#define HIGHBD_MASKSADMXN_SSSE3(m, n)                                         \
+  unsigned int aom_highbd_masked_sad##m##x##n##_ssse3(                        \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      const uint8_t *msk, int msk_stride) {                                   \
+    return highbd_masked_sad_ssse3(src, src_stride, ref, ref_stride, msk,     \
+                                   msk_stride, m, n);                         \
+  }
+
+#if CONFIG_EXT_PARTITION
+HIGHBD_MASKSADMXN_SSSE3(128, 128)
+HIGHBD_MASKSADMXN_SSSE3(128, 64)
+HIGHBD_MASKSADMXN_SSSE3(64, 128)
+#endif  // CONFIG_EXT_PARTITION
+HIGHBD_MASKSADMXN_SSSE3(64, 64)
+HIGHBD_MASKSADMXN_SSSE3(64, 32)
+HIGHBD_MASKSADMXN_SSSE3(32, 64)
+HIGHBD_MASKSADMXN_SSSE3(32, 32)
+HIGHBD_MASKSADMXN_SSSE3(32, 16)
+HIGHBD_MASKSADMXN_SSSE3(16, 32)
+HIGHBD_MASKSADMXN_SSSE3(16, 16)
+HIGHBD_MASKSADMXN_SSSE3(16, 8)
+HIGHBD_MASKSADMXN_SSSE3(8, 16)
+HIGHBD_MASKSADMXN_SSSE3(8, 8)
+HIGHBD_MASKSADMXN_SSSE3(8, 4)
+
+#define HIGHBD_MASKSAD4XN_SSSE3(n)                                            \
+  unsigned int aom_highbd_masked_sad4x##n##_ssse3(                            \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      const uint8_t *msk, int msk_stride) {                                   \
+    return highbd_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride, msk,  \
+                                      msk_stride, n);                         \
+  }
+
+HIGHBD_MASKSAD4XN_SSSE3(8)
+HIGHBD_MASKSAD4XN_SSSE3(4)
+
+// For width a multiple of 8
+// Assumes values in m are <=64
+static INLINE unsigned int highbd_masked_sad_ssse3(
+    const uint8_t *a8_ptr, int a_stride, const uint8_t *b8_ptr, int b_stride,
+    const uint8_t *m_ptr, int m_stride, int width, int height) {
+  int y, x;
+  __m128i a, b, m, temp1, temp2;
+  const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8_ptr);
+  const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8_ptr);
+  __m128i res = _mm_setzero_si128();
+  // For each row
+  for (y = 0; y < height; y++) {
+    // Covering the full width
+    for (x = 0; x < width; x += 8) {
+      // Load a, b, m in xmm registers
+      a = _mm_loadu_si128((const __m128i *)(a_ptr + x));
+      b = _mm_loadu_si128((const __m128i *)(b_ptr + x));
+      m = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(m_ptr + x)),
+                            _mm_setzero_si128());
+
+      // Calculate the difference between a & b
+      temp1 = _mm_subs_epu16(a, b);
+      temp2 = _mm_subs_epu16(b, a);
+      temp1 = _mm_or_si128(temp1, temp2);
+
+      // Add result of multiplying by m and add pairs together to running total
+      res = _mm_add_epi32(res, _mm_madd_epi16(temp1, m));
+    }
+    // Move onto the next row
+    a_ptr += a_stride;
+    b_ptr += b_stride;
+    m_ptr += m_stride;
+  }
+  res = _mm_hadd_epi32(res, _mm_setzero_si128());
+  res = _mm_hadd_epi32(res, _mm_setzero_si128());
+  // sad = (sad + 31) >> 6;
+  return (_mm_cvtsi128_si32(res) + 31) >> 6;
+}
+
+static INLINE unsigned int highbd_masked_sad4xh_ssse3(
+    const uint8_t *a8_ptr, int a_stride, const uint8_t *b8_ptr, int b_stride,
+    const uint8_t *m_ptr, int m_stride, int height) {
+  int y;
+  __m128i a, b, m, temp1, temp2;
+  const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8_ptr);
+  const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8_ptr);
+  __m128i res = _mm_setzero_si128();
+  // Add the masked SAD for 2 rows at a time
+  for (y = 0; y < height; y += 2) {
+    // Load a, b, m in xmm registers
+    a = highbd_width4_load_2rows(a_ptr, a_stride);
+    b = highbd_width4_load_2rows(b_ptr, b_stride);
+    temp1 = _mm_loadl_epi64((const __m128i *)m_ptr);
+    temp2 = _mm_loadl_epi64((const __m128i *)(m_ptr + m_stride));
+    m = _mm_unpacklo_epi8(_mm_unpacklo_epi32(temp1, temp2),
+                          _mm_setzero_si128());
+
+    // Calculate the difference between a & b
+    temp1 = _mm_subs_epu16(a, b);
+    temp2 = _mm_subs_epu16(b, a);
+    temp1 = _mm_or_si128(temp1, temp2);
+
+    // Multiply by m and add together
+    res = _mm_add_epi32(res, _mm_madd_epi16(temp1, m));
+
+    // Move onto the next rows
+    a_ptr += a_stride * 2;
+    b_ptr += b_stride * 2;
+    m_ptr += m_stride * 2;
+  }
+  res = _mm_hadd_epi32(res, _mm_setzero_si128());
+  res = _mm_hadd_epi32(res, _mm_setzero_si128());
+  // sad = (sad + 31) >> 6;
+  return (_mm_cvtsi128_si32(res) + 31) >> 6;
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c
new file mode 100644
index 0000000000..fe14597f64
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c
@@ -0,0 +1,1948 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdlib.h>
+#include <emmintrin.h>
+#include <tmmintrin.h>
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_filter.h"
+
+// Half pixel shift
+#define HALF_PIXEL_OFFSET (BIL_SUBPEL_SHIFTS / 2)
+
+/*****************************************************************************
+ * Horizontal additions
+ *****************************************************************************/
+
+static INLINE int32_t hsum_epi32_si32(__m128i v_d) {
+  v_d = _mm_hadd_epi32(v_d, v_d);
+  v_d = _mm_hadd_epi32(v_d, v_d);
+  return _mm_cvtsi128_si32(v_d);
+}
+
+static INLINE int64_t hsum_epi64_si64(__m128i v_q) {
+  v_q = _mm_add_epi64(v_q, _mm_srli_si128(v_q, 8));
+#if ARCH_X86_64
+  return _mm_cvtsi128_si64(v_q);
+#else
+  {
+    int64_t tmp;
+    _mm_storel_epi64((__m128i *)&tmp, v_q);
+    return tmp;
+  }
+#endif
+}
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE int64_t hsum_epi32_si64(__m128i v_d) {
+  const __m128i v_sign_d = _mm_cmplt_epi32(v_d, _mm_setzero_si128());
+  const __m128i v_0_q = _mm_unpacklo_epi32(v_d, v_sign_d);
+  const __m128i v_1_q = _mm_unpackhi_epi32(v_d, v_sign_d);
+  return hsum_epi64_si64(_mm_add_epi64(v_0_q, v_1_q));
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static INLINE uint32_t calc_masked_variance(__m128i v_sum_d, __m128i v_sse_q,
+                                            uint32_t *sse, int w, int h) {
+  int64_t sum64;
+  uint64_t sse64;
+
+  // Horizontal sum
+  sum64 = hsum_epi32_si32(v_sum_d);
+  sse64 = hsum_epi64_si64(v_sse_q);
+
+  sum64 = (sum64 >= 0) ? sum64 : -sum64;
+
+  // Round
+  sum64 = ROUND_POWER_OF_TWO(sum64, 6);
+  sse64 = ROUND_POWER_OF_TWO(sse64, 12);
+
+  // Store the SSE
+  *sse = (uint32_t)sse64;
+  // Compute the variance
+  return *sse - (uint32_t)((sum64 * sum64) / (w * h));
+}
+
+/*****************************************************************************
+ * n*16 Wide versions
+ *****************************************************************************/
+
+static INLINE unsigned int masked_variancewxh_ssse3(
+    const uint8_t *a, int a_stride, const uint8_t *b, int b_stride,
+    const uint8_t *m, int m_stride, int w, int h, unsigned int *sse) {
+  int ii, jj;
+
+  const __m128i v_zero = _mm_setzero_si128();
+
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+
+  assert((w % 16) == 0);
+
+  for (ii = 0; ii < h; ii++) {
+    for (jj = 0; jj < w; jj += 16) {
+      // Load inputs - 8 bits
+      const __m128i v_a_b = _mm_loadu_si128((const __m128i *)(a + jj));
+      const __m128i v_b_b = _mm_loadu_si128((const __m128i *)(b + jj));
+      const __m128i v_m_b = _mm_loadu_si128((const __m128i *)(m + jj));
+
+      // Unpack to 16 bits - still containing max 8 bits
+      const __m128i v_a0_w = _mm_unpacklo_epi8(v_a_b, v_zero);
+      const __m128i v_b0_w = _mm_unpacklo_epi8(v_b_b, v_zero);
+      const __m128i v_m0_w = _mm_unpacklo_epi8(v_m_b, v_zero);
+      const __m128i v_a1_w = _mm_unpackhi_epi8(v_a_b, v_zero);
+      const __m128i v_b1_w = _mm_unpackhi_epi8(v_b_b, v_zero);
+      const __m128i v_m1_w = _mm_unpackhi_epi8(v_m_b, v_zero);
+
+      // Difference: [-255, 255]
+      const __m128i v_d0_w = _mm_sub_epi16(v_a0_w, v_b0_w);
+      const __m128i v_d1_w = _mm_sub_epi16(v_a1_w, v_b1_w);
+
+      // Error - [-255, 255] * [0, 64] = [0xc040, 0x3fc0] => fits in 15 bits
+      const __m128i v_e0_w = _mm_mullo_epi16(v_d0_w, v_m0_w);
+      const __m128i v_e0_d = _mm_madd_epi16(v_d0_w, v_m0_w);
+      const __m128i v_e1_w = _mm_mullo_epi16(v_d1_w, v_m1_w);
+      const __m128i v_e1_d = _mm_madd_epi16(v_d1_w, v_m1_w);
+
+      // Squared error - using madd it's max (15 bits * 15 bits) * 2 = 31 bits
+      const __m128i v_se0_d = _mm_madd_epi16(v_e0_w, v_e0_w);
+      const __m128i v_se1_d = _mm_madd_epi16(v_e1_w, v_e1_w);
+
+      // Sum of v_se{0,1}_d - 31 bits + 31 bits = 32 bits
+      const __m128i v_se_d = _mm_add_epi32(v_se0_d, v_se1_d);
+
+      // Unpack Squared error to 64 bits
+      const __m128i v_se_lo_q = _mm_unpacklo_epi32(v_se_d, v_zero);
+      const __m128i v_se_hi_q = _mm_unpackhi_epi32(v_se_d, v_zero);
+
+      // Accumulate
+      v_sum_d = _mm_add_epi32(v_sum_d, v_e0_d);
+      v_sum_d = _mm_add_epi32(v_sum_d, v_e1_d);
+      v_sse_q = _mm_add_epi64(v_sse_q, v_se_lo_q);
+      v_sse_q = _mm_add_epi64(v_sse_q, v_se_hi_q);
+    }
+
+    // Move on to next row
+    a += a_stride;
+    b += b_stride;
+    m += m_stride;
+  }
+
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, w, h);
+}
+
+#define MASKED_VARWXH(W, H)                                                   \
+  unsigned int aom_masked_variance##W##x##H##_ssse3(                          \
+      const uint8_t *a, int a_stride, const uint8_t *b, int b_stride,         \
+      const uint8_t *m, int m_stride, unsigned int *sse) {                    \
+    return masked_variancewxh_ssse3(a, a_stride, b, b_stride, m, m_stride, W, \
+                                    H, sse);                                  \
+  }
+
+MASKED_VARWXH(16, 8)
+MASKED_VARWXH(16, 16)
+MASKED_VARWXH(16, 32)
+MASKED_VARWXH(32, 16)
+MASKED_VARWXH(32, 32)
+MASKED_VARWXH(32, 64)
+MASKED_VARWXH(64, 32)
+MASKED_VARWXH(64, 64)
+#if CONFIG_EXT_PARTITION
+MASKED_VARWXH(64, 128)
+MASKED_VARWXH(128, 64)
+MASKED_VARWXH(128, 128)
+#endif  // CONFIG_EXT_PARTITION
+
+/*****************************************************************************
+ * 8 Wide versions
+ *****************************************************************************/
+
+static INLINE unsigned int masked_variance8xh_ssse3(
+    const uint8_t *a, int a_stride, const uint8_t *b, int b_stride,
+    const uint8_t *m, int m_stride, int h, unsigned int *sse) {
+  int ii;
+
+  const __m128i v_zero = _mm_setzero_si128();
+
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+
+  for (ii = 0; ii < h; ii++) {
+    // Load inputs - 8 bits
+    const __m128i v_a_b = _mm_loadl_epi64((const __m128i *)a);
+    const __m128i v_b_b = _mm_loadl_epi64((const __m128i *)b);
+    const __m128i v_m_b = _mm_loadl_epi64((const __m128i *)m);
+
+    // Unpack to 16 bits - still containing max 8 bits
+    const __m128i v_a_w = _mm_unpacklo_epi8(v_a_b, v_zero);
+    const __m128i v_b_w = _mm_unpacklo_epi8(v_b_b, v_zero);
+    const __m128i v_m_w = _mm_unpacklo_epi8(v_m_b, v_zero);
+
+    // Difference: [-255, 255]
+    const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w);
+
+    // Error - [-255, 255] * [0, 64] = [0xc040, 0x3fc0] => fits in 15 bits
+    const __m128i v_e_w = _mm_mullo_epi16(v_d_w, v_m_w);
+    const __m128i v_e_d = _mm_madd_epi16(v_d_w, v_m_w);
+
+    // Squared error - using madd it's max (15 bits * 15 bits) * 2 = 31 bits
+    const __m128i v_se_d = _mm_madd_epi16(v_e_w, v_e_w);
+
+    // Unpack Squared error to 64 bits
+    const __m128i v_se_lo_q = _mm_unpacklo_epi32(v_se_d, v_zero);
+    const __m128i v_se_hi_q = _mm_unpackhi_epi32(v_se_d, v_zero);
+
+    // Accumulate
+    v_sum_d = _mm_add_epi32(v_sum_d, v_e_d);
+    v_sse_q = _mm_add_epi64(v_sse_q, v_se_lo_q);
+    v_sse_q = _mm_add_epi64(v_sse_q, v_se_hi_q);
+
+    // Move on to next row
+    a += a_stride;
+    b += b_stride;
+    m += m_stride;
+  }
+
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, 8, h);
+}
+
+#define MASKED_VAR8XH(H)                                                      \
+  unsigned int aom_masked_variance8x##H##_ssse3(                              \
+      const uint8_t *a, int a_stride, const uint8_t *b, int b_stride,         \
+      const uint8_t *m, int m_stride, unsigned int *sse) {                    \
+    return masked_variance8xh_ssse3(a, a_stride, b, b_stride, m, m_stride, H, \
+                                    sse);                                     \
+  }
+
+MASKED_VAR8XH(4)
+MASKED_VAR8XH(8)
+MASKED_VAR8XH(16)
+
+/*****************************************************************************
+ * 4 Wide versions
+ *****************************************************************************/
+
+static INLINE unsigned int masked_variance4xh_ssse3(
+    const uint8_t *a, int a_stride, const uint8_t *b, int b_stride,
+    const uint8_t *m, int m_stride, int h, unsigned int *sse) {
+  int ii;
+
+  const __m128i v_zero = _mm_setzero_si128();
+
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+
+  assert((h % 2) == 0);
+
+  for (ii = 0; ii < h / 2; ii++) {
+    // Load 2 input rows - 8 bits
+    const __m128i v_a0_b = _mm_cvtsi32_si128(*(const uint32_t *)a);
+    const __m128i v_b0_b = _mm_cvtsi32_si128(*(const uint32_t *)b);
+    const __m128i v_m0_b = _mm_cvtsi32_si128(*(const uint32_t *)m);
+    const __m128i v_a1_b = _mm_cvtsi32_si128(*(const uint32_t *)(a + a_stride));
+    const __m128i v_b1_b = _mm_cvtsi32_si128(*(const uint32_t *)(b + b_stride));
+    const __m128i v_m1_b = _mm_cvtsi32_si128(*(const uint32_t *)(m + m_stride));
+
+    // Interleave 2 rows into a single register
+    const __m128i v_a_b = _mm_unpacklo_epi32(v_a0_b, v_a1_b);
+    const __m128i v_b_b = _mm_unpacklo_epi32(v_b0_b, v_b1_b);
+    const __m128i v_m_b = _mm_unpacklo_epi32(v_m0_b, v_m1_b);
+
+    // Unpack to 16 bits - still containing max 8 bits
+    const __m128i v_a_w = _mm_unpacklo_epi8(v_a_b, v_zero);
+    const __m128i v_b_w = _mm_unpacklo_epi8(v_b_b, v_zero);
+    const __m128i v_m_w = _mm_unpacklo_epi8(v_m_b, v_zero);
+
+    // Difference: [-255, 255]
+    const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w);
+
+    // Error - [-255, 255] * [0, 64] = [0xc040, 0x3fc0] => fits in 15 bits
+    const __m128i v_e_w = _mm_mullo_epi16(v_d_w, v_m_w);
+    const __m128i v_e_d = _mm_madd_epi16(v_d_w, v_m_w);
+
+    // Squared error - using madd it's max (15 bits * 15 bits) * 2 = 31 bits
+    const __m128i v_se_d = _mm_madd_epi16(v_e_w, v_e_w);
+
+    // Unpack Squared error to 64 bits
+    const __m128i v_se_lo_q = _mm_unpacklo_epi32(v_se_d, v_zero);
+    const __m128i v_se_hi_q = _mm_unpackhi_epi32(v_se_d, v_zero);
+
+    // Accumulate
+    v_sum_d = _mm_add_epi32(v_sum_d, v_e_d);
+    v_sse_q = _mm_add_epi64(v_sse_q, v_se_lo_q);
+    v_sse_q = _mm_add_epi64(v_sse_q, v_se_hi_q);
+
+    // Move on to next 2 row
+    a += a_stride * 2;
+    b += b_stride * 2;
+    m += m_stride * 2;
+  }
+
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, 4, h);
+}
+
+#define MASKED_VAR4XH(H)                                                      \
+  unsigned int aom_masked_variance4x##H##_ssse3(                              \
+      const uint8_t *a, int a_stride, const uint8_t *b, int b_stride,         \
+      const uint8_t *m, int m_stride, unsigned int *sse) {                    \
+    return masked_variance4xh_ssse3(a, a_stride, b, b_stride, m, m_stride, H, \
+                                    sse);                                     \
+  }
+
+MASKED_VAR4XH(4)
+MASKED_VAR4XH(8)
+
+#if CONFIG_HIGHBITDEPTH
+
+// Main calculation for n*8 wide blocks
+static INLINE void highbd_masked_variance64_ssse3(
+    const uint16_t *a, int a_stride, const uint16_t *b, int b_stride,
+    const uint8_t *m, int m_stride, int w, int h, int64_t *sum, uint64_t *sse) {
+  int ii, jj;
+
+  const __m128i v_zero = _mm_setzero_si128();
+
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+
+  assert((w % 8) == 0);
+
+  for (ii = 0; ii < h; ii++) {
+    for (jj = 0; jj < w; jj += 8) {
+      // Load inputs - 8 bits
+      const __m128i v_a_w = _mm_loadu_si128((const __m128i *)(a + jj));
+      const __m128i v_b_w = _mm_loadu_si128((const __m128i *)(b + jj));
+      const __m128i v_m_b = _mm_loadl_epi64((const __m128i *)(m + jj));
+
+      // Unpack m to 16 bits - still containing max 8 bits
+      const __m128i v_m_w = _mm_unpacklo_epi8(v_m_b, v_zero);
+
+      // Difference: [-4095, 4095]
+      const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w);
+
+      // Error - [-4095, 4095] * [0, 64] => sum of 2 of these fits in 19 bits
+      const __m128i v_e_d = _mm_madd_epi16(v_d_w, v_m_w);
+
+      // Squared error - max (18 bits * 18 bits) = 36 bits (no sign bit)
+      const __m128i v_absd_w = _mm_abs_epi16(v_d_w);
+      const __m128i v_dlo_d = _mm_unpacklo_epi16(v_absd_w, v_zero);
+      const __m128i v_mlo_d = _mm_unpacklo_epi16(v_m_w, v_zero);
+      const __m128i v_elo_d = _mm_madd_epi16(v_dlo_d, v_mlo_d);
+      const __m128i v_dhi_d = _mm_unpackhi_epi16(v_absd_w, v_zero);
+      const __m128i v_mhi_d = _mm_unpackhi_epi16(v_m_w, v_zero);
+      const __m128i v_ehi_d = _mm_madd_epi16(v_dhi_d, v_mhi_d);
+      // Square and sum the errors -> 36bits * 4 = 38bits
+      __m128i v_se0_q, v_se1_q, v_se2_q, v_se3_q, v_se_q, v_elo1_d, v_ehi3_d;
+      v_se0_q = _mm_mul_epu32(v_elo_d, v_elo_d);
+      v_elo1_d = _mm_srli_si128(v_elo_d, 4);
+      v_se1_q = _mm_mul_epu32(v_elo1_d, v_elo1_d);
+      v_se0_q = _mm_add_epi64(v_se0_q, v_se1_q);
+      v_se2_q = _mm_mul_epu32(v_ehi_d, v_ehi_d);
+      v_ehi3_d = _mm_srli_si128(v_ehi_d, 4);
+      v_se3_q = _mm_mul_epu32(v_ehi3_d, v_ehi3_d);
+      v_se1_q = _mm_add_epi64(v_se2_q, v_se3_q);
+      v_se_q = _mm_add_epi64(v_se0_q, v_se1_q);
+
+      // Accumulate
+      v_sum_d = _mm_add_epi32(v_sum_d, v_e_d);
+      v_sse_q = _mm_add_epi64(v_sse_q, v_se_q);
+    }
+
+    // Move on to next row
+    a += a_stride;
+    b += b_stride;
+    m += m_stride;
+  }
+
+  // Horizontal sum
+  *sum = hsum_epi32_si64(v_sum_d);
+  *sse = hsum_epi64_si64(v_sse_q);
+
+  // Round
+  *sum = (*sum >= 0) ? *sum : -*sum;
+  *sum = ROUND_POWER_OF_TWO(*sum, 6);
+  *sse = ROUND_POWER_OF_TWO(*sse, 12);
+}
+
+// Main calculation for 4 wide blocks
+static INLINE void highbd_masked_variance64_4wide_ssse3(
+    const uint16_t *a, int a_stride, const uint16_t *b, int b_stride,
+    const uint8_t *m, int m_stride, int h, int64_t *sum, uint64_t *sse) {
+  int ii;
+
+  const __m128i v_zero = _mm_setzero_si128();
+
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+
+  assert((h % 2) == 0);
+
+  for (ii = 0; ii < h / 2; ii++) {
+    // Load 2 input rows - 8 bits
+    const __m128i v_a0_w = _mm_loadl_epi64((const __m128i *)a);
+    const __m128i v_b0_w = _mm_loadl_epi64((const __m128i *)b);
+    const __m128i v_m0_b = _mm_cvtsi32_si128(*(const uint32_t *)m);
+    const __m128i v_a1_w = _mm_loadl_epi64((const __m128i *)(a + a_stride));
+    const __m128i v_b1_w = _mm_loadl_epi64((const __m128i *)(b + b_stride));
+    const __m128i v_m1_b = _mm_cvtsi32_si128(*(const uint32_t *)(m + m_stride));
+
+    // Interleave 2 rows into a single register
+    const __m128i v_a_w = _mm_unpacklo_epi64(v_a0_w, v_a1_w);
+    const __m128i v_b_w = _mm_unpacklo_epi64(v_b0_w, v_b1_w);
+    const __m128i v_m_b = _mm_unpacklo_epi32(v_m0_b, v_m1_b);
+
+    // Unpack to 16 bits - still containing max 8 bits
+    const __m128i v_m_w = _mm_unpacklo_epi8(v_m_b, v_zero);
+
+    // Difference: [-4095, 4095]
+    const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w);
+
+    // Error - [-4095, 4095] * [0, 64] => fits in 19 bits (incld sign bit)
+    const __m128i v_e_d = _mm_madd_epi16(v_d_w, v_m_w);
+
+    // Squared error - max (18 bits * 18 bits) = 36 bits (no sign bit)
+    const __m128i v_absd_w = _mm_abs_epi16(v_d_w);
+    const __m128i v_dlo_d = _mm_unpacklo_epi16(v_absd_w, v_zero);
+    const __m128i v_mlo_d = _mm_unpacklo_epi16(v_m_w, v_zero);
+    const __m128i v_elo_d = _mm_madd_epi16(v_dlo_d, v_mlo_d);
+    const __m128i v_dhi_d = _mm_unpackhi_epi16(v_absd_w, v_zero);
+    const __m128i v_mhi_d = _mm_unpackhi_epi16(v_m_w, v_zero);
+    const __m128i v_ehi_d = _mm_madd_epi16(v_dhi_d, v_mhi_d);
+    // Square and sum the errors -> 36bits * 4 = 38bits
+    __m128i v_se0_q, v_se1_q, v_se2_q, v_se3_q, v_se_q, v_elo1_d, v_ehi3_d;
+    v_se0_q = _mm_mul_epu32(v_elo_d, v_elo_d);
+    v_elo1_d = _mm_srli_si128(v_elo_d, 4);
+    v_se1_q = _mm_mul_epu32(v_elo1_d, v_elo1_d);
+    v_se0_q = _mm_add_epi64(v_se0_q, v_se1_q);
+    v_se2_q = _mm_mul_epu32(v_ehi_d, v_ehi_d);
+    v_ehi3_d = _mm_srli_si128(v_ehi_d, 4);
+    v_se3_q = _mm_mul_epu32(v_ehi3_d, v_ehi3_d);
+    v_se1_q = _mm_add_epi64(v_se2_q, v_se3_q);
+    v_se_q = _mm_add_epi64(v_se0_q, v_se1_q);
+
+    // Accumulate
+    v_sum_d = _mm_add_epi32(v_sum_d, v_e_d);
+    v_sse_q = _mm_add_epi64(v_sse_q, v_se_q);
+
+    // Move on to next row
+    a += a_stride * 2;
+    b += b_stride * 2;
+    m += m_stride * 2;
+  }
+
+  // Horizontal sum
+  *sum = hsum_epi32_si32(v_sum_d);
+  *sse = hsum_epi64_si64(v_sse_q);
+
+  // Round
+  *sum = (*sum >= 0) ? *sum : -*sum;
+  *sum = ROUND_POWER_OF_TWO(*sum, 6);
+  *sse = ROUND_POWER_OF_TWO(*sse, 12);
+}
+
+static INLINE unsigned int highbd_masked_variancewxh_ssse3(
+    const uint16_t *a, int a_stride, const uint16_t *b, int b_stride,
+    const uint8_t *m, int m_stride, int w, int h, unsigned int *sse) {
+  uint64_t sse64;
+  int64_t sum64;
+
+  if (w == 4)
+    highbd_masked_variance64_4wide_ssse3(a, a_stride, b, b_stride, m, m_stride,
+                                         h, &sum64, &sse64);
+  else
+    highbd_masked_variance64_ssse3(a, a_stride, b, b_stride, m, m_stride, w, h,
+                                   &sum64, &sse64);
+
+  // Store the SSE
+  *sse = (uint32_t)sse64;
+  // Compute and return variance
+  return *sse - (uint32_t)((sum64 * sum64) / (w * h));
+}
+
+static INLINE unsigned int highbd_10_masked_variancewxh_ssse3(
+    const uint16_t *a, int a_stride, const uint16_t *b, int b_stride,
+    const uint8_t *m, int m_stride, int w, int h, unsigned int *sse) {
+  uint64_t sse64;
+  int64_t sum64;
+
+  if (w == 4)
+    highbd_masked_variance64_4wide_ssse3(a, a_stride, b, b_stride, m, m_stride,
+                                         h, &sum64, &sse64);
+  else
+    highbd_masked_variance64_ssse3(a, a_stride, b, b_stride, m, m_stride, w, h,
+                                   &sum64, &sse64);
+
+  // Normalise
+  sum64 = ROUND_POWER_OF_TWO(sum64, 2);
+  sse64 = ROUND_POWER_OF_TWO(sse64, 4);
+
+  // Store the SSE
+  *sse = (uint32_t)sse64;
+  // Compute and return variance
+  return *sse - (uint32_t)((sum64 * sum64) / (w * h));
+}
+
+static INLINE unsigned int highbd_12_masked_variancewxh_ssse3(
+    const uint16_t *a, int a_stride, const uint16_t *b, int b_stride,
+    const uint8_t *m, int m_stride, int w, int h, unsigned int *sse) {
+  uint64_t sse64;
+  int64_t sum64;
+
+  if (w == 4)
+    highbd_masked_variance64_4wide_ssse3(a, a_stride, b, b_stride, m, m_stride,
+                                         h, &sum64, &sse64);
+  else
+    highbd_masked_variance64_ssse3(a, a_stride, b, b_stride, m, m_stride, w, h,
+                                   &sum64, &sse64);
+
+  sum64 = ROUND_POWER_OF_TWO(sum64, 4);
+  sse64 = ROUND_POWER_OF_TWO(sse64, 8);
+
+  // Store the SSE
+  *sse = (uint32_t)sse64;
+  // Compute and return variance
+  return *sse - (uint32_t)((sum64 * sum64) / (w * h));
+}
+
+#define HIGHBD_MASKED_VARWXH(W, H)                                         \
+  unsigned int aom_highbd_masked_variance##W##x##H##_ssse3(                \
+      const uint8_t *a8, int a_stride, const uint8_t *b8, int b_stride,    \
+      const uint8_t *m, int m_stride, unsigned int *sse) {                 \
+    uint16_t *a = CONVERT_TO_SHORTPTR(a8);                                 \
+    uint16_t *b = CONVERT_TO_SHORTPTR(b8);                                 \
+    return highbd_masked_variancewxh_ssse3(a, a_stride, b, b_stride, m,    \
+                                           m_stride, W, H, sse);           \
+  }                                                                        \
+                                                                           \
+  unsigned int aom_highbd_10_masked_variance##W##x##H##_ssse3(             \
+      const uint8_t *a8, int a_stride, const uint8_t *b8, int b_stride,    \
+      const uint8_t *m, int m_stride, unsigned int *sse) {                 \
+    uint16_t *a = CONVERT_TO_SHORTPTR(a8);                                 \
+    uint16_t *b = CONVERT_TO_SHORTPTR(b8);                                 \
+    return highbd_10_masked_variancewxh_ssse3(a, a_stride, b, b_stride, m, \
+                                              m_stride, W, H, sse);        \
+  }                                                                        \
+                                                                           \
+  unsigned int aom_highbd_12_masked_variance##W##x##H##_ssse3(             \
+      const uint8_t *a8, int a_stride, const uint8_t *b8, int b_stride,    \
+      const uint8_t *m, int m_stride, unsigned int *sse) {                 \
+    uint16_t *a = CONVERT_TO_SHORTPTR(a8);                                 \
+    uint16_t *b = CONVERT_TO_SHORTPTR(b8);                                 \
+    return highbd_12_masked_variancewxh_ssse3(a, a_stride, b, b_stride, m, \
+                                              m_stride, W, H, sse);        \
+  }
+
+HIGHBD_MASKED_VARWXH(4, 4)
+HIGHBD_MASKED_VARWXH(4, 8)
+HIGHBD_MASKED_VARWXH(8, 4)
+HIGHBD_MASKED_VARWXH(8, 8)
+HIGHBD_MASKED_VARWXH(8, 16)
+HIGHBD_MASKED_VARWXH(16, 8)
+HIGHBD_MASKED_VARWXH(16, 16)
+HIGHBD_MASKED_VARWXH(16, 32)
+HIGHBD_MASKED_VARWXH(32, 16)
+HIGHBD_MASKED_VARWXH(32, 32)
+HIGHBD_MASKED_VARWXH(32, 64)
+HIGHBD_MASKED_VARWXH(64, 32)
+HIGHBD_MASKED_VARWXH(64, 64)
+#if CONFIG_EXT_PARTITION
+HIGHBD_MASKED_VARWXH(64, 128)
+HIGHBD_MASKED_VARWXH(128, 64)
+HIGHBD_MASKED_VARWXH(128, 128)
+#endif  // CONFIG_EXT_PARTITION
+
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+// Sub pixel versions
+//////////////////////////////////////////////////////////////////////////////
+
+typedef __m128i (*filter_fn_t)(__m128i v_a_b, __m128i v_b_b,
+                               __m128i v_filter_b);
+
+static INLINE __m128i apply_filter_avg(const __m128i v_a_b, const __m128i v_b_b,
+                                       const __m128i v_filter_b) {
+  (void)v_filter_b;
+  return _mm_avg_epu8(v_a_b, v_b_b);
+}
+
+static INLINE __m128i apply_filter(const __m128i v_a_b, const __m128i v_b_b,
+                                   const __m128i v_filter_b) {
+  const __m128i v_rounding_w = _mm_set1_epi16(1 << (FILTER_BITS - 1));
+  __m128i v_input_lo_b = _mm_unpacklo_epi8(v_a_b, v_b_b);
+  __m128i v_input_hi_b = _mm_unpackhi_epi8(v_a_b, v_b_b);
+  __m128i v_temp0_w = _mm_maddubs_epi16(v_input_lo_b, v_filter_b);
+  __m128i v_temp1_w = _mm_maddubs_epi16(v_input_hi_b, v_filter_b);
+  __m128i v_res_lo_w =
+      _mm_srai_epi16(_mm_add_epi16(v_temp0_w, v_rounding_w), FILTER_BITS);
+  __m128i v_res_hi_w =
+      _mm_srai_epi16(_mm_add_epi16(v_temp1_w, v_rounding_w), FILTER_BITS);
+  return _mm_packus_epi16(v_res_lo_w, v_res_hi_w);
+}
+
+// Apply the filter to the contents of the lower half of a and b
+static INLINE void apply_filter_lo(const __m128i v_a_lo_b,
+                                   const __m128i v_b_lo_b,
+                                   const __m128i v_filter_b, __m128i *v_res_w) {
+  const __m128i v_rounding_w = _mm_set1_epi16(1 << (FILTER_BITS - 1));
+  __m128i v_input_b = _mm_unpacklo_epi8(v_a_lo_b, v_b_lo_b);
+  __m128i v_temp0_w = _mm_maddubs_epi16(v_input_b, v_filter_b);
+  *v_res_w =
+      _mm_srai_epi16(_mm_add_epi16(v_temp0_w, v_rounding_w), FILTER_BITS);
+}
+
+static void sum_and_sse(const __m128i v_a_b, const __m128i v_b_b,
+                        const __m128i v_m_b, __m128i *v_sum_d,
+                        __m128i *v_sse_q) {
+  const __m128i v_zero = _mm_setzero_si128();
+  // Unpack to 16 bits - still containing max 8 bits
+  const __m128i v_a0_w = _mm_unpacklo_epi8(v_a_b, v_zero);
+  const __m128i v_b0_w = _mm_unpacklo_epi8(v_b_b, v_zero);
+  const __m128i v_m0_w = _mm_unpacklo_epi8(v_m_b, v_zero);
+  const __m128i v_a1_w = _mm_unpackhi_epi8(v_a_b, v_zero);
+  const __m128i v_b1_w = _mm_unpackhi_epi8(v_b_b, v_zero);
+  const __m128i v_m1_w = _mm_unpackhi_epi8(v_m_b, v_zero);
+
+  // Difference: [-255, 255]
+  const __m128i v_d0_w = _mm_sub_epi16(v_a0_w, v_b0_w);
+  const __m128i v_d1_w = _mm_sub_epi16(v_a1_w, v_b1_w);
+
+  // Error - [-255, 255] * [0, 64] = [0xc040, 0x3fc0] => fits in 15 bits
+  const __m128i v_e0_w = _mm_mullo_epi16(v_d0_w, v_m0_w);
+  const __m128i v_e0_d = _mm_madd_epi16(v_d0_w, v_m0_w);
+  const __m128i v_e1_w = _mm_mullo_epi16(v_d1_w, v_m1_w);
+  const __m128i v_e1_d = _mm_madd_epi16(v_d1_w, v_m1_w);
+
+  // Squared error - using madd it's max (15 bits * 15 bits) * 2 = 31 bits
+  const __m128i v_se0_d = _mm_madd_epi16(v_e0_w, v_e0_w);
+  const __m128i v_se1_d = _mm_madd_epi16(v_e1_w, v_e1_w);
+
+  // Sum of v_se{0,1}_d - 31 bits + 31 bits = 32 bits
+  const __m128i v_se_d = _mm_add_epi32(v_se0_d, v_se1_d);
+
+  // Unpack Squared error to 64 bits
+  const __m128i v_se_lo_q = _mm_unpacklo_epi32(v_se_d, v_zero);
+  const __m128i v_se_hi_q = _mm_unpackhi_epi32(v_se_d, v_zero);
+
+  // Accumulate
+  *v_sum_d = _mm_add_epi32(*v_sum_d, v_e0_d);
+  *v_sum_d = _mm_add_epi32(*v_sum_d, v_e1_d);
+  *v_sse_q = _mm_add_epi64(*v_sse_q, v_se_lo_q);
+  *v_sse_q = _mm_add_epi64(*v_sse_q, v_se_hi_q);
+}
+
+// Functions for width (W) >= 16
+unsigned int aom_masked_subpel_varWxH_xzero(const uint8_t *src, int src_stride,
+                                            int yoffset, const uint8_t *dst,
+                                            int dst_stride, const uint8_t *msk,
+                                            int msk_stride, unsigned int *sse,
+                                            int w, int h,
+                                            filter_fn_t filter_fn) {
+  int i, j;
+  __m128i v_src0_b, v_src1_b, v_res_b, v_dst_b, v_msk_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  const __m128i v_filter_b = _mm_set1_epi16(
+      (bilinear_filters_2t[yoffset][1] << 8) + bilinear_filters_2t[yoffset][0]);
+  assert(yoffset < BIL_SUBPEL_SHIFTS);
+  for (j = 0; j < w; j += 16) {
+    // Load the first row ready
+    v_src0_b = _mm_loadu_si128((const __m128i *)(src + j));
+    // Process 2 rows at a time
+    for (i = 0; i < h; i += 2) {
+      // Load the next row apply the filter
+      v_src1_b = _mm_loadu_si128((const __m128i *)(src + j + src_stride));
+      v_res_b = filter_fn(v_src0_b, v_src1_b, v_filter_b);
+      // Load the dst and msk for the variance calculation
+      v_dst_b = _mm_loadu_si128((const __m128i *)(dst + j));
+      v_msk_b = _mm_loadu_si128((const __m128i *)(msk + j));
+      sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q);
+
+      // Load the next row apply the filter
+      v_src0_b = _mm_loadu_si128((const __m128i *)(src + j + src_stride * 2));
+      v_res_b = filter_fn(v_src1_b, v_src0_b, v_filter_b);
+      // Load the dst and msk for the variance calculation
+      v_dst_b = _mm_loadu_si128((const __m128i *)(dst + j + dst_stride));
+      v_msk_b = _mm_loadu_si128((const __m128i *)(msk + j + msk_stride));
+      sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q);
+      // Move onto the next block of rows
+      src += src_stride * 2;
+      dst += dst_stride * 2;
+      msk += msk_stride * 2;
+    }
+    // Reset to the top of the block
+    src -= src_stride * h;
+    dst -= dst_stride * h;
+    msk -= msk_stride * h;
+  }
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, w, h);
+}
+unsigned int aom_masked_subpel_varWxH_yzero(const uint8_t *src, int src_stride,
+                                            int xoffset, const uint8_t *dst,
+                                            int dst_stride, const uint8_t *msk,
+                                            int msk_stride, unsigned int *sse,
+                                            int w, int h,
+                                            filter_fn_t filter_fn) {
+  int i, j;
+  __m128i v_src0_b, v_src1_b, v_res_b, v_dst_b, v_msk_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  const __m128i v_filter_b = _mm_set1_epi16(
+      (bilinear_filters_2t[xoffset][1] << 8) + bilinear_filters_2t[xoffset][0]);
+  assert(xoffset < BIL_SUBPEL_SHIFTS);
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j += 16) {
+      // Load this row and one below & apply the filter to them
+      v_src0_b = _mm_loadu_si128((const __m128i *)(src + j));
+      v_src1_b = _mm_loadu_si128((const __m128i *)(src + j + 1));
+      v_res_b = filter_fn(v_src0_b, v_src1_b, v_filter_b);
+
+      // Load the dst and msk for the variance calculation
+      v_dst_b = _mm_loadu_si128((const __m128i *)(dst + j));
+      v_msk_b = _mm_loadu_si128((const __m128i *)(msk + j));
+      sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q);
+    }
+    src += src_stride;
+    dst += dst_stride;
+    msk += msk_stride;
+  }
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, w, h);
+}
+unsigned int aom_masked_subpel_varWxH_xnonzero_ynonzero(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride,
+    unsigned int *sse, int w, int h, filter_fn_t xfilter_fn,
+    filter_fn_t yfilter_fn) {
+  int i, j;
+  __m128i v_src0_b, v_src1_b, v_src2_b, v_src3_b;
+  __m128i v_filtered0_b, v_filtered1_b, v_res_b, v_dst_b, v_msk_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  const __m128i v_filterx_b = _mm_set1_epi16(
+      (bilinear_filters_2t[xoffset][1] << 8) + bilinear_filters_2t[xoffset][0]);
+  const __m128i v_filtery_b = _mm_set1_epi16(
+      (bilinear_filters_2t[yoffset][1] << 8) + bilinear_filters_2t[yoffset][0]);
+  assert(yoffset < BIL_SUBPEL_SHIFTS);
+  assert(xoffset < BIL_SUBPEL_SHIFTS);
+  for (j = 0; j < w; j += 16) {
+    // Load the first row ready
+    v_src0_b = _mm_loadu_si128((const __m128i *)(src + j));
+    v_src1_b = _mm_loadu_si128((const __m128i *)(src + j + 1));
+    v_filtered0_b = xfilter_fn(v_src0_b, v_src1_b, v_filterx_b);
+    // Process 2 rows at a time
+    for (i = 0; i < h; i += 2) {
+      // Load the next row & apply the filter
+      v_src2_b = _mm_loadu_si128((const __m128i *)(src + src_stride + j));
+      v_src3_b = _mm_loadu_si128((const __m128i *)(src + src_stride + j + 1));
+      v_filtered1_b = xfilter_fn(v_src2_b, v_src3_b, v_filterx_b);
+      // Load the dst and msk for the variance calculation
+      v_dst_b = _mm_loadu_si128((const __m128i *)(dst + j));
+      v_msk_b = _mm_loadu_si128((const __m128i *)(msk + j));
+      // Complete the calculation for this row and add it to the running total
+      v_res_b = yfilter_fn(v_filtered0_b, v_filtered1_b, v_filtery_b);
+      sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q);
+
+      // Load the next row & apply the filter
+      v_src0_b = _mm_loadu_si128((const __m128i *)(src + src_stride * 2 + j));
+      v_src1_b =
+          _mm_loadu_si128((const __m128i *)(src + src_stride * 2 + j + 1));
+      v_filtered0_b = xfilter_fn(v_src0_b, v_src1_b, v_filterx_b);
+      // Load the dst and msk for the variance calculation
+      v_dst_b = _mm_loadu_si128((const __m128i *)(dst + dst_stride + j));
+      v_msk_b = _mm_loadu_si128((const __m128i *)(msk + msk_stride + j));
+      // Complete the calculation for this row and add it to the running total
+      v_res_b = yfilter_fn(v_filtered1_b, v_filtered0_b, v_filtery_b);
+      sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q);
+      // Move onto the next block of rows
+      src += src_stride * 2;
+      dst += dst_stride * 2;
+      msk += msk_stride * 2;
+    }
+    // Reset to the top of the block
+    src -= src_stride * h;
+    dst -= dst_stride * h;
+    msk -= msk_stride * h;
+  }
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, w, h);
+}
+
+// Note order in which rows loaded xmm[127:96] = row 1, xmm[95:64] = row 2,
+// xmm[63:32] = row 3, xmm[31:0] = row 4
+unsigned int aom_masked_subpel_var4xH_xzero(const uint8_t *src, int src_stride,
+                                            int yoffset, const uint8_t *dst,
+                                            int dst_stride, const uint8_t *msk,
+                                            int msk_stride, unsigned int *sse,
+                                            int h) {
+  int i;
+  __m128i v_src0_b, v_src1_b, v_src2_b, v_src3_b, v_filtered1_w, v_filtered2_w;
+  __m128i v_dst0_b, v_dst1_b, v_dst2_b, v_dst3_b;
+  __m128i v_msk0_b, v_msk1_b, v_msk2_b, v_msk3_b, v_res_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  __m128i v_filter_b = _mm_set1_epi16((bilinear_filters_2t[yoffset][1] << 8) +
+                                      bilinear_filters_2t[yoffset][0]);
+  assert(yoffset < BIL_SUBPEL_SHIFTS);
+  // Load the first row of src data ready
+  v_src0_b = _mm_loadl_epi64((const __m128i *)src);
+  for (i = 0; i < h; i += 4) {
+    // Load the rest of the source data for these rows
+    v_src1_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 1));
+    v_src1_b = _mm_unpacklo_epi32(v_src1_b, v_src0_b);
+    v_src2_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2));
+    v_src3_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 3));
+    v_src3_b = _mm_unpacklo_epi32(v_src3_b, v_src2_b);
+    v_src0_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 4));
+    // Load the dst data
+    v_dst0_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 0));
+    v_dst1_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 1));
+    v_dst0_b = _mm_unpacklo_epi32(v_dst1_b, v_dst0_b);
+    v_dst2_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 2));
+    v_dst3_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 3));
+    v_dst2_b = _mm_unpacklo_epi32(v_dst3_b, v_dst2_b);
+    v_dst0_b = _mm_unpacklo_epi64(v_dst2_b, v_dst0_b);
+    // Load the mask data
+    v_msk0_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 0));
+    v_msk1_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 1));
+    v_msk0_b = _mm_unpacklo_epi32(v_msk1_b, v_msk0_b);
+    v_msk2_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 2));
+    v_msk3_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 3));
+    v_msk2_b = _mm_unpacklo_epi32(v_msk3_b, v_msk2_b);
+    v_msk0_b = _mm_unpacklo_epi64(v_msk2_b, v_msk0_b);
+    // Apply the y filter
+    if (yoffset == HALF_PIXEL_OFFSET) {
+      v_src1_b = _mm_unpacklo_epi64(v_src3_b, v_src1_b);
+      v_src2_b =
+          _mm_or_si128(_mm_slli_si128(v_src1_b, 4),
+                       _mm_and_si128(v_src0_b, _mm_setr_epi32(-1, 0, 0, 0)));
+      v_res_b = _mm_avg_epu8(v_src1_b, v_src2_b);
+    } else {
+      v_src2_b =
+          _mm_or_si128(_mm_slli_si128(v_src1_b, 4),
+                       _mm_and_si128(v_src2_b, _mm_setr_epi32(-1, 0, 0, 0)));
+      apply_filter_lo(v_src1_b, v_src2_b, v_filter_b, &v_filtered1_w);
+      v_src2_b =
+          _mm_or_si128(_mm_slli_si128(v_src3_b, 4),
+                       _mm_and_si128(v_src0_b, _mm_setr_epi32(-1, 0, 0, 0)));
+      apply_filter_lo(v_src3_b, v_src2_b, v_filter_b, &v_filtered2_w);
+      v_res_b = _mm_packus_epi16(v_filtered2_w, v_filtered1_w);
+    }
+    // Compute the sum and SSE
+    sum_and_sse(v_res_b, v_dst0_b, v_msk0_b, &v_sum_d, &v_sse_q);
+    // Move onto the next set of rows
+    src += src_stride * 4;
+    dst += dst_stride * 4;
+    msk += msk_stride * 4;
+  }
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, 4, h);
+}
+
+// Note order in which rows loaded xmm[127:64] = row 1, xmm[63:0] = row 2
+unsigned int aom_masked_subpel_var8xH_xzero(const uint8_t *src, int src_stride,
+                                            int yoffset, const uint8_t *dst,
+                                            int dst_stride, const uint8_t *msk,
+                                            int msk_stride, unsigned int *sse,
+                                            int h) {
+  int i;
+  __m128i v_src0_b, v_src1_b, v_filtered0_w, v_filtered1_w, v_res_b;
+  __m128i v_dst_b = _mm_setzero_si128();
+  __m128i v_msk_b = _mm_setzero_si128();
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  __m128i v_filter_b = _mm_set1_epi16((bilinear_filters_2t[yoffset][1] << 8) +
+                                      bilinear_filters_2t[yoffset][0]);
+  assert(yoffset < BIL_SUBPEL_SHIFTS);
+  // Load the first row of src data ready
+  v_src0_b = _mm_loadl_epi64((const __m128i *)src);
+  for (i = 0; i < h; i += 2) {
+    if (yoffset == HALF_PIXEL_OFFSET) {
+      // Load the rest of the source data for these rows
+      v_src1_b = _mm_or_si128(
+          _mm_slli_si128(v_src0_b, 8),
+          _mm_loadl_epi64((const __m128i *)(src + src_stride * 1)));
+      v_src0_b = _mm_or_si128(
+          _mm_slli_si128(v_src1_b, 8),
+          _mm_loadl_epi64((const __m128i *)(src + src_stride * 2)));
+      // Apply the y filter
+      v_res_b = _mm_avg_epu8(v_src1_b, v_src0_b);
+    } else {
+      // Load the data and apply the y filter
+      v_src1_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 1));
+      apply_filter_lo(v_src0_b, v_src1_b, v_filter_b, &v_filtered0_w);
+      v_src0_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2));
+      apply_filter_lo(v_src1_b, v_src0_b, v_filter_b, &v_filtered1_w);
+      v_res_b = _mm_packus_epi16(v_filtered1_w, v_filtered0_w);
+    }
+    // Load the dst data
+    v_dst_b = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1)),
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0)));
+    // Load the mask data
+    v_msk_b = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1)),
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0)));
+    // Compute the sum and SSE
+    sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q);
+    // Move onto the next set of rows
+    src += src_stride * 2;
+    dst += dst_stride * 2;
+    msk += msk_stride * 2;
+  }
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, 8, h);
+}
+
+// Note order in which rows loaded xmm[127:96] = row 1, xmm[95:64] = row 2,
+// xmm[63:32] = row 3, xmm[31:0] = row 4
+unsigned int aom_masked_subpel_var4xH_yzero(const uint8_t *src, int src_stride,
+                                            int xoffset, const uint8_t *dst,
+                                            int dst_stride, const uint8_t *msk,
+                                            int msk_stride, unsigned int *sse,
+                                            int h) {
+  int i;
+  __m128i v_src0_b, v_src1_b, v_src2_b, v_src3_b, v_filtered0_w, v_filtered2_w;
+  __m128i v_src0_shift_b, v_src1_shift_b, v_src2_shift_b, v_src3_shift_b;
+  __m128i v_dst0_b, v_dst1_b, v_dst2_b, v_dst3_b;
+  __m128i v_msk0_b, v_msk1_b, v_msk2_b, v_msk3_b, v_res_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  __m128i v_filter_b = _mm_set1_epi16((bilinear_filters_2t[xoffset][1] << 8) +
+                                      bilinear_filters_2t[xoffset][0]);
+  assert(xoffset < BIL_SUBPEL_SHIFTS);
+  for (i = 0; i < h; i += 4) {
+    // Load the src data
+    v_src0_b = _mm_loadl_epi64((const __m128i *)src);
+    v_src0_shift_b = _mm_srli_si128(v_src0_b, 1);
+    v_src1_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 1));
+    v_src0_b = _mm_unpacklo_epi32(v_src1_b, v_src0_b);
+    v_src1_shift_b = _mm_srli_si128(v_src1_b, 1);
+    v_src2_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2));
+    v_src0_shift_b = _mm_unpacklo_epi32(v_src1_shift_b, v_src0_shift_b);
+    v_src2_shift_b = _mm_srli_si128(v_src2_b, 1);
+    v_src3_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 3));
+    v_src2_b = _mm_unpacklo_epi32(v_src3_b, v_src2_b);
+    v_src3_shift_b = _mm_srli_si128(v_src3_b, 1);
+    v_src2_shift_b = _mm_unpacklo_epi32(v_src3_shift_b, v_src2_shift_b);
+    // Load the dst data
+    v_dst0_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 0));
+    v_dst1_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 1));
+    v_dst0_b = _mm_unpacklo_epi32(v_dst1_b, v_dst0_b);
+    v_dst2_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 2));
+    v_dst3_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 3));
+    v_dst2_b = _mm_unpacklo_epi32(v_dst3_b, v_dst2_b);
+    v_dst0_b = _mm_unpacklo_epi64(v_dst2_b, v_dst0_b);
+    // Load the mask data
+    v_msk0_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 0));
+    v_msk1_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 1));
+    v_msk0_b = _mm_unpacklo_epi32(v_msk1_b, v_msk0_b);
+    v_msk2_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 2));
+    v_msk3_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 3));
+    v_msk2_b = _mm_unpacklo_epi32(v_msk3_b, v_msk2_b);
+    v_msk0_b = _mm_unpacklo_epi64(v_msk2_b, v_msk0_b);
+    // Apply the x filter
+    if (xoffset == HALF_PIXEL_OFFSET) {
+      v_src0_b = _mm_unpacklo_epi64(v_src2_b, v_src0_b);
+      v_src0_shift_b = _mm_unpacklo_epi64(v_src2_shift_b, v_src0_shift_b);
+      v_res_b = _mm_avg_epu8(v_src0_b, v_src0_shift_b);
+    } else {
+      apply_filter_lo(v_src0_b, v_src0_shift_b, v_filter_b, &v_filtered0_w);
+      apply_filter_lo(v_src2_b, v_src2_shift_b, v_filter_b, &v_filtered2_w);
+      v_res_b = _mm_packus_epi16(v_filtered2_w, v_filtered0_w);
+    }
+    // Compute the sum and SSE
+    sum_and_sse(v_res_b, v_dst0_b, v_msk0_b, &v_sum_d, &v_sse_q);
+    // Move onto the next set of rows
+    src += src_stride * 4;
+    dst += dst_stride * 4;
+    msk += msk_stride * 4;
+  }
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, 4, h);
+}
+
+unsigned int aom_masked_subpel_var8xH_yzero(const uint8_t *src, int src_stride,
+                                            int xoffset, const uint8_t *dst,
+                                            int dst_stride, const uint8_t *msk,
+                                            int msk_stride, unsigned int *sse,
+                                            int h) {
+  int i;
+  __m128i v_src0_b, v_src1_b, v_filtered0_w, v_filtered1_w;
+  __m128i v_src0_shift_b, v_src1_shift_b, v_res_b, v_dst_b, v_msk_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  __m128i v_filter_b = _mm_set1_epi16((bilinear_filters_2t[xoffset][1] << 8) +
+                                      bilinear_filters_2t[xoffset][0]);
+  assert(xoffset < BIL_SUBPEL_SHIFTS);
+  for (i = 0; i < h; i += 2) {
+    // Load the src data
+    v_src0_b = _mm_loadu_si128((const __m128i *)(src));
+    v_src0_shift_b = _mm_srli_si128(v_src0_b, 1);
+    v_src1_b = _mm_loadu_si128((const __m128i *)(src + src_stride));
+    v_src1_shift_b = _mm_srli_si128(v_src1_b, 1);
+    // Apply the x filter
+    if (xoffset == HALF_PIXEL_OFFSET) {
+      v_src1_b = _mm_unpacklo_epi64(v_src0_b, v_src1_b);
+      v_src1_shift_b = _mm_unpacklo_epi64(v_src0_shift_b, v_src1_shift_b);
+      v_res_b = _mm_avg_epu8(v_src1_b, v_src1_shift_b);
+    } else {
+      apply_filter_lo(v_src0_b, v_src0_shift_b, v_filter_b, &v_filtered0_w);
+      apply_filter_lo(v_src1_b, v_src1_shift_b, v_filter_b, &v_filtered1_w);
+      v_res_b = _mm_packus_epi16(v_filtered0_w, v_filtered1_w);
+    }
+    // Load the dst data
+    v_dst_b = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0)),
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1)));
+    // Load the mask data
+    v_msk_b = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0)),
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1)));
+    // Compute the sum and SSE
+    sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q);
+    // Move onto the next set of rows
+    src += src_stride * 2;
+    dst += dst_stride * 2;
+    msk += msk_stride * 2;
+  }
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, 8, h);
+}
+
+// Note order in which rows loaded xmm[127:96] = row 1, xmm[95:64] = row 2,
+// xmm[63:32] = row 3, xmm[31:0] = row 4
+unsigned int aom_masked_subpel_var4xH_xnonzero_ynonzero(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride,
+    unsigned int *sse, int h) {
+  int i;
+  __m128i v_src0_b, v_src1_b, v_src2_b, v_src3_b, v_filtered0_w, v_filtered2_w;
+  __m128i v_src0_shift_b, v_src1_shift_b, v_src2_shift_b, v_src3_shift_b;
+  __m128i v_dst0_b, v_dst1_b, v_dst2_b, v_dst3_b, v_temp_b;
+  __m128i v_msk0_b, v_msk1_b, v_msk2_b, v_msk3_b, v_extra_row_b, v_res_b;
+  __m128i v_xres_b[2];
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  __m128i v_filterx_b = _mm_set1_epi16((bilinear_filters_2t[xoffset][1] << 8) +
+                                       bilinear_filters_2t[xoffset][0]);
+  __m128i v_filtery_b = _mm_set1_epi16((bilinear_filters_2t[yoffset][1] << 8) +
+                                       bilinear_filters_2t[yoffset][0]);
+  assert(xoffset < BIL_SUBPEL_SHIFTS);
+  assert(yoffset < BIL_SUBPEL_SHIFTS);
+  for (i = 0; i < h; i += 4) {
+    // Load the src data
+    v_src0_b = _mm_loadl_epi64((const __m128i *)src);
+    v_src0_shift_b = _mm_srli_si128(v_src0_b, 1);
+    v_src1_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 1));
+    v_src0_b = _mm_unpacklo_epi32(v_src1_b, v_src0_b);
+    v_src1_shift_b = _mm_srli_si128(v_src1_b, 1);
+    v_src2_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2));
+    v_src0_shift_b = _mm_unpacklo_epi32(v_src1_shift_b, v_src0_shift_b);
+    v_src2_shift_b = _mm_srli_si128(v_src2_b, 1);
+    v_src3_b = _mm_loadl_epi64((const __m128i *)(src + src_stride * 3));
+    v_src2_b = _mm_unpacklo_epi32(v_src3_b, v_src2_b);
+    v_src3_shift_b = _mm_srli_si128(v_src3_b, 1);
+    v_src2_shift_b = _mm_unpacklo_epi32(v_src3_shift_b, v_src2_shift_b);
+    // Apply the x filter
+    if (xoffset == HALF_PIXEL_OFFSET) {
+      v_src0_b = _mm_unpacklo_epi64(v_src2_b, v_src0_b);
+      v_src0_shift_b = _mm_unpacklo_epi64(v_src2_shift_b, v_src0_shift_b);
+      v_xres_b[i == 0 ? 0 : 1] = _mm_avg_epu8(v_src0_b, v_src0_shift_b);
+    } else {
+      apply_filter_lo(v_src0_b, v_src0_shift_b, v_filterx_b, &v_filtered0_w);
+      apply_filter_lo(v_src2_b, v_src2_shift_b, v_filterx_b, &v_filtered2_w);
+      v_xres_b[i == 0 ? 0 : 1] = _mm_packus_epi16(v_filtered2_w, v_filtered0_w);
+    }
+    // Move onto the next set of rows
+    src += src_stride * 4;
+  }
+  // Load one more row to be used in the y filter
+  v_src0_b = _mm_loadl_epi64((const __m128i *)src);
+  v_src0_shift_b = _mm_srli_si128(v_src0_b, 1);
+  // Apply the x filter
+  if (xoffset == HALF_PIXEL_OFFSET) {
+    v_extra_row_b = _mm_and_si128(_mm_avg_epu8(v_src0_b, v_src0_shift_b),
+                                  _mm_setr_epi32(-1, 0, 0, 0));
+  } else {
+    apply_filter_lo(v_src0_b, v_src0_shift_b, v_filterx_b, &v_filtered0_w);
+    v_extra_row_b =
+        _mm_and_si128(_mm_packus_epi16(v_filtered0_w, _mm_setzero_si128()),
+                      _mm_setr_epi32(-1, 0, 0, 0));
+  }
+
+  for (i = 0; i < h; i += 4) {
+    if (h == 8 && i == 0) {
+      v_temp_b = _mm_or_si128(_mm_slli_si128(v_xres_b[0], 4),
+                              _mm_srli_si128(v_xres_b[1], 12));
+    } else {
+      v_temp_b = _mm_or_si128(_mm_slli_si128(v_xres_b[i == 0 ? 0 : 1], 4),
+                              v_extra_row_b);
+    }
+    // Apply the y filter
+    if (yoffset == HALF_PIXEL_OFFSET) {
+      v_res_b = _mm_avg_epu8(v_xres_b[i == 0 ? 0 : 1], v_temp_b);
+    } else {
+      v_res_b = apply_filter(v_xres_b[i == 0 ? 0 : 1], v_temp_b, v_filtery_b);
+    }
+
+    // Load the dst data
+    v_dst0_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 0));
+    v_dst1_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 1));
+    v_dst0_b = _mm_unpacklo_epi32(v_dst1_b, v_dst0_b);
+    v_dst2_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 2));
+    v_dst3_b = _mm_cvtsi32_si128(*(const uint32_t *)(dst + dst_stride * 3));
+    v_dst2_b = _mm_unpacklo_epi32(v_dst3_b, v_dst2_b);
+    v_dst0_b = _mm_unpacklo_epi64(v_dst2_b, v_dst0_b);
+    // Load the mask data
+    v_msk0_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 0));
+    v_msk1_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 1));
+    v_msk0_b = _mm_unpacklo_epi32(v_msk1_b, v_msk0_b);
+    v_msk2_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 2));
+    v_msk3_b = _mm_cvtsi32_si128(*(const uint32_t *)(msk + msk_stride * 3));
+    v_msk2_b = _mm_unpacklo_epi32(v_msk3_b, v_msk2_b);
+    v_msk0_b = _mm_unpacklo_epi64(v_msk2_b, v_msk0_b);
+    // Compute the sum and SSE
+    sum_and_sse(v_res_b, v_dst0_b, v_msk0_b, &v_sum_d, &v_sse_q);
+    // Move onto the next set of rows
+    dst += dst_stride * 4;
+    msk += msk_stride * 4;
+  }
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, 4, h);
+}
+
+unsigned int aom_masked_subpel_var8xH_xnonzero_ynonzero(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride,
+    unsigned int *sse, int h) {
+  int i;
+  __m128i v_src0_b, v_src1_b, v_filtered0_w, v_filtered1_w, v_dst_b, v_msk_b;
+  __m128i v_src0_shift_b, v_src1_shift_b;
+  __m128i v_xres0_b, v_xres1_b, v_res_b, v_temp_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  __m128i v_filterx_b = _mm_set1_epi16((bilinear_filters_2t[xoffset][1] << 8) +
+                                       bilinear_filters_2t[xoffset][0]);
+  __m128i v_filtery_b = _mm_set1_epi16((bilinear_filters_2t[yoffset][1] << 8) +
+                                       bilinear_filters_2t[yoffset][0]);
+  assert(xoffset < BIL_SUBPEL_SHIFTS);
+  assert(yoffset < BIL_SUBPEL_SHIFTS);
+  // Load the first block of src data
+  v_src0_b = _mm_loadu_si128((const __m128i *)(src));
+  v_src0_shift_b = _mm_srli_si128(v_src0_b, 1);
+  v_src1_b = _mm_loadu_si128((const __m128i *)(src + src_stride));
+  v_src1_shift_b = _mm_srli_si128(v_src1_b, 1);
+  // Apply the x filter
+  if (xoffset == HALF_PIXEL_OFFSET) {
+    v_src1_b = _mm_unpacklo_epi64(v_src0_b, v_src1_b);
+    v_src1_shift_b = _mm_unpacklo_epi64(v_src0_shift_b, v_src1_shift_b);
+    v_xres0_b = _mm_avg_epu8(v_src1_b, v_src1_shift_b);
+  } else {
+    apply_filter_lo(v_src0_b, v_src0_shift_b, v_filterx_b, &v_filtered0_w);
+    apply_filter_lo(v_src1_b, v_src1_shift_b, v_filterx_b, &v_filtered1_w);
+    v_xres0_b = _mm_packus_epi16(v_filtered0_w, v_filtered1_w);
+  }
+  for (i = 0; i < h; i += 4) {
+    // Load the next block of src data
+    v_src0_b = _mm_loadu_si128((const __m128i *)(src + src_stride * 2));
+    v_src0_shift_b = _mm_srli_si128(v_src0_b, 1);
+    v_src1_b = _mm_loadu_si128((const __m128i *)(src + src_stride * 3));
+    v_src1_shift_b = _mm_srli_si128(v_src1_b, 1);
+    // Apply the x filter
+    if (xoffset == HALF_PIXEL_OFFSET) {
+      v_src1_b = _mm_unpacklo_epi64(v_src0_b, v_src1_b);
+      v_src1_shift_b = _mm_unpacklo_epi64(v_src0_shift_b, v_src1_shift_b);
+      v_xres1_b = _mm_avg_epu8(v_src1_b, v_src1_shift_b);
+    } else {
+      apply_filter_lo(v_src0_b, v_src0_shift_b, v_filterx_b, &v_filtered0_w);
+      apply_filter_lo(v_src1_b, v_src1_shift_b, v_filterx_b, &v_filtered1_w);
+      v_xres1_b = _mm_packus_epi16(v_filtered0_w, v_filtered1_w);
+    }
+    // Apply the y filter to the previous block
+    v_temp_b = _mm_or_si128(_mm_srli_si128(v_xres0_b, 8),
+                            _mm_slli_si128(v_xres1_b, 8));
+    if (yoffset == HALF_PIXEL_OFFSET) {
+      v_res_b = _mm_avg_epu8(v_xres0_b, v_temp_b);
+    } else {
+      v_res_b = apply_filter(v_xres0_b, v_temp_b, v_filtery_b);
+    }
+    // Load the dst data
+    v_dst_b = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0)),
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1)));
+    // Load the mask data
+    v_msk_b = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0)),
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1)));
+    // Compute the sum and SSE
+    sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q);
+
+    // Load the next block of src data
+    v_src0_b = _mm_loadu_si128((const __m128i *)(src + src_stride * 4));
+    v_src0_shift_b = _mm_srli_si128(v_src0_b, 1);
+    v_src1_b = _mm_loadu_si128((const __m128i *)(src + src_stride * 5));
+    v_src1_shift_b = _mm_srli_si128(v_src1_b, 1);
+    // Apply the x filter
+    if (xoffset == HALF_PIXEL_OFFSET) {
+      v_src1_b = _mm_unpacklo_epi64(v_src0_b, v_src1_b);
+      v_src1_shift_b = _mm_unpacklo_epi64(v_src0_shift_b, v_src1_shift_b);
+      v_xres0_b = _mm_avg_epu8(v_src1_b, v_src1_shift_b);
+    } else {
+      apply_filter_lo(v_src0_b, v_src0_shift_b, v_filterx_b, &v_filtered0_w);
+      apply_filter_lo(v_src1_b, v_src1_shift_b, v_filterx_b, &v_filtered1_w);
+      v_xres0_b = _mm_packus_epi16(v_filtered0_w, v_filtered1_w);
+    }
+    // Apply the y filter to the previous block
+    v_temp_b = _mm_or_si128(_mm_srli_si128(v_xres1_b, 8),
+                            _mm_slli_si128(v_xres0_b, 8));
+    if (yoffset == HALF_PIXEL_OFFSET) {
+      v_res_b = _mm_avg_epu8(v_xres1_b, v_temp_b);
+    } else {
+      v_res_b = apply_filter(v_xres1_b, v_temp_b, v_filtery_b);
+    }
+    // Load the dst data
+    v_dst_b = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 2)),
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 3)));
+    // Load the mask data
+    v_msk_b = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 2)),
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 3)));
+    // Compute the sum and SSE
+    sum_and_sse(v_res_b, v_dst_b, v_msk_b, &v_sum_d, &v_sse_q);
+    // Move onto the next set of rows
+    src += src_stride * 4;
+    dst += dst_stride * 4;
+    msk += msk_stride * 4;
+  }
+  return calc_masked_variance(v_sum_d, v_sse_q, sse, 8, h);
+}
+
+// For W >=16
+#define MASK_SUBPIX_VAR_LARGE(W, H)                                            \
+  unsigned int aom_masked_sub_pixel_variance##W##x##H##_ssse3(                 \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,            \
+      const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride,  \
+      unsigned int *sse) {                                                     \
+    assert(W % 16 == 0);                                                       \
+    if (xoffset == 0) {                                                        \
+      if (yoffset == 0)                                                        \
+        return aom_masked_variance##W##x##H##_ssse3(                           \
+            src, src_stride, dst, dst_stride, msk, msk_stride, sse);           \
+      else if (yoffset == HALF_PIXEL_OFFSET)                                   \
+        return aom_masked_subpel_varWxH_xzero(                                 \
+            src, src_stride, HALF_PIXEL_OFFSET, dst, dst_stride, msk,          \
+            msk_stride, sse, W, H, apply_filter_avg);                          \
+      else                                                                     \
+        return aom_masked_subpel_varWxH_xzero(src, src_stride, yoffset, dst,   \
+                                              dst_stride, msk, msk_stride,     \
+                                              sse, W, H, apply_filter);        \
+    } else if (yoffset == 0) {                                                 \
+      if (xoffset == HALF_PIXEL_OFFSET)                                        \
+        return aom_masked_subpel_varWxH_yzero(                                 \
+            src, src_stride, HALF_PIXEL_OFFSET, dst, dst_stride, msk,          \
+            msk_stride, sse, W, H, apply_filter_avg);                          \
+      else                                                                     \
+        return aom_masked_subpel_varWxH_yzero(src, src_stride, xoffset, dst,   \
+                                              dst_stride, msk, msk_stride,     \
+                                              sse, W, H, apply_filter);        \
+    } else if (xoffset == HALF_PIXEL_OFFSET) {                                 \
+      if (yoffset == HALF_PIXEL_OFFSET)                                        \
+        return aom_masked_subpel_varWxH_xnonzero_ynonzero(                     \
+            src, src_stride, HALF_PIXEL_OFFSET, HALF_PIXEL_OFFSET, dst,        \
+            dst_stride, msk, msk_stride, sse, W, H, apply_filter_avg,          \
+            apply_filter_avg);                                                 \
+      else                                                                     \
+        return aom_masked_subpel_varWxH_xnonzero_ynonzero(                     \
+            src, src_stride, HALF_PIXEL_OFFSET, yoffset, dst, dst_stride, msk, \
+            msk_stride, sse, W, H, apply_filter_avg, apply_filter);            \
+    } else {                                                                   \
+      if (yoffset == HALF_PIXEL_OFFSET)                                        \
+        return aom_masked_subpel_varWxH_xnonzero_ynonzero(                     \
+            src, src_stride, xoffset, HALF_PIXEL_OFFSET, dst, dst_stride, msk, \
+            msk_stride, sse, W, H, apply_filter, apply_filter_avg);            \
+      else                                                                     \
+        return aom_masked_subpel_varWxH_xnonzero_ynonzero(                     \
+            src, src_stride, xoffset, yoffset, dst, dst_stride, msk,           \
+            msk_stride, sse, W, H, apply_filter, apply_filter);                \
+    }                                                                          \
+  }
+
+// For W < 16
+#define MASK_SUBPIX_VAR_SMALL(W, H)                                            \
+  unsigned int aom_masked_sub_pixel_variance##W##x##H##_ssse3(                 \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,            \
+      const uint8_t *dst, int dst_stride, const uint8_t *msk, int msk_stride,  \
+      unsigned int *sse) {                                                     \
+    assert(W == 4 || W == 8);                                                  \
+    if (xoffset == 0 && yoffset == 0)                                          \
+      return aom_masked_variance##W##x##H##_ssse3(                             \
+          src, src_stride, dst, dst_stride, msk, msk_stride, sse);             \
+    else if (xoffset == 0)                                                     \
+      return aom_masked_subpel_var##W##xH_xzero(                               \
+          src, src_stride, yoffset, dst, dst_stride, msk, msk_stride, sse, H); \
+    else if (yoffset == 0)                                                     \
+      return aom_masked_subpel_var##W##xH_yzero(                               \
+          src, src_stride, xoffset, dst, dst_stride, msk, msk_stride, sse, H); \
+    else                                                                       \
+      return aom_masked_subpel_var##W##xH_xnonzero_ynonzero(                   \
+          src, src_stride, xoffset, yoffset, dst, dst_stride, msk, msk_stride, \
+          sse, H);                                                             \
+  }
+
+MASK_SUBPIX_VAR_SMALL(4, 4)
+MASK_SUBPIX_VAR_SMALL(4, 8)
+MASK_SUBPIX_VAR_SMALL(8, 4)
+MASK_SUBPIX_VAR_SMALL(8, 8)
+MASK_SUBPIX_VAR_SMALL(8, 16)
+MASK_SUBPIX_VAR_LARGE(16, 8)
+MASK_SUBPIX_VAR_LARGE(16, 16)
+MASK_SUBPIX_VAR_LARGE(16, 32)
+MASK_SUBPIX_VAR_LARGE(32, 16)
+MASK_SUBPIX_VAR_LARGE(32, 32)
+MASK_SUBPIX_VAR_LARGE(32, 64)
+MASK_SUBPIX_VAR_LARGE(64, 32)
+MASK_SUBPIX_VAR_LARGE(64, 64)
+#if CONFIG_EXT_PARTITION
+MASK_SUBPIX_VAR_LARGE(64, 128)
+MASK_SUBPIX_VAR_LARGE(128, 64)
+MASK_SUBPIX_VAR_LARGE(128, 128)
+#endif  // CONFIG_EXT_PARTITION
+
+#if CONFIG_HIGHBITDEPTH
+typedef uint32_t (*highbd_calc_masked_var_t)(__m128i v_sum_d, __m128i v_sse_q,
+                                             uint32_t *sse, int w, int h);
+typedef unsigned int (*highbd_variance_fn_t)(const uint8_t *a8, int a_stride,
+                                             const uint8_t *b8, int b_stride,
+                                             const uint8_t *m, int m_stride,
+                                             unsigned int *sse);
+typedef __m128i (*highbd_filter_fn_t)(__m128i v_a_w, __m128i v_b_w,
+                                      __m128i v_filter_w);
+
+static INLINE __m128i highbd_apply_filter_avg(const __m128i v_a_w,
+                                              const __m128i v_b_w,
+                                              const __m128i v_filter_w) {
+  (void)v_filter_w;
+  return _mm_avg_epu16(v_a_w, v_b_w);
+}
+
+static INLINE __m128i highbd_apply_filter(const __m128i v_a_w,
+                                          const __m128i v_b_w,
+                                          const __m128i v_filter_w) {
+  const __m128i v_rounding_d = _mm_set1_epi32(1 << (FILTER_BITS - 1));
+  __m128i v_input_lo_w = _mm_unpacklo_epi16(v_a_w, v_b_w);
+  __m128i v_input_hi_w = _mm_unpackhi_epi16(v_a_w, v_b_w);
+  __m128i v_temp0_d = _mm_madd_epi16(v_input_lo_w, v_filter_w);
+  __m128i v_temp1_d = _mm_madd_epi16(v_input_hi_w, v_filter_w);
+  __m128i v_res_lo_d =
+      _mm_srai_epi32(_mm_add_epi32(v_temp0_d, v_rounding_d), FILTER_BITS);
+  __m128i v_res_hi_d =
+      _mm_srai_epi32(_mm_add_epi32(v_temp1_d, v_rounding_d), FILTER_BITS);
+  return _mm_packs_epi32(v_res_lo_d, v_res_hi_d);
+}
+// Apply the filter to the contents of the lower half of a and b
+static INLINE void highbd_apply_filter_lo(const __m128i v_a_lo_w,
+                                          const __m128i v_b_lo_w,
+                                          const __m128i v_filter_w,
+                                          __m128i *v_res_d) {
+  const __m128i v_rounding_d = _mm_set1_epi32(1 << (FILTER_BITS - 1));
+  __m128i v_input_w = _mm_unpacklo_epi16(v_a_lo_w, v_b_lo_w);
+  __m128i v_temp0_d = _mm_madd_epi16(v_input_w, v_filter_w);
+  *v_res_d =
+      _mm_srai_epi32(_mm_add_epi32(v_temp0_d, v_rounding_d), FILTER_BITS);
+}
+
+static void highbd_sum_and_sse(const __m128i v_a_w, const __m128i v_b_w,
+                               const __m128i v_m_b, __m128i *v_sum_d,
+                               __m128i *v_sse_q) {
+  const __m128i v_zero = _mm_setzero_si128();
+  const __m128i v_m_w = _mm_unpacklo_epi8(v_m_b, v_zero);
+
+  // Difference: [-2^12, 2^12] => 13 bits (incld sign bit)
+  const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w);
+
+  // Error - [-4095, 4095] * [0, 64] & sum pairs => fits in 19 + 1 bits
+  const __m128i v_e_d = _mm_madd_epi16(v_d_w, v_m_w);
+
+  // Squared error - max (18 bits * 18 bits) = 36 bits (no sign bit)
+  const __m128i v_absd_w = _mm_abs_epi16(v_d_w);
+  const __m128i v_dlo_d = _mm_unpacklo_epi16(v_absd_w, v_zero);
+  const __m128i v_mlo_d = _mm_unpacklo_epi16(v_m_w, v_zero);
+  const __m128i v_elo_d = _mm_madd_epi16(v_dlo_d, v_mlo_d);
+  const __m128i v_dhi_d = _mm_unpackhi_epi16(v_absd_w, v_zero);
+  const __m128i v_mhi_d = _mm_unpackhi_epi16(v_m_w, v_zero);
+  const __m128i v_ehi_d = _mm_madd_epi16(v_dhi_d, v_mhi_d);
+  // Square and sum the errors -> 36bits * 4 = 38bits
+  __m128i v_se0_q, v_se1_q, v_se2_q, v_se3_q, v_se_q, v_elo1_d, v_ehi3_d;
+  v_se0_q = _mm_mul_epu32(v_elo_d, v_elo_d);
+  v_elo1_d = _mm_srli_si128(v_elo_d, 4);
+  v_se1_q = _mm_mul_epu32(v_elo1_d, v_elo1_d);
+  v_se0_q = _mm_add_epi64(v_se0_q, v_se1_q);
+  v_se2_q = _mm_mul_epu32(v_ehi_d, v_ehi_d);
+  v_ehi3_d = _mm_srli_si128(v_ehi_d, 4);
+  v_se3_q = _mm_mul_epu32(v_ehi3_d, v_ehi3_d);
+  v_se1_q = _mm_add_epi64(v_se2_q, v_se3_q);
+  v_se_q = _mm_add_epi64(v_se0_q, v_se1_q);
+
+  // Accumulate
+  *v_sum_d = _mm_add_epi32(*v_sum_d, v_e_d);
+  *v_sse_q = _mm_add_epi64(*v_sse_q, v_se_q);
+}
+
+static INLINE uint32_t highbd_10_calc_masked_variance(__m128i v_sum_d,
+                                                      __m128i v_sse_q,
+                                                      uint32_t *sse, int w,
+                                                      int h) {
+  int64_t sum64;
+  uint64_t sse64;
+
+  // Horizontal sum
+  sum64 = hsum_epi32_si32(v_sum_d);
+  sse64 = hsum_epi64_si64(v_sse_q);
+
+  sum64 = (sum64 >= 0) ? sum64 : -sum64;
+
+  // Round
+  sum64 = ROUND_POWER_OF_TWO(sum64, 6);
+  sse64 = ROUND_POWER_OF_TWO(sse64, 12);
+
+  // Normalise
+  sum64 = ROUND_POWER_OF_TWO(sum64, 2);
+  sse64 = ROUND_POWER_OF_TWO(sse64, 4);
+
+  // Store the SSE
+  *sse = (uint32_t)sse64;
+  // Compute the variance
+  return *sse - (uint32_t)((sum64 * sum64) / (w * h));
+}
+static INLINE uint32_t highbd_12_calc_masked_variance(__m128i v_sum_d,
+                                                      __m128i v_sse_q,
+                                                      uint32_t *sse, int w,
+                                                      int h) {
+  int64_t sum64;
+  uint64_t sse64;
+
+  // Horizontal sum
+  sum64 = hsum_epi32_si64(v_sum_d);
+  sse64 = hsum_epi64_si64(v_sse_q);
+
+  sum64 = (sum64 >= 0) ? sum64 : -sum64;
+
+  // Round
+  sum64 = ROUND_POWER_OF_TWO(sum64, 6);
+  sse64 = ROUND_POWER_OF_TWO(sse64, 12);
+
+  // Normalise
+  sum64 = ROUND_POWER_OF_TWO(sum64, 4);
+  sse64 = ROUND_POWER_OF_TWO(sse64, 8);
+
+  // Store the SSE
+  *sse = (uint32_t)sse64;
+  // Compute the variance
+  return *sse - (uint32_t)((sum64 * sum64) / (w * h));
+}
+
+// High bit depth functions for width (W) >= 8
+unsigned int aom_highbd_masked_subpel_varWxH_xzero(
+    const uint16_t *src, int src_stride, int yoffset, const uint16_t *dst,
+    int dst_stride, const uint8_t *msk, int msk_stride, unsigned int *sse,
+    int w, int h, highbd_filter_fn_t filter_fn,
+    highbd_calc_masked_var_t calc_var) {
+  int i, j;
+  __m128i v_src0_w, v_src1_w, v_res_w, v_dst_w, v_msk_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  const __m128i v_filter_w =
+      _mm_set1_epi32((bilinear_filters_2t[yoffset][1] << 16) +
+                     bilinear_filters_2t[yoffset][0]);
+  assert(yoffset < BIL_SUBPEL_SHIFTS);
+  for (j = 0; j < w; j += 8) {
+    // Load the first row ready
+    v_src0_w = _mm_loadu_si128((const __m128i *)(src + j));
+    // Process 2 rows at a time
+    for (i = 0; i < h; i += 2) {
+      // Load the next row apply the filter
+      v_src1_w = _mm_loadu_si128((const __m128i *)(src + j + src_stride));
+      v_res_w = filter_fn(v_src0_w, v_src1_w, v_filter_w);
+      // Load the dst and msk for the variance calculation
+      v_dst_w = _mm_loadu_si128((const __m128i *)(dst + j));
+      v_msk_b = _mm_loadl_epi64((const __m128i *)(msk + j));
+      highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q);
+
+      // Load the next row apply the filter
+      v_src0_w = _mm_loadu_si128((const __m128i *)(src + j + src_stride * 2));
+      v_res_w = filter_fn(v_src1_w, v_src0_w, v_filter_w);
+      // Load the dst and msk for the variance calculation
+      v_dst_w = _mm_loadu_si128((const __m128i *)(dst + j + dst_stride));
+      v_msk_b = _mm_loadl_epi64((const __m128i *)(msk + j + msk_stride));
+      highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q);
+      // Move onto the next block of rows
+      src += src_stride * 2;
+      dst += dst_stride * 2;
+      msk += msk_stride * 2;
+    }
+    // Reset to the top of the block
+    src -= src_stride * h;
+    dst -= dst_stride * h;
+    msk -= msk_stride * h;
+  }
+  return calc_var(v_sum_d, v_sse_q, sse, w, h);
+}
+unsigned int aom_highbd_masked_subpel_varWxH_yzero(
+    const uint16_t *src, int src_stride, int xoffset, const uint16_t *dst,
+    int dst_stride, const uint8_t *msk, int msk_stride, unsigned int *sse,
+    int w, int h, highbd_filter_fn_t filter_fn,
+    highbd_calc_masked_var_t calc_var) {
+  int i, j;
+  __m128i v_src0_w, v_src1_w, v_res_w, v_dst_w, v_msk_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  const __m128i v_filter_w =
+      _mm_set1_epi32((bilinear_filters_2t[xoffset][1] << 16) +
+                     bilinear_filters_2t[xoffset][0]);
+  assert(xoffset < BIL_SUBPEL_SHIFTS);
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j += 8) {
+      // Load this row & apply the filter to them
+      v_src0_w = _mm_loadu_si128((const __m128i *)(src + j));
+      v_src1_w = _mm_loadu_si128((const __m128i *)(src + j + 1));
+      v_res_w = filter_fn(v_src0_w, v_src1_w, v_filter_w);
+
+      // Load the dst and msk for the variance calculation
+      v_dst_w = _mm_loadu_si128((const __m128i *)(dst + j));
+      v_msk_b = _mm_loadl_epi64((const __m128i *)(msk + j));
+      highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q);
+    }
+    src += src_stride;
+    dst += dst_stride;
+    msk += msk_stride;
+  }
+  return calc_var(v_sum_d, v_sse_q, sse, w, h);
+}
+
+unsigned int aom_highbd_masked_subpel_varWxH_xnonzero_ynonzero(
+    const uint16_t *src, int src_stride, int xoffset, int yoffset,
+    const uint16_t *dst, int dst_stride, const uint8_t *msk, int msk_stride,
+    unsigned int *sse, int w, int h, highbd_filter_fn_t xfilter_fn,
+    highbd_filter_fn_t yfilter_fn, highbd_calc_masked_var_t calc_var) {
+  int i, j;
+  __m128i v_src0_w, v_src1_w, v_src2_w, v_src3_w;
+  __m128i v_filtered0_w, v_filtered1_w, v_res_w, v_dst_w, v_msk_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  const __m128i v_filterx_w =
+      _mm_set1_epi32((bilinear_filters_2t[xoffset][1] << 16) +
+                     bilinear_filters_2t[xoffset][0]);
+  const __m128i v_filtery_w =
+      _mm_set1_epi32((bilinear_filters_2t[yoffset][1] << 16) +
+                     bilinear_filters_2t[yoffset][0]);
+  assert(xoffset < BIL_SUBPEL_SHIFTS);
+  assert(yoffset < BIL_SUBPEL_SHIFTS);
+  for (j = 0; j < w; j += 8) {
+    // Load the first row ready
+    v_src0_w = _mm_loadu_si128((const __m128i *)(src + j));
+    v_src1_w = _mm_loadu_si128((const __m128i *)(src + j + 1));
+    v_filtered0_w = xfilter_fn(v_src0_w, v_src1_w, v_filterx_w);
+    // Process 2 rows at a time
+    for (i = 0; i < h; i += 2) {
+      // Load the next row & apply the filter
+      v_src2_w = _mm_loadu_si128((const __m128i *)(src + src_stride + j));
+      v_src3_w = _mm_loadu_si128((const __m128i *)(src + src_stride + j + 1));
+      v_filtered1_w = xfilter_fn(v_src2_w, v_src3_w, v_filterx_w);
+      // Load the dst and msk for the variance calculation
+      v_dst_w = _mm_loadu_si128((const __m128i *)(dst + j));
+      v_msk_b = _mm_loadl_epi64((const __m128i *)(msk + j));
+      // Complete the calculation for this row and add it to the running total
+      v_res_w = yfilter_fn(v_filtered0_w, v_filtered1_w, v_filtery_w);
+      highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q);
+
+      // Load the next row & apply the filter
+      v_src0_w = _mm_loadu_si128((const __m128i *)(src + src_stride * 2 + j));
+      v_src1_w =
+          _mm_loadu_si128((const __m128i *)(src + src_stride * 2 + j + 1));
+      v_filtered0_w = xfilter_fn(v_src0_w, v_src1_w, v_filterx_w);
+      // Load the dst and msk for the variance calculation
+      v_dst_w = _mm_loadu_si128((const __m128i *)(dst + dst_stride + j));
+      v_msk_b = _mm_loadl_epi64((const __m128i *)(msk + msk_stride + j));
+      // Complete the calculation for this row and add it to the running total
+      v_res_w = yfilter_fn(v_filtered1_w, v_filtered0_w, v_filtery_w);
+      highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q);
+      // Move onto the next block of rows
+      src += src_stride * 2;
+      dst += dst_stride * 2;
+      msk += msk_stride * 2;
+    }
+    // Reset to the top of the block
+    src -= src_stride * h;
+    dst -= dst_stride * h;
+    msk -= msk_stride * h;
+  }
+  return calc_var(v_sum_d, v_sse_q, sse, w, h);
+}
+
+// Note order in which rows loaded xmm[127:64] = row 1, xmm[63:0] = row 2
+unsigned int aom_highbd_masked_subpel_var4xH_xzero(
+    const uint16_t *src, int src_stride, int yoffset, const uint16_t *dst,
+    int dst_stride, const uint8_t *msk, int msk_stride, unsigned int *sse,
+    int h, highbd_calc_masked_var_t calc_var) {
+  int i;
+  __m128i v_src0_w, v_src1_w, v_filtered0_d, v_filtered1_d, v_res_w;
+  __m128i v_dst_w, v_msk_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  __m128i v_filter_w = _mm_set1_epi32((bilinear_filters_2t[yoffset][1] << 16) +
+                                      bilinear_filters_2t[yoffset][0]);
+  assert(yoffset < BIL_SUBPEL_SHIFTS);
+  // Load the first row of src data ready
+  v_src0_w = _mm_loadl_epi64((const __m128i *)src);
+  for (i = 0; i < h; i += 2) {
+    if (yoffset == HALF_PIXEL_OFFSET) {
+      // Load the rest of the source data for these rows
+      v_src1_w = _mm_or_si128(
+          _mm_slli_si128(v_src0_w, 8),
+          _mm_loadl_epi64((const __m128i *)(src + src_stride * 1)));
+      v_src0_w = _mm_or_si128(
+          _mm_slli_si128(v_src1_w, 8),
+          _mm_loadl_epi64((const __m128i *)(src + src_stride * 2)));
+      // Apply the y filter
+      v_res_w = _mm_avg_epu16(v_src1_w, v_src0_w);
+    } else {
+      // Load the data and apply the y filter
+      v_src1_w = _mm_loadl_epi64((const __m128i *)(src + src_stride * 1));
+      highbd_apply_filter_lo(v_src0_w, v_src1_w, v_filter_w, &v_filtered0_d);
+      v_src0_w = _mm_loadl_epi64((const __m128i *)(src + src_stride * 2));
+      highbd_apply_filter_lo(v_src1_w, v_src0_w, v_filter_w, &v_filtered1_d);
+      v_res_w = _mm_packs_epi32(v_filtered1_d, v_filtered0_d);
+    }
+    // Load the dst data
+    v_dst_w = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1)),
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0)));
+    // Load the mask data
+    v_msk_b = _mm_unpacklo_epi32(
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1)),
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0)));
+    // Compute the sum and SSE
+    highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q);
+    // Move onto the next set of rows
+    src += src_stride * 2;
+    dst += dst_stride * 2;
+    msk += msk_stride * 2;
+  }
+  return calc_var(v_sum_d, v_sse_q, sse, 4, h);
+}
+
+unsigned int aom_highbd_masked_subpel_var4xH_yzero(
+    const uint16_t *src, int src_stride, int xoffset, const uint16_t *dst,
+    int dst_stride, const uint8_t *msk, int msk_stride, unsigned int *sse,
+    int h, highbd_calc_masked_var_t calc_var) {
+  int i;
+  __m128i v_src0_w, v_src1_w, v_filtered0_d, v_filtered1_d;
+  __m128i v_src0_shift_w, v_src1_shift_w, v_res_w, v_dst_w, v_msk_b;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  __m128i v_filter_w = _mm_set1_epi32((bilinear_filters_2t[xoffset][1] << 16) +
+                                      bilinear_filters_2t[xoffset][0]);
+  assert(xoffset < BIL_SUBPEL_SHIFTS);
+  for (i = 0; i < h; i += 2) {
+    // Load the src data
+    v_src0_w = _mm_loadu_si128((const __m128i *)(src));
+    v_src0_shift_w = _mm_srli_si128(v_src0_w, 2);
+    v_src1_w = _mm_loadu_si128((const __m128i *)(src + src_stride));
+    v_src1_shift_w = _mm_srli_si128(v_src1_w, 2);
+    // Apply the x filter
+    if (xoffset == HALF_PIXEL_OFFSET) {
+      v_src1_w = _mm_unpacklo_epi64(v_src0_w, v_src1_w);
+      v_src1_shift_w = _mm_unpacklo_epi64(v_src0_shift_w, v_src1_shift_w);
+      v_res_w = _mm_avg_epu16(v_src1_w, v_src1_shift_w);
+    } else {
+      highbd_apply_filter_lo(v_src0_w, v_src0_shift_w, v_filter_w,
+                             &v_filtered0_d);
+      highbd_apply_filter_lo(v_src1_w, v_src1_shift_w, v_filter_w,
+                             &v_filtered1_d);
+      v_res_w = _mm_packs_epi32(v_filtered0_d, v_filtered1_d);
+    }
+    // Load the dst data
+    v_dst_w = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0)),
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1)));
+    // Load the mask data
+    v_msk_b = _mm_unpacklo_epi32(
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0)),
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1)));
+    // Compute the sum and SSE
+    highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q);
+    // Move onto the next set of rows
+    src += src_stride * 2;
+    dst += dst_stride * 2;
+    msk += msk_stride * 2;
+  }
+  return calc_var(v_sum_d, v_sse_q, sse, 4, h);
+}
+
+unsigned int aom_highbd_masked_subpel_var4xH_xnonzero_ynonzero(
+    const uint16_t *src, int src_stride, int xoffset, int yoffset,
+    const uint16_t *dst, int dst_stride, const uint8_t *msk, int msk_stride,
+    unsigned int *sse, int h, highbd_calc_masked_var_t calc_var) {
+  int i;
+  __m128i v_src0_w, v_src1_w, v_filtered0_d, v_filtered1_d, v_dst_w, v_msk_b;
+  __m128i v_src0_shift_w, v_src1_shift_w;
+  __m128i v_xres0_w, v_xres1_w, v_res_w, v_temp_w;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_q = _mm_setzero_si128();
+  __m128i v_filterx_w = _mm_set1_epi32((bilinear_filters_2t[xoffset][1] << 16) +
+                                       bilinear_filters_2t[xoffset][0]);
+  __m128i v_filtery_w = _mm_set1_epi32((bilinear_filters_2t[yoffset][1] << 16) +
+                                       bilinear_filters_2t[yoffset][0]);
+  assert(xoffset < BIL_SUBPEL_SHIFTS);
+  assert(yoffset < BIL_SUBPEL_SHIFTS);
+  // Load the first block of src data
+  v_src0_w = _mm_loadu_si128((const __m128i *)(src));
+  v_src0_shift_w = _mm_srli_si128(v_src0_w, 2);
+  v_src1_w = _mm_loadu_si128((const __m128i *)(src + src_stride));
+  v_src1_shift_w = _mm_srli_si128(v_src1_w, 2);
+  // Apply the x filter
+  if (xoffset == HALF_PIXEL_OFFSET) {
+    v_src1_w = _mm_unpacklo_epi64(v_src0_w, v_src1_w);
+    v_src1_shift_w = _mm_unpacklo_epi64(v_src0_shift_w, v_src1_shift_w);
+    v_xres0_w = _mm_avg_epu16(v_src1_w, v_src1_shift_w);
+  } else {
+    highbd_apply_filter_lo(v_src0_w, v_src0_shift_w, v_filterx_w,
+                           &v_filtered0_d);
+    highbd_apply_filter_lo(v_src1_w, v_src1_shift_w, v_filterx_w,
+                           &v_filtered1_d);
+    v_xres0_w = _mm_packs_epi32(v_filtered0_d, v_filtered1_d);
+  }
+  for (i = 0; i < h; i += 4) {
+    // Load the next block of src data
+    v_src0_w = _mm_loadu_si128((const __m128i *)(src + src_stride * 2));
+    v_src0_shift_w = _mm_srli_si128(v_src0_w, 2);
+    v_src1_w = _mm_loadu_si128((const __m128i *)(src + src_stride * 3));
+    v_src1_shift_w = _mm_srli_si128(v_src1_w, 2);
+    // Apply the x filter
+    if (xoffset == HALF_PIXEL_OFFSET) {
+      v_src1_w = _mm_unpacklo_epi64(v_src0_w, v_src1_w);
+      v_src1_shift_w = _mm_unpacklo_epi64(v_src0_shift_w, v_src1_shift_w);
+      v_xres1_w = _mm_avg_epu16(v_src1_w, v_src1_shift_w);
+    } else {
+      highbd_apply_filter_lo(v_src0_w, v_src0_shift_w, v_filterx_w,
+                             &v_filtered0_d);
+      highbd_apply_filter_lo(v_src1_w, v_src1_shift_w, v_filterx_w,
+                             &v_filtered1_d);
+      v_xres1_w = _mm_packs_epi32(v_filtered0_d, v_filtered1_d);
+    }
+    // Apply the y filter to the previous block
+    v_temp_w = _mm_or_si128(_mm_srli_si128(v_xres0_w, 8),
+                            _mm_slli_si128(v_xres1_w, 8));
+    if (yoffset == HALF_PIXEL_OFFSET) {
+      v_res_w = _mm_avg_epu16(v_xres0_w, v_temp_w);
+    } else {
+      v_res_w = highbd_apply_filter(v_xres0_w, v_temp_w, v_filtery_w);
+    }
+    // Load the dst data
+    v_dst_w = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 0)),
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 1)));
+    // Load the mask data
+    v_msk_b = _mm_unpacklo_epi32(
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 0)),
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 1)));
+    // Compute the sum and SSE
+    highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q);
+
+    // Load the next block of src data
+    v_src0_w = _mm_loadu_si128((const __m128i *)(src + src_stride * 4));
+    v_src0_shift_w = _mm_srli_si128(v_src0_w, 2);
+    v_src1_w = _mm_loadu_si128((const __m128i *)(src + src_stride * 5));
+    v_src1_shift_w = _mm_srli_si128(v_src1_w, 2);
+    // Apply the x filter
+    if (xoffset == HALF_PIXEL_OFFSET) {
+      v_src1_w = _mm_unpacklo_epi64(v_src0_w, v_src1_w);
+      v_src1_shift_w = _mm_unpacklo_epi64(v_src0_shift_w, v_src1_shift_w);
+      v_xres0_w = _mm_avg_epu16(v_src1_w, v_src1_shift_w);
+    } else {
+      highbd_apply_filter_lo(v_src0_w, v_src0_shift_w, v_filterx_w,
+                             &v_filtered0_d);
+      highbd_apply_filter_lo(v_src1_w, v_src1_shift_w, v_filterx_w,
+                             &v_filtered1_d);
+      v_xres0_w = _mm_packs_epi32(v_filtered0_d, v_filtered1_d);
+    }
+    // Apply the y filter to the previous block
+    v_temp_w = _mm_or_si128(_mm_srli_si128(v_xres1_w, 8),
+                            _mm_slli_si128(v_xres0_w, 8));
+    if (yoffset == HALF_PIXEL_OFFSET) {
+      v_res_w = _mm_avg_epu16(v_xres1_w, v_temp_w);
+    } else {
+      v_res_w = highbd_apply_filter(v_xres1_w, v_temp_w, v_filtery_w);
+    }
+    // Load the dst data
+    v_dst_w = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 2)),
+        _mm_loadl_epi64((const __m128i *)(dst + dst_stride * 3)));
+    // Load the mask data
+    v_msk_b = _mm_unpacklo_epi32(
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 2)),
+        _mm_loadl_epi64((const __m128i *)(msk + msk_stride * 3)));
+    // Compute the sum and SSE
+    highbd_sum_and_sse(v_res_w, v_dst_w, v_msk_b, &v_sum_d, &v_sse_q);
+    // Move onto the next set of rows
+    src += src_stride * 4;
+    dst += dst_stride * 4;
+    msk += msk_stride * 4;
+  }
+  return calc_var(v_sum_d, v_sse_q, sse, 4, h);
+}
+
+// For W >=8
+#define HIGHBD_MASK_SUBPIX_VAR_LARGE(W, H)                                     \
+  unsigned int highbd_masked_sub_pixel_variance##W##x##H##_ssse3(              \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *dst8, int dst_stride, const uint8_t *msk, int msk_stride, \
+      unsigned int *sse, highbd_calc_masked_var_t calc_var,                    \
+      highbd_variance_fn_t full_variance_function) {                           \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    assert(W % 8 == 0);                                                        \
+    if (xoffset == 0) {                                                        \
+      if (yoffset == 0)                                                        \
+        return full_variance_function(src8, src_stride, dst8, dst_stride, msk, \
+                                      msk_stride, sse);                        \
+      else if (yoffset == HALF_PIXEL_OFFSET)                                   \
+        return aom_highbd_masked_subpel_varWxH_xzero(                          \
+            src, src_stride, HALF_PIXEL_OFFSET, dst, dst_stride, msk,          \
+            msk_stride, sse, W, H, highbd_apply_filter_avg, calc_var);         \
+      else                                                                     \
+        return aom_highbd_masked_subpel_varWxH_xzero(                          \
+            src, src_stride, yoffset, dst, dst_stride, msk, msk_stride, sse,   \
+            W, H, highbd_apply_filter, calc_var);                              \
+    } else if (yoffset == 0) {                                                 \
+      if (xoffset == HALF_PIXEL_OFFSET)                                        \
+        return aom_highbd_masked_subpel_varWxH_yzero(                          \
+            src, src_stride, HALF_PIXEL_OFFSET, dst, dst_stride, msk,          \
+            msk_stride, sse, W, H, highbd_apply_filter_avg, calc_var);         \
+      else                                                                     \
+        return aom_highbd_masked_subpel_varWxH_yzero(                          \
+            src, src_stride, xoffset, dst, dst_stride, msk, msk_stride, sse,   \
+            W, H, highbd_apply_filter, calc_var);                              \
+    } else if (xoffset == HALF_PIXEL_OFFSET) {                                 \
+      if (yoffset == HALF_PIXEL_OFFSET)                                        \
+        return aom_highbd_masked_subpel_varWxH_xnonzero_ynonzero(              \
+            src, src_stride, HALF_PIXEL_OFFSET, HALF_PIXEL_OFFSET, dst,        \
+            dst_stride, msk, msk_stride, sse, W, H, highbd_apply_filter_avg,   \
+            highbd_apply_filter_avg, calc_var);                                \
+      else                                                                     \
+        return aom_highbd_masked_subpel_varWxH_xnonzero_ynonzero(              \
+            src, src_stride, HALF_PIXEL_OFFSET, yoffset, dst, dst_stride, msk, \
+            msk_stride, sse, W, H, highbd_apply_filter_avg,                    \
+            highbd_apply_filter, calc_var);                                    \
+    } else {                                                                   \
+      if (yoffset == HALF_PIXEL_OFFSET)                                        \
+        return aom_highbd_masked_subpel_varWxH_xnonzero_ynonzero(              \
+            src, src_stride, xoffset, HALF_PIXEL_OFFSET, dst, dst_stride, msk, \
+            msk_stride, sse, W, H, highbd_apply_filter,                        \
+            highbd_apply_filter_avg, calc_var);                                \
+      else                                                                     \
+        return aom_highbd_masked_subpel_varWxH_xnonzero_ynonzero(              \
+            src, src_stride, xoffset, yoffset, dst, dst_stride, msk,           \
+            msk_stride, sse, W, H, highbd_apply_filter, highbd_apply_filter,   \
+            calc_var);                                                         \
+    }                                                                          \
+  }
+
+// For W < 8
+#define HIGHBD_MASK_SUBPIX_VAR_SMALL(W, H)                                     \
+  unsigned int highbd_masked_sub_pixel_variance##W##x##H##_ssse3(              \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *dst8, int dst_stride, const uint8_t *msk, int msk_stride, \
+      unsigned int *sse, highbd_calc_masked_var_t calc_var,                    \
+      highbd_variance_fn_t full_variance_function) {                           \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    assert(W == 4);                                                            \
+    if (xoffset == 0 && yoffset == 0)                                          \
+      return full_variance_function(src8, src_stride, dst8, dst_stride, msk,   \
+                                    msk_stride, sse);                          \
+    else if (xoffset == 0)                                                     \
+      return aom_highbd_masked_subpel_var4xH_xzero(                            \
+          src, src_stride, yoffset, dst, dst_stride, msk, msk_stride, sse, H,  \
+          calc_var);                                                           \
+    else if (yoffset == 0)                                                     \
+      return aom_highbd_masked_subpel_var4xH_yzero(                            \
+          src, src_stride, xoffset, dst, dst_stride, msk, msk_stride, sse, H,  \
+          calc_var);                                                           \
+    else                                                                       \
+      return aom_highbd_masked_subpel_var4xH_xnonzero_ynonzero(                \
+          src, src_stride, xoffset, yoffset, dst, dst_stride, msk, msk_stride, \
+          sse, H, calc_var);                                                   \
+  }
+
+#define HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(W, H)                                  \
+  unsigned int aom_highbd_masked_sub_pixel_variance##W##x##H##_ssse3(          \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *dst8, int dst_stride, const uint8_t *msk, int msk_stride, \
+      unsigned int *sse) {                                                     \
+    return highbd_masked_sub_pixel_variance##W##x##H##_ssse3(                  \
+        src8, src_stride, xoffset, yoffset, dst8, dst_stride, msk, msk_stride, \
+        sse, calc_masked_variance,                                             \
+        aom_highbd_masked_variance##W##x##H##_ssse3);                          \
+  }                                                                            \
+  unsigned int aom_highbd_10_masked_sub_pixel_variance##W##x##H##_ssse3(       \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *dst8, int dst_stride, const uint8_t *msk, int msk_stride, \
+      unsigned int *sse) {                                                     \
+    return highbd_masked_sub_pixel_variance##W##x##H##_ssse3(                  \
+        src8, src_stride, xoffset, yoffset, dst8, dst_stride, msk, msk_stride, \
+        sse, highbd_10_calc_masked_variance,                                   \
+        aom_highbd_10_masked_variance##W##x##H##_ssse3);                       \
+  }                                                                            \
+  unsigned int aom_highbd_12_masked_sub_pixel_variance##W##x##H##_ssse3(       \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *dst8, int dst_stride, const uint8_t *msk, int msk_stride, \
+      unsigned int *sse) {                                                     \
+    return highbd_masked_sub_pixel_variance##W##x##H##_ssse3(                  \
+        src8, src_stride, xoffset, yoffset, dst8, dst_stride, msk, msk_stride, \
+        sse, highbd_12_calc_masked_variance,                                   \
+        aom_highbd_12_masked_variance##W##x##H##_ssse3);                       \
+  }
+
+HIGHBD_MASK_SUBPIX_VAR_SMALL(4, 4)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(4, 4)
+HIGHBD_MASK_SUBPIX_VAR_SMALL(4, 8)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(4, 8)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(8, 4)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(8, 4)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(8, 8)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(8, 8)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(8, 16)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(8, 16)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(16, 8)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(16, 8)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(16, 16)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(16, 16)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(16, 32)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(16, 32)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(32, 16)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(32, 16)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(32, 32)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(32, 32)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(32, 64)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(32, 64)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(64, 32)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(64, 32)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(64, 64)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(64, 64)
+#if CONFIG_EXT_PARTITION
+HIGHBD_MASK_SUBPIX_VAR_LARGE(64, 128)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(64, 128)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(128, 64)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(128, 64)
+HIGHBD_MASK_SUBPIX_VAR_LARGE(128, 128)
+HIGHBD_MASK_SUBPIX_VAR_WRAPPERS(128, 128)
+#endif  // CONFIG_EXT_PARTITION
+#endif
diff --git a/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c b/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c
new file mode 100644
index 0000000000..ad77f974c7
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c
@@ -0,0 +1,262 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <immintrin.h>
+
+#include "./aom_config.h"
+#include "aom_ports/mem.h"
+#include "aom/aom_integer.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/x86/synonyms.h"
+
+////////////////////////////////////////////////////////////////////////////////
+// 8 bit
+////////////////////////////////////////////////////////////////////////////////
+
+static INLINE unsigned int obmc_sad_w4(const uint8_t *pre, const int pre_stride,
+                                       const int32_t *wsrc, const int32_t *mask,
+                                       const int height) {
+  const int pre_step = pre_stride - 4;
+  int n = 0;
+  __m128i v_sad_d = _mm_setzero_si128();
+
+  do {
+    const __m128i v_p_b = xx_loadl_32(pre + n);
+    const __m128i v_m_d = xx_load_128(mask + n);
+    const __m128i v_w_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p_d = _mm_cvtepu8_epi32(v_p_b);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d);
+
+    const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d);
+    const __m128i v_absdiff_d = _mm_abs_epi32(v_diff_d);
+
+    // Rounded absolute difference
+    const __m128i v_rad_d = xx_roundn_epu32(v_absdiff_d, 12);
+
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad_d);
+
+    n += 4;
+
+    if (n % 4 == 0) pre += pre_step;
+  } while (n < 4 * height);
+
+  return xx_hsum_epi32_si32(v_sad_d);
+}
+
+static INLINE unsigned int obmc_sad_w8n(const uint8_t *pre,
+                                        const int pre_stride,
+                                        const int32_t *wsrc,
+                                        const int32_t *mask, const int width,
+                                        const int height) {
+  const int pre_step = pre_stride - width;
+  int n = 0;
+  __m128i v_sad_d = _mm_setzero_si128();
+
+  assert(width >= 8);
+  assert(IS_POWER_OF_TWO(width));
+
+  do {
+    const __m128i v_p1_b = xx_loadl_32(pre + n + 4);
+    const __m128i v_m1_d = xx_load_128(mask + n + 4);
+    const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
+    const __m128i v_p0_b = xx_loadl_32(pre + n);
+    const __m128i v_m0_d = xx_load_128(mask + n);
+    const __m128i v_w0_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p0_d = _mm_cvtepu8_epi32(v_p0_b);
+    const __m128i v_p1_d = _mm_cvtepu8_epi32(v_p1_b);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
+    const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);
+
+    const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
+    const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
+    const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d);
+    const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d);
+
+    // Rounded absolute difference
+    const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12);
+    const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12);
+
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d);
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d);
+
+    n += 8;
+
+    if (n % width == 0) pre += pre_step;
+  } while (n < width * height);
+
+  return xx_hsum_epi32_si32(v_sad_d);
+}
+
+#define OBMCSADWXH(w, h)                                       \
+  unsigned int aom_obmc_sad##w##x##h##_sse4_1(                 \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc, \
+      const int32_t *msk) {                                    \
+    if (w == 4) {                                              \
+      return obmc_sad_w4(pre, pre_stride, wsrc, msk, h);       \
+    } else {                                                   \
+      return obmc_sad_w8n(pre, pre_stride, wsrc, msk, w, h);   \
+    }                                                          \
+  }
+
+#if CONFIG_EXT_PARTITION
+OBMCSADWXH(128, 128)
+OBMCSADWXH(128, 64)
+OBMCSADWXH(64, 128)
+#endif  // CONFIG_EXT_PARTITION
+OBMCSADWXH(64, 64)
+OBMCSADWXH(64, 32)
+OBMCSADWXH(32, 64)
+OBMCSADWXH(32, 32)
+OBMCSADWXH(32, 16)
+OBMCSADWXH(16, 32)
+OBMCSADWXH(16, 16)
+OBMCSADWXH(16, 8)
+OBMCSADWXH(8, 16)
+OBMCSADWXH(8, 8)
+OBMCSADWXH(8, 4)
+OBMCSADWXH(4, 8)
+OBMCSADWXH(4, 4)
+
+////////////////////////////////////////////////////////////////////////////////
+// High bit-depth
+////////////////////////////////////////////////////////////////////////////////
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE unsigned int hbd_obmc_sad_w4(const uint8_t *pre8,
+                                           const int pre_stride,
+                                           const int32_t *wsrc,
+                                           const int32_t *mask,
+                                           const int height) {
+  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
+  const int pre_step = pre_stride - 4;
+  int n = 0;
+  __m128i v_sad_d = _mm_setzero_si128();
+
+  do {
+    const __m128i v_p_w = xx_loadl_64(pre + n);
+    const __m128i v_m_d = xx_load_128(mask + n);
+    const __m128i v_w_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p_d = _mm_cvtepu16_epi32(v_p_w);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d);
+
+    const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d);
+    const __m128i v_absdiff_d = _mm_abs_epi32(v_diff_d);
+
+    // Rounded absolute difference
+    const __m128i v_rad_d = xx_roundn_epu32(v_absdiff_d, 12);
+
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad_d);
+
+    n += 4;
+
+    if (n % 4 == 0) pre += pre_step;
+  } while (n < 4 * height);
+
+  return xx_hsum_epi32_si32(v_sad_d);
+}
+
+static INLINE unsigned int hbd_obmc_sad_w8n(const uint8_t *pre8,
+                                            const int pre_stride,
+                                            const int32_t *wsrc,
+                                            const int32_t *mask,
+                                            const int width, const int height) {
+  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
+  const int pre_step = pre_stride - width;
+  int n = 0;
+  __m128i v_sad_d = _mm_setzero_si128();
+
+  assert(width >= 8);
+  assert(IS_POWER_OF_TWO(width));
+
+  do {
+    const __m128i v_p1_w = xx_loadl_64(pre + n + 4);
+    const __m128i v_m1_d = xx_load_128(mask + n + 4);
+    const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
+    const __m128i v_p0_w = xx_loadl_64(pre + n);
+    const __m128i v_m0_d = xx_load_128(mask + n);
+    const __m128i v_w0_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p0_d = _mm_cvtepu16_epi32(v_p0_w);
+    const __m128i v_p1_d = _mm_cvtepu16_epi32(v_p1_w);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
+    const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);
+
+    const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
+    const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
+    const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d);
+    const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d);
+
+    // Rounded absolute difference
+    const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12);
+    const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12);
+
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d);
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d);
+
+    n += 8;
+
+    if (n % width == 0) pre += pre_step;
+  } while (n < width * height);
+
+  return xx_hsum_epi32_si32(v_sad_d);
+}
+
+#define HBD_OBMCSADWXH(w, h)                                      \
+  unsigned int aom_highbd_obmc_sad##w##x##h##_sse4_1(             \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,    \
+      const int32_t *mask) {                                      \
+    if (w == 4) {                                                 \
+      return hbd_obmc_sad_w4(pre, pre_stride, wsrc, mask, h);     \
+    } else {                                                      \
+      return hbd_obmc_sad_w8n(pre, pre_stride, wsrc, mask, w, h); \
+    }                                                             \
+  }
+
+#if CONFIG_EXT_PARTITION
+HBD_OBMCSADWXH(128, 128)
+HBD_OBMCSADWXH(128, 64)
+HBD_OBMCSADWXH(64, 128)
+#endif  // CONFIG_EXT_PARTITION
+HBD_OBMCSADWXH(64, 64)
+HBD_OBMCSADWXH(64, 32)
+HBD_OBMCSADWXH(32, 64)
+HBD_OBMCSADWXH(32, 32)
+HBD_OBMCSADWXH(32, 16)
+HBD_OBMCSADWXH(16, 32)
+HBD_OBMCSADWXH(16, 16)
+HBD_OBMCSADWXH(16, 8)
+HBD_OBMCSADWXH(8, 16)
+HBD_OBMCSADWXH(8, 8)
+HBD_OBMCSADWXH(8, 4)
+HBD_OBMCSADWXH(4, 8)
+HBD_OBMCSADWXH(4, 4)
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c
new file mode 100644
index 0000000000..efb3659cfc
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c
@@ -0,0 +1,355 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <immintrin.h>
+
+#include "./aom_config.h"
+#include "aom_ports/mem.h"
+#include "aom/aom_integer.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/aom_filter.h"
+
+////////////////////////////////////////////////////////////////////////////////
+// 8 bit
+////////////////////////////////////////////////////////////////////////////////
+
+static INLINE void obmc_variance_w4(const uint8_t *pre, const int pre_stride,
+                                    const int32_t *wsrc, const int32_t *mask,
+                                    unsigned int *const sse, int *const sum,
+                                    const int h) {
+  const int pre_step = pre_stride - 4;
+  int n = 0;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_d = _mm_setzero_si128();
+
+  assert(IS_POWER_OF_TWO(h));
+
+  do {
+    const __m128i v_p_b = xx_loadl_32(pre + n);
+    const __m128i v_m_d = xx_load_128(mask + n);
+    const __m128i v_w_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p_d = _mm_cvtepu8_epi32(v_p_b);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d);
+
+    const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d);
+    const __m128i v_rdiff_d = xx_roundn_epi32(v_diff_d, 12);
+    const __m128i v_sqrdiff_d = _mm_mullo_epi32(v_rdiff_d, v_rdiff_d);
+
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d);
+    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);
+
+    n += 4;
+
+    if (n % 4 == 0) pre += pre_step;
+  } while (n < 4 * h);
+
+  *sum = xx_hsum_epi32_si32(v_sum_d);
+  *sse = xx_hsum_epi32_si32(v_sse_d);
+}
+
+static INLINE void obmc_variance_w8n(const uint8_t *pre, const int pre_stride,
+                                     const int32_t *wsrc, const int32_t *mask,
+                                     unsigned int *const sse, int *const sum,
+                                     const int w, const int h) {
+  const int pre_step = pre_stride - w;
+  int n = 0;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_d = _mm_setzero_si128();
+
+  assert(w >= 8);
+  assert(IS_POWER_OF_TWO(w));
+  assert(IS_POWER_OF_TWO(h));
+
+  do {
+    const __m128i v_p1_b = xx_loadl_32(pre + n + 4);
+    const __m128i v_m1_d = xx_load_128(mask + n + 4);
+    const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
+    const __m128i v_p0_b = xx_loadl_32(pre + n);
+    const __m128i v_m0_d = xx_load_128(mask + n);
+    const __m128i v_w0_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p0_d = _mm_cvtepu8_epi32(v_p0_b);
+    const __m128i v_p1_d = _mm_cvtepu8_epi32(v_p1_b);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
+    const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);
+
+    const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
+    const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
+
+    const __m128i v_rdiff0_d = xx_roundn_epi32(v_diff0_d, 12);
+    const __m128i v_rdiff1_d = xx_roundn_epi32(v_diff1_d, 12);
+    const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff0_d, v_rdiff1_d);
+    const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w);
+
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff0_d);
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d);
+    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);
+
+    n += 8;
+
+    if (n % w == 0) pre += pre_step;
+  } while (n < w * h);
+
+  *sum = xx_hsum_epi32_si32(v_sum_d);
+  *sse = xx_hsum_epi32_si32(v_sse_d);
+}
+
+#define OBMCVARWXH(W, H)                                               \
+  unsigned int aom_obmc_variance##W##x##H##_sse4_1(                    \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,         \
+      const int32_t *mask, unsigned int *sse) {                        \
+    int sum;                                                           \
+    if (W == 4) {                                                      \
+      obmc_variance_w4(pre, pre_stride, wsrc, mask, sse, &sum, H);     \
+    } else {                                                           \
+      obmc_variance_w8n(pre, pre_stride, wsrc, mask, sse, &sum, W, H); \
+    }                                                                  \
+    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H));      \
+  }
+
+#if CONFIG_EXT_PARTITION
+OBMCVARWXH(128, 128)
+OBMCVARWXH(128, 64)
+OBMCVARWXH(64, 128)
+#endif  // CONFIG_EXT_PARTITION
+OBMCVARWXH(64, 64)
+OBMCVARWXH(64, 32)
+OBMCVARWXH(32, 64)
+OBMCVARWXH(32, 32)
+OBMCVARWXH(32, 16)
+OBMCVARWXH(16, 32)
+OBMCVARWXH(16, 16)
+OBMCVARWXH(16, 8)
+OBMCVARWXH(8, 16)
+OBMCVARWXH(8, 8)
+OBMCVARWXH(8, 4)
+OBMCVARWXH(4, 8)
+OBMCVARWXH(4, 4)
+
+////////////////////////////////////////////////////////////////////////////////
+// High bit-depth
+////////////////////////////////////////////////////////////////////////////////
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE void hbd_obmc_variance_w4(
+    const uint8_t *pre8, const int pre_stride, const int32_t *wsrc,
+    const int32_t *mask, uint64_t *const sse, int64_t *const sum, const int h) {
+  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
+  const int pre_step = pre_stride - 4;
+  int n = 0;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_d = _mm_setzero_si128();
+
+  assert(IS_POWER_OF_TWO(h));
+
+  do {
+    const __m128i v_p_w = xx_loadl_64(pre + n);
+    const __m128i v_m_d = xx_load_128(mask + n);
+    const __m128i v_w_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p_d = _mm_cvtepu16_epi32(v_p_w);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d);
+
+    const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d);
+    const __m128i v_rdiff_d = xx_roundn_epi32(v_diff_d, 12);
+    const __m128i v_sqrdiff_d = _mm_mullo_epi32(v_rdiff_d, v_rdiff_d);
+
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d);
+    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);
+
+    n += 4;
+
+    if (n % 4 == 0) pre += pre_step;
+  } while (n < 4 * h);
+
+  *sum = xx_hsum_epi32_si32(v_sum_d);
+  *sse = xx_hsum_epi32_si32(v_sse_d);
+}
+
+static INLINE void hbd_obmc_variance_w8n(
+    const uint8_t *pre8, const int pre_stride, const int32_t *wsrc,
+    const int32_t *mask, uint64_t *const sse, int64_t *const sum, const int w,
+    const int h) {
+  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
+  const int pre_step = pre_stride - w;
+  int n = 0;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_d = _mm_setzero_si128();
+
+  assert(w >= 8);
+  assert(IS_POWER_OF_TWO(w));
+  assert(IS_POWER_OF_TWO(h));
+
+  do {
+    const __m128i v_p1_w = xx_loadl_64(pre + n + 4);
+    const __m128i v_m1_d = xx_load_128(mask + n + 4);
+    const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
+    const __m128i v_p0_w = xx_loadl_64(pre + n);
+    const __m128i v_m0_d = xx_load_128(mask + n);
+    const __m128i v_w0_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p0_d = _mm_cvtepu16_epi32(v_p0_w);
+    const __m128i v_p1_d = _mm_cvtepu16_epi32(v_p1_w);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
+    const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);
+
+    const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
+    const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
+
+    const __m128i v_rdiff0_d = xx_roundn_epi32(v_diff0_d, 12);
+    const __m128i v_rdiff1_d = xx_roundn_epi32(v_diff1_d, 12);
+    const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff0_d, v_rdiff1_d);
+    const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w);
+
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff0_d);
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d);
+    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);
+
+    n += 8;
+
+    if (n % w == 0) pre += pre_step;
+  } while (n < w * h);
+
+  *sum += xx_hsum_epi32_si64(v_sum_d);
+  *sse += xx_hsum_epi32_si64(v_sse_d);
+}
+
+static INLINE void highbd_obmc_variance(const uint8_t *pre8, int pre_stride,
+                                        const int32_t *wsrc,
+                                        const int32_t *mask, int w, int h,
+                                        unsigned int *sse, int *sum) {
+  int64_t sum64 = 0;
+  uint64_t sse64 = 0;
+  if (w == 4) {
+    hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h);
+  } else {
+    hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h);
+  }
+  *sum = (int)sum64;
+  *sse = (unsigned int)sse64;
+}
+
+static INLINE void highbd_10_obmc_variance(const uint8_t *pre8, int pre_stride,
+                                           const int32_t *wsrc,
+                                           const int32_t *mask, int w, int h,
+                                           unsigned int *sse, int *sum) {
+  int64_t sum64 = 0;
+  uint64_t sse64 = 0;
+  if (w == 4) {
+    hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h);
+  } else {
+    hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h);
+  }
+  *sum = (int)ROUND_POWER_OF_TWO(sum64, 2);
+  *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 4);
+}
+
+static INLINE void highbd_12_obmc_variance(const uint8_t *pre8, int pre_stride,
+                                           const int32_t *wsrc,
+                                           const int32_t *mask, int w, int h,
+                                           unsigned int *sse, int *sum) {
+  int64_t sum64 = 0;
+  uint64_t sse64 = 0;
+  if (w == 128) {
+    do {
+      hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, 128,
+                            32);
+      pre8 += 32 * pre_stride;
+      wsrc += 32 * 128;
+      mask += 32 * 128;
+      h -= 32;
+    } while (h > 0);
+  } else if (w == 64 && h >= 128) {
+    do {
+      hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, 64,
+                            64);
+      pre8 += 64 * pre_stride;
+      wsrc += 64 * 64;
+      mask += 64 * 64;
+      h -= 64;
+    } while (h > 0);
+  } else if (w == 4) {
+    hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h);
+  } else {
+    hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h);
+  }
+  *sum = (int)ROUND_POWER_OF_TWO(sum64, 4);
+  *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 8);
+}
+
+#define HBD_OBMCVARWXH(W, H)                                               \
+  unsigned int aom_highbd_obmc_variance##W##x##H##_sse4_1(                 \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
+      const int32_t *mask, unsigned int *sse) {                            \
+    int sum;                                                               \
+    highbd_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum);    \
+    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H));          \
+  }                                                                        \
+                                                                           \
+  unsigned int aom_highbd_10_obmc_variance##W##x##H##_sse4_1(              \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
+      const int32_t *mask, unsigned int *sse) {                            \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    highbd_10_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));              \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }                                                                        \
+                                                                           \
+  unsigned int aom_highbd_12_obmc_variance##W##x##H##_sse4_1(              \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
+      const int32_t *mask, unsigned int *sse) {                            \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    highbd_12_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));              \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }
+
+#if CONFIG_EXT_PARTITION
+HBD_OBMCVARWXH(128, 128)
+HBD_OBMCVARWXH(128, 64)
+HBD_OBMCVARWXH(64, 128)
+#endif  // CONFIG_EXT_PARTITION
+HBD_OBMCVARWXH(64, 64)
+HBD_OBMCVARWXH(64, 32)
+HBD_OBMCVARWXH(32, 64)
+HBD_OBMCVARWXH(32, 32)
+HBD_OBMCVARWXH(32, 16)
+HBD_OBMCVARWXH(16, 32)
+HBD_OBMCVARWXH(16, 16)
+HBD_OBMCVARWXH(16, 8)
+HBD_OBMCVARWXH(8, 16)
+HBD_OBMCVARWXH(8, 8)
+HBD_OBMCVARWXH(8, 4)
+HBD_OBMCVARWXH(4, 8)
+HBD_OBMCVARWXH(4, 4)
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/quantize_avx_x86_64.asm b/third_party/aom/aom_dsp/x86/quantize_avx_x86_64.asm
new file mode 100644
index 0000000000..954a95b983
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/quantize_avx_x86_64.asm
@@ -0,0 +1,547 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro QUANTIZE_FN 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, \
+                                eob, scan, iscan
+
+  vzeroupper
+
+  ; If we can skip this block, then just zero the output
+  cmp                         skipmp, 0
+  jne .blank
+
+%ifnidn %1, b_32x32
+
+  ; Special case for ncoeff == 16, as it is frequent and we can save on
+  ; not setting up a loop.
+  cmp                       ncoeffmp, 16
+  jne .generic
+
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+  ;; Special case of ncoeff == 16
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+.single:
+
+  movifnidn                   coeffq, coeffmp
+  movifnidn                    zbinq, zbinmp
+  mova                            m0, [zbinq]              ; m0 = zbin
+
+  ; Get DC and first 15 AC coeffs - in this special case, that is all.
+%if CONFIG_HIGHBITDEPTH
+  ; coeff stored as 32bit numbers but we process them as 16 bit numbers
+  mova                            m9, [coeffq]
+  packssdw                        m9, [coeffq+16]          ; m9 = c[i]
+  mova                           m10, [coeffq+32]
+  packssdw                       m10, [coeffq+48]          ; m10 = c[i]
+%else
+  mova                            m9, [coeffq]             ; m9 = c[i]
+  mova                           m10, [coeffq+16]          ; m10 = c[i]
+%endif
+
+  mov                             r0, eobmp                ; Output pointer
+  mov                             r1, qcoeffmp             ; Output pointer
+  mov                             r2, dqcoeffmp            ; Output pointer
+
+  pxor                            m5, m5                   ; m5 = dedicated zero
+
+  pcmpeqw                         m4, m4                   ; All word lanes -1
+  paddw                           m0, m4                   ; m0 = zbin - 1
+
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  punpckhqdq                      m0, m0
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+
+  ; Check if all coeffs are less than zbin. If yes, we just write zeros
+  ; to the outputs and we are done.
+  por                            m14, m7, m12
+  ptest                          m14, m14
+  jnz .single_nonzero
+
+%if CONFIG_HIGHBITDEPTH
+  mova                       [r1   ], ymm5
+  mova                       [r1+32], ymm5
+  mova                       [r2   ], ymm5
+  mova                       [r2+32], ymm5
+%else
+  mova                          [r1], ymm5
+  mova                          [r2], ymm5
+%endif
+  mov                           [r0], word 0
+
+  vzeroupper
+  RET
+
+.single_nonzero:
+
+  ; Actual quantization of size 16 block - setup pointers, rounders, etc.
+  movifnidn                       r4, roundmp
+  movifnidn                       r5, quantmp
+  mov                             r3, dequantmp
+  mov                             r6, shiftmp
+  mova                            m1, [r4]              ; m1 = round
+  mova                            m2, [r5]              ; m2 = quant
+  mova                            m3, [r3]              ; m3 = dequant
+  mova                            m4, [r6]              ; m4 = shift
+
+  mov                             r3, iscanmp
+
+  DEFINE_ARGS eob, qcoeff, dqcoeff, iscan
+
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                           m8, m6                   ; m8 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                          m8, m4                   ; m8 = m8*qsh>>16
+  punpckhqdq                      m4, m4
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                            m8, m7
+  pand                           m13, m12
+
+%if CONFIG_HIGHBITDEPTH
+  ; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m8
+  punpckhwd                       m6, m8, m6
+  pmovsxwd                       m11, m8
+  mova                  [qcoeffq   ], m11
+  mova                  [qcoeffq+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova                  [qcoeffq+32], m11
+  mova                  [qcoeffq+48], m6
+%else
+  mova                  [qcoeffq   ], m8
+  mova                  [qcoeffq+16], m13
+%endif
+
+  pmullw                          m8, m3                   ; dqc[i] = qc[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+
+%if CONFIG_HIGHBITDEPTH
+  ; Store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m8
+  punpckhwd                       m6, m8, m6
+  pmovsxwd                       m11, m8
+  mova                 [dqcoeffq   ], m11
+  mova                 [dqcoeffq+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova                 [dqcoeffq+32], m11
+  mova                 [dqcoeffq+48], m6
+%else
+  mova                 [dqcoeffq   ], m8
+  mova                 [dqcoeffq+16], m13
+%endif
+
+  mova                            m6, [iscanq]            ; m6 = scan[i]
+  mova                           m11, [iscanq+16]         ; m11 = scan[i]
+
+  pcmpeqw                         m8,  m8,  m5            ; m8 = c[i] == 0
+  pcmpeqw                        m13, m13,  m5            ; m13 = c[i] == 0
+  psubw                           m6,  m6,  m7            ; m6 = scan[i] + 1
+  psubw                          m11, m11, m12            ; m11 = scan[i] + 1
+  pandn                           m8,  m8,  m6            ; m8 = max(eob)
+  pandn                          m13, m13, m11            ; m13 = max(eob)
+  pmaxsw                          m8,  m8, m13
+
+  ; Horizontally accumulate/max eobs and write into [eob] memory pointer
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  movq                           rax, m8
+  mov                         [eobq], ax
+
+  vzeroupper
+  RET
+
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+  ;; Generic case of ncoeff != 16
+  ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+.generic:
+
+%endif ; %ifnidn %1, b_32x32
+
+DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
+            qcoeff, dqcoeff, dequant, eob, scan, iscan
+
+  ; Actual quantization loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, dequantmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  mova                            m0, [zbinq]              ; m0 = zbin
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+  mova                            m3, [r2]                 ; m3 = dequant
+  pcmpeqw                         m4, m4                   ; All lanes -1
+%ifidn %1, b_32x32
+  psubw                           m0, m4
+  psubw                           m1, m4
+  psrlw                           m0, 1                    ; m0 = (m0 + 1) / 2
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  paddw                           m0, m4                   ; m0 = m0 + 1
+
+  mov                             r2, shiftmp
+  mov                             r3, qcoeffmp
+  mova                            m4, [r2]                 ; m4 = shift
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+%ifidn %1, b_32x32
+  psllw                           m4, 1
+%endif
+  pxor                            m5, m5                   ; m5 = dedicated zero
+
+  DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
+
+%if CONFIG_HIGHBITDEPTH
+  lea                         coeffq, [  coeffq+ncoeffq*4]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*4]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*4]
+%else
+  lea                         coeffq, [  coeffq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+%endif
+  lea                         iscanq, [  iscanq+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+%if CONFIG_HIGHBITDEPTH
+  ; coeff stored as 32bit numbers & require 16bit numbers
+  mova                            m9, [coeffq+ncoeffq*4+ 0]
+  packssdw                        m9, [coeffq+ncoeffq*4+16]
+  mova                           m10, [coeffq+ncoeffq*4+32]
+  packssdw                       m10, [coeffq+ncoeffq*4+48]
+%else
+  mova                            m9, [coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [coeffq+ncoeffq*2+16] ; m10 = c[i]
+%endif
+
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  punpckhqdq                      m0, m0
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+
+  ; Check if all coeffs are less than zbin. If yes, skip forward quickly.
+  por                            m14, m7, m12
+  ptest                          m14, m14
+  jnz .first_nonzero
+
+%if CONFIG_HIGHBITDEPTH
+  mova        [qcoeffq+ncoeffq*4   ], ymm5
+  mova        [qcoeffq+ncoeffq*4+32], ymm5
+  mova       [dqcoeffq+ncoeffq*4   ], ymm5
+  mova       [dqcoeffq+ncoeffq*4+32], ymm5
+%else
+  mova           [qcoeffq+ncoeffq*2], ymm5
+  mova          [dqcoeffq+ncoeffq*2], ymm5
+%endif
+
+  add                        ncoeffq, mmsize
+
+  punpckhqdq                      m1, m1
+  punpckhqdq                      m2, m2
+  punpckhqdq                      m3, m3
+  punpckhqdq                      m4, m4
+  pxor                            m8, m8
+
+  jmp .ac_only_loop
+
+.first_nonzero:
+
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                           m8, m6                   ; m8 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                          m8, m4                   ; m8 = m8*qsh>>16
+  punpckhqdq                      m4, m4
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                            m8, m7
+  pand                           m13, m12
+
+%if CONFIG_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m8
+  punpckhwd                       m6, m8, m6
+  pmovsxwd                       m11, m8
+  mova        [qcoeffq+ncoeffq*4+ 0], m11
+  mova        [qcoeffq+ncoeffq*4+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova        [qcoeffq+ncoeffq*4+32], m11
+  mova        [qcoeffq+ncoeffq*4+48], m6
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], m8
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%endif
+
+%ifidn %1, b_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; dqc[i] = qc[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+%endif
+
+%if CONFIG_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m8
+  punpckhwd                       m6, m8, m6
+  pmovsxwd                       m11, m8
+  mova       [dqcoeffq+ncoeffq*4+ 0], m11
+  mova       [dqcoeffq+ncoeffq*4+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova       [dqcoeffq+ncoeffq*4+32], m11
+  mova       [dqcoeffq+ncoeffq*4+48], m6
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], m8
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+%endif
+
+  pcmpeqw                         m8, m5                    ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                    ; m13 = c[i] == 0
+  mova                            m6, [iscanq+ncoeffq*2]    ; m6 = scan[i]
+  mova                           m11, [iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                    ; m6 = scan[i] + 1
+  psubw                          m11, m12                   ; m11 = scan[i] + 1
+  pandn                           m8, m6                    ; m8 = max(eob)
+  pandn                          m13, m11                   ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+
+.ac_only_loop:
+
+%if CONFIG_HIGHBITDEPTH
+  ; pack coeff from 32bit to 16bit array
+  mova                            m9, [coeffq+ncoeffq*4+ 0]
+  packssdw                        m9, [coeffq+ncoeffq*4+16]
+  mova                           m10, [coeffq+ncoeffq*4+32]
+  packssdw                       m10, [coeffq+ncoeffq*4+48]
+%else
+  mova                            m9, [coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [coeffq+ncoeffq*2+16] ; m10 = c[i]
+%endif
+
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+
+  ; Check if all coeffs are less than zbin. If yes, skip this itertion.
+  ; And just write zeros as the result would be.
+  por                            m14, m7, m12
+  ptest                          m14, m14
+  jnz .rest_nonzero
+
+%if CONFIG_HIGHBITDEPTH
+  mova        [qcoeffq+ncoeffq*4+ 0], ymm5
+  mova        [qcoeffq+ncoeffq*4+32], ymm5
+  mova       [dqcoeffq+ncoeffq*4+ 0], ymm5
+  mova       [dqcoeffq+ncoeffq*4+32], ymm5
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], ymm5
+  mova       [dqcoeffq+ncoeffq*2+ 0], ymm5
+%endif
+  add                        ncoeffq, mmsize
+  jnz .ac_only_loop
+
+  ; Horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  movq                           rax, m8
+  mov                           [r2], ax
+  vzeroupper
+  RET
+
+.rest_nonzero:
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                          m14, m6                   ; m14 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                         m14, m4                   ; m14 = m14*qsh>>16
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                           m14, m7
+  pand                           m13, m12
+
+%if CONFIG_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m14
+  punpckhwd                       m6, m14, m6
+  pmovsxwd                       m11, m14
+  mova        [qcoeffq+ncoeffq*4+ 0], m11
+  mova        [qcoeffq+ncoeffq*4+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova        [qcoeffq+ncoeffq*4+32], m11
+  mova        [qcoeffq+ncoeffq*4+48], m6
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], m14
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%endif
+
+%ifidn %1, b_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; dqc[i] = qc[i] * q
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+
+%if CONFIG_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pcmpgtw                         m6, m5, m14
+  punpckhwd                       m6, m14, m6
+  pmovsxwd                       m11, m14
+  mova       [dqcoeffq+ncoeffq*4+ 0], m11
+  mova       [dqcoeffq+ncoeffq*4+16], m6
+  pcmpgtw                         m6, m5, m13
+  punpckhwd                       m6, m13, m6
+  pmovsxwd                       m11, m13
+  mova       [dqcoeffq+ncoeffq*4+32], m11
+  mova       [dqcoeffq+ncoeffq*4+48], m6
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], m14
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+%endif
+
+  pcmpeqw                        m14, m5                    ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                    ; m13 = c[i] == 0
+  mova                            m6, [iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                    ; m6 = scan[i] + 1
+  psubw                          m11, m12                   ; m11 = scan[i] + 1
+  pandn                          m14, m6                    ; m14 = max(eob)
+  pandn                          m13, m11                   ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jnz .ac_only_loop
+
+  ; Horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  movq                           rax, m8
+  mov                           [r2], ax
+  vzeroupper
+  RET
+
+  ; Skip-block, i.e. just write all zeroes
+.blank:
+
+DEFINE_ARGS coeff, ncoeff, skip, zbin, round, quant, shift, \
+            qcoeff, dqcoeff, dequant, eob, scan, iscan
+
+  mov                             r0, dqcoeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, qcoeffmp
+  mov                             r3, eobmp
+
+DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
+
+%if CONFIG_HIGHBITDEPTH
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*4]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*4]
+%else
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+%endif
+
+  neg                        ncoeffq
+  pxor                            m7, m7
+
+.blank_loop:
+%if CONFIG_HIGHBITDEPTH
+  mova       [dqcoeffq+ncoeffq*4+ 0], ymm7
+  mova       [dqcoeffq+ncoeffq*4+32], ymm7
+  mova        [qcoeffq+ncoeffq*4+ 0], ymm7
+  mova        [qcoeffq+ncoeffq*4+32], ymm7
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], ymm7
+  mova        [qcoeffq+ncoeffq*2+ 0], ymm7
+%endif
+  add                        ncoeffq, mmsize
+  jl .blank_loop
+
+  mov                         [eobq], word 0
+
+  vzeroupper
+  RET
+%endmacro
+
+INIT_XMM avx
+QUANTIZE_FN b, 7
+QUANTIZE_FN b_32x32, 7
+
+END
diff --git a/third_party/aom/aom_dsp/x86/quantize_sse2.c b/third_party/aom/aom_dsp/x86/quantize_sse2.c
new file mode 100644
index 0000000000..890c1f01e8
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/quantize_sse2.c
@@ -0,0 +1,249 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>
+#include <xmmintrin.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+static INLINE __m128i load_coefficients(const tran_low_t *coeff_ptr) {
+#if CONFIG_HIGHBITDEPTH
+  return _mm_setr_epi16((int16_t)coeff_ptr[0], (int16_t)coeff_ptr[1],
+                        (int16_t)coeff_ptr[2], (int16_t)coeff_ptr[3],
+                        (int16_t)coeff_ptr[4], (int16_t)coeff_ptr[5],
+                        (int16_t)coeff_ptr[6], (int16_t)coeff_ptr[7]);
+#else
+  return _mm_load_si128((const __m128i *)coeff_ptr);
+#endif
+}
+
+static INLINE void store_coefficients(__m128i coeff_vals,
+                                      tran_low_t *coeff_ptr) {
+#if CONFIG_HIGHBITDEPTH
+  __m128i one = _mm_set1_epi16(1);
+  __m128i coeff_vals_hi = _mm_mulhi_epi16(coeff_vals, one);
+  __m128i coeff_vals_lo = _mm_mullo_epi16(coeff_vals, one);
+  __m128i coeff_vals_1 = _mm_unpacklo_epi16(coeff_vals_lo, coeff_vals_hi);
+  __m128i coeff_vals_2 = _mm_unpackhi_epi16(coeff_vals_lo, coeff_vals_hi);
+  _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals_1);
+  _mm_store_si128((__m128i *)(coeff_ptr + 4), coeff_vals_2);
+#else
+  _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals);
+#endif
+}
+
+void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                         int skip_block, const int16_t *zbin_ptr,
+                         const int16_t *round_ptr, const int16_t *quant_ptr,
+                         const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+                         tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                         uint16_t *eob_ptr, const int16_t *scan_ptr,
+                         const int16_t *iscan_ptr) {
+  __m128i zero;
+  (void)scan_ptr;
+
+  coeff_ptr += n_coeffs;
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+  if (!skip_block) {
+    __m128i eob;
+    __m128i zbin;
+    __m128i round, quant, dequant, shift;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        __m128i pw_1;
+        zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+        round = _mm_load_si128((const __m128i *)round_ptr);
+        quant = _mm_load_si128((const __m128i *)quant_ptr);
+        pw_1 = _mm_set1_epi16(1);
+        zbin = _mm_sub_epi16(zbin, pw_1);
+        dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+        shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        __m128i cmp_mask0, cmp_mask1;
+        // Do DC and first 15 AC
+        coeff0 = load_coefficients(coeff_ptr + n_coeffs);
+        coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+        zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+        cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+        qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
+        qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
+        qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
+        shift = _mm_unpackhi_epi64(shift, shift);
+        qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        // Mask out zbin threshold coeffs
+        qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+        qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+        store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs);
+        store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        store_coefficients(coeff0, dqcoeff_ptr + n_coeffs);
+        store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // AC only loop
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        __m128i cmp_mask0, cmp_mask1;
+
+        coeff0 = load_coefficients(coeff_ptr + n_coeffs);
+        coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+        cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+        qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
+        qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
+        qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
+        qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        // Mask out zbin threshold coeffs
+        qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+        qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+        store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs);
+        store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        store_coefficients(coeff0, dqcoeff_ptr + n_coeffs);
+        store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      store_coefficients(zero, dqcoeff_ptr + n_coeffs);
+      store_coefficients(zero, dqcoeff_ptr + n_coeffs + 8);
+      store_coefficients(zero, qcoeff_ptr + n_coeffs);
+      store_coefficients(zero, qcoeff_ptr + n_coeffs + 8);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm b/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm
new file mode 100644
index 0000000000..36b4dddbdf
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm
@@ -0,0 +1,349 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_1: times 8 dw 1
+
+SECTION .text
+
+; TODO(yunqingwang)fix quantize_b code for skip=1 case.
+%macro QUANTIZE_FN 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, \
+                                eob, scan, iscan
+  cmp                    dword skipm, 0
+  jne .blank
+
+  ; actual quantize loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, dequantmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  mova                            m0, [zbinq]              ; m0 = zbin
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+%ifidn %1, b_32x32
+  pcmpeqw                         m5, m5
+  psrlw                           m5, 15
+  paddw                           m0, m5
+  paddw                           m1, m5
+  psrlw                           m0, 1                    ; m0 = (m0 + 1) / 2
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  mova                            m3, [r2q]                ; m3 = dequant
+  psubw                           m0, [pw_1]
+  mov                             r2, shiftmp
+  mov                             r3, qcoeffmp
+  mova                            m4, [r2]                 ; m4 = shift
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+%ifidn %1, b_32x32
+  psllw                           m4, 1
+%endif
+  pxor                            m5, m5                   ; m5 = dedicated zero
+  DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
+%if CONFIG_HIGHBITDEPTH
+  lea                         coeffq, [  coeffq+ncoeffq*4]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*4]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*4]
+%else
+  lea                         coeffq, [  coeffq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+%endif
+  lea                         iscanq, [  iscanq+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+%if CONFIG_HIGHBITDEPTH
+  ; coeff stored as 32bit numbers & require 16bit numbers
+  mova                            m9, [  coeffq+ncoeffq*4+ 0]
+  packssdw                        m9, [  coeffq+ncoeffq*4+16]
+  mova                           m10, [  coeffq+ncoeffq*4+32]
+  packssdw                       m10, [  coeffq+ncoeffq*4+48]
+%else
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+%endif
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  punpckhqdq                      m0, m0
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                           m8, m6                   ; m8 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                          m8, m4                   ; m8 = m8*qsh>>16
+  punpckhqdq                      m4, m4
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                            m8, m7
+  pand                           m13, m12
+%if CONFIG_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  mova                           m11, m8
+  mova                            m6, m8
+  pcmpgtw                         m5, m8
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+ 0], m11
+  mova        [qcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+32], m11
+  mova        [qcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5             ; reset m5 to zero register
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], m8
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%endif
+%ifidn %1, b_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; dqc[i] = qc[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+%endif
+%if CONFIG_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  mova                            m11, m8
+  mova                            m6, m8
+  pcmpgtw                         m5, m8
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+ 0], m11
+  mova       [dqcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+32], m11
+  mova       [dqcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5             ; reset m5 to zero register
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], m8
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+%endif
+  pcmpeqw                         m8, m5                   ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m12                  ; m11 = scan[i] + 1
+  pandn                           m8, m6                   ; m8 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jz .accumulate_eob
+
+.ac_only_loop:
+%if CONFIG_HIGHBITDEPTH
+  ; pack coeff from 32bit to 16bit array
+  mova                            m9, [  coeffq+ncoeffq*4+ 0]
+  packssdw                        m9, [  coeffq+ncoeffq*4+16]
+  mova                           m10, [  coeffq+ncoeffq*4+32]
+  packssdw                       m10, [  coeffq+ncoeffq*4+48]
+%else
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+%endif
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+%ifidn %1, b_32x32
+  pmovmskb                       r6d, m7
+  pmovmskb                       r2d, m12
+  or                              r6, r2
+  jz .skip_iter
+%endif
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                          m14, m6                   ; m14 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                         m14, m4                   ; m14 = m14*qsh>>16
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                           m14, m7
+  pand                           m13, m12
+%if CONFIG_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pxor                           m11, m11
+  mova                           m11, m14
+  mova                            m6, m14
+  pcmpgtw                         m5, m14
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+ 0], m11
+  mova        [qcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+32], m11
+  mova        [qcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5             ; reset m5 to zero register
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], m14
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%endif
+%ifidn %1, b_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; dqc[i] = qc[i] * q
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+%if CONFIG_HIGHBITDEPTH
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  mova                           m11, m14
+  mova                            m6, m14
+  pcmpgtw                         m5, m14
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+ 0], m11
+  mova       [dqcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+32], m11
+  mova       [dqcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], m14
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+%endif
+  pcmpeqw                        m14, m5                   ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m12                  ; m11 = scan[i] + 1
+  pandn                          m14, m6                   ; m14 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+%ifidn %1, b_32x32
+  jmp .accumulate_eob
+.skip_iter:
+%if CONFIG_HIGHBITDEPTH
+  mova        [qcoeffq+ncoeffq*4+ 0], m5
+  mova        [qcoeffq+ncoeffq*4+16], m5
+  mova        [qcoeffq+ncoeffq*4+32], m5
+  mova        [qcoeffq+ncoeffq*4+48], m5
+  mova       [dqcoeffq+ncoeffq*4+ 0], m5
+  mova       [dqcoeffq+ncoeffq*4+16], m5
+  mova       [dqcoeffq+ncoeffq*4+32], m5
+  mova       [dqcoeffq+ncoeffq*4+48], m5
+%else
+  mova        [qcoeffq+ncoeffq*2+ 0], m5
+  mova        [qcoeffq+ncoeffq*2+16], m5
+  mova       [dqcoeffq+ncoeffq*2+ 0], m5
+  mova       [dqcoeffq+ncoeffq*2+16], m5
+%endif
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+%endif
+
+.accumulate_eob:
+  ; horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  pextrw                          r6, m8, 0
+  mov                             [r2], r6
+  RET
+
+  ; skip-block, i.e. just write all zeroes
+.blank:
+  mov                             r0, dqcoeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, qcoeffmp
+  mov                             r3, eobmp
+  DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
+%if CONFIG_HIGHBITDEPTH
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*4]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*4]
+%else
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+%endif
+  neg                        ncoeffq
+  pxor                            m7, m7
+.blank_loop:
+%if CONFIG_HIGHBITDEPTH
+  mova       [dqcoeffq+ncoeffq*4+ 0], m7
+  mova       [dqcoeffq+ncoeffq*4+16], m7
+  mova       [dqcoeffq+ncoeffq*4+32], m7
+  mova       [dqcoeffq+ncoeffq*4+48], m7
+  mova        [qcoeffq+ncoeffq*4+ 0], m7
+  mova        [qcoeffq+ncoeffq*4+16], m7
+  mova        [qcoeffq+ncoeffq*4+32], m7
+  mova        [qcoeffq+ncoeffq*4+48], m7
+%else
+  mova       [dqcoeffq+ncoeffq*2+ 0], m7
+  mova       [dqcoeffq+ncoeffq*2+16], m7
+  mova        [qcoeffq+ncoeffq*2+ 0], m7
+  mova        [qcoeffq+ncoeffq*2+16], m7
+%endif
+  add                        ncoeffq, mmsize
+  jl .blank_loop
+  mov                    word [eobq], 0
+  RET
+%endmacro
+
+INIT_XMM ssse3
+QUANTIZE_FN b, 7
+QUANTIZE_FN b_32x32, 7
diff --git a/third_party/aom/aom_dsp/x86/sad4d_avx2.c b/third_party/aom/aom_dsp/x86/sad4d_avx2.c
new file mode 100644
index 0000000000..e60f518b4c
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad4d_avx2.c
@@ -0,0 +1,216 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <immintrin.h>  // AVX2
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+void aom_sad32x32x4d_avx2(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t res[4]) {
+  __m256i src_reg, ref0_reg, ref1_reg, ref2_reg, ref3_reg;
+  __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
+  __m256i sum_mlow, sum_mhigh;
+  int i;
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+  sum_ref0 = _mm256_set1_epi16(0);
+  sum_ref1 = _mm256_set1_epi16(0);
+  sum_ref2 = _mm256_set1_epi16(0);
+  sum_ref3 = _mm256_set1_epi16(0);
+  for (i = 0; i < 32; i++) {
+    // load src and all refs
+    src_reg = _mm256_loadu_si256((const __m256i *)src);
+    ref0_reg = _mm256_loadu_si256((const __m256i *)ref0);
+    ref1_reg = _mm256_loadu_si256((const __m256i *)ref1);
+    ref2_reg = _mm256_loadu_si256((const __m256i *)ref2);
+    ref3_reg = _mm256_loadu_si256((const __m256i *)ref3);
+    // sum of the absolute differences between every ref-i to src
+    ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
+    ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
+    ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
+    ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
+    // sum every ref-i
+    sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
+    sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
+    sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
+    sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
+
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+  {
+    __m128i sum;
+    // in sum_ref-i the result is saved in the first 4 bytes
+    // the other 4 bytes are zeroed.
+    // sum_ref1 and sum_ref3 are shifted left by 4 bytes
+    sum_ref1 = _mm256_slli_si256(sum_ref1, 4);
+    sum_ref3 = _mm256_slli_si256(sum_ref3, 4);
+
+    // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
+    sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1);
+    sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3);
+
+    // merge every 64 bit from each sum_ref-i
+    sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2);
+    sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2);
+
+    // add the low 64 bit to the high 64 bit
+    sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh);
+
+    // add the low 128 bit to the high 128 bit
+    sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow),
+                        _mm256_extractf128_si256(sum_mlow, 1));
+
+    _mm_storeu_si128((__m128i *)(res), sum);
+  }
+  _mm256_zeroupper();
+}
+
+void aom_sad64x64x4d_avx2(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t res[4]) {
+  __m256i src_reg, srcnext_reg, ref0_reg, ref0next_reg;
+  __m256i ref1_reg, ref1next_reg, ref2_reg, ref2next_reg;
+  __m256i ref3_reg, ref3next_reg;
+  __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
+  __m256i sum_mlow, sum_mhigh;
+  int i;
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+  sum_ref0 = _mm256_set1_epi16(0);
+  sum_ref1 = _mm256_set1_epi16(0);
+  sum_ref2 = _mm256_set1_epi16(0);
+  sum_ref3 = _mm256_set1_epi16(0);
+  for (i = 0; i < 64; i++) {
+    // load 64 bytes from src and all refs
+    src_reg = _mm256_loadu_si256((const __m256i *)src);
+    srcnext_reg = _mm256_loadu_si256((const __m256i *)(src + 32));
+    ref0_reg = _mm256_loadu_si256((const __m256i *)ref0);
+    ref0next_reg = _mm256_loadu_si256((const __m256i *)(ref0 + 32));
+    ref1_reg = _mm256_loadu_si256((const __m256i *)ref1);
+    ref1next_reg = _mm256_loadu_si256((const __m256i *)(ref1 + 32));
+    ref2_reg = _mm256_loadu_si256((const __m256i *)ref2);
+    ref2next_reg = _mm256_loadu_si256((const __m256i *)(ref2 + 32));
+    ref3_reg = _mm256_loadu_si256((const __m256i *)ref3);
+    ref3next_reg = _mm256_loadu_si256((const __m256i *)(ref3 + 32));
+    // sum of the absolute differences between every ref-i to src
+    ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
+    ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
+    ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
+    ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
+    ref0next_reg = _mm256_sad_epu8(ref0next_reg, srcnext_reg);
+    ref1next_reg = _mm256_sad_epu8(ref1next_reg, srcnext_reg);
+    ref2next_reg = _mm256_sad_epu8(ref2next_reg, srcnext_reg);
+    ref3next_reg = _mm256_sad_epu8(ref3next_reg, srcnext_reg);
+
+    // sum every ref-i
+    sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
+    sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
+    sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
+    sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
+    sum_ref0 = _mm256_add_epi32(sum_ref0, ref0next_reg);
+    sum_ref1 = _mm256_add_epi32(sum_ref1, ref1next_reg);
+    sum_ref2 = _mm256_add_epi32(sum_ref2, ref2next_reg);
+    sum_ref3 = _mm256_add_epi32(sum_ref3, ref3next_reg);
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+  {
+    __m128i sum;
+
+    // in sum_ref-i the result is saved in the first 4 bytes
+    // the other 4 bytes are zeroed.
+    // sum_ref1 and sum_ref3 are shifted left by 4 bytes
+    sum_ref1 = _mm256_slli_si256(sum_ref1, 4);
+    sum_ref3 = _mm256_slli_si256(sum_ref3, 4);
+
+    // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
+    sum_ref0 = _mm256_or_si256(sum_ref0, sum_ref1);
+    sum_ref2 = _mm256_or_si256(sum_ref2, sum_ref3);
+
+    // merge every 64 bit from each sum_ref-i
+    sum_mlow = _mm256_unpacklo_epi64(sum_ref0, sum_ref2);
+    sum_mhigh = _mm256_unpackhi_epi64(sum_ref0, sum_ref2);
+
+    // add the low 64 bit to the high 64 bit
+    sum_mlow = _mm256_add_epi32(sum_mlow, sum_mhigh);
+
+    // add the low 128 bit to the high 128 bit
+    sum = _mm_add_epi32(_mm256_castsi256_si128(sum_mlow),
+                        _mm256_extractf128_si256(sum_mlow, 1));
+
+    _mm_storeu_si128((__m128i *)(res), sum);
+  }
+  _mm256_zeroupper();
+}
+
+void aom_sad32x64x4d_avx2(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t res[4]) {
+  const uint8_t *rf[4];
+  uint32_t sum0[4];
+  uint32_t sum1[4];
+
+  rf[0] = ref[0];
+  rf[1] = ref[1];
+  rf[2] = ref[2];
+  rf[3] = ref[3];
+  aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum0);
+  src += src_stride << 5;
+  rf[0] += ref_stride << 5;
+  rf[1] += ref_stride << 5;
+  rf[2] += ref_stride << 5;
+  rf[3] += ref_stride << 5;
+  aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum1);
+  res[0] = sum0[0] + sum1[0];
+  res[1] = sum0[1] + sum1[1];
+  res[2] = sum0[2] + sum1[2];
+  res[3] = sum0[3] + sum1[3];
+}
+
+void aom_sad64x32x4d_avx2(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t res[4]) {
+  const uint8_t *rf[4];
+  uint32_t sum0[4];
+  uint32_t sum1[4];
+  unsigned int half_width = 32;
+
+  rf[0] = ref[0];
+  rf[1] = ref[1];
+  rf[2] = ref[2];
+  rf[3] = ref[3];
+  aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum0);
+  src += half_width;
+  rf[0] += half_width;
+  rf[1] += half_width;
+  rf[2] += half_width;
+  rf[3] += half_width;
+  aom_sad32x32x4d_avx2(src, src_stride, rf, ref_stride, sum1);
+  res[0] = sum0[0] + sum1[0];
+  res[1] = sum0[1] + sum1[1];
+  res[2] = sum0[2] + sum1[2];
+  res[3] = sum0[3] + sum1[3];
+}
diff --git a/third_party/aom/aom_dsp/x86/sad4d_sse2.asm b/third_party/aom/aom_dsp/x86/sad4d_sse2.asm
new file mode 100644
index 0000000000..8f04ef2f3c
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad4d_sse2.asm
@@ -0,0 +1,253 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; PROCESS_4x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_4x2x4 5-6 0
+  movd                  m0, [srcq +%2]
+%if %1 == 1
+  movd                  m6, [ref1q+%3]
+  movd                  m4, [ref2q+%3]
+  movd                  m7, [ref3q+%3]
+  movd                  m5, [ref4q+%3]
+  movd                  m1, [srcq +%4]
+  movd                  m2, [ref1q+%5]
+  punpckldq             m0, m1
+  punpckldq             m6, m2
+  movd                  m1, [ref2q+%5]
+  movd                  m2, [ref3q+%5]
+  movd                  m3, [ref4q+%5]
+  punpckldq             m4, m1
+  punpckldq             m7, m2
+  punpckldq             m5, m3
+  movlhps               m0, m0
+  movlhps               m6, m4
+  movlhps               m7, m5
+  psadbw                m6, m0
+  psadbw                m7, m0
+%else
+  movd                  m1, [ref1q+%3]
+  movd                  m5, [ref1q+%5]
+  movd                  m2, [ref2q+%3]
+  movd                  m4, [ref2q+%5]
+  punpckldq             m1, m5
+  punpckldq             m2, m4
+  movd                  m3, [ref3q+%3]
+  movd                  m5, [ref3q+%5]
+  punpckldq             m3, m5
+  movd                  m4, [ref4q+%3]
+  movd                  m5, [ref4q+%5]
+  punpckldq             m4, m5
+  movd                  m5, [srcq +%4]
+  punpckldq             m0, m5
+  movlhps               m0, m0
+  movlhps               m1, m2
+  movlhps               m3, m4
+  psadbw                m1, m0
+  psadbw                m3, m0
+  paddd                 m6, m1
+  paddd                 m7, m3
+%endif
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*2]
+  lea                ref1q, [ref1q+ref_strideq*2]
+  lea                ref2q, [ref2q+ref_strideq*2]
+  lea                ref3q, [ref3q+ref_strideq*2]
+  lea                ref4q, [ref4q+ref_strideq*2]
+%endif
+%endmacro
+
+; PROCESS_8x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_8x2x4 5-6 0
+  movh                  m0, [srcq +%2]
+%if %1 == 1
+  movh                  m4, [ref1q+%3]
+  movh                  m5, [ref2q+%3]
+  movh                  m6, [ref3q+%3]
+  movh                  m7, [ref4q+%3]
+  movhps                m0, [srcq +%4]
+  movhps                m4, [ref1q+%5]
+  movhps                m5, [ref2q+%5]
+  movhps                m6, [ref3q+%5]
+  movhps                m7, [ref4q+%5]
+  psadbw                m4, m0
+  psadbw                m5, m0
+  psadbw                m6, m0
+  psadbw                m7, m0
+%else
+  movh                  m1, [ref1q+%3]
+  movh                  m2, [ref2q+%3]
+  movh                  m3, [ref3q+%3]
+  movhps                m0, [srcq +%4]
+  movhps                m1, [ref1q+%5]
+  movhps                m2, [ref2q+%5]
+  movhps                m3, [ref3q+%5]
+  psadbw                m1, m0
+  psadbw                m2, m0
+  psadbw                m3, m0
+  paddd                 m4, m1
+  movh                  m1, [ref4q+%3]
+  movhps                m1, [ref4q+%5]
+  paddd                 m5, m2
+  paddd                 m6, m3
+  psadbw                m1, m0
+  paddd                 m7, m1
+%endif
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*2]
+  lea                ref1q, [ref1q+ref_strideq*2]
+  lea                ref2q, [ref2q+ref_strideq*2]
+  lea                ref3q, [ref3q+ref_strideq*2]
+  lea                ref4q, [ref4q+ref_strideq*2]
+%endif
+%endmacro
+
+; PROCESS_16x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_16x2x4 5-6 0
+  ; 1st 16 px
+  mova                  m0, [srcq +%2]
+%if %1 == 1
+  movu                  m4, [ref1q+%3]
+  movu                  m5, [ref2q+%3]
+  movu                  m6, [ref3q+%3]
+  movu                  m7, [ref4q+%3]
+  psadbw                m4, m0
+  psadbw                m5, m0
+  psadbw                m6, m0
+  psadbw                m7, m0
+%else
+  movu                  m1, [ref1q+%3]
+  movu                  m2, [ref2q+%3]
+  movu                  m3, [ref3q+%3]
+  psadbw                m1, m0
+  psadbw                m2, m0
+  psadbw                m3, m0
+  paddd                 m4, m1
+  movu                  m1, [ref4q+%3]
+  paddd                 m5, m2
+  paddd                 m6, m3
+  psadbw                m1, m0
+  paddd                 m7, m1
+%endif
+
+  ; 2nd 16 px
+  mova                  m0, [srcq +%4]
+  movu                  m1, [ref1q+%5]
+  movu                  m2, [ref2q+%5]
+  movu                  m3, [ref3q+%5]
+  psadbw                m1, m0
+  psadbw                m2, m0
+  psadbw                m3, m0
+  paddd                 m4, m1
+  movu                  m1, [ref4q+%5]
+  paddd                 m5, m2
+  paddd                 m6, m3
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*2]
+  lea                ref1q, [ref1q+ref_strideq*2]
+  lea                ref2q, [ref2q+ref_strideq*2]
+  lea                ref3q, [ref3q+ref_strideq*2]
+  lea                ref4q, [ref4q+ref_strideq*2]
+%endif
+  psadbw                m1, m0
+  paddd                 m7, m1
+%endmacro
+
+; PROCESS_32x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_32x2x4 5-6 0
+  PROCESS_16x2x4 %1, %2, %3, %2 + 16, %3 + 16
+  PROCESS_16x2x4  0, %4, %5, %4 + 16, %5 + 16, %6
+%endmacro
+
+; PROCESS_64x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_64x2x4 5-6 0
+  PROCESS_32x2x4 %1, %2, %3, %2 + 32, %3 + 32
+  PROCESS_32x2x4  0, %4, %5, %4 + 32, %5 + 32, %6
+%endmacro
+
+; PROCESS_128x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro PROCESS_128x2x4 5-6 0
+  PROCESS_64x2x4 %1, %2, %3, %2 + 64, %3 + 64
+  PROCESS_64x2x4  0, %4, %5, %4 + 64, %5 + 64, %6
+%endmacro
+
+; void aom_sadNxNx4d_sse2(uint8_t *src,    int src_stride,
+;                         uint8_t *ref[4], int ref_stride,
+;                         uint32_t res[4]);
+; where NxN = 64x64, 32x32, 16x16, 16x8, 8x16, 8x8, 8x4, 4x8 and 4x4
+%macro SADNXN4D 2
+%if UNIX64
+cglobal sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \
+                              res, ref2, ref3, ref4
+%else
+cglobal sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \
+                              ref2, ref3, ref4
+%endif
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+  mov                ref2q, [ref1q+gprsize*1]
+  mov                ref3q, [ref1q+gprsize*2]
+  mov                ref4q, [ref1q+gprsize*3]
+  mov                ref1q, [ref1q+gprsize*0]
+
+  PROCESS_%1x2x4 1, 0, 0, src_strideq, ref_strideq, 1
+%rep (%2-4)/2
+  PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 1
+%endrep
+  PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 0
+
+%if %1 > 4
+  pslldq                m5, 4
+  pslldq                m7, 4
+  por                   m4, m5
+  por                   m6, m7
+  mova                  m5, m4
+  mova                  m7, m6
+  punpcklqdq            m4, m6
+  punpckhqdq            m5, m7
+  movifnidn             r4, r4mp
+  paddd                 m4, m5
+  movu                [r4], m4
+  RET
+%else
+  movifnidn             r4, r4mp
+  pshufd            m6, m6, 0x08
+  pshufd            m7, m7, 0x08
+  movq              [r4+0], m6
+  movq              [r4+8], m7
+  RET
+%endif
+%endmacro
+
+INIT_XMM sse2
+%if CONFIG_EXT_PARTITION
+SADNXN4D 128, 128
+SADNXN4D 128, 64
+SADNXN4D 64,  128
+%endif
+SADNXN4D 64, 64
+SADNXN4D 64, 32
+SADNXN4D 32, 64
+SADNXN4D 32, 32
+SADNXN4D 32, 16
+SADNXN4D 16, 32
+SADNXN4D 16, 16
+SADNXN4D 16,  8
+SADNXN4D  8, 16
+SADNXN4D  8,  8
+SADNXN4D  8,  4
+SADNXN4D  4,  8
+SADNXN4D  4,  4
diff --git a/third_party/aom/aom_dsp/x86/sad_avx2.c b/third_party/aom/aom_dsp/x86/sad_avx2.c
new file mode 100644
index 0000000000..efba612896
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad_avx2.c
@@ -0,0 +1,187 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <immintrin.h>
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/mem.h"
+
+#define FSAD64_H(h)                                                           \
+  unsigned int aom_sad64x##h##_avx2(const uint8_t *src_ptr, int src_stride,   \
+                                    const uint8_t *ref_ptr, int ref_stride) { \
+    int i, res;                                                               \
+    __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;                           \
+    __m256i sum_sad = _mm256_setzero_si256();                                 \
+    __m256i sum_sad_h;                                                        \
+    __m128i sum_sad128;                                                       \
+    for (i = 0; i < h; i++) {                                                 \
+      ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);                \
+      ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32));         \
+      sad1_reg = _mm256_sad_epu8(                                             \
+          ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));            \
+      sad2_reg = _mm256_sad_epu8(                                             \
+          ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32)));     \
+      sum_sad =                                                               \
+          _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));    \
+      ref_ptr += ref_stride;                                                  \
+      src_ptr += src_stride;                                                  \
+    }                                                                         \
+    sum_sad_h = _mm256_srli_si256(sum_sad, 8);                                \
+    sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);                           \
+    sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);                        \
+    sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);  \
+    res = _mm_cvtsi128_si32(sum_sad128);                                      \
+    _mm256_zeroupper();                                                       \
+    return res;                                                               \
+  }
+
+#define FSAD32_H(h)                                                           \
+  unsigned int aom_sad32x##h##_avx2(const uint8_t *src_ptr, int src_stride,   \
+                                    const uint8_t *ref_ptr, int ref_stride) { \
+    int i, res;                                                               \
+    __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;                           \
+    __m256i sum_sad = _mm256_setzero_si256();                                 \
+    __m256i sum_sad_h;                                                        \
+    __m128i sum_sad128;                                                       \
+    int ref2_stride = ref_stride << 1;                                        \
+    int src2_stride = src_stride << 1;                                        \
+    int max = h >> 1;                                                         \
+    for (i = 0; i < max; i++) {                                               \
+      ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);                \
+      ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \
+      sad1_reg = _mm256_sad_epu8(                                             \
+          ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));            \
+      sad2_reg = _mm256_sad_epu8(                                             \
+          ref2_reg,                                                           \
+          _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride)));       \
+      sum_sad =                                                               \
+          _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));    \
+      ref_ptr += ref2_stride;                                                 \
+      src_ptr += src2_stride;                                                 \
+    }                                                                         \
+    sum_sad_h = _mm256_srli_si256(sum_sad, 8);                                \
+    sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);                           \
+    sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);                        \
+    sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);  \
+    res = _mm_cvtsi128_si32(sum_sad128);                                      \
+    _mm256_zeroupper();                                                       \
+    return res;                                                               \
+  }
+
+#define FSAD64  \
+  FSAD64_H(64); \
+  FSAD64_H(32);
+
+#define FSAD32  \
+  FSAD32_H(64); \
+  FSAD32_H(32); \
+  FSAD32_H(16);
+
+/* clang-format off */
+FSAD64
+FSAD32
+/* clang-format on */
+
+#undef FSAD64
+#undef FSAD32
+#undef FSAD64_H
+#undef FSAD32_H
+
+#define FSADAVG64_H(h)                                                        \
+  unsigned int aom_sad64x##h##_avg_avx2(                                      \
+      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr,         \
+      int ref_stride, const uint8_t *second_pred) {                           \
+    int i, res;                                                               \
+    __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;                           \
+    __m256i sum_sad = _mm256_setzero_si256();                                 \
+    __m256i sum_sad_h;                                                        \
+    __m128i sum_sad128;                                                       \
+    for (i = 0; i < h; i++) {                                                 \
+      ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);                \
+      ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32));         \
+      ref1_reg = _mm256_avg_epu8(                                             \
+          ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred));        \
+      ref2_reg = _mm256_avg_epu8(                                             \
+          ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32))); \
+      sad1_reg = _mm256_sad_epu8(                                             \
+          ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));            \
+      sad2_reg = _mm256_sad_epu8(                                             \
+          ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32)));     \
+      sum_sad =                                                               \
+          _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));    \
+      ref_ptr += ref_stride;                                                  \
+      src_ptr += src_stride;                                                  \
+      second_pred += 64;                                                      \
+    }                                                                         \
+    sum_sad_h = _mm256_srli_si256(sum_sad, 8);                                \
+    sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);                           \
+    sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);                        \
+    sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);  \
+    res = _mm_cvtsi128_si32(sum_sad128);                                      \
+    _mm256_zeroupper();                                                       \
+    return res;                                                               \
+  }
+
+#define FSADAVG32_H(h)                                                        \
+  unsigned int aom_sad32x##h##_avg_avx2(                                      \
+      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr,         \
+      int ref_stride, const uint8_t *second_pred) {                           \
+    int i, res;                                                               \
+    __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;                           \
+    __m256i sum_sad = _mm256_setzero_si256();                                 \
+    __m256i sum_sad_h;                                                        \
+    __m128i sum_sad128;                                                       \
+    int ref2_stride = ref_stride << 1;                                        \
+    int src2_stride = src_stride << 1;                                        \
+    int max = h >> 1;                                                         \
+    for (i = 0; i < max; i++) {                                               \
+      ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);                \
+      ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \
+      ref1_reg = _mm256_avg_epu8(                                             \
+          ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred));        \
+      ref2_reg = _mm256_avg_epu8(                                             \
+          ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32))); \
+      sad1_reg = _mm256_sad_epu8(                                             \
+          ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));            \
+      sad2_reg = _mm256_sad_epu8(                                             \
+          ref2_reg,                                                           \
+          _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride)));       \
+      sum_sad =                                                               \
+          _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));    \
+      ref_ptr += ref2_stride;                                                 \
+      src_ptr += src2_stride;                                                 \
+      second_pred += 64;                                                      \
+    }                                                                         \
+    sum_sad_h = _mm256_srli_si256(sum_sad, 8);                                \
+    sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);                           \
+    sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);                        \
+    sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);  \
+    res = _mm_cvtsi128_si32(sum_sad128);                                      \
+    _mm256_zeroupper();                                                       \
+    return res;                                                               \
+  }
+
+#define FSADAVG64  \
+  FSADAVG64_H(64); \
+  FSADAVG64_H(32);
+
+#define FSADAVG32  \
+  FSADAVG32_H(64); \
+  FSADAVG32_H(32); \
+  FSADAVG32_H(16);
+
+/* clang-format off */
+FSADAVG64
+FSADAVG32
+/* clang-format on */
+
+#undef FSADAVG64
+#undef FSADAVG32
+#undef FSADAVG64_H
+#undef FSADAVG32_H
diff --git a/third_party/aom/aom_dsp/x86/sad_highbd_avx2.c b/third_party/aom/aom_dsp/x86/sad_highbd_avx2.c
new file mode 100644
index 0000000000..196394379a
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad_highbd_avx2.c
@@ -0,0 +1,1043 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+// SAD
+static INLINE unsigned int get_sad_from_mm256_epi32(const __m256i *v) {
+  // input 8 32-bit summation
+  __m128i lo128, hi128;
+  __m256i u = _mm256_srli_si256(*v, 8);
+  u = _mm256_add_epi32(u, *v);
+
+  // 4 32-bit summation
+  hi128 = _mm256_extracti128_si256(u, 1);
+  lo128 = _mm256_castsi256_si128(u);
+  lo128 = _mm_add_epi32(hi128, lo128);
+
+  // 2 32-bit summation
+  hi128 = _mm_srli_si128(lo128, 4);
+  lo128 = _mm_add_epi32(lo128, hi128);
+
+  return (unsigned int)_mm_cvtsi128_si32(lo128);
+}
+
+unsigned int aom_highbd_sad16x8_avx2(const uint8_t *src, int src_stride,
+                                     const uint8_t *ref, int ref_stride) {
+  const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src);
+  const uint16_t *ref_ptr = CONVERT_TO_SHORTPTR(ref);
+
+  // first 4 rows
+  __m256i s0 = _mm256_loadu_si256((const __m256i *)src_ptr);
+  __m256i s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride));
+  __m256i s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride));
+  __m256i s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride));
+
+  __m256i r0 = _mm256_loadu_si256((const __m256i *)ref_ptr);
+  __m256i r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride));
+  __m256i r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride));
+  __m256i r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride));
+
+  __m256i u0 = _mm256_sub_epi16(s0, r0);
+  __m256i u1 = _mm256_sub_epi16(s1, r1);
+  __m256i u2 = _mm256_sub_epi16(s2, r2);
+  __m256i u3 = _mm256_sub_epi16(s3, r3);
+  __m256i zero = _mm256_setzero_si256();
+  __m256i sum0, sum1;
+
+  u0 = _mm256_abs_epi16(u0);
+  u1 = _mm256_abs_epi16(u1);
+  u2 = _mm256_abs_epi16(u2);
+  u3 = _mm256_abs_epi16(u3);
+
+  sum0 = _mm256_add_epi16(u0, u1);
+  sum0 = _mm256_add_epi16(sum0, u2);
+  sum0 = _mm256_add_epi16(sum0, u3);
+
+  // second 4 rows
+  src_ptr += src_stride << 2;
+  ref_ptr += ref_stride << 2;
+  s0 = _mm256_loadu_si256((const __m256i *)src_ptr);
+  s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride));
+  s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride));
+  s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride));
+
+  r0 = _mm256_loadu_si256((const __m256i *)ref_ptr);
+  r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride));
+  r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride));
+  r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride));
+
+  u0 = _mm256_sub_epi16(s0, r0);
+  u1 = _mm256_sub_epi16(s1, r1);
+  u2 = _mm256_sub_epi16(s2, r2);
+  u3 = _mm256_sub_epi16(s3, r3);
+
+  u0 = _mm256_abs_epi16(u0);
+  u1 = _mm256_abs_epi16(u1);
+  u2 = _mm256_abs_epi16(u2);
+  u3 = _mm256_abs_epi16(u3);
+
+  sum1 = _mm256_add_epi16(u0, u1);
+  sum1 = _mm256_add_epi16(sum1, u2);
+  sum1 = _mm256_add_epi16(sum1, u3);
+
+  // find out the SAD
+  s0 = _mm256_unpacklo_epi16(sum0, zero);
+  s1 = _mm256_unpackhi_epi16(sum0, zero);
+  r0 = _mm256_unpacklo_epi16(sum1, zero);
+  r1 = _mm256_unpackhi_epi16(sum1, zero);
+  s0 = _mm256_add_epi32(s0, s1);
+  r0 = _mm256_add_epi32(r0, r1);
+  sum0 = _mm256_add_epi32(s0, r0);
+  // 8 32-bit summation
+
+  return (unsigned int)get_sad_from_mm256_epi32(&sum0);
+}
+
+unsigned int aom_highbd_sad16x16_avx2(const uint8_t *src, int src_stride,
+                                      const uint8_t *ref, int ref_stride) {
+  const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src);
+  const uint16_t *ref_ptr = CONVERT_TO_SHORTPTR(ref);
+  __m256i s0, s1, s2, s3, r0, r1, r2, r3, u0, u1, u2, u3;
+  __m256i sum0;
+  __m256i sum = _mm256_setzero_si256();
+  const __m256i zero = _mm256_setzero_si256();
+  int row = 0;
+
+  // Loop for every 4 rows
+  while (row < 16) {
+    s0 = _mm256_loadu_si256((const __m256i *)src_ptr);
+    s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride));
+    s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride));
+    s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride));
+
+    r0 = _mm256_loadu_si256((const __m256i *)ref_ptr);
+    r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride));
+    r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride));
+    r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride));
+
+    u0 = _mm256_sub_epi16(s0, r0);
+    u1 = _mm256_sub_epi16(s1, r1);
+    u2 = _mm256_sub_epi16(s2, r2);
+    u3 = _mm256_sub_epi16(s3, r3);
+
+    u0 = _mm256_abs_epi16(u0);
+    u1 = _mm256_abs_epi16(u1);
+    u2 = _mm256_abs_epi16(u2);
+    u3 = _mm256_abs_epi16(u3);
+
+    sum0 = _mm256_add_epi16(u0, u1);
+    sum0 = _mm256_add_epi16(sum0, u2);
+    sum0 = _mm256_add_epi16(sum0, u3);
+
+    s0 = _mm256_unpacklo_epi16(sum0, zero);
+    s1 = _mm256_unpackhi_epi16(sum0, zero);
+    sum = _mm256_add_epi32(sum, s0);
+    sum = _mm256_add_epi32(sum, s1);
+    // 8 32-bit summation
+
+    row += 4;
+    src_ptr += src_stride << 2;
+    ref_ptr += ref_stride << 2;
+  }
+  return get_sad_from_mm256_epi32(&sum);
+}
+
+static void sad32x4(const uint16_t *src_ptr, int src_stride,
+                    const uint16_t *ref_ptr, int ref_stride,
+                    const uint16_t *sec_ptr, __m256i *sad_acc) {
+  __m256i s0, s1, s2, s3, r0, r1, r2, r3;
+  const __m256i zero = _mm256_setzero_si256();
+  int row_sections = 0;
+
+  while (row_sections < 2) {
+    s0 = _mm256_loadu_si256((const __m256i *)src_ptr);
+    s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + 16));
+    s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride));
+    s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 16));
+
+    r0 = _mm256_loadu_si256((const __m256i *)ref_ptr);
+    r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16));
+    r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride));
+    r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 16));
+
+    if (sec_ptr) {
+      r0 = _mm256_avg_epu16(r0, _mm256_loadu_si256((const __m256i *)sec_ptr));
+      r1 = _mm256_avg_epu16(
+          r1, _mm256_loadu_si256((const __m256i *)(sec_ptr + 16)));
+      r2 = _mm256_avg_epu16(
+          r2, _mm256_loadu_si256((const __m256i *)(sec_ptr + 32)));
+      r3 = _mm256_avg_epu16(
+          r3, _mm256_loadu_si256((const __m256i *)(sec_ptr + 48)));
+    }
+    s0 = _mm256_sub_epi16(s0, r0);
+    s1 = _mm256_sub_epi16(s1, r1);
+    s2 = _mm256_sub_epi16(s2, r2);
+    s3 = _mm256_sub_epi16(s3, r3);
+
+    s0 = _mm256_abs_epi16(s0);
+    s1 = _mm256_abs_epi16(s1);
+    s2 = _mm256_abs_epi16(s2);
+    s3 = _mm256_abs_epi16(s3);
+
+    s0 = _mm256_add_epi16(s0, s1);
+    s0 = _mm256_add_epi16(s0, s2);
+    s0 = _mm256_add_epi16(s0, s3);
+
+    r0 = _mm256_unpacklo_epi16(s0, zero);
+    r1 = _mm256_unpackhi_epi16(s0, zero);
+
+    r0 = _mm256_add_epi32(r0, r1);
+    *sad_acc = _mm256_add_epi32(*sad_acc, r0);
+
+    row_sections += 1;
+    src_ptr += src_stride << 1;
+    ref_ptr += ref_stride << 1;
+    if (sec_ptr) sec_ptr += 32 << 1;
+  }
+}
+
+unsigned int aom_highbd_sad32x16_avx2(const uint8_t *src, int src_stride,
+                                      const uint8_t *ref, int ref_stride) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  const int left_shift = 2;
+  int row_section = 0;
+
+  while (row_section < 4) {
+    sad32x4(srcp, src_stride, refp, ref_stride, NULL, &sad);
+    srcp += src_stride << left_shift;
+    refp += ref_stride << left_shift;
+    row_section += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad16x32_avx2(const uint8_t *src, int src_stride,
+                                      const uint8_t *ref, int ref_stride) {
+  uint32_t sum = aom_highbd_sad16x16_avx2(src, src_stride, ref, ref_stride);
+  src += src_stride << 4;
+  ref += ref_stride << 4;
+  sum += aom_highbd_sad16x16_avx2(src, src_stride, ref, ref_stride);
+  return sum;
+}
+
+unsigned int aom_highbd_sad32x32_avx2(const uint8_t *src, int src_stride,
+                                      const uint8_t *ref, int ref_stride) {
+  uint32_t sum = aom_highbd_sad32x16_avx2(src, src_stride, ref, ref_stride);
+  src += src_stride << 4;
+  ref += ref_stride << 4;
+  sum += aom_highbd_sad32x16_avx2(src, src_stride, ref, ref_stride);
+  return sum;
+}
+
+unsigned int aom_highbd_sad32x64_avx2(const uint8_t *src, int src_stride,
+                                      const uint8_t *ref, int ref_stride) {
+  uint32_t sum = aom_highbd_sad32x32_avx2(src, src_stride, ref, ref_stride);
+  src += src_stride << 5;
+  ref += ref_stride << 5;
+  sum += aom_highbd_sad32x32_avx2(src, src_stride, ref, ref_stride);
+  return sum;
+}
+
+static void sad64x2(const uint16_t *src_ptr, int src_stride,
+                    const uint16_t *ref_ptr, int ref_stride,
+                    const uint16_t *sec_ptr, __m256i *sad_acc) {
+  __m256i s[8], r[8];
+  const __m256i zero = _mm256_setzero_si256();
+
+  s[0] = _mm256_loadu_si256((const __m256i *)src_ptr);
+  s[1] = _mm256_loadu_si256((const __m256i *)(src_ptr + 16));
+  s[2] = _mm256_loadu_si256((const __m256i *)(src_ptr + 32));
+  s[3] = _mm256_loadu_si256((const __m256i *)(src_ptr + 48));
+  s[4] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride));
+  s[5] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 16));
+  s[6] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 32));
+  s[7] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 48));
+
+  r[0] = _mm256_loadu_si256((const __m256i *)ref_ptr);
+  r[1] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16));
+  r[2] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 32));
+  r[3] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 48));
+  r[4] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride));
+  r[5] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 16));
+  r[6] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 32));
+  r[7] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 48));
+
+  if (sec_ptr) {
+    r[0] = _mm256_avg_epu16(r[0], _mm256_loadu_si256((const __m256i *)sec_ptr));
+    r[1] = _mm256_avg_epu16(
+        r[1], _mm256_loadu_si256((const __m256i *)(sec_ptr + 16)));
+    r[2] = _mm256_avg_epu16(
+        r[2], _mm256_loadu_si256((const __m256i *)(sec_ptr + 32)));
+    r[3] = _mm256_avg_epu16(
+        r[3], _mm256_loadu_si256((const __m256i *)(sec_ptr + 48)));
+    r[4] = _mm256_avg_epu16(
+        r[4], _mm256_loadu_si256((const __m256i *)(sec_ptr + 64)));
+    r[5] = _mm256_avg_epu16(
+        r[5], _mm256_loadu_si256((const __m256i *)(sec_ptr + 80)));
+    r[6] = _mm256_avg_epu16(
+        r[6], _mm256_loadu_si256((const __m256i *)(sec_ptr + 96)));
+    r[7] = _mm256_avg_epu16(
+        r[7], _mm256_loadu_si256((const __m256i *)(sec_ptr + 112)));
+  }
+
+  s[0] = _mm256_sub_epi16(s[0], r[0]);
+  s[1] = _mm256_sub_epi16(s[1], r[1]);
+  s[2] = _mm256_sub_epi16(s[2], r[2]);
+  s[3] = _mm256_sub_epi16(s[3], r[3]);
+  s[4] = _mm256_sub_epi16(s[4], r[4]);
+  s[5] = _mm256_sub_epi16(s[5], r[5]);
+  s[6] = _mm256_sub_epi16(s[6], r[6]);
+  s[7] = _mm256_sub_epi16(s[7], r[7]);
+
+  s[0] = _mm256_abs_epi16(s[0]);
+  s[1] = _mm256_abs_epi16(s[1]);
+  s[2] = _mm256_abs_epi16(s[2]);
+  s[3] = _mm256_abs_epi16(s[3]);
+  s[4] = _mm256_abs_epi16(s[4]);
+  s[5] = _mm256_abs_epi16(s[5]);
+  s[6] = _mm256_abs_epi16(s[6]);
+  s[7] = _mm256_abs_epi16(s[7]);
+
+  s[0] = _mm256_add_epi16(s[0], s[1]);
+  s[0] = _mm256_add_epi16(s[0], s[2]);
+  s[0] = _mm256_add_epi16(s[0], s[3]);
+
+  s[4] = _mm256_add_epi16(s[4], s[5]);
+  s[4] = _mm256_add_epi16(s[4], s[6]);
+  s[4] = _mm256_add_epi16(s[4], s[7]);
+
+  r[0] = _mm256_unpacklo_epi16(s[0], zero);
+  r[1] = _mm256_unpackhi_epi16(s[0], zero);
+  r[2] = _mm256_unpacklo_epi16(s[4], zero);
+  r[3] = _mm256_unpackhi_epi16(s[4], zero);
+
+  r[0] = _mm256_add_epi32(r[0], r[1]);
+  r[0] = _mm256_add_epi32(r[0], r[2]);
+  r[0] = _mm256_add_epi32(r[0], r[3]);
+  *sad_acc = _mm256_add_epi32(*sad_acc, r[0]);
+}
+
+unsigned int aom_highbd_sad64x32_avx2(const uint8_t *src, int src_stride,
+                                      const uint8_t *ref, int ref_stride) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  const int left_shift = 1;
+  int row_section = 0;
+
+  while (row_section < 16) {
+    sad64x2(srcp, src_stride, refp, ref_stride, NULL, &sad);
+    srcp += src_stride << left_shift;
+    refp += ref_stride << left_shift;
+    row_section += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad64x64_avx2(const uint8_t *src, int src_stride,
+                                      const uint8_t *ref, int ref_stride) {
+  uint32_t sum = aom_highbd_sad64x32_avx2(src, src_stride, ref, ref_stride);
+  src += src_stride << 5;
+  ref += ref_stride << 5;
+  sum += aom_highbd_sad64x32_avx2(src, src_stride, ref, ref_stride);
+  return sum;
+}
+
+#if CONFIG_EXT_PARTITION
+static void sad128x1(const uint16_t *src_ptr, const uint16_t *ref_ptr,
+                     const uint16_t *sec_ptr, __m256i *sad_acc) {
+  __m256i s[8], r[8];
+  const __m256i zero = _mm256_setzero_si256();
+
+  s[0] = _mm256_loadu_si256((const __m256i *)src_ptr);
+  s[1] = _mm256_loadu_si256((const __m256i *)(src_ptr + 16));
+  s[2] = _mm256_loadu_si256((const __m256i *)(src_ptr + 32));
+  s[3] = _mm256_loadu_si256((const __m256i *)(src_ptr + 48));
+  s[4] = _mm256_loadu_si256((const __m256i *)(src_ptr + 64));
+  s[5] = _mm256_loadu_si256((const __m256i *)(src_ptr + 80));
+  s[6] = _mm256_loadu_si256((const __m256i *)(src_ptr + 96));
+  s[7] = _mm256_loadu_si256((const __m256i *)(src_ptr + 112));
+
+  r[0] = _mm256_loadu_si256((const __m256i *)ref_ptr);
+  r[1] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16));
+  r[2] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 32));
+  r[3] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 48));
+  r[4] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 64));
+  r[5] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 80));
+  r[6] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 96));
+  r[7] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 112));
+
+  if (sec_ptr) {
+    r[0] = _mm256_avg_epu16(r[0], _mm256_loadu_si256((const __m256i *)sec_ptr));
+    r[1] = _mm256_avg_epu16(
+        r[1], _mm256_loadu_si256((const __m256i *)(sec_ptr + 16)));
+    r[2] = _mm256_avg_epu16(
+        r[2], _mm256_loadu_si256((const __m256i *)(sec_ptr + 32)));
+    r[3] = _mm256_avg_epu16(
+        r[3], _mm256_loadu_si256((const __m256i *)(sec_ptr + 48)));
+    r[4] = _mm256_avg_epu16(
+        r[4], _mm256_loadu_si256((const __m256i *)(sec_ptr + 64)));
+    r[5] = _mm256_avg_epu16(
+        r[5], _mm256_loadu_si256((const __m256i *)(sec_ptr + 80)));
+    r[6] = _mm256_avg_epu16(
+        r[6], _mm256_loadu_si256((const __m256i *)(sec_ptr + 96)));
+    r[7] = _mm256_avg_epu16(
+        r[7], _mm256_loadu_si256((const __m256i *)(sec_ptr + 112)));
+  }
+
+  s[0] = _mm256_sub_epi16(s[0], r[0]);
+  s[1] = _mm256_sub_epi16(s[1], r[1]);
+  s[2] = _mm256_sub_epi16(s[2], r[2]);
+  s[3] = _mm256_sub_epi16(s[3], r[3]);
+  s[4] = _mm256_sub_epi16(s[4], r[4]);
+  s[5] = _mm256_sub_epi16(s[5], r[5]);
+  s[6] = _mm256_sub_epi16(s[6], r[6]);
+  s[7] = _mm256_sub_epi16(s[7], r[7]);
+
+  s[0] = _mm256_abs_epi16(s[0]);
+  s[1] = _mm256_abs_epi16(s[1]);
+  s[2] = _mm256_abs_epi16(s[2]);
+  s[3] = _mm256_abs_epi16(s[3]);
+  s[4] = _mm256_abs_epi16(s[4]);
+  s[5] = _mm256_abs_epi16(s[5]);
+  s[6] = _mm256_abs_epi16(s[6]);
+  s[7] = _mm256_abs_epi16(s[7]);
+
+  s[0] = _mm256_add_epi16(s[0], s[1]);
+  s[0] = _mm256_add_epi16(s[0], s[2]);
+  s[0] = _mm256_add_epi16(s[0], s[3]);
+
+  s[4] = _mm256_add_epi16(s[4], s[5]);
+  s[4] = _mm256_add_epi16(s[4], s[6]);
+  s[4] = _mm256_add_epi16(s[4], s[7]);
+
+  r[0] = _mm256_unpacklo_epi16(s[0], zero);
+  r[1] = _mm256_unpackhi_epi16(s[0], zero);
+  r[2] = _mm256_unpacklo_epi16(s[4], zero);
+  r[3] = _mm256_unpackhi_epi16(s[4], zero);
+
+  r[0] = _mm256_add_epi32(r[0], r[1]);
+  r[0] = _mm256_add_epi32(r[0], r[2]);
+  r[0] = _mm256_add_epi32(r[0], r[3]);
+  *sad_acc = _mm256_add_epi32(*sad_acc, r[0]);
+}
+
+unsigned int aom_highbd_sad128x64_avx2(const uint8_t *src, int src_stride,
+                                       const uint8_t *ref, int ref_stride) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  int row = 0;
+  while (row < 64) {
+    sad128x1(srcp, refp, NULL, &sad);
+    srcp += src_stride;
+    refp += ref_stride;
+    row += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad64x128_avx2(const uint8_t *src, int src_stride,
+                                       const uint8_t *ref, int ref_stride) {
+  uint32_t sum = aom_highbd_sad64x64_avx2(src, src_stride, ref, ref_stride);
+  src += src_stride << 6;
+  ref += ref_stride << 6;
+  sum += aom_highbd_sad64x64_avx2(src, src_stride, ref, ref_stride);
+  return sum;
+}
+
+unsigned int aom_highbd_sad128x128_avx2(const uint8_t *src, int src_stride,
+                                        const uint8_t *ref, int ref_stride) {
+  uint32_t sum = aom_highbd_sad128x64_avx2(src, src_stride, ref, ref_stride);
+  src += src_stride << 6;
+  ref += ref_stride << 6;
+  sum += aom_highbd_sad128x64_avx2(src, src_stride, ref, ref_stride);
+  return sum;
+}
+#endif  // CONFIG_EXT_PARTITION
+
+// If sec_ptr = 0, calculate regular SAD. Otherwise, calculate average SAD.
+static INLINE void sad16x4(const uint16_t *src_ptr, int src_stride,
+                           const uint16_t *ref_ptr, int ref_stride,
+                           const uint16_t *sec_ptr, __m256i *sad_acc) {
+  __m256i s0, s1, s2, s3, r0, r1, r2, r3;
+  const __m256i zero = _mm256_setzero_si256();
+
+  s0 = _mm256_loadu_si256((const __m256i *)src_ptr);
+  s1 = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride));
+  s2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride));
+  s3 = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride));
+
+  r0 = _mm256_loadu_si256((const __m256i *)ref_ptr);
+  r1 = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride));
+  r2 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride));
+  r3 = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride));
+
+  if (sec_ptr) {
+    r0 = _mm256_avg_epu16(r0, _mm256_loadu_si256((const __m256i *)sec_ptr));
+    r1 = _mm256_avg_epu16(r1,
+                          _mm256_loadu_si256((const __m256i *)(sec_ptr + 16)));
+    r2 = _mm256_avg_epu16(r2,
+                          _mm256_loadu_si256((const __m256i *)(sec_ptr + 32)));
+    r3 = _mm256_avg_epu16(r3,
+                          _mm256_loadu_si256((const __m256i *)(sec_ptr + 48)));
+  }
+
+  s0 = _mm256_sub_epi16(s0, r0);
+  s1 = _mm256_sub_epi16(s1, r1);
+  s2 = _mm256_sub_epi16(s2, r2);
+  s3 = _mm256_sub_epi16(s3, r3);
+
+  s0 = _mm256_abs_epi16(s0);
+  s1 = _mm256_abs_epi16(s1);
+  s2 = _mm256_abs_epi16(s2);
+  s3 = _mm256_abs_epi16(s3);
+
+  s0 = _mm256_add_epi16(s0, s1);
+  s0 = _mm256_add_epi16(s0, s2);
+  s0 = _mm256_add_epi16(s0, s3);
+
+  r0 = _mm256_unpacklo_epi16(s0, zero);
+  r1 = _mm256_unpackhi_epi16(s0, zero);
+
+  r0 = _mm256_add_epi32(r0, r1);
+  *sad_acc = _mm256_add_epi32(*sad_acc, r0);
+}
+
+unsigned int aom_highbd_sad16x8_avg_avx2(const uint8_t *src, int src_stride,
+                                         const uint8_t *ref, int ref_stride,
+                                         const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+
+  sad16x4(srcp, src_stride, refp, ref_stride, secp, &sad);
+
+  // Next 4 rows
+  srcp += src_stride << 2;
+  refp += ref_stride << 2;
+  secp += 64;
+  sad16x4(srcp, src_stride, refp, ref_stride, secp, &sad);
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad16x16_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 3;
+  uint32_t sum = aom_highbd_sad16x8_avg_avx2(src, src_stride, ref, ref_stride,
+                                             second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 16 << left_shift;
+  sum += aom_highbd_sad16x8_avg_avx2(src, src_stride, ref, ref_stride,
+                                     second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad16x32_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 4;
+  uint32_t sum = aom_highbd_sad16x16_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 16 << left_shift;
+  sum += aom_highbd_sad16x16_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad32x16_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+  const int left_shift = 2;
+  int row_section = 0;
+
+  while (row_section < 4) {
+    sad32x4(srcp, src_stride, refp, ref_stride, secp, &sad);
+    srcp += src_stride << left_shift;
+    refp += ref_stride << left_shift;
+    secp += 32 << left_shift;
+    row_section += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad32x32_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 4;
+  uint32_t sum = aom_highbd_sad32x16_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 32 << left_shift;
+  sum += aom_highbd_sad32x16_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad32x64_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 5;
+  uint32_t sum = aom_highbd_sad32x32_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 32 << left_shift;
+  sum += aom_highbd_sad32x32_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad64x32_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+  const int left_shift = 1;
+  int row_section = 0;
+
+  while (row_section < 16) {
+    sad64x2(srcp, src_stride, refp, ref_stride, secp, &sad);
+    srcp += src_stride << left_shift;
+    refp += ref_stride << left_shift;
+    secp += 64 << left_shift;
+    row_section += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad64x64_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 5;
+  uint32_t sum = aom_highbd_sad64x32_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 64 << left_shift;
+  sum += aom_highbd_sad64x32_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+#if CONFIG_EXT_PARTITION
+unsigned int aom_highbd_sad64x128_avg_avx2(const uint8_t *src, int src_stride,
+                                           const uint8_t *ref, int ref_stride,
+                                           const uint8_t *second_pred) {
+  const int left_shift = 6;
+  uint32_t sum = aom_highbd_sad64x64_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 64 << left_shift;
+  sum += aom_highbd_sad64x64_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad128x64_avg_avx2(const uint8_t *src, int src_stride,
+                                           const uint8_t *ref, int ref_stride,
+                                           const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+  int row = 0;
+  while (row < 64) {
+    sad128x1(srcp, refp, secp, &sad);
+    srcp += src_stride;
+    refp += ref_stride;
+    secp += 16 << 3;
+    row += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad128x128_avg_avx2(const uint8_t *src, int src_stride,
+                                            const uint8_t *ref, int ref_stride,
+                                            const uint8_t *second_pred) {
+  unsigned int sum;
+  const int left_shift = 6;
+
+  sum = aom_highbd_sad128x64_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 128 << left_shift;
+  sum += aom_highbd_sad128x64_avg_avx2(src, src_stride, ref, ref_stride,
+                                       second_pred);
+  return sum;
+}
+#endif  // CONFIG_EXT_PARTITION
+
+// SAD 4D
+// Combine 4 __m256i vectors to uint32_t result[4]
+static INLINE void get_4d_sad_from_mm256_epi32(const __m256i *v,
+                                               uint32_t *res) {
+  __m256i u0, u1, u2, u3;
+  const __m256i mask = _mm256_set1_epi64x(UINT32_MAX);
+  __m128i sad;
+
+  // 8 32-bit summation
+  u0 = _mm256_srli_si256(v[0], 4);
+  u1 = _mm256_srli_si256(v[1], 4);
+  u2 = _mm256_srli_si256(v[2], 4);
+  u3 = _mm256_srli_si256(v[3], 4);
+
+  u0 = _mm256_add_epi32(u0, v[0]);
+  u1 = _mm256_add_epi32(u1, v[1]);
+  u2 = _mm256_add_epi32(u2, v[2]);
+  u3 = _mm256_add_epi32(u3, v[3]);
+
+  u0 = _mm256_and_si256(u0, mask);
+  u1 = _mm256_and_si256(u1, mask);
+  u2 = _mm256_and_si256(u2, mask);
+  u3 = _mm256_and_si256(u3, mask);
+  // 4 32-bit summation, evenly positioned
+
+  u1 = _mm256_slli_si256(u1, 4);
+  u3 = _mm256_slli_si256(u3, 4);
+
+  u0 = _mm256_or_si256(u0, u1);
+  u2 = _mm256_or_si256(u2, u3);
+  // 8 32-bit summation, interleaved
+
+  u1 = _mm256_unpacklo_epi64(u0, u2);
+  u3 = _mm256_unpackhi_epi64(u0, u2);
+
+  u0 = _mm256_add_epi32(u1, u3);
+  sad = _mm_add_epi32(_mm256_extractf128_si256(u0, 1),
+                      _mm256_castsi256_si128(u0));
+  _mm_storeu_si128((__m128i *)res, sad);
+}
+
+static void convert_pointers(const uint8_t *const ref8[],
+                             const uint16_t *ref[]) {
+  ref[0] = CONVERT_TO_SHORTPTR(ref8[0]);
+  ref[1] = CONVERT_TO_SHORTPTR(ref8[1]);
+  ref[2] = CONVERT_TO_SHORTPTR(ref8[2]);
+  ref[3] = CONVERT_TO_SHORTPTR(ref8[3]);
+}
+
+static void init_sad(__m256i *s) {
+  s[0] = _mm256_setzero_si256();
+  s[1] = _mm256_setzero_si256();
+  s[2] = _mm256_setzero_si256();
+  s[3] = _mm256_setzero_si256();
+}
+
+void aom_highbd_sad16x8x4d_avx2(const uint8_t *src, int src_stride,
+                                const uint8_t *const ref_array[],
+                                int ref_stride, uint32_t *sad_array) {
+  __m256i sad_vec[4];
+  const uint16_t *refp[4];
+  const uint16_t *keep = CONVERT_TO_SHORTPTR(src);
+  const uint16_t *srcp;
+  const int shift_for_4_rows = 2;
+  int i;
+
+  init_sad(sad_vec);
+  convert_pointers(ref_array, refp);
+
+  for (i = 0; i < 4; ++i) {
+    srcp = keep;
+    sad16x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]);
+    srcp += src_stride << shift_for_4_rows;
+    refp[i] += ref_stride << shift_for_4_rows;
+    sad16x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]);
+  }
+  get_4d_sad_from_mm256_epi32(sad_vec, sad_array);
+}
+
+void aom_highbd_sad16x16x4d_avx2(const uint8_t *src, int src_stride,
+                                 const uint8_t *const ref_array[],
+                                 int ref_stride, uint32_t *sad_array) {
+  uint32_t first8rows[4];
+  uint32_t second8rows[4];
+  const uint8_t *ref[4];
+  const int shift_for_8_rows = 3;
+
+  ref[0] = ref_array[0];
+  ref[1] = ref_array[1];
+  ref[2] = ref_array[2];
+  ref[3] = ref_array[3];
+
+  aom_highbd_sad16x8x4d_avx2(src, src_stride, ref, ref_stride, first8rows);
+  src += src_stride << shift_for_8_rows;
+  ref[0] += ref_stride << shift_for_8_rows;
+  ref[1] += ref_stride << shift_for_8_rows;
+  ref[2] += ref_stride << shift_for_8_rows;
+  ref[3] += ref_stride << shift_for_8_rows;
+  aom_highbd_sad16x8x4d_avx2(src, src_stride, ref, ref_stride, second8rows);
+  sad_array[0] = first8rows[0] + second8rows[0];
+  sad_array[1] = first8rows[1] + second8rows[1];
+  sad_array[2] = first8rows[2] + second8rows[2];
+  sad_array[3] = first8rows[3] + second8rows[3];
+}
+
+void aom_highbd_sad16x32x4d_avx2(const uint8_t *src, int src_stride,
+                                 const uint8_t *const ref_array[],
+                                 int ref_stride, uint32_t *sad_array) {
+  uint32_t first_half[4];
+  uint32_t second_half[4];
+  const uint8_t *ref[4];
+  const int shift_for_rows = 4;
+
+  ref[0] = ref_array[0];
+  ref[1] = ref_array[1];
+  ref[2] = ref_array[2];
+  ref[3] = ref_array[3];
+
+  aom_highbd_sad16x16x4d_avx2(src, src_stride, ref, ref_stride, first_half);
+  src += src_stride << shift_for_rows;
+  ref[0] += ref_stride << shift_for_rows;
+  ref[1] += ref_stride << shift_for_rows;
+  ref[2] += ref_stride << shift_for_rows;
+  ref[3] += ref_stride << shift_for_rows;
+  aom_highbd_sad16x16x4d_avx2(src, src_stride, ref, ref_stride, second_half);
+  sad_array[0] = first_half[0] + second_half[0];
+  sad_array[1] = first_half[1] + second_half[1];
+  sad_array[2] = first_half[2] + second_half[2];
+  sad_array[3] = first_half[3] + second_half[3];
+}
+
+void aom_highbd_sad32x16x4d_avx2(const uint8_t *src, int src_stride,
+                                 const uint8_t *const ref_array[],
+                                 int ref_stride, uint32_t *sad_array) {
+  __m256i sad_vec[4];
+  const uint16_t *refp[4];
+  const uint16_t *keep = CONVERT_TO_SHORTPTR(src);
+  const uint16_t *srcp;
+  const int shift_for_4_rows = 2;
+  int i;
+  int rows_section;
+
+  init_sad(sad_vec);
+  convert_pointers(ref_array, refp);
+
+  for (i = 0; i < 4; ++i) {
+    srcp = keep;
+    rows_section = 0;
+    while (rows_section < 4) {
+      sad32x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]);
+      srcp += src_stride << shift_for_4_rows;
+      refp[i] += ref_stride << shift_for_4_rows;
+      rows_section++;
+    }
+  }
+  get_4d_sad_from_mm256_epi32(sad_vec, sad_array);
+}
+
+void aom_highbd_sad32x32x4d_avx2(const uint8_t *src, int src_stride,
+                                 const uint8_t *const ref_array[],
+                                 int ref_stride, uint32_t *sad_array) {
+  uint32_t first_half[4];
+  uint32_t second_half[4];
+  const uint8_t *ref[4];
+  const int shift_for_rows = 4;
+
+  ref[0] = ref_array[0];
+  ref[1] = ref_array[1];
+  ref[2] = ref_array[2];
+  ref[3] = ref_array[3];
+
+  aom_highbd_sad32x16x4d_avx2(src, src_stride, ref, ref_stride, first_half);
+  src += src_stride << shift_for_rows;
+  ref[0] += ref_stride << shift_for_rows;
+  ref[1] += ref_stride << shift_for_rows;
+  ref[2] += ref_stride << shift_for_rows;
+  ref[3] += ref_stride << shift_for_rows;
+  aom_highbd_sad32x16x4d_avx2(src, src_stride, ref, ref_stride, second_half);
+  sad_array[0] = first_half[0] + second_half[0];
+  sad_array[1] = first_half[1] + second_half[1];
+  sad_array[2] = first_half[2] + second_half[2];
+  sad_array[3] = first_half[3] + second_half[3];
+}
+
+void aom_highbd_sad32x64x4d_avx2(const uint8_t *src, int src_stride,
+                                 const uint8_t *const ref_array[],
+                                 int ref_stride, uint32_t *sad_array) {
+  uint32_t first_half[4];
+  uint32_t second_half[4];
+  const uint8_t *ref[4];
+  const int shift_for_rows = 5;
+
+  ref[0] = ref_array[0];
+  ref[1] = ref_array[1];
+  ref[2] = ref_array[2];
+  ref[3] = ref_array[3];
+
+  aom_highbd_sad32x32x4d_avx2(src, src_stride, ref, ref_stride, first_half);
+  src += src_stride << shift_for_rows;
+  ref[0] += ref_stride << shift_for_rows;
+  ref[1] += ref_stride << shift_for_rows;
+  ref[2] += ref_stride << shift_for_rows;
+  ref[3] += ref_stride << shift_for_rows;
+  aom_highbd_sad32x32x4d_avx2(src, src_stride, ref, ref_stride, second_half);
+  sad_array[0] = first_half[0] + second_half[0];
+  sad_array[1] = first_half[1] + second_half[1];
+  sad_array[2] = first_half[2] + second_half[2];
+  sad_array[3] = first_half[3] + second_half[3];
+}
+
+void aom_highbd_sad64x32x4d_avx2(const uint8_t *src, int src_stride,
+                                 const uint8_t *const ref_array[],
+                                 int ref_stride, uint32_t *sad_array) {
+  __m256i sad_vec[4];
+  const uint16_t *refp[4];
+  const uint16_t *keep = CONVERT_TO_SHORTPTR(src);
+  const uint16_t *srcp;
+  const int shift_for_rows = 1;
+  int i;
+  int rows_section;
+
+  init_sad(sad_vec);
+  convert_pointers(ref_array, refp);
+
+  for (i = 0; i < 4; ++i) {
+    srcp = keep;
+    rows_section = 0;
+    while (rows_section < 16) {
+      sad64x2(srcp, src_stride, refp[i], ref_stride, NULL, &sad_vec[i]);
+      srcp += src_stride << shift_for_rows;
+      refp[i] += ref_stride << shift_for_rows;
+      rows_section++;
+    }
+  }
+  get_4d_sad_from_mm256_epi32(sad_vec, sad_array);
+}
+
+void aom_highbd_sad64x64x4d_avx2(const uint8_t *src, int src_stride,
+                                 const uint8_t *const ref_array[],
+                                 int ref_stride, uint32_t *sad_array) {
+  uint32_t first_half[4];
+  uint32_t second_half[4];
+  const uint8_t *ref[4];
+  const int shift_for_rows = 5;
+
+  ref[0] = ref_array[0];
+  ref[1] = ref_array[1];
+  ref[2] = ref_array[2];
+  ref[3] = ref_array[3];
+
+  aom_highbd_sad64x32x4d_avx2(src, src_stride, ref, ref_stride, first_half);
+  src += src_stride << shift_for_rows;
+  ref[0] += ref_stride << shift_for_rows;
+  ref[1] += ref_stride << shift_for_rows;
+  ref[2] += ref_stride << shift_for_rows;
+  ref[3] += ref_stride << shift_for_rows;
+  aom_highbd_sad64x32x4d_avx2(src, src_stride, ref, ref_stride, second_half);
+  sad_array[0] = first_half[0] + second_half[0];
+  sad_array[1] = first_half[1] + second_half[1];
+  sad_array[2] = first_half[2] + second_half[2];
+  sad_array[3] = first_half[3] + second_half[3];
+}
+
+#if CONFIG_EXT_PARTITION
+void aom_highbd_sad64x128x4d_avx2(const uint8_t *src, int src_stride,
+                                  const uint8_t *const ref_array[],
+                                  int ref_stride, uint32_t *sad_array) {
+  uint32_t first_half[4];
+  uint32_t second_half[4];
+  const uint8_t *ref[4];
+  const int shift_for_rows = 6;
+
+  ref[0] = ref_array[0];
+  ref[1] = ref_array[1];
+  ref[2] = ref_array[2];
+  ref[3] = ref_array[3];
+
+  aom_highbd_sad64x64x4d_avx2(src, src_stride, ref, ref_stride, first_half);
+  src += src_stride << shift_for_rows;
+  ref[0] += ref_stride << shift_for_rows;
+  ref[1] += ref_stride << shift_for_rows;
+  ref[2] += ref_stride << shift_for_rows;
+  ref[3] += ref_stride << shift_for_rows;
+  aom_highbd_sad64x64x4d_avx2(src, src_stride, ref, ref_stride, second_half);
+  sad_array[0] = first_half[0] + second_half[0];
+  sad_array[1] = first_half[1] + second_half[1];
+  sad_array[2] = first_half[2] + second_half[2];
+  sad_array[3] = first_half[3] + second_half[3];
+}
+
+void aom_highbd_sad128x64x4d_avx2(const uint8_t *src, int src_stride,
+                                  const uint8_t *const ref_array[],
+                                  int ref_stride, uint32_t *sad_array) {
+  __m256i sad_vec[4];
+  const uint16_t *refp[4];
+  const uint16_t *keep = CONVERT_TO_SHORTPTR(src);
+  const uint16_t *srcp;
+  int i;
+  int rows_section;
+
+  init_sad(sad_vec);
+  convert_pointers(ref_array, refp);
+
+  for (i = 0; i < 4; ++i) {
+    srcp = keep;
+    rows_section = 0;
+    while (rows_section < 64) {
+      sad128x1(srcp, refp[i], NULL, &sad_vec[i]);
+      srcp += src_stride;
+      refp[i] += ref_stride;
+      rows_section++;
+    }
+  }
+  get_4d_sad_from_mm256_epi32(sad_vec, sad_array);
+}
+
+void aom_highbd_sad128x128x4d_avx2(const uint8_t *src, int src_stride,
+                                   const uint8_t *const ref_array[],
+                                   int ref_stride, uint32_t *sad_array) {
+  uint32_t first_half[4];
+  uint32_t second_half[4];
+  const uint8_t *ref[4];
+  const int shift_for_rows = 6;
+
+  ref[0] = ref_array[0];
+  ref[1] = ref_array[1];
+  ref[2] = ref_array[2];
+  ref[3] = ref_array[3];
+
+  aom_highbd_sad128x64x4d_avx2(src, src_stride, ref, ref_stride, first_half);
+  src += src_stride << shift_for_rows;
+  ref[0] += ref_stride << shift_for_rows;
+  ref[1] += ref_stride << shift_for_rows;
+  ref[2] += ref_stride << shift_for_rows;
+  ref[3] += ref_stride << shift_for_rows;
+  aom_highbd_sad128x64x4d_avx2(src, src_stride, ref, ref_stride, second_half);
+  sad_array[0] = first_half[0] + second_half[0];
+  sad_array[1] = first_half[1] + second_half[1];
+  sad_array[2] = first_half[2] + second_half[2];
+  sad_array[3] = first_half[3] + second_half[3];
+}
+#endif  // CONFIG_EXT_PARTITION
diff --git a/third_party/aom/aom_dsp/x86/sad_impl_avx2.c b/third_party/aom/aom_dsp/x86/sad_impl_avx2.c
new file mode 100644
index 0000000000..4419c65b29
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad_impl_avx2.c
@@ -0,0 +1,233 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>
+#include "./aom_dsp_rtcd.h"
+
+static unsigned int sad32x32(const uint8_t *src_ptr, int src_stride,
+                             const uint8_t *ref_ptr, int ref_stride) {
+  __m256i s1, s2, r1, r2;
+  __m256i sum = _mm256_setzero_si256();
+  __m128i sum_i128;
+  int i;
+
+  for (i = 0; i < 16; ++i) {
+    r1 = _mm256_loadu_si256((__m256i const *)ref_ptr);
+    r2 = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride));
+    s1 = _mm256_sad_epu8(r1, _mm256_loadu_si256((__m256i const *)src_ptr));
+    s2 = _mm256_sad_epu8(
+        r2, _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride)));
+    sum = _mm256_add_epi32(sum, _mm256_add_epi32(s1, s2));
+    ref_ptr += ref_stride << 1;
+    src_ptr += src_stride << 1;
+  }
+
+  sum = _mm256_add_epi32(sum, _mm256_srli_si256(sum, 8));
+  sum_i128 = _mm_add_epi32(_mm256_extracti128_si256(sum, 1),
+                           _mm256_castsi256_si128(sum));
+  return _mm_cvtsi128_si32(sum_i128);
+}
+
+static unsigned int sad64x32(const uint8_t *src_ptr, int src_stride,
+                             const uint8_t *ref_ptr, int ref_stride) {
+  unsigned int half_width = 32;
+  uint32_t sum = sad32x32(src_ptr, src_stride, ref_ptr, ref_stride);
+  src_ptr += half_width;
+  ref_ptr += half_width;
+  sum += sad32x32(src_ptr, src_stride, ref_ptr, ref_stride);
+  return sum;
+}
+
+static unsigned int sad64x64(const uint8_t *src_ptr, int src_stride,
+                             const uint8_t *ref_ptr, int ref_stride) {
+  uint32_t sum = sad64x32(src_ptr, src_stride, ref_ptr, ref_stride);
+  src_ptr += src_stride << 5;
+  ref_ptr += ref_stride << 5;
+  sum += sad64x32(src_ptr, src_stride, ref_ptr, ref_stride);
+  return sum;
+}
+
+unsigned int aom_sad128x64_avx2(const uint8_t *src_ptr, int src_stride,
+                                const uint8_t *ref_ptr, int ref_stride) {
+  unsigned int half_width = 64;
+  uint32_t sum = sad64x64(src_ptr, src_stride, ref_ptr, ref_stride);
+  src_ptr += half_width;
+  ref_ptr += half_width;
+  sum += sad64x64(src_ptr, src_stride, ref_ptr, ref_stride);
+  return sum;
+}
+
+unsigned int aom_sad64x128_avx2(const uint8_t *src_ptr, int src_stride,
+                                const uint8_t *ref_ptr, int ref_stride) {
+  uint32_t sum = sad64x64(src_ptr, src_stride, ref_ptr, ref_stride);
+  src_ptr += src_stride << 6;
+  ref_ptr += ref_stride << 6;
+  sum += sad64x64(src_ptr, src_stride, ref_ptr, ref_stride);
+  return sum;
+}
+
+unsigned int aom_sad128x128_avx2(const uint8_t *src_ptr, int src_stride,
+                                 const uint8_t *ref_ptr, int ref_stride) {
+  uint32_t sum = aom_sad128x64_avx2(src_ptr, src_stride, ref_ptr, ref_stride);
+  src_ptr += src_stride << 6;
+  ref_ptr += ref_stride << 6;
+  sum += aom_sad128x64_avx2(src_ptr, src_stride, ref_ptr, ref_stride);
+  return sum;
+}
+
+static void sad64x64x4d(const uint8_t *src, int src_stride,
+                        const uint8_t *const ref[4], int ref_stride,
+                        __m128i *res) {
+  uint32_t sum[4];
+  aom_sad64x64x4d_avx2(src, src_stride, ref, ref_stride, sum);
+  *res = _mm_loadu_si128((const __m128i *)sum);
+}
+
+void aom_sad64x128x4d_avx2(const uint8_t *src, int src_stride,
+                           const uint8_t *const ref[4], int ref_stride,
+                           uint32_t res[4]) {
+  __m128i sum0, sum1;
+  const uint8_t *rf[4];
+
+  rf[0] = ref[0];
+  rf[1] = ref[1];
+  rf[2] = ref[2];
+  rf[3] = ref[3];
+  sad64x64x4d(src, src_stride, rf, ref_stride, &sum0);
+  src += src_stride << 6;
+  rf[0] += ref_stride << 6;
+  rf[1] += ref_stride << 6;
+  rf[2] += ref_stride << 6;
+  rf[3] += ref_stride << 6;
+  sad64x64x4d(src, src_stride, rf, ref_stride, &sum1);
+  sum0 = _mm_add_epi32(sum0, sum1);
+  _mm_storeu_si128((__m128i *)res, sum0);
+}
+
+void aom_sad128x64x4d_avx2(const uint8_t *src, int src_stride,
+                           const uint8_t *const ref[4], int ref_stride,
+                           uint32_t res[4]) {
+  __m128i sum0, sum1;
+  unsigned int half_width = 64;
+  const uint8_t *rf[4];
+
+  rf[0] = ref[0];
+  rf[1] = ref[1];
+  rf[2] = ref[2];
+  rf[3] = ref[3];
+  sad64x64x4d(src, src_stride, rf, ref_stride, &sum0);
+  src += half_width;
+  rf[0] += half_width;
+  rf[1] += half_width;
+  rf[2] += half_width;
+  rf[3] += half_width;
+  sad64x64x4d(src, src_stride, rf, ref_stride, &sum1);
+  sum0 = _mm_add_epi32(sum0, sum1);
+  _mm_storeu_si128((__m128i *)res, sum0);
+}
+
+void aom_sad128x128x4d_avx2(const uint8_t *src, int src_stride,
+                            const uint8_t *const ref[4], int ref_stride,
+                            uint32_t res[4]) {
+  const uint8_t *rf[4];
+  uint32_t sum0[4];
+  uint32_t sum1[4];
+
+  rf[0] = ref[0];
+  rf[1] = ref[1];
+  rf[2] = ref[2];
+  rf[3] = ref[3];
+  aom_sad128x64x4d_avx2(src, src_stride, rf, ref_stride, sum0);
+  src += src_stride << 6;
+  rf[0] += ref_stride << 6;
+  rf[1] += ref_stride << 6;
+  rf[2] += ref_stride << 6;
+  rf[3] += ref_stride << 6;
+  aom_sad128x64x4d_avx2(src, src_stride, rf, ref_stride, sum1);
+  res[0] = sum0[0] + sum1[0];
+  res[1] = sum0[1] + sum1[1];
+  res[2] = sum0[2] + sum1[2];
+  res[3] = sum0[3] + sum1[3];
+}
+
+static unsigned int sad_w64_avg_avx2(const uint8_t *src_ptr, int src_stride,
+                                     const uint8_t *ref_ptr, int ref_stride,
+                                     const int h, const uint8_t *second_pred,
+                                     const int second_pred_stride) {
+  int i, res;
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;
+  __m256i sum_sad = _mm256_setzero_si256();
+  __m256i sum_sad_h;
+  __m128i sum_sad128;
+  for (i = 0; i < h; i++) {
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32));
+    ref1_reg = _mm256_avg_epu8(
+        ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred));
+    ref2_reg = _mm256_avg_epu8(
+        ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32)));
+    sad1_reg =
+        _mm256_sad_epu8(ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));
+    sad2_reg = _mm256_sad_epu8(
+        ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32)));
+    sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));
+    ref_ptr += ref_stride;
+    src_ptr += src_stride;
+    second_pred += second_pred_stride;
+  }
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8);
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);
+  res = _mm_cvtsi128_si32(sum_sad128);
+
+  return res;
+}
+
+unsigned int aom_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride,
+                                    const uint8_t *ref_ptr, int ref_stride,
+                                    const uint8_t *second_pred) {
+  uint32_t sum = sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64,
+                                  second_pred, 64);
+  src_ptr += src_stride << 6;
+  ref_ptr += ref_stride << 6;
+  second_pred += 64 << 6;
+  sum += sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64,
+                          second_pred, 64);
+  return sum;
+}
+
+unsigned int aom_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride,
+                                    const uint8_t *ref_ptr, int ref_stride,
+                                    const uint8_t *second_pred) {
+  unsigned int half_width = 64;
+  uint32_t sum = sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64,
+                                  second_pred, 128);
+  src_ptr += half_width;
+  ref_ptr += half_width;
+  second_pred += half_width;
+  sum += sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64,
+                          second_pred, 128);
+  return sum;
+}
+
+unsigned int aom_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride,
+                                     const uint8_t *ref_ptr, int ref_stride,
+                                     const uint8_t *second_pred) {
+  uint32_t sum = aom_sad128x64_avg_avx2(src_ptr, src_stride, ref_ptr,
+                                        ref_stride, second_pred);
+  src_ptr += src_stride << 6;
+  ref_ptr += ref_stride << 6;
+  second_pred += 128 << 6;
+  sum += aom_sad128x64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride,
+                                second_pred);
+  return sum;
+}
diff --git a/third_party/aom/aom_dsp/x86/sad_sse2.asm b/third_party/aom/aom_dsp/x86/sad_sse2.asm
new file mode 100644
index 0000000000..e45457a572
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad_sse2.asm
@@ -0,0 +1,345 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro SAD_FN 4
+%if %4 == 0
+%if %3 == 5
+cglobal sad%1x%2, 4, %3, 5, src, src_stride, ref, ref_stride, n_rows
+%else ; %3 == 7
+cglobal sad%1x%2, 4, %3, 6, src, src_stride, ref, ref_stride, \
+                            src_stride3, ref_stride3, n_rows
+%endif ; %3 == 5/7
+%else ; avg
+%if %3 == 5
+cglobal sad%1x%2_avg, 5, 1 + %3, 5, src, src_stride, ref, ref_stride, \
+                                    second_pred, n_rows
+%else ; %3 == 7
+cglobal sad%1x%2_avg, 5, ARCH_X86_64 + %3, 6, src, src_stride, \
+                                              ref, ref_stride, \
+                                              second_pred, \
+                                              src_stride3, ref_stride3
+%if ARCH_X86_64
+%define n_rowsd r7d
+%else ; x86-32
+%define n_rowsd dword r0m
+%endif ; x86-32/64
+%endif ; %3 == 5/7
+%endif ; avg/sad
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+%if %3 == 7
+  lea         src_stride3q, [src_strideq*3]
+  lea         ref_stride3q, [ref_strideq*3]
+%endif ; %3 == 7
+%endmacro
+
+%if CONFIG_EXT_PARTITION
+; unsigned int aom_sad128x128_sse2(uint8_t *src, int src_stride,
+;                                  uint8_t *ref, int ref_stride);
+%macro SAD128XN 1-2 0
+  SAD_FN 128, %1, 5, %2
+  mov              n_rowsd, %1
+  pxor                  m0, m0
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+16]
+  psadbw                m3, [srcq+32]
+  psadbw                m4, [srcq+48]
+
+  paddd                 m1, m2
+  paddd                 m3, m4
+  paddd                 m0, m1
+  paddd                 m0, m3
+
+  movu                  m1, [refq+64]
+  movu                  m2, [refq+80]
+  movu                  m3, [refq+96]
+  movu                  m4, [refq+112]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*4]
+  pavgb                 m2, [second_predq+mmsize*5]
+  pavgb                 m3, [second_predq+mmsize*6]
+  pavgb                 m4, [second_predq+mmsize*7]
+  lea         second_predq, [second_predq+mmsize*8]
+%endif
+  psadbw                m1, [srcq+64]
+  psadbw                m2, [srcq+80]
+  psadbw                m3, [srcq+96]
+  psadbw                m4, [srcq+112]
+
+  add                 refq, ref_strideq
+  add                 srcq, src_strideq
+
+  paddd                 m1, m2
+  paddd                 m3, m4
+  paddd                 m0, m1
+  paddd                 m0, m3
+
+  sub              n_rowsd, 1
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD128XN 128     ; sad128x128_sse2
+SAD128XN 128, 1  ; sad128x128_avg_sse2
+SAD128XN 64      ; sad128x64_sse2
+SAD128XN 64, 1   ; sad128x64_avg_sse2
+%endif
+
+
+; unsigned int aom_sad64x64_sse2(uint8_t *src, int src_stride,
+;                                uint8_t *ref, int ref_stride);
+%macro SAD64XN 1-2 0
+  SAD_FN 64, %1, 5, %2
+  mov              n_rowsd, %1
+  pxor                  m0, m0
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+16]
+  psadbw                m3, [srcq+32]
+  psadbw                m4, [srcq+48]
+  paddd                 m1, m2
+  paddd                 m3, m4
+  add                 refq, ref_strideq
+  paddd                 m0, m1
+  add                 srcq, src_strideq
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+%if CONFIG_EXT_PARTITION
+SAD64XN 128     ; sad64x128_sse2
+SAD64XN 128, 1  ; sad64x128_avg_sse2
+%endif
+SAD64XN 64 ; sad64x64_sse2
+SAD64XN 32 ; sad64x32_sse2
+SAD64XN 64, 1 ; sad64x64_avg_sse2
+SAD64XN 32, 1 ; sad64x32_avg_sse2
+
+; unsigned int aom_sad32x32_sse2(uint8_t *src, int src_stride,
+;                                uint8_t *ref, int ref_stride);
+%macro SAD32XN 1-2 0
+  SAD_FN 32, %1, 5, %2
+  mov              n_rowsd, %1/2
+  pxor                  m0, m0
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+ref_strideq]
+  movu                  m4, [refq+ref_strideq+16]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+16]
+  psadbw                m3, [srcq+src_strideq]
+  psadbw                m4, [srcq+src_strideq+16]
+  paddd                 m1, m2
+  paddd                 m3, m4
+  lea                 refq, [refq+ref_strideq*2]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*2]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD32XN 64 ; sad32x64_sse2
+SAD32XN 32 ; sad32x32_sse2
+SAD32XN 16 ; sad32x16_sse2
+SAD32XN 64, 1 ; sad32x64_avg_sse2
+SAD32XN 32, 1 ; sad32x32_avg_sse2
+SAD32XN 16, 1 ; sad32x16_avg_sse2
+
+; unsigned int aom_sad16x{8,16}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro SAD16XN 1-2 0
+  SAD_FN 16, %1, 7, %2
+  mov              n_rowsd, %1/4
+  pxor                  m0, m0
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+ref_strideq]
+  movu                  m3, [refq+ref_strideq*2]
+  movu                  m4, [refq+ref_stride3q]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+src_strideq]
+  psadbw                m3, [srcq+src_strideq*2]
+  psadbw                m4, [srcq+src_stride3q]
+  paddd                 m1, m2
+  paddd                 m3, m4
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD16XN 32 ; sad16x32_sse2
+SAD16XN 16 ; sad16x16_sse2
+SAD16XN  8 ; sad16x8_sse2
+SAD16XN 32, 1 ; sad16x32_avg_sse2
+SAD16XN 16, 1 ; sad16x16_avg_sse2
+SAD16XN  8, 1 ; sad16x8_avg_sse2
+
+; unsigned int aom_sad8x{8,16}_sse2(uint8_t *src, int src_stride,
+;                                   uint8_t *ref, int ref_stride);
+%macro SAD8XN 1-2 0
+  SAD_FN 8, %1, 7, %2
+  mov              n_rowsd, %1/4
+  pxor                  m0, m0
+
+.loop:
+  movh                  m1, [refq]
+  movhps                m1, [refq+ref_strideq]
+  movh                  m2, [refq+ref_strideq*2]
+  movhps                m2, [refq+ref_stride3q]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  lea         second_predq, [second_predq+mmsize*2]
+%endif
+  movh                  m3, [srcq]
+  movhps                m3, [srcq+src_strideq]
+  movh                  m4, [srcq+src_strideq*2]
+  movhps                m4, [srcq+src_stride3q]
+  psadbw                m1, m3
+  psadbw                m2, m4
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  paddd                 m0, m2
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD8XN 16 ; sad8x16_sse2
+SAD8XN  8 ; sad8x8_sse2
+SAD8XN  4 ; sad8x4_sse2
+SAD8XN 16, 1 ; sad8x16_avg_sse2
+SAD8XN  8, 1 ; sad8x8_avg_sse2
+SAD8XN  4, 1 ; sad8x4_avg_sse2
+
+; unsigned int aom_sad4x{4, 8}_sse2(uint8_t *src, int src_stride,
+;                                   uint8_t *ref, int ref_stride);
+%macro SAD4XN 1-2 0
+  SAD_FN 4, %1, 7, %2
+  mov              n_rowsd, %1/4
+  pxor                  m0, m0
+
+.loop:
+  movd                  m1, [refq]
+  movd                  m2, [refq+ref_strideq]
+  movd                  m3, [refq+ref_strideq*2]
+  movd                  m4, [refq+ref_stride3q]
+  punpckldq             m1, m2
+  punpckldq             m3, m4
+  movlhps               m1, m3
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  lea         second_predq, [second_predq+mmsize*1]
+%endif
+  movd                  m2, [srcq]
+  movd                  m5, [srcq+src_strideq]
+  movd                  m4, [srcq+src_strideq*2]
+  movd                  m3, [srcq+src_stride3q]
+  punpckldq             m2, m5
+  punpckldq             m4, m3
+  movlhps               m2, m4
+  psadbw                m1, m2
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD4XN  8 ; sad4x8_sse
+SAD4XN  4 ; sad4x4_sse
+SAD4XN  8, 1 ; sad4x8_avg_sse
+SAD4XN  4, 1 ; sad4x4_avg_sse
diff --git a/third_party/aom/aom_dsp/x86/sad_sse3.asm b/third_party/aom/aom_dsp/x86/sad_sse3.asm
new file mode 100644
index 0000000000..f6c27c855a
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad_sse3.asm
@@ -0,0 +1,377 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "aom_ports/x86_abi_support.asm"
+
+%macro STACK_FRAME_CREATE_X3 0
+%if ABI_IS_32BIT
+  %define     src_ptr       rsi
+  %define     src_stride    rax
+  %define     ref_ptr       rdi
+  %define     ref_stride    rdx
+  %define     end_ptr       rcx
+  %define     ret_var       rbx
+  %define     result_ptr    arg(4)
+  %define     height        dword ptr arg(4)
+    push        rbp
+    mov         rbp,        rsp
+    push        rsi
+    push        rdi
+    push        rbx
+
+    mov         rsi,        arg(0)              ; src_ptr
+    mov         rdi,        arg(2)              ; ref_ptr
+
+    movsxd      rax,        dword ptr arg(1)    ; src_stride
+    movsxd      rdx,        dword ptr arg(3)    ; ref_stride
+%else
+  %if LIBAOM_YASM_WIN64
+    SAVE_XMM 7, u
+    %define     src_ptr     rcx
+    %define     src_stride  rdx
+    %define     ref_ptr     r8
+    %define     ref_stride  r9
+    %define     end_ptr     r10
+    %define     ret_var     r11
+    %define     result_ptr  [rsp+xmm_stack_space+8+4*8]
+    %define     height      dword ptr [rsp+xmm_stack_space+8+4*8]
+  %else
+    %define     src_ptr     rdi
+    %define     src_stride  rsi
+    %define     ref_ptr     rdx
+    %define     ref_stride  rcx
+    %define     end_ptr     r9
+    %define     ret_var     r10
+    %define     result_ptr  r8
+    %define     height      r8
+  %endif
+%endif
+
+%endmacro
+
+%macro STACK_FRAME_DESTROY_X3 0
+  %define     src_ptr
+  %define     src_stride
+  %define     ref_ptr
+  %define     ref_stride
+  %define     end_ptr
+  %define     ret_var
+  %define     result_ptr
+  %define     height
+
+%if ABI_IS_32BIT
+    pop         rbx
+    pop         rdi
+    pop         rsi
+    pop         rbp
+%else
+  %if LIBAOM_YASM_WIN64
+    RESTORE_XMM
+  %endif
+%endif
+    ret
+%endmacro
+
+%macro PROCESS_16X2X3 5
+%if %1==0
+        movdqa          xmm0,       XMMWORD PTR [%2]
+        lddqu           xmm5,       XMMWORD PTR [%3]
+        lddqu           xmm6,       XMMWORD PTR [%3+1]
+        lddqu           xmm7,       XMMWORD PTR [%3+2]
+
+        psadbw          xmm5,       xmm0
+        psadbw          xmm6,       xmm0
+        psadbw          xmm7,       xmm0
+%else
+        movdqa          xmm0,       XMMWORD PTR [%2]
+        lddqu           xmm1,       XMMWORD PTR [%3]
+        lddqu           xmm2,       XMMWORD PTR [%3+1]
+        lddqu           xmm3,       XMMWORD PTR [%3+2]
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endif
+        movdqa          xmm0,       XMMWORD PTR [%2+%4]
+        lddqu           xmm1,       XMMWORD PTR [%3+%5]
+        lddqu           xmm2,       XMMWORD PTR [%3+%5+1]
+        lddqu           xmm3,       XMMWORD PTR [%3+%5+2]
+
+%if %1==0 || %1==1
+        lea             %2,         [%2+%4*2]
+        lea             %3,         [%3+%5*2]
+%endif
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endmacro
+
+%macro PROCESS_8X2X3 5
+%if %1==0
+        movq            mm0,       QWORD PTR [%2]
+        movq            mm5,       QWORD PTR [%3]
+        movq            mm6,       QWORD PTR [%3+1]
+        movq            mm7,       QWORD PTR [%3+2]
+
+        psadbw          mm5,       mm0
+        psadbw          mm6,       mm0
+        psadbw          mm7,       mm0
+%else
+        movq            mm0,       QWORD PTR [%2]
+        movq            mm1,       QWORD PTR [%3]
+        movq            mm2,       QWORD PTR [%3+1]
+        movq            mm3,       QWORD PTR [%3+2]
+
+        psadbw          mm1,       mm0
+        psadbw          mm2,       mm0
+        psadbw          mm3,       mm0
+
+        paddw           mm5,       mm1
+        paddw           mm6,       mm2
+        paddw           mm7,       mm3
+%endif
+        movq            mm0,       QWORD PTR [%2+%4]
+        movq            mm1,       QWORD PTR [%3+%5]
+        movq            mm2,       QWORD PTR [%3+%5+1]
+        movq            mm3,       QWORD PTR [%3+%5+2]
+
+%if %1==0 || %1==1
+        lea             %2,        [%2+%4*2]
+        lea             %3,        [%3+%5*2]
+%endif
+
+        psadbw          mm1,       mm0
+        psadbw          mm2,       mm0
+        psadbw          mm3,       mm0
+
+        paddw           mm5,       mm1
+        paddw           mm6,       mm2
+        paddw           mm7,       mm3
+%endmacro
+
+;void int aom_sad16x16x3_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(aom_sad16x16x3_sse3) PRIVATE
+sym(aom_sad16x16x3_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+        PROCESS_16X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
+
+        mov             rcx,        result_ptr
+
+        movq            xmm0,       xmm5
+        psrldq          xmm5,       8
+
+        paddw           xmm0,       xmm5
+        movd            [rcx],      xmm0
+;-
+        movq            xmm0,       xmm6
+        psrldq          xmm6,       8
+
+        paddw           xmm0,       xmm6
+        movd            [rcx+4],    xmm0
+;-
+        movq            xmm0,       xmm7
+        psrldq          xmm7,       8
+
+        paddw           xmm0,       xmm7
+        movd            [rcx+8],    xmm0
+
+    STACK_FRAME_DESTROY_X3
+
+;void int aom_sad16x8x3_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(aom_sad16x8x3_sse3) PRIVATE
+sym(aom_sad16x8x3_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+        PROCESS_16X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_16X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
+
+        mov             rcx,        result_ptr
+
+        movq            xmm0,       xmm5
+        psrldq          xmm5,       8
+
+        paddw           xmm0,       xmm5
+        movd            [rcx],      xmm0
+;-
+        movq            xmm0,       xmm6
+        psrldq          xmm6,       8
+
+        paddw           xmm0,       xmm6
+        movd            [rcx+4],    xmm0
+;-
+        movq            xmm0,       xmm7
+        psrldq          xmm7,       8
+
+        paddw           xmm0,       xmm7
+        movd            [rcx+8],    xmm0
+
+    STACK_FRAME_DESTROY_X3
+
+;void int aom_sad8x16x3_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(aom_sad8x16x3_sse3) PRIVATE
+sym(aom_sad8x16x3_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+        PROCESS_8X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
+
+        mov             rcx,        result_ptr
+
+        punpckldq       mm5,        mm6
+
+        movq            [rcx],      mm5
+        movd            [rcx+8],    mm7
+
+    STACK_FRAME_DESTROY_X3
+
+;void int aom_sad8x8x3_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(aom_sad8x8x3_sse3) PRIVATE
+sym(aom_sad8x8x3_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+        PROCESS_8X2X3 0, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 1, src_ptr, ref_ptr, src_stride, ref_stride
+        PROCESS_8X2X3 2, src_ptr, ref_ptr, src_stride, ref_stride
+
+        mov             rcx,        result_ptr
+
+        punpckldq       mm5,        mm6
+
+        movq            [rcx],      mm5
+        movd            [rcx+8],    mm7
+
+    STACK_FRAME_DESTROY_X3
+
+;void int aom_sad4x4x3_sse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(aom_sad4x4x3_sse3) PRIVATE
+sym(aom_sad4x4x3_sse3):
+
+    STACK_FRAME_CREATE_X3
+
+        movd            mm0,        DWORD PTR [src_ptr]
+        movd            mm1,        DWORD PTR [ref_ptr]
+
+        movd            mm2,        DWORD PTR [src_ptr+src_stride]
+        movd            mm3,        DWORD PTR [ref_ptr+ref_stride]
+
+        punpcklbw       mm0,        mm2
+        punpcklbw       mm1,        mm3
+
+        movd            mm4,        DWORD PTR [ref_ptr+1]
+        movd            mm5,        DWORD PTR [ref_ptr+2]
+
+        movd            mm2,        DWORD PTR [ref_ptr+ref_stride+1]
+        movd            mm3,        DWORD PTR [ref_ptr+ref_stride+2]
+
+        psadbw          mm1,        mm0
+
+        punpcklbw       mm4,        mm2
+        punpcklbw       mm5,        mm3
+
+        psadbw          mm4,        mm0
+        psadbw          mm5,        mm0
+
+        lea             src_ptr,    [src_ptr+src_stride*2]
+        lea             ref_ptr,    [ref_ptr+ref_stride*2]
+
+        movd            mm0,        DWORD PTR [src_ptr]
+        movd            mm2,        DWORD PTR [ref_ptr]
+
+        movd            mm3,        DWORD PTR [src_ptr+src_stride]
+        movd            mm6,        DWORD PTR [ref_ptr+ref_stride]
+
+        punpcklbw       mm0,        mm3
+        punpcklbw       mm2,        mm6
+
+        movd            mm3,        DWORD PTR [ref_ptr+1]
+        movd            mm7,        DWORD PTR [ref_ptr+2]
+
+        psadbw          mm2,        mm0
+
+        paddw           mm1,        mm2
+
+        movd            mm2,        DWORD PTR [ref_ptr+ref_stride+1]
+        movd            mm6,        DWORD PTR [ref_ptr+ref_stride+2]
+
+        punpcklbw       mm3,        mm2
+        punpcklbw       mm7,        mm6
+
+        psadbw          mm3,        mm0
+        psadbw          mm7,        mm0
+
+        paddw           mm3,        mm4
+        paddw           mm7,        mm5
+
+        mov             rcx,        result_ptr
+
+        punpckldq       mm1,        mm3
+
+        movq            [rcx],      mm1
+        movd            [rcx+8],    mm7
+
+    STACK_FRAME_DESTROY_X3
diff --git a/third_party/aom/aom_dsp/x86/sad_sse4.asm b/third_party/aom/aom_dsp/x86/sad_sse4.asm
new file mode 100644
index 0000000000..5e9c758451
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad_sse4.asm
@@ -0,0 +1,362 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+%include "aom_ports/x86_abi_support.asm"
+
+%macro PROCESS_16X2X8 1
+%if %1
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        movq            xmm1,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        movq            xmm2,       MMWORD PTR [rdi+16]
+        punpcklqdq      xmm1,       xmm3
+        punpcklqdq      xmm3,       xmm2
+
+        movdqa          xmm2,       xmm1
+        mpsadbw         xmm1,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+
+        psrldq          xmm0,       8
+
+        movdqa          xmm4,       xmm3
+        mpsadbw         xmm3,       xmm0,  0x0
+        mpsadbw         xmm4,       xmm0,  0x5
+
+        paddw           xmm1,       xmm2
+        paddw           xmm1,       xmm3
+        paddw           xmm1,       xmm4
+%else
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        movq            xmm5,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        movq            xmm2,       MMWORD PTR [rdi+16]
+        punpcklqdq      xmm5,       xmm3
+        punpcklqdq      xmm3,       xmm2
+
+        movdqa          xmm2,       xmm5
+        mpsadbw         xmm5,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+
+        psrldq          xmm0,       8
+
+        movdqa          xmm4,       xmm3
+        mpsadbw         xmm3,       xmm0,  0x0
+        mpsadbw         xmm4,       xmm0,  0x5
+
+        paddw           xmm5,       xmm2
+        paddw           xmm5,       xmm3
+        paddw           xmm5,       xmm4
+
+        paddw           xmm1,       xmm5
+%endif
+        movdqa          xmm0,       XMMWORD PTR [rsi + rax]
+        movq            xmm5,       MMWORD PTR [rdi+ rdx]
+        movq            xmm3,       MMWORD PTR [rdi+ rdx+8]
+        movq            xmm2,       MMWORD PTR [rdi+ rdx+16]
+        punpcklqdq      xmm5,       xmm3
+        punpcklqdq      xmm3,       xmm2
+
+        lea             rsi,        [rsi+rax*2]
+        lea             rdi,        [rdi+rdx*2]
+
+        movdqa          xmm2,       xmm5
+        mpsadbw         xmm5,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+
+        psrldq          xmm0,       8
+        movdqa          xmm4,       xmm3
+        mpsadbw         xmm3,       xmm0,  0x0
+        mpsadbw         xmm4,       xmm0,  0x5
+
+        paddw           xmm5,       xmm2
+        paddw           xmm5,       xmm3
+        paddw           xmm5,       xmm4
+
+        paddw           xmm1,       xmm5
+%endmacro
+
+%macro PROCESS_8X2X8 1
+%if %1
+        movq            xmm0,       MMWORD PTR [rsi]
+        movq            xmm1,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        punpcklqdq      xmm1,       xmm3
+
+        movdqa          xmm2,       xmm1
+        mpsadbw         xmm1,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+        paddw           xmm1,       xmm2
+%else
+        movq            xmm0,       MMWORD PTR [rsi]
+        movq            xmm5,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        punpcklqdq      xmm5,       xmm3
+
+        movdqa          xmm2,       xmm5
+        mpsadbw         xmm5,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+        paddw           xmm5,       xmm2
+
+        paddw           xmm1,       xmm5
+%endif
+        movq            xmm0,       MMWORD PTR [rsi + rax]
+        movq            xmm5,       MMWORD PTR [rdi+ rdx]
+        movq            xmm3,       MMWORD PTR [rdi+ rdx+8]
+        punpcklqdq      xmm5,       xmm3
+
+        lea             rsi,        [rsi+rax*2]
+        lea             rdi,        [rdi+rdx*2]
+
+        movdqa          xmm2,       xmm5
+        mpsadbw         xmm5,       xmm0,  0x0
+        mpsadbw         xmm2,       xmm0,  0x5
+        paddw           xmm5,       xmm2
+
+        paddw           xmm1,       xmm5
+%endmacro
+
+%macro PROCESS_4X2X8 1
+%if %1
+        movd            xmm0,       [rsi]
+        movq            xmm1,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        punpcklqdq      xmm1,       xmm3
+
+        mpsadbw         xmm1,       xmm0,  0x0
+%else
+        movd            xmm0,       [rsi]
+        movq            xmm5,       MMWORD PTR [rdi]
+        movq            xmm3,       MMWORD PTR [rdi+8]
+        punpcklqdq      xmm5,       xmm3
+
+        mpsadbw         xmm5,       xmm0,  0x0
+
+        paddw           xmm1,       xmm5
+%endif
+        movd            xmm0,       [rsi + rax]
+        movq            xmm5,       MMWORD PTR [rdi+ rdx]
+        movq            xmm3,       MMWORD PTR [rdi+ rdx+8]
+        punpcklqdq      xmm5,       xmm3
+
+        lea             rsi,        [rsi+rax*2]
+        lea             rdi,        [rdi+rdx*2]
+
+        mpsadbw         xmm5,       xmm0,  0x0
+
+        paddw           xmm1,       xmm5
+%endmacro
+
+%macro WRITE_AS_INTS 0
+    mov             rdi,        arg(4)           ;Results
+    pxor            xmm0, xmm0
+    movdqa          xmm2, xmm1
+    punpcklwd       xmm1, xmm0
+    punpckhwd       xmm2, xmm0
+
+    movdqa          [rdi],    xmm1
+    movdqa          [rdi + 16],    xmm2
+%endmacro
+
+;void aom_sad16x16x8_sse4_1(
+;    const unsigned char *src_ptr,
+;    int  src_stride,
+;    const unsigned char *ref_ptr,
+;    int  ref_stride,
+;    unsigned short *sad_array);
+global sym(aom_sad16x16x8_sse4_1) PRIVATE
+sym(aom_sad16x16x8_sse4_1):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0)           ;src_ptr
+    mov             rdi,        arg(2)           ;ref_ptr
+
+    movsxd          rax,        dword ptr arg(1) ;src_stride
+    movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+    PROCESS_16X2X8 1
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+
+    WRITE_AS_INTS
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void aom_sad16x8x8_sse4_1(
+;    const unsigned char *src_ptr,
+;    int  src_stride,
+;    const unsigned char *ref_ptr,
+;    int  ref_stride,
+;    unsigned short *sad_array
+;);
+global sym(aom_sad16x8x8_sse4_1) PRIVATE
+sym(aom_sad16x8x8_sse4_1):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0)           ;src_ptr
+    mov             rdi,        arg(2)           ;ref_ptr
+
+    movsxd          rax,        dword ptr arg(1) ;src_stride
+    movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+    PROCESS_16X2X8 1
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+    PROCESS_16X2X8 0
+
+    WRITE_AS_INTS
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void aom_sad8x8x8_sse4_1(
+;    const unsigned char *src_ptr,
+;    int  src_stride,
+;    const unsigned char *ref_ptr,
+;    int  ref_stride,
+;    unsigned short *sad_array
+;);
+global sym(aom_sad8x8x8_sse4_1) PRIVATE
+sym(aom_sad8x8x8_sse4_1):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0)           ;src_ptr
+    mov             rdi,        arg(2)           ;ref_ptr
+
+    movsxd          rax,        dword ptr arg(1) ;src_stride
+    movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+    PROCESS_8X2X8 1
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+
+    WRITE_AS_INTS
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void aom_sad8x16x8_sse4_1(
+;    const unsigned char *src_ptr,
+;    int  src_stride,
+;    const unsigned char *ref_ptr,
+;    int  ref_stride,
+;    unsigned short *sad_array
+;);
+global sym(aom_sad8x16x8_sse4_1) PRIVATE
+sym(aom_sad8x16x8_sse4_1):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0)           ;src_ptr
+    mov             rdi,        arg(2)           ;ref_ptr
+
+    movsxd          rax,        dword ptr arg(1) ;src_stride
+    movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+    PROCESS_8X2X8 1
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+    PROCESS_8X2X8 0
+
+    WRITE_AS_INTS
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;void aom_sad4x4x8_sse4_1(
+;    const unsigned char *src_ptr,
+;    int  src_stride,
+;    const unsigned char *ref_ptr,
+;    int  ref_stride,
+;    unsigned short *sad_array
+;);
+global sym(aom_sad4x4x8_sse4_1) PRIVATE
+sym(aom_sad4x4x8_sse4_1):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0)           ;src_ptr
+    mov             rdi,        arg(2)           ;ref_ptr
+
+    movsxd          rax,        dword ptr arg(1) ;src_stride
+    movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+    PROCESS_4X2X8 1
+    PROCESS_4X2X8 0
+
+    WRITE_AS_INTS
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+
+
diff --git a/third_party/aom/aom_dsp/x86/sad_ssse3.asm b/third_party/aom/aom_dsp/x86/sad_ssse3.asm
new file mode 100644
index 0000000000..96b64b040b
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad_ssse3.asm
@@ -0,0 +1,373 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+%include "aom_ports/x86_abi_support.asm"
+
+%macro PROCESS_16X2X3 1
+%if %1
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        lddqu           xmm5,       XMMWORD PTR [rdi]
+        lddqu           xmm6,       XMMWORD PTR [rdi+1]
+        lddqu           xmm7,       XMMWORD PTR [rdi+2]
+
+        psadbw          xmm5,       xmm0
+        psadbw          xmm6,       xmm0
+        psadbw          xmm7,       xmm0
+%else
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        lddqu           xmm1,       XMMWORD PTR [rdi]
+        lddqu           xmm2,       XMMWORD PTR [rdi+1]
+        lddqu           xmm3,       XMMWORD PTR [rdi+2]
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endif
+        movdqa          xmm0,       XMMWORD PTR [rsi+rax]
+        lddqu           xmm1,       XMMWORD PTR [rdi+rdx]
+        lddqu           xmm2,       XMMWORD PTR [rdi+rdx+1]
+        lddqu           xmm3,       XMMWORD PTR [rdi+rdx+2]
+
+        lea             rsi,        [rsi+rax*2]
+        lea             rdi,        [rdi+rdx*2]
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endmacro
+
+%macro PROCESS_16X2X3_OFFSET 2
+%if %1
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        movdqa          xmm4,       XMMWORD PTR [rdi]
+        movdqa          xmm7,       XMMWORD PTR [rdi+16]
+
+        movdqa          xmm5,       xmm7
+        palignr         xmm5,       xmm4,       %2
+
+        movdqa          xmm6,       xmm7
+        palignr         xmm6,       xmm4,       (%2+1)
+
+        palignr         xmm7,       xmm4,       (%2+2)
+
+        psadbw          xmm5,       xmm0
+        psadbw          xmm6,       xmm0
+        psadbw          xmm7,       xmm0
+%else
+        movdqa          xmm0,       XMMWORD PTR [rsi]
+        movdqa          xmm4,       XMMWORD PTR [rdi]
+        movdqa          xmm3,       XMMWORD PTR [rdi+16]
+
+        movdqa          xmm1,       xmm3
+        palignr         xmm1,       xmm4,       %2
+
+        movdqa          xmm2,       xmm3
+        palignr         xmm2,       xmm4,       (%2+1)
+
+        palignr         xmm3,       xmm4,       (%2+2)
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endif
+        movdqa          xmm0,       XMMWORD PTR [rsi+rax]
+        movdqa          xmm4,       XMMWORD PTR [rdi+rdx]
+        movdqa          xmm3,       XMMWORD PTR [rdi+rdx+16]
+
+        movdqa          xmm1,       xmm3
+        palignr         xmm1,       xmm4,       %2
+
+        movdqa          xmm2,       xmm3
+        palignr         xmm2,       xmm4,       (%2+1)
+
+        palignr         xmm3,       xmm4,       (%2+2)
+
+        lea             rsi,        [rsi+rax*2]
+        lea             rdi,        [rdi+rdx*2]
+
+        psadbw          xmm1,       xmm0
+        psadbw          xmm2,       xmm0
+        psadbw          xmm3,       xmm0
+
+        paddw           xmm5,       xmm1
+        paddw           xmm6,       xmm2
+        paddw           xmm7,       xmm3
+%endmacro
+
+%macro PROCESS_16X16X3_OFFSET 2
+%2_aligned_by_%1:
+
+        sub             rdi,        %1
+
+        PROCESS_16X2X3_OFFSET 1, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+
+        jmp             %2_store_off
+
+%endmacro
+
+%macro PROCESS_16X8X3_OFFSET 2
+%2_aligned_by_%1:
+
+        sub             rdi,        %1
+
+        PROCESS_16X2X3_OFFSET 1, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+        PROCESS_16X2X3_OFFSET 0, %1
+
+        jmp             %2_store_off
+
+%endmacro
+
+;void int aom_sad16x16x3_ssse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(aom_sad16x16x3_ssse3) PRIVATE
+sym(aom_sad16x16x3_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rcx
+    ; end prolog
+
+        mov             rsi,        arg(0) ;src_ptr
+        mov             rdi,        arg(2) ;ref_ptr
+
+        mov             rdx,        0xf
+        and             rdx,        rdi
+
+        jmp .aom_sad16x16x3_ssse3_skiptable
+.aom_sad16x16x3_ssse3_jumptable:
+        dd .aom_sad16x16x3_ssse3_aligned_by_0  - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_1  - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_2  - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_3  - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_4  - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_5  - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_6  - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_7  - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_8  - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_9  - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_10 - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_11 - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_12 - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_13 - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_14 - .aom_sad16x16x3_ssse3_do_jump
+        dd .aom_sad16x16x3_ssse3_aligned_by_15 - .aom_sad16x16x3_ssse3_do_jump
+.aom_sad16x16x3_ssse3_skiptable:
+
+        call .aom_sad16x16x3_ssse3_do_jump
+.aom_sad16x16x3_ssse3_do_jump:
+        pop             rcx                         ; get the address of do_jump
+        mov             rax,  .aom_sad16x16x3_ssse3_jumptable - .aom_sad16x16x3_ssse3_do_jump
+        add             rax,  rcx  ; get the absolute address of aom_sad16x16x3_ssse3_jumptable
+
+        movsxd          rax,  dword [rax + 4*rdx]   ; get the 32 bit offset from the jumptable
+        add             rcx,        rax
+
+        movsxd          rax,        dword ptr arg(1) ;src_stride
+        movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+        jmp             rcx
+
+        PROCESS_16X16X3_OFFSET 0,  .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 1,  .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 2,  .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 3,  .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 4,  .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 5,  .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 6,  .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 7,  .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 8,  .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 9,  .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 10, .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 11, .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 12, .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 13, .aom_sad16x16x3_ssse3
+        PROCESS_16X16X3_OFFSET 14, .aom_sad16x16x3_ssse3
+
+.aom_sad16x16x3_ssse3_aligned_by_15:
+        PROCESS_16X2X3 1
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+
+.aom_sad16x16x3_ssse3_store_off:
+        mov             rdi,        arg(4) ;Results
+
+        movq            xmm0,       xmm5
+        psrldq          xmm5,       8
+
+        paddw           xmm0,       xmm5
+        movd            [rdi],      xmm0
+;-
+        movq            xmm0,       xmm6
+        psrldq          xmm6,       8
+
+        paddw           xmm0,       xmm6
+        movd            [rdi+4],    xmm0
+;-
+        movq            xmm0,       xmm7
+        psrldq          xmm7,       8
+
+        paddw           xmm0,       xmm7
+        movd            [rdi+8],    xmm0
+
+    ; begin epilog
+    pop         rcx
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void int aom_sad16x8x3_ssse3(
+;    unsigned char *src_ptr,
+;    int  src_stride,
+;    unsigned char *ref_ptr,
+;    int  ref_stride,
+;    int  *results)
+global sym(aom_sad16x8x3_ssse3) PRIVATE
+sym(aom_sad16x8x3_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 5
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rcx
+    ; end prolog
+
+        mov             rsi,        arg(0) ;src_ptr
+        mov             rdi,        arg(2) ;ref_ptr
+
+        mov             rdx,        0xf
+        and             rdx,        rdi
+
+        jmp .aom_sad16x8x3_ssse3_skiptable
+.aom_sad16x8x3_ssse3_jumptable:
+        dd .aom_sad16x8x3_ssse3_aligned_by_0  - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_1  - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_2  - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_3  - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_4  - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_5  - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_6  - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_7  - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_8  - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_9  - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_10 - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_11 - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_12 - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_13 - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_14 - .aom_sad16x8x3_ssse3_do_jump
+        dd .aom_sad16x8x3_ssse3_aligned_by_15 - .aom_sad16x8x3_ssse3_do_jump
+.aom_sad16x8x3_ssse3_skiptable:
+
+        call .aom_sad16x8x3_ssse3_do_jump
+.aom_sad16x8x3_ssse3_do_jump:
+        pop             rcx                         ; get the address of do_jump
+        mov             rax,  .aom_sad16x8x3_ssse3_jumptable - .aom_sad16x8x3_ssse3_do_jump
+        add             rax,  rcx  ; get the absolute address of aom_sad16x8x3_ssse3_jumptable
+
+        movsxd          rax,  dword [rax + 4*rdx]   ; get the 32 bit offset from the jumptable
+        add             rcx,        rax
+
+        movsxd          rax,        dword ptr arg(1) ;src_stride
+        movsxd          rdx,        dword ptr arg(3) ;ref_stride
+
+        jmp             rcx
+
+        PROCESS_16X8X3_OFFSET 0,  .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 1,  .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 2,  .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 3,  .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 4,  .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 5,  .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 6,  .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 7,  .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 8,  .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 9,  .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 10, .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 11, .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 12, .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 13, .aom_sad16x8x3_ssse3
+        PROCESS_16X8X3_OFFSET 14, .aom_sad16x8x3_ssse3
+
+.aom_sad16x8x3_ssse3_aligned_by_15:
+
+        PROCESS_16X2X3 1
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+        PROCESS_16X2X3 0
+
+.aom_sad16x8x3_ssse3_store_off:
+        mov             rdi,        arg(4) ;Results
+
+        movq            xmm0,       xmm5
+        psrldq          xmm5,       8
+
+        paddw           xmm0,       xmm5
+        movd            [rdi],      xmm0
+;-
+        movq            xmm0,       xmm6
+        psrldq          xmm6,       8
+
+        paddw           xmm0,       xmm6
+        movd            [rdi+4],    xmm0
+;-
+        movq            xmm0,       xmm7
+        psrldq          xmm7,       8
+
+        paddw           xmm0,       xmm7
+        movd            [rdi+8],    xmm0
+
+    ; begin epilog
+    pop         rcx
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm b/third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm
new file mode 100644
index 0000000000..aa70106c84
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/ssim_opt_x86_64.asm
@@ -0,0 +1,219 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "aom_ports/x86_abi_support.asm"
+
+; tabulate_ssim - sums sum_s,sum_r,sum_sq_s,sum_sq_r, sum_sxr
+%macro TABULATE_SSIM 0
+        paddusw         xmm15, xmm3  ; sum_s
+        paddusw         xmm14, xmm4  ; sum_r
+        movdqa          xmm1, xmm3
+        pmaddwd         xmm1, xmm1
+        paddd           xmm13, xmm1 ; sum_sq_s
+        movdqa          xmm2, xmm4
+        pmaddwd         xmm2, xmm2
+        paddd           xmm12, xmm2 ; sum_sq_r
+        pmaddwd         xmm3, xmm4
+        paddd           xmm11, xmm3  ; sum_sxr
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_Q 1
+        movdqa          xmm2,%1
+        punpckldq       %1,xmm0
+        punpckhdq       xmm2,xmm0
+        paddq           %1,xmm2
+        movdqa          xmm2,%1
+        punpcklqdq      %1,xmm0
+        punpckhqdq      xmm2,xmm0
+        paddq           %1,xmm2
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_W 1
+        movdqa          xmm1, %1
+        punpcklwd       %1,xmm0
+        punpckhwd       xmm1,xmm0
+        paddd           %1, xmm1
+        SUM_ACROSS_Q    %1
+%endmacro
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    uint32_t *sum_s,
+;    uint32_t *sum_r,
+;    uint32_t *sum_sq_s,
+;    uint32_t *sum_sq_r,
+;    uint32_t *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+global sym(aom_ssim_parms_16x16_sse2) PRIVATE
+sym(aom_ssim_parms_16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 16      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movdqu          xmm5, [rsi]
+    movdqu          xmm6, [rdi]
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpckhbw       xmm3, xmm0 ; high_s
+    punpckhbw       xmm4, xmm0 ; high_r
+
+    TABULATE_SSIM
+
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    uint32_t *sum_s,
+;    uint32_t *sum_r,
+;    uint32_t *sum_sq_s,
+;    uint32_t *sum_sq_r,
+;    uint32_t *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+global sym(aom_ssim_parms_8x8_sse2) PRIVATE
+sym(aom_ssim_parms_8x8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 8      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movq            xmm3, [rsi]
+    movq            xmm4, [rdi]
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm b/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm
new file mode 100644
index 0000000000..d3feb7ec03
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm
@@ -0,0 +1,1489 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_8: times  8 dw  8
+bilin_filter_m_sse2: times  8 dw 16
+                     times  8 dw  0
+                     times  8 dw 14
+                     times  8 dw  2
+                     times  8 dw 12
+                     times  8 dw  4
+                     times  8 dw 10
+                     times  8 dw  6
+                     times 16 dw  8
+                     times  8 dw  6
+                     times  8 dw 10
+                     times  8 dw  4
+                     times  8 dw 12
+                     times  8 dw  2
+                     times  8 dw 14
+
+bilin_filter_m_ssse3: times  8 db 16,  0
+                      times  8 db 14,  2
+                      times  8 db 12,  4
+                      times  8 db 10,  6
+                      times 16 db  8
+                      times  8 db  6, 10
+                      times  8 db  4, 12
+                      times  8 db  2, 14
+
+SECTION .text
+
+; int aom_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride,
+;                               int x_offset, int y_offset,
+;                               const uint8_t *dst, ptrdiff_t dst_stride,
+;                               int height, unsigned int *sse);
+;
+; This function returns the SE and stores SSE in the given pointer.
+
+%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse
+  psubw                %3, %4
+  psubw                %1, %2
+  paddw                %5, %3
+  pmaddwd              %3, %3
+  paddw                %5, %1
+  pmaddwd              %1, %1
+  paddd                %6, %3
+  paddd                %6, %1
+%endmacro
+
+%macro STORE_AND_RET 1
+%if %1 > 4
+  ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit
+  ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg.
+  ; We have to sign-extend it before adding the words within the register
+  ; and outputing to a dword.
+  pcmpgtw              m5, m6           ; mask for 0 > x
+  movhlps              m3, m7
+  punpcklwd            m4, m6, m5
+  punpckhwd            m6, m5           ; sign-extend m6 word->dword
+  paddd                m7, m3
+  paddd                m6, m4
+  pshufd               m3, m7, 0x1
+  movhlps              m4, m6
+  paddd                m7, m3
+  paddd                m6, m4
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  pshufd               m4, m6, 0x1
+  movd               [r1], m7           ; store sse
+  paddd                m6, m4
+  movd               raxd, m6           ; store sum as return value
+%else ; 4xh
+  pshuflw              m4, m6, 0xe
+  pshuflw              m3, m7, 0xe
+  paddw                m6, m4
+  paddd                m7, m3
+  pcmpgtw              m5, m6           ; mask for 0 > x
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  punpcklwd            m6, m5           ; sign-extend m6 word->dword
+  movd               [r1], m7           ; store sse
+  pshuflw              m4, m6, 0xe
+  paddd                m6, m4
+  movd               raxd, m6           ; store sum as return value
+%endif
+  RET
+%endmacro
+
+%macro INC_SRC_BY_SRC_STRIDE  0
+%if ARCH_X86=1 && CONFIG_PIC=1
+  add                srcq, src_stridemp
+%else
+  add                srcq, src_strideq
+%endif
+%endmacro
+
+%macro SUBPEL_VARIANCE 1-2 0 ; W
+%if cpuflag(ssse3)
+%define bilin_filter_m bilin_filter_m_ssse3
+%define filter_idx_shift 4
+%else
+%define bilin_filter_m bilin_filter_m_sse2
+%define filter_idx_shift 5
+%endif
+; FIXME(rbultje) only bilinear filters use >8 registers, and ssse3 only uses
+; 11, not 13, if the registers are ordered correctly. May make a minor speed
+; difference on Win64
+
+%ifdef PIC    ; 64bit PIC
+  %if %2 == 1 ; avg
+    cglobal sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \
+                                      x_offset, y_offset, \
+                                      dst, dst_stride, \
+                                      sec, sec_stride, height, sse
+    %define sec_str sec_strideq
+  %else
+    cglobal sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, x_offset, \
+                                  y_offset, dst, dst_stride, height, sse
+  %endif
+  %define block_height heightd
+  %define bilin_filter sseq
+%else
+  %if ARCH_X86=1 && CONFIG_PIC=1
+    %if %2 == 1 ; avg
+      cglobal sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \
+                                  x_offset, y_offset, \
+                                  dst, dst_stride, \
+                                  sec, sec_stride, \
+                                  height, sse, g_bilin_filter, g_pw_8
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+
+      ;Store bilin_filter and pw_8 location in stack
+      %if GET_GOT_DEFINED == 1
+        GET_GOT eax
+        add esp, 4                ; restore esp
+      %endif
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %else
+      cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, x_offset, \
+                                y_offset, dst, dst_stride, height, sse, \
+                                g_bilin_filter, g_pw_8
+      %define block_height heightd
+
+      ;Store bilin_filter and pw_8 location in stack
+      %if GET_GOT_DEFINED == 1
+        GET_GOT eax
+        add esp, 4                ; restore esp
+      %endif
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %endif
+  %else
+    %if %2 == 1 ; avg
+      cglobal sub_pixel_avg_variance%1xh, 7 + 2 * ARCH_X86_64, \
+                        7 + 2 * ARCH_X86_64, 13, src, src_stride, \
+                                             x_offset, y_offset, \
+                                             dst, dst_stride, \
+                                             sec, sec_stride, \
+                                             height, sse
+      %if ARCH_X86_64
+      %define block_height heightd
+      %define sec_str sec_strideq
+      %else
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+      %endif
+    %else
+      cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, x_offset, \
+                              y_offset, dst, dst_stride, height, sse
+      %define block_height heightd
+    %endif
+
+    %define bilin_filter bilin_filter_m
+  %endif
+%endif
+
+%if %1 == 4
+  %define movx movd
+%else
+  %define movx movh
+%endif
+
+  ASSERT               %1 <= 16         ; m6 overflows if w > 16
+  pxor                 m6, m6           ; sum
+  pxor                 m7, m7           ; sse
+  ; FIXME(rbultje) if both filters are bilinear, we don't actually use m5; we
+  ; could perhaps use it for something more productive then
+  pxor                 m5, m5           ; dedicated zero register
+%if %1 < 16
+  sar                   block_height, 1
+%if %2 == 1 ; avg
+  shl             sec_str, 1
+%endif
+%endif
+
+  ; FIXME(rbultje) replace by jumptable?
+  test          x_offsetd, x_offsetd
+  jnz .x_nonzero
+  ; x_offset == 0
+  test          y_offsetd, y_offsetd
+  jnz .x_zero_y_nonzero
+
+  ; x_offset == 0 && y_offset == 0
+.x_zero_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  mova                 m1, [dstq]
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%endif
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+
+%if %2 == 0 ; !avg
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m0, [srcq+src_strideq]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m0, m1
+%endif
+%else ; !avg
+  movx                 m2, [srcq+src_strideq]
+%endif
+
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+
+%if %2 == 1 ; avg
+%if %1 > 4
+  pavgb                m0, [secq]
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+%endif
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%if %1 > 4
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_zero_y_zero_loop
+  STORE_AND_RET %1
+
+.x_zero_y_nonzero:
+  cmp           y_offsetd, 4
+  jne .x_zero_y_nonhalf
+
+  ; x_offset == 0 && y_offset == 0.5
+.x_zero_y_half_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+src_strideq]
+  mova                 m1, [dstq]
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+%endif
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m2, [srcq+src_strideq]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m2, [srcq+src_strideq*2]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq*2]
+  punpckldq            m2, m1
+%endif
+  movx                 m1, [dstq]
+%if %1 > 4
+  movlhps              m0, m2
+%else ; 4xh
+  punpckldq            m0, m2
+%endif
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m0, m2
+  punpcklbw            m1, m5
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpcklbw            m3, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m4, [secq]
+  pavgb                m0, m4
+  punpcklbw            m3, m5
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  movx                 m4, [srcq+src_strideq*2]
+  movx                 m1, [dstq]
+  pavgb                m0, m2
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m2, m4
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_zero_y_half_loop
+  STORE_AND_RET %1
+
+.x_zero_y_nonhalf:
+  ; x_offset == 0 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else ; x86-32 or mmx
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0, reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_zero_y_other_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+src_strideq]
+  mova                 m1, [dstq]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  punpcklbw            m0, m5
+  punpcklbw            m4, m5
+  ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can
+  ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of
+  ; instructions is the same (5), but it is 1 mul instead of 2, so might be
+  ; slightly faster because of pmullw latency. It would also cut our rodata
+  ; tables in half for this function, and save 1-2 registers on x86-64.
+  pmullw               m2, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m2, m3
+  paddw                m0, m4
+%endif
+  psraw                m2, 4
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m2, [srcq+src_strideq]
+  movx                 m4, [srcq+src_strideq*2]
+  movx                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  movx                 m1, [dstq]
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m1, m2, filter_y_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, m1
+  paddw                m2, filter_rnd
+  movx                 m1, [dstq]
+  paddw                m2, m4
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_zero_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonzero:
+  cmp           x_offsetd, 4
+  jne .x_nonhalf
+  ; x_offset == 0.5
+  test          y_offsetd, y_offsetd
+  jnz .x_half_y_nonzero
+
+  ; x_offset == 0.5 && y_offset == 0
+.x_half_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+%endif
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m4, [srcq+1]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m0, [srcq+src_strideq]
+  movhps               m4, [srcq+src_strideq+1]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m0, m1
+  movx                 m2, [srcq+src_strideq+1]
+  punpckldq            m4, m2
+%endif
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m0, m4
+  punpcklbw            m3, m5
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m1, m5
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  movx                 m2, [srcq+src_strideq]
+  movx                 m1, [dstq]
+  pavgb                m0, m4
+  movx                 m4, [srcq+src_strideq+1]
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m2, m4
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_half_y_zero_loop
+  STORE_AND_RET %1
+
+.x_half_y_nonzero:
+  cmp           y_offsetd, 4
+  jne .x_half_y_nonhalf
+
+  ; x_offset == 0.5 && y_offset == 0.5
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_half_loop:
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m4, m3
+  punpckhbw            m3, m1, m5
+  pavgb                m0, m4
+%if %2 == 1 ; avg
+  punpcklbw            m1, m5
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_half_loop:
+  movx                 m2, [srcq]
+  movx                 m3, [srcq+1]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m2, [srcq+src_strideq]
+  movhps               m3, [srcq+src_strideq+1]
+%else
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m2, m1
+  movx                 m1, [srcq+src_strideq+1]
+  punpckldq            m3, m1
+%endif
+  pavgb                m2, m3
+%if %1 > 4
+  movlhps              m0, m2
+  movhlps              m4, m2
+%else ; 4xh
+  punpckldq            m0, m2
+  pshuflw              m4, m2, 0xe
+%endif
+  movx                 m1, [dstq]
+  pavgb                m0, m2
+  movx                 m3, [dstq+dst_strideq]
+%if %1 > 4
+  pavgb                m0, [secq]
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+%endif
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%if %1 > 4
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  movx                 m4, [srcq+src_strideq]
+  movx                 m1, [srcq+src_strideq+1]
+  pavgb                m2, m3
+  pavgb                m4, m1
+  pavgb                m0, m2
+  pavgb                m2, m4
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_half_y_half_loop
+  STORE_AND_RET %1
+
+.x_half_y_nonhalf:
+  ; x_offset == 0.5 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else  ;x86_32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0.5. We can reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_other_loop:
+  movu                 m4, [srcq]
+  movu                 m2, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m4, m2
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  pmullw               m2, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, filter_rnd
+  punpcklbw            m0, m5
+  paddw                m2, m3
+  punpcklbw            m3, m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  paddw                m0, m3
+%endif
+  punpckhbw            m3, m1, m5
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+%if notcpuflag(ssse3)
+  punpcklbw            m0, m5
+%endif
+.x_half_y_other_loop:
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m4, [srcq+src_strideq]
+  movx                 m3, [srcq+src_strideq+1]
+  pavgb                m2, m1
+  pavgb                m4, m3
+  movx                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  movx                 m1, [dstq]
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m1, m2, filter_y_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  paddw                m0, m1
+  pmullw               m1, m4, filter_y_b
+  paddw                m2, filter_rnd
+  paddw                m2, m1
+  movx                 m1, [dstq]
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_half_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonhalf:
+  test          y_offsetd, y_offsetd
+  jnz .x_nonhalf_y_nonzero
+
+  ; x_offset == bilin interpolation && y_offset == 0
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+;y_offset == 0. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_other_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+1]
+  mova                 m1, [dstq]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  punpcklbw            m0, m5
+  punpcklbw            m4, m5
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m2, m3
+  paddw                m0, m4
+%endif
+  psraw                m2, 4
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m2, [srcq+src_strideq]
+  movx                 m4, [srcq+src_strideq+1]
+  movx                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  movx                 m1, [dstq]
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_x_a
+  pmaddubsw            m2, filter_x_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m0, m1
+  paddw                m2, filter_rnd
+  movx                 m1, [dstq]
+  paddw                m2, m4
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_other_y_zero_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonhalf_y_nonzero:
+  cmp           y_offsetd, 4
+  jne .x_nonhalf_y_nonhalf
+
+  ; x_offset == bilin interpolation && y_offset == 0.5
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0.5. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m1
+  punpcklbw            m0, m1
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m1, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m0, m1
+  paddw                m2, m3
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+  add                srcq, src_strideq
+  packuswb             m0, m2
+.x_other_y_half_loop:
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+%if cpuflag(ssse3)
+  mova                 m1, [dstq]
+  punpckhbw            m2, m4, m3
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m4, m2
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%else
+  punpckhbw            m2, m4, m5
+  punpckhbw            m1, m3, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m4, m3
+  paddw                m2, m1
+  mova                 m1, [dstq]
+  psraw                m4, 4
+  psraw                m2, 4
+  punpckhbw            m3, m1, m5
+  ; FIXME(rbultje) the repeated pack/unpack here around m0/m2 is because we
+  ; have a 1-register shortage to be able to store the backup of the bilin
+  ; filtered second line as words as cache for the next line. Packing into
+  ; a byte costs 1 pack and 2 unpacks, but saves a register.
+  packuswb             m4, m2
+  punpcklbw            m1, m5
+  pavgb                m0, m4
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  pavgb                m0, [secq]
+%endif
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  pmaddubsw            m0, filter_x_a
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m1
+%endif
+  add                srcq, src_strideq
+  psraw                m0, 4
+.x_other_y_half_loop:
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m4, [srcq+src_strideq]
+  movx                 m3, [srcq+src_strideq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m2, m1
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m1, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  paddw                m2, m1
+  movx                 m1, [dstq]
+  paddw                m4, m3
+  movx                 m3, [dstq+dst_strideq]
+%endif
+  psraw                m2, 4
+  psraw                m4, 4
+  pavgw                m0, m2
+  pavgw                m2, m4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline - also consider going to bytes here
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_other_y_half_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonhalf_y_nonhalf:
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                m11, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m12, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_y_a m10
+%define filter_y_b m11
+%define filter_rnd m12
+%else   ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; In this case, there is NO unused register. Used src_stride register. Later,
+; src_stride has to be loaded from stack when it is needed.
+%define tempq src_strideq
+  mov tempq, g_bilin_filterm
+  add           x_offsetq, tempq
+  add           y_offsetq, tempq
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+  add           y_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+  ; x_offset == bilin interpolation && y_offset == bilin interpolation
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m1
+  punpcklbw            m0, m1
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m1, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m0, m1
+  paddw                m2, m3
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+  packuswb             m0, m2
+.x_other_y_other_loop:
+%if cpuflag(ssse3)
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+  mova                 m1, [dstq]
+  punpckhbw            m2, m4, m3
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  punpckhbw            m3, m1, m5
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m4, m2
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  punpcklbw            m1, m5
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  psraw                m0, 4
+%else
+  movu                 m3, [srcq]
+  movu                 m4, [srcq+1]
+  punpckhbw            m1, m3, m5
+  punpckhbw            m2, m4, m5
+  punpcklbw            m3, m5
+  punpcklbw            m4, m5
+  pmullw               m3, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m3, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m3, m4
+  paddw                m1, m2
+  psraw                m3, 4
+  psraw                m1, 4
+  packuswb             m4, m3, m1
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  pmullw               m2, filter_y_a
+  pmullw               m1, filter_y_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, m1
+  mova                 m1, [dstq]
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  paddw                m0, m3
+  punpckhbw            m3, m1, m5
+  psraw                m0, 4
+  punpcklbw            m1, m5
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  INC_SRC_BY_SRC_STRIDE
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  pmaddubsw            m0, filter_x_a
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m1
+%endif
+  psraw                m0, 4
+%if cpuflag(ssse3)
+  packuswb             m0, m0
+%endif
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+
+  INC_SRC_BY_SRC_STRIDE
+  movx                 m4, [srcq]
+  movx                 m3, [srcq+1]
+
+%if cpuflag(ssse3)
+  punpcklbw            m2, m1
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  movx                 m3, [dstq+dst_strideq]
+  movx                 m1, [dstq]
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m2, m2
+  packuswb             m4, m4
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+  psraw                m0, 4
+  psraw                m2, 4
+  punpcklbw            m1, m5
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m1, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  paddw                m2, m1
+  paddw                m4, m3
+  psraw                m2, 4
+  psraw                m4, 4
+  pmullw               m0, filter_y_a
+  pmullw               m3, m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  pmullw               m1, m4, filter_y_b
+  paddw                m2, filter_rnd
+  paddw                m0, m3
+  movx                 m3, [dstq+dst_strideq]
+  paddw                m2, m1
+  movx                 m1, [dstq]
+  psraw                m0, 4
+  psraw                m2, 4
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_other_y_other_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+%undef movx
+  STORE_AND_RET %1
+%endmacro
+
+; FIXME(rbultje) the non-bilinear versions (i.e. x=0,8&&y=0,8) are identical
+; between the ssse3 and non-ssse3 version. It may make sense to merge their
+; code in the sense that the ssse3 version would jump to the appropriate
+; location in the sse/2 version, rather than duplicating that code in the
+; binary.
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  4
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_XMM ssse3
+SUBPEL_VARIANCE  4
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  4, 1
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
+
+INIT_XMM ssse3
+SUBPEL_VARIANCE  4, 1
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
diff --git a/third_party/aom/aom_dsp/x86/subtract_sse2.asm b/third_party/aom/aom_dsp/x86/subtract_sse2.asm
new file mode 100644
index 0000000000..7bd5b23ad2
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/subtract_sse2.asm
@@ -0,0 +1,150 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; void aom_subtract_block(int rows, int cols,
+;                         int16_t *diff, ptrdiff_t diff_stride,
+;                         const uint8_t *src, ptrdiff_t src_stride,
+;                         const uint8_t *pred, ptrdiff_t pred_stride)
+
+INIT_XMM sse2
+cglobal subtract_block, 7, 7, 8, \
+                        rows, cols, diff, diff_stride, src, src_stride, \
+                        pred, pred_stride
+%define pred_str colsq
+  pxor                  m7, m7         ; dedicated zero register
+  cmp                colsd, 4
+  je .case_4
+  cmp                colsd, 8
+  je .case_8
+  cmp                colsd, 16
+  je .case_16
+  cmp                colsd, 32
+  je .case_32
+%if CONFIG_EXT_PARTITION
+  cmp                colsd, 64
+  je .case_64
+%endif
+
+%macro loop16 6
+  mova                  m0, [srcq+%1]
+  mova                  m4, [srcq+%2]
+  mova                  m1, [predq+%3]
+  mova                  m5, [predq+%4]
+  punpckhbw             m2, m0, m7
+  punpckhbw             m3, m1, m7
+  punpcklbw             m0, m7
+  punpcklbw             m1, m7
+  psubw                 m2, m3
+  psubw                 m0, m1
+  punpckhbw             m1, m4, m7
+  punpckhbw             m3, m5, m7
+  punpcklbw             m4, m7
+  punpcklbw             m5, m7
+  psubw                 m1, m3
+  psubw                 m4, m5
+  mova [diffq+mmsize*0+%5], m0
+  mova [diffq+mmsize*1+%5], m2
+  mova [diffq+mmsize*0+%6], m4
+  mova [diffq+mmsize*1+%6], m1
+%endmacro
+
+%if CONFIG_EXT_PARTITION
+  mov             pred_str, pred_stridemp
+.loop_128:
+  loop16 0*mmsize, 1*mmsize, 0*mmsize, 1*mmsize,  0*mmsize,  2*mmsize
+  loop16 2*mmsize, 3*mmsize, 2*mmsize, 3*mmsize,  4*mmsize,  6*mmsize
+  loop16 4*mmsize, 5*mmsize, 4*mmsize, 5*mmsize,  8*mmsize, 10*mmsize
+  loop16 6*mmsize, 7*mmsize, 6*mmsize, 7*mmsize, 12*mmsize, 14*mmsize
+  lea                diffq, [diffq+diff_strideq*2]
+  add                predq, pred_str
+  add                 srcq, src_strideq
+  sub                rowsd, 1
+  jnz .loop_128
+  RET
+
+.case_64:
+%endif
+  mov             pred_str, pred_stridemp
+.loop_64:
+  loop16 0*mmsize, 1*mmsize, 0*mmsize, 1*mmsize, 0*mmsize, 2*mmsize
+  loop16 2*mmsize, 3*mmsize, 2*mmsize, 3*mmsize, 4*mmsize, 6*mmsize
+  lea                diffq, [diffq+diff_strideq*2]
+  add                predq, pred_str
+  add                 srcq, src_strideq
+  dec                rowsd
+  jg .loop_64
+  RET
+
+.case_32:
+  mov             pred_str, pred_stridemp
+.loop_32:
+  loop16 0, mmsize, 0, mmsize, 0, 2*mmsize
+  lea                diffq, [diffq+diff_strideq*2]
+  add                predq, pred_str
+  add                 srcq, src_strideq
+  dec                rowsd
+  jg .loop_32
+  RET
+
+.case_16:
+  mov             pred_str, pred_stridemp
+.loop_16:
+  loop16 0, src_strideq, 0, pred_str, 0, diff_strideq*2
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                predq, [predq+pred_str*2]
+  lea                 srcq, [srcq+src_strideq*2]
+  sub                rowsd, 2
+  jg .loop_16
+  RET
+
+%macro loop_h 0
+  movh                  m0, [srcq]
+  movh                  m2, [srcq+src_strideq]
+  movh                  m1, [predq]
+  movh                  m3, [predq+pred_str]
+  punpcklbw             m0, m7
+  punpcklbw             m1, m7
+  punpcklbw             m2, m7
+  punpcklbw             m3, m7
+  psubw                 m0, m1
+  psubw                 m2, m3
+  mova             [diffq], m0
+  mova [diffq+diff_strideq*2], m2
+%endmacro
+
+.case_8:
+  mov             pred_str, pred_stridemp
+.loop_8:
+  loop_h
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                 srcq, [srcq+src_strideq*2]
+  lea                predq, [predq+pred_str*2]
+  sub                rowsd, 2
+  jg .loop_8
+  RET
+
+INIT_MMX
+.case_4:
+  mov             pred_str, pred_stridemp
+.loop_4:
+  loop_h
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                 srcq, [srcq+src_strideq*2]
+  lea                predq, [predq+pred_str*2]
+  sub                rowsd, 2
+  jg .loop_4
+  RET
diff --git a/third_party/aom/aom_dsp/x86/sum_squares_sse2.c b/third_party/aom/aom_dsp/x86/sum_squares_sse2.c
new file mode 100644
index 0000000000..6be99fbcac
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sum_squares_sse2.c
@@ -0,0 +1,210 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>
+#include <stdio.h>
+
+#include "aom_dsp/x86/synonyms.h"
+
+#include "./aom_dsp_rtcd.h"
+
+static uint64_t aom_sum_squares_2d_i16_4x4_sse2(const int16_t *src,
+                                                int stride) {
+  const __m128i v_val_0_w =
+      _mm_loadl_epi64((const __m128i *)(src + 0 * stride));
+  const __m128i v_val_1_w =
+      _mm_loadl_epi64((const __m128i *)(src + 1 * stride));
+  const __m128i v_val_2_w =
+      _mm_loadl_epi64((const __m128i *)(src + 2 * stride));
+  const __m128i v_val_3_w =
+      _mm_loadl_epi64((const __m128i *)(src + 3 * stride));
+
+  const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
+  const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
+  const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
+  const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
+
+  const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
+  const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
+  const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
+
+  const __m128i v_sum_d =
+      _mm_add_epi32(v_sum_0123_d, _mm_srli_epi64(v_sum_0123_d, 32));
+
+  return (uint64_t)_mm_cvtsi128_si32(v_sum_d);
+}
+
+#ifdef __GNUC__
+// This prevents GCC/Clang from inlining this function into
+// aom_sum_squares_2d_i16_sse2, which in turn saves some stack
+// maintenance instructions in the common case of 4x4.
+__attribute__((noinline))
+#endif
+static uint64_t
+aom_sum_squares_2d_i16_nxn_sse2(const int16_t *src, int stride, int width,
+                                int height) {
+  int r, c;
+
+  const __m128i v_zext_mask_q = _mm_set_epi32(0, 0xffffffff, 0, 0xffffffff);
+  __m128i v_acc_q = _mm_setzero_si128();
+
+  for (r = 0; r < height; r += 8) {
+    __m128i v_acc_d = _mm_setzero_si128();
+
+    for (c = 0; c < width; c += 8) {
+      const int16_t *b = src + c;
+
+      const __m128i v_val_0_w =
+          _mm_load_si128((const __m128i *)(b + 0 * stride));
+      const __m128i v_val_1_w =
+          _mm_load_si128((const __m128i *)(b + 1 * stride));
+      const __m128i v_val_2_w =
+          _mm_load_si128((const __m128i *)(b + 2 * stride));
+      const __m128i v_val_3_w =
+          _mm_load_si128((const __m128i *)(b + 3 * stride));
+      const __m128i v_val_4_w =
+          _mm_load_si128((const __m128i *)(b + 4 * stride));
+      const __m128i v_val_5_w =
+          _mm_load_si128((const __m128i *)(b + 5 * stride));
+      const __m128i v_val_6_w =
+          _mm_load_si128((const __m128i *)(b + 6 * stride));
+      const __m128i v_val_7_w =
+          _mm_load_si128((const __m128i *)(b + 7 * stride));
+
+      const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
+      const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
+      const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
+      const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
+      const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w);
+      const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w);
+      const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w);
+      const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w);
+
+      const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
+      const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
+      const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d);
+      const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d);
+
+      const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
+      const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d);
+
+      v_acc_d = _mm_add_epi32(v_acc_d, v_sum_0123_d);
+      v_acc_d = _mm_add_epi32(v_acc_d, v_sum_4567_d);
+    }
+
+    v_acc_q = _mm_add_epi64(v_acc_q, _mm_and_si128(v_acc_d, v_zext_mask_q));
+    v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_epi64(v_acc_d, 32));
+
+    src += 8 * stride;
+  }
+
+  v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_si128(v_acc_q, 8));
+
+#if ARCH_X86_64
+  return (uint64_t)_mm_cvtsi128_si64(v_acc_q);
+#else
+  {
+    uint64_t tmp;
+    _mm_storel_epi64((__m128i *)&tmp, v_acc_q);
+    return tmp;
+  }
+#endif
+}
+
+uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width,
+                                     int height) {
+  // 4 elements per row only requires half an XMM register, so this
+  // must be a special case, but also note that over 75% of all calls
+  // are with size == 4, so it is also the common case.
+  if (LIKELY(width == 4 && height == 4)) {
+    return aom_sum_squares_2d_i16_4x4_sse2(src, stride);
+  } else if (LIKELY(width % 8 == 0 && height % 8 == 0)) {
+    // Generic case
+    return aom_sum_squares_2d_i16_nxn_sse2(src, stride, width, height);
+  } else {
+    return aom_sum_squares_2d_i16_c(src, stride, width, height);
+  }
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// 1D version
+//////////////////////////////////////////////////////////////////////////////
+
+static uint64_t aom_sum_squares_i16_64n_sse2(const int16_t *src, uint32_t n) {
+  const __m128i v_zext_mask_q = _mm_set_epi32(0, 0xffffffff, 0, 0xffffffff);
+  __m128i v_acc0_q = _mm_setzero_si128();
+  __m128i v_acc1_q = _mm_setzero_si128();
+
+  const int16_t *const end = src + n;
+
+  assert(n % 64 == 0);
+
+  while (src < end) {
+    const __m128i v_val_0_w = xx_load_128(src);
+    const __m128i v_val_1_w = xx_load_128(src + 8);
+    const __m128i v_val_2_w = xx_load_128(src + 16);
+    const __m128i v_val_3_w = xx_load_128(src + 24);
+    const __m128i v_val_4_w = xx_load_128(src + 32);
+    const __m128i v_val_5_w = xx_load_128(src + 40);
+    const __m128i v_val_6_w = xx_load_128(src + 48);
+    const __m128i v_val_7_w = xx_load_128(src + 56);
+
+    const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
+    const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
+    const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
+    const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
+    const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w);
+    const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w);
+    const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w);
+    const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w);
+
+    const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
+    const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
+    const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d);
+    const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d);
+
+    const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
+    const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d);
+
+    const __m128i v_sum_d = _mm_add_epi32(v_sum_0123_d, v_sum_4567_d);
+
+    v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_and_si128(v_sum_d, v_zext_mask_q));
+    v_acc1_q = _mm_add_epi64(v_acc1_q, _mm_srli_epi64(v_sum_d, 32));
+
+    src += 64;
+  }
+
+  v_acc0_q = _mm_add_epi64(v_acc0_q, v_acc1_q);
+  v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_srli_si128(v_acc0_q, 8));
+
+#if ARCH_X86_64
+  return (uint64_t)_mm_cvtsi128_si64(v_acc0_q);
+#else
+  {
+    uint64_t tmp;
+    _mm_storel_epi64((__m128i *)&tmp, v_acc0_q);
+    return tmp;
+  }
+#endif
+}
+
+uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t n) {
+  if (n % 64 == 0) {
+    return aom_sum_squares_i16_64n_sse2(src, n);
+  } else if (n > 64) {
+    int k = n & ~(64 - 1);
+    return aom_sum_squares_i16_64n_sse2(src, k) +
+           aom_sum_squares_i16_c(src + k, n - k);
+  } else {
+    return aom_sum_squares_i16_c(src, n);
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/synonyms.h b/third_party/aom/aom_dsp/x86/synonyms.h
new file mode 100644
index 0000000000..bef606dae7
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/synonyms.h
@@ -0,0 +1,120 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_X86_SYNONYMS_H_
+#define AOM_DSP_X86_SYNONYMS_H_
+
+#include <immintrin.h>
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+/**
+ * Various reusable shorthands for x86 SIMD intrinsics.
+ *
+ * Intrinsics prefixed with xx_ operate on or return 128bit XMM registers.
+ * Intrinsics prefixed with yy_ operate on or return 256bit YMM registers.
+ */
+
+// Loads and stores to do away with the tedium of casting the address
+// to the right type.
+static INLINE __m128i xx_loadl_32(const void *a) {
+  return _mm_cvtsi32_si128(*(const uint32_t *)a);
+}
+
+static INLINE __m128i xx_loadl_64(const void *a) {
+  return _mm_loadl_epi64((const __m128i *)a);
+}
+
+static INLINE __m128i xx_load_128(const void *a) {
+  return _mm_load_si128((const __m128i *)a);
+}
+
+static INLINE __m128i xx_loadu_128(const void *a) {
+  return _mm_loadu_si128((const __m128i *)a);
+}
+
+static INLINE void xx_storel_32(void *const a, const __m128i v) {
+  *(uint32_t *)a = _mm_cvtsi128_si32(v);
+}
+
+static INLINE void xx_storel_64(void *const a, const __m128i v) {
+  _mm_storel_epi64((__m128i *)a, v);
+}
+
+static INLINE void xx_store_128(void *const a, const __m128i v) {
+  _mm_store_si128((__m128i *)a, v);
+}
+
+static INLINE void xx_storeu_128(void *const a, const __m128i v) {
+  _mm_storeu_si128((__m128i *)a, v);
+}
+
+static INLINE __m128i xx_round_epu16(__m128i v_val_w) {
+  return _mm_avg_epu16(v_val_w, _mm_setzero_si128());
+}
+
+static INLINE __m128i xx_roundn_epu16(__m128i v_val_w, int bits) {
+  const __m128i v_s_w = _mm_srli_epi16(v_val_w, bits - 1);
+  return _mm_avg_epu16(v_s_w, _mm_setzero_si128());
+}
+
+static INLINE __m128i xx_roundn_epu32(__m128i v_val_d, int bits) {
+  const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1);
+  const __m128i v_tmp_d = _mm_add_epi32(v_val_d, v_bias_d);
+  return _mm_srli_epi32(v_tmp_d, bits);
+}
+
+// This is equivalent to ROUND_POWER_OF_TWO(v_val_d, bits)
+static INLINE __m128i xx_roundn_epi32_unsigned(__m128i v_val_d, int bits) {
+  const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1);
+  const __m128i v_tmp_d = _mm_add_epi32(v_val_d, v_bias_d);
+  return _mm_srai_epi32(v_tmp_d, bits);
+}
+
+// This is equivalent to ROUND_POWER_OF_TWO_SIGNED(v_val_d, bits)
+static INLINE __m128i xx_roundn_epi32(__m128i v_val_d, int bits) {
+  const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1);
+  const __m128i v_sign_d = _mm_srai_epi32(v_val_d, 31);
+  const __m128i v_tmp_d =
+      _mm_add_epi32(_mm_add_epi32(v_val_d, v_bias_d), v_sign_d);
+  return _mm_srai_epi32(v_tmp_d, bits);
+}
+
+#ifdef __SSSE3__
+static INLINE int32_t xx_hsum_epi32_si32(__m128i v_d) {
+  v_d = _mm_hadd_epi32(v_d, v_d);
+  v_d = _mm_hadd_epi32(v_d, v_d);
+  return _mm_cvtsi128_si32(v_d);
+}
+
+static INLINE int64_t xx_hsum_epi64_si64(__m128i v_q) {
+  v_q = _mm_add_epi64(v_q, _mm_srli_si128(v_q, 8));
+#if ARCH_X86_64
+  return _mm_cvtsi128_si64(v_q);
+#else
+  {
+    int64_t tmp;
+    _mm_storel_epi64((__m128i *)&tmp, v_q);
+    return tmp;
+  }
+#endif
+}
+
+static INLINE int64_t xx_hsum_epi32_si64(__m128i v_d) {
+  const __m128i v_sign_d = _mm_cmplt_epi32(v_d, _mm_setzero_si128());
+  const __m128i v_0_q = _mm_unpacklo_epi32(v_d, v_sign_d);
+  const __m128i v_1_q = _mm_unpackhi_epi32(v_d, v_sign_d);
+  return xx_hsum_epi64_si64(_mm_add_epi64(v_0_q, v_1_q));
+}
+#endif  // __SSSE3__
+
+#endif  // AOM_DSP_X86_SYNONYMS_H_
diff --git a/third_party/aom/aom_dsp/x86/txfm_common_avx2.h b/third_party/aom/aom_dsp/x86/txfm_common_avx2.h
new file mode 100644
index 0000000000..39e9b8e2ad
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/txfm_common_avx2.h
@@ -0,0 +1,204 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_X86_TXFM_COMMON_AVX2_H
+#define AOM_DSP_X86_TXFM_COMMON_AVX2_H
+
+#include <immintrin.h>
+
+#include "aom_dsp/txfm_common.h"
+
+#define pair256_set_epi16(a, b)                                            \
+  _mm256_set_epi16((int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                   (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                   (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                   (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a))
+
+#define pair256_set_epi32(a, b)                                                \
+  _mm256_set_epi32((int)(b), (int)(a), (int)(b), (int)(a), (int)(b), (int)(a), \
+                   (int)(b), (int)(a))
+
+static INLINE void mm256_reverse_epi16(__m256i *u) {
+  const __m256i control = _mm256_set_epi16(
+      0x0100, 0x0302, 0x0504, 0x0706, 0x0908, 0x0B0A, 0x0D0C, 0x0F0E, 0x0100,
+      0x0302, 0x0504, 0x0706, 0x0908, 0x0B0A, 0x0D0C, 0x0F0E);
+  __m256i v = _mm256_shuffle_epi8(*u, control);
+  *u = _mm256_permute2x128_si256(v, v, 1);
+}
+
+static INLINE void mm256_transpose_16x16(__m256i *in) {
+  __m256i tr0_0 = _mm256_unpacklo_epi16(in[0], in[1]);
+  __m256i tr0_1 = _mm256_unpackhi_epi16(in[0], in[1]);
+  __m256i tr0_2 = _mm256_unpacklo_epi16(in[2], in[3]);
+  __m256i tr0_3 = _mm256_unpackhi_epi16(in[2], in[3]);
+  __m256i tr0_4 = _mm256_unpacklo_epi16(in[4], in[5]);
+  __m256i tr0_5 = _mm256_unpackhi_epi16(in[4], in[5]);
+  __m256i tr0_6 = _mm256_unpacklo_epi16(in[6], in[7]);
+  __m256i tr0_7 = _mm256_unpackhi_epi16(in[6], in[7]);
+
+  __m256i tr0_8 = _mm256_unpacklo_epi16(in[8], in[9]);
+  __m256i tr0_9 = _mm256_unpackhi_epi16(in[8], in[9]);
+  __m256i tr0_a = _mm256_unpacklo_epi16(in[10], in[11]);
+  __m256i tr0_b = _mm256_unpackhi_epi16(in[10], in[11]);
+  __m256i tr0_c = _mm256_unpacklo_epi16(in[12], in[13]);
+  __m256i tr0_d = _mm256_unpackhi_epi16(in[12], in[13]);
+  __m256i tr0_e = _mm256_unpacklo_epi16(in[14], in[15]);
+  __m256i tr0_f = _mm256_unpackhi_epi16(in[14], in[15]);
+
+  // 00 10 01 11 02 12 03 13  08 18 09 19 0a 1a 0b 1b
+  // 04 14 05 15 06 16 07 17  0c 1c 0d 1d 0e 1e 0f 1f
+  // 20 30 21 31 22 32 23 33  28 38 29 39 2a 3a 2b 3b
+  // 24 34 25 35 26 36 27 37  2c 3c 2d 3d 2e 3e 2f 3f
+  // 40 50 41 51 42 52 43 53  48 58 49 59 4a 5a 4b 5b
+  // 44 54 45 55 46 56 47 57  4c 5c 4d 5d 4e 5e 4f 5f
+  // 60 70 61 71 62 72 63 73  68 78 69 79 6a 7a 6b 7b
+  // 64 74 65 75 66 76 67 77  6c 7c 6d 7d 6e 7e 6f 7f
+
+  // 80 90 81 91 82 92 83 93  88 98 89 99 8a 9a 8b 9b
+  // 84 94 85 95 86 96 87 97  8c 9c 8d 9d 8e 9e 8f 9f
+  // a0 b0 a1 b1 a2 b2 a3 b3  a8 b8 a9 b9 aa ba ab bb
+  // a4 b4 a5 b5 a6 b6 a7 b7  ac bc ad bd ae be af bf
+  // c0 d0 c1 d1 c2 d2 c3 d3  c8 d8 c9 d9 ca da cb db
+  // c4 d4 c5 d5 c6 d6 c7 d7  cc dc cd dd ce de cf df
+  // e0 f0 e1 f1 e2 f2 e3 f3  e8 f8 e9 f9 ea fa eb fb
+  // e4 f4 e5 f5 e6 f6 e7 f7  ec fc ed fd ee fe ef ff
+
+  __m256i tr1_0 = _mm256_unpacklo_epi32(tr0_0, tr0_2);
+  __m256i tr1_1 = _mm256_unpackhi_epi32(tr0_0, tr0_2);
+  __m256i tr1_2 = _mm256_unpacklo_epi32(tr0_1, tr0_3);
+  __m256i tr1_3 = _mm256_unpackhi_epi32(tr0_1, tr0_3);
+  __m256i tr1_4 = _mm256_unpacklo_epi32(tr0_4, tr0_6);
+  __m256i tr1_5 = _mm256_unpackhi_epi32(tr0_4, tr0_6);
+  __m256i tr1_6 = _mm256_unpacklo_epi32(tr0_5, tr0_7);
+  __m256i tr1_7 = _mm256_unpackhi_epi32(tr0_5, tr0_7);
+
+  __m256i tr1_8 = _mm256_unpacklo_epi32(tr0_8, tr0_a);
+  __m256i tr1_9 = _mm256_unpackhi_epi32(tr0_8, tr0_a);
+  __m256i tr1_a = _mm256_unpacklo_epi32(tr0_9, tr0_b);
+  __m256i tr1_b = _mm256_unpackhi_epi32(tr0_9, tr0_b);
+  __m256i tr1_c = _mm256_unpacklo_epi32(tr0_c, tr0_e);
+  __m256i tr1_d = _mm256_unpackhi_epi32(tr0_c, tr0_e);
+  __m256i tr1_e = _mm256_unpacklo_epi32(tr0_d, tr0_f);
+  __m256i tr1_f = _mm256_unpackhi_epi32(tr0_d, tr0_f);
+
+  // 00 10 20 30 01 11 21 31  08 18 28 38 09 19 29 39
+  // 02 12 22 32 03 13 23 33  0a 1a 2a 3a 0b 1b 2b 3b
+  // 04 14 24 34 05 15 25 35  0c 1c 2c 3c 0d 1d 2d 3d
+  // 06 16 26 36 07 17 27 37  0e 1e 2e 3e 0f 1f 2f 3f
+  // 40 50 60 70 41 51 61 71  48 58 68 78 49 59 69 79
+  // 42 52 62 72 43 53 63 73  4a 5a 6a 7a 4b 5b 6b 7b
+  // 44 54 64 74 45 55 65 75  4c 5c 6c 7c 4d 5d 6d 7d
+  // 46 56 66 76 47 57 67 77  4e 5e 6e 7e 4f 5f 6f 7f
+
+  // 80 90 a0 b0 81 91 a1 b1  88 98 a8 b8 89 99 a9 b9
+  // 82 92 a2 b2 83 93 a3 b3  8a 9a aa ba 8b 9b ab bb
+  // 84 94 a4 b4 85 95 a5 b5  8c 9c ac bc 8d 9d ad bd
+  // 86 96 a6 b6 87 97 a7 b7  8e ae 9e be 8f 9f af bf
+  // c0 d0 e0 f0 c1 d1 e1 f1  c8 d8 e8 f8 c9 d9 e9 f9
+  // c2 d2 e2 f2 c3 d3 e3 f3  ca da ea fa cb db eb fb
+  // c4 d4 e4 f4 c5 d5 e5 f5  cc dc ef fc cd dd ed fd
+  // c6 d6 e6 f6 c7 d7 e7 f7  ce de ee fe cf df ef ff
+
+  tr0_0 = _mm256_unpacklo_epi64(tr1_0, tr1_4);
+  tr0_1 = _mm256_unpackhi_epi64(tr1_0, tr1_4);
+  tr0_2 = _mm256_unpacklo_epi64(tr1_1, tr1_5);
+  tr0_3 = _mm256_unpackhi_epi64(tr1_1, tr1_5);
+  tr0_4 = _mm256_unpacklo_epi64(tr1_2, tr1_6);
+  tr0_5 = _mm256_unpackhi_epi64(tr1_2, tr1_6);
+  tr0_6 = _mm256_unpacklo_epi64(tr1_3, tr1_7);
+  tr0_7 = _mm256_unpackhi_epi64(tr1_3, tr1_7);
+
+  tr0_8 = _mm256_unpacklo_epi64(tr1_8, tr1_c);
+  tr0_9 = _mm256_unpackhi_epi64(tr1_8, tr1_c);
+  tr0_a = _mm256_unpacklo_epi64(tr1_9, tr1_d);
+  tr0_b = _mm256_unpackhi_epi64(tr1_9, tr1_d);
+  tr0_c = _mm256_unpacklo_epi64(tr1_a, tr1_e);
+  tr0_d = _mm256_unpackhi_epi64(tr1_a, tr1_e);
+  tr0_e = _mm256_unpacklo_epi64(tr1_b, tr1_f);
+  tr0_f = _mm256_unpackhi_epi64(tr1_b, tr1_f);
+
+  // 00 10 20 30 40 50 60 70  08 18 28 38 48 58 68 78
+  // 01 11 21 31 41 51 61 71  09 19 29 39 49 59 69 79
+  // 02 12 22 32 42 52 62 72  0a 1a 2a 3a 4a 5a 6a 7a
+  // 03 13 23 33 43 53 63 73  0b 1b 2b 3b 4b 5b 6b 7b
+  // 04 14 24 34 44 54 64 74  0c 1c 2c 3c 4c 5c 6c 7c
+  // 05 15 25 35 45 55 65 75  0d 1d 2d 3d 4d 5d 6d 7d
+  // 06 16 26 36 46 56 66 76  0e 1e 2e 3e 4e 5e 6e 7e
+  // 07 17 27 37 47 57 67 77  0f 1f 2f 3f 4f 5f 6f 7f
+
+  // 80 90 a0 b0 c0 d0 e0 f0  88 98 a8 b8 c8 d8 e8 f8
+  // 81 91 a1 b1 c1 d1 e1 f1  89 99 a9 b9 c9 d9 e9 f9
+  // 82 92 a2 b2 c2 d2 e2 f2  8a 9a aa ba ca da ea fa
+  // 83 93 a3 b3 c3 d3 e3 f3  8b 9b ab bb cb db eb fb
+  // 84 94 a4 b4 c4 d4 e4 f4  8c 9c ac bc cc dc ef fc
+  // 85 95 a5 b5 c5 d5 e5 f5  8d 9d ad bd cd dd ed fd
+  // 86 96 a6 b6 c6 d6 e6 f6  8e ae 9e be ce de ee fe
+  // 87 97 a7 b7 c7 d7 e7 f7  8f 9f af bf cf df ef ff
+
+  in[0] = _mm256_permute2x128_si256(tr0_0, tr0_8, 0x20);  // 0010 0000
+  in[8] = _mm256_permute2x128_si256(tr0_0, tr0_8, 0x31);  // 0011 0001
+  in[1] = _mm256_permute2x128_si256(tr0_1, tr0_9, 0x20);
+  in[9] = _mm256_permute2x128_si256(tr0_1, tr0_9, 0x31);
+  in[2] = _mm256_permute2x128_si256(tr0_2, tr0_a, 0x20);
+  in[10] = _mm256_permute2x128_si256(tr0_2, tr0_a, 0x31);
+  in[3] = _mm256_permute2x128_si256(tr0_3, tr0_b, 0x20);
+  in[11] = _mm256_permute2x128_si256(tr0_3, tr0_b, 0x31);
+
+  in[4] = _mm256_permute2x128_si256(tr0_4, tr0_c, 0x20);
+  in[12] = _mm256_permute2x128_si256(tr0_4, tr0_c, 0x31);
+  in[5] = _mm256_permute2x128_si256(tr0_5, tr0_d, 0x20);
+  in[13] = _mm256_permute2x128_si256(tr0_5, tr0_d, 0x31);
+  in[6] = _mm256_permute2x128_si256(tr0_6, tr0_e, 0x20);
+  in[14] = _mm256_permute2x128_si256(tr0_6, tr0_e, 0x31);
+  in[7] = _mm256_permute2x128_si256(tr0_7, tr0_f, 0x20);
+  in[15] = _mm256_permute2x128_si256(tr0_7, tr0_f, 0x31);
+}
+
+static INLINE __m256i butter_fly(__m256i a0, __m256i a1, const __m256i cospi) {
+  const __m256i dct_rounding = _mm256_set1_epi32(DCT_CONST_ROUNDING);
+  __m256i y0 = _mm256_madd_epi16(a0, cospi);
+  __m256i y1 = _mm256_madd_epi16(a1, cospi);
+
+  y0 = _mm256_add_epi32(y0, dct_rounding);
+  y1 = _mm256_add_epi32(y1, dct_rounding);
+  y0 = _mm256_srai_epi32(y0, DCT_CONST_BITS);
+  y1 = _mm256_srai_epi32(y1, DCT_CONST_BITS);
+
+  return _mm256_packs_epi32(y0, y1);
+}
+
+static INLINE void txfm_scaling16_avx2(const int16_t c, __m256i *in) {
+  const __m256i zero = _mm256_setzero_si256();
+  const __m256i sqrt2_epi16 = _mm256_set1_epi16(c);
+  const __m256i dct_const_rounding = _mm256_set1_epi32(DCT_CONST_ROUNDING);
+  __m256i u0, u1;
+  int i = 0;
+
+  while (i < 16) {
+    in[i] = _mm256_slli_epi16(in[i], 1);
+
+    u0 = _mm256_unpacklo_epi16(zero, in[i]);
+    u1 = _mm256_unpackhi_epi16(zero, in[i]);
+
+    u0 = _mm256_madd_epi16(u0, sqrt2_epi16);
+    u1 = _mm256_madd_epi16(u1, sqrt2_epi16);
+
+    u0 = _mm256_add_epi32(u0, dct_const_rounding);
+    u1 = _mm256_add_epi32(u1, dct_const_rounding);
+
+    u0 = _mm256_srai_epi32(u0, DCT_CONST_BITS);
+    u1 = _mm256_srai_epi32(u1, DCT_CONST_BITS);
+    in[i] = _mm256_packs_epi32(u0, u1);
+    i++;
+  }
+}
+
+#endif  // AOM_DSP_X86_TXFM_COMMON_AVX2_H
diff --git a/third_party/aom/aom_dsp/x86/txfm_common_intrin.h b/third_party/aom/aom_dsp/x86/txfm_common_intrin.h
new file mode 100644
index 0000000000..e4ac56339c
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/txfm_common_intrin.h
@@ -0,0 +1,31 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _AOM_DSP_X86_TXFM_COMMON_INTRIN_H_
+#define _AOM_DSP_X86_TXFM_COMMON_INTRIN_H_
+
+// Note:
+//  This header file should be put below any x86 intrinsics head file
+
+static INLINE void storeu_output(const __m128i *poutput, tran_low_t *dst_ptr) {
+#if CONFIG_HIGHBITDEPTH
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  _mm_storeu_si128((__m128i *)(dst_ptr), out0);
+  _mm_storeu_si128((__m128i *)(dst_ptr + 4), out1);
+#else
+  _mm_storeu_si128((__m128i *)(dst_ptr), *poutput);
+#endif  // CONFIG_HIGHBITDEPTH
+}
+
+#endif  // _AOM_DSP_X86_TXFM_COMMON_INTRIN_H_
diff --git a/third_party/aom/aom_dsp/x86/txfm_common_sse2.h b/third_party/aom/aom_dsp/x86/txfm_common_sse2.h
new file mode 100644
index 0000000000..4257d8b9ca
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/txfm_common_sse2.h
@@ -0,0 +1,326 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_DSP_X86_TXFM_COMMON_SSE2_H_
+#define AOM_DSP_X86_TXFM_COMMON_SSE2_H_
+
+#include <emmintrin.h>
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/synonyms.h"
+
+#define pair_set_epi16(a, b)                                            \
+  _mm_set_epi16((int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a))
+
+#define dual_set_epi16(a, b)                                            \
+  _mm_set_epi16((int16_t)(b), (int16_t)(b), (int16_t)(b), (int16_t)(b), \
+                (int16_t)(a), (int16_t)(a), (int16_t)(a), (int16_t)(a))
+
+#define octa_set_epi16(a, b, c, d, e, f, g, h)                           \
+  _mm_setr_epi16((int16_t)(a), (int16_t)(b), (int16_t)(c), (int16_t)(d), \
+                 (int16_t)(e), (int16_t)(f), (int16_t)(g), (int16_t)(h))
+
+// Reverse the 8 16 bit words in __m128i
+static INLINE __m128i mm_reverse_epi16(const __m128i x) {
+  const __m128i a = _mm_shufflelo_epi16(x, 0x1b);
+  const __m128i b = _mm_shufflehi_epi16(a, 0x1b);
+  return _mm_shuffle_epi32(b, 0x4e);
+}
+
+#if CONFIG_EXT_TX
+// Identity transform (both forward and inverse).
+static INLINE void idtx16_8col(__m128i *in) {
+  const __m128i k__zero_epi16 = _mm_set1_epi16((int16_t)0);
+  const __m128i k__sqrt2_epi16 = _mm_set1_epi16((int16_t)Sqrt2);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i y0, y1, y2, y3, y4, y5, y6, y7;
+
+  in[0] = _mm_slli_epi16(in[0], 1);
+  in[1] = _mm_slli_epi16(in[1], 1);
+  in[2] = _mm_slli_epi16(in[2], 1);
+  in[3] = _mm_slli_epi16(in[3], 1);
+  in[4] = _mm_slli_epi16(in[4], 1);
+  in[5] = _mm_slli_epi16(in[5], 1);
+  in[6] = _mm_slli_epi16(in[6], 1);
+  in[7] = _mm_slli_epi16(in[7], 1);
+  in[8] = _mm_slli_epi16(in[8], 1);
+  in[9] = _mm_slli_epi16(in[9], 1);
+  in[10] = _mm_slli_epi16(in[10], 1);
+  in[11] = _mm_slli_epi16(in[11], 1);
+  in[12] = _mm_slli_epi16(in[12], 1);
+  in[13] = _mm_slli_epi16(in[13], 1);
+  in[14] = _mm_slli_epi16(in[14], 1);
+  in[15] = _mm_slli_epi16(in[15], 1);
+
+  v0 = _mm_unpacklo_epi16(in[0], k__zero_epi16);
+  v1 = _mm_unpacklo_epi16(in[1], k__zero_epi16);
+  v2 = _mm_unpacklo_epi16(in[2], k__zero_epi16);
+  v3 = _mm_unpacklo_epi16(in[3], k__zero_epi16);
+  v4 = _mm_unpacklo_epi16(in[4], k__zero_epi16);
+  v5 = _mm_unpacklo_epi16(in[5], k__zero_epi16);
+  v6 = _mm_unpacklo_epi16(in[6], k__zero_epi16);
+  v7 = _mm_unpacklo_epi16(in[7], k__zero_epi16);
+
+  u0 = _mm_unpacklo_epi16(in[8], k__zero_epi16);
+  u1 = _mm_unpacklo_epi16(in[9], k__zero_epi16);
+  u2 = _mm_unpacklo_epi16(in[10], k__zero_epi16);
+  u3 = _mm_unpacklo_epi16(in[11], k__zero_epi16);
+  u4 = _mm_unpacklo_epi16(in[12], k__zero_epi16);
+  u5 = _mm_unpacklo_epi16(in[13], k__zero_epi16);
+  u6 = _mm_unpacklo_epi16(in[14], k__zero_epi16);
+  u7 = _mm_unpacklo_epi16(in[15], k__zero_epi16);
+
+  x0 = _mm_unpackhi_epi16(in[0], k__zero_epi16);
+  x1 = _mm_unpackhi_epi16(in[1], k__zero_epi16);
+  x2 = _mm_unpackhi_epi16(in[2], k__zero_epi16);
+  x3 = _mm_unpackhi_epi16(in[3], k__zero_epi16);
+  x4 = _mm_unpackhi_epi16(in[4], k__zero_epi16);
+  x5 = _mm_unpackhi_epi16(in[5], k__zero_epi16);
+  x6 = _mm_unpackhi_epi16(in[6], k__zero_epi16);
+  x7 = _mm_unpackhi_epi16(in[7], k__zero_epi16);
+
+  y0 = _mm_unpackhi_epi16(in[8], k__zero_epi16);
+  y1 = _mm_unpackhi_epi16(in[9], k__zero_epi16);
+  y2 = _mm_unpackhi_epi16(in[10], k__zero_epi16);
+  y3 = _mm_unpackhi_epi16(in[11], k__zero_epi16);
+  y4 = _mm_unpackhi_epi16(in[12], k__zero_epi16);
+  y5 = _mm_unpackhi_epi16(in[13], k__zero_epi16);
+  y6 = _mm_unpackhi_epi16(in[14], k__zero_epi16);
+  y7 = _mm_unpackhi_epi16(in[15], k__zero_epi16);
+
+  v0 = _mm_madd_epi16(v0, k__sqrt2_epi16);
+  v1 = _mm_madd_epi16(v1, k__sqrt2_epi16);
+  v2 = _mm_madd_epi16(v2, k__sqrt2_epi16);
+  v3 = _mm_madd_epi16(v3, k__sqrt2_epi16);
+  v4 = _mm_madd_epi16(v4, k__sqrt2_epi16);
+  v5 = _mm_madd_epi16(v5, k__sqrt2_epi16);
+  v6 = _mm_madd_epi16(v6, k__sqrt2_epi16);
+  v7 = _mm_madd_epi16(v7, k__sqrt2_epi16);
+
+  x0 = _mm_madd_epi16(x0, k__sqrt2_epi16);
+  x1 = _mm_madd_epi16(x1, k__sqrt2_epi16);
+  x2 = _mm_madd_epi16(x2, k__sqrt2_epi16);
+  x3 = _mm_madd_epi16(x3, k__sqrt2_epi16);
+  x4 = _mm_madd_epi16(x4, k__sqrt2_epi16);
+  x5 = _mm_madd_epi16(x5, k__sqrt2_epi16);
+  x6 = _mm_madd_epi16(x6, k__sqrt2_epi16);
+  x7 = _mm_madd_epi16(x7, k__sqrt2_epi16);
+
+  u0 = _mm_madd_epi16(u0, k__sqrt2_epi16);
+  u1 = _mm_madd_epi16(u1, k__sqrt2_epi16);
+  u2 = _mm_madd_epi16(u2, k__sqrt2_epi16);
+  u3 = _mm_madd_epi16(u3, k__sqrt2_epi16);
+  u4 = _mm_madd_epi16(u4, k__sqrt2_epi16);
+  u5 = _mm_madd_epi16(u5, k__sqrt2_epi16);
+  u6 = _mm_madd_epi16(u6, k__sqrt2_epi16);
+  u7 = _mm_madd_epi16(u7, k__sqrt2_epi16);
+
+  y0 = _mm_madd_epi16(y0, k__sqrt2_epi16);
+  y1 = _mm_madd_epi16(y1, k__sqrt2_epi16);
+  y2 = _mm_madd_epi16(y2, k__sqrt2_epi16);
+  y3 = _mm_madd_epi16(y3, k__sqrt2_epi16);
+  y4 = _mm_madd_epi16(y4, k__sqrt2_epi16);
+  y5 = _mm_madd_epi16(y5, k__sqrt2_epi16);
+  y6 = _mm_madd_epi16(y6, k__sqrt2_epi16);
+  y7 = _mm_madd_epi16(y7, k__sqrt2_epi16);
+
+  v0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  x0 = _mm_add_epi32(x0, k__DCT_CONST_ROUNDING);
+  x1 = _mm_add_epi32(x1, k__DCT_CONST_ROUNDING);
+  x2 = _mm_add_epi32(x2, k__DCT_CONST_ROUNDING);
+  x3 = _mm_add_epi32(x3, k__DCT_CONST_ROUNDING);
+  x4 = _mm_add_epi32(x4, k__DCT_CONST_ROUNDING);
+  x5 = _mm_add_epi32(x5, k__DCT_CONST_ROUNDING);
+  x6 = _mm_add_epi32(x6, k__DCT_CONST_ROUNDING);
+  x7 = _mm_add_epi32(x7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+
+  y0 = _mm_add_epi32(y0, k__DCT_CONST_ROUNDING);
+  y1 = _mm_add_epi32(y1, k__DCT_CONST_ROUNDING);
+  y2 = _mm_add_epi32(y2, k__DCT_CONST_ROUNDING);
+  y3 = _mm_add_epi32(y3, k__DCT_CONST_ROUNDING);
+  y4 = _mm_add_epi32(y4, k__DCT_CONST_ROUNDING);
+  y5 = _mm_add_epi32(y5, k__DCT_CONST_ROUNDING);
+  y6 = _mm_add_epi32(y6, k__DCT_CONST_ROUNDING);
+  y7 = _mm_add_epi32(y7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  x0 = _mm_srai_epi32(x0, DCT_CONST_BITS);
+  x1 = _mm_srai_epi32(x1, DCT_CONST_BITS);
+  x2 = _mm_srai_epi32(x2, DCT_CONST_BITS);
+  x3 = _mm_srai_epi32(x3, DCT_CONST_BITS);
+  x4 = _mm_srai_epi32(x4, DCT_CONST_BITS);
+  x5 = _mm_srai_epi32(x5, DCT_CONST_BITS);
+  x6 = _mm_srai_epi32(x6, DCT_CONST_BITS);
+  x7 = _mm_srai_epi32(x7, DCT_CONST_BITS);
+
+  u0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  y0 = _mm_srai_epi32(y0, DCT_CONST_BITS);
+  y1 = _mm_srai_epi32(y1, DCT_CONST_BITS);
+  y2 = _mm_srai_epi32(y2, DCT_CONST_BITS);
+  y3 = _mm_srai_epi32(y3, DCT_CONST_BITS);
+  y4 = _mm_srai_epi32(y4, DCT_CONST_BITS);
+  y5 = _mm_srai_epi32(y5, DCT_CONST_BITS);
+  y6 = _mm_srai_epi32(y6, DCT_CONST_BITS);
+  y7 = _mm_srai_epi32(y7, DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(v0, x0);
+  in[1] = _mm_packs_epi32(v1, x1);
+  in[2] = _mm_packs_epi32(v2, x2);
+  in[3] = _mm_packs_epi32(v3, x3);
+  in[4] = _mm_packs_epi32(v4, x4);
+  in[5] = _mm_packs_epi32(v5, x5);
+  in[6] = _mm_packs_epi32(v6, x6);
+  in[7] = _mm_packs_epi32(v7, x7);
+
+  in[8] = _mm_packs_epi32(u0, y0);
+  in[9] = _mm_packs_epi32(u1, y1);
+  in[10] = _mm_packs_epi32(u2, y2);
+  in[11] = _mm_packs_epi32(u3, y3);
+  in[12] = _mm_packs_epi32(u4, y4);
+  in[13] = _mm_packs_epi32(u5, y5);
+  in[14] = _mm_packs_epi32(u6, y6);
+  in[15] = _mm_packs_epi32(u7, y7);
+}
+#endif  // CONFIG_EXT_TX
+
+static INLINE void scale_sqrt2_8x4(__m128i *in) {
+  // Implements ROUND_POWER_OF_TWO(input * Sqrt2, DCT_CONST_BITS), for 32
+  // consecutive elements.
+  const __m128i v_scale_w = _mm_set1_epi16((int16_t)Sqrt2);
+
+  const __m128i v_p0l_w = _mm_mullo_epi16(in[0], v_scale_w);
+  const __m128i v_p0h_w = _mm_mulhi_epi16(in[0], v_scale_w);
+  const __m128i v_p1l_w = _mm_mullo_epi16(in[1], v_scale_w);
+  const __m128i v_p1h_w = _mm_mulhi_epi16(in[1], v_scale_w);
+  const __m128i v_p2l_w = _mm_mullo_epi16(in[2], v_scale_w);
+  const __m128i v_p2h_w = _mm_mulhi_epi16(in[2], v_scale_w);
+  const __m128i v_p3l_w = _mm_mullo_epi16(in[3], v_scale_w);
+  const __m128i v_p3h_w = _mm_mulhi_epi16(in[3], v_scale_w);
+
+  const __m128i v_p0a_d = _mm_unpacklo_epi16(v_p0l_w, v_p0h_w);
+  const __m128i v_p0b_d = _mm_unpackhi_epi16(v_p0l_w, v_p0h_w);
+  const __m128i v_p1a_d = _mm_unpacklo_epi16(v_p1l_w, v_p1h_w);
+  const __m128i v_p1b_d = _mm_unpackhi_epi16(v_p1l_w, v_p1h_w);
+  const __m128i v_p2a_d = _mm_unpacklo_epi16(v_p2l_w, v_p2h_w);
+  const __m128i v_p2b_d = _mm_unpackhi_epi16(v_p2l_w, v_p2h_w);
+  const __m128i v_p3a_d = _mm_unpacklo_epi16(v_p3l_w, v_p3h_w);
+  const __m128i v_p3b_d = _mm_unpackhi_epi16(v_p3l_w, v_p3h_w);
+
+  in[0] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p0a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p0b_d, DCT_CONST_BITS));
+  in[1] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p1a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p1b_d, DCT_CONST_BITS));
+  in[2] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p2a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p2b_d, DCT_CONST_BITS));
+  in[3] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p3a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p3b_d, DCT_CONST_BITS));
+}
+
+static INLINE void scale_sqrt2_8x8(__m128i *in) {
+  // Implements 'ROUND_POWER_OF_TWO_SIGNED(input * Sqrt2, DCT_CONST_BITS)'
+  // for each element.
+  const __m128i v_scale_w = _mm_set1_epi16((int16_t)Sqrt2);
+
+  const __m128i v_p0l_w = _mm_mullo_epi16(in[0], v_scale_w);
+  const __m128i v_p0h_w = _mm_mulhi_epi16(in[0], v_scale_w);
+  const __m128i v_p1l_w = _mm_mullo_epi16(in[1], v_scale_w);
+  const __m128i v_p1h_w = _mm_mulhi_epi16(in[1], v_scale_w);
+  const __m128i v_p2l_w = _mm_mullo_epi16(in[2], v_scale_w);
+  const __m128i v_p2h_w = _mm_mulhi_epi16(in[2], v_scale_w);
+  const __m128i v_p3l_w = _mm_mullo_epi16(in[3], v_scale_w);
+  const __m128i v_p3h_w = _mm_mulhi_epi16(in[3], v_scale_w);
+  const __m128i v_p4l_w = _mm_mullo_epi16(in[4], v_scale_w);
+  const __m128i v_p4h_w = _mm_mulhi_epi16(in[4], v_scale_w);
+  const __m128i v_p5l_w = _mm_mullo_epi16(in[5], v_scale_w);
+  const __m128i v_p5h_w = _mm_mulhi_epi16(in[5], v_scale_w);
+  const __m128i v_p6l_w = _mm_mullo_epi16(in[6], v_scale_w);
+  const __m128i v_p6h_w = _mm_mulhi_epi16(in[6], v_scale_w);
+  const __m128i v_p7l_w = _mm_mullo_epi16(in[7], v_scale_w);
+  const __m128i v_p7h_w = _mm_mulhi_epi16(in[7], v_scale_w);
+
+  const __m128i v_p0a_d = _mm_unpacklo_epi16(v_p0l_w, v_p0h_w);
+  const __m128i v_p0b_d = _mm_unpackhi_epi16(v_p0l_w, v_p0h_w);
+  const __m128i v_p1a_d = _mm_unpacklo_epi16(v_p1l_w, v_p1h_w);
+  const __m128i v_p1b_d = _mm_unpackhi_epi16(v_p1l_w, v_p1h_w);
+  const __m128i v_p2a_d = _mm_unpacklo_epi16(v_p2l_w, v_p2h_w);
+  const __m128i v_p2b_d = _mm_unpackhi_epi16(v_p2l_w, v_p2h_w);
+  const __m128i v_p3a_d = _mm_unpacklo_epi16(v_p3l_w, v_p3h_w);
+  const __m128i v_p3b_d = _mm_unpackhi_epi16(v_p3l_w, v_p3h_w);
+  const __m128i v_p4a_d = _mm_unpacklo_epi16(v_p4l_w, v_p4h_w);
+  const __m128i v_p4b_d = _mm_unpackhi_epi16(v_p4l_w, v_p4h_w);
+  const __m128i v_p5a_d = _mm_unpacklo_epi16(v_p5l_w, v_p5h_w);
+  const __m128i v_p5b_d = _mm_unpackhi_epi16(v_p5l_w, v_p5h_w);
+  const __m128i v_p6a_d = _mm_unpacklo_epi16(v_p6l_w, v_p6h_w);
+  const __m128i v_p6b_d = _mm_unpackhi_epi16(v_p6l_w, v_p6h_w);
+  const __m128i v_p7a_d = _mm_unpacklo_epi16(v_p7l_w, v_p7h_w);
+  const __m128i v_p7b_d = _mm_unpackhi_epi16(v_p7l_w, v_p7h_w);
+
+  in[0] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p0a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p0b_d, DCT_CONST_BITS));
+  in[1] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p1a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p1b_d, DCT_CONST_BITS));
+  in[2] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p2a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p2b_d, DCT_CONST_BITS));
+  in[3] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p3a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p3b_d, DCT_CONST_BITS));
+  in[4] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p4a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p4b_d, DCT_CONST_BITS));
+  in[5] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p5a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p5b_d, DCT_CONST_BITS));
+  in[6] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p6a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p6b_d, DCT_CONST_BITS));
+  in[7] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p7a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p7b_d, DCT_CONST_BITS));
+}
+
+static INLINE void scale_sqrt2_8x16(__m128i *in) {
+  scale_sqrt2_8x8(in);
+  scale_sqrt2_8x8(in + 8);
+}
+
+#endif  // AOM_DSP_X86_TXFM_COMMON_SSE2_H_
diff --git a/third_party/aom/aom_dsp/x86/variance_avx2.c b/third_party/aom/aom_dsp/x86/variance_avx2.c
new file mode 100644
index 0000000000..18a70dffe7
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/variance_avx2.c
@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>
+#include "./aom_dsp_rtcd.h"
+
+typedef void (*get_var_avx2)(const uint8_t *src, int src_stride,
+                             const uint8_t *ref, int ref_stride,
+                             unsigned int *sse, int *sum);
+
+void aom_get32x32var_avx2(const uint8_t *src, int src_stride,
+                          const uint8_t *ref, int ref_stride, unsigned int *sse,
+                          int *sum);
+
+static void variance_avx2(const uint8_t *src, int src_stride,
+                          const uint8_t *ref, int ref_stride, int w, int h,
+                          unsigned int *sse, int *sum, get_var_avx2 var_fn,
+                          int block_size) {
+  int i, j;
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i += 16) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(&src[src_stride * i + j], src_stride, &ref[ref_stride * i + j],
+             ref_stride, &sse0, &sum0);
+      *sse += sse0;
+      *sum += sum0;
+    }
+  }
+}
+
+unsigned int aom_variance16x16_avx2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  unsigned int variance;
+  variance_avx2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum,
+                aom_get16x16var_avx2, 16);
+
+  variance = *sse - (((uint32_t)((int64_t)sum * sum)) >> 8);
+  _mm256_zeroupper();
+  return variance;
+}
+
+unsigned int aom_mse16x16_avx2(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride,
+                               unsigned int *sse) {
+  int sum;
+  aom_get16x16var_avx2(src, src_stride, ref, ref_stride, sse, &sum);
+  _mm256_zeroupper();
+  return *sse;
+}
+
+unsigned int aom_variance32x16_avx2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  unsigned int variance;
+  variance_avx2(src, src_stride, ref, ref_stride, 32, 16, sse, &sum,
+                aom_get32x32var_avx2, 32);
+
+  variance = *sse - (uint32_t)(((int64_t)sum * sum) >> 9);
+  _mm256_zeroupper();
+  return variance;
+}
+
+unsigned int aom_variance32x32_avx2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  unsigned int variance;
+  variance_avx2(src, src_stride, ref, ref_stride, 32, 32, sse, &sum,
+                aom_get32x32var_avx2, 32);
+
+  variance = *sse - (uint32_t)(((int64_t)sum * sum) >> 10);
+  _mm256_zeroupper();
+  return variance;
+}
+
+unsigned int aom_variance64x64_avx2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  unsigned int variance;
+  variance_avx2(src, src_stride, ref, ref_stride, 64, 64, sse, &sum,
+                aom_get32x32var_avx2, 32);
+
+  variance = *sse - (uint32_t)(((int64_t)sum * sum) >> 12);
+  _mm256_zeroupper();
+  return variance;
+}
+
+unsigned int aom_variance64x32_avx2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  unsigned int variance;
+  variance_avx2(src, src_stride, ref, ref_stride, 64, 32, sse, &sum,
+                aom_get32x32var_avx2, 32);
+
+  variance = *sse - (uint32_t)(((int64_t)sum * sum) >> 11);
+  _mm256_zeroupper();
+  return variance;
+}
+
+unsigned int aom_sub_pixel_variance32xh_avx2(const uint8_t *src, int src_stride,
+                                             int x_offset, int y_offset,
+                                             const uint8_t *dst, int dst_stride,
+                                             int height, unsigned int *sse);
+
+unsigned int aom_sub_pixel_avg_variance32xh_avx2(
+    const uint8_t *src, int src_stride, int x_offset, int y_offset,
+    const uint8_t *dst, int dst_stride, const uint8_t *sec, int sec_stride,
+    int height, unsigned int *sseptr);
+
+unsigned int aom_sub_pixel_variance64x64_avx2(const uint8_t *src,
+                                              int src_stride, int x_offset,
+                                              int y_offset, const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  unsigned int sse1;
+  const int se1 = aom_sub_pixel_variance32xh_avx2(
+      src, src_stride, x_offset, y_offset, dst, dst_stride, 64, &sse1);
+  unsigned int sse2;
+  const int se2 =
+      aom_sub_pixel_variance32xh_avx2(src + 32, src_stride, x_offset, y_offset,
+                                      dst + 32, dst_stride, 64, &sse2);
+  const int se = se1 + se2;
+  unsigned int variance;
+  *sse = sse1 + sse2;
+
+  variance = *sse - (uint32_t)(((int64_t)se * se) >> 12);
+  _mm256_zeroupper();
+  return variance;
+}
+
+unsigned int aom_sub_pixel_variance32x32_avx2(const uint8_t *src,
+                                              int src_stride, int x_offset,
+                                              int y_offset, const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  const int se = aom_sub_pixel_variance32xh_avx2(
+      src, src_stride, x_offset, y_offset, dst, dst_stride, 32, sse);
+
+  const unsigned int variance = *sse - (uint32_t)(((int64_t)se * se) >> 10);
+  _mm256_zeroupper();
+  return variance;
+}
+
+unsigned int aom_sub_pixel_avg_variance64x64_avx2(
+    const uint8_t *src, int src_stride, int x_offset, int y_offset,
+    const uint8_t *dst, int dst_stride, unsigned int *sse, const uint8_t *sec) {
+  unsigned int sse1;
+  const int se1 = aom_sub_pixel_avg_variance32xh_avx2(
+      src, src_stride, x_offset, y_offset, dst, dst_stride, sec, 64, 64, &sse1);
+  unsigned int sse2;
+  const int se2 = aom_sub_pixel_avg_variance32xh_avx2(
+      src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, sec + 32,
+      64, 64, &sse2);
+  const int se = se1 + se2;
+  unsigned int variance;
+
+  *sse = sse1 + sse2;
+
+  variance = *sse - (uint32_t)(((int64_t)se * se) >> 12);
+  _mm256_zeroupper();
+  return variance;
+}
+
+unsigned int aom_sub_pixel_avg_variance32x32_avx2(
+    const uint8_t *src, int src_stride, int x_offset, int y_offset,
+    const uint8_t *dst, int dst_stride, unsigned int *sse, const uint8_t *sec) {
+  // Process 32 elements in parallel.
+  const int se = aom_sub_pixel_avg_variance32xh_avx2(
+      src, src_stride, x_offset, y_offset, dst, dst_stride, sec, 32, 32, sse);
+
+  const unsigned int variance = *sse - (uint32_t)(((int64_t)se * se) >> 10);
+  _mm256_zeroupper();
+  return variance;
+}
diff --git a/third_party/aom/aom_dsp/x86/variance_impl_avx2.c b/third_party/aom/aom_dsp/x86/variance_impl_avx2.c
new file mode 100644
index 0000000000..999b541e35
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/variance_impl_avx2.c
@@ -0,0 +1,713 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/mem.h"
+
+/* clang-format off */
+DECLARE_ALIGNED(32, static const uint8_t, bilinear_filters_avx2[512]) = {
+  16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0,
+  16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0,
+  14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2,
+  14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2,
+  12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4,
+  12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4,
+  10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6,
+  10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6,
+   8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+   8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+   6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,
+   6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,
+   4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,
+   4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,
+   2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,
+   2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,
+};
+/* clang-format on */
+
+void aom_get16x16var_avx2(const unsigned char *src_ptr, int source_stride,
+                          const unsigned char *ref_ptr, int recon_stride,
+                          unsigned int *SSE, int *Sum) {
+  __m256i src, src_expand_low, src_expand_high, ref, ref_expand_low;
+  __m256i ref_expand_high, madd_low, madd_high;
+  unsigned int i, src_2strides, ref_2strides;
+  __m256i zero_reg = _mm256_set1_epi16(0);
+  __m256i sum_ref_src = _mm256_set1_epi16(0);
+  __m256i madd_ref_src = _mm256_set1_epi16(0);
+
+  // processing two strides in a 256 bit register reducing the number
+  // of loop stride by half (comparing to the sse2 code)
+  src_2strides = source_stride << 1;
+  ref_2strides = recon_stride << 1;
+  for (i = 0; i < 8; i++) {
+    src = _mm256_castsi128_si256(_mm_loadu_si128((__m128i const *)(src_ptr)));
+    src = _mm256_inserti128_si256(
+        src, _mm_loadu_si128((__m128i const *)(src_ptr + source_stride)), 1);
+
+    ref = _mm256_castsi128_si256(_mm_loadu_si128((__m128i const *)(ref_ptr)));
+    ref = _mm256_inserti128_si256(
+        ref, _mm_loadu_si128((__m128i const *)(ref_ptr + recon_stride)), 1);
+
+    // expanding to 16 bit each lane
+    src_expand_low = _mm256_unpacklo_epi8(src, zero_reg);
+    src_expand_high = _mm256_unpackhi_epi8(src, zero_reg);
+
+    ref_expand_low = _mm256_unpacklo_epi8(ref, zero_reg);
+    ref_expand_high = _mm256_unpackhi_epi8(ref, zero_reg);
+
+    // src-ref
+    src_expand_low = _mm256_sub_epi16(src_expand_low, ref_expand_low);
+    src_expand_high = _mm256_sub_epi16(src_expand_high, ref_expand_high);
+
+    // madd low (src - ref)
+    madd_low = _mm256_madd_epi16(src_expand_low, src_expand_low);
+
+    // add high to low
+    src_expand_low = _mm256_add_epi16(src_expand_low, src_expand_high);
+
+    // madd high (src - ref)
+    madd_high = _mm256_madd_epi16(src_expand_high, src_expand_high);
+
+    sum_ref_src = _mm256_add_epi16(sum_ref_src, src_expand_low);
+
+    // add high to low
+    madd_ref_src =
+        _mm256_add_epi32(madd_ref_src, _mm256_add_epi32(madd_low, madd_high));
+
+    src_ptr += src_2strides;
+    ref_ptr += ref_2strides;
+  }
+
+  {
+    __m128i sum_res, madd_res;
+    __m128i expand_sum_low, expand_sum_high, expand_sum;
+    __m128i expand_madd_low, expand_madd_high, expand_madd;
+    __m128i ex_expand_sum_low, ex_expand_sum_high, ex_expand_sum;
+
+    // extract the low lane and add it to the high lane
+    sum_res = _mm_add_epi16(_mm256_castsi256_si128(sum_ref_src),
+                            _mm256_extractf128_si256(sum_ref_src, 1));
+
+    madd_res = _mm_add_epi32(_mm256_castsi256_si128(madd_ref_src),
+                             _mm256_extractf128_si256(madd_ref_src, 1));
+
+    // padding each 2 bytes with another 2 zeroed bytes
+    expand_sum_low =
+        _mm_unpacklo_epi16(_mm256_castsi256_si128(zero_reg), sum_res);
+    expand_sum_high =
+        _mm_unpackhi_epi16(_mm256_castsi256_si128(zero_reg), sum_res);
+
+    // shifting the sign 16 bits right
+    expand_sum_low = _mm_srai_epi32(expand_sum_low, 16);
+    expand_sum_high = _mm_srai_epi32(expand_sum_high, 16);
+
+    expand_sum = _mm_add_epi32(expand_sum_low, expand_sum_high);
+
+    // expand each 32 bits of the madd result to 64 bits
+    expand_madd_low =
+        _mm_unpacklo_epi32(madd_res, _mm256_castsi256_si128(zero_reg));
+    expand_madd_high =
+        _mm_unpackhi_epi32(madd_res, _mm256_castsi256_si128(zero_reg));
+
+    expand_madd = _mm_add_epi32(expand_madd_low, expand_madd_high);
+
+    ex_expand_sum_low =
+        _mm_unpacklo_epi32(expand_sum, _mm256_castsi256_si128(zero_reg));
+    ex_expand_sum_high =
+        _mm_unpackhi_epi32(expand_sum, _mm256_castsi256_si128(zero_reg));
+
+    ex_expand_sum = _mm_add_epi32(ex_expand_sum_low, ex_expand_sum_high);
+
+    // shift 8 bytes eight
+    madd_res = _mm_srli_si128(expand_madd, 8);
+    sum_res = _mm_srli_si128(ex_expand_sum, 8);
+
+    madd_res = _mm_add_epi32(madd_res, expand_madd);
+    sum_res = _mm_add_epi32(sum_res, ex_expand_sum);
+
+    *((int *)SSE) = _mm_cvtsi128_si32(madd_res);
+
+    *((int *)Sum) = _mm_cvtsi128_si32(sum_res);
+  }
+  _mm256_zeroupper();
+}
+
+void aom_get32x32var_avx2(const unsigned char *src_ptr, int source_stride,
+                          const unsigned char *ref_ptr, int recon_stride,
+                          unsigned int *SSE, int *Sum) {
+  __m256i src, src_expand_low, src_expand_high, ref, ref_expand_low;
+  __m256i ref_expand_high, madd_low, madd_high;
+  unsigned int i;
+  __m256i zero_reg = _mm256_set1_epi16(0);
+  __m256i sum_ref_src = _mm256_set1_epi16(0);
+  __m256i madd_ref_src = _mm256_set1_epi16(0);
+
+  // processing 32 elements in parallel
+  for (i = 0; i < 16; i++) {
+    src = _mm256_loadu_si256((__m256i const *)(src_ptr));
+
+    ref = _mm256_loadu_si256((__m256i const *)(ref_ptr));
+
+    // expanding to 16 bit each lane
+    src_expand_low = _mm256_unpacklo_epi8(src, zero_reg);
+    src_expand_high = _mm256_unpackhi_epi8(src, zero_reg);
+
+    ref_expand_low = _mm256_unpacklo_epi8(ref, zero_reg);
+    ref_expand_high = _mm256_unpackhi_epi8(ref, zero_reg);
+
+    // src-ref
+    src_expand_low = _mm256_sub_epi16(src_expand_low, ref_expand_low);
+    src_expand_high = _mm256_sub_epi16(src_expand_high, ref_expand_high);
+
+    // madd low (src - ref)
+    madd_low = _mm256_madd_epi16(src_expand_low, src_expand_low);
+
+    // add high to low
+    src_expand_low = _mm256_add_epi16(src_expand_low, src_expand_high);
+
+    // madd high (src - ref)
+    madd_high = _mm256_madd_epi16(src_expand_high, src_expand_high);
+
+    sum_ref_src = _mm256_add_epi16(sum_ref_src, src_expand_low);
+
+    // add high to low
+    madd_ref_src =
+        _mm256_add_epi32(madd_ref_src, _mm256_add_epi32(madd_low, madd_high));
+
+    src_ptr += source_stride;
+    ref_ptr += recon_stride;
+  }
+
+  {
+    __m256i expand_sum_low, expand_sum_high, expand_sum;
+    __m256i expand_madd_low, expand_madd_high, expand_madd;
+    __m256i ex_expand_sum_low, ex_expand_sum_high, ex_expand_sum;
+
+    // padding each 2 bytes with another 2 zeroed bytes
+    expand_sum_low = _mm256_unpacklo_epi16(zero_reg, sum_ref_src);
+    expand_sum_high = _mm256_unpackhi_epi16(zero_reg, sum_ref_src);
+
+    // shifting the sign 16 bits right
+    expand_sum_low = _mm256_srai_epi32(expand_sum_low, 16);
+    expand_sum_high = _mm256_srai_epi32(expand_sum_high, 16);
+
+    expand_sum = _mm256_add_epi32(expand_sum_low, expand_sum_high);
+
+    // expand each 32 bits of the madd result to 64 bits
+    expand_madd_low = _mm256_unpacklo_epi32(madd_ref_src, zero_reg);
+    expand_madd_high = _mm256_unpackhi_epi32(madd_ref_src, zero_reg);
+
+    expand_madd = _mm256_add_epi32(expand_madd_low, expand_madd_high);
+
+    ex_expand_sum_low = _mm256_unpacklo_epi32(expand_sum, zero_reg);
+    ex_expand_sum_high = _mm256_unpackhi_epi32(expand_sum, zero_reg);
+
+    ex_expand_sum = _mm256_add_epi32(ex_expand_sum_low, ex_expand_sum_high);
+
+    // shift 8 bytes eight
+    madd_ref_src = _mm256_srli_si256(expand_madd, 8);
+    sum_ref_src = _mm256_srli_si256(ex_expand_sum, 8);
+
+    madd_ref_src = _mm256_add_epi32(madd_ref_src, expand_madd);
+    sum_ref_src = _mm256_add_epi32(sum_ref_src, ex_expand_sum);
+
+    // extract the low lane and the high lane and add the results
+    *((int *)SSE) =
+        _mm_cvtsi128_si32(_mm256_castsi256_si128(madd_ref_src)) +
+        _mm_cvtsi128_si32(_mm256_extractf128_si256(madd_ref_src, 1));
+
+    *((int *)Sum) = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_ref_src)) +
+                    _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_ref_src, 1));
+  }
+  _mm256_zeroupper();
+}
+
+#define FILTER_SRC(filter)                               \
+  /* filter the source */                                \
+  exp_src_lo = _mm256_maddubs_epi16(exp_src_lo, filter); \
+  exp_src_hi = _mm256_maddubs_epi16(exp_src_hi, filter); \
+                                                         \
+  /* add 8 to source */                                  \
+  exp_src_lo = _mm256_add_epi16(exp_src_lo, pw8);        \
+  exp_src_hi = _mm256_add_epi16(exp_src_hi, pw8);        \
+                                                         \
+  /* divide source by 16 */                              \
+  exp_src_lo = _mm256_srai_epi16(exp_src_lo, 4);         \
+  exp_src_hi = _mm256_srai_epi16(exp_src_hi, 4);
+
+#define MERGE_WITH_SRC(src_reg, reg)               \
+  exp_src_lo = _mm256_unpacklo_epi8(src_reg, reg); \
+  exp_src_hi = _mm256_unpackhi_epi8(src_reg, reg);
+
+#define LOAD_SRC_DST                                    \
+  /* load source and destination */                     \
+  src_reg = _mm256_loadu_si256((__m256i const *)(src)); \
+  dst_reg = _mm256_loadu_si256((__m256i const *)(dst));
+
+#define AVG_NEXT_SRC(src_reg, size_stride)                                 \
+  src_next_reg = _mm256_loadu_si256((__m256i const *)(src + size_stride)); \
+  /* average between current and next stride source */                     \
+  src_reg = _mm256_avg_epu8(src_reg, src_next_reg);
+
+#define MERGE_NEXT_SRC(src_reg, size_stride)                               \
+  src_next_reg = _mm256_loadu_si256((__m256i const *)(src + size_stride)); \
+  MERGE_WITH_SRC(src_reg, src_next_reg)
+
+#define CALC_SUM_SSE_INSIDE_LOOP                          \
+  /* expand each byte to 2 bytes */                       \
+  exp_dst_lo = _mm256_unpacklo_epi8(dst_reg, zero_reg);   \
+  exp_dst_hi = _mm256_unpackhi_epi8(dst_reg, zero_reg);   \
+  /* source - dest */                                     \
+  exp_src_lo = _mm256_sub_epi16(exp_src_lo, exp_dst_lo);  \
+  exp_src_hi = _mm256_sub_epi16(exp_src_hi, exp_dst_hi);  \
+  /* caculate sum */                                      \
+  sum_reg = _mm256_add_epi16(sum_reg, exp_src_lo);        \
+  exp_src_lo = _mm256_madd_epi16(exp_src_lo, exp_src_lo); \
+  sum_reg = _mm256_add_epi16(sum_reg, exp_src_hi);        \
+  exp_src_hi = _mm256_madd_epi16(exp_src_hi, exp_src_hi); \
+  /* calculate sse */                                     \
+  sse_reg = _mm256_add_epi32(sse_reg, exp_src_lo);        \
+  sse_reg = _mm256_add_epi32(sse_reg, exp_src_hi);
+
+// final calculation to sum and sse
+#define CALC_SUM_AND_SSE                                                   \
+  res_cmp = _mm256_cmpgt_epi16(zero_reg, sum_reg);                         \
+  sse_reg_hi = _mm256_srli_si256(sse_reg, 8);                              \
+  sum_reg_lo = _mm256_unpacklo_epi16(sum_reg, res_cmp);                    \
+  sum_reg_hi = _mm256_unpackhi_epi16(sum_reg, res_cmp);                    \
+  sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi);                         \
+  sum_reg = _mm256_add_epi32(sum_reg_lo, sum_reg_hi);                      \
+                                                                           \
+  sse_reg_hi = _mm256_srli_si256(sse_reg, 4);                              \
+  sum_reg_hi = _mm256_srli_si256(sum_reg, 8);                              \
+                                                                           \
+  sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi);                         \
+  sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi);                         \
+  *((int *)sse) = _mm_cvtsi128_si32(_mm256_castsi256_si128(sse_reg)) +     \
+                  _mm_cvtsi128_si32(_mm256_extractf128_si256(sse_reg, 1)); \
+  sum_reg_hi = _mm256_srli_si256(sum_reg, 4);                              \
+  sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi);                         \
+  sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_reg)) +               \
+        _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_reg, 1));
+
+unsigned int aom_sub_pixel_variance32xh_avx2(const uint8_t *src, int src_stride,
+                                             int x_offset, int y_offset,
+                                             const uint8_t *dst, int dst_stride,
+                                             int height, unsigned int *sse) {
+  __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi;
+  __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi;
+  __m256i zero_reg;
+  int i, sum;
+  sum_reg = _mm256_set1_epi16(0);
+  sse_reg = _mm256_set1_epi16(0);
+  zero_reg = _mm256_set1_epi16(0);
+
+  // x_offset = 0 and y_offset = 0
+  if (x_offset == 0) {
+    if (y_offset == 0) {
+      for (i = 0; i < height; i++) {
+        LOAD_SRC_DST
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src += src_stride;
+        dst += dst_stride;
+      }
+      // x_offset = 0 and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i src_next_reg;
+      for (i = 0; i < height; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, src_stride)
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src += src_stride;
+        dst += dst_stride;
+      }
+      // x_offset = 0 and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg;
+
+      y_offset <<= 5;
+      filter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, src_stride)
+        FILTER_SRC(filter)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src += src_stride;
+        dst += dst_stride;
+      }
+    }
+    // x_offset = 8  and y_offset = 0
+  } else if (x_offset == 8) {
+    if (y_offset == 0) {
+      __m256i src_next_reg;
+      for (i = 0; i < height; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src += src_stride;
+        dst += dst_stride;
+      }
+      // x_offset = 8  and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i src_next_reg, src_avg;
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *)(src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height; i++) {
+        src_avg = src_reg;
+        src += src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        // average between previous average to current average
+        src_avg = _mm256_avg_epu8(src_avg, src_reg);
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_avg, zero_reg)
+        // save current source average
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst += dst_stride;
+      }
+      // x_offset = 8  and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg, src_avg;
+      y_offset <<= 5;
+      filter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *)(src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height; i++) {
+        // save current source average
+        src_avg = src_reg;
+        src += src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        MERGE_WITH_SRC(src_avg, src_reg)
+        FILTER_SRC(filter)
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst += dst_stride;
+      }
+    }
+    // x_offset = bilin interpolation and y_offset = 0
+  } else {
+    if (y_offset == 0) {
+      __m256i filter, pw8, src_next_reg;
+      x_offset <<= 5;
+      filter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src += src_stride;
+        dst += dst_stride;
+      }
+      // x_offset = bilin interpolation and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i filter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      filter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      src_reg = _mm256_loadu_si256((__m256i const *)(src));
+      MERGE_NEXT_SRC(src_reg, 1)
+      FILTER_SRC(filter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height; i++) {
+        src += src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // average between previous pack to the current
+        src_pack = _mm256_avg_epu8(src_pack, src_reg);
+        MERGE_WITH_SRC(src_pack, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src_pack = src_reg;
+        dst += dst_stride;
+      }
+      // x_offset = bilin interpolation and y_offset = bilin interpolation
+    } else {
+      __m256i xfilter, yfilter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      xfilter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + x_offset));
+      y_offset <<= 5;
+      yfilter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *)(src));
+      MERGE_NEXT_SRC(src_reg, 1)
+
+      FILTER_SRC(xfilter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height; i++) {
+        src += src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(xfilter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // merge previous pack to current pack source
+        MERGE_WITH_SRC(src_pack, src_reg)
+        // filter the source
+        FILTER_SRC(yfilter)
+        src_pack = src_reg;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst += dst_stride;
+      }
+    }
+  }
+  CALC_SUM_AND_SSE
+  _mm256_zeroupper();
+  return sum;
+}
+
+unsigned int aom_sub_pixel_avg_variance32xh_avx2(
+    const uint8_t *src, int src_stride, int x_offset, int y_offset,
+    const uint8_t *dst, int dst_stride, const uint8_t *sec, int sec_stride,
+    int height, unsigned int *sse) {
+  __m256i sec_reg;
+  __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi;
+  __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi;
+  __m256i zero_reg;
+  int i, sum;
+  sum_reg = _mm256_set1_epi16(0);
+  sse_reg = _mm256_set1_epi16(0);
+  zero_reg = _mm256_set1_epi16(0);
+
+  // x_offset = 0 and y_offset = 0
+  if (x_offset == 0) {
+    if (y_offset == 0) {
+      for (i = 0; i < height; i++) {
+        LOAD_SRC_DST
+        sec_reg = _mm256_loadu_si256((__m256i const *)(sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec += sec_stride;
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src += src_stride;
+        dst += dst_stride;
+      }
+    } else if (y_offset == 8) {
+      __m256i src_next_reg;
+      for (i = 0; i < height; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, src_stride)
+        sec_reg = _mm256_loadu_si256((__m256i const *)(sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec += sec_stride;
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src += src_stride;
+        dst += dst_stride;
+      }
+      // x_offset = 0 and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg;
+
+      y_offset <<= 5;
+      filter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, src_stride)
+        FILTER_SRC(filter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *)(sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec += sec_stride;
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src += src_stride;
+        dst += dst_stride;
+      }
+    }
+    // x_offset = 8  and y_offset = 0
+  } else if (x_offset == 8) {
+    if (y_offset == 0) {
+      __m256i src_next_reg;
+      for (i = 0; i < height; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        sec_reg = _mm256_loadu_si256((__m256i const *)(sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec += sec_stride;
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src += src_stride;
+        dst += dst_stride;
+      }
+      // x_offset = 8  and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i src_next_reg, src_avg;
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *)(src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height; i++) {
+        // save current source average
+        src_avg = src_reg;
+        src += src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        // average between previous average to current average
+        src_avg = _mm256_avg_epu8(src_avg, src_reg);
+        sec_reg = _mm256_loadu_si256((__m256i const *)(sec));
+        src_avg = _mm256_avg_epu8(src_avg, sec_reg);
+        sec += sec_stride;
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_avg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst += dst_stride;
+      }
+      // x_offset = 8  and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg, src_avg;
+      y_offset <<= 5;
+      filter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *)(src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height; i++) {
+        // save current source average
+        src_avg = src_reg;
+        src += src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        MERGE_WITH_SRC(src_avg, src_reg)
+        FILTER_SRC(filter)
+        src_avg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *)(sec));
+        src_avg = _mm256_avg_epu8(src_avg, sec_reg);
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_avg, zero_reg)
+        sec += sec_stride;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst += dst_stride;
+      }
+    }
+    // x_offset = bilin interpolation and y_offset = 0
+  } else {
+    if (y_offset == 0) {
+      __m256i filter, pw8, src_next_reg;
+      x_offset <<= 5;
+      filter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *)(sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        sec += sec_stride;
+        CALC_SUM_SSE_INSIDE_LOOP
+        src += src_stride;
+        dst += dst_stride;
+      }
+      // x_offset = bilin interpolation and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i filter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      filter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      src_reg = _mm256_loadu_si256((__m256i const *)(src));
+      MERGE_NEXT_SRC(src_reg, 1)
+      FILTER_SRC(filter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height; i++) {
+        src += src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // average between previous pack to the current
+        src_pack = _mm256_avg_epu8(src_pack, src_reg);
+        sec_reg = _mm256_loadu_si256((__m256i const *)(sec));
+        src_pack = _mm256_avg_epu8(src_pack, sec_reg);
+        sec += sec_stride;
+        MERGE_WITH_SRC(src_pack, zero_reg)
+        src_pack = src_reg;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst += dst_stride;
+      }
+      // x_offset = bilin interpolation and y_offset = bilin interpolation
+    } else {
+      __m256i xfilter, yfilter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      xfilter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + x_offset));
+      y_offset <<= 5;
+      yfilter = _mm256_load_si256(
+          (__m256i const *)(bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *)(src));
+      MERGE_NEXT_SRC(src_reg, 1)
+
+      FILTER_SRC(xfilter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height; i++) {
+        src += src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(xfilter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // merge previous pack to current pack source
+        MERGE_WITH_SRC(src_pack, src_reg)
+        // filter the source
+        FILTER_SRC(yfilter)
+        src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *)(sec));
+        src_pack = _mm256_avg_epu8(src_pack, sec_reg);
+        MERGE_WITH_SRC(src_pack, zero_reg)
+        src_pack = src_reg;
+        sec += sec_stride;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst += dst_stride;
+      }
+    }
+  }
+  CALC_SUM_AND_SSE
+  _mm256_zeroupper();
+  return sum;
+}
diff --git a/third_party/aom/aom_dsp/x86/variance_sse2.c b/third_party/aom/aom_dsp/x86/variance_sse2.c
new file mode 100644
index 0000000000..d9563aa7f1
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/variance_sse2.c
@@ -0,0 +1,690 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>  // SSE2
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom_ports/mem.h"
+
+typedef void (*getNxMvar_fn_t)(const unsigned char *src, int src_stride,
+                               const unsigned char *ref, int ref_stride,
+                               unsigned int *sse, int *sum);
+
+unsigned int aom_get_mb_ss_sse2(const int16_t *src) {
+  __m128i vsum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 32; ++i) {
+    const __m128i v = _mm_loadu_si128((const __m128i *)src);
+    vsum = _mm_add_epi32(vsum, _mm_madd_epi16(v, v));
+    src += 8;
+  }
+
+  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4));
+  return _mm_cvtsi128_si32(vsum);
+}
+
+#define READ64(p, stride, i)                                  \
+  _mm_unpacklo_epi8(                                          \
+      _mm_cvtsi32_si128(*(const uint32_t *)(p + i * stride)), \
+      _mm_cvtsi32_si128(*(const uint32_t *)(p + (i + 1) * stride)))
+
+static void get4x4var_sse2(const uint8_t *src, int src_stride,
+                           const uint8_t *ref, int ref_stride,
+                           unsigned int *sse, int *sum) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i src0 = _mm_unpacklo_epi8(READ64(src, src_stride, 0), zero);
+  const __m128i src1 = _mm_unpacklo_epi8(READ64(src, src_stride, 2), zero);
+  const __m128i ref0 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 0), zero);
+  const __m128i ref1 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 2), zero);
+  const __m128i diff0 = _mm_sub_epi16(src0, ref0);
+  const __m128i diff1 = _mm_sub_epi16(src1, ref1);
+
+  // sum
+  __m128i vsum = _mm_add_epi16(diff0, diff1);
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
+  *sum = (int16_t)_mm_extract_epi16(vsum, 0);
+
+  // sse
+  vsum =
+      _mm_add_epi32(_mm_madd_epi16(diff0, diff0), _mm_madd_epi16(diff1, diff1));
+  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4));
+  *sse = _mm_cvtsi128_si32(vsum);
+}
+
+void aom_get8x8var_sse2(const uint8_t *src, int src_stride, const uint8_t *ref,
+                        int ref_stride, unsigned int *sse, int *sum) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i vsum = _mm_setzero_si128();
+  __m128i vsse = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 8; i += 2) {
+    const __m128i src0 = _mm_unpacklo_epi8(
+        _mm_loadl_epi64((const __m128i *)(src + i * src_stride)), zero);
+    const __m128i ref0 = _mm_unpacklo_epi8(
+        _mm_loadl_epi64((const __m128i *)(ref + i * ref_stride)), zero);
+    const __m128i diff0 = _mm_sub_epi16(src0, ref0);
+
+    const __m128i src1 = _mm_unpacklo_epi8(
+        _mm_loadl_epi64((const __m128i *)(src + (i + 1) * src_stride)), zero);
+    const __m128i ref1 = _mm_unpacklo_epi8(
+        _mm_loadl_epi64((const __m128i *)(ref + (i + 1) * ref_stride)), zero);
+    const __m128i diff1 = _mm_sub_epi16(src1, ref1);
+
+    vsum = _mm_add_epi16(vsum, diff0);
+    vsum = _mm_add_epi16(vsum, diff1);
+    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
+    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
+  }
+
+  // sum
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
+  *sum = (int16_t)_mm_extract_epi16(vsum, 0);
+
+  // sse
+  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
+  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
+  *sse = _mm_cvtsi128_si32(vsse);
+}
+
+void aom_get16x16var_sse2(const uint8_t *src, int src_stride,
+                          const uint8_t *ref, int ref_stride, unsigned int *sse,
+                          int *sum) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i vsum = _mm_setzero_si128();
+  __m128i vsse = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 16; ++i) {
+    const __m128i s = _mm_loadu_si128((const __m128i *)src);
+    const __m128i r = _mm_loadu_si128((const __m128i *)ref);
+
+    const __m128i src0 = _mm_unpacklo_epi8(s, zero);
+    const __m128i ref0 = _mm_unpacklo_epi8(r, zero);
+    const __m128i diff0 = _mm_sub_epi16(src0, ref0);
+
+    const __m128i src1 = _mm_unpackhi_epi8(s, zero);
+    const __m128i ref1 = _mm_unpackhi_epi8(r, zero);
+    const __m128i diff1 = _mm_sub_epi16(src1, ref1);
+
+    vsum = _mm_add_epi16(vsum, diff0);
+    vsum = _mm_add_epi16(vsum, diff1);
+    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
+    vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
+
+    src += src_stride;
+    ref += ref_stride;
+  }
+
+  // sum
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
+  *sum =
+      (int16_t)_mm_extract_epi16(vsum, 0) + (int16_t)_mm_extract_epi16(vsum, 1);
+
+  // sse
+  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
+  vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
+  *sse = _mm_cvtsi128_si32(vsse);
+}
+
+static void variance_sse2(const unsigned char *src, int src_stride,
+                          const unsigned char *ref, int ref_stride, int w,
+                          int h, unsigned int *sse, int *sum,
+                          getNxMvar_fn_t var_fn, int block_size) {
+  int i, j;
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
+             ref_stride, &sse0, &sum0);
+      *sse += sse0;
+      *sum += sum0;
+    }
+  }
+}
+
+unsigned int aom_variance4x4_sse2(const uint8_t *src, int src_stride,
+                                  const uint8_t *ref, int ref_stride,
+                                  unsigned int *sse) {
+  int sum;
+  get4x4var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
+  assert(sum <= 255 * 4 * 4);
+  assert(sum >= -255 * 4 * 4);
+  return *sse - ((sum * sum) >> 4);
+}
+
+unsigned int aom_variance8x4_sse2(const uint8_t *src, int src_stride,
+                                  const uint8_t *ref, int ref_stride,
+                                  unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 8, 4, sse, &sum,
+                get4x4var_sse2, 4);
+  assert(sum <= 255 * 8 * 4);
+  assert(sum >= -255 * 8 * 4);
+  return *sse - ((sum * sum) >> 5);
+}
+
+unsigned int aom_variance4x8_sse2(const uint8_t *src, int src_stride,
+                                  const uint8_t *ref, int ref_stride,
+                                  unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 4, 8, sse, &sum,
+                get4x4var_sse2, 4);
+  assert(sum <= 255 * 8 * 4);
+  assert(sum >= -255 * 8 * 4);
+  return *sse - ((sum * sum) >> 5);
+}
+
+unsigned int aom_variance8x8_sse2(const unsigned char *src, int src_stride,
+                                  const unsigned char *ref, int ref_stride,
+                                  unsigned int *sse) {
+  int sum;
+  aom_get8x8var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
+  assert(sum <= 255 * 8 * 8);
+  assert(sum >= -255 * 8 * 8);
+  return *sse - ((sum * sum) >> 6);
+}
+
+unsigned int aom_variance16x8_sse2(const unsigned char *src, int src_stride,
+                                   const unsigned char *ref, int ref_stride,
+                                   unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 16, 8, sse, &sum,
+                aom_get8x8var_sse2, 8);
+  assert(sum <= 255 * 16 * 8);
+  assert(sum >= -255 * 16 * 8);
+  return *sse - ((sum * sum) >> 7);
+}
+
+unsigned int aom_variance8x16_sse2(const unsigned char *src, int src_stride,
+                                   const unsigned char *ref, int ref_stride,
+                                   unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 8, 16, sse, &sum,
+                aom_get8x8var_sse2, 8);
+  assert(sum <= 255 * 16 * 8);
+  assert(sum >= -255 * 16 * 8);
+  return *sse - ((sum * sum) >> 7);
+}
+
+unsigned int aom_variance16x16_sse2(const unsigned char *src, int src_stride,
+                                    const unsigned char *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  aom_get16x16var_sse2(src, src_stride, ref, ref_stride, sse, &sum);
+  assert(sum <= 255 * 16 * 16);
+  assert(sum >= -255 * 16 * 16);
+  return *sse - ((uint32_t)((int64_t)sum * sum) >> 8);
+}
+
+unsigned int aom_variance32x32_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 32, 32, sse, &sum,
+                aom_get16x16var_sse2, 16);
+  assert(sum <= 255 * 32 * 32);
+  assert(sum >= -255 * 32 * 32);
+  return *sse - (unsigned int)(((int64_t)sum * sum) >> 10);
+}
+
+unsigned int aom_variance32x16_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 32, 16, sse, &sum,
+                aom_get16x16var_sse2, 16);
+  assert(sum <= 255 * 32 * 16);
+  assert(sum >= -255 * 32 * 16);
+  return *sse - (unsigned int)(((int64_t)sum * sum) >> 9);
+}
+
+unsigned int aom_variance16x32_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 16, 32, sse, &sum,
+                aom_get16x16var_sse2, 16);
+  assert(sum <= 255 * 32 * 16);
+  assert(sum >= -255 * 32 * 16);
+  return *sse - (unsigned int)(((int64_t)sum * sum) >> 9);
+}
+
+unsigned int aom_variance64x64_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 64, 64, sse, &sum,
+                aom_get16x16var_sse2, 16);
+  assert(sum <= 255 * 64 * 64);
+  assert(sum >= -255 * 64 * 64);
+  return *sse - (unsigned int)(((int64_t)sum * sum) >> 12);
+}
+
+unsigned int aom_variance64x32_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 64, 32, sse, &sum,
+                aom_get16x16var_sse2, 16);
+  assert(sum <= 255 * 64 * 32);
+  assert(sum >= -255 * 64 * 32);
+  return *sse - (unsigned int)(((int64_t)sum * sum) >> 11);
+}
+
+unsigned int aom_variance32x64_sse2(const uint8_t *src, int src_stride,
+                                    const uint8_t *ref, int ref_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_sse2(src, src_stride, ref, ref_stride, 32, 64, sse, &sum,
+                aom_get16x16var_sse2, 16);
+  assert(sum <= 255 * 64 * 32);
+  assert(sum >= -255 * 64 * 32);
+  return *sse - (unsigned int)(((int64_t)sum * sum) >> 11);
+}
+
+unsigned int aom_mse8x8_sse2(const uint8_t *src, int src_stride,
+                             const uint8_t *ref, int ref_stride,
+                             unsigned int *sse) {
+  aom_variance8x8_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+unsigned int aom_mse8x16_sse2(const uint8_t *src, int src_stride,
+                              const uint8_t *ref, int ref_stride,
+                              unsigned int *sse) {
+  aom_variance8x16_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+unsigned int aom_mse16x8_sse2(const uint8_t *src, int src_stride,
+                              const uint8_t *ref, int ref_stride,
+                              unsigned int *sse) {
+  aom_variance16x8_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+unsigned int aom_mse16x16_sse2(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride,
+                               unsigned int *sse) {
+  aom_variance16x16_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+// The 2 unused parameters are place holders for PIC enabled build.
+// These definitions are for functions defined in subpel_variance.asm
+#define DECL(w, opt)                                                           \
+  int aom_sub_pixel_variance##w##xh_##opt(                                     \
+      const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset,    \
+      const uint8_t *dst, ptrdiff_t dst_stride, int height, unsigned int *sse, \
+      void *unused0, void *unused)
+#define DECLS(opt1, opt2) \
+  DECL(4, opt1);          \
+  DECL(8, opt1);          \
+  DECL(16, opt1)
+
+DECLS(sse2, sse2);
+DECLS(ssse3, ssse3);
+#undef DECLS
+#undef DECL
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast)                       \
+  unsigned int aom_sub_pixel_variance##w##x##h##_##opt(                        \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,          \
+      const uint8_t *dst, int dst_stride, unsigned int *sse_ptr) {             \
+    unsigned int sse;                                                          \
+    int se = aom_sub_pixel_variance##wf##xh_##opt(src, src_stride, x_offset,   \
+                                                  y_offset, dst, dst_stride,   \
+                                                  h, &sse, NULL, NULL);        \
+    if (w > wf) {                                                              \
+      unsigned int sse2;                                                       \
+      int se2 = aom_sub_pixel_variance##wf##xh_##opt(                          \
+          src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, h,   \
+          &sse2, NULL, NULL);                                                  \
+      se += se2;                                                               \
+      sse += sse2;                                                             \
+      if (w > wf * 2) {                                                        \
+        se2 = aom_sub_pixel_variance##wf##xh_##opt(                            \
+            src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, h, \
+            &sse2, NULL, NULL);                                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+        se2 = aom_sub_pixel_variance##wf##xh_##opt(                            \
+            src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, h, \
+            &sse2, NULL, NULL);                                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+      }                                                                        \
+    }                                                                          \
+    *sse_ptr = sse;                                                            \
+    return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2));   \
+  }
+
+#define FNS(opt1, opt2)                              \
+  FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t));  \
+  FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t));  \
+  FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t));  \
+  FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t));  \
+  FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t));  \
+  FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t));  \
+  FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \
+  FN(16, 8, 16, 4, 3, opt1, (int32_t), (int32_t));   \
+  FN(8, 16, 8, 3, 4, opt1, (int32_t), (int32_t));    \
+  FN(8, 8, 8, 3, 3, opt1, (int32_t), (int32_t));     \
+  FN(8, 4, 8, 3, 2, opt1, (int32_t), (int32_t));     \
+  FN(4, 8, 4, 2, 3, opt1, (int32_t), (int32_t));     \
+  FN(4, 4, 4, 2, 2, opt1, (int32_t), (int32_t))
+
+FNS(sse2, sse2);
+FNS(ssse3, ssse3);
+
+#undef FNS
+#undef FN
+
+// The 2 unused parameters are place holders for PIC enabled build.
+#define DECL(w, opt)                                                        \
+  int aom_sub_pixel_avg_variance##w##xh_##opt(                              \
+      const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
+      const uint8_t *dst, ptrdiff_t dst_stride, const uint8_t *sec,         \
+      ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0,   \
+      void *unused)
+#define DECLS(opt1, opt2) \
+  DECL(4, opt1);          \
+  DECL(8, opt1);          \
+  DECL(16, opt1)
+
+DECLS(sse2, sse2);
+DECLS(ssse3, ssse3);
+#undef DECL
+#undef DECLS
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast)                       \
+  unsigned int aom_sub_pixel_avg_variance##w##x##h##_##opt(                    \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,          \
+      const uint8_t *dst, int dst_stride, unsigned int *sseptr,                \
+      const uint8_t *sec) {                                                    \
+    unsigned int sse;                                                          \
+    int se = aom_sub_pixel_avg_variance##wf##xh_##opt(                         \
+        src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \
+        NULL, NULL);                                                           \
+    if (w > wf) {                                                              \
+      unsigned int sse2;                                                       \
+      int se2 = aom_sub_pixel_avg_variance##wf##xh_##opt(                      \
+          src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride,      \
+          sec + 16, w, h, &sse2, NULL, NULL);                                  \
+      se += se2;                                                               \
+      sse += sse2;                                                             \
+      if (w > wf * 2) {                                                        \
+        se2 = aom_sub_pixel_avg_variance##wf##xh_##opt(                        \
+            src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride,    \
+            sec + 32, w, h, &sse2, NULL, NULL);                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+        se2 = aom_sub_pixel_avg_variance##wf##xh_##opt(                        \
+            src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride,    \
+            sec + 48, w, h, &sse2, NULL, NULL);                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+      }                                                                        \
+    }                                                                          \
+    *sseptr = sse;                                                             \
+    return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2));   \
+  }
+
+#define FNS(opt1, opt2)                              \
+  FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t));  \
+  FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t));  \
+  FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t));  \
+  FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t));  \
+  FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t));  \
+  FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t));  \
+  FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \
+  FN(16, 8, 16, 4, 3, opt1, (uint32_t), (int32_t));  \
+  FN(8, 16, 8, 3, 4, opt1, (uint32_t), (int32_t));   \
+  FN(8, 8, 8, 3, 3, opt1, (uint32_t), (int32_t));    \
+  FN(8, 4, 8, 3, 2, opt1, (uint32_t), (int32_t));    \
+  FN(4, 8, 4, 2, 3, opt1, (uint32_t), (int32_t));    \
+  FN(4, 4, 4, 2, 2, opt1, (uint32_t), (int32_t))
+
+FNS(sse2, sse);
+FNS(ssse3, ssse3);
+
+#undef FNS
+#undef FN
+
+void aom_upsampled_pred_sse2(uint8_t *comp_pred, int width, int height,
+                             const uint8_t *ref, int ref_stride) {
+  int i, j;
+  int stride = ref_stride << 3;
+
+  if (width >= 16) {
+    // read 16 points at one time
+    for (i = 0; i < height; i++) {
+      for (j = 0; j < width; j += 16) {
+        __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
+        __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
+        __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32));
+        __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48));
+        __m128i s4 = _mm_loadu_si128((const __m128i *)(ref + 64));
+        __m128i s5 = _mm_loadu_si128((const __m128i *)(ref + 80));
+        __m128i s6 = _mm_loadu_si128((const __m128i *)(ref + 96));
+        __m128i s7 = _mm_loadu_si128((const __m128i *)(ref + 112));
+        __m128i t0, t1, t2, t3;
+
+        t0 = _mm_unpacklo_epi8(s0, s1);
+        s1 = _mm_unpackhi_epi8(s0, s1);
+        t1 = _mm_unpacklo_epi8(s2, s3);
+        s3 = _mm_unpackhi_epi8(s2, s3);
+        t2 = _mm_unpacklo_epi8(s4, s5);
+        s5 = _mm_unpackhi_epi8(s4, s5);
+        t3 = _mm_unpacklo_epi8(s6, s7);
+        s7 = _mm_unpackhi_epi8(s6, s7);
+
+        s0 = _mm_unpacklo_epi8(t0, s1);
+        s2 = _mm_unpacklo_epi8(t1, s3);
+        s4 = _mm_unpacklo_epi8(t2, s5);
+        s6 = _mm_unpacklo_epi8(t3, s7);
+        s0 = _mm_unpacklo_epi32(s0, s2);
+        s4 = _mm_unpacklo_epi32(s4, s6);
+        s0 = _mm_unpacklo_epi64(s0, s4);
+
+        _mm_storeu_si128((__m128i *)(comp_pred), s0);
+        comp_pred += 16;
+        ref += 16 * 8;
+      }
+      ref += stride - (width << 3);
+    }
+  } else if (width >= 8) {
+    // read 8 points at one time
+    for (i = 0; i < height; i++) {
+      for (j = 0; j < width; j += 8) {
+        __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
+        __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
+        __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32));
+        __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48));
+        __m128i t0, t1;
+
+        t0 = _mm_unpacklo_epi8(s0, s1);
+        s1 = _mm_unpackhi_epi8(s0, s1);
+        t1 = _mm_unpacklo_epi8(s2, s3);
+        s3 = _mm_unpackhi_epi8(s2, s3);
+
+        s0 = _mm_unpacklo_epi8(t0, s1);
+        s2 = _mm_unpacklo_epi8(t1, s3);
+        s0 = _mm_unpacklo_epi32(s0, s2);
+
+        _mm_storel_epi64((__m128i *)(comp_pred), s0);
+        comp_pred += 8;
+        ref += 8 * 8;
+      }
+      ref += stride - (width << 3);
+    }
+  } else {
+    // read 4 points at one time
+    for (i = 0; i < height; i++) {
+      for (j = 0; j < width; j += 4) {
+        __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
+        __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
+        __m128i t0;
+
+        t0 = _mm_unpacklo_epi8(s0, s1);
+        s1 = _mm_unpackhi_epi8(s0, s1);
+        s0 = _mm_unpacklo_epi8(t0, s1);
+
+        *(int *)comp_pred = _mm_cvtsi128_si32(s0);
+        comp_pred += 4;
+        ref += 4 * 8;
+      }
+      ref += stride - (width << 3);
+    }
+  }
+}
+
+void aom_comp_avg_upsampled_pred_sse2(uint8_t *comp_pred, const uint8_t *pred,
+                                      int width, int height, const uint8_t *ref,
+                                      int ref_stride) {
+  const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi16(1);
+  int i, j;
+  int stride = ref_stride << 3;
+
+  if (width >= 16) {
+    // read 16 points at one time
+    for (i = 0; i < height; i++) {
+      for (j = 0; j < width; j += 16) {
+        __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
+        __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
+        __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32));
+        __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48));
+        __m128i s4 = _mm_loadu_si128((const __m128i *)(ref + 64));
+        __m128i s5 = _mm_loadu_si128((const __m128i *)(ref + 80));
+        __m128i s6 = _mm_loadu_si128((const __m128i *)(ref + 96));
+        __m128i s7 = _mm_loadu_si128((const __m128i *)(ref + 112));
+        __m128i p0 = _mm_loadu_si128((const __m128i *)pred);
+        __m128i p1;
+        __m128i t0, t1, t2, t3;
+
+        t0 = _mm_unpacklo_epi8(s0, s1);
+        s1 = _mm_unpackhi_epi8(s0, s1);
+        t1 = _mm_unpacklo_epi8(s2, s3);
+        s3 = _mm_unpackhi_epi8(s2, s3);
+        t2 = _mm_unpacklo_epi8(s4, s5);
+        s5 = _mm_unpackhi_epi8(s4, s5);
+        t3 = _mm_unpacklo_epi8(s6, s7);
+        s7 = _mm_unpackhi_epi8(s6, s7);
+
+        s0 = _mm_unpacklo_epi8(t0, s1);
+        s2 = _mm_unpacklo_epi8(t1, s3);
+        s4 = _mm_unpacklo_epi8(t2, s5);
+        s6 = _mm_unpacklo_epi8(t3, s7);
+
+        s0 = _mm_unpacklo_epi32(s0, s2);
+        s4 = _mm_unpacklo_epi32(s4, s6);
+        s0 = _mm_unpacklo_epi8(s0, zero);
+        s4 = _mm_unpacklo_epi8(s4, zero);
+
+        p1 = _mm_unpackhi_epi8(p0, zero);
+        p0 = _mm_unpacklo_epi8(p0, zero);
+        p0 = _mm_adds_epu16(s0, p0);
+        p1 = _mm_adds_epu16(s4, p1);
+        p0 = _mm_adds_epu16(p0, one);
+        p1 = _mm_adds_epu16(p1, one);
+
+        p0 = _mm_srli_epi16(p0, 1);
+        p1 = _mm_srli_epi16(p1, 1);
+        p0 = _mm_packus_epi16(p0, p1);
+
+        _mm_storeu_si128((__m128i *)(comp_pred), p0);
+        comp_pred += 16;
+        pred += 16;
+        ref += 16 * 8;
+      }
+      ref += stride - (width << 3);
+    }
+  } else if (width >= 8) {
+    // read 8 points at one time
+    for (i = 0; i < height; i++) {
+      for (j = 0; j < width; j += 8) {
+        __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
+        __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
+        __m128i s2 = _mm_loadu_si128((const __m128i *)(ref + 32));
+        __m128i s3 = _mm_loadu_si128((const __m128i *)(ref + 48));
+        __m128i p0 = _mm_loadl_epi64((const __m128i *)pred);
+        __m128i t0, t1;
+
+        t0 = _mm_unpacklo_epi8(s0, s1);
+        s1 = _mm_unpackhi_epi8(s0, s1);
+        t1 = _mm_unpacklo_epi8(s2, s3);
+        s3 = _mm_unpackhi_epi8(s2, s3);
+
+        s0 = _mm_unpacklo_epi8(t0, s1);
+        s2 = _mm_unpacklo_epi8(t1, s3);
+        s0 = _mm_unpacklo_epi32(s0, s2);
+        s0 = _mm_unpacklo_epi8(s0, zero);
+
+        p0 = _mm_unpacklo_epi8(p0, zero);
+        p0 = _mm_adds_epu16(s0, p0);
+        p0 = _mm_adds_epu16(p0, one);
+        p0 = _mm_srli_epi16(p0, 1);
+        p0 = _mm_packus_epi16(p0, zero);
+
+        _mm_storel_epi64((__m128i *)(comp_pred), p0);
+        comp_pred += 8;
+        pred += 8;
+        ref += 8 * 8;
+      }
+      ref += stride - (width << 3);
+    }
+  } else {
+    // read 4 points at one time
+    for (i = 0; i < height; i++) {
+      for (j = 0; j < width; j += 4) {
+        __m128i s0 = _mm_loadu_si128((const __m128i *)ref);
+        __m128i s1 = _mm_loadu_si128((const __m128i *)(ref + 16));
+        __m128i p0 = _mm_cvtsi32_si128(*(const uint32_t *)pred);
+        __m128i t0;
+
+        t0 = _mm_unpacklo_epi8(s0, s1);
+        s1 = _mm_unpackhi_epi8(s0, s1);
+        s0 = _mm_unpacklo_epi8(t0, s1);
+        s0 = _mm_unpacklo_epi8(s0, zero);
+
+        p0 = _mm_unpacklo_epi8(p0, zero);
+        p0 = _mm_adds_epu16(s0, p0);
+        p0 = _mm_adds_epu16(p0, one);
+        p0 = _mm_srli_epi16(p0, 1);
+        p0 = _mm_packus_epi16(p0, zero);
+
+        *(int *)comp_pred = _mm_cvtsi128_si32(p0);
+        comp_pred += 4;
+        pred += 4;
+        ref += 4 * 8;
+      }
+      ref += stride - (width << 3);
+    }
+  }
+}
diff --git a/third_party/aom/aom_mem/aom_mem.c b/third_party/aom/aom_mem/aom_mem.c
new file mode 100644
index 0000000000..66a0c08dea
--- /dev/null
+++ b/third_party/aom/aom_mem/aom_mem.c
@@ -0,0 +1,74 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#define __AOM_MEM_C__
+
+#include "aom_mem.h"
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "include/aom_mem_intrnl.h"
+#include "aom/aom_integer.h"
+
+static size_t GetAlignedMallocSize(size_t size, size_t align) {
+  return size + align - 1 + ADDRESS_STORAGE_SIZE;
+}
+
+static size_t *GetMallocAddressLocation(void *const mem) {
+  return ((size_t *)mem) - 1;
+}
+
+static void SetActualMallocAddress(void *const mem,
+                                   const void *const malloc_addr) {
+  size_t *const malloc_addr_location = GetMallocAddressLocation(mem);
+  *malloc_addr_location = (size_t)malloc_addr;
+}
+
+static void *GetActualMallocAddress(void *const mem) {
+  const size_t *const malloc_addr_location = GetMallocAddressLocation(mem);
+  return (void *)(*malloc_addr_location);
+}
+
+void *aom_memalign(size_t align, size_t size) {
+  void *x = NULL;
+  const size_t aligned_size = GetAlignedMallocSize(size, align);
+  void *const addr = malloc(aligned_size);
+  if (addr) {
+    x = align_addr((unsigned char *)addr + ADDRESS_STORAGE_SIZE, align);
+    SetActualMallocAddress(x, addr);
+  }
+  return x;
+}
+
+void *aom_malloc(size_t size) { return aom_memalign(DEFAULT_ALIGNMENT, size); }
+
+void *aom_calloc(size_t num, size_t size) {
+  const size_t total_size = num * size;
+  void *const x = aom_malloc(total_size);
+  if (x) memset(x, 0, total_size);
+  return x;
+}
+
+void aom_free(void *memblk) {
+  if (memblk) {
+    void *addr = GetActualMallocAddress(memblk);
+    free(addr);
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void *aom_memset16(void *dest, int val, size_t length) {
+  size_t i;
+  uint16_t *dest16 = (uint16_t *)dest;
+  for (i = 0; i < length; i++) *dest16++ = val;
+  return dest;
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/aom_mem/aom_mem.cmake b/third_party/aom/aom_mem/aom_mem.cmake
new file mode 100644
index 0000000000..0375b09e06
--- /dev/null
+++ b/third_party/aom/aom_mem/aom_mem.cmake
@@ -0,0 +1,22 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+set(AOM_MEM_SOURCES
+    "${AOM_ROOT}/aom_mem/aom_mem.c"
+    "${AOM_ROOT}/aom_mem/aom_mem.h"
+    "${AOM_ROOT}/aom_mem/include/aom_mem_intrnl.h")
+
+# Creates the aom_mem build target and makes libaom depend on it. The libaom
+# target must exist before this function is called.
+function (setup_aom_mem_targets)
+  add_library(aom_mem OBJECT ${AOM_MEM_SOURCES})
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_mem PARENT_SCOPE)
+  target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_mem>)
+endfunction ()
diff --git a/third_party/aom/aom_mem/aom_mem.h b/third_party/aom/aom_mem/aom_mem.h
new file mode 100644
index 0000000000..75bd4be65c
--- /dev/null
+++ b/third_party/aom/aom_mem/aom_mem.h
@@ -0,0 +1,46 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_MEM_AOM_MEM_H_
+#define AOM_MEM_AOM_MEM_H_
+
+#include "aom_config.h"
+#if defined(__uClinux__)
+#include <lddk.h>
+#endif
+
+#include <stdlib.h>
+#include <stddef.h>
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+void *aom_memalign(size_t align, size_t size);
+void *aom_malloc(size_t size);
+void *aom_calloc(size_t num, size_t size);
+void aom_free(void *memblk);
+
+#if CONFIG_HIGHBITDEPTH
+void *aom_memset16(void *dest, int val, size_t length);
+#endif
+
+#include <string.h>
+
+#ifdef AOM_MEM_PLTFRM
+#include AOM_MEM_PLTFRM
+#endif
+
+#if defined(__cplusplus)
+}
+#endif
+
+#endif  // AOM_MEM_AOM_MEM_H_
diff --git a/third_party/aom/aom_mem/aom_mem.mk b/third_party/aom/aom_mem/aom_mem.mk
new file mode 100644
index 0000000000..e9162c2842
--- /dev/null
+++ b/third_party/aom/aom_mem/aom_mem.mk
@@ -0,0 +1,4 @@
+MEM_SRCS-yes += aom_mem.mk
+MEM_SRCS-yes += aom_mem.c
+MEM_SRCS-yes += aom_mem.h
+MEM_SRCS-yes += include/aom_mem_intrnl.h
diff --git a/third_party/aom/aom_mem/include/aom_mem_intrnl.h b/third_party/aom/aom_mem/include/aom_mem_intrnl.h
new file mode 100644
index 0000000000..3cdfbe08df
--- /dev/null
+++ b/third_party/aom/aom_mem/include/aom_mem_intrnl.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_MEM_INCLUDE_AOM_MEM_INTRNL_H_
+#define AOM_MEM_INCLUDE_AOM_MEM_INTRNL_H_
+#include "./aom_config.h"
+
+#define ADDRESS_STORAGE_SIZE sizeof(size_t)
+
+#ifndef DEFAULT_ALIGNMENT
+#if defined(VXWORKS)
+/*default addr alignment to use in calls to aom_* functions other than
+  aom_memalign*/
+#define DEFAULT_ALIGNMENT 32
+#else
+#define DEFAULT_ALIGNMENT (2 * sizeof(void *)) /* NOLINT */
+#endif
+#endif
+
+/*returns an addr aligned to the byte boundary specified by align*/
+#define align_addr(addr, align) \
+  (void *)(((size_t)(addr) + ((align)-1)) & ~(size_t)((align)-1))
+
+#endif  // AOM_MEM_INCLUDE_AOM_MEM_INTRNL_H_
diff --git a/third_party/aom/aom_ports/aom_once.h b/third_party/aom/aom_ports/aom_once.h
new file mode 100644
index 0000000000..3cfd2fd958
--- /dev/null
+++ b/third_party/aom/aom_ports/aom_once.h
@@ -0,0 +1,141 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_AOM_ONCE_H_
+#define AOM_PORTS_AOM_ONCE_H_
+
+#include "aom_config.h"
+
+/* Implement a function wrapper to guarantee initialization
+ * thread-safety for library singletons.
+ *
+ * NOTE: These functions use static locks, and can only be
+ * used with one common argument per compilation unit. So
+ *
+ * file1.c:
+ *   aom_once(foo);
+ *   ...
+ *   aom_once(foo);
+ *
+ *   file2.c:
+ *     aom_once(bar);
+ *
+ * will ensure foo() and bar() are each called only once, but in
+ *
+ * file1.c:
+ *   aom_once(foo);
+ *   aom_once(bar):
+ *
+ * bar() will never be called because the lock is used up
+ * by the call to foo().
+ */
+
+#if CONFIG_MULTITHREAD && defined(_WIN32)
+#include <windows.h>
+#include <stdlib.h>
+/* Declare a per-compilation-unit state variable to track the progress
+ * of calling func() only once. This must be at global scope because
+ * local initializers are not thread-safe in MSVC prior to Visual
+ * Studio 2015.
+ *
+ * As a static, once_state will be zero-initialized as program start.
+ */
+static LONG once_state;
+static void once(void (*func)(void)) {
+  /* Try to advance once_state from its initial value of 0 to 1.
+   * Only one thread can succeed in doing so.
+   */
+  if (InterlockedCompareExchange(&once_state, 1, 0) == 0) {
+    /* We're the winning thread, having set once_state to 1.
+     * Call our function. */
+    func();
+    /* Now advance once_state to 2, unblocking any other threads. */
+    InterlockedIncrement(&once_state);
+    return;
+  }
+
+  /* We weren't the winning thread, but we want to block on
+   * the state variable so we don't return before func()
+   * has finished executing elsewhere.
+   *
+   * Try to advance once_state from 2 to 2, which is only possible
+   * after the winning thead advances it from 1 to 2.
+   */
+  while (InterlockedCompareExchange(&once_state, 2, 2) != 2) {
+    /* State isn't yet 2. Try again.
+     *
+     * We are used for singleton initialization functions,
+     * which should complete quickly. Contention will likewise
+     * be rare, so it's worthwhile to use a simple but cpu-
+     * intensive busy-wait instead of successive backoff,
+     * waiting on a kernel object, or another heavier-weight scheme.
+     *
+     * We can at least yield our timeslice.
+     */
+    Sleep(0);
+  }
+
+  /* We've seen once_state advance to 2, so we know func()
+   * has been called. And we've left once_state as we found it,
+   * so other threads will have the same experience.
+   *
+   * It's safe to return now.
+   */
+  return;
+}
+
+#elif CONFIG_MULTITHREAD && defined(__OS2__)
+#define INCL_DOS
+#include <os2.h>
+static void once(void (*func)(void)) {
+  static int done;
+
+  /* If the initialization is complete, return early. */
+  if (done) return;
+
+  /* Causes all other threads in the process to block themselves
+   * and give up their time slice.
+   */
+  DosEnterCritSec();
+
+  if (!done) {
+    func();
+    done = 1;
+  }
+
+  /* Restores normal thread dispatching for the current process. */
+  DosExitCritSec();
+}
+
+#elif CONFIG_MULTITHREAD && HAVE_PTHREAD_H
+#include <pthread.h>
+static void once(void (*func)(void)) {
+  static pthread_once_t lock = PTHREAD_ONCE_INIT;
+  pthread_once(&lock, func);
+}
+
+#else
+/* No-op version that performs no synchronization. *_rtcd() is idempotent,
+ * so as long as your platform provides atomic loads/stores of pointers
+ * no synchronization is strictly necessary.
+ */
+
+static void once(void (*func)(void)) {
+  static int done;
+
+  if (!done) {
+    func();
+    done = 1;
+  }
+}
+#endif
+
+#endif  // AOM_PORTS_AOM_ONCE_H_
diff --git a/third_party/aom/aom_ports/aom_ports.cmake b/third_party/aom/aom_ports/aom_ports.cmake
new file mode 100644
index 0000000000..3e8883d564
--- /dev/null
+++ b/third_party/aom/aom_ports/aom_ports.cmake
@@ -0,0 +1,66 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+set(AOM_PORTS_INCLUDES
+    "${AOM_ROOT}/aom_ports/aom_once.h"
+    "${AOM_ROOT}/aom_ports/aom_timer.h"
+    "${AOM_ROOT}/aom_ports/bitops.h"
+    "${AOM_ROOT}/aom_ports/emmintrin_compat.h"
+    "${AOM_ROOT}/aom_ports/mem.h"
+    "${AOM_ROOT}/aom_ports/mem_ops.h"
+    "${AOM_ROOT}/aom_ports/mem_ops_aligned.h"
+    "${AOM_ROOT}/aom_ports/msvc.h"
+    "${AOM_ROOT}/aom_ports/system_state.h")
+
+set(AOM_PORTS_INCLUDES_X86
+    "${AOM_ROOT}/aom_ports/x86_abi_support.asm")
+
+set(AOM_PORTS_ASM_MMX "${AOM_ROOT}/aom_ports/emms.asm")
+
+set(AOM_PORTS_SOURCES_ARM
+    "${AOM_ROOT}/aom_ports/arm.h"
+    "${AOM_ROOT}/aom_ports/arm_cpudetect.c")
+
+# For arm targets and targets where HAVE_MMX is true:
+#   Creates the aom_ports build target, adds the includes in aom_ports to the
+#   target, and makes libaom depend on it.
+# Otherwise:
+#   Adds the includes in aom_ports to the libaom target.
+# For all target platforms:
+#   The libaom target must exist before this function is called.
+function (setup_aom_ports_targets)
+  if (HAVE_MMX)
+    add_asm_library("aom_ports" "AOM_PORTS_ASM_MMX" "aom")
+    set(aom_ports_has_symbols 1)
+  elseif ("${AOM_TARGET_CPU}" MATCHES "arm")
+    add_library(aom_ports OBJECT ${AOM_PORTS_SOURCES_ARM})
+    set(aom_ports_has_symbols 1)
+    list(APPEND AOM_LIB_TARGETS aom_ports)
+    target_sources(aom PRIVATE $<TARGET_OBJECTS:aom_ports>)
+  endif ()
+
+  if (aom_ports_has_symbols)
+    target_sources(aom_ports PUBLIC ${AOM_PORTS_INCLUDES})
+
+    if ("${AOM_TARGET_CPU}" STREQUAL "x86" OR
+        "${AOM_TARGET_CPU}" STREQUAL "x86_64")
+      target_sources(aom_ports PUBLIC ${AOM_PORTS_INCLUDES_X86})
+    endif ()
+
+    set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} PARENT_SCOPE)
+  else ()
+    target_sources(aom PUBLIC ${AOM_PORTS_INCLUDES})
+
+    if ("${AOM_TARGET_CPU}" STREQUAL "x86" OR
+        "${AOM_TARGET_CPU}" STREQUAL "x86_64")
+      target_sources(aom PUBLIC ${AOM_PORTS_INCLUDES_X86})
+    endif ()
+  endif ()
+endfunction ()
diff --git a/third_party/aom/aom_ports/aom_ports.mk b/third_party/aom/aom_ports/aom_ports.mk
new file mode 100644
index 0000000000..1f18f6bd16
--- /dev/null
+++ b/third_party/aom/aom_ports/aom_ports.mk
@@ -0,0 +1,29 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+
+PORTS_SRCS-yes += aom_ports.mk
+
+PORTS_SRCS-yes += bitops.h
+PORTS_SRCS-yes += mem.h
+PORTS_SRCS-yes += msvc.h
+PORTS_SRCS-yes += system_state.h
+PORTS_SRCS-yes += aom_timer.h
+
+ifeq ($(ARCH_X86)$(ARCH_X86_64),yes)
+PORTS_SRCS-yes += emms.asm
+PORTS_SRCS-yes += x86.h
+PORTS_SRCS-yes += x86_abi_support.asm
+endif
+
+PORTS_SRCS-$(ARCH_ARM) += arm_cpudetect.c
+PORTS_SRCS-$(ARCH_ARM) += arm.h
diff --git a/third_party/aom/aom_ports/aom_timer.h b/third_party/aom/aom_ports/aom_timer.h
new file mode 100644
index 0000000000..904f2fe51c
--- /dev/null
+++ b/third_party/aom/aom_ports/aom_timer.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_AOM_TIMER_H_
+#define AOM_PORTS_AOM_TIMER_H_
+
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+
+#if CONFIG_OS_SUPPORT
+
+#if defined(_WIN32)
+/*
+ * Win32 specific includes
+ */
+#ifndef WIN32_LEAN_AND_MEAN
+#define WIN32_LEAN_AND_MEAN
+#endif
+#include <windows.h>
+#else
+/*
+ * POSIX specific includes
+ */
+#include <sys/time.h>
+
+/* timersub is not provided by msys at this time. */
+#ifndef timersub
+#define timersub(a, b, result)                       \
+  do {                                               \
+    (result)->tv_sec = (a)->tv_sec - (b)->tv_sec;    \
+    (result)->tv_usec = (a)->tv_usec - (b)->tv_usec; \
+    if ((result)->tv_usec < 0) {                     \
+      --(result)->tv_sec;                            \
+      (result)->tv_usec += 1000000;                  \
+    }                                                \
+  } while (0)
+#endif
+#endif
+
+struct aom_usec_timer {
+#if defined(_WIN32)
+  LARGE_INTEGER begin, end;
+#else
+  struct timeval begin, end;
+#endif
+};
+
+static INLINE void aom_usec_timer_start(struct aom_usec_timer *t) {
+#if defined(_WIN32)
+  QueryPerformanceCounter(&t->begin);
+#else
+  gettimeofday(&t->begin, NULL);
+#endif
+}
+
+static INLINE void aom_usec_timer_mark(struct aom_usec_timer *t) {
+#if defined(_WIN32)
+  QueryPerformanceCounter(&t->end);
+#else
+  gettimeofday(&t->end, NULL);
+#endif
+}
+
+static INLINE int64_t aom_usec_timer_elapsed(struct aom_usec_timer *t) {
+#if defined(_WIN32)
+  LARGE_INTEGER freq, diff;
+
+  diff.QuadPart = t->end.QuadPart - t->begin.QuadPart;
+
+  QueryPerformanceFrequency(&freq);
+  return diff.QuadPart * 1000000 / freq.QuadPart;
+#else
+  struct timeval diff;
+
+  timersub(&t->end, &t->begin, &diff);
+  return diff.tv_sec * 1000000 + diff.tv_usec;
+#endif
+}
+
+#else /* CONFIG_OS_SUPPORT = 0*/
+
+/* Empty timer functions if CONFIG_OS_SUPPORT = 0 */
+#ifndef timersub
+#define timersub(a, b, result)
+#endif
+
+struct aom_usec_timer {
+  void *dummy;
+};
+
+static INLINE void aom_usec_timer_start(struct aom_usec_timer *t) { (void)t; }
+
+static INLINE void aom_usec_timer_mark(struct aom_usec_timer *t) { (void)t; }
+
+static INLINE int aom_usec_timer_elapsed(struct aom_usec_timer *t) {
+  (void)t;
+  return 0;
+}
+
+#endif /* CONFIG_OS_SUPPORT */
+
+#endif  // AOM_PORTS_AOM_TIMER_H_
diff --git a/third_party/aom/aom_ports/arm.h b/third_party/aom/aom_ports/arm.h
new file mode 100644
index 0000000000..448a70dcc6
--- /dev/null
+++ b/third_party/aom/aom_ports/arm.h
@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_ARM_H_
+#define AOM_PORTS_ARM_H_
+#include <stdlib.h>
+#include "aom_config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*ARMv5TE "Enhanced DSP" instructions.*/
+#define HAS_EDSP 0x01
+/*ARMv6 "Parallel" or "Media" instructions.*/
+#define HAS_MEDIA 0x02
+/*ARMv7 optional NEON instructions.*/
+#define HAS_NEON 0x04
+
+int arm_cpu_caps(void);
+
+// Earlier gcc compilers have issues with some neon intrinsics
+#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ == 4 && \
+    __GNUC_MINOR__ <= 6
+#define AOM_INCOMPATIBLE_GCC
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_PORTS_ARM_H_
diff --git a/third_party/aom/aom_ports/arm_cpudetect.c b/third_party/aom/aom_ports/arm_cpudetect.c
new file mode 100644
index 0000000000..65ba846c9c
--- /dev/null
+++ b/third_party/aom/aom_ports/arm_cpudetect.c
@@ -0,0 +1,176 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+#include "aom_ports/arm.h"
+#include "./aom_config.h"
+
+#ifdef WINAPI_FAMILY
+#include <winapifamily.h>
+#if !WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
+#define getenv(x) NULL
+#endif
+#endif
+
+static int arm_cpu_env_flags(int *flags) {
+  char *env;
+  env = getenv("AOM_SIMD_CAPS");
+  if (env && *env) {
+    *flags = (int)strtol(env, NULL, 0);
+    return 0;
+  }
+  *flags = 0;
+  return -1;
+}
+
+static int arm_cpu_env_mask(void) {
+  char *env;
+  env = getenv("AOM_SIMD_CAPS_MASK");
+  return env && *env ? (int)strtol(env, NULL, 0) : ~0;
+}
+
+#if !CONFIG_RUNTIME_CPU_DETECT
+
+int arm_cpu_caps(void) {
+  /* This function should actually be a no-op. There is no way to adjust any of
+   * these because the RTCD tables do not exist: the functions are called
+   * statically */
+  int flags;
+  int mask;
+  if (!arm_cpu_env_flags(&flags)) {
+    return flags;
+  }
+  mask = arm_cpu_env_mask();
+#if HAVE_MEDIA
+  flags |= HAS_MEDIA;
+#endif /* HAVE_MEDIA */
+#if HAVE_NEON || HAVE_NEON_ASM
+  flags |= HAS_NEON;
+#endif /* HAVE_NEON  || HAVE_NEON_ASM */
+  return flags & mask;
+}
+
+#elif defined(_MSC_VER) /* end !CONFIG_RUNTIME_CPU_DETECT */
+/*For GetExceptionCode() and EXCEPTION_ILLEGAL_INSTRUCTION.*/
+#define WIN32_LEAN_AND_MEAN
+#define WIN32_EXTRA_LEAN
+#include <windows.h>
+
+int arm_cpu_caps(void) {
+  int flags;
+  int mask;
+  if (!arm_cpu_env_flags(&flags)) {
+    return flags;
+  }
+  mask = arm_cpu_env_mask();
+/* MSVC has no inline __asm support for ARM, but it does let you __emit
+ *  instructions via their assembled hex code.
+ * All of these instructions should be essentially nops.
+ */
+#if HAVE_MEDIA
+  if (mask & HAS_MEDIA) {
+    __try {
+      /*SHADD8 r3,r3,r3*/
+      __emit(0xE6333F93);
+      flags |= HAS_MEDIA;
+    } __except (GetExceptionCode() == EXCEPTION_ILLEGAL_INSTRUCTION) {
+      /*Ignore exception.*/
+    }
+  }
+#endif /* HAVE_MEDIA */
+#if HAVE_NEON || HAVE_NEON_ASM
+  if (mask & HAS_NEON) {
+    __try {
+      /*VORR q0,q0,q0*/
+      __emit(0xF2200150);
+      flags |= HAS_NEON;
+    } __except (GetExceptionCode() == EXCEPTION_ILLEGAL_INSTRUCTION) {
+      /*Ignore exception.*/
+    }
+  }
+#endif /* HAVE_NEON || HAVE_NEON_ASM */
+  return flags & mask;
+}
+
+#elif defined(__ANDROID__) /* end _MSC_VER */
+#include <cpu-features.h>
+
+int arm_cpu_caps(void) {
+  int flags;
+  int mask;
+  uint64_t features;
+  if (!arm_cpu_env_flags(&flags)) {
+    return flags;
+  }
+  mask = arm_cpu_env_mask();
+  features = android_getCpuFeatures();
+
+#if HAVE_MEDIA
+  flags |= HAS_MEDIA;
+#endif /* HAVE_MEDIA */
+#if HAVE_NEON || HAVE_NEON_ASM
+  if (features & ANDROID_CPU_ARM_FEATURE_NEON) flags |= HAS_NEON;
+#endif /* HAVE_NEON || HAVE_NEON_ASM */
+  return flags & mask;
+}
+
+#elif defined(__linux__) /* end __ANDROID__ */
+
+#include <stdio.h>
+
+int arm_cpu_caps(void) {
+  FILE *fin;
+  int flags;
+  int mask;
+  if (!arm_cpu_env_flags(&flags)) {
+    return flags;
+  }
+  mask = arm_cpu_env_mask();
+  /* Reading /proc/self/auxv would be easier, but that doesn't work reliably
+   *  on Android.
+   * This also means that detection will fail in Scratchbox.
+   */
+  fin = fopen("/proc/cpuinfo", "r");
+  if (fin != NULL) {
+    /* 512 should be enough for anybody (it's even enough for all the flags
+     * that x86 has accumulated... so far).
+     */
+    char buf[512];
+    while (fgets(buf, 511, fin) != NULL) {
+#if HAVE_NEON || HAVE_NEON_ASM
+      if (memcmp(buf, "Features", 8) == 0) {
+        char *p;
+        p = strstr(buf, " neon");
+        if (p != NULL && (p[5] == ' ' || p[5] == '\n')) {
+          flags |= HAS_NEON;
+        }
+      }
+#endif /* HAVE_NEON || HAVE_NEON_ASM */
+#if HAVE_MEDIA
+      if (memcmp(buf, "CPU architecture:", 17) == 0) {
+        int version;
+        version = atoi(buf + 17);
+        if (version >= 6) {
+          flags |= HAS_MEDIA;
+        }
+      }
+#endif /* HAVE_MEDIA */
+    }
+    fclose(fin);
+  }
+  return flags & mask;
+}
+#else  /* end __linux__ */
+#error \
+    "--enable-runtime-cpu-detect selected, but no CPU detection method " \
+"available for your platform. Reconfigure with --disable-runtime-cpu-detect."
+#endif
diff --git a/third_party/aom/aom_ports/bitops.h b/third_party/aom/aom_ports/bitops.h
new file mode 100644
index 0000000000..36f5bd487b
--- /dev/null
+++ b/third_party/aom/aom_ports/bitops.h
@@ -0,0 +1,76 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_BITOPS_H_
+#define AOM_PORTS_BITOPS_H_
+
+#include <assert.h>
+
+#include "aom_ports/msvc.h"
+
+#ifdef _MSC_VER
+#if defined(_M_X64) || defined(_M_IX86)
+#include <intrin.h>
+#define USE_MSC_INTRINSICS
+#endif
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// These versions of get_msb() are only valid when n != 0 because all
+// of the optimized versions are undefined when n == 0:
+// https://gcc.gnu.org/onlinedocs/gcc/Other-Builtins.html
+
+// use GNU builtins where available.
+#if defined(__GNUC__) && \
+    ((__GNUC__ == 3 && __GNUC_MINOR__ >= 4) || __GNUC__ >= 4)
+static INLINE int get_msb(unsigned int n) {
+  assert(n != 0);
+  return 31 ^ __builtin_clz(n);
+}
+#elif defined(USE_MSC_INTRINSICS)
+#pragma intrinsic(_BitScanReverse)
+
+static INLINE int get_msb(unsigned int n) {
+  unsigned long first_set_bit;
+  assert(n != 0);
+  _BitScanReverse(&first_set_bit, n);
+  return first_set_bit;
+}
+#undef USE_MSC_INTRINSICS
+#else
+// Returns (int)floor(log2(n)). n must be > 0.
+static INLINE int get_msb(unsigned int n) {
+  int log = 0;
+  unsigned int value = n;
+  int i;
+
+  assert(n != 0);
+
+  for (i = 4; i >= 0; --i) {
+    const int shift = (1 << i);
+    const unsigned int x = value >> shift;
+    if (x != 0) {
+      value = x;
+      log += shift;
+    }
+  }
+  return log;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_PORTS_BITOPS_H_
diff --git a/third_party/aom/aom_ports/config.h b/third_party/aom/aom_ports/config.h
new file mode 100644
index 0000000000..462ec66fb0
--- /dev/null
+++ b/third_party/aom/aom_ports/config.h
@@ -0,0 +1,17 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_CONFIG_H_
+#define AOM_PORTS_CONFIG_H_
+
+#include "aom_config.h"
+
+#endif  // AOM_PORTS_CONFIG_H_
diff --git a/third_party/aom/aom_ports/emmintrin_compat.h b/third_party/aom/aom_ports/emmintrin_compat.h
new file mode 100644
index 0000000000..f9d44c647e
--- /dev/null
+++ b/third_party/aom/aom_ports/emmintrin_compat.h
@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_EMMINTRIN_COMPAT_H_
+#define AOM_PORTS_EMMINTRIN_COMPAT_H_
+
+#if defined(__GNUC__) && __GNUC__ < 4
+/* From emmintrin.h (gcc 4.5.3) */
+/* Casts between various SP, DP, INT vector types.  Note that these do no
+   conversion of values, they just change the type.  */
+extern __inline __m128
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+    _mm_castpd_ps(__m128d __A) {
+  return (__m128)__A;
+}
+
+extern __inline __m128i
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+    _mm_castpd_si128(__m128d __A) {
+  return (__m128i)__A;
+}
+
+extern __inline __m128d
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+    _mm_castps_pd(__m128 __A) {
+  return (__m128d)__A;
+}
+
+extern __inline __m128i
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+    _mm_castps_si128(__m128 __A) {
+  return (__m128i)__A;
+}
+
+extern __inline __m128
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+    _mm_castsi128_ps(__m128i __A) {
+  return (__m128)__A;
+}
+
+extern __inline __m128d
+    __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+    _mm_castsi128_pd(__m128i __A) {
+  return (__m128d)__A;
+}
+#endif
+
+#endif  // AOM_PORTS_EMMINTRIN_COMPAT_H_
diff --git a/third_party/aom/aom_ports/emms.asm b/third_party/aom/aom_ports/emms.asm
new file mode 100644
index 0000000000..90776bacb3
--- /dev/null
+++ b/third_party/aom/aom_ports/emms.asm
@@ -0,0 +1,41 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+%include "aom_ports/x86_abi_support.asm"
+
+section .text
+global sym(aom_reset_mmx_state) PRIVATE
+sym(aom_reset_mmx_state):
+    emms
+    ret
+
+
+%if LIBAOM_YASM_WIN64
+global sym(aom_winx64_fldcw) PRIVATE
+sym(aom_winx64_fldcw):
+    sub   rsp, 8
+    mov   [rsp], rcx ; win x64 specific
+    fldcw [rsp]
+    add   rsp, 8
+    ret
+
+
+global sym(aom_winx64_fstcw) PRIVATE
+sym(aom_winx64_fstcw):
+    sub   rsp, 8
+    fstcw [rsp]
+    mov   rax, [rsp]
+    add   rsp, 8
+    ret
+%endif
diff --git a/third_party/aom/aom_ports/mem.h b/third_party/aom/aom_ports/mem.h
new file mode 100644
index 0000000000..bd881cdc8c
--- /dev/null
+++ b/third_party/aom/aom_ports/mem.h
@@ -0,0 +1,65 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_MEM_H_
+#define AOM_PORTS_MEM_H_
+
+#include "aom_config.h"
+#include "aom/aom_integer.h"
+
+#if (defined(__GNUC__) && __GNUC__) || defined(__SUNPRO_C)
+#define DECLARE_ALIGNED(n, typ, val) typ val __attribute__((aligned(n)))
+#elif defined(_MSC_VER)
+#define DECLARE_ALIGNED(n, typ, val) __declspec(align(n)) typ val
+#else
+#warning No alignment directives known for this compiler.
+#define DECLARE_ALIGNED(n, typ, val) typ val
+#endif
+
+/* Indicates that the usage of the specified variable has been audited to assure
+ * that it's safe to use uninitialized. Silences 'may be used uninitialized'
+ * warnings on gcc.
+ */
+#if defined(__GNUC__) && __GNUC__
+#define UNINITIALIZED_IS_SAFE(x) x = x
+#else
+#define UNINITIALIZED_IS_SAFE(x) x
+#endif
+
+#if HAVE_NEON && defined(_MSC_VER)
+#define __builtin_prefetch(x)
+#endif
+
+/* Shift down with rounding for use when n >= 0, value >= 0 */
+#define ROUND_POWER_OF_TWO(value, n) (((value) + (((1 << (n)) >> 1))) >> (n))
+
+/* Shift down with rounding for signed integers, for use when n >= 0 */
+#define ROUND_POWER_OF_TWO_SIGNED(value, n)           \
+  (((value) < 0) ? -ROUND_POWER_OF_TWO(-(value), (n)) \
+                 : ROUND_POWER_OF_TWO((value), (n)))
+
+/* Shift down with rounding for use when n >= 0, value >= 0 for (64 bit) */
+#define ROUND_POWER_OF_TWO_64(value, n) \
+  (((value) + ((((int64_t)1 << (n)) >> 1))) >> (n))
+/* Shift down with rounding for signed integers, for use when n >= 0 (64 bit) */
+#define ROUND_POWER_OF_TWO_SIGNED_64(value, n)           \
+  (((value) < 0) ? -ROUND_POWER_OF_TWO_64(-(value), (n)) \
+                 : ROUND_POWER_OF_TWO_64((value), (n)))
+
+#define ALIGN_POWER_OF_TWO(value, n) \
+  (((value) + ((1 << (n)) - 1)) & ~((1 << (n)) - 1))
+
+#define CONVERT_TO_SHORTPTR(x) ((uint16_t *)(((uintptr_t)(x)) << 1))
+#if CONFIG_HIGHBITDEPTH
+#define CONVERT_TO_BYTEPTR(x) ((uint8_t *)(((uintptr_t)(x)) >> 1))
+#endif  // CONFIG_HIGHBITDEPTH
+
+#endif  // AOM_PORTS_MEM_H_
diff --git a/third_party/aom/aom_ports/mem_ops.h b/third_party/aom/aom_ports/mem_ops.h
new file mode 100644
index 0000000000..62126755c8
--- /dev/null
+++ b/third_party/aom/aom_ports/mem_ops.h
@@ -0,0 +1,229 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_MEM_OPS_H_
+#define AOM_PORTS_MEM_OPS_H_
+
+/* \file
+ * \brief Provides portable memory access primitives
+ *
+ * This function provides portable primitives for getting and setting of
+ * signed and unsigned integers in 16, 24, and 32 bit sizes. The operations
+ * can be performed on unaligned data regardless of hardware support for
+ * unaligned accesses.
+ *
+ * The type used to pass the integral values may be changed by defining
+ * MEM_VALUE_T with the appropriate type. The type given must be an integral
+ * numeric type.
+ *
+ * The actual functions instantiated have the MEM_VALUE_T type name pasted
+ * on to the symbol name. This allows the developer to instantiate these
+ * operations for multiple types within the same translation unit. This is
+ * of somewhat questionable utility, but the capability exists nonetheless.
+ * Users not making use of this functionality should call the functions
+ * without the type name appended, and the preprocessor will take care of
+ * it.
+ *
+ * NOTE: This code is not supported on platforms where char > 1 octet ATM.
+ */
+
+#ifndef MAU_T
+/* Minimum Access Unit for this target */
+#define MAU_T unsigned char
+#endif
+
+#ifndef MEM_VALUE_T
+#define MEM_VALUE_T int
+#endif
+
+#undef MEM_VALUE_T_SZ_BITS
+#define MEM_VALUE_T_SZ_BITS (sizeof(MEM_VALUE_T) << 3)
+
+#undef mem_ops_wrap_symbol
+#define mem_ops_wrap_symbol(fn) mem_ops_wrap_symbol2(fn, MEM_VALUE_T)
+#undef mem_ops_wrap_symbol2
+#define mem_ops_wrap_symbol2(fn, typ) mem_ops_wrap_symbol3(fn, typ)
+#undef mem_ops_wrap_symbol3
+#define mem_ops_wrap_symbol3(fn, typ) fn##_as_##typ
+
+/*
+ * Include aligned access routines
+ */
+#define INCLUDED_BY_MEM_OPS_H
+#include "mem_ops_aligned.h"
+#undef INCLUDED_BY_MEM_OPS_H
+
+#undef mem_get_be16
+#define mem_get_be16 mem_ops_wrap_symbol(mem_get_be16)
+static unsigned MEM_VALUE_T mem_get_be16(const void *vmem) {
+  unsigned MEM_VALUE_T val;
+  const MAU_T *mem = (const MAU_T *)vmem;
+
+  val = mem[0] << 8;
+  val |= mem[1];
+  return val;
+}
+
+#undef mem_get_be24
+#define mem_get_be24 mem_ops_wrap_symbol(mem_get_be24)
+static unsigned MEM_VALUE_T mem_get_be24(const void *vmem) {
+  unsigned MEM_VALUE_T val;
+  const MAU_T *mem = (const MAU_T *)vmem;
+
+  val = mem[0] << 16;
+  val |= mem[1] << 8;
+  val |= mem[2];
+  return val;
+}
+
+#undef mem_get_be32
+#define mem_get_be32 mem_ops_wrap_symbol(mem_get_be32)
+static unsigned MEM_VALUE_T mem_get_be32(const void *vmem) {
+  unsigned MEM_VALUE_T val;
+  const MAU_T *mem = (const MAU_T *)vmem;
+
+  val = ((unsigned MEM_VALUE_T)mem[0]) << 24;
+  val |= mem[1] << 16;
+  val |= mem[2] << 8;
+  val |= mem[3];
+  return val;
+}
+
+#undef mem_get_le16
+#define mem_get_le16 mem_ops_wrap_symbol(mem_get_le16)
+static unsigned MEM_VALUE_T mem_get_le16(const void *vmem) {
+  unsigned MEM_VALUE_T val;
+  const MAU_T *mem = (const MAU_T *)vmem;
+
+  val = mem[1] << 8;
+  val |= mem[0];
+  return val;
+}
+
+#undef mem_get_le24
+#define mem_get_le24 mem_ops_wrap_symbol(mem_get_le24)
+static unsigned MEM_VALUE_T mem_get_le24(const void *vmem) {
+  unsigned MEM_VALUE_T val;
+  const MAU_T *mem = (const MAU_T *)vmem;
+
+  val = mem[2] << 16;
+  val |= mem[1] << 8;
+  val |= mem[0];
+  return val;
+}
+
+#undef mem_get_le32
+#define mem_get_le32 mem_ops_wrap_symbol(mem_get_le32)
+static unsigned MEM_VALUE_T mem_get_le32(const void *vmem) {
+  unsigned MEM_VALUE_T val;
+  const MAU_T *mem = (const MAU_T *)vmem;
+
+  val = ((unsigned MEM_VALUE_T)mem[3]) << 24;
+  val |= mem[2] << 16;
+  val |= mem[1] << 8;
+  val |= mem[0];
+  return val;
+}
+
+#define mem_get_s_generic(end, sz)                                            \
+  static AOM_INLINE signed MEM_VALUE_T mem_get_s##end##sz(const void *vmem) { \
+    const MAU_T *mem = (const MAU_T *)vmem;                                   \
+    signed MEM_VALUE_T val = mem_get_##end##sz(mem);                          \
+    return (val << (MEM_VALUE_T_SZ_BITS - sz)) >> (MEM_VALUE_T_SZ_BITS - sz); \
+  }
+
+/* clang-format off */
+#undef  mem_get_sbe16
+#define mem_get_sbe16 mem_ops_wrap_symbol(mem_get_sbe16)
+mem_get_s_generic(be, 16)
+
+#undef  mem_get_sbe24
+#define mem_get_sbe24 mem_ops_wrap_symbol(mem_get_sbe24)
+mem_get_s_generic(be, 24)
+
+#undef  mem_get_sbe32
+#define mem_get_sbe32 mem_ops_wrap_symbol(mem_get_sbe32)
+mem_get_s_generic(be, 32)
+
+#undef  mem_get_sle16
+#define mem_get_sle16 mem_ops_wrap_symbol(mem_get_sle16)
+mem_get_s_generic(le, 16)
+
+#undef  mem_get_sle24
+#define mem_get_sle24 mem_ops_wrap_symbol(mem_get_sle24)
+mem_get_s_generic(le, 24)
+
+#undef  mem_get_sle32
+#define mem_get_sle32 mem_ops_wrap_symbol(mem_get_sle32)
+mem_get_s_generic(le, 32)
+
+#undef  mem_put_be16
+#define mem_put_be16 mem_ops_wrap_symbol(mem_put_be16)
+static AOM_INLINE void mem_put_be16(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >> 8) & 0xff);
+  mem[1] = (MAU_T)((val >> 0) & 0xff);
+}
+
+#undef  mem_put_be24
+#define mem_put_be24 mem_ops_wrap_symbol(mem_put_be24)
+static AOM_INLINE void mem_put_be24(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >> 16) & 0xff);
+  mem[1] = (MAU_T)((val >>  8) & 0xff);
+  mem[2] = (MAU_T)((val >>  0) & 0xff);
+}
+
+#undef  mem_put_be32
+#define mem_put_be32 mem_ops_wrap_symbol(mem_put_be32)
+static AOM_INLINE void mem_put_be32(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >> 24) & 0xff);
+  mem[1] = (MAU_T)((val >> 16) & 0xff);
+  mem[2] = (MAU_T)((val >>  8) & 0xff);
+  mem[3] = (MAU_T)((val >>  0) & 0xff);
+}
+
+#undef  mem_put_le16
+#define mem_put_le16 mem_ops_wrap_symbol(mem_put_le16)
+static AOM_INLINE void mem_put_le16(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >> 0) & 0xff);
+  mem[1] = (MAU_T)((val >> 8) & 0xff);
+}
+
+#undef  mem_put_le24
+#define mem_put_le24 mem_ops_wrap_symbol(mem_put_le24)
+static AOM_INLINE void mem_put_le24(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >>  0) & 0xff);
+  mem[1] = (MAU_T)((val >>  8) & 0xff);
+  mem[2] = (MAU_T)((val >> 16) & 0xff);
+}
+
+#undef  mem_put_le32
+#define mem_put_le32 mem_ops_wrap_symbol(mem_put_le32)
+static AOM_INLINE void mem_put_le32(void *vmem, MEM_VALUE_T val) {
+  MAU_T *mem = (MAU_T *)vmem;
+
+  mem[0] = (MAU_T)((val >>  0) & 0xff);
+  mem[1] = (MAU_T)((val >>  8) & 0xff);
+  mem[2] = (MAU_T)((val >> 16) & 0xff);
+  mem[3] = (MAU_T)((val >> 24) & 0xff);
+}
+/* clang-format on */
+
+#endif  // AOM_PORTS_MEM_OPS_H_
diff --git a/third_party/aom/aom_ports/mem_ops_aligned.h b/third_party/aom/aom_ports/mem_ops_aligned.h
new file mode 100644
index 0000000000..8c3ab1cb1a
--- /dev/null
+++ b/third_party/aom/aom_ports/mem_ops_aligned.h
@@ -0,0 +1,172 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_MEM_OPS_ALIGNED_H_
+#define AOM_PORTS_MEM_OPS_ALIGNED_H_
+
+#include "aom/aom_integer.h"
+
+/* \file
+ * \brief Provides portable memory access primitives for operating on aligned
+ *        data
+ *
+ * This file is split from mem_ops.h for easier maintenance. See mem_ops.h
+ * for a more detailed description of these primitives.
+ */
+#ifndef INCLUDED_BY_MEM_OPS_H
+#error Include mem_ops.h, not mem_ops_aligned.h directly.
+#endif
+
+/* Architectures that provide instructions for doing this byte swapping
+ * could redefine these macros.
+ */
+#define swap_endian_16(val, raw)                                     \
+  do {                                                               \
+    val = (uint16_t)(((raw >> 8) & 0x00ff) | ((raw << 8) & 0xff00)); \
+  } while (0)
+#define swap_endian_32(val, raw)                                   \
+  do {                                                             \
+    val = ((raw >> 24) & 0x000000ff) | ((raw >> 8) & 0x0000ff00) | \
+          ((raw << 8) & 0x00ff0000) | ((raw << 24) & 0xff000000);  \
+  } while (0)
+#define swap_endian_16_se(val, raw) \
+  do {                              \
+    swap_endian_16(val, raw);       \
+    val = ((val << 16) >> 16);      \
+  } while (0)
+#define swap_endian_32_se(val, raw) swap_endian_32(val, raw)
+
+#define mem_get_ne_aligned_generic(end, sz)                           \
+  static AOM_INLINE unsigned MEM_VALUE_T mem_get_##end##sz##_aligned( \
+      const void *vmem) {                                             \
+    const uint##sz##_t *mem = (const uint##sz##_t *)vmem;             \
+    return *mem;                                                      \
+  }
+
+#define mem_get_sne_aligned_generic(end, sz)                         \
+  static AOM_INLINE signed MEM_VALUE_T mem_get_s##end##sz##_aligned( \
+      const void *vmem) {                                            \
+    const int##sz##_t *mem = (const int##sz##_t *)vmem;              \
+    return *mem;                                                     \
+  }
+
+#define mem_get_se_aligned_generic(end, sz)                           \
+  static AOM_INLINE unsigned MEM_VALUE_T mem_get_##end##sz##_aligned( \
+      const void *vmem) {                                             \
+    const uint##sz##_t *mem = (const uint##sz##_t *)vmem;             \
+    unsigned MEM_VALUE_T val, raw = *mem;                             \
+    swap_endian_##sz(val, raw);                                       \
+    return val;                                                       \
+  }
+
+#define mem_get_sse_aligned_generic(end, sz)                         \
+  static AOM_INLINE signed MEM_VALUE_T mem_get_s##end##sz##_aligned( \
+      const void *vmem) {                                            \
+    const int##sz##_t *mem = (const int##sz##_t *)vmem;              \
+    unsigned MEM_VALUE_T val, raw = *mem;                            \
+    swap_endian_##sz##_se(val, raw);                                 \
+    return val;                                                      \
+  }
+
+#define mem_put_ne_aligned_generic(end, sz)                             \
+  static AOM_INLINE void mem_put_##end##sz##_aligned(void *vmem,        \
+                                                     MEM_VALUE_T val) { \
+    uint##sz##_t *mem = (uint##sz##_t *)vmem;                           \
+    *mem = (uint##sz##_t)val;                                           \
+  }
+
+#define mem_put_se_aligned_generic(end, sz)                             \
+  static AOM_INLINE void mem_put_##end##sz##_aligned(void *vmem,        \
+                                                     MEM_VALUE_T val) { \
+    uint##sz##_t *mem = (uint##sz##_t *)vmem, raw;                      \
+    swap_endian_##sz(raw, val);                                         \
+    *mem = (uint##sz##_t)raw;                                           \
+  }
+
+#include "aom_config.h"
+#if CONFIG_BIG_ENDIAN
+#define mem_get_be_aligned_generic(sz) mem_get_ne_aligned_generic(be, sz)
+#define mem_get_sbe_aligned_generic(sz) mem_get_sne_aligned_generic(be, sz)
+#define mem_get_le_aligned_generic(sz) mem_get_se_aligned_generic(le, sz)
+#define mem_get_sle_aligned_generic(sz) mem_get_sse_aligned_generic(le, sz)
+#define mem_put_be_aligned_generic(sz) mem_put_ne_aligned_generic(be, sz)
+#define mem_put_le_aligned_generic(sz) mem_put_se_aligned_generic(le, sz)
+#else
+#define mem_get_be_aligned_generic(sz) mem_get_se_aligned_generic(be, sz)
+#define mem_get_sbe_aligned_generic(sz) mem_get_sse_aligned_generic(be, sz)
+#define mem_get_le_aligned_generic(sz) mem_get_ne_aligned_generic(le, sz)
+#define mem_get_sle_aligned_generic(sz) mem_get_sne_aligned_generic(le, sz)
+#define mem_put_be_aligned_generic(sz) mem_put_se_aligned_generic(be, sz)
+#define mem_put_le_aligned_generic(sz) mem_put_ne_aligned_generic(le, sz)
+#endif
+
+/* clang-format off */
+#undef  mem_get_be16_aligned
+#define mem_get_be16_aligned mem_ops_wrap_symbol(mem_get_be16_aligned)
+mem_get_be_aligned_generic(16)
+
+#undef  mem_get_be32_aligned
+#define mem_get_be32_aligned mem_ops_wrap_symbol(mem_get_be32_aligned)
+mem_get_be_aligned_generic(32)
+
+#undef  mem_get_le16_aligned
+#define mem_get_le16_aligned mem_ops_wrap_symbol(mem_get_le16_aligned)
+mem_get_le_aligned_generic(16)
+
+#undef  mem_get_le32_aligned
+#define mem_get_le32_aligned mem_ops_wrap_symbol(mem_get_le32_aligned)
+mem_get_le_aligned_generic(32)
+
+#undef  mem_get_sbe16_aligned
+#define mem_get_sbe16_aligned mem_ops_wrap_symbol(mem_get_sbe16_aligned)
+mem_get_sbe_aligned_generic(16)
+
+#undef  mem_get_sbe32_aligned
+#define mem_get_sbe32_aligned mem_ops_wrap_symbol(mem_get_sbe32_aligned)
+mem_get_sbe_aligned_generic(32)
+
+#undef  mem_get_sle16_aligned
+#define mem_get_sle16_aligned mem_ops_wrap_symbol(mem_get_sle16_aligned)
+mem_get_sle_aligned_generic(16)
+
+#undef  mem_get_sle32_aligned
+#define mem_get_sle32_aligned mem_ops_wrap_symbol(mem_get_sle32_aligned)
+mem_get_sle_aligned_generic(32)
+
+#undef  mem_put_be16_aligned
+#define mem_put_be16_aligned mem_ops_wrap_symbol(mem_put_be16_aligned)
+mem_put_be_aligned_generic(16)
+
+#undef  mem_put_be32_aligned
+#define mem_put_be32_aligned mem_ops_wrap_symbol(mem_put_be32_aligned)
+mem_put_be_aligned_generic(32)
+
+#undef  mem_put_le16_aligned
+#define mem_put_le16_aligned mem_ops_wrap_symbol(mem_put_le16_aligned)
+mem_put_le_aligned_generic(16)
+
+#undef  mem_put_le32_aligned
+#define mem_put_le32_aligned mem_ops_wrap_symbol(mem_put_le32_aligned)
+mem_put_le_aligned_generic(32)
+
+#undef mem_get_ne_aligned_generic
+#undef mem_get_se_aligned_generic
+#undef mem_get_sne_aligned_generic
+#undef mem_get_sse_aligned_generic
+#undef mem_put_ne_aligned_generic
+#undef mem_put_se_aligned_generic
+#undef swap_endian_16
+#undef swap_endian_32
+#undef swap_endian_16_se
+#undef swap_endian_32_se
+/* clang-format on */
+
+#endif  // AOM_PORTS_MEM_OPS_ALIGNED_H_
diff --git a/third_party/aom/aom_ports/msvc.h b/third_party/aom/aom_ports/msvc.h
new file mode 100644
index 0000000000..2d3ab9b653
--- /dev/null
+++ b/third_party/aom/aom_ports/msvc.h
@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_MSVC_H_
+#define AOM_PORTS_MSVC_H_
+#ifdef _MSC_VER
+
+#include "./aom_config.h"
+
+#if _MSC_VER < 1900  // VS2015 provides snprintf
+#define snprintf _snprintf
+#endif  // _MSC_VER < 1900
+
+#if _MSC_VER < 1800  // VS2013 provides round
+#include <math.h>
+static INLINE double round(double x) {
+  if (x < 0)
+    return ceil(x - 0.5);
+  else
+    return floor(x + 0.5);
+}
+
+static INLINE float roundf(float x) {
+  if (x < 0)
+    return (float)ceil(x - 0.5f);
+  else
+    return (float)floor(x + 0.5f);
+}
+
+static INLINE long lroundf(float x) {
+  if (x < 0)
+    return (long)(x - 0.5f);
+  else
+    return (long)(x + 0.5f);
+}
+#endif  // _MSC_VER < 1800
+
+#endif  // _MSC_VER
+#endif  // AOM_PORTS_MSVC_H_
diff --git a/third_party/aom/aom_ports/system_state.h b/third_party/aom/aom_ports/system_state.h
new file mode 100644
index 0000000000..5d40d4cb82
--- /dev/null
+++ b/third_party/aom/aom_ports/system_state.h
@@ -0,0 +1,23 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_SYSTEM_STATE_H_
+#define AOM_PORTS_SYSTEM_STATE_H_
+
+#include "./aom_config.h"
+
+#if ARCH_X86 || ARCH_X86_64
+void aom_reset_mmx_state(void);
+#define aom_clear_system_state() aom_reset_mmx_state()
+#else
+#define aom_clear_system_state()
+#endif  // ARCH_X86 || ARCH_X86_64
+#endif  // AOM_PORTS_SYSTEM_STATE_H_
diff --git a/third_party/aom/aom_ports/x86.h b/third_party/aom/aom_ports/x86.h
new file mode 100644
index 0000000000..e5680ca934
--- /dev/null
+++ b/third_party/aom/aom_ports/x86.h
@@ -0,0 +1,322 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_PORTS_X86_H_
+#define AOM_PORTS_X86_H_
+#include <stdlib.h>
+
+#if defined(_MSC_VER)
+#include <intrin.h> /* For __cpuidex, __rdtsc */
+#endif
+
+#include "aom_config.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef enum {
+  AOM_CPU_UNKNOWN = -1,
+  AOM_CPU_AMD,
+  AOM_CPU_AMD_OLD,
+  AOM_CPU_CENTAUR,
+  AOM_CPU_CYRIX,
+  AOM_CPU_INTEL,
+  AOM_CPU_NEXGEN,
+  AOM_CPU_NSC,
+  AOM_CPU_RISE,
+  AOM_CPU_SIS,
+  AOM_CPU_TRANSMETA,
+  AOM_CPU_TRANSMETA_OLD,
+  AOM_CPU_UMC,
+  AOM_CPU_VIA,
+
+  AOM_CPU_LAST
+} aom_cpu_t;
+
+#if defined(__GNUC__) && __GNUC__ || defined(__ANDROID__)
+#if ARCH_X86_64
+#define cpuid(func, func2, ax, bx, cx, dx)                      \
+  __asm__ __volatile__("cpuid           \n\t"                   \
+                       : "=a"(ax), "=b"(bx), "=c"(cx), "=d"(dx) \
+                       : "a"(func), "c"(func2));
+#else
+#define cpuid(func, func2, ax, bx, cx, dx)     \
+  __asm__ __volatile__(                        \
+      "mov %%ebx, %%edi   \n\t"                \
+      "cpuid              \n\t"                \
+      "xchg %%edi, %%ebx  \n\t"                \
+      : "=a"(ax), "=D"(bx), "=c"(cx), "=d"(dx) \
+      : "a"(func), "c"(func2));
+#endif
+#elif defined(__SUNPRO_C) || \
+    defined(__SUNPRO_CC) /* end __GNUC__ or __ANDROID__*/
+#if ARCH_X86_64
+#define cpuid(func, func2, ax, bx, cx, dx)     \
+  asm volatile(                                \
+      "xchg %rsi, %rbx \n\t"                   \
+      "cpuid           \n\t"                   \
+      "movl %ebx, %edi \n\t"                   \
+      "xchg %rsi, %rbx \n\t"                   \
+      : "=a"(ax), "=D"(bx), "=c"(cx), "=d"(dx) \
+      : "a"(func), "c"(func2));
+#else
+#define cpuid(func, func2, ax, bx, cx, dx)     \
+  asm volatile(                                \
+      "pushl %ebx       \n\t"                  \
+      "cpuid            \n\t"                  \
+      "movl %ebx, %edi  \n\t"                  \
+      "popl %ebx        \n\t"                  \
+      : "=a"(ax), "=D"(bx), "=c"(cx), "=d"(dx) \
+      : "a"(func), "c"(func2));
+#endif
+#else /* end __SUNPRO__ */
+#if ARCH_X86_64
+#if defined(_MSC_VER) && _MSC_VER > 1500
+#define cpuid(func, func2, a, b, c, d) \
+  do {                                 \
+    int regs[4];                       \
+    __cpuidex(regs, func, func2);      \
+    a = regs[0];                       \
+    b = regs[1];                       \
+    c = regs[2];                       \
+    d = regs[3];                       \
+  } while (0)
+#else
+#define cpuid(func, func2, a, b, c, d) \
+  do {                                 \
+    int regs[4];                       \
+    __cpuid(regs, func);               \
+    a = regs[0];                       \
+    b = regs[1];                       \
+    c = regs[2];                       \
+    d = regs[3];                       \
+  } while (0)
+#endif
+#else
+/* clang-format off */
+#define cpuid(func, func2, a, b, c, d) \
+  __asm mov eax, func                  \
+  __asm mov ecx, func2                 \
+  __asm cpuid                          \
+  __asm mov a, eax                     \
+  __asm mov b, ebx                     \
+  __asm mov c, ecx                     \
+  __asm mov d, edx
+#endif
+/* clang-format on */
+#endif /* end others */
+
+// NaCl has no support for xgetbv or the raw opcode.
+#if !defined(__native_client__) && (defined(__i386__) || defined(__x86_64__))
+static INLINE uint64_t xgetbv(void) {
+  const uint32_t ecx = 0;
+  uint32_t eax, edx;
+  // Use the raw opcode for xgetbv for compatibility with older toolchains.
+  __asm__ volatile(".byte 0x0f, 0x01, 0xd0\n"
+                   : "=a"(eax), "=d"(edx)
+                   : "c"(ecx));
+  return ((uint64_t)edx << 32) | eax;
+}
+#elif (defined(_M_X64) || defined(_M_IX86)) && defined(_MSC_FULL_VER) && \
+    _MSC_FULL_VER >= 160040219  // >= VS2010 SP1
+#include <immintrin.h>
+#define xgetbv() _xgetbv(0)
+#elif defined(_MSC_VER) && defined(_M_IX86)
+static INLINE uint64_t xgetbv(void) {
+  uint32_t eax_, edx_;
+  __asm {
+    xor ecx, ecx  // ecx = 0
+    // Use the raw opcode for xgetbv for compatibility with older toolchains.
+    __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0
+    mov eax_, eax
+    mov edx_, edx
+  }
+  return ((uint64_t)edx_ << 32) | eax_;
+}
+#else
+#define xgetbv() 0U  // no AVX for older x64 or unrecognized toolchains.
+#endif
+
+#if defined(_MSC_VER) && _MSC_VER >= 1700
+#include <windows.h>
+#if WINAPI_FAMILY_PARTITION(WINAPI_FAMILY_APP)
+#define getenv(x) NULL
+#endif
+#endif
+
+#define HAS_MMX 0x01
+#define HAS_SSE 0x02
+#define HAS_SSE2 0x04
+#define HAS_SSE3 0x08
+#define HAS_SSSE3 0x10
+#define HAS_SSE4_1 0x20
+#define HAS_AVX 0x40
+#define HAS_AVX2 0x80
+#ifndef BIT
+#define BIT(n) (1 << n)
+#endif
+
+static INLINE int x86_simd_caps(void) {
+  unsigned int flags = 0;
+  unsigned int mask = ~0;
+  unsigned int max_cpuid_val, reg_eax, reg_ebx, reg_ecx, reg_edx;
+  char *env;
+  (void)reg_ebx;
+
+  /* See if the CPU capabilities are being overridden by the environment */
+  env = getenv("AOM_SIMD_CAPS");
+
+  if (env && *env) return (int)strtol(env, NULL, 0);
+
+  env = getenv("AOM_SIMD_CAPS_MASK");
+
+  if (env && *env) mask = (unsigned int)strtoul(env, NULL, 0);
+
+  /* Ensure that the CPUID instruction supports extended features */
+  cpuid(0, 0, max_cpuid_val, reg_ebx, reg_ecx, reg_edx);
+
+  if (max_cpuid_val < 1) return 0;
+
+  /* Get the standard feature flags */
+  cpuid(1, 0, reg_eax, reg_ebx, reg_ecx, reg_edx);
+
+  if (reg_edx & BIT(23)) flags |= HAS_MMX;
+
+  if (reg_edx & BIT(25)) flags |= HAS_SSE; /* aka xmm */
+
+  if (reg_edx & BIT(26)) flags |= HAS_SSE2; /* aka wmt */
+
+  if (reg_ecx & BIT(0)) flags |= HAS_SSE3;
+
+  if (reg_ecx & BIT(9)) flags |= HAS_SSSE3;
+
+  if (reg_ecx & BIT(19)) flags |= HAS_SSE4_1;
+
+  // bits 27 (OSXSAVE) & 28 (256-bit AVX)
+  if ((reg_ecx & (BIT(27) | BIT(28))) == (BIT(27) | BIT(28))) {
+    if ((xgetbv() & 0x6) == 0x6) {
+      flags |= HAS_AVX;
+
+      if (max_cpuid_val >= 7) {
+        /* Get the leaf 7 feature flags. Needed to check for AVX2 support */
+        cpuid(7, 0, reg_eax, reg_ebx, reg_ecx, reg_edx);
+
+        if (reg_ebx & BIT(5)) flags |= HAS_AVX2;
+      }
+    }
+  }
+
+  return flags & mask;
+}
+
+// Note:
+//  32-bit CPU cycle counter is light-weighted for most function performance
+//  measurement. For large function (CPU time > a couple of seconds), 64-bit
+//  counter should be used.
+// 32-bit CPU cycle counter
+static INLINE unsigned int x86_readtsc(void) {
+#if defined(__GNUC__) && __GNUC__
+  unsigned int tsc;
+  __asm__ __volatile__("rdtsc\n\t" : "=a"(tsc) :);
+  return tsc;
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+  unsigned int tsc;
+  asm volatile("rdtsc\n\t" : "=a"(tsc) :);
+  return tsc;
+#else
+#if ARCH_X86_64
+  return (unsigned int)__rdtsc();
+#else
+  __asm rdtsc;
+#endif
+#endif
+}
+// 64-bit CPU cycle counter
+static INLINE uint64_t x86_readtsc64(void) {
+#if defined(__GNUC__) && __GNUC__
+  uint32_t hi, lo;
+  __asm__ __volatile__("rdtsc" : "=a"(lo), "=d"(hi));
+  return ((uint64_t)hi << 32) | lo;
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+  uint_t hi, lo;
+  asm volatile("rdtsc\n\t" : "=a"(lo), "=d"(hi));
+  return ((uint64_t)hi << 32) | lo;
+#else
+#if ARCH_X86_64
+  return (uint64_t)__rdtsc();
+#else
+  __asm rdtsc;
+#endif
+#endif
+}
+
+#if defined(__GNUC__) && __GNUC__
+#define x86_pause_hint() __asm__ __volatile__("pause \n\t")
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+#define x86_pause_hint() asm volatile("pause \n\t")
+#else
+#if ARCH_X86_64
+#define x86_pause_hint() _mm_pause();
+#else
+#define x86_pause_hint() __asm pause
+#endif
+#endif
+
+#if defined(__GNUC__) && __GNUC__
+static void x87_set_control_word(unsigned short mode) {
+  __asm__ __volatile__("fldcw %0" : : "m"(*&mode));
+}
+static unsigned short x87_get_control_word(void) {
+  unsigned short mode;
+  __asm__ __volatile__("fstcw %0\n\t" : "=m"(*&mode) :);
+  return mode;
+}
+#elif defined(__SUNPRO_C) || defined(__SUNPRO_CC)
+static void x87_set_control_word(unsigned short mode) {
+  asm volatile("fldcw %0" : : "m"(*&mode));
+}
+static unsigned short x87_get_control_word(void) {
+  unsigned short mode;
+  asm volatile("fstcw %0\n\t" : "=m"(*&mode) :);
+  return mode;
+}
+#elif ARCH_X86_64
+/* No fldcw intrinsics on Windows x64, punt to external asm */
+extern void aom_winx64_fldcw(unsigned short mode);
+extern unsigned short aom_winx64_fstcw(void);
+#define x87_set_control_word aom_winx64_fldcw
+#define x87_get_control_word aom_winx64_fstcw
+#else
+static void x87_set_control_word(unsigned short mode) {
+  __asm { fldcw mode }
+}
+static unsigned short x87_get_control_word(void) {
+  unsigned short mode;
+  __asm { fstcw mode }
+  return mode;
+}
+#endif
+
+static INLINE unsigned int x87_set_double_precision(void) {
+  unsigned int mode = x87_get_control_word();
+  x87_set_control_word((mode & ~0x300) | 0x200);
+  return mode;
+}
+
+extern void aom_reset_mmx_state(void);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_PORTS_X86_H_
diff --git a/third_party/aom/aom_ports/x86_abi_support.asm b/third_party/aom/aom_ports/x86_abi_support.asm
new file mode 100644
index 0000000000..6aeee60a06
--- /dev/null
+++ b/third_party/aom/aom_ports/x86_abi_support.asm
@@ -0,0 +1,395 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+%include "aom_config.asm"
+
+; 32/64 bit compatibility macros
+;
+; In general, we make the source use 64 bit syntax, then twiddle with it using
+; the preprocessor to get the 32 bit syntax on 32 bit platforms.
+;
+%ifidn __OUTPUT_FORMAT__,elf32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,macho32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,win32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,aout
+%define ABI_IS_32BIT 1
+%else
+%define ABI_IS_32BIT 0
+%endif
+
+%if ABI_IS_32BIT
+%define rax eax
+%define rbx ebx
+%define rcx ecx
+%define rdx edx
+%define rsi esi
+%define rdi edi
+%define rsp esp
+%define rbp ebp
+%define movsxd mov
+%macro movq 2
+  %ifidn %1,eax
+    movd %1,%2
+  %elifidn %2,eax
+    movd %1,%2
+  %elifidn %1,ebx
+    movd %1,%2
+  %elifidn %2,ebx
+    movd %1,%2
+  %elifidn %1,ecx
+    movd %1,%2
+  %elifidn %2,ecx
+    movd %1,%2
+  %elifidn %1,edx
+    movd %1,%2
+  %elifidn %2,edx
+    movd %1,%2
+  %elifidn %1,esi
+    movd %1,%2
+  %elifidn %2,esi
+    movd %1,%2
+  %elifidn %1,edi
+    movd %1,%2
+  %elifidn %2,edi
+    movd %1,%2
+  %elifidn %1,esp
+    movd %1,%2
+  %elifidn %2,esp
+    movd %1,%2
+  %elifidn %1,ebp
+    movd %1,%2
+  %elifidn %2,ebp
+    movd %1,%2
+  %else
+    movq %1,%2
+  %endif
+%endmacro
+%endif
+
+
+; LIBAOM_YASM_WIN64
+; Set LIBAOM_YASM_WIN64 if output is Windows 64bit so the code will work if x64
+; or win64 is defined on the Yasm command line.
+%ifidn __OUTPUT_FORMAT__,win64
+%define LIBAOM_YASM_WIN64 1
+%elifidn __OUTPUT_FORMAT__,x64
+%define LIBAOM_YASM_WIN64 1
+%else
+%define LIBAOM_YASM_WIN64 0
+%endif
+
+; sym()
+; Return the proper symbol name for the target ABI.
+;
+; Certain ABIs, notably MS COFF and Darwin MACH-O, require that symbols
+; with C linkage be prefixed with an underscore.
+;
+%ifidn   __OUTPUT_FORMAT__,elf32
+%define sym(x) x
+%elifidn __OUTPUT_FORMAT__,elf64
+%define sym(x) x
+%elifidn __OUTPUT_FORMAT__,elfx32
+%define sym(x) x
+%elif LIBAOM_YASM_WIN64
+%define sym(x) x
+%else
+%define sym(x) _ %+ x
+%endif
+
+;  PRIVATE
+;  Macro for the attribute to hide a global symbol for the target ABI.
+;  This is only active if CHROMIUM is defined.
+;
+;  Chromium doesn't like exported global symbols due to symbol clashing with
+;  plugins among other things.
+;
+;  Requires Chromium's patched copy of yasm:
+;    http://src.chromium.org/viewvc/chrome?view=rev&revision=73761
+;    http://www.tortall.net/projects/yasm/ticket/236
+;
+%ifdef CHROMIUM
+  %ifidn   __OUTPUT_FORMAT__,elf32
+    %define PRIVATE :hidden
+  %elifidn __OUTPUT_FORMAT__,elf64
+    %define PRIVATE :hidden
+  %elifidn __OUTPUT_FORMAT__,elfx32
+    %define PRIVATE :hidden
+  %elif LIBAOM_YASM_WIN64
+    %define PRIVATE
+  %else
+    %define PRIVATE :private_extern
+  %endif
+%else
+  %define PRIVATE
+%endif
+
+; arg()
+; Return the address specification of the given argument
+;
+%if ABI_IS_32BIT
+  %define arg(x) [ebp+8+4*x]
+%else
+  ; 64 bit ABI passes arguments in registers. This is a workaround to get up
+  ; and running quickly. Relies on SHADOW_ARGS_TO_STACK
+  %if LIBAOM_YASM_WIN64
+    %define arg(x) [rbp+16+8*x]
+  %else
+    %define arg(x) [rbp-8-8*x]
+  %endif
+%endif
+
+; REG_SZ_BYTES, REG_SZ_BITS
+; Size of a register
+%if ABI_IS_32BIT
+%define REG_SZ_BYTES 4
+%define REG_SZ_BITS  32
+%else
+%define REG_SZ_BYTES 8
+%define REG_SZ_BITS  64
+%endif
+
+
+; ALIGN_STACK <alignment> <register>
+; This macro aligns the stack to the given alignment (in bytes). The stack
+; is left such that the previous value of the stack pointer is the first
+; argument on the stack (ie, the inverse of this macro is 'pop rsp.')
+; This macro uses one temporary register, which is not preserved, and thus
+; must be specified as an argument.
+%macro ALIGN_STACK 2
+    mov         %2, rsp
+    and         rsp, -%1
+    lea         rsp, [rsp - (%1 - REG_SZ_BYTES)]
+    push        %2
+%endmacro
+
+
+;
+; The Microsoft assembler tries to impose a certain amount of type safety in
+; its register usage. YASM doesn't recognize these directives, so we just
+; %define them away to maintain as much compatibility as possible with the
+; original inline assembler we're porting from.
+;
+%idefine PTR
+%idefine XMMWORD
+%idefine MMWORD
+
+; PIC macros
+;
+%if ABI_IS_32BIT
+  %if CONFIG_PIC=1
+  %ifidn __OUTPUT_FORMAT__,elf32
+    %define WRT_PLT wrt ..plt
+    %macro GET_GOT 1
+      extern _GLOBAL_OFFSET_TABLE_
+      push %1
+      call %%get_got
+      %%sub_offset:
+      jmp %%exitGG
+      %%get_got:
+      mov %1, [esp]
+      add %1, _GLOBAL_OFFSET_TABLE_ + $$ - %%sub_offset wrt ..gotpc
+      ret
+      %%exitGG:
+      %undef GLOBAL
+      %define GLOBAL(x) x + %1 wrt ..gotoff
+      %undef RESTORE_GOT
+      %define RESTORE_GOT pop %1
+    %endmacro
+  %elifidn __OUTPUT_FORMAT__,macho32
+    %macro GET_GOT 1
+      push %1
+      call %%get_got
+      %%get_got:
+      pop  %1
+      %undef GLOBAL
+      %define GLOBAL(x) x + %1 - %%get_got
+      %undef RESTORE_GOT
+      %define RESTORE_GOT pop %1
+    %endmacro
+  %endif
+  %endif
+
+  %ifdef CHROMIUM
+    %ifidn __OUTPUT_FORMAT__,macho32
+      %define HIDDEN_DATA(x) x:private_extern
+    %else
+      %define HIDDEN_DATA(x) x
+    %endif
+  %else
+    %define HIDDEN_DATA(x) x
+  %endif
+%else
+  %macro GET_GOT 1
+  %endmacro
+  %define GLOBAL(x) rel x
+  %ifidn __OUTPUT_FORMAT__,elf64
+    %define WRT_PLT wrt ..plt
+    %define HIDDEN_DATA(x) x:data hidden
+  %elifidn __OUTPUT_FORMAT__,elfx32
+    %define WRT_PLT wrt ..plt
+    %define HIDDEN_DATA(x) x:data hidden
+  %elifidn __OUTPUT_FORMAT__,macho64
+    %ifdef CHROMIUM
+      %define HIDDEN_DATA(x) x:private_extern
+    %else
+      %define HIDDEN_DATA(x) x
+    %endif
+  %else
+    %define HIDDEN_DATA(x) x
+  %endif
+%endif
+%ifnmacro GET_GOT
+    %macro GET_GOT 1
+    %endmacro
+    %define GLOBAL(x) x
+%endif
+%ifndef RESTORE_GOT
+%define RESTORE_GOT
+%endif
+%ifndef WRT_PLT
+%define WRT_PLT
+%endif
+
+%if ABI_IS_32BIT
+  %macro SHADOW_ARGS_TO_STACK 1
+  %endm
+  %define UNSHADOW_ARGS
+%else
+%if LIBAOM_YASM_WIN64
+  %macro SHADOW_ARGS_TO_STACK 1 ; argc
+    %if %1 > 0
+        mov arg(0),rcx
+    %endif
+    %if %1 > 1
+        mov arg(1),rdx
+    %endif
+    %if %1 > 2
+        mov arg(2),r8
+    %endif
+    %if %1 > 3
+        mov arg(3),r9
+    %endif
+  %endm
+%else
+  %macro SHADOW_ARGS_TO_STACK 1 ; argc
+    %if %1 > 0
+        push rdi
+    %endif
+    %if %1 > 1
+        push rsi
+    %endif
+    %if %1 > 2
+        push rdx
+    %endif
+    %if %1 > 3
+        push rcx
+    %endif
+    %if %1 > 4
+        push r8
+    %endif
+    %if %1 > 5
+        push r9
+    %endif
+    %if %1 > 6
+      %assign i %1-6
+      %assign off 16
+      %rep i
+        mov rax,[rbp+off]
+        push rax
+        %assign off off+8
+      %endrep
+    %endif
+  %endm
+%endif
+  %define UNSHADOW_ARGS mov rsp, rbp
+%endif
+
+; Win64 ABI requires that XMM6:XMM15 are callee saved
+; SAVE_XMM n, [u]
+; store registers 6-n on the stack
+; if u is specified, use unaligned movs.
+; Win64 ABI requires 16 byte stack alignment, but then pushes an 8 byte return
+; value. Typically we follow this up with 'push rbp' - re-aligning the stack -
+; but in some cases this is not done and unaligned movs must be used.
+%if LIBAOM_YASM_WIN64
+%macro SAVE_XMM 1-2 a
+  %if %1 < 6
+    %error Only xmm registers 6-15 must be preserved
+  %else
+    %assign last_xmm %1
+    %define movxmm movdq %+ %2
+    %assign xmm_stack_space ((last_xmm - 5) * 16)
+    sub rsp, xmm_stack_space
+    %assign i 6
+    %rep (last_xmm - 5)
+      movxmm [rsp + ((i - 6) * 16)], xmm %+ i
+      %assign i i+1
+    %endrep
+  %endif
+%endmacro
+%macro RESTORE_XMM 0
+  %ifndef last_xmm
+    %error RESTORE_XMM must be paired with SAVE_XMM n
+  %else
+    %assign i last_xmm
+    %rep (last_xmm - 5)
+      movxmm xmm %+ i, [rsp +((i - 6) * 16)]
+      %assign i i-1
+    %endrep
+    add rsp, xmm_stack_space
+    ; there are a couple functions which return from multiple places.
+    ; otherwise, we could uncomment these:
+    ; %undef last_xmm
+    ; %undef xmm_stack_space
+    ; %undef movxmm
+  %endif
+%endmacro
+%else
+%macro SAVE_XMM 1-2
+%endmacro
+%macro RESTORE_XMM 0
+%endmacro
+%endif
+
+; Name of the rodata section
+;
+; .rodata seems to be an elf-ism, as it doesn't work on OSX.
+;
+%ifidn __OUTPUT_FORMAT__,macho64
+%define SECTION_RODATA section .text
+%elifidn __OUTPUT_FORMAT__,macho32
+%macro SECTION_RODATA 0
+section .text
+%endmacro
+%elifidn __OUTPUT_FORMAT__,aout
+%define SECTION_RODATA section .data
+%else
+%define SECTION_RODATA section .rodata
+%endif
+
+
+; Tell GNU ld that we don't require an executable stack.
+%ifidn __OUTPUT_FORMAT__,elf32
+section .note.GNU-stack noalloc noexec nowrite progbits
+section .text
+%elifidn __OUTPUT_FORMAT__,elf64
+section .note.GNU-stack noalloc noexec nowrite progbits
+section .text
+%elifidn __OUTPUT_FORMAT__,elfx32
+section .note.GNU-stack noalloc noexec nowrite progbits
+section .text
+%endif
diff --git a/third_party/aom/aom_scale/aom_scale.cmake b/third_party/aom/aom_scale/aom_scale.cmake
new file mode 100644
index 0000000000..a6aa31afce
--- /dev/null
+++ b/third_party/aom/aom_scale/aom_scale.cmake
@@ -0,0 +1,34 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+set(AOM_SCALE_SOURCES
+    "${AOM_ROOT}/aom_scale/aom_scale.h"
+    "${AOM_ROOT}/aom_scale/generic/aom_scale.c"
+    "${AOM_ROOT}/aom_scale/generic/gen_scalers.c"
+    "${AOM_ROOT}/aom_scale/generic/yv12config.c"
+    "${AOM_ROOT}/aom_scale/generic/yv12extend.c"
+    "${AOM_ROOT}/aom_scale/yv12config.h")
+
+set(AOM_SCALE_INTRIN_DSPR2
+    "${AOM_ROOT}/aom_scale/mips/dspr2/yv12extend_dspr2.c")
+
+# Creates the aom_scale build target and makes libaom depend on it. The libaom
+# target must exist before this function is called.
+function (setup_aom_scale_targets)
+  add_library(aom_scale OBJECT ${AOM_SCALE_SOURCES})
+  target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_scale>)
+
+  if (HAVE_DSPR2)
+    add_intrinsics_object_library("" "dspr2" "aom_scale"
+                                  "AOM_SCALE_INTRIN_DSPR2")
+  endif ()
+
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_scale PARENT_SCOPE)
+endfunction ()
diff --git a/third_party/aom/aom_scale/aom_scale.h b/third_party/aom/aom_scale/aom_scale.h
new file mode 100644
index 0000000000..6e089f5aaa
--- /dev/null
+++ b/third_party/aom/aom_scale/aom_scale.h
@@ -0,0 +1,23 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_SCALE_AOM_SCALE_H_
+#define AOM_SCALE_AOM_SCALE_H_
+
+#include "aom_scale/yv12config.h"
+
+extern void aom_scale_frame(YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst,
+                            unsigned char *temp_area, unsigned char temp_height,
+                            unsigned int hscale, unsigned int hratio,
+                            unsigned int vscale, unsigned int vratio,
+                            unsigned int interlaced);
+
+#endif  // AOM_SCALE_AOM_SCALE_H_
diff --git a/third_party/aom/aom_scale/aom_scale.mk b/third_party/aom/aom_scale/aom_scale.mk
new file mode 100644
index 0000000000..e3a68cfcf0
--- /dev/null
+++ b/third_party/aom/aom_scale/aom_scale.mk
@@ -0,0 +1,16 @@
+SCALE_SRCS-yes += aom_scale.mk
+SCALE_SRCS-yes += yv12config.h
+SCALE_SRCS-$(CONFIG_SPATIAL_RESAMPLING) += aom_scale.h
+SCALE_SRCS-$(CONFIG_SPATIAL_RESAMPLING) += generic/aom_scale.c
+SCALE_SRCS-yes += generic/yv12config.c
+SCALE_SRCS-yes += generic/yv12extend.c
+SCALE_SRCS-$(CONFIG_SPATIAL_RESAMPLING) += generic/gen_scalers.c
+SCALE_SRCS-yes += aom_scale_rtcd.c
+SCALE_SRCS-yes += aom_scale_rtcd.pl
+
+#mips(dspr2)
+SCALE_SRCS-$(HAVE_DSPR2)  += mips/dspr2/yv12extend_dspr2.c
+
+SCALE_SRCS-no += $(SCALE_SRCS_REMOVE-yes)
+
+$(eval $(call rtcd_h_template,aom_scale_rtcd,aom_scale/aom_scale_rtcd.pl))
diff --git a/third_party/aom/aom_scale/aom_scale_rtcd.c b/third_party/aom/aom_scale/aom_scale_rtcd.c
new file mode 100644
index 0000000000..dec23735bf
--- /dev/null
+++ b/third_party/aom/aom_scale/aom_scale_rtcd.c
@@ -0,0 +1,16 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "./aom_config.h"
+#define RTCD_C
+#include "./aom_scale_rtcd.h"
+#include "aom_ports/aom_once.h"
+
+void aom_scale_rtcd() { once(setup_rtcd_internal); }
diff --git a/third_party/aom/aom_scale/aom_scale_rtcd.pl b/third_party/aom/aom_scale/aom_scale_rtcd.pl
new file mode 100644
index 0000000000..9d9a1a29ae
--- /dev/null
+++ b/third_party/aom/aom_scale/aom_scale_rtcd.pl
@@ -0,0 +1,38 @@
+sub aom_scale_forward_decls() {
+print <<EOF
+struct yv12_buffer_config;
+EOF
+}
+forward_decls qw/aom_scale_forward_decls/;
+
+# Scaler functions
+if (aom_config("CONFIG_SPATIAL_RESAMPLING") eq "yes") {
+    add_proto qw/void aom_horizontal_line_5_4_scale/, "const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width";
+    add_proto qw/void aom_vertical_band_5_4_scale/, "unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width";
+    add_proto qw/void aom_horizontal_line_5_3_scale/, "const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width";
+    add_proto qw/void aom_vertical_band_5_3_scale/, "unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width";
+    add_proto qw/void aom_horizontal_line_2_1_scale/, "const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width";
+    add_proto qw/void aom_vertical_band_2_1_scale/, "unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width";
+    add_proto qw/void aom_vertical_band_2_1_scale_i/, "unsigned char *source, int src_pitch, unsigned char *dest, int dest_pitch, unsigned int dest_width";
+}
+
+add_proto qw/void aom_yv12_extend_frame_borders/, "struct yv12_buffer_config *ybf";
+
+add_proto qw/void aom_yv12_copy_frame/, "const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc";
+
+add_proto qw/void aom_yv12_copy_y/, "const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc";
+
+add_proto qw/void aom_yv12_copy_u/, "const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc";
+
+add_proto qw/void aom_yv12_copy_v/, "const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc";
+
+if (aom_config("CONFIG_AV1") eq "yes") {
+    add_proto qw/void aom_extend_frame_borders/, "struct yv12_buffer_config *ybf";
+    specialize qw/aom_extend_frame_borders dspr2/;
+
+    add_proto qw/void aom_extend_frame_inner_borders/, "struct yv12_buffer_config *ybf";
+    specialize qw/aom_extend_frame_inner_borders dspr2/;
+
+    add_proto qw/void aom_extend_frame_borders_y/, "struct yv12_buffer_config *ybf";
+}
+1;
diff --git a/third_party/aom/aom_scale/generic/aom_scale.c b/third_party/aom/aom_scale/generic/aom_scale.c
new file mode 100644
index 0000000000..14f3ae0da6
--- /dev/null
+++ b/third_party/aom/aom_scale/generic/aom_scale.c
@@ -0,0 +1,529 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/****************************************************************************
+ *
+ *   Module Title :     scale.c
+ *
+ *   Description  :     Image scaling functions.
+ *
+ ***************************************************************************/
+
+/****************************************************************************
+*  Header Files
+****************************************************************************/
+#include "./aom_scale_rtcd.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_scale/aom_scale.h"
+#include "aom_scale/yv12config.h"
+
+typedef struct {
+  int expanded_frame_width;
+  int expanded_frame_height;
+
+  int HScale;
+  int HRatio;
+  int VScale;
+  int VRatio;
+
+  YV12_BUFFER_CONFIG *src_yuv_config;
+  YV12_BUFFER_CONFIG *dst_yuv_config;
+
+} SCALE_VARS;
+
+/****************************************************************************
+ *
+ *  ROUTINE       : scale1d_2t1_i
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to data to be scaled.
+ *                  int source_step             : Number of pixels to step on
+ *                                                in source.
+ *                  unsigned int source_scale   : Scale for source (UNUSED).
+ *                  unsigned int source_length  : Length of source (UNUSED).
+ *                  unsigned char *dest         : Pointer to output data array.
+ *                  int dest_step               : Number of pixels to step on
+ *                                                in destination.
+ *                  unsigned int dest_scale     : Scale for destination
+ *                                                (UNUSED).
+ *                  unsigned int dest_length    : Length of destination.
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Performs 2-to-1 interpolated scaling.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ ****************************************************************************/
+static void scale1d_2t1_i(const unsigned char *source, int source_step,
+                          unsigned int source_scale, unsigned int source_length,
+                          unsigned char *dest, int dest_step,
+                          unsigned int dest_scale, unsigned int dest_length) {
+  const unsigned char *const dest_end = dest + dest_length * dest_step;
+  (void)source_length;
+  (void)source_scale;
+  (void)dest_scale;
+
+  source_step *= 2;  // Every other row.
+
+  dest[0] = source[0];  // Special case: 1st pixel.
+  source += source_step;
+  dest += dest_step;
+
+  while (dest < dest_end) {
+    const unsigned int a = 3 * source[-source_step];
+    const unsigned int b = 10 * source[0];
+    const unsigned int c = 3 * source[source_step];
+    *dest = (unsigned char)((8 + a + b + c) >> 4);
+    source += source_step;
+    dest += dest_step;
+  }
+}
+
+/****************************************************************************
+ *
+ *  ROUTINE       : scale1d_2t1_ps
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to data to be scaled.
+ *                  int source_step             : Number of pixels to step on
+ *                                                in source.
+ *                  unsigned int source_scale   : Scale for source (UNUSED).
+ *                  unsigned int source_length  : Length of source (UNUSED).
+ *                  unsigned char *dest         : Pointer to output data array.
+ *                  int dest_step               : Number of pixels to step on
+ *                                                in destination.
+ *                  unsigned int dest_scale     : Scale for destination
+ *                                                (UNUSED).
+ *                  unsigned int dest_length    : Length of destination.
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Performs 2-to-1 point subsampled scaling.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ ****************************************************************************/
+static void scale1d_2t1_ps(const unsigned char *source, int source_step,
+                           unsigned int source_scale,
+                           unsigned int source_length, unsigned char *dest,
+                           int dest_step, unsigned int dest_scale,
+                           unsigned int dest_length) {
+  const unsigned char *const dest_end = dest + dest_length * dest_step;
+  (void)source_length;
+  (void)source_scale;
+  (void)dest_scale;
+
+  source_step *= 2;  // Every other row.
+
+  while (dest < dest_end) {
+    *dest = *source;
+    source += source_step;
+    dest += dest_step;
+  }
+}
+/****************************************************************************
+ *
+ *  ROUTINE       : scale1d_c
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to data to be scaled.
+ *                  int source_step             : Number of pixels to step on
+ *                                                in source.
+ *                  unsigned int source_scale   : Scale for source.
+ *                  unsigned int source_length  : Length of source (UNUSED).
+ *                  unsigned char *dest         : Pointer to output data array.
+ *                  int dest_step               : Number of pixels to step on
+ *                                                in destination.
+ *                  unsigned int dest_scale     : Scale for destination.
+ *                  unsigned int dest_length    : Length of destination.
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Performs linear interpolation in one dimension.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ ****************************************************************************/
+static void scale1d_c(const unsigned char *source, int source_step,
+                      unsigned int source_scale, unsigned int source_length,
+                      unsigned char *dest, int dest_step,
+                      unsigned int dest_scale, unsigned int dest_length) {
+  const unsigned char *const dest_end = dest + dest_length * dest_step;
+  const unsigned int round_value = dest_scale / 2;
+  unsigned int left_modifier = dest_scale;
+  unsigned int right_modifier = 0;
+  unsigned char left_pixel = source[0];
+  unsigned char right_pixel = source[source_step];
+
+  (void)source_length;
+
+  /* These asserts are needed if there are boundary issues... */
+  /* assert ( dest_scale > source_scale );*/
+  /* assert ( (source_length - 1) * dest_scale >= (dest_length - 1) *
+   * source_scale);*/
+
+  while (dest < dest_end) {
+    *dest = (unsigned char)((left_modifier * left_pixel +
+                             right_modifier * right_pixel + round_value) /
+                            dest_scale);
+
+    right_modifier += source_scale;
+
+    while (right_modifier > dest_scale) {
+      right_modifier -= dest_scale;
+      source += source_step;
+      left_pixel = source[0];
+      right_pixel = source[source_step];
+    }
+
+    left_modifier = dest_scale - right_modifier;
+  }
+}
+
+/****************************************************************************
+ *
+ *  ROUTINE       : Scale2D
+ *
+ *  INPUTS        : const unsigned char *source    : Pointer to data to be
+ *                                                   scaled.
+ *                  int source_pitch               : Stride of source image.
+ *                  unsigned int source_width      : Width of input image.
+ *                  unsigned int source_height     : Height of input image.
+ *                  unsigned char *dest            : Pointer to output data
+ *                                                   array.
+ *                  int dest_pitch                 : Stride of destination
+ *                                                   image.
+ *                  unsigned int dest_width        : Width of destination image.
+ *                  unsigned int dest_height       : Height of destination
+ *                                                   image.
+ *                  unsigned char *temp_area       : Pointer to temp work area.
+ *                  unsigned char temp_area_height : Height of temp work area.
+ *                  unsigned int hscale            : Horizontal scale factor
+ *                                                   numerator.
+ *                  unsigned int hratio            : Horizontal scale factor
+ *                                                   denominator.
+ *                  unsigned int vscale            : Vertical scale factor
+ *                                                   numerator.
+ *                  unsigned int vratio            : Vertical scale factor
+ *                                                   denominator.
+ *                  unsigned int interlaced        : Interlace flag.
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Performs 2-tap linear interpolation in two dimensions.
+ *
+ *  SPECIAL NOTES : Expansion is performed one band at a time to help with
+ *                  caching.
+ *
+ ****************************************************************************/
+static void Scale2D(
+    /*const*/
+    unsigned char *source, int source_pitch, unsigned int source_width,
+    unsigned int source_height, unsigned char *dest, int dest_pitch,
+    unsigned int dest_width, unsigned int dest_height, unsigned char *temp_area,
+    unsigned char temp_area_height, unsigned int hscale, unsigned int hratio,
+    unsigned int vscale, unsigned int vratio, unsigned int interlaced) {
+  unsigned int i, j, k;
+  unsigned int bands;
+  unsigned int dest_band_height;
+  unsigned int source_band_height;
+
+  typedef void (*Scale1D)(const unsigned char *source, int source_step,
+                          unsigned int source_scale, unsigned int source_length,
+                          unsigned char *dest, int dest_step,
+                          unsigned int dest_scale, unsigned int dest_length);
+
+  Scale1D Scale1Dv = scale1d_c;
+  Scale1D Scale1Dh = scale1d_c;
+
+  void (*horiz_line_scale)(const unsigned char *, unsigned int, unsigned char *,
+                           unsigned int) = NULL;
+  void (*vert_band_scale)(unsigned char *, int, unsigned char *, int,
+                          unsigned int) = NULL;
+
+  int ratio_scalable = 1;
+  int interpolation = 0;
+
+  unsigned char *source_base;
+  unsigned char *line_src;
+
+  source_base = (unsigned char *)source;
+
+  if (source_pitch < 0) {
+    int offset;
+
+    offset = (source_height - 1);
+    offset *= source_pitch;
+
+    source_base += offset;
+  }
+
+  /* find out the ratio for each direction */
+  switch (hratio * 10 / hscale) {
+    case 8:
+      /* 4-5 Scale in Width direction */
+      horiz_line_scale = aom_horizontal_line_5_4_scale;
+      break;
+    case 6:
+      /* 3-5 Scale in Width direction */
+      horiz_line_scale = aom_horizontal_line_5_3_scale;
+      break;
+    case 5:
+      /* 1-2 Scale in Width direction */
+      horiz_line_scale = aom_horizontal_line_2_1_scale;
+      break;
+    default:
+      /* The ratio is not acceptable now */
+      /* throw("The ratio is not acceptable for now!"); */
+      ratio_scalable = 0;
+      break;
+  }
+
+  switch (vratio * 10 / vscale) {
+    case 8:
+      /* 4-5 Scale in vertical direction */
+      vert_band_scale = aom_vertical_band_5_4_scale;
+      source_band_height = 5;
+      dest_band_height = 4;
+      break;
+    case 6:
+      /* 3-5 Scale in vertical direction */
+      vert_band_scale = aom_vertical_band_5_3_scale;
+      source_band_height = 5;
+      dest_band_height = 3;
+      break;
+    case 5:
+      /* 1-2 Scale in vertical direction */
+
+      if (interlaced) {
+        /* if the content is interlaced, point sampling is used */
+        vert_band_scale = aom_vertical_band_2_1_scale;
+      } else {
+        interpolation = 1;
+        /* if the content is progressive, interplo */
+        vert_band_scale = aom_vertical_band_2_1_scale_i;
+      }
+
+      source_band_height = 2;
+      dest_band_height = 1;
+      break;
+    default:
+      /* The ratio is not acceptable now */
+      /* throw("The ratio is not acceptable for now!"); */
+      ratio_scalable = 0;
+      break;
+  }
+
+  if (ratio_scalable) {
+    if (source_height == dest_height) {
+      /* for each band of the image */
+      for (k = 0; k < dest_height; ++k) {
+        horiz_line_scale(source, source_width, dest, dest_width);
+        source += source_pitch;
+        dest += dest_pitch;
+      }
+
+      return;
+    }
+
+    if (interpolation) {
+      if (source < source_base) source = source_base;
+
+      horiz_line_scale(source, source_width, temp_area, dest_width);
+    }
+
+    for (k = 0; k < (dest_height + dest_band_height - 1) / dest_band_height;
+         ++k) {
+      /* scale one band horizontally */
+      for (i = 0; i < source_band_height; ++i) {
+        /* Trap case where we could read off the base of the source buffer */
+
+        line_src = source + i * source_pitch;
+
+        if (line_src < source_base) line_src = source_base;
+
+        horiz_line_scale(line_src, source_width,
+                         temp_area + (i + 1) * dest_pitch, dest_width);
+      }
+
+      /* Vertical scaling is in place */
+      vert_band_scale(temp_area + dest_pitch, dest_pitch, dest, dest_pitch,
+                      dest_width);
+
+      if (interpolation)
+        memcpy(temp_area, temp_area + source_band_height * dest_pitch,
+               dest_width);
+
+      /* Next band... */
+      source += (unsigned long)source_band_height * source_pitch;
+      dest += (unsigned long)dest_band_height * dest_pitch;
+    }
+
+    return;
+  }
+
+  if (hscale == 2 && hratio == 1) Scale1Dh = scale1d_2t1_ps;
+
+  if (vscale == 2 && vratio == 1) {
+    if (interlaced)
+      Scale1Dv = scale1d_2t1_ps;
+    else
+      Scale1Dv = scale1d_2t1_i;
+  }
+
+  if (source_height == dest_height) {
+    /* for each band of the image */
+    for (k = 0; k < dest_height; ++k) {
+      Scale1Dh(source, 1, hscale, source_width + 1, dest, 1, hratio,
+               dest_width);
+      source += source_pitch;
+      dest += dest_pitch;
+    }
+
+    return;
+  }
+
+  if (dest_height > source_height) {
+    dest_band_height = temp_area_height - 1;
+    source_band_height = dest_band_height * source_height / dest_height;
+  } else {
+    source_band_height = temp_area_height - 1;
+    dest_band_height = source_band_height * vratio / vscale;
+  }
+
+  /* first row needs to be done so that we can stay one row ahead for vertical
+   * zoom */
+  Scale1Dh(source, 1, hscale, source_width + 1, temp_area, 1, hratio,
+           dest_width);
+
+  /* for each band of the image */
+  bands = (dest_height + dest_band_height - 1) / dest_band_height;
+
+  for (k = 0; k < bands; ++k) {
+    /* scale one band horizontally */
+    for (i = 1; i < source_band_height + 1; ++i) {
+      if (k * source_band_height + i < source_height) {
+        Scale1Dh(source + i * source_pitch, 1, hscale, source_width + 1,
+                 temp_area + i * dest_pitch, 1, hratio, dest_width);
+      } else { /*  Duplicate the last row */
+        /* copy temp_area row 0 over from last row in the past */
+        memcpy(temp_area + i * dest_pitch, temp_area + (i - 1) * dest_pitch,
+               dest_pitch);
+      }
+    }
+
+    /* scale one band vertically */
+    for (j = 0; j < dest_width; ++j) {
+      Scale1Dv(&temp_area[j], dest_pitch, vscale, source_band_height + 1,
+               &dest[j], dest_pitch, vratio, dest_band_height);
+    }
+
+    /* copy temp_area row 0 over from last row in the past */
+    memcpy(temp_area, temp_area + source_band_height * dest_pitch, dest_pitch);
+
+    /* move to the next band */
+    source += source_band_height * source_pitch;
+    dest += dest_band_height * dest_pitch;
+  }
+}
+
+/****************************************************************************
+ *
+ *  ROUTINE       : aom_scale_frame
+ *
+ *  INPUTS        : YV12_BUFFER_CONFIG *src        : Pointer to frame to be
+ *                                                   scaled.
+ *                  YV12_BUFFER_CONFIG *dst        : Pointer to buffer to hold
+ *                                                   scaled frame.
+ *                  unsigned char *temp_area       : Pointer to temp work area.
+ *                  unsigned char temp_area_height : Height of temp work area.
+ *                  unsigned int hscale            : Horizontal scale factor
+ *                                                   numerator.
+ *                  unsigned int hratio            : Horizontal scale factor
+ *                                                   denominator.
+ *                  unsigned int vscale            : Vertical scale factor
+ *                                                   numerator.
+ *                  unsigned int vratio            : Vertical scale factor
+ *                                                   denominator.
+ *                  unsigned int interlaced        : Interlace flag.
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Performs 2-tap linear interpolation in two dimensions.
+ *
+ *  SPECIAL NOTES : Expansion is performed one band at a time to help with
+ *                  caching.
+ *
+ ****************************************************************************/
+void aom_scale_frame(YV12_BUFFER_CONFIG *src, YV12_BUFFER_CONFIG *dst,
+                     unsigned char *temp_area, unsigned char temp_height,
+                     unsigned int hscale, unsigned int hratio,
+                     unsigned int vscale, unsigned int vratio,
+                     unsigned int interlaced) {
+  int i;
+  int dw = (hscale - 1 + src->y_width * hratio) / hscale;
+  int dh = (vscale - 1 + src->y_height * vratio) / vscale;
+
+  /* call our internal scaling routines!! */
+  Scale2D((unsigned char *)src->y_buffer, src->y_stride, src->y_width,
+          src->y_height, (unsigned char *)dst->y_buffer, dst->y_stride, dw, dh,
+          temp_area, temp_height, hscale, hratio, vscale, vratio, interlaced);
+
+  if (dw < (int)dst->y_width)
+    for (i = 0; i < dh; ++i)
+      memset(dst->y_buffer + i * dst->y_stride + dw - 1,
+             dst->y_buffer[i * dst->y_stride + dw - 2], dst->y_width - dw + 1);
+
+  if (dh < (int)dst->y_height)
+    for (i = dh - 1; i < (int)dst->y_height; ++i)
+      memcpy(dst->y_buffer + i * dst->y_stride,
+             dst->y_buffer + (dh - 2) * dst->y_stride, dst->y_width + 1);
+
+  Scale2D((unsigned char *)src->u_buffer, src->uv_stride, src->uv_width,
+          src->uv_height, (unsigned char *)dst->u_buffer, dst->uv_stride,
+          dw / 2, dh / 2, temp_area, temp_height, hscale, hratio, vscale,
+          vratio, interlaced);
+
+  if (dw / 2 < (int)dst->uv_width)
+    for (i = 0; i < dst->uv_height; ++i)
+      memset(dst->u_buffer + i * dst->uv_stride + dw / 2 - 1,
+             dst->u_buffer[i * dst->uv_stride + dw / 2 - 2],
+             dst->uv_width - dw / 2 + 1);
+
+  if (dh / 2 < (int)dst->uv_height)
+    for (i = dh / 2 - 1; i < (int)dst->y_height / 2; ++i)
+      memcpy(dst->u_buffer + i * dst->uv_stride,
+             dst->u_buffer + (dh / 2 - 2) * dst->uv_stride, dst->uv_width);
+
+  Scale2D((unsigned char *)src->v_buffer, src->uv_stride, src->uv_width,
+          src->uv_height, (unsigned char *)dst->v_buffer, dst->uv_stride,
+          dw / 2, dh / 2, temp_area, temp_height, hscale, hratio, vscale,
+          vratio, interlaced);
+
+  if (dw / 2 < (int)dst->uv_width)
+    for (i = 0; i < dst->uv_height; ++i)
+      memset(dst->v_buffer + i * dst->uv_stride + dw / 2 - 1,
+             dst->v_buffer[i * dst->uv_stride + dw / 2 - 2],
+             dst->uv_width - dw / 2 + 1);
+
+  if (dh / 2 < (int)dst->uv_height)
+    for (i = dh / 2 - 1; i < (int)dst->y_height / 2; ++i)
+      memcpy(dst->v_buffer + i * dst->uv_stride,
+             dst->v_buffer + (dh / 2 - 2) * dst->uv_stride, dst->uv_width);
+}
diff --git a/third_party/aom/aom_scale/generic/gen_scalers.c b/third_party/aom/aom_scale/generic/gen_scalers.c
new file mode 100644
index 0000000000..71fa82f308
--- /dev/null
+++ b/third_party/aom/aom_scale/generic/gen_scalers.c
@@ -0,0 +1,200 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_scale_rtcd.h"
+#include "aom_scale/aom_scale.h"
+#include "aom_mem/aom_mem.h"
+/****************************************************************************
+*  Imports
+****************************************************************************/
+
+/****************************************************************************
+ *
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to source data.
+ *                  unsigned int source_width   : Stride of source.
+ *                  unsigned char *dest         : Pointer to destination data.
+ *                  unsigned int dest_width     : Stride of destination
+ *                                                (NOT USED).
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Copies horizontal line of pixels from source to
+ *                  destination scaling up by 4 to 5.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ ****************************************************************************/
+void aom_horizontal_line_5_4_scale_c(const unsigned char *source,
+                                     unsigned int source_width,
+                                     unsigned char *dest,
+                                     unsigned int dest_width) {
+  const unsigned char *const source_end = source + source_width;
+  (void)dest_width;
+
+  while (source < source_end) {
+    const unsigned int a = source[0];
+    const unsigned int b = source[1];
+    const unsigned int c = source[2];
+    const unsigned int d = source[3];
+    const unsigned int e = source[4];
+
+    dest[0] = (unsigned char)a;
+    dest[1] = (unsigned char)((b * 192 + c * 64 + 128) >> 8);
+    dest[2] = (unsigned char)((c * 128 + d * 128 + 128) >> 8);
+    dest[3] = (unsigned char)((d * 64 + e * 192 + 128) >> 8);
+
+    source += 5;
+    dest += 4;
+  }
+}
+
+void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch,
+                                   unsigned char *dest, int dest_pitch,
+                                   unsigned int dest_width) {
+  const unsigned char *const dest_end = dest + dest_width;
+  while (dest < dest_end) {
+    const unsigned int a = source[0 * src_pitch];
+    const unsigned int b = source[1 * src_pitch];
+    const unsigned int c = source[2 * src_pitch];
+    const unsigned int d = source[3 * src_pitch];
+    const unsigned int e = source[4 * src_pitch];
+
+    dest[0 * dest_pitch] = (unsigned char)a;
+    dest[1 * dest_pitch] = (unsigned char)((b * 192 + c * 64 + 128) >> 8);
+    dest[2 * dest_pitch] = (unsigned char)((c * 128 + d * 128 + 128) >> 8);
+    dest[3 * dest_pitch] = (unsigned char)((d * 64 + e * 192 + 128) >> 8);
+
+    ++source;
+    ++dest;
+  }
+}
+
+/*7***************************************************************************
+ *
+ *  ROUTINE       : aom_horizontal_line_3_5_scale_c
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to source data.
+ *                  unsigned int source_width   : Stride of source.
+ *                  unsigned char *dest         : Pointer to destination data.
+ *                  unsigned int dest_width     : Stride of destination
+ *                                                (NOT USED).
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Copies horizontal line of pixels from source to
+ *                  destination scaling up by 3 to 5.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ *
+ ****************************************************************************/
+void aom_horizontal_line_5_3_scale_c(const unsigned char *source,
+                                     unsigned int source_width,
+                                     unsigned char *dest,
+                                     unsigned int dest_width) {
+  const unsigned char *const source_end = source + source_width;
+  (void)dest_width;
+  while (source < source_end) {
+    const unsigned int a = source[0];
+    const unsigned int b = source[1];
+    const unsigned int c = source[2];
+    const unsigned int d = source[3];
+    const unsigned int e = source[4];
+
+    dest[0] = (unsigned char)a;
+    dest[1] = (unsigned char)((b * 85 + c * 171 + 128) >> 8);
+    dest[2] = (unsigned char)((d * 171 + e * 85 + 128) >> 8);
+
+    source += 5;
+    dest += 3;
+  }
+}
+
+void aom_vertical_band_5_3_scale_c(unsigned char *source, int src_pitch,
+                                   unsigned char *dest, int dest_pitch,
+                                   unsigned int dest_width) {
+  const unsigned char *const dest_end = dest + dest_width;
+  while (dest < dest_end) {
+    const unsigned int a = source[0 * src_pitch];
+    const unsigned int b = source[1 * src_pitch];
+    const unsigned int c = source[2 * src_pitch];
+    const unsigned int d = source[3 * src_pitch];
+    const unsigned int e = source[4 * src_pitch];
+
+    dest[0 * dest_pitch] = (unsigned char)a;
+    dest[1 * dest_pitch] = (unsigned char)((b * 85 + c * 171 + 128) >> 8);
+    dest[2 * dest_pitch] = (unsigned char)((d * 171 + e * 85 + 128) >> 8);
+
+    ++source;
+    ++dest;
+  }
+}
+
+/****************************************************************************
+ *
+ *  ROUTINE       : aom_horizontal_line_1_2_scale_c
+ *
+ *  INPUTS        : const unsigned char *source : Pointer to source data.
+ *                  unsigned int source_width   : Stride of source.
+ *                  unsigned char *dest         : Pointer to destination data.
+ *                  unsigned int dest_width     : Stride of destination
+ *                                                (NOT USED).
+ *
+ *  OUTPUTS       : None.
+ *
+ *  RETURNS       : void
+ *
+ *  FUNCTION      : Copies horizontal line of pixels from source to
+ *                  destination scaling up by 1 to 2.
+ *
+ *  SPECIAL NOTES : None.
+ *
+ ****************************************************************************/
+void aom_horizontal_line_2_1_scale_c(const unsigned char *source,
+                                     unsigned int source_width,
+                                     unsigned char *dest,
+                                     unsigned int dest_width) {
+  const unsigned char *const source_end = source + source_width;
+  (void)dest_width;
+  while (source < source_end) {
+    dest[0] = source[0];
+    source += 2;
+    ++dest;
+  }
+}
+
+void aom_vertical_band_2_1_scale_c(unsigned char *source, int src_pitch,
+                                   unsigned char *dest, int dest_pitch,
+                                   unsigned int dest_width) {
+  (void)dest_pitch;
+  (void)src_pitch;
+  memcpy(dest, source, dest_width);
+}
+
+void aom_vertical_band_2_1_scale_i_c(unsigned char *source, int src_pitch,
+                                     unsigned char *dest, int dest_pitch,
+                                     unsigned int dest_width) {
+  const unsigned char *const dest_end = dest + dest_width;
+  (void)dest_pitch;
+  while (dest < dest_end) {
+    const unsigned int a = source[-src_pitch] * 3;
+    const unsigned int b = source[0] * 10;
+    const unsigned int c = source[src_pitch] * 3;
+    dest[0] = (unsigned char)((8 + a + b + c) >> 4);
+    ++source;
+    ++dest;
+  }
+}
diff --git a/third_party/aom/aom_scale/generic/yv12config.c b/third_party/aom/aom_scale/generic/yv12config.c
new file mode 100644
index 0000000000..ee15ae103c
--- /dev/null
+++ b/third_party/aom/aom_scale/generic/yv12config.c
@@ -0,0 +1,199 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "aom_scale/yv12config.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+/****************************************************************************
+*  Exports
+****************************************************************************/
+
+/****************************************************************************
+ *
+ ****************************************************************************/
+#define yv12_align_addr(addr, align) \
+  (void *)(((size_t)(addr) + ((align)-1)) & (size_t) - (align))
+
+#if CONFIG_AV1
+// TODO(jkoleszar): Maybe replace this with struct aom_image
+
+int aom_free_frame_buffer(YV12_BUFFER_CONFIG *ybf) {
+  if (ybf) {
+    if (ybf->buffer_alloc_sz > 0) {
+      aom_free(ybf->buffer_alloc);
+    }
+
+#if CONFIG_HIGHBITDEPTH && CONFIG_GLOBAL_MOTION
+    if (ybf->y_buffer_8bit) free(ybf->y_buffer_8bit);
+#endif
+
+    /* buffer_alloc isn't accessed by most functions.  Rather y_buffer,
+      u_buffer and v_buffer point to buffer_alloc and are used.  Clear out
+      all of this so that a freed pointer isn't inadvertently used */
+    memset(ybf, 0, sizeof(YV12_BUFFER_CONFIG));
+  } else {
+    return -1;
+  }
+
+  return 0;
+}
+
+int aom_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf, int width, int height,
+                             int ss_x, int ss_y,
+#if CONFIG_HIGHBITDEPTH
+                             int use_highbitdepth,
+#endif
+                             int border, int byte_alignment,
+                             aom_codec_frame_buffer_t *fb,
+                             aom_get_frame_buffer_cb_fn_t cb, void *cb_priv) {
+  if (ybf) {
+    const int aom_byte_align = (byte_alignment == 0) ? 1 : byte_alignment;
+    const int aligned_width = (width + 7) & ~7;
+    const int aligned_height = (height + 7) & ~7;
+    const int y_stride = ((aligned_width + 2 * border) + 31) & ~31;
+    const uint64_t yplane_size =
+        (aligned_height + 2 * border) * (uint64_t)y_stride + byte_alignment;
+    const int uv_width = aligned_width >> ss_x;
+    const int uv_height = aligned_height >> ss_y;
+    const int uv_stride = y_stride >> ss_x;
+    const int uv_border_w = border >> ss_x;
+    const int uv_border_h = border >> ss_y;
+    const uint64_t uvplane_size =
+        (uv_height + 2 * uv_border_h) * (uint64_t)uv_stride + byte_alignment;
+
+#if CONFIG_HIGHBITDEPTH
+    const uint64_t frame_size =
+        (1 + use_highbitdepth) * (yplane_size + 2 * uvplane_size);
+#else
+    const uint64_t frame_size = yplane_size + 2 * uvplane_size;
+#endif  // CONFIG_HIGHBITDEPTH
+
+    uint8_t *buf = NULL;
+
+    if (cb != NULL) {
+      const int align_addr_extra_size = 31;
+      const uint64_t external_frame_size = frame_size + align_addr_extra_size;
+
+      assert(fb != NULL);
+
+      if (external_frame_size != (size_t)external_frame_size) return -1;
+
+      // Allocation to hold larger frame, or first allocation.
+      if (cb(cb_priv, (size_t)external_frame_size, fb) < 0) return -1;
+
+      if (fb->data == NULL || fb->size < external_frame_size) return -1;
+
+      ybf->buffer_alloc = (uint8_t *)yv12_align_addr(fb->data, 32);
+
+#if defined(__has_feature)
+#if __has_feature(memory_sanitizer)
+      // This memset is needed for fixing the issue of using uninitialized
+      // value in msan test. It will cause a perf loss, so only do this for
+      // msan test.
+      memset(ybf->buffer_alloc, 0, (int)frame_size);
+#endif
+#endif
+    } else if (frame_size > (size_t)ybf->buffer_alloc_sz) {
+      // Allocation to hold larger frame, or first allocation.
+      aom_free(ybf->buffer_alloc);
+      ybf->buffer_alloc = NULL;
+
+      if (frame_size != (size_t)frame_size) return -1;
+
+      ybf->buffer_alloc = (uint8_t *)aom_memalign(32, (size_t)frame_size);
+      if (!ybf->buffer_alloc) return -1;
+
+      ybf->buffer_alloc_sz = (size_t)frame_size;
+
+      // This memset is needed for fixing valgrind error from C loop filter
+      // due to access uninitialized memory in frame border. It could be
+      // removed if border is totally removed.
+      memset(ybf->buffer_alloc, 0, ybf->buffer_alloc_sz);
+    }
+
+    /* Only support allocating buffers that have a border that's a multiple
+     * of 32. The border restriction is required to get 16-byte alignment of
+     * the start of the chroma rows without introducing an arbitrary gap
+     * between planes, which would break the semantics of things like
+     * aom_img_set_rect(). */
+    if (border & 0x1f) return -3;
+
+    ybf->y_crop_width = width;
+    ybf->y_crop_height = height;
+    ybf->y_width = aligned_width;
+    ybf->y_height = aligned_height;
+    ybf->y_stride = y_stride;
+
+    ybf->uv_crop_width = (width + ss_x) >> ss_x;
+    ybf->uv_crop_height = (height + ss_y) >> ss_y;
+    ybf->uv_width = uv_width;
+    ybf->uv_height = uv_height;
+    ybf->uv_stride = uv_stride;
+
+    ybf->border = border;
+    ybf->frame_size = (size_t)frame_size;
+    ybf->subsampling_x = ss_x;
+    ybf->subsampling_y = ss_y;
+
+    buf = ybf->buffer_alloc;
+#if CONFIG_HIGHBITDEPTH
+    if (use_highbitdepth) {
+      // Store uint16 addresses when using 16bit framebuffers
+      buf = CONVERT_TO_BYTEPTR(ybf->buffer_alloc);
+      ybf->flags = YV12_FLAG_HIGHBITDEPTH;
+    } else {
+      ybf->flags = 0;
+    }
+#endif  // CONFIG_HIGHBITDEPTH
+
+    ybf->y_buffer = (uint8_t *)yv12_align_addr(
+        buf + (border * y_stride) + border, aom_byte_align);
+    ybf->u_buffer = (uint8_t *)yv12_align_addr(
+        buf + yplane_size + (uv_border_h * uv_stride) + uv_border_w,
+        aom_byte_align);
+    ybf->v_buffer =
+        (uint8_t *)yv12_align_addr(buf + yplane_size + uvplane_size +
+                                       (uv_border_h * uv_stride) + uv_border_w,
+                                   aom_byte_align);
+
+#if CONFIG_HIGHBITDEPTH && CONFIG_GLOBAL_MOTION
+    if (ybf->y_buffer_8bit) {
+      free(ybf->y_buffer_8bit);
+      ybf->y_buffer_8bit = NULL;
+    }
+#endif
+
+    ybf->corrupted = 0; /* assume not corrupted by errors */
+    return 0;
+  }
+  return -2;
+}
+
+int aom_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf, int width, int height,
+                           int ss_x, int ss_y,
+#if CONFIG_HIGHBITDEPTH
+                           int use_highbitdepth,
+#endif
+                           int border, int byte_alignment) {
+  if (ybf) {
+    aom_free_frame_buffer(ybf);
+    return aom_realloc_frame_buffer(ybf, width, height, ss_x, ss_y,
+#if CONFIG_HIGHBITDEPTH
+                                    use_highbitdepth,
+#endif
+                                    border, byte_alignment, NULL, NULL, NULL);
+  }
+  return -2;
+}
+#endif
diff --git a/third_party/aom/aom_scale/generic/yv12extend.c b/third_party/aom/aom_scale/generic/yv12extend.c
new file mode 100644
index 0000000000..05e4633625
--- /dev/null
+++ b/third_party/aom/aom_scale/generic/yv12extend.c
@@ -0,0 +1,382 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include "./aom_config.h"
+#include "./aom_scale_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_scale/yv12config.h"
+
+static void extend_plane(uint8_t *const src, int src_stride, int width,
+                         int height, int extend_top, int extend_left,
+                         int extend_bottom, int extend_right) {
+  int i;
+  const int linesize = extend_left + extend_right + width;
+
+  /* copy the left and right most columns out */
+  uint8_t *src_ptr1 = src;
+  uint8_t *src_ptr2 = src + width - 1;
+  uint8_t *dst_ptr1 = src - extend_left;
+  uint8_t *dst_ptr2 = src + width;
+
+  for (i = 0; i < height; ++i) {
+    memset(dst_ptr1, src_ptr1[0], extend_left);
+    memset(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_stride;
+    src_ptr2 += src_stride;
+    dst_ptr1 += src_stride;
+    dst_ptr2 += src_stride;
+  }
+
+  /* Now copy the top and bottom lines into each line of the respective
+   * borders
+   */
+  src_ptr1 = src - extend_left;
+  src_ptr2 = src + src_stride * (height - 1) - extend_left;
+  dst_ptr1 = src + src_stride * -extend_top - extend_left;
+  dst_ptr2 = src + src_stride * height - extend_left;
+
+  for (i = 0; i < extend_top; ++i) {
+    memcpy(dst_ptr1, src_ptr1, linesize);
+    dst_ptr1 += src_stride;
+  }
+
+  for (i = 0; i < extend_bottom; ++i) {
+    memcpy(dst_ptr2, src_ptr2, linesize);
+    dst_ptr2 += src_stride;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void extend_plane_high(uint8_t *const src8, int src_stride, int width,
+                              int height, int extend_top, int extend_left,
+                              int extend_bottom, int extend_right) {
+  int i;
+  const int linesize = extend_left + extend_right + width;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+
+  /* copy the left and right most columns out */
+  uint16_t *src_ptr1 = src;
+  uint16_t *src_ptr2 = src + width - 1;
+  uint16_t *dst_ptr1 = src - extend_left;
+  uint16_t *dst_ptr2 = src + width;
+
+  for (i = 0; i < height; ++i) {
+    aom_memset16(dst_ptr1, src_ptr1[0], extend_left);
+    aom_memset16(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_stride;
+    src_ptr2 += src_stride;
+    dst_ptr1 += src_stride;
+    dst_ptr2 += src_stride;
+  }
+
+  /* Now copy the top and bottom lines into each line of the respective
+   * borders
+   */
+  src_ptr1 = src - extend_left;
+  src_ptr2 = src + src_stride * (height - 1) - extend_left;
+  dst_ptr1 = src + src_stride * -extend_top - extend_left;
+  dst_ptr2 = src + src_stride * height - extend_left;
+
+  for (i = 0; i < extend_top; ++i) {
+    memcpy(dst_ptr1, src_ptr1, linesize * sizeof(uint16_t));
+    dst_ptr1 += src_stride;
+  }
+
+  for (i = 0; i < extend_bottom; ++i) {
+    memcpy(dst_ptr2, src_ptr2, linesize * sizeof(uint16_t));
+    dst_ptr2 += src_stride;
+  }
+}
+#endif
+
+void aom_yv12_extend_frame_borders_c(YV12_BUFFER_CONFIG *ybf) {
+  const int uv_border = ybf->border / 2;
+
+  assert(ybf->border % 2 == 0);
+  assert(ybf->y_height - ybf->y_crop_height < 16);
+  assert(ybf->y_width - ybf->y_crop_width < 16);
+  assert(ybf->y_height - ybf->y_crop_height >= 0);
+  assert(ybf->y_width - ybf->y_crop_width >= 0);
+
+#if CONFIG_HIGHBITDEPTH
+  if (ybf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    extend_plane_high(ybf->y_buffer, ybf->y_stride, ybf->y_crop_width,
+                      ybf->y_crop_height, ybf->border, ybf->border,
+                      ybf->border + ybf->y_height - ybf->y_crop_height,
+                      ybf->border + ybf->y_width - ybf->y_crop_width);
+
+    extend_plane_high(ybf->u_buffer, ybf->uv_stride, ybf->uv_crop_width,
+                      ybf->uv_crop_height, uv_border, uv_border,
+                      uv_border + ybf->uv_height - ybf->uv_crop_height,
+                      uv_border + ybf->uv_width - ybf->uv_crop_width);
+
+    extend_plane_high(ybf->v_buffer, ybf->uv_stride, ybf->uv_crop_width,
+                      ybf->uv_crop_height, uv_border, uv_border,
+                      uv_border + ybf->uv_height - ybf->uv_crop_height,
+                      uv_border + ybf->uv_width - ybf->uv_crop_width);
+    return;
+  }
+#endif
+  extend_plane(ybf->y_buffer, ybf->y_stride, ybf->y_crop_width,
+               ybf->y_crop_height, ybf->border, ybf->border,
+               ybf->border + ybf->y_height - ybf->y_crop_height,
+               ybf->border + ybf->y_width - ybf->y_crop_width);
+
+  extend_plane(ybf->u_buffer, ybf->uv_stride, ybf->uv_crop_width,
+               ybf->uv_crop_height, uv_border, uv_border,
+               uv_border + ybf->uv_height - ybf->uv_crop_height,
+               uv_border + ybf->uv_width - ybf->uv_crop_width);
+
+  extend_plane(ybf->v_buffer, ybf->uv_stride, ybf->uv_crop_width,
+               ybf->uv_crop_height, uv_border, uv_border,
+               uv_border + ybf->uv_height - ybf->uv_crop_height,
+               uv_border + ybf->uv_width - ybf->uv_crop_width);
+}
+
+#if CONFIG_AV1
+static void extend_frame(YV12_BUFFER_CONFIG *const ybf, int ext_size) {
+  const int c_w = ybf->uv_crop_width;
+  const int c_h = ybf->uv_crop_height;
+  const int ss_x = ybf->uv_width < ybf->y_width;
+  const int ss_y = ybf->uv_height < ybf->y_height;
+  const int c_et = ext_size >> ss_y;
+  const int c_el = ext_size >> ss_x;
+  const int c_eb = c_et + ybf->uv_height - ybf->uv_crop_height;
+  const int c_er = c_el + ybf->uv_width - ybf->uv_crop_width;
+
+  assert(ybf->y_height - ybf->y_crop_height < 16);
+  assert(ybf->y_width - ybf->y_crop_width < 16);
+  assert(ybf->y_height - ybf->y_crop_height >= 0);
+  assert(ybf->y_width - ybf->y_crop_width >= 0);
+
+#if CONFIG_HIGHBITDEPTH
+  if (ybf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    extend_plane_high(ybf->y_buffer, ybf->y_stride, ybf->y_crop_width,
+                      ybf->y_crop_height, ext_size, ext_size,
+                      ext_size + ybf->y_height - ybf->y_crop_height,
+                      ext_size + ybf->y_width - ybf->y_crop_width);
+    extend_plane_high(ybf->u_buffer, ybf->uv_stride, c_w, c_h, c_et, c_el, c_eb,
+                      c_er);
+    extend_plane_high(ybf->v_buffer, ybf->uv_stride, c_w, c_h, c_et, c_el, c_eb,
+                      c_er);
+    return;
+  }
+#endif
+  extend_plane(ybf->y_buffer, ybf->y_stride, ybf->y_crop_width,
+               ybf->y_crop_height, ext_size, ext_size,
+               ext_size + ybf->y_height - ybf->y_crop_height,
+               ext_size + ybf->y_width - ybf->y_crop_width);
+
+  extend_plane(ybf->u_buffer, ybf->uv_stride, c_w, c_h, c_et, c_el, c_eb, c_er);
+
+  extend_plane(ybf->v_buffer, ybf->uv_stride, c_w, c_h, c_et, c_el, c_eb, c_er);
+}
+
+void aom_extend_frame_borders_c(YV12_BUFFER_CONFIG *ybf) {
+  extend_frame(ybf, ybf->border);
+}
+
+void aom_extend_frame_inner_borders_c(YV12_BUFFER_CONFIG *ybf) {
+  const int inner_bw = (ybf->border > AOMINNERBORDERINPIXELS)
+                           ? AOMINNERBORDERINPIXELS
+                           : ybf->border;
+  extend_frame(ybf, inner_bw);
+}
+
+void aom_extend_frame_borders_y_c(YV12_BUFFER_CONFIG *ybf) {
+  int ext_size = ybf->border;
+  assert(ybf->y_height - ybf->y_crop_height < 16);
+  assert(ybf->y_width - ybf->y_crop_width < 16);
+  assert(ybf->y_height - ybf->y_crop_height >= 0);
+  assert(ybf->y_width - ybf->y_crop_width >= 0);
+
+#if CONFIG_HIGHBITDEPTH
+  if (ybf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    extend_plane_high(ybf->y_buffer, ybf->y_stride, ybf->y_crop_width,
+                      ybf->y_crop_height, ext_size, ext_size,
+                      ext_size + ybf->y_height - ybf->y_crop_height,
+                      ext_size + ybf->y_width - ybf->y_crop_width);
+    return;
+  }
+#endif
+  extend_plane(ybf->y_buffer, ybf->y_stride, ybf->y_crop_width,
+               ybf->y_crop_height, ext_size, ext_size,
+               ext_size + ybf->y_height - ybf->y_crop_height,
+               ext_size + ybf->y_width - ybf->y_crop_width);
+}
+#endif  // CONFIG_AV1
+
+#if CONFIG_HIGHBITDEPTH
+static void memcpy_short_addr(uint8_t *dst8, const uint8_t *src8, int num) {
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  memcpy(dst, src, num * sizeof(uint16_t));
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+// Copies the source image into the destination image and updates the
+// destination's UMV borders.
+// Note: The frames are assumed to be identical in size.
+void aom_yv12_copy_frame_c(const YV12_BUFFER_CONFIG *src_bc,
+                           YV12_BUFFER_CONFIG *dst_bc) {
+  int row;
+  const uint8_t *src = src_bc->y_buffer;
+  uint8_t *dst = dst_bc->y_buffer;
+
+#if 0
+  /* These assertions are valid in the codec, but the libaom-tester uses
+   * this code slightly differently.
+   */
+  assert(src_bc->y_width == dst_bc->y_width);
+  assert(src_bc->y_height == dst_bc->y_height);
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+  if (src_bc->flags & YV12_FLAG_HIGHBITDEPTH) {
+    assert(dst_bc->flags & YV12_FLAG_HIGHBITDEPTH);
+    for (row = 0; row < src_bc->y_height; ++row) {
+      memcpy_short_addr(dst, src, src_bc->y_width);
+      src += src_bc->y_stride;
+      dst += dst_bc->y_stride;
+    }
+
+    src = src_bc->u_buffer;
+    dst = dst_bc->u_buffer;
+
+    for (row = 0; row < src_bc->uv_height; ++row) {
+      memcpy_short_addr(dst, src, src_bc->uv_width);
+      src += src_bc->uv_stride;
+      dst += dst_bc->uv_stride;
+    }
+
+    src = src_bc->v_buffer;
+    dst = dst_bc->v_buffer;
+
+    for (row = 0; row < src_bc->uv_height; ++row) {
+      memcpy_short_addr(dst, src, src_bc->uv_width);
+      src += src_bc->uv_stride;
+      dst += dst_bc->uv_stride;
+    }
+
+    aom_yv12_extend_frame_borders_c(dst_bc);
+    return;
+  } else {
+    assert(!(dst_bc->flags & YV12_FLAG_HIGHBITDEPTH));
+  }
+#endif
+
+  for (row = 0; row < src_bc->y_height; ++row) {
+    memcpy(dst, src, src_bc->y_width);
+    src += src_bc->y_stride;
+    dst += dst_bc->y_stride;
+  }
+
+  src = src_bc->u_buffer;
+  dst = dst_bc->u_buffer;
+
+  for (row = 0; row < src_bc->uv_height; ++row) {
+    memcpy(dst, src, src_bc->uv_width);
+    src += src_bc->uv_stride;
+    dst += dst_bc->uv_stride;
+  }
+
+  src = src_bc->v_buffer;
+  dst = dst_bc->v_buffer;
+
+  for (row = 0; row < src_bc->uv_height; ++row) {
+    memcpy(dst, src, src_bc->uv_width);
+    src += src_bc->uv_stride;
+    dst += dst_bc->uv_stride;
+  }
+
+  aom_yv12_extend_frame_borders_c(dst_bc);
+}
+
+void aom_yv12_copy_y_c(const YV12_BUFFER_CONFIG *src_ybc,
+                       YV12_BUFFER_CONFIG *dst_ybc) {
+  int row;
+  const uint8_t *src = src_ybc->y_buffer;
+  uint8_t *dst = dst_ybc->y_buffer;
+
+#if CONFIG_HIGHBITDEPTH
+  if (src_ybc->flags & YV12_FLAG_HIGHBITDEPTH) {
+    const uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+    uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
+    for (row = 0; row < src_ybc->y_height; ++row) {
+      memcpy(dst16, src16, src_ybc->y_width * sizeof(uint16_t));
+      src16 += src_ybc->y_stride;
+      dst16 += dst_ybc->y_stride;
+    }
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  for (row = 0; row < src_ybc->y_height; ++row) {
+    memcpy(dst, src, src_ybc->y_width);
+    src += src_ybc->y_stride;
+    dst += dst_ybc->y_stride;
+  }
+}
+
+void aom_yv12_copy_u_c(const YV12_BUFFER_CONFIG *src_bc,
+                       YV12_BUFFER_CONFIG *dst_bc) {
+  int row;
+  const uint8_t *src = src_bc->u_buffer;
+  uint8_t *dst = dst_bc->u_buffer;
+
+#if CONFIG_HIGHBITDEPTH
+  if (src_bc->flags & YV12_FLAG_HIGHBITDEPTH) {
+    const uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+    uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
+    for (row = 0; row < src_bc->uv_height; ++row) {
+      memcpy(dst16, src16, src_bc->uv_width * sizeof(uint16_t));
+      src16 += src_bc->uv_stride;
+      dst16 += dst_bc->uv_stride;
+    }
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  for (row = 0; row < src_bc->uv_height; ++row) {
+    memcpy(dst, src, src_bc->uv_width);
+    src += src_bc->uv_stride;
+    dst += dst_bc->uv_stride;
+  }
+}
+
+void aom_yv12_copy_v_c(const YV12_BUFFER_CONFIG *src_bc,
+                       YV12_BUFFER_CONFIG *dst_bc) {
+  int row;
+  const uint8_t *src = src_bc->v_buffer;
+  uint8_t *dst = dst_bc->v_buffer;
+
+#if CONFIG_HIGHBITDEPTH
+  if (src_bc->flags & YV12_FLAG_HIGHBITDEPTH) {
+    const uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+    uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
+    for (row = 0; row < src_bc->uv_height; ++row) {
+      memcpy(dst16, src16, src_bc->uv_width * sizeof(uint16_t));
+      src16 += src_bc->uv_stride;
+      dst16 += dst_bc->uv_stride;
+    }
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  for (row = 0; row < src_bc->uv_height; ++row) {
+    memcpy(dst, src, src_bc->uv_width);
+    src += src_bc->uv_stride;
+    dst += dst_bc->uv_stride;
+  }
+}
diff --git a/third_party/aom/aom_scale/mips/dspr2/yv12extend_dspr2.c b/third_party/aom/aom_scale/mips/dspr2/yv12extend_dspr2.c
new file mode 100644
index 0000000000..51192f7b93
--- /dev/null
+++ b/third_party/aom/aom_scale/mips/dspr2/yv12extend_dspr2.c
@@ -0,0 +1,139 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "aom_scale/yv12config.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_scale/aom_scale.h"
+
+#if HAVE_DSPR2
+static void extend_plane(uint8_t *const src, int src_stride, int width,
+                         int height, int extend_top, int extend_left,
+                         int extend_bottom, int extend_right) {
+  int i, j;
+  uint8_t *left_src, *right_src;
+  uint8_t *left_dst_start, *right_dst_start;
+  uint8_t *left_dst, *right_dst;
+  uint8_t *top_src, *bot_src;
+  uint8_t *top_dst, *bot_dst;
+  uint32_t left_pix;
+  uint32_t right_pix;
+  uint32_t linesize;
+
+  /* copy the left and right most columns out */
+  left_src = src;
+  right_src = src + width - 1;
+  left_dst_start = src - extend_left;
+  right_dst_start = src + width;
+
+  for (i = height; i--;) {
+    left_dst = left_dst_start;
+    right_dst = right_dst_start;
+
+    __asm__ __volatile__(
+        "lb        %[left_pix],     0(%[left_src])      \n\t"
+        "lb        %[right_pix],    0(%[right_src])     \n\t"
+        "replv.qb  %[left_pix],     %[left_pix]         \n\t"
+        "replv.qb  %[right_pix],    %[right_pix]        \n\t"
+
+        : [left_pix] "=&r"(left_pix), [right_pix] "=&r"(right_pix)
+        : [left_src] "r"(left_src), [right_src] "r"(right_src));
+
+    for (j = extend_left / 4; j--;) {
+      __asm__ __volatile__(
+          "sw     %[left_pix],    0(%[left_dst])     \n\t"
+          "sw     %[right_pix],   0(%[right_dst])    \n\t"
+
+          :
+          : [left_dst] "r"(left_dst), [left_pix] "r"(left_pix),
+            [right_dst] "r"(right_dst), [right_pix] "r"(right_pix));
+
+      left_dst += 4;
+      right_dst += 4;
+    }
+
+    for (j = extend_left % 4; j--;) {
+      __asm__ __volatile__(
+          "sb     %[left_pix],    0(%[left_dst])     \n\t"
+          "sb     %[right_pix],   0(%[right_dst])     \n\t"
+
+          :
+          : [left_dst] "r"(left_dst), [left_pix] "r"(left_pix),
+            [right_dst] "r"(right_dst), [right_pix] "r"(right_pix));
+
+      left_dst += 1;
+      right_dst += 1;
+    }
+
+    left_src += src_stride;
+    right_src += src_stride;
+    left_dst_start += src_stride;
+    right_dst_start += src_stride;
+  }
+
+  /* Now copy the top and bottom lines into each line of the respective
+   * borders
+   */
+  top_src = src - extend_left;
+  bot_src = src + src_stride * (height - 1) - extend_left;
+  top_dst = src + src_stride * (-extend_top) - extend_left;
+  bot_dst = src + src_stride * (height)-extend_left;
+  linesize = extend_left + extend_right + width;
+
+  for (i = 0; i < extend_top; i++) {
+    memcpy(top_dst, top_src, linesize);
+    top_dst += src_stride;
+  }
+
+  for (i = 0; i < extend_bottom; i++) {
+    memcpy(bot_dst, bot_src, linesize);
+    bot_dst += src_stride;
+  }
+}
+
+static void extend_frame(YV12_BUFFER_CONFIG *const ybf, int ext_size) {
+  const int c_w = ybf->uv_crop_width;
+  const int c_h = ybf->uv_crop_height;
+  const int ss_x = ybf->uv_width < ybf->y_width;
+  const int ss_y = ybf->uv_height < ybf->y_height;
+  const int c_et = ext_size >> ss_y;
+  const int c_el = ext_size >> ss_x;
+  const int c_eb = c_et + ybf->uv_height - ybf->uv_crop_height;
+  const int c_er = c_el + ybf->uv_width - ybf->uv_crop_width;
+
+  assert(ybf->y_height - ybf->y_crop_height < 16);
+  assert(ybf->y_width - ybf->y_crop_width < 16);
+  assert(ybf->y_height - ybf->y_crop_height >= 0);
+  assert(ybf->y_width - ybf->y_crop_width >= 0);
+
+  extend_plane(ybf->y_buffer, ybf->y_stride, ybf->y_crop_width,
+               ybf->y_crop_height, ext_size, ext_size,
+               ext_size + ybf->y_height - ybf->y_crop_height,
+               ext_size + ybf->y_width - ybf->y_crop_width);
+
+  extend_plane(ybf->u_buffer, ybf->uv_stride, c_w, c_h, c_et, c_el, c_eb, c_er);
+
+  extend_plane(ybf->v_buffer, ybf->uv_stride, c_w, c_h, c_et, c_el, c_eb, c_er);
+}
+
+void aom_extend_frame_borders_dspr2(YV12_BUFFER_CONFIG *ybf) {
+  extend_frame(ybf, ybf->border);
+}
+
+void aom_extend_frame_inner_borders_dspr2(YV12_BUFFER_CONFIG *ybf) {
+  const int inner_bw = (ybf->border > AOMINNERBORDERINPIXELS)
+                           ? AOMINNERBORDERINPIXELS
+                           : ybf->border;
+  extend_frame(ybf, inner_bw);
+}
+#endif
diff --git a/third_party/aom/aom_scale/yv12config.h b/third_party/aom/aom_scale/yv12config.h
new file mode 100644
index 0000000000..434dc7b4a3
--- /dev/null
+++ b/third_party/aom/aom_scale/yv12config.h
@@ -0,0 +1,110 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_SCALE_YV12CONFIG_H_
+#define AOM_SCALE_YV12CONFIG_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./aom_config.h"
+#include "aom/aom_codec.h"
+#include "aom/aom_frame_buffer.h"
+#include "aom/aom_integer.h"
+
+#if CONFIG_EXT_PARTITION
+#define AOMINNERBORDERINPIXELS 160
+#else
+#define AOMINNERBORDERINPIXELS 96
+#endif  // CONFIG_EXT_PARTITION
+#define AOM_INTERP_EXTEND 4
+
+// TODO(jingning): Use unified inter predictor for encoder and
+// decoder during the development process. Revisit the frame border
+// to improve the decoder performance.
+#define AOM_BORDER_IN_PIXELS 160
+
+typedef struct yv12_buffer_config {
+  int y_width;
+  int y_height;
+  int y_crop_width;
+  int y_crop_height;
+  int y_stride;
+
+  int uv_width;
+  int uv_height;
+  int uv_crop_width;
+  int uv_crop_height;
+  int uv_stride;
+
+  int alpha_width;
+  int alpha_height;
+  int alpha_stride;
+
+  uint8_t *y_buffer;
+  uint8_t *u_buffer;
+  uint8_t *v_buffer;
+  uint8_t *alpha_buffer;
+
+#if CONFIG_HIGHBITDEPTH && CONFIG_GLOBAL_MOTION
+  // If the frame is stored in a 16-bit buffer, this stores an 8-bit version
+  // for use in global motion detection. It is allocated on-demand.
+  uint8_t *y_buffer_8bit;
+#endif
+
+  uint8_t *buffer_alloc;
+  size_t buffer_alloc_sz;
+  int border;
+  size_t frame_size;
+  int subsampling_x;
+  int subsampling_y;
+  unsigned int bit_depth;
+  aom_color_space_t color_space;
+  aom_color_range_t color_range;
+  int render_width;
+  int render_height;
+
+  int corrupted;
+  int flags;
+} YV12_BUFFER_CONFIG;
+
+#define YV12_FLAG_HIGHBITDEPTH 8
+
+int aom_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf, int width, int height,
+                           int ss_x, int ss_y,
+#if CONFIG_HIGHBITDEPTH
+                           int use_highbitdepth,
+#endif
+                           int border, int byte_alignment);
+
+// Updates the yv12 buffer config with the frame buffer. |byte_alignment| must
+// be a power of 2, from 32 to 1024. 0 sets legacy alignment. If cb is not
+// NULL, then libaom is using the frame buffer callbacks to handle memory.
+// If cb is not NULL, libaom will call cb with minimum size in bytes needed
+// to decode the current frame. If cb is NULL, libaom will allocate memory
+// internally to decode the current frame. Returns 0 on success. Returns < 0
+// on failure.
+int aom_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf, int width, int height,
+                             int ss_x, int ss_y,
+#if CONFIG_HIGHBITDEPTH
+                             int use_highbitdepth,
+#endif
+                             int border, int byte_alignment,
+                             aom_codec_frame_buffer_t *fb,
+                             aom_get_frame_buffer_cb_fn_t cb, void *cb_priv);
+int aom_free_frame_buffer(YV12_BUFFER_CONFIG *ybf);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // AOM_SCALE_YV12CONFIG_H_
diff --git a/third_party/aom/aom_util/aom_thread.c b/third_party/aom/aom_util/aom_thread.c
new file mode 100644
index 0000000000..954b8f99c3
--- /dev/null
+++ b/third_party/aom/aom_util/aom_thread.c
@@ -0,0 +1,183 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+//
+// Multi-threaded worker
+//
+// Original source:
+//  https://chromium.googlesource.com/webm/libwebp
+
+#include <assert.h>
+#include <string.h>  // for memset()
+#include "./aom_thread.h"
+#include "aom_mem/aom_mem.h"
+
+#if CONFIG_MULTITHREAD
+
+struct AVxWorkerImpl {
+  pthread_mutex_t mutex_;
+  pthread_cond_t condition_;
+  pthread_t thread_;
+};
+
+//------------------------------------------------------------------------------
+
+static void execute(AVxWorker *const worker);  // Forward declaration.
+
+static THREADFN thread_loop(void *ptr) {
+  AVxWorker *const worker = (AVxWorker *)ptr;
+  int done = 0;
+  while (!done) {
+    pthread_mutex_lock(&worker->impl_->mutex_);
+    while (worker->status_ == OK) {  // wait in idling mode
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
+    }
+    if (worker->status_ == WORK) {
+      execute(worker);
+      worker->status_ = OK;
+    } else if (worker->status_ == NOT_OK) {  // finish the worker
+      done = 1;
+    }
+    // signal to the main thread that we're done (for sync())
+    pthread_cond_signal(&worker->impl_->condition_);
+    pthread_mutex_unlock(&worker->impl_->mutex_);
+  }
+  return THREAD_RETURN(NULL);  // Thread is finished
+}
+
+// main thread state control
+static void change_state(AVxWorker *const worker, AVxWorkerStatus new_status) {
+  // No-op when attempting to change state on a thread that didn't come up.
+  // Checking status_ without acquiring the lock first would result in a data
+  // race.
+  if (worker->impl_ == NULL) return;
+
+  pthread_mutex_lock(&worker->impl_->mutex_);
+  if (worker->status_ >= OK) {
+    // wait for the worker to finish
+    while (worker->status_ != OK) {
+      pthread_cond_wait(&worker->impl_->condition_, &worker->impl_->mutex_);
+    }
+    // assign new status and release the working thread if needed
+    if (new_status != OK) {
+      worker->status_ = new_status;
+      pthread_cond_signal(&worker->impl_->condition_);
+    }
+  }
+  pthread_mutex_unlock(&worker->impl_->mutex_);
+}
+
+#endif  // CONFIG_MULTITHREAD
+
+//------------------------------------------------------------------------------
+
+static void init(AVxWorker *const worker) {
+  memset(worker, 0, sizeof(*worker));
+  worker->status_ = NOT_OK;
+}
+
+static int sync(AVxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  change_state(worker, OK);
+#endif
+  assert(worker->status_ <= OK);
+  return !worker->had_error;
+}
+
+static int reset(AVxWorker *const worker) {
+  int ok = 1;
+  worker->had_error = 0;
+  if (worker->status_ < OK) {
+#if CONFIG_MULTITHREAD
+    worker->impl_ = (AVxWorkerImpl *)aom_calloc(1, sizeof(*worker->impl_));
+    if (worker->impl_ == NULL) {
+      return 0;
+    }
+    if (pthread_mutex_init(&worker->impl_->mutex_, NULL)) {
+      goto Error;
+    }
+    if (pthread_cond_init(&worker->impl_->condition_, NULL)) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      goto Error;
+    }
+    pthread_mutex_lock(&worker->impl_->mutex_);
+    ok = !pthread_create(&worker->impl_->thread_, NULL, thread_loop, worker);
+    if (ok) worker->status_ = OK;
+    pthread_mutex_unlock(&worker->impl_->mutex_);
+    if (!ok) {
+      pthread_mutex_destroy(&worker->impl_->mutex_);
+      pthread_cond_destroy(&worker->impl_->condition_);
+    Error:
+      aom_free(worker->impl_);
+      worker->impl_ = NULL;
+      return 0;
+    }
+#else
+    worker->status_ = OK;
+#endif
+  } else if (worker->status_ > OK) {
+    ok = sync(worker);
+  }
+  assert(!ok || (worker->status_ == OK));
+  return ok;
+}
+
+static void execute(AVxWorker *const worker) {
+  if (worker->hook != NULL) {
+    worker->had_error |= !worker->hook(worker->data1, worker->data2);
+  }
+}
+
+static void launch(AVxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  change_state(worker, WORK);
+#else
+  execute(worker);
+#endif
+}
+
+static void end(AVxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  if (worker->impl_ != NULL) {
+    change_state(worker, NOT_OK);
+    pthread_join(worker->impl_->thread_, NULL);
+    pthread_mutex_destroy(&worker->impl_->mutex_);
+    pthread_cond_destroy(&worker->impl_->condition_);
+    aom_free(worker->impl_);
+    worker->impl_ = NULL;
+  }
+#else
+  worker->status_ = NOT_OK;
+  assert(worker->impl_ == NULL);
+#endif
+  assert(worker->status_ == NOT_OK);
+}
+
+//------------------------------------------------------------------------------
+
+static AVxWorkerInterface g_worker_interface = { init,   reset,   sync,
+                                                 launch, execute, end };
+
+int aom_set_worker_interface(const AVxWorkerInterface *const winterface) {
+  if (winterface == NULL || winterface->init == NULL ||
+      winterface->reset == NULL || winterface->sync == NULL ||
+      winterface->launch == NULL || winterface->execute == NULL ||
+      winterface->end == NULL) {
+    return 0;
+  }
+  g_worker_interface = *winterface;
+  return 1;
+}
+
+const AVxWorkerInterface *aom_get_worker_interface(void) {
+  return &g_worker_interface;
+}
+
+//------------------------------------------------------------------------------
diff --git a/third_party/aom/aom_util/aom_thread.h b/third_party/aom/aom_util/aom_thread.h
new file mode 100644
index 0000000000..0ae8f2f49d
--- /dev/null
+++ b/third_party/aom/aom_util/aom_thread.h
@@ -0,0 +1,412 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+//
+// Multi-threaded worker
+//
+// Original source:
+//  https://chromium.googlesource.com/webm/libwebp
+
+#ifndef AOM_THREAD_H_
+#define AOM_THREAD_H_
+
+#include "./aom_config.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Set maximum decode threads to be 8 due to the limit of frame buffers
+// and not enough semaphores in the emulation layer on windows.
+#define MAX_DECODE_THREADS 8
+
+#if CONFIG_MULTITHREAD
+
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+#include <errno.h>    // NOLINT
+#include <process.h>  // NOLINT
+#include <windows.h>  // NOLINT
+typedef HANDLE pthread_t;
+typedef CRITICAL_SECTION pthread_mutex_t;
+
+#if _WIN32_WINNT >= 0x0600  // Windows Vista / Server 2008 or greater
+#define USE_WINDOWS_CONDITION_VARIABLE
+typedef CONDITION_VARIABLE pthread_cond_t;
+#else
+typedef struct {
+  HANDLE waiting_sem_;
+  HANDLE received_sem_;
+  HANDLE signal_event_;
+} pthread_cond_t;
+#endif  // _WIN32_WINNT >= 0x600
+
+#ifndef WINAPI_FAMILY_PARTITION
+#define WINAPI_PARTITION_DESKTOP 1
+#define WINAPI_FAMILY_PARTITION(x) x
+#endif
+
+#if !WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
+#define USE_CREATE_THREAD
+#endif
+
+//------------------------------------------------------------------------------
+// simplistic pthread emulation layer
+
+// _beginthreadex requires __stdcall
+#define THREADFN unsigned int __stdcall
+#define THREAD_RETURN(val) (unsigned int)((DWORD_PTR)val)
+
+#if _WIN32_WINNT >= 0x0501  // Windows XP or greater
+#define WaitForSingleObject(obj, timeout) \
+  WaitForSingleObjectEx(obj, timeout, FALSE /*bAlertable*/)
+#endif
+
+static INLINE int pthread_create(pthread_t *const thread, const void *attr,
+                                 unsigned int(__stdcall *start)(void *),
+                                 void *arg) {
+  (void)attr;
+#ifdef USE_CREATE_THREAD
+  *thread = CreateThread(NULL,          /* lpThreadAttributes */
+                         0,             /* dwStackSize */
+                         start, arg, 0, /* dwStackSize */
+                         NULL);         /* lpThreadId */
+#else
+  *thread = (pthread_t)_beginthreadex(NULL,          /* void *security */
+                                      0,             /* unsigned stack_size */
+                                      start, arg, 0, /* unsigned initflag */
+                                      NULL);         /* unsigned *thrdaddr */
+#endif
+  if (*thread == NULL) return 1;
+  SetThreadPriority(*thread, THREAD_PRIORITY_ABOVE_NORMAL);
+  return 0;
+}
+
+static INLINE int pthread_join(pthread_t thread, void **value_ptr) {
+  (void)value_ptr;
+  return (WaitForSingleObject(thread, INFINITE) != WAIT_OBJECT_0 ||
+          CloseHandle(thread) == 0);
+}
+
+// Mutex
+static INLINE int pthread_mutex_init(pthread_mutex_t *const mutex,
+                                     void *mutexattr) {
+  (void)mutexattr;
+#if _WIN32_WINNT >= 0x0600  // Windows Vista / Server 2008 or greater
+  InitializeCriticalSectionEx(mutex, 0 /*dwSpinCount*/, 0 /*Flags*/);
+#else
+  InitializeCriticalSection(mutex);
+#endif
+  return 0;
+}
+
+static INLINE int pthread_mutex_trylock(pthread_mutex_t *const mutex) {
+  return TryEnterCriticalSection(mutex) ? 0 : EBUSY;
+}
+
+static INLINE int pthread_mutex_lock(pthread_mutex_t *const mutex) {
+  EnterCriticalSection(mutex);
+  return 0;
+}
+
+static INLINE int pthread_mutex_unlock(pthread_mutex_t *const mutex) {
+  LeaveCriticalSection(mutex);
+  return 0;
+}
+
+static INLINE int pthread_mutex_destroy(pthread_mutex_t *const mutex) {
+  DeleteCriticalSection(mutex);
+  return 0;
+}
+
+// Condition
+static INLINE int pthread_cond_destroy(pthread_cond_t *const condition) {
+  int ok = 1;
+#ifdef USE_WINDOWS_CONDITION_VARIABLE
+  (void)condition;
+#else
+  ok &= (CloseHandle(condition->waiting_sem_) != 0);
+  ok &= (CloseHandle(condition->received_sem_) != 0);
+  ok &= (CloseHandle(condition->signal_event_) != 0);
+#endif
+  return !ok;
+}
+
+static INLINE int pthread_cond_init(pthread_cond_t *const condition,
+                                    void *cond_attr) {
+  (void)cond_attr;
+#ifdef USE_WINDOWS_CONDITION_VARIABLE
+  InitializeConditionVariable(condition);
+#else
+  condition->waiting_sem_ = CreateSemaphore(NULL, 0, MAX_DECODE_THREADS, NULL);
+  condition->received_sem_ = CreateSemaphore(NULL, 0, MAX_DECODE_THREADS, NULL);
+  condition->signal_event_ = CreateEvent(NULL, FALSE, FALSE, NULL);
+  if (condition->waiting_sem_ == NULL || condition->received_sem_ == NULL ||
+      condition->signal_event_ == NULL) {
+    pthread_cond_destroy(condition);
+    return 1;
+  }
+#endif
+  return 0;
+}
+
+static INLINE int pthread_cond_signal(pthread_cond_t *const condition) {
+  int ok = 1;
+#ifdef USE_WINDOWS_CONDITION_VARIABLE
+  WakeConditionVariable(condition);
+#else
+  if (WaitForSingleObject(condition->waiting_sem_, 0) == WAIT_OBJECT_0) {
+    // a thread is waiting in pthread_cond_wait: allow it to be notified
+    ok = SetEvent(condition->signal_event_);
+    // wait until the event is consumed so the signaler cannot consume
+    // the event via its own pthread_cond_wait.
+    ok &= (WaitForSingleObject(condition->received_sem_, INFINITE) !=
+           WAIT_OBJECT_0);
+  }
+#endif
+  return !ok;
+}
+
+static INLINE int pthread_cond_wait(pthread_cond_t *const condition,
+                                    pthread_mutex_t *const mutex) {
+  int ok;
+#ifdef USE_WINDOWS_CONDITION_VARIABLE
+  ok = SleepConditionVariableCS(condition, mutex, INFINITE);
+#else
+  // note that there is a consumer available so the signal isn't dropped in
+  // pthread_cond_signal
+  if (!ReleaseSemaphore(condition->waiting_sem_, 1, NULL)) return 1;
+  // now unlock the mutex so pthread_cond_signal may be issued
+  pthread_mutex_unlock(mutex);
+  ok = (WaitForSingleObject(condition->signal_event_, INFINITE) ==
+        WAIT_OBJECT_0);
+  ok &= ReleaseSemaphore(condition->received_sem_, 1, NULL);
+  pthread_mutex_lock(mutex);
+#endif
+  return !ok;
+}
+#elif defined(__OS2__)
+#define INCL_DOS
+#include <os2.h>  // NOLINT
+
+#include <errno.h>        // NOLINT
+#include <stdlib.h>       // NOLINT
+#include <sys/builtin.h>  // NOLINT
+
+#define pthread_t TID
+#define pthread_mutex_t HMTX
+
+typedef struct {
+  HEV event_sem_;
+  HEV ack_sem_;
+  volatile unsigned wait_count_;
+} pthread_cond_t;
+
+//------------------------------------------------------------------------------
+// simplistic pthread emulation layer
+
+#define THREADFN void *
+#define THREAD_RETURN(val) (val)
+
+typedef struct {
+  void *(*start_)(void *);
+  void *arg_;
+} thread_arg;
+
+static void thread_start(void *arg) {
+  thread_arg targ = *(thread_arg *)arg;
+  free(arg);
+
+  targ.start_(targ.arg_);
+}
+
+static INLINE int pthread_create(pthread_t *const thread, const void *attr,
+                                 void *(*start)(void *), void *arg) {
+  int tid;
+  thread_arg *targ = (thread_arg *)malloc(sizeof(*targ));
+  if (targ == NULL) return 1;
+
+  (void)attr;
+
+  targ->start_ = start;
+  targ->arg_ = arg;
+  tid = (pthread_t)_beginthread(thread_start, NULL, 1024 * 1024, targ);
+  if (tid == -1) {
+    free(targ);
+    return 1;
+  }
+
+  *thread = tid;
+  return 0;
+}
+
+static INLINE int pthread_join(pthread_t thread, void **value_ptr) {
+  (void)value_ptr;
+  return DosWaitThread(&thread, DCWW_WAIT) != 0;
+}
+
+// Mutex
+static INLINE int pthread_mutex_init(pthread_mutex_t *const mutex,
+                                     void *mutexattr) {
+  (void)mutexattr;
+  return DosCreateMutexSem(NULL, mutex, 0, FALSE) != 0;
+}
+
+static INLINE int pthread_mutex_trylock(pthread_mutex_t *const mutex) {
+  return DosRequestMutexSem(*mutex, SEM_IMMEDIATE_RETURN) == 0 ? 0 : EBUSY;
+}
+
+static INLINE int pthread_mutex_lock(pthread_mutex_t *const mutex) {
+  return DosRequestMutexSem(*mutex, SEM_INDEFINITE_WAIT) != 0;
+}
+
+static INLINE int pthread_mutex_unlock(pthread_mutex_t *const mutex) {
+  return DosReleaseMutexSem(*mutex) != 0;
+}
+
+static INLINE int pthread_mutex_destroy(pthread_mutex_t *const mutex) {
+  return DosCloseMutexSem(*mutex) != 0;
+}
+
+// Condition
+static INLINE int pthread_cond_destroy(pthread_cond_t *const condition) {
+  int ok = 1;
+  ok &= DosCloseEventSem(condition->event_sem_) == 0;
+  ok &= DosCloseEventSem(condition->ack_sem_) == 0;
+  return !ok;
+}
+
+static INLINE int pthread_cond_init(pthread_cond_t *const condition,
+                                    void *cond_attr) {
+  int ok = 1;
+  (void)cond_attr;
+
+  ok &=
+      DosCreateEventSem(NULL, &condition->event_sem_, DCE_POSTONE, FALSE) == 0;
+  ok &= DosCreateEventSem(NULL, &condition->ack_sem_, DCE_POSTONE, FALSE) == 0;
+  if (!ok) {
+    pthread_cond_destroy(condition);
+    return 1;
+  }
+  condition->wait_count_ = 0;
+  return 0;
+}
+
+static INLINE int pthread_cond_signal(pthread_cond_t *const condition) {
+  int ok = 1;
+
+  if (!__atomic_cmpxchg32(&condition->wait_count_, 0, 0)) {
+    ok &= DosPostEventSem(condition->event_sem_) == 0;
+    ok &= DosWaitEventSem(condition->ack_sem_, SEM_INDEFINITE_WAIT) == 0;
+  }
+
+  return !ok;
+}
+
+static INLINE int pthread_cond_broadcast(pthread_cond_t *const condition) {
+  int ok = 1;
+
+  while (!__atomic_cmpxchg32(&condition->wait_count_, 0, 0))
+    ok &= pthread_cond_signal(condition) == 0;
+
+  return !ok;
+}
+
+static INLINE int pthread_cond_wait(pthread_cond_t *const condition,
+                                    pthread_mutex_t *const mutex) {
+  int ok = 1;
+
+  __atomic_increment(&condition->wait_count_);
+
+  ok &= pthread_mutex_unlock(mutex) == 0;
+
+  ok &= DosWaitEventSem(condition->event_sem_, SEM_INDEFINITE_WAIT) == 0;
+
+  __atomic_decrement(&condition->wait_count_);
+
+  ok &= DosPostEventSem(condition->ack_sem_) == 0;
+
+  pthread_mutex_lock(mutex);
+
+  return !ok;
+}
+#else                 // _WIN32
+#include <pthread.h>  // NOLINT
+#define THREADFN void *
+#define THREAD_RETURN(val) val
+#endif
+
+#endif  // CONFIG_MULTITHREAD
+
+// State of the worker thread object
+typedef enum {
+  NOT_OK = 0,  // object is unusable
+  OK,          // ready to work
+  WORK         // busy finishing the current task
+} AVxWorkerStatus;
+
+// Function to be called by the worker thread. Takes two opaque pointers as
+// arguments (data1 and data2), and should return false in case of error.
+typedef int (*AVxWorkerHook)(void *, void *);
+
+// Platform-dependent implementation details for the worker.
+typedef struct AVxWorkerImpl AVxWorkerImpl;
+
+// Synchronization object used to launch job in the worker thread
+typedef struct {
+  AVxWorkerImpl *impl_;
+  AVxWorkerStatus status_;
+  AVxWorkerHook hook;  // hook to call
+  void *data1;         // first argument passed to 'hook'
+  void *data2;         // second argument passed to 'hook'
+  int had_error;       // return value of the last call to 'hook'
+} AVxWorker;
+
+// The interface for all thread-worker related functions. All these functions
+// must be implemented.
+typedef struct {
+  // Must be called first, before any other method.
+  void (*init)(AVxWorker *const worker);
+  // Must be called to initialize the object and spawn the thread. Re-entrant.
+  // Will potentially launch the thread. Returns false in case of error.
+  int (*reset)(AVxWorker *const worker);
+  // Makes sure the previous work is finished. Returns true if worker->had_error
+  // was not set and no error condition was triggered by the working thread.
+  int (*sync)(AVxWorker *const worker);
+  // Triggers the thread to call hook() with data1 and data2 arguments. These
+  // hook/data1/data2 values can be changed at any time before calling this
+  // function, but not be changed afterward until the next call to Sync().
+  void (*launch)(AVxWorker *const worker);
+  // This function is similar to launch() except that it calls the
+  // hook directly instead of using a thread. Convenient to bypass the thread
+  // mechanism while still using the AVxWorker structs. sync() must
+  // still be called afterward (for error reporting).
+  void (*execute)(AVxWorker *const worker);
+  // Kill the thread and terminate the object. To use the object again, one
+  // must call reset() again.
+  void (*end)(AVxWorker *const worker);
+} AVxWorkerInterface;
+
+// Install a new set of threading functions, overriding the defaults. This
+// should be done before any workers are started, i.e., before any encoding or
+// decoding takes place. The contents of the interface struct are copied, it
+// is safe to free the corresponding memory after this call. This function is
+// not thread-safe. Return false in case of invalid pointer or methods.
+int aom_set_worker_interface(const AVxWorkerInterface *const winterface);
+
+// Retrieve the currently set thread worker interface.
+const AVxWorkerInterface *aom_get_worker_interface(void);
+
+//------------------------------------------------------------------------------
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_THREAD_H_
diff --git a/third_party/aom/aom_util/aom_util.cmake b/third_party/aom/aom_util/aom_util.cmake
new file mode 100644
index 0000000000..484d9d9a04
--- /dev/null
+++ b/third_party/aom/aom_util/aom_util.cmake
@@ -0,0 +1,29 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+set(AOM_UTIL_SOURCES
+    "${AOM_ROOT}/aom_util/aom_thread.c"
+    "${AOM_ROOT}/aom_util/aom_thread.h"
+    "${AOM_ROOT}/aom_util/endian_inl.h")
+
+if (CONFIG_BITSTREAM_DEBUG)
+  set(AOM_UTIL_SOURCES
+      ${AOM_UTIL_SOURCES}
+      "${AOM_ROOT}/aom_util/debug_util.c"
+      "${AOM_ROOT}/aom_util/debug_util.h")
+endif ()
+
+# Creates the aom_util build target and makes libaom depend on it. The libaom
+# target must exist before this function is called.
+function (setup_aom_util_targets)
+  add_library(aom_util OBJECT ${AOM_UTIL_SOURCES})
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_util PARENT_SCOPE)
+  target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_util>)
+endfunction ()
diff --git a/third_party/aom/aom_util/aom_util.mk b/third_party/aom/aom_util/aom_util.mk
new file mode 100644
index 0000000000..14b484a153
--- /dev/null
+++ b/third_party/aom/aom_util/aom_util.mk
@@ -0,0 +1,18 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+UTIL_SRCS-yes += aom_util.mk
+UTIL_SRCS-yes += aom_thread.c
+UTIL_SRCS-yes += aom_thread.h
+UTIL_SRCS-$(CONFIG_BITSTREAM_DEBUG) += debug_util.c
+UTIL_SRCS-$(CONFIG_BITSTREAM_DEBUG) += debug_util.h
+UTIL_SRCS-yes += endian_inl.h
diff --git a/third_party/aom/aom_util/debug_util.c b/third_party/aom/aom_util/debug_util.c
new file mode 100644
index 0000000000..071d66976c
--- /dev/null
+++ b/third_party/aom/aom_util/debug_util.c
@@ -0,0 +1,98 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+#include <string.h>
+#include "aom_util/debug_util.h"
+
+#define QUEUE_MAX_SIZE 2000000
+static int result_queue[QUEUE_MAX_SIZE];
+#if CONFIG_DAALA_EC
+static int nsymbs_queue[QUEUE_MAX_SIZE];
+static aom_cdf_prob cdf_queue[QUEUE_MAX_SIZE][16];
+#else
+static int prob_queue[QUEUE_MAX_SIZE];
+#endif
+
+static int queue_r = 0;
+static int queue_w = 0;
+static int queue_prev_w = -1;
+static int skip_r = 0;
+static int skip_w = 0;
+
+static int frame_idx_w = 0;
+
+static int frame_idx_r = 0;
+
+void bitstream_queue_set_frame_write(int frame_idx) { frame_idx_w = frame_idx; }
+
+int bitstream_queue_get_frame_write(void) { return frame_idx_w; }
+
+void bitstream_queue_set_frame_read(int frame_idx) { frame_idx_r = frame_idx; }
+
+int bitstream_queue_get_frame_read(void) { return frame_idx_r; }
+
+void bitstream_queue_set_skip_write(int skip) { skip_w = skip; }
+
+void bitstream_queue_set_skip_read(int skip) { skip_r = skip; }
+
+void bitstream_queue_record_write(void) { queue_prev_w = queue_w; }
+
+void bitstream_queue_reset_write(void) { queue_w = queue_prev_w; }
+
+int bitstream_queue_get_write(void) { return queue_w; }
+
+int bitstream_queue_get_read(void) { return queue_r; }
+
+void bitstream_queue_pop(int *result,
+#if CONFIG_DAALA_EC
+                         aom_cdf_prob *cdf, int *nsymbs) {
+#else
+                         int *prob) {
+#endif  // CONFIG_DAALA_EC
+  if (!skip_r) {
+    if (queue_w == queue_r) {
+      printf("buffer underflow queue_w %d queue_r %d\n", queue_w, queue_r);
+      assert(0);
+    }
+    *result = result_queue[queue_r];
+#if CONFIG_DAALA_EC
+    *nsymbs = nsymbs_queue[queue_r];
+    memcpy(cdf, cdf_queue[queue_r], *nsymbs * sizeof(*cdf));
+#else
+    *prob = prob_queue[queue_r];
+#endif  // CONFIG_DAALA_EC
+    queue_r = (queue_r + 1) % QUEUE_MAX_SIZE;
+  }
+}
+
+void bitstream_queue_push(int result,
+#if CONFIG_DAALA_EC
+                          const aom_cdf_prob *cdf, int nsymbs) {
+#else
+                          int prob) {
+#endif  // CONFIG_DAALA_EC
+  if (!skip_w) {
+    result_queue[queue_w] = result;
+#if CONFIG_DAALA_EC
+    nsymbs_queue[queue_w] = nsymbs;
+    memcpy(cdf_queue[queue_w], cdf, nsymbs * sizeof(*cdf));
+#else
+    prob_queue[queue_w] = prob;
+#endif  // CONFIG_DAALA_EC
+    queue_w = (queue_w + 1) % QUEUE_MAX_SIZE;
+    if (queue_w == queue_r) {
+      printf("buffer overflow queue_w %d queue_r %d\n", queue_w, queue_r);
+      assert(0);
+    }
+  }
+}
diff --git a/third_party/aom/aom_util/debug_util.h b/third_party/aom/aom_util/debug_util.h
new file mode 100644
index 0000000000..2ed56ea22b
--- /dev/null
+++ b/third_party/aom/aom_util/debug_util.h
@@ -0,0 +1,54 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOM_UTIL_DEBUG_UTIL_H_
+#define AOM_UTIL_DEBUG_UTIL_H_
+
+#include "./aom_config.h"
+#if CONFIG_DAALA_EC
+#include "aom_dsp/prob.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* This is a debug tool used to detect bitstream error. On encoder side, it
+ * pushes each bit and probability into a queue before the bit is written into
+ * the Arithmetic coder. On decoder side, whenever a bit is read out from the
+ * Arithmetic coder, it pops out the reference bit and probability from the
+ * queue as well. If the two results do not match, this debug tool will report
+ * an error.  This tool can be used to pin down the bitstream error precisely.
+ * By combining gdb's backtrace method, we can detect which module causes the
+ * bitstream error. */
+int bitstream_queue_get_write(void);
+int bitstream_queue_get_read(void);
+void bitstream_queue_record_write(void);
+void bitstream_queue_reset_write(void);
+#if CONFIG_DAALA_EC
+void bitstream_queue_pop(int *result, aom_cdf_prob *cdf, int *nsymbs);
+void bitstream_queue_push(int result, const aom_cdf_prob *cdf, int nsymbs);
+#else
+void bitstream_queue_pop(int *result, int *prob);
+void bitstream_queue_push(int result, int prob);
+#endif
+void bitstream_queue_set_skip_write(int skip);
+void bitstream_queue_set_skip_read(int skip);
+void bitstream_queue_set_frame_write(int frame_idx);
+int bitstream_queue_get_frame_write(void);
+void bitstream_queue_set_frame_read(int frame_idx);
+int bitstream_queue_get_frame_read(void);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_UTIL_DEBUG_UTIL_H_
diff --git a/third_party/aom/aom_util/endian_inl.h b/third_party/aom/aom_util/endian_inl.h
new file mode 100644
index 0000000000..17a238649c
--- /dev/null
+++ b/third_party/aom/aom_util/endian_inl.h
@@ -0,0 +1,120 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+//
+// Endian related functions.
+
+#ifndef AOM_UTIL_ENDIAN_INL_H_
+#define AOM_UTIL_ENDIAN_INL_H_
+
+#include <stdlib.h>
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+#if defined(__GNUC__)
+#define LOCAL_GCC_VERSION ((__GNUC__ << 8) | __GNUC_MINOR__)
+#define LOCAL_GCC_PREREQ(maj, min) (LOCAL_GCC_VERSION >= (((maj) << 8) | (min)))
+#else
+#define LOCAL_GCC_VERSION 0
+#define LOCAL_GCC_PREREQ(maj, min) 0
+#endif
+
+// handle clang compatibility
+#ifndef __has_builtin
+#define __has_builtin(x) 0
+#endif
+
+// some endian fix (e.g.: mips-gcc doesn't define __BIG_ENDIAN__)
+#if !defined(WORDS_BIGENDIAN) &&                   \
+    (defined(__BIG_ENDIAN__) || defined(_M_PPC) || \
+     (defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)))
+#define WORDS_BIGENDIAN
+#endif
+
+#if defined(WORDS_BIGENDIAN)
+#define HToLE32 BSwap32
+#define HToLE16 BSwap16
+#define HToBE64(x) (x)
+#define HToBE32(x) (x)
+#else
+#define HToLE32(x) (x)
+#define HToLE16(x) (x)
+#define HToBE64(X) BSwap64(X)
+#define HToBE32(X) BSwap32(X)
+#endif
+
+#if LOCAL_GCC_PREREQ(4, 8) || __has_builtin(__builtin_bswap16)
+#define HAVE_BUILTIN_BSWAP16
+#endif
+
+#if LOCAL_GCC_PREREQ(4, 3) || __has_builtin(__builtin_bswap32)
+#define HAVE_BUILTIN_BSWAP32
+#endif
+
+#if LOCAL_GCC_PREREQ(4, 3) || __has_builtin(__builtin_bswap64)
+#define HAVE_BUILTIN_BSWAP64
+#endif
+
+#if HAVE_MIPS32 && defined(__mips__) && !defined(__mips64) && \
+    defined(__mips_isa_rev) && (__mips_isa_rev >= 2) && (__mips_isa_rev < 6)
+#define AOM_USE_MIPS32_R2
+#endif
+
+static INLINE uint16_t BSwap16(uint16_t x) {
+#if defined(HAVE_BUILTIN_BSWAP16)
+  return __builtin_bswap16(x);
+#elif defined(_MSC_VER)
+  return _byteswap_ushort(x);
+#else
+  // gcc will recognize a 'rorw $8, ...' here:
+  return (x >> 8) | ((x & 0xff) << 8);
+#endif  // HAVE_BUILTIN_BSWAP16
+}
+
+static INLINE uint32_t BSwap32(uint32_t x) {
+#if defined(AOM_USE_MIPS32_R2)
+  uint32_t ret;
+  __asm__ volatile(
+      "wsbh   %[ret], %[x]          \n\t"
+      "rotr   %[ret], %[ret],  16   \n\t"
+      : [ret] "=r"(ret)
+      : [x] "r"(x));
+  return ret;
+#elif defined(HAVE_BUILTIN_BSWAP32)
+  return __builtin_bswap32(x);
+#elif defined(__i386__) || defined(__x86_64__)
+  uint32_t swapped_bytes;
+  __asm__ volatile("bswap %0" : "=r"(swapped_bytes) : "0"(x));
+  return swapped_bytes;
+#elif defined(_MSC_VER)
+  return (uint32_t)_byteswap_ulong(x);
+#else
+  return (x >> 24) | ((x >> 8) & 0xff00) | ((x << 8) & 0xff0000) | (x << 24);
+#endif  // HAVE_BUILTIN_BSWAP32
+}
+
+static INLINE uint64_t BSwap64(uint64_t x) {
+#if defined(HAVE_BUILTIN_BSWAP64)
+  return __builtin_bswap64(x);
+#elif defined(__x86_64__)
+  uint64_t swapped_bytes;
+  __asm__ volatile("bswapq %0" : "=r"(swapped_bytes) : "0"(x));
+  return swapped_bytes;
+#elif defined(_MSC_VER)
+  return (uint64_t)_byteswap_uint64(x);
+#else   // generic code for swapping 64-bit values (suggested by bdb@)
+  x = ((x & 0xffffffff00000000ull) >> 32) | ((x & 0x00000000ffffffffull) << 32);
+  x = ((x & 0xffff0000ffff0000ull) >> 16) | ((x & 0x0000ffff0000ffffull) << 16);
+  x = ((x & 0xff00ff00ff00ff00ull) >> 8) | ((x & 0x00ff00ff00ff00ffull) << 8);
+  return x;
+#endif  // HAVE_BUILTIN_BSWAP64
+}
+
+#endif  // AOM_UTIL_ENDIAN_INL_H_
diff --git a/third_party/aom/aomdec.c b/third_party/aom/aomdec.c
new file mode 100644
index 0000000000..f74baddb7a
--- /dev/null
+++ b/third_party/aom/aomdec.c
@@ -0,0 +1,1070 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdarg.h>
+#include <string.h>
+#include <limits.h>
+
+#include "./aom_config.h"
+
+#if CONFIG_OS_SUPPORT
+#if HAVE_UNISTD_H
+#include <unistd.h>  // NOLINT
+#elif !defined(STDOUT_FILENO)
+#define STDOUT_FILENO 1
+#endif
+#endif
+
+#if CONFIG_LIBYUV
+#include "third_party/libyuv/include/libyuv/scale.h"
+#endif
+
+#include "./args.h"
+#include "./ivfdec.h"
+
+#include "aom/aom_decoder.h"
+#include "aom_ports/mem_ops.h"
+#include "aom_ports/aom_timer.h"
+
+#if CONFIG_AV1_DECODER
+#include "aom/aomdx.h"
+#endif
+
+#include "./md5_utils.h"
+
+#include "./tools_common.h"
+#if CONFIG_WEBM_IO
+#include "./webmdec.h"
+#endif
+#include "./y4menc.h"
+
+static const char *exec_name;
+
+struct AvxDecInputContext {
+  struct AvxInputContext *aom_input_ctx;
+  struct WebmInputContext *webm_ctx;
+};
+
+static const arg_def_t looparg =
+    ARG_DEF(NULL, "loops", 1, "Number of times to decode the file");
+static const arg_def_t codecarg = ARG_DEF(NULL, "codec", 1, "Codec to use");
+static const arg_def_t use_yv12 =
+    ARG_DEF(NULL, "yv12", 0, "Output raw YV12 frames");
+static const arg_def_t use_i420 =
+    ARG_DEF(NULL, "i420", 0, "Output raw I420 frames");
+static const arg_def_t flipuvarg =
+    ARG_DEF(NULL, "flipuv", 0, "Flip the chroma planes in the output");
+static const arg_def_t rawvideo =
+    ARG_DEF(NULL, "rawvideo", 0, "Output raw YUV frames");
+static const arg_def_t noblitarg =
+    ARG_DEF(NULL, "noblit", 0, "Don't process the decoded frames");
+static const arg_def_t progressarg =
+    ARG_DEF(NULL, "progress", 0, "Show progress after each frame decodes");
+static const arg_def_t limitarg =
+    ARG_DEF(NULL, "limit", 1, "Stop decoding after n frames");
+static const arg_def_t skiparg =
+    ARG_DEF(NULL, "skip", 1, "Skip the first n input frames");
+static const arg_def_t postprocarg =
+    ARG_DEF(NULL, "postproc", 0, "Postprocess decoded frames");
+static const arg_def_t summaryarg =
+    ARG_DEF(NULL, "summary", 0, "Show timing summary");
+static const arg_def_t outputfile =
+    ARG_DEF("o", "output", 1, "Output file name pattern (see below)");
+static const arg_def_t threadsarg =
+    ARG_DEF("t", "threads", 1, "Max threads to use");
+static const arg_def_t frameparallelarg =
+    ARG_DEF(NULL, "frame-parallel", 0, "Frame parallel decode");
+static const arg_def_t verbosearg =
+    ARG_DEF("v", "verbose", 0, "Show version string");
+static const arg_def_t error_concealment =
+    ARG_DEF(NULL, "error-concealment", 0, "Enable decoder error-concealment");
+static const arg_def_t scalearg =
+    ARG_DEF("S", "scale", 0, "Scale output frames uniformly");
+static const arg_def_t continuearg =
+    ARG_DEF("k", "keep-going", 0, "(debug) Continue decoding after error");
+static const arg_def_t fb_arg =
+    ARG_DEF(NULL, "frame-buffers", 1, "Number of frame buffers to use");
+static const arg_def_t md5arg =
+    ARG_DEF(NULL, "md5", 0, "Compute the MD5 sum of the decoded frame");
+static const arg_def_t framestatsarg =
+    ARG_DEF(NULL, "framestats", 1, "Output per-frame stats (.csv format)");
+#if CONFIG_HIGHBITDEPTH
+static const arg_def_t outbitdeptharg =
+    ARG_DEF(NULL, "output-bit-depth", 1, "Output bit-depth for decoded frames");
+#endif
+#if CONFIG_EXT_TILE
+static const arg_def_t tiler = ARG_DEF(NULL, "tile-row", 1,
+                                       "Row index of tile to decode "
+                                       "(-1 for all rows)");
+static const arg_def_t tilec = ARG_DEF(NULL, "tile-column", 1,
+                                       "Column index of tile to decode "
+                                       "(-1 for all columns)");
+#endif  // CONFIG_EXT_TILE
+
+static const arg_def_t *all_args[] = { &codecarg,
+                                       &use_yv12,
+                                       &use_i420,
+                                       &flipuvarg,
+                                       &rawvideo,
+                                       &noblitarg,
+                                       &progressarg,
+                                       &limitarg,
+                                       &skiparg,
+                                       &postprocarg,
+                                       &summaryarg,
+                                       &outputfile,
+                                       &threadsarg,
+                                       &frameparallelarg,
+                                       &verbosearg,
+                                       &scalearg,
+                                       &fb_arg,
+                                       &md5arg,
+                                       &framestatsarg,
+                                       &error_concealment,
+                                       &continuearg,
+#if CONFIG_HIGHBITDEPTH
+                                       &outbitdeptharg,
+#endif
+#if CONFIG_EXT_TILE
+                                       &tiler,
+                                       &tilec,
+#endif  // CONFIG_EXT_TILE
+                                       NULL };
+
+#if CONFIG_LIBYUV
+static INLINE int libyuv_scale(aom_image_t *src, aom_image_t *dst,
+                               FilterModeEnum mode) {
+#if CONFIG_HIGHBITDEPTH
+  if (src->fmt == AOM_IMG_FMT_I42016) {
+    assert(dst->fmt == AOM_IMG_FMT_I42016);
+    return I420Scale_16(
+        (uint16_t *)src->planes[AOM_PLANE_Y], src->stride[AOM_PLANE_Y] / 2,
+        (uint16_t *)src->planes[AOM_PLANE_U], src->stride[AOM_PLANE_U] / 2,
+        (uint16_t *)src->planes[AOM_PLANE_V], src->stride[AOM_PLANE_V] / 2,
+        src->d_w, src->d_h, (uint16_t *)dst->planes[AOM_PLANE_Y],
+        dst->stride[AOM_PLANE_Y] / 2, (uint16_t *)dst->planes[AOM_PLANE_U],
+        dst->stride[AOM_PLANE_U] / 2, (uint16_t *)dst->planes[AOM_PLANE_V],
+        dst->stride[AOM_PLANE_V] / 2, dst->d_w, dst->d_h, mode);
+  }
+#endif
+  assert(src->fmt == AOM_IMG_FMT_I420);
+  assert(dst->fmt == AOM_IMG_FMT_I420);
+  return I420Scale(src->planes[AOM_PLANE_Y], src->stride[AOM_PLANE_Y],
+                   src->planes[AOM_PLANE_U], src->stride[AOM_PLANE_U],
+                   src->planes[AOM_PLANE_V], src->stride[AOM_PLANE_V], src->d_w,
+                   src->d_h, dst->planes[AOM_PLANE_Y], dst->stride[AOM_PLANE_Y],
+                   dst->planes[AOM_PLANE_U], dst->stride[AOM_PLANE_U],
+                   dst->planes[AOM_PLANE_V], dst->stride[AOM_PLANE_V], dst->d_w,
+                   dst->d_h, mode);
+}
+#endif
+
+void usage_exit(void) {
+  int i;
+
+  fprintf(stderr,
+          "Usage: %s <options> filename\n\n"
+          "Options:\n",
+          exec_name);
+  arg_show_usage(stderr, all_args);
+  fprintf(stderr,
+          "\nOutput File Patterns:\n\n"
+          "  The -o argument specifies the name of the file(s) to "
+          "write to. If the\n  argument does not include any escape "
+          "characters, the output will be\n  written to a single file. "
+          "Otherwise, the filename will be calculated by\n  expanding "
+          "the following escape characters:\n");
+  fprintf(stderr,
+          "\n\t%%w   - Frame width"
+          "\n\t%%h   - Frame height"
+          "\n\t%%<n> - Frame number, zero padded to <n> places (1..9)"
+          "\n\n  Pattern arguments are only supported in conjunction "
+          "with the --yv12 and\n  --i420 options. If the -o option is "
+          "not specified, the output will be\n  directed to stdout.\n");
+  fprintf(stderr, "\nIncluded decoders:\n\n");
+
+  for (i = 0; i < get_aom_decoder_count(); ++i) {
+    const AvxInterface *const decoder = get_aom_decoder_by_index(i);
+    fprintf(stderr, "    %-6s - %s\n", decoder->name,
+            aom_codec_iface_name(decoder->codec_interface()));
+  }
+
+  exit(EXIT_FAILURE);
+}
+
+static int raw_read_frame(FILE *infile, uint8_t **buffer, size_t *bytes_read,
+                          size_t *buffer_size) {
+  char raw_hdr[RAW_FRAME_HDR_SZ];
+  size_t frame_size = 0;
+
+  if (fread(raw_hdr, RAW_FRAME_HDR_SZ, 1, infile) != 1) {
+    if (!feof(infile)) warn("Failed to read RAW frame size\n");
+  } else {
+    const size_t kCorruptFrameThreshold = 256 * 1024 * 1024;
+    const size_t kFrameTooSmallThreshold = 256 * 1024;
+    frame_size = mem_get_le32(raw_hdr);
+
+    if (frame_size > kCorruptFrameThreshold) {
+      warn("Read invalid frame size (%u)\n", (unsigned int)frame_size);
+      frame_size = 0;
+    }
+
+    if (frame_size < kFrameTooSmallThreshold) {
+      warn("Warning: Read invalid frame size (%u) - not a raw file?\n",
+           (unsigned int)frame_size);
+    }
+
+    if (frame_size > *buffer_size) {
+      uint8_t *new_buf = realloc(*buffer, 2 * frame_size);
+      if (new_buf) {
+        *buffer = new_buf;
+        *buffer_size = 2 * frame_size;
+      } else {
+        warn("Failed to allocate compressed data buffer\n");
+        frame_size = 0;
+      }
+    }
+  }
+
+  if (!feof(infile)) {
+    if (fread(*buffer, 1, frame_size, infile) != frame_size) {
+      warn("Failed to read full frame\n");
+      return 1;
+    }
+    *bytes_read = frame_size;
+  }
+
+  return 0;
+}
+
+static int read_frame(struct AvxDecInputContext *input, uint8_t **buf,
+                      size_t *bytes_in_buffer, size_t *buffer_size) {
+  switch (input->aom_input_ctx->file_type) {
+#if CONFIG_WEBM_IO
+    case FILE_TYPE_WEBM:
+      return webm_read_frame(input->webm_ctx, buf, bytes_in_buffer);
+#endif
+    case FILE_TYPE_RAW:
+      return raw_read_frame(input->aom_input_ctx->file, buf, bytes_in_buffer,
+                            buffer_size);
+    case FILE_TYPE_IVF:
+      return ivf_read_frame(input->aom_input_ctx->file, buf, bytes_in_buffer,
+                            buffer_size);
+    default: return 1;
+  }
+}
+
+static void update_image_md5(const aom_image_t *img, const int planes[3],
+                             MD5Context *md5) {
+  int i, y;
+
+  for (i = 0; i < 3; ++i) {
+    const int plane = planes[i];
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int w = aom_img_plane_width(img, plane) *
+                  ((img->fmt & AOM_IMG_FMT_HIGHBITDEPTH) ? 2 : 1);
+    const int h = aom_img_plane_height(img, plane);
+
+    for (y = 0; y < h; ++y) {
+      MD5Update(md5, buf, w);
+      buf += stride;
+    }
+  }
+}
+
+static void write_image_file(const aom_image_t *img, const int planes[3],
+                             FILE *file) {
+  int i, y;
+#if CONFIG_HIGHBITDEPTH
+  const int bytes_per_sample = ((img->fmt & AOM_IMG_FMT_HIGHBITDEPTH) ? 2 : 1);
+#else
+  const int bytes_per_sample = 1;
+#endif
+
+  for (i = 0; i < 3; ++i) {
+    const int plane = planes[i];
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int w = aom_img_plane_width(img, plane);
+    const int h = aom_img_plane_height(img, plane);
+
+    for (y = 0; y < h; ++y) {
+      fwrite(buf, bytes_per_sample, w, file);
+      buf += stride;
+    }
+  }
+}
+
+static int file_is_raw(struct AvxInputContext *input) {
+  uint8_t buf[32];
+  int is_raw = 0;
+  aom_codec_stream_info_t si;
+
+  si.sz = sizeof(si);
+
+  if (fread(buf, 1, 32, input->file) == 32) {
+    int i;
+
+    if (mem_get_le32(buf) < 256 * 1024 * 1024) {
+      for (i = 0; i < get_aom_decoder_count(); ++i) {
+        const AvxInterface *const decoder = get_aom_decoder_by_index(i);
+        if (!aom_codec_peek_stream_info(decoder->codec_interface(), buf + 4,
+                                        32 - 4, &si)) {
+          is_raw = 1;
+          input->fourcc = decoder->fourcc;
+          input->width = si.w;
+          input->height = si.h;
+          input->framerate.numerator = 30;
+          input->framerate.denominator = 1;
+          break;
+        }
+      }
+    }
+  }
+
+  rewind(input->file);
+  return is_raw;
+}
+
+static void show_progress(int frame_in, int frame_out, uint64_t dx_time) {
+  fprintf(stderr,
+          "%d decoded frames/%d showed frames in %" PRId64 " us (%.2f fps)\r",
+          frame_in, frame_out, dx_time,
+          (double)frame_out * 1000000.0 / (double)dx_time);
+}
+
+struct ExternalFrameBuffer {
+  uint8_t *data;
+  size_t size;
+  int in_use;
+};
+
+struct ExternalFrameBufferList {
+  int num_external_frame_buffers;
+  struct ExternalFrameBuffer *ext_fb;
+};
+
+// Callback used by libaom to request an external frame buffer. |cb_priv|
+// Application private data passed into the set function. |min_size| is the
+// minimum size in bytes needed to decode the next frame. |fb| pointer to the
+// frame buffer.
+static int get_av1_frame_buffer(void *cb_priv, size_t min_size,
+                                aom_codec_frame_buffer_t *fb) {
+  int i;
+  struct ExternalFrameBufferList *const ext_fb_list =
+      (struct ExternalFrameBufferList *)cb_priv;
+  if (ext_fb_list == NULL) return -1;
+
+  // Find a free frame buffer.
+  for (i = 0; i < ext_fb_list->num_external_frame_buffers; ++i) {
+    if (!ext_fb_list->ext_fb[i].in_use) break;
+  }
+
+  if (i == ext_fb_list->num_external_frame_buffers) return -1;
+
+  if (ext_fb_list->ext_fb[i].size < min_size) {
+    free(ext_fb_list->ext_fb[i].data);
+    ext_fb_list->ext_fb[i].data = (uint8_t *)calloc(min_size, sizeof(uint8_t));
+    if (!ext_fb_list->ext_fb[i].data) return -1;
+
+    ext_fb_list->ext_fb[i].size = min_size;
+  }
+
+  fb->data = ext_fb_list->ext_fb[i].data;
+  fb->size = ext_fb_list->ext_fb[i].size;
+  ext_fb_list->ext_fb[i].in_use = 1;
+
+  // Set the frame buffer's private data to point at the external frame buffer.
+  fb->priv = &ext_fb_list->ext_fb[i];
+  return 0;
+}
+
+// Callback used by libaom when there are no references to the frame buffer.
+// |cb_priv| user private data passed into the set function. |fb| pointer
+// to the frame buffer.
+static int release_av1_frame_buffer(void *cb_priv,
+                                    aom_codec_frame_buffer_t *fb) {
+  struct ExternalFrameBuffer *const ext_fb =
+      (struct ExternalFrameBuffer *)fb->priv;
+  (void)cb_priv;
+  ext_fb->in_use = 0;
+  return 0;
+}
+
+static void generate_filename(const char *pattern, char *out, size_t q_len,
+                              unsigned int d_w, unsigned int d_h,
+                              unsigned int frame_in) {
+  const char *p = pattern;
+  char *q = out;
+
+  do {
+    char *next_pat = strchr(p, '%');
+
+    if (p == next_pat) {
+      size_t pat_len;
+
+      /* parse the pattern */
+      q[q_len - 1] = '\0';
+      switch (p[1]) {
+        case 'w': snprintf(q, q_len - 1, "%d", d_w); break;
+        case 'h': snprintf(q, q_len - 1, "%d", d_h); break;
+        case '1': snprintf(q, q_len - 1, "%d", frame_in); break;
+        case '2': snprintf(q, q_len - 1, "%02d", frame_in); break;
+        case '3': snprintf(q, q_len - 1, "%03d", frame_in); break;
+        case '4': snprintf(q, q_len - 1, "%04d", frame_in); break;
+        case '5': snprintf(q, q_len - 1, "%05d", frame_in); break;
+        case '6': snprintf(q, q_len - 1, "%06d", frame_in); break;
+        case '7': snprintf(q, q_len - 1, "%07d", frame_in); break;
+        case '8': snprintf(q, q_len - 1, "%08d", frame_in); break;
+        case '9': snprintf(q, q_len - 1, "%09d", frame_in); break;
+        default: die("Unrecognized pattern %%%c\n", p[1]); break;
+      }
+
+      pat_len = strlen(q);
+      if (pat_len >= q_len - 1) die("Output filename too long.\n");
+      q += pat_len;
+      p += 2;
+      q_len -= pat_len;
+    } else {
+      size_t copy_len;
+
+      /* copy the next segment */
+      if (!next_pat)
+        copy_len = strlen(p);
+      else
+        copy_len = next_pat - p;
+
+      if (copy_len >= q_len - 1) die("Output filename too long.\n");
+
+      memcpy(q, p, copy_len);
+      q[copy_len] = '\0';
+      q += copy_len;
+      p += copy_len;
+      q_len -= copy_len;
+    }
+  } while (*p);
+}
+
+static int is_single_file(const char *outfile_pattern) {
+  const char *p = outfile_pattern;
+
+  do {
+    p = strchr(p, '%');
+    if (p && p[1] >= '1' && p[1] <= '9')
+      return 0;  // pattern contains sequence number, so it's not unique
+    if (p) p++;
+  } while (p);
+
+  return 1;
+}
+
+static void print_md5(unsigned char digest[16], const char *filename) {
+  int i;
+
+  for (i = 0; i < 16; ++i) printf("%02x", digest[i]);
+  printf("  %s\n", filename);
+}
+
+static FILE *open_outfile(const char *name) {
+  if (strcmp("-", name) == 0) {
+    set_binary_mode(stdout);
+    return stdout;
+  } else {
+    FILE *file = fopen(name, "wb");
+    if (!file) fatal("Failed to open output file '%s'", name);
+    return file;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static int img_shifted_realloc_required(const aom_image_t *img,
+                                        const aom_image_t *shifted,
+                                        aom_img_fmt_t required_fmt) {
+  return img->d_w != shifted->d_w || img->d_h != shifted->d_h ||
+         required_fmt != shifted->fmt;
+}
+#endif
+
+static int main_loop(int argc, const char **argv_) {
+  aom_codec_ctx_t decoder;
+  char *fn = NULL;
+  int i;
+  int ret = EXIT_FAILURE;
+  uint8_t *buf = NULL;
+  size_t bytes_in_buffer = 0, buffer_size = 0;
+  FILE *infile;
+  int frame_in = 0, frame_out = 0, flipuv = 0, noblit = 0;
+  int do_md5 = 0, progress = 0, frame_parallel = 0;
+  int stop_after = 0, postproc = 0, summary = 0, quiet = 1;
+  int arg_skip = 0;
+  int ec_enabled = 0;
+  int keep_going = 0;
+  const AvxInterface *interface = NULL;
+  const AvxInterface *fourcc_interface = NULL;
+  uint64_t dx_time = 0;
+  struct arg arg;
+  char **argv, **argi, **argj;
+
+  int single_file;
+  int use_y4m = 1;
+  int opt_yv12 = 0;
+  int opt_i420 = 0;
+  aom_codec_dec_cfg_t cfg = { 0, 0, 0 };
+#if CONFIG_HIGHBITDEPTH
+  unsigned int output_bit_depth = 0;
+#endif
+#if CONFIG_EXT_TILE
+  int tile_row = -1;
+  int tile_col = -1;
+#endif  // CONFIG_EXT_TILE
+  int frames_corrupted = 0;
+  int dec_flags = 0;
+  int do_scale = 0;
+  aom_image_t *scaled_img = NULL;
+#if CONFIG_HIGHBITDEPTH
+  aom_image_t *img_shifted = NULL;
+#endif
+  int frame_avail, got_data, flush_decoder = 0;
+  int num_external_frame_buffers = 0;
+  struct ExternalFrameBufferList ext_fb_list = { 0, NULL };
+
+  const char *outfile_pattern = NULL;
+  char outfile_name[PATH_MAX] = { 0 };
+  FILE *outfile = NULL;
+
+  FILE *framestats_file = NULL;
+
+  MD5Context md5_ctx;
+  unsigned char md5_digest[16];
+
+  struct AvxDecInputContext input = { NULL, NULL };
+  struct AvxInputContext aom_input_ctx;
+#if CONFIG_WEBM_IO
+  struct WebmInputContext webm_ctx;
+  memset(&(webm_ctx), 0, sizeof(webm_ctx));
+  input.webm_ctx = &webm_ctx;
+#endif
+  input.aom_input_ctx = &aom_input_ctx;
+
+  /* Parse command line */
+  exec_name = argv_[0];
+  argv = argv_dup(argc - 1, argv_ + 1);
+
+  for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
+    memset(&arg, 0, sizeof(arg));
+    arg.argv_step = 1;
+
+    if (arg_match(&arg, &codecarg, argi)) {
+      interface = get_aom_decoder_by_name(arg.val);
+      if (!interface)
+        die("Error: Unrecognized argument (%s) to --codec\n", arg.val);
+    } else if (arg_match(&arg, &looparg, argi)) {
+      // no-op
+    } else if (arg_match(&arg, &outputfile, argi)) {
+      outfile_pattern = arg.val;
+    } else if (arg_match(&arg, &use_yv12, argi)) {
+      use_y4m = 0;
+      flipuv = 1;
+      opt_yv12 = 1;
+    } else if (arg_match(&arg, &use_i420, argi)) {
+      use_y4m = 0;
+      flipuv = 0;
+      opt_i420 = 1;
+    } else if (arg_match(&arg, &rawvideo, argi)) {
+      use_y4m = 0;
+    } else if (arg_match(&arg, &flipuvarg, argi)) {
+      flipuv = 1;
+    } else if (arg_match(&arg, &noblitarg, argi)) {
+      noblit = 1;
+    } else if (arg_match(&arg, &progressarg, argi)) {
+      progress = 1;
+    } else if (arg_match(&arg, &limitarg, argi)) {
+      stop_after = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &skiparg, argi)) {
+      arg_skip = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &postprocarg, argi)) {
+      postproc = 1;
+    } else if (arg_match(&arg, &md5arg, argi)) {
+      do_md5 = 1;
+    } else if (arg_match(&arg, &framestatsarg, argi)) {
+      framestats_file = fopen(arg.val, "w");
+      if (!framestats_file) {
+        die("Error: Could not open --framestats file (%s) for writing.\n",
+            arg.val);
+      }
+    } else if (arg_match(&arg, &summaryarg, argi)) {
+      summary = 1;
+    } else if (arg_match(&arg, &threadsarg, argi)) {
+      cfg.threads = arg_parse_uint(&arg);
+    }
+#if CONFIG_AV1_DECODER
+    else if (arg_match(&arg, &frameparallelarg, argi))
+      frame_parallel = 1;
+#endif
+    else if (arg_match(&arg, &verbosearg, argi))
+      quiet = 0;
+    else if (arg_match(&arg, &scalearg, argi))
+      do_scale = 1;
+    else if (arg_match(&arg, &fb_arg, argi))
+      num_external_frame_buffers = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &continuearg, argi))
+      keep_going = 1;
+#if CONFIG_HIGHBITDEPTH
+    else if (arg_match(&arg, &outbitdeptharg, argi)) {
+      output_bit_depth = arg_parse_uint(&arg);
+    }
+#endif
+#if CONFIG_EXT_TILE
+    else if (arg_match(&arg, &tiler, argi))
+      tile_row = arg_parse_int(&arg);
+    else if (arg_match(&arg, &tilec, argi))
+      tile_col = arg_parse_int(&arg);
+#endif  // CONFIG_EXT_TILE
+    else
+      argj++;
+  }
+
+  /* Check for unrecognized options */
+  for (argi = argv; *argi; argi++)
+    if (argi[0][0] == '-' && strlen(argi[0]) > 1)
+      die("Error: Unrecognized option %s\n", *argi);
+
+  /* Handle non-option arguments */
+  fn = argv[0];
+
+  if (!fn) {
+    free(argv);
+    usage_exit();
+  }
+  /* Open file */
+  infile = strcmp(fn, "-") ? fopen(fn, "rb") : set_binary_mode(stdin);
+
+  if (!infile) {
+    fatal("Failed to open input file '%s'", strcmp(fn, "-") ? fn : "stdin");
+  }
+#if CONFIG_OS_SUPPORT
+  /* Make sure we don't dump to the terminal, unless forced to with -o - */
+  if (!outfile_pattern && isatty(STDOUT_FILENO) && !do_md5 && !noblit) {
+    fprintf(stderr,
+            "Not dumping raw video to your terminal. Use '-o -' to "
+            "override.\n");
+    return EXIT_FAILURE;
+  }
+#endif
+  input.aom_input_ctx->file = infile;
+  if (file_is_ivf(input.aom_input_ctx))
+    input.aom_input_ctx->file_type = FILE_TYPE_IVF;
+#if CONFIG_WEBM_IO
+  else if (file_is_webm(input.webm_ctx, input.aom_input_ctx))
+    input.aom_input_ctx->file_type = FILE_TYPE_WEBM;
+#endif
+  else if (file_is_raw(input.aom_input_ctx))
+    input.aom_input_ctx->file_type = FILE_TYPE_RAW;
+  else {
+    fprintf(stderr, "Unrecognized input file type.\n");
+#if !CONFIG_WEBM_IO
+    fprintf(stderr, "aomdec was built without WebM container support.\n");
+#endif
+    return EXIT_FAILURE;
+  }
+
+  outfile_pattern = outfile_pattern ? outfile_pattern : "-";
+  single_file = is_single_file(outfile_pattern);
+
+  if (!noblit && single_file) {
+    generate_filename(outfile_pattern, outfile_name, PATH_MAX,
+                      aom_input_ctx.width, aom_input_ctx.height, 0);
+    if (do_md5)
+      MD5Init(&md5_ctx);
+    else
+      outfile = open_outfile(outfile_name);
+  }
+
+  if (use_y4m && !noblit) {
+    if (!single_file) {
+      fprintf(stderr,
+              "YUV4MPEG2 not supported with output patterns,"
+              " try --i420 or --yv12 or --rawvideo.\n");
+      return EXIT_FAILURE;
+    }
+
+#if CONFIG_WEBM_IO
+    if (aom_input_ctx.file_type == FILE_TYPE_WEBM) {
+      if (webm_guess_framerate(input.webm_ctx, input.aom_input_ctx)) {
+        fprintf(stderr,
+                "Failed to guess framerate -- error parsing "
+                "webm file?\n");
+        return EXIT_FAILURE;
+      }
+    }
+#endif
+  }
+
+  fourcc_interface = get_aom_decoder_by_fourcc(aom_input_ctx.fourcc);
+  if (interface && fourcc_interface && interface != fourcc_interface)
+    warn("Header indicates codec: %s\n", fourcc_interface->name);
+  else
+    interface = fourcc_interface;
+
+  if (!interface) interface = get_aom_decoder_by_index(0);
+
+  dec_flags = (postproc ? AOM_CODEC_USE_POSTPROC : 0) |
+              (ec_enabled ? AOM_CODEC_USE_ERROR_CONCEALMENT : 0) |
+              (frame_parallel ? AOM_CODEC_USE_FRAME_THREADING : 0);
+  if (aom_codec_dec_init(&decoder, interface->codec_interface(), &cfg,
+                         dec_flags)) {
+    fprintf(stderr, "Failed to initialize decoder: %s\n",
+            aom_codec_error(&decoder));
+    goto fail2;
+  }
+
+  if (!quiet) fprintf(stderr, "%s\n", decoder.name);
+
+#if CONFIG_AV1_DECODER && CONFIG_EXT_TILE
+  if (aom_codec_control(&decoder, AV1_SET_DECODE_TILE_ROW, tile_row)) {
+    fprintf(stderr, "Failed to set decode_tile_row: %s\n",
+            aom_codec_error(&decoder));
+    goto fail;
+  }
+
+  if (aom_codec_control(&decoder, AV1_SET_DECODE_TILE_COL, tile_col)) {
+    fprintf(stderr, "Failed to set decode_tile_col: %s\n",
+            aom_codec_error(&decoder));
+    goto fail;
+  }
+#endif
+
+  if (arg_skip) fprintf(stderr, "Skipping first %d frames.\n", arg_skip);
+  while (arg_skip) {
+    if (read_frame(&input, &buf, &bytes_in_buffer, &buffer_size)) break;
+    arg_skip--;
+  }
+
+  if (num_external_frame_buffers > 0) {
+    ext_fb_list.num_external_frame_buffers = num_external_frame_buffers;
+    ext_fb_list.ext_fb = (struct ExternalFrameBuffer *)calloc(
+        num_external_frame_buffers, sizeof(*ext_fb_list.ext_fb));
+    if (aom_codec_set_frame_buffer_functions(&decoder, get_av1_frame_buffer,
+                                             release_av1_frame_buffer,
+                                             &ext_fb_list)) {
+      fprintf(stderr, "Failed to configure external frame buffers: %s\n",
+              aom_codec_error(&decoder));
+      goto fail;
+    }
+  }
+
+  frame_avail = 1;
+  got_data = 0;
+
+  if (framestats_file) fprintf(framestats_file, "bytes,qp\r\n");
+
+  /* Decode file */
+  while (frame_avail || got_data) {
+    aom_codec_iter_t iter = NULL;
+    aom_image_t *img;
+    struct aom_usec_timer timer;
+    int corrupted = 0;
+
+    frame_avail = 0;
+    if (!stop_after || frame_in < stop_after) {
+      if (!read_frame(&input, &buf, &bytes_in_buffer, &buffer_size)) {
+        frame_avail = 1;
+        frame_in++;
+
+        aom_usec_timer_start(&timer);
+
+        if (aom_codec_decode(&decoder, buf, (unsigned int)bytes_in_buffer, NULL,
+                             0)) {
+          const char *detail = aom_codec_error_detail(&decoder);
+          warn("Failed to decode frame %d: %s", frame_in,
+               aom_codec_error(&decoder));
+
+          if (detail) warn("Additional information: %s", detail);
+          if (!keep_going) goto fail;
+        }
+
+        if (framestats_file) {
+          int qp;
+          if (aom_codec_control(&decoder, AOMD_GET_LAST_QUANTIZER, &qp)) {
+            warn("Failed AOMD_GET_LAST_QUANTIZER: %s",
+                 aom_codec_error(&decoder));
+            if (!keep_going) goto fail;
+          }
+          fprintf(framestats_file, "%d,%d\r\n", (int)bytes_in_buffer, qp);
+        }
+
+        aom_usec_timer_mark(&timer);
+        dx_time += aom_usec_timer_elapsed(&timer);
+      } else {
+        flush_decoder = 1;
+      }
+    } else {
+      flush_decoder = 1;
+    }
+
+    aom_usec_timer_start(&timer);
+
+    if (flush_decoder) {
+      // Flush the decoder in frame parallel decode.
+      if (aom_codec_decode(&decoder, NULL, 0, NULL, 0)) {
+        warn("Failed to flush decoder: %s", aom_codec_error(&decoder));
+      }
+    }
+
+    got_data = 0;
+    if ((img = aom_codec_get_frame(&decoder, &iter))) {
+      ++frame_out;
+      got_data = 1;
+    }
+
+    aom_usec_timer_mark(&timer);
+    dx_time += (unsigned int)aom_usec_timer_elapsed(&timer);
+
+    if (!frame_parallel &&
+        aom_codec_control(&decoder, AOMD_GET_FRAME_CORRUPTED, &corrupted)) {
+      warn("Failed AOM_GET_FRAME_CORRUPTED: %s", aom_codec_error(&decoder));
+      if (!keep_going) goto fail;
+    }
+    frames_corrupted += corrupted;
+
+    if (progress) show_progress(frame_in, frame_out, dx_time);
+
+    if (!noblit && img) {
+      const int PLANES_YUV[] = { AOM_PLANE_Y, AOM_PLANE_U, AOM_PLANE_V };
+      const int PLANES_YVU[] = { AOM_PLANE_Y, AOM_PLANE_V, AOM_PLANE_U };
+      const int *planes = flipuv ? PLANES_YVU : PLANES_YUV;
+
+      if (do_scale) {
+        if (frame_out == 1) {
+          // If the output frames are to be scaled to a fixed display size then
+          // use the width and height specified in the container. If either of
+          // these is set to 0, use the display size set in the first frame
+          // header. If that is unavailable, use the raw decoded size of the
+          // first decoded frame.
+          int render_width = aom_input_ctx.width;
+          int render_height = aom_input_ctx.height;
+          if (!render_width || !render_height) {
+            int render_size[2];
+            if (aom_codec_control(&decoder, AV1D_GET_DISPLAY_SIZE,
+                                  render_size)) {
+              // As last resort use size of first frame as display size.
+              render_width = img->d_w;
+              render_height = img->d_h;
+            } else {
+              render_width = render_size[0];
+              render_height = render_size[1];
+            }
+          }
+          scaled_img =
+              aom_img_alloc(NULL, img->fmt, render_width, render_height, 16);
+          scaled_img->bit_depth = img->bit_depth;
+        }
+
+        if (img->d_w != scaled_img->d_w || img->d_h != scaled_img->d_h) {
+#if CONFIG_LIBYUV
+          libyuv_scale(img, scaled_img, kFilterBox);
+          img = scaled_img;
+#else
+          fprintf(stderr,
+                  "Failed  to scale output frame: %s.\n"
+                  "Scaling is disabled in this configuration. "
+                  "To enable scaling, configure with --enable-libyuv\n",
+                  aom_codec_error(&decoder));
+          goto fail;
+#endif
+        }
+      }
+#if CONFIG_HIGHBITDEPTH
+      // Default to codec bit depth if output bit depth not set
+      if (!output_bit_depth && single_file && !do_md5) {
+        output_bit_depth = img->bit_depth;
+      }
+      // Shift up or down if necessary
+      if (output_bit_depth != 0 && output_bit_depth != img->bit_depth) {
+        const aom_img_fmt_t shifted_fmt =
+            output_bit_depth == 8
+                ? img->fmt ^ (img->fmt & AOM_IMG_FMT_HIGHBITDEPTH)
+                : img->fmt | AOM_IMG_FMT_HIGHBITDEPTH;
+        if (img_shifted &&
+            img_shifted_realloc_required(img, img_shifted, shifted_fmt)) {
+          aom_img_free(img_shifted);
+          img_shifted = NULL;
+        }
+        if (!img_shifted) {
+          img_shifted =
+              aom_img_alloc(NULL, shifted_fmt, img->d_w, img->d_h, 16);
+          img_shifted->bit_depth = output_bit_depth;
+        }
+        if (output_bit_depth > img->bit_depth) {
+          aom_img_upshift(img_shifted, img, output_bit_depth - img->bit_depth);
+        } else {
+          aom_img_downshift(img_shifted, img,
+                            img->bit_depth - output_bit_depth);
+        }
+        img = img_shifted;
+      }
+#endif
+
+#if CONFIG_EXT_TILE
+      aom_input_ctx.width = img->d_w;
+      aom_input_ctx.height = img->d_h;
+#endif  // CONFIG_EXT_TILE
+
+      if (single_file) {
+        if (use_y4m) {
+          char y4m_buf[Y4M_BUFFER_SIZE] = { 0 };
+          size_t len = 0;
+          if (img->fmt == AOM_IMG_FMT_I440 || img->fmt == AOM_IMG_FMT_I44016) {
+            fprintf(stderr, "Cannot produce y4m output for 440 sampling.\n");
+            goto fail;
+          }
+          if (frame_out == 1) {
+            // Y4M file header
+            len = y4m_write_file_header(
+                y4m_buf, sizeof(y4m_buf), aom_input_ctx.width,
+                aom_input_ctx.height, &aom_input_ctx.framerate, img->fmt,
+                img->bit_depth);
+            if (do_md5) {
+              MD5Update(&md5_ctx, (md5byte *)y4m_buf, (unsigned int)len);
+            } else {
+              fputs(y4m_buf, outfile);
+            }
+          }
+
+          // Y4M frame header
+          len = y4m_write_frame_header(y4m_buf, sizeof(y4m_buf));
+          if (do_md5) {
+            MD5Update(&md5_ctx, (md5byte *)y4m_buf, (unsigned int)len);
+          } else {
+            fputs(y4m_buf, outfile);
+          }
+        } else {
+          if (frame_out == 1) {
+            // Check if --yv12 or --i420 options are consistent with the
+            // bit-stream decoded
+            if (opt_i420) {
+              if (img->fmt != AOM_IMG_FMT_I420 &&
+                  img->fmt != AOM_IMG_FMT_I42016) {
+                fprintf(stderr, "Cannot produce i420 output for bit-stream.\n");
+                goto fail;
+              }
+            }
+            if (opt_yv12) {
+              if ((img->fmt != AOM_IMG_FMT_I420 &&
+                   img->fmt != AOM_IMG_FMT_YV12) ||
+                  img->bit_depth != 8) {
+                fprintf(stderr, "Cannot produce yv12 output for bit-stream.\n");
+                goto fail;
+              }
+            }
+          }
+        }
+
+        if (do_md5) {
+          update_image_md5(img, planes, &md5_ctx);
+        } else {
+          write_image_file(img, planes, outfile);
+        }
+      } else {
+        generate_filename(outfile_pattern, outfile_name, PATH_MAX, img->d_w,
+                          img->d_h, frame_in);
+        if (do_md5) {
+          MD5Init(&md5_ctx);
+          update_image_md5(img, planes, &md5_ctx);
+          MD5Final(md5_digest, &md5_ctx);
+          print_md5(md5_digest, outfile_name);
+        } else {
+          outfile = open_outfile(outfile_name);
+          write_image_file(img, planes, outfile);
+          fclose(outfile);
+        }
+      }
+    }
+  }
+
+  if (summary || progress) {
+    show_progress(frame_in, frame_out, dx_time);
+    fprintf(stderr, "\n");
+  }
+
+  if (frames_corrupted) {
+    fprintf(stderr, "WARNING: %d frames corrupted.\n", frames_corrupted);
+  } else {
+    ret = EXIT_SUCCESS;
+  }
+
+fail:
+
+  if (aom_codec_destroy(&decoder)) {
+    fprintf(stderr, "Failed to destroy decoder: %s\n",
+            aom_codec_error(&decoder));
+  }
+
+fail2:
+
+  if (!noblit && single_file) {
+    if (do_md5) {
+      MD5Final(md5_digest, &md5_ctx);
+      print_md5(md5_digest, outfile_name);
+    } else {
+      fclose(outfile);
+    }
+  }
+
+#if CONFIG_WEBM_IO
+  if (input.aom_input_ctx->file_type == FILE_TYPE_WEBM)
+    webm_free(input.webm_ctx);
+#endif
+
+  if (input.aom_input_ctx->file_type != FILE_TYPE_WEBM) free(buf);
+
+  if (scaled_img) aom_img_free(scaled_img);
+#if CONFIG_HIGHBITDEPTH
+  if (img_shifted) aom_img_free(img_shifted);
+#endif
+
+  for (i = 0; i < ext_fb_list.num_external_frame_buffers; ++i) {
+    free(ext_fb_list.ext_fb[i].data);
+  }
+  free(ext_fb_list.ext_fb);
+
+  fclose(infile);
+  if (framestats_file) fclose(framestats_file);
+
+  free(argv);
+
+  return ret;
+}
+
+int main(int argc, const char **argv_) {
+  unsigned int loops = 1, i;
+  char **argv, **argi, **argj;
+  struct arg arg;
+  int error = 0;
+
+  argv = argv_dup(argc - 1, argv_ + 1);
+  for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
+    memset(&arg, 0, sizeof(arg));
+    arg.argv_step = 1;
+
+    if (arg_match(&arg, &looparg, argi)) {
+      loops = arg_parse_uint(&arg);
+      break;
+    }
+  }
+  free(argv);
+  for (i = 0; !error && i < loops; i++) error = main_loop(argc, argv_);
+  return error;
+}
diff --git a/third_party/aom/aomenc.c b/third_party/aom/aomenc.c
new file mode 100644
index 0000000000..f4bf888ce5
--- /dev/null
+++ b/third_party/aom/aomenc.c
@@ -0,0 +1,2144 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aomenc.h"
+#include "./aom_config.h"
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#if CONFIG_LIBYUV
+#include "third_party/libyuv/include/libyuv/scale.h"
+#endif
+
+#include "aom/aom_encoder.h"
+#if CONFIG_DECODERS
+#include "aom/aom_decoder.h"
+#endif
+
+#include "./args.h"
+#include "./ivfenc.h"
+#include "./tools_common.h"
+#include "examples/encoder_util.h"
+
+#if CONFIG_AV1_ENCODER
+#include "aom/aomcx.h"
+#endif
+#if CONFIG_AV1_DECODER
+#include "aom/aomdx.h"
+#endif
+
+#include "./aomstats.h"
+#include "./rate_hist.h"
+#include "./warnings.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/aom_timer.h"
+#include "aom_ports/mem_ops.h"
+#if CONFIG_WEBM_IO
+#include "./webmenc.h"
+#endif
+#include "./y4minput.h"
+
+/* Swallow warnings about unused results of fread/fwrite */
+static size_t wrap_fread(void *ptr, size_t size, size_t nmemb, FILE *stream) {
+  return fread(ptr, size, nmemb, stream);
+}
+#define fread wrap_fread
+
+static size_t wrap_fwrite(const void *ptr, size_t size, size_t nmemb,
+                          FILE *stream) {
+  return fwrite(ptr, size, nmemb, stream);
+}
+#define fwrite wrap_fwrite
+
+static const char *exec_name;
+
+static void warn_or_exit_on_errorv(aom_codec_ctx_t *ctx, int fatal,
+                                   const char *s, va_list ap) {
+  if (ctx->err) {
+    const char *detail = aom_codec_error_detail(ctx);
+
+    vfprintf(stderr, s, ap);
+    fprintf(stderr, ": %s\n", aom_codec_error(ctx));
+
+    if (detail) fprintf(stderr, "    %s\n", detail);
+
+    if (fatal) exit(EXIT_FAILURE);
+  }
+}
+
+static void ctx_exit_on_error(aom_codec_ctx_t *ctx, const char *s, ...) {
+  va_list ap;
+
+  va_start(ap, s);
+  warn_or_exit_on_errorv(ctx, 1, s, ap);
+  va_end(ap);
+}
+
+static void warn_or_exit_on_error(aom_codec_ctx_t *ctx, int fatal,
+                                  const char *s, ...) {
+  va_list ap;
+
+  va_start(ap, s);
+  warn_or_exit_on_errorv(ctx, fatal, s, ap);
+  va_end(ap);
+}
+
+static int read_frame(struct AvxInputContext *input_ctx, aom_image_t *img) {
+  FILE *f = input_ctx->file;
+  y4m_input *y4m = &input_ctx->y4m;
+  int shortread = 0;
+
+  if (input_ctx->file_type == FILE_TYPE_Y4M) {
+    if (y4m_input_fetch_frame(y4m, f, img) < 1) return 0;
+  } else {
+    shortread = read_yuv_frame(input_ctx, img);
+  }
+
+  return !shortread;
+}
+
+static int file_is_y4m(const char detect[4]) {
+  if (memcmp(detect, "YUV4", 4) == 0) {
+    return 1;
+  }
+  return 0;
+}
+
+static int fourcc_is_ivf(const char detect[4]) {
+  if (memcmp(detect, "DKIF", 4) == 0) {
+    return 1;
+  }
+  return 0;
+}
+
+static const arg_def_t debugmode =
+    ARG_DEF("D", "debug", 0, "Debug mode (makes output deterministic)");
+static const arg_def_t outputfile =
+    ARG_DEF("o", "output", 1, "Output filename");
+static const arg_def_t use_yv12 =
+    ARG_DEF(NULL, "yv12", 0, "Input file is YV12 ");
+static const arg_def_t use_i420 =
+    ARG_DEF(NULL, "i420", 0, "Input file is I420 (default)");
+static const arg_def_t use_i422 =
+    ARG_DEF(NULL, "i422", 0, "Input file is I422");
+static const arg_def_t use_i444 =
+    ARG_DEF(NULL, "i444", 0, "Input file is I444");
+static const arg_def_t use_i440 =
+    ARG_DEF(NULL, "i440", 0, "Input file is I440");
+static const arg_def_t codecarg = ARG_DEF(NULL, "codec", 1, "Codec to use");
+static const arg_def_t passes =
+    ARG_DEF("p", "passes", 1, "Number of passes (1/2)");
+static const arg_def_t pass_arg =
+    ARG_DEF(NULL, "pass", 1, "Pass to execute (1/2)");
+static const arg_def_t fpf_name =
+    ARG_DEF(NULL, "fpf", 1, "First pass statistics file name");
+#if CONFIG_FP_MB_STATS
+static const arg_def_t fpmbf_name =
+    ARG_DEF(NULL, "fpmbf", 1, "First pass block statistics file name");
+#endif
+static const arg_def_t limit =
+    ARG_DEF(NULL, "limit", 1, "Stop encoding after n input frames");
+static const arg_def_t skip =
+    ARG_DEF(NULL, "skip", 1, "Skip the first n input frames");
+static const arg_def_t deadline =
+    ARG_DEF("d", "deadline", 1, "Deadline per frame (usec)");
+static const arg_def_t good_dl =
+    ARG_DEF(NULL, "good", 0, "Use Good Quality Deadline");
+static const arg_def_t quietarg =
+    ARG_DEF("q", "quiet", 0, "Do not print encode progress");
+static const arg_def_t verbosearg =
+    ARG_DEF("v", "verbose", 0, "Show encoder parameters");
+static const arg_def_t psnrarg =
+    ARG_DEF(NULL, "psnr", 0, "Show PSNR in status line");
+
+static const struct arg_enum_list test_decode_enum[] = {
+  { "off", TEST_DECODE_OFF },
+  { "fatal", TEST_DECODE_FATAL },
+  { "warn", TEST_DECODE_WARN },
+  { NULL, 0 }
+};
+static const arg_def_t recontest = ARG_DEF_ENUM(
+    NULL, "test-decode", 1, "Test encode/decode mismatch", test_decode_enum);
+static const arg_def_t framerate =
+    ARG_DEF(NULL, "fps", 1, "Stream frame rate (rate/scale)");
+static const arg_def_t use_webm =
+    ARG_DEF(NULL, "webm", 0, "Output WebM (default when WebM IO is enabled)");
+static const arg_def_t use_ivf = ARG_DEF(NULL, "ivf", 0, "Output IVF");
+static const arg_def_t out_part =
+    ARG_DEF("P", "output-partitions", 0,
+            "Makes encoder output partitions. Requires IVF output!");
+static const arg_def_t q_hist_n =
+    ARG_DEF(NULL, "q-hist", 1, "Show quantizer histogram (n-buckets)");
+static const arg_def_t rate_hist_n =
+    ARG_DEF(NULL, "rate-hist", 1, "Show rate histogram (n-buckets)");
+static const arg_def_t disable_warnings =
+    ARG_DEF(NULL, "disable-warnings", 0,
+            "Disable warnings about potentially incorrect encode settings.");
+static const arg_def_t disable_warning_prompt =
+    ARG_DEF("y", "disable-warning-prompt", 0,
+            "Display warnings, but do not prompt user to continue.");
+
+#if CONFIG_HIGHBITDEPTH
+static const arg_def_t test16bitinternalarg = ARG_DEF(
+    NULL, "test-16bit-internal", 0, "Force use of 16 bit internal buffer");
+
+static const struct arg_enum_list bitdepth_enum[] = {
+  { "8", AOM_BITS_8 }, { "10", AOM_BITS_10 }, { "12", AOM_BITS_12 }, { NULL, 0 }
+};
+
+static const arg_def_t bitdeptharg = ARG_DEF_ENUM(
+    "b", "bit-depth", 1,
+    "Bit depth for codec (8 for version <=1, 10 or 12 for version 2)",
+    bitdepth_enum);
+static const arg_def_t inbitdeptharg =
+    ARG_DEF(NULL, "input-bit-depth", 1, "Bit depth of input");
+#endif
+
+static const arg_def_t *main_args[] = { &debugmode,
+                                        &outputfile,
+                                        &codecarg,
+                                        &passes,
+                                        &pass_arg,
+                                        &fpf_name,
+                                        &limit,
+                                        &skip,
+                                        &deadline,
+                                        &good_dl,
+                                        &quietarg,
+                                        &verbosearg,
+                                        &psnrarg,
+                                        &use_webm,
+                                        &use_ivf,
+                                        &out_part,
+                                        &q_hist_n,
+                                        &rate_hist_n,
+                                        &disable_warnings,
+                                        &disable_warning_prompt,
+                                        &recontest,
+                                        NULL };
+
+static const arg_def_t usage =
+    ARG_DEF("u", "usage", 1, "Usage profile number to use");
+static const arg_def_t threads =
+    ARG_DEF("t", "threads", 1, "Max number of threads to use");
+static const arg_def_t profile =
+    ARG_DEF(NULL, "profile", 1, "Bitstream profile number to use");
+static const arg_def_t width = ARG_DEF("w", "width", 1, "Frame width");
+static const arg_def_t height = ARG_DEF("h", "height", 1, "Frame height");
+#if CONFIG_WEBM_IO
+static const struct arg_enum_list stereo_mode_enum[] = {
+  { "mono", STEREO_FORMAT_MONO },
+  { "left-right", STEREO_FORMAT_LEFT_RIGHT },
+  { "bottom-top", STEREO_FORMAT_BOTTOM_TOP },
+  { "top-bottom", STEREO_FORMAT_TOP_BOTTOM },
+  { "right-left", STEREO_FORMAT_RIGHT_LEFT },
+  { NULL, 0 }
+};
+static const arg_def_t stereo_mode = ARG_DEF_ENUM(
+    NULL, "stereo-mode", 1, "Stereo 3D video format", stereo_mode_enum);
+#endif
+static const arg_def_t timebase = ARG_DEF(
+    NULL, "timebase", 1, "Output timestamp precision (fractional seconds)");
+static const arg_def_t error_resilient =
+    ARG_DEF(NULL, "error-resilient", 1, "Enable error resiliency features");
+static const arg_def_t lag_in_frames =
+    ARG_DEF(NULL, "lag-in-frames", 1, "Max number of frames to lag");
+
+static const arg_def_t *global_args[] = { &use_yv12,
+                                          &use_i420,
+                                          &use_i422,
+                                          &use_i444,
+                                          &use_i440,
+                                          &usage,
+                                          &threads,
+                                          &profile,
+                                          &width,
+                                          &height,
+#if CONFIG_WEBM_IO
+                                          &stereo_mode,
+#endif
+                                          &timebase,
+                                          &framerate,
+                                          &error_resilient,
+#if CONFIG_HIGHBITDEPTH
+                                          &test16bitinternalarg,
+                                          &bitdeptharg,
+#endif
+                                          &lag_in_frames,
+                                          NULL };
+
+static const arg_def_t dropframe_thresh =
+    ARG_DEF(NULL, "drop-frame", 1, "Temporal resampling threshold (buf %)");
+static const arg_def_t resize_allowed =
+    ARG_DEF(NULL, "resize-allowed", 1, "Spatial resampling enabled (bool)");
+static const arg_def_t resize_width =
+    ARG_DEF(NULL, "resize-width", 1, "Width of encoded frame");
+static const arg_def_t resize_height =
+    ARG_DEF(NULL, "resize-height", 1, "Height of encoded frame");
+static const arg_def_t resize_up_thresh =
+    ARG_DEF(NULL, "resize-up", 1, "Upscale threshold (buf %)");
+static const arg_def_t resize_down_thresh =
+    ARG_DEF(NULL, "resize-down", 1, "Downscale threshold (buf %)");
+static const struct arg_enum_list end_usage_enum[] = { { "vbr", AOM_VBR },
+                                                       { "cbr", AOM_CBR },
+                                                       { "cq", AOM_CQ },
+                                                       { "q", AOM_Q },
+                                                       { NULL, 0 } };
+static const arg_def_t end_usage =
+    ARG_DEF_ENUM(NULL, "end-usage", 1, "Rate control mode", end_usage_enum);
+static const arg_def_t target_bitrate =
+    ARG_DEF(NULL, "target-bitrate", 1, "Bitrate (kbps)");
+static const arg_def_t min_quantizer =
+    ARG_DEF(NULL, "min-q", 1, "Minimum (best) quantizer");
+static const arg_def_t max_quantizer =
+    ARG_DEF(NULL, "max-q", 1, "Maximum (worst) quantizer");
+static const arg_def_t undershoot_pct =
+    ARG_DEF(NULL, "undershoot-pct", 1, "Datarate undershoot (min) target (%)");
+static const arg_def_t overshoot_pct =
+    ARG_DEF(NULL, "overshoot-pct", 1, "Datarate overshoot (max) target (%)");
+static const arg_def_t buf_sz =
+    ARG_DEF(NULL, "buf-sz", 1, "Client buffer size (ms)");
+static const arg_def_t buf_initial_sz =
+    ARG_DEF(NULL, "buf-initial-sz", 1, "Client initial buffer size (ms)");
+static const arg_def_t buf_optimal_sz =
+    ARG_DEF(NULL, "buf-optimal-sz", 1, "Client optimal buffer size (ms)");
+static const arg_def_t *rc_args[] = {
+  &dropframe_thresh, &resize_allowed,     &resize_width,   &resize_height,
+  &resize_up_thresh, &resize_down_thresh, &end_usage,      &target_bitrate,
+  &min_quantizer,    &max_quantizer,      &undershoot_pct, &overshoot_pct,
+  &buf_sz,           &buf_initial_sz,     &buf_optimal_sz, NULL
+};
+
+static const arg_def_t bias_pct =
+    ARG_DEF(NULL, "bias-pct", 1, "CBR/VBR bias (0=CBR, 100=VBR)");
+static const arg_def_t minsection_pct =
+    ARG_DEF(NULL, "minsection-pct", 1, "GOP min bitrate (% of target)");
+static const arg_def_t maxsection_pct =
+    ARG_DEF(NULL, "maxsection-pct", 1, "GOP max bitrate (% of target)");
+static const arg_def_t *rc_twopass_args[] = { &bias_pct, &minsection_pct,
+                                              &maxsection_pct, NULL };
+
+static const arg_def_t kf_min_dist =
+    ARG_DEF(NULL, "kf-min-dist", 1, "Minimum keyframe interval (frames)");
+static const arg_def_t kf_max_dist =
+    ARG_DEF(NULL, "kf-max-dist", 1, "Maximum keyframe interval (frames)");
+static const arg_def_t kf_disabled =
+    ARG_DEF(NULL, "disable-kf", 0, "Disable keyframe placement");
+static const arg_def_t *kf_args[] = { &kf_min_dist, &kf_max_dist, &kf_disabled,
+                                      NULL };
+
+static const arg_def_t noise_sens =
+    ARG_DEF(NULL, "noise-sensitivity", 1, "Noise sensitivity (frames to blur)");
+static const arg_def_t sharpness =
+    ARG_DEF(NULL, "sharpness", 1, "Loop filter sharpness (0..7)");
+static const arg_def_t static_thresh =
+    ARG_DEF(NULL, "static-thresh", 1, "Motion detection threshold");
+static const arg_def_t auto_altref =
+    ARG_DEF(NULL, "auto-alt-ref", 1, "Enable automatic alt reference frames");
+static const arg_def_t arnr_maxframes =
+    ARG_DEF(NULL, "arnr-maxframes", 1, "AltRef max frames (0..15)");
+static const arg_def_t arnr_strength =
+    ARG_DEF(NULL, "arnr-strength", 1, "AltRef filter strength (0..6)");
+static const struct arg_enum_list tuning_enum[] = {
+  { "psnr", AOM_TUNE_PSNR }, { "ssim", AOM_TUNE_SSIM }, { NULL, 0 }
+};
+static const arg_def_t tune_ssim =
+    ARG_DEF_ENUM(NULL, "tune", 1, "Material to favor", tuning_enum);
+static const arg_def_t cq_level =
+    ARG_DEF(NULL, "cq-level", 1, "Constant/Constrained Quality level");
+static const arg_def_t max_intra_rate_pct =
+    ARG_DEF(NULL, "max-intra-rate", 1, "Max I-frame bitrate (pct)");
+
+#if CONFIG_AV1_ENCODER
+static const arg_def_t cpu_used_av1 =
+    ARG_DEF(NULL, "cpu-used", 1, "CPU Used (0..8)");
+static const arg_def_t tile_cols =
+    ARG_DEF(NULL, "tile-columns", 1, "Number of tile columns to use, log2");
+static const arg_def_t tile_rows =
+    ARG_DEF(NULL, "tile-rows", 1,
+            "Number of tile rows to use, log2 (set to 0 while threads > 1)");
+#if CONFIG_EXT_TILE
+static const arg_def_t tile_encoding_mode =
+    ARG_DEF(NULL, "tile-encoding-mode", 1,
+            "Tile encoding mode (0: normal"
+            " (default), 1: vr)");
+#endif
+#if CONFIG_DEPENDENT_HORZTILES
+static const arg_def_t tile_dependent_rows =
+    ARG_DEF(NULL, "tile-dependent-rows", 1, "Enable dependent Tile rows");
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+static const arg_def_t tile_loopfilter = ARG_DEF(
+    NULL, "tile-loopfilter", 1, "Enable loop filter across tile boundary");
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+static const arg_def_t lossless =
+    ARG_DEF(NULL, "lossless", 1, "Lossless mode (0: false (default), 1: true)");
+#if CONFIG_AOM_QM
+static const arg_def_t enable_qm =
+    ARG_DEF(NULL, "enable-qm", 1,
+            "Enable quantisation matrices (0: false (default), 1: true)");
+static const arg_def_t qm_min = ARG_DEF(
+    NULL, "qm-min", 1, "Min quant matrix flatness (0..15), default is 8");
+static const arg_def_t qm_max = ARG_DEF(
+    NULL, "qm-max", 1, "Max quant matrix flatness (0..15), default is 16");
+#endif
+#if CONFIG_TILE_GROUPS
+static const arg_def_t num_tg = ARG_DEF(
+    NULL, "num-tile-groups", 1, "Maximum number of tile groups, default is 1");
+static const arg_def_t mtu_size =
+    ARG_DEF(NULL, "mtu-size", 1,
+            "MTU size for a tile group, default is 0 (no MTU targeting), "
+            "overrides maximum number of tile groups");
+#endif
+#if CONFIG_TEMPMV_SIGNALING
+static const arg_def_t disable_tempmv = ARG_DEF(
+    NULL, "disable-tempmv", 1, "Disable temporal mv prediction (default is 0)");
+#endif
+static const arg_def_t frame_parallel_decoding =
+    ARG_DEF(NULL, "frame-parallel", 1,
+            "Enable frame parallel decodability features "
+            "(0: false (default), 1: true)");
+#if CONFIG_DELTA_Q && !CONFIG_EXT_DELTA_Q
+static const arg_def_t aq_mode = ARG_DEF(
+    NULL, "aq-mode", 1,
+    "Adaptive quantization mode (0: off (default), 1: variance 2: complexity, "
+    "3: cyclic refresh, 4: delta quant)");
+#else
+static const arg_def_t aq_mode = ARG_DEF(
+    NULL, "aq-mode", 1,
+    "Adaptive quantization mode (0: off (default), 1: variance 2: complexity, "
+    "3: cyclic refresh)");
+#endif
+#if CONFIG_EXT_DELTA_Q
+static const arg_def_t deltaq_mode = ARG_DEF(
+    NULL, "deltaq-mode", 1,
+    "Delta qindex mode (0: off (default), 1: deltaq 2: deltaq + deltalf)");
+#endif
+static const arg_def_t frame_periodic_boost =
+    ARG_DEF(NULL, "frame-boost", 1,
+            "Enable frame periodic boost (0: off (default), 1: on)");
+static const arg_def_t gf_cbr_boost_pct = ARG_DEF(
+    NULL, "gf-cbr-boost", 1, "Boost for Golden Frame in CBR mode (pct)");
+static const arg_def_t max_inter_rate_pct =
+    ARG_DEF(NULL, "max-inter-rate", 1, "Max P-frame bitrate (pct)");
+static const arg_def_t min_gf_interval = ARG_DEF(
+    NULL, "min-gf-interval", 1,
+    "min gf/arf frame interval (default 0, indicating in-built behavior)");
+static const arg_def_t max_gf_interval = ARG_DEF(
+    NULL, "max-gf-interval", 1,
+    "max gf/arf frame interval (default 0, indicating in-built behavior)");
+
+static const struct arg_enum_list color_space_enum[] = {
+  { "unknown", AOM_CS_UNKNOWN },
+  { "bt601", AOM_CS_BT_601 },
+  { "bt709", AOM_CS_BT_709 },
+  { "smpte170", AOM_CS_SMPTE_170 },
+  { "smpte240", AOM_CS_SMPTE_240 },
+  { "bt2020", AOM_CS_BT_2020 },
+  { "reserved", AOM_CS_RESERVED },
+  { "sRGB", AOM_CS_SRGB },
+  { NULL, 0 }
+};
+
+static const arg_def_t input_color_space =
+    ARG_DEF_ENUM(NULL, "color-space", 1, "The color space of input content:",
+                 color_space_enum);
+
+static const struct arg_enum_list tune_content_enum[] = {
+  { "default", AOM_CONTENT_DEFAULT },
+  { "screen", AOM_CONTENT_SCREEN },
+  { NULL, 0 }
+};
+
+static const arg_def_t tune_content = ARG_DEF_ENUM(
+    NULL, "tune-content", 1, "Tune content type", tune_content_enum);
+#endif
+
+#if CONFIG_AV1_ENCODER
+#if CONFIG_EXT_PARTITION
+static const struct arg_enum_list superblock_size_enum[] = {
+  { "dynamic", AOM_SUPERBLOCK_SIZE_DYNAMIC },
+  { "64", AOM_SUPERBLOCK_SIZE_64X64 },
+  { "128", AOM_SUPERBLOCK_SIZE_128X128 },
+  { NULL, 0 }
+};
+static const arg_def_t superblock_size = ARG_DEF_ENUM(
+    NULL, "sb-size", 1, "Superblock size to use", superblock_size_enum);
+#endif  // CONFIG_EXT_PARTITION
+
+static const arg_def_t *av1_args[] = { &cpu_used_av1,
+                                       &auto_altref,
+                                       &sharpness,
+                                       &static_thresh,
+                                       &tile_cols,
+                                       &tile_rows,
+#if CONFIG_EXT_TILE
+                                       &tile_encoding_mode,
+#endif
+#if CONFIG_DEPENDENT_HORZTILES
+                                       &tile_dependent_rows,
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+                                       &tile_loopfilter,
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+                                       &arnr_maxframes,
+                                       &arnr_strength,
+                                       &tune_ssim,
+                                       &cq_level,
+                                       &max_intra_rate_pct,
+                                       &max_inter_rate_pct,
+                                       &gf_cbr_boost_pct,
+                                       &lossless,
+#if CONFIG_AOM_QM
+                                       &enable_qm,
+                                       &qm_min,
+                                       &qm_max,
+#endif
+                                       &frame_parallel_decoding,
+                                       &aq_mode,
+#if CONFIG_EXT_DELTA_Q
+                                       &deltaq_mode,
+#endif
+                                       &frame_periodic_boost,
+                                       &noise_sens,
+                                       &tune_content,
+                                       &input_color_space,
+                                       &min_gf_interval,
+                                       &max_gf_interval,
+#if CONFIG_EXT_PARTITION
+                                       &superblock_size,
+#endif  // CONFIG_EXT_PARTITION
+#if CONFIG_TILE_GROUPS
+                                       &num_tg,
+                                       &mtu_size,
+#endif
+#if CONFIG_TEMPMV_SIGNALING
+                                       &disable_tempmv,
+#endif
+#if CONFIG_HIGHBITDEPTH
+                                       &bitdeptharg,
+                                       &inbitdeptharg,
+#endif  // CONFIG_HIGHBITDEPTH
+                                       NULL };
+static const int av1_arg_ctrl_map[] = { AOME_SET_CPUUSED,
+                                        AOME_SET_ENABLEAUTOALTREF,
+                                        AOME_SET_SHARPNESS,
+                                        AOME_SET_STATIC_THRESHOLD,
+                                        AV1E_SET_TILE_COLUMNS,
+                                        AV1E_SET_TILE_ROWS,
+#if CONFIG_EXT_TILE
+                                        AV1E_SET_TILE_ENCODING_MODE,
+#endif
+#if CONFIG_DEPENDENT_HORZTILES
+                                        AV1E_SET_TILE_DEPENDENT_ROWS,
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+                                        AV1E_SET_TILE_LOOPFILTER,
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+                                        AOME_SET_ARNR_MAXFRAMES,
+                                        AOME_SET_ARNR_STRENGTH,
+                                        AOME_SET_TUNING,
+                                        AOME_SET_CQ_LEVEL,
+                                        AOME_SET_MAX_INTRA_BITRATE_PCT,
+                                        AV1E_SET_MAX_INTER_BITRATE_PCT,
+                                        AV1E_SET_GF_CBR_BOOST_PCT,
+                                        AV1E_SET_LOSSLESS,
+#if CONFIG_AOM_QM
+                                        AV1E_SET_ENABLE_QM,
+                                        AV1E_SET_QM_MIN,
+                                        AV1E_SET_QM_MAX,
+#endif
+                                        AV1E_SET_FRAME_PARALLEL_DECODING,
+                                        AV1E_SET_AQ_MODE,
+#if CONFIG_EXT_DELTA_Q
+                                        AV1E_SET_DELTAQ_MODE,
+#endif
+                                        AV1E_SET_FRAME_PERIODIC_BOOST,
+                                        AV1E_SET_NOISE_SENSITIVITY,
+                                        AV1E_SET_TUNE_CONTENT,
+                                        AV1E_SET_COLOR_SPACE,
+                                        AV1E_SET_MIN_GF_INTERVAL,
+                                        AV1E_SET_MAX_GF_INTERVAL,
+#if CONFIG_EXT_PARTITION
+                                        AV1E_SET_SUPERBLOCK_SIZE,
+#endif  // CONFIG_EXT_PARTITION
+#if CONFIG_TILE_GROUPS
+                                        AV1E_SET_NUM_TG,
+                                        AV1E_SET_MTU,
+#endif
+#if CONFIG_TEMPMV_SIGNALING
+                                        AV1E_SET_DISABLE_TEMPMV,
+#endif
+                                        0 };
+#endif
+
+static const arg_def_t *no_args[] = { NULL };
+
+void usage_exit(void) {
+  int i;
+  const int num_encoder = get_aom_encoder_count();
+
+  fprintf(stderr, "Usage: %s <options> -o dst_filename src_filename \n",
+          exec_name);
+
+  fprintf(stderr, "\nOptions:\n");
+  arg_show_usage(stderr, main_args);
+  fprintf(stderr, "\nEncoder Global Options:\n");
+  arg_show_usage(stderr, global_args);
+  fprintf(stderr, "\nRate Control Options:\n");
+  arg_show_usage(stderr, rc_args);
+  fprintf(stderr, "\nTwopass Rate Control Options:\n");
+  arg_show_usage(stderr, rc_twopass_args);
+  fprintf(stderr, "\nKeyframe Placement Options:\n");
+  arg_show_usage(stderr, kf_args);
+#if CONFIG_AV1_ENCODER
+  fprintf(stderr, "\nAV1 Specific Options:\n");
+  arg_show_usage(stderr, av1_args);
+#endif
+  fprintf(stderr,
+          "\nStream timebase (--timebase):\n"
+          "  The desired precision of timestamps in the output, expressed\n"
+          "  in fractional seconds. Default is 1/1000.\n");
+  fprintf(stderr, "\nIncluded encoders:\n\n");
+
+  for (i = 0; i < num_encoder; ++i) {
+    const AvxInterface *const encoder = get_aom_encoder_by_index(i);
+    const char *defstr = (i == (num_encoder - 1)) ? "(default)" : "";
+    fprintf(stderr, "    %-6s - %s %s\n", encoder->name,
+            aom_codec_iface_name(encoder->codec_interface()), defstr);
+  }
+  fprintf(stderr, "\n        ");
+  fprintf(stderr, "Use --codec to switch to a non-default encoder.\n\n");
+
+  exit(EXIT_FAILURE);
+}
+
+#define NELEMENTS(x) (sizeof(x) / sizeof(x[0]))
+#if CONFIG_AV1_ENCODER
+#define ARG_CTRL_CNT_MAX NELEMENTS(av1_arg_ctrl_map)
+#endif
+
+#if !CONFIG_WEBM_IO
+typedef int stereo_format_t;
+struct WebmOutputContext {
+  int debug;
+};
+#endif
+
+/* Per-stream configuration */
+struct stream_config {
+  struct aom_codec_enc_cfg cfg;
+  const char *out_fn;
+  const char *stats_fn;
+#if CONFIG_FP_MB_STATS
+  const char *fpmb_stats_fn;
+#endif
+  stereo_format_t stereo_fmt;
+  int arg_ctrls[ARG_CTRL_CNT_MAX][2];
+  int arg_ctrl_cnt;
+  int write_webm;
+#if CONFIG_HIGHBITDEPTH
+  // whether to use 16bit internal buffers
+  int use_16bit_internal;
+#endif
+};
+
+struct stream_state {
+  int index;
+  struct stream_state *next;
+  struct stream_config config;
+  FILE *file;
+  struct rate_hist *rate_hist;
+  struct WebmOutputContext webm_ctx;
+  uint64_t psnr_sse_total;
+  uint64_t psnr_samples_total;
+  double psnr_totals[4];
+  int psnr_count;
+  int counts[64];
+  aom_codec_ctx_t encoder;
+  unsigned int frames_out;
+  uint64_t cx_time;
+  size_t nbytes;
+  stats_io_t stats;
+#if CONFIG_FP_MB_STATS
+  stats_io_t fpmb_stats;
+#endif
+  struct aom_image *img;
+  aom_codec_ctx_t decoder;
+  int mismatch_seen;
+};
+
+static void validate_positive_rational(const char *msg,
+                                       struct aom_rational *rat) {
+  if (rat->den < 0) {
+    rat->num *= -1;
+    rat->den *= -1;
+  }
+
+  if (rat->num < 0) die("Error: %s must be positive\n", msg);
+
+  if (!rat->den) die("Error: %s has zero denominator\n", msg);
+}
+
+static void parse_global_config(struct AvxEncoderConfig *global, char **argv) {
+  char **argi, **argj;
+  struct arg arg;
+  const int num_encoder = get_aom_encoder_count();
+
+  if (num_encoder < 1) die("Error: no valid encoder available\n");
+
+  /* Initialize default parameters */
+  memset(global, 0, sizeof(*global));
+  global->codec = get_aom_encoder_by_index(num_encoder - 1);
+  global->passes = 0;
+  global->color_type = I420;
+  /* Assign default deadline to good quality */
+  global->deadline = AOM_DL_GOOD_QUALITY;
+
+  for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
+    arg.argv_step = 1;
+
+    if (arg_match(&arg, &codecarg, argi)) {
+      global->codec = get_aom_encoder_by_name(arg.val);
+      if (!global->codec)
+        die("Error: Unrecognized argument (%s) to --codec\n", arg.val);
+    } else if (arg_match(&arg, &passes, argi)) {
+      global->passes = arg_parse_uint(&arg);
+
+      if (global->passes < 1 || global->passes > 2)
+        die("Error: Invalid number of passes (%d)\n", global->passes);
+    } else if (arg_match(&arg, &pass_arg, argi)) {
+      global->pass = arg_parse_uint(&arg);
+
+      if (global->pass < 1 || global->pass > 2)
+        die("Error: Invalid pass selected (%d)\n", global->pass);
+    } else if (arg_match(&arg, &usage, argi))
+      global->usage = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &deadline, argi))
+      global->deadline = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &good_dl, argi))
+      global->deadline = AOM_DL_GOOD_QUALITY;
+    else if (arg_match(&arg, &use_yv12, argi))
+      global->color_type = YV12;
+    else if (arg_match(&arg, &use_i420, argi))
+      global->color_type = I420;
+    else if (arg_match(&arg, &use_i422, argi))
+      global->color_type = I422;
+    else if (arg_match(&arg, &use_i444, argi))
+      global->color_type = I444;
+    else if (arg_match(&arg, &use_i440, argi))
+      global->color_type = I440;
+    else if (arg_match(&arg, &quietarg, argi))
+      global->quiet = 1;
+    else if (arg_match(&arg, &verbosearg, argi))
+      global->verbose = 1;
+    else if (arg_match(&arg, &limit, argi))
+      global->limit = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &skip, argi))
+      global->skip_frames = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &psnrarg, argi))
+      global->show_psnr = 1;
+    else if (arg_match(&arg, &recontest, argi))
+      global->test_decode = arg_parse_enum_or_int(&arg);
+    else if (arg_match(&arg, &framerate, argi)) {
+      global->framerate = arg_parse_rational(&arg);
+      validate_positive_rational(arg.name, &global->framerate);
+      global->have_framerate = 1;
+    } else if (arg_match(&arg, &out_part, argi))
+      global->out_part = 1;
+    else if (arg_match(&arg, &debugmode, argi))
+      global->debug = 1;
+    else if (arg_match(&arg, &q_hist_n, argi))
+      global->show_q_hist_buckets = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &rate_hist_n, argi))
+      global->show_rate_hist_buckets = arg_parse_uint(&arg);
+    else if (arg_match(&arg, &disable_warnings, argi))
+      global->disable_warnings = 1;
+    else if (arg_match(&arg, &disable_warning_prompt, argi))
+      global->disable_warning_prompt = 1;
+    else
+      argj++;
+  }
+
+  if (global->pass) {
+    /* DWIM: Assume the user meant passes=2 if pass=2 is specified */
+    if (global->pass > global->passes) {
+      warn("Assuming --pass=%d implies --passes=%d\n", global->pass,
+           global->pass);
+      global->passes = global->pass;
+    }
+  }
+  /* Validate global config */
+  if (global->passes == 0) {
+#if CONFIG_AV1_ENCODER
+    // Make default AV1 passes = 2 until there is a better quality 1-pass
+    // encoder
+    if (global->codec != NULL && global->codec->name != NULL)
+      global->passes = (strcmp(global->codec->name, "av1") == 0) ? 2 : 1;
+#else
+    global->passes = 1;
+#endif
+  }
+}
+
+static void open_input_file(struct AvxInputContext *input) {
+  /* Parse certain options from the input file, if possible */
+  input->file = strcmp(input->filename, "-") ? fopen(input->filename, "rb")
+                                             : set_binary_mode(stdin);
+
+  if (!input->file) fatal("Failed to open input file");
+
+  if (!fseeko(input->file, 0, SEEK_END)) {
+    /* Input file is seekable. Figure out how long it is, so we can get
+     * progress info.
+     */
+    input->length = ftello(input->file);
+    rewind(input->file);
+  }
+
+  /* Default to 1:1 pixel aspect ratio. */
+  input->pixel_aspect_ratio.numerator = 1;
+  input->pixel_aspect_ratio.denominator = 1;
+
+  /* For RAW input sources, these bytes will applied on the first frame
+   *  in read_frame().
+   */
+  input->detect.buf_read = fread(input->detect.buf, 1, 4, input->file);
+  input->detect.position = 0;
+
+  if (input->detect.buf_read == 4 && file_is_y4m(input->detect.buf)) {
+    if (y4m_input_open(&input->y4m, input->file, input->detect.buf, 4,
+                       input->only_i420) >= 0) {
+      input->file_type = FILE_TYPE_Y4M;
+      input->width = input->y4m.pic_w;
+      input->height = input->y4m.pic_h;
+      input->pixel_aspect_ratio.numerator = input->y4m.par_n;
+      input->pixel_aspect_ratio.denominator = input->y4m.par_d;
+      input->framerate.numerator = input->y4m.fps_n;
+      input->framerate.denominator = input->y4m.fps_d;
+      input->fmt = input->y4m.aom_fmt;
+      input->bit_depth = input->y4m.bit_depth;
+    } else
+      fatal("Unsupported Y4M stream.");
+  } else if (input->detect.buf_read == 4 && fourcc_is_ivf(input->detect.buf)) {
+    fatal("IVF is not supported as input.");
+  } else {
+    input->file_type = FILE_TYPE_RAW;
+  }
+}
+
+static void close_input_file(struct AvxInputContext *input) {
+  fclose(input->file);
+  if (input->file_type == FILE_TYPE_Y4M) y4m_input_close(&input->y4m);
+}
+
+static struct stream_state *new_stream(struct AvxEncoderConfig *global,
+                                       struct stream_state *prev) {
+  struct stream_state *stream;
+
+  stream = calloc(1, sizeof(*stream));
+  if (stream == NULL) {
+    fatal("Failed to allocate new stream.");
+  }
+
+  if (prev) {
+    memcpy(stream, prev, sizeof(*stream));
+    stream->index++;
+    prev->next = stream;
+  } else {
+    aom_codec_err_t res;
+
+    /* Populate encoder configuration */
+    res = aom_codec_enc_config_default(global->codec->codec_interface(),
+                                       &stream->config.cfg, global->usage);
+    if (res) fatal("Failed to get config: %s\n", aom_codec_err_to_string(res));
+
+    /* Change the default timebase to a high enough value so that the
+     * encoder will always create strictly increasing timestamps.
+     */
+    stream->config.cfg.g_timebase.den = 1000;
+
+    /* Never use the library's default resolution, require it be parsed
+     * from the file or set on the command line.
+     */
+    stream->config.cfg.g_w = 0;
+    stream->config.cfg.g_h = 0;
+
+    /* Initialize remaining stream parameters */
+    stream->config.write_webm = 1;
+#if CONFIG_WEBM_IO
+    stream->config.stereo_fmt = STEREO_FORMAT_MONO;
+    stream->webm_ctx.last_pts_ns = -1;
+    stream->webm_ctx.writer = NULL;
+    stream->webm_ctx.segment = NULL;
+#endif
+
+    /* Allows removal of the application version from the EBML tags */
+    stream->webm_ctx.debug = global->debug;
+  }
+
+  /* Output files must be specified for each stream */
+  stream->config.out_fn = NULL;
+
+  stream->next = NULL;
+  return stream;
+}
+
+static int parse_stream_params(struct AvxEncoderConfig *global,
+                               struct stream_state *stream, char **argv) {
+  char **argi, **argj;
+  struct arg arg;
+  static const arg_def_t **ctrl_args = no_args;
+  static const int *ctrl_args_map = NULL;
+  struct stream_config *config = &stream->config;
+  int eos_mark_found = 0;
+#if CONFIG_HIGHBITDEPTH
+  int test_16bit_internal = 0;
+#endif
+
+  // Handle codec specific options
+  if (0) {
+#if CONFIG_AV1_ENCODER
+  } else if (strcmp(global->codec->name, "av1") == 0) {
+    // TODO(jingning): Reuse AV1 specific encoder configuration parameters.
+    // Consider to expand this set for AV1 encoder control.
+    ctrl_args = av1_args;
+    ctrl_args_map = av1_arg_ctrl_map;
+#endif
+  }
+
+  for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
+    arg.argv_step = 1;
+
+    /* Once we've found an end-of-stream marker (--) we want to continue
+     * shifting arguments but not consuming them.
+     */
+    if (eos_mark_found) {
+      argj++;
+      continue;
+    } else if (!strcmp(*argj, "--")) {
+      eos_mark_found = 1;
+      continue;
+    }
+
+    if (arg_match(&arg, &outputfile, argi)) {
+      config->out_fn = arg.val;
+    } else if (arg_match(&arg, &fpf_name, argi)) {
+      config->stats_fn = arg.val;
+#if CONFIG_FP_MB_STATS
+    } else if (arg_match(&arg, &fpmbf_name, argi)) {
+      config->fpmb_stats_fn = arg.val;
+#endif
+    } else if (arg_match(&arg, &use_webm, argi)) {
+#if CONFIG_WEBM_IO
+      config->write_webm = 1;
+#else
+      die("Error: --webm specified but webm is disabled.");
+#endif
+    } else if (arg_match(&arg, &use_ivf, argi)) {
+      config->write_webm = 0;
+    } else if (arg_match(&arg, &threads, argi)) {
+      config->cfg.g_threads = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &profile, argi)) {
+      config->cfg.g_profile = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &width, argi)) {
+      config->cfg.g_w = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &height, argi)) {
+      config->cfg.g_h = arg_parse_uint(&arg);
+#if CONFIG_HIGHBITDEPTH
+    } else if (arg_match(&arg, &bitdeptharg, argi)) {
+      config->cfg.g_bit_depth = arg_parse_enum_or_int(&arg);
+    } else if (arg_match(&arg, &inbitdeptharg, argi)) {
+      config->cfg.g_input_bit_depth = arg_parse_uint(&arg);
+#endif
+#if CONFIG_WEBM_IO
+    } else if (arg_match(&arg, &stereo_mode, argi)) {
+      config->stereo_fmt = arg_parse_enum_or_int(&arg);
+#endif
+    } else if (arg_match(&arg, &timebase, argi)) {
+      config->cfg.g_timebase = arg_parse_rational(&arg);
+      validate_positive_rational(arg.name, &config->cfg.g_timebase);
+    } else if (arg_match(&arg, &error_resilient, argi)) {
+      config->cfg.g_error_resilient = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &lag_in_frames, argi)) {
+      config->cfg.g_lag_in_frames = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &dropframe_thresh, argi)) {
+      config->cfg.rc_dropframe_thresh = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_allowed, argi)) {
+      config->cfg.rc_resize_allowed = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_width, argi)) {
+      config->cfg.rc_scaled_width = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_height, argi)) {
+      config->cfg.rc_scaled_height = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_up_thresh, argi)) {
+      config->cfg.rc_resize_up_thresh = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &resize_down_thresh, argi)) {
+      config->cfg.rc_resize_down_thresh = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &end_usage, argi)) {
+      config->cfg.rc_end_usage = arg_parse_enum_or_int(&arg);
+    } else if (arg_match(&arg, &target_bitrate, argi)) {
+      config->cfg.rc_target_bitrate = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &min_quantizer, argi)) {
+      config->cfg.rc_min_quantizer = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &max_quantizer, argi)) {
+      config->cfg.rc_max_quantizer = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &undershoot_pct, argi)) {
+      config->cfg.rc_undershoot_pct = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &overshoot_pct, argi)) {
+      config->cfg.rc_overshoot_pct = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &buf_sz, argi)) {
+      config->cfg.rc_buf_sz = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &buf_initial_sz, argi)) {
+      config->cfg.rc_buf_initial_sz = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &buf_optimal_sz, argi)) {
+      config->cfg.rc_buf_optimal_sz = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &bias_pct, argi)) {
+      config->cfg.rc_2pass_vbr_bias_pct = arg_parse_uint(&arg);
+      if (global->passes < 2)
+        warn("option %s ignored in one-pass mode.\n", arg.name);
+    } else if (arg_match(&arg, &minsection_pct, argi)) {
+      config->cfg.rc_2pass_vbr_minsection_pct = arg_parse_uint(&arg);
+
+      if (global->passes < 2)
+        warn("option %s ignored in one-pass mode.\n", arg.name);
+    } else if (arg_match(&arg, &maxsection_pct, argi)) {
+      config->cfg.rc_2pass_vbr_maxsection_pct = arg_parse_uint(&arg);
+
+      if (global->passes < 2)
+        warn("option %s ignored in one-pass mode.\n", arg.name);
+    } else if (arg_match(&arg, &kf_min_dist, argi)) {
+      config->cfg.kf_min_dist = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &kf_max_dist, argi)) {
+      config->cfg.kf_max_dist = arg_parse_uint(&arg);
+    } else if (arg_match(&arg, &kf_disabled, argi)) {
+      config->cfg.kf_mode = AOM_KF_DISABLED;
+#if CONFIG_HIGHBITDEPTH
+    } else if (arg_match(&arg, &test16bitinternalarg, argi)) {
+      if (strcmp(global->codec->name, "av1") == 0 ||
+          strcmp(global->codec->name, "av1") == 0) {
+        test_16bit_internal = 1;
+      }
+#endif
+    } else {
+      int i, match = 0;
+      for (i = 0; ctrl_args[i]; i++) {
+        if (arg_match(&arg, ctrl_args[i], argi)) {
+          int j;
+          match = 1;
+
+          /* Point either to the next free element or the first
+          * instance of this control.
+          */
+          for (j = 0; j < config->arg_ctrl_cnt; j++)
+            if (ctrl_args_map != NULL &&
+                config->arg_ctrls[j][0] == ctrl_args_map[i])
+              break;
+
+          /* Update/insert */
+          assert(j < (int)ARG_CTRL_CNT_MAX);
+          if (ctrl_args_map != NULL && j < (int)ARG_CTRL_CNT_MAX) {
+            config->arg_ctrls[j][0] = ctrl_args_map[i];
+            config->arg_ctrls[j][1] = arg_parse_enum_or_int(&arg);
+            if (j == config->arg_ctrl_cnt) config->arg_ctrl_cnt++;
+          }
+        }
+      }
+      if (!match) argj++;
+    }
+  }
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_LOWBITDEPTH
+  if (strcmp(global->codec->name, "av1") == 0 ||
+      strcmp(global->codec->name, "av1") == 0) {
+    config->use_16bit_internal =
+        test_16bit_internal | (config->cfg.g_profile > 1);
+  }
+#else
+  config->use_16bit_internal = 1;
+#endif
+#endif
+  return eos_mark_found;
+}
+
+#define FOREACH_STREAM(func)                                \
+  do {                                                      \
+    struct stream_state *stream;                            \
+    for (stream = streams; stream; stream = stream->next) { \
+      func;                                                 \
+    }                                                       \
+  } while (0)
+
+static void validate_stream_config(const struct stream_state *stream,
+                                   const struct AvxEncoderConfig *global) {
+  const struct stream_state *streami;
+  (void)global;
+
+  if (!stream->config.cfg.g_w || !stream->config.cfg.g_h)
+    fatal(
+        "Stream %d: Specify stream dimensions with --width (-w) "
+        " and --height (-h)",
+        stream->index);
+
+  // Check that the codec bit depth is greater than the input bit depth.
+  if (stream->config.cfg.g_input_bit_depth >
+      (unsigned int)stream->config.cfg.g_bit_depth) {
+    fatal("Stream %d: codec bit depth (%d) less than input bit depth (%d)",
+          stream->index, (int)stream->config.cfg.g_bit_depth,
+          stream->config.cfg.g_input_bit_depth);
+  }
+
+  for (streami = stream; streami; streami = streami->next) {
+    /* All streams require output files */
+    if (!streami->config.out_fn)
+      fatal("Stream %d: Output file is required (specify with -o)",
+            streami->index);
+
+    /* Check for two streams outputting to the same file */
+    if (streami != stream) {
+      const char *a = stream->config.out_fn;
+      const char *b = streami->config.out_fn;
+      if (!strcmp(a, b) && strcmp(a, "/dev/null") && strcmp(a, ":nul"))
+        fatal("Stream %d: duplicate output file (from stream %d)",
+              streami->index, stream->index);
+    }
+
+    /* Check for two streams sharing a stats file. */
+    if (streami != stream) {
+      const char *a = stream->config.stats_fn;
+      const char *b = streami->config.stats_fn;
+      if (a && b && !strcmp(a, b))
+        fatal("Stream %d: duplicate stats file (from stream %d)",
+              streami->index, stream->index);
+    }
+
+#if CONFIG_FP_MB_STATS
+    /* Check for two streams sharing a mb stats file. */
+    if (streami != stream) {
+      const char *a = stream->config.fpmb_stats_fn;
+      const char *b = streami->config.fpmb_stats_fn;
+      if (a && b && !strcmp(a, b))
+        fatal("Stream %d: duplicate mb stats file (from stream %d)",
+              streami->index, stream->index);
+    }
+#endif
+  }
+}
+
+static void set_stream_dimensions(struct stream_state *stream, unsigned int w,
+                                  unsigned int h) {
+  if (!stream->config.cfg.g_w) {
+    if (!stream->config.cfg.g_h)
+      stream->config.cfg.g_w = w;
+    else
+      stream->config.cfg.g_w = w * stream->config.cfg.g_h / h;
+  }
+  if (!stream->config.cfg.g_h) {
+    stream->config.cfg.g_h = h * stream->config.cfg.g_w / w;
+  }
+}
+
+static const char *file_type_to_string(enum VideoFileType t) {
+  switch (t) {
+    case FILE_TYPE_RAW: return "RAW";
+    case FILE_TYPE_Y4M: return "Y4M";
+    default: return "Other";
+  }
+}
+
+static const char *image_format_to_string(aom_img_fmt_t f) {
+  switch (f) {
+    case AOM_IMG_FMT_I420: return "I420";
+    case AOM_IMG_FMT_I422: return "I422";
+    case AOM_IMG_FMT_I444: return "I444";
+    case AOM_IMG_FMT_I440: return "I440";
+    case AOM_IMG_FMT_YV12: return "YV12";
+    case AOM_IMG_FMT_I42016: return "I42016";
+    case AOM_IMG_FMT_I42216: return "I42216";
+    case AOM_IMG_FMT_I44416: return "I44416";
+    case AOM_IMG_FMT_I44016: return "I44016";
+    default: return "Other";
+  }
+}
+
+static void show_stream_config(struct stream_state *stream,
+                               struct AvxEncoderConfig *global,
+                               struct AvxInputContext *input) {
+#define SHOW(field) \
+  fprintf(stderr, "    %-28s = %d\n", #field, stream->config.cfg.field)
+
+  if (stream->index == 0) {
+    fprintf(stderr, "Codec: %s\n",
+            aom_codec_iface_name(global->codec->codec_interface()));
+    fprintf(stderr, "Source file: %s File Type: %s Format: %s\n",
+            input->filename, file_type_to_string(input->file_type),
+            image_format_to_string(input->fmt));
+  }
+  if (stream->next || stream->index)
+    fprintf(stderr, "\nStream Index: %d\n", stream->index);
+  fprintf(stderr, "Destination file: %s\n", stream->config.out_fn);
+  fprintf(stderr, "Encoder parameters:\n");
+
+  SHOW(g_usage);
+  SHOW(g_threads);
+  SHOW(g_profile);
+  SHOW(g_w);
+  SHOW(g_h);
+  SHOW(g_bit_depth);
+  SHOW(g_input_bit_depth);
+  SHOW(g_timebase.num);
+  SHOW(g_timebase.den);
+  SHOW(g_error_resilient);
+  SHOW(g_pass);
+  SHOW(g_lag_in_frames);
+  SHOW(rc_dropframe_thresh);
+  SHOW(rc_resize_allowed);
+  SHOW(rc_scaled_width);
+  SHOW(rc_scaled_height);
+  SHOW(rc_resize_up_thresh);
+  SHOW(rc_resize_down_thresh);
+  SHOW(rc_end_usage);
+  SHOW(rc_target_bitrate);
+  SHOW(rc_min_quantizer);
+  SHOW(rc_max_quantizer);
+  SHOW(rc_undershoot_pct);
+  SHOW(rc_overshoot_pct);
+  SHOW(rc_buf_sz);
+  SHOW(rc_buf_initial_sz);
+  SHOW(rc_buf_optimal_sz);
+  SHOW(rc_2pass_vbr_bias_pct);
+  SHOW(rc_2pass_vbr_minsection_pct);
+  SHOW(rc_2pass_vbr_maxsection_pct);
+  SHOW(kf_mode);
+  SHOW(kf_min_dist);
+  SHOW(kf_max_dist);
+}
+
+static void open_output_file(struct stream_state *stream,
+                             struct AvxEncoderConfig *global,
+                             const struct AvxRational *pixel_aspect_ratio) {
+  const char *fn = stream->config.out_fn;
+  const struct aom_codec_enc_cfg *const cfg = &stream->config.cfg;
+
+  if (cfg->g_pass == AOM_RC_FIRST_PASS) return;
+
+  stream->file = strcmp(fn, "-") ? fopen(fn, "wb") : set_binary_mode(stdout);
+
+  if (!stream->file) fatal("Failed to open output file");
+
+  if (stream->config.write_webm && fseek(stream->file, 0, SEEK_CUR))
+    fatal("WebM output to pipes not supported.");
+
+#if CONFIG_WEBM_IO
+  if (stream->config.write_webm) {
+    stream->webm_ctx.stream = stream->file;
+    write_webm_file_header(&stream->webm_ctx, cfg, stream->config.stereo_fmt,
+                           global->codec->fourcc, pixel_aspect_ratio);
+  }
+#else
+  (void)pixel_aspect_ratio;
+#endif
+
+  if (!stream->config.write_webm) {
+    ivf_write_file_header(stream->file, cfg, global->codec->fourcc, 0);
+  }
+}
+
+static void close_output_file(struct stream_state *stream,
+                              unsigned int fourcc) {
+  const struct aom_codec_enc_cfg *const cfg = &stream->config.cfg;
+
+  if (cfg->g_pass == AOM_RC_FIRST_PASS) return;
+
+#if CONFIG_WEBM_IO
+  if (stream->config.write_webm) {
+    write_webm_file_footer(&stream->webm_ctx);
+  }
+#endif
+
+  if (!stream->config.write_webm) {
+    if (!fseek(stream->file, 0, SEEK_SET))
+      ivf_write_file_header(stream->file, &stream->config.cfg, fourcc,
+                            stream->frames_out);
+  }
+
+  fclose(stream->file);
+}
+
+static void setup_pass(struct stream_state *stream,
+                       struct AvxEncoderConfig *global, int pass) {
+  if (stream->config.stats_fn) {
+    if (!stats_open_file(&stream->stats, stream->config.stats_fn, pass))
+      fatal("Failed to open statistics store");
+  } else {
+    if (!stats_open_mem(&stream->stats, pass))
+      fatal("Failed to open statistics store");
+  }
+
+#if CONFIG_FP_MB_STATS
+  if (stream->config.fpmb_stats_fn) {
+    if (!stats_open_file(&stream->fpmb_stats, stream->config.fpmb_stats_fn,
+                         pass))
+      fatal("Failed to open mb statistics store");
+  } else {
+    if (!stats_open_mem(&stream->fpmb_stats, pass))
+      fatal("Failed to open mb statistics store");
+  }
+#endif
+
+  stream->config.cfg.g_pass = global->passes == 2
+                                  ? pass ? AOM_RC_LAST_PASS : AOM_RC_FIRST_PASS
+                                  : AOM_RC_ONE_PASS;
+  if (pass) {
+    stream->config.cfg.rc_twopass_stats_in = stats_get(&stream->stats);
+#if CONFIG_FP_MB_STATS
+    stream->config.cfg.rc_firstpass_mb_stats_in =
+        stats_get(&stream->fpmb_stats);
+#endif
+  }
+
+  stream->cx_time = 0;
+  stream->nbytes = 0;
+  stream->frames_out = 0;
+}
+
+static void initialize_encoder(struct stream_state *stream,
+                               struct AvxEncoderConfig *global) {
+  int i;
+  int flags = 0;
+
+  flags |= global->show_psnr ? AOM_CODEC_USE_PSNR : 0;
+  flags |= global->out_part ? AOM_CODEC_USE_OUTPUT_PARTITION : 0;
+#if CONFIG_HIGHBITDEPTH
+  flags |= stream->config.use_16bit_internal ? AOM_CODEC_USE_HIGHBITDEPTH : 0;
+#endif
+
+  /* Construct Encoder Context */
+  aom_codec_enc_init(&stream->encoder, global->codec->codec_interface(),
+                     &stream->config.cfg, flags);
+  ctx_exit_on_error(&stream->encoder, "Failed to initialize encoder");
+
+  /* Note that we bypass the aom_codec_control wrapper macro because
+   * we're being clever to store the control IDs in an array. Real
+   * applications will want to make use of the enumerations directly
+   */
+  for (i = 0; i < stream->config.arg_ctrl_cnt; i++) {
+    int ctrl = stream->config.arg_ctrls[i][0];
+    int value = stream->config.arg_ctrls[i][1];
+    if (aom_codec_control_(&stream->encoder, ctrl, value))
+      fprintf(stderr, "Error: Tried to set control %d = %d\n", ctrl, value);
+
+    ctx_exit_on_error(&stream->encoder, "Failed to control codec");
+  }
+
+#if CONFIG_DECODERS
+  if (global->test_decode != TEST_DECODE_OFF) {
+    const AvxInterface *decoder = get_aom_decoder_by_name(global->codec->name);
+    aom_codec_dec_cfg_t cfg = { 0, 0, 0 };
+    aom_codec_dec_init(&stream->decoder, decoder->codec_interface(), &cfg, 0);
+
+#if CONFIG_AV1_DECODER && CONFIG_EXT_TILE
+    if (strcmp(global->codec->name, "av1") == 0) {
+      aom_codec_control(&stream->decoder, AV1_SET_DECODE_TILE_ROW, -1);
+      ctx_exit_on_error(&stream->decoder, "Failed to set decode_tile_row");
+
+      aom_codec_control(&stream->decoder, AV1_SET_DECODE_TILE_COL, -1);
+      ctx_exit_on_error(&stream->decoder, "Failed to set decode_tile_col");
+    }
+#endif
+  }
+#endif
+}
+
+static void encode_frame(struct stream_state *stream,
+                         struct AvxEncoderConfig *global, struct aom_image *img,
+                         unsigned int frames_in) {
+  aom_codec_pts_t frame_start, next_frame_start;
+  struct aom_codec_enc_cfg *cfg = &stream->config.cfg;
+  struct aom_usec_timer timer;
+
+  frame_start =
+      (cfg->g_timebase.den * (int64_t)(frames_in - 1) * global->framerate.den) /
+      cfg->g_timebase.num / global->framerate.num;
+  next_frame_start =
+      (cfg->g_timebase.den * (int64_t)(frames_in)*global->framerate.den) /
+      cfg->g_timebase.num / global->framerate.num;
+
+/* Scale if necessary */
+#if CONFIG_HIGHBITDEPTH
+  if (img) {
+    if ((img->fmt & AOM_IMG_FMT_HIGHBITDEPTH) &&
+        (img->d_w != cfg->g_w || img->d_h != cfg->g_h)) {
+      if (img->fmt != AOM_IMG_FMT_I42016) {
+        fprintf(stderr, "%s can only scale 4:2:0 inputs\n", exec_name);
+        exit(EXIT_FAILURE);
+      }
+#if CONFIG_LIBYUV
+      if (!stream->img) {
+        stream->img =
+            aom_img_alloc(NULL, AOM_IMG_FMT_I42016, cfg->g_w, cfg->g_h, 16);
+      }
+      I420Scale_16(
+          (uint16 *)img->planes[AOM_PLANE_Y], img->stride[AOM_PLANE_Y] / 2,
+          (uint16 *)img->planes[AOM_PLANE_U], img->stride[AOM_PLANE_U] / 2,
+          (uint16 *)img->planes[AOM_PLANE_V], img->stride[AOM_PLANE_V] / 2,
+          img->d_w, img->d_h, (uint16 *)stream->img->planes[AOM_PLANE_Y],
+          stream->img->stride[AOM_PLANE_Y] / 2,
+          (uint16 *)stream->img->planes[AOM_PLANE_U],
+          stream->img->stride[AOM_PLANE_U] / 2,
+          (uint16 *)stream->img->planes[AOM_PLANE_V],
+          stream->img->stride[AOM_PLANE_V] / 2, stream->img->d_w,
+          stream->img->d_h, kFilterBox);
+      img = stream->img;
+#else
+      stream->encoder.err = 1;
+      ctx_exit_on_error(&stream->encoder,
+                        "Stream %d: Failed to encode frame.\n"
+                        "Scaling disabled in this configuration. \n"
+                        "To enable, configure with --enable-libyuv\n",
+                        stream->index);
+#endif
+    }
+  }
+#endif
+  if (img && (img->d_w != cfg->g_w || img->d_h != cfg->g_h)) {
+    if (img->fmt != AOM_IMG_FMT_I420 && img->fmt != AOM_IMG_FMT_YV12) {
+      fprintf(stderr, "%s can only scale 4:2:0 8bpp inputs\n", exec_name);
+      exit(EXIT_FAILURE);
+    }
+#if CONFIG_LIBYUV
+    if (!stream->img)
+      stream->img =
+          aom_img_alloc(NULL, AOM_IMG_FMT_I420, cfg->g_w, cfg->g_h, 16);
+    I420Scale(
+        img->planes[AOM_PLANE_Y], img->stride[AOM_PLANE_Y],
+        img->planes[AOM_PLANE_U], img->stride[AOM_PLANE_U],
+        img->planes[AOM_PLANE_V], img->stride[AOM_PLANE_V], img->d_w, img->d_h,
+        stream->img->planes[AOM_PLANE_Y], stream->img->stride[AOM_PLANE_Y],
+        stream->img->planes[AOM_PLANE_U], stream->img->stride[AOM_PLANE_U],
+        stream->img->planes[AOM_PLANE_V], stream->img->stride[AOM_PLANE_V],
+        stream->img->d_w, stream->img->d_h, kFilterBox);
+    img = stream->img;
+#else
+    stream->encoder.err = 1;
+    ctx_exit_on_error(&stream->encoder,
+                      "Stream %d: Failed to encode frame.\n"
+                      "Scaling disabled in this configuration. \n"
+                      "To enable, configure with --enable-libyuv\n",
+                      stream->index);
+#endif
+  }
+
+  aom_usec_timer_start(&timer);
+  aom_codec_encode(&stream->encoder, img, frame_start,
+                   (unsigned long)(next_frame_start - frame_start), 0,
+                   global->deadline);
+  aom_usec_timer_mark(&timer);
+  stream->cx_time += aom_usec_timer_elapsed(&timer);
+  ctx_exit_on_error(&stream->encoder, "Stream %d: Failed to encode frame",
+                    stream->index);
+}
+
+static void update_quantizer_histogram(struct stream_state *stream) {
+  if (stream->config.cfg.g_pass != AOM_RC_FIRST_PASS) {
+    int q;
+
+    aom_codec_control(&stream->encoder, AOME_GET_LAST_QUANTIZER_64, &q);
+    ctx_exit_on_error(&stream->encoder, "Failed to read quantizer");
+    stream->counts[q]++;
+  }
+}
+
+static void get_cx_data(struct stream_state *stream,
+                        struct AvxEncoderConfig *global, int *got_data) {
+  const aom_codec_cx_pkt_t *pkt;
+  const struct aom_codec_enc_cfg *cfg = &stream->config.cfg;
+  aom_codec_iter_t iter = NULL;
+
+  *got_data = 0;
+  while ((pkt = aom_codec_get_cx_data(&stream->encoder, &iter))) {
+    static size_t fsize = 0;
+    static FileOffset ivf_header_pos = 0;
+
+    switch (pkt->kind) {
+      case AOM_CODEC_CX_FRAME_PKT:
+        if (!(pkt->data.frame.flags & AOM_FRAME_IS_FRAGMENT)) {
+          stream->frames_out++;
+        }
+        if (!global->quiet)
+          fprintf(stderr, " %6luF", (unsigned long)pkt->data.frame.sz);
+
+        update_rate_histogram(stream->rate_hist, cfg, pkt);
+#if CONFIG_WEBM_IO
+        if (stream->config.write_webm) {
+          write_webm_block(&stream->webm_ctx, cfg, pkt);
+        }
+#endif
+        if (!stream->config.write_webm) {
+          if (pkt->data.frame.partition_id <= 0) {
+            ivf_header_pos = ftello(stream->file);
+            fsize = pkt->data.frame.sz;
+
+            ivf_write_frame_header(stream->file, pkt->data.frame.pts, fsize);
+          } else {
+            fsize += pkt->data.frame.sz;
+
+            if (!(pkt->data.frame.flags & AOM_FRAME_IS_FRAGMENT)) {
+              const FileOffset currpos = ftello(stream->file);
+              fseeko(stream->file, ivf_header_pos, SEEK_SET);
+              ivf_write_frame_size(stream->file, fsize);
+              fseeko(stream->file, currpos, SEEK_SET);
+            }
+          }
+
+          (void)fwrite(pkt->data.frame.buf, 1, pkt->data.frame.sz,
+                       stream->file);
+        }
+        stream->nbytes += pkt->data.raw.sz;
+
+        *got_data = 1;
+#if CONFIG_DECODERS
+        if (global->test_decode != TEST_DECODE_OFF && !stream->mismatch_seen) {
+          aom_codec_decode(&stream->decoder, pkt->data.frame.buf,
+                           (unsigned int)pkt->data.frame.sz, NULL, 0);
+          if (stream->decoder.err) {
+            warn_or_exit_on_error(&stream->decoder,
+                                  global->test_decode == TEST_DECODE_FATAL,
+                                  "Failed to decode frame %d in stream %d",
+                                  stream->frames_out + 1, stream->index);
+            stream->mismatch_seen = stream->frames_out + 1;
+          }
+        }
+#endif
+        break;
+      case AOM_CODEC_STATS_PKT:
+        stream->frames_out++;
+        stats_write(&stream->stats, pkt->data.twopass_stats.buf,
+                    pkt->data.twopass_stats.sz);
+        stream->nbytes += pkt->data.raw.sz;
+        break;
+#if CONFIG_FP_MB_STATS
+      case AOM_CODEC_FPMB_STATS_PKT:
+        stats_write(&stream->fpmb_stats, pkt->data.firstpass_mb_stats.buf,
+                    pkt->data.firstpass_mb_stats.sz);
+        stream->nbytes += pkt->data.raw.sz;
+        break;
+#endif
+      case AOM_CODEC_PSNR_PKT:
+
+        if (global->show_psnr) {
+          int i;
+
+          stream->psnr_sse_total += pkt->data.psnr.sse[0];
+          stream->psnr_samples_total += pkt->data.psnr.samples[0];
+          for (i = 0; i < 4; i++) {
+            if (!global->quiet)
+              fprintf(stderr, "%.3f ", pkt->data.psnr.psnr[i]);
+            stream->psnr_totals[i] += pkt->data.psnr.psnr[i];
+          }
+          stream->psnr_count++;
+        }
+
+        break;
+      default: break;
+    }
+  }
+}
+
+static void show_psnr(struct stream_state *stream, double peak) {
+  int i;
+  double ovpsnr;
+
+  if (!stream->psnr_count) return;
+
+  fprintf(stderr, "Stream %d PSNR (Overall/Avg/Y/U/V)", stream->index);
+  ovpsnr = sse_to_psnr((double)stream->psnr_samples_total, peak,
+                       (double)stream->psnr_sse_total);
+  fprintf(stderr, " %.3f", ovpsnr);
+
+  for (i = 0; i < 4; i++) {
+    fprintf(stderr, " %.3f", stream->psnr_totals[i] / stream->psnr_count);
+  }
+  fprintf(stderr, "\n");
+}
+
+static float usec_to_fps(uint64_t usec, unsigned int frames) {
+  return (float)(usec > 0 ? frames * 1000000.0 / (float)usec : 0);
+}
+
+static void test_decode(struct stream_state *stream,
+                        enum TestDecodeFatality fatal,
+                        const AvxInterface *codec) {
+  aom_image_t enc_img, dec_img;
+
+  if (stream->mismatch_seen) return;
+
+  /* Get the internal reference frame */
+  if (strcmp(codec->name, "vp8") == 0) {
+    struct aom_ref_frame ref_enc, ref_dec;
+    const unsigned int frame_width = (stream->config.cfg.g_w + 15) & ~15;
+    const unsigned int frame_height = (stream->config.cfg.g_h + 15) & ~15;
+    aom_img_alloc(&ref_enc.img, AOM_IMG_FMT_I420, frame_width, frame_height, 1);
+    enc_img = ref_enc.img;
+    aom_img_alloc(&ref_dec.img, AOM_IMG_FMT_I420, frame_width, frame_height, 1);
+    dec_img = ref_dec.img;
+
+    ref_enc.frame_type = AOM_LAST_FRAME;
+    ref_dec.frame_type = AOM_LAST_FRAME;
+    aom_codec_control(&stream->encoder, AOM_COPY_REFERENCE, &ref_enc);
+    aom_codec_control(&stream->decoder, AOM_COPY_REFERENCE, &ref_dec);
+  } else {
+    aom_codec_control(&stream->encoder, AV1_GET_NEW_FRAME_IMAGE, &enc_img);
+    aom_codec_control(&stream->decoder, AV1_GET_NEW_FRAME_IMAGE, &dec_img);
+
+#if CONFIG_HIGHBITDEPTH
+    if ((enc_img.fmt & AOM_IMG_FMT_HIGHBITDEPTH) !=
+        (dec_img.fmt & AOM_IMG_FMT_HIGHBITDEPTH)) {
+      if (enc_img.fmt & AOM_IMG_FMT_HIGHBITDEPTH) {
+        aom_image_t enc_hbd_img;
+        aom_img_alloc(&enc_hbd_img, enc_img.fmt - AOM_IMG_FMT_HIGHBITDEPTH,
+                      enc_img.d_w, enc_img.d_h, 16);
+        aom_img_truncate_16_to_8(&enc_hbd_img, &enc_img);
+        enc_img = enc_hbd_img;
+      }
+      if (dec_img.fmt & AOM_IMG_FMT_HIGHBITDEPTH) {
+        aom_image_t dec_hbd_img;
+        aom_img_alloc(&dec_hbd_img, dec_img.fmt - AOM_IMG_FMT_HIGHBITDEPTH,
+                      dec_img.d_w, dec_img.d_h, 16);
+        aom_img_truncate_16_to_8(&dec_hbd_img, &dec_img);
+        dec_img = dec_hbd_img;
+      }
+    }
+#endif
+  }
+  ctx_exit_on_error(&stream->encoder, "Failed to get encoder reference frame");
+  ctx_exit_on_error(&stream->decoder, "Failed to get decoder reference frame");
+
+  if (!aom_compare_img(&enc_img, &dec_img)) {
+    int y[4], u[4], v[4];
+#if CONFIG_HIGHBITDEPTH
+    if (enc_img.fmt & AOM_IMG_FMT_HIGHBITDEPTH) {
+      aom_find_mismatch_high(&enc_img, &dec_img, y, u, v);
+    } else {
+      aom_find_mismatch(&enc_img, &dec_img, y, u, v);
+    }
+#else
+    aom_find_mismatch(&enc_img, &dec_img, y, u, v);
+#endif
+    stream->decoder.err = 1;
+    warn_or_exit_on_error(&stream->decoder, fatal == TEST_DECODE_FATAL,
+                          "Stream %d: Encode/decode mismatch on frame %d at"
+                          " Y[%d, %d] {%d/%d},"
+                          " U[%d, %d] {%d/%d},"
+                          " V[%d, %d] {%d/%d}",
+                          stream->index, stream->frames_out, y[0], y[1], y[2],
+                          y[3], u[0], u[1], u[2], u[3], v[0], v[1], v[2], v[3]);
+    stream->mismatch_seen = stream->frames_out;
+  }
+
+  aom_img_free(&enc_img);
+  aom_img_free(&dec_img);
+}
+
+static void print_time(const char *label, int64_t etl) {
+  int64_t hours;
+  int64_t mins;
+  int64_t secs;
+
+  if (etl >= 0) {
+    hours = etl / 3600;
+    etl -= hours * 3600;
+    mins = etl / 60;
+    etl -= mins * 60;
+    secs = etl;
+
+    fprintf(stderr, "[%3s %2" PRId64 ":%02" PRId64 ":%02" PRId64 "] ", label,
+            hours, mins, secs);
+  } else {
+    fprintf(stderr, "[%3s  unknown] ", label);
+  }
+}
+
+int main(int argc, const char **argv_) {
+  int pass;
+  aom_image_t raw;
+#if CONFIG_HIGHBITDEPTH
+  aom_image_t raw_shift;
+  int allocated_raw_shift = 0;
+  int use_16bit_internal = 0;
+  int input_shift = 0;
+#endif
+  int frame_avail, got_data;
+
+  struct AvxInputContext input;
+  struct AvxEncoderConfig global;
+  struct stream_state *streams = NULL;
+  char **argv, **argi;
+  uint64_t cx_time = 0;
+  int stream_cnt = 0;
+  int res = 0;
+  int profile_updated = 0;
+
+  memset(&input, 0, sizeof(input));
+  exec_name = argv_[0];
+
+  if (argc < 3) usage_exit();
+
+  /* Setup default input stream settings */
+  input.framerate.numerator = 30;
+  input.framerate.denominator = 1;
+  input.only_i420 = 1;
+  input.bit_depth = 0;
+
+  /* First parse the global configuration values, because we want to apply
+   * other parameters on top of the default configuration provided by the
+   * codec.
+   */
+  argv = argv_dup(argc - 1, argv_ + 1);
+  parse_global_config(&global, argv);
+
+  switch (global.color_type) {
+    case I420: input.fmt = AOM_IMG_FMT_I420; break;
+    case I422: input.fmt = AOM_IMG_FMT_I422; break;
+    case I444: input.fmt = AOM_IMG_FMT_I444; break;
+    case I440: input.fmt = AOM_IMG_FMT_I440; break;
+    case YV12: input.fmt = AOM_IMG_FMT_YV12; break;
+  }
+
+  {
+    /* Now parse each stream's parameters. Using a local scope here
+     * due to the use of 'stream' as loop variable in FOREACH_STREAM
+     * loops
+     */
+    struct stream_state *stream = NULL;
+
+    do {
+      stream = new_stream(&global, stream);
+      stream_cnt++;
+      if (!streams) streams = stream;
+    } while (parse_stream_params(&global, stream, argv));
+  }
+
+  /* Check for unrecognized options */
+  for (argi = argv; *argi; argi++)
+    if (argi[0][0] == '-' && argi[0][1])
+      die("Error: Unrecognized option %s\n", *argi);
+
+  FOREACH_STREAM(check_encoder_config(global.disable_warning_prompt, &global,
+                                      &stream->config.cfg););
+
+  /* Handle non-option arguments */
+  input.filename = argv[0];
+
+  if (!input.filename) usage_exit();
+
+  /* Decide if other chroma subsamplings than 4:2:0 are supported */
+  if (global.codec->fourcc == AV1_FOURCC) input.only_i420 = 0;
+
+  for (pass = global.pass ? global.pass - 1 : 0; pass < global.passes; pass++) {
+    int frames_in = 0, seen_frames = 0;
+    int64_t estimated_time_left = -1;
+    int64_t average_rate = -1;
+    int64_t lagged_count = 0;
+
+    open_input_file(&input);
+
+    /* If the input file doesn't specify its w/h (raw files), try to get
+     * the data from the first stream's configuration.
+     */
+    if (!input.width || !input.height) {
+      FOREACH_STREAM({
+        if (stream->config.cfg.g_w && stream->config.cfg.g_h) {
+          input.width = stream->config.cfg.g_w;
+          input.height = stream->config.cfg.g_h;
+          break;
+        }
+      });
+    }
+
+    /* Update stream configurations from the input file's parameters */
+    if (!input.width || !input.height)
+      fatal(
+          "Specify stream dimensions with --width (-w) "
+          " and --height (-h)");
+
+    /* If input file does not specify bit-depth but input-bit-depth parameter
+     * exists, assume that to be the input bit-depth. However, if the
+     * input-bit-depth paramter does not exist, assume the input bit-depth
+     * to be the same as the codec bit-depth.
+     */
+    if (!input.bit_depth) {
+      FOREACH_STREAM({
+        if (stream->config.cfg.g_input_bit_depth)
+          input.bit_depth = stream->config.cfg.g_input_bit_depth;
+        else
+          input.bit_depth = stream->config.cfg.g_input_bit_depth =
+              (int)stream->config.cfg.g_bit_depth;
+      });
+      if (input.bit_depth > 8) input.fmt |= AOM_IMG_FMT_HIGHBITDEPTH;
+    } else {
+      FOREACH_STREAM(
+          { stream->config.cfg.g_input_bit_depth = input.bit_depth; });
+    }
+
+#if CONFIG_HIGHBITDEPTH
+    FOREACH_STREAM({
+      if (input.fmt != AOM_IMG_FMT_I420 && input.fmt != AOM_IMG_FMT_I42016) {
+        /* Automatically upgrade if input is non-4:2:0 but a 4:2:0 profile
+           was selected. */
+        switch (stream->config.cfg.g_profile) {
+          case 0:
+            stream->config.cfg.g_profile = 1;
+            profile_updated = 1;
+            break;
+          case 2:
+            stream->config.cfg.g_profile = 3;
+            profile_updated = 1;
+            break;
+          default: break;
+        }
+      }
+      /* Automatically set the codec bit depth to match the input bit depth.
+       * Upgrade the profile if required. */
+      if (stream->config.cfg.g_input_bit_depth >
+          (unsigned int)stream->config.cfg.g_bit_depth) {
+        stream->config.cfg.g_bit_depth = stream->config.cfg.g_input_bit_depth;
+      }
+      if (stream->config.cfg.g_bit_depth > 8) {
+        switch (stream->config.cfg.g_profile) {
+          case 0:
+            stream->config.cfg.g_profile = 2;
+            profile_updated = 1;
+            break;
+          case 1:
+            stream->config.cfg.g_profile = 3;
+            profile_updated = 1;
+            break;
+          default: break;
+        }
+      }
+      if (stream->config.cfg.g_profile > 1) {
+        stream->config.use_16bit_internal = 1;
+      }
+      if (profile_updated && !global.quiet) {
+        fprintf(stderr,
+                "Warning: automatically upgrading to profile %d to "
+                "match input format.\n",
+                stream->config.cfg.g_profile);
+      }
+    });
+#else
+    FOREACH_STREAM({
+      if (input.fmt != AOM_IMG_FMT_I420 && input.fmt != AOM_IMG_FMT_I42016) {
+        /* Automatically upgrade if input is non-4:2:0 but a 4:2:0 profile
+           was selected. */
+        switch (stream->config.cfg.g_profile) {
+          case 0:
+            stream->config.cfg.g_profile = 1;
+            profile_updated = 1;
+            break;
+          case 2:
+            stream->config.cfg.g_profile = 3;
+            profile_updated = 1;
+            break;
+          default: break;
+        }
+      }
+      if (profile_updated && !global.quiet) {
+        fprintf(stderr,
+                "Warning: automatically upgrading to profile %d to "
+                "match input format.\n",
+                stream->config.cfg.g_profile);
+      }
+    });
+#endif
+
+    FOREACH_STREAM(set_stream_dimensions(stream, input.width, input.height));
+    FOREACH_STREAM(validate_stream_config(stream, &global));
+
+    /* Ensure that --passes and --pass are consistent. If --pass is set and
+     * --passes=2, ensure --fpf was set.
+     */
+    if (global.pass && global.passes == 2)
+      FOREACH_STREAM({
+        if (!stream->config.stats_fn)
+          die("Stream %d: Must specify --fpf when --pass=%d"
+              " and --passes=2\n",
+              stream->index, global.pass);
+      });
+
+#if !CONFIG_WEBM_IO
+    FOREACH_STREAM({
+      if (stream->config.write_webm) {
+        stream->config.write_webm = 0;
+        warn(
+            "aomenc was compiled without WebM container support."
+            "Producing IVF output");
+      }
+    });
+#endif
+
+    /* Use the frame rate from the file only if none was specified
+     * on the command-line.
+     */
+    if (!global.have_framerate) {
+      global.framerate.num = input.framerate.numerator;
+      global.framerate.den = input.framerate.denominator;
+      FOREACH_STREAM(stream->config.cfg.g_timebase.den = global.framerate.num;
+                     stream->config.cfg.g_timebase.num = global.framerate.den);
+    }
+
+    /* Show configuration */
+    if (global.verbose && pass == 0)
+      FOREACH_STREAM(show_stream_config(stream, &global, &input));
+
+    if (pass == (global.pass ? global.pass - 1 : 0)) {
+      if (input.file_type == FILE_TYPE_Y4M)
+        /*The Y4M reader does its own allocation.
+          Just initialize this here to avoid problems if we never read any
+           frames.*/
+        memset(&raw, 0, sizeof(raw));
+      else
+        aom_img_alloc(&raw, input.fmt, input.width, input.height, 32);
+
+      FOREACH_STREAM(stream->rate_hist = init_rate_histogram(
+                         &stream->config.cfg, &global.framerate));
+    }
+
+    FOREACH_STREAM(setup_pass(stream, &global, pass));
+    FOREACH_STREAM(
+        open_output_file(stream, &global, &input.pixel_aspect_ratio));
+    FOREACH_STREAM(initialize_encoder(stream, &global));
+
+#if CONFIG_HIGHBITDEPTH
+    if (strcmp(global.codec->name, "av1") == 0 ||
+        strcmp(global.codec->name, "av1") == 0) {
+      // Check to see if at least one stream uses 16 bit internal.
+      // Currently assume that the bit_depths for all streams using
+      // highbitdepth are the same.
+      FOREACH_STREAM({
+        if (stream->config.use_16bit_internal) {
+          use_16bit_internal = 1;
+        }
+        if (stream->config.cfg.g_profile == 0) {
+          input_shift = 0;
+        } else {
+          input_shift = (int)stream->config.cfg.g_bit_depth -
+                        stream->config.cfg.g_input_bit_depth;
+        }
+      });
+    }
+#endif
+
+    frame_avail = 1;
+    got_data = 0;
+
+    while (frame_avail || got_data) {
+      struct aom_usec_timer timer;
+
+      if (!global.limit || frames_in < global.limit) {
+        frame_avail = read_frame(&input, &raw);
+
+        if (frame_avail) frames_in++;
+        seen_frames =
+            frames_in > global.skip_frames ? frames_in - global.skip_frames : 0;
+
+        if (!global.quiet) {
+          float fps = usec_to_fps(cx_time, seen_frames);
+          fprintf(stderr, "\rPass %d/%d ", pass + 1, global.passes);
+
+          if (stream_cnt == 1)
+            fprintf(stderr, "frame %4d/%-4d %7" PRId64 "B ", frames_in,
+                    streams->frames_out, (int64_t)streams->nbytes);
+          else
+            fprintf(stderr, "frame %4d ", frames_in);
+
+          fprintf(stderr, "%7" PRId64 " %s %.2f %s ",
+                  cx_time > 9999999 ? cx_time / 1000 : cx_time,
+                  cx_time > 9999999 ? "ms" : "us", fps >= 1.0 ? fps : fps * 60,
+                  fps >= 1.0 ? "fps" : "fpm");
+          print_time("ETA", estimated_time_left);
+        }
+
+      } else {
+        frame_avail = 0;
+      }
+
+      if (frames_in > global.skip_frames) {
+#if CONFIG_HIGHBITDEPTH
+        aom_image_t *frame_to_encode;
+        if (input_shift || (use_16bit_internal && input.bit_depth == 8)) {
+          assert(use_16bit_internal);
+          // Input bit depth and stream bit depth do not match, so up
+          // shift frame to stream bit depth
+          if (!allocated_raw_shift) {
+            aom_img_alloc(&raw_shift, raw.fmt | AOM_IMG_FMT_HIGHBITDEPTH,
+                          input.width, input.height, 32);
+            allocated_raw_shift = 1;
+          }
+          aom_img_upshift(&raw_shift, &raw, input_shift);
+          frame_to_encode = &raw_shift;
+        } else {
+          frame_to_encode = &raw;
+        }
+        aom_usec_timer_start(&timer);
+        if (use_16bit_internal) {
+          assert(frame_to_encode->fmt & AOM_IMG_FMT_HIGHBITDEPTH);
+          FOREACH_STREAM({
+            if (stream->config.use_16bit_internal)
+              encode_frame(stream, &global,
+                           frame_avail ? frame_to_encode : NULL, frames_in);
+            else
+              assert(0);
+          });
+        } else {
+          assert((frame_to_encode->fmt & AOM_IMG_FMT_HIGHBITDEPTH) == 0);
+          FOREACH_STREAM(encode_frame(stream, &global,
+                                      frame_avail ? frame_to_encode : NULL,
+                                      frames_in));
+        }
+#else
+        aom_usec_timer_start(&timer);
+        FOREACH_STREAM(encode_frame(stream, &global, frame_avail ? &raw : NULL,
+                                    frames_in));
+#endif
+        aom_usec_timer_mark(&timer);
+        cx_time += aom_usec_timer_elapsed(&timer);
+
+        FOREACH_STREAM(update_quantizer_histogram(stream));
+
+        got_data = 0;
+        FOREACH_STREAM(get_cx_data(stream, &global, &got_data));
+
+        if (!got_data && input.length && streams != NULL &&
+            !streams->frames_out) {
+          lagged_count = global.limit ? seen_frames : ftello(input.file);
+        } else if (input.length) {
+          int64_t remaining;
+          int64_t rate;
+
+          if (global.limit) {
+            const int64_t frame_in_lagged = (seen_frames - lagged_count) * 1000;
+
+            rate = cx_time ? frame_in_lagged * (int64_t)1000000 / cx_time : 0;
+            remaining = 1000 * (global.limit - global.skip_frames -
+                                seen_frames + lagged_count);
+          } else {
+            const int64_t input_pos = ftello(input.file);
+            const int64_t input_pos_lagged = input_pos - lagged_count;
+            const int64_t input_limit = input.length;
+
+            rate = cx_time ? input_pos_lagged * (int64_t)1000000 / cx_time : 0;
+            remaining = input_limit - input_pos + lagged_count;
+          }
+
+          average_rate =
+              (average_rate <= 0) ? rate : (average_rate * 7 + rate) / 8;
+          estimated_time_left = average_rate ? remaining / average_rate : -1;
+        }
+
+        if (got_data && global.test_decode != TEST_DECODE_OFF)
+          FOREACH_STREAM(test_decode(stream, global.test_decode, global.codec));
+      }
+
+      fflush(stdout);
+      if (!global.quiet) fprintf(stderr, "\033[K");
+    }
+
+    if (stream_cnt > 1) fprintf(stderr, "\n");
+
+    if (!global.quiet) {
+      FOREACH_STREAM(fprintf(
+          stderr, "\rPass %d/%d frame %4d/%-4d %7" PRId64 "B %7" PRId64
+                  "b/f %7" PRId64 "b/s"
+                  " %7" PRId64 " %s (%.2f fps)\033[K\n",
+          pass + 1, global.passes, frames_in, stream->frames_out,
+          (int64_t)stream->nbytes,
+          seen_frames ? (int64_t)(stream->nbytes * 8 / seen_frames) : 0,
+          seen_frames
+              ? (int64_t)stream->nbytes * 8 * (int64_t)global.framerate.num /
+                    global.framerate.den / seen_frames
+              : 0,
+          stream->cx_time > 9999999 ? stream->cx_time / 1000 : stream->cx_time,
+          stream->cx_time > 9999999 ? "ms" : "us",
+          usec_to_fps(stream->cx_time, seen_frames)));
+    }
+
+    if (global.show_psnr) {
+      if (global.codec->fourcc == AV1_FOURCC) {
+        FOREACH_STREAM(
+            show_psnr(stream, (1 << stream->config.cfg.g_input_bit_depth) - 1));
+      } else {
+        FOREACH_STREAM(show_psnr(stream, 255.0));
+      }
+    }
+
+    FOREACH_STREAM(aom_codec_destroy(&stream->encoder));
+
+    if (global.test_decode != TEST_DECODE_OFF) {
+      FOREACH_STREAM(aom_codec_destroy(&stream->decoder));
+    }
+
+    close_input_file(&input);
+
+    if (global.test_decode == TEST_DECODE_FATAL) {
+      FOREACH_STREAM(res |= stream->mismatch_seen);
+    }
+    FOREACH_STREAM(close_output_file(stream, global.codec->fourcc));
+
+    FOREACH_STREAM(stats_close(&stream->stats, global.passes - 1));
+
+#if CONFIG_FP_MB_STATS
+    FOREACH_STREAM(stats_close(&stream->fpmb_stats, global.passes - 1));
+#endif
+
+    if (global.pass) break;
+  }
+
+  if (global.show_q_hist_buckets)
+    FOREACH_STREAM(
+        show_q_histogram(stream->counts, global.show_q_hist_buckets));
+
+  if (global.show_rate_hist_buckets)
+    FOREACH_STREAM(show_rate_histogram(stream->rate_hist, &stream->config.cfg,
+                                       global.show_rate_hist_buckets));
+  FOREACH_STREAM(destroy_rate_histogram(stream->rate_hist));
+
+#if CONFIG_INTERNAL_STATS
+  /* TODO(jkoleszar): This doesn't belong in this executable. Do it for now,
+   * to match some existing utilities.
+   */
+  if (!(global.pass == 1 && global.passes == 2))
+    FOREACH_STREAM({
+      FILE *f = fopen("opsnr.stt", "a");
+      if (stream->mismatch_seen) {
+        fprintf(f, "First mismatch occurred in frame %d\n",
+                stream->mismatch_seen);
+      } else {
+        fprintf(f, "No mismatch detected in recon buffers\n");
+      }
+      fclose(f);
+    });
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+  if (allocated_raw_shift) aom_img_free(&raw_shift);
+#endif
+  aom_img_free(&raw);
+  free(argv);
+  free(streams);
+  return res ? EXIT_FAILURE : EXIT_SUCCESS;
+}
diff --git a/third_party/aom/aomenc.h b/third_party/aom/aomenc.h
new file mode 100644
index 0000000000..248e58356e
--- /dev/null
+++ b/third_party/aom/aomenc.h
@@ -0,0 +1,65 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOMENC_H_
+#define AOMENC_H_
+
+#include "aom/aom_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum TestDecodeFatality {
+  TEST_DECODE_OFF,
+  TEST_DECODE_FATAL,
+  TEST_DECODE_WARN,
+};
+
+typedef enum {
+  I420,  // 4:2:0 8+ bit-depth
+  I422,  // 4:2:2 8+ bit-depth
+  I444,  // 4:4:4 8+ bit-depth
+  I440,  // 4:4:0 8+ bit-depth
+  YV12,  // 4:2:0 with uv flipped, only 8-bit depth
+} ColorInputType;
+
+struct AvxInterface;
+
+/* Configuration elements common to all streams. */
+struct AvxEncoderConfig {
+  const struct AvxInterface *codec;
+  int passes;
+  int pass;
+  int usage;
+  int deadline;
+  ColorInputType color_type;
+  int quiet;
+  int verbose;
+  int limit;
+  int skip_frames;
+  int show_psnr;
+  enum TestDecodeFatality test_decode;
+  int have_framerate;
+  struct aom_rational framerate;
+  int out_part;
+  int debug;
+  int show_q_hist_buckets;
+  int show_rate_hist_buckets;
+  int disable_warnings;
+  int disable_warning_prompt;
+  int experimental_bitstream;
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOMENC_H_
diff --git a/third_party/aom/aomstats.c b/third_party/aom/aomstats.c
new file mode 100644
index 0000000000..0cfeea2f14
--- /dev/null
+++ b/third_party/aom/aomstats.c
@@ -0,0 +1,106 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aomstats.h"
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "./tools_common.h"
+
+int stats_open_file(stats_io_t *stats, const char *fpf, int pass) {
+  int res;
+  stats->pass = pass;
+
+  if (pass == 0) {
+    stats->file = fopen(fpf, "wb");
+    stats->buf.sz = 0;
+    stats->buf.buf = NULL;
+    res = (stats->file != NULL);
+  } else {
+    size_t nbytes;
+
+    stats->file = fopen(fpf, "rb");
+
+    if (stats->file == NULL) fatal("First-pass stats file does not exist!");
+
+    if (fseek(stats->file, 0, SEEK_END))
+      fatal("First-pass stats file must be seekable!");
+
+    stats->buf.sz = stats->buf_alloc_sz = ftell(stats->file);
+    rewind(stats->file);
+
+    stats->buf.buf = malloc(stats->buf_alloc_sz);
+
+    if (!stats->buf.buf)
+      fatal("Failed to allocate first-pass stats buffer (%lu bytes)",
+            (unsigned int)stats->buf_alloc_sz);
+
+    nbytes = fread(stats->buf.buf, 1, stats->buf.sz, stats->file);
+    res = (nbytes == stats->buf.sz);
+  }
+
+  return res;
+}
+
+int stats_open_mem(stats_io_t *stats, int pass) {
+  int res;
+  stats->pass = pass;
+
+  if (!pass) {
+    stats->buf.sz = 0;
+    stats->buf_alloc_sz = 64 * 1024;
+    stats->buf.buf = malloc(stats->buf_alloc_sz);
+  }
+
+  stats->buf_ptr = stats->buf.buf;
+  res = (stats->buf.buf != NULL);
+  return res;
+}
+
+void stats_close(stats_io_t *stats, int last_pass) {
+  if (stats->file) {
+    if (stats->pass == last_pass) {
+      free(stats->buf.buf);
+    }
+
+    fclose(stats->file);
+    stats->file = NULL;
+  } else {
+    if (stats->pass == last_pass) free(stats->buf.buf);
+  }
+}
+
+void stats_write(stats_io_t *stats, const void *pkt, size_t len) {
+  if (stats->file) {
+    (void)fwrite(pkt, 1, len, stats->file);
+  } else {
+    if (stats->buf.sz + len > stats->buf_alloc_sz) {
+      size_t new_sz = stats->buf_alloc_sz + 64 * 1024;
+      char *new_ptr = realloc(stats->buf.buf, new_sz);
+
+      if (new_ptr) {
+        stats->buf_ptr = new_ptr + (stats->buf_ptr - (char *)stats->buf.buf);
+        stats->buf.buf = new_ptr;
+        stats->buf_alloc_sz = new_sz;
+      } else {
+        fatal("Failed to realloc firstpass stats buffer.");
+      }
+    }
+
+    memcpy(stats->buf_ptr, pkt, len);
+    stats->buf.sz += len;
+    stats->buf_ptr += len;
+  }
+}
+
+aom_fixed_buf_t stats_get(stats_io_t *stats) { return stats->buf; }
diff --git a/third_party/aom/aomstats.h b/third_party/aom/aomstats.h
new file mode 100644
index 0000000000..6438093447
--- /dev/null
+++ b/third_party/aom/aomstats.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AOMSTATS_H_
+#define AOMSTATS_H_
+
+#include <stdio.h>
+
+#include "aom/aom_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* This structure is used to abstract the different ways of handling
+ * first pass statistics
+ */
+typedef struct {
+  aom_fixed_buf_t buf;
+  int pass;
+  FILE *file;
+  char *buf_ptr;
+  size_t buf_alloc_sz;
+} stats_io_t;
+
+int stats_open_file(stats_io_t *stats, const char *fpf, int pass);
+int stats_open_mem(stats_io_t *stats, int pass);
+void stats_close(stats_io_t *stats, int last_pass);
+void stats_write(stats_io_t *stats, const void *pkt, size_t len);
+aom_fixed_buf_t stats_get(stats_io_t *stats);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOMSTATS_H_
diff --git a/third_party/aom/args.c b/third_party/aom/args.c
new file mode 100644
index 0000000000..571103595f
--- /dev/null
+++ b/third_party/aom/args.c
@@ -0,0 +1,212 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+#include "args.h"
+
+#include "aom/aom_integer.h"
+#include "aom_ports/msvc.h"
+
+#if defined(__GNUC__) && __GNUC__
+extern void die(const char *fmt, ...) __attribute__((noreturn));
+#else
+extern void die(const char *fmt, ...);
+#endif
+
+struct arg arg_init(char **argv) {
+  struct arg a;
+
+  a.argv = argv;
+  a.argv_step = 1;
+  a.name = NULL;
+  a.val = NULL;
+  a.def = NULL;
+  return a;
+}
+
+int arg_match(struct arg *arg_, const struct arg_def *def, char **argv) {
+  struct arg arg;
+
+  if (!argv[0] || argv[0][0] != '-') return 0;
+
+  arg = arg_init(argv);
+
+  if (def->short_name && strlen(arg.argv[0]) == strlen(def->short_name) + 1 &&
+      !strcmp(arg.argv[0] + 1, def->short_name)) {
+    arg.name = arg.argv[0] + 1;
+    arg.val = def->has_val ? arg.argv[1] : NULL;
+    arg.argv_step = def->has_val ? 2 : 1;
+  } else if (def->long_name) {
+    const size_t name_len = strlen(def->long_name);
+
+    if (strlen(arg.argv[0]) >= name_len + 2 && arg.argv[0][1] == '-' &&
+        !strncmp(arg.argv[0] + 2, def->long_name, name_len) &&
+        (arg.argv[0][name_len + 2] == '=' ||
+         arg.argv[0][name_len + 2] == '\0')) {
+      arg.name = arg.argv[0] + 2;
+      arg.val = arg.name[name_len] == '=' ? arg.name + name_len + 1 : NULL;
+      arg.argv_step = 1;
+    }
+  }
+
+  if (arg.name && !arg.val && def->has_val)
+    die("Error: option %s requires argument.\n", arg.name);
+
+  if (arg.name && arg.val && !def->has_val)
+    die("Error: option %s requires no argument.\n", arg.name);
+
+  if (arg.name && (arg.val || !def->has_val)) {
+    arg.def = def;
+    *arg_ = arg;
+    return 1;
+  }
+
+  return 0;
+}
+
+const char *arg_next(struct arg *arg) {
+  if (arg->argv[0]) arg->argv += arg->argv_step;
+
+  return *arg->argv;
+}
+
+char **argv_dup(int argc, const char **argv) {
+  char **new_argv = malloc((argc + 1) * sizeof(*argv));
+
+  memcpy(new_argv, argv, argc * sizeof(*argv));
+  new_argv[argc] = NULL;
+  return new_argv;
+}
+
+void arg_show_usage(FILE *fp, const struct arg_def *const *defs) {
+  char option_text[40] = { 0 };
+
+  for (; *defs; defs++) {
+    const struct arg_def *def = *defs;
+    char *short_val = def->has_val ? " <arg>" : "";
+    char *long_val = def->has_val ? "=<arg>" : "";
+
+    if (def->short_name && def->long_name) {
+      char *comma = def->has_val ? "," : ",      ";
+
+      snprintf(option_text, 37, "-%s%s%s --%s%6s", def->short_name, short_val,
+               comma, def->long_name, long_val);
+    } else if (def->short_name)
+      snprintf(option_text, 37, "-%s%s", def->short_name, short_val);
+    else if (def->long_name)
+      snprintf(option_text, 37, "          --%s%s", def->long_name, long_val);
+
+    fprintf(fp, "  %-37s\t%s\n", option_text, def->desc);
+
+    if (def->enums) {
+      const struct arg_enum_list *listptr;
+
+      fprintf(fp, "  %-37s\t  ", "");
+
+      for (listptr = def->enums; listptr->name; listptr++)
+        fprintf(fp, "%s%s", listptr->name, listptr[1].name ? ", " : "\n");
+    }
+  }
+}
+
+unsigned int arg_parse_uint(const struct arg *arg) {
+  char *endptr;
+  const unsigned long rawval = strtoul(arg->val, &endptr, 10);  // NOLINT
+
+  if (arg->val[0] != '\0' && endptr[0] == '\0') {
+    if (rawval <= UINT_MAX) return (unsigned int)rawval;
+
+    die("Option %s: Value %lu out of range for unsigned int\n", arg->name,
+        rawval);
+  }
+
+  die("Option %s: Invalid character '%c'\n", arg->name, *endptr);
+  return 0;
+}
+
+int arg_parse_int(const struct arg *arg) {
+  char *endptr;
+  const long rawval = strtol(arg->val, &endptr, 10);  // NOLINT
+
+  if (arg->val[0] != '\0' && endptr[0] == '\0') {
+    if (rawval >= INT_MIN && rawval <= INT_MAX) return (int)rawval;
+
+    die("Option %s: Value %ld out of range for signed int\n", arg->name,
+        rawval);
+  }
+
+  die("Option %s: Invalid character '%c'\n", arg->name, *endptr);
+  return 0;
+}
+
+struct aom_rational {
+  int num; /**< fraction numerator */
+  int den; /**< fraction denominator */
+};
+struct aom_rational arg_parse_rational(const struct arg *arg) {
+  long int rawval;
+  char *endptr;
+  struct aom_rational rat;
+
+  /* parse numerator */
+  rawval = strtol(arg->val, &endptr, 10);
+
+  if (arg->val[0] != '\0' && endptr[0] == '/') {
+    if (rawval >= INT_MIN && rawval <= INT_MAX)
+      rat.num = (int)rawval;
+    else
+      die("Option %s: Value %ld out of range for signed int\n", arg->name,
+          rawval);
+  } else
+    die("Option %s: Expected / at '%c'\n", arg->name, *endptr);
+
+  /* parse denominator */
+  rawval = strtol(endptr + 1, &endptr, 10);
+
+  if (arg->val[0] != '\0' && endptr[0] == '\0') {
+    if (rawval >= INT_MIN && rawval <= INT_MAX)
+      rat.den = (int)rawval;
+    else
+      die("Option %s: Value %ld out of range for signed int\n", arg->name,
+          rawval);
+  } else
+    die("Option %s: Invalid character '%c'\n", arg->name, *endptr);
+
+  return rat;
+}
+
+int arg_parse_enum(const struct arg *arg) {
+  const struct arg_enum_list *listptr;
+  long int rawval;
+  char *endptr;
+
+  /* First see if the value can be parsed as a raw value */
+  rawval = strtol(arg->val, &endptr, 10);
+  if (arg->val[0] != '\0' && endptr[0] == '\0') {
+    /* Got a raw value, make sure it's valid */
+    for (listptr = arg->def->enums; listptr->name; listptr++)
+      if (listptr->val == rawval) return (int)rawval;
+  }
+
+  /* Next see if it can be parsed as a string */
+  for (listptr = arg->def->enums; listptr->name; listptr++)
+    if (!strcmp(arg->val, listptr->name)) return listptr->val;
+
+  die("Option %s: Invalid value '%s'\n", arg->name, arg->val);
+  return 0;
+}
+
+int arg_parse_enum_or_int(const struct arg *arg) {
+  if (arg->def->enums) return arg_parse_enum(arg);
+  return arg_parse_int(arg);
+}
diff --git a/third_party/aom/args.h b/third_party/aom/args.h
new file mode 100644
index 0000000000..e7841fc646
--- /dev/null
+++ b/third_party/aom/args.h
@@ -0,0 +1,64 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef ARGS_H_
+#define ARGS_H_
+#include <stdio.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct arg {
+  char **argv;
+  const char *name;
+  const char *val;
+  unsigned int argv_step;
+  const struct arg_def *def;
+};
+
+struct arg_enum_list {
+  const char *name;
+  int val;
+};
+#define ARG_ENUM_LIST_END \
+  { 0 }
+
+typedef struct arg_def {
+  const char *short_name;
+  const char *long_name;
+  int has_val;
+  const char *desc;
+  const struct arg_enum_list *enums;
+} arg_def_t;
+#define ARG_DEF(s, l, v, d) \
+  { s, l, v, d, NULL }
+#define ARG_DEF_ENUM(s, l, v, d, e) \
+  { s, l, v, d, e }
+#define ARG_DEF_LIST_END \
+  { 0 }
+
+struct arg arg_init(char **argv);
+int arg_match(struct arg *arg_, const struct arg_def *def, char **argv);
+const char *arg_next(struct arg *arg);
+void arg_show_usage(FILE *fp, const struct arg_def *const *defs);
+char **argv_dup(int argc, const char **argv);
+
+unsigned int arg_parse_uint(const struct arg *arg);
+int arg_parse_int(const struct arg *arg);
+struct aom_rational arg_parse_rational(const struct arg *arg);
+int arg_parse_enum(const struct arg *arg);
+int arg_parse_enum_or_int(const struct arg *arg);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // ARGS_H_
diff --git a/third_party/aom/av1/av1.cmake b/third_party/aom/av1/av1.cmake
new file mode 100644
index 0000000000..00f687a0df
--- /dev/null
+++ b/third_party/aom/av1/av1.cmake
@@ -0,0 +1,518 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+set(AOM_AV1_COMMON_SOURCES
+    "${AOM_ROOT}/av1/av1_iface_common.h"
+    "${AOM_ROOT}/av1/common/alloccommon.c"
+    "${AOM_ROOT}/av1/common/alloccommon.h"
+    # TODO(tomfinegan): Foward transform belongs in encoder.
+    "${AOM_ROOT}/av1/common/av1_fwd_txfm1d.c"
+    "${AOM_ROOT}/av1/common/av1_fwd_txfm1d.h"
+    "${AOM_ROOT}/av1/common/av1_fwd_txfm2d.c"
+    "${AOM_ROOT}/av1/common/av1_fwd_txfm2d_cfg.h"
+    "${AOM_ROOT}/av1/common/av1_inv_txfm1d.c"
+    "${AOM_ROOT}/av1/common/av1_inv_txfm1d.h"
+    "${AOM_ROOT}/av1/common/av1_inv_txfm2d.c"
+    "${AOM_ROOT}/av1/common/av1_inv_txfm2d_cfg.h"
+    "${AOM_ROOT}/av1/common/av1_loopfilter.c"
+    "${AOM_ROOT}/av1/common/av1_loopfilter.h"
+    "${AOM_ROOT}/av1/common/av1_txfm.h"
+    "${AOM_ROOT}/av1/common/blockd.c"
+    "${AOM_ROOT}/av1/common/blockd.h"
+    "${AOM_ROOT}/av1/common/common.h"
+    "${AOM_ROOT}/av1/common/common_data.h"
+    "${AOM_ROOT}/av1/common/convolve.c"
+    "${AOM_ROOT}/av1/common/convolve.h"
+    "${AOM_ROOT}/av1/common/debugmodes.c"
+    "${AOM_ROOT}/av1/common/entropy.c"
+    "${AOM_ROOT}/av1/common/entropy.h"
+    "${AOM_ROOT}/av1/common/entropymode.c"
+    "${AOM_ROOT}/av1/common/entropymode.h"
+    "${AOM_ROOT}/av1/common/entropymv.c"
+    "${AOM_ROOT}/av1/common/entropymv.h"
+    "${AOM_ROOT}/av1/common/enums.h"
+    "${AOM_ROOT}/av1/common/filter.c"
+    "${AOM_ROOT}/av1/common/filter.h"
+    "${AOM_ROOT}/av1/common/frame_buffers.c"
+    "${AOM_ROOT}/av1/common/frame_buffers.h"
+    "${AOM_ROOT}/av1/common/idct.c"
+    "${AOM_ROOT}/av1/common/idct.h"
+    "${AOM_ROOT}/av1/common/mv.h"
+    "${AOM_ROOT}/av1/common/mvref_common.c"
+    "${AOM_ROOT}/av1/common/mvref_common.h"
+    "${AOM_ROOT}/av1/common/odintrin.c"
+    "${AOM_ROOT}/av1/common/odintrin.h"
+    "${AOM_ROOT}/av1/common/onyxc_int.h"
+    "${AOM_ROOT}/av1/common/pred_common.c"
+    "${AOM_ROOT}/av1/common/pred_common.h"
+    "${AOM_ROOT}/av1/common/quant_common.c"
+    "${AOM_ROOT}/av1/common/quant_common.h"
+    "${AOM_ROOT}/av1/common/reconinter.c"
+    "${AOM_ROOT}/av1/common/reconinter.h"
+    "${AOM_ROOT}/av1/common/reconintra.c"
+    "${AOM_ROOT}/av1/common/reconintra.h"
+    "${AOM_ROOT}/av1/common/resize.c"
+    "${AOM_ROOT}/av1/common/resize.h"
+    "${AOM_ROOT}/av1/common/restoration.h"
+    "${AOM_ROOT}/av1/common/scale.c"
+    "${AOM_ROOT}/av1/common/scale.h"
+    "${AOM_ROOT}/av1/common/scan.c"
+    "${AOM_ROOT}/av1/common/scan.h"
+    "${AOM_ROOT}/av1/common/seg_common.c"
+    "${AOM_ROOT}/av1/common/seg_common.h"
+    "${AOM_ROOT}/av1/common/thread_common.c"
+    "${AOM_ROOT}/av1/common/thread_common.h"
+    "${AOM_ROOT}/av1/common/tile_common.c"
+    "${AOM_ROOT}/av1/common/tile_common.h")
+
+set(AOM_AV1_DECODER_SOURCES
+    "${AOM_ROOT}/av1/av1_dx_iface.c"
+    "${AOM_ROOT}/av1/decoder/decodeframe.c"
+    "${AOM_ROOT}/av1/decoder/decodeframe.h"
+    "${AOM_ROOT}/av1/decoder/decodemv.c"
+    "${AOM_ROOT}/av1/decoder/decodemv.h"
+    "${AOM_ROOT}/av1/decoder/decoder.c"
+    "${AOM_ROOT}/av1/decoder/decoder.h"
+    "${AOM_ROOT}/av1/decoder/detokenize.c"
+    "${AOM_ROOT}/av1/decoder/detokenize.h"
+    "${AOM_ROOT}/av1/decoder/dsubexp.c"
+    "${AOM_ROOT}/av1/decoder/dsubexp.h"
+    "${AOM_ROOT}/av1/decoder/dthread.c"
+    "${AOM_ROOT}/av1/decoder/dthread.h")
+
+set(AOM_AV1_ENCODER_SOURCES
+    "${AOM_ROOT}/av1/av1_cx_iface.c"
+    "${AOM_ROOT}/av1/encoder/aq_complexity.c"
+    "${AOM_ROOT}/av1/encoder/aq_complexity.h"
+    "${AOM_ROOT}/av1/encoder/aq_cyclicrefresh.c"
+    "${AOM_ROOT}/av1/encoder/aq_cyclicrefresh.h"
+    "${AOM_ROOT}/av1/encoder/aq_variance.c"
+    "${AOM_ROOT}/av1/encoder/aq_variance.h"
+    "${AOM_ROOT}/av1/encoder/av1_quantize.c"
+    "${AOM_ROOT}/av1/encoder/av1_quantize.h"
+    "${AOM_ROOT}/av1/encoder/bitstream.c"
+    "${AOM_ROOT}/av1/encoder/bitstream.h"
+    "${AOM_ROOT}/av1/encoder/block.h"
+    "${AOM_ROOT}/av1/encoder/context_tree.c"
+    "${AOM_ROOT}/av1/encoder/context_tree.h"
+    "${AOM_ROOT}/av1/encoder/cost.c"
+    "${AOM_ROOT}/av1/encoder/cost.h"
+    "${AOM_ROOT}/av1/encoder/dct.c"
+    "${AOM_ROOT}/av1/encoder/encodeframe.c"
+    "${AOM_ROOT}/av1/encoder/encodeframe.h"
+    "${AOM_ROOT}/av1/encoder/encodemb.c"
+    "${AOM_ROOT}/av1/encoder/encodemb.h"
+    "${AOM_ROOT}/av1/encoder/encodemv.c"
+    "${AOM_ROOT}/av1/encoder/encodemv.h"
+    "${AOM_ROOT}/av1/encoder/encoder.c"
+    "${AOM_ROOT}/av1/encoder/encoder.h"
+    "${AOM_ROOT}/av1/encoder/ethread.c"
+    "${AOM_ROOT}/av1/encoder/ethread.h"
+    "${AOM_ROOT}/av1/encoder/extend.c"
+    "${AOM_ROOT}/av1/encoder/extend.h"
+    "${AOM_ROOT}/av1/encoder/firstpass.c"
+    "${AOM_ROOT}/av1/encoder/firstpass.h"
+    "${AOM_ROOT}/av1/encoder/hybrid_fwd_txfm.c"
+    "${AOM_ROOT}/av1/encoder/hybrid_fwd_txfm.h"
+    "${AOM_ROOT}/av1/encoder/lookahead.c"
+    "${AOM_ROOT}/av1/encoder/lookahead.h"
+    "${AOM_ROOT}/av1/encoder/mbgraph.c"
+    "${AOM_ROOT}/av1/encoder/mbgraph.h"
+    "${AOM_ROOT}/av1/encoder/mcomp.c"
+    "${AOM_ROOT}/av1/encoder/mcomp.h"
+    "${AOM_ROOT}/av1/encoder/picklpf.c"
+    "${AOM_ROOT}/av1/encoder/picklpf.h"
+    "${AOM_ROOT}/av1/encoder/ratectrl.c"
+    "${AOM_ROOT}/av1/encoder/ratectrl.h"
+    "${AOM_ROOT}/av1/encoder/rd.c"
+    "${AOM_ROOT}/av1/encoder/rd.h"
+    "${AOM_ROOT}/av1/encoder/rdopt.c"
+    "${AOM_ROOT}/av1/encoder/rdopt.h"
+    "${AOM_ROOT}/av1/encoder/segmentation.c"
+    "${AOM_ROOT}/av1/encoder/segmentation.h"
+    "${AOM_ROOT}/av1/encoder/speed_features.c"
+    "${AOM_ROOT}/av1/encoder/speed_features.h"
+    "${AOM_ROOT}/av1/encoder/subexp.c"
+    "${AOM_ROOT}/av1/encoder/subexp.h"
+    "${AOM_ROOT}/av1/encoder/temporal_filter.c"
+    "${AOM_ROOT}/av1/encoder/temporal_filter.h"
+    "${AOM_ROOT}/av1/encoder/tokenize.c"
+    "${AOM_ROOT}/av1/encoder/tokenize.h"
+    "${AOM_ROOT}/av1/encoder/treewriter.c"
+    "${AOM_ROOT}/av1/encoder/treewriter.h"
+    "${AOM_ROOT}/av1/encoder/variance_tree.c"
+    "${AOM_ROOT}/av1/encoder/variance_tree.h")
+
+set(AOM_AV1_COMMON_INTRIN_SSE2
+    # Requires CONFIG_GLOBAL_MOTION or CONFIG_WARPED_MOTION
+    #"${AOM_ROOT}/av1/common/x86/warp_plane_sse2.c"
+    "${AOM_ROOT}/av1/common/x86/idct_intrin_sse2.c")
+
+set(AOM_AV1_COMMON_INTRIN_SSSE3
+    "${AOM_ROOT}/av1/common/x86/av1_convolve_ssse3.c")
+
+set(AOM_AV1_COMMON_INTRIN_SSE4_1
+    "${AOM_ROOT}/av1/common/x86/av1_fwd_txfm1d_sse4.c"
+    "${AOM_ROOT}/av1/common/x86/av1_fwd_txfm2d_sse4.c")
+
+set(AOM_AV1_COMMON_INTRIN_AVX2
+    "${AOM_ROOT}/av1/common/x86/hybrid_inv_txfm_avx2.c")
+
+set(AOM_AV1_COMMON_INTRIN_DSPR2
+    "${AOM_ROOT}/av1/common/mips/dspr2/av1_itrans16_dspr2.c"
+    "${AOM_ROOT}/av1/common/mips/dspr2/av1_itrans4_dspr2.c"
+    "${AOM_ROOT}/av1/common/mips/dspr2/av1_itrans8_dspr2.c")
+
+set(AOM_AV1_COMMON_INTRIN_MSA
+    "${AOM_ROOT}/av1/common/mips/msa/av1_idct16x16_msa.c"
+    "${AOM_ROOT}/av1/common/mips/msa/av1_idct4x4_msa.c"
+    "${AOM_ROOT}/av1/common/mips/msa/av1_idct8x8_msa.c")
+
+set(AOM_AV1_ENCODER_ASM_SSE2
+    "${AOM_ROOT}/av1/encoder/x86/dct_sse2.asm"
+    "${AOM_ROOT}/av1/encoder/x86/error_sse2.asm"
+    "${AOM_ROOT}/av1/encoder/x86/temporal_filter_apply_sse2.asm")
+
+set(AOM_AV1_ENCODER_INTRIN_SSE2
+    "${AOM_ROOT}/av1/encoder/x86/dct_intrin_sse2.c"
+    "${AOM_ROOT}/av1/encoder/x86/highbd_block_error_intrin_sse2.c"
+    "${AOM_ROOT}/av1/encoder/x86/av1_quantize_sse2.c")
+
+set(AOM_AV1_ENCODER_ASM_SSSE3_X86_64
+    "${AOM_ROOT}/av1/encoder/x86/av1_quantize_ssse3_x86_64.asm")
+
+set(AOM_AV1_ENCODER_INTRIN_SSSE3
+    "${AOM_ROOT}/av1/encoder/x86/dct_ssse3.c")
+
+set(AOM_AV1_ENCODER_INTRIN_AVX2
+    "${AOM_ROOT}/av1/encoder/x86/error_intrin_avx2.c"
+    "${AOM_ROOT}/av1/encoder/x86/hybrid_fwd_txfm_avx2.c")
+
+set(AOM_AV1_ENCODER_INTRIN_NEON
+    "${AOM_ROOT}/av1/encoder/arm/neon/quantize_neon.c")
+
+set(AOM_AV1_ENCODER_INTRIN_MSA
+    "${AOM_ROOT}/av1/encoder/mips/msa/error_msa.c"
+    "${AOM_ROOT}/av1/encoder/mips/msa/fdct16x16_msa.c"
+    "${AOM_ROOT}/av1/encoder/mips/msa/fdct4x4_msa.c"
+    "${AOM_ROOT}/av1/encoder/mips/msa/fdct8x8_msa.c"
+    "${AOM_ROOT}/av1/encoder/mips/msa/fdct_msa.h"
+    "${AOM_ROOT}/av1/encoder/mips/msa/temporal_filter_msa.c")
+
+if (CONFIG_HIGHBITDEPTH)
+  set(AOM_AV1_COMMON_INTRIN_SSE4_1
+      ${AOM_AV1_COMMON_INTRIN_SSE4_1}
+      "${AOM_ROOT}/av1/common/x86/av1_highbd_convolve_sse4.c"
+      "${AOM_ROOT}/av1/common/x86/highbd_inv_txfm_sse4.c")
+
+  set(AOM_AV1_COMMON_INTRIN_AVX2
+      ${AOM_AV1_COMMON_INTRIN_AVX2}
+      "${AOM_ROOT}/av1/common/x86/highbd_inv_txfm_avx2.c")
+
+  set(AOM_AV1_ENCODER_INTRIN_SSE4_1
+      ${AOM_AV1_ENCODER_INTRIN_SSE4_1}
+      "${AOM_ROOT}/av1/encoder/x86/av1_highbd_quantize_sse4.c"
+      "${AOM_ROOT}/av1/encoder/x86/highbd_fwd_txfm_sse4.c")
+else ()
+  set(AOM_AV1_COMMON_INTRIN_NEON
+      ${AOM_AV1_COMMON_INTRIN_NEON}
+      "${AOM_ROOT}/av1/encoder/arm/neon/dct_neon.c"
+      "${AOM_ROOT}/av1/common/arm/neon/iht4x4_add_neon.c"
+      "${AOM_ROOT}/av1/common/arm/neon/iht8x8_add_neon.c")
+
+  set(AOM_AV1_ENCODER_INTRIN_NEON
+      ${AOM_AV1_ENCODER_INTRIN_NEON}
+      "${AOM_ROOT}/av1/encoder/arm/neon/error_neon.c")
+endif ()
+
+if (CONFIG_CDEF)
+  set(AOM_AV1_COMMON_SOURCES
+      ${AOM_AV1_COMMON_SOURCES}
+      "${AOM_ROOT}/av1/common/clpf.c"
+      "${AOM_ROOT}/av1/common/clpf.h"
+      "${AOM_ROOT}/av1/common/clpf_simd.h"
+      "${AOM_ROOT}/av1/common/cdef_simd.h"
+      "${AOM_ROOT}/av1/common/cdef.c"
+      "${AOM_ROOT}/av1/common/cdef.h"
+      "${AOM_ROOT}/av1/common/od_dering.c"
+      "${AOM_ROOT}/av1/common/od_dering.h"
+      "${AOM_ROOT}/av1/common/od_dering_simd.h")
+
+  set(AOM_AV1_ENCODER_SOURCES
+      ${AOM_AV1_ENCODER_SOURCES}
+      "${AOM_ROOT}/av1/encoder/pickcdef.c")
+
+  set(AOM_AV1_COMMON_INTRIN_SSE2
+      ${AOM_AV1_COMMON_INTRIN_SSE2}
+      "${AOM_ROOT}/av1/common/clpf_sse2.c"
+      "${AOM_ROOT}/av1/common/od_dering_sse2.c")
+
+  set(AOM_AV1_COMMON_INTRIN_SSSE3
+      ${AOM_AV1_COMMON_INTRIN_SSSE3}
+      "${AOM_ROOT}/av1/common/clpf_ssse3.c"
+      "${AOM_ROOT}/av1/common/od_dering_ssse3.c")
+
+  set(AOM_AV1_COMMON_INTRIN_SSE4_1
+      ${AOM_AV1_COMMON_INTRIN_SSE4_1}
+      "${AOM_ROOT}/av1/common/clpf_sse4.c"
+      "${AOM_ROOT}/av1/common/od_dering_sse4.c")
+
+  set(AOM_AV1_COMMON_INTRIN_NEON
+      ${AOM_AV1_COMMON_INTRIN_NEON}
+      "${AOM_ROOT}/av1/common/clpf_neon.c"
+      "${AOM_ROOT}/av1/common/od_dering_neon.c")
+endif ()
+
+if (CONFIG_EXT_INTER)
+  set(AOM_AV1_ENCODER_SOURCES
+      ${AOM_AV1_ENCODER_SOURCES}
+      "${AOM_ROOT}/av1/encoder/wedge_utils.c")
+
+  set(AOM_AV1_ENCODER_INTRIN_SSE2
+      ${AOM_AV1_ENCODER_INTRIN_SSE2}
+      "${AOM_ROOT}/av1/encoder/x86/wedge_utils_sse2.c")
+endif ()
+
+if (CONFIG_FILTER_INTRA)
+  set(AOM_AV1_COMMON_INTRIN_SSE4_1
+      ${AOM_AV1_COMMON_INTRIN_SSE4_1}
+      "${AOM_ROOT}/av1/common/x86/filterintra_sse4.c")
+endif ()
+
+if (CONFIG_ACCOUNTING)
+  set(AOM_AV1_DECODER_SOURCES
+      ${AOM_AV1_DECODER_SOURCES}
+      "${AOM_ROOT}/av1/decoder/accounting.c"
+      "${AOM_ROOT}/av1/decoder/accounting.h")
+endif ()
+
+if (CONFIG_INSPECTION)
+  set(AOM_AV1_DECODER_SOURCES
+      ${AOM_AV1_DECODER_SOURCES}
+      "${AOM_ROOT}/av1/decoder/inspection.c"
+      "${AOM_ROOT}/av1/decoder/inspection.h")
+endif ()
+
+if (CONFIG_INTERNAL_STATS)
+  set(AOM_AV1_ENCODER_SOURCES
+      ${AOM_AV1_ENCODER_SOURCES}
+      "${AOM_ROOT}/av1/encoder/blockiness.c")
+endif ()
+
+if (CONFIG_PALETTE)
+  set(AOM_AV1_ENCODER_SOURCES
+      ${AOM_AV1_ENCODER_SOURCES}
+      "${AOM_ROOT}/av1/encoder/palette.c"
+      "${AOM_ROOT}/av1/encoder/palette.h")
+endif ()
+
+if (CONFIG_CFL)
+  set(AOM_AV1_COMMON_SOURCES
+      ${AOM_AV1_COMMON_SOURCES}
+    "${AOM_ROOT}/av1/common/cfl.c"
+    "${AOM_ROOT}/av1/common/cfl.h")
+endif ()
+
+if (CONFIG_PVQ)
+  set(AOM_AV1_COMMON_SOURCES
+      ${AOM_AV1_COMMON_SOURCES}
+      "${AOM_ROOT}/av1/common/laplace_tables.c"
+      "${AOM_ROOT}/av1/common/pvq.c"
+      "${AOM_ROOT}/av1/common/pvq.h"
+      "${AOM_ROOT}/av1/common/pvq_state.c"
+      "${AOM_ROOT}/av1/common/pvq_state.h"
+      "${AOM_ROOT}/av1/common/partition.c"
+      "${AOM_ROOT}/av1/common/partition.h"
+      "${AOM_ROOT}/av1/common/generic_code.c"
+      "${AOM_ROOT}/av1/common/generic_code.h"
+      "${AOM_ROOT}/av1/common/zigzag4.c"
+      "${AOM_ROOT}/av1/common/zigzag8.c"
+      "${AOM_ROOT}/av1/common/zigzag16.c"
+      "${AOM_ROOT}/av1/common/zigzag32.c")
+
+    set(AOM_AV1_DECODER_SOURCES
+        ${AOM_AV1_DECODER_SOURCES}
+        "${AOM_ROOT}/av1/decoder/decint.h"
+        "${AOM_ROOT}/av1/decoder/pvq_decoder.c"
+        "${AOM_ROOT}/av1/decoder/pvq_decoder.h"
+        "${AOM_ROOT}/av1/decoder/generic_decoder.c"
+        "${AOM_ROOT}/av1/decoder/laplace_decoder.c")
+
+    set(AOM_AV1_ENCODER_SOURCES
+        ${AOM_AV1_ENCODER_SOURCES}
+        "${AOM_ROOT}/av1/encoder/daala_compat_enc.c"
+        "${AOM_ROOT}/av1/encoder/encint.h"
+        "${AOM_ROOT}/av1/encoder/pvq_encoder.c"
+        "${AOM_ROOT}/av1/encoder/pvq_encoder.h"
+        "${AOM_ROOT}/av1/encoder/generic_encoder.c"
+        "${AOM_ROOT}/av1/encoder/laplace_encoder.c")
+
+    set(AOM_AV1_COMMON_INTRIN_SSE4_1
+        ${AOM_AV1_COMMON_INTRIN_SSE4_1}
+        "${AOM_ROOT}/av1/common/x86/pvq_sse4.c"
+        "${AOM_ROOT}/av1/common/x86/pvq_sse4.h")
+
+    if (NOT CONFIG_AV1_ENCODER)
+      # TODO(tomfinegan): These should probably be in av1/common, and in a
+      # common source list. For now this mirrors the original build system.
+      set(AOM_AV1_DECODER_SOURCES
+          ${AOM_AV1_DECODER_SOURCES}
+          "${AOM_ROOT}/av1/encoder/dct.c"
+          "${AOM_ROOT}/av1/encoder/hybrid_fwd_txfm.c"
+          "${AOM_ROOT}/av1/encoder/hybrid_fwd_txfm.h")
+
+      set(AOM_AV1_DECODER_ASM_SSE2
+          ${AOM_AV1_DECODER_ASM_SSE2}
+          "${AOM_ROOT}/av1/encoder/x86/dct_sse2.asm")
+
+      set(AOM_AV1_DECODER_INTRIN_SSE2
+          ${AOM_AV1_DECODER_INTRIN_SSE2}
+          "${AOM_ROOT}/av1/encoder/x86/dct_intrin_sse2.c")
+
+      set(AOM_AV1_DECODER_INTRIN_SSSE3
+          ${AOM_AV1_DECODER_INTRIN_SSSE3}
+          "${AOM_ROOT}/av1/encoder/x86/dct_ssse3.c")
+    endif ()
+endif ()
+
+if (CONFIG_WARPED_MOTION)
+  set(AOM_AV1_COMMON_SOURCES
+      ${AOM_AV1_COMMON_SOURCES}
+      "${AOM_ROOT}/av1/common/warped_motion.c"
+      "${AOM_ROOT}/av1/common/warped_motion.h")
+
+  set(AOM_AV1_COMMON_INTRIN_SSE2
+      ${AOM_AV1_COMMON_INTRIN_SSE2}
+      "${AOM_ROOT}/av1/common/x86/warp_plane_sse2.c")
+endif ()
+
+# Setup AV1 common/decoder/encoder targets. The libaom target must exist before
+# this function is called.
+function (setup_av1_targets)
+  add_library(aom_av1_common OBJECT ${AOM_AV1_COMMON_SOURCES})
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_av1_common)
+  target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_av1_common>)
+
+  if (CONFIG_AV1_DECODER)
+    add_library(aom_av1_decoder OBJECT ${AOM_AV1_DECODER_SOURCES})
+    set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_av1_decoder)
+    target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_av1_decoder>)
+  endif ()
+
+  if (CONFIG_AV1_ENCODER)
+    add_library(aom_av1_encoder OBJECT ${AOM_AV1_ENCODER_SOURCES})
+    set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} aom_av1_encoder)
+    target_sources(aom PUBLIC $<TARGET_OBJECTS:aom_av1_encoder>)
+  endif ()
+
+  if (HAVE_SSE2)
+    require_flag_nomsvc("-msse2" NO)
+    add_intrinsics_object_library("-msse2" "sse2" "aom_av1_common"
+                                  "AOM_AV1_COMMON_INTRIN_SSE2")
+    if (CONFIG_AV1_DECODER)
+      if (AOM_AV1_DECODER_ASM_SSE2)
+        add_asm_library("aom_av1_decoder_sse2" "AOM_AV1_DECODER_ASM_SSE2" "aom")
+      endif ()
+
+      if (AOM_AV1_DECODER_INTRIN_SSE2)
+        add_intrinsics_object_library("-msse2" "sse2" "aom_av1_decoder"
+                                      "AOM_AV1_DECODER_INTRIN_SSE2")
+      endif ()
+    endif ()
+
+    if (CONFIG_AV1_ENCODER)
+      add_asm_library("aom_av1_encoder_sse2" "AOM_AV1_ENCODER_ASM_SSE2" "aom")
+      add_intrinsics_object_library("-msse2" "sse2" "aom_av1_encoder"
+                                    "AOM_AV1_ENCODER_INTRIN_SSE2")
+    endif ()
+  endif ()
+
+  if (HAVE_SSSE3)
+    require_flag_nomsvc("-mssse3" NO)
+    add_intrinsics_object_library("-mssse3" "ssse3" "aom_av1_common"
+                                  "AOM_AV1_COMMON_INTRIN_SSSE3")
+
+    if (CONFIG_AV1_DECODER)
+      if (AOM_AV1_DECODER_INTRIN_SSSE3)
+        add_intrinsics_object_library("-mssse3" "ssse3" "aom_av1_decoder"
+                                      "AOM_AV1_DECODER_INTRIN_SSSE3")
+      endif ()
+    endif ()
+
+    if (CONFIG_AV1_ENCODER)
+      add_intrinsics_object_library("-mssse3" "ssse3" "aom_av1_encoder"
+                                    "AOM_AV1_ENCODER_INTRIN_SSSE3")
+    endif ()
+  endif ()
+
+  if (HAVE_SSE4_1)
+    require_flag_nomsvc("-msse4.1" NO)
+    add_intrinsics_object_library("-msse4.1" "sse4" "aom_av1_common"
+                                  "AOM_AV1_COMMON_INTRIN_SSE4_1")
+
+    if (CONFIG_AV1_ENCODER)
+      if ("${AOM_TARGET_CPU}" STREQUAL "x86_64")
+        add_asm_library("aom_av1_encoder_ssse3"
+                        "AOM_AV1_ENCODER_ASM_SSSE3_X86_64" "aom")
+      endif ()
+
+      if (AOM_AV1_ENCODER_INTRIN_SSE4_1)
+        add_intrinsics_object_library("-msse4.1" "sse4" "aom_av1_encoder"
+                                      "AOM_AV1_ENCODER_INTRIN_SSE4_1")
+      endif ()
+    endif ()
+  endif ()
+
+  if (HAVE_AVX2)
+    require_flag_nomsvc("-mavx2" NO)
+    add_intrinsics_object_library("-mavx2" "avx2" "aom_av1_common"
+                                  "AOM_AV1_COMMON_INTRIN_AVX2")
+
+    if (CONFIG_AV1_ENCODER)
+      add_intrinsics_object_library("-mavx2" "avx2" "aom_av1_encoder"
+                                    "AOM_AV1_ENCODER_INTRIN_AVX2")
+    endif ()
+  endif ()
+
+  if (HAVE_NEON)
+    if (AOM_AV1_COMMON_INTRIN_NEON)
+      add_intrinsics_object_library("${AOM_INTRIN_NEON_FLAG}"
+                                    "neon"
+                                    "aom_av1_common"
+                                    "AOM_AV1_COMMON_INTRIN_NEON")
+    endif ()
+
+    if (AOM_AV1_ENCODER_INTRIN_NEON)
+      add_intrinsics_object_library("${AOM_INTRIN_NEON_FLAG}"
+                                    "neon"
+                                    "aom_av1_encoder"
+                                    "AOM_AV1_ENCODER_INTRIN_NEON")
+    endif ()
+  endif ()
+
+  if (HAVE_DSPR2)
+    add_intrinsics_object_library("" "dspr2" "aom_av1_common"
+                                  "AOM_AV1_COMMON_INTRIN_DSPR2")
+  endif ()
+
+  if (HAVE_MSA)
+    add_intrinsics_object_library("" "msa" "aom_av1_common"
+                                  "AOM_AV1_COMMON_INTRIN_MSA")
+    add_intrinsics_object_library("" "msa" "aom_av1_encoder"
+                                  "AOM_AV1_ENCODER_INTRIN_MSA")
+  endif ()
+
+  # Pass the new lib targets up to the parent scope instance of
+  # $AOM_LIB_TARGETS.
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} PARENT_SCOPE)
+endfunction ()
+
+function (setup_av1_test_targets)
+endfunction ()
diff --git a/third_party/aom/av1/av1_common.mk b/third_party/aom/av1/av1_common.mk
new file mode 100644
index 0000000000..6b9a289aff
--- /dev/null
+++ b/third_party/aom/av1/av1_common.mk
@@ -0,0 +1,180 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+AV1_COMMON_SRCS-yes += av1_common.mk
+AV1_COMMON_SRCS-yes += av1_iface_common.h
+AV1_COMMON_SRCS-yes += common/alloccommon.c
+AV1_COMMON_SRCS-yes += common/av1_loopfilter.c
+AV1_COMMON_SRCS-yes += common/av1_loopfilter.h
+AV1_COMMON_SRCS-yes += common/blockd.c
+AV1_COMMON_SRCS-yes += common/debugmodes.c
+AV1_COMMON_SRCS-yes += common/entropy.c
+AV1_COMMON_SRCS-yes += common/entropymode.c
+AV1_COMMON_SRCS-yes += common/entropymv.c
+AV1_COMMON_SRCS-yes += common/frame_buffers.c
+AV1_COMMON_SRCS-yes += common/frame_buffers.h
+AV1_COMMON_SRCS-yes += common/alloccommon.h
+AV1_COMMON_SRCS-yes += common/blockd.h
+AV1_COMMON_SRCS-yes += common/common.h
+AV1_COMMON_SRCS-yes += common/entropy.h
+AV1_COMMON_SRCS-yes += common/entropymode.h
+AV1_COMMON_SRCS-yes += common/entropymv.h
+AV1_COMMON_SRCS-yes += common/enums.h
+AV1_COMMON_SRCS-yes += common/filter.h
+AV1_COMMON_SRCS-yes += common/filter.c
+AV1_COMMON_SRCS-yes += common/idct.h
+AV1_COMMON_SRCS-yes += common/idct.c
+AV1_COMMON_SRCS-yes += common/thread_common.h
+AV1_COMMON_SRCS-$(CONFIG_LV_MAP) += common/txb_common.h
+AV1_COMMON_SRCS-$(CONFIG_LV_MAP) += common/txb_common.c
+AV1_COMMON_SRCS-yes += common/mv.h
+AV1_COMMON_SRCS-yes += common/onyxc_int.h
+AV1_COMMON_SRCS-yes += common/pred_common.h
+AV1_COMMON_SRCS-yes += common/pred_common.c
+AV1_COMMON_SRCS-yes += common/quant_common.h
+AV1_COMMON_SRCS-yes += common/reconinter.h
+AV1_COMMON_SRCS-yes += common/reconintra.h
+AV1_COMMON_SRCS-yes += common/av1_rtcd.c
+AV1_COMMON_SRCS-yes += common/av1_rtcd_defs.pl
+AV1_COMMON_SRCS-yes += common/scale.h
+AV1_COMMON_SRCS-yes += common/scale.c
+AV1_COMMON_SRCS-yes += common/seg_common.h
+AV1_COMMON_SRCS-yes += common/seg_common.c
+AV1_COMMON_SRCS-yes += common/tile_common.h
+AV1_COMMON_SRCS-yes += common/tile_common.c
+AV1_COMMON_SRCS-yes += common/thread_common.c
+AV1_COMMON_SRCS-yes += common/mvref_common.c
+AV1_COMMON_SRCS-yes += common/mvref_common.h
+AV1_COMMON_SRCS-yes += common/quant_common.c
+AV1_COMMON_SRCS-yes += common/reconinter.c
+AV1_COMMON_SRCS-yes += common/reconintra.c
+AV1_COMMON_SRCS-yes += common/resize.c
+AV1_COMMON_SRCS-yes += common/resize.h
+AV1_COMMON_SRCS-yes += common/restoration.h
+AV1_COMMON_SRCS-yes += common/common_data.h
+AV1_COMMON_SRCS-yes += common/scan.c
+AV1_COMMON_SRCS-yes += common/scan.h
+# TODO(angiebird) the forward transform belongs under encoder/
+AV1_COMMON_SRCS-yes += common/av1_txfm.h
+AV1_COMMON_SRCS-yes += common/av1_fwd_txfm1d.h
+AV1_COMMON_SRCS-yes += common/av1_fwd_txfm1d.c
+AV1_COMMON_SRCS-yes += common/av1_inv_txfm1d.h
+AV1_COMMON_SRCS-yes += common/av1_inv_txfm1d.c
+AV1_COMMON_SRCS-yes += common/av1_fwd_txfm2d.c
+AV1_COMMON_SRCS-yes += common/av1_fwd_txfm2d_cfg.h
+AV1_COMMON_SRCS-yes += common/av1_inv_txfm2d.c
+AV1_COMMON_SRCS-yes += common/av1_inv_txfm2d_cfg.h
+AV1_COMMON_SRCS-$(HAVE_SSSE3) += common/x86/av1_convolve_ssse3.c
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/av1_highbd_convolve_sse4.c
+endif
+AV1_COMMON_SRCS-yes += common/convolve.c
+AV1_COMMON_SRCS-yes += common/convolve.h
+ifeq ($(CONFIG_LOOP_RESTORATION),yes)
+AV1_COMMON_SRCS-yes += common/restoration.h
+AV1_COMMON_SRCS-yes += common/restoration.c
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/selfguided_sse4.c
+endif
+ifeq (yes,$(filter $(CONFIG_GLOBAL_MOTION) $(CONFIG_WARPED_MOTION),yes))
+AV1_COMMON_SRCS-yes += common/warped_motion.h
+AV1_COMMON_SRCS-yes += common/warped_motion.c
+endif
+ifeq ($(CONFIG_CDEF),yes)
+AV1_COMMON_SRCS-yes += common/clpf.c
+AV1_COMMON_SRCS-yes += common/clpf.h
+AV1_COMMON_SRCS-yes += common/clpf_simd.h
+AV1_COMMON_SRCS-yes += common/cdef_simd.h
+AV1_COMMON_SRCS-$(HAVE_SSE2) += common/clpf_sse2.c
+AV1_COMMON_SRCS-$(HAVE_SSSE3) += common/clpf_ssse3.c
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/clpf_sse4.c
+AV1_COMMON_SRCS-$(HAVE_NEON) += common/clpf_neon.c
+AV1_COMMON_SRCS-$(HAVE_SSE2) += common/od_dering_sse2.c
+AV1_COMMON_SRCS-$(HAVE_SSSE3) += common/od_dering_ssse3.c
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/od_dering_sse4.c
+AV1_COMMON_SRCS-$(HAVE_NEON) += common/od_dering_neon.c
+AV1_COMMON_SRCS-yes += common/od_dering.c
+AV1_COMMON_SRCS-yes += common/od_dering.h
+AV1_COMMON_SRCS-yes += common/od_dering_simd.h
+AV1_COMMON_SRCS-yes += common/cdef.c
+AV1_COMMON_SRCS-yes += common/cdef.h
+endif
+AV1_COMMON_SRCS-yes += common/odintrin.c
+AV1_COMMON_SRCS-yes += common/odintrin.h
+
+ifeq ($(CONFIG_CFL),yes)
+AV1_COMMON_SRCS-yes += common/cfl.h
+AV1_COMMON_SRCS-yes += common/cfl.c
+endif
+
+ifeq ($(CONFIG_PVQ),yes)
+# PVQ from daala
+AV1_COMMON_SRCS-yes += common/pvq.c
+AV1_COMMON_SRCS-yes += common/partition.c
+AV1_COMMON_SRCS-yes += common/partition.h
+AV1_COMMON_SRCS-yes += common/zigzag4.c
+AV1_COMMON_SRCS-yes += common/zigzag8.c
+AV1_COMMON_SRCS-yes += common/zigzag16.c
+AV1_COMMON_SRCS-yes += common/zigzag32.c
+AV1_COMMON_SRCS-yes += common/zigzag.h
+AV1_COMMON_SRCS-yes += common/generic_code.c
+AV1_COMMON_SRCS-yes += common/pvq_state.c
+AV1_COMMON_SRCS-yes += common/laplace_tables.c
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/pvq_sse4.c
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/pvq_sse4.h
+endif
+ifneq ($(findstring yes,$(CONFIG_PVQ)$(CONFIG_DAALA_DIST)$(CONFIG_XIPHRC)),)
+AV1_COMMON_SRCS-yes += common/pvq.h
+AV1_COMMON_SRCS-yes += common/pvq_state.h
+AV1_COMMON_SRCS-yes += common/generic_code.h
+endif
+
+ifneq ($(CONFIG_HIGHBITDEPTH),yes)
+AV1_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/av1_itrans4_dspr2.c
+AV1_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/av1_itrans8_dspr2.c
+AV1_COMMON_SRCS-$(HAVE_DSPR2)  += common/mips/dspr2/av1_itrans16_dspr2.c
+endif
+
+# common (msa)
+AV1_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/av1_idct4x4_msa.c
+AV1_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/av1_idct8x8_msa.c
+AV1_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/av1_idct16x16_msa.c
+
+AV1_COMMON_SRCS-$(HAVE_SSE2) += common/x86/idct_intrin_sse2.c
+AV1_COMMON_SRCS-$(HAVE_AVX2) += common/x86/hybrid_inv_txfm_avx2.c
+
+ifeq ($(CONFIG_AV1_ENCODER),yes)
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/av1_txfm1d_sse4.h
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/av1_fwd_txfm1d_sse4.c
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/av1_fwd_txfm2d_sse4.c
+endif
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/highbd_txfm_utility_sse4.h
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/highbd_inv_txfm_sse4.c
+AV1_COMMON_SRCS-$(HAVE_AVX2) += common/x86/highbd_inv_txfm_avx2.c
+endif
+
+ifneq ($(CONFIG_HIGHBITDEPTH),yes)
+AV1_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/iht4x4_add_neon.c
+AV1_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/iht8x8_add_neon.c
+endif
+
+ifeq ($(CONFIG_FILTER_INTRA),yes)
+AV1_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/filterintra_sse4.c
+endif
+
+ifneq ($(findstring yes,$(CONFIG_GLOBAL_MOTION) $(CONFIG_WARPED_MOTION)),)
+AV1_COMMON_SRCS-$(HAVE_SSE2) += common/x86/warp_plane_sse2.c
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+AV1_COMMON_SRCS-$(HAVE_SSSE3) += common/x86/highbd_warp_plane_ssse3.c
+endif
+endif
+
+$(eval $(call rtcd_h_template,av1_rtcd,av1/common/av1_rtcd_defs.pl))
diff --git a/third_party/aom/av1/av1_cx.mk b/third_party/aom/av1/av1_cx.mk
new file mode 100644
index 0000000000..0a0d770ce2
--- /dev/null
+++ b/third_party/aom/av1/av1_cx.mk
@@ -0,0 +1,165 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+AV1_CX_EXPORTS += exports_enc
+
+AV1_CX_SRCS-yes += $(AV1_COMMON_SRCS-yes)
+AV1_CX_SRCS-no  += $(AV1_COMMON_SRCS-no)
+AV1_CX_SRCS_REMOVE-yes += $(AV1_COMMON_SRCS_REMOVE-yes)
+AV1_CX_SRCS_REMOVE-no  += $(AV1_COMMON_SRCS_REMOVE-no)
+
+AV1_CX_SRCS-yes += av1_cx_iface.c
+
+AV1_CX_SRCS-yes += encoder/av1_quantize.c
+AV1_CX_SRCS-yes += encoder/av1_quantize.h
+AV1_CX_SRCS-yes += encoder/bitstream.c
+AV1_CX_SRCS-yes += encoder/context_tree.c
+AV1_CX_SRCS-yes += encoder/context_tree.h
+AV1_CX_SRCS-yes += encoder/variance_tree.c
+AV1_CX_SRCS-yes += encoder/variance_tree.h
+AV1_CX_SRCS-yes += encoder/cost.h
+AV1_CX_SRCS-yes += encoder/cost.c
+AV1_CX_SRCS-yes += encoder/dct.c
+AV1_CX_SRCS-yes += encoder/hybrid_fwd_txfm.c
+AV1_CX_SRCS-yes += encoder/hybrid_fwd_txfm.h
+AV1_CX_SRCS-yes += encoder/encodeframe.c
+AV1_CX_SRCS-yes += encoder/encodeframe.h
+AV1_CX_SRCS-yes += encoder/encodemb.c
+AV1_CX_SRCS-yes += encoder/encodemv.c
+AV1_CX_SRCS-yes += encoder/ethread.h
+AV1_CX_SRCS-yes += encoder/ethread.c
+AV1_CX_SRCS-yes += encoder/extend.c
+AV1_CX_SRCS-yes += encoder/firstpass.c
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += ../third_party/fastfeat/fast.h
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += ../third_party/fastfeat/nonmax.c
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += ../third_party/fastfeat/fast_9.c
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += ../third_party/fastfeat/fast.c
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += encoder/corner_match.c
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += encoder/corner_match.h
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += encoder/corner_detect.c
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += encoder/corner_detect.h
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += encoder/global_motion.c
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += encoder/global_motion.h
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += encoder/ransac.c
+AV1_CX_SRCS-$(CONFIG_GLOBAL_MOTION) += encoder/ransac.h
+AV1_CX_SRCS-yes += encoder/block.h
+AV1_CX_SRCS-yes += encoder/bitstream.h
+AV1_CX_SRCS-yes += encoder/encodemb.h
+AV1_CX_SRCS-yes += encoder/encodemv.h
+AV1_CX_SRCS-$(CONFIG_LV_MAP) += encoder/encodetxb.c
+AV1_CX_SRCS-$(CONFIG_LV_MAP) += encoder/encodetxb.h
+AV1_CX_SRCS-yes += encoder/extend.h
+AV1_CX_SRCS-yes += encoder/firstpass.h
+AV1_CX_SRCS-yes += encoder/lookahead.c
+AV1_CX_SRCS-yes += encoder/lookahead.h
+AV1_CX_SRCS-yes += encoder/mcomp.h
+AV1_CX_SRCS-yes += encoder/encoder.h
+AV1_CX_SRCS-yes += encoder/ratectrl.h
+ifeq ($(CONFIG_XIPHRC),yes)
+AV1_CX_SRCS-yes += encoder/ratectrl_xiph.h
+endif
+AV1_CX_SRCS-yes += encoder/rd.h
+AV1_CX_SRCS-yes += encoder/rdopt.h
+AV1_CX_SRCS-yes += encoder/tokenize.h
+AV1_CX_SRCS-yes += encoder/treewriter.h
+AV1_CX_SRCS-yes += encoder/mcomp.c
+AV1_CX_SRCS-yes += encoder/encoder.c
+ifeq ($(CONFIG_PALETTE),yes)
+AV1_CX_SRCS-yes += encoder/palette.h
+AV1_CX_SRCS-yes += encoder/palette.c
+endif
+AV1_CX_SRCS-yes += encoder/picklpf.c
+AV1_CX_SRCS-yes += encoder/picklpf.h
+AV1_CX_SRCS-$(CONFIG_LOOP_RESTORATION) += encoder/pickrst.c
+AV1_CX_SRCS-$(CONFIG_LOOP_RESTORATION) += encoder/pickrst.h
+AV1_CX_SRCS-yes += encoder/ratectrl.c
+ifeq ($(CONFIG_XIPHRC),yes)
+AV1_CX_SRCS-yes += encoder/ratectrl_xiph.c
+endif
+AV1_CX_SRCS-yes += encoder/rd.c
+AV1_CX_SRCS-yes += encoder/rdopt.c
+AV1_CX_SRCS-yes += encoder/segmentation.c
+AV1_CX_SRCS-yes += encoder/segmentation.h
+AV1_CX_SRCS-yes += encoder/speed_features.c
+AV1_CX_SRCS-yes += encoder/speed_features.h
+AV1_CX_SRCS-yes += encoder/subexp.c
+AV1_CX_SRCS-yes += encoder/subexp.h
+AV1_CX_SRCS-$(CONFIG_INTERNAL_STATS) += encoder/blockiness.c
+
+AV1_CX_SRCS-yes += encoder/tokenize.c
+AV1_CX_SRCS-yes += encoder/treewriter.c
+AV1_CX_SRCS-yes += encoder/aq_variance.c
+AV1_CX_SRCS-yes += encoder/aq_variance.h
+AV1_CX_SRCS-yes += encoder/aq_cyclicrefresh.c
+AV1_CX_SRCS-yes += encoder/aq_cyclicrefresh.h
+AV1_CX_SRCS-yes += encoder/aq_complexity.c
+AV1_CX_SRCS-yes += encoder/aq_complexity.h
+AV1_CX_SRCS-yes += encoder/temporal_filter.c
+AV1_CX_SRCS-yes += encoder/temporal_filter.h
+AV1_CX_SRCS-yes += encoder/mbgraph.c
+AV1_CX_SRCS-yes += encoder/mbgraph.h
+ifeq ($(CONFIG_CDEF),yes)
+AV1_CX_SRCS-yes += encoder/pickcdef.c
+endif
+ifeq ($(CONFIG_PVQ),yes)
+# PVQ from daala
+AV1_CX_SRCS-yes += encoder/daala_compat_enc.c
+AV1_CX_SRCS-yes += encoder/pvq_encoder.c
+AV1_CX_SRCS-yes += encoder/pvq_encoder.h
+AV1_CX_SRCS-yes += encoder/generic_encoder.c
+AV1_CX_SRCS-yes += encoder/laplace_encoder.c
+endif
+ifneq ($(findstring yes,$(CONFIG_XIPHRC)$(CONFIG_PVQ)),)
+AV1_CX_SRCS-yes += encoder/encint.h
+endif
+
+AV1_CX_SRCS-$(HAVE_SSE2) += encoder/x86/av1_quantize_sse2.c
+AV1_CX_SRCS-$(HAVE_SSE2) += encoder/x86/temporal_filter_apply_sse2.asm
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+AV1_CX_SRCS-$(HAVE_SSE2) += encoder/x86/highbd_block_error_intrin_sse2.c
+endif
+
+AV1_CX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_sse2.asm
+AV1_CX_SRCS-$(HAVE_SSE2) += encoder/x86/error_sse2.asm
+
+ifeq ($(ARCH_X86_64),yes)
+AV1_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/av1_quantize_ssse3_x86_64.asm
+endif
+
+AV1_CX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_intrin_sse2.c
+AV1_CX_SRCS-$(HAVE_SSSE3) += encoder/x86/dct_ssse3.c
+AV1_CX_SRCS-$(HAVE_AVX2) += encoder/x86/hybrid_fwd_txfm_avx2.c
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+AV1_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/av1_highbd_quantize_sse4.c
+AV1_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/highbd_fwd_txfm_sse4.c
+endif
+
+ifeq ($(CONFIG_EXT_INTER),yes)
+AV1_CX_SRCS-yes += encoder/wedge_utils.c
+AV1_CX_SRCS-$(HAVE_SSE2) += encoder/x86/wedge_utils_sse2.c
+endif
+
+AV1_CX_SRCS-$(HAVE_AVX2) += encoder/x86/error_intrin_avx2.c
+
+ifneq ($(CONFIG_HIGHBITDEPTH),yes)
+AV1_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/dct_neon.c
+AV1_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/error_neon.c
+endif
+AV1_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/quantize_neon.c
+
+AV1_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/error_msa.c
+AV1_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/fdct4x4_msa.c
+AV1_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/fdct8x8_msa.c
+AV1_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/fdct16x16_msa.c
+AV1_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/fdct_msa.h
+AV1_CX_SRCS-$(HAVE_MSA) += encoder/mips/msa/temporal_filter_msa.c
+
+AV1_CX_SRCS-yes := $(filter-out $(AV1_CX_SRCS_REMOVE-yes),$(AV1_CX_SRCS-yes))
diff --git a/third_party/aom/av1/av1_cx_iface.c b/third_party/aom/av1/av1_cx_iface.c
new file mode 100644
index 0000000000..d4832a15c2
--- /dev/null
+++ b/third_party/aom/av1/av1_cx_iface.c
@@ -0,0 +1,1605 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "./aom_config.h"
+#include "aom/aom_encoder.h"
+#include "aom_ports/aom_once.h"
+#include "aom_ports/system_state.h"
+#include "aom/internal/aom_codec_internal.h"
+#include "./aom_version.h"
+#include "av1/encoder/encoder.h"
+#include "aom/aomcx.h"
+#include "av1/encoder/firstpass.h"
+#include "av1/av1_iface_common.h"
+
+struct av1_extracfg {
+  int cpu_used;  // available cpu percentage in 1/16
+  unsigned int enable_auto_alt_ref;
+#if CONFIG_EXT_REFS
+  unsigned int enable_auto_bwd_ref;
+#endif  // CONFIG_EXT_REFS
+  unsigned int noise_sensitivity;
+  unsigned int sharpness;
+  unsigned int static_thresh;
+  unsigned int tile_columns;
+  unsigned int tile_rows;
+#if CONFIG_DEPENDENT_HORZTILES
+  unsigned int dependent_horz_tiles;
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  unsigned int loop_filter_across_tiles_enabled;
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+  unsigned int arnr_max_frames;
+  unsigned int arnr_strength;
+  unsigned int min_gf_interval;
+  unsigned int max_gf_interval;
+  aom_tune_metric tuning;
+  unsigned int cq_level;  // constrained quality level
+  unsigned int rc_max_intra_bitrate_pct;
+  unsigned int rc_max_inter_bitrate_pct;
+  unsigned int gf_cbr_boost_pct;
+  unsigned int lossless;
+#if CONFIG_AOM_QM
+  unsigned int enable_qm;
+  unsigned int qm_min;
+  unsigned int qm_max;
+#endif
+#if CONFIG_TILE_GROUPS
+  unsigned int num_tg;
+  unsigned int mtu_size;
+#endif
+#if CONFIG_TEMPMV_SIGNALING
+  unsigned int disable_tempmv;
+#endif
+  unsigned int frame_parallel_decoding_mode;
+  AQ_MODE aq_mode;
+#if CONFIG_EXT_DELTA_Q
+  DELTAQ_MODE deltaq_mode;
+#endif
+  unsigned int frame_periodic_boost;
+  aom_bit_depth_t bit_depth;
+  aom_tune_content content;
+  aom_color_space_t color_space;
+  int color_range;
+  int render_width;
+  int render_height;
+  aom_superblock_size_t superblock_size;
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  int ans_window_size_log2;
+#endif
+#if CONFIG_EXT_TILE
+  unsigned int tile_encoding_mode;
+#endif  // CONFIG_EXT_TILE
+
+  unsigned int motion_vector_unit_test;
+};
+
+static struct av1_extracfg default_extra_cfg = {
+  0,  // cpu_used
+  1,  // enable_auto_alt_ref
+#if CONFIG_EXT_REFS
+  0,    // enable_auto_bwd_ref
+#endif  // CONFIG_EXT_REFS
+  0,    // noise_sensitivity
+  0,    // sharpness
+  0,    // static_thresh
+#if CONFIG_EXT_TILE
+  UINT_MAX,  // tile_columns
+  UINT_MAX,  // tile_rows
+#else
+  0,  // tile_columns
+  0,  // tile_rows
+#endif  // CONFIG_EXT_TILE
+#if CONFIG_DEPENDENT_HORZTILES
+  0,  // Depdendent Horizontal tiles
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  1,              // loop_filter_across_tiles_enabled
+#endif            // CONFIG_LOOPFILTERING_ACROSS_TILES
+  7,              // arnr_max_frames
+  5,              // arnr_strength
+  0,              // min_gf_interval; 0 -> default decision
+  0,              // max_gf_interval; 0 -> default decision
+  AOM_TUNE_PSNR,  // tuning
+  10,             // cq_level
+  0,              // rc_max_intra_bitrate_pct
+  0,              // rc_max_inter_bitrate_pct
+  0,              // gf_cbr_boost_pct
+  0,              // lossless
+#if CONFIG_AOM_QM
+  0,                 // enable_qm
+  DEFAULT_QM_FIRST,  // qm_min
+  DEFAULT_QM_LAST,   // qm_max
+#endif
+#if CONFIG_TILE_GROUPS
+  1,  // max number of tile groups
+  0,  // mtu_size
+#endif
+#if CONFIG_TEMPMV_SIGNALING
+  0,  // disable temporal mv prediction
+#endif
+  1,      // frame_parallel_decoding_mode
+  NO_AQ,  // aq_mode
+#if CONFIG_EXT_DELTA_Q
+  NO_DELTA_Q,  // deltaq_mode
+#endif
+  CONFIG_XIPHRC,                // frame_periodic_delta_q
+  AOM_BITS_8,                   // Bit depth
+  AOM_CONTENT_DEFAULT,          // content
+  AOM_CS_UNKNOWN,               // color space
+  0,                            // color range
+  0,                            // render width
+  0,                            // render height
+  AOM_SUPERBLOCK_SIZE_DYNAMIC,  // superblock_size
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  23,  // ans_window_size_log2
+#endif
+#if CONFIG_EXT_TILE
+  0,    // Tile encoding mode is TILE_NORMAL by default.
+#endif  // CONFIG_EXT_TILE
+
+  0,  // motion_vector_unit_test
+};
+
+struct aom_codec_alg_priv {
+  aom_codec_priv_t base;
+  aom_codec_enc_cfg_t cfg;
+  struct av1_extracfg extra_cfg;
+  AV1EncoderConfig oxcf;
+  AV1_COMP *cpi;
+  unsigned char *cx_data;
+  size_t cx_data_sz;
+  unsigned char *pending_cx_data;
+  size_t pending_cx_data_sz;
+  int pending_frame_count;
+  size_t pending_frame_sizes[8];
+  aom_image_t preview_img;
+  aom_enc_frame_flags_t next_frame_flags;
+  aom_postproc_cfg_t preview_ppcfg;
+  aom_codec_pkt_list_decl(256) pkt_list;
+  unsigned int fixed_kf_cntr;
+  // BufferPool that holds all reference frames.
+  BufferPool *buffer_pool;
+};
+
+static aom_codec_err_t update_error_state(
+    aom_codec_alg_priv_t *ctx, const struct aom_internal_error_info *error) {
+  const aom_codec_err_t res = error->error_code;
+
+  if (res != AOM_CODEC_OK)
+    ctx->base.err_detail = error->has_detail ? error->detail : NULL;
+
+  return res;
+}
+
+#undef ERROR
+#define ERROR(str)                  \
+  do {                              \
+    ctx->base.err_detail = str;     \
+    return AOM_CODEC_INVALID_PARAM; \
+  } while (0)
+
+#define RANGE_CHECK(p, memb, lo, hi)                   \
+  do {                                                 \
+    if (!((p)->memb >= (lo) && (p)->memb <= (hi)))     \
+      ERROR(#memb " out of range [" #lo ".." #hi "]"); \
+  } while (0)
+
+#define RANGE_CHECK_HI(p, memb, hi)                                     \
+  do {                                                                  \
+    if (!((p)->memb <= (hi))) ERROR(#memb " out of range [.." #hi "]"); \
+  } while (0)
+
+#define RANGE_CHECK_LO(p, memb, lo)                                     \
+  do {                                                                  \
+    if (!((p)->memb >= (lo))) ERROR(#memb " out of range [" #lo "..]"); \
+  } while (0)
+
+#define RANGE_CHECK_BOOL(p, memb)                                     \
+  do {                                                                \
+    if (!!((p)->memb) != (p)->memb) ERROR(#memb " expected boolean"); \
+  } while (0)
+
+static aom_codec_err_t validate_config(aom_codec_alg_priv_t *ctx,
+                                       const aom_codec_enc_cfg_t *cfg,
+                                       const struct av1_extracfg *extra_cfg) {
+  RANGE_CHECK(cfg, g_w, 1, 65535);  // 16 bits available
+  RANGE_CHECK(cfg, g_h, 1, 65535);  // 16 bits available
+  RANGE_CHECK(cfg, g_timebase.den, 1, 1000000000);
+  RANGE_CHECK(cfg, g_timebase.num, 1, cfg->g_timebase.den);
+  RANGE_CHECK_HI(cfg, g_profile, 3);
+
+  RANGE_CHECK_HI(cfg, rc_max_quantizer, 63);
+  RANGE_CHECK_HI(cfg, rc_min_quantizer, cfg->rc_max_quantizer);
+  RANGE_CHECK_BOOL(extra_cfg, lossless);
+  RANGE_CHECK(extra_cfg, aq_mode, 0, AQ_MODE_COUNT - 1);
+#if CONFIG_EXT_DELTA_Q
+  RANGE_CHECK(extra_cfg, deltaq_mode, 0, DELTAQ_MODE_COUNT - 1);
+#endif
+  RANGE_CHECK_HI(extra_cfg, frame_periodic_boost, 1);
+  RANGE_CHECK_HI(cfg, g_threads, 64);
+  RANGE_CHECK_HI(cfg, g_lag_in_frames, MAX_LAG_BUFFERS);
+  RANGE_CHECK(cfg, rc_end_usage, AOM_VBR, AOM_Q);
+  RANGE_CHECK_HI(cfg, rc_undershoot_pct, 100);
+  RANGE_CHECK_HI(cfg, rc_overshoot_pct, 100);
+  RANGE_CHECK_HI(cfg, rc_2pass_vbr_bias_pct, 100);
+  RANGE_CHECK(cfg, kf_mode, AOM_KF_DISABLED, AOM_KF_AUTO);
+  RANGE_CHECK_BOOL(cfg, rc_resize_allowed);
+  RANGE_CHECK_HI(cfg, rc_dropframe_thresh, 100);
+  RANGE_CHECK_HI(cfg, rc_resize_up_thresh, 100);
+  RANGE_CHECK_HI(cfg, rc_resize_down_thresh, 100);
+  RANGE_CHECK(cfg, g_pass, AOM_RC_ONE_PASS, AOM_RC_LAST_PASS);
+  RANGE_CHECK_HI(extra_cfg, min_gf_interval, MAX_LAG_BUFFERS - 1);
+  RANGE_CHECK_HI(extra_cfg, max_gf_interval, MAX_LAG_BUFFERS - 1);
+  if (extra_cfg->max_gf_interval > 0) {
+    RANGE_CHECK(extra_cfg, max_gf_interval, 2, (MAX_LAG_BUFFERS - 1));
+  }
+  if (extra_cfg->min_gf_interval > 0 && extra_cfg->max_gf_interval > 0) {
+    RANGE_CHECK(extra_cfg, max_gf_interval, extra_cfg->min_gf_interval,
+                (MAX_LAG_BUFFERS - 1));
+  }
+
+  if (cfg->rc_resize_allowed == 1) {
+    RANGE_CHECK_HI(cfg, rc_scaled_width, cfg->g_w);
+    RANGE_CHECK_HI(cfg, rc_scaled_height, cfg->g_h);
+  }
+
+  // AV1 does not support a lower bound on the keyframe interval in
+  // automatic keyframe placement mode.
+  if (cfg->kf_mode != AOM_KF_DISABLED && cfg->kf_min_dist != cfg->kf_max_dist &&
+      cfg->kf_min_dist > 0)
+    ERROR(
+        "kf_min_dist not supported in auto mode, use 0 "
+        "or kf_max_dist instead.");
+
+  RANGE_CHECK_HI(extra_cfg, motion_vector_unit_test, 2);
+  RANGE_CHECK_HI(extra_cfg, enable_auto_alt_ref, 2);
+#if CONFIG_EXT_REFS
+  RANGE_CHECK_HI(extra_cfg, enable_auto_bwd_ref, 2);
+#endif  // CONFIG_EXT_REFS
+  RANGE_CHECK(extra_cfg, cpu_used, 0, 8);
+  RANGE_CHECK_HI(extra_cfg, noise_sensitivity, 6);
+  RANGE_CHECK(extra_cfg, superblock_size, AOM_SUPERBLOCK_SIZE_64X64,
+              AOM_SUPERBLOCK_SIZE_DYNAMIC);
+#if CONFIG_EXT_TILE
+// TODO(any): Waring. If CONFIG_EXT_TILE is true, tile_columns really
+// means tile_width, and tile_rows really means tile_hight. The interface
+// should be sanitized.
+#if CONFIG_EXT_PARTITION
+  if (extra_cfg->superblock_size != AOM_SUPERBLOCK_SIZE_64X64) {
+    if (extra_cfg->tile_columns != UINT_MAX)
+      RANGE_CHECK(extra_cfg, tile_columns, 1, 32);
+    if (extra_cfg->tile_rows != UINT_MAX)
+      RANGE_CHECK(extra_cfg, tile_rows, 1, 32);
+  } else
+#endif  // CONFIG_EXT_PARTITION
+  {
+    if (extra_cfg->tile_columns != UINT_MAX)
+      RANGE_CHECK(extra_cfg, tile_columns, 1, 64);
+    if (extra_cfg->tile_rows != UINT_MAX)
+      RANGE_CHECK(extra_cfg, tile_rows, 1, 64);
+  }
+  RANGE_CHECK_HI(extra_cfg, tile_encoding_mode, 1);
+#else
+  RANGE_CHECK_HI(extra_cfg, tile_columns, 6);
+  RANGE_CHECK_HI(extra_cfg, tile_rows, 2);
+#endif  // CONFIG_EXT_TILE
+#if CONFIG_DEPENDENT_HORZTILES
+  RANGE_CHECK_HI(extra_cfg, dependent_horz_tiles, 1);
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  RANGE_CHECK_HI(extra_cfg, loop_filter_across_tiles_enabled, 1);
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+  RANGE_CHECK_HI(extra_cfg, sharpness, 7);
+  RANGE_CHECK_HI(extra_cfg, arnr_max_frames, 15);
+  RANGE_CHECK_HI(extra_cfg, arnr_strength, 6);
+  RANGE_CHECK_HI(extra_cfg, cq_level, 63);
+  RANGE_CHECK(cfg, g_bit_depth, AOM_BITS_8, AOM_BITS_12);
+  RANGE_CHECK(cfg, g_input_bit_depth, 8, 12);
+  RANGE_CHECK(extra_cfg, content, AOM_CONTENT_DEFAULT, AOM_CONTENT_INVALID - 1);
+
+  // TODO(yaowu): remove this when ssim tuning is implemented for av1
+  if (extra_cfg->tuning == AOM_TUNE_SSIM)
+    ERROR("Option --tune=ssim is not currently supported in AV1.");
+
+  if (cfg->g_pass == AOM_RC_LAST_PASS) {
+#if !CONFIG_XIPHRC
+    const size_t packet_sz = sizeof(FIRSTPASS_STATS);
+    const int n_packets = (int)(cfg->rc_twopass_stats_in.sz / packet_sz);
+    const FIRSTPASS_STATS *stats;
+#endif
+
+    if (cfg->rc_twopass_stats_in.buf == NULL)
+      ERROR("rc_twopass_stats_in.buf not set.");
+
+#if !CONFIG_XIPHRC
+    if (cfg->rc_twopass_stats_in.sz % packet_sz)
+      ERROR("rc_twopass_stats_in.sz indicates truncated packet.");
+
+    if (cfg->rc_twopass_stats_in.sz < 2 * packet_sz)
+      ERROR("rc_twopass_stats_in requires at least two packets.");
+
+    stats =
+        (const FIRSTPASS_STATS *)cfg->rc_twopass_stats_in.buf + n_packets - 1;
+
+    if ((int)(stats->count + 0.5) != n_packets - 1)
+      ERROR("rc_twopass_stats_in missing EOS stats packet");
+#endif
+  }
+
+#if !CONFIG_HIGHBITDEPTH
+  if (cfg->g_profile > (unsigned int)PROFILE_1) {
+    ERROR("Profile > 1 not supported in this build configuration");
+  }
+#endif
+  if (cfg->g_profile <= (unsigned int)PROFILE_1 &&
+      cfg->g_bit_depth > AOM_BITS_8) {
+    ERROR("Codec high bit-depth not supported in profile < 2");
+  }
+  if (cfg->g_profile <= (unsigned int)PROFILE_1 && cfg->g_input_bit_depth > 8) {
+    ERROR("Source high bit-depth not supported in profile < 2");
+  }
+  if (cfg->g_profile > (unsigned int)PROFILE_1 &&
+      cfg->g_bit_depth == AOM_BITS_8) {
+    ERROR("Codec bit-depth 8 not supported in profile > 1");
+  }
+  RANGE_CHECK(extra_cfg, color_space, AOM_CS_UNKNOWN, AOM_CS_SRGB);
+  RANGE_CHECK(extra_cfg, color_range, 0, 1);
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  RANGE_CHECK(extra_cfg, ans_window_size_log2, 8, 23);
+#endif
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t validate_img(aom_codec_alg_priv_t *ctx,
+                                    const aom_image_t *img) {
+  switch (img->fmt) {
+    case AOM_IMG_FMT_YV12:
+    case AOM_IMG_FMT_I420:
+    case AOM_IMG_FMT_I42016: break;
+    case AOM_IMG_FMT_I422:
+    case AOM_IMG_FMT_I444:
+    case AOM_IMG_FMT_I440:
+      if (ctx->cfg.g_profile != (unsigned int)PROFILE_1) {
+        ERROR(
+            "Invalid image format. I422, I444, I440 images are "
+            "not supported in profile.");
+      }
+      break;
+    case AOM_IMG_FMT_I42216:
+    case AOM_IMG_FMT_I44416:
+    case AOM_IMG_FMT_I44016:
+      if (ctx->cfg.g_profile != (unsigned int)PROFILE_1 &&
+          ctx->cfg.g_profile != (unsigned int)PROFILE_3) {
+        ERROR(
+            "Invalid image format. 16-bit I422, I444, I440 images are "
+            "not supported in profile.");
+      }
+      break;
+    default:
+      ERROR(
+          "Invalid image format. Only YV12, I420, I422, I444 images are "
+          "supported.");
+      break;
+  }
+
+  if (img->d_w != ctx->cfg.g_w || img->d_h != ctx->cfg.g_h)
+    ERROR("Image size must match encoder init configuration size");
+
+  return AOM_CODEC_OK;
+}
+
+static int get_image_bps(const aom_image_t *img) {
+  switch (img->fmt) {
+    case AOM_IMG_FMT_YV12:
+    case AOM_IMG_FMT_I420: return 12;
+    case AOM_IMG_FMT_I422: return 16;
+    case AOM_IMG_FMT_I444: return 24;
+    case AOM_IMG_FMT_I440: return 16;
+    case AOM_IMG_FMT_I42016: return 24;
+    case AOM_IMG_FMT_I42216: return 32;
+    case AOM_IMG_FMT_I44416: return 48;
+    case AOM_IMG_FMT_I44016: return 32;
+    default: assert(0 && "Invalid image format"); break;
+  }
+  return 0;
+}
+
+static aom_codec_err_t set_encoder_config(
+    AV1EncoderConfig *oxcf, const aom_codec_enc_cfg_t *cfg,
+    const struct av1_extracfg *extra_cfg) {
+  const int is_vbr = cfg->rc_end_usage == AOM_VBR;
+  oxcf->profile = cfg->g_profile;
+  oxcf->max_threads = (int)cfg->g_threads;
+  oxcf->width = cfg->g_w;
+  oxcf->height = cfg->g_h;
+  oxcf->bit_depth = cfg->g_bit_depth;
+  oxcf->input_bit_depth = cfg->g_input_bit_depth;
+  // guess a frame rate if out of whack, use 30
+  oxcf->init_framerate = (double)cfg->g_timebase.den / cfg->g_timebase.num;
+  if (oxcf->init_framerate > 180) oxcf->init_framerate = 30;
+
+  oxcf->mode = GOOD;
+
+  switch (cfg->g_pass) {
+    case AOM_RC_ONE_PASS: oxcf->pass = 0; break;
+    case AOM_RC_FIRST_PASS: oxcf->pass = 1; break;
+    case AOM_RC_LAST_PASS: oxcf->pass = 2; break;
+  }
+
+  oxcf->lag_in_frames =
+      cfg->g_pass == AOM_RC_FIRST_PASS ? 0 : cfg->g_lag_in_frames;
+  oxcf->rc_mode = cfg->rc_end_usage;
+
+  // Convert target bandwidth from Kbit/s to Bit/s
+  oxcf->target_bandwidth = 1000 * cfg->rc_target_bitrate;
+  oxcf->rc_max_intra_bitrate_pct = extra_cfg->rc_max_intra_bitrate_pct;
+  oxcf->rc_max_inter_bitrate_pct = extra_cfg->rc_max_inter_bitrate_pct;
+  oxcf->gf_cbr_boost_pct = extra_cfg->gf_cbr_boost_pct;
+
+  oxcf->best_allowed_q =
+      extra_cfg->lossless ? 0 : av1_quantizer_to_qindex(cfg->rc_min_quantizer);
+  oxcf->worst_allowed_q =
+      extra_cfg->lossless ? 0 : av1_quantizer_to_qindex(cfg->rc_max_quantizer);
+  oxcf->cq_level = av1_quantizer_to_qindex(extra_cfg->cq_level);
+  oxcf->fixed_q = -1;
+
+#if CONFIG_AOM_QM
+  oxcf->using_qm = extra_cfg->enable_qm;
+  oxcf->qm_minlevel = extra_cfg->qm_min;
+  oxcf->qm_maxlevel = extra_cfg->qm_max;
+#endif
+
+#if CONFIG_TILE_GROUPS
+  oxcf->num_tile_groups = extra_cfg->num_tg;
+  oxcf->mtu = extra_cfg->mtu_size;
+#endif
+
+#if CONFIG_TEMPMV_SIGNALING
+  oxcf->disable_tempmv = extra_cfg->disable_tempmv;
+#endif
+  oxcf->under_shoot_pct = cfg->rc_undershoot_pct;
+  oxcf->over_shoot_pct = cfg->rc_overshoot_pct;
+
+  oxcf->scaled_frame_width = cfg->rc_scaled_width;
+  oxcf->scaled_frame_height = cfg->rc_scaled_height;
+  if (cfg->rc_resize_allowed == 1) {
+    oxcf->resize_mode =
+        (oxcf->scaled_frame_width == 0 || oxcf->scaled_frame_height == 0)
+            ? RESIZE_DYNAMIC
+            : RESIZE_FIXED;
+  } else {
+    oxcf->resize_mode = RESIZE_NONE;
+  }
+
+  oxcf->maximum_buffer_size_ms = is_vbr ? 240000 : cfg->rc_buf_sz;
+  oxcf->starting_buffer_level_ms = is_vbr ? 60000 : cfg->rc_buf_initial_sz;
+  oxcf->optimal_buffer_level_ms = is_vbr ? 60000 : cfg->rc_buf_optimal_sz;
+
+  oxcf->drop_frames_water_mark = cfg->rc_dropframe_thresh;
+
+  oxcf->two_pass_vbrbias = cfg->rc_2pass_vbr_bias_pct;
+  oxcf->two_pass_vbrmin_section = cfg->rc_2pass_vbr_minsection_pct;
+  oxcf->two_pass_vbrmax_section = cfg->rc_2pass_vbr_maxsection_pct;
+
+  oxcf->auto_key =
+      cfg->kf_mode == AOM_KF_AUTO && cfg->kf_min_dist != cfg->kf_max_dist;
+
+  oxcf->key_freq = cfg->kf_max_dist;
+
+  oxcf->speed = extra_cfg->cpu_used;
+  oxcf->enable_auto_arf = extra_cfg->enable_auto_alt_ref;
+#if CONFIG_EXT_REFS
+  oxcf->enable_auto_brf = extra_cfg->enable_auto_bwd_ref;
+#endif  // CONFIG_EXT_REFS
+  oxcf->noise_sensitivity = extra_cfg->noise_sensitivity;
+  oxcf->sharpness = extra_cfg->sharpness;
+
+  oxcf->two_pass_stats_in = cfg->rc_twopass_stats_in;
+
+#if CONFIG_FP_MB_STATS
+  oxcf->firstpass_mb_stats_in = cfg->rc_firstpass_mb_stats_in;
+#endif
+
+  oxcf->color_space = extra_cfg->color_space;
+  oxcf->color_range = extra_cfg->color_range;
+  oxcf->render_width = extra_cfg->render_width;
+  oxcf->render_height = extra_cfg->render_height;
+  oxcf->arnr_max_frames = extra_cfg->arnr_max_frames;
+  oxcf->arnr_strength = extra_cfg->arnr_strength;
+  oxcf->min_gf_interval = extra_cfg->min_gf_interval;
+  oxcf->max_gf_interval = extra_cfg->max_gf_interval;
+
+  oxcf->tuning = extra_cfg->tuning;
+  oxcf->content = extra_cfg->content;
+
+#if CONFIG_EXT_PARTITION
+  oxcf->superblock_size = extra_cfg->superblock_size;
+#endif  // CONFIG_EXT_PARTITION
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  oxcf->ans_window_size_log2 = extra_cfg->ans_window_size_log2;
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+
+#if CONFIG_EXT_TILE
+  {
+#if CONFIG_EXT_PARTITION
+    const unsigned int max =
+        extra_cfg->superblock_size == AOM_SUPERBLOCK_SIZE_64X64 ? 64 : 32;
+#else
+    const unsigned int max = 64;
+#endif  // CONFIG_EXT_PARTITION
+    oxcf->tile_columns = AOMMIN(extra_cfg->tile_columns, max);
+    oxcf->tile_rows = AOMMIN(extra_cfg->tile_rows, max);
+    oxcf->tile_encoding_mode = extra_cfg->tile_encoding_mode;
+  }
+#else
+  oxcf->tile_columns = extra_cfg->tile_columns;
+  oxcf->tile_rows = extra_cfg->tile_rows;
+#endif  // CONFIG_EXT_TILE
+#if CONFIG_DEPENDENT_HORZTILES
+  oxcf->dependent_horz_tiles = extra_cfg->dependent_horz_tiles;
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  oxcf->loop_filter_across_tiles_enabled =
+      extra_cfg->loop_filter_across_tiles_enabled;
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+  oxcf->error_resilient_mode = cfg->g_error_resilient;
+  oxcf->frame_parallel_decoding_mode = extra_cfg->frame_parallel_decoding_mode;
+
+  oxcf->aq_mode = extra_cfg->aq_mode;
+#if CONFIG_EXT_DELTA_Q
+  oxcf->deltaq_mode = extra_cfg->deltaq_mode;
+#endif
+
+  oxcf->frame_periodic_boost = extra_cfg->frame_periodic_boost;
+
+  oxcf->motion_vector_unit_test = extra_cfg->motion_vector_unit_test;
+  /*
+  printf("Current AV1 Settings: \n");
+  printf("target_bandwidth: %d\n", oxcf->target_bandwidth);
+  printf("noise_sensitivity: %d\n", oxcf->noise_sensitivity);
+  printf("sharpness: %d\n",    oxcf->sharpness);
+  printf("cpu_used: %d\n",  oxcf->cpu_used);
+  printf("Mode: %d\n",     oxcf->mode);
+  printf("auto_key: %d\n",  oxcf->auto_key);
+  printf("key_freq: %d\n", oxcf->key_freq);
+  printf("end_usage: %d\n", oxcf->end_usage);
+  printf("under_shoot_pct: %d\n", oxcf->under_shoot_pct);
+  printf("over_shoot_pct: %d\n", oxcf->over_shoot_pct);
+  printf("starting_buffer_level: %d\n", oxcf->starting_buffer_level);
+  printf("optimal_buffer_level: %d\n",  oxcf->optimal_buffer_level);
+  printf("maximum_buffer_size: %d\n", oxcf->maximum_buffer_size);
+  printf("fixed_q: %d\n",  oxcf->fixed_q);
+  printf("worst_allowed_q: %d\n", oxcf->worst_allowed_q);
+  printf("best_allowed_q: %d\n", oxcf->best_allowed_q);
+  printf("allow_spatial_resampling: %d\n", oxcf->allow_spatial_resampling);
+  printf("scaled_frame_width: %d\n", oxcf->scaled_frame_width);
+  printf("scaled_frame_height: %d\n", oxcf->scaled_frame_height);
+  printf("two_pass_vbrbias: %d\n",  oxcf->two_pass_vbrbias);
+  printf("two_pass_vbrmin_section: %d\n", oxcf->two_pass_vbrmin_section);
+  printf("two_pass_vbrmax_section: %d\n", oxcf->two_pass_vbrmax_section);
+  printf("lag_in_frames: %d\n", oxcf->lag_in_frames);
+  printf("enable_auto_arf: %d\n", oxcf->enable_auto_arf);
+  printf("Version: %d\n", oxcf->Version);
+  printf("error resilient: %d\n", oxcf->error_resilient_mode);
+  printf("frame parallel detokenization: %d\n",
+         oxcf->frame_parallel_decoding_mode);
+  */
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t encoder_set_config(aom_codec_alg_priv_t *ctx,
+                                          const aom_codec_enc_cfg_t *cfg) {
+  aom_codec_err_t res;
+  int force_key = 0;
+
+  if (cfg->g_w != ctx->cfg.g_w || cfg->g_h != ctx->cfg.g_h) {
+    if (cfg->g_lag_in_frames > 1 || cfg->g_pass != AOM_RC_ONE_PASS)
+      ERROR("Cannot change width or height after initialization");
+    if (!valid_ref_frame_size(ctx->cfg.g_w, ctx->cfg.g_h, cfg->g_w, cfg->g_h) ||
+        (ctx->cpi->initial_width && (int)cfg->g_w > ctx->cpi->initial_width) ||
+        (ctx->cpi->initial_height && (int)cfg->g_h > ctx->cpi->initial_height))
+      force_key = 1;
+  }
+
+  // Prevent increasing lag_in_frames. This check is stricter than it needs
+  // to be -- the limit is not increasing past the first lag_in_frames
+  // value, but we don't track the initial config, only the last successful
+  // config.
+  if (cfg->g_lag_in_frames > ctx->cfg.g_lag_in_frames)
+    ERROR("Cannot increase lag_in_frames");
+
+  res = validate_config(ctx, cfg, &ctx->extra_cfg);
+
+  if (res == AOM_CODEC_OK) {
+    ctx->cfg = *cfg;
+    set_encoder_config(&ctx->oxcf, &ctx->cfg, &ctx->extra_cfg);
+    // On profile change, request a key frame
+    force_key |= ctx->cpi->common.profile != ctx->oxcf.profile;
+    av1_change_config(ctx->cpi, &ctx->oxcf);
+  }
+
+  if (force_key) ctx->next_frame_flags |= AOM_EFLAG_FORCE_KF;
+
+  return res;
+}
+
+static aom_codec_err_t ctrl_get_quantizer(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL) return AOM_CODEC_INVALID_PARAM;
+  *arg = av1_get_quantizer(ctx->cpi);
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t ctrl_get_quantizer64(aom_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL) return AOM_CODEC_INVALID_PARAM;
+  *arg = av1_qindex_to_quantizer(av1_get_quantizer(ctx->cpi));
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t update_extra_cfg(aom_codec_alg_priv_t *ctx,
+                                        const struct av1_extracfg *extra_cfg) {
+  const aom_codec_err_t res = validate_config(ctx, &ctx->cfg, extra_cfg);
+  if (res == AOM_CODEC_OK) {
+    ctx->extra_cfg = *extra_cfg;
+    set_encoder_config(&ctx->oxcf, &ctx->cfg, &ctx->extra_cfg);
+    av1_change_config(ctx->cpi, &ctx->oxcf);
+  }
+  return res;
+}
+
+static aom_codec_err_t ctrl_set_cpuused(aom_codec_alg_priv_t *ctx,
+                                        va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.cpu_used = CAST(AOME_SET_CPUUSED, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_enable_auto_alt_ref(aom_codec_alg_priv_t *ctx,
+                                                    va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.enable_auto_alt_ref = CAST(AOME_SET_ENABLEAUTOALTREF, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+#if CONFIG_EXT_REFS
+static aom_codec_err_t ctrl_set_enable_auto_bwd_ref(aom_codec_alg_priv_t *ctx,
+                                                    va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.enable_auto_bwd_ref = CAST(AOME_SET_ENABLEAUTOBWDREF, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+#endif  // CONFIG_EXT_REFS
+
+static aom_codec_err_t ctrl_set_noise_sensitivity(aom_codec_alg_priv_t *ctx,
+                                                  va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.noise_sensitivity = CAST(AV1E_SET_NOISE_SENSITIVITY, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_sharpness(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.sharpness = CAST(AOME_SET_SHARPNESS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_static_thresh(aom_codec_alg_priv_t *ctx,
+                                              va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.static_thresh = CAST(AOME_SET_STATIC_THRESHOLD, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_tile_columns(aom_codec_alg_priv_t *ctx,
+                                             va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.tile_columns = CAST(AV1E_SET_TILE_COLUMNS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_tile_rows(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.tile_rows = CAST(AV1E_SET_TILE_ROWS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+#if CONFIG_DEPENDENT_HORZTILES
+static aom_codec_err_t ctrl_set_tile_dependent_rows(aom_codec_alg_priv_t *ctx,
+                                                    va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.dependent_horz_tiles = CAST(AV1E_SET_TILE_DEPENDENT_ROWS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+static aom_codec_err_t ctrl_set_tile_loopfilter(aom_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.loop_filter_across_tiles_enabled =
+      CAST(AV1E_SET_TILE_LOOPFILTER, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+static aom_codec_err_t ctrl_set_arnr_max_frames(aom_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.arnr_max_frames = CAST(AOME_SET_ARNR_MAXFRAMES, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_arnr_strength(aom_codec_alg_priv_t *ctx,
+                                              va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.arnr_strength = CAST(AOME_SET_ARNR_STRENGTH, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_tuning(aom_codec_alg_priv_t *ctx,
+                                       va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.tuning = CAST(AOME_SET_TUNING, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_cq_level(aom_codec_alg_priv_t *ctx,
+                                         va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.cq_level = CAST(AOME_SET_CQ_LEVEL, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_rc_max_intra_bitrate_pct(
+    aom_codec_alg_priv_t *ctx, va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.rc_max_intra_bitrate_pct =
+      CAST(AOME_SET_MAX_INTRA_BITRATE_PCT, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_rc_max_inter_bitrate_pct(
+    aom_codec_alg_priv_t *ctx, va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.rc_max_inter_bitrate_pct =
+      CAST(AOME_SET_MAX_INTER_BITRATE_PCT, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_rc_gf_cbr_boost_pct(aom_codec_alg_priv_t *ctx,
+                                                    va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.gf_cbr_boost_pct = CAST(AV1E_SET_GF_CBR_BOOST_PCT, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_lossless(aom_codec_alg_priv_t *ctx,
+                                         va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.lossless = CAST(AV1E_SET_LOSSLESS, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+#if CONFIG_AOM_QM
+static aom_codec_err_t ctrl_set_enable_qm(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.enable_qm = CAST(AV1E_SET_ENABLE_QM, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_qm_min(aom_codec_alg_priv_t *ctx,
+                                       va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.qm_min = CAST(AV1E_SET_QM_MIN, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_qm_max(aom_codec_alg_priv_t *ctx,
+                                       va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.qm_max = CAST(AV1E_SET_QM_MAX, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+#endif
+
+#if CONFIG_TILE_GROUPS
+static aom_codec_err_t ctrl_set_num_tg(aom_codec_alg_priv_t *ctx,
+                                       va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.num_tg = CAST(AV1E_SET_NUM_TG, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_mtu(aom_codec_alg_priv_t *ctx, va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.mtu_size = CAST(AV1E_SET_MTU, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+#endif
+#if CONFIG_TEMPMV_SIGNALING
+static aom_codec_err_t ctrl_set_disable_tempmv(aom_codec_alg_priv_t *ctx,
+                                               va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.disable_tempmv = CAST(AV1E_SET_DISABLE_TEMPMV, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+#endif
+static aom_codec_err_t ctrl_set_frame_parallel_decoding_mode(
+    aom_codec_alg_priv_t *ctx, va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.frame_parallel_decoding_mode =
+      CAST(AV1E_SET_FRAME_PARALLEL_DECODING, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+#if CONFIG_EXT_TILE
+static aom_codec_err_t ctrl_set_tile_encoding_mode(aom_codec_alg_priv_t *ctx,
+                                                   va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.tile_encoding_mode = CAST(AV1E_SET_TILE_ENCODING_MODE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+#endif  // CONFIG_EXT_TILE
+
+static aom_codec_err_t ctrl_set_aq_mode(aom_codec_alg_priv_t *ctx,
+                                        va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.aq_mode = CAST(AV1E_SET_AQ_MODE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+#if CONFIG_EXT_DELTA_Q
+static aom_codec_err_t ctrl_set_deltaq_mode(aom_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.deltaq_mode = CAST(AV1E_SET_DELTAQ_MODE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+#endif
+static aom_codec_err_t ctrl_set_min_gf_interval(aom_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.min_gf_interval = CAST(AV1E_SET_MIN_GF_INTERVAL, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_max_gf_interval(aom_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.max_gf_interval = CAST(AV1E_SET_MAX_GF_INTERVAL, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_frame_periodic_boost(aom_codec_alg_priv_t *ctx,
+                                                     va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.frame_periodic_boost = CAST(AV1E_SET_FRAME_PERIODIC_BOOST, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_enable_motion_vector_unit_test(
+    aom_codec_alg_priv_t *ctx, va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.motion_vector_unit_test =
+      CAST(AV1E_ENABLE_MOTION_VECTOR_UNIT_TEST, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t encoder_init(aom_codec_ctx_t *ctx,
+                                    aom_codec_priv_enc_mr_cfg_t *data) {
+  aom_codec_err_t res = AOM_CODEC_OK;
+  (void)data;
+
+  if (ctx->priv == NULL) {
+    aom_codec_alg_priv_t *const priv = aom_calloc(1, sizeof(*priv));
+    if (priv == NULL) return AOM_CODEC_MEM_ERROR;
+
+    ctx->priv = (aom_codec_priv_t *)priv;
+    ctx->priv->init_flags = ctx->init_flags;
+    ctx->priv->enc.total_encoders = 1;
+    priv->buffer_pool = (BufferPool *)aom_calloc(1, sizeof(BufferPool));
+    if (priv->buffer_pool == NULL) return AOM_CODEC_MEM_ERROR;
+
+#if CONFIG_MULTITHREAD
+    if (pthread_mutex_init(&priv->buffer_pool->pool_mutex, NULL)) {
+      return AOM_CODEC_MEM_ERROR;
+    }
+#endif
+
+    if (ctx->config.enc) {
+      // Update the reference to the config structure to an internal copy.
+      priv->cfg = *ctx->config.enc;
+      ctx->config.enc = &priv->cfg;
+    }
+
+    priv->extra_cfg = default_extra_cfg;
+    once(av1_initialize_enc);
+
+    res = validate_config(priv, &priv->cfg, &priv->extra_cfg);
+
+    if (res == AOM_CODEC_OK) {
+      set_encoder_config(&priv->oxcf, &priv->cfg, &priv->extra_cfg);
+#if CONFIG_HIGHBITDEPTH
+      priv->oxcf.use_highbitdepth =
+          (ctx->init_flags & AOM_CODEC_USE_HIGHBITDEPTH) ? 1 : 0;
+#endif
+      priv->cpi = av1_create_compressor(&priv->oxcf, priv->buffer_pool);
+      if (priv->cpi == NULL)
+        res = AOM_CODEC_MEM_ERROR;
+      else
+        priv->cpi->output_pkt_list = &priv->pkt_list.head;
+    }
+  }
+
+  return res;
+}
+
+static aom_codec_err_t encoder_destroy(aom_codec_alg_priv_t *ctx) {
+  free(ctx->cx_data);
+  av1_remove_compressor(ctx->cpi);
+#if CONFIG_MULTITHREAD
+  pthread_mutex_destroy(&ctx->buffer_pool->pool_mutex);
+#endif
+  aom_free(ctx->buffer_pool);
+  aom_free(ctx);
+  return AOM_CODEC_OK;
+}
+
+static void pick_quickcompress_mode(aom_codec_alg_priv_t *ctx,
+                                    unsigned long deadline) {
+  MODE new_mode = GOOD;
+
+  switch (ctx->cfg.g_pass) {
+    case AOM_RC_ONE_PASS:
+      switch (deadline) {
+        default: new_mode = GOOD; break;
+      }
+      break;
+    case AOM_RC_FIRST_PASS: break;
+    case AOM_RC_LAST_PASS: new_mode = GOOD;
+  }
+
+  if (ctx->oxcf.mode != new_mode) {
+    ctx->oxcf.mode = new_mode;
+    av1_change_config(ctx->cpi, &ctx->oxcf);
+  }
+}
+
+// Turn on to test if supplemental superframe data breaks decoding
+// #define TEST_SUPPLEMENTAL_SUPERFRAME_DATA
+static int write_superframe_index(aom_codec_alg_priv_t *ctx) {
+  uint8_t marker = 0xc0;
+  unsigned int mask;
+  int mag, index_sz;
+  int i;
+  size_t max_frame_sz = 0;
+
+  assert(ctx->pending_frame_count);
+  assert(ctx->pending_frame_count <= 8);
+
+  // Add the number of frames to the marker byte
+  marker |= ctx->pending_frame_count - 1;
+  for (i = 0; i < ctx->pending_frame_count - 1; i++) {
+    const size_t frame_sz = (unsigned int)ctx->pending_frame_sizes[i] - 1;
+    max_frame_sz = frame_sz > max_frame_sz ? frame_sz : max_frame_sz;
+  }
+
+  // Choose the magnitude
+  for (mag = 0, mask = 0xff; mag < 4; mag++) {
+    if (max_frame_sz <= mask) break;
+    mask <<= 8;
+    mask |= 0xff;
+  }
+  marker |= mag << 3;
+
+  // Write the index
+  index_sz = 2 + (mag + 1) * (ctx->pending_frame_count - 1);
+  if (ctx->pending_cx_data_sz + index_sz < ctx->cx_data_sz) {
+    uint8_t *x = ctx->pending_cx_data + ctx->pending_cx_data_sz;
+#ifdef TEST_SUPPLEMENTAL_SUPERFRAME_DATA
+    uint8_t marker_test = 0xc0;
+    int mag_test = 2;     // 1 - 4
+    int frames_test = 4;  // 1 - 8
+    int index_sz_test = 2 + mag_test * frames_test;
+    marker_test |= frames_test - 1;
+    marker_test |= (mag_test - 1) << 3;
+    *x++ = marker_test;
+    for (i = 0; i < mag_test * frames_test; ++i)
+      *x++ = 0;  // fill up with arbitrary data
+    *x++ = marker_test;
+    ctx->pending_cx_data_sz += index_sz_test;
+    printf("Added supplemental superframe data\n");
+#endif
+
+    *x++ = marker;
+    for (i = 0; i < ctx->pending_frame_count - 1; i++) {
+      unsigned int this_sz;
+      int j;
+
+      assert(ctx->pending_frame_sizes[i] > 0);
+      this_sz = (unsigned int)ctx->pending_frame_sizes[i] - 1;
+      for (j = 0; j <= mag; j++) {
+        *x++ = this_sz & 0xff;
+        this_sz >>= 8;
+      }
+    }
+    *x++ = marker;
+    ctx->pending_cx_data_sz += index_sz;
+#ifdef TEST_SUPPLEMENTAL_SUPERFRAME_DATA
+    index_sz += index_sz_test;
+#endif
+  }
+  return index_sz;
+}
+
+// av1 uses 10,000,000 ticks/second as time stamp
+#define TICKS_PER_SEC 10000000LL
+
+static int64_t timebase_units_to_ticks(const aom_rational_t *timebase,
+                                       int64_t n) {
+  return n * TICKS_PER_SEC * timebase->num / timebase->den;
+}
+
+static int64_t ticks_to_timebase_units(const aom_rational_t *timebase,
+                                       int64_t n) {
+  const int64_t round = TICKS_PER_SEC * timebase->num / 2 - 1;
+  return (n * timebase->den + round) / timebase->num / TICKS_PER_SEC;
+}
+
+static aom_codec_frame_flags_t get_frame_pkt_flags(const AV1_COMP *cpi,
+                                                   unsigned int lib_flags) {
+  aom_codec_frame_flags_t flags = lib_flags << 16;
+
+  if (lib_flags & FRAMEFLAGS_KEY) flags |= AOM_FRAME_IS_KEY;
+
+  if (cpi->droppable) flags |= AOM_FRAME_IS_DROPPABLE;
+
+  return flags;
+}
+
+static aom_codec_err_t encoder_encode(aom_codec_alg_priv_t *ctx,
+                                      const aom_image_t *img,
+                                      aom_codec_pts_t pts,
+                                      unsigned long duration,
+                                      aom_enc_frame_flags_t enc_flags,
+                                      unsigned long deadline) {
+  const size_t kMinCompressedSize = 8192;
+  volatile aom_codec_err_t res = AOM_CODEC_OK;
+  volatile aom_enc_frame_flags_t flags = enc_flags;
+  AV1_COMP *const cpi = ctx->cpi;
+  const aom_rational_t *const timebase = &ctx->cfg.g_timebase;
+  size_t data_sz;
+
+  if (cpi == NULL) return AOM_CODEC_INVALID_PARAM;
+
+  if (img != NULL) {
+    res = validate_img(ctx, img);
+    // TODO(jzern) the checks related to cpi's validity should be treated as a
+    // failure condition, encoder setup is done fully in init() currently.
+    if (res == AOM_CODEC_OK) {
+#if CONFIG_EXT_REFS
+      data_sz = ALIGN_POWER_OF_TWO(ctx->cfg.g_w, 5) *
+                ALIGN_POWER_OF_TWO(ctx->cfg.g_h, 5) * get_image_bps(img);
+#else
+      // There's no codec control for multiple alt-refs so check the encoder
+      // instance for its status to determine the compressed data size.
+      data_sz = ALIGN_POWER_OF_TWO(ctx->cfg.g_w, 5) *
+                ALIGN_POWER_OF_TWO(ctx->cfg.g_h, 5) * get_image_bps(img) / 8 *
+                (cpi->multi_arf_allowed ? 8 : 2);
+#endif  // CONFIG_EXT_REFS
+      if (data_sz < kMinCompressedSize) data_sz = kMinCompressedSize;
+      if (ctx->cx_data == NULL || ctx->cx_data_sz < data_sz) {
+        ctx->cx_data_sz = data_sz;
+        free(ctx->cx_data);
+        ctx->cx_data = (unsigned char *)malloc(ctx->cx_data_sz);
+        if (ctx->cx_data == NULL) {
+          return AOM_CODEC_MEM_ERROR;
+        }
+      }
+    }
+  }
+
+  pick_quickcompress_mode(ctx, deadline);
+  aom_codec_pkt_list_init(&ctx->pkt_list);
+
+  // Handle Flags
+  if (((flags & AOM_EFLAG_NO_UPD_GF) && (flags & AOM_EFLAG_FORCE_GF)) ||
+      ((flags & AOM_EFLAG_NO_UPD_ARF) && (flags & AOM_EFLAG_FORCE_ARF))) {
+    ctx->base.err_detail = "Conflicting flags.";
+    return AOM_CODEC_INVALID_PARAM;
+  }
+
+  if (setjmp(cpi->common.error.jmp)) {
+    cpi->common.error.setjmp = 0;
+    res = update_error_state(ctx, &cpi->common.error);
+    aom_clear_system_state();
+    return res;
+  }
+  cpi->common.error.setjmp = 1;
+
+  av1_apply_encoding_flags(cpi, flags);
+
+  // Handle fixed keyframe intervals
+  if (ctx->cfg.kf_mode == AOM_KF_AUTO &&
+      ctx->cfg.kf_min_dist == ctx->cfg.kf_max_dist) {
+    if (++ctx->fixed_kf_cntr > ctx->cfg.kf_min_dist) {
+      flags |= AOM_EFLAG_FORCE_KF;
+      ctx->fixed_kf_cntr = 1;
+    }
+  }
+
+  if (res == AOM_CODEC_OK) {
+    unsigned int lib_flags = 0;
+    YV12_BUFFER_CONFIG sd;
+    int64_t dst_time_stamp = timebase_units_to_ticks(timebase, pts);
+    int64_t dst_end_time_stamp =
+        timebase_units_to_ticks(timebase, pts + duration);
+    size_t size, cx_data_sz;
+    unsigned char *cx_data;
+
+    // Set up internal flags
+    if (ctx->base.init_flags & AOM_CODEC_USE_PSNR) cpi->b_calculate_psnr = 1;
+
+    if (img != NULL) {
+      res = image2yuvconfig(img, &sd);
+
+      // Store the original flags in to the frame buffer. Will extract the
+      // key frame flag when we actually encode this frame.
+      if (av1_receive_raw_frame(cpi, flags | ctx->next_frame_flags, &sd,
+                                dst_time_stamp, dst_end_time_stamp)) {
+        res = update_error_state(ctx, &cpi->common.error);
+      }
+      ctx->next_frame_flags = 0;
+    }
+
+    cx_data = ctx->cx_data;
+    cx_data_sz = ctx->cx_data_sz;
+
+    /* Any pending invisible frames? */
+    if (ctx->pending_cx_data) {
+      memmove(cx_data, ctx->pending_cx_data, ctx->pending_cx_data_sz);
+      ctx->pending_cx_data = cx_data;
+      cx_data += ctx->pending_cx_data_sz;
+      cx_data_sz -= ctx->pending_cx_data_sz;
+
+      /* TODO: this is a minimal check, the underlying codec doesn't respect
+       * the buffer size anyway.
+       */
+      if (cx_data_sz < ctx->cx_data_sz / 2) {
+        aom_internal_error(&cpi->common.error, AOM_CODEC_ERROR,
+                           "Compressed data buffer too small");
+        return AOM_CODEC_ERROR;
+      }
+    }
+
+    while (cx_data_sz >= ctx->cx_data_sz / 2 &&
+           -1 != av1_get_compressed_data(cpi, &lib_flags, &size, cx_data,
+                                         &dst_time_stamp, &dst_end_time_stamp,
+                                         !img)) {
+#if CONFIG_REFERENCE_BUFFER
+      if (cpi->common.invalid_delta_frame_id_minus1) {
+        ctx->base.err_detail = "Invalid delta_frame_id_minus1";
+        return AOM_CODEC_ERROR;
+      }
+#endif
+      if (size) {
+        aom_codec_cx_pkt_t pkt;
+
+        // Pack invisible frames with the next visible frame
+        if (!cpi->common.show_frame) {
+          if (ctx->pending_cx_data == 0) ctx->pending_cx_data = cx_data;
+          ctx->pending_cx_data_sz += size;
+          ctx->pending_frame_sizes[ctx->pending_frame_count++] = size;
+          cx_data += size;
+          cx_data_sz -= size;
+
+          continue;
+        }
+
+        // Add the frame packet to the list of returned packets.
+        pkt.kind = AOM_CODEC_CX_FRAME_PKT;
+        pkt.data.frame.pts = ticks_to_timebase_units(timebase, dst_time_stamp);
+        pkt.data.frame.duration = (unsigned long)ticks_to_timebase_units(
+            timebase, dst_end_time_stamp - dst_time_stamp);
+        pkt.data.frame.flags = get_frame_pkt_flags(cpi, lib_flags);
+
+        if (ctx->pending_cx_data) {
+          ctx->pending_frame_sizes[ctx->pending_frame_count++] = size;
+          ctx->pending_cx_data_sz += size;
+          size += write_superframe_index(ctx);
+          pkt.data.frame.buf = ctx->pending_cx_data;
+          pkt.data.frame.sz = ctx->pending_cx_data_sz;
+          ctx->pending_cx_data = NULL;
+          ctx->pending_cx_data_sz = 0;
+          ctx->pending_frame_count = 0;
+        } else {
+          pkt.data.frame.buf = cx_data;
+          pkt.data.frame.sz = size;
+        }
+        pkt.data.frame.partition_id = -1;
+
+        aom_codec_pkt_list_add(&ctx->pkt_list.head, &pkt);
+
+        cx_data += size;
+        cx_data_sz -= size;
+      }
+    }
+  }
+
+  cpi->common.error.setjmp = 0;
+  return res;
+}
+
+static const aom_codec_cx_pkt_t *encoder_get_cxdata(aom_codec_alg_priv_t *ctx,
+                                                    aom_codec_iter_t *iter) {
+  return aom_codec_pkt_list_get(&ctx->pkt_list.head, iter);
+}
+
+static aom_codec_err_t ctrl_set_reference(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  aom_ref_frame_t *const frame = va_arg(args, aom_ref_frame_t *);
+
+  if (frame != NULL) {
+    YV12_BUFFER_CONFIG sd;
+
+    image2yuvconfig(&frame->img, &sd);
+    av1_set_reference_enc(ctx->cpi, ref_frame_to_av1_reframe(frame->frame_type),
+                          &sd);
+    return AOM_CODEC_OK;
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_copy_reference(aom_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  aom_ref_frame_t *const frame = va_arg(args, aom_ref_frame_t *);
+
+  if (frame != NULL) {
+    YV12_BUFFER_CONFIG sd;
+
+    image2yuvconfig(&frame->img, &sd);
+    av1_copy_reference_enc(ctx->cpi,
+                           ref_frame_to_av1_reframe(frame->frame_type), &sd);
+    return AOM_CODEC_OK;
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_get_reference(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  av1_ref_frame_t *const frame = va_arg(args, av1_ref_frame_t *);
+
+  if (frame != NULL) {
+    YV12_BUFFER_CONFIG *fb = get_ref_frame(&ctx->cpi->common, frame->idx);
+    if (fb == NULL) return AOM_CODEC_ERROR;
+
+    yuvconfig2image(&frame->img, fb, NULL);
+    return AOM_CODEC_OK;
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_get_new_frame_image(aom_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  aom_image_t *const new_img = va_arg(args, aom_image_t *);
+
+  if (new_img != NULL) {
+    YV12_BUFFER_CONFIG new_frame;
+
+    if (av1_get_last_show_frame(ctx->cpi, &new_frame) == 0) {
+      yuvconfig2image(new_img, &new_frame, NULL);
+      return AOM_CODEC_OK;
+    } else {
+      return AOM_CODEC_ERROR;
+    }
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_set_previewpp(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  (void)ctx;
+  (void)args;
+  return AOM_CODEC_INCAPABLE;
+}
+
+static aom_image_t *encoder_get_preview(aom_codec_alg_priv_t *ctx) {
+  YV12_BUFFER_CONFIG sd;
+
+  if (av1_get_preview_raw_frame(ctx->cpi, &sd) == 0) {
+    yuvconfig2image(&ctx->preview_img, &sd, NULL);
+    return &ctx->preview_img;
+  } else {
+    return NULL;
+  }
+}
+
+static aom_codec_err_t ctrl_use_reference(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  const int reference_flag = va_arg(args, int);
+
+  av1_use_as_reference(ctx->cpi, reference_flag);
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t ctrl_set_roi_map(aom_codec_alg_priv_t *ctx,
+                                        va_list args) {
+  (void)ctx;
+  (void)args;
+
+  // TODO(yaowu): Need to re-implement and test for AV1.
+  return AOM_CODEC_INVALID_PARAM;
+}
+
+static aom_codec_err_t ctrl_set_active_map(aom_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  aom_active_map_t *const map = va_arg(args, aom_active_map_t *);
+
+  if (map) {
+    if (!av1_set_active_map(ctx->cpi, map->active_map, (int)map->rows,
+                            (int)map->cols))
+      return AOM_CODEC_OK;
+    else
+      return AOM_CODEC_INVALID_PARAM;
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_get_active_map(aom_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  aom_active_map_t *const map = va_arg(args, aom_active_map_t *);
+
+  if (map) {
+    if (!av1_get_active_map(ctx->cpi, map->active_map, (int)map->rows,
+                            (int)map->cols))
+      return AOM_CODEC_OK;
+    else
+      return AOM_CODEC_INVALID_PARAM;
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_set_scale_mode(aom_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  aom_scaling_mode_t *const mode = va_arg(args, aom_scaling_mode_t *);
+
+  if (mode) {
+    const int res =
+        av1_set_internal_size(ctx->cpi, (AOM_SCALING)mode->h_scaling_mode,
+                              (AOM_SCALING)mode->v_scaling_mode);
+    return (res == 0) ? AOM_CODEC_OK : AOM_CODEC_INVALID_PARAM;
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_set_tune_content(aom_codec_alg_priv_t *ctx,
+                                             va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.content = CAST(AV1E_SET_TUNE_CONTENT, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_color_space(aom_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.color_space = CAST(AV1E_SET_COLOR_SPACE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_color_range(aom_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.color_range = CAST(AV1E_SET_COLOR_RANGE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_render_size(aom_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  int *const render_size = va_arg(args, int *);
+  extra_cfg.render_width = render_size[0];
+  extra_cfg.render_height = render_size[1];
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static aom_codec_err_t ctrl_set_superblock_size(aom_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.superblock_size = CAST(AV1E_SET_SUPERBLOCK_SIZE, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+static aom_codec_err_t ctrl_set_ans_window_size_log2(aom_codec_alg_priv_t *ctx,
+                                                     va_list args) {
+  struct av1_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.ans_window_size_log2 = CAST(AV1E_SET_ANS_WINDOW_SIZE_LOG2, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+#endif
+
+static aom_codec_ctrl_fn_map_t encoder_ctrl_maps[] = {
+  { AOM_COPY_REFERENCE, ctrl_copy_reference },
+  { AOME_USE_REFERENCE, ctrl_use_reference },
+
+  // Setters
+  { AOM_SET_REFERENCE, ctrl_set_reference },
+  { AOM_SET_POSTPROC, ctrl_set_previewpp },
+  { AOME_SET_ROI_MAP, ctrl_set_roi_map },
+  { AOME_SET_ACTIVEMAP, ctrl_set_active_map },
+  { AOME_SET_SCALEMODE, ctrl_set_scale_mode },
+  { AOME_SET_CPUUSED, ctrl_set_cpuused },
+  { AOME_SET_ENABLEAUTOALTREF, ctrl_set_enable_auto_alt_ref },
+#if CONFIG_EXT_REFS
+  { AOME_SET_ENABLEAUTOBWDREF, ctrl_set_enable_auto_bwd_ref },
+#endif  // CONFIG_EXT_REFS
+  { AOME_SET_SHARPNESS, ctrl_set_sharpness },
+  { AOME_SET_STATIC_THRESHOLD, ctrl_set_static_thresh },
+  { AV1E_SET_TILE_COLUMNS, ctrl_set_tile_columns },
+  { AV1E_SET_TILE_ROWS, ctrl_set_tile_rows },
+#if CONFIG_DEPENDENT_HORZTILES
+  { AV1E_SET_TILE_DEPENDENT_ROWS, ctrl_set_tile_dependent_rows },
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  { AV1E_SET_TILE_LOOPFILTER, ctrl_set_tile_loopfilter },
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+  { AOME_SET_ARNR_MAXFRAMES, ctrl_set_arnr_max_frames },
+  { AOME_SET_ARNR_STRENGTH, ctrl_set_arnr_strength },
+  { AOME_SET_TUNING, ctrl_set_tuning },
+  { AOME_SET_CQ_LEVEL, ctrl_set_cq_level },
+  { AOME_SET_MAX_INTRA_BITRATE_PCT, ctrl_set_rc_max_intra_bitrate_pct },
+  { AV1E_SET_MAX_INTER_BITRATE_PCT, ctrl_set_rc_max_inter_bitrate_pct },
+  { AV1E_SET_GF_CBR_BOOST_PCT, ctrl_set_rc_gf_cbr_boost_pct },
+  { AV1E_SET_LOSSLESS, ctrl_set_lossless },
+#if CONFIG_AOM_QM
+  { AV1E_SET_ENABLE_QM, ctrl_set_enable_qm },
+  { AV1E_SET_QM_MIN, ctrl_set_qm_min },
+  { AV1E_SET_QM_MAX, ctrl_set_qm_max },
+#endif
+#if CONFIG_TILE_GROUPS
+  { AV1E_SET_NUM_TG, ctrl_set_num_tg },
+  { AV1E_SET_MTU, ctrl_set_mtu },
+#endif
+#if CONFIG_TEMPMV_SIGNALING
+  { AV1E_SET_DISABLE_TEMPMV, ctrl_set_disable_tempmv },
+#endif
+  { AV1E_SET_FRAME_PARALLEL_DECODING, ctrl_set_frame_parallel_decoding_mode },
+  { AV1E_SET_AQ_MODE, ctrl_set_aq_mode },
+#if CONFIG_EXT_DELTA_Q
+  { AV1E_SET_DELTAQ_MODE, ctrl_set_deltaq_mode },
+#endif
+  { AV1E_SET_FRAME_PERIODIC_BOOST, ctrl_set_frame_periodic_boost },
+  { AV1E_SET_TUNE_CONTENT, ctrl_set_tune_content },
+  { AV1E_SET_COLOR_SPACE, ctrl_set_color_space },
+  { AV1E_SET_COLOR_RANGE, ctrl_set_color_range },
+  { AV1E_SET_NOISE_SENSITIVITY, ctrl_set_noise_sensitivity },
+  { AV1E_SET_MIN_GF_INTERVAL, ctrl_set_min_gf_interval },
+  { AV1E_SET_MAX_GF_INTERVAL, ctrl_set_max_gf_interval },
+  { AV1E_SET_RENDER_SIZE, ctrl_set_render_size },
+  { AV1E_SET_SUPERBLOCK_SIZE, ctrl_set_superblock_size },
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  { AV1E_SET_ANS_WINDOW_SIZE_LOG2, ctrl_set_ans_window_size_log2 },
+#endif
+#if CONFIG_EXT_TILE
+  { AV1E_SET_TILE_ENCODING_MODE, ctrl_set_tile_encoding_mode },
+#endif  // CONFIG_EXT_TILE
+  { AV1E_ENABLE_MOTION_VECTOR_UNIT_TEST, ctrl_enable_motion_vector_unit_test },
+
+  // Getters
+  { AOME_GET_LAST_QUANTIZER, ctrl_get_quantizer },
+  { AOME_GET_LAST_QUANTIZER_64, ctrl_get_quantizer64 },
+  { AV1_GET_REFERENCE, ctrl_get_reference },
+  { AV1E_GET_ACTIVEMAP, ctrl_get_active_map },
+  { AV1_GET_NEW_FRAME_IMAGE, ctrl_get_new_frame_image },
+
+  { -1, NULL },
+};
+
+static aom_codec_enc_cfg_map_t encoder_usage_cfg_map[] = {
+  { 0,
+    {
+        // NOLINT
+        0,  // g_usage
+        8,  // g_threads
+        0,  // g_profile
+
+        320,         // g_width
+        240,         // g_height
+        AOM_BITS_8,  // g_bit_depth
+        8,           // g_input_bit_depth
+
+        { 1, 30 },  // g_timebase
+
+        0,  // g_error_resilient
+
+        AOM_RC_ONE_PASS,  // g_pass
+
+        25,  // g_lag_in_frames
+
+        0,   // rc_dropframe_thresh
+        0,   // rc_resize_allowed
+        0,   // rc_scaled_width
+        0,   // rc_scaled_height
+        60,  // rc_resize_down_thresold
+        30,  // rc_resize_up_thresold
+
+        AOM_VBR,      // rc_end_usage
+        { NULL, 0 },  // rc_twopass_stats_in
+        { NULL, 0 },  // rc_firstpass_mb_stats_in
+        256,          // rc_target_bandwidth
+        0,            // rc_min_quantizer
+        63,           // rc_max_quantizer
+        25,           // rc_undershoot_pct
+        25,           // rc_overshoot_pct
+
+        6000,  // rc_max_buffer_size
+        4000,  // rc_buffer_initial_size
+        5000,  // rc_buffer_optimal_size
+
+        50,    // rc_two_pass_vbrbias
+        0,     // rc_two_pass_vbrmin_section
+        2000,  // rc_two_pass_vbrmax_section
+
+        // keyframing settings (kf)
+        AOM_KF_AUTO,  // g_kfmode
+        0,            // kf_min_dist
+        9999,         // kf_max_dist
+    } },
+};
+
+#ifndef VERSION_STRING
+#define VERSION_STRING
+#endif
+CODEC_INTERFACE(aom_codec_av1_cx) = {
+  "AOMedia Project AV1 Encoder" VERSION_STRING,
+  AOM_CODEC_INTERNAL_ABI_VERSION,
+#if CONFIG_HIGHBITDEPTH
+  AOM_CODEC_CAP_HIGHBITDEPTH |
+#endif
+      AOM_CODEC_CAP_ENCODER | AOM_CODEC_CAP_PSNR,  // aom_codec_caps_t
+  encoder_init,                                    // aom_codec_init_fn_t
+  encoder_destroy,                                 // aom_codec_destroy_fn_t
+  encoder_ctrl_maps,                               // aom_codec_ctrl_fn_map_t
+  {
+      // NOLINT
+      NULL,  // aom_codec_peek_si_fn_t
+      NULL,  // aom_codec_get_si_fn_t
+      NULL,  // aom_codec_decode_fn_t
+      NULL,  // aom_codec_frame_get_fn_t
+      NULL   // aom_codec_set_fb_fn_t
+  },
+  {
+      // NOLINT
+      1,                      // 1 cfg map
+      encoder_usage_cfg_map,  // aom_codec_enc_cfg_map_t
+      encoder_encode,         // aom_codec_encode_fn_t
+      encoder_get_cxdata,     // aom_codec_get_cx_data_fn_t
+      encoder_set_config,     // aom_codec_enc_config_set_fn_t
+      NULL,                   // aom_codec_get_global_headers_fn_t
+      encoder_get_preview,    // aom_codec_get_preview_frame_fn_t
+      NULL                    // aom_codec_enc_mr_get_mem_loc_fn_t
+  }
+};
diff --git a/third_party/aom/av1/av1_dx.mk b/third_party/aom/av1/av1_dx.mk
new file mode 100644
index 0000000000..1a54ea22a2
--- /dev/null
+++ b/third_party/aom/av1/av1_dx.mk
@@ -0,0 +1,71 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+AV1_DX_EXPORTS += exports_dec
+
+AV1_DX_SRCS-yes += $(AV1_COMMON_SRCS-yes)
+AV1_DX_SRCS-no  += $(AV1_COMMON_SRCS-no)
+AV1_DX_SRCS_REMOVE-yes += $(AV1_COMMON_SRCS_REMOVE-yes)
+AV1_DX_SRCS_REMOVE-no  += $(AV1_COMMON_SRCS_REMOVE-no)
+
+AV1_DX_SRCS-yes += av1_dx_iface.c
+
+AV1_DX_SRCS-yes += decoder/decodemv.c
+AV1_DX_SRCS-yes += decoder/decodeframe.c
+AV1_DX_SRCS-yes += decoder/decodeframe.h
+AV1_DX_SRCS-yes += decoder/detokenize.c
+AV1_DX_SRCS-yes += decoder/decodemv.h
+AV1_DX_SRCS-$(CONFIG_LV_MAP) += decoder/decodetxb.c
+AV1_DX_SRCS-$(CONFIG_LV_MAP) += decoder/decodetxb.h
+AV1_DX_SRCS-yes += decoder/detokenize.h
+AV1_DX_SRCS-yes += decoder/dthread.c
+AV1_DX_SRCS-yes += decoder/dthread.h
+AV1_DX_SRCS-yes += decoder/decoder.c
+AV1_DX_SRCS-yes += decoder/decoder.h
+AV1_DX_SRCS-yes += decoder/dsubexp.c
+AV1_DX_SRCS-yes += decoder/dsubexp.h
+
+ifeq ($(CONFIG_ACCOUNTING),yes)
+AV1_DX_SRCS-yes += decoder/accounting.h
+AV1_DX_SRCS-yes += decoder/accounting.c
+endif
+
+ifeq ($(CONFIG_INSPECTION),yes)
+AV1_DX_SRCS-yes += decoder/inspection.c
+AV1_DX_SRCS-yes += decoder/inspection.h
+endif
+
+ifeq ($(CONFIG_PVQ),yes)
+# PVQ from daala
+AV1_DX_SRCS-yes += decoder/pvq_decoder.c
+AV1_DX_SRCS-yes += decoder/pvq_decoder.h
+AV1_DX_SRCS-yes += decoder/decint.h
+AV1_DX_SRCS-yes += decoder/generic_decoder.c
+AV1_DX_SRCS-yes += decoder/laplace_decoder.c
+AV1_DX_SRCS-yes += encoder/hybrid_fwd_txfm.c
+AV1_DX_SRCS-yes += encoder/hybrid_fwd_txfm.h
+
+AV1_DX_SRCS-yes += encoder/dct.c
+AV1_DX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_sse2.asm
+AV1_DX_SRCS-$(HAVE_SSE2) += encoder/x86/dct_intrin_sse2.c
+AV1_DX_SRCS-$(HAVE_SSSE3) += encoder/x86/dct_ssse3.c
+
+ifneq ($(CONFIG_HIGHBITDEPTH),yes)
+AV1_DX_SRCS-$(HAVE_NEON) += encoder/arm/neon/dct_neon.c
+endif
+
+AV1_DX_SRCS-$(HAVE_MSA) += encoder/mips/msa/fdct4x4_msa.c
+AV1_DX_SRCS-$(HAVE_MSA) += encoder/mips/msa/fdct8x8_msa.c
+AV1_DX_SRCS-$(HAVE_MSA) += encoder/mips/msa/fdct16x16_msa.c
+AV1_DX_SRCS-$(HAVE_MSA) += encoder/mips/msa/fdct_msa.h
+endif
+
+AV1_DX_SRCS-yes := $(filter-out $(AV1_DX_SRCS_REMOVE-yes),$(AV1_DX_SRCS-yes))
diff --git a/third_party/aom/av1/av1_dx_iface.c b/third_party/aom/av1/av1_dx_iface.c
new file mode 100644
index 0000000000..f20ea4815a
--- /dev/null
+++ b/third_party/aom/av1/av1_dx_iface.c
@@ -0,0 +1,1223 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "./aom_config.h"
+#include "./aom_version.h"
+
+#include "aom/internal/aom_codec_internal.h"
+#include "aom/aomdx.h"
+#include "aom/aom_decoder.h"
+#include "aom_dsp/bitreader_buffer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_util/aom_thread.h"
+
+#include "av1/common/alloccommon.h"
+#include "av1/common/frame_buffers.h"
+#include "av1/common/enums.h"
+
+#include "av1/decoder/decoder.h"
+#include "av1/decoder/decodeframe.h"
+
+#include "av1/av1_iface_common.h"
+
+typedef aom_codec_stream_info_t av1_stream_info_t;
+
+// This limit is due to framebuffer numbers.
+// TODO(hkuang): Remove this limit after implementing ondemand framebuffers.
+#define FRAME_CACHE_SIZE 6  // Cache maximum 6 decoded frames.
+
+typedef struct cache_frame {
+  int fb_idx;
+  aom_image_t img;
+} cache_frame;
+
+struct aom_codec_alg_priv {
+  aom_codec_priv_t base;
+  aom_codec_dec_cfg_t cfg;
+  av1_stream_info_t si;
+  int postproc_cfg_set;
+  aom_postproc_cfg_t postproc_cfg;
+  aom_decrypt_cb decrypt_cb;
+  void *decrypt_state;
+  aom_image_t img;
+  int img_avail;
+  int flushed;
+  int invert_tile_order;
+  int last_show_frame;  // Index of last output frame.
+  int byte_alignment;
+  int skip_loop_filter;
+  int decode_tile_row;
+  int decode_tile_col;
+
+  // Frame parallel related.
+  int frame_parallel_decode;  // frame-based threading.
+  AVxWorker *frame_workers;
+  int num_frame_workers;
+  int next_submit_worker_id;
+  int last_submit_worker_id;
+  int next_output_worker_id;
+  int available_threads;
+  cache_frame frame_cache[FRAME_CACHE_SIZE];
+  int frame_cache_write;
+  int frame_cache_read;
+  int num_cache_frames;
+  int need_resync;  // wait for key/intra-only frame
+  // BufferPool that holds all reference frames. Shared by all the FrameWorkers.
+  BufferPool *buffer_pool;
+
+  // External frame buffer info to save for AV1 common.
+  void *ext_priv;  // Private data associated with the external frame buffers.
+  aom_get_frame_buffer_cb_fn_t get_ext_fb_cb;
+  aom_release_frame_buffer_cb_fn_t release_ext_fb_cb;
+
+#if CONFIG_INSPECTION
+  aom_inspect_cb inspect_cb;
+  void *inspect_ctx;
+#endif
+};
+
+static aom_codec_err_t decoder_init(aom_codec_ctx_t *ctx,
+                                    aom_codec_priv_enc_mr_cfg_t *data) {
+  // This function only allocates space for the aom_codec_alg_priv_t
+  // structure. More memory may be required at the time the stream
+  // information becomes known.
+  (void)data;
+
+  if (!ctx->priv) {
+    aom_codec_alg_priv_t *const priv =
+        (aom_codec_alg_priv_t *)aom_calloc(1, sizeof(*priv));
+    if (priv == NULL) return AOM_CODEC_MEM_ERROR;
+
+    ctx->priv = (aom_codec_priv_t *)priv;
+    ctx->priv->init_flags = ctx->init_flags;
+    priv->si.sz = sizeof(priv->si);
+    priv->flushed = 0;
+    // Only do frame parallel decode when threads > 1.
+    priv->frame_parallel_decode =
+        (ctx->config.dec && (ctx->config.dec->threads > 1) &&
+         (ctx->init_flags & AOM_CODEC_USE_FRAME_THREADING))
+            ? 1
+            : 0;
+    if (ctx->config.dec) {
+      priv->cfg = *ctx->config.dec;
+      ctx->config.dec = &priv->cfg;
+    }
+  }
+
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t decoder_destroy(aom_codec_alg_priv_t *ctx) {
+  if (ctx->frame_workers != NULL) {
+    int i;
+    for (i = 0; i < ctx->num_frame_workers; ++i) {
+      AVxWorker *const worker = &ctx->frame_workers[i];
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      aom_get_worker_interface()->end(worker);
+      av1_remove_common(&frame_worker_data->pbi->common);
+#if CONFIG_LOOP_RESTORATION
+      av1_free_restoration_buffers(&frame_worker_data->pbi->common);
+#endif  // CONFIG_LOOP_RESTORATION
+      av1_decoder_remove(frame_worker_data->pbi);
+      aom_free(frame_worker_data->scratch_buffer);
+#if CONFIG_MULTITHREAD
+      pthread_mutex_destroy(&frame_worker_data->stats_mutex);
+      pthread_cond_destroy(&frame_worker_data->stats_cond);
+#endif
+      aom_free(frame_worker_data);
+    }
+#if CONFIG_MULTITHREAD
+    pthread_mutex_destroy(&ctx->buffer_pool->pool_mutex);
+#endif
+  }
+
+  if (ctx->buffer_pool) {
+    av1_free_ref_frame_buffers(ctx->buffer_pool);
+    av1_free_internal_frame_buffers(&ctx->buffer_pool->int_frame_buffers);
+  }
+
+  aom_free(ctx->frame_workers);
+  aom_free(ctx->buffer_pool);
+  aom_free(ctx);
+  return AOM_CODEC_OK;
+}
+
+static int parse_bitdepth_colorspace_sampling(BITSTREAM_PROFILE profile,
+                                              struct aom_read_bit_buffer *rb) {
+  aom_color_space_t color_space;
+  if (profile >= PROFILE_2) rb->bit_offset += 1;  // Bit-depth 10 or 12.
+  color_space = (aom_color_space_t)aom_rb_read_literal(rb, 3);
+  if (color_space != AOM_CS_SRGB) {
+    rb->bit_offset += 1;  // [16,235] (including xvycc) vs [0,255] range.
+    if (profile == PROFILE_1 || profile == PROFILE_3) {
+      rb->bit_offset += 2;  // subsampling x/y.
+      rb->bit_offset += 1;  // unused.
+    }
+  } else {
+    if (profile == PROFILE_1 || profile == PROFILE_3) {
+      rb->bit_offset += 1;  // unused
+    } else {
+      // RGB is only available in version 1.
+      return 0;
+    }
+  }
+  return 1;
+}
+
+static aom_codec_err_t decoder_peek_si_internal(
+    const uint8_t *data, unsigned int data_sz, aom_codec_stream_info_t *si,
+    int *is_intra_only, aom_decrypt_cb decrypt_cb, void *decrypt_state) {
+  int intra_only_flag = 0;
+  uint8_t clear_buffer[9];
+
+  if (data + data_sz <= data) return AOM_CODEC_INVALID_PARAM;
+
+  si->is_kf = 0;
+  si->w = si->h = 0;
+
+  if (decrypt_cb) {
+    data_sz = AOMMIN(sizeof(clear_buffer), data_sz);
+    decrypt_cb(decrypt_state, data, clear_buffer, data_sz);
+    data = clear_buffer;
+  }
+
+  {
+    int show_frame;
+    int error_resilient;
+    struct aom_read_bit_buffer rb = { data, data + data_sz, 0, NULL, NULL };
+    const int frame_marker = aom_rb_read_literal(&rb, 2);
+    const BITSTREAM_PROFILE profile = av1_read_profile(&rb);
+
+    if (frame_marker != AOM_FRAME_MARKER) return AOM_CODEC_UNSUP_BITSTREAM;
+
+    if (profile >= MAX_PROFILES) return AOM_CODEC_UNSUP_BITSTREAM;
+
+    if ((profile >= 2 && data_sz <= 1) || data_sz < 1)
+      return AOM_CODEC_UNSUP_BITSTREAM;
+
+    if (aom_rb_read_bit(&rb)) {     // show an existing frame
+      aom_rb_read_literal(&rb, 3);  // Frame buffer to show.
+      return AOM_CODEC_OK;
+    }
+
+    if (data_sz <= 8) return AOM_CODEC_UNSUP_BITSTREAM;
+
+    si->is_kf = !aom_rb_read_bit(&rb);
+    show_frame = aom_rb_read_bit(&rb);
+    error_resilient = aom_rb_read_bit(&rb);
+#if CONFIG_REFERENCE_BUFFER
+    {
+      /* TODO: Move outside frame loop or inside key-frame branch */
+      int frame_id_len;
+      SequenceHeader seq_params;
+      read_sequence_header(&seq_params);
+      if (seq_params.frame_id_numbers_present_flag) {
+        frame_id_len = seq_params.frame_id_length_minus7 + 7;
+        aom_rb_read_literal(&rb, frame_id_len);
+      }
+    }
+#endif
+    if (si->is_kf) {
+      if (!av1_read_sync_code(&rb)) return AOM_CODEC_UNSUP_BITSTREAM;
+
+      if (!parse_bitdepth_colorspace_sampling(profile, &rb))
+        return AOM_CODEC_UNSUP_BITSTREAM;
+      av1_read_frame_size(&rb, (int *)&si->w, (int *)&si->h);
+    } else {
+      intra_only_flag = show_frame ? 0 : aom_rb_read_bit(&rb);
+
+      rb.bit_offset += error_resilient ? 0 : 2;  // reset_frame_context
+
+      if (intra_only_flag) {
+        if (!av1_read_sync_code(&rb)) return AOM_CODEC_UNSUP_BITSTREAM;
+        if (profile > PROFILE_0) {
+          if (!parse_bitdepth_colorspace_sampling(profile, &rb))
+            return AOM_CODEC_UNSUP_BITSTREAM;
+        }
+        rb.bit_offset += REF_FRAMES;  // refresh_frame_flags
+        av1_read_frame_size(&rb, (int *)&si->w, (int *)&si->h);
+      }
+    }
+  }
+  if (is_intra_only != NULL) *is_intra_only = intra_only_flag;
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t decoder_peek_si(const uint8_t *data,
+                                       unsigned int data_sz,
+                                       aom_codec_stream_info_t *si) {
+  return decoder_peek_si_internal(data, data_sz, si, NULL, NULL, NULL);
+}
+
+static aom_codec_err_t decoder_get_si(aom_codec_alg_priv_t *ctx,
+                                      aom_codec_stream_info_t *si) {
+  const size_t sz = (si->sz >= sizeof(av1_stream_info_t))
+                        ? sizeof(av1_stream_info_t)
+                        : sizeof(aom_codec_stream_info_t);
+  memcpy(si, &ctx->si, sz);
+  si->sz = (unsigned int)sz;
+
+  return AOM_CODEC_OK;
+}
+
+static void set_error_detail(aom_codec_alg_priv_t *ctx,
+                             const char *const error) {
+  ctx->base.err_detail = error;
+}
+
+static aom_codec_err_t update_error_state(
+    aom_codec_alg_priv_t *ctx, const struct aom_internal_error_info *error) {
+  if (error->error_code)
+    set_error_detail(ctx, error->has_detail ? error->detail : NULL);
+
+  return error->error_code;
+}
+
+static void init_buffer_callbacks(aom_codec_alg_priv_t *ctx) {
+  int i;
+
+  for (i = 0; i < ctx->num_frame_workers; ++i) {
+    AVxWorker *const worker = &ctx->frame_workers[i];
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    AV1_COMMON *const cm = &frame_worker_data->pbi->common;
+    BufferPool *const pool = cm->buffer_pool;
+
+    cm->new_fb_idx = INVALID_IDX;
+    cm->byte_alignment = ctx->byte_alignment;
+    cm->skip_loop_filter = ctx->skip_loop_filter;
+
+    if (ctx->get_ext_fb_cb != NULL && ctx->release_ext_fb_cb != NULL) {
+      pool->get_fb_cb = ctx->get_ext_fb_cb;
+      pool->release_fb_cb = ctx->release_ext_fb_cb;
+      pool->cb_priv = ctx->ext_priv;
+    } else {
+      pool->get_fb_cb = av1_get_frame_buffer;
+      pool->release_fb_cb = av1_release_frame_buffer;
+
+      if (av1_alloc_internal_frame_buffers(&pool->int_frame_buffers))
+        aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                           "Failed to initialize internal frame buffers");
+
+      pool->cb_priv = &pool->int_frame_buffers;
+    }
+  }
+}
+
+static void set_default_ppflags(aom_postproc_cfg_t *cfg) {
+  cfg->post_proc_flag = AOM_DEBLOCK | AOM_DEMACROBLOCK;
+  cfg->deblocking_level = 4;
+  cfg->noise_level = 0;
+}
+
+static int frame_worker_hook(void *arg1, void *arg2) {
+  FrameWorkerData *const frame_worker_data = (FrameWorkerData *)arg1;
+  const uint8_t *data = frame_worker_data->data;
+  (void)arg2;
+
+  frame_worker_data->result = av1_receive_compressed_data(
+      frame_worker_data->pbi, frame_worker_data->data_size, &data);
+  frame_worker_data->data_end = data;
+
+  if (frame_worker_data->pbi->common.frame_parallel_decode) {
+    // In frame parallel decoding, a worker thread must successfully decode all
+    // the compressed data.
+    if (frame_worker_data->result != 0 ||
+        frame_worker_data->data + frame_worker_data->data_size - 1 > data) {
+      AVxWorker *const worker = frame_worker_data->pbi->frame_worker_owner;
+      BufferPool *const pool = frame_worker_data->pbi->common.buffer_pool;
+      // Signal all the other threads that are waiting for this frame.
+      av1_frameworker_lock_stats(worker);
+      frame_worker_data->frame_context_ready = 1;
+      lock_buffer_pool(pool);
+      frame_worker_data->pbi->cur_buf->buf.corrupted = 1;
+      unlock_buffer_pool(pool);
+      frame_worker_data->pbi->need_resync = 1;
+      av1_frameworker_signal_stats(worker);
+      av1_frameworker_unlock_stats(worker);
+      return 0;
+    }
+  } else if (frame_worker_data->result != 0) {
+    // Check decode result in serial decode.
+    frame_worker_data->pbi->cur_buf->buf.corrupted = 1;
+    frame_worker_data->pbi->need_resync = 1;
+  }
+  return !frame_worker_data->result;
+}
+
+static aom_codec_err_t init_decoder(aom_codec_alg_priv_t *ctx) {
+  int i;
+  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+
+  ctx->last_show_frame = -1;
+  ctx->next_submit_worker_id = 0;
+  ctx->last_submit_worker_id = 0;
+  ctx->next_output_worker_id = 0;
+  ctx->frame_cache_read = 0;
+  ctx->frame_cache_write = 0;
+  ctx->num_cache_frames = 0;
+  ctx->need_resync = 1;
+  ctx->num_frame_workers =
+      (ctx->frame_parallel_decode == 1) ? ctx->cfg.threads : 1;
+  if (ctx->num_frame_workers > MAX_DECODE_THREADS)
+    ctx->num_frame_workers = MAX_DECODE_THREADS;
+  ctx->available_threads = ctx->num_frame_workers;
+  ctx->flushed = 0;
+
+  ctx->buffer_pool = (BufferPool *)aom_calloc(1, sizeof(BufferPool));
+  if (ctx->buffer_pool == NULL) return AOM_CODEC_MEM_ERROR;
+
+#if CONFIG_MULTITHREAD
+  if (pthread_mutex_init(&ctx->buffer_pool->pool_mutex, NULL)) {
+    set_error_detail(ctx, "Failed to allocate buffer pool mutex");
+    return AOM_CODEC_MEM_ERROR;
+  }
+#endif
+
+  ctx->frame_workers = (AVxWorker *)aom_malloc(ctx->num_frame_workers *
+                                               sizeof(*ctx->frame_workers));
+  if (ctx->frame_workers == NULL) {
+    set_error_detail(ctx, "Failed to allocate frame_workers");
+    return AOM_CODEC_MEM_ERROR;
+  }
+
+  for (i = 0; i < ctx->num_frame_workers; ++i) {
+    AVxWorker *const worker = &ctx->frame_workers[i];
+    FrameWorkerData *frame_worker_data = NULL;
+    winterface->init(worker);
+    worker->data1 = aom_memalign(32, sizeof(FrameWorkerData));
+    if (worker->data1 == NULL) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data");
+      return AOM_CODEC_MEM_ERROR;
+    }
+    frame_worker_data = (FrameWorkerData *)worker->data1;
+    frame_worker_data->pbi = av1_decoder_create(ctx->buffer_pool);
+    if (frame_worker_data->pbi == NULL) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data");
+      return AOM_CODEC_MEM_ERROR;
+    }
+    frame_worker_data->pbi->frame_worker_owner = worker;
+    frame_worker_data->worker_id = i;
+    frame_worker_data->scratch_buffer = NULL;
+    frame_worker_data->scratch_buffer_size = 0;
+    frame_worker_data->frame_context_ready = 0;
+    frame_worker_data->received_frame = 0;
+#if CONFIG_MULTITHREAD
+    if (pthread_mutex_init(&frame_worker_data->stats_mutex, NULL)) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data mutex");
+      return AOM_CODEC_MEM_ERROR;
+    }
+
+    if (pthread_cond_init(&frame_worker_data->stats_cond, NULL)) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data cond");
+      return AOM_CODEC_MEM_ERROR;
+    }
+#endif
+    // If decoding in serial mode, FrameWorker thread could create tile worker
+    // thread or loopfilter thread.
+    frame_worker_data->pbi->max_threads =
+        (ctx->frame_parallel_decode == 0) ? ctx->cfg.threads : 0;
+
+    frame_worker_data->pbi->inv_tile_order = ctx->invert_tile_order;
+    frame_worker_data->pbi->common.frame_parallel_decode =
+        ctx->frame_parallel_decode;
+    worker->hook = (AVxWorkerHook)frame_worker_hook;
+    if (!winterface->reset(worker)) {
+      set_error_detail(ctx, "Frame Worker thread creation failed");
+      return AOM_CODEC_MEM_ERROR;
+    }
+  }
+
+  // If postprocessing was enabled by the application and a
+  // configuration has not been provided, default it.
+  if (!ctx->postproc_cfg_set && (ctx->base.init_flags & AOM_CODEC_USE_POSTPROC))
+    set_default_ppflags(&ctx->postproc_cfg);
+
+  init_buffer_callbacks(ctx);
+
+  return AOM_CODEC_OK;
+}
+
+static INLINE void check_resync(aom_codec_alg_priv_t *const ctx,
+                                const AV1Decoder *const pbi) {
+  // Clear resync flag if worker got a key frame or intra only frame.
+  if (ctx->need_resync == 1 && pbi->need_resync == 0 &&
+      (pbi->common.intra_only || pbi->common.frame_type == KEY_FRAME))
+    ctx->need_resync = 0;
+}
+
+static aom_codec_err_t decode_one(aom_codec_alg_priv_t *ctx,
+                                  const uint8_t **data, unsigned int data_sz,
+                                  void *user_priv, int64_t deadline) {
+  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+  (void)deadline;
+
+  // Determine the stream parameters. Note that we rely on peek_si to
+  // validate that we have a buffer that does not wrap around the top
+  // of the heap.
+  if (!ctx->si.h) {
+    int is_intra_only = 0;
+    const aom_codec_err_t res =
+        decoder_peek_si_internal(*data, data_sz, &ctx->si, &is_intra_only,
+                                 ctx->decrypt_cb, ctx->decrypt_state);
+    if (res != AOM_CODEC_OK) return res;
+
+    if (!ctx->si.is_kf && !is_intra_only) return AOM_CODEC_ERROR;
+  }
+
+  if (!ctx->frame_parallel_decode) {
+    AVxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    frame_worker_data->data = *data;
+    frame_worker_data->data_size = data_sz;
+    frame_worker_data->user_priv = user_priv;
+    frame_worker_data->received_frame = 1;
+
+    // Set these even if already initialized.  The caller may have changed the
+    // decrypt config between frames.
+    frame_worker_data->pbi->decrypt_cb = ctx->decrypt_cb;
+    frame_worker_data->pbi->decrypt_state = ctx->decrypt_state;
+#if CONFIG_INSPECTION
+    frame_worker_data->pbi->inspect_cb = ctx->inspect_cb;
+    frame_worker_data->pbi->inspect_ctx = ctx->inspect_ctx;
+#endif
+
+#if CONFIG_EXT_TILE
+    frame_worker_data->pbi->dec_tile_row = ctx->decode_tile_row;
+    frame_worker_data->pbi->dec_tile_col = ctx->decode_tile_col;
+#endif  // CONFIG_EXT_TILE
+
+    worker->had_error = 0;
+    winterface->execute(worker);
+
+    // Update data pointer after decode.
+    *data = frame_worker_data->data_end;
+
+    if (worker->had_error)
+      return update_error_state(ctx, &frame_worker_data->pbi->common.error);
+
+    check_resync(ctx, frame_worker_data->pbi);
+  } else {
+    AVxWorker *const worker = &ctx->frame_workers[ctx->next_submit_worker_id];
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    // Copy context from last worker thread to next worker thread.
+    if (ctx->next_submit_worker_id != ctx->last_submit_worker_id)
+      av1_frameworker_copy_context(
+          &ctx->frame_workers[ctx->next_submit_worker_id],
+          &ctx->frame_workers[ctx->last_submit_worker_id]);
+
+    frame_worker_data->pbi->ready_for_new_data = 0;
+    // Copy the compressed data into worker's internal buffer.
+    // TODO(hkuang): Will all the workers allocate the same size
+    // as the size of the first intra frame be better? This will
+    // avoid too many deallocate and allocate.
+    if (frame_worker_data->scratch_buffer_size < data_sz) {
+      aom_free(frame_worker_data->scratch_buffer);
+      frame_worker_data->scratch_buffer = (uint8_t *)aom_malloc(data_sz);
+      if (frame_worker_data->scratch_buffer == NULL) {
+        set_error_detail(ctx, "Failed to reallocate scratch buffer");
+        return AOM_CODEC_MEM_ERROR;
+      }
+      frame_worker_data->scratch_buffer_size = data_sz;
+    }
+    frame_worker_data->data_size = data_sz;
+    memcpy(frame_worker_data->scratch_buffer, *data, data_sz);
+
+    frame_worker_data->frame_decoded = 0;
+    frame_worker_data->frame_context_ready = 0;
+    frame_worker_data->received_frame = 1;
+    frame_worker_data->data = frame_worker_data->scratch_buffer;
+    frame_worker_data->user_priv = user_priv;
+
+    if (ctx->next_submit_worker_id != ctx->last_submit_worker_id)
+      ctx->last_submit_worker_id =
+          (ctx->last_submit_worker_id + 1) % ctx->num_frame_workers;
+
+    ctx->next_submit_worker_id =
+        (ctx->next_submit_worker_id + 1) % ctx->num_frame_workers;
+    --ctx->available_threads;
+    worker->had_error = 0;
+    winterface->launch(worker);
+  }
+
+  return AOM_CODEC_OK;
+}
+
+static void wait_worker_and_cache_frame(aom_codec_alg_priv_t *ctx) {
+  YV12_BUFFER_CONFIG sd;
+  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+  AVxWorker *const worker = &ctx->frame_workers[ctx->next_output_worker_id];
+  FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+  ctx->next_output_worker_id =
+      (ctx->next_output_worker_id + 1) % ctx->num_frame_workers;
+  // TODO(hkuang): Add worker error handling here.
+  winterface->sync(worker);
+  frame_worker_data->received_frame = 0;
+  ++ctx->available_threads;
+
+  check_resync(ctx, frame_worker_data->pbi);
+
+  if (av1_get_raw_frame(frame_worker_data->pbi, &sd) == 0) {
+    AV1_COMMON *const cm = &frame_worker_data->pbi->common;
+    RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+    ctx->frame_cache[ctx->frame_cache_write].fb_idx = cm->new_fb_idx;
+    yuvconfig2image(&ctx->frame_cache[ctx->frame_cache_write].img, &sd,
+                    frame_worker_data->user_priv);
+    ctx->frame_cache[ctx->frame_cache_write].img.fb_priv =
+        frame_bufs[cm->new_fb_idx].raw_frame_buffer.priv;
+    ctx->frame_cache_write = (ctx->frame_cache_write + 1) % FRAME_CACHE_SIZE;
+    ++ctx->num_cache_frames;
+  }
+}
+
+static aom_codec_err_t decoder_decode(aom_codec_alg_priv_t *ctx,
+                                      const uint8_t *data, unsigned int data_sz,
+                                      void *user_priv, long deadline) {
+  const uint8_t *data_start = data;
+  const uint8_t *const data_end = data + data_sz;
+  aom_codec_err_t res;
+  uint32_t frame_sizes[8];
+  int frame_count;
+
+  if (data == NULL && data_sz == 0) {
+    ctx->flushed = 1;
+    return AOM_CODEC_OK;
+  }
+
+  // Reset flushed when receiving a valid frame.
+  ctx->flushed = 0;
+
+  // Initialize the decoder workers on the first frame.
+  if (ctx->frame_workers == NULL) {
+    res = init_decoder(ctx);
+    if (res != AOM_CODEC_OK) return res;
+  }
+
+  res = av1_parse_superframe_index(data, data_sz, frame_sizes, &frame_count,
+                                   ctx->decrypt_cb, ctx->decrypt_state);
+  if (res != AOM_CODEC_OK) return res;
+
+  if (ctx->frame_parallel_decode) {
+    // Decode in frame parallel mode. When decoding in this mode, the frame
+    // passed to the decoder must be either a normal frame or a superframe with
+    // superframe index so the decoder could get each frame's start position
+    // in the superframe.
+    if (frame_count > 0) {
+      int i;
+
+      for (i = 0; i < frame_count; ++i) {
+        const uint8_t *data_start_copy = data_start;
+        const uint32_t frame_size = frame_sizes[i];
+        if (data_start < data ||
+            frame_size > (uint32_t)(data_end - data_start)) {
+          set_error_detail(ctx, "Invalid frame size in index");
+          return AOM_CODEC_CORRUPT_FRAME;
+        }
+
+        if (ctx->available_threads == 0) {
+          // No more threads for decoding. Wait until the next output worker
+          // finishes decoding. Then copy the decoded frame into cache.
+          if (ctx->num_cache_frames < FRAME_CACHE_SIZE) {
+            wait_worker_and_cache_frame(ctx);
+          } else {
+            // TODO(hkuang): Add unit test to test this path.
+            set_error_detail(ctx, "Frame output cache is full.");
+            return AOM_CODEC_ERROR;
+          }
+        }
+
+        res =
+            decode_one(ctx, &data_start_copy, frame_size, user_priv, deadline);
+        if (res != AOM_CODEC_OK) return res;
+        data_start += frame_size;
+      }
+    } else {
+      if (ctx->available_threads == 0) {
+        // No more threads for decoding. Wait until the next output worker
+        // finishes decoding. Then copy the decoded frame into cache.
+        if (ctx->num_cache_frames < FRAME_CACHE_SIZE) {
+          wait_worker_and_cache_frame(ctx);
+        } else {
+          // TODO(hkuang): Add unit test to test this path.
+          set_error_detail(ctx, "Frame output cache is full.");
+          return AOM_CODEC_ERROR;
+        }
+      }
+
+      res = decode_one(ctx, &data, data_sz, user_priv, deadline);
+      if (res != AOM_CODEC_OK) return res;
+    }
+  } else {
+    // Decode in serial mode.
+    if (frame_count > 0) {
+      int i;
+
+      for (i = 0; i < frame_count; ++i) {
+        const uint8_t *data_start_copy = data_start;
+        const uint32_t frame_size = frame_sizes[i];
+        if (data_start < data ||
+            frame_size > (uint32_t)(data_end - data_start)) {
+          set_error_detail(ctx, "Invalid frame size in index");
+          return AOM_CODEC_CORRUPT_FRAME;
+        }
+
+        res =
+            decode_one(ctx, &data_start_copy, frame_size, user_priv, deadline);
+        if (res != AOM_CODEC_OK) return res;
+
+        data_start += frame_size;
+      }
+    } else {
+      while (data_start < data_end) {
+        const uint32_t frame_size = (uint32_t)(data_end - data_start);
+        res = decode_one(ctx, &data_start, frame_size, user_priv, deadline);
+        if (res != AOM_CODEC_OK) return res;
+
+        // Account for suboptimal termination by the encoder.
+        while (data_start < data_end) {
+          const uint8_t marker =
+              read_marker(ctx->decrypt_cb, ctx->decrypt_state, data_start);
+          if (marker) break;
+          ++data_start;
+        }
+      }
+    }
+  }
+
+  return res;
+}
+
+static void release_last_output_frame(aom_codec_alg_priv_t *ctx) {
+  RefCntBuffer *const frame_bufs = ctx->buffer_pool->frame_bufs;
+  // Decrease reference count of last output frame in frame parallel mode.
+  if (ctx->frame_parallel_decode && ctx->last_show_frame >= 0) {
+    BufferPool *const pool = ctx->buffer_pool;
+    lock_buffer_pool(pool);
+    decrease_ref_count(ctx->last_show_frame, frame_bufs, pool);
+    unlock_buffer_pool(pool);
+  }
+}
+
+static aom_image_t *decoder_get_frame(aom_codec_alg_priv_t *ctx,
+                                      aom_codec_iter_t *iter) {
+  aom_image_t *img = NULL;
+
+  // Only return frame when all the cpu are busy or
+  // application fluhsed the decoder in frame parallel decode.
+  if (ctx->frame_parallel_decode && ctx->available_threads > 0 &&
+      !ctx->flushed) {
+    return NULL;
+  }
+
+  // Output the frames in the cache first.
+  if (ctx->num_cache_frames > 0) {
+    release_last_output_frame(ctx);
+    ctx->last_show_frame = ctx->frame_cache[ctx->frame_cache_read].fb_idx;
+    if (ctx->need_resync) return NULL;
+    img = &ctx->frame_cache[ctx->frame_cache_read].img;
+    ctx->frame_cache_read = (ctx->frame_cache_read + 1) % FRAME_CACHE_SIZE;
+    --ctx->num_cache_frames;
+    return img;
+  }
+
+  // iter acts as a flip flop, so an image is only returned on the first
+  // call to get_frame.
+  if (*iter == NULL && ctx->frame_workers != NULL) {
+    do {
+      YV12_BUFFER_CONFIG sd;
+      const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+      AVxWorker *const worker = &ctx->frame_workers[ctx->next_output_worker_id];
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      ctx->next_output_worker_id =
+          (ctx->next_output_worker_id + 1) % ctx->num_frame_workers;
+      // Wait for the frame from worker thread.
+      if (winterface->sync(worker)) {
+        // Check if worker has received any frames.
+        if (frame_worker_data->received_frame == 1) {
+          ++ctx->available_threads;
+          frame_worker_data->received_frame = 0;
+          check_resync(ctx, frame_worker_data->pbi);
+        }
+        if (av1_get_raw_frame(frame_worker_data->pbi, &sd) == 0) {
+          AV1_COMMON *const cm = &frame_worker_data->pbi->common;
+          RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+          release_last_output_frame(ctx);
+          ctx->last_show_frame = frame_worker_data->pbi->common.new_fb_idx;
+          if (ctx->need_resync) return NULL;
+          yuvconfig2image(&ctx->img, &sd, frame_worker_data->user_priv);
+
+#if CONFIG_EXT_TILE
+          if (cm->tile_encoding_mode &&
+              frame_worker_data->pbi->dec_tile_row >= 0) {
+            const int tile_row =
+                AOMMIN(frame_worker_data->pbi->dec_tile_row, cm->tile_rows - 1);
+            const int mi_row = tile_row * cm->tile_height;
+            const int ssy = ctx->img.y_chroma_shift;
+            int plane;
+            ctx->img.planes[0] += mi_row * MI_SIZE * ctx->img.stride[0];
+            for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
+              ctx->img.planes[plane] +=
+                  mi_row * (MI_SIZE >> ssy) * ctx->img.stride[plane];
+            }
+            ctx->img.d_h =
+                AOMMIN(cm->tile_height, cm->mi_rows - mi_row) * MI_SIZE;
+          }
+
+          if (cm->tile_encoding_mode &&
+              frame_worker_data->pbi->dec_tile_col >= 0) {
+            const int tile_col =
+                AOMMIN(frame_worker_data->pbi->dec_tile_col, cm->tile_cols - 1);
+            const int mi_col = tile_col * cm->tile_width;
+            const int ssx = ctx->img.x_chroma_shift;
+            int plane;
+            ctx->img.planes[0] += mi_col * MI_SIZE;
+            for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
+              ctx->img.planes[plane] += mi_col * (MI_SIZE >> ssx);
+            }
+            ctx->img.d_w =
+                AOMMIN(cm->tile_width, cm->mi_cols - mi_col) * MI_SIZE;
+          }
+#endif  // CONFIG_EXT_TILE
+
+          ctx->img.fb_priv = frame_bufs[cm->new_fb_idx].raw_frame_buffer.priv;
+          img = &ctx->img;
+          return img;
+        }
+      } else {
+        // Decoding failed. Release the worker thread.
+        frame_worker_data->received_frame = 0;
+        ++ctx->available_threads;
+        ctx->need_resync = 1;
+        if (ctx->flushed != 1) return NULL;
+      }
+    } while (ctx->next_output_worker_id != ctx->next_submit_worker_id);
+  }
+  return NULL;
+}
+
+static aom_codec_err_t decoder_set_fb_fn(
+    aom_codec_alg_priv_t *ctx, aom_get_frame_buffer_cb_fn_t cb_get,
+    aom_release_frame_buffer_cb_fn_t cb_release, void *cb_priv) {
+  if (cb_get == NULL || cb_release == NULL) {
+    return AOM_CODEC_INVALID_PARAM;
+  } else if (ctx->frame_workers == NULL) {
+    // If the decoder has already been initialized, do not accept changes to
+    // the frame buffer functions.
+    ctx->get_ext_fb_cb = cb_get;
+    ctx->release_ext_fb_cb = cb_release;
+    ctx->ext_priv = cb_priv;
+    return AOM_CODEC_OK;
+  }
+
+  return AOM_CODEC_ERROR;
+}
+
+static aom_codec_err_t ctrl_set_reference(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  aom_ref_frame_t *const data = va_arg(args, aom_ref_frame_t *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return AOM_CODEC_INCAPABLE;
+  }
+
+  if (data) {
+    aom_ref_frame_t *const frame = (aom_ref_frame_t *)data;
+    YV12_BUFFER_CONFIG sd;
+    AVxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    image2yuvconfig(&frame->img, &sd);
+    return av1_set_reference_dec(&frame_worker_data->pbi->common,
+                                 ref_frame_to_av1_reframe(frame->frame_type),
+                                 &sd);
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_copy_reference(aom_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  const aom_ref_frame_t *const frame = va_arg(args, aom_ref_frame_t *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return AOM_CODEC_INCAPABLE;
+  }
+
+  if (frame) {
+    YV12_BUFFER_CONFIG sd;
+    AVxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    image2yuvconfig(&frame->img, &sd);
+    return av1_copy_reference_dec(frame_worker_data->pbi,
+                                  (AOM_REFFRAME)frame->frame_type, &sd);
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_get_reference(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  av1_ref_frame_t *data = va_arg(args, av1_ref_frame_t *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return AOM_CODEC_INCAPABLE;
+  }
+
+  if (data) {
+    YV12_BUFFER_CONFIG *fb;
+    AVxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    fb = get_ref_frame(&frame_worker_data->pbi->common, data->idx);
+    if (fb == NULL) return AOM_CODEC_ERROR;
+    yuvconfig2image(&data->img, fb, NULL);
+    return AOM_CODEC_OK;
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_get_new_frame_image(aom_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  aom_image_t *new_img = va_arg(args, aom_image_t *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return AOM_CODEC_INCAPABLE;
+  }
+
+  if (new_img) {
+    YV12_BUFFER_CONFIG new_frame;
+    AVxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+
+    if (av1_get_frame_to_show(frame_worker_data->pbi, &new_frame) == 0) {
+      yuvconfig2image(new_img, &new_frame, NULL);
+      return AOM_CODEC_OK;
+    } else {
+      return AOM_CODEC_ERROR;
+    }
+  } else {
+    return AOM_CODEC_INVALID_PARAM;
+  }
+}
+
+static aom_codec_err_t ctrl_set_postproc(aom_codec_alg_priv_t *ctx,
+                                         va_list args) {
+  (void)ctx;
+  (void)args;
+  return AOM_CODEC_INCAPABLE;
+}
+
+static aom_codec_err_t ctrl_set_dbg_options(aom_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  (void)ctx;
+  (void)args;
+  return AOM_CODEC_INCAPABLE;
+}
+
+static aom_codec_err_t ctrl_get_last_ref_updates(aom_codec_alg_priv_t *ctx,
+                                                 va_list args) {
+  int *const update_info = va_arg(args, int *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return AOM_CODEC_INCAPABLE;
+  }
+
+  if (update_info) {
+    if (ctx->frame_workers) {
+      AVxWorker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      *update_info = frame_worker_data->pbi->refresh_frame_flags;
+      return AOM_CODEC_OK;
+    } else {
+      return AOM_CODEC_ERROR;
+    }
+  }
+
+  return AOM_CODEC_INVALID_PARAM;
+}
+
+static aom_codec_err_t ctrl_get_last_quantizer(aom_codec_alg_priv_t *ctx,
+                                               va_list args) {
+  int *const arg = va_arg(args, int *);
+  if (arg == NULL) return AOM_CODEC_INVALID_PARAM;
+  *arg =
+      ((FrameWorkerData *)ctx->frame_workers[0].data1)->pbi->common.base_qindex;
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t ctrl_get_frame_corrupted(aom_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  int *corrupted = va_arg(args, int *);
+
+  if (corrupted) {
+    if (ctx->frame_workers) {
+      AVxWorker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      RefCntBuffer *const frame_bufs =
+          frame_worker_data->pbi->common.buffer_pool->frame_bufs;
+      if (frame_worker_data->pbi->common.frame_to_show == NULL)
+        return AOM_CODEC_ERROR;
+      if (ctx->last_show_frame >= 0)
+        *corrupted = frame_bufs[ctx->last_show_frame].buf.corrupted;
+      return AOM_CODEC_OK;
+    } else {
+      return AOM_CODEC_ERROR;
+    }
+  }
+
+  return AOM_CODEC_INVALID_PARAM;
+}
+
+static aom_codec_err_t ctrl_get_frame_size(aom_codec_alg_priv_t *ctx,
+                                           va_list args) {
+  int *const frame_size = va_arg(args, int *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return AOM_CODEC_INCAPABLE;
+  }
+
+  if (frame_size) {
+    if (ctx->frame_workers) {
+      AVxWorker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      const AV1_COMMON *const cm = &frame_worker_data->pbi->common;
+      frame_size[0] = cm->width;
+      frame_size[1] = cm->height;
+      return AOM_CODEC_OK;
+    } else {
+      return AOM_CODEC_ERROR;
+    }
+  }
+
+  return AOM_CODEC_INVALID_PARAM;
+}
+
+static aom_codec_err_t ctrl_get_render_size(aom_codec_alg_priv_t *ctx,
+                                            va_list args) {
+  int *const render_size = va_arg(args, int *);
+
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return AOM_CODEC_INCAPABLE;
+  }
+
+  if (render_size) {
+    if (ctx->frame_workers) {
+      AVxWorker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      const AV1_COMMON *const cm = &frame_worker_data->pbi->common;
+      render_size[0] = cm->render_width;
+      render_size[1] = cm->render_height;
+      return AOM_CODEC_OK;
+    } else {
+      return AOM_CODEC_ERROR;
+    }
+  }
+
+  return AOM_CODEC_INVALID_PARAM;
+}
+
+static aom_codec_err_t ctrl_get_bit_depth(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  unsigned int *const bit_depth = va_arg(args, unsigned int *);
+  AVxWorker *const worker = &ctx->frame_workers[ctx->next_output_worker_id];
+
+  if (bit_depth) {
+    if (worker) {
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      const AV1_COMMON *const cm = &frame_worker_data->pbi->common;
+      *bit_depth = cm->bit_depth;
+      return AOM_CODEC_OK;
+    } else {
+      return AOM_CODEC_ERROR;
+    }
+  }
+
+  return AOM_CODEC_INVALID_PARAM;
+}
+
+static aom_codec_err_t ctrl_set_invert_tile_order(aom_codec_alg_priv_t *ctx,
+                                                  va_list args) {
+  ctx->invert_tile_order = va_arg(args, int);
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t ctrl_set_decryptor(aom_codec_alg_priv_t *ctx,
+                                          va_list args) {
+  aom_decrypt_init *init = va_arg(args, aom_decrypt_init *);
+  ctx->decrypt_cb = init ? init->decrypt_cb : NULL;
+  ctx->decrypt_state = init ? init->decrypt_state : NULL;
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t ctrl_set_byte_alignment(aom_codec_alg_priv_t *ctx,
+                                               va_list args) {
+  const int legacy_byte_alignment = 0;
+  const int min_byte_alignment = 32;
+  const int max_byte_alignment = 1024;
+  const int byte_alignment = va_arg(args, int);
+
+  if (byte_alignment != legacy_byte_alignment &&
+      (byte_alignment < min_byte_alignment ||
+       byte_alignment > max_byte_alignment ||
+       (byte_alignment & (byte_alignment - 1)) != 0))
+    return AOM_CODEC_INVALID_PARAM;
+
+  ctx->byte_alignment = byte_alignment;
+  if (ctx->frame_workers) {
+    AVxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    frame_worker_data->pbi->common.byte_alignment = byte_alignment;
+  }
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t ctrl_set_skip_loop_filter(aom_codec_alg_priv_t *ctx,
+                                                 va_list args) {
+  ctx->skip_loop_filter = va_arg(args, int);
+
+  if (ctx->frame_workers) {
+    AVxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    frame_worker_data->pbi->common.skip_loop_filter = ctx->skip_loop_filter;
+  }
+
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t ctrl_get_accounting(aom_codec_alg_priv_t *ctx,
+                                           va_list args) {
+#if !CONFIG_ACCOUNTING
+  (void)ctx;
+  (void)args;
+  return AOM_CODEC_INCAPABLE;
+#else
+  if (ctx->frame_workers) {
+    AVxWorker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    AV1Decoder *pbi = frame_worker_data->pbi;
+    Accounting **acct = va_arg(args, Accounting **);
+    *acct = &pbi->accounting;
+    return AOM_CODEC_OK;
+  }
+  return AOM_CODEC_ERROR;
+#endif
+}
+static aom_codec_err_t ctrl_set_decode_tile_row(aom_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  ctx->decode_tile_row = va_arg(args, int);
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t ctrl_set_decode_tile_col(aom_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  ctx->decode_tile_col = va_arg(args, int);
+  return AOM_CODEC_OK;
+}
+
+static aom_codec_err_t ctrl_set_inspection_callback(aom_codec_alg_priv_t *ctx,
+                                                    va_list args) {
+#if !CONFIG_INSPECTION
+  (void)ctx;
+  (void)args;
+  return AOM_CODEC_INCAPABLE;
+#else
+  aom_inspect_init *init = va_arg(args, aom_inspect_init *);
+  ctx->inspect_cb = init->inspect_cb;
+  ctx->inspect_ctx = init->inspect_ctx;
+  return AOM_CODEC_OK;
+#endif
+}
+
+static aom_codec_ctrl_fn_map_t decoder_ctrl_maps[] = {
+  { AOM_COPY_REFERENCE, ctrl_copy_reference },
+
+  // Setters
+  { AOM_SET_REFERENCE, ctrl_set_reference },
+  { AOM_SET_POSTPROC, ctrl_set_postproc },
+  { AOM_SET_DBG_COLOR_REF_FRAME, ctrl_set_dbg_options },
+  { AOM_SET_DBG_COLOR_MB_MODES, ctrl_set_dbg_options },
+  { AOM_SET_DBG_COLOR_B_MODES, ctrl_set_dbg_options },
+  { AOM_SET_DBG_DISPLAY_MV, ctrl_set_dbg_options },
+  { AV1_INVERT_TILE_DECODE_ORDER, ctrl_set_invert_tile_order },
+  { AOMD_SET_DECRYPTOR, ctrl_set_decryptor },
+  { AV1_SET_BYTE_ALIGNMENT, ctrl_set_byte_alignment },
+  { AV1_SET_SKIP_LOOP_FILTER, ctrl_set_skip_loop_filter },
+  { AV1_SET_DECODE_TILE_ROW, ctrl_set_decode_tile_row },
+  { AV1_SET_DECODE_TILE_COL, ctrl_set_decode_tile_col },
+  { AV1_SET_INSPECTION_CALLBACK, ctrl_set_inspection_callback },
+
+  // Getters
+  { AOMD_GET_FRAME_CORRUPTED, ctrl_get_frame_corrupted },
+  { AOMD_GET_LAST_QUANTIZER, ctrl_get_last_quantizer },
+  { AOMD_GET_LAST_REF_UPDATES, ctrl_get_last_ref_updates },
+  { AV1D_GET_BIT_DEPTH, ctrl_get_bit_depth },
+  { AV1D_GET_DISPLAY_SIZE, ctrl_get_render_size },
+  { AV1D_GET_FRAME_SIZE, ctrl_get_frame_size },
+  { AV1_GET_ACCOUNTING, ctrl_get_accounting },
+  { AV1_GET_NEW_FRAME_IMAGE, ctrl_get_new_frame_image },
+  { AV1_GET_REFERENCE, ctrl_get_reference },
+
+  { -1, NULL },
+};
+
+#ifndef VERSION_STRING
+#define VERSION_STRING
+#endif
+CODEC_INTERFACE(aom_codec_av1_dx) = {
+  "AOMedia Project AV1 Decoder" VERSION_STRING,
+  AOM_CODEC_INTERNAL_ABI_VERSION,
+  AOM_CODEC_CAP_DECODER |
+      AOM_CODEC_CAP_EXTERNAL_FRAME_BUFFER,  // aom_codec_caps_t
+  decoder_init,                             // aom_codec_init_fn_t
+  decoder_destroy,                          // aom_codec_destroy_fn_t
+  decoder_ctrl_maps,                        // aom_codec_ctrl_fn_map_t
+  {
+      // NOLINT
+      decoder_peek_si,    // aom_codec_peek_si_fn_t
+      decoder_get_si,     // aom_codec_get_si_fn_t
+      decoder_decode,     // aom_codec_decode_fn_t
+      decoder_get_frame,  // aom_codec_frame_get_fn_t
+      decoder_set_fb_fn,  // aom_codec_set_fb_fn_t
+  },
+  {
+      // NOLINT
+      0,
+      NULL,  // aom_codec_enc_cfg_map_t
+      NULL,  // aom_codec_encode_fn_t
+      NULL,  // aom_codec_get_cx_data_fn_t
+      NULL,  // aom_codec_enc_config_set_fn_t
+      NULL,  // aom_codec_get_global_headers_fn_t
+      NULL,  // aom_codec_get_preview_frame_fn_t
+      NULL   // aom_codec_enc_mr_get_mem_loc_fn_t
+  }
+};
diff --git a/third_party/aom/av1/av1_iface_common.h b/third_party/aom/av1/av1_iface_common.h
new file mode 100644
index 0000000000..df3614212b
--- /dev/null
+++ b/third_party/aom/av1/av1_iface_common.h
@@ -0,0 +1,146 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AV1_AV1_IFACE_COMMON_H_
+#define AV1_AV1_IFACE_COMMON_H_
+
+#include "aom_ports/mem.h"
+
+static void yuvconfig2image(aom_image_t *img, const YV12_BUFFER_CONFIG *yv12,
+                            void *user_priv) {
+  /** aom_img_wrap() doesn't allow specifying independent strides for
+    * the Y, U, and V planes, nor other alignment adjustments that
+    * might be representable by a YV12_BUFFER_CONFIG, so we just
+    * initialize all the fields.*/
+  int bps;
+  if (!yv12->subsampling_y) {
+    if (!yv12->subsampling_x) {
+      img->fmt = AOM_IMG_FMT_I444;
+      bps = 24;
+    } else {
+      img->fmt = AOM_IMG_FMT_I422;
+      bps = 16;
+    }
+  } else {
+    if (!yv12->subsampling_x) {
+      img->fmt = AOM_IMG_FMT_I440;
+      bps = 16;
+    } else {
+      img->fmt = AOM_IMG_FMT_I420;
+      bps = 12;
+    }
+  }
+  img->cs = yv12->color_space;
+  img->range = yv12->color_range;
+  img->bit_depth = 8;
+  img->w = yv12->y_stride;
+  img->h = ALIGN_POWER_OF_TWO(yv12->y_height + 2 * AOM_BORDER_IN_PIXELS, 3);
+  img->d_w = yv12->y_crop_width;
+  img->d_h = yv12->y_crop_height;
+  img->r_w = yv12->render_width;
+  img->r_h = yv12->render_height;
+  img->x_chroma_shift = yv12->subsampling_x;
+  img->y_chroma_shift = yv12->subsampling_y;
+  img->planes[AOM_PLANE_Y] = yv12->y_buffer;
+  img->planes[AOM_PLANE_U] = yv12->u_buffer;
+  img->planes[AOM_PLANE_V] = yv12->v_buffer;
+  img->planes[AOM_PLANE_ALPHA] = NULL;
+  img->stride[AOM_PLANE_Y] = yv12->y_stride;
+  img->stride[AOM_PLANE_U] = yv12->uv_stride;
+  img->stride[AOM_PLANE_V] = yv12->uv_stride;
+  img->stride[AOM_PLANE_ALPHA] = yv12->y_stride;
+#if CONFIG_HIGHBITDEPTH
+  if (yv12->flags & YV12_FLAG_HIGHBITDEPTH) {
+    // aom_image_t uses byte strides and a pointer to the first byte
+    // of the image.
+    img->fmt = (aom_img_fmt_t)(img->fmt | AOM_IMG_FMT_HIGHBITDEPTH);
+    img->bit_depth = yv12->bit_depth;
+    img->planes[AOM_PLANE_Y] = (uint8_t *)CONVERT_TO_SHORTPTR(yv12->y_buffer);
+    img->planes[AOM_PLANE_U] = (uint8_t *)CONVERT_TO_SHORTPTR(yv12->u_buffer);
+    img->planes[AOM_PLANE_V] = (uint8_t *)CONVERT_TO_SHORTPTR(yv12->v_buffer);
+    img->planes[AOM_PLANE_ALPHA] = NULL;
+    img->stride[AOM_PLANE_Y] = 2 * yv12->y_stride;
+    img->stride[AOM_PLANE_U] = 2 * yv12->uv_stride;
+    img->stride[AOM_PLANE_V] = 2 * yv12->uv_stride;
+    img->stride[AOM_PLANE_ALPHA] = 2 * yv12->y_stride;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  img->bps = bps;
+  img->user_priv = user_priv;
+  img->img_data = yv12->buffer_alloc;
+  img->img_data_owner = 0;
+  img->self_allocd = 0;
+}
+
+static aom_codec_err_t image2yuvconfig(const aom_image_t *img,
+                                       YV12_BUFFER_CONFIG *yv12) {
+  yv12->y_buffer = img->planes[AOM_PLANE_Y];
+  yv12->u_buffer = img->planes[AOM_PLANE_U];
+  yv12->v_buffer = img->planes[AOM_PLANE_V];
+
+  yv12->y_crop_width = img->d_w;
+  yv12->y_crop_height = img->d_h;
+  yv12->render_width = img->r_w;
+  yv12->render_height = img->r_h;
+  yv12->y_width = img->d_w;
+  yv12->y_height = img->d_h;
+
+  yv12->uv_width =
+      img->x_chroma_shift == 1 ? (1 + yv12->y_width) / 2 : yv12->y_width;
+  yv12->uv_height =
+      img->y_chroma_shift == 1 ? (1 + yv12->y_height) / 2 : yv12->y_height;
+  yv12->uv_crop_width = yv12->uv_width;
+  yv12->uv_crop_height = yv12->uv_height;
+
+  yv12->y_stride = img->stride[AOM_PLANE_Y];
+  yv12->uv_stride = img->stride[AOM_PLANE_U];
+  yv12->color_space = img->cs;
+  yv12->color_range = img->range;
+
+#if CONFIG_HIGHBITDEPTH
+  if (img->fmt & AOM_IMG_FMT_HIGHBITDEPTH) {
+    // In aom_image_t
+    //     planes point to uint8 address of start of data
+    //     stride counts uint8s to reach next row
+    // In YV12_BUFFER_CONFIG
+    //     y_buffer, u_buffer, v_buffer point to uint16 address of data
+    //     stride and border counts in uint16s
+    // This means that all the address calculations in the main body of code
+    // should work correctly.
+    // However, before we do any pixel operations we need to cast the address
+    // to a uint16 ponter and double its value.
+    yv12->y_buffer = CONVERT_TO_BYTEPTR(yv12->y_buffer);
+    yv12->u_buffer = CONVERT_TO_BYTEPTR(yv12->u_buffer);
+    yv12->v_buffer = CONVERT_TO_BYTEPTR(yv12->v_buffer);
+    yv12->y_stride >>= 1;
+    yv12->uv_stride >>= 1;
+    yv12->flags = YV12_FLAG_HIGHBITDEPTH;
+  } else {
+    yv12->flags = 0;
+  }
+  yv12->border = (yv12->y_stride - img->w) / 2;
+#else
+  yv12->border = (img->stride[AOM_PLANE_Y] - img->w) / 2;
+#endif  // CONFIG_HIGHBITDEPTH
+  yv12->subsampling_x = img->x_chroma_shift;
+  yv12->subsampling_y = img->y_chroma_shift;
+  return AOM_CODEC_OK;
+}
+
+static AOM_REFFRAME ref_frame_to_av1_reframe(aom_ref_frame_type_t frame) {
+  switch (frame) {
+    case AOM_LAST_FRAME: return AOM_LAST_FLAG;
+    case AOM_GOLD_FRAME: return AOM_GOLD_FLAG;
+    case AOM_ALTR_FRAME: return AOM_ALT_FLAG;
+  }
+  assert(0 && "Invalid Reference Frame");
+  return AOM_LAST_FLAG;
+}
+#endif  // AV1_AV1_IFACE_COMMON_H_
diff --git a/third_party/aom/av1/common/alloccommon.c b/third_party/aom/av1/common/alloccommon.c
new file mode 100644
index 0000000000..79d41a9c82
--- /dev/null
+++ b/third_party/aom/av1/common/alloccommon.c
@@ -0,0 +1,209 @@
+/*
+ *
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "aom_mem/aom_mem.h"
+
+#include "av1/common/alloccommon.h"
+#include "av1/common/blockd.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/entropymv.h"
+#include "av1/common/onyxc_int.h"
+
+void av1_set_mb_mi(AV1_COMMON *cm, int width, int height) {
+  // TODO(jingning): Fine tune the loop filter operations and bring this
+  // back to integer multiple of 4 for cb4x4.
+  const int aligned_width = ALIGN_POWER_OF_TWO(width, 3);
+  const int aligned_height = ALIGN_POWER_OF_TWO(height, 3);
+
+  cm->mi_cols = aligned_width >> MI_SIZE_LOG2;
+  cm->mi_rows = aligned_height >> MI_SIZE_LOG2;
+  cm->mi_stride = calc_mi_size(cm->mi_cols);
+
+#if CONFIG_CB4X4
+  cm->mb_cols = (cm->mi_cols + 2) >> 2;
+  cm->mb_rows = (cm->mi_rows + 2) >> 2;
+#else
+  cm->mb_cols = (cm->mi_cols + 1) >> 1;
+  cm->mb_rows = (cm->mi_rows + 1) >> 1;
+#endif
+  cm->MBs = cm->mb_rows * cm->mb_cols;
+}
+
+static int alloc_seg_map(AV1_COMMON *cm, int seg_map_size) {
+  int i;
+
+  for (i = 0; i < NUM_PING_PONG_BUFFERS; ++i) {
+    cm->seg_map_array[i] = (uint8_t *)aom_calloc(seg_map_size, 1);
+    if (cm->seg_map_array[i] == NULL) return 1;
+  }
+  cm->seg_map_alloc_size = seg_map_size;
+
+  // Init the index.
+  cm->seg_map_idx = 0;
+  cm->prev_seg_map_idx = 1;
+
+  cm->current_frame_seg_map = cm->seg_map_array[cm->seg_map_idx];
+  if (!cm->frame_parallel_decode)
+    cm->last_frame_seg_map = cm->seg_map_array[cm->prev_seg_map_idx];
+
+  return 0;
+}
+
+static void free_seg_map(AV1_COMMON *cm) {
+  int i;
+
+  for (i = 0; i < NUM_PING_PONG_BUFFERS; ++i) {
+    aom_free(cm->seg_map_array[i]);
+    cm->seg_map_array[i] = NULL;
+  }
+
+  cm->current_frame_seg_map = NULL;
+
+  if (!cm->frame_parallel_decode) {
+    cm->last_frame_seg_map = NULL;
+  }
+}
+
+void av1_free_ref_frame_buffers(BufferPool *pool) {
+  int i;
+
+  for (i = 0; i < FRAME_BUFFERS; ++i) {
+    if (pool->frame_bufs[i].ref_count > 0 &&
+        pool->frame_bufs[i].raw_frame_buffer.data != NULL) {
+      pool->release_fb_cb(pool->cb_priv, &pool->frame_bufs[i].raw_frame_buffer);
+      pool->frame_bufs[i].ref_count = 0;
+    }
+    aom_free(pool->frame_bufs[i].mvs);
+    pool->frame_bufs[i].mvs = NULL;
+    aom_free_frame_buffer(&pool->frame_bufs[i].buf);
+  }
+}
+
+#if CONFIG_LOOP_RESTORATION
+// Assumes cm->rst_info[p].restoration_tilesize is already initialized
+void av1_alloc_restoration_buffers(AV1_COMMON *cm) {
+  int p;
+  av1_alloc_restoration_struct(cm, &cm->rst_info[0], cm->width, cm->height);
+  for (p = 1; p < MAX_MB_PLANE; ++p)
+    av1_alloc_restoration_struct(
+        cm, &cm->rst_info[p], ROUND_POWER_OF_TWO(cm->width, cm->subsampling_x),
+        ROUND_POWER_OF_TWO(cm->height, cm->subsampling_y));
+  aom_free(cm->rst_internal.tmpbuf);
+  CHECK_MEM_ERROR(cm, cm->rst_internal.tmpbuf,
+                  (int32_t *)aom_memalign(16, RESTORATION_TMPBUF_SIZE));
+}
+
+void av1_free_restoration_buffers(AV1_COMMON *cm) {
+  int p;
+  for (p = 0; p < MAX_MB_PLANE; ++p)
+    av1_free_restoration_struct(&cm->rst_info[p]);
+  aom_free(cm->rst_internal.tmpbuf);
+  cm->rst_internal.tmpbuf = NULL;
+}
+#endif  // CONFIG_LOOP_RESTORATION
+
+void av1_free_context_buffers(AV1_COMMON *cm) {
+  int i;
+  cm->free_mi(cm);
+  free_seg_map(cm);
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    aom_free(cm->above_context[i]);
+    cm->above_context[i] = NULL;
+  }
+  aom_free(cm->above_seg_context);
+  cm->above_seg_context = NULL;
+#if CONFIG_VAR_TX
+  aom_free(cm->above_txfm_context);
+  cm->above_txfm_context = NULL;
+#endif
+}
+
+int av1_alloc_context_buffers(AV1_COMMON *cm, int width, int height) {
+  int new_mi_size;
+
+  av1_set_mb_mi(cm, width, height);
+  new_mi_size = cm->mi_stride * calc_mi_size(cm->mi_rows);
+  if (cm->mi_alloc_size < new_mi_size) {
+    cm->free_mi(cm);
+    if (cm->alloc_mi(cm, new_mi_size)) goto fail;
+  }
+
+  if (cm->seg_map_alloc_size < cm->mi_rows * cm->mi_cols) {
+    // Create the segmentation map structure and set to 0.
+    free_seg_map(cm);
+    if (alloc_seg_map(cm, cm->mi_rows * cm->mi_cols)) goto fail;
+  }
+
+  if (cm->above_context_alloc_cols < cm->mi_cols) {
+    // TODO(geza.lore): These are bigger than they need to be.
+    // cm->tile_width would be enough but it complicates indexing a
+    // little elsewhere.
+    const int aligned_mi_cols =
+        ALIGN_POWER_OF_TWO(cm->mi_cols, MAX_MIB_SIZE_LOG2);
+    int i;
+
+    for (i = 0; i < MAX_MB_PLANE; i++) {
+      aom_free(cm->above_context[i]);
+      cm->above_context[i] = (ENTROPY_CONTEXT *)aom_calloc(
+          2 * aligned_mi_cols, sizeof(*cm->above_context[0]));
+      if (!cm->above_context[i]) goto fail;
+    }
+
+    aom_free(cm->above_seg_context);
+    cm->above_seg_context = (PARTITION_CONTEXT *)aom_calloc(
+        aligned_mi_cols, sizeof(*cm->above_seg_context));
+    if (!cm->above_seg_context) goto fail;
+
+#if CONFIG_VAR_TX
+    aom_free(cm->above_txfm_context);
+    cm->above_txfm_context = (TXFM_CONTEXT *)aom_calloc(
+        aligned_mi_cols, sizeof(*cm->above_txfm_context));
+    if (!cm->above_txfm_context) goto fail;
+#endif
+
+    cm->above_context_alloc_cols = aligned_mi_cols;
+  }
+
+  return 0;
+
+fail:
+  // clear the mi_* values to force a realloc on resync
+  av1_set_mb_mi(cm, 0, 0);
+  av1_free_context_buffers(cm);
+  return 1;
+}
+
+void av1_remove_common(AV1_COMMON *cm) {
+  av1_free_context_buffers(cm);
+
+  aom_free(cm->fc);
+  cm->fc = NULL;
+  aom_free(cm->frame_contexts);
+  cm->frame_contexts = NULL;
+}
+
+void av1_init_context_buffers(AV1_COMMON *cm) {
+  cm->setup_mi(cm);
+  if (cm->last_frame_seg_map && !cm->frame_parallel_decode)
+    memset(cm->last_frame_seg_map, 0, cm->mi_rows * cm->mi_cols);
+}
+
+void av1_swap_current_and_last_seg_map(AV1_COMMON *cm) {
+  // Swap indices.
+  const int tmp = cm->seg_map_idx;
+  cm->seg_map_idx = cm->prev_seg_map_idx;
+  cm->prev_seg_map_idx = tmp;
+
+  cm->current_frame_seg_map = cm->seg_map_array[cm->seg_map_idx];
+  cm->last_frame_seg_map = cm->seg_map_array[cm->prev_seg_map_idx];
+}
diff --git a/third_party/aom/av1/common/alloccommon.h b/third_party/aom/av1/common/alloccommon.h
new file mode 100644
index 0000000000..51863cd045
--- /dev/null
+++ b/third_party/aom/av1/common/alloccommon.h
@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_ALLOCCOMMON_H_
+#define AV1_COMMON_ALLOCCOMMON_H_
+
+#define INVALID_IDX -1  // Invalid buffer index.
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1Common;
+struct BufferPool;
+
+void av1_remove_common(struct AV1Common *cm);
+
+int av1_alloc_context_buffers(struct AV1Common *cm, int width, int height);
+void av1_init_context_buffers(struct AV1Common *cm);
+void av1_free_context_buffers(struct AV1Common *cm);
+
+void av1_free_ref_frame_buffers(struct BufferPool *pool);
+#if CONFIG_LOOP_RESTORATION
+void av1_alloc_restoration_buffers(struct AV1Common *cm);
+void av1_free_restoration_buffers(struct AV1Common *cm);
+#endif  // CONFIG_LOOP_RESTORATION
+
+int av1_alloc_state_buffers(struct AV1Common *cm, int width, int height);
+void av1_free_state_buffers(struct AV1Common *cm);
+
+void av1_set_mb_mi(struct AV1Common *cm, int width, int height);
+
+void av1_swap_current_and_last_seg_map(struct AV1Common *cm);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_ALLOCCOMMON_H_
diff --git a/third_party/aom/av1/common/arm/neon/iht4x4_add_neon.c b/third_party/aom/av1/common/arm/neon/iht4x4_add_neon.c
new file mode 100644
index 0000000000..02572d4059
--- /dev/null
+++ b/third_party/aom/av1/common/arm/neon/iht4x4_add_neon.c
@@ -0,0 +1,227 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "aom_dsp/txfm_common.h"
+#include "av1/common/common.h"
+
+static INLINE void TRANSPOSE4X4(int16x8_t *q8s16, int16x8_t *q9s16) {
+  int32x4_t q8s32, q9s32;
+  int16x4x2_t d0x2s16, d1x2s16;
+  int32x4x2_t q0x2s32;
+
+  d0x2s16 = vtrn_s16(vget_low_s16(*q8s16), vget_high_s16(*q8s16));
+  d1x2s16 = vtrn_s16(vget_low_s16(*q9s16), vget_high_s16(*q9s16));
+
+  q8s32 = vreinterpretq_s32_s16(vcombine_s16(d0x2s16.val[0], d0x2s16.val[1]));
+  q9s32 = vreinterpretq_s32_s16(vcombine_s16(d1x2s16.val[0], d1x2s16.val[1]));
+  q0x2s32 = vtrnq_s32(q8s32, q9s32);
+
+  *q8s16 = vreinterpretq_s16_s32(q0x2s32.val[0]);
+  *q9s16 = vreinterpretq_s16_s32(q0x2s32.val[1]);
+  return;
+}
+
+static INLINE void GENERATE_COSINE_CONSTANTS(int16x4_t *d0s16, int16x4_t *d1s16,
+                                             int16x4_t *d2s16) {
+  *d0s16 = vdup_n_s16((int16_t)cospi_8_64);
+  *d1s16 = vdup_n_s16((int16_t)cospi_16_64);
+  *d2s16 = vdup_n_s16((int16_t)cospi_24_64);
+  return;
+}
+
+static INLINE void GENERATE_SINE_CONSTANTS(int16x4_t *d3s16, int16x4_t *d4s16,
+                                           int16x4_t *d5s16, int16x8_t *q3s16) {
+  *d3s16 = vdup_n_s16((int16_t)sinpi_1_9);
+  *d4s16 = vdup_n_s16((int16_t)sinpi_2_9);
+  *q3s16 = vdupq_n_s16((int16_t)sinpi_3_9);
+  *d5s16 = vdup_n_s16((int16_t)sinpi_4_9);
+  return;
+}
+
+static INLINE void IDCT4x4_1D(int16x4_t *d0s16, int16x4_t *d1s16,
+                              int16x4_t *d2s16, int16x8_t *q8s16,
+                              int16x8_t *q9s16) {
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d23s16, d24s16;
+  int16x4_t d26s16, d27s16, d28s16, d29s16;
+  int32x4_t q10s32, q13s32, q14s32, q15s32;
+  int16x8_t q13s16, q14s16;
+
+  d16s16 = vget_low_s16(*q8s16);
+  d17s16 = vget_high_s16(*q8s16);
+  d18s16 = vget_low_s16(*q9s16);
+  d19s16 = vget_high_s16(*q9s16);
+
+  d23s16 = vadd_s16(d16s16, d18s16);
+  d24s16 = vsub_s16(d16s16, d18s16);
+
+  q15s32 = vmull_s16(d17s16, *d2s16);
+  q10s32 = vmull_s16(d17s16, *d0s16);
+  q13s32 = vmull_s16(d23s16, *d1s16);
+  q14s32 = vmull_s16(d24s16, *d1s16);
+  q15s32 = vmlsl_s16(q15s32, d19s16, *d0s16);
+  q10s32 = vmlal_s16(q10s32, d19s16, *d2s16);
+
+  d26s16 = vqrshrn_n_s32(q13s32, 14);
+  d27s16 = vqrshrn_n_s32(q14s32, 14);
+  d29s16 = vqrshrn_n_s32(q15s32, 14);
+  d28s16 = vqrshrn_n_s32(q10s32, 14);
+
+  q13s16 = vcombine_s16(d26s16, d27s16);
+  q14s16 = vcombine_s16(d28s16, d29s16);
+  *q8s16 = vaddq_s16(q13s16, q14s16);
+  *q9s16 = vsubq_s16(q13s16, q14s16);
+  *q9s16 = vcombine_s16(vget_high_s16(*q9s16), vget_low_s16(*q9s16));  // vswp
+  return;
+}
+
+static INLINE void IADST4x4_1D(int16x4_t *d3s16, int16x4_t *d4s16,
+                               int16x4_t *d5s16, int16x8_t *q3s16,
+                               int16x8_t *q8s16, int16x8_t *q9s16) {
+  int16x4_t d6s16, d16s16, d17s16, d18s16, d19s16;
+  int32x4_t q8s32, q9s32, q10s32, q11s32, q12s32, q13s32, q14s32, q15s32;
+
+  d6s16 = vget_low_s16(*q3s16);
+
+  d16s16 = vget_low_s16(*q8s16);
+  d17s16 = vget_high_s16(*q8s16);
+  d18s16 = vget_low_s16(*q9s16);
+  d19s16 = vget_high_s16(*q9s16);
+
+  q10s32 = vmull_s16(*d3s16, d16s16);
+  q11s32 = vmull_s16(*d4s16, d16s16);
+  q12s32 = vmull_s16(d6s16, d17s16);
+  q13s32 = vmull_s16(*d5s16, d18s16);
+  q14s32 = vmull_s16(*d3s16, d18s16);
+  q15s32 = vmovl_s16(d16s16);
+  q15s32 = vaddw_s16(q15s32, d19s16);
+  q8s32 = vmull_s16(*d4s16, d19s16);
+  q15s32 = vsubw_s16(q15s32, d18s16);
+  q9s32 = vmull_s16(*d5s16, d19s16);
+
+  q10s32 = vaddq_s32(q10s32, q13s32);
+  q10s32 = vaddq_s32(q10s32, q8s32);
+  q11s32 = vsubq_s32(q11s32, q14s32);
+  q8s32 = vdupq_n_s32((int32_t)sinpi_3_9);
+  q11s32 = vsubq_s32(q11s32, q9s32);
+  q15s32 = vmulq_s32(q15s32, q8s32);
+
+  q13s32 = vaddq_s32(q10s32, q12s32);
+  q10s32 = vaddq_s32(q10s32, q11s32);
+  q14s32 = vaddq_s32(q11s32, q12s32);
+  q10s32 = vsubq_s32(q10s32, q12s32);
+
+  d16s16 = vqrshrn_n_s32(q13s32, 14);
+  d17s16 = vqrshrn_n_s32(q14s32, 14);
+  d18s16 = vqrshrn_n_s32(q15s32, 14);
+  d19s16 = vqrshrn_n_s32(q10s32, 14);
+
+  *q8s16 = vcombine_s16(d16s16, d17s16);
+  *q9s16 = vcombine_s16(d18s16, d19s16);
+  return;
+}
+
+void av1_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest,
+                            int dest_stride, int tx_type) {
+  uint8x8_t d26u8, d27u8;
+  int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16;
+  uint32x2_t d26u32, d27u32;
+  int16x8_t q3s16, q8s16, q9s16;
+  uint16x8_t q8u16, q9u16;
+
+  d26u32 = d27u32 = vdup_n_u32(0);
+
+  q8s16 = vld1q_s16(input);
+  q9s16 = vld1q_s16(input + 8);
+
+  TRANSPOSE4X4(&q8s16, &q9s16);
+
+  switch (tx_type) {
+    case 0:  // idct_idct is not supported. Fall back to C
+      av1_iht4x4_16_add_c(input, dest, dest_stride, tx_type);
+      return;
+      break;
+    case 1:  // iadst_idct
+      // generate constants
+      GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
+      GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
+
+      // first transform rows
+      IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
+
+      // transpose the matrix
+      TRANSPOSE4X4(&q8s16, &q9s16);
+
+      // then transform columns
+      IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+      break;
+    case 2:  // idct_iadst
+      // generate constantsyy
+      GENERATE_COSINE_CONSTANTS(&d0s16, &d1s16, &d2s16);
+      GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
+
+      // first transform rows
+      IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+
+      // transpose the matrix
+      TRANSPOSE4X4(&q8s16, &q9s16);
+
+      // then transform columns
+      IDCT4x4_1D(&d0s16, &d1s16, &d2s16, &q8s16, &q9s16);
+      break;
+    case 3:  // iadst_iadst
+      // generate constants
+      GENERATE_SINE_CONSTANTS(&d3s16, &d4s16, &d5s16, &q3s16);
+
+      // first transform rows
+      IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+
+      // transpose the matrix
+      TRANSPOSE4X4(&q8s16, &q9s16);
+
+      // then transform columns
+      IADST4x4_1D(&d3s16, &d4s16, &d5s16, &q3s16, &q8s16, &q9s16);
+      break;
+    default:  // iadst_idct
+      assert(0);
+      break;
+  }
+
+  q8s16 = vrshrq_n_s16(q8s16, 4);
+  q9s16 = vrshrq_n_s16(q9s16, 4);
+
+  d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 0);
+  dest += dest_stride;
+  d26u32 = vld1_lane_u32((const uint32_t *)dest, d26u32, 1);
+  dest += dest_stride;
+  d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 0);
+  dest += dest_stride;
+  d27u32 = vld1_lane_u32((const uint32_t *)dest, d27u32, 1);
+
+  q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u32(d26u32));
+  q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u32(d27u32));
+
+  d26u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+  d27u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+
+  vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 1);
+  dest -= dest_stride;
+  vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d27u8), 0);
+  dest -= dest_stride;
+  vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 1);
+  dest -= dest_stride;
+  vst1_lane_u32((uint32_t *)dest, vreinterpret_u32_u8(d26u8), 0);
+  return;
+}
diff --git a/third_party/aom/av1/common/arm/neon/iht8x8_add_neon.c b/third_party/aom/av1/common/arm/neon/iht8x8_add_neon.c
new file mode 100644
index 0000000000..86798ccf14
--- /dev/null
+++ b/third_party/aom/av1/common/arm/neon/iht8x8_add_neon.c
@@ -0,0 +1,593 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "aom_dsp/txfm_common.h"
+#include "av1/common/common.h"
+
+static INLINE void TRANSPOSE8X8(int16x8_t *q8s16, int16x8_t *q9s16,
+                                int16x8_t *q10s16, int16x8_t *q11s16,
+                                int16x8_t *q12s16, int16x8_t *q13s16,
+                                int16x8_t *q14s16, int16x8_t *q15s16) {
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  int32x4x2_t q0x2s32, q1x2s32, q2x2s32, q3x2s32;
+  int16x8x2_t q0x2s16, q1x2s16, q2x2s16, q3x2s16;
+
+  d16s16 = vget_low_s16(*q8s16);
+  d17s16 = vget_high_s16(*q8s16);
+  d18s16 = vget_low_s16(*q9s16);
+  d19s16 = vget_high_s16(*q9s16);
+  d20s16 = vget_low_s16(*q10s16);
+  d21s16 = vget_high_s16(*q10s16);
+  d22s16 = vget_low_s16(*q11s16);
+  d23s16 = vget_high_s16(*q11s16);
+  d24s16 = vget_low_s16(*q12s16);
+  d25s16 = vget_high_s16(*q12s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+  d30s16 = vget_low_s16(*q15s16);
+  d31s16 = vget_high_s16(*q15s16);
+
+  *q8s16 = vcombine_s16(d16s16, d24s16);   // vswp d17, d24
+  *q9s16 = vcombine_s16(d18s16, d26s16);   // vswp d19, d26
+  *q10s16 = vcombine_s16(d20s16, d28s16);  // vswp d21, d28
+  *q11s16 = vcombine_s16(d22s16, d30s16);  // vswp d23, d30
+  *q12s16 = vcombine_s16(d17s16, d25s16);
+  *q13s16 = vcombine_s16(d19s16, d27s16);
+  *q14s16 = vcombine_s16(d21s16, d29s16);
+  *q15s16 = vcombine_s16(d23s16, d31s16);
+
+  q0x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q8s16), vreinterpretq_s32_s16(*q10s16));
+  q1x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q9s16), vreinterpretq_s32_s16(*q11s16));
+  q2x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q12s16), vreinterpretq_s32_s16(*q14s16));
+  q3x2s32 =
+      vtrnq_s32(vreinterpretq_s32_s16(*q13s16), vreinterpretq_s32_s16(*q15s16));
+
+  q0x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[0]),   // q8
+                      vreinterpretq_s16_s32(q1x2s32.val[0]));  // q9
+  q1x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q0x2s32.val[1]),   // q10
+                      vreinterpretq_s16_s32(q1x2s32.val[1]));  // q11
+  q2x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[0]),   // q12
+                      vreinterpretq_s16_s32(q3x2s32.val[0]));  // q13
+  q3x2s16 = vtrnq_s16(vreinterpretq_s16_s32(q2x2s32.val[1]),   // q14
+                      vreinterpretq_s16_s32(q3x2s32.val[1]));  // q15
+
+  *q8s16 = q0x2s16.val[0];
+  *q9s16 = q0x2s16.val[1];
+  *q10s16 = q1x2s16.val[0];
+  *q11s16 = q1x2s16.val[1];
+  *q12s16 = q2x2s16.val[0];
+  *q13s16 = q2x2s16.val[1];
+  *q14s16 = q3x2s16.val[0];
+  *q15s16 = q3x2s16.val[1];
+  return;
+}
+
+static INLINE void IDCT8x8_1D(int16x8_t *q8s16, int16x8_t *q9s16,
+                              int16x8_t *q10s16, int16x8_t *q11s16,
+                              int16x8_t *q12s16, int16x8_t *q13s16,
+                              int16x8_t *q14s16, int16x8_t *q15s16) {
+  int16x4_t d0s16, d1s16, d2s16, d3s16;
+  int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  int16x8_t q0s16, q1s16, q2s16, q3s16, q4s16, q5s16, q6s16, q7s16;
+  int32x4_t q2s32, q3s32, q5s32, q6s32, q8s32, q9s32;
+  int32x4_t q10s32, q11s32, q12s32, q13s32, q15s32;
+
+  d0s16 = vdup_n_s16((int16_t)cospi_28_64);
+  d1s16 = vdup_n_s16((int16_t)cospi_4_64);
+  d2s16 = vdup_n_s16((int16_t)cospi_12_64);
+  d3s16 = vdup_n_s16((int16_t)cospi_20_64);
+
+  d16s16 = vget_low_s16(*q8s16);
+  d17s16 = vget_high_s16(*q8s16);
+  d18s16 = vget_low_s16(*q9s16);
+  d19s16 = vget_high_s16(*q9s16);
+  d20s16 = vget_low_s16(*q10s16);
+  d21s16 = vget_high_s16(*q10s16);
+  d22s16 = vget_low_s16(*q11s16);
+  d23s16 = vget_high_s16(*q11s16);
+  d24s16 = vget_low_s16(*q12s16);
+  d25s16 = vget_high_s16(*q12s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+  d30s16 = vget_low_s16(*q15s16);
+  d31s16 = vget_high_s16(*q15s16);
+
+  q2s32 = vmull_s16(d18s16, d0s16);
+  q3s32 = vmull_s16(d19s16, d0s16);
+  q5s32 = vmull_s16(d26s16, d2s16);
+  q6s32 = vmull_s16(d27s16, d2s16);
+
+  q2s32 = vmlsl_s16(q2s32, d30s16, d1s16);
+  q3s32 = vmlsl_s16(q3s32, d31s16, d1s16);
+  q5s32 = vmlsl_s16(q5s32, d22s16, d3s16);
+  q6s32 = vmlsl_s16(q6s32, d23s16, d3s16);
+
+  d8s16 = vqrshrn_n_s32(q2s32, 14);
+  d9s16 = vqrshrn_n_s32(q3s32, 14);
+  d10s16 = vqrshrn_n_s32(q5s32, 14);
+  d11s16 = vqrshrn_n_s32(q6s32, 14);
+  q4s16 = vcombine_s16(d8s16, d9s16);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+
+  q2s32 = vmull_s16(d18s16, d1s16);
+  q3s32 = vmull_s16(d19s16, d1s16);
+  q9s32 = vmull_s16(d26s16, d3s16);
+  q13s32 = vmull_s16(d27s16, d3s16);
+
+  q2s32 = vmlal_s16(q2s32, d30s16, d0s16);
+  q3s32 = vmlal_s16(q3s32, d31s16, d0s16);
+  q9s32 = vmlal_s16(q9s32, d22s16, d2s16);
+  q13s32 = vmlal_s16(q13s32, d23s16, d2s16);
+
+  d14s16 = vqrshrn_n_s32(q2s32, 14);
+  d15s16 = vqrshrn_n_s32(q3s32, 14);
+  d12s16 = vqrshrn_n_s32(q9s32, 14);
+  d13s16 = vqrshrn_n_s32(q13s32, 14);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+  q7s16 = vcombine_s16(d14s16, d15s16);
+
+  d0s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q2s32 = vmull_s16(d16s16, d0s16);
+  q3s32 = vmull_s16(d17s16, d0s16);
+  q13s32 = vmull_s16(d16s16, d0s16);
+  q15s32 = vmull_s16(d17s16, d0s16);
+
+  q2s32 = vmlal_s16(q2s32, d24s16, d0s16);
+  q3s32 = vmlal_s16(q3s32, d25s16, d0s16);
+  q13s32 = vmlsl_s16(q13s32, d24s16, d0s16);
+  q15s32 = vmlsl_s16(q15s32, d25s16, d0s16);
+
+  d0s16 = vdup_n_s16((int16_t)cospi_24_64);
+  d1s16 = vdup_n_s16((int16_t)cospi_8_64);
+
+  d18s16 = vqrshrn_n_s32(q2s32, 14);
+  d19s16 = vqrshrn_n_s32(q3s32, 14);
+  d22s16 = vqrshrn_n_s32(q13s32, 14);
+  d23s16 = vqrshrn_n_s32(q15s32, 14);
+  *q9s16 = vcombine_s16(d18s16, d19s16);
+  *q11s16 = vcombine_s16(d22s16, d23s16);
+
+  q2s32 = vmull_s16(d20s16, d0s16);
+  q3s32 = vmull_s16(d21s16, d0s16);
+  q8s32 = vmull_s16(d20s16, d1s16);
+  q12s32 = vmull_s16(d21s16, d1s16);
+
+  q2s32 = vmlsl_s16(q2s32, d28s16, d1s16);
+  q3s32 = vmlsl_s16(q3s32, d29s16, d1s16);
+  q8s32 = vmlal_s16(q8s32, d28s16, d0s16);
+  q12s32 = vmlal_s16(q12s32, d29s16, d0s16);
+
+  d26s16 = vqrshrn_n_s32(q2s32, 14);
+  d27s16 = vqrshrn_n_s32(q3s32, 14);
+  d30s16 = vqrshrn_n_s32(q8s32, 14);
+  d31s16 = vqrshrn_n_s32(q12s32, 14);
+  *q13s16 = vcombine_s16(d26s16, d27s16);
+  *q15s16 = vcombine_s16(d30s16, d31s16);
+
+  q0s16 = vaddq_s16(*q9s16, *q15s16);
+  q1s16 = vaddq_s16(*q11s16, *q13s16);
+  q2s16 = vsubq_s16(*q11s16, *q13s16);
+  q3s16 = vsubq_s16(*q9s16, *q15s16);
+
+  *q13s16 = vsubq_s16(q4s16, q5s16);
+  q4s16 = vaddq_s16(q4s16, q5s16);
+  *q14s16 = vsubq_s16(q7s16, q6s16);
+  q7s16 = vaddq_s16(q7s16, q6s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+
+  d16s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  q9s32 = vmull_s16(d28s16, d16s16);
+  q10s32 = vmull_s16(d29s16, d16s16);
+  q11s32 = vmull_s16(d28s16, d16s16);
+  q12s32 = vmull_s16(d29s16, d16s16);
+
+  q9s32 = vmlsl_s16(q9s32, d26s16, d16s16);
+  q10s32 = vmlsl_s16(q10s32, d27s16, d16s16);
+  q11s32 = vmlal_s16(q11s32, d26s16, d16s16);
+  q12s32 = vmlal_s16(q12s32, d27s16, d16s16);
+
+  d10s16 = vqrshrn_n_s32(q9s32, 14);
+  d11s16 = vqrshrn_n_s32(q10s32, 14);
+  d12s16 = vqrshrn_n_s32(q11s32, 14);
+  d13s16 = vqrshrn_n_s32(q12s32, 14);
+  q5s16 = vcombine_s16(d10s16, d11s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  *q8s16 = vaddq_s16(q0s16, q7s16);
+  *q9s16 = vaddq_s16(q1s16, q6s16);
+  *q10s16 = vaddq_s16(q2s16, q5s16);
+  *q11s16 = vaddq_s16(q3s16, q4s16);
+  *q12s16 = vsubq_s16(q3s16, q4s16);
+  *q13s16 = vsubq_s16(q2s16, q5s16);
+  *q14s16 = vsubq_s16(q1s16, q6s16);
+  *q15s16 = vsubq_s16(q0s16, q7s16);
+  return;
+}
+
+static INLINE void IADST8X8_1D(int16x8_t *q8s16, int16x8_t *q9s16,
+                               int16x8_t *q10s16, int16x8_t *q11s16,
+                               int16x8_t *q12s16, int16x8_t *q13s16,
+                               int16x8_t *q14s16, int16x8_t *q15s16) {
+  int16x4_t d0s16, d1s16, d2s16, d3s16, d4s16, d5s16, d6s16, d7s16;
+  int16x4_t d8s16, d9s16, d10s16, d11s16, d12s16, d13s16, d14s16, d15s16;
+  int16x4_t d16s16, d17s16, d18s16, d19s16, d20s16, d21s16, d22s16, d23s16;
+  int16x4_t d24s16, d25s16, d26s16, d27s16, d28s16, d29s16, d30s16, d31s16;
+  int16x8_t q2s16, q4s16, q5s16, q6s16;
+  int32x4_t q0s32, q1s32, q2s32, q3s32, q4s32, q5s32, q6s32, q7s32, q8s32;
+  int32x4_t q9s32, q10s32, q11s32, q12s32, q13s32, q14s32, q15s32;
+
+  d16s16 = vget_low_s16(*q8s16);
+  d17s16 = vget_high_s16(*q8s16);
+  d18s16 = vget_low_s16(*q9s16);
+  d19s16 = vget_high_s16(*q9s16);
+  d20s16 = vget_low_s16(*q10s16);
+  d21s16 = vget_high_s16(*q10s16);
+  d22s16 = vget_low_s16(*q11s16);
+  d23s16 = vget_high_s16(*q11s16);
+  d24s16 = vget_low_s16(*q12s16);
+  d25s16 = vget_high_s16(*q12s16);
+  d26s16 = vget_low_s16(*q13s16);
+  d27s16 = vget_high_s16(*q13s16);
+  d28s16 = vget_low_s16(*q14s16);
+  d29s16 = vget_high_s16(*q14s16);
+  d30s16 = vget_low_s16(*q15s16);
+  d31s16 = vget_high_s16(*q15s16);
+
+  d14s16 = vdup_n_s16((int16_t)cospi_2_64);
+  d15s16 = vdup_n_s16((int16_t)cospi_30_64);
+
+  q1s32 = vmull_s16(d30s16, d14s16);
+  q2s32 = vmull_s16(d31s16, d14s16);
+  q3s32 = vmull_s16(d30s16, d15s16);
+  q4s32 = vmull_s16(d31s16, d15s16);
+
+  d30s16 = vdup_n_s16((int16_t)cospi_18_64);
+  d31s16 = vdup_n_s16((int16_t)cospi_14_64);
+
+  q1s32 = vmlal_s16(q1s32, d16s16, d15s16);
+  q2s32 = vmlal_s16(q2s32, d17s16, d15s16);
+  q3s32 = vmlsl_s16(q3s32, d16s16, d14s16);
+  q4s32 = vmlsl_s16(q4s32, d17s16, d14s16);
+
+  q5s32 = vmull_s16(d22s16, d30s16);
+  q6s32 = vmull_s16(d23s16, d30s16);
+  q7s32 = vmull_s16(d22s16, d31s16);
+  q8s32 = vmull_s16(d23s16, d31s16);
+
+  q5s32 = vmlal_s16(q5s32, d24s16, d31s16);
+  q6s32 = vmlal_s16(q6s32, d25s16, d31s16);
+  q7s32 = vmlsl_s16(q7s32, d24s16, d30s16);
+  q8s32 = vmlsl_s16(q8s32, d25s16, d30s16);
+
+  q11s32 = vaddq_s32(q1s32, q5s32);
+  q12s32 = vaddq_s32(q2s32, q6s32);
+  q1s32 = vsubq_s32(q1s32, q5s32);
+  q2s32 = vsubq_s32(q2s32, q6s32);
+
+  d22s16 = vqrshrn_n_s32(q11s32, 14);
+  d23s16 = vqrshrn_n_s32(q12s32, 14);
+  *q11s16 = vcombine_s16(d22s16, d23s16);
+
+  q12s32 = vaddq_s32(q3s32, q7s32);
+  q15s32 = vaddq_s32(q4s32, q8s32);
+  q3s32 = vsubq_s32(q3s32, q7s32);
+  q4s32 = vsubq_s32(q4s32, q8s32);
+
+  d2s16 = vqrshrn_n_s32(q1s32, 14);
+  d3s16 = vqrshrn_n_s32(q2s32, 14);
+  d24s16 = vqrshrn_n_s32(q12s32, 14);
+  d25s16 = vqrshrn_n_s32(q15s32, 14);
+  d6s16 = vqrshrn_n_s32(q3s32, 14);
+  d7s16 = vqrshrn_n_s32(q4s32, 14);
+  *q12s16 = vcombine_s16(d24s16, d25s16);
+
+  d0s16 = vdup_n_s16((int16_t)cospi_10_64);
+  d1s16 = vdup_n_s16((int16_t)cospi_22_64);
+  q4s32 = vmull_s16(d26s16, d0s16);
+  q5s32 = vmull_s16(d27s16, d0s16);
+  q2s32 = vmull_s16(d26s16, d1s16);
+  q6s32 = vmull_s16(d27s16, d1s16);
+
+  d30s16 = vdup_n_s16((int16_t)cospi_26_64);
+  d31s16 = vdup_n_s16((int16_t)cospi_6_64);
+
+  q4s32 = vmlal_s16(q4s32, d20s16, d1s16);
+  q5s32 = vmlal_s16(q5s32, d21s16, d1s16);
+  q2s32 = vmlsl_s16(q2s32, d20s16, d0s16);
+  q6s32 = vmlsl_s16(q6s32, d21s16, d0s16);
+
+  q0s32 = vmull_s16(d18s16, d30s16);
+  q13s32 = vmull_s16(d19s16, d30s16);
+
+  q0s32 = vmlal_s16(q0s32, d28s16, d31s16);
+  q13s32 = vmlal_s16(q13s32, d29s16, d31s16);
+
+  q10s32 = vmull_s16(d18s16, d31s16);
+  q9s32 = vmull_s16(d19s16, d31s16);
+
+  q10s32 = vmlsl_s16(q10s32, d28s16, d30s16);
+  q9s32 = vmlsl_s16(q9s32, d29s16, d30s16);
+
+  q14s32 = vaddq_s32(q2s32, q10s32);
+  q15s32 = vaddq_s32(q6s32, q9s32);
+  q2s32 = vsubq_s32(q2s32, q10s32);
+  q6s32 = vsubq_s32(q6s32, q9s32);
+
+  d28s16 = vqrshrn_n_s32(q14s32, 14);
+  d29s16 = vqrshrn_n_s32(q15s32, 14);
+  d4s16 = vqrshrn_n_s32(q2s32, 14);
+  d5s16 = vqrshrn_n_s32(q6s32, 14);
+  *q14s16 = vcombine_s16(d28s16, d29s16);
+
+  q9s32 = vaddq_s32(q4s32, q0s32);
+  q10s32 = vaddq_s32(q5s32, q13s32);
+  q4s32 = vsubq_s32(q4s32, q0s32);
+  q5s32 = vsubq_s32(q5s32, q13s32);
+
+  d30s16 = vdup_n_s16((int16_t)cospi_8_64);
+  d31s16 = vdup_n_s16((int16_t)cospi_24_64);
+
+  d18s16 = vqrshrn_n_s32(q9s32, 14);
+  d19s16 = vqrshrn_n_s32(q10s32, 14);
+  d8s16 = vqrshrn_n_s32(q4s32, 14);
+  d9s16 = vqrshrn_n_s32(q5s32, 14);
+  *q9s16 = vcombine_s16(d18s16, d19s16);
+
+  q5s32 = vmull_s16(d2s16, d30s16);
+  q6s32 = vmull_s16(d3s16, d30s16);
+  q7s32 = vmull_s16(d2s16, d31s16);
+  q0s32 = vmull_s16(d3s16, d31s16);
+
+  q5s32 = vmlal_s16(q5s32, d6s16, d31s16);
+  q6s32 = vmlal_s16(q6s32, d7s16, d31s16);
+  q7s32 = vmlsl_s16(q7s32, d6s16, d30s16);
+  q0s32 = vmlsl_s16(q0s32, d7s16, d30s16);
+
+  q1s32 = vmull_s16(d4s16, d30s16);
+  q3s32 = vmull_s16(d5s16, d30s16);
+  q10s32 = vmull_s16(d4s16, d31s16);
+  q2s32 = vmull_s16(d5s16, d31s16);
+
+  q1s32 = vmlsl_s16(q1s32, d8s16, d31s16);
+  q3s32 = vmlsl_s16(q3s32, d9s16, d31s16);
+  q10s32 = vmlal_s16(q10s32, d8s16, d30s16);
+  q2s32 = vmlal_s16(q2s32, d9s16, d30s16);
+
+  *q8s16 = vaddq_s16(*q11s16, *q9s16);
+  *q11s16 = vsubq_s16(*q11s16, *q9s16);
+  q4s16 = vaddq_s16(*q12s16, *q14s16);
+  *q12s16 = vsubq_s16(*q12s16, *q14s16);
+
+  q14s32 = vaddq_s32(q5s32, q1s32);
+  q15s32 = vaddq_s32(q6s32, q3s32);
+  q5s32 = vsubq_s32(q5s32, q1s32);
+  q6s32 = vsubq_s32(q6s32, q3s32);
+
+  d18s16 = vqrshrn_n_s32(q14s32, 14);
+  d19s16 = vqrshrn_n_s32(q15s32, 14);
+  d10s16 = vqrshrn_n_s32(q5s32, 14);
+  d11s16 = vqrshrn_n_s32(q6s32, 14);
+  *q9s16 = vcombine_s16(d18s16, d19s16);
+
+  q1s32 = vaddq_s32(q7s32, q10s32);
+  q3s32 = vaddq_s32(q0s32, q2s32);
+  q7s32 = vsubq_s32(q7s32, q10s32);
+  q0s32 = vsubq_s32(q0s32, q2s32);
+
+  d28s16 = vqrshrn_n_s32(q1s32, 14);
+  d29s16 = vqrshrn_n_s32(q3s32, 14);
+  d14s16 = vqrshrn_n_s32(q7s32, 14);
+  d15s16 = vqrshrn_n_s32(q0s32, 14);
+  *q14s16 = vcombine_s16(d28s16, d29s16);
+
+  d30s16 = vdup_n_s16((int16_t)cospi_16_64);
+
+  d22s16 = vget_low_s16(*q11s16);
+  d23s16 = vget_high_s16(*q11s16);
+  q2s32 = vmull_s16(d22s16, d30s16);
+  q3s32 = vmull_s16(d23s16, d30s16);
+  q13s32 = vmull_s16(d22s16, d30s16);
+  q1s32 = vmull_s16(d23s16, d30s16);
+
+  d24s16 = vget_low_s16(*q12s16);
+  d25s16 = vget_high_s16(*q12s16);
+  q2s32 = vmlal_s16(q2s32, d24s16, d30s16);
+  q3s32 = vmlal_s16(q3s32, d25s16, d30s16);
+  q13s32 = vmlsl_s16(q13s32, d24s16, d30s16);
+  q1s32 = vmlsl_s16(q1s32, d25s16, d30s16);
+
+  d4s16 = vqrshrn_n_s32(q2s32, 14);
+  d5s16 = vqrshrn_n_s32(q3s32, 14);
+  d24s16 = vqrshrn_n_s32(q13s32, 14);
+  d25s16 = vqrshrn_n_s32(q1s32, 14);
+  q2s16 = vcombine_s16(d4s16, d5s16);
+  *q12s16 = vcombine_s16(d24s16, d25s16);
+
+  q13s32 = vmull_s16(d10s16, d30s16);
+  q1s32 = vmull_s16(d11s16, d30s16);
+  q11s32 = vmull_s16(d10s16, d30s16);
+  q0s32 = vmull_s16(d11s16, d30s16);
+
+  q13s32 = vmlal_s16(q13s32, d14s16, d30s16);
+  q1s32 = vmlal_s16(q1s32, d15s16, d30s16);
+  q11s32 = vmlsl_s16(q11s32, d14s16, d30s16);
+  q0s32 = vmlsl_s16(q0s32, d15s16, d30s16);
+
+  d20s16 = vqrshrn_n_s32(q13s32, 14);
+  d21s16 = vqrshrn_n_s32(q1s32, 14);
+  d12s16 = vqrshrn_n_s32(q11s32, 14);
+  d13s16 = vqrshrn_n_s32(q0s32, 14);
+  *q10s16 = vcombine_s16(d20s16, d21s16);
+  q6s16 = vcombine_s16(d12s16, d13s16);
+
+  q5s16 = vdupq_n_s16(0);
+
+  *q9s16 = vsubq_s16(q5s16, *q9s16);
+  *q11s16 = vsubq_s16(q5s16, q2s16);
+  *q13s16 = vsubq_s16(q5s16, q6s16);
+  *q15s16 = vsubq_s16(q5s16, q4s16);
+  return;
+}
+
+void av1_iht8x8_64_add_neon(const tran_low_t *input, uint8_t *dest,
+                            int dest_stride, int tx_type) {
+  int i;
+  uint8_t *d1, *d2;
+  uint8x8_t d0u8, d1u8, d2u8, d3u8;
+  uint64x1_t d0u64, d1u64, d2u64, d3u64;
+  int16x8_t q8s16, q9s16, q10s16, q11s16, q12s16, q13s16, q14s16, q15s16;
+  uint16x8_t q8u16, q9u16, q10u16, q11u16;
+
+  q8s16 = vld1q_s16(input);
+  q9s16 = vld1q_s16(input + 8);
+  q10s16 = vld1q_s16(input + 8 * 2);
+  q11s16 = vld1q_s16(input + 8 * 3);
+  q12s16 = vld1q_s16(input + 8 * 4);
+  q13s16 = vld1q_s16(input + 8 * 5);
+  q14s16 = vld1q_s16(input + 8 * 6);
+  q15s16 = vld1q_s16(input + 8 * 7);
+
+  TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+               &q15s16);
+
+  switch (tx_type) {
+    case 0:  // idct_idct is not supported. Fall back to C
+      av1_iht8x8_64_add_c(input, dest, dest_stride, tx_type);
+      return;
+      break;
+    case 1:  // iadst_idct
+      // generate IDCT constants
+      // GENERATE_IDCT_CONSTANTS
+
+      // first transform rows
+      IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+                 &q15s16);
+
+      // transpose the matrix
+      TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+                   &q15s16);
+
+      // generate IADST constants
+      // GENERATE_IADST_CONSTANTS
+
+      // then transform columns
+      IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+                  &q15s16);
+      break;
+    case 2:  // idct_iadst
+      // generate IADST constants
+      // GENERATE_IADST_CONSTANTS
+
+      // first transform rows
+      IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+                  &q15s16);
+
+      // transpose the matrix
+      TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+                   &q15s16);
+
+      // generate IDCT constants
+      // GENERATE_IDCT_CONSTANTS
+
+      // then transform columns
+      IDCT8x8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+                 &q15s16);
+      break;
+    case 3:  // iadst_iadst
+      // generate IADST constants
+      // GENERATE_IADST_CONSTANTS
+
+      // first transform rows
+      IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+                  &q15s16);
+
+      // transpose the matrix
+      TRANSPOSE8X8(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+                   &q15s16);
+
+      // then transform columns
+      IADST8X8_1D(&q8s16, &q9s16, &q10s16, &q11s16, &q12s16, &q13s16, &q14s16,
+                  &q15s16);
+      break;
+    default:  // iadst_idct
+      assert(0);
+      break;
+  }
+
+  q8s16 = vrshrq_n_s16(q8s16, 5);
+  q9s16 = vrshrq_n_s16(q9s16, 5);
+  q10s16 = vrshrq_n_s16(q10s16, 5);
+  q11s16 = vrshrq_n_s16(q11s16, 5);
+  q12s16 = vrshrq_n_s16(q12s16, 5);
+  q13s16 = vrshrq_n_s16(q13s16, 5);
+  q14s16 = vrshrq_n_s16(q14s16, 5);
+  q15s16 = vrshrq_n_s16(q15s16, 5);
+
+  for (d1 = d2 = dest, i = 0; i < 2; i++) {
+    if (i != 0) {
+      q8s16 = q12s16;
+      q9s16 = q13s16;
+      q10s16 = q14s16;
+      q11s16 = q15s16;
+    }
+
+    d0u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d1u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d2u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+    d3u64 = vld1_u64((uint64_t *)d1);
+    d1 += dest_stride;
+
+    q8u16 = vaddw_u8(vreinterpretq_u16_s16(q8s16), vreinterpret_u8_u64(d0u64));
+    q9u16 = vaddw_u8(vreinterpretq_u16_s16(q9s16), vreinterpret_u8_u64(d1u64));
+    q10u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q10s16), vreinterpret_u8_u64(d2u64));
+    q11u16 =
+        vaddw_u8(vreinterpretq_u16_s16(q11s16), vreinterpret_u8_u64(d3u64));
+
+    d0u8 = vqmovun_s16(vreinterpretq_s16_u16(q8u16));
+    d1u8 = vqmovun_s16(vreinterpretq_s16_u16(q9u16));
+    d2u8 = vqmovun_s16(vreinterpretq_s16_u16(q10u16));
+    d3u8 = vqmovun_s16(vreinterpretq_s16_u16(q11u16));
+
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d0u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d1u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d2u8));
+    d2 += dest_stride;
+    vst1_u64((uint64_t *)d2, vreinterpret_u64_u8(d3u8));
+    d2 += dest_stride;
+  }
+  return;
+}
diff --git a/third_party/aom/av1/common/av1_fwd_txfm1d.c b/third_party/aom/av1/common/av1_fwd_txfm1d.c
new file mode 100644
index 0000000000..7a691e03f5
--- /dev/null
+++ b/third_party/aom/av1/common/av1_fwd_txfm1d.c
@@ -0,0 +1,2312 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include "av1/common/av1_fwd_txfm1d.h"
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+
+void range_check_func(int32_t stage, const int32_t *input, const int32_t *buf,
+                      int32_t size, int8_t bit);
+
+#define range_check(stage, input, buf, size, bit) \
+  range_check_func(stage, input, buf, size, bit)
+#else
+#define range_check(stage, input, buf, size, bit) \
+  {                                               \
+    (void)stage;                                  \
+    (void)input;                                  \
+    (void)buf;                                    \
+    (void)size;                                   \
+    (void)bit;                                    \
+  }
+#endif
+
+// TODO(angiebird): Make 1-d txfm functions static
+void av1_fdct4_new(const int32_t *input, int32_t *output, const int8_t *cos_bit,
+                   const int8_t *stage_range) {
+  const int32_t size = 4;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[4];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0] + input[3];
+  bf1[1] = input[1] + input[2];
+  bf1[2] = -input[2] + input[1];
+  bf1[3] = -input[3] + input[0];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[32], bf0[1], cospi[32], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[48], bf0[3], -cospi[16], bf0[2], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[2];
+  bf1[2] = bf0[1];
+  bf1[3] = bf0[3];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_fdct8_new(const int32_t *input, int32_t *output, const int8_t *cos_bit,
+                   const int8_t *stage_range) {
+  const int32_t size = 8;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[8];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0] + input[7];
+  bf1[1] = input[1] + input[6];
+  bf1[2] = input[2] + input[5];
+  bf1[3] = input[3] + input[4];
+  bf1[4] = -input[4] + input[3];
+  bf1[5] = -input[5] + input[2];
+  bf1[6] = -input[6] + input[1];
+  bf1[7] = -input[7] + input[0];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[3];
+  bf1[1] = bf0[1] + bf0[2];
+  bf1[2] = -bf0[2] + bf0[1];
+  bf1[3] = -bf0[3] + bf0[0];
+  bf1[4] = bf0[4];
+  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[5], cos_bit[stage]);
+  bf1[7] = bf0[7];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[32], bf0[1], cospi[32], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[48], bf0[3], -cospi[16], bf0[2], cos_bit[stage]);
+  bf1[4] = bf0[4] + bf0[5];
+  bf1[5] = -bf0[5] + bf0[4];
+  bf1[6] = -bf0[6] + bf0[7];
+  bf1[7] = bf0[7] + bf0[6];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[56], bf0[4], cospi[8], bf0[7], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[24], bf0[5], cospi[40], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[24], bf0[6], -cospi[40], bf0[5], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[56], bf0[7], -cospi[8], bf0[4], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[4];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[6];
+  bf1[4] = bf0[1];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[3];
+  bf1[7] = bf0[7];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_fdct16_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 16;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[16];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0] + input[15];
+  bf1[1] = input[1] + input[14];
+  bf1[2] = input[2] + input[13];
+  bf1[3] = input[3] + input[12];
+  bf1[4] = input[4] + input[11];
+  bf1[5] = input[5] + input[10];
+  bf1[6] = input[6] + input[9];
+  bf1[7] = input[7] + input[8];
+  bf1[8] = -input[8] + input[7];
+  bf1[9] = -input[9] + input[6];
+  bf1[10] = -input[10] + input[5];
+  bf1[11] = -input[11] + input[4];
+  bf1[12] = -input[12] + input[3];
+  bf1[13] = -input[13] + input[2];
+  bf1[14] = -input[14] + input[1];
+  bf1[15] = -input[15] + input[0];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[7];
+  bf1[1] = bf0[1] + bf0[6];
+  bf1[2] = bf0[2] + bf0[5];
+  bf1[3] = bf0[3] + bf0[4];
+  bf1[4] = -bf0[4] + bf0[3];
+  bf1[5] = -bf0[5] + bf0[2];
+  bf1[6] = -bf0[6] + bf0[1];
+  bf1[7] = -bf0[7] + bf0[0];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(-cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[32], bf0[12], cospi[32], bf0[11], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[32], bf0[13], cospi[32], bf0[10], cos_bit[stage]);
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[3];
+  bf1[1] = bf0[1] + bf0[2];
+  bf1[2] = -bf0[2] + bf0[1];
+  bf1[3] = -bf0[3] + bf0[0];
+  bf1[4] = bf0[4];
+  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[5], cos_bit[stage]);
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8] + bf0[11];
+  bf1[9] = bf0[9] + bf0[10];
+  bf1[10] = -bf0[10] + bf0[9];
+  bf1[11] = -bf0[11] + bf0[8];
+  bf1[12] = -bf0[12] + bf0[15];
+  bf1[13] = -bf0[13] + bf0[14];
+  bf1[14] = bf0[14] + bf0[13];
+  bf1[15] = bf0[15] + bf0[12];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[32], bf0[1], cospi[32], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[48], bf0[3], -cospi[16], bf0[2], cos_bit[stage]);
+  bf1[4] = bf0[4] + bf0[5];
+  bf1[5] = -bf0[5] + bf0[4];
+  bf1[6] = -bf0[6] + bf0[7];
+  bf1[7] = bf0[7] + bf0[6];
+  bf1[8] = bf0[8];
+  bf1[9] = half_btf(-cospi[16], bf0[9], cospi[48], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(-cospi[48], bf0[10], -cospi[16], bf0[13], cos_bit[stage]);
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = half_btf(cospi[48], bf0[13], -cospi[16], bf0[10], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[16], bf0[14], cospi[48], bf0[9], cos_bit[stage]);
+  bf1[15] = bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[56], bf0[4], cospi[8], bf0[7], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[24], bf0[5], cospi[40], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[24], bf0[6], -cospi[40], bf0[5], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[56], bf0[7], -cospi[8], bf0[4], cos_bit[stage]);
+  bf1[8] = bf0[8] + bf0[9];
+  bf1[9] = -bf0[9] + bf0[8];
+  bf1[10] = -bf0[10] + bf0[11];
+  bf1[11] = bf0[11] + bf0[10];
+  bf1[12] = bf0[12] + bf0[13];
+  bf1[13] = -bf0[13] + bf0[12];
+  bf1[14] = -bf0[14] + bf0[15];
+  bf1[15] = bf0[15] + bf0[14];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[60], bf0[8], cospi[4], bf0[15], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[28], bf0[9], cospi[36], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[44], bf0[10], cospi[20], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[12], bf0[11], cospi[52], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[12], bf0[12], -cospi[52], bf0[11], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[44], bf0[13], -cospi[20], bf0[10], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[28], bf0[14], -cospi[36], bf0[9], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[60], bf0[15], -cospi[4], bf0[8], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[8];
+  bf1[2] = bf0[4];
+  bf1[3] = bf0[12];
+  bf1[4] = bf0[2];
+  bf1[5] = bf0[10];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[14];
+  bf1[8] = bf0[1];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[5];
+  bf1[11] = bf0[13];
+  bf1[12] = bf0[3];
+  bf1[13] = bf0[11];
+  bf1[14] = bf0[7];
+  bf1[15] = bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_fdct32_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 32;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[32];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0] + input[31];
+  bf1[1] = input[1] + input[30];
+  bf1[2] = input[2] + input[29];
+  bf1[3] = input[3] + input[28];
+  bf1[4] = input[4] + input[27];
+  bf1[5] = input[5] + input[26];
+  bf1[6] = input[6] + input[25];
+  bf1[7] = input[7] + input[24];
+  bf1[8] = input[8] + input[23];
+  bf1[9] = input[9] + input[22];
+  bf1[10] = input[10] + input[21];
+  bf1[11] = input[11] + input[20];
+  bf1[12] = input[12] + input[19];
+  bf1[13] = input[13] + input[18];
+  bf1[14] = input[14] + input[17];
+  bf1[15] = input[15] + input[16];
+  bf1[16] = -input[16] + input[15];
+  bf1[17] = -input[17] + input[14];
+  bf1[18] = -input[18] + input[13];
+  bf1[19] = -input[19] + input[12];
+  bf1[20] = -input[20] + input[11];
+  bf1[21] = -input[21] + input[10];
+  bf1[22] = -input[22] + input[9];
+  bf1[23] = -input[23] + input[8];
+  bf1[24] = -input[24] + input[7];
+  bf1[25] = -input[25] + input[6];
+  bf1[26] = -input[26] + input[5];
+  bf1[27] = -input[27] + input[4];
+  bf1[28] = -input[28] + input[3];
+  bf1[29] = -input[29] + input[2];
+  bf1[30] = -input[30] + input[1];
+  bf1[31] = -input[31] + input[0];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[15];
+  bf1[1] = bf0[1] + bf0[14];
+  bf1[2] = bf0[2] + bf0[13];
+  bf1[3] = bf0[3] + bf0[12];
+  bf1[4] = bf0[4] + bf0[11];
+  bf1[5] = bf0[5] + bf0[10];
+  bf1[6] = bf0[6] + bf0[9];
+  bf1[7] = bf0[7] + bf0[8];
+  bf1[8] = -bf0[8] + bf0[7];
+  bf1[9] = -bf0[9] + bf0[6];
+  bf1[10] = -bf0[10] + bf0[5];
+  bf1[11] = -bf0[11] + bf0[4];
+  bf1[12] = -bf0[12] + bf0[3];
+  bf1[13] = -bf0[13] + bf0[2];
+  bf1[14] = -bf0[14] + bf0[1];
+  bf1[15] = -bf0[15] + bf0[0];
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = half_btf(-cospi[32], bf0[20], cospi[32], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[32], bf0[21], cospi[32], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[32], bf0[22], cospi[32], bf0[25], cos_bit[stage]);
+  bf1[23] = half_btf(-cospi[32], bf0[23], cospi[32], bf0[24], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[32], bf0[24], cospi[32], bf0[23], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[32], bf0[25], cospi[32], bf0[22], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[32], bf0[26], cospi[32], bf0[21], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[32], bf0[27], cospi[32], bf0[20], cos_bit[stage]);
+  bf1[28] = bf0[28];
+  bf1[29] = bf0[29];
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[7];
+  bf1[1] = bf0[1] + bf0[6];
+  bf1[2] = bf0[2] + bf0[5];
+  bf1[3] = bf0[3] + bf0[4];
+  bf1[4] = -bf0[4] + bf0[3];
+  bf1[5] = -bf0[5] + bf0[2];
+  bf1[6] = -bf0[6] + bf0[1];
+  bf1[7] = -bf0[7] + bf0[0];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(-cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[32], bf0[12], cospi[32], bf0[11], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[32], bf0[13], cospi[32], bf0[10], cos_bit[stage]);
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = bf0[16] + bf0[23];
+  bf1[17] = bf0[17] + bf0[22];
+  bf1[18] = bf0[18] + bf0[21];
+  bf1[19] = bf0[19] + bf0[20];
+  bf1[20] = -bf0[20] + bf0[19];
+  bf1[21] = -bf0[21] + bf0[18];
+  bf1[22] = -bf0[22] + bf0[17];
+  bf1[23] = -bf0[23] + bf0[16];
+  bf1[24] = -bf0[24] + bf0[31];
+  bf1[25] = -bf0[25] + bf0[30];
+  bf1[26] = -bf0[26] + bf0[29];
+  bf1[27] = -bf0[27] + bf0[28];
+  bf1[28] = bf0[28] + bf0[27];
+  bf1[29] = bf0[29] + bf0[26];
+  bf1[30] = bf0[30] + bf0[25];
+  bf1[31] = bf0[31] + bf0[24];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[3];
+  bf1[1] = bf0[1] + bf0[2];
+  bf1[2] = -bf0[2] + bf0[1];
+  bf1[3] = -bf0[3] + bf0[0];
+  bf1[4] = bf0[4];
+  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[5], cos_bit[stage]);
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8] + bf0[11];
+  bf1[9] = bf0[9] + bf0[10];
+  bf1[10] = -bf0[10] + bf0[9];
+  bf1[11] = -bf0[11] + bf0[8];
+  bf1[12] = -bf0[12] + bf0[15];
+  bf1[13] = -bf0[13] + bf0[14];
+  bf1[14] = bf0[14] + bf0[13];
+  bf1[15] = bf0[15] + bf0[12];
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = half_btf(-cospi[16], bf0[18], cospi[48], bf0[29], cos_bit[stage]);
+  bf1[19] = half_btf(-cospi[16], bf0[19], cospi[48], bf0[28], cos_bit[stage]);
+  bf1[20] = half_btf(-cospi[48], bf0[20], -cospi[16], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[48], bf0[21], -cospi[16], bf0[26], cos_bit[stage]);
+  bf1[22] = bf0[22];
+  bf1[23] = bf0[23];
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = half_btf(cospi[48], bf0[26], -cospi[16], bf0[21], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[48], bf0[27], -cospi[16], bf0[20], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[16], bf0[28], cospi[48], bf0[19], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[16], bf0[29], cospi[48], bf0[18], cos_bit[stage]);
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[32], bf0[1], cospi[32], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[48], bf0[3], -cospi[16], bf0[2], cos_bit[stage]);
+  bf1[4] = bf0[4] + bf0[5];
+  bf1[5] = -bf0[5] + bf0[4];
+  bf1[6] = -bf0[6] + bf0[7];
+  bf1[7] = bf0[7] + bf0[6];
+  bf1[8] = bf0[8];
+  bf1[9] = half_btf(-cospi[16], bf0[9], cospi[48], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(-cospi[48], bf0[10], -cospi[16], bf0[13], cos_bit[stage]);
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = half_btf(cospi[48], bf0[13], -cospi[16], bf0[10], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[16], bf0[14], cospi[48], bf0[9], cos_bit[stage]);
+  bf1[15] = bf0[15];
+  bf1[16] = bf0[16] + bf0[19];
+  bf1[17] = bf0[17] + bf0[18];
+  bf1[18] = -bf0[18] + bf0[17];
+  bf1[19] = -bf0[19] + bf0[16];
+  bf1[20] = -bf0[20] + bf0[23];
+  bf1[21] = -bf0[21] + bf0[22];
+  bf1[22] = bf0[22] + bf0[21];
+  bf1[23] = bf0[23] + bf0[20];
+  bf1[24] = bf0[24] + bf0[27];
+  bf1[25] = bf0[25] + bf0[26];
+  bf1[26] = -bf0[26] + bf0[25];
+  bf1[27] = -bf0[27] + bf0[24];
+  bf1[28] = -bf0[28] + bf0[31];
+  bf1[29] = -bf0[29] + bf0[30];
+  bf1[30] = bf0[30] + bf0[29];
+  bf1[31] = bf0[31] + bf0[28];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[56], bf0[4], cospi[8], bf0[7], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[24], bf0[5], cospi[40], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[24], bf0[6], -cospi[40], bf0[5], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[56], bf0[7], -cospi[8], bf0[4], cos_bit[stage]);
+  bf1[8] = bf0[8] + bf0[9];
+  bf1[9] = -bf0[9] + bf0[8];
+  bf1[10] = -bf0[10] + bf0[11];
+  bf1[11] = bf0[11] + bf0[10];
+  bf1[12] = bf0[12] + bf0[13];
+  bf1[13] = -bf0[13] + bf0[12];
+  bf1[14] = -bf0[14] + bf0[15];
+  bf1[15] = bf0[15] + bf0[14];
+  bf1[16] = bf0[16];
+  bf1[17] = half_btf(-cospi[8], bf0[17], cospi[56], bf0[30], cos_bit[stage]);
+  bf1[18] = half_btf(-cospi[56], bf0[18], -cospi[8], bf0[29], cos_bit[stage]);
+  bf1[19] = bf0[19];
+  bf1[20] = bf0[20];
+  bf1[21] = half_btf(-cospi[40], bf0[21], cospi[24], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[24], bf0[22], -cospi[40], bf0[25], cos_bit[stage]);
+  bf1[23] = bf0[23];
+  bf1[24] = bf0[24];
+  bf1[25] = half_btf(cospi[24], bf0[25], -cospi[40], bf0[22], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[40], bf0[26], cospi[24], bf0[21], cos_bit[stage]);
+  bf1[27] = bf0[27];
+  bf1[28] = bf0[28];
+  bf1[29] = half_btf(cospi[56], bf0[29], -cospi[8], bf0[18], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[8], bf0[30], cospi[56], bf0[17], cos_bit[stage]);
+  bf1[31] = bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[60], bf0[8], cospi[4], bf0[15], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[28], bf0[9], cospi[36], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[44], bf0[10], cospi[20], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[12], bf0[11], cospi[52], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[12], bf0[12], -cospi[52], bf0[11], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[44], bf0[13], -cospi[20], bf0[10], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[28], bf0[14], -cospi[36], bf0[9], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[60], bf0[15], -cospi[4], bf0[8], cos_bit[stage]);
+  bf1[16] = bf0[16] + bf0[17];
+  bf1[17] = -bf0[17] + bf0[16];
+  bf1[18] = -bf0[18] + bf0[19];
+  bf1[19] = bf0[19] + bf0[18];
+  bf1[20] = bf0[20] + bf0[21];
+  bf1[21] = -bf0[21] + bf0[20];
+  bf1[22] = -bf0[22] + bf0[23];
+  bf1[23] = bf0[23] + bf0[22];
+  bf1[24] = bf0[24] + bf0[25];
+  bf1[25] = -bf0[25] + bf0[24];
+  bf1[26] = -bf0[26] + bf0[27];
+  bf1[27] = bf0[27] + bf0[26];
+  bf1[28] = bf0[28] + bf0[29];
+  bf1[29] = -bf0[29] + bf0[28];
+  bf1[30] = -bf0[30] + bf0[31];
+  bf1[31] = bf0[31] + bf0[30];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 8
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = half_btf(cospi[62], bf0[16], cospi[2], bf0[31], cos_bit[stage]);
+  bf1[17] = half_btf(cospi[30], bf0[17], cospi[34], bf0[30], cos_bit[stage]);
+  bf1[18] = half_btf(cospi[46], bf0[18], cospi[18], bf0[29], cos_bit[stage]);
+  bf1[19] = half_btf(cospi[14], bf0[19], cospi[50], bf0[28], cos_bit[stage]);
+  bf1[20] = half_btf(cospi[54], bf0[20], cospi[10], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(cospi[22], bf0[21], cospi[42], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(cospi[38], bf0[22], cospi[26], bf0[25], cos_bit[stage]);
+  bf1[23] = half_btf(cospi[6], bf0[23], cospi[58], bf0[24], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[6], bf0[24], -cospi[58], bf0[23], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[38], bf0[25], -cospi[26], bf0[22], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[22], bf0[26], -cospi[42], bf0[21], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[54], bf0[27], -cospi[10], bf0[20], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[14], bf0[28], -cospi[50], bf0[19], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[46], bf0[29], -cospi[18], bf0[18], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[30], bf0[30], -cospi[34], bf0[17], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[62], bf0[31], -cospi[2], bf0[16], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 9
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[16];
+  bf1[2] = bf0[8];
+  bf1[3] = bf0[24];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[20];
+  bf1[6] = bf0[12];
+  bf1[7] = bf0[28];
+  bf1[8] = bf0[2];
+  bf1[9] = bf0[18];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[26];
+  bf1[12] = bf0[6];
+  bf1[13] = bf0[22];
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[30];
+  bf1[16] = bf0[1];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[9];
+  bf1[19] = bf0[25];
+  bf1[20] = bf0[5];
+  bf1[21] = bf0[21];
+  bf1[22] = bf0[13];
+  bf1[23] = bf0[29];
+  bf1[24] = bf0[3];
+  bf1[25] = bf0[19];
+  bf1[26] = bf0[11];
+  bf1[27] = bf0[27];
+  bf1[28] = bf0[7];
+  bf1[29] = bf0[23];
+  bf1[30] = bf0[15];
+  bf1[31] = bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_fadst4_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 4;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[4];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[3];
+  bf1[1] = input[0];
+  bf1[2] = input[1];
+  bf1[3] = input[2];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[8], bf0[0], cospi[56], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[8], bf0[1], cospi[56], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[40], bf0[2], cospi[24], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[40], bf0[3], cospi[24], bf0[2], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[2];
+  bf1[1] = bf0[1] + bf0[3];
+  bf1[2] = -bf0[2] + bf0[0];
+  bf1[3] = -bf0[3] + bf0[1];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = half_btf(cospi[32], bf0[2], cospi[32], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[32], bf0[3], cospi[32], bf0[2], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = -bf0[2];
+  bf1[2] = bf0[3];
+  bf1[3] = -bf0[1];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_fadst8_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 8;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[8];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[7];
+  bf1[1] = input[0];
+  bf1[2] = input[5];
+  bf1[3] = input[2];
+  bf1[4] = input[3];
+  bf1[5] = input[4];
+  bf1[6] = input[1];
+  bf1[7] = input[6];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[4], bf0[0], cospi[60], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[4], bf0[1], cospi[60], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[20], bf0[2], cospi[44], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[20], bf0[3], cospi[44], bf0[2], cos_bit[stage]);
+  bf1[4] = half_btf(cospi[36], bf0[4], cospi[28], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(-cospi[36], bf0[5], cospi[28], bf0[4], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[52], bf0[6], cospi[12], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(-cospi[52], bf0[7], cospi[12], bf0[6], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[4];
+  bf1[1] = bf0[1] + bf0[5];
+  bf1[2] = bf0[2] + bf0[6];
+  bf1[3] = bf0[3] + bf0[7];
+  bf1[4] = -bf0[4] + bf0[0];
+  bf1[5] = -bf0[5] + bf0[1];
+  bf1[6] = -bf0[6] + bf0[2];
+  bf1[7] = -bf0[7] + bf0[3];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[16], bf0[4], cospi[48], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(-cospi[16], bf0[5], cospi[48], bf0[4], cos_bit[stage]);
+  bf1[6] = half_btf(-cospi[48], bf0[6], cospi[16], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[48], bf0[7], cospi[16], bf0[6], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[2];
+  bf1[1] = bf0[1] + bf0[3];
+  bf1[2] = -bf0[2] + bf0[0];
+  bf1[3] = -bf0[3] + bf0[1];
+  bf1[4] = bf0[4] + bf0[6];
+  bf1[5] = bf0[5] + bf0[7];
+  bf1[6] = -bf0[6] + bf0[4];
+  bf1[7] = -bf0[7] + bf0[5];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = half_btf(cospi[32], bf0[2], cospi[32], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[32], bf0[3], cospi[32], bf0[2], cos_bit[stage]);
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(-cospi[32], bf0[7], cospi[32], bf0[6], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = -bf0[4];
+  bf1[2] = bf0[6];
+  bf1[3] = -bf0[2];
+  bf1[4] = bf0[3];
+  bf1[5] = -bf0[7];
+  bf1[6] = bf0[5];
+  bf1[7] = -bf0[1];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_fadst16_new(const int32_t *input, int32_t *output,
+                     const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 16;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[16];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[15];
+  bf1[1] = input[0];
+  bf1[2] = input[13];
+  bf1[3] = input[2];
+  bf1[4] = input[11];
+  bf1[5] = input[4];
+  bf1[6] = input[9];
+  bf1[7] = input[6];
+  bf1[8] = input[7];
+  bf1[9] = input[8];
+  bf1[10] = input[5];
+  bf1[11] = input[10];
+  bf1[12] = input[3];
+  bf1[13] = input[12];
+  bf1[14] = input[1];
+  bf1[15] = input[14];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[2], bf0[0], cospi[62], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[2], bf0[1], cospi[62], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[10], bf0[2], cospi[54], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[10], bf0[3], cospi[54], bf0[2], cos_bit[stage]);
+  bf1[4] = half_btf(cospi[18], bf0[4], cospi[46], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(-cospi[18], bf0[5], cospi[46], bf0[4], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[26], bf0[6], cospi[38], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(-cospi[26], bf0[7], cospi[38], bf0[6], cos_bit[stage]);
+  bf1[8] = half_btf(cospi[34], bf0[8], cospi[30], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(-cospi[34], bf0[9], cospi[30], bf0[8], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[42], bf0[10], cospi[22], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[42], bf0[11], cospi[22], bf0[10], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[50], bf0[12], cospi[14], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(-cospi[50], bf0[13], cospi[14], bf0[12], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[58], bf0[14], cospi[6], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(-cospi[58], bf0[15], cospi[6], bf0[14], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[8];
+  bf1[1] = bf0[1] + bf0[9];
+  bf1[2] = bf0[2] + bf0[10];
+  bf1[3] = bf0[3] + bf0[11];
+  bf1[4] = bf0[4] + bf0[12];
+  bf1[5] = bf0[5] + bf0[13];
+  bf1[6] = bf0[6] + bf0[14];
+  bf1[7] = bf0[7] + bf0[15];
+  bf1[8] = -bf0[8] + bf0[0];
+  bf1[9] = -bf0[9] + bf0[1];
+  bf1[10] = -bf0[10] + bf0[2];
+  bf1[11] = -bf0[11] + bf0[3];
+  bf1[12] = -bf0[12] + bf0[4];
+  bf1[13] = -bf0[13] + bf0[5];
+  bf1[14] = -bf0[14] + bf0[6];
+  bf1[15] = -bf0[15] + bf0[7];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[8], bf0[8], cospi[56], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(-cospi[8], bf0[9], cospi[56], bf0[8], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[40], bf0[10], cospi[24], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[40], bf0[11], cospi[24], bf0[10], cos_bit[stage]);
+  bf1[12] = half_btf(-cospi[56], bf0[12], cospi[8], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[56], bf0[13], cospi[8], bf0[12], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[24], bf0[14], cospi[40], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[24], bf0[15], cospi[40], bf0[14], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[4];
+  bf1[1] = bf0[1] + bf0[5];
+  bf1[2] = bf0[2] + bf0[6];
+  bf1[3] = bf0[3] + bf0[7];
+  bf1[4] = -bf0[4] + bf0[0];
+  bf1[5] = -bf0[5] + bf0[1];
+  bf1[6] = -bf0[6] + bf0[2];
+  bf1[7] = -bf0[7] + bf0[3];
+  bf1[8] = bf0[8] + bf0[12];
+  bf1[9] = bf0[9] + bf0[13];
+  bf1[10] = bf0[10] + bf0[14];
+  bf1[11] = bf0[11] + bf0[15];
+  bf1[12] = -bf0[12] + bf0[8];
+  bf1[13] = -bf0[13] + bf0[9];
+  bf1[14] = -bf0[14] + bf0[10];
+  bf1[15] = -bf0[15] + bf0[11];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[16], bf0[4], cospi[48], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(-cospi[16], bf0[5], cospi[48], bf0[4], cos_bit[stage]);
+  bf1[6] = half_btf(-cospi[48], bf0[6], cospi[16], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[48], bf0[7], cospi[16], bf0[6], cos_bit[stage]);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = half_btf(cospi[16], bf0[12], cospi[48], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(-cospi[16], bf0[13], cospi[48], bf0[12], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[48], bf0[14], cospi[16], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[48], bf0[15], cospi[16], bf0[14], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[2];
+  bf1[1] = bf0[1] + bf0[3];
+  bf1[2] = -bf0[2] + bf0[0];
+  bf1[3] = -bf0[3] + bf0[1];
+  bf1[4] = bf0[4] + bf0[6];
+  bf1[5] = bf0[5] + bf0[7];
+  bf1[6] = -bf0[6] + bf0[4];
+  bf1[7] = -bf0[7] + bf0[5];
+  bf1[8] = bf0[8] + bf0[10];
+  bf1[9] = bf0[9] + bf0[11];
+  bf1[10] = -bf0[10] + bf0[8];
+  bf1[11] = -bf0[11] + bf0[9];
+  bf1[12] = bf0[12] + bf0[14];
+  bf1[13] = bf0[13] + bf0[15];
+  bf1[14] = -bf0[14] + bf0[12];
+  bf1[15] = -bf0[15] + bf0[13];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 8
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = half_btf(cospi[32], bf0[2], cospi[32], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[32], bf0[3], cospi[32], bf0[2], cos_bit[stage]);
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(-cospi[32], bf0[7], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(cospi[32], bf0[10], cospi[32], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[10], cos_bit[stage]);
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = half_btf(cospi[32], bf0[14], cospi[32], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(-cospi[32], bf0[15], cospi[32], bf0[14], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 9
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = -bf0[8];
+  bf1[2] = bf0[12];
+  bf1[3] = -bf0[4];
+  bf1[4] = bf0[6];
+  bf1[5] = -bf0[14];
+  bf1[6] = bf0[10];
+  bf1[7] = -bf0[2];
+  bf1[8] = bf0[3];
+  bf1[9] = -bf0[11];
+  bf1[10] = bf0[15];
+  bf1[11] = -bf0[7];
+  bf1[12] = bf0[5];
+  bf1[13] = -bf0[13];
+  bf1[14] = bf0[9];
+  bf1[15] = -bf0[1];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_fadst32_new(const int32_t *input, int32_t *output,
+                     const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 32;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[32];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[31];
+  bf1[1] = input[0];
+  bf1[2] = input[29];
+  bf1[3] = input[2];
+  bf1[4] = input[27];
+  bf1[5] = input[4];
+  bf1[6] = input[25];
+  bf1[7] = input[6];
+  bf1[8] = input[23];
+  bf1[9] = input[8];
+  bf1[10] = input[21];
+  bf1[11] = input[10];
+  bf1[12] = input[19];
+  bf1[13] = input[12];
+  bf1[14] = input[17];
+  bf1[15] = input[14];
+  bf1[16] = input[15];
+  bf1[17] = input[16];
+  bf1[18] = input[13];
+  bf1[19] = input[18];
+  bf1[20] = input[11];
+  bf1[21] = input[20];
+  bf1[22] = input[9];
+  bf1[23] = input[22];
+  bf1[24] = input[7];
+  bf1[25] = input[24];
+  bf1[26] = input[5];
+  bf1[27] = input[26];
+  bf1[28] = input[3];
+  bf1[29] = input[28];
+  bf1[30] = input[1];
+  bf1[31] = input[30];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[1], bf0[0], cospi[63], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[1], bf0[1], cospi[63], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[5], bf0[2], cospi[59], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[5], bf0[3], cospi[59], bf0[2], cos_bit[stage]);
+  bf1[4] = half_btf(cospi[9], bf0[4], cospi[55], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(-cospi[9], bf0[5], cospi[55], bf0[4], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[13], bf0[6], cospi[51], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(-cospi[13], bf0[7], cospi[51], bf0[6], cos_bit[stage]);
+  bf1[8] = half_btf(cospi[17], bf0[8], cospi[47], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(-cospi[17], bf0[9], cospi[47], bf0[8], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[21], bf0[10], cospi[43], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[21], bf0[11], cospi[43], bf0[10], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[25], bf0[12], cospi[39], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(-cospi[25], bf0[13], cospi[39], bf0[12], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[29], bf0[14], cospi[35], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(-cospi[29], bf0[15], cospi[35], bf0[14], cos_bit[stage]);
+  bf1[16] = half_btf(cospi[33], bf0[16], cospi[31], bf0[17], cos_bit[stage]);
+  bf1[17] = half_btf(-cospi[33], bf0[17], cospi[31], bf0[16], cos_bit[stage]);
+  bf1[18] = half_btf(cospi[37], bf0[18], cospi[27], bf0[19], cos_bit[stage]);
+  bf1[19] = half_btf(-cospi[37], bf0[19], cospi[27], bf0[18], cos_bit[stage]);
+  bf1[20] = half_btf(cospi[41], bf0[20], cospi[23], bf0[21], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[41], bf0[21], cospi[23], bf0[20], cos_bit[stage]);
+  bf1[22] = half_btf(cospi[45], bf0[22], cospi[19], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(-cospi[45], bf0[23], cospi[19], bf0[22], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[49], bf0[24], cospi[15], bf0[25], cos_bit[stage]);
+  bf1[25] = half_btf(-cospi[49], bf0[25], cospi[15], bf0[24], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[53], bf0[26], cospi[11], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(-cospi[53], bf0[27], cospi[11], bf0[26], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[57], bf0[28], cospi[7], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(-cospi[57], bf0[29], cospi[7], bf0[28], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[61], bf0[30], cospi[3], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(-cospi[61], bf0[31], cospi[3], bf0[30], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[16];
+  bf1[1] = bf0[1] + bf0[17];
+  bf1[2] = bf0[2] + bf0[18];
+  bf1[3] = bf0[3] + bf0[19];
+  bf1[4] = bf0[4] + bf0[20];
+  bf1[5] = bf0[5] + bf0[21];
+  bf1[6] = bf0[6] + bf0[22];
+  bf1[7] = bf0[7] + bf0[23];
+  bf1[8] = bf0[8] + bf0[24];
+  bf1[9] = bf0[9] + bf0[25];
+  bf1[10] = bf0[10] + bf0[26];
+  bf1[11] = bf0[11] + bf0[27];
+  bf1[12] = bf0[12] + bf0[28];
+  bf1[13] = bf0[13] + bf0[29];
+  bf1[14] = bf0[14] + bf0[30];
+  bf1[15] = bf0[15] + bf0[31];
+  bf1[16] = -bf0[16] + bf0[0];
+  bf1[17] = -bf0[17] + bf0[1];
+  bf1[18] = -bf0[18] + bf0[2];
+  bf1[19] = -bf0[19] + bf0[3];
+  bf1[20] = -bf0[20] + bf0[4];
+  bf1[21] = -bf0[21] + bf0[5];
+  bf1[22] = -bf0[22] + bf0[6];
+  bf1[23] = -bf0[23] + bf0[7];
+  bf1[24] = -bf0[24] + bf0[8];
+  bf1[25] = -bf0[25] + bf0[9];
+  bf1[26] = -bf0[26] + bf0[10];
+  bf1[27] = -bf0[27] + bf0[11];
+  bf1[28] = -bf0[28] + bf0[12];
+  bf1[29] = -bf0[29] + bf0[13];
+  bf1[30] = -bf0[30] + bf0[14];
+  bf1[31] = -bf0[31] + bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = half_btf(cospi[4], bf0[16], cospi[60], bf0[17], cos_bit[stage]);
+  bf1[17] = half_btf(-cospi[4], bf0[17], cospi[60], bf0[16], cos_bit[stage]);
+  bf1[18] = half_btf(cospi[20], bf0[18], cospi[44], bf0[19], cos_bit[stage]);
+  bf1[19] = half_btf(-cospi[20], bf0[19], cospi[44], bf0[18], cos_bit[stage]);
+  bf1[20] = half_btf(cospi[36], bf0[20], cospi[28], bf0[21], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[36], bf0[21], cospi[28], bf0[20], cos_bit[stage]);
+  bf1[22] = half_btf(cospi[52], bf0[22], cospi[12], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(-cospi[52], bf0[23], cospi[12], bf0[22], cos_bit[stage]);
+  bf1[24] = half_btf(-cospi[60], bf0[24], cospi[4], bf0[25], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[60], bf0[25], cospi[4], bf0[24], cos_bit[stage]);
+  bf1[26] = half_btf(-cospi[44], bf0[26], cospi[20], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[44], bf0[27], cospi[20], bf0[26], cos_bit[stage]);
+  bf1[28] = half_btf(-cospi[28], bf0[28], cospi[36], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[28], bf0[29], cospi[36], bf0[28], cos_bit[stage]);
+  bf1[30] = half_btf(-cospi[12], bf0[30], cospi[52], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[12], bf0[31], cospi[52], bf0[30], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[8];
+  bf1[1] = bf0[1] + bf0[9];
+  bf1[2] = bf0[2] + bf0[10];
+  bf1[3] = bf0[3] + bf0[11];
+  bf1[4] = bf0[4] + bf0[12];
+  bf1[5] = bf0[5] + bf0[13];
+  bf1[6] = bf0[6] + bf0[14];
+  bf1[7] = bf0[7] + bf0[15];
+  bf1[8] = -bf0[8] + bf0[0];
+  bf1[9] = -bf0[9] + bf0[1];
+  bf1[10] = -bf0[10] + bf0[2];
+  bf1[11] = -bf0[11] + bf0[3];
+  bf1[12] = -bf0[12] + bf0[4];
+  bf1[13] = -bf0[13] + bf0[5];
+  bf1[14] = -bf0[14] + bf0[6];
+  bf1[15] = -bf0[15] + bf0[7];
+  bf1[16] = bf0[16] + bf0[24];
+  bf1[17] = bf0[17] + bf0[25];
+  bf1[18] = bf0[18] + bf0[26];
+  bf1[19] = bf0[19] + bf0[27];
+  bf1[20] = bf0[20] + bf0[28];
+  bf1[21] = bf0[21] + bf0[29];
+  bf1[22] = bf0[22] + bf0[30];
+  bf1[23] = bf0[23] + bf0[31];
+  bf1[24] = -bf0[24] + bf0[16];
+  bf1[25] = -bf0[25] + bf0[17];
+  bf1[26] = -bf0[26] + bf0[18];
+  bf1[27] = -bf0[27] + bf0[19];
+  bf1[28] = -bf0[28] + bf0[20];
+  bf1[29] = -bf0[29] + bf0[21];
+  bf1[30] = -bf0[30] + bf0[22];
+  bf1[31] = -bf0[31] + bf0[23];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[8], bf0[8], cospi[56], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(-cospi[8], bf0[9], cospi[56], bf0[8], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[40], bf0[10], cospi[24], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[40], bf0[11], cospi[24], bf0[10], cos_bit[stage]);
+  bf1[12] = half_btf(-cospi[56], bf0[12], cospi[8], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[56], bf0[13], cospi[8], bf0[12], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[24], bf0[14], cospi[40], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[24], bf0[15], cospi[40], bf0[14], cos_bit[stage]);
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = bf0[20];
+  bf1[21] = bf0[21];
+  bf1[22] = bf0[22];
+  bf1[23] = bf0[23];
+  bf1[24] = half_btf(cospi[8], bf0[24], cospi[56], bf0[25], cos_bit[stage]);
+  bf1[25] = half_btf(-cospi[8], bf0[25], cospi[56], bf0[24], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[40], bf0[26], cospi[24], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(-cospi[40], bf0[27], cospi[24], bf0[26], cos_bit[stage]);
+  bf1[28] = half_btf(-cospi[56], bf0[28], cospi[8], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[56], bf0[29], cospi[8], bf0[28], cos_bit[stage]);
+  bf1[30] = half_btf(-cospi[24], bf0[30], cospi[40], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[24], bf0[31], cospi[40], bf0[30], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[4];
+  bf1[1] = bf0[1] + bf0[5];
+  bf1[2] = bf0[2] + bf0[6];
+  bf1[3] = bf0[3] + bf0[7];
+  bf1[4] = -bf0[4] + bf0[0];
+  bf1[5] = -bf0[5] + bf0[1];
+  bf1[6] = -bf0[6] + bf0[2];
+  bf1[7] = -bf0[7] + bf0[3];
+  bf1[8] = bf0[8] + bf0[12];
+  bf1[9] = bf0[9] + bf0[13];
+  bf1[10] = bf0[10] + bf0[14];
+  bf1[11] = bf0[11] + bf0[15];
+  bf1[12] = -bf0[12] + bf0[8];
+  bf1[13] = -bf0[13] + bf0[9];
+  bf1[14] = -bf0[14] + bf0[10];
+  bf1[15] = -bf0[15] + bf0[11];
+  bf1[16] = bf0[16] + bf0[20];
+  bf1[17] = bf0[17] + bf0[21];
+  bf1[18] = bf0[18] + bf0[22];
+  bf1[19] = bf0[19] + bf0[23];
+  bf1[20] = -bf0[20] + bf0[16];
+  bf1[21] = -bf0[21] + bf0[17];
+  bf1[22] = -bf0[22] + bf0[18];
+  bf1[23] = -bf0[23] + bf0[19];
+  bf1[24] = bf0[24] + bf0[28];
+  bf1[25] = bf0[25] + bf0[29];
+  bf1[26] = bf0[26] + bf0[30];
+  bf1[27] = bf0[27] + bf0[31];
+  bf1[28] = -bf0[28] + bf0[24];
+  bf1[29] = -bf0[29] + bf0[25];
+  bf1[30] = -bf0[30] + bf0[26];
+  bf1[31] = -bf0[31] + bf0[27];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 8
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[16], bf0[4], cospi[48], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(-cospi[16], bf0[5], cospi[48], bf0[4], cos_bit[stage]);
+  bf1[6] = half_btf(-cospi[48], bf0[6], cospi[16], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[48], bf0[7], cospi[16], bf0[6], cos_bit[stage]);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = half_btf(cospi[16], bf0[12], cospi[48], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(-cospi[16], bf0[13], cospi[48], bf0[12], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[48], bf0[14], cospi[16], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[48], bf0[15], cospi[16], bf0[14], cos_bit[stage]);
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = half_btf(cospi[16], bf0[20], cospi[48], bf0[21], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[16], bf0[21], cospi[48], bf0[20], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[48], bf0[22], cospi[16], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(cospi[48], bf0[23], cospi[16], bf0[22], cos_bit[stage]);
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = bf0[26];
+  bf1[27] = bf0[27];
+  bf1[28] = half_btf(cospi[16], bf0[28], cospi[48], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(-cospi[16], bf0[29], cospi[48], bf0[28], cos_bit[stage]);
+  bf1[30] = half_btf(-cospi[48], bf0[30], cospi[16], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[48], bf0[31], cospi[16], bf0[30], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 9
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[2];
+  bf1[1] = bf0[1] + bf0[3];
+  bf1[2] = -bf0[2] + bf0[0];
+  bf1[3] = -bf0[3] + bf0[1];
+  bf1[4] = bf0[4] + bf0[6];
+  bf1[5] = bf0[5] + bf0[7];
+  bf1[6] = -bf0[6] + bf0[4];
+  bf1[7] = -bf0[7] + bf0[5];
+  bf1[8] = bf0[8] + bf0[10];
+  bf1[9] = bf0[9] + bf0[11];
+  bf1[10] = -bf0[10] + bf0[8];
+  bf1[11] = -bf0[11] + bf0[9];
+  bf1[12] = bf0[12] + bf0[14];
+  bf1[13] = bf0[13] + bf0[15];
+  bf1[14] = -bf0[14] + bf0[12];
+  bf1[15] = -bf0[15] + bf0[13];
+  bf1[16] = bf0[16] + bf0[18];
+  bf1[17] = bf0[17] + bf0[19];
+  bf1[18] = -bf0[18] + bf0[16];
+  bf1[19] = -bf0[19] + bf0[17];
+  bf1[20] = bf0[20] + bf0[22];
+  bf1[21] = bf0[21] + bf0[23];
+  bf1[22] = -bf0[22] + bf0[20];
+  bf1[23] = -bf0[23] + bf0[21];
+  bf1[24] = bf0[24] + bf0[26];
+  bf1[25] = bf0[25] + bf0[27];
+  bf1[26] = -bf0[26] + bf0[24];
+  bf1[27] = -bf0[27] + bf0[25];
+  bf1[28] = bf0[28] + bf0[30];
+  bf1[29] = bf0[29] + bf0[31];
+  bf1[30] = -bf0[30] + bf0[28];
+  bf1[31] = -bf0[31] + bf0[29];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 10
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = half_btf(cospi[32], bf0[2], cospi[32], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(-cospi[32], bf0[3], cospi[32], bf0[2], cos_bit[stage]);
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(-cospi[32], bf0[7], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(cospi[32], bf0[10], cospi[32], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[10], cos_bit[stage]);
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = half_btf(cospi[32], bf0[14], cospi[32], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(-cospi[32], bf0[15], cospi[32], bf0[14], cos_bit[stage]);
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = half_btf(cospi[32], bf0[18], cospi[32], bf0[19], cos_bit[stage]);
+  bf1[19] = half_btf(-cospi[32], bf0[19], cospi[32], bf0[18], cos_bit[stage]);
+  bf1[20] = bf0[20];
+  bf1[21] = bf0[21];
+  bf1[22] = half_btf(cospi[32], bf0[22], cospi[32], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(-cospi[32], bf0[23], cospi[32], bf0[22], cos_bit[stage]);
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = half_btf(cospi[32], bf0[26], cospi[32], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(-cospi[32], bf0[27], cospi[32], bf0[26], cos_bit[stage]);
+  bf1[28] = bf0[28];
+  bf1[29] = bf0[29];
+  bf1[30] = half_btf(cospi[32], bf0[30], cospi[32], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(-cospi[32], bf0[31], cospi[32], bf0[30], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 11
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = -bf0[16];
+  bf1[2] = bf0[24];
+  bf1[3] = -bf0[8];
+  bf1[4] = bf0[12];
+  bf1[5] = -bf0[28];
+  bf1[6] = bf0[20];
+  bf1[7] = -bf0[4];
+  bf1[8] = bf0[6];
+  bf1[9] = -bf0[22];
+  bf1[10] = bf0[30];
+  bf1[11] = -bf0[14];
+  bf1[12] = bf0[10];
+  bf1[13] = -bf0[26];
+  bf1[14] = bf0[18];
+  bf1[15] = -bf0[2];
+  bf1[16] = bf0[3];
+  bf1[17] = -bf0[19];
+  bf1[18] = bf0[27];
+  bf1[19] = -bf0[11];
+  bf1[20] = bf0[15];
+  bf1[21] = -bf0[31];
+  bf1[22] = bf0[23];
+  bf1[23] = -bf0[7];
+  bf1[24] = bf0[5];
+  bf1[25] = -bf0[21];
+  bf1[26] = bf0[29];
+  bf1[27] = -bf0[13];
+  bf1[28] = bf0[9];
+  bf1[29] = -bf0[25];
+  bf1[30] = bf0[17];
+  bf1[31] = -bf0[1];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+#if CONFIG_TX64X64
+void av1_fdct64_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 64;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[64];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf1 = output;
+  bf1[0] = input[0] + input[63];
+  bf1[1] = input[1] + input[62];
+  bf1[2] = input[2] + input[61];
+  bf1[3] = input[3] + input[60];
+  bf1[4] = input[4] + input[59];
+  bf1[5] = input[5] + input[58];
+  bf1[6] = input[6] + input[57];
+  bf1[7] = input[7] + input[56];
+  bf1[8] = input[8] + input[55];
+  bf1[9] = input[9] + input[54];
+  bf1[10] = input[10] + input[53];
+  bf1[11] = input[11] + input[52];
+  bf1[12] = input[12] + input[51];
+  bf1[13] = input[13] + input[50];
+  bf1[14] = input[14] + input[49];
+  bf1[15] = input[15] + input[48];
+  bf1[16] = input[16] + input[47];
+  bf1[17] = input[17] + input[46];
+  bf1[18] = input[18] + input[45];
+  bf1[19] = input[19] + input[44];
+  bf1[20] = input[20] + input[43];
+  bf1[21] = input[21] + input[42];
+  bf1[22] = input[22] + input[41];
+  bf1[23] = input[23] + input[40];
+  bf1[24] = input[24] + input[39];
+  bf1[25] = input[25] + input[38];
+  bf1[26] = input[26] + input[37];
+  bf1[27] = input[27] + input[36];
+  bf1[28] = input[28] + input[35];
+  bf1[29] = input[29] + input[34];
+  bf1[30] = input[30] + input[33];
+  bf1[31] = input[31] + input[32];
+  bf1[32] = -input[32] + input[31];
+  bf1[33] = -input[33] + input[30];
+  bf1[34] = -input[34] + input[29];
+  bf1[35] = -input[35] + input[28];
+  bf1[36] = -input[36] + input[27];
+  bf1[37] = -input[37] + input[26];
+  bf1[38] = -input[38] + input[25];
+  bf1[39] = -input[39] + input[24];
+  bf1[40] = -input[40] + input[23];
+  bf1[41] = -input[41] + input[22];
+  bf1[42] = -input[42] + input[21];
+  bf1[43] = -input[43] + input[20];
+  bf1[44] = -input[44] + input[19];
+  bf1[45] = -input[45] + input[18];
+  bf1[46] = -input[46] + input[17];
+  bf1[47] = -input[47] + input[16];
+  bf1[48] = -input[48] + input[15];
+  bf1[49] = -input[49] + input[14];
+  bf1[50] = -input[50] + input[13];
+  bf1[51] = -input[51] + input[12];
+  bf1[52] = -input[52] + input[11];
+  bf1[53] = -input[53] + input[10];
+  bf1[54] = -input[54] + input[9];
+  bf1[55] = -input[55] + input[8];
+  bf1[56] = -input[56] + input[7];
+  bf1[57] = -input[57] + input[6];
+  bf1[58] = -input[58] + input[5];
+  bf1[59] = -input[59] + input[4];
+  bf1[60] = -input[60] + input[3];
+  bf1[61] = -input[61] + input[2];
+  bf1[62] = -input[62] + input[1];
+  bf1[63] = -input[63] + input[0];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[31];
+  bf1[1] = bf0[1] + bf0[30];
+  bf1[2] = bf0[2] + bf0[29];
+  bf1[3] = bf0[3] + bf0[28];
+  bf1[4] = bf0[4] + bf0[27];
+  bf1[5] = bf0[5] + bf0[26];
+  bf1[6] = bf0[6] + bf0[25];
+  bf1[7] = bf0[7] + bf0[24];
+  bf1[8] = bf0[8] + bf0[23];
+  bf1[9] = bf0[9] + bf0[22];
+  bf1[10] = bf0[10] + bf0[21];
+  bf1[11] = bf0[11] + bf0[20];
+  bf1[12] = bf0[12] + bf0[19];
+  bf1[13] = bf0[13] + bf0[18];
+  bf1[14] = bf0[14] + bf0[17];
+  bf1[15] = bf0[15] + bf0[16];
+  bf1[16] = -bf0[16] + bf0[15];
+  bf1[17] = -bf0[17] + bf0[14];
+  bf1[18] = -bf0[18] + bf0[13];
+  bf1[19] = -bf0[19] + bf0[12];
+  bf1[20] = -bf0[20] + bf0[11];
+  bf1[21] = -bf0[21] + bf0[10];
+  bf1[22] = -bf0[22] + bf0[9];
+  bf1[23] = -bf0[23] + bf0[8];
+  bf1[24] = -bf0[24] + bf0[7];
+  bf1[25] = -bf0[25] + bf0[6];
+  bf1[26] = -bf0[26] + bf0[5];
+  bf1[27] = -bf0[27] + bf0[4];
+  bf1[28] = -bf0[28] + bf0[3];
+  bf1[29] = -bf0[29] + bf0[2];
+  bf1[30] = -bf0[30] + bf0[1];
+  bf1[31] = -bf0[31] + bf0[0];
+  bf1[32] = bf0[32];
+  bf1[33] = bf0[33];
+  bf1[34] = bf0[34];
+  bf1[35] = bf0[35];
+  bf1[36] = bf0[36];
+  bf1[37] = bf0[37];
+  bf1[38] = bf0[38];
+  bf1[39] = bf0[39];
+  bf1[40] = half_btf(-cospi[32], bf0[40], cospi[32], bf0[55], cos_bit[stage]);
+  bf1[41] = half_btf(-cospi[32], bf0[41], cospi[32], bf0[54], cos_bit[stage]);
+  bf1[42] = half_btf(-cospi[32], bf0[42], cospi[32], bf0[53], cos_bit[stage]);
+  bf1[43] = half_btf(-cospi[32], bf0[43], cospi[32], bf0[52], cos_bit[stage]);
+  bf1[44] = half_btf(-cospi[32], bf0[44], cospi[32], bf0[51], cos_bit[stage]);
+  bf1[45] = half_btf(-cospi[32], bf0[45], cospi[32], bf0[50], cos_bit[stage]);
+  bf1[46] = half_btf(-cospi[32], bf0[46], cospi[32], bf0[49], cos_bit[stage]);
+  bf1[47] = half_btf(-cospi[32], bf0[47], cospi[32], bf0[48], cos_bit[stage]);
+  bf1[48] = half_btf(cospi[32], bf0[48], cospi[32], bf0[47], cos_bit[stage]);
+  bf1[49] = half_btf(cospi[32], bf0[49], cospi[32], bf0[46], cos_bit[stage]);
+  bf1[50] = half_btf(cospi[32], bf0[50], cospi[32], bf0[45], cos_bit[stage]);
+  bf1[51] = half_btf(cospi[32], bf0[51], cospi[32], bf0[44], cos_bit[stage]);
+  bf1[52] = half_btf(cospi[32], bf0[52], cospi[32], bf0[43], cos_bit[stage]);
+  bf1[53] = half_btf(cospi[32], bf0[53], cospi[32], bf0[42], cos_bit[stage]);
+  bf1[54] = half_btf(cospi[32], bf0[54], cospi[32], bf0[41], cos_bit[stage]);
+  bf1[55] = half_btf(cospi[32], bf0[55], cospi[32], bf0[40], cos_bit[stage]);
+  bf1[56] = bf0[56];
+  bf1[57] = bf0[57];
+  bf1[58] = bf0[58];
+  bf1[59] = bf0[59];
+  bf1[60] = bf0[60];
+  bf1[61] = bf0[61];
+  bf1[62] = bf0[62];
+  bf1[63] = bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[15];
+  bf1[1] = bf0[1] + bf0[14];
+  bf1[2] = bf0[2] + bf0[13];
+  bf1[3] = bf0[3] + bf0[12];
+  bf1[4] = bf0[4] + bf0[11];
+  bf1[5] = bf0[5] + bf0[10];
+  bf1[6] = bf0[6] + bf0[9];
+  bf1[7] = bf0[7] + bf0[8];
+  bf1[8] = -bf0[8] + bf0[7];
+  bf1[9] = -bf0[9] + bf0[6];
+  bf1[10] = -bf0[10] + bf0[5];
+  bf1[11] = -bf0[11] + bf0[4];
+  bf1[12] = -bf0[12] + bf0[3];
+  bf1[13] = -bf0[13] + bf0[2];
+  bf1[14] = -bf0[14] + bf0[1];
+  bf1[15] = -bf0[15] + bf0[0];
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = half_btf(-cospi[32], bf0[20], cospi[32], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[32], bf0[21], cospi[32], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[32], bf0[22], cospi[32], bf0[25], cos_bit[stage]);
+  bf1[23] = half_btf(-cospi[32], bf0[23], cospi[32], bf0[24], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[32], bf0[24], cospi[32], bf0[23], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[32], bf0[25], cospi[32], bf0[22], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[32], bf0[26], cospi[32], bf0[21], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[32], bf0[27], cospi[32], bf0[20], cos_bit[stage]);
+  bf1[28] = bf0[28];
+  bf1[29] = bf0[29];
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[31];
+  bf1[32] = bf0[32] + bf0[47];
+  bf1[33] = bf0[33] + bf0[46];
+  bf1[34] = bf0[34] + bf0[45];
+  bf1[35] = bf0[35] + bf0[44];
+  bf1[36] = bf0[36] + bf0[43];
+  bf1[37] = bf0[37] + bf0[42];
+  bf1[38] = bf0[38] + bf0[41];
+  bf1[39] = bf0[39] + bf0[40];
+  bf1[40] = -bf0[40] + bf0[39];
+  bf1[41] = -bf0[41] + bf0[38];
+  bf1[42] = -bf0[42] + bf0[37];
+  bf1[43] = -bf0[43] + bf0[36];
+  bf1[44] = -bf0[44] + bf0[35];
+  bf1[45] = -bf0[45] + bf0[34];
+  bf1[46] = -bf0[46] + bf0[33];
+  bf1[47] = -bf0[47] + bf0[32];
+  bf1[48] = -bf0[48] + bf0[63];
+  bf1[49] = -bf0[49] + bf0[62];
+  bf1[50] = -bf0[50] + bf0[61];
+  bf1[51] = -bf0[51] + bf0[60];
+  bf1[52] = -bf0[52] + bf0[59];
+  bf1[53] = -bf0[53] + bf0[58];
+  bf1[54] = -bf0[54] + bf0[57];
+  bf1[55] = -bf0[55] + bf0[56];
+  bf1[56] = bf0[56] + bf0[55];
+  bf1[57] = bf0[57] + bf0[54];
+  bf1[58] = bf0[58] + bf0[53];
+  bf1[59] = bf0[59] + bf0[52];
+  bf1[60] = bf0[60] + bf0[51];
+  bf1[61] = bf0[61] + bf0[50];
+  bf1[62] = bf0[62] + bf0[49];
+  bf1[63] = bf0[63] + bf0[48];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[7];
+  bf1[1] = bf0[1] + bf0[6];
+  bf1[2] = bf0[2] + bf0[5];
+  bf1[3] = bf0[3] + bf0[4];
+  bf1[4] = -bf0[4] + bf0[3];
+  bf1[5] = -bf0[5] + bf0[2];
+  bf1[6] = -bf0[6] + bf0[1];
+  bf1[7] = -bf0[7] + bf0[0];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(-cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[32], bf0[12], cospi[32], bf0[11], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[32], bf0[13], cospi[32], bf0[10], cos_bit[stage]);
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = bf0[16] + bf0[23];
+  bf1[17] = bf0[17] + bf0[22];
+  bf1[18] = bf0[18] + bf0[21];
+  bf1[19] = bf0[19] + bf0[20];
+  bf1[20] = -bf0[20] + bf0[19];
+  bf1[21] = -bf0[21] + bf0[18];
+  bf1[22] = -bf0[22] + bf0[17];
+  bf1[23] = -bf0[23] + bf0[16];
+  bf1[24] = -bf0[24] + bf0[31];
+  bf1[25] = -bf0[25] + bf0[30];
+  bf1[26] = -bf0[26] + bf0[29];
+  bf1[27] = -bf0[27] + bf0[28];
+  bf1[28] = bf0[28] + bf0[27];
+  bf1[29] = bf0[29] + bf0[26];
+  bf1[30] = bf0[30] + bf0[25];
+  bf1[31] = bf0[31] + bf0[24];
+  bf1[32] = bf0[32];
+  bf1[33] = bf0[33];
+  bf1[34] = bf0[34];
+  bf1[35] = bf0[35];
+  bf1[36] = half_btf(-cospi[16], bf0[36], cospi[48], bf0[59], cos_bit[stage]);
+  bf1[37] = half_btf(-cospi[16], bf0[37], cospi[48], bf0[58], cos_bit[stage]);
+  bf1[38] = half_btf(-cospi[16], bf0[38], cospi[48], bf0[57], cos_bit[stage]);
+  bf1[39] = half_btf(-cospi[16], bf0[39], cospi[48], bf0[56], cos_bit[stage]);
+  bf1[40] = half_btf(-cospi[48], bf0[40], -cospi[16], bf0[55], cos_bit[stage]);
+  bf1[41] = half_btf(-cospi[48], bf0[41], -cospi[16], bf0[54], cos_bit[stage]);
+  bf1[42] = half_btf(-cospi[48], bf0[42], -cospi[16], bf0[53], cos_bit[stage]);
+  bf1[43] = half_btf(-cospi[48], bf0[43], -cospi[16], bf0[52], cos_bit[stage]);
+  bf1[44] = bf0[44];
+  bf1[45] = bf0[45];
+  bf1[46] = bf0[46];
+  bf1[47] = bf0[47];
+  bf1[48] = bf0[48];
+  bf1[49] = bf0[49];
+  bf1[50] = bf0[50];
+  bf1[51] = bf0[51];
+  bf1[52] = half_btf(cospi[48], bf0[52], -cospi[16], bf0[43], cos_bit[stage]);
+  bf1[53] = half_btf(cospi[48], bf0[53], -cospi[16], bf0[42], cos_bit[stage]);
+  bf1[54] = half_btf(cospi[48], bf0[54], -cospi[16], bf0[41], cos_bit[stage]);
+  bf1[55] = half_btf(cospi[48], bf0[55], -cospi[16], bf0[40], cos_bit[stage]);
+  bf1[56] = half_btf(cospi[16], bf0[56], cospi[48], bf0[39], cos_bit[stage]);
+  bf1[57] = half_btf(cospi[16], bf0[57], cospi[48], bf0[38], cos_bit[stage]);
+  bf1[58] = half_btf(cospi[16], bf0[58], cospi[48], bf0[37], cos_bit[stage]);
+  bf1[59] = half_btf(cospi[16], bf0[59], cospi[48], bf0[36], cos_bit[stage]);
+  bf1[60] = bf0[60];
+  bf1[61] = bf0[61];
+  bf1[62] = bf0[62];
+  bf1[63] = bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[3];
+  bf1[1] = bf0[1] + bf0[2];
+  bf1[2] = -bf0[2] + bf0[1];
+  bf1[3] = -bf0[3] + bf0[0];
+  bf1[4] = bf0[4];
+  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[5], cos_bit[stage]);
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8] + bf0[11];
+  bf1[9] = bf0[9] + bf0[10];
+  bf1[10] = -bf0[10] + bf0[9];
+  bf1[11] = -bf0[11] + bf0[8];
+  bf1[12] = -bf0[12] + bf0[15];
+  bf1[13] = -bf0[13] + bf0[14];
+  bf1[14] = bf0[14] + bf0[13];
+  bf1[15] = bf0[15] + bf0[12];
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = half_btf(-cospi[16], bf0[18], cospi[48], bf0[29], cos_bit[stage]);
+  bf1[19] = half_btf(-cospi[16], bf0[19], cospi[48], bf0[28], cos_bit[stage]);
+  bf1[20] = half_btf(-cospi[48], bf0[20], -cospi[16], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[48], bf0[21], -cospi[16], bf0[26], cos_bit[stage]);
+  bf1[22] = bf0[22];
+  bf1[23] = bf0[23];
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = half_btf(cospi[48], bf0[26], -cospi[16], bf0[21], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[48], bf0[27], -cospi[16], bf0[20], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[16], bf0[28], cospi[48], bf0[19], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[16], bf0[29], cospi[48], bf0[18], cos_bit[stage]);
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[31];
+  bf1[32] = bf0[32] + bf0[39];
+  bf1[33] = bf0[33] + bf0[38];
+  bf1[34] = bf0[34] + bf0[37];
+  bf1[35] = bf0[35] + bf0[36];
+  bf1[36] = -bf0[36] + bf0[35];
+  bf1[37] = -bf0[37] + bf0[34];
+  bf1[38] = -bf0[38] + bf0[33];
+  bf1[39] = -bf0[39] + bf0[32];
+  bf1[40] = -bf0[40] + bf0[47];
+  bf1[41] = -bf0[41] + bf0[46];
+  bf1[42] = -bf0[42] + bf0[45];
+  bf1[43] = -bf0[43] + bf0[44];
+  bf1[44] = bf0[44] + bf0[43];
+  bf1[45] = bf0[45] + bf0[42];
+  bf1[46] = bf0[46] + bf0[41];
+  bf1[47] = bf0[47] + bf0[40];
+  bf1[48] = bf0[48] + bf0[55];
+  bf1[49] = bf0[49] + bf0[54];
+  bf1[50] = bf0[50] + bf0[53];
+  bf1[51] = bf0[51] + bf0[52];
+  bf1[52] = -bf0[52] + bf0[51];
+  bf1[53] = -bf0[53] + bf0[50];
+  bf1[54] = -bf0[54] + bf0[49];
+  bf1[55] = -bf0[55] + bf0[48];
+  bf1[56] = -bf0[56] + bf0[63];
+  bf1[57] = -bf0[57] + bf0[62];
+  bf1[58] = -bf0[58] + bf0[61];
+  bf1[59] = -bf0[59] + bf0[60];
+  bf1[60] = bf0[60] + bf0[59];
+  bf1[61] = bf0[61] + bf0[58];
+  bf1[62] = bf0[62] + bf0[57];
+  bf1[63] = bf0[63] + bf0[56];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(-cospi[32], bf0[1], cospi[32], bf0[0], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[48], bf0[3], -cospi[16], bf0[2], cos_bit[stage]);
+  bf1[4] = bf0[4] + bf0[5];
+  bf1[5] = -bf0[5] + bf0[4];
+  bf1[6] = -bf0[6] + bf0[7];
+  bf1[7] = bf0[7] + bf0[6];
+  bf1[8] = bf0[8];
+  bf1[9] = half_btf(-cospi[16], bf0[9], cospi[48], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(-cospi[48], bf0[10], -cospi[16], bf0[13], cos_bit[stage]);
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = half_btf(cospi[48], bf0[13], -cospi[16], bf0[10], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[16], bf0[14], cospi[48], bf0[9], cos_bit[stage]);
+  bf1[15] = bf0[15];
+  bf1[16] = bf0[16] + bf0[19];
+  bf1[17] = bf0[17] + bf0[18];
+  bf1[18] = -bf0[18] + bf0[17];
+  bf1[19] = -bf0[19] + bf0[16];
+  bf1[20] = -bf0[20] + bf0[23];
+  bf1[21] = -bf0[21] + bf0[22];
+  bf1[22] = bf0[22] + bf0[21];
+  bf1[23] = bf0[23] + bf0[20];
+  bf1[24] = bf0[24] + bf0[27];
+  bf1[25] = bf0[25] + bf0[26];
+  bf1[26] = -bf0[26] + bf0[25];
+  bf1[27] = -bf0[27] + bf0[24];
+  bf1[28] = -bf0[28] + bf0[31];
+  bf1[29] = -bf0[29] + bf0[30];
+  bf1[30] = bf0[30] + bf0[29];
+  bf1[31] = bf0[31] + bf0[28];
+  bf1[32] = bf0[32];
+  bf1[33] = bf0[33];
+  bf1[34] = half_btf(-cospi[8], bf0[34], cospi[56], bf0[61], cos_bit[stage]);
+  bf1[35] = half_btf(-cospi[8], bf0[35], cospi[56], bf0[60], cos_bit[stage]);
+  bf1[36] = half_btf(-cospi[56], bf0[36], -cospi[8], bf0[59], cos_bit[stage]);
+  bf1[37] = half_btf(-cospi[56], bf0[37], -cospi[8], bf0[58], cos_bit[stage]);
+  bf1[38] = bf0[38];
+  bf1[39] = bf0[39];
+  bf1[40] = bf0[40];
+  bf1[41] = bf0[41];
+  bf1[42] = half_btf(-cospi[40], bf0[42], cospi[24], bf0[53], cos_bit[stage]);
+  bf1[43] = half_btf(-cospi[40], bf0[43], cospi[24], bf0[52], cos_bit[stage]);
+  bf1[44] = half_btf(-cospi[24], bf0[44], -cospi[40], bf0[51], cos_bit[stage]);
+  bf1[45] = half_btf(-cospi[24], bf0[45], -cospi[40], bf0[50], cos_bit[stage]);
+  bf1[46] = bf0[46];
+  bf1[47] = bf0[47];
+  bf1[48] = bf0[48];
+  bf1[49] = bf0[49];
+  bf1[50] = half_btf(cospi[24], bf0[50], -cospi[40], bf0[45], cos_bit[stage]);
+  bf1[51] = half_btf(cospi[24], bf0[51], -cospi[40], bf0[44], cos_bit[stage]);
+  bf1[52] = half_btf(cospi[40], bf0[52], cospi[24], bf0[43], cos_bit[stage]);
+  bf1[53] = half_btf(cospi[40], bf0[53], cospi[24], bf0[42], cos_bit[stage]);
+  bf1[54] = bf0[54];
+  bf1[55] = bf0[55];
+  bf1[56] = bf0[56];
+  bf1[57] = bf0[57];
+  bf1[58] = half_btf(cospi[56], bf0[58], -cospi[8], bf0[37], cos_bit[stage]);
+  bf1[59] = half_btf(cospi[56], bf0[59], -cospi[8], bf0[36], cos_bit[stage]);
+  bf1[60] = half_btf(cospi[8], bf0[60], cospi[56], bf0[35], cos_bit[stage]);
+  bf1[61] = half_btf(cospi[8], bf0[61], cospi[56], bf0[34], cos_bit[stage]);
+  bf1[62] = bf0[62];
+  bf1[63] = bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[56], bf0[4], cospi[8], bf0[7], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[24], bf0[5], cospi[40], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[24], bf0[6], -cospi[40], bf0[5], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[56], bf0[7], -cospi[8], bf0[4], cos_bit[stage]);
+  bf1[8] = bf0[8] + bf0[9];
+  bf1[9] = -bf0[9] + bf0[8];
+  bf1[10] = -bf0[10] + bf0[11];
+  bf1[11] = bf0[11] + bf0[10];
+  bf1[12] = bf0[12] + bf0[13];
+  bf1[13] = -bf0[13] + bf0[12];
+  bf1[14] = -bf0[14] + bf0[15];
+  bf1[15] = bf0[15] + bf0[14];
+  bf1[16] = bf0[16];
+  bf1[17] = half_btf(-cospi[8], bf0[17], cospi[56], bf0[30], cos_bit[stage]);
+  bf1[18] = half_btf(-cospi[56], bf0[18], -cospi[8], bf0[29], cos_bit[stage]);
+  bf1[19] = bf0[19];
+  bf1[20] = bf0[20];
+  bf1[21] = half_btf(-cospi[40], bf0[21], cospi[24], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[24], bf0[22], -cospi[40], bf0[25], cos_bit[stage]);
+  bf1[23] = bf0[23];
+  bf1[24] = bf0[24];
+  bf1[25] = half_btf(cospi[24], bf0[25], -cospi[40], bf0[22], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[40], bf0[26], cospi[24], bf0[21], cos_bit[stage]);
+  bf1[27] = bf0[27];
+  bf1[28] = bf0[28];
+  bf1[29] = half_btf(cospi[56], bf0[29], -cospi[8], bf0[18], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[8], bf0[30], cospi[56], bf0[17], cos_bit[stage]);
+  bf1[31] = bf0[31];
+  bf1[32] = bf0[32] + bf0[35];
+  bf1[33] = bf0[33] + bf0[34];
+  bf1[34] = -bf0[34] + bf0[33];
+  bf1[35] = -bf0[35] + bf0[32];
+  bf1[36] = -bf0[36] + bf0[39];
+  bf1[37] = -bf0[37] + bf0[38];
+  bf1[38] = bf0[38] + bf0[37];
+  bf1[39] = bf0[39] + bf0[36];
+  bf1[40] = bf0[40] + bf0[43];
+  bf1[41] = bf0[41] + bf0[42];
+  bf1[42] = -bf0[42] + bf0[41];
+  bf1[43] = -bf0[43] + bf0[40];
+  bf1[44] = -bf0[44] + bf0[47];
+  bf1[45] = -bf0[45] + bf0[46];
+  bf1[46] = bf0[46] + bf0[45];
+  bf1[47] = bf0[47] + bf0[44];
+  bf1[48] = bf0[48] + bf0[51];
+  bf1[49] = bf0[49] + bf0[50];
+  bf1[50] = -bf0[50] + bf0[49];
+  bf1[51] = -bf0[51] + bf0[48];
+  bf1[52] = -bf0[52] + bf0[55];
+  bf1[53] = -bf0[53] + bf0[54];
+  bf1[54] = bf0[54] + bf0[53];
+  bf1[55] = bf0[55] + bf0[52];
+  bf1[56] = bf0[56] + bf0[59];
+  bf1[57] = bf0[57] + bf0[58];
+  bf1[58] = -bf0[58] + bf0[57];
+  bf1[59] = -bf0[59] + bf0[56];
+  bf1[60] = -bf0[60] + bf0[63];
+  bf1[61] = -bf0[61] + bf0[62];
+  bf1[62] = bf0[62] + bf0[61];
+  bf1[63] = bf0[63] + bf0[60];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 8
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[60], bf0[8], cospi[4], bf0[15], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[28], bf0[9], cospi[36], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[44], bf0[10], cospi[20], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[12], bf0[11], cospi[52], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[12], bf0[12], -cospi[52], bf0[11], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[44], bf0[13], -cospi[20], bf0[10], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[28], bf0[14], -cospi[36], bf0[9], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[60], bf0[15], -cospi[4], bf0[8], cos_bit[stage]);
+  bf1[16] = bf0[16] + bf0[17];
+  bf1[17] = -bf0[17] + bf0[16];
+  bf1[18] = -bf0[18] + bf0[19];
+  bf1[19] = bf0[19] + bf0[18];
+  bf1[20] = bf0[20] + bf0[21];
+  bf1[21] = -bf0[21] + bf0[20];
+  bf1[22] = -bf0[22] + bf0[23];
+  bf1[23] = bf0[23] + bf0[22];
+  bf1[24] = bf0[24] + bf0[25];
+  bf1[25] = -bf0[25] + bf0[24];
+  bf1[26] = -bf0[26] + bf0[27];
+  bf1[27] = bf0[27] + bf0[26];
+  bf1[28] = bf0[28] + bf0[29];
+  bf1[29] = -bf0[29] + bf0[28];
+  bf1[30] = -bf0[30] + bf0[31];
+  bf1[31] = bf0[31] + bf0[30];
+  bf1[32] = bf0[32];
+  bf1[33] = half_btf(-cospi[4], bf0[33], cospi[60], bf0[62], cos_bit[stage]);
+  bf1[34] = half_btf(-cospi[60], bf0[34], -cospi[4], bf0[61], cos_bit[stage]);
+  bf1[35] = bf0[35];
+  bf1[36] = bf0[36];
+  bf1[37] = half_btf(-cospi[36], bf0[37], cospi[28], bf0[58], cos_bit[stage]);
+  bf1[38] = half_btf(-cospi[28], bf0[38], -cospi[36], bf0[57], cos_bit[stage]);
+  bf1[39] = bf0[39];
+  bf1[40] = bf0[40];
+  bf1[41] = half_btf(-cospi[20], bf0[41], cospi[44], bf0[54], cos_bit[stage]);
+  bf1[42] = half_btf(-cospi[44], bf0[42], -cospi[20], bf0[53], cos_bit[stage]);
+  bf1[43] = bf0[43];
+  bf1[44] = bf0[44];
+  bf1[45] = half_btf(-cospi[52], bf0[45], cospi[12], bf0[50], cos_bit[stage]);
+  bf1[46] = half_btf(-cospi[12], bf0[46], -cospi[52], bf0[49], cos_bit[stage]);
+  bf1[47] = bf0[47];
+  bf1[48] = bf0[48];
+  bf1[49] = half_btf(cospi[12], bf0[49], -cospi[52], bf0[46], cos_bit[stage]);
+  bf1[50] = half_btf(cospi[52], bf0[50], cospi[12], bf0[45], cos_bit[stage]);
+  bf1[51] = bf0[51];
+  bf1[52] = bf0[52];
+  bf1[53] = half_btf(cospi[44], bf0[53], -cospi[20], bf0[42], cos_bit[stage]);
+  bf1[54] = half_btf(cospi[20], bf0[54], cospi[44], bf0[41], cos_bit[stage]);
+  bf1[55] = bf0[55];
+  bf1[56] = bf0[56];
+  bf1[57] = half_btf(cospi[28], bf0[57], -cospi[36], bf0[38], cos_bit[stage]);
+  bf1[58] = half_btf(cospi[36], bf0[58], cospi[28], bf0[37], cos_bit[stage]);
+  bf1[59] = bf0[59];
+  bf1[60] = bf0[60];
+  bf1[61] = half_btf(cospi[60], bf0[61], -cospi[4], bf0[34], cos_bit[stage]);
+  bf1[62] = half_btf(cospi[4], bf0[62], cospi[60], bf0[33], cos_bit[stage]);
+  bf1[63] = bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 9
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = half_btf(cospi[62], bf0[16], cospi[2], bf0[31], cos_bit[stage]);
+  bf1[17] = half_btf(cospi[30], bf0[17], cospi[34], bf0[30], cos_bit[stage]);
+  bf1[18] = half_btf(cospi[46], bf0[18], cospi[18], bf0[29], cos_bit[stage]);
+  bf1[19] = half_btf(cospi[14], bf0[19], cospi[50], bf0[28], cos_bit[stage]);
+  bf1[20] = half_btf(cospi[54], bf0[20], cospi[10], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(cospi[22], bf0[21], cospi[42], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(cospi[38], bf0[22], cospi[26], bf0[25], cos_bit[stage]);
+  bf1[23] = half_btf(cospi[6], bf0[23], cospi[58], bf0[24], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[6], bf0[24], -cospi[58], bf0[23], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[38], bf0[25], -cospi[26], bf0[22], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[22], bf0[26], -cospi[42], bf0[21], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[54], bf0[27], -cospi[10], bf0[20], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[14], bf0[28], -cospi[50], bf0[19], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[46], bf0[29], -cospi[18], bf0[18], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[30], bf0[30], -cospi[34], bf0[17], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[62], bf0[31], -cospi[2], bf0[16], cos_bit[stage]);
+  bf1[32] = bf0[32] + bf0[33];
+  bf1[33] = -bf0[33] + bf0[32];
+  bf1[34] = -bf0[34] + bf0[35];
+  bf1[35] = bf0[35] + bf0[34];
+  bf1[36] = bf0[36] + bf0[37];
+  bf1[37] = -bf0[37] + bf0[36];
+  bf1[38] = -bf0[38] + bf0[39];
+  bf1[39] = bf0[39] + bf0[38];
+  bf1[40] = bf0[40] + bf0[41];
+  bf1[41] = -bf0[41] + bf0[40];
+  bf1[42] = -bf0[42] + bf0[43];
+  bf1[43] = bf0[43] + bf0[42];
+  bf1[44] = bf0[44] + bf0[45];
+  bf1[45] = -bf0[45] + bf0[44];
+  bf1[46] = -bf0[46] + bf0[47];
+  bf1[47] = bf0[47] + bf0[46];
+  bf1[48] = bf0[48] + bf0[49];
+  bf1[49] = -bf0[49] + bf0[48];
+  bf1[50] = -bf0[50] + bf0[51];
+  bf1[51] = bf0[51] + bf0[50];
+  bf1[52] = bf0[52] + bf0[53];
+  bf1[53] = -bf0[53] + bf0[52];
+  bf1[54] = -bf0[54] + bf0[55];
+  bf1[55] = bf0[55] + bf0[54];
+  bf1[56] = bf0[56] + bf0[57];
+  bf1[57] = -bf0[57] + bf0[56];
+  bf1[58] = -bf0[58] + bf0[59];
+  bf1[59] = bf0[59] + bf0[58];
+  bf1[60] = bf0[60] + bf0[61];
+  bf1[61] = -bf0[61] + bf0[60];
+  bf1[62] = -bf0[62] + bf0[63];
+  bf1[63] = bf0[63] + bf0[62];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 10
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = bf0[20];
+  bf1[21] = bf0[21];
+  bf1[22] = bf0[22];
+  bf1[23] = bf0[23];
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = bf0[26];
+  bf1[27] = bf0[27];
+  bf1[28] = bf0[28];
+  bf1[29] = bf0[29];
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[31];
+  bf1[32] = half_btf(cospi[63], bf0[32], cospi[1], bf0[63], cos_bit[stage]);
+  bf1[33] = half_btf(cospi[31], bf0[33], cospi[33], bf0[62], cos_bit[stage]);
+  bf1[34] = half_btf(cospi[47], bf0[34], cospi[17], bf0[61], cos_bit[stage]);
+  bf1[35] = half_btf(cospi[15], bf0[35], cospi[49], bf0[60], cos_bit[stage]);
+  bf1[36] = half_btf(cospi[55], bf0[36], cospi[9], bf0[59], cos_bit[stage]);
+  bf1[37] = half_btf(cospi[23], bf0[37], cospi[41], bf0[58], cos_bit[stage]);
+  bf1[38] = half_btf(cospi[39], bf0[38], cospi[25], bf0[57], cos_bit[stage]);
+  bf1[39] = half_btf(cospi[7], bf0[39], cospi[57], bf0[56], cos_bit[stage]);
+  bf1[40] = half_btf(cospi[59], bf0[40], cospi[5], bf0[55], cos_bit[stage]);
+  bf1[41] = half_btf(cospi[27], bf0[41], cospi[37], bf0[54], cos_bit[stage]);
+  bf1[42] = half_btf(cospi[43], bf0[42], cospi[21], bf0[53], cos_bit[stage]);
+  bf1[43] = half_btf(cospi[11], bf0[43], cospi[53], bf0[52], cos_bit[stage]);
+  bf1[44] = half_btf(cospi[51], bf0[44], cospi[13], bf0[51], cos_bit[stage]);
+  bf1[45] = half_btf(cospi[19], bf0[45], cospi[45], bf0[50], cos_bit[stage]);
+  bf1[46] = half_btf(cospi[35], bf0[46], cospi[29], bf0[49], cos_bit[stage]);
+  bf1[47] = half_btf(cospi[3], bf0[47], cospi[61], bf0[48], cos_bit[stage]);
+  bf1[48] = half_btf(cospi[3], bf0[48], -cospi[61], bf0[47], cos_bit[stage]);
+  bf1[49] = half_btf(cospi[35], bf0[49], -cospi[29], bf0[46], cos_bit[stage]);
+  bf1[50] = half_btf(cospi[19], bf0[50], -cospi[45], bf0[45], cos_bit[stage]);
+  bf1[51] = half_btf(cospi[51], bf0[51], -cospi[13], bf0[44], cos_bit[stage]);
+  bf1[52] = half_btf(cospi[11], bf0[52], -cospi[53], bf0[43], cos_bit[stage]);
+  bf1[53] = half_btf(cospi[43], bf0[53], -cospi[21], bf0[42], cos_bit[stage]);
+  bf1[54] = half_btf(cospi[27], bf0[54], -cospi[37], bf0[41], cos_bit[stage]);
+  bf1[55] = half_btf(cospi[59], bf0[55], -cospi[5], bf0[40], cos_bit[stage]);
+  bf1[56] = half_btf(cospi[7], bf0[56], -cospi[57], bf0[39], cos_bit[stage]);
+  bf1[57] = half_btf(cospi[39], bf0[57], -cospi[25], bf0[38], cos_bit[stage]);
+  bf1[58] = half_btf(cospi[23], bf0[58], -cospi[41], bf0[37], cos_bit[stage]);
+  bf1[59] = half_btf(cospi[55], bf0[59], -cospi[9], bf0[36], cos_bit[stage]);
+  bf1[60] = half_btf(cospi[15], bf0[60], -cospi[49], bf0[35], cos_bit[stage]);
+  bf1[61] = half_btf(cospi[47], bf0[61], -cospi[17], bf0[34], cos_bit[stage]);
+  bf1[62] = half_btf(cospi[31], bf0[62], -cospi[33], bf0[33], cos_bit[stage]);
+  bf1[63] = half_btf(cospi[63], bf0[63], -cospi[1], bf0[32], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 11
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[32];
+  bf1[2] = bf0[16];
+  bf1[3] = bf0[48];
+  bf1[4] = bf0[8];
+  bf1[5] = bf0[40];
+  bf1[6] = bf0[24];
+  bf1[7] = bf0[56];
+  bf1[8] = bf0[4];
+  bf1[9] = bf0[36];
+  bf1[10] = bf0[20];
+  bf1[11] = bf0[52];
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[44];
+  bf1[14] = bf0[28];
+  bf1[15] = bf0[60];
+  bf1[16] = bf0[2];
+  bf1[17] = bf0[34];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[50];
+  bf1[20] = bf0[10];
+  bf1[21] = bf0[42];
+  bf1[22] = bf0[26];
+  bf1[23] = bf0[58];
+  bf1[24] = bf0[6];
+  bf1[25] = bf0[38];
+  bf1[26] = bf0[22];
+  bf1[27] = bf0[54];
+  bf1[28] = bf0[14];
+  bf1[29] = bf0[46];
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[62];
+  bf1[32] = bf0[1];
+  bf1[33] = bf0[33];
+  bf1[34] = bf0[17];
+  bf1[35] = bf0[49];
+  bf1[36] = bf0[9];
+  bf1[37] = bf0[41];
+  bf1[38] = bf0[25];
+  bf1[39] = bf0[57];
+  bf1[40] = bf0[5];
+  bf1[41] = bf0[37];
+  bf1[42] = bf0[21];
+  bf1[43] = bf0[53];
+  bf1[44] = bf0[13];
+  bf1[45] = bf0[45];
+  bf1[46] = bf0[29];
+  bf1[47] = bf0[61];
+  bf1[48] = bf0[3];
+  bf1[49] = bf0[35];
+  bf1[50] = bf0[19];
+  bf1[51] = bf0[51];
+  bf1[52] = bf0[11];
+  bf1[53] = bf0[43];
+  bf1[54] = bf0[27];
+  bf1[55] = bf0[59];
+  bf1[56] = bf0[7];
+  bf1[57] = bf0[39];
+  bf1[58] = bf0[23];
+  bf1[59] = bf0[55];
+  bf1[60] = bf0[15];
+  bf1[61] = bf0[47];
+  bf1[62] = bf0[31];
+  bf1[63] = bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+#endif  // CONFIG_TX64X64
diff --git a/third_party/aom/av1/common/av1_fwd_txfm1d.h b/third_party/aom/av1/common/av1_fwd_txfm1d.h
new file mode 100644
index 0000000000..9f246717e6
--- /dev/null
+++ b/third_party/aom/av1/common/av1_fwd_txfm1d.h
@@ -0,0 +1,45 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_FWD_TXFM1D_H_
+#define AV1_FWD_TXFM1D_H_
+
+#include "av1/common/av1_txfm.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_fdct4_new(const int32_t *input, int32_t *output, const int8_t *cos_bit,
+                   const int8_t *stage_range);
+void av1_fdct8_new(const int32_t *input, int32_t *output, const int8_t *cos_bit,
+                   const int8_t *stage_range);
+void av1_fdct16_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fdct32_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fdct64_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
+
+void av1_fadst4_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fadst8_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fadst16_new(const int32_t *input, int32_t *output,
+                     const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fadst32_new(const int32_t *input, int32_t *output,
+                     const int8_t *cos_bit, const int8_t *stage_range);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // AV1_FWD_TXFM1D_H_
diff --git a/third_party/aom/av1/common/av1_fwd_txfm2d.c b/third_party/aom/av1/common/av1_fwd_txfm2d.c
new file mode 100644
index 0000000000..d1dba82ca1
--- /dev/null
+++ b/third_party/aom/av1/common/av1_fwd_txfm2d.c
@@ -0,0 +1,230 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "./av1_rtcd.h"
+#include "av1/common/enums.h"
+#include "av1/common/av1_fwd_txfm1d.h"
+#include "av1/common/av1_fwd_txfm2d_cfg.h"
+#include "av1/common/av1_txfm.h"
+
+static INLINE TxfmFunc fwd_txfm_type_to_func(TXFM_TYPE txfm_type) {
+  switch (txfm_type) {
+    case TXFM_TYPE_DCT4: return av1_fdct4_new;
+    case TXFM_TYPE_DCT8: return av1_fdct8_new;
+    case TXFM_TYPE_DCT16: return av1_fdct16_new;
+    case TXFM_TYPE_DCT32: return av1_fdct32_new;
+    case TXFM_TYPE_ADST4: return av1_fadst4_new;
+    case TXFM_TYPE_ADST8: return av1_fadst8_new;
+    case TXFM_TYPE_ADST16: return av1_fadst16_new;
+    case TXFM_TYPE_ADST32: return av1_fadst32_new;
+    default: assert(0); return NULL;
+  }
+}
+
+static INLINE void fwd_txfm2d_c(const int16_t *input, int32_t *output,
+                                const int stride, const TXFM_2D_FLIP_CFG *cfg,
+                                int32_t *buf) {
+  int c, r;
+  const int txfm_size = cfg->cfg->txfm_size;
+  const int8_t *shift = cfg->cfg->shift;
+  const int8_t *stage_range_col = cfg->cfg->stage_range_col;
+  const int8_t *stage_range_row = cfg->cfg->stage_range_row;
+  const int8_t *cos_bit_col = cfg->cfg->cos_bit_col;
+  const int8_t *cos_bit_row = cfg->cfg->cos_bit_row;
+  const TxfmFunc txfm_func_col = fwd_txfm_type_to_func(cfg->cfg->txfm_type_col);
+  const TxfmFunc txfm_func_row = fwd_txfm_type_to_func(cfg->cfg->txfm_type_row);
+
+  // use output buffer as temp buffer
+  int32_t *temp_in = output;
+  int32_t *temp_out = output + txfm_size;
+
+  // Columns
+  for (c = 0; c < txfm_size; ++c) {
+    if (cfg->ud_flip == 0) {
+      for (r = 0; r < txfm_size; ++r) temp_in[r] = input[r * stride + c];
+    } else {
+      for (r = 0; r < txfm_size; ++r)
+        // flip upside down
+        temp_in[r] = input[(txfm_size - r - 1) * stride + c];
+    }
+    round_shift_array(temp_in, txfm_size, -shift[0]);
+    txfm_func_col(temp_in, temp_out, cos_bit_col, stage_range_col);
+    round_shift_array(temp_out, txfm_size, -shift[1]);
+    if (cfg->lr_flip == 0) {
+      for (r = 0; r < txfm_size; ++r) buf[r * txfm_size + c] = temp_out[r];
+    } else {
+      for (r = 0; r < txfm_size; ++r)
+        // flip from left to right
+        buf[r * txfm_size + (txfm_size - c - 1)] = temp_out[r];
+    }
+  }
+
+  // Rows
+  for (r = 0; r < txfm_size; ++r) {
+    txfm_func_row(buf + r * txfm_size, output + r * txfm_size, cos_bit_row,
+                  stage_range_row);
+    round_shift_array(output + r * txfm_size, txfm_size, -shift[2]);
+  }
+}
+
+void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride,
+                          int tx_type, int bd) {
+  int32_t txfm_buf[4 * 4];
+  TXFM_2D_FLIP_CFG cfg = av1_get_fwd_txfm_cfg(tx_type, TX_4X4);
+  (void)bd;
+  fwd_txfm2d_c(input, output, stride, &cfg, txfm_buf);
+}
+
+void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride,
+                          int tx_type, int bd) {
+  int32_t txfm_buf[8 * 8];
+  TXFM_2D_FLIP_CFG cfg = av1_get_fwd_txfm_cfg(tx_type, TX_8X8);
+  (void)bd;
+  fwd_txfm2d_c(input, output, stride, &cfg, txfm_buf);
+}
+
+void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride,
+                            int tx_type, int bd) {
+  int32_t txfm_buf[16 * 16];
+  TXFM_2D_FLIP_CFG cfg = av1_get_fwd_txfm_cfg(tx_type, TX_16X16);
+  (void)bd;
+  fwd_txfm2d_c(input, output, stride, &cfg, txfm_buf);
+}
+
+void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride,
+                            int tx_type, int bd) {
+  int32_t txfm_buf[32 * 32];
+  TXFM_2D_FLIP_CFG cfg = av1_get_fwd_txfm_cfg(tx_type, TX_32X32);
+  (void)bd;
+  fwd_txfm2d_c(input, output, stride, &cfg, txfm_buf);
+}
+
+void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride,
+                            int tx_type, int bd) {
+  int32_t txfm_buf[64 * 64];
+  TXFM_2D_FLIP_CFG cfg = av1_get_fwd_txfm_64x64_cfg(tx_type);
+  (void)bd;
+  fwd_txfm2d_c(input, output, stride, &cfg, txfm_buf);
+}
+
+#if CONFIG_EXT_TX
+static const TXFM_2D_CFG *fwd_txfm_cfg_ls[FLIPADST_ADST + 1][TX_SIZES] = {
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_dct_dct_4, &fwd_txfm_2d_cfg_dct_dct_8,
+      &fwd_txfm_2d_cfg_dct_dct_16, &fwd_txfm_2d_cfg_dct_dct_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_adst_dct_4, &fwd_txfm_2d_cfg_adst_dct_8,
+      &fwd_txfm_2d_cfg_adst_dct_16, &fwd_txfm_2d_cfg_adst_dct_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_dct_adst_4, &fwd_txfm_2d_cfg_dct_adst_8,
+      &fwd_txfm_2d_cfg_dct_adst_16, &fwd_txfm_2d_cfg_dct_adst_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_adst_adst_4, &fwd_txfm_2d_cfg_adst_adst_8,
+      &fwd_txfm_2d_cfg_adst_adst_16, &fwd_txfm_2d_cfg_adst_adst_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_adst_dct_4, &fwd_txfm_2d_cfg_adst_dct_8,
+      &fwd_txfm_2d_cfg_adst_dct_16, &fwd_txfm_2d_cfg_adst_dct_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_dct_adst_4, &fwd_txfm_2d_cfg_dct_adst_8,
+      &fwd_txfm_2d_cfg_dct_adst_16, &fwd_txfm_2d_cfg_dct_adst_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_adst_adst_4, &fwd_txfm_2d_cfg_adst_adst_8,
+      &fwd_txfm_2d_cfg_adst_adst_16, &fwd_txfm_2d_cfg_adst_adst_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_adst_adst_4, &fwd_txfm_2d_cfg_adst_adst_8,
+      &fwd_txfm_2d_cfg_adst_adst_16, &fwd_txfm_2d_cfg_adst_adst_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_adst_adst_4, &fwd_txfm_2d_cfg_adst_adst_8,
+      &fwd_txfm_2d_cfg_adst_adst_16, &fwd_txfm_2d_cfg_adst_adst_32 },
+};
+#else  // CONFIG_EXT_TX
+static const TXFM_2D_CFG *fwd_txfm_cfg_ls[TX_TYPES][TX_SIZES] = {
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_dct_dct_4, &fwd_txfm_2d_cfg_dct_dct_8,
+      &fwd_txfm_2d_cfg_dct_dct_16, &fwd_txfm_2d_cfg_dct_dct_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_adst_dct_4, &fwd_txfm_2d_cfg_adst_dct_8,
+      &fwd_txfm_2d_cfg_adst_dct_16, &fwd_txfm_2d_cfg_adst_dct_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_dct_adst_4, &fwd_txfm_2d_cfg_dct_adst_8,
+      &fwd_txfm_2d_cfg_dct_adst_16, &fwd_txfm_2d_cfg_dct_adst_32 },
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &fwd_txfm_2d_cfg_adst_adst_4, &fwd_txfm_2d_cfg_adst_adst_8,
+      &fwd_txfm_2d_cfg_adst_adst_16, &fwd_txfm_2d_cfg_adst_adst_32 },
+};
+#endif  // CONFIG_EXT_TX
+
+TXFM_2D_FLIP_CFG av1_get_fwd_txfm_cfg(int tx_type, int tx_size) {
+  TXFM_2D_FLIP_CFG cfg;
+  set_flip_cfg(tx_type, &cfg);
+  cfg.cfg = fwd_txfm_cfg_ls[tx_type][tx_size];
+  return cfg;
+}
+
+TXFM_2D_FLIP_CFG av1_get_fwd_txfm_64x64_cfg(int tx_type) {
+  TXFM_2D_FLIP_CFG cfg;
+  switch (tx_type) {
+    case DCT_DCT:
+      cfg.cfg = &fwd_txfm_2d_cfg_dct_dct_64;
+      cfg.ud_flip = 0;
+      cfg.lr_flip = 0;
+      break;
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    default:
+      cfg.ud_flip = 0;
+      cfg.lr_flip = 0;
+      assert(0);
+  }
+  return cfg;
+}
diff --git a/third_party/aom/av1/common/av1_fwd_txfm2d_cfg.h b/third_party/aom/av1/common/av1_fwd_txfm2d_cfg.h
new file mode 100644
index 0000000000..b5c828286b
--- /dev/null
+++ b/third_party/aom/av1/common/av1_fwd_txfm2d_cfg.h
@@ -0,0 +1,444 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_FWD_TXFM2D_CFG_H_
+#define AV1_FWD_TXFM2D_CFG_H_
+#include "av1/common/enums.h"
+#include "av1/common/av1_fwd_txfm1d.h"
+//  ---------------- config fwd_dct_dct_4 ----------------
+static const int8_t fwd_shift_dct_dct_4[3] = { 2, 0, 0 };
+static const int8_t fwd_stage_range_col_dct_dct_4[4] = { 15, 16, 17, 17 };
+static const int8_t fwd_stage_range_row_dct_dct_4[4] = { 17, 18, 18, 18 };
+static const int8_t fwd_cos_bit_col_dct_dct_4[4] = { 13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_dct_dct_4[4] = { 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_dct_dct_4 = {
+  4,  // .txfm_size
+  4,  // .stage_num_col
+  4,  // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_dct_dct_4,            // .shift
+  fwd_stage_range_col_dct_dct_4,  // .stage_range_col
+  fwd_stage_range_row_dct_dct_4,  // .stage_range_row
+  fwd_cos_bit_col_dct_dct_4,      // .cos_bit_col
+  fwd_cos_bit_row_dct_dct_4,      // .cos_bit_row
+  TXFM_TYPE_DCT4,                 // .txfm_type_col
+  TXFM_TYPE_DCT4
+};  // .txfm_type_row
+
+//  ---------------- config fwd_dct_dct_8 ----------------
+static const int8_t fwd_shift_dct_dct_8[3] = { 2, -1, 0 };
+static const int8_t fwd_stage_range_col_dct_dct_8[6] = {
+  15, 16, 17, 18, 18, 18
+};
+static const int8_t fwd_stage_range_row_dct_dct_8[6] = {
+  17, 18, 19, 19, 19, 19
+};
+static const int8_t fwd_cos_bit_col_dct_dct_8[6] = { 13, 13, 13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_dct_dct_8[6] = { 13, 13, 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_dct_dct_8 = {
+  8,  // .txfm_size
+  6,  // .stage_num_col
+  6,  // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_dct_dct_8,            // .shift
+  fwd_stage_range_col_dct_dct_8,  // .stage_range_col
+  fwd_stage_range_row_dct_dct_8,  // .stage_range_row
+  fwd_cos_bit_col_dct_dct_8,      // .cos_bit_col
+  fwd_cos_bit_row_dct_dct_8,      // .cos_bit_row
+  TXFM_TYPE_DCT8,                 // .txfm_type_col
+  TXFM_TYPE_DCT8
+};  // .txfm_type_row
+
+//  ---------------- config fwd_dct_dct_16 ----------------
+static const int8_t fwd_shift_dct_dct_16[3] = { 2, -2, 0 };
+static const int8_t fwd_stage_range_col_dct_dct_16[8] = { 15, 16, 17, 18,
+                                                          19, 19, 19, 19 };
+static const int8_t fwd_stage_range_row_dct_dct_16[8] = { 17, 18, 19, 20,
+                                                          20, 20, 20, 20 };
+static const int8_t fwd_cos_bit_col_dct_dct_16[8] = { 13, 13, 13, 13,
+                                                      13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_dct_dct_16[8] = { 12, 12, 12, 12,
+                                                      12, 12, 12, 12 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_dct_dct_16 = {
+  16,  // .txfm_size
+  8,   // .stage_num_col
+  8,   // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_dct_dct_16,            // .shift
+  fwd_stage_range_col_dct_dct_16,  // .stage_range_col
+  fwd_stage_range_row_dct_dct_16,  // .stage_range_row
+  fwd_cos_bit_col_dct_dct_16,      // .cos_bit_col
+  fwd_cos_bit_row_dct_dct_16,      // .cos_bit_row
+  TXFM_TYPE_DCT16,                 // .txfm_type_col
+  TXFM_TYPE_DCT16
+};  // .txfm_type_row
+
+//  ---------------- config fwd_dct_dct_32 ----------------
+static const int8_t fwd_shift_dct_dct_32[3] = { 2, -4, 0 };
+static const int8_t fwd_stage_range_col_dct_dct_32[10] = { 15, 16, 17, 18, 19,
+                                                           20, 20, 20, 20, 20 };
+static const int8_t fwd_stage_range_row_dct_dct_32[10] = { 16, 17, 18, 19, 20,
+                                                           20, 20, 20, 20, 20 };
+static const int8_t fwd_cos_bit_col_dct_dct_32[10] = { 12, 12, 12, 12, 12,
+                                                       12, 12, 12, 12, 12 };
+static const int8_t fwd_cos_bit_row_dct_dct_32[10] = { 12, 12, 12, 12, 12,
+                                                       12, 12, 12, 12, 12 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_dct_dct_32 = {
+  32,  // .txfm_size
+  10,  // .stage_num_col
+  10,  // .stage_num_row
+  // 1,  // .log_scale
+  fwd_shift_dct_dct_32,            // .shift
+  fwd_stage_range_col_dct_dct_32,  // .stage_range_col
+  fwd_stage_range_row_dct_dct_32,  // .stage_range_row
+  fwd_cos_bit_col_dct_dct_32,      // .cos_bit_col
+  fwd_cos_bit_row_dct_dct_32,      // .cos_bit_row
+  TXFM_TYPE_DCT32,                 // .txfm_type_col
+  TXFM_TYPE_DCT32
+};  // .txfm_type_row
+
+//  ---------------- config fwd_dct_dct_64 ----------------
+static const int8_t fwd_shift_dct_dct_64[3] = { 0, -2, -2 };
+static const int8_t fwd_stage_range_col_dct_dct_64[12] = {
+  13, 14, 15, 16, 17, 18, 19, 19, 19, 19, 19, 19
+};
+static const int8_t fwd_stage_range_row_dct_dct_64[12] = {
+  17, 18, 19, 20, 21, 22, 22, 22, 22, 22, 22, 22
+};
+static const int8_t fwd_cos_bit_col_dct_dct_64[12] = { 15, 15, 15, 15, 15, 14,
+                                                       13, 13, 13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_dct_dct_64[12] = { 15, 14, 13, 12, 11, 10,
+                                                       10, 10, 10, 10, 10, 10 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_dct_dct_64 = {
+  64,                              // .txfm_size
+  12,                              // .stage_num_col
+  12,                              // .stage_num_row
+  fwd_shift_dct_dct_64,            // .shift
+  fwd_stage_range_col_dct_dct_64,  // .stage_range_col
+  fwd_stage_range_row_dct_dct_64,  // .stage_range_row
+  fwd_cos_bit_col_dct_dct_64,      // .cos_bit_col
+  fwd_cos_bit_row_dct_dct_64,      // .cos_bit_row
+  TXFM_TYPE_DCT64,                 // .txfm_type_col
+  TXFM_TYPE_DCT64
+};  // .txfm_type_row
+
+//  ---------------- config fwd_dct_adst_4 ----------------
+static const int8_t fwd_shift_dct_adst_4[3] = { 2, 0, 0 };
+static const int8_t fwd_stage_range_col_dct_adst_4[4] = { 15, 16, 17, 17 };
+static const int8_t fwd_stage_range_row_dct_adst_4[6] = {
+  17, 17, 17, 18, 18, 18
+};
+static const int8_t fwd_cos_bit_col_dct_adst_4[4] = { 13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_dct_adst_4[6] = { 13, 13, 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_dct_adst_4 = {
+  4,  // .txfm_size
+  4,  // .stage_num_col
+  6,  // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_dct_adst_4,            // .shift
+  fwd_stage_range_col_dct_adst_4,  // .stage_range_col
+  fwd_stage_range_row_dct_adst_4,  // .stage_range_row
+  fwd_cos_bit_col_dct_adst_4,      // .cos_bit_col
+  fwd_cos_bit_row_dct_adst_4,      // .cos_bit_row
+  TXFM_TYPE_DCT4,                  // .txfm_type_col
+  TXFM_TYPE_ADST4
+};  // .txfm_type_row
+
+//  ---------------- config fwd_dct_adst_8 ----------------
+static const int8_t fwd_shift_dct_adst_8[3] = { 2, -1, 0 };
+static const int8_t fwd_stage_range_col_dct_adst_8[6] = {
+  15, 16, 17, 18, 18, 18
+};
+static const int8_t fwd_stage_range_row_dct_adst_8[8] = { 17, 17, 17, 18,
+                                                          18, 19, 19, 19 };
+static const int8_t fwd_cos_bit_col_dct_adst_8[6] = { 13, 13, 13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_dct_adst_8[8] = { 13, 13, 13, 13,
+                                                      13, 13, 13, 13 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_dct_adst_8 = {
+  8,  // .txfm_size
+  6,  // .stage_num_col
+  8,  // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_dct_adst_8,            // .shift
+  fwd_stage_range_col_dct_adst_8,  // .stage_range_col
+  fwd_stage_range_row_dct_adst_8,  // .stage_range_row
+  fwd_cos_bit_col_dct_adst_8,      // .cos_bit_col
+  fwd_cos_bit_row_dct_adst_8,      // .cos_bit_row
+  TXFM_TYPE_DCT8,                  // .txfm_type_col
+  TXFM_TYPE_ADST8
+};  // .txfm_type_row
+
+//  ---------------- config fwd_dct_adst_16 ----------------
+static const int8_t fwd_shift_dct_adst_16[3] = { 2, -2, 0 };
+static const int8_t fwd_stage_range_col_dct_adst_16[8] = { 15, 16, 17, 18,
+                                                           19, 19, 19, 19 };
+static const int8_t fwd_stage_range_row_dct_adst_16[10] = {
+  17, 17, 17, 18, 18, 19, 19, 20, 20, 20
+};
+static const int8_t fwd_cos_bit_col_dct_adst_16[8] = { 13, 13, 13, 13,
+                                                       13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_dct_adst_16[10] = { 12, 12, 12, 12, 12,
+                                                        12, 12, 12, 12, 12 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_dct_adst_16 = {
+  16,  // .txfm_size
+  8,   // .stage_num_col
+  10,  // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_dct_adst_16,            // .shift
+  fwd_stage_range_col_dct_adst_16,  // .stage_range_col
+  fwd_stage_range_row_dct_adst_16,  // .stage_range_row
+  fwd_cos_bit_col_dct_adst_16,      // .cos_bit_col
+  fwd_cos_bit_row_dct_adst_16,      // .cos_bit_row
+  TXFM_TYPE_DCT16,                  // .txfm_type_col
+  TXFM_TYPE_ADST16
+};  // .txfm_type_row
+
+//  ---------------- config fwd_dct_adst_32 ----------------
+static const int8_t fwd_shift_dct_adst_32[3] = { 2, -4, 0 };
+static const int8_t fwd_stage_range_col_dct_adst_32[10] = {
+  15, 16, 17, 18, 19, 20, 20, 20, 20, 20
+};
+static const int8_t fwd_stage_range_row_dct_adst_32[12] = {
+  16, 16, 16, 17, 17, 18, 18, 19, 19, 20, 20, 20
+};
+static const int8_t fwd_cos_bit_col_dct_adst_32[10] = { 12, 12, 12, 12, 12,
+                                                        12, 12, 12, 12, 12 };
+static const int8_t fwd_cos_bit_row_dct_adst_32[12] = {
+  12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12
+};
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_dct_adst_32 = {
+  32,  // .txfm_size
+  10,  // .stage_num_col
+  12,  // .stage_num_row
+  // 1,  // .log_scale
+  fwd_shift_dct_adst_32,            // .shift
+  fwd_stage_range_col_dct_adst_32,  // .stage_range_col
+  fwd_stage_range_row_dct_adst_32,  // .stage_range_row
+  fwd_cos_bit_col_dct_adst_32,      // .cos_bit_col
+  fwd_cos_bit_row_dct_adst_32,      // .cos_bit_row
+  TXFM_TYPE_DCT32,                  // .txfm_type_col
+  TXFM_TYPE_ADST32
+};  // .txfm_type_row
+//  ---------------- config fwd_adst_adst_4 ----------------
+static const int8_t fwd_shift_adst_adst_4[3] = { 2, 0, 0 };
+static const int8_t fwd_stage_range_col_adst_adst_4[6] = { 15, 15, 16,
+                                                           17, 17, 17 };
+static const int8_t fwd_stage_range_row_adst_adst_4[6] = { 17, 17, 17,
+                                                           18, 18, 18 };
+static const int8_t fwd_cos_bit_col_adst_adst_4[6] = { 13, 13, 13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_adst_adst_4[6] = { 13, 13, 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_adst_adst_4 = {
+  4,  // .txfm_size
+  6,  // .stage_num_col
+  6,  // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_adst_adst_4,            // .shift
+  fwd_stage_range_col_adst_adst_4,  // .stage_range_col
+  fwd_stage_range_row_adst_adst_4,  // .stage_range_row
+  fwd_cos_bit_col_adst_adst_4,      // .cos_bit_col
+  fwd_cos_bit_row_adst_adst_4,      // .cos_bit_row
+  TXFM_TYPE_ADST4,                  // .txfm_type_col
+  TXFM_TYPE_ADST4
+};  // .txfm_type_row
+
+//  ---------------- config fwd_adst_adst_8 ----------------
+static const int8_t fwd_shift_adst_adst_8[3] = { 2, -1, 0 };
+static const int8_t fwd_stage_range_col_adst_adst_8[8] = { 15, 15, 16, 17,
+                                                           17, 18, 18, 18 };
+static const int8_t fwd_stage_range_row_adst_adst_8[8] = { 17, 17, 17, 18,
+                                                           18, 19, 19, 19 };
+static const int8_t fwd_cos_bit_col_adst_adst_8[8] = { 13, 13, 13, 13,
+                                                       13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_adst_adst_8[8] = { 13, 13, 13, 13,
+                                                       13, 13, 13, 13 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_adst_adst_8 = {
+  8,  // .txfm_size
+  8,  // .stage_num_col
+  8,  // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_adst_adst_8,            // .shift
+  fwd_stage_range_col_adst_adst_8,  // .stage_range_col
+  fwd_stage_range_row_adst_adst_8,  // .stage_range_row
+  fwd_cos_bit_col_adst_adst_8,      // .cos_bit_col
+  fwd_cos_bit_row_adst_adst_8,      // .cos_bit_row
+  TXFM_TYPE_ADST8,                  // .txfm_type_col
+  TXFM_TYPE_ADST8
+};  // .txfm_type_row
+
+//  ---------------- config fwd_adst_adst_16 ----------------
+static const int8_t fwd_shift_adst_adst_16[3] = { 2, -2, 0 };
+static const int8_t fwd_stage_range_col_adst_adst_16[10] = {
+  15, 15, 16, 17, 17, 18, 18, 19, 19, 19
+};
+static const int8_t fwd_stage_range_row_adst_adst_16[10] = {
+  17, 17, 17, 18, 18, 19, 19, 20, 20, 20
+};
+static const int8_t fwd_cos_bit_col_adst_adst_16[10] = { 13, 13, 13, 13, 13,
+                                                         13, 13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_adst_adst_16[10] = { 12, 12, 12, 12, 12,
+                                                         12, 12, 12, 12, 12 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_adst_adst_16 = {
+  16,  // .txfm_size
+  10,  // .stage_num_col
+  10,  // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_adst_adst_16,            // .shift
+  fwd_stage_range_col_adst_adst_16,  // .stage_range_col
+  fwd_stage_range_row_adst_adst_16,  // .stage_range_row
+  fwd_cos_bit_col_adst_adst_16,      // .cos_bit_col
+  fwd_cos_bit_row_adst_adst_16,      // .cos_bit_row
+  TXFM_TYPE_ADST16,                  // .txfm_type_col
+  TXFM_TYPE_ADST16
+};  // .txfm_type_row
+
+//  ---------------- config fwd_adst_adst_32 ----------------
+static const int8_t fwd_shift_adst_adst_32[3] = { 2, -4, 0 };
+static const int8_t fwd_stage_range_col_adst_adst_32[12] = {
+  15, 15, 16, 17, 17, 18, 18, 19, 19, 20, 20, 20
+};
+static const int8_t fwd_stage_range_row_adst_adst_32[12] = {
+  16, 16, 16, 17, 17, 18, 18, 19, 19, 20, 20, 20
+};
+static const int8_t fwd_cos_bit_col_adst_adst_32[12] = {
+  12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12
+};
+static const int8_t fwd_cos_bit_row_adst_adst_32[12] = {
+  12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12
+};
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_adst_adst_32 = {
+  32,  // .txfm_size
+  12,  // .stage_num_col
+  12,  // .stage_num_row
+  // 1,  // .log_scale
+  fwd_shift_adst_adst_32,            // .shift
+  fwd_stage_range_col_adst_adst_32,  // .stage_range_col
+  fwd_stage_range_row_adst_adst_32,  // .stage_range_row
+  fwd_cos_bit_col_adst_adst_32,      // .cos_bit_col
+  fwd_cos_bit_row_adst_adst_32,      // .cos_bit_row
+  TXFM_TYPE_ADST32,                  // .txfm_type_col
+  TXFM_TYPE_ADST32
+};  // .txfm_type_row
+
+//  ---------------- config fwd_adst_dct_4 ----------------
+static const int8_t fwd_shift_adst_dct_4[3] = { 2, 0, 0 };
+static const int8_t fwd_stage_range_col_adst_dct_4[6] = {
+  15, 15, 16, 17, 17, 17
+};
+static const int8_t fwd_stage_range_row_adst_dct_4[4] = { 17, 18, 18, 18 };
+static const int8_t fwd_cos_bit_col_adst_dct_4[6] = { 13, 13, 13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_adst_dct_4[4] = { 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_adst_dct_4 = {
+  4,  // .txfm_size
+  6,  // .stage_num_col
+  4,  // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_adst_dct_4,            // .shift
+  fwd_stage_range_col_adst_dct_4,  // .stage_range_col
+  fwd_stage_range_row_adst_dct_4,  // .stage_range_row
+  fwd_cos_bit_col_adst_dct_4,      // .cos_bit_col
+  fwd_cos_bit_row_adst_dct_4,      // .cos_bit_row
+  TXFM_TYPE_ADST4,                 // .txfm_type_col
+  TXFM_TYPE_DCT4
+};  // .txfm_type_row
+
+//  ---------------- config fwd_adst_dct_8 ----------------
+static const int8_t fwd_shift_adst_dct_8[3] = { 2, -1, 0 };
+static const int8_t fwd_stage_range_col_adst_dct_8[8] = { 15, 15, 16, 17,
+                                                          17, 18, 18, 18 };
+static const int8_t fwd_stage_range_row_adst_dct_8[6] = {
+  17, 18, 19, 19, 19, 19
+};
+static const int8_t fwd_cos_bit_col_adst_dct_8[8] = { 13, 13, 13, 13,
+                                                      13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_adst_dct_8[6] = { 13, 13, 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_adst_dct_8 = {
+  8,  // .txfm_size
+  8,  // .stage_num_col
+  6,  // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_adst_dct_8,            // .shift
+  fwd_stage_range_col_adst_dct_8,  // .stage_range_col
+  fwd_stage_range_row_adst_dct_8,  // .stage_range_row
+  fwd_cos_bit_col_adst_dct_8,      // .cos_bit_col
+  fwd_cos_bit_row_adst_dct_8,      // .cos_bit_row
+  TXFM_TYPE_ADST8,                 // .txfm_type_col
+  TXFM_TYPE_DCT8
+};  // .txfm_type_row
+
+//  ---------------- config fwd_adst_dct_16 ----------------
+static const int8_t fwd_shift_adst_dct_16[3] = { 2, -2, 0 };
+static const int8_t fwd_stage_range_col_adst_dct_16[10] = {
+  15, 15, 16, 17, 17, 18, 18, 19, 19, 19
+};
+static const int8_t fwd_stage_range_row_adst_dct_16[8] = { 17, 18, 19, 20,
+                                                           20, 20, 20, 20 };
+static const int8_t fwd_cos_bit_col_adst_dct_16[10] = { 13, 13, 13, 13, 13,
+                                                        13, 13, 13, 13, 13 };
+static const int8_t fwd_cos_bit_row_adst_dct_16[8] = { 12, 12, 12, 12,
+                                                       12, 12, 12, 12 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_adst_dct_16 = {
+  16,  // .txfm_size
+  10,  // .stage_num_col
+  8,   // .stage_num_row
+  // 0,  // .log_scale
+  fwd_shift_adst_dct_16,            // .shift
+  fwd_stage_range_col_adst_dct_16,  // .stage_range_col
+  fwd_stage_range_row_adst_dct_16,  // .stage_range_row
+  fwd_cos_bit_col_adst_dct_16,      // .cos_bit_col
+  fwd_cos_bit_row_adst_dct_16,      // .cos_bit_row
+  TXFM_TYPE_ADST16,                 // .txfm_type_col
+  TXFM_TYPE_DCT16
+};  // .txfm_type_row
+
+//  ---------------- config fwd_adst_dct_32 ----------------
+static const int8_t fwd_shift_adst_dct_32[3] = { 2, -4, 0 };
+static const int8_t fwd_stage_range_col_adst_dct_32[12] = {
+  15, 15, 16, 17, 17, 18, 18, 19, 19, 20, 20, 20
+};
+static const int8_t fwd_stage_range_row_adst_dct_32[10] = {
+  16, 17, 18, 19, 20, 20, 20, 20, 20, 20
+};
+static const int8_t fwd_cos_bit_col_adst_dct_32[12] = {
+  12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12
+};
+static const int8_t fwd_cos_bit_row_adst_dct_32[10] = { 12, 12, 12, 12, 12,
+                                                        12, 12, 12, 12, 12 };
+
+static const TXFM_2D_CFG fwd_txfm_2d_cfg_adst_dct_32 = {
+  32,  // .txfm_size
+  12,  // .stage_num_col
+  10,  // .stage_num_row
+  // 1,  // .log_scale
+  fwd_shift_adst_dct_32,            // .shift
+  fwd_stage_range_col_adst_dct_32,  // .stage_range_col
+  fwd_stage_range_row_adst_dct_32,  // .stage_range_row
+  fwd_cos_bit_col_adst_dct_32,      // .cos_bit_col
+  fwd_cos_bit_row_adst_dct_32,      // .cos_bit_row
+  TXFM_TYPE_ADST32,                 // .txfm_type_col
+  TXFM_TYPE_DCT32
+};      // .txfm_type_row
+#endif  // AV1_FWD_TXFM2D_CFG_H_
diff --git a/third_party/aom/av1/common/av1_inv_txfm1d.c b/third_party/aom/av1/common/av1_inv_txfm1d.c
new file mode 100644
index 0000000000..54bbe9adf1
--- /dev/null
+++ b/third_party/aom/av1/common/av1_inv_txfm1d.c
@@ -0,0 +1,2334 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include "av1/common/av1_inv_txfm1d.h"
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+
+void range_check_func(int32_t stage, const int32_t *input, const int32_t *buf,
+                      int32_t size, int8_t bit) {
+  const int64_t maxValue = (1LL << (bit - 1)) - 1;
+  const int64_t minValue = -(1LL << (bit - 1));
+
+  for (int i = 0; i < size; ++i) {
+    if (buf[i] < minValue || buf[i] > maxValue) {
+      fprintf(stderr, "Error: coeffs contain out-of-range values\n");
+      fprintf(stderr, "stage: %d\n", stage);
+      fprintf(stderr, "node: %d\n", i);
+      fprintf(stderr, "allowed range: [%" PRId64 ";%" PRId64 "]\n", minValue,
+              maxValue);
+      fprintf(stderr, "coeffs: ");
+
+      fprintf(stderr, "[");
+      for (int j = 0; j < size; j++) {
+        if (j > 0) fprintf(stderr, ", ");
+        fprintf(stderr, "%d", input[j]);
+      }
+      fprintf(stderr, "]\n");
+      assert(0);
+    }
+  }
+}
+
+#define range_check(stage, input, buf, size, bit) \
+  range_check_func(stage, input, buf, size, bit)
+#else
+#define range_check(stage, input, buf, size, bit) \
+  {                                               \
+    (void)stage;                                  \
+    (void)input;                                  \
+    (void)buf;                                    \
+    (void)size;                                   \
+    (void)bit;                                    \
+  }
+#endif
+
+// TODO(angiebird): Make 1-d txfm functions static
+void av1_idct4_new(const int32_t *input, int32_t *output, const int8_t *cos_bit,
+                   const int8_t *stage_range) {
+  const int32_t size = 4;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[4];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = input[2];
+  bf1[2] = input[1];
+  bf1[3] = input[3];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(cospi[32], bf0[0], -cospi[32], bf0[1], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], -cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[16], bf0[2], cospi[48], bf0[3], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[3];
+  bf1[1] = bf0[1] + bf0[2];
+  bf1[2] = bf0[1] - bf0[2];
+  bf1[3] = bf0[0] - bf0[3];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_idct8_new(const int32_t *input, int32_t *output, const int8_t *cos_bit,
+                   const int8_t *stage_range) {
+  const int32_t size = 8;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[8];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = input[4];
+  bf1[2] = input[2];
+  bf1[3] = input[6];
+  bf1[4] = input[1];
+  bf1[5] = input[5];
+  bf1[6] = input[3];
+  bf1[7] = input[7];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[56], bf0[4], -cospi[8], bf0[7], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[24], bf0[5], -cospi[40], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[40], bf0[5], cospi[24], bf0[6], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[8], bf0[4], cospi[56], bf0[7], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(cospi[32], bf0[0], -cospi[32], bf0[1], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], -cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[16], bf0[2], cospi[48], bf0[3], cos_bit[stage]);
+  bf1[4] = bf0[4] + bf0[5];
+  bf1[5] = bf0[4] - bf0[5];
+  bf1[6] = -bf0[6] + bf0[7];
+  bf1[7] = bf0[6] + bf0[7];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[3];
+  bf1[1] = bf0[1] + bf0[2];
+  bf1[2] = bf0[1] - bf0[2];
+  bf1[3] = bf0[0] - bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[7] = bf0[7];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[7];
+  bf1[1] = bf0[1] + bf0[6];
+  bf1[2] = bf0[2] + bf0[5];
+  bf1[3] = bf0[3] + bf0[4];
+  bf1[4] = bf0[3] - bf0[4];
+  bf1[5] = bf0[2] - bf0[5];
+  bf1[6] = bf0[1] - bf0[6];
+  bf1[7] = bf0[0] - bf0[7];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_idct16_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 16;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[16];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = input[8];
+  bf1[2] = input[4];
+  bf1[3] = input[12];
+  bf1[4] = input[2];
+  bf1[5] = input[10];
+  bf1[6] = input[6];
+  bf1[7] = input[14];
+  bf1[8] = input[1];
+  bf1[9] = input[9];
+  bf1[10] = input[5];
+  bf1[11] = input[13];
+  bf1[12] = input[3];
+  bf1[13] = input[11];
+  bf1[14] = input[7];
+  bf1[15] = input[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[60], bf0[8], -cospi[4], bf0[15], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[28], bf0[9], -cospi[36], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[44], bf0[10], -cospi[20], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[12], bf0[11], -cospi[52], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[52], bf0[11], cospi[12], bf0[12], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[20], bf0[10], cospi[44], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[36], bf0[9], cospi[28], bf0[14], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[4], bf0[8], cospi[60], bf0[15], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[56], bf0[4], -cospi[8], bf0[7], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[24], bf0[5], -cospi[40], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[40], bf0[5], cospi[24], bf0[6], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[8], bf0[4], cospi[56], bf0[7], cos_bit[stage]);
+  bf1[8] = bf0[8] + bf0[9];
+  bf1[9] = bf0[8] - bf0[9];
+  bf1[10] = -bf0[10] + bf0[11];
+  bf1[11] = bf0[10] + bf0[11];
+  bf1[12] = bf0[12] + bf0[13];
+  bf1[13] = bf0[12] - bf0[13];
+  bf1[14] = -bf0[14] + bf0[15];
+  bf1[15] = bf0[14] + bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(cospi[32], bf0[0], -cospi[32], bf0[1], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], -cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[16], bf0[2], cospi[48], bf0[3], cos_bit[stage]);
+  bf1[4] = bf0[4] + bf0[5];
+  bf1[5] = bf0[4] - bf0[5];
+  bf1[6] = -bf0[6] + bf0[7];
+  bf1[7] = bf0[6] + bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = half_btf(-cospi[16], bf0[9], cospi[48], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(-cospi[48], bf0[10], -cospi[16], bf0[13], cos_bit[stage]);
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = half_btf(-cospi[16], bf0[10], cospi[48], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[48], bf0[9], cospi[16], bf0[14], cos_bit[stage]);
+  bf1[15] = bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[3];
+  bf1[1] = bf0[1] + bf0[2];
+  bf1[2] = bf0[1] - bf0[2];
+  bf1[3] = bf0[0] - bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8] + bf0[11];
+  bf1[9] = bf0[9] + bf0[10];
+  bf1[10] = bf0[9] - bf0[10];
+  bf1[11] = bf0[8] - bf0[11];
+  bf1[12] = -bf0[12] + bf0[15];
+  bf1[13] = -bf0[13] + bf0[14];
+  bf1[14] = bf0[13] + bf0[14];
+  bf1[15] = bf0[12] + bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[7];
+  bf1[1] = bf0[1] + bf0[6];
+  bf1[2] = bf0[2] + bf0[5];
+  bf1[3] = bf0[3] + bf0[4];
+  bf1[4] = bf0[3] - bf0[4];
+  bf1[5] = bf0[2] - bf0[5];
+  bf1[6] = bf0[1] - bf0[6];
+  bf1[7] = bf0[0] - bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(-cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[15];
+  bf1[1] = bf0[1] + bf0[14];
+  bf1[2] = bf0[2] + bf0[13];
+  bf1[3] = bf0[3] + bf0[12];
+  bf1[4] = bf0[4] + bf0[11];
+  bf1[5] = bf0[5] + bf0[10];
+  bf1[6] = bf0[6] + bf0[9];
+  bf1[7] = bf0[7] + bf0[8];
+  bf1[8] = bf0[7] - bf0[8];
+  bf1[9] = bf0[6] - bf0[9];
+  bf1[10] = bf0[5] - bf0[10];
+  bf1[11] = bf0[4] - bf0[11];
+  bf1[12] = bf0[3] - bf0[12];
+  bf1[13] = bf0[2] - bf0[13];
+  bf1[14] = bf0[1] - bf0[14];
+  bf1[15] = bf0[0] - bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_idct32_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 32;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[32];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = input[16];
+  bf1[2] = input[8];
+  bf1[3] = input[24];
+  bf1[4] = input[4];
+  bf1[5] = input[20];
+  bf1[6] = input[12];
+  bf1[7] = input[28];
+  bf1[8] = input[2];
+  bf1[9] = input[18];
+  bf1[10] = input[10];
+  bf1[11] = input[26];
+  bf1[12] = input[6];
+  bf1[13] = input[22];
+  bf1[14] = input[14];
+  bf1[15] = input[30];
+  bf1[16] = input[1];
+  bf1[17] = input[17];
+  bf1[18] = input[9];
+  bf1[19] = input[25];
+  bf1[20] = input[5];
+  bf1[21] = input[21];
+  bf1[22] = input[13];
+  bf1[23] = input[29];
+  bf1[24] = input[3];
+  bf1[25] = input[19];
+  bf1[26] = input[11];
+  bf1[27] = input[27];
+  bf1[28] = input[7];
+  bf1[29] = input[23];
+  bf1[30] = input[15];
+  bf1[31] = input[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = half_btf(cospi[62], bf0[16], -cospi[2], bf0[31], cos_bit[stage]);
+  bf1[17] = half_btf(cospi[30], bf0[17], -cospi[34], bf0[30], cos_bit[stage]);
+  bf1[18] = half_btf(cospi[46], bf0[18], -cospi[18], bf0[29], cos_bit[stage]);
+  bf1[19] = half_btf(cospi[14], bf0[19], -cospi[50], bf0[28], cos_bit[stage]);
+  bf1[20] = half_btf(cospi[54], bf0[20], -cospi[10], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(cospi[22], bf0[21], -cospi[42], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(cospi[38], bf0[22], -cospi[26], bf0[25], cos_bit[stage]);
+  bf1[23] = half_btf(cospi[6], bf0[23], -cospi[58], bf0[24], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[58], bf0[23], cospi[6], bf0[24], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[26], bf0[22], cospi[38], bf0[25], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[42], bf0[21], cospi[22], bf0[26], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[10], bf0[20], cospi[54], bf0[27], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[50], bf0[19], cospi[14], bf0[28], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[18], bf0[18], cospi[46], bf0[29], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[34], bf0[17], cospi[30], bf0[30], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[2], bf0[16], cospi[62], bf0[31], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[60], bf0[8], -cospi[4], bf0[15], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[28], bf0[9], -cospi[36], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[44], bf0[10], -cospi[20], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[12], bf0[11], -cospi[52], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[52], bf0[11], cospi[12], bf0[12], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[20], bf0[10], cospi[44], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[36], bf0[9], cospi[28], bf0[14], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[4], bf0[8], cospi[60], bf0[15], cos_bit[stage]);
+  bf1[16] = bf0[16] + bf0[17];
+  bf1[17] = bf0[16] - bf0[17];
+  bf1[18] = -bf0[18] + bf0[19];
+  bf1[19] = bf0[18] + bf0[19];
+  bf1[20] = bf0[20] + bf0[21];
+  bf1[21] = bf0[20] - bf0[21];
+  bf1[22] = -bf0[22] + bf0[23];
+  bf1[23] = bf0[22] + bf0[23];
+  bf1[24] = bf0[24] + bf0[25];
+  bf1[25] = bf0[24] - bf0[25];
+  bf1[26] = -bf0[26] + bf0[27];
+  bf1[27] = bf0[26] + bf0[27];
+  bf1[28] = bf0[28] + bf0[29];
+  bf1[29] = bf0[28] - bf0[29];
+  bf1[30] = -bf0[30] + bf0[31];
+  bf1[31] = bf0[30] + bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[56], bf0[4], -cospi[8], bf0[7], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[24], bf0[5], -cospi[40], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[40], bf0[5], cospi[24], bf0[6], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[8], bf0[4], cospi[56], bf0[7], cos_bit[stage]);
+  bf1[8] = bf0[8] + bf0[9];
+  bf1[9] = bf0[8] - bf0[9];
+  bf1[10] = -bf0[10] + bf0[11];
+  bf1[11] = bf0[10] + bf0[11];
+  bf1[12] = bf0[12] + bf0[13];
+  bf1[13] = bf0[12] - bf0[13];
+  bf1[14] = -bf0[14] + bf0[15];
+  bf1[15] = bf0[14] + bf0[15];
+  bf1[16] = bf0[16];
+  bf1[17] = half_btf(-cospi[8], bf0[17], cospi[56], bf0[30], cos_bit[stage]);
+  bf1[18] = half_btf(-cospi[56], bf0[18], -cospi[8], bf0[29], cos_bit[stage]);
+  bf1[19] = bf0[19];
+  bf1[20] = bf0[20];
+  bf1[21] = half_btf(-cospi[40], bf0[21], cospi[24], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[24], bf0[22], -cospi[40], bf0[25], cos_bit[stage]);
+  bf1[23] = bf0[23];
+  bf1[24] = bf0[24];
+  bf1[25] = half_btf(-cospi[40], bf0[22], cospi[24], bf0[25], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[24], bf0[21], cospi[40], bf0[26], cos_bit[stage]);
+  bf1[27] = bf0[27];
+  bf1[28] = bf0[28];
+  bf1[29] = half_btf(-cospi[8], bf0[18], cospi[56], bf0[29], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[56], bf0[17], cospi[8], bf0[30], cos_bit[stage]);
+  bf1[31] = bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(cospi[32], bf0[0], -cospi[32], bf0[1], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], -cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[16], bf0[2], cospi[48], bf0[3], cos_bit[stage]);
+  bf1[4] = bf0[4] + bf0[5];
+  bf1[5] = bf0[4] - bf0[5];
+  bf1[6] = -bf0[6] + bf0[7];
+  bf1[7] = bf0[6] + bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = half_btf(-cospi[16], bf0[9], cospi[48], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(-cospi[48], bf0[10], -cospi[16], bf0[13], cos_bit[stage]);
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = half_btf(-cospi[16], bf0[10], cospi[48], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[48], bf0[9], cospi[16], bf0[14], cos_bit[stage]);
+  bf1[15] = bf0[15];
+  bf1[16] = bf0[16] + bf0[19];
+  bf1[17] = bf0[17] + bf0[18];
+  bf1[18] = bf0[17] - bf0[18];
+  bf1[19] = bf0[16] - bf0[19];
+  bf1[20] = -bf0[20] + bf0[23];
+  bf1[21] = -bf0[21] + bf0[22];
+  bf1[22] = bf0[21] + bf0[22];
+  bf1[23] = bf0[20] + bf0[23];
+  bf1[24] = bf0[24] + bf0[27];
+  bf1[25] = bf0[25] + bf0[26];
+  bf1[26] = bf0[25] - bf0[26];
+  bf1[27] = bf0[24] - bf0[27];
+  bf1[28] = -bf0[28] + bf0[31];
+  bf1[29] = -bf0[29] + bf0[30];
+  bf1[30] = bf0[29] + bf0[30];
+  bf1[31] = bf0[28] + bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[3];
+  bf1[1] = bf0[1] + bf0[2];
+  bf1[2] = bf0[1] - bf0[2];
+  bf1[3] = bf0[0] - bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8] + bf0[11];
+  bf1[9] = bf0[9] + bf0[10];
+  bf1[10] = bf0[9] - bf0[10];
+  bf1[11] = bf0[8] - bf0[11];
+  bf1[12] = -bf0[12] + bf0[15];
+  bf1[13] = -bf0[13] + bf0[14];
+  bf1[14] = bf0[13] + bf0[14];
+  bf1[15] = bf0[12] + bf0[15];
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = half_btf(-cospi[16], bf0[18], cospi[48], bf0[29], cos_bit[stage]);
+  bf1[19] = half_btf(-cospi[16], bf0[19], cospi[48], bf0[28], cos_bit[stage]);
+  bf1[20] = half_btf(-cospi[48], bf0[20], -cospi[16], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[48], bf0[21], -cospi[16], bf0[26], cos_bit[stage]);
+  bf1[22] = bf0[22];
+  bf1[23] = bf0[23];
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = half_btf(-cospi[16], bf0[21], cospi[48], bf0[26], cos_bit[stage]);
+  bf1[27] = half_btf(-cospi[16], bf0[20], cospi[48], bf0[27], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[48], bf0[19], cospi[16], bf0[28], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[48], bf0[18], cospi[16], bf0[29], cos_bit[stage]);
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[7];
+  bf1[1] = bf0[1] + bf0[6];
+  bf1[2] = bf0[2] + bf0[5];
+  bf1[3] = bf0[3] + bf0[4];
+  bf1[4] = bf0[3] - bf0[4];
+  bf1[5] = bf0[2] - bf0[5];
+  bf1[6] = bf0[1] - bf0[6];
+  bf1[7] = bf0[0] - bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(-cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = bf0[16] + bf0[23];
+  bf1[17] = bf0[17] + bf0[22];
+  bf1[18] = bf0[18] + bf0[21];
+  bf1[19] = bf0[19] + bf0[20];
+  bf1[20] = bf0[19] - bf0[20];
+  bf1[21] = bf0[18] - bf0[21];
+  bf1[22] = bf0[17] - bf0[22];
+  bf1[23] = bf0[16] - bf0[23];
+  bf1[24] = -bf0[24] + bf0[31];
+  bf1[25] = -bf0[25] + bf0[30];
+  bf1[26] = -bf0[26] + bf0[29];
+  bf1[27] = -bf0[27] + bf0[28];
+  bf1[28] = bf0[27] + bf0[28];
+  bf1[29] = bf0[26] + bf0[29];
+  bf1[30] = bf0[25] + bf0[30];
+  bf1[31] = bf0[24] + bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 8
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[15];
+  bf1[1] = bf0[1] + bf0[14];
+  bf1[2] = bf0[2] + bf0[13];
+  bf1[3] = bf0[3] + bf0[12];
+  bf1[4] = bf0[4] + bf0[11];
+  bf1[5] = bf0[5] + bf0[10];
+  bf1[6] = bf0[6] + bf0[9];
+  bf1[7] = bf0[7] + bf0[8];
+  bf1[8] = bf0[7] - bf0[8];
+  bf1[9] = bf0[6] - bf0[9];
+  bf1[10] = bf0[5] - bf0[10];
+  bf1[11] = bf0[4] - bf0[11];
+  bf1[12] = bf0[3] - bf0[12];
+  bf1[13] = bf0[2] - bf0[13];
+  bf1[14] = bf0[1] - bf0[14];
+  bf1[15] = bf0[0] - bf0[15];
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = half_btf(-cospi[32], bf0[20], cospi[32], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[32], bf0[21], cospi[32], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[32], bf0[22], cospi[32], bf0[25], cos_bit[stage]);
+  bf1[23] = half_btf(-cospi[32], bf0[23], cospi[32], bf0[24], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[32], bf0[23], cospi[32], bf0[24], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[32], bf0[22], cospi[32], bf0[25], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[32], bf0[21], cospi[32], bf0[26], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[32], bf0[20], cospi[32], bf0[27], cos_bit[stage]);
+  bf1[28] = bf0[28];
+  bf1[29] = bf0[29];
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 9
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[31];
+  bf1[1] = bf0[1] + bf0[30];
+  bf1[2] = bf0[2] + bf0[29];
+  bf1[3] = bf0[3] + bf0[28];
+  bf1[4] = bf0[4] + bf0[27];
+  bf1[5] = bf0[5] + bf0[26];
+  bf1[6] = bf0[6] + bf0[25];
+  bf1[7] = bf0[7] + bf0[24];
+  bf1[8] = bf0[8] + bf0[23];
+  bf1[9] = bf0[9] + bf0[22];
+  bf1[10] = bf0[10] + bf0[21];
+  bf1[11] = bf0[11] + bf0[20];
+  bf1[12] = bf0[12] + bf0[19];
+  bf1[13] = bf0[13] + bf0[18];
+  bf1[14] = bf0[14] + bf0[17];
+  bf1[15] = bf0[15] + bf0[16];
+  bf1[16] = bf0[15] - bf0[16];
+  bf1[17] = bf0[14] - bf0[17];
+  bf1[18] = bf0[13] - bf0[18];
+  bf1[19] = bf0[12] - bf0[19];
+  bf1[20] = bf0[11] - bf0[20];
+  bf1[21] = bf0[10] - bf0[21];
+  bf1[22] = bf0[9] - bf0[22];
+  bf1[23] = bf0[8] - bf0[23];
+  bf1[24] = bf0[7] - bf0[24];
+  bf1[25] = bf0[6] - bf0[25];
+  bf1[26] = bf0[5] - bf0[26];
+  bf1[27] = bf0[4] - bf0[27];
+  bf1[28] = bf0[3] - bf0[28];
+  bf1[29] = bf0[2] - bf0[29];
+  bf1[30] = bf0[1] - bf0[30];
+  bf1[31] = bf0[0] - bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_iadst4_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 4;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[4];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = -input[3];
+  bf1[2] = -input[1];
+  bf1[3] = input[2];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = half_btf(cospi[32], bf0[2], cospi[32], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[32], bf0[2], -cospi[32], bf0[3], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[2];
+  bf1[1] = bf0[1] + bf0[3];
+  bf1[2] = bf0[0] - bf0[2];
+  bf1[3] = bf0[1] - bf0[3];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[8], bf0[0], cospi[56], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(cospi[56], bf0[0], -cospi[8], bf0[1], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[40], bf0[2], cospi[24], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[24], bf0[2], -cospi[40], bf0[3], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[1];
+  bf1[1] = bf0[2];
+  bf1[2] = bf0[3];
+  bf1[3] = bf0[0];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_iadst8_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 8;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[8];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = -input[7];
+  bf1[2] = -input[3];
+  bf1[3] = input[4];
+  bf1[4] = -input[1];
+  bf1[5] = input[6];
+  bf1[6] = input[2];
+  bf1[7] = -input[5];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = half_btf(cospi[32], bf0[2], cospi[32], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[32], bf0[2], -cospi[32], bf0[3], cos_bit[stage]);
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[32], bf0[6], -cospi[32], bf0[7], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[2];
+  bf1[1] = bf0[1] + bf0[3];
+  bf1[2] = bf0[0] - bf0[2];
+  bf1[3] = bf0[1] - bf0[3];
+  bf1[4] = bf0[4] + bf0[6];
+  bf1[5] = bf0[5] + bf0[7];
+  bf1[6] = bf0[4] - bf0[6];
+  bf1[7] = bf0[5] - bf0[7];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[16], bf0[4], cospi[48], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[48], bf0[4], -cospi[16], bf0[5], cos_bit[stage]);
+  bf1[6] = half_btf(-cospi[48], bf0[6], cospi[16], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[16], bf0[6], cospi[48], bf0[7], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[4];
+  bf1[1] = bf0[1] + bf0[5];
+  bf1[2] = bf0[2] + bf0[6];
+  bf1[3] = bf0[3] + bf0[7];
+  bf1[4] = bf0[0] - bf0[4];
+  bf1[5] = bf0[1] - bf0[5];
+  bf1[6] = bf0[2] - bf0[6];
+  bf1[7] = bf0[3] - bf0[7];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[4], bf0[0], cospi[60], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(cospi[60], bf0[0], -cospi[4], bf0[1], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[20], bf0[2], cospi[44], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[44], bf0[2], -cospi[20], bf0[3], cos_bit[stage]);
+  bf1[4] = half_btf(cospi[36], bf0[4], cospi[28], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[28], bf0[4], -cospi[36], bf0[5], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[52], bf0[6], cospi[12], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[12], bf0[6], -cospi[52], bf0[7], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[1];
+  bf1[1] = bf0[6];
+  bf1[2] = bf0[3];
+  bf1[3] = bf0[4];
+  bf1[4] = bf0[5];
+  bf1[5] = bf0[2];
+  bf1[6] = bf0[7];
+  bf1[7] = bf0[0];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_iadst16_new(const int32_t *input, int32_t *output,
+                     const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 16;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[16];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = -input[15];
+  bf1[2] = -input[7];
+  bf1[3] = input[8];
+  bf1[4] = -input[3];
+  bf1[5] = input[12];
+  bf1[6] = input[4];
+  bf1[7] = -input[11];
+  bf1[8] = -input[1];
+  bf1[9] = input[14];
+  bf1[10] = input[6];
+  bf1[11] = -input[9];
+  bf1[12] = input[2];
+  bf1[13] = -input[13];
+  bf1[14] = -input[5];
+  bf1[15] = input[10];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = half_btf(cospi[32], bf0[2], cospi[32], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[32], bf0[2], -cospi[32], bf0[3], cos_bit[stage]);
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[32], bf0[6], -cospi[32], bf0[7], cos_bit[stage]);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(cospi[32], bf0[10], cospi[32], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[32], bf0[10], -cospi[32], bf0[11], cos_bit[stage]);
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = half_btf(cospi[32], bf0[14], cospi[32], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[32], bf0[14], -cospi[32], bf0[15], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[2];
+  bf1[1] = bf0[1] + bf0[3];
+  bf1[2] = bf0[0] - bf0[2];
+  bf1[3] = bf0[1] - bf0[3];
+  bf1[4] = bf0[4] + bf0[6];
+  bf1[5] = bf0[5] + bf0[7];
+  bf1[6] = bf0[4] - bf0[6];
+  bf1[7] = bf0[5] - bf0[7];
+  bf1[8] = bf0[8] + bf0[10];
+  bf1[9] = bf0[9] + bf0[11];
+  bf1[10] = bf0[8] - bf0[10];
+  bf1[11] = bf0[9] - bf0[11];
+  bf1[12] = bf0[12] + bf0[14];
+  bf1[13] = bf0[13] + bf0[15];
+  bf1[14] = bf0[12] - bf0[14];
+  bf1[15] = bf0[13] - bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[16], bf0[4], cospi[48], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[48], bf0[4], -cospi[16], bf0[5], cos_bit[stage]);
+  bf1[6] = half_btf(-cospi[48], bf0[6], cospi[16], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[16], bf0[6], cospi[48], bf0[7], cos_bit[stage]);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = half_btf(cospi[16], bf0[12], cospi[48], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[48], bf0[12], -cospi[16], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[48], bf0[14], cospi[16], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[16], bf0[14], cospi[48], bf0[15], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[4];
+  bf1[1] = bf0[1] + bf0[5];
+  bf1[2] = bf0[2] + bf0[6];
+  bf1[3] = bf0[3] + bf0[7];
+  bf1[4] = bf0[0] - bf0[4];
+  bf1[5] = bf0[1] - bf0[5];
+  bf1[6] = bf0[2] - bf0[6];
+  bf1[7] = bf0[3] - bf0[7];
+  bf1[8] = bf0[8] + bf0[12];
+  bf1[9] = bf0[9] + bf0[13];
+  bf1[10] = bf0[10] + bf0[14];
+  bf1[11] = bf0[11] + bf0[15];
+  bf1[12] = bf0[8] - bf0[12];
+  bf1[13] = bf0[9] - bf0[13];
+  bf1[14] = bf0[10] - bf0[14];
+  bf1[15] = bf0[11] - bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[8], bf0[8], cospi[56], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[56], bf0[8], -cospi[8], bf0[9], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[40], bf0[10], cospi[24], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[24], bf0[10], -cospi[40], bf0[11], cos_bit[stage]);
+  bf1[12] = half_btf(-cospi[56], bf0[12], cospi[8], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[8], bf0[12], cospi[56], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[24], bf0[14], cospi[40], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[40], bf0[14], cospi[24], bf0[15], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[8];
+  bf1[1] = bf0[1] + bf0[9];
+  bf1[2] = bf0[2] + bf0[10];
+  bf1[3] = bf0[3] + bf0[11];
+  bf1[4] = bf0[4] + bf0[12];
+  bf1[5] = bf0[5] + bf0[13];
+  bf1[6] = bf0[6] + bf0[14];
+  bf1[7] = bf0[7] + bf0[15];
+  bf1[8] = bf0[0] - bf0[8];
+  bf1[9] = bf0[1] - bf0[9];
+  bf1[10] = bf0[2] - bf0[10];
+  bf1[11] = bf0[3] - bf0[11];
+  bf1[12] = bf0[4] - bf0[12];
+  bf1[13] = bf0[5] - bf0[13];
+  bf1[14] = bf0[6] - bf0[14];
+  bf1[15] = bf0[7] - bf0[15];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 8
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[2], bf0[0], cospi[62], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(cospi[62], bf0[0], -cospi[2], bf0[1], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[10], bf0[2], cospi[54], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[54], bf0[2], -cospi[10], bf0[3], cos_bit[stage]);
+  bf1[4] = half_btf(cospi[18], bf0[4], cospi[46], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[46], bf0[4], -cospi[18], bf0[5], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[26], bf0[6], cospi[38], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[38], bf0[6], -cospi[26], bf0[7], cos_bit[stage]);
+  bf1[8] = half_btf(cospi[34], bf0[8], cospi[30], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[30], bf0[8], -cospi[34], bf0[9], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[42], bf0[10], cospi[22], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[22], bf0[10], -cospi[42], bf0[11], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[50], bf0[12], cospi[14], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[14], bf0[12], -cospi[50], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[58], bf0[14], cospi[6], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[6], bf0[14], -cospi[58], bf0[15], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 9
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[1];
+  bf1[1] = bf0[14];
+  bf1[2] = bf0[3];
+  bf1[3] = bf0[12];
+  bf1[4] = bf0[5];
+  bf1[5] = bf0[10];
+  bf1[6] = bf0[7];
+  bf1[7] = bf0[8];
+  bf1[8] = bf0[9];
+  bf1[9] = bf0[6];
+  bf1[10] = bf0[11];
+  bf1[11] = bf0[4];
+  bf1[12] = bf0[13];
+  bf1[13] = bf0[2];
+  bf1[14] = bf0[15];
+  bf1[15] = bf0[0];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+void av1_iadst32_new(const int32_t *input, int32_t *output,
+                     const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 32;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[32];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = -input[31];
+  bf1[2] = -input[15];
+  bf1[3] = input[16];
+  bf1[4] = -input[7];
+  bf1[5] = input[24];
+  bf1[6] = input[8];
+  bf1[7] = -input[23];
+  bf1[8] = -input[3];
+  bf1[9] = input[28];
+  bf1[10] = input[12];
+  bf1[11] = -input[19];
+  bf1[12] = input[4];
+  bf1[13] = -input[27];
+  bf1[14] = -input[11];
+  bf1[15] = input[20];
+  bf1[16] = -input[1];
+  bf1[17] = input[30];
+  bf1[18] = input[14];
+  bf1[19] = -input[17];
+  bf1[20] = input[6];
+  bf1[21] = -input[25];
+  bf1[22] = -input[9];
+  bf1[23] = input[22];
+  bf1[24] = input[2];
+  bf1[25] = -input[29];
+  bf1[26] = -input[13];
+  bf1[27] = input[18];
+  bf1[28] = -input[5];
+  bf1[29] = input[26];
+  bf1[30] = input[10];
+  bf1[31] = -input[21];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = half_btf(cospi[32], bf0[2], cospi[32], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[32], bf0[2], -cospi[32], bf0[3], cos_bit[stage]);
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = half_btf(cospi[32], bf0[6], cospi[32], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[32], bf0[6], -cospi[32], bf0[7], cos_bit[stage]);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(cospi[32], bf0[10], cospi[32], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[32], bf0[10], -cospi[32], bf0[11], cos_bit[stage]);
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = half_btf(cospi[32], bf0[14], cospi[32], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[32], bf0[14], -cospi[32], bf0[15], cos_bit[stage]);
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = half_btf(cospi[32], bf0[18], cospi[32], bf0[19], cos_bit[stage]);
+  bf1[19] = half_btf(cospi[32], bf0[18], -cospi[32], bf0[19], cos_bit[stage]);
+  bf1[20] = bf0[20];
+  bf1[21] = bf0[21];
+  bf1[22] = half_btf(cospi[32], bf0[22], cospi[32], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(cospi[32], bf0[22], -cospi[32], bf0[23], cos_bit[stage]);
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = half_btf(cospi[32], bf0[26], cospi[32], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[32], bf0[26], -cospi[32], bf0[27], cos_bit[stage]);
+  bf1[28] = bf0[28];
+  bf1[29] = bf0[29];
+  bf1[30] = half_btf(cospi[32], bf0[30], cospi[32], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[32], bf0[30], -cospi[32], bf0[31], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[2];
+  bf1[1] = bf0[1] + bf0[3];
+  bf1[2] = bf0[0] - bf0[2];
+  bf1[3] = bf0[1] - bf0[3];
+  bf1[4] = bf0[4] + bf0[6];
+  bf1[5] = bf0[5] + bf0[7];
+  bf1[6] = bf0[4] - bf0[6];
+  bf1[7] = bf0[5] - bf0[7];
+  bf1[8] = bf0[8] + bf0[10];
+  bf1[9] = bf0[9] + bf0[11];
+  bf1[10] = bf0[8] - bf0[10];
+  bf1[11] = bf0[9] - bf0[11];
+  bf1[12] = bf0[12] + bf0[14];
+  bf1[13] = bf0[13] + bf0[15];
+  bf1[14] = bf0[12] - bf0[14];
+  bf1[15] = bf0[13] - bf0[15];
+  bf1[16] = bf0[16] + bf0[18];
+  bf1[17] = bf0[17] + bf0[19];
+  bf1[18] = bf0[16] - bf0[18];
+  bf1[19] = bf0[17] - bf0[19];
+  bf1[20] = bf0[20] + bf0[22];
+  bf1[21] = bf0[21] + bf0[23];
+  bf1[22] = bf0[20] - bf0[22];
+  bf1[23] = bf0[21] - bf0[23];
+  bf1[24] = bf0[24] + bf0[26];
+  bf1[25] = bf0[25] + bf0[27];
+  bf1[26] = bf0[24] - bf0[26];
+  bf1[27] = bf0[25] - bf0[27];
+  bf1[28] = bf0[28] + bf0[30];
+  bf1[29] = bf0[29] + bf0[31];
+  bf1[30] = bf0[28] - bf0[30];
+  bf1[31] = bf0[29] - bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[16], bf0[4], cospi[48], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[48], bf0[4], -cospi[16], bf0[5], cos_bit[stage]);
+  bf1[6] = half_btf(-cospi[48], bf0[6], cospi[16], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[16], bf0[6], cospi[48], bf0[7], cos_bit[stage]);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = half_btf(cospi[16], bf0[12], cospi[48], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[48], bf0[12], -cospi[16], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[48], bf0[14], cospi[16], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[16], bf0[14], cospi[48], bf0[15], cos_bit[stage]);
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = half_btf(cospi[16], bf0[20], cospi[48], bf0[21], cos_bit[stage]);
+  bf1[21] = half_btf(cospi[48], bf0[20], -cospi[16], bf0[21], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[48], bf0[22], cospi[16], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(cospi[16], bf0[22], cospi[48], bf0[23], cos_bit[stage]);
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = bf0[26];
+  bf1[27] = bf0[27];
+  bf1[28] = half_btf(cospi[16], bf0[28], cospi[48], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[48], bf0[28], -cospi[16], bf0[29], cos_bit[stage]);
+  bf1[30] = half_btf(-cospi[48], bf0[30], cospi[16], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[16], bf0[30], cospi[48], bf0[31], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[4];
+  bf1[1] = bf0[1] + bf0[5];
+  bf1[2] = bf0[2] + bf0[6];
+  bf1[3] = bf0[3] + bf0[7];
+  bf1[4] = bf0[0] - bf0[4];
+  bf1[5] = bf0[1] - bf0[5];
+  bf1[6] = bf0[2] - bf0[6];
+  bf1[7] = bf0[3] - bf0[7];
+  bf1[8] = bf0[8] + bf0[12];
+  bf1[9] = bf0[9] + bf0[13];
+  bf1[10] = bf0[10] + bf0[14];
+  bf1[11] = bf0[11] + bf0[15];
+  bf1[12] = bf0[8] - bf0[12];
+  bf1[13] = bf0[9] - bf0[13];
+  bf1[14] = bf0[10] - bf0[14];
+  bf1[15] = bf0[11] - bf0[15];
+  bf1[16] = bf0[16] + bf0[20];
+  bf1[17] = bf0[17] + bf0[21];
+  bf1[18] = bf0[18] + bf0[22];
+  bf1[19] = bf0[19] + bf0[23];
+  bf1[20] = bf0[16] - bf0[20];
+  bf1[21] = bf0[17] - bf0[21];
+  bf1[22] = bf0[18] - bf0[22];
+  bf1[23] = bf0[19] - bf0[23];
+  bf1[24] = bf0[24] + bf0[28];
+  bf1[25] = bf0[25] + bf0[29];
+  bf1[26] = bf0[26] + bf0[30];
+  bf1[27] = bf0[27] + bf0[31];
+  bf1[28] = bf0[24] - bf0[28];
+  bf1[29] = bf0[25] - bf0[29];
+  bf1[30] = bf0[26] - bf0[30];
+  bf1[31] = bf0[27] - bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[8], bf0[8], cospi[56], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[56], bf0[8], -cospi[8], bf0[9], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[40], bf0[10], cospi[24], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[24], bf0[10], -cospi[40], bf0[11], cos_bit[stage]);
+  bf1[12] = half_btf(-cospi[56], bf0[12], cospi[8], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[8], bf0[12], cospi[56], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(-cospi[24], bf0[14], cospi[40], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[40], bf0[14], cospi[24], bf0[15], cos_bit[stage]);
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = bf0[20];
+  bf1[21] = bf0[21];
+  bf1[22] = bf0[22];
+  bf1[23] = bf0[23];
+  bf1[24] = half_btf(cospi[8], bf0[24], cospi[56], bf0[25], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[56], bf0[24], -cospi[8], bf0[25], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[40], bf0[26], cospi[24], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[24], bf0[26], -cospi[40], bf0[27], cos_bit[stage]);
+  bf1[28] = half_btf(-cospi[56], bf0[28], cospi[8], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[8], bf0[28], cospi[56], bf0[29], cos_bit[stage]);
+  bf1[30] = half_btf(-cospi[24], bf0[30], cospi[40], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[40], bf0[30], cospi[24], bf0[31], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[8];
+  bf1[1] = bf0[1] + bf0[9];
+  bf1[2] = bf0[2] + bf0[10];
+  bf1[3] = bf0[3] + bf0[11];
+  bf1[4] = bf0[4] + bf0[12];
+  bf1[5] = bf0[5] + bf0[13];
+  bf1[6] = bf0[6] + bf0[14];
+  bf1[7] = bf0[7] + bf0[15];
+  bf1[8] = bf0[0] - bf0[8];
+  bf1[9] = bf0[1] - bf0[9];
+  bf1[10] = bf0[2] - bf0[10];
+  bf1[11] = bf0[3] - bf0[11];
+  bf1[12] = bf0[4] - bf0[12];
+  bf1[13] = bf0[5] - bf0[13];
+  bf1[14] = bf0[6] - bf0[14];
+  bf1[15] = bf0[7] - bf0[15];
+  bf1[16] = bf0[16] + bf0[24];
+  bf1[17] = bf0[17] + bf0[25];
+  bf1[18] = bf0[18] + bf0[26];
+  bf1[19] = bf0[19] + bf0[27];
+  bf1[20] = bf0[20] + bf0[28];
+  bf1[21] = bf0[21] + bf0[29];
+  bf1[22] = bf0[22] + bf0[30];
+  bf1[23] = bf0[23] + bf0[31];
+  bf1[24] = bf0[16] - bf0[24];
+  bf1[25] = bf0[17] - bf0[25];
+  bf1[26] = bf0[18] - bf0[26];
+  bf1[27] = bf0[19] - bf0[27];
+  bf1[28] = bf0[20] - bf0[28];
+  bf1[29] = bf0[21] - bf0[29];
+  bf1[30] = bf0[22] - bf0[30];
+  bf1[31] = bf0[23] - bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 8
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = half_btf(cospi[4], bf0[16], cospi[60], bf0[17], cos_bit[stage]);
+  bf1[17] = half_btf(cospi[60], bf0[16], -cospi[4], bf0[17], cos_bit[stage]);
+  bf1[18] = half_btf(cospi[20], bf0[18], cospi[44], bf0[19], cos_bit[stage]);
+  bf1[19] = half_btf(cospi[44], bf0[18], -cospi[20], bf0[19], cos_bit[stage]);
+  bf1[20] = half_btf(cospi[36], bf0[20], cospi[28], bf0[21], cos_bit[stage]);
+  bf1[21] = half_btf(cospi[28], bf0[20], -cospi[36], bf0[21], cos_bit[stage]);
+  bf1[22] = half_btf(cospi[52], bf0[22], cospi[12], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(cospi[12], bf0[22], -cospi[52], bf0[23], cos_bit[stage]);
+  bf1[24] = half_btf(-cospi[60], bf0[24], cospi[4], bf0[25], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[4], bf0[24], cospi[60], bf0[25], cos_bit[stage]);
+  bf1[26] = half_btf(-cospi[44], bf0[26], cospi[20], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[20], bf0[26], cospi[44], bf0[27], cos_bit[stage]);
+  bf1[28] = half_btf(-cospi[28], bf0[28], cospi[36], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[36], bf0[28], cospi[28], bf0[29], cos_bit[stage]);
+  bf1[30] = half_btf(-cospi[12], bf0[30], cospi[52], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[52], bf0[30], cospi[12], bf0[31], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 9
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[16];
+  bf1[1] = bf0[1] + bf0[17];
+  bf1[2] = bf0[2] + bf0[18];
+  bf1[3] = bf0[3] + bf0[19];
+  bf1[4] = bf0[4] + bf0[20];
+  bf1[5] = bf0[5] + bf0[21];
+  bf1[6] = bf0[6] + bf0[22];
+  bf1[7] = bf0[7] + bf0[23];
+  bf1[8] = bf0[8] + bf0[24];
+  bf1[9] = bf0[9] + bf0[25];
+  bf1[10] = bf0[10] + bf0[26];
+  bf1[11] = bf0[11] + bf0[27];
+  bf1[12] = bf0[12] + bf0[28];
+  bf1[13] = bf0[13] + bf0[29];
+  bf1[14] = bf0[14] + bf0[30];
+  bf1[15] = bf0[15] + bf0[31];
+  bf1[16] = bf0[0] - bf0[16];
+  bf1[17] = bf0[1] - bf0[17];
+  bf1[18] = bf0[2] - bf0[18];
+  bf1[19] = bf0[3] - bf0[19];
+  bf1[20] = bf0[4] - bf0[20];
+  bf1[21] = bf0[5] - bf0[21];
+  bf1[22] = bf0[6] - bf0[22];
+  bf1[23] = bf0[7] - bf0[23];
+  bf1[24] = bf0[8] - bf0[24];
+  bf1[25] = bf0[9] - bf0[25];
+  bf1[26] = bf0[10] - bf0[26];
+  bf1[27] = bf0[11] - bf0[27];
+  bf1[28] = bf0[12] - bf0[28];
+  bf1[29] = bf0[13] - bf0[29];
+  bf1[30] = bf0[14] - bf0[30];
+  bf1[31] = bf0[15] - bf0[31];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 10
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[1], bf0[0], cospi[63], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(cospi[63], bf0[0], -cospi[1], bf0[1], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[5], bf0[2], cospi[59], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[59], bf0[2], -cospi[5], bf0[3], cos_bit[stage]);
+  bf1[4] = half_btf(cospi[9], bf0[4], cospi[55], bf0[5], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[55], bf0[4], -cospi[9], bf0[5], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[13], bf0[6], cospi[51], bf0[7], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[51], bf0[6], -cospi[13], bf0[7], cos_bit[stage]);
+  bf1[8] = half_btf(cospi[17], bf0[8], cospi[47], bf0[9], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[47], bf0[8], -cospi[17], bf0[9], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[21], bf0[10], cospi[43], bf0[11], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[43], bf0[10], -cospi[21], bf0[11], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[25], bf0[12], cospi[39], bf0[13], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[39], bf0[12], -cospi[25], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[29], bf0[14], cospi[35], bf0[15], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[35], bf0[14], -cospi[29], bf0[15], cos_bit[stage]);
+  bf1[16] = half_btf(cospi[33], bf0[16], cospi[31], bf0[17], cos_bit[stage]);
+  bf1[17] = half_btf(cospi[31], bf0[16], -cospi[33], bf0[17], cos_bit[stage]);
+  bf1[18] = half_btf(cospi[37], bf0[18], cospi[27], bf0[19], cos_bit[stage]);
+  bf1[19] = half_btf(cospi[27], bf0[18], -cospi[37], bf0[19], cos_bit[stage]);
+  bf1[20] = half_btf(cospi[41], bf0[20], cospi[23], bf0[21], cos_bit[stage]);
+  bf1[21] = half_btf(cospi[23], bf0[20], -cospi[41], bf0[21], cos_bit[stage]);
+  bf1[22] = half_btf(cospi[45], bf0[22], cospi[19], bf0[23], cos_bit[stage]);
+  bf1[23] = half_btf(cospi[19], bf0[22], -cospi[45], bf0[23], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[49], bf0[24], cospi[15], bf0[25], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[15], bf0[24], -cospi[49], bf0[25], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[53], bf0[26], cospi[11], bf0[27], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[11], bf0[26], -cospi[53], bf0[27], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[57], bf0[28], cospi[7], bf0[29], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[7], bf0[28], -cospi[57], bf0[29], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[61], bf0[30], cospi[3], bf0[31], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[3], bf0[30], -cospi[61], bf0[31], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 11
+  stage++;
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[1];
+  bf1[1] = bf0[30];
+  bf1[2] = bf0[3];
+  bf1[3] = bf0[28];
+  bf1[4] = bf0[5];
+  bf1[5] = bf0[26];
+  bf1[6] = bf0[7];
+  bf1[7] = bf0[24];
+  bf1[8] = bf0[9];
+  bf1[9] = bf0[22];
+  bf1[10] = bf0[11];
+  bf1[11] = bf0[20];
+  bf1[12] = bf0[13];
+  bf1[13] = bf0[18];
+  bf1[14] = bf0[15];
+  bf1[15] = bf0[16];
+  bf1[16] = bf0[17];
+  bf1[17] = bf0[14];
+  bf1[18] = bf0[19];
+  bf1[19] = bf0[12];
+  bf1[20] = bf0[21];
+  bf1[21] = bf0[10];
+  bf1[22] = bf0[23];
+  bf1[23] = bf0[8];
+  bf1[24] = bf0[25];
+  bf1[25] = bf0[6];
+  bf1[26] = bf0[27];
+  bf1[27] = bf0[4];
+  bf1[28] = bf0[29];
+  bf1[29] = bf0[2];
+  bf1[30] = bf0[31];
+  bf1[31] = bf0[0];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+
+#if CONFIG_TX64X64
+void av1_idct64_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range) {
+  const int32_t size = 64;
+  const int32_t *cospi;
+
+  int32_t stage = 0;
+  int32_t *bf0, *bf1;
+  int32_t step[64];
+
+  // stage 0;
+  range_check(stage, input, input, size, stage_range[stage]);
+
+  // stage 1;
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf1 = output;
+  bf1[0] = input[0];
+  bf1[1] = input[32];
+  bf1[2] = input[16];
+  bf1[3] = input[48];
+  bf1[4] = input[8];
+  bf1[5] = input[40];
+  bf1[6] = input[24];
+  bf1[7] = input[56];
+  bf1[8] = input[4];
+  bf1[9] = input[36];
+  bf1[10] = input[20];
+  bf1[11] = input[52];
+  bf1[12] = input[12];
+  bf1[13] = input[44];
+  bf1[14] = input[28];
+  bf1[15] = input[60];
+  bf1[16] = input[2];
+  bf1[17] = input[34];
+  bf1[18] = input[18];
+  bf1[19] = input[50];
+  bf1[20] = input[10];
+  bf1[21] = input[42];
+  bf1[22] = input[26];
+  bf1[23] = input[58];
+  bf1[24] = input[6];
+  bf1[25] = input[38];
+  bf1[26] = input[22];
+  bf1[27] = input[54];
+  bf1[28] = input[14];
+  bf1[29] = input[46];
+  bf1[30] = input[30];
+  bf1[31] = input[62];
+  bf1[32] = input[1];
+  bf1[33] = input[33];
+  bf1[34] = input[17];
+  bf1[35] = input[49];
+  bf1[36] = input[9];
+  bf1[37] = input[41];
+  bf1[38] = input[25];
+  bf1[39] = input[57];
+  bf1[40] = input[5];
+  bf1[41] = input[37];
+  bf1[42] = input[21];
+  bf1[43] = input[53];
+  bf1[44] = input[13];
+  bf1[45] = input[45];
+  bf1[46] = input[29];
+  bf1[47] = input[61];
+  bf1[48] = input[3];
+  bf1[49] = input[35];
+  bf1[50] = input[19];
+  bf1[51] = input[51];
+  bf1[52] = input[11];
+  bf1[53] = input[43];
+  bf1[54] = input[27];
+  bf1[55] = input[59];
+  bf1[56] = input[7];
+  bf1[57] = input[39];
+  bf1[58] = input[23];
+  bf1[59] = input[55];
+  bf1[60] = input[15];
+  bf1[61] = input[47];
+  bf1[62] = input[31];
+  bf1[63] = input[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 2
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = bf0[20];
+  bf1[21] = bf0[21];
+  bf1[22] = bf0[22];
+  bf1[23] = bf0[23];
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = bf0[26];
+  bf1[27] = bf0[27];
+  bf1[28] = bf0[28];
+  bf1[29] = bf0[29];
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[31];
+  bf1[32] = half_btf(cospi[63], bf0[32], -cospi[1], bf0[63], cos_bit[stage]);
+  bf1[33] = half_btf(cospi[31], bf0[33], -cospi[33], bf0[62], cos_bit[stage]);
+  bf1[34] = half_btf(cospi[47], bf0[34], -cospi[17], bf0[61], cos_bit[stage]);
+  bf1[35] = half_btf(cospi[15], bf0[35], -cospi[49], bf0[60], cos_bit[stage]);
+  bf1[36] = half_btf(cospi[55], bf0[36], -cospi[9], bf0[59], cos_bit[stage]);
+  bf1[37] = half_btf(cospi[23], bf0[37], -cospi[41], bf0[58], cos_bit[stage]);
+  bf1[38] = half_btf(cospi[39], bf0[38], -cospi[25], bf0[57], cos_bit[stage]);
+  bf1[39] = half_btf(cospi[7], bf0[39], -cospi[57], bf0[56], cos_bit[stage]);
+  bf1[40] = half_btf(cospi[59], bf0[40], -cospi[5], bf0[55], cos_bit[stage]);
+  bf1[41] = half_btf(cospi[27], bf0[41], -cospi[37], bf0[54], cos_bit[stage]);
+  bf1[42] = half_btf(cospi[43], bf0[42], -cospi[21], bf0[53], cos_bit[stage]);
+  bf1[43] = half_btf(cospi[11], bf0[43], -cospi[53], bf0[52], cos_bit[stage]);
+  bf1[44] = half_btf(cospi[51], bf0[44], -cospi[13], bf0[51], cos_bit[stage]);
+  bf1[45] = half_btf(cospi[19], bf0[45], -cospi[45], bf0[50], cos_bit[stage]);
+  bf1[46] = half_btf(cospi[35], bf0[46], -cospi[29], bf0[49], cos_bit[stage]);
+  bf1[47] = half_btf(cospi[3], bf0[47], -cospi[61], bf0[48], cos_bit[stage]);
+  bf1[48] = half_btf(cospi[61], bf0[47], cospi[3], bf0[48], cos_bit[stage]);
+  bf1[49] = half_btf(cospi[29], bf0[46], cospi[35], bf0[49], cos_bit[stage]);
+  bf1[50] = half_btf(cospi[45], bf0[45], cospi[19], bf0[50], cos_bit[stage]);
+  bf1[51] = half_btf(cospi[13], bf0[44], cospi[51], bf0[51], cos_bit[stage]);
+  bf1[52] = half_btf(cospi[53], bf0[43], cospi[11], bf0[52], cos_bit[stage]);
+  bf1[53] = half_btf(cospi[21], bf0[42], cospi[43], bf0[53], cos_bit[stage]);
+  bf1[54] = half_btf(cospi[37], bf0[41], cospi[27], bf0[54], cos_bit[stage]);
+  bf1[55] = half_btf(cospi[5], bf0[40], cospi[59], bf0[55], cos_bit[stage]);
+  bf1[56] = half_btf(cospi[57], bf0[39], cospi[7], bf0[56], cos_bit[stage]);
+  bf1[57] = half_btf(cospi[25], bf0[38], cospi[39], bf0[57], cos_bit[stage]);
+  bf1[58] = half_btf(cospi[41], bf0[37], cospi[23], bf0[58], cos_bit[stage]);
+  bf1[59] = half_btf(cospi[9], bf0[36], cospi[55], bf0[59], cos_bit[stage]);
+  bf1[60] = half_btf(cospi[49], bf0[35], cospi[15], bf0[60], cos_bit[stage]);
+  bf1[61] = half_btf(cospi[17], bf0[34], cospi[47], bf0[61], cos_bit[stage]);
+  bf1[62] = half_btf(cospi[33], bf0[33], cospi[31], bf0[62], cos_bit[stage]);
+  bf1[63] = half_btf(cospi[1], bf0[32], cospi[63], bf0[63], cos_bit[stage]);
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 3
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = bf0[10];
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = bf0[13];
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = half_btf(cospi[62], bf0[16], -cospi[2], bf0[31], cos_bit[stage]);
+  bf1[17] = half_btf(cospi[30], bf0[17], -cospi[34], bf0[30], cos_bit[stage]);
+  bf1[18] = half_btf(cospi[46], bf0[18], -cospi[18], bf0[29], cos_bit[stage]);
+  bf1[19] = half_btf(cospi[14], bf0[19], -cospi[50], bf0[28], cos_bit[stage]);
+  bf1[20] = half_btf(cospi[54], bf0[20], -cospi[10], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(cospi[22], bf0[21], -cospi[42], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(cospi[38], bf0[22], -cospi[26], bf0[25], cos_bit[stage]);
+  bf1[23] = half_btf(cospi[6], bf0[23], -cospi[58], bf0[24], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[58], bf0[23], cospi[6], bf0[24], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[26], bf0[22], cospi[38], bf0[25], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[42], bf0[21], cospi[22], bf0[26], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[10], bf0[20], cospi[54], bf0[27], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[50], bf0[19], cospi[14], bf0[28], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[18], bf0[18], cospi[46], bf0[29], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[34], bf0[17], cospi[30], bf0[30], cos_bit[stage]);
+  bf1[31] = half_btf(cospi[2], bf0[16], cospi[62], bf0[31], cos_bit[stage]);
+  bf1[32] = bf0[32] + bf0[33];
+  bf1[33] = bf0[32] - bf0[33];
+  bf1[34] = -bf0[34] + bf0[35];
+  bf1[35] = bf0[34] + bf0[35];
+  bf1[36] = bf0[36] + bf0[37];
+  bf1[37] = bf0[36] - bf0[37];
+  bf1[38] = -bf0[38] + bf0[39];
+  bf1[39] = bf0[38] + bf0[39];
+  bf1[40] = bf0[40] + bf0[41];
+  bf1[41] = bf0[40] - bf0[41];
+  bf1[42] = -bf0[42] + bf0[43];
+  bf1[43] = bf0[42] + bf0[43];
+  bf1[44] = bf0[44] + bf0[45];
+  bf1[45] = bf0[44] - bf0[45];
+  bf1[46] = -bf0[46] + bf0[47];
+  bf1[47] = bf0[46] + bf0[47];
+  bf1[48] = bf0[48] + bf0[49];
+  bf1[49] = bf0[48] - bf0[49];
+  bf1[50] = -bf0[50] + bf0[51];
+  bf1[51] = bf0[50] + bf0[51];
+  bf1[52] = bf0[52] + bf0[53];
+  bf1[53] = bf0[52] - bf0[53];
+  bf1[54] = -bf0[54] + bf0[55];
+  bf1[55] = bf0[54] + bf0[55];
+  bf1[56] = bf0[56] + bf0[57];
+  bf1[57] = bf0[56] - bf0[57];
+  bf1[58] = -bf0[58] + bf0[59];
+  bf1[59] = bf0[58] + bf0[59];
+  bf1[60] = bf0[60] + bf0[61];
+  bf1[61] = bf0[60] - bf0[61];
+  bf1[62] = -bf0[62] + bf0[63];
+  bf1[63] = bf0[62] + bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 4
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] = half_btf(cospi[60], bf0[8], -cospi[4], bf0[15], cos_bit[stage]);
+  bf1[9] = half_btf(cospi[28], bf0[9], -cospi[36], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(cospi[44], bf0[10], -cospi[20], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(cospi[12], bf0[11], -cospi[52], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[52], bf0[11], cospi[12], bf0[12], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[20], bf0[10], cospi[44], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[36], bf0[9], cospi[28], bf0[14], cos_bit[stage]);
+  bf1[15] = half_btf(cospi[4], bf0[8], cospi[60], bf0[15], cos_bit[stage]);
+  bf1[16] = bf0[16] + bf0[17];
+  bf1[17] = bf0[16] - bf0[17];
+  bf1[18] = -bf0[18] + bf0[19];
+  bf1[19] = bf0[18] + bf0[19];
+  bf1[20] = bf0[20] + bf0[21];
+  bf1[21] = bf0[20] - bf0[21];
+  bf1[22] = -bf0[22] + bf0[23];
+  bf1[23] = bf0[22] + bf0[23];
+  bf1[24] = bf0[24] + bf0[25];
+  bf1[25] = bf0[24] - bf0[25];
+  bf1[26] = -bf0[26] + bf0[27];
+  bf1[27] = bf0[26] + bf0[27];
+  bf1[28] = bf0[28] + bf0[29];
+  bf1[29] = bf0[28] - bf0[29];
+  bf1[30] = -bf0[30] + bf0[31];
+  bf1[31] = bf0[30] + bf0[31];
+  bf1[32] = bf0[32];
+  bf1[33] = half_btf(-cospi[4], bf0[33], cospi[60], bf0[62], cos_bit[stage]);
+  bf1[34] = half_btf(-cospi[60], bf0[34], -cospi[4], bf0[61], cos_bit[stage]);
+  bf1[35] = bf0[35];
+  bf1[36] = bf0[36];
+  bf1[37] = half_btf(-cospi[36], bf0[37], cospi[28], bf0[58], cos_bit[stage]);
+  bf1[38] = half_btf(-cospi[28], bf0[38], -cospi[36], bf0[57], cos_bit[stage]);
+  bf1[39] = bf0[39];
+  bf1[40] = bf0[40];
+  bf1[41] = half_btf(-cospi[20], bf0[41], cospi[44], bf0[54], cos_bit[stage]);
+  bf1[42] = half_btf(-cospi[44], bf0[42], -cospi[20], bf0[53], cos_bit[stage]);
+  bf1[43] = bf0[43];
+  bf1[44] = bf0[44];
+  bf1[45] = half_btf(-cospi[52], bf0[45], cospi[12], bf0[50], cos_bit[stage]);
+  bf1[46] = half_btf(-cospi[12], bf0[46], -cospi[52], bf0[49], cos_bit[stage]);
+  bf1[47] = bf0[47];
+  bf1[48] = bf0[48];
+  bf1[49] = half_btf(-cospi[52], bf0[46], cospi[12], bf0[49], cos_bit[stage]);
+  bf1[50] = half_btf(cospi[12], bf0[45], cospi[52], bf0[50], cos_bit[stage]);
+  bf1[51] = bf0[51];
+  bf1[52] = bf0[52];
+  bf1[53] = half_btf(-cospi[20], bf0[42], cospi[44], bf0[53], cos_bit[stage]);
+  bf1[54] = half_btf(cospi[44], bf0[41], cospi[20], bf0[54], cos_bit[stage]);
+  bf1[55] = bf0[55];
+  bf1[56] = bf0[56];
+  bf1[57] = half_btf(-cospi[36], bf0[38], cospi[28], bf0[57], cos_bit[stage]);
+  bf1[58] = half_btf(cospi[28], bf0[37], cospi[36], bf0[58], cos_bit[stage]);
+  bf1[59] = bf0[59];
+  bf1[60] = bf0[60];
+  bf1[61] = half_btf(-cospi[4], bf0[34], cospi[60], bf0[61], cos_bit[stage]);
+  bf1[62] = half_btf(cospi[60], bf0[33], cospi[4], bf0[62], cos_bit[stage]);
+  bf1[63] = bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 5
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = half_btf(cospi[56], bf0[4], -cospi[8], bf0[7], cos_bit[stage]);
+  bf1[5] = half_btf(cospi[24], bf0[5], -cospi[40], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[40], bf0[5], cospi[24], bf0[6], cos_bit[stage]);
+  bf1[7] = half_btf(cospi[8], bf0[4], cospi[56], bf0[7], cos_bit[stage]);
+  bf1[8] = bf0[8] + bf0[9];
+  bf1[9] = bf0[8] - bf0[9];
+  bf1[10] = -bf0[10] + bf0[11];
+  bf1[11] = bf0[10] + bf0[11];
+  bf1[12] = bf0[12] + bf0[13];
+  bf1[13] = bf0[12] - bf0[13];
+  bf1[14] = -bf0[14] + bf0[15];
+  bf1[15] = bf0[14] + bf0[15];
+  bf1[16] = bf0[16];
+  bf1[17] = half_btf(-cospi[8], bf0[17], cospi[56], bf0[30], cos_bit[stage]);
+  bf1[18] = half_btf(-cospi[56], bf0[18], -cospi[8], bf0[29], cos_bit[stage]);
+  bf1[19] = bf0[19];
+  bf1[20] = bf0[20];
+  bf1[21] = half_btf(-cospi[40], bf0[21], cospi[24], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[24], bf0[22], -cospi[40], bf0[25], cos_bit[stage]);
+  bf1[23] = bf0[23];
+  bf1[24] = bf0[24];
+  bf1[25] = half_btf(-cospi[40], bf0[22], cospi[24], bf0[25], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[24], bf0[21], cospi[40], bf0[26], cos_bit[stage]);
+  bf1[27] = bf0[27];
+  bf1[28] = bf0[28];
+  bf1[29] = half_btf(-cospi[8], bf0[18], cospi[56], bf0[29], cos_bit[stage]);
+  bf1[30] = half_btf(cospi[56], bf0[17], cospi[8], bf0[30], cos_bit[stage]);
+  bf1[31] = bf0[31];
+  bf1[32] = bf0[32] + bf0[35];
+  bf1[33] = bf0[33] + bf0[34];
+  bf1[34] = bf0[33] - bf0[34];
+  bf1[35] = bf0[32] - bf0[35];
+  bf1[36] = -bf0[36] + bf0[39];
+  bf1[37] = -bf0[37] + bf0[38];
+  bf1[38] = bf0[37] + bf0[38];
+  bf1[39] = bf0[36] + bf0[39];
+  bf1[40] = bf0[40] + bf0[43];
+  bf1[41] = bf0[41] + bf0[42];
+  bf1[42] = bf0[41] - bf0[42];
+  bf1[43] = bf0[40] - bf0[43];
+  bf1[44] = -bf0[44] + bf0[47];
+  bf1[45] = -bf0[45] + bf0[46];
+  bf1[46] = bf0[45] + bf0[46];
+  bf1[47] = bf0[44] + bf0[47];
+  bf1[48] = bf0[48] + bf0[51];
+  bf1[49] = bf0[49] + bf0[50];
+  bf1[50] = bf0[49] - bf0[50];
+  bf1[51] = bf0[48] - bf0[51];
+  bf1[52] = -bf0[52] + bf0[55];
+  bf1[53] = -bf0[53] + bf0[54];
+  bf1[54] = bf0[53] + bf0[54];
+  bf1[55] = bf0[52] + bf0[55];
+  bf1[56] = bf0[56] + bf0[59];
+  bf1[57] = bf0[57] + bf0[58];
+  bf1[58] = bf0[57] - bf0[58];
+  bf1[59] = bf0[56] - bf0[59];
+  bf1[60] = -bf0[60] + bf0[63];
+  bf1[61] = -bf0[61] + bf0[62];
+  bf1[62] = bf0[61] + bf0[62];
+  bf1[63] = bf0[60] + bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 6
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = half_btf(cospi[32], bf0[0], cospi[32], bf0[1], cos_bit[stage]);
+  bf1[1] = half_btf(cospi[32], bf0[0], -cospi[32], bf0[1], cos_bit[stage]);
+  bf1[2] = half_btf(cospi[48], bf0[2], -cospi[16], bf0[3], cos_bit[stage]);
+  bf1[3] = half_btf(cospi[16], bf0[2], cospi[48], bf0[3], cos_bit[stage]);
+  bf1[4] = bf0[4] + bf0[5];
+  bf1[5] = bf0[4] - bf0[5];
+  bf1[6] = -bf0[6] + bf0[7];
+  bf1[7] = bf0[6] + bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = half_btf(-cospi[16], bf0[9], cospi[48], bf0[14], cos_bit[stage]);
+  bf1[10] = half_btf(-cospi[48], bf0[10], -cospi[16], bf0[13], cos_bit[stage]);
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] = half_btf(-cospi[16], bf0[10], cospi[48], bf0[13], cos_bit[stage]);
+  bf1[14] = half_btf(cospi[48], bf0[9], cospi[16], bf0[14], cos_bit[stage]);
+  bf1[15] = bf0[15];
+  bf1[16] = bf0[16] + bf0[19];
+  bf1[17] = bf0[17] + bf0[18];
+  bf1[18] = bf0[17] - bf0[18];
+  bf1[19] = bf0[16] - bf0[19];
+  bf1[20] = -bf0[20] + bf0[23];
+  bf1[21] = -bf0[21] + bf0[22];
+  bf1[22] = bf0[21] + bf0[22];
+  bf1[23] = bf0[20] + bf0[23];
+  bf1[24] = bf0[24] + bf0[27];
+  bf1[25] = bf0[25] + bf0[26];
+  bf1[26] = bf0[25] - bf0[26];
+  bf1[27] = bf0[24] - bf0[27];
+  bf1[28] = -bf0[28] + bf0[31];
+  bf1[29] = -bf0[29] + bf0[30];
+  bf1[30] = bf0[29] + bf0[30];
+  bf1[31] = bf0[28] + bf0[31];
+  bf1[32] = bf0[32];
+  bf1[33] = bf0[33];
+  bf1[34] = half_btf(-cospi[8], bf0[34], cospi[56], bf0[61], cos_bit[stage]);
+  bf1[35] = half_btf(-cospi[8], bf0[35], cospi[56], bf0[60], cos_bit[stage]);
+  bf1[36] = half_btf(-cospi[56], bf0[36], -cospi[8], bf0[59], cos_bit[stage]);
+  bf1[37] = half_btf(-cospi[56], bf0[37], -cospi[8], bf0[58], cos_bit[stage]);
+  bf1[38] = bf0[38];
+  bf1[39] = bf0[39];
+  bf1[40] = bf0[40];
+  bf1[41] = bf0[41];
+  bf1[42] = half_btf(-cospi[40], bf0[42], cospi[24], bf0[53], cos_bit[stage]);
+  bf1[43] = half_btf(-cospi[40], bf0[43], cospi[24], bf0[52], cos_bit[stage]);
+  bf1[44] = half_btf(-cospi[24], bf0[44], -cospi[40], bf0[51], cos_bit[stage]);
+  bf1[45] = half_btf(-cospi[24], bf0[45], -cospi[40], bf0[50], cos_bit[stage]);
+  bf1[46] = bf0[46];
+  bf1[47] = bf0[47];
+  bf1[48] = bf0[48];
+  bf1[49] = bf0[49];
+  bf1[50] = half_btf(-cospi[40], bf0[45], cospi[24], bf0[50], cos_bit[stage]);
+  bf1[51] = half_btf(-cospi[40], bf0[44], cospi[24], bf0[51], cos_bit[stage]);
+  bf1[52] = half_btf(cospi[24], bf0[43], cospi[40], bf0[52], cos_bit[stage]);
+  bf1[53] = half_btf(cospi[24], bf0[42], cospi[40], bf0[53], cos_bit[stage]);
+  bf1[54] = bf0[54];
+  bf1[55] = bf0[55];
+  bf1[56] = bf0[56];
+  bf1[57] = bf0[57];
+  bf1[58] = half_btf(-cospi[8], bf0[37], cospi[56], bf0[58], cos_bit[stage]);
+  bf1[59] = half_btf(-cospi[8], bf0[36], cospi[56], bf0[59], cos_bit[stage]);
+  bf1[60] = half_btf(cospi[56], bf0[35], cospi[8], bf0[60], cos_bit[stage]);
+  bf1[61] = half_btf(cospi[56], bf0[34], cospi[8], bf0[61], cos_bit[stage]);
+  bf1[62] = bf0[62];
+  bf1[63] = bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 7
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[3];
+  bf1[1] = bf0[1] + bf0[2];
+  bf1[2] = bf0[1] - bf0[2];
+  bf1[3] = bf0[0] - bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = half_btf(-cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[6] = half_btf(cospi[32], bf0[5], cospi[32], bf0[6], cos_bit[stage]);
+  bf1[7] = bf0[7];
+  bf1[8] = bf0[8] + bf0[11];
+  bf1[9] = bf0[9] + bf0[10];
+  bf1[10] = bf0[9] - bf0[10];
+  bf1[11] = bf0[8] - bf0[11];
+  bf1[12] = -bf0[12] + bf0[15];
+  bf1[13] = -bf0[13] + bf0[14];
+  bf1[14] = bf0[13] + bf0[14];
+  bf1[15] = bf0[12] + bf0[15];
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = half_btf(-cospi[16], bf0[18], cospi[48], bf0[29], cos_bit[stage]);
+  bf1[19] = half_btf(-cospi[16], bf0[19], cospi[48], bf0[28], cos_bit[stage]);
+  bf1[20] = half_btf(-cospi[48], bf0[20], -cospi[16], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[48], bf0[21], -cospi[16], bf0[26], cos_bit[stage]);
+  bf1[22] = bf0[22];
+  bf1[23] = bf0[23];
+  bf1[24] = bf0[24];
+  bf1[25] = bf0[25];
+  bf1[26] = half_btf(-cospi[16], bf0[21], cospi[48], bf0[26], cos_bit[stage]);
+  bf1[27] = half_btf(-cospi[16], bf0[20], cospi[48], bf0[27], cos_bit[stage]);
+  bf1[28] = half_btf(cospi[48], bf0[19], cospi[16], bf0[28], cos_bit[stage]);
+  bf1[29] = half_btf(cospi[48], bf0[18], cospi[16], bf0[29], cos_bit[stage]);
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[31];
+  bf1[32] = bf0[32] + bf0[39];
+  bf1[33] = bf0[33] + bf0[38];
+  bf1[34] = bf0[34] + bf0[37];
+  bf1[35] = bf0[35] + bf0[36];
+  bf1[36] = bf0[35] - bf0[36];
+  bf1[37] = bf0[34] - bf0[37];
+  bf1[38] = bf0[33] - bf0[38];
+  bf1[39] = bf0[32] - bf0[39];
+  bf1[40] = -bf0[40] + bf0[47];
+  bf1[41] = -bf0[41] + bf0[46];
+  bf1[42] = -bf0[42] + bf0[45];
+  bf1[43] = -bf0[43] + bf0[44];
+  bf1[44] = bf0[43] + bf0[44];
+  bf1[45] = bf0[42] + bf0[45];
+  bf1[46] = bf0[41] + bf0[46];
+  bf1[47] = bf0[40] + bf0[47];
+  bf1[48] = bf0[48] + bf0[55];
+  bf1[49] = bf0[49] + bf0[54];
+  bf1[50] = bf0[50] + bf0[53];
+  bf1[51] = bf0[51] + bf0[52];
+  bf1[52] = bf0[51] - bf0[52];
+  bf1[53] = bf0[50] - bf0[53];
+  bf1[54] = bf0[49] - bf0[54];
+  bf1[55] = bf0[48] - bf0[55];
+  bf1[56] = -bf0[56] + bf0[63];
+  bf1[57] = -bf0[57] + bf0[62];
+  bf1[58] = -bf0[58] + bf0[61];
+  bf1[59] = -bf0[59] + bf0[60];
+  bf1[60] = bf0[59] + bf0[60];
+  bf1[61] = bf0[58] + bf0[61];
+  bf1[62] = bf0[57] + bf0[62];
+  bf1[63] = bf0[56] + bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 8
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[7];
+  bf1[1] = bf0[1] + bf0[6];
+  bf1[2] = bf0[2] + bf0[5];
+  bf1[3] = bf0[3] + bf0[4];
+  bf1[4] = bf0[3] - bf0[4];
+  bf1[5] = bf0[2] - bf0[5];
+  bf1[6] = bf0[1] - bf0[6];
+  bf1[7] = bf0[0] - bf0[7];
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] = half_btf(-cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
+  bf1[11] = half_btf(-cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
+  bf1[12] = half_btf(cospi[32], bf0[11], cospi[32], bf0[12], cos_bit[stage]);
+  bf1[13] = half_btf(cospi[32], bf0[10], cospi[32], bf0[13], cos_bit[stage]);
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  bf1[16] = bf0[16] + bf0[23];
+  bf1[17] = bf0[17] + bf0[22];
+  bf1[18] = bf0[18] + bf0[21];
+  bf1[19] = bf0[19] + bf0[20];
+  bf1[20] = bf0[19] - bf0[20];
+  bf1[21] = bf0[18] - bf0[21];
+  bf1[22] = bf0[17] - bf0[22];
+  bf1[23] = bf0[16] - bf0[23];
+  bf1[24] = -bf0[24] + bf0[31];
+  bf1[25] = -bf0[25] + bf0[30];
+  bf1[26] = -bf0[26] + bf0[29];
+  bf1[27] = -bf0[27] + bf0[28];
+  bf1[28] = bf0[27] + bf0[28];
+  bf1[29] = bf0[26] + bf0[29];
+  bf1[30] = bf0[25] + bf0[30];
+  bf1[31] = bf0[24] + bf0[31];
+  bf1[32] = bf0[32];
+  bf1[33] = bf0[33];
+  bf1[34] = bf0[34];
+  bf1[35] = bf0[35];
+  bf1[36] = half_btf(-cospi[16], bf0[36], cospi[48], bf0[59], cos_bit[stage]);
+  bf1[37] = half_btf(-cospi[16], bf0[37], cospi[48], bf0[58], cos_bit[stage]);
+  bf1[38] = half_btf(-cospi[16], bf0[38], cospi[48], bf0[57], cos_bit[stage]);
+  bf1[39] = half_btf(-cospi[16], bf0[39], cospi[48], bf0[56], cos_bit[stage]);
+  bf1[40] = half_btf(-cospi[48], bf0[40], -cospi[16], bf0[55], cos_bit[stage]);
+  bf1[41] = half_btf(-cospi[48], bf0[41], -cospi[16], bf0[54], cos_bit[stage]);
+  bf1[42] = half_btf(-cospi[48], bf0[42], -cospi[16], bf0[53], cos_bit[stage]);
+  bf1[43] = half_btf(-cospi[48], bf0[43], -cospi[16], bf0[52], cos_bit[stage]);
+  bf1[44] = bf0[44];
+  bf1[45] = bf0[45];
+  bf1[46] = bf0[46];
+  bf1[47] = bf0[47];
+  bf1[48] = bf0[48];
+  bf1[49] = bf0[49];
+  bf1[50] = bf0[50];
+  bf1[51] = bf0[51];
+  bf1[52] = half_btf(-cospi[16], bf0[43], cospi[48], bf0[52], cos_bit[stage]);
+  bf1[53] = half_btf(-cospi[16], bf0[42], cospi[48], bf0[53], cos_bit[stage]);
+  bf1[54] = half_btf(-cospi[16], bf0[41], cospi[48], bf0[54], cos_bit[stage]);
+  bf1[55] = half_btf(-cospi[16], bf0[40], cospi[48], bf0[55], cos_bit[stage]);
+  bf1[56] = half_btf(cospi[48], bf0[39], cospi[16], bf0[56], cos_bit[stage]);
+  bf1[57] = half_btf(cospi[48], bf0[38], cospi[16], bf0[57], cos_bit[stage]);
+  bf1[58] = half_btf(cospi[48], bf0[37], cospi[16], bf0[58], cos_bit[stage]);
+  bf1[59] = half_btf(cospi[48], bf0[36], cospi[16], bf0[59], cos_bit[stage]);
+  bf1[60] = bf0[60];
+  bf1[61] = bf0[61];
+  bf1[62] = bf0[62];
+  bf1[63] = bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 9
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[15];
+  bf1[1] = bf0[1] + bf0[14];
+  bf1[2] = bf0[2] + bf0[13];
+  bf1[3] = bf0[3] + bf0[12];
+  bf1[4] = bf0[4] + bf0[11];
+  bf1[5] = bf0[5] + bf0[10];
+  bf1[6] = bf0[6] + bf0[9];
+  bf1[7] = bf0[7] + bf0[8];
+  bf1[8] = bf0[7] - bf0[8];
+  bf1[9] = bf0[6] - bf0[9];
+  bf1[10] = bf0[5] - bf0[10];
+  bf1[11] = bf0[4] - bf0[11];
+  bf1[12] = bf0[3] - bf0[12];
+  bf1[13] = bf0[2] - bf0[13];
+  bf1[14] = bf0[1] - bf0[14];
+  bf1[15] = bf0[0] - bf0[15];
+  bf1[16] = bf0[16];
+  bf1[17] = bf0[17];
+  bf1[18] = bf0[18];
+  bf1[19] = bf0[19];
+  bf1[20] = half_btf(-cospi[32], bf0[20], cospi[32], bf0[27], cos_bit[stage]);
+  bf1[21] = half_btf(-cospi[32], bf0[21], cospi[32], bf0[26], cos_bit[stage]);
+  bf1[22] = half_btf(-cospi[32], bf0[22], cospi[32], bf0[25], cos_bit[stage]);
+  bf1[23] = half_btf(-cospi[32], bf0[23], cospi[32], bf0[24], cos_bit[stage]);
+  bf1[24] = half_btf(cospi[32], bf0[23], cospi[32], bf0[24], cos_bit[stage]);
+  bf1[25] = half_btf(cospi[32], bf0[22], cospi[32], bf0[25], cos_bit[stage]);
+  bf1[26] = half_btf(cospi[32], bf0[21], cospi[32], bf0[26], cos_bit[stage]);
+  bf1[27] = half_btf(cospi[32], bf0[20], cospi[32], bf0[27], cos_bit[stage]);
+  bf1[28] = bf0[28];
+  bf1[29] = bf0[29];
+  bf1[30] = bf0[30];
+  bf1[31] = bf0[31];
+  bf1[32] = bf0[32] + bf0[47];
+  bf1[33] = bf0[33] + bf0[46];
+  bf1[34] = bf0[34] + bf0[45];
+  bf1[35] = bf0[35] + bf0[44];
+  bf1[36] = bf0[36] + bf0[43];
+  bf1[37] = bf0[37] + bf0[42];
+  bf1[38] = bf0[38] + bf0[41];
+  bf1[39] = bf0[39] + bf0[40];
+  bf1[40] = bf0[39] - bf0[40];
+  bf1[41] = bf0[38] - bf0[41];
+  bf1[42] = bf0[37] - bf0[42];
+  bf1[43] = bf0[36] - bf0[43];
+  bf1[44] = bf0[35] - bf0[44];
+  bf1[45] = bf0[34] - bf0[45];
+  bf1[46] = bf0[33] - bf0[46];
+  bf1[47] = bf0[32] - bf0[47];
+  bf1[48] = -bf0[48] + bf0[63];
+  bf1[49] = -bf0[49] + bf0[62];
+  bf1[50] = -bf0[50] + bf0[61];
+  bf1[51] = -bf0[51] + bf0[60];
+  bf1[52] = -bf0[52] + bf0[59];
+  bf1[53] = -bf0[53] + bf0[58];
+  bf1[54] = -bf0[54] + bf0[57];
+  bf1[55] = -bf0[55] + bf0[56];
+  bf1[56] = bf0[55] + bf0[56];
+  bf1[57] = bf0[54] + bf0[57];
+  bf1[58] = bf0[53] + bf0[58];
+  bf1[59] = bf0[52] + bf0[59];
+  bf1[60] = bf0[51] + bf0[60];
+  bf1[61] = bf0[50] + bf0[61];
+  bf1[62] = bf0[49] + bf0[62];
+  bf1[63] = bf0[48] + bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 10
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = output;
+  bf1 = step;
+  bf1[0] = bf0[0] + bf0[31];
+  bf1[1] = bf0[1] + bf0[30];
+  bf1[2] = bf0[2] + bf0[29];
+  bf1[3] = bf0[3] + bf0[28];
+  bf1[4] = bf0[4] + bf0[27];
+  bf1[5] = bf0[5] + bf0[26];
+  bf1[6] = bf0[6] + bf0[25];
+  bf1[7] = bf0[7] + bf0[24];
+  bf1[8] = bf0[8] + bf0[23];
+  bf1[9] = bf0[9] + bf0[22];
+  bf1[10] = bf0[10] + bf0[21];
+  bf1[11] = bf0[11] + bf0[20];
+  bf1[12] = bf0[12] + bf0[19];
+  bf1[13] = bf0[13] + bf0[18];
+  bf1[14] = bf0[14] + bf0[17];
+  bf1[15] = bf0[15] + bf0[16];
+  bf1[16] = bf0[15] - bf0[16];
+  bf1[17] = bf0[14] - bf0[17];
+  bf1[18] = bf0[13] - bf0[18];
+  bf1[19] = bf0[12] - bf0[19];
+  bf1[20] = bf0[11] - bf0[20];
+  bf1[21] = bf0[10] - bf0[21];
+  bf1[22] = bf0[9] - bf0[22];
+  bf1[23] = bf0[8] - bf0[23];
+  bf1[24] = bf0[7] - bf0[24];
+  bf1[25] = bf0[6] - bf0[25];
+  bf1[26] = bf0[5] - bf0[26];
+  bf1[27] = bf0[4] - bf0[27];
+  bf1[28] = bf0[3] - bf0[28];
+  bf1[29] = bf0[2] - bf0[29];
+  bf1[30] = bf0[1] - bf0[30];
+  bf1[31] = bf0[0] - bf0[31];
+  bf1[32] = bf0[32];
+  bf1[33] = bf0[33];
+  bf1[34] = bf0[34];
+  bf1[35] = bf0[35];
+  bf1[36] = bf0[36];
+  bf1[37] = bf0[37];
+  bf1[38] = bf0[38];
+  bf1[39] = bf0[39];
+  bf1[40] = half_btf(-cospi[32], bf0[40], cospi[32], bf0[55], cos_bit[stage]);
+  bf1[41] = half_btf(-cospi[32], bf0[41], cospi[32], bf0[54], cos_bit[stage]);
+  bf1[42] = half_btf(-cospi[32], bf0[42], cospi[32], bf0[53], cos_bit[stage]);
+  bf1[43] = half_btf(-cospi[32], bf0[43], cospi[32], bf0[52], cos_bit[stage]);
+  bf1[44] = half_btf(-cospi[32], bf0[44], cospi[32], bf0[51], cos_bit[stage]);
+  bf1[45] = half_btf(-cospi[32], bf0[45], cospi[32], bf0[50], cos_bit[stage]);
+  bf1[46] = half_btf(-cospi[32], bf0[46], cospi[32], bf0[49], cos_bit[stage]);
+  bf1[47] = half_btf(-cospi[32], bf0[47], cospi[32], bf0[48], cos_bit[stage]);
+  bf1[48] = half_btf(cospi[32], bf0[47], cospi[32], bf0[48], cos_bit[stage]);
+  bf1[49] = half_btf(cospi[32], bf0[46], cospi[32], bf0[49], cos_bit[stage]);
+  bf1[50] = half_btf(cospi[32], bf0[45], cospi[32], bf0[50], cos_bit[stage]);
+  bf1[51] = half_btf(cospi[32], bf0[44], cospi[32], bf0[51], cos_bit[stage]);
+  bf1[52] = half_btf(cospi[32], bf0[43], cospi[32], bf0[52], cos_bit[stage]);
+  bf1[53] = half_btf(cospi[32], bf0[42], cospi[32], bf0[53], cos_bit[stage]);
+  bf1[54] = half_btf(cospi[32], bf0[41], cospi[32], bf0[54], cos_bit[stage]);
+  bf1[55] = half_btf(cospi[32], bf0[40], cospi[32], bf0[55], cos_bit[stage]);
+  bf1[56] = bf0[56];
+  bf1[57] = bf0[57];
+  bf1[58] = bf0[58];
+  bf1[59] = bf0[59];
+  bf1[60] = bf0[60];
+  bf1[61] = bf0[61];
+  bf1[62] = bf0[62];
+  bf1[63] = bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+
+  // stage 11
+  stage++;
+  cospi = cospi_arr[cos_bit[stage] - cos_bit_min];
+  bf0 = step;
+  bf1 = output;
+  bf1[0] = bf0[0] + bf0[63];
+  bf1[1] = bf0[1] + bf0[62];
+  bf1[2] = bf0[2] + bf0[61];
+  bf1[3] = bf0[3] + bf0[60];
+  bf1[4] = bf0[4] + bf0[59];
+  bf1[5] = bf0[5] + bf0[58];
+  bf1[6] = bf0[6] + bf0[57];
+  bf1[7] = bf0[7] + bf0[56];
+  bf1[8] = bf0[8] + bf0[55];
+  bf1[9] = bf0[9] + bf0[54];
+  bf1[10] = bf0[10] + bf0[53];
+  bf1[11] = bf0[11] + bf0[52];
+  bf1[12] = bf0[12] + bf0[51];
+  bf1[13] = bf0[13] + bf0[50];
+  bf1[14] = bf0[14] + bf0[49];
+  bf1[15] = bf0[15] + bf0[48];
+  bf1[16] = bf0[16] + bf0[47];
+  bf1[17] = bf0[17] + bf0[46];
+  bf1[18] = bf0[18] + bf0[45];
+  bf1[19] = bf0[19] + bf0[44];
+  bf1[20] = bf0[20] + bf0[43];
+  bf1[21] = bf0[21] + bf0[42];
+  bf1[22] = bf0[22] + bf0[41];
+  bf1[23] = bf0[23] + bf0[40];
+  bf1[24] = bf0[24] + bf0[39];
+  bf1[25] = bf0[25] + bf0[38];
+  bf1[26] = bf0[26] + bf0[37];
+  bf1[27] = bf0[27] + bf0[36];
+  bf1[28] = bf0[28] + bf0[35];
+  bf1[29] = bf0[29] + bf0[34];
+  bf1[30] = bf0[30] + bf0[33];
+  bf1[31] = bf0[31] + bf0[32];
+  bf1[32] = bf0[31] - bf0[32];
+  bf1[33] = bf0[30] - bf0[33];
+  bf1[34] = bf0[29] - bf0[34];
+  bf1[35] = bf0[28] - bf0[35];
+  bf1[36] = bf0[27] - bf0[36];
+  bf1[37] = bf0[26] - bf0[37];
+  bf1[38] = bf0[25] - bf0[38];
+  bf1[39] = bf0[24] - bf0[39];
+  bf1[40] = bf0[23] - bf0[40];
+  bf1[41] = bf0[22] - bf0[41];
+  bf1[42] = bf0[21] - bf0[42];
+  bf1[43] = bf0[20] - bf0[43];
+  bf1[44] = bf0[19] - bf0[44];
+  bf1[45] = bf0[18] - bf0[45];
+  bf1[46] = bf0[17] - bf0[46];
+  bf1[47] = bf0[16] - bf0[47];
+  bf1[48] = bf0[15] - bf0[48];
+  bf1[49] = bf0[14] - bf0[49];
+  bf1[50] = bf0[13] - bf0[50];
+  bf1[51] = bf0[12] - bf0[51];
+  bf1[52] = bf0[11] - bf0[52];
+  bf1[53] = bf0[10] - bf0[53];
+  bf1[54] = bf0[9] - bf0[54];
+  bf1[55] = bf0[8] - bf0[55];
+  bf1[56] = bf0[7] - bf0[56];
+  bf1[57] = bf0[6] - bf0[57];
+  bf1[58] = bf0[5] - bf0[58];
+  bf1[59] = bf0[4] - bf0[59];
+  bf1[60] = bf0[3] - bf0[60];
+  bf1[61] = bf0[2] - bf0[61];
+  bf1[62] = bf0[1] - bf0[62];
+  bf1[63] = bf0[0] - bf0[63];
+  range_check(stage, input, bf1, size, stage_range[stage]);
+}
+#endif  // CONFIG_TX64X64
diff --git a/third_party/aom/av1/common/av1_inv_txfm1d.h b/third_party/aom/av1/common/av1_inv_txfm1d.h
new file mode 100644
index 0000000000..9e7a2323ba
--- /dev/null
+++ b/third_party/aom/av1/common/av1_inv_txfm1d.h
@@ -0,0 +1,45 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_INV_TXFM1D_H_
+#define AV1_INV_TXFM1D_H_
+
+#include "av1/common/av1_txfm.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_idct4_new(const int32_t *input, int32_t *output, const int8_t *cos_bit,
+                   const int8_t *stage_range);
+void av1_idct8_new(const int32_t *input, int32_t *output, const int8_t *cos_bit,
+                   const int8_t *stage_range);
+void av1_idct16_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
+void av1_idct32_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
+void av1_idct64_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
+
+void av1_iadst4_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
+void av1_iadst8_new(const int32_t *input, int32_t *output,
+                    const int8_t *cos_bit, const int8_t *stage_range);
+void av1_iadst16_new(const int32_t *input, int32_t *output,
+                     const int8_t *cos_bit, const int8_t *stage_range);
+void av1_iadst32_new(const int32_t *input, int32_t *output,
+                     const int8_t *cos_bit, const int8_t *stage_range);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // AV1_INV_TXFM1D_H_
diff --git a/third_party/aom/av1/common/av1_inv_txfm2d.c b/third_party/aom/av1/common/av1_inv_txfm2d.c
new file mode 100644
index 0000000000..d56c7d11f9
--- /dev/null
+++ b/third_party/aom/av1/common/av1_inv_txfm2d.c
@@ -0,0 +1,256 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "av1/common/enums.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/av1_inv_txfm1d.h"
+#include "av1/common/av1_inv_txfm2d_cfg.h"
+
+static INLINE TxfmFunc inv_txfm_type_to_func(TXFM_TYPE txfm_type) {
+  switch (txfm_type) {
+    case TXFM_TYPE_DCT4: return av1_idct4_new;
+    case TXFM_TYPE_DCT8: return av1_idct8_new;
+    case TXFM_TYPE_DCT16: return av1_idct16_new;
+    case TXFM_TYPE_DCT32: return av1_idct32_new;
+    case TXFM_TYPE_ADST4: return av1_iadst4_new;
+    case TXFM_TYPE_ADST8: return av1_iadst8_new;
+    case TXFM_TYPE_ADST16: return av1_iadst16_new;
+    case TXFM_TYPE_ADST32: return av1_iadst32_new;
+    default: assert(0); return NULL;
+  }
+}
+
+const TXFM_2D_CFG *inv_txfm_cfg_ls[TX_TYPES][TX_SIZES] = {
+  // DCT_DCT
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &inv_txfm_2d_cfg_dct_dct_4, &inv_txfm_2d_cfg_dct_dct_8,
+      &inv_txfm_2d_cfg_dct_dct_16, &inv_txfm_2d_cfg_dct_dct_32 },
+  // ADST_DCT
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &inv_txfm_2d_cfg_adst_dct_4, &inv_txfm_2d_cfg_adst_dct_8,
+      &inv_txfm_2d_cfg_adst_dct_16, &inv_txfm_2d_cfg_adst_dct_32 },
+  // DCT_ADST
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &inv_txfm_2d_cfg_dct_adst_4, &inv_txfm_2d_cfg_dct_adst_8,
+      &inv_txfm_2d_cfg_dct_adst_16, &inv_txfm_2d_cfg_dct_adst_32 },
+  // ADST_ADST
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &inv_txfm_2d_cfg_adst_adst_4, &inv_txfm_2d_cfg_adst_adst_8,
+      &inv_txfm_2d_cfg_adst_adst_16, &inv_txfm_2d_cfg_adst_adst_32 },
+#if CONFIG_EXT_TX
+  // FLIPADST_DCT
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &inv_txfm_2d_cfg_adst_dct_4, &inv_txfm_2d_cfg_adst_dct_8,
+      &inv_txfm_2d_cfg_adst_dct_16, &inv_txfm_2d_cfg_adst_dct_32 },
+  // DCT_FLIPADST
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &inv_txfm_2d_cfg_dct_adst_4, &inv_txfm_2d_cfg_dct_adst_8,
+      &inv_txfm_2d_cfg_dct_adst_16, &inv_txfm_2d_cfg_dct_adst_32 },
+  // FLIPADST_FLIPADST
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &inv_txfm_2d_cfg_adst_adst_4, &inv_txfm_2d_cfg_adst_adst_8,
+      &inv_txfm_2d_cfg_adst_adst_16, &inv_txfm_2d_cfg_adst_adst_32 },
+  // ADST_FLIPADST
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &inv_txfm_2d_cfg_adst_adst_4, &inv_txfm_2d_cfg_adst_adst_8,
+      &inv_txfm_2d_cfg_adst_adst_16, &inv_txfm_2d_cfg_adst_adst_32 },
+  // FLIPADST_ADST
+  {
+#if CONFIG_CB4X4
+      NULL,
+#endif
+      &inv_txfm_2d_cfg_adst_adst_4, &inv_txfm_2d_cfg_adst_adst_8,
+      &inv_txfm_2d_cfg_adst_adst_16, &inv_txfm_2d_cfg_adst_adst_32 },
+  { // IDTX
+#if CONFIG_CB4X4
+    NULL,
+#endif
+    &inv_txfm_2d_cfg_adst_adst_4, &inv_txfm_2d_cfg_adst_adst_8,
+    &inv_txfm_2d_cfg_adst_adst_16, &inv_txfm_2d_cfg_adst_adst_32 },
+  { // V_DCT
+#if CONFIG_CB4X4
+    NULL,
+#endif
+    &inv_txfm_2d_cfg_dct_adst_4, &inv_txfm_2d_cfg_dct_adst_8,
+    &inv_txfm_2d_cfg_dct_adst_16, &inv_txfm_2d_cfg_dct_adst_32 },
+  { // H_DCT
+#if CONFIG_CB4X4
+    NULL,
+#endif
+    &inv_txfm_2d_cfg_adst_dct_4, &inv_txfm_2d_cfg_adst_dct_8,
+    &inv_txfm_2d_cfg_adst_dct_16, &inv_txfm_2d_cfg_adst_dct_32 },
+  { // V_ADST
+#if CONFIG_CB4X4
+    NULL,
+#endif
+    &inv_txfm_2d_cfg_adst_adst_4, &inv_txfm_2d_cfg_adst_adst_8,
+    &inv_txfm_2d_cfg_adst_adst_16, &inv_txfm_2d_cfg_adst_adst_32 },
+  { // H_ADST
+#if CONFIG_CB4X4
+    NULL,
+#endif
+    &inv_txfm_2d_cfg_adst_adst_4, &inv_txfm_2d_cfg_adst_adst_8,
+    &inv_txfm_2d_cfg_adst_adst_16, &inv_txfm_2d_cfg_adst_adst_32 },
+  { // V_FLIP_ADST
+#if CONFIG_CB4X4
+    NULL,
+#endif
+    &inv_txfm_2d_cfg_adst_adst_4, &inv_txfm_2d_cfg_adst_adst_8,
+    &inv_txfm_2d_cfg_adst_adst_16, &inv_txfm_2d_cfg_adst_adst_32 },
+  { // H_FLIP_ADST
+#if CONFIG_CB4X4
+    NULL,
+#endif
+    &inv_txfm_2d_cfg_adst_adst_4, &inv_txfm_2d_cfg_adst_adst_8,
+    &inv_txfm_2d_cfg_adst_adst_16, &inv_txfm_2d_cfg_adst_adst_32 },
+#endif  // CONFIG_EXT_TX
+};
+
+TXFM_2D_FLIP_CFG av1_get_inv_txfm_cfg(int tx_type, int tx_size) {
+  TXFM_2D_FLIP_CFG cfg;
+  set_flip_cfg(tx_type, &cfg);
+  cfg.cfg = inv_txfm_cfg_ls[tx_type][tx_size];
+  return cfg;
+}
+
+TXFM_2D_FLIP_CFG av1_get_inv_txfm_64x64_cfg(int tx_type) {
+  TXFM_2D_FLIP_CFG cfg = { 0, 0, NULL };
+  switch (tx_type) {
+    case DCT_DCT:
+      cfg.cfg = &inv_txfm_2d_cfg_dct_dct_64;
+      set_flip_cfg(tx_type, &cfg);
+      break;
+    default: assert(0);
+  }
+  return cfg;
+}
+
+static INLINE void inv_txfm2d_add_c(const int32_t *input, int16_t *output,
+                                    int stride, TXFM_2D_FLIP_CFG *cfg,
+                                    int32_t *txfm_buf) {
+  const int txfm_size = cfg->cfg->txfm_size;
+  const int8_t *shift = cfg->cfg->shift;
+  const int8_t *stage_range_col = cfg->cfg->stage_range_col;
+  const int8_t *stage_range_row = cfg->cfg->stage_range_row;
+  const int8_t *cos_bit_col = cfg->cfg->cos_bit_col;
+  const int8_t *cos_bit_row = cfg->cfg->cos_bit_row;
+  const TxfmFunc txfm_func_col = inv_txfm_type_to_func(cfg->cfg->txfm_type_col);
+  const TxfmFunc txfm_func_row = inv_txfm_type_to_func(cfg->cfg->txfm_type_row);
+
+  // txfm_buf's length is  txfm_size * txfm_size + 2 * txfm_size
+  // it is used for intermediate data buffering
+  int32_t *temp_in = txfm_buf;
+  int32_t *temp_out = temp_in + txfm_size;
+  int32_t *buf = temp_out + txfm_size;
+  int32_t *buf_ptr = buf;
+  int c, r;
+
+  // Rows
+  for (r = 0; r < txfm_size; ++r) {
+    txfm_func_row(input, buf_ptr, cos_bit_row, stage_range_row);
+    round_shift_array(buf_ptr, txfm_size, -shift[0]);
+    input += txfm_size;
+    buf_ptr += txfm_size;
+  }
+
+  // Columns
+  for (c = 0; c < txfm_size; ++c) {
+    if (cfg->lr_flip == 0) {
+      for (r = 0; r < txfm_size; ++r) temp_in[r] = buf[r * txfm_size + c];
+    } else {
+      // flip left right
+      for (r = 0; r < txfm_size; ++r)
+        temp_in[r] = buf[r * txfm_size + (txfm_size - c - 1)];
+    }
+    txfm_func_col(temp_in, temp_out, cos_bit_col, stage_range_col);
+    round_shift_array(temp_out, txfm_size, -shift[1]);
+    if (cfg->ud_flip == 0) {
+      for (r = 0; r < txfm_size; ++r) output[r * stride + c] += temp_out[r];
+    } else {
+      // flip upside down
+      for (r = 0; r < txfm_size; ++r)
+        output[r * stride + c] += temp_out[txfm_size - r - 1];
+    }
+  }
+}
+
+static INLINE void inv_txfm2d_add_facade(const int32_t *input, uint16_t *output,
+                                         int stride, int32_t *txfm_buf,
+                                         int tx_type, int tx_size, int bd) {
+  // output contains the prediction signal which is always positive and smaller
+  // than (1 << bd) - 1
+  // since bd < 16-1, therefore we can treat the uint16_t* output buffer as an
+  // int16_t*
+  TXFM_2D_FLIP_CFG cfg = av1_get_inv_txfm_cfg(tx_type, tx_size);
+  inv_txfm2d_add_c(input, (int16_t *)output, stride, &cfg, txfm_buf);
+  clamp_block((int16_t *)output, cfg.cfg->txfm_size, stride, 0, (1 << bd) - 1);
+}
+
+void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output,
+                              int stride, int tx_type, int bd) {
+  int txfm_buf[4 * 4 + 4 + 4];
+  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_4X4, bd);
+}
+
+void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output,
+                              int stride, int tx_type, int bd) {
+  int txfm_buf[8 * 8 + 8 + 8];
+  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_8X8, bd);
+}
+
+void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output,
+                                int stride, int tx_type, int bd) {
+  int txfm_buf[16 * 16 + 16 + 16];
+  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_16X16, bd);
+}
+
+void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output,
+                                int stride, int tx_type, int bd) {
+  int txfm_buf[32 * 32 + 32 + 32];
+  inv_txfm2d_add_facade(input, output, stride, txfm_buf, tx_type, TX_32X32, bd);
+}
+
+void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output,
+                                int stride, int tx_type, int bd) {
+  int txfm_buf[64 * 64 + 64 + 64];
+  // output contains the prediction signal which is always positive and smaller
+  // than (1 << bd) - 1
+  // since bd < 16-1, therefore we can treat the uint16_t* output buffer as an
+  // int16_t*
+  TXFM_2D_FLIP_CFG cfg = av1_get_inv_txfm_64x64_cfg(tx_type);
+  inv_txfm2d_add_c(input, (int16_t *)output, stride, &cfg, txfm_buf);
+  clamp_block((int16_t *)output, 64, stride, 0, (1 << bd) - 1);
+}
diff --git a/third_party/aom/av1/common/av1_inv_txfm2d_cfg.h b/third_party/aom/av1/common/av1_inv_txfm2d_cfg.h
new file mode 100644
index 0000000000..9eabc2e5a2
--- /dev/null
+++ b/third_party/aom/av1/common/av1_inv_txfm2d_cfg.h
@@ -0,0 +1,447 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_INV_TXFM2D_CFG_H_
+#define AV1_INV_TXFM2D_CFG_H_
+#include "av1/common/av1_inv_txfm1d.h"
+//  ---------------- config inv_dct_dct_4 ----------------
+static const int8_t inv_shift_dct_dct_4[2] = { 0, -4 };
+static const int8_t inv_stage_range_col_dct_dct_4[4] = { 18, 18, 17, 17 };
+static const int8_t inv_stage_range_row_dct_dct_4[4] = { 18, 18, 18, 18 };
+static const int8_t inv_cos_bit_col_dct_dct_4[4] = { 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_dct_dct_4[4] = { 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_dct_dct_4 = {
+  4,  // .txfm_size
+  4,  // .stage_num_col
+  4,  // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_dct_dct_4,            // .shift
+  inv_stage_range_col_dct_dct_4,  // .stage_range_col
+  inv_stage_range_row_dct_dct_4,  // .stage_range_row
+  inv_cos_bit_col_dct_dct_4,      // .cos_bit_col
+  inv_cos_bit_row_dct_dct_4,      // .cos_bit_row
+  TXFM_TYPE_DCT4,                 // .txfm_type_col
+  TXFM_TYPE_DCT4
+};  // .txfm_type_row
+
+//  ---------------- config inv_dct_dct_8 ----------------
+static const int8_t inv_shift_dct_dct_8[2] = { 0, -5 };
+static const int8_t inv_stage_range_col_dct_dct_8[6] = {
+  19, 19, 19, 19, 18, 18
+};
+static const int8_t inv_stage_range_row_dct_dct_8[6] = {
+  19, 19, 19, 19, 19, 19
+};
+static const int8_t inv_cos_bit_col_dct_dct_8[6] = { 13, 13, 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_dct_dct_8[6] = { 13, 13, 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_dct_dct_8 = {
+  8,  // .txfm_size
+  6,  // .stage_num_col
+  6,  // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_dct_dct_8,            // .shift
+  inv_stage_range_col_dct_dct_8,  // .stage_range_col
+  inv_stage_range_row_dct_dct_8,  // .stage_range_row
+  inv_cos_bit_col_dct_dct_8,      // .cos_bit_col
+  inv_cos_bit_row_dct_dct_8,      // .cos_bit_row
+  TXFM_TYPE_DCT8,                 // .txfm_type_col
+  TXFM_TYPE_DCT8
+};  // .txfm_type_row
+
+//  ---------------- config inv_dct_dct_16 ----------------
+static const int8_t inv_shift_dct_dct_16[2] = { -1, -5 };
+static const int8_t inv_stage_range_col_dct_dct_16[8] = { 19, 19, 19, 19,
+                                                          19, 19, 18, 18 };
+static const int8_t inv_stage_range_row_dct_dct_16[8] = { 20, 20, 20, 20,
+                                                          20, 20, 20, 20 };
+static const int8_t inv_cos_bit_col_dct_dct_16[8] = { 13, 13, 13, 13,
+                                                      13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_dct_dct_16[8] = { 12, 12, 12, 12,
+                                                      12, 12, 12, 12 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_dct_dct_16 = {
+  16,  // .txfm_size
+  8,   // .stage_num_col
+  8,   // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_dct_dct_16,            // .shift
+  inv_stage_range_col_dct_dct_16,  // .stage_range_col
+  inv_stage_range_row_dct_dct_16,  // .stage_range_row
+  inv_cos_bit_col_dct_dct_16,      // .cos_bit_col
+  inv_cos_bit_row_dct_dct_16,      // .cos_bit_row
+  TXFM_TYPE_DCT16,                 // .txfm_type_col
+  TXFM_TYPE_DCT16
+};  // .txfm_type_row
+
+//  ---------------- config inv_dct_dct_32 ----------------
+static const int8_t inv_shift_dct_dct_32[2] = { -1, -5 };
+static const int8_t inv_stage_range_col_dct_dct_32[10] = { 19, 19, 19, 19, 19,
+                                                           19, 19, 19, 18, 18 };
+static const int8_t inv_stage_range_row_dct_dct_32[10] = { 20, 20, 20, 20, 20,
+                                                           20, 20, 20, 20, 20 };
+static const int8_t inv_cos_bit_col_dct_dct_32[10] = { 13, 13, 13, 13, 13,
+                                                       13, 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_dct_dct_32[10] = { 12, 12, 12, 12, 12,
+                                                       12, 12, 12, 12, 12 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_dct_dct_32 = {
+  32,  // .txfm_size
+  10,  // .stage_num_col
+  10,  // .stage_num_row
+  // 1,  // .log_scale
+  inv_shift_dct_dct_32,            // .shift
+  inv_stage_range_col_dct_dct_32,  // .stage_range_col
+  inv_stage_range_row_dct_dct_32,  // .stage_range_row
+  inv_cos_bit_col_dct_dct_32,      // .cos_bit_col
+  inv_cos_bit_row_dct_dct_32,      // .cos_bit_row
+  TXFM_TYPE_DCT32,                 // .txfm_type_col
+  TXFM_TYPE_DCT32
+};  // .txfm_type_row
+
+//  ---------------- config inv_dct_dct_64 ----------------
+static const int8_t inv_shift_dct_dct_64[2] = { -1, -7 };
+static const int8_t inv_stage_range_col_dct_dct_64[12] = {
+  19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 18, 18
+};
+static const int8_t inv_stage_range_row_dct_dct_64[12] = {
+  20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20
+};
+static const int8_t inv_cos_bit_col_dct_dct_64[12] = { 13, 13, 13, 13, 13, 13,
+                                                       13, 13, 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_dct_dct_64[12] = { 12, 12, 12, 12, 12, 12,
+                                                       12, 12, 12, 12, 12, 12 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_dct_dct_64 = {
+  64,                              // .txfm_size
+  12,                              // .stage_num_col
+  12,                              // .stage_num_row
+  inv_shift_dct_dct_64,            // .shift
+  inv_stage_range_col_dct_dct_64,  // .stage_range_col
+  inv_stage_range_row_dct_dct_64,  // .stage_range_row
+  inv_cos_bit_col_dct_dct_64,      // .cos_bit_col
+  inv_cos_bit_row_dct_dct_64,      // .cos_bit_row
+  TXFM_TYPE_DCT64,                 // .txfm_type_col
+  TXFM_TYPE_DCT64
+};  // .txfm_type_row
+
+//  ---------------- config inv_dct_adst_4 ----------------
+static const int8_t inv_shift_dct_adst_4[2] = { 0, -4 };
+static const int8_t inv_stage_range_col_dct_adst_4[4] = { 18, 18, 17, 17 };
+static const int8_t inv_stage_range_row_dct_adst_4[6] = {
+  18, 18, 18, 18, 18, 18
+};
+static const int8_t inv_cos_bit_col_dct_adst_4[4] = { 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_dct_adst_4[6] = { 13, 13, 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_dct_adst_4 = {
+  4,  // .txfm_size
+  4,  // .stage_num_col
+  6,  // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_dct_adst_4,            // .shift
+  inv_stage_range_col_dct_adst_4,  // .stage_range_col
+  inv_stage_range_row_dct_adst_4,  // .stage_range_row
+  inv_cos_bit_col_dct_adst_4,      // .cos_bit_col
+  inv_cos_bit_row_dct_adst_4,      // .cos_bit_row
+  TXFM_TYPE_DCT4,                  // .txfm_type_col
+  TXFM_TYPE_ADST4
+};  // .txfm_type_row
+
+//  ---------------- config inv_dct_adst_8 ----------------
+static const int8_t inv_shift_dct_adst_8[2] = { 0, -5 };
+static const int8_t inv_stage_range_col_dct_adst_8[6] = {
+  19, 19, 19, 19, 18, 18
+};
+static const int8_t inv_stage_range_row_dct_adst_8[8] = { 19, 19, 19, 19,
+                                                          19, 19, 19, 19 };
+static const int8_t inv_cos_bit_col_dct_adst_8[6] = { 13, 13, 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_dct_adst_8[8] = { 13, 13, 13, 13,
+                                                      13, 13, 13, 13 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_dct_adst_8 = {
+  8,  // .txfm_size
+  6,  // .stage_num_col
+  8,  // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_dct_adst_8,            // .shift
+  inv_stage_range_col_dct_adst_8,  // .stage_range_col
+  inv_stage_range_row_dct_adst_8,  // .stage_range_row
+  inv_cos_bit_col_dct_adst_8,      // .cos_bit_col
+  inv_cos_bit_row_dct_adst_8,      // .cos_bit_row
+  TXFM_TYPE_DCT8,                  // .txfm_type_col
+  TXFM_TYPE_ADST8
+};  // .txfm_type_row
+
+//  ---------------- config inv_dct_adst_16 ----------------
+static const int8_t inv_shift_dct_adst_16[2] = { -1, -5 };
+static const int8_t inv_stage_range_col_dct_adst_16[8] = { 19, 19, 19, 19,
+                                                           19, 19, 18, 18 };
+static const int8_t inv_stage_range_row_dct_adst_16[10] = {
+  20, 20, 20, 20, 20, 20, 20, 20, 20, 20
+};
+static const int8_t inv_cos_bit_col_dct_adst_16[8] = { 13, 13, 13, 13,
+                                                       13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_dct_adst_16[10] = { 12, 12, 12, 12, 12,
+                                                        12, 12, 12, 12, 12 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_dct_adst_16 = {
+  16,  // .txfm_size
+  8,   // .stage_num_col
+  10,  // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_dct_adst_16,            // .shift
+  inv_stage_range_col_dct_adst_16,  // .stage_range_col
+  inv_stage_range_row_dct_adst_16,  // .stage_range_row
+  inv_cos_bit_col_dct_adst_16,      // .cos_bit_col
+  inv_cos_bit_row_dct_adst_16,      // .cos_bit_row
+  TXFM_TYPE_DCT16,                  // .txfm_type_col
+  TXFM_TYPE_ADST16
+};  // .txfm_type_row
+
+//  ---------------- config inv_dct_adst_32 ----------------
+static const int8_t inv_shift_dct_adst_32[2] = { -1, -5 };
+static const int8_t inv_stage_range_col_dct_adst_32[10] = {
+  19, 19, 19, 19, 19, 19, 19, 19, 18, 18
+};
+static const int8_t inv_stage_range_row_dct_adst_32[12] = {
+  20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20
+};
+static const int8_t inv_cos_bit_col_dct_adst_32[10] = { 13, 13, 13, 13, 13,
+                                                        13, 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_dct_adst_32[12] = {
+  12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12
+};
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_dct_adst_32 = {
+  32,  // .txfm_size
+  10,  // .stage_num_col
+  12,  // .stage_num_row
+  // 1,  // .log_scale
+  inv_shift_dct_adst_32,            // .shift
+  inv_stage_range_col_dct_adst_32,  // .stage_range_col
+  inv_stage_range_row_dct_adst_32,  // .stage_range_row
+  inv_cos_bit_col_dct_adst_32,      // .cos_bit_col
+  inv_cos_bit_row_dct_adst_32,      // .cos_bit_row
+  TXFM_TYPE_DCT32,                  // .txfm_type_col
+  TXFM_TYPE_ADST32
+};  // .txfm_type_row
+
+//  ---------------- config inv_adst_adst_4 ----------------
+static const int8_t inv_shift_adst_adst_4[2] = { 0, -4 };
+static const int8_t inv_stage_range_col_adst_adst_4[6] = { 18, 18, 18,
+                                                           18, 17, 17 };
+static const int8_t inv_stage_range_row_adst_adst_4[6] = { 18, 18, 18,
+                                                           18, 18, 18 };
+static const int8_t inv_cos_bit_col_adst_adst_4[6] = { 13, 13, 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_adst_adst_4[6] = { 13, 13, 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_adst_adst_4 = {
+  4,  // .txfm_size
+  6,  // .stage_num_col
+  6,  // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_adst_adst_4,            // .shift
+  inv_stage_range_col_adst_adst_4,  // .stage_range_col
+  inv_stage_range_row_adst_adst_4,  // .stage_range_row
+  inv_cos_bit_col_adst_adst_4,      // .cos_bit_col
+  inv_cos_bit_row_adst_adst_4,      // .cos_bit_row
+  TXFM_TYPE_ADST4,                  // .txfm_type_col
+  TXFM_TYPE_ADST4
+};  // .txfm_type_row
+
+//  ---------------- config inv_adst_adst_8 ----------------
+static const int8_t inv_shift_adst_adst_8[2] = { 0, -5 };
+static const int8_t inv_stage_range_col_adst_adst_8[8] = { 19, 19, 19, 19,
+                                                           19, 19, 18, 18 };
+static const int8_t inv_stage_range_row_adst_adst_8[8] = { 19, 19, 19, 19,
+                                                           19, 19, 19, 19 };
+static const int8_t inv_cos_bit_col_adst_adst_8[8] = { 13, 13, 13, 13,
+                                                       13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_adst_adst_8[8] = { 13, 13, 13, 13,
+                                                       13, 13, 13, 13 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_adst_adst_8 = {
+  8,  // .txfm_size
+  8,  // .stage_num_col
+  8,  // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_adst_adst_8,            // .shift
+  inv_stage_range_col_adst_adst_8,  // .stage_range_col
+  inv_stage_range_row_adst_adst_8,  // .stage_range_row
+  inv_cos_bit_col_adst_adst_8,      // .cos_bit_col
+  inv_cos_bit_row_adst_adst_8,      // .cos_bit_row
+  TXFM_TYPE_ADST8,                  // .txfm_type_col
+  TXFM_TYPE_ADST8
+};  // .txfm_type_row
+
+//  ---------------- config inv_adst_adst_16 ----------------
+static const int8_t inv_shift_adst_adst_16[2] = { -1, -5 };
+static const int8_t inv_stage_range_col_adst_adst_16[10] = {
+  19, 19, 19, 19, 19, 19, 19, 19, 18, 18
+};
+static const int8_t inv_stage_range_row_adst_adst_16[10] = {
+  20, 20, 20, 20, 20, 20, 20, 20, 20, 20
+};
+static const int8_t inv_cos_bit_col_adst_adst_16[10] = { 13, 13, 13, 13, 13,
+                                                         13, 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_adst_adst_16[10] = { 12, 12, 12, 12, 12,
+                                                         12, 12, 12, 12, 12 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_adst_adst_16 = {
+  16,  // .txfm_size
+  10,  // .stage_num_col
+  10,  // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_adst_adst_16,            // .shift
+  inv_stage_range_col_adst_adst_16,  // .stage_range_col
+  inv_stage_range_row_adst_adst_16,  // .stage_range_row
+  inv_cos_bit_col_adst_adst_16,      // .cos_bit_col
+  inv_cos_bit_row_adst_adst_16,      // .cos_bit_row
+  TXFM_TYPE_ADST16,                  // .txfm_type_col
+  TXFM_TYPE_ADST16
+};  // .txfm_type_row
+
+//  ---------------- config inv_adst_adst_32 ----------------
+static const int8_t inv_shift_adst_adst_32[2] = { -1, -5 };
+static const int8_t inv_stage_range_col_adst_adst_32[12] = {
+  19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 18, 18
+};
+static const int8_t inv_stage_range_row_adst_adst_32[12] = {
+  20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20, 20
+};
+static const int8_t inv_cos_bit_col_adst_adst_32[12] = {
+  13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13
+};
+static const int8_t inv_cos_bit_row_adst_adst_32[12] = {
+  12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12
+};
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_adst_adst_32 = {
+  32,  // .txfm_size
+  12,  // .stage_num_col
+  12,  // .stage_num_row
+  // 1,  // .log_scale
+  inv_shift_adst_adst_32,            // .shift
+  inv_stage_range_col_adst_adst_32,  // .stage_range_col
+  inv_stage_range_row_adst_adst_32,  // .stage_range_row
+  inv_cos_bit_col_adst_adst_32,      // .cos_bit_col
+  inv_cos_bit_row_adst_adst_32,      // .cos_bit_row
+  TXFM_TYPE_ADST32,                  // .txfm_type_col
+  TXFM_TYPE_ADST32
+};  // .txfm_type_row
+
+//  ---------------- config inv_adst_dct_4 ----------------
+static const int8_t inv_shift_adst_dct_4[2] = { 0, -4 };
+static const int8_t inv_stage_range_col_adst_dct_4[6] = {
+  18, 18, 18, 18, 17, 17
+};
+static const int8_t inv_stage_range_row_adst_dct_4[4] = { 18, 18, 18, 18 };
+static const int8_t inv_cos_bit_col_adst_dct_4[6] = { 13, 13, 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_adst_dct_4[4] = { 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_adst_dct_4 = {
+  4,  // .txfm_size
+  6,  // .stage_num_col
+  4,  // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_adst_dct_4,            // .shift
+  inv_stage_range_col_adst_dct_4,  // .stage_range_col
+  inv_stage_range_row_adst_dct_4,  // .stage_range_row
+  inv_cos_bit_col_adst_dct_4,      // .cos_bit_col
+  inv_cos_bit_row_adst_dct_4,      // .cos_bit_row
+  TXFM_TYPE_ADST4,                 // .txfm_type_col
+  TXFM_TYPE_DCT4
+};  // .txfm_type_row
+
+//  ---------------- config inv_adst_dct_8 ----------------
+static const int8_t inv_shift_adst_dct_8[2] = { 0, -5 };
+static const int8_t inv_stage_range_col_adst_dct_8[8] = { 19, 19, 19, 19,
+                                                          19, 19, 18, 18 };
+static const int8_t inv_stage_range_row_adst_dct_8[6] = {
+  19, 19, 19, 19, 19, 19
+};
+static const int8_t inv_cos_bit_col_adst_dct_8[8] = { 13, 13, 13, 13,
+                                                      13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_adst_dct_8[6] = { 13, 13, 13, 13, 13, 13 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_adst_dct_8 = {
+  8,  // .txfm_size
+  8,  // .stage_num_col
+  6,  // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_adst_dct_8,            // .shift
+  inv_stage_range_col_adst_dct_8,  // .stage_range_col
+  inv_stage_range_row_adst_dct_8,  // .stage_range_row
+  inv_cos_bit_col_adst_dct_8,      // .cos_bit_col
+  inv_cos_bit_row_adst_dct_8,      // .cos_bit_row
+  TXFM_TYPE_ADST8,                 // .txfm_type_col
+  TXFM_TYPE_DCT8
+};  // .txfm_type_row
+
+//  ---------------- config inv_adst_dct_16 ----------------
+static const int8_t inv_shift_adst_dct_16[2] = { -1, -5 };
+static const int8_t inv_stage_range_col_adst_dct_16[10] = {
+  19, 19, 19, 19, 19, 19, 19, 19, 18, 18
+};
+static const int8_t inv_stage_range_row_adst_dct_16[8] = { 20, 20, 20, 20,
+                                                           20, 20, 20, 20 };
+static const int8_t inv_cos_bit_col_adst_dct_16[10] = { 13, 13, 13, 13, 13,
+                                                        13, 13, 13, 13, 13 };
+static const int8_t inv_cos_bit_row_adst_dct_16[8] = { 12, 12, 12, 12,
+                                                       12, 12, 12, 12 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_adst_dct_16 = {
+  16,  // .txfm_size
+  10,  // .stage_num_col
+  8,   // .stage_num_row
+  // 0,  // .log_scale
+  inv_shift_adst_dct_16,            // .shift
+  inv_stage_range_col_adst_dct_16,  // .stage_range_col
+  inv_stage_range_row_adst_dct_16,  // .stage_range_row
+  inv_cos_bit_col_adst_dct_16,      // .cos_bit_col
+  inv_cos_bit_row_adst_dct_16,      // .cos_bit_row
+  TXFM_TYPE_ADST16,                 // .txfm_type_col
+  TXFM_TYPE_DCT16
+};  // .txfm_type_row
+
+//  ---------------- config inv_adst_dct_32 ----------------
+static const int8_t inv_shift_adst_dct_32[2] = { -1, -5 };
+static const int8_t inv_stage_range_col_adst_dct_32[12] = {
+  19, 19, 19, 19, 19, 19, 19, 19, 19, 19, 18, 18
+};
+static const int8_t inv_stage_range_row_adst_dct_32[10] = {
+  20, 20, 20, 20, 20, 20, 20, 20, 20, 20
+};
+static const int8_t inv_cos_bit_col_adst_dct_32[12] = {
+  13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13
+};
+static const int8_t inv_cos_bit_row_adst_dct_32[10] = { 12, 12, 12, 12, 12,
+                                                        12, 12, 12, 12, 12 };
+
+static const TXFM_2D_CFG inv_txfm_2d_cfg_adst_dct_32 = {
+  32,  // .txfm_size
+  12,  // .stage_num_col
+  10,  // .stage_num_row
+  // 1,  // .log_scale
+  inv_shift_adst_dct_32,            // .shift
+  inv_stage_range_col_adst_dct_32,  // .stage_range_col
+  inv_stage_range_row_adst_dct_32,  // .stage_range_row
+  inv_cos_bit_col_adst_dct_32,      // .cos_bit_col
+  inv_cos_bit_row_adst_dct_32,      // .cos_bit_row
+  TXFM_TYPE_ADST32,                 // .txfm_type_col
+  TXFM_TYPE_DCT32
+};  // .txfm_type_row
+
+extern const TXFM_2D_CFG *inv_txfm_cfg_ls[TX_TYPES][TX_SIZES];
+
+#endif  // AV1_INV_TXFM2D_CFG_H_
diff --git a/third_party/aom/av1/common/av1_loopfilter.c b/third_party/aom/av1/common/av1_loopfilter.c
new file mode 100644
index 0000000000..5308717954
--- /dev/null
+++ b/third_party/aom/av1/common/av1_loopfilter.c
@@ -0,0 +1,2336 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "av1/common/av1_loopfilter.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/reconinter.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#include "av1/common/seg_common.h"
+
+#define CONFIG_PARALLEL_DEBLOCKING_15TAPLUMAONLY 0
+
+// 64 bit masks for left transform size. Each 1 represents a position where
+// we should apply a loop filter across the left border of an 8x8 block
+// boundary.
+//
+// In the case of TX_16X16->  ( in low order byte first we end up with
+// a mask that looks like this
+//
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//    10101010
+//
+// A loopfilter should be applied to every other 8x8 horizontally.
+static const uint64_t left_64x64_txform_mask[TX_SIZES] = {
+#if CONFIG_CB4X4
+  0xffffffffffffffffULL,  // TX_2X2
+#endif
+  0xffffffffffffffffULL,  // TX_4X4
+  0xffffffffffffffffULL,  // TX_8x8
+  0x5555555555555555ULL,  // TX_16x16
+  0x1111111111111111ULL,  // TX_32x32
+#if CONFIG_TX64X64
+  0x0101010101010101ULL,  // TX_64x64
+#endif                    // CONFIG_TX64X64
+};
+
+// 64 bit masks for above transform size. Each 1 represents a position where
+// we should apply a loop filter across the top border of an 8x8 block
+// boundary.
+//
+// In the case of TX_32x32 ->  ( in low order byte first we end up with
+// a mask that looks like this
+//
+//    11111111
+//    00000000
+//    00000000
+//    00000000
+//    11111111
+//    00000000
+//    00000000
+//    00000000
+//
+// A loopfilter should be applied to every other 4 the row vertically.
+static const uint64_t above_64x64_txform_mask[TX_SIZES] = {
+#if CONFIG_CB4X4
+  0xffffffffffffffffULL,  // TX_4X4
+#endif
+  0xffffffffffffffffULL,  // TX_4X4
+  0xffffffffffffffffULL,  // TX_8x8
+  0x00ff00ff00ff00ffULL,  // TX_16x16
+  0x000000ff000000ffULL,  // TX_32x32
+#if CONFIG_TX64X64
+  0x00000000000000ffULL,  // TX_64x64
+#endif                    // CONFIG_TX64X64
+};
+
+// 64 bit masks for prediction sizes (left). Each 1 represents a position
+// where left border of an 8x8 block. These are aligned to the right most
+// appropriate bit, and then shifted into place.
+//
+// In the case of TX_16x32 ->  ( low order byte first ) we end up with
+// a mask that looks like this :
+//
+//  10000000
+//  10000000
+//  10000000
+//  10000000
+//  00000000
+//  00000000
+//  00000000
+//  00000000
+static const uint64_t left_prediction_mask[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0x0000000000000001ULL,  // BLOCK_2X2,
+  0x0000000000000001ULL,  // BLOCK_2X4,
+  0x0000000000000001ULL,  // BLOCK_4X2,
+#endif
+  0x0000000000000001ULL,  // BLOCK_4X4,
+  0x0000000000000001ULL,  // BLOCK_4X8,
+  0x0000000000000001ULL,  // BLOCK_8X4,
+  0x0000000000000001ULL,  // BLOCK_8X8,
+  0x0000000000000101ULL,  // BLOCK_8X16,
+  0x0000000000000001ULL,  // BLOCK_16X8,
+  0x0000000000000101ULL,  // BLOCK_16X16,
+  0x0000000001010101ULL,  // BLOCK_16X32,
+  0x0000000000000101ULL,  // BLOCK_32X16,
+  0x0000000001010101ULL,  // BLOCK_32X32,
+  0x0101010101010101ULL,  // BLOCK_32X64,
+  0x0000000001010101ULL,  // BLOCK_64X32,
+  0x0101010101010101ULL,  // BLOCK_64X64
+};
+
+// 64 bit mask to shift and set for each prediction size.
+static const uint64_t above_prediction_mask[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0x0000000000000001ULL,  // BLOCK_2X2
+  0x0000000000000001ULL,  // BLOCK_2X4
+  0x0000000000000001ULL,  // BLOCK_4X2
+#endif
+  0x0000000000000001ULL,  // BLOCK_4X4
+  0x0000000000000001ULL,  // BLOCK_4X8
+  0x0000000000000001ULL,  // BLOCK_8X4
+  0x0000000000000001ULL,  // BLOCK_8X8
+  0x0000000000000001ULL,  // BLOCK_8X16,
+  0x0000000000000003ULL,  // BLOCK_16X8
+  0x0000000000000003ULL,  // BLOCK_16X16
+  0x0000000000000003ULL,  // BLOCK_16X32,
+  0x000000000000000fULL,  // BLOCK_32X16,
+  0x000000000000000fULL,  // BLOCK_32X32,
+  0x000000000000000fULL,  // BLOCK_32X64,
+  0x00000000000000ffULL,  // BLOCK_64X32,
+  0x00000000000000ffULL,  // BLOCK_64X64
+};
+// 64 bit mask to shift and set for each prediction size. A bit is set for
+// each 8x8 block that would be in the left most block of the given block
+// size in the 64x64 block.
+static const uint64_t size_mask[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0x0000000000000001ULL,  // BLOCK_2X2
+  0x0000000000000001ULL,  // BLOCK_2X4
+  0x0000000000000001ULL,  // BLOCK_4X2
+#endif
+  0x0000000000000001ULL,  // BLOCK_4X4
+  0x0000000000000001ULL,  // BLOCK_4X8
+  0x0000000000000001ULL,  // BLOCK_8X4
+  0x0000000000000001ULL,  // BLOCK_8X8
+  0x0000000000000101ULL,  // BLOCK_8X16,
+  0x0000000000000003ULL,  // BLOCK_16X8
+  0x0000000000000303ULL,  // BLOCK_16X16
+  0x0000000003030303ULL,  // BLOCK_16X32,
+  0x0000000000000f0fULL,  // BLOCK_32X16,
+  0x000000000f0f0f0fULL,  // BLOCK_32X32,
+  0x0f0f0f0f0f0f0f0fULL,  // BLOCK_32X64,
+  0x00000000ffffffffULL,  // BLOCK_64X32,
+  0xffffffffffffffffULL,  // BLOCK_64X64
+};
+
+// These are used for masking the left and above 32x32 borders.
+static const uint64_t left_border = 0x1111111111111111ULL;
+static const uint64_t above_border = 0x000000ff000000ffULL;
+
+// 16 bit masks for uv transform sizes.
+static const uint16_t left_64x64_txform_mask_uv[TX_SIZES] = {
+#if CONFIG_CB4X4
+  0xffff,  // TX_2X2
+#endif
+  0xffff,  // TX_4X4
+  0xffff,  // TX_8x8
+  0x5555,  // TX_16x16
+  0x1111,  // TX_32x32
+#if CONFIG_TX64X64
+  0x0101,  // TX_64x64, never used
+#endif     // CONFIG_TX64X64
+};
+
+static const uint16_t above_64x64_txform_mask_uv[TX_SIZES] = {
+#if CONFIG_CB4X4
+  0xffff,  // TX_2X2
+#endif
+  0xffff,  // TX_4X4
+  0xffff,  // TX_8x8
+  0x0f0f,  // TX_16x16
+  0x000f,  // TX_32x32
+#if CONFIG_TX64X64
+  0x0003,  // TX_64x64, never used
+#endif     // CONFIG_TX64X64
+};
+
+// 16 bit left mask to shift and set for each uv prediction size.
+static const uint16_t left_prediction_mask_uv[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0x0001,  // BLOCK_2X2,
+  0x0001,  // BLOCK_2X4,
+  0x0001,  // BLOCK_4X2,
+#endif
+  0x0001,  // BLOCK_4X4,
+  0x0001,  // BLOCK_4X8,
+  0x0001,  // BLOCK_8X4,
+  0x0001,  // BLOCK_8X8,
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8,
+  0x0001,  // BLOCK_16X16,
+  0x0011,  // BLOCK_16X32,
+  0x0001,  // BLOCK_32X16,
+  0x0011,  // BLOCK_32X32,
+  0x1111,  // BLOCK_32X64
+  0x0011,  // BLOCK_64X32,
+  0x1111,  // BLOCK_64X64
+};
+// 16 bit above mask to shift and set for uv each prediction size.
+static const uint16_t above_prediction_mask_uv[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0x0001,  // BLOCK_2X2
+  0x0001,  // BLOCK_2X4
+  0x0001,  // BLOCK_4X2
+#endif
+  0x0001,  // BLOCK_4X4
+  0x0001,  // BLOCK_4X8
+  0x0001,  // BLOCK_8X4
+  0x0001,  // BLOCK_8X8
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8
+  0x0001,  // BLOCK_16X16
+  0x0001,  // BLOCK_16X32,
+  0x0003,  // BLOCK_32X16,
+  0x0003,  // BLOCK_32X32,
+  0x0003,  // BLOCK_32X64,
+  0x000f,  // BLOCK_64X32,
+  0x000f,  // BLOCK_64X64
+};
+
+// 64 bit mask to shift and set for each uv prediction size
+static const uint16_t size_mask_uv[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0x0001,  // BLOCK_2X2
+  0x0001,  // BLOCK_2X4
+  0x0001,  // BLOCK_4X2
+#endif
+  0x0001,  // BLOCK_4X4
+  0x0001,  // BLOCK_4X8
+  0x0001,  // BLOCK_8X4
+  0x0001,  // BLOCK_8X8
+  0x0001,  // BLOCK_8X16,
+  0x0001,  // BLOCK_16X8
+  0x0001,  // BLOCK_16X16
+  0x0011,  // BLOCK_16X32,
+  0x0003,  // BLOCK_32X16,
+  0x0033,  // BLOCK_32X32,
+  0x3333,  // BLOCK_32X64,
+  0x00ff,  // BLOCK_64X32,
+  0xffff,  // BLOCK_64X64
+};
+static const uint16_t left_border_uv = 0x1111;
+static const uint16_t above_border_uv = 0x000f;
+
+static const int mode_lf_lut[] = {
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0,  // INTRA_MODES
+#if CONFIG_ALT_INTRA
+  0,
+#endif
+  1, 1, 0, 1,  // INTER_MODES (ZEROMV == 0)
+#if CONFIG_EXT_INTER
+  1, 1, 1, 1, 1, 1, 1, 1, 0, 1  // INTER_COMPOUND_MODES (ZERO_ZEROMV == 0)
+#endif                          // CONFIG_EXT_INTER
+};
+
+static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
+  int lvl;
+
+  // For each possible value for the loop filter fill out limits
+  for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++) {
+    // Set loop filter parameters that control sharpness.
+    int block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));
+
+    if (sharpness_lvl > 0) {
+      if (block_inside_limit > (9 - sharpness_lvl))
+        block_inside_limit = (9 - sharpness_lvl);
+    }
+
+    if (block_inside_limit < 1) block_inside_limit = 1;
+
+    memset(lfi->lfthr[lvl].lim, block_inside_limit, SIMD_WIDTH);
+    memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
+           SIMD_WIDTH);
+  }
+}
+#if CONFIG_EXT_DELTA_Q
+static uint8_t get_filter_level(const AV1_COMMON *cm,
+                                const loop_filter_info_n *lfi_n,
+                                const MB_MODE_INFO *mbmi) {
+#if CONFIG_SUPERTX
+  const int segment_id = AOMMIN(mbmi->segment_id, mbmi->segment_id_supertx);
+  assert(
+      IMPLIES(supertx_enabled(mbmi), mbmi->segment_id_supertx != MAX_SEGMENTS));
+  assert(IMPLIES(supertx_enabled(mbmi),
+                 mbmi->segment_id_supertx <= mbmi->segment_id));
+#else
+  const int segment_id = mbmi->segment_id;
+#endif  // CONFIG_SUPERTX
+  if (cm->delta_lf_present_flag) {
+    int lvl_seg = clamp(mbmi->current_delta_lf_from_base + cm->lf.filter_level,
+                        0, MAX_LOOP_FILTER);
+    const int scale = 1 << (lvl_seg >> 5);
+    if (segfeature_active(&cm->seg, segment_id, SEG_LVL_ALT_LF)) {
+      const int data = get_segdata(&cm->seg, segment_id, SEG_LVL_ALT_LF);
+      lvl_seg =
+          clamp(cm->seg.abs_delta == SEGMENT_ABSDATA ? data : lvl_seg + data, 0,
+                MAX_LOOP_FILTER);
+    }
+
+    if (cm->lf.mode_ref_delta_enabled) {
+      lvl_seg += cm->lf.ref_deltas[mbmi->ref_frame[0]] * scale;
+      if (mbmi->ref_frame[0] > INTRA_FRAME)
+        lvl_seg += cm->lf.mode_deltas[mode_lf_lut[mbmi->mode]] * scale;
+      lvl_seg = clamp(lvl_seg, 0, MAX_LOOP_FILTER);
+    }
+    return lvl_seg;
+  } else {
+    return lfi_n->lvl[segment_id][mbmi->ref_frame[0]][mode_lf_lut[mbmi->mode]];
+  }
+}
+#else
+static uint8_t get_filter_level(const loop_filter_info_n *lfi_n,
+                                const MB_MODE_INFO *mbmi) {
+#if CONFIG_SUPERTX
+  const int segment_id = AOMMIN(mbmi->segment_id, mbmi->segment_id_supertx);
+  assert(
+      IMPLIES(supertx_enabled(mbmi), mbmi->segment_id_supertx != MAX_SEGMENTS));
+  assert(IMPLIES(supertx_enabled(mbmi),
+                 mbmi->segment_id_supertx <= mbmi->segment_id));
+#else
+  const int segment_id = mbmi->segment_id;
+#endif  // CONFIG_SUPERTX
+  return lfi_n->lvl[segment_id][mbmi->ref_frame[0]][mode_lf_lut[mbmi->mode]];
+}
+#endif
+
+#define NELEMENTS(x) (sizeof((x)) / sizeof((x)[0]))
+
+void av1_loop_filter_init(AV1_COMMON *cm) {
+  assert(MB_MODE_COUNT == NELEMENTS(mode_lf_lut));
+  loop_filter_info_n *lfi = &cm->lf_info;
+  struct loopfilter *lf = &cm->lf;
+  int lvl;
+
+  // init limits for given sharpness
+  update_sharpness(lfi, lf->sharpness_level);
+  lf->last_sharpness_level = lf->sharpness_level;
+
+  // init hev threshold const vectors
+  for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++)
+    memset(lfi->lfthr[lvl].hev_thr, (lvl >> 4), SIMD_WIDTH);
+}
+
+void av1_loop_filter_frame_init(AV1_COMMON *cm, int default_filt_lvl) {
+  int seg_id;
+  // n_shift is the multiplier for lf_deltas
+  // the multiplier is 1 for when filter_lvl is between 0 and 31;
+  // 2 when filter_lvl is between 32 and 63
+  const int scale = 1 << (default_filt_lvl >> 5);
+  loop_filter_info_n *const lfi = &cm->lf_info;
+  struct loopfilter *const lf = &cm->lf;
+  const struct segmentation *const seg = &cm->seg;
+
+  // update limits if sharpness has changed
+  if (lf->last_sharpness_level != lf->sharpness_level) {
+    update_sharpness(lfi, lf->sharpness_level);
+    lf->last_sharpness_level = lf->sharpness_level;
+  }
+
+  for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
+    int lvl_seg = default_filt_lvl;
+    if (segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
+      const int data = get_segdata(seg, seg_id, SEG_LVL_ALT_LF);
+      lvl_seg = clamp(
+          seg->abs_delta == SEGMENT_ABSDATA ? data : default_filt_lvl + data, 0,
+          MAX_LOOP_FILTER);
+    }
+
+    if (!lf->mode_ref_delta_enabled) {
+      // we could get rid of this if we assume that deltas are set to
+      // zero when not in use; encoder always uses deltas
+      memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id]));
+    } else {
+      int ref, mode;
+      const int intra_lvl = lvl_seg + lf->ref_deltas[INTRA_FRAME] * scale;
+      lfi->lvl[seg_id][INTRA_FRAME][0] = clamp(intra_lvl, 0, MAX_LOOP_FILTER);
+
+      for (ref = LAST_FRAME; ref < TOTAL_REFS_PER_FRAME; ++ref) {
+        for (mode = 0; mode < MAX_MODE_LF_DELTAS; ++mode) {
+          const int inter_lvl = lvl_seg + lf->ref_deltas[ref] * scale +
+                                lf->mode_deltas[mode] * scale;
+          lfi->lvl[seg_id][ref][mode] = clamp(inter_lvl, 0, MAX_LOOP_FILTER);
+        }
+      }
+    }
+  }
+}
+
+static void filter_selectively_vert_row2(int subsampling_factor, uint8_t *s,
+                                         int pitch, unsigned int mask_16x16_l,
+                                         unsigned int mask_8x8_l,
+                                         unsigned int mask_4x4_l,
+                                         unsigned int mask_4x4_int_l,
+                                         const loop_filter_info_n *lfi_n,
+                                         const uint8_t *lfl) {
+  const int mask_shift = subsampling_factor ? 4 : 8;
+  const int mask_cutoff = subsampling_factor ? 0xf : 0xff;
+  const int lfl_forward = subsampling_factor ? 4 : 8;
+
+  unsigned int mask_16x16_0 = mask_16x16_l & mask_cutoff;
+  unsigned int mask_8x8_0 = mask_8x8_l & mask_cutoff;
+  unsigned int mask_4x4_0 = mask_4x4_l & mask_cutoff;
+  unsigned int mask_4x4_int_0 = mask_4x4_int_l & mask_cutoff;
+  unsigned int mask_16x16_1 = (mask_16x16_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_8x8_1 = (mask_8x8_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_4x4_1 = (mask_4x4_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_4x4_int_1 = (mask_4x4_int_l >> mask_shift) & mask_cutoff;
+  unsigned int mask;
+
+  for (mask = mask_16x16_0 | mask_8x8_0 | mask_4x4_0 | mask_4x4_int_0 |
+              mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
+       mask; mask >>= 1) {
+    const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
+    const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);
+
+    if (mask & 1) {
+      if ((mask_16x16_0 | mask_16x16_1) & 1) {
+        if ((mask_16x16_0 & mask_16x16_1) & 1) {
+          aom_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                   lfi0->hev_thr);
+        } else if (mask_16x16_0 & 1) {
+          aom_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr);
+        } else {
+          aom_lpf_vertical_16(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
+                              lfi1->hev_thr);
+        }
+      }
+
+      if ((mask_8x8_0 | mask_8x8_1) & 1) {
+        if ((mask_8x8_0 & mask_8x8_1) & 1) {
+          aom_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                  lfi1->hev_thr);
+        } else if (mask_8x8_0 & 1) {
+          aom_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr);
+        } else {
+          aom_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
+                             lfi1->hev_thr);
+        }
+      }
+
+      if ((mask_4x4_0 | mask_4x4_1) & 1) {
+        if ((mask_4x4_0 & mask_4x4_1) & 1) {
+          aom_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                  lfi1->hev_thr);
+        } else if (mask_4x4_0 & 1) {
+          aom_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr);
+        } else {
+          aom_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
+                             lfi1->hev_thr);
+        }
+      }
+
+      if ((mask_4x4_int_0 | mask_4x4_int_1) & 1) {
+        if ((mask_4x4_int_0 & mask_4x4_int_1) & 1) {
+          aom_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
+                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                  lfi1->hev_thr);
+        } else if (mask_4x4_int_0 & 1) {
+          aom_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
+                             lfi0->hev_thr);
+        } else {
+          aom_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim, lfi1->lim,
+                             lfi1->hev_thr);
+        }
+      }
+    }
+
+    s += 8;
+    lfl += 1;
+    mask_16x16_0 >>= 1;
+    mask_8x8_0 >>= 1;
+    mask_4x4_0 >>= 1;
+    mask_4x4_int_0 >>= 1;
+    mask_16x16_1 >>= 1;
+    mask_8x8_1 >>= 1;
+    mask_4x4_1 >>= 1;
+    mask_4x4_int_1 >>= 1;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_filter_selectively_vert_row2(
+    int subsampling_factor, uint16_t *s, int pitch, unsigned int mask_16x16_l,
+    unsigned int mask_8x8_l, unsigned int mask_4x4_l,
+    unsigned int mask_4x4_int_l, const loop_filter_info_n *lfi_n,
+    const uint8_t *lfl, int bd) {
+  const int mask_shift = subsampling_factor ? 4 : 8;
+  const int mask_cutoff = subsampling_factor ? 0xf : 0xff;
+  const int lfl_forward = subsampling_factor ? 4 : 8;
+
+  unsigned int mask_16x16_0 = mask_16x16_l & mask_cutoff;
+  unsigned int mask_8x8_0 = mask_8x8_l & mask_cutoff;
+  unsigned int mask_4x4_0 = mask_4x4_l & mask_cutoff;
+  unsigned int mask_4x4_int_0 = mask_4x4_int_l & mask_cutoff;
+  unsigned int mask_16x16_1 = (mask_16x16_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_8x8_1 = (mask_8x8_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_4x4_1 = (mask_4x4_l >> mask_shift) & mask_cutoff;
+  unsigned int mask_4x4_int_1 = (mask_4x4_int_l >> mask_shift) & mask_cutoff;
+  unsigned int mask;
+
+  for (mask = mask_16x16_0 | mask_8x8_0 | mask_4x4_0 | mask_4x4_int_0 |
+              mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
+       mask; mask >>= 1) {
+    const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
+    const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);
+
+    if (mask & 1) {
+      if ((mask_16x16_0 | mask_16x16_1) & 1) {
+        if ((mask_16x16_0 & mask_16x16_1) & 1) {
+          aom_highbd_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                          lfi0->hev_thr, bd);
+        } else if (mask_16x16_0 & 1) {
+          aom_highbd_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
+                                     lfi0->hev_thr, bd);
+        } else {
+          aom_highbd_lpf_vertical_16(s + 8 * pitch, pitch, lfi1->mblim,
+                                     lfi1->lim, lfi1->hev_thr, bd);
+        }
+      }
+
+      if ((mask_8x8_0 | mask_8x8_1) & 1) {
+        if ((mask_8x8_0 & mask_8x8_1) & 1) {
+          aom_highbd_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                         lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                         lfi1->hev_thr, bd);
+        } else if (mask_8x8_0 & 1) {
+          aom_highbd_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim,
+                                    lfi0->hev_thr, bd);
+        } else {
+          aom_highbd_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim,
+                                    lfi1->lim, lfi1->hev_thr, bd);
+        }
+      }
+
+      if ((mask_4x4_0 | mask_4x4_1) & 1) {
+        if ((mask_4x4_0 & mask_4x4_1) & 1) {
+          aom_highbd_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
+                                         lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                         lfi1->hev_thr, bd);
+        } else if (mask_4x4_0 & 1) {
+          aom_highbd_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim,
+                                    lfi0->hev_thr, bd);
+        } else {
+          aom_highbd_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim,
+                                    lfi1->lim, lfi1->hev_thr, bd);
+        }
+      }
+
+      if ((mask_4x4_int_0 | mask_4x4_int_1) & 1) {
+        if ((mask_4x4_int_0 & mask_4x4_int_1) & 1) {
+          aom_highbd_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
+                                         lfi0->hev_thr, lfi1->mblim, lfi1->lim,
+                                         lfi1->hev_thr, bd);
+        } else if (mask_4x4_int_0 & 1) {
+          aom_highbd_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
+                                    lfi0->hev_thr, bd);
+        } else {
+          aom_highbd_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim,
+                                    lfi1->lim, lfi1->hev_thr, bd);
+        }
+      }
+    }
+
+    s += 8;
+    lfl += 1;
+    mask_16x16_0 >>= 1;
+    mask_8x8_0 >>= 1;
+    mask_4x4_0 >>= 1;
+    mask_4x4_int_0 >>= 1;
+    mask_16x16_1 >>= 1;
+    mask_8x8_1 >>= 1;
+    mask_4x4_1 >>= 1;
+    mask_4x4_int_1 >>= 1;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static void filter_selectively_horiz(
+    uint8_t *s, int pitch, unsigned int mask_16x16, unsigned int mask_8x8,
+    unsigned int mask_4x4, unsigned int mask_4x4_int,
+    const loop_filter_info_n *lfi_n, const uint8_t *lfl) {
+  unsigned int mask;
+  int count;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; mask;
+       mask >>= count) {
+    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
+
+    count = 1;
+    if (mask & 1) {
+      if (mask_16x16 & 1) {
+        if ((mask_16x16 & 3) == 3) {
+          aom_lpf_horizontal_edge_16(s, pitch, lfi->mblim, lfi->lim,
+                                     lfi->hev_thr);
+          count = 2;
+        } else {
+          aom_lpf_horizontal_edge_8(s, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr);
+        }
+      } else if (mask_8x8 & 1) {
+        if ((mask_8x8 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
+
+          aom_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr, lfin->mblim, lfin->lim,
+                                    lfin->hev_thr);
+
+          if ((mask_4x4_int & 3) == 3) {
+            aom_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
+                                      lfi->lim, lfi->hev_thr, lfin->mblim,
+                                      lfin->lim, lfin->hev_thr);
+          } else {
+            if (mask_4x4_int & 1)
+              aom_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                   lfi->hev_thr);
+            else if (mask_4x4_int & 2)
+              aom_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                   lfin->lim, lfin->hev_thr);
+          }
+          count = 2;
+        } else {
+          aom_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+
+          if (mask_4x4_int & 1)
+            aom_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                 lfi->hev_thr);
+        }
+      } else if (mask_4x4 & 1) {
+        if ((mask_4x4 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
+
+          aom_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr, lfin->mblim, lfin->lim,
+                                    lfin->hev_thr);
+          if ((mask_4x4_int & 3) == 3) {
+            aom_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
+                                      lfi->lim, lfi->hev_thr, lfin->mblim,
+                                      lfin->lim, lfin->hev_thr);
+          } else {
+            if (mask_4x4_int & 1)
+              aom_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                   lfi->hev_thr);
+            else if (mask_4x4_int & 2)
+              aom_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                   lfin->lim, lfin->hev_thr);
+          }
+          count = 2;
+        } else {
+          aom_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+
+          if (mask_4x4_int & 1)
+            aom_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                 lfi->hev_thr);
+        }
+      } else if (mask_4x4_int & 1) {
+        aom_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                             lfi->hev_thr);
+      }
+    }
+    s += 8 * count;
+    lfl += count;
+    mask_16x16 >>= count;
+    mask_8x8 >>= count;
+    mask_4x4 >>= count;
+    mask_4x4_int >>= count;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_filter_selectively_horiz(
+    uint16_t *s, int pitch, unsigned int mask_16x16, unsigned int mask_8x8,
+    unsigned int mask_4x4, unsigned int mask_4x4_int,
+    const loop_filter_info_n *lfi_n, const uint8_t *lfl, int bd) {
+  unsigned int mask;
+  int count;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; mask;
+       mask >>= count) {
+    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
+
+    count = 1;
+    if (mask & 1) {
+      if (mask_16x16 & 1) {
+        if ((mask_16x16 & 3) == 3) {
+          aom_highbd_lpf_horizontal_edge_16(s, pitch, lfi->mblim, lfi->lim,
+                                            lfi->hev_thr, bd);
+          count = 2;
+        } else {
+          aom_highbd_lpf_horizontal_edge_8(s, pitch, lfi->mblim, lfi->lim,
+                                           lfi->hev_thr, bd);
+        }
+      } else if (mask_8x8 & 1) {
+        if ((mask_8x8 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
+
+          aom_highbd_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
+                                           lfi->hev_thr, lfin->mblim, lfin->lim,
+                                           lfin->hev_thr, bd);
+
+          if ((mask_4x4_int & 3) == 3) {
+            aom_highbd_lpf_horizontal_4_dual(
+                s + 4 * pitch, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
+                lfin->mblim, lfin->lim, lfin->hev_thr, bd);
+          } else {
+            if (mask_4x4_int & 1) {
+              aom_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                          lfi->lim, lfi->hev_thr, bd);
+            } else if (mask_4x4_int & 2) {
+              aom_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                          lfin->lim, lfin->hev_thr, bd);
+            }
+          }
+          count = 2;
+        } else {
+          aom_highbd_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim,
+                                      lfi->hev_thr, bd);
+
+          if (mask_4x4_int & 1) {
+            aom_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                        lfi->lim, lfi->hev_thr, bd);
+          }
+        }
+      } else if (mask_4x4 & 1) {
+        if ((mask_4x4 & 3) == 3) {
+          // Next block's thresholds.
+          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);
+
+          aom_highbd_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
+                                           lfi->hev_thr, lfin->mblim, lfin->lim,
+                                           lfin->hev_thr, bd);
+          if ((mask_4x4_int & 3) == 3) {
+            aom_highbd_lpf_horizontal_4_dual(
+                s + 4 * pitch, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
+                lfin->mblim, lfin->lim, lfin->hev_thr, bd);
+          } else {
+            if (mask_4x4_int & 1) {
+              aom_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                          lfi->lim, lfi->hev_thr, bd);
+            } else if (mask_4x4_int & 2) {
+              aom_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
+                                          lfin->lim, lfin->hev_thr, bd);
+            }
+          }
+          count = 2;
+        } else {
+          aom_highbd_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim,
+                                      lfi->hev_thr, bd);
+
+          if (mask_4x4_int & 1) {
+            aom_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
+                                        lfi->lim, lfi->hev_thr, bd);
+          }
+        }
+      } else if (mask_4x4_int & 1) {
+        aom_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
+                                    lfi->hev_thr, bd);
+      }
+    }
+    s += 8 * count;
+    lfl += count;
+    mask_16x16 >>= count;
+    mask_8x8 >>= count;
+    mask_4x4 >>= count;
+    mask_4x4_int >>= count;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+// This function ors into the current lfm structure, where to do loop
+// filters for the specific mi we are looking at. It uses information
+// including the block_size_type (32x16, 32x32, etc.), the transform size,
+// whether there were any coefficients encoded, and the loop filter strength
+// block we are currently looking at. Shift is used to position the
+// 1's we produce.
+// TODO(JBB) Need another function for different resolution color..
+static void build_masks(AV1_COMMON *const cm,
+                        const loop_filter_info_n *const lfi_n,
+                        const MODE_INFO *mi, const int shift_y,
+                        const int shift_uv, LOOP_FILTER_MASK *lfm) {
+  const MB_MODE_INFO *mbmi = &mi->mbmi;
+  const BLOCK_SIZE block_size = mbmi->sb_type;
+  // TODO(debargha): Check if masks can be setup correctly when
+  // rectangular transfroms are used with the EXT_TX expt.
+  const TX_SIZE tx_size_y = txsize_sqr_map[mbmi->tx_size];
+  const TX_SIZE tx_size_y_left = txsize_horz_map[mbmi->tx_size];
+  const TX_SIZE tx_size_y_above = txsize_vert_map[mbmi->tx_size];
+  const TX_SIZE tx_size_uv =
+      txsize_sqr_map[uv_txsize_lookup[block_size][mbmi->tx_size][1][1]];
+  const TX_SIZE tx_size_uv_left =
+      txsize_horz_map[uv_txsize_lookup[block_size][mbmi->tx_size][1][1]];
+  const TX_SIZE tx_size_uv_above =
+      txsize_vert_map[uv_txsize_lookup[block_size][mbmi->tx_size][1][1]];
+#if CONFIG_EXT_DELTA_Q
+  const int filter_level = get_filter_level(cm, lfi_n, mbmi);
+#else
+  const int filter_level = get_filter_level(lfi_n, mbmi);
+  (void)cm;
+#endif
+  uint64_t *const left_y = &lfm->left_y[tx_size_y_left];
+  uint64_t *const above_y = &lfm->above_y[tx_size_y_above];
+  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
+  uint16_t *const left_uv = &lfm->left_uv[tx_size_uv_left];
+  uint16_t *const above_uv = &lfm->above_uv[tx_size_uv_above];
+  uint16_t *const int_4x4_uv = &lfm->left_int_4x4_uv;
+  int i;
+
+  // If filter level is 0 we don't loop filter.
+  if (!filter_level) {
+    return;
+  } else {
+    const int w = num_8x8_blocks_wide_lookup[block_size];
+    const int h = num_8x8_blocks_high_lookup[block_size];
+    const int row = (shift_y >> MAX_MIB_SIZE_LOG2);
+    const int col = shift_y - (row << MAX_MIB_SIZE_LOG2);
+
+    for (i = 0; i < h; i++) memset(&lfm->lfl_y[row + i][col], filter_level, w);
+  }
+
+  // These set 1 in the current block size for the block size edges.
+  // For instance if the block size is 32x16, we'll set:
+  //    above =   1111
+  //              0000
+  //    and
+  //    left  =   1000
+  //          =   1000
+  // NOTE : In this example the low bit is left most ( 1000 ) is stored as
+  //        1,  not 8...
+  //
+  // U and V set things on a 16 bit scale.
+  //
+  *above_y |= above_prediction_mask[block_size] << shift_y;
+  *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
+  *left_y |= left_prediction_mask[block_size] << shift_y;
+  *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;
+
+  // If the block has no coefficients and is not intra we skip applying
+  // the loop filter on block edges.
+  if (mbmi->skip && is_inter_block(mbmi)) return;
+
+  // Here we are adding a mask for the transform size. The transform
+  // size mask is set to be correct for a 64x64 prediction block size. We
+  // mask to match the size of the block we are working on and then shift it
+  // into place..
+  *above_y |= (size_mask[block_size] & above_64x64_txform_mask[tx_size_y_above])
+              << shift_y;
+  *above_uv |=
+      (size_mask_uv[block_size] & above_64x64_txform_mask_uv[tx_size_uv_above])
+      << shift_uv;
+
+  *left_y |= (size_mask[block_size] & left_64x64_txform_mask[tx_size_y_left])
+             << shift_y;
+  *left_uv |=
+      (size_mask_uv[block_size] & left_64x64_txform_mask_uv[tx_size_uv_left])
+      << shift_uv;
+
+  // Here we are trying to determine what to do with the internal 4x4 block
+  // boundaries.  These differ from the 4x4 boundaries on the outside edge of
+  // an 8x8 in that the internal ones can be skipped and don't depend on
+  // the prediction block size.
+  if (tx_size_y == TX_4X4)
+    *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y;
+
+  if (tx_size_uv == TX_4X4)
+    *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
+}
+
+// This function does the same thing as the one above with the exception that
+// it only affects the y masks. It exists because for blocks < 16x16 in size,
+// we only update u and v masks on the first block.
+static void build_y_mask(AV1_COMMON *const cm,
+                         const loop_filter_info_n *const lfi_n,
+                         const MODE_INFO *mi, const int shift_y,
+#if CONFIG_SUPERTX
+                         int supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         LOOP_FILTER_MASK *lfm) {
+  const MB_MODE_INFO *mbmi = &mi->mbmi;
+  const TX_SIZE tx_size_y = txsize_sqr_map[mbmi->tx_size];
+  const TX_SIZE tx_size_y_left = txsize_horz_map[mbmi->tx_size];
+  const TX_SIZE tx_size_y_above = txsize_vert_map[mbmi->tx_size];
+#if CONFIG_SUPERTX
+  const BLOCK_SIZE block_size =
+      supertx_enabled ? (BLOCK_SIZE)(3 * tx_size_y) : mbmi->sb_type;
+#else
+  const BLOCK_SIZE block_size = mbmi->sb_type;
+#endif
+#if CONFIG_EXT_DELTA_Q
+  const int filter_level = get_filter_level(cm, lfi_n, mbmi);
+#else
+  const int filter_level = get_filter_level(lfi_n, mbmi);
+  (void)cm;
+#endif
+  uint64_t *const left_y = &lfm->left_y[tx_size_y_left];
+  uint64_t *const above_y = &lfm->above_y[tx_size_y_above];
+  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
+  int i;
+
+  if (!filter_level) {
+    return;
+  } else {
+    const int w = num_8x8_blocks_wide_lookup[block_size];
+    const int h = num_8x8_blocks_high_lookup[block_size];
+    const int row = (shift_y >> MAX_MIB_SIZE_LOG2);
+    const int col = shift_y - (row << MAX_MIB_SIZE_LOG2);
+
+    for (i = 0; i < h; i++) memset(&lfm->lfl_y[row + i][col], filter_level, w);
+  }
+
+  *above_y |= above_prediction_mask[block_size] << shift_y;
+  *left_y |= left_prediction_mask[block_size] << shift_y;
+
+  if (mbmi->skip && is_inter_block(mbmi)) return;
+
+  *above_y |= (size_mask[block_size] & above_64x64_txform_mask[tx_size_y_above])
+              << shift_y;
+
+  *left_y |= (size_mask[block_size] & left_64x64_txform_mask[tx_size_y_left])
+             << shift_y;
+
+  if (tx_size_y == TX_4X4)
+    *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y;
+}
+
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+// This function update the bit masks for the entire 64x64 region represented
+// by mi_row, mi_col. In case one of the edge is a tile boundary, loop filtering
+// for that edge is disabled. This function only check the tile boundary info
+// for the top left corner mi to determine the boundary information for the
+// top and left edge of the whole super block
+static void update_tile_boundary_filter_mask(AV1_COMMON *const cm,
+                                             const int mi_row, const int mi_col,
+                                             LOOP_FILTER_MASK *lfm) {
+  int i;
+  MODE_INFO *const mi = cm->mi + mi_row * cm->mi_stride + mi_col;
+
+  if (mi->mbmi.boundary_info & TILE_LEFT_BOUNDARY) {
+    for (i = 0; i <= TX_32X32; i++) {
+      lfm->left_y[i] &= 0xfefefefefefefefeULL;
+      lfm->left_uv[i] &= 0xeeee;
+    }
+  }
+
+  if (mi->mbmi.boundary_info & TILE_ABOVE_BOUNDARY) {
+    for (i = 0; i <= TX_32X32; i++) {
+      lfm->above_y[i] &= 0xffffffffffffff00ULL;
+      lfm->above_uv[i] &= 0xfff0;
+    }
+  }
+}
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+// This function sets up the bit masks for the entire 64x64 region represented
+// by mi_row, mi_col.
+// TODO(JBB): This function only works for yv12.
+void av1_setup_mask(AV1_COMMON *const cm, const int mi_row, const int mi_col,
+                    MODE_INFO **mi, const int mode_info_stride,
+                    LOOP_FILTER_MASK *lfm) {
+  int idx_32, idx_16, idx_8;
+  const loop_filter_info_n *const lfi_n = &cm->lf_info;
+  MODE_INFO **mip = mi;
+  MODE_INFO **mip2 = mi;
+
+  // These are offsets to the next mi in the 64x64 block. It is what gets
+  // added to the mi ptr as we go through each loop. It helps us to avoid
+  // setting up special row and column counters for each index. The last step
+  // brings us out back to the starting position.
+  const int offset_32[] = { 4, (mode_info_stride << 2) - 4, 4,
+                            -(mode_info_stride << 2) - 4 };
+  const int offset_16[] = { 2, (mode_info_stride << 1) - 2, 2,
+                            -(mode_info_stride << 1) - 2 };
+  const int offset[] = { 1, mode_info_stride - 1, 1, -mode_info_stride - 1 };
+
+  // Following variables represent shifts to position the current block
+  // mask over the appropriate block. A shift of 36 to the left will move
+  // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left
+  // 4 rows to the appropriate spot.
+  const int shift_32_y[] = { 0, 4, 32, 36 };
+  const int shift_16_y[] = { 0, 2, 16, 18 };
+  const int shift_8_y[] = { 0, 1, 8, 9 };
+  const int shift_32_uv[] = { 0, 2, 8, 10 };
+  const int shift_16_uv[] = { 0, 1, 4, 5 };
+  int i;
+  const int max_rows = AOMMIN(cm->mi_rows - mi_row, MAX_MIB_SIZE);
+  const int max_cols = AOMMIN(cm->mi_cols - mi_col, MAX_MIB_SIZE);
+#if CONFIG_EXT_PARTITION
+  assert(0 && "Not yet updated");
+#endif  // CONFIG_EXT_PARTITION
+
+  av1_zero(*lfm);
+  assert(mip[0] != NULL);
+
+  // TODO(jimbankoski): Try moving most of the following code into decode
+  // loop and storing lfm in the mbmi structure so that we don't have to go
+  // through the recursive loop structure multiple times.
+  switch (mip[0]->mbmi.sb_type) {
+    case BLOCK_64X64: build_masks(cm, lfi_n, mip[0], 0, 0, lfm); break;
+    case BLOCK_64X32: build_masks(cm, lfi_n, mip[0], 0, 0, lfm);
+#if CONFIG_SUPERTX && CONFIG_TX64X64
+      if (supertx_enabled(&mip[0]->mbmi)) break;
+#endif  // CONFIG_SUPERTX && CONFIG_TX64X64
+      mip2 = mip + mode_info_stride * 4;
+      if (4 >= max_rows) break;
+      build_masks(cm, lfi_n, mip2[0], 32, 8, lfm);
+      break;
+    case BLOCK_32X64: build_masks(cm, lfi_n, mip[0], 0, 0, lfm);
+#if CONFIG_SUPERTX && CONFIG_TX64X64
+      if (supertx_enabled(&mip[0]->mbmi)) break;
+#endif  // CONFIG_SUPERTX && CONFIG_TX64X64
+      mip2 = mip + 4;
+      if (4 >= max_cols) break;
+      build_masks(cm, lfi_n, mip2[0], 4, 2, lfm);
+      break;
+    default:
+#if CONFIG_SUPERTX && CONFIG_TX64X64
+      if (mip[0]->mbmi.tx_size == TX_64X64) {
+        build_masks(cm, lfi_n, mip[0], 0, 0, lfm);
+      } else {
+#endif  // CONFIG_SUPERTX && CONFIG_TX64X64
+        for (idx_32 = 0; idx_32 < 4; mip += offset_32[idx_32], ++idx_32) {
+          const int shift_y_32 = shift_32_y[idx_32];
+          const int shift_uv_32 = shift_32_uv[idx_32];
+          const int mi_32_col_offset = ((idx_32 & 1) << 2);
+          const int mi_32_row_offset = ((idx_32 >> 1) << 2);
+          if (mi_32_col_offset >= max_cols || mi_32_row_offset >= max_rows)
+            continue;
+          switch (mip[0]->mbmi.sb_type) {
+            case BLOCK_32X32:
+              build_masks(cm, lfi_n, mip[0], shift_y_32, shift_uv_32, lfm);
+              break;
+            case BLOCK_32X16:
+              build_masks(cm, lfi_n, mip[0], shift_y_32, shift_uv_32, lfm);
+#if CONFIG_SUPERTX
+              if (supertx_enabled(&mip[0]->mbmi)) break;
+#endif
+              if (mi_32_row_offset + 2 >= max_rows) continue;
+              mip2 = mip + mode_info_stride * 2;
+              build_masks(cm, lfi_n, mip2[0], shift_y_32 + 16, shift_uv_32 + 4,
+                          lfm);
+              break;
+            case BLOCK_16X32:
+              build_masks(cm, lfi_n, mip[0], shift_y_32, shift_uv_32, lfm);
+#if CONFIG_SUPERTX
+              if (supertx_enabled(&mip[0]->mbmi)) break;
+#endif
+              if (mi_32_col_offset + 2 >= max_cols) continue;
+              mip2 = mip + 2;
+              build_masks(cm, lfi_n, mip2[0], shift_y_32 + 2, shift_uv_32 + 1,
+                          lfm);
+              break;
+            default:
+#if CONFIG_SUPERTX
+              if (mip[0]->mbmi.tx_size == TX_32X32) {
+                build_masks(cm, lfi_n, mip[0], shift_y_32, shift_uv_32, lfm);
+                break;
+              }
+#endif
+              for (idx_16 = 0; idx_16 < 4; mip += offset_16[idx_16], ++idx_16) {
+                const int shift_y_32_16 = shift_y_32 + shift_16_y[idx_16];
+                const int shift_uv_32_16 = shift_uv_32 + shift_16_uv[idx_16];
+                const int mi_16_col_offset =
+                    mi_32_col_offset + ((idx_16 & 1) << 1);
+                const int mi_16_row_offset =
+                    mi_32_row_offset + ((idx_16 >> 1) << 1);
+
+                if (mi_16_col_offset >= max_cols ||
+                    mi_16_row_offset >= max_rows)
+                  continue;
+
+                switch (mip[0]->mbmi.sb_type) {
+                  case BLOCK_16X16:
+                    build_masks(cm, lfi_n, mip[0], shift_y_32_16,
+                                shift_uv_32_16, lfm);
+                    break;
+                  case BLOCK_16X8:
+#if CONFIG_SUPERTX
+                    if (supertx_enabled(&mip[0]->mbmi)) break;
+#endif
+                    build_masks(cm, lfi_n, mip[0], shift_y_32_16,
+                                shift_uv_32_16, lfm);
+                    if (mi_16_row_offset + 1 >= max_rows) continue;
+                    mip2 = mip + mode_info_stride;
+                    build_y_mask(cm, lfi_n, mip2[0], shift_y_32_16 + 8,
+#if CONFIG_SUPERTX
+                                 0,
+#endif
+                                 lfm);
+                    break;
+                  case BLOCK_8X16:
+#if CONFIG_SUPERTX
+                    if (supertx_enabled(&mip[0]->mbmi)) break;
+#endif
+                    build_masks(cm, lfi_n, mip[0], shift_y_32_16,
+                                shift_uv_32_16, lfm);
+                    if (mi_16_col_offset + 1 >= max_cols) continue;
+                    mip2 = mip + 1;
+                    build_y_mask(cm, lfi_n, mip2[0], shift_y_32_16 + 1,
+#if CONFIG_SUPERTX
+                                 0,
+#endif
+                                 lfm);
+                    break;
+                  default: {
+                    const int shift_y_32_16_8_zero =
+                        shift_y_32_16 + shift_8_y[0];
+#if CONFIG_SUPERTX
+                    if (mip[0]->mbmi.tx_size == TX_16X16) {
+                      build_masks(cm, lfi_n, mip[0], shift_y_32_16_8_zero,
+                                  shift_uv_32_16, lfm);
+                      break;
+                    }
+#endif
+                    build_masks(cm, lfi_n, mip[0], shift_y_32_16_8_zero,
+                                shift_uv_32_16, lfm);
+                    mip += offset[0];
+                    for (idx_8 = 1; idx_8 < 4; mip += offset[idx_8], ++idx_8) {
+                      const int shift_y_32_16_8 =
+                          shift_y_32_16 + shift_8_y[idx_8];
+                      const int mi_8_col_offset =
+                          mi_16_col_offset + ((idx_8 & 1));
+                      const int mi_8_row_offset =
+                          mi_16_row_offset + ((idx_8 >> 1));
+
+                      if (mi_8_col_offset >= max_cols ||
+                          mi_8_row_offset >= max_rows)
+                        continue;
+                      build_y_mask(cm, lfi_n, mip[0], shift_y_32_16_8,
+#if CONFIG_SUPERTX
+                                   supertx_enabled(&mip[0]->mbmi),
+#endif
+                                   lfm);
+                    }
+                    break;
+                  }
+                }
+              }
+              break;
+          }
+        }
+#if CONFIG_SUPERTX && CONFIG_TX64X64
+      }
+#endif  // CONFIG_SUPERTX && CONFIG_TX64X64
+      break;
+  }
+  // The largest loopfilter we have is 16x16 so we use the 16x16 mask
+  // for 32x32 transforms also.
+  lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32];
+  lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32];
+  lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32];
+  lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32];
+
+  // We do at least 8 tap filter on every 32x32 even if the transform size
+  // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and
+  // remove it from the 4x4.
+  lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border;
+  lfm->left_y[TX_4X4] &= ~left_border;
+  lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border;
+  lfm->above_y[TX_4X4] &= ~above_border;
+  lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv;
+  lfm->left_uv[TX_4X4] &= ~left_border_uv;
+  lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv;
+  lfm->above_uv[TX_4X4] &= ~above_border_uv;
+
+  // We do some special edge handling.
+  if (mi_row + MAX_MIB_SIZE > cm->mi_rows) {
+    const uint64_t rows = cm->mi_rows - mi_row;
+
+    // Each pixel inside the border gets a 1,
+    const uint64_t mask_y = (((uint64_t)1 << (rows << MAX_MIB_SIZE_LOG2)) - 1);
+    const uint16_t mask_uv =
+        (((uint16_t)1 << (((rows + 1) >> 1) << (MAX_MIB_SIZE_LOG2 - 1))) - 1);
+
+    // Remove values completely outside our border.
+    for (i = 0; i < TX_32X32; i++) {
+      lfm->left_y[i] &= mask_y;
+      lfm->above_y[i] &= mask_y;
+      lfm->left_uv[i] &= mask_uv;
+      lfm->above_uv[i] &= mask_uv;
+    }
+    lfm->int_4x4_y &= mask_y;
+    lfm->above_int_4x4_uv = lfm->left_int_4x4_uv & mask_uv;
+
+    // We don't apply a wide loop filter on the last uv block row. If set
+    // apply the shorter one instead.
+    if (rows == 1) {
+      lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16];
+      lfm->above_uv[TX_16X16] = 0;
+    }
+    if (rows == 5) {
+      lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00;
+      lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00);
+    }
+  } else {
+    lfm->above_int_4x4_uv = lfm->left_int_4x4_uv;
+  }
+
+  if (mi_col + MAX_MIB_SIZE > cm->mi_cols) {
+    const uint64_t columns = cm->mi_cols - mi_col;
+
+    // Each pixel inside the border gets a 1, the multiply copies the border
+    // to where we need it.
+    const uint64_t mask_y = (((1 << columns) - 1)) * 0x0101010101010101ULL;
+    const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111;
+
+    // Internal edges are not applied on the last column of the image so
+    // we mask 1 more for the internal edges
+    const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111;
+
+    // Remove the bits outside the image edge.
+    for (i = 0; i < TX_32X32; i++) {
+      lfm->left_y[i] &= mask_y;
+      lfm->above_y[i] &= mask_y;
+      lfm->left_uv[i] &= mask_uv;
+      lfm->above_uv[i] &= mask_uv;
+    }
+    lfm->int_4x4_y &= mask_y;
+    lfm->left_int_4x4_uv &= mask_uv_int;
+
+    // We don't apply a wide loop filter on the last uv column. If set
+    // apply the shorter one instead.
+    if (columns == 1) {
+      lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16];
+      lfm->left_uv[TX_16X16] = 0;
+    }
+    if (columns == 5) {
+      lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc);
+      lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc);
+    }
+  }
+  // We don't apply a loop filter on the first column in the image, mask that
+  // out.
+  if (mi_col == 0) {
+    for (i = 0; i < TX_32X32; i++) {
+      lfm->left_y[i] &= 0xfefefefefefefefeULL;
+      lfm->left_uv[i] &= 0xeeee;
+    }
+  }
+
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  if (av1_disable_loopfilter_on_tile_boundary(cm)) {
+    update_tile_boundary_filter_mask(cm, mi_row, mi_col, lfm);
+  }
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+  // Assert if we try to apply 2 different loop filters at the same position.
+  assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_8X8]));
+  assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_4X4]));
+  assert(!(lfm->left_y[TX_8X8] & lfm->left_y[TX_4X4]));
+  assert(!(lfm->int_4x4_y & lfm->left_y[TX_16X16]));
+  assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_8X8]));
+  assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_4X4]));
+  assert(!(lfm->left_uv[TX_8X8] & lfm->left_uv[TX_4X4]));
+  assert(!(lfm->left_int_4x4_uv & lfm->left_uv[TX_16X16]));
+  assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_8X8]));
+  assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_4X4]));
+  assert(!(lfm->above_y[TX_8X8] & lfm->above_y[TX_4X4]));
+  assert(!(lfm->int_4x4_y & lfm->above_y[TX_16X16]));
+  assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_8X8]));
+  assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_4X4]));
+  assert(!(lfm->above_uv[TX_8X8] & lfm->above_uv[TX_4X4]));
+  assert(!(lfm->above_int_4x4_uv & lfm->above_uv[TX_16X16]));
+}
+
+static void filter_selectively_vert(
+    uint8_t *s, int pitch, unsigned int mask_16x16, unsigned int mask_8x8,
+    unsigned int mask_4x4, unsigned int mask_4x4_int,
+    const loop_filter_info_n *lfi_n, const uint8_t *lfl) {
+  unsigned int mask;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; mask;
+       mask >>= 1) {
+    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
+
+    if (mask & 1) {
+      if (mask_16x16 & 1) {
+        aom_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+      } else if (mask_8x8 & 1) {
+        aom_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+      } else if (mask_4x4 & 1) {
+        aom_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+      }
+    }
+    if (mask_4x4_int & 1)
+      aom_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
+    s += 8;
+    lfl += 1;
+    mask_16x16 >>= 1;
+    mask_8x8 >>= 1;
+    mask_4x4 >>= 1;
+    mask_4x4_int >>= 1;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_filter_selectively_vert(
+    uint16_t *s, int pitch, unsigned int mask_16x16, unsigned int mask_8x8,
+    unsigned int mask_4x4, unsigned int mask_4x4_int,
+    const loop_filter_info_n *lfi_n, const uint8_t *lfl, int bd) {
+  unsigned int mask;
+
+  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; mask;
+       mask >>= 1) {
+    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;
+
+    if (mask & 1) {
+      if (mask_16x16 & 1) {
+        aom_highbd_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
+                                   bd);
+      } else if (mask_8x8 & 1) {
+        aom_highbd_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
+                                  bd);
+      } else if (mask_4x4 & 1) {
+        aom_highbd_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
+                                  bd);
+      }
+    }
+    if (mask_4x4_int & 1)
+      aom_highbd_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim,
+                                lfi->hev_thr, bd);
+    s += 8;
+    lfl += 1;
+    mask_16x16 >>= 1;
+    mask_8x8 >>= 1;
+    mask_4x4 >>= 1;
+    mask_4x4_int >>= 1;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+typedef struct {
+  unsigned int m16x16;
+  unsigned int m8x8;
+  unsigned int m4x4;
+} FilterMasks;
+
+// Get filter level and masks for the given row index 'idx_r'. (Only used for
+// the non420 case).
+// Note: 'row_masks_ptr' and/or 'col_masks_ptr' can be passed NULL.
+static void get_filter_level_and_masks_non420(
+    AV1_COMMON *const cm, const struct macroblockd_plane *const plane,
+    MODE_INFO **mib, int mi_row, int mi_col, int idx_r, uint8_t *const lfl_r,
+    unsigned int *const mask_4x4_int_r, FilterMasks *const row_masks_ptr,
+    FilterMasks *const col_masks_ptr) {
+  const int ss_x = plane->subsampling_x;
+  const int ss_y = plane->subsampling_y;
+  const int col_step = mi_size_wide[BLOCK_8X8] << ss_x;
+  FilterMasks row_masks, col_masks;
+  memset(&row_masks, 0, sizeof(row_masks));
+  memset(&col_masks, 0, sizeof(col_masks));
+  *mask_4x4_int_r = 0;
+  const int r = idx_r >> mi_height_log2_lookup[BLOCK_8X8];
+
+  // Determine the vertical edges that need filtering
+  int idx_c;
+  for (idx_c = 0; idx_c < cm->mib_size && mi_col + idx_c < cm->mi_cols;
+       idx_c += col_step) {
+    const MODE_INFO *mi = mib[idx_r * cm->mi_stride + idx_c];
+    const MB_MODE_INFO *mbmi = &mi[0].mbmi;
+    const BLOCK_SIZE sb_type = mbmi->sb_type;
+    const int skip_this = mbmi->skip && is_inter_block(mbmi);
+    // Map index to 8x8 unit
+    const int c = idx_c >> mi_width_log2_lookup[BLOCK_8X8];
+
+    const int blk_row = r & (num_8x8_blocks_high_lookup[sb_type] - 1);
+    const int blk_col = c & (num_8x8_blocks_wide_lookup[sb_type] - 1);
+
+    // left edge of current unit is block/partition edge -> no skip
+    const int block_edge_left =
+        (num_4x4_blocks_wide_lookup[sb_type] > 1) ? !blk_col : 1;
+    const int skip_this_c = skip_this && !block_edge_left;
+    // top edge of current unit is block/partition edge -> no skip
+    const int block_edge_above =
+        (num_4x4_blocks_high_lookup[sb_type] > 1) ? !blk_row : 1;
+    const int skip_this_r = skip_this && !block_edge_above;
+
+#if CONFIG_VAR_TX
+    const TX_SIZE mb_tx_size = mbmi->inter_tx_size[blk_row][blk_col];
+#endif
+
+    TX_SIZE tx_size = (plane->plane_type == PLANE_TYPE_UV)
+                          ? get_uv_tx_size(mbmi, plane)
+                          : mbmi->tx_size;
+
+    const int skip_border_4x4_c =
+        ss_x && mi_col + idx_c >= cm->mi_cols - mi_size_wide[BLOCK_8X8];
+    const int skip_border_4x4_r =
+        ss_y && mi_row + idx_r >= cm->mi_rows - mi_size_high[BLOCK_8X8];
+
+    TX_SIZE tx_size_c = txsize_horz_map[tx_size];
+    TX_SIZE tx_size_r = txsize_vert_map[tx_size];
+
+    int tx_size_mask = 0;
+    const int c_step = (c >> ss_x);
+    const int r_step = (r >> ss_y);
+    const int col_mask = 1 << c_step;
+
+#if CONFIG_VAR_TX
+    if (is_inter_block(mbmi) && !mbmi->skip) {
+      tx_size = (plane->plane_type == PLANE_TYPE_UV)
+                    ? uv_txsize_lookup[sb_type][mb_tx_size][ss_x][ss_y]
+                    : mb_tx_size;
+    }
+#endif
+
+// Filter level can vary per MI
+#if CONFIG_EXT_DELTA_Q
+    if (!(lfl_r[c_step] = get_filter_level(cm, &cm->lf_info, mbmi))) continue;
+#else
+    if (!(lfl_r[c_step] = get_filter_level(&cm->lf_info, mbmi))) continue;
+#endif
+
+#if CONFIG_VAR_TX
+    tx_size_r = AOMMIN(tx_size, cm->above_txfm_context[mi_col + c]);
+    tx_size_c =
+        AOMMIN(tx_size, cm->left_txfm_context[(mi_row + r) & MAX_MIB_MASK]);
+
+    cm->above_txfm_context[mi_col + c] = tx_size;
+    cm->left_txfm_context[(mi_row + r) & MAX_MIB_MASK] = tx_size;
+#endif  // CONFIG_VAR_TX
+
+    if (tx_size_c == TX_32X32)
+      tx_size_mask = 3;
+    else if (tx_size_c == TX_16X16)
+      tx_size_mask = 1;
+    else
+      tx_size_mask = 0;
+
+    // Build masks based on the transform size of each block
+    // handle vertical mask
+    if (tx_size_c == TX_32X32) {
+      if (!skip_this_c && (c_step & tx_size_mask) == 0) {
+        if (!skip_border_4x4_c)
+          col_masks.m16x16 |= col_mask;
+        else
+          col_masks.m8x8 |= col_mask;
+      }
+    } else if (tx_size_c == TX_16X16) {
+      if (!skip_this_c && (c_step & tx_size_mask) == 0) {
+        if (!skip_border_4x4_c)
+          col_masks.m16x16 |= col_mask;
+        else
+          col_masks.m8x8 |= col_mask;
+      }
+    } else {
+      // force 8x8 filtering on 32x32 boundaries
+      if (!skip_this_c && (c_step & tx_size_mask) == 0) {
+        if (tx_size_c == TX_8X8 || ((c >> ss_x) & 3) == 0)
+          col_masks.m8x8 |= col_mask;
+        else
+          col_masks.m4x4 |= col_mask;
+      }
+
+      if (!skip_this && tx_size_c < TX_8X8 && !skip_border_4x4_c &&
+          (c_step & tx_size_mask) == 0)
+        *mask_4x4_int_r |= col_mask;
+    }
+
+    if (tx_size_r == TX_32X32)
+      tx_size_mask = 3;
+    else if (tx_size_r == TX_16X16)
+      tx_size_mask = 1;
+    else
+      tx_size_mask = 0;
+
+    // set horizontal mask
+    if (tx_size_r == TX_32X32) {
+      if (!skip_this_r && (r_step & tx_size_mask) == 0) {
+        if (!skip_border_4x4_r)
+          row_masks.m16x16 |= col_mask;
+        else
+          row_masks.m8x8 |= col_mask;
+      }
+    } else if (tx_size_r == TX_16X16) {
+      if (!skip_this_r && (r_step & tx_size_mask) == 0) {
+        if (!skip_border_4x4_r)
+          row_masks.m16x16 |= col_mask;
+        else
+          row_masks.m8x8 |= col_mask;
+      }
+    } else {
+      // force 8x8 filtering on 32x32 boundaries
+      if (!skip_this_r && (r_step & tx_size_mask) == 0) {
+        if (tx_size_r == TX_8X8 || (r_step & 3) == 0)
+          row_masks.m8x8 |= col_mask;
+        else
+          row_masks.m4x4 |= col_mask;
+      }
+
+      if (!skip_this && tx_size_r < TX_8X8 && !skip_border_4x4_c &&
+          ((r >> ss_y) & tx_size_mask) == 0)
+        *mask_4x4_int_r |= col_mask;
+    }
+  }
+
+  if (row_masks_ptr) *row_masks_ptr = row_masks;
+  if (col_masks_ptr) *col_masks_ptr = col_masks;
+}
+
+void av1_filter_block_plane_non420_ver(AV1_COMMON *const cm,
+                                       struct macroblockd_plane *plane,
+                                       MODE_INFO **mib, int mi_row,
+                                       int mi_col) {
+  const int ss_y = plane->subsampling_y;
+  const int row_step = mi_size_high[BLOCK_8X8] << ss_y;
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t *const dst0 = dst->buf;
+  uint8_t lfl[MAX_MIB_SIZE][MAX_MIB_SIZE] = { { 0 } };
+
+  int idx_r;
+  for (idx_r = 0; idx_r < cm->mib_size && mi_row + idx_r < cm->mi_rows;
+       idx_r += row_step) {
+    unsigned int mask_4x4_int;
+    FilterMasks col_masks;
+    const int r = idx_r >> mi_height_log2_lookup[BLOCK_8X8];
+    get_filter_level_and_masks_non420(cm, plane, mib, mi_row, mi_col, idx_r,
+                                      &lfl[r][0], &mask_4x4_int, NULL,
+                                      &col_masks);
+
+    // Disable filtering on the leftmost column or tile boundary
+    unsigned int border_mask = ~(mi_col == 0);
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+    if (av1_disable_loopfilter_on_tile_boundary(cm) &&
+        ((mib[0]->mbmi.boundary_info & TILE_LEFT_BOUNDARY) != 0)) {
+      border_mask = 0xfffffffe;
+    }
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+#if CONFIG_HIGHBITDEPTH
+    if (cm->use_highbitdepth)
+      highbd_filter_selectively_vert(
+          CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
+          col_masks.m16x16 & border_mask, col_masks.m8x8 & border_mask,
+          col_masks.m4x4 & border_mask, mask_4x4_int, &cm->lf_info, &lfl[r][0],
+          (int)cm->bit_depth);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      filter_selectively_vert(
+          dst->buf, dst->stride, col_masks.m16x16 & border_mask,
+          col_masks.m8x8 & border_mask, col_masks.m4x4 & border_mask,
+          mask_4x4_int, &cm->lf_info, &lfl[r][0]);
+    dst->buf += 8 * dst->stride;
+  }
+
+  // Now do horizontal pass
+  dst->buf = dst0;
+}
+
+void av1_filter_block_plane_non420_hor(AV1_COMMON *const cm,
+                                       struct macroblockd_plane *plane,
+                                       MODE_INFO **mib, int mi_row,
+                                       int mi_col) {
+  const int ss_y = plane->subsampling_y;
+  const int row_step = mi_size_high[BLOCK_8X8] << ss_y;
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t *const dst0 = dst->buf;
+  FilterMasks row_masks_array[MAX_MIB_SIZE];
+  unsigned int mask_4x4_int[MAX_MIB_SIZE] = { 0 };
+  uint8_t lfl[MAX_MIB_SIZE][MAX_MIB_SIZE] = { { 0 } };
+  int idx_r;
+  for (idx_r = 0; idx_r < cm->mib_size && mi_row + idx_r < cm->mi_rows;
+       idx_r += row_step) {
+    const int r = idx_r >> mi_height_log2_lookup[BLOCK_8X8];
+    get_filter_level_and_masks_non420(cm, plane, mib, mi_row, mi_col, idx_r,
+                                      &lfl[r][0], mask_4x4_int + r,
+                                      row_masks_array + r, NULL);
+  }
+  for (idx_r = 0; idx_r < cm->mib_size && mi_row + idx_r < cm->mi_rows;
+       idx_r += row_step) {
+    const int skip_border_4x4_r =
+        ss_y && mi_row + idx_r >= cm->mi_rows - mi_size_wide[BLOCK_8X8];
+    const int r = idx_r >> mi_width_log2_lookup[BLOCK_8X8];
+    const unsigned int mask_4x4_int_r = skip_border_4x4_r ? 0 : mask_4x4_int[r];
+    FilterMasks row_masks;
+
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+    // Disable filtering on the abovemost row or tile boundary
+    const MODE_INFO *mi = cm->mi + (mi_row + r) * cm->mi_stride;
+    if ((av1_disable_loopfilter_on_tile_boundary(cm) &&
+         (mi->mbmi.boundary_info & TILE_ABOVE_BOUNDARY)) ||
+        (mi_row + idx_r == 0)) {
+      memset(&row_masks, 0, sizeof(row_masks));
+#else
+    if (mi_row + idx_r == 0) {
+      memset(&row_masks, 0, sizeof(row_masks));
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+    } else {
+      memcpy(&row_masks, row_masks_array + r, sizeof(row_masks));
+    }
+#if CONFIG_HIGHBITDEPTH
+    if (cm->use_highbitdepth)
+      highbd_filter_selectively_horiz(
+          CONVERT_TO_SHORTPTR(dst->buf), dst->stride, row_masks.m16x16,
+          row_masks.m8x8, row_masks.m4x4, mask_4x4_int_r, &cm->lf_info,
+          &lfl[r][0], (int)cm->bit_depth);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      filter_selectively_horiz(dst->buf, dst->stride, row_masks.m16x16,
+                               row_masks.m8x8, row_masks.m4x4, mask_4x4_int_r,
+                               &cm->lf_info, &lfl[r][0]);
+    dst->buf += 8 * dst->stride;
+  }
+  dst->buf = dst0;
+}
+
+void av1_filter_block_plane_ss00_ver(AV1_COMMON *const cm,
+                                     struct macroblockd_plane *const plane,
+                                     int mi_row, LOOP_FILTER_MASK *lfm) {
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t *const dst0 = dst->buf;
+  int r;
+  uint64_t mask_16x16 = lfm->left_y[TX_16X16];
+  uint64_t mask_8x8 = lfm->left_y[TX_8X8];
+  uint64_t mask_4x4 = lfm->left_y[TX_4X4];
+  uint64_t mask_4x4_int = lfm->int_4x4_y;
+
+  assert(plane->subsampling_x == 0 && plane->subsampling_y == 0);
+
+  // Vertical pass: do 2 rows at one time
+  for (r = 0; r < cm->mib_size && mi_row + r < cm->mi_rows; r += 2) {
+    unsigned int mask_16x16_l = mask_16x16 & 0xffff;
+    unsigned int mask_8x8_l = mask_8x8 & 0xffff;
+    unsigned int mask_4x4_l = mask_4x4 & 0xffff;
+    unsigned int mask_4x4_int_l = mask_4x4_int & 0xffff;
+
+// Disable filtering on the leftmost column.
+#if CONFIG_HIGHBITDEPTH
+    if (cm->use_highbitdepth)
+      highbd_filter_selectively_vert_row2(
+          plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
+          mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
+          &lfm->lfl_y[r][0], (int)cm->bit_depth);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      filter_selectively_vert_row2(
+          plane->subsampling_x, dst->buf, dst->stride, mask_16x16_l, mask_8x8_l,
+          mask_4x4_l, mask_4x4_int_l, &cm->lf_info, &lfm->lfl_y[r][0]);
+
+    dst->buf += 2 * MI_SIZE * dst->stride;
+    mask_16x16 >>= 2 * MI_SIZE;
+    mask_8x8 >>= 2 * MI_SIZE;
+    mask_4x4 >>= 2 * MI_SIZE;
+    mask_4x4_int >>= 2 * MI_SIZE;
+  }
+
+  // Horizontal pass
+  dst->buf = dst0;
+}
+
+void av1_filter_block_plane_ss00_hor(AV1_COMMON *const cm,
+                                     struct macroblockd_plane *const plane,
+                                     int mi_row, LOOP_FILTER_MASK *lfm) {
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t *const dst0 = dst->buf;
+  int r;
+  uint64_t mask_16x16 = lfm->above_y[TX_16X16];
+  uint64_t mask_8x8 = lfm->above_y[TX_8X8];
+  uint64_t mask_4x4 = lfm->above_y[TX_4X4];
+  uint64_t mask_4x4_int = lfm->int_4x4_y;
+
+  assert(plane->subsampling_x == 0 && plane->subsampling_y == 0);
+
+  for (r = 0; r < cm->mib_size && mi_row + r < cm->mi_rows; r++) {
+    unsigned int mask_16x16_r;
+    unsigned int mask_8x8_r;
+    unsigned int mask_4x4_r;
+
+    if (mi_row + r == 0) {
+      mask_16x16_r = 0;
+      mask_8x8_r = 0;
+      mask_4x4_r = 0;
+    } else {
+      mask_16x16_r = mask_16x16 & 0xff;
+      mask_8x8_r = mask_8x8 & 0xff;
+      mask_4x4_r = mask_4x4 & 0xff;
+    }
+
+#if CONFIG_HIGHBITDEPTH
+    if (cm->use_highbitdepth)
+      highbd_filter_selectively_horiz(
+          CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_r, mask_8x8_r,
+          mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info, &lfm->lfl_y[r][0],
+          (int)cm->bit_depth);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
+                               mask_4x4_r, mask_4x4_int & 0xff, &cm->lf_info,
+                               &lfm->lfl_y[r][0]);
+
+    dst->buf += MI_SIZE * dst->stride;
+    mask_16x16 >>= MI_SIZE;
+    mask_8x8 >>= MI_SIZE;
+    mask_4x4 >>= MI_SIZE;
+    mask_4x4_int >>= MI_SIZE;
+  }
+  // restore the buf pointer in case there is additional filter pass.
+  dst->buf = dst0;
+}
+
+void av1_filter_block_plane_ss11_ver(AV1_COMMON *const cm,
+                                     struct macroblockd_plane *const plane,
+                                     int mi_row, LOOP_FILTER_MASK *lfm) {
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t *const dst0 = dst->buf;
+  int r, c;
+
+  uint16_t mask_16x16 = lfm->left_uv[TX_16X16];
+  uint16_t mask_8x8 = lfm->left_uv[TX_8X8];
+  uint16_t mask_4x4 = lfm->left_uv[TX_4X4];
+  uint16_t mask_4x4_int = lfm->left_int_4x4_uv;
+
+  assert(plane->subsampling_x == 1 && plane->subsampling_y == 1);
+  assert(plane->plane_type == PLANE_TYPE_UV);
+  memset(lfm->lfl_uv, 0, sizeof(lfm->lfl_uv));
+
+  // Vertical pass: do 2 rows at one time
+  for (r = 0; r < cm->mib_size && mi_row + r < cm->mi_rows; r += 4) {
+    for (c = 0; c < (cm->mib_size >> 1); c++) {
+      lfm->lfl_uv[r >> 1][c] = lfm->lfl_y[r][c << 1];
+      lfm->lfl_uv[(r + 2) >> 1][c] = lfm->lfl_y[r + 2][c << 1];
+    }
+
+    {
+      unsigned int mask_16x16_l = mask_16x16 & 0xff;
+      unsigned int mask_8x8_l = mask_8x8 & 0xff;
+      unsigned int mask_4x4_l = mask_4x4 & 0xff;
+      unsigned int mask_4x4_int_l = mask_4x4_int & 0xff;
+
+// Disable filtering on the leftmost column.
+#if CONFIG_HIGHBITDEPTH
+      if (cm->use_highbitdepth)
+        highbd_filter_selectively_vert_row2(
+            plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
+            mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
+            &lfm->lfl_uv[r >> 1][0], (int)cm->bit_depth);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        filter_selectively_vert_row2(plane->subsampling_x, dst->buf,
+                                     dst->stride, mask_16x16_l, mask_8x8_l,
+                                     mask_4x4_l, mask_4x4_int_l, &cm->lf_info,
+                                     &lfm->lfl_uv[r >> 1][0]);
+
+      dst->buf += 2 * MI_SIZE * dst->stride;
+      mask_16x16 >>= MI_SIZE;
+      mask_8x8 >>= MI_SIZE;
+      mask_4x4 >>= MI_SIZE;
+      mask_4x4_int >>= MI_SIZE;
+    }
+  }
+
+  // Horizontal pass
+  dst->buf = dst0;
+}
+
+void av1_filter_block_plane_ss11_hor(AV1_COMMON *const cm,
+                                     struct macroblockd_plane *const plane,
+                                     int mi_row, LOOP_FILTER_MASK *lfm) {
+  struct buf_2d *const dst = &plane->dst;
+  uint8_t *const dst0 = dst->buf;
+  int r, c;
+  uint64_t mask_16x16 = lfm->above_uv[TX_16X16];
+  uint64_t mask_8x8 = lfm->above_uv[TX_8X8];
+  uint64_t mask_4x4 = lfm->above_uv[TX_4X4];
+  uint64_t mask_4x4_int = lfm->above_int_4x4_uv;
+
+  assert(plane->subsampling_x == 1 && plane->subsampling_y == 1);
+  memset(lfm->lfl_uv, 0, sizeof(lfm->lfl_uv));
+
+  // re-porpulate the filter level for uv, same as the code for vertical
+  // filter in av1_filter_block_plane_ss11_ver
+  for (r = 0; r < cm->mib_size && mi_row + r < cm->mi_rows; r += 4) {
+    for (c = 0; c < (cm->mib_size >> 1); c++) {
+      lfm->lfl_uv[r >> 1][c] = lfm->lfl_y[r][c << 1];
+      lfm->lfl_uv[(r + 2) >> 1][c] = lfm->lfl_y[r + 2][c << 1];
+    }
+  }
+
+  for (r = 0; r < cm->mib_size && mi_row + r < cm->mi_rows; r += 2) {
+    const int skip_border_4x4_r = mi_row + r == cm->mi_rows - 1;
+    const unsigned int mask_4x4_int_r =
+        skip_border_4x4_r ? 0 : (mask_4x4_int & 0xf);
+    unsigned int mask_16x16_r;
+    unsigned int mask_8x8_r;
+    unsigned int mask_4x4_r;
+
+    if (mi_row + r == 0) {
+      mask_16x16_r = 0;
+      mask_8x8_r = 0;
+      mask_4x4_r = 0;
+    } else {
+      mask_16x16_r = mask_16x16 & 0xf;
+      mask_8x8_r = mask_8x8 & 0xf;
+      mask_4x4_r = mask_4x4 & 0xf;
+    }
+
+#if CONFIG_HIGHBITDEPTH
+    if (cm->use_highbitdepth)
+      highbd_filter_selectively_horiz(
+          CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_r, mask_8x8_r,
+          mask_4x4_r, mask_4x4_int_r, &cm->lf_info, &lfm->lfl_uv[r >> 1][0],
+          (int)cm->bit_depth);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
+                               mask_4x4_r, mask_4x4_int_r, &cm->lf_info,
+                               &lfm->lfl_uv[r >> 1][0]);
+
+    dst->buf += MI_SIZE * dst->stride;
+    mask_16x16 >>= MI_SIZE / 2;
+    mask_8x8 >>= MI_SIZE / 2;
+    mask_4x4 >>= MI_SIZE / 2;
+    mask_4x4_int >>= MI_SIZE / 2;
+  }
+  // restore the buf pointer in case there is additional filter pass.
+  dst->buf = dst0;
+}
+
+#if !(CONFIG_VAR_TX || CONFIG_EXT_PARTITION || CONFIG_EXT_PARTITION_TYPES || \
+      CONFIG_CB4X4)
+#if CONFIG_PARALLEL_DEBLOCKING
+typedef enum EDGE_DIR { VERT_EDGE = 0, HORZ_EDGE = 1, NUM_EDGE_DIRS } EDGE_DIR;
+static const uint32_t av1_prediction_masks[NUM_EDGE_DIRS][BLOCK_SIZES] = {
+  // mask for vertical edges filtering
+  {
+#if CONFIG_CB4X4
+      2 - 1,   // BLOCK_2X2
+      2 - 1,   // BLOCK_2X4
+      4 - 1,   // BLOCK_4X2
+#endif         // CONFIG_CB4X4
+      4 - 1,   // BLOCK_4X4
+      4 - 1,   // BLOCK_4X8
+      8 - 1,   // BLOCK_8X4
+      8 - 1,   // BLOCK_8X8
+      8 - 1,   // BLOCK_8X16
+      16 - 1,  // BLOCK_16X8
+      16 - 1,  // BLOCK_16X16
+      16 - 1,  // BLOCK_16X32
+      32 - 1,  // BLOCK_32X16
+      32 - 1,  // BLOCK_32X32
+      32 - 1,  // BLOCK_32X64
+      64 - 1,  // BLOCK_64X32
+      64 - 1,  // BLOCK_64X64
+#if CONFIG_EXT_PARTITION
+      64 - 1,   // BLOCK_64X128
+      128 - 1,  // BLOCK_128X64
+      128 - 1   // BLOCK_128X128
+#endif          // CONFIG_EXT_PARTITION
+  },
+  // mask for horizontal edges filtering
+  {
+#if CONFIG_CB4X4
+      2 - 1,   // BLOCK_2X2
+      4 - 1,   // BLOCK_2X4
+      2 - 1,   // BLOCK_4X2
+#endif         // CONFIG_CB4X4
+      4 - 1,   // BLOCK_4X4
+      8 - 1,   // BLOCK_4X8
+      4 - 1,   // BLOCK_8X4
+      8 - 1,   // BLOCK_8X8
+      16 - 1,  // BLOCK_8X16
+      8 - 1,   // BLOCK_16X8
+      16 - 1,  // BLOCK_16X16
+      32 - 1,  // BLOCK_16X32
+      16 - 1,  // BLOCK_32X16
+      32 - 1,  // BLOCK_32X32
+      64 - 1,  // BLOCK_32X64
+      32 - 1,  // BLOCK_64X32
+      64 - 1,  // BLOCK_64X64
+#if CONFIG_EXT_PARTITION
+      128 - 1,  // BLOCK_64X128
+      64 - 1,   // BLOCK_128X64
+      128 - 1   // BLOCK_128X128
+#endif          // CONFIG_EXT_PARTITION
+  },
+};
+
+static const uint32_t av1_transform_masks[NUM_EDGE_DIRS][TX_SIZES_ALL] = {
+  {
+#if CONFIG_CB4X4
+      2 - 1,  // TX_2X2
+#endif
+      4 - 1,   // TX_4X4
+      8 - 1,   // TX_8X8
+      16 - 1,  // TX_16X16
+      32 - 1,  // TX_32X32
+#if CONFIG_TX64X64
+      64 - 1,  // TX_64X64
+#endif         // CONFIG_TX64X64
+      4 - 1,   // TX_4X8
+      8 - 1,   // TX_8X4
+      8 - 1,   // TX_8X16
+      16 - 1,  // TX_16X8
+      16 - 1,  // TX_16X32
+      32 - 1,  // TX_32X16
+      4 - 1,   // TX_4X16
+      16 - 1,  // TX_16X4
+      8 - 1,   // TX_8X32
+      32 - 1   // TX_32X8
+  },
+  {
+#if CONFIG_CB4X4
+      2 - 1,  // TX_2X2
+#endif
+      4 - 1,   // TX_4X4
+      8 - 1,   // TX_8X8
+      16 - 1,  // TX_16X16
+      32 - 1,  // TX_32X32
+#if CONFIG_TX64X64
+      64 - 1,  // TX_64X64
+#endif         // CONFIG_TX64X64
+      8 - 1,   // TX_4X8
+      4 - 1,   // TX_8X4
+      16 - 1,  // TX_8X16
+      8 - 1,   // TX_16X8
+      32 - 1,  // TX_16X32
+      16 - 1,  // TX_32X16
+      16 - 1,  // TX_4X16
+      4 - 1,   // TX_16X4
+      32 - 1,  // TX_8X32
+      8 - 1    // TX_32X8
+  }
+};
+
+static TX_SIZE av1_get_transform_size(const MODE_INFO *const pCurr,
+                                      const EDGE_DIR edgeDir,
+                                      const uint32_t scaleHorz,
+                                      const uint32_t scaleVert) {
+  const BLOCK_SIZE bs = pCurr->mbmi.sb_type;
+  TX_SIZE txSize;
+  // since in case of chrominance or non-square transorm need to convert
+  // transform size into transform size in particular direction.
+  txSize = uv_txsize_lookup[bs][pCurr->mbmi.tx_size][scaleHorz][scaleVert];
+  if (VERT_EDGE == edgeDir) {
+    txSize = txsize_horz_map[txSize];
+  } else {
+    txSize = txsize_vert_map[txSize];
+  }
+  return txSize;
+}
+
+typedef struct AV1_DEBLOCKING_PARAMETERS {
+  // length of the filter applied to the outer edge
+  uint32_t filterLength;
+  // length of the filter applied to the inner edge
+  uint32_t filterLengthInternal;
+  // deblocking limits
+  const uint8_t *lim;
+  const uint8_t *mblim;
+  const uint8_t *hev_thr;
+} AV1_DEBLOCKING_PARAMETERS;
+
+static void set_lpf_parameters(AV1_DEBLOCKING_PARAMETERS *const pParams,
+                               const MODE_INFO **const ppCurr,
+                               const ptrdiff_t modeStep,
+                               const AV1_COMMON *const cm,
+                               const EDGE_DIR edgeDir, const uint32_t x,
+                               const uint32_t y, const uint32_t width,
+                               const uint32_t height, const uint32_t scaleHorz,
+                               const uint32_t scaleVert) {
+  // reset to initial values
+  pParams->filterLength = 0;
+  pParams->filterLengthInternal = 0;
+  // no deblocking is required
+  if ((width <= x) || (height <= y)) {
+    return;
+  }
+#if CONFIG_EXT_PARTITION
+  // not sure if changes are required.
+  assert(0 && "Not yet updated");
+#endif  // CONFIG_EXT_PARTITION
+  {
+    const TX_SIZE ts =
+        av1_get_transform_size(ppCurr[0], edgeDir, scaleHorz, scaleVert);
+    const uint32_t currLevel = get_filter_level(&cm->lf_info, &ppCurr[0]->mbmi);
+    const int currSkipped =
+        ppCurr[0]->mbmi.skip && is_inter_block(&ppCurr[0]->mbmi);
+    const uint32_t coord = (VERT_EDGE == edgeDir) ? (x) : (y);
+    uint32_t level = currLevel;
+    // prepare outer edge parameters. deblock the edge if it's an edge of a TU
+    if (coord) {
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+      if (!av1_disable_loopfilter_on_tile_boundary(cm) ||
+          ((VERT_EDGE == edgeDir) &&
+           (0 == (ppCurr[0]->mbmi.boundary_info & TILE_LEFT_BOUNDARY))) ||
+          ((HORZ_EDGE == edgeDir) &&
+           (0 == (ppCurr[0]->mbmi.boundary_info & TILE_ABOVE_BOUNDARY))))
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+      {
+        const int32_t tuEdge =
+            (coord & av1_transform_masks[edgeDir][ts]) ? (0) : (1);
+        if (tuEdge) {
+          const MODE_INFO *const pPrev = *(ppCurr - modeStep);
+          const TX_SIZE pvTs =
+              av1_get_transform_size(pPrev, edgeDir, scaleHorz, scaleVert);
+          const uint32_t pvLvl = get_filter_level(&cm->lf_info, &pPrev->mbmi);
+          const int pvSkip = pPrev->mbmi.skip && is_inter_block(&pPrev->mbmi);
+          const int32_t puEdge =
+              (coord &
+               av1_prediction_masks[edgeDir]
+                                   [ss_size_lookup[ppCurr[0]->mbmi.sb_type]
+                                                  [scaleHorz][scaleVert]])
+                  ? (0)
+                  : (1);
+          // if the current and the previous blocks are skipped,
+          // deblock the edge if the edge belongs to a PU's edge only.
+          if ((currLevel || pvLvl) && (!pvSkip || !currSkipped || puEdge)) {
+#if CONFIG_PARALLEL_DEBLOCKING_15TAP || CONFIG_PARALLEL_DEBLOCKING_15TAPLUMAONLY
+            const TX_SIZE minTs = AOMMIN(ts, pvTs);
+            if (TX_4X4 >= minTs) {
+              pParams->filterLength = 4;
+            } else if (TX_8X8 == minTs) {
+              pParams->filterLength = 8;
+            } else {
+              pParams->filterLength = 16;
+#if CONFIG_PARALLEL_DEBLOCKING_15TAPLUMAONLY
+              // No wide filtering for chroma plane
+              if (scaleHorz || scaleVert) {
+                pParams->filterLength = 8;
+              }
+#endif
+            }
+#else
+            pParams->filterLength = (TX_4X4 >= AOMMIN(ts, pvTs)) ? (4) : (8);
+
+#endif  // CONFIG_PARALLEL_DEBLOCKING_15TAP
+
+            // update the level if the current block is skipped,
+            // but the previous one is not
+            level = (currLevel) ? (currLevel) : (pvLvl);
+          }
+        }
+      }
+      // prepare internal edge parameters
+      if (currLevel && !currSkipped) {
+        pParams->filterLengthInternal = (TX_4X4 >= ts) ? (4) : (0);
+      }
+      // prepare common parameters
+      if (pParams->filterLength || pParams->filterLengthInternal) {
+        const loop_filter_thresh *const limits = cm->lf_info.lfthr + level;
+        pParams->lim = limits->lim;
+        pParams->mblim = limits->mblim;
+        pParams->hev_thr = limits->hev_thr;
+      }
+    }
+  }
+}
+
+static void av1_filter_block_plane_vert(const AV1_COMMON *const cm,
+                                        const MACROBLOCKD_PLANE *const pPlane,
+                                        const MODE_INFO **ppModeInfo,
+                                        const ptrdiff_t modeStride,
+                                        const uint32_t cuX,
+                                        const uint32_t cuY) {
+  const uint32_t scaleHorz = pPlane->subsampling_x;
+  const uint32_t scaleVert = pPlane->subsampling_y;
+  const uint32_t width = pPlane->dst.width;
+  const uint32_t height = pPlane->dst.height;
+  uint8_t *const pDst = pPlane->dst.buf;
+  const int dstStride = pPlane->dst.stride;
+  for (int y = 0; y < (MAX_MIB_SIZE >> scaleVert); y += 1) {
+    uint8_t *p = pDst + y * MI_SIZE * dstStride;
+    for (int x = 0; x < (MAX_MIB_SIZE >> scaleHorz); x += 1) {
+      const MODE_INFO **const pCurr =
+          ppModeInfo + (y << scaleVert) * modeStride + (x << scaleHorz);
+      AV1_DEBLOCKING_PARAMETERS params;
+      memset(&params, 0, sizeof(params));
+      set_lpf_parameters(&params, pCurr, ((ptrdiff_t)1 << scaleHorz), cm,
+                         VERT_EDGE, cuX + x * MI_SIZE, cuY + y * MI_SIZE, width,
+                         height, scaleHorz, scaleVert);
+      switch (params.filterLength) {
+        // apply 4-tap filtering
+        case 4:
+          aom_lpf_vertical_4(p, dstStride, params.mblim, params.lim,
+                             params.hev_thr);
+          break;
+        // apply 8-tap filtering
+        case 8:
+          aom_lpf_vertical_8(p, dstStride, params.mblim, params.lim,
+                             params.hev_thr);
+          break;
+#if CONFIG_PARALLEL_DEBLOCKING_15TAP || CONFIG_PARALLEL_DEBLOCKING_15TAPLUMAONLY
+        // apply 16-tap filtering
+        case 16:
+          aom_lpf_vertical_16(p, dstStride, params.mblim, params.lim,
+                              params.hev_thr);
+          break;
+#endif  // CONFIG_PARALLEL_DEBLOCKING_15TAP
+        // no filtering
+        default: break;
+      }
+      // process the internal edge
+      if (params.filterLengthInternal) {
+        aom_lpf_vertical_4(p + 4, dstStride, params.mblim, params.lim,
+                           params.hev_thr);
+      }
+      // advance the destination pointer
+      p += 8;
+    }
+  }
+}
+
+static void av1_filter_block_plane_horz(const AV1_COMMON *const cm,
+                                        const MACROBLOCKD_PLANE *const pPlane,
+                                        const MODE_INFO **ppModeInfo,
+                                        const ptrdiff_t modeStride,
+                                        const uint32_t cuX,
+                                        const uint32_t cuY) {
+  const uint32_t scaleHorz = pPlane->subsampling_x;
+  const uint32_t scaleVert = pPlane->subsampling_y;
+  const uint32_t width = pPlane->dst.width;
+  const uint32_t height = pPlane->dst.height;
+  uint8_t *const pDst = pPlane->dst.buf;
+  const int dstStride = pPlane->dst.stride;
+  for (int y = 0; y < (MAX_MIB_SIZE >> scaleVert); y += 1) {
+    uint8_t *p = pDst + y * MI_SIZE * dstStride;
+    for (int x = 0; x < (MAX_MIB_SIZE >> scaleHorz); x += 1) {
+      const MODE_INFO **const pCurr =
+          ppModeInfo + (y << scaleVert) * modeStride + (x << scaleHorz);
+      AV1_DEBLOCKING_PARAMETERS params;
+      memset(&params, 0, sizeof(params));
+      set_lpf_parameters(&params, pCurr, (modeStride << scaleVert), cm,
+                         HORZ_EDGE, cuX + x * MI_SIZE, cuY + y * MI_SIZE, width,
+                         height, scaleHorz, scaleVert);
+      switch (params.filterLength) {
+        // apply 4-tap filtering
+        case 4:
+          aom_lpf_horizontal_4(p, dstStride, params.mblim, params.lim,
+                               params.hev_thr);
+          break;
+        // apply 8-tap filtering
+        case 8:
+          aom_lpf_horizontal_8(p, dstStride, params.mblim, params.lim,
+                               params.hev_thr);
+          break;
+#if CONFIG_PARALLEL_DEBLOCKING_15TAP || CONFIG_PARALLEL_DEBLOCKING_15TAPLUMAONLY
+        // apply 16-tap filtering
+        case 16:
+          aom_lpf_horizontal_edge_16(p, dstStride, params.mblim, params.lim,
+                                     params.hev_thr);
+          break;
+#endif  // CONFIG_PARALLEL_DEBLOCKING_15TAP
+        // no filtering
+        default: break;
+      }
+      // process the internal edge
+      if (params.filterLengthInternal) {
+        aom_lpf_horizontal_4(p + 4 * dstStride, dstStride, params.mblim,
+                             params.lim, params.hev_thr);
+      }
+      // advance the destination pointer
+      p += 8;
+    }
+  }
+}
+#endif  // CONFIG_PARALLEL_DEBLOCKING
+#endif
+
+void av1_loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer, AV1_COMMON *cm,
+                          struct macroblockd_plane planes[MAX_MB_PLANE],
+                          int start, int stop, int y_only) {
+#if CONFIG_VAR_TX || CONFIG_EXT_PARTITION || CONFIG_EXT_PARTITION_TYPES || \
+    CONFIG_CB4X4
+  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
+  int mi_row, mi_col;
+
+#if CONFIG_VAR_TX
+  memset(cm->above_txfm_context, TX_SIZES, cm->mi_cols);
+#endif  // CONFIG_VAR_TX
+  for (mi_row = start; mi_row < stop; mi_row += cm->mib_size) {
+    MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+#if CONFIG_VAR_TX
+    memset(cm->left_txfm_context, TX_SIZES, MAX_MIB_SIZE);
+#endif  // CONFIG_VAR_TX
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += cm->mib_size) {
+      int plane;
+
+      av1_setup_dst_planes(planes, cm->sb_size, frame_buffer, mi_row, mi_col);
+
+      for (plane = 0; plane < num_planes; ++plane) {
+        av1_filter_block_plane_non420_ver(cm, &planes[plane], mi + mi_col,
+                                          mi_row, mi_col);
+        av1_filter_block_plane_non420_hor(cm, &planes[plane], mi + mi_col,
+                                          mi_row, mi_col);
+      }
+    }
+  }
+#else  // CONFIG_VAR_TX || CONFIG_EXT_PARTITION || CONFIG_EXT_PARTITION_TYPES
+  const int num_planes = y_only ? 1 : MAX_MB_PLANE;
+  int mi_row, mi_col;
+#if !CONFIG_PARALLEL_DEBLOCKING
+  enum lf_path path;
+  LOOP_FILTER_MASK lfm;
+
+  if (y_only)
+    path = LF_PATH_444;
+  else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
+    path = LF_PATH_420;
+  else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
+    path = LF_PATH_444;
+  else
+    path = LF_PATH_SLOW;
+#endif
+#if CONFIG_PARALLEL_DEBLOCKING
+  for (mi_row = start; mi_row < stop; mi_row += MAX_MIB_SIZE) {
+    MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MAX_MIB_SIZE) {
+      av1_setup_dst_planes(planes, cm->sb_size, frame_buffer, mi_row, mi_col);
+      for (int planeIdx = 0; planeIdx < num_planes; planeIdx += 1) {
+        const int32_t scaleHorz = planes[planeIdx].subsampling_x;
+        const int32_t scaleVert = planes[planeIdx].subsampling_y;
+        av1_filter_block_plane_vert(
+            cm, planes + planeIdx, (const MODE_INFO **)(mi + mi_col),
+            cm->mi_stride, (mi_col * MI_SIZE) >> scaleHorz,
+            (mi_row * MI_SIZE) >> scaleVert);
+      }
+    }
+  }
+  for (mi_row = start; mi_row < stop; mi_row += MAX_MIB_SIZE) {
+    MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MAX_MIB_SIZE) {
+      av1_setup_dst_planes(planes, cm->sb_size, frame_buffer, mi_row, mi_col);
+      for (int planeIdx = 0; planeIdx < num_planes; planeIdx += 1) {
+        const int32_t scaleHorz = planes[planeIdx].subsampling_x;
+        const int32_t scaleVert = planes[planeIdx].subsampling_y;
+        av1_filter_block_plane_horz(
+            cm, planes + planeIdx, (const MODE_INFO **)(mi + mi_col),
+            cm->mi_stride, (mi_col * MI_SIZE) >> scaleHorz,
+            (mi_row * MI_SIZE) >> scaleVert);
+      }
+    }
+  }
+#else   // CONFIG_PARALLEL_DEBLOCKING
+  for (mi_row = start; mi_row < stop; mi_row += MAX_MIB_SIZE) {
+    MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MAX_MIB_SIZE) {
+      int plane;
+
+      av1_setup_dst_planes(planes, cm->sb_size, frame_buffer, mi_row, mi_col);
+
+      // TODO(JBB): Make setup_mask work for non 420.
+      av1_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, &lfm);
+
+      av1_filter_block_plane_ss00_ver(cm, &planes[0], mi_row, &lfm);
+      av1_filter_block_plane_ss00_hor(cm, &planes[0], mi_row, &lfm);
+      for (plane = 1; plane < num_planes; ++plane) {
+        switch (path) {
+          case LF_PATH_420:
+            av1_filter_block_plane_ss11_ver(cm, &planes[plane], mi_row, &lfm);
+            av1_filter_block_plane_ss11_hor(cm, &planes[plane], mi_row, &lfm);
+            break;
+          case LF_PATH_444:
+            av1_filter_block_plane_ss00_ver(cm, &planes[plane], mi_row, &lfm);
+            av1_filter_block_plane_ss00_hor(cm, &planes[plane], mi_row, &lfm);
+            break;
+          case LF_PATH_SLOW:
+            av1_filter_block_plane_non420_ver(cm, &planes[plane], mi + mi_col,
+                                              mi_row, mi_col);
+            av1_filter_block_plane_non420_hor(cm, &planes[plane], mi + mi_col,
+                                              mi_row, mi_col);
+
+            break;
+        }
+      }
+    }
+  }
+#endif  // CONFIG_PARALLEL_DEBLOCKING
+#endif  // CONFIG_VAR_TX || CONFIG_EXT_PARTITION || CONFIG_EXT_PARTITION_TYPES
+}
+
+void av1_loop_filter_frame(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
+                           MACROBLOCKD *xd, int frame_filter_level, int y_only,
+                           int partial_frame) {
+  int start_mi_row, end_mi_row, mi_rows_to_filter;
+#if CONFIG_EXT_DELTA_Q
+  int orig_filter_level = cm->lf.filter_level;
+#endif
+  if (!frame_filter_level) return;
+  start_mi_row = 0;
+  mi_rows_to_filter = cm->mi_rows;
+  if (partial_frame && cm->mi_rows > 8) {
+    start_mi_row = cm->mi_rows >> 1;
+    start_mi_row &= 0xfffffff8;
+    mi_rows_to_filter = AOMMAX(cm->mi_rows / 8, 8);
+  }
+  end_mi_row = start_mi_row + mi_rows_to_filter;
+  av1_loop_filter_frame_init(cm, frame_filter_level);
+#if CONFIG_EXT_DELTA_Q
+  cm->lf.filter_level = frame_filter_level;
+#endif
+  av1_loop_filter_rows(frame, cm, xd->plane, start_mi_row, end_mi_row, y_only);
+#if CONFIG_EXT_DELTA_Q
+  cm->lf.filter_level = orig_filter_level;
+#endif
+}
+
+void av1_loop_filter_data_reset(
+    LFWorkerData *lf_data, YV12_BUFFER_CONFIG *frame_buffer,
+    struct AV1Common *cm, const struct macroblockd_plane planes[MAX_MB_PLANE]) {
+  lf_data->frame_buffer = frame_buffer;
+  lf_data->cm = cm;
+  lf_data->start = 0;
+  lf_data->stop = 0;
+  lf_data->y_only = 0;
+  memcpy(lf_data->planes, planes, sizeof(lf_data->planes));
+}
+
+int av1_loop_filter_worker(LFWorkerData *const lf_data, void *unused) {
+  (void)unused;
+  av1_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
+                       lf_data->start, lf_data->stop, lf_data->y_only);
+  return 1;
+}
diff --git a/third_party/aom/av1/common/av1_loopfilter.h b/third_party/aom/av1/common/av1_loopfilter.h
new file mode 100644
index 0000000000..8ac5d99e64
--- /dev/null
+++ b/third_party/aom/av1/common/av1_loopfilter.h
@@ -0,0 +1,162 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_LOOPFILTER_H_
+#define AV1_COMMON_LOOPFILTER_H_
+
+#include "aom_ports/mem.h"
+#include "./aom_config.h"
+
+#include "av1/common/blockd.h"
+#include "av1/common/seg_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_LOOP_FILTER 63
+#define MAX_SHARPNESS 7
+
+#define SIMD_WIDTH 16
+
+#define MAX_MODE_LF_DELTAS 2
+
+enum lf_path {
+  LF_PATH_420,
+  LF_PATH_444,
+  LF_PATH_SLOW,
+};
+
+struct loopfilter {
+  int filter_level;
+
+  int sharpness_level;
+  int last_sharpness_level;
+
+  uint8_t mode_ref_delta_enabled;
+  uint8_t mode_ref_delta_update;
+
+  // 0 = Intra, Last, Last2+Last3(CONFIG_EXT_REFS),
+  // GF, BRF(CONFIG_EXT_REFS), ARF
+  signed char ref_deltas[TOTAL_REFS_PER_FRAME];
+  signed char last_ref_deltas[TOTAL_REFS_PER_FRAME];
+
+  // 0 = ZERO_MV, MV
+  signed char mode_deltas[MAX_MODE_LF_DELTAS];
+  signed char last_mode_deltas[MAX_MODE_LF_DELTAS];
+};
+
+// Need to align this structure so when it is declared and
+// passed it can be loaded into vector registers.
+typedef struct {
+  DECLARE_ALIGNED(SIMD_WIDTH, uint8_t, mblim[SIMD_WIDTH]);
+  DECLARE_ALIGNED(SIMD_WIDTH, uint8_t, lim[SIMD_WIDTH]);
+  DECLARE_ALIGNED(SIMD_WIDTH, uint8_t, hev_thr[SIMD_WIDTH]);
+} loop_filter_thresh;
+
+typedef struct {
+  loop_filter_thresh lfthr[MAX_LOOP_FILTER + 1];
+  uint8_t lvl[MAX_SEGMENTS][TOTAL_REFS_PER_FRAME][MAX_MODE_LF_DELTAS];
+} loop_filter_info_n;
+
+// This structure holds bit masks for all 8x8 blocks in a 64x64 region.
+// Each 1 bit represents a position in which we want to apply the loop filter.
+// Left_ entries refer to whether we apply a filter on the border to the
+// left of the block.   Above_ entries refer to whether or not to apply a
+// filter on the above border.   Int_ entries refer to whether or not to
+// apply borders on the 4x4 edges within the 8x8 block that each bit
+// represents.
+// Since each transform is accompanied by a potentially different type of
+// loop filter there is a different entry in the array for each transform size.
+typedef struct {
+  uint64_t left_y[TX_SIZES];
+  uint64_t above_y[TX_SIZES];
+  uint64_t int_4x4_y;
+  uint16_t left_uv[TX_SIZES];
+  uint16_t above_uv[TX_SIZES];
+  uint16_t left_int_4x4_uv;
+  uint16_t above_int_4x4_uv;
+  uint8_t lfl_y[MAX_MIB_SIZE][MAX_MIB_SIZE];
+  uint8_t lfl_uv[MAX_MIB_SIZE / 2][MAX_MIB_SIZE / 2];
+} LOOP_FILTER_MASK;
+
+/* assorted loopfilter functions which get used elsewhere */
+struct AV1Common;
+struct macroblockd;
+struct AV1LfSyncData;
+
+// This function sets up the bit masks for the entire 64x64 region represented
+// by mi_row, mi_col.
+void av1_setup_mask(struct AV1Common *const cm, const int mi_row,
+                    const int mi_col, MODE_INFO **mi_8x8,
+                    const int mode_info_stride, LOOP_FILTER_MASK *lfm);
+
+void av1_filter_block_plane_ss00_ver(struct AV1Common *const cm,
+                                     struct macroblockd_plane *const plane,
+                                     int mi_row, LOOP_FILTER_MASK *lfm);
+void av1_filter_block_plane_ss00_hor(struct AV1Common *const cm,
+                                     struct macroblockd_plane *const plane,
+                                     int mi_row, LOOP_FILTER_MASK *lfm);
+void av1_filter_block_plane_ss11_ver(struct AV1Common *const cm,
+                                     struct macroblockd_plane *const plane,
+                                     int mi_row, LOOP_FILTER_MASK *lfm);
+void av1_filter_block_plane_ss11_hor(struct AV1Common *const cm,
+                                     struct macroblockd_plane *const plane,
+                                     int mi_row, LOOP_FILTER_MASK *lfm);
+
+void av1_filter_block_plane_non420_ver(struct AV1Common *const cm,
+                                       struct macroblockd_plane *plane,
+                                       MODE_INFO **mi_8x8, int mi_row,
+                                       int mi_col);
+void av1_filter_block_plane_non420_hor(struct AV1Common *const cm,
+                                       struct macroblockd_plane *plane,
+                                       MODE_INFO **mi_8x8, int mi_row,
+                                       int mi_col);
+
+void av1_loop_filter_init(struct AV1Common *cm);
+
+// Update the loop filter for the current frame.
+// This should be called before av1_loop_filter_rows(),
+// av1_loop_filter_frame()
+// calls this function directly.
+void av1_loop_filter_frame_init(struct AV1Common *cm, int default_filt_lvl);
+
+void av1_loop_filter_frame(YV12_BUFFER_CONFIG *frame, struct AV1Common *cm,
+                           struct macroblockd *mbd, int filter_level,
+                           int y_only, int partial_frame);
+
+// Apply the loop filter to [start, stop) macro block rows in frame_buffer.
+void av1_loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer,
+                          struct AV1Common *cm,
+                          struct macroblockd_plane planes[MAX_MB_PLANE],
+                          int start, int stop, int y_only);
+
+typedef struct LoopFilterWorkerData {
+  YV12_BUFFER_CONFIG *frame_buffer;
+  struct AV1Common *cm;
+  struct macroblockd_plane planes[MAX_MB_PLANE];
+
+  int start;
+  int stop;
+  int y_only;
+} LFWorkerData;
+
+void av1_loop_filter_data_reset(
+    LFWorkerData *lf_data, YV12_BUFFER_CONFIG *frame_buffer,
+    struct AV1Common *cm, const struct macroblockd_plane planes[MAX_MB_PLANE]);
+
+// Operates on the rows described by 'lf_data'.
+int av1_loop_filter_worker(LFWorkerData *const lf_data, void *unused);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_LOOPFILTER_H_
diff --git a/third_party/aom/av1/common/av1_rtcd.c b/third_party/aom/av1/common/av1_rtcd.c
new file mode 100644
index 0000000000..f9ccd19795
--- /dev/null
+++ b/third_party/aom/av1/common/av1_rtcd.c
@@ -0,0 +1,20 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "./aom_config.h"
+#define RTCD_C
+#include "./av1_rtcd.h"
+#include "aom_ports/aom_once.h"
+
+void av1_rtcd() {
+  // TODO(JBB): Remove this once, by insuring that both the encoder and
+  // decoder setup functions are protected by once();
+  once(setup_rtcd_internal);
+}
diff --git a/third_party/aom/av1/common/av1_rtcd_defs.pl b/third_party/aom/av1/common/av1_rtcd_defs.pl
new file mode 100755
index 0000000000..1dca10c526
--- /dev/null
+++ b/third_party/aom/av1/common/av1_rtcd_defs.pl
@@ -0,0 +1,644 @@
+sub av1_common_forward_decls() {
+print <<EOF
+/*
+ * AV1
+ */
+
+#include "aom/aom_integer.h"
+#include "av1/common/common.h"
+#include "av1/common/enums.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/filter.h"
+#include "av1/common/convolve.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/odintrin.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct aom_variance_vtable;
+struct search_site_config;
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+typedef uint16_t od_dering_in;
+EOF
+}
+forward_decls qw/av1_common_forward_decls/;
+
+# functions that are 64 bit only.
+$mmx_x86_64 = $sse2_x86_64 = $ssse3_x86_64 = $avx_x86_64 = $avx2_x86_64 = '';
+if ($opts{arch} eq "x86_64") {
+  $mmx_x86_64 = 'mmx';
+  $sse2_x86_64 = 'sse2';
+  $ssse3_x86_64 = 'ssse3';
+  $avx_x86_64 = 'avx';
+  $avx2_x86_64 = 'avx2';
+}
+
+#
+# 10/12-tap convolution filters
+#
+add_proto qw/void av1_lowbd_convolve_init/, "void";
+specialize qw/av1_lowbd_convolve_init ssse3/;
+
+add_proto qw/void av1_convolve_horiz/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params";
+specialize qw/av1_convolve_horiz ssse3/;
+
+add_proto qw/void av1_convolve_vert/, "const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params";
+specialize qw/av1_convolve_vert ssse3/;
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void av1_highbd_convolve_init/, "void";
+  specialize qw/av1_highbd_convolve_init sse4_1/;
+  add_proto qw/void av1_highbd_convolve_horiz/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd";
+  specialize qw/av1_highbd_convolve_horiz sse4_1/;
+  add_proto qw/void av1_highbd_convolve_vert/, "const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd";
+  specialize qw/av1_highbd_convolve_vert sse4_1/;
+}
+
+#
+# Inverse dct
+#
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  {
+    add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht4x4_16_add sse2/;
+
+    add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht4x8_32_add sse2/;
+
+    add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht8x4_32_add sse2/;
+
+    add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht8x16_128_add sse2/;
+
+    add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht16x8_128_add sse2/;
+
+    add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht16x32_512_add sse2/;
+
+    add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht32x16_512_add sse2/;
+
+    add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+
+    add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+
+    add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+
+    add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+
+    add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht8x8_64_add sse2/;
+
+    add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+      specialize qw/av1_iht16x16_256_add sse2 avx2/;
+
+    add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+  }
+} else {
+  {
+    add_proto qw/void av1_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+    specialize qw/av1_iht4x4_16_add sse2 neon dspr2/;
+
+    add_proto qw/void av1_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht4x8_32_add sse2/;
+
+    add_proto qw/void av1_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht8x4_32_add sse2/;
+
+    add_proto qw/void av1_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht8x16_128_add sse2/;
+
+    add_proto qw/void av1_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht16x8_128_add sse2/;
+
+    add_proto qw/void av1_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht16x32_512_add sse2/;
+
+    add_proto qw/void av1_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht32x16_512_add sse2/;
+
+    add_proto qw/void av1_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+
+    add_proto qw/void av1_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+
+    add_proto qw/void av1_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+
+    add_proto qw/void av1_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+
+    add_proto qw/void av1_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type";
+      specialize qw/av1_iht8x8_64_add sse2 neon dspr2/;
+
+    add_proto qw/void av1_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+      specialize qw/av1_iht16x16_256_add sse2 avx2 dspr2/;
+
+    add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+
+    if (aom_config("CONFIG_EXT_TX") ne "yes") {
+      specialize qw/av1_iht4x4_16_add msa/;
+      specialize qw/av1_iht8x8_64_add msa/;
+      specialize qw/av1_iht16x16_256_add msa/;
+    }
+  }
+}
+
+add_proto qw/void av1_iht32x32_1024_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+
+if (aom_config("CONFIG_TX64X64") eq "yes") {
+  add_proto qw/void av1_iht64x64_4096_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type";
+}
+
+if (aom_config("CONFIG_NEW_QUANT") eq "yes") {
+  add_proto qw/void quantize_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+
+  add_proto qw/void quantize_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+
+  add_proto qw/void quantize_32x32_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+
+  add_proto qw/void quantize_32x32_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+
+  if (aom_config("CONFIG_TX64X64") eq "yes") {
+    add_proto qw/void quantize_64x64_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+
+    add_proto qw/void quantize_64x64_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+  }
+}
+
+# FILTER_INTRA predictor functions
+if (aom_config("CONFIG_FILTER_INTRA") eq "yes") {
+  add_proto qw/void av1_dc_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
+  add_proto qw/void av1_v_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
+  add_proto qw/void av1_h_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
+  add_proto qw/void av1_d45_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
+  add_proto qw/void av1_d135_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
+  add_proto qw/void av1_d117_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
+  add_proto qw/void av1_d153_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
+  add_proto qw/void av1_d207_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
+  add_proto qw/void av1_d63_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
+  add_proto qw/void av1_tm_filter_predictor/, "uint8_t *dst, ptrdiff_t stride, int bs, const uint8_t *above, const uint8_t *left";
+  # High bitdepth functions
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    add_proto qw/void av1_highbd_dc_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
+    add_proto qw/void av1_highbd_v_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
+    add_proto qw/void av1_highbd_h_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
+    add_proto qw/void av1_highbd_d45_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
+    add_proto qw/void av1_highbd_d135_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
+    add_proto qw/void av1_highbd_d117_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
+    add_proto qw/void av1_highbd_d153_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
+    add_proto qw/void av1_highbd_d207_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
+    add_proto qw/void av1_highbd_d63_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
+    add_proto qw/void av1_highbd_tm_filter_predictor/, "uint16_t *dst, ptrdiff_t stride, int bs, const uint16_t *above, const uint16_t *left, int bd";
+  }
+}
+
+# High bitdepth functions
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  #
+  # Sub Pixel Filters
+  #
+  add_proto qw/void av1_highbd_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+
+  add_proto qw/void av1_highbd_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+
+  add_proto qw/void av1_highbd_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/av1_highbd_convolve8/, "$sse2_x86_64";
+
+  add_proto qw/void av1_highbd_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/av1_highbd_convolve8_horiz/, "$sse2_x86_64";
+
+  add_proto qw/void av1_highbd_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/av1_highbd_convolve8_vert/, "$sse2_x86_64";
+
+  add_proto qw/void av1_highbd_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/av1_highbd_convolve8_avg/, "$sse2_x86_64";
+
+  add_proto qw/void av1_highbd_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/av1_highbd_convolve8_avg_horiz/, "$sse2_x86_64";
+
+  add_proto qw/void av1_highbd_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/av1_highbd_convolve8_avg_vert/, "$sse2_x86_64";
+
+  #
+  # dct
+  #
+  add_proto qw/void av1_highbd_iht4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht4x8_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht8x4_32_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht8x16_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht16x8_128_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht16x32_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht32x16_512_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht4x16_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht16x4_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht8x32_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht32x8_256_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht8x8_64_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd";
+
+  add_proto qw/void av1_highbd_iht16x16_256_add/, "const tran_low_t *input, uint8_t *output, int pitch, int tx_type, int bd";
+}
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  #inv txfm
+  add_proto qw/void av1_inv_txfm2d_add_4x4/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
+  specialize qw/av1_inv_txfm2d_add_4x4 sse4_1/;
+  add_proto qw/void av1_inv_txfm2d_add_8x8/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
+  specialize qw/av1_inv_txfm2d_add_8x8 sse4_1/;
+  add_proto qw/void av1_inv_txfm2d_add_16x16/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
+  specialize qw/av1_inv_txfm2d_add_16x16 sse4_1/;
+  add_proto qw/void av1_inv_txfm2d_add_32x32/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
+  specialize qw/av1_inv_txfm2d_add_32x32 avx2/;
+  add_proto qw/void av1_inv_txfm2d_add_64x64/, "const int32_t *input, uint16_t *output, int stride, int tx_type, int bd";
+}
+
+#
+# Encoder functions below this point.
+#
+if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
+
+# ENCODEMB INVOKE
+
+if (aom_config("CONFIG_AOM_QM") eq "yes") {
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    # the transform coefficients are held in 32-bit
+    # values, so the assembler code for  av1_block_error can no longer be used.
+    add_proto qw/int64_t av1_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
+
+    add_proto qw/void av1_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
+
+    add_proto qw/void av1_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
+
+    if (aom_config("CONFIG_TX64X64") eq "yes") {
+      add_proto qw/void av1_quantize_fp_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
+    }
+
+    add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
+  } else {
+    add_proto qw/int64_t av1_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
+    specialize qw/av1_block_error avx2 msa/, "$sse2_x86inc";
+
+    add_proto qw/int64_t av1_block_error_fp/, "const int16_t *coeff, const int16_t *dqcoeff, int block_size";
+    specialize qw/av1_block_error_fp neon/, "$sse2_x86inc";
+
+    add_proto qw/void av1_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
+
+    add_proto qw/void av1_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
+
+    if (aom_config("CONFIG_TX64X64") eq "yes") {
+      add_proto qw/void av1_quantize_fp_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
+    }
+
+    add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t *iqm_ptr";
+  }
+} else {
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    # the transform coefficients are held in 32-bit
+    # values, so the assembler code for  av1_block_error can no longer be used.
+    add_proto qw/int64_t av1_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
+
+    add_proto qw/void av1_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+
+    add_proto qw/void av1_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+
+    if (aom_config("CONFIG_TX64X64") eq "yes") {
+      add_proto qw/void av1_quantize_fp_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    }
+
+    add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+  } else {
+    add_proto qw/int64_t av1_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz";
+    specialize qw/av1_block_error sse2 avx2 msa/;
+
+    add_proto qw/int64_t av1_block_error_fp/, "const int16_t *coeff, const int16_t *dqcoeff, int block_size";
+    specialize qw/av1_block_error_fp neon sse2/;
+
+    add_proto qw/void av1_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    specialize qw/av1_quantize_fp neon sse2/, "$ssse3_x86_64";
+
+    add_proto qw/void av1_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    specialize qw/av1_quantize_fp_32x32/, "$ssse3_x86_64";
+
+    if (aom_config("CONFIG_TX64X64") eq "yes") {
+      add_proto qw/void av1_quantize_fp_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    }
+
+    add_proto qw/void av1_fdct8x8_quant/, "const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan";
+    specialize qw/av1_fdct8x8_quant sse2 ssse3 neon/;
+  }
+
+}
+
+# fdct functions
+
+add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+specialize qw/av1_fht4x4 sse2/;
+
+add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
+
+add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+specialize qw/av1_fht8x8 sse2/;
+
+add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+specialize qw/av1_fht16x16 sse2 avx2/;
+
+add_proto qw/void av1_fht32x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+specialize qw/av1_fht32x32 sse2 avx2/;
+
+if (aom_config("CONFIG_TX64X64") eq "yes") {
+  add_proto qw/void av1_fht64x64/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+}
+
+add_proto qw/void av1_fht4x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+specialize qw/av1_fht4x8 sse2/;
+
+add_proto qw/void av1_fht8x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+specialize qw/av1_fht8x4 sse2/;
+
+add_proto qw/void av1_fht8x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+specialize qw/av1_fht8x16 sse2/;
+
+add_proto qw/void av1_fht16x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+specialize qw/av1_fht16x8 sse2/;
+
+add_proto qw/void av1_fht16x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+specialize qw/av1_fht16x32 sse2/;
+
+add_proto qw/void av1_fht32x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+specialize qw/av1_fht32x16 sse2/;
+
+add_proto qw/void av1_fht4x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+add_proto qw/void av1_fht16x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+add_proto qw/void av1_fht8x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+add_proto qw/void av1_fht32x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+if (aom_config("CONFIG_HIGHBITDEPTH") ne "yes") {
+  if (aom_config("CONFIG_EXT_TX") ne "yes") {
+    specialize qw/av1_fht4x4 msa/;
+    specialize qw/av1_fht8x8 msa/;
+    specialize qw/av1_fht16x16 msa/;
+  }
+}
+
+add_proto qw/void av1_fwd_idtx/, "const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type";
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  #fwd txfm
+  add_proto qw/void av1_fwd_txfm2d_4x4/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
+  specialize qw/av1_fwd_txfm2d_4x4 sse4_1/;
+  add_proto qw/void av1_fwd_txfm2d_8x8/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
+  specialize qw/av1_fwd_txfm2d_8x8 sse4_1/;
+  add_proto qw/void av1_fwd_txfm2d_16x16/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
+  specialize qw/av1_fwd_txfm2d_16x16 sse4_1/;
+  add_proto qw/void av1_fwd_txfm2d_32x32/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
+  specialize qw/av1_fwd_txfm2d_32x32 sse4_1/;
+  add_proto qw/void av1_fwd_txfm2d_64x64/, "const int16_t *input, int32_t *output, int stride, int tx_type, int bd";
+  specialize qw/av1_fwd_txfm2d_64x64 sse4_1/;
+}
+
+#
+# Motion search
+#
+add_proto qw/int av1_full_search_sad/, "const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv";
+specialize qw/av1_full_search_sad sse3 sse4_1/;
+$av1_full_search_sad_sse3=av1_full_search_sadx3;
+$av1_full_search_sad_sse4_1=av1_full_search_sadx8;
+
+add_proto qw/int av1_diamond_search_sad/, "struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv";
+
+add_proto qw/int av1_full_range_search/, "const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv";
+
+add_proto qw/void av1_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
+specialize qw/av1_temporal_filter_apply sse2 msa/;
+
+if (aom_config("CONFIG_AOM_QM") eq "yes") {
+  add_proto qw/void av1_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
+} else {
+  add_proto qw/void av1_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale";
+}
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+
+  # ENCODEMB INVOKE
+  if (aom_config("CONFIG_NEW_QUANT") eq "yes") {
+    add_proto qw/void highbd_quantize_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+
+    add_proto qw/void highbd_quantize_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+
+    add_proto qw/void highbd_quantize_32x32_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+
+    add_proto qw/void highbd_quantize_32x32_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+
+    if (aom_config("CONFIG_TX64X64") eq "yes") {
+      add_proto qw/void highbd_quantize_64x64_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+
+      add_proto qw/void highbd_quantize_64x64_fp_nuq/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *quant_ptr, const int16_t *dequant_ptr, const cuml_bins_type_nuq *cuml_bins_ptr, const dequant_val_type_nuq *dequant_val, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const int16_t *scan, const uint8_t *band";
+    }
+  }
+
+  add_proto qw/int64_t av1_highbd_block_error/, "const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd";
+  specialize qw/av1_highbd_block_error sse2/;
+
+  if (aom_config("CONFIG_AOM_QM") eq "yes") {
+    add_proto qw/void av1_highbd_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
+
+    add_proto qw/void av1_highbd_quantize_fp_32x32/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
+
+    if (aom_config("CONFIG_TX64X64") eq "yes") {
+      add_proto qw/void av1_highbd_quantize_fp_64x64/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
+    }
+
+    add_proto qw/void av1_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale";
+  } else {
+    add_proto qw/void av1_highbd_quantize_fp/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale";
+    specialize qw/av1_highbd_quantize_fp sse4_1/;
+
+    add_proto qw/void av1_highbd_quantize_b/, "const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale";
+  }
+
+  # fdct functions
+  add_proto qw/void av1_highbd_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/av1_highbd_fht4x4 sse4_1/;
+
+  add_proto qw/void av1_highbd_fht4x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht8x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht8x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht16x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht16x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht32x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht4x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht16x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht8x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht32x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  add_proto qw/void av1_highbd_fht32x32/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+
+  if (aom_config("CONFIG_TX64X64") eq "yes") {
+    add_proto qw/void av1_highbd_fht64x64/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  }
+
+  add_proto qw/void av1_highbd_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
+
+  add_proto qw/void av1_highbd_temporal_filter_apply/, "uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count";
+
+}
+# End av1_high encoder functions
+
+if (aom_config("CONFIG_EXT_INTER") eq "yes") {
+  add_proto qw/uint64_t av1_wedge_sse_from_residuals/, "const int16_t *r1, const int16_t *d, const uint8_t *m, int N";
+  specialize qw/av1_wedge_sse_from_residuals sse2/;
+  add_proto qw/int av1_wedge_sign_from_residuals/, "const int16_t *ds, const uint8_t *m, int N, int64_t limit";
+  specialize qw/av1_wedge_sign_from_residuals sse2/;
+  add_proto qw/void av1_wedge_compute_delta_squares/, "int16_t *d, const int16_t *a, const int16_t *b, int N";
+  specialize qw/av1_wedge_compute_delta_squares sse2/;
+}
+
+}
+# end encoder functions
+
+# If PVQ is enabled, fwd transforms are required by decoder
+if (aom_config("CONFIG_PVQ") eq "yes") {
+# fdct functions
+
+if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+  add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/av1_fht4x4 sse2/;
+
+  add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/av1_fht8x8 sse2/;
+
+  add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/av1_fht16x16 sse2/;
+
+  add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/av1_fwht4x4 sse2/;
+} else {
+  add_proto qw/void av1_fht4x4/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/av1_fht4x4 sse2 msa/;
+
+  add_proto qw/void av1_fht8x8/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/av1_fht8x8 sse2 msa/;
+
+  add_proto qw/void av1_fht16x16/, "const int16_t *input, tran_low_t *output, int stride, int tx_type";
+  specialize qw/av1_fht16x16 sse2 msa/;
+
+  add_proto qw/void av1_fwht4x4/, "const int16_t *input, tran_low_t *output, int stride";
+  specialize qw/av1_fwht4x4 msa sse2/;
+}
+
+}
+
+# Deringing Functions
+
+if (aom_config("CONFIG_CDEF") eq "yes") {
+  add_proto qw/void aom_clpf_block_hbd/, "uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd";
+  add_proto qw/void aom_clpf_hblock_hbd/, "uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd";
+  add_proto qw/void aom_clpf_block/, "uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd";
+  add_proto qw/void aom_clpf_hblock/, "uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd";
+  add_proto qw/int od_dir_find8/, "const od_dering_in *img, int stride, int32_t *var, int coeff_shift";
+  add_proto qw/void od_filter_dering_direction_4x4/, "uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping";
+  add_proto qw/void od_filter_dering_direction_8x8/, "uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping";
+
+  add_proto qw/void copy_8x8_16bit_to_8bit/, "uint8_t *dst, int dstride, const uint16_t *src, int sstride";
+  add_proto qw/void copy_4x4_16bit_to_8bit/, "uint8_t *dst, int dstride, const uint16_t *src, int sstride";
+  add_proto qw/void copy_8x8_16bit_to_16bit/, "uint16_t *dst, int dstride, const uint16_t *src, int sstride";
+  add_proto qw/void copy_4x4_16bit_to_16bit/, "uint16_t *dst, int dstride, const uint16_t *src, int sstride";
+  add_proto qw/void copy_rect8_8bit_to_16bit/, "uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h";
+  add_proto qw/void copy_rect8_16bit_to_16bit/, "uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h";
+
+# VS compiling for 32 bit targets does not support vector types in
+  # structs as arguments, which makes the v256 type of the intrinsics
+  # hard to support, so optimizations for this target are disabled.
+  if ($opts{config} !~ /libs-x86-win32-vs.*/) {
+    specialize qw/aom_clpf_block_hbd sse2 ssse3 sse4_1 neon/;
+    specialize qw/aom_clpf_hblock_hbd sse2 ssse3 sse4_1 neon/;
+    specialize qw/aom_clpf_block sse2 ssse3 sse4_1 neon/;
+    specialize qw/aom_clpf_hblock sse2 ssse3 sse4_1 neon/;
+    specialize qw/od_dir_find8 sse2 ssse3 sse4_1 neon/;
+    specialize qw/od_filter_dering_direction_4x4 sse2 ssse3 sse4_1 neon/;
+    specialize qw/od_filter_dering_direction_8x8 sse2 ssse3 sse4_1 neon/;
+
+    specialize qw/copy_8x8_16bit_to_8bit sse2 ssse3 sse4_1 neon/;
+    specialize qw/copy_4x4_16bit_to_8bit sse2 ssse3 sse4_1 neon/;
+    specialize qw/copy_8x8_16bit_to_16bit sse2 ssse3 sse4_1 neon/;
+    specialize qw/copy_4x4_16bit_to_16bit sse2 ssse3 sse4_1 neon/;
+    specialize qw/copy_rect8_8bit_to_16bit sse2 ssse3 sse4_1 neon/;
+    specialize qw/copy_rect8_16bit_to_16bit sse2 ssse3 sse4_1 neon/;
+  }
+}
+
+# PVQ Functions
+
+if (aom_config("CONFIG_PVQ") eq "yes") {
+  add_proto qw/double pvq_search_rdo_double/, "const od_val16 *xcoeff, int n, int k, int *ypulse, double g2, double pvq_norm_lambda, int prev_k";
+  specialize qw/pvq_search_rdo_double sse4_1/;
+}
+
+# WARPED_MOTION / GLOBAL_MOTION functions
+
+if ((aom_config("CONFIG_WARPED_MOTION") eq "yes") ||
+    (aom_config("CONFIG_GLOBAL_MOTION") eq "yes")) {
+  add_proto qw/void av1_warp_affine/, "int32_t *mat, uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta";
+  specialize qw/av1_warp_affine sse2/;
+
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    add_proto qw/void av1_highbd_warp_affine/, "int32_t *mat, uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, int ref_frm, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta";
+    specialize qw/av1_highbd_warp_affine ssse3/;
+  }
+}
+
+# LOOP_RESTORATION functions
+
+if (aom_config("CONFIG_LOOP_RESTORATION") eq "yes") {
+  add_proto qw/void apply_selfguided_restoration/, "uint8_t *dat, int width, int height, int stride, int eps, int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf";
+  specialize qw/apply_selfguided_restoration sse4_1/;
+
+  add_proto qw/void av1_selfguided_restoration/, "uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps, int32_t *tmpbuf";
+  specialize qw/av1_selfguided_restoration sse4_1/;
+
+  add_proto qw/void av1_highpass_filter/, "uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps";
+  specialize qw/av1_highpass_filter sse4_1/;
+
+  if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
+    add_proto qw/void apply_selfguided_restoration_highbd/, "uint16_t *dat, int width, int height, int stride, int bit_depth, int eps, int *xqd, uint16_t *dst, int dst_stride, int32_t *tmpbuf";
+    specialize qw/apply_selfguided_restoration_highbd sse4_1/;
+
+    add_proto qw/void av1_selfguided_restoration_highbd/, "uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int bit_depth, int r, int eps, int32_t *tmpbuf";
+    specialize qw/av1_selfguided_restoration_highbd sse4_1/;
+
+    add_proto qw/void av1_highpass_filter_highbd/, "uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps";
+    specialize qw/av1_highpass_filter_highbd sse4_1/;
+  }
+}
+
+1;
diff --git a/third_party/aom/av1/common/av1_txfm.h b/third_party/aom/av1/common/av1_txfm.h
new file mode 100644
index 0000000000..6987317aee
--- /dev/null
+++ b/third_party/aom/av1/common/av1_txfm.h
@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_TXFM_H_
+#define AV1_TXFM_H_
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "av1/common/enums.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+static const int cos_bit_min = 10;
+static const int cos_bit_max = 16;
+
+// cospi_arr[i][j] = (int)round(cos(M_PI*j/128) * (1<<(cos_bit_min+i)));
+static const int32_t cospi_arr[7][64] = {
+  { 1024, 1024, 1023, 1021, 1019, 1016, 1013, 1009, 1004, 999, 993, 987, 980,
+    972,  964,  955,  946,  936,  926,  915,  903,  891,  878, 865, 851, 837,
+    822,  807,  792,  775,  759,  742,  724,  706,  688,  669, 650, 630, 610,
+    590,  569,  548,  526,  505,  483,  460,  438,  415,  392, 369, 345, 321,
+    297,  273,  249,  224,  200,  175,  150,  125,  100,  75,  50,  25 },
+  { 2048, 2047, 2046, 2042, 2038, 2033, 2026, 2018, 2009, 1998, 1987,
+    1974, 1960, 1945, 1928, 1911, 1892, 1872, 1851, 1829, 1806, 1782,
+    1757, 1730, 1703, 1674, 1645, 1615, 1583, 1551, 1517, 1483, 1448,
+    1412, 1375, 1338, 1299, 1260, 1220, 1179, 1138, 1096, 1053, 1009,
+    965,  921,  876,  830,  784,  737,  690,  642,  595,  546,  498,
+    449,  400,  350,  301,  251,  201,  151,  100,  50 },
+  { 4096, 4095, 4091, 4085, 4076, 4065, 4052, 4036, 4017, 3996, 3973,
+    3948, 3920, 3889, 3857, 3822, 3784, 3745, 3703, 3659, 3612, 3564,
+    3513, 3461, 3406, 3349, 3290, 3229, 3166, 3102, 3035, 2967, 2896,
+    2824, 2751, 2675, 2598, 2520, 2440, 2359, 2276, 2191, 2106, 2019,
+    1931, 1842, 1751, 1660, 1567, 1474, 1380, 1285, 1189, 1092, 995,
+    897,  799,  700,  601,  501,  401,  301,  201,  101 },
+  { 8192, 8190, 8182, 8170, 8153, 8130, 8103, 8071, 8035, 7993, 7946,
+    7895, 7839, 7779, 7713, 7643, 7568, 7489, 7405, 7317, 7225, 7128,
+    7027, 6921, 6811, 6698, 6580, 6458, 6333, 6203, 6070, 5933, 5793,
+    5649, 5501, 5351, 5197, 5040, 4880, 4717, 4551, 4383, 4212, 4038,
+    3862, 3683, 3503, 3320, 3135, 2948, 2760, 2570, 2378, 2185, 1990,
+    1795, 1598, 1401, 1202, 1003, 803,  603,  402,  201 },
+  { 16384, 16379, 16364, 16340, 16305, 16261, 16207, 16143, 16069, 15986, 15893,
+    15791, 15679, 15557, 15426, 15286, 15137, 14978, 14811, 14635, 14449, 14256,
+    14053, 13842, 13623, 13395, 13160, 12916, 12665, 12406, 12140, 11866, 11585,
+    11297, 11003, 10702, 10394, 10080, 9760,  9434,  9102,  8765,  8423,  8076,
+    7723,  7366,  7005,  6639,  6270,  5897,  5520,  5139,  4756,  4370,  3981,
+    3590,  3196,  2801,  2404,  2006,  1606,  1205,  804,   402 },
+  { 32768, 32758, 32729, 32679, 32610, 32522, 32413, 32286, 32138, 31972, 31786,
+    31581, 31357, 31114, 30853, 30572, 30274, 29957, 29622, 29269, 28899, 28511,
+    28106, 27684, 27246, 26791, 26320, 25833, 25330, 24812, 24279, 23732, 23170,
+    22595, 22006, 21403, 20788, 20160, 19520, 18868, 18205, 17531, 16846, 16151,
+    15447, 14733, 14010, 13279, 12540, 11793, 11039, 10279, 9512,  8740,  7962,
+    7180,  6393,  5602,  4808,  4011,  3212,  2411,  1608,  804 },
+  { 65536, 65516, 65457, 65358, 65220, 65043, 64827, 64571, 64277, 63944, 63572,
+    63162, 62714, 62228, 61705, 61145, 60547, 59914, 59244, 58538, 57798, 57022,
+    56212, 55368, 54491, 53581, 52639, 51665, 50660, 49624, 48559, 47464, 46341,
+    45190, 44011, 42806, 41576, 40320, 39040, 37736, 36410, 35062, 33692, 32303,
+    30893, 29466, 28020, 26558, 25080, 23586, 22078, 20557, 19024, 17479, 15924,
+    14359, 12785, 11204, 9616,  8022,  6424,  4821,  3216,  1608 }
+};
+
+static INLINE int32_t round_shift(int32_t value, int bit) {
+  assert(bit >= 1);
+  return (value + (1 << (bit - 1))) >> bit;
+}
+
+static INLINE void round_shift_array(int32_t *arr, int size, int bit) {
+  int i;
+  if (bit == 0) {
+    return;
+  } else {
+    if (bit > 0) {
+      for (i = 0; i < size; i++) {
+        arr[i] = round_shift(arr[i], bit);
+      }
+    } else {
+      for (i = 0; i < size; i++) {
+        arr[i] = arr[i] * (1 << (-bit));
+      }
+    }
+  }
+}
+
+static INLINE int32_t half_btf(int32_t w0, int32_t in0, int32_t w1, int32_t in1,
+                               int bit) {
+  int32_t result_32 = w0 * in0 + w1 * in1;
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+  int64_t result_64 = (int64_t)w0 * (int64_t)in0 + (int64_t)w1 * (int64_t)in1;
+  if (result_64 < INT32_MIN || result_64 > INT32_MAX) {
+    printf("%s %d overflow result_32: %d result_64: %" PRId64
+           " w0: %d in0: %d w1: %d in1: "
+           "%d\n",
+           __FILE__, __LINE__, result_32, result_64, w0, in0, w1, in1);
+    assert(0 && "half_btf overflow");
+  }
+#endif
+  return round_shift(result_32, bit);
+}
+
+static INLINE int get_max_bit(int x) {
+  int max_bit = -1;
+  while (x) {
+    x = x >> 1;
+    max_bit++;
+  }
+  return max_bit;
+}
+
+// TODO(angiebird): implement SSE
+static INLINE void clamp_block(int16_t *block, int block_size, int stride,
+                               int low, int high) {
+  int i, j;
+  for (i = 0; i < block_size; ++i) {
+    for (j = 0; j < block_size; ++j) {
+      block[i * stride + j] = clamp(block[i * stride + j], low, high);
+    }
+  }
+}
+
+typedef void (*TxfmFunc)(const int32_t *input, int32_t *output,
+                         const int8_t *cos_bit, const int8_t *stage_range);
+
+typedef enum TXFM_TYPE {
+  TXFM_TYPE_DCT4,
+  TXFM_TYPE_DCT8,
+  TXFM_TYPE_DCT16,
+  TXFM_TYPE_DCT32,
+  TXFM_TYPE_DCT64,
+  TXFM_TYPE_ADST4,
+  TXFM_TYPE_ADST8,
+  TXFM_TYPE_ADST16,
+  TXFM_TYPE_ADST32,
+} TXFM_TYPE;
+
+typedef struct TXFM_2D_CFG {
+  const int txfm_size;
+  const int stage_num_col;
+  const int stage_num_row;
+
+  const int8_t *shift;
+  const int8_t *stage_range_col;
+  const int8_t *stage_range_row;
+  const int8_t *cos_bit_col;
+  const int8_t *cos_bit_row;
+  const TXFM_TYPE txfm_type_col;
+  const TXFM_TYPE txfm_type_row;
+} TXFM_2D_CFG;
+
+typedef struct TXFM_2D_FLIP_CFG {
+  int ud_flip;  // flip upside down
+  int lr_flip;  // flip left to right
+  const TXFM_2D_CFG *cfg;
+} TXFM_2D_FLIP_CFG;
+
+static INLINE void set_flip_cfg(int tx_type, TXFM_2D_FLIP_CFG *cfg) {
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+      cfg->ud_flip = 0;
+      cfg->lr_flip = 0;
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      cfg->ud_flip = 1;
+      cfg->lr_flip = 0;
+      break;
+    case DCT_FLIPADST:
+      cfg->ud_flip = 0;
+      cfg->lr_flip = 1;
+      break;
+    case FLIPADST_FLIPADST:
+      cfg->ud_flip = 1;
+      cfg->lr_flip = 1;
+      break;
+    case ADST_FLIPADST:
+      cfg->ud_flip = 0;
+      cfg->lr_flip = 1;
+      break;
+    case FLIPADST_ADST:
+      cfg->ud_flip = 1;
+      cfg->lr_flip = 0;
+      break;
+#endif  // CONFIG_EXT_TX
+    default:
+      cfg->ud_flip = 0;
+      cfg->lr_flip = 0;
+      assert(0);
+  }
+}
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+TXFM_2D_FLIP_CFG av1_get_fwd_txfm_cfg(int tx_type, int tx_size);
+TXFM_2D_FLIP_CFG av1_get_fwd_txfm_64x64_cfg(int tx_type);
+#ifdef __cplusplus
+}
+#endif  // __cplusplus
+
+#endif  // AV1_TXFM_H_
diff --git a/third_party/aom/av1/common/blockd.c b/third_party/aom/av1/common/blockd.c
new file mode 100644
index 0000000000..4eb6f01eaf
--- /dev/null
+++ b/third_party/aom/av1/common/blockd.c
@@ -0,0 +1,296 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+
+#include "aom_ports/system_state.h"
+
+#include "av1/common/blockd.h"
+#include "av1/common/onyxc_int.h"
+
+PREDICTION_MODE av1_left_block_mode(const MODE_INFO *cur_mi,
+                                    const MODE_INFO *left_mi, int b) {
+  if (b == 0 || b == 2) {
+    if (!left_mi || is_inter_block(&left_mi->mbmi)) return DC_PRED;
+
+    return get_y_mode(left_mi, b + 1);
+  } else {
+    assert(b == 1 || b == 3);
+    return cur_mi->bmi[b - 1].as_mode;
+  }
+}
+
+PREDICTION_MODE av1_above_block_mode(const MODE_INFO *cur_mi,
+                                     const MODE_INFO *above_mi, int b) {
+  if (b == 0 || b == 1) {
+    if (!above_mi || is_inter_block(&above_mi->mbmi)) return DC_PRED;
+
+    return get_y_mode(above_mi, b + 2);
+  } else {
+    assert(b == 2 || b == 3);
+    return cur_mi->bmi[b - 2].as_mode;
+  }
+}
+
+#if CONFIG_COEF_INTERLEAVE
+void av1_foreach_transformed_block_interleave(
+    const MACROBLOCKD *const xd, BLOCK_SIZE bsize,
+    foreach_transformed_block_visitor visit, void *arg) {
+  const struct macroblockd_plane *const pd_y = &xd->plane[0];
+  const struct macroblockd_plane *const pd_c = &xd->plane[1];
+  const MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+
+  const TX_SIZE tx_log2_y = mbmi->tx_size;
+  const TX_SIZE tx_log2_c = get_uv_tx_size(mbmi, pd_c);
+  const int tx_sz_y = (1 << tx_log2_y);
+  const int tx_sz_c = (1 << tx_log2_c);
+
+  const BLOCK_SIZE plane_bsize_y = get_plane_block_size(bsize, pd_y);
+  const BLOCK_SIZE plane_bsize_c = get_plane_block_size(bsize, pd_c);
+
+  const int num_4x4_w_y = num_4x4_blocks_wide_lookup[plane_bsize_y];
+  const int num_4x4_w_c = num_4x4_blocks_wide_lookup[plane_bsize_c];
+  const int num_4x4_h_y = num_4x4_blocks_high_lookup[plane_bsize_y];
+  const int num_4x4_h_c = num_4x4_blocks_high_lookup[plane_bsize_c];
+
+  const int step_y = 1 << (tx_log2_y << 1);
+  const int step_c = 1 << (tx_log2_c << 1);
+
+  const int max_4x4_w_y =
+      get_max_4x4_size(num_4x4_w_y, xd->mb_to_right_edge, pd_y->subsampling_x);
+  const int max_4x4_h_y =
+      get_max_4x4_size(num_4x4_h_y, xd->mb_to_bottom_edge, pd_y->subsampling_y);
+
+  const int extra_step_y = ((num_4x4_w_y - max_4x4_w_y) >> tx_log2_y) * step_y;
+
+  const int max_4x4_w_c =
+      get_max_4x4_size(num_4x4_w_c, xd->mb_to_right_edge, pd_c->subsampling_x);
+  const int max_4x4_h_c =
+      get_max_4x4_size(num_4x4_h_c, xd->mb_to_bottom_edge, pd_c->subsampling_y);
+
+  const int extra_step_c = ((num_4x4_w_c - max_4x4_w_c) >> tx_log2_c) * step_c;
+
+  // The max_4x4_w/h may be smaller than tx_sz under some corner cases,
+  // i.e. when the SB is splitted by tile boundaries.
+  const int tu_num_w_y = (max_4x4_w_y + tx_sz_y - 1) / tx_sz_y;
+  const int tu_num_h_y = (max_4x4_h_y + tx_sz_y - 1) / tx_sz_y;
+  const int tu_num_w_c = (max_4x4_w_c + tx_sz_c - 1) / tx_sz_c;
+  const int tu_num_h_c = (max_4x4_h_c + tx_sz_c - 1) / tx_sz_c;
+  const int tu_num_c = tu_num_w_c * tu_num_h_c;
+
+  int tu_idx_c = 0;
+  int offset_y, row_y, col_y;
+  int offset_c, row_c, col_c;
+
+  for (row_y = 0; row_y < tu_num_h_y; row_y++) {
+    for (col_y = 0; col_y < tu_num_w_y; col_y++) {
+      // luma
+      offset_y = (row_y * tu_num_w_y + col_y) * step_y + row_y * extra_step_y;
+      visit(0, offset_y, row_y * tx_sz_y, col_y * tx_sz_y, plane_bsize_y,
+            tx_log2_y, arg);
+      // chroma
+      if (tu_idx_c < tu_num_c) {
+        row_c = (tu_idx_c / tu_num_w_c) * tx_sz_c;
+        col_c = (tu_idx_c % tu_num_w_c) * tx_sz_c;
+        offset_c = tu_idx_c * step_c + (tu_idx_c / tu_num_w_c) * extra_step_c;
+        visit(1, offset_c, row_c, col_c, plane_bsize_c, tx_log2_c, arg);
+        visit(2, offset_c, row_c, col_c, plane_bsize_c, tx_log2_c, arg);
+        tu_idx_c++;
+      }
+    }
+  }
+
+  // In 422 case, it's possible that Chroma has more TUs than Luma
+  while (tu_idx_c < tu_num_c) {
+    row_c = (tu_idx_c / tu_num_w_c) * tx_sz_c;
+    col_c = (tu_idx_c % tu_num_w_c) * tx_sz_c;
+    offset_c = tu_idx_c * step_c + row_c * extra_step_c;
+    visit(1, offset_c, row_c, col_c, plane_bsize_c, tx_log2_c, arg);
+    visit(2, offset_c, row_c, col_c, plane_bsize_c, tx_log2_c, arg);
+    tu_idx_c++;
+  }
+}
+#endif
+
+void av1_foreach_transformed_block_in_plane(
+    const MACROBLOCKD *const xd, BLOCK_SIZE bsize, int plane,
+    foreach_transformed_block_visitor visit, void *arg) {
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  // block and transform sizes, in number of 4x4 blocks log 2 ("*_b")
+  // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8
+  // transform size varies per plane, look it up in a common way.
+  const TX_SIZE tx_size = get_tx_size(plane, xd);
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+  const BLOCK_SIZE plane_bsize =
+      AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#else
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#endif
+  const uint8_t txw_unit = tx_size_wide_unit[tx_size];
+  const uint8_t txh_unit = tx_size_high_unit[tx_size];
+  const int step = txw_unit * txh_unit;
+  int i = 0, r, c;
+
+  // If mb_to_right_edge is < 0 we are in a situation in which
+  // the current block size extends into the UMV and we won't
+  // visit the sub blocks that are wholly within the UMV.
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+
+  // Keep track of the row and column of the blocks we use so that we know
+  // if we are in the unrestricted motion border.
+  for (r = 0; r < max_blocks_high; r += txh_unit) {
+    // Skip visiting the sub blocks that are wholly within the UMV.
+    for (c = 0; c < max_blocks_wide; c += txw_unit) {
+      visit(plane, i, r, c, plane_bsize, tx_size, arg);
+      i += step;
+    }
+  }
+}
+
+#if CONFIG_LV_MAP
+void av1_foreach_transformed_block(const MACROBLOCKD *const xd,
+                                   BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                   foreach_transformed_block_visitor visit,
+                                   void *arg) {
+  int plane;
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4
+    if (!is_chroma_reference(mi_row, mi_col, bsize,
+                             xd->plane[plane].subsampling_x,
+                             xd->plane[plane].subsampling_y))
+      continue;
+#else
+    (void)mi_row;
+    (void)mi_col;
+#endif
+    av1_foreach_transformed_block_in_plane(xd, bsize, plane, visit, arg);
+  }
+}
+#endif
+
+#if CONFIG_DAALA_DIST
+void av1_foreach_8x8_transformed_block_in_plane(
+    const MACROBLOCKD *const xd, BLOCK_SIZE bsize, int plane,
+    foreach_transformed_block_visitor visit,
+    foreach_transformed_block_visitor mi_visit, void *arg) {
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  // block and transform sizes, in number of 4x4 blocks log 2 ("*_b")
+  // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8
+  // transform size varies per plane, look it up in a common way.
+  const TX_SIZE tx_size = get_tx_size(plane, xd);
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+  const uint8_t txw_unit = tx_size_wide_unit[tx_size];
+  const uint8_t txh_unit = tx_size_high_unit[tx_size];
+  const int step = txw_unit * txh_unit;
+  int i = 0, r, c;
+
+  // If mb_to_right_edge is < 0 we are in a situation in which
+  // the current block size extends into the UMV and we won't
+  // visit the sub blocks that are wholly within the UMV.
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+
+  // Keep track of the row and column of the blocks we use so that we know
+  // if we are in the unrestricted motion border.
+  for (r = 0; r < max_blocks_high; r += txh_unit) {
+    // Skip visiting the sub blocks that are wholly within the UMV.
+    for (c = 0; c < max_blocks_wide; c += txw_unit) {
+      visit(plane, i, r, c, plane_bsize, tx_size, arg);
+      // Call whenever each 8x8 block is done
+      if ((r & 1) && (c & 1))
+        mi_visit(plane, i, r - 1, c - 1, plane_bsize, TX_8X8, arg);
+      i += step;
+    }
+  }
+}
+#endif
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+void av1_set_contexts(const MACROBLOCKD *xd, struct macroblockd_plane *pd,
+                      int plane, TX_SIZE tx_size, int has_eob, int aoff,
+                      int loff) {
+  ENTROPY_CONTEXT *const a = pd->above_context + aoff;
+  ENTROPY_CONTEXT *const l = pd->left_context + loff;
+  const int txs_wide = tx_size_wide_unit[tx_size];
+  const int txs_high = tx_size_high_unit[tx_size];
+#if CONFIG_CB4X4
+  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+#else
+  const BLOCK_SIZE bsize = AOMMAX(xd->mi[0]->mbmi.sb_type, BLOCK_8X8);
+#endif
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+
+  // above
+  if (has_eob && xd->mb_to_right_edge < 0) {
+    int i;
+    const int blocks_wide = max_block_wide(xd, plane_bsize, plane);
+    int above_contexts = txs_wide;
+    if (above_contexts + aoff > blocks_wide)
+      above_contexts = blocks_wide - aoff;
+
+    for (i = 0; i < above_contexts; ++i) a[i] = has_eob;
+    for (i = above_contexts; i < txs_wide; ++i) a[i] = 0;
+  } else {
+    memset(a, has_eob, sizeof(ENTROPY_CONTEXT) * txs_wide);
+  }
+
+  // left
+  if (has_eob && xd->mb_to_bottom_edge < 0) {
+    int i;
+    const int blocks_high = max_block_high(xd, plane_bsize, plane);
+    int left_contexts = txs_high;
+    if (left_contexts + loff > blocks_high) left_contexts = blocks_high - loff;
+
+    for (i = 0; i < left_contexts; ++i) l[i] = has_eob;
+    for (i = left_contexts; i < txs_high; ++i) l[i] = 0;
+  } else {
+    memset(l, has_eob, sizeof(ENTROPY_CONTEXT) * txs_high);
+  }
+}
+#endif
+
+void av1_setup_block_planes(MACROBLOCKD *xd, int ss_x, int ss_y) {
+  int i;
+
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    xd->plane[i].plane_type = get_plane_type(i);
+    xd->plane[i].subsampling_x = i ? ss_x : 0;
+    xd->plane[i].subsampling_y = i ? ss_y : 0;
+  }
+}
+
+#if CONFIG_EXT_INTRA
+const int16_t dr_intra_derivative[90] = {
+  1,    14666, 7330, 4884, 3660, 2926, 2435, 2084, 1821, 1616, 1451, 1317, 1204,
+  1108, 1026,  955,  892,  837,  787,  743,  703,  666,  633,  603,  574,  548,
+  524,  502,   481,  461,  443,  426,  409,  394,  379,  365,  352,  339,  327,
+  316,  305,   294,  284,  274,  265,  256,  247,  238,  230,  222,  214,  207,
+  200,  192,   185,  179,  172,  166,  159,  153,  147,  141,  136,  130,  124,
+  119,  113,   108,  103,  98,   93,   88,   83,   78,   73,   68,   63,   59,
+  54,   49,    45,   40,   35,   31,   26,   22,   17,   13,   8,    4,
+};
+
+#if CONFIG_INTRA_INTERP
+int av1_is_intra_filter_switchable(int angle) {
+  assert(angle > 0 && angle < 270);
+  if (angle % 45 == 0) return 0;
+  if (angle > 90 && angle < 180) {
+    return 1;
+  } else {
+    return ((angle < 90 ? dr_intra_derivative[angle]
+                        : dr_intra_derivative[270 - angle]) &
+            0xFF) > 0;
+  }
+}
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
diff --git a/third_party/aom/av1/common/blockd.h b/third_party/aom/av1/common/blockd.h
new file mode 100644
index 0000000000..0acab965d8
--- /dev/null
+++ b/third_party/aom/av1/common/blockd.h
@@ -0,0 +1,1371 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_BLOCKD_H_
+#define AV1_COMMON_BLOCKD_H_
+
+#include "./aom_config.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+#include "aom_scale/yv12config.h"
+
+#include "av1/common/common_data.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/mv.h"
+#include "av1/common/scale.h"
+#include "av1/common/seg_common.h"
+#include "av1/common/tile_common.h"
+#if CONFIG_PVQ
+#include "av1/common/pvq.h"
+#include "av1/common/pvq_state.h"
+#include "av1/decoder/decint.h"
+#endif
+#if CONFIG_CFL
+#include "av1/common/cfl.h"
+#endif
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define SUB8X8_COMP_REF (!(CONFIG_CB4X4 && CONFIG_CHROMA_2X2))
+
+#define MAX_MB_PLANE 3
+
+#if CONFIG_EXT_INTER
+
+#if CONFIG_COMPOUND_SEGMENT
+// Set COMPOUND_SEGMENT_TYPE to one of the three
+// 0: Uniform
+// 1: Difference weighted
+#define COMPOUND_SEGMENT_TYPE 1
+
+#if COMPOUND_SEGMENT_TYPE == 0
+#define MAX_SEG_MASK_BITS 1
+// SEG_MASK_TYPES should not surpass 1 << MAX_SEG_MASK_BITS
+typedef enum {
+  UNIFORM_45 = 0,
+  UNIFORM_45_INV,
+  SEG_MASK_TYPES,
+} SEG_MASK_TYPE;
+
+#elif COMPOUND_SEGMENT_TYPE == 1
+#define MAX_SEG_MASK_BITS 1
+// SEG_MASK_TYPES should not surpass 1 << MAX_SEG_MASK_BITS
+typedef enum {
+  DIFFWTD_42 = 0,
+  DIFFWTD_42_INV,
+  SEG_MASK_TYPES,
+} SEG_MASK_TYPE;
+
+#endif  // COMPOUND_SEGMENT_TYPE
+#endif  // CONFIG_COMPOUND_SEGMENT
+#endif  // CONFIG_EXT_INTER
+
+typedef enum {
+  KEY_FRAME = 0,
+  INTER_FRAME = 1,
+  FRAME_TYPES,
+} FRAME_TYPE;
+
+static INLINE int is_inter_mode(PREDICTION_MODE mode) {
+#if CONFIG_EXT_INTER
+  return mode >= NEARESTMV && mode <= NEW_NEWMV;
+#else
+  return mode >= NEARESTMV && mode <= NEWMV;
+#endif  // CONFIG_EXT_INTER
+}
+
+#if CONFIG_PVQ
+typedef struct PVQ_INFO {
+  int theta[PVQ_MAX_PARTITIONS];
+  int qg[PVQ_MAX_PARTITIONS];
+  int k[PVQ_MAX_PARTITIONS];
+  od_coeff y[OD_TXSIZE_MAX * OD_TXSIZE_MAX];
+  int nb_bands;
+  int off[PVQ_MAX_PARTITIONS];
+  int size[PVQ_MAX_PARTITIONS];
+  int skip_rest;
+  int skip_dir;
+  int bs;  // log of the block size minus two,
+           // i.e. equivalent to aom's TX_SIZE
+  // Block skip info, indicating whether DC/AC, is coded.
+  PVQ_SKIP_TYPE ac_dc_coded;  // bit0: DC coded, bit1 : AC coded (1 means coded)
+  tran_low_t dq_dc_residue;
+} PVQ_INFO;
+
+typedef struct PVQ_QUEUE {
+  PVQ_INFO *buf;  // buffer for pvq info, stored in encoding order
+  int curr_pos;   // curr position to write PVQ_INFO
+  int buf_len;    // allocated buffer length
+  int last_pos;   // last written position of PVQ_INFO in a tile
+} PVQ_QUEUE;
+#endif
+
+typedef struct {
+  uint8_t *plane[MAX_MB_PLANE];
+  int stride[MAX_MB_PLANE];
+} BUFFER_SET;
+
+#if CONFIG_EXT_INTER
+static INLINE int is_inter_singleref_mode(PREDICTION_MODE mode) {
+  return mode >= NEARESTMV && mode <= NEWMV;
+}
+#if CONFIG_COMPOUND_SINGLEREF
+static INLINE int is_inter_singleref_comp_mode(PREDICTION_MODE mode) {
+  return mode >= SR_NEAREST_NEARMV && mode <= SR_NEW_NEWMV;
+}
+#endif  // CONFIG_COMPOUND_SINGLEREF
+static INLINE int is_inter_compound_mode(PREDICTION_MODE mode) {
+  return mode >= NEAREST_NEARESTMV && mode <= NEW_NEWMV;
+}
+
+static INLINE PREDICTION_MODE compound_ref0_mode(PREDICTION_MODE mode) {
+  static PREDICTION_MODE lut[MB_MODE_COUNT] = {
+    MB_MODE_COUNT,  // DC_PRED
+    MB_MODE_COUNT,  // V_PRED
+    MB_MODE_COUNT,  // H_PRED
+    MB_MODE_COUNT,  // D45_PRED
+    MB_MODE_COUNT,  // D135_PRED
+    MB_MODE_COUNT,  // D117_PRED
+    MB_MODE_COUNT,  // D153_PRED
+    MB_MODE_COUNT,  // D207_PRED
+    MB_MODE_COUNT,  // D63_PRED
+#if CONFIG_ALT_INTRA
+    MB_MODE_COUNT,  // SMOOTH_PRED
+#endif              // CONFIG_ALT_INTRA
+    MB_MODE_COUNT,  // TM_PRED
+    MB_MODE_COUNT,  // NEARESTMV
+    MB_MODE_COUNT,  // NEARMV
+    MB_MODE_COUNT,  // ZEROMV
+    MB_MODE_COUNT,  // NEWMV
+#if CONFIG_COMPOUND_SINGLEREF
+    NEARESTMV,  // SR_NEAREST_NEARMV
+    NEARESTMV,  // SR_NEAREST_NEWMV
+    NEARMV,     // SR_NEAR_NEWMV
+    ZEROMV,     // SR_ZERO_NEWMV
+    NEWMV,      // SR_NEW_NEWMV
+#endif          // CONFIG_COMPOUND_SINGLEREF
+    NEARESTMV,  // NEAREST_NEARESTMV
+    NEARESTMV,  // NEAREST_NEARMV
+    NEARMV,     // NEAR_NEARESTMV
+    NEARMV,     // NEAR_NEARMV
+    NEARESTMV,  // NEAREST_NEWMV
+    NEWMV,      // NEW_NEARESTMV
+    NEARMV,     // NEAR_NEWMV
+    NEWMV,      // NEW_NEARMV
+    ZEROMV,     // ZERO_ZEROMV
+    NEWMV,      // NEW_NEWMV
+  };
+  assert(is_inter_compound_mode(mode));
+  return lut[mode];
+}
+
+static INLINE PREDICTION_MODE compound_ref1_mode(PREDICTION_MODE mode) {
+  static PREDICTION_MODE lut[MB_MODE_COUNT] = {
+    MB_MODE_COUNT,  // DC_PRED
+    MB_MODE_COUNT,  // V_PRED
+    MB_MODE_COUNT,  // H_PRED
+    MB_MODE_COUNT,  // D45_PRED
+    MB_MODE_COUNT,  // D135_PRED
+    MB_MODE_COUNT,  // D117_PRED
+    MB_MODE_COUNT,  // D153_PRED
+    MB_MODE_COUNT,  // D207_PRED
+    MB_MODE_COUNT,  // D63_PRED
+#if CONFIG_ALT_INTRA
+    MB_MODE_COUNT,  // SMOOTH_PRED
+#endif              // CONFIG_ALT_INTRA
+    MB_MODE_COUNT,  // TM_PRED
+    MB_MODE_COUNT,  // NEARESTMV
+    MB_MODE_COUNT,  // NEARMV
+    MB_MODE_COUNT,  // ZEROMV
+    MB_MODE_COUNT,  // NEWMV
+#if CONFIG_COMPOUND_SINGLEREF
+    NEARMV,     // SR_NEAREST_NEARMV
+    NEWMV,      // SR_NEAREST_NEWMV
+    NEWMV,      // SR_NEAR_NEWMV
+    NEWMV,      // SR_ZERO_NEWMV
+    NEWMV,      // SR_NEW_NEWMV
+#endif          // CONFIG_COMPOUND_SINGLEREF
+    NEARESTMV,  // NEAREST_NEARESTMV
+    NEARMV,     // NEAREST_NEARMV
+    NEARESTMV,  // NEAR_NEARESTMV
+    NEARMV,     // NEAR_NEARMV
+    NEWMV,      // NEAREST_NEWMV
+    NEARESTMV,  // NEW_NEARESTMV
+    NEWMV,      // NEAR_NEWMV
+    NEARMV,     // NEW_NEARMV
+    ZEROMV,     // ZERO_ZEROMV
+    NEWMV,      // NEW_NEWMV
+  };
+  assert(is_inter_compound_mode(mode));
+  return lut[mode];
+}
+
+static INLINE int have_nearmv_in_inter_mode(PREDICTION_MODE mode) {
+  return (mode == NEARMV || mode == NEAR_NEARMV || mode == NEAREST_NEARMV ||
+          mode == NEAR_NEARESTMV || mode == NEAR_NEWMV || mode == NEW_NEARMV);
+}
+
+static INLINE int have_newmv_in_inter_mode(PREDICTION_MODE mode) {
+  return (mode == NEWMV || mode == NEW_NEWMV || mode == NEAREST_NEWMV ||
+          mode == NEW_NEARESTMV || mode == NEAR_NEWMV || mode == NEW_NEARMV);
+}
+
+static INLINE int use_masked_motion_search(COMPOUND_TYPE type) {
+#if CONFIG_WEDGE
+  return (type == COMPOUND_WEDGE);
+#else
+  (void)type;
+  return 0;
+#endif
+}
+
+static INLINE int is_masked_compound_type(COMPOUND_TYPE type) {
+#if CONFIG_COMPOUND_SEGMENT && CONFIG_WEDGE
+  return (type == COMPOUND_WEDGE || type == COMPOUND_SEG);
+#elif !CONFIG_COMPOUND_SEGMENT && CONFIG_WEDGE
+  return (type == COMPOUND_WEDGE);
+#elif CONFIG_COMPOUND_SEGMENT && !CONFIG_WEDGE
+  return (type == COMPOUND_SEG);
+#endif  // CONFIG_COMPOUND_SEGMENT
+  (void)type;
+  return 0;
+}
+#else
+
+static INLINE int have_nearmv_in_inter_mode(PREDICTION_MODE mode) {
+  return (mode == NEARMV);
+}
+
+static INLINE int have_newmv_in_inter_mode(PREDICTION_MODE mode) {
+  return (mode == NEWMV);
+}
+#endif  // CONFIG_EXT_INTER
+
+/* For keyframes, intra block modes are predicted by the (already decoded)
+   modes for the Y blocks to the left and above us; for interframes, there
+   is a single probability table. */
+
+typedef struct {
+  PREDICTION_MODE as_mode;
+  int_mv as_mv[2];  // first, second inter predictor motion vectors
+#if CONFIG_REF_MV
+  int_mv pred_mv[2];
+#endif
+#if CONFIG_EXT_INTER
+  int_mv ref_mv[2];
+#endif  // CONFIG_EXT_INTER
+} b_mode_info;
+
+typedef int8_t MV_REFERENCE_FRAME;
+
+#if CONFIG_PALETTE
+typedef struct {
+  // Number of base colors for Y (0) and UV (1)
+  uint8_t palette_size[2];
+// Value of base colors for Y, U, and V
+#if CONFIG_HIGHBITDEPTH
+  uint16_t palette_colors[3 * PALETTE_MAX_SIZE];
+#else
+  uint8_t palette_colors[3 * PALETTE_MAX_SIZE];
+#endif  // CONFIG_HIGHBITDEPTH
+  // Only used by encoder to store the color index of the top left pixel.
+  // TODO(huisu): move this to encoder
+  uint8_t palette_first_color_idx[2];
+} PALETTE_MODE_INFO;
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_FILTER_INTRA
+#define USE_3TAP_INTRA_FILTER 1  // 0: 4-tap; 1: 3-tap
+typedef struct {
+  // 1: an ext intra mode is used; 0: otherwise.
+  uint8_t use_filter_intra_mode[PLANE_TYPES];
+  FILTER_INTRA_MODE filter_intra_mode[PLANE_TYPES];
+} FILTER_INTRA_MODE_INFO;
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_VAR_TX
+#if CONFIG_RD_DEBUG
+#define TXB_COEFF_COST_MAP_SIZE (2 * MAX_MIB_SIZE)
+#endif
+#endif
+
+typedef struct RD_STATS {
+  int rate;
+  int64_t dist;
+  // Please be careful of using rdcost, it's not guaranteed to be set all the
+  // time.
+  // TODO(angiebird): Create a set of functions to manipulate the RD_STATS. In
+  // these functions, make sure rdcost is always up-to-date according to
+  // rate/dist.
+  int64_t rdcost;
+  int64_t sse;
+  int skip;  // sse should equal to dist when skip == 1
+#if CONFIG_RD_DEBUG
+  int txb_coeff_cost[MAX_MB_PLANE];
+#if CONFIG_VAR_TX
+  int txb_coeff_cost_map[MAX_MB_PLANE][TXB_COEFF_COST_MAP_SIZE]
+                        [TXB_COEFF_COST_MAP_SIZE];
+#endif  // CONFIG_VAR_TX
+#endif  // CONFIG_RD_DEBUG
+} RD_STATS;
+
+#if CONFIG_EXT_INTER
+// This struct is used to group function args that are commonly
+// sent together in functions related to interinter compound modes
+typedef struct {
+#if CONFIG_WEDGE
+  int wedge_index;
+  int wedge_sign;
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+  SEG_MASK_TYPE mask_type;
+  uint8_t *seg_mask;
+#endif  // CONFIG_COMPOUND_SEGMENT
+  COMPOUND_TYPE interinter_compound_type;
+} INTERINTER_COMPOUND_DATA;
+#endif  // CONFIG_EXT_INTER
+
+// This structure now relates to 8x8 block regions.
+typedef struct {
+  // Common for both INTER and INTRA blocks
+  BLOCK_SIZE sb_type;
+  PREDICTION_MODE mode;
+  TX_SIZE tx_size;
+#if CONFIG_VAR_TX
+  // TODO(jingning): This effectively assigned a separate entry for each
+  // 8x8 block. Apparently it takes much more space than needed.
+  TX_SIZE inter_tx_size[MAX_MIB_SIZE][MAX_MIB_SIZE];
+  TX_SIZE min_tx_size;
+#endif
+  int8_t skip;
+  int8_t segment_id;
+#if CONFIG_SUPERTX
+  // Minimum of all segment IDs under the current supertx block.
+  int8_t segment_id_supertx;
+#endif                      // CONFIG_SUPERTX
+  int8_t seg_id_predicted;  // valid only when temporal_update is enabled
+
+  // Only for INTRA blocks
+  PREDICTION_MODE uv_mode;
+#if CONFIG_PALETTE
+  PALETTE_MODE_INFO palette_mode_info;
+#endif  // CONFIG_PALETTE
+#if CONFIG_INTRABC
+  uint8_t use_intrabc;
+#endif  // CONFIG_INTRABC
+
+// Only for INTER blocks
+#if CONFIG_DUAL_FILTER
+  InterpFilter interp_filter[4];
+#else
+  InterpFilter interp_filter;
+#endif
+  MV_REFERENCE_FRAME ref_frame[2];
+  TX_TYPE tx_type;
+#if CONFIG_TXK_SEL
+  TX_TYPE txk_type[MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
+#endif
+
+#if CONFIG_FILTER_INTRA
+  FILTER_INTRA_MODE_INFO filter_intra_mode_info;
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_EXT_INTRA
+  // The actual prediction angle is the base angle + (angle_delta * step).
+  int8_t angle_delta[2];
+#if CONFIG_INTRA_INTERP
+  // To-Do (huisu): this may be replaced by interp_filter
+  INTRA_FILTER intra_filter;
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+
+#if CONFIG_EXT_INTER
+  // interintra members
+  INTERINTRA_MODE interintra_mode;
+  // TODO(debargha): Consolidate these flags
+  int use_wedge_interintra;
+  int interintra_wedge_index;
+  int interintra_wedge_sign;
+  // interinter members
+  COMPOUND_TYPE interinter_compound_type;
+#if CONFIG_WEDGE
+  int wedge_index;
+  int wedge_sign;
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+  SEG_MASK_TYPE mask_type;
+#endif  // CONFIG_COMPOUND_SEGMENT
+#endif  // CONFIG_EXT_INTER
+  MOTION_MODE motion_mode;
+#if CONFIG_MOTION_VAR
+  int overlappable_neighbors[2];
+#endif  // CONFIG_MOTION_VAR
+  int_mv mv[2];
+  int_mv pred_mv[2];
+#if CONFIG_REF_MV
+  uint8_t ref_mv_idx;
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+  PARTITION_TYPE partition;
+#endif
+#if CONFIG_NEW_QUANT
+  int dq_off_index;
+  int send_dq_bit;
+#endif  // CONFIG_NEW_QUANT
+  /* deringing gain *per-superblock* */
+  int8_t cdef_strength;
+#if CONFIG_DELTA_Q
+  int current_q_index;
+#if CONFIG_EXT_DELTA_Q
+  int current_delta_lf_from_base;
+#endif
+#endif
+#if CONFIG_RD_DEBUG
+  RD_STATS rd_stats;
+  int mi_row;
+  int mi_col;
+#endif
+#if CONFIG_WARPED_MOTION
+  int num_proj_ref[2];
+  WarpedMotionParams wm_params[2];
+#endif  // CONFIG_WARPED_MOTION
+
+  BOUNDARY_TYPE boundary_info;
+} MB_MODE_INFO;
+
+typedef struct MODE_INFO {
+  MB_MODE_INFO mbmi;
+  b_mode_info bmi[4];
+} MODE_INFO;
+
+#if CONFIG_INTRABC
+static INLINE int is_intrabc_block(const MB_MODE_INFO *mbmi) {
+  return mbmi->use_intrabc;
+}
+#endif
+
+static INLINE PREDICTION_MODE get_y_mode(const MODE_INFO *mi, int block) {
+#if CONFIG_CB4X4
+  (void)block;
+  return mi->mbmi.mode;
+#else
+  return mi->mbmi.sb_type < BLOCK_8X8 ? mi->bmi[block].as_mode : mi->mbmi.mode;
+#endif
+}
+
+static INLINE int is_inter_block(const MB_MODE_INFO *mbmi) {
+#if CONFIG_INTRABC
+  if (is_intrabc_block(mbmi)) return 1;
+#endif
+  return mbmi->ref_frame[0] > INTRA_FRAME;
+}
+
+static INLINE int has_second_ref(const MB_MODE_INFO *mbmi) {
+  return mbmi->ref_frame[1] > INTRA_FRAME;
+}
+
+PREDICTION_MODE av1_left_block_mode(const MODE_INFO *cur_mi,
+                                    const MODE_INFO *left_mi, int b);
+
+PREDICTION_MODE av1_above_block_mode(const MODE_INFO *cur_mi,
+                                     const MODE_INFO *above_mi, int b);
+
+#if CONFIG_GLOBAL_MOTION
+static INLINE int is_global_mv_block(const MODE_INFO *mi, int block,
+                                     TransformationType type) {
+  PREDICTION_MODE mode = get_y_mode(mi, block);
+#if GLOBAL_SUB8X8_USED
+  const int block_size_allowed = 1;
+#else
+  const BLOCK_SIZE bsize = mi->mbmi.sb_type;
+  const int block_size_allowed = (bsize >= BLOCK_8X8);
+#endif  // GLOBAL_SUB8X8_USED
+#if CONFIG_EXT_INTER
+  return (mode == ZEROMV || mode == ZERO_ZEROMV) && type > TRANSLATION &&
+         block_size_allowed;
+#else
+  return mode == ZEROMV && type > TRANSLATION && block_size_allowed;
+#endif  // CONFIG_EXT_INTER
+}
+#endif  // CONFIG_GLOBAL_MOTION
+
+enum mv_precision { MV_PRECISION_Q3, MV_PRECISION_Q4 };
+
+struct buf_2d {
+  uint8_t *buf;
+  uint8_t *buf0;
+  int width;
+  int height;
+  int stride;
+};
+
+typedef struct macroblockd_plane {
+  tran_low_t *dqcoeff;
+  PLANE_TYPE plane_type;
+  int subsampling_x;
+  int subsampling_y;
+  struct buf_2d dst;
+  struct buf_2d pre[2];
+  ENTROPY_CONTEXT *above_context;
+  ENTROPY_CONTEXT *left_context;
+  int16_t seg_dequant[MAX_SEGMENTS][2];
+#if CONFIG_NEW_QUANT
+  dequant_val_type_nuq seg_dequant_nuq[MAX_SEGMENTS][QUANT_PROFILES]
+                                      [COEF_BANDS];
+#endif
+#if CONFIG_PALETTE
+  uint8_t *color_index_map;
+#endif  // CONFIG_PALETTE
+
+  // number of 4x4s in current block
+  uint16_t n4_w, n4_h;
+  // log2 of n4_w, n4_h
+  uint8_t n4_wl, n4_hl;
+  // block size in pixels
+  uint8_t width, height;
+
+#if CONFIG_AOM_QM
+  const qm_val_t *seg_iqmatrix[MAX_SEGMENTS][2][TX_SIZES];
+#endif
+  // encoder
+  const int16_t *dequant;
+#if CONFIG_NEW_QUANT
+  const dequant_val_type_nuq *dequant_val_nuq[QUANT_PROFILES];
+#endif  // CONFIG_NEW_QUANT
+#if CONFIG_AOM_QM
+  const qm_val_t *seg_qmatrix[MAX_SEGMENTS][2][TX_SIZES];
+#endif
+
+#if CONFIG_PVQ || CONFIG_DAALA_DIST
+  DECLARE_ALIGNED(16, int16_t, pred[MAX_SB_SQUARE]);
+  // PVQ: forward transformed predicted image, a reference for PVQ.
+  tran_low_t *pvq_ref_coeff;
+#endif
+} MACROBLOCKD_PLANE;
+
+#define BLOCK_OFFSET(x, i) \
+  ((x) + (i) * (1 << (tx_size_wide_log2[0] + tx_size_high_log2[0])))
+
+typedef struct RefBuffer {
+  // TODO(dkovalev): idx is not really required and should be removed, now it
+  // is used in av1_onyxd_if.c
+  int idx;
+  YV12_BUFFER_CONFIG *buf;
+  struct scale_factors sf;
+} RefBuffer;
+
+typedef int16_t EobThresholdMD[TX_SIZES_ALL][TX_TYPES];
+
+typedef struct macroblockd {
+  struct macroblockd_plane plane[MAX_MB_PLANE];
+  uint8_t bmode_blocks_wl;
+  uint8_t bmode_blocks_hl;
+
+  FRAME_COUNTS *counts;
+  TileInfo tile;
+
+  int mi_stride;
+
+  MODE_INFO **mi;
+  MODE_INFO *left_mi;
+  MODE_INFO *above_mi;
+  MB_MODE_INFO *left_mbmi;
+  MB_MODE_INFO *above_mbmi;
+
+  int up_available;
+  int left_available;
+#if CONFIG_CHROMA_SUB8X8
+  int chroma_up_available;
+  int chroma_left_available;
+#endif
+
+  const aom_prob (*partition_probs)[PARTITION_TYPES - 1];
+
+  /* Distance of MB away from frame edges */
+  int mb_to_left_edge;
+  int mb_to_right_edge;
+  int mb_to_top_edge;
+  int mb_to_bottom_edge;
+
+  FRAME_CONTEXT *fc;
+
+  /* pointers to reference frames */
+  const RefBuffer *block_refs[2];
+
+  /* pointer to current frame */
+  const YV12_BUFFER_CONFIG *cur_buf;
+
+  ENTROPY_CONTEXT *above_context[MAX_MB_PLANE];
+  ENTROPY_CONTEXT left_context[MAX_MB_PLANE][2 * MAX_MIB_SIZE];
+
+  PARTITION_CONTEXT *above_seg_context;
+  PARTITION_CONTEXT left_seg_context[MAX_MIB_SIZE];
+
+#if CONFIG_VAR_TX
+  TXFM_CONTEXT *above_txfm_context;
+  TXFM_CONTEXT *left_txfm_context;
+  TXFM_CONTEXT left_txfm_context_buffer[MAX_MIB_SIZE];
+
+  TX_SIZE max_tx_size;
+#if CONFIG_SUPERTX
+  TX_SIZE supertx_size;
+#endif
+#endif
+
+  // block dimension in the unit of mode_info.
+  uint8_t n8_w, n8_h;
+
+#if CONFIG_REF_MV
+  uint8_t ref_mv_count[MODE_CTX_REF_FRAMES];
+  CANDIDATE_MV ref_mv_stack[MODE_CTX_REF_FRAMES][MAX_REF_MV_STACK_SIZE];
+  uint8_t is_sec_rect;
+#endif
+
+#if CONFIG_PVQ
+  daala_dec_ctx daala_dec;
+#endif
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *tile_ctx;
+#endif
+#if CONFIG_HIGHBITDEPTH
+  /* Bit depth: 8, 10, 12 */
+  int bd;
+#endif
+
+  int qindex[MAX_SEGMENTS];
+  int lossless[MAX_SEGMENTS];
+  int corrupted;
+
+  struct aom_internal_error_info *error_info;
+#if CONFIG_GLOBAL_MOTION
+  WarpedMotionParams *global_motion;
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_DELTA_Q
+  int prev_qindex;
+  int delta_qindex;
+  int current_qindex;
+#if CONFIG_EXT_DELTA_Q
+  // Since actual frame level loop filtering level value is not available
+  // at the beginning of the tile (only available during actual filtering)
+  // at encoder side.we record the delta_lf (against the frame level loop
+  // filtering level) and code the delta between previous superblock's delta
+  // lf and current delta lf. It is equivalent to the delta between previous
+  // superblock's actual lf and current lf.
+  int prev_delta_lf_from_base;
+  int current_delta_lf_from_base;
+#endif
+#endif
+#if CONFIG_ADAPT_SCAN
+  const EobThresholdMD *eob_threshold_md;
+#endif
+
+#if CONFIG_EXT_INTER && CONFIG_COMPOUND_SEGMENT
+  DECLARE_ALIGNED(16, uint8_t, seg_mask[2 * MAX_SB_SQUARE]);
+#endif  // CONFIG_EXT_INTER && CONFIG_COMPOUND_SEGMENT
+
+#if CONFIG_CFL
+  CFL_CTX *cfl;
+#endif
+} MACROBLOCKD;
+
+static INLINE BLOCK_SIZE get_subsize(BLOCK_SIZE bsize,
+                                     PARTITION_TYPE partition) {
+  if (partition == PARTITION_INVALID)
+    return BLOCK_INVALID;
+  else
+    return subsize_lookup[partition][bsize];
+}
+
+static const TX_TYPE intra_mode_to_tx_type_context[INTRA_MODES] = {
+  DCT_DCT,    // DC
+  ADST_DCT,   // V
+  DCT_ADST,   // H
+  DCT_DCT,    // D45
+  ADST_ADST,  // D135
+  ADST_DCT,   // D117
+  DCT_ADST,   // D153
+  DCT_ADST,   // D207
+  ADST_DCT,   // D63
+#if CONFIG_ALT_INTRA
+  ADST_ADST,  // SMOOTH
+#endif        // CONFIG_ALT_INTRA
+  ADST_ADST,  // TM
+};
+
+#if CONFIG_SUPERTX
+static INLINE int supertx_enabled(const MB_MODE_INFO *mbmi) {
+  TX_SIZE max_tx_size = txsize_sqr_map[mbmi->tx_size];
+  return tx_size_wide[max_tx_size] >
+         AOMMIN(block_size_wide[mbmi->sb_type], block_size_high[mbmi->sb_type]);
+}
+#endif  // CONFIG_SUPERTX
+
+#define USE_TXTYPE_SEARCH_FOR_SUB8X8_IN_CB4X4 1
+
+#if CONFIG_RECT_TX
+static INLINE int is_rect_tx(TX_SIZE tx_size) { return tx_size >= TX_SIZES; }
+#endif  // CONFIG_RECT_TX
+
+#if CONFIG_EXT_TX
+#define ALLOW_INTRA_EXT_TX 1
+
+typedef enum {
+  // DCT only
+  EXT_TX_SET_DCTONLY = 0,
+  // DCT + Identity only
+  EXT_TX_SET_DCT_IDTX = 1,
+  // Discrete Trig transforms w/o flip (4) + Identity (1)
+  EXT_TX_SET_DTT4_IDTX = 2,
+  // Discrete Trig transforms w/o flip (4) + Identity (1) + 1D Hor/vert DCT (2)
+  EXT_TX_SET_DTT4_IDTX_1DDCT = 3,
+  // Discrete Trig transforms w/ flip (9) + Identity (1) + 1D Hor/Ver DCT (2)
+  EXT_TX_SET_DTT9_IDTX_1DDCT = 4,
+  // Discrete Trig transforms w/ flip (9) + Identity (1) + 1D Hor/Ver (6)
+  EXT_TX_SET_ALL16 = 5,
+  EXT_TX_SET_TYPES
+} TxSetType;
+
+// Number of transform types in each set type
+static const int num_ext_tx_set[EXT_TX_SET_TYPES] = { 1, 2, 5, 7, 12, 16 };
+
+// Maps intra set index to the set type
+static const int ext_tx_set_type_intra[EXT_TX_SETS_INTRA] = {
+  EXT_TX_SET_DCTONLY, EXT_TX_SET_DTT4_IDTX_1DDCT, EXT_TX_SET_DTT4_IDTX
+};
+
+// Maps inter set index to the set type
+static const int ext_tx_set_type_inter[EXT_TX_SETS_INTER] = {
+  EXT_TX_SET_DCTONLY, EXT_TX_SET_ALL16, EXT_TX_SET_DTT9_IDTX_1DDCT,
+  EXT_TX_SET_DCT_IDTX
+};
+
+// Maps set types above to the indices used for intra
+static const int ext_tx_set_index_intra[EXT_TX_SET_TYPES] = { 0, -1, 2,
+                                                              1, -1, -1 };
+
+// Maps set types above to the indices used for inter
+static const int ext_tx_set_index_inter[EXT_TX_SET_TYPES] = {
+  0, 3, -1, -1, 2, 1
+};
+
+static INLINE TxSetType get_ext_tx_set_type(TX_SIZE tx_size, BLOCK_SIZE bs,
+                                            int is_inter, int use_reduced_set) {
+  const TX_SIZE tx_size2 = txsize_sqr_up_map[tx_size];
+  tx_size = txsize_sqr_map[tx_size];
+#if CONFIG_CB4X4 && USE_TXTYPE_SEARCH_FOR_SUB8X8_IN_CB4X4
+  (void)bs;
+  if (tx_size > TX_32X32) return EXT_TX_SET_DCTONLY;
+#else
+  if (tx_size > TX_32X32 || bs < BLOCK_8X8) return EXT_TX_SET_DCTONLY;
+#endif
+  if (use_reduced_set)
+    return is_inter ? EXT_TX_SET_DCT_IDTX : EXT_TX_SET_DTT4_IDTX;
+  if (tx_size2 == TX_32X32)
+    return is_inter ? EXT_TX_SET_DCT_IDTX : EXT_TX_SET_DCTONLY;
+  if (is_inter)
+    return (tx_size == TX_16X16 ? EXT_TX_SET_DTT9_IDTX_1DDCT
+                                : EXT_TX_SET_ALL16);
+  else
+    return (tx_size == TX_16X16 ? EXT_TX_SET_DTT4_IDTX
+                                : EXT_TX_SET_DTT4_IDTX_1DDCT);
+}
+
+static INLINE int get_ext_tx_set(TX_SIZE tx_size, BLOCK_SIZE bs, int is_inter,
+                                 int use_reduced_set) {
+  const TxSetType set_type =
+      get_ext_tx_set_type(tx_size, bs, is_inter, use_reduced_set);
+  return is_inter ? ext_tx_set_index_inter[set_type]
+                  : ext_tx_set_index_intra[set_type];
+}
+
+static const int use_intra_ext_tx_for_txsize[EXT_TX_SETS_INTRA][EXT_TX_SIZES] =
+    {
+#if CONFIG_CB4X4
+      { 1, 1, 1, 1, 1 },  // unused
+      { 0, 1, 1, 0, 0 },
+      { 0, 0, 0, 1, 0 },
+#else
+      { 1, 1, 1, 1 },  // unused
+      { 1, 1, 0, 0 },
+      { 0, 0, 1, 0 },
+#endif  // CONFIG_CB4X4
+    };
+
+static const int use_inter_ext_tx_for_txsize[EXT_TX_SETS_INTER][EXT_TX_SIZES] =
+    {
+#if CONFIG_CB4X4
+      { 1, 1, 1, 1, 1 },  // unused
+      { 0, 1, 1, 0, 0 },
+      { 0, 0, 0, 1, 0 },
+      { 0, 0, 0, 0, 1 },
+#else
+      { 1, 1, 1, 1 },  // unused
+      { 1, 1, 0, 0 },
+      { 0, 0, 1, 0 },
+      { 0, 0, 0, 1 },
+#endif  // CONFIG_CB4X4
+    };
+
+// Transform types used in each intra set
+static const int ext_tx_used_intra[EXT_TX_SETS_INTRA][TX_TYPES] = {
+  { 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0 },
+  { 1, 1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0 },
+};
+
+// Numbers of transform types used in each intra set
+static const int ext_tx_cnt_intra[EXT_TX_SETS_INTRA] = { 1, 7, 5 };
+
+// Transform types used in each inter set
+static const int ext_tx_used_inter[EXT_TX_SETS_INTER][TX_TYPES] = {
+  { 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 },
+  { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0 },
+  { 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0 },
+};
+
+// Numbers of transform types used in each inter set
+static const int ext_tx_cnt_inter[EXT_TX_SETS_INTER] = { 1, 16, 12, 2 };
+
+// 1D Transforms used in inter set, this needs to be changed if
+// ext_tx_used_inter is changed
+static const int ext_tx_used_inter_1D[EXT_TX_SETS_INTER][TX_TYPES_1D] = {
+  { 1, 0, 0, 0 }, { 1, 1, 1, 1 }, { 1, 1, 1, 1 }, { 1, 0, 0, 1 },
+};
+
+static INLINE int get_ext_tx_types(TX_SIZE tx_size, BLOCK_SIZE bs, int is_inter,
+                                   int use_reduced_set) {
+  const int set_type =
+      get_ext_tx_set_type(tx_size, bs, is_inter, use_reduced_set);
+  return num_ext_tx_set[set_type];
+}
+
+#if CONFIG_RECT_TX
+static INLINE int is_rect_tx_allowed_bsize(BLOCK_SIZE bsize) {
+  static const char LUT[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+    0,  // BLOCK_2X2
+    0,  // BLOCK_2X4
+    0,  // BLOCK_4X2
+#endif
+    0,  // BLOCK_4X4
+    1,  // BLOCK_4X8
+    1,  // BLOCK_8X4
+    0,  // BLOCK_8X8
+    1,  // BLOCK_8X16
+    1,  // BLOCK_16X8
+    0,  // BLOCK_16X16
+    1,  // BLOCK_16X32
+    1,  // BLOCK_32X16
+    0,  // BLOCK_32X32
+    0,  // BLOCK_32X64
+    0,  // BLOCK_64X32
+    0,  // BLOCK_64X64
+#if CONFIG_EXT_PARTITION
+    0,  // BLOCK_64X128
+    0,  // BLOCK_128X64
+    0,  // BLOCK_128X128
+#endif  // CONFIG_EXT_PARTITION
+  };
+
+  return LUT[bsize];
+}
+
+static INLINE int is_rect_tx_allowed(const MACROBLOCKD *xd,
+                                     const MB_MODE_INFO *mbmi) {
+  return is_rect_tx_allowed_bsize(mbmi->sb_type) &&
+         !xd->lossless[mbmi->segment_id];
+}
+#endif  // CONFIG_RECT_TX
+#endif  // CONFIG_EXT_TX
+
+static INLINE TX_SIZE tx_size_from_tx_mode(BLOCK_SIZE bsize, TX_MODE tx_mode,
+                                           int is_inter) {
+  const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+#if (CONFIG_VAR_TX || CONFIG_EXT_TX) && CONFIG_RECT_TX
+  const TX_SIZE max_rect_tx_size = max_txsize_rect_lookup[bsize];
+#else
+  const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
+#endif  // (CONFIG_VAR_TX || CONFIG_EXT_TX) && CONFIG_RECT_TX
+  (void)is_inter;
+#if CONFIG_VAR_TX && CONFIG_RECT_TX
+#if CONFIG_CB4X4
+  if (bsize == BLOCK_4X4)
+    return AOMMIN(max_txsize_lookup[bsize], largest_tx_size);
+#else
+  if (bsize < BLOCK_8X8)
+    return AOMMIN(max_txsize_lookup[bsize], largest_tx_size);
+#endif
+  if (txsize_sqr_map[max_rect_tx_size] <= largest_tx_size)
+    return max_rect_tx_size;
+  else
+    return largest_tx_size;
+#elif CONFIG_EXT_TX && CONFIG_RECT_TX
+  if (txsize_sqr_up_map[max_rect_tx_size] <= largest_tx_size) {
+    return max_rect_tx_size;
+  } else {
+    return largest_tx_size;
+  }
+#else
+  return AOMMIN(max_tx_size, largest_tx_size);
+#endif  // CONFIG_VAR_TX && CONFIG_RECT_TX
+}
+
+#if CONFIG_EXT_INTRA
+#define MAX_ANGLE_DELTA 3
+#define ANGLE_STEP 3
+extern const int16_t dr_intra_derivative[90];
+static const uint8_t mode_to_angle_map[INTRA_MODES] = {
+  0, 90, 180, 45, 135, 111, 157, 203, 67, 0,
+};
+#if CONFIG_INTRA_INTERP
+// Returns whether filter selection is needed for a given
+// intra prediction angle.
+int av1_is_intra_filter_switchable(int angle);
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+
+#define FIXED_TX_TYPE 0
+
+// Converts block_index for given transform size to index of the block in raster
+// order.
+static INLINE int av1_block_index_to_raster_order(TX_SIZE tx_size,
+                                                  int block_idx) {
+  // For transform size 4x8, the possible block_idx values are 0 & 2, because
+  // block_idx values are incremented in steps of size 'tx_width_unit x
+  // tx_height_unit'. But, for this transform size, block_idx = 2 corresponds to
+  // block number 1 in raster order, inside an 8x8 MI block.
+  // For any other transform size, the two indices are equivalent.
+  return (tx_size == TX_4X8 && block_idx == 2) ? 1 : block_idx;
+}
+
+// Inverse of above function.
+// Note: only implemented for transform sizes 4x4, 4x8 and 8x4 right now.
+static INLINE int av1_raster_order_to_block_index(TX_SIZE tx_size,
+                                                  int raster_order) {
+  assert(tx_size == TX_4X4 || tx_size == TX_4X8 || tx_size == TX_8X4);
+  // We ensure that block indices are 0 & 2 if tx size is 4x8 or 8x4.
+  return (tx_size == TX_4X4) ? raster_order : (raster_order > 0) ? 2 : 0;
+}
+
+static INLINE TX_TYPE get_default_tx_type(PLANE_TYPE plane_type,
+                                          const MACROBLOCKD *xd, int block_idx,
+                                          TX_SIZE tx_size) {
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+
+  if (is_inter_block(mbmi) || plane_type != PLANE_TYPE_Y ||
+      xd->lossless[mbmi->segment_id] || tx_size >= TX_32X32)
+    return DCT_DCT;
+
+  return intra_mode_to_tx_type_context[plane_type == PLANE_TYPE_Y
+                                           ? get_y_mode(xd->mi[0], block_idx)
+                                           : mbmi->uv_mode];
+}
+
+static INLINE TX_TYPE get_tx_type(PLANE_TYPE plane_type, const MACROBLOCKD *xd,
+                                  int block, TX_SIZE tx_size) {
+  const MODE_INFO *const mi = xd->mi[0];
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+#if CONFIG_INTRABC
+  // TODO(aconverse@google.com): Revisit this decision
+  if (is_intrabc_block(mbmi)) return DCT_DCT;
+#endif  // CONFIG_INTRABC
+#if !CONFIG_TXK_SEL
+#if FIXED_TX_TYPE
+  const int block_raster_idx = av1_block_index_to_raster_order(tx_size, block);
+  return get_default_tx_type(plane_type, xd, block_raster_idx, tx_size);
+#elif CONFIG_EXT_TX
+#if !CONFIG_CB4X4
+  const int block_raster_idx = av1_block_index_to_raster_order(tx_size, block);
+#endif  // !CONFIG_CB4X4
+  if (xd->lossless[mbmi->segment_id] || txsize_sqr_map[tx_size] > TX_32X32 ||
+      (txsize_sqr_map[tx_size] >= TX_32X32 && !is_inter_block(mbmi)))
+    return DCT_DCT;
+  if (mbmi->sb_type >= BLOCK_8X8 || CONFIG_CB4X4) {
+    if (plane_type == PLANE_TYPE_Y) {
+#if !ALLOW_INTRA_EXT_TX
+      if (is_inter_block(mbmi))
+#endif  // ALLOW_INTRA_EXT_TX
+        return mbmi->tx_type;
+    }
+
+    if (is_inter_block(mbmi)) {
+// UV Inter only
+#if CONFIG_CB4X4
+      if (tx_size < TX_4X4) return DCT_DCT;
+#endif
+      return (mbmi->tx_type == IDTX && txsize_sqr_map[tx_size] >= TX_32X32)
+                 ? DCT_DCT
+                 : mbmi->tx_type;
+    }
+  }
+
+#if CONFIG_CB4X4
+  (void)block;
+  if (tx_size < TX_4X4)
+    return DCT_DCT;
+  else
+    return intra_mode_to_tx_type_context[mbmi->uv_mode];
+#else
+
+  // Sub8x8-Inter/Intra OR UV-Intra
+  if (is_inter_block(mbmi))  // Sub8x8-Inter
+    return DCT_DCT;
+  else  // Sub8x8 Intra OR UV-Intra
+    return intra_mode_to_tx_type_context[plane_type == PLANE_TYPE_Y
+                                             ? get_y_mode(mi, block_raster_idx)
+                                             : mbmi->uv_mode];
+#endif  // CONFIG_CB4X4
+#else   // CONFIG_EXT_TX
+  (void)block;
+  if (plane_type != PLANE_TYPE_Y || xd->lossless[mbmi->segment_id] ||
+      txsize_sqr_map[tx_size] >= TX_32X32)
+    return DCT_DCT;
+  return mbmi->tx_type;
+#endif  // CONFIG_EXT_TX
+#else   // !CONFIG_TXK_SEL
+  (void)tx_size;
+  TX_TYPE tx_type;
+  if (plane_type != PLANE_TYPE_Y || xd->lossless[mbmi->segment_id] ||
+      mbmi->tx_size >= TX_32X32) {
+    tx_type = DCT_DCT;
+  } else {
+    tx_type = mbmi->txk_type[block];
+  }
+  assert(tx_type >= DCT_DCT && tx_type < TX_TYPES);
+  return tx_type;
+#endif  // !CONFIG_TXK_SEL
+}
+
+void av1_setup_block_planes(MACROBLOCKD *xd, int ss_x, int ss_y);
+
+static INLINE int tx_size_to_depth(TX_SIZE tx_size) {
+  return (int)(tx_size - TX_4X4);
+}
+
+static INLINE TX_SIZE depth_to_tx_size(int depth) {
+  return (TX_SIZE)(depth + TX_4X4);
+}
+
+static INLINE TX_SIZE get_uv_tx_size(const MB_MODE_INFO *mbmi,
+                                     const struct macroblockd_plane *pd) {
+  TX_SIZE uv_txsize;
+#if CONFIG_CB4X4
+  assert(mbmi->tx_size > TX_2X2);
+#endif
+
+#if CONFIG_SUPERTX
+  if (supertx_enabled(mbmi))
+    return uvsupertx_size_lookup[txsize_sqr_map[mbmi->tx_size]]
+                                [pd->subsampling_x][pd->subsampling_y];
+#endif  // CONFIG_SUPERTX
+
+  uv_txsize = uv_txsize_lookup[mbmi->sb_type][mbmi->tx_size][pd->subsampling_x]
+                              [pd->subsampling_y];
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+  uv_txsize = AOMMAX(uv_txsize, TX_4X4);
+#endif
+  assert(uv_txsize != TX_INVALID);
+  return uv_txsize;
+}
+
+static INLINE TX_SIZE get_tx_size(int plane, const MACROBLOCKD *xd) {
+  const MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const MACROBLOCKD_PLANE *pd = &xd->plane[plane];
+  const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi, pd) : mbmi->tx_size;
+  return tx_size;
+}
+
+static INLINE BLOCK_SIZE
+get_plane_block_size(BLOCK_SIZE bsize, const struct macroblockd_plane *pd) {
+  return ss_size_lookup[bsize][pd->subsampling_x][pd->subsampling_y];
+}
+
+static INLINE void reset_skip_context(MACROBLOCKD *xd, BLOCK_SIZE bsize) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    struct macroblockd_plane *const pd = &xd->plane[i];
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+    const int txs_wide = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+    const int txs_high = block_size_high[plane_bsize] >> tx_size_high_log2[0];
+    memset(pd->above_context, 0, sizeof(ENTROPY_CONTEXT) * txs_wide);
+    memset(pd->left_context, 0, sizeof(ENTROPY_CONTEXT) * txs_high);
+  }
+}
+
+typedef void (*foreach_transformed_block_visitor)(int plane, int block,
+                                                  int blk_row, int blk_col,
+                                                  BLOCK_SIZE plane_bsize,
+                                                  TX_SIZE tx_size, void *arg);
+
+void av1_foreach_transformed_block_in_plane(
+    const MACROBLOCKD *const xd, BLOCK_SIZE bsize, int plane,
+    foreach_transformed_block_visitor visit, void *arg);
+
+#if CONFIG_LV_MAP
+void av1_foreach_transformed_block(const MACROBLOCKD *const xd,
+                                   BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                   foreach_transformed_block_visitor visit,
+                                   void *arg);
+#endif
+
+#if CONFIG_DAALA_DIST
+void av1_foreach_8x8_transformed_block_in_plane(
+    const MACROBLOCKD *const xd, BLOCK_SIZE bsize, int plane,
+    foreach_transformed_block_visitor visit,
+    foreach_transformed_block_visitor mi_visit, void *arg);
+#endif
+
+#if CONFIG_COEF_INTERLEAVE
+static INLINE int get_max_4x4_size(int num_4x4, int mb_to_edge,
+                                   int subsampling) {
+  return num_4x4 + (mb_to_edge >= 0 ? 0 : mb_to_edge >> (5 + subsampling));
+}
+
+void av1_foreach_transformed_block_interleave(
+    const MACROBLOCKD *const xd, BLOCK_SIZE bsize,
+    foreach_transformed_block_visitor visit, void *arg);
+#endif
+
+void av1_set_contexts(const MACROBLOCKD *xd, struct macroblockd_plane *pd,
+                      int plane, TX_SIZE tx_size, int has_eob, int aoff,
+                      int loff);
+
+#if CONFIG_EXT_INTER
+static INLINE int is_interintra_allowed_bsize(const BLOCK_SIZE bsize) {
+#if CONFIG_INTERINTRA
+  // TODO(debargha): Should this be bsize < BLOCK_LARGEST?
+  return (bsize >= BLOCK_8X8) && (bsize < BLOCK_64X64);
+#else
+  (void)bsize;
+  return 0;
+#endif  // CONFIG_INTERINTRA
+}
+
+static INLINE int is_interintra_allowed_mode(const PREDICTION_MODE mode) {
+#if CONFIG_INTERINTRA
+  return (mode >= NEARESTMV) && (mode <= NEWMV);
+#else
+  (void)mode;
+  return 0;
+#endif  // CONFIG_INTERINTRA
+}
+
+static INLINE int is_interintra_allowed_ref(const MV_REFERENCE_FRAME rf[2]) {
+#if CONFIG_INTERINTRA
+  return (rf[0] > INTRA_FRAME) && (rf[1] <= INTRA_FRAME);
+#else
+  (void)rf;
+  return 0;
+#endif  // CONFIG_INTERINTRA
+}
+
+static INLINE int is_interintra_allowed(const MB_MODE_INFO *mbmi) {
+  return is_interintra_allowed_bsize(mbmi->sb_type) &&
+         is_interintra_allowed_mode(mbmi->mode) &&
+         is_interintra_allowed_ref(mbmi->ref_frame);
+}
+
+static INLINE int is_interintra_allowed_bsize_group(int group) {
+  int i;
+  for (i = 0; i < BLOCK_SIZES; i++) {
+    if (size_group_lookup[i] == group &&
+        is_interintra_allowed_bsize((BLOCK_SIZE)i)) {
+      return 1;
+    }
+  }
+  return 0;
+}
+
+static INLINE int is_interintra_pred(const MB_MODE_INFO *mbmi) {
+  return (mbmi->ref_frame[1] == INTRA_FRAME) && is_interintra_allowed(mbmi);
+}
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_VAR_TX
+static INLINE int get_vartx_max_txsize(const MB_MODE_INFO *const mbmi,
+                                       BLOCK_SIZE bsize) {
+#if CONFIG_CB4X4
+  (void)mbmi;
+  return max_txsize_rect_lookup[bsize];
+#endif  // CONFIG_C4X4
+  return mbmi->sb_type < BLOCK_8X8 ? max_txsize_rect_lookup[mbmi->sb_type]
+                                   : max_txsize_rect_lookup[bsize];
+}
+#endif  // CONFIG_VAR_TX
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+static INLINE int is_motion_variation_allowed_bsize(BLOCK_SIZE bsize) {
+  return (bsize >= BLOCK_8X8);
+}
+
+static INLINE int is_motion_variation_allowed_compound(
+    const MB_MODE_INFO *mbmi) {
+  if (!has_second_ref(mbmi))
+    return 1;
+  else
+    return 0;
+}
+
+#if CONFIG_MOTION_VAR
+// input: log2 of length, 0(4), 1(8), ...
+static const int max_neighbor_obmc[6] = { 0, 1, 2, 3, 4, 4 };
+
+static INLINE int check_num_overlappable_neighbors(const MB_MODE_INFO *mbmi) {
+  return !(mbmi->overlappable_neighbors[0] == 0 &&
+           mbmi->overlappable_neighbors[1] == 0);
+}
+#endif
+
+static INLINE MOTION_MODE motion_mode_allowed(
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+    int block, const WarpedMotionParams *gm_params,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+    const MODE_INFO *mi) {
+  const MB_MODE_INFO *mbmi = &mi->mbmi;
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+  const TransformationType gm_type = gm_params[mbmi->ref_frame[0]].wmtype;
+  if (is_global_mv_block(mi, block, gm_type)) return SIMPLE_TRANSLATION;
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+#if CONFIG_EXT_INTER
+  if (is_motion_variation_allowed_bsize(mbmi->sb_type) &&
+      is_inter_mode(mbmi->mode) && mbmi->ref_frame[1] != INTRA_FRAME &&
+      is_motion_variation_allowed_compound(mbmi)) {
+#else
+  if (is_motion_variation_allowed_bsize(mbmi->sb_type) &&
+      is_inter_mode(mbmi->mode) && is_motion_variation_allowed_compound(mbmi)) {
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR
+    if (!check_num_overlappable_neighbors(mbmi)) return SIMPLE_TRANSLATION;
+#endif
+#if CONFIG_WARPED_MOTION
+    if (!has_second_ref(mbmi) && mbmi->num_proj_ref[0] >= 1)
+      return WARPED_CAUSAL;
+    else
+#endif  // CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+      return OBMC_CAUSAL;
+#else
+    return SIMPLE_TRANSLATION;
+#endif  // CONFIG_MOTION_VAR
+  } else {
+    return SIMPLE_TRANSLATION;
+  }
+}
+
+static INLINE void assert_motion_mode_valid(MOTION_MODE mode,
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+                                            int block,
+                                            const WarpedMotionParams *gm_params,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+                                            const MODE_INFO *mi) {
+  const MOTION_MODE last_motion_mode_allowed = motion_mode_allowed(
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      block, gm_params,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      mi);
+  // Check that the input mode is not illegal
+  if (last_motion_mode_allowed < mode)
+    assert(0 && "Illegal motion mode selected");
+}
+
+#if CONFIG_MOTION_VAR
+static INLINE int is_neighbor_overlappable(const MB_MODE_INFO *mbmi) {
+  return (is_inter_block(mbmi));
+}
+#endif  // CONFIG_MOTION_VAR
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+// Returns sub-sampled dimensions of the given block.
+// The output values for 'rows_within_bounds' and 'cols_within_bounds' will
+// differ from 'height' and 'width' when part of the block is outside the right
+// and/or bottom image boundary.
+static INLINE void av1_get_block_dimensions(BLOCK_SIZE bsize, int plane,
+                                            const MACROBLOCKD *xd, int *width,
+                                            int *height,
+                                            int *rows_within_bounds,
+                                            int *cols_within_bounds) {
+  const int block_height = block_size_high[bsize];
+  const int block_width = block_size_wide[bsize];
+  const int block_rows = (xd->mb_to_bottom_edge >= 0)
+                             ? block_height
+                             : (xd->mb_to_bottom_edge >> 3) + block_height;
+  const int block_cols = (xd->mb_to_right_edge >= 0)
+                             ? block_width
+                             : (xd->mb_to_right_edge >> 3) + block_width;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  assert(IMPLIES(plane == PLANE_TYPE_Y, pd->subsampling_x == 0));
+  assert(IMPLIES(plane == PLANE_TYPE_Y, pd->subsampling_y == 0));
+  assert(block_width >= block_cols);
+  assert(block_height >= block_rows);
+  if (width) *width = block_width >> pd->subsampling_x;
+  if (height) *height = block_height >> pd->subsampling_y;
+  if (rows_within_bounds) *rows_within_bounds = block_rows >> pd->subsampling_y;
+  if (cols_within_bounds) *cols_within_bounds = block_cols >> pd->subsampling_x;
+}
+
+#if CONFIG_GLOBAL_MOTION
+static INLINE int is_nontrans_global_motion(const MACROBLOCKD *xd) {
+  const MODE_INFO *mi = xd->mi[0];
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  int ref;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+  // First check if all modes are ZEROMV
+  if (mbmi->sb_type >= BLOCK_8X8 || unify_bsize) {
+#if CONFIG_EXT_INTER
+    if (mbmi->mode != ZEROMV && mbmi->mode != ZERO_ZEROMV) return 0;
+#else
+    if (mbmi->mode != ZEROMV) return 0;
+#endif  // CONFIG_EXT_INTER
+  } else {
+#if CONFIG_EXT_INTER
+    if (mi->bmi[0].as_mode != ZEROMV || mi->bmi[1].as_mode != ZEROMV ||
+        mi->bmi[2].as_mode != ZEROMV || mi->bmi[3].as_mode != ZEROMV ||
+        mi->bmi[0].as_mode != ZERO_ZEROMV ||
+        mi->bmi[1].as_mode != ZERO_ZEROMV ||
+        mi->bmi[2].as_mode != ZERO_ZEROMV || mi->bmi[3].as_mode != ZERO_ZEROMV)
+      return 0;
+#else
+    if (mi->bmi[0].as_mode != ZEROMV || mi->bmi[1].as_mode != ZEROMV ||
+        mi->bmi[2].as_mode != ZEROMV || mi->bmi[3].as_mode != ZEROMV)
+      return 0;
+#endif  // CONFIG_EXT_INTER
+  }
+
+#if !GLOBAL_SUB8X8_USED
+  if (mbmi->sb_type < BLOCK_8X8) return 0;
+#endif
+
+  // Now check if all global motion is non translational
+  for (ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) {
+    if (xd->global_motion[mbmi->ref_frame[ref]].wmtype <= TRANSLATION) return 0;
+  }
+  return 1;
+}
+#endif  // CONFIG_GLOBAL_MOTION
+
+static INLINE PLANE_TYPE get_plane_type(int plane) {
+  return (plane == 0) ? PLANE_TYPE_Y : PLANE_TYPE_UV;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_BLOCKD_H_
diff --git a/third_party/aom/av1/common/cdef.c b/third_party/aom/av1/common/cdef.c
new file mode 100644
index 0000000000..53dff98b78
--- /dev/null
+++ b/third_party/aom/av1/common/cdef.c
@@ -0,0 +1,445 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <string.h>
+
+#include "./aom_scale_rtcd.h"
+#include "aom/aom_integer.h"
+#include "av1/common/cdef.h"
+#include "av1/common/od_dering.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/reconinter.h"
+
+int sb_all_skip(const AV1_COMMON *const cm, int mi_row, int mi_col) {
+  int r, c;
+  int maxc, maxr;
+  int skip = 1;
+  maxc = cm->mi_cols - mi_col;
+  maxr = cm->mi_rows - mi_row;
+#if CONFIG_EXT_PARTITION
+  if (maxr > cm->mib_size_log2) maxr = cm->mib_size_log2;
+  if (maxc > cm->mib_size_log2) maxc = cm->mib_size_log2;
+#else
+  if (maxr > MAX_MIB_SIZE) maxr = MAX_MIB_SIZE;
+  if (maxc > MAX_MIB_SIZE) maxc = MAX_MIB_SIZE;
+#endif
+
+  for (r = 0; r < maxr; r++) {
+    for (c = 0; c < maxc; c++) {
+      skip = skip &&
+             cm->mi_grid_visible[(mi_row + r) * cm->mi_stride + mi_col + c]
+                 ->mbmi.skip;
+    }
+  }
+  return skip;
+}
+
+static int is_8x8_block_skip(MODE_INFO **grid, int mi_row, int mi_col,
+                             int mi_stride) {
+  int is_skip = 1;
+  for (int r = 0; r < mi_size_high[BLOCK_8X8]; ++r)
+    for (int c = 0; c < mi_size_wide[BLOCK_8X8]; ++c)
+      is_skip &= grid[(mi_row + r) * mi_stride + (mi_col + c)]->mbmi.skip;
+
+  return is_skip;
+}
+
+int sb_compute_dering_list(const AV1_COMMON *const cm, int mi_row, int mi_col,
+                           dering_list *dlist, int filter_skip) {
+  int r, c;
+  int maxc, maxr;
+  MODE_INFO **grid;
+  int count = 0;
+  grid = cm->mi_grid_visible;
+  maxc = cm->mi_cols - mi_col;
+  maxr = cm->mi_rows - mi_row;
+#if CONFIG_EXT_PARTITION
+  if (maxr > cm->mib_size_log2) maxr = cm->mib_size_log2;
+  if (maxc > cm->mib_size_log2) maxc = cm->mib_size_log2;
+#else
+  if (maxr > MAX_MIB_SIZE) maxr = MAX_MIB_SIZE;
+  if (maxc > MAX_MIB_SIZE) maxc = MAX_MIB_SIZE;
+#endif
+
+  const int r_step = mi_size_high[BLOCK_8X8];
+  const int c_step = mi_size_wide[BLOCK_8X8];
+  const int r_shift = (r_step == 2);
+  const int c_shift = (c_step == 2);
+
+  assert(r_step == 1 || r_step == 2);
+  assert(c_step == 1 || c_step == 2);
+
+  if (filter_skip) {
+    for (r = 0; r < maxr; r += r_step) {
+      for (c = 0; c < maxc; c += c_step) {
+        dlist[count].by = r >> r_shift;
+        dlist[count].bx = c >> c_shift;
+        dlist[count].skip =
+            is_8x8_block_skip(grid, mi_row + r, mi_col + c, cm->mi_stride);
+        count++;
+      }
+    }
+  } else {
+    for (r = 0; r < maxr; r += r_step) {
+      for (c = 0; c < maxc; c += c_step) {
+        if (!is_8x8_block_skip(grid, mi_row + r, mi_col + c, cm->mi_stride)) {
+          dlist[count].by = r >> r_shift;
+          dlist[count].bx = c >> c_shift;
+          dlist[count].skip = 0;
+          count++;
+        }
+      }
+    }
+  }
+  return count;
+}
+
+void copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src,
+                                int sstride, int v, int h) {
+  int i, j;
+  for (i = 0; i < v; i++) {
+    for (j = 0; j < h; j++) {
+      dst[i * dstride + j] = src[i * sstride + j];
+    }
+  }
+}
+
+void copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride,
+                                 const uint16_t *src, int sstride, int v,
+                                 int h) {
+  int i, j;
+  for (i = 0; i < v; i++) {
+    for (j = 0; j < h; j++) {
+      dst[i * dstride + j] = src[i * sstride + j];
+    }
+  }
+}
+
+void copy_sb8_16(UNUSED AV1_COMMON *cm, uint16_t *dst, int dstride,
+                 const uint8_t *src, int src_voffset, int src_hoffset,
+                 int sstride, int vsize, int hsize) {
+#if CONFIG_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    const uint16_t *base =
+        &CONVERT_TO_SHORTPTR(src)[src_voffset * sstride + src_hoffset];
+    copy_rect8_16bit_to_16bit(dst, dstride, base, sstride, vsize, hsize);
+  } else {
+#endif
+    const uint8_t *base = &src[src_voffset * sstride + src_hoffset];
+    copy_rect8_8bit_to_16bit(dst, dstride, base, sstride, vsize, hsize);
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif
+}
+
+static INLINE void fill_rect(uint16_t *dst, int dstride, int v, int h,
+                             uint16_t x) {
+  int i, j;
+  for (i = 0; i < v; i++) {
+    for (j = 0; j < h; j++) {
+      dst[i * dstride + j] = x;
+    }
+  }
+}
+
+static INLINE void copy_rect(uint16_t *dst, int dstride, const uint16_t *src,
+                             int sstride, int v, int h) {
+  int i, j;
+  for (i = 0; i < v; i++) {
+    for (j = 0; j < h; j++) {
+      dst[i * dstride + j] = src[i * sstride + j];
+    }
+  }
+}
+
+void av1_cdef_frame(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
+                    MACROBLOCKD *xd) {
+  int sbr, sbc;
+  int nhsb, nvsb;
+  uint16_t src[OD_DERING_INBUF_SIZE];
+  uint16_t *linebuf[3];
+  uint16_t *colbuf[3];
+  dering_list dlist[MAX_MIB_SIZE * MAX_MIB_SIZE];
+  unsigned char *row_dering, *prev_row_dering, *curr_row_dering;
+  int dering_count;
+  int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS] = { { 0 } };
+  int var[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS] = { { 0 } };
+  int stride;
+  int mi_wide_l2[3];
+  int mi_high_l2[3];
+  int xdec[3];
+  int ydec[3];
+  int pli;
+  int dering_left;
+  int coeff_shift = AOMMAX(cm->bit_depth - 8, 0);
+  int nplanes = 3;
+  int chroma_dering =
+      xd->plane[1].subsampling_x == xd->plane[1].subsampling_y &&
+      xd->plane[2].subsampling_x == xd->plane[2].subsampling_y;
+  nvsb = (cm->mi_rows + MAX_MIB_SIZE - 1) / MAX_MIB_SIZE;
+  nhsb = (cm->mi_cols + MAX_MIB_SIZE - 1) / MAX_MIB_SIZE;
+  av1_setup_dst_planes(xd->plane, cm->sb_size, frame, 0, 0);
+  row_dering = aom_malloc(sizeof(*row_dering) * (nhsb + 2) * 2);
+  memset(row_dering, 1, sizeof(*row_dering) * (nhsb + 2) * 2);
+  prev_row_dering = row_dering + 1;
+  curr_row_dering = prev_row_dering + nhsb + 2;
+  for (pli = 0; pli < nplanes; pli++) {
+    xdec[pli] = xd->plane[pli].subsampling_x;
+    ydec[pli] = xd->plane[pli].subsampling_y;
+    mi_wide_l2[pli] = MI_SIZE_LOG2 - xd->plane[pli].subsampling_x;
+    mi_high_l2[pli] = MI_SIZE_LOG2 - xd->plane[pli].subsampling_y;
+  }
+  stride = (cm->mi_cols << MI_SIZE_LOG2) + 2 * OD_FILT_HBORDER;
+  for (pli = 0; pli < nplanes; pli++) {
+    linebuf[pli] = aom_malloc(sizeof(*linebuf) * OD_FILT_VBORDER * stride);
+    colbuf[pli] =
+        aom_malloc(sizeof(*colbuf) *
+                   ((MAX_SB_SIZE << mi_high_l2[pli]) + 2 * OD_FILT_VBORDER) *
+                   OD_FILT_HBORDER);
+  }
+  for (sbr = 0; sbr < nvsb; sbr++) {
+    for (pli = 0; pli < nplanes; pli++) {
+      const int block_height =
+          (MAX_MIB_SIZE << mi_high_l2[pli]) + 2 * OD_FILT_VBORDER;
+      fill_rect(colbuf[pli], OD_FILT_HBORDER, block_height, OD_FILT_HBORDER,
+                OD_DERING_VERY_LARGE);
+    }
+    dering_left = 1;
+    for (sbc = 0; sbc < nhsb; sbc++) {
+      int level, clpf_strength;
+      int uv_level, uv_clpf_strength;
+      int nhb, nvb;
+      int cstart = 0;
+      curr_row_dering[sbc] = 0;
+      if (cm->mi_grid_visible[MAX_MIB_SIZE * sbr * cm->mi_stride +
+                              MAX_MIB_SIZE * sbc] == NULL ||
+          cm->mi_grid_visible[MAX_MIB_SIZE * sbr * cm->mi_stride +
+                              MAX_MIB_SIZE * sbc]
+                  ->mbmi.cdef_strength == -1) {
+        dering_left = 0;
+        continue;
+      }
+      if (!dering_left) cstart = -OD_FILT_HBORDER;
+      nhb = AOMMIN(MAX_MIB_SIZE, cm->mi_cols - MAX_MIB_SIZE * sbc);
+      nvb = AOMMIN(MAX_MIB_SIZE, cm->mi_rows - MAX_MIB_SIZE * sbr);
+      int tile_top, tile_left, tile_bottom, tile_right;
+      int mi_idx = MAX_MIB_SIZE * sbr * cm->mi_stride + MAX_MIB_SIZE * sbc;
+      BOUNDARY_TYPE boundary_tl =
+          cm->mi_grid_visible[MAX_MIB_SIZE * sbr * cm->mi_stride +
+                              MAX_MIB_SIZE * sbc]
+              ->mbmi.boundary_info;
+      tile_top = boundary_tl & TILE_ABOVE_BOUNDARY;
+      tile_left = boundary_tl & TILE_LEFT_BOUNDARY;
+      /* Right and bottom information appear unreliable, so we use the top
+         and left flags for the next superblocks. */
+      if (sbr != nvsb - 1 &&
+          cm->mi_grid_visible[mi_idx + MAX_MIB_SIZE * cm->mi_stride])
+        tile_bottom = cm->mi_grid_visible[mi_idx + MAX_MIB_SIZE * cm->mi_stride]
+                          ->mbmi.boundary_info &
+                      TILE_ABOVE_BOUNDARY;
+      else
+        tile_bottom = 1;
+      if (sbc != nhsb - 1 && cm->mi_grid_visible[mi_idx + MAX_MIB_SIZE])
+        tile_right =
+            cm->mi_grid_visible[mi_idx + MAX_MIB_SIZE]->mbmi.boundary_info &
+            TILE_LEFT_BOUNDARY;
+      else
+        tile_right = 1;
+      const int mbmi_cdef_strength =
+          cm->mi_grid_visible[MAX_MIB_SIZE * sbr * cm->mi_stride +
+                              MAX_MIB_SIZE * sbc]
+              ->mbmi.cdef_strength;
+      level = cm->cdef_strengths[mbmi_cdef_strength] / CLPF_STRENGTHS;
+      clpf_strength = cm->cdef_strengths[mbmi_cdef_strength] % CLPF_STRENGTHS;
+      clpf_strength += clpf_strength == 3;
+      uv_level = cm->cdef_uv_strengths[mbmi_cdef_strength] / CLPF_STRENGTHS;
+      uv_clpf_strength =
+          cm->cdef_uv_strengths[mbmi_cdef_strength] % CLPF_STRENGTHS;
+      uv_clpf_strength += uv_clpf_strength == 3;
+      if ((level == 0 && clpf_strength == 0 && uv_level == 0 &&
+           uv_clpf_strength == 0) ||
+          (dering_count = sb_compute_dering_list(
+               cm, sbr * MAX_MIB_SIZE, sbc * MAX_MIB_SIZE, dlist,
+               get_filter_skip(level) || get_filter_skip(uv_level))) == 0) {
+        dering_left = 0;
+        continue;
+      }
+
+      curr_row_dering[sbc] = 1;
+      for (pli = 0; pli < nplanes; pli++) {
+        uint16_t dst[MAX_SB_SIZE * MAX_SB_SIZE];
+        int coffset;
+        int rend, cend;
+        int clpf_damping = cm->cdef_clpf_damping;
+        int dering_damping = cm->cdef_dering_damping;
+        int hsize = nhb << mi_wide_l2[pli];
+        int vsize = nvb << mi_high_l2[pli];
+
+        if (pli) {
+          if (chroma_dering)
+            level = uv_level;
+          else
+            level = 0;
+          clpf_strength = uv_clpf_strength;
+        }
+
+        if (sbc == nhsb - 1)
+          cend = hsize;
+        else
+          cend = hsize + OD_FILT_HBORDER;
+
+        if (sbr == nvsb - 1)
+          rend = vsize;
+        else
+          rend = vsize + OD_FILT_VBORDER;
+
+        coffset = sbc * MAX_MIB_SIZE << mi_wide_l2[pli];
+        if (sbc == nhsb - 1) {
+          /* On the last superblock column, fill in the right border with
+             OD_DERING_VERY_LARGE to avoid filtering with the outside. */
+          fill_rect(&src[cend + OD_FILT_HBORDER], OD_FILT_BSTRIDE,
+                    rend + OD_FILT_VBORDER, hsize + OD_FILT_HBORDER - cend,
+                    OD_DERING_VERY_LARGE);
+        }
+        if (sbr == nvsb - 1) {
+          /* On the last superblock row, fill in the bottom border with
+             OD_DERING_VERY_LARGE to avoid filtering with the outside. */
+          fill_rect(&src[(rend + OD_FILT_VBORDER) * OD_FILT_BSTRIDE],
+                    OD_FILT_BSTRIDE, OD_FILT_VBORDER,
+                    hsize + 2 * OD_FILT_HBORDER, OD_DERING_VERY_LARGE);
+        }
+        /* Copy in the pixels we need from the current superblock for
+           deringing.*/
+        copy_sb8_16(
+            cm,
+            &src[OD_FILT_VBORDER * OD_FILT_BSTRIDE + OD_FILT_HBORDER + cstart],
+            OD_FILT_BSTRIDE, xd->plane[pli].dst.buf,
+            (MAX_MIB_SIZE << mi_high_l2[pli]) * sbr, coffset + cstart,
+            xd->plane[pli].dst.stride, rend, cend - cstart);
+        if (!prev_row_dering[sbc]) {
+          copy_sb8_16(cm, &src[OD_FILT_HBORDER], OD_FILT_BSTRIDE,
+                      xd->plane[pli].dst.buf,
+                      (MAX_MIB_SIZE << mi_high_l2[pli]) * sbr - OD_FILT_VBORDER,
+                      coffset, xd->plane[pli].dst.stride, OD_FILT_VBORDER,
+                      hsize);
+        } else if (sbr > 0) {
+          copy_rect(&src[OD_FILT_HBORDER], OD_FILT_BSTRIDE,
+                    &linebuf[pli][coffset], stride, OD_FILT_VBORDER, hsize);
+        } else {
+          fill_rect(&src[OD_FILT_HBORDER], OD_FILT_BSTRIDE, OD_FILT_VBORDER,
+                    hsize, OD_DERING_VERY_LARGE);
+        }
+        if (!prev_row_dering[sbc - 1]) {
+          copy_sb8_16(cm, src, OD_FILT_BSTRIDE, xd->plane[pli].dst.buf,
+                      (MAX_MIB_SIZE << mi_high_l2[pli]) * sbr - OD_FILT_VBORDER,
+                      coffset - OD_FILT_HBORDER, xd->plane[pli].dst.stride,
+                      OD_FILT_VBORDER, OD_FILT_HBORDER);
+        } else if (sbr > 0 && sbc > 0) {
+          copy_rect(src, OD_FILT_BSTRIDE,
+                    &linebuf[pli][coffset - OD_FILT_HBORDER], stride,
+                    OD_FILT_VBORDER, OD_FILT_HBORDER);
+        } else {
+          fill_rect(src, OD_FILT_BSTRIDE, OD_FILT_VBORDER, OD_FILT_HBORDER,
+                    OD_DERING_VERY_LARGE);
+        }
+        if (!prev_row_dering[sbc + 1]) {
+          copy_sb8_16(cm, &src[OD_FILT_HBORDER + (nhb << mi_wide_l2[pli])],
+                      OD_FILT_BSTRIDE, xd->plane[pli].dst.buf,
+                      (MAX_MIB_SIZE << mi_high_l2[pli]) * sbr - OD_FILT_VBORDER,
+                      coffset + hsize, xd->plane[pli].dst.stride,
+                      OD_FILT_VBORDER, OD_FILT_HBORDER);
+        } else if (sbr > 0 && sbc < nhsb - 1) {
+          copy_rect(&src[hsize + OD_FILT_HBORDER], OD_FILT_BSTRIDE,
+                    &linebuf[pli][coffset + hsize], stride, OD_FILT_VBORDER,
+                    OD_FILT_HBORDER);
+        } else {
+          fill_rect(&src[hsize + OD_FILT_HBORDER], OD_FILT_BSTRIDE,
+                    OD_FILT_VBORDER, OD_FILT_HBORDER, OD_DERING_VERY_LARGE);
+        }
+        if (dering_left) {
+          /* If we deringed the superblock on the left then we need to copy in
+             saved pixels. */
+          copy_rect(src, OD_FILT_BSTRIDE, colbuf[pli], OD_FILT_HBORDER,
+                    rend + OD_FILT_VBORDER, OD_FILT_HBORDER);
+        }
+        /* Saving pixels in case we need to dering the superblock on the
+            right. */
+        copy_rect(colbuf[pli], OD_FILT_HBORDER, src + hsize, OD_FILT_BSTRIDE,
+                  rend + OD_FILT_VBORDER, OD_FILT_HBORDER);
+        copy_sb8_16(
+            cm, &linebuf[pli][coffset], stride, xd->plane[pli].dst.buf,
+            (MAX_MIB_SIZE << mi_high_l2[pli]) * (sbr + 1) - OD_FILT_VBORDER,
+            coffset, xd->plane[pli].dst.stride, OD_FILT_VBORDER, hsize);
+
+        if (tile_top) {
+          fill_rect(src, OD_FILT_BSTRIDE, OD_FILT_VBORDER,
+                    hsize + 2 * OD_FILT_HBORDER, OD_DERING_VERY_LARGE);
+        }
+        if (tile_left) {
+          fill_rect(src, OD_FILT_BSTRIDE, vsize + 2 * OD_FILT_VBORDER,
+                    OD_FILT_HBORDER, OD_DERING_VERY_LARGE);
+        }
+        if (tile_bottom) {
+          fill_rect(&src[(vsize + OD_FILT_VBORDER) * OD_FILT_BSTRIDE],
+                    OD_FILT_BSTRIDE, OD_FILT_VBORDER,
+                    hsize + 2 * OD_FILT_HBORDER, OD_DERING_VERY_LARGE);
+        }
+        if (tile_right) {
+          fill_rect(&src[hsize + OD_FILT_HBORDER], OD_FILT_BSTRIDE,
+                    vsize + 2 * OD_FILT_VBORDER, OD_FILT_HBORDER,
+                    OD_DERING_VERY_LARGE);
+        }
+#if CONFIG_HIGHBITDEPTH
+        if (cm->use_highbitdepth) {
+          od_dering(
+              (uint8_t *)&CONVERT_TO_SHORTPTR(
+                  xd->plane[pli]
+                      .dst.buf)[xd->plane[pli].dst.stride *
+                                    (MAX_MIB_SIZE * sbr << mi_high_l2[pli]) +
+                                (sbc * MAX_MIB_SIZE << mi_wide_l2[pli])],
+              xd->plane[pli].dst.stride, dst,
+              &src[OD_FILT_VBORDER * OD_FILT_BSTRIDE + OD_FILT_HBORDER],
+              xdec[pli], ydec[pli], dir, NULL, var, pli, dlist, dering_count,
+              level, clpf_strength, clpf_damping, dering_damping, coeff_shift,
+              0, 1);
+        } else {
+#endif
+          od_dering(&xd->plane[pli]
+                         .dst.buf[xd->plane[pli].dst.stride *
+                                      (MAX_MIB_SIZE * sbr << mi_high_l2[pli]) +
+                                  (sbc * MAX_MIB_SIZE << mi_wide_l2[pli])],
+                    xd->plane[pli].dst.stride, dst,
+                    &src[OD_FILT_VBORDER * OD_FILT_BSTRIDE + OD_FILT_HBORDER],
+                    xdec[pli], ydec[pli], dir, NULL, var, pli, dlist,
+                    dering_count, level, clpf_strength, clpf_damping,
+                    dering_damping, coeff_shift, 0, 0);
+
+#if CONFIG_HIGHBITDEPTH
+        }
+#endif
+      }
+      dering_left = 1;
+    }
+    {
+      unsigned char *tmp;
+      tmp = prev_row_dering;
+      prev_row_dering = curr_row_dering;
+      curr_row_dering = tmp;
+    }
+  }
+  aom_free(row_dering);
+  for (pli = 0; pli < nplanes; pli++) {
+    aom_free(linebuf[pli]);
+    aom_free(colbuf[pli]);
+  }
+}
diff --git a/third_party/aom/av1/common/cdef.h b/third_party/aom/av1/common/cdef.h
new file mode 100644
index 0000000000..08c438de6e
--- /dev/null
+++ b/third_party/aom/av1/common/cdef.h
@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AV1_COMMON_DERING_H_
+#define AV1_COMMON_DERING_H_
+
+#define CDEF_STRENGTH_BITS 7
+
+#define DERING_STRENGTHS 32
+#define CLPF_STRENGTHS 4
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "av1/common/od_dering.h"
+#include "av1/common/onyxc_int.h"
+#include "./od_dering.h"
+
+static INLINE int sign(int i) { return i < 0 ? -1 : 1; }
+
+static INLINE int constrain(int diff, int threshold, unsigned int damping) {
+  return threshold
+             ? sign(diff) *
+                   AOMMIN(
+                       abs(diff),
+                       AOMMAX(0, threshold - (abs(diff) >>
+                                              (damping - get_msb(threshold)))))
+             : 0;
+}
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int sb_all_skip(const AV1_COMMON *const cm, int mi_row, int mi_col);
+int sb_compute_dering_list(const AV1_COMMON *const cm, int mi_row, int mi_col,
+                           dering_list *dlist, int filter_skip);
+void av1_cdef_frame(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm, MACROBLOCKD *xd);
+
+void av1_cdef_search(YV12_BUFFER_CONFIG *frame, const YV12_BUFFER_CONFIG *ref,
+                     AV1_COMMON *cm, MACROBLOCKD *xd);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+#endif  // AV1_COMMON_DERING_H_
diff --git a/third_party/aom/av1/common/cdef_simd.h b/third_party/aom/av1/common/cdef_simd.h
new file mode 100644
index 0000000000..2649099a21
--- /dev/null
+++ b/third_party/aom/av1/common/cdef_simd.h
@@ -0,0 +1,27 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AV1_COMMON_CDEF_SIMD_H_
+#define AV1_COMMON_CDEF_SIMD_H_
+
+#include "aom_dsp/aom_simd.h"
+
+// sign(a-b) * min(abs(a-b), max(0, threshold - (abs(a-b) >> adjdamp)))
+SIMD_INLINE v128 constrain16(v128 a, v128 b, unsigned int threshold,
+                             unsigned int adjdamp) {
+  v128 diff = v128_sub_16(a, b);
+  const v128 sign = v128_shr_n_s16(diff, 15);
+  diff = v128_abs_s16(diff);
+  const v128 s =
+      v128_ssub_u16(v128_dup_16(threshold), v128_shr_u16(diff, adjdamp));
+  return v128_xor(v128_add_16(sign, v128_min_s16(diff, s)), sign);
+}
+
+#endif  // AV1_COMMON_CDEF_SIMD_H_
diff --git a/third_party/aom/av1/common/cfl.c b/third_party/aom/av1/common/cfl.c
new file mode 100644
index 0000000000..d66a989ad6
--- /dev/null
+++ b/third_party/aom/av1/common/cfl.c
@@ -0,0 +1,240 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/cfl.h"
+#include "av1/common/common_data.h"
+#include "av1/common/onyxc_int.h"
+
+#include "aom/internal/aom_codec_internal.h"
+
+void cfl_init(CFL_CTX *cfl, AV1_COMMON *cm, int subsampling_x,
+              int subsampling_y) {
+  if (!((subsampling_x == 0 && subsampling_y == 0) ||
+        (subsampling_x == 1 && subsampling_y == 1))) {
+    aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                       "Only 4:4:4 and 4:2:0 are currently supported by CfL");
+  }
+  memset(&cfl->y_pix, 0, sizeof(uint8_t) * MAX_SB_SQUARE);
+  cfl->subsampling_x = subsampling_x;
+  cfl->subsampling_y = subsampling_y;
+}
+
+// CfL computes its own block-level DC_PRED. This is required to compute both
+// alpha_cb and alpha_cr before the prediction are computed.
+void cfl_dc_pred(MACROBLOCKD *xd, BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
+  const struct macroblockd_plane *const pd_u = &xd->plane[AOM_PLANE_U];
+  const struct macroblockd_plane *const pd_v = &xd->plane[AOM_PLANE_V];
+
+  const uint8_t *const dst_u = pd_u->dst.buf;
+  const uint8_t *const dst_v = pd_v->dst.buf;
+
+  const int dst_u_stride = pd_u->dst.stride;
+  const int dst_v_stride = pd_v->dst.stride;
+
+  const int block_width = (plane_bsize != BLOCK_INVALID)
+                              ? block_size_wide[plane_bsize]
+                              : tx_size_wide[tx_size];
+  const int block_height = (plane_bsize != BLOCK_INVALID)
+                               ? block_size_high[plane_bsize]
+                               : tx_size_high[tx_size];
+
+  // Number of pixel on the top and left borders.
+  const int num_pel = block_width + block_height;
+
+  int sum_u = 0;
+  int sum_v = 0;
+
+  // Match behavior of build_intra_predictors (reconintra.c) at superblock
+  // boundaries:
+  //
+  // 127 127 127 .. 127 127 127 127 127 127
+  // 129  A   B  ..  Y   Z
+  // 129  C   D  ..  W   X
+  // 129  E   F  ..  U   V
+  // 129  G   H  ..  S   T   T   T   T   T
+  // ..
+
+  // TODO(ltrudeau) replace this with DC_PRED assembly
+  if (xd->up_available && xd->mb_to_right_edge >= 0) {
+    for (int i = 0; i < block_width; i++) {
+      sum_u += dst_u[-dst_u_stride + i];
+      sum_v += dst_v[-dst_v_stride + i];
+    }
+  } else {
+    sum_u = block_width * 127;
+    sum_v = block_width * 127;
+  }
+
+  if (xd->left_available && xd->mb_to_bottom_edge >= 0) {
+    for (int i = 0; i < block_height; i++) {
+      sum_u += dst_u[i * dst_u_stride - 1];
+      sum_v += dst_v[i * dst_v_stride - 1];
+    }
+  } else {
+    sum_u += block_height * 129;
+    sum_v += block_height * 129;
+  }
+
+  xd->cfl->dc_pred[CFL_PRED_U] = (sum_u + (num_pel >> 1)) / num_pel;
+  xd->cfl->dc_pred[CFL_PRED_V] = (sum_v + (num_pel >> 1)) / num_pel;
+}
+
+// Predict the current transform block using CfL.
+// it is assumed that dst points at the start of the transform block
+void cfl_predict_block(const CFL_CTX *cfl, uint8_t *dst, int dst_stride,
+                       int row, int col, TX_SIZE tx_size, int dc_pred) {
+  const int tx_block_width = tx_size_wide[tx_size];
+  const int tx_block_height = tx_size_high[tx_size];
+
+  // TODO(ltrudeau) implement alpha
+  // Place holder for alpha
+  const double alpha = 0;
+  const double y_avg = cfl_load(cfl, dst, dst_stride, row, col, tx_size);
+
+  for (int j = 0; j < tx_block_height; j++) {
+    for (int i = 0; i < tx_block_width; i++) {
+      dst[i] = (uint8_t)(alpha * y_avg + dc_pred + 0.5);
+    }
+    dst += dst_stride;
+  }
+}
+
+void cfl_store(CFL_CTX *cfl, const uint8_t *input, int input_stride, int row,
+               int col, TX_SIZE tx_size) {
+  const int tx_width = tx_size_wide[tx_size];
+  const int tx_height = tx_size_high[tx_size];
+  const int tx_off_log2 = tx_size_wide_log2[0];
+
+  // Store the input into the CfL pixel buffer
+  uint8_t *y_pix = &cfl->y_pix[(row * MAX_SB_SIZE + col) << tx_off_log2];
+
+  // Check that we remain inside the pixel buffer.
+  assert(MAX_SB_SIZE * (row + tx_height - 1) + col + tx_width - 1 <
+         MAX_SB_SQUARE);
+
+  for (int j = 0; j < tx_height; j++) {
+    for (int i = 0; i < tx_width; i++) {
+      y_pix[i] = input[i];
+    }
+    y_pix += MAX_SB_SIZE;
+    input += input_stride;
+  }
+
+  // Store the surface of the pixel buffer that was written to, this way we
+  // can manage chroma overrun (e.g. when the chroma surfaces goes beyond the
+  // frame boundary)
+  if (col == 0 && row == 0) {
+    cfl->y_width = tx_width;
+    cfl->y_height = tx_height;
+  } else {
+    cfl->y_width = OD_MAXI((col << tx_off_log2) + tx_width, cfl->y_width);
+    cfl->y_height = OD_MAXI((row << tx_off_log2) + tx_height, cfl->y_height);
+  }
+}
+
+// Load from the CfL pixel buffer into output
+double cfl_load(const CFL_CTX *cfl, uint8_t *output, int output_stride, int row,
+                int col, TX_SIZE tx_size) {
+  const int tx_width = tx_size_wide[tx_size];
+  const int tx_height = tx_size_high[tx_size];
+  const int sub_x = cfl->subsampling_x;
+  const int sub_y = cfl->subsampling_y;
+  const int tx_off_log2 = tx_size_wide_log2[0];
+
+  const uint8_t *y_pix;
+
+  int diff_width = 0;
+  int diff_height = 0;
+
+  int pred_row_offset = 0;
+  int output_row_offset = 0;
+  int top_left, bot_left;
+
+  // TODO(ltrudeau) add support for 4:2:2
+  if (sub_y == 0 && sub_x == 0) {
+    y_pix = &cfl->y_pix[(row * MAX_SB_SIZE + col) << tx_off_log2];
+    int uv_width = (col << tx_off_log2) + tx_width;
+    diff_width = uv_width - cfl->y_width;
+    int uv_height = (row << tx_off_log2) + tx_width;
+    diff_height = uv_height - cfl->y_height;
+    for (int j = 0; j < tx_height; j++) {
+      for (int i = 0; i < tx_width; i++) {
+        // In 4:4:4, pixels match 1 to 1
+        output[output_row_offset + i] = y_pix[pred_row_offset + i];
+      }
+      pred_row_offset += MAX_SB_SIZE;
+      output_row_offset += output_stride;
+    }
+  } else if (sub_y == 1 && sub_x == 1) {
+    y_pix = &cfl->y_pix[(row * MAX_SB_SIZE + col) << (tx_off_log2 + sub_y)];
+    int uv_width = ((col << tx_off_log2) + tx_width) << sub_x;
+    diff_width = (uv_width - cfl->y_width) >> sub_x;
+    int uv_height = ((row << tx_off_log2) + tx_width) << sub_y;
+    diff_height = (uv_height - cfl->y_height) >> sub_y;
+    for (int j = 0; j < tx_height; j++) {
+      for (int i = 0; i < tx_width; i++) {
+        top_left = (pred_row_offset + i) << sub_y;
+        bot_left = top_left + MAX_SB_SIZE;
+        // In 4:2:0, average pixels in 2x2 grid
+        output[output_row_offset + i] = OD_SHR_ROUND(
+            y_pix[top_left] + y_pix[top_left + 1]        // Top row
+                + y_pix[bot_left] + y_pix[bot_left + 1]  // Bottom row
+            ,
+            2);
+      }
+      pred_row_offset += MAX_SB_SIZE;
+      output_row_offset += output_stride;
+    }
+  } else {
+    assert(0);  // Unsupported chroma subsampling
+  }
+  // Due to frame boundary issues, it is possible that the total area of
+  // covered by Chroma exceeds that of Luma. When this happens, we write over
+  // the broken data by repeating the last columns and/or rows.
+  //
+  // Note that in order to manage the case where both rows and columns
+  // overrun,
+  // we apply rows first. This way, when the rows overrun the bottom of the
+  // frame, the columns will be copied over them.
+  if (diff_width > 0) {
+    int last_pixel;
+    output_row_offset = tx_width - diff_width;
+
+    for (int j = 0; j < tx_height; j++) {
+      last_pixel = output_row_offset - 1;
+      for (int i = 0; i < diff_width; i++) {
+        output[output_row_offset + i] = output[last_pixel];
+      }
+      output_row_offset += output_stride;
+    }
+  }
+
+  if (diff_height > 0) {
+    output_row_offset = diff_height * output_stride;
+    const int last_row_offset = output_row_offset - output_stride;
+    for (int j = 0; j < diff_height; j++) {
+      for (int i = 0; i < tx_width; i++) {
+        output[output_row_offset + i] = output[last_row_offset + i];
+      }
+      output_row_offset += output_stride;
+    }
+  }
+
+  int avg = 0;
+  output_row_offset = 0;
+  for (int j = 0; j < tx_height; j++) {
+    for (int i = 0; i < tx_width; i++) {
+      avg += output[output_row_offset + i];
+    }
+    output_row_offset += output_stride;
+  }
+  return avg / (double)(tx_width * tx_height);
+}
diff --git a/third_party/aom/av1/common/cfl.h b/third_party/aom/av1/common/cfl.h
new file mode 100644
index 0000000000..371df70bea
--- /dev/null
+++ b/third_party/aom/av1/common/cfl.h
@@ -0,0 +1,52 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_CFL_H_
+#define AV1_COMMON_CFL_H_
+
+#include "av1/common/enums.h"
+
+// Forward declaration of AV1_COMMON, in order to avoid creating a cyclic
+// dependency by importing av1/common/onyxc_int.h
+typedef struct AV1Common AV1_COMMON;
+
+// Forward declaration of MACROBLOCK, in order to avoid creating a cyclic
+// dependency by importing av1/common/blockd.h
+typedef struct macroblockd MACROBLOCKD;
+
+typedef struct {
+  // Pixel buffer containing the luma pixels used as prediction for chroma
+  uint8_t y_pix[MAX_SB_SQUARE];
+
+  // Height and width of the luma prediction block currently in the pixel buffer
+  int y_height, y_width;
+
+  // Chroma subsampling
+  int subsampling_x, subsampling_y;
+
+  // CfL Performs its own block level DC_PRED for each chromatic plane
+  int dc_pred[CFL_PRED_PLANES];
+} CFL_CTX;
+
+void cfl_init(CFL_CTX *cfl, AV1_COMMON *cm, int subsampling_x,
+              int subsampling_y);
+
+void cfl_dc_pred(MACROBLOCKD *xd, BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+
+void cfl_predict_block(const CFL_CTX *cfl, uint8_t *dst, int dst_stride,
+                       int row, int col, TX_SIZE tx_size, int dc_pred);
+
+void cfl_store(CFL_CTX *cfl, const uint8_t *input, int input_stride, int row,
+               int col, TX_SIZE tx_size);
+
+double cfl_load(const CFL_CTX *cfl, uint8_t *output, int output_stride, int row,
+                int col, TX_SIZE tx_size);
+#endif  // AV1_COMMON_CFL_H_
diff --git a/third_party/aom/av1/common/clpf.c b/third_party/aom/av1/common/clpf.c
new file mode 100644
index 0000000000..3637deeea9
--- /dev/null
+++ b/third_party/aom/av1/common/clpf.c
@@ -0,0 +1,116 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./clpf.h"
+#include "./av1_rtcd.h"
+#include "./cdef.h"
+#include "aom/aom_image.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+int av1_clpf_sample(int X, int A, int B, int C, int D, int E, int F, int G,
+                    int H, int s, unsigned int dmp) {
+  int delta = 1 * constrain(A - X, s, dmp) + 3 * constrain(B - X, s, dmp) +
+              1 * constrain(C - X, s, dmp) + 3 * constrain(D - X, s, dmp) +
+              3 * constrain(E - X, s, dmp) + 1 * constrain(F - X, s, dmp) +
+              3 * constrain(G - X, s, dmp) + 1 * constrain(H - X, s, dmp);
+  return (8 + delta - (delta < 0)) >> 4;
+}
+
+int av1_clpf_hsample(int X, int A, int B, int C, int D, int s,
+                     unsigned int dmp) {
+  int delta = 1 * constrain(A - X, s, dmp) + 3 * constrain(B - X, s, dmp) +
+              3 * constrain(C - X, s, dmp) + 1 * constrain(D - X, s, dmp);
+  return (4 + delta - (delta < 0)) >> 3;
+}
+
+void aom_clpf_block_c(uint8_t *dst, const uint16_t *src, int dstride,
+                      int sstride, int sizex, int sizey, unsigned int strength,
+                      unsigned int damping) {
+  int x, y;
+
+  for (y = 0; y < sizey; y++) {
+    for (x = 0; x < sizex; x++) {
+      const int X = src[y * sstride + x];
+      const int A = src[(y - 2) * sstride + x];
+      const int B = src[(y - 1) * sstride + x];
+      const int C = src[y * sstride + x - 2];
+      const int D = src[y * sstride + x - 1];
+      const int E = src[y * sstride + x + 1];
+      const int F = src[y * sstride + x + 2];
+      const int G = src[(y + 1) * sstride + x];
+      const int H = src[(y + 2) * sstride + x];
+      const int delta =
+          av1_clpf_sample(X, A, B, C, D, E, F, G, H, strength, damping);
+      dst[y * dstride + x] = X + delta;
+    }
+  }
+}
+
+// Identical to aom_clpf_block_c() apart from "dst".
+void aom_clpf_block_hbd_c(uint16_t *dst, const uint16_t *src, int dstride,
+                          int sstride, int sizex, int sizey,
+                          unsigned int strength, unsigned int damping) {
+  int x, y;
+
+  for (y = 0; y < sizey; y++) {
+    for (x = 0; x < sizex; x++) {
+      const int X = src[y * sstride + x];
+      const int A = src[(y - 2) * sstride + x];
+      const int B = src[(y - 1) * sstride + x];
+      const int C = src[y * sstride + x - 2];
+      const int D = src[y * sstride + x - 1];
+      const int E = src[y * sstride + x + 1];
+      const int F = src[y * sstride + x + 2];
+      const int G = src[(y + 1) * sstride + x];
+      const int H = src[(y + 2) * sstride + x];
+      const int delta =
+          av1_clpf_sample(X, A, B, C, D, E, F, G, H, strength, damping);
+      dst[y * dstride + x] = X + delta;
+    }
+  }
+}
+
+// Vertically restricted filter
+void aom_clpf_hblock_c(uint8_t *dst, const uint16_t *src, int dstride,
+                       int sstride, int sizex, int sizey, unsigned int strength,
+                       unsigned int damping) {
+  int x, y;
+
+  for (y = 0; y < sizey; y++) {
+    for (x = 0; x < sizex; x++) {
+      const int X = src[y * sstride + x];
+      const int A = src[y * sstride + x - 2];
+      const int B = src[y * sstride + x - 1];
+      const int C = src[y * sstride + x + 1];
+      const int D = src[y * sstride + x + 2];
+      const int delta = av1_clpf_hsample(X, A, B, C, D, strength, damping);
+      dst[y * dstride + x] = X + delta;
+    }
+  }
+}
+
+void aom_clpf_hblock_hbd_c(uint16_t *dst, const uint16_t *src, int dstride,
+                           int sstride, int sizex, int sizey,
+                           unsigned int strength, unsigned int damping) {
+  int x, y;
+
+  for (y = 0; y < sizey; y++) {
+    for (x = 0; x < sizex; x++) {
+      const int X = src[y * sstride + x];
+      const int A = src[y * sstride + x - 2];
+      const int B = src[y * sstride + x - 1];
+      const int C = src[y * sstride + x + 1];
+      const int D = src[y * sstride + x + 2];
+      const int delta = av1_clpf_hsample(X, A, B, C, D, strength, damping);
+      dst[y * dstride + x] = X + delta;
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/clpf.h b/third_party/aom/av1/common/clpf.h
new file mode 100644
index 0000000000..d6348deb0f
--- /dev/null
+++ b/third_party/aom/av1/common/clpf.h
@@ -0,0 +1,18 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AV1_COMMON_CLPF_H_
+#define AV1_COMMON_CLPF_H_
+
+#include "av1/common/reconinter.h"
+
+int av1_clpf_sample(int X, int A, int B, int C, int D, int E, int F, int G,
+                    int H, int b, unsigned int dmp);
+#endif
diff --git a/third_party/aom/av1/common/clpf_neon.c b/third_party/aom/av1/common/clpf_neon.c
new file mode 100644
index 0000000000..f1a004c2c6
--- /dev/null
+++ b/third_party/aom/av1/common/clpf_neon.c
@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_neon
+#include "./clpf_simd.h"
diff --git a/third_party/aom/av1/common/clpf_simd.h b/third_party/aom/av1/common/clpf_simd.h
new file mode 100644
index 0000000000..a615b5ed38
--- /dev/null
+++ b/third_party/aom/av1/common/clpf_simd.h
@@ -0,0 +1,446 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "./cdef_simd.h"
+#include "aom_ports/bitops.h"
+#include "aom_ports/mem.h"
+
+// sign(a - b) * min(abs(a - b), max(0, strength - (abs(a - b) >> adjdamp)))
+SIMD_INLINE v128 constrain(v256 a, v256 b, unsigned int strength,
+                           unsigned int adjdamp) {
+  const v256 diff16 = v256_sub_16(a, b);
+  v128 diff = v128_pack_s16_s8(v256_high_v128(diff16), v256_low_v128(diff16));
+  const v128 sign = v128_cmplt_s8(diff, v128_zero());
+  diff = v128_abs_s8(diff);
+  return v128_xor(
+      v128_add_8(sign,
+                 v128_min_u8(diff, v128_ssub_u8(v128_dup_8(strength),
+                                                v128_shr_u8(diff, adjdamp)))),
+      sign);
+}
+
+// delta = 1/16 * constrain(a, x, s, d) + 3/16 * constrain(b, x, s, d) +
+//         1/16 * constrain(c, x, s, d) + 3/16 * constrain(d, x, s, d) +
+//         3/16 * constrain(e, x, s, d) + 1/16 * constrain(f, x, s, d) +
+//         3/16 * constrain(g, x, s, d) + 1/16 * constrain(h, x, s, d)
+SIMD_INLINE v128 calc_delta(v256 x, v256 a, v256 b, v256 c, v256 d, v256 e,
+                            v256 f, v256 g, v256 h, unsigned int s,
+                            unsigned int dmp) {
+  const v128 bdeg =
+      v128_add_8(v128_add_8(constrain(b, x, s, dmp), constrain(d, x, s, dmp)),
+                 v128_add_8(constrain(e, x, s, dmp), constrain(g, x, s, dmp)));
+  const v128 delta = v128_add_8(
+      v128_add_8(v128_add_8(constrain(a, x, s, dmp), constrain(c, x, s, dmp)),
+                 v128_add_8(constrain(f, x, s, dmp), constrain(h, x, s, dmp))),
+      v128_add_8(v128_add_8(bdeg, bdeg), bdeg));
+  return v128_add_8(
+      v128_pack_s16_u8(v256_high_v128(x), v256_low_v128(x)),
+      v128_shr_s8(
+          v128_add_8(v128_dup_8(8),
+                     v128_add_8(delta, v128_cmplt_s8(delta, v128_zero()))),
+          4));
+}
+
+// delta = 1/8 * constrain(a, x, s, d) + 3/8 * constrain(b, x, s, d) +
+//         3/8 * constrain(c, x, s, d) + 1/8 * constrain(d, x, s, d) +
+SIMD_INLINE v128 calc_hdelta(v256 x, v256 a, v256 b, v256 c, v256 d,
+                             unsigned int s, unsigned int dmp) {
+  const v128 bc = v128_add_8(constrain(b, x, s, dmp), constrain(c, x, s, dmp));
+  const v128 delta =
+      v128_add_8(v128_add_8(constrain(a, x, s, dmp), constrain(d, x, s, dmp)),
+                 v128_add_8(v128_add_8(bc, bc), bc));
+  return v128_add_8(
+      v128_pack_s16_u8(v256_high_v128(x), v256_low_v128(x)),
+      v128_shr_s8(
+          v128_add_8(v128_dup_8(4),
+                     v128_add_8(delta, v128_cmplt_s8(delta, v128_zero()))),
+          3));
+}
+
+// Process blocks of width 8, two lines at a time, 8 bit.
+static void SIMD_FUNC(clpf_block8)(uint8_t *dst, const uint16_t *src,
+                                   int dstride, int sstride, int sizey,
+                                   unsigned int strength,
+                                   unsigned int adjdamp) {
+  int y;
+
+  for (y = 0; y < sizey; y += 2) {
+    const v128 l1 = v128_load_aligned(src);
+    const v128 l2 = v128_load_aligned(src + sstride);
+    const v128 l3 = v128_load_aligned(src - sstride);
+    const v128 l4 = v128_load_aligned(src + 2 * sstride);
+    const v256 a = v256_from_v128(v128_load_aligned(src - 2 * sstride), l3);
+    const v256 b = v256_from_v128(l3, l1);
+    const v256 g = v256_from_v128(l2, l4);
+    const v256 h = v256_from_v128(l4, v128_load_aligned(src + 3 * sstride));
+    const v256 c = v256_from_v128(v128_load_unaligned(src - 2),
+                                  v128_load_unaligned(src - 2 + sstride));
+    const v256 d = v256_from_v128(v128_load_unaligned(src - 1),
+                                  v128_load_unaligned(src - 1 + sstride));
+    const v256 e = v256_from_v128(v128_load_unaligned(src + 1),
+                                  v128_load_unaligned(src + 1 + sstride));
+    const v256 f = v256_from_v128(v128_load_unaligned(src + 2),
+                                  v128_load_unaligned(src + 2 + sstride));
+    const v128 o = calc_delta(v256_from_v128(l1, l2), a, b, c, d, e, f, g, h,
+                              strength, adjdamp);
+
+    v64_store_aligned(dst, v128_high_v64(o));
+    v64_store_aligned(dst + dstride, v128_low_v64(o));
+    src += sstride * 2;
+    dst += dstride * 2;
+  }
+}
+
+// Process blocks of width 4, four lines at a time, 8 bit.
+static void SIMD_FUNC(clpf_block4)(uint8_t *dst, const uint16_t *src,
+                                   int dstride, int sstride, int sizey,
+                                   unsigned int strength,
+                                   unsigned int adjdamp) {
+  int y;
+
+  for (y = 0; y < sizey; y += 4) {
+    const v64 l0 = v64_load_aligned(src - 2 * sstride);
+    const v64 l1 = v64_load_aligned(src - sstride);
+    const v64 l2 = v64_load_aligned(src);
+    const v64 l3 = v64_load_aligned(src + sstride);
+    const v64 l4 = v64_load_aligned(src + 2 * sstride);
+    const v64 l5 = v64_load_aligned(src + 3 * sstride);
+    const v64 l6 = v64_load_aligned(src + 4 * sstride);
+    const v64 l7 = v64_load_aligned(src + 5 * sstride);
+    const v128 o =
+        calc_delta(v256_from_v64(l2, l3, l4, l5), v256_from_v64(l0, l1, l2, l3),
+                   v256_from_v64(l1, l2, l3, l4),
+                   v256_from_v64(v64_load_unaligned(src - 2),
+                                 v64_load_unaligned(src + sstride - 2),
+                                 v64_load_unaligned(src + 2 * sstride - 2),
+                                 v64_load_unaligned(src + 3 * sstride - 2)),
+                   v256_from_v64(v64_load_unaligned(src - 1),
+                                 v64_load_unaligned(src + sstride - 1),
+                                 v64_load_unaligned(src + 2 * sstride - 1),
+                                 v64_load_unaligned(src + 3 * sstride - 1)),
+                   v256_from_v64(v64_load_unaligned(src + 1),
+                                 v64_load_unaligned(src + sstride + 1),
+                                 v64_load_unaligned(src + 2 * sstride + 1),
+                                 v64_load_unaligned(src + 3 * sstride + 1)),
+                   v256_from_v64(v64_load_unaligned(src + 2),
+                                 v64_load_unaligned(src + sstride + 2),
+                                 v64_load_unaligned(src + 2 * sstride + 2),
+                                 v64_load_unaligned(src + 3 * sstride + 2)),
+                   v256_from_v64(l3, l4, l5, l6), v256_from_v64(l4, l5, l6, l7),
+                   strength, adjdamp);
+
+    u32_store_aligned(dst, v128_low_u32(v128_shr_n_byte(o, 12)));
+    u32_store_aligned(dst + dstride, v128_low_u32(v128_shr_n_byte(o, 8)));
+    u32_store_aligned(dst + 2 * dstride, v128_low_u32(v128_shr_n_byte(o, 4)));
+    u32_store_aligned(dst + 3 * dstride, v128_low_u32(o));
+
+    dst += 4 * dstride;
+    src += 4 * sstride;
+  }
+}
+
+static void SIMD_FUNC(clpf_hblock8)(uint8_t *dst, const uint16_t *src,
+                                    int dstride, int sstride, int sizey,
+                                    unsigned int strength,
+                                    unsigned int adjdamp) {
+  int y;
+
+  for (y = 0; y < sizey; y += 2) {
+    const v256 x = v256_from_v128(v128_load_aligned(src),
+                                  v128_load_aligned(src + sstride));
+    const v256 a = v256_from_v128(v128_load_unaligned(src - 2),
+                                  v128_load_unaligned(src - 2 + sstride));
+    const v256 b = v256_from_v128(v128_load_unaligned(src - 1),
+                                  v128_load_unaligned(src - 1 + sstride));
+    const v256 c = v256_from_v128(v128_load_unaligned(src + 1),
+                                  v128_load_unaligned(src + 1 + sstride));
+    const v256 d = v256_from_v128(v128_load_unaligned(src + 2),
+                                  v128_load_unaligned(src + 2 + sstride));
+    const v128 o = calc_hdelta(x, a, b, c, d, strength, adjdamp);
+
+    v64_store_aligned(dst, v128_high_v64(o));
+    v64_store_aligned(dst + dstride, v128_low_v64(o));
+    src += sstride * 2;
+    dst += dstride * 2;
+  }
+}
+
+// Process blocks of width 4, four lines at a time, 8 bit.
+static void SIMD_FUNC(clpf_hblock4)(uint8_t *dst, const uint16_t *src,
+                                    int dstride, int sstride, int sizey,
+                                    unsigned int strength,
+                                    unsigned int adjdamp) {
+  int y;
+
+  for (y = 0; y < sizey; y += 4) {
+    const v256 a = v256_from_v64(v64_load_unaligned(src - 2),
+                                 v64_load_unaligned(src + sstride - 2),
+                                 v64_load_unaligned(src + 2 * sstride - 2),
+                                 v64_load_unaligned(src + 3 * sstride - 2));
+    const v256 b = v256_from_v64(v64_load_unaligned(src - 1),
+                                 v64_load_unaligned(src + sstride - 1),
+                                 v64_load_unaligned(src + 2 * sstride - 1),
+                                 v64_load_unaligned(src + 3 * sstride - 1));
+    const v256 c = v256_from_v64(v64_load_unaligned(src + 1),
+                                 v64_load_unaligned(src + sstride + 1),
+                                 v64_load_unaligned(src + 2 * sstride + 1),
+                                 v64_load_unaligned(src + 3 * sstride + 1));
+    const v256 d = v256_from_v64(v64_load_unaligned(src + 2),
+                                 v64_load_unaligned(src + sstride + 2),
+                                 v64_load_unaligned(src + 2 * sstride + 2),
+                                 v64_load_unaligned(src + 3 * sstride + 2));
+
+    const v128 o = calc_hdelta(
+        v256_from_v64(v64_load_aligned(src), v64_load_aligned(src + sstride),
+                      v64_load_aligned(src + 2 * sstride),
+                      v64_load_aligned(src + 3 * sstride)),
+        a, b, c, d, strength, adjdamp);
+
+    u32_store_aligned(dst, v128_low_u32(v128_shr_n_byte(o, 12)));
+    u32_store_aligned(dst + dstride, v128_low_u32(v128_shr_n_byte(o, 8)));
+    u32_store_aligned(dst + 2 * dstride, v128_low_u32(v128_shr_n_byte(o, 4)));
+    u32_store_aligned(dst + 3 * dstride, v128_low_u32(o));
+
+    dst += 4 * dstride;
+    src += 4 * sstride;
+  }
+}
+
+void SIMD_FUNC(aom_clpf_block)(uint8_t *dst, const uint16_t *src, int dstride,
+                               int sstride, int sizex, int sizey,
+                               unsigned int strength, unsigned int dmp) {
+  if ((sizex != 4 && sizex != 8) || ((sizey & 3) && sizex == 4)) {
+    // Fallback to C for odd sizes:
+    // * block widths not 4 or 8
+    // * block heights not a multiple of 4 if the block width is 4
+    aom_clpf_block_c(dst, src, dstride, sstride, sizex, sizey, strength, dmp);
+  } else {
+    (sizex == 4 ? SIMD_FUNC(clpf_block4) : SIMD_FUNC(clpf_block8))(
+        dst, src, dstride, sstride, sizey, strength, dmp - get_msb(strength));
+  }
+}
+
+void SIMD_FUNC(aom_clpf_hblock)(uint8_t *dst, const uint16_t *src, int dstride,
+                                int sstride, int sizex, int sizey,
+                                unsigned int strength, unsigned int dmp) {
+  if ((sizex != 4 && sizex != 8) || ((sizey & 3) && sizex == 4)) {
+    // Fallback to C for odd sizes:
+    // * block widths not 4 or 8
+    // * block heights not a multiple of 4 if the block width is 4
+    aom_clpf_hblock_c(dst, src, dstride, sstride, sizex, sizey, strength, dmp);
+  } else {
+    (sizex == 4 ? SIMD_FUNC(clpf_hblock4) : SIMD_FUNC(clpf_hblock8))(
+        dst, src, dstride, sstride, sizey, strength, dmp - get_msb(strength));
+  }
+}
+
+// delta = 1/16 * constrain(a, x, s, d) + 3/16 * constrain(b, x, s, d) +
+//         1/16 * constrain(c, x, s, d) + 3/16 * constrain(d, x, s, d) +
+//         3/16 * constrain(e, x, s, d) + 1/16 * constrain(f, x, s, d) +
+//         3/16 * constrain(g, x, s, d) + 1/16 * constrain(h, x, s, d)
+SIMD_INLINE v128 calc_delta_hbd(v128 x, v128 a, v128 b, v128 c, v128 d, v128 e,
+                                v128 f, v128 g, v128 h, unsigned int s,
+                                unsigned int dmp) {
+  const v128 bdeg = v128_add_16(
+      v128_add_16(constrain16(b, x, s, dmp), constrain16(d, x, s, dmp)),
+      v128_add_16(constrain16(e, x, s, dmp), constrain16(g, x, s, dmp)));
+  const v128 delta = v128_add_16(
+      v128_add_16(
+          v128_add_16(constrain16(a, x, s, dmp), constrain16(c, x, s, dmp)),
+          v128_add_16(constrain16(f, x, s, dmp), constrain16(h, x, s, dmp))),
+      v128_add_16(v128_add_16(bdeg, bdeg), bdeg));
+  return v128_add_16(
+      x,
+      v128_shr_s16(
+          v128_add_16(v128_dup_16(8),
+                      v128_add_16(delta, v128_cmplt_s16(delta, v128_zero()))),
+          4));
+}
+
+static void calc_delta_hbd4(v128 o, v128 a, v128 b, v128 c, v128 d, v128 e,
+                            v128 f, v128 g, v128 h, uint16_t *dst,
+                            unsigned int s, unsigned int dmp, int dstride) {
+  o = calc_delta_hbd(o, a, b, c, d, e, f, g, h, s, dmp);
+  v64_store_aligned(dst, v128_high_v64(o));
+  v64_store_aligned(dst + dstride, v128_low_v64(o));
+}
+
+static void calc_delta_hbd8(v128 o, v128 a, v128 b, v128 c, v128 d, v128 e,
+                            v128 f, v128 g, v128 h, uint16_t *dst,
+                            unsigned int s, unsigned int adjdamp) {
+  v128_store_aligned(dst,
+                     calc_delta_hbd(o, a, b, c, d, e, f, g, h, s, adjdamp));
+}
+
+// delta = 1/16 * constrain(a, x, s, dmp) + 3/16 * constrain(b, x, s, dmp) +
+//         3/16 * constrain(c, x, s, dmp) + 1/16 * constrain(d, x, s, dmp)
+SIMD_INLINE v128 calc_hdelta_hbd(v128 x, v128 a, v128 b, v128 c, v128 d,
+                                 unsigned int s, unsigned int dmp) {
+  const v128 bc =
+      v128_add_16(constrain16(b, x, s, dmp), constrain16(c, x, s, dmp));
+  const v128 delta = v128_add_16(
+      v128_add_16(constrain16(a, x, s, dmp), constrain16(d, x, s, dmp)),
+      v128_add_16(v128_add_16(bc, bc), bc));
+  return v128_add_16(
+      x,
+      v128_shr_s16(
+          v128_add_16(v128_dup_16(4),
+                      v128_add_16(delta, v128_cmplt_s16(delta, v128_zero()))),
+          3));
+}
+
+static void calc_hdelta_hbd4(v128 o, v128 a, v128 b, v128 c, v128 d,
+                             uint16_t *dst, unsigned int s,
+                             unsigned int adjdamp, int dstride) {
+  o = calc_hdelta_hbd(o, a, b, c, d, s, adjdamp);
+  v64_store_aligned(dst, v128_high_v64(o));
+  v64_store_aligned(dst + dstride, v128_low_v64(o));
+}
+
+static void calc_hdelta_hbd8(v128 o, v128 a, v128 b, v128 c, v128 d,
+                             uint16_t *dst, unsigned int s,
+                             unsigned int adjdamp) {
+  v128_store_aligned(dst, calc_hdelta_hbd(o, a, b, c, d, s, adjdamp));
+}
+
+// Process blocks of width 4, two lines at time.
+static void SIMD_FUNC(clpf_block_hbd4)(uint16_t *dst, const uint16_t *src,
+                                       int dstride, int sstride, int sizey,
+                                       unsigned int strength,
+                                       unsigned int adjdamp) {
+  int y;
+
+  for (y = 0; y < sizey; y += 2) {
+    const v64 l1 = v64_load_aligned(src);
+    const v64 l2 = v64_load_aligned(src + sstride);
+    const v64 l3 = v64_load_aligned(src - sstride);
+    const v64 l4 = v64_load_aligned(src + 2 * sstride);
+    const v128 a = v128_from_v64(v64_load_aligned(src - 2 * sstride), l3);
+    const v128 b = v128_from_v64(l3, l1);
+    const v128 g = v128_from_v64(l2, l4);
+    const v128 h = v128_from_v64(l4, v64_load_aligned(src + 3 * sstride));
+    const v128 c = v128_from_v64(v64_load_unaligned(src - 2),
+                                 v64_load_unaligned(src - 2 + sstride));
+    const v128 d = v128_from_v64(v64_load_unaligned(src - 1),
+                                 v64_load_unaligned(src - 1 + sstride));
+    const v128 e = v128_from_v64(v64_load_unaligned(src + 1),
+                                 v64_load_unaligned(src + 1 + sstride));
+    const v128 f = v128_from_v64(v64_load_unaligned(src + 2),
+                                 v64_load_unaligned(src + 2 + sstride));
+
+    calc_delta_hbd4(v128_from_v64(l1, l2), a, b, c, d, e, f, g, h, dst,
+                    strength, adjdamp, dstride);
+    src += sstride * 2;
+    dst += dstride * 2;
+  }
+}
+
+// The most simple case.  Start here if you need to understand the functions.
+static void SIMD_FUNC(clpf_block_hbd)(uint16_t *dst, const uint16_t *src,
+                                      int dstride, int sstride, int sizey,
+                                      unsigned int strength,
+                                      unsigned int adjdamp) {
+  int y;
+
+  for (y = 0; y < sizey; y++) {
+    const v128 o = v128_load_aligned(src);
+    const v128 a = v128_load_aligned(src - 2 * sstride);
+    const v128 b = v128_load_aligned(src - 1 * sstride);
+    const v128 g = v128_load_aligned(src + sstride);
+    const v128 h = v128_load_aligned(src + 2 * sstride);
+    const v128 c = v128_load_unaligned(src - 2);
+    const v128 d = v128_load_unaligned(src - 1);
+    const v128 e = v128_load_unaligned(src + 1);
+    const v128 f = v128_load_unaligned(src + 2);
+
+    calc_delta_hbd8(o, a, b, c, d, e, f, g, h, dst, strength, adjdamp);
+    src += sstride;
+    dst += dstride;
+  }
+}
+
+// Process blocks of width 4, horizontal filter, two lines at time.
+static void SIMD_FUNC(clpf_hblock_hbd4)(uint16_t *dst, const uint16_t *src,
+                                        int dstride, int sstride, int sizey,
+                                        unsigned int strength,
+                                        unsigned int adjdamp) {
+  int y;
+
+  for (y = 0; y < sizey; y += 2) {
+    const v128 a = v128_from_v64(v64_load_unaligned(src - 2),
+                                 v64_load_unaligned(src - 2 + sstride));
+    const v128 b = v128_from_v64(v64_load_unaligned(src - 1),
+                                 v64_load_unaligned(src - 1 + sstride));
+    const v128 c = v128_from_v64(v64_load_unaligned(src + 1),
+                                 v64_load_unaligned(src + 1 + sstride));
+    const v128 d = v128_from_v64(v64_load_unaligned(src + 2),
+                                 v64_load_unaligned(src + 2 + sstride));
+
+    calc_hdelta_hbd4(v128_from_v64(v64_load_unaligned(src),
+                                   v64_load_unaligned(src + sstride)),
+                     a, b, c, d, dst, strength, adjdamp, dstride);
+    src += sstride * 2;
+    dst += dstride * 2;
+  }
+}
+
+// Process blocks of width 8, horizontal filter, two lines at time.
+static void SIMD_FUNC(clpf_hblock_hbd)(uint16_t *dst, const uint16_t *src,
+                                       int dstride, int sstride, int sizey,
+                                       unsigned int strength,
+                                       unsigned int adjdamp) {
+  int y;
+
+  for (y = 0; y < sizey; y++) {
+    const v128 o = v128_load_aligned(src);
+    const v128 a = v128_load_unaligned(src - 2);
+    const v128 b = v128_load_unaligned(src - 1);
+    const v128 c = v128_load_unaligned(src + 1);
+    const v128 d = v128_load_unaligned(src + 2);
+
+    calc_hdelta_hbd8(o, a, b, c, d, dst, strength, adjdamp);
+    src += sstride;
+    dst += dstride;
+  }
+}
+
+void SIMD_FUNC(aom_clpf_block_hbd)(uint16_t *dst, const uint16_t *src,
+                                   int dstride, int sstride, int sizex,
+                                   int sizey, unsigned int strength,
+                                   unsigned int dmp) {
+  if ((sizex != 4 && sizex != 8) || ((sizey & 1) && sizex == 4)) {
+    // Fallback to C for odd sizes:
+    // * block width not 4 or 8
+    // * block heights not a multiple of 2 if the block width is 4
+    aom_clpf_block_hbd_c(dst, src, dstride, sstride, sizex, sizey, strength,
+                         dmp);
+  } else {
+    (sizex == 4 ? SIMD_FUNC(clpf_block_hbd4) : SIMD_FUNC(clpf_block_hbd))(
+        dst, src, dstride, sstride, sizey, strength, dmp - get_msb(strength));
+  }
+}
+
+void SIMD_FUNC(aom_clpf_hblock_hbd)(uint16_t *dst, const uint16_t *src,
+                                    int dstride, int sstride, int sizex,
+                                    int sizey, unsigned int strength,
+                                    unsigned int dmp) {
+  if ((sizex != 4 && sizex != 8) || ((sizey & 1) && sizex == 4)) {
+    // Fallback to C for odd sizes:
+    // * block width not 4 or 8
+    // * block heights not a multiple of 2 if the block width is 4
+    aom_clpf_hblock_hbd_c(dst, src, dstride, sstride, sizex, sizey, strength,
+                          dmp);
+  } else {
+    (sizex == 4 ? SIMD_FUNC(clpf_hblock_hbd4) : SIMD_FUNC(clpf_hblock_hbd))(
+        dst, src, dstride, sstride, sizey, strength, dmp - get_msb(strength));
+  }
+}
diff --git a/third_party/aom/av1/common/clpf_sse2.c b/third_party/aom/av1/common/clpf_sse2.c
new file mode 100644
index 0000000000..e29c2ab7ea
--- /dev/null
+++ b/third_party/aom/av1/common/clpf_sse2.c
@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_sse2
+#include "./clpf_simd.h"
diff --git a/third_party/aom/av1/common/clpf_sse4.c b/third_party/aom/av1/common/clpf_sse4.c
new file mode 100644
index 0000000000..537139f17a
--- /dev/null
+++ b/third_party/aom/av1/common/clpf_sse4.c
@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_sse4_1
+#include "./clpf_simd.h"
diff --git a/third_party/aom/av1/common/clpf_ssse3.c b/third_party/aom/av1/common/clpf_ssse3.c
new file mode 100644
index 0000000000..d7ed8dec5d
--- /dev/null
+++ b/third_party/aom/av1/common/clpf_ssse3.c
@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_ssse3
+#include "./clpf_simd.h"
diff --git a/third_party/aom/av1/common/common.h b/third_party/aom/av1/common/common.h
new file mode 100644
index 0000000000..551055a76d
--- /dev/null
+++ b/third_party/aom/av1/common/common.h
@@ -0,0 +1,64 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_COMMON_H_
+#define AV1_COMMON_COMMON_H_
+
+/* Interface header for common constant data structures and lookup tables */
+
+#include <assert.h>
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/bitops.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define PI 3.141592653589793238462643383279502884
+
+// Only need this for fixed-size arrays, for structs just assign.
+#define av1_copy(dest, src)              \
+  {                                      \
+    assert(sizeof(dest) == sizeof(src)); \
+    memcpy(dest, src, sizeof(src));      \
+  }
+
+// Use this for variably-sized arrays.
+#define av1_copy_array(dest, src, n)           \
+  {                                            \
+    assert(sizeof(*(dest)) == sizeof(*(src))); \
+    memcpy(dest, src, n * sizeof(*(src)));     \
+  }
+
+#define av1_zero(dest) memset(&(dest), 0, sizeof(dest))
+#define av1_zero_array(dest, n) memset(dest, 0, n * sizeof(*(dest)))
+
+static INLINE int get_unsigned_bits(unsigned int num_values) {
+  return num_values > 0 ? get_msb(num_values) + 1 : 0;
+}
+
+#define CHECK_MEM_ERROR(cm, lval, expr) \
+  AOM_CHECK_MEM_ERROR(&cm->error, lval, expr)
+// TODO(yaowu: validate the usage of these codes or develop new ones.)
+#define AV1_SYNC_CODE_0 0x49
+#define AV1_SYNC_CODE_1 0x83
+#define AV1_SYNC_CODE_2 0x43
+
+#define AOM_FRAME_MARKER 0x2
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_COMMON_H_
diff --git a/third_party/aom/av1/common/common_data.h b/third_party/aom/av1/common/common_data.h
new file mode 100644
index 0000000000..415d5cf73e
--- /dev/null
+++ b/third_party/aom/av1/common/common_data.h
@@ -0,0 +1,1405 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_COMMON_DATA_H_
+#define AV1_COMMON_COMMON_DATA_H_
+
+#include "av1/common/enums.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_EXT_PARTITION
+#define IF_EXT_PARTITION(...) __VA_ARGS__
+#else
+#define IF_EXT_PARTITION(...)
+#endif
+
+// Log 2 conversion lookup tables for block width and height
+static const uint8_t b_width_log2_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0, 0, 0,
+#endif
+  0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, IF_EXT_PARTITION(4, 5, 5)
+};
+static const uint8_t b_height_log2_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0, 0, 0,
+#endif
+  0, 1, 0, 1, 2, 1, 2, 3, 2, 3, 4, 3, 4, IF_EXT_PARTITION(5, 4, 5)
+};
+// Log 2 conversion lookup tables for modeinfo width and height
+static const uint8_t mi_width_log2_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0, 0, 0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, IF_EXT_PARTITION(4, 5, 5)
+#else
+  0, 0, 0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, IF_EXT_PARTITION(3, 4, 4)
+#endif
+};
+static const uint8_t mi_height_log2_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0, 0, 0, 0, 1, 0, 1, 2, 1, 2, 3, 2, 3, 4, 3, 4, IF_EXT_PARTITION(5, 4, 5)
+#else
+  0, 0, 0, 0, 1, 0, 1, 2, 1, 2, 3, 2, 3, IF_EXT_PARTITION(4, 3, 4)
+#endif
+};
+
+static const uint8_t mi_size_wide[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  1, 1, 1, 1, 1, 2, 2, 2, 4, 4, 4, 8, 8, 8, 16, 16, IF_EXT_PARTITION(16, 32, 32)
+#else
+  1, 1, 1, 1, 1, 2, 2, 2, 4, 4, 4, 8, 8, IF_EXT_PARTITION(8, 16, 16)
+#endif
+};
+static const uint8_t mi_size_high[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  1, 1, 1, 1, 2, 1, 2, 4, 2, 4, 8, 4, 8, 16, 8, 16, IF_EXT_PARTITION(32, 16, 32)
+#else
+  1, 1, 1, 1, 2, 1, 2, 4, 2, 4, 8, 4, 8, IF_EXT_PARTITION(16, 8, 16)
+#endif
+};
+
+// Width/height lookup tables in units of various block sizes
+static const uint8_t block_size_wide[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  2, 2, 4,
+#endif
+  4, 4, 8, 8, 8, 16, 16, 16, 32, 32, 32, 64, 64, IF_EXT_PARTITION(64, 128, 128)
+};
+
+static const uint8_t block_size_high[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  2, 4, 2,
+#endif
+  4, 8, 4, 8, 16, 8, 16, 32, 16, 32, 64, 32, 64, IF_EXT_PARTITION(128, 64, 128)
+};
+
+static const uint8_t num_4x4_blocks_wide_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  1, 1, 1,
+#endif
+  1, 1, 2, 2, 2, 4, 4, 4, 8, 8, 8, 16, 16, IF_EXT_PARTITION(16, 32, 32)
+};
+static const uint8_t num_4x4_blocks_high_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  1, 1, 1,
+#endif
+  1, 2, 1, 2, 4, 2, 4, 8, 4, 8, 16, 8, 16, IF_EXT_PARTITION(32, 16, 32)
+};
+static const uint8_t num_8x8_blocks_wide_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  1, 1, 1,
+#endif
+  1, 1, 1, 1, 1, 2, 2, 2, 4, 4, 4, 8, 8, IF_EXT_PARTITION(8, 16, 16)
+};
+static const uint8_t num_8x8_blocks_high_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  1, 1, 1,
+#endif
+  1, 1, 1, 1, 2, 1, 2, 4, 2, 4, 8, 4, 8, IF_EXT_PARTITION(16, 8, 16)
+};
+static const uint8_t num_16x16_blocks_wide_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  1, 1, 1,
+#endif
+  1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4, IF_EXT_PARTITION(4, 8, 8)
+};
+static const uint8_t num_16x16_blocks_high_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  1, 1, 1,
+#endif
+  1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4, IF_EXT_PARTITION(8, 4, 8)
+};
+
+// AOMMIN(3, AOMMIN(b_width_log2(bsize), b_height_log2(bsize)))
+static const uint8_t size_group_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  0, 0, 0,
+#endif
+  0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 3, IF_EXT_PARTITION(3, 3, 3)
+};
+
+static const uint8_t num_pels_log2_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  2, 3, 3,
+#endif
+  4, 5, 5, 6, 7, 7, 8, 9, 9, 10, 11, 11, 12, IF_EXT_PARTITION(13, 13, 14)
+};
+
+/* clang-format off */
+static const PARTITION_TYPE
+  partition_lookup[MAX_SB_SIZE_LOG2 - 1][BLOCK_SIZES] = {
+  {     // 4X4 ->
+#if CONFIG_CB4X4
+    // 2X2,            2X4,               4X2,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#endif
+    //                                    4X4
+                                          PARTITION_NONE,
+    // 4X8,            8X4,               8X8
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    // 8X16,           16X8,              16X16
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    // 16X32,          32X16,             32X32
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    // 32X64,          64X32,             64X64
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#if CONFIG_EXT_PARTITION
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#endif  // CONFIG_EXT_PARTITION
+  }, {  // 8X8 ->
+#if CONFIG_CB4X4
+    // 2X2,            2X4,               4X2,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#endif
+    //                                    4X4
+                                          PARTITION_SPLIT,
+    // 4X8,            8X4,               8X8
+    PARTITION_VERT,    PARTITION_HORZ,    PARTITION_NONE,
+    // 8X16,           16X8,              16X16
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    // 16X32,          32X16,             32X32
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    // 32X64,          64X32,             64X64
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#if CONFIG_EXT_PARTITION
+    // 64x128,         128x64,            128x128
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#endif  // CONFIG_EXT_PARTITION
+  }, {  // 16X16 ->
+#if CONFIG_CB4X4
+    // 2X2,            2X4,               4X2,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#endif
+    //                                    4X4
+                                          PARTITION_SPLIT,
+    // 4X8,            8X4,               8X8
+    PARTITION_SPLIT,   PARTITION_SPLIT,   PARTITION_SPLIT,
+    // 8X16,           16X8,              16X16
+    PARTITION_VERT,    PARTITION_HORZ,    PARTITION_NONE,
+    // 16X32,          32X16,             32X32
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+    // 32X64,          64X32,             64X64
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#if CONFIG_EXT_PARTITION
+    // 64x128,         128x64,            128x128
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#endif  // CONFIG_EXT_PARTITION
+  }, {  // 32X32 ->
+#if CONFIG_CB4X4
+    // 2X2,            2X4,               4X2,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#endif
+    //                                    4X4
+                                          PARTITION_SPLIT,
+    // 4X8,            8X4,               8X8
+    PARTITION_SPLIT,   PARTITION_SPLIT,   PARTITION_SPLIT,
+    // 8X16,           16X8,              16X16
+    PARTITION_SPLIT,   PARTITION_SPLIT,   PARTITION_SPLIT,
+    // 16X32,          32X16,             32X32
+    PARTITION_VERT,    PARTITION_HORZ,    PARTITION_NONE,
+    // 32X64,          64X32,             64X64
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#if CONFIG_EXT_PARTITION
+    // 64x128,         128x64,            128x128
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#endif  // CONFIG_EXT_PARTITION
+  }, {  // 64X64 ->
+#if CONFIG_CB4X4
+    // 2X2,            2X4,               4X2,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#endif
+    //                                    4X4
+                                          PARTITION_SPLIT,
+    // 4X8,            8X4,               8X8
+    PARTITION_SPLIT,   PARTITION_SPLIT,   PARTITION_SPLIT,
+    // 8X16,           16X8,              16X16
+    PARTITION_SPLIT,   PARTITION_SPLIT,   PARTITION_SPLIT,
+    // 16X32,          32X16,             32X32
+    PARTITION_SPLIT,   PARTITION_SPLIT,   PARTITION_SPLIT,
+    // 32X64,          64X32,             64X64
+    PARTITION_VERT,    PARTITION_HORZ,    PARTITION_NONE,
+#if CONFIG_EXT_PARTITION
+    // 64x128,         128x64,            128x128
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+  }, {  // 128x128 ->
+#if CONFIG_CB4X4
+    // 2X2,            2X4,               4X2,
+    PARTITION_INVALID, PARTITION_INVALID, PARTITION_INVALID,
+#endif
+    //                                    4X4
+                                          PARTITION_SPLIT,
+    // 4X8,            8X4,               8X8
+    PARTITION_SPLIT,   PARTITION_SPLIT,   PARTITION_SPLIT,
+    // 8X16,           16X8,              16X16
+    PARTITION_SPLIT,   PARTITION_SPLIT,   PARTITION_SPLIT,
+    // 16X32,          32X16,             32X32
+    PARTITION_SPLIT,   PARTITION_SPLIT,   PARTITION_SPLIT,
+    // 32X64,          64X32,             64X64
+    PARTITION_SPLIT,   PARTITION_SPLIT,   PARTITION_SPLIT,
+    // 64x128,         128x64,            128x128
+    PARTITION_VERT,    PARTITION_HORZ,    PARTITION_NONE,
+#endif  // CONFIG_EXT_PARTITION
+  }
+};
+
+#if CONFIG_EXT_PARTITION_TYPES
+static const BLOCK_SIZE subsize_lookup[EXT_PARTITION_TYPES][BLOCK_SIZES] =
+#else
+static const BLOCK_SIZE subsize_lookup[PARTITION_TYPES][BLOCK_SIZES] =
+#endif  // CONFIG_EXT_PARTITION_TYPES
+{
+  {     // PARTITION_NONE
+#if CONFIG_CB4X4
+    // 2X2,        2X4,           4X2,
+    BLOCK_2X2,     BLOCK_2X4,     BLOCK_4X2,
+#endif
+    //                            4X4
+                                  BLOCK_4X4,
+    // 4X8,        8X4,           8X8
+    BLOCK_4X8,     BLOCK_8X4,     BLOCK_8X8,
+    // 8X16,       16X8,          16X16
+    BLOCK_8X16,    BLOCK_16X8,    BLOCK_16X16,
+    // 16X32,      32X16,         32X32
+    BLOCK_16X32,   BLOCK_32X16,   BLOCK_32X32,
+    // 32X64,      64X32,         64X64
+    BLOCK_32X64,   BLOCK_64X32,   BLOCK_64X64,
+#if CONFIG_EXT_PARTITION
+    // 64x128,     128x64,        128x128
+    BLOCK_64X128,  BLOCK_128X64,  BLOCK_128X128,
+#endif  // CONFIG_EXT_PARTITION
+  }, {  // PARTITION_HORZ
+#if CONFIG_CB4X4
+    // 2X2,        2X4,           4X2,
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+    //                            4X4
+                                  BLOCK_4X2,
+#else
+    //                            4X4
+                                  BLOCK_INVALID,
+#endif
+    // 4X8,        8X4,           8X8
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_8X4,
+    // 8X16,       16X8,          16X16
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_16X8,
+    // 16X32,      32X16,         32X32
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_32X16,
+    // 32X64,      64X32,         64X64
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_64X32,
+#if CONFIG_EXT_PARTITION
+    // 64x128,     128x64,        128x128
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_128X64,
+#endif  // CONFIG_EXT_PARTITION
+  }, {  // PARTITION_VERT
+#if CONFIG_CB4X4
+    // 2X2,        2X4,           4X2,
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+    //                            4X4
+                                  BLOCK_2X4,
+#else
+    //                            4X4
+                                  BLOCK_INVALID,
+#endif
+    // 4X8,        8X4,           8X8
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_4X8,
+    // 8X16,       16X8,          16X16
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_8X16,
+    // 16X32,      32X16,         32X32
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_16X32,
+    // 32X64,      64X32,         64X64
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_32X64,
+#if CONFIG_EXT_PARTITION
+    // 64x128,     128x64,        128x128
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_64X128,
+#endif  // CONFIG_EXT_PARTITION
+  }, {  // PARTITION_SPLIT
+#if CONFIG_CB4X4
+    // 2X2,        2X4,           4X2,
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+#endif
+    //                            4X4
+                                  BLOCK_INVALID,
+    // 4X8,        8X4,           8X8
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_4X4,
+    // 8X16,       16X8,          16X16
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_8X8,
+    // 16X32,      32X16,         32X32
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_16X16,
+    // 32X64,      64X32,         64X64
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_32X32,
+#if CONFIG_EXT_PARTITION
+    // 64x128,     128x64,        128x128
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_64X64,
+#endif  // CONFIG_EXT_PARTITION
+#if CONFIG_EXT_PARTITION_TYPES
+  }, {  // PARTITION_HORZ_A
+#if CONFIG_CB4X4
+    // 2X2,        2X4,           4X2,
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+#endif
+    //                            4X4
+                                  BLOCK_INVALID,
+    // 4X8,        8X4,           8X8
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_8X4,
+    // 8X16,       16X8,          16X16
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_16X8,
+    // 16X32,      32X16,         32X32
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_32X16,
+    // 32X64,      64X32,         64X64
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_64X32,
+#if CONFIG_EXT_PARTITION
+    // 64x128,     128x64,        128x128
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_128X64,
+#endif  // CONFIG_EXT_PARTITION
+  }, {  // PARTITION_HORZ_B
+#if CONFIG_CB4X4
+    // 2X2,        2X4,           4X2,
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+#endif
+    //                            4X4
+                                  BLOCK_INVALID,
+    // 4X8,        8X4,           8X8
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_8X4,
+    // 8X16,       16X8,          16X16
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_16X8,
+    // 16X32,      32X16,         32X32
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_32X16,
+    // 32X64,      64X32,         64X64
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_64X32,
+#if CONFIG_EXT_PARTITION
+    // 64x128,     128x64,        128x128
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_128X64,
+#endif  // CONFIG_EXT_PARTITION
+  }, {  // PARTITION_VERT_A
+#if CONFIG_CB4X4
+    // 2X2,        2X4,           4X2,
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+#endif
+    //                            4X4
+                                  BLOCK_INVALID,
+    // 4X8,        8X4,           8X8
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_4X8,
+    // 8X16,       16X8,          16X16
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_8X16,
+    // 16X32,      32X16,         32X32
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_16X32,
+    // 32X64,      64X32,         64X64
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_32X64,
+#if CONFIG_EXT_PARTITION
+    // 64x128,     128x64,        128x128
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_64X128,
+#endif  // CONFIG_EXT_PARTITION
+  }, {  // PARTITION_VERT_B
+#if CONFIG_CB4X4
+    // 2X2,        2X4,           4X2,
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_INVALID,
+#endif
+    //                            4X4
+                                  BLOCK_INVALID,
+    // 4X8,        8X4,           8X8
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_4X8,
+    // 8X16,       16X8,          16X16
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_8X16,
+    // 16X32,      32X16,         32X32
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_16X32,
+    // 32X64,      64X32,         64X64
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_32X64,
+#if CONFIG_EXT_PARTITION
+    // 64x128,     128x64,        128x128
+    BLOCK_INVALID, BLOCK_INVALID, BLOCK_64X128,
+#endif  // CONFIG_EXT_PARTITION
+#endif  // CONFIG_EXT_PARTITION_TYPES
+  }
+};
+
+static const TX_SIZE max_txsize_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  // 2X2,    2X4,      4X2,
+  TX_2X2,    TX_2X2,   TX_2X2,
+#endif
+  //                   4X4
+                       TX_4X4,
+  // 4X8,    8X4,      8X8
+  TX_4X4,    TX_4X4,   TX_8X8,
+  // 8X16,   16X8,     16X16
+  TX_8X8,    TX_8X8,   TX_16X16,
+  // 16X32,  32X16,    32X32
+  TX_16X16,  TX_16X16, TX_32X32,
+  // 32X64,  64X32,
+  TX_32X32,  TX_32X32,
+#if CONFIG_TX64X64
+  // 64X64
+  TX_64X64,
+#if CONFIG_EXT_PARTITION
+  // 64x128, 128x64,   128x128
+  TX_64X64,  TX_64X64, TX_64X64,
+#endif  // CONFIG_EXT_PARTITION
+#else
+  // 64X64
+  TX_32X32,
+#if CONFIG_EXT_PARTITION
+  // 64x128, 128x64,   128x128
+  TX_32X32,  TX_32X32, TX_32X32,
+#endif  // CONFIG_EXT_PARTITION
+#endif  // CONFIG_TX64X64
+};
+
+#if CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+static const TX_SIZE max_txsize_rect_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  // 2X2,    2X4,      4X2,
+  TX_2X2,    TX_2X2,   TX_2X2,
+#endif  // CONFIG_CB4X4
+  //                   4X4
+                       TX_4X4,
+  // 4X8,    8X4,      8X8
+  TX_4X8,    TX_8X4,   TX_8X8,
+  // 8X16,   16X8,     16X16
+  TX_8X16,   TX_16X8,  TX_16X16,
+  // 16X32,  32X16,    32X32
+  TX_16X32,  TX_32X16, TX_32X32,
+  // 32X64,  64X32,
+  TX_32X32,  TX_32X32,
+#if CONFIG_TX64X64
+  // 64X64
+  TX_64X64,
+#if CONFIG_EXT_PARTITION
+  // 64x128, 128x64,   128x128
+  TX_64X64,  TX_64X64, TX_64X64,
+#endif  // CONFIG_EXT_PARTITION
+#else
+  // 64X64
+  TX_32X32,
+#if CONFIG_EXT_PARTITION
+  // 64x128, 128x64,   128x128
+  TX_32X32,  TX_32X32, TX_32X32,
+#endif  // CONFIG_EXT_PARTITION
+#endif  // CONFIG_TX64X64
+};
+#else
+#define max_txsize_rect_lookup max_txsize_lookup
+#endif  // CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+
+#if CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+// Same as "max_txsize_lookup[bsize] - TX_8X8", except for rectangular
+// block which may use a rectangular transform, in which  case it is
+// "(max_txsize_lookup[bsize] + 1) - TX_8X8", invalid for bsize < 8X8
+static const int32_t intra_tx_size_cat_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  // 2X2,             2X4,                4X2,
+  INT32_MIN,          INT32_MIN,          INT32_MIN,
+  //                                      4X4,
+                                          INT32_MIN,
+  // 4X8,             8X4,                8X8,
+  TX_8X8 - TX_8X8,    TX_8X8 - TX_8X8,    TX_8X8 - TX_8X8,
+#else
+  //                                      4X4
+                                          INT32_MIN,
+  // 4X8,             8X4,                8X8
+  INT32_MIN,          INT32_MIN,          TX_8X8 - TX_8X8,
+#endif  // CONFIG_CB4X4
+  // 8X16,            16X8,               16X16
+  TX_16X16 - TX_8X8,  TX_16X16 - TX_8X8,  TX_16X16 - TX_8X8,
+  // 16X32,           32X16,              32X32
+  TX_32X32 - TX_8X8,  TX_32X32 - TX_8X8,  TX_32X32 - TX_8X8,
+  // 32X64,           64X32,
+  TX_32X32 - TX_8X8,  TX_32X32 - TX_8X8,
+#if CONFIG_TX64X64
+  // 64X64
+  TX_64X64 - TX_8X8,
+#if CONFIG_EXT_PARTITION
+  // 64x128,          128x64,             128x128
+  TX_64X64 - TX_8X8,  TX_64X64 - TX_8X8,  TX_64X64 - TX_8X8,
+#endif  // CONFIG_EXT_PARTITION
+#else
+  // 64X64
+  TX_32X32 - TX_8X8,
+#if CONFIG_EXT_PARTITION
+  // 64x128,          128x64,             128x128
+  TX_32X32 - TX_8X8,  TX_32X32 - TX_8X8,  TX_32X32 - TX_8X8,
+#endif  // CONFIG_EXT_PARTITION
+#endif  // CONFIG_TX64X64
+};
+#else
+// Same as "max_txsize_lookup[bsize] - TX_8X8", invalid for bsize < 8X8
+static const int32_t intra_tx_size_cat_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  // 2X2,             2X4,                4X2,
+  INT32_MIN,          INT32_MIN,          INT32_MIN,
+#endif
+  //                                      4X4
+                                          INT32_MIN,
+  // 4X8,             8X4,                8X8
+  INT32_MIN,          INT32_MIN,          TX_8X8 - TX_8X8,
+  // 8X16,            16X8,               16X16
+  TX_8X8 - TX_8X8,    TX_8X8 - TX_8X8,    TX_16X16 - TX_8X8,
+  // 16X32,           32X16,              32X32
+  TX_16X16 - TX_8X8,  TX_16X16 - TX_8X8,  TX_32X32 - TX_8X8,
+  // 32X64,           64X32,
+  TX_32X32 - TX_8X8,  TX_32X32 - TX_8X8,
+#if CONFIG_TX64X64
+  // 64X64
+  TX_64X64 - TX_8X8,
+#if CONFIG_EXT_PARTITION
+  // 64x128,          128x64,             128x128
+  TX_64X64 - TX_8X8,  TX_64X64 - TX_8X8,  TX_64X64 - TX_8X8,
+#endif  // CONFIG_EXT_PARTITION
+#else
+  // 64X64
+  TX_32X32 - TX_8X8,
+#if CONFIG_EXT_PARTITION
+  // 64x128,          128x64,             128x128
+  TX_32X32 - TX_8X8,  TX_32X32 - TX_8X8,  TX_32X32 - TX_8X8,
+#endif  // CONFIG_EXT_PARTITION
+#endif  // CONFIG_TX64X64
+};
+#endif  // CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+
+#define inter_tx_size_cat_lookup intra_tx_size_cat_lookup
+
+/* clang-format on */
+
+static const TX_SIZE sub_tx_size_map[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  TX_2X2,  // TX_2X2
+#endif
+  TX_4X4,    // TX_4X4
+  TX_4X4,    // TX_8X8
+  TX_8X8,    // TX_16X16
+  TX_16X16,  // TX_32X32
+#if CONFIG_TX64X64
+  TX_32X32,  // TX_64X64
+#endif       // CONFIG_TX64X64
+  TX_4X4,    // TX_4X8
+  TX_4X4,    // TX_8X4
+  TX_8X8,    // TX_8X16
+  TX_8X8,    // TX_16X8
+  TX_16X16,  // TX_16X32
+  TX_16X16,  // TX_32X16
+  TX_4X4,    // TX_4X16
+  TX_4X4,    // TX_16X4
+  TX_8X8,    // TX_8X32
+  TX_8X8,    // TX_32X8
+};
+
+static const TX_SIZE txsize_horz_map[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  TX_2X2,  // TX_2X2
+#endif
+  TX_4X4,    // TX_4X4
+  TX_8X8,    // TX_8X8
+  TX_16X16,  // TX_16X16
+  TX_32X32,  // TX_32X32
+#if CONFIG_TX64X64
+  TX_64X64,  // TX_64X64
+#endif       // CONFIG_TX64X64
+  TX_4X4,    // TX_4X8
+  TX_8X8,    // TX_8X4
+  TX_8X8,    // TX_8X16
+  TX_16X16,  // TX_16X8
+  TX_16X16,  // TX_16X32
+  TX_32X32,  // TX_32X16
+  TX_4X4,    // TX_4X16
+  TX_16X16,  // TX_16X4
+  TX_8X8,    // TX_8X32
+  TX_32X32,  // TX_32X8
+};
+
+static const TX_SIZE txsize_vert_map[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  TX_2X2,  // TX_2X2
+#endif
+  TX_4X4,    // TX_4X4
+  TX_8X8,    // TX_8X8
+  TX_16X16,  // TX_16X16
+  TX_32X32,  // TX_32X32
+#if CONFIG_TX64X64
+  TX_64X64,  // TX_64X64
+#endif       // CONFIG_TX64X64
+  TX_8X8,    // TX_4X8
+  TX_4X4,    // TX_8X4
+  TX_16X16,  // TX_8X16
+  TX_8X8,    // TX_16X8
+  TX_32X32,  // TX_16X32
+  TX_16X16,  // TX_32X16
+  TX_16X16,  // TX_4X16
+  TX_4X4,    // TX_16X4
+  TX_32X32,  // TX_8X32
+  TX_8X8,    // TX_32X8
+};
+
+#if CONFIG_CB4X4
+#define TX_SIZE_W_MIN 2
+#else
+#define TX_SIZE_W_MIN 4
+#endif
+
+// Transform block width in pixels
+static const int tx_size_wide[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  2,
+#endif
+  4,  8, 16, 32,
+#if CONFIG_TX64X64
+  64,
+#endif  // CONFIG_TX64X64
+  4,  8, 8,  16, 16, 32, 4, 16, 8, 32
+};
+
+#if CONFIG_CB4X4
+#define TX_SIZE_H_MIN 2
+#else
+#define TX_SIZE_H_MIN 4
+#endif
+
+// Transform block height in pixels
+static const int tx_size_high[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  2,
+#endif
+  4,  8, 16, 32,
+#if CONFIG_TX64X64
+  64,
+#endif  // CONFIG_TX64X64
+  8,  4, 16, 8,  32, 16, 16, 4, 32, 8
+};
+
+// Transform block width in unit
+static const int tx_size_wide_unit[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  1,  2, 4, 8, 16,
+#if CONFIG_TX64X64
+  32,
+#endif  // CONFIG_TX64X64
+  2,  4, 4, 8, 8,  16, 2, 8, 4, 16
+#else  // CONFIG_CB4X4
+  1,  2, 4, 8,
+#if CONFIG_TX64X64
+  16,
+#endif  // CONFIG_TX64X64
+  1,  2, 2, 4, 4, 8, 1, 4, 2, 8
+#endif  // CONFIG_CB4X4
+};
+
+// Transform block height in unit
+static const int tx_size_high_unit[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  1,  2, 4, 8, 16,
+#if CONFIG_TX64X64
+  32,
+#endif  // CONFIG_TX64X64
+  4,  2, 8, 4, 16, 8, 8, 2, 16, 4
+#else  // CONFIG_CB4X4
+  1,  2, 4, 8,
+#if CONFIG_TX64X64
+  16,
+#endif  // CONFIG_TX64X64
+  2,  1, 4, 2, 8, 4, 4, 1, 8, 2
+#endif  // CONFIG_CB4X4
+};
+
+// Transform block width in log2
+static const int tx_size_wide_log2[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  1,
+#endif
+  2, 3, 4, 5,
+#if CONFIG_TX64X64
+  6,
+#endif  // CONFIG_TX64X64
+  2, 3, 3, 4, 4, 5, 2, 4, 3, 5
+};
+
+// Transform block height in log2
+static const int tx_size_high_log2[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  1,
+#endif
+  2, 3, 4, 5,
+#if CONFIG_TX64X64
+  6,
+#endif  // CONFIG_TX64X64
+  3, 2, 4, 3, 5, 4, 4, 2, 5, 3
+};
+
+static const int tx_size_2d[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  4,
+#endif
+  16,   64, 256, 1024,
+#if CONFIG_TX64X64
+  4096,
+#endif  // CONFIG_TX64X64
+  32,   32, 128, 128,  512, 512, 64, 64, 256, 256
+};
+
+static const BLOCK_SIZE txsize_to_bsize[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  BLOCK_2X2,  // TX_2X2
+#endif
+  BLOCK_4X4,    // TX_4X4
+  BLOCK_8X8,    // TX_8X8
+  BLOCK_16X16,  // TX_16X16
+  BLOCK_32X32,  // TX_32X32
+#if CONFIG_TX64X64
+  BLOCK_64X64,    // TX_64X64
+#endif            // CONFIG_TX64X64
+  BLOCK_4X8,      // TX_4X8
+  BLOCK_8X4,      // TX_8X4
+  BLOCK_8X16,     // TX_8X16
+  BLOCK_16X8,     // TX_16X8
+  BLOCK_16X32,    // TX_16X32
+  BLOCK_32X16,    // TX_32X16
+  BLOCK_INVALID,  // TX_4X16
+  BLOCK_INVALID,  // TX_16X4
+  BLOCK_INVALID,  // TX_8X32
+  BLOCK_INVALID,  // TX_32X8
+};
+
+static const TX_SIZE txsize_sqr_map[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  TX_2X2,  // TX_2X2
+#endif
+  TX_4X4,    // TX_4X4
+  TX_8X8,    // TX_8X8
+  TX_16X16,  // TX_16X16
+  TX_32X32,  // TX_32X32
+#if CONFIG_TX64X64
+  TX_64X64,  // TX_64X64
+#endif       // CONFIG_TX64X64
+  TX_4X4,    // TX_4X8
+  TX_4X4,    // TX_8X4
+  TX_8X8,    // TX_8X16
+  TX_8X8,    // TX_16X8
+  TX_16X16,  // TX_16X32
+  TX_16X16,  // TX_32X16
+  TX_4X4,    // TX_4X16
+  TX_4X4,    // TX_16X4
+  TX_8X8,    // TX_8X32
+  TX_8X8,    // TX_32X8
+};
+
+static const TX_SIZE txsize_sqr_up_map[TX_SIZES_ALL] = {
+#if CONFIG_CB4X4
+  TX_2X2,  // TX_2X2
+#endif
+  TX_4X4,    // TX_4X4
+  TX_8X8,    // TX_8X8
+  TX_16X16,  // TX_16X16
+  TX_32X32,  // TX_32X32
+#if CONFIG_TX64X64
+  TX_64X64,  // TX_64X64
+#endif       // CONFIG_TX64X64
+  TX_8X8,    // TX_4X8
+  TX_8X8,    // TX_8X4
+  TX_16X16,  // TX_8X16
+  TX_16X16,  // TX_16X8
+  TX_32X32,  // TX_16X32
+  TX_32X32,  // TX_32X16
+  TX_16X16,  // TX_4X16
+  TX_16X16,  // TX_16X4
+  TX_32X32,  // TX_8X32
+  TX_32X32,  // TX_32X8
+};
+
+/* clang-format off */
+static const TX_SIZE tx_mode_to_biggest_tx_size[TX_MODES] = {
+  TX_4X4,    // ONLY_4X4
+  TX_8X8,    // ALLOW_8X8
+  TX_16X16,  // ALLOW_16X16
+  TX_32X32,  // ALLOW_32X32
+#if CONFIG_TX64X64
+  TX_64X64,  // ALLOW_64X64
+  TX_64X64,  // TX_MODE_SELECT
+#else
+  TX_32X32,  // TX_MODE_SELECT
+#endif  // CONFIG_TX64X64
+};
+/* clang-format on */
+
+static const BLOCK_SIZE ss_size_lookup[BLOCK_SIZES][2][2] = {
+//  ss_x == 0    ss_x == 0        ss_x == 1      ss_x == 1
+//  ss_y == 0    ss_y == 1        ss_y == 0      ss_y == 1
+#if CONFIG_CB4X4
+  { { BLOCK_2X2, BLOCK_INVALID }, { BLOCK_INVALID, BLOCK_INVALID } },
+  { { BLOCK_2X4, BLOCK_INVALID }, { BLOCK_INVALID, BLOCK_INVALID } },
+  { { BLOCK_4X2, BLOCK_INVALID }, { BLOCK_INVALID, BLOCK_INVALID } },
+  { { BLOCK_4X4, BLOCK_4X2 }, { BLOCK_2X4, BLOCK_2X2 } },
+  { { BLOCK_4X8, BLOCK_4X4 }, { BLOCK_INVALID, BLOCK_2X4 } },
+  { { BLOCK_8X4, BLOCK_INVALID }, { BLOCK_4X4, BLOCK_4X2 } },
+#else
+  { { BLOCK_4X4, BLOCK_INVALID }, { BLOCK_INVALID, BLOCK_INVALID } },
+  { { BLOCK_4X8, BLOCK_4X4 }, { BLOCK_INVALID, BLOCK_INVALID } },
+  { { BLOCK_8X4, BLOCK_INVALID }, { BLOCK_4X4, BLOCK_INVALID } },
+#endif
+  { { BLOCK_8X8, BLOCK_8X4 }, { BLOCK_4X8, BLOCK_4X4 } },
+  { { BLOCK_8X16, BLOCK_8X8 }, { BLOCK_INVALID, BLOCK_4X8 } },
+  { { BLOCK_16X8, BLOCK_INVALID }, { BLOCK_8X8, BLOCK_8X4 } },
+  { { BLOCK_16X16, BLOCK_16X8 }, { BLOCK_8X16, BLOCK_8X8 } },
+  { { BLOCK_16X32, BLOCK_16X16 }, { BLOCK_INVALID, BLOCK_8X16 } },
+  { { BLOCK_32X16, BLOCK_INVALID }, { BLOCK_16X16, BLOCK_16X8 } },
+  { { BLOCK_32X32, BLOCK_32X16 }, { BLOCK_16X32, BLOCK_16X16 } },
+  { { BLOCK_32X64, BLOCK_32X32 }, { BLOCK_INVALID, BLOCK_16X32 } },
+  { { BLOCK_64X32, BLOCK_INVALID }, { BLOCK_32X32, BLOCK_32X16 } },
+  { { BLOCK_64X64, BLOCK_64X32 }, { BLOCK_32X64, BLOCK_32X32 } },
+#if CONFIG_EXT_PARTITION
+  { { BLOCK_64X128, BLOCK_64X64 }, { BLOCK_INVALID, BLOCK_32X64 } },
+  { { BLOCK_128X64, BLOCK_INVALID }, { BLOCK_64X64, BLOCK_64X32 } },
+  { { BLOCK_128X128, BLOCK_128X64 }, { BLOCK_64X128, BLOCK_64X64 } },
+#endif  // CONFIG_EXT_PARTITION
+};
+
+static const TX_SIZE uv_txsize_lookup[BLOCK_SIZES][TX_SIZES_ALL][2][2] = {
+//  ss_x == 0    ss_x == 0        ss_x == 1      ss_x == 1
+//  ss_y == 0    ss_y == 1        ss_y == 0      ss_y == 1
+#if CONFIG_CB4X4
+  {
+      // BLOCK_2X2
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+#if CONFIG_TX64X64
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+#endif  // CONFIG_TX64X64
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+      // BLOCK_2X4
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+#if CONFIG_TX64X64
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+#endif  // CONFIG_TX64X64
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+      // BLOCK_2X4
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+#if CONFIG_TX64X64
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+#endif  // CONFIG_TX64X64
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+#endif
+  {
+// BLOCK_4X4
+#if CONFIG_CB4X4
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_4X4, TX_2X2 }, { TX_2X2, TX_2X2 } },
+#else
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif  // CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_TX64X64
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_4X8
+#if CONFIG_CB4X4
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_4X4, TX_2X2 }, { TX_2X2, TX_2X2 } },
+#else
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_TX64X64
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif  // CONFIG_TX64X64
+#if CONFIG_CB4X4
+      { { TX_4X8, TX_4X4 }, { TX_2X2, TX_2X2 } },  // used
+#else
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },  // used
+#endif
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_8X4
+#if CONFIG_CB4X4
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+      { { TX_4X4, TX_2X2 }, { TX_2X2, TX_2X2 } },
+#else
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_TX64X64
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif  // CONFIG_TX64X64
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_CB4X4
+      { { TX_8X4, TX_2X2 }, { TX_4X4, TX_2X2 } },  // used
+#else
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },  // used
+#endif
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_8X8
+#if CONFIG_CB4X4
+      { { TX_2X2, TX_2X2 }, { TX_2X2, TX_2X2 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_TX64X64
+      { { TX_8X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_8X4, TX_8X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_8X8, TX_8X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_8X8, TX_8X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_8X8, TX_8X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_8X16
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_TX64X64
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X16, TX_8X8 }, { TX_4X8, TX_4X8 } },  // used
+      { { TX_8X16, TX_8X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X16, TX_8X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X16, TX_8X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_16X8
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_4X4 }, { TX_8X8, TX_4X4 } },
+      { { TX_8X8, TX_4X4 }, { TX_8X8, TX_4X4 } },
+      { { TX_8X8, TX_4X4 }, { TX_8X8, TX_4X4 } },
+#if CONFIG_TX64X64
+      { { TX_8X8, TX_4X4 }, { TX_8X8, TX_4X4 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_16X8, TX_8X4 }, { TX_8X8, TX_8X4 } },
+      { { TX_16X8, TX_8X4 }, { TX_8X8, TX_8X4 } },  // used
+      { { TX_16X8, TX_8X4 }, { TX_8X8, TX_8X4 } },
+      { { TX_16X8, TX_8X4 }, { TX_8X8, TX_8X4 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_16X16
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_8X8 }, { TX_8X8, TX_8X8 } },
+#if CONFIG_TX64X64
+      { { TX_16X16, TX_8X8 }, { TX_8X8, TX_8X8 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X8 }, { TX_8X16, TX_8X8 } },
+      { { TX_16X8, TX_16X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X8 }, { TX_8X16, TX_8X8 } },
+      { { TX_16X16, TX_16X8 }, { TX_8X16, TX_8X8 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_16X32
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_8X8, TX_8X8 } },
+#if CONFIG_TX64X64
+      { { TX_16X16, TX_16X16 }, { TX_8X8, TX_8X8 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X32, TX_16X16 }, { TX_8X16, TX_8X16 } },  // used
+      { { TX_16X32, TX_16X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_32X16
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_8X8 }, { TX_16X16, TX_8X8 } },
+      { { TX_16X16, TX_8X8 }, { TX_16X16, TX_8X8 } },
+#if CONFIG_TX64X64
+      { { TX_16X16, TX_8X8 }, { TX_16X16, TX_8X8 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X8 }, { TX_8X16, TX_8X8 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_32X16, TX_16X8 }, { TX_16X16, TX_16X8 } },
+      { { TX_32X16, TX_16X8 }, { TX_16X16, TX_16X8 } },  // used
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_32X32
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_16X16 }, { TX_16X16, TX_16X16 } },
+#if CONFIG_TX64X64
+      { { TX_32X32, TX_16X16 }, { TX_16X16, TX_16X16 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X16 }, { TX_16X32, TX_16X16 } },
+      { { TX_32X16, TX_32X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_32X64
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_32X32 }, { TX_16X16, TX_16X16 } },
+#if CONFIG_TX64X64
+      { { TX_32X32, TX_32X32 }, { TX_16X16, TX_16X16 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X32 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X16, TX_32X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_64X32
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_16X16 }, { TX_32X32, TX_16X16 } },
+#if CONFIG_TX64X64
+      { { TX_32X32, TX_16X16 }, { TX_32X32, TX_16X16 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X16 }, { TX_16X32, TX_16X16 } },
+      { { TX_32X16, TX_16X16 }, { TX_32X16, TX_16X16 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_64X64
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_32X32 }, { TX_32X32, TX_32X32 } },
+#if CONFIG_TX64X64
+      { { TX_64X64, TX_32X32 }, { TX_32X32, TX_32X32 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X32 }, { TX_16X32, TX_16X32 } },
+      { { TX_32X16, TX_32X16 }, { TX_32X16, TX_16X16 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+#if CONFIG_EXT_PARTITION
+  {
+// BLOCK_64X128
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_32X32 }, { TX_32X32, TX_32X32 } },
+#if CONFIG_TX64X64
+      { { TX_64X64, TX_64X64 }, { TX_32X32, TX_32X32 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X32 }, { TX_16X32, TX_16X32 } },
+      { { TX_32X16, TX_32X16 }, { TX_32X16, TX_32X16 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_128X64
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_32X32 }, { TX_32X32, TX_32X32 } },
+#if CONFIG_TX64X64
+      { { TX_64X64, TX_32X32 }, { TX_64X64, TX_32X32 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X32 }, { TX_16X32, TX_16X32 } },
+      { { TX_32X16, TX_32X16 }, { TX_32X16, TX_32X16 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+  {
+// BLOCK_128X128
+#if CONFIG_CB4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_32X32 }, { TX_32X32, TX_32X32 } },
+#if CONFIG_TX64X64
+      { { TX_64X64, TX_64X64 }, { TX_64X64, TX_64X64 } },
+#endif  // CONFIG_TX64X64
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X32 }, { TX_16X32, TX_16X32 } },
+      { { TX_32X16, TX_32X16 }, { TX_32X16, TX_32X16 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+  },
+#endif  // CONFIG_EXT_PARTITION
+};
+
+// Generates 4 bit field in which each bit set to 1 represents
+// a blocksize partition  1111 means we split 64x64, 32x32, 16x16
+// and 8x8.  1000 means we just split the 64x64 to 32x32
+/* clang-format off */
+static const struct {
+  PARTITION_CONTEXT above;
+  PARTITION_CONTEXT left;
+} partition_context_lookup[BLOCK_SIZES] = {
+#if CONFIG_EXT_PARTITION
+#if CONFIG_CB4X4
+  { 31, 31 },  // 2X2   - {0b11111, 0b11111}
+  { 31, 31 },  // 2X4   - {0b11111, 0b11111}
+  { 31, 31 },  // 4X2   - {0b11111, 0b11111}
+#endif
+  { 31, 31 },  // 4X4   - {0b11111, 0b11111}
+  { 31, 30 },  // 4X8   - {0b11111, 0b11110}
+  { 30, 31 },  // 8X4   - {0b11110, 0b11111}
+  { 30, 30 },  // 8X8   - {0b11110, 0b11110}
+  { 30, 28 },  // 8X16  - {0b11110, 0b11100}
+  { 28, 30 },  // 16X8  - {0b11100, 0b11110}
+  { 28, 28 },  // 16X16 - {0b11100, 0b11100}
+  { 28, 24 },  // 16X32 - {0b11100, 0b11000}
+  { 24, 28 },  // 32X16 - {0b11000, 0b11100}
+  { 24, 24 },  // 32X32 - {0b11000, 0b11000}
+  { 24, 16 },  // 32X64 - {0b11000, 0b10000}
+  { 16, 24 },  // 64X32 - {0b10000, 0b11000}
+  { 16, 16 },  // 64X64 - {0b10000, 0b10000}
+  { 16, 0 },   // 64X128- {0b10000, 0b00000}
+  { 0, 16 },   // 128X64- {0b00000, 0b10000}
+  { 0, 0 },    // 128X128-{0b00000, 0b00000}
+#else
+#if CONFIG_CB4X4
+  { 15, 15 },  // 2X2   - {0b1111, 0b1111}
+  { 15, 15 },  // 2X4   - {0b1111, 0b1111}
+  { 15, 15 },  // 4X2   - {0b1111, 0b1111}
+#endif
+  { 15, 15 },  // 4X4   - {0b1111, 0b1111}
+  { 15, 14 },  // 4X8   - {0b1111, 0b1110}
+  { 14, 15 },  // 8X4   - {0b1110, 0b1111}
+  { 14, 14 },  // 8X8   - {0b1110, 0b1110}
+  { 14, 12 },  // 8X16  - {0b1110, 0b1100}
+  { 12, 14 },  // 16X8  - {0b1100, 0b1110}
+  { 12, 12 },  // 16X16 - {0b1100, 0b1100}
+  { 12, 8 },   // 16X32 - {0b1100, 0b1000}
+  { 8, 12 },   // 32X16 - {0b1000, 0b1100}
+  { 8, 8 },    // 32X32 - {0b1000, 0b1000}
+  { 8, 0 },    // 32X64 - {0b1000, 0b0000}
+  { 0, 8 },    // 64X32 - {0b0000, 0b1000}
+  { 0, 0 },    // 64X64 - {0b0000, 0b0000}
+#endif  // CONFIG_EXT_PARTITION
+};
+/* clang-format on */
+
+#if CONFIG_SUPERTX
+static const TX_SIZE uvsupertx_size_lookup[TX_SIZES][2][2] = {
+//  ss_x == 0 ss_x == 0   ss_x == 1 ss_x == 1
+//  ss_y == 0 ss_y == 1   ss_y == 0 ss_y == 1
+#if CONFIG_CB4X4
+  { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif
+  { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+  { { TX_8X8, TX_4X4 }, { TX_4X4, TX_4X4 } },
+  { { TX_16X16, TX_8X8 }, { TX_8X8, TX_8X8 } },
+  { { TX_32X32, TX_16X16 }, { TX_16X16, TX_16X16 } },
+#if CONFIG_TX64X64
+  { { TX_64X64, TX_32X32 }, { TX_32X32, TX_32X32 } },
+#endif  // CONFIG_TX64X64
+};
+
+#if CONFIG_EXT_PARTITION_TYPES
+static const int partition_supertx_context_lookup[EXT_PARTITION_TYPES] = {
+  -1, 0, 0, 1, 0, 0, 0, 0
+};
+
+#else
+static const int partition_supertx_context_lookup[PARTITION_TYPES] = { -1, 0, 0,
+                                                                       1 };
+#endif  // CONFIG_EXT_PARTITION_TYPES
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_ADAPT_SCAN
+#define EOB_THRESHOLD_NUM 2
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_COMMON_DATA_H_
diff --git a/third_party/aom/av1/common/convolve.c b/third_party/aom/av1/common/convolve.c
new file mode 100644
index 0000000000..eab6fe7a35
--- /dev/null
+++ b/third_party/aom/av1/common/convolve.c
@@ -0,0 +1,775 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <string.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+#include "av1/common/convolve.h"
+#include "av1/common/filter.h"
+#include "av1/common/onyxc_int.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+#define MAX_BLOCK_WIDTH (MAX_SB_SIZE)
+#define MAX_BLOCK_HEIGHT (MAX_SB_SIZE)
+#define MAX_STEP (32)
+
+void av1_convolve_horiz_c(const uint8_t *src, int src_stride, uint8_t *dst,
+                          int dst_stride, int w, int h,
+                          const InterpFilterParams filter_params,
+                          const int subpel_x_q4, int x_step_q4,
+                          ConvolveParams *conv_params) {
+  int x, y;
+  int filter_size = filter_params.taps;
+  assert(conv_params->round == CONVOLVE_OPT_ROUND);
+  src -= filter_size / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = subpel_x_q4;
+    for (x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+          filter_params, x_q4 & SUBPEL_MASK);
+      int k, sum = 0;
+      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];
+
+      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      if (conv_params->ref)
+        dst[x] = ROUND_POWER_OF_TWO(dst[x] + sum, 1);
+      else
+        dst[x] = sum;
+
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void av1_convolve_vert_c(const uint8_t *src, int src_stride, uint8_t *dst,
+                         int dst_stride, int w, int h,
+                         const InterpFilterParams filter_params,
+                         const int subpel_y_q4, int y_step_q4,
+                         ConvolveParams *conv_params) {
+  int x, y;
+  int filter_size = filter_params.taps;
+  assert(conv_params->round == CONVOLVE_OPT_ROUND);
+  src -= src_stride * (filter_size / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = subpel_y_q4;
+    for (y = 0; y < h; ++y) {
+      const uint8_t *const src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+          filter_params, y_q4 & SUBPEL_MASK);
+      int k, sum = 0;
+      for (k = 0; k < filter_size; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+
+      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      if (conv_params->ref)
+        dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] + sum, 1);
+      else
+        dst[y * dst_stride] = sum;
+
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void convolve_copy(const uint8_t *src, int src_stride, uint8_t *dst,
+                          int dst_stride, int w, int h,
+                          ConvolveParams *conv_params) {
+  assert(conv_params->round == CONVOLVE_OPT_ROUND);
+  if (conv_params->ref == 0) {
+    int r;
+    for (r = 0; r < h; ++r) {
+      memcpy(dst, src, w);
+      src += src_stride;
+      dst += dst_stride;
+    }
+  } else {
+    int r, c;
+    for (r = 0; r < h; ++r) {
+      for (c = 0; c < w; ++c) {
+        dst[c] = clip_pixel(ROUND_POWER_OF_TWO(dst[c] + src[c], 1));
+      }
+      src += src_stride;
+      dst += dst_stride;
+    }
+  }
+}
+
+void av1_convolve_horiz_facade(const uint8_t *src, int src_stride, uint8_t *dst,
+                               int dst_stride, int w, int h,
+                               const InterpFilterParams filter_params,
+                               const int subpel_x_q4, int x_step_q4,
+                               ConvolveParams *conv_params) {
+  assert(conv_params->round == CONVOLVE_OPT_ROUND);
+  if (filter_params.taps == SUBPEL_TAPS) {
+    const int16_t *filter_x =
+        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
+    if (conv_params->ref == 0)
+      aom_convolve8_horiz(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
+                          NULL, -1, w, h);
+    else
+      aom_convolve8_avg_horiz(src, src_stride, dst, dst_stride, filter_x,
+                              x_step_q4, NULL, -1, w, h);
+  } else {
+    av1_convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params,
+                       subpel_x_q4, x_step_q4, conv_params);
+  }
+}
+
+void av1_convolve_horiz_facade_c(const uint8_t *src, int src_stride,
+                                 uint8_t *dst, int dst_stride, int w, int h,
+                                 const InterpFilterParams filter_params,
+                                 const int subpel_x_q4, int x_step_q4,
+                                 ConvolveParams *conv_params) {
+  assert(conv_params->round == CONVOLVE_OPT_ROUND);
+  if (filter_params.taps == SUBPEL_TAPS) {
+    const int16_t *filter_x =
+        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
+    if (conv_params->ref == 0)
+      aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                            x_step_q4, NULL, -1, w, h);
+    else
+      aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+                                x_step_q4, NULL, -1, w, h);
+  } else {
+    av1_convolve_horiz_c(src, src_stride, dst, dst_stride, w, h, filter_params,
+                         subpel_x_q4, x_step_q4, conv_params);
+  }
+}
+
+void av1_convolve_vert_facade(const uint8_t *src, int src_stride, uint8_t *dst,
+                              int dst_stride, int w, int h,
+                              const InterpFilterParams filter_params,
+                              const int subpel_y_q4, int y_step_q4,
+                              ConvolveParams *conv_params) {
+  assert(conv_params->round == CONVOLVE_OPT_ROUND);
+  if (filter_params.taps == SUBPEL_TAPS) {
+    const int16_t *filter_y =
+        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
+    if (conv_params->ref == 0) {
+      aom_convolve8_vert(src, src_stride, dst, dst_stride, NULL, -1, filter_y,
+                         y_step_q4, w, h);
+    } else {
+      aom_convolve8_avg_vert(src, src_stride, dst, dst_stride, NULL, -1,
+                             filter_y, y_step_q4, w, h);
+    }
+  } else {
+    av1_convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params,
+                      subpel_y_q4, y_step_q4, conv_params);
+  }
+}
+
+void av1_convolve_vert_facade_c(const uint8_t *src, int src_stride,
+                                uint8_t *dst, int dst_stride, int w, int h,
+                                const InterpFilterParams filter_params,
+                                const int subpel_y_q4, int y_step_q4,
+                                ConvolveParams *conv_params) {
+  assert(conv_params->round == CONVOLVE_OPT_ROUND);
+  if (filter_params.taps == SUBPEL_TAPS) {
+    const int16_t *filter_y =
+        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
+    if (conv_params->ref == 0) {
+      aom_convolve8_vert_c(src, src_stride, dst, dst_stride, NULL, -1, filter_y,
+                           y_step_q4, w, h);
+    } else {
+      aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, NULL, -1,
+                               filter_y, y_step_q4, w, h);
+    }
+  } else {
+    av1_convolve_vert_c(src, src_stride, dst, dst_stride, w, h, filter_params,
+                        subpel_y_q4, y_step_q4, conv_params);
+  }
+}
+
+#if CONFIG_CONVOLVE_ROUND
+void av1_convolve_rounding(const int32_t *src, int src_stride, uint8_t *dst,
+                           int dst_stride, int w, int h, int bits) {
+  int r, c;
+  for (r = 0; r < h; ++r) {
+    for (c = 0; c < w; ++c) {
+      dst[r * dst_stride + c] =
+          clip_pixel(ROUND_POWER_OF_TWO_SIGNED(src[r * src_stride + c], bits));
+    }
+  }
+}
+
+void av1_convolve_2d(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst,
+                     int dst_stride, int w, int h,
+                     InterpFilterParams *filter_params_x,
+                     InterpFilterParams *filter_params_y, const int subpel_x_q4,
+                     const int subpel_y_q4, ConvolveParams *conv_params) {
+  int x, y, k;
+  CONV_BUF_TYPE im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = w;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  (void)conv_params;
+  // horizontal filter
+  const uint8_t *src_horiz = src - fo_vert * src_stride;
+  const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+  for (y = 0; y < im_h; ++y) {
+    for (x = 0; x < w; ++x) {
+      CONV_BUF_TYPE sum = 0;
+      for (k = 0; k < filter_params_x->taps; ++k) {
+        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
+      }
+#if CONFIG_COMPOUND_ROUND
+      im_block[y * im_stride + x] =
+          clip_pixel(ROUND_POWER_OF_TWO_SIGNED(sum, conv_params->round_0));
+#else
+      im_block[y * im_stride + x] =
+          ROUND_POWER_OF_TWO_SIGNED(sum, conv_params->round_0);
+#endif
+    }
+  }
+
+  // vertical filter
+  CONV_BUF_TYPE *src_vert = im_block + fo_vert * im_stride;
+  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      CONV_BUF_TYPE sum = 0;
+      for (k = 0; k < filter_params_y->taps; ++k) {
+        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
+      }
+      dst[y * dst_stride + x] +=
+          ROUND_POWER_OF_TWO_SIGNED(sum, conv_params->round_1);
+    }
+  }
+}
+
+static INLINE void transpose_uint8(uint8_t *dst, int dst_stride,
+                                   const uint8_t *src, int src_stride, int w,
+                                   int h) {
+  int r, c;
+  for (r = 0; r < h; ++r)
+    for (c = 0; c < w; ++c)
+      dst[c * (dst_stride) + r] = src[r * (src_stride) + c];
+}
+
+static INLINE void transpose_int32(int32_t *dst, int dst_stride,
+                                   const int32_t *src, int src_stride, int w,
+                                   int h) {
+  int r, c;
+  for (r = 0; r < h; ++r)
+    for (c = 0; c < w; ++c)
+      dst[c * (dst_stride) + r] = src[r * (src_stride) + c];
+}
+
+void av1_convolve_2d_facade(const uint8_t *src, int src_stride, uint8_t *dst,
+                            int dst_stride, int w, int h,
+                            const InterpFilter *interp_filter,
+                            const int subpel_x_q4, int x_step_q4,
+                            const int subpel_y_q4, int y_step_q4,
+                            ConvolveParams *conv_params) {
+  (void)x_step_q4;
+  (void)y_step_q4;
+  (void)dst;
+  (void)dst_stride;
+#if CONFIG_DUAL_FILTER
+  InterpFilterParams filter_params_x =
+      av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]);
+  InterpFilterParams filter_params_y =
+      av1_get_interp_filter_params(interp_filter[0 + 2 * conv_params->ref]);
+
+  if (filter_params_x.interp_filter == MULTITAP_SHARP &&
+      filter_params_y.interp_filter == MULTITAP_SHARP) {
+    // Avoid two directions both using 12-tap filter.
+    // This will reduce hardware implementation cost.
+    filter_params_y = av1_get_interp_filter_params(EIGHTTAP_SHARP);
+  }
+#else
+  InterpFilterParams filter_params_x =
+      av1_get_interp_filter_params(*interp_filter);
+  InterpFilterParams filter_params_y =
+      av1_get_interp_filter_params(*interp_filter);
+#endif
+
+  if (filter_params_y.taps < filter_params_x.taps) {
+    uint8_t tr_src[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) *
+                   (MAX_SB_SIZE + MAX_FILTER_TAP - 1)];
+    int tr_src_stride = MAX_SB_SIZE + MAX_FILTER_TAP - 1;
+    CONV_BUF_TYPE tr_dst[MAX_SB_SIZE * MAX_SB_SIZE];
+    int tr_dst_stride = MAX_SB_SIZE;
+    int fo_vert = filter_params_y.taps / 2 - 1;
+    int fo_horiz = filter_params_x.taps / 2 - 1;
+
+    transpose_uint8(tr_src, tr_src_stride,
+                    src - fo_vert * src_stride - fo_horiz, src_stride,
+                    w + filter_params_x.taps - 1, h + filter_params_y.taps - 1);
+    transpose_int32(tr_dst, tr_dst_stride, conv_params->dst,
+                    conv_params->dst_stride, w, h);
+
+    // horizontal and vertical parameters are swapped because of the transpose
+    av1_convolve_2d(tr_src + fo_horiz * tr_src_stride + fo_vert, tr_src_stride,
+                    tr_dst, tr_dst_stride, h, w, &filter_params_y,
+                    &filter_params_x, subpel_y_q4, subpel_x_q4, conv_params);
+    transpose_int32(conv_params->dst, conv_params->dst_stride, tr_dst,
+                    tr_dst_stride, h, w);
+  } else {
+    av1_convolve_2d(src, src_stride, conv_params->dst, conv_params->dst_stride,
+                    w, h, &filter_params_x, &filter_params_y, subpel_x_q4,
+                    subpel_y_q4, conv_params);
+  }
+}
+
+#endif  // CONFIG_CONVOLVE_ROUND
+
+typedef void (*ConvolveFunc)(const uint8_t *src, int src_stride, uint8_t *dst,
+                             int dst_stride, int w, int h,
+                             const InterpFilterParams filter_params,
+                             const int subpel_q4, int step_q4,
+                             ConvolveParams *conv_params);
+
+static void convolve_helper(const uint8_t *src, int src_stride, uint8_t *dst,
+                            int dst_stride, int w, int h,
+#if CONFIG_DUAL_FILTER
+                            const InterpFilter *interp_filter,
+#else
+                            const InterpFilter interp_filter,
+#endif
+                            const int subpel_x_q4, int x_step_q4,
+                            const int subpel_y_q4, int y_step_q4,
+                            ConvolveParams *conv_params,
+                            ConvolveFunc convolve_horiz,
+                            ConvolveFunc convolve_vert) {
+  int ignore_horiz = x_step_q4 == 16 && subpel_x_q4 == 0;
+  int ignore_vert = y_step_q4 == 16 && subpel_y_q4 == 0;
+#if CONFIG_DUAL_FILTER
+  InterpFilterParams filter_params_x =
+      av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]);
+  InterpFilterParams filter_params_y =
+      av1_get_interp_filter_params(interp_filter[0 + 2 * conv_params->ref]);
+  InterpFilterParams filter_params;
+#else
+  InterpFilterParams filter_params =
+      av1_get_interp_filter_params(interp_filter);
+#endif
+  assert(conv_params->round == CONVOLVE_OPT_ROUND);
+
+  assert(w <= MAX_BLOCK_WIDTH);
+  assert(h <= MAX_BLOCK_HEIGHT);
+  assert(y_step_q4 <= MAX_STEP);
+  assert(x_step_q4 <= MAX_STEP);
+
+  if (ignore_horiz && ignore_vert) {
+    convolve_copy(src, src_stride, dst, dst_stride, w, h, conv_params);
+  } else if (ignore_vert) {
+#if CONFIG_DUAL_FILTER
+    filter_params = filter_params_x;
+#endif
+    assert(filter_params.taps <= MAX_FILTER_TAP);
+    convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params,
+                   subpel_x_q4, x_step_q4, conv_params);
+  } else if (ignore_horiz) {
+#if CONFIG_DUAL_FILTER
+    filter_params = filter_params_y;
+#endif
+    assert(filter_params.taps <= MAX_FILTER_TAP);
+    convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params,
+                  subpel_y_q4, y_step_q4, conv_params);
+  } else {
+    // temp's size is set to a 256 aligned value to facilitate SIMD
+    // implementation. The value is greater than (maximum possible intermediate
+    // height or width) * MAX_SB_SIZE
+    DECLARE_ALIGNED(16, uint8_t,
+                    temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]);
+    int max_intermediate_size = ((MAX_SB_SIZE * 2 + 16) + 16);
+    int filter_size;
+#if CONFIG_DUAL_FILTER
+    if (interp_filter[0 + 2 * conv_params->ref] == MULTITAP_SHARP &&
+        interp_filter[1 + 2 * conv_params->ref] == MULTITAP_SHARP) {
+      // Avoid two directions both using 12-tap filter.
+      // This will reduce hardware implementation cost.
+      filter_params_y = av1_get_interp_filter_params(EIGHTTAP_SHARP);
+    }
+
+    // we do filter with fewer taps first to reduce hardware implementation
+    // complexity
+    if (filter_params_y.taps < filter_params_x.taps) {
+      int intermediate_width;
+      int temp_stride = max_intermediate_size;
+      ConvolveParams temp_conv_params;
+      temp_conv_params.ref = 0;
+      temp_conv_params.round = CONVOLVE_OPT_ROUND;
+      filter_params = filter_params_y;
+      filter_size = filter_params_x.taps;
+      intermediate_width =
+          (((w - 1) * x_step_q4 + subpel_x_q4) >> SUBPEL_BITS) + filter_size;
+      assert(intermediate_width <= max_intermediate_size);
+
+      assert(filter_params.taps <= MAX_FILTER_TAP);
+
+      convolve_vert(src - (filter_size / 2 - 1), src_stride, temp, temp_stride,
+                    intermediate_width, h, filter_params, subpel_y_q4,
+                    y_step_q4, &temp_conv_params);
+
+      filter_params = filter_params_x;
+      assert(filter_params.taps <= MAX_FILTER_TAP);
+      convolve_horiz(temp + (filter_size / 2 - 1), temp_stride, dst, dst_stride,
+                     w, h, filter_params, subpel_x_q4, x_step_q4, conv_params);
+    } else
+#endif  // CONFIG_DUAL_FILTER
+    {
+      int intermediate_height;
+      int temp_stride = MAX_SB_SIZE;
+      ConvolveParams temp_conv_params;
+      temp_conv_params.ref = 0;
+      temp_conv_params.round = CONVOLVE_OPT_ROUND;
+#if CONFIG_DUAL_FILTER
+      filter_params = filter_params_x;
+      filter_size = filter_params_y.taps;
+#else
+      filter_size = filter_params.taps;
+#endif
+      intermediate_height =
+          (((h - 1) * y_step_q4 + subpel_y_q4) >> SUBPEL_BITS) + filter_size;
+      assert(intermediate_height <= max_intermediate_size);
+      (void)max_intermediate_size;
+
+      assert(filter_params.taps <= MAX_FILTER_TAP);
+
+      convolve_horiz(src - src_stride * (filter_size / 2 - 1), src_stride, temp,
+                     temp_stride, w, intermediate_height, filter_params,
+                     subpel_x_q4, x_step_q4, &temp_conv_params);
+
+#if CONFIG_DUAL_FILTER
+      filter_params = filter_params_y;
+#endif
+      assert(filter_params.taps <= MAX_FILTER_TAP);
+
+      convolve_vert(temp + temp_stride * (filter_size / 2 - 1), temp_stride,
+                    dst, dst_stride, w, h, filter_params, subpel_y_q4,
+                    y_step_q4, conv_params);
+    }
+  }
+}
+
+void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
+                  int dst_stride, int w, int h,
+#if CONFIG_DUAL_FILTER
+                  const InterpFilter *interp_filter,
+#else
+                  const InterpFilter interp_filter,
+#endif
+                  const int subpel_x_q4, int x_step_q4, const int subpel_y_q4,
+                  int y_step_q4, ConvolveParams *conv_params) {
+  convolve_helper(src, src_stride, dst, dst_stride, w, h, interp_filter,
+                  subpel_x_q4, x_step_q4, subpel_y_q4, y_step_q4, conv_params,
+                  av1_convolve_horiz_facade, av1_convolve_vert_facade);
+}
+
+void av1_convolve_c(const uint8_t *src, int src_stride, uint8_t *dst,
+                    int dst_stride, int w, int h,
+#if CONFIG_DUAL_FILTER
+                    const InterpFilter *interp_filter,
+#else
+                    const InterpFilter interp_filter,
+#endif
+                    const int subpel_x_q4, int x_step_q4, const int subpel_y_q4,
+                    int y_step_q4, ConvolveParams *conv_params) {
+  convolve_helper(src, src_stride, dst, dst_stride, w, h, interp_filter,
+                  subpel_x_q4, x_step_q4, subpel_y_q4, y_step_q4, conv_params,
+                  av1_convolve_horiz_facade_c, av1_convolve_vert_facade_c);
+}
+
+void av1_lowbd_convolve_init_c(void) {
+  // A placeholder for SIMD initialization
+  return;
+}
+
+void av1_highbd_convolve_init_c(void) {
+  // A placeholder for SIMD initialization
+  return;
+}
+
+void av1_convolve_init(AV1_COMMON *cm) {
+#if CONFIG_HIGHBITDEPTH
+  if (cm->use_highbitdepth)
+    av1_highbd_convolve_init();
+  else
+    av1_lowbd_convolve_init();
+#else
+  (void)cm;
+  av1_lowbd_convolve_init();
+#endif
+  return;
+}
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_convolve_horiz_c(const uint16_t *src, int src_stride,
+                                 uint16_t *dst, int dst_stride, int w, int h,
+                                 const InterpFilterParams filter_params,
+                                 const int subpel_x_q4, int x_step_q4, int avg,
+                                 int bd) {
+  int x, y;
+  int filter_size = filter_params.taps;
+  src -= filter_size / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = subpel_x_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+          filter_params, x_q4 & SUBPEL_MASK);
+      int k, sum = 0;
+      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];
+      if (avg)
+        dst[x] = ROUND_POWER_OF_TWO(
+            dst[x] +
+                clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
+            1);
+      else
+        dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void av1_highbd_convolve_vert_c(const uint16_t *src, int src_stride,
+                                uint16_t *dst, int dst_stride, int w, int h,
+                                const InterpFilterParams filter_params,
+                                const int subpel_y_q4, int y_step_q4, int avg,
+                                int bd) {
+  int x, y;
+  int filter_size = filter_params.taps;
+  src -= src_stride * (filter_size / 2 - 1);
+
+  for (x = 0; x < w; ++x) {
+    int y_q4 = subpel_y_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *const src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+          filter_params, y_q4 & SUBPEL_MASK);
+      int k, sum = 0;
+      for (k = 0; k < filter_size; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      if (avg) {
+        dst[y * dst_stride] = ROUND_POWER_OF_TWO(
+            dst[y * dst_stride] +
+                clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
+            1);
+      } else {
+        dst[y * dst_stride] =
+            clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      }
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void highbd_convolve_copy(const uint16_t *src, int src_stride,
+                                 uint16_t *dst, int dst_stride, int w, int h,
+                                 int avg, int bd) {
+  if (avg == 0) {
+    int r;
+    for (r = 0; r < h; ++r) {
+      memcpy(dst, src, w * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+    }
+  } else {
+    int r, c;
+    for (r = 0; r < h; ++r) {
+      for (c = 0; c < w; ++c) {
+        dst[c] = clip_pixel_highbd(ROUND_POWER_OF_TWO(dst[c] + src[c], 1), bd);
+      }
+      src += src_stride;
+      dst += dst_stride;
+    }
+  }
+}
+
+void av1_highbd_convolve_horiz_facade(const uint8_t *src8, int src_stride,
+                                      uint8_t *dst8, int dst_stride, int w,
+                                      int h,
+                                      const InterpFilterParams filter_params,
+                                      const int subpel_x_q4, int x_step_q4,
+                                      int avg, int bd) {
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  if (filter_params.taps == SUBPEL_TAPS) {
+    const int16_t *filter_x =
+        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
+    if (avg == 0)
+      aom_highbd_convolve8_horiz(src8, src_stride, dst8, dst_stride, filter_x,
+                                 x_step_q4, NULL, -1, w, h, bd);
+    else
+      aom_highbd_convolve8_avg_horiz(src8, src_stride, dst8, dst_stride,
+                                     filter_x, x_step_q4, NULL, -1, w, h, bd);
+  } else {
+    av1_highbd_convolve_horiz(src, src_stride, dst, dst_stride, w, h,
+                              filter_params, subpel_x_q4, x_step_q4, avg, bd);
+  }
+}
+
+void av1_highbd_convolve_vert_facade(const uint8_t *src8, int src_stride,
+                                     uint8_t *dst8, int dst_stride, int w,
+                                     int h,
+                                     const InterpFilterParams filter_params,
+                                     const int subpel_y_q4, int y_step_q4,
+                                     int avg, int bd) {
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+
+  if (filter_params.taps == SUBPEL_TAPS) {
+    const int16_t *filter_y =
+        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
+    if (avg == 0) {
+      aom_highbd_convolve8_vert(src8, src_stride, dst8, dst_stride, NULL, -1,
+                                filter_y, y_step_q4, w, h, bd);
+    } else {
+      aom_highbd_convolve8_avg_vert(src8, src_stride, dst8, dst_stride, NULL,
+                                    -1, filter_y, y_step_q4, w, h, bd);
+    }
+  } else {
+    av1_highbd_convolve_vert(src, src_stride, dst, dst_stride, w, h,
+                             filter_params, subpel_y_q4, y_step_q4, avg, bd);
+  }
+}
+
+void av1_highbd_convolve(const uint8_t *src8, int src_stride, uint8_t *dst8,
+                         int dst_stride, int w, int h,
+#if CONFIG_DUAL_FILTER
+                         const InterpFilter *interp_filter,
+#else
+                         const InterpFilter interp_filter,
+#endif
+                         const int subpel_x_q4, int x_step_q4,
+                         const int subpel_y_q4, int y_step_q4, int ref_idx,
+                         int bd) {
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  int ignore_horiz = x_step_q4 == 16 && subpel_x_q4 == 0;
+  int ignore_vert = y_step_q4 == 16 && subpel_y_q4 == 0;
+
+  assert(w <= MAX_BLOCK_WIDTH);
+  assert(h <= MAX_BLOCK_HEIGHT);
+  assert(y_step_q4 <= MAX_STEP);
+  assert(x_step_q4 <= MAX_STEP);
+
+  if (ignore_horiz && ignore_vert) {
+    highbd_convolve_copy(src, src_stride, dst, dst_stride, w, h, ref_idx, bd);
+  } else if (ignore_vert) {
+#if CONFIG_DUAL_FILTER
+    InterpFilterParams filter_params =
+        av1_get_interp_filter_params(interp_filter[1 + 2 * ref_idx]);
+#else
+    InterpFilterParams filter_params =
+        av1_get_interp_filter_params(interp_filter);
+#endif
+    av1_highbd_convolve_horiz_facade(src8, src_stride, dst8, dst_stride, w, h,
+                                     filter_params, subpel_x_q4, x_step_q4,
+                                     ref_idx, bd);
+  } else if (ignore_horiz) {
+#if CONFIG_DUAL_FILTER
+    InterpFilterParams filter_params =
+        av1_get_interp_filter_params(interp_filter[0 + 2 * ref_idx]);
+#else
+    InterpFilterParams filter_params =
+        av1_get_interp_filter_params(interp_filter);
+#endif
+    av1_highbd_convolve_vert_facade(src8, src_stride, dst8, dst_stride, w, h,
+                                    filter_params, subpel_y_q4, y_step_q4,
+                                    ref_idx, bd);
+  } else {
+    // temp's size is set to a 256 aligned value to facilitate SIMD
+    // implementation. The value is greater than (maximum possible intermediate
+    // height or width) * MAX_SB_SIZE
+    DECLARE_ALIGNED(16, uint16_t,
+                    temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]);
+    uint8_t *temp8 = CONVERT_TO_BYTEPTR(temp);
+    int max_intermediate_size = ((MAX_SB_SIZE * 2 + 16) + 16);
+    int filter_size;
+    InterpFilterParams filter_params;
+#if CONFIG_DUAL_FILTER
+    InterpFilterParams filter_params_x =
+        av1_get_interp_filter_params(interp_filter[1 + 2 * ref_idx]);
+    InterpFilterParams filter_params_y =
+        av1_get_interp_filter_params(interp_filter[0 + 2 * ref_idx]);
+    if (interp_filter[0 + 2 * ref_idx] == MULTITAP_SHARP &&
+        interp_filter[1 + 2 * ref_idx] == MULTITAP_SHARP) {
+      // Avoid two directions both using 12-tap filter.
+      // This will reduce hardware implementation cost.
+      filter_params_y = av1_get_interp_filter_params(EIGHTTAP_SHARP);
+    }
+#endif
+
+#if CONFIG_DUAL_FILTER
+    if (filter_params_y.taps < filter_params_x.taps) {
+      int intermediate_width;
+      int temp_stride = max_intermediate_size;
+      filter_params = filter_params_y;
+      filter_size = filter_params_x.taps;
+      intermediate_width =
+          (((w - 1) * x_step_q4 + subpel_x_q4) >> SUBPEL_BITS) + filter_size;
+      assert(intermediate_width <= max_intermediate_size);
+
+      assert(filter_params.taps <= MAX_FILTER_TAP);
+
+      av1_highbd_convolve_vert_facade(
+          src8 - (filter_size / 2 - 1), src_stride, temp8, temp_stride,
+          intermediate_width, h, filter_params, subpel_y_q4, y_step_q4, 0, bd);
+
+      filter_params = filter_params_x;
+      assert(filter_params.taps <= MAX_FILTER_TAP);
+
+      av1_highbd_convolve_horiz_facade(
+          temp8 + (filter_size / 2 - 1), temp_stride, dst8, dst_stride, w, h,
+          filter_params, subpel_x_q4, x_step_q4, ref_idx, bd);
+    } else
+#endif  // CONFIG_DUAL_FILTER
+    {
+      int intermediate_height;
+      int temp_stride = MAX_SB_SIZE;
+#if CONFIG_DUAL_FILTER
+      filter_params = filter_params_x;
+      filter_size = filter_params_y.taps;
+#else
+      filter_params = av1_get_interp_filter_params(interp_filter);
+      filter_size = filter_params.taps;
+#endif
+      intermediate_height =
+          (((h - 1) * y_step_q4 + subpel_y_q4) >> SUBPEL_BITS) + filter_size;
+      assert(intermediate_height <= max_intermediate_size);
+      (void)max_intermediate_size;
+
+      av1_highbd_convolve_horiz_facade(
+          src8 - src_stride * (filter_size / 2 - 1), src_stride, temp8,
+          temp_stride, w, intermediate_height, filter_params, subpel_x_q4,
+          x_step_q4, 0, bd);
+
+#if CONFIG_DUAL_FILTER
+      filter_params = filter_params_y;
+#endif
+      filter_size = filter_params.taps;
+      assert(filter_params.taps <= MAX_FILTER_TAP);
+
+      av1_highbd_convolve_vert_facade(
+          temp8 + temp_stride * (filter_size / 2 - 1), temp_stride, dst8,
+          dst_stride, w, h, filter_params, subpel_y_q4, y_step_q4, ref_idx, bd);
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/av1/common/convolve.h b/third_party/aom/av1/common/convolve.h
new file mode 100644
index 0000000000..4a4dd8cdb1
--- /dev/null
+++ b/third_party/aom/av1/common/convolve.h
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_AV1_CONVOLVE_H_
+#define AV1_COMMON_AV1_CONVOLVE_H_
+#include "av1/common/filter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef enum CONVOLVE_OPT {
+  // indicate the results in dst buf is rounded by FILTER_BITS or not
+  CONVOLVE_OPT_ROUND,
+  CONVOLVE_OPT_NO_ROUND,
+} CONVOLVE_OPT;
+
+typedef int32_t CONV_BUF_TYPE;
+
+typedef struct ConvolveParams {
+  int ref;
+  CONVOLVE_OPT round;
+  CONV_BUF_TYPE *dst;
+  int dst_stride;
+  int round_0;
+  int round_1;
+  int plane;
+} ConvolveParams;
+
+static INLINE ConvolveParams get_conv_params(int ref, int plane) {
+  ConvolveParams conv_params;
+  conv_params.ref = ref;
+  conv_params.round = CONVOLVE_OPT_ROUND;
+  conv_params.plane = plane;
+  return conv_params;
+}
+struct AV1Common;
+void av1_convolve_init(struct AV1Common *cm);
+#if CONFIG_CONVOLVE_ROUND
+void av1_convolve_2d(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst,
+                     int dst_stride, int w, int h,
+                     InterpFilterParams *filter_params_x,
+                     InterpFilterParams *filter_params_y, const int subpel_x_q4,
+                     const int subpel_y_q4, ConvolveParams *conv_params);
+
+void av1_convolve_2d_facade(const uint8_t *src, int src_stride, uint8_t *dst,
+                            int dst_stride, int w, int h,
+                            const InterpFilter *interp_filter,
+                            const int subpel_x_q4, int x_step_q4,
+                            const int subpel_y_q4, int y_step_q4,
+                            ConvolveParams *conv_params);
+
+static INLINE ConvolveParams get_conv_params_no_round(int ref, int plane,
+                                                      int32_t *dst,
+                                                      int dst_stride) {
+  ConvolveParams conv_params;
+  conv_params.ref = ref;
+  conv_params.round = CONVOLVE_OPT_NO_ROUND;
+#if CONFIG_COMPOUND_ROUND
+  conv_params.round_0 = FILTER_BITS;
+#else
+  conv_params.round_0 = 5;
+#endif
+  conv_params.round_1 = 0;
+  conv_params.dst = dst;
+  conv_params.dst_stride = dst_stride;
+  conv_params.plane = plane;
+  return conv_params;
+}
+
+void av1_convolve_rounding(const int32_t *src, int src_stride, uint8_t *dst,
+                           int dst_stride, int w, int h, int bits);
+#endif  // CONFIG_CONVOLVE_ROUND
+
+void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
+                  int dst_stride, int w, int h,
+#if CONFIG_DUAL_FILTER
+                  const InterpFilter *interp_filter,
+#else
+                  const InterpFilter interp_filter,
+#endif
+                  const int subpel_x, int xstep, const int subpel_y, int ystep,
+                  ConvolveParams *conv_params);
+
+void av1_convolve_c(const uint8_t *src, int src_stride, uint8_t *dst,
+                    int dst_stride, int w, int h,
+#if CONFIG_DUAL_FILTER
+                    const InterpFilter *interp_filter,
+#else
+                    const InterpFilter interp_filter,
+#endif
+                    const int subpel_x, int xstep, const int subpel_y,
+                    int ystep, ConvolveParams *conv_params);
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
+                         int dst_stride, int w, int h,
+#if CONFIG_DUAL_FILTER
+                         const InterpFilter *interp_filter,
+#else
+                         const InterpFilter interp_filter,
+#endif
+                         const int subpel_x, int xstep, const int subpel_y,
+                         int ystep, int avg, int bd);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_AV1_CONVOLVE_H_
diff --git a/third_party/aom/av1/common/debugmodes.c b/third_party/aom/av1/common/debugmodes.c
new file mode 100644
index 0000000000..d7b31c1e4c
--- /dev/null
+++ b/third_party/aom/av1/common/debugmodes.c
@@ -0,0 +1,90 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdio.h>
+
+#include "av1/common/blockd.h"
+#include "av1/common/onyxc_int.h"
+
+static void log_frame_info(AV1_COMMON *cm, const char *str, FILE *f) {
+  fprintf(f, "%s", str);
+  fprintf(f, "(Frame %d, Show:%d, Q:%d): \n", cm->current_video_frame,
+          cm->show_frame, cm->base_qindex);
+}
+/* This function dereferences a pointer to the mbmi structure
+ * and uses the passed in member offset to print out the value of an integer
+ * for each mbmi member value in the mi structure.
+ */
+static void print_mi_data(AV1_COMMON *cm, FILE *file, const char *descriptor,
+                          size_t member_offset) {
+  int mi_row, mi_col;
+  MODE_INFO **mi = cm->mi_grid_visible;
+  int rows = cm->mi_rows;
+  int cols = cm->mi_cols;
+  char prefix = descriptor[0];
+
+  log_frame_info(cm, descriptor, file);
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    fprintf(file, "%c ", prefix);
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      fprintf(file, "%2d ", *((int *)((char *)(&mi[0]->mbmi) + member_offset)));
+      mi++;
+    }
+    fprintf(file, "\n");
+    mi += 8;
+  }
+  fprintf(file, "\n");
+}
+
+void av1_print_modes_and_motion_vectors(AV1_COMMON *cm, const char *file) {
+  int mi_row;
+  int mi_col;
+  FILE *mvs = fopen(file, "a");
+  MODE_INFO **mi = cm->mi_grid_visible;
+  int rows = cm->mi_rows;
+  int cols = cm->mi_cols;
+
+  print_mi_data(cm, mvs, "Partitions:", offsetof(MB_MODE_INFO, sb_type));
+  print_mi_data(cm, mvs, "Modes:", offsetof(MB_MODE_INFO, mode));
+  print_mi_data(cm, mvs, "Ref frame:", offsetof(MB_MODE_INFO, ref_frame[0]));
+  print_mi_data(cm, mvs, "Transform:", offsetof(MB_MODE_INFO, tx_size));
+  print_mi_data(cm, mvs, "UV Modes:", offsetof(MB_MODE_INFO, uv_mode));
+
+  // output skip infomation.
+  log_frame_info(cm, "Skips:", mvs);
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    fprintf(mvs, "S ");
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      fprintf(mvs, "%2d ", mi[0]->mbmi.skip);
+      mi++;
+    }
+    fprintf(mvs, "\n");
+    mi += 8;
+  }
+  fprintf(mvs, "\n");
+
+  // output motion vectors.
+  log_frame_info(cm, "Vectors ", mvs);
+  mi = cm->mi_grid_visible;
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    fprintf(mvs, "V ");
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      fprintf(mvs, "%4d:%4d ", mi[0]->mbmi.mv[0].as_mv.row,
+              mi[0]->mbmi.mv[0].as_mv.col);
+      mi++;
+    }
+    fprintf(mvs, "\n");
+    mi += 8;
+  }
+  fprintf(mvs, "\n");
+
+  fclose(mvs);
+}
diff --git a/third_party/aom/av1/common/entropy.c b/third_party/aom/av1/common/entropy.c
new file mode 100644
index 0000000000..14ab53ca00
--- /dev/null
+++ b/third_party/aom/av1/common/entropy.c
@@ -0,0 +1,6438 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_mem/aom_mem.h"
+#include "av1/common/blockd.h"
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/scan.h"
+#if CONFIG_LV_MAP
+#include "av1/common/txb_common.h"
+#endif
+
+// Unconstrained Node Tree
+/* clang-format off */
+const aom_tree_index av1_coef_con_tree[TREE_SIZE(ENTROPY_TOKENS)] = {
+  2, 6,                                // 0 = LOW_VAL
+  -TWO_TOKEN, 4,                       // 1 = TWO
+  -THREE_TOKEN, -FOUR_TOKEN,           // 2 = THREE
+  8, 10,                               // 3 = HIGH_LOW
+  -CATEGORY1_TOKEN, -CATEGORY2_TOKEN,  // 4 = CAT_ONE
+  12, 14,                              // 5 = CAT_THREEFOUR
+  -CATEGORY3_TOKEN, -CATEGORY4_TOKEN,  // 6 = CAT_THREE
+  -CATEGORY5_TOKEN, -CATEGORY6_TOKEN   // 7 = CAT_FIVE
+};
+/* clang-format on */
+
+#if CONFIG_NEW_MULTISYMBOL
+/* Extra bits coded from LSB to MSB */
+const aom_cdf_prob av1_cat1_cdf0[CDF_SIZE(2)] = { AOM_ICDF(20352),
+                                                  AOM_ICDF(32768), 0 };
+const aom_cdf_prob *av1_cat1_cdf[] = { av1_cat1_cdf0 };
+
+const aom_cdf_prob av1_cat2_cdf0[CDF_SIZE(4)] = {
+  AOM_ICDF(11963), AOM_ICDF(21121), AOM_ICDF(27719), AOM_ICDF(32768), 0
+};
+const aom_cdf_prob *av1_cat2_cdf[] = { av1_cat2_cdf0 };
+const aom_cdf_prob av1_cat3_cdf0[CDF_SIZE(8)] = {
+  AOM_ICDF(7001),  AOM_ICDF(12802), AOM_ICDF(17911),
+  AOM_ICDF(22144), AOM_ICDF(25503), AOM_ICDF(28286),
+  AOM_ICDF(30737), AOM_ICDF(32768), 0
+};
+const aom_cdf_prob *av1_cat3_cdf[] = { av1_cat3_cdf0 };
+
+const aom_cdf_prob av1_cat4_cdf0[CDF_SIZE(16)] = { AOM_ICDF(3934),
+                                                   AOM_ICDF(7460),
+                                                   AOM_ICDF(10719),
+                                                   AOM_ICDF(13640),
+                                                   AOM_ICDF(16203),
+                                                   AOM_ICDF(18500),
+                                                   AOM_ICDF(20624),
+                                                   AOM_ICDF(22528),
+                                                   AOM_ICDF(24316),
+                                                   AOM_ICDF(25919),
+                                                   AOM_ICDF(27401),
+                                                   AOM_ICDF(28729),
+                                                   AOM_ICDF(29894),
+                                                   AOM_ICDF(30938),
+                                                   AOM_ICDF(31903),
+                                                   AOM_ICDF(32768),
+                                                   0 };
+const aom_cdf_prob *av1_cat4_cdf[] = { av1_cat4_cdf0 };
+
+const aom_cdf_prob av1_cat5_cdf0[CDF_SIZE(16)] = { AOM_ICDF(2942),
+                                                   AOM_ICDF(5794),
+                                                   AOM_ICDF(8473),
+                                                   AOM_ICDF(11069),
+                                                   AOM_ICDF(13469),
+                                                   AOM_ICDF(15795),
+                                                   AOM_ICDF(17980),
+                                                   AOM_ICDF(20097),
+                                                   AOM_ICDF(21952),
+                                                   AOM_ICDF(23750),
+                                                   AOM_ICDF(25439),
+                                                   AOM_ICDF(27076),
+                                                   AOM_ICDF(28589),
+                                                   AOM_ICDF(30056),
+                                                   AOM_ICDF(31434),
+                                                   AOM_ICDF(32768),
+                                                   0 };
+const aom_cdf_prob av1_cat5_cdf1[CDF_SIZE(2)] = { AOM_ICDF(23040),
+                                                  AOM_ICDF(32768), 0 };
+const aom_cdf_prob *av1_cat5_cdf[] = { av1_cat5_cdf0, av1_cat5_cdf1 };
+
+const aom_cdf_prob av1_cat6_cdf0[CDF_SIZE(16)] = {
+  AOM_ICDF(2382),  AOM_ICDF(4727),  AOM_ICDF(7036),  AOM_ICDF(9309),
+  AOM_ICDF(11512), AOM_ICDF(13681), AOM_ICDF(15816), AOM_ICDF(17918),
+  AOM_ICDF(19892), AOM_ICDF(21835), AOM_ICDF(23748), AOM_ICDF(25632),
+  AOM_ICDF(27458), AOM_ICDF(29255), AOM_ICDF(31024), AOM_ICDF(32768)
+};
+const aom_cdf_prob av1_cat6_cdf1[CDF_SIZE(16)] = {
+  AOM_ICDF(9314),  AOM_ICDF(15584), AOM_ICDF(19741), AOM_ICDF(22540),
+  AOM_ICDF(25391), AOM_ICDF(27310), AOM_ICDF(28583), AOM_ICDF(29440),
+  AOM_ICDF(30493), AOM_ICDF(31202), AOM_ICDF(31672), AOM_ICDF(31988),
+  AOM_ICDF(32310), AOM_ICDF(32527), AOM_ICDF(32671), AOM_ICDF(32768)
+};
+const aom_cdf_prob av1_cat6_cdf2[CDF_SIZE(16)] = {
+  AOM_ICDF(29548), AOM_ICDF(31129), AOM_ICDF(31960), AOM_ICDF(32004),
+  AOM_ICDF(32473), AOM_ICDF(32498), AOM_ICDF(32511), AOM_ICDF(32512),
+  AOM_ICDF(32745), AOM_ICDF(32757), AOM_ICDF(32763), AOM_ICDF(32764),
+  AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)
+};
+const aom_cdf_prob av1_cat6_cdf3[CDF_SIZE(16)] = {
+  AOM_ICDF(32006), AOM_ICDF(32258), AOM_ICDF(32510), AOM_ICDF(32512),
+  AOM_ICDF(32638), AOM_ICDF(32639), AOM_ICDF(32640), AOM_ICDF(32641),
+  AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+  AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)
+};
+const aom_cdf_prob av1_cat6_cdf4[CDF_SIZE(4)] = {
+  AOM_ICDF(32513), AOM_ICDF(32641), AOM_ICDF(32767), AOM_ICDF(32768)
+};
+const aom_cdf_prob *av1_cat6_cdf[] = {
+  av1_cat6_cdf0, av1_cat6_cdf1, av1_cat6_cdf2, av1_cat6_cdf3, av1_cat6_cdf4
+};
+#endif
+/* Extra bits coded from MSB to LSB */
+const aom_prob av1_cat1_prob[] = { 159 };
+const aom_prob av1_cat2_prob[] = { 165, 145 };
+const aom_prob av1_cat3_prob[] = { 173, 148, 140 };
+const aom_prob av1_cat4_prob[] = { 176, 155, 140, 135 };
+const aom_prob av1_cat5_prob[] = { 180, 157, 141, 134, 130 };
+const aom_prob av1_cat6_prob[] = {
+  255, 255, 255, 255, 254, 254, 254, 252, 249,
+  243, 230, 196, 177, 153, 140, 133, 130, 129
+};
+
+const uint16_t band_count_table[TX_SIZES_ALL][8] = {
+#if CONFIG_CB4X4
+  { 1, 2, 2, 3, 0, 0, 0 },
+#endif
+  { 1, 2, 3, 4, 3, 16 - 13, 0 },    { 1, 2, 3, 4, 11, 64 - 21, 0 },
+  { 1, 2, 3, 4, 11, 256 - 21, 0 },  { 1, 2, 3, 4, 11, 1024 - 21, 0 },
+#if CONFIG_TX64X64
+  { 1, 2, 3, 4, 11, 4096 - 21, 0 },
+#endif  // CONFIG_TX64X64
+  { 1, 2, 3, 4, 8, 32 - 18, 0 },    { 1, 2, 3, 4, 8, 32 - 18, 0 },
+  { 1, 2, 3, 4, 11, 128 - 21, 0 },  { 1, 2, 3, 4, 11, 128 - 21, 0 },
+  { 1, 2, 3, 4, 11, 512 - 21, 0 },  { 1, 2, 3, 4, 11, 512 - 21, 0 },
+  { 1, 2, 3, 4, 11, 64 - 21, 0 },   { 1, 2, 3, 4, 11, 64 - 21, 0 },
+  { 1, 2, 3, 4, 11, 256 - 21, 0 },  { 1, 2, 3, 4, 11, 256 - 21, 0 },
+};
+
+const uint16_t band_cum_count_table[TX_SIZES_ALL][8] = {
+#if CONFIG_CB4X4
+  { 0, 1, 3, 6, 10, 13, 16, 0 },
+#endif
+  { 0, 1, 3, 6, 10, 13, 16, 0 },   { 0, 1, 3, 6, 10, 21, 64, 0 },
+  { 0, 1, 3, 6, 10, 21, 256, 0 },  { 0, 1, 3, 6, 10, 21, 1024, 0 },
+#if CONFIG_TX64X64
+  { 0, 1, 3, 6, 10, 21, 4096, 0 },
+#endif  // CONFIG_TX64X64
+  { 0, 1, 3, 6, 10, 18, 32, 0 },   { 0, 1, 3, 6, 10, 18, 32, 0 },
+  { 0, 1, 3, 6, 10, 21, 128, 0 },  { 0, 1, 3, 6, 10, 21, 128, 0 },
+  { 0, 1, 3, 6, 10, 21, 512, 0 },  { 0, 1, 3, 6, 10, 21, 512, 0 },
+  { 0, 1, 3, 6, 10, 21, 64, 0 },   { 0, 1, 3, 6, 10, 21, 64, 0 },
+  { 0, 1, 3, 6, 10, 21, 256, 0 },  { 0, 1, 3, 6, 10, 21, 256, 0 },
+};
+
+const uint8_t av1_coefband_trans_8x8plus[MAX_TX_SQUARE] = {
+  0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 5,
+  // beyond MAXBAND_INDEX+1 all values are filled as 5
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+#if CONFIG_TX64X64
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5
+#endif  // CONFIG_TX64X64
+};
+
+const uint8_t av1_coefband_trans_4x8_8x4[32] = {
+  0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4,
+  4, 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+};
+
+const uint8_t av1_coefband_trans_4x4[16] = {
+  0, 1, 1, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4, 5, 5, 5,
+};
+
+const uint8_t av1_pt_energy_class[ENTROPY_TOKENS] = { 0, 1, 2, 3, 3, 4,
+                                                      4, 5, 5, 5, 5, 5 };
+
+// Model obtained from a 2-sided zero-centered distribution derived
+// from a Pareto distribution. The cdf of the distribution is:
+// cdf(x) = 0.5 + 0.5 * sgn(x) * [1 - {alpha/(alpha + |x|)} ^ beta]
+//
+// For a given beta and a given probablity of the 1-node, the alpha
+// is first solved, and then the {alpha, beta} pair is used to generate
+// the probabilities for the rest of the nodes.
+
+// beta = 8
+
+// Every odd line in this table can be generated from the even lines
+// by averaging :
+// av1_pareto8_full[l][node] = (av1_pareto8_full[l-1][node] +
+//                              av1_pareto8_full[l+1][node] ) >> 1;
+// Values for tokens ONE_TOKEN through CATEGORY6_TOKEN included here.
+const aom_prob av1_pareto8_full[COEFF_PROB_MODELS][MODEL_NODES] = {
+  { 3, 86, 128, 6, 86, 23, 88, 29 },
+  { 6, 86, 128, 11, 87, 42, 91, 52 },
+  { 9, 86, 129, 17, 88, 61, 94, 76 },
+  { 12, 86, 129, 22, 88, 77, 97, 93 },
+  { 15, 87, 129, 28, 89, 93, 100, 110 },
+  { 17, 87, 129, 33, 90, 105, 103, 123 },
+  { 20, 88, 130, 38, 91, 118, 106, 136 },
+  { 23, 88, 130, 43, 91, 128, 108, 146 },
+  { 26, 89, 131, 48, 92, 139, 111, 156 },
+  { 28, 89, 131, 53, 93, 147, 114, 163 },
+  { 31, 90, 131, 58, 94, 156, 117, 171 },
+  { 34, 90, 131, 62, 94, 163, 119, 177 },
+  { 37, 90, 132, 66, 95, 171, 122, 184 },
+  { 39, 90, 132, 70, 96, 177, 124, 189 },
+  { 42, 91, 132, 75, 97, 183, 127, 194 },
+  { 44, 91, 132, 79, 97, 188, 129, 198 },
+  { 47, 92, 133, 83, 98, 193, 132, 202 },
+  { 49, 92, 133, 86, 99, 197, 134, 205 },
+  { 52, 93, 133, 90, 100, 201, 137, 208 },
+  { 54, 93, 133, 94, 100, 204, 139, 211 },
+  { 57, 94, 134, 98, 101, 208, 142, 214 },
+  { 59, 94, 134, 101, 102, 211, 144, 216 },
+  { 62, 94, 135, 105, 103, 214, 146, 218 },
+  { 64, 94, 135, 108, 103, 216, 148, 220 },
+  { 66, 95, 135, 111, 104, 219, 151, 222 },
+  { 68, 95, 135, 114, 105, 221, 153, 223 },
+  { 71, 96, 136, 117, 106, 224, 155, 225 },
+  { 73, 96, 136, 120, 106, 225, 157, 226 },
+  { 76, 97, 136, 123, 107, 227, 159, 228 },
+  { 78, 97, 136, 126, 108, 229, 160, 229 },
+  { 80, 98, 137, 129, 109, 231, 162, 231 },
+  { 82, 98, 137, 131, 109, 232, 164, 232 },
+  { 84, 98, 138, 134, 110, 234, 166, 233 },
+  { 86, 98, 138, 137, 111, 235, 168, 234 },
+  { 89, 99, 138, 140, 112, 236, 170, 235 },
+  { 91, 99, 138, 142, 112, 237, 171, 235 },
+  { 93, 100, 139, 145, 113, 238, 173, 236 },
+  { 95, 100, 139, 147, 114, 239, 174, 237 },
+  { 97, 101, 140, 149, 115, 240, 176, 238 },
+  { 99, 101, 140, 151, 115, 241, 177, 238 },
+  { 101, 102, 140, 154, 116, 242, 179, 239 },
+  { 103, 102, 140, 156, 117, 242, 180, 239 },
+  { 105, 103, 141, 158, 118, 243, 182, 240 },
+  { 107, 103, 141, 160, 118, 243, 183, 240 },
+  { 109, 104, 141, 162, 119, 244, 185, 241 },
+  { 111, 104, 141, 164, 119, 244, 186, 241 },
+  { 113, 104, 142, 166, 120, 245, 187, 242 },
+  { 114, 104, 142, 168, 121, 245, 188, 242 },
+  { 116, 105, 143, 170, 122, 246, 190, 243 },
+  { 118, 105, 143, 171, 122, 246, 191, 243 },
+  { 120, 106, 143, 173, 123, 247, 192, 244 },
+  { 121, 106, 143, 175, 124, 247, 193, 244 },
+  { 123, 107, 144, 177, 125, 248, 195, 244 },
+  { 125, 107, 144, 178, 125, 248, 196, 244 },
+  { 127, 108, 145, 180, 126, 249, 197, 245 },
+  { 128, 108, 145, 181, 127, 249, 198, 245 },
+  { 130, 109, 145, 183, 128, 249, 199, 245 },
+  { 132, 109, 145, 184, 128, 249, 200, 245 },
+  { 134, 110, 146, 186, 129, 250, 201, 246 },
+  { 135, 110, 146, 187, 130, 250, 202, 246 },
+  { 137, 111, 147, 189, 131, 251, 203, 246 },
+  { 138, 111, 147, 190, 131, 251, 204, 246 },
+  { 140, 112, 147, 192, 132, 251, 205, 247 },
+  { 141, 112, 147, 193, 132, 251, 206, 247 },
+  { 143, 113, 148, 194, 133, 251, 207, 247 },
+  { 144, 113, 148, 195, 134, 251, 207, 247 },
+  { 146, 114, 149, 197, 135, 252, 208, 248 },
+  { 147, 114, 149, 198, 135, 252, 209, 248 },
+  { 149, 115, 149, 199, 136, 252, 210, 248 },
+  { 150, 115, 149, 200, 137, 252, 210, 248 },
+  { 152, 115, 150, 201, 138, 252, 211, 248 },
+  { 153, 115, 150, 202, 138, 252, 212, 248 },
+  { 155, 116, 151, 204, 139, 253, 213, 249 },
+  { 156, 116, 151, 205, 139, 253, 213, 249 },
+  { 158, 117, 151, 206, 140, 253, 214, 249 },
+  { 159, 117, 151, 207, 141, 253, 215, 249 },
+  { 161, 118, 152, 208, 142, 253, 216, 249 },
+  { 162, 118, 152, 209, 142, 253, 216, 249 },
+  { 163, 119, 153, 210, 143, 253, 217, 249 },
+  { 164, 119, 153, 211, 143, 253, 217, 249 },
+  { 166, 120, 153, 212, 144, 254, 218, 250 },
+  { 167, 120, 153, 212, 145, 254, 219, 250 },
+  { 168, 121, 154, 213, 146, 254, 220, 250 },
+  { 169, 121, 154, 214, 146, 254, 220, 250 },
+  { 171, 122, 155, 215, 147, 254, 221, 250 },
+  { 172, 122, 155, 216, 147, 254, 221, 250 },
+  { 173, 123, 155, 217, 148, 254, 222, 250 },
+  { 174, 123, 155, 217, 149, 254, 222, 250 },
+  { 176, 124, 156, 218, 150, 254, 223, 250 },
+  { 177, 124, 156, 219, 150, 254, 223, 250 },
+  { 178, 125, 157, 220, 151, 254, 224, 251 },
+  { 179, 125, 157, 220, 151, 254, 224, 251 },
+  { 180, 126, 157, 221, 152, 254, 225, 251 },
+  { 181, 126, 157, 221, 152, 254, 225, 251 },
+  { 183, 127, 158, 222, 153, 254, 226, 251 },
+  { 184, 127, 158, 223, 154, 254, 226, 251 },
+  { 185, 128, 159, 224, 155, 255, 227, 251 },
+  { 186, 128, 159, 224, 155, 255, 227, 251 },
+  { 187, 129, 160, 225, 156, 255, 228, 251 },
+  { 188, 130, 160, 225, 156, 255, 228, 251 },
+  { 189, 131, 160, 226, 157, 255, 228, 251 },
+  { 190, 131, 160, 226, 158, 255, 228, 251 },
+  { 191, 132, 161, 227, 159, 255, 229, 251 },
+  { 192, 132, 161, 227, 159, 255, 229, 251 },
+  { 193, 133, 162, 228, 160, 255, 230, 252 },
+  { 194, 133, 162, 229, 160, 255, 230, 252 },
+  { 195, 134, 163, 230, 161, 255, 231, 252 },
+  { 196, 134, 163, 230, 161, 255, 231, 252 },
+  { 197, 135, 163, 231, 162, 255, 231, 252 },
+  { 198, 135, 163, 231, 162, 255, 231, 252 },
+  { 199, 136, 164, 232, 163, 255, 232, 252 },
+  { 200, 136, 164, 232, 164, 255, 232, 252 },
+  { 201, 137, 165, 233, 165, 255, 233, 252 },
+  { 201, 137, 165, 233, 165, 255, 233, 252 },
+  { 202, 138, 166, 233, 166, 255, 233, 252 },
+  { 203, 138, 166, 233, 166, 255, 233, 252 },
+  { 204, 139, 166, 234, 167, 255, 234, 252 },
+  { 205, 139, 166, 234, 167, 255, 234, 252 },
+  { 206, 140, 167, 235, 168, 255, 235, 252 },
+  { 206, 140, 167, 235, 168, 255, 235, 252 },
+  { 207, 141, 168, 236, 169, 255, 235, 252 },
+  { 208, 141, 168, 236, 170, 255, 235, 252 },
+  { 209, 142, 169, 237, 171, 255, 236, 252 },
+  { 209, 143, 169, 237, 171, 255, 236, 252 },
+  { 210, 144, 169, 237, 172, 255, 236, 252 },
+  { 211, 144, 169, 237, 172, 255, 236, 252 },
+  { 212, 145, 170, 238, 173, 255, 237, 252 },
+  { 213, 145, 170, 238, 173, 255, 237, 252 },
+  { 214, 146, 171, 239, 174, 255, 237, 253 },
+  { 214, 146, 171, 239, 174, 255, 237, 253 },
+  { 215, 147, 172, 240, 175, 255, 238, 253 },
+  { 215, 147, 172, 240, 175, 255, 238, 253 },
+  { 216, 148, 173, 240, 176, 255, 238, 253 },
+  { 217, 148, 173, 240, 176, 255, 238, 253 },
+  { 218, 149, 173, 241, 177, 255, 239, 253 },
+  { 218, 149, 173, 241, 178, 255, 239, 253 },
+  { 219, 150, 174, 241, 179, 255, 239, 253 },
+  { 219, 151, 174, 241, 179, 255, 239, 253 },
+  { 220, 152, 175, 242, 180, 255, 240, 253 },
+  { 221, 152, 175, 242, 180, 255, 240, 253 },
+  { 222, 153, 176, 242, 181, 255, 240, 253 },
+  { 222, 153, 176, 242, 181, 255, 240, 253 },
+  { 223, 154, 177, 243, 182, 255, 240, 253 },
+  { 223, 154, 177, 243, 182, 255, 240, 253 },
+  { 224, 155, 178, 244, 183, 255, 241, 253 },
+  { 224, 155, 178, 244, 183, 255, 241, 253 },
+  { 225, 156, 178, 244, 184, 255, 241, 253 },
+  { 225, 157, 178, 244, 184, 255, 241, 253 },
+  { 226, 158, 179, 244, 185, 255, 242, 253 },
+  { 227, 158, 179, 244, 185, 255, 242, 253 },
+  { 228, 159, 180, 245, 186, 255, 242, 253 },
+  { 228, 159, 180, 245, 186, 255, 242, 253 },
+  { 229, 160, 181, 245, 187, 255, 242, 253 },
+  { 229, 160, 181, 245, 187, 255, 242, 253 },
+  { 230, 161, 182, 246, 188, 255, 243, 253 },
+  { 230, 162, 182, 246, 188, 255, 243, 253 },
+  { 231, 163, 183, 246, 189, 255, 243, 253 },
+  { 231, 163, 183, 246, 189, 255, 243, 253 },
+  { 232, 164, 184, 247, 190, 255, 243, 253 },
+  { 232, 164, 184, 247, 190, 255, 243, 253 },
+  { 233, 165, 185, 247, 191, 255, 244, 253 },
+  { 233, 165, 185, 247, 191, 255, 244, 253 },
+  { 234, 166, 185, 247, 192, 255, 244, 253 },
+  { 234, 167, 185, 247, 192, 255, 244, 253 },
+  { 235, 168, 186, 248, 193, 255, 244, 253 },
+  { 235, 168, 186, 248, 193, 255, 244, 253 },
+  { 236, 169, 187, 248, 194, 255, 244, 253 },
+  { 236, 169, 187, 248, 194, 255, 244, 253 },
+  { 236, 170, 188, 248, 195, 255, 245, 253 },
+  { 236, 170, 188, 248, 195, 255, 245, 253 },
+  { 237, 171, 189, 249, 196, 255, 245, 254 },
+  { 237, 172, 189, 249, 196, 255, 245, 254 },
+  { 238, 173, 190, 249, 197, 255, 245, 254 },
+  { 238, 173, 190, 249, 197, 255, 245, 254 },
+  { 239, 174, 191, 249, 198, 255, 245, 254 },
+  { 239, 174, 191, 249, 198, 255, 245, 254 },
+  { 240, 175, 192, 249, 199, 255, 246, 254 },
+  { 240, 176, 192, 249, 199, 255, 246, 254 },
+  { 240, 177, 193, 250, 200, 255, 246, 254 },
+  { 240, 177, 193, 250, 200, 255, 246, 254 },
+  { 241, 178, 194, 250, 201, 255, 246, 254 },
+  { 241, 178, 194, 250, 201, 255, 246, 254 },
+  { 242, 179, 195, 250, 202, 255, 246, 254 },
+  { 242, 180, 195, 250, 202, 255, 246, 254 },
+  { 242, 181, 196, 250, 203, 255, 247, 254 },
+  { 242, 181, 196, 250, 203, 255, 247, 254 },
+  { 243, 182, 197, 251, 204, 255, 247, 254 },
+  { 243, 183, 197, 251, 204, 255, 247, 254 },
+  { 244, 184, 198, 251, 205, 255, 247, 254 },
+  { 244, 184, 198, 251, 205, 255, 247, 254 },
+  { 244, 185, 199, 251, 206, 255, 247, 254 },
+  { 244, 185, 199, 251, 206, 255, 247, 254 },
+  { 245, 186, 200, 251, 207, 255, 247, 254 },
+  { 245, 187, 200, 251, 207, 255, 247, 254 },
+  { 246, 188, 201, 252, 207, 255, 248, 254 },
+  { 246, 188, 201, 252, 207, 255, 248, 254 },
+  { 246, 189, 202, 252, 208, 255, 248, 254 },
+  { 246, 190, 202, 252, 208, 255, 248, 254 },
+  { 247, 191, 203, 252, 209, 255, 248, 254 },
+  { 247, 191, 203, 252, 209, 255, 248, 254 },
+  { 247, 192, 204, 252, 210, 255, 248, 254 },
+  { 247, 193, 204, 252, 210, 255, 248, 254 },
+  { 248, 194, 205, 252, 211, 255, 248, 254 },
+  { 248, 194, 205, 252, 211, 255, 248, 254 },
+  { 248, 195, 206, 252, 212, 255, 249, 254 },
+  { 248, 196, 206, 252, 212, 255, 249, 254 },
+  { 249, 197, 207, 253, 213, 255, 249, 254 },
+  { 249, 197, 207, 253, 213, 255, 249, 254 },
+  { 249, 198, 208, 253, 214, 255, 249, 254 },
+  { 249, 199, 209, 253, 214, 255, 249, 254 },
+  { 250, 200, 210, 253, 215, 255, 249, 254 },
+  { 250, 200, 210, 253, 215, 255, 249, 254 },
+  { 250, 201, 211, 253, 215, 255, 249, 254 },
+  { 250, 202, 211, 253, 215, 255, 249, 254 },
+  { 250, 203, 212, 253, 216, 255, 249, 254 },
+  { 250, 203, 212, 253, 216, 255, 249, 254 },
+  { 251, 204, 213, 253, 217, 255, 250, 254 },
+  { 251, 205, 213, 253, 217, 255, 250, 254 },
+  { 251, 206, 214, 254, 218, 255, 250, 254 },
+  { 251, 206, 215, 254, 218, 255, 250, 254 },
+  { 252, 207, 216, 254, 219, 255, 250, 254 },
+  { 252, 208, 216, 254, 219, 255, 250, 254 },
+  { 252, 209, 217, 254, 220, 255, 250, 254 },
+  { 252, 210, 217, 254, 220, 255, 250, 254 },
+  { 252, 211, 218, 254, 221, 255, 250, 254 },
+  { 252, 212, 218, 254, 221, 255, 250, 254 },
+  { 253, 213, 219, 254, 222, 255, 250, 254 },
+  { 253, 213, 220, 254, 222, 255, 250, 254 },
+  { 253, 214, 221, 254, 223, 255, 250, 254 },
+  { 253, 215, 221, 254, 223, 255, 250, 254 },
+  { 253, 216, 222, 254, 224, 255, 251, 254 },
+  { 253, 217, 223, 254, 224, 255, 251, 254 },
+  { 253, 218, 224, 254, 225, 255, 251, 254 },
+  { 253, 219, 224, 254, 225, 255, 251, 254 },
+  { 254, 220, 225, 254, 225, 255, 251, 254 },
+  { 254, 221, 226, 254, 225, 255, 251, 254 },
+  { 254, 222, 227, 255, 226, 255, 251, 254 },
+  { 254, 223, 227, 255, 226, 255, 251, 254 },
+  { 254, 224, 228, 255, 227, 255, 251, 254 },
+  { 254, 225, 229, 255, 227, 255, 251, 254 },
+  { 254, 226, 230, 255, 228, 255, 251, 254 },
+  { 254, 227, 230, 255, 229, 255, 251, 254 },
+  { 255, 228, 231, 255, 230, 255, 251, 254 },
+  { 255, 229, 232, 255, 230, 255, 251, 254 },
+  { 255, 230, 233, 255, 231, 255, 252, 254 },
+  { 255, 231, 234, 255, 231, 255, 252, 254 },
+  { 255, 232, 235, 255, 232, 255, 252, 254 },
+  { 255, 233, 236, 255, 232, 255, 252, 254 },
+  { 255, 235, 237, 255, 233, 255, 252, 254 },
+  { 255, 236, 238, 255, 234, 255, 252, 254 },
+  { 255, 238, 240, 255, 235, 255, 252, 255 },
+  { 255, 239, 241, 255, 235, 255, 252, 254 },
+  { 255, 241, 243, 255, 236, 255, 252, 254 },
+  { 255, 243, 245, 255, 237, 255, 252, 254 },
+  { 255, 246, 247, 255, 239, 255, 253, 255 },
+};
+
+// Model obtained from a 2-sided zero-centered distribution derived
+// from a Pareto distribution. The cdf of the distribution is:
+// cdf(x) = 0.5 + 0.5 * sgn(x) * [1 - {alpha/(alpha + |x|)} ^ beta]
+//
+// For a given beta and a given probability of the 1-node, the alpha
+// is first solved, and then the {alpha, beta} pair is used to generate
+// the probabilities for the rest of the nodes.
+//
+// The full source code of the generating program is available in:
+// tools/gen_constrained_tokenset.py
+//
+#if CONFIG_NEW_TOKENSET
+// Values for tokens TWO_TOKEN through CATEGORY6_TOKEN included
+// in the table here : the ONE_TOKEN probability is
+// removed and the probabilities rescaled.
+//
+// ZERO_TOKEN and ONE_TOKEN are coded as one CDF,
+// and EOB_TOKEN is coded as flags outside this coder.
+const aom_cdf_prob av1_pareto8_tail_probs[COEFF_PROB_MODELS][TAIL_NODES] = {
+  { 128, 127, 127, 252, 497, 969, 1839, 3318, 25511 },
+  { 256, 254, 251, 496, 966, 1834, 3308, 5408, 19995 },
+  { 383, 378, 373, 732, 1408, 2605, 4470, 6646, 15773 },
+  { 511, 502, 493, 961, 1824, 3289, 5373, 7298, 12517 },
+  { 638, 625, 611, 1182, 2215, 3894, 6064, 7548, 9991 },
+  { 766, 746, 726, 1396, 2582, 4428, 6578, 7529, 8017 },
+  { 893, 866, 839, 1603, 2927, 4896, 6945, 7332, 6467 },
+  { 1020, 984, 950, 1803, 3250, 5305, 7191, 7022, 5243 },
+  { 1147, 1102, 1059, 1996, 3552, 5659, 7338, 6646, 4269 },
+  { 1274, 1218, 1166, 2183, 3835, 5963, 7403, 6234, 3492 },
+  { 1400, 1334, 1270, 2363, 4099, 6223, 7401, 5809, 2869 },
+  { 1527, 1447, 1372, 2537, 4345, 6442, 7346, 5386, 2366 },
+  { 1654, 1560, 1473, 2704, 4574, 6624, 7247, 4973, 1959 },
+  { 1780, 1672, 1571, 2866, 4787, 6771, 7114, 4579, 1628 },
+  { 1906, 1782, 1667, 3022, 4984, 6889, 6954, 4206, 1358 },
+  { 2032, 1891, 1762, 3172, 5167, 6979, 6773, 3856, 1136 },
+  { 2158, 2000, 1854, 3316, 5335, 7044, 6577, 3530, 954 },
+  { 2284, 2106, 1944, 3455, 5490, 7087, 6370, 3229, 803 },
+  { 2410, 2212, 2032, 3588, 5632, 7109, 6155, 2951, 679 },
+  { 2535, 2317, 2119, 3717, 5761, 7113, 5936, 2695, 575 },
+  { 2661, 2420, 2203, 3840, 5880, 7101, 5714, 2461, 488 },
+  { 2786, 2522, 2286, 3958, 5987, 7074, 5493, 2246, 416 },
+  { 2911, 2624, 2367, 4072, 6083, 7033, 5273, 2050, 355 },
+  { 3037, 2724, 2446, 4180, 6170, 6981, 5055, 1871, 304 },
+  { 3162, 2822, 2523, 4284, 6247, 6919, 4842, 1708, 261 },
+  { 3286, 2920, 2599, 4384, 6315, 6848, 4633, 1559, 224 },
+  { 3411, 3017, 2672, 4478, 6374, 6768, 4430, 1424, 194 },
+  { 3536, 3112, 2745, 4569, 6426, 6681, 4232, 1300, 167 },
+  { 3660, 3207, 2815, 4656, 6469, 6588, 4040, 1188, 145 },
+  { 3785, 3300, 2883, 4738, 6505, 6490, 3855, 1086, 126 },
+  { 3909, 3392, 2950, 4817, 6534, 6387, 3677, 993, 109 },
+  { 4033, 3483, 3015, 4891, 6557, 6281, 3505, 908, 95 },
+  { 4157, 3573, 3079, 4962, 6573, 6170, 3340, 831, 83 },
+  { 4281, 3662, 3141, 5029, 6584, 6058, 3181, 760, 72 },
+  { 4405, 3750, 3201, 5093, 6588, 5943, 3029, 696, 63 },
+  { 4529, 3837, 3260, 5152, 6587, 5826, 2883, 638, 56 },
+  { 4652, 3922, 3317, 5209, 6582, 5709, 2744, 584, 49 },
+  { 4775, 4007, 3373, 5262, 6572, 5590, 2610, 536, 43 },
+  { 4899, 4090, 3427, 5312, 6557, 5470, 2483, 492, 38 },
+  { 5022, 4173, 3480, 5359, 6538, 5351, 2361, 451, 33 },
+  { 5145, 4254, 3531, 5403, 6515, 5231, 2246, 414, 29 },
+  { 5268, 4334, 3581, 5443, 6489, 5112, 2135, 380, 26 },
+  { 5391, 4414, 3629, 5481, 6458, 4993, 2029, 350, 23 },
+  { 5514, 4492, 3676, 5515, 6425, 4875, 1929, 321, 21 },
+  { 5637, 4569, 3721, 5548, 6388, 4758, 1833, 296, 18 },
+  { 5759, 4645, 3766, 5577, 6349, 4642, 1742, 272, 16 },
+  { 5881, 4720, 3808, 5604, 6307, 4528, 1656, 250, 14 },
+  { 6004, 4794, 3849, 5628, 6262, 4414, 1573, 231, 13 },
+  { 6126, 4867, 3890, 5649, 6215, 4302, 1495, 213, 11 },
+  { 6248, 4939, 3928, 5669, 6166, 4192, 1420, 196, 10 },
+  { 6370, 5010, 3966, 5686, 6114, 4083, 1349, 181, 9 },
+  { 6492, 5080, 4002, 5700, 6061, 3976, 1282, 167, 8 },
+  { 6614, 5149, 4037, 5712, 6006, 3871, 1218, 154, 7 },
+  { 6735, 5217, 4070, 5723, 5950, 3767, 1157, 142, 7 },
+  { 6857, 5284, 4103, 5731, 5891, 3666, 1099, 131, 6 },
+  { 6978, 5351, 4134, 5737, 5832, 3566, 1044, 121, 5 },
+  { 7099, 5415, 4164, 5741, 5771, 3469, 992, 112, 5 },
+  { 7221, 5479, 4192, 5743, 5709, 3373, 943, 104, 4 },
+  { 7342, 5542, 4220, 5743, 5646, 3279, 896, 96, 4 },
+  { 7462, 5604, 4246, 5742, 5583, 3187, 851, 89, 4 },
+  { 7584, 5665, 4272, 5739, 5518, 3097, 808, 82, 3 },
+  { 7704, 5725, 4296, 5734, 5453, 3009, 768, 76, 3 },
+  { 7825, 5784, 4318, 5727, 5386, 2924, 730, 71, 3 },
+  { 7945, 5843, 4341, 5719, 5320, 2840, 693, 65, 2 },
+  { 8066, 5900, 4361, 5709, 5252, 2758, 659, 61, 2 },
+  { 8186, 5956, 4381, 5698, 5185, 2678, 626, 56, 2 },
+  { 8306, 6011, 4400, 5685, 5117, 2600, 595, 52, 2 },
+  { 8426, 6066, 4418, 5671, 5049, 2523, 565, 48, 2 },
+  { 8547, 6119, 4434, 5655, 4981, 2449, 537, 45, 1 },
+  { 8666, 6171, 4450, 5638, 4912, 2377, 511, 42, 1 },
+  { 8786, 6223, 4465, 5620, 4843, 2306, 485, 39, 1 },
+  { 8906, 6274, 4478, 5600, 4775, 2237, 461, 36, 1 },
+  { 9025, 6323, 4491, 5580, 4706, 2170, 438, 34, 1 },
+  { 9144, 6372, 4503, 5558, 4637, 2105, 417, 31, 1 },
+  { 9264, 6420, 4514, 5535, 4568, 2041, 396, 29, 1 },
+  { 9383, 6467, 4524, 5511, 4500, 1979, 376, 27, 1 },
+  { 9502, 6513, 4532, 5486, 4432, 1919, 358, 25, 1 },
+  { 9621, 6558, 4541, 5460, 4364, 1860, 340, 23, 1 },
+  { 9740, 6602, 4548, 5433, 4296, 1803, 323, 22, 1 },
+  { 9859, 6645, 4554, 5405, 4229, 1748, 307, 20, 1 },
+  { 9978, 6688, 4559, 5376, 4161, 1694, 292, 19, 1 },
+  { 10096, 6729, 4564, 5347, 4094, 1641, 278, 18, 1 },
+  { 10215, 6770, 4568, 5316, 4028, 1590, 264, 16, 1 },
+  { 10333, 6809, 4571, 5285, 3962, 1541, 251, 15, 1 },
+  { 10452, 6848, 4573, 5253, 3896, 1492, 239, 14, 1 },
+  { 10570, 6886, 4574, 5220, 3831, 1446, 227, 13, 1 },
+  { 10688, 6923, 4575, 5186, 3767, 1400, 216, 12, 1 },
+  { 10806, 6959, 4575, 5152, 3702, 1356, 205, 12, 1 },
+  { 10924, 6994, 4574, 5117, 3639, 1313, 195, 11, 1 },
+  { 11041, 7029, 4572, 5082, 3576, 1271, 186, 10, 1 },
+  { 11159, 7062, 4570, 5046, 3513, 1231, 177, 9, 1 },
+  { 11277, 7095, 4566, 5009, 3451, 1192, 168, 9, 1 },
+  { 11394, 7127, 4563, 4972, 3390, 1153, 160, 8, 1 },
+  { 11512, 7158, 4558, 4934, 3329, 1116, 152, 8, 1 },
+  { 11629, 7188, 4553, 4896, 3269, 1080, 145, 7, 1 },
+  { 11746, 7217, 4547, 4857, 3210, 1045, 138, 7, 1 },
+  { 11864, 7245, 4540, 4818, 3151, 1012, 131, 6, 1 },
+  { 11980, 7273, 4533, 4779, 3093, 979, 124, 6, 1 },
+  { 12097, 7300, 4525, 4739, 3035, 947, 118, 6, 1 },
+  { 12215, 7326, 4516, 4698, 2978, 916, 113, 5, 1 },
+  { 12331, 7351, 4507, 4658, 2922, 886, 107, 5, 1 },
+  { 12448, 7375, 4497, 4617, 2866, 857, 102, 5, 1 },
+  { 12564, 7398, 4487, 4576, 2812, 829, 97, 4, 1 },
+  { 12681, 7421, 4476, 4534, 2757, 802, 92, 4, 1 },
+  { 12797, 7443, 4464, 4492, 2704, 775, 88, 4, 1 },
+  { 12914, 7464, 4452, 4450, 2651, 749, 84, 3, 1 },
+  { 13030, 7484, 4439, 4408, 2599, 725, 79, 3, 1 },
+  { 13147, 7503, 4426, 4365, 2547, 700, 76, 3, 1 },
+  { 13262, 7522, 4412, 4322, 2497, 677, 72, 3, 1 },
+  { 13378, 7539, 4398, 4280, 2447, 654, 68, 3, 1 },
+  { 13494, 7556, 4383, 4237, 2397, 632, 65, 3, 1 },
+  { 13610, 7573, 4368, 4193, 2348, 611, 62, 2, 1 },
+  { 13726, 7588, 4352, 4150, 2300, 590, 59, 2, 1 },
+  { 13841, 7602, 4335, 4107, 2253, 571, 56, 2, 1 },
+  { 13957, 7616, 4318, 4063, 2207, 551, 53, 2, 1 },
+  { 14072, 7629, 4301, 4019, 2161, 532, 51, 2, 1 },
+  { 14188, 7641, 4283, 3976, 2115, 514, 48, 2, 1 },
+  { 14302, 7652, 4265, 3932, 2071, 497, 46, 2, 1 },
+  { 14418, 7663, 4246, 3888, 2027, 480, 44, 1, 1 },
+  { 14533, 7673, 4227, 3844, 1984, 463, 42, 1, 1 },
+  { 14649, 7682, 4207, 3800, 1941, 447, 40, 1, 1 },
+  { 14763, 7690, 4187, 3757, 1899, 432, 38, 1, 1 },
+  { 14878, 7698, 4166, 3713, 1858, 417, 36, 1, 1 },
+  { 14993, 7705, 4146, 3669, 1817, 402, 34, 1, 1 },
+  { 15109, 7711, 4124, 3625, 1777, 388, 32, 1, 1 },
+  { 15223, 7715, 4103, 3581, 1738, 375, 31, 1, 1 },
+  { 15337, 7720, 4081, 3538, 1699, 362, 29, 1, 1 },
+  { 15452, 7724, 4058, 3494, 1661, 349, 28, 1, 1 },
+  { 15567, 7727, 4035, 3450, 1624, 337, 26, 1, 1 },
+  { 15681, 7729, 4012, 3407, 1587, 325, 25, 1, 1 },
+  { 15795, 7730, 3989, 3364, 1551, 313, 24, 1, 1 },
+  { 15909, 7731, 3965, 3320, 1516, 302, 23, 1, 1 },
+  { 16024, 7731, 3940, 3277, 1481, 291, 22, 1, 1 },
+  { 16138, 7730, 3916, 3234, 1446, 281, 21, 1, 1 },
+  { 16252, 7728, 3891, 3191, 1413, 271, 20, 1, 1 },
+  { 16366, 7726, 3866, 3148, 1380, 261, 19, 1, 1 },
+  { 16480, 7723, 3840, 3106, 1347, 252, 18, 1, 1 },
+  { 16594, 7720, 3814, 3063, 1315, 243, 17, 1, 1 },
+  { 16708, 7715, 3788, 3021, 1284, 234, 16, 1, 1 },
+  { 16822, 7710, 3762, 2979, 1253, 225, 15, 1, 1 },
+  { 16936, 7704, 3735, 2937, 1223, 217, 14, 1, 1 },
+  { 17050, 7697, 3708, 2895, 1193, 209, 14, 1, 1 },
+  { 17162, 7690, 3681, 2854, 1164, 202, 13, 1, 1 },
+  { 17276, 7682, 3654, 2812, 1136, 194, 12, 1, 1 },
+  { 17389, 7673, 3626, 2771, 1108, 187, 12, 1, 1 },
+  { 17504, 7663, 3598, 2730, 1080, 180, 11, 1, 1 },
+  { 17617, 7653, 3570, 2689, 1053, 173, 11, 1, 1 },
+  { 17730, 7642, 3541, 2649, 1027, 167, 10, 1, 1 },
+  { 17843, 7630, 3513, 2608, 1001, 161, 10, 1, 1 },
+  { 17957, 7618, 3484, 2569, 975, 154, 9, 1, 1 },
+  { 18069, 7605, 3455, 2529, 950, 149, 9, 1, 1 },
+  { 18183, 7591, 3426, 2489, 926, 143, 8, 1, 1 },
+  { 18296, 7576, 3396, 2450, 902, 138, 8, 1, 1 },
+  { 18410, 7562, 3366, 2411, 878, 132, 7, 1, 1 },
+  { 18523, 7545, 3337, 2372, 855, 127, 7, 1, 1 },
+  { 18636, 7529, 3306, 2333, 833, 122, 7, 1, 1 },
+  { 18749, 7511, 3276, 2295, 811, 118, 6, 1, 1 },
+  { 18862, 7493, 3246, 2257, 789, 113, 6, 1, 1 },
+  { 18975, 7474, 3215, 2219, 768, 109, 6, 1, 1 },
+  { 19088, 7455, 3185, 2182, 747, 104, 5, 1, 1 },
+  { 19201, 7435, 3154, 2144, 727, 100, 5, 1, 1 },
+  { 19314, 7414, 3123, 2107, 707, 96, 5, 1, 1 },
+  { 19427, 7392, 3092, 2071, 687, 92, 5, 1, 1 },
+  { 19541, 7370, 3060, 2034, 668, 89, 4, 1, 1 },
+  { 19654, 7347, 3029, 1998, 649, 85, 4, 1, 1 },
+  { 19766, 7323, 2997, 1963, 631, 82, 4, 1, 1 },
+  { 19878, 7299, 2966, 1927, 613, 79, 4, 1, 1 },
+  { 19991, 7274, 2934, 1892, 596, 75, 4, 1, 1 },
+  { 20105, 7248, 2902, 1857, 579, 72, 3, 1, 1 },
+  { 20218, 7222, 2870, 1822, 562, 69, 3, 1, 1 },
+  { 20331, 7195, 2838, 1788, 545, 66, 3, 1, 1 },
+  { 20443, 7167, 2806, 1754, 529, 64, 3, 1, 1 },
+  { 20556, 7138, 2774, 1720, 514, 61, 3, 1, 1 },
+  { 20670, 7109, 2741, 1687, 498, 58, 3, 1, 1 },
+  { 20783, 7079, 2709, 1654, 483, 56, 2, 1, 1 },
+  { 20895, 7049, 2676, 1621, 469, 54, 2, 1, 1 },
+  { 21008, 7017, 2644, 1589, 455, 51, 2, 1, 1 },
+  { 21121, 6985, 2611, 1557, 441, 49, 2, 1, 1 },
+  { 21234, 6953, 2578, 1525, 427, 47, 2, 1, 1 },
+  { 21347, 6919, 2545, 1494, 414, 45, 2, 1, 1 },
+  { 21460, 6885, 2513, 1462, 401, 43, 2, 1, 1 },
+  { 21573, 6850, 2480, 1432, 388, 41, 2, 1, 1 },
+  { 21687, 6815, 2447, 1401, 375, 39, 2, 1, 1 },
+  { 21801, 6778, 2414, 1371, 363, 38, 1, 1, 1 },
+  { 21914, 6741, 2381, 1341, 352, 36, 1, 1, 1 },
+  { 22028, 6704, 2348, 1311, 340, 34, 1, 1, 1 },
+  { 22141, 6665, 2315, 1282, 329, 33, 1, 1, 1 },
+  { 22255, 6626, 2282, 1253, 318, 31, 1, 1, 1 },
+  { 22368, 6586, 2249, 1225, 307, 30, 1, 1, 1 },
+  { 22482, 6546, 2216, 1196, 297, 28, 1, 1, 1 },
+  { 22595, 6505, 2183, 1169, 286, 27, 1, 1, 1 },
+  { 22709, 6463, 2149, 1141, 277, 26, 1, 1, 1 },
+  { 22823, 6420, 2116, 1114, 267, 25, 1, 1, 1 },
+  { 22938, 6377, 2083, 1087, 257, 23, 1, 1, 1 },
+  { 23053, 6332, 2050, 1060, 248, 22, 1, 1, 1 },
+  { 23167, 6287, 2017, 1034, 239, 21, 1, 1, 1 },
+  { 23280, 6242, 1984, 1008, 231, 20, 1, 1, 1 },
+  { 23396, 6195, 1951, 982, 222, 19, 1, 1, 1 },
+  { 23510, 6148, 1918, 957, 214, 18, 1, 1, 1 },
+  { 23625, 6100, 1885, 932, 206, 17, 1, 1, 1 },
+  { 23741, 6051, 1852, 907, 198, 16, 1, 1, 1 },
+  { 23855, 6002, 1819, 883, 190, 16, 1, 1, 1 },
+  { 23971, 5951, 1786, 859, 183, 15, 1, 1, 1 },
+  { 24087, 5900, 1753, 835, 176, 14, 1, 1, 1 },
+  { 24203, 5848, 1720, 812, 169, 13, 1, 1, 1 },
+  { 24318, 5796, 1687, 789, 162, 13, 1, 1, 1 },
+  { 24435, 5742, 1655, 766, 155, 12, 1, 1, 1 },
+  { 24552, 5688, 1622, 743, 149, 11, 1, 1, 1 },
+  { 24669, 5632, 1589, 721, 143, 11, 1, 1, 1 },
+  { 24786, 5576, 1557, 699, 137, 10, 1, 1, 1 },
+  { 24903, 5519, 1524, 678, 131, 10, 1, 1, 1 },
+  { 25021, 5462, 1491, 657, 125, 9, 1, 1, 1 },
+  { 25139, 5403, 1459, 636, 120, 8, 1, 1, 1 },
+  { 25258, 5343, 1427, 615, 114, 8, 1, 1, 1 },
+  { 25376, 5283, 1394, 595, 109, 8, 1, 1, 1 },
+  { 25496, 5221, 1362, 575, 104, 7, 1, 1, 1 },
+  { 25614, 5159, 1330, 556, 99, 7, 1, 1, 1 },
+  { 25735, 5096, 1298, 536, 94, 6, 1, 1, 1 },
+  { 25856, 5031, 1265, 517, 90, 6, 1, 1, 1 },
+  { 25977, 4966, 1233, 499, 85, 5, 1, 1, 1 },
+  { 26098, 4899, 1202, 480, 81, 5, 1, 1, 1 },
+  { 26220, 4831, 1170, 462, 77, 5, 1, 1, 1 },
+  { 26343, 4763, 1138, 444, 73, 4, 1, 1, 1 },
+  { 26466, 4693, 1106, 427, 69, 4, 1, 1, 1 },
+  { 26589, 4622, 1075, 410, 65, 4, 1, 1, 1 },
+  { 26713, 4550, 1043, 393, 62, 4, 1, 1, 1 },
+  { 26840, 4476, 1012, 376, 58, 3, 1, 1, 1 },
+  { 26966, 4401, 980, 360, 55, 3, 1, 1, 1 },
+  { 27092, 4325, 949, 344, 52, 3, 1, 1, 1 },
+  { 27220, 4248, 918, 328, 48, 3, 1, 1, 1 },
+  { 27350, 4169, 886, 313, 45, 2, 1, 1, 1 },
+  { 27480, 4088, 855, 298, 42, 2, 1, 1, 1 },
+  { 27610, 4006, 824, 283, 40, 2, 1, 1, 1 },
+  { 27743, 3922, 793, 268, 37, 2, 1, 1, 1 },
+  { 27876, 3837, 762, 254, 34, 2, 1, 1, 1 },
+  { 28011, 3749, 731, 240, 32, 2, 1, 1, 1 },
+  { 28147, 3659, 701, 227, 30, 1, 1, 1, 1 },
+  { 28286, 3568, 670, 213, 27, 1, 1, 1, 1 },
+  { 28426, 3474, 639, 200, 25, 1, 1, 1, 1 },
+  { 28569, 3377, 608, 187, 23, 1, 1, 1, 1 },
+  { 28714, 3278, 577, 174, 21, 1, 1, 1, 1 },
+  { 28860, 3176, 547, 162, 19, 1, 1, 1, 1 },
+  { 29010, 3071, 516, 150, 17, 1, 1, 1, 1 },
+  { 29163, 2962, 485, 138, 16, 1, 1, 1, 1 },
+  { 29320, 2849, 454, 127, 14, 1, 1, 1, 1 },
+  { 29483, 2731, 423, 115, 12, 1, 1, 1, 1 },
+  { 29650, 2608, 391, 104, 11, 1, 1, 1, 1 },
+  { 29823, 2479, 360, 93, 9, 1, 1, 1, 1 },
+  { 30002, 2343, 328, 83, 8, 1, 1, 1, 1 },
+  { 30192, 2198, 295, 72, 7, 1, 1, 1, 1 },
+  { 30393, 2041, 262, 62, 6, 1, 1, 1, 1 },
+  { 30612, 1869, 227, 52, 4, 1, 1, 1, 1 },
+  { 30853, 1676, 191, 41, 3, 1, 1, 1, 1 },
+  { 31131, 1448, 152, 31, 2, 1, 1, 1, 1 },
+  { 31486, 1150, 107, 20, 1, 1, 1, 1, 1 },
+};
+#elif CONFIG_EC_MULTISYMBOL
+// Values for tokens ONE_TOKEN through CATEGORY6_TOKEN included here.
+// ZERO_TOKEN and EOB_TOKEN are coded as flags outside this coder.
+const aom_cdf_prob
+    av1_pareto8_token_probs[COEFF_PROB_MODELS][ENTROPY_TOKENS - 2] = {
+      { 128, 127, 127, 126, 251, 495, 965, 1832, 3305, 25412 },
+      { 256, 254, 252, 249, 492, 959, 1820, 3283, 5365, 19838 },
+      { 384, 379, 374, 369, 724, 1392, 2574, 4417, 6568, 15587 },
+      { 512, 503, 494, 486, 946, 1795, 3238, 5289, 7184, 12321 },
+      { 640, 626, 612, 599, 1159, 2172, 3818, 5946, 7401, 9795 },
+      { 768, 748, 728, 709, 1363, 2522, 4324, 6424, 7352, 7830 },
+      { 896, 869, 842, 816, 1559, 2847, 4762, 6755, 7131, 6291 },
+      { 1024, 988, 954, 921, 1747, 3148, 5139, 6966, 6803, 5078 },
+      { 1152, 1107, 1063, 1022, 1926, 3427, 5460, 7080, 6412, 4119 },
+      { 1280, 1224, 1171, 1120, 2098, 3685, 5730, 7113, 5991, 3356 },
+      { 1408, 1340, 1276, 1216, 2261, 3923, 5955, 7083, 5560, 2746 },
+      { 1536, 1455, 1380, 1308, 2418, 4142, 6140, 7001, 5133, 2255 },
+      { 1664, 1569, 1481, 1398, 2567, 4342, 6287, 6879, 4721, 1860 },
+      { 1792, 1683, 1580, 1485, 2709, 4525, 6401, 6725, 4329, 1539 },
+      { 1920, 1794, 1678, 1570, 2845, 4692, 6486, 6546, 3959, 1278 },
+      { 2048, 1905, 1773, 1651, 2974, 4844, 6543, 6350, 3615, 1065 },
+      { 2176, 2015, 1867, 1731, 3096, 4980, 6576, 6140, 3296, 891 },
+      { 2304, 2123, 1958, 1807, 3212, 5104, 6589, 5922, 3002, 747 },
+      { 2432, 2231, 2048, 1882, 3322, 5214, 6581, 5698, 2732, 628 },
+      { 2560, 2337, 2136, 1953, 3427, 5311, 6557, 5472, 2485, 530 },
+      { 2688, 2442, 2222, 2023, 3525, 5397, 6518, 5246, 2259, 448 },
+      { 2816, 2547, 2306, 2090, 3618, 5472, 6465, 5021, 2053, 380 },
+      { 2944, 2650, 2388, 2154, 3706, 5537, 6401, 4799, 1866, 323 },
+      { 3072, 2752, 2468, 2217, 3788, 5591, 6327, 4581, 1696, 276 },
+      { 3200, 2853, 2547, 2277, 3866, 5637, 6243, 4369, 1541, 235 },
+      { 3328, 2952, 2624, 2335, 3938, 5673, 6152, 4163, 1401, 202 },
+      { 3456, 3051, 2699, 2391, 4006, 5702, 6054, 3962, 1274, 173 },
+      { 3584, 3149, 2772, 2444, 4070, 5723, 5950, 3769, 1158, 149 },
+      { 3712, 3246, 2843, 2496, 4128, 5736, 5842, 3583, 1054, 128 },
+      { 3840, 3341, 2913, 2545, 4183, 5743, 5729, 3404, 959, 111 },
+      { 3968, 3436, 2981, 2593, 4233, 5743, 5614, 3232, 872, 96 },
+      { 4096, 3529, 3048, 2638, 4280, 5737, 5496, 3067, 794, 83 },
+      { 4224, 3621, 3113, 2682, 4322, 5726, 5375, 2909, 724, 72 },
+      { 4352, 3712, 3176, 2724, 4361, 5709, 5253, 2759, 659, 63 },
+      { 4480, 3803, 3237, 2764, 4396, 5687, 5130, 2615, 601, 55 },
+      { 4608, 3892, 3297, 2801, 4428, 5661, 5007, 2478, 548, 48 },
+      { 4736, 3980, 3355, 2838, 4456, 5631, 4883, 2347, 500, 42 },
+      { 4864, 4067, 3412, 2872, 4481, 5596, 4760, 2223, 456, 37 },
+      { 4992, 4152, 3467, 2905, 4503, 5558, 4637, 2105, 417, 32 },
+      { 5120, 4237, 3521, 2936, 4521, 5516, 4515, 1993, 381, 28 },
+      { 5248, 4321, 3573, 2966, 4537, 5471, 4393, 1886, 348, 25 },
+      { 5376, 4404, 3623, 2993, 4550, 5424, 4273, 1785, 318, 22 },
+      { 5504, 4486, 3672, 3020, 4560, 5373, 4155, 1688, 291, 19 },
+      { 5632, 4566, 3720, 3044, 4568, 5321, 4037, 1597, 266, 17 },
+      { 5760, 4646, 3766, 3067, 4572, 5265, 3922, 1511, 244, 15 },
+      { 5888, 4724, 3811, 3089, 4575, 5208, 3808, 1429, 223, 13 },
+      { 6016, 4802, 3854, 3109, 4575, 5148, 3696, 1352, 204, 12 },
+      { 6144, 4878, 3895, 3128, 4573, 5088, 3587, 1278, 187, 10 },
+      { 6272, 4953, 3936, 3145, 4568, 5025, 3479, 1209, 172, 9 },
+      { 6400, 5028, 3975, 3161, 4561, 4961, 3373, 1143, 158, 8 },
+      { 6528, 5101, 4012, 3175, 4553, 4896, 3270, 1081, 145, 7 },
+      { 6656, 5173, 4048, 3189, 4542, 4830, 3168, 1022, 133, 7 },
+      { 6784, 5244, 4083, 3201, 4530, 4763, 3069, 966, 122, 6 },
+      { 6912, 5314, 4117, 3212, 4516, 4694, 2973, 913, 112, 5 },
+      { 7040, 5383, 4149, 3221, 4500, 4626, 2878, 863, 103, 5 },
+      { 7168, 5452, 4180, 3229, 4482, 4556, 2786, 816, 95, 4 },
+      { 7296, 5519, 4210, 3236, 4463, 4486, 2696, 771, 87, 4 },
+      { 7424, 5585, 4238, 3242, 4442, 4416, 2609, 729, 80, 3 },
+      { 7552, 5650, 4265, 3247, 4420, 4345, 2523, 689, 74, 3 },
+      { 7680, 5714, 4291, 3251, 4396, 4274, 2440, 651, 68, 3 },
+      { 7808, 5777, 4315, 3254, 4371, 4203, 2359, 616, 63, 2 },
+      { 7936, 5838, 4339, 3255, 4345, 4132, 2281, 582, 58, 2 },
+      { 8064, 5899, 4361, 3256, 4318, 4061, 2204, 550, 53, 2 },
+      { 8192, 5959, 4382, 3255, 4289, 3990, 2130, 520, 49, 2 },
+      { 8320, 6018, 4402, 3254, 4259, 3919, 2057, 492, 45, 2 },
+      { 8448, 6075, 4421, 3252, 4229, 3848, 1987, 465, 42, 1 },
+      { 8576, 6133, 4438, 3248, 4197, 3778, 1919, 439, 39, 1 },
+      { 8704, 6188, 4455, 3244, 4164, 3708, 1853, 415, 36, 1 },
+      { 8832, 6243, 4470, 3239, 4131, 3638, 1789, 392, 33, 1 },
+      { 8960, 6297, 4484, 3233, 4096, 3569, 1727, 371, 30, 1 },
+      { 9088, 6349, 4497, 3226, 4061, 3500, 1667, 351, 28, 1 },
+      { 9216, 6401, 4509, 3219, 4025, 3432, 1608, 331, 26, 1 },
+      { 9344, 6452, 4520, 3210, 3989, 3364, 1551, 313, 24, 1 },
+      { 9472, 6501, 4530, 3201, 3952, 3297, 1496, 296, 22, 1 },
+      { 9600, 6550, 4539, 3191, 3914, 3230, 1443, 280, 20, 1 },
+      { 9728, 6597, 4547, 3180, 3875, 3164, 1392, 265, 19, 1 },
+      { 9856, 6644, 4554, 3169, 3836, 3098, 1342, 250, 18, 1 },
+      { 9984, 6690, 4560, 3157, 3796, 3034, 1293, 237, 16, 1 },
+      { 10112, 6734, 4565, 3144, 3756, 2970, 1247, 224, 15, 1 },
+      { 10240, 6778, 4568, 3131, 3716, 2907, 1202, 211, 14, 1 },
+      { 10368, 6821, 4571, 3117, 3675, 2844, 1158, 200, 13, 1 },
+      { 10496, 6862, 4573, 3102, 3634, 2783, 1116, 189, 12, 1 },
+      { 10624, 6903, 4574, 3087, 3592, 2722, 1075, 179, 11, 1 },
+      { 10752, 6942, 4575, 3071, 3551, 2662, 1035, 169, 10, 1 },
+      { 10880, 6981, 4574, 3054, 3508, 2603, 997, 160, 10, 1 },
+      { 11008, 7019, 4572, 3038, 3466, 2544, 960, 151, 9, 1 },
+      { 11136, 7055, 4570, 3020, 3424, 2487, 924, 143, 8, 1 },
+      { 11264, 7091, 4566, 3002, 3381, 2430, 890, 135, 8, 1 },
+      { 11392, 7126, 4563, 2984, 3338, 2374, 856, 127, 7, 1 },
+      { 11520, 7159, 4557, 2965, 3295, 2319, 824, 121, 7, 1 },
+      { 11648, 7193, 4552, 2945, 3252, 2264, 793, 114, 6, 1 },
+      { 11776, 7224, 4545, 2925, 3209, 2211, 763, 108, 6, 1 },
+      { 11904, 7255, 4538, 2905, 3165, 2159, 734, 102, 5, 1 },
+      { 12032, 7285, 4530, 2884, 3122, 2107, 706, 96, 5, 1 },
+      { 12160, 7314, 4520, 2863, 3079, 2056, 679, 91, 5, 1 },
+      { 12288, 7341, 4511, 2842, 3036, 2006, 653, 86, 4, 1 },
+      { 12416, 7368, 4500, 2820, 2993, 1957, 628, 81, 4, 1 },
+      { 12544, 7394, 4489, 2797, 2949, 1909, 604, 77, 4, 1 },
+      { 12672, 7419, 4477, 2775, 2906, 1861, 581, 73, 3, 1 },
+      { 12800, 7443, 4464, 2752, 2863, 1815, 558, 69, 3, 1 },
+      { 12928, 7466, 4451, 2729, 2820, 1769, 536, 65, 3, 1 },
+      { 13056, 7488, 4437, 2705, 2777, 1724, 516, 61, 3, 1 },
+      { 13184, 7509, 4422, 2682, 2734, 1680, 495, 58, 3, 1 },
+      { 13312, 7529, 4406, 2658, 2692, 1637, 476, 55, 2, 1 },
+      { 13440, 7548, 4390, 2633, 2650, 1595, 457, 52, 2, 1 },
+      { 13568, 7567, 4373, 2609, 2607, 1553, 439, 49, 2, 1 },
+      { 13696, 7583, 4356, 2584, 2565, 1513, 422, 46, 2, 1 },
+      { 13824, 7600, 4337, 2559, 2523, 1473, 405, 44, 2, 1 },
+      { 13952, 7615, 4319, 2533, 2482, 1434, 389, 41, 2, 1 },
+      { 14080, 7629, 4300, 2508, 2441, 1395, 373, 39, 2, 1 },
+      { 14208, 7643, 4280, 2482, 2400, 1358, 358, 37, 1, 1 },
+      { 14336, 7655, 4259, 2457, 2359, 1321, 344, 35, 1, 1 },
+      { 14464, 7667, 4238, 2431, 2318, 1285, 330, 33, 1, 1 },
+      { 14592, 7677, 4217, 2405, 2278, 1250, 316, 31, 1, 1 },
+      { 14720, 7687, 4195, 2378, 2238, 1215, 304, 29, 1, 1 },
+      { 14848, 7696, 4172, 2352, 2198, 1181, 291, 28, 1, 1 },
+      { 14976, 7703, 4149, 2326, 2159, 1148, 279, 26, 1, 1 },
+      { 15104, 7710, 4125, 2299, 2119, 1116, 268, 25, 1, 1 },
+      { 15232, 7715, 4101, 2272, 2081, 1085, 257, 23, 1, 1 },
+      { 15360, 7721, 4076, 2245, 2042, 1054, 246, 22, 1, 1 },
+      { 15488, 7724, 4051, 2219, 2004, 1023, 236, 21, 1, 1 },
+      { 15616, 7727, 4025, 2192, 1966, 994, 226, 20, 1, 1 },
+      { 15744, 7729, 3999, 2164, 1929, 965, 217, 19, 1, 1 },
+      { 15872, 7731, 3972, 2137, 1892, 937, 207, 18, 1, 1 },
+      { 16000, 7731, 3945, 2110, 1855, 909, 199, 17, 1, 1 },
+      { 16128, 7730, 3918, 2083, 1819, 882, 190, 16, 1, 1 },
+      { 16256, 7728, 3890, 2056, 1783, 856, 182, 15, 1, 1 },
+      { 16384, 7725, 3862, 2029, 1747, 831, 174, 14, 1, 1 },
+      { 16512, 7721, 3833, 2002, 1712, 806, 167, 13, 1, 1 },
+      { 16640, 7717, 3804, 1975, 1677, 781, 160, 12, 1, 1 },
+      { 16768, 7712, 3775, 1947, 1642, 757, 153, 12, 1, 1 },
+      { 16896, 7706, 3745, 1920, 1608, 734, 146, 11, 1, 1 },
+      { 17024, 7699, 3714, 1893, 1575, 711, 140, 10, 1, 1 },
+      { 17152, 7690, 3684, 1866, 1541, 689, 134, 10, 1, 1 },
+      { 17280, 7681, 3653, 1839, 1508, 668, 128, 9, 1, 1 },
+      { 17408, 7671, 3621, 1812, 1476, 647, 122, 9, 1, 1 },
+      { 17536, 7660, 3590, 1785, 1444, 626, 117, 8, 1, 1 },
+      { 17664, 7648, 3558, 1758, 1412, 606, 112, 8, 1, 1 },
+      { 17792, 7635, 3526, 1731, 1381, 587, 107, 7, 1, 1 },
+      { 17920, 7622, 3493, 1704, 1350, 568, 102, 7, 1, 1 },
+      { 18048, 7607, 3461, 1678, 1319, 549, 98, 6, 1, 1 },
+      { 18176, 7592, 3428, 1651, 1289, 531, 93, 6, 1, 1 },
+      { 18304, 7575, 3394, 1625, 1259, 514, 89, 6, 1, 1 },
+      { 18432, 7558, 3361, 1598, 1230, 497, 85, 5, 1, 1 },
+      { 18560, 7540, 3327, 1572, 1201, 480, 81, 5, 1, 1 },
+      { 18688, 7520, 3293, 1546, 1173, 464, 77, 5, 1, 1 },
+      { 18816, 7500, 3258, 1520, 1145, 448, 74, 5, 1, 1 },
+      { 18944, 7480, 3224, 1494, 1117, 433, 70, 4, 1, 1 },
+      { 19072, 7458, 3189, 1468, 1090, 418, 67, 4, 1, 1 },
+      { 19200, 7435, 3154, 1442, 1063, 404, 64, 4, 1, 1 },
+      { 19328, 7410, 3119, 1417, 1037, 390, 61, 4, 1, 1 },
+      { 19456, 7386, 3084, 1392, 1011, 376, 58, 3, 1, 1 },
+      { 19584, 7361, 3048, 1366, 986, 363, 55, 3, 1, 1 },
+      { 19712, 7335, 3012, 1341, 960, 350, 53, 3, 1, 1 },
+      { 19840, 7307, 2977, 1316, 936, 337, 50, 3, 1, 1 },
+      { 19968, 7279, 2941, 1291, 911, 325, 48, 3, 1, 1 },
+      { 20096, 7251, 2905, 1267, 887, 313, 45, 2, 1, 1 },
+      { 20224, 7220, 2868, 1243, 864, 302, 43, 2, 1, 1 },
+      { 20352, 7189, 2832, 1218, 841, 291, 41, 2, 1, 1 },
+      { 20480, 7158, 2795, 1194, 818, 280, 39, 2, 1, 1 },
+      { 20608, 7124, 2759, 1170, 796, 270, 37, 2, 1, 1 },
+      { 20736, 7091, 2722, 1147, 774, 259, 35, 2, 1, 1 },
+      { 20864, 7056, 2685, 1123, 752, 250, 34, 2, 1, 1 },
+      { 20992, 7021, 2648, 1100, 731, 240, 32, 2, 1, 1 },
+      { 21120, 6985, 2612, 1077, 710, 231, 30, 1, 1, 1 },
+      { 21248, 6948, 2574, 1054, 690, 222, 29, 1, 1, 1 },
+      { 21376, 6911, 2537, 1031, 670, 213, 27, 1, 1, 1 },
+      { 21504, 6872, 2500, 1008, 650, 205, 26, 1, 1, 1 },
+      { 21632, 6831, 2463, 986, 631, 197, 25, 1, 1, 1 },
+      { 21760, 6791, 2426, 964, 612, 189, 23, 1, 1, 1 },
+      { 21888, 6749, 2389, 942, 594, 181, 22, 1, 1, 1 },
+      { 22016, 6707, 2351, 921, 575, 174, 21, 1, 1, 1 },
+      { 22144, 6663, 2314, 899, 558, 167, 20, 1, 1, 1 },
+      { 22272, 6619, 2277, 878, 540, 160, 19, 1, 1, 1 },
+      { 22400, 6574, 2240, 857, 523, 153, 18, 1, 1, 1 },
+      { 22528, 6529, 2202, 836, 507, 146, 17, 1, 1, 1 },
+      { 22656, 6482, 2165, 816, 490, 140, 16, 1, 1, 1 },
+      { 22784, 6435, 2128, 795, 474, 134, 15, 1, 1, 1 },
+      { 22912, 6386, 2091, 775, 459, 128, 14, 1, 1, 1 },
+      { 23040, 6336, 2054, 756, 443, 123, 13, 1, 1, 1 },
+      { 23168, 6286, 2017, 736, 428, 117, 13, 1, 1, 1 },
+      { 23296, 6234, 1980, 717, 414, 112, 12, 1, 1, 1 },
+      { 23424, 6183, 1943, 698, 399, 107, 11, 1, 1, 1 },
+      { 23552, 6130, 1906, 679, 385, 102, 11, 1, 1, 1 },
+      { 23680, 6077, 1869, 660, 372, 97, 10, 1, 1, 1 },
+      { 23808, 6022, 1833, 642, 358, 93, 9, 1, 1, 1 },
+      { 23936, 5966, 1796, 624, 345, 89, 9, 1, 1, 1 },
+      { 24064, 5910, 1760, 606, 333, 84, 8, 1, 1, 1 },
+      { 24192, 5853, 1724, 588, 320, 80, 8, 1, 1, 1 },
+      { 24320, 5796, 1687, 571, 308, 76, 7, 1, 1, 1 },
+      { 24448, 5735, 1651, 554, 297, 73, 7, 1, 1, 1 },
+      { 24576, 5677, 1615, 537, 285, 69, 6, 1, 1, 1 },
+      { 24704, 5615, 1579, 521, 274, 66, 6, 1, 1, 1 },
+      { 24832, 5554, 1544, 504, 263, 62, 6, 1, 1, 1 },
+      { 24960, 5492, 1508, 488, 253, 59, 5, 1, 1, 1 },
+      { 25088, 5428, 1473, 473, 242, 56, 5, 1, 1, 1 },
+      { 25216, 5364, 1438, 457, 232, 53, 5, 1, 1, 1 },
+      { 25344, 5300, 1403, 442, 222, 50, 4, 1, 1, 1 },
+      { 25472, 5233, 1368, 427, 213, 48, 4, 1, 1, 1 },
+      { 25600, 5166, 1334, 412, 204, 45, 4, 1, 1, 1 },
+      { 25728, 5098, 1299, 398, 195, 43, 4, 1, 1, 1 },
+      { 25856, 5030, 1266, 384, 186, 40, 3, 1, 1, 1 },
+      { 25984, 4960, 1232, 370, 178, 38, 3, 1, 1, 1 },
+      { 26112, 4890, 1198, 356, 170, 36, 3, 1, 1, 1 },
+      { 26240, 4819, 1164, 343, 162, 34, 3, 1, 1, 1 },
+      { 26368, 4748, 1132, 329, 154, 32, 2, 1, 1, 1 },
+      { 26496, 4675, 1098, 317, 147, 30, 2, 1, 1, 1 },
+      { 26624, 4602, 1066, 304, 139, 28, 2, 1, 1, 1 },
+      { 26752, 4527, 1034, 292, 132, 26, 2, 1, 1, 1 },
+      { 26880, 4451, 1001, 280, 126, 25, 2, 1, 1, 1 },
+      { 27008, 4375, 970, 268, 119, 23, 2, 1, 1, 1 },
+      { 27136, 4299, 938, 256, 113, 21, 2, 1, 1, 1 },
+      { 27264, 4221, 907, 245, 107, 20, 1, 1, 1, 1 },
+      { 27392, 4142, 876, 234, 101, 19, 1, 1, 1, 1 },
+      { 27520, 4063, 846, 223, 95, 17, 1, 1, 1, 1 },
+      { 27648, 3982, 815, 213, 90, 16, 1, 1, 1, 1 },
+      { 27776, 3900, 786, 202, 85, 15, 1, 1, 1, 1 },
+      { 27904, 3818, 756, 192, 80, 14, 1, 1, 1, 1 },
+      { 28032, 3734, 727, 183, 75, 13, 1, 1, 1, 1 },
+      { 28160, 3651, 698, 173, 70, 12, 1, 1, 1, 1 },
+      { 28288, 3566, 669, 164, 66, 11, 1, 1, 1, 1 },
+      { 28416, 3481, 641, 155, 61, 10, 1, 1, 1, 1 },
+      { 28544, 3393, 614, 147, 57, 9, 1, 1, 1, 1 },
+      { 28672, 3306, 586, 138, 53, 9, 1, 1, 1, 1 },
+      { 28800, 3217, 559, 130, 50, 8, 1, 1, 1, 1 },
+      { 28928, 3128, 533, 122, 46, 7, 1, 1, 1, 1 },
+      { 29056, 3037, 507, 114, 43, 7, 1, 1, 1, 1 },
+      { 29184, 2947, 481, 107, 39, 6, 1, 1, 1, 1 },
+      { 29312, 2855, 456, 100, 36, 5, 1, 1, 1, 1 },
+      { 29440, 2762, 431, 93, 33, 5, 1, 1, 1, 1 },
+      { 29568, 2668, 407, 86, 31, 4, 1, 1, 1, 1 },
+      { 29696, 2573, 383, 80, 28, 4, 1, 1, 1, 1 },
+      { 29824, 2478, 359, 74, 25, 4, 1, 1, 1, 1 },
+      { 29952, 2381, 337, 68, 23, 3, 1, 1, 1, 1 },
+      { 30080, 2284, 314, 62, 21, 3, 1, 1, 1, 1 },
+      { 30208, 2185, 293, 57, 19, 2, 1, 1, 1, 1 },
+      { 30336, 2086, 271, 52, 17, 2, 1, 1, 1, 1 },
+      { 30464, 1986, 250, 47, 15, 2, 1, 1, 1, 1 },
+      { 30592, 1885, 230, 42, 13, 2, 1, 1, 1, 1 },
+      { 30720, 1782, 211, 38, 12, 1, 1, 1, 1, 1 },
+      { 30848, 1679, 192, 34, 10, 1, 1, 1, 1, 1 },
+      { 30976, 1575, 173, 30, 9, 1, 1, 1, 1, 1 },
+      { 31104, 1469, 156, 26, 8, 1, 1, 1, 1, 1 },
+      { 31232, 1364, 138, 23, 6, 1, 1, 1, 1, 1 },
+      { 31360, 1257, 122, 19, 5, 1, 1, 1, 1, 1 },
+      { 31488, 1149, 106, 16, 4, 1, 1, 1, 1, 1 },
+      { 31616, 1038, 91, 14, 4, 1, 1, 1, 1, 1 },
+      { 31744, 928, 77, 11, 3, 1, 1, 1, 1, 1 },
+      { 31872, 816, 64, 9, 2, 1, 1, 1, 1, 1 },
+      { 32000, 703, 51, 7, 2, 1, 1, 1, 1, 1 },
+      { 32128, 589, 40, 5, 1, 1, 1, 1, 1, 1 },
+      { 32256, 473, 29, 4, 1, 1, 1, 1, 1, 1 },
+      { 32384, 357, 19, 2, 1, 1, 1, 1, 1, 1 },
+      { 32512, 238, 11, 1, 1, 1, 1, 1, 1, 1 },
+      { 32640, 117, 4, 1, 1, 1, 1, 1, 1, 1 },
+    };
+#endif  // CONFIG_NEW_TOKENSET
+
+/* clang-format off */
+#if CONFIG_Q_ADAPT_PROBS
+const av1_coeff_probs_model
+default_qctx_coef_probs[QCTX_BINS][TX_SIZES][PLANE_TYPES] = {
+    {  // Q_Index 0
+#if CONFIG_CB4X4
+        {  // TX_SIZE 0
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        {182,  34, 137}, { 79,  39, 103}, { 10,  28,  51},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 45,  88, 147}, { 46,  80, 140}, { 25,  69, 119},
+                        { 12,  57,  96}, {  4,  41,  65}, {  1,  20,  31},
+                    },
+                    {  // band 2
+                        { 58, 124, 190}, { 39, 106, 178}, { 16,  86, 147},
+                        {  7,  69, 114}, {  3,  50,  80}, {  1,  25,  42},
+                    },
+                    {  // band 3
+                        { 90, 138, 215}, { 54, 116, 198}, { 18,  86, 155},
+                        {  5,  62, 112}, {  1,  38,  68}, {  1,  17,  30},
+                    },
+                    {  // band 4
+                        {126, 149, 231}, { 82, 114, 211}, { 21,  80, 157},
+                        {  6,  56, 105}, {  1,  36,  64}, {  1,  17,  31},
+                    },
+                    {  // band 5
+                        {171,  56, 236}, {140,  54, 219}, { 57,  45, 167},
+                        { 26,  36, 113}, { 11,  29,  72}, {  3,  18,  39},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        {153, 122, 186}, {106, 109, 171}, { 36,  84, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 27, 151, 201}, { 34, 131, 199}, { 23, 102, 161},
+                        { 10,  80, 120}, {  4,  52,  78}, {  1,  24,  37},
+                    },
+                    {  // band 2
+                        { 43, 158, 213}, { 35, 133, 203}, {  8,  92, 151},
+                        {  2,  64, 106}, {  1,  36,  60}, {  1,  13,  24},
+                    },
+                    {  // band 3
+                        { 68, 167, 223}, { 36, 135, 211}, {  9,  94, 157},
+                        {  2,  67, 112}, {  1,  40,  68}, {  1,  17,  31},
+                    },
+                    {  // band 4
+                        {131, 146, 237}, { 72, 119, 223}, { 17,  82, 164},
+                        {  4,  55, 107}, {  1,  34,  63}, {  1,  16,  29},
+                    },
+                    {  // band 5
+                        {184,  68, 244}, {153,  59, 232}, { 68,  51, 179},
+                        { 31,  40, 123}, { 13,  29,  77}, {  4,  17,  37},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {203,  41, 203}, {127,  56, 174}, { 49,  56, 127},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {110, 121, 217}, {119, 113, 213}, { 64,  95, 185},
+                        { 30,  72, 144}, {  8,  42,  76}, {  2,  17,  25},
+                    },
+                    {  // band 2
+                        {127, 159, 229}, {115, 134, 223}, { 36, 100, 189},
+                        { 11,  75, 142}, {  3,  48,  83}, {  1,  19,  33},
+                    },
+                    {  // band 3
+                        {150, 172, 241}, { 90, 133, 231}, { 28, 102, 192},
+                        {  7,  81, 147}, {  1,  53,  91}, {  1,  25,  42},
+                    },
+                    {  // band 4
+                        {184, 144, 248}, {114, 117, 237}, { 37,  89, 192},
+                        { 10,  63, 130}, {  4,  42,  76}, {  1,  19,  38},
+                    },
+                    {  // band 5
+                        {207,  79, 250}, {179,  74, 241}, { 83,  67, 199},
+                        { 38,  51, 142}, { 17,  37,  97}, { 10,  14,  55},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {220,  82, 232}, {150,  93, 214}, { 66,  95, 177},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {116, 160, 227}, {136, 141, 227}, { 67, 114, 190},
+                        { 40,  94, 148}, { 21,  70, 107}, { 10,  43,  63},
+                    },
+                    {  // band 2
+                        {124, 173, 235}, {105, 147, 226}, { 27, 107, 184},
+                        { 10,  80, 142}, {  3,  50,  86}, {  1,  16,  32},
+                    },
+                    {  // band 3
+                        {149, 179, 243}, { 89, 147, 234}, { 29, 112, 193},
+                        {  9,  94, 157}, {  1,  64, 111}, {  1,  25,  43},
+                    },
+                    {  // band 4
+                        {187, 153, 248}, {127, 130, 241}, { 52,  99, 202},
+                        { 20,  79, 152}, {  4,  50,  93}, {  1,  19,  32},
+                    },
+                    {  // band 5
+                        {215,  82, 251}, {195,  80, 246}, { 93,  70, 204},
+                        { 39,  54, 147}, { 14,  33,  88}, {  6,  14,  39},
+                    },
+                },
+            },
+        },
+#endif
+        {  // TX_SIZE 0
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        {182,  34, 137}, { 79,  39, 103}, { 10,  28,  51},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 45,  88, 147}, { 46,  80, 140}, { 25,  69, 119},
+                        { 12,  57,  96}, {  4,  41,  65}, {  1,  20,  31},
+                    },
+                    {  // band 2
+                        { 58, 124, 190}, { 39, 106, 178}, { 16,  86, 147},
+                        {  7,  69, 114}, {  3,  50,  80}, {  1,  25,  42},
+                    },
+                    {  // band 3
+                        { 90, 138, 215}, { 54, 116, 198}, { 18,  86, 155},
+                        {  5,  62, 112}, {  1,  38,  68}, {  1,  17,  30},
+                    },
+                    {  // band 4
+                        {126, 149, 231}, { 82, 114, 211}, { 21,  80, 157},
+                        {  6,  56, 105}, {  1,  36,  64}, {  1,  17,  31},
+                    },
+                    {  // band 5
+                        {171,  56, 236}, {140,  54, 219}, { 57,  45, 167},
+                        { 26,  36, 113}, { 11,  29,  72}, {  3,  18,  39},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        {153, 122, 186}, {106, 109, 171}, { 36,  84, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 27, 151, 201}, { 34, 131, 199}, { 23, 102, 161},
+                        { 10,  80, 120}, {  4,  52,  78}, {  1,  24,  37},
+                    },
+                    {  // band 2
+                        { 43, 158, 213}, { 35, 133, 203}, {  8,  92, 151},
+                        {  2,  64, 106}, {  1,  36,  60}, {  1,  13,  24},
+                    },
+                    {  // band 3
+                        { 68, 167, 223}, { 36, 135, 211}, {  9,  94, 157},
+                        {  2,  67, 112}, {  1,  40,  68}, {  1,  17,  31},
+                    },
+                    {  // band 4
+                        {131, 146, 237}, { 72, 119, 223}, { 17,  82, 164},
+                        {  4,  55, 107}, {  1,  34,  63}, {  1,  16,  29},
+                    },
+                    {  // band 5
+                        {184,  68, 244}, {153,  59, 232}, { 68,  51, 179},
+                        { 31,  40, 123}, { 13,  29,  77}, {  4,  17,  37},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {203,  41, 203}, {127,  56, 174}, { 49,  56, 127},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {110, 121, 217}, {119, 113, 213}, { 64,  95, 185},
+                        { 30,  72, 144}, {  8,  42,  76}, {  2,  17,  25},
+                    },
+                    {  // band 2
+                        {127, 159, 229}, {115, 134, 223}, { 36, 100, 189},
+                        { 11,  75, 142}, {  3,  48,  83}, {  1,  19,  33},
+                    },
+                    {  // band 3
+                        {150, 172, 241}, { 90, 133, 231}, { 28, 102, 192},
+                        {  7,  81, 147}, {  1,  53,  91}, {  1,  25,  42},
+                    },
+                    {  // band 4
+                        {184, 144, 248}, {114, 117, 237}, { 37,  89, 192},
+                        { 10,  63, 130}, {  4,  42,  76}, {  1,  19,  38},
+                    },
+                    {  // band 5
+                        {207,  79, 250}, {179,  74, 241}, { 83,  67, 199},
+                        { 38,  51, 142}, { 17,  37,  97}, { 10,  14,  55},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {220,  82, 232}, {150,  93, 214}, { 66,  95, 177},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {116, 160, 227}, {136, 141, 227}, { 67, 114, 190},
+                        { 40,  94, 148}, { 21,  70, 107}, { 10,  43,  63},
+                    },
+                    {  // band 2
+                        {124, 173, 235}, {105, 147, 226}, { 27, 107, 184},
+                        { 10,  80, 142}, {  3,  50,  86}, {  1,  16,  32},
+                    },
+                    {  // band 3
+                        {149, 179, 243}, { 89, 147, 234}, { 29, 112, 193},
+                        {  9,  94, 157}, {  1,  64, 111}, {  1,  25,  43},
+                    },
+                    {  // band 4
+                        {187, 153, 248}, {127, 130, 241}, { 52,  99, 202},
+                        { 20,  79, 152}, {  4,  50,  93}, {  1,  19,  32},
+                    },
+                    {  // band 5
+                        {215,  82, 251}, {195,  80, 246}, { 93,  70, 204},
+                        { 39,  54, 147}, { 14,  33,  88}, {  6,  14,  39},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 1
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        {116,  43, 131}, { 39,  41,  94}, {  4,  28,  47},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 28, 101, 141}, { 27,  95, 140}, { 18,  80, 121},
+                        { 10,  61,  95}, {  4,  39,  60}, {  1,  19,  26},
+                    },
+                    {  // band 2
+                        { 29, 150, 183}, { 19, 127, 175}, {  8,  98, 147},
+                        {  3,  76, 115}, {  1,  55,  84}, {  1,  29,  43},
+                    },
+                    {  // band 3
+                        { 26, 168, 202}, { 12, 138, 188}, {  2,  98, 149},
+                        {  1,  69, 110}, {  1,  40,  65}, {  1,  17,  25},
+                    },
+                    {  // band 4
+                        { 33, 188, 225}, { 12, 155, 207}, {  2, 101, 155},
+                        {  1,  65, 106}, {  1,  36,  60}, {  1,  18,  26},
+                    },
+                    {  // band 5
+                        { 79, 205, 242}, { 30, 168, 224}, {  5, 106, 164},
+                        {  1,  68, 110}, {  1,  39,  65}, {  1,  18,  28},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 96,  80, 201}, { 51,  88, 168}, { 14,  78, 116},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {  6, 167, 216}, { 32, 152, 211}, { 24, 121, 182},
+                        { 13,  98, 149}, { 12,  76, 108}, {  8,  48,  62},
+                    },
+                    {  // band 2
+                        { 17, 176, 225}, { 13, 147, 209}, {  3,  96, 155},
+                        {  1,  65, 108}, {  2,  43,  63}, {  2,  23,  25},
+                    },
+                    {  // band 3
+                        { 18, 183, 232}, { 10, 153, 214}, {  1,  96, 154},
+                        {  1,  63, 105}, {  1,  39,  59}, {  1,  21,  24},
+                    },
+                    {  // band 4
+                        { 23, 191, 239}, {  8, 159, 221}, {  1,  97, 158},
+                        {  1,  61, 105}, {  1,  37,  60}, {  1,  20,  26},
+                    },
+                    {  // band 5
+                        { 70, 201, 243}, { 29, 163, 228}, {  4, 102, 169},
+                        {  1,  67, 114}, {  1,  39,  66}, {  1,  17,  29},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {181,  38, 192}, { 95,  47, 151}, { 29,  49, 102},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 72, 131, 202}, { 93, 120, 205}, { 50, 103, 179},
+                        { 24,  79, 143}, { 11,  47,  78}, {  7,  19,  25},
+                    },
+                    {  // band 2
+                        { 84, 176, 221}, { 56, 144, 214}, { 21, 108, 182},
+                        {  8,  83, 139}, {  3,  55,  90}, {  2,  27,  41},
+                    },
+                    {  // band 3
+                        { 84, 195, 234}, { 42, 156, 222}, { 10, 109, 180},
+                        {  4,  77, 133}, {  1,  48,  80}, {  1,  23,  35},
+                    },
+                    {  // band 4
+                        { 89, 210, 238}, { 35, 165, 221}, {  6, 106, 172},
+                        {  2,  70, 123}, {  1,  44,  74}, {  1,  21,  30},
+                    },
+                    {  // band 5
+                        {114, 221, 247}, { 49, 170, 234}, {  7, 113, 184},
+                        {  2,  77, 132}, {  1,  48,  79}, {  1,  25,  33},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {192,  66, 237}, {113,  84, 211}, { 35,  84, 154},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 81, 180, 234}, {127, 165, 229}, { 58, 137, 204},
+                        { 41, 114, 174}, { 44,  94, 136}, { 29,  66,  86},
+                    },
+                    {  // band 2
+                        { 82, 193, 240}, { 39, 162, 223}, {  8, 113, 179},
+                        {  3,  83, 136}, {  6,  62,  84}, {  5,  45,  45},
+                    },
+                    {  // band 3
+                        { 78, 203, 242}, { 31, 170, 227}, {  4, 115, 181},
+                        {  1,  82, 135}, {  2,  59,  82}, {  1,  45,  47},
+                    },
+                    {  // band 4
+                        { 76, 210, 239}, { 25, 170, 213}, {  2,  99, 152},
+                        {  1,  69, 115}, {  1,  49,  80}, {  1,  47,  57},
+                    },
+                    {  // band 5
+                        {103, 217, 250}, { 42, 180, 237}, {  3, 124, 191},
+                        {  1,  90, 150}, {  1,  69, 116}, {  1,  52,  46},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 2
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 58,  38,  99}, {  9,  26,  51}, {  1,  14,  22},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 14,  78, 109}, { 16,  73, 105}, { 11,  62,  92},
+                        {  6,  47,  72}, {  2,  29,  45}, {  1,  12,  18},
+                    },
+                    {  // band 2
+                        { 17, 131, 148}, { 11, 112, 140}, {  5,  87, 118},
+                        {  2,  63,  90}, {  1,  42,  63}, {  1,  19,  31},
+                    },
+                    {  // band 3
+                        { 12, 151, 168}, {  6, 116, 152}, {  1,  76, 115},
+                        {  1,  50,  81}, {  1,  32,  52}, {  1,  14,  23},
+                    },
+                    {  // band 4
+                        { 10, 174, 191}, {  3, 130, 172}, {  1,  80, 126},
+                        {  1,  53,  88}, {  1,  32,  55}, {  1,  14,  24},
+                    },
+                    {  // band 5
+                        { 19, 219, 237}, {  3, 168, 211}, {  1,  90, 142},
+                        {  1,  53,  91}, {  1,  29,  51}, {  1,  12,  21},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 21,  46, 184}, { 10,  53, 130}, {  2,  49,  78},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {  3, 169, 198}, { 37, 165, 196}, { 26, 134, 176},
+                        { 11, 108, 149}, {  5,  81, 112}, {  3,  47,  64},
+                    },
+                    {  // band 2
+                        { 11, 183, 215}, {  8, 142, 192}, {  2,  91, 141},
+                        {  1,  62, 100}, {  1,  38,  62}, {  1,  17,  28},
+                    },
+                    {  // band 3
+                        { 12, 190, 223}, {  6, 149, 199}, {  1,  88, 139},
+                        {  1,  56,  93}, {  1,  31,  54}, {  1,  13,  21},
+                    },
+                    {  // band 4
+                        { 11, 197, 230}, {  3, 154, 204}, {  1,  83, 134},
+                        {  1,  50,  86}, {  1,  28,  49}, {  1,  12,  21},
+                    },
+                    {  // band 5
+                        { 17, 211, 240}, {  2, 167, 217}, {  1,  88, 143},
+                        {  1,  53,  91}, {  1,  30,  53}, {  1,  14,  24},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {151,  30, 151}, { 50,  36, 105}, {  8,  34,  66},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 39, 111, 160}, { 62, 111, 165}, { 37,  99, 147},
+                        { 15,  77, 118}, {  3,  47,  73}, {  1,  17,  27},
+                    },
+                    {  // band 2
+                        { 48, 170, 190}, { 32, 135, 180}, { 11, 100, 149},
+                        {  4,  76, 116}, {  1,  51,  80}, {  1,  22,  36},
+                    },
+                    {  // band 3
+                        { 39, 191, 208}, { 18, 141, 191}, {  3,  96, 150},
+                        {  1,  66, 110}, {  1,  41,  69}, {  1,  17,  28},
+                    },
+                    {  // band 4
+                        { 32, 209, 219}, {  8, 152, 201}, {  1,  96, 153},
+                        {  1,  63, 106}, {  1,  38,  66}, {  1,  17,  29},
+                    },
+                    {  // band 5
+                        { 33, 230, 237}, {  5, 173, 214}, {  1, 100, 155},
+                        {  1,  62, 105}, {  1,  38,  66}, {  1,  18,  32},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {149,  38, 231}, { 59,  51, 186}, { 12,  54, 117},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 53, 179, 226}, {126, 176, 223}, { 58, 147, 202},
+                        { 28, 118, 174}, { 15,  94, 138}, { 14,  63,  87},
+                    },
+                    {  // band 2
+                        { 58, 196, 232}, { 26, 158, 213}, {  5, 106, 166},
+                        {  1,  75, 124}, {  1,  46,  79}, {  1,  23,  39},
+                    },
+                    {  // band 3
+                        { 46, 203, 235}, { 17, 162, 213}, {  2, 104, 165},
+                        {  1,  72, 120}, {  1,  44,  74}, {  1,  22,  33},
+                    },
+                    {  // band 4
+                        { 37, 213, 238}, {  8, 167, 216}, {  1, 104, 168},
+                        {  1,  68, 119}, {  1,  40,  67}, {  1,  17,  29},
+                    },
+                    {  // band 5
+                        { 30, 228, 239}, {  4, 181, 213}, {  1, 103, 153},
+                        {  1,  65, 110}, {  1,  43,  79}, {  1,  27,  56},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 3
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 76,  25,  53}, {  9,  18,  32}, {  1,  12,  18},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 29,  55,  91}, { 19,  58,  95}, { 15,  57,  89},
+                        { 12,  49,  77}, {  3,  29,  44}, {  1,   8,  12},
+                    },
+                    {  // band 2
+                        { 32, 160, 148}, { 33, 143, 146}, { 19, 122, 132},
+                        {  6,  90, 102}, {  1,  58,  70}, {  1,  17,  24},
+                    },
+                    {  // band 3
+                        { 16, 181, 181}, {  6, 142, 165}, {  1,  90, 120},
+                        {  1,  50,  71}, {  1,  25,  38}, {  1,   9,  14},
+                    },
+                    {  // band 4
+                        { 13, 203, 203}, {  3, 154, 176}, {  1,  80, 108},
+                        {  1,  41,  61}, {  1,  24,  37}, {  1,  11,  17},
+                    },
+                    {  // band 5
+                        {  6, 234, 240}, {  1, 178, 204}, {  1,  80, 119},
+                        {  1,  45,  71}, {  1,  26,  42}, {  1,  12,  19},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 78,  20, 135}, { 25,  18, 101}, {  5,  19,  57},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {  7, 144, 183}, {117, 151, 195}, {109, 151, 187},
+                        { 39, 130, 168}, { 11, 100, 125}, {  4,  59,  64},
+                    },
+                    {  // band 2
+                        { 20, 184, 212}, { 12, 148, 191}, {  2,  98, 141},
+                        {  1,  65, 100}, {  1,  39,  61}, {  1,  14,  22},
+                    },
+                    {  // band 3
+                        { 15, 194, 222}, {  6, 153, 198}, {  1,  92, 138},
+                        {  1,  58,  91}, {  1,  32,  52}, {  1,  12,  18},
+                    },
+                    {  // band 4
+                        { 14, 206, 232}, {  3, 162, 206}, {  1,  89, 134},
+                        {  1,  52,  83}, {  1,  28,  46}, {  1,  11,  17},
+                    },
+                    {  // band 5
+                        {  6, 225, 241}, {  1, 175, 210}, {  1,  81, 125},
+                        {  1,  48,  78}, {  1,  28,  46}, {  1,  13,  21},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {124,  23,  93}, { 31,  24,  63}, {  6,  24,  46},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 23,  86, 126}, { 45,  90, 145}, { 31,  91, 133},
+                        { 19,  80, 114}, {  7,  53,  72}, {  1,  20,  27},
+                    },
+                    {  // band 2
+                        { 51, 186, 189}, { 48, 159, 182}, { 33, 128, 156},
+                        { 15,  92, 124}, {  2,  62,  83}, {  1,  29,  43},
+                    },
+                    {  // band 3
+                        { 36, 198, 211}, { 15, 156, 187}, {  3,  97, 137},
+                        {  1,  61,  93}, {  1,  35,  57}, {  1,  15,  23},
+                    },
+                    {  // band 4
+                        { 34, 219, 223}, {  9, 162, 193}, {  1,  91, 136},
+                        {  1,  58,  92}, {  1,  35,  54}, {  1,  14,  23},
+                    },
+                    {  // band 5
+                        { 19, 243, 243}, {  3, 191, 208}, {  1,  91, 137},
+                        {  1,  56,  90}, {  1,  34,  55}, {  1,  16,  24},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {119,  20, 197}, { 19,  29, 156}, {  3,  30, 107},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 24, 192, 226}, {161, 193, 227}, { 97, 185, 222},
+                        { 31, 158, 204}, { 16, 122, 165}, { 17,  84, 112},
+                    },
+                    {  // band 2
+                        { 26, 202, 229}, { 11, 165, 210}, {  2, 103, 152},
+                        {  1,  68, 104}, {  1,  42,  70}, {  1,  16,  36},
+                    },
+                    {  // band 3
+                        { 24, 209, 237}, {  6, 169, 214}, {  1, 102, 154},
+                        {  1,  65, 107}, {  1,  45,  68}, {  1,  17,  24},
+                    },
+                    {  // band 4
+                        { 19, 219, 243}, {  4, 183, 226}, {  1, 115, 172},
+                        {  1,  73, 119}, {  1,  43,  77}, {  1,  15,  37},
+                    },
+                    {  // band 5
+                        { 11, 237, 241}, {  2, 190, 216}, {  1, 108, 146},
+                        {  1,  59,  94}, {  1,  40,  67}, {  1,  30,  53},
+                    },
+                },
+            },
+        },
+#if CONFIG_TX64X64
+        {  // TX_SIZE 4
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 76,  25,  53}, {  9,  18,  32}, {  1,  12,  18},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 29,  55,  91}, { 19,  58,  95}, { 15,  57,  89},
+                        { 12,  49,  77}, {  3,  29,  44}, {  1,   8,  12},
+                    },
+                    {  // band 2
+                        { 32, 160, 148}, { 33, 143, 146}, { 19, 122, 132},
+                        {  6,  90, 102}, {  1,  58,  70}, {  1,  17,  24},
+                    },
+                    {  // band 3
+                        { 16, 181, 181}, {  6, 142, 165}, {  1,  90, 120},
+                        {  1,  50,  71}, {  1,  25,  38}, {  1,   9,  14},
+                    },
+                    {  // band 4
+                        { 13, 203, 203}, {  3, 154, 176}, {  1,  80, 108},
+                        {  1,  41,  61}, {  1,  24,  37}, {  1,  11,  17},
+                    },
+                    {  // band 5
+                        {  6, 234, 240}, {  1, 178, 204}, {  1,  80, 119},
+                        {  1,  45,  71}, {  1,  26,  42}, {  1,  12,  19},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 78,  20, 135}, { 25,  18, 101}, {  5,  19,  57},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {  7, 144, 183}, {117, 151, 195}, {109, 151, 187},
+                        { 39, 130, 168}, { 11, 100, 125}, {  4,  59,  64},
+                    },
+                    {  // band 2
+                        { 20, 184, 212}, { 12, 148, 191}, {  2,  98, 141},
+                        {  1,  65, 100}, {  1,  39,  61}, {  1,  14,  22},
+                    },
+                    {  // band 3
+                        { 15, 194, 222}, {  6, 153, 198}, {  1,  92, 138},
+                        {  1,  58,  91}, {  1,  32,  52}, {  1,  12,  18},
+                    },
+                    {  // band 4
+                        { 14, 206, 232}, {  3, 162, 206}, {  1,  89, 134},
+                        {  1,  52,  83}, {  1,  28,  46}, {  1,  11,  17},
+                    },
+                    {  // band 5
+                        {  6, 225, 241}, {  1, 175, 210}, {  1,  81, 125},
+                        {  1,  48,  78}, {  1,  28,  46}, {  1,  13,  21},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {124,  23,  93}, { 31,  24,  63}, {  6,  24,  46},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 23,  86, 126}, { 45,  90, 145}, { 31,  91, 133},
+                        { 19,  80, 114}, {  7,  53,  72}, {  1,  20,  27},
+                    },
+                    {  // band 2
+                        { 51, 186, 189}, { 48, 159, 182}, { 33, 128, 156},
+                        { 15,  92, 124}, {  2,  62,  83}, {  1,  29,  43},
+                    },
+                    {  // band 3
+                        { 36, 198, 211}, { 15, 156, 187}, {  3,  97, 137},
+                        {  1,  61,  93}, {  1,  35,  57}, {  1,  15,  23},
+                    },
+                    {  // band 4
+                        { 34, 219, 223}, {  9, 162, 193}, {  1,  91, 136},
+                        {  1,  58,  92}, {  1,  35,  54}, {  1,  14,  23},
+                    },
+                    {  // band 5
+                        { 19, 243, 243}, {  3, 191, 208}, {  1,  91, 137},
+                        {  1,  56,  90}, {  1,  34,  55}, {  1,  16,  24},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {119,  20, 197}, { 19,  29, 156}, {  3,  30, 107},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 24, 192, 226}, {161, 193, 227}, { 97, 185, 222},
+                        { 31, 158, 204}, { 16, 122, 165}, { 17,  84, 112},
+                    },
+                    {  // band 2
+                        { 26, 202, 229}, { 11, 165, 210}, {  2, 103, 152},
+                        {  1,  68, 104}, {  1,  42,  70}, {  1,  16,  36},
+                    },
+                    {  // band 3
+                        { 24, 209, 237}, {  6, 169, 214}, {  1, 102, 154},
+                        {  1,  65, 107}, {  1,  45,  68}, {  1,  17,  24},
+                    },
+                    {  // band 4
+                        { 19, 219, 243}, {  4, 183, 226}, {  1, 115, 172},
+                        {  1,  73, 119}, {  1,  43,  77}, {  1,  15,  37},
+                    },
+                    {  // band 5
+                        { 11, 237, 241}, {  2, 190, 216}, {  1, 108, 146},
+                        {  1,  59,  94}, {  1,  40,  67}, {  1,  30,  53},
+                    },
+                },
+            },
+        },
+#endif  // CONFIG_TX64X64
+    },
+    {  // Q_Index 1
+#if CONFIG_CB4X4
+        {  // TX_SIZE 0
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        {174,  30, 159}, { 76,  38, 115}, { 15,  33,  65},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 60,  80, 153}, { 72,  75, 147}, { 36,  68, 129},
+                        { 15,  59, 104}, {  4,  45,  74}, {  1,  28,  45},
+                    },
+                    {  // band 2
+                        { 70, 122, 186}, { 55, 104, 175}, { 21,  83, 144},
+                        {  8,  67, 112}, {  2,  51,  82}, {  1,  34,  57},
+                    },
+                    {  // band 3
+                        { 97, 144, 207}, { 52, 109, 195}, { 16,  77, 153},
+                        {  4,  58, 113}, {  1,  43,  77}, {  1,  27,  48},
+                    },
+                    {  // band 4
+                        {128, 148, 229}, { 76, 104, 210}, { 18,  77, 159},
+                        {  4,  65, 110}, {  1,  52,  82}, {  1,  31,  55},
+                    },
+                    {  // band 5
+                        {165,  51, 238}, {128,  50, 230}, { 57,  49, 185},
+                        { 28,  47, 130}, { 12,  44,  96}, {  3,  36,  60},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        {169, 103, 203}, {117,  96, 176}, { 56,  81, 137},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 31, 150, 224}, { 49, 128, 212}, { 19,  92, 165},
+                        {  6,  67, 116}, {  2,  43,  71}, {  1,  21,  36},
+                    },
+                    {  // band 2
+                        { 58, 156, 230}, { 47, 130, 215}, {  7,  87, 158},
+                        {  2,  63, 114}, {  1,  39,  71}, {  1,  18,  36},
+                    },
+                    {  // band 3
+                        { 85, 176, 240}, { 43, 138, 226}, {  8,  93, 172},
+                        {  2,  70, 127}, {  1,  46,  81}, {  1,  26,  47},
+                    },
+                    {  // band 4
+                        {155, 144, 248}, { 93, 116, 235}, { 21,  83, 180},
+                        {  4,  59, 119}, {  1,  43,  80}, {  1,  25,  50},
+                    },
+                    {  // band 5
+                        {203,  61, 250}, {171,  57, 243}, { 71,  57, 199},
+                        { 31,  49, 144}, { 13,  42,  96}, {  7,  30,  52},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {204,  44, 204}, {137,  57, 184}, { 72,  62, 152},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {145, 117, 236}, {151, 112, 231}, { 87,  95, 208},
+                        { 31,  77, 165}, {  5,  49,  98}, {  1,  24,  39},
+                    },
+                    {  // band 2
+                        {146, 152, 241}, {140, 132, 236}, { 41, 103, 209},
+                        { 10,  86, 165}, {  2,  55, 106}, {  1,  25,  58},
+                    },
+                    {  // band 3
+                        {154, 181, 249}, { 84, 143, 240}, { 23, 114, 210},
+                        {  6, 102, 182}, {  2,  71, 137}, {  1,  35,  90},
+                    },
+                    {  // band 4
+                        {184, 150, 251}, {115, 130, 244}, { 34, 105, 215},
+                        { 15,  89, 173}, {  1,  51, 141}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {211,  71, 253}, {193,  78, 249}, {106,  91, 232},
+                        { 61,  87, 198}, { 85, 153, 254}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {232, 104, 242}, {165, 114, 227}, { 96, 120, 206},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {137, 178, 250}, {146, 153, 245}, { 74, 108, 205},
+                        { 41,  81, 149}, { 24,  55, 104}, { 13,  36,  68},
+                    },
+                    {  // band 2
+                        {147, 185, 252}, {127, 161, 246}, { 30, 104, 208},
+                        { 11,  74, 154}, {  6,  54, 100}, {  2,  29,  63},
+                    },
+                    {  // band 3
+                        {163, 191, 254}, {101, 161, 249}, { 22, 114, 215},
+                        {  6,  89, 173}, {  1,  65, 120}, {  1,   1, 170},
+                    },
+                    {  // band 4
+                        {197, 160, 254}, {142, 141, 251}, { 39, 102, 218},
+                        { 10,  76, 158}, {  1,  56, 122}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {224,  76, 254}, {215,  84, 253}, {107,  85, 232},
+                        { 43,  71, 177}, {  1,   1, 254}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+#endif
+        {  // TX_SIZE 0
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        {174,  30, 159}, { 76,  38, 115}, { 15,  33,  65},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 60,  80, 153}, { 72,  75, 147}, { 36,  68, 129},
+                        { 15,  59, 104}, {  4,  45,  74}, {  1,  28,  45},
+                    },
+                    {  // band 2
+                        { 70, 122, 186}, { 55, 104, 175}, { 21,  83, 144},
+                        {  8,  67, 112}, {  2,  51,  82}, {  1,  34,  57},
+                    },
+                    {  // band 3
+                        { 97, 144, 207}, { 52, 109, 195}, { 16,  77, 153},
+                        {  4,  58, 113}, {  1,  43,  77}, {  1,  27,  48},
+                    },
+                    {  // band 4
+                        {128, 148, 229}, { 76, 104, 210}, { 18,  77, 159},
+                        {  4,  65, 110}, {  1,  52,  82}, {  1,  31,  55},
+                    },
+                    {  // band 5
+                        {165,  51, 238}, {128,  50, 230}, { 57,  49, 185},
+                        { 28,  47, 130}, { 12,  44,  96}, {  3,  36,  60},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        {169, 103, 203}, {117,  96, 176}, { 56,  81, 137},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 31, 150, 224}, { 49, 128, 212}, { 19,  92, 165},
+                        {  6,  67, 116}, {  2,  43,  71}, {  1,  21,  36},
+                    },
+                    {  // band 2
+                        { 58, 156, 230}, { 47, 130, 215}, {  7,  87, 158},
+                        {  2,  63, 114}, {  1,  39,  71}, {  1,  18,  36},
+                    },
+                    {  // band 3
+                        { 85, 176, 240}, { 43, 138, 226}, {  8,  93, 172},
+                        {  2,  70, 127}, {  1,  46,  81}, {  1,  26,  47},
+                    },
+                    {  // band 4
+                        {155, 144, 248}, { 93, 116, 235}, { 21,  83, 180},
+                        {  4,  59, 119}, {  1,  43,  80}, {  1,  25,  50},
+                    },
+                    {  // band 5
+                        {203,  61, 250}, {171,  57, 243}, { 71,  57, 199},
+                        { 31,  49, 144}, { 13,  42,  96}, {  7,  30,  52},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {204,  44, 204}, {137,  57, 184}, { 72,  62, 152},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {145, 117, 236}, {151, 112, 231}, { 87,  95, 208},
+                        { 31,  77, 165}, {  5,  49,  98}, {  1,  24,  39},
+                    },
+                    {  // band 2
+                        {146, 152, 241}, {140, 132, 236}, { 41, 103, 209},
+                        { 10,  86, 165}, {  2,  55, 106}, {  1,  25,  58},
+                    },
+                    {  // band 3
+                        {154, 181, 249}, { 84, 143, 240}, { 23, 114, 210},
+                        {  6, 102, 182}, {  2,  71, 137}, {  1,  35,  90},
+                    },
+                    {  // band 4
+                        {184, 150, 251}, {115, 130, 244}, { 34, 105, 215},
+                        { 15,  89, 173}, {  1,  51, 141}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {211,  71, 253}, {193,  78, 249}, {106,  91, 232},
+                        { 61,  87, 198}, { 85, 153, 254}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {232, 104, 242}, {165, 114, 227}, { 96, 120, 206},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {137, 178, 250}, {146, 153, 245}, { 74, 108, 205},
+                        { 41,  81, 149}, { 24,  55, 104}, { 13,  36,  68},
+                    },
+                    {  // band 2
+                        {147, 185, 252}, {127, 161, 246}, { 30, 104, 208},
+                        { 11,  74, 154}, {  6,  54, 100}, {  2,  29,  63},
+                    },
+                    {  // band 3
+                        {163, 191, 254}, {101, 161, 249}, { 22, 114, 215},
+                        {  6,  89, 173}, {  1,  65, 120}, {  1,   1, 170},
+                    },
+                    {  // band 4
+                        {197, 160, 254}, {142, 141, 251}, { 39, 102, 218},
+                        { 10,  76, 158}, {  1,  56, 122}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {224,  76, 254}, {215,  84, 253}, {107,  85, 232},
+                        { 43,  71, 177}, {  1,   1, 254}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 1
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 68,  37, 120}, { 21,  34,  82}, {  5,  26,  49},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 41,  89, 138}, { 56,  83, 132}, { 31,  73, 115},
+                        { 16,  62,  92}, {  5,  45,  62}, {  1,  24,  32},
+                    },
+                    {  // band 2
+                        { 48, 139, 165}, { 30, 114, 160}, { 13,  92, 132},
+                        {  6,  72, 103}, {  3,  49,  72}, {  1,  26,  41},
+                    },
+                    {  // band 3
+                        { 44, 162, 191}, { 20, 127, 175}, {  5,  90, 137},
+                        {  1,  62, 100}, {  1,  38,  63}, {  1,  20,  32},
+                    },
+                    {  // band 4
+                        { 51, 184, 213}, { 16, 137, 193}, {  2,  89, 143},
+                        {  1,  60, 102}, {  1,  39,  66}, {  1,  23,  37},
+                    },
+                    {  // band 5
+                        { 76, 200, 235}, { 27, 150, 216}, {  3,  99, 164},
+                        {  1,  70, 119}, {  1,  45,  77}, {  1,  22,  38},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 81, 112, 199}, { 49, 101, 164}, { 19,  80, 119},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 12, 181, 217}, { 48, 151, 212}, { 38, 118, 180},
+                        { 22,  95, 140}, { 11,  67,  92}, { 13,  46,  44},
+                    },
+                    {  // band 2
+                        { 29, 188, 226}, { 19, 147, 210}, {  5,  95, 154},
+                        {  4,  68, 106}, {  3,  44,  60}, {  1,  24,  27},
+                    },
+                    {  // band 3
+                        { 30, 195, 234}, { 15, 153, 216}, {  3,  95, 156},
+                        {  2,  66, 108}, {  2,  44,  62}, {  1,  24,  29},
+                    },
+                    {  // band 4
+                        { 36, 203, 243}, { 12, 162, 225}, {  2,  98, 163},
+                        {  2,  67, 113}, {  2,  45,  68}, {  1,  24,  34},
+                    },
+                    {  // band 5
+                        { 86, 207, 248}, { 35, 165, 236}, {  3, 107, 180},
+                        {  1,  73, 128}, {  1,  45,  78}, {  1,  20,  34},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {188,  37, 205}, {118,  51, 172}, { 56,  57, 135},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {116, 135, 225}, {144, 123, 221}, { 72, 103, 197},
+                        { 35,  77, 153}, { 15,  47,  82}, {  6,  25,  34},
+                    },
+                    {  // band 2
+                        {128, 171, 233}, { 82, 142, 226}, { 31, 106, 191},
+                        { 16,  82, 146}, {  9,  59,  98}, {  4,  33,  54},
+                    },
+                    {  // band 3
+                        {126, 197, 241}, { 66, 155, 230}, { 18, 108, 190},
+                        {  7,  82, 148}, {  3,  58,  98}, {  1,  25,  50},
+                    },
+                    {  // band 4
+                        {117, 207, 244}, { 44, 163, 233}, {  9, 112, 191},
+                        {  5,  84, 148}, {  3,  61,  87}, {  1,  28,  38},
+                    },
+                    {  // band 5
+                        {112, 214, 249}, { 39, 174, 240}, {  6, 125, 205},
+                        {  4,  96, 163}, {  5,  66, 100}, {  1, 128, 254},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {227,  70, 234}, {145,  91, 213}, { 61, 100, 173},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {108, 198, 243}, {171, 172, 240}, {118, 130, 210},
+                        {104, 107, 165}, { 64,  85, 114}, { 55,  64,  60},
+                    },
+                    {  // band 2
+                        {110, 208, 247}, { 64, 175, 237}, { 24, 112, 187},
+                        { 24,  81, 133}, { 24,  63,  83}, { 21,  47,  53},
+                    },
+                    {  // band 3
+                        { 91, 218, 249}, { 46, 188, 238}, {  8, 113, 184},
+                        {  5,  83, 137}, {  6,  62,  95}, { 17,  44,  94},
+                    },
+                    {  // band 4
+                        { 84, 216, 248}, { 30, 187, 237}, {  2, 117, 188},
+                        {  1,  88, 141}, {  3,  63,  98}, {  1,   1,   1},
+                    },
+                    {  // band 5
+                        {116, 218, 252}, { 47, 186, 242}, {  2, 132, 204},
+                        {  1, 106, 175}, {  1,  88, 104}, {  1, 254, 128},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 2
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 35,  41, 129}, { 12,  30,  70}, {  2,  19,  32},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 30,  77, 116}, { 39,  70, 110}, { 20,  58,  96},
+                        {  8,  47,  77}, {  2,  33,  52}, {  1,  17,  26},
+                    },
+                    {  // band 2
+                        { 31, 123, 146}, { 18, 103, 140}, {  7,  81, 119},
+                        {  2,  62,  95}, {  1,  44,  70}, {  1,  26,  42},
+                    },
+                    {  // band 3
+                        { 21, 149, 170}, {  9, 114, 158}, {  2,  80, 126},
+                        {  1,  57,  94}, {  1,  36,  61}, {  1,  18,  31},
+                    },
+                    {  // band 4
+                        { 20, 178, 199}, {  6, 134, 183}, {  1,  87, 139},
+                        {  1,  60, 100}, {  1,  37,  64}, {  1,  18,  31},
+                    },
+                    {  // band 5
+                        { 36, 218, 233}, {  6, 160, 207}, {  1,  92, 147},
+                        {  1,  59, 101}, {  1,  35,  62}, {  1,  18,  31},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 17,  62, 211}, { 14,  62, 153}, {  5,  50,  84},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 11, 180, 205}, { 87, 160, 205}, { 53, 128, 184},
+                        { 27, 106, 156}, { 13,  79, 115}, {  6,  46,  67},
+                    },
+                    {  // band 2
+                        { 32, 194, 220}, { 20, 145, 202}, {  4,  96, 152},
+                        {  1,  67, 111}, {  1,  42,  70}, {  1,  21,  37},
+                    },
+                    {  // band 3
+                        { 30, 204, 228}, { 14, 152, 207}, {  1,  92, 149},
+                        {  1,  61, 103}, {  1,  34,  59}, {  1,  16,  28},
+                    },
+                    {  // band 4
+                        { 27, 213, 235}, {  7, 159, 210}, {  1,  88, 143},
+                        {  1,  55,  94}, {  1,  31,  53}, {  1,  16,  27},
+                    },
+                    {  // band 5
+                        { 28, 223, 243}, {  4, 173, 217}, {  1,  91, 146},
+                        {  1,  58,  98}, {  1,  35,  60}, {  1,  19,  33},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {172,  37, 202}, { 83,  51, 156}, { 24,  53, 110},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 76, 134, 206}, {110, 124, 200}, { 47, 106, 180},
+                        { 15,  82, 145}, {  3,  48,  83}, {  1,  19,  32},
+                    },
+                    {  // band 2
+                        { 80, 176, 220}, { 49, 145, 212}, { 17, 112, 180},
+                        {  7,  84, 140}, {  1,  53,  89}, {  1,  27,  43},
+                    },
+                    {  // band 3
+                        { 74, 201, 232}, { 38, 158, 221}, {  8, 112, 179},
+                        {  2,  79, 132}, {  1,  47,  82}, {  1,  26,  42},
+                    },
+                    {  // band 4
+                        { 73, 215, 239}, { 28, 169, 227}, {  3, 112, 176},
+                        {  1,  74, 126}, {  1,  48,  79}, {  1,  27,  44},
+                    },
+                    {  // band 5
+                        { 71, 233, 244}, { 18, 180, 230}, {  1, 114, 180},
+                        {  1,  80, 134}, {  1,  51,  85}, {  1,  26,  36},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {213,  34, 244}, {126,  57, 212}, { 46,  67, 151},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {120, 202, 245}, {198, 173, 241}, {119, 146, 224},
+                        { 76, 126, 195}, { 44, 102, 159}, { 40,  76, 115},
+                    },
+                    {  // band 2
+                        {120, 215, 248}, { 69, 171, 237}, { 23, 119, 194},
+                        { 10,  86, 147}, {  2,  56,  94}, {  1,  25,  44},
+                    },
+                    {  // band 3
+                        {102, 226, 250}, { 53, 183, 239}, {  9, 118, 188},
+                        {  2,  78, 131}, {  1,  48,  89}, {  1,  17,  36},
+                    },
+                    {  // band 4
+                        { 86, 235, 252}, { 34, 194, 240}, {  2, 109, 173},
+                        {  1,  68, 118}, {  1,  44,  79}, {  1,   1,  38},
+                    },
+                    {  // band 5
+                        { 59, 236, 243}, { 11, 189, 228}, {  1, 112, 187},
+                        {  1,  88, 145}, {  1,  55,  92}, {  1,   1, 128},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 3
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 41,  40, 104}, { 12,  31,  64}, {  2,  16,  28},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 65,  58, 132}, { 50,  61, 130}, { 40,  57, 116},
+                        { 22,  46,  87}, {  2,  28,  44}, {  1,  11,  17},
+                    },
+                    {  // band 2
+                        { 55, 139, 135}, { 46, 122, 132}, { 21,  89, 110},
+                        {  6,  60,  78}, {  1,  38,  54}, {  1,  17,  27},
+                    },
+                    {  // band 3
+                        { 29, 167, 161}, { 10, 120, 141}, {  1,  69,  98},
+                        {  1,  42,  66}, {  1,  28,  44}, {  1,  15,  24},
+                    },
+                    {  // band 4
+                        { 19, 191, 180}, {  4, 125, 154}, {  1,  70, 107},
+                        {  1,  48,  77}, {  1,  33,  53}, {  1,  17,  28},
+                    },
+                    {  // band 5
+                        { 16, 238, 231}, {  2, 163, 198}, {  1,  85, 134},
+                        {  1,  54,  90}, {  1,  34,  57}, {  1,  17,  29},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 70,  15, 216}, { 40,  18, 164}, { 14,  17,  83},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 25, 150, 200}, {185, 154, 211}, {123, 137, 199},
+                        { 67, 119, 177}, { 31,  96, 137}, { 18,  63,  86},
+                    },
+                    {  // band 2
+                        { 57, 187, 223}, { 35, 148, 207}, {  7, 104, 159},
+                        {  2,  72, 113}, {  1,  44,  71}, {  1,  20,  34},
+                    },
+                    {  // band 3
+                        { 44, 203, 233}, { 18, 157, 212}, {  1,  98, 150},
+                        {  1,  61, 102}, {  1,  38,  62}, {  1,  19,  31},
+                    },
+                    {  // band 4
+                        { 41, 215, 238}, { 11, 166, 215}, {  1,  94, 146},
+                        {  1,  60, 101}, {  1,  37,  63}, {  1,  17,  28},
+                    },
+                    {  // band 5
+                        { 19, 236, 246}, {  3, 188, 223}, {  1,  95, 146},
+                        {  1,  58,  95}, {  1,  34,  56}, {  1,  17,  27},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {146,  27, 156}, { 49,  32, 116}, { 10,  39,  77},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 47, 101, 172}, { 93, 100, 178}, { 58,  91, 165},
+                        { 26,  75, 134}, {  4,  49,  82}, {  2,  22,  33},
+                    },
+                    {  // band 2
+                        { 60, 158, 196}, { 44, 135, 186}, { 25, 106, 157},
+                        {  8,  81, 124}, {  2,  56,  86}, {  1,  28,  45},
+                    },
+                    {  // band 3
+                        { 44, 169, 212}, { 15, 138, 196}, {  2, 100, 157},
+                        {  1,  74, 119}, {  1,  49,  76}, {  1,  20,  34},
+                    },
+                    {  // band 4
+                        { 38, 199, 231}, { 11, 158, 214}, {  1, 111, 167},
+                        {  1,  76, 122}, {  1,  44,  76}, {  1,  17,  39},
+                    },
+                    {  // band 5
+                        { 40, 236, 246}, { 10, 187, 230}, {  1, 115, 175},
+                        {  1,  74, 122}, {  1,  42,  71}, {  1,  14,  59},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {161,  26, 237}, { 65,  46, 209}, { 21,  46, 161},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 87, 229, 245}, {206, 214, 244}, {148, 186, 236},
+                        { 89, 165, 221}, { 41, 132, 186}, { 37,  93, 141},
+                    },
+                    {  // band 2
+                        { 93, 231, 246}, { 47, 181, 231}, {  8, 117, 188},
+                        {  2,  84, 138}, {  1,  43,  87}, {  1,  27,  41},
+                    },
+                    {  // band 3
+                        { 80, 239, 250}, { 28, 190, 236}, {  1, 119, 183},
+                        {  1,  84, 135}, {  1,  81,  69}, {  1, 102,   1},
+                    },
+                    {  // band 4
+                        { 67, 245, 252}, { 22, 206, 242}, {  1, 130, 195},
+                        {  1,  77, 136}, {  1,  35,  88}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        { 43, 250, 228}, { 31, 185, 204}, {  6, 101, 183},
+                        {  1,  92, 151}, {  1,  84, 137}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+#if CONFIG_TX64X64
+        {  // TX_SIZE 4
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 41,  40, 104}, { 12,  31,  64}, {  2,  16,  28},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 65,  58, 132}, { 50,  61, 130}, { 40,  57, 116},
+                        { 22,  46,  87}, {  2,  28,  44}, {  1,  11,  17},
+                    },
+                    {  // band 2
+                        { 55, 139, 135}, { 46, 122, 132}, { 21,  89, 110},
+                        {  6,  60,  78}, {  1,  38,  54}, {  1,  17,  27},
+                    },
+                    {  // band 3
+                        { 29, 167, 161}, { 10, 120, 141}, {  1,  69,  98},
+                        {  1,  42,  66}, {  1,  28,  44}, {  1,  15,  24},
+                    },
+                    {  // band 4
+                        { 19, 191, 180}, {  4, 125, 154}, {  1,  70, 107},
+                        {  1,  48,  77}, {  1,  33,  53}, {  1,  17,  28},
+                    },
+                    {  // band 5
+                        { 16, 238, 231}, {  2, 163, 198}, {  1,  85, 134},
+                        {  1,  54,  90}, {  1,  34,  57}, {  1,  17,  29},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 70,  15, 216}, { 40,  18, 164}, { 14,  17,  83},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 25, 150, 200}, {185, 154, 211}, {123, 137, 199},
+                        { 67, 119, 177}, { 31,  96, 137}, { 18,  63,  86},
+                    },
+                    {  // band 2
+                        { 57, 187, 223}, { 35, 148, 207}, {  7, 104, 159},
+                        {  2,  72, 113}, {  1,  44,  71}, {  1,  20,  34},
+                    },
+                    {  // band 3
+                        { 44, 203, 233}, { 18, 157, 212}, {  1,  98, 150},
+                        {  1,  61, 102}, {  1,  38,  62}, {  1,  19,  31},
+                    },
+                    {  // band 4
+                        { 41, 215, 238}, { 11, 166, 215}, {  1,  94, 146},
+                        {  1,  60, 101}, {  1,  37,  63}, {  1,  17,  28},
+                    },
+                    {  // band 5
+                        { 19, 236, 246}, {  3, 188, 223}, {  1,  95, 146},
+                        {  1,  58,  95}, {  1,  34,  56}, {  1,  17,  27},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {146,  27, 156}, { 49,  32, 116}, { 10,  39,  77},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 47, 101, 172}, { 93, 100, 178}, { 58,  91, 165},
+                        { 26,  75, 134}, {  4,  49,  82}, {  2,  22,  33},
+                    },
+                    {  // band 2
+                        { 60, 158, 196}, { 44, 135, 186}, { 25, 106, 157},
+                        {  8,  81, 124}, {  2,  56,  86}, {  1,  28,  45},
+                    },
+                    {  // band 3
+                        { 44, 169, 212}, { 15, 138, 196}, {  2, 100, 157},
+                        {  1,  74, 119}, {  1,  49,  76}, {  1,  20,  34},
+                    },
+                    {  // band 4
+                        { 38, 199, 231}, { 11, 158, 214}, {  1, 111, 167},
+                        {  1,  76, 122}, {  1,  44,  76}, {  1,  17,  39},
+                    },
+                    {  // band 5
+                        { 40, 236, 246}, { 10, 187, 230}, {  1, 115, 175},
+                        {  1,  74, 122}, {  1,  42,  71}, {  1,  14,  59},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {161,  26, 237}, { 65,  46, 209}, { 21,  46, 161},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 87, 229, 245}, {206, 214, 244}, {148, 186, 236},
+                        { 89, 165, 221}, { 41, 132, 186}, { 37,  93, 141},
+                    },
+                    {  // band 2
+                        { 93, 231, 246}, { 47, 181, 231}, {  8, 117, 188},
+                        {  2,  84, 138}, {  1,  43,  87}, {  1,  27,  41},
+                    },
+                    {  // band 3
+                        { 80, 239, 250}, { 28, 190, 236}, {  1, 119, 183},
+                        {  1,  84, 135}, {  1,  81,  69}, {  1, 102,   1},
+                    },
+                    {  // band 4
+                        { 67, 245, 252}, { 22, 206, 242}, {  1, 130, 195},
+                        {  1,  77, 136}, {  1,  35,  88}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        { 43, 250, 228}, { 31, 185, 204}, {  6, 101, 183},
+                        {  1,  92, 151}, {  1,  84, 137}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+#endif  // CONFIG_TX64X64
+    },
+    {  // Q_Index 2
+#if CONFIG_CB4X4
+        {  // TX_SIZE 0
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        {181,  22, 175}, { 96,  37, 147}, { 35,  41, 105},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 80,  95, 197}, {111,  92, 193}, { 59,  87, 175},
+                        { 29,  79, 150}, { 10,  65, 118}, {  2,  47,  82},
+                    },
+                    {  // band 2
+                        { 90, 141, 216}, { 77, 120, 210}, { 23,  95, 184},
+                        { 11,  81, 151}, {  6,  75, 130}, {  2,  58, 113},
+                    },
+                    {  // band 3
+                        {122, 167, 231}, { 66, 119, 225}, { 26,  87, 189},
+                        {  7,  76, 151}, {  2,  63, 125}, {  1,  59,  77},
+                    },
+                    {  // band 4
+                        {162, 147, 244}, {110,  97, 236}, { 32,  88, 204},
+                        { 11,  89, 174}, {  5,  78, 151}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {205,  59, 251}, {176,  68, 248}, { 90,  71, 223},
+                        { 49,  72, 188}, { 17,  74, 203}, {128, 128, 128},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        {188,  70, 207}, {140,  73, 189}, { 85,  73, 163},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 59, 144, 239}, { 79, 126, 237}, { 31, 102, 202},
+                        { 10,  81, 153}, {  3,  56, 102}, {  2,  33,  59},
+                    },
+                    {  // band 2
+                        {100, 152, 243}, { 80, 129, 236}, { 14,  94, 194},
+                        {  4,  72, 150}, {  1,  50, 103}, {  1,  35,  60},
+                    },
+                    {  // band 3
+                        {130, 183, 247}, { 70, 139, 242}, { 19, 100, 203},
+                        {  4,  83, 159}, {  1,  59, 119}, {  1,  44,  72},
+                    },
+                    {  // band 4
+                        {197, 138, 252}, {135, 107, 247}, { 31,  86, 210},
+                        {  7,  74, 160}, {  1,  53, 107}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {229,  54, 254}, {200,  51, 251}, { 83,  61, 226},
+                        { 33,  55, 177}, { 12,  74, 145}, {128, 128, 128},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {229,  20, 235}, {183,  37, 221}, {127,  47, 198},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {188, 115, 251}, {208, 110, 250}, {101,  99, 235},
+                        { 38,  81, 197}, {  9,  56, 132}, {  9,  52,  63},
+                    },
+                    {  // band 2
+                        {189, 150, 252}, {186, 137, 251}, { 54, 107, 236},
+                        { 14,  90, 195}, {  1,  89, 104}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {209, 180, 254}, {142, 145, 253}, { 51, 130, 236},
+                        {  6, 128, 214}, {  1, 128, 254}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {231, 140, 254}, {194, 128, 254}, { 75, 119, 233},
+                        {128,  23, 230}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {244,  59, 254}, {239,  81, 254}, {128,  85, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {246,  55, 247}, {197,  64, 235}, {141,  74, 218},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {178, 163, 254}, {192, 138, 252}, { 85, 103, 231},
+                        { 49,  81, 179}, { 32,  54, 133}, { 12,  26,  98},
+                    },
+                    {  // band 2
+                        {189, 173, 254}, {179, 150, 253}, { 60,  94, 237},
+                        { 34,  81, 198}, { 20,  53, 187}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {202, 191, 254}, {157, 160, 254}, { 57, 117, 240},
+                        { 28, 105, 211}, {  1, 128,   1}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {231, 146, 254}, {208, 133, 254}, { 66,  78, 233},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {246,  49, 254}, {246,  63, 254}, { 85, 142, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+#endif
+        {  // TX_SIZE 0
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        {181,  22, 175}, { 96,  37, 147}, { 35,  41, 105},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 80,  95, 197}, {111,  92, 193}, { 59,  87, 175},
+                        { 29,  79, 150}, { 10,  65, 118}, {  2,  47,  82},
+                    },
+                    {  // band 2
+                        { 90, 141, 216}, { 77, 120, 210}, { 23,  95, 184},
+                        { 11,  81, 151}, {  6,  75, 130}, {  2,  58, 113},
+                    },
+                    {  // band 3
+                        {122, 167, 231}, { 66, 119, 225}, { 26,  87, 189},
+                        {  7,  76, 151}, {  2,  63, 125}, {  1,  59,  77},
+                    },
+                    {  // band 4
+                        {162, 147, 244}, {110,  97, 236}, { 32,  88, 204},
+                        { 11,  89, 174}, {  5,  78, 151}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {205,  59, 251}, {176,  68, 248}, { 90,  71, 223},
+                        { 49,  72, 188}, { 17,  74, 203}, {128, 128, 128},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        {188,  70, 207}, {140,  73, 189}, { 85,  73, 163},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 59, 144, 239}, { 79, 126, 237}, { 31, 102, 202},
+                        { 10,  81, 153}, {  3,  56, 102}, {  2,  33,  59},
+                    },
+                    {  // band 2
+                        {100, 152, 243}, { 80, 129, 236}, { 14,  94, 194},
+                        {  4,  72, 150}, {  1,  50, 103}, {  1,  35,  60},
+                    },
+                    {  // band 3
+                        {130, 183, 247}, { 70, 139, 242}, { 19, 100, 203},
+                        {  4,  83, 159}, {  1,  59, 119}, {  1,  44,  72},
+                    },
+                    {  // band 4
+                        {197, 138, 252}, {135, 107, 247}, { 31,  86, 210},
+                        {  7,  74, 160}, {  1,  53, 107}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {229,  54, 254}, {200,  51, 251}, { 83,  61, 226},
+                        { 33,  55, 177}, { 12,  74, 145}, {128, 128, 128},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {229,  20, 235}, {183,  37, 221}, {127,  47, 198},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {188, 115, 251}, {208, 110, 250}, {101,  99, 235},
+                        { 38,  81, 197}, {  9,  56, 132}, {  9,  52,  63},
+                    },
+                    {  // band 2
+                        {189, 150, 252}, {186, 137, 251}, { 54, 107, 236},
+                        { 14,  90, 195}, {  1,  89, 104}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {209, 180, 254}, {142, 145, 253}, { 51, 130, 236},
+                        {  6, 128, 214}, {  1, 128, 254}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {231, 140, 254}, {194, 128, 254}, { 75, 119, 233},
+                        {128,  23, 230}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {244,  59, 254}, {239,  81, 254}, {128,  85, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {246,  55, 247}, {197,  64, 235}, {141,  74, 218},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {178, 163, 254}, {192, 138, 252}, { 85, 103, 231},
+                        { 49,  81, 179}, { 32,  54, 133}, { 12,  26,  98},
+                    },
+                    {  // band 2
+                        {189, 173, 254}, {179, 150, 253}, { 60,  94, 237},
+                        { 34,  81, 198}, { 20,  53, 187}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {202, 191, 254}, {157, 160, 254}, { 57, 117, 240},
+                        { 28, 105, 211}, {  1, 128,   1}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {231, 146, 254}, {208, 133, 254}, { 66,  78, 233},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {246,  49, 254}, {246,  63, 254}, { 85, 142, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 1
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 45,  28, 124}, { 23,  35, 107}, { 10,  34,  78},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 53,  99, 177}, { 82,  96, 174}, { 46,  89, 158},
+                        { 21,  76, 133}, {  6,  56,  94}, {  1,  33,  54},
+                    },
+                    {  // band 2
+                        { 68, 147, 201}, { 42, 124, 195}, { 17,  98, 166},
+                        {  7,  75, 131}, {  2,  53,  93}, {  1,  33,  59},
+                    },
+                    {  // band 3
+                        { 65, 176, 217}, { 30, 137, 206}, {  6,  97, 167},
+                        {  2,  70, 128}, {  1,  47,  88}, {  1,  29,  46},
+                    },
+                    {  // band 4
+                        { 69, 195, 232}, { 24, 146, 218}, {  4, 100, 175},
+                        {  2,  72, 134}, {  1,  51,  93}, {  1,  29,  52},
+                    },
+                    {  // band 5
+                        { 96, 212, 246}, { 39, 158, 234}, {  6, 109, 192},
+                        {  2,  77, 144}, {  1,  50,  95}, {  1,  20,  45},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 71,  80, 213}, { 53,  73, 181}, { 25,  66, 141},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 35, 168, 231}, { 91, 150, 229}, { 49, 122, 202},
+                        { 22,  97, 162}, { 10,  68, 108}, {  9,  48,  57},
+                    },
+                    {  // band 2
+                        { 56, 178, 236}, { 32, 148, 225}, {  9,  99, 176},
+                        {  4,  69, 127}, {  2,  44,  78}, {  1,  25,  41},
+                    },
+                    {  // band 3
+                        { 57, 191, 242}, { 27, 155, 230}, {  5, 102, 180},
+                        {  2,  71, 133}, {  1,  44,  78}, {  1,  27,  41},
+                    },
+                    {  // band 4
+                        { 67, 201, 247}, { 24, 162, 237}, {  3, 106, 188},
+                        {  3,  74, 137}, {  1,  46,  85}, {  1,  34,  48},
+                    },
+                    {  // band 5
+                        {111, 210, 251}, { 47, 166, 244}, {  3, 113, 199},
+                        {  2,  77, 146}, {  1,  48,  93}, {  1,  38,  22},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {206,  21, 221}, {150,  36, 195}, { 94,  44, 164},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {147, 128, 239}, {194, 122, 238}, { 95, 104, 220},
+                        { 39,  81, 183}, { 13,  53, 111}, {  3,  24,  49},
+                    },
+                    {  // band 2
+                        {164, 163, 244}, {106, 142, 239}, { 50, 112, 215},
+                        { 26,  90, 177}, { 12,  67, 130}, {  1,   1,  64},
+                    },
+                    {  // band 3
+                        {155, 193, 249}, { 88, 158, 244}, { 26, 124, 220},
+                        { 10,  98, 173}, {  1,  77, 126}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {141, 205, 252}, { 64, 174, 248}, { 17, 124, 221},
+                        { 12,  92, 176}, {  1,  29, 148}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {150, 217, 254}, { 74, 191, 252}, { 30, 144, 215},
+                        {  1, 106, 137}, {128,   1, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {241,  37, 242}, {175,  48, 223}, { 99,  53, 189},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {153, 183, 248}, {212, 156, 247}, {134, 124, 221},
+                        { 88, 103, 184}, { 59,  86, 132}, { 29,  61,  67},
+                    },
+                    {  // band 2
+                        {162, 199, 250}, {106, 167, 247}, { 56, 110, 207},
+                        { 32,  85, 165}, { 16,  71, 130}, {  1,  93, 254},
+                    },
+                    {  // band 3
+                        {143, 213, 252}, { 86, 187, 250}, { 23, 124, 220},
+                        {  7,  95, 176}, {  1, 109, 102}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {130, 219, 254}, { 70, 201, 253}, { 15, 128, 215},
+                        {  1, 101, 201}, {  1,  64, 170}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {155, 219, 254}, {105, 207, 254}, { 28, 155, 229},
+                        {  1, 153, 191}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 2
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 18,  26, 117}, { 10,  29,  82}, {  3,  25,  52},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 35,  88, 152}, { 62,  85, 150}, { 36,  77, 137},
+                        { 16,  66, 116}, {  4,  47,  81}, {  1,  26,  44},
+                    },
+                    {  // band 2
+                        { 55, 141, 182}, { 32, 119, 177}, { 12,  93, 154},
+                        {  4,  71, 123}, {  1,  51,  89}, {  1,  32,  56},
+                    },
+                    {  // band 3
+                        { 46, 171, 202}, { 21, 130, 191}, {  5,  91, 154},
+                        {  1,  64, 115}, {  1,  42,  77}, {  1,  25,  41},
+                    },
+                    {  // band 4
+                        { 43, 195, 219}, { 12, 142, 203}, {  1,  91, 156},
+                        {  1,  63, 115}, {  1,  41,  77}, {  1,  22,  43},
+                    },
+                    {  // band 5
+                        { 42, 221, 238}, {  8, 162, 219}, {  1,  98, 167},
+                        {  1,  67, 123}, {  1,  43,  83}, {  1,  25,  38},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 16,  51, 216}, { 20,  48, 168}, {  9,  44, 109},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 34, 164, 226}, {124, 148, 226}, { 72, 127, 207},
+                        { 36, 107, 175}, { 15,  81, 129}, {  6,  51,  79},
+                    },
+                    {  // band 2
+                        { 61, 182, 234}, { 35, 148, 220}, {  9, 101, 178},
+                        {  4,  71, 134}, {  1,  46,  90}, {  1,  24,  51},
+                    },
+                    {  // band 3
+                        { 54, 198, 239}, { 25, 156, 224}, {  3,  98, 173},
+                        {  1,  66, 124}, {  1,  41,  78}, {  1,  15,  37},
+                    },
+                    {  // band 4
+                        { 48, 209, 242}, { 12, 162, 226}, {  1,  96, 169},
+                        {  1,  63, 119}, {  1,  40,  78}, {  1,  18,  45},
+                    },
+                    {  // band 5
+                        { 44, 223, 247}, {  6, 173, 232}, {  1, 105, 178},
+                        {  1,  71, 131}, {  1,  44,  84}, {  1,  13,  46},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {188,  26, 214}, {121,  42, 181}, { 66,  49, 149},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {136, 128, 233}, {172, 124, 230}, { 80, 106, 211},
+                        { 27,  81, 174}, {  6,  49,  98}, {  8,  28,  49},
+                    },
+                    {  // band 2
+                        {145, 166, 239}, { 92, 141, 229}, { 28, 108, 196},
+                        {  8,  87, 154}, {  1,  58, 105}, {  1,  27,  59},
+                    },
+                    {  // band 3
+                        {131, 193, 242}, { 66, 151, 231}, { 13, 112, 192},
+                        {  2,  81, 152}, {  1,  66, 121}, {  1,  23,  64},
+                    },
+                    {  // band 4
+                        {112, 211, 246}, { 41, 164, 235}, {  5, 117, 202},
+                        {  1,  83, 162}, {  1,  64, 111}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        { 96, 230, 250}, { 28, 185, 243}, {  2, 132, 204},
+                        {  1,  91, 166}, {  1,  85,  46}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {238,  23, 242}, {157,  29, 215}, { 73,  27, 162},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {165, 173, 250}, {222, 151, 247}, {152, 134, 235},
+                        {114, 120, 210}, { 86, 109, 176}, { 53,  88, 145},
+                    },
+                    {  // band 2
+                        {164, 194, 249}, {100, 158, 241}, { 35, 111, 212},
+                        { 17,  85, 167}, {  1,  52, 112}, {  1,  73,   1},
+                    },
+                    {  // band 3
+                        {151, 215, 252}, { 83, 172, 245}, { 16, 122, 208},
+                        {  6, 101, 165}, {  1,  74, 113}, {  1,   1,   1},
+                    },
+                    {  // band 4
+                        {138, 230, 253}, { 65, 184, 248}, {  8, 128, 212},
+                        {  1, 111, 182}, {128,   1,   1}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {123, 240, 253}, { 36, 201, 250}, {  3, 127, 211},
+                        {  1,  68, 204}, {128,   1,   1}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 3
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 51,  21, 156}, { 30,  23,  86}, {  4,  18,  37},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 38,  77, 129}, { 79,  76, 129}, { 40,  66, 117},
+                        { 12,  54,  95}, {  1,  36,  60}, {  1,  17,  29},
+                    },
+                    {  // band 2
+                        { 44, 133, 149}, { 24, 107, 143}, {  8,  78, 121},
+                        {  3,  59,  97}, {  1,  42,  71}, {  1,  22,  37},
+                    },
+                    {  // band 3
+                        { 29, 160, 171}, {  9, 114, 158}, {  1,  76, 125},
+                        {  1,  54,  93}, {  1,  36,  63}, {  1,  20,  35},
+                    },
+                    {  // band 4
+                        { 22, 188, 205}, {  6, 132, 186}, {  1,  87, 144},
+                        {  1,  62, 107}, {  1,  41,  72}, {  1,  23,  41},
+                    },
+                    {  // band 5
+                        { 25, 233, 236}, {  5, 165, 214}, {  1,  96, 158},
+                        {  1,  63, 112}, {  1,  40,  73}, {  1,  23,  40},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 48,  20, 231}, { 37,  21, 179}, { 15,  18, 109},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 41, 154, 216}, {196, 142, 221}, {131, 125, 207},
+                        { 84, 111, 181}, { 45,  91, 142}, { 27,  62,  89},
+                    },
+                    {  // band 2
+                        { 72, 181, 230}, { 41, 147, 215}, { 10, 102, 173},
+                        {  3,  73, 132}, {  1,  47,  89}, {  1,  23,  50},
+                    },
+                    {  // band 3
+                        { 60, 201, 236}, { 23, 157, 219}, {  2,  99, 167},
+                        {  1,  69, 124}, {  1,  43,  80}, {  1,  22,  39},
+                    },
+                    {  // band 4
+                        { 53, 214, 242}, { 15, 165, 224}, {  1, 101, 173},
+                        {  1,  70, 131}, {  1,  44,  83}, {  1,  23,  49},
+                    },
+                    {  // band 5
+                        { 39, 239, 248}, {  7, 186, 233}, {  1, 108, 174},
+                        {  1,  70, 123}, {  1,  43,  77}, {  1,  16,  42},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {161,  26, 204}, { 77,  40, 160}, { 26,  50, 117},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 80, 140, 218}, {136, 133, 215}, { 63, 117, 197},
+                        { 20,  93, 170}, {  7,  55, 102}, { 13,  32,  52},
+                    },
+                    {  // band 2
+                        { 86, 173, 231}, { 46, 150, 220}, { 18, 118, 190},
+                        {  8,  90, 150}, {  2,  60,  95}, {  1,  39,  41},
+                    },
+                    {  // band 3
+                        { 80, 183, 242}, { 37, 160, 231}, {  6, 120, 182},
+                        {  1,  86, 137}, {  1,  46,  78}, {  1,  15,  24},
+                    },
+                    {  // band 4
+                        { 88, 215, 247}, { 42, 179, 235}, {  4, 116, 182},
+                        {  2,  80, 133}, {  1,  46,  85}, {  1,  64,  43},
+                    },
+                    {  // band 5
+                        {100, 236, 250}, { 31, 186, 234}, {  1, 114, 181},
+                        {  1,  85, 135}, {  1,  78,  64}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {213,  13, 245}, {106,  16, 211}, { 32,  11, 156},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {140, 214, 247}, {241, 186, 243}, {177, 172, 235},
+                        {128, 156, 219}, {106, 130, 191}, { 99, 105, 152},
+                    },
+                    {  // band 2
+                        {125, 218, 248}, { 75, 167, 239}, { 29, 111, 212},
+                        {  6,  66, 152}, {  1,  42,  96}, {  1,  85, 128},
+                    },
+                    {  // band 3
+                        {120, 232, 252}, { 60, 189, 247}, {  8, 141, 200},
+                        {  1,  89, 134}, {  1,  32, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {111, 238, 253}, { 56, 198, 245}, {  1, 123, 208},
+                        {  1,  93, 176}, {  1,   1,  73}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        { 98, 251, 249}, { 56, 189, 244}, { 17, 113, 220},
+                        {  1, 109, 179}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+#if CONFIG_TX64X64
+        {  // TX_SIZE 4
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 51,  21, 156}, { 30,  23,  86}, {  4,  18,  37},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 38,  77, 129}, { 79,  76, 129}, { 40,  66, 117},
+                        { 12,  54,  95}, {  1,  36,  60}, {  1,  17,  29},
+                    },
+                    {  // band 2
+                        { 44, 133, 149}, { 24, 107, 143}, {  8,  78, 121},
+                        {  3,  59,  97}, {  1,  42,  71}, {  1,  22,  37},
+                    },
+                    {  // band 3
+                        { 29, 160, 171}, {  9, 114, 158}, {  1,  76, 125},
+                        {  1,  54,  93}, {  1,  36,  63}, {  1,  20,  35},
+                    },
+                    {  // band 4
+                        { 22, 188, 205}, {  6, 132, 186}, {  1,  87, 144},
+                        {  1,  62, 107}, {  1,  41,  72}, {  1,  23,  41},
+                    },
+                    {  // band 5
+                        { 25, 233, 236}, {  5, 165, 214}, {  1,  96, 158},
+                        {  1,  63, 112}, {  1,  40,  73}, {  1,  23,  40},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 48,  20, 231}, { 37,  21, 179}, { 15,  18, 109},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 41, 154, 216}, {196, 142, 221}, {131, 125, 207},
+                        { 84, 111, 181}, { 45,  91, 142}, { 27,  62,  89},
+                    },
+                    {  // band 2
+                        { 72, 181, 230}, { 41, 147, 215}, { 10, 102, 173},
+                        {  3,  73, 132}, {  1,  47,  89}, {  1,  23,  50},
+                    },
+                    {  // band 3
+                        { 60, 201, 236}, { 23, 157, 219}, {  2,  99, 167},
+                        {  1,  69, 124}, {  1,  43,  80}, {  1,  22,  39},
+                    },
+                    {  // band 4
+                        { 53, 214, 242}, { 15, 165, 224}, {  1, 101, 173},
+                        {  1,  70, 131}, {  1,  44,  83}, {  1,  23,  49},
+                    },
+                    {  // band 5
+                        { 39, 239, 248}, {  7, 186, 233}, {  1, 108, 174},
+                        {  1,  70, 123}, {  1,  43,  77}, {  1,  16,  42},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {161,  26, 204}, { 77,  40, 160}, { 26,  50, 117},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 80, 140, 218}, {136, 133, 215}, { 63, 117, 197},
+                        { 20,  93, 170}, {  7,  55, 102}, { 13,  32,  52},
+                    },
+                    {  // band 2
+                        { 86, 173, 231}, { 46, 150, 220}, { 18, 118, 190},
+                        {  8,  90, 150}, {  2,  60,  95}, {  1,  39,  41},
+                    },
+                    {  // band 3
+                        { 80, 183, 242}, { 37, 160, 231}, {  6, 120, 182},
+                        {  1,  86, 137}, {  1,  46,  78}, {  1,  15,  24},
+                    },
+                    {  // band 4
+                        { 88, 215, 247}, { 42, 179, 235}, {  4, 116, 182},
+                        {  2,  80, 133}, {  1,  46,  85}, {  1,  64,  43},
+                    },
+                    {  // band 5
+                        {100, 236, 250}, { 31, 186, 234}, {  1, 114, 181},
+                        {  1,  85, 135}, {  1,  78,  64}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {213,  13, 245}, {106,  16, 211}, { 32,  11, 156},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {140, 214, 247}, {241, 186, 243}, {177, 172, 235},
+                        {128, 156, 219}, {106, 130, 191}, { 99, 105, 152},
+                    },
+                    {  // band 2
+                        {125, 218, 248}, { 75, 167, 239}, { 29, 111, 212},
+                        {  6,  66, 152}, {  1,  42,  96}, {  1,  85, 128},
+                    },
+                    {  // band 3
+                        {120, 232, 252}, { 60, 189, 247}, {  8, 141, 200},
+                        {  1,  89, 134}, {  1,  32, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {111, 238, 253}, { 56, 198, 245}, {  1, 123, 208},
+                        {  1,  93, 176}, {  1,   1,  73}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        { 98, 251, 249}, { 56, 189, 244}, { 17, 113, 220},
+                        {  1, 109, 179}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+#endif  // CONFIG_TX64X64
+    },
+    {  // Q_Index 3
+#if CONFIG_CB4X4
+        {  // TX_SIZE 0
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        {186,  16, 200}, {122,  31, 187}, { 78,  40, 161},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {146, 119, 245}, {182, 115, 244}, {130, 113, 238},
+                        { 88, 110, 225}, { 47, 103, 208}, {  5, 102, 188},
+                    },
+                    {  // band 2
+                        {164, 157, 248}, {155, 141, 250}, { 71, 116, 243},
+                        { 88, 129, 233}, { 50,  99, 228}, { 26, 148, 191},
+                    },
+                    {  // band 3
+                        {200, 158, 253}, {177, 118, 252}, { 99, 113, 245},
+                        { 77, 120, 210}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {231, 104, 254}, {209,  82, 254}, {143, 112, 252},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {250,  36, 254}, {243,  55, 254}, {223, 170, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        {207,  37, 226}, {164,  46, 218}, {122,  58, 201},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {149, 154, 253}, {170, 137, 253}, { 94, 123, 247},
+                        { 42, 113, 222}, { 16,  97, 174}, { 49,  98, 159},
+                    },
+                    {  // band 2
+                        {177, 162, 253}, {165, 142, 252}, { 51, 108, 243},
+                        { 18, 108, 213}, {  1,  98, 254}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {211, 152, 254}, {184, 116, 254}, { 70, 110, 244},
+                        {  8, 108, 237}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {236,  89, 254}, {210,  67, 254}, {112, 111, 248},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {246,  26, 254}, {233,  35, 254}, {128,   1, 254},
+                        {254, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {247,   2, 247}, {226,   8, 242}, {191,  14, 235},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {231,  94, 254}, {248,  91, 254}, {186,  89, 252},
+                        {128,  92, 244}, { 79, 112, 254}, {128, 128, 128},
+                    },
+                    {  // band 2
+                        {228, 145, 253}, {240, 130, 254}, {223, 105, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {245, 153, 253}, {240, 120, 254}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {254, 128, 254}, {204, 128, 254}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {253,   7, 249}, {224,   9, 244}, {182,  13, 231},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {234, 109, 254}, {242, 104, 254}, {160,  98, 254},
+                        {123,  85, 243}, { 82,  43, 217}, {128, 128, 128},
+                    },
+                    {  // band 2
+                        {243, 137, 254}, {240, 118, 254}, {136,  53, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {251, 173, 254}, {229, 129, 250}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {254, 119, 254}, {254, 128, 128}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+#endif
+        {  // TX_SIZE 0
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        {186,  16, 200}, {122,  31, 187}, { 78,  40, 161},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {146, 119, 245}, {182, 115, 244}, {130, 113, 238},
+                        { 88, 110, 225}, { 47, 103, 208}, {  5, 102, 188},
+                    },
+                    {  // band 2
+                        {164, 157, 248}, {155, 141, 250}, { 71, 116, 243},
+                        { 88, 129, 233}, { 50,  99, 228}, { 26, 148, 191},
+                    },
+                    {  // band 3
+                        {200, 158, 253}, {177, 118, 252}, { 99, 113, 245},
+                        { 77, 120, 210}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {231, 104, 254}, {209,  82, 254}, {143, 112, 252},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {250,  36, 254}, {243,  55, 254}, {223, 170, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        {207,  37, 226}, {164,  46, 218}, {122,  58, 201},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {149, 154, 253}, {170, 137, 253}, { 94, 123, 247},
+                        { 42, 113, 222}, { 16,  97, 174}, { 49,  98, 159},
+                    },
+                    {  // band 2
+                        {177, 162, 253}, {165, 142, 252}, { 51, 108, 243},
+                        { 18, 108, 213}, {  1,  98, 254}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {211, 152, 254}, {184, 116, 254}, { 70, 110, 244},
+                        {  8, 108, 237}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {236,  89, 254}, {210,  67, 254}, {112, 111, 248},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {246,  26, 254}, {233,  35, 254}, {128,   1, 254},
+                        {254, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {247,   2, 247}, {226,   8, 242}, {191,  14, 235},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {231,  94, 254}, {248,  91, 254}, {186,  89, 252},
+                        {128,  92, 244}, { 79, 112, 254}, {128, 128, 128},
+                    },
+                    {  // band 2
+                        {228, 145, 253}, {240, 130, 254}, {223, 105, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {245, 153, 253}, {240, 120, 254}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {254, 128, 254}, {204, 128, 254}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {253,   7, 249}, {224,   9, 244}, {182,  13, 231},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {234, 109, 254}, {242, 104, 254}, {160,  98, 254},
+                        {123,  85, 243}, { 82,  43, 217}, {128, 128, 128},
+                    },
+                    {  // band 2
+                        {243, 137, 254}, {240, 118, 254}, {136,  53, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {251, 173, 254}, {229, 129, 250}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {254, 119, 254}, {254, 128, 128}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 1
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 49,  26, 159}, { 36,  34, 150}, { 26,  38, 124},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 99, 122, 226}, {143, 119, 225}, { 90, 113, 213},
+                        { 46, 102, 193}, { 14,  84, 157}, {  3,  59, 107},
+                    },
+                    {  // band 2
+                        {109, 164, 237}, { 74, 142, 233}, { 29, 112, 216},
+                        { 14,  92, 184}, { 10,  80, 156}, {  1,  52, 137},
+                    },
+                    {  // band 3
+                        {110, 191, 245}, { 59, 156, 240}, { 18, 121, 220},
+                        {  8,  97, 184}, {  3,  84, 150}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {115, 203, 250}, { 59, 167, 246}, { 16, 130, 226},
+                        {  7,  97, 192}, {  1,  71,  99}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {149, 218, 253}, { 93, 171, 251}, { 28, 125, 233},
+                        { 28,  99, 192}, {128,  85,  85}, {128, 128, 128},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 97,  45, 229}, { 79,  52, 205}, { 46,  58, 171},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 99, 180, 249}, {156, 165, 249}, { 73, 141, 237},
+                        { 31, 116, 208}, { 13,  81, 153}, {  5,  42,  86},
+                    },
+                    {  // band 2
+                        {113, 188, 251}, { 68, 161, 244}, { 16, 108, 216},
+                        {  6,  81, 168}, {  2,  65, 118}, {128,   1,   1},
+                    },
+                    {  // band 3
+                        {117, 201, 252}, { 62, 171, 248}, { 12, 119, 221},
+                        {  5,  90, 182}, {  4,  66, 116}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {128, 207, 253}, { 70, 176, 251}, { 11, 126, 228},
+                        {  6,  89, 189}, {  1,  44, 148}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {162, 218, 254}, {107, 170, 253}, { 22, 131, 238},
+                        {  1,  77, 182}, {  1, 254, 128}, {128, 128, 128},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {235,   5, 238}, {194,  14, 223}, {152,  22, 205},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {200, 121, 251}, {241, 115, 252}, {167, 108, 248},
+                        { 93,  93, 233}, { 36,  66, 189}, {128, 128, 128},
+                    },
+                    {  // band 2
+                        {220, 151, 253}, {176, 135, 252}, { 95, 124, 254},
+                        { 64, 105, 217}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {225, 189, 254}, {175, 155, 254}, {102, 119, 254},
+                        {  1,   1,   1}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {218, 195, 254}, {125, 157, 253}, {128, 128, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {221, 197, 254}, { 85, 210, 254}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {250,   9, 246}, {204,  13, 234}, {144,  18, 211},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {213, 157, 253}, {243, 138, 253}, {170, 117, 250},
+                        {109,  91, 233}, { 66,  77, 163}, { 64,  85, 254},
+                    },
+                    {  // band 2
+                        {221, 169, 254}, {182, 141, 253}, {112, 120, 239},
+                        { 85, 165, 254}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {226, 192, 254}, {189, 174, 251}, {153, 128, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {232, 192, 254}, {195, 187, 247}, {  1, 191, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {247, 185, 254}, {254,  93, 254}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 2
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 14,  30, 136}, { 15,  33, 120}, { 10,  33,  90},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 92, 109, 209}, {113, 108, 207}, { 77, 102, 193},
+                        { 39,  91, 171}, { 11,  70, 129}, {  2,  44,  77},
+                    },
+                    {  // band 2
+                        { 99, 158, 223}, { 66, 135, 217}, { 23, 109, 194},
+                        {  9,  85, 160}, {  3,  66, 124}, {  1,  51, 100},
+                    },
+                    {  // band 3
+                        { 89, 189, 234}, { 46, 149, 225}, { 10, 110, 194},
+                        {  2,  83, 156}, {  1,  57, 113}, {  1,  47,  73},
+                    },
+                    {  // band 4
+                        { 78, 206, 242}, { 28, 161, 232}, {  3, 114, 200},
+                        {  1,  86, 161}, {  1,  62, 118}, {  1,   1,   1},
+                    },
+                    {  // band 5
+                        { 72, 227, 250}, { 20, 182, 242}, {  3, 126, 210},
+                        {  2,  91, 166}, {  1,  64, 126}, {128, 128, 128},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 23,  42, 227}, { 41,  43, 195}, { 25,  45, 146},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {100, 172, 245}, {165, 158, 246}, { 88, 137, 234},
+                        { 44, 116, 203}, { 18,  85, 149}, {  7,  56,  92},
+                    },
+                    {  // band 2
+                        {117, 188, 247}, { 70, 155, 239}, { 18, 105, 204},
+                        {  7,  78, 158}, {  2,  50, 111}, {  1,  38,  77},
+                    },
+                    {  // band 3
+                        {104, 207, 250}, { 54, 166, 241}, {  6, 110, 199},
+                        {  1,  78, 155}, {  1,  45, 100}, {  1,   1,   1},
+                    },
+                    {  // band 4
+                        { 87, 216, 251}, { 30, 177, 243}, {  1, 114, 203},
+                        {  1,  85, 157}, {  1,  53, 108}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        { 80, 230, 253}, { 23, 193, 248}, {  1, 127, 215},
+                        {  1,  94, 170}, {  1,  71,  59}, {128, 128, 128},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {222,   9, 234}, {161,  20, 210}, {113,  30, 185},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {195, 120, 248}, {231, 124, 247}, {148, 116, 238},
+                        { 64,  98, 207}, { 20,  70, 147}, { 87,  68, 100},
+                    },
+                    {  // band 2
+                        {186, 161, 250}, {124, 148, 245}, { 44, 123, 230},
+                        { 23, 107, 205}, {  1,  80, 131}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {172, 196, 252}, {110, 160, 248}, { 37, 134, 235},
+                        { 23, 125, 200}, {128, 254, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {173, 209, 253}, {103, 175, 250}, {  1, 120, 240},
+                        {  1, 146, 254}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {184, 235, 254}, { 81, 186, 251}, {128, 109, 254},
+                        {128, 254, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {248,   8, 243}, {185,  11, 225}, {108,  11, 189},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {208, 158, 254}, {244, 147, 252}, {195, 132, 248},
+                        {161, 122, 224}, {129, 114, 188}, { 59, 119, 159},
+                    },
+                    {  // band 2
+                        {202, 182, 253}, {143, 161, 251}, { 73, 115, 247},
+                        {146, 175, 204}, {128,   1, 254}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {202, 204, 254}, {131, 174, 251}, { 18, 153, 207},
+                        {128, 254, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {192, 221, 254}, {114, 190, 254}, {128, 170, 254},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {166, 236, 254}, {119, 200, 254}, {128, 128, 128},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+        {  // TX_SIZE 3
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 30,  32, 144}, { 21,  35,  96}, {  4,  27,  55},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 35, 107, 172}, { 61, 104, 170}, { 33,  94, 160},
+                        { 13,  80, 139}, {  2,  55,  97}, {  1,  28,  49},
+                    },
+                    {  // band 2
+                        { 51, 153, 195}, { 29, 129, 189}, {  9,  99, 163},
+                        {  3,  75, 129}, {  1,  49,  88}, {  1,  29,  50},
+                    },
+                    {  // band 3
+                        { 53, 164, 210}, { 21, 134, 201}, {  3,  97, 164},
+                        {  1,  69, 124}, {  1,  45,  82}, {  1,  31,  58},
+                    },
+                    {  // band 4
+                        { 47, 205, 234}, { 18, 158, 220}, {  2, 109, 177},
+                        {  1,  78, 137}, {  1,  53, 101}, {  1,  34,  70},
+                    },
+                    {  // band 5
+                        { 55, 233, 245}, { 16, 179, 233}, {  1, 116, 191},
+                        {  1,  79, 145}, {  1,  53, 101}, {  1,  37,  58},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 36,  33, 227}, { 39,  28, 190}, { 18,  27, 134},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 76, 156, 235}, {184, 147, 235}, {114, 130, 220},
+                        { 72, 112, 191}, { 42,  87, 144}, { 21,  65,  93},
+                    },
+                    {  // band 2
+                        { 96, 179, 240}, { 51, 149, 228}, { 12, 105, 191},
+                        {  6,  74, 148}, {  1,  47, 100}, {  1,  29,  53},
+                    },
+                    {  // band 3
+                        { 88, 191, 242}, { 35, 154, 231}, {  3, 106, 187},
+                        {  1,  74, 140}, {  1,  41,  84}, {  1,  25,  38},
+                    },
+                    {  // band 4
+                        { 77, 212, 249}, { 28, 171, 239}, {  2, 117, 199},
+                        {  1,  79, 151}, {  1,  45,  99}, {  1,   1,   1},
+                    },
+                    {  // band 5
+                        { 77, 236, 252}, { 27, 190, 246}, {  2, 120, 203},
+                        {  1,  78, 147}, {  1,  42,  72}, {128, 128, 128},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {185,  11, 227}, {113,  30, 182}, { 57,  44, 144},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {151, 139, 244}, {212, 139, 241}, {124, 126, 231},
+                        { 59, 104, 213}, { 26,  73, 158}, { 20,  45,  95},
+                    },
+                    {  // band 2
+                        {155, 163, 247}, {108, 152, 239}, { 39, 124, 214},
+                        {  7, 109, 162}, { 29,  57, 128}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {158, 176, 250}, { 89, 164, 243}, { 11, 114, 196},
+                        {  1,  96, 141}, {  1,  81, 118}, {128,   1,   1},
+                    },
+                    {  // band 4
+                        {148, 212, 251}, { 59, 174, 240}, {  2, 130, 203},
+                        {  1,  70, 168}, {  1,  51, 106}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {104, 237, 252}, { 39, 190, 246}, {  1, 154, 220},
+                        {128, 102,   1}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {236,   6, 242}, {111,   6, 206}, { 36,   5, 161},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {193, 193, 252}, {248, 182, 251}, {218, 150, 246},
+                        {182, 134, 244}, {151, 137, 227}, { 45, 102, 195},
+                    },
+                    {  // band 2
+                        {188, 202, 251}, {125, 165, 249}, { 64,  75, 218},
+                        {  1, 128, 254}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {178, 225, 254}, {107, 188, 231}, { 21, 135, 233},
+                        {128,   1, 254}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {164, 227, 253}, { 55, 193, 251}, {  1, 111, 225},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {151, 243, 254}, { 50, 203, 254}, {128, 179, 254},
+                        {128,   1, 254}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+#if CONFIG_TX64X64
+        {  // TX_SIZE 4
+            {  // Y plane
+                {  // Intra
+                    {  // band 0
+                        { 30,  32, 144}, { 21,  35,  96}, {  4,  27,  55},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 35, 107, 172}, { 61, 104, 170}, { 33,  94, 160},
+                        { 13,  80, 139}, {  2,  55,  97}, {  1,  28,  49},
+                    },
+                    {  // band 2
+                        { 51, 153, 195}, { 29, 129, 189}, {  9,  99, 163},
+                        {  3,  75, 129}, {  1,  49,  88}, {  1,  29,  50},
+                    },
+                    {  // band 3
+                        { 53, 164, 210}, { 21, 134, 201}, {  3,  97, 164},
+                        {  1,  69, 124}, {  1,  45,  82}, {  1,  31,  58},
+                    },
+                    {  // band 4
+                        { 47, 205, 234}, { 18, 158, 220}, {  2, 109, 177},
+                        {  1,  78, 137}, {  1,  53, 101}, {  1,  34,  70},
+                    },
+                    {  // band 5
+                        { 55, 233, 245}, { 16, 179, 233}, {  1, 116, 191},
+                        {  1,  79, 145}, {  1,  53, 101}, {  1,  37,  58},
+                    },
+                },
+                {  // Intra
+                    {  // band 0
+                        { 36,  33, 227}, { 39,  28, 190}, { 18,  27, 134},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        { 76, 156, 235}, {184, 147, 235}, {114, 130, 220},
+                        { 72, 112, 191}, { 42,  87, 144}, { 21,  65,  93},
+                    },
+                    {  // band 2
+                        { 96, 179, 240}, { 51, 149, 228}, { 12, 105, 191},
+                        {  6,  74, 148}, {  1,  47, 100}, {  1,  29,  53},
+                    },
+                    {  // band 3
+                        { 88, 191, 242}, { 35, 154, 231}, {  3, 106, 187},
+                        {  1,  74, 140}, {  1,  41,  84}, {  1,  25,  38},
+                    },
+                    {  // band 4
+                        { 77, 212, 249}, { 28, 171, 239}, {  2, 117, 199},
+                        {  1,  79, 151}, {  1,  45,  99}, {  1,   1,   1},
+                    },
+                    {  // band 5
+                        { 77, 236, 252}, { 27, 190, 246}, {  2, 120, 203},
+                        {  1,  78, 147}, {  1,  42,  72}, {128, 128, 128},
+                    },
+                },
+            },
+            {  // UV plane
+                {  // Inter
+                    {  // band 0
+                        {185,  11, 227}, {113,  30, 182}, { 57,  44, 144},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {151, 139, 244}, {212, 139, 241}, {124, 126, 231},
+                        { 59, 104, 213}, { 26,  73, 158}, { 20,  45,  95},
+                    },
+                    {  // band 2
+                        {155, 163, 247}, {108, 152, 239}, { 39, 124, 214},
+                        {  7, 109, 162}, { 29,  57, 128}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {158, 176, 250}, { 89, 164, 243}, { 11, 114, 196},
+                        {  1,  96, 141}, {  1,  81, 118}, {128,   1,   1},
+                    },
+                    {  // band 4
+                        {148, 212, 251}, { 59, 174, 240}, {  2, 130, 203},
+                        {  1,  70, 168}, {  1,  51, 106}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {104, 237, 252}, { 39, 190, 246}, {  1, 154, 220},
+                        {128, 102,   1}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+                {  // Inter
+                    {  // band 0
+                        {236,   6, 242}, {111,   6, 206}, { 36,   5, 161},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 1
+                        {193, 193, 252}, {248, 182, 251}, {218, 150, 246},
+                        {182, 134, 244}, {151, 137, 227}, { 45, 102, 195},
+                    },
+                    {  // band 2
+                        {188, 202, 251}, {125, 165, 249}, { 64,  75, 218},
+                        {  1, 128, 254}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 3
+                        {178, 225, 254}, {107, 188, 231}, { 21, 135, 233},
+                        {128,   1, 254}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 4
+                        {164, 227, 253}, { 55, 193, 251}, {  1, 111, 225},
+                        {128, 128, 128}, {128, 128, 128}, {128, 128, 128},
+                    },
+                    {  // band 5
+                        {151, 243, 254}, { 50, 203, 254}, {128, 179, 254},
+                        {128,   1, 254}, {128, 128, 128}, {128, 128, 128},
+                    },
+                },
+            },
+        },
+#endif  // CONFIG_TX64X64
+    },
+};
+#else
+#if CONFIG_NEW_TOKENSET
+static const av1_coeff_probs_model default_coef_probs_4x4[PLANE_TYPES] = {
+  { // Y plane
+    { // Intra
+      { // Band 0
+        {97, 27, 144}, {81, 38, 128}, {51, 42, 99}
+      },
+      { // Band 1
+        {74, 113, 204}, {68, 101, 199}, {50, 87, 173},
+        {31, 76, 133}, {13, 55, 86}, {3, 30, 39}
+      },
+      { // Band 2
+        {83, 156, 222}, {74, 127, 215}, {46, 101, 179},
+        {30, 80, 129}, {14, 57, 81}, {3, 27, 37}
+      },
+      { // Band 3
+        {105, 164, 233}, {84, 128, 224}, {49, 92, 175},
+        {28, 60, 114}, {12, 34, 53}, {20, 59, 98}
+      },
+      { // Band 4
+        {131, 159, 243}, {98, 123, 228}, {40, 78, 151},
+        {19, 46, 97}, {13, 47, 19}, {19, 16, 19}
+      },
+      { // Band 5
+        {192, 71, 241}, {174, 70, 226}, {125, 46, 153},
+        {108, 49, 116}, {82, 24, 46}, {60, 14, 30}
+      }
+    },
+    { // Inter
+      { // Band 0
+        {111, 66, 218}, {107, 87, 211}, {93, 99, 207}
+      },
+      { // Band 1
+        {107, 166, 250}, {107, 143, 247}, {73, 119, 221},
+        {43, 91, 166}, {17, 74, 102}, {3, 70, 53}
+      },
+      { // Band 2
+        {126, 177, 251}, {109, 148, 246}, {64, 99, 204},
+        {42, 68, 140}, {28, 52, 84}, {20, 34, 1}
+      },
+      { // Band 3
+        {143, 178, 252}, {114, 144, 245}, {46, 92, 188},
+        {45, 65, 104}, {40, 44, 76}, {1, 1, 1}
+      },
+      { // Band 4
+        {163, 159, 251}, {120, 131, 243}, {47, 81, 182},
+        {32, 39, 128}, {33, 44, 56}, {1, 17, 34}
+      },
+      { // Band 5
+        {209, 94, 251}, {190, 81, 241}, {139, 45, 147},
+        {123, 35, 73}, {118, 1, 118}, {3, 16, 42}
+      }
+    }
+  },
+  { // UV plane
+    { // Intra
+      { // Band 0
+        {189, 37, 229}, {145, 68, 205}, {99, 74, 171}
+      },
+      { // Band 1
+        {153, 139, 242}, {135, 125, 235}, {84, 100, 200},
+        {49, 75, 162}, {9, 21, 84}, {3, 31, 69}
+      },
+      { // Band 2
+        {165, 165, 244}, {128, 144, 240}, {68, 94, 204},
+        {39, 72, 132}, {22, 44, 93}, {26, 73, 26}
+      },
+      { // Band 3
+        {181, 174, 246}, {142, 132, 241}, {81, 96, 212},
+        {41, 70, 166}, {9, 48, 92}, {1, 19, 38}
+      },
+      { // Band 4
+        {197, 159, 251}, {168, 121, 245}, {107, 75, 218},
+        {70, 43, 158}, {1, 128, 1}, {1, 18, 37}
+      },
+      { // Band 5
+        {231, 79, 255}, {211, 74, 249}, {157, 104, 210},
+        {128, 102, 213}, {12, 34, 96}, {2, 20, 47}
+      }
+    },
+    { // Inter
+      { // Band 0
+        {220, 53, 252}, {191, 80, 248}, {154, 100, 245}
+      },
+      { // Band 1
+        {205, 153, 255}, {182, 147, 254}, {110, 131, 231},
+        {68, 114, 161}, {50, 114, 140}, {1, 33, 57}
+      },
+      { // Band 2
+        {213, 171, 255}, {184, 163, 254}, {116, 104, 235},
+        {79, 71, 207}, {1, 41, 79}, {1, 20, 39}
+      },
+      { // Band 3
+        {223, 158, 255}, {203, 137, 255}, {111, 142, 244},
+        {2, 255, 133}, {1, 44, 85}, {1, 22, 47}
+      },
+      { // Band 4
+        {232, 148, 255}, {222, 123, 255}, {255, 128, 255},
+        {3, 61, 124}, {1, 41, 84}, {1, 21, 52}
+      },
+      { // Band 5
+        {248, 92, 255}, {248, 96, 255}, {69, 58, 184},
+        {31, 44, 137}, {14, 38, 105}, {8, 23, 61}
+      }
+    }
+  }
+};
+static const av1_coeff_probs_model default_coef_probs_8x8[PLANE_TYPES] = {
+  { // Y plane
+    { // Intra
+      { // Band 0
+        {112, 31, 159}, {72, 37, 119}, {22, 35, 68}
+      },
+      { // Band 1
+        {42, 109, 174}, {45, 99, 172}, {32, 84, 149},
+        {18, 69, 119}, {6, 46, 76}, {1, 19, 31}
+      },
+      { // Band 2
+        {40, 154, 202}, {35, 126, 191}, {19, 98, 160},
+        {10, 75, 122}, {5, 53, 82}, {1, 23, 39}
+      },
+      { // Band 3
+        {39, 176, 215}, {28, 135, 200}, {11, 93, 156},
+        {5, 63, 109}, {1, 36, 64}, {1, 14, 26}
+      },
+      { // Band 4
+        {41, 191, 230}, {25, 147, 212}, {9, 97, 160},
+        {3, 65, 109}, {1, 33, 58}, {1, 14, 20}
+      },
+      { // Band 5
+        {68, 203, 242}, {40, 159, 220}, {12, 97, 153},
+        {5, 58, 97}, {1, 29, 55}, {1, 11, 18}
+      }
+    },
+    { // Inter
+      { // Band 0
+        {99, 67, 221}, {86, 80, 204}, {60, 87, 184}
+      },
+      { // Band 1
+        {73, 169, 246}, {79, 158, 242}, {50, 135, 220},
+        {30, 113, 181}, {18, 76, 126}, {5, 54, 85}
+      },
+      { // Band 2
+        {90, 184, 250}, {78, 162, 243}, {47, 118, 214},
+        {35, 85, 171}, {32, 53, 115}, {20, 28, 76}
+      },
+      { // Band 3
+        {109, 197, 252}, {89, 172, 247}, {52, 119, 217},
+        {37, 80, 161}, {23, 44, 100}, {1, 18, 34}
+      },
+      { // Band 4
+        {132, 202, 254}, {110, 175, 251}, {63, 128, 228},
+        {37, 86, 168}, {64, 91, 102}, {1, 17, 34}
+      },
+      { // Band 5
+        {126, 204, 253}, {100, 174, 250}, {50, 148, 237},
+        {25, 90, 133}, {1, 64, 85}, {3, 16, 42}
+      }
+    }
+  },
+  { // UV plane
+    { // Intra
+      { // Band 0
+        {195, 35, 235}, {137, 63, 201}, {62, 70, 145}
+      },
+      { // Band 1
+        {110, 158, 233}, {102, 143, 227}, {60, 120, 199},
+        {30, 85, 156}, {9, 50, 90}, {1, 16, 33}
+      },
+      { // Band 2
+        {102, 185, 233}, {71, 152, 224}, {29, 111, 187},
+        {18, 74, 138}, {4, 56, 87}, {1, 18, 46}
+      },
+      { // Band 3
+        {101, 205, 239}, {66, 161, 229}, {23, 109, 183},
+        {9, 85, 135}, {5, 71, 142}, {1, 1, 102}
+      },
+      { // Band 4
+        {109, 216, 243}, {69, 168, 233}, {23, 119, 191},
+        {8, 137, 115}, {1, 54, 98}, {1, 1, 255}
+      },
+      { // Band 5
+        {139, 224, 249}, {98, 176, 238}, {55, 129, 187},
+        {25, 101, 131}, {26, 59, 154}, {2, 20, 47}
+      }
+    },
+    { // Inter
+      { // Band 0
+        {220, 72, 254}, {176, 108, 251}, {114, 132, 247}
+      },
+      { // Band 1
+        {161, 185, 255}, {141, 185, 254}, {131, 180, 249},
+        {111, 164, 186}, {50, 98, 142}, {1, 128, 1}
+      },
+      { // Band 2
+        {171, 195, 255}, {133, 184, 254}, {68, 140, 231},
+        {102, 96, 205}, {1, 1, 128}, {1, 20, 39}
+      },
+      { // Band 3
+        {180, 206, 255}, {148, 191, 254}, {83, 157, 241},
+        {128, 171, 128}, {1, 44, 85}, {1, 22, 47}
+      },
+      { // Band 4
+        {194, 214, 255}, {159, 188, 255}, {122, 148, 250},
+        {3, 255, 124}, {1, 41, 84}, {1, 21, 52}
+      },
+      { // Band 5
+        {231, 217, 255}, {209, 149, 255}, {205, 145, 205},
+        {31, 44, 137}, {14, 38, 105}, {8, 23, 61}
+      }
+    }
+  }
+};
+static const av1_coeff_probs_model default_coef_probs_16x16[PLANE_TYPES] = {
+  { // Y plane
+    { // Intra
+      { // Band 0
+        {91, 31, 117}, {49, 31, 89}, {14, 25, 48}
+      },
+      { // Band 1
+        {31, 97, 151}, {33, 89, 148}, {28, 76, 133},
+        {17, 60, 106}, {7, 42, 72}, {1, 19, 32}
+      },
+      { // Band 2
+        {28, 152, 182}, {28, 120, 174}, {15, 93, 146},
+        {9, 72, 116}, {5, 47, 82}, {1, 21, 37}
+      },
+      { // Band 3
+        {29, 174, 203}, {23, 127, 187}, {9, 89, 145},
+        {2, 56, 100}, {1, 31, 56}, {1, 12, 25}
+      },
+      { // Band 4
+        {28, 193, 220}, {17, 141, 197}, {4, 87, 142},
+        {1, 54, 95}, {1, 31, 56}, {1, 12, 26}
+      },
+      { // Band 5
+        {29, 221, 240}, {11, 167, 215}, {2, 93, 149},
+        {1, 58, 100}, {1, 35, 61}, {1, 16, 28}
+      }
+    },
+    { // Inter
+      { // Band 0
+        {108, 52, 214}, {84, 60, 186}, {45, 69, 161}
+      },
+      { // Band 1
+        {43, 164, 236}, {57, 161, 233}, {38, 146, 214},
+        {24, 120, 182}, {15, 80, 126}, {5, 28, 66}
+      },
+      { // Band 2
+        {58, 187, 242}, {47, 163, 234}, {28, 118, 204},
+        {26, 82, 165}, {21, 54, 112}, {4, 28, 55}
+      },
+      { // Band 3
+        {65, 201, 248}, {51, 170, 239}, {22, 117, 204},
+        {11, 81, 159}, {10, 43, 102}, {1, 1, 1}
+      },
+      { // Band 4
+        {80, 206, 252}, {57, 179, 245}, {25, 129, 214},
+        {16, 97, 170}, {6, 60, 130}, {1, 128, 1}
+      },
+      { // Band 5
+        {97, 217, 253}, {68, 186, 250}, {26, 138, 216},
+        {20, 105, 166}, {11, 78, 111}, {3, 16, 42}
+      }
+    }
+  },
+  { // UV plane
+    { // Intra
+      { // Band 0
+        {181, 37, 233}, {121, 55, 192}, {46, 52, 124}
+      },
+      { // Band 1
+        {108, 157, 221}, {98, 140, 215}, {59, 124, 187},
+        {34, 92, 158}, {9, 68, 112}, {1, 41, 70}
+      },
+      { // Band 2
+        {80, 188, 223}, {46, 153, 204}, {25, 91, 173},
+        {11, 73, 131}, {5, 43, 82}, {1, 17, 91}
+      },
+      { // Band 3
+        {63, 209, 228}, {31, 157, 206}, {8, 104, 167},
+        {3, 63, 122}, {1, 44, 87}, {1, 43, 51}
+      },
+      { // Band 4
+        {52, 220, 234}, {22, 165, 216}, {4, 104, 163},
+        {2, 62, 129}, {1, 33, 50}, {1, 26, 28}
+      },
+      { // Band 5
+        {58, 238, 242}, {24, 183, 224}, {4, 109, 172},
+        {2, 87, 141}, {1, 52, 79}, {1, 51, 64}
+      }
+    },
+    { // Inter
+      { // Band 0
+        {224, 52, 250}, {188, 81, 239}, {138, 114, 228}
+      },
+      { // Band 1
+        {131, 206, 255}, {128, 193, 254}, {119, 173, 247},
+        {106, 127, 187}, {50, 100, 124}, {1, 96, 1}
+      },
+      { // Band 2
+        {123, 214, 254}, {86, 194, 254}, {64, 119, 221},
+        {43, 51, 128}, {1, 32, 110}, {1, 20, 39}
+      },
+      { // Band 3
+        {115, 223, 255}, {78, 200, 254}, {75, 164, 203},
+        {128, 85, 255}, {1, 44, 85}, {1, 22, 47}
+      },
+      { // Band 4
+        {132, 226, 255}, {88, 207, 254}, {20, 140, 225},
+        {3, 61, 124}, {1, 41, 84}, {1, 21, 52}
+      },
+      { // Band 5
+        {180, 236, 255}, {138, 223, 254}, {73, 166, 238},
+        {31, 255, 137}, {14, 38, 105}, {8, 23, 61}
+      }
+    }
+  }
+};
+static const av1_coeff_probs_model default_coef_probs_32x32[PLANE_TYPES] = {
+  { // Y plane
+    { // Intra
+      { // Band 0
+        {163, 26, 188}, {78, 29, 105}, {22, 22, 48}
+      },
+      { // Band 1
+        {72, 93, 168}, {74, 91, 170}, {62, 72, 151},
+        {37, 55, 112}, {10, 33, 63}, {1, 14, 23}
+      },
+      { // Band 2
+        {41, 163, 182}, {36, 136, 177}, {20, 102, 153},
+        {10, 76, 114}, {5, 45, 71}, {1, 17, 27}
+      },
+      { // Band 3
+        {43, 202, 213}, {28, 142, 193}, {10, 90, 141},
+        {2, 51, 93}, {1, 24, 48}, {1, 10, 19}
+      },
+      { // Band 4
+        {46, 216, 220}, {26, 150, 199}, {7, 87, 136},
+        {2, 49, 86}, {1, 28, 47}, {1, 12, 24}
+      },
+      { // Band 5
+        {19, 241, 237}, {5, 172, 200}, {1, 82, 126},
+        {1, 47, 79}, {1, 29, 47}, {1, 14, 25}
+      }
+    },
+    { // Inter
+      { // Band 0
+        {185, 20, 226}, {151, 26, 187}, {109, 34, 144}
+      },
+      { // Band 1
+        {56, 151, 227}, {76, 165, 232}, {62, 161, 222},
+        {47, 139, 201}, {29, 106, 150}, {14, 61, 98}
+      },
+      { // Band 2
+        {57, 200, 237}, {43, 164, 227}, {22, 106, 190},
+        {14, 68, 140}, {10, 48, 90}, {1, 15, 40}
+      },
+      { // Band 3
+        {46, 209, 238}, {28, 165, 225}, {7, 107, 180},
+        {2, 69, 125}, {2, 36, 94}, {1, 1, 1}
+      },
+      { // Band 4
+        {55, 225, 248}, {28, 181, 237}, {7, 117, 198},
+        {6, 77, 144}, {3, 60, 90}, {1, 1, 1}
+      },
+      { // Band 5
+        {63, 243, 251}, {27, 193, 242}, {4, 124, 200},
+        {1, 58, 153}, {1, 59, 124}, {3, 16, 42}
+      }
+    }
+  },
+  { // UV plane
+    { // Intra
+      { // Band 0
+        {208, 28, 218}, {183, 32, 188}, {169, 21, 189}
+      },
+      { // Band 1
+        {205, 124, 247}, {190, 96, 240}, {233, 89, 233},
+        {177, 44, 212}, {59, 58, 59}, {32, 33, 38}
+      },
+      { // Band 2
+        {194, 195, 250}, {179, 190, 226}, {32, 174, 128},
+        {32, 85, 128}, {12, 64, 122}, {1, 85, 90}
+      },
+      { // Band 3
+        {149, 232, 249}, {95, 159, 227}, {28, 91, 171},
+        {28, 102, 114}, {1, 1, 73}, {1, 19, 38}
+      },
+      { // Band 4
+        {154, 239, 246}, {138, 151, 235}, {1, 123, 138},
+        {128, 183, 255}, {1, 128, 1}, {1, 18, 37}
+      },
+      { // Band 5
+        {157, 255, 253}, {75, 171, 241}, {43, 102, 171},
+        {30, 44, 136}, {12, 34, 96}, {2, 20, 47}
+      }
+    },
+    { // Inter
+      { // Band 0
+        {249, 13, 248}, {238, 14, 220}, {225, 16, 174}
+      },
+      { // Band 1
+        {190, 189, 254}, {169, 134, 253}, {124, 179, 248},
+        {138, 131, 223}, {64, 133, 192}, {1, 85, 128}
+      },
+      { // Band 2
+        {139, 212, 254}, {126, 177, 255}, {93, 39, 186},
+        {1, 1, 171}, {1, 41, 79}, {1, 20, 39}
+      },
+      { // Band 3
+        {153, 216, 255}, {165, 204, 255}, {1, 1, 255},
+        {2, 73, 133}, {1, 1, 1}, {1, 22, 47}
+      },
+      { // Band 4
+        {147, 226, 254}, {119, 196, 255}, {1, 128, 255},
+        {1, 1, 171}, {1, 1, 1}, {1, 21, 52}
+      },
+      { // Band 5
+        {168, 240, 255}, {95, 179, 255}, {1, 171, 1},
+        {31, 44, 137}, {14, 38, 105}, {8, 23, 61}
+      }
+    }
+  }
+};
+#else  // CONFIG_NEW_TOKENSET
+static const av1_coeff_probs_model default_coef_probs_4x4[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        { 195,  29, 183 }, {  84,  49, 136 }, {   8,  42,  71 }
+      }, {  // Band 1
+        {  31, 107, 169 }, {  35,  99, 159 }, {  17,  82, 140 },
+        {   8,  66, 114 }, {   2,  44,  76 }, {   1,  19,  32 }
+      }, {  // Band 2
+        {  40, 132, 201 }, {  29, 114, 187 }, {  13,  91, 157 },
+        {   7,  75, 127 }, {   3,  58,  95 }, {   1,  28,  47 }
+      }, {  // Band 3
+        {  69, 142, 221 }, {  42, 122, 201 }, {  15,  91, 159 },
+        {   6,  67, 121 }, {   1,  42,  77 }, {   1,  17,  31 }
+      }, {  // Band 4
+        { 102, 148, 228 }, {  67, 117, 204 }, {  17,  82, 154 },
+        {   6,  59, 114 }, {   2,  39,  75 }, {   1,  15,  29 }
+      }, {  // Band 5
+        { 156,  57, 233 }, { 119,  57, 212 }, {  58,  48, 163 },
+        {  29,  40, 124 }, {  12,  30,  81 }, {   3,  12,  31 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 191, 107, 226 }, { 124, 117, 204 }, {  25,  99, 155 }
+      }, {  // Band 1
+        {  29, 148, 210 }, {  37, 126, 194 }, {   8,  93, 157 },
+        {   2,  68, 118 }, {   1,  39,  69 }, {   1,  17,  33 }
+      }, {  // Band 2
+        {  41, 151, 213 }, {  27, 123, 193 }, {   3,  82, 144 },
+        {   1,  58, 105 }, {   1,  32,  60 }, {   1,  13,  26 }
+      }, {  // Band 3
+        {  59, 159, 220 }, {  23, 126, 198 }, {   4,  88, 151 },
+        {   1,  66, 114 }, {   1,  38,  71 }, {   1,  18,  34 }
+      }, {  // Band 4
+        { 114, 136, 232 }, {  51, 114, 207 }, {  11,  83, 155 },
+        {   3,  56, 105 }, {   1,  33,  65 }, {   1,  17,  34 }
+      }, {  // Band 5
+        { 149,  65, 234 }, { 121,  57, 215 }, {  61,  49, 166 },
+        {  28,  36, 114 }, {  12,  25,  76 }, {   3,  16,  42 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 214,  49, 220 }, { 132,  63, 188 }, {  42,  65, 137 }
+      }, {  // Band 1
+        {  85, 137, 221 }, { 104, 131, 216 }, {  49, 111, 192 },
+        {  21,  87, 155 }, {   2,  49,  87 }, {   1,  16,  28 }
+      }, {  // Band 2
+        {  89, 163, 230 }, {  90, 137, 220 }, {  29, 100, 183 },
+        {  10,  70, 135 }, {   2,  42,  81 }, {   1,  17,  33 }
+      }, {  // Band 3
+        { 108, 167, 237 }, {  55, 133, 222 }, {  15,  97, 179 },
+        {   4,  72, 135 }, {   1,  45,  85 }, {   1,  19,  38 }
+      }, {  // Band 4
+        { 124, 146, 240 }, {  66, 124, 224 }, {  17,  88, 175 },
+        {   4,  58, 122 }, {   1,  36,  75 }, {   1,  18,  37 }
+      }, {  //  Band 5
+        { 141,  79, 241 }, { 126,  70, 227 }, {  66,  58, 182 },
+        {  30,  44, 136 }, {  12,  34,  96 }, {   2,  20,  47 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 229,  99, 249 }, { 143, 111, 235 }, {  46, 109, 192 }
+      }, {  // Band 1
+        {  82, 158, 236 }, {  94, 146, 224 }, {  25, 117, 191 },
+        {   9,  87, 149 }, {   3,  56,  99 }, {   1,  33,  57 }
+      }, {  // Band 2
+        {  83, 167, 237 }, {  68, 145, 222 }, {  10, 103, 177 },
+        {   2,  72, 131 }, {   1,  41,  79 }, {   1,  20,  39 }
+      }, {  // Band 3
+        {  99, 167, 239 }, {  47, 141, 224 }, {  10, 104, 178 },
+        {   2,  73, 133 }, {   1,  44,  85 }, {   1,  22,  47 }
+      }, {  // Band 4
+        { 127, 145, 243 }, {  71, 129, 228 }, {  17,  93, 177 },
+        {   3,  61, 124 }, {   1,  41,  84 }, {   1,  21,  52 }
+      }, {  // Band 5
+        { 157,  78, 244 }, { 140,  72, 231 }, {  69,  58, 184 },
+        {  31,  44, 137 }, {  14,  38, 105 }, {   8,  23,  61 }
+      }
+    }
+  }
+};
+
+static const av1_coeff_probs_model default_coef_probs_8x8[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        { 125,  34, 187 }, {  52,  41, 133 }, {   6,  31,  56 }
+      }, {  // Band 1
+        {  37, 109, 153 }, {  51, 102, 147 }, {  23,  87, 128 },
+        {   8,  67, 101 }, {   1,  41,  63 }, {   1,  19,  29 }
+      }, {  // Band 2
+        {  31, 154, 185 }, {  17, 127, 175 }, {   6,  96, 145 },
+        {   2,  73, 114 }, {   1,  51,  82 }, {   1,  28,  45 }
+      }, {  // Band 3
+        {  23, 163, 200 }, {  10, 131, 185 }, {   2,  93, 148 },
+        {   1,  67, 111 }, {   1,  41,  69 }, {   1,  14,  24 }
+      }, {  // Band 4
+        {  29, 176, 217 }, {  12, 145, 201 }, {   3, 101, 156 },
+        {   1,  69, 111 }, {   1,  39,  63 }, {   1,  14,  23 }
+      }, {  // Band 5
+        {  57, 192, 233 }, {  25, 154, 215 }, {   6, 109, 167 },
+        {   3,  78, 118 }, {   1,  48,  69 }, {   1,  21,  29 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 202, 105, 245 }, { 108, 106, 216 }, {  18,  90, 144 }
+      }, {  // Band 1
+        {  33, 172, 219 }, {  64, 149, 206 }, {  14, 117, 177 },
+        {   5,  90, 141 }, {   2,  61,  95 }, {   1,  37,  57 }
+      }, {  // Band 2
+        {  33, 179, 220 }, {  11, 140, 198 }, {   1,  89, 148 },
+        {   1,  60, 104 }, {   1,  33,  57 }, {   1,  12,  21 }
+      }, {  // Band 3
+        {  30, 181, 221 }, {   8, 141, 198 }, {   1,  87, 145 },
+        {   1,  58, 100 }, {   1,  31,  55 }, {   1,  12,  20 }
+      }, {  // Band 4
+        {  32, 186, 224 }, {   7, 142, 198 }, {   1,  86, 143 },
+        {   1,  58, 100 }, {   1,  31,  55 }, {   1,  12,  22 }
+      }, {  // Band 5
+        {  57, 192, 227 }, {  20, 143, 204 }, {   3,  96, 154 },
+        {   1,  68, 112 }, {   1,  42,  69 }, {   1,  19,  32 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 212,  35, 215 }, { 113,  47, 169 }, {  29,  48, 105 }
+      }, {  // Band 1
+        {  74, 129, 203 }, { 106, 120, 203 }, {  49, 107, 178 },
+        {  19,  84, 144 }, {   4,  50,  84 }, {   1,  15,  25 }
+      }, {  // Band 2
+        {  71, 172, 217 }, {  44, 141, 209 }, {  15, 102, 173 },
+        {   6,  76, 133 }, {   2,  51,  89 }, {   1,  24,  42 }
+      }, {  // Band 3
+        {  64, 185, 231 }, {  31, 148, 216 }, {   8, 103, 175 },
+        {   3,  74, 131 }, {   1,  46,  81 }, {   1,  18,  30 }
+      }, {  // Band 4
+        {  65, 196, 235 }, {  25, 157, 221 }, {   5, 105, 174 },
+        {   1,  67, 120 }, {   1,  38,  69 }, {   1,  15,  30 }
+      }, {  // Band 5
+        {  65, 204, 238 }, {  30, 156, 224 }, {   7, 107, 177 },
+        {   2,  70, 124 }, {   1,  42,  73 }, {   1,  18,  34 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 225,  86, 251 }, { 144, 104, 235 }, {  42,  99, 181 }
+      }, {  // Band 1
+        {  85, 175, 239 }, { 112, 165, 229 }, {  29, 136, 200 },
+        {  12, 103, 162 }, {   6,  77, 123 }, {   2,  53,  84 }
+      }, {  // Band 2
+        {  75, 183, 239 }, {  30, 155, 221 }, {   3, 106, 171 },
+        {   1,  74, 128 }, {   1,  44,  76 }, {   1,  17,  28 }
+      }, {  // Band 3
+        {  73, 185, 240 }, {  27, 159, 222 }, {   2, 107, 172 },
+        {   1,  75, 127 }, {   1,  42,  73 }, {   1,  17,  29 }
+      }, {  // Band 4
+        {  62, 190, 238 }, {  21, 159, 222 }, {   2, 107, 172 },
+        {   1,  72, 122 }, {   1,  40,  71 }, {   1,  18,  32 }
+      }, {  // Band 5
+        {  61, 199, 240 }, {  27, 161, 226 }, {   4, 113, 180 },
+        {   1,  76, 129 }, {   1,  46,  80 }, {   1,  23,  41 }
+      }
+    }
+  }
+};
+
+static const av1_coeff_probs_model default_coef_probs_16x16[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        {   7,  27, 153 }, {   5,  30,  95 }, {   1,  16,  30 }
+      }, {  // Band 1
+        {  50,  75, 127 }, {  57,  75, 124 }, {  27,  67, 108 },
+        {  10,  54,  86 }, {   1,  33,  52 }, {   1,  12,  18 }
+      }, {  // Band 2
+        {  43, 125, 151 }, {  26, 108, 148 }, {   7,  83, 122 },
+        {   2,  59,  89 }, {   1,  38,  60 }, {   1,  17,  27 }
+      }, {  // Band 3
+        {  23, 144, 163 }, {  13, 112, 154 }, {   2,  75, 117 },
+        {   1,  50,  81 }, {   1,  31,  51 }, {   1,  14,  23 }
+      }, {  // Band 4
+        {  18, 162, 185 }, {   6, 123, 171 }, {   1,  78, 125 },
+        {   1,  51,  86 }, {   1,  31,  54 }, {   1,  14,  23 }
+      }, {  // Band 5
+        {  15, 199, 227 }, {   3, 150, 204 }, {   1,  91, 146 },
+        {   1,  55,  95 }, {   1,  30,  53 }, {   1,  11,  20 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        {  19,  55, 240 }, {  19,  59, 196 }, {   3,  52, 105 }
+      }, {  // Band 1
+        {  41, 166, 207 }, { 104, 153, 199 }, {  31, 123, 181 },
+        {  14, 101, 152 }, {   5,  72, 106 }, {   1,  36,  52 }
+      }, {  // Band 2
+        {  35, 176, 211 }, {  12, 131, 190 }, {   2,  88, 144 },
+        {   1,  60, 101 }, {   1,  36,  60 }, {   1,  16,  28 }
+      }, {  // Band 3
+        {  28, 183, 213 }, {   8, 134, 191 }, {   1,  86, 142 },
+        {   1,  56,  96 }, {   1,  30,  53 }, {   1,  12,  20 }
+      }, {  // Band 4
+        {  20, 190, 215 }, {   4, 135, 192 }, {   1,  84, 139 },
+        {   1,  53,  91 }, {   1,  28,  49 }, {   1,  11,  20 }
+      }, {  // Band 5
+        {  13, 196, 216 }, {   2, 137, 192 }, {   1,  86, 143 },
+        {   1,  57,  99 }, {   1,  32,  56 }, {   1,  13,  24 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 211,  29, 217 }, {  96,  47, 156 }, {  22,  43,  87 }
+      }, {  // Band 1
+        {  78, 120, 193 }, { 111, 116, 186 }, {  46, 102, 164 },
+        {  15,  80, 128 }, {   2,  49,  76 }, {   1,  18,  28 }
+      }, {  // Band 2
+        {  71, 161, 203 }, {  42, 132, 192 }, {  10,  98, 150 },
+        {   3,  69, 109 }, {   1,  44,  70 }, {   1,  18,  29 }
+      }, {  // Band 3
+        {  57, 186, 211 }, {  30, 140, 196 }, {   4,  93, 146 },
+        {   1,  62, 102 }, {   1,  38,  65 }, {   1,  16,  27 }
+      }, {  // Band 4
+        {  47, 199, 217 }, {  14, 145, 196 }, {   1,  88, 142 },
+        {   1,  57,  98 }, {   1,  36,  62 }, {   1,  15,  26 }
+      }, {  // Band 5
+        {  26, 219, 229 }, {   5, 155, 207 }, {   1,  94, 151 },
+        {   1,  60, 104 }, {   1,  36,  62 }, {   1,  16,  28 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 233,  29, 248 }, { 146,  47, 220 }, {  43,  52, 140 }
+      }, {  // Band 1
+        { 100, 163, 232 }, { 179, 161, 222 }, {  63, 142, 204 },
+        {  37, 113, 174 }, {  26,  89, 137 }, {  18,  68,  97 }
+      }, {  // Band 2
+        {  85, 181, 230 }, {  32, 146, 209 }, {   7, 100, 164 },
+        {   3,  71, 121 }, {   1,  45,  77 }, {   1,  18,  30 }
+      }, {  // Band 3
+        {  65, 187, 230 }, {  20, 148, 207 }, {   2,  97, 159 },
+        {   1,  68, 116 }, {   1,  40,  70 }, {   1,  14,  29 }
+      }, {  // Band 4
+        {  40, 194, 227 }, {   8, 147, 204 }, {   1,  94, 155 },
+        {   1,  65, 112 }, {   1,  39,  66 }, {   1,  14,  26 }
+      }, {  // Band 5
+        {  16, 208, 228 }, {   3, 151, 207 }, {   1,  98, 160 },
+        {   1,  67, 117 }, {   1,  41,  74 }, {   1,  17,  31 }
+      }
+    }
+  }
+};
+
+static const av1_coeff_probs_model default_coef_probs_32x32[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        {  17,  38, 140 }, {   7,  34,  80 }, {   1,  17,  29 }
+      }, {  // Band 1
+        {  37,  75, 128 }, {  41,  76, 128 }, {  26,  66, 116 },
+        {  12,  52,  94 }, {   2,  32,  55 }, {   1,  10,  16 }
+      }, {  // Band 2
+        {  50, 127, 154 }, {  37, 109, 152 }, {  16,  82, 121 },
+        {   5,  59,  85 }, {   1,  35,  54 }, {   1,  13,  20 }
+      }, {  // Band 3
+        {  40, 142, 167 }, {  17, 110, 157 }, {   2,  71, 112 },
+        {   1,  44,  72 }, {   1,  27,  45 }, {   1,  11,  17 }
+      }, {  // Band 4
+        {  30, 175, 188 }, {   9, 124, 169 }, {   1,  74, 116 },
+        {   1,  48,  78 }, {   1,  30,  49 }, {   1,  11,  18 }
+      }, {  // Band 5
+        {  10, 222, 223 }, {   2, 150, 194 }, {   1,  83, 128 },
+        {   1,  48,  79 }, {   1,  27,  45 }, {   1,  11,  17 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        {  36,  41, 235 }, {  29,  36, 193 }, {  10,  27, 111 }
+      }, {  // Band 1
+        {  85, 165, 222 }, { 177, 162, 215 }, { 110, 135, 195 },
+        {  57, 113, 168 }, {  23,  83, 120 }, {  10,  49,  61 }
+      }, {  // Band 2
+        {  85, 190, 223 }, {  36, 139, 200 }, {   5,  90, 146 },
+        {   1,  60, 103 }, {   1,  38,  65 }, {   1,  18,  30 }
+      }, {  // Band 3
+        {  72, 202, 223 }, {  23, 141, 199 }, {   2,  86, 140 },
+        {   1,  56,  97 }, {   1,  36,  61 }, {   1,  16,  27 }
+      }, {  // Band 4
+        {  55, 218, 225 }, {  13, 145, 200 }, {   1,  86, 141 },
+        {   1,  57,  99 }, {   1,  35,  61 }, {   1,  13,  22 }
+      }, {  // Band 5
+        {  15, 235, 212 }, {   1, 132, 184 }, {   1,  84, 139 },
+        {   1,  57,  97 }, {   1,  34,  56 }, {   1,  14,  23 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 181,  21, 201 }, {  61,  37, 123 }, {  10,  38,  71 }
+      }, {  // Band 1
+        {  47, 106, 172 }, {  95, 104, 173 }, {  42,  93, 159 },
+        {  18,  77, 131 }, {   4,  50,  81 }, {   1,  17,  23 }
+      }, {  // Band 2
+        {  62, 147, 199 }, {  44, 130, 189 }, {  28, 102, 154 },
+        {  18,  75, 115 }, {   2,  44,  65 }, {   1,  12,  19 }
+      }, {  // Band 3
+        {  55, 153, 210 }, {  24, 130, 194 }, {   3,  93, 146 },
+        {   1,  61,  97 }, {   1,  31,  50 }, {   1,  10,  16 }
+      }, {  // Band 4
+        {  49, 186, 223 }, {  17, 148, 204 }, {   1,  96, 142 },
+        {   1,  53,  83 }, {   1,  26,  44 }, {   1,  11,  17 }
+      }, {  // Band 5
+        {  13, 217, 212 }, {   2, 136, 180 }, {   1,  78, 124 },
+        {   1,  50,  83 }, {   1,  29,  49 }, {   1,  14,  23 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 197,  13, 247 }, {  82,  17, 222 }, {  25,  17, 162 }
+      }, {  // Band 1
+        { 126, 186, 247 }, { 234, 191, 243 }, { 176, 177, 234 },
+        { 104, 158, 220 }, {  66, 128, 186 }, {  55,  90, 137 }
+      }, {  // Band 2
+        { 111, 197, 242 }, {  46, 158, 219 }, {   9, 104, 171 },
+        {   2,  65, 125 }, {   1,  44,  80 }, {   1,  17,  91 }
+      }, {  // Band 3
+        { 104, 208, 245 }, {  39, 168, 224 }, {   3, 109, 162 },
+        {   1,  79, 124 }, {   1,  50, 102 }, {   1,  43, 102 }
+      }, {  // Band 4
+        {  84, 220, 246 }, {  31, 177, 231 }, {   2, 115, 180 },
+        {   1,  79, 134 }, {   1,  55,  77 }, {   1,  60,  79 }
+      }, {  // Band 5
+        {  43, 243, 240 }, {   8, 180, 217 }, {   1, 115, 166 },
+        {   1,  84, 121 }, {   1,  51,  67 }, {   1,  16,   6 }
+      }
+    }
+  }
+};
+#endif  // CONFIG_NEW_TOKENSET
+
+#if CONFIG_TX64X64
+// FIXME. Optimize for EC_MULTISYMBOL
+static const av1_coeff_probs_model default_coef_probs_64x64[PLANE_TYPES] = {
+  {  // Y plane
+    {  // Intra
+      {  // Band 0
+        {  17,  38, 140 }, {   7,  34,  80 }, {   1,  17,  29 }
+      }, {  // Band 1
+        {  37,  75, 128 }, {  41,  76, 128 }, {  26,  66, 116 },
+        {  12,  52,  94 }, {   2,  32,  55 }, {   1,  10,  16 }
+      }, {  // Band 2
+        {  50, 127, 154 }, {  37, 109, 152 }, {  16,  82, 121 },
+        {   5,  59,  85 }, {   1,  35,  54 }, {   1,  13,  20 }
+      }, {  // Band 3
+        {  40, 142, 167 }, {  17, 110, 157 }, {   2,  71, 112 },
+        {   1,  44,  72 }, {   1,  27,  45 }, {   1,  11,  17 }
+      }, {  // Band 4
+        {  30, 175, 188 }, {   9, 124, 169 }, {   1,  74, 116 },
+        {   1,  48,  78 }, {   1,  30,  49 }, {   1,  11,  18 }
+      }, {  // Band 5
+        {  10, 222, 223 }, {   2, 150, 194 }, {   1,  83, 128 },
+        {   1,  48,  79 }, {   1,  27,  45 }, {   1,  11,  17 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        {  36,  41, 235 }, {  29,  36, 193 }, {  10,  27, 111 }
+      }, {  // Band 1
+        {  85, 165, 222 }, { 177, 162, 215 }, { 110, 135, 195 },
+        {  57, 113, 168 }, {  23,  83, 120 }, {  10,  49,  61 }
+      }, {  // Band 2
+        {  85, 190, 223 }, {  36, 139, 200 }, {   5,  90, 146 },
+        {   1,  60, 103 }, {   1,  38,  65 }, {   1,  18,  30 }
+      }, {  // Band 3
+        {  72, 202, 223 }, {  23, 141, 199 }, {   2,  86, 140 },
+        {   1,  56,  97 }, {   1,  36,  61 }, {   1,  16,  27 }
+      }, {  // Band 4
+        {  55, 218, 225 }, {  13, 145, 200 }, {   1,  86, 141 },
+        {   1,  57,  99 }, {   1,  35,  61 }, {   1,  13,  22 }
+      }, {  // Band 5
+        {  15, 235, 212 }, {   1, 132, 184 }, {   1,  84, 139 },
+        {   1,  57,  97 }, {   1,  34,  56 }, {   1,  14,  23 }
+      }
+    }
+  }, {  // UV plane
+    {  // Intra
+      {  // Band 0
+        { 181,  21, 201 }, {  61,  37, 123 }, {  10,  38,  71 }
+      }, {  // Band 1
+        {  47, 106, 172 }, {  95, 104, 173 }, {  42,  93, 159 },
+        {  18,  77, 131 }, {   4,  50,  81 }, {   1,  17,  23 }
+      }, {  // Band 2
+        {  62, 147, 199 }, {  44, 130, 189 }, {  28, 102, 154 },
+        {  18,  75, 115 }, {   2,  44,  65 }, {   1,  12,  19 }
+      }, {  // Band 3
+        {  55, 153, 210 }, {  24, 130, 194 }, {   3,  93, 146 },
+        {   1,  61,  97 }, {   1,  31,  50 }, {   1,  10,  16 }
+      }, {  // Band 4
+        {  49, 186, 223 }, {  17, 148, 204 }, {   1,  96, 142 },
+        {   1,  53,  83 }, {   1,  26,  44 }, {   1,  11,  17 }
+      }, {  // Band 5
+        {  13, 217, 212 }, {   2, 136, 180 }, {   1,  78, 124 },
+        {   1,  50,  83 }, {   1,  29,  49 }, {   1,  14,  23 }
+      }
+    }, {  // Inter
+      {  // Band 0
+        { 197,  13, 247 }, {  82,  17, 222 }, {  25,  17, 162 }
+      }, {  // Band 1
+        { 126, 186, 247 }, { 234, 191, 243 }, { 176, 177, 234 },
+        { 104, 158, 220 }, {  66, 128, 186 }, {  55,  90, 137 }
+      }, {  // Band 2
+        { 111, 197, 242 }, {  46, 158, 219 }, {   9, 104, 171 },
+        {   2,  65, 125 }, {   1,  44,  80 }, {   1,  17,  91 }
+      }, {  // Band 3
+        { 104, 208, 245 }, {  39, 168, 224 }, {   3, 109, 162 },
+        {   1,  79, 124 }, {   1,  50, 102 }, {   1,  43, 102 }
+      }, {  // Band 4
+        {  84, 220, 246 }, {  31, 177, 231 }, {   2, 115, 180 },
+        {   1,  79, 134 }, {   1,  55,  77 }, {   1,  60,  79 }
+      }, {  // Band 5
+        {  43, 243, 240 }, {   8, 180, 217 }, {   1, 115, 166 },
+        {   1,  84, 121 }, {   1,  51,  67 }, {   1,  16,   6 }
+      }
+    }
+  }
+};
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_Q_ADAPT_PROBS
+#if CONFIG_NEW_TOKENSET
+static const aom_prob av1_default_blockzero_probs[TX_SIZES][PLANE_TYPES]
+                                           [REF_TYPES][BLOCKZ_CONTEXTS] = {
+  { // TX_4x4
+    { // Y plane
+      { 195, 84, 8, },  // Intra
+      { 191, 124, 25, },  // Inter
+    },
+    { // UV plane
+      { 214, 132, 42, },  // Intra
+      { 229, 143, 46, },  // Inter
+    },
+  },
+  { // TX_8x8
+    { // Y plane
+      { 125, 52, 6, },  // Intra
+      { 202, 108, 18, },  // Inter
+    },
+    { // UV plane
+      { 212, 113, 29, },  // Intra
+      { 225, 144, 42, },  // Inter
+    },
+  },
+  { // TX_16x16
+    { // Y plane
+      { 7, 5, 1, },  // Intra
+      { 19, 19, 3, },  // Inter
+    },
+    { // UV plane
+      { 211, 96, 22, },  // Intra
+      { 233, 146, 43, },  // Inter
+    },
+  },
+  { // TX_32x32
+    { //  Y plane
+      { 17, 7, 1, },  // Intra
+      { 36, 29, 10, },  // Inter
+    },
+    { // UV plane
+      { 181, 61, 10, },  // Intra
+      { 197, 82, 25, },  // Inter
+    },
+  },
+#if CONFIG_TX64X64
+  { // TX_64x64 FIXME: currently the same as 32x32
+    { //  Y plane
+      { 17, 7, 1, },  // Intra
+      { 36, 29, 10, },  // Inter
+    },
+    { // UV plane
+      { 181, 61, 10, },  // Intra
+      { 197, 82, 25, },  // Inter
+    },
+  },
+#endif
+};
+
+static const coeff_cdf_model default_coef_head_cdf_4x4[PLANE_TYPES] = {
+  {     // Y plane
+    {   // Intra
+      { // Band 0
+        { AOM_ICDF(25024), AOM_ICDF(25863), AOM_ICDF(27361), AOM_ICDF(29796),
+          AOM_ICDF(30374), AOM_ICDF(32768) },
+        { AOM_ICDF(10816), AOM_ICDF(14127), AOM_ICDF(17116), AOM_ICDF(23516),
+          AOM_ICDF(24999), AOM_ICDF(32768) },
+        { AOM_ICDF(1088), AOM_ICDF(6358), AOM_ICDF(8428), AOM_ICDF(16648),
+          AOM_ICDF(18276), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(14529), AOM_ICDF(18769), AOM_ICDF(29100), AOM_ICDF(29634),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12993), AOM_ICDF(17117), AOM_ICDF(28404), AOM_ICDF(28988),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11201), AOM_ICDF(14084), AOM_ICDF(25818), AOM_ICDF(26504),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9793), AOM_ICDF(11267), AOM_ICDF(21775), AOM_ICDF(22451),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7105), AOM_ICDF(7562), AOM_ICDF(15777), AOM_ICDF(16225),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3905), AOM_ICDF(3966), AOM_ICDF(8359), AOM_ICDF(8526),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(20033), AOM_ICDF(23643), AOM_ICDF(31102), AOM_ICDF(31374),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16321), AOM_ICDF(20350), AOM_ICDF(30167), AOM_ICDF(30546),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12993), AOM_ICDF(15512), AOM_ICDF(26859), AOM_ICDF(27396),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10305), AOM_ICDF(11659), AOM_ICDF(21669), AOM_ICDF(22330),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7361), AOM_ICDF(7819), AOM_ICDF(15450), AOM_ICDF(15940),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3521), AOM_ICDF(3580), AOM_ICDF(7805), AOM_ICDF(7976),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(21057), AOM_ICDF(25460), AOM_ICDF(31740), AOM_ICDF(31952),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16449), AOM_ICDF(21173), AOM_ICDF(30761), AOM_ICDF(31092),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11841), AOM_ICDF(14615), AOM_ICDF(26188), AOM_ICDF(26824),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7745), AOM_ICDF(8991), AOM_ICDF(18937), AOM_ICDF(19707),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4417), AOM_ICDF(4706), AOM_ICDF(10342), AOM_ICDF(10890),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7617), AOM_ICDF(8392), AOM_ICDF(17295), AOM_ICDF(17915),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(20417), AOM_ICDF(26452), AOM_ICDF(32166), AOM_ICDF(32321),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15809), AOM_ICDF(21634), AOM_ICDF(30947), AOM_ICDF(31298),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10049), AOM_ICDF(12176), AOM_ICDF(23495), AOM_ICDF(24229),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5953), AOM_ICDF(6731), AOM_ICDF(16166), AOM_ICDF(16798),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6081), AOM_ICDF(6188), AOM_ICDF(8114), AOM_ICDF(8764),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2113), AOM_ICDF(2291), AOM_ICDF(4448), AOM_ICDF(5527),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(9153), AOM_ICDF(25905), AOM_ICDF(31431), AOM_ICDF(31934),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9025), AOM_ICDF(23345), AOM_ICDF(30033), AOM_ICDF(30965),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5953), AOM_ICDF(13835), AOM_ICDF(22032), AOM_ICDF(24664),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6337), AOM_ICDF(11435), AOM_ICDF(18366), AOM_ICDF(21418),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3137), AOM_ICDF(4871), AOM_ICDF(8519), AOM_ICDF(12426),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1857), AOM_ICDF(2727), AOM_ICDF(5540), AOM_ICDF(8757),
+         AOM_ICDF(32768) } } },
+    {   // Intra
+      { // Band 0
+        { AOM_ICDF(24512), AOM_ICDF(26673), AOM_ICDF(28962), AOM_ICDF(31929),
+          AOM_ICDF(32126), AOM_ICDF(32768) },
+        { AOM_ICDF(15936), AOM_ICDF(21711), AOM_ICDF(25569), AOM_ICDF(30899),
+          AOM_ICDF(31305), AOM_ICDF(32768) },
+        { AOM_ICDF(3264), AOM_ICDF(14756), AOM_ICDF(20107), AOM_ICDF(29407),
+          AOM_ICDF(30032), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(21313), AOM_ICDF(26020), AOM_ICDF(32523), AOM_ICDF(32575),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18369), AOM_ICDF(24215), AOM_ICDF(32291), AOM_ICDF(32391),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15297), AOM_ICDF(19637), AOM_ICDF(30414), AOM_ICDF(30752),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11713), AOM_ICDF(14040), AOM_ICDF(25408), AOM_ICDF(26033),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9537), AOM_ICDF(10173), AOM_ICDF(18839), AOM_ICDF(19315),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9025), AOM_ICDF(9093), AOM_ICDF(13987), AOM_ICDF(14115),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(22721), AOM_ICDF(27599), AOM_ICDF(32592), AOM_ICDF(32636),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(19009), AOM_ICDF(24676), AOM_ICDF(32258), AOM_ICDF(32367),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12737), AOM_ICDF(16769), AOM_ICDF(28739), AOM_ICDF(29247),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8769), AOM_ICDF(10956), AOM_ICDF(21941), AOM_ICDF(22840),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6721), AOM_ICDF(7678), AOM_ICDF(15319), AOM_ICDF(16290),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4417), AOM_ICDF(4430), AOM_ICDF(4583), AOM_ICDF(5712),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(22849), AOM_ICDF(28333), AOM_ICDF(32633), AOM_ICDF(32671),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18497), AOM_ICDF(24619), AOM_ICDF(32184), AOM_ICDF(32315),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11841), AOM_ICDF(14640), AOM_ICDF(27251), AOM_ICDF(27752),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8385), AOM_ICDF(10154), AOM_ICDF(18339), AOM_ICDF(19621),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(6977), AOM_ICDF(13787), AOM_ICDF(15289),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(194), AOM_ICDF(384), AOM_ICDF(479),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(20417), AOM_ICDF(28167), AOM_ICDF(32552), AOM_ICDF(32621),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16833), AOM_ICDF(23968), AOM_ICDF(31991), AOM_ICDF(32174),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10433), AOM_ICDF(13387), AOM_ICDF(26356), AOM_ICDF(26951),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5057), AOM_ICDF(6823), AOM_ICDF(18967), AOM_ICDF(19843),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(6479), AOM_ICDF(11672), AOM_ICDF(13052),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2241), AOM_ICDF(2265), AOM_ICDF(6355), AOM_ICDF(6432),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(12097), AOM_ICDF(28717), AOM_ICDF(32406), AOM_ICDF(32555),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10433), AOM_ICDF(26113), AOM_ICDF(31504), AOM_ICDF(31975),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5825), AOM_ICDF(14284), AOM_ICDF(21349), AOM_ICDF(24461),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4545), AOM_ICDF(8454), AOM_ICDF(12648), AOM_ICDF(17501),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(7173), AOM_ICDF(15272), AOM_ICDF(19322),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2113), AOM_ICDF(2183), AOM_ICDF(7202), AOM_ICDF(7377),
+         AOM_ICDF(32768) } } } },
+  {     // UV plane
+    {   // Inter
+      { // Band 0
+        { AOM_ICDF(27456), AOM_ICDF(28244), AOM_ICDF(31289), AOM_ICDF(32358),
+          AOM_ICDF(32534), AOM_ICDF(32768) },
+        { AOM_ICDF(16960), AOM_ICDF(21207), AOM_ICDF(26511), AOM_ICDF(30539),
+          AOM_ICDF(31190), AOM_ICDF(32768) },
+        { AOM_ICDF(5440), AOM_ICDF(13412), AOM_ICDF(18469), AOM_ICDF(26423),
+          AOM_ICDF(27669), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(17857), AOM_ICDF(26327), AOM_ICDF(31983), AOM_ICDF(32219),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16065), AOM_ICDF(24198), AOM_ICDF(31431), AOM_ICDF(31785),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12865), AOM_ICDF(18011), AOM_ICDF(28454), AOM_ICDF(29166),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9665), AOM_ICDF(12501), AOM_ICDF(24331), AOM_ICDF(25147),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2753), AOM_ICDF(3121), AOM_ICDF(12661), AOM_ICDF(13034),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4033), AOM_ICDF(4140), AOM_ICDF(11834), AOM_ICDF(11977),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(21185), AOM_ICDF(28338), AOM_ICDF(32249), AOM_ICDF(32417),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18497), AOM_ICDF(25227), AOM_ICDF(31905), AOM_ICDF(32122),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12097), AOM_ICDF(16516), AOM_ICDF(28610), AOM_ICDF(29166),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9281), AOM_ICDF(11157), AOM_ICDF(21438), AOM_ICDF(22312),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(6566), AOM_ICDF(15585), AOM_ICDF(16340),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9409), AOM_ICDF(9659), AOM_ICDF(11827), AOM_ICDF(12911),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(22337), AOM_ICDF(29459), AOM_ICDF(32382), AOM_ICDF(32519),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16961), AOM_ICDF(25262), AOM_ICDF(31874), AOM_ICDF(32123),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12353), AOM_ICDF(17748), AOM_ICDF(29300), AOM_ICDF(29852),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9025), AOM_ICDF(11528), AOM_ICDF(24468), AOM_ICDF(25141),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6209), AOM_ICDF(6565), AOM_ICDF(15806), AOM_ICDF(16121),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2497), AOM_ICDF(2524), AOM_ICDF(7050), AOM_ICDF(7125),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(20417), AOM_ICDF(29779), AOM_ICDF(32552), AOM_ICDF(32636),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15553), AOM_ICDF(26420), AOM_ICDF(32063), AOM_ICDF(32295),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9665), AOM_ICDF(17946), AOM_ICDF(29385), AOM_ICDF(30096),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5569), AOM_ICDF(10207), AOM_ICDF(22410), AOM_ICDF(23836),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16449), AOM_ICDF(16450), AOM_ICDF(16545), AOM_ICDF(16593),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2369), AOM_ICDF(2395), AOM_ICDF(6822), AOM_ICDF(6898),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(10177), AOM_ICDF(30567), AOM_ICDF(32725), AOM_ICDF(32745),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9537), AOM_ICDF(28243), AOM_ICDF(32179), AOM_ICDF(32423),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13377), AOM_ICDF(23187), AOM_ICDF(29322), AOM_ICDF(30382),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13121), AOM_ICDF(21346), AOM_ICDF(29507), AOM_ICDF(30326),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4417), AOM_ICDF(4939), AOM_ICDF(15104), AOM_ICDF(15535),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2625), AOM_ICDF(2680), AOM_ICDF(8218), AOM_ICDF(8338),
+         AOM_ICDF(32768) } } },
+    {   // Inter
+      { // Band 0
+        { AOM_ICDF(29376), AOM_ICDF(30098), AOM_ICDF(32421), AOM_ICDF(32766),
+          AOM_ICDF(32767), AOM_ICDF(32768) },
+        { AOM_ICDF(18368), AOM_ICDF(22916), AOM_ICDF(30116), AOM_ICDF(32541),
+          AOM_ICDF(32650), AOM_ICDF(32768) },
+        { AOM_ICDF(5952), AOM_ICDF(16505), AOM_ICDF(25955), AOM_ICDF(32163),
+          AOM_ICDF(32365), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(19649), AOM_ICDF(30160), AOM_ICDF(32743), AOM_ICDF(32753),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18881), AOM_ICDF(28724), AOM_ICDF(32688), AOM_ICDF(32717),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16833), AOM_ICDF(23053), AOM_ICDF(31244), AOM_ICDF(31573),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(14657), AOM_ICDF(17714), AOM_ICDF(26083), AOM_ICDF(26978),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(14657), AOM_ICDF(16618), AOM_ICDF(24597), AOM_ICDF(25403),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4289), AOM_ICDF(4326), AOM_ICDF(10686), AOM_ICDF(10751),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(21953), AOM_ICDF(30956), AOM_ICDF(32748), AOM_ICDF(32757),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20929), AOM_ICDF(29412), AOM_ICDF(32700), AOM_ICDF(32725),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13377), AOM_ICDF(21495), AOM_ICDF(31216), AOM_ICDF(31569),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9153), AOM_ICDF(15097), AOM_ICDF(28295), AOM_ICDF(28990),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5313), AOM_ICDF(5363), AOM_ICDF(13839), AOM_ICDF(13894),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2625), AOM_ICDF(2652), AOM_ICDF(7276), AOM_ICDF(7351),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(20289), AOM_ICDF(31164), AOM_ICDF(32745), AOM_ICDF(32755),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17601), AOM_ICDF(29635), AOM_ICDF(32739), AOM_ICDF(32751),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18241), AOM_ICDF(24284), AOM_ICDF(32116), AOM_ICDF(32258),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(32705), AOM_ICDF(32706), AOM_ICDF(32739), AOM_ICDF(32740),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(5750), AOM_ICDF(14739), AOM_ICDF(14792),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2881), AOM_ICDF(2913), AOM_ICDF(8427), AOM_ICDF(8498),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(19009), AOM_ICDF(31481), AOM_ICDF(32742), AOM_ICDF(32754),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15809), AOM_ICDF(30521), AOM_ICDF(32736), AOM_ICDF(32750),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16449), AOM_ICDF(32705), AOM_ICDF(32737), AOM_ICDF(32753),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7873), AOM_ICDF(8039), AOM_ICDF(19981), AOM_ICDF(20068),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5313), AOM_ICDF(5366), AOM_ICDF(14376), AOM_ICDF(14430),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2753), AOM_ICDF(2789), AOM_ICDF(8909), AOM_ICDF(8979),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(11841), AOM_ICDF(32116), AOM_ICDF(32728), AOM_ICDF(32748),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12353), AOM_ICDF(32132), AOM_ICDF(32729), AOM_ICDF(32748),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7489), AOM_ICDF(12435), AOM_ICDF(25708), AOM_ICDF(26666),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(7486), AOM_ICDF(20238), AOM_ICDF(21009),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4929), AOM_ICDF(5579), AOM_ICDF(16402), AOM_ICDF(16866),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3009), AOM_ICDF(3246), AOM_ICDF(10158), AOM_ICDF(10533),
+         AOM_ICDF(32768) } } } }
+};
+static const coeff_cdf_model default_coef_head_cdf_8x8[PLANE_TYPES] = {
+  {     // Y plane
+    {   // Intra
+      { // Band 0
+        { AOM_ICDF(16064), AOM_ICDF(18127), AOM_ICDF(22153), AOM_ICDF(27289),
+          AOM_ICDF(28507), AOM_ICDF(32768) },
+        { AOM_ICDF(6720), AOM_ICDF(10545), AOM_ICDF(13491), AOM_ICDF(20948),
+          AOM_ICDF(22631), AOM_ICDF(32768) },
+        { AOM_ICDF(832), AOM_ICDF(5270), AOM_ICDF(5918), AOM_ICDF(12645),
+          AOM_ICDF(13532), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(14017), AOM_ICDF(16139), AOM_ICDF(26799), AOM_ICDF(27295),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12737), AOM_ICDF(15136), AOM_ICDF(26235), AOM_ICDF(26816),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10817), AOM_ICDF(12445), AOM_ICDF(23637), AOM_ICDF(24217),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8897), AOM_ICDF(9702), AOM_ICDF(20040), AOM_ICDF(20500),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5953), AOM_ICDF(6156), AOM_ICDF(13966), AOM_ICDF(14205),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2497), AOM_ICDF(2519), AOM_ICDF(6222), AOM_ICDF(6300),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(19777), AOM_ICDF(21403), AOM_ICDF(30054), AOM_ICDF(30269),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16193), AOM_ICDF(17913), AOM_ICDF(28593), AOM_ICDF(28883),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12609), AOM_ICDF(13572), AOM_ICDF(25248), AOM_ICDF(25534),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9665), AOM_ICDF(10118), AOM_ICDF(20721), AOM_ICDF(20968),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6849), AOM_ICDF(7028), AOM_ICDF(15202), AOM_ICDF(15391),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3009), AOM_ICDF(3036), AOM_ICDF(7601), AOM_ICDF(7675),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(22593), AOM_ICDF(23915), AOM_ICDF(31159), AOM_ICDF(31283),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17345), AOM_ICDF(18690), AOM_ICDF(29425), AOM_ICDF(29611),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11969), AOM_ICDF(12540), AOM_ICDF(24685), AOM_ICDF(24867),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8129), AOM_ICDF(8355), AOM_ICDF(18668), AOM_ICDF(18819),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4673), AOM_ICDF(4714), AOM_ICDF(11752), AOM_ICDF(11814),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1857), AOM_ICDF(1876), AOM_ICDF(5057), AOM_ICDF(5138),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(24513), AOM_ICDF(25718), AOM_ICDF(31947), AOM_ICDF(32014),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18881), AOM_ICDF(20029), AOM_ICDF(30409), AOM_ICDF(30527),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12481), AOM_ICDF(12953), AOM_ICDF(25201), AOM_ICDF(25341),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8385), AOM_ICDF(8528), AOM_ICDF(18815), AOM_ICDF(18910),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4289), AOM_ICDF(4327), AOM_ICDF(10797), AOM_ICDF(10861),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1857), AOM_ICDF(1872), AOM_ICDF(4332), AOM_ICDF(4415),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(26049), AOM_ICDF(27752), AOM_ICDF(32415), AOM_ICDF(32462),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20417), AOM_ICDF(22100), AOM_ICDF(31056), AOM_ICDF(31192),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12481), AOM_ICDF(13075), AOM_ICDF(24646), AOM_ICDF(24844),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7489), AOM_ICDF(7696), AOM_ICDF(17117), AOM_ICDF(17285),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3777), AOM_ICDF(3814), AOM_ICDF(10062), AOM_ICDF(10129),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1473), AOM_ICDF(1486), AOM_ICDF(3735), AOM_ICDF(3820),
+         AOM_ICDF(32768) } } },
+    {   // Intra
+      { // Band 0
+        { AOM_ICDF(25920), AOM_ICDF(27743), AOM_ICDF(29455), AOM_ICDF(32147),
+          AOM_ICDF(32280), AOM_ICDF(32768) },
+        { AOM_ICDF(13888), AOM_ICDF(19845), AOM_ICDF(23350), AOM_ICDF(30219),
+          AOM_ICDF(30660), AOM_ICDF(32768) },
+        { AOM_ICDF(2368), AOM_ICDF(12781), AOM_ICDF(16196), AOM_ICDF(27232),
+          AOM_ICDF(27894), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(21697), AOM_ICDF(24758), AOM_ICDF(32358), AOM_ICDF(32417),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20289), AOM_ICDF(23960), AOM_ICDF(32111), AOM_ICDF(32213),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17345), AOM_ICDF(19966), AOM_ICDF(30630), AOM_ICDF(30841),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(14529), AOM_ICDF(16070), AOM_ICDF(27461), AOM_ICDF(27777),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9793), AOM_ICDF(10613), AOM_ICDF(21146), AOM_ICDF(21566),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6977), AOM_ICDF(7162), AOM_ICDF(15591), AOM_ICDF(15776),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(23617), AOM_ICDF(26783), AOM_ICDF(32572), AOM_ICDF(32607),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20801), AOM_ICDF(24292), AOM_ICDF(32185), AOM_ICDF(32275),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15169), AOM_ICDF(17905), AOM_ICDF(29916), AOM_ICDF(30181),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10945), AOM_ICDF(12972), AOM_ICDF(25565), AOM_ICDF(26064),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6849), AOM_ICDF(8334), AOM_ICDF(18543), AOM_ICDF(19446),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3649), AOM_ICDF(4346), AOM_ICDF(12351), AOM_ICDF(13169),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(25281), AOM_ICDF(28440), AOM_ICDF(32667), AOM_ICDF(32689),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22081), AOM_ICDF(25694), AOM_ICDF(32414), AOM_ICDF(32476),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15297), AOM_ICDF(18341), AOM_ICDF(30141), AOM_ICDF(30410),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10305), AOM_ICDF(12381), AOM_ICDF(24477), AOM_ICDF(25084),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(6673), AOM_ICDF(16325), AOM_ICDF(17080),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2369), AOM_ICDF(2393), AOM_ICDF(6466), AOM_ICDF(6543),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(25921), AOM_ICDF(29445), AOM_ICDF(32729), AOM_ICDF(32739),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22465), AOM_ICDF(26834), AOM_ICDF(32588), AOM_ICDF(32627),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16449), AOM_ICDF(20062), AOM_ICDF(31016), AOM_ICDF(31233),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11073), AOM_ICDF(13165), AOM_ICDF(25353), AOM_ICDF(25896),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11713), AOM_ICDF(13837), AOM_ICDF(20144), AOM_ICDF(21734),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2241), AOM_ICDF(2265), AOM_ICDF(6355), AOM_ICDF(6432),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(26177), AOM_ICDF(29403), AOM_ICDF(32705), AOM_ICDF(32721),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22337), AOM_ICDF(26344), AOM_ICDF(32545), AOM_ICDF(32589),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(19009), AOM_ICDF(21527), AOM_ICDF(31775), AOM_ICDF(31873),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11585), AOM_ICDF(12685), AOM_ICDF(22632), AOM_ICDF(23137),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8257), AOM_ICDF(8305), AOM_ICDF(16444), AOM_ICDF(16492),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2113), AOM_ICDF(2183), AOM_ICDF(7202), AOM_ICDF(7377),
+         AOM_ICDF(32768) } } } },
+  {     // UV plane
+    {   // Inter
+      { // Band 0
+        { AOM_ICDF(27200), AOM_ICDF(27981), AOM_ICDF(31389), AOM_ICDF(32444),
+          AOM_ICDF(32592), AOM_ICDF(32768) },
+        { AOM_ICDF(14528), AOM_ICDF(19068), AOM_ICDF(24887), AOM_ICDF(29901),
+          AOM_ICDF(30688), AOM_ICDF(32768) },
+        { AOM_ICDF(3776), AOM_ICDF(11778), AOM_ICDF(14700), AOM_ICDF(23745),
+          AOM_ICDF(24854), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(20289), AOM_ICDF(25202), AOM_ICDF(31672), AOM_ICDF(31909),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18369), AOM_ICDF(23493), AOM_ICDF(31166), AOM_ICDF(31487),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15425), AOM_ICDF(18619), AOM_ICDF(28941), AOM_ICDF(29393),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10945), AOM_ICDF(12535), AOM_ICDF(24287), AOM_ICDF(24792),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6465), AOM_ICDF(6810), AOM_ICDF(15764), AOM_ICDF(16080),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2113), AOM_ICDF(2137), AOM_ICDF(6125), AOM_ICDF(6203),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(23745), AOM_ICDF(27041), AOM_ICDF(31976), AOM_ICDF(32135),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(19521), AOM_ICDF(22766), AOM_ICDF(31139), AOM_ICDF(31367),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(14273), AOM_ICDF(15834), AOM_ICDF(27820), AOM_ICDF(28105),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9537), AOM_ICDF(10445), AOM_ICDF(22106), AOM_ICDF(22491),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7233), AOM_ICDF(7386), AOM_ICDF(15961), AOM_ICDF(16109),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2369), AOM_ICDF(2401), AOM_ICDF(7891), AOM_ICDF(7964),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(26305), AOM_ICDF(28703), AOM_ICDF(32352), AOM_ICDF(32435),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20673), AOM_ICDF(23490), AOM_ICDF(31517), AOM_ICDF(31680),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(14017), AOM_ICDF(15251), AOM_ICDF(27458), AOM_ICDF(27702),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10945), AOM_ICDF(11374), AOM_ICDF(22496), AOM_ICDF(22687),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9153), AOM_ICDF(9435), AOM_ICDF(22299), AOM_ICDF(22411),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(269), AOM_ICDF(13236), AOM_ICDF(13293),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(27713), AOM_ICDF(29770), AOM_ICDF(32522), AOM_ICDF(32575),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21569), AOM_ICDF(24342), AOM_ICDF(31785), AOM_ICDF(31919),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15297), AOM_ICDF(16497), AOM_ICDF(28367), AOM_ICDF(28569),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17601), AOM_ICDF(17828), AOM_ICDF(24444), AOM_ICDF(24582),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6977), AOM_ICDF(7035), AOM_ICDF(16901), AOM_ICDF(16947),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(384), AOM_ICDF(32706), AOM_ICDF(32707),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(28737), AOM_ICDF(30879), AOM_ICDF(32667), AOM_ICDF(32695),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22593), AOM_ICDF(26241), AOM_ICDF(32073), AOM_ICDF(32207),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16577), AOM_ICDF(19148), AOM_ICDF(28436), AOM_ICDF(28906),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12993), AOM_ICDF(14005), AOM_ICDF(23151), AOM_ICDF(23630),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7617), AOM_ICDF(9188), AOM_ICDF(22797), AOM_ICDF(23313),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2625), AOM_ICDF(2680), AOM_ICDF(8218), AOM_ICDF(8338),
+         AOM_ICDF(32768) } } },
+    {   // Inter
+      { // Band 0
+        { AOM_ICDF(28864), AOM_ICDF(29988), AOM_ICDF(32423), AOM_ICDF(32766),
+          AOM_ICDF(32767), AOM_ICDF(32768) },
+        { AOM_ICDF(18496), AOM_ICDF(24572), AOM_ICDF(30167), AOM_ICDF(32687),
+          AOM_ICDF(32737), AOM_ICDF(32768) },
+        { AOM_ICDF(5440), AOM_ICDF(19618), AOM_ICDF(25332), AOM_ICDF(32393),
+          AOM_ICDF(32491), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(23745), AOM_ICDF(29427), AOM_ICDF(32751), AOM_ICDF(32757),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(23745), AOM_ICDF(28704), AOM_ICDF(32716), AOM_ICDF(32731),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(23105), AOM_ICDF(27943), AOM_ICDF(32524), AOM_ICDF(32587),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21057), AOM_ICDF(24773), AOM_ICDF(29589), AOM_ICDF(30282),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12609), AOM_ICDF(14823), AOM_ICDF(23831), AOM_ICDF(24713),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16449), AOM_ICDF(16450), AOM_ICDF(16545), AOM_ICDF(16593),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(25025), AOM_ICDF(30203), AOM_ICDF(32754), AOM_ICDF(32759),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(23617), AOM_ICDF(28361), AOM_ICDF(32715), AOM_ICDF(32729),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17985), AOM_ICDF(21562), AOM_ICDF(31354), AOM_ICDF(31543),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12353), AOM_ICDF(18915), AOM_ICDF(28742), AOM_ICDF(29548),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(289), AOM_ICDF(16545), AOM_ICDF(16593),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2625), AOM_ICDF(2652), AOM_ICDF(7276), AOM_ICDF(7351),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(26433), AOM_ICDF(30892), AOM_ICDF(32757), AOM_ICDF(32761),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(24513), AOM_ICDF(29274), AOM_ICDF(32721), AOM_ICDF(32735),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20161), AOM_ICDF(24040), AOM_ICDF(32055), AOM_ICDF(32171),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21953), AOM_ICDF(24678), AOM_ICDF(27382), AOM_ICDF(28734),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(5750), AOM_ICDF(14739), AOM_ICDF(14792),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2881), AOM_ICDF(2913), AOM_ICDF(8427), AOM_ICDF(8498),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(27457), AOM_ICDF(31485), AOM_ICDF(32759), AOM_ICDF(32763),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(24129), AOM_ICDF(29502), AOM_ICDF(32752), AOM_ICDF(32757),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(19009), AOM_ICDF(25452), AOM_ICDF(32473), AOM_ICDF(32544),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(32705), AOM_ICDF(32706), AOM_ICDF(32737), AOM_ICDF(32738),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5313), AOM_ICDF(5366), AOM_ICDF(14376), AOM_ICDF(14430),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2753), AOM_ICDF(2789), AOM_ICDF(8909), AOM_ICDF(8979),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(27841), AOM_ICDF(32288), AOM_ICDF(32759), AOM_ICDF(32764),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(19137), AOM_ICDF(30271), AOM_ICDF(32742), AOM_ICDF(32753),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18625), AOM_ICDF(27739), AOM_ICDF(29979), AOM_ICDF(31099),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(7486), AOM_ICDF(20238), AOM_ICDF(21009),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4929), AOM_ICDF(5579), AOM_ICDF(16402), AOM_ICDF(16866),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3009), AOM_ICDF(3246), AOM_ICDF(10158), AOM_ICDF(10533),
+         AOM_ICDF(32768) } } } }
+};
+static const coeff_cdf_model default_coef_head_cdf_16x16[PLANE_TYPES] = {
+  {     // Y plane
+    {   // Intra
+      { // Band 0
+        { AOM_ICDF(960), AOM_ICDF(4882), AOM_ICDF(9467), AOM_ICDF(17710),
+          AOM_ICDF(20412), AOM_ICDF(32768) },
+        { AOM_ICDF(704), AOM_ICDF(4657), AOM_ICDF(6561), AOM_ICDF(14507),
+          AOM_ICDF(16279), AOM_ICDF(32768) },
+        { AOM_ICDF(192), AOM_ICDF(3443), AOM_ICDF(3759), AOM_ICDF(9011),
+          AOM_ICDF(9685), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(12481), AOM_ICDF(13958), AOM_ICDF(24487), AOM_ICDF(24997),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11457), AOM_ICDF(13075), AOM_ICDF(23820), AOM_ICDF(24406),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9793), AOM_ICDF(11127), AOM_ICDF(21775), AOM_ICDF(22387),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7745), AOM_ICDF(8457), AOM_ICDF(18155), AOM_ICDF(18655),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5441), AOM_ICDF(5668), AOM_ICDF(13180), AOM_ICDF(13467),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2497), AOM_ICDF(2520), AOM_ICDF(6340), AOM_ICDF(6417),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(19521), AOM_ICDF(20572), AOM_ICDF(28965), AOM_ICDF(29177),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15425), AOM_ICDF(16741), AOM_ICDF(27247), AOM_ICDF(27554),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11969), AOM_ICDF(12690), AOM_ICDF(23872), AOM_ICDF(24141),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9281), AOM_ICDF(9678), AOM_ICDF(19970), AOM_ICDF(20207),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6081), AOM_ICDF(6266), AOM_ICDF(14682), AOM_ICDF(14876),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2753), AOM_ICDF(2779), AOM_ICDF(7150), AOM_ICDF(7225),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(22337), AOM_ICDF(23293), AOM_ICDF(30630), AOM_ICDF(30753),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16321), AOM_ICDF(17427), AOM_ICDF(28368), AOM_ICDF(28570),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11457), AOM_ICDF(11907), AOM_ICDF(23570), AOM_ICDF(23741),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7233), AOM_ICDF(7331), AOM_ICDF(17258), AOM_ICDF(17334),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4033), AOM_ICDF(4070), AOM_ICDF(10375), AOM_ICDF(10441),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1601), AOM_ICDF(1619), AOM_ICDF(4706), AOM_ICDF(4788),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(24769), AOM_ICDF(25536), AOM_ICDF(31660), AOM_ICDF(31722),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18113), AOM_ICDF(18886), AOM_ICDF(29420), AOM_ICDF(29534),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11201), AOM_ICDF(11412), AOM_ICDF(23207), AOM_ICDF(23291),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6977), AOM_ICDF(7033), AOM_ICDF(16599), AOM_ICDF(16646),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4033), AOM_ICDF(4070), AOM_ICDF(10375), AOM_ICDF(10441),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1601), AOM_ICDF(1620), AOM_ICDF(4827), AOM_ICDF(4909),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(28353), AOM_ICDF(28831), AOM_ICDF(32502), AOM_ICDF(32517),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21441), AOM_ICDF(21869), AOM_ICDF(30977), AOM_ICDF(31017),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11969), AOM_ICDF(12088), AOM_ICDF(24116), AOM_ICDF(24158),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7489), AOM_ICDF(7547), AOM_ICDF(17413), AOM_ICDF(17458),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4545), AOM_ICDF(4585), AOM_ICDF(11325), AOM_ICDF(11388),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2113), AOM_ICDF(2133), AOM_ICDF(5526), AOM_ICDF(5606),
+         AOM_ICDF(32768) } } },
+    {   // Intra
+      { // Band 0
+        { AOM_ICDF(2496), AOM_ICDF(8717), AOM_ICDF(17280), AOM_ICDF(28922),
+          AOM_ICDF(29751), AOM_ICDF(32768) },
+        { AOM_ICDF(2496), AOM_ICDF(9665), AOM_ICDF(15235), AOM_ICDF(26542),
+          AOM_ICDF(27580), AOM_ICDF(32768) },
+        { AOM_ICDF(448), AOM_ICDF(9240), AOM_ICDF(11886), AOM_ICDF(24124),
+          AOM_ICDF(24898), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(21057), AOM_ICDF(22896), AOM_ICDF(31877), AOM_ICDF(31953),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20673), AOM_ICDF(23151), AOM_ICDF(31706), AOM_ICDF(31825),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18753), AOM_ICDF(20519), AOM_ICDF(30497), AOM_ICDF(30668),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15425), AOM_ICDF(16608), AOM_ICDF(27789), AOM_ICDF(28027),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10305), AOM_ICDF(10977), AOM_ICDF(21405), AOM_ICDF(21749),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3649), AOM_ICDF(3812), AOM_ICDF(11213), AOM_ICDF(11445),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(24001), AOM_ICDF(25899), AOM_ICDF(32307), AOM_ICDF(32360),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20929), AOM_ICDF(22941), AOM_ICDF(31775), AOM_ICDF(31867),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15169), AOM_ICDF(16734), AOM_ICDF(29228), AOM_ICDF(29425),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10561), AOM_ICDF(12047), AOM_ICDF(24918), AOM_ICDF(25324),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6977), AOM_ICDF(7929), AOM_ICDF(18311), AOM_ICDF(18918),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3649), AOM_ICDF(3760), AOM_ICDF(9962), AOM_ICDF(10162),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(25793), AOM_ICDF(27526), AOM_ICDF(32565), AOM_ICDF(32591),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21825), AOM_ICDF(23885), AOM_ICDF(32064), AOM_ICDF(32135),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15041), AOM_ICDF(16286), AOM_ICDF(29203), AOM_ICDF(29360),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10433), AOM_ICDF(11058), AOM_ICDF(24349), AOM_ICDF(24538),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5569), AOM_ICDF(6016), AOM_ICDF(16460), AOM_ICDF(16794),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(194), AOM_ICDF(384), AOM_ICDF(479),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(26433), AOM_ICDF(28398), AOM_ICDF(32682), AOM_ICDF(32696),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22977), AOM_ICDF(25086), AOM_ICDF(32367), AOM_ICDF(32412),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16577), AOM_ICDF(17928), AOM_ICDF(30144), AOM_ICDF(30275),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12481), AOM_ICDF(13352), AOM_ICDF(25993), AOM_ICDF(26211),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7745), AOM_ICDF(8069), AOM_ICDF(20501), AOM_ICDF(20657),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16449), AOM_ICDF(16450), AOM_ICDF(16545), AOM_ICDF(16593),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(27841), AOM_ICDF(29700), AOM_ICDF(32721), AOM_ICDF(32730),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(23873), AOM_ICDF(26202), AOM_ICDF(32578), AOM_ICDF(32604),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17729), AOM_ICDF(19046), AOM_ICDF(30448), AOM_ICDF(30568),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13505), AOM_ICDF(14508), AOM_ICDF(26034), AOM_ICDF(26304),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10049), AOM_ICDF(10494), AOM_ICDF(19945), AOM_ICDF(20233),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2113), AOM_ICDF(2183), AOM_ICDF(7202), AOM_ICDF(7377),
+         AOM_ICDF(32768) } } } },
+  {     // UV plane
+    {   // Inter
+      { // Band 0
+        { AOM_ICDF(27072), AOM_ICDF(27916), AOM_ICDF(31095), AOM_ICDF(32400),
+          AOM_ICDF(32553), AOM_ICDF(32768) },
+        { AOM_ICDF(12352), AOM_ICDF(16792), AOM_ICDF(22516), AOM_ICDF(28853),
+          AOM_ICDF(29797), AOM_ICDF(32768) },
+        { AOM_ICDF(2880), AOM_ICDF(9023), AOM_ICDF(11126), AOM_ICDF(20602),
+          AOM_ICDF(21713), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(20161), AOM_ICDF(24785), AOM_ICDF(31070), AOM_ICDF(31430),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17985), AOM_ICDF(22773), AOM_ICDF(30430), AOM_ICDF(30880),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15937), AOM_ICDF(18802), AOM_ICDF(28265), AOM_ICDF(28788),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11841), AOM_ICDF(13587), AOM_ICDF(24798), AOM_ICDF(25335),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8769), AOM_ICDF(9160), AOM_ICDF(19316), AOM_ICDF(19566),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5313), AOM_ICDF(5357), AOM_ICDF(12874), AOM_ICDF(12932),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(24129), AOM_ICDF(26501), AOM_ICDF(31672), AOM_ICDF(31844),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(19649), AOM_ICDF(21553), AOM_ICDF(30130), AOM_ICDF(30370),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11713), AOM_ICDF(13134), AOM_ICDF(25983), AOM_ICDF(26321),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9409), AOM_ICDF(9948), AOM_ICDF(21408), AOM_ICDF(21663),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5569), AOM_ICDF(5757), AOM_ICDF(14335), AOM_ICDF(14533),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2241), AOM_ICDF(2305), AOM_ICDF(13152), AOM_ICDF(13209),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(26817), AOM_ICDF(28135), AOM_ICDF(32130), AOM_ICDF(32209),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20161), AOM_ICDF(21412), AOM_ICDF(30331), AOM_ICDF(30481),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13377), AOM_ICDF(13798), AOM_ICDF(26065), AOM_ICDF(26176),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8129), AOM_ICDF(8290), AOM_ICDF(19920), AOM_ICDF(20008),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(5751), AOM_ICDF(14950), AOM_ICDF(15002),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5569), AOM_ICDF(5601), AOM_ICDF(11041), AOM_ICDF(11105),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(28225), AOM_ICDF(29079), AOM_ICDF(32387), AOM_ICDF(32426),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21185), AOM_ICDF(22046), AOM_ICDF(30982), AOM_ICDF(31061),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13377), AOM_ICDF(13595), AOM_ICDF(25762), AOM_ICDF(25824),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8001), AOM_ICDF(8123), AOM_ICDF(20530), AOM_ICDF(20590),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4289), AOM_ICDF(4322), AOM_ICDF(9907), AOM_ICDF(9974),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3393), AOM_ICDF(3412), AOM_ICDF(6663), AOM_ICDF(6739),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(30529), AOM_ICDF(31014), AOM_ICDF(32651), AOM_ICDF(32664),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(23489), AOM_ICDF(24268), AOM_ICDF(31627), AOM_ICDF(31682),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(14017), AOM_ICDF(14239), AOM_ICDF(26653), AOM_ICDF(26707),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11201), AOM_ICDF(11317), AOM_ICDF(23122), AOM_ICDF(23169),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6721), AOM_ICDF(6768), AOM_ICDF(14810), AOM_ICDF(14863),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6593), AOM_ICDF(6632), AOM_ICDF(13188), AOM_ICDF(13245),
+         AOM_ICDF(32768) } } },
+    {   // Inter
+      { // Band 0
+        { AOM_ICDF(29888), AOM_ICDF(30492), AOM_ICDF(32500), AOM_ICDF(32766),
+          AOM_ICDF(32767), AOM_ICDF(32768) },
+        { AOM_ICDF(18752), AOM_ICDF(23235), AOM_ICDF(29846), AOM_ICDF(32214),
+          AOM_ICDF(32442), AOM_ICDF(32768) },
+        { AOM_ICDF(5568), AOM_ICDF(17762), AOM_ICDF(25039), AOM_ICDF(31213),
+          AOM_ICDF(31651), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(26433), AOM_ICDF(29681), AOM_ICDF(32757), AOM_ICDF(32760),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(24769), AOM_ICDF(28761), AOM_ICDF(32722), AOM_ICDF(32734),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22209), AOM_ICDF(26975), AOM_ICDF(32418), AOM_ICDF(32500),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16321), AOM_ICDF(21333), AOM_ICDF(28368), AOM_ICDF(29283),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12865), AOM_ICDF(14775), AOM_ICDF(22545), AOM_ICDF(23553),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12353), AOM_ICDF(12354), AOM_ICDF(12473), AOM_ICDF(12532),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(27457), AOM_ICDF(30005), AOM_ICDF(32738), AOM_ICDF(32745),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(24897), AOM_ICDF(27541), AOM_ICDF(32723), AOM_ICDF(32731),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15297), AOM_ICDF(19106), AOM_ICDF(30414), AOM_ICDF(30711),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6593), AOM_ICDF(8826), AOM_ICDF(19732), AOM_ICDF(20840),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4161), AOM_ICDF(4233), AOM_ICDF(16509), AOM_ICDF(16557),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2625), AOM_ICDF(2652), AOM_ICDF(7276), AOM_ICDF(7351),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(28609), AOM_ICDF(30482), AOM_ICDF(32761), AOM_ICDF(32763),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(25665), AOM_ICDF(27830), AOM_ICDF(32727), AOM_ICDF(32733),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21057), AOM_ICDF(23803), AOM_ICDF(30367), AOM_ICDF(30721),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10945), AOM_ICDF(21878), AOM_ICDF(32726), AOM_ICDF(32737),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(5750), AOM_ICDF(14739), AOM_ICDF(14792),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2881), AOM_ICDF(2913), AOM_ICDF(8427), AOM_ICDF(8498),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(28993), AOM_ICDF(30944), AOM_ICDF(32762), AOM_ICDF(32764),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(26561), AOM_ICDF(28695), AOM_ICDF(32733), AOM_ICDF(32739),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17985), AOM_ICDF(19028), AOM_ICDF(31008), AOM_ICDF(31079),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7873), AOM_ICDF(8039), AOM_ICDF(19981), AOM_ICDF(20068),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5313), AOM_ICDF(5366), AOM_ICDF(14376), AOM_ICDF(14430),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2753), AOM_ICDF(2789), AOM_ICDF(8909), AOM_ICDF(8979),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(30273), AOM_ICDF(32029), AOM_ICDF(32764), AOM_ICDF(32766),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(28609), AOM_ICDF(30847), AOM_ICDF(32745), AOM_ICDF(32751),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21313), AOM_ICDF(24377), AOM_ICDF(31986), AOM_ICDF(32098),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(32705), AOM_ICDF(32709), AOM_ICDF(32739), AOM_ICDF(32741),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4929), AOM_ICDF(5579), AOM_ICDF(16402), AOM_ICDF(16866),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3009), AOM_ICDF(3246), AOM_ICDF(10158), AOM_ICDF(10533),
+         AOM_ICDF(32768) } } } }
+};
+static const coeff_cdf_model default_coef_head_cdf_32x32[PLANE_TYPES] = {
+  {     // Y plane
+    {   // Intra
+      { // Band 0
+        { AOM_ICDF(2240), AOM_ICDF(5407), AOM_ICDF(18304), AOM_ICDF(25601),
+          AOM_ICDF(27911), AOM_ICDF(32768) },
+        { AOM_ICDF(960), AOM_ICDF(4633), AOM_ICDF(8197), AOM_ICDF(16254),
+          AOM_ICDF(18796), AOM_ICDF(32768) },
+        { AOM_ICDF(192), AOM_ICDF(3061), AOM_ICDF(3557), AOM_ICDF(8701),
+          AOM_ICDF(9762), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(11969), AOM_ICDF(15846), AOM_ICDF(25660), AOM_ICDF(26667),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11713), AOM_ICDF(15794), AOM_ICDF(25737), AOM_ICDF(26760),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9281), AOM_ICDF(12675), AOM_ICDF(23181), AOM_ICDF(24351),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7105), AOM_ICDF(8757), AOM_ICDF(18383), AOM_ICDF(19437),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4289), AOM_ICDF(4579), AOM_ICDF(11353), AOM_ICDF(11792),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1857), AOM_ICDF(1874), AOM_ICDF(4695), AOM_ICDF(4777),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(20929), AOM_ICDF(22297), AOM_ICDF(29370), AOM_ICDF(29646),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17473), AOM_ICDF(18985), AOM_ICDF(28079), AOM_ICDF(28413),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13121), AOM_ICDF(14064), AOM_ICDF(24902), AOM_ICDF(25217),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9793), AOM_ICDF(10214), AOM_ICDF(20069), AOM_ICDF(20329),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5825), AOM_ICDF(5987), AOM_ICDF(13350), AOM_ICDF(13559),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2241), AOM_ICDF(2260), AOM_ICDF(5520), AOM_ICDF(5600),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(25921), AOM_ICDF(26891), AOM_ICDF(31632), AOM_ICDF(31729),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(18241), AOM_ICDF(19463), AOM_ICDF(29222), AOM_ICDF(29419),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11585), AOM_ICDF(12065), AOM_ICDF(23294), AOM_ICDF(23488),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6593), AOM_ICDF(6686), AOM_ICDF(16153), AOM_ICDF(16234),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3137), AOM_ICDF(3170), AOM_ICDF(8751), AOM_ICDF(8821),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1345), AOM_ICDF(1359), AOM_ICDF(3739), AOM_ICDF(3824),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(27713), AOM_ICDF(28504), AOM_ICDF(32068), AOM_ICDF(32132),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(19265), AOM_ICDF(20354), AOM_ICDF(29789), AOM_ICDF(29943),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11201), AOM_ICDF(11538), AOM_ICDF(22701), AOM_ICDF(22848),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6337), AOM_ICDF(6424), AOM_ICDF(15268), AOM_ICDF(15353),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3649), AOM_ICDF(3681), AOM_ICDF(9052), AOM_ICDF(9121),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1601), AOM_ICDF(1618), AOM_ICDF(4584), AOM_ICDF(4667),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(30913), AOM_ICDF(31044), AOM_ICDF(32635), AOM_ICDF(32640),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22081), AOM_ICDF(22261), AOM_ICDF(30452), AOM_ICDF(30477),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10561), AOM_ICDF(10625), AOM_ICDF(21535), AOM_ICDF(21568),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6081), AOM_ICDF(6130), AOM_ICDF(14369), AOM_ICDF(14423),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3777), AOM_ICDF(3809), AOM_ICDF(9156), AOM_ICDF(9225),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1857), AOM_ICDF(1875), AOM_ICDF(4936), AOM_ICDF(5018),
+         AOM_ICDF(32768) } } },
+    {   // Intra
+      { // Band 0
+        { AOM_ICDF(4672), AOM_ICDF(6927), AOM_ICDF(23534), AOM_ICDF(29846),
+          AOM_ICDF(30928), AOM_ICDF(32768) },
+        { AOM_ICDF(3776), AOM_ICDF(6784), AOM_ICDF(18075), AOM_ICDF(25863),
+          AOM_ICDF(27926), AOM_ICDF(32768) },
+        { AOM_ICDF(1344), AOM_ICDF(5588), AOM_ICDF(12166), AOM_ICDF(20966),
+          AOM_ICDF(23504), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(19393), AOM_ICDF(22016), AOM_ICDF(31280), AOM_ICDF(31444),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21185), AOM_ICDF(24329), AOM_ICDF(31706), AOM_ICDF(31865),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20673), AOM_ICDF(23240), AOM_ICDF(31186), AOM_ICDF(31379),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17857), AOM_ICDF(20035), AOM_ICDF(29594), AOM_ICDF(29889),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13633), AOM_ICDF(14929), AOM_ICDF(24883), AOM_ICDF(25337),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7873), AOM_ICDF(8416), AOM_ICDF(17452), AOM_ICDF(17886),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(25665), AOM_ICDF(27145), AOM_ICDF(32256), AOM_ICDF(32314),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21057), AOM_ICDF(22826), AOM_ICDF(31465), AOM_ICDF(31576),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13633), AOM_ICDF(14885), AOM_ICDF(27873), AOM_ICDF(28088),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8769), AOM_ICDF(9515), AOM_ICDF(21941), AOM_ICDF(22248),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(6209), AOM_ICDF(6594), AOM_ICDF(15598), AOM_ICDF(15950),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(1985), AOM_ICDF(2014), AOM_ICDF(6855), AOM_ICDF(6931),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(26817), AOM_ICDF(27824), AOM_ICDF(32362), AOM_ICDF(32399),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21185), AOM_ICDF(22321), AOM_ICDF(31389), AOM_ICDF(31466),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13761), AOM_ICDF(14154), AOM_ICDF(27163), AOM_ICDF(27245),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8897), AOM_ICDF(9011), AOM_ICDF(20600), AOM_ICDF(20659),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4673), AOM_ICDF(4774), AOM_ICDF(15044), AOM_ICDF(15131),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(194), AOM_ICDF(384), AOM_ICDF(479),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(28865), AOM_ICDF(29687), AOM_ICDF(32655), AOM_ICDF(32667),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(23233), AOM_ICDF(24218), AOM_ICDF(32080), AOM_ICDF(32118),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15041), AOM_ICDF(15444), AOM_ICDF(28787), AOM_ICDF(28845),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9921), AOM_ICDF(10248), AOM_ICDF(22818), AOM_ICDF(22944),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7745), AOM_ICDF(7866), AOM_ICDF(16591), AOM_ICDF(16702),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(194), AOM_ICDF(384), AOM_ICDF(479),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(31169), AOM_ICDF(31559), AOM_ICDF(32741), AOM_ICDF(32744),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(24769), AOM_ICDF(25583), AOM_ICDF(32347), AOM_ICDF(32370),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15937), AOM_ICDF(16169), AOM_ICDF(29120), AOM_ICDF(29152),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7489), AOM_ICDF(7578), AOM_ICDF(22647), AOM_ICDF(22677),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7617), AOM_ICDF(7689), AOM_ICDF(19849), AOM_ICDF(19887),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2113), AOM_ICDF(2183), AOM_ICDF(7202), AOM_ICDF(7377),
+         AOM_ICDF(32768) } } } },
+  {     // UV plane
+    {   // Inter
+      { // Band 0
+        { AOM_ICDF(23232), AOM_ICDF(24301), AOM_ICDF(30231), AOM_ICDF(31582),
+          AOM_ICDF(32091), AOM_ICDF(32768) },
+        { AOM_ICDF(7872), AOM_ICDF(11041), AOM_ICDF(22542), AOM_ICDF(27086),
+          AOM_ICDF(29145), AOM_ICDF(32768) },
+        { AOM_ICDF(1344), AOM_ICDF(3989), AOM_ICDF(18125), AOM_ICDF(25340),
+          AOM_ICDF(27820), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(15937), AOM_ICDF(29000), AOM_ICDF(32210), AOM_ICDF(32434),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(12353), AOM_ICDF(26626), AOM_ICDF(31533), AOM_ICDF(31993),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11457), AOM_ICDF(29187), AOM_ICDF(30896), AOM_ICDF(31750),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(21278), AOM_ICDF(28169), AOM_ICDF(29764),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(7489), AOM_ICDF(8855), AOM_ICDF(13365), AOM_ICDF(15620),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4289), AOM_ICDF(4833), AOM_ICDF(8572), AOM_ICDF(10108),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(25025), AOM_ICDF(30783), AOM_ICDF(32603), AOM_ICDF(32666),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(24385), AOM_ICDF(29586), AOM_ICDF(31803), AOM_ICDF(32142),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22337), AOM_ICDF(23002), AOM_ICDF(27573), AOM_ICDF(27903),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10945), AOM_ICDF(12336), AOM_ICDF(21900), AOM_ICDF(22590),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(8257), AOM_ICDF(8830), AOM_ICDF(19986), AOM_ICDF(20298),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10945), AOM_ICDF(10990), AOM_ICDF(18660), AOM_ICDF(18701),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(29761), AOM_ICDF(31473), AOM_ICDF(32693), AOM_ICDF(32715),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(20417), AOM_ICDF(24512), AOM_ICDF(31394), AOM_ICDF(31650),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(11713), AOM_ICDF(13283), AOM_ICDF(25819), AOM_ICDF(26206),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13121), AOM_ICDF(14099), AOM_ICDF(21909), AOM_ICDF(22514),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(248), AOM_ICDF(9546), AOM_ICDF(9614),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2497), AOM_ICDF(2524), AOM_ICDF(7050), AOM_ICDF(7125),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(30657), AOM_ICDF(31885), AOM_ICDF(32691), AOM_ICDF(32715),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(19393), AOM_ICDF(26050), AOM_ICDF(31698), AOM_ICDF(31988),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(15809), AOM_ICDF(15863), AOM_ICDF(24985), AOM_ICDF(25008),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(23489), AOM_ICDF(28138), AOM_ICDF(32751), AOM_ICDF(32756),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16449), AOM_ICDF(16450), AOM_ICDF(16545), AOM_ICDF(16593),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2369), AOM_ICDF(2395), AOM_ICDF(6822), AOM_ICDF(6898),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(32705), AOM_ICDF(32744), AOM_ICDF(32766), AOM_ICDF(32767),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21953), AOM_ICDF(24962), AOM_ICDF(32156), AOM_ICDF(32246),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(13121), AOM_ICDF(15358), AOM_ICDF(26284), AOM_ICDF(26835),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(7417), AOM_ICDF(20132), AOM_ICDF(20885),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4417), AOM_ICDF(4939), AOM_ICDF(15104), AOM_ICDF(15535),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2625), AOM_ICDF(2680), AOM_ICDF(8218), AOM_ICDF(8338),
+         AOM_ICDF(32768) } } },
+    {   // Inter
+      { // Band 0
+        { AOM_ICDF(25280), AOM_ICDF(25678), AOM_ICDF(32446), AOM_ICDF(32622),
+          AOM_ICDF(32724), AOM_ICDF(32768) },
+        { AOM_ICDF(10560), AOM_ICDF(11822), AOM_ICDF(28682), AOM_ICDF(29919),
+          AOM_ICDF(31276), AOM_ICDF(32768) },
+        { AOM_ICDF(3264), AOM_ICDF(5170), AOM_ICDF(21779), AOM_ICDF(24026),
+          AOM_ICDF(27905), AOM_ICDF(32768) } },
+      { // Band 1
+        {AOM_ICDF(24257), AOM_ICDF(30554), AOM_ICDF(32719), AOM_ICDF(32738),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17217), AOM_ICDF(27413), AOM_ICDF(32617), AOM_ICDF(32667),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22977), AOM_ICDF(27600), AOM_ICDF(32482), AOM_ICDF(32552),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16833), AOM_ICDF(24360), AOM_ICDF(30746), AOM_ICDF(31293),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(17089), AOM_ICDF(20060), AOM_ICDF(28880), AOM_ICDF(29370),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(10945), AOM_ICDF(11009), AOM_ICDF(21900), AOM_ICDF(21932),
+         AOM_ICDF(32768) } },
+      { // Band 2
+        {AOM_ICDF(27201), AOM_ICDF(30217), AOM_ICDF(32736), AOM_ICDF(32745),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22721), AOM_ICDF(27676), AOM_ICDF(32749), AOM_ICDF(32754),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5057), AOM_ICDF(12431), AOM_ICDF(25246), AOM_ICDF(26620),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(321), AOM_ICDF(22016), AOM_ICDF(22048),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5313), AOM_ICDF(5363), AOM_ICDF(13839), AOM_ICDF(13894),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2625), AOM_ICDF(2652), AOM_ICDF(7276), AOM_ICDF(7351),
+         AOM_ICDF(32768) } },
+      { // Band 3
+        {AOM_ICDF(27713), AOM_ICDF(30739), AOM_ICDF(32759), AOM_ICDF(32762),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(26177), AOM_ICDF(30430), AOM_ICDF(32756), AOM_ICDF(32760),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(384), AOM_ICDF(32706), AOM_ICDF(32707),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(9409), AOM_ICDF(9528), AOM_ICDF(21591), AOM_ICDF(21646),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(194), AOM_ICDF(384), AOM_ICDF(479),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2881), AOM_ICDF(2913), AOM_ICDF(8427), AOM_ICDF(8498),
+         AOM_ICDF(32768) } },
+      { // Band 4
+        {AOM_ICDF(28993), AOM_ICDF(31156), AOM_ICDF(32747), AOM_ICDF(32753),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(25153), AOM_ICDF(28701), AOM_ICDF(32754), AOM_ICDF(32758),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(16449), AOM_ICDF(16544), AOM_ICDF(32737), AOM_ICDF(32738),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(321), AOM_ICDF(22016), AOM_ICDF(22048),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(193), AOM_ICDF(194), AOM_ICDF(384), AOM_ICDF(479),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(2753), AOM_ICDF(2789), AOM_ICDF(8909), AOM_ICDF(8979),
+         AOM_ICDF(32768) } },
+      { // Band 5
+        {AOM_ICDF(30785), AOM_ICDF(32088), AOM_ICDF(32765), AOM_ICDF(32766),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(22977), AOM_ICDF(26623), AOM_ICDF(32750), AOM_ICDF(32754),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(21953), AOM_ICDF(21954), AOM_ICDF(22017), AOM_ICDF(22049),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(5697), AOM_ICDF(7486), AOM_ICDF(20238), AOM_ICDF(21009),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(4929), AOM_ICDF(5579), AOM_ICDF(16402), AOM_ICDF(16866),
+         AOM_ICDF(32768) },
+        {AOM_ICDF(3009), AOM_ICDF(3246), AOM_ICDF(10158), AOM_ICDF(10533),
+         AOM_ICDF(32768) } } } }
+};
+#endif  // CONFIG_NEW_TOKENSET
+
+/* clang-format on */
+
+static void extend_to_full_distribution(aom_prob *probs, aom_prob p) {
+  assert(p != 0);
+  memcpy(probs, av1_pareto8_full[p - 1], MODEL_NODES * sizeof(aom_prob));
+}
+
+void av1_model_to_full_probs(const aom_prob *model, aom_prob *full) {
+  if (full != model)
+    memcpy(full, model, sizeof(aom_prob) * UNCONSTRAINED_NODES);
+  extend_to_full_distribution(&full[UNCONSTRAINED_NODES], model[PIVOT_NODE]);
+}
+
+#if CONFIG_NEW_TOKENSET
+
+static void build_tail_cdfs(aom_cdf_prob cdf_tail[CDF_SIZE(ENTROPY_TOKENS)],
+                            aom_cdf_prob cdf_head[CDF_SIZE(ENTROPY_TOKENS)],
+                            int band_zero) {
+  int probNZ, prob1, prob_idx, i;
+  int phead[HEAD_TOKENS + 1], sum;
+  const int is_dc = !!band_zero;
+  aom_cdf_prob prev_cdf;
+  prev_cdf = 0;
+  for (i = 0; i < HEAD_TOKENS + is_dc; ++i) {
+    phead[i] = AOM_ICDF(cdf_head[i]) - prev_cdf;
+    prev_cdf = AOM_ICDF(cdf_head[i]);
+  }
+  // Do the tail
+  probNZ = CDF_PROB_TOP - phead[ZERO_TOKEN + is_dc] - (is_dc ? phead[0] : 0);
+  prob1 = phead[is_dc + ONE_TOKEN_EOB] + phead[is_dc + ONE_TOKEN_NEOB];
+  prob_idx =
+      AOMMIN(COEFF_PROB_MODELS - 1, AOMMAX(0, ((256 * prob1) / probNZ) - 1));
+
+  sum = 0;
+  for (i = 0; i < TAIL_TOKENS; ++i) {
+    sum += av1_pareto8_tail_probs[prob_idx][i];
+    cdf_tail[i] = AOM_ICDF(sum);
+  }
+}
+
+static void build_head_cdfs(const aom_prob *pdf_model,
+                            const aom_prob *blockz_model,
+                            aom_cdf_prob cdf_head[ENTROPY_TOKENS + 1]) {
+  int i, p, p1, p2, phead[6], prob_NZ, prob_EOB_1, prob_EOB_2p, prob_NEOB_1,
+      prob_NEOB_2p;
+  int prob8_blocknz;
+  // We have the first coefficient position and so an extended CDF
+  const int is_dc = blockz_model != NULL;
+  const int last_head_val = HEAD_TOKENS - 1 + is_dc;
+
+  assert(pdf_model != NULL);
+  assert(pdf_model[2] != 0);
+
+  /* FIXME: maintain true CDF counts. */
+
+  /* Values are 0=BLOCK_ZERO 1=ZERO_TOKEN, 2=ONE_TOKEN_EOB
+     3=ONE_TOKEN_NEOB, 4=TWO_TOKEN_PLUS_EOB, 5=TWO_TOKEN_PLUS_NEOB
+     */
+  // Block zero probability
+  if (is_dc) {
+    phead[0] =
+        ((*blockz_model) << (CDF_PROB_BITS - 8)) + (1 << (CDF_PROB_BITS - 9));
+    phead[0] = AOMMIN(CDF_PROB_TOP - (HEAD_TOKENS + 1), AOMMAX(1, phead[0]));
+  }
+
+  // Will scale the remaining probabilities by the probability of the block
+  // being non-zero
+  prob8_blocknz = is_dc ? (256 - *blockz_model) : 256;
+
+  // Probability of zero
+  phead[is_dc + ZERO_TOKEN] =
+      (pdf_model[1] << (CDF_PROB_BITS - 8)) + (1 << (CDF_PROB_BITS - 9));
+
+  // Will scale the non-zero values
+  prob_NZ = CDF_PROB_TOP - phead[is_dc + ZERO_TOKEN];
+
+  // Will scale the EOBs by the probability of and EOB_TOKEN ..
+  prob_EOB_1 =
+      (pdf_model[0] << (CDF_PROB_BITS - 8)) + (1 << (CDF_PROB_BITS - 9));
+  // .. use a lower probability of EOB for larger values
+  prob_EOB_2p = prob_EOB_1 / 2;
+
+  prob_NEOB_1 = CDF_PROB_TOP - prob_EOB_1;
+  prob_NEOB_2p = CDF_PROB_TOP - prob_EOB_2p;
+  if (prob_NZ == 0 || prob_NZ == CDF_PROB_TOP) abort();
+  if (prob_EOB_1 == 0 || prob_EOB_1 == CDF_PROB_TOP) abort();
+  if (prob_EOB_2p == 0 || prob_EOB_2p == CDF_PROB_TOP) abort();
+
+  // ONE_CONTEXT_NODE prob
+  p = (pdf_model[2] << (CDF_PROB_BITS - 8)) + (1 << (CDF_PROB_BITS - 9));
+  // Scale by the non-zero factor to get the probability of token = 1
+  p1 = ROUND_POWER_OF_TWO(prob_NZ * p, 15);
+
+  // Scale by the EOB factors
+  phead[is_dc + ONE_TOKEN_EOB] = ROUND_POWER_OF_TWO(p1 * prob_EOB_1, 15);
+  phead[is_dc + ONE_TOKEN_NEOB] = ROUND_POWER_OF_TWO(p1 * prob_NEOB_1, 15);
+
+  // Probability token is 2 or more
+  p2 = CDF_PROB_TOP - p1 - phead[is_dc + ZERO_TOKEN];
+
+  phead[is_dc + TWO_TOKEN_PLUS_EOB] = ROUND_POWER_OF_TWO(p2 * prob_EOB_2p, 15);
+  phead[is_dc + TWO_TOKEN_PLUS_NEOB] =
+      ROUND_POWER_OF_TWO(p2 * prob_NEOB_2p, 15);
+
+  // Now use block non-zerp prob to scale the values
+  for (i = is_dc; i < last_head_val; ++i) {
+    phead[i] = (prob8_blocknz * phead[i] + 128) >> 8;
+  }
+
+  for (i = 0; i < last_head_val; ++i) {
+    int c0;
+    c0 = i > 0 ? AOM_ICDF(cdf_head[i - 1]) : 0;
+    p = AOMMAX(1, AOMMIN(CDF_PROB_TOP - (last_head_val - i) - c0, phead[i]));
+    cdf_head[i] = AOM_ICDF(c0 + p);
+  }
+  cdf_head[last_head_val] = AOM_ICDF(CDF_PROB_TOP);
+}
+
+static void av1_default_coef_cdfs(FRAME_CONTEXT *fc) {
+  int i, j, k, l;
+  for (i = 0; i < PLANE_TYPES; ++i)
+    for (j = 0; j < REF_TYPES; ++j)
+      for (k = 0; k < COEF_BANDS; ++k)
+        for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+#if CONFIG_CB4X4
+          av1_copy(fc->coef_head_cdfs[TX_2X2][i][j][k][l],
+                   default_coef_head_cdf_4x4[i][j][k][l]);
+#endif
+          av1_copy(fc->coef_head_cdfs[TX_4X4][i][j][k][l],
+                   default_coef_head_cdf_4x4[i][j][k][l]);
+          av1_copy(fc->coef_head_cdfs[TX_8X8][i][j][k][l],
+                   default_coef_head_cdf_8x8[i][j][k][l]);
+          av1_copy(fc->coef_head_cdfs[TX_16X16][i][j][k][l],
+                   default_coef_head_cdf_16x16[i][j][k][l]);
+          av1_copy(fc->coef_head_cdfs[TX_32X32][i][j][k][l],
+                   default_coef_head_cdf_32x32[i][j][k][l]);
+#if CONFIG_TX64X64
+          av1_copy(fc->coef_head_cdfs[TX_64X64][i][j][k][l],
+                   default_coef_head_cdf_32x32[i][j][k][l]);
+#endif
+        }
+}
+
+void av1_coef_head_cdfs(FRAME_CONTEXT *fc) {
+  TX_SIZE t;
+  int i, j, k, l;
+  for (t = 0; t < TX_SIZES; ++t)
+    for (i = 0; i < PLANE_TYPES; ++i)
+      for (j = 0; j < REF_TYPES; ++j)
+        for (k = 0; k < COEF_BANDS; ++k)
+          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+            build_head_cdfs(fc->coef_probs[t][i][j][k][l],
+                            k == 0 ? &fc->blockzero_probs[t][i][j][l] : NULL,
+                            fc->coef_head_cdfs[t][i][j][k][l]);
+          }
+}
+
+#elif CONFIG_EC_MULTISYMBOL
+static void build_token_cdfs(const aom_prob *pdf_model,
+                             aom_cdf_prob cdf[ENTROPY_TOKENS + 1]) {
+  int i, sum = 0;
+  assert(pdf_model[2] != 0);
+  for (i = 0; i < ENTROPY_TOKENS - 2; ++i) {
+    sum += av1_pareto8_token_probs[pdf_model[2] - 1][i];
+    cdf[i] = AOM_ICDF(sum);
+  }
+}
+#endif  // CONFIG_NEW_TOKENSET
+
+#if CONFIG_EC_MULTISYMBOL
+void av1_coef_pareto_cdfs(FRAME_CONTEXT *fc) {
+  /* Build the tail based on a Pareto distribution */
+  TX_SIZE t;
+  int i, j, k, l;
+  for (t = 0; t < TX_SIZES; ++t)
+    for (i = 0; i < PLANE_TYPES; ++i)
+      for (j = 0; j < REF_TYPES; ++j)
+        for (k = 0; k < COEF_BANDS; ++k)
+          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
+#if CONFIG_NEW_TOKENSET
+            build_tail_cdfs(fc->coef_tail_cdfs[t][i][j][k][l],
+                            fc->coef_head_cdfs[t][i][j][k][l], k == 0);
+#else
+            build_token_cdfs(fc->coef_probs[t][i][j][k][l],
+                             fc->coef_cdfs[t][i][j][k][l]);
+#endif
+}
+#endif
+
+void av1_default_coef_probs(AV1_COMMON *cm) {
+#if CONFIG_Q_ADAPT_PROBS
+  const int index = AOMMIN(
+      ROUND_POWER_OF_TWO(cm->base_qindex, 8 - QCTX_BIN_BITS), QCTX_BINS - 1);
+  av1_copy(cm->fc->coef_probs, default_qctx_coef_probs[index]);
+#else
+#if CONFIG_CB4X4
+  av1_copy(cm->fc->coef_probs[TX_2X2], default_coef_probs_4x4);
+#endif
+  av1_copy(cm->fc->coef_probs[TX_4X4], default_coef_probs_4x4);
+  av1_copy(cm->fc->coef_probs[TX_8X8], default_coef_probs_8x8);
+  av1_copy(cm->fc->coef_probs[TX_16X16], default_coef_probs_16x16);
+  av1_copy(cm->fc->coef_probs[TX_32X32], default_coef_probs_32x32);
+#if CONFIG_TX64X64
+  av1_copy(cm->fc->coef_probs[TX_64X64], default_coef_probs_64x64);
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_Q_ADAPT_PROBS
+#if CONFIG_NEW_TOKENSET
+  av1_copy(cm->fc->blockzero_probs, av1_default_blockzero_probs);
+#endif
+#if CONFIG_NEW_TOKENSET
+  /* Load the head tokens */
+  av1_default_coef_cdfs(cm->fc);
+#endif
+#if CONFIG_EC_MULTISYMBOL
+  av1_coef_pareto_cdfs(cm->fc);
+#endif  // CONFIG_EC_MULTISYMBOL
+}
+
+#if !CONFIG_LV_MAP
+static void adapt_coef_probs(AV1_COMMON *cm, TX_SIZE tx_size,
+                             unsigned int count_sat,
+                             unsigned int update_factor) {
+  const FRAME_CONTEXT *pre_fc = &cm->frame_contexts[cm->frame_context_idx];
+  av1_coeff_probs_model *const probs = cm->fc->coef_probs[tx_size];
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  const av1_coeff_probs_model *const pre_probs =
+      cm->partial_prob_update
+          ? (const av1_coeff_probs_model *)cm->starting_coef_probs[tx_size]
+          : pre_fc->coef_probs[tx_size];
+#else
+  const av1_coeff_probs_model *const pre_probs = pre_fc->coef_probs[tx_size];
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+  const av1_coeff_count_model *const counts =
+      (const av1_coeff_count_model *)cm->counts.coef[tx_size];
+  const unsigned int(*eob_counts)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
+      (const unsigned int(*)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS])
+          cm->counts.eob_branch[tx_size];
+#if CONFIG_NEW_TOKENSET
+  const av1_blockz_probs_model *const pre_blockz_probs =
+      pre_fc->blockzero_probs[tx_size];
+  av1_blockz_probs_model *const blockz_probs = cm->fc->blockzero_probs[tx_size];
+  const av1_blockz_count_model *const blockz_counts =
+      (const av1_blockz_count_model *)&cm->counts.blockz_count[tx_size][0];
+#endif
+  int i, j, k, l, m;
+#if CONFIG_RECT_TX
+  assert(!is_rect_tx(tx_size));
+#endif  // CONFIG_RECT_TX
+
+  for (i = 0; i < PLANE_TYPES; ++i)
+    for (j = 0; j < REF_TYPES; ++j)
+      for (k = 0; k < COEF_BANDS; ++k)
+        for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+          const int n0 = counts[i][j][k][l][ZERO_TOKEN];
+          const int n1 = counts[i][j][k][l][ONE_TOKEN];
+          const int n2 = counts[i][j][k][l][TWO_TOKEN];
+          const int neob = counts[i][j][k][l][EOB_MODEL_TOKEN];
+          const unsigned int branch_ct[UNCONSTRAINED_NODES][2] = {
+            { neob, eob_counts[i][j][k][l] - neob }, { n0, n1 + n2 }, { n1, n2 }
+          };
+          for (m = 0; m < UNCONSTRAINED_NODES; ++m)
+            probs[i][j][k][l][m] =
+                av1_merge_probs(pre_probs[i][j][k][l][m], branch_ct[m],
+                                count_sat, update_factor);
+        }
+
+#if CONFIG_NEW_TOKENSET
+  for (i = 0; i < PLANE_TYPES; ++i) {
+    for (j = 0; j < REF_TYPES; ++j) {
+      for (k = 0; k < BLOCKZ_CONTEXTS; ++k) {
+        const int n0 = blockz_counts[i][j][k][0];
+        const int n1 = blockz_counts[i][j][k][1];
+        const unsigned int branch_ct[2] = { n0, n1 };
+        blockz_probs[i][j][k] = av1_merge_probs(
+            pre_blockz_probs[i][j][k], branch_ct, count_sat, update_factor);
+      }
+    }
+  }
+#endif
+}
+#endif  // !CONFIG_LV_MAP
+
+void av1_adapt_coef_probs(AV1_COMMON *cm) {
+  unsigned int count_sat, update_factor;
+
+  if (!frame_is_intra_only(cm) && cm->last_frame_type == KEY_FRAME) {
+    update_factor = COEF_MAX_UPDATE_FACTOR_AFTER_KEY; /* adapt quickly */
+    count_sat = COEF_COUNT_SAT_AFTER_KEY;
+  } else {
+    update_factor = COEF_MAX_UPDATE_FACTOR;
+    count_sat = COEF_COUNT_SAT;
+  }
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  if (cm->partial_prob_update == 1) update_factor = COEF_MAX_UPDATE_FACTOR;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+#if CONFIG_LV_MAP
+  av1_adapt_txb_probs(cm, count_sat, update_factor);
+#else
+  TX_SIZE tx_size;
+  for (tx_size = 0; tx_size < TX_SIZES; tx_size++)
+    adapt_coef_probs(cm, tx_size, count_sat, update_factor);
+#endif
+}
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+void av1_partial_adapt_probs(AV1_COMMON *cm, int mi_row, int mi_col) {
+  (void)mi_row;
+  (void)mi_col;
+
+  if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+    cm->partial_prob_update = 1;
+    av1_adapt_coef_probs(cm);
+  }
+}
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+#if CONFIG_EC_ADAPT
+static void av1_average_cdf(aom_cdf_prob *cdf_ptr[], aom_cdf_prob *fc_cdf_ptr,
+                            int cdf_size, const int num_tiles) {
+  int i;
+  for (i = 0; i < cdf_size;) {
+    do {
+      int sum = 0;
+      int j;
+      assert(i < cdf_size);
+      for (j = 0; j < num_tiles; ++j) sum += AOM_ICDF(cdf_ptr[j][i]);
+      fc_cdf_ptr[i] = AOM_ICDF(sum / num_tiles);
+    } while (fc_cdf_ptr[i++] != AOM_ICDF(CDF_PROB_TOP));
+    // Zero symbol counts for the next frame
+    assert(i < cdf_size);
+    fc_cdf_ptr[i++] = 0;
+    // Skip trailing zeros until the start of the next CDF.
+    for (; i < cdf_size && fc_cdf_ptr[i] == 0; ++i) {
+    }
+  }
+}
+
+#define AVERAGE_TILE_CDFS(cname)                            \
+  for (i = 0; i < num_tiles; ++i)                           \
+    cdf_ptr[i] = (aom_cdf_prob *)&ec_ctxs[i]->cname;        \
+  fc_cdf_ptr = (aom_cdf_prob *)&fc->cname;                  \
+  cdf_size = (int)sizeof(fc->cname) / sizeof(aom_cdf_prob); \
+  av1_average_cdf(cdf_ptr, fc_cdf_ptr, cdf_size, num_tiles);
+
+void av1_average_tile_coef_cdfs(FRAME_CONTEXT *fc, FRAME_CONTEXT *ec_ctxs[],
+                                aom_cdf_prob *cdf_ptr[], const int num_tiles) {
+  int i, cdf_size;
+
+  aom_cdf_prob *fc_cdf_ptr;
+
+#if CONFIG_NEW_TOKENSET
+  AVERAGE_TILE_CDFS(coef_head_cdfs)
+  AVERAGE_TILE_CDFS(coef_tail_cdfs)
+#else
+  AVERAGE_TILE_CDFS(coef_cdfs)
+#endif
+}
+
+void av1_average_tile_mv_cdfs(FRAME_CONTEXT *fc, FRAME_CONTEXT *ec_ctxs[],
+                              aom_cdf_prob *cdf_ptr[], const int num_tiles) {
+  int i, k, cdf_size;
+
+  aom_cdf_prob *fc_cdf_ptr;
+
+#if CONFIG_REF_MV
+  int j;
+  for (j = 0; j < NMV_CONTEXTS; ++j) {
+    AVERAGE_TILE_CDFS(nmvc[j].joint_cdf)
+
+    for (k = 0; k < 2; ++k) {
+      AVERAGE_TILE_CDFS(nmvc[j].comps[k].class_cdf);
+      AVERAGE_TILE_CDFS(nmvc[j].comps[k].class0_fp_cdf);
+      AVERAGE_TILE_CDFS(nmvc[j].comps[k].fp_cdf);
+    }
+  }
+#else
+  AVERAGE_TILE_CDFS(nmvc.joint_cdf)
+
+  for (k = 0; k < 2; ++k) {
+    AVERAGE_TILE_CDFS(nmvc.comps[k].class_cdf)
+    AVERAGE_TILE_CDFS(nmvc.comps[k].class0_fp_cdf)
+    AVERAGE_TILE_CDFS(nmvc.comps[k].fp_cdf)
+  }
+#endif
+}
+
+void av1_average_tile_intra_cdfs(FRAME_CONTEXT *fc, FRAME_CONTEXT *ec_ctxs[],
+                                 aom_cdf_prob *cdf_ptr[], const int num_tiles) {
+  int i, cdf_size;
+
+  aom_cdf_prob *fc_cdf_ptr;
+
+  AVERAGE_TILE_CDFS(tx_size_cdf);
+
+#if CONFIG_VAR_TX
+// FIXME: txfm_partition probs
+#endif
+
+  // FIXME: skip probs
+
+  AVERAGE_TILE_CDFS(intra_ext_tx_cdf)
+  AVERAGE_TILE_CDFS(inter_ext_tx_cdf);
+
+  AVERAGE_TILE_CDFS(seg.tree_cdf)
+  AVERAGE_TILE_CDFS(uv_mode_cdf)
+
+  AVERAGE_TILE_CDFS(partition_cdf)
+
+#if CONFIG_DELTA_Q
+  AVERAGE_TILE_CDFS(delta_q_cdf)
+#if CONFIG_EXT_DELTA_Q
+  AVERAGE_TILE_CDFS(delta_lf_cdf)
+#endif
+#endif
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+  AVERAGE_TILE_CDFS(intra_filter_cdf)
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+#if CONFIG_FILTER_INTRA
+#endif  // CONFIG_FILTER_INTRA
+}
+
+void av1_average_tile_inter_cdfs(AV1_COMMON *cm, FRAME_CONTEXT *fc,
+                                 FRAME_CONTEXT *ec_ctxs[],
+                                 aom_cdf_prob *cdf_ptr[], const int num_tiles) {
+  int i, cdf_size;
+
+  aom_cdf_prob *fc_cdf_ptr;
+
+// FIXME: comp_inter_cdf not defined
+
+// FIXME: comp_ref_cdf and comp_bwd_ref not defined
+
+// FIXME: single_ref_cdf not defined
+
+#if CONFIG_REF_MV
+// FIXME: cdfs not defined for newmv_mode, zeromv_mode, drl_mode, new2mv_mode
+#else
+  AVERAGE_TILE_CDFS(inter_mode_cdf)
+#endif
+
+  // FIXME: cdfs not defined for motion_mode_prob, obmc_prob
+
+  // FIXME: cdfs not defined for super_tx
+
+  // FIXME: CONFIG_EXT_INTER cdfs not defined for inter_compound_mode,
+  // interintra_mode etc
+
+  AVERAGE_TILE_CDFS(y_mode_cdf)
+
+  if (cm->interp_filter == SWITCHABLE) {
+    AVERAGE_TILE_CDFS(switchable_interp_cdf)
+  }
+}
+
+#if CONFIG_PVQ
+// Averaging PVQ's expected values for symbol coding
+static void av1_average_pvq_ex(int *cxt_ptr[], int *fc_cxt_ptr, int cxt_size,
+                               const int num_tiles) {
+  int i, j;
+  for (i = 0; i < cxt_size; ++i) {
+    int sum = 0;
+    for (j = 0; j < num_tiles; ++j) sum += cxt_ptr[j][i];
+    fc_cxt_ptr[i] = sum / num_tiles;
+  }
+}
+
+#define AVERAGE_TILE_PVQ_EX(cname)                                        \
+  for (i = 0; i < num_tiles; ++i) cxt_ptr[i] = (int *)&ec_ctxs[i]->cname; \
+  fc_cxt_ptr = (int *)&fc->cname;                                         \
+  cxt_size = (int)sizeof(fc->cname) / sizeof(int);                        \
+  av1_average_pvq_ex(cxt_ptr, fc_cxt_ptr, cxt_size, num_tiles);
+
+void av1_default_pvq_probs(AV1_COMMON *cm) {
+  od_adapt_ctx *adapt = &cm->fc->pvq_context;
+
+  // Init with flat probabilities.
+  od_adapt_ctx_reset(adapt, 0);
+
+  // TODO(yushin): Prepare offline cdf and context table for PVQ,
+  // i.e. od_adapt_ctx, then load them from table,
+  // for example od_adapt_ctx default_pvq_context.
+  // Then do sth like this:
+  // av1_copy(cm->fc->pvq_context, default_pvq_context);
+}
+
+void av1_average_tile_pvq_cdfs(FRAME_CONTEXT *fc, FRAME_CONTEXT *ec_ctxs[],
+                               const int num_tiles) {
+  int i, j, cdf_size, cxt_size;
+
+  aom_cdf_prob *cdf_ptr[MAX_TILE_ROWS * MAX_TILE_COLS];
+  aom_cdf_prob *fc_cdf_ptr;
+  int *cxt_ptr[MAX_TILE_ROWS * MAX_TILE_COLS];
+  int *fc_cxt_ptr;
+
+  AVERAGE_TILE_PVQ_EX(pvq_context.ex_dc)
+  AVERAGE_TILE_PVQ_EX(pvq_context.ex_g)
+
+  for (j = 0; j < OD_NPLANES_MAX; j++) {
+    AVERAGE_TILE_CDFS(pvq_context.model_dc[j].cdf)
+  }
+
+  AVERAGE_TILE_CDFS(pvq_context.skip_cdf)
+
+  AVERAGE_TILE_PVQ_EX(pvq_context.pvq.pvq_codeword_ctx.pvq_adapt)
+  AVERAGE_TILE_CDFS(pvq_context.pvq.pvq_codeword_ctx.pvq_k1_cdf)
+  AVERAGE_TILE_CDFS(pvq_context.pvq.pvq_codeword_ctx.pvq_split_cdf)
+
+  for (j = 0; j < 3; j++) {
+    AVERAGE_TILE_CDFS(pvq_context.pvq.pvq_param_model[j].cdf)
+  }
+
+  AVERAGE_TILE_PVQ_EX(pvq_context.pvq.pvq_ext)
+  AVERAGE_TILE_PVQ_EX(pvq_context.pvq.pvq_exg)
+  AVERAGE_TILE_CDFS(pvq_context.pvq.pvq_gaintheta_cdf)
+  AVERAGE_TILE_CDFS(pvq_context.pvq.pvq_skip_dir_cdf)
+}
+#endif  // CONFIG_PVQ
+#endif  // CONFIG_EC_ADAPT
diff --git a/third_party/aom/av1/common/entropy.h b/third_party/aom/av1/common/entropy.h
new file mode 100644
index 0000000000..b02d41bff1
--- /dev/null
+++ b/third_party/aom/av1/common/entropy.h
@@ -0,0 +1,428 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_ENTROPY_H_
+#define AV1_COMMON_ENTROPY_H_
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/prob.h"
+
+#include "av1/common/common.h"
+#include "av1/common/common_data.h"
+#include "av1/common/enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define DIFF_UPDATE_PROB 252
+#define GROUP_DIFF_UPDATE_PROB 252
+
+#if CONFIG_Q_ADAPT_PROBS
+#define QCTX_BIN_BITS 2
+#define QCTX_BINS (1 << QCTX_BIN_BITS)
+#endif  // CONFIG_Q_ADAPT_PROBS
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+#define COEF_PROBS_BUFS 16
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+// Coefficient token alphabet
+#define ZERO_TOKEN 0        // 0     Extra Bits 0+0
+#define ONE_TOKEN 1         // 1     Extra Bits 0+1
+#define TWO_TOKEN 2         // 2     Extra Bits 0+1
+#define THREE_TOKEN 3       // 3     Extra Bits 0+1
+#define FOUR_TOKEN 4        // 4     Extra Bits 0+1
+#define CATEGORY1_TOKEN 5   // 5-6   Extra Bits 1+1
+#define CATEGORY2_TOKEN 6   // 7-10  Extra Bits 2+1
+#define CATEGORY3_TOKEN 7   // 11-18 Extra Bits 3+1
+#define CATEGORY4_TOKEN 8   // 19-34 Extra Bits 4+1
+#define CATEGORY5_TOKEN 9   // 35-66 Extra Bits 5+1
+#define CATEGORY6_TOKEN 10  // 67+   Extra Bits 14+1
+#define EOB_TOKEN 11        // EOB   Extra Bits 0+0
+#if CONFIG_NEW_TOKENSET
+#define NO_EOB 0           // Not an end-of-block
+#define EARLY_EOB 1        // End of block before the last position
+#define LAST_EOB 2         // End of block in the last position (implicit)
+#define BLOCK_Z_TOKEN 255  // block zero
+#define HEAD_TOKENS 5
+#define TAIL_TOKENS 9
+#define ONE_TOKEN_EOB 1
+#define ONE_TOKEN_NEOB 2
+#define TWO_TOKEN_PLUS_EOB 3
+#define TWO_TOKEN_PLUS_NEOB 4
+#endif
+#define ENTROPY_TOKENS 12
+
+#define ENTROPY_NODES 11
+
+#if CONFIG_LV_MAP
+#define TXB_SKIP_CONTEXTS 13
+#define SIG_COEF_CONTEXTS 20
+#define EOB_COEF_CONTEXTS 25
+#define COEFF_BASE_CONTEXTS 42
+#define DC_SIGN_CONTEXTS 3
+
+#define BR_TMP_OFFSET 12
+#define BR_REF_CAT 4
+#define LEVEL_CONTEXTS (BR_TMP_OFFSET * BR_REF_CAT)
+
+#define NUM_BASE_LEVELS 2
+#define COEFF_BASE_RANGE (15 - NUM_BASE_LEVELS)
+
+#define COEFF_CONTEXT_BITS 6
+#define COEFF_CONTEXT_MASK ((1 << COEFF_CONTEXT_BITS) - 1)
+#endif
+
+DECLARE_ALIGNED(16, extern const uint8_t, av1_pt_energy_class[ENTROPY_TOKENS]);
+
+#define CAT1_MIN_VAL 5
+#define CAT2_MIN_VAL 7
+#define CAT3_MIN_VAL 11
+#define CAT4_MIN_VAL 19
+#define CAT5_MIN_VAL 35
+#define CAT6_MIN_VAL 67
+
+// Extra bit probabilities.
+DECLARE_ALIGNED(16, extern const uint8_t, av1_cat1_prob[1]);
+DECLARE_ALIGNED(16, extern const uint8_t, av1_cat2_prob[2]);
+DECLARE_ALIGNED(16, extern const uint8_t, av1_cat3_prob[3]);
+DECLARE_ALIGNED(16, extern const uint8_t, av1_cat4_prob[4]);
+DECLARE_ALIGNED(16, extern const uint8_t, av1_cat5_prob[5]);
+DECLARE_ALIGNED(16, extern const uint8_t, av1_cat6_prob[18]);
+#if CONFIG_NEW_MULTISYMBOL
+extern const aom_cdf_prob *av1_cat1_cdf[];
+extern const aom_cdf_prob *av1_cat2_cdf[];
+extern const aom_cdf_prob *av1_cat3_cdf[];
+extern const aom_cdf_prob *av1_cat4_cdf[];
+extern const aom_cdf_prob *av1_cat5_cdf[];
+extern const aom_cdf_prob *av1_cat6_cdf[];
+#endif
+
+#define EOB_MODEL_TOKEN 3
+
+typedef struct {
+#if CONFIG_NEW_MULTISYMBOL
+  const aom_cdf_prob **cdf;
+#else
+  const aom_prob *prob;
+#endif
+  int len;
+  int base_val;
+  const int16_t *cost;
+} av1_extra_bit;
+
+// indexed by token value
+extern const av1_extra_bit av1_extra_bits[ENTROPY_TOKENS];
+
+static INLINE int av1_get_cat6_extrabits_size(TX_SIZE tx_size,
+                                              aom_bit_depth_t bit_depth) {
+  tx_size = txsize_sqr_up_map[tx_size];
+#if CONFIG_TX64X64
+  // TODO(debargha): Does TX_64X64 require an additional extrabit?
+  if (tx_size > TX_32X32) tx_size = TX_32X32;
+#endif
+#if CONFIG_CB4X4
+  int tx_offset = (tx_size < TX_4X4) ? 0 : (int)(tx_size - TX_4X4);
+#else
+  int tx_offset = (int)(tx_size - TX_4X4);
+#endif
+  int bits = (int)bit_depth + 3 + tx_offset;
+#if CONFIG_NEW_MULTISYMBOL
+  // Round up
+  bits = AOMMIN((int)sizeof(av1_cat6_prob), ((bits + 3) & ~3));
+#endif
+  assert(bits <= (int)sizeof(av1_cat6_prob));
+  return bits;
+}
+
+#define DCT_MAX_VALUE 16384
+#if CONFIG_HIGHBITDEPTH
+#define DCT_MAX_VALUE_HIGH10 65536
+#define DCT_MAX_VALUE_HIGH12 262144
+#endif  // CONFIG_HIGHBITDEPTH
+
+/* Coefficients are predicted via a 3-dimensional probability table. */
+
+#define REF_TYPES 2  // intra=0, inter=1
+
+/* Middle dimension reflects the coefficient position within the transform. */
+#define COEF_BANDS 6
+
+/* Inside dimension is measure of nearby complexity, that reflects the energy
+   of nearby coefficients are nonzero.  For the first coefficient (DC, unless
+   block type is 0), we look at the (already encoded) blocks above and to the
+   left of the current block.  The context index is then the number (0,1,or 2)
+   of these blocks having nonzero coefficients.
+   After decoding a coefficient, the measure is determined by the size of the
+   most recently decoded coefficient.
+   Note that the intuitive meaning of this measure changes as coefficients
+   are decoded, e.g., prior to the first token, a zero means that my neighbors
+   are empty while, after the first token, because of the use of end-of-block,
+   a zero means we just decoded a zero and hence guarantees that a non-zero
+   coefficient will appear later in this block.  However, this shift
+   in meaning is perfectly OK because our context depends also on the
+   coefficient band (and since zigzag positions 0, 1, and 2 are in
+   distinct bands). */
+
+#define COEFF_CONTEXTS 6
+#if CONFIG_EC_MULTISYMBOL
+#define BLOCKZ_CONTEXTS 3
+#endif
+#define COEFF_CONTEXTS0 3  // for band 0
+#define BAND_COEFF_CONTEXTS(band) \
+  ((band) == 0 ? COEFF_CONTEXTS0 : COEFF_CONTEXTS)
+
+// #define ENTROPY_STATS
+
+typedef unsigned int av1_coeff_count[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS]
+                                    [ENTROPY_TOKENS];
+typedef unsigned int av1_coeff_stats[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS]
+                                    [ENTROPY_NODES][2];
+
+#define SUBEXP_PARAM 4   /* Subexponential code parameter */
+#define MODULUS_PARAM 13 /* Modulus parameter */
+
+struct AV1Common;
+struct frame_contexts;
+void av1_default_coef_probs(struct AV1Common *cm);
+void av1_adapt_coef_probs(struct AV1Common *cm);
+#if CONFIG_EC_ADAPT
+void av1_adapt_coef_cdfs(struct AV1Common *cm, struct frame_contexts *pre_fc);
+#endif
+#if CONFIG_SUBFRAME_PROB_UPDATE
+void av1_partial_adapt_probs(struct AV1Common *cm, int mi_row, int mi_col);
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+// This is the index in the scan order beyond which all coefficients for
+// 8x8 transform and above are in the top band.
+// This macro is currently unused but may be used by certain implementations
+#define MAXBAND_INDEX 21
+
+DECLARE_ALIGNED(16, extern const uint8_t,
+                av1_coefband_trans_8x8plus[MAX_TX_SQUARE]);
+DECLARE_ALIGNED(16, extern const uint8_t, av1_coefband_trans_4x8_8x4[32]);
+DECLARE_ALIGNED(16, extern const uint8_t, av1_coefband_trans_4x4[16]);
+
+DECLARE_ALIGNED(16, extern const uint16_t, band_count_table[TX_SIZES_ALL][8]);
+DECLARE_ALIGNED(16, extern const uint16_t,
+                band_cum_count_table[TX_SIZES_ALL][8]);
+
+static INLINE const uint8_t *get_band_translate(TX_SIZE tx_size) {
+  switch (tx_size) {
+    case TX_4X4: return av1_coefband_trans_4x4;
+    case TX_8X4:
+    case TX_4X8: return av1_coefband_trans_4x8_8x4;
+    default: return av1_coefband_trans_8x8plus;
+  }
+}
+
+// 128 lists of probabilities are stored for the following ONE node probs:
+// 1, 3, 5, 7, ..., 253, 255
+// In between probabilities are interpolated linearly
+
+#define COEFF_PROB_MODELS 255
+
+#define UNCONSTRAINED_NODES 3
+
+#define PIVOT_NODE 2  // which node is pivot
+
+#define MODEL_NODES (ENTROPY_NODES - UNCONSTRAINED_NODES)
+#define TAIL_NODES (MODEL_NODES + 1)
+extern const aom_tree_index av1_coef_con_tree[TREE_SIZE(ENTROPY_TOKENS)];
+extern const aom_prob av1_pareto8_full[COEFF_PROB_MODELS][MODEL_NODES];
+
+typedef aom_prob av1_coeff_probs_model[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS]
+                                      [UNCONSTRAINED_NODES];
+
+typedef unsigned int av1_coeff_count_model[REF_TYPES][COEF_BANDS]
+                                          [COEFF_CONTEXTS]
+                                          [UNCONSTRAINED_NODES + 1];
+
+void av1_model_to_full_probs(const aom_prob *model, aom_prob *full);
+
+#if CONFIG_EC_MULTISYMBOL
+typedef aom_cdf_prob coeff_cdf_model[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS]
+                                    [CDF_SIZE(ENTROPY_TOKENS)];
+typedef aom_prob av1_blockz_probs_model[REF_TYPES][BLOCKZ_CONTEXTS];
+typedef unsigned int av1_blockz_count_model[REF_TYPES][BLOCKZ_CONTEXTS][2];
+extern const aom_cdf_prob av1_pareto8_token_probs[COEFF_PROB_MODELS]
+                                                 [ENTROPY_TOKENS - 2];
+extern const aom_cdf_prob av1_pareto8_tail_probs[COEFF_PROB_MODELS]
+                                                [ENTROPY_TOKENS - 3];
+struct frame_contexts;
+#if CONFIG_NEW_TOKENSET
+void av1_coef_head_cdfs(struct frame_contexts *fc);
+#endif
+void av1_coef_pareto_cdfs(struct frame_contexts *fc);
+#endif  // CONFIG_EC_MULTISYMBOL
+
+typedef char ENTROPY_CONTEXT;
+
+static INLINE int combine_entropy_contexts(ENTROPY_CONTEXT a,
+                                           ENTROPY_CONTEXT b) {
+  return (a != 0) + (b != 0);
+}
+
+static INLINE int get_entropy_context(TX_SIZE tx_size, const ENTROPY_CONTEXT *a,
+                                      const ENTROPY_CONTEXT *l) {
+  ENTROPY_CONTEXT above_ec = 0, left_ec = 0;
+
+#if CONFIG_CB4X4
+  switch (tx_size) {
+    case TX_2X2:
+      above_ec = a[0] != 0;
+      left_ec = l[0] != 0;
+      break;
+    case TX_4X4:
+      above_ec = !!*(const uint16_t *)a;
+      left_ec = !!*(const uint16_t *)l;
+      break;
+    case TX_4X8:
+      above_ec = !!*(const uint16_t *)a;
+      left_ec = !!*(const uint32_t *)l;
+      break;
+    case TX_8X4:
+      above_ec = !!*(const uint32_t *)a;
+      left_ec = !!*(const uint16_t *)l;
+      break;
+    case TX_8X8:
+      above_ec = !!*(const uint32_t *)a;
+      left_ec = !!*(const uint32_t *)l;
+      break;
+    case TX_8X16:
+      above_ec = !!*(const uint32_t *)a;
+      left_ec = !!*(const uint64_t *)l;
+      break;
+    case TX_16X8:
+      above_ec = !!*(const uint64_t *)a;
+      left_ec = !!*(const uint32_t *)l;
+      break;
+    case TX_16X16:
+      above_ec = !!*(const uint64_t *)a;
+      left_ec = !!*(const uint64_t *)l;
+      break;
+    case TX_16X32:
+      above_ec = !!*(const uint64_t *)a;
+      left_ec = !!(*(const uint64_t *)l | *(const uint64_t *)(l + 8));
+      break;
+    case TX_32X16:
+      above_ec = !!(*(const uint64_t *)a | *(const uint64_t *)(a + 8));
+      left_ec = !!*(const uint64_t *)l;
+      break;
+    case TX_32X32:
+      above_ec = !!(*(const uint64_t *)a | *(const uint64_t *)(a + 8));
+      left_ec = !!(*(const uint64_t *)l | *(const uint64_t *)(l + 8));
+      break;
+    default: assert(0 && "Invalid transform size."); break;
+  }
+  return combine_entropy_contexts(above_ec, left_ec);
+#endif
+
+  switch (tx_size) {
+    case TX_4X4:
+      above_ec = a[0] != 0;
+      left_ec = l[0] != 0;
+      break;
+    case TX_4X8:
+      above_ec = a[0] != 0;
+      left_ec = !!*(const uint16_t *)l;
+      break;
+    case TX_8X4:
+      above_ec = !!*(const uint16_t *)a;
+      left_ec = l[0] != 0;
+      break;
+    case TX_8X16:
+      above_ec = !!*(const uint16_t *)a;
+      left_ec = !!*(const uint32_t *)l;
+      break;
+    case TX_16X8:
+      above_ec = !!*(const uint32_t *)a;
+      left_ec = !!*(const uint16_t *)l;
+      break;
+    case TX_16X32:
+      above_ec = !!*(const uint32_t *)a;
+      left_ec = !!*(const uint64_t *)l;
+      break;
+    case TX_32X16:
+      above_ec = !!*(const uint64_t *)a;
+      left_ec = !!*(const uint32_t *)l;
+      break;
+    case TX_8X8:
+      above_ec = !!*(const uint16_t *)a;
+      left_ec = !!*(const uint16_t *)l;
+      break;
+    case TX_16X16:
+      above_ec = !!*(const uint32_t *)a;
+      left_ec = !!*(const uint32_t *)l;
+      break;
+    case TX_32X32:
+      above_ec = !!*(const uint64_t *)a;
+      left_ec = !!*(const uint64_t *)l;
+      break;
+#if CONFIG_TX64X64
+    case TX_64X64:
+      above_ec = !!(*(const uint64_t *)a | *(const uint64_t *)(a + 8));
+      left_ec = !!(*(const uint64_t *)l | *(const uint64_t *)(l + 8));
+      break;
+#endif  // CONFIG_TX64X64
+    default: assert(0 && "Invalid transform size."); break;
+  }
+  return combine_entropy_contexts(above_ec, left_ec);
+}
+
+#define COEF_COUNT_SAT 24
+#define COEF_MAX_UPDATE_FACTOR 112
+#define COEF_COUNT_SAT_AFTER_KEY 24
+#define COEF_MAX_UPDATE_FACTOR_AFTER_KEY 128
+
+#if CONFIG_ADAPT_SCAN
+#define ADAPT_SCAN_UPDATE_RATE_16 (1 << 13)
+#endif
+
+static INLINE aom_prob av1_merge_probs(aom_prob pre_prob,
+                                       const unsigned int ct[2],
+                                       unsigned int count_sat,
+                                       unsigned int max_update_factor) {
+  return merge_probs(pre_prob, ct, count_sat, max_update_factor);
+}
+
+static INLINE aom_prob av1_mode_mv_merge_probs(aom_prob pre_prob,
+                                               const unsigned int ct[2]) {
+  return mode_mv_merge_probs(pre_prob, ct);
+}
+
+#if CONFIG_EC_ADAPT
+void av1_average_tile_coef_cdfs(struct frame_contexts *fc,
+                                struct frame_contexts *ec_ctxs[],
+                                aom_cdf_prob *cdf_ptrs[], int num_tiles);
+void av1_average_tile_mv_cdfs(struct frame_contexts *fc,
+                              struct frame_contexts *ec_ctxs[],
+                              aom_cdf_prob *cdf_ptrs[], int num_tiles);
+void av1_average_tile_intra_cdfs(struct frame_contexts *fc,
+                                 struct frame_contexts *ec_ctxs[],
+                                 aom_cdf_prob *cdf_ptrs[], int num_tiles);
+void av1_average_tile_inter_cdfs(struct AV1Common *cm,
+                                 struct frame_contexts *fc,
+                                 struct frame_contexts *ec_ctxs[],
+                                 aom_cdf_prob *cdf_ptrs[], int num_tiles);
+#if CONFIG_PVQ
+void av1_default_pvq_probs(struct AV1Common *cm);
+void av1_average_tile_pvq_cdfs(struct frame_contexts *fc,
+                               struct frame_contexts *ec_ctxs[], int num_tiles);
+#endif  // CONFIG_PVQ
+#endif  // CONFIG_EC_ADAPT
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_ENTROPY_H_
diff --git a/third_party/aom/av1/common/entropymode.c b/third_party/aom/av1/common/entropymode.c
new file mode 100644
index 0000000000..0fcf762d16
--- /dev/null
+++ b/third_party/aom/av1/common/entropymode.c
@@ -0,0 +1,3792 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_mem/aom_mem.h"
+
+#include "av1/common/reconinter.h"
+#include "av1/common/scan.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/seg_common.h"
+
+#if CONFIG_LV_MAP
+const aom_prob default_txb_skip[TX_SIZES][TXB_SKIP_CONTEXTS] = {
+#if CONFIG_CB4X4
+  { 252, 71, 126, 184, 178, 218, 251, 49, 133, 221, 27, 92, 197 },
+#endif
+  { 252, 71, 126, 184, 178, 218, 251, 49, 133, 221, 27, 92, 197 },
+  { 252, 71, 126, 184, 178, 218, 251, 49, 133, 221, 27, 92, 197 },
+  { 252, 71, 126, 184, 178, 218, 251, 49, 133, 221, 27, 92, 197 },
+  { 252, 71, 126, 184, 178, 218, 251, 49, 133, 221, 27, 92, 197 },
+};
+const aom_prob default_dc_sign[PLANE_TYPES][DC_SIGN_CONTEXTS] = {
+  { 125, 102, 147 }, { 119, 101, 135 },
+};
+
+const aom_prob default_coeff_base
+    [TX_SIZES][PLANE_TYPES][NUM_BASE_LEVELS][COEFF_BASE_CONTEXTS] = {
+#if CONFIG_CB4X4
+      { // TX_2X2
+        {
+            { 73,  128, 131, 204, 165, 226, 169, 236, 18,  128, 51,
+              153, 97,  179, 123, 201, 145, 226, 20,  128, 59,  153,
+              107, 181, 129, 201, 142, 226, 3,   128, 19,  99,  46,
+              135, 92,  166, 129, 190, 157, 217, 128, 128 },
+
+            { 128, 128, 178, 218, 192, 236, 186, 243, 55,  128, 110,
+              183, 151, 205, 168, 221, 180, 238, 65,  128, 116, 178,
+              157, 206, 172, 222, 183, 238, 24,  128, 65,  127, 104,
+              164, 154, 195, 187, 216, 205, 230, 128, 128 },
+        },
+        {
+            { 73,  128, 131, 204, 165, 226, 169, 236, 18,  128, 51,
+              153, 97,  179, 123, 201, 145, 226, 20,  128, 59,  153,
+              107, 181, 129, 201, 142, 226, 3,   128, 19,  99,  46,
+              135, 92,  166, 129, 190, 157, 217, 128, 128 },
+
+            { 128, 128, 178, 218, 192, 236, 186, 243, 55,  128, 110,
+              183, 151, 205, 168, 221, 180, 238, 65,  128, 116, 178,
+              157, 206, 172, 222, 183, 238, 24,  128, 65,  127, 104,
+              164, 154, 195, 187, 216, 205, 230, 128, 128 },
+        } },
+#endif
+      { // TX_4X4
+        {
+            // PLANE_Y
+            { 73,  128, 131, 204, 165, 226, 169, 236, 18,  128, 51,
+              153, 97,  179, 123, 201, 145, 226, 20,  128, 59,  153,
+              107, 181, 129, 201, 142, 226, 3,   128, 19,  99,  46,
+              135, 92,  166, 129, 190, 157, 217, 128, 128 },
+
+            { 128, 128, 178, 218, 192, 236, 186, 243, 55,  128, 110,
+              183, 151, 205, 168, 221, 180, 238, 65,  128, 116, 178,
+              157, 206, 172, 222, 183, 238, 24,  128, 65,  127, 104,
+              164, 154, 195, 187, 216, 205, 230, 128, 128 },
+        },
+        {
+            // PLANE_UV
+            { 47,  128, 100, 176, 140, 207, 150, 223, 11,  128, 35,
+              133, 79,  165, 115, 186, 129, 210, 8,   128, 30,  114,
+              80,  159, 116, 187, 146, 214, 2,   128, 9,   59,  28,
+              86,  71,  131, 117, 165, 149, 188, 128, 128 },
+
+            { 83,  128, 152, 205, 168, 227, 192, 238, 42,  128, 92,
+              169, 138, 193, 165, 209, 128, 206, 36,  128, 86,  159,
+              141, 198, 181, 213, 102, 223, 18,  128, 50,  132, 90,
+              144, 141, 169, 180, 191, 128, 217, 128, 128 },
+        } },
+      {
+          // TX_8X8
+          {
+              // PLANE_Y
+              { 82,  128, 143, 203, 177, 225, 186, 237, 7,   128, 37,
+                109, 78,  151, 110, 182, 139, 213, 25,  128, 51,  115,
+                86,  146, 111, 175, 125, 205, 3,   128, 12,  55,  32,
+                78,  63,  111, 96,  148, 123, 185, 146, 206 },
+
+              { 136, 128, 182, 220, 201, 236, 205, 243, 46,  128, 101,
+                164, 147, 194, 170, 218, 177, 234, 62,  128, 104, 146,
+                143, 183, 165, 207, 183, 228, 30,  128, 60,  95,  95,
+                128, 135, 163, 166, 196, 175, 219, 192, 231 },
+          },
+          {
+              // PLANE_UV
+              { 47,  128, 112, 189, 164, 202, 163, 218, 8,   128, 32,
+                110, 68,  151, 102, 179, 134, 195, 5,   128, 22,  76,
+                54,  103, 80,  146, 101, 182, 1,   128, 5,   39,  17,
+                53,  46,  93,  79,  127, 112, 161, 64,  195 },
+
+              { 90,  128, 156, 210, 183, 225, 128, 236, 39,  128, 98,
+                164, 146, 201, 209, 219, 171, 208, 32,  128, 68,  123,
+                119, 169, 154, 184, 128, 213, 15,  128, 38,  111, 83,
+                112, 120, 163, 180, 170, 154, 213, 128, 205 },
+          },
+      },
+
+      {
+          // TX_16X16
+          {
+              // PLANE_Y
+              { 96,  128, 169, 218, 208, 233, 187, 244, 10,  128, 34,
+                101, 82,  153, 113, 184, 137, 212, 6,   128, 34,  104,
+                81,  145, 109, 176, 147, 202, 1,   128, 3,   43,  15,
+                53,  43,  89,  79,  129, 108, 168, 110, 194 },
+
+              { 156, 128, 206, 232, 218, 240, 128, 251, 39,  128, 108,
+                161, 156, 202, 187, 216, 179, 234, 40,  128, 103, 152,
+                144, 185, 159, 208, 205, 227, 14,  128, 39,  84,  76,
+                110, 121, 151, 157, 187, 201, 206, 64,  216 },
+          },
+          {
+              // PLANE_UV
+              { 42, 128, 139, 211, 180, 230, 199, 238, 3,   128, 32,
+                96, 69,  145, 102, 186, 117, 212, 4,   128, 25,  72,
+                55, 111, 81,  159, 116, 198, 1,   128, 4,   22,  16,
+                34, 35,  68,  63,  116, 89,  165, 102, 199 },
+
+              { 135, 128, 193, 227, 182, 239, 128, 246, 42,  128, 115,
+                156, 146, 203, 188, 216, 128, 229, 32,  128, 82,  127,
+                120, 178, 165, 203, 213, 229, 11,  128, 32,  73,  79,
+                111, 129, 158, 162, 187, 156, 209, 85,  222 },
+          },
+      },
+
+      {
+          // TX_32X32
+          {
+              // PLANE_Y
+              { 97,  128, 163, 232, 191, 246, 219, 252, 3,   128, 41,
+                108, 91,  147, 104, 183, 118, 225, 6,   128, 45,  91,
+                83,  125, 92,  160, 99,  215, 1,   128, 11,  36,  28,
+                46,  43,  59,  57,  86,  73,  145, 91,  210 },
+
+              { 127, 128, 201, 239, 247, 248, 128, 254, 40,  128, 103,
+                152, 158, 199, 186, 225, 181, 242, 38,  128, 92,  112,
+                146, 189, 162, 217, 112, 239, 17,  128, 30,  47,  63,
+                89,  113, 146, 147, 187, 168, 217, 150, 233 },
+          },
+          {
+              // PLANE_UV
+              { 65,  128, 155, 223, 166, 235, 154, 244, 15,  128, 57,
+                154, 110, 199, 159, 224, 149, 239, 9,   128, 57,  140,
+                97,  185, 148, 218, 176, 236, 1,   128, 3,   43,  19,
+                42,  64,  98,  117, 167, 154, 199, 128, 158 },
+
+              { 130, 128, 189, 231, 171, 247, 128, 246, 63,  128, 132,
+                222, 186, 224, 199, 244, 128, 247, 55,  128, 113, 211,
+                164, 230, 225, 243, 128, 239, 7,   128, 31,  102, 106,
+                138, 147, 183, 171, 223, 171, 224, 128, 128 },
+          },
+      },
+    };
+
+const aom_prob default_nz_map[TX_SIZES][PLANE_TYPES][SIG_COEF_CONTEXTS] = {
+#if CONFIG_CB4X4
+  {
+      { 34, 103, 61, 106, 62,  160, 112, 54, 173, 121,
+        75, 157, 92, 75,  157, 129, 94,  65, 52,  37 },
+      { 52,  124, 84,  136, 107, 197, 161, 82, 183, 151,
+        109, 153, 140, 103, 152, 134, 109, 81, 69,  50 },
+  },
+#endif
+  {
+      { 34, 103, 61, 106, 62,  160, 112, 54, 173, 121,
+        75, 157, 92, 75,  157, 129, 94,  65, 52,  37 },
+      { 52,  124, 84,  136, 107, 197, 161, 82, 183, 151,
+        109, 153, 140, 103, 152, 134, 109, 81, 69,  50 },
+  },
+  {
+      { 34, 127, 74,  124, 74,  204, 153, 76,  226, 162,
+        92, 207, 126, 91,  227, 192, 149, 108, 85,  55 },
+      { 43,  136, 115, 158, 130, 212, 187, 112, 231, 180,
+        130, 202, 164, 130, 236, 204, 168, 139, 112, 114 },
+  },
+  {
+      { 25,  117, 70,  120, 77,  215, 171, 102, 234, 156,
+        105, 235, 155, 109, 247, 220, 176, 127, 92,  72 },
+      { 24,  88,  49,  100, 62,  202, 148, 62,  237, 178,
+        102, 233, 168, 105, 244, 198, 162, 127, 103, 71 },
+  },
+  {
+      { 11, 54,  17,  69, 26,  128, 125, 56,  232, 130,
+        60, 237, 121, 66, 250, 168, 134, 114, 93,  53 },
+      { 21, 52,  32,  95,  64,  171, 152, 70,  247, 159,
+        81, 252, 177, 100, 252, 221, 192, 143, 195, 146 },
+  },
+};
+
+const aom_prob default_eob_flag[TX_SIZES][PLANE_TYPES][EOB_COEF_CONTEXTS] = {
+#if CONFIG_CB4X4
+  {
+      { 229, 236, 231, 222, 239, 236, 214, 201, 236, 226, 195, 134, 228,
+        210, 150, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 182, 186, 172, 176, 207, 213, 152, 122, 187, 171, 131, 65, 170,
+        134, 101, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+  },
+#endif
+  {
+      { 229, 236, 231, 222, 239, 236, 214, 201, 236, 226, 195, 134, 228,
+        210, 150, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+      { 182, 186, 172, 176, 207, 213, 152, 122, 187, 171, 131, 65, 170,
+        134, 101, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 },
+  },
+  {
+      { 225, 234, 244, 236, 205, 242, 246, 247, 246, 234, 191, 242, 237,
+        215, 142, 224, 206, 142, 73,  128, 128, 128, 128, 128, 128 },
+      { 154, 171, 187, 175, 62,  199, 202, 206, 215, 200, 111, 197, 199,
+        174, 100, 135, 105, 104, 45,  128, 128, 128, 128, 128, 128 },
+  },
+  {
+      { 180, 213, 216, 229, 233, 232, 240, 235, 220, 178, 239, 238, 225,
+        187, 229, 214, 226, 200, 183, 141, 158, 179, 128, 128, 128 },
+      { 190, 225, 234, 248, 249, 248, 253, 251, 232, 110, 254, 252, 236,
+        57,  253, 248, 232, 85,  244, 189, 112, 64,  128, 128, 128 },
+  },
+  {
+      { 248, 224, 246, 244, 239, 245, 251, 246, 251, 255, 255, 255, 249,
+        255, 255, 255, 229, 255, 255, 255, 228, 255, 255, 247, 137 },
+      { 204, 207, 233, 215, 193, 228, 239, 221, 227, 250, 236, 207, 135,
+        236, 186, 182, 57,  209, 140, 128, 85,  184, 110, 128, 128 },
+  },
+};
+
+const aom_prob default_coeff_lps[TX_SIZES][PLANE_TYPES][LEVEL_CONTEXTS] = {
+#if CONFIG_CB4X4
+  {
+      { 164, 128, 134, 165, 128, 137, 168, 128, 97,  136, 167, 128,
+        182, 205, 143, 172, 200, 145, 173, 193, 103, 137, 170, 191,
+        198, 214, 162, 187, 209, 162, 187, 207, 128, 156, 183, 201,
+        219, 230, 204, 210, 225, 201, 209, 225, 187, 190, 203, 214 },
+      { 106, 128, 98,  126, 128, 87,  122, 128, 54,  89,  131, 128,
+        142, 180, 123, 154, 189, 115, 149, 175, 79,  115, 157, 182,
+        175, 197, 147, 174, 199, 145, 174, 201, 89,  135, 173, 194,
+        212, 222, 206, 203, 223, 188, 201, 220, 128, 144, 202, 206 },
+  },
+#endif
+  {
+      { 164, 128, 134, 165, 128, 137, 168, 128, 97,  136, 167, 128,
+        182, 205, 143, 172, 200, 145, 173, 193, 103, 137, 170, 191,
+        198, 214, 162, 187, 209, 162, 187, 207, 128, 156, 183, 201,
+        219, 230, 204, 210, 225, 201, 209, 225, 187, 190, 203, 214 },
+      { 106, 128, 98,  126, 128, 87,  122, 128, 54,  89,  131, 128,
+        142, 180, 123, 154, 189, 115, 149, 175, 79,  115, 157, 182,
+        175, 197, 147, 174, 199, 145, 174, 201, 89,  135, 173, 194,
+        212, 222, 206, 203, 223, 188, 201, 220, 128, 144, 202, 206 },
+  },
+  {
+      { 171, 128, 123, 169, 128, 121, 165, 128, 82,  125, 168, 128,
+        191, 213, 143, 177, 199, 136, 170, 194, 95,  135, 171, 195,
+        206, 222, 166, 191, 212, 154, 184, 207, 115, 149, 180, 204,
+        223, 237, 196, 215, 231, 186, 209, 228, 158, 178, 201, 222 },
+      { 115, 128, 115, 146, 128, 91,  147, 128, 55,  93,  139, 128,
+        147, 190, 141, 176, 201, 123, 156, 173, 68,  114, 156, 195,
+        186, 205, 153, 191, 214, 141, 179, 205, 107, 132, 166, 184,
+        215, 225, 200, 212, 230, 102, 207, 222, 128, 119, 200, 212 },
+  },
+  {
+      { 185, 128, 134, 198, 128, 128, 195, 128, 58,  110, 162, 128,
+        208, 227, 154, 196, 206, 144, 188, 209, 83,  130, 168, 198,
+        219, 232, 167, 205, 222, 158, 196, 216, 107, 143, 178, 204,
+        233, 244, 202, 226, 238, 191, 217, 234, 153, 178, 200, 223 },
+      { 160, 128, 154, 197, 128, 129, 178, 128, 53,  112, 157, 128,
+        185, 214, 169, 196, 221, 134, 179, 186, 82,  131, 168, 194,
+        204, 220, 176, 209, 221, 173, 194, 209, 107, 154, 181, 203,
+        230, 241, 202, 226, 237, 185, 223, 234, 162, 187, 203, 222 },
+  },
+  {
+      { 177, 128, 165, 226, 128, 152, 219, 128, 45,  129, 188, 128,
+        198, 218, 179, 220, 228, 163, 214, 220, 72,  134, 181, 206,
+        216, 225, 177, 218, 231, 158, 213, 223, 112, 150, 185, 210,
+        245, 251, 204, 234, 247, 195, 231, 243, 163, 186, 213, 235 },
+      { 161, 128, 174, 205, 128, 146, 182, 128, 59,  125, 179, 128,
+        183, 208, 199, 220, 239, 184, 213, 217, 71,  141, 196, 217,
+        213, 219, 215, 230, 237, 171, 224, 238, 112, 173, 193, 221,
+        239, 246, 168, 243, 249, 93,  241, 247, 128, 195, 216, 233 },
+  },
+};
+#endif  // CONFIG_LV_MAP
+
+#if CONFIG_ALT_INTRA
+
+const aom_prob av1_kf_y_mode_prob[INTRA_MODES][INTRA_MODES][INTRA_MODES - 1] = {
+  {
+      // above = dc
+      { 121, 30, 54, 128, 164, 158, 45, 41, 57, 91 },   // left = dc
+      { 91, 38, 101, 102, 124, 141, 49, 48, 45, 73 },   // left = v
+      { 66, 28, 27, 177, 225, 178, 32, 27, 52, 114 },   // left = h
+      { 106, 23, 50, 101, 134, 148, 64, 50, 49, 107 },  // left = d45
+      { 75, 24, 32, 118, 66, 143, 42, 28, 57, 74 },     // left = d135
+      { 95, 24, 40, 142, 56, 141, 72, 121, 129, 255 },  // left = d117
+      { 71, 14, 25, 126, 117, 201, 28, 21, 117, 89 },   // left = d153
+      { 85, 16, 37, 110, 163, 178, 41, 28, 48, 134 },   // left = d207
+      { 86, 25, 32, 83, 105, 133, 58, 81, 46, 95 },     // left = d63
+      { 79, 25, 38, 75, 150, 255, 30, 49, 34, 51 },     // left = smooth
+      { 68, 59, 48, 122, 193, 158, 43, 46, 46, 112 },   // left = paeth
+  },
+  {
+      // above = v
+      { 66, 21, 118, 111, 145, 107, 27, 50, 27, 54 },    // left = dc
+      { 52, 25, 167, 81, 120, 101, 34, 55, 19, 32 },     // left = v
+      { 56, 18, 72, 134, 208, 139, 31, 34, 27, 89 },     // left = h
+      { 75, 21, 94, 88, 134, 123, 49, 57, 30, 68 },      // left = d45
+      { 54, 18, 95, 96, 78, 107, 33, 49, 28, 65 },       // left = d135
+      { 61, 19, 121, 131, 58, 101, 56, 143, 120, 255 },  // left = d117
+      { 53, 13, 78, 103, 110, 147, 31, 41, 64, 77 },     // left = d153
+      { 69, 14, 78, 93, 167, 121, 31, 39, 25, 113 },     // left = d207
+      { 64, 18, 103, 79, 90, 108, 34, 73, 27, 69 },      // left = d63
+      { 52, 20, 103, 61, 161, 255, 22, 42, 16, 35 },     // left = smooth
+      { 50, 31, 124, 92, 161, 120, 50, 53, 23, 60 },     // left = paeth
+  },
+  {
+      // above = h
+      { 94, 29, 31, 158, 214, 178, 35, 31, 72, 111 },   // left = dc
+      { 72, 37, 72, 149, 184, 177, 43, 40, 53, 105 },   // left = v
+      { 53, 21, 14, 196, 242, 209, 29, 19, 55, 145 },   // left = h
+      { 93, 36, 36, 104, 176, 166, 56, 37, 49, 141 },   // left = d45
+      { 84, 32, 27, 124, 108, 143, 38, 36, 76, 134 },   // left = d135
+      { 82, 31, 47, 142, 122, 161, 83, 73, 126, 255 },  // left = d117
+      { 66, 16, 20, 133, 148, 210, 30, 17, 113, 104 },  // left = d153
+      { 76, 16, 17, 129, 207, 181, 41, 20, 46, 163 },   // left = d207
+      { 72, 38, 21, 100, 142, 171, 37, 70, 49, 111 },   // left = d63
+      { 61, 30, 27, 115, 208, 255, 27, 31, 44, 63 },    // left = smooth
+      { 53, 45, 29, 157, 222, 185, 49, 37, 55, 102 },   // left = paeth
+  },
+  {
+      // above = d45
+      { 96, 18, 37, 98, 138, 154, 68, 56, 59, 96 },    // left = dc
+      { 73, 18, 92, 81, 125, 132, 75, 64, 27, 67 },    // left = v
+      { 73, 17, 27, 128, 213, 154, 56, 44, 32, 105 },  // left = h
+      { 101, 20, 21, 75, 138, 138, 82, 56, 23, 154 },  // left = d45
+      { 71, 15, 33, 91, 70, 150, 62, 55, 38, 118 },    // left = d135
+      { 80, 19, 38, 116, 69, 122, 88, 132, 92, 255 },  // left = d117
+      { 68, 11, 22, 101, 116, 179, 52, 44, 85, 96 },   // left = d153
+      { 101, 8, 59, 77, 151, 170, 53, 41, 35, 172 },   // left = d207
+      { 82, 19, 24, 81, 172, 129, 82, 128, 43, 108 },  // left = d63
+      { 66, 18, 42, 64, 143, 255, 52, 52, 25, 83 },    // left = smooth
+      { 57, 24, 42, 85, 169, 145, 104, 71, 34, 86 },   // left = paeth
+  },
+  {
+      // above = d135
+      { 85, 15, 29, 113, 83, 176, 26, 29, 70, 110 },    // left = dc
+      { 78, 28, 49, 111, 91, 141, 30, 42, 48, 75 },     // left = v
+      { 56, 21, 16, 146, 190, 178, 23, 31, 49, 92 },    // left = h
+      { 70, 19, 20, 65, 90, 173, 97, 36, 57, 98 },      // left = d45
+      { 77, 14, 26, 110, 51, 156, 34, 35, 54, 74 },     // left = d135
+      { 78, 18, 36, 153, 47, 131, 62, 102, 155, 255 },  // left = d117
+      { 56, 11, 15, 115, 85, 196, 32, 45, 81, 96 },     // left = d153
+      { 90, 18, 24, 95, 126, 159, 34, 31, 46, 136 },    // left = d207
+      { 80, 23, 28, 90, 75, 141, 39, 50, 46, 87 },      // left = d63
+      { 63, 22, 31, 91, 110, 255, 26, 43, 51, 51 },     // left = smooth
+      { 66, 32, 31, 122, 145, 165, 40, 43, 56, 79 },    // left = paeth
+  },
+  {
+      // above = d117
+      { 81, 16, 61, 170, 74, 105, 54, 105, 113, 255 },  // left = dc
+      { 74, 20, 86, 163, 64, 97, 65, 129, 101, 255 },   // left = v
+      { 63, 15, 47, 168, 141, 176, 69, 77, 77, 255 },   // left = h
+      { 70, 17, 59, 97, 78, 114, 74, 122, 80, 255 },    // left = d45
+      { 78, 13, 50, 153, 34, 126, 75, 114, 120, 255 },  // left = d135
+      { 72, 16, 69, 159, 28, 108, 63, 134, 107, 255 },  // left = d117
+      { 66, 9, 47, 131, 79, 148, 41, 88, 105, 255 },    // left = d153
+      { 78, 12, 60, 119, 105, 133, 47, 95, 63, 255 },   // left = d207
+      { 82, 21, 58, 128, 61, 98, 64, 136, 91, 255 },    // left = d63
+      { 23, 26, 28, 96, 85, 128, 51, 64, 85, 128 },     // left = smooth
+      { 58, 27, 62, 162, 109, 151, 75, 106, 78, 255 },  // left = paeth
+  },
+  {
+      // above = d153
+      { 91, 18, 25, 121, 166, 173, 25, 25, 128, 102 },  // left = dc
+      { 80, 27, 51, 111, 141, 147, 45, 38, 70, 85 },    // left = v
+      { 53, 12, 11, 154, 197, 225, 17, 17, 74, 145 },   // left = h
+      { 93, 27, 23, 111, 143, 188, 43, 39, 69, 112 },   // left = d45
+      { 83, 15, 21, 118, 67, 178, 40, 33, 73, 92 },     // left = d135
+      { 94, 13, 31, 132, 66, 110, 61, 82, 148, 255 },   // left = d117
+      { 76, 9, 11, 96, 105, 201, 16, 13, 157, 97 },     // left = d153
+      { 70, 10, 12, 100, 172, 201, 23, 17, 53, 158 },   // left = d207
+      { 114, 25, 21, 104, 108, 163, 30, 47, 53, 111 },  // left = d63
+      { 70, 16, 21, 80, 157, 255, 25, 30, 81, 69 },     // left = smooth
+      { 87, 32, 26, 120, 191, 168, 32, 33, 70, 118 },   // left = paeth
+  },
+  {
+      // above = d207
+      { 98, 20, 39, 122, 168, 188, 38, 36, 54, 132 },   // left = dc
+      { 81, 37, 62, 97, 122, 153, 38, 43, 36, 118 },    // left = v
+      { 71, 21, 22, 154, 227, 183, 37, 31, 46, 140 },   // left = h
+      { 90, 34, 19, 93, 144, 194, 65, 47, 41, 163 },    // left = d45
+      { 78, 20, 27, 91, 93, 173, 57, 52, 49, 113 },     // left = d135
+      { 79, 25, 45, 121, 101, 147, 69, 56, 122, 255 },  // left = d117
+      { 73, 13, 19, 105, 122, 206, 40, 28, 91, 126 },   // left = d153
+      { 101, 14, 22, 87, 153, 169, 33, 25, 26, 175 },   // left = d207
+      { 81, 28, 23, 86, 115, 169, 48, 56, 41, 111 },    // left = d63
+      { 70, 24, 30, 90, 180, 255, 38, 26, 36, 82 },     // left = smooth
+      { 61, 37, 30, 94, 189, 163, 76, 50, 36, 127 },    // left = paeth
+  },
+  {
+      // above = d63
+      { 77, 13, 46, 86, 138, 117, 55, 88, 34, 68 },     // left = dc
+      { 68, 17, 80, 64, 105, 108, 66, 115, 32, 45 },    // left = v
+      { 62, 13, 37, 124, 210, 131, 46, 57, 28, 103 },   // left = h
+      { 88, 15, 45, 73, 134, 145, 73, 101, 37, 87 },    // left = d45
+      { 68, 16, 35, 78, 81, 133, 54, 71, 33, 67 },      // left = d135
+      { 71, 16, 57, 108, 61, 135, 71, 184, 113, 255 },  // left = d117
+      { 55, 10, 27, 69, 107, 158, 39, 76, 82, 95 },     // left = d153
+      { 80, 9, 38, 78, 153, 145, 50, 63, 28, 123 },     // left = d207
+      { 86, 12, 33, 49, 107, 135, 64, 134, 57, 89 },    // left = d63
+      { 56, 19, 55, 60, 163, 255, 38, 84, 22, 36 },     // left = smooth
+      { 53, 17, 60, 69, 151, 126, 73, 113, 26, 80 },    // left = paeth
+  },
+  {
+      // above = smooth
+      { 79, 16, 46, 89, 167, 255, 22, 36, 29, 42 },   // left = dc
+      { 63, 22, 88, 71, 131, 255, 26, 41, 21, 35 },   // left = v
+      { 51, 18, 28, 142, 232, 255, 26, 25, 25, 75 },  // left = h
+      { 75, 18, 43, 70, 140, 255, 37, 49, 34, 89 },   // left = d45
+      { 70, 14, 35, 87, 83, 255, 30, 36, 34, 50 },    // left = d135
+      { 23, 26, 28, 96, 85, 128, 51, 64, 85, 128 },   // left = d117
+      { 74, 12, 33, 83, 128, 255, 27, 33, 58, 68 },   // left = d153
+      { 66, 11, 30, 77, 179, 255, 21, 27, 23, 113 },  // left = d207
+      { 68, 22, 40, 65, 118, 255, 28, 61, 30, 50 },   // left = d63
+      { 60, 18, 44, 69, 141, 255, 18, 32, 22, 40 },   // left = smooth
+      { 52, 32, 54, 96, 194, 255, 33, 37, 25, 53 },   // left = paeth
+  },
+  {
+      // above = paeth
+      { 76, 47, 67, 123, 182, 150, 41, 52, 55, 97 },    // left = dc
+      { 69, 40, 125, 102, 138, 138, 42, 55, 32, 70 },   // left = v
+      { 46, 28, 27, 160, 232, 169, 34, 21, 32, 122 },   // left = h
+      { 78, 35, 41, 99, 128, 124, 49, 43, 35, 111 },    // left = d45
+      { 66, 28, 47, 100, 113, 145, 37, 40, 72, 93 },    // left = d135
+      { 77, 37, 76, 134, 124, 124, 65, 122, 88, 255 },  // left = d117
+      { 53, 23, 38, 108, 128, 204, 26, 32, 115, 114 },  // left = d153
+      { 65, 20, 29, 101, 202, 186, 29, 24, 29, 188 },   // left = d207
+      { 71, 24, 49, 81, 126, 151, 36, 65, 28, 93 },     // left = d63
+      { 54, 36, 53, 94, 193, 255, 25, 38, 20, 64 },     // left = smooth
+      { 52, 54, 60, 108, 176, 168, 47, 44, 50, 105 },   // left = paeth
+  },
+};
+
+static const aom_prob default_if_y_probs[BLOCK_SIZE_GROUPS][INTRA_MODES - 1] = {
+  { 88, 16, 47, 133, 143, 150, 70, 48, 84, 122 },  // block_size < 8x8
+  { 75, 26, 51, 120, 158, 157, 44, 45, 56, 102 },  // block_size < 16x16
+  { 73, 24, 60, 115, 184, 164, 26, 36, 32, 63 },   // block_size < 32x32
+  { 96, 27, 50, 107, 221, 148, 16, 22, 14, 39 },   // block_size >= 32x32
+};
+
+static const aom_prob default_uv_probs[INTRA_MODES][INTRA_MODES - 1] = {
+  { 199, 3, 79, 179, 220, 109, 38, 50, 68, 138 },   // y = dc
+  { 17, 2, 219, 136, 131, 58, 21, 106, 23, 41 },    // y = v
+  { 26, 1, 5, 244, 253, 138, 16, 21, 68, 205 },     // y = h
+  { 183, 3, 66, 94, 195, 97, 101, 104, 41, 178 },   // y = d45
+  { 178, 2, 36, 158, 99, 175, 21, 29, 105, 77 },    // y = d135
+  { 154, 3, 65, 219, 40, 48, 45, 95, 146, 255 },    // y = d117
+  { 167, 1, 16, 160, 214, 187, 10, 10, 200, 155 },  // y = d153
+  { 154, 2, 18, 178, 238, 132, 25, 21, 34, 221 },   // y = d207
+  { 153, 4, 76, 85, 157, 90, 38, 165, 46, 104 },    // y = d63
+  { 163, 3, 68, 87, 190, 255, 19, 27, 25, 46 },     // y = smooth
+  { 185, 7, 113, 171, 203, 57, 18, 69, 49, 104 },   // y = paeth
+};
+
+#else
+
+const aom_prob av1_kf_y_mode_prob[INTRA_MODES][INTRA_MODES][INTRA_MODES - 1] = {
+  {
+      // above = dc
+      { 137, 30, 42, 148, 151, 207, 70, 52, 91 },   // left = dc
+      { 92, 45, 102, 136, 116, 180, 74, 90, 100 },  // left = v
+      { 73, 32, 19, 187, 222, 215, 46, 34, 100 },   // left = h
+      { 91, 30, 32, 116, 121, 186, 93, 86, 94 },    // left = d45
+      { 72, 35, 36, 149, 68, 206, 68, 63, 105 },    // left = d135
+      { 73, 31, 28, 138, 57, 124, 55, 122, 151 },   // left = d117
+      { 67, 23, 21, 140, 126, 197, 40, 37, 171 },   // left = d153
+      { 86, 27, 28, 128, 154, 212, 45, 43, 53 },    // left = d207
+      { 74, 32, 27, 107, 86, 160, 63, 134, 102 },   // left = d63
+      { 59, 67, 44, 140, 161, 202, 78, 67, 119 }    // left = tm
+  },
+  {
+      // above = v
+      { 63, 36, 126, 146, 123, 158, 60, 90, 96 },   // left = dc
+      { 43, 46, 168, 134, 107, 128, 69, 142, 92 },  // left = v
+      { 44, 29, 68, 159, 201, 177, 50, 57, 77 },    // left = h
+      { 58, 38, 76, 114, 97, 172, 78, 133, 92 },    // left = d45
+      { 46, 41, 76, 140, 63, 184, 69, 112, 57 },    // left = d135
+      { 38, 32, 85, 140, 46, 112, 54, 151, 133 },   // left = d117
+      { 39, 27, 61, 131, 110, 175, 44, 75, 136 },   // left = d153
+      { 52, 30, 74, 113, 130, 175, 51, 64, 58 },    // left = d207
+      { 47, 35, 80, 100, 74, 143, 64, 163, 74 },    // left = d63
+      { 36, 61, 116, 114, 128, 162, 80, 125, 82 }   // left = tm
+  },
+  {
+      // above = h
+      { 82, 26, 26, 171, 208, 204, 44, 32, 105 },  // left = dc
+      { 55, 44, 68, 166, 179, 192, 57, 57, 108 },  // left = v
+      { 42, 26, 11, 199, 241, 228, 23, 15, 85 },   // left = h
+      { 68, 42, 19, 131, 160, 199, 55, 52, 83 },   // left = d45
+      { 58, 50, 25, 139, 115, 232, 39, 52, 118 },  // left = d135
+      { 50, 35, 33, 153, 104, 162, 64, 59, 131 },  // left = d117
+      { 44, 24, 16, 150, 177, 202, 33, 19, 156 },  // left = d153
+      { 55, 27, 12, 153, 203, 218, 26, 27, 49 },   // left = d207
+      { 53, 49, 21, 110, 116, 168, 59, 80, 76 },   // left = d63
+      { 38, 72, 19, 168, 203, 212, 50, 50, 107 }   // left = tm
+  },
+  {
+      // above = d45
+      { 103, 26, 36, 129, 132, 201, 83, 80, 93 },  // left = dc
+      { 59, 38, 83, 112, 103, 162, 98, 136, 90 },  // left = v
+      { 62, 30, 23, 158, 200, 207, 59, 57, 50 },   // left = h
+      { 67, 30, 29, 84, 86, 191, 102, 91, 59 },    // left = d45
+      { 60, 32, 33, 112, 71, 220, 64, 89, 104 },   // left = d135
+      { 53, 26, 34, 130, 56, 149, 84, 120, 103 },  // left = d117
+      { 53, 21, 23, 133, 109, 210, 56, 77, 172 },  // left = d153
+      { 77, 19, 29, 112, 142, 228, 55, 66, 36 },   // left = d207
+      { 61, 29, 29, 93, 97, 165, 83, 175, 162 },   // left = d63
+      { 47, 47, 43, 114, 137, 181, 100, 99, 95 }   // left = tm
+  },
+  {
+      // above = d135
+      { 69, 23, 29, 128, 83, 199, 46, 44, 101 },   // left = dc
+      { 53, 40, 55, 139, 69, 183, 61, 80, 110 },   // left = v
+      { 40, 29, 19, 161, 180, 207, 43, 24, 91 },   // left = h
+      { 60, 34, 19, 105, 61, 198, 53, 64, 89 },    // left = d45
+      { 52, 31, 22, 158, 40, 209, 58, 62, 89 },    // left = d135
+      { 44, 31, 29, 147, 46, 158, 56, 102, 198 },  // left = d117
+      { 35, 19, 12, 135, 87, 209, 41, 45, 167 },   // left = d153
+      { 55, 25, 21, 118, 95, 215, 38, 39, 66 },    // left = d207
+      { 51, 38, 25, 113, 58, 164, 70, 93, 97 },    // left = d63
+      { 47, 54, 34, 146, 108, 203, 72, 103, 151 }  // left = tm
+  },
+  {
+      // above = d117
+      { 64, 19, 37, 156, 66, 138, 49, 95, 133 },   // left = dc
+      { 46, 27, 80, 150, 55, 124, 55, 121, 135 },  // left = v
+      { 36, 23, 27, 165, 149, 166, 54, 64, 118 },  // left = h
+      { 53, 21, 36, 131, 63, 163, 60, 109, 81 },   // left = d45
+      { 40, 26, 35, 154, 40, 185, 51, 97, 123 },   // left = d135
+      { 35, 19, 34, 179, 19, 97, 48, 129, 124 },   // left = d117
+      { 36, 20, 26, 136, 62, 164, 33, 77, 154 },   // left = d153
+      { 45, 18, 32, 130, 90, 157, 40, 79, 91 },    // left = d207
+      { 45, 26, 28, 129, 45, 129, 49, 147, 123 },  // left = d63
+      { 38, 44, 51, 136, 74, 162, 57, 97, 121 }    // left = tm
+  },
+  {
+      // above = d153
+      { 75, 17, 22, 136, 138, 185, 32, 34, 166 },  // left = dc
+      { 56, 39, 58, 133, 117, 173, 48, 53, 187 },  // left = v
+      { 35, 21, 12, 161, 212, 207, 20, 23, 145 },  // left = h
+      { 56, 29, 19, 117, 109, 181, 55, 68, 112 },  // left = d45
+      { 47, 29, 17, 153, 64, 220, 59, 51, 114 },   // left = d135
+      { 46, 16, 24, 136, 76, 147, 41, 64, 172 },   // left = d117
+      { 34, 17, 11, 108, 152, 187, 13, 15, 209 },  // left = d153
+      { 51, 24, 14, 115, 133, 209, 32, 26, 104 },  // left = d207
+      { 55, 30, 18, 122, 79, 179, 44, 88, 116 },   // left = d63
+      { 37, 49, 25, 129, 168, 164, 41, 54, 148 }   // left = tm
+  },
+  {
+      // above = d207
+      { 82, 22, 32, 127, 143, 213, 39, 41, 70 },   // left = dc
+      { 62, 44, 61, 123, 105, 189, 48, 57, 64 },   // left = v
+      { 47, 25, 17, 175, 222, 220, 24, 30, 86 },   // left = h
+      { 68, 36, 17, 106, 102, 206, 59, 74, 74 },   // left = d45
+      { 57, 39, 23, 151, 68, 216, 55, 63, 58 },    // left = d135
+      { 49, 30, 35, 141, 70, 168, 82, 40, 115 },   // left = d117
+      { 51, 25, 15, 136, 129, 202, 38, 35, 139 },  // left = d153
+      { 68, 26, 16, 111, 141, 215, 29, 28, 28 },   // left = d207
+      { 59, 39, 19, 114, 75, 180, 77, 104, 42 },   // left = d63
+      { 40, 61, 26, 126, 152, 206, 61, 59, 93 }    // left = tm
+  },
+  {
+      // above = d63
+      { 78, 23, 39, 111, 117, 170, 74, 124, 94 },   // left = dc
+      { 48, 34, 86, 101, 92, 146, 78, 179, 134 },   // left = v
+      { 47, 22, 24, 138, 187, 178, 68, 69, 59 },    // left = h
+      { 56, 25, 33, 105, 112, 187, 95, 177, 129 },  // left = d45
+      { 48, 31, 27, 114, 63, 183, 82, 116, 56 },    // left = d135
+      { 43, 28, 37, 121, 63, 123, 61, 192, 169 },   // left = d117
+      { 42, 17, 24, 109, 97, 177, 56, 76, 122 },    // left = d153
+      { 58, 18, 28, 105, 139, 182, 70, 92, 63 },    // left = d207
+      { 46, 23, 32, 74, 86, 150, 67, 183, 88 },     // left = d63
+      { 36, 38, 48, 92, 122, 165, 88, 137, 91 }     // left = tm
+  },
+  {
+      // above = tm
+      { 65, 70, 60, 155, 159, 199, 61, 60, 81 },    // left = dc
+      { 44, 78, 115, 132, 119, 173, 71, 112, 93 },  // left = v
+      { 39, 38, 21, 184, 227, 206, 42, 32, 64 },    // left = h
+      { 58, 47, 36, 124, 137, 193, 80, 82, 78 },    // left = d45
+      { 49, 50, 35, 144, 95, 205, 63, 78, 59 },     // left = d135
+      { 41, 53, 52, 148, 71, 142, 65, 128, 51 },    // left = d117
+      { 40, 36, 28, 143, 143, 202, 40, 55, 137 },   // left = d153
+      { 52, 34, 29, 129, 183, 227, 42, 35, 43 },    // left = d207
+      { 42, 44, 44, 104, 105, 164, 64, 130, 80 },   // left = d63
+      { 43, 81, 53, 140, 169, 204, 68, 84, 72 }     // left = tm
+  }
+};
+
+// Default probabilities for signaling Intra mode for Y plane -- used only for
+// inter frames. ('av1_kf_y_mode_prob' is used for intra-only frames).
+// Context used: block size group.
+static const aom_prob default_if_y_probs[BLOCK_SIZE_GROUPS][INTRA_MODES - 1] = {
+  { 65, 32, 18, 144, 162, 194, 41, 51, 98 },   // block_size < 8x8
+  { 132, 68, 18, 165, 217, 196, 45, 40, 78 },  // block_size < 16x16
+  { 173, 80, 19, 176, 240, 193, 64, 35, 46 },  // block_size < 32x32
+  { 221, 135, 38, 194, 248, 121, 96, 85, 29 }  // block_size >= 32x32
+};
+
+// Default probabilities for signaling Intra mode for UV plane -- common for
+// both intra and inter frames.
+// Context used: Intra mode used by Y plane of the same block.
+static const aom_prob default_uv_probs[INTRA_MODES][INTRA_MODES - 1] = {
+  { 120, 7, 76, 176, 208, 126, 28, 54, 103 },   // y = dc
+  { 48, 12, 154, 155, 139, 90, 34, 117, 119 },  // y = v
+  { 67, 6, 25, 204, 243, 158, 13, 21, 96 },     // y = h
+  { 97, 5, 44, 131, 176, 139, 48, 68, 97 },     // y = d45
+  { 83, 5, 42, 156, 111, 152, 26, 49, 152 },    // y = d135
+  { 80, 5, 58, 178, 74, 83, 33, 62, 145 },      // y = d117
+  { 86, 5, 32, 154, 192, 168, 14, 22, 163 },    // y = d153
+  { 85, 5, 32, 156, 216, 148, 19, 29, 73 },     // y = d207
+  { 77, 7, 64, 116, 132, 122, 37, 126, 120 },   // y = d63
+  { 101, 21, 107, 181, 192, 103, 19, 67, 125 }  // y = tm
+};
+
+#endif  // CONFIG_ALT_INTRA
+
+#if CONFIG_EXT_PARTITION_TYPES
+static const aom_prob
+    default_partition_probs[PARTITION_CONTEXTS][EXT_PARTITION_TYPES - 1] = {
+      // 8x8 -> 4x4
+      { 199, 122, 141, 128, 128, 128, 128 },  // a/l both not split
+      { 147, 63, 159, 128, 128, 128, 128 },   // a split, l not split
+      { 148, 133, 118, 128, 128, 128, 128 },  // l split, a not split
+      { 121, 104, 114, 128, 128, 128, 128 },  // a/l both split
+      // 16x16 -> 8x8
+      { 174, 73, 87, 128, 128, 128, 128 },  // a/l both not split
+      { 92, 41, 83, 128, 128, 128, 128 },   // a split, l not split
+      { 82, 99, 50, 128, 128, 128, 128 },   // l split, a not split
+      { 53, 39, 39, 128, 128, 128, 128 },   // a/l both split
+      // 32x32 -> 16x16
+      { 177, 58, 59, 128, 128, 128, 128 },  // a/l both not split
+      { 68, 26, 63, 128, 128, 128, 128 },   // a split, l not split
+      { 52, 79, 25, 128, 128, 128, 128 },   // l split, a not split
+      { 17, 14, 12, 128, 128, 128, 128 },   // a/l both split
+      // 64x64 -> 32x32
+      { 222, 34, 30, 128, 128, 128, 128 },  // a/l both not split
+      { 72, 16, 44, 128, 128, 128, 128 },   // a split, l not split
+      { 58, 32, 12, 128, 128, 128, 128 },   // l split, a not split
+      { 10, 7, 6, 128, 128, 128, 128 },     // a/l both split
+#if CONFIG_EXT_PARTITION
+      // 128x128 -> 64x64
+      { 222, 34, 30, 128, 128, 128, 128 },  // a/l both not split
+      { 72, 16, 44, 128, 128, 128, 128 },   // a split, l not split
+      { 58, 32, 12, 128, 128, 128, 128 },   // l split, a not split
+      { 10, 7, 6, 128, 128, 128, 128 },     // a/l both split
+#endif                                      // CONFIG_EXT_PARTITION
+#if CONFIG_UNPOISON_PARTITION_CTX
+      { 0, 0, 141, 0, 0, 0, 0 },  // 8x8 -> 4x4
+      { 0, 0, 87, 0, 0, 0, 0 },   // 16x16 -> 8x8
+      { 0, 0, 59, 0, 0, 0, 0 },   // 32x32 -> 16x16
+      { 0, 0, 30, 0, 0, 0, 0 },   // 64x64 -> 32x32
+#if CONFIG_EXT_PARTITION
+      { 0, 0, 30, 0, 0, 0, 0 },   // 128x128 -> 64x64
+#endif                            // CONFIG_EXT_PARTITION
+      { 0, 122, 0, 0, 0, 0, 0 },  // 8x8 -> 4x4
+      { 0, 73, 0, 0, 0, 0, 0 },   // 16x16 -> 8x8
+      { 0, 58, 0, 0, 0, 0, 0 },   // 32x32 -> 16x16
+      { 0, 34, 0, 0, 0, 0, 0 },   // 64x64 -> 32x32
+#if CONFIG_EXT_PARTITION
+      { 0, 34, 0, 0, 0, 0, 0 },  // 128x128 -> 64x64
+#endif                           // CONFIG_EXT_PARTITION
+#endif                           // CONFIG_UNPOISON_PARTITION_CTX
+    };
+#else
+static const aom_prob
+    default_partition_probs[PARTITION_CONTEXTS][PARTITION_TYPES - 1] = {
+      // 8x8 -> 4x4
+      { 199, 122, 141 },  // a/l both not split
+      { 147, 63, 159 },   // a split, l not split
+      { 148, 133, 118 },  // l split, a not split
+      { 121, 104, 114 },  // a/l both split
+      // 16x16 -> 8x8
+      { 174, 73, 87 },  // a/l both not split
+      { 92, 41, 83 },   // a split, l not split
+      { 82, 99, 50 },   // l split, a not split
+      { 53, 39, 39 },   // a/l both split
+      // 32x32 -> 16x16
+      { 177, 58, 59 },  // a/l both not split
+      { 68, 26, 63 },   // a split, l not split
+      { 52, 79, 25 },   // l split, a not split
+      { 17, 14, 12 },   // a/l both split
+      // 64x64 -> 32x32
+      { 222, 34, 30 },  // a/l both not split
+      { 72, 16, 44 },   // a split, l not split
+      { 58, 32, 12 },   // l split, a not split
+      { 10, 7, 6 },     // a/l both split
+#if CONFIG_EXT_PARTITION
+      // 128x128 -> 64x64
+      { 222, 34, 30 },  // a/l both not split
+      { 72, 16, 44 },   // a split, l not split
+      { 58, 32, 12 },   // l split, a not split
+      { 10, 7, 6 },     // a/l both split
+#endif  // CONFIG_EXT_PARTITION
+#if CONFIG_UNPOISON_PARTITION_CTX
+      { 0, 0, 141 },    // 8x8 -> 4x4
+      { 0, 0, 87 },     // 16x16 -> 8x8
+      { 0, 0, 59 },     // 32x32 -> 16x16
+      { 0, 0, 30 },     // 64x64 -> 32x32
+#if CONFIG_EXT_PARTITION
+      { 0, 0, 30 },     // 128x128 -> 64x64
+#endif  // CONFIG_EXT_PARTITION
+      { 0, 122, 0 },    // 8x8 -> 4x4
+      { 0, 73, 0 },     // 16x16 -> 8x8
+      { 0, 58, 0 },     // 32x32 -> 16x16
+      { 0, 34, 0 },     // 64x64 -> 32x32
+#if CONFIG_EXT_PARTITION
+      { 0, 34, 0 },     // 128x128 -> 64x64
+#endif  // CONFIG_EXT_PARTITION
+#endif  // CONFIG_UNPOISON_PARTITION_CTX
+    };
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+#if CONFIG_REF_MV
+static const aom_prob default_newmv_prob[NEWMV_MODE_CONTEXTS] = {
+  200, 180, 150, 150, 110, 70, 60,
+};
+
+static const aom_prob default_zeromv_prob[ZEROMV_MODE_CONTEXTS] = {
+  192, 64,
+};
+
+static const aom_prob default_refmv_prob[REFMV_MODE_CONTEXTS] = {
+  220, 220, 200, 200, 180, 128, 30, 220, 30,
+};
+
+static const aom_prob default_drl_prob[DRL_MODE_CONTEXTS] = { 128, 160, 180,
+                                                              128, 160 };
+#endif  // CONFIG_REF_MV
+
+static const aom_prob
+    default_inter_mode_probs[INTER_MODE_CONTEXTS][INTER_MODES - 1] = {
+      { 2, 173, 34 },  // 0 = both zero mv
+      { 7, 145, 85 },  // 1 = one zero mv + one a predicted mv
+      { 7, 166, 63 },  // 2 = two predicted mvs
+      { 7, 94, 66 },   // 3 = one predicted/zero and one new mv
+      { 8, 64, 46 },   // 4 = two new mvs
+      { 17, 81, 31 },  // 5 = one intra neighbour + x
+      { 25, 29, 30 },  // 6 = two intra neighbours
+    };
+
+#if CONFIG_EXT_INTER
+static const aom_prob default_inter_compound_mode_probs
+    [INTER_MODE_CONTEXTS][INTER_COMPOUND_MODES - 1] = {
+      { 2, 173, 68, 192, 64, 192, 128, 180, 180 },   // 0 = both zero mv
+      { 7, 145, 160, 192, 64, 192, 128, 180, 180 },  // 1 = 1 zero + 1 predicted
+      { 7, 166, 126, 192, 64, 192, 128, 180, 180 },  // 2 = two predicted mvs
+      { 7, 94, 132, 192, 64, 192, 128, 180, 180 },   // 3 = 1 pred/zero, 1 new
+      { 8, 64, 64, 192, 64, 192, 128, 180, 180 },    // 4 = two new mvs
+      { 17, 81, 52, 192, 64, 192, 128, 180, 180 },   // 5 = one intra neighbour
+      { 25, 29, 50, 192, 64, 192, 128, 180, 180 },   // 6 = two intra neighbours
+    };
+
+#if CONFIG_COMPOUND_SINGLEREF
+// TODO(zoeliu): Default values to be further adjusted based on the collected
+//               stats.
+static const aom_prob default_inter_singleref_comp_mode_probs
+    [INTER_MODE_CONTEXTS][INTER_SINGLEREF_COMP_MODES - 1] = {
+      { 2, 173, 68, 180 },   // 0 = both zero mv
+      { 7, 145, 160, 180 },  // 1 = 1 zero + 1 predicted
+      { 7, 166, 126, 180 },  // 2 = two predicted mvs
+      { 7, 94, 132, 180 },   // 3 = 1 pred/zero, 1 new
+      { 8, 64, 64, 180 },    // 4 = two new mvs
+      { 17, 81, 52, 180 },   // 5 = one intra neighbour
+      { 25, 29, 50, 180 },   // 6 = two intra neighbours
+    };
+#endif  // CONFIG_COMPOUND_SINGLEREF
+
+#if CONFIG_COMPOUND_SEGMENT && CONFIG_WEDGE
+static const aom_prob
+    default_compound_type_probs[BLOCK_SIZES][COMPOUND_TYPES - 1] = {
+#if CONFIG_CB4X4
+      { 255, 255 }, { 255, 255 }, { 255, 255 },
+#endif
+      { 208, 200 }, { 208, 200 }, { 208, 200 }, { 208, 200 }, { 208, 200 },
+      { 208, 200 }, { 216, 200 }, { 216, 200 }, { 216, 200 }, { 224, 200 },
+      { 224, 200 }, { 240, 200 }, { 240, 200 },
+#if CONFIG_EXT_PARTITION
+      { 255, 200 }, { 255, 200 }, { 255, 200 },
+#endif  // CONFIG_EXT_PARTITION
+    };
+#elif !CONFIG_COMPOUND_SEGMENT && CONFIG_WEDGE
+static const aom_prob
+    default_compound_type_probs[BLOCK_SIZES][COMPOUND_TYPES - 1] = {
+#if CONFIG_CB4X4
+      { 208 }, { 208 }, { 208 },
+#endif
+      { 208 }, { 208 }, { 208 }, { 208 }, { 208 }, { 208 }, { 216 },
+      { 216 }, { 216 }, { 224 }, { 224 }, { 240 }, { 240 },
+#if CONFIG_EXT_PARTITION
+      { 255 }, { 255 }, { 255 },
+#endif  // CONFIG_EXT_PARTITION
+    };
+#elif CONFIG_COMPOUND_SEGMENT && !CONFIG_WEDGE
+static const aom_prob
+    default_compound_type_probs[BLOCK_SIZES][COMPOUND_TYPES - 1] = {
+#if CONFIG_CB4X4
+      { 208 }, { 208 }, { 208 },
+#endif
+      { 208 }, { 208 }, { 208 }, { 208 }, { 208 }, { 208 }, { 216 },
+      { 216 }, { 216 }, { 224 }, { 224 }, { 240 }, { 240 },
+#if CONFIG_EXT_PARTITION
+      { 255 }, { 255 }, { 255 },
+#endif  // CONFIG_EXT_PARTITION
+    };
+#else
+static const aom_prob default_compound_type_probs[BLOCK_SIZES]
+                                                 [COMPOUND_TYPES - 1];
+#endif  // CONFIG_COMPOUND_SEGMENT && CONFIG_WEDGE
+
+static const aom_prob default_interintra_prob[BLOCK_SIZE_GROUPS] = {
+  208, 208, 208, 208,
+};
+
+static const aom_prob
+    default_interintra_mode_prob[BLOCK_SIZE_GROUPS][INTERINTRA_MODES - 1] = {
+      { 65, 32, 18, 144, 162, 194, 41, 51, 98 },   // block_size < 8x8
+      { 132, 68, 18, 165, 217, 196, 45, 40, 78 },  // block_size < 16x16
+      { 173, 80, 19, 176, 240, 193, 64, 35, 46 },  // block_size < 32x32
+      { 221, 135, 38, 194, 248, 121, 96, 85, 29 }  // block_size >= 32x32
+    };
+
+static const aom_prob default_wedge_interintra_prob[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  208, 208, 208,
+#endif
+  208, 208, 208, 208, 208, 208, 216, 216, 216, 224, 224, 224, 240,
+#if CONFIG_EXT_PARTITION
+  208, 208, 208
+#endif  // CONFIG_EXT_PARTITION
+};
+#endif  // CONFIG_EXT_INTER
+
+// Change this section appropriately once warped motion is supported
+#if CONFIG_MOTION_VAR && !CONFIG_WARPED_MOTION
+const aom_tree_index av1_motion_mode_tree[TREE_SIZE(MOTION_MODES)] = {
+  -SIMPLE_TRANSLATION, -OBMC_CAUSAL
+};
+static const aom_prob default_motion_mode_prob[BLOCK_SIZES][MOTION_MODES - 1] =
+    {
+#if CONFIG_CB4X4
+      { 255 }, { 255 }, { 255 },
+#endif
+      { 255 }, { 255 }, { 255 }, { 151 }, { 153 }, { 144 }, { 178 },
+      { 165 }, { 160 }, { 207 }, { 195 }, { 168 }, { 244 },
+#if CONFIG_EXT_PARTITION
+      { 252 }, { 252 }, { 252 },
+#endif  // CONFIG_EXT_PARTITION
+    };
+
+#elif !CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+
+const aom_tree_index av1_motion_mode_tree[TREE_SIZE(MOTION_MODES)] = {
+  -SIMPLE_TRANSLATION, -WARPED_CAUSAL
+};
+
+static const aom_prob default_motion_mode_prob[BLOCK_SIZES][MOTION_MODES - 1] =
+    {
+#if CONFIG_CB4X4
+      { 255 }, { 255 }, { 255 },
+#endif
+      { 255 }, { 255 }, { 255 }, { 151 }, { 153 }, { 144 }, { 178 },
+      { 165 }, { 160 }, { 207 }, { 195 }, { 168 }, { 244 },
+#if CONFIG_EXT_PARTITION
+      { 252 }, { 252 }, { 252 },
+#endif  // CONFIG_EXT_PARTITION
+    };
+
+#elif CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+
+const aom_tree_index av1_motion_mode_tree[TREE_SIZE(MOTION_MODES)] = {
+  -SIMPLE_TRANSLATION, 2, -OBMC_CAUSAL, -WARPED_CAUSAL,
+};
+static const aom_prob default_motion_mode_prob[BLOCK_SIZES][MOTION_MODES - 1] =
+    {
+#if CONFIG_CB4X4
+      { 255, 200 }, { 255, 200 }, { 255, 200 },
+#endif
+      { 255, 200 }, { 255, 200 }, { 255, 200 }, { 151, 200 }, { 153, 200 },
+      { 144, 200 }, { 178, 200 }, { 165, 200 }, { 160, 200 }, { 207, 200 },
+      { 195, 200 }, { 168, 200 }, { 244, 200 },
+#if CONFIG_EXT_PARTITION
+      { 252, 200 }, { 252, 200 }, { 252, 200 },
+#endif  // CONFIG_EXT_PARTITION
+    };
+
+// Probability for the case that only 1 additional motion mode is allowed
+static const aom_prob default_obmc_prob[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  255, 255, 255,
+#endif
+  255, 255, 255, 151, 153, 144, 178, 165, 160, 207, 195, 168, 244,
+#if CONFIG_EXT_PARTITION
+  252, 252, 252,
+#endif  // CONFIG_EXT_PARTITION
+};
+#endif
+
+#if CONFIG_DELTA_Q
+static const aom_prob default_delta_q_probs[DELTA_Q_PROBS] = { 220, 220, 220 };
+#if CONFIG_EC_MULTISYMBOL
+static const aom_cdf_prob default_delta_q_cdf[CDF_SIZE(DELTA_Q_PROBS + 1)] = {
+  AOM_ICDF(28160), AOM_ICDF(32120), AOM_ICDF(32677), AOM_ICDF(32768), 0
+};
+#endif
+#if CONFIG_EXT_DELTA_Q
+static const aom_prob default_delta_lf_probs[DELTA_LF_PROBS] = { 220, 220,
+                                                                 220 };
+#if CONFIG_EC_MULTISYMBOL
+static const aom_cdf_prob default_delta_lf_cdf[CDF_SIZE(DELTA_LF_PROBS + 1)] = {
+  28160, 32120, 32677, 32768, 0
+};
+#endif
+#endif
+#endif
+#if CONFIG_EC_MULTISYMBOL
+int av1_intra_mode_ind[INTRA_MODES];
+int av1_intra_mode_inv[INTRA_MODES];
+int av1_inter_mode_ind[INTER_MODES];
+int av1_inter_mode_inv[INTER_MODES];
+#if CONFIG_EXT_TX
+int av1_ext_tx_intra_ind[EXT_TX_SETS_INTRA][TX_TYPES];
+int av1_ext_tx_intra_inv[EXT_TX_SETS_INTRA][TX_TYPES];
+int av1_ext_tx_inter_ind[EXT_TX_SETS_INTER][TX_TYPES];
+int av1_ext_tx_inter_inv[EXT_TX_SETS_INTER][TX_TYPES];
+#endif
+#endif
+
+#if CONFIG_ALT_INTRA
+const aom_tree_index av1_intra_mode_tree[TREE_SIZE(INTRA_MODES)] = {
+  -DC_PRED,   2,            /* 0 = DC_NODE */
+  -TM_PRED,   4,            /* 1 = TM_NODE */
+  -V_PRED,    6,            /* 2 = V_NODE */
+  8,          12,           /* 3 = COM_NODE */
+  -H_PRED,    10,           /* 4 = H_NODE */
+  -D135_PRED, -D117_PRED,   /* 5 = D135_NODE */
+  -D45_PRED,  14,           /* 6 = D45_NODE */
+  -D63_PRED,  16,           /* 7 = D63_NODE */
+  -D153_PRED, 18,           /* 8 = D153_NODE */
+  -D207_PRED, -SMOOTH_PRED, /* 9 = D207_NODE */
+};
+#else
+const aom_tree_index av1_intra_mode_tree[TREE_SIZE(INTRA_MODES)] = {
+  -DC_PRED,   2,          /* 0 = DC_NODE */
+  -TM_PRED,   4,          /* 1 = TM_NODE */
+  -V_PRED,    6,          /* 2 = V_NODE */
+  8,          12,         /* 3 = COM_NODE */
+  -H_PRED,    10,         /* 4 = H_NODE */
+  -D135_PRED, -D117_PRED, /* 5 = D135_NODE */
+  -D45_PRED,  14,         /* 6 = D45_NODE */
+  -D63_PRED,  16,         /* 7 = D63_NODE */
+  -D153_PRED, -D207_PRED  /* 8 = D153_NODE */
+};
+#endif  // CONFIG_ALT_INTRA
+
+const aom_tree_index av1_inter_mode_tree[TREE_SIZE(INTER_MODES)] = {
+  -INTER_OFFSET(ZEROMV), 2, -INTER_OFFSET(NEARESTMV), 4, -INTER_OFFSET(NEARMV),
+  -INTER_OFFSET(NEWMV)
+};
+
+#if CONFIG_EXT_INTER
+/* clang-format off */
+const aom_tree_index av1_interintra_mode_tree[TREE_SIZE(INTERINTRA_MODES)] = {
+  -II_DC_PRED, 2,                   /* 0 = II_DC_NODE     */
+  -II_TM_PRED, 4,                   /* 1 = II_TM_NODE     */
+  -II_V_PRED, 6,                    /* 2 = II_V_NODE      */
+  8, 12,                            /* 3 = II_COM_NODE    */
+  -II_H_PRED, 10,                   /* 4 = II_H_NODE      */
+  -II_D135_PRED, -II_D117_PRED,     /* 5 = II_D135_NODE   */
+  -II_D45_PRED, 14,                 /* 6 = II_D45_NODE    */
+  -II_D63_PRED, 16,                 /* 7 = II_D63_NODE    */
+  -II_D153_PRED, -II_D207_PRED      /* 8 = II_D153_NODE   */
+};
+
+const aom_tree_index av1_inter_compound_mode_tree
+    [TREE_SIZE(INTER_COMPOUND_MODES)] = {
+  -INTER_COMPOUND_OFFSET(ZERO_ZEROMV), 2,
+  -INTER_COMPOUND_OFFSET(NEAREST_NEARESTMV), 4,
+  6, -INTER_COMPOUND_OFFSET(NEW_NEWMV),
+  8, 12,
+  -INTER_COMPOUND_OFFSET(NEAR_NEARMV), 10,
+  -INTER_COMPOUND_OFFSET(NEAREST_NEARMV),
+      -INTER_COMPOUND_OFFSET(NEAR_NEARESTMV),
+  14, 16,
+  -INTER_COMPOUND_OFFSET(NEAREST_NEWMV), -INTER_COMPOUND_OFFSET(NEW_NEARESTMV),
+  -INTER_COMPOUND_OFFSET(NEAR_NEWMV), -INTER_COMPOUND_OFFSET(NEW_NEARMV)
+};
+
+#if CONFIG_COMPOUND_SINGLEREF
+const aom_tree_index av1_inter_singleref_comp_mode_tree
+    [TREE_SIZE(INTER_SINGLEREF_COMP_MODES)] = {
+  -INTER_SINGLEREF_COMP_OFFSET(SR_ZERO_NEWMV), 2,
+  -INTER_SINGLEREF_COMP_OFFSET(SR_NEAREST_NEARMV), 4,
+  6, -INTER_SINGLEREF_COMP_OFFSET(SR_NEW_NEWMV),
+  -INTER_SINGLEREF_COMP_OFFSET(SR_NEAREST_NEWMV),
+  -INTER_SINGLEREF_COMP_OFFSET(SR_NEAR_NEWMV)
+};
+#endif  // CONFIG_COMPOUND_SINGLEREF
+
+#if CONFIG_COMPOUND_SEGMENT && CONFIG_WEDGE
+const aom_tree_index av1_compound_type_tree[TREE_SIZE(COMPOUND_TYPES)] = {
+  -COMPOUND_AVERAGE, 2, -COMPOUND_WEDGE, -COMPOUND_SEG
+};
+#elif !CONFIG_COMPOUND_SEGMENT && CONFIG_WEDGE
+const aom_tree_index av1_compound_type_tree[TREE_SIZE(COMPOUND_TYPES)] = {
+  -COMPOUND_AVERAGE, -COMPOUND_WEDGE
+};
+#elif CONFIG_COMPOUND_SEGMENT && !CONFIG_WEDGE
+const aom_tree_index av1_compound_type_tree[TREE_SIZE(COMPOUND_TYPES)] = {
+  -COMPOUND_AVERAGE, -COMPOUND_SEG
+};
+#else
+const aom_tree_index av1_compound_type_tree[TREE_SIZE(COMPOUND_TYPES)] = {};
+#endif  // CONFIG_COMPOUND_SEGMENT && CONFIG_WEDGE
+/* clang-format on */
+#endif  // CONFIG_EXT_INTER
+
+const aom_tree_index av1_partition_tree[TREE_SIZE(PARTITION_TYPES)] = {
+  -PARTITION_NONE, 2, -PARTITION_HORZ, 4, -PARTITION_VERT, -PARTITION_SPLIT
+};
+
+#if CONFIG_EXT_PARTITION_TYPES
+/* clang-format off */
+const aom_tree_index av1_ext_partition_tree[TREE_SIZE(EXT_PARTITION_TYPES)] = {
+  -PARTITION_NONE, 2,
+  6, 4,
+  8, -PARTITION_SPLIT,
+  -PARTITION_HORZ, 10,
+  -PARTITION_VERT, 12,
+  -PARTITION_HORZ_A, -PARTITION_HORZ_B,
+  -PARTITION_VERT_A, -PARTITION_VERT_B
+};
+/* clang-format on */
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+static const aom_prob default_intra_inter_p[INTRA_INTER_CONTEXTS] = {
+  9, 102, 187, 225
+};
+
+static const aom_prob default_comp_inter_p[COMP_INTER_CONTEXTS] = {
+  239, 183, 119, 96, 41
+};
+
+#if CONFIG_EXT_REFS
+static const aom_prob default_comp_ref_p[REF_CONTEXTS][FWD_REFS - 1] = {
+  // TODO(zoeliu): To adjust the initial prob values.
+  { 33, 16, 16 },
+  { 77, 74, 74 },
+  { 142, 142, 142 },
+  { 172, 170, 170 },
+  { 238, 247, 247 }
+};
+static const aom_prob default_comp_bwdref_p[REF_CONTEXTS][BWD_REFS - 1] = {
+  { 16 }, { 74 }, { 142 }, { 170 }, { 247 }
+};
+#else
+static const aom_prob default_comp_ref_p[REF_CONTEXTS][COMP_REFS - 1] = {
+  { 50 }, { 126 }, { 123 }, { 221 }, { 226 }
+};
+#endif  // CONFIG_EXT_REFS
+
+static const aom_prob default_single_ref_p[REF_CONTEXTS][SINGLE_REFS - 1] = {
+#if CONFIG_EXT_REFS
+  { 33, 16, 16, 16, 16 },
+  { 77, 74, 74, 74, 74 },
+  { 142, 142, 142, 142, 142 },
+  { 172, 170, 170, 170, 170 },
+  { 238, 247, 247, 247, 247 }
+#else
+  { 33, 16 }, { 77, 74 }, { 142, 142 }, { 172, 170 }, { 238, 247 }
+#endif  // CONFIG_EXT_REFS
+};
+
+#if CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+// TODO(zoeliu): Default values to be further adjusted based on the collected
+//               stats.
+static const aom_prob default_comp_inter_mode_p[COMP_INTER_MODE_CONTEXTS] = {
+  41, 119, 187, 225
+};
+#endif  // CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+
+#if CONFIG_PALETTE
+
+// Tree to code palette size (number of colors in a palette) and the
+// corresponding probabilities for Y and UV planes.
+const aom_tree_index av1_palette_size_tree[TREE_SIZE(PALETTE_SIZES)] = {
+  -TWO_COLORS,  2, -THREE_COLORS, 4,  -FOUR_COLORS,  6,
+  -FIVE_COLORS, 8, -SIX_COLORS,   10, -SEVEN_COLORS, -EIGHT_COLORS,
+};
+
+// TODO(huisu): tune these probs
+const aom_prob
+    av1_default_palette_y_size_prob[PALETTE_BLOCK_SIZES][PALETTE_SIZES - 1] = {
+      { 96, 89, 100, 64, 77, 130 },   { 22, 15, 44, 16, 34, 82 },
+      { 30, 19, 57, 18, 38, 86 },     { 94, 36, 104, 23, 43, 92 },
+      { 116, 76, 107, 46, 65, 105 },  { 112, 82, 94, 40, 70, 112 },
+      { 147, 124, 123, 58, 69, 103 }, { 180, 113, 136, 49, 45, 114 },
+      { 107, 70, 87, 49, 154, 156 },  { 98, 105, 142, 63, 64, 152 },
+#if CONFIG_EXT_PARTITION
+      { 98, 105, 142, 63, 64, 152 },  { 98, 105, 142, 63, 64, 152 },
+      { 98, 105, 142, 63, 64, 152 },
+#endif  // CONFIG_EXT_PARTITION
+    };
+
+const aom_prob
+    av1_default_palette_uv_size_prob[PALETTE_BLOCK_SIZES][PALETTE_SIZES - 1] = {
+      { 160, 196, 228, 213, 175, 230 }, { 87, 148, 208, 141, 166, 163 },
+      { 72, 151, 204, 139, 155, 161 },  { 78, 135, 171, 104, 120, 173 },
+      { 59, 92, 131, 78, 92, 142 },     { 75, 118, 149, 84, 90, 128 },
+      { 89, 87, 92, 66, 66, 128 },      { 67, 53, 54, 55, 66, 93 },
+      { 120, 130, 83, 171, 75, 214 },   { 72, 55, 66, 68, 79, 107 },
+#if CONFIG_EXT_PARTITION
+      { 72, 55, 66, 68, 79, 107 },      { 72, 55, 66, 68, 79, 107 },
+      { 72, 55, 66, 68, 79, 107 },
+#endif  // CONFIG_EXT_PARTITION
+    };
+
+// When palette mode is enabled, following probability tables indicate the
+// probabilities to code the "is_palette" bit (i.e. the bit that indicates
+// if this block uses palette mode or DC_PRED mode).
+const aom_prob av1_default_palette_y_mode_prob
+    [PALETTE_BLOCK_SIZES][PALETTE_Y_MODE_CONTEXTS] = {
+      { 240, 180, 100 }, { 240, 180, 100 }, { 240, 180, 100 },
+      { 240, 180, 100 }, { 240, 180, 100 }, { 240, 180, 100 },
+      { 240, 180, 100 }, { 240, 180, 100 }, { 240, 180, 100 },
+      { 240, 180, 100 },
+#if CONFIG_EXT_PARTITION
+      { 240, 180, 100 }, { 240, 180, 100 }, { 240, 180, 100 },
+#endif  // CONFIG_EXT_PARTITION
+    };
+
+const aom_prob av1_default_palette_uv_mode_prob[PALETTE_UV_MODE_CONTEXTS] = {
+  253, 229
+};
+
+// Trees to code palette color indices (for various palette sizes), and the
+// corresponding probability tables for Y and UV planes.
+const aom_tree_index
+    av1_palette_color_index_tree[PALETTE_SIZES][TREE_SIZE(PALETTE_COLORS)] = {
+      { // 2 colors
+        -PALETTE_COLOR_ONE, -PALETTE_COLOR_TWO },
+      { // 3 colors
+        -PALETTE_COLOR_ONE, 2, -PALETTE_COLOR_TWO, -PALETTE_COLOR_THREE },
+      { // 4 colors
+        -PALETTE_COLOR_ONE, 2, -PALETTE_COLOR_TWO, 4, -PALETTE_COLOR_THREE,
+        -PALETTE_COLOR_FOUR },
+      { // 5 colors
+        -PALETTE_COLOR_ONE, 2, -PALETTE_COLOR_TWO, 4, -PALETTE_COLOR_THREE, 6,
+        -PALETTE_COLOR_FOUR, -PALETTE_COLOR_FIVE },
+      { // 6 colors
+        -PALETTE_COLOR_ONE, 2, -PALETTE_COLOR_TWO, 4, -PALETTE_COLOR_THREE, 6,
+        -PALETTE_COLOR_FOUR, 8, -PALETTE_COLOR_FIVE, -PALETTE_COLOR_SIX },
+      { // 7 colors
+        -PALETTE_COLOR_ONE, 2, -PALETTE_COLOR_TWO, 4, -PALETTE_COLOR_THREE, 6,
+        -PALETTE_COLOR_FOUR, 8, -PALETTE_COLOR_FIVE, 10, -PALETTE_COLOR_SIX,
+        -PALETTE_COLOR_SEVEN },
+      { // 8 colors
+        -PALETTE_COLOR_ONE, 2, -PALETTE_COLOR_TWO, 4, -PALETTE_COLOR_THREE, 6,
+        -PALETTE_COLOR_FOUR, 8, -PALETTE_COLOR_FIVE, 10, -PALETTE_COLOR_SIX, 12,
+        -PALETTE_COLOR_SEVEN, -PALETTE_COLOR_EIGHT },
+    };
+
+// Note: Has to be non-zero to avoid any asserts triggering.
+#define UNUSED_PROB 128
+
+const aom_prob av1_default_palette_y_color_index_prob
+    [PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS][PALETTE_COLORS - 1] = {
+      {
+          // 2 colors
+          { 231, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB, UNUSED_PROB },
+          { UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB, UNUSED_PROB },
+          { 69, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+          { 224, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB, UNUSED_PROB },
+          { 249, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB, UNUSED_PROB },
+      },
+      {
+          // 3 colors
+          { 219, 124, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+          { 91, 191, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+          { 34, 237, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+          { 184, 118, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+          { 252, 124, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+      },
+      {
+          // 4 colors
+          { 204, 87, 97, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 74, 144, 129, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 52, 191, 134, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 151, 85, 147, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 248, 60, 115, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+      },
+      {
+          // 5 colors
+          { 218, 69, 62, 106, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 76, 143, 89, 127, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 21, 233, 94, 131, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 172, 72, 89, 112, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 253, 66, 65, 128, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+      },
+      {
+          // 6 colors
+          { 190, 60, 47, 54, 74, UNUSED_PROB, UNUSED_PROB },
+          { 62, 106, 51, 95, 110, UNUSED_PROB, UNUSED_PROB },
+          { 52, 180, 69, 72, 107, UNUSED_PROB, UNUSED_PROB },
+          { 156, 83, 72, 83, 101, UNUSED_PROB, UNUSED_PROB },
+          { 245, 45, 37, 52, 91, UNUSED_PROB, UNUSED_PROB },
+      },
+      {
+          // 7 colors
+          { 206, 56, 42, 42, 53, 85, UNUSED_PROB },
+          { 70, 100, 45, 68, 77, 94, UNUSED_PROB },
+          { 57, 169, 51, 62, 74, 119, UNUSED_PROB },
+          { 172, 76, 71, 40, 59, 76, UNUSED_PROB },
+          { 248, 47, 36, 53, 61, 110, UNUSED_PROB },
+      },
+      {
+          // 8 colors
+          { 208, 52, 38, 34, 34, 44, 66 },
+          { 52, 107, 34, 73, 69, 82, 87 },
+          { 28, 208, 53, 43, 62, 70, 102 },
+          { 184, 64, 45, 37, 37, 69, 105 },
+          { 251, 18, 31, 45, 47, 61, 104 },
+      },
+    };
+
+const aom_prob av1_default_palette_uv_color_index_prob
+    [PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS][PALETTE_COLORS - 1] = {
+      {
+          // 2 colors
+          { 233, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB, UNUSED_PROB },
+          { UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB, UNUSED_PROB },
+          { 69, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+          { 240, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB, UNUSED_PROB },
+          { 248, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB, UNUSED_PROB },
+      },
+      {
+          // 3 colors
+          { 216, 128, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+          { 110, 171, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+          { 40, 239, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+          { 191, 104, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+          { 247, 134, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB,
+            UNUSED_PROB },
+      },
+      {
+          // 4 colors
+          { 202, 89, 132, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 90, 132, 136, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 63, 195, 149, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 152, 84, 152, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 241, 87, 136, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+      },
+      {
+          // 5 colors
+          { 209, 54, 82, 134, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 94, 173, 180, 93, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 10, 251, 127, 84, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 183, 20, 150, 47, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+          { 252, 73, 111, 150, UNUSED_PROB, UNUSED_PROB, UNUSED_PROB },
+      },
+      {
+          // 6 colors
+          { 192, 67, 59, 46, 184, UNUSED_PROB, UNUSED_PROB },
+          { 59, 92, 61, 100, 130, UNUSED_PROB, UNUSED_PROB },
+          { 49, 162, 68, 91, 150, UNUSED_PROB, UNUSED_PROB },
+          { 133, 29, 36, 153, 101, UNUSED_PROB, UNUSED_PROB },
+          { 247, 71, 44, 90, 129, UNUSED_PROB, UNUSED_PROB },
+      },
+      {
+          // 7 colors
+          { 182, 62, 80, 78, 46, 116, UNUSED_PROB },
+          { 59, 62, 39, 81, 65, 99, UNUSED_PROB },
+          { 54, 177, 48, 58, 93, 104, UNUSED_PROB },
+          { 137, 79, 54, 55, 44, 134, UNUSED_PROB },
+          { 239, 82, 79, 44, 69, 71, UNUSED_PROB },
+      },
+      {
+          // 8 colors
+          { 172, 53, 27, 67, 30, 79, 113 },
+          { 63, 57, 45, 81, 62, 35, 47 },
+          { 51, 200, 36, 47, 82, 165, 129 },
+          { 141, 100, 47, 29, 33, 37, 129 },
+          { 236, 42, 50, 91, 24, 154, 65 },
+      },
+    };
+
+#undef UNUSED_PROB
+
+#define MAX_COLOR_CONTEXT_HASH 8
+// Negative values are invalid
+static const int palette_color_index_context_lookup[MAX_COLOR_CONTEXT_HASH +
+                                                    1] = { -1, -1, 0, -1, -1,
+                                                           4,  3,  2, 1 };
+
+#endif  // CONFIG_PALETTE
+
+// The transform size is coded as an offset to the smallest transform
+// block size.
+const aom_tree_index av1_tx_size_tree[MAX_TX_DEPTH][TREE_SIZE(TX_SIZES)] = {
+  {
+      // Max tx_size is 8X8
+      -0, -1,
+  },
+  {
+      // Max tx_size is 16X16
+      -0, 2, -1, -2,
+  },
+  {
+      // Max tx_size is 32X32
+      -0, 2, -1, 4, -2, -3,
+  },
+#if CONFIG_TX64X64
+  {
+      // Max tx_size is 64X64
+      -0, 2, -1, 4, -2, 6, -3, -4,
+  },
+#endif  // CONFIG_TX64X64
+};
+
+static const aom_prob default_tx_size_prob[MAX_TX_DEPTH][TX_SIZE_CONTEXTS]
+                                          [MAX_TX_DEPTH] = {
+                                            {
+                                                // Max tx_size is 8X8
+                                                { 100 },
+                                                { 66 },
+                                            },
+                                            {
+                                                // Max tx_size is 16X16
+                                                { 20, 152 },
+                                                { 15, 101 },
+                                            },
+                                            {
+                                                // Max tx_size is 32X32
+                                                { 3, 136, 37 },
+                                                { 5, 52, 13 },
+                                            },
+#if CONFIG_TX64X64
+                                            {
+                                                // Max tx_size is 64X64
+                                                { 1, 64, 136, 127 },
+                                                { 1, 32, 52, 67 },
+                                            },
+#endif  // CONFIG_TX64X64
+                                          };
+
+#if CONFIG_LOOP_RESTORATION
+const aom_tree_index
+    av1_switchable_restore_tree[TREE_SIZE(RESTORE_SWITCHABLE_TYPES)] = {
+      -RESTORE_NONE, 2, -RESTORE_WIENER, -RESTORE_SGRPROJ,
+    };
+
+static const aom_prob
+    default_switchable_restore_prob[RESTORE_SWITCHABLE_TYPES - 1] = {
+      32, 128,
+    };
+#endif  // CONFIG_LOOP_RESTORATION
+
+#if CONFIG_PALETTE
+#define NUM_PALETTE_NEIGHBORS 3  // left, top-left and top.
+int av1_get_palette_color_index_context(const uint8_t *color_map, int stride,
+                                        int r, int c, int palette_size,
+                                        uint8_t *color_order, int *color_idx) {
+  int i;
+  // The +10 below should not be needed. But we get a warning "array subscript
+  // is above array bounds [-Werror=array-bounds]" without it, possibly due to
+  // this (or similar) bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59124
+  int scores[PALETTE_MAX_SIZE + 10];
+  const int weights[NUM_PALETTE_NEIGHBORS] = { 2, 1, 2 };
+  const int hash_multipliers[NUM_PALETTE_NEIGHBORS] = { 1, 2, 2 };
+  int color_index_ctx_hash;
+  int color_index_ctx;
+  int color_neighbors[NUM_PALETTE_NEIGHBORS];
+  int inverse_color_order[PALETTE_MAX_SIZE];
+  assert(palette_size <= PALETTE_MAX_SIZE);
+  assert(r > 0 || c > 0);
+
+  // Get color indices of neighbors.
+  color_neighbors[0] = (c - 1 >= 0) ? color_map[r * stride + c - 1] : -1;
+  color_neighbors[1] =
+      (c - 1 >= 0 && r - 1 >= 0) ? color_map[(r - 1) * stride + c - 1] : -1;
+  color_neighbors[2] = (r - 1 >= 0) ? color_map[(r - 1) * stride + c] : -1;
+
+  for (i = 0; i < PALETTE_MAX_SIZE; ++i) {
+    color_order[i] = i;
+    inverse_color_order[i] = i;
+  }
+  memset(scores, 0, PALETTE_MAX_SIZE * sizeof(scores[0]));
+  for (i = 0; i < NUM_PALETTE_NEIGHBORS; ++i) {
+    if (color_neighbors[i] >= 0) {
+      scores[color_neighbors[i]] += weights[i];
+    }
+  }
+
+  // Get the top NUM_PALETTE_NEIGHBORS scores (sorted from large to small).
+  for (i = 0; i < NUM_PALETTE_NEIGHBORS; ++i) {
+    int max = scores[i];
+    int max_idx = i;
+    int j;
+    for (j = i + 1; j < palette_size; ++j) {
+      if (scores[j] > max) {
+        max = scores[j];
+        max_idx = j;
+      }
+    }
+    if (max_idx != i) {
+      // Move the score at index 'max_idx' to index 'i', and shift the scores
+      // from 'i' to 'max_idx - 1' by 1.
+      const int max_score = scores[max_idx];
+      const uint8_t max_color_order = color_order[max_idx];
+      int k;
+      for (k = max_idx; k > i; --k) {
+        scores[k] = scores[k - 1];
+        color_order[k] = color_order[k - 1];
+        inverse_color_order[color_order[k]] = k;
+      }
+      scores[i] = max_score;
+      color_order[i] = max_color_order;
+      inverse_color_order[color_order[i]] = i;
+    }
+  }
+
+  // Get hash value of context.
+  color_index_ctx_hash = 0;
+  for (i = 0; i < NUM_PALETTE_NEIGHBORS; ++i) {
+    color_index_ctx_hash += scores[i] * hash_multipliers[i];
+  }
+  assert(color_index_ctx_hash > 0);
+  assert(color_index_ctx_hash <= MAX_COLOR_CONTEXT_HASH);
+
+  // Lookup context from hash.
+  color_index_ctx = palette_color_index_context_lookup[color_index_ctx_hash];
+  assert(color_index_ctx >= 0);
+  assert(color_index_ctx < PALETTE_COLOR_INDEX_CONTEXTS);
+
+  if (color_idx != NULL) {
+    *color_idx = inverse_color_order[color_map[r * stride + c]];
+  }
+  return color_index_ctx;
+}
+#undef NUM_PALETTE_NEIGHBORS
+#undef MAX_COLOR_CONTEXT_HASH
+
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_VAR_TX
+static const aom_prob default_txfm_partition_probs[TXFM_PARTITION_CONTEXTS] = {
+  250, 231, 212, 241, 166, 66, 241, 230, 135, 243, 154, 64, 248, 161, 63, 128,
+};
+#endif
+
+static const aom_prob default_skip_probs[SKIP_CONTEXTS] = { 192, 128, 64 };
+
+#if CONFIG_DUAL_FILTER
+static const aom_prob default_switchable_interp_prob
+    [SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS - 1] = {
+      { 235, 192, 128 }, { 36, 243, 48 },   { 34, 16, 128 },
+      { 34, 16, 128 },   { 149, 160, 128 }, { 235, 192, 128 },
+      { 36, 243, 48 },   { 34, 16, 128 },   { 34, 16, 128 },
+      { 149, 160, 128 }, { 235, 192, 128 }, { 36, 243, 48 },
+      { 34, 16, 128 },   { 34, 16, 128 },   { 149, 160, 128 },
+      { 235, 192, 128 }, { 36, 243, 48 },   { 34, 16, 128 },
+      { 34, 16, 128 },   { 149, 160, 128 },
+    };
+#else   // CONFIG_DUAL_FILTER
+static const aom_prob default_switchable_interp_prob[SWITCHABLE_FILTER_CONTEXTS]
+                                                    [SWITCHABLE_FILTERS - 1] = {
+                                                      { 235, 162 },
+                                                      { 36, 255 },
+                                                      { 34, 3 },
+                                                      { 149, 144 },
+                                                    };
+#endif  // CONFIG_DUAL_FILTER
+
+#if CONFIG_EXT_TX
+/* clang-format off */
+const aom_tree_index av1_ext_tx_inter_tree[EXT_TX_SETS_INTER]
+                                           [TREE_SIZE(TX_TYPES)] = {
+  { // ToDo(yaowu): remove used entry 0.
+    0
+  }, {
+    -IDTX, 2,
+    4, 14,
+    6, 8,
+    -V_DCT, -H_DCT,
+    10, 12,
+    -V_ADST, -H_ADST,
+    -V_FLIPADST, -H_FLIPADST,
+    -DCT_DCT, 16,
+    18, 24,
+    20, 22,
+    -ADST_DCT, -DCT_ADST,
+    -FLIPADST_DCT, -DCT_FLIPADST,
+    26, 28,
+    -ADST_ADST, -FLIPADST_FLIPADST,
+    -ADST_FLIPADST, -FLIPADST_ADST
+  }, {
+    -IDTX, 2,
+    4, 6,
+    -V_DCT, -H_DCT,
+    -DCT_DCT, 8,
+    10, 16,
+    12, 14,
+    -ADST_DCT, -DCT_ADST,
+    -FLIPADST_DCT, -DCT_FLIPADST,
+    18, 20,
+    -ADST_ADST, -FLIPADST_FLIPADST,
+    -ADST_FLIPADST, -FLIPADST_ADST
+  }, {
+    -IDTX, -DCT_DCT,
+  }
+};
+
+const aom_tree_index av1_ext_tx_intra_tree[EXT_TX_SETS_INTRA]
+                                           [TREE_SIZE(TX_TYPES)] = {
+  {  // ToDo(yaowu): remove unused entry 0.
+    0
+  }, {
+    -IDTX, 2,
+    -DCT_DCT, 4,
+    6, 8,
+    -V_DCT, -H_DCT,
+    -ADST_ADST, 10,
+    -ADST_DCT, -DCT_ADST,
+  }, {
+    -IDTX, 2,
+    -DCT_DCT, 4,
+    -ADST_ADST, 6,
+    -ADST_DCT, -DCT_ADST,
+  }
+};
+/* clang-format on */
+
+static const aom_prob
+    default_inter_ext_tx_prob[EXT_TX_SETS_INTER][EXT_TX_SIZES][TX_TYPES - 1] = {
+      {
+// ToDo(yaowu): remove unused entry 0.
+#if CONFIG_CB4X4
+          { 0 },
+#endif
+          { 0 },
+          { 0 },
+          { 0 },
+          { 0 },
+      },
+      {
+#if CONFIG_CB4X4
+          { 0 },
+#endif
+          { 10, 24, 30, 128, 128, 128, 128, 112, 160, 128, 128, 128, 128, 128,
+            128 },
+          { 10, 24, 30, 128, 128, 128, 128, 112, 160, 128, 128, 128, 128, 128,
+            128 },
+          { 10, 24, 30, 128, 128, 128, 128, 112, 160, 128, 128, 128, 128, 128,
+            128 },
+          { 10, 24, 30, 128, 128, 128, 128, 112, 160, 128, 128, 128, 128, 128,
+            128 },
+      },
+      {
+#if CONFIG_CB4X4
+          { 0 },
+#endif
+          { 10, 30, 128, 112, 160, 128, 128, 128, 128, 128, 128 },
+          { 10, 30, 128, 112, 160, 128, 128, 128, 128, 128, 128 },
+          { 10, 30, 128, 112, 160, 128, 128, 128, 128, 128, 128 },
+          { 10, 30, 128, 112, 160, 128, 128, 128, 128, 128, 128 },
+      },
+      {
+#if CONFIG_CB4X4
+          { 0 },
+#endif
+          { 12 },
+          { 12 },
+          { 12 },
+          { 12 },
+      }
+    };
+
+// TODO(urvang): 3rd context should be tx_type instead of intra mode just like
+// the baseline.
+static const aom_prob
+    default_intra_ext_tx_prob[EXT_TX_SETS_INTRA][EXT_TX_SIZES][INTRA_MODES]
+                             [TX_TYPES - 1] = {
+                               {
+// ToDo(yaowu): remove unused entry 0.
+#if CONFIG_CB4X4
+                                   {
+                                       { 0 },
+                                   },
+#endif
+                                   {
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+#if CONFIG_ALT_INTRA
+                                       { 0 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                                   {
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+#if CONFIG_ALT_INTRA
+                                       { 0 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                                   {
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+#if CONFIG_ALT_INTRA
+                                       { 0 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                                   {
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+                                       { 0 },
+#if CONFIG_ALT_INTRA
+                                       { 0 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                               },
+                               {
+#if CONFIG_CB4X4
+                                   {
+                                       { 0 },
+                                   },
+#endif
+                                   {
+                                       { 8, 224, 32, 128, 64, 128 },
+                                       { 10, 32, 32, 128, 16, 192 },
+                                       { 10, 32, 32, 128, 16, 64 },
+                                       { 9, 200, 32, 128, 64, 128 },
+                                       { 8, 8, 32, 128, 224, 128 },
+                                       { 10, 32, 32, 128, 16, 192 },
+                                       { 10, 32, 32, 128, 16, 64 },
+                                       { 10, 23, 32, 128, 80, 176 },
+                                       { 10, 23, 32, 128, 80, 176 },
+                                       { 10, 32, 32, 128, 16, 64 },
+#if CONFIG_ALT_INTRA
+                                       { 10, 32, 32, 128, 16, 64 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                                   {
+                                       { 8, 224, 32, 128, 64, 128 },
+                                       { 10, 32, 32, 128, 16, 192 },
+                                       { 10, 32, 32, 128, 16, 64 },
+                                       { 9, 200, 32, 128, 64, 128 },
+                                       { 8, 8, 32, 128, 224, 128 },
+                                       { 10, 32, 32, 128, 16, 192 },
+                                       { 10, 32, 32, 128, 16, 64 },
+                                       { 10, 23, 32, 128, 80, 176 },
+                                       { 10, 23, 32, 128, 80, 176 },
+                                       { 10, 32, 32, 128, 16, 64 },
+#if CONFIG_ALT_INTRA
+                                       { 10, 32, 32, 128, 16, 64 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                                   {
+                                       { 8, 224, 32, 128, 64, 128 },
+                                       { 10, 32, 32, 128, 16, 192 },
+                                       { 10, 32, 32, 128, 16, 64 },
+                                       { 9, 200, 32, 128, 64, 128 },
+                                       { 8, 8, 32, 128, 224, 128 },
+                                       { 10, 32, 32, 128, 16, 192 },
+                                       { 10, 32, 32, 128, 16, 64 },
+                                       { 10, 23, 32, 128, 80, 176 },
+                                       { 10, 23, 32, 128, 80, 176 },
+                                       { 10, 32, 32, 128, 16, 64 },
+#if CONFIG_ALT_INTRA
+                                       { 10, 32, 32, 128, 16, 64 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                                   {
+                                       { 8, 224, 32, 128, 64, 128 },
+                                       { 10, 32, 32, 128, 16, 192 },
+                                       { 10, 32, 32, 128, 16, 64 },
+                                       { 9, 200, 32, 128, 64, 128 },
+                                       { 8, 8, 32, 128, 224, 128 },
+                                       { 10, 32, 32, 128, 16, 192 },
+                                       { 10, 32, 32, 128, 16, 64 },
+                                       { 10, 23, 32, 128, 80, 176 },
+                                       { 10, 23, 32, 128, 80, 176 },
+                                       { 10, 32, 32, 128, 16, 64 },
+#if CONFIG_ALT_INTRA
+                                       { 10, 32, 32, 128, 16, 64 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                               },
+                               {
+#if CONFIG_CB4X4
+                                   {
+                                       { 0 },
+                                   },
+#endif
+                                   {
+                                       { 8, 224, 64, 128 },
+                                       { 10, 32, 16, 192 },
+                                       { 10, 32, 16, 64 },
+                                       { 9, 200, 64, 128 },
+                                       { 8, 8, 224, 128 },
+                                       { 10, 32, 16, 192 },
+                                       { 10, 32, 16, 64 },
+                                       { 10, 23, 80, 176 },
+                                       { 10, 23, 80, 176 },
+                                       { 10, 32, 16, 64 },
+#if CONFIG_ALT_INTRA
+                                       { 10, 32, 16, 64 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                                   {
+                                       { 8, 224, 64, 128 },
+                                       { 10, 32, 16, 192 },
+                                       { 10, 32, 16, 64 },
+                                       { 9, 200, 64, 128 },
+                                       { 8, 8, 224, 128 },
+                                       { 10, 32, 16, 192 },
+                                       { 10, 32, 16, 64 },
+                                       { 10, 23, 80, 176 },
+                                       { 10, 23, 80, 176 },
+                                       { 10, 32, 16, 64 },
+#if CONFIG_ALT_INTRA
+                                       { 10, 32, 16, 64 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                                   {
+                                       { 8, 224, 64, 128 },
+                                       { 10, 32, 16, 192 },
+                                       { 10, 32, 16, 64 },
+                                       { 9, 200, 64, 128 },
+                                       { 8, 8, 224, 128 },
+                                       { 10, 32, 16, 192 },
+                                       { 10, 32, 16, 64 },
+                                       { 10, 23, 80, 176 },
+                                       { 10, 23, 80, 176 },
+                                       { 10, 32, 16, 64 },
+#if CONFIG_ALT_INTRA
+                                       { 10, 32, 16, 64 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                                   {
+                                       { 8, 224, 64, 128 },
+                                       { 10, 32, 16, 192 },
+                                       { 10, 32, 16, 64 },
+                                       { 9, 200, 64, 128 },
+                                       { 8, 8, 224, 128 },
+                                       { 10, 32, 16, 192 },
+                                       { 10, 32, 16, 64 },
+                                       { 10, 23, 80, 176 },
+                                       { 10, 23, 80, 176 },
+                                       { 10, 32, 16, 64 },
+#if CONFIG_ALT_INTRA
+                                       { 10, 32, 16, 64 },
+#endif  // CONFIG_ALT_INTRA
+                                   },
+                               },
+                             };
+#else  // !CONFIG_EXT_TX
+
+/* clang-format off */
+const aom_tree_index av1_ext_tx_tree[TREE_SIZE(TX_TYPES)] = {
+  -DCT_DCT, 2,
+  -ADST_ADST, 4,
+  -ADST_DCT, -DCT_ADST
+};
+/* clang-format on */
+
+int av1_ext_tx_ind[TX_TYPES];
+int av1_ext_tx_inv[TX_TYPES];
+
+static const aom_prob
+    default_intra_ext_tx_prob[EXT_TX_SIZES][TX_TYPES][TX_TYPES - 1] = {
+#if CONFIG_CB4X4
+      { { 240, 85, 128 }, { 4, 1, 248 }, { 4, 1, 8 }, { 4, 248, 128 } },
+#endif
+      { { 240, 85, 128 }, { 4, 1, 248 }, { 4, 1, 8 }, { 4, 248, 128 } },
+      { { 244, 85, 128 }, { 8, 2, 248 }, { 8, 2, 8 }, { 8, 248, 128 } },
+      { { 248, 85, 128 }, { 16, 4, 248 }, { 16, 4, 8 }, { 16, 248, 128 } },
+    };
+
+static const aom_prob default_inter_ext_tx_prob[EXT_TX_SIZES][TX_TYPES - 1] = {
+#if CONFIG_CB4X4
+  { 160, 85, 128 },
+#endif
+  { 160, 85, 128 },
+  { 176, 85, 128 },
+  { 192, 85, 128 },
+};
+#endif  // CONFIG_EXT_TX
+
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+static const aom_prob
+    default_intra_filter_probs[INTRA_FILTERS + 1][INTRA_FILTERS - 1] = {
+      { 98, 63, 60 }, { 98, 82, 80 }, { 94, 65, 103 },
+      { 49, 25, 24 }, { 72, 38, 50 },
+    };
+const aom_tree_index av1_intra_filter_tree[TREE_SIZE(INTRA_FILTERS)] = {
+  -INTRA_FILTER_LINEAR,      2, -INTRA_FILTER_8TAP, 4, -INTRA_FILTER_8TAP_SHARP,
+  -INTRA_FILTER_8TAP_SMOOTH,
+};
+#if CONFIG_EC_MULTISYMBOL
+int av1_intra_filter_ind[INTRA_FILTERS];
+int av1_intra_filter_inv[INTRA_FILTERS];
+#endif  // CONFIG_EC_MULTISYMBOL
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+
+#if CONFIG_FILTER_INTRA
+static const aom_prob default_filter_intra_probs[2] = { 230, 230 };
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_SUPERTX
+static const aom_prob
+    default_supertx_prob[PARTITION_SUPERTX_CONTEXTS][TX_SIZES] = {
+#if CONFIG_CB4X4
+#if CONFIG_TX64X64
+      { 1, 1, 160, 160, 170, 180 }, { 1, 1, 200, 200, 210, 220 },
+#else
+      { 1, 1, 160, 160, 170 }, { 1, 1, 200, 200, 210 },
+#endif  // CONFIG_TX64X64
+#else
+#if CONFIG_TX64X64
+      { 1, 160, 160, 170, 180 }, { 1, 200, 200, 210, 220 },
+#else
+      { 1, 160, 160, 170 }, { 1, 200, 200, 210 },
+#endif  // CONFIG_CB4X4
+#endif  // CONFIG_TX64X64
+    };
+#endif  // CONFIG_SUPERTX
+
+// FIXME(someone) need real defaults here
+static const aom_prob default_segment_tree_probs[SEG_TREE_PROBS] = {
+  128, 128, 128, 128, 128, 128, 128
+};
+// clang-format off
+static const aom_prob default_segment_pred_probs[PREDICTION_PROBS] = {
+  128, 128, 128
+};
+// clang-format on
+
+#if CONFIG_EC_MULTISYMBOL
+#if CONFIG_DUAL_FILTER
+static const aom_cdf_prob
+    default_switchable_interp_cdf[SWITCHABLE_FILTER_CONTEXTS][CDF_SIZE(
+        SWITCHABLE_FILTERS)] = {
+      { AOM_ICDF(30080), AOM_ICDF(31088), AOM_ICDF(32096), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4608), AOM_ICDF(9620), AOM_ICDF(31338), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4352), AOM_ICDF(5240), AOM_ICDF(6128), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4352), AOM_ICDF(5240), AOM_ICDF(6128), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(19072), AOM_ICDF(23352), AOM_ICDF(27632), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(30080), AOM_ICDF(31088), AOM_ICDF(32096), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4608), AOM_ICDF(9620), AOM_ICDF(31338), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4352), AOM_ICDF(5240), AOM_ICDF(6128), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4352), AOM_ICDF(5240), AOM_ICDF(6128), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(19072), AOM_ICDF(23352), AOM_ICDF(27632), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(30080), AOM_ICDF(31088), AOM_ICDF(32096), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4608), AOM_ICDF(9620), AOM_ICDF(31338), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4352), AOM_ICDF(5240), AOM_ICDF(6128), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4352), AOM_ICDF(5240), AOM_ICDF(6128), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(19072), AOM_ICDF(23352), AOM_ICDF(27632), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(30080), AOM_ICDF(31088), AOM_ICDF(32096), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4608), AOM_ICDF(9620), AOM_ICDF(31338), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4352), AOM_ICDF(5240), AOM_ICDF(6128), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4352), AOM_ICDF(5240), AOM_ICDF(6128), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(19072), AOM_ICDF(23352), AOM_ICDF(27632), AOM_ICDF(32768), 0 }
+    };
+#else
+static const aom_cdf_prob
+    default_switchable_interp_cdf[SWITCHABLE_FILTER_CONTEXTS][CDF_SIZE(
+        SWITCHABLE_FILTERS)] = {
+      { AOM_ICDF(30080), AOM_ICDF(31781), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4608), AOM_ICDF(32658), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(4352), AOM_ICDF(4685), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(19072), AOM_ICDF(26776), AOM_ICDF(32768), 0 },
+    };
+#endif
+
+static const aom_cdf_prob default_seg_tree_cdf[CDF_SIZE(MAX_SEGMENTS)] = {
+  AOM_ICDF(4096),  AOM_ICDF(8192),  AOM_ICDF(12288),
+  AOM_ICDF(16384), AOM_ICDF(20480), AOM_ICDF(24576),
+  AOM_ICDF(28672), AOM_ICDF(32768), 0
+};
+
+static const aom_cdf_prob
+    default_tx_size_cdf[MAX_TX_DEPTH][TX_SIZE_CONTEXTS][CDF_SIZE(MAX_TX_DEPTH +
+                                                                 1)] = {
+      { { AOM_ICDF(12800), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(8448), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(2560), AOM_ICDF(20496), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(1920), AOM_ICDF(14091), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(384), AOM_ICDF(17588), AOM_ICDF(19782), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(640), AOM_ICDF(7166), AOM_ICDF(8466), AOM_ICDF(32768), 0 } },
+#if CONFIG_TX64X64
+      { { AOM_ICDF(128), AOM_ICDF(8288), AOM_ICDF(21293), AOM_ICDF(26986),
+          AOM_ICDF(32768), 0 },
+        { AOM_ICDF(128), AOM_ICDF(4208), AOM_ICDF(10009), AOM_ICDF(15965),
+          AOM_ICDF(32768), 0 } },
+#endif
+    };
+
+#if CONFIG_ALT_INTRA
+static const aom_cdf_prob
+    default_if_y_mode_cdf[BLOCK_SIZE_GROUPS][CDF_SIZE(INTRA_MODES)] = {
+      { AOM_ICDF(11264), AOM_ICDF(12608), AOM_ICDF(16309), AOM_ICDF(21086),
+        AOM_ICDF(23297), AOM_ICDF(24860), AOM_ICDF(27022), AOM_ICDF(28099),
+        AOM_ICDF(29631), AOM_ICDF(31126), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(9600), AOM_ICDF(11953), AOM_ICDF(16100), AOM_ICDF(20922),
+        AOM_ICDF(22756), AOM_ICDF(23913), AOM_ICDF(25435), AOM_ICDF(26724),
+        AOM_ICDF(28046), AOM_ICDF(29927), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(9344), AOM_ICDF(11540), AOM_ICDF(16515), AOM_ICDF(21763),
+        AOM_ICDF(23078), AOM_ICDF(23816), AOM_ICDF(24725), AOM_ICDF(25856),
+        AOM_ICDF(26720), AOM_ICDF(28208), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(12288), AOM_ICDF(14448), AOM_ICDF(18026), AOM_ICDF(23346),
+        AOM_ICDF(23833), AOM_ICDF(24188), AOM_ICDF(24724), AOM_ICDF(25415),
+        AOM_ICDF(25817), AOM_ICDF(26876), AOM_ICDF(32768), 0 },
+    };
+
+static const aom_cdf_prob
+    default_uv_mode_cdf[INTRA_MODES][CDF_SIZE(INTRA_MODES)] = {
+      { AOM_ICDF(25472), AOM_ICDF(25558), AOM_ICDF(27783), AOM_ICDF(30779),
+        AOM_ICDF(30988), AOM_ICDF(31269), AOM_ICDF(31492), AOM_ICDF(31741),
+        AOM_ICDF(32014), AOM_ICDF(32420), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(2176), AOM_ICDF(2415), AOM_ICDF(28381), AOM_ICDF(29574),
+        AOM_ICDF(29832), AOM_ICDF(30712), AOM_ICDF(30881), AOM_ICDF(31662),
+        AOM_ICDF(31761), AOM_ICDF(31922), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(3328), AOM_ICDF(3443), AOM_ICDF(4016), AOM_ICDF(31099),
+        AOM_ICDF(31272), AOM_ICDF(31420), AOM_ICDF(31504), AOM_ICDF(31608),
+        AOM_ICDF(31916), AOM_ICDF(32598), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(23424), AOM_ICDF(23534), AOM_ICDF(25915), AOM_ICDF(27831),
+        AOM_ICDF(28058), AOM_ICDF(28431), AOM_ICDF(30142), AOM_ICDF(31209),
+        AOM_ICDF(31459), AOM_ICDF(32369), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(22784), AOM_ICDF(22862), AOM_ICDF(24255), AOM_ICDF(26287),
+        AOM_ICDF(28490), AOM_ICDF(29509), AOM_ICDF(29776), AOM_ICDF(30115),
+        AOM_ICDF(31203), AOM_ICDF(31674), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(19712), AOM_ICDF(19865), AOM_ICDF(23141), AOM_ICDF(24428),
+        AOM_ICDF(25731), AOM_ICDF(31377), AOM_ICDF(31622), AOM_ICDF(32047),
+        AOM_ICDF(32458), AOM_ICDF(32767), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(21376), AOM_ICDF(21421), AOM_ICDF(22130), AOM_ICDF(27688),
+        AOM_ICDF(28485), AOM_ICDF(28779), AOM_ICDF(28935), AOM_ICDF(29085),
+        AOM_ICDF(31962), AOM_ICDF(32450), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(19712), AOM_ICDF(19814), AOM_ICDF(20725), AOM_ICDF(28510),
+        AOM_ICDF(28814), AOM_ICDF(29099), AOM_ICDF(29457), AOM_ICDF(29729),
+        AOM_ICDF(30133), AOM_ICDF(32408), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(19584), AOM_ICDF(19790), AOM_ICDF(23643), AOM_ICDF(25501),
+        AOM_ICDF(25913), AOM_ICDF(26673), AOM_ICDF(27578), AOM_ICDF(30923),
+        AOM_ICDF(31255), AOM_ICDF(31870), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(20864), AOM_ICDF(21004), AOM_ICDF(24129), AOM_ICDF(26308),
+        AOM_ICDF(27062), AOM_ICDF(27065), AOM_ICDF(27488), AOM_ICDF(28045),
+        AOM_ICDF(28506), AOM_ICDF(29272), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(23680), AOM_ICDF(23929), AOM_ICDF(27831), AOM_ICDF(30446),
+        AOM_ICDF(30598), AOM_ICDF(31129), AOM_ICDF(31244), AOM_ICDF(31655),
+        AOM_ICDF(31868), AOM_ICDF(32234), AOM_ICDF(32768), 0 },
+    };
+#else   // !CONFIG_ALT_INTRA
+static const aom_cdf_prob
+    default_if_y_mode_cdf[BLOCK_SIZE_GROUPS][CDF_SIZE(INTRA_MODES)] = {
+      { AOM_ICDF(8320), AOM_ICDF(11376), AOM_ICDF(12880), AOM_ICDF(19959),
+        AOM_ICDF(23072), AOM_ICDF(24067), AOM_ICDF(25461), AOM_ICDF(26917),
+        AOM_ICDF(29157), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(16896), AOM_ICDF(21112), AOM_ICDF(21932), AOM_ICDF(27852),
+        AOM_ICDF(28667), AOM_ICDF(28916), AOM_ICDF(29593), AOM_ICDF(30089),
+        AOM_ICDF(30905), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(22144), AOM_ICDF(25464), AOM_ICDF(26006), AOM_ICDF(30364),
+        AOM_ICDF(30583), AOM_ICDF(30655), AOM_ICDF(31183), AOM_ICDF(31400),
+        AOM_ICDF(31646), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(28288), AOM_ICDF(30650), AOM_ICDF(30964), AOM_ICDF(32288),
+        AOM_ICDF(32308), AOM_ICDF(32331), AOM_ICDF(32495), AOM_ICDF(32586),
+        AOM_ICDF(32607), AOM_ICDF(32768), 0 },
+    };
+
+static const aom_cdf_prob
+    default_uv_mode_cdf[INTRA_MODES][CDF_SIZE(INTRA_MODES)] = {
+      { AOM_ICDF(15360), AOM_ICDF(15836), AOM_ICDF(20863), AOM_ICDF(27513),
+        AOM_ICDF(28269), AOM_ICDF(29048), AOM_ICDF(29455), AOM_ICDF(30154),
+        AOM_ICDF(31206), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(6144), AOM_ICDF(7392), AOM_ICDF(22657), AOM_ICDF(25981),
+        AOM_ICDF(26965), AOM_ICDF(28779), AOM_ICDF(29309), AOM_ICDF(30890),
+        AOM_ICDF(31763), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(8576), AOM_ICDF(9143), AOM_ICDF(11450), AOM_ICDF(27575),
+        AOM_ICDF(28108), AOM_ICDF(28438), AOM_ICDF(28658), AOM_ICDF(28995),
+        AOM_ICDF(30410), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(12416), AOM_ICDF(12814), AOM_ICDF(16244), AOM_ICDF(22057),
+        AOM_ICDF(23492), AOM_ICDF(24700), AOM_ICDF(26213), AOM_ICDF(27954),
+        AOM_ICDF(29778), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(10624), AOM_ICDF(11057), AOM_ICDF(14619), AOM_ICDF(19415),
+        AOM_ICDF(23134), AOM_ICDF(25679), AOM_ICDF(26399), AOM_ICDF(27618),
+        AOM_ICDF(30676), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(10240), AOM_ICDF(10680), AOM_ICDF(15684), AOM_ICDF(19118),
+        AOM_ICDF(21856), AOM_ICDF(27563), AOM_ICDF(28234), AOM_ICDF(29332),
+        AOM_ICDF(31278), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(11008), AOM_ICDF(11433), AOM_ICDF(14100), AOM_ICDF(22522),
+        AOM_ICDF(24365), AOM_ICDF(25330), AOM_ICDF(25737), AOM_ICDF(26341),
+        AOM_ICDF(30433), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(10880), AOM_ICDF(11308), AOM_ICDF(13991), AOM_ICDF(23645),
+        AOM_ICDF(24679), AOM_ICDF(25433), AOM_ICDF(25977), AOM_ICDF(26746),
+        AOM_ICDF(28463), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(9856), AOM_ICDF(10483), AOM_ICDF(16054), AOM_ICDF(19959),
+        AOM_ICDF(21708), AOM_ICDF(23628), AOM_ICDF(24949), AOM_ICDF(28797),
+        AOM_ICDF(30658), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(12928), AOM_ICDF(14556), AOM_ICDF(22168), AOM_ICDF(27789),
+        AOM_ICDF(28543), AOM_ICDF(29663), AOM_ICDF(29893), AOM_ICDF(30645),
+        AOM_ICDF(31682), AOM_ICDF(32768), 0 },
+    };
+#endif  // CONFIG_ALT_INTRA
+
+#if CONFIG_EXT_PARTITION_TYPES
+static const aom_cdf_prob
+    default_partition_cdf[PARTITION_CONTEXTS][CDF_SIZE(EXT_PARTITION_TYPES)] = {
+      // 8x8 -> 4x4 only supports the four legacy partition types
+      { AOM_ICDF(25472), AOM_ICDF(28949), AOM_ICDF(31052), AOM_ICDF(32768), 0,
+        0, 0, 0, 0 },
+      { AOM_ICDF(18816), AOM_ICDF(22250), AOM_ICDF(28783), AOM_ICDF(32768), 0,
+        0, 0, 0, 0 },
+      { AOM_ICDF(18944), AOM_ICDF(26126), AOM_ICDF(29188), AOM_ICDF(32768), 0,
+        0, 0, 0, 0 },
+      { AOM_ICDF(15488), AOM_ICDF(22508), AOM_ICDF(27077), AOM_ICDF(32768), 0,
+        0, 0, 0, 0 },
+      { AOM_ICDF(22272), AOM_ICDF(23768), AOM_ICDF(25043), AOM_ICDF(29996),
+        AOM_ICDF(30744), AOM_ICDF(31493), AOM_ICDF(32130), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(11776), AOM_ICDF(13457), AOM_ICDF(16315), AOM_ICDF(28229),
+        AOM_ICDF(29069), AOM_ICDF(29910), AOM_ICDF(31339), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(10496), AOM_ICDF(14802), AOM_ICDF(16136), AOM_ICDF(27127),
+        AOM_ICDF(29280), AOM_ICDF(31434), AOM_ICDF(32101), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(6784), AOM_ICDF(8763), AOM_ICDF(10440), AOM_ICDF(29110),
+        AOM_ICDF(30100), AOM_ICDF(31090), AOM_ICDF(31929), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(22656), AOM_ICDF(23801), AOM_ICDF(24702), AOM_ICDF(30721),
+        AOM_ICDF(31294), AOM_ICDF(31867), AOM_ICDF(32317), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(8704), AOM_ICDF(9926), AOM_ICDF(12586), AOM_ICDF(28885),
+        AOM_ICDF(29496), AOM_ICDF(30107), AOM_ICDF(31437), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(6656), AOM_ICDF(10685), AOM_ICDF(11566), AOM_ICDF(27857),
+        AOM_ICDF(29871), AOM_ICDF(31886), AOM_ICDF(32327), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(2176), AOM_ICDF(3012), AOM_ICDF(3690), AOM_ICDF(31253),
+        AOM_ICDF(31671), AOM_ICDF(32090), AOM_ICDF(32429), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(28416), AOM_ICDF(28705), AOM_ICDF(28926), AOM_ICDF(32258),
+        AOM_ICDF(32402), AOM_ICDF(32547), AOM_ICDF(32657), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(9216), AOM_ICDF(9952), AOM_ICDF(11849), AOM_ICDF(30134),
+        AOM_ICDF(30502), AOM_ICDF(30870), AOM_ICDF(31819), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(7424), AOM_ICDF(9008), AOM_ICDF(9528), AOM_ICDF(30664),
+        AOM_ICDF(31456), AOM_ICDF(32248), AOM_ICDF(32508), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(1280), AOM_ICDF(1710), AOM_ICDF(2069), AOM_ICDF(31978),
+        AOM_ICDF(32193), AOM_ICDF(32409), AOM_ICDF(32588), AOM_ICDF(32768), 0 },
+#if CONFIG_EXT_PARTITION
+      { AOM_ICDF(28416), AOM_ICDF(28705), AOM_ICDF(28926), AOM_ICDF(32258),
+        AOM_ICDF(32402), AOM_ICDF(32547), AOM_ICDF(32657), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(9216), AOM_ICDF(9952), AOM_ICDF(11849), AOM_ICDF(30134),
+        AOM_ICDF(30502), AOM_ICDF(30870), AOM_ICDF(31819), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(7424), AOM_ICDF(9008), AOM_ICDF(9528), AOM_ICDF(30664),
+        AOM_ICDF(31456), AOM_ICDF(32248), AOM_ICDF(32508), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(1280), AOM_ICDF(1710), AOM_ICDF(2069), AOM_ICDF(31978),
+        AOM_ICDF(32193), AOM_ICDF(32409), AOM_ICDF(32588), AOM_ICDF(32768), 0 },
+#endif
+    };
+#else
+static const aom_cdf_prob
+    default_partition_cdf[PARTITION_CONTEXTS][CDF_SIZE(PARTITION_TYPES)] = {
+      { AOM_ICDF(25472), AOM_ICDF(28949), AOM_ICDF(31052), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(18816), AOM_ICDF(22250), AOM_ICDF(28783), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(18944), AOM_ICDF(26126), AOM_ICDF(29188), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(15488), AOM_ICDF(22508), AOM_ICDF(27077), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(22272), AOM_ICDF(25265), AOM_ICDF(27815), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(11776), AOM_ICDF(15138), AOM_ICDF(20854), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(10496), AOM_ICDF(19109), AOM_ICDF(21777), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(6784), AOM_ICDF(10743), AOM_ICDF(14098), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(22656), AOM_ICDF(24947), AOM_ICDF(26749), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(8704), AOM_ICDF(11148), AOM_ICDF(16469), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(6656), AOM_ICDF(14714), AOM_ICDF(16477), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(2176), AOM_ICDF(3849), AOM_ICDF(5205), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(28416), AOM_ICDF(28994), AOM_ICDF(29436), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(9216), AOM_ICDF(10688), AOM_ICDF(14483), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(7424), AOM_ICDF(10592), AOM_ICDF(11632), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(1280), AOM_ICDF(2141), AOM_ICDF(2859), AOM_ICDF(32768), 0 },
+#if CONFIG_EXT_PARTITION
+      { AOM_ICDF(28416), AOM_ICDF(28994), AOM_ICDF(29436), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(9216), AOM_ICDF(10688), AOM_ICDF(14483), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(7424), AOM_ICDF(10592), AOM_ICDF(11632), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(1280), AOM_ICDF(2141), AOM_ICDF(2859), AOM_ICDF(32768), 0 },
+#endif
+    };
+#endif
+
+static const aom_cdf_prob
+    default_inter_mode_cdf[INTER_MODE_CONTEXTS][CDF_SIZE(INTER_MODES)] = {
+      { AOM_ICDF(256), AOM_ICDF(22227), AOM_ICDF(23627), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(896), AOM_ICDF(18948), AOM_ICDF(23537), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(896), AOM_ICDF(21563), AOM_ICDF(24320), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(896), AOM_ICDF(12599), AOM_ICDF(17799), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(1024), AOM_ICDF(8960), AOM_ICDF(13238), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(2176), AOM_ICDF(11856), AOM_ICDF(14388), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(3200), AOM_ICDF(6550), AOM_ICDF(9622), AOM_ICDF(32768), 0 },
+    };
+
+#if CONFIG_EXT_TX
+static const aom_cdf_prob default_intra_ext_tx_cdf
+    [EXT_TX_SETS_INTRA][EXT_TX_SIZES][INTRA_MODES][CDF_SIZE(TX_TYPES)] = {
+      {
+// FIXME: unused zero positions, from uncoded trivial transform set
+#if CONFIG_CB4X4
+          {
+              { 0 },
+          },
+#endif
+          { { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+#if CONFIG_ALT_INTRA
+            { 0 }
+#endif
+          },
+          { { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+#if CONFIG_ALT_INTRA
+            { 0 }
+#endif
+          },
+          { { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+#if CONFIG_ALT_INTRA
+            { 0 }
+#endif
+          },
+          { { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+            { 0 },
+#if CONFIG_ALT_INTRA
+            { 0 }
+#endif
+          },
+      },
+      {
+          { { AOM_ICDF(1024), AOM_ICDF(28800), AOM_ICDF(29048), AOM_ICDF(29296),
+              AOM_ICDF(30164), AOM_ICDF(31466), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(27118), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1152), AOM_ICDF(25852), AOM_ICDF(26284), AOM_ICDF(26717),
+              AOM_ICDF(28230), AOM_ICDF(30499), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1024), AOM_ICDF(2016), AOM_ICDF(3938), AOM_ICDF(5860),
+              AOM_ICDF(29404), AOM_ICDF(31086), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(27118), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(5900), AOM_ICDF(7691),
+              AOM_ICDF(15528), AOM_ICDF(27380), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(5900), AOM_ICDF(7691),
+              AOM_ICDF(15528), AOM_ICDF(27380), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+#if CONFIG_ALT_INTRA
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 }
+#endif
+          },
+          { { AOM_ICDF(1024), AOM_ICDF(28800), AOM_ICDF(29048), AOM_ICDF(29296),
+              AOM_ICDF(30164), AOM_ICDF(31466), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(27118), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1152), AOM_ICDF(25852), AOM_ICDF(26284), AOM_ICDF(26717),
+              AOM_ICDF(28230), AOM_ICDF(30499), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1024), AOM_ICDF(2016), AOM_ICDF(3938), AOM_ICDF(5860),
+              AOM_ICDF(29404), AOM_ICDF(31086), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(27118), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(5900), AOM_ICDF(7691),
+              AOM_ICDF(15528), AOM_ICDF(27380), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(5900), AOM_ICDF(7691),
+              AOM_ICDF(15528), AOM_ICDF(27380), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+#if CONFIG_ALT_INTRA
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 }
+#endif
+          },
+          { { AOM_ICDF(1024), AOM_ICDF(28800), AOM_ICDF(29048), AOM_ICDF(29296),
+              AOM_ICDF(30164), AOM_ICDF(31466), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(27118), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1152), AOM_ICDF(25852), AOM_ICDF(26284), AOM_ICDF(26717),
+              AOM_ICDF(28230), AOM_ICDF(30499), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1024), AOM_ICDF(2016), AOM_ICDF(3938), AOM_ICDF(5860),
+              AOM_ICDF(29404), AOM_ICDF(31086), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(27118), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(5900), AOM_ICDF(7691),
+              AOM_ICDF(15528), AOM_ICDF(27380), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(5900), AOM_ICDF(7691),
+              AOM_ICDF(15528), AOM_ICDF(27380), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+#if CONFIG_ALT_INTRA
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 }
+#endif
+          },
+          { { AOM_ICDF(1024), AOM_ICDF(28800), AOM_ICDF(29048), AOM_ICDF(29296),
+              AOM_ICDF(30164), AOM_ICDF(31466), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(27118), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1152), AOM_ICDF(25852), AOM_ICDF(26284), AOM_ICDF(26717),
+              AOM_ICDF(28230), AOM_ICDF(30499), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1024), AOM_ICDF(2016), AOM_ICDF(3938), AOM_ICDF(5860),
+              AOM_ICDF(29404), AOM_ICDF(31086), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(27118), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(5900), AOM_ICDF(7691),
+              AOM_ICDF(15528), AOM_ICDF(27380), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(5900), AOM_ICDF(7691),
+              AOM_ICDF(15528), AOM_ICDF(27380), AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 },
+#if CONFIG_ALT_INTRA
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(8660),
+              AOM_ICDF(10167), AOM_ICDF(15817), AOM_ICDF(32768), 0 }
+#endif
+          },
+      },
+      {
+          { { AOM_ICDF(1024), AOM_ICDF(28800), AOM_ICDF(29792), AOM_ICDF(31280),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(26310),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1152), AOM_ICDF(25852), AOM_ICDF(27581), AOM_ICDF(30174),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1024), AOM_ICDF(2016), AOM_ICDF(28924), AOM_ICDF(30846),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(26310),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(13065), AOM_ICDF(26611),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(13065), AOM_ICDF(26611),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+#if CONFIG_ALT_INTRA
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 }
+#endif
+          },
+          { { AOM_ICDF(1024), AOM_ICDF(28800), AOM_ICDF(29792), AOM_ICDF(31280),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(26310),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1152), AOM_ICDF(25852), AOM_ICDF(27581), AOM_ICDF(30174),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1024), AOM_ICDF(2016), AOM_ICDF(28924), AOM_ICDF(30846),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(26310),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(13065), AOM_ICDF(26611),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(13065), AOM_ICDF(26611),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+#if CONFIG_ALT_INTRA
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 }
+#endif
+          },
+          { { AOM_ICDF(1024), AOM_ICDF(28800), AOM_ICDF(29792), AOM_ICDF(31280),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(26310),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1152), AOM_ICDF(25852), AOM_ICDF(27581), AOM_ICDF(30174),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1024), AOM_ICDF(2016), AOM_ICDF(28924), AOM_ICDF(30846),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(26310),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(13065), AOM_ICDF(26611),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(13065), AOM_ICDF(26611),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+#if CONFIG_ALT_INTRA
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 }
+#endif
+          },
+          { { AOM_ICDF(1024), AOM_ICDF(28800), AOM_ICDF(29792), AOM_ICDF(31280),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(26310),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1152), AOM_ICDF(25852), AOM_ICDF(27581), AOM_ICDF(30174),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1024), AOM_ICDF(2016), AOM_ICDF(28924), AOM_ICDF(30846),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(26310),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(13065), AOM_ICDF(26611),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(4109), AOM_ICDF(13065), AOM_ICDF(26611),
+              AOM_ICDF(32768), 0 },
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 },
+#if CONFIG_ALT_INTRA
+            { AOM_ICDF(1280), AOM_ICDF(5216), AOM_ICDF(6938), AOM_ICDF(13396),
+              AOM_ICDF(32768), 0 }
+#endif
+          },
+      }
+    };
+static const aom_cdf_prob
+    default_inter_ext_tx_cdf[EXT_TX_SETS_INTER][EXT_TX_SIZES][CDF_SIZE(
+        TX_TYPES)] = {
+      {
+#if CONFIG_CB4X4
+          { 0 },
+#endif
+          { 0 },
+          { 0 },
+          { 0 },
+          { 0 } },
+      {
+#if CONFIG_CB4X4
+          { 0 },
+#endif
+          { AOM_ICDF(1280), AOM_ICDF(1453), AOM_ICDF(1626), AOM_ICDF(2277),
+            AOM_ICDF(2929), AOM_ICDF(3580), AOM_ICDF(4232), AOM_ICDF(16717),
+            AOM_ICDF(19225), AOM_ICDF(21733), AOM_ICDF(24241), AOM_ICDF(26749),
+            AOM_ICDF(28253), AOM_ICDF(29758), AOM_ICDF(31263), AOM_ICDF(32768),
+            0 },
+          { AOM_ICDF(1280), AOM_ICDF(1453), AOM_ICDF(1626), AOM_ICDF(2277),
+            AOM_ICDF(2929), AOM_ICDF(3580), AOM_ICDF(4232), AOM_ICDF(16717),
+            AOM_ICDF(19225), AOM_ICDF(21733), AOM_ICDF(24241), AOM_ICDF(26749),
+            AOM_ICDF(28253), AOM_ICDF(29758), AOM_ICDF(31263), AOM_ICDF(32768),
+            0 },
+          { AOM_ICDF(1280), AOM_ICDF(1453), AOM_ICDF(1626), AOM_ICDF(2277),
+            AOM_ICDF(2929), AOM_ICDF(3580), AOM_ICDF(4232), AOM_ICDF(16717),
+            AOM_ICDF(19225), AOM_ICDF(21733), AOM_ICDF(24241), AOM_ICDF(26749),
+            AOM_ICDF(28253), AOM_ICDF(29758), AOM_ICDF(31263), AOM_ICDF(32768),
+            0 },
+          { AOM_ICDF(1280), AOM_ICDF(1453), AOM_ICDF(1626), AOM_ICDF(2277),
+            AOM_ICDF(2929), AOM_ICDF(3580), AOM_ICDF(4232), AOM_ICDF(16717),
+            AOM_ICDF(19225), AOM_ICDF(21733), AOM_ICDF(24241), AOM_ICDF(26749),
+            AOM_ICDF(28253), AOM_ICDF(29758), AOM_ICDF(31263), AOM_ICDF(32768),
+            0 } },
+      {
+#if CONFIG_CB4X4
+          { 0 },
+#endif
+          { AOM_ICDF(1280), AOM_ICDF(3125), AOM_ICDF(4970), AOM_ICDF(17132),
+            AOM_ICDF(19575), AOM_ICDF(22018), AOM_ICDF(24461), AOM_ICDF(26904),
+            AOM_ICDF(28370), AOM_ICDF(29836), AOM_ICDF(31302), AOM_ICDF(32768),
+            0 },
+          { AOM_ICDF(1280), AOM_ICDF(3125), AOM_ICDF(4970), AOM_ICDF(17132),
+            AOM_ICDF(19575), AOM_ICDF(22018), AOM_ICDF(24461), AOM_ICDF(26904),
+            AOM_ICDF(28370), AOM_ICDF(29836), AOM_ICDF(31302), AOM_ICDF(32768),
+            0 },
+          { AOM_ICDF(1280), AOM_ICDF(3125), AOM_ICDF(4970), AOM_ICDF(17132),
+            AOM_ICDF(19575), AOM_ICDF(22018), AOM_ICDF(24461), AOM_ICDF(26904),
+            AOM_ICDF(28370), AOM_ICDF(29836), AOM_ICDF(31302), AOM_ICDF(32768),
+            0 },
+          { AOM_ICDF(1280), AOM_ICDF(3125), AOM_ICDF(4970), AOM_ICDF(17132),
+            AOM_ICDF(19575), AOM_ICDF(22018), AOM_ICDF(24461), AOM_ICDF(26904),
+            AOM_ICDF(28370), AOM_ICDF(29836), AOM_ICDF(31302), AOM_ICDF(32768),
+            0 } },
+      {
+#if CONFIG_CB4X4
+          { 0 },
+#endif
+          { AOM_ICDF(1536), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(1536), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(1536), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(1536), AOM_ICDF(32768), 0 } }
+    };
+#else
+static const aom_cdf_prob
+    default_intra_ext_tx_cdf[EXT_TX_SIZES][TX_TYPES][CDF_SIZE(TX_TYPES)] = {
+#if CONFIG_CB4X4
+      { { AOM_ICDF(30720), AOM_ICDF(31400), AOM_ICDF(32084), AOM_ICDF(32768),
+          0 },
+        { AOM_ICDF(512), AOM_ICDF(638), AOM_ICDF(31764), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(512), AOM_ICDF(638), AOM_ICDF(1642), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(512), AOM_ICDF(31760), AOM_ICDF(32264), AOM_ICDF(32768),
+          0 } },
+#endif
+      { { AOM_ICDF(30720), AOM_ICDF(31400), AOM_ICDF(32084), AOM_ICDF(32768),
+          0 },
+        { AOM_ICDF(512), AOM_ICDF(638), AOM_ICDF(31764), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(512), AOM_ICDF(638), AOM_ICDF(1642), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(512), AOM_ICDF(31760), AOM_ICDF(32264), AOM_ICDF(32768),
+          0 } },
+
+      { { AOM_ICDF(31232), AOM_ICDF(31742), AOM_ICDF(32255), AOM_ICDF(32768),
+          0 },
+        { AOM_ICDF(1024), AOM_ICDF(1272), AOM_ICDF(31784), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(1024), AOM_ICDF(1272), AOM_ICDF(2256), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(1024), AOM_ICDF(31776), AOM_ICDF(32272), AOM_ICDF(32768),
+          0 } },
+      { { AOM_ICDF(31744), AOM_ICDF(32084), AOM_ICDF(32426), AOM_ICDF(32768),
+          0 },
+        { AOM_ICDF(2048), AOM_ICDF(2528), AOM_ICDF(31823), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(2048), AOM_ICDF(2528), AOM_ICDF(3473), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(2048), AOM_ICDF(31808), AOM_ICDF(32288), AOM_ICDF(32768),
+          0 } },
+    };
+
+static const aom_cdf_prob
+    default_inter_ext_tx_cdf[EXT_TX_SIZES][CDF_SIZE(TX_TYPES)] = {
+#if CONFIG_CB4X4
+      { AOM_ICDF(20480), AOM_ICDF(24560), AOM_ICDF(28664), AOM_ICDF(32768), 0 },
+#endif
+      { AOM_ICDF(20480), AOM_ICDF(24560), AOM_ICDF(28664), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(22528), AOM_ICDF(25928), AOM_ICDF(29348), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(24576), AOM_ICDF(27296), AOM_ICDF(30032), AOM_ICDF(32768), 0 },
+    };
+#endif  // !CONFIG_EXT_TX
+
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+static const aom_cdf_prob
+    default_intra_filter_cdf[INTRA_FILTERS + 1][CDF_SIZE(INTRA_FILTERS)] = {
+      { AOM_ICDF(12544), AOM_ICDF(17521), AOM_ICDF(21095), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(12544), AOM_ICDF(19022), AOM_ICDF(23318), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(12032), AOM_ICDF(17297), AOM_ICDF(23522), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(6272), AOM_ICDF(8860), AOM_ICDF(11101), AOM_ICDF(32768), 0 },
+      { AOM_ICDF(9216), AOM_ICDF(12712), AOM_ICDF(16629), AOM_ICDF(32768), 0 },
+    };
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+
+// CDF version of 'av1_kf_y_mode_prob'.
+const aom_cdf_prob
+    av1_kf_y_mode_cdf[INTRA_MODES][INTRA_MODES][CDF_SIZE(INTRA_MODES)] = {
+#if CONFIG_ALT_INTRA
+      {
+          { AOM_ICDF(15488), AOM_ICDF(17513), AOM_ICDF(20731), AOM_ICDF(24586),
+            AOM_ICDF(25921), AOM_ICDF(26749), AOM_ICDF(27807), AOM_ICDF(28602),
+            AOM_ICDF(29530), AOM_ICDF(30681), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(11648), AOM_ICDF(14783), AOM_ICDF(21879), AOM_ICDF(23981),
+            AOM_ICDF(25213), AOM_ICDF(26218), AOM_ICDF(27472), AOM_ICDF(28465),
+            AOM_ICDF(29221), AOM_ICDF(30232), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8448), AOM_ICDF(11108), AOM_ICDF(13392), AOM_ICDF(25167),
+            AOM_ICDF(26295), AOM_ICDF(26789), AOM_ICDF(27536), AOM_ICDF(28088),
+            AOM_ICDF(29039), AOM_ICDF(30700), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(13568), AOM_ICDF(15293), AOM_ICDF(18706), AOM_ICDF(21610),
+            AOM_ICDF(23139), AOM_ICDF(24254), AOM_ICDF(26383), AOM_ICDF(27630),
+            AOM_ICDF(28613), AOM_ICDF(30350), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9600), AOM_ICDF(11772), AOM_ICDF(14397), AOM_ICDF(16580),
+            AOM_ICDF(20091), AOM_ICDF(22865), AOM_ICDF(24490), AOM_ICDF(25395),
+            AOM_ICDF(27037), AOM_ICDF(28694), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(12160), AOM_ICDF(14092), AOM_ICDF(17010), AOM_ICDF(18922),
+            AOM_ICDF(22683), AOM_ICDF(25751), AOM_ICDF(27725), AOM_ICDF(30109),
+            AOM_ICDF(31449), AOM_ICDF(32763), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9088), AOM_ICDF(10383), AOM_ICDF(12569), AOM_ICDF(17113),
+            AOM_ICDF(21351), AOM_ICDF(22511), AOM_ICDF(23633), AOM_ICDF(24382),
+            AOM_ICDF(28215), AOM_ICDF(29798), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10880), AOM_ICDF(12248), AOM_ICDF(15214), AOM_ICDF(20017),
+            AOM_ICDF(21922), AOM_ICDF(22757), AOM_ICDF(24360), AOM_ICDF(25280),
+            AOM_ICDF(26684), AOM_ICDF(29869), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(11008), AOM_ICDF(13133), AOM_ICDF(15587), AOM_ICDF(17872),
+            AOM_ICDF(19579), AOM_ICDF(21157), AOM_ICDF(23788), AOM_ICDF(26629),
+            AOM_ICDF(27732), AOM_ICDF(29601), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10112), AOM_ICDF(12325), AOM_ICDF(15360), AOM_ICDF(18348),
+            AOM_ICDF(20452), AOM_ICDF(20460), AOM_ICDF(21902), AOM_ICDF(23982),
+            AOM_ICDF(25149), AOM_ICDF(26667), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8704), AOM_ICDF(14250), AOM_ICDF(17722), AOM_ICDF(23128),
+            AOM_ICDF(24217), AOM_ICDF(24892), AOM_ICDF(26215), AOM_ICDF(27392),
+            AOM_ICDF(28358), AOM_ICDF(30287), AOM_ICDF(32768), 0 },
+      },
+      {
+          { AOM_ICDF(8448), AOM_ICDF(10443), AOM_ICDF(20733), AOM_ICDF(23689),
+            AOM_ICDF(24634), AOM_ICDF(25951), AOM_ICDF(26670), AOM_ICDF(27861),
+            AOM_ICDF(28379), AOM_ICDF(29305), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6656), AOM_ICDF(9206), AOM_ICDF(24577), AOM_ICDF(25792),
+            AOM_ICDF(26335), AOM_ICDF(27169), AOM_ICDF(27913), AOM_ICDF(28956),
+            AOM_ICDF(29239), AOM_ICDF(29680), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7168), AOM_ICDF(8968), AOM_ICDF(15662), AOM_ICDF(22937),
+            AOM_ICDF(23849), AOM_ICDF(24616), AOM_ICDF(25603), AOM_ICDF(26555),
+            AOM_ICDF(27210), AOM_ICDF(29142), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9600), AOM_ICDF(11501), AOM_ICDF(19310), AOM_ICDF(21731),
+            AOM_ICDF(22790), AOM_ICDF(23936), AOM_ICDF(25627), AOM_ICDF(27217),
+            AOM_ICDF(27868), AOM_ICDF(29170), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6912), AOM_ICDF(8730), AOM_ICDF(17650), AOM_ICDF(19377),
+            AOM_ICDF(21025), AOM_ICDF(23319), AOM_ICDF(24537), AOM_ICDF(26112),
+            AOM_ICDF(26840), AOM_ICDF(28345), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7808), AOM_ICDF(9661), AOM_ICDF(20583), AOM_ICDF(21996),
+            AOM_ICDF(23898), AOM_ICDF(26818), AOM_ICDF(28120), AOM_ICDF(30716),
+            AOM_ICDF(31678), AOM_ICDF(32764), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6784), AOM_ICDF(8104), AOM_ICDF(15619), AOM_ICDF(18584),
+            AOM_ICDF(20844), AOM_ICDF(22519), AOM_ICDF(23760), AOM_ICDF(25203),
+            AOM_ICDF(27094), AOM_ICDF(28801), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8832), AOM_ICDF(10141), AOM_ICDF(17035), AOM_ICDF(20764),
+            AOM_ICDF(21703), AOM_ICDF(22751), AOM_ICDF(23964), AOM_ICDF(25305),
+            AOM_ICDF(26034), AOM_ICDF(29006), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8192), AOM_ICDF(9920), AOM_ICDF(19113), AOM_ICDF(20594),
+            AOM_ICDF(21747), AOM_ICDF(23327), AOM_ICDF(24581), AOM_ICDF(26916),
+            AOM_ICDF(27533), AOM_ICDF(28944), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6656), AOM_ICDF(8696), AOM_ICDF(18381), AOM_ICDF(20537),
+            AOM_ICDF(21804), AOM_ICDF(21809), AOM_ICDF(22751), AOM_ICDF(24394),
+            AOM_ICDF(24917), AOM_ICDF(25990), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6400), AOM_ICDF(9593), AOM_ICDF(20818), AOM_ICDF(23519),
+            AOM_ICDF(24266), AOM_ICDF(25113), AOM_ICDF(26608), AOM_ICDF(27883),
+            AOM_ICDF(28322), AOM_ICDF(29364), AOM_ICDF(32768), 0 },
+      },
+      {
+          { AOM_ICDF(12032), AOM_ICDF(14381), AOM_ICDF(16608), AOM_ICDF(24946),
+            AOM_ICDF(26084), AOM_ICDF(26582), AOM_ICDF(27428), AOM_ICDF(28075),
+            AOM_ICDF(29395), AOM_ICDF(30858), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9216), AOM_ICDF(12620), AOM_ICDF(18287), AOM_ICDF(24345),
+            AOM_ICDF(25984), AOM_ICDF(26715), AOM_ICDF(27732), AOM_ICDF(28519),
+            AOM_ICDF(29399), AOM_ICDF(30781), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6784), AOM_ICDF(8916), AOM_ICDF(10220), AOM_ICDF(26539),
+            AOM_ICDF(27310), AOM_ICDF(27483), AOM_ICDF(28082), AOM_ICDF(28430),
+            AOM_ICDF(29362), AOM_ICDF(31291), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(11904), AOM_ICDF(14838), AOM_ICDF(17359), AOM_ICDF(21663),
+            AOM_ICDF(22931), AOM_ICDF(23619), AOM_ICDF(25620), AOM_ICDF(26653),
+            AOM_ICDF(27823), AOM_ICDF(30547), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10752), AOM_ICDF(13504), AOM_ICDF(15536), AOM_ICDF(19057),
+            AOM_ICDF(21753), AOM_ICDF(23883), AOM_ICDF(25202), AOM_ICDF(26266),
+            AOM_ICDF(28196), AOM_ICDF(30589), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10496), AOM_ICDF(13193), AOM_ICDF(16787), AOM_ICDF(21011),
+            AOM_ICDF(23929), AOM_ICDF(25651), AOM_ICDF(27958), AOM_ICDF(29330),
+            AOM_ICDF(31022), AOM_ICDF(32761), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8448), AOM_ICDF(9968), AOM_ICDF(11749), AOM_ICDF(18062),
+            AOM_ICDF(21841), AOM_ICDF(22669), AOM_ICDF(23852), AOM_ICDF(24444),
+            AOM_ICDF(28118), AOM_ICDF(30007), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9728), AOM_ICDF(11168), AOM_ICDF(12602), AOM_ICDF(20819),
+            AOM_ICDF(22194), AOM_ICDF(22764), AOM_ICDF(24366), AOM_ICDF(25022),
+            AOM_ICDF(26414), AOM_ICDF(30460), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9216), AOM_ICDF(12712), AOM_ICDF(14357), AOM_ICDF(18346),
+            AOM_ICDF(20486), AOM_ICDF(21549), AOM_ICDF(23170), AOM_ICDF(25794),
+            AOM_ICDF(27129), AOM_ICDF(29574), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7808), AOM_ICDF(10733), AOM_ICDF(13057), AOM_ICDF(20252),
+            AOM_ICDF(21906), AOM_ICDF(21912), AOM_ICDF(23057), AOM_ICDF(24233),
+            AOM_ICDF(25700), AOM_ICDF(27439), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6784), AOM_ICDF(11352), AOM_ICDF(13778), AOM_ICDF(23877),
+            AOM_ICDF(24995), AOM_ICDF(25424), AOM_ICDF(26830), AOM_ICDF(27688),
+            AOM_ICDF(28779), AOM_ICDF(30368), AOM_ICDF(32768), 0 },
+      },
+      {
+          { AOM_ICDF(12288), AOM_ICDF(13728), AOM_ICDF(16480), AOM_ICDF(19841),
+            AOM_ICDF(21570), AOM_ICDF(22715), AOM_ICDF(25385), AOM_ICDF(27000),
+            AOM_ICDF(28329), AOM_ICDF(29994), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9344), AOM_ICDF(10991), AOM_ICDF(18817), AOM_ICDF(20972),
+            AOM_ICDF(22137), AOM_ICDF(23231), AOM_ICDF(26025), AOM_ICDF(27711),
+            AOM_ICDF(28244), AOM_ICDF(29428), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9344), AOM_ICDF(10900), AOM_ICDF(13206), AOM_ICDF(21344),
+            AOM_ICDF(22332), AOM_ICDF(22987), AOM_ICDF(25127), AOM_ICDF(26440),
+            AOM_ICDF(27231), AOM_ICDF(29502), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(12928), AOM_ICDF(14478), AOM_ICDF(15978), AOM_ICDF(18630),
+            AOM_ICDF(19852), AOM_ICDF(20897), AOM_ICDF(24699), AOM_ICDF(26464),
+            AOM_ICDF(27030), AOM_ICDF(30482), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9088), AOM_ICDF(10476), AOM_ICDF(13350), AOM_ICDF(15237),
+            AOM_ICDF(18175), AOM_ICDF(20252), AOM_ICDF(23283), AOM_ICDF(25321),
+            AOM_ICDF(26426), AOM_ICDF(29349), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10240), AOM_ICDF(11912), AOM_ICDF(15008), AOM_ICDF(17177),
+            AOM_ICDF(19979), AOM_ICDF(23056), AOM_ICDF(26395), AOM_ICDF(29681),
+            AOM_ICDF(30790), AOM_ICDF(32760), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8704), AOM_ICDF(9738), AOM_ICDF(11717), AOM_ICDF(15480),
+            AOM_ICDF(18656), AOM_ICDF(20022), AOM_ICDF(22611), AOM_ICDF(24357),
+            AOM_ICDF(27150), AOM_ICDF(29257), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(12928), AOM_ICDF(13548), AOM_ICDF(17978), AOM_ICDF(20602),
+            AOM_ICDF(21814), AOM_ICDF(22427), AOM_ICDF(24568), AOM_ICDF(25881),
+            AOM_ICDF(26823), AOM_ICDF(30817), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10496), AOM_ICDF(12149), AOM_ICDF(14082), AOM_ICDF(18054),
+            AOM_ICDF(19032), AOM_ICDF(19994), AOM_ICDF(24086), AOM_ICDF(28427),
+            AOM_ICDF(29156), AOM_ICDF(30680), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8448), AOM_ICDF(10158), AOM_ICDF(13867), AOM_ICDF(16506),
+            AOM_ICDF(18584), AOM_ICDF(18592), AOM_ICDF(21472), AOM_ICDF(23767),
+            AOM_ICDF(24646), AOM_ICDF(27279), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7296), AOM_ICDF(9684), AOM_ICDF(13471), AOM_ICDF(17701),
+            AOM_ICDF(18934), AOM_ICDF(19878), AOM_ICDF(25115), AOM_ICDF(27238),
+            AOM_ICDF(27972), AOM_ICDF(29583), AOM_ICDF(32768), 0 },
+      },
+      {
+          { AOM_ICDF(10880), AOM_ICDF(12163), AOM_ICDF(14497), AOM_ICDF(17112),
+            AOM_ICDF(20859), AOM_ICDF(22562), AOM_ICDF(23599), AOM_ICDF(24638),
+            AOM_ICDF(26861), AOM_ICDF(29399), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9984), AOM_ICDF(12476), AOM_ICDF(16360), AOM_ICDF(18889),
+            AOM_ICDF(21414), AOM_ICDF(23474), AOM_ICDF(24563), AOM_ICDF(25909),
+            AOM_ICDF(27195), AOM_ICDF(28828), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7168), AOM_ICDF(9268), AOM_ICDF(10737), AOM_ICDF(20063),
+            AOM_ICDF(22315), AOM_ICDF(23302), AOM_ICDF(24152), AOM_ICDF(25195),
+            AOM_ICDF(26645), AOM_ICDF(28845), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8960), AOM_ICDF(10727), AOM_ICDF(12449), AOM_ICDF(14263),
+            AOM_ICDF(16523), AOM_ICDF(17608), AOM_ICDF(23352), AOM_ICDF(24676),
+            AOM_ICDF(26478), AOM_ICDF(28886), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9856), AOM_ICDF(11109), AOM_ICDF(13309), AOM_ICDF(14975),
+            AOM_ICDF(19055), AOM_ICDF(21670), AOM_ICDF(23144), AOM_ICDF(24460),
+            AOM_ICDF(26212), AOM_ICDF(28107), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9984), AOM_ICDF(11586), AOM_ICDF(14565), AOM_ICDF(16562),
+            AOM_ICDF(21107), AOM_ICDF(25444), AOM_ICDF(27218), AOM_ICDF(29429),
+            AOM_ICDF(31451), AOM_ICDF(32763), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7168), AOM_ICDF(8268), AOM_ICDF(9704), AOM_ICDF(13144),
+            AOM_ICDF(18443), AOM_ICDF(20065), AOM_ICDF(21653), AOM_ICDF(23607),
+            AOM_ICDF(26506), AOM_ICDF(28854), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(11520), AOM_ICDF(13014), AOM_ICDF(14866), AOM_ICDF(18136),
+            AOM_ICDF(20231), AOM_ICDF(21509), AOM_ICDF(23004), AOM_ICDF(24186),
+            AOM_ICDF(25728), AOM_ICDF(29468), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10240), AOM_ICDF(12264), AOM_ICDF(14507), AOM_ICDF(16388),
+            AOM_ICDF(18888), AOM_ICDF(20927), AOM_ICDF(22731), AOM_ICDF(24691),
+            AOM_ICDF(26142), AOM_ICDF(28394), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8064), AOM_ICDF(10187), AOM_ICDF(12921), AOM_ICDF(15952),
+            AOM_ICDF(19960), AOM_ICDF(19976), AOM_ICDF(21275), AOM_ICDF(23205),
+            AOM_ICDF(25110), AOM_ICDF(26636), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8448), AOM_ICDF(11488), AOM_ICDF(14065), AOM_ICDF(19113),
+            AOM_ICDF(21604), AOM_ICDF(22978), AOM_ICDF(24508), AOM_ICDF(25895),
+            AOM_ICDF(27398), AOM_ICDF(29055), AOM_ICDF(32768), 0 },
+      },
+      {
+          { AOM_ICDF(10368), AOM_ICDF(11768), AOM_ICDF(16772), AOM_ICDF(19842),
+            AOM_ICDF(22940), AOM_ICDF(27394), AOM_ICDF(28528), AOM_ICDF(30267),
+            AOM_ICDF(31371), AOM_ICDF(32763), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9472), AOM_ICDF(11292), AOM_ICDF(18507), AOM_ICDF(20777),
+            AOM_ICDF(23357), AOM_ICDF(27587), AOM_ICDF(28902), AOM_ICDF(30850),
+            AOM_ICDF(31607), AOM_ICDF(32763), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8064), AOM_ICDF(9512), AOM_ICDF(13782), AOM_ICDF(20645),
+            AOM_ICDF(24493), AOM_ICDF(26242), AOM_ICDF(28001), AOM_ICDF(29435),
+            AOM_ICDF(30438), AOM_ICDF(32759), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8960), AOM_ICDF(10541), AOM_ICDF(15664), AOM_ICDF(17639),
+            AOM_ICDF(19646), AOM_ICDF(22145), AOM_ICDF(25216), AOM_ICDF(28815),
+            AOM_ICDF(30050), AOM_ICDF(32757), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9984), AOM_ICDF(11141), AOM_ICDF(15365), AOM_ICDF(16746),
+            AOM_ICDF(21186), AOM_ICDF(25766), AOM_ICDF(27817), AOM_ICDF(30022),
+            AOM_ICDF(31309), AOM_ICDF(32762), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9216), AOM_ICDF(10688), AOM_ICDF(16639), AOM_ICDF(17735),
+            AOM_ICDF(21499), AOM_ICDF(26657), AOM_ICDF(28161), AOM_ICDF(30572),
+            AOM_ICDF(31490), AOM_ICDF(32763), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8448), AOM_ICDF(9303), AOM_ICDF(13611), AOM_ICDF(16636),
+            AOM_ICDF(20555), AOM_ICDF(23414), AOM_ICDF(24912), AOM_ICDF(27613),
+            AOM_ICDF(29727), AOM_ICDF(32756), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9984), AOM_ICDF(11052), AOM_ICDF(16142), AOM_ICDF(19312),
+            AOM_ICDF(21680), AOM_ICDF(23870), AOM_ICDF(25504), AOM_ICDF(28200),
+            AOM_ICDF(29324), AOM_ICDF(32755), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10496), AOM_ICDF(12323), AOM_ICDF(16955), AOM_ICDF(18839),
+            AOM_ICDF(21144), AOM_ICDF(24861), AOM_ICDF(26838), AOM_ICDF(29988),
+            AOM_ICDF(30976), AOM_ICDF(32761), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(2944), AOM_ICDF(5973), AOM_ICDF(8904), AOM_ICDF(11875),
+            AOM_ICDF(14864), AOM_ICDF(17853), AOM_ICDF(20824), AOM_ICDF(23810),
+            AOM_ICDF(26784), AOM_ICDF(29776), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7424), AOM_ICDF(10097), AOM_ICDF(15588), AOM_ICDF(20217),
+            AOM_ICDF(23899), AOM_ICDF(26460), AOM_ICDF(28308), AOM_ICDF(30155),
+            AOM_ICDF(30951), AOM_ICDF(32761), AOM_ICDF(32768), 0 },
+      },
+      {
+          { AOM_ICDF(11648), AOM_ICDF(13133), AOM_ICDF(15050), AOM_ICDF(20481),
+            AOM_ICDF(22470), AOM_ICDF(23425), AOM_ICDF(24337), AOM_ICDF(25160),
+            AOM_ICDF(28964), AOM_ICDF(30480), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10240), AOM_ICDF(12616), AOM_ICDF(16631), AOM_ICDF(20485),
+            AOM_ICDF(22290), AOM_ICDF(23628), AOM_ICDF(25235), AOM_ICDF(26353),
+            AOM_ICDF(28107), AOM_ICDF(29655), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6784), AOM_ICDF(8002), AOM_ICDF(9066), AOM_ICDF(20038),
+            AOM_ICDF(22926), AOM_ICDF(23324), AOM_ICDF(23951), AOM_ICDF(24537),
+            AOM_ICDF(26916), AOM_ICDF(30231), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(11904), AOM_ICDF(14105), AOM_ICDF(15782), AOM_ICDF(19896),
+            AOM_ICDF(22283), AOM_ICDF(23147), AOM_ICDF(24763), AOM_ICDF(25983),
+            AOM_ICDF(27812), AOM_ICDF(29980), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10624), AOM_ICDF(11922), AOM_ICDF(13632), AOM_ICDF(15941),
+            AOM_ICDF(20469), AOM_ICDF(22453), AOM_ICDF(24065), AOM_ICDF(25187),
+            AOM_ICDF(27349), AOM_ICDF(29296), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(12032), AOM_ICDF(13085), AOM_ICDF(15468), AOM_ICDF(17768),
+            AOM_ICDF(20613), AOM_ICDF(24388), AOM_ICDF(26385), AOM_ICDF(28430),
+            AOM_ICDF(30938), AOM_ICDF(32761), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9728), AOM_ICDF(10538), AOM_ICDF(11493), AOM_ICDF(14765),
+            AOM_ICDF(18460), AOM_ICDF(19471), AOM_ICDF(20302), AOM_ICDF(20935),
+            AOM_ICDF(28192), AOM_ICDF(29926), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8960), AOM_ICDF(9890), AOM_ICDF(10962), AOM_ICDF(16685),
+            AOM_ICDF(18880), AOM_ICDF(19480), AOM_ICDF(20674), AOM_ICDF(21477),
+            AOM_ICDF(23815), AOM_ICDF(29341), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(14592), AOM_ICDF(16367), AOM_ICDF(17712), AOM_ICDF(20293),
+            AOM_ICDF(22544), AOM_ICDF(23829), AOM_ICDF(24877), AOM_ICDF(26326),
+            AOM_ICDF(27660), AOM_ICDF(29875), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8960), AOM_ICDF(10448), AOM_ICDF(12279), AOM_ICDF(16206),
+            AOM_ICDF(18672), AOM_ICDF(18682), AOM_ICDF(20058), AOM_ICDF(21547),
+            AOM_ICDF(25097), AOM_ICDF(27165), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(11136), AOM_ICDF(13840), AOM_ICDF(15762), AOM_ICDF(21710),
+            AOM_ICDF(23038), AOM_ICDF(23734), AOM_ICDF(24863), AOM_ICDF(25882),
+            AOM_ICDF(27765), AOM_ICDF(30071), AOM_ICDF(32768), 0 },
+      },
+      {
+          { AOM_ICDF(12544), AOM_ICDF(14124), AOM_ICDF(16964), AOM_ICDF(21907),
+            AOM_ICDF(23808), AOM_ICDF(24496), AOM_ICDF(25724), AOM_ICDF(26715),
+            AOM_ICDF(27992), AOM_ICDF(30455), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10368), AOM_ICDF(13606), AOM_ICDF(18247), AOM_ICDF(20869),
+            AOM_ICDF(22590), AOM_ICDF(23749), AOM_ICDF(25088), AOM_ICDF(26378),
+            AOM_ICDF(27277), AOM_ICDF(29808), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9088), AOM_ICDF(11031), AOM_ICDF(12899), AOM_ICDF(23497),
+            AOM_ICDF(24465), AOM_ICDF(24851), AOM_ICDF(25995), AOM_ICDF(26815),
+            AOM_ICDF(27885), AOM_ICDF(30555), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(11520), AOM_ICDF(14342), AOM_ICDF(15710), AOM_ICDF(19196),
+            AOM_ICDF(21250), AOM_ICDF(21907), AOM_ICDF(24665), AOM_ICDF(26153),
+            AOM_ICDF(27212), AOM_ICDF(30750), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9984), AOM_ICDF(11764), AOM_ICDF(13979), AOM_ICDF(16405),
+            AOM_ICDF(19279), AOM_ICDF(20658), AOM_ICDF(23354), AOM_ICDF(25266),
+            AOM_ICDF(26702), AOM_ICDF(29380), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10112), AOM_ICDF(12325), AOM_ICDF(15918), AOM_ICDF(19060),
+            AOM_ICDF(21829), AOM_ICDF(23882), AOM_ICDF(26277), AOM_ICDF(27697),
+            AOM_ICDF(30114), AOM_ICDF(32758), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9344), AOM_ICDF(10534), AOM_ICDF(12184), AOM_ICDF(16208),
+            AOM_ICDF(19764), AOM_ICDF(20627), AOM_ICDF(22524), AOM_ICDF(23644),
+            AOM_ICDF(26887), AOM_ICDF(29782), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(12928), AOM_ICDF(14013), AOM_ICDF(15625), AOM_ICDF(19107),
+            AOM_ICDF(20654), AOM_ICDF(21451), AOM_ICDF(22910), AOM_ICDF(23873),
+            AOM_ICDF(24776), AOM_ICDF(30239), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10368), AOM_ICDF(12818), AOM_ICDF(14610), AOM_ICDF(17350),
+            AOM_ICDF(19568), AOM_ICDF(20710), AOM_ICDF(22971), AOM_ICDF(25114),
+            AOM_ICDF(26340), AOM_ICDF(29127), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8960), AOM_ICDF(11192), AOM_ICDF(13720), AOM_ICDF(18429),
+            AOM_ICDF(20409), AOM_ICDF(20417), AOM_ICDF(22250), AOM_ICDF(23318),
+            AOM_ICDF(24647), AOM_ICDF(27248), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7808), AOM_ICDF(11416), AOM_ICDF(13918), AOM_ICDF(19028),
+            AOM_ICDF(20181), AOM_ICDF(20839), AOM_ICDF(24380), AOM_ICDF(26018),
+            AOM_ICDF(26967), AOM_ICDF(29845), AOM_ICDF(32768), 0 },
+      },
+      {
+          { AOM_ICDF(9856), AOM_ICDF(11020), AOM_ICDF(14928), AOM_ICDF(18159),
+            AOM_ICDF(19421), AOM_ICDF(20921), AOM_ICDF(23466), AOM_ICDF(26664),
+            AOM_ICDF(27475), AOM_ICDF(28881), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8704), AOM_ICDF(10302), AOM_ICDF(17323), AOM_ICDF(18907),
+            AOM_ICDF(19868), AOM_ICDF(21184), AOM_ICDF(24171), AOM_ICDF(28033),
+            AOM_ICDF(28625), AOM_ICDF(29353), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7936), AOM_ICDF(9197), AOM_ICDF(12604), AOM_ICDF(20616),
+            AOM_ICDF(21514), AOM_ICDF(22371), AOM_ICDF(24239), AOM_ICDF(26138),
+            AOM_ICDF(26863), AOM_ICDF(29239), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(11264), AOM_ICDF(12524), AOM_ICDF(16083), AOM_ICDF(18574),
+            AOM_ICDF(19858), AOM_ICDF(20841), AOM_ICDF(24242), AOM_ICDF(27606),
+            AOM_ICDF(28352), AOM_ICDF(29853), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8704), AOM_ICDF(10208), AOM_ICDF(13292), AOM_ICDF(15170),
+            AOM_ICDF(17277), AOM_ICDF(19226), AOM_ICDF(22083), AOM_ICDF(25046),
+            AOM_ICDF(26041), AOM_ICDF(27802), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9088), AOM_ICDF(10568), AOM_ICDF(15511), AOM_ICDF(17246),
+            AOM_ICDF(20170), AOM_ICDF(22791), AOM_ICDF(25558), AOM_ICDF(30740),
+            AOM_ICDF(31635), AOM_ICDF(32764), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7040), AOM_ICDF(8045), AOM_ICDF(10653), AOM_ICDF(13145),
+            AOM_ICDF(15286), AOM_ICDF(16614), AOM_ICDF(19075), AOM_ICDF(23140),
+            AOM_ICDF(26224), AOM_ICDF(28652), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(10240), AOM_ICDF(11032), AOM_ICDF(14258), AOM_ICDF(17629),
+            AOM_ICDF(18914), AOM_ICDF(19898), AOM_ICDF(22412), AOM_ICDF(24961),
+            AOM_ICDF(25815), AOM_ICDF(29156), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(11008), AOM_ICDF(12028), AOM_ICDF(14702), AOM_ICDF(16147),
+            AOM_ICDF(17209), AOM_ICDF(18160), AOM_ICDF(21812), AOM_ICDF(27547),
+            AOM_ICDF(28709), AOM_ICDF(30120), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7168), AOM_ICDF(9068), AOM_ICDF(14160), AOM_ICDF(16937),
+            AOM_ICDF(18515), AOM_ICDF(18521), AOM_ICDF(20636), AOM_ICDF(24617),
+            AOM_ICDF(25317), AOM_ICDF(26365), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6784), AOM_ICDF(8510), AOM_ICDF(14195), AOM_ICDF(17148),
+            AOM_ICDF(18158), AOM_ICDF(19201), AOM_ICDF(23070), AOM_ICDF(27351),
+            AOM_ICDF(27901), AOM_ICDF(29422), AOM_ICDF(32768), 0 },
+      },
+      {
+          { AOM_ICDF(10112), AOM_ICDF(11528), AOM_ICDF(15345), AOM_ICDF(19296),
+            AOM_ICDF(21394), AOM_ICDF(21402), AOM_ICDF(22379), AOM_ICDF(23840),
+            AOM_ICDF(24851), AOM_ICDF(26150), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8064), AOM_ICDF(10187), AOM_ICDF(17949), AOM_ICDF(20052),
+            AOM_ICDF(22051), AOM_ICDF(22059), AOM_ICDF(23147), AOM_ICDF(24688),
+            AOM_ICDF(25351), AOM_ICDF(26365), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6528), AOM_ICDF(8373), AOM_ICDF(11041), AOM_ICDF(21963),
+            AOM_ICDF(23089), AOM_ICDF(23093), AOM_ICDF(24076), AOM_ICDF(24925),
+            AOM_ICDF(25691), AOM_ICDF(27764), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9600), AOM_ICDF(11229), AOM_ICDF(14847), AOM_ICDF(17527),
+            AOM_ICDF(19738), AOM_ICDF(19747), AOM_ICDF(21629), AOM_ICDF(23761),
+            AOM_ICDF(24957), AOM_ICDF(27673), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8960), AOM_ICDF(10262), AOM_ICDF(13339), AOM_ICDF(15480),
+            AOM_ICDF(19925), AOM_ICDF(19942), AOM_ICDF(21445), AOM_ICDF(23037),
+            AOM_ICDF(24329), AOM_ICDF(25977), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(2944), AOM_ICDF(5973), AOM_ICDF(8904), AOM_ICDF(11875),
+            AOM_ICDF(14864), AOM_ICDF(17853), AOM_ICDF(20824), AOM_ICDF(23810),
+            AOM_ICDF(26784), AOM_ICDF(29776), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9472), AOM_ICDF(10564), AOM_ICDF(13426), AOM_ICDF(16561),
+            AOM_ICDF(19685), AOM_ICDF(19697), AOM_ICDF(21076), AOM_ICDF(22583),
+            AOM_ICDF(24891), AOM_ICDF(26983), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8448), AOM_ICDF(9493), AOM_ICDF(12221), AOM_ICDF(16542),
+            AOM_ICDF(18394), AOM_ICDF(18401), AOM_ICDF(19580), AOM_ICDF(20971),
+            AOM_ICDF(22031), AOM_ICDF(26770), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8704), AOM_ICDF(10772), AOM_ICDF(14209), AOM_ICDF(16381),
+            AOM_ICDF(18911), AOM_ICDF(18921), AOM_ICDF(20436), AOM_ICDF(23374),
+            AOM_ICDF(24475), AOM_ICDF(26095), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(7680), AOM_ICDF(9444), AOM_ICDF(13453), AOM_ICDF(16320),
+            AOM_ICDF(18650), AOM_ICDF(18659), AOM_ICDF(19651), AOM_ICDF(21291),
+            AOM_ICDF(22277), AOM_ICDF(23916), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6656), AOM_ICDF(9920), AOM_ICDF(14740), AOM_ICDF(19864),
+            AOM_ICDF(21495), AOM_ICDF(21501), AOM_ICDF(22953), AOM_ICDF(24372),
+            AOM_ICDF(25192), AOM_ICDF(26760), AOM_ICDF(32768), 0 },
+      },
+      {
+          { AOM_ICDF(9728), AOM_ICDF(13958), AOM_ICDF(18881), AOM_ICDF(23624),
+            AOM_ICDF(24754), AOM_ICDF(25553), AOM_ICDF(26709), AOM_ICDF(27940),
+            AOM_ICDF(28977), AOM_ICDF(30413), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8832), AOM_ICDF(12572), AOM_ICDF(22433), AOM_ICDF(24653),
+            AOM_ICDF(25676), AOM_ICDF(26551), AOM_ICDF(27571), AOM_ICDF(28688),
+            AOM_ICDF(29198), AOM_ICDF(30174), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(5888), AOM_ICDF(8828), AOM_ICDF(11353), AOM_ICDF(23482),
+            AOM_ICDF(24310), AOM_ICDF(24737), AOM_ICDF(25804), AOM_ICDF(26375),
+            AOM_ICDF(27174), AOM_ICDF(29840), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9984), AOM_ICDF(13099), AOM_ICDF(16249), AOM_ICDF(19443),
+            AOM_ICDF(20990), AOM_ICDF(22637), AOM_ICDF(24576), AOM_ICDF(25952),
+            AOM_ICDF(26884), AOM_ICDF(29435), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8448), AOM_ICDF(11108), AOM_ICDF(15085), AOM_ICDF(18134),
+            AOM_ICDF(20319), AOM_ICDF(21992), AOM_ICDF(23549), AOM_ICDF(24989),
+            AOM_ICDF(27177), AOM_ICDF(29208), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9856), AOM_ICDF(13168), AOM_ICDF(18987), AOM_ICDF(22481),
+            AOM_ICDF(24282), AOM_ICDF(26200), AOM_ICDF(27868), AOM_ICDF(30203),
+            AOM_ICDF(31085), AOM_ICDF(32761), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6784), AOM_ICDF(9119), AOM_ICDF(12629), AOM_ICDF(16877),
+            AOM_ICDF(20262), AOM_ICDF(21125), AOM_ICDF(22307), AOM_ICDF(23615),
+            AOM_ICDF(27727), AOM_ICDF(29972), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(8320), AOM_ICDF(10230), AOM_ICDF(12783), AOM_ICDF(19005),
+            AOM_ICDF(20213), AOM_ICDF(20668), AOM_ICDF(22039), AOM_ICDF(23045),
+            AOM_ICDF(24146), AOM_ICDF(30478), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(9088), AOM_ICDF(11308), AOM_ICDF(15416), AOM_ICDF(18118),
+            AOM_ICDF(19762), AOM_ICDF(20906), AOM_ICDF(22574), AOM_ICDF(25162),
+            AOM_ICDF(25994), AOM_ICDF(28455), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6912), AOM_ICDF(10548), AOM_ICDF(15148), AOM_ICDF(20026),
+            AOM_ICDF(21612), AOM_ICDF(21618), AOM_ICDF(22707), AOM_ICDF(24200),
+            AOM_ICDF(24869), AOM_ICDF(26844), AOM_ICDF(32768), 0 },
+          { AOM_ICDF(6656), AOM_ICDF(12164), AOM_ICDF(16993), AOM_ICDF(21568),
+            AOM_ICDF(22933), AOM_ICDF(23648), AOM_ICDF(25322), AOM_ICDF(26602),
+            AOM_ICDF(27806), AOM_ICDF(29841), AOM_ICDF(32768), 0 },
+      },
+#else   // !CONFIG_ALT_INTRA
+      { { AOM_ICDF(17536), AOM_ICDF(19321), AOM_ICDF(21527), AOM_ICDF(25360),
+          AOM_ICDF(27516), AOM_ICDF(28026), AOM_ICDF(29323), AOM_ICDF(30023),
+          AOM_ICDF(30999), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(11776), AOM_ICDF(15466), AOM_ICDF(22360), AOM_ICDF(24865),
+          AOM_ICDF(26991), AOM_ICDF(27889), AOM_ICDF(29299), AOM_ICDF(30519),
+          AOM_ICDF(31398), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(9344), AOM_ICDF(12272), AOM_ICDF(13793), AOM_ICDF(25813),
+          AOM_ICDF(27359), AOM_ICDF(27654), AOM_ICDF(28573), AOM_ICDF(29130),
+          AOM_ICDF(30551), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(11648), AOM_ICDF(14123), AOM_ICDF(16454), AOM_ICDF(19948),
+          AOM_ICDF(22780), AOM_ICDF(23846), AOM_ICDF(27087), AOM_ICDF(28995),
+          AOM_ICDF(30380), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(9216), AOM_ICDF(12436), AOM_ICDF(15295), AOM_ICDF(17996),
+          AOM_ICDF(24006), AOM_ICDF(25465), AOM_ICDF(27405), AOM_ICDF(28725),
+          AOM_ICDF(30383), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(9344), AOM_ICDF(12181), AOM_ICDF(14433), AOM_ICDF(16634),
+          AOM_ICDF(20355), AOM_ICDF(24317), AOM_ICDF(26133), AOM_ICDF(29295),
+          AOM_ICDF(31344), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(8576), AOM_ICDF(10750), AOM_ICDF(12556), AOM_ICDF(17996),
+          AOM_ICDF(22315), AOM_ICDF(23609), AOM_ICDF(25040), AOM_ICDF(26157),
+          AOM_ICDF(30573), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(11008), AOM_ICDF(13303), AOM_ICDF(15432), AOM_ICDF(20646),
+          AOM_ICDF(23506), AOM_ICDF(24100), AOM_ICDF(25624), AOM_ICDF(26824),
+          AOM_ICDF(28055), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(9472), AOM_ICDF(12384), AOM_ICDF(14534), AOM_ICDF(17094),
+          AOM_ICDF(20257), AOM_ICDF(22155), AOM_ICDF(24767), AOM_ICDF(28955),
+          AOM_ICDF(30474), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7552), AOM_ICDF(14152), AOM_ICDF(17352), AOM_ICDF(22654),
+          AOM_ICDF(25123), AOM_ICDF(25783), AOM_ICDF(27911), AOM_ICDF(29182),
+          AOM_ICDF(30849), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(8064), AOM_ICDF(11538), AOM_ICDF(21987), AOM_ICDF(24941),
+          AOM_ICDF(26913), AOM_ICDF(28136), AOM_ICDF(29222), AOM_ICDF(30469),
+          AOM_ICDF(31331), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5504), AOM_ICDF(10403), AOM_ICDF(25080), AOM_ICDF(26762),
+          AOM_ICDF(27933), AOM_ICDF(29104), AOM_ICDF(30092), AOM_ICDF(31576),
+          AOM_ICDF(32004), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5632), AOM_ICDF(8706), AOM_ICDF(15097), AOM_ICDF(23714),
+          AOM_ICDF(25344), AOM_ICDF(26072), AOM_ICDF(27380), AOM_ICDF(28580),
+          AOM_ICDF(29840), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7424), AOM_ICDF(11186), AOM_ICDF(17593), AOM_ICDF(20154),
+          AOM_ICDF(22974), AOM_ICDF(24351), AOM_ICDF(26916), AOM_ICDF(29956),
+          AOM_ICDF(30967), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5888), AOM_ICDF(10193), AOM_ICDF(16895), AOM_ICDF(19031),
+          AOM_ICDF(23735), AOM_ICDF(25576), AOM_ICDF(27514), AOM_ICDF(29813),
+          AOM_ICDF(30471), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4864), AOM_ICDF(8352), AOM_ICDF(16459), AOM_ICDF(18062),
+          AOM_ICDF(21263), AOM_ICDF(25378), AOM_ICDF(26937), AOM_ICDF(30376),
+          AOM_ICDF(31619), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4992), AOM_ICDF(7922), AOM_ICDF(13842), AOM_ICDF(18004),
+          AOM_ICDF(21779), AOM_ICDF(23527), AOM_ICDF(25115), AOM_ICDF(27357),
+          AOM_ICDF(30232), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6656), AOM_ICDF(9716), AOM_ICDF(16379), AOM_ICDF(20053),
+          AOM_ICDF(22487), AOM_ICDF(23613), AOM_ICDF(25437), AOM_ICDF(27270),
+          AOM_ICDF(28516), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6016), AOM_ICDF(9674), AOM_ICDF(16891), AOM_ICDF(18684),
+          AOM_ICDF(21147), AOM_ICDF(23093), AOM_ICDF(25512), AOM_ICDF(30132),
+          AOM_ICDF(30894), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4608), AOM_ICDF(11318), AOM_ICDF(21038), AOM_ICDF(23650),
+          AOM_ICDF(25303), AOM_ICDF(26262), AOM_ICDF(28295), AOM_ICDF(30479),
+          AOM_ICDF(31212), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(10496), AOM_ICDF(12758), AOM_ICDF(14790), AOM_ICDF(24547),
+          AOM_ICDF(26342), AOM_ICDF(26799), AOM_ICDF(27825), AOM_ICDF(28443),
+          AOM_ICDF(30217), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7040), AOM_ICDF(11462), AOM_ICDF(17121), AOM_ICDF(24215),
+          AOM_ICDF(26504), AOM_ICDF(27267), AOM_ICDF(28492), AOM_ICDF(29444),
+          AOM_ICDF(30846), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5376), AOM_ICDF(8158), AOM_ICDF(9215), AOM_ICDF(26451),
+          AOM_ICDF(27407), AOM_ICDF(27524), AOM_ICDF(27995), AOM_ICDF(28275),
+          AOM_ICDF(29767), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(8704), AOM_ICDF(12652), AOM_ICDF(14145), AOM_ICDF(20101),
+          AOM_ICDF(22879), AOM_ICDF(23675), AOM_ICDF(25629), AOM_ICDF(27079),
+          AOM_ICDF(28923), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7424), AOM_ICDF(12374), AOM_ICDF(14366), AOM_ICDF(18855),
+          AOM_ICDF(23842), AOM_ICDF(24358), AOM_ICDF(25639), AOM_ICDF(27087),
+          AOM_ICDF(29706), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6400), AOM_ICDF(10005), AOM_ICDF(12939), AOM_ICDF(17753),
+          AOM_ICDF(22206), AOM_ICDF(24790), AOM_ICDF(26785), AOM_ICDF(28164),
+          AOM_ICDF(30520), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5632), AOM_ICDF(8176), AOM_ICDF(9713), AOM_ICDF(19053),
+          AOM_ICDF(22343), AOM_ICDF(23222), AOM_ICDF(24453), AOM_ICDF(25070),
+          AOM_ICDF(29761), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7040), AOM_ICDF(9754), AOM_ICDF(10833), AOM_ICDF(21229),
+          AOM_ICDF(23540), AOM_ICDF(23943), AOM_ICDF(24839), AOM_ICDF(25675),
+          AOM_ICDF(27033), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6784), AOM_ICDF(11758), AOM_ICDF(13481), AOM_ICDF(17236),
+          AOM_ICDF(20210), AOM_ICDF(21768), AOM_ICDF(24303), AOM_ICDF(26948),
+          AOM_ICDF(28676), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4864), AOM_ICDF(12712), AOM_ICDF(14201), AOM_ICDF(23863),
+          AOM_ICDF(25952), AOM_ICDF(26386), AOM_ICDF(27632), AOM_ICDF(28635),
+          AOM_ICDF(30362), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(13184), AOM_ICDF(15173), AOM_ICDF(17647), AOM_ICDF(21576),
+          AOM_ICDF(24474), AOM_ICDF(25267), AOM_ICDF(27699), AOM_ICDF(29283),
+          AOM_ICDF(30549), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7552), AOM_ICDF(11295), AOM_ICDF(18257), AOM_ICDF(20811),
+          AOM_ICDF(23213), AOM_ICDF(24606), AOM_ICDF(27731), AOM_ICDF(30407),
+          AOM_ICDF(31237), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7936), AOM_ICDF(10846), AOM_ICDF(12816), AOM_ICDF(22436),
+          AOM_ICDF(24614), AOM_ICDF(25130), AOM_ICDF(26890), AOM_ICDF(28199),
+          AOM_ICDF(29091), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(8576), AOM_ICDF(11411), AOM_ICDF(13830), AOM_ICDF(15918),
+          AOM_ICDF(18996), AOM_ICDF(20044), AOM_ICDF(25114), AOM_ICDF(27835),
+          AOM_ICDF(28972), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7680), AOM_ICDF(10816), AOM_ICDF(13646), AOM_ICDF(15966),
+          AOM_ICDF(21162), AOM_ICDF(22012), AOM_ICDF(24701), AOM_ICDF(27506),
+          AOM_ICDF(29644), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6784), AOM_ICDF(9423), AOM_ICDF(12524), AOM_ICDF(14773),
+          AOM_ICDF(19447), AOM_ICDF(22804), AOM_ICDF(26073), AOM_ICDF(29211),
+          AOM_ICDF(30642), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6784), AOM_ICDF(8916), AOM_ICDF(11059), AOM_ICDF(15861),
+          AOM_ICDF(21174), AOM_ICDF(22338), AOM_ICDF(24620), AOM_ICDF(27071),
+          AOM_ICDF(30899), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(9856), AOM_ICDF(11557), AOM_ICDF(13960), AOM_ICDF(18525),
+          AOM_ICDF(21788), AOM_ICDF(22189), AOM_ICDF(24462), AOM_ICDF(26603),
+          AOM_ICDF(27470), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7808), AOM_ICDF(10636), AOM_ICDF(13143), AOM_ICDF(15844),
+          AOM_ICDF(18698), AOM_ICDF(20272), AOM_ICDF(24323), AOM_ICDF(30096),
+          AOM_ICDF(31787), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6016), AOM_ICDF(10928), AOM_ICDF(14596), AOM_ICDF(18926),
+          AOM_ICDF(21586), AOM_ICDF(22688), AOM_ICDF(26626), AOM_ICDF(29001),
+          AOM_ICDF(30399), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(8832), AOM_ICDF(10983), AOM_ICDF(13451), AOM_ICDF(16582),
+          AOM_ICDF(21656), AOM_ICDF(23109), AOM_ICDF(24845), AOM_ICDF(26207),
+          AOM_ICDF(28796), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6784), AOM_ICDF(10844), AOM_ICDF(15554), AOM_ICDF(18073),
+          AOM_ICDF(22954), AOM_ICDF(24901), AOM_ICDF(26776), AOM_ICDF(28649),
+          AOM_ICDF(30419), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5120), AOM_ICDF(8252), AOM_ICDF(10072), AOM_ICDF(20108),
+          AOM_ICDF(23535), AOM_ICDF(24346), AOM_ICDF(25761), AOM_ICDF(26418),
+          AOM_ICDF(28675), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7680), AOM_ICDF(11012), AOM_ICDF(12627), AOM_ICDF(14595),
+          AOM_ICDF(19462), AOM_ICDF(20888), AOM_ICDF(23348), AOM_ICDF(25703),
+          AOM_ICDF(28159), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6656), AOM_ICDF(9818), AOM_ICDF(11790), AOM_ICDF(13813),
+          AOM_ICDF(22731), AOM_ICDF(24737), AOM_ICDF(26557), AOM_ICDF(28061),
+          AOM_ICDF(29697), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5632), AOM_ICDF(8918), AOM_ICDF(11620), AOM_ICDF(13802),
+          AOM_ICDF(19950), AOM_ICDF(23764), AOM_ICDF(25734), AOM_ICDF(28537),
+          AOM_ICDF(31809), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4480), AOM_ICDF(6580), AOM_ICDF(7808), AOM_ICDF(12281),
+          AOM_ICDF(19375), AOM_ICDF(20970), AOM_ICDF(22860), AOM_ICDF(24602),
+          AOM_ICDF(29929), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7040), AOM_ICDF(9553), AOM_ICDF(11457), AOM_ICDF(15102),
+          AOM_ICDF(20291), AOM_ICDF(21280), AOM_ICDF(22985), AOM_ICDF(24475),
+          AOM_ICDF(26613), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6528), AOM_ICDF(10423), AOM_ICDF(12605), AOM_ICDF(14621),
+          AOM_ICDF(19031), AOM_ICDF(21505), AOM_ICDF(24585), AOM_ICDF(27558),
+          AOM_ICDF(29532), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6016), AOM_ICDF(11659), AOM_ICDF(14463), AOM_ICDF(18867),
+          AOM_ICDF(23653), AOM_ICDF(24903), AOM_ICDF(27115), AOM_ICDF(29389),
+          AOM_ICDF(31382), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(8192), AOM_ICDF(10016), AOM_ICDF(13304), AOM_ICDF(16362),
+          AOM_ICDF(21107), AOM_ICDF(25165), AOM_ICDF(26620), AOM_ICDF(28901),
+          AOM_ICDF(30910), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5888), AOM_ICDF(8723), AOM_ICDF(16237), AOM_ICDF(18318),
+          AOM_ICDF(22002), AOM_ICDF(25923), AOM_ICDF(27394), AOM_ICDF(29934),
+          AOM_ICDF(31428), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4608), AOM_ICDF(7138), AOM_ICDF(9841), AOM_ICDF(18442),
+          AOM_ICDF(22447), AOM_ICDF(24618), AOM_ICDF(26337), AOM_ICDF(27945),
+          AOM_ICDF(30168), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6784), AOM_ICDF(8916), AOM_ICDF(12270), AOM_ICDF(14851),
+          AOM_ICDF(19886), AOM_ICDF(22759), AOM_ICDF(25105), AOM_ICDF(28368),
+          AOM_ICDF(29760), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5120), AOM_ICDF(7928), AOM_ICDF(11324), AOM_ICDF(13340),
+          AOM_ICDF(21205), AOM_ICDF(24224), AOM_ICDF(25926), AOM_ICDF(28518),
+          AOM_ICDF(30560), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4480), AOM_ICDF(6580), AOM_ICDF(10058), AOM_ICDF(11237),
+          AOM_ICDF(16807), AOM_ICDF(25937), AOM_ICDF(27218), AOM_ICDF(30015),
+          AOM_ICDF(31348), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4608), AOM_ICDF(6808), AOM_ICDF(9445), AOM_ICDF(12446),
+          AOM_ICDF(18461), AOM_ICDF(21835), AOM_ICDF(23244), AOM_ICDF(26109),
+          AOM_ICDF(30115), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5760), AOM_ICDF(7659), AOM_ICDF(10798), AOM_ICDF(14720),
+          AOM_ICDF(19157), AOM_ICDF(21955), AOM_ICDF(23645), AOM_ICDF(26460),
+          AOM_ICDF(28702), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5760), AOM_ICDF(8503), AOM_ICDF(11157), AOM_ICDF(13071),
+          AOM_ICDF(17594), AOM_ICDF(22047), AOM_ICDF(24099), AOM_ICDF(29077),
+          AOM_ICDF(30850), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4864), AOM_ICDF(9660), AOM_ICDF(14264), AOM_ICDF(17105),
+          AOM_ICDF(21528), AOM_ICDF(24094), AOM_ICDF(26025), AOM_ICDF(28580),
+          AOM_ICDF(30559), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(9600), AOM_ICDF(11139), AOM_ICDF(12998), AOM_ICDF(18660),
+          AOM_ICDF(22158), AOM_ICDF(23501), AOM_ICDF(24659), AOM_ICDF(25736),
+          AOM_ICDF(30296), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7168), AOM_ICDF(11068), AOM_ICDF(15984), AOM_ICDF(19969),
+          AOM_ICDF(23169), AOM_ICDF(24704), AOM_ICDF(26216), AOM_ICDF(27572),
+          AOM_ICDF(31368), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4480), AOM_ICDF(6801), AOM_ICDF(8018), AOM_ICDF(20908),
+          AOM_ICDF(23071), AOM_ICDF(23583), AOM_ICDF(24301), AOM_ICDF(25062),
+          AOM_ICDF(29427), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7168), AOM_ICDF(10068), AOM_ICDF(11753), AOM_ICDF(15843),
+          AOM_ICDF(19742), AOM_ICDF(21358), AOM_ICDF(23809), AOM_ICDF(26189),
+          AOM_ICDF(29067), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6016), AOM_ICDF(9047), AOM_ICDF(10622), AOM_ICDF(13931),
+          AOM_ICDF(22462), AOM_ICDF(23858), AOM_ICDF(25911), AOM_ICDF(27277),
+          AOM_ICDF(29722), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5888), AOM_ICDF(7568), AOM_ICDF(9931), AOM_ICDF(13533),
+          AOM_ICDF(18431), AOM_ICDF(22063), AOM_ICDF(23777), AOM_ICDF(26025),
+          AOM_ICDF(30555), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4352), AOM_ICDF(6239), AOM_ICDF(7379), AOM_ICDF(13739),
+          AOM_ICDF(16917), AOM_ICDF(18090), AOM_ICDF(18835), AOM_ICDF(19651),
+          AOM_ICDF(30360), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6528), AOM_ICDF(8988), AOM_ICDF(10288), AOM_ICDF(15534),
+          AOM_ICDF(19495), AOM_ICDF(20386), AOM_ICDF(21934), AOM_ICDF(23034),
+          AOM_ICDF(26988), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7040), AOM_ICDF(10055), AOM_ICDF(11652), AOM_ICDF(14757),
+          AOM_ICDF(19622), AOM_ICDF(21715), AOM_ICDF(23615), AOM_ICDF(26761),
+          AOM_ICDF(29483), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4736), AOM_ICDF(10102), AOM_ICDF(12315), AOM_ICDF(19078),
+          AOM_ICDF(21348), AOM_ICDF(22621), AOM_ICDF(24246), AOM_ICDF(26044),
+          AOM_ICDF(29931), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(10496), AOM_ICDF(12410), AOM_ICDF(14955), AOM_ICDF(19891),
+          AOM_ICDF(23137), AOM_ICDF(23792), AOM_ICDF(25159), AOM_ICDF(26378),
+          AOM_ICDF(28125), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7936), AOM_ICDF(12204), AOM_ICDF(17104), AOM_ICDF(20191),
+          AOM_ICDF(23468), AOM_ICDF(24630), AOM_ICDF(26156), AOM_ICDF(27628),
+          AOM_ICDF(28913), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6016), AOM_ICDF(8629), AOM_ICDF(10232), AOM_ICDF(23591),
+          AOM_ICDF(25349), AOM_ICDF(25637), AOM_ICDF(26306), AOM_ICDF(27063),
+          AOM_ICDF(28980), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(8704), AOM_ICDF(12088), AOM_ICDF(13461), AOM_ICDF(16646),
+          AOM_ICDF(20516), AOM_ICDF(21455), AOM_ICDF(24062), AOM_ICDF(26579),
+          AOM_ICDF(28368), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7296), AOM_ICDF(11177), AOM_ICDF(13117), AOM_ICDF(16196),
+          AOM_ICDF(23378), AOM_ICDF(24708), AOM_ICDF(26440), AOM_ICDF(27997),
+          AOM_ICDF(29078), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6272), AOM_ICDF(9377), AOM_ICDF(12575), AOM_ICDF(15616),
+          AOM_ICDF(20919), AOM_ICDF(23697), AOM_ICDF(26603), AOM_ICDF(27566),
+          AOM_ICDF(29903), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6528), AOM_ICDF(9091), AOM_ICDF(10478), AOM_ICDF(16445),
+          AOM_ICDF(21081), AOM_ICDF(22320), AOM_ICDF(23871), AOM_ICDF(25087),
+          AOM_ICDF(29258), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(8704), AOM_ICDF(11148), AOM_ICDF(12499), AOM_ICDF(17340),
+          AOM_ICDF(20656), AOM_ICDF(21288), AOM_ICDF(22588), AOM_ICDF(23701),
+          AOM_ICDF(24693), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7552), AOM_ICDF(11394), AOM_ICDF(12980), AOM_ICDF(15562),
+          AOM_ICDF(19942), AOM_ICDF(21792), AOM_ICDF(25093), AOM_ICDF(28211),
+          AOM_ICDF(28959), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5120), AOM_ICDF(11708), AOM_ICDF(13847), AOM_ICDF(19377),
+          AOM_ICDF(22421), AOM_ICDF(23160), AOM_ICDF(25449), AOM_ICDF(27136),
+          AOM_ICDF(29182), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(9984), AOM_ICDF(12031), AOM_ICDF(15190), AOM_ICDF(18673),
+          AOM_ICDF(21422), AOM_ICDF(22812), AOM_ICDF(25690), AOM_ICDF(29118),
+          AOM_ICDF(30458), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6144), AOM_ICDF(9680), AOM_ICDF(17436), AOM_ICDF(19610),
+          AOM_ICDF(21820), AOM_ICDF(23485), AOM_ICDF(26313), AOM_ICDF(30826),
+          AOM_ICDF(31843), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6016), AOM_ICDF(8315), AOM_ICDF(10607), AOM_ICDF(19333),
+          AOM_ICDF(21572), AOM_ICDF(22553), AOM_ICDF(25266), AOM_ICDF(27288),
+          AOM_ICDF(28551), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7168), AOM_ICDF(9668), AOM_ICDF(12646), AOM_ICDF(16257),
+          AOM_ICDF(19648), AOM_ICDF(20899), AOM_ICDF(25304), AOM_ICDF(30465),
+          AOM_ICDF(31625), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6144), AOM_ICDF(9368), AOM_ICDF(11836), AOM_ICDF(14130),
+          AOM_ICDF(19153), AOM_ICDF(21157), AOM_ICDF(24876), AOM_ICDF(28452),
+          AOM_ICDF(29396), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5504), AOM_ICDF(8486), AOM_ICDF(11996), AOM_ICDF(14412),
+          AOM_ICDF(17968), AOM_ICDF(21814), AOM_ICDF(24424), AOM_ICDF(30682),
+          AOM_ICDF(32059), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5376), AOM_ICDF(7195), AOM_ICDF(9592), AOM_ICDF(13331),
+          AOM_ICDF(17569), AOM_ICDF(19460), AOM_ICDF(22371), AOM_ICDF(25458),
+          AOM_ICDF(28942), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7424), AOM_ICDF(9206), AOM_ICDF(11783), AOM_ICDF(16456),
+          AOM_ICDF(19253), AOM_ICDF(20390), AOM_ICDF(23775), AOM_ICDF(27007),
+          AOM_ICDF(28425), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5888), AOM_ICDF(8303), AOM_ICDF(11361), AOM_ICDF(13440),
+          AOM_ICDF(15848), AOM_ICDF(17549), AOM_ICDF(21532), AOM_ICDF(29564),
+          AOM_ICDF(30665), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4608), AOM_ICDF(8788), AOM_ICDF(13284), AOM_ICDF(16621),
+          AOM_ICDF(18983), AOM_ICDF(20286), AOM_ICDF(24577), AOM_ICDF(28960),
+          AOM_ICDF(30314), AOM_ICDF(32768), 0 } },
+      { { AOM_ICDF(8320), AOM_ICDF(15005), AOM_ICDF(19168), AOM_ICDF(24282),
+          AOM_ICDF(26707), AOM_ICDF(27402), AOM_ICDF(28681), AOM_ICDF(29639),
+          AOM_ICDF(30629), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5632), AOM_ICDF(13900), AOM_ICDF(22376), AOM_ICDF(24867),
+          AOM_ICDF(26804), AOM_ICDF(27734), AOM_ICDF(29130), AOM_ICDF(30722),
+          AOM_ICDF(31465), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(4992), AOM_ICDF(9115), AOM_ICDF(11055), AOM_ICDF(24893),
+          AOM_ICDF(26316), AOM_ICDF(26661), AOM_ICDF(27663), AOM_ICDF(28301),
+          AOM_ICDF(29418), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(7424), AOM_ICDF(12077), AOM_ICDF(14987), AOM_ICDF(19596),
+          AOM_ICDF(22615), AOM_ICDF(23600), AOM_ICDF(26465), AOM_ICDF(28484),
+          AOM_ICDF(29789), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6272), AOM_ICDF(11447), AOM_ICDF(14362), AOM_ICDF(18204),
+          AOM_ICDF(23418), AOM_ICDF(24715), AOM_ICDF(26697), AOM_ICDF(28547),
+          AOM_ICDF(29520), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5248), AOM_ICDF(10946), AOM_ICDF(15379), AOM_ICDF(18167),
+          AOM_ICDF(22197), AOM_ICDF(25432), AOM_ICDF(27295), AOM_ICDF(30031),
+          AOM_ICDF(30576), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5120), AOM_ICDF(9008), AOM_ICDF(11607), AOM_ICDF(18210),
+          AOM_ICDF(22327), AOM_ICDF(23427), AOM_ICDF(24887), AOM_ICDF(26580),
+          AOM_ICDF(29892), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(6656), AOM_ICDF(10124), AOM_ICDF(12689), AOM_ICDF(19922),
+          AOM_ICDF(22480), AOM_ICDF(22807), AOM_ICDF(24441), AOM_ICDF(25579),
+          AOM_ICDF(26787), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5376), AOM_ICDF(10084), AOM_ICDF(13983), AOM_ICDF(17113),
+          AOM_ICDF(19996), AOM_ICDF(21614), AOM_ICDF(24403), AOM_ICDF(28651),
+          AOM_ICDF(29938), AOM_ICDF(32768), 0 },
+        { AOM_ICDF(5504), AOM_ICDF(14131), AOM_ICDF(17989), AOM_ICDF(23324),
+          AOM_ICDF(25513), AOM_ICDF(26071), AOM_ICDF(27850), AOM_ICDF(29464),
+          AOM_ICDF(30393), AOM_ICDF(32768), 0 } },
+#endif  // CONFIG_ALT_INTRA
+    };
+#endif  // CONFIG_EC_MULTISYMBOL
+
+static void init_mode_probs(FRAME_CONTEXT *fc) {
+  av1_copy(fc->uv_mode_prob, default_uv_probs);
+  av1_copy(fc->y_mode_prob, default_if_y_probs);
+  av1_copy(fc->switchable_interp_prob, default_switchable_interp_prob);
+  av1_copy(fc->partition_prob, default_partition_probs);
+  av1_copy(fc->intra_inter_prob, default_intra_inter_p);
+  av1_copy(fc->comp_inter_prob, default_comp_inter_p);
+  av1_copy(fc->comp_ref_prob, default_comp_ref_p);
+#if CONFIG_LV_MAP
+  av1_copy(fc->txb_skip, default_txb_skip);
+  av1_copy(fc->nz_map, default_nz_map);
+  av1_copy(fc->eob_flag, default_eob_flag);
+  av1_copy(fc->dc_sign, default_dc_sign);
+  av1_copy(fc->coeff_base, default_coeff_base);
+  av1_copy(fc->coeff_lps, default_coeff_lps);
+#endif
+#if CONFIG_EXT_REFS
+  av1_copy(fc->comp_bwdref_prob, default_comp_bwdref_p);
+#endif  // CONFIG_EXT_REFS
+  av1_copy(fc->single_ref_prob, default_single_ref_p);
+#if CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+  av1_copy(fc->comp_inter_mode_prob, default_comp_inter_mode_p);
+#endif  // CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+  av1_copy(fc->tx_size_probs, default_tx_size_prob);
+#if CONFIG_VAR_TX
+  av1_copy(fc->txfm_partition_prob, default_txfm_partition_probs);
+#endif
+  av1_copy(fc->skip_probs, default_skip_probs);
+#if CONFIG_REF_MV
+  av1_copy(fc->newmv_prob, default_newmv_prob);
+  av1_copy(fc->zeromv_prob, default_zeromv_prob);
+  av1_copy(fc->refmv_prob, default_refmv_prob);
+  av1_copy(fc->drl_prob, default_drl_prob);
+#endif  // CONFIG_REF_MV
+  av1_copy(fc->inter_mode_probs, default_inter_mode_probs);
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  av1_copy(fc->motion_mode_prob, default_motion_mode_prob);
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  av1_copy(fc->obmc_prob, default_obmc_prob);
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+#if CONFIG_EXT_INTER
+  av1_copy(fc->inter_compound_mode_probs, default_inter_compound_mode_probs);
+#if CONFIG_COMPOUND_SINGLEREF
+  av1_copy(fc->inter_singleref_comp_mode_probs,
+           default_inter_singleref_comp_mode_probs);
+#endif  // CONFIG_COMPOUND_SINGLEREF
+  av1_copy(fc->compound_type_prob, default_compound_type_probs);
+  av1_copy(fc->interintra_prob, default_interintra_prob);
+  av1_copy(fc->interintra_mode_prob, default_interintra_mode_prob);
+  av1_copy(fc->wedge_interintra_prob, default_wedge_interintra_prob);
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_SUPERTX
+  av1_copy(fc->supertx_prob, default_supertx_prob);
+#endif  // CONFIG_SUPERTX
+  av1_copy(fc->seg.tree_probs, default_segment_tree_probs);
+  av1_copy(fc->seg.pred_probs, default_segment_pred_probs);
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+  av1_copy(fc->intra_filter_probs, default_intra_filter_probs);
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  av1_copy(fc->filter_intra_probs, default_filter_intra_probs);
+#endif  // CONFIG_FILTER_INTRA
+  av1_copy(fc->inter_ext_tx_prob, default_inter_ext_tx_prob);
+  av1_copy(fc->intra_ext_tx_prob, default_intra_ext_tx_prob);
+#if CONFIG_LOOP_RESTORATION
+  av1_copy(fc->switchable_restore_prob, default_switchable_restore_prob);
+#endif  // CONFIG_LOOP_RESTORATION
+#if CONFIG_EC_MULTISYMBOL
+  av1_copy(fc->y_mode_cdf, default_if_y_mode_cdf);
+  av1_copy(fc->uv_mode_cdf, default_uv_mode_cdf);
+  av1_copy(fc->switchable_interp_cdf, default_switchable_interp_cdf);
+  av1_copy(fc->partition_cdf, default_partition_cdf);
+  av1_copy(fc->inter_mode_cdf, default_inter_mode_cdf);
+  av1_copy(fc->intra_ext_tx_cdf, default_intra_ext_tx_cdf);
+  av1_copy(fc->inter_ext_tx_cdf, default_inter_ext_tx_cdf);
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+  av1_copy(fc->intra_filter_cdf, default_intra_filter_cdf);
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+  av1_copy(fc->seg.tree_cdf, default_seg_tree_cdf);
+  av1_copy(fc->tx_size_cdf, default_tx_size_cdf);
+#endif  // CONFIG_EC_MULTISYMBOL
+#if CONFIG_DELTA_Q
+  av1_copy(fc->delta_q_prob, default_delta_q_probs);
+#if CONFIG_EC_MULTISYMBOL
+  av1_copy(fc->delta_q_cdf, default_delta_q_cdf);
+#endif  // CONFIG_EC_MULTISYMBOL
+#if CONFIG_EXT_DELTA_Q
+  av1_copy(fc->delta_lf_prob, default_delta_lf_probs);
+#if CONFIG_EC_MULTISYMBOL
+  av1_copy(fc->delta_lf_cdf, default_delta_lf_cdf);
+#endif  // CONFIG_EC_MULTISYMBOL
+#endif
+#endif  // CONFIG_DELTA_Q
+}
+
+#if CONFIG_EC_MULTISYMBOL
+int av1_switchable_interp_ind[SWITCHABLE_FILTERS];
+int av1_switchable_interp_inv[SWITCHABLE_FILTERS];
+
+#if !CONFIG_EC_ADAPT
+void av1_set_mode_cdfs(struct AV1Common *cm) {
+  FRAME_CONTEXT *fc = cm->fc;
+  int i, j;
+  if (cm->seg.enabled && cm->seg.update_map) {
+    av1_tree_to_cdf(av1_segment_tree, cm->fc->seg.tree_probs,
+                    cm->fc->seg.tree_cdf);
+  }
+
+  for (i = 0; i < INTRA_MODES; ++i)
+    av1_tree_to_cdf(av1_intra_mode_tree, fc->uv_mode_prob[i],
+                    fc->uv_mode_cdf[i]);
+#if CONFIG_EXT_PARTITION_TYPES
+  for (i = 0; i < PARTITION_PLOFFSET; ++i)
+    av1_tree_to_cdf(av1_partition_tree, fc->partition_prob[i],
+                    fc->partition_cdf[i]);
+  // Logical index (enum value) to inorder index (tree_to_cdf order)
+  aom_cdf_prob inorder_partition_cdf[CDF_SIZE(EXT_PARTITION_TYPES)] = {};
+  // TODO(aconverse): Generate this dynamically. The assumptions that
+  // av1_indices_from_tree() makes don't hold for this tree.
+  static const uint8_t av1_ext_partition_index_map[EXT_PARTITION_TYPES] = {
+    0, 1, 4, 7, 2, 3, 5, 6,
+  };
+  for (; i < PARTITION_CONTEXTS; ++i) {
+    av1_tree_to_cdf(av1_ext_partition_tree, fc->partition_prob[i],
+                    inorder_partition_cdf);
+    aom_cdf_prob cum_prob = 0;
+    for (j = 0; j < EXT_PARTITION_TYPES; ++j) {
+      int inorder_idx = av1_ext_partition_index_map[j];
+      aom_cdf_prob prob =
+          AOM_ICDF(inorder_partition_cdf[inorder_idx]) -
+          (inorder_idx > 0 ? AOM_ICDF(inorder_partition_cdf[inorder_idx - 1])
+                           : 0);
+      cum_prob += prob;
+      fc->partition_cdf[i][j] = AOM_ICDF(cum_prob);
+    }
+    assert(cum_prob == CDF_PROB_TOP);
+  }
+#else
+  for (i = 0; i < PARTITION_CONTEXTS; ++i)
+    av1_tree_to_cdf(av1_partition_tree, fc->partition_prob[i],
+                    fc->partition_cdf[i]);
+#endif
+
+  for (i = 0; i < INTRA_MODES; ++i)
+    for (j = 0; j < INTRA_MODES; ++j)
+      av1_tree_to_cdf(av1_intra_mode_tree, cm->kf_y_prob[i][j],
+                      cm->fc->kf_y_cdf[i][j]);
+
+  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
+    av1_tree_to_cdf(av1_switchable_interp_tree, fc->switchable_interp_prob[j],
+                    fc->switchable_interp_cdf[j]);
+
+  for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
+    av1_tree_to_cdf(av1_inter_mode_tree, fc->inter_mode_probs[i],
+                    fc->inter_mode_cdf[i]);
+
+  for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
+    av1_tree_to_cdf(av1_intra_mode_tree, fc->y_mode_prob[i], fc->y_mode_cdf[i]);
+
+#if CONFIG_EXT_TX
+  int s;
+  for (s = 0; s < EXT_TX_SETS_INTRA; ++s)
+    for (i = TX_4X4; i < EXT_TX_SIZES; ++i)
+      for (j = 0; j < INTRA_MODES; ++j)
+        av1_tree_to_cdf(av1_ext_tx_intra_tree[s],
+                        fc->intra_ext_tx_prob[s][i][j],
+                        fc->intra_ext_tx_cdf[s][i][j]);
+
+  for (s = 0; s < EXT_TX_SETS_INTER; ++s)
+    for (i = TX_4X4; i < EXT_TX_SIZES; ++i)
+      av1_tree_to_cdf(av1_ext_tx_inter_tree[s], fc->inter_ext_tx_prob[s][i],
+                      fc->inter_ext_tx_cdf[s][i]);
+#else
+  for (i = TX_4X4; i < EXT_TX_SIZES; ++i)
+    for (j = 0; j < TX_TYPES; ++j)
+      av1_tree_to_cdf(av1_ext_tx_tree, fc->intra_ext_tx_prob[i][j],
+                      fc->intra_ext_tx_cdf[i][j]);
+
+  for (i = TX_4X4; i < EXT_TX_SIZES; ++i)
+    av1_tree_to_cdf(av1_ext_tx_tree, fc->inter_ext_tx_prob[i],
+                    fc->inter_ext_tx_cdf[i]);
+#endif
+  for (i = 0; i < MAX_TX_DEPTH; i++) {
+    for (j = 0; j < TX_SIZE_CONTEXTS; j++) {
+      av1_tree_to_cdf(av1_tx_size_tree[i], fc->tx_size_probs[i][j],
+                      fc->tx_size_cdf[i][j]);
+    }
+  }
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+  for (i = 0; i < INTRA_FILTERS + 1; ++i) {
+    av1_tree_to_cdf(av1_intra_filter_tree, fc->intra_filter_probs[i],
+                    fc->intra_filter_cdf[i]);
+  }
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+}
+#endif  // !CONFIG_EC_ADAPT
+#endif  // CONFIG_EC_MULTISYMBOL
+
+#if CONFIG_DUAL_FILTER
+const aom_tree_index av1_switchable_interp_tree[TREE_SIZE(SWITCHABLE_FILTERS)] =
+    {
+      -EIGHTTAP_REGULAR, 2, 4, -MULTITAP_SHARP, -EIGHTTAP_SMOOTH,
+      -EIGHTTAP_SMOOTH2,
+    };
+#else
+const aom_tree_index av1_switchable_interp_tree[TREE_SIZE(SWITCHABLE_FILTERS)] =
+    { -EIGHTTAP_REGULAR, 2, -EIGHTTAP_SMOOTH, -MULTITAP_SHARP };
+#endif  // CONFIG_DUAL_FILTER
+
+void av1_adapt_inter_frame_probs(AV1_COMMON *cm) {
+  int i, j;
+  FRAME_CONTEXT *fc = cm->fc;
+  const FRAME_CONTEXT *pre_fc = &cm->frame_contexts[cm->frame_context_idx];
+  const FRAME_COUNTS *counts = &cm->counts;
+
+  for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+    fc->intra_inter_prob[i] = av1_mode_mv_merge_probs(
+        pre_fc->intra_inter_prob[i], counts->intra_inter[i]);
+
+  for (i = 0; i < COMP_INTER_CONTEXTS; i++)
+    fc->comp_inter_prob[i] = av1_mode_mv_merge_probs(pre_fc->comp_inter_prob[i],
+                                                     counts->comp_inter[i]);
+
+#if CONFIG_EXT_REFS
+  for (i = 0; i < REF_CONTEXTS; i++)
+    for (j = 0; j < (FWD_REFS - 1); j++)
+      fc->comp_ref_prob[i][j] = mode_mv_merge_probs(pre_fc->comp_ref_prob[i][j],
+                                                    counts->comp_ref[i][j]);
+  for (i = 0; i < REF_CONTEXTS; i++)
+    for (j = 0; j < (BWD_REFS - 1); j++)
+      fc->comp_bwdref_prob[i][j] = mode_mv_merge_probs(
+          pre_fc->comp_bwdref_prob[i][j], counts->comp_bwdref[i][j]);
+#else
+  for (i = 0; i < REF_CONTEXTS; i++)
+    for (j = 0; j < (COMP_REFS - 1); j++)
+      fc->comp_ref_prob[i][j] = mode_mv_merge_probs(pre_fc->comp_ref_prob[i][j],
+                                                    counts->comp_ref[i][j]);
+#endif  // CONFIG_EXT_REFS
+
+  for (i = 0; i < REF_CONTEXTS; i++)
+    for (j = 0; j < (SINGLE_REFS - 1); j++)
+      fc->single_ref_prob[i][j] = av1_mode_mv_merge_probs(
+          pre_fc->single_ref_prob[i][j], counts->single_ref[i][j]);
+
+#if CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+  for (i = 0; i < COMP_INTER_MODE_CONTEXTS; i++)
+    fc->comp_inter_mode_prob[i] = av1_mode_mv_merge_probs(
+        pre_fc->comp_inter_mode_prob[i], counts->comp_inter_mode[i]);
+
+#endif  // CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+
+#if CONFIG_REF_MV
+  for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i)
+    fc->newmv_prob[i] =
+        av1_mode_mv_merge_probs(pre_fc->newmv_prob[i], counts->newmv_mode[i]);
+  for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i)
+    fc->zeromv_prob[i] =
+        av1_mode_mv_merge_probs(pre_fc->zeromv_prob[i], counts->zeromv_mode[i]);
+  for (i = 0; i < REFMV_MODE_CONTEXTS; ++i)
+    fc->refmv_prob[i] =
+        av1_mode_mv_merge_probs(pre_fc->refmv_prob[i], counts->refmv_mode[i]);
+
+  for (i = 0; i < DRL_MODE_CONTEXTS; ++i)
+    fc->drl_prob[i] =
+        av1_mode_mv_merge_probs(pre_fc->drl_prob[i], counts->drl_mode[i]);
+#else
+  for (i = 0; i < INTER_MODE_CONTEXTS; i++)
+    aom_tree_merge_probs(av1_inter_mode_tree, pre_fc->inter_mode_probs[i],
+                         counts->inter_mode[i], fc->inter_mode_probs[i]);
+#endif
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  for (i = BLOCK_8X8; i < BLOCK_SIZES; ++i)
+    aom_tree_merge_probs(av1_motion_mode_tree, pre_fc->motion_mode_prob[i],
+                         counts->motion_mode[i], fc->motion_mode_prob[i]);
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  for (i = BLOCK_8X8; i < BLOCK_SIZES; ++i)
+    fc->obmc_prob[i] =
+        av1_mode_mv_merge_probs(pre_fc->obmc_prob[i], counts->obmc[i]);
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_SUPERTX
+  for (i = 0; i < PARTITION_SUPERTX_CONTEXTS; ++i) {
+    for (j = TX_8X8; j < TX_SIZES; ++j) {
+      fc->supertx_prob[i][j] = av1_mode_mv_merge_probs(
+          pre_fc->supertx_prob[i][j], counts->supertx[i][j]);
+    }
+  }
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_EXT_INTER
+  for (i = 0; i < INTER_MODE_CONTEXTS; i++)
+    aom_tree_merge_probs(
+        av1_inter_compound_mode_tree, pre_fc->inter_compound_mode_probs[i],
+        counts->inter_compound_mode[i], fc->inter_compound_mode_probs[i]);
+#if CONFIG_COMPOUND_SINGLEREF
+  for (i = 0; i < INTER_MODE_CONTEXTS; i++)
+    aom_tree_merge_probs(av1_inter_singleref_comp_mode_tree,
+                         pre_fc->inter_singleref_comp_mode_probs[i],
+                         counts->inter_singleref_comp_mode[i],
+                         fc->inter_singleref_comp_mode_probs[i]);
+#endif  // CONFIG_COMPOUND_SINGLEREF
+  for (i = 0; i < BLOCK_SIZE_GROUPS; ++i) {
+    if (is_interintra_allowed_bsize_group(i))
+      fc->interintra_prob[i] = av1_mode_mv_merge_probs(
+          pre_fc->interintra_prob[i], counts->interintra[i]);
+  }
+  for (i = 0; i < BLOCK_SIZE_GROUPS; i++) {
+    aom_tree_merge_probs(
+        av1_interintra_mode_tree, pre_fc->interintra_mode_prob[i],
+        counts->interintra_mode[i], fc->interintra_mode_prob[i]);
+  }
+  for (i = 0; i < BLOCK_SIZES; ++i) {
+    if (is_interintra_allowed_bsize(i) && is_interintra_wedge_used(i))
+      fc->wedge_interintra_prob[i] = av1_mode_mv_merge_probs(
+          pre_fc->wedge_interintra_prob[i], counts->wedge_interintra[i]);
+  }
+
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+  for (i = 0; i < BLOCK_SIZES; ++i) {
+    aom_tree_merge_probs(av1_compound_type_tree, pre_fc->compound_type_prob[i],
+                         counts->compound_interinter[i],
+                         fc->compound_type_prob[i]);
+  }
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#endif  // CONFIG_EXT_INTER
+
+  for (i = 0; i < BLOCK_SIZE_GROUPS; i++)
+    aom_tree_merge_probs(av1_intra_mode_tree, pre_fc->y_mode_prob[i],
+                         counts->y_mode[i], fc->y_mode_prob[i]);
+
+  if (cm->interp_filter == SWITCHABLE) {
+    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
+      aom_tree_merge_probs(
+          av1_switchable_interp_tree, pre_fc->switchable_interp_prob[i],
+          counts->switchable_interp[i], fc->switchable_interp_prob[i]);
+  }
+}
+
+void av1_adapt_intra_frame_probs(AV1_COMMON *cm) {
+  int i, j;
+  FRAME_CONTEXT *fc = cm->fc;
+  const FRAME_CONTEXT *pre_fc = &cm->frame_contexts[cm->frame_context_idx];
+  const FRAME_COUNTS *counts = &cm->counts;
+
+  if (cm->tx_mode == TX_MODE_SELECT) {
+    for (i = 0; i < MAX_TX_DEPTH; ++i) {
+      for (j = 0; j < TX_SIZE_CONTEXTS; ++j)
+        aom_tree_merge_probs(av1_tx_size_tree[i], pre_fc->tx_size_probs[i][j],
+                             counts->tx_size[i][j], fc->tx_size_probs[i][j]);
+    }
+  }
+
+#if CONFIG_VAR_TX
+  if (cm->tx_mode == TX_MODE_SELECT) {
+    for (i = 0; i < TXFM_PARTITION_CONTEXTS; ++i)
+      fc->txfm_partition_prob[i] = av1_mode_mv_merge_probs(
+          pre_fc->txfm_partition_prob[i], counts->txfm_partition[i]);
+  }
+#endif
+
+  for (i = 0; i < SKIP_CONTEXTS; ++i)
+    fc->skip_probs[i] =
+        av1_mode_mv_merge_probs(pre_fc->skip_probs[i], counts->skip[i]);
+
+#if CONFIG_EXT_TX
+  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+    int s;
+    for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
+      if (use_inter_ext_tx_for_txsize[s][i]) {
+        aom_tree_merge_probs(
+            av1_ext_tx_inter_tree[s], pre_fc->inter_ext_tx_prob[s][i],
+            counts->inter_ext_tx[s][i], fc->inter_ext_tx_prob[s][i]);
+      }
+    }
+    for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
+      if (use_intra_ext_tx_for_txsize[s][i]) {
+        for (j = 0; j < INTRA_MODES; ++j)
+          aom_tree_merge_probs(
+              av1_ext_tx_intra_tree[s], pre_fc->intra_ext_tx_prob[s][i][j],
+              counts->intra_ext_tx[s][i][j], fc->intra_ext_tx_prob[s][i][j]);
+      }
+    }
+  }
+#else
+  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+    for (j = 0; j < TX_TYPES; ++j) {
+      aom_tree_merge_probs(av1_ext_tx_tree, pre_fc->intra_ext_tx_prob[i][j],
+                           counts->intra_ext_tx[i][j],
+                           fc->intra_ext_tx_prob[i][j]);
+    }
+  }
+  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+    aom_tree_merge_probs(av1_ext_tx_tree, pre_fc->inter_ext_tx_prob[i],
+                         counts->inter_ext_tx[i], fc->inter_ext_tx_prob[i]);
+  }
+#endif  // CONFIG_EXT_TX
+
+  if (cm->seg.temporal_update) {
+    for (i = 0; i < PREDICTION_PROBS; i++)
+      fc->seg.pred_probs[i] = av1_mode_mv_merge_probs(pre_fc->seg.pred_probs[i],
+                                                      counts->seg.pred[i]);
+
+    aom_tree_merge_probs(av1_segment_tree, pre_fc->seg.tree_probs,
+                         counts->seg.tree_mispred, fc->seg.tree_probs);
+  } else {
+    aom_tree_merge_probs(av1_segment_tree, pre_fc->seg.tree_probs,
+                         counts->seg.tree_total, fc->seg.tree_probs);
+  }
+
+  for (i = 0; i < INTRA_MODES; ++i)
+    aom_tree_merge_probs(av1_intra_mode_tree, pre_fc->uv_mode_prob[i],
+                         counts->uv_mode[i], fc->uv_mode_prob[i]);
+
+#if CONFIG_EXT_PARTITION_TYPES
+  for (i = 0; i < PARTITION_PLOFFSET; ++i)
+    aom_tree_merge_probs(av1_partition_tree, pre_fc->partition_prob[i],
+                         counts->partition[i], fc->partition_prob[i]);
+  for (; i < PARTITION_CONTEXTS_PRIMARY; ++i)
+    aom_tree_merge_probs(av1_ext_partition_tree, pre_fc->partition_prob[i],
+                         counts->partition[i], fc->partition_prob[i]);
+#else
+  for (i = 0; i < PARTITION_CONTEXTS_PRIMARY; ++i) {
+    aom_tree_merge_probs(av1_partition_tree, pre_fc->partition_prob[i],
+                         counts->partition[i], fc->partition_prob[i]);
+  }
+#endif  // CONFIG_EXT_PARTITION_TYPES
+#if CONFIG_UNPOISON_PARTITION_CTX
+  for (i = PARTITION_CONTEXTS_PRIMARY;
+       i < PARTITION_CONTEXTS_PRIMARY + PARTITION_BLOCK_SIZES; ++i) {
+    unsigned int ct[2] = { counts->partition[i][PARTITION_VERT],
+                           counts->partition[i][PARTITION_SPLIT] };
+    assert(counts->partition[i][PARTITION_NONE] == 0);
+    assert(counts->partition[i][PARTITION_HORZ] == 0);
+    assert(fc->partition_prob[i][PARTITION_NONE] == 0);
+    assert(fc->partition_prob[i][PARTITION_HORZ] == 0);
+    fc->partition_prob[i][PARTITION_VERT] =
+        av1_mode_mv_merge_probs(pre_fc->partition_prob[i][PARTITION_VERT], ct);
+  }
+  for (i = PARTITION_CONTEXTS_PRIMARY + PARTITION_BLOCK_SIZES;
+       i < PARTITION_CONTEXTS_PRIMARY + 2 * PARTITION_BLOCK_SIZES; ++i) {
+    unsigned int ct[2] = { counts->partition[i][PARTITION_HORZ],
+                           counts->partition[i][PARTITION_SPLIT] };
+    assert(counts->partition[i][PARTITION_NONE] == 0);
+    assert(counts->partition[i][PARTITION_VERT] == 0);
+    assert(fc->partition_prob[i][PARTITION_NONE] == 0);
+    assert(fc->partition_prob[i][PARTITION_VERT] == 0);
+    fc->partition_prob[i][PARTITION_HORZ] =
+        av1_mode_mv_merge_probs(pre_fc->partition_prob[i][PARTITION_HORZ], ct);
+  }
+#endif
+#if CONFIG_DELTA_Q
+  for (i = 0; i < DELTA_Q_PROBS; ++i)
+    fc->delta_q_prob[i] =
+        mode_mv_merge_probs(pre_fc->delta_q_prob[i], counts->delta_q[i]);
+#if CONFIG_EXT_DELTA_Q
+  for (i = 0; i < DELTA_LF_PROBS; ++i)
+    fc->delta_lf_prob[i] =
+        mode_mv_merge_probs(pre_fc->delta_lf_prob[i], counts->delta_lf[i]);
+#endif  // CONFIG_EXT_DELTA_Q
+#endif
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+  for (i = 0; i < INTRA_FILTERS + 1; ++i) {
+    aom_tree_merge_probs(av1_intra_filter_tree, pre_fc->intra_filter_probs[i],
+                         counts->intra_filter[i], fc->intra_filter_probs[i]);
+  }
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  for (i = 0; i < PLANE_TYPES; ++i) {
+    fc->filter_intra_probs[i] = av1_mode_mv_merge_probs(
+        pre_fc->filter_intra_probs[i], counts->filter_intra[i]);
+  }
+#endif  // CONFIG_FILTER_INTRA
+}
+
+static void set_default_lf_deltas(struct loopfilter *lf) {
+  lf->mode_ref_delta_enabled = 1;
+  lf->mode_ref_delta_update = 1;
+
+  lf->ref_deltas[INTRA_FRAME] = 1;
+  lf->ref_deltas[LAST_FRAME] = 0;
+#if CONFIG_EXT_REFS
+  lf->ref_deltas[LAST2_FRAME] = lf->ref_deltas[LAST_FRAME];
+  lf->ref_deltas[LAST3_FRAME] = lf->ref_deltas[LAST_FRAME];
+  lf->ref_deltas[BWDREF_FRAME] = lf->ref_deltas[LAST_FRAME];
+#endif  // CONFIG_EXT_REFS
+  lf->ref_deltas[GOLDEN_FRAME] = -1;
+  lf->ref_deltas[ALTREF_FRAME] = -1;
+
+  lf->mode_deltas[0] = 0;
+  lf->mode_deltas[1] = 0;
+}
+
+void av1_setup_past_independence(AV1_COMMON *cm) {
+  // Reset the segment feature data to the default stats:
+  // Features disabled, 0, with delta coding (Default state).
+  struct loopfilter *const lf = &cm->lf;
+
+  int i;
+  av1_clearall_segfeatures(&cm->seg);
+  cm->seg.abs_delta = SEGMENT_DELTADATA;
+
+  if (cm->last_frame_seg_map && !cm->frame_parallel_decode)
+    memset(cm->last_frame_seg_map, 0, (cm->mi_rows * cm->mi_cols));
+
+  if (cm->current_frame_seg_map)
+    memset(cm->current_frame_seg_map, 0, (cm->mi_rows * cm->mi_cols));
+
+  // Reset the mode ref deltas for loop filter
+  av1_zero(lf->last_ref_deltas);
+  av1_zero(lf->last_mode_deltas);
+  set_default_lf_deltas(lf);
+
+  // To force update of the sharpness
+  lf->last_sharpness_level = -1;
+
+  av1_default_coef_probs(cm);
+  init_mode_probs(cm->fc);
+  av1_init_mv_probs(cm);
+#if CONFIG_PVQ
+  av1_default_pvq_probs(cm);
+#endif  // CONFIG_PVQ
+#if CONFIG_ADAPT_SCAN
+  av1_init_scan_order(cm);
+#endif
+  av1_convolve_init(cm);
+  cm->fc->initialized = 1;
+
+  if (cm->frame_type == KEY_FRAME || cm->error_resilient_mode ||
+      cm->reset_frame_context == RESET_FRAME_CONTEXT_ALL) {
+    // Reset all frame contexts.
+    for (i = 0; i < FRAME_CONTEXTS; ++i) cm->frame_contexts[i] = *cm->fc;
+  } else if (cm->reset_frame_context == RESET_FRAME_CONTEXT_CURRENT) {
+    // Reset only the frame context specified in the frame header.
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+  }
+
+  // prev_mip will only be allocated in encoder.
+  if (frame_is_intra_only(cm) && cm->prev_mip && !cm->frame_parallel_decode)
+    memset(cm->prev_mip, 0,
+           cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->prev_mip));
+
+  cm->frame_context_idx = 0;
+}
diff --git a/third_party/aom/av1/common/entropymode.h b/third_party/aom/av1/common/entropymode.h
new file mode 100644
index 0000000000..9c3a78d611
--- /dev/null
+++ b/third_party/aom/av1/common/entropymode.h
@@ -0,0 +1,575 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_ENTROPYMODE_H_
+#define AV1_COMMON_ENTROPYMODE_H_
+
+#include "av1/common/entropy.h"
+#include "av1/common/entropymv.h"
+#include "av1/common/filter.h"
+#include "av1/common/seg_common.h"
+#include "aom_dsp/aom_filter.h"
+
+#if CONFIG_PVQ
+#include "av1/common/pvq.h"
+#include "av1/common/pvq_state.h"
+#include "av1/common/generic_code.h"
+#endif  // CONFIG_PVQ
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define BLOCK_SIZE_GROUPS 4
+
+#define TX_SIZE_CONTEXTS 2
+
+#define INTER_OFFSET(mode) ((mode)-NEARESTMV)
+#if CONFIG_EXT_INTER
+#if CONFIG_COMPOUND_SINGLEREF
+#define INTER_SINGLEREF_COMP_OFFSET(mode) ((mode)-SR_NEAREST_NEARMV)
+#endif  // CONFIG_COMPOUND_SINGLEREF
+#define INTER_COMPOUND_OFFSET(mode) ((mode)-NEAREST_NEARESTMV)
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_PALETTE
+// Number of possible contexts for a color index.
+// As can be seen from av1_get_palette_color_index_context(), the possible
+// contexts are (2,0,0), (2,2,1), (3,2,0), (4,1,0), (5,0,0). These are mapped to
+// a value from 0 to 4 using 'palette_color_index_context_lookup' table.
+#define PALETTE_COLOR_INDEX_CONTEXTS 5
+
+// Maximum number of colors in a palette.
+#define PALETTE_MAX_SIZE 8
+// Minimum number of colors in a palette.
+#define PALETTE_MIN_SIZE 2
+
+// Palette mode is available for block sizes >= 8x8.
+#define PALETTE_BLOCK_SIZES (BLOCK_LARGEST - BLOCK_8X8 + 1)
+
+// Palette Y mode context for a block is determined by number of neighboring
+// blocks (top and/or left) using a palette for Y plane. So, possible Y mode'
+// context values are:
+// 0 if neither left nor top block uses palette for Y plane,
+// 1 if exactly one of left or top block uses palette for Y plane, and
+// 2 if both left and top blocks use palette for Y plane.
+#define PALETTE_Y_MODE_CONTEXTS 3
+
+// Palette UV mode context for a block is determined by whether this block uses
+// palette for the Y plane. So, possible values are:
+// 0 if this block doesn't use palette for Y plane.
+// 1 if this block uses palette for Y plane (i.e. Y palette size > 0).
+#define PALETTE_UV_MODE_CONTEXTS 2
+
+#define PALETTE_MAX_BLOCK_SIZE (64 * 64)
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_INTRABC
+#define INTRABC_PROB 192
+#endif  // CONFIG_INTRABC
+
+struct AV1Common;
+
+typedef struct {
+  const int16_t *scan;
+  const int16_t *iscan;
+  const int16_t *neighbors;
+} SCAN_ORDER;
+
+struct seg_counts {
+  unsigned int tree_total[MAX_SEGMENTS];
+  unsigned int tree_mispred[MAX_SEGMENTS];
+  unsigned int pred[PREDICTION_PROBS][2];
+};
+
+typedef struct frame_contexts {
+  aom_prob y_mode_prob[BLOCK_SIZE_GROUPS][INTRA_MODES - 1];
+  aom_prob uv_mode_prob[INTRA_MODES][INTRA_MODES - 1];
+#if CONFIG_EXT_PARTITION_TYPES
+  aom_prob partition_prob[PARTITION_CONTEXTS][EXT_PARTITION_TYPES - 1];
+#else
+  aom_prob partition_prob[PARTITION_CONTEXTS][PARTITION_TYPES - 1];
+#endif
+  av1_coeff_probs_model coef_probs[TX_SIZES][PLANE_TYPES];
+#if CONFIG_NEW_TOKENSET
+  coeff_cdf_model coef_tail_cdfs[TX_SIZES][PLANE_TYPES];
+  coeff_cdf_model coef_head_cdfs[TX_SIZES][PLANE_TYPES];
+  aom_prob blockzero_probs[TX_SIZES][PLANE_TYPES][REF_TYPES][BLOCKZ_CONTEXTS];
+#elif CONFIG_EC_MULTISYMBOL
+  coeff_cdf_model coef_cdfs[TX_SIZES][PLANE_TYPES];
+#endif  // CONFIG_NEW_TOKENSET
+  aom_prob switchable_interp_prob[SWITCHABLE_FILTER_CONTEXTS]
+                                 [SWITCHABLE_FILTERS - 1];
+#if CONFIG_ADAPT_SCAN
+// TODO(angiebird): try aom_prob
+#if CONFIG_CB4X4
+  uint32_t non_zero_prob_2x2[TX_TYPES][4];
+#endif
+  uint32_t non_zero_prob_4X4[TX_TYPES][16];
+  uint32_t non_zero_prob_8X8[TX_TYPES][64];
+  uint32_t non_zero_prob_16X16[TX_TYPES][256];
+  uint32_t non_zero_prob_32X32[TX_TYPES][1024];
+
+  uint32_t non_zero_prob_4X8[TX_TYPES][32];
+  uint32_t non_zero_prob_8X4[TX_TYPES][32];
+  uint32_t non_zero_prob_16X8[TX_TYPES][128];
+  uint32_t non_zero_prob_8X16[TX_TYPES][128];
+  uint32_t non_zero_prob_32X16[TX_TYPES][512];
+  uint32_t non_zero_prob_16X32[TX_TYPES][512];
+
+#if CONFIG_CB4X4
+  DECLARE_ALIGNED(16, int16_t, scan_2x2[TX_TYPES][4]);
+#endif
+  DECLARE_ALIGNED(16, int16_t, scan_4X4[TX_TYPES][16]);
+  DECLARE_ALIGNED(16, int16_t, scan_8X8[TX_TYPES][64]);
+  DECLARE_ALIGNED(16, int16_t, scan_16X16[TX_TYPES][256]);
+  DECLARE_ALIGNED(16, int16_t, scan_32X32[TX_TYPES][1024]);
+
+  DECLARE_ALIGNED(16, int16_t, scan_4X8[TX_TYPES][32]);
+  DECLARE_ALIGNED(16, int16_t, scan_8X4[TX_TYPES][32]);
+  DECLARE_ALIGNED(16, int16_t, scan_8X16[TX_TYPES][128]);
+  DECLARE_ALIGNED(16, int16_t, scan_16X8[TX_TYPES][128]);
+  DECLARE_ALIGNED(16, int16_t, scan_16X32[TX_TYPES][512]);
+  DECLARE_ALIGNED(16, int16_t, scan_32X16[TX_TYPES][512]);
+
+#if CONFIG_CB4X4
+  DECLARE_ALIGNED(16, int16_t, iscan_2x2[TX_TYPES][4]);
+#endif
+  DECLARE_ALIGNED(16, int16_t, iscan_4X4[TX_TYPES][16]);
+  DECLARE_ALIGNED(16, int16_t, iscan_8X8[TX_TYPES][64]);
+  DECLARE_ALIGNED(16, int16_t, iscan_16X16[TX_TYPES][256]);
+  DECLARE_ALIGNED(16, int16_t, iscan_32X32[TX_TYPES][1024]);
+
+  DECLARE_ALIGNED(16, int16_t, iscan_4X8[TX_TYPES][32]);
+  DECLARE_ALIGNED(16, int16_t, iscan_8X4[TX_TYPES][32]);
+  DECLARE_ALIGNED(16, int16_t, iscan_8X16[TX_TYPES][128]);
+  DECLARE_ALIGNED(16, int16_t, iscan_16X8[TX_TYPES][128]);
+  DECLARE_ALIGNED(16, int16_t, iscan_16X32[TX_TYPES][512]);
+  DECLARE_ALIGNED(16, int16_t, iscan_32X16[TX_TYPES][512]);
+
+#if CONFIG_CB4X4
+  int16_t nb_2x2[TX_TYPES][(4 + 1) * 2];
+#endif
+  int16_t nb_4X4[TX_TYPES][(16 + 1) * 2];
+  int16_t nb_8X8[TX_TYPES][(64 + 1) * 2];
+  int16_t nb_16X16[TX_TYPES][(256 + 1) * 2];
+  int16_t nb_32X32[TX_TYPES][(1024 + 1) * 2];
+
+  int16_t nb_4X8[TX_TYPES][(32 + 1) * 2];
+  int16_t nb_8X4[TX_TYPES][(32 + 1) * 2];
+  int16_t nb_8X16[TX_TYPES][(128 + 1) * 2];
+  int16_t nb_16X8[TX_TYPES][(128 + 1) * 2];
+  int16_t nb_16X32[TX_TYPES][(512 + 1) * 2];
+  int16_t nb_32X16[TX_TYPES][(512 + 1) * 2];
+
+  SCAN_ORDER sc[TX_SIZES_ALL][TX_TYPES];
+
+  int16_t eob_threshold[TX_SIZES_ALL][TX_TYPES][EOB_THRESHOLD_NUM];
+#endif  // CONFIG_ADAPT_SCAN
+
+#if CONFIG_LV_MAP
+  aom_prob txb_skip[TX_SIZES][TXB_SKIP_CONTEXTS];
+  aom_prob nz_map[TX_SIZES][PLANE_TYPES][SIG_COEF_CONTEXTS];
+  aom_prob eob_flag[TX_SIZES][PLANE_TYPES][EOB_COEF_CONTEXTS];
+  aom_prob dc_sign[PLANE_TYPES][DC_SIGN_CONTEXTS];
+  aom_prob coeff_base[TX_SIZES][PLANE_TYPES][NUM_BASE_LEVELS]
+                     [COEFF_BASE_CONTEXTS];
+  aom_prob coeff_lps[TX_SIZES][PLANE_TYPES][LEVEL_CONTEXTS];
+#endif
+
+#if CONFIG_REF_MV
+  aom_prob newmv_prob[NEWMV_MODE_CONTEXTS];
+  aom_prob zeromv_prob[ZEROMV_MODE_CONTEXTS];
+  aom_prob refmv_prob[REFMV_MODE_CONTEXTS];
+  aom_prob drl_prob[DRL_MODE_CONTEXTS];
+#endif  // CONFIG_REF_MV
+
+  aom_prob inter_mode_probs[INTER_MODE_CONTEXTS][INTER_MODES - 1];
+#if CONFIG_EXT_INTER
+  aom_prob inter_compound_mode_probs[INTER_MODE_CONTEXTS]
+                                    [INTER_COMPOUND_MODES - 1];
+#if CONFIG_COMPOUND_SINGLEREF
+  aom_prob inter_singleref_comp_mode_probs[INTER_MODE_CONTEXTS]
+                                          [INTER_SINGLEREF_COMP_MODES - 1];
+#endif  // CONFIG_COMPOUND_SINGLEREF
+  aom_prob compound_type_prob[BLOCK_SIZES][COMPOUND_TYPES - 1];
+  aom_prob interintra_prob[BLOCK_SIZE_GROUPS];
+  aom_prob interintra_mode_prob[BLOCK_SIZE_GROUPS][INTERINTRA_MODES - 1];
+  aom_prob wedge_interintra_prob[BLOCK_SIZES];
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  aom_prob motion_mode_prob[BLOCK_SIZES][MOTION_MODES - 1];
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  aom_prob obmc_prob[BLOCK_SIZES];
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  aom_prob intra_inter_prob[INTRA_INTER_CONTEXTS];
+  aom_prob comp_inter_prob[COMP_INTER_CONTEXTS];
+  aom_prob single_ref_prob[REF_CONTEXTS][SINGLE_REFS - 1];
+#if CONFIG_EXT_REFS
+  aom_prob comp_ref_prob[REF_CONTEXTS][FWD_REFS - 1];
+  aom_prob comp_bwdref_prob[REF_CONTEXTS][BWD_REFS - 1];
+#else
+  aom_prob comp_ref_prob[REF_CONTEXTS][COMP_REFS - 1];
+#endif  // CONFIG_EXT_REFS
+#if CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+  aom_prob comp_inter_mode_prob[COMP_INTER_MODE_CONTEXTS];
+#endif  // CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+  aom_prob tx_size_probs[MAX_TX_DEPTH][TX_SIZE_CONTEXTS][MAX_TX_DEPTH];
+#if CONFIG_VAR_TX
+  aom_prob txfm_partition_prob[TXFM_PARTITION_CONTEXTS];
+#endif
+  aom_prob skip_probs[SKIP_CONTEXTS];
+#if CONFIG_REF_MV
+  nmv_context nmvc[NMV_CONTEXTS];
+#else
+  nmv_context nmvc;
+#endif
+#if CONFIG_INTRABC
+  nmv_context ndvc;
+#endif
+  int initialized;
+#if CONFIG_EXT_TX
+  aom_prob inter_ext_tx_prob[EXT_TX_SETS_INTER][EXT_TX_SIZES][TX_TYPES - 1];
+  aom_prob intra_ext_tx_prob[EXT_TX_SETS_INTRA][EXT_TX_SIZES][INTRA_MODES]
+                            [TX_TYPES - 1];
+#else
+  aom_prob intra_ext_tx_prob[EXT_TX_SIZES][TX_TYPES][TX_TYPES - 1];
+  aom_prob inter_ext_tx_prob[EXT_TX_SIZES][TX_TYPES - 1];
+#endif  // CONFIG_EXT_TX
+#if CONFIG_SUPERTX
+  aom_prob supertx_prob[PARTITION_SUPERTX_CONTEXTS][TX_SIZES];
+#endif  // CONFIG_SUPERTX
+  struct segmentation_probs seg;
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+  aom_prob intra_filter_probs[INTRA_FILTERS + 1][INTRA_FILTERS - 1];
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  aom_prob filter_intra_probs[PLANE_TYPES];
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_GLOBAL_MOTION
+  aom_prob global_motion_types_prob[GLOBAL_TRANS_TYPES - 1];
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_LOOP_RESTORATION
+  aom_prob switchable_restore_prob[RESTORE_SWITCHABLE_TYPES - 1];
+#endif  // CONFIG_LOOP_RESTORATION
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob y_mode_cdf[BLOCK_SIZE_GROUPS][CDF_SIZE(INTRA_MODES)];
+  aom_cdf_prob uv_mode_cdf[INTRA_MODES][CDF_SIZE(INTRA_MODES)];
+#if CONFIG_EXT_PARTITION_TYPES
+  aom_cdf_prob partition_cdf[PARTITION_CONTEXTS][CDF_SIZE(EXT_PARTITION_TYPES)];
+#else
+  aom_cdf_prob partition_cdf[PARTITION_CONTEXTS][CDF_SIZE(PARTITION_TYPES)];
+#endif
+  aom_cdf_prob switchable_interp_cdf[SWITCHABLE_FILTER_CONTEXTS]
+                                    [CDF_SIZE(SWITCHABLE_FILTERS)];
+  aom_cdf_prob inter_mode_cdf[INTER_MODE_CONTEXTS][CDF_SIZE(INTER_MODES)];
+  /* Keep track of kf_y_cdf here, as this makes handling
+     multiple copies for adaptation in tiles easier */
+  aom_cdf_prob kf_y_cdf[INTRA_MODES][INTRA_MODES][CDF_SIZE(INTRA_MODES)];
+  aom_cdf_prob tx_size_cdf[MAX_TX_DEPTH][TX_SIZE_CONTEXTS]
+                          [CDF_SIZE(MAX_TX_DEPTH + 1)];
+#if CONFIG_DELTA_Q
+  aom_cdf_prob delta_q_cdf[CDF_SIZE(DELTA_Q_PROBS + 1)];
+#if CONFIG_EXT_DELTA_Q
+  aom_cdf_prob delta_lf_cdf[CDF_SIZE(DELTA_LF_PROBS + 1)];
+#endif
+#endif  // CONFIG_DELTA_Q
+#if CONFIG_EXT_TX
+  aom_cdf_prob intra_ext_tx_cdf[EXT_TX_SETS_INTRA][EXT_TX_SIZES][INTRA_MODES]
+                               [CDF_SIZE(TX_TYPES)];
+  aom_cdf_prob inter_ext_tx_cdf[EXT_TX_SETS_INTER][EXT_TX_SIZES]
+                               [CDF_SIZE(TX_TYPES)];
+#else
+  aom_cdf_prob intra_ext_tx_cdf[EXT_TX_SIZES][TX_TYPES][CDF_SIZE(TX_TYPES)];
+  aom_cdf_prob inter_ext_tx_cdf[EXT_TX_SIZES][CDF_SIZE(TX_TYPES)];
+#endif  // CONFIG_EXT_TX
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+  aom_cdf_prob intra_filter_cdf[INTRA_FILTERS + 1][CDF_SIZE(INTRA_FILTERS)];
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+#endif  // CONFIG_EC_MULTISYMBOL
+#if CONFIG_DELTA_Q
+  aom_prob delta_q_prob[DELTA_Q_PROBS];
+#if CONFIG_EXT_DELTA_Q
+  aom_prob delta_lf_prob[DELTA_LF_PROBS];
+#endif
+#endif
+#if CONFIG_PVQ
+  // TODO(any): If PVQ is enabled, most of coefficient related cdf,
+  // such as coef_cdfs[], coef_tail_cdfs[], and coef_heaf_cdfs[] can be removed.
+  od_adapt_ctx pvq_context;
+#endif  // CONFIG_PVQ
+} FRAME_CONTEXT;
+
+typedef struct FRAME_COUNTS {
+  // Note: This structure should only contain 'unsigned int' fields, or
+  // aggregates built solely from 'unsigned int' fields/elements
+  unsigned int kf_y_mode[INTRA_MODES][INTRA_MODES][INTRA_MODES];
+  unsigned int y_mode[BLOCK_SIZE_GROUPS][INTRA_MODES];
+  unsigned int uv_mode[INTRA_MODES][INTRA_MODES];
+#if CONFIG_EXT_PARTITION_TYPES
+  unsigned int partition[PARTITION_CONTEXTS][EXT_PARTITION_TYPES];
+#else
+  unsigned int partition[PARTITION_CONTEXTS][PARTITION_TYPES];
+#endif
+  av1_coeff_count_model coef[TX_SIZES][PLANE_TYPES];
+  unsigned int eob_branch[TX_SIZES][PLANE_TYPES][REF_TYPES][COEF_BANDS]
+                         [COEFF_CONTEXTS];
+  unsigned int switchable_interp[SWITCHABLE_FILTER_CONTEXTS]
+                                [SWITCHABLE_FILTERS];
+#if CONFIG_ADAPT_SCAN
+#if CONFIG_CB4X4
+  unsigned int non_zero_count_2x2[TX_TYPES][4];
+#endif  // CONFIG_CB4X4
+  unsigned int non_zero_count_4X4[TX_TYPES][16];
+  unsigned int non_zero_count_8X8[TX_TYPES][64];
+  unsigned int non_zero_count_16X16[TX_TYPES][256];
+  unsigned int non_zero_count_32X32[TX_TYPES][1024];
+
+  unsigned int non_zero_count_4x8[TX_TYPES][32];
+  unsigned int non_zero_count_8x4[TX_TYPES][32];
+  unsigned int non_zero_count_8x16[TX_TYPES][128];
+  unsigned int non_zero_count_16x8[TX_TYPES][128];
+  unsigned int non_zero_count_16x32[TX_TYPES][512];
+  unsigned int non_zero_count_32x16[TX_TYPES][512];
+
+  unsigned int txb_count[TX_SIZES_ALL][TX_TYPES];
+#endif  // CONFIG_ADAPT_SCAN
+
+#if CONFIG_LV_MAP
+  unsigned int txb_skip[TX_SIZES][TXB_SKIP_CONTEXTS][2];
+  unsigned int nz_map[TX_SIZES][PLANE_TYPES][SIG_COEF_CONTEXTS][2];
+  unsigned int eob_flag[TX_SIZES][PLANE_TYPES][EOB_COEF_CONTEXTS][2];
+  unsigned int dc_sign[PLANE_TYPES][DC_SIGN_CONTEXTS][2];
+  unsigned int coeff_base[TX_SIZES][PLANE_TYPES][NUM_BASE_LEVELS]
+                         [COEFF_BASE_CONTEXTS][2];
+  unsigned int coeff_lps[TX_SIZES][PLANE_TYPES][LEVEL_CONTEXTS][2];
+#endif  // CONFIG_LV_MAP
+
+#if CONFIG_EC_MULTISYMBOL
+  av1_blockz_count_model blockz_count[TX_SIZES][PLANE_TYPES];
+#endif
+
+#if CONFIG_REF_MV
+  unsigned int newmv_mode[NEWMV_MODE_CONTEXTS][2];
+  unsigned int zeromv_mode[ZEROMV_MODE_CONTEXTS][2];
+  unsigned int refmv_mode[REFMV_MODE_CONTEXTS][2];
+  unsigned int drl_mode[DRL_MODE_CONTEXTS][2];
+#endif
+
+  unsigned int inter_mode[INTER_MODE_CONTEXTS][INTER_MODES];
+#if CONFIG_EXT_INTER
+  unsigned int inter_compound_mode[INTER_MODE_CONTEXTS][INTER_COMPOUND_MODES];
+#if CONFIG_COMPOUND_SINGLEREF
+  unsigned int inter_singleref_comp_mode[INTER_MODE_CONTEXTS]
+                                        [INTER_SINGLEREF_COMP_MODES];
+#endif  // CONFIG_COMPOUND_SINGLEREF
+  unsigned int interintra[BLOCK_SIZE_GROUPS][2];
+  unsigned int interintra_mode[BLOCK_SIZE_GROUPS][INTERINTRA_MODES];
+  unsigned int wedge_interintra[BLOCK_SIZES][2];
+  unsigned int compound_interinter[BLOCK_SIZES][COMPOUND_TYPES];
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  unsigned int motion_mode[BLOCK_SIZES][MOTION_MODES];
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  unsigned int obmc[BLOCK_SIZES][2];
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  unsigned int intra_inter[INTRA_INTER_CONTEXTS][2];
+  unsigned int comp_inter[COMP_INTER_CONTEXTS][2];
+  unsigned int single_ref[REF_CONTEXTS][SINGLE_REFS - 1][2];
+#if CONFIG_EXT_REFS
+  unsigned int comp_ref[REF_CONTEXTS][FWD_REFS - 1][2];
+  unsigned int comp_bwdref[REF_CONTEXTS][BWD_REFS - 1][2];
+#else
+  unsigned int comp_ref[REF_CONTEXTS][COMP_REFS - 1][2];
+#endif  // CONFIG_EXT_REFS
+#if CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+  unsigned int comp_inter_mode[COMP_INTER_MODE_CONTEXTS][2];
+#endif  // CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+  // TODO(any): tx_size_totals is only used by the encoder to decide whether
+  // to use forward updates for the coeff probs, and as such it does not really
+  // belong into this structure.
+  unsigned int tx_size_totals[TX_SIZES];
+  unsigned int tx_size[MAX_TX_DEPTH][TX_SIZE_CONTEXTS][TX_SIZES];
+#if CONFIG_VAR_TX
+  unsigned int txfm_partition[TXFM_PARTITION_CONTEXTS][2];
+#endif
+  unsigned int skip[SKIP_CONTEXTS][2];
+#if CONFIG_REF_MV
+  nmv_context_counts mv[NMV_CONTEXTS];
+#else
+  nmv_context_counts mv;
+#endif
+#if CONFIG_INTRABC
+  nmv_context_counts dv;
+#endif
+#if CONFIG_DELTA_Q
+  unsigned int delta_q[DELTA_Q_PROBS][2];
+#if CONFIG_EXT_DELTA_Q
+  unsigned int delta_lf[DELTA_LF_PROBS][2];
+#endif
+#endif
+#if CONFIG_EXT_TX
+#if CONFIG_RECT_TX
+  unsigned int tx_size_implied[TX_SIZES][TX_SIZES];
+#endif  // CONFIG_RECT_TX
+  unsigned int inter_ext_tx[EXT_TX_SETS_INTER][EXT_TX_SIZES][TX_TYPES];
+  unsigned int intra_ext_tx[EXT_TX_SETS_INTRA][EXT_TX_SIZES][INTRA_MODES]
+                           [TX_TYPES];
+#else
+  unsigned int intra_ext_tx[EXT_TX_SIZES][TX_TYPES][TX_TYPES];
+  unsigned int inter_ext_tx[EXT_TX_SIZES][TX_TYPES];
+#endif  // CONFIG_EXT_TX
+#if CONFIG_SUPERTX
+  unsigned int supertx[PARTITION_SUPERTX_CONTEXTS][TX_SIZES][2];
+  unsigned int supertx_size[TX_SIZES];
+#endif  // CONFIG_SUPERTX
+  struct seg_counts seg;
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+  unsigned int intra_filter[INTRA_FILTERS + 1][INTRA_FILTERS];
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  unsigned int filter_intra[PLANE_TYPES][2];
+#endif  // CONFIG_FILTER_INTRA
+} FRAME_COUNTS;
+
+// Default probabilities for signaling Intra mode for Y plane -- used only for
+// intra-only frames. ('default_if_y_probs' is used for inter frames).
+// Contexts used: Intra mode (Y plane) of 'above' and 'left' blocks.
+extern const aom_prob av1_kf_y_mode_prob[INTRA_MODES][INTRA_MODES]
+                                        [INTRA_MODES - 1];
+#if CONFIG_EC_MULTISYMBOL
+// CDF version of 'av1_kf_y_mode_prob'.
+extern const aom_cdf_prob av1_kf_y_mode_cdf[INTRA_MODES][INTRA_MODES]
+                                           [CDF_SIZE(INTRA_MODES)];
+#endif
+
+#if CONFIG_PALETTE
+extern const aom_prob av1_default_palette_y_mode_prob[PALETTE_BLOCK_SIZES]
+                                                     [PALETTE_Y_MODE_CONTEXTS];
+extern const aom_prob
+    av1_default_palette_uv_mode_prob[PALETTE_UV_MODE_CONTEXTS];
+extern const aom_prob av1_default_palette_y_size_prob[PALETTE_BLOCK_SIZES]
+                                                     [PALETTE_SIZES - 1];
+extern const aom_prob av1_default_palette_uv_size_prob[PALETTE_BLOCK_SIZES]
+                                                      [PALETTE_SIZES - 1];
+extern const aom_prob av1_default_palette_y_color_index_prob
+    [PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS][PALETTE_COLORS - 1];
+extern const aom_prob av1_default_palette_uv_color_index_prob
+    [PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS][PALETTE_COLORS - 1];
+#endif  // CONFIG_PALETTE
+
+extern const aom_tree_index av1_intra_mode_tree[TREE_SIZE(INTRA_MODES)];
+extern const aom_tree_index av1_inter_mode_tree[TREE_SIZE(INTER_MODES)];
+#if CONFIG_EC_MULTISYMBOL
+extern int av1_intra_mode_ind[INTRA_MODES];
+extern int av1_intra_mode_inv[INTRA_MODES];
+extern int av1_inter_mode_ind[INTER_MODES];
+extern int av1_inter_mode_inv[INTER_MODES];
+#if CONFIG_EXT_TX
+extern int av1_ext_tx_intra_ind[EXT_TX_SETS_INTRA][TX_TYPES];
+extern int av1_ext_tx_intra_inv[EXT_TX_SETS_INTRA][TX_TYPES];
+extern int av1_ext_tx_inter_ind[EXT_TX_SETS_INTER][TX_TYPES];
+extern int av1_ext_tx_inter_inv[EXT_TX_SETS_INTER][TX_TYPES];
+#endif
+#endif
+
+#if CONFIG_EXT_INTER
+extern const aom_tree_index
+    av1_interintra_mode_tree[TREE_SIZE(INTERINTRA_MODES)];
+extern const aom_tree_index
+    av1_inter_compound_mode_tree[TREE_SIZE(INTER_COMPOUND_MODES)];
+#if CONFIG_COMPOUND_SINGLEREF
+extern const aom_tree_index
+    av1_inter_singleref_comp_mode_tree[TREE_SIZE(INTER_SINGLEREF_COMP_MODES)];
+#endif  // CONFIG_COMPOUND_SINGLEREF
+extern const aom_tree_index av1_compound_type_tree[TREE_SIZE(COMPOUND_TYPES)];
+#endif  // CONFIG_EXT_INTER
+extern const aom_tree_index av1_partition_tree[TREE_SIZE(PARTITION_TYPES)];
+#if CONFIG_EXT_PARTITION_TYPES
+extern const aom_tree_index
+    av1_ext_partition_tree[TREE_SIZE(EXT_PARTITION_TYPES)];
+#endif
+extern const aom_tree_index
+    av1_switchable_interp_tree[TREE_SIZE(SWITCHABLE_FILTERS)];
+#if CONFIG_PALETTE
+extern const aom_tree_index av1_palette_size_tree[TREE_SIZE(PALETTE_SIZES)];
+extern const aom_tree_index
+    av1_palette_color_index_tree[PALETTE_SIZES][TREE_SIZE(PALETTE_COLORS)];
+#endif  // CONFIG_PALETTE
+extern const aom_tree_index av1_tx_size_tree[MAX_TX_DEPTH][TREE_SIZE(TX_SIZES)];
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+extern const aom_tree_index av1_intra_filter_tree[TREE_SIZE(INTRA_FILTERS)];
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+#if CONFIG_EXT_TX
+extern const aom_tree_index av1_ext_tx_inter_tree[EXT_TX_SETS_INTER]
+                                                 [TREE_SIZE(TX_TYPES)];
+extern const aom_tree_index av1_ext_tx_intra_tree[EXT_TX_SETS_INTRA]
+                                                 [TREE_SIZE(TX_TYPES)];
+#else
+extern const aom_tree_index av1_ext_tx_tree[TREE_SIZE(TX_TYPES)];
+#endif  // CONFIG_EXT_TX
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+extern const aom_tree_index av1_motion_mode_tree[TREE_SIZE(MOTION_MODES)];
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_LOOP_RESTORATION
+#define RESTORE_NONE_SGRPROJ_PROB 64
+#define RESTORE_NONE_BILATERAL_PROB 16
+#define RESTORE_NONE_WIENER_PROB 64
+#define RESTORE_NONE_DOMAINTXFMRF_PROB 64
+extern const aom_tree_index
+    av1_switchable_restore_tree[TREE_SIZE(RESTORE_SWITCHABLE_TYPES)];
+#endif  // CONFIG_LOOP_RESTORATION
+#if CONFIG_EC_MULTISYMBOL
+extern int av1_switchable_interp_ind[SWITCHABLE_FILTERS];
+extern int av1_switchable_interp_inv[SWITCHABLE_FILTERS];
+
+void av1_set_mode_cdfs(struct AV1Common *cm);
+#endif
+
+void av1_setup_past_independence(struct AV1Common *cm);
+
+void av1_adapt_intra_frame_probs(struct AV1Common *cm);
+void av1_adapt_inter_frame_probs(struct AV1Common *cm);
+#if CONFIG_EC_MULTISYMBOL && !CONFIG_EXT_TX
+extern int av1_ext_tx_ind[TX_TYPES];
+extern int av1_ext_tx_inv[TX_TYPES];
+#endif
+
+static INLINE int av1_ceil_log2(int n) {
+  int i = 1, p = 2;
+  while (p < n) {
+    i++;
+    p = p << 1;
+  }
+  return i;
+}
+
+#if CONFIG_PALETTE
+// Returns the context for palette color index at row 'r' and column 'c',
+// along with the 'color_order' of neighbors and the 'color_idx'.
+// The 'color_map' is a 2D array with the given 'stride'.
+int av1_get_palette_color_index_context(const uint8_t *color_map, int stride,
+                                        int r, int c, int palette_size,
+                                        uint8_t *color_order, int *color_idx);
+#endif  // CONFIG_PALETTE
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_ENTROPYMODE_H_
diff --git a/third_party/aom/av1/common/entropymv.c b/third_party/aom/av1/common/entropymv.c
new file mode 100644
index 0000000000..9c162d2c50
--- /dev/null
+++ b/third_party/aom/av1/common/entropymv.c
@@ -0,0 +1,315 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/onyxc_int.h"
+#include "av1/common/entropymv.h"
+
+// Integer pel reference mv threshold for use of high-precision 1/8 mv
+#define COMPANDED_MVREF_THRESH 8
+
+const aom_tree_index av1_mv_joint_tree[TREE_SIZE(MV_JOINTS)] = {
+  -MV_JOINT_ZERO, 2, -MV_JOINT_HNZVZ, 4, -MV_JOINT_HZVNZ, -MV_JOINT_HNZVNZ
+};
+
+/* clang-format off */
+const aom_tree_index av1_mv_class_tree[TREE_SIZE(MV_CLASSES)] = {
+  -MV_CLASS_0, 2,
+  -MV_CLASS_1, 4,
+  6, 8,
+  -MV_CLASS_2, -MV_CLASS_3,
+  10, 12,
+  -MV_CLASS_4, -MV_CLASS_5,
+  -MV_CLASS_6, 14,
+  16, 18,
+  -MV_CLASS_7, -MV_CLASS_8,
+  -MV_CLASS_9, -MV_CLASS_10,
+};
+/* clang-format on */
+
+const aom_tree_index av1_mv_class0_tree[TREE_SIZE(CLASS0_SIZE)] = {
+  -0, -1,
+};
+
+const aom_tree_index av1_mv_fp_tree[TREE_SIZE(MV_FP_SIZE)] = { -0, 2,  -1,
+                                                               4,  -2, -3 };
+
+static const nmv_context default_nmv_context = {
+  { 32, 64, 96 },  // joints
+#if CONFIG_EC_MULTISYMBOL
+  { AOM_ICDF(4096), AOM_ICDF(11264), AOM_ICDF(19328), AOM_ICDF(32768),
+    0 },  // joint_cdf
+#endif
+  { {
+        // Vertical component
+        128,                                                   // sign
+        { 224, 144, 192, 168, 192, 176, 192, 198, 198, 245 },  // class
+#if CONFIG_EC_MULTISYMBOL
+        { AOM_ICDF(28672), AOM_ICDF(30976), AOM_ICDF(31858), AOM_ICDF(32320),
+          AOM_ICDF(32551), AOM_ICDF(32656), AOM_ICDF(32740), AOM_ICDF(32757),
+          AOM_ICDF(32762), AOM_ICDF(32767), AOM_ICDF(32768), 0 },  // class_cdf
+#endif
+        { 216 },                                               // class0
+        { 136, 140, 148, 160, 176, 192, 224, 234, 234, 240 },  // bits
+        { { 128, 128, 64 }, { 96, 112, 64 } },                 // class0_fp
+        { 64, 96, 64 },                                        // fp
+#if CONFIG_EC_MULTISYMBOL
+        { { AOM_ICDF(16384), AOM_ICDF(24576), AOM_ICDF(26624), AOM_ICDF(32768),
+            0 },
+          { AOM_ICDF(12288), AOM_ICDF(21248), AOM_ICDF(24128), AOM_ICDF(32768),
+            0 } },  // class0_fp_cdf
+        { AOM_ICDF(8192), AOM_ICDF(17408), AOM_ICDF(21248), AOM_ICDF(32768),
+          0 },  // fp_cdf
+#endif
+        160,  // class0_hp bit
+        128,  // hp
+    },
+    {
+        // Horizontal component
+        128,                                                   // sign
+        { 216, 128, 176, 160, 176, 176, 192, 198, 198, 208 },  // class
+#if CONFIG_EC_MULTISYMBOL
+        { AOM_ICDF(28672), AOM_ICDF(30976), AOM_ICDF(31858), AOM_ICDF(32320),
+          AOM_ICDF(32551), AOM_ICDF(32656), AOM_ICDF(32740), AOM_ICDF(32757),
+          AOM_ICDF(32762), AOM_ICDF(32767), AOM_ICDF(32768), 0 },  // class_cdf
+#endif
+        { 208 },                                               // class0
+        { 136, 140, 148, 160, 176, 192, 224, 234, 234, 240 },  // bits
+        { { 128, 128, 64 }, { 96, 112, 64 } },                 // class0_fp
+        { 64, 96, 64 },                                        // fp
+#if CONFIG_EC_MULTISYMBOL
+        { { AOM_ICDF(16384), AOM_ICDF(24576), AOM_ICDF(26624), AOM_ICDF(32768),
+            0 },
+          { AOM_ICDF(12288), AOM_ICDF(21248), AOM_ICDF(24128), AOM_ICDF(32768),
+            0 } },  // class0_fp_cdf
+        { AOM_ICDF(8192), AOM_ICDF(17408), AOM_ICDF(21248), AOM_ICDF(32768),
+          0 },  // fp_cdf
+#endif
+        160,  // class0_hp bit
+        128,  // hp
+    } },
+};
+
+static const uint8_t log_in_base_2[] = {
+  0, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
+  4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5,
+  5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6,
+  6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7,
+  7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9,
+  9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 10
+};
+
+#if CONFIG_GLOBAL_MOTION
+#if GLOBAL_TRANS_TYPES == 7  // All models
+const aom_tree_index av1_global_motion_types_tree[TREE_SIZE(
+    GLOBAL_TRANS_TYPES)] = { -IDENTITY,   2,  -TRANSLATION,  4,
+                             -ROTZOOM,    6,  -AFFINE,       8,
+                             -HOMOGRAPHY, 10, -HORTRAPEZOID, -VERTRAPEZOID };
+
+static const aom_prob default_global_motion_types_prob[GLOBAL_TRANS_TYPES - 1] =
+    { 224, 128, 192, 192, 32, 128 };
+
+#elif GLOBAL_TRANS_TYPES == 6  // Do not allow full homography
+const aom_tree_index
+    av1_global_motion_types_tree[TREE_SIZE(GLOBAL_TRANS_TYPES)] = {
+      -IDENTITY,    2, -TRANSLATION, 4, -ROTZOOM, 6, -AFFINE, 8, -HORTRAPEZOID,
+      -VERTRAPEZOID
+    };
+
+static const aom_prob default_global_motion_types_prob[GLOBAL_TRANS_TYPES - 1] =
+    { 224, 128, 192, 192, 128 };
+
+#elif GLOBAL_TRANS_TYPES == 4  // Upto Affine
+const aom_tree_index av1_global_motion_types_tree[TREE_SIZE(
+    GLOBAL_TRANS_TYPES)] = { -IDENTITY, 2, -TRANSLATION, 4, -ROTZOOM, -AFFINE };
+
+static const aom_prob default_global_motion_types_prob[GLOBAL_TRANS_TYPES - 1] =
+    { 224, 128, 240 };
+
+#elif GLOBAL_TRANS_TYPES == 3  // Upto rotation-zoom
+
+const aom_tree_index av1_global_motion_types_tree[TREE_SIZE(
+    GLOBAL_TRANS_TYPES)] = { -IDENTITY, 2, -TRANSLATION, -ROTZOOM };
+
+static const aom_prob default_global_motion_types_prob[GLOBAL_TRANS_TYPES - 1] =
+    { 224, 128 };
+#endif                         // GLOBAL_TRANS_TYPES
+#endif                         // CONFIG_GLOBAL_MOTION
+
+static INLINE int mv_class_base(MV_CLASS_TYPE c) {
+  return c ? CLASS0_SIZE << (c + 2) : 0;
+}
+
+MV_CLASS_TYPE av1_get_mv_class(int z, int *offset) {
+  const MV_CLASS_TYPE c = (z >= CLASS0_SIZE * 4096)
+                              ? MV_CLASS_10
+                              : (MV_CLASS_TYPE)log_in_base_2[z >> 3];
+  if (offset) *offset = z - mv_class_base(c);
+  return c;
+}
+
+static void inc_mv_component(int v, nmv_component_counts *comp_counts, int incr,
+                             int usehp) {
+  int s, z, c, o, d, e, f;
+  assert(v != 0); /* should not be zero */
+  s = v < 0;
+  comp_counts->sign[s] += incr;
+  z = (s ? -v : v) - 1; /* magnitude - 1 */
+
+  c = av1_get_mv_class(z, &o);
+  comp_counts->classes[c] += incr;
+
+  d = (o >> 3);     /* int mv data */
+  f = (o >> 1) & 3; /* fractional pel mv data */
+  e = (o & 1);      /* high precision mv data */
+
+  if (c == MV_CLASS_0) {
+    comp_counts->class0[d] += incr;
+    comp_counts->class0_fp[d][f] += incr;
+    if (usehp) comp_counts->class0_hp[e] += incr;
+  } else {
+    int i;
+    int b = c + CLASS0_BITS - 1;  // number of bits
+    for (i = 0; i < b; ++i) comp_counts->bits[i][((d >> i) & 1)] += incr;
+    comp_counts->fp[f] += incr;
+    if (usehp) comp_counts->hp[e] += incr;
+  }
+}
+
+void av1_inc_mv(const MV *mv, nmv_context_counts *counts, const int usehp) {
+  if (counts != NULL) {
+    const MV_JOINT_TYPE j = av1_get_mv_joint(mv);
+    ++counts->joints[j];
+
+    if (mv_joint_vertical(j))
+      inc_mv_component(mv->row, &counts->comps[0], 1, usehp);
+
+    if (mv_joint_horizontal(j))
+      inc_mv_component(mv->col, &counts->comps[1], 1, usehp);
+  }
+}
+
+void av1_adapt_mv_probs(AV1_COMMON *cm, int allow_hp) {
+  int i, j;
+#if CONFIG_REF_MV
+  int idx;
+  for (idx = 0; idx < NMV_CONTEXTS; ++idx) {
+    nmv_context *fc = &cm->fc->nmvc[idx];
+    const nmv_context *pre_fc =
+        &cm->frame_contexts[cm->frame_context_idx].nmvc[idx];
+    const nmv_context_counts *counts = &cm->counts.mv[idx];
+#else
+  nmv_context *fc = &cm->fc->nmvc;
+  const nmv_context *pre_fc = &cm->frame_contexts[cm->frame_context_idx].nmvc;
+  const nmv_context_counts *counts = &cm->counts.mv;
+#endif  // CONFIG_REF_MV
+    aom_tree_merge_probs(av1_mv_joint_tree, pre_fc->joints, counts->joints,
+                         fc->joints);
+    for (i = 0; i < 2; ++i) {
+      nmv_component *comp = &fc->comps[i];
+      const nmv_component *pre_comp = &pre_fc->comps[i];
+      const nmv_component_counts *c = &counts->comps[i];
+
+      comp->sign = av1_mode_mv_merge_probs(pre_comp->sign, c->sign);
+      aom_tree_merge_probs(av1_mv_class_tree, pre_comp->classes, c->classes,
+                           comp->classes);
+      aom_tree_merge_probs(av1_mv_class0_tree, pre_comp->class0, c->class0,
+                           comp->class0);
+
+      for (j = 0; j < MV_OFFSET_BITS; ++j)
+        comp->bits[j] = av1_mode_mv_merge_probs(pre_comp->bits[j], c->bits[j]);
+
+      for (j = 0; j < CLASS0_SIZE; ++j)
+        aom_tree_merge_probs(av1_mv_fp_tree, pre_comp->class0_fp[j],
+                             c->class0_fp[j], comp->class0_fp[j]);
+
+      aom_tree_merge_probs(av1_mv_fp_tree, pre_comp->fp, c->fp, comp->fp);
+
+      if (allow_hp) {
+        comp->class0_hp =
+            av1_mode_mv_merge_probs(pre_comp->class0_hp, c->class0_hp);
+        comp->hp = av1_mode_mv_merge_probs(pre_comp->hp, c->hp);
+      }
+    }
+#if CONFIG_REF_MV
+  }
+#endif  // CONFIG_REF_MV
+}
+
+#if CONFIG_EC_MULTISYMBOL && !CONFIG_EC_ADAPT
+void av1_set_mv_cdfs(nmv_context *ctx) {
+  int i;
+  int j;
+  av1_tree_to_cdf(av1_mv_joint_tree, ctx->joints, ctx->joint_cdf);
+
+  for (i = 0; i < 2; ++i) {
+    nmv_component *const comp_ctx = &ctx->comps[i];
+    av1_tree_to_cdf(av1_mv_class_tree, comp_ctx->classes, comp_ctx->class_cdf);
+
+    for (j = 0; j < CLASS0_SIZE; ++j) {
+      av1_tree_to_cdf(av1_mv_fp_tree, comp_ctx->class0_fp[j],
+                      comp_ctx->class0_fp_cdf[j]);
+    }
+    av1_tree_to_cdf(av1_mv_fp_tree, comp_ctx->fp, comp_ctx->fp_cdf);
+  }
+}
+#endif
+
+void av1_init_mv_probs(AV1_COMMON *cm) {
+#if CONFIG_REF_MV
+  int i;
+  for (i = 0; i < NMV_CONTEXTS; ++i) {
+    // NB: this sets CDFs too
+    cm->fc->nmvc[i] = default_nmv_context;
+  }
+#else
+  cm->fc->nmvc = default_nmv_context;
+#endif  // CONFIG_REF_MV
+#if CONFIG_INTRABC
+  cm->fc->ndvc = default_nmv_context;
+#endif  // CONFIG_INTRABC
+#if CONFIG_GLOBAL_MOTION
+  av1_copy(cm->fc->global_motion_types_prob, default_global_motion_types_prob);
+#endif  // CONFIG_GLOBAL_MOTION
+}
diff --git a/third_party/aom/av1/common/entropymv.h b/third_party/aom/av1/common/entropymv.h
new file mode 100644
index 0000000000..2c79d447a5
--- /dev/null
+++ b/third_party/aom/av1/common/entropymv.h
@@ -0,0 +1,149 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_ENTROPYMV_H_
+#define AV1_COMMON_ENTROPYMV_H_
+
+#include "./aom_config.h"
+
+#include "aom_dsp/prob.h"
+
+#include "av1/common/mv.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1Common;
+
+void av1_init_mv_probs(struct AV1Common *cm);
+
+void av1_adapt_mv_probs(struct AV1Common *cm, int usehp);
+
+#define MV_UPDATE_PROB 252
+
+/* Symbols for coding which components are zero jointly */
+#define MV_JOINTS 4
+typedef enum {
+  MV_JOINT_ZERO = 0,   /* Zero vector */
+  MV_JOINT_HNZVZ = 1,  /* Vert zero, hor nonzero */
+  MV_JOINT_HZVNZ = 2,  /* Hor zero, vert nonzero */
+  MV_JOINT_HNZVNZ = 3, /* Both components nonzero */
+} MV_JOINT_TYPE;
+
+static INLINE int mv_joint_vertical(MV_JOINT_TYPE type) {
+  return type == MV_JOINT_HZVNZ || type == MV_JOINT_HNZVNZ;
+}
+
+static INLINE int mv_joint_horizontal(MV_JOINT_TYPE type) {
+  return type == MV_JOINT_HNZVZ || type == MV_JOINT_HNZVNZ;
+}
+
+/* Symbols for coding magnitude class of nonzero components */
+#define MV_CLASSES 11
+typedef enum {
+  MV_CLASS_0 = 0,   /* (0, 2]     integer pel */
+  MV_CLASS_1 = 1,   /* (2, 4]     integer pel */
+  MV_CLASS_2 = 2,   /* (4, 8]     integer pel */
+  MV_CLASS_3 = 3,   /* (8, 16]    integer pel */
+  MV_CLASS_4 = 4,   /* (16, 32]   integer pel */
+  MV_CLASS_5 = 5,   /* (32, 64]   integer pel */
+  MV_CLASS_6 = 6,   /* (64, 128]  integer pel */
+  MV_CLASS_7 = 7,   /* (128, 256] integer pel */
+  MV_CLASS_8 = 8,   /* (256, 512] integer pel */
+  MV_CLASS_9 = 9,   /* (512, 1024] integer pel */
+  MV_CLASS_10 = 10, /* (1024,2048] integer pel */
+} MV_CLASS_TYPE;
+
+#define CLASS0_BITS 1 /* bits at integer precision for class 0 */
+#define CLASS0_SIZE (1 << CLASS0_BITS)
+#define MV_OFFSET_BITS (MV_CLASSES + CLASS0_BITS - 2)
+#define MV_FP_SIZE 4
+
+#define MV_MAX_BITS (MV_CLASSES + CLASS0_BITS + 2)
+#define MV_MAX ((1 << MV_MAX_BITS) - 1)
+#define MV_VALS ((MV_MAX << 1) + 1)
+
+#define MV_IN_USE_BITS 14
+#define MV_UPP ((1 << MV_IN_USE_BITS) - 1)
+#define MV_LOW (-(1 << MV_IN_USE_BITS))
+
+extern const aom_tree_index av1_mv_joint_tree[];
+extern const aom_tree_index av1_mv_class_tree[];
+extern const aom_tree_index av1_mv_class0_tree[];
+extern const aom_tree_index av1_mv_fp_tree[];
+
+typedef struct {
+  aom_prob sign;
+  aom_prob classes[MV_CLASSES - 1];
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob class_cdf[CDF_SIZE(MV_CLASSES)];
+#endif
+  aom_prob class0[CLASS0_SIZE - 1];
+  aom_prob bits[MV_OFFSET_BITS];
+  aom_prob class0_fp[CLASS0_SIZE][MV_FP_SIZE - 1];
+  aom_prob fp[MV_FP_SIZE - 1];
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob class0_fp_cdf[CLASS0_SIZE][CDF_SIZE(MV_FP_SIZE)];
+  aom_cdf_prob fp_cdf[CDF_SIZE(MV_FP_SIZE)];
+#endif
+  aom_prob class0_hp;
+  aom_prob hp;
+} nmv_component;
+
+typedef struct {
+  aom_prob joints[MV_JOINTS - 1];
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob joint_cdf[CDF_SIZE(MV_JOINTS)];
+#endif
+  nmv_component comps[2];
+} nmv_context;
+
+static INLINE MV_JOINT_TYPE av1_get_mv_joint(const MV *mv) {
+  if (mv->row == 0) {
+    return mv->col == 0 ? MV_JOINT_ZERO : MV_JOINT_HNZVZ;
+  } else {
+    return mv->col == 0 ? MV_JOINT_HZVNZ : MV_JOINT_HNZVNZ;
+  }
+}
+
+MV_CLASS_TYPE av1_get_mv_class(int z, int *offset);
+
+typedef struct {
+  unsigned int sign[2];
+  unsigned int classes[MV_CLASSES];
+  unsigned int class0[CLASS0_SIZE];
+  unsigned int bits[MV_OFFSET_BITS][2];
+  unsigned int class0_fp[CLASS0_SIZE][MV_FP_SIZE];
+  unsigned int fp[MV_FP_SIZE];
+  unsigned int class0_hp[2];
+  unsigned int hp[2];
+} nmv_component_counts;
+
+typedef struct {
+  unsigned int joints[MV_JOINTS];
+  nmv_component_counts comps[2];
+} nmv_context_counts;
+
+void av1_inc_mv(const MV *mv, nmv_context_counts *mvctx, const int usehp);
+#if CONFIG_GLOBAL_MOTION
+extern const aom_tree_index
+    av1_global_motion_types_tree[TREE_SIZE(GLOBAL_TRANS_TYPES)];
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_EC_MULTISYMBOL
+void av1_set_mv_cdfs(nmv_context *ctx);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_ENTROPYMV_H_
diff --git a/third_party/aom/av1/common/enums.h b/third_party/aom/av1/common/enums.h
new file mode 100644
index 0000000000..054bd40be1
--- /dev/null
+++ b/third_party/aom/av1/common/enums.h
@@ -0,0 +1,543 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_ENUMS_H_
+#define AV1_COMMON_ENUMS_H_
+
+#include "./aom_config.h"
+#include "aom/aom_codec.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#undef MAX_SB_SIZE
+
+// Max superblock size
+#if CONFIG_EXT_PARTITION
+#define MAX_SB_SIZE_LOG2 7
+#else
+#define MAX_SB_SIZE_LOG2 6
+#endif  // CONFIG_EXT_PARTITION
+#define MAX_SB_SIZE (1 << MAX_SB_SIZE_LOG2)
+#define MAX_SB_SQUARE (MAX_SB_SIZE * MAX_SB_SIZE)
+
+// Min superblock size
+#define MIN_SB_SIZE_LOG2 6
+
+// Pixels per Mode Info (MI) unit
+#if CONFIG_CB4X4
+#define MI_SIZE_LOG2 2
+#else
+#define MI_SIZE_LOG2 3
+#endif
+#define MI_SIZE (1 << MI_SIZE_LOG2)
+
+// MI-units per max superblock (MI Block - MIB)
+#define MAX_MIB_SIZE_LOG2 (MAX_SB_SIZE_LOG2 - MI_SIZE_LOG2)
+#define MAX_MIB_SIZE (1 << MAX_MIB_SIZE_LOG2)
+
+// MI-units per min superblock
+#define MIN_MIB_SIZE_LOG2 (MIN_SB_SIZE_LOG2 - MI_SIZE_LOG2)
+
+// Mask to extract MI offset within max MIB
+#define MAX_MIB_MASK (MAX_MIB_SIZE - 1)
+#define MAX_MIB_MASK_2 (MAX_MIB_SIZE * 2 - 1)
+
+// Maximum number of tile rows and tile columns
+#if CONFIG_EXT_TILE
+#define TILE_NORMAL 0
+#define TILE_VR 1
+
+#define MAX_TILE_ROWS 1024
+#define MAX_TILE_COLS 1024
+#else
+#define MAX_TILE_ROWS 4
+#define MAX_TILE_COLS 64
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_VAR_TX
+#define MAX_VARTX_DEPTH 2
+#endif
+
+// Bitstream profiles indicated by 2-3 bits in the uncompressed header.
+// 00: Profile 0.  8-bit 4:2:0 only.
+// 10: Profile 1.  8-bit 4:4:4, 4:2:2, and 4:4:0.
+// 01: Profile 2.  10-bit and 12-bit color only, with 4:2:0 sampling.
+// 110: Profile 3. 10-bit and 12-bit color only, with 4:2:2/4:4:4/4:4:0
+//                 sampling.
+// 111: Undefined profile.
+typedef enum BITSTREAM_PROFILE {
+  PROFILE_0,
+  PROFILE_1,
+  PROFILE_2,
+  PROFILE_3,
+  MAX_PROFILES
+} BITSTREAM_PROFILE;
+
+// Note: Some enums use the attribute 'packed' to use smallest possible integer
+// type, so that we can save memory when they are used in structs/arrays.
+
+typedef enum ATTRIBUTE_PACKED {
+#if CONFIG_CB4X4
+  BLOCK_2X2,
+  BLOCK_2X4,
+  BLOCK_4X2,
+#endif
+  BLOCK_4X4,
+  BLOCK_4X8,
+  BLOCK_8X4,
+  BLOCK_8X8,
+  BLOCK_8X16,
+  BLOCK_16X8,
+  BLOCK_16X16,
+  BLOCK_16X32,
+  BLOCK_32X16,
+  BLOCK_32X32,
+  BLOCK_32X64,
+  BLOCK_64X32,
+  BLOCK_64X64,
+#if CONFIG_EXT_PARTITION
+  BLOCK_64X128,
+  BLOCK_128X64,
+  BLOCK_128X128,
+#endif  // CONFIG_EXT_PARTITION
+  BLOCK_SIZES,
+  BLOCK_INVALID = BLOCK_SIZES,
+  BLOCK_LARGEST = (BLOCK_SIZES - 1)
+} BLOCK_SIZE;
+
+typedef enum {
+  PARTITION_NONE,
+  PARTITION_HORZ,
+  PARTITION_VERT,
+  PARTITION_SPLIT,
+#if CONFIG_EXT_PARTITION_TYPES
+  PARTITION_HORZ_A,  // HORZ split and the left partition is split again
+  PARTITION_HORZ_B,  // HORZ split and the right partition is split again
+  PARTITION_VERT_A,  // VERT split and the top partition is split again
+  PARTITION_VERT_B,  // VERT split and the bottom partition is split again
+  EXT_PARTITION_TYPES,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+  PARTITION_TYPES = PARTITION_SPLIT + 1,
+  PARTITION_INVALID = 255
+} PARTITION_TYPE;
+
+typedef char PARTITION_CONTEXT;
+#define PARTITION_PLOFFSET 4  // number of probability models per block size
+#define PARTITION_BLOCK_SIZES (4 + CONFIG_EXT_PARTITION)
+#define PARTITION_CONTEXTS_PRIMARY (PARTITION_BLOCK_SIZES * PARTITION_PLOFFSET)
+#if CONFIG_UNPOISON_PARTITION_CTX
+#define PARTITION_CONTEXTS \
+  (PARTITION_CONTEXTS_PRIMARY + 2 * PARTITION_BLOCK_SIZES)
+#else
+#define PARTITION_CONTEXTS PARTITION_CONTEXTS_PRIMARY
+#endif
+
+// block transform size
+typedef enum ATTRIBUTE_PACKED {
+#if CONFIG_CB4X4
+  TX_2X2,  // 2x2 transform
+#endif
+  TX_4X4,    // 4x4 transform
+  TX_8X8,    // 8x8 transform
+  TX_16X16,  // 16x16 transform
+  TX_32X32,  // 32x32 transform
+#if CONFIG_TX64X64
+  TX_64X64,           // 64x64 transform
+#endif                // CONFIG_TX64X64
+  TX_4X8,             // 4x8 transform
+  TX_8X4,             // 8x4 transform
+  TX_8X16,            // 8x16 transform
+  TX_16X8,            // 16x8 transform
+  TX_16X32,           // 16x32 transform
+  TX_32X16,           // 32x16 transform
+  TX_4X16,            // 4x16 transform
+  TX_16X4,            // 16x4 transform
+  TX_8X32,            // 8x32 transform
+  TX_32X8,            // 32x8 transform
+  TX_SIZES_ALL,       // Includes rectangular transforms
+  TX_SIZES = TX_4X8,  // Does NOT include rectangular transforms
+  TX_INVALID = 255    // Invalid transform size
+} TX_SIZE;
+
+#define MAX_TX_DEPTH (TX_SIZES - 1 - TX_4X4)
+
+#define MAX_TX_SIZE_LOG2 (5 + CONFIG_TX64X64)
+#define MAX_TX_SIZE (1 << MAX_TX_SIZE_LOG2)
+#define MIN_TX_SIZE_LOG2 2
+#define MIN_TX_SIZE (1 << MIN_TX_SIZE_LOG2)
+#define MAX_TX_SQUARE (MAX_TX_SIZE * MAX_TX_SIZE)
+
+// Number of maxium size transform blocks in the maximum size superblock
+#define MAX_TX_BLOCKS_IN_MAX_SB_LOG2 ((MAX_SB_SIZE_LOG2 - MAX_TX_SIZE_LOG2) * 2)
+#define MAX_TX_BLOCKS_IN_MAX_SB (1 << MAX_TX_BLOCKS_IN_MAX_SB_LOG2)
+
+#define MAX_NUM_TXB (1 << (MAX_SB_SIZE_LOG2 - MIN_TX_SIZE_LOG2))
+
+// frame transform mode
+typedef enum {
+  ONLY_4X4 = 0,     // only 4x4 transform used
+  ALLOW_8X8 = 1,    // allow block transform size up to 8x8
+  ALLOW_16X16 = 2,  // allow block transform size up to 16x16
+  ALLOW_32X32 = 3,  // allow block transform size up to 32x32
+#if CONFIG_TX64X64
+  ALLOW_64X64 = 4,  // allow block transform size up to 64x64
+#endif
+  TX_MODE_SELECT,  // transform specified for each block
+  TX_MODES,
+} TX_MODE;
+
+// 1D tx types
+typedef enum {
+  DCT_1D = 0,
+  ADST_1D = 1,
+  FLIPADST_1D = 2,
+  IDTX_1D = 3,
+  TX_TYPES_1D = 4,
+} TX_TYPE_1D;
+
+typedef enum {
+  DCT_DCT = 0,    // DCT  in both horizontal and vertical
+  ADST_DCT = 1,   // ADST in vertical, DCT in horizontal
+  DCT_ADST = 2,   // DCT  in vertical, ADST in horizontal
+  ADST_ADST = 3,  // ADST in both directions
+#if CONFIG_EXT_TX
+  FLIPADST_DCT = 4,
+  DCT_FLIPADST = 5,
+  FLIPADST_FLIPADST = 6,
+  ADST_FLIPADST = 7,
+  FLIPADST_ADST = 8,
+  IDTX = 9,
+  V_DCT = 10,
+  H_DCT = 11,
+  V_ADST = 12,
+  H_ADST = 13,
+  V_FLIPADST = 14,
+  H_FLIPADST = 15,
+#endif  // CONFIG_EXT_TX
+  TX_TYPES,
+} TX_TYPE;
+
+typedef enum {
+  TILE_LEFT_BOUNDARY = 1,
+  TILE_RIGHT_BOUNDARY = 2,
+  TILE_ABOVE_BOUNDARY = 4,
+  TILE_BOTTOM_BOUNDARY = 8,
+  FRAME_LEFT_BOUNDARY = 16,
+  FRAME_RIGHT_BOUNDARY = 32,
+  FRAME_ABOVE_BOUNDARY = 64,
+  FRAME_BOTTOM_BOUNDARY = 128,
+} BOUNDARY_TYPE;
+
+#if CONFIG_EXT_TX
+#if CONFIG_CB4X4
+#define EXT_TX_SIZES 5  // number of sizes that use extended transforms
+#else
+#define EXT_TX_SIZES 4       // number of sizes that use extended transforms
+#endif                       // CONFIG_CB4X4
+#define EXT_TX_SETS_INTER 4  // Sets of transform selections for INTER
+#define EXT_TX_SETS_INTRA 3  // Sets of transform selections for INTRA
+#else
+#if CONFIG_CB4X4
+#define EXT_TX_SIZES 4  // number of sizes that use extended transforms
+#else
+#define EXT_TX_SIZES 3  // number of sizes that use extended transforms
+#endif
+#endif  // CONFIG_EXT_TX
+
+typedef enum {
+  AOM_LAST_FLAG = 1 << 0,
+#if CONFIG_EXT_REFS
+  AOM_LAST2_FLAG = 1 << 1,
+  AOM_LAST3_FLAG = 1 << 2,
+  AOM_GOLD_FLAG = 1 << 3,
+  AOM_BWD_FLAG = 1 << 4,
+  AOM_ALT_FLAG = 1 << 5,
+  AOM_REFFRAME_ALL = (1 << 6) - 1
+#else
+  AOM_GOLD_FLAG = 1 << 1,
+  AOM_ALT_FLAG = 1 << 2,
+  AOM_REFFRAME_ALL = (1 << 3) - 1
+#endif  // CONFIG_EXT_REFS
+} AOM_REFFRAME;
+
+typedef enum { PLANE_TYPE_Y = 0, PLANE_TYPE_UV = 1, PLANE_TYPES } PLANE_TYPE;
+
+#if CONFIG_CFL
+typedef enum { CFL_PRED_U = 0, CFL_PRED_V = 1, CFL_PRED_PLANES } CFL_PRED_TYPE;
+#endif
+
+#if CONFIG_PALETTE
+typedef enum {
+  TWO_COLORS,
+  THREE_COLORS,
+  FOUR_COLORS,
+  FIVE_COLORS,
+  SIX_COLORS,
+  SEVEN_COLORS,
+  EIGHT_COLORS,
+  PALETTE_SIZES
+} PALETTE_SIZE;
+
+typedef enum {
+  PALETTE_COLOR_ONE,
+  PALETTE_COLOR_TWO,
+  PALETTE_COLOR_THREE,
+  PALETTE_COLOR_FOUR,
+  PALETTE_COLOR_FIVE,
+  PALETTE_COLOR_SIX,
+  PALETTE_COLOR_SEVEN,
+  PALETTE_COLOR_EIGHT,
+  PALETTE_COLORS
+} PALETTE_COLOR;
+#endif  // CONFIG_PALETTE
+
+typedef enum ATTRIBUTE_PACKED {
+  DC_PRED,    // Average of above and left pixels
+  V_PRED,     // Vertical
+  H_PRED,     // Horizontal
+  D45_PRED,   // Directional 45  deg = round(arctan(1/1) * 180/pi)
+  D135_PRED,  // Directional 135 deg = 180 - 45
+  D117_PRED,  // Directional 117 deg = 180 - 63
+  D153_PRED,  // Directional 153 deg = 180 - 27
+  D207_PRED,  // Directional 207 deg = 180 + 27
+  D63_PRED,   // Directional 63  deg = round(arctan(2/1) * 180/pi)
+#if CONFIG_ALT_INTRA
+  SMOOTH_PRED,  // Combination of horizontal and vertical interpolation
+#endif          // CONFIG_ALT_INTRA
+  TM_PRED,      // True-motion
+  NEARESTMV,
+  NEARMV,
+  ZEROMV,
+  NEWMV,
+#if CONFIG_EXT_INTER
+#if CONFIG_COMPOUND_SINGLEREF
+  // Single ref compound modes
+  SR_NEAREST_NEARMV,
+  SR_NEAREST_NEWMV,
+  SR_NEAR_NEWMV,
+  SR_ZERO_NEWMV,
+  SR_NEW_NEWMV,
+#endif  // CONFIG_COMPOUND_SINGLEREF
+  // Compound ref compound modes
+  NEAREST_NEARESTMV,
+  NEAREST_NEARMV,
+  NEAR_NEARESTMV,
+  NEAR_NEARMV,
+  NEAREST_NEWMV,
+  NEW_NEARESTMV,
+  NEAR_NEWMV,
+  NEW_NEARMV,
+  ZERO_ZEROMV,
+  NEW_NEWMV,
+#endif  // CONFIG_EXT_INTER
+  MB_MODE_COUNT,
+  INTRA_MODES = TM_PRED + 1,
+  INTRA_INVALID = MB_MODE_COUNT  // For uv_mode in inter blocks
+} PREDICTION_MODE;
+
+typedef enum {
+  SIMPLE_TRANSLATION = 0,
+#if CONFIG_MOTION_VAR
+  OBMC_CAUSAL,  // 2-sided OBMC
+#endif          // CONFIG_MOTION_VAR
+#if CONFIG_WARPED_MOTION
+  WARPED_CAUSAL,  // 2-sided WARPED
+#endif            // CONFIG_WARPED_MOTION
+  MOTION_MODES
+} MOTION_MODE;
+
+// TODO(urvang): Consider adding II_SMOOTH_PRED if it's helpful.
+
+#if CONFIG_EXT_INTER
+typedef enum {
+  II_DC_PRED = 0,
+  II_V_PRED,
+  II_H_PRED,
+  II_D45_PRED,
+  II_D135_PRED,
+  II_D117_PRED,
+  II_D153_PRED,
+  II_D207_PRED,
+  II_D63_PRED,
+  II_TM_PRED,
+  INTERINTRA_MODES
+} INTERINTRA_MODE;
+
+typedef enum {
+  COMPOUND_AVERAGE = 0,
+#if CONFIG_WEDGE
+  COMPOUND_WEDGE,
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+  COMPOUND_SEG,
+#endif  // CONFIG_COMPOUND_SEGMENT
+  COMPOUND_TYPES,
+} COMPOUND_TYPE;
+#endif  // CONFIG_EXT_INTER
+
+// TODO(huisu): Consider adding FILTER_SMOOTH_PRED to "FILTER_INTRA_MODE".
+#if CONFIG_FILTER_INTRA
+typedef enum {
+  FILTER_DC_PRED,
+  FILTER_V_PRED,
+  FILTER_H_PRED,
+  FILTER_D45_PRED,
+  FILTER_D135_PRED,
+  FILTER_D117_PRED,
+  FILTER_D153_PRED,
+  FILTER_D207_PRED,
+  FILTER_D63_PRED,
+  FILTER_TM_PRED,
+  FILTER_INTRA_MODES,
+} FILTER_INTRA_MODE;
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_EXT_INTRA
+#define DIRECTIONAL_MODES (INTRA_MODES - 2)
+#endif  // CONFIG_EXT_INTRA
+
+#define INTER_MODES (1 + NEWMV - NEARESTMV)
+
+#if CONFIG_EXT_INTER
+#if CONFIG_COMPOUND_SINGLEREF
+#define INTER_SINGLEREF_COMP_MODES (1 + SR_NEW_NEWMV - SR_NEAREST_NEARMV)
+#endif  // CONFIG_COMPOUND_SINGLEREF
+
+#define INTER_COMPOUND_MODES (1 + NEW_NEWMV - NEAREST_NEARESTMV)
+#endif  // CONFIG_EXT_INTER
+
+#define SKIP_CONTEXTS 3
+
+#if CONFIG_REF_MV
+#define NMV_CONTEXTS 3
+
+#define NEWMV_MODE_CONTEXTS 7
+#define ZEROMV_MODE_CONTEXTS 2
+#define REFMV_MODE_CONTEXTS 9
+#define DRL_MODE_CONTEXTS 5
+
+#define ZEROMV_OFFSET 3
+#define REFMV_OFFSET 4
+
+#define NEWMV_CTX_MASK ((1 << ZEROMV_OFFSET) - 1)
+#define ZEROMV_CTX_MASK ((1 << (REFMV_OFFSET - ZEROMV_OFFSET)) - 1)
+#define REFMV_CTX_MASK ((1 << (8 - REFMV_OFFSET)) - 1)
+
+#define ALL_ZERO_FLAG_OFFSET 8
+#define SKIP_NEARESTMV_OFFSET 9
+#define SKIP_NEARMV_OFFSET 10
+#define SKIP_NEARESTMV_SUB8X8_OFFSET 11
+#endif
+
+#define INTER_MODE_CONTEXTS 7
+#if CONFIG_DELTA_Q
+#define DELTA_Q_SMALL 3
+#define DELTA_Q_PROBS (DELTA_Q_SMALL)
+#define DEFAULT_DELTA_Q_RES 4
+#if CONFIG_EXT_DELTA_Q
+#define DELTA_LF_SMALL 3
+#define DELTA_LF_PROBS (DELTA_LF_SMALL)
+#define DEFAULT_DELTA_LF_RES 2
+#endif
+#endif
+
+/* Segment Feature Masks */
+#define MAX_MV_REF_CANDIDATES 2
+
+#if CONFIG_REF_MV
+#define MAX_REF_MV_STACK_SIZE 16
+#if CONFIG_EXT_PARTITION
+#define REF_CAT_LEVEL 640
+#else
+#define REF_CAT_LEVEL 255
+#endif  // CONFIG_EXT_PARTITION
+#endif  // CONFIG_REF_MV
+
+#define INTRA_INTER_CONTEXTS 4
+#define COMP_INTER_CONTEXTS 5
+#define REF_CONTEXTS 5
+#if CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+#define COMP_INTER_MODE_CONTEXTS 4
+#endif  // CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+
+#if CONFIG_VAR_TX
+#define TXFM_PARTITION_CONTEXTS ((TX_SIZES - TX_8X8) * 6 - 2)
+typedef uint8_t TXFM_CONTEXT;
+#endif
+
+#define NONE_FRAME -1
+#define INTRA_FRAME 0
+#define LAST_FRAME 1
+
+#if CONFIG_EXT_REFS
+#define LAST2_FRAME 2
+#define LAST3_FRAME 3
+#define GOLDEN_FRAME 4
+#define BWDREF_FRAME 5
+#define ALTREF_FRAME 6
+#define LAST_REF_FRAMES (LAST3_FRAME - LAST_FRAME + 1)
+#else
+#define GOLDEN_FRAME 2
+#define ALTREF_FRAME 3
+#endif  // CONFIG_EXT_REFS
+
+#define INTER_REFS_PER_FRAME (ALTREF_FRAME - LAST_FRAME + 1)
+#define TOTAL_REFS_PER_FRAME (ALTREF_FRAME - INTRA_FRAME + 1)
+
+#define FWD_REFS (GOLDEN_FRAME - LAST_FRAME + 1)
+#define FWD_RF_OFFSET(ref) (ref - LAST_FRAME)
+#if CONFIG_EXT_REFS
+#define BWD_REFS (ALTREF_FRAME - BWDREF_FRAME + 1)
+#define BWD_RF_OFFSET(ref) (ref - BWDREF_FRAME)
+#else
+#define BWD_REFS 1
+#define BWD_RF_OFFSET(ref) (ref - ALTREF_FRAME)
+#endif  // CONFIG_EXT_REFS
+
+#define SINGLE_REFS (FWD_REFS + BWD_REFS)
+#define COMP_REFS (FWD_REFS * BWD_REFS)
+
+#if CONFIG_REF_MV
+#define MODE_CTX_REF_FRAMES (TOTAL_REFS_PER_FRAME + COMP_REFS)
+#else
+#define MODE_CTX_REF_FRAMES TOTAL_REFS_PER_FRAME
+#endif
+
+#if CONFIG_SUPERTX
+#define PARTITION_SUPERTX_CONTEXTS 2
+#define MAX_SUPERTX_BLOCK_SIZE BLOCK_32X32
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_LOOP_RESTORATION
+typedef enum {
+  RESTORE_NONE = 0,
+  RESTORE_WIENER = 1,
+  RESTORE_SGRPROJ = 2,
+  RESTORE_SWITCHABLE,
+  RESTORE_SWITCHABLE_TYPES = RESTORE_SWITCHABLE,
+  RESTORE_TYPES,
+} RestorationType;
+#endif  // CONFIG_LOOP_RESTORATION
+
+#if CONFIG_FRAME_SUPERRES
+#define SUPERRES_SCALE_DENOMINATOR 16
+#define SUPERRES_SCALE_BITS 3
+#define SUPERRES_SCALE_NUMERATOR_MIN 8
+#endif  // CONFIG_FRAME_SUPERRES
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_ENUMS_H_
diff --git a/third_party/aom/av1/common/filter.c b/third_party/aom/av1/common/filter.c
new file mode 100644
index 0000000000..9f0c58866c
--- /dev/null
+++ b/third_party/aom/av1/common/filter.c
@@ -0,0 +1,360 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/filter.h"
+
+DECLARE_ALIGNED(256, static const InterpKernel,
+                bilinear_filters[SUBPEL_SHIFTS]) = {
+  { 0, 0, 0, 128, 0, 0, 0, 0 },  { 0, 0, 0, 120, 8, 0, 0, 0 },
+  { 0, 0, 0, 112, 16, 0, 0, 0 }, { 0, 0, 0, 104, 24, 0, 0, 0 },
+  { 0, 0, 0, 96, 32, 0, 0, 0 },  { 0, 0, 0, 88, 40, 0, 0, 0 },
+  { 0, 0, 0, 80, 48, 0, 0, 0 },  { 0, 0, 0, 72, 56, 0, 0, 0 },
+  { 0, 0, 0, 64, 64, 0, 0, 0 },  { 0, 0, 0, 56, 72, 0, 0, 0 },
+  { 0, 0, 0, 48, 80, 0, 0, 0 },  { 0, 0, 0, 40, 88, 0, 0, 0 },
+  { 0, 0, 0, 32, 96, 0, 0, 0 },  { 0, 0, 0, 24, 104, 0, 0, 0 },
+  { 0, 0, 0, 16, 112, 0, 0, 0 }, { 0, 0, 0, 8, 120, 0, 0, 0 }
+};
+
+#if USE_TEMPORALFILTER_12TAP
+DECLARE_ALIGNED(16, static const int16_t,
+                sub_pel_filters_temporalfilter_12[SUBPEL_SHIFTS][12]) = {
+  // intfilt 0.8
+  { 0, 0, 0, 0, 0, 128, 0, 0, 0, 0, 0, 0 },
+  { 0, 1, -1, 3, -7, 127, 8, -4, 2, -1, 0, 0 },
+  { 0, 1, -3, 5, -12, 124, 18, -8, 4, -2, 1, 0 },
+  { -1, 2, -4, 8, -17, 120, 28, -11, 6, -3, 1, -1 },
+  { -1, 2, -4, 10, -21, 114, 38, -15, 8, -4, 2, -1 },
+  { -1, 3, -5, 11, -23, 107, 49, -18, 9, -5, 2, -1 },
+  { -1, 3, -6, 12, -25, 99, 60, -21, 11, -6, 3, -1 },
+  { -1, 3, -6, 12, -25, 90, 70, -23, 12, -6, 3, -1 },
+  { -1, 3, -6, 12, -24, 80, 80, -24, 12, -6, 3, -1 },
+  { -1, 3, -6, 12, -23, 70, 90, -25, 12, -6, 3, -1 },
+  { -1, 3, -6, 11, -21, 60, 99, -25, 12, -6, 3, -1 },
+  { -1, 2, -5, 9, -18, 49, 107, -23, 11, -5, 3, -1 },
+  { -1, 2, -4, 8, -15, 38, 114, -21, 10, -4, 2, -1 },
+  { -1, 1, -3, 6, -11, 28, 120, -17, 8, -4, 2, -1 },
+  { 0, 1, -2, 4, -8, 18, 124, -12, 5, -3, 1, 0 },
+  { 0, 0, -1, 2, -4, 8, 127, -7, 3, -1, 1, 0 },
+};
+#endif  // USE_TEMPORALFILTER_12TAP
+
+#if CONFIG_DUAL_FILTER
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8[SUBPEL_SHIFTS]) = {
+  // intfilt 0.575
+  { 0, 0, 0, 128, 0, 0, 0, 0 },        { 0, 1, -5, 126, 8, -3, 1, 0 },
+  { -1, 3, -10, 123, 18, -6, 2, -1 },  { -1, 4, -14, 118, 27, -9, 3, 0 },
+  { -1, 5, -16, 112, 37, -12, 4, -1 }, { -1, 5, -18, 105, 48, -14, 4, -1 },
+  { -1, 6, -19, 97, 58, -17, 5, -1 },  { -1, 6, -20, 88, 68, -18, 6, -1 },
+  { -1, 6, -19, 78, 78, -19, 6, -1 },  { -1, 6, -18, 68, 88, -20, 6, -1 },
+  { -1, 5, -17, 58, 97, -19, 6, -1 },  { -1, 4, -14, 48, 105, -18, 5, -1 },
+  { -1, 4, -12, 37, 112, -16, 5, -1 }, { 0, 3, -9, 27, 118, -14, 4, -1 },
+  { -1, 2, -6, 18, 123, -10, 3, -1 },  { 0, 1, -3, 8, 126, -5, 1, 0 },
+};
+
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_regular_uv[SUBPEL_SHIFTS]) = {
+  // intfilt 0.575
+  { 0, 0, 0, 128, 0, 0, 0, 0 },        { 0, 1, -5, 126, 8, -3, 1, 0 },
+  { -1, 3, -10, 123, 18, -6, 2, -1 },  { -1, 4, -14, 118, 27, -9, 3, 0 },
+  { -1, 5, -16, 112, 37, -12, 4, -1 }, { -1, 5, -18, 105, 48, -14, 4, -1 },
+  { -1, 6, -19, 97, 58, -17, 5, -1 },  { -1, 6, -20, 88, 68, -18, 6, -1 },
+  { -1, 6, -19, 78, 78, -19, 6, -1 },  { -1, 6, -18, 68, 88, -20, 6, -1 },
+  { -1, 5, -17, 58, 97, -19, 6, -1 },  { -1, 4, -14, 48, 105, -18, 5, -1 },
+  { -1, 4, -12, 37, 112, -16, 5, -1 }, { 0, 3, -9, 27, 118, -14, 4, -1 },
+  { -1, 2, -6, 18, 123, -10, 3, -1 },  { 0, 1, -3, 8, 126, -5, 1, 0 },
+};
+
+#if USE_12TAP_FILTER
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8sharp[SUBPEL_SHIFTS]) = {
+  // intfilt 0.8
+  { 0, 0, 0, 128, 0, 0, 0, 0 },         { -1, 2, -6, 127, 9, -4, 2, -1 },
+  { -2, 5, -12, 124, 18, -7, 4, -2 },   { -2, 7, -16, 119, 28, -11, 5, -2 },
+  { -3, 8, -19, 114, 38, -14, 7, -3 },  { -3, 9, -22, 107, 49, -17, 8, -3 },
+  { -4, 10, -23, 99, 60, -20, 10, -4 }, { -4, 11, -23, 90, 70, -22, 10, -4 },
+  { -4, 11, -23, 80, 80, -23, 11, -4 }, { -4, 10, -22, 70, 90, -23, 11, -4 },
+  { -4, 10, -20, 60, 99, -23, 10, -4 }, { -3, 8, -17, 49, 107, -22, 9, -3 },
+  { -3, 7, -14, 38, 114, -19, 8, -3 },  { -2, 5, -11, 28, 119, -16, 7, -2 },
+  { -2, 4, -7, 18, 124, -12, 5, -2 },   { -1, 2, -4, 9, 127, -6, 2, -1 },
+};
+
+DECLARE_ALIGNED(256, static const int16_t,
+                sub_pel_filters_10sharp[SUBPEL_SHIFTS][12]) = {
+  // intfilt 0.85
+  { 0, 0, 0, 0, 0, 128, 0, 0, 0, 0, 0, 0 },
+  { 0, 1, -2, 3, -7, 127, 8, -4, 2, -1, 1, 0 },
+  { 0, 1, -3, 6, -13, 124, 18, -8, 4, -2, 1, 0 },
+  { 0, 2, -4, 8, -18, 120, 28, -12, 6, -4, 2, 0 },
+  { 0, 2, -5, 10, -21, 114, 38, -15, 8, -5, 2, 0 },
+  { 0, 3, -6, 11, -24, 107, 49, -19, 10, -6, 3, 0 },
+  { 0, 3, -7, 12, -25, 99, 59, -21, 11, -6, 3, 0 },
+  { 0, 3, -7, 12, -25, 90, 70, -23, 12, -7, 3, 0 },
+  { 0, 3, -7, 12, -25, 81, 81, -25, 12, -7, 3, 0 },
+  { 0, 3, -7, 12, -23, 70, 90, -25, 12, -7, 3, 0 },
+  { 0, 3, -6, 11, -21, 59, 99, -25, 12, -7, 3, 0 },
+  { 0, 3, -6, 10, -19, 49, 107, -24, 11, -6, 3, 0 },
+  { 0, 2, -5, 8, -15, 38, 114, -21, 10, -5, 2, 0 },
+  { 0, 2, -4, 6, -12, 28, 120, -18, 8, -4, 2, 0 },
+  { 0, 1, -2, 4, -8, 18, 124, -13, 6, -3, 1, 0 },
+  { 0, 1, -1, 2, -4, 8, 127, -7, 3, -2, 1, 0 },
+};
+#else
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8sharp[SUBPEL_SHIFTS]) = {
+#if CONFIG_FILTER_7BIT
+  { 0, 0, 0, 128, 0, 0, 0, 0 },         { -2, 2, -6, 126, 8, -2, 2, 0 },
+  { -2, 6, -12, 124, 16, -6, 4, -2 },   { -2, 8, -18, 120, 26, -10, 6, -2 },
+  { -4, 10, -22, 116, 38, -14, 6, -2 }, { -4, 10, -22, 108, 48, -18, 8, -2 },
+  { -4, 10, -24, 100, 60, -20, 8, -2 }, { -4, 10, -24, 90, 70, -22, 10, -2 },
+  { -4, 12, -24, 80, 80, -24, 12, -4 }, { -2, 10, -22, 70, 90, -24, 10, -4 },
+  { -2, 8, -20, 60, 100, -24, 10, -4 }, { -2, 8, -18, 48, 108, -22, 10, -4 },
+  { -2, 6, -14, 38, 116, -22, 10, -4 }, { -2, 6, -10, 26, 120, -18, 8, -2 },
+  { -2, 4, -6, 16, 124, -12, 6, -2 },   { 0, 2, -2, 8, 126, -6, 2, -2 }
+#else
+  { 0, 0, 0, 128, 0, 0, 0, 0 },         { -1, 3, -7, 127, 8, -3, 1, 0 },
+  { -2, 5, -13, 125, 17, -6, 3, -1 },   { -3, 7, -17, 121, 27, -10, 5, -2 },
+  { -4, 9, -20, 115, 37, -13, 6, -2 },  { -4, 10, -23, 108, 48, -16, 8, -3 },
+  { -4, 10, -24, 100, 59, -19, 9, -3 }, { -4, 11, -24, 90, 70, -21, 10, -4 },
+  { -4, 11, -23, 80, 80, -23, 11, -4 }, { -4, 10, -21, 70, 90, -24, 11, -4 },
+  { -3, 9, -19, 59, 100, -24, 10, -4 }, { -3, 8, -16, 48, 108, -23, 10, -4 },
+  { -2, 6, -13, 37, 115, -20, 9, -4 },  { -2, 5, -10, 27, 121, -17, 7, -3 },
+  { -1, 3, -6, 17, 125, -13, 5, -2 },   { 0, 1, -3, 8, 127, -7, 3, -1 }
+#endif
+};
+#endif
+
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8smooth2[SUBPEL_SHIFTS]) = {
+  // freqmultiplier = 0.2
+  { 0, 0, 0, 128, 0, 0, 0, 0 },   { 0, 9, 30, 44, 32, 11, 2, 0 },
+  { 0, 8, 28, 44, 34, 12, 2, 0 }, { 0, 7, 27, 44, 35, 13, 2, 0 },
+  { 0, 6, 26, 43, 37, 14, 2, 0 }, { 0, 5, 24, 43, 38, 16, 2, 0 },
+  { 0, 5, 23, 42, 38, 17, 3, 0 }, { 0, 4, 21, 41, 40, 19, 3, 0 },
+  { 0, 4, 20, 40, 40, 20, 4, 0 }, { 0, 3, 19, 40, 41, 21, 4, 0 },
+  { 0, 3, 17, 38, 42, 23, 5, 0 }, { 0, 2, 16, 38, 43, 24, 5, 0 },
+  { 0, 2, 14, 37, 43, 26, 6, 0 }, { 0, 2, 13, 35, 44, 27, 7, 0 },
+  { 0, 2, 12, 34, 44, 28, 8, 0 }, { 0, 2, 11, 32, 44, 30, 9, 0 },
+};
+
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_smooth2_uv[SUBPEL_SHIFTS]) = {
+  // freqmultiplier = 0.2
+  { 0, 0, 0, 128, 0, 0, 0, 0 },   { 0, 9, 30, 44, 32, 11, 2, 0 },
+  { 0, 8, 28, 44, 34, 12, 2, 0 }, { 0, 7, 27, 44, 35, 13, 2, 0 },
+  { 0, 6, 26, 43, 37, 14, 2, 0 }, { 0, 5, 24, 43, 38, 16, 2, 0 },
+  { 0, 5, 23, 42, 38, 17, 3, 0 }, { 0, 4, 21, 41, 40, 19, 3, 0 },
+  { 0, 4, 20, 40, 40, 20, 4, 0 }, { 0, 3, 19, 40, 41, 21, 4, 0 },
+  { 0, 3, 17, 38, 42, 23, 5, 0 }, { 0, 2, 16, 38, 43, 24, 5, 0 },
+  { 0, 2, 14, 37, 43, 26, 6, 0 }, { 0, 2, 13, 35, 44, 27, 7, 0 },
+  { 0, 2, 12, 34, 44, 28, 8, 0 }, { 0, 2, 11, 32, 44, 30, 9, 0 },
+};
+
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8smooth[SUBPEL_SHIFTS]) = {
+  // freqmultiplier = 0.8
+  { 0, 0, 0, 128, 0, 0, 0, 0 },    { 0, -5, 13, 102, 24, -7, 1, 0 },
+  { 0, -4, 8, 100, 31, -8, 1, 0 }, { 0, -3, 4, 97, 37, -8, 1, 0 },
+  { 0, -2, 0, 94, 44, -9, 1, 0 },  { 0, -2, -3, 90, 51, -9, 1, 0 },
+  { 0, -1, -5, 84, 59, -9, 0, 0 }, { 0, 0, -7, 79, 65, -9, 0, 0 },
+  { 0, 0, -8, 72, 72, -8, 0, 0 },  { 0, 0, -9, 65, 79, -7, 0, 0 },
+  { 0, 0, -9, 59, 84, -5, -1, 0 }, { 0, 1, -9, 51, 90, -3, -2, 0 },
+  { 0, 1, -9, 44, 94, 0, -2, 0 },  { 0, 1, -8, 37, 97, 4, -3, 0 },
+  { 0, 1, -8, 31, 100, 8, -4, 0 }, { 0, 1, -7, 24, 102, 13, -5, 0 },
+};
+
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_smooth_uv[SUBPEL_SHIFTS]) = {
+  // freqmultiplier = 0.8
+  { 0, 0, 0, 128, 0, 0, 0, 0 },    { 0, -5, 13, 102, 24, -7, 1, 0 },
+  { 0, -4, 8, 100, 31, -8, 1, 0 }, { 0, -3, 4, 97, 37, -8, 1, 0 },
+  { 0, -2, 0, 94, 44, -9, 1, 0 },  { 0, -2, -3, 90, 51, -9, 1, 0 },
+  { 0, -1, -5, 84, 59, -9, 0, 0 }, { 0, 0, -7, 79, 65, -9, 0, 0 },
+  { 0, 0, -8, 72, 72, -8, 0, 0 },  { 0, 0, -9, 65, 79, -7, 0, 0 },
+  { 0, 0, -9, 59, 84, -5, -1, 0 }, { 0, 1, -9, 51, 90, -3, -2, 0 },
+  { 0, 1, -9, 44, 94, 0, -2, 0 },  { 0, 1, -8, 37, 97, 4, -3, 0 },
+  { 0, 1, -8, 31, 100, 8, -4, 0 }, { 0, 1, -7, 24, 102, 13, -5, 0 },
+};
+#else  // CONFIG_DUAL_FILTER
+
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8[SUBPEL_SHIFTS]) = {
+#if CONFIG_FILTER_7BIT
+  { 0, 0, 0, 128, 0, 0, 0, 0 },      { 0, 2, -6, 126, 8, -2, 0, 0 },
+  { 0, 2, -10, 122, 18, -4, 0, 0 },  { 0, 2, -12, 116, 28, -8, 2, 0 },
+  { 0, 2, -14, 110, 38, -10, 2, 0 }, { 0, 2, -14, 102, 48, -12, 2, 0 },
+  { 0, 2, -16, 94, 58, -12, 2, 0 },  { 0, 2, -14, 84, 66, -12, 2, 0 },
+  { 0, 2, -14, 76, 76, -14, 2, 0 },  { 0, 2, -12, 66, 84, -14, 2, 0 },
+  { 0, 2, -12, 58, 94, -16, 2, 0 },  { 0, 2, -12, 48, 102, -14, 2, 0 },
+  { 0, 2, -10, 38, 110, -14, 2, 0 }, { 0, 2, -8, 28, 116, -12, 2, 0 },
+  { 0, 0, -4, 18, 122, -10, 2, 0 },  { 0, 0, -2, 8, 126, -6, 2, 0 }
+#else
+  { 0, 0, 0, 128, 0, 0, 0, 0 },        { 0, 1, -5, 126, 8, -3, 1, 0 },
+  { -1, 3, -10, 122, 18, -6, 2, 0 },   { -1, 4, -13, 118, 27, -9, 3, -1 },
+  { -1, 4, -16, 112, 37, -11, 4, -1 }, { -1, 5, -18, 105, 48, -14, 4, -1 },
+  { -1, 5, -19, 97, 58, -16, 5, -1 },  { -1, 6, -19, 88, 68, -18, 5, -1 },
+  { -1, 6, -19, 78, 78, -19, 6, -1 },  { -1, 5, -18, 68, 88, -19, 6, -1 },
+  { -1, 5, -16, 58, 97, -19, 5, -1 },  { -1, 4, -14, 48, 105, -18, 5, -1 },
+  { -1, 4, -11, 37, 112, -16, 4, -1 }, { -1, 3, -9, 27, 118, -13, 4, -1 },
+  { 0, 2, -6, 18, 122, -10, 3, -1 },   { 0, 1, -3, 8, 126, -5, 1, 0 }
+#endif
+};
+
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8sharp[SUBPEL_SHIFTS]) = {
+#if CONFIG_FILTER_7BIT
+  { 0, 0, 0, 128, 0, 0, 0, 0 },         { -2, 2, -6, 126, 8, -2, 2, 0 },
+  { -2, 6, -12, 124, 16, -6, 4, -2 },   { -2, 8, -18, 120, 26, -10, 6, -2 },
+  { -4, 10, -22, 116, 38, -14, 6, -2 }, { -4, 10, -22, 108, 48, -18, 8, -2 },
+  { -4, 10, -24, 100, 60, -20, 8, -2 }, { -4, 10, -24, 90, 70, -22, 10, -2 },
+  { -4, 12, -24, 80, 80, -24, 12, -4 }, { -2, 10, -22, 70, 90, -24, 10, -4 },
+  { -2, 8, -20, 60, 100, -24, 10, -4 }, { -2, 8, -18, 48, 108, -22, 10, -4 },
+  { -2, 6, -14, 38, 116, -22, 10, -4 }, { -2, 6, -10, 26, 120, -18, 8, -2 },
+  { -2, 4, -6, 16, 124, -12, 6, -2 },   { 0, 2, -2, 8, 126, -6, 2, -2 }
+#else
+  { 0, 0, 0, 128, 0, 0, 0, 0 },         { -1, 3, -7, 127, 8, -3, 1, 0 },
+  { -2, 5, -13, 125, 17, -6, 3, -1 },   { -3, 7, -17, 121, 27, -10, 5, -2 },
+  { -4, 9, -20, 115, 37, -13, 6, -2 },  { -4, 10, -23, 108, 48, -16, 8, -3 },
+  { -4, 10, -24, 100, 59, -19, 9, -3 }, { -4, 11, -24, 90, 70, -21, 10, -4 },
+  { -4, 11, -23, 80, 80, -23, 11, -4 }, { -4, 10, -21, 70, 90, -24, 11, -4 },
+  { -3, 9, -19, 59, 100, -24, 10, -4 }, { -3, 8, -16, 48, 108, -23, 10, -4 },
+  { -2, 6, -13, 37, 115, -20, 9, -4 },  { -2, 5, -10, 27, 121, -17, 7, -3 },
+  { -1, 3, -6, 17, 125, -13, 5, -2 },   { 0, 1, -3, 8, 127, -7, 3, -1 }
+#endif
+};
+
+DECLARE_ALIGNED(256, static const InterpKernel,
+                sub_pel_filters_8smooth[SUBPEL_SHIFTS]) = {
+#if CONFIG_FILTER_7BIT
+  { 0, 0, 0, 128, 0, 0, 0, 0 },     { 0, 2, 28, 62, 34, 2, 0, 0 },
+  { 0, 0, 26, 62, 36, 4, 0, 0 },    { 0, 0, 22, 62, 40, 4, 0, 0 },
+  { 0, 0, 20, 60, 42, 6, 0, 0 },    { 0, 0, 18, 58, 44, 8, 0, 0 },
+  { 0, 0, 16, 56, 46, 10, 0, 0 },   { 0, -2, 16, 54, 48, 12, 0, 0 },
+  { 0, -2, 14, 52, 52, 14, -2, 0 }, { 0, 0, 12, 48, 54, 16, -2, 0 },
+  { 0, 0, 10, 46, 56, 16, 0, 0 },   { 0, 0, 8, 44, 58, 18, 0, 0 },
+  { 0, 0, 6, 42, 60, 20, 0, 0 },    { 0, 0, 4, 40, 62, 22, 0, 0 },
+  { 0, 0, 4, 36, 62, 26, 0, 0 },    { 0, 0, 2, 34, 62, 28, 2, 0 }
+#else
+  { 0, 0, 0, 128, 0, 0, 0, 0 },       { -3, -1, 32, 64, 38, 1, -3, 0 },
+  { -2, -2, 29, 63, 41, 2, -3, 0 },   { -2, -2, 26, 63, 43, 4, -4, 0 },
+  { -2, -3, 24, 62, 46, 5, -4, 0 },   { -2, -3, 21, 60, 49, 7, -4, 0 },
+  { -1, -4, 18, 59, 51, 9, -4, 0 },   { -1, -4, 16, 57, 53, 12, -4, -1 },
+  { -1, -4, 14, 55, 55, 14, -4, -1 }, { -1, -4, 12, 53, 57, 16, -4, -1 },
+  { 0, -4, 9, 51, 59, 18, -4, -1 },   { 0, -4, 7, 49, 60, 21, -3, -2 },
+  { 0, -4, 5, 46, 62, 24, -3, -2 },   { 0, -4, 4, 43, 63, 26, -2, -2 },
+  { 0, -3, 2, 41, 63, 29, -2, -2 },   { 0, -3, 1, 38, 64, 32, -1, -3 }
+#endif
+};
+#endif  // CONFIG_DUAL_FILTER
+
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+const InterpKernel *av1_intra_filter_kernels[INTRA_FILTERS] = {
+  bilinear_filters,         // INTRA_FILTER_LINEAR
+  sub_pel_filters_8,        // INTRA_FILTER_8TAP
+  sub_pel_filters_8sharp,   // INTRA_FILTER_8TAP_SHARP
+  sub_pel_filters_8smooth,  // INTRA_FILTER_8TAP_SMOOTH
+};
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+
+#if CONFIG_DUAL_FILTER
+static const InterpFilterParams
+    av1_interp_filter_params_list[SWITCHABLE_FILTERS + EXTRA_FILTERS] = {
+      { (const int16_t *)sub_pel_filters_8, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        EIGHTTAP_REGULAR },
+      { (const int16_t *)sub_pel_filters_8smooth, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        EIGHTTAP_SMOOTH },
+#if USE_12TAP_FILTER
+      { (const int16_t *)sub_pel_filters_10sharp, 12, SUBPEL_SHIFTS,
+        MULTITAP_SHARP },
+#else
+      { (const int16_t *)sub_pel_filters_8sharp, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        EIGHTTAP_SHARP },
+#endif
+      { (const int16_t *)sub_pel_filters_8smooth2, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        EIGHTTAP_SMOOTH2 },
+      { (const int16_t *)bilinear_filters, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        BILINEAR },
+      { (const int16_t *)sub_pel_filters_8sharp, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        EIGHTTAP_SHARP },
+      { (const int16_t *)sub_pel_filters_regular_uv, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        FILTER_REGULAR_UV },
+      { (const int16_t *)sub_pel_filters_smooth_uv, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        FILTER_SMOOTH_UV },
+      { (const int16_t *)sub_pel_filters_8sharp, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        FILTER_SHARP_UV },
+      { (const int16_t *)sub_pel_filters_smooth2_uv, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        FILTER_SMOOTH2_UV },
+    };
+#else
+static const InterpFilterParams
+    av1_interp_filter_params_list[SWITCHABLE_FILTERS + 1] = {
+      { (const int16_t *)sub_pel_filters_8, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        EIGHTTAP_REGULAR },
+      { (const int16_t *)sub_pel_filters_8smooth, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        EIGHTTAP_SMOOTH },
+      { (const int16_t *)sub_pel_filters_8sharp, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        MULTITAP_SHARP },
+      { (const int16_t *)bilinear_filters, SUBPEL_TAPS, SUBPEL_SHIFTS,
+        BILINEAR }
+    };
+#endif  // CONFIG_DUAL_FILTER
+
+#if USE_TEMPORALFILTER_12TAP
+static const InterpFilterParams av1_interp_temporalfilter_12tap = {
+  (const int16_t *)sub_pel_filters_temporalfilter_12, 12, SUBPEL_SHIFTS,
+  TEMPORALFILTER_12TAP
+};
+#endif  // USE_TEMPORALFILTER_12TAP
+
+InterpFilterParams av1_get_interp_filter_params(
+    const InterpFilter interp_filter) {
+#if USE_TEMPORALFILTER_12TAP
+  if (interp_filter == TEMPORALFILTER_12TAP)
+    return av1_interp_temporalfilter_12tap;
+#endif  // USE_TEMPORALFILTER_12TAP
+  return av1_interp_filter_params_list[interp_filter];
+}
+
+const int16_t *av1_get_interp_filter_kernel(const InterpFilter interp_filter) {
+#if USE_TEMPORALFILTER_12TAP
+  if (interp_filter == TEMPORALFILTER_12TAP)
+    return av1_interp_temporalfilter_12tap.filter_ptr;
+#endif  // USE_TEMPORALFILTER_12TAP
+  return (const int16_t *)av1_interp_filter_params_list[interp_filter]
+      .filter_ptr;
+}
+
+#if CONFIG_DUAL_FILTER
+InterpFilter av1_get_plane_interp_filter(InterpFilter interp_filter,
+                                         int plane) {
+#if USE_TEMPORALFILTER_12TAP
+  assert(interp_filter <= EIGHTTAP_SHARP ||
+         interp_filter == TEMPORALFILTER_12TAP);
+#else
+  assert(interp_filter <= EIGHTTAP_SHARP);
+#endif
+  if (plane == 0) {
+    return interp_filter;
+  } else {
+    switch (interp_filter) {
+      case EIGHTTAP_REGULAR: return FILTER_REGULAR_UV;
+      case EIGHTTAP_SMOOTH: return FILTER_SMOOTH_UV;
+      case MULTITAP_SHARP: return FILTER_SHARP_UV;
+      case EIGHTTAP_SMOOTH2: return FILTER_SMOOTH2_UV;
+      default: return interp_filter;
+    }
+  }
+}
+#endif
diff --git a/third_party/aom/av1/common/filter.h b/third_party/aom/av1/common/filter.h
new file mode 100644
index 0000000000..693a46902d
--- /dev/null
+++ b/third_party/aom/av1/common/filter.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_FILTER_H_
+#define AV1_COMMON_FILTER_H_
+
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define USE_TEMPORALFILTER_12TAP 1
+#define MAX_FILTER_TAP 12
+
+#define USE_12TAP_FILTER 0
+
+typedef enum {
+  EIGHTTAP_REGULAR,
+  EIGHTTAP_SMOOTH,
+  MULTITAP_SHARP,
+#if CONFIG_DUAL_FILTER
+  EIGHTTAP_SMOOTH2,
+#endif  // CONFIG_DUAL_FILTER
+  BILINEAR,
+#if CONFIG_DUAL_FILTER
+  EIGHTTAP_SHARP,
+  FILTER_REGULAR_UV,
+  FILTER_SMOOTH_UV,
+  FILTER_SHARP_UV,
+  FILTER_SMOOTH2_UV,
+#endif  // CONFIG_DUAL_FILTER
+  INTERP_FILTERS_ALL,
+  SWITCHABLE_FILTERS = BILINEAR,
+  SWITCHABLE = SWITCHABLE_FILTERS + 1, /* the last switchable one */
+  EXTRA_FILTERS = INTERP_FILTERS_ALL - SWITCHABLE_FILTERS,
+#if USE_TEMPORALFILTER_12TAP
+  TEMPORALFILTER_12TAP = SWITCHABLE_FILTERS + EXTRA_FILTERS,
+#endif
+} InterpFilter;
+
+#if CONFIG_DUAL_FILTER
+#define MAX_SUBPEL_TAPS 12
+#define LOG_SWITCHABLE_FILTERS \
+  3 /* (1 << LOG_SWITCHABLE_FILTERS) > SWITCHABLE_FILTERS */
+#define SWITCHABLE_FILTER_CONTEXTS ((SWITCHABLE_FILTERS + 1) * 4)
+#define INTER_FILTER_COMP_OFFSET (SWITCHABLE_FILTERS + 1)
+#define INTER_FILTER_DIR_OFFSET ((SWITCHABLE_FILTERS + 1) * 2)
+#else  // CONFIG_DUAL_FILTER
+#define LOG_SWITCHABLE_FILTERS \
+  2 /* (1 << LOG_SWITCHABLE_FILTERS) > SWITCHABLE_FILTERS */
+#define SWITCHABLE_FILTER_CONTEXTS (SWITCHABLE_FILTERS + 1)
+#endif  // CONFIG_DUAL_FILTER
+
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+typedef enum {
+  INTRA_FILTER_LINEAR,
+  INTRA_FILTER_8TAP,
+  INTRA_FILTER_8TAP_SHARP,
+  INTRA_FILTER_8TAP_SMOOTH,
+  INTRA_FILTERS,
+} INTRA_FILTER;
+
+extern const InterpKernel *av1_intra_filter_kernels[INTRA_FILTERS];
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+
+typedef struct InterpFilterParams {
+  const int16_t *filter_ptr;
+  uint16_t taps;
+  uint16_t subpel_shifts;
+  InterpFilter interp_filter;
+} InterpFilterParams;
+
+InterpFilterParams av1_get_interp_filter_params(
+    const InterpFilter interp_filter);
+
+const int16_t *av1_get_interp_filter_kernel(const InterpFilter interp_filter);
+
+static INLINE const int16_t *av1_get_interp_filter_subpel_kernel(
+    const InterpFilterParams filter_params, const int subpel) {
+  return filter_params.filter_ptr + filter_params.taps * subpel;
+}
+
+static INLINE int av1_is_interpolating_filter(
+    const InterpFilter interp_filter) {
+  const InterpFilterParams ip = av1_get_interp_filter_params(interp_filter);
+  return (ip.filter_ptr[ip.taps / 2 - 1] == 128);
+}
+
+#if CONFIG_DUAL_FILTER
+InterpFilter av1_get_plane_interp_filter(InterpFilter interp_filter, int plane);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_FILTER_H_
diff --git a/third_party/aom/av1/common/frame_buffers.c b/third_party/aom/av1/common/frame_buffers.c
new file mode 100644
index 0000000000..0b6b78e3d4
--- /dev/null
+++ b/third_party/aom/av1/common/frame_buffers.c
@@ -0,0 +1,79 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/frame_buffers.h"
+#include "aom_mem/aom_mem.h"
+
+int av1_alloc_internal_frame_buffers(InternalFrameBufferList *list) {
+  assert(list != NULL);
+  av1_free_internal_frame_buffers(list);
+
+  list->num_internal_frame_buffers =
+      AOM_MAXIMUM_REF_BUFFERS + AOM_MAXIMUM_WORK_BUFFERS;
+  list->int_fb = (InternalFrameBuffer *)aom_calloc(
+      list->num_internal_frame_buffers, sizeof(*list->int_fb));
+  return (list->int_fb == NULL);
+}
+
+void av1_free_internal_frame_buffers(InternalFrameBufferList *list) {
+  int i;
+
+  assert(list != NULL);
+
+  for (i = 0; i < list->num_internal_frame_buffers; ++i) {
+    aom_free(list->int_fb[i].data);
+    list->int_fb[i].data = NULL;
+  }
+  aom_free(list->int_fb);
+  list->int_fb = NULL;
+}
+
+int av1_get_frame_buffer(void *cb_priv, size_t min_size,
+                         aom_codec_frame_buffer_t *fb) {
+  int i;
+  InternalFrameBufferList *const int_fb_list =
+      (InternalFrameBufferList *)cb_priv;
+  if (int_fb_list == NULL) return -1;
+
+  // Find a free frame buffer.
+  for (i = 0; i < int_fb_list->num_internal_frame_buffers; ++i) {
+    if (!int_fb_list->int_fb[i].in_use) break;
+  }
+
+  if (i == int_fb_list->num_internal_frame_buffers) return -1;
+
+  if (int_fb_list->int_fb[i].size < min_size) {
+    aom_free(int_fb_list->int_fb[i].data);
+    // The data must be zeroed to fix a valgrind error from the C loop filter
+    // due to access uninitialized memory in frame border. It could be
+    // skipped if border were totally removed.
+    int_fb_list->int_fb[i].data = (uint8_t *)aom_calloc(1, min_size);
+    if (!int_fb_list->int_fb[i].data) return -1;
+    int_fb_list->int_fb[i].size = min_size;
+  }
+
+  fb->data = int_fb_list->int_fb[i].data;
+  fb->size = int_fb_list->int_fb[i].size;
+  int_fb_list->int_fb[i].in_use = 1;
+
+  // Set the frame buffer's private data to point at the internal frame buffer.
+  fb->priv = &int_fb_list->int_fb[i];
+  return 0;
+}
+
+int av1_release_frame_buffer(void *cb_priv, aom_codec_frame_buffer_t *fb) {
+  InternalFrameBuffer *const int_fb = (InternalFrameBuffer *)fb->priv;
+  (void)cb_priv;
+  if (int_fb) int_fb->in_use = 0;
+  return 0;
+}
diff --git a/third_party/aom/av1/common/frame_buffers.h b/third_party/aom/av1/common/frame_buffers.h
new file mode 100644
index 0000000000..e7341cfdd3
--- /dev/null
+++ b/third_party/aom/av1/common/frame_buffers.h
@@ -0,0 +1,54 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_FRAME_BUFFERS_H_
+#define AV1_COMMON_FRAME_BUFFERS_H_
+
+#include "aom/aom_frame_buffer.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct InternalFrameBuffer {
+  uint8_t *data;
+  size_t size;
+  int in_use;
+} InternalFrameBuffer;
+
+typedef struct InternalFrameBufferList {
+  int num_internal_frame_buffers;
+  InternalFrameBuffer *int_fb;
+} InternalFrameBufferList;
+
+// Initializes |list|. Returns 0 on success.
+int av1_alloc_internal_frame_buffers(InternalFrameBufferList *list);
+
+// Free any data allocated to the frame buffers.
+void av1_free_internal_frame_buffers(InternalFrameBufferList *list);
+
+// Callback used by libaom to request an external frame buffer. |cb_priv|
+// Callback private data, which points to an InternalFrameBufferList.
+// |min_size| is the minimum size in bytes needed to decode the next frame.
+// |fb| pointer to the frame buffer.
+int av1_get_frame_buffer(void *cb_priv, size_t min_size,
+                         aom_codec_frame_buffer_t *fb);
+
+// Callback used by libaom when there are no references to the frame buffer.
+// |cb_priv| is not used. |fb| pointer to the frame buffer.
+int av1_release_frame_buffer(void *cb_priv, aom_codec_frame_buffer_t *fb);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_FRAME_BUFFERS_H_
diff --git a/third_party/aom/av1/common/generic_code.c b/third_party/aom/av1/common/generic_code.c
new file mode 100644
index 0000000000..2955a695f5
--- /dev/null
+++ b/third_party/aom/av1/common/generic_code.c
@@ -0,0 +1,114 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include "generic_code.h"
+
+void aom_cdf_init_q15_1D(uint16_t *cdf, int nsyms, int cdf_size) {
+  int i;
+  for (i = 0; i < nsyms; i++)
+    cdf[i] = AOM_ICDF((i + 1)*CDF_PROB_TOP/nsyms);
+
+#if CONFIG_EC_ADAPT
+  cdf[cdf_size - 1] = 0;
+#endif
+}
+
+/** Adapts a Q15 cdf after encoding/decoding a symbol. */
+void aom_cdf_adapt_q15(int val, uint16_t *cdf, int n, int *count, int rate) {
+  int i;
+  *count = OD_MINI(*count + 1, 1 << rate);
+  OD_ASSERT(AOM_ICDF(cdf[n - 1]) == 32768);
+  if (*count >= 1 << rate) {
+    /* Steady-state adaptation based on a simple IIR with dyadic rate. */
+    for (i = 0; i < n; i++) {
+      int tmp;
+      /* When (i < val), we want the adjustment ((cdf[i] - tmp) >> rate) to be
+         positive so long as (cdf[i] > i + 1), and 0 when (cdf[i] == i + 1),
+         to ensure we don't drive any probabilities to 0. Replacing cdf[i] with
+         (i + 2) and solving ((i + 2 - tmp) >> rate == 1) for tmp produces
+         tmp == i + 2 - (1 << rate). Using this value of tmp with
+         cdf[i] == i + 1 instead gives an adjustment of 0 as desired.
+
+         When (i >= val), we want ((cdf[i] - tmp) >> rate) to be negative so
+         long as cdf[i] < 32768 - (n - 1 - i), and 0 when
+         cdf[i] == 32768 - (n - 1 - i), again to ensure we don't drive any
+         probabilities to 0. Since right-shifting any negative value is still
+         negative, we can solve (32768 - (n - 1 - i) - tmp == 0) for tmp,
+         producing tmp = 32769 - n + i. Using this value of tmp with smaller
+         values of cdf[i] instead gives negative adjustments, as desired.
+
+         Combining the two cases gives the expression below. These could be
+         stored in a lookup table indexed by n and rate to avoid the
+         arithmetic. */
+      tmp = 2 - (1<<rate) + i + (32767 + (1<<rate) - n)*(i >= val);
+      cdf[i] = AOM_ICDF(AOM_ICDF(cdf[i]) - ((AOM_ICDF(cdf[i]) - tmp) >> rate));
+    }
+  }
+  else {
+    int alpha;
+    /* Initial adaptation for the first symbols. The adaptation rate is
+       computed to be equivalent to what od_{en,de}code_cdf_adapt() does
+       when the initial cdf is set to increment/4. */
+    alpha = 4*32768/(n + 4**count);
+    for (i = 0; i < n; i++) {
+      int tmp;
+      tmp = (32768 - n)*(i >= val) + i + 1;
+      cdf[i] = AOM_ICDF(AOM_ICDF(cdf[i])
+          - (((AOM_ICDF(cdf[i]) - tmp)*alpha) >> 15));
+    }
+  }
+  OD_ASSERT(AOM_ICDF(cdf[n - 1]) == 32768);
+}
+
+/** Takes the base-2 log of E(x) in Q1.
+ *
+ * @param [in] ExQ16 expectation of x in Q16
+ *
+ * @retval 2*log2(ExQ16/2^16)
+ */
+int log_ex(int ex_q16) {
+  int lg;
+  int lg_q1;
+  int odd;
+  lg = OD_ILOG(ex_q16);
+  if (lg < 15) {
+    odd = ex_q16*ex_q16 > 2 << 2*lg;
+  }
+  else {
+    int tmp;
+    tmp = ex_q16 >> (lg - 8);
+    odd = tmp*tmp > (1 << 15);
+  }
+  lg_q1 = OD_MAXI(0, 2*lg - 33 + odd);
+  return lg_q1;
+}
+
+/** Updates the probability model based on the encoded/decoded value
+ *
+ * @param [in,out] model generic prob model
+ * @param [in,out] ExQ16 expectation of x
+ * @param [in]     x     variable encoded/decoded (used for ExQ16)
+ * @param [in]     xs    variable x after shift (used for the model)
+ * @param [in]     id    id of the icdf to adapt
+ * @param [in]     integration integration period of ExQ16 (leaky average over
+ * 1<<integration samples)
+ */
+void generic_model_update(int *ex_q16, int x, int integration) {
+  /* We could have saturated ExQ16 directly, but this is safe and simpler */
+  x = OD_MINI(x, 32767);
+  OD_IIR_DIADIC(*ex_q16, x << 16, integration);
+}
diff --git a/third_party/aom/av1/common/generic_code.h b/third_party/aom/av1/common/generic_code.h
new file mode 100644
index 0000000000..c9d87799d1
--- /dev/null
+++ b/third_party/aom/av1/common/generic_code.h
@@ -0,0 +1,81 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_generic_code_H)
+# define _generic_code_H
+
+# include "aom_dsp/bitreader.h"
+# include "aom_dsp/bitwriter.h"
+
+# define GENERIC_TABLES 12
+
+#define generic_decode(r, model, ex_q16, integration, ACCT_STR_NAME) \
+  generic_decode_(r, model, ex_q16, integration ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_decode_cdf_adapt_q15(r, cdf, n, count, rate, ACCT_STR_NAME) \
+  aom_decode_cdf_adapt_q15_(r, cdf, n, count, rate ACCT_STR_ARG(ACCT_STR_NAME))
+#define aom_decode_cdf_adapt(r, cdf, n, increment, ACCT_STR_NAME) \
+  aom_decode_cdf_adapt_(r, cdf, n, increment ACCT_STR_ARG(ACCT_STR_NAME))
+
+typedef struct {
+  /** cdf for multiple expectations of x */
+  uint16_t cdf[GENERIC_TABLES][CDF_SIZE(16)];
+} generic_encoder;
+
+#define OD_IIR_DIADIC(y, x, shift) ((y) += ((x) - (y)) >> (shift))
+
+void generic_model_init(generic_encoder *model);
+
+/* Initialize a CDF for use by aom_write_symbol_pvq()/aom_read_symbol_pvq().
+   This is used for CDFs whose size might not match the declared array size.
+   The only real requirement is that the first value of every CDF be zero.
+   Then aom_cdf_init_q15_1D() will be called with the real size the first time
+   the CDF is used. */
+#define OD_CDFS_INIT_DYNAMIC(cdf) (memset(cdf, 0, sizeof(cdf)))
+
+// WARNING: DO NOT USE this init function,
+// if the size of cdf is different from what is declared by code.
+#define OD_CDFS_INIT_Q15(cdfs) \
+  { int n_cdfs = sizeof(cdfs)/sizeof(cdfs[0]); \
+    int cdf_size = sizeof(cdfs[0])/sizeof(cdfs[0][0]); \
+    int nsyms = cdf_size - CONFIG_EC_ADAPT; \
+    int i_; \
+    for (i_ = 0; i_ < n_cdfs; i_++) \
+      aom_cdf_init_q15_1D(cdfs[i_], nsyms, cdf_size); \
+  }
+
+void aom_cdf_init(uint16_t *cdf, int ncdfs, int nsyms, int val, int first);
+
+void aom_cdf_init_q15_1D(uint16_t *cdf, int nsyms, int cdf_size);
+
+void aom_cdf_adapt_q15(int val, uint16_t *cdf, int n, int *count, int rate);
+
+void aom_encode_cdf_adapt_q15(aom_writer *w, int val, uint16_t *cdf, int n,
+ int *count, int rate);
+
+void generic_encode(aom_writer *w, generic_encoder *model, int x,
+ int *ex_q16, int integration);
+double generic_encode_cost(generic_encoder *model, int x, int *ex_q16);
+
+double od_encode_cdf_cost(int val, uint16_t *cdf, int n);
+
+int aom_decode_cdf_adapt_q15_(aom_reader *r, uint16_t *cdf, int n,
+ int *count, int rate ACCT_STR_PARAM);
+
+int generic_decode_(aom_reader *r, generic_encoder *model,
+ int *ex_q16, int integration ACCT_STR_PARAM);
+
+int log_ex(int ex_q16);
+
+void generic_model_update(int *ex_q16, int x, int integration);
+
+#endif
diff --git a/third_party/aom/av1/common/idct.c b/third_party/aom/av1/common/idct.c
new file mode 100644
index 0000000000..0ea58bfe68
--- /dev/null
+++ b/third_party/aom/av1/common/idct.c
@@ -0,0 +1,3067 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+#include "aom_dsp/inv_txfm.h"
+#include "aom_ports/mem.h"
+#include "av1/common/av1_inv_txfm2d_cfg.h"
+#include "av1/common/blockd.h"
+#include "av1/common/enums.h"
+#include "av1/common/idct.h"
+
+int av1_get_tx_scale(const TX_SIZE tx_size) {
+  if (txsize_sqr_up_map[tx_size] == TX_32X32) return 1;
+#if CONFIG_TX64X64
+  else if (txsize_sqr_up_map[tx_size] == TX_64X64)
+    return 2;
+#endif  // CONFIG_TX64X64
+  else
+    return 0;
+}
+
+// NOTE: The implementation of all inverses need to be aware of the fact
+// that input and output could be the same buffer.
+
+#if CONFIG_EXT_TX
+static void iidtx4_c(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 4; ++i)
+    output[i] = (tran_low_t)dct_const_round_shift(input[i] * Sqrt2);
+}
+
+static void iidtx8_c(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 8; ++i) output[i] = input[i] * 2;
+}
+
+static void iidtx16_c(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 16; ++i)
+    output[i] = (tran_low_t)dct_const_round_shift(input[i] * 2 * Sqrt2);
+}
+
+static void iidtx32_c(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 32; ++i) output[i] = input[i] * 4;
+}
+
+#if CONFIG_TX64X64
+static void iidtx64_c(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 64; ++i)
+    output[i] = (tran_low_t)dct_const_round_shift(input[i] * 4 * Sqrt2);
+}
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_EXT_TX
+
+// For use in lieu of ADST
+static void ihalfright32_c(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  tran_low_t inputhalf[16];
+  // Multiply input by sqrt(2)
+  for (i = 0; i < 16; ++i) {
+    inputhalf[i] = (tran_low_t)dct_const_round_shift(input[i] * Sqrt2);
+  }
+  for (i = 0; i < 16; ++i) {
+    output[i] = input[16 + i] * 4;
+  }
+  aom_idct16_c(inputhalf, output + 16);
+  // Note overall scaling factor is 4 times orthogonal
+}
+
+#if CONFIG_TX64X64
+static void idct64_col_c(const tran_low_t *input, tran_low_t *output) {
+  int32_t in[64], out[64];
+  int i;
+  for (i = 0; i < 64; ++i) in[i] = (int32_t)input[i];
+  av1_idct64_new(in, out, inv_cos_bit_col_dct_dct_64,
+                 inv_stage_range_col_dct_dct_64);
+  for (i = 0; i < 64; ++i) output[i] = (tran_low_t)out[i];
+}
+
+static void idct64_row_c(const tran_low_t *input, tran_low_t *output) {
+  int32_t in[64], out[64];
+  int i;
+  for (i = 0; i < 64; ++i) in[i] = (int32_t)input[i];
+  av1_idct64_new(in, out, inv_cos_bit_row_dct_dct_64,
+                 inv_stage_range_row_dct_dct_64);
+  for (i = 0; i < 64; ++i) output[i] = (tran_low_t)out[i];
+}
+
+// For use in lieu of ADST
+static void ihalfright64_c(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  tran_low_t inputhalf[32];
+  // Multiply input by sqrt(2)
+  for (i = 0; i < 32; ++i) {
+    inputhalf[i] = (tran_low_t)dct_const_round_shift(input[i] * Sqrt2);
+  }
+  for (i = 0; i < 32; ++i) {
+    output[i] = (tran_low_t)dct_const_round_shift(input[32 + i] * 4 * Sqrt2);
+  }
+  aom_idct32_c(inputhalf, output + 32);
+  // Note overall scaling factor is 4 * sqrt(2)  times orthogonal
+}
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_idct4(const tran_low_t *input, tran_low_t *output,
+                         const int8_t *cos_bit, const int8_t *stage_range,
+                         int bd) {
+  (void)bd;
+  av1_idct4_new(input, output, cos_bit, stage_range);
+}
+
+static void highbd_idct8(const tran_low_t *input, tran_low_t *output,
+                         const int8_t *cos_bit, const int8_t *stage_range,
+                         int bd) {
+  (void)bd;
+  av1_idct8_new(input, output, cos_bit, stage_range);
+}
+
+static void highbd_idct16(const tran_low_t *input, tran_low_t *output,
+                          const int8_t *cos_bit, const int8_t *stage_range,
+                          int bd) {
+  (void)bd;
+  av1_idct16_new(input, output, cos_bit, stage_range);
+}
+
+static void highbd_idct32(const tran_low_t *input, tran_low_t *output,
+                          const int8_t *cos_bit, const int8_t *stage_range,
+                          int bd) {
+  (void)bd;
+  av1_idct32_new(input, output, cos_bit, stage_range);
+}
+
+static void highbd_iadst4(const tran_low_t *input, tran_low_t *output,
+                          const int8_t *cos_bit, const int8_t *stage_range,
+                          int bd) {
+  (void)bd;
+  av1_iadst4_new(input, output, cos_bit, stage_range);
+}
+
+static void highbd_iadst8(const tran_low_t *input, tran_low_t *output,
+                          const int8_t *cos_bit, const int8_t *stage_range,
+                          int bd) {
+  (void)bd;
+  av1_iadst8_new(input, output, cos_bit, stage_range);
+}
+
+static void highbd_iadst16(const tran_low_t *input, tran_low_t *output,
+                           const int8_t *cos_bit, const int8_t *stage_range,
+                           int bd) {
+  (void)bd;
+  av1_iadst16_new(input, output, cos_bit, stage_range);
+}
+
+#if CONFIG_EXT_TX
+static void highbd_iidtx4_c(const tran_low_t *input, tran_low_t *output,
+                            const int8_t *cos_bit, const int8_t *stage_range,
+                            int bd) {
+  int i;
+  (void)cos_bit;
+  (void)stage_range;
+  for (i = 0; i < 4; ++i)
+    output[i] = HIGHBD_WRAPLOW(dct_const_round_shift(input[i] * Sqrt2), bd);
+}
+
+static void highbd_iidtx8_c(const tran_low_t *input, tran_low_t *output,
+                            const int8_t *cos_bit, const int8_t *stage_range,
+                            int bd) {
+  int i;
+  (void)bd;
+  (void)cos_bit;
+  (void)stage_range;
+  for (i = 0; i < 8; ++i) output[i] = input[i] * 2;
+}
+
+static void highbd_iidtx16_c(const tran_low_t *input, tran_low_t *output,
+                             const int8_t *cos_bit, const int8_t *stage_range,
+                             int bd) {
+  int i;
+  (void)cos_bit;
+  (void)stage_range;
+  for (i = 0; i < 16; ++i)
+    output[i] = HIGHBD_WRAPLOW(dct_const_round_shift(input[i] * 2 * Sqrt2), bd);
+}
+
+static void highbd_iidtx32_c(const tran_low_t *input, tran_low_t *output,
+                             const int8_t *cos_bit, const int8_t *stage_range,
+                             int bd) {
+  int i;
+  (void)bd;
+  (void)cos_bit;
+  (void)stage_range;
+  for (i = 0; i < 32; ++i) output[i] = input[i] * 4;
+}
+#endif  // CONFIG_EXT_TX
+
+static void highbd_ihalfright32_c(const tran_low_t *input, tran_low_t *output,
+                                  const int8_t *cos_bit,
+                                  const int8_t *stage_range, int bd) {
+  int i;
+  tran_low_t inputhalf[16];
+  // Multiply input by sqrt(2)
+  for (i = 0; i < 16; ++i) {
+    inputhalf[i] = HIGHBD_WRAPLOW(dct_const_round_shift(input[i] * Sqrt2), bd);
+  }
+  for (i = 0; i < 16; ++i) {
+    output[i] = input[16 + i] * 4;
+  }
+  highbd_idct16(inputhalf, output + 16, cos_bit, stage_range, bd);
+  // Note overall scaling factor is 4 times orthogonal
+}
+
+#if CONFIG_EXT_TX
+#if CONFIG_TX64X64
+static void highbd_iidtx64_c(const tran_low_t *input, tran_low_t *output,
+                             const int8_t *cos_bit, const int8_t *stage_range,
+                             int bd) {
+  (void)cos_bit;
+  (void)stage_range;
+  int i;
+  for (i = 0; i < 64; ++i)
+    output[i] = HIGHBD_WRAPLOW(dct_const_round_shift(input[i] * 4 * Sqrt2), bd);
+}
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_EXT_TX
+
+#if CONFIG_TX64X64
+// For use in lieu of ADST
+static void highbd_ihalfright64_c(const tran_low_t *input, tran_low_t *output,
+                                  const int8_t *cos_bit,
+                                  const int8_t *stage_range, int bd) {
+  int i;
+  tran_low_t inputhalf[32];
+  // Multiply input by sqrt(2)
+  for (i = 0; i < 32; ++i) {
+    inputhalf[i] = HIGHBD_WRAPLOW(dct_const_round_shift(input[i] * Sqrt2), bd);
+  }
+  for (i = 0; i < 32; ++i) {
+    output[i] =
+        HIGHBD_WRAPLOW(dct_const_round_shift(input[32 + i] * 4 * Sqrt2), bd);
+  }
+  highbd_idct32(inputhalf, output + 32, cos_bit, stage_range, bd);
+  // Note overall scaling factor is 4 * sqrt(2)  times orthogonal
+}
+
+static void highbd_idct64_col_c(const tran_low_t *input, tran_low_t *output,
+                                const int8_t *cos_bit,
+                                const int8_t *stage_range, int bd) {
+  int32_t in[64], out[64];
+  int i;
+  (void)cos_bit;
+  (void)stage_range;
+  (void)bd;
+  for (i = 0; i < 64; ++i) in[i] = (int32_t)input[i];
+  av1_idct64_new(in, out, inv_cos_bit_col_dct_dct_64,
+                 inv_stage_range_col_dct_dct_64);
+  for (i = 0; i < 64; ++i) output[i] = (tran_low_t)out[i];
+}
+
+static void highbd_idct64_row_c(const tran_low_t *input, tran_low_t *output,
+                                const int8_t *cos_bit,
+                                const int8_t *stage_range, int bd) {
+  int32_t in[64], out[64];
+  int i;
+  (void)cos_bit;
+  (void)stage_range;
+  (void)bd;
+  for (i = 0; i < 64; ++i) in[i] = (int32_t)input[i];
+  av1_idct64_new(in, out, inv_cos_bit_row_dct_dct_64,
+                 inv_stage_range_row_dct_dct_64);
+  for (i = 0; i < 64; ++i) output[i] = (tran_low_t)out[i];
+}
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_HIGHBITDEPTH
+
+// Inverse identity transform and add.
+#if CONFIG_EXT_TX
+static void inv_idtx_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                           int bs, int tx_type) {
+  int r, c;
+  const int shift = bs < 32 ? 3 : (bs < 64 ? 2 : 1);
+  if (tx_type == IDTX) {
+    for (r = 0; r < bs; ++r) {
+      for (c = 0; c < bs; ++c)
+        dest[c] = clip_pixel_add(dest[c], input[c] >> shift);
+      dest += stride;
+      input += bs;
+    }
+  }
+}
+#endif  // CONFIG_EXT_TX
+
+#define FLIPUD_PTR(dest, stride, size)       \
+  do {                                       \
+    (dest) = (dest) + ((size)-1) * (stride); \
+    (stride) = -(stride);                    \
+  } while (0)
+
+#if CONFIG_EXT_TX
+static void maybe_flip_strides(uint8_t **dst, int *dstride, tran_low_t **src,
+                               int *sstride, int tx_type, int sizey,
+                               int sizex) {
+  // Note that the transpose of src will be added to dst. In order to LR
+  // flip the addends (in dst coordinates), we UD flip the src. To UD flip
+  // the addends, we UD flip the dst.
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case IDTX:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST: break;
+    case FLIPADST_DCT:
+    case FLIPADST_ADST:
+    case V_FLIPADST:
+      // flip UD
+      FLIPUD_PTR(*dst, *dstride, sizey);
+      break;
+    case DCT_FLIPADST:
+    case ADST_FLIPADST:
+    case H_FLIPADST:
+      // flip LR
+      FLIPUD_PTR(*src, *sstride, sizex);
+      break;
+    case FLIPADST_FLIPADST:
+      // flip UD
+      FLIPUD_PTR(*dst, *dstride, sizey);
+      // flip LR
+      FLIPUD_PTR(*src, *sstride, sizex);
+      break;
+    default: assert(0); break;
+  }
+}
+#endif  // CONFIG_EXT_TX
+
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_EXT_TX
+static void highbd_inv_idtx_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int bs, int tx_type, int bd) {
+  int r, c;
+  const int shift = bs < 32 ? 3 : 2;
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  if (tx_type == IDTX) {
+    for (r = 0; r < bs; ++r) {
+      for (c = 0; c < bs; ++c)
+        dest[c] = highbd_clip_pixel_add(dest[c], input[c] >> shift, bd);
+      dest += stride;
+      input += bs;
+    }
+  }
+}
+
+static void maybe_flip_strides16(uint16_t **dst, int *dstride, tran_low_t **src,
+                                 int *sstride, int tx_type, int sizey,
+                                 int sizex) {
+  // Note that the transpose of src will be added to dst. In order to LR
+  // flip the addends (in dst coordinates), we UD flip the src. To UD flip
+  // the addends, we UD flip the dst.
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case IDTX:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST: break;
+    case FLIPADST_DCT:
+    case FLIPADST_ADST:
+    case V_FLIPADST:
+      // flip UD
+      FLIPUD_PTR(*dst, *dstride, sizey);
+      break;
+    case DCT_FLIPADST:
+    case ADST_FLIPADST:
+    case H_FLIPADST:
+      // flip LR
+      FLIPUD_PTR(*src, *sstride, sizex);
+      break;
+    case FLIPADST_FLIPADST:
+      // flip UD
+      FLIPUD_PTR(*dst, *dstride, sizey);
+      // flip LR
+      FLIPUD_PTR(*src, *sstride, sizex);
+      break;
+    default: assert(0); break;
+  }
+}
+#endif  // CONFIG_EXT_TX
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                         int tx_type) {
+  static const transform_2d IHT_4[] = {
+    { aom_idct4_c, aom_idct4_c },    // DCT_DCT  = 0
+    { aom_iadst4_c, aom_idct4_c },   // ADST_DCT = 1
+    { aom_idct4_c, aom_iadst4_c },   // DCT_ADST = 2
+    { aom_iadst4_c, aom_iadst4_c },  // ADST_ADST = 3
+#if CONFIG_EXT_TX
+    { aom_iadst4_c, aom_idct4_c },   // FLIPADST_DCT
+    { aom_idct4_c, aom_iadst4_c },   // DCT_FLIPADST
+    { aom_iadst4_c, aom_iadst4_c },  // FLIPADST_FLIPADST
+    { aom_iadst4_c, aom_iadst4_c },  // ADST_FLIPADST
+    { aom_iadst4_c, aom_iadst4_c },  // FLIPADST_ADST
+    { iidtx4_c, iidtx4_c },          // IDTX
+    { aom_idct4_c, iidtx4_c },       // V_DCT
+    { iidtx4_c, aom_idct4_c },       // H_DCT
+    { aom_iadst4_c, iidtx4_c },      // V_ADST
+    { iidtx4_c, aom_iadst4_c },      // H_ADST
+    { aom_iadst4_c, iidtx4_c },      // V_FLIPADST
+    { iidtx4_c, aom_iadst4_c },      // H_FLIPADST
+#endif                               // CONFIG_EXT_TX
+  };
+
+  int i, j;
+  tran_low_t tmp;
+  tran_low_t out[4][4];
+  tran_low_t *outp = &out[0][0];
+  int outstride = 4;
+
+  // inverse transform row vectors
+  for (i = 0; i < 4; ++i) {
+    IHT_4[tx_type].rows(input, out[i]);
+    input += 4;
+  }
+
+  // transpose
+  for (i = 1; i < 4; i++) {
+    for (j = 0; j < i; j++) {
+      tmp = out[i][j];
+      out[i][j] = out[j][i];
+      out[j][i] = tmp;
+    }
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 4; ++i) {
+    IHT_4[tx_type].cols(out[i], out[i]);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, 4, 4);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 4));
+    }
+  }
+}
+
+void av1_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                         int tx_type) {
+  static const transform_2d IHT_4x8[] = {
+    { aom_idct8_c, aom_idct4_c },    // DCT_DCT
+    { aom_iadst8_c, aom_idct4_c },   // ADST_DCT
+    { aom_idct8_c, aom_iadst4_c },   // DCT_ADST
+    { aom_iadst8_c, aom_iadst4_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { aom_iadst8_c, aom_idct4_c },   // FLIPADST_DCT
+    { aom_idct8_c, aom_iadst4_c },   // DCT_FLIPADST
+    { aom_iadst8_c, aom_iadst4_c },  // FLIPADST_FLIPADST
+    { aom_iadst8_c, aom_iadst4_c },  // ADST_FLIPADST
+    { aom_iadst8_c, aom_iadst4_c },  // FLIPADST_ADST
+    { iidtx8_c, iidtx4_c },          // IDTX
+    { aom_idct8_c, iidtx4_c },       // V_DCT
+    { iidtx8_c, aom_idct4_c },       // H_DCT
+    { aom_iadst8_c, iidtx4_c },      // V_ADST
+    { iidtx8_c, aom_iadst4_c },      // H_ADST
+    { aom_iadst8_c, iidtx4_c },      // V_FLIPADST
+    { iidtx8_c, aom_iadst4_c },      // H_FLIPADST
+#endif
+  };
+
+  const int n = 4;
+  const int n2 = 8;
+  int i, j;
+  tran_low_t out[4][8], outtmp[4];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n2;
+
+  // inverse transform row vectors and transpose
+  for (i = 0; i < n2; ++i) {
+    IHT_4x8[tx_type].rows(input, outtmp);
+    for (j = 0; j < n; ++j)
+      out[j][i] = (tran_low_t)dct_const_round_shift(outtmp[j] * Sqrt2);
+    input += n;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n; ++i) {
+    IHT_4x8[tx_type].cols(out[i], out[i]);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n2, n);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5));
+    }
+  }
+}
+
+void av1_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                         int tx_type) {
+  static const transform_2d IHT_8x4[] = {
+    { aom_idct4_c, aom_idct8_c },    // DCT_DCT
+    { aom_iadst4_c, aom_idct8_c },   // ADST_DCT
+    { aom_idct4_c, aom_iadst8_c },   // DCT_ADST
+    { aom_iadst4_c, aom_iadst8_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { aom_iadst4_c, aom_idct8_c },   // FLIPADST_DCT
+    { aom_idct4_c, aom_iadst8_c },   // DCT_FLIPADST
+    { aom_iadst4_c, aom_iadst8_c },  // FLIPADST_FLIPADST
+    { aom_iadst4_c, aom_iadst8_c },  // ADST_FLIPADST
+    { aom_iadst4_c, aom_iadst8_c },  // FLIPADST_ADST
+    { iidtx4_c, iidtx8_c },          // IDTX
+    { aom_idct4_c, iidtx8_c },       // V_DCT
+    { iidtx4_c, aom_idct8_c },       // H_DCT
+    { aom_iadst4_c, iidtx8_c },      // V_ADST
+    { iidtx4_c, aom_iadst8_c },      // H_ADST
+    { aom_iadst4_c, iidtx8_c },      // V_FLIPADST
+    { iidtx4_c, aom_iadst8_c },      // H_FLIPADST
+#endif
+  };
+  const int n = 4;
+  const int n2 = 8;
+
+  int i, j;
+  tran_low_t out[8][4], outtmp[8];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n;
+
+  // inverse transform row vectors and transpose
+  for (i = 0; i < n; ++i) {
+    IHT_8x4[tx_type].rows(input, outtmp);
+    for (j = 0; j < n2; ++j)
+      out[j][i] = (tran_low_t)dct_const_round_shift(outtmp[j] * Sqrt2);
+    input += n2;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n2; ++i) {
+    IHT_8x4[tx_type].cols(out[i], out[i]);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n, n2);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5));
+    }
+  }
+}
+
+void av1_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                          int tx_type) {
+  static const transform_2d IHT_4x16[] = {
+    { aom_idct16_c, aom_idct4_c },    // DCT_DCT
+    { aom_iadst16_c, aom_idct4_c },   // ADST_DCT
+    { aom_idct16_c, aom_iadst4_c },   // DCT_ADST
+    { aom_iadst16_c, aom_iadst4_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { aom_iadst16_c, aom_idct4_c },   // FLIPADST_DCT
+    { aom_idct16_c, aom_iadst4_c },   // DCT_FLIPADST
+    { aom_iadst16_c, aom_iadst4_c },  // FLIPADST_FLIPADST
+    { aom_iadst16_c, aom_iadst4_c },  // ADST_FLIPADST
+    { aom_iadst16_c, aom_iadst4_c },  // FLIPADST_ADST
+    { iidtx16_c, iidtx4_c },          // IDTX
+    { aom_idct16_c, iidtx4_c },       // V_DCT
+    { iidtx16_c, aom_idct4_c },       // H_DCT
+    { aom_iadst16_c, iidtx4_c },      // V_ADST
+    { iidtx16_c, aom_iadst4_c },      // H_ADST
+    { aom_iadst16_c, iidtx4_c },      // V_FLIPADST
+    { iidtx16_c, aom_iadst4_c },      // H_FLIPADST
+#endif
+  };
+
+  const int n = 4;
+  const int n4 = 16;
+  int i, j;
+  tran_low_t out[4][16], outtmp[4];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n4;
+
+  // inverse transform row vectors and transpose
+  for (i = 0; i < n4; ++i) {
+    IHT_4x16[tx_type].rows(input, outtmp);
+    for (j = 0; j < n; ++j) out[j][i] = outtmp[j];
+    input += n;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n; ++i) IHT_4x16[tx_type].cols(out[i], out[i]);
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n4, n);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < n4; ++i) {
+    for (j = 0; j < n; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5));
+    }
+  }
+}
+
+void av1_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                          int tx_type) {
+  static const transform_2d IHT_16x4[] = {
+    { aom_idct4_c, aom_idct16_c },    // DCT_DCT
+    { aom_iadst4_c, aom_idct16_c },   // ADST_DCT
+    { aom_idct4_c, aom_iadst16_c },   // DCT_ADST
+    { aom_iadst4_c, aom_iadst16_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { aom_iadst4_c, aom_idct16_c },   // FLIPADST_DCT
+    { aom_idct4_c, aom_iadst16_c },   // DCT_FLIPADST
+    { aom_iadst4_c, aom_iadst16_c },  // FLIPADST_FLIPADST
+    { aom_iadst4_c, aom_iadst16_c },  // ADST_FLIPADST
+    { aom_iadst4_c, aom_iadst16_c },  // FLIPADST_ADST
+    { iidtx4_c, iidtx16_c },          // IDTX
+    { aom_idct4_c, iidtx16_c },       // V_DCT
+    { iidtx4_c, aom_idct16_c },       // H_DCT
+    { aom_iadst4_c, iidtx16_c },      // V_ADST
+    { iidtx4_c, aom_iadst16_c },      // H_ADST
+    { aom_iadst4_c, iidtx16_c },      // V_FLIPADST
+    { iidtx4_c, aom_iadst16_c },      // H_FLIPADST
+#endif
+  };
+  const int n = 4;
+  const int n4 = 16;
+
+  int i, j;
+  tran_low_t out[16][4], outtmp[16];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n;
+
+  // inverse transform row vectors and transpose
+  for (i = 0; i < n; ++i) {
+    IHT_16x4[tx_type].rows(input, outtmp);
+    for (j = 0; j < n4; ++j) out[j][i] = outtmp[j];
+    input += n4;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n4; ++i) IHT_16x4[tx_type].cols(out[i], out[i]);
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n, n4);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n4; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5));
+    }
+  }
+}
+
+void av1_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                           int tx_type) {
+  static const transform_2d IHT_8x16[] = {
+    { aom_idct16_c, aom_idct8_c },    // DCT_DCT
+    { aom_iadst16_c, aom_idct8_c },   // ADST_DCT
+    { aom_idct16_c, aom_iadst8_c },   // DCT_ADST
+    { aom_iadst16_c, aom_iadst8_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { aom_iadst16_c, aom_idct8_c },   // FLIPADST_DCT
+    { aom_idct16_c, aom_iadst8_c },   // DCT_FLIPADST
+    { aom_iadst16_c, aom_iadst8_c },  // FLIPADST_FLIPADST
+    { aom_iadst16_c, aom_iadst8_c },  // ADST_FLIPADST
+    { aom_iadst16_c, aom_iadst8_c },  // FLIPADST_ADST
+    { iidtx16_c, iidtx8_c },          // IDTX
+    { aom_idct16_c, iidtx8_c },       // V_DCT
+    { iidtx16_c, aom_idct8_c },       // H_DCT
+    { aom_iadst16_c, iidtx8_c },      // V_ADST
+    { iidtx16_c, aom_iadst8_c },      // H_ADST
+    { aom_iadst16_c, iidtx8_c },      // V_FLIPADST
+    { iidtx16_c, aom_iadst8_c },      // H_FLIPADST
+#endif
+  };
+
+  const int n = 8;
+  const int n2 = 16;
+  int i, j;
+  tran_low_t out[8][16], outtmp[8];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n2;
+
+  // inverse transform row vectors and transpose
+  for (i = 0; i < n2; ++i) {
+    IHT_8x16[tx_type].rows(input, outtmp);
+    for (j = 0; j < n; ++j)
+      out[j][i] = (tran_low_t)dct_const_round_shift(outtmp[j] * Sqrt2);
+    input += n;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n; ++i) {
+    IHT_8x16[tx_type].cols(out[i], out[i]);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n2, n);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6));
+    }
+  }
+}
+
+void av1_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                           int tx_type) {
+  static const transform_2d IHT_16x8[] = {
+    { aom_idct8_c, aom_idct16_c },    // DCT_DCT
+    { aom_iadst8_c, aom_idct16_c },   // ADST_DCT
+    { aom_idct8_c, aom_iadst16_c },   // DCT_ADST
+    { aom_iadst8_c, aom_iadst16_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { aom_iadst8_c, aom_idct16_c },   // FLIPADST_DCT
+    { aom_idct8_c, aom_iadst16_c },   // DCT_FLIPADST
+    { aom_iadst8_c, aom_iadst16_c },  // FLIPADST_FLIPADST
+    { aom_iadst8_c, aom_iadst16_c },  // ADST_FLIPADST
+    { aom_iadst8_c, aom_iadst16_c },  // FLIPADST_ADST
+    { iidtx8_c, iidtx16_c },          // IDTX
+    { aom_idct8_c, iidtx16_c },       // V_DCT
+    { iidtx8_c, aom_idct16_c },       // H_DCT
+    { aom_iadst8_c, iidtx16_c },      // V_ADST
+    { iidtx8_c, aom_iadst16_c },      // H_ADST
+    { aom_iadst8_c, iidtx16_c },      // V_FLIPADST
+    { iidtx8_c, aom_iadst16_c },      // H_FLIPADST
+#endif
+  };
+  const int n = 8;
+  const int n2 = 16;
+
+  int i, j;
+  tran_low_t out[16][8], outtmp[16];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n;
+
+  // inverse transform row vectors and transpose
+  for (i = 0; i < n; ++i) {
+    IHT_16x8[tx_type].rows(input, outtmp);
+    for (j = 0; j < n2; ++j)
+      out[j][i] = (tran_low_t)dct_const_round_shift(outtmp[j] * Sqrt2);
+    input += n2;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n2; ++i) {
+    IHT_16x8[tx_type].cols(out[i], out[i]);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n, n2);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6));
+    }
+  }
+}
+
+void av1_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                           int tx_type) {
+  static const transform_2d IHT_8x32[] = {
+    { aom_idct32_c, aom_idct8_c },     // DCT_DCT
+    { ihalfright32_c, aom_idct8_c },   // ADST_DCT
+    { aom_idct32_c, aom_iadst8_c },    // DCT_ADST
+    { ihalfright32_c, aom_iadst8_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { ihalfright32_c, aom_idct8_c },   // FLIPADST_DCT
+    { aom_idct32_c, aom_iadst8_c },    // DCT_FLIPADST
+    { ihalfright32_c, aom_iadst8_c },  // FLIPADST_FLIPADST
+    { ihalfright32_c, aom_iadst8_c },  // ADST_FLIPADST
+    { ihalfright32_c, aom_iadst8_c },  // FLIPADST_ADST
+    { iidtx32_c, iidtx8_c },           // IDTX
+    { aom_idct32_c, iidtx8_c },        // V_DCT
+    { iidtx32_c, aom_idct8_c },        // H_DCT
+    { ihalfright32_c, iidtx8_c },      // V_ADST
+    { iidtx32_c, aom_iadst8_c },       // H_ADST
+    { ihalfright32_c, iidtx8_c },      // V_FLIPADST
+    { iidtx32_c, aom_iadst8_c },       // H_FLIPADST
+#endif
+  };
+
+  const int n = 8;
+  const int n4 = 32;
+  int i, j;
+  tran_low_t out[8][32], outtmp[8];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n4;
+
+  // inverse transform row vectors and transpose
+  for (i = 0; i < n4; ++i) {
+    IHT_8x32[tx_type].rows(input, outtmp);
+    for (j = 0; j < n; ++j) out[j][i] = outtmp[j];
+    input += n;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n; ++i) IHT_8x32[tx_type].cols(out[i], out[i]);
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n4, n);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < n4; ++i) {
+    for (j = 0; j < n; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6));
+    }
+  }
+}
+
+void av1_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                           int tx_type) {
+  static const transform_2d IHT_32x8[] = {
+    { aom_idct8_c, aom_idct32_c },     // DCT_DCT
+    { aom_iadst8_c, aom_idct32_c },    // ADST_DCT
+    { aom_idct8_c, ihalfright32_c },   // DCT_ADST
+    { aom_iadst8_c, ihalfright32_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { aom_iadst8_c, aom_idct32_c },    // FLIPADST_DCT
+    { aom_idct8_c, ihalfright32_c },   // DCT_FLIPADST
+    { aom_iadst8_c, ihalfright32_c },  // FLIPADST_FLIPADST
+    { aom_iadst8_c, ihalfright32_c },  // ADST_FLIPADST
+    { aom_iadst8_c, ihalfright32_c },  // FLIPADST_ADST
+    { iidtx8_c, iidtx32_c },           // IDTX
+    { aom_idct8_c, iidtx32_c },        // V_DCT
+    { iidtx8_c, aom_idct32_c },        // H_DCT
+    { aom_iadst8_c, iidtx32_c },       // V_ADST
+    { iidtx8_c, ihalfright32_c },      // H_ADST
+    { aom_iadst8_c, iidtx32_c },       // V_FLIPADST
+    { iidtx8_c, ihalfright32_c },      // H_FLIPADST
+#endif
+  };
+  const int n = 8;
+  const int n4 = 32;
+
+  int i, j;
+  tran_low_t out[32][8], outtmp[32];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n;
+
+  // inverse transform row vectors and transpose
+  for (i = 0; i < n; ++i) {
+    IHT_32x8[tx_type].rows(input, outtmp);
+    for (j = 0; j < n4; ++j) out[j][i] = outtmp[j];
+    input += n4;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n4; ++i) IHT_32x8[tx_type].cols(out[i], out[i]);
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n, n4);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n4; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6));
+    }
+  }
+}
+
+void av1_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  static const transform_2d IHT_16x32[] = {
+    { aom_idct32_c, aom_idct16_c },     // DCT_DCT
+    { ihalfright32_c, aom_idct16_c },   // ADST_DCT
+    { aom_idct32_c, aom_iadst16_c },    // DCT_ADST
+    { ihalfright32_c, aom_iadst16_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { ihalfright32_c, aom_idct16_c },   // FLIPADST_DCT
+    { aom_idct32_c, aom_iadst16_c },    // DCT_FLIPADST
+    { ihalfright32_c, aom_iadst16_c },  // FLIPADST_FLIPADST
+    { ihalfright32_c, aom_iadst16_c },  // ADST_FLIPADST
+    { ihalfright32_c, aom_iadst16_c },  // FLIPADST_ADST
+    { iidtx32_c, iidtx16_c },           // IDTX
+    { aom_idct32_c, iidtx16_c },        // V_DCT
+    { iidtx32_c, aom_idct16_c },        // H_DCT
+    { ihalfright32_c, iidtx16_c },      // V_ADST
+    { iidtx32_c, aom_iadst16_c },       // H_ADST
+    { ihalfright32_c, iidtx16_c },      // V_FLIPADST
+    { iidtx32_c, aom_iadst16_c },       // H_FLIPADST
+#endif
+  };
+
+  const int n = 16;
+  const int n2 = 32;
+  int i, j;
+  tran_low_t out[16][32], outtmp[16];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n2;
+
+  // inverse transform row vectors and transpose
+  for (i = 0; i < n2; ++i) {
+    IHT_16x32[tx_type].rows(input, outtmp);
+    for (j = 0; j < n; ++j)
+      out[j][i] = (tran_low_t)dct_const_round_shift(outtmp[j] * Sqrt2);
+    input += n;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n; ++i) {
+    IHT_16x32[tx_type].cols(out[i], out[i]);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n2, n);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6));
+    }
+  }
+}
+
+void av1_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  static const transform_2d IHT_32x16[] = {
+    { aom_idct16_c, aom_idct32_c },     // DCT_DCT
+    { aom_iadst16_c, aom_idct32_c },    // ADST_DCT
+    { aom_idct16_c, ihalfright32_c },   // DCT_ADST
+    { aom_iadst16_c, ihalfright32_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { aom_iadst16_c, aom_idct32_c },    // FLIPADST_DCT
+    { aom_idct16_c, ihalfright32_c },   // DCT_FLIPADST
+    { aom_iadst16_c, ihalfright32_c },  // FLIPADST_FLIPADST
+    { aom_iadst16_c, ihalfright32_c },  // ADST_FLIPADST
+    { aom_iadst16_c, ihalfright32_c },  // FLIPADST_ADST
+    { iidtx16_c, iidtx32_c },           // IDTX
+    { aom_idct16_c, iidtx32_c },        // V_DCT
+    { iidtx16_c, aom_idct32_c },        // H_DCT
+    { aom_iadst16_c, iidtx32_c },       // V_ADST
+    { iidtx16_c, ihalfright32_c },      // H_ADST
+    { aom_iadst16_c, iidtx32_c },       // V_FLIPADST
+    { iidtx16_c, ihalfright32_c },      // H_FLIPADST
+#endif
+  };
+  const int n = 16;
+  const int n2 = 32;
+
+  int i, j;
+  tran_low_t out[32][16], outtmp[32];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n;
+
+  // inverse transform row vectors and transpose
+  for (i = 0; i < n; ++i) {
+    IHT_32x16[tx_type].rows(input, outtmp);
+    for (j = 0; j < n2; ++j)
+      out[j][i] = (tran_low_t)dct_const_round_shift(outtmp[j] * Sqrt2);
+    input += n2;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n2; ++i) {
+    IHT_32x16[tx_type].cols(out[i], out[i]);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, n, n2);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6));
+    }
+  }
+}
+
+void av1_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                         int tx_type) {
+  static const transform_2d IHT_8[] = {
+    { aom_idct8_c, aom_idct8_c },    // DCT_DCT  = 0
+    { aom_iadst8_c, aom_idct8_c },   // ADST_DCT = 1
+    { aom_idct8_c, aom_iadst8_c },   // DCT_ADST = 2
+    { aom_iadst8_c, aom_iadst8_c },  // ADST_ADST = 3
+#if CONFIG_EXT_TX
+    { aom_iadst8_c, aom_idct8_c },   // FLIPADST_DCT
+    { aom_idct8_c, aom_iadst8_c },   // DCT_FLIPADST
+    { aom_iadst8_c, aom_iadst8_c },  // FLIPADST_FLIPADST
+    { aom_iadst8_c, aom_iadst8_c },  // ADST_FLIPADST
+    { aom_iadst8_c, aom_iadst8_c },  // FLIPADST_ADST
+    { iidtx8_c, iidtx8_c },          // IDTX
+    { aom_idct8_c, iidtx8_c },       // V_DCT
+    { iidtx8_c, aom_idct8_c },       // H_DCT
+    { aom_iadst8_c, iidtx8_c },      // V_ADST
+    { iidtx8_c, aom_iadst8_c },      // H_ADST
+    { aom_iadst8_c, iidtx8_c },      // V_FLIPADST
+    { iidtx8_c, aom_iadst8_c },      // H_FLIPADST
+#endif                               // CONFIG_EXT_TX
+  };
+
+  int i, j;
+  tran_low_t tmp;
+  tran_low_t out[8][8];
+  tran_low_t *outp = &out[0][0];
+  int outstride = 8;
+
+  // inverse transform row vectors
+  for (i = 0; i < 8; ++i) {
+    IHT_8[tx_type].rows(input, out[i]);
+    input += 8;
+  }
+
+  // transpose
+  for (i = 1; i < 8; i++) {
+    for (j = 0; j < i; j++) {
+      tmp = out[i][j];
+      out[i][j] = out[j][i];
+      out[j][i] = tmp;
+    }
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 8; ++i) {
+    IHT_8[tx_type].cols(out[i], out[i]);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, 8, 8);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5));
+    }
+  }
+}
+
+void av1_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  static const transform_2d IHT_16[] = {
+    { aom_idct16_c, aom_idct16_c },    // DCT_DCT  = 0
+    { aom_iadst16_c, aom_idct16_c },   // ADST_DCT = 1
+    { aom_idct16_c, aom_iadst16_c },   // DCT_ADST = 2
+    { aom_iadst16_c, aom_iadst16_c },  // ADST_ADST = 3
+#if CONFIG_EXT_TX
+    { aom_iadst16_c, aom_idct16_c },   // FLIPADST_DCT
+    { aom_idct16_c, aom_iadst16_c },   // DCT_FLIPADST
+    { aom_iadst16_c, aom_iadst16_c },  // FLIPADST_FLIPADST
+    { aom_iadst16_c, aom_iadst16_c },  // ADST_FLIPADST
+    { aom_iadst16_c, aom_iadst16_c },  // FLIPADST_ADST
+    { iidtx16_c, iidtx16_c },          // IDTX
+    { aom_idct16_c, iidtx16_c },       // V_DCT
+    { iidtx16_c, aom_idct16_c },       // H_DCT
+    { aom_iadst16_c, iidtx16_c },      // V_ADST
+    { iidtx16_c, aom_iadst16_c },      // H_ADST
+    { aom_iadst16_c, iidtx16_c },      // V_FLIPADST
+    { iidtx16_c, aom_iadst16_c },      // H_FLIPADST
+#endif                                 // CONFIG_EXT_TX
+  };
+
+  int i, j;
+  tran_low_t tmp;
+  tran_low_t out[16][16];
+  tran_low_t *outp = &out[0][0];
+  int outstride = 16;
+
+  // inverse transform row vectors
+  for (i = 0; i < 16; ++i) {
+    IHT_16[tx_type].rows(input, out[i]);
+    input += 16;
+  }
+
+  // transpose
+  for (i = 1; i < 16; i++) {
+    for (j = 0; j < i; j++) {
+      tmp = out[i][j];
+      out[i][j] = out[j][i];
+      out[j][i] = tmp;
+    }
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 16; ++i) {
+    IHT_16[tx_type].cols(out[i], out[i]);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, 16, 16);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6));
+    }
+  }
+}
+
+#if CONFIG_EXT_TX
+void av1_iht32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                             int tx_type) {
+  static const transform_2d IHT_32[] = {
+    { aom_idct32_c, aom_idct32_c },      // DCT_DCT
+    { ihalfright32_c, aom_idct32_c },    // ADST_DCT
+    { aom_idct32_c, ihalfright32_c },    // DCT_ADST
+    { ihalfright32_c, ihalfright32_c },  // ADST_ADST
+    { ihalfright32_c, aom_idct32_c },    // FLIPADST_DCT
+    { aom_idct32_c, ihalfright32_c },    // DCT_FLIPADST
+    { ihalfright32_c, ihalfright32_c },  // FLIPADST_FLIPADST
+    { ihalfright32_c, ihalfright32_c },  // ADST_FLIPADST
+    { ihalfright32_c, ihalfright32_c },  // FLIPADST_ADST
+    { iidtx32_c, iidtx32_c },            // IDTX
+    { aom_idct32_c, iidtx32_c },         // V_DCT
+    { iidtx32_c, aom_idct32_c },         // H_DCT
+    { ihalfright32_c, iidtx32_c },       // V_ADST
+    { iidtx32_c, ihalfright32_c },       // H_ADST
+    { ihalfright32_c, iidtx32_c },       // V_FLIPADST
+    { iidtx32_c, ihalfright32_c },       // H_FLIPADST
+  };
+
+  int i, j;
+  tran_low_t tmp;
+  tran_low_t out[32][32];
+  tran_low_t *outp = &out[0][0];
+  int outstride = 32;
+
+  // inverse transform row vectors
+  for (i = 0; i < 32; ++i) {
+    IHT_32[tx_type].rows(input, out[i]);
+    input += 32;
+  }
+
+  // transpose
+  for (i = 1; i < 32; i++) {
+    for (j = 0; j < i; j++) {
+      tmp = out[i][j];
+      out[i][j] = out[j][i];
+      out[j][i] = tmp;
+    }
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 32; ++i) {
+    IHT_32[tx_type].cols(out[i], out[i]);
+  }
+
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, 32, 32);
+
+  // Sum with the destination
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6));
+    }
+  }
+}
+#endif  // CONFIG_EXT_TX
+
+#if CONFIG_TX64X64
+void av1_iht64x64_4096_add_c(const tran_low_t *input, uint8_t *dest, int stride,
+                             int tx_type) {
+  static const transform_2d IHT_64[] = {
+    { idct64_col_c, idct64_row_c },      // DCT_DCT
+    { ihalfright64_c, idct64_row_c },    // ADST_DCT
+    { idct64_col_c, ihalfright64_c },    // DCT_ADST
+    { ihalfright64_c, ihalfright64_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { ihalfright64_c, idct64_row_c },    // FLIPADST_DCT
+    { idct64_col_c, ihalfright64_c },    // DCT_FLIPADST
+    { ihalfright64_c, ihalfright64_c },  // FLIPADST_FLIPADST
+    { ihalfright64_c, ihalfright64_c },  // ADST_FLIPADST
+    { ihalfright64_c, ihalfright64_c },  // FLIPADST_ADST
+    { iidtx64_c, iidtx64_c },            // IDTX
+    { idct64_col_c, iidtx64_c },         // V_DCT
+    { iidtx64_c, idct64_row_c },         // H_DCT
+    { ihalfright64_c, iidtx64_c },       // V_ADST
+    { iidtx64_c, ihalfright64_c },       // H_ADST
+    { ihalfright64_c, iidtx64_c },       // V_FLIPADST
+    { iidtx64_c, ihalfright64_c },       // H_FLIPADST
+#endif                                   // CONFIG_EXT_TX
+  };
+
+  int i, j;
+  tran_low_t tmp;
+  tran_low_t out[64][64];
+  tran_low_t *outp = &out[0][0];
+  int outstride = 64;
+
+  // inverse transform row vectors
+  for (i = 0; i < 64; ++i) {
+    IHT_64[tx_type].rows(input, out[i]);
+    for (j = 0; j < 64; ++j) out[i][j] = ROUND_POWER_OF_TWO(out[i][j], 1);
+    input += 64;
+  }
+
+  // transpose
+  for (i = 1; i < 64; i++) {
+    for (j = 0; j < i; j++) {
+      tmp = out[i][j];
+      out[i][j] = out[j][i];
+      out[j][i] = tmp;
+    }
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 64; ++i) {
+    IHT_64[tx_type].cols(out[i], out[i]);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides(&dest, &stride, &outp, &outstride, tx_type, 64, 64);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < 64; ++i) {
+    for (j = 0; j < 64; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] = clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5));
+    }
+  }
+}
+#endif  // CONFIG_TX64X64
+
+// idct
+void av1_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob) {
+  if (eob > 1)
+    aom_idct4x4_16_add(input, dest, stride);
+  else
+    aom_idct4x4_1_add(input, dest, stride);
+}
+
+void av1_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob) {
+  if (eob > 1)
+    aom_iwht4x4_16_add(input, dest, stride);
+  else
+    aom_iwht4x4_1_add(input, dest, stride);
+}
+
+static void idct8x8_add(const tran_low_t *input, uint8_t *dest, int stride,
+                        int eob) {
+  // If dc is 1, then input[0] is the reconstructed value, do not need
+  // dequantization. Also, when dc is 1, dc is counted in eobs, namely eobs >=1.
+
+  // The calculation can be simplified if there are not many non-zero dct
+  // coefficients. Use eobs to decide what to do.
+  // TODO(yunqingwang): "eobs = 1" case is also handled in av1_short_idct8x8_c.
+  // Combine that with code here.
+  if (eob == 1)
+    // DC only DCT coefficient
+    aom_idct8x8_1_add(input, dest, stride);
+#if !CONFIG_ADAPT_SCAN
+  else if (eob <= 12)
+    aom_idct8x8_12_add(input, dest, stride);
+#endif
+  else
+    aom_idct8x8_64_add(input, dest, stride);
+}
+
+static void idct16x16_add(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob) {
+  /* The calculation can be simplified if there are not many non-zero dct
+   * coefficients. Use eobs to separate different cases. */
+  if (eob == 1) /* DC only DCT coefficient. */
+    aom_idct16x16_1_add(input, dest, stride);
+#if !CONFIG_ADAPT_SCAN
+  else if (eob <= 10)
+    aom_idct16x16_10_add(input, dest, stride);
+#endif
+  else
+    aom_idct16x16_256_add(input, dest, stride);
+}
+
+static void idct32x32_add(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob) {
+  if (eob == 1) aom_idct32x32_1_add(input, dest, stride);
+#if !CONFIG_ADAPT_SCAN
+  else if (eob <= 34)
+    // non-zero coeff only in upper-left 8x8
+    aom_idct32x32_34_add(input, dest, stride);
+#endif
+  else
+    aom_idct32x32_1024_add(input, dest, stride);
+}
+
+#if CONFIG_TX64X64
+static void idct64x64_add(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob) {
+  (void)eob;
+  av1_iht64x64_4096_add(input, dest, stride, DCT_DCT);
+}
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_CB4X4
+static void inv_txfm_add_2x2(const tran_low_t *input, uint8_t *dest, int stride,
+                             int eob, TX_TYPE tx_type, int lossless) {
+  tran_high_t a1 = input[0] >> UNIT_QUANT_SHIFT;
+  tran_high_t b1 = input[1] >> UNIT_QUANT_SHIFT;
+  tran_high_t c1 = input[2] >> UNIT_QUANT_SHIFT;
+  tran_high_t d1 = input[3] >> UNIT_QUANT_SHIFT;
+
+  tran_high_t a2 = a1 + c1;
+  tran_high_t b2 = b1 + d1;
+  tran_high_t c2 = a1 - c1;
+  tran_high_t d2 = b1 - d1;
+
+  (void)tx_type;
+  (void)lossless;
+  (void)eob;
+
+  a1 = (a2 + b2) >> 2;
+  b1 = (a2 - b2) >> 2;
+  c1 = (c2 + d2) >> 2;
+  d1 = (c2 - d2) >> 2;
+
+  dest[0] = clip_pixel_add(dest[0], WRAPLOW(a1));
+  dest[1] = clip_pixel_add(dest[1], WRAPLOW(b1));
+  dest[stride] = clip_pixel_add(dest[stride], WRAPLOW(c1));
+  dest[stride + 1] = clip_pixel_add(dest[stride + 1], WRAPLOW(d1));
+}
+#endif
+
+void av1_inv_txfm_add_4x4(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, TX_TYPE tx_type, int lossless) {
+  if (lossless) {
+    assert(tx_type == DCT_DCT);
+    av1_iwht4x4_add(input, dest, stride, eob);
+    return;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT: av1_idct4x4_add(input, dest, stride, eob); break;
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST: av1_iht4x4_16_add(input, dest, stride, tx_type); break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST: av1_iht4x4_16_add(input, dest, stride, tx_type); break;
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      // Use C version since DST only exists in C code
+      av1_iht4x4_16_add_c(input, dest, stride, tx_type);
+      break;
+    case IDTX: inv_idtx_add_c(input, dest, stride, 4, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+void av1_inv_txfm_add_4x8(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, TX_TYPE tx_type) {
+  (void)eob;
+  av1_iht4x8_32_add(input, dest, stride, tx_type);
+}
+
+void av1_inv_txfm_add_8x4(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, TX_TYPE tx_type) {
+  (void)eob;
+  av1_iht8x4_32_add(input, dest, stride, tx_type);
+}
+
+// These will be used by the masked-tx experiment in the future.
+#if CONFIG_MASKED_TX && 0
+static void inv_txfm_add_4x16(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, TX_TYPE tx_type) {
+  (void)eob;
+  av1_iht4x16_64_add(input, dest, stride, tx_type);
+}
+
+static void inv_txfm_add_16x4(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, TX_TYPE tx_type) {
+  (void)eob;
+  av1_iht16x4_64_add(input, dest, stride, tx_type);
+}
+
+static void inv_txfm_add_8x32(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, TX_TYPE tx_type) {
+  (void)eob;
+  av1_iht8x32_256_add(input, dest, stride, tx_type);
+}
+
+static void inv_txfm_add_32x8(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, TX_TYPE tx_type) {
+  (void)eob;
+  av1_iht32x8_256_add(input, dest, stride, tx_type);
+}
+#endif  // CONFIG_MASKED_TX
+
+static void inv_txfm_add_8x16(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, TX_TYPE tx_type) {
+  (void)eob;
+  av1_iht8x16_128_add(input, dest, stride, tx_type);
+}
+
+static void inv_txfm_add_16x8(const tran_low_t *input, uint8_t *dest,
+                              int stride, int eob, TX_TYPE tx_type) {
+  (void)eob;
+  av1_iht16x8_128_add(input, dest, stride, tx_type);
+}
+
+static void inv_txfm_add_16x32(const tran_low_t *input, uint8_t *dest,
+                               int stride, int eob, TX_TYPE tx_type) {
+  (void)eob;
+  av1_iht16x32_512_add(input, dest, stride, tx_type);
+}
+
+static void inv_txfm_add_32x16(const tran_low_t *input, uint8_t *dest,
+                               int stride, int eob, TX_TYPE tx_type) {
+  (void)eob;
+  av1_iht32x16_512_add(input, dest, stride, tx_type);
+}
+
+static void inv_txfm_add_8x8(const tran_low_t *input, uint8_t *dest, int stride,
+                             int eob, TX_TYPE tx_type) {
+  switch (tx_type) {
+    case DCT_DCT: idct8x8_add(input, dest, stride, eob); break;
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST: av1_iht8x8_64_add(input, dest, stride, tx_type); break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST: av1_iht8x8_64_add(input, dest, stride, tx_type); break;
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      // Use C version since DST only exists in C code
+      av1_iht8x8_64_add_c(input, dest, stride, tx_type);
+      break;
+    case IDTX: inv_idtx_add_c(input, dest, stride, 8, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+static void inv_txfm_add_16x16(const tran_low_t *input, uint8_t *dest,
+                               int stride, int eob, TX_TYPE tx_type) {
+  switch (tx_type) {
+    case DCT_DCT: idct16x16_add(input, dest, stride, eob); break;
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST: av1_iht16x16_256_add(input, dest, stride, tx_type); break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST: av1_iht16x16_256_add(input, dest, stride, tx_type); break;
+    case IDTX: inv_idtx_add_c(input, dest, stride, 16, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+static void inv_txfm_add_32x32(const tran_low_t *input, uint8_t *dest,
+                               int stride, int eob, TX_TYPE tx_type) {
+  switch (tx_type) {
+    case DCT_DCT: idct32x32_add(input, dest, stride, eob); break;
+#if CONFIG_EXT_TX
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      av1_iht32x32_1024_add_c(input, dest, stride, tx_type);
+      break;
+    case IDTX: inv_idtx_add_c(input, dest, stride, 32, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+#if CONFIG_TX64X64
+static void inv_txfm_add_64x64(const tran_low_t *input, uint8_t *dest,
+                               int stride, int eob, TX_TYPE tx_type) {
+  switch (tx_type) {
+    case DCT_DCT: idct64x64_add(input, dest, stride, eob); break;
+#if CONFIG_EXT_TX
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      av1_iht64x64_4096_add_c(input, dest, stride, tx_type);
+      break;
+    case IDTX: inv_idtx_add_c(input, dest, stride, 64, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_HIGHBITDEPTH
+
+const TXFM_2D_CFG *inv_txfm_cfg_ls[TX_TYPES][TX_SIZES];
+
+typedef struct {
+  const int8_t *cos_bit;
+  const int8_t *stage_range;
+} tx_1d_cfg;
+
+typedef struct {
+  tx_1d_cfg row;
+  tx_1d_cfg col;
+} tx_2d_cfg;
+
+tx_2d_cfg inv_tx_cfg(int tx_type, int tx_size_row, int tx_size_col) {
+  const TXFM_2D_CFG *cfg_row = inv_txfm_cfg_ls[tx_type][tx_size_row];
+  const int8_t *stage_range_row = cfg_row->stage_range_row;
+  const int8_t *cos_bit_row = cfg_row->cos_bit_row;
+
+  const TXFM_2D_CFG *cfg_col = inv_txfm_cfg_ls[tx_type][tx_size_col];
+  const int8_t *stage_range_col = cfg_col->stage_range_col;
+  const int8_t *cos_bit_col = cfg_col->cos_bit_col;
+
+  tx_2d_cfg cfg = {
+    { cos_bit_row, stage_range_row }, { cos_bit_col, stage_range_col },
+  };
+  return cfg;
+}
+
+void av1_highbd_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_4[] = {
+    { highbd_idct4, highbd_idct4 },    // DCT_DCT
+    { highbd_iadst4, highbd_idct4 },   // ADST_DCT
+    { highbd_idct4, highbd_iadst4 },   // DCT_ADST
+    { highbd_iadst4, highbd_iadst4 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst4, highbd_idct4 },       // FLIPADST_DCT
+    { highbd_idct4, highbd_iadst4 },       // DCT_FLIPADST
+    { highbd_iadst4, highbd_iadst4 },      // FLIPADST_FLIPADST
+    { highbd_iadst4, highbd_iadst4 },      // ADST_FLIPADST
+    { highbd_iadst4, highbd_iadst4 },      // FLIPADST_ADST
+    { highbd_iidtx4_c, highbd_iidtx4_c },  // IDTX
+    { highbd_idct4, highbd_iidtx4_c },     // V_DCT
+    { highbd_iidtx4_c, highbd_idct4 },     // H_DCT
+    { highbd_iadst4, highbd_iidtx4_c },    // V_ADST
+    { highbd_iidtx4_c, highbd_iadst4 },    // H_ADST
+    { highbd_iadst4, highbd_iidtx4_c },    // V_FLIPADST
+    { highbd_iidtx4_c, highbd_iadst4 },    // H_FLIPADST
+#endif                                     // CONFIG_EXT_TX
+  };
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t tmp;
+  tran_low_t out[4][4];
+  tran_low_t *outp = &out[0][0];
+  int outstride = 4;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_4X4, TX_4X4);
+
+  // inverse transform row vectors
+  for (i = 0; i < 4; ++i) {
+    HIGH_IHT_4[tx_type].rows(input, out[i], cfg.row.cos_bit,
+                             cfg.row.stage_range, bd);
+    input += 4;
+  }
+
+  // transpose
+  for (i = 1; i < 4; i++) {
+    for (j = 0; j < i; j++) {
+      tmp = out[i][j];
+      out[i][j] = out[j][i];
+      out[j][i] = tmp;
+    }
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 4; ++i) {
+    HIGH_IHT_4[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                             cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, 4, 4);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < 4; ++i) {
+    for (j = 0; j < 4; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 4), bd);
+    }
+  }
+}
+
+void av1_highbd_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_4x8[] = {
+    { highbd_idct8, highbd_idct4 },    // DCT_DCT
+    { highbd_iadst8, highbd_idct4 },   // ADST_DCT
+    { highbd_idct8, highbd_iadst4 },   // DCT_ADST
+    { highbd_iadst8, highbd_iadst4 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst8, highbd_idct4 },       // FLIPADST_DCT
+    { highbd_idct8, highbd_iadst4 },       // DCT_FLIPADST
+    { highbd_iadst8, highbd_iadst4 },      // FLIPADST_FLIPADST
+    { highbd_iadst8, highbd_iadst4 },      // ADST_FLIPADST
+    { highbd_iadst8, highbd_iadst4 },      // FLIPADST_ADST
+    { highbd_iidtx8_c, highbd_iidtx4_c },  // IDTX
+    { highbd_idct8, highbd_iidtx4_c },     // V_DCT
+    { highbd_iidtx8_c, highbd_idct4 },     // H_DCT
+    { highbd_iadst8, highbd_iidtx4_c },    // V_ADST
+    { highbd_iidtx8_c, highbd_iadst4 },    // H_ADST
+    { highbd_iadst8, highbd_iidtx4_c },    // V_FLIPADST
+    { highbd_iidtx8_c, highbd_iadst4 },    // H_FLIPADST
+#endif                                     // CONFIG_EXT_TX
+  };
+  const int n = 4;
+  const int n2 = 8;
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[4][8], outtmp[4];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n2;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_4X4, TX_8X8);
+
+  // inverse transform row vectors, and transpose
+  for (i = 0; i < n2; ++i) {
+    HIGH_IHT_4x8[tx_type].rows(input, outtmp, cfg.row.cos_bit,
+                               cfg.row.stage_range, bd);
+    for (j = 0; j < n; ++j) {
+      out[j][i] = HIGHBD_WRAPLOW(dct_const_round_shift(outtmp[j] * Sqrt2), bd);
+    }
+    input += n;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n; ++i) {
+    HIGH_IHT_4x8[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                               cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, n2, n);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5), bd);
+    }
+  }
+}
+
+void av1_highbd_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_8x4[] = {
+    { highbd_idct4, highbd_idct8 },    // DCT_DCT
+    { highbd_iadst4, highbd_idct8 },   // ADST_DCT
+    { highbd_idct4, highbd_iadst8 },   // DCT_ADST
+    { highbd_iadst4, highbd_iadst8 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst4, highbd_idct8 },       // FLIPADST_DCT
+    { highbd_idct4, highbd_iadst8 },       // DCT_FLIPADST
+    { highbd_iadst4, highbd_iadst8 },      // FLIPADST_FLIPADST
+    { highbd_iadst4, highbd_iadst8 },      // ADST_FLIPADST
+    { highbd_iadst4, highbd_iadst8 },      // FLIPADST_ADST
+    { highbd_iidtx4_c, highbd_iidtx8_c },  // IDTX
+    { highbd_idct4, highbd_iidtx8_c },     // V_DCT
+    { highbd_iidtx4_c, highbd_idct8 },     // H_DCT
+    { highbd_iadst4, highbd_iidtx8_c },    // V_ADST
+    { highbd_iidtx4_c, highbd_iadst8 },    // H_ADST
+    { highbd_iadst4, highbd_iidtx8_c },    // V_FLIPADST
+    { highbd_iidtx4_c, highbd_iadst8 },    // H_FLIPADST
+#endif                                     // CONFIG_EXT_TX
+  };
+  const int n = 4;
+  const int n2 = 8;
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[8][4], outtmp[8];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_8X8, TX_4X4);
+
+  // inverse transform row vectors, and transpose
+  for (i = 0; i < n; ++i) {
+    HIGH_IHT_8x4[tx_type].rows(input, outtmp, cfg.row.cos_bit,
+                               cfg.row.stage_range, bd);
+    for (j = 0; j < n2; ++j) {
+      out[j][i] = HIGHBD_WRAPLOW(dct_const_round_shift(outtmp[j] * Sqrt2), bd);
+    }
+    input += n2;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n2; ++i) {
+    HIGH_IHT_8x4[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                               cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, n, n2);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5), bd);
+    }
+  }
+}
+
+void av1_highbd_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_4x16[] = {
+    { highbd_idct16, highbd_idct4 },    // DCT_DCT
+    { highbd_iadst16, highbd_idct4 },   // ADST_DCT
+    { highbd_idct16, highbd_iadst4 },   // DCT_ADST
+    { highbd_iadst16, highbd_iadst4 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst16, highbd_idct4 },       // FLIPADST_DCT
+    { highbd_idct16, highbd_iadst4 },       // DCT_FLIPADST
+    { highbd_iadst16, highbd_iadst4 },      // FLIPADST_FLIPADST
+    { highbd_iadst16, highbd_iadst4 },      // ADST_FLIPADST
+    { highbd_iadst16, highbd_iadst4 },      // FLIPADST_ADST
+    { highbd_iidtx16_c, highbd_iidtx4_c },  // IDTX
+    { highbd_idct16, highbd_iidtx4_c },     // V_DCT
+    { highbd_iidtx16_c, highbd_idct4 },     // H_DCT
+    { highbd_iadst16, highbd_iidtx4_c },    // V_ADST
+    { highbd_iidtx16_c, highbd_iadst4 },    // H_ADST
+    { highbd_iadst16, highbd_iidtx4_c },    // V_FLIPADST
+    { highbd_iidtx16_c, highbd_iadst4 },    // H_FLIPADST
+#endif                                      // CONFIG_EXT_TX
+  };
+  const int n = 4;
+  const int n4 = 16;
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[4][16], outtmp[4];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n4;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_4X4, TX_16X16);
+
+  // inverse transform row vectors, and transpose
+  for (i = 0; i < n4; ++i) {
+    HIGH_IHT_4x16[tx_type].rows(input, outtmp, cfg.row.cos_bit,
+                                cfg.row.stage_range, bd);
+    for (j = 0; j < n; ++j) out[j][i] = outtmp[j];
+    input += n;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n; ++i)
+    HIGH_IHT_4x16[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                                cfg.col.stage_range, bd);
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, n4, n);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < n4; ++i) {
+    for (j = 0; j < n; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5), bd);
+    }
+  }
+}
+
+void av1_highbd_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest8,
+                                 int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_16x4[] = {
+    { highbd_idct4, highbd_idct16 },    // DCT_DCT
+    { highbd_iadst4, highbd_idct16 },   // ADST_DCT
+    { highbd_idct4, highbd_iadst16 },   // DCT_ADST
+    { highbd_iadst4, highbd_iadst16 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst4, highbd_idct16 },       // FLIPADST_DCT
+    { highbd_idct4, highbd_iadst16 },       // DCT_FLIPADST
+    { highbd_iadst4, highbd_iadst16 },      // FLIPADST_FLIPADST
+    { highbd_iadst4, highbd_iadst16 },      // ADST_FLIPADST
+    { highbd_iadst4, highbd_iadst16 },      // FLIPADST_ADST
+    { highbd_iidtx4_c, highbd_iidtx16_c },  // IDTX
+    { highbd_idct4, highbd_iidtx16_c },     // V_DCT
+    { highbd_iidtx4_c, highbd_idct16 },     // H_DCT
+    { highbd_iadst4, highbd_iidtx16_c },    // V_ADST
+    { highbd_iidtx4_c, highbd_iadst16 },    // H_ADST
+    { highbd_iadst4, highbd_iidtx16_c },    // V_FLIPADST
+    { highbd_iidtx4_c, highbd_iadst16 },    // H_FLIPADST
+#endif                                      // CONFIG_EXT_TX
+  };
+  const int n = 4;
+  const int n4 = 16;
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[16][4], outtmp[16];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_16X16, TX_4X4);
+
+  // inverse transform row vectors, and transpose
+  for (i = 0; i < n; ++i) {
+    HIGH_IHT_16x4[tx_type].rows(input, outtmp, cfg.row.cos_bit,
+                                cfg.row.stage_range, bd);
+    for (j = 0; j < n4; ++j) out[j][i] = outtmp[j];
+    input += n4;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n4; ++i) {
+    HIGH_IHT_16x4[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                                cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, n, n4);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n4; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5), bd);
+    }
+  }
+}
+
+void av1_highbd_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_8x16[] = {
+    { highbd_idct16, highbd_idct8 },    // DCT_DCT
+    { highbd_iadst16, highbd_idct8 },   // ADST_DCT
+    { highbd_idct16, highbd_iadst8 },   // DCT_ADST
+    { highbd_iadst16, highbd_iadst8 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst16, highbd_idct8 },       // FLIPADST_DCT
+    { highbd_idct16, highbd_iadst8 },       // DCT_FLIPADST
+    { highbd_iadst16, highbd_iadst8 },      // FLIPADST_FLIPADST
+    { highbd_iadst16, highbd_iadst8 },      // ADST_FLIPADST
+    { highbd_iadst16, highbd_iadst8 },      // FLIPADST_ADST
+    { highbd_iidtx16_c, highbd_iidtx8_c },  // IDTX
+    { highbd_idct16, highbd_iidtx8_c },     // V_DCT
+    { highbd_iidtx16_c, highbd_idct8 },     // H_DCT
+    { highbd_iadst16, highbd_iidtx8_c },    // V_ADST
+    { highbd_iidtx16_c, highbd_iadst8 },    // H_ADST
+    { highbd_iadst16, highbd_iidtx8_c },    // V_FLIPADST
+    { highbd_iidtx16_c, highbd_iadst8 },    // H_FLIPADST
+#endif                                      // CONFIG_EXT_TX
+  };
+  const int n = 8;
+  const int n2 = 16;
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[8][16], outtmp[8];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n2;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_8X8, TX_16X16);
+
+  // inverse transform row vectors, and transpose
+  for (i = 0; i < n2; ++i) {
+    HIGH_IHT_8x16[tx_type].rows(input, outtmp, cfg.row.cos_bit,
+                                cfg.row.stage_range, bd);
+    for (j = 0; j < n; ++j)
+      out[j][i] = HIGHBD_WRAPLOW(dct_const_round_shift(outtmp[j] * Sqrt2), bd);
+    input += n;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n; ++i) {
+    HIGH_IHT_8x16[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                                cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, n2, n);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6), bd);
+    }
+  }
+}
+
+void av1_highbd_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_16x8[] = {
+    { highbd_idct8, highbd_idct16 },    // DCT_DCT
+    { highbd_iadst8, highbd_idct16 },   // ADST_DCT
+    { highbd_idct8, highbd_iadst16 },   // DCT_ADST
+    { highbd_iadst8, highbd_iadst16 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst8, highbd_idct16 },       // FLIPADST_DCT
+    { highbd_idct8, highbd_iadst16 },       // DCT_FLIPADST
+    { highbd_iadst8, highbd_iadst16 },      // FLIPADST_FLIPADST
+    { highbd_iadst8, highbd_iadst16 },      // ADST_FLIPADST
+    { highbd_iadst8, highbd_iadst16 },      // FLIPADST_ADST
+    { highbd_iidtx8_c, highbd_iidtx16_c },  // IDTX
+    { highbd_idct8, highbd_iidtx16_c },     // V_DCT
+    { highbd_iidtx8_c, highbd_idct16 },     // H_DCT
+    { highbd_iadst8, highbd_iidtx16_c },    // V_ADST
+    { highbd_iidtx8_c, highbd_iadst16 },    // H_ADST
+    { highbd_iadst8, highbd_iidtx16_c },    // V_FLIPADST
+    { highbd_iidtx8_c, highbd_iadst16 },    // H_FLIPADST
+#endif                                      // CONFIG_EXT_TX
+  };
+  const int n = 8;
+  const int n2 = 16;
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[16][8], outtmp[16];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_16X16, TX_8X8);
+
+  // inverse transform row vectors, and transpose
+  for (i = 0; i < n; ++i) {
+    HIGH_IHT_16x8[tx_type].rows(input, outtmp, cfg.row.cos_bit,
+                                cfg.row.stage_range, bd);
+    for (j = 0; j < n2; ++j)
+      out[j][i] = HIGHBD_WRAPLOW(dct_const_round_shift(outtmp[j] * Sqrt2), bd);
+    input += n2;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n2; ++i) {
+    HIGH_IHT_16x8[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                                cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, n, n2);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6), bd);
+    }
+  }
+}
+
+void av1_highbd_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_8x32[] = {
+    { highbd_idct32, highbd_idct8 },           // DCT_DCT
+    { highbd_ihalfright32_c, highbd_idct8 },   // ADST_DCT
+    { highbd_idct32, highbd_iadst8 },          // DCT_ADST
+    { highbd_ihalfright32_c, highbd_iadst8 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_ihalfright32_c, highbd_idct8 },     // FLIPADST_DCT
+    { highbd_idct32, highbd_iadst8 },            // DCT_FLIPADST
+    { highbd_ihalfright32_c, highbd_iadst8 },    // FLIPADST_FLIPADST
+    { highbd_ihalfright32_c, highbd_iadst8 },    // ADST_FLIPADST
+    { highbd_ihalfright32_c, highbd_iadst8 },    // FLIPADST_ADST
+    { highbd_iidtx32_c, highbd_iidtx8_c },       // IDTX
+    { highbd_idct32, highbd_iidtx8_c },          // V_DCT
+    { highbd_iidtx32_c, highbd_idct8 },          // H_DCT
+    { highbd_ihalfright32_c, highbd_iidtx8_c },  // V_ADST
+    { highbd_iidtx32_c, highbd_iadst8 },         // H_ADST
+    { highbd_ihalfright32_c, highbd_iidtx8_c },  // V_FLIPADST
+    { highbd_iidtx32_c, highbd_iadst8 },         // H_FLIPADST
+#endif                                           // CONFIG_EXT_TX
+  };
+  const int n = 8;
+  const int n4 = 32;
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[8][32], outtmp[8];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n4;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_8X8, TX_32X32);
+
+  // inverse transform row vectors, and transpose
+  for (i = 0; i < n4; ++i) {
+    HIGH_IHT_8x32[tx_type].rows(input, outtmp, cfg.row.cos_bit,
+                                cfg.row.stage_range, bd);
+    for (j = 0; j < n; ++j) out[j][i] = outtmp[j];
+    input += n;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n; ++i)
+    HIGH_IHT_8x32[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                                cfg.col.stage_range, bd);
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, n4, n);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < n4; ++i) {
+    for (j = 0; j < n; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6), bd);
+    }
+  }
+}
+
+void av1_highbd_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest8,
+                                  int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_32x8[] = {
+    { highbd_idct8, highbd_idct32 },           // DCT_DCT
+    { highbd_iadst8, highbd_idct32 },          // ADST_DCT
+    { highbd_idct8, highbd_ihalfright32_c },   // DCT_ADST
+    { highbd_iadst8, highbd_ihalfright32_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst8, highbd_idct32 },            // FLIPADST_DCT
+    { highbd_idct8, highbd_ihalfright32_c },     // DCT_FLIPADST
+    { highbd_iadst8, highbd_ihalfright32_c },    // FLIPADST_FLIPADST
+    { highbd_iadst8, highbd_ihalfright32_c },    // ADST_FLIPADST
+    { highbd_iadst8, highbd_ihalfright32_c },    // FLIPADST_ADST
+    { highbd_iidtx8_c, highbd_iidtx32_c },       // IDTX
+    { highbd_idct8, highbd_iidtx32_c },          // V_DCT
+    { highbd_iidtx8_c, highbd_idct32 },          // H_DCT
+    { highbd_iadst8, highbd_iidtx32_c },         // V_ADST
+    { highbd_iidtx8_c, highbd_ihalfright32_c },  // H_ADST
+    { highbd_iadst8, highbd_iidtx32_c },         // V_FLIPADST
+    { highbd_iidtx8_c, highbd_ihalfright32_c },  // H_FLIPADST
+#endif                                           // CONFIG_EXT_TX
+  };
+  const int n = 8;
+  const int n4 = 32;
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[32][8], outtmp[32];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_32X32, TX_8X8);
+
+  // inverse transform row vectors, and transpose
+  for (i = 0; i < n; ++i) {
+    HIGH_IHT_32x8[tx_type].rows(input, outtmp, cfg.row.cos_bit,
+                                cfg.row.stage_range, bd);
+    for (j = 0; j < n4; ++j) out[j][i] = outtmp[j];
+    input += n4;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n4; ++i)
+    HIGH_IHT_32x8[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                                cfg.col.stage_range, bd);
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, n, n4);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n4; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6), bd);
+    }
+  }
+}
+
+void av1_highbd_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_16x32[] = {
+    { highbd_idct32, highbd_idct16 },           // DCT_DCT
+    { highbd_ihalfright32_c, highbd_idct16 },   // ADST_DCT
+    { highbd_idct32, highbd_iadst16 },          // DCT_ADST
+    { highbd_ihalfright32_c, highbd_iadst16 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_ihalfright32_c, highbd_idct16 },     // FLIPADST_DCT
+    { highbd_idct32, highbd_iadst16 },            // DCT_FLIPADST
+    { highbd_ihalfright32_c, highbd_iadst16 },    // FLIPADST_FLIPADST
+    { highbd_ihalfright32_c, highbd_iadst16 },    // ADST_FLIPADST
+    { highbd_ihalfright32_c, highbd_iadst16 },    // FLIPADST_ADST
+    { highbd_iidtx32_c, highbd_iidtx16_c },       // IDTX
+    { highbd_idct32, highbd_iidtx16_c },          // V_DCT
+    { highbd_iidtx32_c, highbd_idct16 },          // H_DCT
+    { highbd_ihalfright32_c, highbd_iidtx16_c },  // V_ADST
+    { highbd_iidtx32_c, highbd_iadst16 },         // H_ADST
+    { highbd_ihalfright32_c, highbd_iidtx16_c },  // V_FLIPADST
+    { highbd_iidtx32_c, highbd_iadst16 },         // H_FLIPADST
+#endif                                            // CONFIG_EXT_TX
+  };
+  const int n = 16;
+  const int n2 = 32;
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[16][32], outtmp[16];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n2;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_16X16, TX_32X32);
+
+  // inverse transform row vectors, and transpose
+  for (i = 0; i < n2; ++i) {
+    HIGH_IHT_16x32[tx_type].rows(input, outtmp, cfg.row.cos_bit,
+                                 cfg.row.stage_range, bd);
+    for (j = 0; j < n; ++j)
+      out[j][i] = HIGHBD_WRAPLOW(dct_const_round_shift(outtmp[j] * Sqrt2), bd);
+    input += n;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n; ++i) {
+    HIGH_IHT_16x32[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                                 cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, n2, n);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6), bd);
+    }
+  }
+}
+
+void av1_highbd_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_32x16[] = {
+    { highbd_idct16, highbd_idct32 },           // DCT_DCT
+    { highbd_iadst16, highbd_idct32 },          // ADST_DCT
+    { highbd_idct16, highbd_ihalfright32_c },   // DCT_ADST
+    { highbd_iadst16, highbd_ihalfright32_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst16, highbd_idct32 },            // FLIPADST_DCT
+    { highbd_idct16, highbd_ihalfright32_c },     // DCT_FLIPADST
+    { highbd_iadst16, highbd_ihalfright32_c },    // FLIPADST_FLIPADST
+    { highbd_iadst16, highbd_ihalfright32_c },    // ADST_FLIPADST
+    { highbd_iadst16, highbd_ihalfright32_c },    // FLIPADST_ADST
+    { highbd_iidtx16_c, highbd_iidtx32_c },       // IDTX
+    { highbd_idct16, highbd_iidtx32_c },          // V_DCT
+    { highbd_iidtx16_c, highbd_idct32 },          // H_DCT
+    { highbd_iadst16, highbd_iidtx32_c },         // V_ADST
+    { highbd_iidtx16_c, highbd_ihalfright32_c },  // H_ADST
+    { highbd_iadst16, highbd_iidtx32_c },         // V_FLIPADST
+    { highbd_iidtx16_c, highbd_ihalfright32_c },  // H_FLIPADST
+#endif                                            // CONFIG_EXT_TX
+  };
+  const int n = 16;
+  const int n2 = 32;
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t out[32][16], outtmp[32];
+  tran_low_t *outp = &out[0][0];
+  int outstride = n;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_32X32, TX_16X16);
+
+  // inverse transform row vectors, and transpose
+  for (i = 0; i < n; ++i) {
+    HIGH_IHT_32x16[tx_type].rows(input, outtmp, cfg.row.cos_bit,
+                                 cfg.row.stage_range, bd);
+    for (j = 0; j < n2; ++j)
+      out[j][i] = HIGHBD_WRAPLOW(dct_const_round_shift(outtmp[j] * Sqrt2), bd);
+    input += n2;
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < n2; ++i) {
+    HIGH_IHT_32x16[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                                 cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, n, n2);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6), bd);
+    }
+  }
+}
+
+void av1_highbd_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest8,
+                                int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_8[] = {
+    { highbd_idct8, highbd_idct8 },    // DCT_DCT
+    { highbd_iadst8, highbd_idct8 },   // ADST_DCT
+    { highbd_idct8, highbd_iadst8 },   // DCT_ADST
+    { highbd_iadst8, highbd_iadst8 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst8, highbd_idct8 },       // FLIPADST_DCT
+    { highbd_idct8, highbd_iadst8 },       // DCT_FLIPADST
+    { highbd_iadst8, highbd_iadst8 },      // FLIPADST_FLIPADST
+    { highbd_iadst8, highbd_iadst8 },      // ADST_FLIPADST
+    { highbd_iadst8, highbd_iadst8 },      // FLIPADST_ADST
+    { highbd_iidtx8_c, highbd_iidtx8_c },  // IDTX
+    { highbd_idct8, highbd_iidtx8_c },     // V_DCT
+    { highbd_iidtx8_c, highbd_idct8 },     // H_DCT
+    { highbd_iadst8, highbd_iidtx8_c },    // V_ADST
+    { highbd_iidtx8_c, highbd_iadst8 },    // H_ADST
+    { highbd_iadst8, highbd_iidtx8_c },    // V_FLIPADST
+    { highbd_iidtx8_c, highbd_iadst8 },    // H_FLIPADST
+#endif                                     // CONFIG_EXT_TX
+  };
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t tmp;
+  tran_low_t out[8][8];
+  tran_low_t *outp = &out[0][0];
+  int outstride = 8;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_8X8, TX_8X8);
+
+  // inverse transform row vectors
+  for (i = 0; i < 8; ++i) {
+    HIGH_IHT_8[tx_type].rows(input, out[i], cfg.row.cos_bit,
+                             cfg.row.stage_range, bd);
+    input += 8;
+  }
+
+  // transpose
+  for (i = 1; i < 8; i++) {
+    for (j = 0; j < i; j++) {
+      tmp = out[i][j];
+      out[i][j] = out[j][i];
+      out[j][i] = tmp;
+    }
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 8; ++i) {
+    HIGH_IHT_8[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                             cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, 8, 8);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < 8; ++i) {
+    for (j = 0; j < 8; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5), bd);
+    }
+  }
+}
+
+void av1_highbd_iht16x16_256_add_c(const tran_low_t *input, uint8_t *dest8,
+                                   int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_16[] = {
+    { highbd_idct16, highbd_idct16 },    // DCT_DCT
+    { highbd_iadst16, highbd_idct16 },   // ADST_DCT
+    { highbd_idct16, highbd_iadst16 },   // DCT_ADST
+    { highbd_iadst16, highbd_iadst16 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_iadst16, highbd_idct16 },       // FLIPADST_DCT
+    { highbd_idct16, highbd_iadst16 },       // DCT_FLIPADST
+    { highbd_iadst16, highbd_iadst16 },      // FLIPADST_FLIPADST
+    { highbd_iadst16, highbd_iadst16 },      // ADST_FLIPADST
+    { highbd_iadst16, highbd_iadst16 },      // FLIPADST_ADST
+    { highbd_iidtx16_c, highbd_iidtx16_c },  // IDTX
+    { highbd_idct16, highbd_iidtx16_c },     // V_DCT
+    { highbd_iidtx16_c, highbd_idct16 },     // H_DCT
+    { highbd_iadst16, highbd_iidtx16_c },    // V_ADST
+    { highbd_iidtx16_c, highbd_iadst16 },    // H_ADST
+    { highbd_iadst16, highbd_iidtx16_c },    // V_FLIPADST
+    { highbd_iidtx16_c, highbd_iadst16 },    // H_FLIPADST
+#endif                                       // CONFIG_EXT_TX
+  };
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t tmp;
+  tran_low_t out[16][16];
+  tran_low_t *outp = &out[0][0];
+  int outstride = 16;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_16X16, TX_16X16);
+
+  // inverse transform row vectors
+  for (i = 0; i < 16; ++i) {
+    HIGH_IHT_16[tx_type].rows(input, out[i], cfg.row.cos_bit,
+                              cfg.row.stage_range, bd);
+    input += 16;
+  }
+
+  // transpose
+  for (i = 1; i < 16; i++) {
+    for (j = 0; j < i; j++) {
+      tmp = out[i][j];
+      out[i][j] = out[j][i];
+      out[j][i] = tmp;
+    }
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 16; ++i) {
+    HIGH_IHT_16[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                              cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, 16, 16);
+#endif
+
+  // Sum with the destination
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6), bd);
+    }
+  }
+}
+
+#if CONFIG_EXT_TX
+static void highbd_iht32x32_1024_add_c(const tran_low_t *input, uint8_t *dest8,
+                                       int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_32[] = {
+    { highbd_idct32, highbd_idct32 },                  // DCT_DCT
+    { highbd_ihalfright32_c, highbd_idct32 },          // ADST_DCT
+    { highbd_idct32, highbd_ihalfright32_c },          // DCT_ADST
+    { highbd_ihalfright32_c, highbd_ihalfright32_c },  // ADST_ADST
+    { highbd_ihalfright32_c, highbd_idct32 },          // FLIPADST_DCT
+    { highbd_idct32, highbd_ihalfright32_c },          // DCT_FLIPADST
+    { highbd_ihalfright32_c, highbd_ihalfright32_c },  // FLIPADST_FLIPADST
+    { highbd_ihalfright32_c, highbd_ihalfright32_c },  // ADST_FLIPADST
+    { highbd_ihalfright32_c, highbd_ihalfright32_c },  // FLIPADST_ADST
+    { highbd_iidtx32_c, highbd_iidtx32_c },            // IDTX
+    { highbd_idct32, highbd_iidtx32_c },               // V_DCT
+    { highbd_iidtx32_c, highbd_idct32 },               // H_DCT
+    { highbd_ihalfright32_c, highbd_iidtx32_c },       // V_ADST
+    { highbd_iidtx32_c, highbd_ihalfright32_c },       // H_ADST
+    { highbd_ihalfright32_c, highbd_iidtx32_c },       // V_FLIPADST
+    { highbd_iidtx32_c, highbd_ihalfright32_c },       // H_FLIPADST
+  };
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t tmp;
+  tran_low_t out[32][32];
+  tran_low_t *outp = &out[0][0];
+  int outstride = 32;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_32X32, TX_32X32);
+
+  // inverse transform row vectors
+  for (i = 0; i < 32; ++i) {
+    HIGH_IHT_32[tx_type].rows(input, out[i], cfg.row.cos_bit,
+                              cfg.row.stage_range, bd);
+    input += 32;
+  }
+
+  // transpose
+  for (i = 1; i < 32; i++) {
+    for (j = 0; j < i; j++) {
+      tmp = out[i][j];
+      out[i][j] = out[j][i];
+      out[j][i] = tmp;
+    }
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 32; ++i) {
+    HIGH_IHT_32[tx_type].cols(out[i], out[i], cfg.col.cos_bit,
+                              cfg.col.stage_range, bd);
+  }
+
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, 32, 32);
+
+  // Sum with the destination
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 6), bd);
+    }
+  }
+}
+#endif  // CONFIG_EXT_TX
+
+#if CONFIG_TX64X64
+static void highbd_iht64x64_4096_add_c(const tran_low_t *input, uint8_t *dest8,
+                                       int stride, int tx_type, int bd) {
+  static const highbd_transform_2d HIGH_IHT_64[] = {
+    { highbd_idct64_col_c, highbd_idct64_row_c },      // DCT_DCT
+    { highbd_ihalfright64_c, highbd_idct64_row_c },    // ADST_DCT
+    { highbd_idct64_col_c, highbd_ihalfright64_c },    // DCT_ADST
+    { highbd_ihalfright64_c, highbd_ihalfright64_c },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { highbd_ihalfright64_c, highbd_idct64_row_c },    // FLIPADST_DCT
+    { highbd_idct64_col_c, highbd_ihalfright64_c },    // DCT_FLIPADST
+    { highbd_ihalfright64_c, highbd_ihalfright64_c },  // FLIPADST_FLIPADST
+    { highbd_ihalfright64_c, highbd_ihalfright64_c },  // ADST_FLIPADST
+    { highbd_ihalfright64_c, highbd_ihalfright64_c },  // FLIPADST_ADST
+    { highbd_iidtx64_c, highbd_iidtx64_c },            // IDTX
+    { highbd_idct64_col_c, highbd_iidtx64_c },         // V_DCT
+    { highbd_iidtx64_c, highbd_idct64_row_c },         // H_DCT
+    { highbd_ihalfright64_c, highbd_iidtx64_c },       // V_ADST
+    { highbd_iidtx64_c, highbd_ihalfright64_c },       // H_ADST
+    { highbd_ihalfright64_c, highbd_iidtx64_c },       // V_FLIPADST
+    { highbd_iidtx64_c, highbd_ihalfright64_c },       // H_FLIPADST
+#endif                                                 // CONFIG_EXT_TX
+  };
+
+  uint16_t *dest = CONVERT_TO_SHORTPTR(dest8);
+
+  int i, j;
+  tran_low_t tmp;
+  tran_low_t out[64][64];
+  tran_low_t *outp = &out[0][0];
+  int outstride = 64;
+
+  tx_2d_cfg cfg = inv_tx_cfg(tx_type, TX_64X64, TX_64X64);
+
+  // inverse transform row vectors
+  for (i = 0; i < 64; ++i) {
+    HIGH_IHT_64[tx_type].rows(input, out[i], cfg.row.cos_bit,
+                              cfg.row.stage_range, bd);
+    for (j = 0; j < 64; ++j) out[i][j] = ROUND_POWER_OF_TWO(out[i][j], 1);
+    input += 64;
+  }
+
+  // transpose
+  for (i = 1; i < 64; i++) {
+    for (j = 0; j < i; j++) {
+      tmp = out[i][j];
+      out[i][j] = out[j][i];
+      out[j][i] = tmp;
+    }
+  }
+
+  // inverse transform column vectors
+  for (i = 0; i < 64; ++i) {
+    HIGH_IHT_64[tx_type].cols(out[i], out[i], cfg.col.cos_bit_col,
+                              cfg.col.stage_range, bd);
+  }
+
+#if CONFIG_EXT_TX
+  maybe_flip_strides16(&dest, &stride, &outp, &outstride, tx_type, 64, 64);
+#endif  // CONFIG_EXT_TX
+
+  // Sum with the destination
+  for (i = 0; i < 64; ++i) {
+    for (j = 0; j < 64; ++j) {
+      int d = i * stride + j;
+      int s = j * outstride + i;
+      dest[d] =
+          highbd_clip_pixel_add(dest[d], ROUND_POWER_OF_TWO(outp[s], 5), bd);
+    }
+  }
+}
+#endif  // CONFIG_TX64X64
+
+// idct
+void av1_highbd_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd) {
+  if (eob > 1)
+    aom_highbd_idct4x4_16_add(input, dest, stride, bd);
+  else
+    aom_highbd_idct4x4_1_add(input, dest, stride, bd);
+}
+
+void av1_highbd_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd) {
+  if (eob > 1)
+    aom_highbd_iwht4x4_16_add(input, dest, stride, bd);
+  else
+    aom_highbd_iwht4x4_1_add(input, dest, stride, bd);
+}
+
+#if CONFIG_CB4X4
+static void highbd_inv_txfm_add_2x2(const tran_low_t *input, uint8_t *dest,
+                                    int stride, int eob, int bd,
+                                    TX_TYPE tx_type, int lossless) {
+  tran_high_t a1 = input[0] >> UNIT_QUANT_SHIFT;
+  tran_high_t b1 = input[1] >> UNIT_QUANT_SHIFT;
+  tran_high_t c1 = input[2] >> UNIT_QUANT_SHIFT;
+  tran_high_t d1 = input[3] >> UNIT_QUANT_SHIFT;
+
+  tran_high_t a2 = a1 + c1;
+  tran_high_t b2 = b1 + d1;
+  tran_high_t c2 = a1 - c1;
+  tran_high_t d2 = b1 - d1;
+
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dest);
+
+  (void)tx_type;
+  (void)lossless;
+  (void)eob;
+
+  a1 = (a2 + b2) >> 2;
+  b1 = (a2 - b2) >> 2;
+  c1 = (c2 + d2) >> 2;
+  d1 = (c2 - d2) >> 2;
+
+  dst[0] = highbd_clip_pixel_add(dst[0], a1, bd);
+  dst[1] = highbd_clip_pixel_add(dst[1], b1, bd);
+  dst[stride] = highbd_clip_pixel_add(dst[stride], c1, bd);
+  dst[stride + 1] = highbd_clip_pixel_add(dst[stride + 1], d1, bd);
+}
+#endif
+
+void av1_highbd_inv_txfm_add_4x4(const tran_low_t *input, uint8_t *dest,
+                                 int stride, int eob, int bd, TX_TYPE tx_type,
+                                 int lossless) {
+  if (lossless) {
+    assert(tx_type == DCT_DCT);
+    av1_highbd_iwht4x4_add(input, dest, stride, eob, bd);
+    return;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+#endif  // CONFIG_EXT_TX
+      av1_inv_txfm2d_add_4x4(input, CONVERT_TO_SHORTPTR(dest), stride, tx_type,
+                             bd);
+      break;
+#if CONFIG_EXT_TX
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      // Use C version since DST only exists in C code
+      av1_highbd_iht4x4_16_add_c(input, dest, stride, tx_type, bd);
+      break;
+    case IDTX:
+      highbd_inv_idtx_add_c(input, dest, stride, 4, tx_type, bd);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+void av1_highbd_inv_txfm_add_4x8(const tran_low_t *input, uint8_t *dest,
+                                 int stride, int eob, int bd, TX_TYPE tx_type) {
+  (void)eob;
+  av1_highbd_iht4x8_32_add_c(input, dest, stride, tx_type, bd);
+}
+
+void av1_highbd_inv_txfm_add_8x4(const tran_low_t *input, uint8_t *dest,
+                                 int stride, int eob, int bd, TX_TYPE tx_type) {
+  (void)eob;
+  av1_highbd_iht8x4_32_add_c(input, dest, stride, tx_type, bd);
+}
+
+void av1_highbd_inv_txfm_add_4x16(const tran_low_t *input, uint8_t *dest,
+                                  int stride, int eob, int bd,
+                                  TX_TYPE tx_type) {
+  (void)eob;
+  av1_highbd_iht4x16_64_add_c(input, dest, stride, tx_type, bd);
+}
+
+void av1_highbd_inv_txfm_add_16x4(const tran_low_t *input, uint8_t *dest,
+                                  int stride, int eob, int bd,
+                                  TX_TYPE tx_type) {
+  (void)eob;
+  av1_highbd_iht16x4_64_add_c(input, dest, stride, tx_type, bd);
+}
+
+static void highbd_inv_txfm_add_8x16(const tran_low_t *input, uint8_t *dest,
+                                     int stride, int eob, int bd,
+                                     TX_TYPE tx_type) {
+  (void)eob;
+  av1_highbd_iht8x16_128_add_c(input, dest, stride, tx_type, bd);
+}
+
+static void highbd_inv_txfm_add_16x8(const tran_low_t *input, uint8_t *dest,
+                                     int stride, int eob, int bd,
+                                     TX_TYPE tx_type) {
+  (void)eob;
+  av1_highbd_iht16x8_128_add_c(input, dest, stride, tx_type, bd);
+}
+
+void av1_highbd_inv_txfm_add_8x32(const tran_low_t *input, uint8_t *dest,
+                                  int stride, int eob, int bd,
+                                  TX_TYPE tx_type) {
+  (void)eob;
+  av1_highbd_iht8x32_256_add_c(input, dest, stride, tx_type, bd);
+}
+
+void av1_highbd_inv_txfm_add_32x8(const tran_low_t *input, uint8_t *dest,
+                                  int stride, int eob, int bd,
+                                  TX_TYPE tx_type) {
+  (void)eob;
+  av1_highbd_iht32x8_256_add_c(input, dest, stride, tx_type, bd);
+}
+
+static void highbd_inv_txfm_add_16x32(const tran_low_t *input, uint8_t *dest,
+                                      int stride, int eob, int bd,
+                                      TX_TYPE tx_type) {
+  (void)eob;
+  av1_highbd_iht16x32_512_add_c(input, dest, stride, tx_type, bd);
+}
+
+static void highbd_inv_txfm_add_32x16(const tran_low_t *input, uint8_t *dest,
+                                      int stride, int eob, int bd,
+                                      TX_TYPE tx_type) {
+  (void)eob;
+  av1_highbd_iht32x16_512_add_c(input, dest, stride, tx_type, bd);
+}
+
+static void highbd_inv_txfm_add_8x8(const tran_low_t *input, uint8_t *dest,
+                                    int stride, int eob, int bd,
+                                    TX_TYPE tx_type) {
+  (void)eob;
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+#endif  // CONFIG_EXT_TX
+      av1_inv_txfm2d_add_8x8(input, CONVERT_TO_SHORTPTR(dest), stride, tx_type,
+                             bd);
+      break;
+#if CONFIG_EXT_TX
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      // Use C version since DST only exists in C code
+      av1_highbd_iht8x8_64_add_c(input, dest, stride, tx_type, bd);
+      break;
+    case IDTX:
+      highbd_inv_idtx_add_c(input, dest, stride, 8, tx_type, bd);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+static void highbd_inv_txfm_add_16x16(const tran_low_t *input, uint8_t *dest,
+                                      int stride, int eob, int bd,
+                                      TX_TYPE tx_type) {
+  (void)eob;
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+#endif  // CONFIG_EXT_TX
+      av1_inv_txfm2d_add_16x16(input, CONVERT_TO_SHORTPTR(dest), stride,
+                               tx_type, bd);
+      break;
+#if CONFIG_EXT_TX
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      // Use C version since DST only exists in C code
+      av1_highbd_iht16x16_256_add_c(input, dest, stride, tx_type, bd);
+      break;
+    case IDTX:
+      highbd_inv_idtx_add_c(input, dest, stride, 16, tx_type, bd);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+static void highbd_inv_txfm_add_32x32(const tran_low_t *input, uint8_t *dest,
+                                      int stride, int eob, int bd,
+                                      TX_TYPE tx_type) {
+  (void)eob;
+  switch (tx_type) {
+    case DCT_DCT:
+      av1_inv_txfm2d_add_32x32(input, CONVERT_TO_SHORTPTR(dest), stride,
+                               DCT_DCT, bd);
+      break;
+#if CONFIG_EXT_TX
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      highbd_iht32x32_1024_add_c(input, dest, stride, tx_type, bd);
+      break;
+    case IDTX:
+      highbd_inv_idtx_add_c(input, dest, stride, 32, tx_type, bd);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+#if CONFIG_TX64X64
+static void highbd_inv_txfm_add_64x64(const tran_low_t *input, uint8_t *dest,
+                                      int stride, int eob, int bd,
+                                      TX_TYPE tx_type) {
+  (void)eob;
+  switch (tx_type) {
+    case DCT_DCT:
+      av1_inv_txfm2d_add_64x64(input, CONVERT_TO_SHORTPTR(dest), stride,
+                               DCT_DCT, bd);
+      break;
+#if CONFIG_EXT_TX
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      highbd_iht64x64_4096_add_c(input, dest, stride, tx_type, bd);
+      break;
+    case IDTX:
+      highbd_inv_idtx_add_c(input, dest, stride, 64, tx_type, bd);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_inv_txfm_add(const tran_low_t *input, uint8_t *dest, int stride,
+                      INV_TXFM_PARAM *inv_txfm_param) {
+  const TX_TYPE tx_type = inv_txfm_param->tx_type;
+  const TX_SIZE tx_size = inv_txfm_param->tx_size;
+  const int eob = inv_txfm_param->eob;
+  const int lossless = inv_txfm_param->lossless;
+
+  switch (tx_size) {
+#if CONFIG_TX64X64
+    case TX_64X64: inv_txfm_add_64x64(input, dest, stride, eob, tx_type); break;
+#endif  // CONFIG_TX64X64
+    case TX_32X32: inv_txfm_add_32x32(input, dest, stride, eob, tx_type); break;
+    case TX_16X16: inv_txfm_add_16x16(input, dest, stride, eob, tx_type); break;
+    case TX_8X8: inv_txfm_add_8x8(input, dest, stride, eob, tx_type); break;
+    case TX_4X8: av1_inv_txfm_add_4x8(input, dest, stride, eob, tx_type); break;
+    case TX_8X4: av1_inv_txfm_add_8x4(input, dest, stride, eob, tx_type); break;
+    case TX_8X16: inv_txfm_add_8x16(input, dest, stride, eob, tx_type); break;
+    case TX_16X8: inv_txfm_add_16x8(input, dest, stride, eob, tx_type); break;
+    case TX_16X32: inv_txfm_add_16x32(input, dest, stride, eob, tx_type); break;
+    case TX_32X16: inv_txfm_add_32x16(input, dest, stride, eob, tx_type); break;
+    case TX_4X4:
+      // this is like av1_short_idct4x4 but has a special case around eob<=1
+      // which is significant (not just an optimization) for the lossless
+      // case.
+      av1_inv_txfm_add_4x4(input, dest, stride, eob, tx_type, lossless);
+      break;
+#if CONFIG_CB4X4
+    case TX_2X2:
+      inv_txfm_add_2x2(input, dest, stride, eob, tx_type, lossless);
+      break;
+#endif
+    default: assert(0 && "Invalid transform size"); break;
+  }
+}
+
+static void init_inv_txfm_param(const MACROBLOCKD *xd, TX_SIZE tx_size,
+                                TX_TYPE tx_type, int eob, INV_TXFM_PARAM *inv) {
+  inv->tx_type = tx_type;
+  inv->tx_size = tx_size;
+  inv->eob = eob;
+  inv->lossless = xd->lossless[xd->mi[0]->mbmi.segment_id];
+#if CONFIG_HIGHBITDEPTH
+  inv->bd = xd->bd;
+#endif
+#if CONFIG_ADAPT_SCAN
+  inv->eob_threshold = &xd->eob_threshold_md[tx_size][tx_type][0];
+#endif
+}
+
+void av1_inverse_transform_block(const MACROBLOCKD *xd,
+                                 const tran_low_t *dqcoeff, TX_TYPE tx_type,
+                                 TX_SIZE tx_size, uint8_t *dst, int stride,
+                                 int eob) {
+  if (!eob) return;
+#if CONFIG_PVQ
+  const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
+  const int txb_width = block_size_wide[tx_bsize];
+  const int txb_height = block_size_high[tx_bsize];
+  int r, c;
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
+    for (r = 0; r < txb_height; r++)
+      for (c = 0; c < txb_width; c++)
+        CONVERT_TO_SHORTPTR(dst)[r * stride + c] = 0;
+  } else {
+#endif  // CONFIG_HIGHBITDEPTH
+    for (r = 0; r < txb_height; r++)
+      for (c = 0; c < txb_width; c++) dst[r * stride + c] = 0;
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_PVQ
+  INV_TXFM_PARAM inv_txfm_param;
+  init_inv_txfm_param(xd, tx_size, tx_type, eob, &inv_txfm_param);
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    av1_highbd_inv_txfm_add(dqcoeff, dst, stride, &inv_txfm_param);
+  } else {
+#endif  // CONFIG_HIGHBITDEPTH
+    av1_inv_txfm_add(dqcoeff, dst, stride, &inv_txfm_param);
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+}
+
+void av1_inverse_transform_block_facade(MACROBLOCKD *xd, int plane, int block,
+                                        int blk_row, int blk_col, int eob) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  const PLANE_TYPE plane_type = get_plane_type(plane);
+  const TX_SIZE tx_size = get_tx_size(plane, xd);
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const int dst_stride = pd->dst.stride;
+  uint8_t *dst =
+      &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
+  av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, dst, dst_stride,
+                              eob);
+}
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_inv_txfm_add(const tran_low_t *input, uint8_t *dest, int stride,
+                             INV_TXFM_PARAM *inv_txfm_param) {
+  const TX_TYPE tx_type = inv_txfm_param->tx_type;
+  const TX_SIZE tx_size = inv_txfm_param->tx_size;
+  const int eob = inv_txfm_param->eob;
+  const int bd = inv_txfm_param->bd;
+  const int lossless = inv_txfm_param->lossless;
+
+  switch (tx_size) {
+#if CONFIG_TX64X64
+    case TX_64X64:
+      highbd_inv_txfm_add_64x64(input, dest, stride, eob, bd, tx_type);
+      break;
+#endif  // CONFIG_TX64X64
+    case TX_32X32:
+      highbd_inv_txfm_add_32x32(input, dest, stride, eob, bd, tx_type);
+      break;
+    case TX_16X16:
+      highbd_inv_txfm_add_16x16(input, dest, stride, eob, bd, tx_type);
+      break;
+    case TX_8X8:
+      highbd_inv_txfm_add_8x8(input, dest, stride, eob, bd, tx_type);
+      break;
+    case TX_4X8:
+      av1_highbd_inv_txfm_add_4x8(input, dest, stride, eob, bd, tx_type);
+      break;
+    case TX_8X4:
+      av1_highbd_inv_txfm_add_8x4(input, dest, stride, eob, bd, tx_type);
+      break;
+    case TX_8X16:
+      highbd_inv_txfm_add_8x16(input, dest, stride, eob, bd, tx_type);
+      break;
+    case TX_16X8:
+      highbd_inv_txfm_add_16x8(input, dest, stride, eob, bd, tx_type);
+      break;
+    case TX_16X32:
+      highbd_inv_txfm_add_16x32(input, dest, stride, eob, bd, tx_type);
+      break;
+    case TX_32X16:
+      highbd_inv_txfm_add_32x16(input, dest, stride, eob, bd, tx_type);
+      break;
+    case TX_4X4:
+      // this is like av1_short_idct4x4 but has a special case around eob<=1
+      // which is significant (not just an optimization) for the lossless
+      // case.
+      av1_highbd_inv_txfm_add_4x4(input, dest, stride, eob, bd, tx_type,
+                                  lossless);
+      break;
+#if CONFIG_CB4X4
+    case TX_2X2:
+      highbd_inv_txfm_add_2x2(input, dest, stride, eob, bd, tx_type, lossless);
+      break;
+#endif
+    default: assert(0 && "Invalid transform size"); break;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/av1/common/idct.h b/third_party/aom/av1/common/idct.h
new file mode 100644
index 0000000000..e3a192187f
--- /dev/null
+++ b/third_party/aom/av1/common/idct.h
@@ -0,0 +1,99 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_IDCT_H_
+#define AV1_COMMON_IDCT_H_
+
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "av1/common/blockd.h"
+#include "av1/common/common.h"
+#include "av1/common/enums.h"
+#include "aom_dsp/inv_txfm.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_ports/mem.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct INV_TXFM_PARAM {
+#if CONFIG_ADAPT_SCAN
+  const int16_t *eob_threshold;
+#endif
+  TX_TYPE tx_type;
+  TX_SIZE tx_size;
+  int eob;
+  int lossless;
+#if CONFIG_HIGHBITDEPTH
+  int bd;
+#endif
+} INV_TXFM_PARAM;
+
+typedef void (*transform_1d)(const tran_low_t *, tran_low_t *);
+
+typedef struct {
+  transform_1d cols, rows;  // vertical and horizontal
+} transform_2d;
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*highbd_transform_1d)(const tran_low_t *, tran_low_t *,
+                                    const int8_t *cos_bit,
+                                    const int8_t *stage_range, int bd);
+
+typedef struct {
+  highbd_transform_1d cols, rows;  // vertical and horizontal
+} highbd_transform_2d;
+#endif  // CONFIG_HIGHBITDEPTH
+
+#define MAX_TX_SCALE 1
+int av1_get_tx_scale(const TX_SIZE tx_size);
+
+void av1_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+void av1_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                     int eob);
+
+void av1_inv_txfm_add_4x4(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, TX_TYPE tx_type, int lossless);
+void av1_inv_txfm_add_8x4(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, TX_TYPE tx_type);
+void av1_inv_txfm_add_4x8(const tran_low_t *input, uint8_t *dest, int stride,
+                          int eob, TX_TYPE tx_type);
+void av1_inv_txfm_add(const tran_low_t *input, uint8_t *dest, int stride,
+                      INV_TXFM_PARAM *inv_txfm_param);
+void av1_inverse_transform_block(const MACROBLOCKD *xd,
+                                 const tran_low_t *dqcoeff, TX_TYPE tx_type,
+                                 TX_SIZE tx_size, uint8_t *dst, int stride,
+                                 int eob);
+void av1_inverse_transform_block_facade(MACROBLOCKD *xd, int plane, int block,
+                                        int blk_row, int blk_col, int eob);
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_iwht4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd);
+void av1_highbd_idct4x4_add(const tran_low_t *input, uint8_t *dest, int stride,
+                            int eob, int bd);
+void av1_highbd_inv_txfm_add_4x4(const tran_low_t *input, uint8_t *dest,
+                                 int stride, int eob, int bd, TX_TYPE tx_type,
+                                 int lossless);
+void av1_highbd_inv_txfm_add_4x8(const tran_low_t *input, uint8_t *dest,
+                                 int stride, int eob, int bd, TX_TYPE tx_type);
+void av1_highbd_inv_txfm_add_8x4(const tran_low_t *input, uint8_t *dest,
+                                 int stride, int eob, int bd, TX_TYPE tx_type);
+void av1_highbd_inv_txfm_add(const tran_low_t *input, uint8_t *dest, int stride,
+                             INV_TXFM_PARAM *inv_txfm_param);
+#endif  // CONFIG_HIGHBITDEPTH
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_IDCT_H_
diff --git a/third_party/aom/av1/common/laplace_tables.c b/third_party/aom/av1/common/laplace_tables.c
new file mode 100644
index 0000000000..ab87848955
--- /dev/null
+++ b/third_party/aom/av1/common/laplace_tables.c
@@ -0,0 +1,657 @@
+/* This file is auto-generated using "gen_laplace_tables 128 7" */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include "aom_dsp/prob.h"
+#include "pvq.h"
+
+const uint16_t EXP_CDF_TABLE[128][16] = {
+  {AOM_ICDF(32753), AOM_ICDF(32754), AOM_ICDF(32755), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(32499), AOM_ICDF(32753), AOM_ICDF(32755), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(32243), AOM_ICDF(32747), AOM_ICDF(32755), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(31987), AOM_ICDF(32737), AOM_ICDF(32755), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(31732), AOM_ICDF(32724), AOM_ICDF(32755), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(31476), AOM_ICDF(32706), AOM_ICDF(32754), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(31220), AOM_ICDF(32684), AOM_ICDF(32753), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(30964), AOM_ICDF(32658), AOM_ICDF(32751), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(30708), AOM_ICDF(32628), AOM_ICDF(32748), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(30452), AOM_ICDF(32594), AOM_ICDF(32745), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(30198), AOM_ICDF(32558), AOM_ICDF(32742), AOM_ICDF(32756),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(29941), AOM_ICDF(32515), AOM_ICDF(32736), AOM_ICDF(32755),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(29686), AOM_ICDF(32470), AOM_ICDF(32731), AOM_ICDF(32755),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(29429), AOM_ICDF(32419), AOM_ICDF(32723), AOM_ICDF(32754),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(29174), AOM_ICDF(32366), AOM_ICDF(32715), AOM_ICDF(32753),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(28918), AOM_ICDF(32308), AOM_ICDF(32705), AOM_ICDF(32752),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(28662), AOM_ICDF(32246), AOM_ICDF(32694), AOM_ICDF(32750),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(28406), AOM_ICDF(32180), AOM_ICDF(32681), AOM_ICDF(32748),
+    AOM_ICDF(32757), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(28150), AOM_ICDF(32110), AOM_ICDF(32667), AOM_ICDF(32745),
+    AOM_ICDF(32756), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(27894), AOM_ICDF(32036), AOM_ICDF(32651), AOM_ICDF(32742),
+    AOM_ICDF(32756), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(27639), AOM_ICDF(31959), AOM_ICDF(32634), AOM_ICDF(32739),
+    AOM_ICDF(32755), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(27383), AOM_ICDF(31877), AOM_ICDF(32614), AOM_ICDF(32735),
+    AOM_ICDF(32755), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(27126), AOM_ICDF(31790), AOM_ICDF(32592), AOM_ICDF(32730),
+    AOM_ICDF(32754), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(26871), AOM_ICDF(31701), AOM_ICDF(32569), AOM_ICDF(32725),
+    AOM_ICDF(32753), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(26615), AOM_ICDF(31607), AOM_ICDF(32543), AOM_ICDF(32719),
+    AOM_ICDF(32752), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(26361), AOM_ICDF(31511), AOM_ICDF(32517), AOM_ICDF(32713),
+    AOM_ICDF(32751), AOM_ICDF(32758), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(26104), AOM_ICDF(31408), AOM_ICDF(32485), AOM_ICDF(32704),
+    AOM_ICDF(32748), AOM_ICDF(32757), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(25848), AOM_ICDF(31302), AOM_ICDF(32452), AOM_ICDF(32695),
+    AOM_ICDF(32746), AOM_ICDF(32757), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(25591), AOM_ICDF(31191), AOM_ICDF(32416), AOM_ICDF(32684),
+    AOM_ICDF(32743), AOM_ICDF(32756), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(25336), AOM_ICDF(31078), AOM_ICDF(32379), AOM_ICDF(32674),
+    AOM_ICDF(32741), AOM_ICDF(32756), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(25080), AOM_ICDF(30960), AOM_ICDF(32338), AOM_ICDF(32661),
+    AOM_ICDF(32737), AOM_ICDF(32755), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(24824), AOM_ICDF(30838), AOM_ICDF(32295), AOM_ICDF(32648),
+    AOM_ICDF(32733), AOM_ICDF(32754), AOM_ICDF(32759), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(24568), AOM_ICDF(30712), AOM_ICDF(32248), AOM_ICDF(32632),
+    AOM_ICDF(32728), AOM_ICDF(32752), AOM_ICDF(32758), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(24313), AOM_ICDF(30583), AOM_ICDF(32199), AOM_ICDF(32616),
+    AOM_ICDF(32723), AOM_ICDF(32751), AOM_ICDF(32758), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(24057), AOM_ICDF(30449), AOM_ICDF(32147), AOM_ICDF(32598),
+    AOM_ICDF(32718), AOM_ICDF(32750), AOM_ICDF(32758), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(23801), AOM_ICDF(30311), AOM_ICDF(32091), AOM_ICDF(32578),
+    AOM_ICDF(32711), AOM_ICDF(32747), AOM_ICDF(32757), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(23546), AOM_ICDF(30170), AOM_ICDF(32033), AOM_ICDF(32557),
+    AOM_ICDF(32704), AOM_ICDF(32745), AOM_ICDF(32757), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(23288), AOM_ICDF(30022), AOM_ICDF(31969), AOM_ICDF(32532),
+    AOM_ICDF(32695), AOM_ICDF(32742), AOM_ICDF(32756), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(23033), AOM_ICDF(29873), AOM_ICDF(31904), AOM_ICDF(32507),
+    AOM_ICDF(32686), AOM_ICDF(32739), AOM_ICDF(32755), AOM_ICDF(32760),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(22778), AOM_ICDF(29720), AOM_ICDF(31835), AOM_ICDF(32479),
+    AOM_ICDF(32675), AOM_ICDF(32735), AOM_ICDF(32753), AOM_ICDF(32759),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(22521), AOM_ICDF(29561), AOM_ICDF(31761), AOM_ICDF(32449),
+    AOM_ICDF(32664), AOM_ICDF(32731), AOM_ICDF(32752), AOM_ICDF(32759),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(22267), AOM_ICDF(29401), AOM_ICDF(31686), AOM_ICDF(32418),
+    AOM_ICDF(32652), AOM_ICDF(32727), AOM_ICDF(32751), AOM_ICDF(32759),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(22011), AOM_ICDF(29235), AOM_ICDF(31605), AOM_ICDF(32383),
+    AOM_ICDF(32638), AOM_ICDF(32722), AOM_ICDF(32749), AOM_ICDF(32758),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(21754), AOM_ICDF(29064), AOM_ICDF(31520), AOM_ICDF(32345),
+    AOM_ICDF(32622), AOM_ICDF(32715), AOM_ICDF(32746), AOM_ICDF(32757),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(21501), AOM_ICDF(28893), AOM_ICDF(31434), AOM_ICDF(32307),
+    AOM_ICDF(32607), AOM_ICDF(32710), AOM_ICDF(32745), AOM_ICDF(32757),
+    AOM_ICDF(32761), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(21243), AOM_ICDF(28713), AOM_ICDF(31339), AOM_ICDF(32262),
+    AOM_ICDF(32587), AOM_ICDF(32701), AOM_ICDF(32741), AOM_ICDF(32755),
+    AOM_ICDF(32760), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(20988), AOM_ICDF(28532), AOM_ICDF(31243), AOM_ICDF(32217),
+    AOM_ICDF(32567), AOM_ICDF(32693), AOM_ICDF(32738), AOM_ICDF(32754),
+    AOM_ICDF(32760), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(20730), AOM_ICDF(28344), AOM_ICDF(31140), AOM_ICDF(32167),
+    AOM_ICDF(32544), AOM_ICDF(32682), AOM_ICDF(32733), AOM_ICDF(32752),
+    AOM_ICDF(32759), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(20476), AOM_ICDF(28156), AOM_ICDF(31036), AOM_ICDF(32116),
+    AOM_ICDF(32521), AOM_ICDF(32673), AOM_ICDF(32730), AOM_ICDF(32751),
+    AOM_ICDF(32759), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(20220), AOM_ICDF(27962), AOM_ICDF(30926), AOM_ICDF(32061),
+    AOM_ICDF(32495), AOM_ICDF(32661), AOM_ICDF(32725), AOM_ICDF(32749),
+    AOM_ICDF(32758), AOM_ICDF(32762), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(19963), AOM_ICDF(27763), AOM_ICDF(30810), AOM_ICDF(32000),
+    AOM_ICDF(32465), AOM_ICDF(32647), AOM_ICDF(32718), AOM_ICDF(32746),
+    AOM_ICDF(32757), AOM_ICDF(32761), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(19708), AOM_ICDF(27562), AOM_ICDF(30691), AOM_ICDF(31938),
+    AOM_ICDF(32435), AOM_ICDF(32633), AOM_ICDF(32712), AOM_ICDF(32743),
+    AOM_ICDF(32756), AOM_ICDF(32761), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(19454), AOM_ICDF(27358), AOM_ICDF(30569), AOM_ICDF(31873),
+    AOM_ICDF(32403), AOM_ICDF(32618), AOM_ICDF(32705), AOM_ICDF(32741),
+    AOM_ICDF(32755), AOM_ICDF(32761), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(19196), AOM_ICDF(27146), AOM_ICDF(30438), AOM_ICDF(31801),
+    AOM_ICDF(32365), AOM_ICDF(32599), AOM_ICDF(32696), AOM_ICDF(32736),
+    AOM_ICDF(32753), AOM_ICDF(32760), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(18942), AOM_ICDF(26934), AOM_ICDF(30306), AOM_ICDF(31728),
+    AOM_ICDF(32328), AOM_ICDF(32581), AOM_ICDF(32688), AOM_ICDF(32733),
+    AOM_ICDF(32752), AOM_ICDF(32760), AOM_ICDF(32763), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(18684), AOM_ICDF(26714), AOM_ICDF(30164), AOM_ICDF(31647),
+    AOM_ICDF(32284), AOM_ICDF(32558), AOM_ICDF(32676), AOM_ICDF(32727),
+    AOM_ICDF(32749), AOM_ICDF(32758), AOM_ICDF(32762), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(18429), AOM_ICDF(26493), AOM_ICDF(30021), AOM_ICDF(31565),
+    AOM_ICDF(32240), AOM_ICDF(32535), AOM_ICDF(32664), AOM_ICDF(32721),
+    AOM_ICDF(32746), AOM_ICDF(32757), AOM_ICDF(32762), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(18174), AOM_ICDF(26268), AOM_ICDF(29872), AOM_ICDF(31477),
+    AOM_ICDF(32192), AOM_ICDF(32510), AOM_ICDF(32652), AOM_ICDF(32715),
+    AOM_ICDF(32743), AOM_ICDF(32756), AOM_ICDF(32762), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(17920), AOM_ICDF(26040), AOM_ICDF(29719), AOM_ICDF(31386),
+    AOM_ICDF(32141), AOM_ICDF(32483), AOM_ICDF(32638), AOM_ICDF(32708),
+    AOM_ICDF(32740), AOM_ICDF(32754), AOM_ICDF(32761), AOM_ICDF(32764),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(17661), AOM_ICDF(25803), AOM_ICDF(29556), AOM_ICDF(31286),
+    AOM_ICDF(32083), AOM_ICDF(32451), AOM_ICDF(32620), AOM_ICDF(32698),
+    AOM_ICDF(32734), AOM_ICDF(32751), AOM_ICDF(32759), AOM_ICDF(32763),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(17406), AOM_ICDF(25566), AOM_ICDF(29391), AOM_ICDF(31184),
+    AOM_ICDF(32024), AOM_ICDF(32418), AOM_ICDF(32603), AOM_ICDF(32690),
+    AOM_ICDF(32731), AOM_ICDF(32750), AOM_ICDF(32759), AOM_ICDF(32763),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(17151), AOM_ICDF(25325), AOM_ICDF(29220), AOM_ICDF(31076),
+    AOM_ICDF(31961), AOM_ICDF(32383), AOM_ICDF(32584), AOM_ICDF(32680),
+    AOM_ICDF(32726), AOM_ICDF(32748), AOM_ICDF(32758), AOM_ICDF(32763),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(16896), AOM_ICDF(25080), AOM_ICDF(29044), AOM_ICDF(30964),
+    AOM_ICDF(31894), AOM_ICDF(32344), AOM_ICDF(32562), AOM_ICDF(32668),
+    AOM_ICDF(32719), AOM_ICDF(32744), AOM_ICDF(32756), AOM_ICDF(32762),
+    AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(16639), AOM_ICDF(24829), AOM_ICDF(28860), AOM_ICDF(30844),
+    AOM_ICDF(31821), AOM_ICDF(32302), AOM_ICDF(32539), AOM_ICDF(32655),
+    AOM_ICDF(32712), AOM_ICDF(32740), AOM_ICDF(32754), AOM_ICDF(32761),
+    AOM_ICDF(32764), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(16384), AOM_ICDF(24576), AOM_ICDF(28672), AOM_ICDF(30720),
+    AOM_ICDF(31744), AOM_ICDF(32256), AOM_ICDF(32512), AOM_ICDF(32640),
+    AOM_ICDF(32704), AOM_ICDF(32736), AOM_ICDF(32752), AOM_ICDF(32760),
+    AOM_ICDF(32764), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(16130), AOM_ICDF(24320), AOM_ICDF(28479), AOM_ICDF(30591),
+    AOM_ICDF(31663), AOM_ICDF(32208), AOM_ICDF(32485), AOM_ICDF(32625),
+    AOM_ICDF(32696), AOM_ICDF(32732), AOM_ICDF(32750), AOM_ICDF(32759),
+    AOM_ICDF(32764), AOM_ICDF(32766), AOM_ICDF(32767), AOM_ICDF(32768)},
+  {AOM_ICDF(15872), AOM_ICDF(24056), AOM_ICDF(28276), AOM_ICDF(30452),
+    AOM_ICDF(31574), AOM_ICDF(32152), AOM_ICDF(32450), AOM_ICDF(32604),
+    AOM_ICDF(32683), AOM_ICDF(32724), AOM_ICDF(32745), AOM_ICDF(32756),
+    AOM_ICDF(32762), AOM_ICDF(32765), AOM_ICDF(32766), AOM_ICDF(32768)},
+  {AOM_ICDF(15615), AOM_ICDF(23789), AOM_ICDF(28068), AOM_ICDF(30308),
+    AOM_ICDF(31480), AOM_ICDF(32094), AOM_ICDF(32415), AOM_ICDF(32583),
+    AOM_ICDF(32671), AOM_ICDF(32717), AOM_ICDF(32741), AOM_ICDF(32754),
+    AOM_ICDF(32761), AOM_ICDF(32764), AOM_ICDF(32766), AOM_ICDF(32768)},
+  {AOM_ICDF(15361), AOM_ICDF(23521), AOM_ICDF(27856), AOM_ICDF(30159),
+    AOM_ICDF(31382), AOM_ICDF(32032), AOM_ICDF(32377), AOM_ICDF(32560),
+    AOM_ICDF(32657), AOM_ICDF(32709), AOM_ICDF(32737), AOM_ICDF(32752),
+    AOM_ICDF(32760), AOM_ICDF(32764), AOM_ICDF(32766), AOM_ICDF(32768)},
+  {AOM_ICDF(15103), AOM_ICDF(23245), AOM_ICDF(27634), AOM_ICDF(30000),
+    AOM_ICDF(31275), AOM_ICDF(31963), AOM_ICDF(32334), AOM_ICDF(32534),
+    AOM_ICDF(32642), AOM_ICDF(32700), AOM_ICDF(32731), AOM_ICDF(32748),
+    AOM_ICDF(32757), AOM_ICDF(32762), AOM_ICDF(32765), AOM_ICDF(32768)},
+  {AOM_ICDF(14848), AOM_ICDF(22968), AOM_ICDF(27409), AOM_ICDF(29837),
+    AOM_ICDF(31165), AOM_ICDF(31891), AOM_ICDF(32288), AOM_ICDF(32505),
+    AOM_ICDF(32624), AOM_ICDF(32689), AOM_ICDF(32725), AOM_ICDF(32744),
+    AOM_ICDF(32755), AOM_ICDF(32761), AOM_ICDF(32764), AOM_ICDF(32768)},
+  {AOM_ICDF(14592), AOM_ICDF(22686), AOM_ICDF(27176), AOM_ICDF(29666),
+    AOM_ICDF(31047), AOM_ICDF(31813), AOM_ICDF(32238), AOM_ICDF(32474),
+    AOM_ICDF(32605), AOM_ICDF(32678), AOM_ICDF(32718), AOM_ICDF(32740),
+    AOM_ICDF(32752), AOM_ICDF(32759), AOM_ICDF(32763), AOM_ICDF(32768)},
+  {AOM_ICDF(14336), AOM_ICDF(22400), AOM_ICDF(26936), AOM_ICDF(29488),
+    AOM_ICDF(30923), AOM_ICDF(31730), AOM_ICDF(32184), AOM_ICDF(32439),
+    AOM_ICDF(32583), AOM_ICDF(32664), AOM_ICDF(32709), AOM_ICDF(32735),
+    AOM_ICDF(32749), AOM_ICDF(32757), AOM_ICDF(32762), AOM_ICDF(32768)},
+  {AOM_ICDF(14079), AOM_ICDF(22109), AOM_ICDF(26689), AOM_ICDF(29301),
+    AOM_ICDF(30791), AOM_ICDF(31641), AOM_ICDF(32125), AOM_ICDF(32401),
+    AOM_ICDF(32559), AOM_ICDF(32649), AOM_ICDF(32700), AOM_ICDF(32729),
+    AOM_ICDF(32746), AOM_ICDF(32756), AOM_ICDF(32761), AOM_ICDF(32768)},
+  {AOM_ICDF(13825), AOM_ICDF(21817), AOM_ICDF(26437), AOM_ICDF(29108),
+    AOM_ICDF(30652), AOM_ICDF(31545), AOM_ICDF(32061), AOM_ICDF(32359),
+    AOM_ICDF(32532), AOM_ICDF(32632), AOM_ICDF(32690), AOM_ICDF(32723),
+    AOM_ICDF(32742), AOM_ICDF(32753), AOM_ICDF(32759), AOM_ICDF(32768)},
+  {AOM_ICDF(13568), AOM_ICDF(21518), AOM_ICDF(26176), AOM_ICDF(28905),
+    AOM_ICDF(30504), AOM_ICDF(31441), AOM_ICDF(31990), AOM_ICDF(32312),
+    AOM_ICDF(32501), AOM_ICDF(32611), AOM_ICDF(32676), AOM_ICDF(32714),
+    AOM_ICDF(32736), AOM_ICDF(32749), AOM_ICDF(32757), AOM_ICDF(32768)},
+  {AOM_ICDF(13314), AOM_ICDF(21218), AOM_ICDF(25911), AOM_ICDF(28697),
+    AOM_ICDF(30351), AOM_ICDF(31333), AOM_ICDF(31916), AOM_ICDF(32262),
+    AOM_ICDF(32468), AOM_ICDF(32590), AOM_ICDF(32662), AOM_ICDF(32705),
+    AOM_ICDF(32731), AOM_ICDF(32746), AOM_ICDF(32755), AOM_ICDF(32768)},
+  {AOM_ICDF(13054), AOM_ICDF(20908), AOM_ICDF(25633), AOM_ICDF(28475),
+    AOM_ICDF(30185), AOM_ICDF(31214), AOM_ICDF(31833), AOM_ICDF(32205),
+    AOM_ICDF(32429), AOM_ICDF(32564), AOM_ICDF(32645), AOM_ICDF(32694),
+    AOM_ICDF(32723), AOM_ICDF(32741), AOM_ICDF(32752), AOM_ICDF(32768)},
+  {AOM_ICDF(12803), AOM_ICDF(20603), AOM_ICDF(25356), AOM_ICDF(28252),
+    AOM_ICDF(30017), AOM_ICDF(31093), AOM_ICDF(31748), AOM_ICDF(32147),
+    AOM_ICDF(32390), AOM_ICDF(32538), AOM_ICDF(32628), AOM_ICDF(32683),
+    AOM_ICDF(32717), AOM_ICDF(32737), AOM_ICDF(32749), AOM_ICDF(32768)},
+  {AOM_ICDF(12544), AOM_ICDF(20286), AOM_ICDF(25064), AOM_ICDF(28013),
+    AOM_ICDF(29833), AOM_ICDF(30956), AOM_ICDF(31649), AOM_ICDF(32077),
+    AOM_ICDF(32341), AOM_ICDF(32504), AOM_ICDF(32605), AOM_ICDF(32667),
+    AOM_ICDF(32705), AOM_ICDF(32729), AOM_ICDF(32744), AOM_ICDF(32768)},
+  {AOM_ICDF(12288), AOM_ICDF(19968), AOM_ICDF(24768), AOM_ICDF(27768),
+    AOM_ICDF(29643), AOM_ICDF(30815), AOM_ICDF(31547), AOM_ICDF(32005),
+    AOM_ICDF(32291), AOM_ICDF(32470), AOM_ICDF(32582), AOM_ICDF(32652),
+    AOM_ICDF(32696), AOM_ICDF(32723), AOM_ICDF(32740), AOM_ICDF(32768)},
+  {AOM_ICDF(12033), AOM_ICDF(19647), AOM_ICDF(24465), AOM_ICDF(27514),
+    AOM_ICDF(29443), AOM_ICDF(30664), AOM_ICDF(31437), AOM_ICDF(31926),
+    AOM_ICDF(32235), AOM_ICDF(32431), AOM_ICDF(32555), AOM_ICDF(32633),
+    AOM_ICDF(32683), AOM_ICDF(32714), AOM_ICDF(32734), AOM_ICDF(32768)},
+  {AOM_ICDF(11777), AOM_ICDF(19321), AOM_ICDF(24154), AOM_ICDF(27250),
+    AOM_ICDF(29233), AOM_ICDF(30504), AOM_ICDF(31318), AOM_ICDF(31839),
+    AOM_ICDF(32173), AOM_ICDF(32387), AOM_ICDF(32524), AOM_ICDF(32612),
+    AOM_ICDF(32668), AOM_ICDF(32704), AOM_ICDF(32727), AOM_ICDF(32768)},
+  {AOM_ICDF(11521), AOM_ICDF(18991), AOM_ICDF(23835), AOM_ICDF(26976),
+    AOM_ICDF(29013), AOM_ICDF(30334), AOM_ICDF(31190), AOM_ICDF(31745),
+    AOM_ICDF(32105), AOM_ICDF(32338), AOM_ICDF(32489), AOM_ICDF(32587),
+    AOM_ICDF(32651), AOM_ICDF(32692), AOM_ICDF(32719), AOM_ICDF(32768)},
+  {AOM_ICDF(11265), AOM_ICDF(18657), AOM_ICDF(23508), AOM_ICDF(26691),
+    AOM_ICDF(28780), AOM_ICDF(30151), AOM_ICDF(31051), AOM_ICDF(31641),
+    AOM_ICDF(32028), AOM_ICDF(32282), AOM_ICDF(32449), AOM_ICDF(32559),
+    AOM_ICDF(32631), AOM_ICDF(32678), AOM_ICDF(32709), AOM_ICDF(32768)},
+  {AOM_ICDF(11006), AOM_ICDF(18316), AOM_ICDF(23170), AOM_ICDF(26394),
+    AOM_ICDF(28535), AOM_ICDF(29957), AOM_ICDF(30901), AOM_ICDF(31528),
+    AOM_ICDF(31944), AOM_ICDF(32220), AOM_ICDF(32404), AOM_ICDF(32526),
+    AOM_ICDF(32607), AOM_ICDF(32661), AOM_ICDF(32697), AOM_ICDF(32768)},
+  {AOM_ICDF(10752), AOM_ICDF(17976), AOM_ICDF(22830), AOM_ICDF(26091),
+    AOM_ICDF(28282), AOM_ICDF(29754), AOM_ICDF(30743), AOM_ICDF(31408),
+    AOM_ICDF(31854), AOM_ICDF(32154), AOM_ICDF(32356), AOM_ICDF(32491),
+    AOM_ICDF(32582), AOM_ICDF(32643), AOM_ICDF(32684), AOM_ICDF(32768)},
+  {AOM_ICDF(10496), AOM_ICDF(17630), AOM_ICDF(22479), AOM_ICDF(25775),
+    AOM_ICDF(28015), AOM_ICDF(29538), AOM_ICDF(30573), AOM_ICDF(31276),
+    AOM_ICDF(31754), AOM_ICDF(32079), AOM_ICDF(32300), AOM_ICDF(32450),
+    AOM_ICDF(32552), AOM_ICDF(32621), AOM_ICDF(32668), AOM_ICDF(32768)},
+  {AOM_ICDF(10240), AOM_ICDF(17280), AOM_ICDF(22120), AOM_ICDF(25448),
+    AOM_ICDF(27736), AOM_ICDF(29309), AOM_ICDF(30390), AOM_ICDF(31133),
+    AOM_ICDF(31644), AOM_ICDF(31995), AOM_ICDF(32237), AOM_ICDF(32403),
+    AOM_ICDF(32517), AOM_ICDF(32595), AOM_ICDF(32649), AOM_ICDF(32768)},
+  { AOM_ICDF(9984), AOM_ICDF(16926), AOM_ICDF(21753), AOM_ICDF(25109),
+    AOM_ICDF(27443), AOM_ICDF(29066), AOM_ICDF(30194), AOM_ICDF(30978),
+    AOM_ICDF(31523), AOM_ICDF(31902), AOM_ICDF(32166), AOM_ICDF(32349),
+    AOM_ICDF(32476), AOM_ICDF(32565), AOM_ICDF(32627), AOM_ICDF(32768)},
+  { AOM_ICDF(9728), AOM_ICDF(16568), AOM_ICDF(21377), AOM_ICDF(24759),
+    AOM_ICDF(27137), AOM_ICDF(28809), AOM_ICDF(29984), AOM_ICDF(30811),
+    AOM_ICDF(31392), AOM_ICDF(31801), AOM_ICDF(32088), AOM_ICDF(32290),
+    AOM_ICDF(32432), AOM_ICDF(32532), AOM_ICDF(32602), AOM_ICDF(32768)},
+  { AOM_ICDF(9474), AOM_ICDF(16208), AOM_ICDF(20995), AOM_ICDF(24399),
+    AOM_ICDF(26819), AOM_ICDF(28539), AOM_ICDF(29762), AOM_ICDF(30631),
+    AOM_ICDF(31249), AOM_ICDF(31688), AOM_ICDF(32000), AOM_ICDF(32222),
+    AOM_ICDF(32380), AOM_ICDF(32492), AOM_ICDF(32572), AOM_ICDF(32768)},
+  { AOM_ICDF(9216), AOM_ICDF(15840), AOM_ICDF(20601), AOM_ICDF(24023),
+    AOM_ICDF(26483), AOM_ICDF(28251), AOM_ICDF(29522), AOM_ICDF(30435),
+    AOM_ICDF(31091), AOM_ICDF(31563), AOM_ICDF(31902), AOM_ICDF(32146),
+    AOM_ICDF(32321), AOM_ICDF(32447), AOM_ICDF(32537), AOM_ICDF(32768)},
+  { AOM_ICDF(8959), AOM_ICDF(15469), AOM_ICDF(20199), AOM_ICDF(23636),
+    AOM_ICDF(26133), AOM_ICDF(27947), AOM_ICDF(29265), AOM_ICDF(30223),
+    AOM_ICDF(30919), AOM_ICDF(31425), AOM_ICDF(31792), AOM_ICDF(32059),
+    AOM_ICDF(32253), AOM_ICDF(32394), AOM_ICDF(32496), AOM_ICDF(32768)},
+  { AOM_ICDF(8705), AOM_ICDF(15097), AOM_ICDF(19791), AOM_ICDF(23238),
+    AOM_ICDF(25770), AOM_ICDF(27629), AOM_ICDF(28994), AOM_ICDF(29997),
+    AOM_ICDF(30733), AOM_ICDF(31274), AOM_ICDF(31671), AOM_ICDF(31963),
+    AOM_ICDF(32177), AOM_ICDF(32334), AOM_ICDF(32449), AOM_ICDF(32768)},
+  { AOM_ICDF(8449), AOM_ICDF(14719), AOM_ICDF(19373), AOM_ICDF(22827),
+    AOM_ICDF(25390), AOM_ICDF(27292), AOM_ICDF(28704), AOM_ICDF(29752),
+    AOM_ICDF(30530), AOM_ICDF(31107), AOM_ICDF(31535), AOM_ICDF(31853),
+    AOM_ICDF(32089), AOM_ICDF(32264), AOM_ICDF(32394), AOM_ICDF(32768)},
+  { AOM_ICDF(8192), AOM_ICDF(14336), AOM_ICDF(18944), AOM_ICDF(22400),
+    AOM_ICDF(24992), AOM_ICDF(26936), AOM_ICDF(28394), AOM_ICDF(29488),
+    AOM_ICDF(30308), AOM_ICDF(30923), AOM_ICDF(31384), AOM_ICDF(31730),
+    AOM_ICDF(31989), AOM_ICDF(32184), AOM_ICDF(32330), AOM_ICDF(32768)},
+  { AOM_ICDF(7936), AOM_ICDF(13950), AOM_ICDF(18507), AOM_ICDF(21961),
+    AOM_ICDF(24578), AOM_ICDF(26561), AOM_ICDF(28064), AOM_ICDF(29203),
+    AOM_ICDF(30066), AOM_ICDF(30720), AOM_ICDF(31216), AOM_ICDF(31592),
+    AOM_ICDF(31877), AOM_ICDF(32093), AOM_ICDF(32256), AOM_ICDF(32768)},
+  { AOM_ICDF(7678), AOM_ICDF(13558), AOM_ICDF(18060), AOM_ICDF(21507),
+    AOM_ICDF(24146), AOM_ICDF(26166), AOM_ICDF(27713), AOM_ICDF(28897),
+    AOM_ICDF(29804), AOM_ICDF(30498), AOM_ICDF(31030), AOM_ICDF(31437),
+    AOM_ICDF(31749), AOM_ICDF(31988), AOM_ICDF(32171), AOM_ICDF(32768)},
+  { AOM_ICDF(7423), AOM_ICDF(13165), AOM_ICDF(17606), AOM_ICDF(21041),
+    AOM_ICDF(23698), AOM_ICDF(25753), AOM_ICDF(27342), AOM_ICDF(28571),
+    AOM_ICDF(29522), AOM_ICDF(30257), AOM_ICDF(30826), AOM_ICDF(31266),
+    AOM_ICDF(31606), AOM_ICDF(31869), AOM_ICDF(32073), AOM_ICDF(32768)},
+  { AOM_ICDF(7168), AOM_ICDF(12768), AOM_ICDF(17143), AOM_ICDF(20561),
+    AOM_ICDF(23231), AOM_ICDF(25317), AOM_ICDF(26947), AOM_ICDF(28220),
+    AOM_ICDF(29215), AOM_ICDF(29992), AOM_ICDF(30599), AOM_ICDF(31073),
+    AOM_ICDF(31444), AOM_ICDF(31734), AOM_ICDF(31960), AOM_ICDF(32768)},
+  { AOM_ICDF(6911), AOM_ICDF(12365), AOM_ICDF(16669), AOM_ICDF(20065),
+    AOM_ICDF(22744), AOM_ICDF(24858), AOM_ICDF(26526), AOM_ICDF(27842),
+    AOM_ICDF(28881), AOM_ICDF(29701), AOM_ICDF(30348), AOM_ICDF(30858),
+    AOM_ICDF(31261), AOM_ICDF(31579), AOM_ICDF(31830), AOM_ICDF(32768)},
+  { AOM_ICDF(6657), AOM_ICDF(11961), AOM_ICDF(16188), AOM_ICDF(19556),
+    AOM_ICDF(22240), AOM_ICDF(24379), AOM_ICDF(26083), AOM_ICDF(27441),
+    AOM_ICDF(28523), AOM_ICDF(29385), AOM_ICDF(30072), AOM_ICDF(30620),
+    AOM_ICDF(31056), AOM_ICDF(31404), AOM_ICDF(31681), AOM_ICDF(32768)},
+  { AOM_ICDF(6400), AOM_ICDF(11550), AOM_ICDF(15694), AOM_ICDF(19029),
+    AOM_ICDF(21712), AOM_ICDF(23871), AOM_ICDF(25609), AOM_ICDF(27007),
+    AOM_ICDF(28132), AOM_ICDF(29037), AOM_ICDF(29766), AOM_ICDF(30352),
+    AOM_ICDF(30824), AOM_ICDF(31204), AOM_ICDF(31509), AOM_ICDF(32768)},
+  { AOM_ICDF(6142), AOM_ICDF(11134), AOM_ICDF(15190), AOM_ICDF(18486),
+    AOM_ICDF(21164), AOM_ICDF(23340), AOM_ICDF(25108), AOM_ICDF(26544),
+    AOM_ICDF(27711), AOM_ICDF(28659), AOM_ICDF(29429), AOM_ICDF(30055),
+    AOM_ICDF(30564), AOM_ICDF(30977), AOM_ICDF(31313), AOM_ICDF(32768)},
+  { AOM_ICDF(5890), AOM_ICDF(10720), AOM_ICDF(14682), AOM_ICDF(17932),
+    AOM_ICDF(20598), AOM_ICDF(22785), AOM_ICDF(24579), AOM_ICDF(26051),
+    AOM_ICDF(27258), AOM_ICDF(28248), AOM_ICDF(29060), AOM_ICDF(29726),
+    AOM_ICDF(30273), AOM_ICDF(30721), AOM_ICDF(31089), AOM_ICDF(32768)},
+  { AOM_ICDF(5631), AOM_ICDF(10295), AOM_ICDF(14157), AOM_ICDF(17356),
+    AOM_ICDF(20005), AOM_ICDF(22199), AOM_ICDF(24016), AOM_ICDF(25520),
+    AOM_ICDF(26766), AOM_ICDF(27798), AOM_ICDF(28652), AOM_ICDF(29359),
+    AOM_ICDF(29945), AOM_ICDF(30430), AOM_ICDF(30832), AOM_ICDF(32768)},
+  { AOM_ICDF(5377), AOM_ICDF(9871), AOM_ICDF(13628), AOM_ICDF(16768),
+    AOM_ICDF(19393), AOM_ICDF(21587), AOM_ICDF(23421), AOM_ICDF(24954),
+    AOM_ICDF(26236), AOM_ICDF(27308), AOM_ICDF(28204), AOM_ICDF(28953),
+    AOM_ICDF(29579), AOM_ICDF(30102), AOM_ICDF(30539), AOM_ICDF(32768)},
+  { AOM_ICDF(5121), AOM_ICDF(9441), AOM_ICDF(13086), AOM_ICDF(16161),
+    AOM_ICDF(18756), AOM_ICDF(20945), AOM_ICDF(22792), AOM_ICDF(24351),
+    AOM_ICDF(25666), AOM_ICDF(26776), AOM_ICDF(27712), AOM_ICDF(28502),
+    AOM_ICDF(29169), AOM_ICDF(29731), AOM_ICDF(30206), AOM_ICDF(32768)},
+  { AOM_ICDF(4865), AOM_ICDF(9007), AOM_ICDF(12534), AOM_ICDF(15538),
+    AOM_ICDF(18096), AOM_ICDF(20274), AOM_ICDF(22129), AOM_ICDF(23708),
+    AOM_ICDF(25053), AOM_ICDF(26198), AOM_ICDF(27173), AOM_ICDF(28004),
+    AOM_ICDF(28711), AOM_ICDF(29313), AOM_ICDF(29826), AOM_ICDF(32768)},
+  { AOM_ICDF(4608), AOM_ICDF(8568), AOM_ICDF(11971), AOM_ICDF(14896),
+    AOM_ICDF(17409), AOM_ICDF(19569), AOM_ICDF(21425), AOM_ICDF(23020),
+    AOM_ICDF(24391), AOM_ICDF(25569), AOM_ICDF(26581), AOM_ICDF(27451),
+    AOM_ICDF(28199), AOM_ICDF(28842), AOM_ICDF(29394), AOM_ICDF(32768)},
+  { AOM_ICDF(4351), AOM_ICDF(8125), AOM_ICDF(11398), AOM_ICDF(14236),
+    AOM_ICDF(16697), AOM_ICDF(18831), AOM_ICDF(20682), AOM_ICDF(22287),
+    AOM_ICDF(23679), AOM_ICDF(24886), AOM_ICDF(25933), AOM_ICDF(26841),
+    AOM_ICDF(27628), AOM_ICDF(28311), AOM_ICDF(28903), AOM_ICDF(32768)},
+  { AOM_ICDF(4096), AOM_ICDF(7680), AOM_ICDF(10816), AOM_ICDF(13560),
+    AOM_ICDF(15961), AOM_ICDF(18062), AOM_ICDF(19900), AOM_ICDF(21508),
+    AOM_ICDF(22915), AOM_ICDF(24146), AOM_ICDF(25224), AOM_ICDF(26167),
+    AOM_ICDF(26992), AOM_ICDF(27714), AOM_ICDF(28346), AOM_ICDF(32768)},
+  { AOM_ICDF(3840), AOM_ICDF(7230), AOM_ICDF(10223), AOM_ICDF(12865),
+    AOM_ICDF(15197), AOM_ICDF(17256), AOM_ICDF(19074), AOM_ICDF(20679),
+    AOM_ICDF(22096), AOM_ICDF(23347), AOM_ICDF(24451), AOM_ICDF(25426),
+    AOM_ICDF(26287), AOM_ICDF(27047), AOM_ICDF(27718), AOM_ICDF(32768)},
+  { AOM_ICDF(3584), AOM_ICDF(6776), AOM_ICDF(9619), AOM_ICDF(12151),
+    AOM_ICDF(14406), AOM_ICDF(16414), AOM_ICDF(18203), AOM_ICDF(19796),
+    AOM_ICDF(21215), AOM_ICDF(22479), AOM_ICDF(23604), AOM_ICDF(24606),
+    AOM_ICDF(25499), AOM_ICDF(26294), AOM_ICDF(27002), AOM_ICDF(32768)},
+  { AOM_ICDF(3328), AOM_ICDF(6318), AOM_ICDF(9004), AOM_ICDF(11417),
+    AOM_ICDF(13585), AOM_ICDF(15533), AOM_ICDF(17283), AOM_ICDF(18856),
+    AOM_ICDF(20269), AOM_ICDF(21538), AOM_ICDF(22678), AOM_ICDF(23703),
+    AOM_ICDF(24624), AOM_ICDF(25451), AOM_ICDF(26194), AOM_ICDF(32768)},
+  { AOM_ICDF(3072), AOM_ICDF(5856), AOM_ICDF(8379), AOM_ICDF(10665),
+    AOM_ICDF(12737), AOM_ICDF(14615), AOM_ICDF(16317), AOM_ICDF(17859),
+    AOM_ICDF(19257), AOM_ICDF(20524), AOM_ICDF(21672), AOM_ICDF(22712),
+    AOM_ICDF(23655), AOM_ICDF(24509), AOM_ICDF(25283), AOM_ICDF(32768)},
+  { AOM_ICDF(2816), AOM_ICDF(5390), AOM_ICDF(7743), AOM_ICDF(9894),
+    AOM_ICDF(11860), AOM_ICDF(13657), AOM_ICDF(15299), AOM_ICDF(16800),
+    AOM_ICDF(18172), AOM_ICDF(19426), AOM_ICDF(20573), AOM_ICDF(21621),
+    AOM_ICDF(22579), AOM_ICDF(23455), AOM_ICDF(24255), AOM_ICDF(32768)},
+  { AOM_ICDF(2560), AOM_ICDF(4920), AOM_ICDF(7096), AOM_ICDF(9102),
+    AOM_ICDF(10951), AOM_ICDF(12656), AOM_ICDF(14227), AOM_ICDF(15676),
+    AOM_ICDF(17011), AOM_ICDF(18242), AOM_ICDF(19377), AOM_ICDF(20423),
+    AOM_ICDF(21388), AOM_ICDF(22277), AOM_ICDF(23097), AOM_ICDF(32768)},
+  { AOM_ICDF(2304), AOM_ICDF(4446), AOM_ICDF(6437), AOM_ICDF(8288),
+    AOM_ICDF(10009), AOM_ICDF(11609), AOM_ICDF(13097), AOM_ICDF(14480),
+    AOM_ICDF(15766), AOM_ICDF(16961), AOM_ICDF(18072), AOM_ICDF(19105),
+    AOM_ICDF(20066), AOM_ICDF(20959), AOM_ICDF(21789), AOM_ICDF(32768)},
+  { AOM_ICDF(2048), AOM_ICDF(3968), AOM_ICDF(5768), AOM_ICDF(7456),
+    AOM_ICDF(9038), AOM_ICDF(10521), AOM_ICDF(11911), AOM_ICDF(13215),
+    AOM_ICDF(14437), AOM_ICDF(15583), AOM_ICDF(16657), AOM_ICDF(17664),
+    AOM_ICDF(18608), AOM_ICDF(19493), AOM_ICDF(20323), AOM_ICDF(32768)},
+  { AOM_ICDF(1792), AOM_ICDF(3486), AOM_ICDF(5087), AOM_ICDF(6601),
+    AOM_ICDF(8032), AOM_ICDF(9385), AOM_ICDF(10664), AOM_ICDF(11873),
+    AOM_ICDF(13016), AOM_ICDF(14096), AOM_ICDF(15117), AOM_ICDF(16082),
+    AOM_ICDF(16995), AOM_ICDF(17858), AOM_ICDF(18673), AOM_ICDF(32768)},
+  { AOM_ICDF(1536), AOM_ICDF(3000), AOM_ICDF(4395), AOM_ICDF(5725),
+    AOM_ICDF(6993), AOM_ICDF(8201), AOM_ICDF(9353), AOM_ICDF(10451),
+    AOM_ICDF(11497), AOM_ICDF(12494), AOM_ICDF(13444), AOM_ICDF(14350),
+    AOM_ICDF(15213), AOM_ICDF(16036), AOM_ICDF(16820), AOM_ICDF(32768)},
+  { AOM_ICDF(1280), AOM_ICDF(2510), AOM_ICDF(3692), AOM_ICDF(4828),
+    AOM_ICDF(5919), AOM_ICDF(6968), AOM_ICDF(7976), AOM_ICDF(8944),
+    AOM_ICDF(9875), AOM_ICDF(10769), AOM_ICDF(11628), AOM_ICDF(12454),
+    AOM_ICDF(13248), AOM_ICDF(14011), AOM_ICDF(14744), AOM_ICDF(32768)},
+  { AOM_ICDF(1024), AOM_ICDF(2016), AOM_ICDF(2977), AOM_ICDF(3908),
+    AOM_ICDF(4810), AOM_ICDF(5684), AOM_ICDF(6530), AOM_ICDF(7350),
+    AOM_ICDF(8144), AOM_ICDF(8913), AOM_ICDF(9658), AOM_ICDF(10380),
+    AOM_ICDF(11080), AOM_ICDF(11758), AOM_ICDF(12415), AOM_ICDF(32768)},
+  {  AOM_ICDF(768), AOM_ICDF(1518), AOM_ICDF(2250), AOM_ICDF(2965),
+    AOM_ICDF(3663), AOM_ICDF(4345), AOM_ICDF(5011), AOM_ICDF(5662),
+    AOM_ICDF(6297), AOM_ICDF(6917), AOM_ICDF(7523), AOM_ICDF(8115),
+    AOM_ICDF(8693), AOM_ICDF(9257), AOM_ICDF(9808), AOM_ICDF(32768)},
+  {  AOM_ICDF(512), AOM_ICDF(1016), AOM_ICDF(1512), AOM_ICDF(2000),
+    AOM_ICDF(2481), AOM_ICDF(2954), AOM_ICDF(3420), AOM_ICDF(3879),
+    AOM_ICDF(4330), AOM_ICDF(4774), AOM_ICDF(5211), AOM_ICDF(5642),
+    AOM_ICDF(6066), AOM_ICDF(6483), AOM_ICDF(6894), AOM_ICDF(32768)},
+  {  AOM_ICDF(256),  AOM_ICDF(510),  AOM_ICDF(762), AOM_ICDF(1012),
+    AOM_ICDF(1260), AOM_ICDF(1506), AOM_ICDF(1750), AOM_ICDF(1992),
+    AOM_ICDF(2232), AOM_ICDF(2471), AOM_ICDF(2708), AOM_ICDF(2943),
+    AOM_ICDF(3176), AOM_ICDF(3407), AOM_ICDF(3636), AOM_ICDF(32768)},
+};
+
+
+const uint16_t LAPLACE_OFFSET[128] = {
+  0,
+  29871,
+  28672,
+  27751,
+  26975,
+  26291,
+  25673,
+  25105,
+  24576,
+  24079,
+  23609,
+  23162,
+  22734,
+  22325,
+  21931,
+  21550,
+  21182,
+  20826,
+  20480,
+  20143,
+  19815,
+  19495,
+  19183,
+  18877,
+  18579,
+  18286,
+  17999,
+  17718,
+  17442,
+  17170,
+  16904,
+  16642,
+  16384,
+  16129,
+  15879,
+  15633,
+  15390,
+  15150,
+  14913,
+  14680,
+  14450,
+  14222,
+  13997,
+  13775,
+  13556,
+  13338,
+  13124,
+  12911,
+  12701,
+  12493,
+  12288,
+  12084,
+  11882,
+  11682,
+  11484,
+  11288,
+  11094,
+  10901,
+  10710,
+  10521,
+  10333,
+  10147,
+  9962,
+  9779,
+  9597,
+  9417,
+  9238,
+  9060,
+  8884,
+  8709,
+  8535,
+  8363,
+  8192,
+  8021,
+  7853,
+  7685,
+  7518,
+  7352,
+  7188,
+  7025,
+  6862,
+  6701,
+  6540,
+  6381,
+  6222,
+  6065,
+  5908,
+  5753,
+  5598,
+  5444,
+  5291,
+  5138,
+  4987,
+  4837,
+  4687,
+  4538,
+  4390,
+  4242,
+  4096,
+  3950,
+  3804,
+  3660,
+  3516,
+  3373,
+  3231,
+  3089,
+  2948,
+  2808,
+  2668,
+  2529,
+  2391,
+  2253,
+  2116,
+  1979,
+  1843,
+  1708,
+  1573,
+  1439,
+  1306,
+  1172,
+  1040,
+  908,
+  777,
+  646,
+  516,
+  386,
+  257,
+  128,
+};
diff --git a/third_party/aom/av1/common/mips/dspr2/av1_itrans16_dspr2.c b/third_party/aom/av1/common/mips/dspr2/av1_itrans16_dspr2.c
new file mode 100644
index 0000000000..79f9338bd1
--- /dev/null
+++ b/third_party/aom/av1/common/mips/dspr2/av1_itrans16_dspr2.c
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "av1/common/common.h"
+#include "av1/common/blockd.h"
+#include "av1/common/idct.h"
+#include "aom_dsp/mips/inv_txfm_dspr2.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+void av1_iht16x16_256_add_dspr2(const int16_t *input, uint8_t *dest, int pitch,
+                                int tx_type) {
+  int i, j;
+  DECLARE_ALIGNED(32, int16_t, out[16 * 16]);
+  int16_t *outptr = out;
+  int16_t temp_out[16];
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp    %[pos],    1    \n\t" : : [pos] "r"(pos));
+
+  switch (tx_type) {
+    case DCT_DCT:  // DCT in both horizontal and vertical
+      idct16_rows_dspr2(input, outptr, 16);
+      idct16_cols_add_blk_dspr2(out, dest, pitch);
+      break;
+    case ADST_DCT:  // ADST in vertical, DCT in horizontal
+      idct16_rows_dspr2(input, outptr, 16);
+
+      outptr = out;
+
+      for (i = 0; i < 16; ++i) {
+        iadst16_dspr2(outptr, temp_out);
+
+        for (j = 0; j < 16; ++j)
+          dest[j * pitch + i] = clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 6) +
+                                           dest[j * pitch + i]);
+        outptr += 16;
+      }
+      break;
+    case DCT_ADST:  // DCT in vertical, ADST in horizontal
+    {
+      int16_t temp_in[16 * 16];
+
+      for (i = 0; i < 16; ++i) {
+        /* prefetch row */
+        prefetch_load((const uint8_t *)(input + 16));
+
+        iadst16_dspr2(input, outptr);
+        input += 16;
+        outptr += 16;
+      }
+
+      for (i = 0; i < 16; ++i)
+        for (j = 0; j < 16; ++j) temp_in[j * 16 + i] = out[i * 16 + j];
+
+      idct16_cols_add_blk_dspr2(temp_in, dest, pitch);
+    } break;
+    case ADST_ADST:  // ADST in both directions
+    {
+      int16_t temp_in[16];
+
+      for (i = 0; i < 16; ++i) {
+        /* prefetch row */
+        prefetch_load((const uint8_t *)(input + 16));
+
+        iadst16_dspr2(input, outptr);
+        input += 16;
+        outptr += 16;
+      }
+
+      for (i = 0; i < 16; ++i) {
+        for (j = 0; j < 16; ++j) temp_in[j] = out[j * 16 + i];
+        iadst16_dspr2(temp_in, temp_out);
+        for (j = 0; j < 16; ++j)
+          dest[j * pitch + i] = clip_pixel(ROUND_POWER_OF_TWO(temp_out[j], 6) +
+                                           dest[j * pitch + i]);
+      }
+    } break;
+    default: printf("av1_short_iht16x16_add_dspr2 : Invalid tx_type\n"); break;
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/av1/common/mips/dspr2/av1_itrans4_dspr2.c b/third_party/aom/av1/common/mips/dspr2/av1_itrans4_dspr2.c
new file mode 100644
index 0000000000..0a9552376d
--- /dev/null
+++ b/third_party/aom/av1/common/mips/dspr2/av1_itrans4_dspr2.c
@@ -0,0 +1,91 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "av1/common/common.h"
+#include "av1/common/blockd.h"
+#include "av1/common/idct.h"
+#include "aom_dsp/mips/inv_txfm_dspr2.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+void av1_iht4x4_16_add_dspr2(const int16_t *input, uint8_t *dest,
+                             int dest_stride, int tx_type) {
+  int i, j;
+  DECLARE_ALIGNED(32, int16_t, out[4 * 4]);
+  int16_t *outptr = out;
+  int16_t temp_in[4 * 4], temp_out[4];
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp      %[pos],     1           \n\t"
+                       :
+                       : [pos] "r"(pos));
+
+  switch (tx_type) {
+    case DCT_DCT:  // DCT in both horizontal and vertical
+      aom_idct4_rows_dspr2(input, outptr);
+      aom_idct4_columns_add_blk_dspr2(&out[0], dest, dest_stride);
+      break;
+    case ADST_DCT:  // ADST in vertical, DCT in horizontal
+      aom_idct4_rows_dspr2(input, outptr);
+
+      outptr = out;
+
+      for (i = 0; i < 4; ++i) {
+        iadst4_dspr2(outptr, temp_out);
+
+        for (j = 0; j < 4; ++j)
+          dest[j * dest_stride + i] = clip_pixel(
+              ROUND_POWER_OF_TWO(temp_out[j], 4) + dest[j * dest_stride + i]);
+
+        outptr += 4;
+      }
+      break;
+    case DCT_ADST:  // DCT in vertical, ADST in horizontal
+      for (i = 0; i < 4; ++i) {
+        iadst4_dspr2(input, outptr);
+        input += 4;
+        outptr += 4;
+      }
+
+      for (i = 0; i < 4; ++i) {
+        for (j = 0; j < 4; ++j) {
+          temp_in[i * 4 + j] = out[j * 4 + i];
+        }
+      }
+      aom_idct4_columns_add_blk_dspr2(&temp_in[0], dest, dest_stride);
+      break;
+    case ADST_ADST:  // ADST in both directions
+      for (i = 0; i < 4; ++i) {
+        iadst4_dspr2(input, outptr);
+        input += 4;
+        outptr += 4;
+      }
+
+      for (i = 0; i < 4; ++i) {
+        for (j = 0; j < 4; ++j) temp_in[j] = out[j * 4 + i];
+        iadst4_dspr2(temp_in, temp_out);
+
+        for (j = 0; j < 4; ++j)
+          dest[j * dest_stride + i] = clip_pixel(
+              ROUND_POWER_OF_TWO(temp_out[j], 4) + dest[j * dest_stride + i]);
+      }
+      break;
+    default: printf("av1_short_iht4x4_add_dspr2 : Invalid tx_type\n"); break;
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/av1/common/mips/dspr2/av1_itrans8_dspr2.c b/third_party/aom/av1/common/mips/dspr2/av1_itrans8_dspr2.c
new file mode 100644
index 0000000000..8bf5b4f0e5
--- /dev/null
+++ b/third_party/aom/av1/common/mips/dspr2/av1_itrans8_dspr2.c
@@ -0,0 +1,85 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "av1/common/common.h"
+#include "av1/common/blockd.h"
+#include "aom_dsp/mips/inv_txfm_dspr2.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_ports/mem.h"
+
+#if HAVE_DSPR2
+void av1_iht8x8_64_add_dspr2(const int16_t *input, uint8_t *dest,
+                             int dest_stride, int tx_type) {
+  int i, j;
+  DECLARE_ALIGNED(32, int16_t, out[8 * 8]);
+  int16_t *outptr = out;
+  int16_t temp_in[8 * 8], temp_out[8];
+  uint32_t pos = 45;
+
+  /* bit positon for extract from acc */
+  __asm__ __volatile__("wrdsp    %[pos],    1    \n\t" : : [pos] "r"(pos));
+
+  switch (tx_type) {
+    case DCT_DCT:  // DCT in both horizontal and vertical
+      idct8_rows_dspr2(input, outptr, 8);
+      idct8_columns_add_blk_dspr2(&out[0], dest, dest_stride);
+      break;
+    case ADST_DCT:  // ADST in vertical, DCT in horizontal
+      idct8_rows_dspr2(input, outptr, 8);
+
+      for (i = 0; i < 8; ++i) {
+        iadst8_dspr2(&out[i * 8], temp_out);
+
+        for (j = 0; j < 8; ++j)
+          dest[j * dest_stride + i] = clip_pixel(
+              ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i]);
+      }
+      break;
+    case DCT_ADST:  // DCT in vertical, ADST in horizontal
+      for (i = 0; i < 8; ++i) {
+        iadst8_dspr2(input, outptr);
+        input += 8;
+        outptr += 8;
+      }
+
+      for (i = 0; i < 8; ++i) {
+        for (j = 0; j < 8; ++j) {
+          temp_in[i * 8 + j] = out[j * 8 + i];
+        }
+      }
+      idct8_columns_add_blk_dspr2(&temp_in[0], dest, dest_stride);
+      break;
+    case ADST_ADST:  // ADST in both directions
+      for (i = 0; i < 8; ++i) {
+        iadst8_dspr2(input, outptr);
+        input += 8;
+        outptr += 8;
+      }
+
+      for (i = 0; i < 8; ++i) {
+        for (j = 0; j < 8; ++j) temp_in[j] = out[j * 8 + i];
+
+        iadst8_dspr2(temp_in, temp_out);
+
+        for (j = 0; j < 8; ++j)
+          dest[j * dest_stride + i] = clip_pixel(
+              ROUND_POWER_OF_TWO(temp_out[j], 5) + dest[j * dest_stride + i]);
+      }
+      break;
+    default: printf("av1_short_iht8x8_add_dspr2 : Invalid tx_type\n"); break;
+  }
+}
+#endif  // #if HAVE_DSPR2
diff --git a/third_party/aom/av1/common/mips/msa/av1_idct16x16_msa.c b/third_party/aom/av1/common/mips/msa/av1_idct16x16_msa.c
new file mode 100644
index 0000000000..4bd0a16356
--- /dev/null
+++ b/third_party/aom/av1/common/mips/msa/av1_idct16x16_msa.c
@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/enums.h"
+#include "aom_dsp/mips/inv_txfm_msa.h"
+
+void av1_iht16x16_256_add_msa(const int16_t *input, uint8_t *dst,
+                              int32_t dst_stride, int32_t tx_type) {
+  int32_t i;
+  DECLARE_ALIGNED(32, int16_t, out[16 * 16]);
+  int16_t *out_ptr = &out[0];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      /* transform rows */
+      for (i = 0; i < 2; ++i) {
+        /* process 16 * 8 block */
+        aom_idct16_1d_rows_msa((input + (i << 7)), (out_ptr + (i << 7)));
+      }
+
+      /* transform columns */
+      for (i = 0; i < 2; ++i) {
+        /* process 8 * 16 block */
+        aom_idct16_1d_columns_addblk_msa((out_ptr + (i << 3)), (dst + (i << 3)),
+                                         dst_stride);
+      }
+      break;
+    case ADST_DCT:
+      /* transform rows */
+      for (i = 0; i < 2; ++i) {
+        /* process 16 * 8 block */
+        aom_idct16_1d_rows_msa((input + (i << 7)), (out_ptr + (i << 7)));
+      }
+
+      /* transform columns */
+      for (i = 0; i < 2; ++i) {
+        aom_iadst16_1d_columns_addblk_msa((out_ptr + (i << 3)),
+                                          (dst + (i << 3)), dst_stride);
+      }
+      break;
+    case DCT_ADST:
+      /* transform rows */
+      for (i = 0; i < 2; ++i) {
+        /* process 16 * 8 block */
+        aom_iadst16_1d_rows_msa((input + (i << 7)), (out_ptr + (i << 7)));
+      }
+
+      /* transform columns */
+      for (i = 0; i < 2; ++i) {
+        /* process 8 * 16 block */
+        aom_idct16_1d_columns_addblk_msa((out_ptr + (i << 3)), (dst + (i << 3)),
+                                         dst_stride);
+      }
+      break;
+    case ADST_ADST:
+      /* transform rows */
+      for (i = 0; i < 2; ++i) {
+        /* process 16 * 8 block */
+        aom_iadst16_1d_rows_msa((input + (i << 7)), (out_ptr + (i << 7)));
+      }
+
+      /* transform columns */
+      for (i = 0; i < 2; ++i) {
+        aom_iadst16_1d_columns_addblk_msa((out_ptr + (i << 3)),
+                                          (dst + (i << 3)), dst_stride);
+      }
+      break;
+    default: assert(0); break;
+  }
+}
diff --git a/third_party/aom/av1/common/mips/msa/av1_idct4x4_msa.c b/third_party/aom/av1/common/mips/msa/av1_idct4x4_msa.c
new file mode 100644
index 0000000000..8364f8dc44
--- /dev/null
+++ b/third_party/aom/av1/common/mips/msa/av1_idct4x4_msa.c
@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/enums.h"
+#include "aom_dsp/mips/inv_txfm_msa.h"
+
+void av1_iht4x4_16_add_msa(const int16_t *input, uint8_t *dst,
+                           int32_t dst_stride, int32_t tx_type) {
+  v8i16 in0, in1, in2, in3;
+
+  /* load vector elements of 4x4 block */
+  LD4x4_SH(input, in0, in1, in2, in3);
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      /* DCT in horizontal */
+      AOM_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      /* DCT in vertical */
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case ADST_DCT:
+      /* DCT in horizontal */
+      AOM_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      /* ADST in vertical */
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_IADST4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case DCT_ADST:
+      /* ADST in horizontal */
+      AOM_IADST4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      /* DCT in vertical */
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_IDCT4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case ADST_ADST:
+      /* ADST in horizontal */
+      AOM_IADST4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      /* ADST in vertical */
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_IADST4x4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    default: assert(0); break;
+  }
+
+  /* final rounding (add 2^3, divide by 2^4) and shift */
+  SRARI_H4_SH(in0, in1, in2, in3, 4);
+  /* add block and store 4x4 */
+  ADDBLK_ST4x4_UB(in0, in1, in2, in3, dst, dst_stride);
+}
diff --git a/third_party/aom/av1/common/mips/msa/av1_idct8x8_msa.c b/third_party/aom/av1/common/mips/msa/av1_idct8x8_msa.c
new file mode 100644
index 0000000000..71117051bc
--- /dev/null
+++ b/third_party/aom/av1/common/mips/msa/av1_idct8x8_msa.c
@@ -0,0 +1,79 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/enums.h"
+#include "aom_dsp/mips/inv_txfm_msa.h"
+
+void av1_iht8x8_64_add_msa(const int16_t *input, uint8_t *dst,
+                           int32_t dst_stride, int32_t tx_type) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  /* load vector elements of 8x8 block */
+  LD_SH8(input, 8, in0, in1, in2, in3, in4, in5, in6, in7);
+
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      /* DCT in horizontal */
+      AOM_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+      /* DCT in vertical */
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+      break;
+    case ADST_DCT:
+      /* DCT in horizontal */
+      AOM_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+      /* ADST in vertical */
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      break;
+    case DCT_ADST:
+      /* ADST in horizontal */
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      /* DCT in vertical */
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_IDCT8x8_1D(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+      break;
+    case ADST_ADST:
+      /* ADST in horizontal */
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      /* ADST in vertical */
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      break;
+    default: assert(0); break;
+  }
+
+  /* final rounding (add 2^4, divide by 2^5) and shift */
+  SRARI_H4_SH(in0, in1, in2, in3, 5);
+  SRARI_H4_SH(in4, in5, in6, in7, 5);
+
+  /* add block and store 8x8 */
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, in0, in1, in2, in3);
+  dst += (4 * dst_stride);
+  AOM_ADDBLK_ST8x4_UB(dst, dst_stride, in4, in5, in6, in7);
+}
diff --git a/third_party/aom/av1/common/mv.h b/third_party/aom/av1/common/mv.h
new file mode 100644
index 0000000000..d4df3790f3
--- /dev/null
+++ b/third_party/aom/av1/common/mv.h
@@ -0,0 +1,302 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_MV_H_
+#define AV1_COMMON_MV_H_
+
+#include "av1/common/common.h"
+#include "av1/common/common_data.h"
+#include "aom_dsp/aom_filter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct mv {
+  int16_t row;
+  int16_t col;
+} MV;
+
+typedef union int_mv {
+  uint32_t as_int;
+  MV as_mv;
+} int_mv; /* facilitates faster equality tests and copies */
+
+typedef struct mv32 {
+  int32_t row;
+  int32_t col;
+} MV32;
+
+#if (CONFIG_WARPED_MOTION || CONFIG_MOTION_VAR) && CONFIG_GLOBAL_MOTION
+#define SEPARATE_GLOBAL_MOTION 1
+#endif  // (CONFIG_WARPED_MOTION || CONFIG_MOTION_VAR) && CONFIG_GLOBAL_MOTION
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+// Bits of precision used for the model
+#define WARPEDMODEL_PREC_BITS 16
+#define WARPEDMODEL_ROW3HOMO_PREC_BITS 16
+
+#define WARPEDMODEL_TRANS_CLAMP (128 << WARPEDMODEL_PREC_BITS)
+#define WARPEDMODEL_DIAGAFFINE_CLAMP (1 << (WARPEDMODEL_PREC_BITS + 1))
+#define WARPEDMODEL_NONDIAGAFFINE_CLAMP (1 << (WARPEDMODEL_PREC_BITS - 1))
+#define WARPEDMODEL_ROW3HOMO_CLAMP (1 << (WARPEDMODEL_PREC_BITS - 1))
+
+// Bits of subpel precision for warped interpolation
+#define WARPEDPIXEL_PREC_BITS 6
+#define WARPEDPIXEL_PREC_SHIFTS (1 << WARPEDPIXEL_PREC_BITS)
+
+// Taps for ntap filter
+#define WARPEDPIXEL_FILTER_TAPS 6
+
+// Precision of filter taps
+#define WARPEDPIXEL_FILTER_BITS 7
+
+// Precision bits reduction after horizontal shear
+#define HORSHEAR_REDUCE_PREC_BITS 5
+#define VERSHEAR_REDUCE_PREC_BITS \
+  (2 * WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS)
+
+#define WARPEDDIFF_PREC_BITS (WARPEDMODEL_PREC_BITS - WARPEDPIXEL_PREC_BITS)
+
+/* clang-format off */
+typedef enum {
+  IDENTITY = 0,      // identity transformation, 0-parameter
+  TRANSLATION = 1,   // translational motion 2-parameter
+  ROTZOOM = 2,       // simplified affine with rotation + zoom only, 4-parameter
+  AFFINE = 3,        // affine, 6-parameter
+  HORTRAPEZOID = 4,  // constrained homography, hor trapezoid, 6-parameter
+  VERTRAPEZOID = 5,  // constrained homography, ver trapezoid, 6-parameter
+  HOMOGRAPHY = 6,    // homography, 8-parameter
+  TRANS_TYPES = 7,
+} TransformationType;
+/* clang-format on */
+
+// Number of types used for global motion (must be >= 3 and <= TRANS_TYPES)
+// The following can be useful:
+// GLOBAL_TRANS_TYPES 3 - up to rotation-zoom
+// GLOBAL_TRANS_TYPES 4 - up to affine
+// GLOBAL_TRANS_TYPES 6 - up to hor/ver trapezoids
+// GLOBAL_TRANS_TYPES 7 - up to full homography
+#define GLOBAL_TRANS_TYPES 4
+
+typedef struct {
+#if CONFIG_GLOBAL_MOTION
+  int global_warp_allowed;
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_WARPED_MOTION
+  int local_warp_allowed;
+#endif  // CONFIG_WARPED_MOTION
+} WarpTypesAllowed;
+
+// number of parameters used by each transformation in TransformationTypes
+static const int trans_model_params[TRANS_TYPES] = { 0, 2, 4, 6, 6, 6, 8 };
+
+// The order of values in the wmmat matrix below is best described
+// by the homography:
+//      [x'     (m2 m3 m0   [x
+//  z .  y'  =   m4 m5 m1 *  y
+//       1]      m6 m7 1)    1]
+typedef struct {
+  TransformationType wmtype;
+  int32_t wmmat[8];
+  int16_t alpha, beta, gamma, delta;
+} WarpedMotionParams;
+
+static INLINE void set_default_warp_params(WarpedMotionParams *wm) {
+  static const int32_t default_wm_mat[8] = {
+    0, 0, (1 << WARPEDMODEL_PREC_BITS), 0, 0, (1 << WARPEDMODEL_PREC_BITS), 0, 0
+  };
+  memset(wm, 0, sizeof(*wm));
+  memcpy(wm->wmmat, default_wm_mat, sizeof(wm->wmmat));
+  wm->wmtype = IDENTITY;
+}
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+
+#if CONFIG_GLOBAL_MOTION
+// The following constants describe the various precisions
+// of different parameters in the global motion experiment.
+//
+// Given the general homography:
+//      [x'     (a  b  c   [x
+//  z .  y'  =   d  e  f *  y
+//       1]      g  h  i)    1]
+//
+// Constants using the name ALPHA here are related to parameters
+// a, b, d, e. Constants using the name TRANS are related
+// to parameters c and f.
+//
+// Anything ending in PREC_BITS is the number of bits of precision
+// to maintain when converting from double to integer.
+//
+// The ABS parameters are used to create an upper and lower bound
+// for each parameter. In other words, after a parameter is integerized
+// it is clamped between -(1 << ABS_XXX_BITS) and (1 << ABS_XXX_BITS).
+//
+// XXX_PREC_DIFF and XXX_DECODE_FACTOR
+// are computed once here to prevent repetitive
+// computation on the decoder side. These are
+// to allow the global motion parameters to be encoded in a lower
+// precision than the warped model precision. This means that they
+// need to be changed to warped precision when they are decoded.
+//
+// XX_MIN, XX_MAX are also computed to avoid repeated computation
+
+#define SUBEXPFIN_K 3
+#define GM_TRANS_PREC_BITS 6
+#define GM_ABS_TRANS_BITS 12
+#define GM_ABS_TRANS_ONLY_BITS (GM_ABS_TRANS_BITS - GM_TRANS_PREC_BITS + 3)
+#define GM_TRANS_PREC_DIFF (WARPEDMODEL_PREC_BITS - GM_TRANS_PREC_BITS)
+#define GM_TRANS_ONLY_PREC_DIFF (WARPEDMODEL_PREC_BITS - 3)
+#define GM_TRANS_DECODE_FACTOR (1 << GM_TRANS_PREC_DIFF)
+#define GM_TRANS_ONLY_DECODE_FACTOR (1 << GM_TRANS_ONLY_PREC_DIFF)
+
+#define GM_ALPHA_PREC_BITS 15
+#define GM_ABS_ALPHA_BITS 12
+#define GM_ALPHA_PREC_DIFF (WARPEDMODEL_PREC_BITS - GM_ALPHA_PREC_BITS)
+#define GM_ALPHA_DECODE_FACTOR (1 << GM_ALPHA_PREC_DIFF)
+
+#define GM_ROW3HOMO_PREC_BITS 16
+#define GM_ABS_ROW3HOMO_BITS 11
+#define GM_ROW3HOMO_PREC_DIFF \
+  (WARPEDMODEL_ROW3HOMO_PREC_BITS - GM_ROW3HOMO_PREC_BITS)
+#define GM_ROW3HOMO_DECODE_FACTOR (1 << GM_ROW3HOMO_PREC_DIFF)
+
+#define GM_TRANS_MAX (1 << GM_ABS_TRANS_BITS)
+#define GM_ALPHA_MAX (1 << GM_ABS_ALPHA_BITS)
+#define GM_ROW3HOMO_MAX (1 << GM_ABS_ROW3HOMO_BITS)
+
+#define GM_TRANS_MIN -GM_TRANS_MAX
+#define GM_ALPHA_MIN -GM_ALPHA_MAX
+#define GM_ROW3HOMO_MIN -GM_ROW3HOMO_MAX
+
+// Use global motion parameters for sub8x8 blocks
+#define GLOBAL_SUB8X8_USED 0
+
+static INLINE int block_center_x(int mi_col, BLOCK_SIZE bs) {
+  const int bw = block_size_wide[bs];
+  return mi_col * MI_SIZE + bw / 2 - 1;
+}
+
+static INLINE int block_center_y(int mi_row, BLOCK_SIZE bs) {
+  const int bh = block_size_high[bs];
+  return mi_row * MI_SIZE + bh / 2 - 1;
+}
+
+static INLINE int convert_to_trans_prec(int allow_hp, int coor) {
+  if (allow_hp)
+    return ROUND_POWER_OF_TWO_SIGNED(coor, WARPEDMODEL_PREC_BITS - 3);
+  else
+    return ROUND_POWER_OF_TWO_SIGNED(coor, WARPEDMODEL_PREC_BITS - 2) * 2;
+}
+
+// Convert a global motion translation vector (which may have more bits than a
+// regular motion vector) into a motion vector
+static INLINE int_mv gm_get_motion_vector(const WarpedMotionParams *gm,
+                                          int allow_hp, BLOCK_SIZE bsize,
+                                          int mi_col, int mi_row,
+                                          int block_idx) {
+  const int unify_bsize = CONFIG_CB4X4;
+  int_mv res;
+  const int32_t *mat = gm->wmmat;
+  int x, y, tx, ty;
+
+  if (gm->wmtype == TRANSLATION) {
+    res.as_mv.row = gm->wmmat[0] >> GM_TRANS_ONLY_PREC_DIFF;
+    res.as_mv.col = gm->wmmat[1] >> GM_TRANS_ONLY_PREC_DIFF;
+    return res;
+  }
+
+  if (bsize >= BLOCK_8X8 || unify_bsize) {
+    x = block_center_x(mi_col, bsize);
+    y = block_center_y(mi_row, bsize);
+  } else {
+    x = block_center_x(mi_col, bsize);
+    y = block_center_y(mi_row, bsize);
+    x += (block_idx & 1) * MI_SIZE / 2;
+    y += (block_idx & 2) * MI_SIZE / 4;
+  }
+
+  if (gm->wmtype == ROTZOOM) {
+    assert(gm->wmmat[5] == gm->wmmat[2]);
+    assert(gm->wmmat[4] == -gm->wmmat[3]);
+  }
+  if (gm->wmtype > AFFINE) {
+    int xc = (int)((int64_t)mat[2] * x + (int64_t)mat[3] * y + mat[0]);
+    int yc = (int)((int64_t)mat[4] * x + (int64_t)mat[5] * y + mat[1]);
+    const int Z = (int)((int64_t)mat[6] * x + (int64_t)mat[7] * y +
+                        (1 << WARPEDMODEL_ROW3HOMO_PREC_BITS));
+    xc *= 1 << (WARPEDMODEL_ROW3HOMO_PREC_BITS - WARPEDMODEL_PREC_BITS);
+    yc *= 1 << (WARPEDMODEL_ROW3HOMO_PREC_BITS - WARPEDMODEL_PREC_BITS);
+    xc = (int)(xc > 0 ? ((int64_t)xc + Z / 2) / Z : ((int64_t)xc - Z / 2) / Z);
+    yc = (int)(yc > 0 ? ((int64_t)yc + Z / 2) / Z : ((int64_t)yc - Z / 2) / Z);
+    tx = convert_to_trans_prec(allow_hp, xc) - (x << 3);
+    ty = convert_to_trans_prec(allow_hp, yc) - (y << 3);
+  } else {
+    const int xc =
+        (mat[2] - (1 << WARPEDMODEL_PREC_BITS)) * x + mat[3] * y + mat[0];
+    const int yc =
+        mat[4] * x + (mat[5] - (1 << WARPEDMODEL_PREC_BITS)) * y + mat[1];
+    tx = convert_to_trans_prec(allow_hp, xc);
+    ty = convert_to_trans_prec(allow_hp, yc);
+  }
+
+  res.as_mv.row = ty;
+  res.as_mv.col = tx;
+  return res;
+}
+
+static INLINE TransformationType get_gmtype(const WarpedMotionParams *gm) {
+  if (gm->wmmat[6] != 0 || gm->wmmat[7] != 0) {
+    if (!gm->wmmat[6] && !gm->wmmat[4]) return HORTRAPEZOID;
+    if (!gm->wmmat[7] && !gm->wmmat[3]) return VERTRAPEZOID;
+    return HOMOGRAPHY;
+  }
+  if (gm->wmmat[5] == (1 << WARPEDMODEL_PREC_BITS) && !gm->wmmat[4] &&
+      gm->wmmat[2] == (1 << WARPEDMODEL_PREC_BITS) && !gm->wmmat[3]) {
+    return ((!gm->wmmat[1] && !gm->wmmat[0]) ? IDENTITY : TRANSLATION);
+  }
+  if (gm->wmmat[2] == gm->wmmat[5] && gm->wmmat[3] == -gm->wmmat[4])
+    return ROTZOOM;
+  else
+    return AFFINE;
+}
+#endif  // CONFIG_GLOBAL_MOTION
+
+#if CONFIG_REF_MV
+typedef struct candidate_mv {
+  int_mv this_mv;
+  int_mv comp_mv;
+  uint8_t pred_diff[2];
+  int weight;
+} CANDIDATE_MV;
+#endif
+
+static INLINE int is_zero_mv(const MV *mv) {
+  return *((const uint32_t *)mv) == 0;
+}
+
+static INLINE int is_equal_mv(const MV *a, const MV *b) {
+  return *((const uint32_t *)a) == *((const uint32_t *)b);
+}
+
+static INLINE void clamp_mv(MV *mv, int min_col, int max_col, int min_row,
+                            int max_row) {
+  mv->col = clamp(mv->col, min_col, max_col);
+  mv->row = clamp(mv->row, min_row, max_row);
+}
+
+static INLINE int mv_has_subpel(const MV *mv) {
+  return (mv->row & SUBPEL_MASK) || (mv->col & SUBPEL_MASK);
+}
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_MV_H_
diff --git a/third_party/aom/av1/common/mvref_common.c b/third_party/aom/av1/common/mvref_common.c
new file mode 100644
index 0000000000..5222948c88
--- /dev/null
+++ b/third_party/aom/av1/common/mvref_common.c
@@ -0,0 +1,1164 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/mvref_common.h"
+#if CONFIG_WARPED_MOTION
+#include "av1/common/warped_motion.h"
+#endif  // CONFIG_WARPED_MOTION
+
+#if CONFIG_REF_MV
+
+static uint8_t add_ref_mv_candidate(
+    const MODE_INFO *const candidate_mi, const MB_MODE_INFO *const candidate,
+    const MV_REFERENCE_FRAME rf[2], uint8_t *refmv_count,
+    CANDIDATE_MV *ref_mv_stack, const int use_hp, int len, int block, int col) {
+  int index = 0, ref;
+  int newmv_count = 0;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+  if (rf[1] == NONE_FRAME) {
+    // single reference frame
+    for (ref = 0; ref < 2; ++ref) {
+      if (candidate->ref_frame[ref] == rf[0]) {
+        int_mv this_refmv = get_sub_block_mv(candidate_mi, ref, col, block);
+        lower_mv_precision(&this_refmv.as_mv, use_hp);
+
+        for (index = 0; index < *refmv_count; ++index)
+          if (ref_mv_stack[index].this_mv.as_int == this_refmv.as_int) break;
+
+        if (index < *refmv_count) ref_mv_stack[index].weight += 2 * len;
+
+        // Add a new item to the list.
+        if (index == *refmv_count) {
+          ref_mv_stack[index].this_mv = this_refmv;
+          ref_mv_stack[index].pred_diff[0] = av1_get_pred_diff_ctx(
+              get_sub_block_pred_mv(candidate_mi, ref, col, block), this_refmv);
+          ref_mv_stack[index].weight = 2 * len;
+          ++(*refmv_count);
+
+          if (candidate->mode == NEWMV) ++newmv_count;
+        }
+
+        if (candidate_mi->mbmi.sb_type < BLOCK_8X8 && block >= 0 &&
+            !unify_bsize) {
+          int alt_block = 3 - block;
+          this_refmv = get_sub_block_mv(candidate_mi, ref, col, alt_block);
+          lower_mv_precision(&this_refmv.as_mv, use_hp);
+
+          for (index = 0; index < *refmv_count; ++index)
+            if (ref_mv_stack[index].this_mv.as_int == this_refmv.as_int) break;
+
+          if (index < *refmv_count) ref_mv_stack[index].weight += len;
+
+          // Add a new item to the list.
+          if (index == *refmv_count) {
+            ref_mv_stack[index].this_mv = this_refmv;
+            ref_mv_stack[index].pred_diff[0] = av1_get_pred_diff_ctx(
+                get_sub_block_pred_mv(candidate_mi, ref, col, alt_block),
+                this_refmv);
+            ref_mv_stack[index].weight = len;
+            ++(*refmv_count);
+
+            if (candidate->mode == NEWMV) ++newmv_count;
+          }
+        }
+      }
+    }
+  } else {
+    // compound reference frame
+    if (candidate->ref_frame[0] == rf[0] && candidate->ref_frame[1] == rf[1]) {
+      int_mv this_refmv[2];
+
+      for (ref = 0; ref < 2; ++ref) {
+        this_refmv[ref] = get_sub_block_mv(candidate_mi, ref, col, block);
+        lower_mv_precision(&this_refmv[ref].as_mv, use_hp);
+      }
+
+      for (index = 0; index < *refmv_count; ++index)
+        if ((ref_mv_stack[index].this_mv.as_int == this_refmv[0].as_int) &&
+            (ref_mv_stack[index].comp_mv.as_int == this_refmv[1].as_int))
+          break;
+
+      if (index < *refmv_count) ref_mv_stack[index].weight += 2 * len;
+
+      // Add a new item to the list.
+      if (index == *refmv_count) {
+        ref_mv_stack[index].this_mv = this_refmv[0];
+        ref_mv_stack[index].comp_mv = this_refmv[1];
+        ref_mv_stack[index].pred_diff[0] = av1_get_pred_diff_ctx(
+            get_sub_block_pred_mv(candidate_mi, 0, col, block), this_refmv[0]);
+        ref_mv_stack[index].pred_diff[1] = av1_get_pred_diff_ctx(
+            get_sub_block_pred_mv(candidate_mi, 1, col, block), this_refmv[1]);
+        ref_mv_stack[index].weight = 2 * len;
+        ++(*refmv_count);
+
+#if CONFIG_EXT_INTER
+        if (candidate->mode == NEW_NEWMV)
+#else
+        if (candidate->mode == NEWMV)
+#endif  // CONFIG_EXT_INTER
+          ++newmv_count;
+      }
+
+      if (candidate_mi->mbmi.sb_type < BLOCK_8X8 && block >= 0 &&
+          !unify_bsize) {
+        int alt_block = 3 - block;
+        this_refmv[0] = get_sub_block_mv(candidate_mi, 0, col, alt_block);
+        this_refmv[1] = get_sub_block_mv(candidate_mi, 1, col, alt_block);
+
+        for (ref = 0; ref < 2; ++ref)
+          lower_mv_precision(&this_refmv[ref].as_mv, use_hp);
+
+        for (index = 0; index < *refmv_count; ++index)
+          if (ref_mv_stack[index].this_mv.as_int == this_refmv[0].as_int &&
+              ref_mv_stack[index].comp_mv.as_int == this_refmv[1].as_int)
+            break;
+
+        if (index < *refmv_count) ref_mv_stack[index].weight += len;
+
+        // Add a new item to the list.
+        if (index == *refmv_count) {
+          ref_mv_stack[index].this_mv = this_refmv[0];
+          ref_mv_stack[index].comp_mv = this_refmv[1];
+          ref_mv_stack[index].pred_diff[0] = av1_get_pred_diff_ctx(
+              get_sub_block_pred_mv(candidate_mi, 0, col, block),
+              this_refmv[0]);
+          ref_mv_stack[index].pred_diff[0] = av1_get_pred_diff_ctx(
+              get_sub_block_pred_mv(candidate_mi, 1, col, block),
+              this_refmv[1]);
+          ref_mv_stack[index].weight = len;
+          ++(*refmv_count);
+
+#if CONFIG_EXT_INTER
+          if (candidate->mode == NEW_NEWMV)
+#else
+          if (candidate->mode == NEWMV)
+#endif  // CONFIG_EXT_INTER
+            ++newmv_count;
+        }
+      }
+    }
+  }
+  return newmv_count;
+}
+
+static uint8_t scan_row_mbmi(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                             const int mi_row, const int mi_col, int block,
+                             const MV_REFERENCE_FRAME rf[2], int row_offset,
+                             CANDIDATE_MV *ref_mv_stack, uint8_t *refmv_count) {
+  const TileInfo *const tile = &xd->tile;
+  int i;
+  uint8_t newmv_count = 0;
+#if CONFIG_CB4X4
+  const int bsize = xd->mi[0]->mbmi.sb_type;
+  const int mi_offset =
+      bsize < BLOCK_8X8 ? mi_size_wide[BLOCK_4X4] : mi_size_wide[BLOCK_8X8];
+  // TODO(jingning): Revisit this part after cb4x4 is stable.
+  if (bsize >= BLOCK_8X8) row_offset *= 2;
+#else
+  const int mi_offset = mi_size_wide[BLOCK_8X8];
+#endif
+
+  for (i = 0; i < xd->n8_w && *refmv_count < MAX_REF_MV_STACK_SIZE;) {
+    POSITION mi_pos;
+#if CONFIG_CB4X4
+    const int use_step_16 = (xd->n8_w >= 16);
+#else
+    const int use_step_16 = (xd->n8_w >= 8);
+#endif
+
+    mi_pos.row = row_offset;
+    mi_pos.col = i;
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, cm, &mi_pos)) {
+      const MODE_INFO *const candidate_mi =
+          xd->mi[mi_pos.row * xd->mi_stride + mi_pos.col];
+      const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
+      int len = AOMMIN(xd->n8_w, mi_size_wide[candidate->sb_type]);
+      if (use_step_16) len = AOMMAX(mi_size_wide[BLOCK_16X16], len);
+      newmv_count += add_ref_mv_candidate(
+          candidate_mi, candidate, rf, refmv_count, ref_mv_stack,
+          cm->allow_high_precision_mv, len, block, mi_pos.col);
+      i += len;
+    } else {
+      if (use_step_16)
+        i += (mi_offset << 1);
+      else
+        i += mi_offset;
+    }
+  }
+
+  return newmv_count;
+}
+
+static uint8_t scan_col_mbmi(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                             const int mi_row, const int mi_col, int block,
+                             const MV_REFERENCE_FRAME rf[2], int col_offset,
+                             CANDIDATE_MV *ref_mv_stack, uint8_t *refmv_count) {
+  const TileInfo *const tile = &xd->tile;
+  int i;
+  uint8_t newmv_count = 0;
+#if CONFIG_CB4X4
+  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  const int mi_offset =
+      (bsize < BLOCK_8X8) ? mi_size_high[BLOCK_4X4] : mi_size_high[BLOCK_8X8];
+  if (bsize >= BLOCK_8X8) col_offset *= 2;
+#else
+  const int mi_offset = mi_size_wide[BLOCK_8X8];
+#endif
+
+  for (i = 0; i < xd->n8_h && *refmv_count < MAX_REF_MV_STACK_SIZE;) {
+    POSITION mi_pos;
+#if CONFIG_CB4X4
+    const int use_step_16 = (xd->n8_h >= 16);
+#else
+    const int use_step_16 = (xd->n8_h >= 8);
+#endif
+
+    mi_pos.row = i;
+    mi_pos.col = col_offset;
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, cm, &mi_pos)) {
+      const MODE_INFO *const candidate_mi =
+          xd->mi[mi_pos.row * xd->mi_stride + mi_pos.col];
+      const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
+      int len = AOMMIN(xd->n8_h, mi_size_high[candidate->sb_type]);
+      if (use_step_16) len = AOMMAX(mi_size_high[BLOCK_16X16], len);
+      newmv_count += add_ref_mv_candidate(
+          candidate_mi, candidate, rf, refmv_count, ref_mv_stack,
+          cm->allow_high_precision_mv, len, block, mi_pos.col);
+      i += len;
+    } else {
+      if (use_step_16)
+        i += (mi_offset << 1);
+      else
+        i += mi_offset;
+    }
+  }
+
+  return newmv_count;
+}
+
+static uint8_t scan_blk_mbmi(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                             const int mi_row, const int mi_col, int block,
+                             const MV_REFERENCE_FRAME rf[2], int row_offset,
+                             int col_offset, CANDIDATE_MV *ref_mv_stack,
+                             uint8_t *refmv_count) {
+  const TileInfo *const tile = &xd->tile;
+  POSITION mi_pos;
+  uint8_t newmv_count = 0;
+
+  mi_pos.row = row_offset;
+  mi_pos.col = col_offset;
+
+  if (is_inside(tile, mi_col, mi_row, cm->mi_rows, cm, &mi_pos) &&
+      *refmv_count < MAX_REF_MV_STACK_SIZE) {
+    const MODE_INFO *const candidate_mi =
+        xd->mi[mi_pos.row * xd->mi_stride + mi_pos.col];
+    const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
+    const int len = mi_size_wide[BLOCK_8X8];
+
+    newmv_count += add_ref_mv_candidate(
+        candidate_mi, candidate, rf, refmv_count, ref_mv_stack,
+        cm->allow_high_precision_mv, len, block, mi_pos.col);
+  }  // Analyze a single 8x8 block motion information.
+
+  return newmv_count;
+}
+
+static int has_top_right(const MACROBLOCKD *xd, int mi_row, int mi_col,
+                         int bs) {
+  const int mask_row = mi_row & MAX_MIB_MASK;
+  const int mask_col = mi_col & MAX_MIB_MASK;
+
+  // In a split partition all apart from the bottom right has a top right
+  int has_tr = !((mask_row & bs) && (mask_col & bs));
+
+  // bs > 0 and bs is a power of 2
+  assert(bs > 0 && !(bs & (bs - 1)));
+
+  // For each 4x4 group of blocks, when the bottom right is decoded the blocks
+  // to the right have not been decoded therefore the bottom right does
+  // not have a top right
+  while (bs < MAX_MIB_SIZE) {
+    if (mask_col & bs) {
+      if ((mask_col & (2 * bs)) && (mask_row & (2 * bs))) {
+        has_tr = 0;
+        break;
+      }
+    } else {
+      break;
+    }
+    bs <<= 1;
+  }
+
+  // The left hand of two vertical rectangles always has a top right (as the
+  // block above will have been decoded)
+  if (xd->n8_w < xd->n8_h)
+    if (!xd->is_sec_rect) has_tr = 1;
+
+  // The bottom of two horizontal rectangles never has a top right (as the block
+  // to the right won't have been decoded)
+  if (xd->n8_w > xd->n8_h)
+    if (xd->is_sec_rect) has_tr = 0;
+
+#if CONFIG_EXT_PARTITION_TYPES
+  // The bottom left square of a Vertical A does not have a top right as it is
+  // decoded before the right hand rectangle of the partition
+  if (xd->mi[0]->mbmi.partition == PARTITION_VERT_A)
+    if ((mask_row & bs) && !(mask_col & bs)) has_tr = 0;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+  return has_tr;
+}
+
+static int add_col_ref_mv(const AV1_COMMON *cm,
+                          const MV_REF *prev_frame_mvs_base,
+                          const MACROBLOCKD *xd, int mi_row, int mi_col,
+                          MV_REFERENCE_FRAME ref_frame, int blk_row,
+                          int blk_col, uint8_t *refmv_count,
+                          CANDIDATE_MV *ref_mv_stack, int16_t *mode_context) {
+  const MV_REF *prev_frame_mvs =
+      prev_frame_mvs_base + blk_row * cm->mi_cols + blk_col;
+  POSITION mi_pos;
+  int ref, idx;
+  int coll_blk_count = 0;
+  const int weight_unit = mi_size_wide[BLOCK_8X8];
+
+#if CONFIG_MV_COMPRESS
+  mi_pos.row = (mi_row & 0x01) ? blk_row : blk_row + 1;
+  mi_pos.col = (mi_col & 0x01) ? blk_col : blk_col + 1;
+#else
+  mi_pos.row = blk_row;
+  mi_pos.col = blk_col;
+#endif
+
+  if (!is_inside(&xd->tile, mi_col, mi_row, cm->mi_rows, cm, &mi_pos))
+    return coll_blk_count;
+  for (ref = 0; ref < 2; ++ref) {
+    if (prev_frame_mvs->ref_frame[ref] == ref_frame) {
+      int_mv this_refmv = prev_frame_mvs->mv[ref];
+      lower_mv_precision(&this_refmv.as_mv, cm->allow_high_precision_mv);
+
+      if (abs(this_refmv.as_mv.row) >= 16 || abs(this_refmv.as_mv.col) >= 16)
+        mode_context[ref_frame] |= (1 << ZEROMV_OFFSET);
+
+      for (idx = 0; idx < *refmv_count; ++idx)
+        if (this_refmv.as_int == ref_mv_stack[idx].this_mv.as_int) break;
+
+      if (idx < *refmv_count) ref_mv_stack[idx].weight += 2 * weight_unit;
+
+      if (idx == *refmv_count && *refmv_count < MAX_REF_MV_STACK_SIZE) {
+        ref_mv_stack[idx].this_mv.as_int = this_refmv.as_int;
+        ref_mv_stack[idx].pred_diff[0] =
+            av1_get_pred_diff_ctx(prev_frame_mvs->pred_mv[ref], this_refmv);
+        ref_mv_stack[idx].weight = 2 * weight_unit;
+        ++(*refmv_count);
+      }
+
+      ++coll_blk_count;
+    }
+  }
+
+  return coll_blk_count;
+}
+
+static void setup_ref_mv_list(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                              MV_REFERENCE_FRAME ref_frame,
+                              uint8_t *refmv_count, CANDIDATE_MV *ref_mv_stack,
+                              int_mv *mv_ref_list, int block, int mi_row,
+                              int mi_col, int16_t *mode_context) {
+  int idx, nearest_refmv_count = 0;
+  uint8_t newmv_count = 0;
+  CANDIDATE_MV tmp_mv;
+  int len, nr_len;
+
+#if CONFIG_MV_COMPRESS
+  const MV_REF *const prev_frame_mvs_base =
+      cm->use_prev_frame_mvs
+          ? cm->prev_frame->mvs + (((mi_row >> 1) << 1) + 1) * cm->mi_cols +
+                ((mi_col >> 1) << 1) + 1
+          : NULL;
+#else
+  const MV_REF *const prev_frame_mvs_base =
+      cm->use_prev_frame_mvs
+          ? cm->prev_frame->mvs + mi_row * cm->mi_cols + mi_col
+          : NULL;
+#endif
+
+  const int bs = AOMMAX(xd->n8_w, xd->n8_h);
+  const int has_tr = has_top_right(xd, mi_row, mi_col, bs);
+  MV_REFERENCE_FRAME rf[2];
+
+  av1_set_ref_frame(rf, ref_frame);
+  mode_context[ref_frame] = 0;
+  *refmv_count = 0;
+
+  // Scan the first above row mode info.
+  newmv_count += scan_row_mbmi(cm, xd, mi_row, mi_col, block, rf, -1,
+                               ref_mv_stack, refmv_count);
+  // Scan the first left column mode info.
+  newmv_count += scan_col_mbmi(cm, xd, mi_row, mi_col, block, rf, -1,
+                               ref_mv_stack, refmv_count);
+
+  // Check top-right boundary
+  if (has_tr)
+    newmv_count += scan_blk_mbmi(cm, xd, mi_row, mi_col, block, rf, -1,
+                                 xd->n8_w, ref_mv_stack, refmv_count);
+
+  nearest_refmv_count = *refmv_count;
+
+  for (idx = 0; idx < nearest_refmv_count; ++idx)
+    ref_mv_stack[idx].weight += REF_CAT_LEVEL;
+#if CONFIG_TEMPMV_SIGNALING
+  if (cm->use_prev_frame_mvs && rf[1] == NONE_FRAME) {
+#else
+  if (prev_frame_mvs_base && cm->show_frame && cm->last_show_frame &&
+      rf[1] == NONE_FRAME) {
+#endif
+    int blk_row, blk_col;
+    int coll_blk_count = 0;
+#if CONFIG_CB4X4
+    const int mi_step = (xd->n8_w == 1 || xd->n8_h == 1)
+                            ? mi_size_wide[BLOCK_8X8]
+                            : mi_size_wide[BLOCK_16X16];
+#else
+    const int mi_step = mi_size_wide[BLOCK_16X16];
+#endif
+
+#if CONFIG_TPL_MV
+    int tpl_sample_pos[5][2] = { { -1, xd->n8_w },
+                                 { 0, xd->n8_w },
+                                 { xd->n8_h, xd->n8_w },
+                                 { xd->n8_h, 0 },
+                                 { xd->n8_h, -1 } };
+    int i;
+#endif
+
+    for (blk_row = 0; blk_row < xd->n8_h; blk_row += mi_step) {
+      for (blk_col = 0; blk_col < xd->n8_w; blk_col += mi_step) {
+        coll_blk_count += add_col_ref_mv(
+            cm, prev_frame_mvs_base, xd, mi_row, mi_col, ref_frame, blk_row,
+            blk_col, refmv_count, ref_mv_stack, mode_context);
+      }
+    }
+
+#if CONFIG_TPL_MV
+    for (i = 0; i < 5; ++i) {
+      blk_row = tpl_sample_pos[i][0];
+      blk_col = tpl_sample_pos[i][1];
+      coll_blk_count += add_col_ref_mv(cm, prev_frame_mvs_base, xd, mi_row,
+                                       mi_col, ref_frame, blk_row, blk_col,
+                                       refmv_count, ref_mv_stack, mode_context);
+    }
+#endif
+
+    if (coll_blk_count == 0) mode_context[ref_frame] |= (1 << ZEROMV_OFFSET);
+  } else {
+    mode_context[ref_frame] |= (1 << ZEROMV_OFFSET);
+  }
+
+  // Scan the second outer area.
+  scan_blk_mbmi(cm, xd, mi_row, mi_col, block, rf, -1, -1, ref_mv_stack,
+                refmv_count);
+  for (idx = 2; idx <= 3; ++idx) {
+    scan_row_mbmi(cm, xd, mi_row, mi_col, block, rf, -idx, ref_mv_stack,
+                  refmv_count);
+    scan_col_mbmi(cm, xd, mi_row, mi_col, block, rf, -idx, ref_mv_stack,
+                  refmv_count);
+  }
+  scan_col_mbmi(cm, xd, mi_row, mi_col, block, rf, -4, ref_mv_stack,
+                refmv_count);
+
+  switch (nearest_refmv_count) {
+    case 0:
+      mode_context[ref_frame] |= 0;
+      if (*refmv_count >= 1) mode_context[ref_frame] |= 1;
+
+      if (*refmv_count == 1)
+        mode_context[ref_frame] |= (1 << REFMV_OFFSET);
+      else if (*refmv_count >= 2)
+        mode_context[ref_frame] |= (2 << REFMV_OFFSET);
+      break;
+    case 1:
+      mode_context[ref_frame] |= (newmv_count > 0) ? 2 : 3;
+
+      if (*refmv_count == 1)
+        mode_context[ref_frame] |= (3 << REFMV_OFFSET);
+      else if (*refmv_count >= 2)
+        mode_context[ref_frame] |= (4 << REFMV_OFFSET);
+      break;
+
+    case 2:
+    default:
+      if (newmv_count >= 2)
+        mode_context[ref_frame] |= 4;
+      else if (newmv_count == 1)
+        mode_context[ref_frame] |= 5;
+      else
+        mode_context[ref_frame] |= 6;
+
+      mode_context[ref_frame] |= (5 << REFMV_OFFSET);
+      break;
+  }
+
+  // Rank the likelihood and assign nearest and near mvs.
+  len = nearest_refmv_count;
+  while (len > 0) {
+    nr_len = 0;
+    for (idx = 1; idx < len; ++idx) {
+      if (ref_mv_stack[idx - 1].weight < ref_mv_stack[idx].weight) {
+        tmp_mv = ref_mv_stack[idx - 1];
+        ref_mv_stack[idx - 1] = ref_mv_stack[idx];
+        ref_mv_stack[idx] = tmp_mv;
+        nr_len = idx;
+      }
+    }
+    len = nr_len;
+  }
+
+  len = *refmv_count;
+  while (len > nearest_refmv_count) {
+    nr_len = nearest_refmv_count;
+    for (idx = nearest_refmv_count + 1; idx < len; ++idx) {
+      if (ref_mv_stack[idx - 1].weight < ref_mv_stack[idx].weight) {
+        tmp_mv = ref_mv_stack[idx - 1];
+        ref_mv_stack[idx - 1] = ref_mv_stack[idx];
+        ref_mv_stack[idx] = tmp_mv;
+        nr_len = idx;
+      }
+    }
+    len = nr_len;
+  }
+
+  if (rf[1] > NONE_FRAME) {
+    for (idx = 0; idx < *refmv_count; ++idx) {
+      clamp_mv_ref(&ref_mv_stack[idx].this_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                   xd->n8_h << MI_SIZE_LOG2, xd);
+      clamp_mv_ref(&ref_mv_stack[idx].comp_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                   xd->n8_h << MI_SIZE_LOG2, xd);
+    }
+  } else {
+    for (idx = 0; idx < AOMMIN(MAX_MV_REF_CANDIDATES, *refmv_count); ++idx) {
+      mv_ref_list[idx].as_int = ref_mv_stack[idx].this_mv.as_int;
+      clamp_mv_ref(&mv_ref_list[idx].as_mv, xd->n8_w << MI_SIZE_LOG2,
+                   xd->n8_h << MI_SIZE_LOG2, xd);
+    }
+  }
+}
+#endif
+
+// This function searches the neighbourhood of a given MB/SB
+// to try and find candidate reference vectors.
+static void find_mv_refs_idx(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                             MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                             int_mv *mv_ref_list, int block, int mi_row,
+                             int mi_col, find_mv_refs_sync sync,
+                             void *const data, int16_t *mode_context,
+                             int_mv zeromv) {
+  const int *ref_sign_bias = cm->ref_frame_sign_bias;
+  int i, refmv_count = 0;
+#if !CONFIG_REF_MV
+  const POSITION *const mv_ref_search = mv_ref_blocks[mi->mbmi.sb_type];
+#endif
+  int different_ref_found = 0;
+  int context_counter = 0;
+#if CONFIG_MV_COMPRESS
+  const TileInfo *const tile_ = &xd->tile;
+  int mi_row_end = tile_->mi_row_end;
+  int mi_col_end = tile_->mi_col_end;
+  const MV_REF *const prev_frame_mvs =
+      cm->use_prev_frame_mvs
+          ? cm->prev_frame->mvs +
+                AOMMIN(((mi_row >> 1) << 1) + 1 + (((xd->n8_h - 1) >> 1) << 1),
+                       mi_row_end - 1) *
+                    cm->mi_cols +
+                AOMMIN(((mi_col >> 1) << 1) + 1 + (((xd->n8_w - 1) >> 1) << 1),
+                       mi_col_end - 1)
+          : NULL;
+#else
+  const MV_REF *const prev_frame_mvs =
+      cm->use_prev_frame_mvs
+          ? cm->prev_frame->mvs + mi_row * cm->mi_cols + mi_col
+          : NULL;
+#endif
+  const TileInfo *const tile = &xd->tile;
+  const BLOCK_SIZE bsize = mi->mbmi.sb_type;
+  const int bw = block_size_wide[AOMMAX(bsize, BLOCK_8X8)];
+  const int bh = block_size_high[AOMMAX(bsize, BLOCK_8X8)];
+#if CONFIG_REF_MV
+  POSITION mv_ref_search[MVREF_NEIGHBOURS];
+  const int num_8x8_blocks_wide = num_8x8_blocks_wide_lookup[bsize];
+  const int num_8x8_blocks_high = num_8x8_blocks_high_lookup[bsize];
+  mv_ref_search[0].row = num_8x8_blocks_high - 1;
+  mv_ref_search[0].col = -1;
+  mv_ref_search[1].row = -1;
+  mv_ref_search[1].col = num_8x8_blocks_wide - 1;
+  mv_ref_search[2].row = -1;
+  mv_ref_search[2].col = (num_8x8_blocks_wide - 1) >> 1;
+  mv_ref_search[3].row = (num_8x8_blocks_high - 1) >> 1;
+  mv_ref_search[3].col = -1;
+  mv_ref_search[4].row = -1;
+  mv_ref_search[4].col = -1;
+#if CONFIG_EXT_PARTITION_TYPES
+  if (num_8x8_blocks_wide == num_8x8_blocks_high) {
+    mv_ref_search[5].row = -1;
+    mv_ref_search[5].col = 0;
+    mv_ref_search[6].row = 0;
+    mv_ref_search[6].col = -1;
+  } else {
+    mv_ref_search[5].row = -1;
+    mv_ref_search[5].col = num_8x8_blocks_wide;
+    mv_ref_search[6].row = num_8x8_blocks_high;
+    mv_ref_search[6].col = -1;
+  }
+#else
+  mv_ref_search[5].row = -1;
+  mv_ref_search[5].col = num_8x8_blocks_wide;
+  mv_ref_search[6].row = num_8x8_blocks_high;
+  mv_ref_search[6].col = -1;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+  mv_ref_search[7].row = -1;
+  mv_ref_search[7].col = -3;
+  mv_ref_search[8].row = num_8x8_blocks_high - 1;
+  mv_ref_search[8].col = -3;
+
+#if CONFIG_CB4X4
+  for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
+    mv_ref_search[i].row *= 2;
+    mv_ref_search[i].col *= 2;
+  }
+#endif  // CONFIG_CB4X4
+#endif  // CONFIG_REF_MV
+
+  // The nearest 2 blocks are treated differently
+  // if the size < 8x8 we get the mv from the bmi substructure,
+  // and we also need to keep a mode count.
+  for (i = 0; i < 2; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, cm, mv_ref)) {
+      const MODE_INFO *const candidate_mi =
+          xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
+      const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
+      // Keep counts for entropy encoding.
+      context_counter += mode_2_counter[candidate->mode];
+      different_ref_found = 1;
+
+      if (candidate->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, block),
+                        refmv_count, mv_ref_list, bw, bh, xd, Done);
+      else if (candidate->ref_frame[1] == ref_frame)
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 1, mv_ref->col, block),
+                        refmv_count, mv_ref_list, bw, bh, xd, Done);
+    }
+  }
+
+  // Check the rest of the neighbors in much the same way
+  // as before except we don't need to keep track of sub blocks or
+  // mode counts.
+  for (; i < MVREF_NEIGHBOURS; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, cm, mv_ref)) {
+      const MB_MODE_INFO *const candidate =
+          !xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride]
+              ? NULL
+              : &xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride]->mbmi;
+#if CONFIG_REF_MV
+      if (candidate == NULL) continue;
+      if ((mi_row % MAX_MIB_SIZE) + mv_ref->row >= MAX_MIB_SIZE ||
+          (mi_col % MAX_MIB_SIZE) + mv_ref->col >= MAX_MIB_SIZE)
+        continue;
+#endif
+      different_ref_found = 1;
+
+      if (candidate->ref_frame[0] == ref_frame)
+        ADD_MV_REF_LIST(candidate->mv[0], refmv_count, mv_ref_list, bw, bh, xd,
+                        Done);
+      else if (candidate->ref_frame[1] == ref_frame)
+        ADD_MV_REF_LIST(candidate->mv[1], refmv_count, mv_ref_list, bw, bh, xd,
+                        Done);
+    }
+  }
+
+// TODO(hkuang): Remove this sync after fixing pthread_cond_broadcast
+// on windows platform. The sync here is unncessary if use_perv_frame_mvs
+// is 0. But after removing it, there will be hang in the unit test on windows
+// due to several threads waiting for a thread's signal.
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+  if (cm->frame_parallel_decode && sync != NULL) {
+    sync(data, mi_row);
+  }
+#endif
+
+  // Check the last frame's mode and mv info.
+  if (cm->use_prev_frame_mvs) {
+    // Synchronize here for frame parallel decode if sync function is provided.
+    if (cm->frame_parallel_decode && sync != NULL) {
+      sync(data, mi_row);
+    }
+
+    if (prev_frame_mvs->ref_frame[0] == ref_frame) {
+      ADD_MV_REF_LIST(prev_frame_mvs->mv[0], refmv_count, mv_ref_list, bw, bh,
+                      xd, Done);
+    } else if (prev_frame_mvs->ref_frame[1] == ref_frame) {
+      ADD_MV_REF_LIST(prev_frame_mvs->mv[1], refmv_count, mv_ref_list, bw, bh,
+                      xd, Done);
+    }
+  }
+
+  // Since we couldn't find 2 mvs from the same reference frame
+  // go back through the neighbors and find motion vectors from
+  // different reference frames.
+  if (different_ref_found) {
+    for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
+      const POSITION *mv_ref = &mv_ref_search[i];
+      if (is_inside(tile, mi_col, mi_row, cm->mi_rows, cm, mv_ref)) {
+        const MB_MODE_INFO *const candidate =
+            !xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride]
+                ? NULL
+                : &xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride]->mbmi;
+#if CONFIG_REF_MV
+        if (candidate == NULL) continue;
+        if ((mi_row % MAX_MIB_SIZE) + mv_ref->row >= MAX_MIB_SIZE ||
+            (mi_col % MAX_MIB_SIZE) + mv_ref->col >= MAX_MIB_SIZE)
+          continue;
+#endif
+
+        // If the candidate is INTRA we don't want to consider its mv.
+        IF_DIFF_REF_FRAME_ADD_MV(candidate, ref_frame, ref_sign_bias,
+                                 refmv_count, mv_ref_list, bw, bh, xd, Done);
+      }
+    }
+  }
+
+  // Since we still don't have a candidate we'll try the last frame.
+  if (cm->use_prev_frame_mvs) {
+    if (prev_frame_mvs->ref_frame[0] != ref_frame &&
+        prev_frame_mvs->ref_frame[0] > INTRA_FRAME) {
+      int_mv mv = prev_frame_mvs->mv[0];
+      if (ref_sign_bias[prev_frame_mvs->ref_frame[0]] !=
+          ref_sign_bias[ref_frame]) {
+        mv.as_mv.row *= -1;
+        mv.as_mv.col *= -1;
+      }
+      ADD_MV_REF_LIST(mv, refmv_count, mv_ref_list, bw, bh, xd, Done);
+    }
+
+    if (prev_frame_mvs->ref_frame[1] > INTRA_FRAME &&
+        prev_frame_mvs->ref_frame[1] != ref_frame) {
+      int_mv mv = prev_frame_mvs->mv[1];
+      if (ref_sign_bias[prev_frame_mvs->ref_frame[1]] !=
+          ref_sign_bias[ref_frame]) {
+        mv.as_mv.row *= -1;
+        mv.as_mv.col *= -1;
+      }
+      ADD_MV_REF_LIST(mv, refmv_count, mv_ref_list, bw, bh, xd, Done);
+    }
+  }
+
+Done:
+  if (mode_context)
+    mode_context[ref_frame] = counter_to_context[context_counter];
+  for (i = refmv_count; i < MAX_MV_REF_CANDIDATES; ++i)
+    mv_ref_list[i].as_int = zeromv.as_int;
+}
+
+#if CONFIG_EXT_INTER
+// This function keeps a mode count for a given MB/SB
+void av1_update_mv_context(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                           MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                           int_mv *mv_ref_list, int block, int mi_row,
+                           int mi_col, int16_t *mode_context) {
+  int i, refmv_count = 0;
+#if !CONFIG_REF_MV
+  const POSITION *const mv_ref_search = mv_ref_blocks[mi->mbmi.sb_type];
+#endif
+  int context_counter = 0;
+  const int bw = block_size_wide[mi->mbmi.sb_type];
+  const int bh = block_size_high[mi->mbmi.sb_type];
+  const TileInfo *const tile = &xd->tile;
+#if CONFIG_REF_MV
+  POSITION mv_ref_search[MVREF_NEIGHBOURS];
+  const int num_8x8_blocks_wide = mi_size_wide[mi->mbmi.sb_type];
+  const int num_8x8_blocks_high = mi_size_high[mi->mbmi.sb_type];
+  mv_ref_search[0].row = num_8x8_blocks_high - 1;
+  mv_ref_search[0].col = -1;
+  mv_ref_search[1].row = -1;
+  mv_ref_search[1].col = num_8x8_blocks_wide - 1;
+  mv_ref_search[2].row = -1;
+  mv_ref_search[2].col = (num_8x8_blocks_wide - 1) >> 1;
+  mv_ref_search[3].row = (num_8x8_blocks_high - 1) >> 1;
+  mv_ref_search[3].col = -1;
+  mv_ref_search[4].row = -1;
+  mv_ref_search[4].col = -1;
+#if CONFIG_EXT_PARTITION_TYPES
+  if (num_8x8_blocks_wide == num_8x8_blocks_high) {
+    mv_ref_search[5].row = -1;
+    mv_ref_search[5].col = 0;
+    mv_ref_search[6].row = 0;
+    mv_ref_search[6].col = -1;
+  } else {
+    mv_ref_search[5].row = -1;
+    mv_ref_search[5].col = num_8x8_blocks_wide;
+    mv_ref_search[6].row = num_8x8_blocks_high;
+    mv_ref_search[6].col = -1;
+  }
+#else
+  mv_ref_search[5].row = -1;
+  mv_ref_search[5].col = num_8x8_blocks_wide;
+  mv_ref_search[6].row = num_8x8_blocks_high;
+  mv_ref_search[6].col = -1;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+  mv_ref_search[7].row = -1;
+  mv_ref_search[7].col = -3;
+  mv_ref_search[8].row = num_8x8_blocks_high - 1;
+  mv_ref_search[8].col = -3;
+#endif
+
+  // Blank the reference vector list
+  memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
+
+  // The nearest 2 blocks are examined only.
+  // If the size < 8x8, we get the mv from the bmi substructure;
+  for (i = 0; i < 2; ++i) {
+    const POSITION *const mv_ref = &mv_ref_search[i];
+    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, cm, mv_ref)) {
+      const MODE_INFO *const candidate_mi =
+          xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
+      const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
+
+      // Keep counts for entropy encoding.
+      context_counter += mode_2_counter[candidate->mode];
+
+      if (candidate->ref_frame[0] == ref_frame) {
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, block),
+                        refmv_count, mv_ref_list, bw, bh, xd, Done);
+      } else if (candidate->ref_frame[1] == ref_frame) {
+        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 1, mv_ref->col, block),
+                        refmv_count, mv_ref_list, bw, bh, xd, Done);
+      }
+    }
+  }
+
+Done:
+
+  if (mode_context)
+    mode_context[ref_frame] = counter_to_context[context_counter];
+}
+#endif  // CONFIG_EXT_INTER
+
+void av1_find_mv_refs(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                      MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+#if CONFIG_REF_MV
+                      uint8_t *ref_mv_count, CANDIDATE_MV *ref_mv_stack,
+#if CONFIG_EXT_INTER
+                      int16_t *compound_mode_context,
+#endif  // CONFIG_EXT_INTER
+#endif
+                      int_mv *mv_ref_list, int mi_row, int mi_col,
+                      find_mv_refs_sync sync, void *const data,
+                      int16_t *mode_context) {
+  int_mv zeromv[2];
+#if CONFIG_GLOBAL_MOTION
+  BLOCK_SIZE bsize = mi->mbmi.sb_type;
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_REF_MV
+  int idx, all_zero = 1;
+#if CONFIG_GLOBAL_MOTION
+  MV_REFERENCE_FRAME rf[2];
+#endif  // CONFIG_GLOBAL_MOTION
+#endif  // CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+  av1_update_mv_context(cm, xd, mi, ref_frame, mv_ref_list, -1, mi_row, mi_col,
+#if CONFIG_REF_MV
+                        compound_mode_context);
+#else
+                        mode_context);
+#endif  // CONFIG_REF_MV
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_GLOBAL_MOTION
+#if CONFIG_REF_MV
+  av1_set_ref_frame(rf, ref_frame);
+  zeromv[0].as_int = gm_get_motion_vector(&cm->global_motion[rf[0]],
+                                          cm->allow_high_precision_mv, bsize,
+                                          mi_col, mi_row, 0)
+                         .as_int;
+  zeromv[1].as_int = (rf[1] != NONE_FRAME)
+                         ? gm_get_motion_vector(&cm->global_motion[rf[1]],
+                                                cm->allow_high_precision_mv,
+                                                bsize, mi_col, mi_row, 0)
+                               .as_int
+                         : 0;
+#else
+  zeromv[0].as_int = gm_get_motion_vector(&cm->global_motion[ref_frame],
+                                          cm->allow_high_precision_mv, bsize,
+                                          mi_col, mi_row, 0)
+                         .as_int;
+  zeromv[1].as_int = 0;
+#endif  // CONFIG_REF_MV
+#else
+  zeromv[0].as_int = zeromv[1].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+
+#if CONFIG_REF_MV
+  if (ref_frame <= ALTREF_FRAME)
+#endif  // CONFIG_REF_MV
+    find_mv_refs_idx(cm, xd, mi, ref_frame, mv_ref_list, -1, mi_row, mi_col,
+                     sync, data, mode_context, zeromv[0]);
+
+#if CONFIG_REF_MV
+  setup_ref_mv_list(cm, xd, ref_frame, ref_mv_count, ref_mv_stack, mv_ref_list,
+                    -1, mi_row, mi_col, mode_context);
+  /* Note: If global motion is enabled, then we want to set the ALL_ZERO flag
+     iff all of the MVs we could generate with NEARMV/NEARESTMV are equivalent
+     to the global motion vector.
+     Note: For the following to work properly, the encoder can't throw away
+     any global motion models after calling this function, even if they are
+     unused. Instead we rely on the recode loop: If any non-IDENTITY model
+     is unused, the whole frame will be re-encoded without it.
+     The problem is that, otherwise, we can end up in the following situation:
+     * Encoder has a global motion model with nonzero translational part,
+       and all candidate MVs are zero. So the ALL_ZERO flag is unset.
+     * Encoder throws away global motion because it is never used.
+     * Decoder sees that there is no global motion and all candidate MVs are
+       zero, so sets the ALL_ZERO flag.
+     * This leads to an encode/decode mismatch.
+  */
+  if (*ref_mv_count >= 2) {
+    for (idx = 0; idx < AOMMIN(3, *ref_mv_count); ++idx) {
+      if (ref_mv_stack[idx].this_mv.as_int != zeromv[0].as_int) all_zero = 0;
+      if (ref_frame > ALTREF_FRAME)
+        if (ref_mv_stack[idx].comp_mv.as_int != zeromv[1].as_int) all_zero = 0;
+    }
+  } else if (ref_frame <= ALTREF_FRAME) {
+    for (idx = 0; idx < MAX_MV_REF_CANDIDATES; ++idx)
+      if (mv_ref_list[idx].as_int != zeromv[0].as_int) all_zero = 0;
+  }
+
+  if (all_zero) mode_context[ref_frame] |= (1 << ALL_ZERO_FLAG_OFFSET);
+#endif
+}
+
+void av1_find_best_ref_mvs(int allow_hp, int_mv *mvlist, int_mv *nearest_mv,
+                           int_mv *near_mv) {
+  int i;
+  // Make sure all the candidates are properly clamped etc
+  for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i) {
+    lower_mv_precision(&mvlist[i].as_mv, allow_hp);
+  }
+  *nearest_mv = mvlist[0];
+  *near_mv = mvlist[1];
+}
+
+void av1_append_sub8x8_mvs_for_idx(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                   int block, int ref, int mi_row, int mi_col,
+#if CONFIG_REF_MV
+                                   CANDIDATE_MV *ref_mv_stack,
+                                   uint8_t *ref_mv_count,
+#endif
+#if CONFIG_EXT_INTER
+                                   int_mv *mv_list,
+#endif  // CONFIG_EXT_INTER
+                                   int_mv *nearest_mv, int_mv *near_mv) {
+#if !CONFIG_EXT_INTER
+  int_mv mv_list[MAX_MV_REF_CANDIDATES];
+#endif  // !CONFIG_EXT_INTER
+  MODE_INFO *const mi = xd->mi[0];
+  b_mode_info *bmi = mi->bmi;
+  int n;
+  int_mv zeromv;
+#if CONFIG_REF_MV
+  CANDIDATE_MV tmp_mv;
+  uint8_t idx;
+  uint8_t above_count = 0, left_count = 0;
+  MV_REFERENCE_FRAME rf[2] = { mi->mbmi.ref_frame[ref], NONE_FRAME };
+  *ref_mv_count = 0;
+#endif
+
+  assert(MAX_MV_REF_CANDIDATES == 2);
+
+#if CONFIG_GLOBAL_MOTION
+  zeromv.as_int =
+      gm_get_motion_vector(&cm->global_motion[ref], cm->allow_high_precision_mv,
+                           mi->mbmi.sb_type, mi_col, mi_row, block)
+          .as_int;
+#else
+  zeromv.as_int = 0;
+#endif
+  find_mv_refs_idx(cm, xd, mi, mi->mbmi.ref_frame[ref], mv_list, block, mi_row,
+                   mi_col, NULL, NULL, NULL, zeromv);
+
+#if CONFIG_REF_MV
+  scan_blk_mbmi(cm, xd, mi_row, mi_col, block, rf, -1, 0, ref_mv_stack,
+                ref_mv_count);
+  above_count = *ref_mv_count;
+
+  scan_blk_mbmi(cm, xd, mi_row, mi_col, block, rf, 0, -1, ref_mv_stack,
+                ref_mv_count);
+  left_count = *ref_mv_count - above_count;
+
+  if (above_count > 1 && left_count > 0) {
+    tmp_mv = ref_mv_stack[1];
+    ref_mv_stack[1] = ref_mv_stack[above_count];
+    ref_mv_stack[above_count] = tmp_mv;
+  }
+
+  for (idx = 0; idx < *ref_mv_count; ++idx)
+    clamp_mv_ref(&ref_mv_stack[idx].this_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                 xd->n8_h << MI_SIZE_LOG2, xd);
+
+  for (idx = 0; idx < AOMMIN(MAX_MV_REF_CANDIDATES, *ref_mv_count); ++idx)
+    mv_list[idx].as_int = ref_mv_stack[idx].this_mv.as_int;
+#endif
+
+  near_mv->as_int = 0;
+  switch (block) {
+    case 0:
+      nearest_mv->as_int = mv_list[0].as_int;
+      near_mv->as_int = mv_list[1].as_int;
+      break;
+    case 1:
+    case 2:
+      nearest_mv->as_int = bmi[0].as_mv[ref].as_int;
+      for (n = 0; n < MAX_MV_REF_CANDIDATES; ++n)
+        if (nearest_mv->as_int != mv_list[n].as_int) {
+          near_mv->as_int = mv_list[n].as_int;
+          break;
+        }
+      break;
+    case 3: {
+      int_mv candidates[2 + MAX_MV_REF_CANDIDATES];
+      candidates[0] = bmi[1].as_mv[ref];
+      candidates[1] = bmi[0].as_mv[ref];
+      candidates[2] = mv_list[0];
+      candidates[3] = mv_list[1];
+
+      nearest_mv->as_int = bmi[2].as_mv[ref].as_int;
+      for (n = 0; n < 2 + MAX_MV_REF_CANDIDATES; ++n)
+        if (nearest_mv->as_int != candidates[n].as_int) {
+          near_mv->as_int = candidates[n].as_int;
+          break;
+        }
+      break;
+    }
+    default: assert(0 && "Invalid block index.");
+  }
+}
+
+#if CONFIG_WARPED_MOTION
+void calc_projection_samples(MB_MODE_INFO *const mbmi, int x, int y,
+                             int *pts_inref) {
+  pts_inref[0] = (x * 8) + mbmi->mv[0].as_mv.col;
+  pts_inref[1] = (y * 8) + mbmi->mv[0].as_mv.row;
+}
+
+// Note: Samples returned are at 1/8-pel precision
+int findSamples(const AV1_COMMON *cm, MACROBLOCKD *xd, int mi_row, int mi_col,
+                int *pts, int *pts_inref) {
+  MB_MODE_INFO *const mbmi0 = &(xd->mi[0]->mbmi);
+  int ref_frame = mbmi0->ref_frame[0];
+  int up_available = xd->up_available;
+  int left_available = xd->left_available;
+  int i, mi_step, np = 0;
+  int global_offset_c = mi_col * MI_SIZE;
+  int global_offset_r = mi_row * MI_SIZE;
+
+  // scan the above row
+  if (up_available) {
+    for (i = 0; i < AOMMIN(xd->n8_w, cm->mi_cols - mi_col); i += mi_step) {
+      int mi_row_offset = -1;
+      int mi_col_offset = i;
+
+      MODE_INFO *mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+      MB_MODE_INFO *mbmi = &mi->mbmi;
+
+      mi_step = AOMMIN(xd->n8_w, mi_size_wide[mbmi->sb_type]);
+
+      if (mbmi->ref_frame[0] == ref_frame && mbmi->ref_frame[1] == NONE_FRAME) {
+        int bw = block_size_wide[mbmi->sb_type];
+        int bh = block_size_high[mbmi->sb_type];
+        int cr_offset = -AOMMAX(bh, MI_SIZE) / 2 - 1;
+        int cc_offset = i * MI_SIZE + AOMMAX(bw, MI_SIZE) / 2 - 1;
+        int x = cc_offset + global_offset_c;
+        int y = cr_offset + global_offset_r;
+
+        pts[0] = (x * 8);
+        pts[1] = (y * 8);
+        calc_projection_samples(mbmi, x, y, pts_inref);
+        pts += 2;
+        pts_inref += 2;
+        np++;
+        if (np >= LEAST_SQUARES_SAMPLES_MAX) return LEAST_SQUARES_SAMPLES_MAX;
+      }
+    }
+  }
+  assert(2 * np <= SAMPLES_ARRAY_SIZE);
+
+  // scan the left column
+  if (left_available) {
+    for (i = 0; i < AOMMIN(xd->n8_h, cm->mi_rows - mi_row); i += mi_step) {
+      int mi_row_offset = i;
+      int mi_col_offset = -1;
+
+      MODE_INFO *mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+      MB_MODE_INFO *mbmi = &mi->mbmi;
+
+      mi_step = AOMMIN(xd->n8_h, mi_size_high[mbmi->sb_type]);
+
+      if (mbmi->ref_frame[0] == ref_frame && mbmi->ref_frame[1] == NONE_FRAME) {
+        int bw = block_size_wide[mbmi->sb_type];
+        int bh = block_size_high[mbmi->sb_type];
+        int cr_offset = i * MI_SIZE + AOMMAX(bh, MI_SIZE) / 2 - 1;
+        int cc_offset = -AOMMAX(bw, MI_SIZE) / 2 - 1;
+        int x = cc_offset + global_offset_c;
+        int y = cr_offset + global_offset_r;
+
+        pts[0] = (x * 8);
+        pts[1] = (y * 8);
+        calc_projection_samples(mbmi, x, y, pts_inref);
+        pts += 2;
+        pts_inref += 2;
+        np++;
+        if (np >= LEAST_SQUARES_SAMPLES_MAX) return LEAST_SQUARES_SAMPLES_MAX;
+      }
+    }
+  }
+  assert(2 * np <= SAMPLES_ARRAY_SIZE);
+
+  if (left_available && up_available) {
+    int mi_row_offset = -1;
+    int mi_col_offset = -1;
+
+    MODE_INFO *mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+    MB_MODE_INFO *mbmi = &mi->mbmi;
+
+    if (mbmi->ref_frame[0] == ref_frame && mbmi->ref_frame[1] == NONE_FRAME) {
+      int bw = block_size_wide[mbmi->sb_type];
+      int bh = block_size_high[mbmi->sb_type];
+      int cr_offset = -AOMMAX(bh, MI_SIZE) / 2 - 1;
+      int cc_offset = -AOMMAX(bw, MI_SIZE) / 2 - 1;
+      int x = cc_offset + global_offset_c;
+      int y = cr_offset + global_offset_r;
+
+      pts[0] = (x * 8);
+      pts[1] = (y * 8);
+      calc_projection_samples(mbmi, x, y, pts_inref);
+      np++;
+    }
+  }
+  assert(2 * np <= SAMPLES_ARRAY_SIZE);
+
+  return np;
+}
+#endif  // CONFIG_WARPED_MOTION
diff --git a/third_party/aom/av1/common/mvref_common.h b/third_party/aom/av1/common/mvref_common.h
new file mode 100644
index 0000000000..01f74b77a7
--- /dev/null
+++ b/third_party/aom/av1/common/mvref_common.h
@@ -0,0 +1,580 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AV1_COMMON_MVREF_COMMON_H_
+#define AV1_COMMON_MVREF_COMMON_H_
+
+#include "av1/common/onyxc_int.h"
+#include "av1/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_REF_MV
+#define MVREF_NEIGHBOURS 9
+#else
+#define MVREF_NEIGHBOURS 8
+#endif
+
+typedef struct position {
+  int row;
+  int col;
+} POSITION;
+
+typedef enum {
+  BOTH_ZERO = 0,
+  ZERO_PLUS_PREDICTED = 1,
+  BOTH_PREDICTED = 2,
+  NEW_PLUS_NON_INTRA = 3,
+  BOTH_NEW = 4,
+  INTRA_PLUS_NON_INTRA = 5,
+  BOTH_INTRA = 6,
+  INVALID_CASE = 9
+} motion_vector_context;
+
+// This is used to figure out a context for the ref blocks. The code flattens
+// an array that would have 3 possible counts (0, 1 & 2) for 3 choices by
+// adding 9 for each intra block, 3 for each zero mv and 1 for each new
+// motion vector. This single number is then converted into a context
+// with a single lookup ( counter_to_context ).
+static const int mode_2_counter[MB_MODE_COUNT] = {
+  9,  // DC_PRED
+  9,  // V_PRED
+  9,  // H_PRED
+  9,  // D45_PRED
+  9,  // D135_PRED
+  9,  // D117_PRED
+  9,  // D153_PRED
+  9,  // D207_PRED
+  9,  // D63_PRED
+#if CONFIG_ALT_INTRA
+  9,    // SMOOTH_PRED
+#endif  // CONFIG_ALT_INTRA
+  9,    // TM_PRED
+  0,    // NEARESTMV
+  0,    // NEARMV
+  3,    // ZEROMV
+  1,    // NEWMV
+#if CONFIG_EXT_INTER
+#if CONFIG_COMPOUND_SINGLEREF
+  0,    // SR_NEAREST_NEARMV
+  1,    // SR_NEAREST_NEWMV
+  1,    // SR_NEAR_NEWMV
+  3,    // SR_ZERO_NEWMV
+  1,    // SR_NEW_NEWMV
+#endif  // CONFIG_COMPOUND_SINGLEREF
+  0,    // NEAREST_NEARESTMV
+  0,    // NEAREST_NEARMV
+  0,    // NEAR_NEARESTMV
+  0,    // NEAR_NEARMV
+  1,    // NEAREST_NEWMV
+  1,    // NEW_NEARESTMV
+  1,    // NEAR_NEWMV
+  1,    // NEW_NEARMV
+  3,    // ZERO_ZEROMV
+  1,    // NEW_NEWMV
+#endif  // CONFIG_EXT_INTER
+};
+
+// There are 3^3 different combinations of 3 counts that can be either 0,1 or
+// 2. However the actual count can never be greater than 2 so the highest
+// counter we need is 18. 9 is an invalid counter that's never used.
+static const int counter_to_context[19] = {
+  BOTH_PREDICTED,        // 0
+  NEW_PLUS_NON_INTRA,    // 1
+  BOTH_NEW,              // 2
+  ZERO_PLUS_PREDICTED,   // 3
+  NEW_PLUS_NON_INTRA,    // 4
+  INVALID_CASE,          // 5
+  BOTH_ZERO,             // 6
+  INVALID_CASE,          // 7
+  INVALID_CASE,          // 8
+  INTRA_PLUS_NON_INTRA,  // 9
+  INTRA_PLUS_NON_INTRA,  // 10
+  INVALID_CASE,          // 11
+  INTRA_PLUS_NON_INTRA,  // 12
+  INVALID_CASE,          // 13
+  INVALID_CASE,          // 14
+  INVALID_CASE,          // 15
+  INVALID_CASE,          // 16
+  INVALID_CASE,          // 17
+  BOTH_INTRA             // 18
+};
+
+#if !CONFIG_REF_MV
+static const POSITION mv_ref_blocks[BLOCK_SIZES][MVREF_NEIGHBOURS] = {
+  // 4X4
+  { { -1, 0 },
+    { 0, -1 },
+    { -1, -1 },
+    { -2, 0 },
+    { 0, -2 },
+    { -2, -1 },
+    { -1, -2 },
+    { -2, -2 } },
+  // 4X8
+  { { -1, 0 },
+    { 0, -1 },
+    { -1, -1 },
+    { -2, 0 },
+    { 0, -2 },
+    { -2, -1 },
+    { -1, -2 },
+    { -2, -2 } },
+  // 8X4
+  { { -1, 0 },
+    { 0, -1 },
+    { -1, -1 },
+    { -2, 0 },
+    { 0, -2 },
+    { -2, -1 },
+    { -1, -2 },
+    { -2, -2 } },
+  // 8X8
+  { { -1, 0 },
+    { 0, -1 },
+    { -1, -1 },
+    { -2, 0 },
+    { 0, -2 },
+    { -2, -1 },
+    { -1, -2 },
+    { -2, -2 } },
+  // 8X16
+  { { 0, -1 },
+    { -1, 0 },
+    { 1, -1 },
+    { -1, -1 },
+    { 0, -2 },
+    { -2, 0 },
+    { -2, -1 },
+    { -1, -2 } },
+  // 16X8
+  { { -1, 0 },
+    { 0, -1 },
+    { -1, 1 },
+    { -1, -1 },
+    { -2, 0 },
+    { 0, -2 },
+    { -1, -2 },
+    { -2, -1 } },
+  // 16X16
+  { { -1, 0 },
+    { 0, -1 },
+    { -1, 1 },
+    { 1, -1 },
+    { -1, -1 },
+    { -3, 0 },
+    { 0, -3 },
+    { -3, -3 } },
+  // 16X32
+  { { 0, -1 },
+    { -1, 0 },
+    { 2, -1 },
+    { -1, -1 },
+    { -1, 1 },
+    { 0, -3 },
+    { -3, 0 },
+    { -3, -3 } },
+  // 32X16
+  { { -1, 0 },
+    { 0, -1 },
+    { -1, 2 },
+    { -1, -1 },
+    { 1, -1 },
+    { -3, 0 },
+    { 0, -3 },
+    { -3, -3 } },
+  // 32X32
+  { { -1, 1 },
+    { 1, -1 },
+    { -1, 2 },
+    { 2, -1 },
+    { -1, -1 },
+    { -3, 0 },
+    { 0, -3 },
+    { -3, -3 } },
+  // 32X64
+  { { 0, -1 },
+    { -1, 0 },
+    { 4, -1 },
+    { -1, 2 },
+    { -1, -1 },
+    { 0, -3 },
+    { -3, 0 },
+    { 2, -1 } },
+  // 64X32
+  { { -1, 0 },
+    { 0, -1 },
+    { -1, 4 },
+    { 2, -1 },
+    { -1, -1 },
+    { -3, 0 },
+    { 0, -3 },
+    { -1, 2 } },
+  // 64X64
+  { { -1, 3 },
+    { 3, -1 },
+    { -1, 4 },
+    { 4, -1 },
+    { -1, -1 },
+    { -1, 0 },
+    { 0, -1 },
+    { -1, 6 } },
+#if CONFIG_EXT_PARTITION
+  // TODO(debargha/jingning) Making them twice the 32x64, .. ones above
+  // 64x128
+  { { 0, -2 },
+    { -2, 0 },
+    { 8, -2 },
+    { -2, 4 },
+    { -2, -2 },
+    { 0, -6 },
+    { -6, 0 },
+    { 4, -2 } },
+  // 128x64
+  { { -2, 0 },
+    { 0, -2 },
+    { -2, 8 },
+    { 4, -2 },
+    { -2, -2 },
+    { -6, 0 },
+    { 0, -6 },
+    { -2, 4 } },
+  // 128x128
+  { { -2, 6 },
+    { 6, -2 },
+    { -2, 8 },
+    { 8, -2 },
+    { -2, -2 },
+    { -2, 0 },
+    { 0, -2 },
+    { -2, 12 } },
+#endif  // CONFIG_EXT_PARTITION
+};
+#endif
+
+static const int idx_n_column_to_subblock[4][2] = {
+  { 1, 2 }, { 1, 3 }, { 3, 2 }, { 3, 3 }
+};
+
+// clamp_mv_ref
+#if CONFIG_EXT_PARTITION
+#define MV_BORDER (16 << 3)  // Allow 16 pels in 1/8th pel units
+#else
+#define MV_BORDER (8 << 3)  // Allow 8 pels in 1/8th pel units
+#endif                      // CONFIG_EXT_PARTITION
+
+static INLINE void clamp_mv_ref(MV *mv, int bw, int bh, const MACROBLOCKD *xd) {
+  clamp_mv(mv, xd->mb_to_left_edge - bw * 8 - MV_BORDER,
+           xd->mb_to_right_edge + bw * 8 + MV_BORDER,
+           xd->mb_to_top_edge - bh * 8 - MV_BORDER,
+           xd->mb_to_bottom_edge + bh * 8 + MV_BORDER);
+}
+
+// This function returns either the appropriate sub block or block's mv
+// on whether the block_size < 8x8 and we have check_sub_blocks set.
+static INLINE int_mv get_sub_block_mv(const MODE_INFO *candidate, int which_mv,
+                                      int search_col, int block_idx) {
+#if CONFIG_REF_MV
+  (void)search_col;
+  (void)block_idx;
+  return candidate->mbmi.mv[which_mv];
+#else
+  return block_idx >= 0 && candidate->mbmi.sb_type < BLOCK_8X8
+             ? candidate
+                   ->bmi[idx_n_column_to_subblock[block_idx][search_col == 0]]
+                   .as_mv[which_mv]
+             : candidate->mbmi.mv[which_mv];
+#endif
+}
+
+#if CONFIG_REF_MV
+static INLINE int_mv get_sub_block_pred_mv(const MODE_INFO *candidate,
+                                           int which_mv, int search_col,
+                                           int block_idx) {
+  (void)search_col;
+  (void)block_idx;
+  return candidate->mbmi.mv[which_mv];
+}
+#endif
+
+// Performs mv sign inversion if indicated by the reference frame combination.
+static INLINE int_mv scale_mv(const MB_MODE_INFO *mbmi, int ref,
+                              const MV_REFERENCE_FRAME this_ref_frame,
+                              const int *ref_sign_bias) {
+  int_mv mv = mbmi->mv[ref];
+  if (ref_sign_bias[mbmi->ref_frame[ref]] != ref_sign_bias[this_ref_frame]) {
+    mv.as_mv.row *= -1;
+    mv.as_mv.col *= -1;
+  }
+  return mv;
+}
+
+#define CLIP_IN_ADD(mv, bw, bh, xd) clamp_mv_ref(mv, bw, bh, xd)
+
+// This macro is used to add a motion vector mv_ref list if it isn't
+// already in the list.  If it's the second motion vector it will also
+// skip all additional processing and jump to done!
+#define ADD_MV_REF_LIST(mv, refmv_count, mv_ref_list, bw, bh, xd, Done)      \
+  do {                                                                       \
+    (mv_ref_list)[(refmv_count)] = (mv);                                     \
+    CLIP_IN_ADD(&(mv_ref_list)[(refmv_count)].as_mv, (bw), (bh), (xd));      \
+    if (refmv_count && (mv_ref_list)[1].as_int != (mv_ref_list)[0].as_int) { \
+      (refmv_count) = 2;                                                     \
+      goto Done;                                                             \
+    }                                                                        \
+    (refmv_count) = 1;                                                       \
+  } while (0)
+
+// If either reference frame is different, not INTRA, and they
+// are different from each other scale and add the mv to our list.
+#define IF_DIFF_REF_FRAME_ADD_MV(mbmi, ref_frame, ref_sign_bias, refmv_count, \
+                                 mv_ref_list, bw, bh, xd, Done)               \
+  do {                                                                        \
+    if (is_inter_block(mbmi)) {                                               \
+      if ((mbmi)->ref_frame[0] != ref_frame)                                  \
+        ADD_MV_REF_LIST(scale_mv((mbmi), 0, ref_frame, ref_sign_bias),        \
+                        refmv_count, mv_ref_list, bw, bh, xd, Done);          \
+      if (has_second_ref(mbmi) && (mbmi)->ref_frame[1] != ref_frame)          \
+        ADD_MV_REF_LIST(scale_mv((mbmi), 1, ref_frame, ref_sign_bias),        \
+                        refmv_count, mv_ref_list, bw, bh, xd, Done);          \
+    }                                                                         \
+  } while (0)
+
+// Checks that the given mi_row, mi_col and search point
+// are inside the borders of the tile.
+static INLINE int is_inside(const TileInfo *const tile, int mi_col, int mi_row,
+                            int mi_rows, const AV1_COMMON *cm,
+                            const POSITION *mi_pos) {
+#if CONFIG_DEPENDENT_HORZTILES
+  const int dependent_horz_tile_flag = cm->dependent_horz_tiles;
+#else
+  const int dependent_horz_tile_flag = 0;
+  (void)cm;
+#endif
+#if CONFIG_TILE_GROUPS
+  if (dependent_horz_tile_flag && !tile->tg_horz_boundary) {
+#else
+  if (dependent_horz_tile_flag) {
+#endif
+    return !(mi_row + mi_pos->row < 0 ||
+             mi_col + mi_pos->col < tile->mi_col_start ||
+             mi_row + mi_pos->row >= mi_rows ||
+             mi_col + mi_pos->col >= tile->mi_col_end);
+  } else {
+    return !(mi_row + mi_pos->row < tile->mi_row_start ||
+             mi_col + mi_pos->col < tile->mi_col_start ||
+             mi_row + mi_pos->row >= tile->mi_row_end ||
+             mi_col + mi_pos->col >= tile->mi_col_end);
+  }
+}
+
+static INLINE void lower_mv_precision(MV *mv, int allow_hp) {
+  if (!allow_hp) {
+    if (mv->row & 1) mv->row += (mv->row > 0 ? -1 : 1);
+    if (mv->col & 1) mv->col += (mv->col > 0 ? -1 : 1);
+  }
+}
+
+#if CONFIG_REF_MV
+static INLINE uint8_t av1_get_pred_diff_ctx(const int_mv pred_mv,
+                                            const int_mv this_mv) {
+  if (abs(this_mv.as_mv.row - pred_mv.as_mv.row) <= 4 &&
+      abs(this_mv.as_mv.col - pred_mv.as_mv.col) <= 4)
+    return 2;
+  else
+    return 1;
+}
+
+static INLINE int av1_nmv_ctx(const uint8_t ref_mv_count,
+                              const CANDIDATE_MV *ref_mv_stack, int ref,
+                              int ref_mv_idx) {
+  if (ref_mv_stack[ref_mv_idx].weight >= REF_CAT_LEVEL && ref_mv_count > 0)
+    return ref_mv_stack[ref_mv_idx].pred_diff[ref];
+
+  return 0;
+}
+
+static INLINE int8_t av1_ref_frame_type(const MV_REFERENCE_FRAME *const rf) {
+  if (rf[1] > INTRA_FRAME) {
+    return TOTAL_REFS_PER_FRAME + FWD_RF_OFFSET(rf[0]) +
+           BWD_RF_OFFSET(rf[1]) * FWD_REFS;
+  }
+
+  return rf[0];
+}
+
+// clang-format off
+static MV_REFERENCE_FRAME ref_frame_map[COMP_REFS][2] = {
+#if CONFIG_EXT_REFS
+  { LAST_FRAME, BWDREF_FRAME },  { LAST2_FRAME, BWDREF_FRAME },
+  { LAST3_FRAME, BWDREF_FRAME }, { GOLDEN_FRAME, BWDREF_FRAME },
+
+  { LAST_FRAME, ALTREF_FRAME },  { LAST2_FRAME, ALTREF_FRAME },
+  { LAST3_FRAME, ALTREF_FRAME }, { GOLDEN_FRAME, ALTREF_FRAME }
+#else
+  { LAST_FRAME, ALTREF_FRAME }, { GOLDEN_FRAME, ALTREF_FRAME }
+#endif
+};
+// clang-format on
+
+static INLINE void av1_set_ref_frame(MV_REFERENCE_FRAME *rf,
+                                     int8_t ref_frame_type) {
+  if (ref_frame_type >= TOTAL_REFS_PER_FRAME) {
+    rf[0] = ref_frame_map[ref_frame_type - TOTAL_REFS_PER_FRAME][0];
+    rf[1] = ref_frame_map[ref_frame_type - TOTAL_REFS_PER_FRAME][1];
+  } else {
+    rf[0] = ref_frame_type;
+    rf[1] = NONE_FRAME;
+    assert(ref_frame_type > INTRA_FRAME &&
+           ref_frame_type < TOTAL_REFS_PER_FRAME);
+  }
+}
+
+static INLINE int16_t av1_mode_context_analyzer(
+    const int16_t *const mode_context, const MV_REFERENCE_FRAME *const rf,
+    BLOCK_SIZE bsize, int block) {
+  int16_t mode_ctx = 0;
+  int8_t ref_frame_type = av1_ref_frame_type(rf);
+
+  if (block >= 0) {
+    mode_ctx = mode_context[rf[0]] & 0x00ff;
+#if !CONFIG_CB4X4
+    if (block > 0 && bsize < BLOCK_8X8 && bsize > BLOCK_4X4)
+      mode_ctx |= (1 << SKIP_NEARESTMV_SUB8X8_OFFSET);
+#else
+    (void)block;
+    (void)bsize;
+#endif
+
+    return mode_ctx;
+  }
+
+  return mode_context[ref_frame_type];
+}
+
+static INLINE uint8_t av1_drl_ctx(const CANDIDATE_MV *ref_mv_stack,
+                                  int ref_idx) {
+  if (ref_mv_stack[ref_idx].weight >= REF_CAT_LEVEL &&
+      ref_mv_stack[ref_idx + 1].weight >= REF_CAT_LEVEL)
+    return 0;
+
+  if (ref_mv_stack[ref_idx].weight >= REF_CAT_LEVEL &&
+      ref_mv_stack[ref_idx + 1].weight < REF_CAT_LEVEL)
+    return 2;
+
+  if (ref_mv_stack[ref_idx].weight < REF_CAT_LEVEL &&
+      ref_mv_stack[ref_idx + 1].weight < REF_CAT_LEVEL)
+    return 3;
+
+  return 0;
+}
+#endif
+
+typedef void (*find_mv_refs_sync)(void *const data, int mi_row);
+void av1_find_mv_refs(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                      MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+#if CONFIG_REF_MV
+                      uint8_t *ref_mv_count, CANDIDATE_MV *ref_mv_stack,
+#if CONFIG_EXT_INTER
+                      int16_t *compound_mode_context,
+#endif  // CONFIG_EXT_INTER
+#endif
+                      int_mv *mv_ref_list, int mi_row, int mi_col,
+                      find_mv_refs_sync sync, void *const data,
+                      int16_t *mode_context);
+
+// check a list of motion vectors by sad score using a number rows of pixels
+// above and a number cols of pixels in the left to select the one with best
+// score to use as ref motion vector
+void av1_find_best_ref_mvs(int allow_hp, int_mv *mvlist, int_mv *nearest_mv,
+                           int_mv *near_mv);
+
+void av1_append_sub8x8_mvs_for_idx(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                   int block, int ref, int mi_row, int mi_col,
+#if CONFIG_REF_MV
+                                   CANDIDATE_MV *ref_mv_stack,
+                                   uint8_t *ref_mv_count,
+#endif
+#if CONFIG_EXT_INTER
+                                   int_mv *mv_list,
+#endif  // CONFIG_EXT_INTER
+                                   int_mv *nearest_mv, int_mv *near_mv);
+
+#if CONFIG_EXT_INTER
+// This function keeps a mode count for a given MB/SB
+void av1_update_mv_context(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                           MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
+                           int_mv *mv_ref_list, int block, int mi_row,
+                           int mi_col, int16_t *mode_context);
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_WARPED_MOTION
+int findSamples(const AV1_COMMON *cm, MACROBLOCKD *xd, int mi_row, int mi_col,
+                int *pts, int *pts_inref);
+#endif  // CONFIG_WARPED_MOTION
+
+#if CONFIG_INTRABC
+static INLINE void av1_find_ref_dv(int_mv *ref_dv, int mi_row, int mi_col) {
+  // TODO(aconverse@google.com): Handle tiles and such
+  (void)mi_col;
+  if (mi_row < MAX_MIB_SIZE) {
+    ref_dv->as_mv.row = 0;
+    ref_dv->as_mv.col = -MI_SIZE * MAX_MIB_SIZE;
+  } else {
+    ref_dv->as_mv.row = -MI_SIZE * MAX_MIB_SIZE;
+    ref_dv->as_mv.col = 0;
+  }
+}
+
+static INLINE int is_dv_valid(const MV dv, const TileInfo *const tile,
+                              int mi_row, int mi_col, BLOCK_SIZE bsize) {
+  const int bw = block_size_wide[bsize];
+  const int bh = block_size_high[bsize];
+  const int SCALE_PX_TO_MV = 8;
+  // Disallow subpixel for now
+  // SUBPEL_MASK is not the correct scale
+  if ((dv.row & (SCALE_PX_TO_MV - 1) || dv.col & (SCALE_PX_TO_MV - 1)))
+    return 0;
+  // Is the source top-left inside the current tile?
+  const int src_top_edge = mi_row * MI_SIZE * SCALE_PX_TO_MV + dv.row;
+  const int tile_top_edge = tile->mi_row_start * MI_SIZE * SCALE_PX_TO_MV;
+  if (src_top_edge < tile_top_edge) return 0;
+  const int src_left_edge = mi_col * MI_SIZE * SCALE_PX_TO_MV + dv.col;
+  const int tile_left_edge = tile->mi_col_start * MI_SIZE * SCALE_PX_TO_MV;
+  if (src_left_edge < tile_left_edge) return 0;
+  // Is the bottom right inside the current tile?
+  const int src_bottom_edge = (mi_row * MI_SIZE + bh) * SCALE_PX_TO_MV + dv.row;
+  const int tile_bottom_edge = tile->mi_row_end * MI_SIZE * SCALE_PX_TO_MV;
+  if (src_bottom_edge > tile_bottom_edge) return 0;
+  const int src_right_edge = (mi_col * MI_SIZE + bw) * SCALE_PX_TO_MV + dv.col;
+  const int tile_right_edge = tile->mi_col_end * MI_SIZE * SCALE_PX_TO_MV;
+  if (src_right_edge > tile_right_edge) return 0;
+  // Is the bottom right within an already coded SB?
+  const int active_sb_top_edge =
+      (mi_row & ~MAX_MIB_MASK) * MI_SIZE * SCALE_PX_TO_MV;
+  const int active_sb_bottom_edge =
+      ((mi_row & ~MAX_MIB_MASK) + MAX_MIB_SIZE) * MI_SIZE * SCALE_PX_TO_MV;
+  const int active_sb_left_edge =
+      (mi_col & ~MAX_MIB_MASK) * MI_SIZE * SCALE_PX_TO_MV;
+  if (src_bottom_edge > active_sb_bottom_edge) return 0;
+  if (src_bottom_edge > active_sb_top_edge &&
+      src_right_edge > active_sb_left_edge)
+    return 0;
+  return 1;
+}
+#endif  // CONFIG_INTRABC
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_MVREF_COMMON_H_
diff --git a/third_party/aom/av1/common/od_dering.c b/third_party/aom/av1/common/od_dering.c
new file mode 100644
index 0000000000..f54f337ef5
--- /dev/null
+++ b/third_party/aom/av1/common/od_dering.c
@@ -0,0 +1,416 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+
+#ifdef HAVE_CONFIG_H
+#include "./config.h"
+#endif
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+#include "./cdef.h"
+
+/* Generated from gen_filter_tables.c. */
+const int OD_DIRECTION_OFFSETS_TABLE[8][3] = {
+  { -1 * OD_FILT_BSTRIDE + 1, -2 * OD_FILT_BSTRIDE + 2,
+    -3 * OD_FILT_BSTRIDE + 3 },
+  { 0 * OD_FILT_BSTRIDE + 1, -1 * OD_FILT_BSTRIDE + 2,
+    -1 * OD_FILT_BSTRIDE + 3 },
+  { 0 * OD_FILT_BSTRIDE + 1, 0 * OD_FILT_BSTRIDE + 2, 0 * OD_FILT_BSTRIDE + 3 },
+  { 0 * OD_FILT_BSTRIDE + 1, 1 * OD_FILT_BSTRIDE + 2, 1 * OD_FILT_BSTRIDE + 3 },
+  { 1 * OD_FILT_BSTRIDE + 1, 2 * OD_FILT_BSTRIDE + 2, 3 * OD_FILT_BSTRIDE + 3 },
+  { 1 * OD_FILT_BSTRIDE + 0, 2 * OD_FILT_BSTRIDE + 1, 3 * OD_FILT_BSTRIDE + 1 },
+  { 1 * OD_FILT_BSTRIDE + 0, 2 * OD_FILT_BSTRIDE + 0, 3 * OD_FILT_BSTRIDE + 0 },
+  { 1 * OD_FILT_BSTRIDE + 0, 2 * OD_FILT_BSTRIDE - 1, 3 * OD_FILT_BSTRIDE - 1 },
+};
+
+/* Detect direction. 0 means 45-degree up-right, 2 is horizontal, and so on.
+   The search minimizes the weighted variance along all the lines in a
+   particular direction, i.e. the squared error between the input and a
+   "predicted" block where each pixel is replaced by the average along a line
+   in a particular direction. Since each direction have the same sum(x^2) term,
+   that term is never computed. See Section 2, step 2, of:
+   http://jmvalin.ca/notes/intra_paint.pdf */
+int od_dir_find8_c(const uint16_t *img, int stride, int32_t *var,
+                   int coeff_shift) {
+  int i;
+  int32_t cost[8] = { 0 };
+  int partial[8][15] = { { 0 } };
+  int32_t best_cost = 0;
+  int best_dir = 0;
+  /* Instead of dividing by n between 2 and 8, we multiply by 3*5*7*8/n.
+     The output is then 840 times larger, but we don't care for finding
+     the max. */
+  static const int div_table[] = { 0, 840, 420, 280, 210, 168, 140, 120, 105 };
+  for (i = 0; i < 8; i++) {
+    int j;
+    for (j = 0; j < 8; j++) {
+      int x;
+      /* We subtract 128 here to reduce the maximum range of the squared
+         partial sums. */
+      x = (img[i * stride + j] >> coeff_shift) - 128;
+      partial[0][i + j] += x;
+      partial[1][i + j / 2] += x;
+      partial[2][i] += x;
+      partial[3][3 + i - j / 2] += x;
+      partial[4][7 + i - j] += x;
+      partial[5][3 - i / 2 + j] += x;
+      partial[6][j] += x;
+      partial[7][i / 2 + j] += x;
+    }
+  }
+  for (i = 0; i < 8; i++) {
+    cost[2] += partial[2][i] * partial[2][i];
+    cost[6] += partial[6][i] * partial[6][i];
+  }
+  cost[2] *= div_table[8];
+  cost[6] *= div_table[8];
+  for (i = 0; i < 7; i++) {
+    cost[0] += (partial[0][i] * partial[0][i] +
+                partial[0][14 - i] * partial[0][14 - i]) *
+               div_table[i + 1];
+    cost[4] += (partial[4][i] * partial[4][i] +
+                partial[4][14 - i] * partial[4][14 - i]) *
+               div_table[i + 1];
+  }
+  cost[0] += partial[0][7] * partial[0][7] * div_table[8];
+  cost[4] += partial[4][7] * partial[4][7] * div_table[8];
+  for (i = 1; i < 8; i += 2) {
+    int j;
+    for (j = 0; j < 4 + 1; j++) {
+      cost[i] += partial[i][3 + j] * partial[i][3 + j];
+    }
+    cost[i] *= div_table[8];
+    for (j = 0; j < 4 - 1; j++) {
+      cost[i] += (partial[i][j] * partial[i][j] +
+                  partial[i][10 - j] * partial[i][10 - j]) *
+                 div_table[2 * j + 2];
+    }
+  }
+  for (i = 0; i < 8; i++) {
+    if (cost[i] > best_cost) {
+      best_cost = cost[i];
+      best_dir = i;
+    }
+  }
+  /* Difference between the optimal variance and the variance along the
+     orthogonal direction. Again, the sum(x^2) terms cancel out. */
+  *var = best_cost - cost[(best_dir + 4) & 7];
+  /* We'd normally divide by 840, but dividing by 1024 is close enough
+     for what we're going to do with this. */
+  *var >>= 10;
+  return best_dir;
+}
+
+/* Smooth in the direction detected. */
+void od_filter_dering_direction_8x8_c(uint16_t *y, int ystride,
+                                      const uint16_t *in, int threshold,
+                                      int dir, int damping) {
+  int i;
+  int j;
+  int k;
+  static const int taps[3] = { 3, 2, 1 };
+  for (i = 0; i < 8; i++) {
+    for (j = 0; j < 8; j++) {
+      int16_t sum;
+      int16_t xx;
+      int16_t yy;
+      xx = in[i * OD_FILT_BSTRIDE + j];
+      sum = 0;
+      for (k = 0; k < 3; k++) {
+        int16_t p0;
+        int16_t p1;
+        p0 = in[i * OD_FILT_BSTRIDE + j + OD_DIRECTION_OFFSETS_TABLE[dir][k]] -
+             xx;
+        p1 = in[i * OD_FILT_BSTRIDE + j - OD_DIRECTION_OFFSETS_TABLE[dir][k]] -
+             xx;
+        sum += taps[k] * constrain(p0, threshold, damping);
+        sum += taps[k] * constrain(p1, threshold, damping);
+      }
+      sum = (sum + 8) >> 4;
+      yy = xx + sum;
+      y[i * ystride + j] = yy;
+    }
+  }
+}
+
+/* Smooth in the direction detected. */
+void od_filter_dering_direction_4x4_c(uint16_t *y, int ystride,
+                                      const uint16_t *in, int threshold,
+                                      int dir, int damping) {
+  int i;
+  int j;
+  int k;
+  static const int taps[2] = { 4, 1 };
+  for (i = 0; i < 4; i++) {
+    for (j = 0; j < 4; j++) {
+      int16_t sum;
+      int16_t xx;
+      int16_t yy;
+      xx = in[i * OD_FILT_BSTRIDE + j];
+      sum = 0;
+      for (k = 0; k < 2; k++) {
+        int16_t p0;
+        int16_t p1;
+        p0 = in[i * OD_FILT_BSTRIDE + j + OD_DIRECTION_OFFSETS_TABLE[dir][k]] -
+             xx;
+        p1 = in[i * OD_FILT_BSTRIDE + j - OD_DIRECTION_OFFSETS_TABLE[dir][k]] -
+             xx;
+        sum += taps[k] * constrain(p0, threshold, damping);
+        sum += taps[k] * constrain(p1, threshold, damping);
+      }
+      sum = (sum + 8) >> 4;
+      yy = xx + sum;
+      y[i * ystride + j] = yy;
+    }
+  }
+}
+
+/* Compute deringing filter threshold for an 8x8 block based on the
+   directional variance difference. A high variance difference means that we
+   have a highly directional pattern (e.g. a high contrast edge), so we can
+   apply more deringing. A low variance means that we either have a low
+   contrast edge, or a non-directional texture, so we want to be careful not
+   to blur. */
+static INLINE int od_adjust_thresh(int threshold, int32_t var) {
+  const int i = var >> 6 ? AOMMIN(get_msb(var >> 6), 12) : 0;
+  /* We use the variance of 8x8 blocks to adjust the threshold. */
+  return var ? (threshold * (4 + i) + 8) >> 4 : 0;
+}
+
+void copy_8x8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src,
+                               int sstride) {
+  int i, j;
+  for (i = 0; i < 8; i++)
+    for (j = 0; j < 8; j++) dst[i * dstride + j] = src[i * sstride + j];
+}
+
+void copy_4x4_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src,
+                               int sstride) {
+  int i, j;
+  for (i = 0; i < 4; i++)
+    for (j = 0; j < 4; j++) dst[i * dstride + j] = src[i * sstride + j];
+}
+
+void copy_dering_16bit_to_16bit(uint16_t *dst, int dstride, uint16_t *src,
+                                dering_list *dlist, int dering_count,
+                                int bsize) {
+  int bi, bx, by;
+
+  if (bsize == BLOCK_8X8) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      copy_8x8_16bit_to_16bit(&dst[(by << 3) * dstride + (bx << 3)], dstride,
+                              &src[bi << (3 + 3)], 8);
+    }
+  } else if (bsize == BLOCK_4X8) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      copy_4x4_16bit_to_16bit(&dst[(by << 3) * dstride + (bx << 2)], dstride,
+                              &src[bi << (3 + 2)], 4);
+      copy_4x4_16bit_to_16bit(&dst[((by << 3) + 4) * dstride + (bx << 2)],
+                              dstride, &src[(bi << (3 + 2)) + 4 * 4], 4);
+    }
+  } else if (bsize == BLOCK_8X4) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      copy_4x4_16bit_to_16bit(&dst[(by << 2) * dstride + (bx << 3)], dstride,
+                              &src[bi << (2 + 3)], 8);
+      copy_4x4_16bit_to_16bit(&dst[(by << 2) * dstride + (bx << 3) + 4],
+                              dstride, &src[(bi << (2 + 3)) + 4], 8);
+    }
+  } else {
+    assert(bsize == BLOCK_4X4);
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      copy_4x4_16bit_to_16bit(&dst[(by << 2) * dstride + (bx << 2)], dstride,
+                              &src[bi << (2 + 2)], 4);
+    }
+  }
+}
+
+void copy_8x8_16bit_to_8bit_c(uint8_t *dst, int dstride, const uint16_t *src,
+                              int sstride) {
+  int i, j;
+  for (i = 0; i < 8; i++)
+    for (j = 0; j < 8; j++)
+      dst[i * dstride + j] = (uint8_t)src[i * sstride + j];
+}
+
+void copy_4x4_16bit_to_8bit_c(uint8_t *dst, int dstride, const uint16_t *src,
+                              int sstride) {
+  int i, j;
+  for (i = 0; i < 4; i++)
+    for (j = 0; j < 4; j++)
+      dst[i * dstride + j] = (uint8_t)src[i * sstride + j];
+}
+
+static void copy_dering_16bit_to_8bit(uint8_t *dst, int dstride,
+                                      const uint16_t *src, dering_list *dlist,
+                                      int dering_count, int bsize) {
+  int bi, bx, by;
+  if (bsize == BLOCK_8X8) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      copy_8x8_16bit_to_8bit(&dst[(by << 3) * dstride + (bx << 3)], dstride,
+                             &src[bi << (3 + 3)], 8);
+    }
+  } else if (bsize == BLOCK_4X8) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      copy_4x4_16bit_to_8bit(&dst[(by << 3) * dstride + (bx << 2)], dstride,
+                             &src[bi << (3 + 2)], 4);
+      copy_4x4_16bit_to_8bit(&dst[((by << 3) + 4) * dstride + (bx << 2)],
+                             dstride, &src[(bi << (3 + 2)) + 4 * 4], 4);
+    }
+  } else if (bsize == BLOCK_8X4) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      copy_4x4_16bit_to_8bit(&dst[(by << 2) * dstride + (bx << 3)], dstride,
+                             &src[bi << (2 + 3)], 8);
+      copy_4x4_16bit_to_8bit(&dst[(by << 2) * dstride + (bx << 3) + 4], dstride,
+                             &src[(bi << (2 + 3)) + 4], 8);
+    }
+  } else {
+    assert(bsize == BLOCK_4X4);
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      copy_4x4_16bit_to_8bit(&dst[(by << 2) * dstride + (bx << 2)], dstride,
+                             &src[bi << (2 * 2)], 4);
+    }
+  }
+}
+
+int get_filter_skip(int level) {
+  int filter_skip = level & 1;
+  if (level == 1) filter_skip = 0;
+  return filter_skip;
+}
+
+void od_dering(uint8_t *dst, int dstride, uint16_t *y, uint16_t *in, int xdec,
+               int ydec, int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS],
+               int *dirinit, int var[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS],
+               int pli, dering_list *dlist, int dering_count, int level,
+               int clpf_strength, int clpf_damping, int dering_damping,
+               int coeff_shift, int skip_dering, int hbd) {
+  int bi;
+  int bx;
+  int by;
+  int bsize, bsizex, bsizey;
+
+  int threshold = (level >> 1) << coeff_shift;
+  int filter_skip = get_filter_skip(level);
+  if (level == 1) threshold = 31 << coeff_shift;
+
+  od_filter_dering_direction_func filter_dering_direction[] = {
+    od_filter_dering_direction_4x4, od_filter_dering_direction_8x8
+  };
+  clpf_damping += coeff_shift - (pli != AOM_PLANE_Y);
+  dering_damping += coeff_shift - (pli != AOM_PLANE_Y);
+  bsize =
+      ydec ? (xdec ? BLOCK_4X4 : BLOCK_8X4) : (xdec ? BLOCK_4X8 : BLOCK_8X8);
+  bsizex = 3 - xdec;
+  bsizey = 3 - ydec;
+
+  if (!skip_dering) {
+    if (pli == 0) {
+      if (!dirinit || !*dirinit) {
+        for (bi = 0; bi < dering_count; bi++) {
+          by = dlist[bi].by;
+          bx = dlist[bi].bx;
+          dir[by][bx] =
+              od_dir_find8(&in[8 * by * OD_FILT_BSTRIDE + 8 * bx],
+                           OD_FILT_BSTRIDE, &var[by][bx], coeff_shift);
+        }
+        if (dirinit) *dirinit = 1;
+      }
+    }
+    // Only run dering for non-zero threshold (which is always the case for
+    // 4:2:2 or 4:4:0). If we don't dering, we still need to eventually write
+    // something out in y[] later.
+    if (threshold != 0) {
+      assert(bsize == BLOCK_8X8 || bsize == BLOCK_4X4);
+      for (bi = 0; bi < dering_count; bi++) {
+        int t = !filter_skip && dlist[bi].skip ? 0 : threshold;
+        by = dlist[bi].by;
+        bx = dlist[bi].bx;
+        (filter_dering_direction[bsize == BLOCK_8X8])(
+            &y[bi << (bsizex + bsizey)], 1 << bsizex,
+            &in[(by * OD_FILT_BSTRIDE << bsizey) + (bx << bsizex)],
+            pli ? t : od_adjust_thresh(t, var[by][bx]), dir[by][bx],
+            dering_damping);
+      }
+    }
+  }
+
+  if (clpf_strength) {
+    if (threshold && !skip_dering)
+      copy_dering_16bit_to_16bit(in, OD_FILT_BSTRIDE, y, dlist, dering_count,
+                                 bsize);
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      int py = by << bsizey;
+      int px = bx << bsizex;
+
+      if (!filter_skip && dlist[bi].skip) continue;
+      if (!dst || hbd) {
+        // 16 bit destination if high bitdepth or 8 bit destination not given
+        (!threshold || (dir[by][bx] < 4 && dir[by][bx]) ? aom_clpf_block_hbd
+                                                        : aom_clpf_hblock_hbd)(
+            dst ? (uint16_t *)dst + py * dstride + px
+                : &y[bi << (bsizex + bsizey)],
+            in + py * OD_FILT_BSTRIDE + px, dst && hbd ? dstride : 1 << bsizex,
+            OD_FILT_BSTRIDE, 1 << bsizex, 1 << bsizey,
+            clpf_strength << coeff_shift, clpf_damping);
+      } else {
+        // Do clpf and write the result to an 8 bit destination
+        (!threshold || (dir[by][bx] < 4 && dir[by][bx]) ? aom_clpf_block
+                                                        : aom_clpf_hblock)(
+            dst + py * dstride + px, in + py * OD_FILT_BSTRIDE + px, dstride,
+            OD_FILT_BSTRIDE, 1 << bsizex, 1 << bsizey,
+            clpf_strength << coeff_shift, clpf_damping);
+      }
+    }
+  } else if (threshold != 0) {
+    // No clpf, so copy instead
+    if (hbd) {
+      copy_dering_16bit_to_16bit((uint16_t *)dst, dstride, y, dlist,
+                                 dering_count, bsize);
+    } else {
+      copy_dering_16bit_to_8bit(dst, dstride, y, dlist, dering_count, bsize);
+    }
+  } else if (dirinit) {
+    // If we're here, both dering and clpf are off, and we still haven't written
+    // anything to y[] yet, so we just copy the input to y[]. This is necessary
+    // only for av1_cdef_search() and only av1_cdef_search() sets dirinit.
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      int iy, ix;
+      // TODO(stemidts/jmvalin): SIMD optimisations
+      for (iy = 0; iy < 1 << bsizey; iy++)
+        for (ix = 0; ix < 1 << bsizex; ix++)
+          y[(bi << (bsizex + bsizey)) + (iy << bsizex) + ix] =
+              in[((by << bsizey) + iy) * OD_FILT_BSTRIDE + (bx << bsizex) + ix];
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/od_dering.h b/third_party/aom/av1/common/od_dering.h
new file mode 100644
index 0000000000..4362001b45
--- /dev/null
+++ b/third_party/aom/av1/common/od_dering.h
@@ -0,0 +1,54 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if !defined(_dering_H)
+#define _dering_H (1)
+
+#include "odintrin.h"
+
+#define OD_DERING_NBLOCKS (MAX_SB_SIZE / 8)
+
+/* We need to buffer three vertical lines. */
+#define OD_FILT_VBORDER (3)
+/* We only need to buffer three horizontal pixels too, but let's align to
+   16 bytes (8 x 16 bits) to make vectorization easier. */
+#define OD_FILT_HBORDER (8)
+#define OD_FILT_BSTRIDE ALIGN_POWER_OF_TWO(MAX_SB_SIZE + 2 * OD_FILT_HBORDER, 3)
+
+#define OD_DERING_VERY_LARGE (30000)
+#define OD_DERING_INBUF_SIZE \
+  (OD_FILT_BSTRIDE * (MAX_SB_SIZE + 2 * OD_FILT_VBORDER))
+
+extern const int OD_DIRECTION_OFFSETS_TABLE[8][3];
+
+typedef struct {
+  uint8_t by;
+  uint8_t bx;
+  uint8_t skip;
+} dering_list;
+
+typedef void (*od_filter_dering_direction_func)(uint16_t *y, int ystride,
+                                                const uint16_t *in,
+                                                int threshold, int dir,
+                                                int damping);
+void copy_dering_16bit_to_16bit(uint16_t *dst, int dstride, uint16_t *src,
+                                dering_list *dlist, int dering_count,
+                                int bsize);
+
+int get_filter_skip(int level);
+
+void od_dering(uint8_t *dst, int dstride, uint16_t *y, uint16_t *in, int xdec,
+               int ydec, int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS],
+               int *dirinit, int var[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS],
+               int pli, dering_list *dlist, int dering_count, int level,
+               int clpf_strength, int clpf_damping, int dering_damping,
+               int coeff_shift, int skip_dering, int hbd);
+#endif
diff --git a/third_party/aom/av1/common/od_dering_neon.c b/third_party/aom/av1/common/od_dering_neon.c
new file mode 100644
index 0000000000..99441050ab
--- /dev/null
+++ b/third_party/aom/av1/common/od_dering_neon.c
@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_neon
+#include "./od_dering_simd.h"
diff --git a/third_party/aom/av1/common/od_dering_simd.h b/third_party/aom/av1/common/od_dering_simd.h
new file mode 100644
index 0000000000..4074e7e509
--- /dev/null
+++ b/third_party/aom/av1/common/od_dering_simd.h
@@ -0,0 +1,390 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "./cdef_simd.h"
+#include "./od_dering.h"
+
+/* partial A is a 16-bit vector of the form:
+   [x8 x7 x6 x5 x4 x3 x2 x1] and partial B has the form:
+   [0  y1 y2 y3 y4 y5 y6 y7].
+   This function computes (x1^2+y1^2)*C1 + (x2^2+y2^2)*C2 + ...
+   (x7^2+y2^7)*C7 + (x8^2+0^2)*C8 where the C1..C8 constants are in const1
+   and const2. */
+static INLINE v128 fold_mul_and_sum(v128 partiala, v128 partialb, v128 const1,
+                                    v128 const2) {
+  v128 tmp;
+  /* Reverse partial B. */
+  partialb = v128_shuffle_8(
+      partialb, v128_from_32(0x0f0e0100, 0x03020504, 0x07060908, 0x0b0a0d0c));
+  /* Interleave the x and y values of identical indices and pair x8 with 0. */
+  tmp = partiala;
+  partiala = v128_ziplo_16(partialb, partiala);
+  partialb = v128_ziphi_16(partialb, tmp);
+  /* Square and add the corresponding x and y values. */
+  partiala = v128_madd_s16(partiala, partiala);
+  partialb = v128_madd_s16(partialb, partialb);
+  /* Multiply by constant. */
+  partiala = v128_mullo_s32(partiala, const1);
+  partialb = v128_mullo_s32(partialb, const2);
+  /* Sum all results. */
+  partiala = v128_add_32(partiala, partialb);
+  return partiala;
+}
+
+static INLINE v128 hsum4(v128 x0, v128 x1, v128 x2, v128 x3) {
+  v128 t0, t1, t2, t3;
+  t0 = v128_ziplo_32(x1, x0);
+  t1 = v128_ziplo_32(x3, x2);
+  t2 = v128_ziphi_32(x1, x0);
+  t3 = v128_ziphi_32(x3, x2);
+  x0 = v128_ziplo_64(t1, t0);
+  x1 = v128_ziphi_64(t1, t0);
+  x2 = v128_ziplo_64(t3, t2);
+  x3 = v128_ziphi_64(t3, t2);
+  return v128_add_32(v128_add_32(x0, x1), v128_add_32(x2, x3));
+}
+
+/* Computes cost for directions 0, 5, 6 and 7. We can call this function again
+   to compute the remaining directions. */
+static INLINE v128 compute_directions(v128 lines[8], int32_t tmp_cost1[4]) {
+  v128 partial4a, partial4b, partial5a, partial5b, partial7a, partial7b;
+  v128 partial6;
+  v128 tmp;
+  /* Partial sums for lines 0 and 1. */
+  partial4a = v128_shl_n_byte(lines[0], 14);
+  partial4b = v128_shr_n_byte(lines[0], 2);
+  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[1], 12));
+  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[1], 4));
+  tmp = v128_add_16(lines[0], lines[1]);
+  partial5a = v128_shl_n_byte(tmp, 10);
+  partial5b = v128_shr_n_byte(tmp, 6);
+  partial7a = v128_shl_n_byte(tmp, 4);
+  partial7b = v128_shr_n_byte(tmp, 12);
+  partial6 = tmp;
+
+  /* Partial sums for lines 2 and 3. */
+  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[2], 10));
+  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[2], 6));
+  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[3], 8));
+  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[3], 8));
+  tmp = v128_add_16(lines[2], lines[3]);
+  partial5a = v128_add_16(partial5a, v128_shl_n_byte(tmp, 8));
+  partial5b = v128_add_16(partial5b, v128_shr_n_byte(tmp, 8));
+  partial7a = v128_add_16(partial7a, v128_shl_n_byte(tmp, 6));
+  partial7b = v128_add_16(partial7b, v128_shr_n_byte(tmp, 10));
+  partial6 = v128_add_16(partial6, tmp);
+
+  /* Partial sums for lines 4 and 5. */
+  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[4], 6));
+  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[4], 10));
+  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[5], 4));
+  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[5], 12));
+  tmp = v128_add_16(lines[4], lines[5]);
+  partial5a = v128_add_16(partial5a, v128_shl_n_byte(tmp, 6));
+  partial5b = v128_add_16(partial5b, v128_shr_n_byte(tmp, 10));
+  partial7a = v128_add_16(partial7a, v128_shl_n_byte(tmp, 8));
+  partial7b = v128_add_16(partial7b, v128_shr_n_byte(tmp, 8));
+  partial6 = v128_add_16(partial6, tmp);
+
+  /* Partial sums for lines 6 and 7. */
+  partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[6], 2));
+  partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[6], 14));
+  partial4a = v128_add_16(partial4a, lines[7]);
+  tmp = v128_add_16(lines[6], lines[7]);
+  partial5a = v128_add_16(partial5a, v128_shl_n_byte(tmp, 4));
+  partial5b = v128_add_16(partial5b, v128_shr_n_byte(tmp, 12));
+  partial7a = v128_add_16(partial7a, v128_shl_n_byte(tmp, 10));
+  partial7b = v128_add_16(partial7b, v128_shr_n_byte(tmp, 6));
+  partial6 = v128_add_16(partial6, tmp);
+
+  /* Compute costs in terms of partial sums. */
+  partial4a =
+      fold_mul_and_sum(partial4a, partial4b, v128_from_32(210, 280, 420, 840),
+                       v128_from_32(105, 120, 140, 168));
+  partial7a =
+      fold_mul_and_sum(partial7a, partial7b, v128_from_32(210, 420, 0, 0),
+                       v128_from_32(105, 105, 105, 140));
+  partial5a =
+      fold_mul_and_sum(partial5a, partial5b, v128_from_32(210, 420, 0, 0),
+                       v128_from_32(105, 105, 105, 140));
+  partial6 = v128_madd_s16(partial6, partial6);
+  partial6 = v128_mullo_s32(partial6, v128_dup_32(105));
+
+  partial4a = hsum4(partial4a, partial5a, partial6, partial7a);
+  v128_store_unaligned(tmp_cost1, partial4a);
+  return partial4a;
+}
+
+/* transpose and reverse the order of the lines -- equivalent to a 90-degree
+   counter-clockwise rotation of the pixels. */
+static INLINE void array_reverse_transpose_8x8(v128 *in, v128 *res) {
+  const v128 tr0_0 = v128_ziplo_16(in[1], in[0]);
+  const v128 tr0_1 = v128_ziplo_16(in[3], in[2]);
+  const v128 tr0_2 = v128_ziphi_16(in[1], in[0]);
+  const v128 tr0_3 = v128_ziphi_16(in[3], in[2]);
+  const v128 tr0_4 = v128_ziplo_16(in[5], in[4]);
+  const v128 tr0_5 = v128_ziplo_16(in[7], in[6]);
+  const v128 tr0_6 = v128_ziphi_16(in[5], in[4]);
+  const v128 tr0_7 = v128_ziphi_16(in[7], in[6]);
+
+  const v128 tr1_0 = v128_ziplo_32(tr0_1, tr0_0);
+  const v128 tr1_1 = v128_ziplo_32(tr0_5, tr0_4);
+  const v128 tr1_2 = v128_ziphi_32(tr0_1, tr0_0);
+  const v128 tr1_3 = v128_ziphi_32(tr0_5, tr0_4);
+  const v128 tr1_4 = v128_ziplo_32(tr0_3, tr0_2);
+  const v128 tr1_5 = v128_ziplo_32(tr0_7, tr0_6);
+  const v128 tr1_6 = v128_ziphi_32(tr0_3, tr0_2);
+  const v128 tr1_7 = v128_ziphi_32(tr0_7, tr0_6);
+
+  res[7] = v128_ziplo_64(tr1_1, tr1_0);
+  res[6] = v128_ziphi_64(tr1_1, tr1_0);
+  res[5] = v128_ziplo_64(tr1_3, tr1_2);
+  res[4] = v128_ziphi_64(tr1_3, tr1_2);
+  res[3] = v128_ziplo_64(tr1_5, tr1_4);
+  res[2] = v128_ziphi_64(tr1_5, tr1_4);
+  res[1] = v128_ziplo_64(tr1_7, tr1_6);
+  res[0] = v128_ziphi_64(tr1_7, tr1_6);
+}
+
+int SIMD_FUNC(od_dir_find8)(const od_dering_in *img, int stride, int32_t *var,
+                            int coeff_shift) {
+  int i;
+  int32_t cost[8];
+  int32_t best_cost = 0;
+  int best_dir = 0;
+  v128 lines[8];
+  for (i = 0; i < 8; i++) {
+    lines[i] = v128_load_unaligned(&img[i * stride]);
+    lines[i] =
+        v128_sub_16(v128_shr_s16(lines[i], coeff_shift), v128_dup_16(128));
+  }
+
+#if defined(__SSE4_1__)
+  /* Compute "mostly vertical" directions. */
+  __m128i dir47 = compute_directions(lines, cost + 4);
+
+  array_reverse_transpose_8x8(lines, lines);
+
+  /* Compute "mostly horizontal" directions. */
+  __m128i dir03 = compute_directions(lines, cost);
+
+  __m128i max = _mm_max_epi32(dir03, dir47);
+  max = _mm_max_epi32(max, _mm_shuffle_epi32(max, _MM_SHUFFLE(1, 0, 3, 2)));
+  max = _mm_max_epi32(max, _mm_shuffle_epi32(max, _MM_SHUFFLE(2, 3, 0, 1)));
+  best_cost = _mm_cvtsi128_si32(max);
+  __m128i t =
+      _mm_packs_epi32(_mm_cmpeq_epi32(max, dir03), _mm_cmpeq_epi32(max, dir47));
+  best_dir = _mm_movemask_epi8(_mm_packs_epi16(t, t));
+  best_dir = get_msb(best_dir ^ (best_dir - 1));  // Count trailing zeros
+#else
+  /* Compute "mostly vertical" directions. */
+  compute_directions(lines, cost + 4);
+
+  array_reverse_transpose_8x8(lines, lines);
+
+  /* Compute "mostly horizontal" directions. */
+  compute_directions(lines, cost);
+
+  for (i = 0; i < 8; i++) {
+    if (cost[i] > best_cost) {
+      best_cost = cost[i];
+      best_dir = i;
+    }
+  }
+#endif
+
+  /* Difference between the optimal variance and the variance along the
+     orthogonal direction. Again, the sum(x^2) terms cancel out. */
+  *var = best_cost - cost[(best_dir + 4) & 7];
+  /* We'd normally divide by 840, but dividing by 1024 is close enough
+     for what we're going to do with this. */
+  *var >>= 10;
+  return best_dir;
+}
+
+void SIMD_FUNC(od_filter_dering_direction_4x4)(uint16_t *y, int ystride,
+                                               const uint16_t *in,
+                                               int threshold, int dir,
+                                               int damping) {
+  int i;
+  v128 p0, p1, sum, row, res;
+  int o1 = OD_DIRECTION_OFFSETS_TABLE[dir][0];
+  int o2 = OD_DIRECTION_OFFSETS_TABLE[dir][1];
+
+  if (threshold) damping -= get_msb(threshold);
+  for (i = 0; i < 4; i += 2) {
+    sum = v128_zero();
+    row = v128_from_v64(v64_load_aligned(&in[i * OD_FILT_BSTRIDE]),
+                        v64_load_aligned(&in[(i + 1) * OD_FILT_BSTRIDE]));
+
+    // p0 = constrain16(in[i*OD_FILT_BSTRIDE + offset], row, threshold, damping)
+    p0 = v128_from_v64(v64_load_unaligned(&in[i * OD_FILT_BSTRIDE + o1]),
+                       v64_load_unaligned(&in[(i + 1) * OD_FILT_BSTRIDE + o1]));
+    p0 = constrain16(p0, row, threshold, damping);
+
+    // p1 = constrain16(in[i*OD_FILT_BSTRIDE - offset], row, threshold, damping)
+    p1 = v128_from_v64(v64_load_unaligned(&in[i * OD_FILT_BSTRIDE - o1]),
+                       v64_load_unaligned(&in[(i + 1) * OD_FILT_BSTRIDE - o1]));
+    p1 = constrain16(p1, row, threshold, damping);
+
+    // sum += 4 * (p0 + p1)
+    sum = v128_add_16(sum, v128_shl_n_16(v128_add_16(p0, p1), 2));
+
+    // p0 = constrain16(in[i*OD_FILT_BSTRIDE + offset], row, threshold, damping)
+    p0 = v128_from_v64(v64_load_unaligned(&in[i * OD_FILT_BSTRIDE + o2]),
+                       v64_load_unaligned(&in[(i + 1) * OD_FILT_BSTRIDE + o2]));
+    p0 = constrain16(p0, row, threshold, damping);
+
+    // p1 = constrain16(in[i*OD_FILT_BSTRIDE - offset], row, threshold, damping)
+    p1 = v128_from_v64(v64_load_unaligned(&in[i * OD_FILT_BSTRIDE - o2]),
+                       v64_load_unaligned(&in[(i + 1) * OD_FILT_BSTRIDE - o2]));
+    p1 = constrain16(p1, row, threshold, damping);
+
+    // sum += 1 * (p0 + p1)
+    sum = v128_add_16(sum, v128_add_16(p0, p1));
+
+    // res = row + ((sum + 8) >> 4)
+    res = v128_add_16(sum, v128_dup_16(8));
+    res = v128_shr_n_s16(res, 4);
+    res = v128_add_16(row, res);
+    v64_store_aligned(&y[i * ystride], v128_high_v64(res));
+    v64_store_aligned(&y[(i + 1) * ystride], v128_low_v64(res));
+  }
+}
+
+void SIMD_FUNC(od_filter_dering_direction_8x8)(uint16_t *y, int ystride,
+                                               const uint16_t *in,
+                                               int threshold, int dir,
+                                               int damping) {
+  int i;
+  v128 sum, p0, p1, row, res;
+  int o1 = OD_DIRECTION_OFFSETS_TABLE[dir][0];
+  int o2 = OD_DIRECTION_OFFSETS_TABLE[dir][1];
+  int o3 = OD_DIRECTION_OFFSETS_TABLE[dir][2];
+
+  if (threshold) damping -= get_msb(threshold);
+  for (i = 0; i < 8; i++) {
+    sum = v128_zero();
+    row = v128_load_aligned(&in[i * OD_FILT_BSTRIDE]);
+
+    // p0 = constrain16(in[i*OD_FILT_BSTRIDE + offset], row, threshold, damping)
+    p0 = v128_load_unaligned(&in[i * OD_FILT_BSTRIDE + o1]);
+    p0 = constrain16(p0, row, threshold, damping);
+
+    // p1 = constrain16(in[i*OD_FILT_BSTRIDE - offset], row, threshold, damping)
+    p1 = v128_load_unaligned(&in[i * OD_FILT_BSTRIDE - o1]);
+    p1 = constrain16(p1, row, threshold, damping);
+
+    // sum += 3 * (p0 + p1)
+    p0 = v128_add_16(p0, p1);
+    p0 = v128_add_16(p0, v128_shl_n_16(p0, 1));
+    sum = v128_add_16(sum, p0);
+
+    // p0 = constrain16(in[i*OD_FILT_BSTRIDE + offset], row, threshold, damping)
+    p0 = v128_load_unaligned(&in[i * OD_FILT_BSTRIDE + o2]);
+    p0 = constrain16(p0, row, threshold, damping);
+
+    // p1 = constrain16(in[i*OD_FILT_BSTRIDE - offset], row, threshold, damping)
+    p1 = v128_load_unaligned(&in[i * OD_FILT_BSTRIDE - o2]);
+    p1 = constrain16(p1, row, threshold, damping);
+
+    // sum += 2 * (p0 + p1)
+    p0 = v128_shl_n_16(v128_add_16(p0, p1), 1);
+    sum = v128_add_16(sum, p0);
+
+    // p0 = constrain16(in[i*OD_FILT_BSTRIDE + offset], row, threshold, damping)
+    p0 = v128_load_unaligned(&in[i * OD_FILT_BSTRIDE + o3]);
+    p0 = constrain16(p0, row, threshold, damping);
+
+    // p1 = constrain16(in[i*OD_FILT_BSTRIDE - offset], row, threshold, damping)
+    p1 = v128_load_unaligned(&in[i * OD_FILT_BSTRIDE - o3]);
+    p1 = constrain16(p1, row, threshold, damping);
+
+    // sum += (p0 + p1)
+    p0 = v128_add_16(p0, p1);
+    sum = v128_add_16(sum, p0);
+
+    // res = row + ((sum + 8) >> 4)
+    res = v128_add_16(sum, v128_dup_16(8));
+    res = v128_shr_n_s16(res, 4);
+    res = v128_add_16(row, res);
+    v128_store_unaligned(&y[i * ystride], res);
+  }
+}
+
+void SIMD_FUNC(copy_8x8_16bit_to_8bit)(uint8_t *dst, int dstride,
+                                       const uint16_t *src, int sstride) {
+  int i;
+  for (i = 0; i < 8; i++) {
+    v128 row = v128_load_unaligned(&src[i * sstride]);
+    row = v128_pack_s16_u8(row, row);
+    v64_store_unaligned(&dst[i * dstride], v128_low_v64(row));
+  }
+}
+
+void SIMD_FUNC(copy_4x4_16bit_to_8bit)(uint8_t *dst, int dstride,
+                                       const uint16_t *src, int sstride) {
+  int i;
+  for (i = 0; i < 4; i++) {
+    v128 row = v128_load_unaligned(&src[i * sstride]);
+    row = v128_pack_s16_u8(row, row);
+    u32_store_unaligned(&dst[i * dstride], v128_low_u32(row));
+  }
+}
+
+void SIMD_FUNC(copy_8x8_16bit_to_16bit)(uint16_t *dst, int dstride,
+                                        const uint16_t *src, int sstride) {
+  int i;
+  for (i = 0; i < 8; i++) {
+    v128 row = v128_load_unaligned(&src[i * sstride]);
+    v128_store_unaligned(&dst[i * dstride], row);
+  }
+}
+
+void SIMD_FUNC(copy_4x4_16bit_to_16bit)(uint16_t *dst, int dstride,
+                                        const uint16_t *src, int sstride) {
+  int i;
+  for (i = 0; i < 4; i++) {
+    v64 row = v64_load_unaligned(&src[i * sstride]);
+    v64_store_unaligned(&dst[i * dstride], row);
+  }
+}
+
+void SIMD_FUNC(copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride,
+                                         const uint8_t *src, int sstride, int v,
+                                         int h) {
+  int i, j;
+  for (i = 0; i < v; i++) {
+    for (j = 0; j < (h & ~0x7); j += 8) {
+      v64 row = v64_load_unaligned(&src[i * sstride + j]);
+      v128_store_unaligned(&dst[i * dstride + j], v128_unpack_u8_s16(row));
+    }
+    for (; j < h; j++) {
+      dst[i * dstride + j] = src[i * sstride + j];
+    }
+  }
+}
+
+void SIMD_FUNC(copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride,
+                                          const uint16_t *src, int sstride,
+                                          int v, int h) {
+  int i, j;
+  for (i = 0; i < v; i++) {
+    for (j = 0; j < (h & ~0x7); j += 8) {
+      v128 row = v128_load_unaligned(&src[i * sstride + j]);
+      v128_store_unaligned(&dst[i * dstride + j], row);
+    }
+    for (; j < h; j++) {
+      dst[i * dstride + j] = src[i * sstride + j];
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/od_dering_sse2.c b/third_party/aom/av1/common/od_dering_sse2.c
new file mode 100644
index 0000000000..8a2a62f6c5
--- /dev/null
+++ b/third_party/aom/av1/common/od_dering_sse2.c
@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_sse2
+#include "./od_dering_simd.h"
diff --git a/third_party/aom/av1/common/od_dering_sse4.c b/third_party/aom/av1/common/od_dering_sse4.c
new file mode 100644
index 0000000000..0769db9fd9
--- /dev/null
+++ b/third_party/aom/av1/common/od_dering_sse4.c
@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_sse4_1
+#include "./od_dering_simd.h"
diff --git a/third_party/aom/av1/common/od_dering_ssse3.c b/third_party/aom/av1/common/od_dering_ssse3.c
new file mode 100644
index 0000000000..99df62b6b5
--- /dev/null
+++ b/third_party/aom/av1/common/od_dering_ssse3.c
@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_simd.h"
+#define SIMD_FUNC(name) name##_ssse3
+#include "./od_dering_simd.h"
diff --git a/third_party/aom/av1/common/odintrin.c b/third_party/aom/av1/common/odintrin.c
new file mode 100644
index 0000000000..868efacc99
--- /dev/null
+++ b/third_party/aom/av1/common/odintrin.c
@@ -0,0 +1,551 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#include "av1/common/odintrin.h"
+
+#if defined(OD_ENABLE_ASSERTIONS)
+# include <stdio.h>
+
+void od_fatal_impl(const char *_str, const char *_file, int _line) {
+  fprintf(stderr, "Fatal (internal) error in %s, line %d: %s\n",
+   _file, _line, _str);
+  abort();
+}
+#endif
+
+/*Constants for use with OD_DIVU_SMALL().
+  See \cite{Rob05} for details on computing these constants.
+  @INPROCEEDINGS{Rob05,
+    author="Arch D. Robison",
+    title="{N}-bit Unsigned Division via {N}-bit Multiply-Add",
+    booktitle="Proc. of the 17th IEEE Symposium on Computer Arithmetic
+     (ARITH'05)",
+    pages="131--139",
+    address="Cape Cod, MA",
+    month=Jun,
+    year=2005
+  }*/
+uint32_t OD_DIVU_SMALL_CONSTS[OD_DIVU_DMAX][2] = {
+  { 0xFFFFFFFF, 0xFFFFFFFF }, { 0xFFFFFFFF, 0xFFFFFFFF },
+  { 0xAAAAAAAB, 0 },          { 0xFFFFFFFF, 0xFFFFFFFF },
+  { 0xCCCCCCCD, 0 },          { 0xAAAAAAAB, 0 },
+  { 0x92492492, 0x92492492 }, { 0xFFFFFFFF, 0xFFFFFFFF },
+  { 0xE38E38E4, 0 },          { 0xCCCCCCCD, 0 },
+  { 0xBA2E8BA3, 0 },          { 0xAAAAAAAB, 0 },
+  { 0x9D89D89E, 0 },          { 0x92492492, 0x92492492 },
+  { 0x88888889, 0 },          { 0xFFFFFFFF, 0xFFFFFFFF },
+  { 0xF0F0F0F1, 0 },          { 0xE38E38E4, 0 },
+  { 0xD79435E5, 0xD79435E5 }, { 0xCCCCCCCD, 0 },
+  { 0xC30C30C3, 0xC30C30C3 }, { 0xBA2E8BA3, 0 },
+  { 0xB21642C9, 0 },          { 0xAAAAAAAB, 0 },
+  { 0xA3D70A3E, 0 },          { 0x9D89D89E, 0 },
+  { 0x97B425ED, 0x97B425ED }, { 0x92492492, 0x92492492 },
+  { 0x8D3DCB09, 0 },          { 0x88888889, 0 },
+  { 0x84210842, 0x84210842 }, { 0xFFFFFFFF, 0xFFFFFFFF },
+  { 0xF83E0F84, 0 },          { 0xF0F0F0F1, 0 },
+  { 0xEA0EA0EA, 0xEA0EA0EA }, { 0xE38E38E4, 0 },
+  { 0xDD67C8A6, 0xDD67C8A6 }, { 0xD79435E5, 0xD79435E5 },
+  { 0xD20D20D2, 0xD20D20D2 }, { 0xCCCCCCCD, 0 },
+  { 0xC7CE0C7D, 0 },          { 0xC30C30C3, 0xC30C30C3 },
+  { 0xBE82FA0C, 0 },          { 0xBA2E8BA3, 0 },
+  { 0xB60B60B6, 0xB60B60B6 }, { 0xB21642C9, 0 },
+  { 0xAE4C415D, 0 },          { 0xAAAAAAAB, 0 },
+  { 0xA72F053A, 0 },          { 0xA3D70A3E, 0 },
+  { 0xA0A0A0A1, 0 },          { 0x9D89D89E, 0 },
+  { 0x9A90E7D9, 0x9A90E7D9 }, { 0x97B425ED, 0x97B425ED },
+  { 0x94F2094F, 0x94F2094F }, { 0x92492492, 0x92492492 },
+  { 0x8FB823EE, 0x8FB823EE }, { 0x8D3DCB09, 0 },
+  { 0x8AD8F2FC, 0 },          { 0x88888889, 0 },
+  { 0x864B8A7E, 0 },          { 0x84210842, 0x84210842 },
+  { 0x82082082, 0x82082082 }, { 0xFFFFFFFF, 0xFFFFFFFF },
+  { 0xFC0FC0FD, 0 },          { 0xF83E0F84, 0 },
+  { 0xF4898D60, 0 },          { 0xF0F0F0F1, 0 },
+  { 0xED7303B6, 0 },          { 0xEA0EA0EA, 0xEA0EA0EA },
+  { 0xE6C2B449, 0 },          { 0xE38E38E4, 0 },
+  { 0xE070381C, 0xE070381C }, { 0xDD67C8A6, 0xDD67C8A6 },
+  { 0xDA740DA8, 0 },          { 0xD79435E5, 0xD79435E5 },
+  { 0xD4C77B04, 0 },          { 0xD20D20D2, 0xD20D20D2 },
+  { 0xCF6474A9, 0 },          { 0xCCCCCCCD, 0 },
+  { 0xCA4587E7, 0 },          { 0xC7CE0C7D, 0 },
+  { 0xC565C87C, 0 },          { 0xC30C30C3, 0xC30C30C3 },
+  { 0xC0C0C0C1, 0 },          { 0xBE82FA0C, 0 },
+  { 0xBC52640C, 0 },          { 0xBA2E8BA3, 0 },
+  { 0xB81702E1, 0 },          { 0xB60B60B6, 0xB60B60B6 },
+  { 0xB40B40B4, 0xB40B40B4 }, { 0xB21642C9, 0 },
+  { 0xB02C0B03, 0 },          { 0xAE4C415D, 0 },
+  { 0xAC769184, 0xAC769184 }, { 0xAAAAAAAB, 0 },
+  { 0xA8E83F57, 0xA8E83F57 }, { 0xA72F053A, 0 },
+  { 0xA57EB503, 0 },          { 0xA3D70A3E, 0 },
+  { 0xA237C32B, 0xA237C32B }, { 0xA0A0A0A1, 0 },
+  { 0x9F1165E7, 0x9F1165E7 }, { 0x9D89D89E, 0 },
+  { 0x9C09C09C, 0x9C09C09C }, { 0x9A90E7D9, 0x9A90E7D9 },
+  { 0x991F1A51, 0x991F1A51 }, { 0x97B425ED, 0x97B425ED },
+  { 0x964FDA6C, 0x964FDA6C }, { 0x94F2094F, 0x94F2094F },
+  { 0x939A85C4, 0x939A85C4 }, { 0x92492492, 0x92492492 },
+  { 0x90FDBC09, 0x90FDBC09 }, { 0x8FB823EE, 0x8FB823EE },
+  { 0x8E78356D, 0x8E78356D }, { 0x8D3DCB09, 0 },
+  { 0x8C08C08C, 0x8C08C08C }, { 0x8AD8F2FC, 0 },
+  { 0x89AE408A, 0 },          { 0x88888889, 0 },
+  { 0x8767AB5F, 0x8767AB5F }, { 0x864B8A7E, 0 },
+  { 0x85340853, 0x85340853 }, { 0x84210842, 0x84210842 },
+  { 0x83126E98, 0 },          { 0x82082082, 0x82082082 },
+  { 0x81020408, 0x81020408 }, { 0xFFFFFFFF, 0xFFFFFFFF },
+  { 0xFE03F810, 0 },          { 0xFC0FC0FD, 0 },
+  { 0xFA232CF3, 0 },          { 0xF83E0F84, 0 },
+  { 0xF6603D99, 0 },          { 0xF4898D60, 0 },
+  { 0xF2B9D649, 0 },          { 0xF0F0F0F1, 0 },
+  { 0xEF2EB720, 0 },          { 0xED7303B6, 0 },
+  { 0xEBBDB2A6, 0 },          { 0xEA0EA0EA, 0xEA0EA0EA },
+  { 0xE865AC7C, 0 },          { 0xE6C2B449, 0 },
+  { 0xE525982B, 0 },          { 0xE38E38E4, 0 },
+  { 0xE1FC780F, 0 },          { 0xE070381C, 0xE070381C },
+  { 0xDEE95C4D, 0 },          { 0xDD67C8A6, 0xDD67C8A6 },
+  { 0xDBEB61EF, 0 },          { 0xDA740DA8, 0 },
+  { 0xD901B204, 0 },          { 0xD79435E5, 0xD79435E5 },
+  { 0xD62B80D7, 0 },          { 0xD4C77B04, 0 },
+  { 0xD3680D37, 0 },          { 0xD20D20D2, 0xD20D20D2 },
+  { 0xD0B69FCC, 0 },          { 0xCF6474A9, 0 },
+  { 0xCE168A77, 0xCE168A77 }, { 0xCCCCCCCD, 0 },
+  { 0xCB8727C1, 0 },          { 0xCA4587E7, 0 },
+  { 0xC907DA4F, 0 },          { 0xC7CE0C7D, 0 },
+  { 0xC6980C6A, 0 },          { 0xC565C87C, 0 },
+  { 0xC4372F86, 0 },          { 0xC30C30C3, 0xC30C30C3 },
+  { 0xC1E4BBD6, 0 },          { 0xC0C0C0C1, 0 },
+  { 0xBFA02FE8, 0xBFA02FE8 }, { 0xBE82FA0C, 0 },
+  { 0xBD691047, 0xBD691047 }, { 0xBC52640C, 0 },
+  { 0xBB3EE722, 0 },          { 0xBA2E8BA3, 0 },
+  { 0xB92143FA, 0xB92143FA }, { 0xB81702E1, 0 },
+  { 0xB70FBB5A, 0xB70FBB5A }, { 0xB60B60B6, 0xB60B60B6 },
+  { 0xB509E68B, 0 },          { 0xB40B40B4, 0xB40B40B4 },
+  { 0xB30F6353, 0 },          { 0xB21642C9, 0 },
+  { 0xB11FD3B8, 0xB11FD3B8 }, { 0xB02C0B03, 0 },
+  { 0xAF3ADDC7, 0 },          { 0xAE4C415D, 0 },
+  { 0xAD602B58, 0xAD602B58 }, { 0xAC769184, 0xAC769184 },
+  { 0xAB8F69E3, 0 },          { 0xAAAAAAAB, 0 },
+  { 0xA9C84A48, 0 },          { 0xA8E83F57, 0xA8E83F57 },
+  { 0xA80A80A8, 0xA80A80A8 }, { 0xA72F053A, 0 },
+  { 0xA655C439, 0xA655C439 }, { 0xA57EB503, 0 },
+  { 0xA4A9CF1E, 0 },          { 0xA3D70A3E, 0 },
+  { 0xA3065E40, 0 },          { 0xA237C32B, 0xA237C32B },
+  { 0xA16B312F, 0 },          { 0xA0A0A0A1, 0 },
+  { 0x9FD809FE, 0 },          { 0x9F1165E7, 0x9F1165E7 },
+  { 0x9E4CAD24, 0 },          { 0x9D89D89E, 0 },
+  { 0x9CC8E161, 0 },          { 0x9C09C09C, 0x9C09C09C },
+  { 0x9B4C6F9F, 0 },          { 0x9A90E7D9, 0x9A90E7D9 },
+  { 0x99D722DB, 0 },          { 0x991F1A51, 0x991F1A51 },
+  { 0x9868C80A, 0 },          { 0x97B425ED, 0x97B425ED },
+  { 0x97012E02, 0x97012E02 }, { 0x964FDA6C, 0x964FDA6C },
+  { 0x95A02568, 0x95A02568 }, { 0x94F2094F, 0x94F2094F },
+  { 0x94458094, 0x94458094 }, { 0x939A85C4, 0x939A85C4 },
+  { 0x92F11384, 0x92F11384 }, { 0x92492492, 0x92492492 },
+  { 0x91A2B3C5, 0 },          { 0x90FDBC09, 0x90FDBC09 },
+  { 0x905A3863, 0x905A3863 }, { 0x8FB823EE, 0x8FB823EE },
+  { 0x8F1779DA, 0 },          { 0x8E78356D, 0x8E78356D },
+  { 0x8DDA5202, 0x8DDA5202 }, { 0x8D3DCB09, 0 },
+  { 0x8CA29C04, 0x8CA29C04 }, { 0x8C08C08C, 0x8C08C08C },
+  { 0x8B70344A, 0x8B70344A }, { 0x8AD8F2FC, 0 },
+  { 0x8A42F870, 0x8A42F870 }, { 0x89AE408A, 0 },
+  { 0x891AC73B, 0 },          { 0x88888889, 0 },
+  { 0x87F78088, 0 },          { 0x8767AB5F, 0x8767AB5F },
+  { 0x86D90545, 0 },          { 0x864B8A7E, 0 },
+  { 0x85BF3761, 0x85BF3761 }, { 0x85340853, 0x85340853 },
+  { 0x84A9F9C8, 0x84A9F9C8 }, { 0x84210842, 0x84210842 },
+  { 0x83993052, 0x83993052 }, { 0x83126E98, 0 },
+  { 0x828CBFBF, 0 },          { 0x82082082, 0x82082082 },
+  { 0x81848DA9, 0 },          { 0x81020408, 0x81020408 },
+  { 0x80808081, 0 },          { 0xFFFFFFFF, 0xFFFFFFFF },
+  { 0xFF00FF01, 0 },          { 0xFE03F810, 0 },
+  { 0xFD08E551, 0 },          { 0xFC0FC0FD, 0 },
+  { 0xFB188566, 0 },          { 0xFA232CF3, 0 },
+  { 0xF92FB222, 0 },          { 0xF83E0F84, 0 },
+  { 0xF74E3FC3, 0 },          { 0xF6603D99, 0 },
+  { 0xF57403D6, 0 },          { 0xF4898D60, 0 },
+  { 0xF3A0D52D, 0 },          { 0xF2B9D649, 0 },
+  { 0xF1D48BCF, 0 },          { 0xF0F0F0F1, 0 },
+  { 0xF00F00F0, 0xF00F00F0 }, { 0xEF2EB720, 0 },
+  { 0xEE500EE5, 0xEE500EE5 }, { 0xED7303B6, 0 },
+  { 0xEC979119, 0 },          { 0xEBBDB2A6, 0 },
+  { 0xEAE56404, 0 },          { 0xEA0EA0EA, 0xEA0EA0EA },
+  { 0xE9396520, 0 },          { 0xE865AC7C, 0 },
+  { 0xE79372E3, 0 },          { 0xE6C2B449, 0 },
+  { 0xE5F36CB0, 0xE5F36CB0 }, { 0xE525982B, 0 },
+  { 0xE45932D8, 0 },          { 0xE38E38E4, 0 },
+  { 0xE2C4A689, 0 },          { 0xE1FC780F, 0 },
+  { 0xE135A9CA, 0 },          { 0xE070381C, 0xE070381C },
+  { 0xDFAC1F75, 0 },          { 0xDEE95C4D, 0 },
+  { 0xDE27EB2D, 0 },          { 0xDD67C8A6, 0xDD67C8A6 },
+  { 0xDCA8F159, 0 },          { 0xDBEB61EF, 0 },
+  { 0xDB2F171E, 0 },          { 0xDA740DA8, 0 },
+  { 0xD9BA4257, 0 },          { 0xD901B204, 0 },
+  { 0xD84A598F, 0 },          { 0xD79435E5, 0xD79435E5 },
+  { 0xD6DF43FD, 0 },          { 0xD62B80D7, 0 },
+  { 0xD578E97D, 0 },          { 0xD4C77B04, 0 },
+  { 0xD417328A, 0 },          { 0xD3680D37, 0 },
+  { 0xD2BA083C, 0 },          { 0xD20D20D2, 0xD20D20D2 },
+  { 0xD161543E, 0xD161543E }, { 0xD0B69FCC, 0 },
+  { 0xD00D00D0, 0xD00D00D0 }, { 0xCF6474A9, 0 },
+  { 0xCEBCF8BC, 0 },          { 0xCE168A77, 0xCE168A77 },
+  { 0xCD712753, 0 },          { 0xCCCCCCCD, 0 },
+  { 0xCC29786D, 0 },          { 0xCB8727C1, 0 },
+  { 0xCAE5D85F, 0xCAE5D85F }, { 0xCA4587E7, 0 },
+  { 0xC9A633FD, 0 },          { 0xC907DA4F, 0 },
+  { 0xC86A7890, 0xC86A7890 }, { 0xC7CE0C7D, 0 },
+  { 0xC73293D8, 0 },          { 0xC6980C6A, 0 },
+  { 0xC5FE7403, 0xC5FE7403 }, { 0xC565C87C, 0 },
+  { 0xC4CE07B0, 0xC4CE07B0 }, { 0xC4372F86, 0 },
+  { 0xC3A13DE6, 0xC3A13DE6 }, { 0xC30C30C3, 0xC30C30C3 },
+  { 0xC2780614, 0 },          { 0xC1E4BBD6, 0 },
+  { 0xC152500C, 0xC152500C }, { 0xC0C0C0C1, 0 },
+  { 0xC0300C03, 0xC0300C03 }, { 0xBFA02FE8, 0xBFA02FE8 },
+  { 0xBF112A8B, 0 },          { 0xBE82FA0C, 0 },
+  { 0xBDF59C92, 0 },          { 0xBD691047, 0xBD691047 },
+  { 0xBCDD535E, 0 },          { 0xBC52640C, 0 },
+  { 0xBBC8408D, 0 },          { 0xBB3EE722, 0 },
+  { 0xBAB65610, 0xBAB65610 }, { 0xBA2E8BA3, 0 },
+  { 0xB9A7862A, 0xB9A7862A }, { 0xB92143FA, 0xB92143FA },
+  { 0xB89BC36D, 0 },          { 0xB81702E1, 0 },
+  { 0xB79300B8, 0 },          { 0xB70FBB5A, 0xB70FBB5A },
+  { 0xB68D3134, 0xB68D3134 }, { 0xB60B60B6, 0xB60B60B6 },
+  { 0xB58A4855, 0xB58A4855 }, { 0xB509E68B, 0 },
+  { 0xB48A39D4, 0xB48A39D4 }, { 0xB40B40B4, 0xB40B40B4 },
+  { 0xB38CF9B0, 0xB38CF9B0 }, { 0xB30F6353, 0 },
+  { 0xB2927C2A, 0 },          { 0xB21642C9, 0 },
+  { 0xB19AB5C5, 0 },          { 0xB11FD3B8, 0xB11FD3B8 },
+  { 0xB0A59B42, 0 },          { 0xB02C0B03, 0 },
+  { 0xAFB321A1, 0xAFB321A1 }, { 0xAF3ADDC7, 0 },
+  { 0xAEC33E20, 0 },          { 0xAE4C415D, 0 },
+  { 0xADD5E632, 0xADD5E632 }, { 0xAD602B58, 0xAD602B58 },
+  { 0xACEB0F89, 0xACEB0F89 }, { 0xAC769184, 0xAC769184 },
+  { 0xAC02B00B, 0 },          { 0xAB8F69E3, 0 },
+  { 0xAB1CBDD4, 0 },          { 0xAAAAAAAB, 0 },
+  { 0xAA392F36, 0 },          { 0xA9C84A48, 0 },
+  { 0xA957FAB5, 0xA957FAB5 }, { 0xA8E83F57, 0xA8E83F57 },
+  { 0xA8791709, 0 },          { 0xA80A80A8, 0xA80A80A8 },
+  { 0xA79C7B17, 0 },          { 0xA72F053A, 0 },
+  { 0xA6C21DF7, 0 },          { 0xA655C439, 0xA655C439 },
+  { 0xA5E9F6ED, 0xA5E9F6ED }, { 0xA57EB503, 0 },
+  { 0xA513FD6C, 0 },          { 0xA4A9CF1E, 0 },
+  { 0xA4402910, 0xA4402910 }, { 0xA3D70A3E, 0 },
+  { 0xA36E71A3, 0 },          { 0xA3065E40, 0 },
+  { 0xA29ECF16, 0xA29ECF16 }, { 0xA237C32B, 0xA237C32B },
+  { 0xA1D13986, 0 },          { 0xA16B312F, 0 },
+  { 0xA105A933, 0 },          { 0xA0A0A0A1, 0 },
+  { 0xA03C1689, 0 },          { 0x9FD809FE, 0 },
+  { 0x9F747A15, 0x9F747A15 }, { 0x9F1165E7, 0x9F1165E7 },
+  { 0x9EAECC8D, 0x9EAECC8D }, { 0x9E4CAD24, 0 },
+  { 0x9DEB06C9, 0x9DEB06C9 }, { 0x9D89D89E, 0 },
+  { 0x9D2921C4, 0 },          { 0x9CC8E161, 0 },
+  { 0x9C69169B, 0x9C69169B }, { 0x9C09C09C, 0x9C09C09C },
+  { 0x9BAADE8E, 0x9BAADE8E }, { 0x9B4C6F9F, 0 },
+  { 0x9AEE72FD, 0 },          { 0x9A90E7D9, 0x9A90E7D9 },
+  { 0x9A33CD67, 0x9A33CD67 }, { 0x99D722DB, 0 },
+  { 0x997AE76B, 0x997AE76B }, { 0x991F1A51, 0x991F1A51 },
+  { 0x98C3BAC7, 0x98C3BAC7 }, { 0x9868C80A, 0 },
+  { 0x980E4156, 0x980E4156 }, { 0x97B425ED, 0x97B425ED },
+  { 0x975A7510, 0 },          { 0x97012E02, 0x97012E02 },
+  { 0x96A8500A, 0 },          { 0x964FDA6C, 0x964FDA6C },
+  { 0x95F7CC73, 0 },          { 0x95A02568, 0x95A02568 },
+  { 0x9548E498, 0 },          { 0x94F2094F, 0x94F2094F },
+  { 0x949B92DE, 0 },          { 0x94458094, 0x94458094 },
+  { 0x93EFD1C5, 0x93EFD1C5 }, { 0x939A85C4, 0x939A85C4 },
+  { 0x93459BE7, 0 },          { 0x92F11384, 0x92F11384 },
+  { 0x929CEBF5, 0 },          { 0x92492492, 0x92492492 },
+  { 0x91F5BCB9, 0 },          { 0x91A2B3C5, 0 },
+  { 0x91500915, 0x91500915 }, { 0x90FDBC09, 0x90FDBC09 },
+  { 0x90ABCC02, 0x90ABCC02 }, { 0x905A3863, 0x905A3863 },
+  { 0x90090090, 0x90090090 }, { 0x8FB823EE, 0x8FB823EE },
+  { 0x8F67A1E4, 0 },          { 0x8F1779DA, 0 },
+  { 0x8EC7AB3A, 0 },          { 0x8E78356D, 0x8E78356D },
+  { 0x8E2917E1, 0 },          { 0x8DDA5202, 0x8DDA5202 },
+  { 0x8D8BE340, 0 },          { 0x8D3DCB09, 0 },
+  { 0x8CF008CF, 0x8CF008CF }, { 0x8CA29C04, 0x8CA29C04 },
+  { 0x8C55841D, 0 },          { 0x8C08C08C, 0x8C08C08C },
+  { 0x8BBC50C9, 0 },          { 0x8B70344A, 0x8B70344A },
+  { 0x8B246A88, 0 },          { 0x8AD8F2FC, 0 },
+  { 0x8A8DCD20, 0 },          { 0x8A42F870, 0x8A42F870 },
+  { 0x89F8746A, 0 },          { 0x89AE408A, 0 },
+  { 0x89645C4F, 0x89645C4F }, { 0x891AC73B, 0 },
+  { 0x88D180CD, 0x88D180CD }, { 0x88888889, 0 },
+  { 0x883FDDF0, 0x883FDDF0 }, { 0x87F78088, 0 },
+  { 0x87AF6FD6, 0 },          { 0x8767AB5F, 0x8767AB5F },
+  { 0x872032AC, 0x872032AC }, { 0x86D90545, 0 },
+  { 0x869222B2, 0 },          { 0x864B8A7E, 0 },
+  { 0x86053C34, 0x86053C34 }, { 0x85BF3761, 0x85BF3761 },
+  { 0x85797B91, 0x85797B91 }, { 0x85340853, 0x85340853 },
+  { 0x84EEDD36, 0 },          { 0x84A9F9C8, 0x84A9F9C8 },
+  { 0x84655D9C, 0 },          { 0x84210842, 0x84210842 },
+  { 0x83DCF94E, 0 },          { 0x83993052, 0x83993052 },
+  { 0x8355ACE4, 0 },          { 0x83126E98, 0 },
+  { 0x82CF7504, 0 },          { 0x828CBFBF, 0 },
+  { 0x824A4E61, 0 },          { 0x82082082, 0x82082082 },
+  { 0x81C635BC, 0x81C635BC }, { 0x81848DA9, 0 },
+  { 0x814327E4, 0 },          { 0x81020408, 0x81020408 },
+  { 0x80C121B3, 0 },          { 0x80808081, 0 },
+  { 0x80402010, 0x80402010 }, { 0xFFFFFFFF, 0xFFFFFFFF },
+  { 0xFF803FE1, 0 },          { 0xFF00FF01, 0 },
+  { 0xFE823CA6, 0 },          { 0xFE03F810, 0 },
+  { 0xFD863087, 0 },          { 0xFD08E551, 0 },
+  { 0xFC8C15B5, 0 },          { 0xFC0FC0FD, 0 },
+  { 0xFB93E673, 0 },          { 0xFB188566, 0 },
+  { 0xFA9D9D20, 0 },          { 0xFA232CF3, 0 },
+  { 0xF9A9342D, 0 },          { 0xF92FB222, 0 },
+  { 0xF8B6A622, 0xF8B6A622 }, { 0xF83E0F84, 0 },
+  { 0xF7C5ED9D, 0 },          { 0xF74E3FC3, 0 },
+  { 0xF6D7054E, 0 },          { 0xF6603D99, 0 },
+  { 0xF5E9E7FD, 0 },          { 0xF57403D6, 0 },
+  { 0xF4FE9083, 0 },          { 0xF4898D60, 0 },
+  { 0xF414F9CE, 0 },          { 0xF3A0D52D, 0 },
+  { 0xF32D1EE0, 0 },          { 0xF2B9D649, 0 },
+  { 0xF246FACC, 0 },          { 0xF1D48BCF, 0 },
+  { 0xF16288B9, 0 },          { 0xF0F0F0F1, 0 },
+  { 0xF07FC3E0, 0xF07FC3E0 }, { 0xF00F00F0, 0xF00F00F0 },
+  { 0xEF9EA78C, 0 },          { 0xEF2EB720, 0 },
+  { 0xEEBF2F19, 0 },          { 0xEE500EE5, 0xEE500EE5 },
+  { 0xEDE155F4, 0 },          { 0xED7303B6, 0 },
+  { 0xED05179C, 0xED05179C }, { 0xEC979119, 0 },
+  { 0xEC2A6FA0, 0xEC2A6FA0 }, { 0xEBBDB2A6, 0 },
+  { 0xEB5159A0, 0 },          { 0xEAE56404, 0 },
+  { 0xEA79D14A, 0 },          { 0xEA0EA0EA, 0xEA0EA0EA },
+  { 0xE9A3D25E, 0xE9A3D25E }, { 0xE9396520, 0 },
+  { 0xE8CF58AB, 0 },          { 0xE865AC7C, 0 },
+  { 0xE7FC600F, 0 },          { 0xE79372E3, 0 },
+  { 0xE72AE476, 0 },          { 0xE6C2B449, 0 },
+  { 0xE65AE1DC, 0 },          { 0xE5F36CB0, 0xE5F36CB0 },
+  { 0xE58C544A, 0 },          { 0xE525982B, 0 },
+  { 0xE4BF37D9, 0 },          { 0xE45932D8, 0 },
+  { 0xE3F388AF, 0 },          { 0xE38E38E4, 0 },
+  { 0xE32942FF, 0 },          { 0xE2C4A689, 0 },
+  { 0xE260630B, 0 },          { 0xE1FC780F, 0 },
+  { 0xE198E520, 0 },          { 0xE135A9CA, 0 },
+  { 0xE0D2C59A, 0 },          { 0xE070381C, 0xE070381C },
+  { 0xE00E00E0, 0xE00E00E0 }, { 0xDFAC1F75, 0 },
+  { 0xDF4A9369, 0 },          { 0xDEE95C4D, 0 },
+  { 0xDE8879B3, 0 },          { 0xDE27EB2D, 0 },
+  { 0xDDC7B04D, 0 },          { 0xDD67C8A6, 0xDD67C8A6 },
+  { 0xDD0833CE, 0 },          { 0xDCA8F159, 0 },
+  { 0xDC4A00DD, 0 },          { 0xDBEB61EF, 0 },
+  { 0xDB8D1428, 0 },          { 0xDB2F171E, 0 },
+  { 0xDAD16A6B, 0 },          { 0xDA740DA8, 0 },
+  { 0xDA17006D, 0xDA17006D }, { 0xD9BA4257, 0 },
+  { 0xD95DD300, 0 },          { 0xD901B204, 0 },
+  { 0xD8A5DEFF, 0 },          { 0xD84A598F, 0 },
+  { 0xD7EF2152, 0 },          { 0xD79435E5, 0xD79435E5 },
+  { 0xD73996E9, 0 },          { 0xD6DF43FD, 0 },
+  { 0xD6853CC1, 0 },          { 0xD62B80D7, 0 },
+  { 0xD5D20FDF, 0 },          { 0xD578E97D, 0 },
+  { 0xD5200D52, 0xD5200D52 }, { 0xD4C77B04, 0 },
+  { 0xD46F3235, 0 },          { 0xD417328A, 0 },
+  { 0xD3BF7BA9, 0 },          { 0xD3680D37, 0 },
+  { 0xD310E6DB, 0 },          { 0xD2BA083C, 0 },
+  { 0xD2637101, 0 },          { 0xD20D20D2, 0xD20D20D2 },
+  { 0xD1B71759, 0 },          { 0xD161543E, 0xD161543E },
+  { 0xD10BD72C, 0 },          { 0xD0B69FCC, 0 },
+  { 0xD061ADCA, 0 },          { 0xD00D00D0, 0xD00D00D0 },
+  { 0xCFB8988C, 0 },          { 0xCF6474A9, 0 },
+  { 0xCF1094D4, 0 },          { 0xCEBCF8BC, 0 },
+  { 0xCE69A00D, 0 },          { 0xCE168A77, 0xCE168A77 },
+  { 0xCDC3B7A9, 0xCDC3B7A9 }, { 0xCD712753, 0 },
+  { 0xCD1ED924, 0 },          { 0xCCCCCCCD, 0 },
+  { 0xCC7B0200, 0 },          { 0xCC29786D, 0 },
+  { 0xCBD82FC7, 0 },          { 0xCB8727C1, 0 },
+  { 0xCB36600D, 0 },          { 0xCAE5D85F, 0xCAE5D85F },
+  { 0xCA95906C, 0 },          { 0xCA4587E7, 0 },
+  { 0xC9F5BE86, 0 },          { 0xC9A633FD, 0 },
+  { 0xC956E803, 0xC956E803 }, { 0xC907DA4F, 0 },
+  { 0xC8B90A96, 0 },          { 0xC86A7890, 0xC86A7890 },
+  { 0xC81C23F5, 0xC81C23F5 }, { 0xC7CE0C7D, 0 },
+  { 0xC78031E0, 0xC78031E0 }, { 0xC73293D8, 0 },
+  { 0xC6E5321D, 0 },          { 0xC6980C6A, 0 },
+  { 0xC64B2278, 0xC64B2278 }, { 0xC5FE7403, 0xC5FE7403 },
+  { 0xC5B200C6, 0 },          { 0xC565C87C, 0 },
+  { 0xC519CAE0, 0xC519CAE0 }, { 0xC4CE07B0, 0xC4CE07B0 },
+  { 0xC4827EA8, 0xC4827EA8 }, { 0xC4372F86, 0 },
+  { 0xC3EC1A06, 0 },          { 0xC3A13DE6, 0xC3A13DE6 },
+  { 0xC3569AE6, 0 },          { 0xC30C30C3, 0xC30C30C3 },
+  { 0xC2C1FF3E, 0 },          { 0xC2780614, 0 },
+  { 0xC22E4507, 0 },          { 0xC1E4BBD6, 0 },
+  { 0xC19B6A42, 0 },          { 0xC152500C, 0xC152500C },
+  { 0xC1096CF6, 0 },          { 0xC0C0C0C1, 0 },
+  { 0xC0784B2F, 0 },          { 0xC0300C03, 0xC0300C03 },
+  { 0xBFE80300, 0 },          { 0xBFA02FE8, 0xBFA02FE8 },
+  { 0xBF589280, 0 },          { 0xBF112A8B, 0 },
+  { 0xBEC9F7CE, 0 },          { 0xBE82FA0C, 0 },
+  { 0xBE3C310C, 0 },          { 0xBDF59C92, 0 },
+  { 0xBDAF3C64, 0 },          { 0xBD691047, 0xBD691047 },
+  { 0xBD231803, 0 },          { 0xBCDD535E, 0 },
+  { 0xBC97C21E, 0xBC97C21E }, { 0xBC52640C, 0 },
+  { 0xBC0D38EE, 0xBC0D38EE }, { 0xBBC8408D, 0 },
+  { 0xBB837AB1, 0 },          { 0xBB3EE722, 0 },
+  { 0xBAFA85A9, 0xBAFA85A9 }, { 0xBAB65610, 0xBAB65610 },
+  { 0xBA725820, 0xBA725820 }, { 0xBA2E8BA3, 0 },
+  { 0xB9EAF063, 0 },          { 0xB9A7862A, 0xB9A7862A },
+  { 0xB9644CC4, 0 },          { 0xB92143FA, 0xB92143FA },
+  { 0xB8DE6B9A, 0 },          { 0xB89BC36D, 0 },
+  { 0xB8594B41, 0 },          { 0xB81702E1, 0 },
+  { 0xB7D4EA19, 0xB7D4EA19 }, { 0xB79300B8, 0 },
+  { 0xB7514689, 0 },          { 0xB70FBB5A, 0xB70FBB5A },
+  { 0xB6CE5EF9, 0xB6CE5EF9 }, { 0xB68D3134, 0xB68D3134 },
+  { 0xB64C31D9, 0 },          { 0xB60B60B6, 0xB60B60B6 },
+  { 0xB5CABD9B, 0 },          { 0xB58A4855, 0xB58A4855 },
+  { 0xB54A00B5, 0xB54A00B5 }, { 0xB509E68B, 0 },
+  { 0xB4C9F9A5, 0 },          { 0xB48A39D4, 0xB48A39D4 },
+  { 0xB44AA6E9, 0xB44AA6E9 }, { 0xB40B40B4, 0xB40B40B4 },
+  { 0xB3CC0706, 0 },          { 0xB38CF9B0, 0xB38CF9B0 },
+  { 0xB34E1884, 0 },          { 0xB30F6353, 0 },
+  { 0xB2D0D9EF, 0 },          { 0xB2927C2A, 0 },
+  { 0xB25449D7, 0 },          { 0xB21642C9, 0 },
+  { 0xB1D866D1, 0xB1D866D1 }, { 0xB19AB5C5, 0 },
+  { 0xB15D2F76, 0 },          { 0xB11FD3B8, 0xB11FD3B8 },
+  { 0xB0E2A260, 0xB0E2A260 }, { 0xB0A59B42, 0 },
+  { 0xB068BE31, 0 },          { 0xB02C0B03, 0 },
+  { 0xAFEF818C, 0 },          { 0xAFB321A1, 0xAFB321A1 },
+  { 0xAF76EB19, 0 },          { 0xAF3ADDC7, 0 },
+  { 0xAEFEF982, 0 },          { 0xAEC33E20, 0 },
+  { 0xAE87AB76, 0xAE87AB76 }, { 0xAE4C415D, 0 },
+  { 0xAE10FFA9, 0 },          { 0xADD5E632, 0xADD5E632 },
+  { 0xAD9AF4D0, 0 },          { 0xAD602B58, 0xAD602B58 },
+  { 0xAD2589A4, 0 },          { 0xACEB0F89, 0xACEB0F89 },
+  { 0xACB0BCE1, 0xACB0BCE1 }, { 0xAC769184, 0xAC769184 },
+  { 0xAC3C8D4A, 0 },          { 0xAC02B00B, 0 },
+  { 0xABC8F9A0, 0xABC8F9A0 }, { 0xAB8F69E3, 0 },
+  { 0xAB5600AC, 0 },          { 0xAB1CBDD4, 0 },
+  { 0xAAE3A136, 0 },          { 0xAAAAAAAB, 0 },
+  { 0xAA71DA0D, 0 },          { 0xAA392F36, 0 },
+  { 0xAA00AA01, 0 },          { 0xA9C84A48, 0 },
+  { 0xA9900FE6, 0 },          { 0xA957FAB5, 0xA957FAB5 },
+  { 0xA9200A92, 0xA9200A92 }, { 0xA8E83F57, 0xA8E83F57 },
+  { 0xA8B098E0, 0xA8B098E0 }, { 0xA8791709, 0 },
+  { 0xA841B9AD, 0 },          { 0xA80A80A8, 0xA80A80A8 },
+  { 0xA7D36BD8, 0 },          { 0xA79C7B17, 0 },
+  { 0xA765AE44, 0 },          { 0xA72F053A, 0 },
+  { 0xA6F87FD6, 0xA6F87FD6 }, { 0xA6C21DF7, 0 },
+  { 0xA68BDF79, 0 },          { 0xA655C439, 0xA655C439 },
+  { 0xA61FCC16, 0xA61FCC16 }, { 0xA5E9F6ED, 0xA5E9F6ED },
+  { 0xA5B4449D, 0 },          { 0xA57EB503, 0 },
+  { 0xA54947FE, 0 },          { 0xA513FD6C, 0 },
+  { 0xA4DED52C, 0xA4DED52C }, { 0xA4A9CF1E, 0 },
+  { 0xA474EB1F, 0xA474EB1F }, { 0xA4402910, 0xA4402910 },
+  { 0xA40B88D0, 0 },          { 0xA3D70A3E, 0 },
+  { 0xA3A2AD39, 0xA3A2AD39 }, { 0xA36E71A3, 0 },
+  { 0xA33A575A, 0xA33A575A }, { 0xA3065E40, 0 },
+  { 0xA2D28634, 0 },          { 0xA29ECF16, 0xA29ECF16 },
+  { 0xA26B38C9, 0 },          { 0xA237C32B, 0xA237C32B },
+  { 0xA2046E1F, 0xA2046E1F }, { 0xA1D13986, 0 },
+  { 0xA19E2540, 0 },          { 0xA16B312F, 0 },
+  { 0xA1385D35, 0 },          { 0xA105A933, 0 },
+  { 0xA0D3150C, 0 },          { 0xA0A0A0A1, 0 },
+  { 0xA06E4BD4, 0xA06E4BD4 }, { 0xA03C1689, 0 },
+  { 0xA00A00A0, 0xA00A00A0 }, { 0x9FD809FE, 0 },
+  { 0x9FA63284, 0 },          { 0x9F747A15, 0x9F747A15 },
+  { 0x9F42E095, 0x9F42E095 }, { 0x9F1165E7, 0x9F1165E7 },
+  { 0x9EE009EE, 0x9EE009EE }, { 0x9EAECC8D, 0x9EAECC8D },
+  { 0x9E7DADA9, 0 },          { 0x9E4CAD24, 0 },
+  { 0x9E1BCAE3, 0 },          { 0x9DEB06C9, 0x9DEB06C9 },
+  { 0x9DBA60BB, 0x9DBA60BB }, { 0x9D89D89E, 0 },
+  { 0x9D596E54, 0x9D596E54 }, { 0x9D2921C4, 0 },
+  { 0x9CF8F2D1, 0x9CF8F2D1 }, { 0x9CC8E161, 0 },
+  { 0x9C98ED58, 0 },          { 0x9C69169B, 0x9C69169B },
+  { 0x9C395D10, 0x9C395D10 }, { 0x9C09C09C, 0x9C09C09C },
+  { 0x9BDA4124, 0x9BDA4124 }, { 0x9BAADE8E, 0x9BAADE8E },
+  { 0x9B7B98C0, 0 },          { 0x9B4C6F9F, 0 },
+  { 0x9B1D6311, 0x9B1D6311 }, { 0x9AEE72FD, 0 },
+  { 0x9ABF9F48, 0x9ABF9F48 }, { 0x9A90E7D9, 0x9A90E7D9 },
+  { 0x9A624C97, 0 },          { 0x9A33CD67, 0x9A33CD67 },
+  { 0x9A056A31, 0 },          { 0x99D722DB, 0 },
+  { 0x99A8F74C, 0 },          { 0x997AE76B, 0x997AE76B },
+  { 0x994CF320, 0x994CF320 }, { 0x991F1A51, 0x991F1A51 },
+  { 0x98F15CE7, 0 },          { 0x98C3BAC7, 0x98C3BAC7 },
+  { 0x989633DB, 0x989633DB }, { 0x9868C80A, 0 },
+  { 0x983B773B, 0 },          { 0x980E4156, 0x980E4156 },
+  { 0x97E12644, 0x97E12644 }, { 0x97B425ED, 0x97B425ED },
+  { 0x97874039, 0 },          { 0x975A7510, 0 },
+  { 0x972DC45B, 0 },          { 0x97012E02, 0x97012E02 },
+  { 0x96D4B1EF, 0 },          { 0x96A8500A, 0 },
+  { 0x967C083B, 0 },          { 0x964FDA6C, 0x964FDA6C },
+  { 0x9623C686, 0x9623C686 }, { 0x95F7CC73, 0 },
+  { 0x95CBEC1B, 0 },          { 0x95A02568, 0x95A02568 },
+  { 0x95747844, 0 },          { 0x9548E498, 0 },
+  { 0x951D6A4E, 0 },          { 0x94F2094F, 0x94F2094F },
+  { 0x94C6C187, 0 },          { 0x949B92DE, 0 },
+  { 0x94707D3F, 0 },          { 0x94458094, 0x94458094 },
+  { 0x941A9CC8, 0x941A9CC8 }, { 0x93EFD1C5, 0x93EFD1C5 },
+  { 0x93C51F76, 0 },          { 0x939A85C4, 0x939A85C4 },
+  { 0x9370049C, 0 },          { 0x93459BE7, 0 },
+  { 0x931B4B91, 0 },          { 0x92F11384, 0x92F11384 },
+  { 0x92C6F3AC, 0x92C6F3AC }, { 0x929CEBF5, 0 },
+  { 0x9272FC48, 0x9272FC48 }, { 0x92492492, 0x92492492 },
+  { 0x921F64BF, 0 },          { 0x91F5BCB9, 0 },
+  { 0x91CC2C6C, 0x91CC2C6C }, { 0x91A2B3C5, 0 },
+  { 0x917952AF, 0 },          { 0x91500915, 0x91500915 },
+  { 0x9126D6E5, 0 },          { 0x90FDBC09, 0x90FDBC09 },
+  { 0x90D4B86F, 0 },          { 0x90ABCC02, 0x90ABCC02 },
+  { 0x9082F6B0, 0 },          { 0x905A3863, 0x905A3863 },
+  { 0x9031910A, 0 },          { 0x90090090, 0x90090090 },
+  { 0x8FE086E3, 0 },          { 0x8FB823EE, 0x8FB823EE },
+  { 0x8F8FD7A0, 0 },          { 0x8F67A1E4, 0 },
+  { 0x8F3F82A8, 0x8F3F82A8 }, { 0x8F1779DA, 0 },
+  { 0x8EEF8766, 0 },          { 0x8EC7AB3A, 0 },
+  { 0x8E9FE542, 0x8E9FE542 }, { 0x8E78356D, 0x8E78356D },
+  { 0x8E509BA8, 0x8E509BA8 }, { 0x8E2917E1, 0 },
+  { 0x8E01AA05, 0 },          { 0x8DDA5202, 0x8DDA5202 },
+  { 0x8DB30FC6, 0x8DB30FC6 }, { 0x8D8BE340, 0 },
+  { 0x8D64CC5C, 0 },          { 0x8D3DCB09, 0 },
+  { 0x8D16DF35, 0x8D16DF35 }, { 0x8CF008CF, 0x8CF008CF },
+  { 0x8CC947C5, 0 },          { 0x8CA29C04, 0x8CA29C04 },
+  { 0x8C7C057D, 0 },          { 0x8C55841D, 0 },
+  { 0x8C2F17D2, 0x8C2F17D2 }, { 0x8C08C08C, 0x8C08C08C },
+  { 0x8BE27E39, 0x8BE27E39 }, { 0x8BBC50C9, 0 },
+  { 0x8B963829, 0x8B963829 }, { 0x8B70344A, 0x8B70344A },
+  { 0x8B4A451A, 0 },          { 0x8B246A88, 0 },
+  { 0x8AFEA483, 0x8AFEA483 }, { 0x8AD8F2FC, 0 },
+  { 0x8AB355E0, 0x8AB355E0 }, { 0x8A8DCD20, 0 },
+  { 0x8A6858AB, 0 },          { 0x8A42F870, 0x8A42F870 },
+  { 0x8A1DAC60, 0x8A1DAC60 }, { 0x89F8746A, 0 },
+  { 0x89D3507D, 0 },          { 0x89AE408A, 0 },
+  { 0x89894480, 0 },          { 0x89645C4F, 0x89645C4F },
+  { 0x893F87E8, 0x893F87E8 }, { 0x891AC73B, 0 },
+  { 0x88F61A37, 0x88F61A37 }, { 0x88D180CD, 0x88D180CD },
+  { 0x88ACFAEE, 0 },          { 0x88888889, 0 },
+  { 0x8864298F, 0 },          { 0x883FDDF0, 0x883FDDF0 },
+  { 0x881BA59E, 0 },          { 0x87F78088, 0 },
+  { 0x87D36EA0, 0 },          { 0x87AF6FD6, 0 },
+  { 0x878B841B, 0 },          { 0x8767AB5F, 0x8767AB5F },
+  { 0x8743E595, 0 },          { 0x872032AC, 0x872032AC },
+  { 0x86FC9296, 0x86FC9296 }, { 0x86D90545, 0 },
+  { 0x86B58AA8, 0 },          { 0x869222B2, 0 },
+  { 0x866ECD53, 0x866ECD53 }, { 0x864B8A7E, 0 },
+  { 0x86285A23, 0x86285A23 }, { 0x86053C34, 0x86053C34 },
+  { 0x85E230A3, 0x85E230A3 }, { 0x85BF3761, 0x85BF3761 },
+  { 0x859C5060, 0x859C5060 }, { 0x85797B91, 0x85797B91 },
+  { 0x8556B8E7, 0x8556B8E7 }, { 0x85340853, 0x85340853 },
+  { 0x851169C7, 0x851169C7 }, { 0x84EEDD36, 0 },
+  { 0x84CC6290, 0 },          { 0x84A9F9C8, 0x84A9F9C8 },
+  { 0x8487A2D1, 0 },          { 0x84655D9C, 0 },
+  { 0x84432A1B, 0x84432A1B }, { 0x84210842, 0x84210842 },
+  { 0x83FEF802, 0x83FEF802 }, { 0x83DCF94E, 0 },
+  { 0x83BB0C18, 0 },          { 0x83993052, 0x83993052 },
+  { 0x837765F0, 0x837765F0 }, { 0x8355ACE4, 0 },
+  { 0x83340520, 0x83340520 }, { 0x83126E98, 0 },
+  { 0x82F0E93D, 0x82F0E93D }, { 0x82CF7504, 0 },
+  { 0x82AE11DE, 0 },          { 0x828CBFBF, 0 },
+  { 0x826B7E99, 0x826B7E99 }, { 0x824A4E61, 0 },
+  { 0x82292F08, 0 },          { 0x82082082, 0x82082082 },
+  { 0x81E722C2, 0x81E722C2 }, { 0x81C635BC, 0x81C635BC },
+  { 0x81A55963, 0 },          { 0x81848DA9, 0 },
+  { 0x8163D283, 0 },          { 0x814327E4, 0 },
+  { 0x81228DBF, 0 },          { 0x81020408, 0x81020408 },
+  { 0x80E18AB3, 0 },          { 0x80C121B3, 0 },
+  { 0x80A0C8FB, 0x80A0C8FB }, { 0x80808081, 0 },
+  { 0x80604836, 0x80604836 }, { 0x80402010, 0x80402010 },
+  { 0x80200802, 0x80200802 }, { 0xFFFFFFFF, 0xFFFFFFFF }
+};
diff --git a/third_party/aom/av1/common/odintrin.h b/third_party/aom/av1/common/odintrin.h
new file mode 100644
index 0000000000..fe99d80039
--- /dev/null
+++ b/third_party/aom/av1/common/odintrin.h
@@ -0,0 +1,267 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifndef AV1_COMMON_ODINTRIN_H_
+#define AV1_COMMON_ODINTRIN_H_
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/bitops.h"
+#include "av1/common/enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+# if !defined(M_PI)
+#  define M_PI      (3.1415926535897932384626433832795)
+# endif
+
+# if !defined(M_SQRT2)
+#  define M_SQRT2 (1.41421356237309504880168872420970)
+# endif
+
+# if !defined(M_SQRT1_2)
+#  define M_SQRT1_2 (0.70710678118654752440084436210485)
+# endif
+
+# if !defined(M_LOG2E)
+#  define M_LOG2E (1.4426950408889634073599246810019)
+# endif
+
+# if !defined(M_LN2)
+#  define M_LN2 (0.69314718055994530941723212145818)
+# endif
+
+/*Smallest blocks are 4x4*/
+#define OD_LOG_BSIZE0 (2)
+/*There are 5 block sizes total (4x4, 8x8, 16x16, 32x32 and 64x64).*/
+#define OD_NBSIZES (5)
+
+/*There are 4 transform sizes total in AV1 (4x4, 8x8, 16x16 and 32x32).*/
+#define OD_TXSIZES TX_SIZES
+/*The log of the maximum length of the side of a transform.*/
+#define OD_LOG_TXSIZE_MAX (OD_LOG_BSIZE0 + OD_TXSIZES - 1)
+/*The maximum length of the side of a transform.*/
+#define OD_TXSIZE_MAX (1 << OD_LOG_TXSIZE_MAX)
+
+/**The maximum number of color planes allowed in a single frame.*/
+# define OD_NPLANES_MAX (3)
+
+# define OD_COEFF_SHIFT (4)
+
+# define OD_DISABLE_CFL (1)
+# define OD_DISABLE_FILTER (1)
+
+#if !defined(NDEBUG)
+# define OD_ENABLE_ASSERTIONS (1)
+#endif
+
+# define OD_LOG(a)
+# define OD_LOG_PARTIAL(a)
+
+/*Possible block sizes, note that OD_BLOCK_NXN = log2(N) - 2.*/
+#define OD_BLOCK_4X4 (0)
+#define OD_BLOCK_8X8 (1)
+#define OD_BLOCK_16X16 (2)
+#define OD_BLOCK_32X32 (3)
+#define OD_BLOCK_SIZES (OD_BLOCK_32X32 + 1)
+
+# define OD_LIMIT_BSIZE_MIN (OD_BLOCK_4X4)
+# define OD_LIMIT_BSIZE_MAX (OD_BLOCK_32X32)
+
+typedef int od_coeff;
+
+#define OD_DIVU_DMAX (1024)
+
+extern uint32_t OD_DIVU_SMALL_CONSTS[OD_DIVU_DMAX][2];
+
+/*Translate unsigned division by small divisors into multiplications.*/
+#define OD_DIVU_SMALL(_x, _d)                                     \
+  ((uint32_t)((OD_DIVU_SMALL_CONSTS[(_d)-1][0] * (uint64_t)(_x) + \
+               OD_DIVU_SMALL_CONSTS[(_d)-1][1]) >>                \
+              32) >>                                              \
+   (OD_ILOG_NZ(_d) - 1))
+
+#define OD_DIVU(_x, _d) \
+  (((_d) < OD_DIVU_DMAX) ? (OD_DIVU_SMALL((_x), (_d))) : ((_x) / (_d)))
+
+#define OD_MINI AOMMIN
+#define OD_MAXI AOMMAX
+#define OD_CLAMPI(min, val, max) (OD_MAXI(min, OD_MINI(val, max)))
+
+#define OD_CLZ0 (1)
+#define OD_CLZ(x) (-get_msb(x))
+#define OD_ILOG_NZ(x) (OD_CLZ0 - OD_CLZ(x))
+/*Note that __builtin_clz is not defined when x == 0, according to the gcc
+   documentation (and that of the x86 BSR instruction that implements it), so
+   we have to special-case it.
+  We define a special version of the macro to use when x can be zero.*/
+#define OD_ILOG(x) ((x) ? OD_ILOG_NZ(x) : 0)
+
+#define OD_LOG2(x) (M_LOG2E*log(x))
+#define OD_EXP2(x) (exp(M_LN2*(x)))
+
+/*Enable special features for gcc and compatible compilers.*/
+#if defined(__GNUC__) && defined(__GNUC_MINOR__) && defined(__GNUC_PATCHLEVEL__)
+#define OD_GNUC_PREREQ(maj, min, pat)                                \
+  ((__GNUC__ << 16) + (__GNUC_MINOR__ << 8) + __GNUC_PATCHLEVEL__ >= \
+   ((maj) << 16) + ((min) << 8) + pat)  // NOLINT
+#else
+#define OD_GNUC_PREREQ(maj, min, pat) (0)
+#endif
+
+#if OD_GNUC_PREREQ(3, 4, 0)
+#define OD_WARN_UNUSED_RESULT __attribute__((__warn_unused_result__))
+#else
+#define OD_WARN_UNUSED_RESULT
+#endif
+
+#if OD_GNUC_PREREQ(3, 4, 0)
+#define OD_ARG_NONNULL(x) __attribute__((__nonnull__(x)))
+#else
+#define OD_ARG_NONNULL(x)
+#endif
+
+#if defined(OD_ENABLE_ASSERTIONS)
+#if OD_GNUC_PREREQ(2, 5, 0)
+__attribute__((noreturn))
+#endif
+void od_fatal_impl(const char *_str, const char *_file, int _line);
+
+#define OD_FATAL(_str) (od_fatal_impl(_str, __FILE__, __LINE__))
+
+#define OD_ASSERT(_cond)                     \
+  do {                                       \
+    if (!(_cond)) {                          \
+      OD_FATAL("assertion failed: " #_cond); \
+    }                                        \
+  } while (0)
+
+#define OD_ASSERT2(_cond, _message)                        \
+  do {                                                     \
+    if (!(_cond)) {                                        \
+      OD_FATAL("assertion failed: " #_cond "\n" _message); \
+    }                                                      \
+  } while (0)
+
+#define OD_ALWAYS_TRUE(_cond) OD_ASSERT(_cond)
+
+#else
+#define OD_ASSERT(_cond)
+#define OD_ASSERT2(_cond, _message)
+#define OD_ALWAYS_TRUE(_cond) ((void)(_cond))
+#endif
+
+/** Copy n elements of memory from src to dst. The 0* term provides
+    compile-time type checking  */
+#if !defined(OVERRIDE_OD_COPY)
+#define OD_COPY(dst, src, n) \
+  (memcpy((dst), (src), sizeof(*(dst)) * (n) + 0 * ((dst) - (src))))
+#endif
+
+/** Copy n elements of memory from src to dst, allowing overlapping regions.
+    The 0* term provides compile-time type checking */
+#if !defined(OVERRIDE_OD_MOVE)
+# define OD_MOVE(dst, src, n) \
+ (memmove((dst), (src), sizeof(*(dst))*(n) + 0*((dst) - (src)) ))
+#endif
+
+/** Linkage will break without this if using a C++ compiler, and will issue
+ * warnings without this for a C compiler*/
+#if defined(__cplusplus)
+# define OD_EXTERN extern
+#else
+# define OD_EXTERN
+#endif
+
+/** Set n elements of dst to zero */
+#if !defined(OVERRIDE_OD_CLEAR)
+# define OD_CLEAR(dst, n) (memset((dst), 0, sizeof(*(dst))*(n)))
+#endif
+
+/** Silence unused parameter/variable warnings */
+# define OD_UNUSED(expr) (void)(expr)
+
+#if defined(OD_FLOAT_PVQ)
+typedef double od_val16;
+typedef double od_val32;
+# define OD_QCONST32(x, bits) (x)
+# define OD_ROUND16(x) (x)
+# define OD_ROUND32(x) (x)
+# define OD_SHL(x, shift) (x)
+# define OD_SHR(x, shift) (x)
+# define OD_SHR_ROUND(x, shift) (x)
+# define OD_ABS(x) (fabs(x))
+# define OD_MULT16_16(a, b) ((a)*(b))
+# define OD_MULT16_32_Q16(a, b) ((a)*(b))
+#else
+typedef int16_t od_val16;
+typedef int32_t od_val32;
+/** Compile-time conversion of float constant to 32-bit value */
+# define OD_QCONST32(x, bits) ((od_val32)(.5 + (x)*(((od_val32)1) << (bits))))
+# define OD_ROUND16(x) (int16_t)(floor(.5 + (x)))
+# define OD_ROUND32(x) (int32_t)(floor(.5 + (x)))
+/*Shift x left by shift*/
+# define OD_SHL(a, shift) ((int32_t)((uint32_t)(a) << (shift)))
+/*Shift x right by shift (without rounding)*/
+# define OD_SHR(x, shift) \
+  ((int32_t)((x) >> (shift)))
+/*Shift x right by shift (with rounding)*/
+# define OD_SHR_ROUND(x, shift) \
+  ((int32_t)(((x) + (1 << (shift) >> 1)) >> (shift)))
+/*Shift x right by shift (without rounding) or left by -shift if shift
+  is negative.*/
+# define OD_VSHR(x, shift) \
+  (((shift) > 0) ? OD_SHR(x, shift) : OD_SHL(x, -(shift)))
+/*Shift x right by shift (with rounding) or left by -shift if shift
+  is negative.*/
+# define OD_VSHR_ROUND(x, shift) \
+  (((shift) > 0) ? OD_SHR_ROUND(x, shift) : OD_SHL(x, -(shift)))
+# define OD_ABS(x) (abs(x))
+/* (od_val32)(od_val16) gives TI compiler a hint that it's 16x16->32 multiply */
+/** 16x16 multiplication where the result fits in 32 bits */
+# define OD_MULT16_16(a, b) \
+ (((od_val32)(od_val16)(a))*((od_val32)(od_val16)(b)))
+/* Multiplies 16-bit a by 32-bit b and keeps bits [16:47]. */
+# define OD_MULT16_32_Q16(a, b) ((int16_t)(a)*(int64_t)(int32_t)(b) >> 16)
+/*16x16 multiplication where the result fits in 16 bits, without rounding.*/
+# define OD_MULT16_16_Q15(a, b) \
+  (((int16_t)(a)*((int32_t)(int16_t)(b))) >> 15)
+/*16x16 multiplication where the result fits in 16 bits, without rounding.*/
+# define OD_MULT16_16_Q16(a, b) \
+  ((((int16_t)(a))*((int32_t)(int16_t)(b))) >> 16)
+#endif
+
+/*All of these macros should expect floats as arguments.*/
+/*These two should compile as a single SSE instruction.*/
+# define OD_MINF(a, b) ((a) < (b) ? (a) : (b))
+# define OD_MAXF(a, b) ((a) > (b) ? (a) : (b))
+
+# define OD_DIV_R0(x, y) (((x) + OD_FLIPSIGNI((((y) + 1) >> 1) - 1, (x)))/(y))
+
+# define OD_SIGNMASK(a) (-((a) < 0))
+# define OD_FLIPSIGNI(a, b) (((a) + OD_SIGNMASK(b)) ^ OD_SIGNMASK(b))
+
+# define OD_MULT16_16_Q15(a, b) \
+  (((int16_t)(a)*((int32_t)(int16_t)(b))) >> 15)
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_ODINTRIN_H_
diff --git a/third_party/aom/av1/common/onyxc_int.h b/third_party/aom/av1/common/onyxc_int.h
new file mode 100644
index 0000000000..7980bde394
--- /dev/null
+++ b/third_party/aom/av1/common/onyxc_int.h
@@ -0,0 +1,1027 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_ONYXC_INT_H_
+#define AV1_COMMON_ONYXC_INT_H_
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "aom/internal/aom_codec_internal.h"
+#include "aom_util/aom_thread.h"
+#if CONFIG_ANS
+#include "aom_dsp/ans.h"
+#endif
+#include "av1/common/alloccommon.h"
+#include "av1/common/av1_loopfilter.h"
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/entropymv.h"
+#include "av1/common/frame_buffers.h"
+#include "av1/common/mv.h"
+#include "av1/common/quant_common.h"
+#if CONFIG_LOOP_RESTORATION
+#include "av1/common/restoration.h"
+#endif  // CONFIG_LOOP_RESTORATION
+#include "av1/common/tile_common.h"
+#include "av1/common/odintrin.h"
+#if CONFIG_PVQ
+#include "av1/common/pvq.h"
+#endif
+#if CONFIG_CFL
+#include "av1/common/cfl.h"
+#endif
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define CDEF_MAX_STRENGTHS 16
+
+#define REF_FRAMES_LOG2 3
+#define REF_FRAMES (1 << REF_FRAMES_LOG2)
+
+// 4 scratch frames for the new frames to support a maximum of 4 cores decoding
+// in parallel, 3 for scaled references on the encoder.
+// TODO(hkuang): Add ondemand frame buffers instead of hardcoding the number
+// of framebuffers.
+// TODO(jkoleszar): These 3 extra references could probably come from the
+// normal reference pool.
+#define FRAME_BUFFERS (REF_FRAMES + 7)
+
+#if CONFIG_REFERENCE_BUFFER
+/* Constant values while waiting for the sequence header */
+#define FRAME_ID_NUMBERS_PRESENT_FLAG 1
+#define FRAME_ID_LENGTH_MINUS7 8         // Allows frame id up to 2^15-1
+#define DELTA_FRAME_ID_LENGTH_MINUS2 12  // Allows frame id deltas up to 2^14-1
+#endif
+
+#if CONFIG_EXT_REFS
+#define FRAME_CONTEXTS_LOG2 3
+#else
+#define FRAME_CONTEXTS_LOG2 2
+#endif
+
+#define FRAME_CONTEXTS (1 << FRAME_CONTEXTS_LOG2)
+
+#define NUM_PING_PONG_BUFFERS 2
+
+typedef enum {
+  SINGLE_REFERENCE = 0,
+  COMPOUND_REFERENCE = 1,
+  REFERENCE_MODE_SELECT = 2,
+  REFERENCE_MODES = 3,
+} REFERENCE_MODE;
+
+typedef enum {
+  RESET_FRAME_CONTEXT_NONE = 0,
+  RESET_FRAME_CONTEXT_CURRENT = 1,
+  RESET_FRAME_CONTEXT_ALL = 2,
+} RESET_FRAME_CONTEXT_MODE;
+
+typedef enum {
+  /**
+   * Update frame context to values resulting from forward probability
+   * updates signaled in the frame header
+   */
+  REFRESH_FRAME_CONTEXT_FORWARD,
+  /**
+   * Update frame context to values resulting from backward probability
+   * updates based on entropy/counts in the decoded frame
+   */
+  REFRESH_FRAME_CONTEXT_BACKWARD,
+} REFRESH_FRAME_CONTEXT_MODE;
+
+typedef struct {
+  int_mv mv[2];
+#if CONFIG_REF_MV
+  int_mv pred_mv[2];
+#endif
+  MV_REFERENCE_FRAME ref_frame[2];
+} MV_REF;
+
+typedef struct {
+  int ref_count;
+  MV_REF *mvs;
+  int mi_rows;
+  int mi_cols;
+#if CONFIG_GLOBAL_MOTION
+  WarpedMotionParams global_motion[TOTAL_REFS_PER_FRAME];
+#endif  // CONFIG_GLOBAL_MOTION
+  aom_codec_frame_buffer_t raw_frame_buffer;
+  YV12_BUFFER_CONFIG buf;
+#if CONFIG_TEMPMV_SIGNALING
+  uint8_t intra_only;
+#endif
+  // The Following variables will only be used in frame parallel decode.
+
+  // frame_worker_owner indicates which FrameWorker owns this buffer. NULL means
+  // that no FrameWorker owns, or is decoding, this buffer.
+  AVxWorker *frame_worker_owner;
+
+  // row and col indicate which position frame has been decoded to in real
+  // pixel unit. They are reset to -1 when decoding begins and set to INT_MAX
+  // when the frame is fully decoded.
+  int row;
+  int col;
+} RefCntBuffer;
+
+typedef struct BufferPool {
+// Protect BufferPool from being accessed by several FrameWorkers at
+// the same time during frame parallel decode.
+// TODO(hkuang): Try to use atomic variable instead of locking the whole pool.
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t pool_mutex;
+#endif
+
+  // Private data associated with the frame buffer callbacks.
+  void *cb_priv;
+
+  aom_get_frame_buffer_cb_fn_t get_fb_cb;
+  aom_release_frame_buffer_cb_fn_t release_fb_cb;
+
+  RefCntBuffer frame_bufs[FRAME_BUFFERS];
+
+  // Frame buffers allocated internally by the codec.
+  InternalFrameBufferList int_frame_buffers;
+} BufferPool;
+
+typedef struct AV1Common {
+  struct aom_internal_error_info error;
+  aom_color_space_t color_space;
+  int color_range;
+  int width;
+  int height;
+  int render_width;
+  int render_height;
+  int last_width;
+  int last_height;
+
+#if CONFIG_FRAME_SUPERRES
+  // The numerator of the superres scale, the denominator is fixed
+  uint8_t superres_scale_numerator;
+  int superres_width, superres_height;
+#endif  // CONFIG_FRAME_SUPERRES
+
+  // TODO(jkoleszar): this implies chroma ss right now, but could vary per
+  // plane. Revisit as part of the future change to YV12_BUFFER_CONFIG to
+  // support additional planes.
+  int subsampling_x;
+  int subsampling_y;
+
+#if CONFIG_HIGHBITDEPTH
+  // Marks if we need to use 16bit frame buffers (1: yes, 0: no).
+  int use_highbitdepth;
+#endif
+  YV12_BUFFER_CONFIG *frame_to_show;
+  RefCntBuffer *prev_frame;
+
+  // TODO(hkuang): Combine this with cur_buf in macroblockd.
+  RefCntBuffer *cur_frame;
+
+  int ref_frame_map[REF_FRAMES]; /* maps fb_idx to reference slot */
+
+  // Prepare ref_frame_map for the next frame.
+  // Only used in frame parallel decode.
+  int next_ref_frame_map[REF_FRAMES];
+
+  // TODO(jkoleszar): could expand active_ref_idx to 4, with 0 as intra, and
+  // roll new_fb_idx into it.
+
+  // Each Inter frame can reference INTER_REFS_PER_FRAME buffers
+  RefBuffer frame_refs[INTER_REFS_PER_FRAME];
+
+  int new_fb_idx;
+
+  FRAME_TYPE last_frame_type; /* last frame's frame type for motion search.*/
+  FRAME_TYPE frame_type;
+
+  int show_frame;
+  int last_show_frame;
+  int show_existing_frame;
+#if CONFIG_EXT_REFS
+  // Flag for a frame used as a reference - not written to the bitstream
+  int is_reference_frame;
+#endif  // CONFIG_EXT_REFS
+
+  // Flag signaling that the frame is encoded using only INTRA modes.
+  uint8_t intra_only;
+  uint8_t last_intra_only;
+
+  int allow_high_precision_mv;
+
+#if CONFIG_PALETTE
+  int allow_screen_content_tools;
+#endif  // CONFIG_PALETTE
+
+  // Flag signaling which frame contexts should be reset to default values.
+  RESET_FRAME_CONTEXT_MODE reset_frame_context;
+
+  // MBs, mb_rows/cols is in 16-pixel units; mi_rows/cols is in
+  // MODE_INFO (8-pixel) units.
+  int MBs;
+  int mb_rows, mi_rows;
+  int mb_cols, mi_cols;
+  int mi_stride;
+
+  /* profile settings */
+  TX_MODE tx_mode;
+
+  int base_qindex;
+  int y_dc_delta_q;
+  int uv_dc_delta_q;
+  int uv_ac_delta_q;
+  int16_t y_dequant[MAX_SEGMENTS][2];
+  int16_t uv_dequant[MAX_SEGMENTS][2];
+
+#if CONFIG_AOM_QM
+  // Global quant matrix tables
+  qm_val_t *giqmatrix[NUM_QM_LEVELS][2][2][TX_SIZES];
+  qm_val_t *gqmatrix[NUM_QM_LEVELS][2][2][TX_SIZES];
+
+  // Local quant matrix tables for each frame
+  qm_val_t *y_iqmatrix[MAX_SEGMENTS][2][TX_SIZES];
+  qm_val_t *uv_iqmatrix[MAX_SEGMENTS][2][TX_SIZES];
+  // Encoder
+  qm_val_t *y_qmatrix[MAX_SEGMENTS][2][TX_SIZES];
+  qm_val_t *uv_qmatrix[MAX_SEGMENTS][2][TX_SIZES];
+
+  int using_qmatrix;
+  int min_qmlevel;
+  int max_qmlevel;
+#endif
+#if CONFIG_NEW_QUANT
+  dequant_val_type_nuq y_dequant_nuq[MAX_SEGMENTS][QUANT_PROFILES][COEF_BANDS];
+  dequant_val_type_nuq uv_dequant_nuq[MAX_SEGMENTS][QUANT_PROFILES][COEF_BANDS];
+#endif
+
+  /* We allocate a MODE_INFO struct for each macroblock, together with
+     an extra row on top and column on the left to simplify prediction. */
+  int mi_alloc_size;
+  MODE_INFO *mip; /* Base of allocated array */
+  MODE_INFO *mi;  /* Corresponds to upper left visible macroblock */
+
+  // TODO(agrange): Move prev_mi into encoder structure.
+  // prev_mip and prev_mi will only be allocated in encoder.
+  MODE_INFO *prev_mip; /* MODE_INFO array 'mip' from last decoded frame */
+  MODE_INFO *prev_mi;  /* 'mi' from last frame (points into prev_mip) */
+
+  // Separate mi functions between encoder and decoder.
+  int (*alloc_mi)(struct AV1Common *cm, int mi_size);
+  void (*free_mi)(struct AV1Common *cm);
+  void (*setup_mi)(struct AV1Common *cm);
+
+  // Grid of pointers to 8x8 MODE_INFO structs.  Any 8x8 not in the visible
+  // area will be NULL.
+  MODE_INFO **mi_grid_base;
+  MODE_INFO **mi_grid_visible;
+  MODE_INFO **prev_mi_grid_base;
+  MODE_INFO **prev_mi_grid_visible;
+
+  // Whether to use previous frame's motion vectors for prediction.
+  int use_prev_frame_mvs;
+
+  // Persistent mb segment id map used in prediction.
+  int seg_map_idx;
+  int prev_seg_map_idx;
+
+  uint8_t *seg_map_array[NUM_PING_PONG_BUFFERS];
+  uint8_t *last_frame_seg_map;
+  uint8_t *current_frame_seg_map;
+  int seg_map_alloc_size;
+
+  InterpFilter interp_filter;
+
+  loop_filter_info_n lf_info;
+#if CONFIG_LOOP_RESTORATION
+  RestorationInfo rst_info[MAX_MB_PLANE];
+  RestorationInternal rst_internal;
+#endif  // CONFIG_LOOP_RESTORATION
+
+  // Flag signaling how frame contexts should be updated at the end of
+  // a frame decode
+  REFRESH_FRAME_CONTEXT_MODE refresh_frame_context;
+
+  int ref_frame_sign_bias[TOTAL_REFS_PER_FRAME]; /* Two state 0, 1 */
+
+  struct loopfilter lf;
+  struct segmentation seg;
+
+  int frame_parallel_decode;  // frame-based threading.
+
+#if CONFIG_EXT_TX
+  int reduced_tx_set_used;
+#endif  // CONFIG_EXT_TX
+
+// Context probabilities for reference frame prediction
+#if CONFIG_EXT_REFS
+  MV_REFERENCE_FRAME comp_fwd_ref[FWD_REFS];
+  MV_REFERENCE_FRAME comp_bwd_ref[BWD_REFS];
+#else
+  MV_REFERENCE_FRAME comp_fixed_ref;
+  MV_REFERENCE_FRAME comp_var_ref[COMP_REFS];
+#endif  // CONFIG_EXT_REFS
+  REFERENCE_MODE reference_mode;
+
+  FRAME_CONTEXT *fc;              /* this frame entropy */
+  FRAME_CONTEXT *frame_contexts;  // FRAME_CONTEXTS
+  unsigned int frame_context_idx; /* Context to use/update */
+  FRAME_COUNTS counts;
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  // The initial probabilities for a frame, before any subframe backward update,
+  // and after forward update.
+  av1_coeff_probs_model starting_coef_probs[TX_SIZES][PLANE_TYPES];
+  // Number of subframe backward updates already done
+  uint8_t coef_probs_update_idx;
+  // Signal if the backward update is subframe or end-of-frame
+  uint8_t partial_prob_update;
+  // Frame level flag to turn on/off subframe backward update
+  uint8_t do_subframe_update;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+  unsigned int current_video_frame;
+  BITSTREAM_PROFILE profile;
+
+  // AOM_BITS_8 in profile 0 or 1, AOM_BITS_10 or AOM_BITS_12 in profile 2 or 3.
+  aom_bit_depth_t bit_depth;
+  aom_bit_depth_t dequant_bit_depth;  // bit_depth of current dequantizer
+
+  int error_resilient_mode;
+
+#if !CONFIG_EXT_TILE
+  int log2_tile_cols, log2_tile_rows;
+#endif  // !CONFIG_EXT_TILE
+  int tile_cols, tile_rows;
+  int tile_width, tile_height;  // In MI units
+#if CONFIG_EXT_TILE
+  unsigned int tile_encoding_mode;
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_DEPENDENT_HORZTILES
+  int dependent_horz_tiles;
+#if CONFIG_TILE_GROUPS
+  int tile_group_start_row[MAX_TILE_ROWS][MAX_TILE_COLS];
+  int tile_group_start_col[MAX_TILE_ROWS][MAX_TILE_COLS];
+#endif
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  int loop_filter_across_tiles_enabled;
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+  int byte_alignment;
+  int skip_loop_filter;
+
+  // Private data associated with the frame buffer callbacks.
+  void *cb_priv;
+  aom_get_frame_buffer_cb_fn_t get_fb_cb;
+  aom_release_frame_buffer_cb_fn_t release_fb_cb;
+
+  // Handles memory for the codec.
+  InternalFrameBufferList int_frame_buffers;
+
+  // External BufferPool passed from outside.
+  BufferPool *buffer_pool;
+
+  PARTITION_CONTEXT *above_seg_context;
+  ENTROPY_CONTEXT *above_context[MAX_MB_PLANE];
+#if CONFIG_VAR_TX
+  TXFM_CONTEXT *above_txfm_context;
+  TXFM_CONTEXT left_txfm_context[MAX_MIB_SIZE];
+#endif
+  int above_context_alloc_cols;
+
+  // scratch memory for intraonly/keyframe forward updates from default tables
+  // - this is intentionally not placed in FRAME_CONTEXT since it's reset upon
+  // each keyframe and not used afterwards
+  aom_prob kf_y_prob[INTRA_MODES][INTRA_MODES][INTRA_MODES - 1];
+#if CONFIG_GLOBAL_MOTION
+  WarpedMotionParams global_motion[TOTAL_REFS_PER_FRAME];
+#endif
+
+  BLOCK_SIZE sb_size;  // Size of the superblock used for this frame
+  int mib_size;        // Size of the superblock in units of MI blocks
+  int mib_size_log2;   // Log 2 of above.
+#if CONFIG_CDEF
+  int cdef_dering_damping;
+  int cdef_clpf_damping;
+  int nb_cdef_strengths;
+  int cdef_strengths[CDEF_MAX_STRENGTHS];
+  int cdef_uv_strengths[CDEF_MAX_STRENGTHS];
+  int cdef_bits;
+#endif
+
+#if CONFIG_DELTA_Q
+  int delta_q_present_flag;
+  // Resolution of delta quant
+  int delta_q_res;
+#if CONFIG_EXT_DELTA_Q
+  int delta_lf_present_flag;
+  // Resolution of delta lf level
+  int delta_lf_res;
+#endif
+#endif
+#if CONFIG_TILE_GROUPS
+  int num_tg;
+#endif
+#if CONFIG_REFERENCE_BUFFER
+  int current_frame_id;
+  int ref_frame_id[REF_FRAMES];
+  int valid_for_referencing[REF_FRAMES];
+  int refresh_mask;
+  int invalid_delta_frame_id_minus1;
+#endif
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  int ans_window_size_log2;
+#endif
+} AV1_COMMON;
+
+#if CONFIG_REFERENCE_BUFFER
+/* Initial version of sequence header structure */
+typedef struct SequenceHeader {
+  int frame_id_numbers_present_flag;
+  int frame_id_length_minus7;
+  int delta_frame_id_length_minus2;
+} SequenceHeader;
+#endif
+
+// TODO(hkuang): Don't need to lock the whole pool after implementing atomic
+// frame reference count.
+static void lock_buffer_pool(BufferPool *const pool) {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_lock(&pool->pool_mutex);
+#else
+  (void)pool;
+#endif
+}
+
+static void unlock_buffer_pool(BufferPool *const pool) {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_unlock(&pool->pool_mutex);
+#else
+  (void)pool;
+#endif
+}
+
+static INLINE YV12_BUFFER_CONFIG *get_ref_frame(AV1_COMMON *cm, int index) {
+  if (index < 0 || index >= REF_FRAMES) return NULL;
+  if (cm->ref_frame_map[index] < 0) return NULL;
+  assert(cm->ref_frame_map[index] < FRAME_BUFFERS);
+  return &cm->buffer_pool->frame_bufs[cm->ref_frame_map[index]].buf;
+}
+
+static INLINE YV12_BUFFER_CONFIG *get_frame_new_buffer(
+    const AV1_COMMON *const cm) {
+  return &cm->buffer_pool->frame_bufs[cm->new_fb_idx].buf;
+}
+
+static INLINE int get_free_fb(AV1_COMMON *cm) {
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+  int i;
+
+  lock_buffer_pool(cm->buffer_pool);
+  for (i = 0; i < FRAME_BUFFERS; ++i)
+    if (frame_bufs[i].ref_count == 0) break;
+
+  if (i != FRAME_BUFFERS) {
+    frame_bufs[i].ref_count = 1;
+  } else {
+    // Reset i to be INVALID_IDX to indicate no free buffer found.
+    i = INVALID_IDX;
+  }
+
+  unlock_buffer_pool(cm->buffer_pool);
+  return i;
+}
+
+static INLINE void ref_cnt_fb(RefCntBuffer *bufs, int *idx, int new_idx) {
+  const int ref_index = *idx;
+
+  if (ref_index >= 0 && bufs[ref_index].ref_count > 0)
+    bufs[ref_index].ref_count--;
+
+  *idx = new_idx;
+
+  bufs[new_idx].ref_count++;
+}
+
+static INLINE int mi_cols_aligned_to_sb(const AV1_COMMON *cm) {
+  return ALIGN_POWER_OF_TWO(cm->mi_cols, cm->mib_size_log2);
+}
+
+static INLINE int mi_rows_aligned_to_sb(const AV1_COMMON *cm) {
+  return ALIGN_POWER_OF_TWO(cm->mi_rows, cm->mib_size_log2);
+}
+
+static INLINE int frame_is_intra_only(const AV1_COMMON *const cm) {
+  return cm->frame_type == KEY_FRAME || cm->intra_only;
+}
+
+static INLINE void av1_init_macroblockd(AV1_COMMON *cm, MACROBLOCKD *xd,
+#if CONFIG_PVQ
+                                        tran_low_t *pvq_ref_coeff,
+#endif
+#if CONFIG_CFL
+                                        CFL_CTX *cfl,
+#endif
+                                        tran_low_t *dqcoeff) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    xd->plane[i].dqcoeff = dqcoeff;
+#if CONFIG_PVQ
+    xd->plane[i].pvq_ref_coeff = pvq_ref_coeff;
+#endif
+#if CONFIG_CFL
+    xd->cfl = cfl;
+    cfl_init(cfl, cm, xd->plane[AOM_PLANE_U].subsampling_x,
+             xd->plane[AOM_PLANE_U].subsampling_y);
+#endif
+    xd->above_context[i] = cm->above_context[i];
+    if (xd->plane[i].plane_type == PLANE_TYPE_Y) {
+      memcpy(xd->plane[i].seg_dequant, cm->y_dequant, sizeof(cm->y_dequant));
+#if CONFIG_AOM_QM
+      memcpy(xd->plane[i].seg_iqmatrix, cm->y_iqmatrix, sizeof(cm->y_iqmatrix));
+#endif
+
+#if CONFIG_NEW_QUANT
+      memcpy(xd->plane[i].seg_dequant_nuq, cm->y_dequant_nuq,
+             sizeof(cm->y_dequant_nuq));
+#endif
+    } else {
+      memcpy(xd->plane[i].seg_dequant, cm->uv_dequant, sizeof(cm->uv_dequant));
+#if CONFIG_AOM_QM
+      memcpy(xd->plane[i].seg_iqmatrix, cm->uv_iqmatrix,
+             sizeof(cm->uv_iqmatrix));
+#endif
+#if CONFIG_NEW_QUANT
+      memcpy(xd->plane[i].seg_dequant_nuq, cm->uv_dequant_nuq,
+             sizeof(cm->uv_dequant_nuq));
+#endif
+    }
+    xd->fc = cm->fc;
+  }
+  xd->above_seg_context = cm->above_seg_context;
+#if CONFIG_VAR_TX
+  xd->above_txfm_context = cm->above_txfm_context;
+#endif
+  xd->mi_stride = cm->mi_stride;
+  xd->error_info = &cm->error;
+}
+
+static INLINE void set_skip_context(MACROBLOCKD *xd, int mi_row, int mi_col) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    struct macroblockd_plane *const pd = &xd->plane[i];
+#if CONFIG_CHROMA_SUB8X8
+    if (xd->mi[0]->mbmi.sb_type < BLOCK_8X8) {
+      // Offset the buffer pointer
+      if (pd->subsampling_y && (mi_row & 0x01)) mi_row -= 1;
+      if (pd->subsampling_x && (mi_col & 0x01)) mi_col -= 1;
+    }
+#endif
+    int above_idx = mi_col * 2;
+    int left_idx = (mi_row * 2) & MAX_MIB_MASK_2;
+    pd->above_context = &xd->above_context[i][above_idx >> pd->subsampling_x];
+    pd->left_context = &xd->left_context[i][left_idx >> pd->subsampling_y];
+  }
+}
+
+static INLINE int calc_mi_size(int len) {
+  // len is in mi units.
+  return len + MAX_MIB_SIZE;
+}
+
+static INLINE void set_plane_n4(MACROBLOCKD *const xd, int bw, int bh) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    xd->plane[i].n4_w = (bw << 1) >> xd->plane[i].subsampling_x;
+    xd->plane[i].n4_h = (bh << 1) >> xd->plane[i].subsampling_y;
+
+    xd->plane[i].width = (bw * MI_SIZE) >> xd->plane[i].subsampling_x;
+    xd->plane[i].height = (bh * MI_SIZE) >> xd->plane[i].subsampling_y;
+
+#if !CONFIG_CHROMA_2X2
+    xd->plane[i].width = AOMMAX(xd->plane[i].width, 4);
+    xd->plane[i].height = AOMMAX(xd->plane[i].height, 4);
+#endif
+  }
+}
+
+static INLINE void set_mi_row_col(MACROBLOCKD *xd, const TileInfo *const tile,
+                                  int mi_row, int bh, int mi_col, int bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                                  int dependent_horz_tile_flag,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                                  int mi_rows, int mi_cols) {
+  xd->mb_to_top_edge = -((mi_row * MI_SIZE) * 8);
+  xd->mb_to_bottom_edge = ((mi_rows - bh - mi_row) * MI_SIZE) * 8;
+  xd->mb_to_left_edge = -((mi_col * MI_SIZE) * 8);
+  xd->mb_to_right_edge = ((mi_cols - bw - mi_col) * MI_SIZE) * 8;
+
+#if CONFIG_DEPENDENT_HORZTILES
+  if (dependent_horz_tile_flag) {
+#if CONFIG_TILE_GROUPS
+    xd->up_available = (mi_row > tile->mi_row_start) || !tile->tg_horz_boundary;
+#else
+    xd->up_available = (mi_row > 0);
+#endif  // CONFIG_TILE_GROUPS
+  } else {
+#endif  // CONFIG_DEPENDENT_HORZTILES
+    // Are edges available for intra prediction?
+    xd->up_available = (mi_row > tile->mi_row_start);
+#if CONFIG_DEPENDENT_HORZTILES
+  }
+#endif  // CONFIG_DEPENDENT_HORZTILES
+
+  xd->left_available = (mi_col > tile->mi_col_start);
+#if CONFIG_CHROMA_SUB8X8
+  xd->chroma_up_available = xd->up_available;
+  xd->chroma_left_available = xd->left_available;
+  if (xd->plane[1].subsampling_x && bw < mi_size_wide[BLOCK_8X8])
+    xd->chroma_left_available = (mi_col - 1) > tile->mi_col_start;
+  if (xd->plane[1].subsampling_y && bh < mi_size_high[BLOCK_8X8])
+    xd->chroma_up_available = (mi_row - 1) > tile->mi_row_start;
+#endif
+  if (xd->up_available) {
+    xd->above_mi = xd->mi[-xd->mi_stride];
+    // above_mi may be NULL in encoder's first pass.
+    xd->above_mbmi = xd->above_mi ? &xd->above_mi->mbmi : NULL;
+  } else {
+    xd->above_mi = NULL;
+    xd->above_mbmi = NULL;
+  }
+
+  if (xd->left_available) {
+    xd->left_mi = xd->mi[-1];
+    // left_mi may be NULL in encoder's first pass.
+    xd->left_mbmi = xd->left_mi ? &xd->left_mi->mbmi : NULL;
+  } else {
+    xd->left_mi = NULL;
+    xd->left_mbmi = NULL;
+  }
+
+  xd->n8_h = bh;
+  xd->n8_w = bw;
+#if CONFIG_REF_MV
+  xd->is_sec_rect = 0;
+  if (xd->n8_w < xd->n8_h)
+    if (mi_col & (xd->n8_h - 1)) xd->is_sec_rect = 1;
+
+  if (xd->n8_w > xd->n8_h)
+    if (mi_row & (xd->n8_w - 1)) xd->is_sec_rect = 1;
+#endif  // CONFIG_REF_MV
+}
+
+static INLINE const aom_prob *get_y_mode_probs(const AV1_COMMON *cm,
+                                               const MODE_INFO *mi,
+                                               const MODE_INFO *above_mi,
+                                               const MODE_INFO *left_mi,
+                                               int block) {
+  const PREDICTION_MODE above = av1_above_block_mode(mi, above_mi, block);
+  const PREDICTION_MODE left = av1_left_block_mode(mi, left_mi, block);
+  return cm->kf_y_prob[above][left];
+}
+
+#if CONFIG_EC_MULTISYMBOL
+static INLINE aom_cdf_prob *get_y_mode_cdf(FRAME_CONTEXT *tile_ctx,
+                                           const MODE_INFO *mi,
+                                           const MODE_INFO *above_mi,
+                                           const MODE_INFO *left_mi,
+                                           int block) {
+  const PREDICTION_MODE above = av1_above_block_mode(mi, above_mi, block);
+  const PREDICTION_MODE left = av1_left_block_mode(mi, left_mi, block);
+  return tile_ctx->kf_y_cdf[above][left];
+}
+#endif
+
+static INLINE void update_partition_context(MACROBLOCKD *xd, int mi_row,
+                                            int mi_col, BLOCK_SIZE subsize,
+                                            BLOCK_SIZE bsize) {
+  PARTITION_CONTEXT *const above_ctx = xd->above_seg_context + mi_col;
+  PARTITION_CONTEXT *const left_ctx =
+      xd->left_seg_context + (mi_row & MAX_MIB_MASK);
+
+#if CONFIG_EXT_PARTITION_TYPES
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  memset(above_ctx, partition_context_lookup[subsize].above, bw);
+  memset(left_ctx, partition_context_lookup[subsize].left, bh);
+#else
+  // num_4x4_blocks_wide_lookup[bsize] / 2
+  const int bs = mi_size_wide[bsize];
+
+  // update the partition context at the end notes. set partition bits
+  // of block sizes larger than the current one to be one, and partition
+  // bits of smaller block sizes to be zero.
+  memset(above_ctx, partition_context_lookup[subsize].above, bs);
+  memset(left_ctx, partition_context_lookup[subsize].left, bs);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+}
+
+#if CONFIG_CB4X4
+static INLINE int is_chroma_reference(int mi_row, int mi_col, BLOCK_SIZE bsize,
+                                      int subsampling_x, int subsampling_y) {
+#if CONFIG_CHROMA_2X2
+  return 1;
+#endif
+
+#if CONFIG_CHROMA_SUB8X8
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+
+  int ref_pos = ((mi_row & 0x01) || !(bh & 0x01) || !subsampling_y) &&
+                ((mi_col & 0x01) || !(bw & 0x01) || !subsampling_x);
+
+  return ref_pos;
+#else
+  int ref_pos = !(((mi_row & 0x01) && subsampling_y) ||
+                  ((mi_col & 0x01) && subsampling_x));
+
+  if (bsize >= BLOCK_8X8) ref_pos = 1;
+
+  return ref_pos;
+#endif
+}
+
+static INLINE BLOCK_SIZE scale_chroma_bsize(BLOCK_SIZE bsize, int subsampling_x,
+                                            int subsampling_y) {
+  BLOCK_SIZE bs = bsize;
+
+  if (bs < BLOCK_8X8) {
+    if (subsampling_x == 1 && subsampling_y == 1)
+      bs = BLOCK_8X8;
+    else if (subsampling_x == 1)
+      bs = BLOCK_8X4;
+    else if (subsampling_y == 1)
+      bs = BLOCK_4X8;
+  }
+
+  return bs;
+}
+#endif
+
+#if CONFIG_EXT_PARTITION_TYPES
+static INLINE void update_ext_partition_context(MACROBLOCKD *xd, int mi_row,
+                                                int mi_col, BLOCK_SIZE subsize,
+                                                BLOCK_SIZE bsize,
+                                                PARTITION_TYPE partition) {
+  if (bsize >= BLOCK_8X8) {
+    const int hbs = mi_size_wide[bsize] / 2;
+    BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+    switch (partition) {
+      case PARTITION_SPLIT:
+        if (bsize != BLOCK_8X8) break;
+      case PARTITION_NONE:
+      case PARTITION_HORZ:
+      case PARTITION_VERT:
+        update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+        break;
+      case PARTITION_HORZ_A:
+        update_partition_context(xd, mi_row, mi_col, bsize2, subsize);
+        update_partition_context(xd, mi_row + hbs, mi_col, subsize, subsize);
+        break;
+      case PARTITION_HORZ_B:
+        update_partition_context(xd, mi_row, mi_col, subsize, subsize);
+        update_partition_context(xd, mi_row + hbs, mi_col, bsize2, subsize);
+        break;
+      case PARTITION_VERT_A:
+        update_partition_context(xd, mi_row, mi_col, bsize2, subsize);
+        update_partition_context(xd, mi_row, mi_col + hbs, subsize, subsize);
+        break;
+      case PARTITION_VERT_B:
+        update_partition_context(xd, mi_row, mi_col, subsize, subsize);
+        update_partition_context(xd, mi_row, mi_col + hbs, bsize2, subsize);
+        break;
+      default: assert(0 && "Invalid partition type");
+    }
+  }
+}
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+static INLINE int partition_plane_context(const MACROBLOCKD *xd, int mi_row,
+                                          int mi_col,
+#if CONFIG_UNPOISON_PARTITION_CTX
+                                          int has_rows, int has_cols,
+#endif
+                                          BLOCK_SIZE bsize) {
+#if CONFIG_UNPOISON_PARTITION_CTX
+  const PARTITION_CONTEXT *above_ctx = xd->above_seg_context + mi_col;
+  const PARTITION_CONTEXT *left_ctx =
+      xd->left_seg_context + (mi_row & MAX_MIB_MASK);
+  // Minimum partition point is 8x8. Offset the bsl accordingly.
+  const int bsl = mi_width_log2_lookup[bsize] - mi_width_log2_lookup[BLOCK_8X8];
+  int above = (*above_ctx >> bsl) & 1, left = (*left_ctx >> bsl) & 1;
+
+  assert(b_width_log2_lookup[bsize] == b_height_log2_lookup[bsize]);
+  assert(bsl >= 0);
+
+  if (has_rows && has_cols)
+    return (left * 2 + above) + bsl * PARTITION_PLOFFSET;
+  else if (has_rows && !has_cols)
+    return PARTITION_CONTEXTS_PRIMARY + bsl;
+  else if (!has_rows && has_cols)
+    return PARTITION_CONTEXTS_PRIMARY + PARTITION_BLOCK_SIZES + bsl;
+  else
+    return PARTITION_CONTEXTS;  // Bogus context, forced SPLIT
+#else
+  const PARTITION_CONTEXT *above_ctx = xd->above_seg_context + mi_col;
+  const PARTITION_CONTEXT *left_ctx =
+      xd->left_seg_context + (mi_row & MAX_MIB_MASK);
+  // Minimum partition point is 8x8. Offset the bsl accordingly.
+  const int bsl = mi_width_log2_lookup[bsize] - mi_width_log2_lookup[BLOCK_8X8];
+  int above = (*above_ctx >> bsl) & 1, left = (*left_ctx >> bsl) & 1;
+
+  assert(b_width_log2_lookup[bsize] == b_height_log2_lookup[bsize]);
+  assert(bsl >= 0);
+
+  return (left * 2 + above) + bsl * PARTITION_PLOFFSET;
+#endif
+}
+
+static INLINE int max_block_wide(const MACROBLOCKD *xd, BLOCK_SIZE bsize,
+                                 int plane) {
+  int max_blocks_wide = block_size_wide[bsize];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+
+  if (xd->mb_to_right_edge < 0)
+    max_blocks_wide += xd->mb_to_right_edge >> (3 + pd->subsampling_x);
+
+  // Scale the width in the transform block unit.
+  return max_blocks_wide >> tx_size_wide_log2[0];
+}
+
+static INLINE int max_block_high(const MACROBLOCKD *xd, BLOCK_SIZE bsize,
+                                 int plane) {
+  int max_blocks_high = block_size_high[bsize];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+
+  if (xd->mb_to_bottom_edge < 0)
+    max_blocks_high += xd->mb_to_bottom_edge >> (3 + pd->subsampling_y);
+
+  // Scale the width in the transform block unit.
+  return max_blocks_high >> tx_size_wide_log2[0];
+}
+
+static INLINE void av1_zero_above_context(AV1_COMMON *const cm,
+                                          int mi_col_start, int mi_col_end) {
+  const int width = mi_col_end - mi_col_start;
+  const int aligned_width = ALIGN_POWER_OF_TWO(width, cm->mib_size_log2);
+
+  const int offset_y = 2 * mi_col_start;
+  const int width_y = 2 * aligned_width;
+  const int offset_uv = offset_y >> cm->subsampling_x;
+  const int width_uv = width_y >> cm->subsampling_x;
+
+  av1_zero_array(cm->above_context[0] + offset_y, width_y);
+  av1_zero_array(cm->above_context[1] + offset_uv, width_uv);
+  av1_zero_array(cm->above_context[2] + offset_uv, width_uv);
+
+  av1_zero_array(cm->above_seg_context + mi_col_start, aligned_width);
+
+#if CONFIG_VAR_TX
+  av1_zero_array(cm->above_txfm_context + mi_col_start, aligned_width);
+#endif  // CONFIG_VAR_TX
+}
+
+static INLINE void av1_zero_left_context(MACROBLOCKD *const xd) {
+  av1_zero(xd->left_context);
+  av1_zero(xd->left_seg_context);
+#if CONFIG_VAR_TX
+  av1_zero(xd->left_txfm_context_buffer);
+#endif
+}
+
+#if CONFIG_VAR_TX
+static INLINE TX_SIZE get_min_tx_size(TX_SIZE tx_size) {
+  if (tx_size >= TX_SIZES_ALL) assert(0);
+  return txsize_sqr_map[tx_size];
+}
+
+static INLINE void set_txfm_ctx(TXFM_CONTEXT *txfm_ctx, uint8_t txs, int len) {
+  int i;
+  for (i = 0; i < len; ++i) txfm_ctx[i] = txs;
+}
+
+static INLINE void set_txfm_ctxs(TX_SIZE tx_size, int n8_w, int n8_h, int skip,
+                                 const MACROBLOCKD *xd) {
+  uint8_t bw = tx_size_wide[tx_size];
+  uint8_t bh = tx_size_high[tx_size];
+
+  if (skip) {
+    bw = n8_w * MI_SIZE;
+    bh = n8_h * MI_SIZE;
+  }
+
+  set_txfm_ctx(xd->above_txfm_context, bw, n8_w);
+  set_txfm_ctx(xd->left_txfm_context, bh, n8_h);
+}
+
+static INLINE void txfm_partition_update(TXFM_CONTEXT *above_ctx,
+                                         TXFM_CONTEXT *left_ctx,
+                                         TX_SIZE tx_size, TX_SIZE txb_size) {
+  BLOCK_SIZE bsize = txsize_to_bsize[txb_size];
+  int bh = mi_size_high[bsize];
+  int bw = mi_size_wide[bsize];
+  uint8_t txw = tx_size_wide[tx_size];
+  uint8_t txh = tx_size_high[tx_size];
+  int i;
+  for (i = 0; i < bh; ++i) left_ctx[i] = txh;
+  for (i = 0; i < bw; ++i) above_ctx[i] = txw;
+}
+
+static INLINE int txfm_partition_context(TXFM_CONTEXT *above_ctx,
+                                         TXFM_CONTEXT *left_ctx,
+                                         BLOCK_SIZE bsize, TX_SIZE tx_size) {
+  const uint8_t txw = tx_size_wide[tx_size];
+  const uint8_t txh = tx_size_high[tx_size];
+  const int above = *above_ctx < txw;
+  const int left = *left_ctx < txh;
+  TX_SIZE max_tx_size = max_txsize_lookup[bsize];
+  int category = TXFM_PARTITION_CONTEXTS - 1;
+
+  // dummy return, not used by others.
+  if (tx_size <= TX_4X4) return 0;
+
+  switch (AOMMAX(block_size_wide[bsize], block_size_high[bsize])) {
+#if CONFIG_EXT_PARTITION
+    case 128:
+#endif
+    case 64:
+    case 32: max_tx_size = TX_32X32; break;
+    case 16: max_tx_size = TX_16X16; break;
+    case 8: max_tx_size = TX_8X8; break;
+    default: assert(0);
+  }
+
+  if (max_tx_size >= TX_8X8) {
+    category = (tx_size != max_tx_size && max_tx_size > TX_8X8) +
+               (TX_SIZES - 1 - max_tx_size) * 2;
+  }
+  if (category == TXFM_PARTITION_CONTEXTS - 1) return category;
+  return category * 3 + above + left;
+}
+#endif
+
+static INLINE PARTITION_TYPE get_partition(const AV1_COMMON *const cm,
+                                           int mi_row, int mi_col,
+                                           BLOCK_SIZE bsize) {
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) {
+    return PARTITION_INVALID;
+  } else {
+    const int offset = mi_row * cm->mi_stride + mi_col;
+    MODE_INFO **mi = cm->mi_grid_visible + offset;
+    const MB_MODE_INFO *const mbmi = &mi[0]->mbmi;
+    const int bsl = b_width_log2_lookup[bsize];
+    const PARTITION_TYPE partition = partition_lookup[bsl][mbmi->sb_type];
+#if !CONFIG_EXT_PARTITION_TYPES
+    return partition;
+#else
+    const int hbs = mi_size_wide[bsize] / 2;
+
+    assert(cm->mi_grid_visible[offset] == &cm->mi[offset]);
+
+    if (partition != PARTITION_NONE && bsize > BLOCK_8X8 &&
+        mi_row + hbs < cm->mi_rows && mi_col + hbs < cm->mi_cols) {
+      const BLOCK_SIZE h = get_subsize(bsize, PARTITION_HORZ_A);
+      const BLOCK_SIZE v = get_subsize(bsize, PARTITION_VERT_A);
+      const MB_MODE_INFO *const mbmi_right = &mi[hbs]->mbmi;
+      const MB_MODE_INFO *const mbmi_below = &mi[hbs * cm->mi_stride]->mbmi;
+      if (mbmi->sb_type == h) {
+        return mbmi_below->sb_type == h ? PARTITION_HORZ : PARTITION_HORZ_B;
+      } else if (mbmi->sb_type == v) {
+        return mbmi_right->sb_type == v ? PARTITION_VERT : PARTITION_VERT_B;
+      } else if (mbmi_below->sb_type == h) {
+        return PARTITION_HORZ_A;
+      } else if (mbmi_right->sb_type == v) {
+        return PARTITION_VERT_A;
+      } else {
+        return PARTITION_SPLIT;
+      }
+    }
+
+    return partition;
+#endif  // !CONFIG_EXT_PARTITION_TYPES
+  }
+}
+
+static INLINE void set_sb_size(AV1_COMMON *const cm, BLOCK_SIZE sb_size) {
+  cm->sb_size = sb_size;
+  cm->mib_size = mi_size_wide[cm->sb_size];
+#if CONFIG_CB4X4
+  cm->mib_size_log2 = b_width_log2_lookup[cm->sb_size];
+#else
+  cm->mib_size_log2 = mi_width_log2_lookup[cm->sb_size];
+#endif
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_ONYXC_INT_H_
diff --git a/third_party/aom/av1/common/partition.c b/third_party/aom/av1/common/partition.c
new file mode 100644
index 0000000000..634a9edd59
--- /dev/null
+++ b/third_party/aom/av1/common/partition.c
@@ -0,0 +1,256 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include "enums.h"
+#include "odintrin.h"
+#include "partition.h"
+#include "zigzag.h"
+
+OD_EXTERN const index_pair *OD_ZIGZAG4[4] = {
+  OD_ZIGZAG4_DCT_DCT,
+  OD_ZIGZAG4_ADST_DCT,
+  OD_ZIGZAG4_DCT_ADST,
+  OD_ZIGZAG4_ADST_ADST
+};
+
+OD_EXTERN const index_pair *OD_ZIGZAG8[4] = {
+  OD_ZIGZAG8_DCT_DCT,
+  OD_ZIGZAG8_ADST_DCT,
+  OD_ZIGZAG8_DCT_ADST,
+  OD_ZIGZAG8_ADST_ADST
+};
+
+OD_EXTERN const index_pair *OD_ZIGZAG16[4] = {
+  OD_ZIGZAG16_DCT_DCT,
+  OD_ZIGZAG16_ADST_DCT,
+  OD_ZIGZAG16_DCT_ADST,
+  OD_ZIGZAG16_ADST_ADST
+};
+
+OD_EXTERN const index_pair *OD_ZIGZAG32[4] = {
+  OD_ZIGZAG32_DCT_DCT,
+  OD_ZIGZAG32_DCT_DCT,
+  OD_ZIGZAG32_DCT_DCT,
+  OD_ZIGZAG32_DCT_DCT
+};
+
+/* The tables below specify how coefficient blocks are translated to
+   and from PVQ partition coding scan order for 4x4, 8x8 and 16x16 */
+
+static const int OD_LAYOUT32_OFFSETS[4] = { 0, 128, 256, 768 };
+const band_layout OD_LAYOUT32 = {
+  OD_ZIGZAG32,
+  32,
+  3,
+  OD_LAYOUT32_OFFSETS
+};
+
+static const int OD_LAYOUT16_OFFSETS[4] = { 0, 32, 64, 192 };
+const band_layout OD_LAYOUT16 = {
+  OD_ZIGZAG16,
+  16,
+  3,
+  OD_LAYOUT16_OFFSETS
+};
+
+const int OD_LAYOUT8_OFFSETS[4] = { 0, 8, 16, 48 };
+const band_layout OD_LAYOUT8 = {
+  OD_ZIGZAG8,
+  8,
+  3,
+  OD_LAYOUT8_OFFSETS
+};
+
+static const int OD_LAYOUT4_OFFSETS[2] = { 0, 15 };
+const band_layout OD_LAYOUT4 = {
+  OD_ZIGZAG4,
+  4,
+  1,
+  OD_LAYOUT4_OFFSETS
+};
+
+/* First element is the number of bands, followed by the list all the band
+  boundaries. */
+static const int OD_BAND_OFFSETS4[] = {1, 1, 16};
+static const int OD_BAND_OFFSETS8[] = {4, 1, 16, 24, 32, 64};
+static const int OD_BAND_OFFSETS16[] = {7, 1, 16, 24, 32, 64, 96, 128, 256};
+static const int OD_BAND_OFFSETS32[] = {10, 1, 16, 24, 32, 64, 96, 128, 256,
+ 384, 512, 1024};
+static const int OD_BAND_OFFSETS64[] = {13, 1, 16, 24, 32, 64, 96, 128, 256,
+ 384, 512, 1024, 1536, 2048, 4096};
+
+const int *const OD_BAND_OFFSETS[OD_TXSIZES + 1] = {
+  OD_BAND_OFFSETS4,
+  OD_BAND_OFFSETS8,
+  OD_BAND_OFFSETS16,
+  OD_BAND_OFFSETS32,
+  OD_BAND_OFFSETS64
+};
+
+/** Perform a single stage of conversion from a coefficient block in
+ * raster order into coding scan order
+ *
+ * @param [in]     layout  scan order specification
+ * @param [out]    dst     destination vector
+ * @param [in]     src     source coefficient block
+ * @param [int]    int     source vector row stride
+ */
+static void od_band_from_raster(const band_layout *layout, tran_low_t *dst,
+ const tran_low_t *src, int stride, TX_TYPE tx_type) {
+  int i;
+  int len;
+  len = layout->band_offsets[layout->nb_bands];
+  for (i = 0; i < len; i++) {
+    dst[i] = src[layout->dst_table[tx_type][i][1]*stride + layout->dst_table[tx_type][i][0]];
+  }
+}
+
+/** Perform a single stage of conversion from a vector in coding scan
+    order back into a coefficient block in raster order
+ *
+ * @param [in]     layout  scan order specification
+ * @param [out]    dst     destination coefficient block
+ * @param [in]     src     source vector
+ * @param [int]    stride  destination vector row stride
+ */
+static void od_raster_from_band(const band_layout *layout, tran_low_t *dst,
+ int stride, TX_TYPE tx_type, const tran_low_t *src) {
+  int i;
+  int len;
+  len = layout->band_offsets[layout->nb_bands];
+  for (i = 0; i < len; i++) {
+    dst[layout->dst_table[tx_type][i][1]*stride + layout->dst_table[tx_type][i][0]] = src[i];
+  }
+}
+
+static const band_layout *const OD_LAYOUTS[] = {&OD_LAYOUT4, &OD_LAYOUT8,
+ &OD_LAYOUT16, &OD_LAYOUT32};
+
+/** Converts a coefficient block in raster order into a vector in
+ * coding scan order with the PVQ partitions laid out one after
+ * another.  This works in stages; the 4x4 conversion is applied to
+ * the coefficients nearest DC, then the 8x8 applied to the 8x8 block
+ * nearest DC that was not already coded by 4x4, then 16x16 following
+ * the same pattern.
+ *
+ * @param [out]    dst        destination vector
+ * @param [in]     n          block size (along one side)
+ * @param [in]     ty_type    transfrom type
+ * @param [in]     src        source coefficient block
+ * @param [in]     stride     source vector row stride
+ */
+void od_raster_to_coding_order(tran_low_t *dst, int n, TX_TYPE ty_type,
+ const tran_low_t *src, int stride) {
+  int bs;
+  /* dst + 1 because DC is not included for 4x4 blocks. */
+  od_band_from_raster(OD_LAYOUTS[0], dst + 1, src, stride, ty_type);
+  for (bs = 1; bs < OD_TXSIZES; bs++) {
+    int size;
+    int offset;
+    /* Length of block size > 4. */
+    size = 1 << (OD_LOG_BSIZE0 + bs);
+    /* Offset is the size of the previous block squared. */
+    offset = 1 << 2*(OD_LOG_BSIZE0 - 1 + bs);
+    if (n >= size) {
+      /* 3 16x16 bands come after 3 8x8 bands, which come after 2 4x4 bands. */
+      od_band_from_raster(OD_LAYOUTS[bs], dst + offset, src, stride, ty_type);
+    }
+  }
+  dst[0] = src[0];
+}
+
+/** Converts a vector in coding scan order witht he PVQ partitions
+ * laid out one after another into a coefficient block in raster
+ * order. This works in stages in the reverse order of raster->scan
+ * order; the 16x16 conversion is applied to the coefficients that
+ * don't appear in an 8x8 block, then the 8x8 applied to the 8x8 block
+ * sans the 4x4 block it contains, then 4x4 is converted sans DC.
+ *
+ * @param [out]    dst        destination coefficient block
+ * @param [in]     stride     destination vector row stride
+ * @param [in]     src        source vector
+ * @param [in]     n          block size (along one side)
+ */
+void od_coding_order_to_raster(tran_low_t *dst, int stride, TX_TYPE ty_type,
+ const tran_low_t *src, int n) {
+  int bs;
+  /* src + 1 because DC is not included for 4x4 blocks. */
+  od_raster_from_band(OD_LAYOUTS[0], dst, stride, ty_type, src + 1);
+  for (bs = 1; bs < OD_TXSIZES; bs++) {
+    int size;
+    int offset;
+    /* Length of block size > 4 */
+    size = 1 << (OD_LOG_BSIZE0 + bs);
+    /* Offset is the size of the previous block squared. */
+    offset = 1 << 2*(OD_LOG_BSIZE0 - 1 + bs);
+    if (n >= size) {
+      /* 3 16x16 bands come after 3 8x8 bands, which come after 2 4x4 bands. */
+      od_raster_from_band(OD_LAYOUTS[bs], dst, stride, ty_type, src + offset);
+    }
+  }
+  dst[0] = src[0];
+}
+
+/** Perform a single stage of conversion from a coefficient block in
+ * raster order into coding scan order
+ *
+ * @param [in]     layout  scan order specification
+ * @param [out]    dst     destination vector
+ * @param [in]     src     source coefficient block
+ * @param [int]    int     source vector row stride
+ */
+static void od_band_from_raster_16(const band_layout *layout, int16_t *dst,
+ const int16_t *src, int stride) {
+  int i;
+  int len;
+  len = layout->band_offsets[layout->nb_bands];
+  for (i = 0; i < len; i++) {
+    dst[i] = src[layout->dst_table[DCT_DCT][i][1]*stride + layout->dst_table[DCT_DCT][i][0]];
+  }
+}
+
+/** Converts a coefficient block in raster order into a vector in
+ * coding scan order with the PVQ partitions laid out one after
+ * another.  This works in stages; the 4x4 conversion is applied to
+ * the coefficients nearest DC, then the 8x8 applied to the 8x8 block
+ * nearest DC that was not already coded by 4x4, then 16x16 following
+ * the same pattern.
+ *
+ * @param [out]    dst        destination vector
+ * @param [in]     n          block size (along one side)
+ * @param [in]     src        source coefficient block
+ * @param [in]     stride     source vector row stride
+ */
+void od_raster_to_coding_order_16(int16_t *dst, int n, const int16_t *src,
+ int stride) {
+  int bs;
+  /* dst + 1 because DC is not included for 4x4 blocks. */
+  od_band_from_raster_16(OD_LAYOUTS[0], dst + 1, src, stride);
+  for (bs = 1; bs < OD_TXSIZES; bs++) {
+    int size;
+    int offset;
+    /* Length of block size > 4. */
+    size = 1 << (OD_LOG_BSIZE0 + bs);
+    /* Offset is the size of the previous block squared. */
+    offset = 1 << 2*(OD_LOG_BSIZE0 - 1 + bs);
+    if (n >= size) {
+      /* 3 16x16 bands come after 3 8x8 bands, which come after 2 4x4 bands. */
+      od_band_from_raster_16(OD_LAYOUTS[bs], dst + offset, src, stride);
+    }
+  }
+  dst[0] = src[0];
+}
diff --git a/third_party/aom/av1/common/partition.h b/third_party/aom/av1/common/partition.h
new file mode 100644
index 0000000000..bd308f94f9
--- /dev/null
+++ b/third_party/aom/av1/common/partition.h
@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_partition_H)
+# define _partition_H
+
+#include "av1/common/enums.h"
+#include "odintrin.h"
+
+typedef unsigned char index_pair[2];
+
+typedef struct {
+  const index_pair **const dst_table;
+  int size;
+  int nb_bands;
+  const int *const band_offsets;
+} band_layout;
+
+extern const int *const OD_BAND_OFFSETS[OD_TXSIZES + 1];
+
+void od_raster_to_coding_order(tran_low_t *dst, int n,  TX_TYPE ty_type,
+ const tran_low_t *src, int stride);
+
+void od_coding_order_to_raster(tran_low_t *dst, int stride,  TX_TYPE ty_type,
+ const tran_low_t *src, int n);
+
+void od_raster_to_coding_order_16(int16_t *dst, int n, const int16_t *src,
+ int stride);
+
+#endif
diff --git a/third_party/aom/av1/common/pred_common.c b/third_party/aom/av1/common/pred_common.c
new file mode 100644
index 0000000000..905dd3afef
--- /dev/null
+++ b/third_party/aom/av1/common/pred_common.c
@@ -0,0 +1,1408 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/common.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/reconinter.h"
+#if CONFIG_EXT_INTRA
+#include "av1/common/reconintra.h"
+#endif  // CONFIG_EXT_INTRA
+#include "av1/common/seg_common.h"
+
+// Returns a context number for the given MB prediction signal
+#if CONFIG_DUAL_FILTER
+static InterpFilter get_ref_filter_type(const MODE_INFO *mi,
+                                        const MACROBLOCKD *xd, int dir,
+                                        MV_REFERENCE_FRAME ref_frame) {
+  InterpFilter ref_type = SWITCHABLE_FILTERS;
+  const MB_MODE_INFO *ref_mbmi = &mi->mbmi;
+  int use_subpel[2] = {
+    has_subpel_mv_component(mi, xd, dir),
+    has_subpel_mv_component(mi, xd, dir + 2),
+  };
+
+  if (ref_mbmi->ref_frame[0] == ref_frame && use_subpel[0])
+    ref_type = ref_mbmi->interp_filter[(dir & 0x01)];
+  else if (ref_mbmi->ref_frame[1] == ref_frame && use_subpel[1])
+    ref_type = ref_mbmi->interp_filter[(dir & 0x01) + 2];
+
+  return ref_type;
+}
+
+int av1_get_pred_context_switchable_interp(const MACROBLOCKD *xd, int dir) {
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int ctx_offset =
+      (mbmi->ref_frame[1] > INTRA_FRAME) * INTER_FILTER_COMP_OFFSET;
+  MV_REFERENCE_FRAME ref_frame =
+      (dir < 2) ? mbmi->ref_frame[0] : mbmi->ref_frame[1];
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  int filter_type_ctx = ctx_offset + (dir & 0x01) * INTER_FILTER_DIR_OFFSET;
+  int left_type = SWITCHABLE_FILTERS;
+  int above_type = SWITCHABLE_FILTERS;
+
+  if (xd->left_available)
+    left_type = get_ref_filter_type(xd->mi[-1], xd, dir, ref_frame);
+
+  if (xd->up_available)
+    above_type =
+        get_ref_filter_type(xd->mi[-xd->mi_stride], xd, dir, ref_frame);
+
+  if (left_type == above_type) {
+    filter_type_ctx += left_type;
+  } else if (left_type == SWITCHABLE_FILTERS) {
+    assert(above_type != SWITCHABLE_FILTERS);
+    filter_type_ctx += above_type;
+  } else if (above_type == SWITCHABLE_FILTERS) {
+    assert(left_type != SWITCHABLE_FILTERS);
+    filter_type_ctx += left_type;
+  } else {
+    filter_type_ctx += SWITCHABLE_FILTERS;
+  }
+
+  return filter_type_ctx;
+}
+#else
+int av1_get_pred_context_switchable_interp(const MACROBLOCKD *xd) {
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int left_type = xd->left_available && is_inter_block(left_mbmi)
+                            ? left_mbmi->interp_filter
+                            : SWITCHABLE_FILTERS;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const int above_type = xd->up_available && is_inter_block(above_mbmi)
+                             ? above_mbmi->interp_filter
+                             : SWITCHABLE_FILTERS;
+
+  if (left_type == above_type) {
+    return left_type;
+  } else if (left_type == SWITCHABLE_FILTERS) {
+    assert(above_type != SWITCHABLE_FILTERS);
+    return above_type;
+  } else if (above_type == SWITCHABLE_FILTERS) {
+    assert(left_type != SWITCHABLE_FILTERS);
+    return left_type;
+  } else {
+    return SWITCHABLE_FILTERS;
+  }
+}
+#endif
+
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+// Obtain the reference filter type from the above/left neighbor blocks.
+static INTRA_FILTER get_ref_intra_filter(const MB_MODE_INFO *ref_mbmi) {
+  INTRA_FILTER ref_type = INTRA_FILTERS;
+
+  if (ref_mbmi->sb_type >= BLOCK_8X8) {
+    const PREDICTION_MODE mode = ref_mbmi->mode;
+    if (is_inter_block(ref_mbmi)) {
+#if CONFIG_DUAL_FILTER
+      switch (ref_mbmi->interp_filter[0]) {
+#else
+      switch (ref_mbmi->interp_filter) {
+#endif
+        case EIGHTTAP_REGULAR: ref_type = INTRA_FILTER_8TAP; break;
+        case EIGHTTAP_SMOOTH: ref_type = INTRA_FILTER_8TAP_SMOOTH; break;
+        case MULTITAP_SHARP: ref_type = INTRA_FILTER_8TAP_SHARP; break;
+        case BILINEAR: ref_type = INTRA_FILTERS; break;
+        default: break;
+      }
+    } else {
+      if (av1_is_directional_mode(mode, ref_mbmi->sb_type)) {
+        const int p_angle =
+            mode_to_angle_map[mode] + ref_mbmi->angle_delta[0] * ANGLE_STEP;
+        if (av1_is_intra_filter_switchable(p_angle)) {
+          ref_type = ref_mbmi->intra_filter;
+        }
+      }
+    }
+  }
+  return ref_type;
+}
+
+int av1_get_pred_context_intra_interp(const MACROBLOCKD *xd) {
+  int left_type = INTRA_FILTERS, above_type = INTRA_FILTERS;
+
+  if (xd->left_available) left_type = get_ref_intra_filter(xd->left_mbmi);
+
+  if (xd->up_available) above_type = get_ref_intra_filter(xd->above_mbmi);
+
+  if (left_type == above_type)
+    return left_type;
+  else if (left_type == INTRA_FILTERS && above_type != INTRA_FILTERS)
+    return above_type;
+  else if (left_type != INTRA_FILTERS && above_type == INTRA_FILTERS)
+    return left_type;
+  else
+    return INTRA_FILTERS;
+}
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+
+// The mode info data structure has a one element border above and to the
+// left of the entries corresponding to real macroblocks.
+// The prediction flags in these dummy entries are initialized to 0.
+// 0 - inter/inter, inter/--, --/inter, --/--
+// 1 - intra/inter, inter/intra
+// 2 - intra/--, --/intra
+// 3 - intra/intra
+int av1_get_intra_inter_context(const MACROBLOCKD *xd) {
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+    return left_intra && above_intra ? 3 : left_intra || above_intra;
+  } else if (has_above || has_left) {  // one edge available
+    return 2 * !is_inter_block(has_above ? above_mbmi : left_mbmi);
+  } else {
+    return 0;
+  }
+}
+
+#if CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+// The compound/single mode info data structure has one element border above and
+// to the left of the entries corresponding to real macroblocks.
+// The prediction flags in these dummy entries are initialized to 0.
+// 0 - single/single
+// 1 - single/--, --/single, --/--
+// 2 - single/comp, comp/single
+// 3 - comp/comp, comp/--, --/comp
+int av1_get_inter_mode_context(const MACROBLOCKD *xd) {
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+  if (has_above && has_left) {  // both edges available (0/2/3)
+    const int above_inter_comp_mode = is_inter_compound_mode(above_mbmi->mode);
+    const int left_inter_comp_mode = is_inter_compound_mode(left_mbmi->mode);
+    return (above_inter_comp_mode && left_inter_comp_mode)
+               ? 3
+               : (above_inter_comp_mode || left_inter_comp_mode) * 2;
+  } else if (has_above || has_left) {  // one edge available (1/3)
+    const MB_MODE_INFO *const edge_mbmi = has_above ? above_mbmi : left_mbmi;
+    return is_inter_compound_mode(edge_mbmi->mode) ? 3 : 1;
+  } else {  // no edge available (1)
+    return 1;
+  }
+}
+#endif  // CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+
+#if CONFIG_EXT_REFS
+#define CHECK_BACKWARD_REFS(ref_frame) \
+  (((ref_frame) >= BWDREF_FRAME) && ((ref_frame) <= ALTREF_FRAME))
+#define IS_BACKWARD_REF_FRAME(ref_frame) CHECK_BACKWARD_REFS(ref_frame)
+#else
+#define IS_BACKWARD_REF_FRAME(ref_frame) ((ref_frame) == cm->comp_fixed_ref)
+#endif  // CONFIG_EXT_REFS
+
+int av1_get_reference_mode_context(const AV1_COMMON *cm,
+                                   const MACROBLOCKD *xd) {
+  int ctx;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+#if CONFIG_EXT_REFS
+  (void)cm;
+#endif  // CONFIG_EXT_REFS
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  if (has_above && has_left) {  // both edges available
+    if (!has_second_ref(above_mbmi) && !has_second_ref(left_mbmi))
+      // neither edge uses comp pred (0/1)
+      ctx = IS_BACKWARD_REF_FRAME(above_mbmi->ref_frame[0]) ^
+            IS_BACKWARD_REF_FRAME(left_mbmi->ref_frame[0]);
+    else if (!has_second_ref(above_mbmi))
+      // one of two edges uses comp pred (2/3)
+      ctx = 2 + (IS_BACKWARD_REF_FRAME(above_mbmi->ref_frame[0]) ||
+                 !is_inter_block(above_mbmi));
+    else if (!has_second_ref(left_mbmi))
+      // one of two edges uses comp pred (2/3)
+      ctx = 2 + (IS_BACKWARD_REF_FRAME(left_mbmi->ref_frame[0]) ||
+                 !is_inter_block(left_mbmi));
+    else  // both edges use comp pred (4)
+      ctx = 4;
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+
+    if (!has_second_ref(edge_mbmi))
+      // edge does not use comp pred (0/1)
+      ctx = IS_BACKWARD_REF_FRAME(edge_mbmi->ref_frame[0]);
+    else
+      // edge uses comp pred (3)
+      ctx = 3;
+  } else {  // no edges available (1)
+    ctx = 1;
+  }
+  assert(ctx >= 0 && ctx < COMP_INTER_CONTEXTS);
+  return ctx;
+}
+
+#if CONFIG_EXT_REFS
+
+// TODO(zoeliu): Future work will be conducted to optimize the context design
+//               for the coding of the reference frames.
+
+#define CHECK_LAST_OR_LAST2(ref_frame) \
+  ((ref_frame == LAST_FRAME) || (ref_frame == LAST2_FRAME))
+
+#define CHECK_GOLDEN_OR_LAST3(ref_frame) \
+  ((ref_frame == GOLDEN_FRAME) || (ref_frame == LAST3_FRAME))
+
+// Returns a context number for the given MB prediction signal
+// Signal the first reference frame for a compound mode be either
+// GOLDEN/LAST3, or LAST/LAST2.
+//
+// NOTE(zoeliu): The probability of ref_frame[0] is either
+//               GOLDEN_FRAME or LAST3_FRAME.
+#if CONFIG_LOWDELAY_COMPOUND
+int av1_get_pred_context_comp_ref_p(UNUSED const AV1_COMMON *cm,
+                                    const MACROBLOCKD *xd) {
+#else
+int av1_get_pred_context_comp_ref_p(const AV1_COMMON *cm,
+                                    const MACROBLOCKD *xd) {
+#endif
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int above_in_image = xd->up_available;
+  const int left_in_image = xd->left_available;
+
+// Note:
+// The mode info data structure has a one element border above and to the
+// left of the entries correpsonding to real macroblocks.
+// The prediction flags in these dummy entries are initialised to 0.
+#if CONFIG_LOWDELAY_COMPOUND  // No change to bitstream
+  // Code seems to assume that signbias of cm->comp_bwd_ref[0] is always 1
+  const int bwd_ref_sign_idx = 1;
+#else
+  const int bwd_ref_sign_idx = cm->ref_frame_sign_bias[cm->comp_bwd_ref[0]];
+#endif
+  const int fwd_ref_sign_idx = !bwd_ref_sign_idx;
+
+  if (above_in_image && left_in_image) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra (2)
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+
+      if (!has_second_ref(edge_mbmi))  // single pred (1/3)
+        pred_context =
+            1 + 2 * (!CHECK_GOLDEN_OR_LAST3(edge_mbmi->ref_frame[0]));
+      else  // comp pred (1/3)
+        pred_context = 1 +
+                       2 * (!CHECK_GOLDEN_OR_LAST3(
+                               edge_mbmi->ref_frame[fwd_ref_sign_idx]));
+    } else {  // inter/inter
+      const int l_sg = !has_second_ref(left_mbmi);
+      const int a_sg = !has_second_ref(above_mbmi);
+      const MV_REFERENCE_FRAME frfa =
+          a_sg ? above_mbmi->ref_frame[0]
+               : above_mbmi->ref_frame[fwd_ref_sign_idx];
+      const MV_REFERENCE_FRAME frfl =
+          l_sg ? left_mbmi->ref_frame[0]
+               : left_mbmi->ref_frame[fwd_ref_sign_idx];
+
+      if (frfa == frfl && CHECK_GOLDEN_OR_LAST3(frfa)) {
+        pred_context = 0;
+      } else if (l_sg && a_sg) {  // single/single
+        if ((CHECK_BACKWARD_REFS(frfa) && CHECK_LAST_OR_LAST2(frfl)) ||
+            (CHECK_BACKWARD_REFS(frfl) && CHECK_LAST_OR_LAST2(frfa))) {
+          pred_context = 4;
+        } else if (CHECK_GOLDEN_OR_LAST3(frfa) || CHECK_GOLDEN_OR_LAST3(frfl)) {
+          pred_context = 1;
+        } else {
+          pred_context = 3;
+        }
+      } else if (l_sg || a_sg) {  // single/comp
+        const MV_REFERENCE_FRAME frfc = l_sg ? frfa : frfl;
+        const MV_REFERENCE_FRAME rfs = a_sg ? frfa : frfl;
+
+        if (CHECK_GOLDEN_OR_LAST3(frfc) && !CHECK_GOLDEN_OR_LAST3(rfs))
+          pred_context = 1;
+        else if (CHECK_GOLDEN_OR_LAST3(rfs) && !CHECK_GOLDEN_OR_LAST3(frfc))
+          pred_context = 2;
+        else
+          pred_context = 4;
+      } else {  // comp/comp
+        if ((CHECK_LAST_OR_LAST2(frfa) && CHECK_LAST_OR_LAST2(frfl))) {
+          pred_context = 4;
+        } else {
+          // NOTE(zoeliu): Following assert may be removed once confirmed.
+          assert(CHECK_GOLDEN_OR_LAST3(frfa) || CHECK_GOLDEN_OR_LAST3(frfl));
+          pred_context = 2;
+        }
+      }
+    }
+  } else if (above_in_image || left_in_image) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = above_in_image ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi)) {
+      pred_context = 2;
+    } else {
+      if (has_second_ref(edge_mbmi))
+        pred_context =
+            4 *
+            (!CHECK_GOLDEN_OR_LAST3(edge_mbmi->ref_frame[fwd_ref_sign_idx]));
+      else
+        pred_context = 3 * (!CHECK_GOLDEN_OR_LAST3(edge_mbmi->ref_frame[0]));
+    }
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+
+  return pred_context;
+}
+
+// Returns a context number for the given MB prediction signal
+// Signal the first reference frame for a compound mode be LAST,
+// conditioning on that it is known either LAST/LAST2.
+//
+// NOTE(zoeliu): The probability of ref_frame[0] is LAST_FRAME,
+// conditioning on it is either LAST_FRAME or LAST2_FRAME.
+#if CONFIG_LOWDELAY_COMPOUND
+int av1_get_pred_context_comp_ref_p1(UNUSED const AV1_COMMON *cm,
+                                     const MACROBLOCKD *xd) {
+#else
+int av1_get_pred_context_comp_ref_p1(const AV1_COMMON *cm,
+                                     const MACROBLOCKD *xd) {
+#endif
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int above_in_image = xd->up_available;
+  const int left_in_image = xd->left_available;
+
+// Note:
+// The mode info data structure has a one element border above and to the
+// left of the entries correpsonding to real macroblocks.
+// The prediction flags in these dummy entries are initialised to 0.
+#if CONFIG_LOWDELAY_COMPOUND  // No change to bitstream
+  // Code seems to assume that signbias of cm->comp_bwd_ref[0] is always 1
+  const int bwd_ref_sign_idx = 1;
+#else
+  const int bwd_ref_sign_idx = cm->ref_frame_sign_bias[cm->comp_bwd_ref[0]];
+#endif
+  const int fwd_ref_sign_idx = !bwd_ref_sign_idx;
+
+  if (above_in_image && left_in_image) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra (2)
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+
+      if (!has_second_ref(edge_mbmi))  // single pred (1/3)
+        pred_context = 1 + 2 * (edge_mbmi->ref_frame[0] != LAST_FRAME);
+      else  // comp pred (1/3)
+        pred_context =
+            1 + 2 * (edge_mbmi->ref_frame[fwd_ref_sign_idx] != LAST_FRAME);
+    } else {  // inter/inter
+      const int l_sg = !has_second_ref(left_mbmi);
+      const int a_sg = !has_second_ref(above_mbmi);
+      const MV_REFERENCE_FRAME frfa =
+          a_sg ? above_mbmi->ref_frame[0]
+               : above_mbmi->ref_frame[fwd_ref_sign_idx];
+      const MV_REFERENCE_FRAME frfl =
+          l_sg ? left_mbmi->ref_frame[0]
+               : left_mbmi->ref_frame[fwd_ref_sign_idx];
+
+      if (frfa == frfl && frfa == LAST_FRAME)
+        pred_context = 0;
+      else if (l_sg && a_sg) {  // single/single
+        if (frfa == LAST_FRAME || frfl == LAST_FRAME)
+          pred_context = 1;
+        else if (CHECK_GOLDEN_OR_LAST3(frfa) || CHECK_GOLDEN_OR_LAST3(frfl))
+          pred_context = 2 + (frfa != frfl);
+        else if (frfa == frfl ||
+                 (CHECK_BACKWARD_REFS(frfa) && CHECK_BACKWARD_REFS(frfl)))
+          pred_context = 3;
+        else
+          pred_context = 4;
+      } else if (l_sg || a_sg) {  // single/comp
+        const MV_REFERENCE_FRAME frfc = l_sg ? frfa : frfl;
+        const MV_REFERENCE_FRAME rfs = a_sg ? frfa : frfl;
+
+        if (frfc == LAST_FRAME && rfs != LAST_FRAME)
+          pred_context = 1;
+        else if (rfs == LAST_FRAME && frfc != LAST_FRAME)
+          pred_context = 2;
+        else
+          pred_context =
+              3 + (frfc == LAST2_FRAME || CHECK_GOLDEN_OR_LAST3(rfs));
+      } else {  // comp/comp
+        if (frfa == LAST_FRAME || frfl == LAST_FRAME)
+          pred_context = 2;
+        else
+          pred_context =
+              3 + (CHECK_GOLDEN_OR_LAST3(frfa) || CHECK_GOLDEN_OR_LAST3(frfl));
+      }
+    }
+  } else if (above_in_image || left_in_image) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = above_in_image ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi)) {
+      pred_context = 2;
+    } else {
+      if (has_second_ref(edge_mbmi)) {
+        pred_context =
+            4 * (edge_mbmi->ref_frame[fwd_ref_sign_idx] != LAST_FRAME);
+      } else {
+        if (edge_mbmi->ref_frame[0] == LAST_FRAME)
+          pred_context = 0;
+        else
+          pred_context = 2 + CHECK_GOLDEN_OR_LAST3(edge_mbmi->ref_frame[0]);
+      }
+    }
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+
+  return pred_context;
+}
+
+// Returns a context number for the given MB prediction signal
+// Signal the first reference frame for a compound mode be GOLDEN,
+// conditioning on that it is known either GOLDEN or LAST3.
+//
+// NOTE(zoeliu): The probability of ref_frame[0] is GOLDEN_FRAME,
+// conditioning on it is either GOLDEN or LAST3.
+#if CONFIG_LOWDELAY_COMPOUND
+int av1_get_pred_context_comp_ref_p2(UNUSED const AV1_COMMON *cm,
+                                     const MACROBLOCKD *xd) {
+#else
+int av1_get_pred_context_comp_ref_p2(const AV1_COMMON *cm,
+                                     const MACROBLOCKD *xd) {
+#endif
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int above_in_image = xd->up_available;
+  const int left_in_image = xd->left_available;
+
+// Note:
+// The mode info data structure has a one element border above and to the
+// left of the entries correpsonding to real macroblocks.
+// The prediction flags in these dummy entries are initialised to 0.
+#if CONFIG_LOWDELAY_COMPOUND  // No change to bitstream
+  // Code seems to assume that signbias of cm->comp_bwd_ref[0] is always 1
+  const int bwd_ref_sign_idx = 1;
+#else
+  const int bwd_ref_sign_idx = cm->ref_frame_sign_bias[cm->comp_bwd_ref[0]];
+#endif
+  const int fwd_ref_sign_idx = !bwd_ref_sign_idx;
+
+  if (above_in_image && left_in_image) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra (2)
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+
+      if (!has_second_ref(edge_mbmi))  // single pred (1/3)
+        pred_context = 1 + 2 * (edge_mbmi->ref_frame[0] != GOLDEN_FRAME);
+      else  // comp pred (1/3)
+        pred_context =
+            1 + 2 * (edge_mbmi->ref_frame[fwd_ref_sign_idx] != GOLDEN_FRAME);
+    } else {  // inter/inter
+      const int l_sg = !has_second_ref(left_mbmi);
+      const int a_sg = !has_second_ref(above_mbmi);
+      const MV_REFERENCE_FRAME frfa =
+          a_sg ? above_mbmi->ref_frame[0]
+               : above_mbmi->ref_frame[fwd_ref_sign_idx];
+      const MV_REFERENCE_FRAME frfl =
+          l_sg ? left_mbmi->ref_frame[0]
+               : left_mbmi->ref_frame[fwd_ref_sign_idx];
+
+      if (frfa == frfl && frfa == GOLDEN_FRAME)
+        pred_context = 0;
+      else if (l_sg && a_sg) {  // single/single
+        if (frfa == GOLDEN_FRAME || frfl == GOLDEN_FRAME)
+          pred_context = 1;
+        else if (CHECK_LAST_OR_LAST2(frfa) || CHECK_LAST_OR_LAST2(frfl))
+          pred_context = 2 + (frfa != frfl);
+        else if (frfa == frfl ||
+                 (CHECK_BACKWARD_REFS(frfa) && CHECK_BACKWARD_REFS(frfl)))
+          pred_context = 3;
+        else
+          pred_context = 4;
+      } else if (l_sg || a_sg) {  // single/comp
+        const MV_REFERENCE_FRAME frfc = l_sg ? frfa : frfl;
+        const MV_REFERENCE_FRAME rfs = a_sg ? frfa : frfl;
+
+        if (frfc == GOLDEN_FRAME && rfs != GOLDEN_FRAME)
+          pred_context = 1;
+        else if (rfs == GOLDEN_FRAME && frfc != GOLDEN_FRAME)
+          pred_context = 2;
+        else
+          pred_context = 3 + (frfc == LAST3_FRAME || CHECK_LAST_OR_LAST2(rfs));
+      } else {  // comp/comp
+        if (frfa == GOLDEN_FRAME || frfl == GOLDEN_FRAME)
+          pred_context = 2;
+        else
+          pred_context =
+              3 + (CHECK_LAST_OR_LAST2(frfa) || CHECK_LAST_OR_LAST2(frfl));
+      }
+    }
+  } else if (above_in_image || left_in_image) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = above_in_image ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi)) {
+      pred_context = 2;
+    } else {
+      if (has_second_ref(edge_mbmi)) {
+        pred_context =
+            4 * (edge_mbmi->ref_frame[fwd_ref_sign_idx] != GOLDEN_FRAME);
+      } else {
+        if (edge_mbmi->ref_frame[0] == GOLDEN_FRAME)
+          pred_context = 0;
+        else
+          pred_context = 2 + CHECK_LAST_OR_LAST2(edge_mbmi->ref_frame[0]);
+      }
+    }
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+
+  return pred_context;
+}
+
+// Returns a context number for the given MB prediction signal
+#if CONFIG_LOWDELAY_COMPOUND
+int av1_get_pred_context_comp_bwdref_p(UNUSED const AV1_COMMON *cm,
+                                       const MACROBLOCKD *xd) {
+#else
+int av1_get_pred_context_comp_bwdref_p(const AV1_COMMON *cm,
+                                       const MACROBLOCKD *xd) {
+#endif
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int above_in_image = xd->up_available;
+  const int left_in_image = xd->left_available;
+
+// Note:
+// The mode info data structure has a one element border above and to the
+// left of the entries corresponding to real macroblocks.
+// The prediction flags in these dummy entries are initialized to 0.
+#if CONFIG_LOWDELAY_COMPOUND  // No change to bitstream
+  // Code seems to assume that signbias of cm->comp_bwd_ref[0] is always 1
+  const int bwd_ref_sign_idx = 1;
+#else
+  const int bwd_ref_sign_idx = cm->ref_frame_sign_bias[cm->comp_bwd_ref[0]];
+#endif
+  const int fwd_ref_sign_idx = !bwd_ref_sign_idx;
+
+  if (above_in_image && left_in_image) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra (2)
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+
+      if (!has_second_ref(edge_mbmi))  // single pred (1/3)
+        pred_context = 1 + 2 * (edge_mbmi->ref_frame[0] != cm->comp_bwd_ref[1]);
+      else  // comp pred (1/3)
+        pred_context =
+            1 +
+            2 * (edge_mbmi->ref_frame[bwd_ref_sign_idx] != cm->comp_bwd_ref[1]);
+    } else {  // inter/inter
+      const int l_comp = has_second_ref(left_mbmi);
+      const int a_comp = has_second_ref(above_mbmi);
+
+      const MV_REFERENCE_FRAME l_brf =
+          l_comp ? left_mbmi->ref_frame[bwd_ref_sign_idx] : NONE_FRAME;
+      const MV_REFERENCE_FRAME a_brf =
+          a_comp ? above_mbmi->ref_frame[bwd_ref_sign_idx] : NONE_FRAME;
+
+      const MV_REFERENCE_FRAME l_frf =
+          !l_comp ? left_mbmi->ref_frame[0]
+                  : left_mbmi->ref_frame[fwd_ref_sign_idx];
+      const MV_REFERENCE_FRAME a_frf =
+          !a_comp ? above_mbmi->ref_frame[0]
+                  : above_mbmi->ref_frame[fwd_ref_sign_idx];
+
+      if (l_comp && a_comp) {  // comp/comp
+        if (l_brf == a_brf && l_brf == cm->comp_bwd_ref[1]) {
+          pred_context = 0;
+        } else if (l_brf == cm->comp_bwd_ref[1] ||
+                   a_brf == cm->comp_bwd_ref[1]) {
+          pred_context = 1;
+        } else {
+          // NOTE: Backward ref should be either BWDREF or ALTREF.
+          assert(l_brf == a_brf && l_brf != cm->comp_bwd_ref[1]);
+          pred_context = 3;
+        }
+      } else if (!l_comp && !a_comp) {  // single/single
+        if (l_frf == a_frf && l_frf == cm->comp_bwd_ref[1]) {
+          pred_context = 0;
+        } else if (l_frf == cm->comp_bwd_ref[1] ||
+                   a_frf == cm->comp_bwd_ref[1]) {
+          pred_context = 1;
+        } else if (l_frf == a_frf) {
+          pred_context = 3;
+        } else {
+          assert(l_frf != a_frf && l_frf != cm->comp_bwd_ref[1] &&
+                 a_frf != cm->comp_bwd_ref[1]);
+          pred_context = 4;
+        }
+      } else {  // comp/single
+        assert((l_comp && !a_comp) || (!l_comp && a_comp));
+
+        if ((l_comp && l_brf == cm->comp_bwd_ref[1] &&
+             a_frf == cm->comp_bwd_ref[1]) ||
+            (a_comp && a_brf == cm->comp_bwd_ref[1] &&
+             l_frf == cm->comp_bwd_ref[1])) {
+          pred_context = 1;
+        } else if ((l_comp && l_brf == cm->comp_bwd_ref[1]) ||
+                   (a_comp && a_brf == cm->comp_bwd_ref[1]) ||
+                   (!l_comp && l_frf == cm->comp_bwd_ref[1]) ||
+                   (!a_comp && a_frf == cm->comp_bwd_ref[1])) {
+          pred_context = 2;
+        } else {
+          pred_context = 4;
+        }
+      }
+    }
+  } else if (above_in_image || left_in_image) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = above_in_image ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi)) {
+      pred_context = 2;
+    } else {
+      if (has_second_ref(edge_mbmi)) {
+        pred_context =
+            4 * (edge_mbmi->ref_frame[bwd_ref_sign_idx] != cm->comp_bwd_ref[1]);
+      } else {
+        pred_context = 3 * (edge_mbmi->ref_frame[0] != cm->comp_bwd_ref[1]);
+      }
+    }
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+
+  return pred_context;
+}
+
+#else  // CONFIG_EXT_REFS
+
+// Returns a context number for the given MB prediction signal
+int av1_get_pred_context_comp_ref_p(const AV1_COMMON *cm,
+                                    const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int above_in_image = xd->up_available;
+  const int left_in_image = xd->left_available;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  const int fix_ref_idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
+  const int var_ref_idx = !fix_ref_idx;
+
+  if (above_in_image && left_in_image) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra (2)
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+
+      if (!has_second_ref(edge_mbmi))  // single pred (1/3)
+        pred_context = 1 + 2 * (edge_mbmi->ref_frame[0] != cm->comp_var_ref[1]);
+      else  // comp pred (1/3)
+        pred_context =
+            1 + 2 * (edge_mbmi->ref_frame[var_ref_idx] != cm->comp_var_ref[1]);
+    } else {  // inter/inter
+      const int l_sg = !has_second_ref(left_mbmi);
+      const int a_sg = !has_second_ref(above_mbmi);
+      const MV_REFERENCE_FRAME vrfa =
+          a_sg ? above_mbmi->ref_frame[0] : above_mbmi->ref_frame[var_ref_idx];
+      const MV_REFERENCE_FRAME vrfl =
+          l_sg ? left_mbmi->ref_frame[0] : left_mbmi->ref_frame[var_ref_idx];
+
+      if (vrfa == vrfl && cm->comp_var_ref[1] == vrfa) {
+        pred_context = 0;
+      } else if (l_sg && a_sg) {  // single/single
+        if ((vrfa == cm->comp_fixed_ref && vrfl == cm->comp_var_ref[0]) ||
+            (vrfl == cm->comp_fixed_ref && vrfa == cm->comp_var_ref[0]))
+          pred_context = 4;
+        else if (vrfa == vrfl)
+          pred_context = 3;
+        else
+          pred_context = 1;
+      } else if (l_sg || a_sg) {  // single/comp
+        const MV_REFERENCE_FRAME vrfc = l_sg ? vrfa : vrfl;
+        const MV_REFERENCE_FRAME rfs = a_sg ? vrfa : vrfl;
+        if (vrfc == cm->comp_var_ref[1] && rfs != cm->comp_var_ref[1])
+          pred_context = 1;
+        else if (rfs == cm->comp_var_ref[1] && vrfc != cm->comp_var_ref[1])
+          pred_context = 2;
+        else
+          pred_context = 4;
+      } else if (vrfa == vrfl) {  // comp/comp
+        pred_context = 4;
+      } else {
+        pred_context = 2;
+      }
+    }
+  } else if (above_in_image || left_in_image) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = above_in_image ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi)) {
+      pred_context = 2;
+    } else {
+      if (has_second_ref(edge_mbmi))
+        pred_context =
+            4 * (edge_mbmi->ref_frame[var_ref_idx] != cm->comp_var_ref[1]);
+      else
+        pred_context = 3 * (edge_mbmi->ref_frame[0] != cm->comp_var_ref[1]);
+    }
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+
+  return pred_context;
+}
+
+#endif  // CONFIG_EXT_REFS
+
+#if CONFIG_EXT_REFS
+
+// For the bit to signal whether the single reference is a ALTREF_FRAME
+// or a BWDREF_FRAME.
+//
+// NOTE(zoeliu): The probability of ref_frame[0] is ALTREF/BWDREF.
+int av1_get_pred_context_single_ref_p1(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries correpsonding to real macroblocks.
+  // The prediction flags in these dummy entries are initialised to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+
+      if (!has_second_ref(edge_mbmi))  // single
+        pred_context = 4 * (!CHECK_BACKWARD_REFS(edge_mbmi->ref_frame[0]));
+      else  // comp
+        pred_context = 2;
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mbmi);
+      const int left_has_second = has_second_ref(left_mbmi);
+
+      const MV_REFERENCE_FRAME above0 = above_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME left0 = left_mbmi->ref_frame[0];
+
+      if (above_has_second && left_has_second) {  // comp/comp
+        pred_context = 2;
+      } else if (above_has_second || left_has_second) {  // single/comp
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+
+        pred_context = (!CHECK_BACKWARD_REFS(rfs)) ? 4 : 1;
+      } else {  // single/single
+        pred_context = 2 * (!CHECK_BACKWARD_REFS(above0)) +
+                       2 * (!CHECK_BACKWARD_REFS(left0));
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+    if (!is_inter_block(edge_mbmi)) {  // intra
+      pred_context = 2;
+    } else {                           // inter
+      if (!has_second_ref(edge_mbmi))  // single
+        pred_context = 4 * (!CHECK_BACKWARD_REFS(edge_mbmi->ref_frame[0]));
+      else  // comp
+        pred_context = 2;
+    }
+  } else {  // no edges available
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
+
+// For the bit to signal whether the single reference is ALTREF_FRAME or
+// BWDREF_FRAME, knowing that it shall be either of these 2 choices.
+//
+// NOTE(zoeliu): The probability of ref_frame[0] is ALTREF_FRAME, conditioning
+// on it is either ALTREF_FRAME/BWDREF_FRAME.
+int av1_get_pred_context_single_ref_p2(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries correpsonding to real macroblocks.
+  // The prediction flags in these dummy entries are initialised to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+      if (!has_second_ref(edge_mbmi)) {  // single
+        if (!CHECK_BACKWARD_REFS(edge_mbmi->ref_frame[0]))
+          pred_context = 3;
+        else
+          pred_context = 4 * (edge_mbmi->ref_frame[0] == BWDREF_FRAME);
+      } else {  // comp
+        pred_context = 1 +
+                       2 * (edge_mbmi->ref_frame[0] == BWDREF_FRAME ||
+                            edge_mbmi->ref_frame[1] == BWDREF_FRAME);
+      }
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mbmi);
+      const int left_has_second = has_second_ref(left_mbmi);
+      const MV_REFERENCE_FRAME above0 = above_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME above1 = above_mbmi->ref_frame[1];
+      const MV_REFERENCE_FRAME left0 = left_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME left1 = left_mbmi->ref_frame[1];
+
+      if (above_has_second && left_has_second) {  // comp/comp
+        if (above0 == left0 && above1 == left1)
+          pred_context =
+              3 * (above0 == BWDREF_FRAME || above1 == BWDREF_FRAME ||
+                   left0 == BWDREF_FRAME || left1 == BWDREF_FRAME);
+        else
+          pred_context = 2;
+      } else if (above_has_second || left_has_second) {  // single/comp
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf1 = above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf2 = above_has_second ? above1 : left1;
+
+        if (rfs == BWDREF_FRAME)
+          pred_context = 3 + (crf1 == BWDREF_FRAME || crf2 == BWDREF_FRAME);
+        else if (rfs == ALTREF_FRAME)
+          pred_context = (crf1 == BWDREF_FRAME || crf2 == BWDREF_FRAME);
+        else
+          pred_context = 1 + 2 * (crf1 == BWDREF_FRAME || crf2 == BWDREF_FRAME);
+      } else {  // single/single
+        if (!CHECK_BACKWARD_REFS(above0) && !CHECK_BACKWARD_REFS(left0)) {
+          pred_context = 2 + (above0 == left0);
+        } else if (!CHECK_BACKWARD_REFS(above0) ||
+                   !CHECK_BACKWARD_REFS(left0)) {
+          const MV_REFERENCE_FRAME edge0 =
+              !CHECK_BACKWARD_REFS(above0) ? left0 : above0;
+          pred_context = 4 * (edge0 == BWDREF_FRAME);
+        } else {
+          pred_context =
+              2 * (above0 == BWDREF_FRAME) + 2 * (left0 == BWDREF_FRAME);
+        }
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi) ||
+        (!CHECK_BACKWARD_REFS(edge_mbmi->ref_frame[0]) &&
+         !has_second_ref(edge_mbmi)))
+      pred_context = 2;
+    else if (!has_second_ref(edge_mbmi))  // single
+      pred_context = 4 * (edge_mbmi->ref_frame[0] == BWDREF_FRAME);
+    else  // comp
+      pred_context = 3 * (edge_mbmi->ref_frame[0] == BWDREF_FRAME ||
+                          edge_mbmi->ref_frame[1] == BWDREF_FRAME);
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
+
+// For the bit to signal whether the single reference is LAST3/GOLDEN or
+// LAST2/LAST, knowing that it shall be either of these 2 choices.
+//
+// NOTE(zoeliu): The probability of ref_frame[0] is LAST3/GOLDEN, conditioning
+// on it is either LAST3/GOLDEN/LAST2/LAST.
+int av1_get_pred_context_single_ref_p3(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries correpsonding to real macroblocks.
+  // The prediction flags in these dummy entries are initialised to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+      if (!has_second_ref(edge_mbmi)) {  // single
+        if (CHECK_BACKWARD_REFS(edge_mbmi->ref_frame[0]))
+          pred_context = 3;
+        else
+          pred_context = 4 * CHECK_LAST_OR_LAST2(edge_mbmi->ref_frame[0]);
+      } else {  // comp
+        pred_context = 1 +
+                       2 * (CHECK_LAST_OR_LAST2(edge_mbmi->ref_frame[0]) ||
+                            CHECK_LAST_OR_LAST2(edge_mbmi->ref_frame[1]));
+      }
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mbmi);
+      const int left_has_second = has_second_ref(left_mbmi);
+      const MV_REFERENCE_FRAME above0 = above_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME above1 = above_mbmi->ref_frame[1];
+      const MV_REFERENCE_FRAME left0 = left_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME left1 = left_mbmi->ref_frame[1];
+
+      if (above_has_second && left_has_second) {  // comp/comp
+        if (above0 == left0 && above1 == left1)
+          pred_context =
+              3 * (CHECK_LAST_OR_LAST2(above0) || CHECK_LAST_OR_LAST2(above1) ||
+                   CHECK_LAST_OR_LAST2(left0) || CHECK_LAST_OR_LAST2(left1));
+        else
+          pred_context = 2;
+      } else if (above_has_second || left_has_second) {  // single/comp
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf1 = above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf2 = above_has_second ? above1 : left1;
+
+        if (CHECK_LAST_OR_LAST2(rfs))
+          pred_context =
+              3 + (CHECK_LAST_OR_LAST2(crf1) || CHECK_LAST_OR_LAST2(crf2));
+        else if (CHECK_GOLDEN_OR_LAST3(rfs))
+          pred_context =
+              (CHECK_LAST_OR_LAST2(crf1) || CHECK_LAST_OR_LAST2(crf2));
+        else
+          pred_context =
+              1 + 2 * (CHECK_LAST_OR_LAST2(crf1) || CHECK_LAST_OR_LAST2(crf2));
+      } else {  // single/single
+        if (CHECK_BACKWARD_REFS(above0) && CHECK_BACKWARD_REFS(left0)) {
+          pred_context = 2 + (above0 == left0);
+        } else if (CHECK_BACKWARD_REFS(above0) || CHECK_BACKWARD_REFS(left0)) {
+          const MV_REFERENCE_FRAME edge0 =
+              CHECK_BACKWARD_REFS(above0) ? left0 : above0;
+          pred_context = 4 * CHECK_LAST_OR_LAST2(edge0);
+        } else {
+          pred_context =
+              2 * CHECK_LAST_OR_LAST2(above0) + 2 * CHECK_LAST_OR_LAST2(left0);
+        }
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi) ||
+        (CHECK_BACKWARD_REFS(edge_mbmi->ref_frame[0]) &&
+         !has_second_ref(edge_mbmi)))
+      pred_context = 2;
+    else if (!has_second_ref(edge_mbmi))  // single
+      pred_context = 4 * (CHECK_LAST_OR_LAST2(edge_mbmi->ref_frame[0]));
+    else  // comp
+      pred_context = 3 * (CHECK_LAST_OR_LAST2(edge_mbmi->ref_frame[0]) ||
+                          CHECK_LAST_OR_LAST2(edge_mbmi->ref_frame[1]));
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
+
+// For the bit to signal whether the single reference is LAST2_FRAME or
+// LAST_FRAME, knowing that it shall be either of these 2 choices.
+//
+// NOTE(zoeliu): The probability of ref_frame[0] is LAST2_FRAME, conditioning
+// on it is either LAST2_FRAME/LAST_FRAME.
+int av1_get_pred_context_single_ref_p4(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries correpsonding to real macroblocks.
+  // The prediction flags in these dummy entries are initialised to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+      if (!has_second_ref(edge_mbmi)) {  // single
+        if (!CHECK_LAST_OR_LAST2(edge_mbmi->ref_frame[0]))
+          pred_context = 3;
+        else
+          pred_context = 4 * (edge_mbmi->ref_frame[0] == LAST_FRAME);
+      } else {  // comp
+        pred_context = 1 +
+                       2 * (edge_mbmi->ref_frame[0] == LAST_FRAME ||
+                            edge_mbmi->ref_frame[1] == LAST_FRAME);
+      }
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mbmi);
+      const int left_has_second = has_second_ref(left_mbmi);
+      const MV_REFERENCE_FRAME above0 = above_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME above1 = above_mbmi->ref_frame[1];
+      const MV_REFERENCE_FRAME left0 = left_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME left1 = left_mbmi->ref_frame[1];
+
+      if (above_has_second && left_has_second) {  // comp/comp
+        if (above0 == left0 && above1 == left1)
+          pred_context = 3 * (above0 == LAST_FRAME || above1 == LAST_FRAME ||
+                              left0 == LAST_FRAME || left1 == LAST_FRAME);
+        else
+          pred_context = 2;
+      } else if (above_has_second || left_has_second) {  // single/comp
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf1 = above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf2 = above_has_second ? above1 : left1;
+
+        if (rfs == LAST_FRAME)
+          pred_context = 3 + (crf1 == LAST_FRAME || crf2 == LAST_FRAME);
+        else if (rfs == LAST2_FRAME)
+          pred_context = (crf1 == LAST_FRAME || crf2 == LAST_FRAME);
+        else
+          pred_context = 1 + 2 * (crf1 == LAST_FRAME || crf2 == LAST_FRAME);
+      } else {  // single/single
+        if (!CHECK_LAST_OR_LAST2(above0) && !CHECK_LAST_OR_LAST2(left0)) {
+          pred_context = 2 + (above0 == left0);
+        } else if (!CHECK_LAST_OR_LAST2(above0) ||
+                   !CHECK_LAST_OR_LAST2(left0)) {
+          const MV_REFERENCE_FRAME edge0 =
+              !CHECK_LAST_OR_LAST2(above0) ? left0 : above0;
+          pred_context = 4 * (edge0 == LAST_FRAME);
+        } else {
+          pred_context = 2 * (above0 == LAST_FRAME) + 2 * (left0 == LAST_FRAME);
+        }
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi) ||
+        (!CHECK_LAST_OR_LAST2(edge_mbmi->ref_frame[0]) &&
+         !has_second_ref(edge_mbmi)))
+      pred_context = 2;
+    else if (!has_second_ref(edge_mbmi))  // single
+      pred_context = 4 * (edge_mbmi->ref_frame[0] == LAST_FRAME);
+    else  // comp
+      pred_context = 3 * (edge_mbmi->ref_frame[0] == LAST_FRAME ||
+                          edge_mbmi->ref_frame[1] == LAST_FRAME);
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
+
+// For the bit to signal whether the single reference is GOLDEN_FRAME or
+// LAST3_FRAME, knowing that it shall be either of these 2 choices.
+//
+// NOTE(zoeliu): The probability of ref_frame[0] is GOLDEN_FRAME, conditioning
+// on it is either GOLDEN_FRAME/LAST3_FRAME.
+int av1_get_pred_context_single_ref_p5(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries correpsonding to real macroblocks.
+  // The prediction flags in these dummy entries are initialised to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+      if (!has_second_ref(edge_mbmi)) {  // single
+        if (!CHECK_GOLDEN_OR_LAST3(edge_mbmi->ref_frame[0]))
+          pred_context = 3;
+        else
+          pred_context = 4 * (edge_mbmi->ref_frame[0] == LAST3_FRAME);
+      } else {  // comp
+        pred_context = 1 +
+                       2 * (edge_mbmi->ref_frame[0] == LAST3_FRAME ||
+                            edge_mbmi->ref_frame[1] == LAST3_FRAME);
+      }
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mbmi);
+      const int left_has_second = has_second_ref(left_mbmi);
+      const MV_REFERENCE_FRAME above0 = above_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME above1 = above_mbmi->ref_frame[1];
+      const MV_REFERENCE_FRAME left0 = left_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME left1 = left_mbmi->ref_frame[1];
+
+      if (above_has_second && left_has_second) {  // comp/comp
+        if (above0 == left0 && above1 == left1)
+          pred_context = 3 * (above0 == LAST3_FRAME || above1 == LAST3_FRAME ||
+                              left0 == LAST3_FRAME || left1 == LAST3_FRAME);
+        else
+          pred_context = 2;
+      } else if (above_has_second || left_has_second) {  // single/comp
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf1 = above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf2 = above_has_second ? above1 : left1;
+
+        if (rfs == LAST3_FRAME)
+          pred_context = 3 + (crf1 == LAST3_FRAME || crf2 == LAST3_FRAME);
+        else if (rfs == GOLDEN_FRAME)
+          pred_context = (crf1 == LAST3_FRAME || crf2 == LAST3_FRAME);
+        else
+          pred_context = 1 + 2 * (crf1 == LAST3_FRAME || crf2 == LAST3_FRAME);
+      } else {  // single/single
+        if (!CHECK_GOLDEN_OR_LAST3(above0) && !CHECK_GOLDEN_OR_LAST3(left0)) {
+          pred_context = 2 + (above0 == left0);
+        } else if (!CHECK_GOLDEN_OR_LAST3(above0) ||
+                   !CHECK_GOLDEN_OR_LAST3(left0)) {
+          const MV_REFERENCE_FRAME edge0 =
+              !CHECK_GOLDEN_OR_LAST3(above0) ? left0 : above0;
+          pred_context = 4 * (edge0 == LAST3_FRAME);
+        } else {
+          pred_context =
+              2 * (above0 == LAST3_FRAME) + 2 * (left0 == LAST3_FRAME);
+        }
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi) ||
+        (!CHECK_GOLDEN_OR_LAST3(edge_mbmi->ref_frame[0]) &&
+         !has_second_ref(edge_mbmi)))
+      pred_context = 2;
+    else if (!has_second_ref(edge_mbmi))  // single
+      pred_context = 4 * (edge_mbmi->ref_frame[0] == LAST3_FRAME);
+    else  // comp
+      pred_context = 3 * (edge_mbmi->ref_frame[0] == LAST3_FRAME ||
+                          edge_mbmi->ref_frame[1] == LAST3_FRAME);
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
+
+#else  // CONFIG_EXT_REFS
+
+int av1_get_pred_context_single_ref_p1(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+      if (!has_second_ref(edge_mbmi))
+        pred_context = 4 * (edge_mbmi->ref_frame[0] == LAST_FRAME);
+      else
+        pred_context = 1 + (edge_mbmi->ref_frame[0] == LAST_FRAME ||
+                            edge_mbmi->ref_frame[1] == LAST_FRAME);
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mbmi);
+      const int left_has_second = has_second_ref(left_mbmi);
+      const MV_REFERENCE_FRAME above0 = above_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME above1 = above_mbmi->ref_frame[1];
+      const MV_REFERENCE_FRAME left0 = left_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME left1 = left_mbmi->ref_frame[1];
+
+      if (above_has_second && left_has_second) {
+        pred_context = 1 + (above0 == LAST_FRAME || above1 == LAST_FRAME ||
+                            left0 == LAST_FRAME || left1 == LAST_FRAME);
+      } else if (above_has_second || left_has_second) {
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf1 = above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf2 = above_has_second ? above1 : left1;
+
+        if (rfs == LAST_FRAME)
+          pred_context = 3 + (crf1 == LAST_FRAME || crf2 == LAST_FRAME);
+        else
+          pred_context = (crf1 == LAST_FRAME || crf2 == LAST_FRAME);
+      } else {
+        pred_context = 2 * (above0 == LAST_FRAME) + 2 * (left0 == LAST_FRAME);
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+    if (!is_inter_block(edge_mbmi)) {  // intra
+      pred_context = 2;
+    } else {  // inter
+      if (!has_second_ref(edge_mbmi))
+        pred_context = 4 * (edge_mbmi->ref_frame[0] == LAST_FRAME);
+      else
+        pred_context = 1 + (edge_mbmi->ref_frame[0] == LAST_FRAME ||
+                            edge_mbmi->ref_frame[1] == LAST_FRAME);
+    }
+  } else {  // no edges available
+    pred_context = 2;
+  }
+
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
+
+int av1_get_pred_context_single_ref_p2(const MACROBLOCKD *xd) {
+  int pred_context;
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+
+  // Note:
+  // The mode info data structure has a one element border above and to the
+  // left of the entries corresponding to real macroblocks.
+  // The prediction flags in these dummy entries are initialized to 0.
+  if (has_above && has_left) {  // both edges available
+    const int above_intra = !is_inter_block(above_mbmi);
+    const int left_intra = !is_inter_block(left_mbmi);
+
+    if (above_intra && left_intra) {  // intra/intra
+      pred_context = 2;
+    } else if (above_intra || left_intra) {  // intra/inter or inter/intra
+      const MB_MODE_INFO *edge_mbmi = above_intra ? left_mbmi : above_mbmi;
+      if (!has_second_ref(edge_mbmi)) {
+        if (edge_mbmi->ref_frame[0] == LAST_FRAME)
+          pred_context = 3;
+        else
+          pred_context = 4 * (edge_mbmi->ref_frame[0] == GOLDEN_FRAME);
+      } else {
+        pred_context = 1 +
+                       2 * (edge_mbmi->ref_frame[0] == GOLDEN_FRAME ||
+                            edge_mbmi->ref_frame[1] == GOLDEN_FRAME);
+      }
+    } else {  // inter/inter
+      const int above_has_second = has_second_ref(above_mbmi);
+      const int left_has_second = has_second_ref(left_mbmi);
+      const MV_REFERENCE_FRAME above0 = above_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME above1 = above_mbmi->ref_frame[1];
+      const MV_REFERENCE_FRAME left0 = left_mbmi->ref_frame[0];
+      const MV_REFERENCE_FRAME left1 = left_mbmi->ref_frame[1];
+
+      if (above_has_second && left_has_second) {
+        if (above0 == left0 && above1 == left1)
+          pred_context =
+              3 * (above0 == GOLDEN_FRAME || above1 == GOLDEN_FRAME ||
+                   left0 == GOLDEN_FRAME || left1 == GOLDEN_FRAME);
+        else
+          pred_context = 2;
+      } else if (above_has_second || left_has_second) {
+        const MV_REFERENCE_FRAME rfs = !above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf1 = above_has_second ? above0 : left0;
+        const MV_REFERENCE_FRAME crf2 = above_has_second ? above1 : left1;
+
+        if (rfs == GOLDEN_FRAME)
+          pred_context = 3 + (crf1 == GOLDEN_FRAME || crf2 == GOLDEN_FRAME);
+        else if (rfs != GOLDEN_FRAME && rfs != LAST_FRAME)
+          pred_context = crf1 == GOLDEN_FRAME || crf2 == GOLDEN_FRAME;
+        else
+          pred_context = 1 + 2 * (crf1 == GOLDEN_FRAME || crf2 == GOLDEN_FRAME);
+      } else {
+        if (above0 == LAST_FRAME && left0 == LAST_FRAME) {
+          pred_context = 3;
+        } else if (above0 == LAST_FRAME || left0 == LAST_FRAME) {
+          const MV_REFERENCE_FRAME edge0 =
+              (above0 == LAST_FRAME) ? left0 : above0;
+          pred_context = 4 * (edge0 == GOLDEN_FRAME);
+        } else {
+          pred_context =
+              2 * (above0 == GOLDEN_FRAME) + 2 * (left0 == GOLDEN_FRAME);
+        }
+      }
+    }
+  } else if (has_above || has_left) {  // one edge available
+    const MB_MODE_INFO *edge_mbmi = has_above ? above_mbmi : left_mbmi;
+
+    if (!is_inter_block(edge_mbmi) ||
+        (edge_mbmi->ref_frame[0] == LAST_FRAME && !has_second_ref(edge_mbmi)))
+      pred_context = 2;
+    else if (!has_second_ref(edge_mbmi))
+      pred_context = 4 * (edge_mbmi->ref_frame[0] == GOLDEN_FRAME);
+    else
+      pred_context = 3 * (edge_mbmi->ref_frame[0] == GOLDEN_FRAME ||
+                          edge_mbmi->ref_frame[1] == GOLDEN_FRAME);
+  } else {  // no edges available (2)
+    pred_context = 2;
+  }
+  assert(pred_context >= 0 && pred_context < REF_CONTEXTS);
+  return pred_context;
+}
+
+#endif  // CONFIG_EXT_REFS
diff --git a/third_party/aom/av1/common/pred_common.h b/third_party/aom/av1/common/pred_common.h
new file mode 100644
index 0000000000..e16ad70f62
--- /dev/null
+++ b/third_party/aom/av1/common/pred_common.h
@@ -0,0 +1,266 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_PRED_COMMON_H_
+#define AV1_COMMON_PRED_COMMON_H_
+
+#include "av1/common/blockd.h"
+#include "av1/common/onyxc_int.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE int get_segment_id(const AV1_COMMON *const cm,
+                                 const uint8_t *segment_ids, BLOCK_SIZE bsize,
+                                 int mi_row, int mi_col) {
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int xmis = AOMMIN(cm->mi_cols - mi_col, bw);
+  const int ymis = AOMMIN(cm->mi_rows - mi_row, bh);
+  int x, y, segment_id = MAX_SEGMENTS;
+
+  for (y = 0; y < ymis; ++y)
+    for (x = 0; x < xmis; ++x)
+      segment_id =
+          AOMMIN(segment_id, segment_ids[mi_offset + y * cm->mi_cols + x]);
+
+  assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
+  return segment_id;
+}
+
+static INLINE int av1_get_pred_context_seg_id(const MACROBLOCKD *xd) {
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int above_sip =
+      (above_mi != NULL) ? above_mi->mbmi.seg_id_predicted : 0;
+  const int left_sip = (left_mi != NULL) ? left_mi->mbmi.seg_id_predicted : 0;
+
+  return above_sip + left_sip;
+}
+
+static INLINE aom_prob av1_get_pred_prob_seg_id(
+    const struct segmentation_probs *segp, const MACROBLOCKD *xd) {
+  return segp->pred_probs[av1_get_pred_context_seg_id(xd)];
+}
+
+static INLINE int av1_get_skip_context(const MACROBLOCKD *xd) {
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const int above_skip = (above_mi != NULL) ? above_mi->mbmi.skip : 0;
+  const int left_skip = (left_mi != NULL) ? left_mi->mbmi.skip : 0;
+  return above_skip + left_skip;
+}
+
+static INLINE aom_prob av1_get_skip_prob(const AV1_COMMON *cm,
+                                         const MACROBLOCKD *xd) {
+  return cm->fc->skip_probs[av1_get_skip_context(xd)];
+}
+
+#if CONFIG_DUAL_FILTER
+int av1_get_pred_context_switchable_interp(const MACROBLOCKD *xd, int dir);
+#else
+int av1_get_pred_context_switchable_interp(const MACROBLOCKD *xd);
+#endif
+
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+int av1_get_pred_context_intra_interp(const MACROBLOCKD *xd);
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+
+int av1_get_intra_inter_context(const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_intra_inter_prob(const AV1_COMMON *cm,
+                                                const MACROBLOCKD *xd) {
+  return cm->fc->intra_inter_prob[av1_get_intra_inter_context(xd)];
+}
+
+int av1_get_reference_mode_context(const AV1_COMMON *cm, const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_reference_mode_prob(const AV1_COMMON *cm,
+                                                   const MACROBLOCKD *xd) {
+  return cm->fc->comp_inter_prob[av1_get_reference_mode_context(cm, xd)];
+}
+
+int av1_get_pred_context_comp_ref_p(const AV1_COMMON *cm,
+                                    const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_pred_prob_comp_ref_p(const AV1_COMMON *cm,
+                                                    const MACROBLOCKD *xd) {
+  const int pred_context = av1_get_pred_context_comp_ref_p(cm, xd);
+  return cm->fc->comp_ref_prob[pred_context][0];
+}
+
+#if CONFIG_EXT_REFS
+int av1_get_pred_context_comp_ref_p1(const AV1_COMMON *cm,
+                                     const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_pred_prob_comp_ref_p1(const AV1_COMMON *cm,
+                                                     const MACROBLOCKD *xd) {
+  const int pred_context = av1_get_pred_context_comp_ref_p1(cm, xd);
+  return cm->fc->comp_ref_prob[pred_context][1];
+}
+
+int av1_get_pred_context_comp_ref_p2(const AV1_COMMON *cm,
+                                     const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_pred_prob_comp_ref_p2(const AV1_COMMON *cm,
+                                                     const MACROBLOCKD *xd) {
+  const int pred_context = av1_get_pred_context_comp_ref_p2(cm, xd);
+  return cm->fc->comp_ref_prob[pred_context][2];
+}
+
+int av1_get_pred_context_comp_bwdref_p(const AV1_COMMON *cm,
+                                       const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_pred_prob_comp_bwdref_p(const AV1_COMMON *cm,
+                                                       const MACROBLOCKD *xd) {
+  const int pred_context = av1_get_pred_context_comp_bwdref_p(cm, xd);
+  return cm->fc->comp_bwdref_prob[pred_context][0];
+}
+#endif  // CONFIG_EXT_REFS
+
+int av1_get_pred_context_single_ref_p1(const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_pred_prob_single_ref_p1(const AV1_COMMON *cm,
+                                                       const MACROBLOCKD *xd) {
+  return cm->fc->single_ref_prob[av1_get_pred_context_single_ref_p1(xd)][0];
+}
+
+int av1_get_pred_context_single_ref_p2(const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_pred_prob_single_ref_p2(const AV1_COMMON *cm,
+                                                       const MACROBLOCKD *xd) {
+  return cm->fc->single_ref_prob[av1_get_pred_context_single_ref_p2(xd)][1];
+}
+
+#if CONFIG_EXT_REFS
+int av1_get_pred_context_single_ref_p3(const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_pred_prob_single_ref_p3(const AV1_COMMON *cm,
+                                                       const MACROBLOCKD *xd) {
+  return cm->fc->single_ref_prob[av1_get_pred_context_single_ref_p3(xd)][2];
+}
+
+int av1_get_pred_context_single_ref_p4(const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_pred_prob_single_ref_p4(const AV1_COMMON *cm,
+                                                       const MACROBLOCKD *xd) {
+  return cm->fc->single_ref_prob[av1_get_pred_context_single_ref_p4(xd)][3];
+}
+
+int av1_get_pred_context_single_ref_p5(const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_pred_prob_single_ref_p5(const AV1_COMMON *cm,
+                                                       const MACROBLOCKD *xd) {
+  return cm->fc->single_ref_prob[av1_get_pred_context_single_ref_p5(xd)][4];
+}
+#endif  // CONFIG_EXT_REFS
+
+#if CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+int av1_get_inter_mode_context(const MACROBLOCKD *xd);
+
+static INLINE aom_prob av1_get_inter_mode_prob(const AV1_COMMON *cm,
+                                               const MACROBLOCKD *xd) {
+  return cm->fc->comp_inter_mode_prob[av1_get_inter_mode_context(xd)];
+}
+#endif  // CONFIG_EXT_INTER && CONFIG_COMPOUND_SINGLEREF
+
+// Returns a context number for the given MB prediction signal
+// The mode info data structure has a one element border above and to the
+// left of the entries corresponding to real blocks.
+// The prediction flags in these dummy entries are initialized to 0.
+static INLINE int get_tx_size_context(const MACROBLOCKD *xd) {
+  const int max_tx_size = max_txsize_lookup[xd->mi[0]->mbmi.sb_type];
+  const MB_MODE_INFO *const above_mbmi = xd->above_mbmi;
+  const MB_MODE_INFO *const left_mbmi = xd->left_mbmi;
+  const int has_above = xd->up_available;
+  const int has_left = xd->left_available;
+  int above_ctx = (has_above && !above_mbmi->skip)
+                      ? (int)txsize_sqr_map[above_mbmi->tx_size]
+                      : max_tx_size;
+  int left_ctx = (has_left && !left_mbmi->skip)
+                     ? (int)txsize_sqr_map[left_mbmi->tx_size]
+                     : max_tx_size;
+
+  if (!has_left) left_ctx = above_ctx;
+
+  if (!has_above) above_ctx = left_ctx;
+#if CONFIG_CB4X4
+  // TODO(jingning): Temporary setup. Will rework this after the cb4x4
+  // framework is up running.
+  return (above_ctx + left_ctx) > max_tx_size + 1;
+#else
+  return (above_ctx + left_ctx) > max_tx_size;
+#endif
+}
+
+#if CONFIG_VAR_TX
+static void update_tx_counts(AV1_COMMON *cm, MACROBLOCKD *xd,
+                             MB_MODE_INFO *mbmi, BLOCK_SIZE plane_bsize,
+                             TX_SIZE tx_size, int blk_row, int blk_col,
+                             TX_SIZE max_tx_size, int ctx) {
+  const struct macroblockd_plane *const pd = &xd->plane[0];
+  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
+  const int tx_row = blk_row >> (1 - pd->subsampling_y);
+  const int tx_col = blk_col >> (1 - pd->subsampling_x);
+  const TX_SIZE plane_tx_size = mbmi->inter_tx_size[tx_row][tx_col];
+  const int max_blocks_high = max_block_high(xd, plane_bsize, 0);
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, 0);
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  if (tx_size == plane_tx_size) {
+    ++xd->counts->tx_size[max_tx_size - TX_8X8][ctx][tx_size];
+    mbmi->tx_size = tx_size;
+  } else {
+    int bsl = b_width_log2_lookup[bsize];
+    int i;
+
+    assert(bsl > 0);
+    --bsl;
+
+    for (i = 0; i < 4; ++i) {
+      const int offsetr = blk_row + ((i >> 1) << bsl);
+      const int offsetc = blk_col + ((i & 0x01) << bsl);
+
+      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
+      update_tx_counts(cm, xd, mbmi, plane_bsize, (TX_SIZE)(tx_size - 1),
+                       offsetr, offsetc, max_tx_size, ctx);
+    }
+  }
+}
+
+static INLINE void inter_block_tx_count_update(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                               MB_MODE_INFO *mbmi,
+                                               BLOCK_SIZE plane_bsize,
+                                               int ctx) {
+  const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+  const int mi_height = block_size_high[plane_bsize] >> tx_size_wide_log2[0];
+  TX_SIZE max_tx_size = max_txsize_lookup[plane_bsize];
+  int bh = tx_size_wide_unit[max_tx_size];
+  int idx, idy;
+
+  for (idy = 0; idy < mi_height; idy += bh)
+    for (idx = 0; idx < mi_width; idx += bh)
+      update_tx_counts(cm, xd, mbmi, plane_bsize, max_tx_size, idy, idx,
+                       max_tx_size, ctx);
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_PRED_COMMON_H_
diff --git a/third_party/aom/av1/common/pvq.c b/third_party/aom/av1/common/pvq.c
new file mode 100644
index 0000000000..75fe761d74
--- /dev/null
+++ b/third_party/aom/av1/common/pvq.c
@@ -0,0 +1,1007 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include "odintrin.h"
+#include "partition.h"
+#include "pvq.h"
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+/* Imported from encode.c in daala */
+/* These are the PVQ equivalent of quantization matrices, except that
+   the values are per-band. */
+#define OD_MASKING_DISABLED 0
+#define OD_MASKING_ENABLED 1
+
+const unsigned char OD_LUMA_QM_Q4[2][OD_QM_SIZE] = {
+/* Flat quantization for PSNR. The DC component isn't 16 because the DC
+   magnitude compensation is done here for inter (Haar DC doesn't need it).
+   Masking disabled: */
+ {
+  16, 16,
+  16, 16, 16, 16,
+  16, 16, 16, 16, 16, 16,
+  16, 16, 16, 16, 16, 16, 16, 16
+ },
+/* The non-flat AC coefficients compensate for the non-linear scaling caused
+   by activity masking. The values are currently hand-tuned so that the rate
+   of each band remains roughly constant when enabling activity masking
+   on intra.
+   Masking enabled: */
+ {
+  16, 16,
+  16, 18, 28, 32,
+  16, 14, 20, 20, 28, 32,
+  16, 11, 14, 14, 17, 17, 22, 28
+ }
+};
+
+const unsigned char OD_CHROMA_QM_Q4[2][OD_QM_SIZE] = {
+/* Chroma quantization is different because of the reduced lapping.
+   FIXME: Use the same matrix as luma for 4:4:4.
+   Masking disabled: */
+ {
+  16, 16,
+  16, 16, 16, 16,
+  16, 16, 16, 16, 16, 16,
+  16, 16, 16, 16, 16, 16, 16, 16
+ },
+/* The AC part is flat for chroma because it has no activity masking.
+   Masking enabled: */
+ {
+  16, 16,
+  16, 16, 16, 16,
+  16, 16, 16, 16, 16, 16,
+  16, 16, 16, 16, 16, 16, 16, 16
+ }
+};
+
+/* No interpolation, always use od_flat_qm_q4, but use a different scale for
+   each plane.
+   FIXME: Add interpolation and properly tune chroma. */
+const od_qm_entry OD_DEFAULT_QMS[2][2][OD_NPLANES_MAX] = {
+  /* Masking disabled */
+  { { { 4, 256, OD_LUMA_QM_Q4[OD_MASKING_DISABLED] },
+      { 4, 256, OD_CHROMA_QM_Q4[OD_MASKING_DISABLED] },
+      { 4, 256, OD_CHROMA_QM_Q4[OD_MASKING_DISABLED] } },
+    { { 0, 0, NULL},
+      { 0, 0, NULL},
+      { 0, 0, NULL} } },
+  /* Masking enabled */
+  { { { 4, 256, OD_LUMA_QM_Q4[OD_MASKING_ENABLED] },
+      { 4, 256, OD_CHROMA_QM_Q4[OD_MASKING_ENABLED] },
+      { 4, 256, OD_CHROMA_QM_Q4[OD_MASKING_ENABLED] } },
+    { { 0, 0, NULL},
+      { 0, 0, NULL},
+      { 0, 0, NULL} } }
+};
+
+/* Constants for the beta parameter, which controls how activity masking is
+   used.
+   beta = 1 / (1 - alpha), so when beta is 1, alpha is 0 and activity
+   masking is disabled. When beta is 1.5, activity masking is used. Note that
+   activity masking is neither used for 4x4 blocks nor for chroma. */
+#define OD_BETA(b) OD_QCONST32(b, OD_BETA_SHIFT)
+static const od_val16 OD_PVQ_BETA4_LUMA[1] = {OD_BETA(1.)};
+static const od_val16 OD_PVQ_BETA8_LUMA[4] = {OD_BETA(1.), OD_BETA(1.),
+ OD_BETA(1.), OD_BETA(1.)};
+static const od_val16 OD_PVQ_BETA16_LUMA[7] = {OD_BETA(1.), OD_BETA(1.),
+ OD_BETA(1.), OD_BETA(1.), OD_BETA(1.), OD_BETA(1.), OD_BETA(1.)};
+static const od_val16 OD_PVQ_BETA32_LUMA[10] = {OD_BETA(1.), OD_BETA(1.),
+ OD_BETA(1.), OD_BETA(1.), OD_BETA(1.), OD_BETA(1.), OD_BETA(1.), OD_BETA(1.),
+ OD_BETA(1.), OD_BETA(1.)};
+
+static const od_val16 OD_PVQ_BETA4_LUMA_MASKING[1] = {OD_BETA(1.)};
+static const od_val16 OD_PVQ_BETA8_LUMA_MASKING[4] = {OD_BETA(1.5),
+ OD_BETA(1.5), OD_BETA(1.5), OD_BETA(1.5)};
+static const od_val16 OD_PVQ_BETA16_LUMA_MASKING[7] = {OD_BETA(1.5),
+ OD_BETA(1.5), OD_BETA(1.5), OD_BETA(1.5), OD_BETA(1.5), OD_BETA(1.5),
+ OD_BETA(1.5)};
+static const od_val16 OD_PVQ_BETA32_LUMA_MASKING[10] = {OD_BETA(1.5),
+ OD_BETA(1.5), OD_BETA(1.5), OD_BETA(1.5), OD_BETA(1.5), OD_BETA(1.5),
+ OD_BETA(1.5), OD_BETA(1.5), OD_BETA(1.5), OD_BETA(1.5)};
+
+static const od_val16 OD_PVQ_BETA4_CHROMA[1] = {OD_BETA(1.)};
+static const od_val16 OD_PVQ_BETA8_CHROMA[4] = {OD_BETA(1.), OD_BETA(1.),
+ OD_BETA(1.), OD_BETA(1.)};
+static const od_val16 OD_PVQ_BETA16_CHROMA[7] = {OD_BETA(1.), OD_BETA(1.),
+ OD_BETA(1.), OD_BETA(1.), OD_BETA(1.), OD_BETA(1.), OD_BETA(1.)};
+static const od_val16 OD_PVQ_BETA32_CHROMA[10] = {OD_BETA(1.), OD_BETA(1.),
+ OD_BETA(1.), OD_BETA(1.), OD_BETA(1.), OD_BETA(1.), OD_BETA(1.), OD_BETA(1.),
+ OD_BETA(1.), OD_BETA(1.)};
+
+const od_val16 *const OD_PVQ_BETA[2][OD_NPLANES_MAX][OD_TXSIZES + 1] = {
+ {{OD_PVQ_BETA4_LUMA, OD_PVQ_BETA8_LUMA,
+   OD_PVQ_BETA16_LUMA, OD_PVQ_BETA32_LUMA},
+  {OD_PVQ_BETA4_CHROMA, OD_PVQ_BETA8_CHROMA,
+   OD_PVQ_BETA16_CHROMA, OD_PVQ_BETA32_CHROMA},
+  {OD_PVQ_BETA4_CHROMA, OD_PVQ_BETA8_CHROMA,
+   OD_PVQ_BETA16_CHROMA, OD_PVQ_BETA32_CHROMA}},
+ {{OD_PVQ_BETA4_LUMA_MASKING, OD_PVQ_BETA8_LUMA_MASKING,
+   OD_PVQ_BETA16_LUMA_MASKING, OD_PVQ_BETA32_LUMA_MASKING},
+  {OD_PVQ_BETA4_CHROMA, OD_PVQ_BETA8_CHROMA,
+   OD_PVQ_BETA16_CHROMA, OD_PVQ_BETA32_CHROMA},
+  {OD_PVQ_BETA4_CHROMA, OD_PVQ_BETA8_CHROMA,
+   OD_PVQ_BETA16_CHROMA, OD_PVQ_BETA32_CHROMA}}
+};
+
+
+void od_interp_qm(unsigned char *out, int q, const od_qm_entry *entry1,
+  const od_qm_entry *entry2) {
+  int i;
+  if (entry2 == NULL || entry2->qm_q4 == NULL
+   || q < entry1->interp_q << OD_COEFF_SHIFT) {
+    /* Use entry1. */
+    for (i = 0; i < OD_QM_SIZE; i++) {
+      out[i] = OD_MINI(255, entry1->qm_q4[i]*entry1->scale_q8 >> 8);
+    }
+  }
+  else if (entry1 == NULL || entry1->qm_q4 == NULL
+   || q > entry2->interp_q << OD_COEFF_SHIFT) {
+    /* Use entry2. */
+    for (i = 0; i < OD_QM_SIZE; i++) {
+      out[i] = OD_MINI(255, entry2->qm_q4[i]*entry2->scale_q8 >> 8);
+    }
+  }
+  else {
+    /* Interpolate between entry1 and entry2. The interpolation is linear
+       in terms of log(q) vs log(m*scale). Considering that we're ultimately
+       multiplying the result it makes sense, but we haven't tried other
+       interpolation methods. */
+    double x;
+    const unsigned char *m1;
+    const unsigned char *m2;
+    int q1;
+    int q2;
+    m1 = entry1->qm_q4;
+    m2 = entry2->qm_q4;
+    q1 = entry1->interp_q << OD_COEFF_SHIFT;
+    q2 = entry2->interp_q << OD_COEFF_SHIFT;
+    x = (log(q)-log(q1))/(log(q2)-log(q1));
+    for (i = 0; i < OD_QM_SIZE; i++) {
+      out[i] = OD_MINI(255, (int)floor(.5 + (1./256)*exp(
+       x*log(m2[i]*entry2->scale_q8) + (1 - x)*log(m1[i]*entry1->scale_q8))));
+    }
+  }
+}
+
+void od_adapt_pvq_ctx_reset(od_pvq_adapt_ctx *state, int is_keyframe) {
+  od_pvq_codeword_ctx *ctx;
+  int i;
+  int pli;
+  int bs;
+  ctx = &state->pvq_codeword_ctx;
+  OD_CDFS_INIT_DYNAMIC(state->pvq_param_model[0].cdf);
+  OD_CDFS_INIT_DYNAMIC(state->pvq_param_model[1].cdf);
+  OD_CDFS_INIT_DYNAMIC(state->pvq_param_model[2].cdf);
+  for (i = 0; i < 2*OD_TXSIZES; i++) {
+    ctx->pvq_adapt[4*i + OD_ADAPT_K_Q8] = 384;
+    ctx->pvq_adapt[4*i + OD_ADAPT_SUM_EX_Q8] = 256;
+    ctx->pvq_adapt[4*i + OD_ADAPT_COUNT_Q8] = 104;
+    ctx->pvq_adapt[4*i + OD_ADAPT_COUNT_EX_Q8] = 128;
+  }
+  OD_CDFS_INIT_DYNAMIC(ctx->pvq_k1_cdf);
+  for (pli = 0; pli < OD_NPLANES_MAX; pli++) {
+    for (bs = 0; bs < OD_TXSIZES; bs++)
+    for (i = 0; i < PVQ_MAX_PARTITIONS; i++) {
+      state->pvq_exg[pli][bs][i] = 2 << 16;
+    }
+  }
+  for (i = 0; i < OD_TXSIZES*PVQ_MAX_PARTITIONS; i++) {
+    state->pvq_ext[i] = is_keyframe ? 24576 : 2 << 16;
+  }
+  OD_CDFS_INIT_DYNAMIC(state->pvq_gaintheta_cdf);
+  OD_CDFS_INIT_Q15(state->pvq_skip_dir_cdf);
+  OD_CDFS_INIT_DYNAMIC(ctx->pvq_split_cdf);
+}
+
+/* QMs are arranged from smallest to largest blocksizes, first for
+   blocks with decimation=0, followed by blocks with decimation=1.*/
+int od_qm_offset(int bs, int xydec)
+{
+    return xydec*OD_QM_STRIDE + OD_QM_OFFSET(bs);
+}
+
+#if defined(OD_FLOAT_PVQ)
+#define OD_DEFAULT_MAG 1.0
+#else
+#define OD_DEFAULT_MAG OD_QM_SCALE
+#endif
+
+/* Initialize the quantization matrix. */
+// Note: When hybrid transform and corresponding scan order is used by PVQ,
+// we don't need seperate qm and qm_inv for each transform type,
+// because AOM does not do magnitude compensation (i.e. simplay x16 for all coeffs).
+void od_init_qm(int16_t *x, int16_t *x_inv, const int *qm) {
+  int i;
+  int j;
+  int16_t y[OD_TXSIZE_MAX*OD_TXSIZE_MAX];
+  int16_t y_inv[OD_TXSIZE_MAX*OD_TXSIZE_MAX];
+  int16_t *x1;
+  int16_t *x1_inv;
+  int off;
+  int bs;
+  int xydec;
+  for (bs = 0; bs < OD_TXSIZES; bs++) {
+    for (xydec = 0; xydec < 2; xydec++) {
+      off = od_qm_offset(bs, xydec);
+      x1 = x + off;
+      x1_inv = x_inv + off;
+      for (i = 0; i < 4 << bs; i++) {
+        for (j = 0; j < 4 << bs; j++) {
+          /*This will ultimately be clamped to fit in 16 bits.*/
+          od_val32 mag;
+          int16_t ytmp;
+          mag = OD_DEFAULT_MAG;
+          if (i != 0 || j != 0) {
+#if defined(OD_FLOAT_PVQ)
+            mag /= 0.0625*qm[(i << 1 >> bs)*8 + (j << 1 >> bs)];
+#else
+            int qmv;
+            qmv = qm[(i << 1 >> bs)*8 + (j << 1 >> bs)];
+            mag *= 16;
+            mag = (mag + (qmv >> 1))/qmv;
+#endif
+            OD_ASSERT(mag > 0.0);
+          }
+          /*Convert to fit in 16 bits.*/
+#if defined(OD_FLOAT_PVQ)
+          y[i*(4 << bs) + j] = (int16_t)OD_MINI(OD_QM_SCALE_MAX,
+           (int32_t)floor(.5 + mag*OD_QM_SCALE));
+          y_inv[i*(4 << bs) + j] = (int16_t)floor(.5
+           + OD_QM_SCALE*OD_QM_INV_SCALE/(double)y[i*(4 << bs) + j]);
+#else
+          y[i*(4 << bs) + j] = (int16_t)OD_MINI(OD_QM_SCALE_MAX, mag);
+          ytmp = y[i*(4 << bs) + j];
+          y_inv[i*(4 << bs) + j] = (int16_t)((OD_QM_SCALE*OD_QM_INV_SCALE
+           + (ytmp >> 1))/ytmp);
+#endif
+        }
+      }
+      od_raster_to_coding_order_16(x1, 4 << bs, y, 4 << bs);
+      od_raster_to_coding_order_16(x1_inv, 4 << bs, y_inv, 4 << bs);
+    }
+  }
+}
+
+/* Maps each possible size (n) in the split k-tokenizer to a different value.
+   Possible values of n are:
+   2, 3, 4, 7, 8, 14, 15, 16, 31, 32, 63, 64, 127, 128
+   Since we don't care about the order (even in the bit-stream) the simplest
+   ordering (implemented here) is:
+   14, 2, 3, 4, 7, 8, 15, 16, 31, 32, 63, 64, 127, 128 */
+int od_pvq_size_ctx(int n) {
+  int logn;
+  int odd;
+  logn = OD_ILOG(n - 1);
+  odd = n & 1;
+  return 2*logn - 1 - odd - 7*(n == 14);
+}
+
+/* Maps a length n to a context for the (k=1, n<=16) coder, with a special
+   case when n is the original length (orig_length=1) of the vector (i.e. we
+   haven't split it yet). For orig_length=0, we use the same mapping as
+   od_pvq_size_ctx() up to n=16. When orig_length=1, we map lengths
+   7, 8, 14, 15 to contexts 8 to 11. */
+int od_pvq_k1_ctx(int n, int orig_length) {
+  if (orig_length) return 8 + 2*(n > 8) + (n & 1);
+  else return od_pvq_size_ctx(n);
+}
+
+/* Indexing for the packed quantization matrices. */
+int od_qm_get_index(int bs, int band) {
+  /* The -band/3 term is due to the fact that we force corresponding horizontal
+     and vertical bands to have the same quantization. */
+  OD_ASSERT(bs >= 0 && bs < OD_TXSIZES);
+  return bs*(bs + 1) + band - band/3;
+}
+
+#if !defined(OD_FLOAT_PVQ)
+/*See celt/mathops.c in Opus and tools/cos_search.c.*/
+static int16_t od_pvq_cos_pi_2(int16_t x)
+{
+  int16_t x2;
+  x2 = OD_MULT16_16_Q15(x, x);
+  return OD_MINI(32767, (1073758164 - x*x + x2*(-7654 + OD_MULT16_16_Q16(x2,
+   16573 + OD_MULT16_16_Q16(-2529, x2)))) >> 15);
+}
+#endif
+
+/*Approximates cos(x) for -pi < x < pi.
+  Input is in OD_THETA_SCALE.*/
+od_val16 od_pvq_cos(od_val32 x) {
+#if defined(OD_FLOAT_PVQ)
+  return cos(x);
+#else
+  /*Wrap x around by masking, since cos is periodic.*/
+  x = x & 0x0001ffff;
+  if (x > (1 << 16)) {
+    x = (1 << 17) - x;
+  }
+  if (x & 0x00007fff) {
+    if (x < (1 << 15)) {
+       return od_pvq_cos_pi_2((int16_t)x);
+    }
+    else {
+      return -od_pvq_cos_pi_2((int16_t)(65536 - x));
+    }
+  }
+  else {
+    if (x & 0x0000ffff) {
+      return 0;
+    }
+    else if (x & 0x0001ffff) {
+      return -32767;
+    }
+    else {
+      return 32767;
+    }
+  }
+#endif
+}
+
+/*Approximates sin(x) for 0 <= x < pi.
+  Input is in OD_THETA_SCALE.*/
+od_val16 od_pvq_sin(od_val32 x) {
+#if defined(OD_FLOAT_PVQ)
+  return sin(x);
+#else
+  return od_pvq_cos(32768 - x);
+#endif
+}
+
+#if !defined(OD_FLOAT_PVQ)
+/* Computes an upper-bound on the number of bits required to store the L2 norm
+   of a vector (excluding sign). */
+int od_vector_log_mag(const od_coeff *x, int n) {
+  int i;
+  int32_t sum;
+  sum = 0;
+  for (i = 0; i < n; i++) {
+    int16_t tmp;
+    tmp = x[i] >> 8;
+    sum += tmp*(int32_t)tmp;
+  }
+  /* We add one full bit (instead of rounding OD_ILOG() up) for safety because
+     the >> 8 above causes the sum to be slightly underestimated. */
+  return 8 + 1 + OD_ILOG(n + sum)/2;
+}
+#endif
+
+/** Computes Householder reflection that aligns the reference r to the
+ *  dimension in r with the greatest absolute value. The reflection
+ *  vector is returned in r.
+ *
+ * @param [in,out]  r      reference vector to be reflected, reflection
+ *                         also returned in r
+ * @param [in]      n      number of dimensions in r
+ * @param [in]      gr     gain of reference vector
+ * @param [out]     sign   sign of reflection
+ * @return                 dimension number to which reflection aligns
+ **/
+int od_compute_householder(od_val16 *r, int n, od_val32 gr, int *sign,
+ int shift) {
+  int m;
+  int i;
+  int s;
+  od_val16 maxr;
+  OD_UNUSED(shift);
+  /* Pick component with largest magnitude. Not strictly
+   * necessary, but it helps numerical stability */
+  m = 0;
+  maxr = 0;
+  for (i = 0; i < n; i++) {
+    if (OD_ABS(r[i]) > maxr) {
+      maxr = OD_ABS(r[i]);
+      m = i;
+    }
+  }
+  s = r[m] > 0 ? 1 : -1;
+  /* This turns r into a Householder reflection vector that would reflect
+   * the original r[] to e_m */
+  r[m] += OD_SHR_ROUND(gr*s, shift);
+  *sign = s;
+  return m;
+}
+
+#if !defined(OD_FLOAT_PVQ)
+#define OD_RCP_INSHIFT 15
+#define OD_RCP_OUTSHIFT 14
+static od_val16 od_rcp(od_val16 x)
+{
+  int i;
+  od_val16 n;
+  od_val16 r;
+  i = OD_ILOG(x) - 1;
+  /*n is Q15 with range [0,1).*/
+  n = OD_VSHR_ROUND(x, i - OD_RCP_INSHIFT) - (1 << OD_RCP_INSHIFT);
+  /*Start with a linear approximation:
+    r = 1.8823529411764706-0.9411764705882353*n.
+    The coefficients and the result are Q14 in the range [15420,30840].*/
+  r = 30840 + OD_MULT16_16_Q15(-15420, n);
+  /*Perform two Newton iterations:
+    r -= r*((r*n)-1.Q15)
+       = r*((r*n)+(r-1.Q15)).*/
+  r = r - OD_MULT16_16_Q15(r, (OD_MULT16_16_Q15(r, n) + r - 32768));
+  /*We subtract an extra 1 in the second iteration to avoid overflow; it also
+     neatly compensates for truncation error in the rest of the process.*/
+  r = r - (1 + OD_MULT16_16_Q15(r, OD_MULT16_16_Q15(r, n) + r - 32768));
+  /*r is now the Q15 solution to 2/(n+1), with a maximum relative error
+     of 7.05346E-5, a (relative) RMSE of 2.14418E-5, and a peak absolute
+     error of 1.24665/32768.*/
+  return OD_VSHR_ROUND(r, i - OD_RCP_OUTSHIFT);
+}
+#endif
+
+/** Applies Householder reflection from compute_householder(). The
+ * reflection is its own inverse.
+ *
+ * @param [out]     out    reflected vector
+ * @param [in]      x      vector to be reflected
+ * @param [in]      r      reflection
+ * @param [in]      n      number of dimensions in x,r
+ */
+void od_apply_householder(od_val16 *out, const od_val16 *x, const od_val16 *r,
+ int n) {
+  int i;
+  od_val32 proj;
+  od_val16 proj_1;
+  od_val32 l2r;
+#if !defined(OD_FLOAT_PVQ)
+  od_val16 proj_norm;
+  od_val16 l2r_norm;
+  od_val16 rcp;
+  int proj_shift;
+  int l2r_shift;
+  int outshift;
+#endif
+  /*FIXME: Can we get l2r and/or l2r_shift from an earlier computation?*/
+  l2r = 0;
+  for (i = 0; i < n; i++) {
+    l2r += OD_MULT16_16(r[i], r[i]);
+  }
+  /* Apply Householder reflection */
+  proj = 0;
+  for (i = 0; i < n; i++) {
+    proj += OD_MULT16_16(r[i], x[i]);
+  }
+#if defined(OD_FLOAT_PVQ)
+  proj_1 = proj*2./(1e-100 + l2r);
+  for (i = 0; i < n; i++) {
+    out[i] = x[i] - r[i]*proj_1;
+  }
+#else
+  /*l2r_norm is [0.5, 1.0[ in Q15.*/
+  l2r_shift = (OD_ILOG(l2r) - 1) - 14;
+  l2r_norm = OD_VSHR_ROUND(l2r, l2r_shift);
+  rcp = od_rcp(l2r_norm);
+  proj_shift = (OD_ILOG(abs(proj)) - 1) - 14;
+  /*proj_norm is [0.5, 1.0[ in Q15.*/
+  proj_norm = OD_VSHR_ROUND(proj, proj_shift);
+  proj_1 = OD_MULT16_16_Q15(proj_norm, rcp);
+  /*The proj*2. in the float code becomes -1 in the final outshift.
+    The sign of l2r_shift is positive since we're taking the reciprocal of
+     l2r_norm and this is a right shift.*/
+  outshift = OD_MINI(30, OD_RCP_OUTSHIFT - proj_shift - 1 + l2r_shift);
+  if (outshift >= 0) {
+    for (i = 0; i < n; i++) {
+      int32_t tmp;
+      tmp = OD_MULT16_16(r[i], proj_1);
+      tmp = OD_SHR_ROUND(tmp, outshift);
+      out[i] = x[i] - tmp;
+    }
+  }
+  else {
+    /*FIXME: Can we make this case impossible?
+      Right now, if r[] is all zeros except for 1, 2, or 3 ones, and
+       if x[] is all zeros except for large values at the same position as the
+       ones in r[], then we can end up with a shift of -1.*/
+    for (i = 0; i < n; i++) {
+      int32_t tmp;
+      tmp = OD_MULT16_16(r[i], proj_1);
+      tmp = OD_SHL(tmp, -outshift);
+      out[i] = x[i] - tmp;
+    }
+  }
+#endif
+}
+
+#if !defined(OD_FLOAT_PVQ)
+static od_val16 od_beta_rcp(od_val16 beta){
+  if (beta == OD_BETA(1.))
+    return OD_BETA(1.);
+  else if (beta == OD_BETA(1.5))
+    return OD_BETA(1./1.5);
+  else {
+    od_val16 rcp_beta;
+    /*Shift by 1 less, transposing beta to range [.5, .75] and thus < 32768.*/
+    rcp_beta = od_rcp(beta << (OD_RCP_INSHIFT - 1 - OD_BETA_SHIFT));
+    return OD_SHR_ROUND(rcp_beta, OD_RCP_OUTSHIFT + 1 - OD_BETA_SHIFT);
+  }
+}
+
+#define OD_EXP2_INSHIFT 15
+#define OD_EXP2_FRACSHIFT 15
+#define OD_EXP2_OUTSHIFT 15
+static const int32_t OD_EXP2_C[5] = {32768, 22709, 7913, 1704, 443};
+/*Output is [1.0, 2.0) in Q(OD_EXP2_FRACSHIFT).
+  It does not include the integer offset, which is added in od_exp2 after the
+   final shift).*/
+static int32_t od_exp2_frac(int32_t x)
+{
+  return OD_MULT16_16_Q15(x, (OD_EXP2_C[1] + OD_MULT16_16_Q15(x,
+   (OD_EXP2_C[2] + OD_MULT16_16_Q15(x, (OD_EXP2_C[3]
+   + OD_MULT16_16_Q15(x, OD_EXP2_C[4])))))));
+}
+
+/** Base-2 exponential approximation (2^x) with Q15 input and output.*/
+static int32_t od_exp2(int32_t x)
+{
+  int integer;
+  int32_t frac;
+  integer = x >> OD_EXP2_INSHIFT;
+  if (integer > 14)
+    return 0x7f000000;
+  else if (integer < -15)
+    return 0;
+  frac = od_exp2_frac(x - OD_SHL(integer, OD_EXP2_INSHIFT));
+  return OD_VSHR_ROUND(OD_EXP2_C[0] + frac, -integer) + 1;
+}
+
+#define OD_LOG2_INSHIFT 15
+#define OD_LOG2_OUTSHIFT 15
+#define OD_LOG2_INSCALE_1 (1./(1 << OD_LOG2_INSHIFT))
+#define OD_LOG2_OUTSCALE (1 << OD_LOG2_OUTSHIFT)
+static int16_t od_log2(int16_t x)
+{
+  return x + OD_MULT16_16_Q15(x, (14482 + OD_MULT16_16_Q15(x, (-23234
+   + OD_MULT16_16_Q15(x, (13643 + OD_MULT16_16_Q15(x, (-6403
+   + OD_MULT16_16_Q15(x, 1515)))))))));
+}
+
+static int32_t od_pow(int32_t x, od_val16 beta)
+{
+  int16_t t;
+  int xshift;
+  int log2_x;
+  od_val32 logr;
+  /*FIXME: this conditional is to avoid doing log2(0).*/
+  if (x == 0)
+    return 0;
+  log2_x = (OD_ILOG(x) - 1);
+  xshift = log2_x - OD_LOG2_INSHIFT;
+  /*t should be in range [0.0, 1.0[ in Q(OD_LOG2_INSHIFT).*/
+  t = OD_VSHR(x, xshift) - (1 << OD_LOG2_INSHIFT);
+  /*log2(g/OD_COMPAND_SCALE) = log2(x) - OD_COMPAND_SHIFT in
+     Q(OD_LOG2_OUTSHIFT).*/
+  logr = od_log2(t) + (log2_x - OD_COMPAND_SHIFT)*OD_LOG2_OUTSCALE;
+  logr = OD_MULT16_32_QBETA(beta, logr);
+  return od_exp2(logr);
+}
+#endif
+
+/** Gain companding: raises gain to the power 1/beta for activity masking.
+ *
+ * @param [in]  g     real (uncompanded) gain
+ * @param [in]  q0    uncompanded quality parameter
+ * @param [in]  beta  activity masking beta param (exponent)
+ * @return            g^(1/beta)
+ */
+static od_val32 od_gain_compand(od_val32 g, int q0, od_val16 beta) {
+#if defined(OD_FLOAT_PVQ)
+  if (beta == 1) return OD_CGAIN_SCALE*g/(double)q0;
+  else {
+    return OD_CGAIN_SCALE*OD_COMPAND_SCALE*pow(g*OD_COMPAND_SCALE_1,
+     1./beta)/(double)q0;
+  }
+#else
+  if (beta == OD_BETA(1)) return (OD_CGAIN_SCALE*g + (q0 >> 1))/q0;
+  else {
+    int32_t expr;
+    expr = od_pow(g, od_beta_rcp(beta));
+    expr <<= OD_CGAIN_SHIFT + OD_COMPAND_SHIFT - OD_EXP2_OUTSHIFT;
+    return (expr + (q0 >> 1))/q0;
+  }
+#endif
+}
+
+#if !defined(OD_FLOAT_PVQ)
+#define OD_SQRT_INSHIFT 16
+#define OD_SQRT_OUTSHIFT 15
+static int16_t od_rsqrt_norm(int16_t x);
+
+static int16_t od_sqrt_norm(int32_t x)
+{
+  OD_ASSERT(x < 65536);
+  return OD_MINI(OD_SHR_ROUND(x*od_rsqrt_norm(x), OD_SQRT_OUTSHIFT), 32767);
+}
+
+static int16_t od_sqrt(int32_t x, int *sqrt_shift)
+{
+  int k;
+  int s;
+  int32_t t;
+  if (x == 0) {
+    *sqrt_shift = 0;
+     return 0;
+  }
+  OD_ASSERT(x < (1 << 30));
+  k = ((OD_ILOG(x) - 1) >> 1);
+  /*t is x in the range [0.25, 1) in QINSHIFT, or x*2^(-s).
+    Shift by log2(x) - log2(0.25*(1 << INSHIFT)) to ensure 0.25 lower bound.*/
+  s = 2*k - (OD_SQRT_INSHIFT - 2);
+  t = OD_VSHR(x, s);
+  /*We want to express od_sqrt() in terms of od_sqrt_norm(), which is
+     defined as (2^OUTSHIFT)*sqrt(t*(2^-INSHIFT)) with t=x*(2^-s).
+    This simplifies to 2^(OUTSHIFT-(INSHIFT/2)-(s/2))*sqrt(x), so the caller
+     needs to shift right by OUTSHIFT - INSHIFT/2 - s/2.*/
+  *sqrt_shift = OD_SQRT_OUTSHIFT - ((s + OD_SQRT_INSHIFT) >> 1);
+  return od_sqrt_norm(t);
+}
+#endif
+
+/** Gain expanding: raises gain to the power beta for activity masking.
+ *
+ * @param [in]  cg    companded gain
+ * @param [in]  q0    uncompanded quality parameter
+ * @param [in]  beta  activity masking beta param (exponent)
+ * @return            g^beta
+ */
+od_val32 od_gain_expand(od_val32 cg0, int q0, od_val16 beta) {
+  if (beta == OD_BETA(1)) {
+    /*The multiply fits into 28 bits because the expanded gain has a range from
+       0 to 2^20.*/
+    return OD_SHR_ROUND(cg0*q0, OD_CGAIN_SHIFT);
+  }
+  else if (beta == OD_BETA(1.5)) {
+#if defined(OD_FLOAT_PVQ)
+    double cg;
+    cg = cg0*OD_CGAIN_SCALE_1;
+    cg *= q0*OD_COMPAND_SCALE_1;
+    return OD_COMPAND_SCALE*cg*sqrt(cg);
+#else
+    int32_t irt;
+    int64_t tmp;
+    int sqrt_inshift;
+    int sqrt_outshift;
+    /*cg0 is in Q(OD_CGAIN_SHIFT) and we need to divide it by
+       2^OD_COMPAND_SHIFT.*/
+    irt = od_sqrt(cg0*q0, &sqrt_outshift);
+    sqrt_inshift = (OD_CGAIN_SHIFT + OD_COMPAND_SHIFT) >> 1;
+    /*tmp is in Q(OD_CGAIN_SHIFT + OD_COMPAND_SHIFT).*/
+    tmp = cg0*q0*(int64_t)irt;
+    /*Expanded gain must be in Q(OD_COMPAND_SHIFT), thus OD_COMPAND_SHIFT is
+       not included here.*/
+    return OD_MAXI(1,
+        OD_VSHR_ROUND(tmp, OD_CGAIN_SHIFT + sqrt_outshift + sqrt_inshift));
+#endif
+  }
+  else {
+#if defined(OD_FLOAT_PVQ)
+    /*Expanded gain must be in Q(OD_COMPAND_SHIFT), hence the multiply by
+       OD_COMPAND_SCALE.*/
+    double cg;
+    cg = cg0*OD_CGAIN_SCALE_1;
+    return OD_COMPAND_SCALE*pow(cg*q0*OD_COMPAND_SCALE_1, beta);
+#else
+    int32_t expr;
+    int32_t cg;
+    cg = OD_SHR_ROUND(cg0*q0, OD_CGAIN_SHIFT);
+    expr = od_pow(cg, beta);
+    /*Expanded gain must be in Q(OD_COMPAND_SHIFT), hence the subtraction by
+       OD_COMPAND_SHIFT.*/
+    return OD_MAXI(1, OD_SHR_ROUND(expr, OD_EXP2_OUTSHIFT - OD_COMPAND_SHIFT));
+#endif
+  }
+}
+
+/** Computes the raw and quantized/companded gain of a given input
+ * vector
+ *
+ * @param [in]      x      vector of input data
+ * @param [in]      n      number of elements in vector x
+ * @param [in]      q0     quantizer
+ * @param [out]     g      raw gain
+ * @param [in]      beta   activity masking beta param
+ * @param [in]      bshift shift to be applied to raw gain
+ * @return                 quantized/companded gain
+ */
+od_val32 od_pvq_compute_gain(const od_val16 *x, int n, int q0, od_val32 *g,
+ od_val16 beta, int bshift) {
+  int i;
+  od_val32 acc;
+#if !defined(OD_FLOAT_PVQ)
+  od_val32 irt;
+  int sqrt_shift;
+#else
+  OD_UNUSED(bshift);
+#endif
+  acc = 0;
+  for (i = 0; i < n; i++) {
+    acc += x[i]*(od_val32)x[i];
+  }
+#if defined(OD_FLOAT_PVQ)
+  *g = sqrt(acc);
+#else
+  irt = od_sqrt(acc, &sqrt_shift);
+  *g = OD_VSHR_ROUND(irt, sqrt_shift - bshift);
+#endif
+  /* Normalize gain by quantization step size and apply companding
+     (if ACTIVITY != 1). */
+  return od_gain_compand(*g, q0, beta);
+}
+
+/** Compute theta quantization range from quantized/companded gain
+ *
+ * @param [in]      qcg    quantized companded gain value
+ * @param [in]      beta   activity masking beta param
+ * @return                 max theta value
+ */
+int od_pvq_compute_max_theta(od_val32 qcg, od_val16 beta){
+  /* Set angular resolution (in ra) to match the encoded gain */
+#if defined(OD_FLOAT_PVQ)
+  int ts = (int)floor(.5 + qcg*OD_CGAIN_SCALE_1*M_PI/(2*beta));
+#else
+  int ts = OD_SHR_ROUND(qcg*OD_MULT16_16_QBETA(OD_QCONST32(M_PI/2,
+   OD_CGAIN_SHIFT), od_beta_rcp(beta)), OD_CGAIN_SHIFT*2);
+#endif
+  /* Special case for low gains -- will need to be tuned anyway */
+  if (qcg < OD_QCONST32(1.4, OD_CGAIN_SHIFT)) ts = 1;
+  return ts;
+}
+
+/** Decode quantized theta value from coded value
+ *
+ * @param [in]      t          quantized companded gain value
+ * @param [in]      max_theta  maximum theta value
+ * @return                     decoded theta value
+ */
+od_val32 od_pvq_compute_theta(int t, int max_theta) {
+  if (max_theta != 0) {
+#if defined(OD_FLOAT_PVQ)
+    return OD_MINI(t, max_theta - 1)*.5*M_PI/max_theta;
+#else
+    return (OD_MAX_THETA_SCALE*OD_MINI(t, max_theta - 1)
+     + (max_theta >> 1))/max_theta;
+#endif
+  }
+  else return 0;
+}
+
+#define OD_SQRT_TBL_SHIFT (10)
+
+#define OD_ITHETA_SHIFT 15
+/** Compute the number of pulses used for PVQ encoding a vector from
+ * available metrics (encode and decode side)
+ *
+ * @param [in]      qcg        quantized companded gain value
+ * @param [in]      itheta     quantized PVQ error angle theta
+ * @param [in]      noref      indicates present or lack of reference
+ *                             (prediction)
+ * @param [in]      n          number of elements to be coded
+ * @param [in]      beta       activity masking beta param
+ * @return                     number of pulses to use for coding
+ */
+int od_pvq_compute_k(od_val32 qcg, int itheta, int noref, int n,
+    od_val16 beta) {
+#if !defined(OD_FLOAT_PVQ)
+  /*Lookup table for sqrt(n+3/2) and sqrt(n+2/2) in Q10.
+    Real max values are 32792 and 32784, but clamped to stay within 16 bits.
+    Update with tools/gen_sqrt_tbl if needed.*/
+  static const od_val16 od_sqrt_table[2][13] = {
+   {0, 0, 0, 0, 2290, 2985, 4222, 0, 8256, 0, 16416, 0, 32767},
+   {0, 0, 0, 0, 2401, 3072, 4284, 0, 8287, 0, 16432, 0, 32767}};
+#endif
+  if (noref) {
+    if (qcg == 0) return 0;
+    if (n == 15 && qcg == OD_CGAIN_SCALE && beta > OD_BETA(1.25)) {
+      return 1;
+    }
+    else {
+#if defined(OD_FLOAT_PVQ)
+      return OD_MAXI(1, (int)floor(.5 + (qcg*OD_CGAIN_SCALE_1 - .2)*
+       sqrt((n + 3)/2)/beta));
+#else
+      od_val16 rt;
+      OD_ASSERT(OD_ILOG(n + 1) < 13);
+      rt = od_sqrt_table[1][OD_ILOG(n + 1)];
+      /*FIXME: get rid of 64-bit mul.*/
+      return OD_MAXI(1, OD_SHR_ROUND((int64_t)((qcg
+       - (int64_t)OD_QCONST32(.2, OD_CGAIN_SHIFT))*
+       OD_MULT16_16_QBETA(od_beta_rcp(beta), rt)), OD_CGAIN_SHIFT
+       + OD_SQRT_TBL_SHIFT));
+#endif
+    }
+  }
+  else {
+    if (itheta == 0) return 0;
+    /* Sets K according to gain and theta, based on the high-rate
+       PVQ distortion curves (see PVQ document). Low-rate will have to be
+       perceptually tuned anyway. We subtract 0.2 from the radius as an
+       approximation for the fact that the coefficients aren't identically
+       distributed within a band so at low gain the number of dimensions that
+       are likely to have a pulse is less than n. */
+#if defined(OD_FLOAT_PVQ)
+    return OD_MAXI(1, (int)floor(.5 + (itheta - .2)*sqrt((n + 2)/2)));
+#else
+    od_val16 rt;
+    OD_ASSERT(OD_ILOG(n + 1) < 13);
+    rt = od_sqrt_table[0][OD_ILOG(n + 1)];
+    /*FIXME: get rid of 64-bit mul.*/
+    return OD_MAXI(1, OD_VSHR_ROUND(((OD_SHL(itheta, OD_ITHETA_SHIFT)
+     - OD_QCONST32(.2, OD_ITHETA_SHIFT)))*(int64_t)rt,
+     OD_SQRT_TBL_SHIFT + OD_ITHETA_SHIFT));
+#endif
+  }
+}
+
+#if !defined(OD_FLOAT_PVQ)
+#define OD_RSQRT_INSHIFT 16
+#define OD_RSQRT_OUTSHIFT 14
+/** Reciprocal sqrt approximation where the input is in the range [0.25,1) in
+     Q16 and the output is in the range (1.0, 2.0] in Q14).
+    Error is always within +/1 of round(1/sqrt(t))*/
+static int16_t od_rsqrt_norm(int16_t t)
+{
+  int16_t n;
+  int32_t r;
+  int32_t r2;
+  int32_t ry;
+  int32_t y;
+  int32_t ret;
+  /* Range of n is [-16384,32767] ([-0.5,1) in Q15).*/
+  n = t - 32768;
+  OD_ASSERT(n >= -16384);
+  /*Get a rough initial guess for the root.
+    The optimal minimax quadratic approximation (using relative error) is
+     r = 1.437799046117536+n*(-0.823394375837328+n*0.4096419668459485).
+    Coefficients here, and the final result r, are Q14.*/
+  r = (23565 + OD_MULT16_16_Q15(n, (-13481 + OD_MULT16_16_Q15(n, 6711))));
+  /*We want y = t*r*r-1 in Q15, but t is 32-bit Q16 and r is Q14.
+    We can compute the result from n and r using Q15 multiplies with some
+     adjustment, carefully done to avoid overflow.*/
+  r2 = r*r;
+  y = (((r2 >> 15)*n + r2) >> 12) - 131077;
+  ry = r*y;
+  /*Apply a 2nd-order Householder iteration: r += r*y*(y*0.375-0.5).
+    This yields the Q14 reciprocal square root of the Q16 t, with a maximum
+     relative error of 1.04956E-4, a (relative) RMSE of 2.80979E-5, and a peak
+     absolute error of 2.26591/16384.*/
+  ret = r + ((((ry >> 16)*(3*y) >> 3) - ry) >> 18);
+  OD_ASSERT(ret >= 16384 && ret < 32768);
+  return (int16_t)ret;
+}
+
+static int16_t od_rsqrt(int32_t x, int *rsqrt_shift)
+{
+   int k;
+   int s;
+   int16_t t;
+   k = (OD_ILOG(x) - 1) >> 1;
+  /*t is x in the range [0.25, 1) in QINSHIFT, or x*2^(-s).
+    Shift by log2(x) - log2(0.25*(1 << INSHIFT)) to ensure 0.25 lower bound.*/
+   s = 2*k - (OD_RSQRT_INSHIFT - 2);
+   t = OD_VSHR(x, s);
+   /*We want to express od_rsqrt() in terms of od_rsqrt_norm(), which is
+      defined as (2^OUTSHIFT)/sqrt(t*(2^-INSHIFT)) with t=x*(2^-s).
+     This simplifies to 2^(OUTSHIFT+(INSHIFT/2)+(s/2))/sqrt(x), so the caller
+      needs to shift right by OUTSHIFT + INSHIFT/2 + s/2.*/
+   *rsqrt_shift = OD_RSQRT_OUTSHIFT + ((s + OD_RSQRT_INSHIFT) >> 1);
+   return od_rsqrt_norm(t);
+}
+#endif
+
+/** Synthesizes one parition of coefficient values from a PVQ-encoded
+ * vector.  This 'partial' version is called by the encode loop where
+ * the Householder reflection has already been computed and there's no
+ * need to recompute it.
+ *
+ * @param [out]     xcoeff  output coefficient partition (x in math doc)
+ * @param [in]      ypulse  PVQ-encoded values (y in the math doc); in
+ *                          the noref case, this vector has n entries,
+ *                          in the reference case it contains n-1 entries
+ *                          (the m-th entry is not included)
+ * @param [in]      r       reference vector (prediction)
+ * @param [in]      n       number of elements in this partition
+ * @param [in]      noref   indicates presence or lack of prediction
+ * @param [in]      g       decoded quantized vector gain
+ * @param [in]      theta   decoded theta (prediction error)
+ * @param [in]      m       alignment dimension of Householder reflection
+ * @param [in]      s       sign of Householder reflection
+ * @param [in]      qm_inv  inverse of the QM with magnitude compensation
+ */
+void od_pvq_synthesis_partial(od_coeff *xcoeff, const od_coeff *ypulse,
+ const od_val16 *r16, int n, int noref, od_val32 g, od_val32 theta, int m, int s,
+ const int16_t *qm_inv) {
+  int i;
+  int yy;
+  od_val32 scale;
+  int nn;
+#if !defined(OD_FLOAT_PVQ)
+  int gshift;
+  int qshift;
+#endif
+  OD_ASSERT(g != 0);
+  nn = n-(!noref); /* when noref==0, vector in is sized n-1 */
+  yy = 0;
+  for (i = 0; i < nn; i++)
+    yy += ypulse[i]*(int32_t)ypulse[i];
+#if !defined(OD_FLOAT_PVQ)
+  /* Shift required for the magnitude of the pre-qm synthesis to be guaranteed
+     to fit in 16 bits. In practice, the range will be 8192-16384 after scaling
+     most of the time. */
+  gshift = OD_MAXI(0, OD_ILOG(g) - 14);
+#endif
+  /*scale is g/sqrt(yy) in Q(16-gshift) so that x[]*scale has a norm that fits
+     in 16 bits.*/
+  if (yy == 0) scale = 0;
+#if defined(OD_FLOAT_PVQ)
+  else {
+    scale = g/sqrt(yy);
+  }
+#else
+  else {
+    int rsqrt_shift;
+    int16_t rsqrt;
+    /*FIXME: should be < int64_t*/
+    int64_t tmp;
+    rsqrt = od_rsqrt(yy, &rsqrt_shift);
+    tmp = rsqrt*(int64_t)g;
+    scale = OD_VSHR_ROUND(tmp, rsqrt_shift + gshift - 16);
+  }
+  /* Shift to apply after multiplying by the inverse QM, taking into account
+     gshift. */
+  qshift = OD_QM_INV_SHIFT - gshift;
+#endif
+  if (noref) {
+    for (i = 0; i < n; i++) {
+      od_val32 x;
+      /* This multiply doesn't round, so it introduces some bias.
+         It would be nice (but not critical) to fix this. */
+      x = OD_MULT16_32_Q16(ypulse[i], scale);
+#if defined(OD_FLOAT_PVQ)
+      xcoeff[i] = (od_coeff)floor(.5
+       + x*(qm_inv[i]*OD_QM_INV_SCALE_1));
+#else
+      xcoeff[i] = OD_SHR_ROUND(x*qm_inv[i], qshift);
+#endif
+    }
+  }
+  else{
+    od_val16 x[MAXN];
+    scale = OD_ROUND32(scale*OD_TRIG_SCALE_1*od_pvq_sin(theta));
+    /* The following multiply doesn't round, but it's probably OK since
+       the Householder reflection is likely to undo most of the resulting
+       bias. */
+    for (i = 0; i < m; i++)
+      x[i] = OD_MULT16_32_Q16(ypulse[i], scale);
+    x[m] = OD_ROUND16(-s*(OD_SHR_ROUND(g, gshift))*OD_TRIG_SCALE_1*
+     od_pvq_cos(theta));
+    for (i = m; i < nn; i++)
+      x[i+1] = OD_MULT16_32_Q16(ypulse[i], scale);
+    od_apply_householder(x, x, r16, n);
+    for (i = 0; i < n; i++) {
+#if defined(OD_FLOAT_PVQ)
+      xcoeff[i] = (od_coeff)floor(.5 + (x[i]*(qm_inv[i]*OD_QM_INV_SCALE_1)));
+#else
+      xcoeff[i] = OD_SHR_ROUND(x[i]*qm_inv[i], qshift);
+#endif
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/pvq.h b/third_party/aom/av1/common/pvq.h
new file mode 100644
index 0000000000..17e54d4c5a
--- /dev/null
+++ b/third_party/aom/av1/common/pvq.h
@@ -0,0 +1,183 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_pvq_H)
+# define _pvq_H (1)
+# include "generic_code.h"
+# include "odintrin.h"
+
+extern const uint16_t EXP_CDF_TABLE[][16];
+extern const uint16_t LAPLACE_OFFSET[];
+
+#if CONFIG_DAALA_DIST
+#define AV1_PVQ_ENABLE_ACTIVITY_MASKING (1)
+#else
+#define AV1_PVQ_ENABLE_ACTIVITY_MASKING (0)
+#endif
+
+# define PVQ_MAX_PARTITIONS (1 + 3*(OD_TXSIZES-1))
+
+# define OD_NOREF_ADAPT_SPEED (4)
+/* Normalized lambda for PVQ quantizer. Since we normalize the gain by q, the
+   distortion is normalized by q^2 and lambda does not need the q^2 factor.
+   At high rate, this would be log(2)/6, but we're using a slightly more
+   aggressive value, closer to:
+   Li, Xiang, et al. "Laplace distribution based Lagrangian rate distortion
+   optimization for hybrid video coding." Circuits and Systems for Video
+   Technology, IEEE Transactions on 19.2 (2009): 193-205.
+   */
+# define OD_PVQ_LAMBDA (.1146)
+
+#define OD_PVQ_SKIP_ZERO 1
+#define OD_PVQ_SKIP_COPY 2
+
+/* Maximum size for coding a PVQ band. */
+#define OD_MAX_PVQ_SIZE (1024)
+
+#if defined(OD_FLOAT_PVQ)
+#define OD_QM_SHIFT (15)
+#else
+#define OD_QM_SHIFT (11)
+#endif
+#define OD_QM_SCALE (1 << OD_QM_SHIFT)
+#if defined(OD_FLOAT_PVQ)
+#define OD_QM_SCALE_1 (1./OD_QM_SCALE)
+#endif
+#define OD_QM_SCALE_MAX 32767
+#define OD_QM_INV_SHIFT (12)
+#define OD_QM_INV_SCALE (1 << OD_QM_INV_SHIFT)
+#if defined(OD_FLOAT_PVQ)
+#define OD_QM_INV_SCALE_1 (1./OD_QM_INV_SCALE)
+#endif
+#define OD_QM_OFFSET(bs) ((((1 << 2*bs) - 1) << 2*OD_LOG_BSIZE0)/3)
+#define OD_QM_STRIDE (OD_QM_OFFSET(OD_TXSIZES))
+#define OD_QM_BUFFER_SIZE (2*OD_QM_STRIDE)
+
+#if !defined(OD_FLOAT_PVQ)
+#define OD_THETA_SHIFT (15)
+#define OD_THETA_SCALE ((1 << OD_THETA_SHIFT)*2./M_PI)
+#define OD_MAX_THETA_SCALE (1 << OD_THETA_SHIFT)
+#define OD_TRIG_SCALE (32768)
+#define OD_BETA_SHIFT (12)
+#define OD_BETA_SCALE_1 (1./(1 << OD_BETA_SHIFT))
+/*Multiplies 16-bit a by 32-bit b and keeps bits [16:64-OD_BETA_SHIFT-1].*/
+#define OD_MULT16_32_QBETA(a, b) \
+ ((int16_t)(a)*(int64_t)(int32_t)(b) >> OD_BETA_SHIFT)
+# define OD_MULT16_16_QBETA(a, b) \
+  ((((int16_t)(a))*((int32_t)(int16_t)(b))) >> OD_BETA_SHIFT)
+#define OD_CGAIN_SHIFT (8)
+#define OD_CGAIN_SCALE (1 << OD_CGAIN_SHIFT)
+#else
+#define OD_BETA_SCALE_1 (1.)
+#define OD_THETA_SCALE (1)
+#define OD_TRIG_SCALE (1)
+#define OD_CGAIN_SCALE (1)
+#endif
+#define OD_THETA_SCALE_1 (1./OD_THETA_SCALE)
+#define OD_TRIG_SCALE_1 (1./OD_TRIG_SCALE)
+#define OD_CGAIN_SCALE_1 (1./OD_CGAIN_SCALE)
+#define OD_CGAIN_SCALE_2 (OD_CGAIN_SCALE_1*OD_CGAIN_SCALE_1)
+
+/* Largest PVQ partition is half the coefficients of largest block size. */
+#define MAXN (OD_TXSIZE_MAX*OD_TXSIZE_MAX/2)
+
+#define OD_COMPAND_SHIFT (8 + OD_COEFF_SHIFT)
+#define OD_COMPAND_SCALE (1 << OD_COMPAND_SHIFT)
+#define OD_COMPAND_SCALE_1 (1./OD_COMPAND_SCALE)
+
+#define OD_QM_SIZE (OD_TXSIZES*(OD_TXSIZES + 1))
+
+#define OD_FLAT_QM 0
+#define OD_HVS_QM  1
+
+# define OD_NSB_ADAPT_CTXS (4)
+
+# define OD_ADAPT_K_Q8        0
+# define OD_ADAPT_SUM_EX_Q8   1
+# define OD_ADAPT_COUNT_Q8    2
+# define OD_ADAPT_COUNT_EX_Q8 3
+
+# define OD_ADAPT_NO_VALUE (-2147483647-1)
+
+typedef enum {
+  PVQ_SKIP = 0x0,
+  DC_CODED = 0x1,
+  AC_CODED = 0x2,
+  AC_DC_CODED = 0x3,
+} PVQ_SKIP_TYPE;
+
+typedef struct od_pvq_adapt_ctx  od_pvq_adapt_ctx;
+typedef struct od_pvq_codeword_ctx od_pvq_codeword_ctx;
+
+struct od_pvq_codeword_ctx {
+  int                 pvq_adapt[2*OD_TXSIZES*OD_NSB_ADAPT_CTXS];
+  /* CDFs are size 16 despite the fact that we're using less than that. */
+  uint16_t            pvq_k1_cdf[12][CDF_SIZE(16)];
+  uint16_t            pvq_split_cdf[22*7][CDF_SIZE(8)];
+};
+
+struct od_pvq_adapt_ctx {
+  od_pvq_codeword_ctx pvq_codeword_ctx;
+  generic_encoder     pvq_param_model[3];
+  int                 pvq_ext[OD_TXSIZES*PVQ_MAX_PARTITIONS];
+  int                 pvq_exg[OD_NPLANES_MAX][OD_TXSIZES][PVQ_MAX_PARTITIONS];
+  uint16_t pvq_gaintheta_cdf[2*OD_TXSIZES*PVQ_MAX_PARTITIONS][CDF_SIZE(16)];
+  uint16_t pvq_skip_dir_cdf[2*(OD_TXSIZES-1)][CDF_SIZE(7)];
+};
+
+typedef struct od_qm_entry {
+  int interp_q;
+  int scale_q8;
+  const unsigned char *qm_q4;
+} od_qm_entry;
+
+extern const od_qm_entry OD_DEFAULT_QMS[2][2][OD_NPLANES_MAX];
+
+void od_adapt_pvq_ctx_reset(od_pvq_adapt_ctx *state, int is_keyframe);
+int od_pvq_size_ctx(int n);
+int od_pvq_k1_ctx(int n, int orig_size);
+
+od_val16 od_pvq_sin(od_val32 x);
+od_val16 od_pvq_cos(od_val32 x);
+#if !defined(OD_FLOAT_PVQ)
+int od_vector_log_mag(const od_coeff *x, int n);
+#endif
+
+void od_interp_qm(unsigned char *out, int q, const od_qm_entry *entry1,
+                  const od_qm_entry *entry2);
+
+int od_qm_get_index(int bs, int band);
+
+extern const od_val16 *const OD_PVQ_BETA[2][OD_NPLANES_MAX][OD_TXSIZES + 1];
+
+void od_init_qm(int16_t *x, int16_t *x_inv, const int *qm);
+int od_compute_householder(od_val16 *r, int n, od_val32 gr, int *sign,
+ int shift);
+void od_apply_householder(od_val16 *out, const od_val16 *x, const od_val16 *r,
+ int n);
+void od_pvq_synthesis_partial(od_coeff *xcoeff, const od_coeff *ypulse,
+                                  const od_val16 *r, int n,
+                                  int noref, od_val32 g,
+                                  od_val32 theta, int m, int s,
+                                  const int16_t *qm_inv);
+od_val32 od_gain_expand(od_val32 cg, int q0, od_val16 beta);
+od_val32 od_pvq_compute_gain(const od_val16 *x, int n, int q0, od_val32 *g,
+ od_val16 beta, int bshift);
+int od_pvq_compute_max_theta(od_val32 qcg, od_val16 beta);
+od_val32 od_pvq_compute_theta(int t, int max_theta);
+int od_pvq_compute_k(od_val32 qcg, int itheta, int noref, int n, od_val16 beta);
+
+int od_vector_is_null(const od_coeff *x, int len);
+int od_qm_offset(int bs, int xydec);
+
+#endif
diff --git a/third_party/aom/av1/common/pvq_state.c b/third_party/aom/av1/common/pvq_state.c
new file mode 100644
index 0000000000..197b9b3a82
--- /dev/null
+++ b/third_party/aom/av1/common/pvq_state.c
@@ -0,0 +1,50 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/pvq_state.h"
+#include "av1/common/odintrin.h"
+
+void od_adapt_ctx_reset(od_adapt_ctx *adapt, int is_keyframe) {
+  int pli;
+  od_adapt_pvq_ctx_reset(&adapt->pvq, is_keyframe);
+  OD_CDFS_INIT_Q15(adapt->skip_cdf);
+  for (pli = 0; pli < OD_NPLANES_MAX; pli++) {
+    int i;
+    OD_CDFS_INIT_DYNAMIC(adapt->model_dc[pli].cdf);
+    for (i = 0; i < OD_TXSIZES; i++) {
+      int j;
+      adapt->ex_g[pli][i] = 8;
+      for (j = 0; j < 3; j++) {
+        adapt->ex_dc[pli][i][j] = pli > 0 ? 8 : 32768;
+      }
+    }
+  }
+}
+
+void od_init_skipped_coeffs(int16_t *d, int16_t *pred, int is_keyframe, int bo,
+                            int n, int w) {
+  int i;
+  int j;
+  if (is_keyframe) {
+    for (i = 0; i < n; i++) {
+      for (j = 0; j < n; j++) {
+        /* skip DC */
+        if (i || j) d[bo + i * w + j] = 0;
+      }
+    }
+  } else {
+    for (i = 0; i < n; i++) {
+      for (j = 0; j < n; j++) {
+        d[bo + i * w + j] = pred[i * n + j];
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/pvq_state.h b/third_party/aom/av1/common/pvq_state.h
new file mode 100644
index 0000000000..84d454e700
--- /dev/null
+++ b/third_party/aom/av1/common/pvq_state.h
@@ -0,0 +1,52 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_state_H)
+# define _state_H (1)
+
+typedef struct od_state     od_state;
+typedef struct od_adapt_ctx od_adapt_ctx;
+
+# include "generic_code.h"
+# include "odintrin.h"
+# include "pvq.h"
+
+/*Adaptation speed of scalar Laplace encoding.*/
+# define OD_SCALAR_ADAPT_SPEED (4)
+
+struct od_adapt_ctx {
+  /* Support for PVQ encode/decode */
+  od_pvq_adapt_ctx pvq;
+
+  generic_encoder model_dc[OD_NPLANES_MAX];
+
+  int ex_dc[OD_NPLANES_MAX][OD_TXSIZES][3];
+  int ex_g[OD_NPLANES_MAX][OD_TXSIZES];
+
+  /* Joint skip flag for DC and AC */
+  uint16_t skip_cdf[OD_TXSIZES*2][CDF_SIZE(4)];
+};
+
+struct od_state {
+  od_adapt_ctx *adapt;
+  unsigned char pvq_qm_q4[OD_NPLANES_MAX][OD_QM_SIZE];
+  /* Quantization matrices and their inverses. */
+  int16_t qm[OD_QM_BUFFER_SIZE];
+  int16_t qm_inv[OD_QM_BUFFER_SIZE];
+};
+
+void od_adapt_ctx_reset(od_adapt_ctx *state, int is_keyframe);
+void od_init_skipped_coeffs(int16_t *d, int16_t *pred, int is_keyframe,
+ int bo, int n, int w);
+
+#endif
diff --git a/third_party/aom/av1/common/quant_common.c b/third_party/aom/av1/common/quant_common.c
new file mode 100644
index 0000000000..763465e480
--- /dev/null
+++ b/third_party/aom/av1/common/quant_common.c
@@ -0,0 +1,11369 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/common.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/entropy.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/seg_common.h"
+#include "av1/common/blockd.h"
+
+#if CONFIG_NEW_QUANT
+// Bin widths expressed as a fraction over 128 of the quant stepsize,
+// for the quantization bins 0-4.
+// So a value x indicates the bin is actually factor x/128 of the
+// nominal quantization step.  For the zero bin, the width is only
+// for one side of zero, so the actual width is twice that.
+//
+// Functions with nuq correspond to "non uniform quantization"
+// TODO(sarahparker, debargha): Optimize these tables
+
+typedef struct {
+  uint8_t knots[NUQ_KNOTS];  // offsets
+  uint8_t doff;              // dequantization
+} qprofile_type;
+
+static const qprofile_type nuq[QUANT_PROFILES][COEF_BANDS] = {
+  {
+      // lossless
+      { { 64, 128, 128 }, 0 },  // dc, band 0
+      { { 64, 128, 128 }, 0 },  // band 1
+      { { 64, 128, 128 }, 0 },  // band 2
+      { { 64, 128, 128 }, 0 },  // band 3
+      { { 64, 128, 128 }, 0 },  // band 4
+      { { 64, 128, 128 }, 0 },  // band 5
+  },
+  {
+      { { 64, 128, 128 }, 4 },   // dc, band 0
+      { { 64, 128, 128 }, 6 },   // band 1
+      { { 64, 128, 128 }, 8 },   // band 2
+      { { 64, 128, 128 }, 10 },  // band 3
+      { { 72, 128, 128 }, 12 },  // band 4
+      { { 80, 128, 128 }, 14 }   // band 5
+  },
+  {
+      { { 64, 128, 128 }, 6 },   // dc, band 0
+      { { 64, 128, 128 }, 8 },   // band 1
+      { { 64, 128, 128 }, 10 },  // band 2
+      { { 64, 128, 128 }, 12 },  // band 3
+      { { 72, 128, 128 }, 14 },  // band 4
+      { { 80, 128, 128 }, 16 }   // band 5
+  },
+  {
+      { { 64, 128, 128 }, 8 },   // dc, band 0
+      { { 64, 128, 128 }, 10 },  // band 1
+      { { 64, 128, 128 }, 12 },  // band 2
+      { { 72, 128, 128 }, 14 },  // band 3
+      { { 76, 128, 128 }, 16 },  // band 4
+      { { 80, 128, 128 }, 18 }   // band 5
+  }
+};
+
+static const uint8_t *get_nuq_knots(int band, int q_profile) {
+  return nuq[q_profile][band].knots;
+}
+
+static INLINE int16_t quant_to_doff_fixed(int band, int q_profile) {
+  return nuq[q_profile][band].doff;
+}
+
+// get cumulative bins
+static INLINE void get_cuml_bins_nuq(int q, int band, tran_low_t *cuml_bins,
+                                     int q_profile) {
+  const uint8_t *knots = get_nuq_knots(band, q_profile);
+  int16_t cuml_knots[NUQ_KNOTS];
+  int i;
+  cuml_knots[0] = knots[0];
+  for (i = 1; i < NUQ_KNOTS; ++i) cuml_knots[i] = cuml_knots[i - 1] + knots[i];
+  for (i = 0; i < NUQ_KNOTS; ++i)
+    cuml_bins[i] = ROUND_POWER_OF_TWO(cuml_knots[i] * q, 7);
+}
+
+void av1_get_dequant_val_nuq(int q, int band, tran_low_t *dq,
+                             tran_low_t *cuml_bins, int q_profile) {
+  const uint8_t *knots = get_nuq_knots(band, q_profile);
+  tran_low_t cuml_bins_[NUQ_KNOTS], *cuml_bins_ptr;
+  tran_low_t doff;
+  int i;
+  cuml_bins_ptr = (cuml_bins ? cuml_bins : cuml_bins_);
+  get_cuml_bins_nuq(q, band, cuml_bins_ptr, q_profile);
+  dq[0] = 0;
+  for (i = 1; i < NUQ_KNOTS; ++i) {
+    doff = quant_to_doff_fixed(band, q_profile);
+    doff = ROUND_POWER_OF_TWO(doff * knots[i], 7);
+    dq[i] =
+        cuml_bins_ptr[i - 1] + ROUND_POWER_OF_TWO((knots[i] - doff * 2) * q, 8);
+  }
+  doff = quant_to_doff_fixed(band, q_profile);
+  dq[NUQ_KNOTS] =
+      cuml_bins_ptr[NUQ_KNOTS - 1] + ROUND_POWER_OF_TWO((64 - doff) * q, 7);
+}
+
+tran_low_t av1_dequant_abscoeff_nuq(int v, int q, const tran_low_t *dq) {
+  if (v <= NUQ_KNOTS)
+    return dq[v];
+  else
+    return dq[NUQ_KNOTS] + (v - NUQ_KNOTS) * q;
+}
+
+tran_low_t av1_dequant_coeff_nuq(int v, int q, const tran_low_t *dq) {
+  tran_low_t dqmag = av1_dequant_abscoeff_nuq(abs(v), q, dq);
+  return (v < 0 ? -dqmag : dqmag);
+}
+#endif  // CONFIG_NEW_QUANT
+
+static const int16_t dc_qlookup[QINDEX_RANGE] = {
+  4,    8,    8,    9,    10,  11,  12,  12,  13,  14,  15,   16,   17,   18,
+  19,   19,   20,   21,   22,  23,  24,  25,  26,  26,  27,   28,   29,   30,
+  31,   32,   32,   33,   34,  35,  36,  37,  38,  38,  39,   40,   41,   42,
+  43,   43,   44,   45,   46,  47,  48,  48,  49,  50,  51,   52,   53,   53,
+  54,   55,   56,   57,   57,  58,  59,  60,  61,  62,  62,   63,   64,   65,
+  66,   66,   67,   68,   69,  70,  70,  71,  72,  73,  74,   74,   75,   76,
+  77,   78,   78,   79,   80,  81,  81,  82,  83,  84,  85,   85,   87,   88,
+  90,   92,   93,   95,   96,  98,  99,  101, 102, 104, 105,  107,  108,  110,
+  111,  113,  114,  116,  117, 118, 120, 121, 123, 125, 127,  129,  131,  134,
+  136,  138,  140,  142,  144, 146, 148, 150, 152, 154, 156,  158,  161,  164,
+  166,  169,  172,  174,  177, 180, 182, 185, 187, 190, 192,  195,  199,  202,
+  205,  208,  211,  214,  217, 220, 223, 226, 230, 233, 237,  240,  243,  247,
+  250,  253,  257,  261,  265, 269, 272, 276, 280, 284, 288,  292,  296,  300,
+  304,  309,  313,  317,  322, 326, 330, 335, 340, 344, 349,  354,  359,  364,
+  369,  374,  379,  384,  389, 395, 400, 406, 411, 417, 423,  429,  435,  441,
+  447,  454,  461,  467,  475, 482, 489, 497, 505, 513, 522,  530,  539,  549,
+  559,  569,  579,  590,  602, 614, 626, 640, 654, 668, 684,  700,  717,  736,
+  755,  775,  796,  819,  843, 869, 896, 925, 955, 988, 1022, 1058, 1098, 1139,
+  1184, 1232, 1282, 1336,
+};
+
+#if CONFIG_HIGHBITDEPTH
+static const int16_t dc_qlookup_10[QINDEX_RANGE] = {
+  4,    9,    10,   13,   15,   17,   20,   22,   25,   28,   31,   34,   37,
+  40,   43,   47,   50,   53,   57,   60,   64,   68,   71,   75,   78,   82,
+  86,   90,   93,   97,   101,  105,  109,  113,  116,  120,  124,  128,  132,
+  136,  140,  143,  147,  151,  155,  159,  163,  166,  170,  174,  178,  182,
+  185,  189,  193,  197,  200,  204,  208,  212,  215,  219,  223,  226,  230,
+  233,  237,  241,  244,  248,  251,  255,  259,  262,  266,  269,  273,  276,
+  280,  283,  287,  290,  293,  297,  300,  304,  307,  310,  314,  317,  321,
+  324,  327,  331,  334,  337,  343,  350,  356,  362,  369,  375,  381,  387,
+  394,  400,  406,  412,  418,  424,  430,  436,  442,  448,  454,  460,  466,
+  472,  478,  484,  490,  499,  507,  516,  525,  533,  542,  550,  559,  567,
+  576,  584,  592,  601,  609,  617,  625,  634,  644,  655,  666,  676,  687,
+  698,  708,  718,  729,  739,  749,  759,  770,  782,  795,  807,  819,  831,
+  844,  856,  868,  880,  891,  906,  920,  933,  947,  961,  975,  988,  1001,
+  1015, 1030, 1045, 1061, 1076, 1090, 1105, 1120, 1137, 1153, 1170, 1186, 1202,
+  1218, 1236, 1253, 1271, 1288, 1306, 1323, 1342, 1361, 1379, 1398, 1416, 1436,
+  1456, 1476, 1496, 1516, 1537, 1559, 1580, 1601, 1624, 1647, 1670, 1692, 1717,
+  1741, 1766, 1791, 1817, 1844, 1871, 1900, 1929, 1958, 1990, 2021, 2054, 2088,
+  2123, 2159, 2197, 2236, 2276, 2319, 2363, 2410, 2458, 2508, 2561, 2616, 2675,
+  2737, 2802, 2871, 2944, 3020, 3102, 3188, 3280, 3375, 3478, 3586, 3702, 3823,
+  3953, 4089, 4236, 4394, 4559, 4737, 4929, 5130, 5347,
+};
+
+static const int16_t dc_qlookup_12[QINDEX_RANGE] = {
+  4,     12,    18,    25,    33,    41,    50,    60,    70,    80,    91,
+  103,   115,   127,   140,   153,   166,   180,   194,   208,   222,   237,
+  251,   266,   281,   296,   312,   327,   343,   358,   374,   390,   405,
+  421,   437,   453,   469,   484,   500,   516,   532,   548,   564,   580,
+  596,   611,   627,   643,   659,   674,   690,   706,   721,   737,   752,
+  768,   783,   798,   814,   829,   844,   859,   874,   889,   904,   919,
+  934,   949,   964,   978,   993,   1008,  1022,  1037,  1051,  1065,  1080,
+  1094,  1108,  1122,  1136,  1151,  1165,  1179,  1192,  1206,  1220,  1234,
+  1248,  1261,  1275,  1288,  1302,  1315,  1329,  1342,  1368,  1393,  1419,
+  1444,  1469,  1494,  1519,  1544,  1569,  1594,  1618,  1643,  1668,  1692,
+  1717,  1741,  1765,  1789,  1814,  1838,  1862,  1885,  1909,  1933,  1957,
+  1992,  2027,  2061,  2096,  2130,  2165,  2199,  2233,  2267,  2300,  2334,
+  2367,  2400,  2434,  2467,  2499,  2532,  2575,  2618,  2661,  2704,  2746,
+  2788,  2830,  2872,  2913,  2954,  2995,  3036,  3076,  3127,  3177,  3226,
+  3275,  3324,  3373,  3421,  3469,  3517,  3565,  3621,  3677,  3733,  3788,
+  3843,  3897,  3951,  4005,  4058,  4119,  4181,  4241,  4301,  4361,  4420,
+  4479,  4546,  4612,  4677,  4742,  4807,  4871,  4942,  5013,  5083,  5153,
+  5222,  5291,  5367,  5442,  5517,  5591,  5665,  5745,  5825,  5905,  5984,
+  6063,  6149,  6234,  6319,  6404,  6495,  6587,  6678,  6769,  6867,  6966,
+  7064,  7163,  7269,  7376,  7483,  7599,  7715,  7832,  7958,  8085,  8214,
+  8352,  8492,  8635,  8788,  8945,  9104,  9275,  9450,  9639,  9832,  10031,
+  10245, 10465, 10702, 10946, 11210, 11482, 11776, 12081, 12409, 12750, 13118,
+  13501, 13913, 14343, 14807, 15290, 15812, 16356, 16943, 17575, 18237, 18949,
+  19718, 20521, 21387,
+};
+#endif
+
+static const int16_t ac_qlookup[QINDEX_RANGE] = {
+  4,    8,    9,    10,   11,   12,   13,   14,   15,   16,   17,   18,   19,
+  20,   21,   22,   23,   24,   25,   26,   27,   28,   29,   30,   31,   32,
+  33,   34,   35,   36,   37,   38,   39,   40,   41,   42,   43,   44,   45,
+  46,   47,   48,   49,   50,   51,   52,   53,   54,   55,   56,   57,   58,
+  59,   60,   61,   62,   63,   64,   65,   66,   67,   68,   69,   70,   71,
+  72,   73,   74,   75,   76,   77,   78,   79,   80,   81,   82,   83,   84,
+  85,   86,   87,   88,   89,   90,   91,   92,   93,   94,   95,   96,   97,
+  98,   99,   100,  101,  102,  104,  106,  108,  110,  112,  114,  116,  118,
+  120,  122,  124,  126,  128,  130,  132,  134,  136,  138,  140,  142,  144,
+  146,  148,  150,  152,  155,  158,  161,  164,  167,  170,  173,  176,  179,
+  182,  185,  188,  191,  194,  197,  200,  203,  207,  211,  215,  219,  223,
+  227,  231,  235,  239,  243,  247,  251,  255,  260,  265,  270,  275,  280,
+  285,  290,  295,  300,  305,  311,  317,  323,  329,  335,  341,  347,  353,
+  359,  366,  373,  380,  387,  394,  401,  408,  416,  424,  432,  440,  448,
+  456,  465,  474,  483,  492,  501,  510,  520,  530,  540,  550,  560,  571,
+  582,  593,  604,  615,  627,  639,  651,  663,  676,  689,  702,  715,  729,
+  743,  757,  771,  786,  801,  816,  832,  848,  864,  881,  898,  915,  933,
+  951,  969,  988,  1007, 1026, 1046, 1066, 1087, 1108, 1129, 1151, 1173, 1196,
+  1219, 1243, 1267, 1292, 1317, 1343, 1369, 1396, 1423, 1451, 1479, 1508, 1537,
+  1567, 1597, 1628, 1660, 1692, 1725, 1759, 1793, 1828,
+};
+
+#if CONFIG_HIGHBITDEPTH
+static const int16_t ac_qlookup_10[QINDEX_RANGE] = {
+  4,    9,    11,   13,   16,   18,   21,   24,   27,   30,   33,   37,   40,
+  44,   48,   51,   55,   59,   63,   67,   71,   75,   79,   83,   88,   92,
+  96,   100,  105,  109,  114,  118,  122,  127,  131,  136,  140,  145,  149,
+  154,  158,  163,  168,  172,  177,  181,  186,  190,  195,  199,  204,  208,
+  213,  217,  222,  226,  231,  235,  240,  244,  249,  253,  258,  262,  267,
+  271,  275,  280,  284,  289,  293,  297,  302,  306,  311,  315,  319,  324,
+  328,  332,  337,  341,  345,  349,  354,  358,  362,  367,  371,  375,  379,
+  384,  388,  392,  396,  401,  409,  417,  425,  433,  441,  449,  458,  466,
+  474,  482,  490,  498,  506,  514,  523,  531,  539,  547,  555,  563,  571,
+  579,  588,  596,  604,  616,  628,  640,  652,  664,  676,  688,  700,  713,
+  725,  737,  749,  761,  773,  785,  797,  809,  825,  841,  857,  873,  889,
+  905,  922,  938,  954,  970,  986,  1002, 1018, 1038, 1058, 1078, 1098, 1118,
+  1138, 1158, 1178, 1198, 1218, 1242, 1266, 1290, 1314, 1338, 1362, 1386, 1411,
+  1435, 1463, 1491, 1519, 1547, 1575, 1603, 1631, 1663, 1695, 1727, 1759, 1791,
+  1823, 1859, 1895, 1931, 1967, 2003, 2039, 2079, 2119, 2159, 2199, 2239, 2283,
+  2327, 2371, 2415, 2459, 2507, 2555, 2603, 2651, 2703, 2755, 2807, 2859, 2915,
+  2971, 3027, 3083, 3143, 3203, 3263, 3327, 3391, 3455, 3523, 3591, 3659, 3731,
+  3803, 3876, 3952, 4028, 4104, 4184, 4264, 4348, 4432, 4516, 4604, 4692, 4784,
+  4876, 4972, 5068, 5168, 5268, 5372, 5476, 5584, 5692, 5804, 5916, 6032, 6148,
+  6268, 6388, 6512, 6640, 6768, 6900, 7036, 7172, 7312,
+};
+
+static const int16_t ac_qlookup_12[QINDEX_RANGE] = {
+  4,     13,    19,    27,    35,    44,    54,    64,    75,    87,    99,
+  112,   126,   139,   154,   168,   183,   199,   214,   230,   247,   263,
+  280,   297,   314,   331,   349,   366,   384,   402,   420,   438,   456,
+  475,   493,   511,   530,   548,   567,   586,   604,   623,   642,   660,
+  679,   698,   716,   735,   753,   772,   791,   809,   828,   846,   865,
+  884,   902,   920,   939,   957,   976,   994,   1012,  1030,  1049,  1067,
+  1085,  1103,  1121,  1139,  1157,  1175,  1193,  1211,  1229,  1246,  1264,
+  1282,  1299,  1317,  1335,  1352,  1370,  1387,  1405,  1422,  1440,  1457,
+  1474,  1491,  1509,  1526,  1543,  1560,  1577,  1595,  1627,  1660,  1693,
+  1725,  1758,  1791,  1824,  1856,  1889,  1922,  1954,  1987,  2020,  2052,
+  2085,  2118,  2150,  2183,  2216,  2248,  2281,  2313,  2346,  2378,  2411,
+  2459,  2508,  2556,  2605,  2653,  2701,  2750,  2798,  2847,  2895,  2943,
+  2992,  3040,  3088,  3137,  3185,  3234,  3298,  3362,  3426,  3491,  3555,
+  3619,  3684,  3748,  3812,  3876,  3941,  4005,  4069,  4149,  4230,  4310,
+  4390,  4470,  4550,  4631,  4711,  4791,  4871,  4967,  5064,  5160,  5256,
+  5352,  5448,  5544,  5641,  5737,  5849,  5961,  6073,  6185,  6297,  6410,
+  6522,  6650,  6778,  6906,  7034,  7162,  7290,  7435,  7579,  7723,  7867,
+  8011,  8155,  8315,  8475,  8635,  8795,  8956,  9132,  9308,  9484,  9660,
+  9836,  10028, 10220, 10412, 10604, 10812, 11020, 11228, 11437, 11661, 11885,
+  12109, 12333, 12573, 12813, 13053, 13309, 13565, 13821, 14093, 14365, 14637,
+  14925, 15213, 15502, 15806, 16110, 16414, 16734, 17054, 17390, 17726, 18062,
+  18414, 18766, 19134, 19502, 19886, 20270, 20670, 21070, 21486, 21902, 22334,
+  22766, 23214, 23662, 24126, 24590, 25070, 25551, 26047, 26559, 27071, 27599,
+  28143, 28687, 29247,
+};
+#endif
+
+int16_t av1_dc_quant(int qindex, int delta, aom_bit_depth_t bit_depth) {
+#if CONFIG_HIGHBITDEPTH
+  switch (bit_depth) {
+    case AOM_BITS_8: return dc_qlookup[clamp(qindex + delta, 0, MAXQ)];
+    case AOM_BITS_10: return dc_qlookup_10[clamp(qindex + delta, 0, MAXQ)];
+    case AOM_BITS_12: return dc_qlookup_12[clamp(qindex + delta, 0, MAXQ)];
+    default:
+      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
+      return -1;
+  }
+#else
+  (void)bit_depth;
+  return dc_qlookup[clamp(qindex + delta, 0, MAXQ)];
+#endif
+}
+
+int16_t av1_ac_quant(int qindex, int delta, aom_bit_depth_t bit_depth) {
+#if CONFIG_HIGHBITDEPTH
+  switch (bit_depth) {
+    case AOM_BITS_8: return ac_qlookup[clamp(qindex + delta, 0, MAXQ)];
+    case AOM_BITS_10: return ac_qlookup_10[clamp(qindex + delta, 0, MAXQ)];
+    case AOM_BITS_12: return ac_qlookup_12[clamp(qindex + delta, 0, MAXQ)];
+    default:
+      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
+      return -1;
+  }
+#else
+  (void)bit_depth;
+  return ac_qlookup[clamp(qindex + delta, 0, MAXQ)];
+#endif
+}
+
+int16_t av1_qindex_from_ac(int ac, aom_bit_depth_t bit_depth) {
+  int i;
+  const int16_t *tab = ac_qlookup;
+  ac *= 4;
+#if CONFIG_HIGHBITDEPTH
+  switch (bit_depth) {
+    case AOM_BITS_10: {
+      tab = ac_qlookup_10;
+      ac *= 4;
+      break;
+    }
+    case AOM_BITS_12: {
+      tab = ac_qlookup_12;
+      ac *= 16;
+      break;
+    }
+    default:
+      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
+      return -1;
+  }
+#endif
+  (void)bit_depth;
+  for (i = 0; i < QINDEX_RANGE; i++) {
+    if (ac <= tab[i]) return i;
+  }
+  return QINDEX_RANGE - 1;
+}
+
+int av1_get_qindex(const struct segmentation *seg, int segment_id,
+                   int base_qindex) {
+  if (segfeature_active(seg, segment_id, SEG_LVL_ALT_Q)) {
+    const int data = get_segdata(seg, segment_id, SEG_LVL_ALT_Q);
+    const int seg_qindex =
+        seg->abs_delta == SEGMENT_ABSDATA ? data : base_qindex + data;
+    return clamp(seg_qindex, 0, MAXQ);
+  } else {
+    return base_qindex;
+  }
+}
+
+#if CONFIG_AOM_QM
+qm_val_t *aom_iqmatrix(AV1_COMMON *cm, int qmlevel, int is_chroma,
+                       int log2sizem2, int is_intra) {
+  return &cm->giqmatrix[qmlevel][!!is_chroma][!!is_intra][log2sizem2][0];
+}
+qm_val_t *aom_qmatrix(AV1_COMMON *cm, int qmlevel, int is_chroma,
+                      int log2sizem2, int is_intra) {
+  return &cm->gqmatrix[qmlevel][!!is_chroma][!!is_intra][log2sizem2][0];
+}
+
+static uint16_t iwt_matrix_ref[NUM_QM_LEVELS][2][2]
+                              [4 * 4 + 8 * 8 + 16 * 16 + 32 * 32];
+static uint16_t wt_matrix_ref[NUM_QM_LEVELS][2][2]
+                             [4 * 4 + 8 * 8 + 16 * 16 + 32 * 32];
+
+void aom_qm_init(AV1_COMMON *cm) {
+  int q, c, f, t, size;
+  int current;
+  for (q = 0; q < NUM_QM_LEVELS; ++q) {
+    for (c = 0; c < 2; ++c) {
+      for (f = 0; f < 2; ++f) {
+        current = 0;
+        for (t = 0; t < TX_SIZES; ++t) {
+          size = 1 << (t + 2);
+          cm->gqmatrix[q][c][f][t] = &wt_matrix_ref[q][c][f][current];
+          cm->giqmatrix[q][c][f][t] = &iwt_matrix_ref[q][c][f][current];
+          current += size * size;
+        }
+      }
+    }
+  }
+}
+
+static uint16_t iwt_matrix_ref[NUM_QM_LEVELS][2][2][4 * 4 + 8 * 8 + 16 * 16 +
+                                                    32 * 32] = {
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 71, 124, 214, 71, 112, 165, 241, 124, 165, 254, 331, 214, 241, 331,
+        414,
+        /* Size 8 */
+        64, 47, 51, 69, 97, 132, 173, 218, 47, 54, 52, 62, 81, 109, 142, 181,
+        51, 52, 75, 90, 108, 133, 165, 201, 69, 62, 90, 119, 144, 169, 198, 232,
+        97, 81, 108, 144, 178, 208, 238, 268, 132, 109, 133, 169, 208, 244, 276,
+        305, 173, 142, 165, 198, 238, 276, 309, 338, 218, 181, 201, 232, 268,
+        305, 338, 367,
+        /* Size 16 */
+        64, 54, 47, 49, 51, 59, 69, 81, 97, 111, 132, 150, 173, 193, 218, 218,
+        54, 52, 50, 51, 51, 58, 65, 75, 88, 101, 119, 135, 156, 175, 198, 198,
+        47, 50, 54, 53, 52, 56, 62, 70, 81, 93, 109, 123, 142, 159, 181, 181,
+        49, 51, 53, 57, 61, 67, 73, 82, 93, 104, 120, 134, 153, 170, 191, 191,
+        51, 51, 52, 61, 75, 82, 90, 98, 108, 119, 133, 147, 165, 181, 201, 201,
+        59, 58, 56, 67, 82, 91, 102, 112, 123, 135, 149, 163, 180, 196, 215,
+        215, 69, 65, 62, 73, 90, 102, 119, 130, 144, 155, 169, 182, 198, 214,
+        232, 232, 81, 75, 70, 82, 98, 112, 130, 143, 159, 172, 186, 200, 216,
+        231, 249, 249, 97, 88, 81, 93, 108, 123, 144, 159, 178, 192, 208, 222,
+        238, 252, 268, 268, 111, 101, 93, 104, 119, 135, 155, 172, 192, 207,
+        225, 239, 255, 269, 285, 285, 132, 119, 109, 120, 133, 149, 169, 186,
+        208, 225, 244, 259, 276, 290, 305, 305, 150, 135, 123, 134, 147, 163,
+        182, 200, 222, 239, 259, 274, 291, 305, 321, 321, 173, 156, 142, 153,
+        165, 180, 198, 216, 238, 255, 276, 291, 309, 323, 338, 338, 193, 175,
+        159, 170, 181, 196, 214, 231, 252, 269, 290, 305, 323, 337, 352, 352,
+        218, 198, 181, 191, 201, 215, 232, 249, 268, 285, 305, 321, 338, 352,
+        367, 367, 218, 198, 181, 191, 201, 215, 232, 249, 268, 285, 305, 321,
+        338, 352, 367, 367,
+        /* Size 32 */
+        64, 59, 54, 50, 47, 48, 49, 50, 51, 55, 59, 63, 69, 74, 81, 88, 97, 104,
+        111, 121, 132, 140, 150, 161, 173, 183, 193, 205, 218, 218, 218, 218,
+        59, 56, 53, 51, 49, 49, 50, 51, 51, 54, 58, 62, 67, 72, 78, 84, 92, 99,
+        106, 115, 125, 133, 142, 152, 164, 173, 183, 195, 208, 208, 208, 208,
+        54, 53, 52, 51, 50, 51, 51, 51, 51, 54, 58, 61, 65, 70, 75, 81, 88, 94,
+        101, 110, 119, 127, 135, 145, 156, 165, 175, 186, 198, 198, 198, 198,
+        50, 51, 51, 52, 52, 52, 52, 52, 52, 54, 57, 60, 63, 68, 72, 78, 85, 90,
+        97, 105, 114, 121, 129, 138, 149, 157, 167, 177, 189, 189, 189, 189, 47,
+        49, 50, 52, 54, 54, 53, 52, 52, 54, 56, 59, 62, 66, 70, 75, 81, 87, 93,
+        100, 109, 115, 123, 132, 142, 150, 159, 170, 181, 181, 181, 181, 48, 49,
+        51, 52, 54, 54, 55, 56, 56, 59, 61, 64, 67, 71, 76, 81, 87, 92, 98, 105,
+        114, 121, 128, 137, 147, 155, 164, 174, 186, 186, 186, 186, 49, 50, 51,
+        52, 53, 55, 57, 59, 61, 64, 67, 70, 73, 77, 82, 87, 93, 98, 104, 111,
+        120, 126, 134, 143, 153, 161, 170, 179, 191, 191, 191, 191, 50, 51, 51,
+        52, 52, 56, 59, 63, 68, 71, 74, 77, 81, 85, 89, 94, 100, 105, 111, 118,
+        126, 133, 140, 149, 158, 166, 175, 185, 196, 196, 196, 196, 51, 51, 51,
+        52, 52, 56, 61, 68, 75, 79, 82, 86, 90, 94, 98, 103, 108, 113, 119, 126,
+        133, 140, 147, 155, 165, 172, 181, 191, 201, 201, 201, 201, 55, 54, 54,
+        54, 54, 59, 64, 71, 79, 82, 86, 91, 96, 100, 105, 110, 115, 120, 126,
+        133, 140, 147, 155, 163, 172, 180, 188, 198, 208, 208, 208, 208, 59, 58,
+        58, 57, 56, 61, 67, 74, 82, 86, 91, 96, 102, 107, 112, 117, 123, 129,
+        135, 141, 149, 156, 163, 171, 180, 188, 196, 205, 215, 215, 215, 215,
+        63, 62, 61, 60, 59, 64, 70, 77, 86, 91, 96, 103, 110, 115, 120, 126,
+        133, 138, 144, 151, 158, 165, 172, 180, 189, 196, 204, 213, 223, 223,
+        223, 223, 69, 67, 65, 63, 62, 67, 73, 81, 90, 96, 102, 110, 119, 124,
+        130, 137, 144, 149, 155, 162, 169, 175, 182, 190, 198, 206, 214, 222,
+        232, 232, 232, 232, 74, 72, 70, 68, 66, 71, 77, 85, 94, 100, 107, 115,
+        124, 130, 136, 143, 151, 157, 163, 170, 177, 184, 191, 199, 207, 214,
+        222, 231, 240, 240, 240, 240, 81, 78, 75, 72, 70, 76, 82, 89, 98, 105,
+        112, 120, 130, 136, 143, 151, 159, 165, 172, 179, 186, 193, 200, 208,
+        216, 223, 231, 240, 249, 249, 249, 249, 88, 84, 81, 78, 75, 81, 87, 94,
+        103, 110, 117, 126, 137, 143, 151, 159, 168, 174, 181, 189, 197, 203,
+        211, 218, 226, 234, 241, 249, 258, 258, 258, 258, 97, 92, 88, 85, 81,
+        87, 93, 100, 108, 115, 123, 133, 144, 151, 159, 168, 178, 184, 192, 200,
+        208, 215, 222, 229, 238, 245, 252, 260, 268, 268, 268, 268, 104, 99, 94,
+        90, 87, 92, 98, 105, 113, 120, 129, 138, 149, 157, 165, 174, 184, 191,
+        199, 207, 216, 223, 230, 238, 246, 253, 260, 268, 276, 276, 276, 276,
+        111, 106, 101, 97, 93, 98, 104, 111, 119, 126, 135, 144, 155, 163, 172,
+        181, 192, 199, 207, 215, 225, 232, 239, 247, 255, 262, 269, 277, 285,
+        285, 285, 285, 121, 115, 110, 105, 100, 105, 111, 118, 126, 133, 141,
+        151, 162, 170, 179, 189, 200, 207, 215, 224, 234, 241, 248, 256, 265,
+        272, 279, 287, 295, 295, 295, 295, 132, 125, 119, 114, 109, 114, 120,
+        126, 133, 140, 149, 158, 169, 177, 186, 197, 208, 216, 225, 234, 244,
+        251, 259, 267, 276, 282, 290, 297, 305, 305, 305, 305, 140, 133, 127,
+        121, 115, 121, 126, 133, 140, 147, 156, 165, 175, 184, 193, 203, 215,
+        223, 232, 241, 251, 258, 266, 275, 283, 290, 297, 305, 313, 313, 313,
+        313, 150, 142, 135, 129, 123, 128, 134, 140, 147, 155, 163, 172, 182,
+        191, 200, 211, 222, 230, 239, 248, 259, 266, 274, 283, 291, 298, 305,
+        313, 321, 321, 321, 321, 161, 152, 145, 138, 132, 137, 143, 149, 155,
+        163, 171, 180, 190, 199, 208, 218, 229, 238, 247, 256, 267, 275, 283,
+        291, 300, 307, 314, 322, 329, 329, 329, 329, 173, 164, 156, 149, 142,
+        147, 153, 158, 165, 172, 180, 189, 198, 207, 216, 226, 238, 246, 255,
+        265, 276, 283, 291, 300, 309, 316, 323, 331, 338, 338, 338, 338, 183,
+        173, 165, 157, 150, 155, 161, 166, 172, 180, 188, 196, 206, 214, 223,
+        234, 245, 253, 262, 272, 282, 290, 298, 307, 316, 323, 330, 337, 345,
+        345, 345, 345, 193, 183, 175, 167, 159, 164, 170, 175, 181, 188, 196,
+        204, 214, 222, 231, 241, 252, 260, 269, 279, 290, 297, 305, 314, 323,
+        330, 337, 345, 352, 352, 352, 352, 205, 195, 186, 177, 170, 174, 179,
+        185, 191, 198, 205, 213, 222, 231, 240, 249, 260, 268, 277, 287, 297,
+        305, 313, 322, 331, 337, 345, 352, 360, 360, 360, 360, 218, 208, 198,
+        189, 181, 186, 191, 196, 201, 208, 215, 223, 232, 240, 249, 258, 268,
+        276, 285, 295, 305, 313, 321, 329, 338, 345, 352, 360, 367, 367, 367,
+        367, 218, 208, 198, 189, 181, 186, 191, 196, 201, 208, 215, 223, 232,
+        240, 249, 258, 268, 276, 285, 295, 305, 313, 321, 329, 338, 345, 352,
+        360, 367, 367, 367, 367, 218, 208, 198, 189, 181, 186, 191, 196, 201,
+        208, 215, 223, 232, 240, 249, 258, 268, 276, 285, 295, 305, 313, 321,
+        329, 338, 345, 352, 360, 367, 367, 367, 367, 218, 208, 198, 189, 181,
+        186, 191, 196, 201, 208, 215, 223, 232, 240, 249, 258, 268, 276, 285,
+        295, 305, 313, 321, 329, 338, 345, 352, 360, 367, 367, 367, 367 },
+      { /* Intra matrices */
+        /* Size 4 */
+        16, 18, 33, 60, 18, 29, 45, 68, 33, 45, 72, 98, 60, 68, 98, 129,
+        /* Size 8 */
+        20, 14, 16, 21, 31, 43, 58, 75, 14, 17, 16, 19, 25, 35, 46, 61, 16, 16,
+        24, 28, 34, 43, 54, 68, 21, 19, 28, 38, 47, 56, 67, 80, 31, 25, 34, 47,
+        59, 71, 83, 95, 43, 35, 43, 56, 71, 85, 99, 112, 58, 46, 54, 67, 83, 99,
+        113, 127, 75, 61, 68, 80, 95, 112, 127, 141,
+        /* Size 16 */
+        19, 16, 14, 14, 15, 17, 20, 24, 29, 34, 41, 47, 55, 62, 71, 71, 16, 15,
+        15, 15, 15, 17, 19, 22, 26, 31, 36, 42, 49, 55, 64, 64, 14, 15, 16, 16,
+        15, 17, 18, 21, 24, 28, 33, 38, 44, 50, 58, 58, 14, 15, 16, 17, 18, 20,
+        22, 24, 28, 32, 37, 41, 48, 54, 61, 61, 15, 15, 15, 18, 22, 24, 27, 30,
+        33, 36, 41, 46, 52, 58, 65, 65, 17, 17, 17, 20, 24, 27, 31, 34, 38, 42,
+        46, 51, 57, 63, 70, 70, 20, 19, 18, 22, 27, 31, 36, 40, 45, 49, 53, 58,
+        64, 70, 76, 76, 24, 22, 21, 24, 30, 34, 40, 44, 50, 54, 60, 65, 71, 76,
+        83, 83, 29, 26, 24, 28, 33, 38, 45, 50, 56, 61, 67, 73, 79, 84, 91, 91,
+        34, 31, 28, 32, 36, 42, 49, 54, 61, 67, 74, 79, 86, 91, 98, 98, 41, 36,
+        33, 37, 41, 46, 53, 60, 67, 74, 81, 87, 94, 100, 106, 106, 47, 42, 38,
+        41, 46, 51, 58, 65, 73, 79, 87, 93, 100, 106, 113, 113, 55, 49, 44, 48,
+        52, 57, 64, 71, 79, 86, 94, 100, 108, 114, 121, 121, 62, 55, 50, 54, 58,
+        63, 70, 76, 84, 91, 100, 106, 114, 120, 127, 127, 71, 64, 58, 61, 65,
+        70, 76, 83, 91, 98, 106, 113, 121, 127, 134, 134, 71, 64, 58, 61, 65,
+        70, 76, 83, 91, 98, 106, 113, 121, 127, 134, 134,
+        /* Size 32 */
+        18, 17, 15, 14, 13, 14, 14, 14, 15, 16, 17, 18, 20, 22, 23, 26, 28, 30,
+        33, 36, 40, 42, 45, 49, 53, 57, 60, 65, 69, 69, 69, 69, 17, 16, 15, 14,
+        14, 14, 14, 14, 15, 16, 17, 18, 19, 21, 23, 25, 27, 29, 31, 34, 37, 40,
+        43, 46, 50, 53, 57, 61, 66, 66, 66, 66, 15, 15, 15, 15, 14, 14, 14, 15,
+        15, 15, 16, 18, 19, 20, 22, 24, 26, 28, 30, 32, 35, 38, 41, 44, 48, 51,
+        54, 58, 62, 62, 62, 62, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 16, 17,
+        18, 19, 21, 23, 25, 26, 28, 31, 34, 36, 39, 42, 45, 48, 51, 55, 59, 59,
+        59, 59, 13, 14, 14, 15, 16, 15, 15, 15, 15, 15, 16, 17, 18, 19, 20, 22,
+        24, 25, 27, 29, 32, 34, 37, 40, 43, 46, 49, 52, 56, 56, 56, 56, 14, 14,
+        14, 15, 15, 16, 16, 16, 16, 17, 18, 18, 19, 20, 22, 23, 25, 27, 29, 31,
+        34, 36, 38, 41, 45, 47, 50, 54, 58, 58, 58, 58, 14, 14, 14, 15, 15, 16,
+        16, 17, 18, 18, 19, 20, 21, 22, 24, 25, 27, 29, 31, 33, 36, 38, 40, 43,
+        46, 49, 52, 56, 60, 60, 60, 60, 14, 14, 15, 15, 15, 16, 17, 18, 19, 20,
+        21, 22, 23, 25, 26, 28, 29, 31, 33, 35, 38, 40, 42, 45, 48, 51, 54, 57,
+        61, 61, 61, 61, 15, 15, 15, 15, 15, 16, 18, 19, 22, 23, 24, 25, 26, 27,
+        29, 30, 32, 34, 35, 38, 40, 42, 45, 47, 50, 53, 56, 59, 63, 63, 63, 63,
+        16, 16, 15, 15, 15, 17, 18, 20, 23, 24, 25, 27, 28, 29, 31, 32, 34, 36,
+        38, 40, 42, 45, 47, 50, 53, 56, 59, 62, 66, 66, 66, 66, 17, 17, 16, 16,
+        16, 18, 19, 21, 24, 25, 27, 28, 30, 32, 33, 35, 37, 39, 41, 43, 45, 47,
+        50, 53, 56, 58, 61, 65, 68, 68, 68, 68, 18, 18, 18, 17, 17, 18, 20, 22,
+        25, 27, 28, 30, 33, 34, 36, 38, 40, 42, 44, 46, 48, 51, 53, 56, 59, 62,
+        64, 68, 71, 71, 71, 71, 20, 19, 19, 18, 18, 19, 21, 23, 26, 28, 30, 33,
+        35, 37, 39, 41, 43, 45, 47, 50, 52, 54, 57, 59, 62, 65, 68, 71, 74, 74,
+        74, 74, 22, 21, 20, 19, 19, 20, 22, 25, 27, 29, 32, 34, 37, 39, 41, 43,
+        46, 48, 50, 52, 55, 57, 60, 62, 65, 68, 71, 74, 77, 77, 77, 77, 23, 23,
+        22, 21, 20, 22, 24, 26, 29, 31, 33, 36, 39, 41, 43, 46, 49, 51, 53, 55,
+        58, 60, 63, 66, 69, 71, 74, 77, 81, 81, 81, 81, 26, 25, 24, 23, 22, 23,
+        25, 28, 30, 32, 35, 38, 41, 43, 46, 48, 52, 54, 56, 59, 62, 64, 67, 69,
+        72, 75, 78, 81, 84, 84, 84, 84, 28, 27, 26, 25, 24, 25, 27, 29, 32, 34,
+        37, 40, 43, 46, 49, 52, 55, 57, 60, 63, 66, 68, 71, 74, 77, 79, 82, 85,
+        88, 88, 88, 88, 30, 29, 28, 26, 25, 27, 29, 31, 34, 36, 39, 42, 45, 48,
+        51, 54, 57, 60, 62, 65, 69, 71, 74, 77, 80, 83, 85, 88, 92, 92, 92, 92,
+        33, 31, 30, 28, 27, 29, 31, 33, 35, 38, 41, 44, 47, 50, 53, 56, 60, 62,
+        65, 68, 72, 74, 77, 80, 83, 86, 89, 92, 95, 95, 95, 95, 36, 34, 32, 31,
+        29, 31, 33, 35, 38, 40, 43, 46, 50, 52, 55, 59, 63, 65, 68, 72, 75, 78,
+        81, 84, 87, 90, 93, 96, 99, 99, 99, 99, 40, 37, 35, 34, 32, 34, 36, 38,
+        40, 42, 45, 48, 52, 55, 58, 62, 66, 69, 72, 75, 79, 82, 85, 88, 91, 94,
+        97, 100, 103, 103, 103, 103, 42, 40, 38, 36, 34, 36, 38, 40, 42, 45, 47,
+        51, 54, 57, 60, 64, 68, 71, 74, 78, 82, 85, 88, 91, 94, 97, 100, 103,
+        107, 107, 107, 107, 45, 43, 41, 39, 37, 38, 40, 42, 45, 47, 50, 53, 57,
+        60, 63, 67, 71, 74, 77, 81, 85, 88, 91, 94, 98, 101, 104, 107, 110, 110,
+        110, 110, 49, 46, 44, 42, 40, 41, 43, 45, 47, 50, 53, 56, 59, 62, 66,
+        69, 74, 77, 80, 84, 88, 91, 94, 98, 101, 104, 107, 110, 114, 114, 114,
+        114, 53, 50, 48, 45, 43, 45, 46, 48, 50, 53, 56, 59, 62, 65, 69, 72, 77,
+        80, 83, 87, 91, 94, 98, 101, 105, 108, 111, 114, 118, 118, 118, 118, 57,
+        53, 51, 48, 46, 47, 49, 51, 53, 56, 58, 62, 65, 68, 71, 75, 79, 83, 86,
+        90, 94, 97, 101, 104, 108, 111, 114, 117, 121, 121, 121, 121, 60, 57,
+        54, 51, 49, 50, 52, 54, 56, 59, 61, 64, 68, 71, 74, 78, 82, 85, 89, 93,
+        97, 100, 104, 107, 111, 114, 117, 120, 124, 124, 124, 124, 65, 61, 58,
+        55, 52, 54, 56, 57, 59, 62, 65, 68, 71, 74, 77, 81, 85, 88, 92, 96, 100,
+        103, 107, 110, 114, 117, 120, 124, 127, 127, 127, 127, 69, 66, 62, 59,
+        56, 58, 60, 61, 63, 66, 68, 71, 74, 77, 81, 84, 88, 92, 95, 99, 103,
+        107, 110, 114, 118, 121, 124, 127, 130, 130, 130, 130, 69, 66, 62, 59,
+        56, 58, 60, 61, 63, 66, 68, 71, 74, 77, 81, 84, 88, 92, 95, 99, 103,
+        107, 110, 114, 118, 121, 124, 127, 130, 130, 130, 130, 69, 66, 62, 59,
+        56, 58, 60, 61, 63, 66, 68, 71, 74, 77, 81, 84, 88, 92, 95, 99, 103,
+        107, 110, 114, 118, 121, 124, 127, 130, 130, 130, 130, 69, 66, 62, 59,
+        56, 58, 60, 61, 63, 66, 68, 71, 74, 77, 81, 84, 88, 92, 95, 99, 103,
+        107, 110, 114, 118, 121, 124, 127, 130, 130, 130, 130 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 106, 117, 154, 106, 131, 141, 167, 117, 141, 191, 225, 154, 167,
+        225, 279,
+        /* Size 8 */
+        64, 51, 98, 104, 113, 128, 148, 172, 51, 76, 100, 89, 92, 103, 118, 136,
+        98, 100, 119, 115, 114, 121, 134, 151, 104, 89, 115, 132, 140, 147, 158,
+        173, 113, 92, 114, 140, 160, 174, 186, 201, 128, 103, 121, 147, 174,
+        195, 213, 229, 148, 118, 134, 158, 186, 213, 236, 256, 172, 136, 151,
+        173, 201, 229, 256, 280,
+        /* Size 16 */
+        64, 57, 51, 67, 98, 101, 104, 108, 113, 120, 128, 137, 148, 159, 172,
+        172, 57, 59, 61, 75, 99, 97, 96, 99, 101, 107, 114, 122, 131, 141, 152,
+        152, 51, 61, 76, 86, 100, 94, 89, 91, 92, 97, 103, 110, 118, 126, 136,
+        136, 67, 75, 86, 96, 109, 104, 100, 101, 102, 106, 111, 118, 125, 134,
+        143, 143, 98, 99, 100, 109, 119, 117, 115, 115, 114, 118, 121, 127, 134,
+        142, 151, 151, 101, 97, 94, 104, 117, 120, 123, 124, 126, 129, 133, 139,
+        145, 153, 161, 161, 104, 96, 89, 100, 115, 123, 132, 136, 140, 144, 147,
+        153, 158, 165, 173, 173, 108, 99, 91, 101, 115, 124, 136, 142, 149, 154,
+        160, 165, 171, 178, 186, 186, 113, 101, 92, 102, 114, 126, 140, 149,
+        160, 167, 174, 180, 186, 193, 201, 201, 120, 107, 97, 106, 118, 129,
+        144, 154, 167, 175, 184, 191, 199, 206, 214, 214, 128, 114, 103, 111,
+        121, 133, 147, 160, 174, 184, 195, 204, 213, 221, 229, 229, 137, 122,
+        110, 118, 127, 139, 153, 165, 180, 191, 204, 213, 224, 233, 242, 242,
+        148, 131, 118, 125, 134, 145, 158, 171, 186, 199, 213, 224, 236, 246,
+        256, 256, 159, 141, 126, 134, 142, 153, 165, 178, 193, 206, 221, 233,
+        246, 256, 267, 267, 172, 152, 136, 143, 151, 161, 173, 186, 201, 214,
+        229, 242, 256, 267, 280, 280, 172, 152, 136, 143, 151, 161, 173, 186,
+        201, 214, 229, 242, 256, 267, 280, 280,
+        /* Size 32 */
+        64, 60, 57, 54, 51, 58, 67, 79, 98, 99, 101, 103, 104, 106, 108, 110,
+        113, 116, 120, 124, 128, 133, 137, 143, 148, 153, 159, 165, 172, 172,
+        172, 172, 60, 59, 58, 56, 55, 62, 71, 82, 98, 99, 99, 100, 100, 102,
+        103, 105, 107, 110, 113, 117, 121, 125, 129, 134, 139, 144, 149, 155,
+        161, 161, 161, 161, 57, 58, 59, 60, 61, 67, 75, 85, 99, 98, 97, 97, 96,
+        97, 99, 100, 101, 104, 107, 110, 114, 118, 122, 126, 131, 136, 141, 146,
+        152, 152, 152, 152, 54, 56, 60, 63, 67, 73, 80, 89, 99, 97, 96, 94, 93,
+        94, 94, 95, 97, 99, 102, 105, 108, 112, 115, 120, 124, 128, 133, 138,
+        144, 144, 144, 144, 51, 55, 61, 67, 76, 81, 86, 92, 100, 97, 94, 92, 89,
+        90, 91, 91, 92, 95, 97, 100, 103, 106, 110, 113, 118, 122, 126, 131,
+        136, 136, 136, 136, 58, 62, 67, 73, 81, 85, 91, 97, 104, 101, 99, 97,
+        94, 95, 96, 96, 97, 99, 101, 104, 107, 110, 113, 117, 121, 125, 130,
+        134, 140, 140, 140, 140, 67, 71, 75, 80, 86, 91, 96, 102, 109, 106, 104,
+        102, 100, 101, 101, 102, 102, 104, 106, 109, 111, 114, 118, 121, 125,
+        129, 134, 138, 143, 143, 143, 143, 79, 82, 85, 89, 92, 97, 102, 108,
+        114, 112, 110, 109, 107, 107, 107, 108, 108, 110, 112, 114, 116, 119,
+        122, 126, 129, 133, 138, 142, 147, 147, 147, 147, 98, 98, 99, 99, 100,
+        104, 109, 114, 119, 118, 117, 116, 115, 115, 115, 114, 114, 116, 118,
+        119, 121, 124, 127, 130, 134, 138, 142, 146, 151, 151, 151, 151, 99, 99,
+        98, 97, 97, 101, 106, 112, 118, 118, 118, 119, 119, 119, 119, 120, 120,
+        122, 123, 125, 127, 130, 133, 136, 139, 143, 147, 151, 156, 156, 156,
+        156, 101, 99, 97, 96, 94, 99, 104, 110, 117, 118, 120, 121, 123, 124,
+        124, 125, 126, 128, 129, 131, 133, 136, 139, 142, 145, 149, 153, 157,
+        161, 161, 161, 161, 103, 100, 97, 94, 92, 97, 102, 109, 116, 119, 121,
+        124, 127, 129, 130, 131, 133, 134, 136, 138, 140, 143, 145, 148, 151,
+        155, 159, 163, 167, 167, 167, 167, 104, 100, 96, 93, 89, 94, 100, 107,
+        115, 119, 123, 127, 132, 134, 136, 138, 140, 142, 144, 146, 147, 150,
+        153, 155, 158, 162, 165, 169, 173, 173, 173, 173, 106, 102, 97, 94, 90,
+        95, 101, 107, 115, 119, 124, 129, 134, 137, 139, 142, 145, 147, 149,
+        151, 153, 156, 159, 162, 164, 168, 172, 175, 179, 179, 179, 179, 108,
+        103, 99, 94, 91, 96, 101, 107, 115, 119, 124, 130, 136, 139, 142, 146,
+        149, 152, 154, 157, 160, 162, 165, 168, 171, 175, 178, 182, 186, 186,
+        186, 186, 110, 105, 100, 95, 91, 96, 102, 108, 114, 120, 125, 131, 138,
+        142, 146, 150, 154, 157, 160, 163, 166, 169, 172, 175, 178, 182, 185,
+        189, 193, 193, 193, 193, 113, 107, 101, 97, 92, 97, 102, 108, 114, 120,
+        126, 133, 140, 145, 149, 154, 160, 163, 167, 170, 174, 177, 180, 183,
+        186, 190, 193, 197, 201, 201, 201, 201, 116, 110, 104, 99, 95, 99, 104,
+        110, 116, 122, 128, 134, 142, 147, 152, 157, 163, 167, 171, 175, 179,
+        182, 185, 189, 192, 196, 199, 203, 207, 207, 207, 207, 120, 113, 107,
+        102, 97, 101, 106, 112, 118, 123, 129, 136, 144, 149, 154, 160, 167,
+        171, 175, 179, 184, 187, 191, 195, 199, 202, 206, 210, 214, 214, 214,
+        214, 124, 117, 110, 105, 100, 104, 109, 114, 119, 125, 131, 138, 146,
+        151, 157, 163, 170, 175, 179, 184, 190, 193, 197, 201, 206, 209, 213,
+        217, 221, 221, 221, 221, 128, 121, 114, 108, 103, 107, 111, 116, 121,
+        127, 133, 140, 147, 153, 160, 166, 174, 179, 184, 190, 195, 200, 204,
+        208, 213, 217, 221, 225, 229, 229, 229, 229, 133, 125, 118, 112, 106,
+        110, 114, 119, 124, 130, 136, 143, 150, 156, 162, 169, 177, 182, 187,
+        193, 200, 204, 209, 213, 218, 222, 227, 231, 235, 235, 235, 235, 137,
+        129, 122, 115, 110, 113, 118, 122, 127, 133, 139, 145, 153, 159, 165,
+        172, 180, 185, 191, 197, 204, 209, 213, 219, 224, 228, 233, 237, 242,
+        242, 242, 242, 143, 134, 126, 120, 113, 117, 121, 126, 130, 136, 142,
+        148, 155, 162, 168, 175, 183, 189, 195, 201, 208, 213, 219, 224, 230,
+        234, 239, 244, 249, 249, 249, 249, 148, 139, 131, 124, 118, 121, 125,
+        129, 134, 139, 145, 151, 158, 164, 171, 178, 186, 192, 199, 206, 213,
+        218, 224, 230, 236, 241, 246, 251, 256, 256, 256, 256, 153, 144, 136,
+        128, 122, 125, 129, 133, 138, 143, 149, 155, 162, 168, 175, 182, 190,
+        196, 202, 209, 217, 222, 228, 234, 241, 246, 251, 256, 262, 262, 262,
+        262, 159, 149, 141, 133, 126, 130, 134, 138, 142, 147, 153, 159, 165,
+        172, 178, 185, 193, 199, 206, 213, 221, 227, 233, 239, 246, 251, 256,
+        262, 267, 267, 267, 267, 165, 155, 146, 138, 131, 134, 138, 142, 146,
+        151, 157, 163, 169, 175, 182, 189, 197, 203, 210, 217, 225, 231, 237,
+        244, 251, 256, 262, 267, 273, 273, 273, 273, 172, 161, 152, 144, 136,
+        140, 143, 147, 151, 156, 161, 167, 173, 179, 186, 193, 201, 207, 214,
+        221, 229, 235, 242, 249, 256, 262, 267, 273, 280, 280, 280, 280, 172,
+        161, 152, 144, 136, 140, 143, 147, 151, 156, 161, 167, 173, 179, 186,
+        193, 201, 207, 214, 221, 229, 235, 242, 249, 256, 262, 267, 273, 280,
+        280, 280, 280, 172, 161, 152, 144, 136, 140, 143, 147, 151, 156, 161,
+        167, 173, 179, 186, 193, 201, 207, 214, 221, 229, 235, 242, 249, 256,
+        262, 267, 273, 280, 280, 280, 280, 172, 161, 152, 144, 136, 140, 143,
+        147, 151, 156, 161, 167, 173, 179, 186, 193, 201, 207, 214, 221, 229,
+        235, 242, 249, 256, 262, 267, 273, 280, 280, 280, 280 },
+      { /* Intra matrices */
+        /* Size 4 */
+        23, 40, 44, 59, 40, 50, 54, 64, 44, 54, 74, 89, 59, 64, 89, 114,
+        /* Size 8 */
+        25, 20, 39, 42, 46, 52, 61, 72, 20, 30, 40, 36, 37, 41, 48, 56, 39, 40,
+        49, 47, 46, 49, 55, 62, 42, 36, 47, 54, 58, 61, 66, 73, 46, 37, 46, 58,
+        67, 73, 79, 85, 52, 41, 49, 61, 73, 83, 91, 99, 61, 48, 55, 66, 79, 91,
+        103, 113, 72, 56, 62, 73, 85, 99, 113, 125,
+        /* Size 16 */
+        24, 22, 19, 26, 38, 39, 41, 42, 44, 47, 51, 55, 59, 64, 69, 69, 22, 22,
+        23, 29, 38, 38, 37, 38, 39, 42, 45, 48, 52, 56, 61, 61, 19, 23, 29, 33,
+        39, 37, 34, 35, 36, 38, 40, 43, 46, 50, 54, 54, 26, 29, 33, 37, 42, 41,
+        39, 39, 40, 42, 43, 46, 49, 53, 57, 57, 38, 38, 39, 42, 47, 46, 45, 45,
+        45, 46, 48, 50, 53, 56, 60, 60, 39, 38, 37, 41, 46, 47, 48, 49, 50, 51,
+        53, 55, 58, 61, 65, 65, 41, 37, 34, 39, 45, 48, 52, 54, 56, 57, 59, 61,
+        64, 67, 70, 70, 42, 38, 35, 39, 45, 49, 54, 57, 60, 62, 64, 67, 69, 72,
+        76, 76, 44, 39, 36, 40, 45, 50, 56, 60, 64, 67, 70, 73, 76, 79, 82, 82,
+        47, 42, 38, 42, 46, 51, 57, 62, 67, 71, 75, 78, 82, 85, 89, 89, 51, 45,
+        40, 43, 48, 53, 59, 64, 70, 75, 80, 84, 88, 92, 96, 96, 55, 48, 43, 46,
+        50, 55, 61, 67, 73, 78, 84, 88, 93, 97, 102, 102, 59, 52, 46, 49, 53,
+        58, 64, 69, 76, 82, 88, 93, 99, 104, 109, 109, 64, 56, 50, 53, 56, 61,
+        67, 72, 79, 85, 92, 97, 104, 109, 114, 114, 69, 61, 54, 57, 60, 65, 70,
+        76, 82, 89, 96, 102, 109, 114, 120, 120, 69, 61, 54, 57, 60, 65, 70, 76,
+        82, 89, 96, 102, 109, 114, 120, 120,
+        /* Size 32 */
+        24, 22, 21, 20, 19, 22, 25, 30, 37, 38, 39, 39, 40, 41, 42, 42, 43, 45,
+        46, 48, 50, 51, 54, 56, 58, 60, 63, 65, 68, 68, 68, 68, 22, 22, 22, 21,
+        21, 23, 27, 31, 37, 38, 38, 38, 38, 39, 40, 40, 41, 42, 44, 45, 47, 48,
+        50, 52, 54, 56, 59, 61, 64, 64, 64, 64, 21, 22, 22, 22, 23, 25, 28, 32,
+        38, 37, 37, 37, 37, 37, 38, 38, 39, 40, 41, 42, 44, 45, 47, 49, 51, 53,
+        55, 57, 60, 60, 60, 60, 20, 21, 22, 24, 25, 28, 30, 34, 38, 37, 36, 36,
+        35, 36, 36, 36, 37, 38, 39, 40, 41, 43, 44, 46, 48, 50, 52, 54, 56, 56,
+        56, 56, 19, 21, 23, 25, 29, 30, 33, 35, 38, 37, 36, 35, 34, 34, 34, 35,
+        35, 36, 37, 38, 39, 41, 42, 44, 45, 47, 49, 51, 53, 53, 53, 53, 22, 23,
+        25, 28, 30, 32, 34, 37, 40, 39, 38, 37, 36, 36, 36, 37, 37, 38, 39, 40,
+        41, 42, 44, 45, 47, 49, 50, 52, 54, 54, 54, 54, 25, 27, 28, 30, 33, 34,
+        37, 39, 42, 41, 40, 39, 38, 38, 39, 39, 39, 40, 41, 42, 43, 44, 45, 47,
+        48, 50, 52, 54, 56, 56, 56, 56, 30, 31, 32, 34, 35, 37, 39, 41, 44, 43,
+        42, 42, 41, 41, 41, 41, 41, 42, 43, 44, 45, 46, 47, 49, 50, 52, 54, 55,
+        57, 57, 57, 57, 37, 37, 38, 38, 38, 40, 42, 44, 46, 46, 45, 45, 44, 44,
+        44, 44, 44, 45, 45, 46, 47, 48, 49, 51, 52, 54, 55, 57, 59, 59, 59, 59,
+        38, 38, 37, 37, 37, 39, 41, 43, 46, 46, 46, 46, 46, 46, 46, 46, 46, 47,
+        48, 48, 49, 50, 52, 53, 54, 56, 58, 59, 61, 61, 61, 61, 39, 38, 37, 36,
+        36, 38, 40, 42, 45, 46, 46, 47, 48, 48, 48, 48, 49, 49, 50, 51, 52, 53,
+        54, 55, 57, 58, 60, 62, 64, 64, 64, 64, 39, 38, 37, 36, 35, 37, 39, 42,
+        45, 46, 47, 48, 49, 50, 50, 51, 52, 52, 53, 54, 55, 56, 57, 58, 59, 61,
+        63, 64, 66, 66, 66, 66, 40, 38, 37, 35, 34, 36, 38, 41, 44, 46, 48, 49,
+        51, 52, 53, 54, 55, 55, 56, 57, 58, 59, 60, 61, 62, 64, 65, 67, 69, 69,
+        69, 69, 41, 39, 37, 36, 34, 36, 38, 41, 44, 46, 48, 50, 52, 53, 54, 55,
+        57, 57, 58, 59, 60, 61, 63, 64, 65, 67, 68, 70, 71, 71, 71, 71, 42, 40,
+        38, 36, 34, 36, 39, 41, 44, 46, 48, 50, 53, 54, 56, 57, 59, 60, 61, 62,
+        63, 64, 65, 67, 68, 69, 71, 73, 74, 74, 74, 74, 42, 40, 38, 36, 35, 37,
+        39, 41, 44, 46, 48, 51, 54, 55, 57, 59, 61, 62, 63, 64, 66, 67, 68, 70,
+        71, 73, 74, 76, 78, 78, 78, 78, 43, 41, 39, 37, 35, 37, 39, 41, 44, 46,
+        49, 52, 55, 57, 59, 61, 63, 64, 66, 67, 69, 70, 72, 73, 75, 76, 78, 79,
+        81, 81, 81, 81, 45, 42, 40, 38, 36, 38, 40, 42, 45, 47, 49, 52, 55, 57,
+        60, 62, 64, 66, 68, 69, 71, 73, 74, 76, 77, 79, 80, 82, 84, 84, 84, 84,
+        46, 44, 41, 39, 37, 39, 41, 43, 45, 48, 50, 53, 56, 58, 61, 63, 66, 68,
+        70, 71, 74, 75, 77, 78, 80, 82, 83, 85, 87, 87, 87, 87, 48, 45, 42, 40,
+        38, 40, 42, 44, 46, 48, 51, 54, 57, 59, 62, 64, 67, 69, 71, 74, 76, 78,
+        79, 81, 83, 85, 87, 89, 90, 90, 90, 90, 50, 47, 44, 41, 39, 41, 43, 45,
+        47, 49, 52, 55, 58, 60, 63, 66, 69, 71, 74, 76, 79, 80, 82, 84, 87, 88,
+        90, 92, 94, 94, 94, 94, 51, 48, 45, 43, 41, 42, 44, 46, 48, 50, 53, 56,
+        59, 61, 64, 67, 70, 73, 75, 78, 80, 82, 85, 87, 89, 91, 93, 95, 97, 97,
+        97, 97, 54, 50, 47, 44, 42, 44, 45, 47, 49, 52, 54, 57, 60, 63, 65, 68,
+        72, 74, 77, 79, 82, 85, 87, 89, 92, 94, 96, 98, 100, 100, 100, 100, 56,
+        52, 49, 46, 44, 45, 47, 49, 51, 53, 55, 58, 61, 64, 67, 70, 73, 76, 78,
+        81, 84, 87, 89, 92, 94, 96, 99, 101, 103, 103, 103, 103, 58, 54, 51, 48,
+        45, 47, 48, 50, 52, 54, 57, 59, 62, 65, 68, 71, 75, 77, 80, 83, 87, 89,
+        92, 94, 97, 99, 102, 104, 107, 107, 107, 107, 60, 56, 53, 50, 47, 49,
+        50, 52, 54, 56, 58, 61, 64, 67, 69, 73, 76, 79, 82, 85, 88, 91, 94, 96,
+        99, 102, 104, 107, 109, 109, 109, 109, 63, 59, 55, 52, 49, 50, 52, 54,
+        55, 58, 60, 63, 65, 68, 71, 74, 78, 80, 83, 87, 90, 93, 96, 99, 102,
+        104, 107, 109, 112, 112, 112, 112, 65, 61, 57, 54, 51, 52, 54, 55, 57,
+        59, 62, 64, 67, 70, 73, 76, 79, 82, 85, 89, 92, 95, 98, 101, 104, 107,
+        109, 112, 115, 115, 115, 115, 68, 64, 60, 56, 53, 54, 56, 57, 59, 61,
+        64, 66, 69, 71, 74, 78, 81, 84, 87, 90, 94, 97, 100, 103, 107, 109, 112,
+        115, 118, 118, 118, 118, 68, 64, 60, 56, 53, 54, 56, 57, 59, 61, 64, 66,
+        69, 71, 74, 78, 81, 84, 87, 90, 94, 97, 100, 103, 107, 109, 112, 115,
+        118, 118, 118, 118, 68, 64, 60, 56, 53, 54, 56, 57, 59, 61, 64, 66, 69,
+        71, 74, 78, 81, 84, 87, 90, 94, 97, 100, 103, 107, 109, 112, 115, 118,
+        118, 118, 118, 68, 64, 60, 56, 53, 54, 56, 57, 59, 61, 64, 66, 69, 71,
+        74, 78, 81, 84, 87, 90, 94, 97, 100, 103, 107, 109, 112, 115, 118, 118,
+        118, 118 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 70, 120, 197, 70, 109, 156, 218, 120, 156, 229, 287, 197, 218, 287,
+        344,
+        /* Size 8 */
+        64, 47, 51, 69, 94, 126, 161, 197, 47, 55, 52, 62, 80, 105, 135, 167,
+        51, 52, 75, 88, 105, 127, 154, 183, 69, 62, 88, 115, 136, 157, 181, 207,
+        94, 80, 105, 136, 165, 189, 212, 234, 126, 105, 127, 157, 189, 216, 240,
+        261, 161, 135, 154, 181, 212, 240, 264, 284, 197, 167, 183, 207, 234,
+        261, 284, 303,
+        /* Size 16 */
+        64, 54, 47, 49, 51, 59, 69, 80, 94, 108, 126, 141, 161, 177, 197, 197,
+        54, 53, 51, 51, 52, 58, 65, 74, 87, 99, 115, 129, 147, 162, 181, 181,
+        47, 51, 55, 53, 52, 57, 62, 70, 80, 91, 105, 118, 135, 149, 167, 167,
+        49, 51, 53, 57, 62, 67, 73, 81, 91, 102, 115, 128, 144, 158, 175, 175,
+        51, 52, 52, 62, 75, 81, 88, 96, 105, 115, 127, 139, 154, 167, 183, 183,
+        59, 58, 57, 67, 81, 89, 100, 108, 118, 128, 140, 152, 166, 179, 195,
+        195, 69, 65, 62, 73, 88, 100, 115, 124, 136, 146, 157, 168, 181, 193,
+        207, 207, 80, 74, 70, 81, 96, 108, 124, 136, 149, 159, 172, 183, 195,
+        207, 220, 220, 94, 87, 80, 91, 105, 118, 136, 149, 165, 176, 189, 200,
+        212, 222, 234, 234, 108, 99, 91, 102, 115, 128, 146, 159, 176, 188, 202,
+        213, 225, 235, 247, 247, 126, 115, 105, 115, 127, 140, 157, 172, 189,
+        202, 216, 228, 240, 250, 261, 261, 141, 129, 118, 128, 139, 152, 168,
+        183, 200, 213, 228, 239, 251, 261, 272, 272, 161, 147, 135, 144, 154,
+        166, 181, 195, 212, 225, 240, 251, 264, 273, 284, 284, 177, 162, 149,
+        158, 167, 179, 193, 207, 222, 235, 250, 261, 273, 283, 293, 293, 197,
+        181, 167, 175, 183, 195, 207, 220, 234, 247, 261, 272, 284, 293, 303,
+        303, 197, 181, 167, 175, 183, 195, 207, 220, 234, 247, 261, 272, 284,
+        293, 303, 303,
+        /* Size 32 */
+        64, 59, 54, 51, 47, 48, 49, 50, 51, 55, 59, 64, 69, 74, 80, 86, 94, 101,
+        108, 116, 126, 133, 141, 150, 161, 169, 177, 186, 197, 197, 197, 197,
+        59, 56, 54, 51, 49, 50, 50, 51, 52, 55, 58, 62, 67, 72, 77, 83, 90, 96,
+        103, 111, 120, 127, 135, 143, 153, 161, 169, 178, 189, 189, 189, 189,
+        54, 54, 53, 52, 51, 51, 51, 52, 52, 55, 58, 61, 65, 69, 74, 80, 87, 92,
+        99, 106, 115, 121, 129, 137, 147, 154, 162, 171, 181, 181, 181, 181, 51,
+        51, 52, 52, 53, 53, 52, 52, 52, 55, 57, 60, 63, 67, 72, 77, 83, 89, 95,
+        102, 110, 116, 123, 131, 141, 148, 155, 164, 174, 174, 174, 174, 47, 49,
+        51, 53, 55, 54, 53, 53, 52, 54, 57, 59, 62, 66, 70, 75, 80, 85, 91, 98,
+        105, 111, 118, 126, 135, 142, 149, 158, 167, 167, 167, 167, 48, 50, 51,
+        53, 54, 55, 55, 56, 57, 59, 61, 64, 67, 71, 75, 80, 85, 90, 96, 102,
+        110, 116, 123, 130, 139, 146, 153, 162, 171, 171, 171, 171, 49, 50, 51,
+        52, 53, 55, 57, 59, 62, 64, 67, 70, 73, 77, 81, 86, 91, 96, 102, 108,
+        115, 121, 128, 135, 144, 150, 158, 166, 175, 175, 175, 175, 50, 51, 52,
+        52, 53, 56, 59, 63, 68, 70, 73, 76, 80, 84, 88, 92, 97, 102, 108, 114,
+        121, 127, 133, 140, 148, 155, 162, 170, 179, 179, 179, 179, 51, 52, 52,
+        52, 52, 57, 62, 68, 75, 78, 81, 84, 88, 92, 96, 100, 105, 109, 115, 120,
+        127, 133, 139, 146, 154, 160, 167, 175, 183, 183, 183, 183, 55, 55, 55,
+        55, 54, 59, 64, 70, 78, 81, 85, 89, 94, 98, 102, 106, 111, 116, 121,
+        127, 133, 139, 145, 152, 160, 166, 173, 181, 189, 189, 189, 189, 59, 58,
+        58, 57, 57, 61, 67, 73, 81, 85, 89, 94, 100, 104, 108, 113, 118, 123,
+        128, 134, 140, 146, 152, 159, 166, 173, 179, 187, 195, 195, 195, 195,
+        64, 62, 61, 60, 59, 64, 70, 76, 84, 89, 94, 100, 107, 111, 116, 121,
+        126, 131, 137, 142, 148, 154, 160, 166, 173, 180, 186, 193, 201, 201,
+        201, 201, 69, 67, 65, 63, 62, 67, 73, 80, 88, 94, 100, 107, 115, 119,
+        124, 130, 136, 141, 146, 151, 157, 163, 168, 175, 181, 187, 193, 200,
+        207, 207, 207, 207, 74, 72, 69, 67, 66, 71, 77, 84, 92, 98, 104, 111,
+        119, 124, 130, 136, 142, 147, 152, 158, 164, 170, 175, 181, 188, 194,
+        200, 206, 213, 213, 213, 213, 80, 77, 74, 72, 70, 75, 81, 88, 96, 102,
+        108, 116, 124, 130, 136, 142, 149, 154, 159, 165, 172, 177, 183, 189,
+        195, 201, 207, 213, 220, 220, 220, 220, 86, 83, 80, 77, 75, 80, 86, 92,
+        100, 106, 113, 121, 130, 136, 142, 149, 156, 162, 167, 173, 180, 185,
+        191, 197, 203, 209, 214, 220, 227, 227, 227, 227, 94, 90, 87, 83, 80,
+        85, 91, 97, 105, 111, 118, 126, 136, 142, 149, 156, 165, 170, 176, 182,
+        189, 194, 200, 205, 212, 217, 222, 228, 234, 234, 234, 234, 101, 96, 92,
+        89, 85, 90, 96, 102, 109, 116, 123, 131, 141, 147, 154, 162, 170, 176,
+        182, 188, 195, 200, 206, 212, 218, 223, 229, 234, 240, 240, 240, 240,
+        108, 103, 99, 95, 91, 96, 102, 108, 115, 121, 128, 137, 146, 152, 159,
+        167, 176, 182, 188, 195, 202, 207, 213, 219, 225, 230, 235, 241, 247,
+        247, 247, 247, 116, 111, 106, 102, 98, 102, 108, 114, 120, 127, 134,
+        142, 151, 158, 165, 173, 182, 188, 195, 201, 209, 214, 220, 226, 232,
+        237, 242, 248, 254, 254, 254, 254, 126, 120, 115, 110, 105, 110, 115,
+        121, 127, 133, 140, 148, 157, 164, 172, 180, 189, 195, 202, 209, 216,
+        222, 228, 234, 240, 245, 250, 255, 261, 261, 261, 261, 133, 127, 121,
+        116, 111, 116, 121, 127, 133, 139, 146, 154, 163, 170, 177, 185, 194,
+        200, 207, 214, 222, 227, 233, 239, 245, 250, 255, 261, 266, 266, 266,
+        266, 141, 135, 129, 123, 118, 123, 128, 133, 139, 145, 152, 160, 168,
+        175, 183, 191, 200, 206, 213, 220, 228, 233, 239, 245, 251, 256, 261,
+        266, 272, 272, 272, 272, 150, 143, 137, 131, 126, 130, 135, 140, 146,
+        152, 159, 166, 175, 181, 189, 197, 205, 212, 219, 226, 234, 239, 245,
+        251, 257, 262, 267, 272, 278, 278, 278, 278, 161, 153, 147, 141, 135,
+        139, 144, 148, 154, 160, 166, 173, 181, 188, 195, 203, 212, 218, 225,
+        232, 240, 245, 251, 257, 264, 268, 273, 278, 284, 284, 284, 284, 169,
+        161, 154, 148, 142, 146, 150, 155, 160, 166, 173, 180, 187, 194, 201,
+        209, 217, 223, 230, 237, 245, 250, 256, 262, 268, 273, 278, 283, 288,
+        288, 288, 288, 177, 169, 162, 155, 149, 153, 158, 162, 167, 173, 179,
+        186, 193, 200, 207, 214, 222, 229, 235, 242, 250, 255, 261, 267, 273,
+        278, 283, 288, 293, 293, 293, 293, 186, 178, 171, 164, 158, 162, 166,
+        170, 175, 181, 187, 193, 200, 206, 213, 220, 228, 234, 241, 248, 255,
+        261, 266, 272, 278, 283, 288, 293, 298, 298, 298, 298, 197, 189, 181,
+        174, 167, 171, 175, 179, 183, 189, 195, 201, 207, 213, 220, 227, 234,
+        240, 247, 254, 261, 266, 272, 278, 284, 288, 293, 298, 303, 303, 303,
+        303, 197, 189, 181, 174, 167, 171, 175, 179, 183, 189, 195, 201, 207,
+        213, 220, 227, 234, 240, 247, 254, 261, 266, 272, 278, 284, 288, 293,
+        298, 303, 303, 303, 303, 197, 189, 181, 174, 167, 171, 175, 179, 183,
+        189, 195, 201, 207, 213, 220, 227, 234, 240, 247, 254, 261, 266, 272,
+        278, 284, 288, 293, 298, 303, 303, 303, 303, 197, 189, 181, 174, 167,
+        171, 175, 179, 183, 189, 195, 201, 207, 213, 220, 227, 234, 240, 247,
+        254, 261, 266, 272, 278, 284, 288, 293, 298, 303, 303, 303, 303 },
+      { /* Intra matrices */
+        /* Size 4 */
+        19, 21, 37, 63, 21, 33, 49, 70, 37, 49, 74, 96, 63, 70, 96, 119,
+        /* Size 8 */
+        23, 17, 18, 25, 34, 47, 61, 77, 17, 19, 18, 22, 29, 38, 50, 64, 18, 18,
+        27, 32, 38, 47, 58, 71, 25, 22, 32, 42, 51, 60, 70, 81, 34, 29, 38, 51,
+        63, 73, 83, 94, 47, 38, 47, 60, 73, 85, 96, 106, 61, 50, 58, 70, 83, 96,
+        108, 118, 77, 64, 71, 81, 94, 106, 118, 127,
+        /* Size 16 */
+        22, 18, 16, 17, 17, 20, 23, 27, 33, 38, 45, 51, 58, 65, 73, 73, 18, 18,
+        17, 17, 18, 20, 22, 25, 30, 34, 40, 46, 53, 59, 67, 67, 16, 17, 19, 18,
+        18, 19, 21, 24, 28, 32, 37, 42, 48, 54, 61, 61, 17, 17, 18, 19, 21, 23,
+        25, 28, 31, 35, 40, 45, 52, 57, 64, 64, 17, 18, 18, 21, 26, 28, 31, 33,
+        37, 40, 45, 50, 55, 61, 68, 68, 20, 20, 19, 23, 28, 31, 35, 38, 42, 46,
+        50, 55, 61, 66, 72, 72, 23, 22, 21, 25, 31, 35, 40, 44, 48, 52, 57, 61,
+        67, 72, 78, 78, 27, 25, 24, 28, 33, 38, 44, 48, 54, 58, 63, 67, 73, 78,
+        83, 83, 33, 30, 28, 31, 37, 42, 48, 54, 60, 64, 70, 74, 80, 84, 90, 90,
+        38, 34, 32, 35, 40, 46, 52, 58, 64, 69, 75, 80, 85, 90, 95, 95, 45, 40,
+        37, 40, 45, 50, 57, 63, 70, 75, 82, 87, 92, 97, 102, 102, 51, 46, 42,
+        45, 50, 55, 61, 67, 74, 80, 87, 92, 97, 102, 107, 107, 58, 53, 48, 52,
+        55, 61, 67, 73, 80, 85, 92, 97, 103, 107, 112, 112, 65, 59, 54, 57, 61,
+        66, 72, 78, 84, 90, 97, 102, 107, 112, 117, 117, 73, 67, 61, 64, 68, 72,
+        78, 83, 90, 95, 102, 107, 112, 117, 122, 122, 73, 67, 61, 64, 68, 72,
+        78, 83, 90, 95, 102, 107, 112, 117, 122, 122,
+        /* Size 32 */
+        21, 19, 18, 17, 16, 16, 16, 17, 17, 18, 20, 21, 23, 25, 27, 29, 32, 34,
+        37, 40, 44, 46, 49, 53, 57, 60, 63, 67, 72, 72, 72, 72, 19, 19, 18, 17,
+        16, 16, 17, 17, 17, 18, 19, 21, 22, 24, 26, 28, 31, 33, 35, 38, 41, 44,
+        47, 50, 54, 57, 60, 64, 68, 68, 68, 68, 18, 18, 17, 17, 17, 17, 17, 17,
+        17, 18, 19, 20, 22, 23, 25, 27, 29, 31, 34, 36, 39, 42, 45, 48, 52, 54,
+        57, 61, 65, 65, 65, 65, 17, 17, 17, 17, 17, 17, 17, 17, 17, 18, 19, 20,
+        21, 22, 24, 26, 28, 30, 32, 35, 38, 40, 43, 46, 49, 52, 55, 58, 62, 62,
+        62, 62, 16, 16, 17, 17, 18, 18, 18, 17, 17, 18, 19, 20, 21, 22, 23, 25,
+        27, 29, 31, 33, 36, 38, 41, 44, 47, 50, 53, 56, 60, 60, 60, 60, 16, 16,
+        17, 17, 18, 18, 18, 18, 19, 19, 20, 21, 22, 24, 25, 27, 29, 31, 33, 35,
+        38, 40, 42, 45, 49, 51, 54, 57, 61, 61, 61, 61, 16, 17, 17, 17, 18, 18,
+        19, 20, 20, 21, 22, 23, 24, 26, 27, 29, 31, 33, 35, 37, 40, 42, 44, 47,
+        50, 53, 56, 59, 63, 63, 63, 63, 17, 17, 17, 17, 17, 18, 20, 21, 22, 23,
+        24, 26, 27, 28, 30, 31, 33, 35, 37, 39, 42, 44, 46, 49, 52, 55, 58, 61,
+        64, 64, 64, 64, 17, 17, 17, 17, 17, 19, 20, 22, 25, 26, 27, 28, 30, 31,
+        32, 34, 36, 37, 39, 42, 44, 46, 49, 51, 54, 57, 60, 63, 66, 66, 66, 66,
+        18, 18, 18, 18, 18, 19, 21, 23, 26, 27, 29, 30, 32, 33, 35, 36, 38, 40,
+        42, 44, 46, 49, 51, 54, 57, 59, 62, 65, 68, 68, 68, 68, 20, 19, 19, 19,
+        19, 20, 22, 24, 27, 29, 30, 32, 34, 35, 37, 39, 41, 43, 45, 47, 49, 51,
+        54, 56, 59, 62, 64, 67, 71, 71, 71, 71, 21, 21, 20, 20, 20, 21, 23, 26,
+        28, 30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 50, 52, 54, 57, 59, 62, 64,
+        67, 70, 73, 73, 73, 73, 23, 22, 22, 21, 21, 22, 24, 27, 30, 32, 34, 36,
+        39, 41, 43, 45, 47, 49, 51, 53, 56, 58, 60, 62, 65, 68, 70, 73, 76, 76,
+        76, 76, 25, 24, 23, 22, 22, 24, 26, 28, 31, 33, 35, 38, 41, 43, 45, 47,
+        50, 52, 54, 56, 58, 60, 63, 65, 68, 70, 73, 76, 78, 78, 78, 78, 27, 26,
+        25, 24, 23, 25, 27, 30, 32, 35, 37, 40, 43, 45, 47, 50, 52, 54, 56, 59,
+        61, 63, 66, 68, 71, 73, 76, 78, 81, 81, 81, 81, 29, 28, 27, 26, 25, 27,
+        29, 31, 34, 36, 39, 42, 45, 47, 50, 52, 55, 57, 60, 62, 65, 67, 69, 72,
+        74, 76, 79, 82, 84, 84, 84, 84, 32, 31, 29, 28, 27, 29, 31, 33, 36, 38,
+        41, 44, 47, 50, 52, 55, 58, 61, 63, 66, 68, 70, 73, 75, 78, 80, 82, 85,
+        88, 88, 88, 88, 34, 33, 31, 30, 29, 31, 33, 35, 37, 40, 43, 46, 49, 52,
+        54, 57, 61, 63, 65, 68, 71, 73, 75, 78, 80, 83, 85, 88, 90, 90, 90, 90,
+        37, 35, 34, 32, 31, 33, 35, 37, 39, 42, 45, 48, 51, 54, 56, 60, 63, 65,
+        68, 71, 74, 76, 78, 81, 83, 86, 88, 90, 93, 93, 93, 93, 40, 38, 36, 35,
+        33, 35, 37, 39, 42, 44, 47, 50, 53, 56, 59, 62, 66, 68, 71, 73, 77, 79,
+        81, 84, 87, 89, 91, 93, 96, 96, 96, 96, 44, 41, 39, 38, 36, 38, 40, 42,
+        44, 46, 49, 52, 56, 58, 61, 65, 68, 71, 74, 77, 80, 82, 85, 87, 90, 92,
+        94, 97, 99, 99, 99, 99, 46, 44, 42, 40, 38, 40, 42, 44, 46, 49, 51, 54,
+        58, 60, 63, 67, 70, 73, 76, 79, 82, 84, 87, 90, 92, 95, 97, 99, 102,
+        102, 102, 102, 49, 47, 45, 43, 41, 42, 44, 46, 49, 51, 54, 57, 60, 63,
+        66, 69, 73, 75, 78, 81, 85, 87, 89, 92, 95, 97, 99, 102, 104, 104, 104,
+        104, 53, 50, 48, 46, 44, 45, 47, 49, 51, 54, 56, 59, 62, 65, 68, 72, 75,
+        78, 81, 84, 87, 90, 92, 95, 98, 100, 102, 105, 107, 107, 107, 107, 57,
+        54, 52, 49, 47, 49, 50, 52, 54, 57, 59, 62, 65, 68, 71, 74, 78, 80, 83,
+        87, 90, 92, 95, 98, 101, 103, 105, 107, 110, 110, 110, 110, 60, 57, 54,
+        52, 50, 51, 53, 55, 57, 59, 62, 64, 68, 70, 73, 76, 80, 83, 86, 89, 92,
+        95, 97, 100, 103, 105, 107, 110, 112, 112, 112, 112, 63, 60, 57, 55, 53,
+        54, 56, 58, 60, 62, 64, 67, 70, 73, 76, 79, 82, 85, 88, 91, 94, 97, 99,
+        102, 105, 107, 109, 112, 114, 114, 114, 114, 67, 64, 61, 58, 56, 57, 59,
+        61, 63, 65, 67, 70, 73, 76, 78, 82, 85, 88, 90, 93, 97, 99, 102, 105,
+        107, 110, 112, 114, 117, 117, 117, 117, 72, 68, 65, 62, 60, 61, 63, 64,
+        66, 68, 71, 73, 76, 78, 81, 84, 88, 90, 93, 96, 99, 102, 104, 107, 110,
+        112, 114, 117, 119, 119, 119, 119, 72, 68, 65, 62, 60, 61, 63, 64, 66,
+        68, 71, 73, 76, 78, 81, 84, 88, 90, 93, 96, 99, 102, 104, 107, 110, 112,
+        114, 117, 119, 119, 119, 119, 72, 68, 65, 62, 60, 61, 63, 64, 66, 68,
+        71, 73, 76, 78, 81, 84, 88, 90, 93, 96, 99, 102, 104, 107, 110, 112,
+        114, 117, 119, 119, 119, 119, 72, 68, 65, 62, 60, 61, 63, 64, 66, 68,
+        71, 73, 76, 78, 81, 84, 88, 90, 93, 96, 99, 102, 104, 107, 110, 112,
+        114, 117, 119, 119, 119, 119 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 104, 114, 146, 104, 126, 136, 158, 114, 136, 178, 207, 146, 158,
+        207, 250,
+        /* Size 8 */
+        64, 51, 96, 102, 109, 123, 141, 161, 51, 75, 98, 88, 91, 100, 114, 131,
+        96, 98, 116, 111, 111, 117, 128, 143, 102, 88, 111, 127, 134, 140, 150,
+        163, 109, 91, 111, 134, 151, 163, 174, 186, 123, 100, 117, 140, 163,
+        181, 196, 209, 141, 114, 128, 150, 174, 196, 214, 230, 161, 131, 143,
+        163, 186, 209, 230, 248,
+        /* Size 16 */
+        64, 57, 51, 67, 96, 99, 102, 106, 109, 116, 123, 132, 141, 151, 161,
+        161, 57, 59, 61, 75, 97, 96, 94, 97, 99, 105, 111, 118, 126, 135, 144,
+        144, 51, 61, 75, 85, 98, 92, 88, 89, 91, 95, 100, 107, 114, 122, 131,
+        131, 67, 75, 85, 94, 106, 102, 98, 99, 100, 104, 108, 114, 121, 128,
+        137, 137, 96, 97, 98, 106, 116, 113, 111, 111, 111, 114, 117, 123, 128,
+        135, 143, 143, 99, 96, 92, 102, 113, 116, 119, 120, 121, 125, 128, 133,
+        138, 145, 152, 152, 102, 94, 88, 98, 111, 119, 127, 131, 134, 137, 140,
+        145, 150, 156, 163, 163, 106, 97, 89, 99, 111, 120, 131, 136, 142, 146,
+        151, 156, 161, 167, 173, 173, 109, 99, 91, 100, 111, 121, 134, 142, 151,
+        157, 163, 168, 174, 179, 186, 186, 116, 105, 95, 104, 114, 125, 137,
+        146, 157, 164, 172, 178, 184, 190, 196, 196, 123, 111, 100, 108, 117,
+        128, 140, 151, 163, 172, 181, 188, 196, 202, 209, 209, 132, 118, 107,
+        114, 123, 133, 145, 156, 168, 178, 188, 196, 205, 211, 219, 219, 141,
+        126, 114, 121, 128, 138, 150, 161, 174, 184, 196, 205, 214, 222, 230,
+        230, 151, 135, 122, 128, 135, 145, 156, 167, 179, 190, 202, 211, 222,
+        230, 238, 238, 161, 144, 131, 137, 143, 152, 163, 173, 186, 196, 209,
+        219, 230, 238, 248, 248, 161, 144, 131, 137, 143, 152, 163, 173, 186,
+        196, 209, 219, 230, 238, 248, 248,
+        /* Size 32 */
+        64, 60, 57, 54, 51, 58, 67, 79, 96, 97, 99, 100, 102, 104, 106, 107,
+        109, 113, 116, 120, 123, 127, 132, 136, 141, 146, 151, 156, 161, 161,
+        161, 161, 60, 59, 58, 57, 56, 62, 70, 81, 96, 97, 97, 98, 98, 99, 101,
+        102, 104, 107, 110, 113, 117, 120, 124, 129, 133, 137, 142, 147, 152,
+        152, 152, 152, 57, 58, 59, 60, 61, 67, 75, 84, 97, 96, 96, 95, 94, 96,
+        97, 98, 99, 102, 105, 107, 111, 114, 118, 122, 126, 130, 135, 139, 144,
+        144, 144, 144, 54, 57, 60, 63, 67, 73, 79, 87, 97, 96, 94, 92, 91, 92,
+        93, 94, 95, 97, 100, 102, 105, 108, 112, 116, 120, 124, 128, 132, 137,
+        137, 137, 137, 51, 56, 61, 67, 75, 80, 85, 91, 98, 95, 92, 90, 88, 89,
+        89, 90, 91, 93, 95, 98, 100, 103, 107, 110, 114, 118, 122, 126, 131,
+        131, 131, 131, 58, 62, 67, 73, 80, 84, 89, 95, 102, 99, 97, 95, 93, 93,
+        94, 94, 95, 97, 99, 102, 104, 107, 110, 114, 117, 121, 125, 129, 134,
+        134, 134, 134, 67, 70, 75, 79, 85, 89, 94, 100, 106, 104, 102, 100, 98,
+        99, 99, 99, 100, 102, 104, 106, 108, 111, 114, 117, 121, 124, 128, 132,
+        137, 137, 137, 137, 79, 81, 84, 87, 91, 95, 100, 105, 110, 109, 107,
+        106, 104, 105, 105, 105, 105, 107, 109, 111, 112, 115, 118, 121, 125,
+        128, 132, 136, 140, 140, 140, 140, 96, 96, 97, 97, 98, 102, 106, 110,
+        116, 114, 113, 112, 111, 111, 111, 111, 111, 112, 114, 116, 117, 120,
+        123, 125, 128, 132, 135, 139, 143, 143, 143, 143, 97, 97, 96, 96, 95,
+        99, 104, 109, 114, 115, 115, 115, 115, 115, 115, 116, 116, 117, 119,
+        121, 122, 125, 128, 130, 133, 137, 140, 144, 148, 148, 148, 148, 99, 97,
+        96, 94, 92, 97, 102, 107, 113, 115, 116, 117, 119, 119, 120, 121, 121,
+        123, 125, 126, 128, 130, 133, 136, 138, 142, 145, 149, 152, 152, 152,
+        152, 100, 98, 95, 92, 90, 95, 100, 106, 112, 115, 117, 120, 123, 124,
+        125, 126, 127, 129, 131, 132, 134, 136, 139, 141, 144, 147, 150, 154,
+        157, 157, 157, 157, 102, 98, 94, 91, 88, 93, 98, 104, 111, 115, 119,
+        123, 127, 129, 131, 132, 134, 136, 137, 139, 140, 143, 145, 147, 150,
+        153, 156, 159, 163, 163, 163, 163, 104, 99, 96, 92, 89, 93, 99, 105,
+        111, 115, 119, 124, 129, 131, 133, 136, 138, 140, 142, 144, 146, 148,
+        150, 153, 155, 158, 161, 164, 168, 168, 168, 168, 106, 101, 97, 93, 89,
+        94, 99, 105, 111, 115, 120, 125, 131, 133, 136, 139, 142, 144, 146, 149,
+        151, 153, 156, 158, 161, 164, 167, 170, 173, 173, 173, 173, 107, 102,
+        98, 94, 90, 94, 99, 105, 111, 116, 121, 126, 132, 136, 139, 143, 147,
+        149, 151, 154, 157, 159, 162, 164, 167, 170, 173, 176, 179, 179, 179,
+        179, 109, 104, 99, 95, 91, 95, 100, 105, 111, 116, 121, 127, 134, 138,
+        142, 147, 151, 154, 157, 160, 163, 166, 168, 171, 174, 177, 179, 182,
+        186, 186, 186, 186, 113, 107, 102, 97, 93, 97, 102, 107, 112, 117, 123,
+        129, 136, 140, 144, 149, 154, 157, 160, 164, 167, 170, 173, 176, 179,
+        182, 185, 188, 191, 191, 191, 191, 116, 110, 105, 100, 95, 99, 104, 109,
+        114, 119, 125, 131, 137, 142, 146, 151, 157, 160, 164, 168, 172, 175,
+        178, 181, 184, 187, 190, 193, 196, 196, 196, 196, 120, 113, 107, 102,
+        98, 102, 106, 111, 116, 121, 126, 132, 139, 144, 149, 154, 160, 164,
+        168, 172, 176, 180, 183, 186, 190, 193, 196, 199, 202, 202, 202, 202,
+        123, 117, 111, 105, 100, 104, 108, 112, 117, 122, 128, 134, 140, 146,
+        151, 157, 163, 167, 172, 176, 181, 185, 188, 192, 196, 199, 202, 205,
+        209, 209, 209, 209, 127, 120, 114, 108, 103, 107, 111, 115, 120, 125,
+        130, 136, 143, 148, 153, 159, 166, 170, 175, 180, 185, 188, 192, 196,
+        200, 203, 207, 210, 214, 214, 214, 214, 132, 124, 118, 112, 107, 110,
+        114, 118, 123, 128, 133, 139, 145, 150, 156, 162, 168, 173, 178, 183,
+        188, 192, 196, 200, 205, 208, 211, 215, 219, 219, 219, 219, 136, 129,
+        122, 116, 110, 114, 117, 121, 125, 130, 136, 141, 147, 153, 158, 164,
+        171, 176, 181, 186, 192, 196, 200, 205, 209, 213, 216, 220, 224, 224,
+        224, 224, 141, 133, 126, 120, 114, 117, 121, 125, 128, 133, 138, 144,
+        150, 155, 161, 167, 174, 179, 184, 190, 196, 200, 205, 209, 214, 218,
+        222, 226, 230, 230, 230, 230, 146, 137, 130, 124, 118, 121, 124, 128,
+        132, 137, 142, 147, 153, 158, 164, 170, 177, 182, 187, 193, 199, 203,
+        208, 213, 218, 222, 226, 230, 234, 234, 234, 234, 151, 142, 135, 128,
+        122, 125, 128, 132, 135, 140, 145, 150, 156, 161, 167, 173, 179, 185,
+        190, 196, 202, 207, 211, 216, 222, 226, 230, 234, 238, 238, 238, 238,
+        156, 147, 139, 132, 126, 129, 132, 136, 139, 144, 149, 154, 159, 164,
+        170, 176, 182, 188, 193, 199, 205, 210, 215, 220, 226, 230, 234, 238,
+        243, 243, 243, 243, 161, 152, 144, 137, 131, 134, 137, 140, 143, 148,
+        152, 157, 163, 168, 173, 179, 186, 191, 196, 202, 209, 214, 219, 224,
+        230, 234, 238, 243, 248, 248, 248, 248, 161, 152, 144, 137, 131, 134,
+        137, 140, 143, 148, 152, 157, 163, 168, 173, 179, 186, 191, 196, 202,
+        209, 214, 219, 224, 230, 234, 238, 243, 248, 248, 248, 248, 161, 152,
+        144, 137, 131, 134, 137, 140, 143, 148, 152, 157, 163, 168, 173, 179,
+        186, 191, 196, 202, 209, 214, 219, 224, 230, 234, 238, 243, 248, 248,
+        248, 248, 161, 152, 144, 137, 131, 134, 137, 140, 143, 148, 152, 157,
+        163, 168, 173, 179, 186, 191, 196, 202, 209, 214, 219, 224, 230, 234,
+        238, 243, 248, 248, 248, 248 },
+      { /* Intra matrices */
+        /* Size 4 */
+        25, 42, 46, 60, 42, 51, 55, 65, 46, 55, 75, 88, 60, 65, 88, 109,
+        /* Size 8 */
+        27, 22, 41, 44, 48, 54, 63, 72, 22, 32, 42, 38, 39, 43, 50, 58, 41, 42,
+        50, 49, 48, 51, 57, 64, 44, 38, 49, 56, 59, 62, 67, 73, 48, 39, 48, 59,
+        67, 73, 78, 84, 54, 43, 51, 62, 73, 82, 90, 96, 63, 50, 57, 67, 78, 90,
+        99, 107, 72, 58, 64, 73, 84, 96, 107, 117,
+        /* Size 16 */
+        26, 23, 21, 27, 40, 41, 43, 44, 46, 49, 52, 56, 61, 65, 70, 70, 23, 24,
+        25, 31, 40, 40, 39, 40, 42, 44, 47, 50, 54, 58, 62, 62, 21, 25, 31, 35,
+        41, 39, 37, 37, 38, 40, 42, 45, 48, 52, 56, 56, 27, 31, 35, 39, 45, 43,
+        41, 41, 42, 44, 46, 48, 51, 55, 58, 58, 40, 40, 41, 45, 49, 48, 47, 47,
+        47, 48, 50, 52, 55, 58, 62, 62, 41, 40, 39, 43, 48, 49, 50, 51, 52, 53,
+        54, 57, 59, 62, 66, 66, 43, 39, 37, 41, 47, 50, 54, 56, 57, 59, 60, 62,
+        65, 68, 71, 71, 44, 40, 37, 41, 47, 51, 56, 58, 61, 63, 65, 67, 70, 73,
+        76, 76, 46, 42, 38, 42, 47, 52, 57, 61, 65, 68, 71, 73, 76, 79, 82, 82,
+        49, 44, 40, 44, 48, 53, 59, 63, 68, 71, 75, 78, 81, 84, 87, 87, 52, 47,
+        42, 46, 50, 54, 60, 65, 71, 75, 80, 83, 87, 90, 93, 93, 56, 50, 45, 48,
+        52, 57, 62, 67, 73, 78, 83, 87, 91, 95, 98, 98, 61, 54, 48, 51, 55, 59,
+        65, 70, 76, 81, 87, 91, 96, 100, 104, 104, 65, 58, 52, 55, 58, 62, 68,
+        73, 79, 84, 90, 95, 100, 104, 108, 108, 70, 62, 56, 58, 62, 66, 71, 76,
+        82, 87, 93, 98, 104, 108, 113, 113, 70, 62, 56, 58, 62, 66, 71, 76, 82,
+        87, 93, 98, 104, 108, 113, 113,
+        /* Size 32 */
+        26, 24, 23, 22, 21, 23, 27, 32, 39, 40, 41, 41, 42, 43, 44, 44, 45, 47,
+        48, 50, 52, 53, 55, 57, 60, 62, 64, 66, 69, 69, 69, 69, 24, 24, 23, 23,
+        22, 25, 29, 33, 40, 40, 40, 40, 40, 41, 42, 42, 43, 44, 46, 47, 49, 50,
+        52, 54, 56, 58, 60, 62, 65, 65, 65, 65, 23, 23, 24, 24, 25, 27, 30, 34,
+        40, 40, 39, 39, 39, 39, 40, 40, 41, 42, 43, 45, 46, 47, 49, 51, 53, 55,
+        57, 59, 61, 61, 61, 61, 22, 23, 24, 26, 27, 30, 32, 36, 40, 39, 39, 38,
+        37, 38, 38, 39, 39, 40, 41, 42, 43, 45, 46, 48, 50, 52, 54, 56, 58, 58,
+        58, 58, 21, 22, 25, 27, 31, 33, 35, 37, 40, 39, 38, 37, 36, 36, 37, 37,
+        37, 38, 39, 40, 41, 43, 44, 46, 47, 49, 51, 53, 55, 55, 55, 55, 23, 25,
+        27, 30, 33, 34, 37, 39, 42, 41, 40, 39, 38, 38, 39, 39, 39, 40, 41, 42,
+        43, 44, 46, 47, 49, 50, 52, 54, 56, 56, 56, 56, 27, 29, 30, 32, 35, 37,
+        39, 41, 44, 43, 42, 41, 40, 41, 41, 41, 41, 42, 43, 44, 45, 46, 47, 49,
+        50, 52, 54, 56, 58, 58, 58, 58, 32, 33, 34, 36, 37, 39, 41, 43, 46, 45,
+        44, 44, 43, 43, 43, 43, 43, 44, 45, 46, 47, 48, 49, 51, 52, 54, 55, 57,
+        59, 59, 59, 59, 39, 40, 40, 40, 40, 42, 44, 46, 48, 48, 47, 47, 46, 46,
+        46, 46, 46, 47, 47, 48, 49, 50, 51, 52, 54, 55, 57, 59, 61, 61, 61, 61,
+        40, 40, 40, 39, 39, 41, 43, 45, 48, 48, 48, 48, 48, 48, 48, 48, 48, 49,
+        50, 50, 51, 52, 53, 55, 56, 57, 59, 61, 63, 63, 63, 63, 41, 40, 39, 39,
+        38, 40, 42, 44, 47, 48, 48, 49, 50, 50, 50, 50, 51, 51, 52, 53, 54, 55,
+        56, 57, 58, 60, 61, 63, 65, 65, 65, 65, 41, 40, 39, 38, 37, 39, 41, 44,
+        47, 48, 49, 50, 51, 52, 52, 53, 53, 54, 55, 56, 56, 57, 58, 60, 61, 62,
+        64, 65, 67, 67, 67, 67, 42, 40, 39, 37, 36, 38, 40, 43, 46, 48, 50, 51,
+        53, 54, 55, 56, 56, 57, 58, 59, 59, 60, 61, 62, 64, 65, 66, 68, 69, 69,
+        69, 69, 43, 41, 39, 38, 36, 38, 41, 43, 46, 48, 50, 52, 54, 55, 56, 57,
+        58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 69, 70, 72, 72, 72, 72, 44, 42,
+        40, 38, 37, 39, 41, 43, 46, 48, 50, 52, 55, 56, 57, 59, 60, 61, 62, 63,
+        64, 65, 66, 67, 69, 70, 71, 73, 74, 74, 74, 74, 44, 42, 40, 39, 37, 39,
+        41, 43, 46, 48, 50, 53, 56, 57, 59, 60, 62, 63, 64, 66, 67, 68, 69, 70,
+        72, 73, 74, 76, 77, 77, 77, 77, 45, 43, 41, 39, 37, 39, 41, 43, 46, 48,
+        51, 53, 56, 58, 60, 62, 64, 65, 67, 68, 70, 71, 72, 73, 75, 76, 77, 79,
+        80, 80, 80, 80, 47, 44, 42, 40, 38, 40, 42, 44, 47, 49, 51, 54, 57, 59,
+        61, 63, 65, 67, 68, 70, 72, 73, 74, 76, 77, 78, 80, 81, 83, 83, 83, 83,
+        48, 46, 43, 41, 39, 41, 43, 45, 47, 50, 52, 55, 58, 60, 62, 64, 67, 68,
+        70, 72, 74, 75, 77, 78, 80, 81, 83, 84, 86, 86, 86, 86, 50, 47, 45, 42,
+        40, 42, 44, 46, 48, 50, 53, 56, 59, 61, 63, 66, 68, 70, 72, 74, 76, 77,
+        79, 81, 82, 84, 85, 87, 89, 89, 89, 89, 52, 49, 46, 43, 41, 43, 45, 47,
+        49, 51, 54, 56, 59, 62, 64, 67, 70, 72, 74, 76, 78, 80, 82, 83, 85, 87,
+        88, 90, 92, 92, 92, 92, 53, 50, 47, 45, 43, 44, 46, 48, 50, 52, 55, 57,
+        60, 63, 65, 68, 71, 73, 75, 77, 80, 82, 83, 85, 87, 89, 91, 92, 94, 94,
+        94, 94, 55, 52, 49, 46, 44, 46, 47, 49, 51, 53, 56, 58, 61, 64, 66, 69,
+        72, 74, 77, 79, 82, 83, 85, 87, 90, 91, 93, 95, 97, 97, 97, 97, 57, 54,
+        51, 48, 46, 47, 49, 51, 52, 55, 57, 60, 62, 65, 67, 70, 73, 76, 78, 81,
+        83, 85, 87, 90, 92, 94, 95, 97, 99, 99, 99, 99, 60, 56, 53, 50, 47, 49,
+        50, 52, 54, 56, 58, 61, 64, 66, 69, 72, 75, 77, 80, 82, 85, 87, 90, 92,
+        94, 96, 98, 100, 102, 102, 102, 102, 62, 58, 55, 52, 49, 50, 52, 54, 55,
+        57, 60, 62, 65, 67, 70, 73, 76, 78, 81, 84, 87, 89, 91, 94, 96, 98, 100,
+        102, 104, 104, 104, 104, 64, 60, 57, 54, 51, 52, 54, 55, 57, 59, 61, 64,
+        66, 69, 71, 74, 77, 80, 83, 85, 88, 91, 93, 95, 98, 100, 102, 104, 107,
+        107, 107, 107, 66, 62, 59, 56, 53, 54, 56, 57, 59, 61, 63, 65, 68, 70,
+        73, 76, 79, 81, 84, 87, 90, 92, 95, 97, 100, 102, 104, 107, 109, 109,
+        109, 109, 69, 65, 61, 58, 55, 56, 58, 59, 61, 63, 65, 67, 69, 72, 74,
+        77, 80, 83, 86, 89, 92, 94, 97, 99, 102, 104, 107, 109, 111, 111, 111,
+        111, 69, 65, 61, 58, 55, 56, 58, 59, 61, 63, 65, 67, 69, 72, 74, 77, 80,
+        83, 86, 89, 92, 94, 97, 99, 102, 104, 107, 109, 111, 111, 111, 111, 69,
+        65, 61, 58, 55, 56, 58, 59, 61, 63, 65, 67, 69, 72, 74, 77, 80, 83, 86,
+        89, 92, 94, 97, 99, 102, 104, 107, 109, 111, 111, 111, 111, 69, 65, 61,
+        58, 55, 56, 58, 59, 61, 63, 65, 67, 69, 72, 74, 77, 80, 83, 86, 89, 92,
+        94, 97, 99, 102, 104, 107, 109, 111, 111, 111, 111 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 70, 116, 182, 70, 106, 148, 199, 116, 148, 207, 251, 182, 199, 251,
+        292,
+        /* Size 8 */
+        64, 48, 52, 69, 92, 120, 150, 179, 48, 55, 53, 62, 79, 102, 128, 155,
+        52, 53, 74, 87, 101, 121, 144, 168, 69, 62, 87, 110, 129, 147, 166, 186,
+        92, 79, 101, 129, 153, 172, 190, 207, 120, 102, 121, 147, 172, 194, 211,
+        226, 150, 128, 144, 166, 190, 211, 228, 242, 179, 155, 168, 186, 207,
+        226, 242, 255,
+        /* Size 16 */
+        64, 55, 48, 50, 52, 59, 69, 79, 92, 104, 120, 133, 150, 163, 179, 179,
+        55, 53, 51, 52, 52, 58, 65, 74, 85, 96, 110, 123, 138, 151, 166, 166,
+        48, 51, 55, 54, 53, 57, 62, 70, 79, 89, 102, 113, 128, 140, 155, 155,
+        50, 52, 54, 58, 62, 67, 72, 80, 89, 99, 111, 122, 135, 147, 161, 161,
+        52, 52, 53, 62, 74, 80, 87, 93, 101, 110, 121, 131, 144, 155, 168, 168,
+        59, 58, 57, 67, 80, 88, 97, 105, 113, 122, 133, 143, 154, 165, 177, 177,
+        69, 65, 62, 72, 87, 97, 110, 119, 129, 137, 147, 156, 166, 176, 186,
+        186, 79, 74, 70, 80, 93, 105, 119, 128, 140, 149, 158, 167, 177, 186,
+        196, 196, 92, 85, 79, 89, 101, 113, 129, 140, 153, 162, 172, 181, 190,
+        198, 207, 207, 104, 96, 89, 99, 110, 122, 137, 149, 162, 171, 182, 191,
+        200, 208, 216, 216, 120, 110, 102, 111, 121, 133, 147, 158, 172, 182,
+        194, 202, 211, 218, 226, 226, 133, 123, 113, 122, 131, 143, 156, 167,
+        181, 191, 202, 210, 219, 227, 234, 234, 150, 138, 128, 135, 144, 154,
+        166, 177, 190, 200, 211, 219, 228, 235, 242, 242, 163, 151, 140, 147,
+        155, 165, 176, 186, 198, 208, 218, 227, 235, 242, 249, 249, 179, 166,
+        155, 161, 168, 177, 186, 196, 207, 216, 226, 234, 242, 249, 255, 255,
+        179, 166, 155, 161, 168, 177, 186, 196, 207, 216, 226, 234, 242, 249,
+        255, 255,
+        /* Size 32 */
+        64, 59, 55, 51, 48, 49, 50, 51, 52, 55, 59, 64, 69, 73, 79, 85, 92, 98,
+        104, 112, 120, 126, 133, 141, 150, 156, 163, 170, 179, 179, 179, 179,
+        59, 56, 54, 52, 50, 50, 51, 52, 52, 55, 59, 62, 67, 71, 76, 82, 89, 94,
+        100, 107, 115, 121, 128, 135, 144, 150, 156, 164, 172, 172, 172, 172,
+        55, 54, 53, 52, 51, 52, 52, 52, 52, 55, 58, 61, 65, 69, 74, 79, 85, 90,
+        96, 103, 110, 116, 123, 130, 138, 144, 151, 158, 166, 166, 166, 166, 51,
+        52, 52, 53, 53, 53, 53, 53, 53, 55, 58, 60, 63, 67, 72, 77, 82, 87, 93,
+        99, 106, 112, 118, 125, 133, 139, 145, 152, 160, 160, 160, 160, 48, 50,
+        51, 53, 55, 55, 54, 53, 53, 55, 57, 59, 62, 66, 70, 74, 79, 84, 89, 95,
+        102, 107, 113, 120, 128, 134, 140, 147, 155, 155, 155, 155, 49, 50, 52,
+        53, 55, 55, 56, 56, 57, 59, 61, 64, 67, 70, 74, 79, 84, 89, 94, 100,
+        106, 112, 117, 124, 131, 137, 144, 150, 158, 158, 158, 158, 50, 51, 52,
+        53, 54, 56, 58, 60, 62, 64, 67, 69, 72, 76, 80, 84, 89, 94, 99, 104,
+        111, 116, 122, 128, 135, 141, 147, 154, 161, 161, 161, 161, 51, 52, 52,
+        53, 53, 56, 60, 63, 67, 70, 73, 76, 79, 82, 86, 90, 95, 99, 104, 110,
+        116, 121, 126, 133, 139, 145, 151, 157, 164, 164, 164, 164, 52, 52, 52,
+        53, 53, 57, 62, 67, 74, 77, 80, 83, 87, 90, 93, 97, 101, 106, 110, 115,
+        121, 126, 131, 137, 144, 149, 155, 161, 168, 168, 168, 168, 55, 55, 55,
+        55, 55, 59, 64, 70, 77, 80, 84, 87, 92, 95, 99, 103, 107, 111, 116, 121,
+        126, 131, 137, 142, 149, 154, 160, 166, 172, 172, 172, 172, 59, 59, 58,
+        58, 57, 61, 67, 73, 80, 84, 88, 92, 97, 101, 105, 109, 113, 118, 122,
+        127, 133, 137, 143, 148, 154, 159, 165, 170, 177, 177, 177, 177, 64, 62,
+        61, 60, 59, 64, 69, 76, 83, 87, 92, 97, 103, 107, 111, 116, 121, 125,
+        129, 134, 139, 144, 149, 154, 160, 165, 170, 176, 181, 181, 181, 181,
+        69, 67, 65, 63, 62, 67, 72, 79, 87, 92, 97, 103, 110, 114, 119, 124,
+        129, 133, 137, 142, 147, 151, 156, 161, 166, 171, 176, 181, 186, 186,
+        186, 186, 73, 71, 69, 67, 66, 70, 76, 82, 90, 95, 101, 107, 114, 119,
+        123, 129, 134, 138, 143, 147, 152, 157, 161, 166, 172, 176, 181, 186,
+        191, 191, 191, 191, 79, 76, 74, 72, 70, 74, 80, 86, 93, 99, 105, 111,
+        119, 123, 128, 134, 140, 144, 149, 153, 158, 163, 167, 172, 177, 182,
+        186, 191, 196, 196, 196, 196, 85, 82, 79, 77, 74, 79, 84, 90, 97, 103,
+        109, 116, 124, 129, 134, 140, 146, 150, 155, 160, 165, 169, 174, 178,
+        183, 188, 192, 197, 202, 202, 202, 202, 92, 89, 85, 82, 79, 84, 89, 95,
+        101, 107, 113, 121, 129, 134, 140, 146, 153, 157, 162, 167, 172, 176,
+        181, 185, 190, 194, 198, 203, 207, 207, 207, 207, 98, 94, 90, 87, 84,
+        89, 94, 99, 106, 111, 118, 125, 133, 138, 144, 150, 157, 162, 167, 172,
+        177, 181, 186, 190, 195, 199, 203, 207, 212, 212, 212, 212, 104, 100,
+        96, 93, 89, 94, 99, 104, 110, 116, 122, 129, 137, 143, 149, 155, 162,
+        167, 171, 177, 182, 186, 191, 195, 200, 204, 208, 212, 216, 216, 216,
+        216, 112, 107, 103, 99, 95, 100, 104, 110, 115, 121, 127, 134, 142, 147,
+        153, 160, 167, 172, 177, 182, 188, 192, 196, 201, 205, 209, 213, 217,
+        221, 221, 221, 221, 120, 115, 110, 106, 102, 106, 111, 116, 121, 126,
+        133, 139, 147, 152, 158, 165, 172, 177, 182, 188, 194, 198, 202, 206,
+        211, 215, 218, 222, 226, 226, 226, 226, 126, 121, 116, 112, 107, 112,
+        116, 121, 126, 131, 137, 144, 151, 157, 163, 169, 176, 181, 186, 192,
+        198, 202, 206, 211, 215, 219, 222, 226, 230, 230, 230, 230, 133, 128,
+        123, 118, 113, 117, 122, 126, 131, 137, 143, 149, 156, 161, 167, 174,
+        181, 186, 191, 196, 202, 206, 210, 215, 219, 223, 227, 230, 234, 234,
+        234, 234, 141, 135, 130, 125, 120, 124, 128, 133, 137, 142, 148, 154,
+        161, 166, 172, 178, 185, 190, 195, 201, 206, 211, 215, 219, 224, 227,
+        231, 234, 238, 238, 238, 238, 150, 144, 138, 133, 128, 131, 135, 139,
+        144, 149, 154, 160, 166, 172, 177, 183, 190, 195, 200, 205, 211, 215,
+        219, 224, 228, 232, 235, 239, 242, 242, 242, 242, 156, 150, 144, 139,
+        134, 137, 141, 145, 149, 154, 159, 165, 171, 176, 182, 188, 194, 199,
+        204, 209, 215, 219, 223, 227, 232, 235, 238, 242, 246, 246, 246, 246,
+        163, 156, 151, 145, 140, 144, 147, 151, 155, 160, 165, 170, 176, 181,
+        186, 192, 198, 203, 208, 213, 218, 222, 227, 231, 235, 238, 242, 245,
+        249, 249, 249, 249, 170, 164, 158, 152, 147, 150, 154, 157, 161, 166,
+        170, 176, 181, 186, 191, 197, 203, 207, 212, 217, 222, 226, 230, 234,
+        239, 242, 245, 249, 252, 252, 252, 252, 179, 172, 166, 160, 155, 158,
+        161, 164, 168, 172, 177, 181, 186, 191, 196, 202, 207, 212, 216, 221,
+        226, 230, 234, 238, 242, 246, 249, 252, 255, 255, 255, 255, 179, 172,
+        166, 160, 155, 158, 161, 164, 168, 172, 177, 181, 186, 191, 196, 202,
+        207, 212, 216, 221, 226, 230, 234, 238, 242, 246, 249, 252, 255, 255,
+        255, 255, 179, 172, 166, 160, 155, 158, 161, 164, 168, 172, 177, 181,
+        186, 191, 196, 202, 207, 212, 216, 221, 226, 230, 234, 238, 242, 246,
+        249, 252, 255, 255, 255, 255, 179, 172, 166, 160, 155, 158, 161, 164,
+        168, 172, 177, 181, 186, 191, 196, 202, 207, 212, 216, 221, 226, 230,
+        234, 238, 242, 246, 249, 252, 255, 255, 255, 255 },
+      { /* Intra matrices */
+        /* Size 4 */
+        21, 23, 40, 65, 23, 36, 52, 72, 40, 52, 75, 93, 65, 72, 93, 111,
+        /* Size 8 */
+        26, 19, 21, 28, 38, 50, 63, 77, 19, 22, 21, 25, 32, 42, 53, 66, 21, 21,
+        30, 35, 42, 50, 61, 72, 28, 25, 35, 46, 54, 62, 71, 81, 38, 32, 42, 54,
+        65, 74, 83, 92, 50, 42, 50, 62, 74, 85, 94, 101, 63, 53, 61, 71, 83, 94,
+        103, 110, 77, 66, 72, 81, 92, 101, 110, 117,
+        /* Size 16 */
+        25, 21, 18, 19, 20, 23, 26, 31, 36, 41, 48, 54, 61, 67, 74, 74, 21, 20,
+        20, 20, 20, 22, 25, 29, 33, 38, 44, 49, 56, 61, 68, 68, 18, 20, 21, 21,
+        20, 22, 24, 27, 31, 35, 40, 45, 51, 57, 63, 63, 19, 20, 21, 22, 24, 26,
+        28, 31, 35, 39, 44, 49, 55, 60, 66, 66, 20, 20, 20, 24, 29, 31, 34, 37,
+        40, 44, 48, 53, 58, 63, 69, 69, 23, 22, 22, 26, 31, 34, 38, 41, 45, 49,
+        53, 58, 63, 68, 73, 73, 26, 25, 24, 28, 34, 38, 44, 47, 52, 55, 60, 64,
+        68, 73, 78, 78, 31, 29, 27, 31, 37, 41, 47, 52, 57, 60, 65, 69, 74, 78,
+        83, 83, 36, 33, 31, 35, 40, 45, 52, 57, 62, 67, 71, 75, 80, 84, 88, 88,
+        41, 38, 35, 39, 44, 49, 55, 60, 67, 71, 76, 80, 84, 88, 92, 92, 48, 44,
+        40, 44, 48, 53, 60, 65, 71, 76, 81, 85, 90, 94, 97, 97, 54, 49, 45, 49,
+        53, 58, 64, 69, 75, 80, 85, 90, 94, 98, 101, 101, 61, 56, 51, 55, 58,
+        63, 68, 74, 80, 84, 90, 94, 99, 102, 106, 106, 67, 61, 57, 60, 63, 68,
+        73, 78, 84, 88, 94, 98, 102, 105, 109, 109, 74, 68, 63, 66, 69, 73, 78,
+        83, 88, 92, 97, 101, 106, 109, 112, 112, 74, 68, 63, 66, 69, 73, 78, 83,
+        88, 92, 97, 101, 106, 109, 112, 112,
+        /* Size 32 */
+        24, 22, 21, 19, 18, 18, 19, 19, 19, 21, 22, 24, 26, 28, 30, 32, 35, 38,
+        40, 43, 47, 50, 53, 56, 60, 63, 66, 69, 73, 73, 73, 73, 22, 21, 20, 19,
+        19, 19, 19, 19, 19, 21, 22, 23, 25, 27, 29, 31, 34, 36, 39, 41, 45, 47,
+        50, 53, 57, 60, 63, 66, 70, 70, 70, 70, 21, 20, 20, 19, 19, 19, 19, 19,
+        20, 21, 22, 23, 25, 26, 28, 30, 33, 35, 37, 40, 43, 45, 48, 51, 55, 57,
+        60, 63, 67, 67, 67, 67, 19, 19, 19, 20, 20, 20, 20, 20, 20, 21, 22, 23,
+        24, 25, 27, 29, 31, 33, 36, 38, 41, 43, 46, 49, 52, 55, 58, 61, 64, 64,
+        64, 64, 18, 19, 19, 20, 21, 20, 20, 20, 20, 21, 21, 22, 23, 25, 26, 28,
+        30, 32, 34, 37, 39, 42, 44, 47, 50, 53, 56, 59, 62, 62, 62, 62, 18, 19,
+        19, 20, 20, 21, 21, 21, 21, 22, 23, 24, 25, 27, 28, 30, 32, 34, 36, 38,
+        41, 43, 46, 49, 52, 54, 57, 60, 63, 63, 63, 63, 19, 19, 19, 20, 20, 21,
+        22, 22, 23, 24, 25, 26, 27, 29, 30, 32, 34, 36, 38, 40, 43, 45, 48, 50,
+        53, 56, 59, 62, 65, 65, 65, 65, 19, 19, 19, 20, 20, 21, 22, 24, 25, 26,
+        27, 29, 30, 31, 33, 35, 36, 38, 40, 43, 45, 47, 50, 52, 55, 58, 60, 63,
+        66, 66, 66, 66, 19, 19, 20, 20, 20, 21, 23, 25, 28, 29, 30, 32, 33, 34,
+        36, 37, 39, 41, 43, 45, 47, 49, 52, 54, 57, 59, 62, 65, 68, 68, 68, 68,
+        21, 21, 21, 21, 21, 22, 24, 26, 29, 30, 32, 33, 35, 37, 38, 40, 42, 43,
+        45, 47, 50, 52, 54, 57, 59, 62, 64, 67, 70, 70, 70, 70, 22, 22, 22, 22,
+        21, 23, 25, 27, 30, 32, 34, 35, 37, 39, 40, 42, 44, 46, 48, 50, 52, 54,
+        57, 59, 62, 64, 66, 69, 72, 72, 72, 72, 24, 23, 23, 23, 22, 24, 26, 29,
+        32, 33, 35, 38, 40, 42, 43, 45, 47, 49, 51, 53, 55, 57, 59, 62, 64, 66,
+        69, 71, 74, 74, 74, 74, 26, 25, 25, 24, 23, 25, 27, 30, 33, 35, 37, 40,
+        43, 45, 46, 48, 51, 52, 54, 56, 58, 60, 62, 65, 67, 69, 71, 74, 76, 76,
+        76, 76, 28, 27, 26, 25, 25, 27, 29, 31, 34, 37, 39, 42, 45, 46, 48, 51,
+        53, 55, 57, 59, 61, 63, 65, 67, 70, 72, 74, 76, 79, 79, 79, 79, 30, 29,
+        28, 27, 26, 28, 30, 33, 36, 38, 40, 43, 46, 48, 51, 53, 55, 57, 59, 61,
+        64, 66, 68, 70, 72, 74, 76, 79, 81, 81, 81, 81, 32, 31, 30, 29, 28, 30,
+        32, 35, 37, 40, 42, 45, 48, 51, 53, 55, 58, 60, 62, 64, 67, 69, 71, 73,
+        75, 77, 79, 81, 83, 83, 83, 83, 35, 34, 33, 31, 30, 32, 34, 36, 39, 42,
+        44, 47, 51, 53, 55, 58, 61, 63, 65, 67, 70, 72, 74, 76, 78, 80, 82, 84,
+        86, 86, 86, 86, 38, 36, 35, 33, 32, 34, 36, 38, 41, 43, 46, 49, 52, 55,
+        57, 60, 63, 65, 67, 70, 72, 74, 76, 78, 80, 82, 84, 86, 88, 88, 88, 88,
+        40, 39, 37, 36, 34, 36, 38, 40, 43, 45, 48, 51, 54, 57, 59, 62, 65, 67,
+        70, 72, 74, 76, 78, 81, 83, 85, 86, 88, 91, 91, 91, 91, 43, 41, 40, 38,
+        37, 38, 40, 43, 45, 47, 50, 53, 56, 59, 61, 64, 67, 70, 72, 74, 77, 79,
+        81, 83, 85, 87, 89, 91, 93, 93, 93, 93, 47, 45, 43, 41, 39, 41, 43, 45,
+        47, 50, 52, 55, 58, 61, 64, 67, 70, 72, 74, 77, 80, 82, 84, 86, 88, 90,
+        92, 94, 95, 95, 95, 95, 50, 47, 45, 43, 42, 43, 45, 47, 49, 52, 54, 57,
+        60, 63, 66, 69, 72, 74, 76, 79, 82, 84, 86, 88, 90, 92, 94, 95, 97, 97,
+        97, 97, 53, 50, 48, 46, 44, 46, 48, 50, 52, 54, 57, 59, 62, 65, 68, 71,
+        74, 76, 78, 81, 84, 86, 88, 90, 92, 94, 96, 97, 99, 99, 99, 99, 56, 53,
+        51, 49, 47, 49, 50, 52, 54, 57, 59, 62, 65, 67, 70, 73, 76, 78, 81, 83,
+        86, 88, 90, 92, 94, 96, 98, 100, 101, 101, 101, 101, 60, 57, 55, 52, 50,
+        52, 53, 55, 57, 59, 62, 64, 67, 70, 72, 75, 78, 80, 83, 85, 88, 90, 92,
+        94, 97, 98, 100, 102, 104, 104, 104, 104, 63, 60, 57, 55, 53, 54, 56,
+        58, 59, 62, 64, 66, 69, 72, 74, 77, 80, 82, 85, 87, 90, 92, 94, 96, 98,
+        100, 102, 103, 105, 105, 105, 105, 66, 63, 60, 58, 56, 57, 59, 60, 62,
+        64, 66, 69, 71, 74, 76, 79, 82, 84, 86, 89, 92, 94, 96, 98, 100, 102,
+        103, 105, 107, 107, 107, 107, 69, 66, 63, 61, 59, 60, 62, 63, 65, 67,
+        69, 71, 74, 76, 79, 81, 84, 86, 88, 91, 94, 95, 97, 100, 102, 103, 105,
+        107, 108, 108, 108, 108, 73, 70, 67, 64, 62, 63, 65, 66, 68, 70, 72, 74,
+        76, 79, 81, 83, 86, 88, 91, 93, 95, 97, 99, 101, 104, 105, 107, 108,
+        110, 110, 110, 110, 73, 70, 67, 64, 62, 63, 65, 66, 68, 70, 72, 74, 76,
+        79, 81, 83, 86, 88, 91, 93, 95, 97, 99, 101, 104, 105, 107, 108, 110,
+        110, 110, 110, 73, 70, 67, 64, 62, 63, 65, 66, 68, 70, 72, 74, 76, 79,
+        81, 83, 86, 88, 91, 93, 95, 97, 99, 101, 104, 105, 107, 108, 110, 110,
+        110, 110, 73, 70, 67, 64, 62, 63, 65, 66, 68, 70, 72, 74, 76, 79, 81,
+        83, 86, 88, 91, 93, 95, 97, 99, 101, 104, 105, 107, 108, 110, 110, 110,
+        110 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 101, 110, 139, 101, 122, 130, 149, 110, 130, 167, 190, 139, 149,
+        190, 225,
+        /* Size 8 */
+        64, 52, 94, 99, 106, 119, 134, 152, 52, 75, 96, 87, 89, 98, 110, 125,
+        94, 96, 112, 108, 108, 113, 123, 136, 99, 87, 108, 122, 128, 134, 142,
+        153, 106, 89, 108, 128, 143, 153, 162, 172, 119, 98, 113, 134, 153, 168,
+        180, 191, 134, 110, 123, 142, 162, 180, 195, 207, 152, 125, 136, 153,
+        172, 191, 207, 221,
+        /* Size 16 */
+        64, 57, 52, 67, 94, 97, 99, 103, 106, 112, 119, 126, 134, 142, 152, 152,
+        57, 59, 61, 74, 95, 94, 93, 95, 97, 102, 107, 114, 121, 129, 137, 137,
+        52, 61, 75, 84, 96, 91, 87, 88, 89, 93, 98, 104, 110, 117, 125, 125, 67,
+        74, 84, 92, 103, 99, 96, 97, 97, 101, 105, 110, 116, 123, 130, 130, 94,
+        95, 96, 103, 112, 110, 108, 108, 108, 110, 113, 118, 123, 129, 136, 136,
+        97, 94, 91, 99, 110, 112, 115, 116, 117, 120, 123, 127, 132, 138, 144,
+        144, 99, 93, 87, 96, 108, 115, 122, 125, 128, 131, 134, 138, 142, 147,
+        153, 153, 103, 95, 88, 97, 108, 116, 125, 130, 135, 139, 143, 147, 151,
+        156, 162, 162, 106, 97, 89, 97, 108, 117, 128, 135, 143, 148, 153, 157,
+        162, 167, 172, 172, 112, 102, 93, 101, 110, 120, 131, 139, 148, 154,
+        160, 165, 171, 176, 181, 181, 119, 107, 98, 105, 113, 123, 134, 143,
+        153, 160, 168, 174, 180, 185, 191, 191, 126, 114, 104, 110, 118, 127,
+        138, 147, 157, 165, 174, 180, 187, 193, 199, 199, 134, 121, 110, 116,
+        123, 132, 142, 151, 162, 171, 180, 187, 195, 201, 207, 207, 142, 129,
+        117, 123, 129, 138, 147, 156, 167, 176, 185, 193, 201, 207, 214, 214,
+        152, 137, 125, 130, 136, 144, 153, 162, 172, 181, 191, 199, 207, 214,
+        221, 221, 152, 137, 125, 130, 136, 144, 153, 162, 172, 181, 191, 199,
+        207, 214, 221, 221,
+        /* Size 32 */
+        64, 60, 57, 54, 52, 58, 67, 78, 94, 95, 97, 98, 99, 101, 103, 104, 106,
+        109, 112, 115, 119, 122, 126, 130, 134, 138, 142, 147, 152, 152, 152,
+        152, 60, 59, 58, 57, 56, 62, 70, 80, 94, 95, 95, 95, 96, 97, 99, 100,
+        101, 104, 107, 110, 113, 116, 120, 123, 127, 131, 135, 139, 144, 144,
+        144, 144, 57, 58, 59, 60, 61, 67, 74, 83, 95, 94, 94, 93, 93, 94, 95,
+        96, 97, 99, 102, 104, 107, 110, 114, 117, 121, 125, 129, 133, 137, 137,
+        137, 137, 54, 57, 60, 63, 67, 72, 79, 86, 95, 94, 92, 91, 89, 90, 91,
+        92, 93, 95, 97, 100, 102, 105, 109, 112, 115, 119, 123, 127, 131, 131,
+        131, 131, 52, 56, 61, 67, 75, 79, 84, 89, 96, 93, 91, 89, 87, 87, 88,
+        88, 89, 91, 93, 96, 98, 101, 104, 107, 110, 114, 117, 121, 125, 125,
+        125, 125, 58, 62, 67, 72, 79, 83, 88, 93, 99, 97, 95, 93, 91, 92, 92,
+        93, 93, 95, 97, 99, 101, 104, 107, 110, 113, 117, 120, 124, 128, 128,
+        128, 128, 67, 70, 74, 79, 84, 88, 92, 97, 103, 101, 99, 98, 96, 96, 97,
+        97, 97, 99, 101, 103, 105, 108, 110, 113, 116, 120, 123, 127, 130, 130,
+        130, 130, 78, 80, 83, 86, 89, 93, 97, 102, 107, 106, 104, 103, 102, 102,
+        102, 102, 102, 104, 106, 107, 109, 112, 114, 117, 120, 123, 126, 130,
+        133, 133, 133, 133, 94, 94, 95, 95, 96, 99, 103, 107, 112, 111, 110,
+        109, 108, 108, 108, 108, 108, 109, 110, 112, 113, 116, 118, 121, 123,
+        126, 129, 133, 136, 136, 136, 136, 95, 95, 94, 94, 93, 97, 101, 106,
+        111, 111, 111, 111, 111, 111, 112, 112, 112, 113, 115, 116, 118, 120,
+        122, 125, 127, 130, 133, 137, 140, 140, 140, 140, 97, 95, 94, 92, 91,
+        95, 99, 104, 110, 111, 112, 113, 115, 115, 116, 116, 117, 118, 120, 121,
+        123, 125, 127, 129, 132, 135, 138, 141, 144, 144, 144, 144, 98, 95, 93,
+        91, 89, 93, 98, 103, 109, 111, 113, 116, 118, 119, 120, 121, 122, 124,
+        125, 127, 128, 130, 132, 134, 137, 139, 142, 145, 148, 148, 148, 148,
+        99, 96, 93, 89, 87, 91, 96, 102, 108, 111, 115, 118, 122, 124, 125, 127,
+        128, 130, 131, 132, 134, 136, 138, 140, 142, 144, 147, 150, 153, 153,
+        153, 153, 101, 97, 94, 90, 87, 92, 96, 102, 108, 111, 115, 119, 124,
+        125, 127, 129, 132, 133, 135, 136, 138, 140, 142, 144, 146, 149, 152,
+        154, 157, 157, 157, 157, 103, 99, 95, 91, 88, 92, 97, 102, 108, 112,
+        116, 120, 125, 127, 130, 133, 135, 137, 139, 141, 143, 145, 147, 149,
+        151, 154, 156, 159, 162, 162, 162, 162, 104, 100, 96, 92, 88, 93, 97,
+        102, 108, 112, 116, 121, 127, 129, 133, 136, 139, 141, 143, 145, 148,
+        150, 152, 154, 156, 159, 161, 164, 167, 167, 167, 167, 106, 101, 97, 93,
+        89, 93, 97, 102, 108, 112, 117, 122, 128, 132, 135, 139, 143, 145, 148,
+        150, 153, 155, 157, 160, 162, 164, 167, 169, 172, 172, 172, 172, 109,
+        104, 99, 95, 91, 95, 99, 104, 109, 113, 118, 124, 130, 133, 137, 141,
+        145, 148, 151, 154, 157, 159, 161, 164, 166, 169, 171, 174, 176, 176,
+        176, 176, 112, 107, 102, 97, 93, 97, 101, 106, 110, 115, 120, 125, 131,
+        135, 139, 143, 148, 151, 154, 157, 160, 163, 165, 168, 171, 173, 176,
+        178, 181, 181, 181, 181, 115, 110, 104, 100, 96, 99, 103, 107, 112, 116,
+        121, 127, 132, 136, 141, 145, 150, 154, 157, 161, 164, 167, 170, 172,
+        175, 178, 180, 183, 186, 186, 186, 186, 119, 113, 107, 102, 98, 101,
+        105, 109, 113, 118, 123, 128, 134, 138, 143, 148, 153, 157, 160, 164,
+        168, 171, 174, 177, 180, 183, 185, 188, 191, 191, 191, 191, 122, 116,
+        110, 105, 101, 104, 108, 112, 116, 120, 125, 130, 136, 140, 145, 150,
+        155, 159, 163, 167, 171, 174, 177, 180, 184, 186, 189, 192, 195, 195,
+        195, 195, 126, 120, 114, 109, 104, 107, 110, 114, 118, 122, 127, 132,
+        138, 142, 147, 152, 157, 161, 165, 170, 174, 177, 180, 184, 187, 190,
+        193, 196, 199, 199, 199, 199, 130, 123, 117, 112, 107, 110, 113, 117,
+        121, 125, 129, 134, 140, 144, 149, 154, 160, 164, 168, 172, 177, 180,
+        184, 187, 191, 194, 197, 200, 203, 203, 203, 203, 134, 127, 121, 115,
+        110, 113, 116, 120, 123, 127, 132, 137, 142, 146, 151, 156, 162, 166,
+        171, 175, 180, 184, 187, 191, 195, 198, 201, 204, 207, 207, 207, 207,
+        138, 131, 125, 119, 114, 117, 120, 123, 126, 130, 135, 139, 144, 149,
+        154, 159, 164, 169, 173, 178, 183, 186, 190, 194, 198, 201, 204, 207,
+        211, 211, 211, 211, 142, 135, 129, 123, 117, 120, 123, 126, 129, 133,
+        138, 142, 147, 152, 156, 161, 167, 171, 176, 180, 185, 189, 193, 197,
+        201, 204, 207, 211, 214, 214, 214, 214, 147, 139, 133, 127, 121, 124,
+        127, 130, 133, 137, 141, 145, 150, 154, 159, 164, 169, 174, 178, 183,
+        188, 192, 196, 200, 204, 207, 211, 214, 217, 217, 217, 217, 152, 144,
+        137, 131, 125, 128, 130, 133, 136, 140, 144, 148, 153, 157, 162, 167,
+        172, 176, 181, 186, 191, 195, 199, 203, 207, 211, 214, 217, 221, 221,
+        221, 221, 152, 144, 137, 131, 125, 128, 130, 133, 136, 140, 144, 148,
+        153, 157, 162, 167, 172, 176, 181, 186, 191, 195, 199, 203, 207, 211,
+        214, 217, 221, 221, 221, 221, 152, 144, 137, 131, 125, 128, 130, 133,
+        136, 140, 144, 148, 153, 157, 162, 167, 172, 176, 181, 186, 191, 195,
+        199, 203, 207, 211, 214, 217, 221, 221, 221, 221, 152, 144, 137, 131,
+        125, 128, 130, 133, 136, 140, 144, 148, 153, 157, 162, 167, 172, 176,
+        181, 186, 191, 195, 199, 203, 207, 211, 214, 217, 221, 221, 221, 221 },
+      { /* Intra matrices */
+        /* Size 4 */
+        27, 44, 48, 61, 44, 53, 57, 66, 48, 57, 74, 86, 61, 66, 86, 103,
+        /* Size 8 */
+        29, 23, 43, 46, 49, 56, 63, 72, 23, 34, 44, 40, 41, 45, 51, 59, 43, 44,
+        52, 50, 50, 53, 58, 64, 46, 40, 50, 57, 60, 63, 67, 73, 49, 41, 50, 60,
+        68, 73, 78, 83, 56, 45, 53, 63, 73, 81, 88, 93, 63, 51, 58, 67, 78, 88,
+        96, 102, 72, 59, 64, 73, 83, 93, 102, 110,
+        /* Size 16 */
+        28, 25, 23, 29, 42, 43, 45, 46, 48, 51, 54, 58, 62, 66, 70, 70, 25, 26,
+        27, 33, 42, 42, 42, 43, 44, 46, 49, 52, 55, 59, 63, 63, 23, 27, 33, 37,
+        43, 41, 39, 39, 40, 42, 44, 47, 50, 53, 57, 57, 29, 33, 37, 41, 46, 45,
+        43, 44, 44, 46, 47, 50, 53, 56, 60, 60, 42, 42, 43, 46, 51, 50, 49, 49,
+        49, 50, 51, 54, 56, 59, 63, 63, 43, 42, 41, 45, 50, 51, 52, 53, 53, 55,
+        56, 58, 61, 63, 67, 67, 45, 42, 39, 43, 49, 52, 56, 57, 59, 60, 61, 63,
+        65, 68, 71, 71, 46, 43, 39, 44, 49, 53, 57, 60, 62, 64, 66, 68, 70, 73,
+        75, 75, 48, 44, 40, 44, 49, 53, 59, 62, 66, 69, 71, 73, 76, 78, 81, 81,
+        51, 46, 42, 46, 50, 55, 60, 64, 69, 72, 75, 77, 80, 83, 85, 85, 54, 49,
+        44, 47, 51, 56, 61, 66, 71, 75, 79, 82, 85, 88, 91, 91, 58, 52, 47, 50,
+        54, 58, 63, 68, 73, 77, 82, 85, 89, 92, 95, 95, 62, 55, 50, 53, 56, 61,
+        65, 70, 76, 80, 85, 89, 93, 96, 99, 99, 66, 59, 53, 56, 59, 63, 68, 73,
+        78, 83, 88, 92, 96, 99, 103, 103, 70, 63, 57, 60, 63, 67, 71, 75, 81,
+        85, 91, 95, 99, 103, 107, 107, 70, 63, 57, 60, 63, 67, 71, 75, 81, 85,
+        91, 95, 99, 103, 107, 107,
+        /* Size 32 */
+        28, 26, 25, 23, 22, 25, 29, 34, 41, 42, 43, 43, 44, 45, 46, 46, 47, 49,
+        50, 52, 53, 55, 57, 59, 61, 63, 65, 67, 69, 69, 69, 69, 26, 26, 25, 25,
+        24, 27, 31, 35, 42, 42, 42, 42, 42, 43, 44, 44, 45, 46, 48, 49, 50, 52,
+        54, 55, 57, 59, 61, 63, 66, 66, 66, 66, 25, 25, 26, 26, 26, 29, 32, 37,
+        42, 42, 41, 41, 41, 41, 42, 42, 43, 44, 45, 46, 48, 49, 51, 53, 54, 56,
+        58, 60, 62, 62, 62, 62, 23, 25, 26, 28, 29, 32, 34, 38, 42, 41, 41, 40,
+        39, 40, 40, 41, 41, 42, 43, 44, 45, 47, 48, 50, 52, 53, 55, 57, 59, 59,
+        59, 59, 22, 24, 26, 29, 33, 35, 37, 39, 42, 41, 40, 39, 38, 38, 39, 39,
+        39, 40, 41, 42, 43, 45, 46, 48, 49, 51, 53, 54, 56, 56, 56, 56, 25, 27,
+        29, 32, 35, 37, 39, 41, 44, 43, 42, 41, 40, 40, 41, 41, 41, 42, 43, 44,
+        45, 46, 48, 49, 51, 52, 54, 56, 58, 58, 58, 58, 29, 31, 32, 34, 37, 39,
+        41, 43, 46, 45, 44, 43, 43, 43, 43, 43, 43, 44, 45, 46, 47, 48, 49, 51,
+        52, 54, 55, 57, 59, 59, 59, 59, 34, 35, 37, 38, 39, 41, 43, 45, 48, 47,
+        46, 46, 45, 45, 45, 45, 45, 46, 47, 48, 49, 50, 51, 52, 54, 55, 57, 58,
+        60, 60, 60, 60, 41, 42, 42, 42, 42, 44, 46, 48, 50, 49, 49, 49, 48, 48,
+        48, 48, 48, 49, 49, 50, 51, 52, 53, 54, 55, 57, 58, 60, 62, 62, 62, 62,
+        42, 42, 42, 41, 41, 43, 45, 47, 49, 50, 50, 50, 50, 50, 50, 50, 50, 51,
+        51, 52, 53, 54, 55, 56, 57, 59, 60, 62, 64, 64, 64, 64, 43, 42, 41, 41,
+        40, 42, 44, 46, 49, 50, 50, 51, 51, 52, 52, 52, 52, 53, 54, 54, 55, 56,
+        57, 58, 60, 61, 62, 64, 65, 65, 65, 65, 43, 42, 41, 40, 39, 41, 43, 46,
+        49, 50, 51, 52, 53, 54, 54, 54, 55, 56, 56, 57, 58, 59, 60, 61, 62, 63,
+        65, 66, 68, 68, 68, 68, 44, 42, 41, 39, 38, 40, 43, 45, 48, 50, 51, 53,
+        55, 56, 56, 57, 58, 58, 59, 60, 61, 61, 62, 63, 64, 66, 67, 68, 70, 70,
+        70, 70, 45, 43, 41, 40, 38, 40, 43, 45, 48, 50, 52, 54, 56, 57, 57, 58,
+        59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 71, 72, 72, 72, 72, 46, 44,
+        42, 40, 39, 41, 43, 45, 48, 50, 52, 54, 56, 57, 59, 60, 61, 62, 63, 64,
+        65, 66, 67, 68, 69, 70, 72, 73, 74, 74, 74, 74, 46, 44, 42, 41, 39, 41,
+        43, 45, 48, 50, 52, 54, 57, 58, 60, 61, 63, 64, 65, 66, 67, 68, 70, 71,
+        72, 73, 74, 75, 77, 77, 77, 77, 47, 45, 43, 41, 39, 41, 43, 45, 48, 50,
+        52, 55, 58, 59, 61, 63, 65, 66, 67, 69, 70, 71, 72, 73, 75, 76, 77, 78,
+        79, 79, 79, 79, 49, 46, 44, 42, 40, 42, 44, 46, 49, 51, 53, 56, 58, 60,
+        62, 64, 66, 68, 69, 70, 72, 73, 74, 75, 77, 78, 79, 80, 82, 82, 82, 82,
+        50, 48, 45, 43, 41, 43, 45, 47, 49, 51, 54, 56, 59, 61, 63, 65, 67, 69,
+        70, 72, 74, 75, 76, 78, 79, 80, 81, 83, 84, 84, 84, 84, 52, 49, 46, 44,
+        42, 44, 46, 48, 50, 52, 54, 57, 60, 62, 64, 66, 69, 70, 72, 74, 76, 77,
+        78, 80, 81, 83, 84, 85, 87, 87, 87, 87, 53, 50, 48, 45, 43, 45, 47, 49,
+        51, 53, 55, 58, 61, 63, 65, 67, 70, 72, 74, 76, 78, 79, 81, 82, 84, 85,
+        86, 88, 89, 89, 89, 89, 55, 52, 49, 47, 45, 46, 48, 50, 52, 54, 56, 59,
+        61, 64, 66, 68, 71, 73, 75, 77, 79, 81, 82, 84, 86, 87, 88, 90, 91, 91,
+        91, 91, 57, 54, 51, 48, 46, 48, 49, 51, 53, 55, 57, 60, 62, 65, 67, 70,
+        72, 74, 76, 78, 81, 82, 84, 86, 87, 89, 90, 92, 93, 93, 93, 93, 59, 55,
+        53, 50, 48, 49, 51, 52, 54, 56, 58, 61, 63, 66, 68, 71, 73, 75, 78, 80,
+        82, 84, 86, 87, 89, 91, 92, 94, 96, 96, 96, 96, 61, 57, 54, 52, 49, 51,
+        52, 54, 55, 57, 60, 62, 64, 67, 69, 72, 75, 77, 79, 81, 84, 86, 87, 89,
+        91, 93, 95, 96, 98, 98, 98, 98, 63, 59, 56, 53, 51, 52, 54, 55, 57, 59,
+        61, 63, 66, 68, 70, 73, 76, 78, 80, 83, 85, 87, 89, 91, 93, 95, 96, 98,
+        100, 100, 100, 100, 65, 61, 58, 55, 53, 54, 55, 57, 58, 60, 62, 65, 67,
+        69, 72, 74, 77, 79, 81, 84, 86, 88, 90, 92, 95, 96, 98, 100, 102, 102,
+        102, 102, 67, 63, 60, 57, 54, 56, 57, 58, 60, 62, 64, 66, 68, 71, 73,
+        75, 78, 80, 83, 85, 88, 90, 92, 94, 96, 98, 100, 102, 103, 103, 103,
+        103, 69, 66, 62, 59, 56, 58, 59, 60, 62, 64, 65, 68, 70, 72, 74, 77, 79,
+        82, 84, 87, 89, 91, 93, 96, 98, 100, 102, 103, 105, 105, 105, 105, 69,
+        66, 62, 59, 56, 58, 59, 60, 62, 64, 65, 68, 70, 72, 74, 77, 79, 82, 84,
+        87, 89, 91, 93, 96, 98, 100, 102, 103, 105, 105, 105, 105, 69, 66, 62,
+        59, 56, 58, 59, 60, 62, 64, 65, 68, 70, 72, 74, 77, 79, 82, 84, 87, 89,
+        91, 93, 96, 98, 100, 102, 103, 105, 105, 105, 105, 69, 66, 62, 59, 56,
+        58, 59, 60, 62, 64, 65, 68, 70, 72, 74, 77, 79, 82, 84, 87, 89, 91, 93,
+        96, 98, 100, 102, 103, 105, 105, 105, 105 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 70, 112, 168, 70, 103, 139, 181, 112, 139, 188, 221, 168, 181, 221,
+        251,
+        /* Size 8 */
+        64, 49, 53, 68, 90, 115, 139, 163, 49, 56, 53, 62, 78, 99, 121, 144, 53,
+        53, 74, 85, 98, 115, 135, 154, 68, 62, 85, 106, 122, 137, 153, 169, 90,
+        78, 98, 122, 142, 158, 172, 185, 115, 99, 115, 137, 158, 174, 188, 199,
+        139, 121, 135, 153, 172, 188, 200, 210, 163, 144, 154, 169, 185, 199,
+        210, 220,
+        /* Size 16 */
+        64, 55, 49, 51, 53, 59, 68, 78, 90, 101, 115, 126, 139, 150, 163, 163,
+        55, 54, 52, 52, 53, 58, 65, 73, 84, 94, 106, 117, 130, 140, 153, 153,
+        49, 52, 56, 54, 53, 57, 62, 69, 78, 87, 99, 109, 121, 132, 144, 144, 51,
+        52, 54, 58, 62, 66, 72, 79, 87, 96, 106, 116, 128, 137, 149, 149, 53,
+        53, 53, 62, 74, 79, 85, 91, 98, 106, 115, 124, 135, 144, 154, 154, 59,
+        58, 57, 66, 79, 86, 94, 101, 109, 117, 125, 134, 143, 152, 161, 161, 68,
+        65, 62, 72, 85, 94, 106, 114, 122, 129, 137, 145, 153, 160, 169, 169,
+        78, 73, 69, 79, 91, 101, 114, 122, 131, 139, 147, 154, 162, 169, 176,
+        176, 90, 84, 78, 87, 98, 109, 122, 131, 142, 149, 158, 164, 172, 178,
+        185, 185, 101, 94, 87, 96, 106, 117, 129, 139, 149, 157, 166, 172, 179,
+        185, 191, 191, 115, 106, 99, 106, 115, 125, 137, 147, 158, 166, 174,
+        181, 188, 193, 199, 199, 126, 117, 109, 116, 124, 134, 145, 154, 164,
+        172, 181, 187, 194, 199, 204, 204, 139, 130, 121, 128, 135, 143, 153,
+        162, 172, 179, 188, 194, 200, 205, 210, 210, 150, 140, 132, 137, 144,
+        152, 160, 169, 178, 185, 193, 199, 205, 210, 215, 215, 163, 153, 144,
+        149, 154, 161, 169, 176, 185, 191, 199, 204, 210, 215, 220, 220, 163,
+        153, 144, 149, 154, 161, 169, 176, 185, 191, 199, 204, 210, 215, 220,
+        220,
+        /* Size 32 */
+        64, 59, 55, 52, 49, 50, 51, 52, 53, 56, 59, 64, 68, 73, 78, 83, 90, 95,
+        101, 107, 115, 120, 126, 132, 139, 145, 150, 156, 163, 163, 163, 163,
+        59, 57, 54, 52, 50, 51, 51, 52, 53, 56, 59, 63, 67, 71, 75, 81, 87, 92,
+        97, 103, 110, 115, 121, 127, 134, 140, 145, 151, 158, 158, 158, 158, 55,
+        54, 54, 53, 52, 52, 52, 53, 53, 55, 58, 62, 65, 69, 73, 78, 84, 88, 94,
+        99, 106, 111, 117, 123, 130, 135, 140, 146, 153, 153, 153, 153, 52, 52,
+        53, 53, 54, 54, 53, 53, 53, 55, 58, 61, 63, 67, 71, 76, 81, 85, 90, 96,
+        102, 107, 113, 119, 125, 130, 136, 142, 148, 148, 148, 148, 49, 50, 52,
+        54, 56, 55, 54, 54, 53, 55, 57, 60, 62, 65, 69, 73, 78, 83, 87, 93, 99,
+        104, 109, 115, 121, 126, 132, 137, 144, 144, 144, 144, 50, 51, 52, 54,
+        55, 56, 56, 57, 57, 59, 62, 64, 67, 70, 74, 78, 82, 87, 91, 97, 102,
+        107, 112, 118, 124, 129, 134, 140, 146, 146, 146, 146, 51, 51, 52, 53,
+        54, 56, 58, 60, 62, 64, 66, 69, 72, 75, 79, 83, 87, 91, 96, 101, 106,
+        111, 116, 122, 128, 132, 137, 143, 149, 149, 149, 149, 52, 52, 53, 53,
+        54, 57, 60, 63, 67, 70, 72, 75, 78, 81, 84, 88, 92, 96, 101, 105, 111,
+        115, 120, 125, 131, 136, 140, 146, 151, 151, 151, 151, 53, 53, 53, 53,
+        53, 57, 62, 67, 74, 76, 79, 82, 85, 88, 91, 95, 98, 102, 106, 111, 115,
+        120, 124, 129, 135, 139, 144, 149, 154, 154, 154, 154, 56, 56, 55, 55,
+        55, 59, 64, 70, 76, 79, 82, 86, 89, 93, 96, 99, 103, 107, 111, 115, 120,
+        124, 129, 134, 139, 143, 148, 152, 158, 158, 158, 158, 59, 59, 58, 58,
+        57, 62, 66, 72, 79, 82, 86, 90, 94, 98, 101, 105, 109, 113, 117, 121,
+        125, 129, 134, 138, 143, 147, 152, 156, 161, 161, 161, 161, 64, 63, 62,
+        61, 60, 64, 69, 75, 82, 86, 90, 95, 100, 103, 107, 111, 115, 119, 122,
+        127, 131, 135, 139, 143, 148, 152, 156, 160, 165, 165, 165, 165, 68, 67,
+        65, 63, 62, 67, 72, 78, 85, 89, 94, 100, 106, 110, 114, 118, 122, 125,
+        129, 133, 137, 141, 145, 149, 153, 157, 160, 165, 169, 169, 169, 169,
+        73, 71, 69, 67, 65, 70, 75, 81, 88, 93, 98, 103, 110, 113, 118, 122,
+        126, 130, 134, 138, 142, 145, 149, 153, 157, 161, 164, 168, 173, 173,
+        173, 173, 78, 75, 73, 71, 69, 74, 79, 84, 91, 96, 101, 107, 114, 118,
+        122, 126, 131, 135, 139, 142, 147, 150, 154, 158, 162, 165, 169, 172,
+        176, 176, 176, 176, 83, 81, 78, 76, 73, 78, 83, 88, 95, 99, 105, 111,
+        118, 122, 126, 131, 136, 140, 144, 148, 152, 155, 159, 163, 166, 170,
+        173, 177, 180, 180, 180, 180, 90, 87, 84, 81, 78, 82, 87, 92, 98, 103,
+        109, 115, 122, 126, 131, 136, 142, 146, 149, 153, 158, 161, 164, 168,
+        172, 175, 178, 181, 185, 185, 185, 185, 95, 92, 88, 85, 83, 87, 91, 96,
+        102, 107, 113, 119, 125, 130, 135, 140, 146, 149, 153, 157, 162, 165,
+        168, 172, 175, 178, 181, 185, 188, 188, 188, 188, 101, 97, 94, 90, 87,
+        91, 96, 101, 106, 111, 117, 122, 129, 134, 139, 144, 149, 153, 157, 161,
+        166, 169, 172, 176, 179, 182, 185, 188, 191, 191, 191, 191, 107, 103,
+        99, 96, 93, 97, 101, 105, 111, 115, 121, 127, 133, 138, 142, 148, 153,
+        157, 161, 165, 170, 173, 176, 180, 183, 186, 189, 192, 195, 195, 195,
+        195, 115, 110, 106, 102, 99, 102, 106, 111, 115, 120, 125, 131, 137,
+        142, 147, 152, 158, 162, 166, 170, 174, 177, 181, 184, 188, 190, 193,
+        196, 199, 199, 199, 199, 120, 115, 111, 107, 104, 107, 111, 115, 120,
+        124, 129, 135, 141, 145, 150, 155, 161, 165, 169, 173, 177, 181, 184,
+        187, 191, 193, 196, 199, 202, 202, 202, 202, 126, 121, 117, 113, 109,
+        112, 116, 120, 124, 129, 134, 139, 145, 149, 154, 159, 164, 168, 172,
+        176, 181, 184, 187, 190, 194, 196, 199, 202, 204, 204, 204, 204, 132,
+        127, 123, 119, 115, 118, 122, 125, 129, 134, 138, 143, 149, 153, 158,
+        163, 168, 172, 176, 180, 184, 187, 190, 194, 197, 199, 202, 205, 207,
+        207, 207, 207, 139, 134, 130, 125, 121, 124, 128, 131, 135, 139, 143,
+        148, 153, 157, 162, 166, 172, 175, 179, 183, 188, 191, 194, 197, 200,
+        203, 205, 208, 210, 210, 210, 210, 145, 140, 135, 130, 126, 129, 132,
+        136, 139, 143, 147, 152, 157, 161, 165, 170, 175, 178, 182, 186, 190,
+        193, 196, 199, 203, 205, 208, 210, 213, 213, 213, 213, 150, 145, 140,
+        136, 132, 134, 137, 140, 144, 148, 152, 156, 160, 164, 169, 173, 178,
+        181, 185, 189, 193, 196, 199, 202, 205, 208, 210, 212, 215, 215, 215,
+        215, 156, 151, 146, 142, 137, 140, 143, 146, 149, 152, 156, 160, 165,
+        168, 172, 177, 181, 185, 188, 192, 196, 199, 202, 205, 208, 210, 212,
+        215, 217, 217, 217, 217, 163, 158, 153, 148, 144, 146, 149, 151, 154,
+        158, 161, 165, 169, 173, 176, 180, 185, 188, 191, 195, 199, 202, 204,
+        207, 210, 213, 215, 217, 220, 220, 220, 220, 163, 158, 153, 148, 144,
+        146, 149, 151, 154, 158, 161, 165, 169, 173, 176, 180, 185, 188, 191,
+        195, 199, 202, 204, 207, 210, 213, 215, 217, 220, 220, 220, 220, 163,
+        158, 153, 148, 144, 146, 149, 151, 154, 158, 161, 165, 169, 173, 176,
+        180, 185, 188, 191, 195, 199, 202, 204, 207, 210, 213, 215, 217, 220,
+        220, 220, 220, 163, 158, 153, 148, 144, 146, 149, 151, 154, 158, 161,
+        165, 169, 173, 176, 180, 185, 188, 191, 195, 199, 202, 204, 207, 210,
+        213, 215, 217, 220, 220, 220, 220 },
+      { /* Intra matrices */
+        /* Size 4 */
+        24, 26, 43, 67, 26, 39, 54, 72, 43, 54, 75, 91, 67, 72, 91, 105,
+        /* Size 8 */
+        28, 21, 23, 30, 41, 53, 65, 77, 21, 25, 23, 27, 35, 45, 56, 68, 23, 23,
+        33, 38, 45, 53, 63, 73, 30, 27, 38, 49, 56, 64, 72, 81, 41, 35, 45, 56,
+        67, 75, 82, 89, 53, 45, 53, 64, 75, 84, 91, 97, 65, 56, 63, 72, 82, 91,
+        98, 104, 77, 68, 73, 81, 89, 97, 104, 109,
+        /* Size 16 */
+        27, 24, 21, 21, 22, 25, 29, 34, 39, 44, 51, 56, 63, 68, 75, 75, 24, 23,
+        22, 22, 22, 25, 28, 32, 36, 41, 47, 52, 58, 63, 70, 70, 21, 22, 24, 23,
+        23, 24, 27, 30, 34, 38, 43, 48, 54, 59, 65, 65, 21, 22, 23, 25, 26, 28,
+        31, 34, 38, 42, 47, 52, 57, 62, 68, 68, 22, 22, 23, 26, 32, 34, 37, 40,
+        43, 47, 51, 55, 61, 65, 70, 70, 25, 25, 24, 28, 34, 37, 41, 44, 48, 52,
+        56, 60, 65, 69, 74, 74, 29, 28, 27, 31, 37, 41, 47, 50, 54, 58, 62, 65,
+        70, 74, 78, 78, 34, 32, 30, 34, 40, 44, 50, 54, 59, 62, 67, 70, 74, 78,
+        82, 82, 39, 36, 34, 38, 43, 48, 54, 59, 64, 68, 72, 76, 79, 83, 86, 86,
+        44, 41, 38, 42, 47, 52, 58, 62, 68, 72, 76, 80, 83, 86, 90, 90, 51, 47,
+        43, 47, 51, 56, 62, 67, 72, 76, 81, 84, 88, 91, 94, 94, 56, 52, 48, 52,
+        55, 60, 65, 70, 76, 80, 84, 87, 91, 94, 97, 97, 63, 58, 54, 57, 61, 65,
+        70, 74, 79, 83, 88, 91, 95, 97, 100, 100, 68, 63, 59, 62, 65, 69, 74,
+        78, 83, 86, 91, 94, 97, 100, 103, 103, 75, 70, 65, 68, 70, 74, 78, 82,
+        86, 90, 94, 97, 100, 103, 105, 105, 75, 70, 65, 68, 70, 74, 78, 82, 86,
+        90, 94, 97, 100, 103, 105, 105,
+        /* Size 32 */
+        27, 25, 23, 22, 20, 21, 21, 21, 22, 23, 25, 27, 29, 31, 33, 36, 39, 41,
+        43, 46, 50, 52, 55, 58, 62, 64, 67, 70, 73, 73, 73, 73, 25, 24, 23, 22,
+        21, 21, 21, 22, 22, 23, 25, 26, 28, 30, 32, 34, 37, 39, 42, 45, 48, 50,
+        53, 56, 59, 62, 65, 67, 71, 71, 71, 71, 23, 23, 22, 22, 22, 22, 22, 22,
+        22, 23, 24, 26, 27, 29, 31, 33, 36, 38, 40, 43, 46, 48, 51, 54, 57, 60,
+        62, 65, 68, 68, 68, 68, 22, 22, 22, 22, 22, 22, 22, 22, 22, 23, 24, 25,
+        27, 28, 30, 32, 34, 36, 39, 41, 44, 46, 49, 52, 55, 57, 60, 63, 66, 66,
+        66, 66, 20, 21, 22, 22, 23, 23, 23, 22, 22, 23, 24, 25, 26, 28, 29, 31,
+        33, 35, 37, 40, 43, 45, 47, 50, 53, 55, 58, 61, 64, 64, 64, 64, 21, 21,
+        22, 22, 23, 23, 23, 24, 24, 25, 26, 27, 28, 30, 31, 33, 35, 37, 39, 42,
+        44, 46, 49, 52, 54, 57, 59, 62, 65, 65, 65, 65, 21, 21, 22, 22, 23, 23,
+        24, 25, 26, 27, 28, 29, 30, 32, 33, 35, 37, 39, 41, 43, 46, 48, 51, 53,
+        56, 58, 61, 63, 66, 66, 66, 66, 21, 22, 22, 22, 22, 24, 25, 27, 28, 29,
+        30, 32, 33, 34, 36, 38, 40, 41, 43, 46, 48, 50, 52, 55, 58, 60, 62, 65,
+        68, 68, 68, 68, 22, 22, 22, 22, 22, 24, 26, 28, 31, 32, 33, 35, 36, 38,
+        39, 41, 42, 44, 46, 48, 50, 52, 54, 57, 59, 62, 64, 66, 69, 69, 69, 69,
+        23, 23, 23, 23, 23, 25, 27, 29, 32, 34, 35, 37, 38, 40, 41, 43, 45, 46,
+        48, 50, 52, 54, 57, 59, 61, 64, 66, 68, 71, 71, 71, 71, 25, 25, 24, 24,
+        24, 26, 28, 30, 33, 35, 37, 39, 41, 42, 44, 45, 47, 49, 51, 53, 55, 57,
+        59, 61, 64, 66, 68, 70, 73, 73, 73, 73, 27, 26, 26, 25, 25, 27, 29, 32,
+        35, 37, 39, 41, 43, 45, 46, 48, 50, 52, 54, 56, 58, 60, 61, 64, 66, 68,
+        70, 72, 74, 74, 74, 74, 29, 28, 27, 27, 26, 28, 30, 33, 36, 38, 41, 43,
+        46, 48, 49, 51, 53, 55, 57, 59, 61, 62, 64, 66, 68, 70, 72, 74, 76, 76,
+        76, 76, 31, 30, 29, 28, 28, 30, 32, 34, 38, 40, 42, 45, 48, 49, 51, 53,
+        56, 57, 59, 61, 63, 65, 66, 68, 71, 72, 74, 76, 78, 78, 78, 78, 33, 32,
+        31, 30, 29, 31, 33, 36, 39, 41, 44, 46, 49, 51, 53, 55, 58, 60, 61, 63,
+        65, 67, 69, 71, 73, 75, 76, 78, 80, 80, 80, 80, 36, 34, 33, 32, 31, 33,
+        35, 38, 41, 43, 45, 48, 51, 53, 55, 58, 60, 62, 64, 66, 68, 70, 71, 73,
+        75, 77, 79, 80, 82, 82, 82, 82, 39, 37, 36, 34, 33, 35, 37, 40, 42, 45,
+        47, 50, 53, 56, 58, 60, 63, 65, 67, 69, 71, 72, 74, 76, 78, 79, 81, 83,
+        85, 85, 85, 85, 41, 39, 38, 36, 35, 37, 39, 41, 44, 46, 49, 52, 55, 57,
+        60, 62, 65, 67, 69, 71, 73, 74, 76, 78, 80, 81, 83, 85, 86, 86, 86, 86,
+        43, 42, 40, 39, 37, 39, 41, 43, 46, 48, 51, 54, 57, 59, 61, 64, 67, 69,
+        70, 73, 75, 76, 78, 80, 82, 83, 85, 86, 88, 88, 88, 88, 46, 45, 43, 41,
+        40, 42, 43, 46, 48, 50, 53, 56, 59, 61, 63, 66, 69, 71, 73, 75, 77, 79,
+        80, 82, 84, 85, 87, 88, 90, 90, 90, 90, 50, 48, 46, 44, 43, 44, 46, 48,
+        50, 52, 55, 58, 61, 63, 65, 68, 71, 73, 75, 77, 79, 81, 83, 84, 86, 87,
+        89, 90, 92, 92, 92, 92, 52, 50, 48, 46, 45, 46, 48, 50, 52, 54, 57, 60,
+        62, 65, 67, 70, 72, 74, 76, 79, 81, 82, 84, 86, 88, 89, 90, 92, 93, 93,
+        93, 93, 55, 53, 51, 49, 47, 49, 51, 52, 54, 57, 59, 61, 64, 66, 69, 71,
+        74, 76, 78, 80, 83, 84, 86, 88, 89, 91, 92, 94, 95, 95, 95, 95, 58, 56,
+        54, 52, 50, 52, 53, 55, 57, 59, 61, 64, 66, 68, 71, 73, 76, 78, 80, 82,
+        84, 86, 88, 89, 91, 92, 94, 95, 97, 97, 97, 97, 62, 59, 57, 55, 53, 54,
+        56, 58, 59, 61, 64, 66, 68, 71, 73, 75, 78, 80, 82, 84, 86, 88, 89, 91,
+        93, 94, 95, 97, 98, 98, 98, 98, 64, 62, 60, 57, 55, 57, 58, 60, 62, 64,
+        66, 68, 70, 72, 75, 77, 79, 81, 83, 85, 87, 89, 91, 92, 94, 95, 97, 98,
+        99, 99, 99, 99, 67, 65, 62, 60, 58, 59, 61, 62, 64, 66, 68, 70, 72, 74,
+        76, 79, 81, 83, 85, 87, 89, 90, 92, 94, 95, 97, 98, 99, 101, 101, 101,
+        101, 70, 67, 65, 63, 61, 62, 63, 65, 66, 68, 70, 72, 74, 76, 78, 80, 83,
+        85, 86, 88, 90, 92, 94, 95, 97, 98, 99, 101, 102, 102, 102, 102, 73, 71,
+        68, 66, 64, 65, 66, 68, 69, 71, 73, 74, 76, 78, 80, 82, 85, 86, 88, 90,
+        92, 93, 95, 97, 98, 99, 101, 102, 103, 103, 103, 103, 73, 71, 68, 66,
+        64, 65, 66, 68, 69, 71, 73, 74, 76, 78, 80, 82, 85, 86, 88, 90, 92, 93,
+        95, 97, 98, 99, 101, 102, 103, 103, 103, 103, 73, 71, 68, 66, 64, 65,
+        66, 68, 69, 71, 73, 74, 76, 78, 80, 82, 85, 86, 88, 90, 92, 93, 95, 97,
+        98, 99, 101, 102, 103, 103, 103, 103, 73, 71, 68, 66, 64, 65, 66, 68,
+        69, 71, 73, 74, 76, 78, 80, 82, 85, 86, 88, 90, 92, 93, 95, 97, 98, 99,
+        101, 102, 103, 103, 103, 103 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 99, 107, 133, 99, 117, 124, 141, 107, 124, 156, 175, 133, 141, 175,
+        203,
+        /* Size 8 */
+        64, 52, 92, 97, 103, 114, 128, 143, 52, 74, 93, 85, 87, 96, 107, 120,
+        92, 93, 108, 105, 104, 109, 118, 129, 97, 85, 105, 117, 122, 127, 134,
+        143, 103, 87, 104, 122, 135, 144, 151, 159, 114, 96, 109, 127, 144, 156,
+        166, 175, 128, 107, 118, 134, 151, 166, 178, 188, 143, 120, 129, 143,
+        159, 175, 188, 198,
+        /* Size 16 */
+        64, 57, 52, 66, 92, 94, 97, 100, 103, 108, 114, 121, 128, 135, 143, 143,
+        57, 59, 61, 74, 93, 92, 91, 93, 95, 99, 104, 110, 116, 123, 130, 130,
+        52, 61, 74, 83, 93, 89, 85, 86, 87, 91, 96, 101, 107, 113, 120, 120, 66,
+        74, 83, 91, 100, 97, 94, 95, 95, 98, 102, 107, 112, 118, 124, 124, 92,
+        93, 93, 100, 108, 106, 105, 104, 104, 107, 109, 114, 118, 123, 129, 129,
+        94, 92, 89, 97, 106, 108, 111, 112, 113, 115, 118, 122, 126, 131, 136,
+        136, 97, 91, 85, 94, 105, 111, 117, 120, 122, 125, 127, 131, 134, 139,
+        143, 143, 100, 93, 86, 95, 104, 112, 120, 124, 129, 132, 135, 139, 142,
+        146, 151, 151, 103, 95, 87, 95, 104, 113, 122, 129, 135, 139, 144, 147,
+        151, 155, 159, 159, 108, 99, 91, 98, 107, 115, 125, 132, 139, 144, 150,
+        154, 158, 162, 167, 167, 114, 104, 96, 102, 109, 118, 127, 135, 144,
+        150, 156, 161, 166, 170, 175, 175, 121, 110, 101, 107, 114, 122, 131,
+        139, 147, 154, 161, 166, 172, 176, 181, 181, 128, 116, 107, 112, 118,
+        126, 134, 142, 151, 158, 166, 172, 178, 183, 188, 188, 135, 123, 113,
+        118, 123, 131, 139, 146, 155, 162, 170, 176, 183, 188, 193, 193, 143,
+        130, 120, 124, 129, 136, 143, 151, 159, 167, 175, 181, 188, 193, 198,
+        198, 143, 130, 120, 124, 129, 136, 143, 151, 159, 167, 175, 181, 188,
+        193, 198, 198,
+        /* Size 32 */
+        64, 61, 57, 55, 52, 58, 66, 77, 92, 93, 94, 96, 97, 98, 100, 101, 103,
+        106, 108, 111, 114, 117, 121, 124, 128, 131, 135, 139, 143, 143, 143,
+        143, 61, 59, 58, 57, 56, 62, 70, 79, 92, 93, 93, 93, 94, 95, 96, 97, 99,
+        101, 104, 106, 109, 112, 115, 118, 122, 125, 129, 132, 136, 136, 136,
+        136, 57, 58, 59, 60, 61, 67, 74, 82, 93, 92, 92, 91, 91, 92, 93, 94, 95,
+        97, 99, 101, 104, 107, 110, 113, 116, 119, 123, 126, 130, 130, 130, 130,
+        55, 57, 60, 63, 67, 72, 78, 85, 93, 92, 90, 89, 88, 89, 89, 90, 91, 93,
+        95, 97, 100, 102, 105, 108, 111, 114, 118, 121, 125, 125, 125, 125, 52,
+        56, 61, 67, 74, 78, 83, 88, 93, 91, 89, 87, 85, 86, 86, 87, 87, 89, 91,
+        93, 96, 98, 101, 104, 107, 110, 113, 116, 120, 120, 120, 120, 58, 62,
+        67, 72, 78, 82, 86, 91, 97, 95, 93, 91, 89, 90, 90, 91, 91, 93, 95, 97,
+        99, 101, 104, 106, 109, 112, 115, 119, 122, 122, 122, 122, 66, 70, 74,
+        78, 83, 86, 91, 95, 100, 99, 97, 95, 94, 94, 95, 95, 95, 97, 98, 100,
+        102, 104, 107, 109, 112, 115, 118, 121, 124, 124, 124, 124, 77, 79, 82,
+        85, 88, 91, 95, 99, 104, 103, 101, 100, 99, 99, 99, 99, 100, 101, 102,
+        104, 106, 108, 110, 113, 115, 118, 121, 124, 127, 127, 127, 127, 92, 92,
+        93, 93, 93, 97, 100, 104, 108, 107, 106, 105, 105, 105, 104, 104, 104,
+        106, 107, 108, 109, 111, 114, 116, 118, 121, 123, 126, 129, 129, 129,
+        129, 93, 93, 92, 92, 91, 95, 99, 103, 107, 107, 107, 107, 108, 108, 108,
+        108, 108, 110, 111, 112, 113, 115, 117, 120, 122, 124, 127, 130, 133,
+        133, 133, 133, 94, 93, 92, 90, 89, 93, 97, 101, 106, 107, 108, 109, 111,
+        111, 112, 112, 113, 114, 115, 116, 118, 120, 122, 124, 126, 128, 131,
+        133, 136, 136, 136, 136, 96, 93, 91, 89, 87, 91, 95, 100, 105, 107, 109,
+        112, 114, 115, 116, 116, 117, 119, 120, 121, 122, 124, 126, 128, 130,
+        132, 135, 137, 140, 140, 140, 140, 97, 94, 91, 88, 85, 89, 94, 99, 105,
+        108, 111, 114, 117, 118, 120, 121, 122, 124, 125, 126, 127, 129, 131,
+        132, 134, 136, 139, 141, 143, 143, 143, 143, 98, 95, 92, 89, 86, 90, 94,
+        99, 105, 108, 111, 115, 118, 120, 122, 124, 125, 127, 128, 130, 131,
+        133, 134, 136, 138, 140, 142, 145, 147, 147, 147, 147, 100, 96, 93, 89,
+        86, 90, 95, 99, 104, 108, 112, 116, 120, 122, 124, 126, 129, 130, 132,
+        133, 135, 137, 139, 140, 142, 144, 146, 149, 151, 151, 151, 151, 101,
+        97, 94, 90, 87, 91, 95, 99, 104, 108, 112, 116, 121, 124, 126, 129, 132,
+        134, 135, 137, 139, 141, 143, 145, 147, 149, 151, 153, 155, 155, 155,
+        155, 103, 99, 95, 91, 87, 91, 95, 100, 104, 108, 113, 117, 122, 125,
+        129, 132, 135, 137, 139, 142, 144, 146, 147, 149, 151, 153, 155, 157,
+        159, 159, 159, 159, 106, 101, 97, 93, 89, 93, 97, 101, 106, 110, 114,
+        119, 124, 127, 130, 134, 137, 140, 142, 144, 147, 149, 151, 153, 155,
+        157, 159, 161, 163, 163, 163, 163, 108, 104, 99, 95, 91, 95, 98, 102,
+        107, 111, 115, 120, 125, 128, 132, 135, 139, 142, 144, 147, 150, 152,
+        154, 156, 158, 160, 162, 164, 167, 167, 167, 167, 111, 106, 101, 97, 93,
+        97, 100, 104, 108, 112, 116, 121, 126, 130, 133, 137, 142, 144, 147,
+        150, 153, 155, 157, 160, 162, 164, 166, 168, 170, 170, 170, 170, 114,
+        109, 104, 100, 96, 99, 102, 106, 109, 113, 118, 122, 127, 131, 135, 139,
+        144, 147, 150, 153, 156, 159, 161, 164, 166, 168, 170, 172, 175, 175,
+        175, 175, 117, 112, 107, 102, 98, 101, 104, 108, 111, 115, 120, 124,
+        129, 133, 137, 141, 146, 149, 152, 155, 159, 161, 164, 166, 169, 171,
+        173, 175, 178, 178, 178, 178, 121, 115, 110, 105, 101, 104, 107, 110,
+        114, 117, 122, 126, 131, 134, 139, 143, 147, 151, 154, 157, 161, 164,
+        166, 169, 172, 174, 176, 178, 181, 181, 181, 181, 124, 118, 113, 108,
+        104, 106, 109, 113, 116, 120, 124, 128, 132, 136, 140, 145, 149, 153,
+        156, 160, 164, 166, 169, 172, 175, 177, 179, 182, 184, 184, 184, 184,
+        128, 122, 116, 111, 107, 109, 112, 115, 118, 122, 126, 130, 134, 138,
+        142, 147, 151, 155, 158, 162, 166, 169, 172, 175, 178, 180, 183, 185,
+        188, 188, 188, 188, 131, 125, 119, 114, 110, 112, 115, 118, 121, 124,
+        128, 132, 136, 140, 144, 149, 153, 157, 160, 164, 168, 171, 174, 177,
+        180, 183, 185, 188, 190, 190, 190, 190, 135, 129, 123, 118, 113, 115,
+        118, 121, 123, 127, 131, 135, 139, 142, 146, 151, 155, 159, 162, 166,
+        170, 173, 176, 179, 183, 185, 188, 190, 193, 193, 193, 193, 139, 132,
+        126, 121, 116, 119, 121, 124, 126, 130, 133, 137, 141, 145, 149, 153,
+        157, 161, 164, 168, 172, 175, 178, 182, 185, 188, 190, 193, 196, 196,
+        196, 196, 143, 136, 130, 125, 120, 122, 124, 127, 129, 133, 136, 140,
+        143, 147, 151, 155, 159, 163, 167, 170, 175, 178, 181, 184, 188, 190,
+        193, 196, 198, 198, 198, 198, 143, 136, 130, 125, 120, 122, 124, 127,
+        129, 133, 136, 140, 143, 147, 151, 155, 159, 163, 167, 170, 175, 178,
+        181, 184, 188, 190, 193, 196, 198, 198, 198, 198, 143, 136, 130, 125,
+        120, 122, 124, 127, 129, 133, 136, 140, 143, 147, 151, 155, 159, 163,
+        167, 170, 175, 178, 181, 184, 188, 190, 193, 196, 198, 198, 198, 198,
+        143, 136, 130, 125, 120, 122, 124, 127, 129, 133, 136, 140, 143, 147,
+        151, 155, 159, 163, 167, 170, 175, 178, 181, 184, 188, 190, 193, 196,
+        198, 198, 198, 198 },
+      { /* Intra matrices */
+        /* Size 4 */
+        29, 46, 49, 62, 46, 54, 58, 66, 49, 58, 74, 84, 62, 66, 84, 99,
+        /* Size 8 */
+        31, 25, 45, 48, 51, 57, 64, 72, 25, 36, 46, 42, 43, 47, 53, 60, 45, 46,
+        54, 52, 52, 54, 59, 65, 48, 42, 52, 59, 61, 64, 68, 73, 51, 43, 52, 61,
+        68, 73, 77, 82, 57, 47, 54, 64, 73, 80, 86, 90, 64, 53, 59, 68, 77, 86,
+        92, 98, 72, 60, 65, 73, 82, 90, 98, 104,
+        /* Size 16 */
+        30, 27, 24, 31, 44, 45, 47, 48, 50, 53, 56, 59, 63, 66, 71, 71, 27, 28,
+        29, 35, 44, 44, 43, 44, 45, 48, 50, 53, 57, 60, 64, 64, 24, 29, 35, 39,
+        45, 43, 41, 41, 42, 44, 46, 49, 52, 55, 58, 58, 31, 35, 39, 43, 48, 47,
+        45, 45, 46, 47, 49, 52, 54, 58, 61, 61, 44, 44, 45, 48, 52, 51, 51, 51,
+        50, 52, 53, 55, 58, 60, 63, 63, 45, 44, 43, 47, 51, 53, 54, 54, 55, 56,
+        57, 59, 62, 64, 67, 67, 47, 43, 41, 45, 51, 54, 57, 58, 60, 61, 62, 64,
+        66, 68, 71, 71, 48, 44, 41, 45, 51, 54, 58, 61, 63, 65, 67, 68, 70, 73,
+        75, 75, 50, 45, 42, 46, 50, 55, 60, 63, 67, 69, 71, 73, 75, 77, 80, 80,
+        53, 48, 44, 47, 52, 56, 61, 65, 69, 72, 74, 77, 79, 81, 84, 84, 56, 50,
+        46, 49, 53, 57, 62, 67, 71, 74, 78, 81, 83, 86, 88, 88, 59, 53, 49, 52,
+        55, 59, 64, 68, 73, 77, 81, 83, 87, 89, 92, 92, 63, 57, 52, 54, 58, 62,
+        66, 70, 75, 79, 83, 87, 90, 93, 95, 95, 66, 60, 55, 58, 60, 64, 68, 73,
+        77, 81, 86, 89, 93, 95, 98, 98, 71, 64, 58, 61, 63, 67, 71, 75, 80, 84,
+        88, 92, 95, 98, 102, 102, 71, 64, 58, 61, 63, 67, 71, 75, 80, 84, 88,
+        92, 95, 98, 102, 102,
+        /* Size 32 */
+        30, 28, 27, 25, 24, 27, 31, 36, 43, 44, 45, 45, 46, 47, 48, 48, 49, 50,
+        52, 53, 55, 56, 58, 60, 62, 64, 65, 67, 70, 70, 70, 70, 28, 28, 27, 27,
+        26, 29, 33, 37, 44, 44, 44, 44, 44, 45, 46, 46, 47, 48, 49, 51, 52, 54,
+        55, 57, 59, 60, 62, 64, 66, 66, 66, 66, 27, 27, 28, 28, 28, 31, 35, 39,
+        44, 44, 43, 43, 43, 43, 44, 44, 45, 46, 47, 48, 50, 51, 53, 54, 56, 58,
+        59, 61, 63, 63, 63, 63, 25, 27, 28, 30, 31, 34, 37, 40, 44, 43, 43, 42,
+        41, 42, 42, 43, 43, 44, 45, 46, 47, 49, 50, 52, 53, 55, 57, 58, 60, 60,
+        60, 60, 24, 26, 28, 31, 35, 37, 39, 41, 44, 43, 42, 41, 40, 40, 41, 41,
+        41, 42, 43, 44, 45, 47, 48, 49, 51, 53, 54, 56, 58, 58, 58, 58, 27, 29,
+        31, 34, 37, 39, 41, 43, 46, 45, 44, 43, 42, 42, 43, 43, 43, 44, 45, 46,
+        47, 48, 49, 51, 52, 54, 55, 57, 59, 59, 59, 59, 31, 33, 35, 37, 39, 41,
+        43, 45, 48, 47, 46, 45, 45, 45, 45, 45, 45, 46, 47, 48, 49, 50, 51, 52,
+        54, 55, 57, 58, 60, 60, 60, 60, 36, 37, 39, 40, 41, 43, 45, 47, 50, 49,
+        48, 48, 47, 47, 47, 47, 47, 48, 49, 50, 50, 52, 53, 54, 55, 57, 58, 60,
+        61, 61, 61, 61, 43, 44, 44, 44, 44, 46, 48, 50, 52, 51, 51, 50, 50, 50,
+        50, 50, 50, 50, 51, 52, 52, 53, 54, 56, 57, 58, 60, 61, 63, 63, 63, 63,
+        44, 44, 44, 43, 43, 45, 47, 49, 51, 51, 51, 51, 51, 51, 52, 52, 52, 52,
+        53, 54, 54, 55, 56, 58, 59, 60, 61, 63, 64, 64, 64, 64, 45, 44, 43, 43,
+        42, 44, 46, 48, 51, 51, 52, 52, 53, 53, 53, 54, 54, 55, 55, 56, 57, 58,
+        59, 60, 61, 62, 63, 65, 66, 66, 66, 66, 45, 44, 43, 42, 41, 43, 45, 48,
+        50, 51, 52, 53, 55, 55, 55, 56, 56, 57, 58, 58, 59, 60, 61, 62, 63, 64,
+        65, 67, 68, 68, 68, 68, 46, 44, 43, 41, 40, 42, 45, 47, 50, 51, 53, 55,
+        56, 57, 58, 58, 59, 60, 60, 61, 62, 62, 63, 64, 65, 66, 68, 69, 70, 70,
+        70, 70, 47, 45, 43, 42, 40, 42, 45, 47, 50, 51, 53, 55, 57, 58, 59, 60,
+        61, 61, 62, 63, 64, 64, 65, 66, 67, 68, 70, 71, 72, 72, 72, 72, 48, 46,
+        44, 42, 41, 43, 45, 47, 50, 52, 53, 55, 58, 59, 60, 61, 62, 63, 64, 65,
+        66, 67, 67, 68, 69, 71, 72, 73, 74, 74, 74, 74, 48, 46, 44, 43, 41, 43,
+        45, 47, 50, 52, 54, 56, 58, 60, 61, 62, 64, 65, 66, 67, 68, 69, 70, 71,
+        72, 73, 74, 75, 76, 76, 76, 76, 49, 47, 45, 43, 41, 43, 45, 47, 50, 52,
+        54, 56, 59, 61, 62, 64, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77,
+        79, 79, 79, 79, 50, 48, 46, 44, 42, 44, 46, 48, 50, 52, 55, 57, 60, 61,
+        63, 65, 67, 68, 69, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 80, 80, 80,
+        52, 49, 47, 45, 43, 45, 47, 49, 51, 53, 55, 58, 60, 62, 64, 66, 68, 69,
+        71, 72, 73, 75, 76, 77, 78, 79, 80, 81, 82, 82, 82, 82, 53, 51, 48, 46,
+        44, 46, 48, 50, 52, 54, 56, 58, 61, 63, 65, 67, 69, 70, 72, 74, 75, 76,
+        78, 79, 80, 81, 82, 83, 85, 85, 85, 85, 55, 52, 50, 47, 45, 47, 49, 50,
+        52, 54, 57, 59, 62, 64, 66, 68, 70, 72, 73, 75, 77, 78, 80, 81, 82, 83,
+        84, 86, 87, 87, 87, 87, 56, 54, 51, 49, 47, 48, 50, 52, 53, 55, 58, 60,
+        62, 64, 67, 69, 71, 73, 75, 76, 78, 80, 81, 82, 84, 85, 86, 87, 89, 89,
+        89, 89, 58, 55, 53, 50, 48, 49, 51, 53, 54, 56, 59, 61, 63, 65, 67, 70,
+        72, 74, 76, 78, 80, 81, 82, 84, 85, 87, 88, 89, 90, 90, 90, 90, 60, 57,
+        54, 52, 49, 51, 52, 54, 56, 58, 60, 62, 64, 66, 68, 71, 73, 75, 77, 79,
+        81, 82, 84, 85, 87, 88, 90, 91, 92, 92, 92, 92, 62, 59, 56, 53, 51, 52,
+        54, 55, 57, 59, 61, 63, 65, 67, 69, 72, 74, 76, 78, 80, 82, 84, 85, 87,
+        89, 90, 91, 93, 94, 94, 94, 94, 64, 60, 58, 55, 53, 54, 55, 57, 58, 60,
+        62, 64, 66, 68, 71, 73, 75, 77, 79, 81, 83, 85, 87, 88, 90, 91, 93, 94,
+        96, 96, 96, 96, 65, 62, 59, 57, 54, 55, 57, 58, 60, 61, 63, 65, 68, 70,
+        72, 74, 76, 78, 80, 82, 84, 86, 88, 90, 91, 93, 94, 96, 97, 97, 97, 97,
+        67, 64, 61, 58, 56, 57, 58, 60, 61, 63, 65, 67, 69, 71, 73, 75, 77, 79,
+        81, 83, 86, 87, 89, 91, 93, 94, 96, 97, 99, 99, 99, 99, 70, 66, 63, 60,
+        58, 59, 60, 61, 63, 64, 66, 68, 70, 72, 74, 76, 79, 80, 82, 85, 87, 89,
+        90, 92, 94, 96, 97, 99, 100, 100, 100, 100, 70, 66, 63, 60, 58, 59, 60,
+        61, 63, 64, 66, 68, 70, 72, 74, 76, 79, 80, 82, 85, 87, 89, 90, 92, 94,
+        96, 97, 99, 100, 100, 100, 100, 70, 66, 63, 60, 58, 59, 60, 61, 63, 64,
+        66, 68, 70, 72, 74, 76, 79, 80, 82, 85, 87, 89, 90, 92, 94, 96, 97, 99,
+        100, 100, 100, 100, 70, 66, 63, 60, 58, 59, 60, 61, 63, 64, 66, 68, 70,
+        72, 74, 76, 79, 80, 82, 85, 87, 89, 90, 92, 94, 96, 97, 99, 100, 100,
+        100, 100 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 69, 108, 155, 69, 100, 131, 166, 108, 131, 171, 197, 155, 166, 197,
+        218,
+        /* Size 8 */
+        64, 49, 53, 68, 88, 109, 130, 149, 49, 56, 54, 62, 77, 96, 115, 134, 53,
+        54, 73, 83, 95, 110, 126, 142, 68, 62, 83, 102, 116, 128, 141, 154, 88,
+        77, 95, 116, 132, 145, 156, 166, 109, 96, 110, 128, 145, 158, 168, 176,
+        130, 115, 126, 141, 156, 168, 177, 185, 149, 134, 142, 154, 166, 176,
+        185, 191,
+        /* Size 16 */
+        64, 56, 49, 51, 53, 60, 68, 77, 88, 97, 109, 119, 130, 139, 149, 149,
+        56, 54, 52, 53, 53, 59, 65, 73, 82, 91, 102, 111, 122, 131, 141, 141,
+        49, 52, 56, 55, 54, 58, 62, 69, 77, 85, 96, 104, 115, 124, 134, 134, 51,
+        53, 55, 58, 62, 66, 71, 78, 85, 93, 102, 111, 120, 128, 138, 138, 53,
+        53, 54, 62, 73, 78, 83, 89, 95, 102, 110, 117, 126, 134, 142, 142, 60,
+        59, 58, 66, 78, 84, 92, 98, 104, 111, 118, 125, 133, 140, 148, 148, 68,
+        65, 62, 71, 83, 92, 102, 108, 116, 122, 128, 134, 141, 147, 154, 154,
+        77, 73, 69, 78, 89, 98, 108, 115, 123, 129, 136, 142, 148, 153, 159,
+        159, 88, 82, 77, 85, 95, 104, 116, 123, 132, 138, 145, 150, 156, 161,
+        166, 166, 97, 91, 85, 93, 102, 111, 122, 129, 138, 144, 151, 156, 162,
+        166, 171, 171, 109, 102, 96, 102, 110, 118, 128, 136, 145, 151, 158,
+        163, 168, 172, 176, 176, 119, 111, 104, 111, 117, 125, 134, 142, 150,
+        156, 163, 168, 173, 176, 180, 180, 130, 122, 115, 120, 126, 133, 141,
+        148, 156, 162, 168, 173, 177, 181, 185, 185, 139, 131, 124, 128, 134,
+        140, 147, 153, 161, 166, 172, 176, 181, 184, 188, 188, 149, 141, 134,
+        138, 142, 148, 154, 159, 166, 171, 176, 180, 185, 188, 191, 191, 149,
+        141, 134, 138, 142, 148, 154, 159, 166, 171, 176, 180, 185, 188, 191,
+        191,
+        /* Size 32 */
+        64, 60, 56, 52, 49, 50, 51, 52, 53, 56, 60, 64, 68, 72, 77, 82, 88, 92,
+        97, 103, 109, 114, 119, 124, 130, 134, 139, 144, 149, 149, 149, 149, 60,
+        57, 55, 53, 51, 51, 52, 53, 53, 56, 59, 63, 66, 70, 75, 79, 85, 89, 94,
+        100, 106, 110, 115, 120, 126, 130, 135, 140, 145, 145, 145, 145, 56, 55,
+        54, 53, 52, 53, 53, 53, 53, 56, 59, 62, 65, 69, 73, 77, 82, 86, 91, 96,
+        102, 106, 111, 116, 122, 126, 131, 136, 141, 141, 141, 141, 52, 53, 53,
+        54, 54, 54, 54, 54, 54, 56, 58, 61, 64, 67, 71, 75, 80, 84, 88, 93, 99,
+        103, 108, 113, 118, 123, 127, 132, 137, 137, 137, 137, 49, 51, 52, 54,
+        56, 55, 55, 54, 54, 56, 58, 60, 62, 65, 69, 73, 77, 81, 85, 90, 96, 100,
+        104, 109, 115, 119, 124, 128, 134, 134, 134, 134, 50, 51, 53, 54, 55,
+        56, 56, 57, 58, 60, 62, 64, 66, 70, 73, 77, 81, 85, 89, 94, 99, 103,
+        107, 112, 118, 122, 126, 131, 136, 136, 136, 136, 51, 52, 53, 54, 55,
+        56, 58, 60, 62, 64, 66, 69, 71, 74, 78, 81, 85, 89, 93, 97, 102, 106,
+        111, 115, 120, 124, 128, 133, 138, 138, 138, 138, 52, 53, 53, 54, 54,
+        57, 60, 63, 67, 69, 72, 74, 77, 80, 83, 86, 90, 94, 97, 101, 106, 110,
+        114, 118, 123, 127, 131, 135, 140, 140, 140, 140, 53, 53, 53, 54, 54,
+        58, 62, 67, 73, 75, 78, 80, 83, 86, 89, 92, 95, 99, 102, 106, 110, 114,
+        117, 122, 126, 130, 134, 138, 142, 142, 142, 142, 56, 56, 56, 56, 56,
+        60, 64, 69, 75, 78, 81, 84, 87, 90, 93, 96, 100, 103, 106, 110, 114,
+        118, 121, 125, 130, 133, 137, 141, 145, 145, 145, 145, 60, 59, 59, 58,
+        58, 62, 66, 72, 78, 81, 84, 88, 92, 95, 98, 101, 104, 108, 111, 115,
+        118, 122, 125, 129, 133, 136, 140, 144, 148, 148, 148, 148, 64, 63, 62,
+        61, 60, 64, 69, 74, 80, 84, 88, 92, 97, 100, 103, 106, 110, 113, 116,
+        119, 123, 126, 130, 133, 137, 140, 143, 147, 151, 151, 151, 151, 68, 66,
+        65, 64, 62, 66, 71, 77, 83, 87, 92, 97, 102, 105, 108, 112, 116, 119,
+        122, 125, 128, 131, 134, 138, 141, 144, 147, 150, 154, 154, 154, 154,
+        72, 70, 69, 67, 65, 70, 74, 80, 86, 90, 95, 100, 105, 108, 112, 116,
+        119, 122, 125, 129, 132, 135, 138, 141, 144, 147, 150, 153, 156, 156,
+        156, 156, 77, 75, 73, 71, 69, 73, 78, 83, 89, 93, 98, 103, 108, 112,
+        115, 119, 123, 126, 129, 133, 136, 139, 142, 145, 148, 151, 153, 156,
+        159, 159, 159, 159, 82, 79, 77, 75, 73, 77, 81, 86, 92, 96, 101, 106,
+        112, 116, 119, 123, 128, 131, 134, 137, 140, 143, 146, 149, 152, 154,
+        157, 160, 162, 162, 162, 162, 88, 85, 82, 80, 77, 81, 85, 90, 95, 100,
+        104, 110, 116, 119, 123, 128, 132, 135, 138, 141, 145, 147, 150, 153,
+        156, 158, 161, 163, 166, 166, 166, 166, 92, 89, 86, 84, 81, 85, 89, 94,
+        99, 103, 108, 113, 119, 122, 126, 131, 135, 138, 141, 144, 148, 150,
+        153, 156, 159, 161, 163, 166, 168, 168, 168, 168, 97, 94, 91, 88, 85,
+        89, 93, 97, 102, 106, 111, 116, 122, 125, 129, 134, 138, 141, 144, 148,
+        151, 154, 156, 159, 162, 164, 166, 168, 171, 171, 171, 171, 103, 100,
+        96, 93, 90, 94, 97, 101, 106, 110, 115, 119, 125, 129, 133, 137, 141,
+        144, 148, 151, 154, 157, 159, 162, 165, 167, 169, 171, 174, 174, 174,
+        174, 109, 106, 102, 99, 96, 99, 102, 106, 110, 114, 118, 123, 128, 132,
+        136, 140, 145, 148, 151, 154, 158, 160, 163, 165, 168, 170, 172, 174,
+        176, 176, 176, 176, 114, 110, 106, 103, 100, 103, 106, 110, 114, 118,
+        122, 126, 131, 135, 139, 143, 147, 150, 154, 157, 160, 163, 165, 168,
+        170, 172, 174, 176, 178, 178, 178, 178, 119, 115, 111, 108, 104, 107,
+        111, 114, 117, 121, 125, 130, 134, 138, 142, 146, 150, 153, 156, 159,
+        163, 165, 168, 170, 173, 174, 176, 178, 180, 180, 180, 180, 124, 120,
+        116, 113, 109, 112, 115, 118, 122, 125, 129, 133, 138, 141, 145, 149,
+        153, 156, 159, 162, 165, 168, 170, 172, 175, 177, 179, 181, 183, 183,
+        183, 183, 130, 126, 122, 118, 115, 118, 120, 123, 126, 130, 133, 137,
+        141, 144, 148, 152, 156, 159, 162, 165, 168, 170, 173, 175, 177, 179,
+        181, 183, 185, 185, 185, 185, 134, 130, 126, 123, 119, 122, 124, 127,
+        130, 133, 136, 140, 144, 147, 151, 154, 158, 161, 164, 167, 170, 172,
+        174, 177, 179, 181, 183, 185, 186, 186, 186, 186, 139, 135, 131, 127,
+        124, 126, 128, 131, 134, 137, 140, 143, 147, 150, 153, 157, 161, 163,
+        166, 169, 172, 174, 176, 179, 181, 183, 184, 186, 188, 188, 188, 188,
+        144, 140, 136, 132, 128, 131, 133, 135, 138, 141, 144, 147, 150, 153,
+        156, 160, 163, 166, 168, 171, 174, 176, 178, 181, 183, 185, 186, 188,
+        190, 190, 190, 190, 149, 145, 141, 137, 134, 136, 138, 140, 142, 145,
+        148, 151, 154, 156, 159, 162, 166, 168, 171, 174, 176, 178, 180, 183,
+        185, 186, 188, 190, 191, 191, 191, 191, 149, 145, 141, 137, 134, 136,
+        138, 140, 142, 145, 148, 151, 154, 156, 159, 162, 166, 168, 171, 174,
+        176, 178, 180, 183, 185, 186, 188, 190, 191, 191, 191, 191, 149, 145,
+        141, 137, 134, 136, 138, 140, 142, 145, 148, 151, 154, 156, 159, 162,
+        166, 168, 171, 174, 176, 178, 180, 183, 185, 186, 188, 190, 191, 191,
+        191, 191, 149, 145, 141, 137, 134, 136, 138, 140, 142, 145, 148, 151,
+        154, 156, 159, 162, 166, 168, 171, 174, 176, 178, 180, 183, 185, 186,
+        188, 190, 191, 191, 191, 191 },
+      { /* Intra matrices */
+        /* Size 4 */
+        26, 29, 46, 68, 29, 42, 57, 73, 46, 57, 75, 88, 68, 73, 88, 99,
+        /* Size 8 */
+        31, 24, 26, 33, 44, 55, 67, 77, 24, 27, 26, 30, 38, 48, 58, 69, 26, 26,
+        36, 41, 47, 55, 64, 73, 33, 30, 41, 51, 59, 65, 73, 80, 44, 38, 47, 59,
+        68, 75, 81, 87, 55, 48, 55, 65, 75, 82, 88, 93, 67, 58, 64, 73, 81, 88,
+        94, 98, 77, 69, 73, 80, 87, 93, 98, 102,
+        /* Size 16 */
+        30, 26, 23, 24, 25, 28, 32, 37, 42, 47, 53, 58, 64, 69, 75, 75, 26, 25,
+        25, 25, 25, 28, 31, 34, 39, 44, 49, 54, 60, 65, 70, 70, 23, 25, 26, 26,
+        25, 27, 29, 33, 37, 41, 46, 51, 56, 61, 66, 66, 24, 25, 26, 27, 29, 31,
+        34, 37, 41, 45, 50, 54, 59, 63, 69, 69, 25, 25, 25, 29, 35, 37, 40, 43,
+        46, 49, 54, 58, 62, 66, 71, 71, 28, 28, 27, 31, 37, 40, 44, 47, 51, 54,
+        58, 62, 66, 70, 74, 74, 32, 31, 29, 34, 40, 44, 50, 53, 57, 60, 63, 67,
+        70, 74, 77, 77, 37, 34, 33, 37, 43, 47, 53, 57, 61, 64, 68, 71, 74, 77,
+        81, 81, 42, 39, 37, 41, 46, 51, 57, 61, 66, 69, 72, 75, 79, 81, 84, 84,
+        47, 44, 41, 45, 49, 54, 60, 64, 69, 72, 76, 79, 82, 84, 87, 87, 53, 49,
+        46, 50, 54, 58, 63, 68, 72, 76, 80, 83, 85, 88, 90, 90, 58, 54, 51, 54,
+        58, 62, 67, 71, 75, 79, 83, 85, 88, 90, 93, 93, 64, 60, 56, 59, 62, 66,
+        70, 74, 79, 82, 85, 88, 91, 93, 95, 95, 69, 65, 61, 63, 66, 70, 74, 77,
+        81, 84, 88, 90, 93, 95, 97, 97, 75, 70, 66, 69, 71, 74, 77, 81, 84, 87,
+        90, 93, 95, 97, 99, 99, 75, 70, 66, 69, 71, 74, 77, 81, 84, 87, 90, 93,
+        95, 97, 99, 99,
+        /* Size 32 */
+        30, 28, 26, 24, 23, 23, 24, 24, 24, 26, 28, 30, 32, 34, 36, 39, 42, 44,
+        46, 49, 52, 55, 57, 60, 63, 66, 68, 71, 73, 73, 73, 73, 28, 26, 25, 24,
+        23, 24, 24, 24, 25, 26, 27, 29, 31, 33, 35, 37, 40, 42, 45, 47, 50, 53,
+        55, 58, 61, 63, 66, 68, 71, 71, 71, 71, 26, 25, 25, 25, 24, 24, 24, 25,
+        25, 26, 27, 29, 30, 32, 34, 36, 39, 41, 43, 46, 49, 51, 53, 56, 59, 61,
+        64, 66, 69, 69, 69, 69, 24, 24, 25, 25, 25, 25, 25, 25, 25, 26, 27, 28,
+        29, 31, 33, 35, 37, 39, 42, 44, 47, 49, 52, 54, 57, 59, 62, 64, 67, 67,
+        67, 67, 23, 23, 24, 25, 26, 26, 25, 25, 25, 26, 27, 28, 29, 30, 32, 34,
+        36, 38, 40, 43, 45, 48, 50, 52, 55, 58, 60, 62, 65, 65, 65, 65, 23, 24,
+        24, 25, 26, 26, 26, 26, 27, 28, 29, 30, 31, 32, 34, 36, 38, 40, 42, 44,
+        47, 49, 51, 54, 57, 59, 61, 64, 66, 66, 66, 66, 24, 24, 24, 25, 25, 26,
+        27, 28, 29, 30, 31, 32, 33, 35, 36, 38, 40, 42, 44, 46, 49, 51, 53, 55,
+        58, 60, 62, 65, 67, 67, 67, 67, 24, 24, 25, 25, 25, 26, 28, 29, 31, 32,
+        33, 35, 36, 37, 39, 41, 43, 44, 46, 48, 51, 53, 55, 57, 60, 62, 64, 66,
+        69, 69, 69, 69, 24, 25, 25, 25, 25, 27, 29, 31, 34, 35, 36, 38, 39, 41,
+        42, 44, 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67, 70, 70, 70, 70,
+        26, 26, 26, 26, 26, 28, 30, 32, 35, 37, 38, 40, 41, 43, 44, 46, 47, 49,
+        51, 53, 55, 57, 59, 61, 63, 65, 67, 69, 71, 71, 71, 71, 28, 27, 27, 27,
+        27, 29, 31, 33, 36, 38, 40, 41, 43, 45, 46, 48, 50, 52, 53, 55, 57, 59,
+        61, 63, 65, 67, 69, 71, 73, 73, 73, 73, 30, 29, 29, 28, 28, 30, 32, 35,
+        38, 40, 41, 44, 46, 47, 49, 51, 53, 54, 56, 58, 60, 61, 63, 65, 67, 69,
+        71, 72, 74, 74, 74, 74, 32, 31, 30, 29, 29, 31, 33, 36, 39, 41, 43, 46,
+        49, 50, 52, 54, 56, 57, 59, 61, 62, 64, 66, 67, 69, 71, 72, 74, 76, 76,
+        76, 76, 34, 33, 32, 31, 30, 32, 35, 37, 41, 43, 45, 47, 50, 52, 54, 56,
+        58, 59, 61, 63, 64, 66, 68, 69, 71, 73, 74, 76, 78, 78, 78, 78, 36, 35,
+        34, 33, 32, 34, 36, 39, 42, 44, 46, 49, 52, 54, 56, 58, 60, 61, 63, 65,
+        66, 68, 70, 71, 73, 74, 76, 78, 79, 79, 79, 79, 39, 37, 36, 35, 34, 36,
+        38, 41, 44, 46, 48, 51, 54, 56, 58, 60, 62, 64, 65, 67, 69, 70, 72, 73,
+        75, 76, 78, 79, 81, 81, 81, 81, 42, 40, 39, 37, 36, 38, 40, 43, 45, 47,
+        50, 53, 56, 58, 60, 62, 64, 66, 68, 69, 71, 73, 74, 76, 77, 79, 80, 81,
+        83, 83, 83, 83, 44, 42, 41, 39, 38, 40, 42, 44, 47, 49, 52, 54, 57, 59,
+        61, 64, 66, 68, 69, 71, 73, 74, 76, 77, 79, 80, 81, 83, 84, 84, 84, 84,
+        46, 45, 43, 42, 40, 42, 44, 46, 49, 51, 53, 56, 59, 61, 63, 65, 68, 69,
+        71, 73, 75, 76, 77, 79, 81, 82, 83, 84, 86, 86, 86, 86, 49, 47, 46, 44,
+        43, 44, 46, 48, 51, 53, 55, 58, 61, 63, 65, 67, 69, 71, 73, 75, 77, 78,
+        79, 81, 82, 83, 85, 86, 87, 87, 87, 87, 52, 50, 49, 47, 45, 47, 49, 51,
+        53, 55, 57, 60, 62, 64, 66, 69, 71, 73, 75, 77, 78, 80, 81, 83, 84, 85,
+        86, 88, 89, 89, 89, 89, 55, 53, 51, 49, 48, 49, 51, 53, 55, 57, 59, 61,
+        64, 66, 68, 70, 73, 74, 76, 78, 80, 81, 82, 84, 85, 86, 88, 89, 90, 90,
+        90, 90, 57, 55, 53, 52, 50, 51, 53, 55, 57, 59, 61, 63, 66, 68, 70, 72,
+        74, 76, 77, 79, 81, 82, 84, 85, 87, 88, 89, 90, 91, 91, 91, 91, 60, 58,
+        56, 54, 52, 54, 55, 57, 59, 61, 63, 65, 67, 69, 71, 73, 76, 77, 79, 81,
+        83, 84, 85, 87, 88, 89, 90, 91, 92, 92, 92, 92, 63, 61, 59, 57, 55, 57,
+        58, 60, 61, 63, 65, 67, 69, 71, 73, 75, 77, 79, 81, 82, 84, 85, 87, 88,
+        89, 90, 92, 93, 94, 94, 94, 94, 66, 63, 61, 59, 58, 59, 60, 62, 63, 65,
+        67, 69, 71, 73, 74, 76, 79, 80, 82, 83, 85, 86, 88, 89, 90, 91, 93, 94,
+        95, 95, 95, 95, 68, 66, 64, 62, 60, 61, 62, 64, 65, 67, 69, 71, 72, 74,
+        76, 78, 80, 81, 83, 85, 86, 88, 89, 90, 92, 93, 94, 95, 96, 96, 96, 96,
+        71, 68, 66, 64, 62, 64, 65, 66, 67, 69, 71, 72, 74, 76, 78, 79, 81, 83,
+        84, 86, 88, 89, 90, 91, 93, 94, 95, 96, 97, 97, 97, 97, 73, 71, 69, 67,
+        65, 66, 67, 69, 70, 71, 73, 74, 76, 78, 79, 81, 83, 84, 86, 87, 89, 90,
+        91, 92, 94, 95, 96, 97, 98, 98, 98, 98, 73, 71, 69, 67, 65, 66, 67, 69,
+        70, 71, 73, 74, 76, 78, 79, 81, 83, 84, 86, 87, 89, 90, 91, 92, 94, 95,
+        96, 97, 98, 98, 98, 98, 73, 71, 69, 67, 65, 66, 67, 69, 70, 71, 73, 74,
+        76, 78, 79, 81, 83, 84, 86, 87, 89, 90, 91, 92, 94, 95, 96, 97, 98, 98,
+        98, 98, 73, 71, 69, 67, 65, 66, 67, 69, 70, 71, 73, 74, 76, 78, 79, 81,
+        83, 84, 86, 87, 89, 90, 91, 92, 94, 95, 96, 97, 98, 98, 98, 98 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 96, 104, 126, 96, 113, 119, 133, 104, 119, 146, 162, 126, 133, 162,
+        183,
+        /* Size 8 */
+        64, 53, 90, 94, 100, 110, 121, 134, 53, 73, 91, 84, 86, 93, 103, 115,
+        90, 91, 104, 101, 101, 105, 113, 123, 94, 84, 101, 112, 117, 121, 127,
+        135, 100, 86, 101, 117, 128, 135, 141, 148, 110, 93, 105, 121, 135, 145,
+        153, 160, 121, 103, 113, 127, 141, 153, 163, 170, 134, 115, 123, 135,
+        148, 160, 170, 179,
+        /* Size 16 */
+        64, 58, 53, 66, 90, 92, 94, 97, 100, 105, 110, 115, 121, 127, 134, 134,
+        58, 59, 61, 73, 90, 90, 89, 90, 92, 96, 101, 106, 112, 117, 123, 123,
+        53, 61, 73, 81, 91, 87, 84, 85, 86, 89, 93, 98, 103, 109, 115, 115, 66,
+        73, 81, 89, 97, 94, 92, 92, 93, 96, 99, 103, 108, 113, 119, 119, 90, 90,
+        91, 97, 104, 103, 101, 101, 101, 103, 105, 109, 113, 118, 123, 123, 92,
+        90, 87, 94, 103, 105, 107, 107, 108, 110, 113, 116, 120, 124, 128, 128,
+        94, 89, 84, 92, 101, 107, 112, 115, 117, 119, 121, 124, 127, 131, 135,
+        135, 97, 90, 85, 92, 101, 107, 115, 118, 122, 125, 128, 131, 134, 137,
+        141, 141, 100, 92, 86, 93, 101, 108, 117, 122, 128, 131, 135, 138, 141,
+        144, 148, 148, 105, 96, 89, 96, 103, 110, 119, 125, 131, 135, 140, 143,
+        147, 150, 154, 154, 110, 101, 93, 99, 105, 113, 121, 128, 135, 140, 145,
+        149, 153, 157, 160, 160, 115, 106, 98, 103, 109, 116, 124, 131, 138,
+        143, 149, 153, 158, 161, 165, 165, 121, 112, 103, 108, 113, 120, 127,
+        134, 141, 147, 153, 158, 163, 166, 170, 170, 127, 117, 109, 113, 118,
+        124, 131, 137, 144, 150, 157, 161, 166, 170, 174, 174, 134, 123, 115,
+        119, 123, 128, 135, 141, 148, 154, 160, 165, 170, 174, 179, 179, 134,
+        123, 115, 119, 123, 128, 135, 141, 148, 154, 160, 165, 170, 174, 179,
+        179,
+        /* Size 32 */
+        64, 61, 58, 55, 53, 59, 66, 76, 90, 91, 92, 93, 94, 96, 97, 98, 100,
+        102, 105, 107, 110, 112, 115, 118, 121, 124, 127, 131, 134, 134, 134,
+        134, 61, 60, 59, 58, 57, 62, 70, 78, 90, 90, 91, 91, 91, 93, 94, 95, 96,
+        98, 100, 103, 105, 108, 110, 113, 116, 119, 122, 125, 129, 129, 129,
+        129, 58, 59, 59, 60, 61, 67, 73, 81, 90, 90, 90, 89, 89, 90, 90, 91, 92,
+        94, 96, 98, 101, 103, 106, 109, 112, 114, 117, 120, 123, 123, 123, 123,
+        55, 58, 60, 63, 67, 72, 77, 83, 91, 90, 88, 87, 86, 87, 87, 88, 89, 91,
+        93, 95, 97, 99, 102, 104, 107, 110, 113, 116, 119, 119, 119, 119, 53,
+        57, 61, 67, 73, 77, 81, 86, 91, 89, 87, 85, 84, 84, 85, 85, 86, 88, 89,
+        91, 93, 95, 98, 100, 103, 106, 109, 111, 115, 115, 115, 115, 59, 62, 67,
+        72, 77, 81, 85, 89, 94, 92, 91, 89, 87, 88, 88, 89, 89, 91, 92, 94, 96,
+        98, 100, 103, 105, 108, 111, 114, 117, 117, 117, 117, 66, 70, 73, 77,
+        81, 85, 89, 93, 97, 96, 94, 93, 92, 92, 92, 92, 93, 94, 96, 97, 99, 101,
+        103, 105, 108, 110, 113, 116, 119, 119, 119, 119, 76, 78, 81, 83, 86,
+        89, 93, 97, 101, 99, 98, 97, 96, 96, 96, 97, 97, 98, 99, 101, 102, 104,
+        106, 108, 110, 113, 115, 118, 121, 121, 121, 121, 90, 90, 90, 91, 91,
+        94, 97, 101, 104, 103, 103, 102, 101, 101, 101, 101, 101, 102, 103, 104,
+        105, 107, 109, 111, 113, 115, 118, 120, 123, 123, 123, 123, 91, 90, 90,
+        90, 89, 92, 96, 99, 103, 104, 104, 104, 104, 104, 104, 104, 105, 106,
+        107, 108, 109, 111, 112, 114, 116, 118, 121, 123, 126, 126, 126, 126,
+        92, 91, 90, 88, 87, 91, 94, 98, 103, 104, 105, 106, 107, 107, 107, 108,
+        108, 109, 110, 112, 113, 114, 116, 118, 120, 122, 124, 126, 128, 128,
+        128, 128, 93, 91, 89, 87, 85, 89, 93, 97, 102, 104, 106, 107, 109, 110,
+        111, 112, 112, 113, 115, 116, 117, 118, 120, 121, 123, 125, 127, 129,
+        131, 131, 131, 131, 94, 91, 89, 86, 84, 87, 92, 96, 101, 104, 107, 109,
+        112, 113, 115, 116, 117, 118, 119, 120, 121, 122, 124, 125, 127, 129,
+        131, 133, 135, 135, 135, 135, 96, 93, 90, 87, 84, 88, 92, 96, 101, 104,
+        107, 110, 113, 115, 116, 118, 119, 121, 122, 123, 124, 126, 127, 129,
+        130, 132, 134, 136, 138, 138, 138, 138, 97, 94, 90, 87, 85, 88, 92, 96,
+        101, 104, 107, 111, 115, 116, 118, 120, 122, 123, 125, 126, 128, 129,
+        131, 132, 134, 135, 137, 139, 141, 141, 141, 141, 98, 95, 91, 88, 85,
+        89, 92, 97, 101, 104, 108, 112, 116, 118, 120, 122, 125, 126, 128, 130,
+        131, 133, 134, 136, 137, 139, 141, 142, 144, 144, 144, 144, 100, 96, 92,
+        89, 86, 89, 93, 97, 101, 105, 108, 112, 117, 119, 122, 125, 128, 130,
+        131, 133, 135, 136, 138, 140, 141, 143, 144, 146, 148, 148, 148, 148,
+        102, 98, 94, 91, 88, 91, 94, 98, 102, 106, 109, 113, 118, 121, 123, 126,
+        130, 131, 133, 135, 137, 139, 141, 142, 144, 146, 147, 149, 151, 151,
+        151, 151, 105, 100, 96, 93, 89, 92, 96, 99, 103, 107, 110, 115, 119,
+        122, 125, 128, 131, 133, 135, 138, 140, 142, 143, 145, 147, 149, 150,
+        152, 154, 154, 154, 154, 107, 103, 98, 95, 91, 94, 97, 101, 104, 108,
+        112, 116, 120, 123, 126, 130, 133, 135, 138, 140, 143, 144, 146, 148,
+        150, 152, 153, 155, 157, 157, 157, 157, 110, 105, 101, 97, 93, 96, 99,
+        102, 105, 109, 113, 117, 121, 124, 128, 131, 135, 137, 140, 143, 145,
+        147, 149, 151, 153, 155, 157, 158, 160, 160, 160, 160, 112, 108, 103,
+        99, 95, 98, 101, 104, 107, 111, 114, 118, 122, 126, 129, 133, 136, 139,
+        142, 144, 147, 149, 151, 153, 156, 157, 159, 161, 162, 162, 162, 162,
+        115, 110, 106, 102, 98, 100, 103, 106, 109, 112, 116, 120, 124, 127,
+        131, 134, 138, 141, 143, 146, 149, 151, 153, 156, 158, 160, 161, 163,
+        165, 165, 165, 165, 118, 113, 109, 104, 100, 103, 105, 108, 111, 114,
+        118, 121, 125, 129, 132, 136, 140, 142, 145, 148, 151, 153, 156, 158,
+        160, 162, 164, 166, 168, 168, 168, 168, 121, 116, 112, 107, 103, 105,
+        108, 110, 113, 116, 120, 123, 127, 130, 134, 137, 141, 144, 147, 150,
+        153, 156, 158, 160, 163, 165, 166, 168, 170, 170, 170, 170, 124, 119,
+        114, 110, 106, 108, 110, 113, 115, 118, 122, 125, 129, 132, 135, 139,
+        143, 146, 149, 152, 155, 157, 160, 162, 165, 166, 168, 170, 172, 172,
+        172, 172, 127, 122, 117, 113, 109, 111, 113, 115, 118, 121, 124, 127,
+        131, 134, 137, 141, 144, 147, 150, 153, 157, 159, 161, 164, 166, 168,
+        170, 172, 174, 174, 174, 174, 131, 125, 120, 116, 111, 114, 116, 118,
+        120, 123, 126, 129, 133, 136, 139, 142, 146, 149, 152, 155, 158, 161,
+        163, 166, 168, 170, 172, 174, 177, 177, 177, 177, 134, 129, 123, 119,
+        115, 117, 119, 121, 123, 126, 128, 131, 135, 138, 141, 144, 148, 151,
+        154, 157, 160, 162, 165, 168, 170, 172, 174, 177, 179, 179, 179, 179,
+        134, 129, 123, 119, 115, 117, 119, 121, 123, 126, 128, 131, 135, 138,
+        141, 144, 148, 151, 154, 157, 160, 162, 165, 168, 170, 172, 174, 177,
+        179, 179, 179, 179, 134, 129, 123, 119, 115, 117, 119, 121, 123, 126,
+        128, 131, 135, 138, 141, 144, 148, 151, 154, 157, 160, 162, 165, 168,
+        170, 172, 174, 177, 179, 179, 179, 179, 134, 129, 123, 119, 115, 117,
+        119, 121, 123, 126, 128, 131, 135, 138, 141, 144, 148, 151, 154, 157,
+        160, 162, 165, 168, 170, 172, 174, 177, 179, 179, 179, 179 },
+      { /* Intra matrices */
+        /* Size 4 */
+        31, 47, 51, 63, 47, 56, 59, 67, 51, 59, 74, 82, 63, 67, 82, 95,
+        /* Size 8 */
+        33, 27, 47, 50, 53, 58, 65, 72, 27, 38, 48, 44, 45, 49, 55, 61, 47, 48,
+        55, 54, 53, 56, 60, 66, 50, 44, 54, 60, 62, 65, 68, 73, 53, 45, 53, 62,
+        69, 73, 76, 80, 58, 49, 56, 65, 73, 79, 84, 88, 65, 55, 60, 68, 76, 84,
+        89, 94, 72, 61, 66, 73, 80, 88, 94, 99,
+        /* Size 16 */
+        32, 29, 26, 33, 46, 47, 48, 50, 51, 54, 57, 60, 63, 67, 70, 70, 29, 30,
+        31, 37, 46, 46, 45, 46, 47, 50, 52, 55, 58, 61, 65, 65, 26, 31, 37, 41,
+        47, 45, 43, 43, 44, 46, 48, 50, 53, 56, 60, 60, 33, 37, 41, 45, 50, 48,
+        47, 47, 48, 49, 51, 53, 56, 59, 62, 62, 46, 46, 47, 50, 54, 53, 52, 52,
+        52, 53, 55, 57, 59, 61, 64, 64, 47, 46, 45, 48, 53, 54, 55, 56, 56, 57,
+        59, 60, 62, 65, 67, 67, 48, 45, 43, 47, 52, 55, 58, 60, 61, 62, 63, 65,
+        67, 69, 71, 71, 50, 46, 43, 47, 52, 56, 60, 62, 64, 65, 67, 69, 70, 72,
+        74, 74, 51, 47, 44, 48, 52, 56, 61, 64, 67, 69, 71, 73, 75, 76, 78, 78,
+        54, 50, 46, 49, 53, 57, 62, 65, 69, 71, 74, 76, 78, 80, 82, 82, 57, 52,
+        48, 51, 55, 59, 63, 67, 71, 74, 77, 79, 82, 84, 86, 86, 60, 55, 50, 53,
+        57, 60, 65, 69, 73, 76, 79, 82, 84, 86, 89, 89, 63, 58, 53, 56, 59, 62,
+        67, 70, 75, 78, 82, 84, 87, 89, 92, 92, 67, 61, 56, 59, 61, 65, 69, 72,
+        76, 80, 84, 86, 89, 92, 94, 94, 70, 65, 60, 62, 64, 67, 71, 74, 78, 82,
+        86, 89, 92, 94, 97, 97, 70, 65, 60, 62, 64, 67, 71, 74, 78, 82, 86, 89,
+        92, 94, 97, 97,
+        /* Size 32 */
+        32, 30, 29, 27, 26, 29, 33, 38, 45, 46, 47, 47, 48, 49, 49, 50, 51, 52,
+        53, 55, 56, 58, 59, 61, 63, 64, 66, 68, 70, 70, 70, 70, 30, 30, 29, 29,
+        28, 31, 35, 39, 46, 46, 46, 46, 46, 47, 47, 48, 49, 50, 51, 52, 54, 55,
+        57, 58, 60, 61, 63, 65, 67, 67, 67, 67, 29, 29, 30, 30, 30, 33, 37, 41,
+        46, 46, 45, 45, 45, 45, 46, 46, 47, 48, 49, 50, 51, 53, 54, 56, 57, 59,
+        60, 62, 64, 64, 64, 64, 27, 29, 30, 32, 33, 36, 39, 42, 46, 45, 45, 44,
+        43, 44, 44, 45, 45, 46, 47, 48, 49, 50, 52, 53, 55, 56, 58, 59, 61, 61,
+        61, 61, 26, 28, 30, 33, 37, 39, 41, 43, 46, 45, 44, 43, 42, 42, 43, 43,
+        43, 44, 45, 46, 47, 48, 50, 51, 53, 54, 56, 57, 59, 59, 59, 59, 29, 31,
+        33, 36, 39, 41, 43, 45, 48, 47, 46, 45, 44, 44, 45, 45, 45, 46, 47, 48,
+        49, 50, 51, 52, 54, 55, 57, 58, 60, 60, 60, 60, 33, 35, 37, 39, 41, 43,
+        45, 47, 49, 49, 48, 47, 46, 47, 47, 47, 47, 48, 49, 49, 50, 51, 53, 54,
+        55, 57, 58, 59, 61, 61, 61, 61, 38, 39, 41, 42, 43, 45, 47, 49, 51, 51,
+        50, 49, 49, 49, 49, 49, 49, 50, 51, 51, 52, 53, 54, 55, 57, 58, 59, 61,
+        62, 62, 62, 62, 45, 46, 46, 46, 46, 48, 49, 51, 53, 53, 52, 52, 52, 52,
+        52, 51, 51, 52, 53, 53, 54, 55, 56, 57, 58, 59, 61, 62, 63, 63, 63, 63,
+        46, 46, 46, 45, 45, 47, 49, 51, 53, 53, 53, 53, 53, 53, 53, 53, 53, 54,
+        55, 55, 56, 57, 58, 59, 60, 61, 62, 64, 65, 65, 65, 65, 47, 46, 45, 45,
+        44, 46, 48, 50, 52, 53, 53, 54, 54, 55, 55, 55, 55, 56, 57, 57, 58, 59,
+        60, 61, 62, 63, 64, 65, 67, 67, 67, 67, 47, 46, 45, 44, 43, 45, 47, 49,
+        52, 53, 54, 55, 56, 56, 57, 57, 58, 58, 59, 59, 60, 61, 62, 63, 64, 65,
+        66, 67, 68, 68, 68, 68, 48, 46, 45, 43, 42, 44, 46, 49, 52, 53, 54, 56,
+        58, 58, 59, 59, 60, 61, 61, 62, 62, 63, 64, 65, 66, 67, 68, 69, 70, 70,
+        70, 70, 49, 47, 45, 44, 42, 44, 47, 49, 52, 53, 55, 56, 58, 59, 60, 61,
+        62, 62, 63, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 72, 72, 72, 49, 47,
+        46, 44, 43, 45, 47, 49, 52, 53, 55, 57, 59, 60, 61, 62, 63, 64, 64, 65,
+        66, 67, 68, 69, 69, 70, 71, 72, 74, 74, 74, 74, 50, 48, 46, 45, 43, 45,
+        47, 49, 51, 53, 55, 57, 59, 61, 62, 63, 65, 65, 66, 67, 68, 69, 70, 71,
+        72, 72, 73, 74, 75, 75, 75, 75, 51, 49, 47, 45, 43, 45, 47, 49, 51, 53,
+        55, 58, 60, 62, 63, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 76,
+        77, 77, 77, 77, 52, 50, 48, 46, 44, 46, 48, 50, 52, 54, 56, 58, 61, 62,
+        64, 65, 67, 68, 69, 70, 72, 72, 73, 74, 75, 76, 77, 78, 79, 79, 79, 79,
+        53, 51, 49, 47, 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 64, 66, 68, 69,
+        70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 81, 81, 81, 55, 52, 50, 48,
+        46, 48, 49, 51, 53, 55, 57, 59, 62, 63, 65, 67, 69, 70, 72, 73, 75, 76,
+        77, 78, 79, 80, 81, 82, 83, 83, 83, 83, 56, 54, 51, 49, 47, 49, 50, 52,
+        54, 56, 58, 60, 62, 64, 66, 68, 70, 72, 73, 75, 76, 77, 78, 79, 81, 82,
+        83, 84, 85, 85, 85, 85, 58, 55, 53, 50, 48, 50, 51, 53, 55, 57, 59, 61,
+        63, 65, 67, 69, 71, 72, 74, 76, 77, 78, 80, 81, 82, 83, 84, 85, 86, 86,
+        86, 86, 59, 57, 54, 52, 50, 51, 53, 54, 56, 58, 60, 62, 64, 66, 68, 70,
+        72, 73, 75, 77, 78, 80, 81, 82, 83, 84, 85, 86, 87, 87, 87, 87, 61, 58,
+        56, 53, 51, 52, 54, 55, 57, 59, 61, 63, 65, 67, 69, 71, 73, 74, 76, 78,
+        79, 81, 82, 83, 85, 86, 87, 88, 89, 89, 89, 89, 63, 60, 57, 55, 53, 54,
+        55, 57, 58, 60, 62, 64, 66, 68, 69, 72, 74, 75, 77, 79, 81, 82, 83, 85,
+        86, 87, 88, 89, 91, 91, 91, 91, 64, 61, 59, 56, 54, 55, 57, 58, 59, 61,
+        63, 65, 67, 69, 70, 72, 75, 76, 78, 80, 82, 83, 84, 86, 87, 88, 89, 91,
+        92, 92, 92, 92, 66, 63, 60, 58, 56, 57, 58, 59, 61, 62, 64, 66, 68, 70,
+        71, 73, 76, 77, 79, 81, 83, 84, 85, 87, 88, 89, 91, 92, 93, 93, 93, 93,
+        68, 65, 62, 59, 57, 58, 59, 61, 62, 64, 65, 67, 69, 71, 72, 74, 76, 78,
+        80, 82, 84, 85, 86, 88, 89, 91, 92, 93, 94, 94, 94, 94, 70, 67, 64, 61,
+        59, 60, 61, 62, 63, 65, 67, 68, 70, 72, 74, 75, 77, 79, 81, 83, 85, 86,
+        87, 89, 91, 92, 93, 94, 96, 96, 96, 96, 70, 67, 64, 61, 59, 60, 61, 62,
+        63, 65, 67, 68, 70, 72, 74, 75, 77, 79, 81, 83, 85, 86, 87, 89, 91, 92,
+        93, 94, 96, 96, 96, 96, 70, 67, 64, 61, 59, 60, 61, 62, 63, 65, 67, 68,
+        70, 72, 74, 75, 77, 79, 81, 83, 85, 86, 87, 89, 91, 92, 93, 94, 96, 96,
+        96, 96, 70, 67, 64, 61, 59, 60, 61, 62, 63, 65, 67, 68, 70, 72, 74, 75,
+        77, 79, 81, 83, 85, 86, 87, 89, 91, 92, 93, 94, 96, 96, 96, 96 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 69, 104, 143, 69, 97, 124, 152, 104, 124, 156, 176, 143, 152, 176,
+        192,
+        /* Size 8 */
+        64, 50, 54, 68, 86, 104, 122, 137, 50, 56, 54, 62, 76, 93, 109, 125, 54,
+        54, 72, 82, 92, 105, 118, 131, 68, 62, 82, 98, 110, 120, 130, 140, 86,
+        76, 92, 110, 123, 133, 142, 150, 104, 93, 105, 120, 133, 144, 151, 158,
+        122, 109, 118, 130, 142, 151, 158, 164, 137, 125, 131, 140, 150, 158,
+        164, 169,
+        /* Size 16 */
+        64, 56, 50, 52, 54, 60, 68, 76, 86, 94, 104, 112, 122, 129, 137, 137,
+        56, 55, 53, 54, 54, 59, 65, 72, 81, 89, 98, 106, 115, 122, 130, 130, 50,
+        53, 56, 55, 54, 58, 62, 69, 76, 84, 93, 100, 109, 116, 125, 125, 52, 54,
+        55, 59, 62, 66, 71, 77, 83, 90, 98, 105, 114, 120, 128, 128, 54, 54, 54,
+        62, 72, 77, 82, 87, 92, 98, 105, 111, 118, 124, 131, 131, 60, 59, 58,
+        66, 77, 82, 89, 94, 100, 106, 112, 118, 124, 130, 136, 136, 68, 65, 62,
+        71, 82, 89, 98, 104, 110, 115, 120, 125, 130, 135, 140, 140, 76, 72, 69,
+        77, 87, 94, 104, 109, 116, 121, 126, 131, 136, 140, 145, 145, 86, 81,
+        76, 83, 92, 100, 110, 116, 123, 128, 133, 138, 142, 146, 150, 150, 94,
+        89, 84, 90, 98, 106, 115, 121, 128, 133, 138, 142, 146, 150, 153, 153,
+        104, 98, 93, 98, 105, 112, 120, 126, 133, 138, 144, 147, 151, 154, 158,
+        158, 112, 106, 100, 105, 111, 118, 125, 131, 138, 142, 147, 151, 155,
+        158, 161, 161, 122, 115, 109, 114, 118, 124, 130, 136, 142, 146, 151,
+        155, 158, 161, 164, 164, 129, 122, 116, 120, 124, 130, 135, 140, 146,
+        150, 154, 158, 161, 164, 166, 166, 137, 130, 125, 128, 131, 136, 140,
+        145, 150, 153, 158, 161, 164, 166, 169, 169, 137, 130, 125, 128, 131,
+        136, 140, 145, 150, 153, 158, 161, 164, 166, 169, 169,
+        /* Size 32 */
+        64, 60, 56, 53, 50, 51, 52, 53, 54, 57, 60, 64, 68, 72, 76, 80, 86, 90,
+        94, 99, 104, 108, 112, 117, 122, 125, 129, 132, 137, 137, 137, 137, 60,
+        58, 55, 53, 52, 52, 53, 53, 54, 57, 59, 63, 66, 70, 74, 78, 83, 87, 91,
+        96, 101, 105, 109, 113, 118, 122, 125, 129, 133, 133, 133, 133, 56, 55,
+        55, 54, 53, 53, 54, 54, 54, 56, 59, 62, 65, 68, 72, 76, 81, 84, 89, 93,
+        98, 102, 106, 110, 115, 118, 122, 126, 130, 130, 130, 130, 53, 53, 54,
+        54, 55, 55, 55, 54, 54, 56, 59, 61, 64, 67, 70, 74, 78, 82, 86, 90, 95,
+        99, 103, 107, 112, 115, 119, 123, 127, 127, 127, 127, 50, 52, 53, 55,
+        56, 56, 55, 55, 54, 56, 58, 60, 62, 65, 69, 72, 76, 80, 84, 88, 93, 96,
+        100, 104, 109, 113, 116, 120, 125, 125, 125, 125, 51, 52, 53, 55, 56,
+        56, 57, 57, 58, 60, 62, 64, 66, 69, 72, 76, 80, 83, 87, 91, 95, 99, 103,
+        107, 111, 115, 118, 122, 126, 126, 126, 126, 52, 53, 54, 55, 55, 57, 59,
+        60, 62, 64, 66, 68, 71, 73, 77, 80, 83, 87, 90, 94, 98, 102, 105, 109,
+        114, 117, 120, 124, 128, 128, 128, 128, 53, 53, 54, 54, 55, 57, 60, 63,
+        67, 69, 71, 73, 76, 78, 81, 84, 88, 91, 94, 98, 101, 105, 108, 112, 116,
+        119, 122, 126, 129, 129, 129, 129, 54, 54, 54, 54, 54, 58, 62, 67, 72,
+        74, 77, 79, 82, 84, 87, 89, 92, 95, 98, 101, 105, 108, 111, 115, 118,
+        121, 124, 128, 131, 131, 131, 131, 57, 57, 56, 56, 56, 60, 64, 69, 74,
+        77, 79, 82, 85, 88, 90, 93, 96, 99, 102, 105, 108, 111, 114, 118, 121,
+        124, 127, 130, 133, 133, 133, 133, 60, 59, 59, 59, 58, 62, 66, 71, 77,
+        79, 82, 86, 89, 92, 94, 97, 100, 103, 106, 109, 112, 115, 118, 121, 124,
+        127, 130, 133, 136, 136, 136, 136, 64, 63, 62, 61, 60, 64, 68, 73, 79,
+        82, 86, 89, 93, 96, 99, 102, 105, 107, 110, 113, 116, 118, 121, 124,
+        127, 130, 132, 135, 138, 138, 138, 138, 68, 66, 65, 64, 62, 66, 71, 76,
+        82, 85, 89, 93, 98, 101, 104, 107, 110, 112, 115, 117, 120, 122, 125,
+        128, 130, 133, 135, 138, 140, 140, 140, 140, 72, 70, 68, 67, 65, 69, 73,
+        78, 84, 88, 92, 96, 101, 104, 106, 110, 113, 115, 118, 120, 123, 125,
+        128, 130, 133, 135, 138, 140, 143, 143, 143, 143, 76, 74, 72, 70, 69,
+        72, 77, 81, 87, 90, 94, 99, 104, 106, 109, 113, 116, 119, 121, 124, 126,
+        129, 131, 133, 136, 138, 140, 142, 145, 145, 145, 145, 80, 78, 76, 74,
+        72, 76, 80, 84, 89, 93, 97, 102, 107, 110, 113, 116, 120, 122, 124, 127,
+        130, 132, 134, 136, 139, 141, 143, 145, 147, 147, 147, 147, 86, 83, 81,
+        78, 76, 80, 83, 88, 92, 96, 100, 105, 110, 113, 116, 120, 123, 126, 128,
+        131, 133, 135, 138, 140, 142, 144, 146, 148, 150, 150, 150, 150, 90, 87,
+        84, 82, 80, 83, 87, 91, 95, 99, 103, 107, 112, 115, 119, 122, 126, 128,
+        131, 133, 136, 138, 140, 142, 144, 146, 148, 150, 151, 151, 151, 151,
+        94, 91, 89, 86, 84, 87, 90, 94, 98, 102, 106, 110, 115, 118, 121, 124,
+        128, 131, 133, 136, 138, 140, 142, 144, 146, 148, 150, 152, 153, 153,
+        153, 153, 99, 96, 93, 90, 88, 91, 94, 98, 101, 105, 109, 113, 117, 120,
+        124, 127, 131, 133, 136, 138, 141, 143, 145, 147, 149, 150, 152, 154,
+        155, 155, 155, 155, 104, 101, 98, 95, 93, 95, 98, 101, 105, 108, 112,
+        116, 120, 123, 126, 130, 133, 136, 138, 141, 144, 145, 147, 149, 151,
+        153, 154, 156, 158, 158, 158, 158, 108, 105, 102, 99, 96, 99, 102, 105,
+        108, 111, 115, 118, 122, 125, 129, 132, 135, 138, 140, 143, 145, 147,
+        149, 151, 153, 154, 156, 158, 159, 159, 159, 159, 112, 109, 106, 103,
+        100, 103, 105, 108, 111, 114, 118, 121, 125, 128, 131, 134, 138, 140,
+        142, 145, 147, 149, 151, 153, 155, 156, 158, 159, 161, 161, 161, 161,
+        117, 113, 110, 107, 104, 107, 109, 112, 115, 118, 121, 124, 128, 130,
+        133, 136, 140, 142, 144, 147, 149, 151, 153, 155, 157, 158, 159, 161,
+        162, 162, 162, 162, 122, 118, 115, 112, 109, 111, 114, 116, 118, 121,
+        124, 127, 130, 133, 136, 139, 142, 144, 146, 149, 151, 153, 155, 157,
+        158, 160, 161, 162, 164, 164, 164, 164, 125, 122, 118, 115, 113, 115,
+        117, 119, 121, 124, 127, 130, 133, 135, 138, 141, 144, 146, 148, 150,
+        153, 154, 156, 158, 160, 161, 162, 164, 165, 165, 165, 165, 129, 125,
+        122, 119, 116, 118, 120, 122, 124, 127, 130, 132, 135, 138, 140, 143,
+        146, 148, 150, 152, 154, 156, 158, 159, 161, 162, 164, 165, 166, 166,
+        166, 166, 132, 129, 126, 123, 120, 122, 124, 126, 128, 130, 133, 135,
+        138, 140, 142, 145, 148, 150, 152, 154, 156, 158, 159, 161, 162, 164,
+        165, 166, 167, 167, 167, 167, 137, 133, 130, 127, 125, 126, 128, 129,
+        131, 133, 136, 138, 140, 143, 145, 147, 150, 151, 153, 155, 158, 159,
+        161, 162, 164, 165, 166, 167, 169, 169, 169, 169, 137, 133, 130, 127,
+        125, 126, 128, 129, 131, 133, 136, 138, 140, 143, 145, 147, 150, 151,
+        153, 155, 158, 159, 161, 162, 164, 165, 166, 167, 169, 169, 169, 169,
+        137, 133, 130, 127, 125, 126, 128, 129, 131, 133, 136, 138, 140, 143,
+        145, 147, 150, 151, 153, 155, 158, 159, 161, 162, 164, 165, 166, 167,
+        169, 169, 169, 169, 137, 133, 130, 127, 125, 126, 128, 129, 131, 133,
+        136, 138, 140, 143, 145, 147, 150, 151, 153, 155, 158, 159, 161, 162,
+        164, 165, 166, 167, 169, 169, 169, 169 },
+      { /* Intra matrices */
+        /* Size 4 */
+        29, 32, 48, 68, 32, 45, 58, 73, 48, 58, 75, 86, 68, 73, 86, 94,
+        /* Size 8 */
+        34, 26, 28, 36, 46, 57, 67, 77, 26, 30, 29, 33, 41, 50, 60, 69, 28, 29,
+        39, 44, 50, 57, 65, 73, 36, 33, 44, 53, 60, 66, 73, 79, 46, 41, 50, 60,
+        68, 75, 80, 85, 57, 50, 57, 66, 75, 81, 86, 90, 67, 60, 65, 73, 80, 86,
+        90, 94, 77, 69, 73, 79, 85, 90, 94, 97,
+        /* Size 16 */
+        33, 29, 26, 27, 28, 31, 35, 39, 45, 50, 55, 60, 65, 70, 74, 74, 29, 28,
+        27, 27, 28, 30, 34, 37, 42, 47, 52, 56, 62, 66, 71, 71, 26, 27, 29, 28,
+        28, 30, 32, 35, 40, 44, 49, 53, 58, 62, 67, 67, 27, 27, 28, 30, 32, 34,
+        37, 40, 44, 47, 52, 56, 61, 65, 69, 69, 28, 28, 28, 32, 37, 40, 43, 45,
+        49, 52, 56, 59, 64, 67, 71, 71, 31, 30, 30, 34, 40, 43, 47, 50, 53, 56,
+        60, 63, 67, 70, 74, 74, 35, 34, 32, 37, 43, 47, 52, 55, 59, 61, 65, 67,
+        71, 74, 77, 77, 39, 37, 35, 40, 45, 50, 55, 58, 62, 65, 68, 71, 74, 77,
+        79, 79, 45, 42, 40, 44, 49, 53, 59, 62, 66, 69, 72, 75, 78, 80, 82, 82,
+        50, 47, 44, 47, 52, 56, 61, 65, 69, 72, 75, 78, 80, 82, 85, 85, 55, 52,
+        49, 52, 56, 60, 65, 68, 72, 75, 79, 81, 83, 85, 87, 87, 60, 56, 53, 56,
+        59, 63, 67, 71, 75, 78, 81, 83, 85, 87, 89, 89, 65, 62, 58, 61, 64, 67,
+        71, 74, 78, 80, 83, 85, 88, 89, 91, 91, 70, 66, 62, 65, 67, 70, 74, 77,
+        80, 82, 85, 87, 89, 91, 93, 93, 74, 71, 67, 69, 71, 74, 77, 79, 82, 85,
+        87, 89, 91, 93, 94, 94, 74, 71, 67, 69, 71, 74, 77, 79, 82, 85, 87, 89,
+        91, 93, 94, 94,
+        /* Size 32 */
+        33, 30, 28, 27, 25, 26, 26, 27, 27, 29, 30, 32, 35, 37, 39, 41, 44, 47,
+        49, 52, 55, 57, 59, 62, 65, 67, 69, 71, 73, 73, 73, 73, 30, 29, 28, 27,
+        26, 26, 27, 27, 27, 29, 30, 32, 34, 36, 38, 40, 43, 45, 47, 50, 53, 55,
+        57, 60, 63, 65, 67, 69, 71, 71, 71, 71, 28, 28, 28, 27, 27, 27, 27, 27,
+        27, 29, 30, 31, 33, 35, 37, 39, 42, 44, 46, 48, 51, 53, 56, 58, 61, 63,
+        65, 67, 70, 70, 70, 70, 27, 27, 27, 27, 28, 28, 28, 27, 27, 28, 30, 31,
+        32, 34, 36, 38, 40, 42, 44, 47, 50, 52, 54, 56, 59, 61, 63, 65, 68, 68,
+        68, 68, 25, 26, 27, 28, 29, 28, 28, 28, 27, 28, 29, 30, 32, 33, 35, 37,
+        39, 41, 43, 45, 48, 50, 52, 55, 57, 59, 61, 64, 66, 66, 66, 66, 26, 26,
+        27, 28, 28, 29, 29, 29, 29, 30, 31, 33, 34, 35, 37, 39, 41, 43, 45, 47,
+        50, 52, 54, 56, 59, 61, 63, 65, 67, 67, 67, 67, 26, 27, 27, 28, 28, 29,
+        30, 31, 32, 33, 34, 35, 36, 38, 39, 41, 43, 45, 47, 49, 51, 53, 55, 57,
+        60, 62, 64, 66, 68, 68, 68, 68, 27, 27, 27, 27, 28, 29, 31, 32, 34, 35,
+        36, 38, 39, 40, 42, 44, 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
+        69, 69, 69, 69, 27, 27, 27, 27, 27, 29, 32, 34, 37, 38, 39, 41, 42, 43,
+        45, 46, 48, 49, 51, 53, 55, 57, 59, 61, 63, 64, 66, 68, 70, 70, 70, 70,
+        29, 29, 29, 28, 28, 30, 33, 35, 38, 39, 41, 42, 44, 45, 47, 48, 50, 52,
+        53, 55, 57, 59, 60, 62, 64, 66, 68, 69, 71, 71, 71, 71, 30, 30, 30, 30,
+        29, 31, 34, 36, 39, 41, 42, 44, 46, 48, 49, 51, 52, 54, 55, 57, 59, 61,
+        62, 64, 66, 68, 69, 71, 73, 73, 73, 73, 32, 32, 31, 31, 30, 33, 35, 38,
+        41, 42, 44, 46, 49, 50, 52, 53, 55, 56, 58, 59, 61, 63, 64, 66, 68, 69,
+        71, 72, 74, 74, 74, 74, 35, 34, 33, 32, 32, 34, 36, 39, 42, 44, 46, 49,
+        51, 53, 54, 56, 58, 59, 60, 62, 64, 65, 66, 68, 70, 71, 72, 74, 75, 75,
+        75, 75, 37, 36, 35, 34, 33, 35, 38, 40, 43, 45, 48, 50, 53, 54, 56, 58,
+        59, 61, 62, 64, 65, 67, 68, 70, 71, 73, 74, 75, 77, 77, 77, 77, 39, 38,
+        37, 36, 35, 37, 39, 42, 45, 47, 49, 52, 54, 56, 58, 59, 61, 63, 64, 66,
+        67, 69, 70, 71, 73, 74, 75, 77, 78, 78, 78, 78, 41, 40, 39, 38, 37, 39,
+        41, 44, 46, 48, 51, 53, 56, 58, 59, 61, 63, 65, 66, 68, 69, 71, 72, 73,
+        75, 76, 77, 78, 80, 80, 80, 80, 44, 43, 42, 40, 39, 41, 43, 45, 48, 50,
+        52, 55, 58, 59, 61, 63, 65, 67, 68, 70, 71, 73, 74, 75, 76, 78, 79, 80,
+        81, 81, 81, 81, 47, 45, 44, 42, 41, 43, 45, 47, 49, 52, 54, 56, 59, 61,
+        63, 65, 67, 68, 70, 71, 73, 74, 75, 76, 78, 79, 80, 81, 82, 82, 82, 82,
+        49, 47, 46, 44, 43, 45, 47, 49, 51, 53, 55, 58, 60, 62, 64, 66, 68, 70,
+        71, 73, 74, 75, 77, 78, 79, 80, 81, 82, 83, 83, 83, 83, 52, 50, 48, 47,
+        45, 47, 49, 51, 53, 55, 57, 59, 62, 64, 66, 68, 70, 71, 73, 74, 76, 77,
+        78, 79, 81, 82, 83, 84, 85, 85, 85, 85, 55, 53, 51, 50, 48, 50, 51, 53,
+        55, 57, 59, 61, 64, 65, 67, 69, 71, 73, 74, 76, 77, 79, 80, 81, 82, 83,
+        84, 85, 86, 86, 86, 86, 57, 55, 53, 52, 50, 52, 53, 55, 57, 59, 61, 63,
+        65, 67, 69, 71, 73, 74, 75, 77, 79, 80, 81, 82, 83, 84, 85, 86, 87, 87,
+        87, 87, 59, 57, 56, 54, 52, 54, 55, 57, 59, 60, 62, 64, 66, 68, 70, 72,
+        74, 75, 77, 78, 80, 81, 82, 83, 84, 85, 86, 87, 88, 88, 88, 88, 62, 60,
+        58, 56, 55, 56, 57, 59, 61, 62, 64, 66, 68, 70, 71, 73, 75, 76, 78, 79,
+        81, 82, 83, 84, 85, 86, 87, 88, 89, 89, 89, 89, 65, 63, 61, 59, 57, 59,
+        60, 61, 63, 64, 66, 68, 70, 71, 73, 75, 76, 78, 79, 81, 82, 83, 84, 85,
+        86, 87, 88, 89, 90, 90, 90, 90, 67, 65, 63, 61, 59, 61, 62, 63, 64, 66,
+        68, 69, 71, 73, 74, 76, 78, 79, 80, 82, 83, 84, 85, 86, 87, 88, 89, 90,
+        90, 90, 90, 90, 69, 67, 65, 63, 61, 63, 64, 65, 66, 68, 69, 71, 72, 74,
+        75, 77, 79, 80, 81, 83, 84, 85, 86, 87, 88, 89, 90, 90, 91, 91, 91, 91,
+        71, 69, 67, 65, 64, 65, 66, 67, 68, 69, 71, 72, 74, 75, 77, 78, 80, 81,
+        82, 84, 85, 86, 87, 88, 89, 90, 90, 91, 92, 92, 92, 92, 73, 71, 70, 68,
+        66, 67, 68, 69, 70, 71, 73, 74, 75, 77, 78, 80, 81, 82, 83, 85, 86, 87,
+        88, 89, 90, 90, 91, 92, 93, 93, 93, 93, 73, 71, 70, 68, 66, 67, 68, 69,
+        70, 71, 73, 74, 75, 77, 78, 80, 81, 82, 83, 85, 86, 87, 88, 89, 90, 90,
+        91, 92, 93, 93, 93, 93, 73, 71, 70, 68, 66, 67, 68, 69, 70, 71, 73, 74,
+        75, 77, 78, 80, 81, 82, 83, 85, 86, 87, 88, 89, 90, 90, 91, 92, 93, 93,
+        93, 93, 73, 71, 70, 68, 66, 67, 68, 69, 70, 71, 73, 74, 75, 77, 78, 80,
+        81, 82, 83, 85, 86, 87, 88, 89, 90, 90, 91, 92, 93, 93, 93, 93 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 94, 100, 120, 94, 108, 114, 126, 100, 114, 136, 149, 120, 126, 149,
+        166,
+        /* Size 8 */
+        64, 53, 88, 92, 97, 105, 115, 126, 53, 73, 89, 82, 84, 91, 100, 109, 88,
+        89, 100, 98, 98, 102, 108, 116, 92, 82, 98, 107, 111, 115, 120, 126, 97,
+        84, 98, 111, 121, 127, 132, 137, 105, 91, 102, 115, 127, 135, 142, 147,
+        115, 100, 108, 120, 132, 142, 149, 155, 126, 109, 116, 126, 137, 147,
+        155, 162,
+        /* Size 16 */
+        64, 58, 53, 66, 88, 90, 92, 94, 97, 101, 105, 110, 115, 120, 126, 126,
+        58, 60, 61, 72, 88, 87, 87, 88, 90, 94, 97, 102, 107, 112, 117, 117, 53,
+        61, 73, 80, 89, 85, 82, 83, 84, 87, 91, 95, 100, 104, 109, 109, 66, 72,
+        80, 87, 94, 92, 89, 90, 90, 93, 96, 100, 104, 108, 113, 113, 88, 88, 89,
+        94, 100, 99, 98, 98, 98, 100, 102, 105, 108, 112, 116, 116, 90, 87, 85,
+        92, 99, 101, 102, 103, 104, 106, 108, 111, 114, 117, 121, 121, 92, 87,
+        82, 89, 98, 102, 107, 109, 111, 113, 115, 117, 120, 123, 126, 126, 94,
+        88, 83, 90, 98, 103, 109, 113, 116, 118, 121, 123, 126, 128, 132, 132,
+        97, 90, 84, 90, 98, 104, 111, 116, 121, 124, 127, 129, 132, 134, 137,
+        137, 101, 94, 87, 93, 100, 106, 113, 118, 124, 127, 131, 134, 136, 139,
+        142, 142, 105, 97, 91, 96, 102, 108, 115, 121, 127, 131, 135, 138, 142,
+        144, 147, 147, 110, 102, 95, 100, 105, 111, 117, 123, 129, 134, 138,
+        142, 145, 148, 151, 151, 115, 107, 100, 104, 108, 114, 120, 126, 132,
+        136, 142, 145, 149, 152, 155, 155, 120, 112, 104, 108, 112, 117, 123,
+        128, 134, 139, 144, 148, 152, 155, 158, 158, 126, 117, 109, 113, 116,
+        121, 126, 132, 137, 142, 147, 151, 155, 158, 162, 162, 126, 117, 109,
+        113, 116, 121, 126, 132, 137, 142, 147, 151, 155, 158, 162, 162,
+        /* Size 32 */
+        64, 61, 58, 56, 53, 59, 66, 75, 88, 89, 90, 91, 92, 93, 94, 95, 97, 99,
+        101, 103, 105, 108, 110, 113, 115, 118, 120, 123, 126, 126, 126, 126,
+        61, 60, 59, 58, 57, 63, 69, 77, 88, 88, 89, 89, 89, 90, 91, 92, 93, 95,
+        97, 99, 101, 103, 106, 108, 111, 113, 116, 118, 121, 121, 121, 121, 58,
+        59, 60, 61, 61, 67, 72, 80, 88, 88, 87, 87, 87, 88, 88, 89, 90, 92, 94,
+        95, 97, 100, 102, 104, 107, 109, 112, 114, 117, 117, 117, 117, 56, 58,
+        61, 63, 67, 71, 76, 82, 89, 88, 86, 85, 84, 85, 86, 86, 87, 89, 90, 92,
+        94, 96, 98, 101, 103, 105, 108, 110, 113, 113, 113, 113, 53, 57, 61, 67,
+        73, 76, 80, 84, 89, 87, 85, 84, 82, 83, 83, 84, 84, 86, 87, 89, 91, 93,
+        95, 97, 100, 102, 104, 107, 109, 109, 109, 109, 59, 63, 67, 71, 76, 80,
+        83, 87, 92, 90, 88, 87, 86, 86, 86, 87, 87, 89, 90, 92, 93, 95, 97, 99,
+        102, 104, 106, 109, 111, 111, 111, 111, 66, 69, 72, 76, 80, 83, 87, 90,
+        94, 93, 92, 91, 89, 90, 90, 90, 90, 92, 93, 94, 96, 98, 100, 102, 104,
+        106, 108, 110, 113, 113, 113, 113, 75, 77, 80, 82, 84, 87, 90, 94, 97,
+        96, 95, 94, 93, 94, 94, 94, 94, 95, 96, 97, 99, 100, 102, 104, 106, 108,
+        110, 112, 115, 115, 115, 115, 88, 88, 88, 89, 89, 92, 94, 97, 100, 100,
+        99, 99, 98, 98, 98, 98, 98, 99, 100, 101, 102, 103, 105, 106, 108, 110,
+        112, 114, 116, 116, 116, 116, 89, 88, 88, 88, 87, 90, 93, 96, 100, 100,
+        100, 100, 100, 100, 100, 101, 101, 102, 103, 104, 105, 106, 108, 109,
+        111, 113, 115, 117, 119, 119, 119, 119, 90, 89, 87, 86, 85, 88, 92, 95,
+        99, 100, 101, 102, 102, 103, 103, 104, 104, 105, 106, 107, 108, 109,
+        111, 112, 114, 116, 117, 119, 121, 121, 121, 121, 91, 89, 87, 85, 84,
+        87, 91, 94, 99, 100, 102, 103, 105, 106, 106, 107, 108, 109, 109, 110,
+        111, 113, 114, 115, 117, 118, 120, 122, 124, 124, 124, 124, 92, 89, 87,
+        84, 82, 86, 89, 93, 98, 100, 102, 105, 107, 108, 109, 110, 111, 112,
+        113, 114, 115, 116, 117, 119, 120, 122, 123, 125, 126, 126, 126, 126,
+        93, 90, 88, 85, 83, 86, 90, 94, 98, 100, 103, 106, 108, 110, 111, 112,
+        114, 115, 116, 117, 118, 119, 120, 121, 123, 124, 126, 127, 129, 129,
+        129, 129, 94, 91, 88, 86, 83, 86, 90, 94, 98, 100, 103, 106, 109, 111,
+        113, 114, 116, 117, 118, 119, 121, 122, 123, 124, 126, 127, 128, 130,
+        132, 132, 132, 132, 95, 92, 89, 86, 84, 87, 90, 94, 98, 101, 104, 107,
+        110, 112, 114, 116, 118, 119, 121, 122, 124, 125, 126, 127, 129, 130,
+        131, 133, 134, 134, 134, 134, 97, 93, 90, 87, 84, 87, 90, 94, 98, 101,
+        104, 108, 111, 114, 116, 118, 121, 122, 124, 125, 127, 128, 129, 130,
+        132, 133, 134, 136, 137, 137, 137, 137, 99, 95, 92, 89, 86, 89, 92, 95,
+        99, 102, 105, 109, 112, 115, 117, 119, 122, 124, 125, 127, 129, 130,
+        131, 133, 134, 135, 137, 138, 139, 139, 139, 139, 101, 97, 94, 90, 87,
+        90, 93, 96, 100, 103, 106, 109, 113, 116, 118, 121, 124, 125, 127, 129,
+        131, 132, 134, 135, 136, 138, 139, 140, 142, 142, 142, 142, 103, 99, 95,
+        92, 89, 92, 94, 97, 101, 104, 107, 110, 114, 117, 119, 122, 125, 127,
+        129, 131, 133, 134, 136, 137, 139, 140, 142, 143, 144, 144, 144, 144,
+        105, 101, 97, 94, 91, 93, 96, 99, 102, 105, 108, 111, 115, 118, 121,
+        124, 127, 129, 131, 133, 135, 137, 138, 140, 142, 143, 144, 146, 147,
+        147, 147, 147, 108, 103, 100, 96, 93, 95, 98, 100, 103, 106, 109, 113,
+        116, 119, 122, 125, 128, 130, 132, 134, 137, 138, 140, 142, 143, 145,
+        146, 148, 149, 149, 149, 149, 110, 106, 102, 98, 95, 97, 100, 102, 105,
+        108, 111, 114, 117, 120, 123, 126, 129, 131, 134, 136, 138, 140, 142,
+        143, 145, 147, 148, 149, 151, 151, 151, 151, 113, 108, 104, 101, 97, 99,
+        102, 104, 106, 109, 112, 115, 119, 121, 124, 127, 130, 133, 135, 137,
+        140, 142, 143, 145, 147, 149, 150, 151, 153, 153, 153, 153, 115, 111,
+        107, 103, 100, 102, 104, 106, 108, 111, 114, 117, 120, 123, 126, 129,
+        132, 134, 136, 139, 142, 143, 145, 147, 149, 151, 152, 154, 155, 155,
+        155, 155, 118, 113, 109, 105, 102, 104, 106, 108, 110, 113, 116, 118,
+        122, 124, 127, 130, 133, 135, 138, 140, 143, 145, 147, 149, 151, 152,
+        154, 155, 157, 157, 157, 157, 120, 116, 112, 108, 104, 106, 108, 110,
+        112, 115, 117, 120, 123, 126, 128, 131, 134, 137, 139, 142, 144, 146,
+        148, 150, 152, 154, 155, 157, 158, 158, 158, 158, 123, 118, 114, 110,
+        107, 109, 110, 112, 114, 117, 119, 122, 125, 127, 130, 133, 136, 138,
+        140, 143, 146, 148, 149, 151, 154, 155, 157, 158, 160, 160, 160, 160,
+        126, 121, 117, 113, 109, 111, 113, 115, 116, 119, 121, 124, 126, 129,
+        132, 134, 137, 139, 142, 144, 147, 149, 151, 153, 155, 157, 158, 160,
+        162, 162, 162, 162, 126, 121, 117, 113, 109, 111, 113, 115, 116, 119,
+        121, 124, 126, 129, 132, 134, 137, 139, 142, 144, 147, 149, 151, 153,
+        155, 157, 158, 160, 162, 162, 162, 162, 126, 121, 117, 113, 109, 111,
+        113, 115, 116, 119, 121, 124, 126, 129, 132, 134, 137, 139, 142, 144,
+        147, 149, 151, 153, 155, 157, 158, 160, 162, 162, 162, 162, 126, 121,
+        117, 113, 109, 111, 113, 115, 116, 119, 121, 124, 126, 129, 132, 134,
+        137, 139, 142, 144, 147, 149, 151, 153, 155, 157, 158, 160, 162, 162,
+        162, 162 },
+      { /* Intra matrices */
+        /* Size 4 */
+        33, 49, 53, 64, 49, 57, 60, 67, 53, 60, 73, 81, 64, 67, 81, 91,
+        /* Size 8 */
+        35, 29, 49, 51, 54, 59, 65, 72, 29, 40, 50, 46, 47, 51, 56, 62, 49, 50,
+        56, 55, 55, 57, 61, 66, 51, 46, 55, 61, 63, 65, 68, 72, 54, 47, 55, 63,
+        69, 72, 76, 79, 59, 51, 57, 65, 72, 78, 82, 85, 65, 56, 61, 68, 76, 82,
+        86, 90, 72, 62, 66, 72, 79, 85, 90, 94,
+        /* Size 16 */
+        34, 31, 28, 36, 48, 49, 50, 52, 53, 55, 58, 61, 64, 67, 70, 70, 31, 32,
+        33, 39, 48, 48, 47, 48, 49, 51, 53, 56, 59, 62, 65, 65, 28, 33, 39, 43,
+        49, 47, 45, 45, 46, 48, 50, 52, 55, 57, 61, 61, 36, 39, 43, 47, 52, 50,
+        49, 49, 49, 51, 53, 55, 57, 60, 63, 63, 48, 48, 49, 52, 55, 55, 54, 54,
+        54, 55, 56, 58, 60, 62, 65, 65, 49, 48, 47, 50, 55, 55, 56, 57, 57, 58,
+        60, 61, 63, 65, 68, 68, 50, 47, 45, 49, 54, 56, 59, 61, 62, 63, 64, 65,
+        67, 69, 71, 71, 52, 48, 45, 49, 54, 57, 61, 62, 64, 66, 67, 69, 70, 72,
+        74, 74, 53, 49, 46, 49, 54, 57, 62, 64, 67, 69, 71, 72, 74, 76, 77, 77,
+        55, 51, 48, 51, 55, 58, 63, 66, 69, 71, 73, 75, 77, 78, 80, 80, 58, 53,
+        50, 53, 56, 60, 64, 67, 71, 73, 76, 78, 80, 82, 83, 83, 61, 56, 52, 55,
+        58, 61, 65, 69, 72, 75, 78, 80, 82, 84, 86, 86, 64, 59, 55, 57, 60, 63,
+        67, 70, 74, 77, 80, 82, 85, 86, 88, 88, 67, 62, 57, 60, 62, 65, 69, 72,
+        76, 78, 82, 84, 86, 88, 90, 90, 70, 65, 61, 63, 65, 68, 71, 74, 77, 80,
+        83, 86, 88, 90, 92, 92, 70, 65, 61, 63, 65, 68, 71, 74, 77, 80, 83, 86,
+        88, 90, 92, 92,
+        /* Size 32 */
+        34, 32, 31, 29, 28, 31, 35, 40, 47, 48, 48, 49, 50, 50, 51, 52, 52, 54,
+        55, 56, 57, 59, 60, 62, 63, 65, 66, 68, 70, 70, 70, 70, 32, 32, 31, 31,
+        30, 33, 37, 42, 47, 48, 48, 48, 48, 49, 49, 50, 50, 52, 53, 54, 55, 56,
+        58, 59, 61, 62, 64, 65, 67, 67, 67, 67, 31, 31, 32, 32, 33, 35, 39, 43,
+        48, 47, 47, 47, 47, 47, 48, 48, 49, 50, 51, 52, 53, 54, 55, 57, 58, 60,
+        61, 63, 64, 64, 64, 64, 29, 31, 32, 34, 35, 38, 41, 44, 48, 47, 47, 46,
+        45, 46, 46, 47, 47, 48, 49, 50, 51, 52, 53, 55, 56, 57, 59, 60, 62, 62,
+        62, 62, 28, 30, 33, 35, 39, 41, 43, 45, 48, 47, 46, 45, 44, 44, 45, 45,
+        45, 46, 47, 48, 49, 50, 51, 53, 54, 55, 57, 58, 60, 60, 60, 60, 31, 33,
+        35, 38, 41, 43, 45, 47, 50, 49, 48, 47, 46, 46, 47, 47, 47, 48, 49, 50,
+        50, 52, 53, 54, 55, 57, 58, 59, 61, 61, 61, 61, 35, 37, 39, 41, 43, 45,
+        47, 49, 51, 50, 50, 49, 48, 48, 49, 49, 49, 50, 50, 51, 52, 53, 54, 55,
+        57, 58, 59, 60, 62, 62, 62, 62, 40, 42, 43, 44, 45, 47, 49, 51, 53, 52,
+        52, 51, 51, 51, 51, 51, 51, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 62,
+        63, 63, 63, 63, 47, 47, 48, 48, 48, 50, 51, 53, 55, 54, 54, 54, 53, 53,
+        53, 53, 53, 54, 54, 55, 55, 56, 57, 58, 59, 60, 61, 63, 64, 64, 64, 64,
+        48, 48, 47, 47, 47, 49, 50, 52, 54, 54, 54, 54, 54, 55, 55, 55, 55, 55,
+        56, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 65, 65, 65, 48, 48, 47, 47,
+        46, 48, 50, 52, 54, 54, 55, 55, 56, 56, 56, 56, 57, 57, 58, 58, 59, 60,
+        61, 61, 62, 63, 65, 66, 67, 67, 67, 67, 49, 48, 47, 46, 45, 47, 49, 51,
+        54, 54, 55, 56, 57, 58, 58, 58, 59, 59, 60, 60, 61, 62, 63, 63, 64, 65,
+        66, 67, 68, 68, 68, 68, 50, 48, 47, 45, 44, 46, 48, 51, 53, 54, 56, 57,
+        59, 59, 60, 60, 61, 62, 62, 63, 63, 64, 65, 65, 66, 67, 68, 69, 70, 70,
+        70, 70, 50, 49, 47, 46, 44, 46, 48, 51, 53, 55, 56, 58, 59, 60, 61, 62,
+        62, 63, 63, 64, 65, 65, 66, 67, 68, 69, 69, 70, 71, 71, 71, 71, 51, 49,
+        48, 46, 45, 47, 49, 51, 53, 55, 56, 58, 60, 61, 62, 63, 64, 64, 65, 66,
+        66, 67, 68, 69, 69, 70, 71, 72, 73, 73, 73, 73, 52, 50, 48, 47, 45, 47,
+        49, 51, 53, 55, 56, 58, 60, 62, 63, 64, 65, 66, 67, 67, 68, 69, 70, 70,
+        71, 72, 73, 74, 75, 75, 75, 75, 52, 50, 49, 47, 45, 47, 49, 51, 53, 55,
+        57, 59, 61, 62, 64, 65, 66, 67, 68, 69, 70, 71, 72, 72, 73, 74, 75, 76,
+        76, 76, 76, 76, 54, 52, 50, 48, 46, 48, 50, 51, 54, 55, 57, 59, 62, 63,
+        64, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 75, 76, 77, 78, 78, 78, 78,
+        55, 53, 51, 49, 47, 49, 50, 52, 54, 56, 58, 60, 62, 63, 65, 67, 68, 69,
+        70, 71, 73, 73, 74, 75, 76, 77, 78, 78, 79, 79, 79, 79, 56, 54, 52, 50,
+        48, 50, 51, 53, 55, 56, 58, 60, 63, 64, 66, 67, 69, 70, 71, 73, 74, 75,
+        76, 77, 77, 78, 79, 80, 81, 81, 81, 81, 57, 55, 53, 51, 49, 50, 52, 54,
+        55, 57, 59, 61, 63, 65, 66, 68, 70, 71, 73, 74, 75, 76, 77, 78, 79, 80,
+        81, 82, 82, 82, 82, 82, 59, 56, 54, 52, 50, 52, 53, 55, 56, 58, 60, 62,
+        64, 65, 67, 69, 71, 72, 73, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 84,
+        84, 84, 60, 58, 55, 53, 51, 53, 54, 56, 57, 59, 61, 63, 65, 66, 68, 70,
+        72, 73, 74, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 85, 85, 85, 62, 59,
+        57, 55, 53, 54, 55, 57, 58, 60, 61, 63, 65, 67, 69, 70, 72, 74, 75, 77,
+        78, 79, 80, 81, 82, 83, 84, 85, 86, 86, 86, 86, 63, 61, 58, 56, 54, 55,
+        57, 58, 59, 61, 62, 64, 66, 68, 69, 71, 73, 75, 76, 77, 79, 80, 81, 82,
+        84, 85, 86, 86, 87, 87, 87, 87, 65, 62, 60, 57, 55, 57, 58, 59, 60, 62,
+        63, 65, 67, 69, 70, 72, 74, 75, 77, 78, 80, 81, 82, 83, 85, 86, 86, 87,
+        88, 88, 88, 88, 66, 64, 61, 59, 57, 58, 59, 60, 61, 63, 65, 66, 68, 69,
+        71, 73, 75, 76, 78, 79, 81, 82, 83, 84, 86, 86, 87, 88, 89, 89, 89, 89,
+        68, 65, 63, 60, 58, 59, 60, 62, 63, 64, 66, 67, 69, 70, 72, 74, 76, 77,
+        78, 80, 82, 83, 84, 85, 86, 87, 88, 89, 90, 90, 90, 90, 70, 67, 64, 62,
+        60, 61, 62, 63, 64, 65, 67, 68, 70, 71, 73, 75, 76, 78, 79, 81, 82, 84,
+        85, 86, 87, 88, 89, 90, 91, 91, 91, 91, 70, 67, 64, 62, 60, 61, 62, 63,
+        64, 65, 67, 68, 70, 71, 73, 75, 76, 78, 79, 81, 82, 84, 85, 86, 87, 88,
+        89, 90, 91, 91, 91, 91, 70, 67, 64, 62, 60, 61, 62, 63, 64, 65, 67, 68,
+        70, 71, 73, 75, 76, 78, 79, 81, 82, 84, 85, 86, 87, 88, 89, 90, 91, 91,
+        91, 91, 70, 67, 64, 62, 60, 61, 62, 63, 64, 65, 67, 68, 70, 71, 73, 75,
+        76, 78, 79, 81, 82, 84, 85, 86, 87, 88, 89, 90, 91, 91, 91, 91 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 69, 100, 132, 69, 93, 117, 139, 100, 117, 142, 157, 132, 139, 157,
+        169,
+        /* Size 8 */
+        64, 51, 54, 67, 84, 100, 114, 126, 51, 57, 55, 62, 75, 89, 104, 116, 54,
+        55, 71, 80, 89, 100, 111, 121, 67, 62, 80, 94, 104, 113, 121, 129, 84,
+        75, 89, 104, 115, 123, 130, 136, 100, 89, 100, 113, 123, 131, 137, 142,
+        114, 104, 111, 121, 130, 137, 142, 146, 126, 116, 121, 129, 136, 142,
+        146, 150,
+        /* Size 16 */
+        64, 57, 51, 53, 54, 60, 67, 75, 84, 91, 100, 106, 114, 119, 126, 126,
+        57, 55, 54, 54, 55, 59, 65, 71, 79, 86, 94, 101, 108, 114, 121, 121, 51,
+        54, 57, 56, 55, 58, 62, 68, 75, 82, 89, 96, 104, 109, 116, 116, 53, 54,
+        56, 59, 62, 66, 70, 75, 82, 88, 94, 100, 107, 113, 119, 119, 54, 55, 55,
+        62, 71, 76, 80, 84, 89, 94, 100, 105, 111, 116, 121, 121, 60, 59, 58,
+        66, 76, 81, 87, 91, 96, 101, 106, 111, 116, 120, 125, 125, 67, 65, 62,
+        70, 80, 87, 94, 99, 104, 108, 113, 116, 121, 125, 129, 129, 75, 71, 68,
+        75, 84, 91, 99, 104, 109, 113, 118, 121, 125, 128, 132, 132, 84, 79, 75,
+        82, 89, 96, 104, 109, 115, 119, 123, 126, 130, 133, 136, 136, 91, 86,
+        82, 88, 94, 101, 108, 113, 119, 123, 127, 130, 133, 136, 139, 139, 100,
+        94, 89, 94, 100, 106, 113, 118, 123, 127, 131, 134, 137, 139, 142, 142,
+        106, 101, 96, 100, 105, 111, 116, 121, 126, 130, 134, 137, 140, 142,
+        144, 144, 114, 108, 104, 107, 111, 116, 121, 125, 130, 133, 137, 140,
+        142, 144, 146, 146, 119, 114, 109, 113, 116, 120, 125, 128, 133, 136,
+        139, 142, 144, 146, 148, 148, 126, 121, 116, 119, 121, 125, 129, 132,
+        136, 139, 142, 144, 146, 148, 150, 150, 126, 121, 116, 119, 121, 125,
+        129, 132, 136, 139, 142, 144, 146, 148, 150, 150,
+        /* Size 32 */
+        64, 60, 57, 54, 51, 52, 53, 54, 54, 57, 60, 64, 67, 71, 75, 79, 84, 87,
+        91, 95, 100, 103, 106, 110, 114, 117, 119, 122, 126, 126, 126, 126, 60,
+        58, 56, 54, 52, 53, 53, 54, 55, 57, 60, 63, 66, 69, 73, 77, 81, 85, 88,
+        92, 97, 100, 103, 107, 111, 114, 117, 120, 123, 123, 123, 123, 57, 56,
+        55, 55, 54, 54, 54, 54, 55, 57, 59, 62, 65, 68, 71, 75, 79, 82, 86, 90,
+        94, 97, 101, 104, 108, 111, 114, 117, 121, 121, 121, 121, 54, 54, 55,
+        55, 55, 55, 55, 55, 55, 57, 59, 61, 64, 66, 70, 73, 77, 80, 84, 88, 92,
+        95, 98, 102, 106, 109, 112, 115, 118, 118, 118, 118, 51, 52, 54, 55, 57,
+        56, 56, 56, 55, 57, 58, 60, 62, 65, 68, 71, 75, 78, 82, 85, 89, 93, 96,
+        100, 104, 106, 109, 113, 116, 116, 116, 116, 52, 53, 54, 55, 56, 57, 57,
+        58, 58, 60, 62, 64, 66, 69, 72, 75, 78, 81, 84, 88, 92, 95, 98, 102,
+        105, 108, 111, 114, 117, 117, 117, 117, 53, 53, 54, 55, 56, 57, 59, 61,
+        62, 64, 66, 68, 70, 73, 75, 78, 82, 84, 88, 91, 94, 97, 100, 104, 107,
+        110, 113, 116, 119, 119, 119, 119, 54, 54, 54, 55, 56, 58, 61, 63, 67,
+        68, 70, 72, 75, 77, 80, 82, 85, 88, 91, 94, 97, 100, 103, 106, 109, 112,
+        114, 117, 120, 120, 120, 120, 54, 55, 55, 55, 55, 58, 62, 67, 71, 73,
+        76, 78, 80, 82, 84, 87, 89, 92, 94, 97, 100, 103, 105, 108, 111, 114,
+        116, 119, 121, 121, 121, 121, 57, 57, 57, 57, 57, 60, 64, 68, 73, 76,
+        78, 80, 83, 85, 88, 90, 92, 95, 97, 100, 103, 105, 108, 111, 113, 116,
+        118, 121, 123, 123, 123, 123, 60, 60, 59, 59, 58, 62, 66, 70, 76, 78,
+        81, 84, 87, 89, 91, 93, 96, 98, 101, 103, 106, 108, 111, 113, 116, 118,
+        120, 122, 125, 125, 125, 125, 64, 63, 62, 61, 60, 64, 68, 72, 78, 80,
+        84, 87, 90, 92, 95, 97, 100, 102, 104, 107, 109, 111, 113, 116, 118,
+        120, 122, 124, 127, 127, 127, 127, 67, 66, 65, 64, 62, 66, 70, 75, 80,
+        83, 87, 90, 94, 97, 99, 101, 104, 106, 108, 110, 113, 114, 116, 119,
+        121, 123, 125, 127, 129, 129, 129, 129, 71, 69, 68, 66, 65, 69, 73, 77,
+        82, 85, 89, 92, 97, 99, 101, 104, 107, 109, 111, 113, 115, 117, 119,
+        121, 123, 125, 126, 128, 130, 130, 130, 130, 75, 73, 71, 70, 68, 72, 75,
+        80, 84, 88, 91, 95, 99, 101, 104, 106, 109, 111, 113, 115, 118, 119,
+        121, 123, 125, 127, 128, 130, 132, 132, 132, 132, 79, 77, 75, 73, 71,
+        75, 78, 82, 87, 90, 93, 97, 101, 104, 106, 109, 112, 114, 116, 118, 120,
+        122, 124, 126, 127, 129, 131, 132, 134, 134, 134, 134, 84, 81, 79, 77,
+        75, 78, 82, 85, 89, 92, 96, 100, 104, 107, 109, 112, 115, 117, 119, 121,
+        123, 125, 126, 128, 130, 131, 133, 134, 136, 136, 136, 136, 87, 85, 82,
+        80, 78, 81, 84, 88, 92, 95, 98, 102, 106, 109, 111, 114, 117, 119, 121,
+        123, 125, 127, 128, 130, 132, 133, 134, 136, 137, 137, 137, 137, 91, 88,
+        86, 84, 82, 84, 88, 91, 94, 97, 101, 104, 108, 111, 113, 116, 119, 121,
+        123, 125, 127, 129, 130, 132, 133, 135, 136, 137, 139, 139, 139, 139,
+        95, 92, 90, 88, 85, 88, 91, 94, 97, 100, 103, 107, 110, 113, 115, 118,
+        121, 123, 125, 127, 129, 131, 132, 134, 135, 136, 138, 139, 140, 140,
+        140, 140, 100, 97, 94, 92, 89, 92, 94, 97, 100, 103, 106, 109, 113, 115,
+        118, 120, 123, 125, 127, 129, 131, 133, 134, 135, 137, 138, 139, 141,
+        142, 142, 142, 142, 103, 100, 97, 95, 93, 95, 97, 100, 103, 105, 108,
+        111, 114, 117, 119, 122, 125, 127, 129, 131, 133, 134, 135, 137, 138,
+        139, 141, 142, 143, 143, 143, 143, 106, 103, 101, 98, 96, 98, 100, 103,
+        105, 108, 111, 113, 116, 119, 121, 124, 126, 128, 130, 132, 134, 135,
+        137, 138, 140, 141, 142, 143, 144, 144, 144, 144, 110, 107, 104, 102,
+        100, 102, 104, 106, 108, 111, 113, 116, 119, 121, 123, 126, 128, 130,
+        132, 134, 135, 137, 138, 140, 141, 142, 143, 144, 145, 145, 145, 145,
+        114, 111, 108, 106, 104, 105, 107, 109, 111, 113, 116, 118, 121, 123,
+        125, 127, 130, 132, 133, 135, 137, 138, 140, 141, 142, 143, 144, 145,
+        146, 146, 146, 146, 117, 114, 111, 109, 106, 108, 110, 112, 114, 116,
+        118, 120, 123, 125, 127, 129, 131, 133, 135, 136, 138, 139, 141, 142,
+        143, 144, 145, 146, 147, 147, 147, 147, 119, 117, 114, 112, 109, 111,
+        113, 114, 116, 118, 120, 122, 125, 126, 128, 131, 133, 134, 136, 138,
+        139, 141, 142, 143, 144, 145, 146, 147, 148, 148, 148, 148, 122, 120,
+        117, 115, 113, 114, 116, 117, 119, 121, 122, 124, 127, 128, 130, 132,
+        134, 136, 137, 139, 141, 142, 143, 144, 145, 146, 147, 148, 149, 149,
+        149, 149, 126, 123, 121, 118, 116, 117, 119, 120, 121, 123, 125, 127,
+        129, 130, 132, 134, 136, 137, 139, 140, 142, 143, 144, 145, 146, 147,
+        148, 149, 150, 150, 150, 150, 126, 123, 121, 118, 116, 117, 119, 120,
+        121, 123, 125, 127, 129, 130, 132, 134, 136, 137, 139, 140, 142, 143,
+        144, 145, 146, 147, 148, 149, 150, 150, 150, 150, 126, 123, 121, 118,
+        116, 117, 119, 120, 121, 123, 125, 127, 129, 130, 132, 134, 136, 137,
+        139, 140, 142, 143, 144, 145, 146, 147, 148, 149, 150, 150, 150, 150,
+        126, 123, 121, 118, 116, 117, 119, 120, 121, 123, 125, 127, 129, 130,
+        132, 134, 136, 137, 139, 140, 142, 143, 144, 145, 146, 147, 148, 149,
+        150, 150, 150, 150 },
+      { /* Intra matrices */
+        /* Size 4 */
+        32, 34, 51, 69, 34, 47, 60, 73, 51, 60, 75, 83, 69, 73, 83, 90,
+        /* Size 8 */
+        37, 29, 31, 39, 49, 59, 68, 76, 29, 33, 31, 36, 44, 53, 61, 70, 31, 31,
+        41, 47, 52, 59, 66, 73, 39, 36, 47, 56, 62, 67, 73, 78, 49, 44, 52, 62,
+        69, 74, 79, 83, 59, 53, 59, 67, 74, 79, 83, 87, 68, 61, 66, 73, 79, 83,
+        87, 90, 76, 70, 73, 78, 83, 87, 90, 92,
+        /* Size 16 */
+        36, 32, 28, 29, 30, 34, 38, 42, 48, 52, 57, 61, 66, 70, 74, 74, 32, 31,
+        30, 30, 30, 33, 36, 40, 45, 49, 54, 58, 63, 67, 71, 71, 28, 30, 32, 31,
+        31, 33, 35, 38, 42, 46, 51, 55, 60, 64, 68, 68, 29, 30, 31, 33, 35, 37,
+        40, 43, 46, 50, 54, 58, 62, 66, 69, 69, 30, 30, 31, 35, 40, 43, 45, 48,
+        51, 54, 58, 61, 65, 68, 71, 71, 34, 33, 33, 37, 43, 46, 49, 52, 55, 58,
+        61, 64, 68, 70, 73, 73, 38, 36, 35, 40, 45, 49, 54, 57, 60, 63, 65, 68,
+        71, 73, 76, 76, 42, 40, 38, 43, 48, 52, 57, 60, 63, 66, 69, 71, 74, 76,
+        78, 78, 48, 45, 42, 46, 51, 55, 60, 63, 67, 70, 72, 74, 77, 78, 80, 80,
+        52, 49, 46, 50, 54, 58, 63, 66, 70, 72, 75, 77, 79, 81, 82, 82, 57, 54,
+        51, 54, 58, 61, 65, 69, 72, 75, 77, 79, 81, 83, 84, 84, 61, 58, 55, 58,
+        61, 64, 68, 71, 74, 77, 79, 81, 83, 84, 86, 86, 66, 63, 60, 62, 65, 68,
+        71, 74, 77, 79, 81, 83, 85, 86, 87, 87, 70, 67, 64, 66, 68, 70, 73, 76,
+        78, 81, 83, 84, 86, 87, 89, 89, 74, 71, 68, 69, 71, 73, 76, 78, 80, 82,
+        84, 86, 87, 89, 90, 90, 74, 71, 68, 69, 71, 73, 76, 78, 80, 82, 84, 86,
+        87, 89, 90, 90,
+        /* Size 32 */
+        35, 33, 31, 30, 28, 28, 29, 29, 30, 31, 33, 35, 37, 39, 42, 44, 47, 49,
+        51, 54, 57, 59, 61, 63, 65, 67, 69, 71, 73, 73, 73, 73, 33, 32, 31, 30,
+        29, 29, 29, 30, 30, 31, 33, 35, 37, 39, 41, 43, 46, 48, 50, 52, 55, 57,
+        59, 61, 64, 65, 67, 69, 71, 71, 71, 71, 31, 31, 30, 30, 30, 30, 30, 30,
+        30, 31, 33, 34, 36, 38, 40, 42, 44, 46, 48, 51, 53, 55, 57, 60, 62, 64,
+        66, 68, 70, 70, 70, 70, 30, 30, 30, 30, 30, 30, 30, 30, 30, 31, 32, 34,
+        35, 37, 39, 41, 43, 45, 47, 49, 52, 54, 56, 58, 61, 62, 64, 66, 68, 68,
+        68, 68, 28, 29, 30, 30, 31, 31, 31, 31, 30, 31, 32, 33, 34, 36, 38, 40,
+        42, 44, 46, 48, 50, 52, 54, 57, 59, 61, 63, 65, 67, 67, 67, 67, 28, 29,
+        30, 30, 31, 31, 32, 32, 32, 33, 34, 35, 37, 38, 40, 42, 44, 46, 47, 50,
+        52, 54, 56, 58, 60, 62, 64, 66, 68, 68, 68, 68, 29, 29, 30, 30, 31, 32,
+        33, 33, 34, 35, 37, 38, 39, 40, 42, 44, 46, 47, 49, 51, 53, 55, 57, 59,
+        61, 63, 65, 67, 68, 68, 68, 68, 29, 30, 30, 30, 31, 32, 33, 35, 37, 38,
+        39, 40, 42, 43, 45, 46, 48, 50, 51, 53, 55, 57, 59, 60, 62, 64, 66, 67,
+        69, 69, 69, 69, 30, 30, 30, 30, 30, 32, 34, 37, 40, 41, 42, 43, 45, 46,
+        47, 49, 50, 52, 53, 55, 57, 58, 60, 62, 64, 65, 67, 68, 70, 70, 70, 70,
+        31, 31, 31, 31, 31, 33, 35, 38, 41, 42, 44, 45, 47, 48, 49, 51, 52, 54,
+        55, 57, 59, 60, 62, 63, 65, 67, 68, 70, 71, 71, 71, 71, 33, 33, 33, 32,
+        32, 34, 37, 39, 42, 44, 45, 47, 49, 50, 51, 53, 54, 56, 57, 59, 60, 62,
+        63, 65, 67, 68, 69, 71, 72, 72, 72, 72, 35, 35, 34, 34, 33, 35, 38, 40,
+        43, 45, 47, 49, 51, 52, 54, 55, 57, 58, 59, 61, 62, 64, 65, 67, 68, 69,
+        71, 72, 74, 74, 74, 74, 37, 37, 36, 35, 34, 37, 39, 42, 45, 47, 49, 51,
+        53, 55, 56, 58, 59, 61, 62, 63, 65, 66, 67, 68, 70, 71, 72, 73, 75, 75,
+        75, 75, 39, 39, 38, 37, 36, 38, 40, 43, 46, 48, 50, 52, 55, 56, 58, 59,
+        61, 62, 63, 65, 66, 67, 69, 70, 71, 72, 73, 75, 76, 76, 76, 76, 42, 41,
+        40, 39, 38, 40, 42, 45, 47, 49, 51, 54, 56, 58, 59, 61, 63, 64, 65, 66,
+        68, 69, 70, 71, 73, 74, 75, 76, 77, 77, 77, 77, 44, 43, 42, 41, 40, 42,
+        44, 46, 49, 51, 53, 55, 58, 59, 61, 63, 64, 66, 67, 68, 69, 71, 72, 73,
+        74, 75, 76, 77, 78, 78, 78, 78, 47, 46, 44, 43, 42, 44, 46, 48, 50, 52,
+        54, 57, 59, 61, 63, 64, 66, 67, 69, 70, 71, 72, 73, 74, 76, 76, 77, 78,
+        79, 79, 79, 79, 49, 48, 46, 45, 44, 46, 47, 50, 52, 54, 56, 58, 61, 62,
+        64, 66, 67, 69, 70, 71, 73, 73, 75, 76, 77, 78, 78, 79, 80, 80, 80, 80,
+        51, 50, 48, 47, 46, 47, 49, 51, 53, 55, 57, 59, 62, 63, 65, 67, 69, 70,
+        71, 72, 74, 75, 76, 77, 78, 79, 79, 80, 81, 81, 81, 81, 54, 52, 51, 49,
+        48, 50, 51, 53, 55, 57, 59, 61, 63, 65, 66, 68, 70, 71, 72, 74, 75, 76,
+        77, 78, 79, 80, 81, 81, 82, 82, 82, 82, 57, 55, 53, 52, 50, 52, 53, 55,
+        57, 59, 60, 62, 65, 66, 68, 69, 71, 73, 74, 75, 76, 77, 78, 79, 80, 81,
+        82, 82, 83, 83, 83, 83, 59, 57, 55, 54, 52, 54, 55, 57, 58, 60, 62, 64,
+        66, 67, 69, 71, 72, 73, 75, 76, 77, 78, 79, 80, 81, 82, 82, 83, 84, 84,
+        84, 84, 61, 59, 57, 56, 54, 56, 57, 59, 60, 62, 63, 65, 67, 69, 70, 72,
+        73, 75, 76, 77, 78, 79, 80, 81, 82, 83, 83, 84, 85, 85, 85, 85, 63, 61,
+        60, 58, 57, 58, 59, 60, 62, 63, 65, 67, 68, 70, 71, 73, 74, 76, 77, 78,
+        79, 80, 81, 82, 83, 83, 84, 85, 85, 85, 85, 85, 65, 64, 62, 61, 59, 60,
+        61, 62, 64, 65, 67, 68, 70, 71, 73, 74, 76, 77, 78, 79, 80, 81, 82, 83,
+        84, 84, 85, 86, 86, 86, 86, 86, 67, 65, 64, 62, 61, 62, 63, 64, 65, 67,
+        68, 69, 71, 72, 74, 75, 76, 78, 79, 80, 81, 82, 83, 83, 84, 85, 86, 86,
+        87, 87, 87, 87, 69, 67, 66, 64, 63, 64, 65, 66, 67, 68, 69, 71, 72, 73,
+        75, 76, 77, 78, 79, 81, 82, 82, 83, 84, 85, 86, 86, 87, 87, 87, 87, 87,
+        71, 69, 68, 66, 65, 66, 67, 67, 68, 70, 71, 72, 73, 75, 76, 77, 78, 79,
+        80, 81, 82, 83, 84, 85, 86, 86, 87, 87, 88, 88, 88, 88, 73, 71, 70, 68,
+        67, 68, 68, 69, 70, 71, 72, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84,
+        85, 85, 86, 87, 87, 88, 89, 89, 89, 89, 73, 71, 70, 68, 67, 68, 68, 69,
+        70, 71, 72, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 85, 86, 87,
+        87, 88, 89, 89, 89, 89, 73, 71, 70, 68, 67, 68, 68, 69, 70, 71, 72, 74,
+        75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 85, 86, 87, 87, 88, 89, 89,
+        89, 89, 73, 71, 70, 68, 67, 68, 68, 69, 70, 71, 72, 74, 75, 76, 77, 78,
+        79, 80, 81, 82, 83, 84, 85, 85, 86, 87, 87, 88, 89, 89, 89, 89 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 91, 97, 113, 91, 104, 108, 119, 97, 108, 127, 138, 113, 119, 138,
+        151,
+        /* Size 8 */
+        64, 54, 86, 89, 93, 101, 109, 118, 54, 72, 87, 81, 82, 88, 96, 104, 86,
+        87, 97, 95, 94, 98, 103, 110, 89, 81, 95, 103, 106, 109, 113, 119, 93,
+        82, 94, 106, 114, 119, 123, 127, 101, 88, 98, 109, 119, 126, 131, 135,
+        109, 96, 103, 113, 123, 131, 137, 142, 118, 104, 110, 119, 127, 135,
+        142, 147,
+        /* Size 16 */
+        64, 58, 54, 66, 86, 87, 89, 91, 93, 97, 101, 105, 109, 114, 118, 118,
+        58, 60, 62, 72, 86, 85, 85, 86, 88, 91, 94, 98, 102, 106, 111, 111, 54,
+        62, 72, 79, 87, 84, 81, 81, 82, 85, 88, 92, 96, 100, 104, 104, 66, 72,
+        79, 85, 91, 89, 87, 87, 88, 90, 93, 96, 100, 103, 107, 107, 86, 86, 87,
+        91, 97, 96, 95, 94, 94, 96, 98, 100, 103, 107, 110, 110, 87, 85, 84, 89,
+        96, 97, 98, 99, 100, 101, 103, 106, 108, 111, 114, 114, 89, 85, 81, 87,
+        95, 98, 103, 104, 106, 108, 109, 111, 113, 116, 119, 119, 91, 86, 81,
+        87, 94, 99, 104, 107, 110, 112, 114, 116, 118, 120, 123, 123, 93, 88,
+        82, 88, 94, 100, 106, 110, 114, 116, 119, 121, 123, 125, 127, 127, 97,
+        91, 85, 90, 96, 101, 108, 112, 116, 119, 122, 124, 127, 129, 131, 131,
+        101, 94, 88, 93, 98, 103, 109, 114, 119, 122, 126, 128, 131, 133, 135,
+        135, 105, 98, 92, 96, 100, 106, 111, 116, 121, 124, 128, 131, 134, 136,
+        138, 138, 109, 102, 96, 100, 103, 108, 113, 118, 123, 127, 131, 134,
+        137, 139, 142, 142, 114, 106, 100, 103, 107, 111, 116, 120, 125, 129,
+        133, 136, 139, 142, 144, 144, 118, 111, 104, 107, 110, 114, 119, 123,
+        127, 131, 135, 138, 142, 144, 147, 147, 118, 111, 104, 107, 110, 114,
+        119, 123, 127, 131, 135, 138, 142, 144, 147, 147,
+        /* Size 32 */
+        64, 61, 58, 56, 54, 59, 66, 75, 86, 86, 87, 88, 89, 90, 91, 92, 93, 95,
+        97, 99, 101, 103, 105, 107, 109, 111, 114, 116, 118, 118, 118, 118, 61,
+        60, 59, 58, 57, 63, 69, 76, 86, 86, 86, 87, 87, 88, 89, 90, 90, 92, 94,
+        96, 97, 99, 101, 103, 106, 108, 110, 112, 114, 114, 114, 114, 58, 59,
+        60, 61, 62, 66, 72, 78, 86, 86, 85, 85, 85, 85, 86, 87, 88, 89, 91, 92,
+        94, 96, 98, 100, 102, 104, 106, 109, 111, 111, 111, 111, 56, 58, 61, 63,
+        66, 71, 75, 80, 86, 85, 84, 84, 83, 83, 84, 84, 85, 86, 88, 89, 91, 93,
+        95, 97, 99, 101, 103, 105, 108, 108, 108, 108, 54, 57, 62, 66, 72, 75,
+        79, 83, 87, 85, 84, 82, 81, 81, 81, 82, 82, 84, 85, 87, 88, 90, 92, 94,
+        96, 98, 100, 102, 104, 104, 104, 104, 59, 63, 66, 71, 75, 78, 82, 85,
+        89, 88, 86, 85, 84, 84, 84, 85, 85, 86, 88, 89, 90, 92, 94, 96, 98, 100,
+        102, 104, 106, 106, 106, 106, 66, 69, 72, 75, 79, 82, 85, 88, 91, 90,
+        89, 88, 87, 87, 87, 88, 88, 89, 90, 91, 93, 94, 96, 98, 100, 101, 103,
+        105, 107, 107, 107, 107, 75, 76, 78, 80, 83, 85, 88, 91, 94, 93, 92, 91,
+        91, 91, 91, 91, 91, 92, 93, 94, 95, 97, 98, 100, 101, 103, 105, 107,
+        109, 109, 109, 109, 86, 86, 86, 86, 87, 89, 91, 94, 97, 96, 96, 95, 95,
+        94, 94, 94, 94, 95, 96, 97, 98, 99, 100, 102, 103, 105, 107, 109, 110,
+        110, 110, 110, 86, 86, 86, 85, 85, 88, 90, 93, 96, 96, 96, 96, 96, 97,
+        97, 97, 97, 98, 99, 99, 100, 102, 103, 104, 106, 107, 109, 111, 112,
+        112, 112, 112, 87, 86, 85, 84, 84, 86, 89, 92, 96, 96, 97, 98, 98, 99,
+        99, 100, 100, 101, 101, 102, 103, 104, 106, 107, 108, 110, 111, 113,
+        114, 114, 114, 114, 88, 87, 85, 84, 82, 85, 88, 91, 95, 96, 98, 99, 101,
+        101, 102, 102, 103, 104, 104, 105, 106, 107, 108, 109, 111, 112, 113,
+        115, 116, 116, 116, 116, 89, 87, 85, 83, 81, 84, 87, 91, 95, 96, 98,
+        101, 103, 104, 104, 105, 106, 107, 108, 108, 109, 110, 111, 112, 113,
+        115, 116, 117, 119, 119, 119, 119, 90, 88, 85, 83, 81, 84, 87, 91, 94,
+        97, 99, 101, 104, 105, 106, 107, 108, 109, 110, 110, 111, 112, 113, 114,
+        116, 117, 118, 119, 121, 121, 121, 121, 91, 89, 86, 84, 81, 84, 87, 91,
+        94, 97, 99, 102, 104, 106, 107, 108, 110, 111, 112, 113, 114, 115, 116,
+        117, 118, 119, 120, 122, 123, 123, 123, 123, 92, 90, 87, 84, 82, 85, 88,
+        91, 94, 97, 100, 102, 105, 107, 108, 110, 112, 113, 114, 115, 116, 117,
+        118, 119, 120, 122, 123, 124, 125, 125, 125, 125, 93, 90, 88, 85, 82,
+        85, 88, 91, 94, 97, 100, 103, 106, 108, 110, 112, 114, 115, 116, 118,
+        119, 120, 121, 122, 123, 124, 125, 126, 127, 127, 127, 127, 95, 92, 89,
+        86, 84, 86, 89, 92, 95, 98, 101, 104, 107, 109, 111, 113, 115, 116, 118,
+        119, 120, 122, 123, 124, 125, 126, 127, 128, 129, 129, 129, 129, 97, 94,
+        91, 88, 85, 88, 90, 93, 96, 99, 101, 104, 108, 110, 112, 114, 116, 118,
+        119, 121, 122, 123, 124, 126, 127, 128, 129, 130, 131, 131, 131, 131,
+        99, 96, 92, 89, 87, 89, 91, 94, 97, 99, 102, 105, 108, 110, 113, 115,
+        118, 119, 121, 122, 124, 125, 126, 128, 129, 130, 131, 132, 133, 133,
+        133, 133, 101, 97, 94, 91, 88, 90, 93, 95, 98, 100, 103, 106, 109, 111,
+        114, 116, 119, 120, 122, 124, 126, 127, 128, 130, 131, 132, 133, 134,
+        135, 135, 135, 135, 103, 99, 96, 93, 90, 92, 94, 97, 99, 102, 104, 107,
+        110, 112, 115, 117, 120, 122, 123, 125, 127, 128, 130, 131, 132, 133,
+        134, 136, 137, 137, 137, 137, 105, 101, 98, 95, 92, 94, 96, 98, 100,
+        103, 106, 108, 111, 113, 116, 118, 121, 123, 124, 126, 128, 130, 131,
+        132, 134, 135, 136, 137, 138, 138, 138, 138, 107, 103, 100, 97, 94, 96,
+        98, 100, 102, 104, 107, 109, 112, 114, 117, 119, 122, 124, 126, 128,
+        130, 131, 132, 134, 135, 136, 138, 139, 140, 140, 140, 140, 109, 106,
+        102, 99, 96, 98, 100, 101, 103, 106, 108, 111, 113, 116, 118, 120, 123,
+        125, 127, 129, 131, 132, 134, 135, 137, 138, 139, 140, 142, 142, 142,
+        142, 111, 108, 104, 101, 98, 100, 101, 103, 105, 107, 110, 112, 115,
+        117, 119, 122, 124, 126, 128, 130, 132, 133, 135, 136, 138, 139, 140,
+        142, 143, 143, 143, 143, 114, 110, 106, 103, 100, 102, 103, 105, 107,
+        109, 111, 113, 116, 118, 120, 123, 125, 127, 129, 131, 133, 134, 136,
+        138, 139, 140, 142, 143, 144, 144, 144, 144, 116, 112, 109, 105, 102,
+        104, 105, 107, 109, 111, 113, 115, 117, 119, 122, 124, 126, 128, 130,
+        132, 134, 136, 137, 139, 140, 142, 143, 144, 145, 145, 145, 145, 118,
+        114, 111, 108, 104, 106, 107, 109, 110, 112, 114, 116, 119, 121, 123,
+        125, 127, 129, 131, 133, 135, 137, 138, 140, 142, 143, 144, 145, 147,
+        147, 147, 147, 118, 114, 111, 108, 104, 106, 107, 109, 110, 112, 114,
+        116, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 138, 140, 142,
+        143, 144, 145, 147, 147, 147, 147, 118, 114, 111, 108, 104, 106, 107,
+        109, 110, 112, 114, 116, 119, 121, 123, 125, 127, 129, 131, 133, 135,
+        137, 138, 140, 142, 143, 144, 145, 147, 147, 147, 147, 118, 114, 111,
+        108, 104, 106, 107, 109, 110, 112, 114, 116, 119, 121, 123, 125, 127,
+        129, 131, 133, 135, 137, 138, 140, 142, 143, 144, 145, 147, 147, 147,
+        147 },
+      { /* Intra matrices */
+        /* Size 4 */
+        35, 51, 54, 64, 51, 58, 61, 67, 54, 61, 72, 79, 64, 67, 79, 87,
+        /* Size 8 */
+        37, 31, 51, 53, 56, 60, 66, 71, 31, 42, 51, 48, 49, 52, 57, 63, 51, 51,
+        58, 56, 56, 58, 62, 66, 53, 48, 56, 61, 64, 66, 68, 72, 56, 49, 56, 64,
+        69, 72, 75, 77, 60, 52, 58, 66, 72, 76, 80, 83, 66, 57, 62, 68, 75, 80,
+        84, 87, 71, 63, 66, 72, 77, 83, 87, 90,
+        /* Size 16 */
+        37, 33, 31, 38, 50, 51, 52, 53, 55, 57, 59, 62, 64, 67, 70, 70, 33, 34,
+        35, 41, 50, 50, 49, 50, 51, 53, 55, 57, 60, 63, 65, 65, 31, 35, 41, 45,
+        50, 48, 47, 47, 48, 49, 51, 54, 56, 59, 61, 61, 38, 41, 45, 49, 53, 52,
+        51, 51, 51, 53, 54, 56, 58, 61, 63, 63, 50, 50, 50, 53, 57, 56, 55, 55,
+        55, 56, 57, 59, 61, 63, 65, 65, 51, 50, 48, 52, 56, 57, 58, 58, 58, 59,
+        60, 62, 64, 66, 68, 68, 52, 49, 47, 51, 55, 58, 60, 61, 62, 63, 64, 66,
+        67, 69, 70, 70, 53, 50, 47, 51, 55, 58, 61, 63, 65, 66, 67, 69, 70, 71,
+        73, 73, 55, 51, 48, 51, 55, 58, 62, 65, 67, 69, 70, 72, 73, 75, 76, 76,
+        57, 53, 49, 53, 56, 59, 63, 66, 69, 71, 73, 74, 76, 77, 78, 78, 59, 55,
+        51, 54, 57, 60, 64, 67, 70, 73, 75, 77, 78, 80, 81, 81, 62, 57, 54, 56,
+        59, 62, 66, 69, 72, 74, 77, 78, 80, 82, 83, 83, 64, 60, 56, 58, 61, 64,
+        67, 70, 73, 76, 78, 80, 82, 84, 85, 85, 67, 63, 59, 61, 63, 66, 69, 71,
+        75, 77, 80, 82, 84, 85, 87, 87, 70, 65, 61, 63, 65, 68, 70, 73, 76, 78,
+        81, 83, 85, 87, 89, 89, 70, 65, 61, 63, 65, 68, 70, 73, 76, 78, 81, 83,
+        85, 87, 89, 89,
+        /* Size 32 */
+        36, 35, 33, 32, 30, 33, 37, 43, 49, 50, 50, 51, 51, 52, 53, 53, 54, 55,
+        56, 57, 58, 60, 61, 62, 64, 65, 66, 68, 69, 69, 69, 69, 35, 34, 33, 33,
+        32, 35, 39, 44, 49, 49, 50, 50, 50, 50, 51, 52, 52, 53, 54, 55, 56, 58,
+        59, 60, 62, 63, 64, 66, 67, 67, 67, 67, 33, 33, 34, 34, 35, 38, 41, 45,
+        49, 49, 49, 49, 49, 49, 49, 50, 50, 51, 52, 53, 54, 56, 57, 58, 59, 61,
+        62, 63, 65, 65, 65, 65, 32, 33, 34, 36, 38, 40, 43, 46, 50, 49, 48, 48,
+        47, 48, 48, 48, 49, 50, 51, 51, 52, 54, 55, 56, 57, 59, 60, 61, 63, 63,
+        63, 63, 30, 32, 35, 38, 41, 43, 45, 47, 50, 49, 48, 47, 46, 46, 47, 47,
+        47, 48, 49, 50, 51, 52, 53, 54, 55, 57, 58, 59, 61, 61, 61, 61, 33, 35,
+        38, 40, 43, 45, 47, 49, 51, 50, 50, 49, 48, 48, 48, 49, 49, 50, 50, 51,
+        52, 53, 54, 55, 57, 58, 59, 60, 62, 62, 62, 62, 37, 39, 41, 43, 45, 47,
+        49, 51, 53, 52, 51, 51, 50, 50, 50, 50, 51, 51, 52, 53, 54, 54, 56, 57,
+        58, 59, 60, 61, 63, 63, 63, 63, 43, 44, 45, 46, 47, 49, 51, 52, 54, 54,
+        53, 53, 52, 52, 52, 52, 52, 53, 54, 54, 55, 56, 57, 58, 59, 60, 61, 62,
+        63, 63, 63, 63, 49, 49, 49, 50, 50, 51, 53, 54, 56, 56, 55, 55, 55, 55,
+        55, 55, 54, 55, 56, 56, 57, 57, 58, 59, 60, 61, 62, 63, 64, 64, 64, 64,
+        50, 49, 49, 49, 49, 50, 52, 54, 56, 56, 56, 56, 56, 56, 56, 56, 56, 57,
+        57, 58, 58, 59, 60, 61, 62, 63, 64, 65, 66, 66, 66, 66, 50, 50, 49, 48,
+        48, 50, 51, 53, 55, 56, 56, 57, 57, 57, 57, 58, 58, 58, 59, 59, 60, 61,
+        61, 62, 63, 64, 65, 66, 67, 67, 67, 67, 51, 50, 49, 48, 47, 49, 51, 53,
+        55, 56, 57, 57, 58, 59, 59, 59, 60, 60, 61, 61, 62, 62, 63, 64, 65, 65,
+        66, 67, 68, 68, 68, 68, 51, 50, 49, 47, 46, 48, 50, 52, 55, 56, 57, 58,
+        60, 60, 61, 61, 62, 62, 63, 63, 64, 64, 65, 66, 66, 67, 68, 69, 70, 70,
+        70, 70, 52, 50, 49, 48, 46, 48, 50, 52, 55, 56, 57, 59, 60, 61, 62, 62,
+        63, 63, 64, 65, 65, 66, 66, 67, 68, 68, 69, 70, 71, 71, 71, 71, 53, 51,
+        49, 48, 47, 48, 50, 52, 55, 56, 57, 59, 61, 62, 62, 63, 64, 65, 65, 66,
+        67, 67, 68, 69, 69, 70, 71, 71, 72, 72, 72, 72, 53, 52, 50, 48, 47, 49,
+        50, 52, 55, 56, 58, 59, 61, 62, 63, 64, 65, 66, 67, 67, 68, 69, 69, 70,
+        71, 71, 72, 73, 74, 74, 74, 74, 54, 52, 50, 49, 47, 49, 51, 52, 54, 56,
+        58, 60, 62, 63, 64, 65, 67, 67, 68, 69, 70, 70, 71, 72, 72, 73, 74, 74,
+        75, 75, 75, 75, 55, 53, 51, 50, 48, 50, 51, 53, 55, 57, 58, 60, 62, 63,
+        65, 66, 67, 68, 69, 70, 71, 71, 72, 73, 74, 74, 75, 76, 76, 76, 76, 76,
+        56, 54, 52, 51, 49, 50, 52, 54, 56, 57, 59, 61, 63, 64, 65, 67, 68, 69,
+        70, 71, 72, 73, 73, 74, 75, 76, 76, 77, 78, 78, 78, 78, 57, 55, 53, 51,
+        50, 51, 53, 54, 56, 58, 59, 61, 63, 65, 66, 67, 69, 70, 71, 72, 73, 74,
+        75, 75, 76, 77, 78, 78, 79, 79, 79, 79, 58, 56, 54, 52, 51, 52, 54, 55,
+        57, 58, 60, 62, 64, 65, 67, 68, 70, 71, 72, 73, 74, 75, 76, 77, 77, 78,
+        79, 80, 80, 80, 80, 80, 60, 58, 56, 54, 52, 53, 54, 56, 57, 59, 61, 62,
+        64, 66, 67, 69, 70, 71, 73, 74, 75, 76, 77, 78, 78, 79, 80, 81, 81, 81,
+        81, 81, 61, 59, 57, 55, 53, 54, 56, 57, 58, 60, 61, 63, 65, 66, 68, 69,
+        71, 72, 73, 75, 76, 77, 78, 78, 79, 80, 81, 82, 82, 82, 82, 82, 62, 60,
+        58, 56, 54, 55, 57, 58, 59, 61, 62, 64, 66, 67, 69, 70, 72, 73, 74, 75,
+        77, 78, 78, 79, 80, 81, 82, 83, 83, 83, 83, 83, 64, 62, 59, 57, 55, 57,
+        58, 59, 60, 62, 63, 65, 66, 68, 69, 71, 72, 74, 75, 76, 77, 78, 79, 80,
+        81, 82, 83, 84, 84, 84, 84, 84, 65, 63, 61, 59, 57, 58, 59, 60, 61, 63,
+        64, 65, 67, 68, 70, 71, 73, 74, 76, 77, 78, 79, 80, 81, 82, 83, 84, 84,
+        85, 85, 85, 85, 66, 64, 62, 60, 58, 59, 60, 61, 62, 64, 65, 66, 68, 69,
+        71, 72, 74, 75, 76, 78, 79, 80, 81, 82, 83, 84, 84, 85, 86, 86, 86, 86,
+        68, 66, 63, 61, 59, 60, 61, 62, 63, 65, 66, 67, 69, 70, 71, 73, 74, 76,
+        77, 78, 80, 81, 82, 83, 84, 84, 85, 86, 87, 87, 87, 87, 69, 67, 65, 63,
+        61, 62, 63, 63, 64, 66, 67, 68, 70, 71, 72, 74, 75, 76, 78, 79, 80, 81,
+        82, 83, 84, 85, 86, 87, 88, 88, 88, 88, 69, 67, 65, 63, 61, 62, 63, 63,
+        64, 66, 67, 68, 70, 71, 72, 74, 75, 76, 78, 79, 80, 81, 82, 83, 84, 85,
+        86, 87, 88, 88, 88, 88, 69, 67, 65, 63, 61, 62, 63, 63, 64, 66, 67, 68,
+        70, 71, 72, 74, 75, 76, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 88,
+        88, 88, 69, 67, 65, 63, 61, 62, 63, 63, 64, 66, 67, 68, 70, 71, 72, 74,
+        75, 76, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 88, 88, 88 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 68, 96, 122, 68, 90, 110, 128, 96, 110, 130, 142, 122, 128, 142,
+        150,
+        /* Size 8 */
+        64, 52, 55, 67, 81, 95, 107, 116, 52, 58, 56, 63, 74, 86, 98, 108, 55,
+        56, 71, 78, 86, 95, 104, 113, 67, 63, 78, 91, 99, 106, 112, 118, 81, 74,
+        86, 99, 108, 114, 119, 124, 95, 86, 95, 106, 114, 120, 125, 128, 107,
+        98, 104, 112, 119, 125, 129, 132, 116, 108, 113, 118, 124, 128, 132,
+        134,
+        /* Size 16 */
+        64, 57, 52, 53, 55, 61, 67, 74, 81, 88, 95, 100, 107, 111, 116, 116, 57,
+        56, 55, 55, 55, 60, 65, 71, 77, 84, 91, 96, 102, 107, 112, 112, 52, 55,
+        58, 57, 56, 59, 63, 68, 74, 80, 86, 92, 98, 103, 108, 108, 53, 55, 57,
+        59, 62, 66, 70, 74, 80, 85, 91, 96, 101, 106, 110, 110, 55, 55, 56, 62,
+        71, 74, 78, 82, 86, 91, 95, 100, 104, 108, 113, 113, 61, 60, 59, 66, 74,
+        79, 84, 88, 92, 96, 100, 104, 108, 112, 115, 115, 67, 65, 63, 70, 78,
+        84, 91, 94, 99, 102, 106, 109, 112, 115, 118, 118, 74, 71, 68, 74, 82,
+        88, 94, 98, 103, 106, 110, 112, 116, 118, 121, 121, 81, 77, 74, 80, 86,
+        92, 99, 103, 108, 111, 114, 117, 119, 121, 124, 124, 88, 84, 80, 85, 91,
+        96, 102, 106, 111, 114, 117, 119, 122, 124, 126, 126, 95, 91, 86, 91,
+        95, 100, 106, 110, 114, 117, 120, 122, 125, 126, 128, 128, 100, 96, 92,
+        96, 100, 104, 109, 112, 117, 119, 122, 124, 127, 128, 130, 130, 107,
+        102, 98, 101, 104, 108, 112, 116, 119, 122, 125, 127, 129, 130, 132,
+        132, 111, 107, 103, 106, 108, 112, 115, 118, 121, 124, 126, 128, 130,
+        131, 133, 133, 116, 112, 108, 110, 113, 115, 118, 121, 124, 126, 128,
+        130, 132, 133, 134, 134, 116, 112, 108, 110, 113, 115, 118, 121, 124,
+        126, 128, 130, 132, 133, 134, 134,
+        /* Size 32 */
+        64, 60, 57, 55, 52, 53, 53, 54, 55, 58, 61, 64, 67, 70, 74, 77, 81, 84,
+        88, 91, 95, 98, 100, 103, 107, 109, 111, 113, 116, 116, 116, 116, 60,
+        58, 57, 55, 53, 54, 54, 55, 55, 58, 60, 63, 66, 69, 72, 76, 79, 82, 86,
+        89, 93, 95, 98, 101, 104, 107, 109, 111, 114, 114, 114, 114, 57, 57, 56,
+        55, 55, 55, 55, 55, 55, 57, 60, 62, 65, 68, 71, 74, 77, 80, 84, 87, 91,
+        93, 96, 99, 102, 105, 107, 109, 112, 112, 112, 112, 55, 55, 55, 56, 56,
+        56, 56, 56, 56, 57, 59, 61, 64, 66, 69, 72, 76, 79, 82, 85, 88, 91, 94,
+        97, 100, 103, 105, 108, 110, 110, 110, 110, 52, 53, 55, 56, 58, 57, 57,
+        56, 56, 57, 59, 61, 63, 65, 68, 71, 74, 77, 80, 83, 86, 89, 92, 95, 98,
+        101, 103, 106, 108, 108, 108, 108, 53, 54, 55, 56, 57, 58, 58, 58, 59,
+        60, 62, 64, 66, 68, 71, 74, 77, 79, 82, 85, 89, 91, 94, 97, 100, 102,
+        104, 107, 109, 109, 109, 109, 53, 54, 55, 56, 57, 58, 59, 61, 62, 64,
+        66, 68, 70, 72, 74, 77, 80, 82, 85, 88, 91, 93, 96, 98, 101, 103, 106,
+        108, 110, 110, 110, 110, 54, 55, 55, 56, 56, 58, 61, 63, 66, 68, 70, 72,
+        74, 76, 78, 80, 83, 85, 88, 90, 93, 95, 98, 100, 103, 105, 107, 109,
+        112, 112, 112, 112, 55, 55, 55, 56, 56, 59, 62, 66, 71, 73, 74, 76, 78,
+        80, 82, 84, 86, 88, 91, 93, 95, 97, 100, 102, 104, 106, 108, 110, 113,
+        113, 113, 113, 58, 58, 57, 57, 57, 60, 64, 68, 73, 74, 77, 79, 81, 83,
+        85, 87, 89, 91, 93, 95, 98, 100, 102, 104, 106, 108, 110, 112, 114, 114,
+        114, 114, 61, 60, 60, 59, 59, 62, 66, 70, 74, 77, 79, 81, 84, 86, 88,
+        90, 92, 94, 96, 98, 100, 102, 104, 106, 108, 110, 112, 113, 115, 115,
+        115, 115, 64, 63, 62, 61, 61, 64, 68, 72, 76, 79, 81, 84, 87, 89, 91,
+        93, 95, 97, 99, 101, 103, 105, 106, 108, 110, 112, 113, 115, 117, 117,
+        117, 117, 67, 66, 65, 64, 63, 66, 70, 74, 78, 81, 84, 87, 91, 92, 94,
+        96, 99, 100, 102, 104, 106, 107, 109, 110, 112, 114, 115, 117, 118, 118,
+        118, 118, 70, 69, 68, 66, 65, 68, 72, 76, 80, 83, 86, 89, 92, 94, 96,
+        99, 101, 102, 104, 106, 108, 109, 111, 112, 114, 115, 117, 118, 119,
+        119, 119, 119, 74, 72, 71, 69, 68, 71, 74, 78, 82, 85, 88, 91, 94, 96,
+        98, 101, 103, 105, 106, 108, 110, 111, 112, 114, 116, 117, 118, 119,
+        121, 121, 121, 121, 77, 76, 74, 72, 71, 74, 77, 80, 84, 87, 90, 93, 96,
+        99, 101, 103, 105, 107, 108, 110, 112, 113, 114, 116, 117, 118, 120,
+        121, 122, 122, 122, 122, 81, 79, 77, 76, 74, 77, 80, 83, 86, 89, 92, 95,
+        99, 101, 103, 105, 108, 109, 111, 112, 114, 115, 117, 118, 119, 120,
+        121, 122, 124, 124, 124, 124, 84, 82, 80, 79, 77, 79, 82, 85, 88, 91,
+        94, 97, 100, 102, 105, 107, 109, 111, 112, 114, 115, 117, 118, 119, 120,
+        121, 123, 124, 125, 125, 125, 125, 88, 86, 84, 82, 80, 82, 85, 88, 91,
+        93, 96, 99, 102, 104, 106, 108, 111, 112, 114, 115, 117, 118, 119, 121,
+        122, 123, 124, 125, 126, 126, 126, 126, 91, 89, 87, 85, 83, 85, 88, 90,
+        93, 95, 98, 101, 104, 106, 108, 110, 112, 114, 115, 117, 119, 120, 121,
+        122, 123, 124, 125, 126, 127, 127, 127, 127, 95, 93, 91, 88, 86, 89, 91,
+        93, 95, 98, 100, 103, 106, 108, 110, 112, 114, 115, 117, 119, 120, 121,
+        122, 123, 125, 125, 126, 127, 128, 128, 128, 128, 98, 95, 93, 91, 89,
+        91, 93, 95, 97, 100, 102, 105, 107, 109, 111, 113, 115, 117, 118, 120,
+        121, 122, 123, 124, 126, 126, 127, 128, 129, 129, 129, 129, 100, 98, 96,
+        94, 92, 94, 96, 98, 100, 102, 104, 106, 109, 111, 112, 114, 117, 118,
+        119, 121, 122, 123, 124, 125, 127, 127, 128, 129, 130, 130, 130, 130,
+        103, 101, 99, 97, 95, 97, 98, 100, 102, 104, 106, 108, 110, 112, 114,
+        116, 118, 119, 121, 122, 123, 124, 125, 127, 128, 128, 129, 130, 131,
+        131, 131, 131, 107, 104, 102, 100, 98, 100, 101, 103, 104, 106, 108,
+        110, 112, 114, 116, 117, 119, 120, 122, 123, 125, 126, 127, 128, 129,
+        129, 130, 131, 132, 132, 132, 132, 109, 107, 105, 103, 101, 102, 103,
+        105, 106, 108, 110, 112, 114, 115, 117, 118, 120, 121, 123, 124, 125,
+        126, 127, 128, 129, 130, 131, 131, 132, 132, 132, 132, 111, 109, 107,
+        105, 103, 104, 106, 107, 108, 110, 112, 113, 115, 117, 118, 120, 121,
+        123, 124, 125, 126, 127, 128, 129, 130, 131, 131, 132, 133, 133, 133,
+        133, 113, 111, 109, 108, 106, 107, 108, 109, 110, 112, 113, 115, 117,
+        118, 119, 121, 122, 124, 125, 126, 127, 128, 129, 130, 131, 131, 132,
+        133, 134, 134, 134, 134, 116, 114, 112, 110, 108, 109, 110, 112, 113,
+        114, 115, 117, 118, 119, 121, 122, 124, 125, 126, 127, 128, 129, 130,
+        131, 132, 132, 133, 134, 134, 134, 134, 134, 116, 114, 112, 110, 108,
+        109, 110, 112, 113, 114, 115, 117, 118, 119, 121, 122, 124, 125, 126,
+        127, 128, 129, 130, 131, 132, 132, 133, 134, 134, 134, 134, 134, 116,
+        114, 112, 110, 108, 109, 110, 112, 113, 114, 115, 117, 118, 119, 121,
+        122, 124, 125, 126, 127, 128, 129, 130, 131, 132, 132, 133, 134, 134,
+        134, 134, 134, 116, 114, 112, 110, 108, 109, 110, 112, 113, 114, 115,
+        117, 118, 119, 121, 122, 124, 125, 126, 127, 128, 129, 130, 131, 132,
+        132, 133, 134, 134, 134, 134, 134 },
+      { /* Intra matrices */
+        /* Size 4 */
+        35, 37, 53, 69, 37, 50, 61, 72, 53, 61, 74, 81, 69, 72, 81, 86,
+        /* Size 8 */
+        40, 32, 34, 42, 51, 60, 68, 75, 32, 36, 34, 39, 46, 55, 63, 70, 34, 34,
+        44, 49, 54, 61, 67, 73, 42, 39, 49, 57, 63, 68, 72, 76, 51, 46, 54, 63,
+        69, 74, 77, 80, 60, 55, 61, 68, 74, 78, 81, 84, 68, 63, 67, 72, 77, 81,
+        84, 86, 75, 70, 73, 76, 80, 84, 86, 88,
+        /* Size 16 */
+        39, 35, 31, 32, 33, 37, 41, 45, 50, 54, 59, 63, 67, 70, 73, 73, 35, 34,
+        33, 33, 33, 36, 39, 43, 47, 51, 56, 60, 64, 67, 70, 70, 31, 33, 35, 34,
+        34, 36, 38, 41, 45, 49, 53, 57, 61, 64, 68, 68, 32, 33, 34, 36, 38, 40,
+        42, 45, 49, 52, 56, 59, 63, 66, 69, 69, 33, 33, 34, 38, 43, 45, 48, 50,
+        53, 56, 59, 62, 65, 68, 71, 71, 37, 36, 36, 40, 45, 48, 52, 54, 57, 60,
+        62, 65, 68, 70, 73, 73, 41, 39, 38, 42, 48, 52, 56, 59, 61, 64, 66, 68,
+        71, 73, 75, 75, 45, 43, 41, 45, 50, 54, 59, 61, 64, 66, 69, 71, 73, 75,
+        77, 77, 50, 47, 45, 49, 53, 57, 61, 64, 67, 70, 72, 74, 75, 77, 78, 78,
+        54, 51, 49, 52, 56, 60, 64, 66, 70, 72, 74, 76, 77, 79, 80, 80, 59, 56,
+        53, 56, 59, 62, 66, 69, 72, 74, 76, 78, 79, 80, 82, 82, 63, 60, 57, 59,
+        62, 65, 68, 71, 74, 76, 78, 79, 81, 82, 83, 83, 67, 64, 61, 63, 65, 68,
+        71, 73, 75, 77, 79, 81, 82, 83, 84, 84, 70, 67, 64, 66, 68, 70, 73, 75,
+        77, 79, 80, 82, 83, 84, 85, 85, 73, 70, 68, 69, 71, 73, 75, 77, 78, 80,
+        82, 83, 84, 85, 86, 86, 73, 70, 68, 69, 71, 73, 75, 77, 78, 80, 82, 83,
+        84, 85, 86, 86,
+        /* Size 32 */
+        38, 36, 34, 32, 31, 31, 32, 32, 33, 34, 36, 38, 40, 42, 44, 47, 49, 51,
+        53, 56, 58, 60, 62, 64, 66, 67, 69, 71, 72, 72, 72, 72, 36, 35, 34, 33,
+        32, 32, 32, 33, 33, 34, 36, 38, 40, 41, 43, 46, 48, 50, 52, 54, 57, 59,
+        60, 62, 64, 66, 68, 69, 71, 71, 71, 71, 34, 34, 33, 33, 32, 33, 33, 33,
+        33, 34, 36, 37, 39, 41, 42, 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65,
+        66, 68, 70, 70, 70, 70, 32, 33, 33, 33, 33, 33, 33, 33, 33, 34, 35, 37,
+        38, 40, 42, 44, 46, 48, 50, 52, 54, 56, 58, 60, 62, 63, 65, 67, 68, 68,
+        68, 68, 31, 32, 32, 33, 34, 34, 34, 33, 33, 34, 35, 36, 37, 39, 41, 43,
+        45, 46, 48, 50, 53, 54, 56, 58, 60, 62, 64, 65, 67, 67, 67, 67, 31, 32,
+        33, 33, 34, 34, 35, 35, 35, 36, 37, 38, 40, 41, 43, 44, 46, 48, 50, 52,
+        54, 56, 57, 59, 61, 63, 64, 66, 68, 68, 68, 68, 32, 32, 33, 33, 34, 35,
+        35, 36, 37, 38, 39, 41, 42, 43, 45, 46, 48, 50, 52, 53, 55, 57, 59, 61,
+        62, 64, 65, 67, 69, 69, 69, 69, 32, 33, 33, 33, 33, 35, 36, 38, 40, 41,
+        42, 43, 44, 46, 47, 49, 50, 52, 53, 55, 57, 58, 60, 62, 63, 65, 66, 68,
+        69, 69, 69, 69, 33, 33, 33, 33, 33, 35, 37, 40, 43, 44, 45, 46, 47, 49,
+        50, 51, 53, 54, 55, 57, 58, 60, 61, 63, 65, 66, 67, 69, 70, 70, 70, 70,
+        34, 34, 34, 34, 34, 36, 38, 41, 44, 45, 46, 48, 49, 50, 52, 53, 54, 56,
+        57, 59, 60, 61, 63, 64, 66, 67, 68, 70, 71, 71, 71, 71, 36, 36, 36, 35,
+        35, 37, 39, 42, 45, 46, 48, 49, 51, 52, 54, 55, 56, 58, 59, 60, 62, 63,
+        64, 66, 67, 68, 69, 71, 72, 72, 72, 72, 38, 38, 37, 37, 36, 38, 41, 43,
+        46, 48, 49, 51, 53, 54, 56, 57, 58, 60, 61, 62, 63, 65, 66, 67, 68, 69,
+        71, 72, 73, 73, 73, 73, 40, 40, 39, 38, 37, 40, 42, 44, 47, 49, 51, 53,
+        55, 57, 58, 59, 61, 62, 63, 64, 65, 66, 67, 69, 70, 71, 72, 73, 74, 74,
+        74, 74, 42, 41, 41, 40, 39, 41, 43, 46, 49, 50, 52, 54, 57, 58, 59, 61,
+        62, 63, 64, 65, 67, 68, 69, 70, 71, 72, 73, 74, 75, 75, 75, 75, 44, 43,
+        42, 42, 41, 43, 45, 47, 50, 52, 54, 56, 58, 59, 61, 62, 64, 65, 66, 67,
+        68, 69, 70, 71, 72, 73, 74, 75, 76, 76, 76, 76, 47, 46, 45, 44, 43, 44,
+        46, 49, 51, 53, 55, 57, 59, 61, 62, 64, 65, 66, 67, 68, 69, 70, 71, 72,
+        73, 74, 75, 76, 77, 77, 77, 77, 49, 48, 47, 46, 45, 46, 48, 50, 53, 54,
+        56, 58, 61, 62, 64, 65, 67, 68, 69, 70, 71, 72, 73, 74, 74, 75, 76, 77,
+        78, 78, 78, 78, 51, 50, 49, 48, 46, 48, 50, 52, 54, 56, 58, 60, 62, 63,
+        65, 66, 68, 69, 70, 71, 72, 73, 74, 75, 75, 76, 77, 78, 78, 78, 78, 78,
+        53, 52, 51, 50, 48, 50, 52, 53, 55, 57, 59, 61, 63, 64, 66, 67, 69, 70,
+        71, 72, 73, 74, 75, 75, 76, 77, 78, 78, 79, 79, 79, 79, 56, 54, 53, 52,
+        50, 52, 53, 55, 57, 59, 60, 62, 64, 65, 67, 68, 70, 71, 72, 73, 74, 75,
+        76, 76, 77, 78, 79, 79, 80, 80, 80, 80, 58, 57, 55, 54, 53, 54, 55, 57,
+        58, 60, 62, 63, 65, 67, 68, 69, 71, 72, 73, 74, 75, 76, 77, 77, 78, 79,
+        79, 80, 81, 81, 81, 81, 60, 59, 57, 56, 54, 56, 57, 58, 60, 61, 63, 65,
+        66, 68, 69, 70, 72, 73, 74, 75, 76, 77, 77, 78, 79, 79, 80, 81, 81, 81,
+        81, 81, 62, 60, 59, 58, 56, 57, 59, 60, 61, 63, 64, 66, 67, 69, 70, 71,
+        73, 74, 75, 76, 77, 77, 78, 79, 80, 80, 81, 81, 82, 82, 82, 82, 64, 62,
+        61, 60, 58, 59, 61, 62, 63, 64, 66, 67, 69, 70, 71, 72, 74, 75, 75, 76,
+        77, 78, 79, 80, 80, 81, 81, 82, 82, 82, 82, 82, 66, 64, 63, 62, 60, 61,
+        62, 63, 65, 66, 67, 68, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 80,
+        81, 82, 82, 83, 83, 83, 83, 83, 67, 66, 65, 63, 62, 63, 64, 65, 66, 67,
+        68, 69, 71, 72, 73, 74, 75, 76, 77, 78, 79, 79, 80, 81, 82, 82, 83, 83,
+        84, 84, 84, 84, 69, 68, 66, 65, 64, 64, 65, 66, 67, 68, 69, 71, 72, 73,
+        74, 75, 76, 77, 78, 79, 79, 80, 81, 81, 82, 83, 83, 84, 84, 84, 84, 84,
+        71, 69, 68, 67, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78,
+        78, 79, 80, 81, 81, 82, 83, 83, 84, 84, 84, 84, 84, 84, 72, 71, 70, 68,
+        67, 68, 69, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 78, 79, 80, 81, 81,
+        82, 82, 83, 84, 84, 84, 85, 85, 85, 85, 72, 71, 70, 68, 67, 68, 69, 69,
+        70, 71, 72, 73, 74, 75, 76, 77, 78, 78, 79, 80, 81, 81, 82, 82, 83, 84,
+        84, 84, 85, 85, 85, 85, 72, 71, 70, 68, 67, 68, 69, 69, 70, 71, 72, 73,
+        74, 75, 76, 77, 78, 78, 79, 80, 81, 81, 82, 82, 83, 84, 84, 84, 85, 85,
+        85, 85, 72, 71, 70, 68, 67, 68, 69, 69, 70, 71, 72, 73, 74, 75, 76, 77,
+        78, 78, 79, 80, 81, 81, 82, 82, 83, 84, 84, 84, 85, 85, 85, 85 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 88, 93, 107, 88, 99, 103, 112, 93, 103, 119, 127, 107, 112, 127,
+        137,
+        /* Size 8 */
+        64, 54, 83, 87, 90, 97, 104, 111, 54, 71, 84, 79, 81, 86, 92, 100, 83,
+        84, 93, 91, 91, 94, 99, 104, 87, 79, 91, 98, 101, 103, 107, 111, 90, 81,
+        91, 101, 107, 111, 115, 118, 97, 86, 94, 103, 111, 117, 121, 124, 104,
+        92, 99, 107, 115, 121, 126, 129, 111, 100, 104, 111, 118, 124, 129, 133,
+        /* Size 16 */
+        64, 59, 54, 66, 83, 85, 87, 88, 90, 93, 97, 100, 104, 107, 111, 111, 59,
+        60, 62, 71, 84, 83, 83, 84, 85, 88, 91, 94, 98, 101, 105, 105, 54, 62,
+        71, 77, 84, 82, 79, 80, 81, 83, 86, 89, 92, 96, 100, 100, 66, 71, 77,
+        83, 88, 87, 85, 85, 85, 87, 90, 92, 95, 99, 102, 102, 83, 84, 84, 88,
+        93, 92, 91, 91, 91, 92, 94, 96, 99, 101, 104, 104, 85, 83, 82, 87, 92,
+        93, 95, 95, 96, 97, 98, 100, 103, 105, 108, 108, 87, 83, 79, 85, 91, 95,
+        98, 100, 101, 102, 103, 105, 107, 109, 111, 111, 88, 84, 80, 85, 91, 95,
+        100, 102, 104, 106, 107, 109, 111, 113, 115, 115, 90, 85, 81, 85, 91,
+        96, 101, 104, 107, 109, 111, 113, 115, 116, 118, 118, 93, 88, 83, 87,
+        92, 97, 102, 106, 109, 112, 114, 116, 118, 120, 121, 121, 97, 91, 86,
+        90, 94, 98, 103, 107, 111, 114, 117, 119, 121, 123, 124, 124, 100, 94,
+        89, 92, 96, 100, 105, 109, 113, 116, 119, 121, 123, 125, 127, 127, 104,
+        98, 92, 95, 99, 103, 107, 111, 115, 118, 121, 123, 126, 128, 129, 129,
+        107, 101, 96, 99, 101, 105, 109, 113, 116, 120, 123, 125, 128, 129, 131,
+        131, 111, 105, 100, 102, 104, 108, 111, 115, 118, 121, 124, 127, 129,
+        131, 133, 133, 111, 105, 100, 102, 104, 108, 111, 115, 118, 121, 124,
+        127, 129, 131, 133, 133,
+        /* Size 32 */
+        64, 61, 59, 57, 54, 60, 66, 74, 83, 84, 85, 86, 87, 87, 88, 89, 90, 92,
+        93, 95, 97, 98, 100, 102, 104, 105, 107, 109, 111, 111, 111, 111, 61,
+        60, 60, 59, 58, 63, 68, 75, 84, 84, 84, 84, 85, 85, 86, 87, 88, 89, 90,
+        92, 94, 95, 97, 99, 101, 102, 104, 106, 108, 108, 108, 108, 59, 60, 60,
+        61, 62, 66, 71, 77, 84, 84, 83, 83, 83, 83, 84, 84, 85, 86, 88, 89, 91,
+        92, 94, 96, 98, 99, 101, 103, 105, 105, 105, 105, 57, 59, 61, 64, 66,
+        70, 74, 79, 84, 83, 82, 82, 81, 81, 82, 82, 83, 84, 85, 87, 88, 90, 91,
+        93, 95, 97, 98, 100, 102, 102, 102, 102, 54, 58, 62, 66, 71, 74, 77, 81,
+        84, 83, 82, 80, 79, 79, 80, 80, 81, 82, 83, 84, 86, 87, 89, 91, 92, 94,
+        96, 98, 100, 100, 100, 100, 60, 63, 66, 70, 74, 77, 80, 83, 86, 85, 84,
+        83, 82, 82, 82, 83, 83, 84, 85, 86, 88, 89, 91, 92, 94, 95, 97, 99, 101,
+        101, 101, 101, 66, 68, 71, 74, 77, 80, 83, 85, 88, 88, 87, 86, 85, 85,
+        85, 85, 85, 86, 87, 89, 90, 91, 92, 94, 95, 97, 99, 100, 102, 102, 102,
+        102, 74, 75, 77, 79, 81, 83, 85, 88, 91, 90, 89, 89, 88, 88, 88, 88, 88,
+        89, 90, 91, 92, 93, 94, 96, 97, 98, 100, 102, 103, 103, 103, 103, 83,
+        84, 84, 84, 84, 86, 88, 91, 93, 93, 92, 92, 91, 91, 91, 91, 91, 92, 92,
+        93, 94, 95, 96, 97, 99, 100, 101, 103, 104, 104, 104, 104, 84, 84, 84,
+        83, 83, 85, 88, 90, 93, 93, 93, 93, 93, 93, 93, 93, 93, 94, 95, 95, 96,
+        97, 98, 99, 101, 102, 103, 105, 106, 106, 106, 106, 85, 84, 83, 82, 82,
+        84, 87, 89, 92, 93, 93, 94, 95, 95, 95, 95, 96, 96, 97, 98, 98, 99, 100,
+        102, 103, 104, 105, 106, 108, 108, 108, 108, 86, 84, 83, 82, 80, 83, 86,
+        89, 92, 93, 94, 95, 96, 97, 97, 98, 98, 99, 100, 100, 101, 102, 103,
+        104, 105, 106, 107, 108, 109, 109, 109, 109, 87, 85, 83, 81, 79, 82, 85,
+        88, 91, 93, 95, 96, 98, 99, 100, 100, 101, 102, 102, 103, 103, 104, 105,
+        106, 107, 108, 109, 110, 111, 111, 111, 111, 87, 85, 83, 81, 79, 82, 85,
+        88, 91, 93, 95, 97, 99, 100, 101, 102, 102, 103, 104, 105, 105, 106,
+        107, 108, 109, 110, 111, 112, 113, 113, 113, 113, 88, 86, 84, 82, 80,
+        82, 85, 88, 91, 93, 95, 97, 100, 101, 102, 103, 104, 105, 106, 106, 107,
+        108, 109, 110, 111, 112, 113, 114, 115, 115, 115, 115, 89, 87, 84, 82,
+        80, 83, 85, 88, 91, 93, 95, 98, 100, 102, 103, 104, 106, 107, 107, 108,
+        109, 110, 111, 112, 113, 114, 115, 115, 116, 116, 116, 116, 90, 88, 85,
+        83, 81, 83, 85, 88, 91, 93, 96, 98, 101, 102, 104, 106, 107, 108, 109,
+        110, 111, 112, 113, 114, 115, 116, 116, 117, 118, 118, 118, 118, 92, 89,
+        86, 84, 82, 84, 86, 89, 92, 94, 96, 99, 102, 103, 105, 107, 108, 109,
+        111, 112, 113, 114, 114, 115, 116, 117, 118, 119, 120, 120, 120, 120,
+        93, 90, 88, 85, 83, 85, 87, 90, 92, 95, 97, 100, 102, 104, 106, 107,
+        109, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 120, 121, 121,
+        121, 121, 95, 92, 89, 87, 84, 86, 89, 91, 93, 95, 98, 100, 103, 105,
+        106, 108, 110, 112, 113, 114, 116, 116, 117, 118, 119, 120, 121, 122,
+        123, 123, 123, 123, 97, 94, 91, 88, 86, 88, 90, 92, 94, 96, 98, 101,
+        103, 105, 107, 109, 111, 113, 114, 116, 117, 118, 119, 120, 121, 122,
+        123, 124, 124, 124, 124, 124, 98, 95, 92, 90, 87, 89, 91, 93, 95, 97,
+        99, 102, 104, 106, 108, 110, 112, 114, 115, 116, 118, 119, 120, 121,
+        122, 123, 124, 125, 126, 126, 126, 126, 100, 97, 94, 91, 89, 91, 92, 94,
+        96, 98, 100, 103, 105, 107, 109, 111, 113, 114, 116, 117, 119, 120, 121,
+        122, 123, 124, 125, 126, 127, 127, 127, 127, 102, 99, 96, 93, 91, 92,
+        94, 96, 97, 99, 102, 104, 106, 108, 110, 112, 114, 115, 117, 118, 120,
+        121, 122, 123, 125, 125, 126, 127, 128, 128, 128, 128, 104, 101, 98, 95,
+        92, 94, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115, 116, 118,
+        119, 121, 122, 123, 125, 126, 127, 128, 128, 129, 129, 129, 129, 105,
+        102, 99, 97, 94, 95, 97, 98, 100, 102, 104, 106, 108, 110, 112, 114,
+        116, 117, 119, 120, 122, 123, 124, 125, 127, 128, 128, 129, 130, 130,
+        130, 130, 107, 104, 101, 98, 96, 97, 99, 100, 101, 103, 105, 107, 109,
+        111, 113, 115, 116, 118, 120, 121, 123, 124, 125, 126, 128, 128, 129,
+        130, 131, 131, 131, 131, 109, 106, 103, 100, 98, 99, 100, 102, 103, 105,
+        106, 108, 110, 112, 114, 115, 117, 119, 120, 122, 124, 125, 126, 127,
+        128, 129, 130, 131, 132, 132, 132, 132, 111, 108, 105, 102, 100, 101,
+        102, 103, 104, 106, 108, 109, 111, 113, 115, 116, 118, 120, 121, 123,
+        124, 126, 127, 128, 129, 130, 131, 132, 133, 133, 133, 133, 111, 108,
+        105, 102, 100, 101, 102, 103, 104, 106, 108, 109, 111, 113, 115, 116,
+        118, 120, 121, 123, 124, 126, 127, 128, 129, 130, 131, 132, 133, 133,
+        133, 133, 111, 108, 105, 102, 100, 101, 102, 103, 104, 106, 108, 109,
+        111, 113, 115, 116, 118, 120, 121, 123, 124, 126, 127, 128, 129, 130,
+        131, 132, 133, 133, 133, 133, 111, 108, 105, 102, 100, 101, 102, 103,
+        104, 106, 108, 109, 111, 113, 115, 116, 118, 120, 121, 123, 124, 126,
+        127, 128, 129, 130, 131, 132, 133, 133, 133, 133 },
+      { /* Intra matrices */
+        /* Size 4 */
+        37, 52, 55, 64, 52, 59, 62, 67, 55, 62, 72, 77, 64, 67, 77, 84,
+        /* Size 8 */
+        40, 33, 52, 54, 57, 61, 66, 71, 33, 44, 53, 49, 50, 54, 58, 63, 52, 53,
+        59, 57, 57, 59, 63, 66, 54, 49, 57, 62, 64, 66, 68, 71, 57, 50, 57, 64,
+        68, 71, 74, 76, 61, 54, 59, 66, 71, 75, 78, 80, 66, 58, 63, 68, 74, 78,
+        81, 84, 71, 63, 66, 71, 76, 80, 84, 86,
+        /* Size 16 */
+        39, 36, 33, 40, 51, 52, 53, 55, 56, 58, 60, 62, 65, 67, 70, 70, 36, 37,
+        38, 43, 52, 51, 51, 52, 53, 54, 56, 58, 61, 63, 66, 66, 33, 38, 44, 47,
+        52, 50, 49, 49, 50, 51, 53, 55, 57, 60, 62, 62, 40, 43, 47, 51, 55, 53,
+        52, 52, 53, 54, 55, 57, 59, 61, 64, 64, 51, 52, 52, 55, 58, 57, 56, 56,
+        56, 57, 58, 60, 61, 63, 65, 65, 52, 51, 50, 53, 57, 58, 59, 59, 59, 60,
+        61, 63, 64, 66, 68, 68, 53, 51, 49, 52, 56, 59, 61, 62, 63, 64, 65, 66,
+        67, 68, 70, 70, 55, 52, 49, 52, 56, 59, 62, 64, 65, 66, 67, 68, 70, 71,
+        72, 72, 56, 53, 50, 53, 56, 59, 63, 65, 67, 69, 70, 71, 72, 73, 75, 75,
+        58, 54, 51, 54, 57, 60, 64, 66, 69, 70, 72, 73, 74, 76, 77, 77, 60, 56,
+        53, 55, 58, 61, 65, 67, 70, 72, 74, 75, 77, 78, 79, 79, 62, 58, 55, 57,
+        60, 63, 66, 68, 71, 73, 75, 77, 78, 79, 81, 81, 65, 61, 57, 59, 61, 64,
+        67, 70, 72, 74, 77, 78, 80, 81, 82, 82, 67, 63, 60, 61, 63, 66, 68, 71,
+        73, 76, 78, 79, 81, 82, 84, 84, 70, 66, 62, 64, 65, 68, 70, 72, 75, 77,
+        79, 81, 82, 84, 85, 85, 70, 66, 62, 64, 65, 68, 70, 72, 75, 77, 79, 81,
+        82, 84, 85, 85,
+        /* Size 32 */
+        39, 37, 35, 34, 33, 36, 40, 45, 51, 51, 52, 52, 53, 54, 54, 55, 55, 56,
+        57, 58, 59, 61, 62, 63, 64, 65, 67, 68, 69, 69, 69, 69, 37, 36, 36, 35,
+        35, 38, 41, 46, 51, 51, 51, 52, 52, 52, 53, 53, 54, 55, 56, 57, 58, 59,
+        60, 61, 62, 63, 64, 66, 67, 67, 67, 67, 35, 36, 36, 37, 37, 40, 43, 47,
+        51, 51, 51, 51, 50, 51, 51, 52, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61,
+        63, 64, 65, 65, 65, 65, 34, 35, 37, 38, 40, 42, 45, 48, 51, 51, 50, 50,
+        49, 50, 50, 50, 51, 51, 52, 53, 54, 55, 56, 57, 58, 60, 61, 62, 63, 63,
+        63, 63, 33, 35, 37, 40, 43, 45, 47, 49, 52, 51, 50, 49, 48, 48, 49, 49,
+        49, 50, 51, 52, 52, 53, 55, 56, 57, 58, 59, 60, 62, 62, 62, 62, 36, 38,
+        40, 42, 45, 47, 49, 51, 53, 52, 51, 51, 50, 50, 50, 50, 51, 51, 52, 53,
+        54, 55, 56, 57, 58, 59, 60, 61, 62, 62, 62, 62, 40, 41, 43, 45, 47, 49,
+        50, 52, 54, 54, 53, 52, 52, 52, 52, 52, 52, 53, 54, 54, 55, 56, 57, 58,
+        59, 60, 61, 62, 63, 63, 63, 63, 45, 46, 47, 48, 49, 51, 52, 54, 56, 55,
+        55, 54, 54, 54, 54, 54, 54, 55, 55, 56, 56, 57, 58, 59, 60, 61, 62, 63,
+        64, 64, 64, 64, 51, 51, 51, 51, 52, 53, 54, 56, 57, 57, 57, 56, 56, 56,
+        56, 56, 56, 56, 57, 57, 58, 59, 59, 60, 61, 62, 63, 64, 65, 65, 65, 65,
+        51, 51, 51, 51, 51, 52, 54, 55, 57, 57, 57, 57, 57, 57, 57, 57, 57, 58,
+        58, 59, 59, 60, 61, 61, 62, 63, 64, 65, 66, 66, 66, 66, 52, 51, 51, 50,
+        50, 51, 53, 55, 57, 57, 57, 58, 58, 58, 59, 59, 59, 59, 60, 60, 61, 61,
+        62, 63, 64, 64, 65, 66, 67, 67, 67, 67, 52, 52, 51, 50, 49, 51, 52, 54,
+        56, 57, 58, 59, 59, 60, 60, 60, 61, 61, 61, 62, 62, 63, 64, 64, 65, 66,
+        66, 67, 68, 68, 68, 68, 53, 52, 50, 49, 48, 50, 52, 54, 56, 57, 58, 59,
+        61, 61, 61, 62, 62, 63, 63, 64, 64, 65, 65, 66, 66, 67, 68, 69, 69, 69,
+        69, 69, 54, 52, 51, 50, 48, 50, 52, 54, 56, 57, 58, 60, 61, 62, 62, 63,
+        63, 64, 64, 65, 65, 66, 66, 67, 68, 68, 69, 70, 70, 70, 70, 70, 54, 53,
+        51, 50, 49, 50, 52, 54, 56, 57, 59, 60, 61, 62, 63, 64, 64, 65, 66, 66,
+        67, 67, 68, 68, 69, 70, 70, 71, 72, 72, 72, 72, 55, 53, 52, 50, 49, 50,
+        52, 54, 56, 57, 59, 60, 62, 63, 64, 65, 66, 66, 67, 67, 68, 69, 69, 70,
+        70, 71, 71, 72, 73, 73, 73, 73, 55, 54, 52, 51, 49, 51, 52, 54, 56, 57,
+        59, 61, 62, 63, 64, 66, 67, 67, 68, 69, 69, 70, 70, 71, 72, 72, 73, 73,
+        74, 74, 74, 74, 56, 55, 53, 51, 50, 51, 53, 55, 56, 58, 59, 61, 63, 64,
+        65, 66, 67, 68, 69, 70, 70, 71, 71, 72, 73, 73, 74, 74, 75, 75, 75, 75,
+        57, 56, 54, 52, 51, 52, 54, 55, 57, 58, 60, 61, 63, 64, 66, 67, 68, 69,
+        70, 70, 71, 72, 72, 73, 74, 74, 75, 75, 76, 76, 76, 76, 58, 57, 55, 53,
+        52, 53, 54, 56, 57, 59, 60, 62, 64, 65, 66, 67, 69, 70, 70, 71, 72, 73,
+        73, 74, 75, 75, 76, 77, 77, 77, 77, 77, 59, 58, 56, 54, 52, 54, 55, 56,
+        58, 59, 61, 62, 64, 65, 67, 68, 69, 70, 71, 72, 73, 74, 75, 75, 76, 76,
+        77, 78, 78, 78, 78, 78, 61, 59, 57, 55, 53, 55, 56, 57, 59, 60, 61, 63,
+        65, 66, 67, 69, 70, 71, 72, 73, 74, 75, 75, 76, 77, 77, 78, 78, 79, 79,
+        79, 79, 62, 60, 58, 56, 55, 56, 57, 58, 59, 61, 62, 64, 65, 66, 68, 69,
+        70, 71, 72, 73, 75, 75, 76, 77, 77, 78, 79, 79, 80, 80, 80, 80, 63, 61,
+        59, 57, 56, 57, 58, 59, 60, 61, 63, 64, 66, 67, 68, 70, 71, 72, 73, 74,
+        75, 76, 77, 77, 78, 79, 79, 80, 81, 81, 81, 81, 64, 62, 60, 58, 57, 58,
+        59, 60, 61, 62, 64, 65, 66, 68, 69, 70, 72, 73, 74, 75, 76, 77, 77, 78,
+        79, 80, 80, 81, 82, 82, 82, 82, 65, 63, 61, 60, 58, 59, 60, 61, 62, 63,
+        64, 66, 67, 68, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 80, 81, 82,
+        82, 82, 82, 82, 67, 64, 63, 61, 59, 60, 61, 62, 63, 64, 65, 66, 68, 69,
+        70, 71, 73, 74, 75, 76, 77, 78, 79, 79, 80, 81, 82, 82, 83, 83, 83, 83,
+        68, 66, 64, 62, 60, 61, 62, 63, 64, 65, 66, 67, 69, 70, 71, 72, 73, 74,
+        75, 77, 78, 78, 79, 80, 81, 82, 82, 83, 84, 84, 84, 84, 69, 67, 65, 63,
+        62, 62, 63, 64, 65, 66, 67, 68, 69, 70, 72, 73, 74, 75, 76, 77, 78, 79,
+        80, 81, 82, 82, 83, 84, 84, 84, 84, 84, 69, 67, 65, 63, 62, 62, 63, 64,
+        65, 66, 67, 68, 69, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 82,
+        83, 84, 84, 84, 84, 84, 69, 67, 65, 63, 62, 62, 63, 64, 65, 66, 67, 68,
+        69, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 82, 83, 84, 84, 84,
+        84, 84, 69, 67, 65, 63, 62, 62, 63, 64, 65, 66, 67, 68, 69, 70, 72, 73,
+        74, 75, 76, 77, 78, 79, 80, 81, 82, 82, 83, 84, 84, 84, 84, 84 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 68, 92, 113, 68, 87, 103, 117, 92, 103, 119, 128, 113, 117, 128,
+        134,
+        /* Size 8 */
+        64, 53, 56, 67, 79, 91, 100, 107, 53, 58, 56, 63, 73, 84, 93, 101, 56,
+        56, 70, 77, 83, 91, 98, 105, 67, 63, 77, 87, 94, 99, 104, 109, 79, 73,
+        83, 94, 101, 106, 110, 113, 91, 84, 91, 99, 106, 110, 114, 116, 100, 93,
+        98, 104, 110, 114, 117, 119, 107, 101, 105, 109, 113, 116, 119, 121,
+        /* Size 16 */
+        64, 58, 53, 54, 56, 61, 67, 72, 79, 84, 91, 95, 100, 103, 107, 107, 58,
+        57, 55, 56, 56, 60, 65, 70, 76, 81, 87, 91, 96, 100, 104, 104, 53, 55,
+        58, 57, 56, 59, 63, 67, 73, 78, 84, 88, 93, 97, 101, 101, 54, 56, 57,
+        60, 63, 66, 69, 73, 78, 82, 87, 91, 96, 99, 103, 103, 56, 56, 56, 63,
+        70, 73, 77, 80, 83, 87, 91, 94, 98, 101, 105, 105, 61, 60, 59, 66, 73,
+        77, 81, 85, 88, 91, 95, 98, 101, 104, 107, 107, 67, 65, 63, 69, 77, 81,
+        87, 90, 94, 96, 99, 102, 104, 106, 109, 109, 72, 70, 67, 73, 80, 85, 90,
+        93, 97, 100, 102, 104, 107, 109, 111, 111, 79, 76, 73, 78, 83, 88, 94,
+        97, 101, 103, 106, 108, 110, 111, 113, 113, 84, 81, 78, 82, 87, 91, 96,
+        100, 103, 105, 108, 110, 112, 113, 115, 115, 91, 87, 84, 87, 91, 95, 99,
+        102, 106, 108, 110, 112, 114, 115, 116, 116, 95, 91, 88, 91, 94, 98,
+        102, 104, 108, 110, 112, 114, 115, 116, 118, 118, 100, 96, 93, 96, 98,
+        101, 104, 107, 110, 112, 114, 115, 117, 118, 119, 119, 103, 100, 97, 99,
+        101, 104, 106, 109, 111, 113, 115, 116, 118, 119, 120, 120, 107, 104,
+        101, 103, 105, 107, 109, 111, 113, 115, 116, 118, 119, 120, 121, 121,
+        107, 104, 101, 103, 105, 107, 109, 111, 113, 115, 116, 118, 119, 120,
+        121, 121,
+        /* Size 32 */
+        64, 61, 58, 55, 53, 54, 54, 55, 56, 58, 61, 64, 67, 70, 72, 76, 79, 82,
+        84, 87, 91, 93, 95, 97, 100, 102, 103, 105, 107, 107, 107, 107, 61, 59,
+        57, 56, 54, 55, 55, 56, 56, 58, 60, 63, 66, 68, 71, 74, 78, 80, 83, 86,
+        89, 91, 93, 96, 98, 100, 102, 104, 106, 106, 106, 106, 58, 57, 57, 56,
+        55, 56, 56, 56, 56, 58, 60, 62, 65, 67, 70, 73, 76, 78, 81, 84, 87, 89,
+        91, 94, 96, 98, 100, 102, 104, 104, 104, 104, 55, 56, 56, 56, 57, 57,
+        57, 56, 56, 58, 60, 62, 64, 66, 69, 71, 74, 77, 79, 82, 85, 87, 90, 92,
+        95, 97, 99, 101, 103, 103, 103, 103, 53, 54, 55, 57, 58, 58, 57, 57, 56,
+        58, 59, 61, 63, 65, 67, 70, 73, 75, 78, 81, 84, 86, 88, 91, 93, 95, 97,
+        99, 101, 101, 101, 101, 54, 55, 56, 57, 58, 58, 59, 59, 59, 61, 62, 64,
+        66, 68, 70, 73, 75, 77, 80, 82, 85, 87, 90, 92, 94, 96, 98, 100, 102,
+        102, 102, 102, 54, 55, 56, 57, 57, 59, 60, 61, 63, 64, 66, 67, 69, 71,
+        73, 75, 78, 80, 82, 84, 87, 89, 91, 93, 96, 97, 99, 101, 103, 103, 103,
+        103, 55, 56, 56, 56, 57, 59, 61, 63, 66, 68, 69, 71, 73, 74, 76, 78, 80,
+        82, 84, 87, 89, 91, 93, 95, 97, 99, 100, 102, 104, 104, 104, 104, 56,
+        56, 56, 56, 56, 59, 63, 66, 70, 72, 73, 75, 77, 78, 80, 81, 83, 85, 87,
+        89, 91, 93, 94, 96, 98, 100, 101, 103, 105, 105, 105, 105, 58, 58, 58,
+        58, 58, 61, 64, 68, 72, 73, 75, 77, 79, 80, 82, 84, 86, 87, 89, 91, 93,
+        94, 96, 98, 100, 101, 103, 104, 106, 106, 106, 106, 61, 60, 60, 60, 59,
+        62, 66, 69, 73, 75, 77, 79, 81, 83, 85, 86, 88, 90, 91, 93, 95, 96, 98,
+        99, 101, 102, 104, 105, 107, 107, 107, 107, 64, 63, 62, 62, 61, 64, 67,
+        71, 75, 77, 79, 82, 84, 86, 87, 89, 91, 92, 94, 95, 97, 98, 100, 101,
+        103, 104, 105, 106, 108, 108, 108, 108, 67, 66, 65, 64, 63, 66, 69, 73,
+        77, 79, 81, 84, 87, 88, 90, 92, 94, 95, 96, 98, 99, 100, 102, 103, 104,
+        105, 106, 108, 109, 109, 109, 109, 70, 68, 67, 66, 65, 68, 71, 74, 78,
+        80, 83, 86, 88, 90, 92, 93, 95, 97, 98, 99, 101, 102, 103, 104, 106,
+        107, 108, 109, 110, 110, 110, 110, 72, 71, 70, 69, 67, 70, 73, 76, 80,
+        82, 85, 87, 90, 92, 93, 95, 97, 98, 100, 101, 102, 103, 104, 106, 107,
+        108, 109, 110, 111, 111, 111, 111, 76, 74, 73, 71, 70, 73, 75, 78, 81,
+        84, 86, 89, 92, 93, 95, 97, 99, 100, 101, 103, 104, 105, 106, 107, 108,
+        109, 110, 111, 112, 112, 112, 112, 79, 78, 76, 74, 73, 75, 78, 80, 83,
+        86, 88, 91, 94, 95, 97, 99, 101, 102, 103, 104, 106, 107, 108, 109, 110,
+        110, 111, 112, 113, 113, 113, 113, 82, 80, 78, 77, 75, 77, 80, 82, 85,
+        87, 90, 92, 95, 97, 98, 100, 102, 103, 104, 106, 107, 108, 109, 110,
+        111, 111, 112, 113, 114, 114, 114, 114, 84, 83, 81, 79, 78, 80, 82, 84,
+        87, 89, 91, 94, 96, 98, 100, 101, 103, 104, 105, 107, 108, 109, 110,
+        111, 112, 112, 113, 114, 115, 115, 115, 115, 87, 86, 84, 82, 81, 82, 84,
+        87, 89, 91, 93, 95, 98, 99, 101, 103, 104, 106, 107, 108, 109, 110, 111,
+        112, 113, 113, 114, 115, 116, 116, 116, 116, 91, 89, 87, 85, 84, 85, 87,
+        89, 91, 93, 95, 97, 99, 101, 102, 104, 106, 107, 108, 109, 110, 111,
+        112, 113, 114, 114, 115, 116, 116, 116, 116, 116, 93, 91, 89, 87, 86,
+        87, 89, 91, 93, 94, 96, 98, 100, 102, 103, 105, 107, 108, 109, 110, 111,
+        112, 113, 114, 114, 115, 116, 116, 117, 117, 117, 117, 95, 93, 91, 90,
+        88, 90, 91, 93, 94, 96, 98, 100, 102, 103, 104, 106, 108, 109, 110, 111,
+        112, 113, 114, 114, 115, 116, 116, 117, 118, 118, 118, 118, 97, 96, 94,
+        92, 91, 92, 93, 95, 96, 98, 99, 101, 103, 104, 106, 107, 109, 110, 111,
+        112, 113, 114, 114, 115, 116, 117, 117, 118, 118, 118, 118, 118, 100,
+        98, 96, 95, 93, 94, 96, 97, 98, 100, 101, 103, 104, 106, 107, 108, 110,
+        111, 112, 113, 114, 114, 115, 116, 117, 117, 118, 118, 119, 119, 119,
+        119, 102, 100, 98, 97, 95, 96, 97, 99, 100, 101, 102, 104, 105, 107,
+        108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 117, 118, 118, 119,
+        119, 119, 119, 119, 103, 102, 100, 99, 97, 98, 99, 100, 101, 103, 104,
+        105, 106, 108, 109, 110, 111, 112, 113, 114, 115, 116, 116, 117, 118,
+        118, 119, 119, 120, 120, 120, 120, 105, 104, 102, 101, 99, 100, 101,
+        102, 103, 104, 105, 106, 108, 109, 110, 111, 112, 113, 114, 115, 116,
+        116, 117, 118, 118, 119, 119, 120, 120, 120, 120, 120, 107, 106, 104,
+        103, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113,
+        114, 115, 116, 116, 117, 118, 118, 119, 119, 120, 120, 121, 121, 121,
+        121, 107, 106, 104, 103, 101, 102, 103, 104, 105, 106, 107, 108, 109,
+        110, 111, 112, 113, 114, 115, 116, 116, 117, 118, 118, 119, 119, 120,
+        120, 121, 121, 121, 121, 107, 106, 104, 103, 101, 102, 103, 104, 105,
+        106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 116, 117, 118,
+        118, 119, 119, 120, 120, 121, 121, 121, 121, 107, 106, 104, 103, 101,
+        102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115,
+        116, 116, 117, 118, 118, 119, 119, 120, 120, 121, 121, 121, 121 },
+      { /* Intra matrices */
+        /* Size 4 */
+        38, 40, 55, 69, 40, 52, 62, 72, 55, 62, 73, 79, 69, 72, 79, 83,
+        /* Size 8 */
+        43, 35, 37, 45, 53, 62, 68, 74, 35, 39, 37, 42, 49, 56, 64, 70, 37, 37,
+        47, 52, 56, 62, 67, 72, 45, 42, 52, 59, 64, 68, 72, 75, 53, 49, 56, 64,
+        69, 73, 76, 78, 62, 56, 62, 68, 73, 76, 79, 81, 68, 64, 67, 72, 76, 79,
+        81, 83, 74, 70, 72, 75, 78, 81, 83, 84,
+        /* Size 16 */
+        42, 38, 34, 35, 36, 40, 44, 48, 52, 56, 60, 63, 67, 70, 72, 72, 38, 37,
+        36, 36, 36, 39, 42, 46, 50, 54, 58, 61, 65, 67, 70, 70, 34, 36, 38, 37,
+        37, 39, 41, 44, 48, 51, 55, 59, 62, 65, 68, 68, 35, 36, 37, 39, 41, 43,
+        45, 48, 51, 54, 58, 61, 64, 67, 69, 69, 36, 36, 37, 41, 46, 48, 50, 53,
+        55, 58, 61, 63, 66, 68, 70, 70, 40, 39, 39, 43, 48, 51, 54, 56, 59, 61,
+        63, 66, 68, 70, 72, 72, 44, 42, 41, 45, 50, 54, 58, 60, 62, 64, 66, 68,
+        70, 72, 74, 74, 48, 46, 44, 48, 53, 56, 60, 62, 65, 67, 69, 70, 72, 74,
+        75, 75, 52, 50, 48, 51, 55, 59, 62, 65, 68, 69, 71, 73, 74, 75, 77, 77,
+        56, 54, 51, 54, 58, 61, 64, 67, 69, 71, 73, 74, 76, 77, 78, 78, 60, 58,
+        55, 58, 61, 63, 66, 69, 71, 73, 75, 76, 77, 78, 79, 79, 63, 61, 59, 61,
+        63, 66, 68, 70, 73, 74, 76, 77, 78, 79, 80, 80, 67, 65, 62, 64, 66, 68,
+        70, 72, 74, 76, 77, 78, 79, 80, 81, 81, 70, 67, 65, 67, 68, 70, 72, 74,
+        75, 77, 78, 79, 80, 81, 82, 82, 72, 70, 68, 69, 70, 72, 74, 75, 77, 78,
+        79, 80, 81, 82, 82, 82, 72, 70, 68, 69, 70, 72, 74, 75, 77, 78, 79, 80,
+        81, 82, 82, 82,
+        /* Size 32 */
+        41, 39, 37, 35, 34, 34, 35, 35, 36, 37, 39, 41, 43, 45, 47, 49, 52, 54,
+        55, 57, 60, 61, 63, 65, 66, 68, 69, 70, 72, 72, 72, 72, 39, 38, 37, 36,
+        35, 35, 35, 36, 36, 37, 39, 41, 43, 44, 46, 48, 51, 52, 54, 56, 58, 60,
+        62, 63, 65, 66, 68, 69, 70, 70, 70, 70, 37, 37, 36, 36, 36, 36, 36, 36,
+        36, 37, 39, 40, 42, 43, 45, 47, 49, 51, 53, 55, 57, 59, 60, 62, 64, 65,
+        67, 68, 69, 69, 69, 69, 35, 36, 36, 36, 36, 36, 36, 36, 36, 37, 38, 40,
+        41, 43, 44, 46, 48, 50, 52, 54, 56, 57, 59, 61, 63, 64, 65, 67, 68, 68,
+        68, 68, 34, 35, 36, 36, 37, 37, 37, 37, 36, 37, 38, 39, 40, 42, 44, 45,
+        47, 49, 51, 53, 55, 56, 58, 60, 62, 63, 64, 66, 67, 67, 67, 67, 34, 35,
+        36, 36, 37, 37, 38, 38, 38, 39, 40, 41, 42, 44, 45, 47, 49, 51, 52, 54,
+        56, 57, 59, 61, 62, 64, 65, 66, 68, 68, 68, 68, 35, 35, 36, 36, 37, 38,
+        38, 39, 40, 41, 42, 44, 45, 46, 48, 49, 51, 52, 54, 55, 57, 59, 60, 62,
+        63, 65, 66, 67, 69, 69, 69, 69, 35, 36, 36, 36, 37, 38, 39, 41, 43, 44,
+        45, 46, 47, 48, 50, 51, 53, 54, 55, 57, 59, 60, 61, 63, 64, 65, 67, 68,
+        69, 69, 69, 69, 36, 36, 36, 36, 36, 38, 40, 43, 45, 46, 48, 49, 50, 51,
+        52, 53, 55, 56, 57, 58, 60, 61, 62, 64, 65, 66, 67, 69, 70, 70, 70, 70,
+        37, 37, 37, 37, 37, 39, 41, 44, 46, 48, 49, 50, 52, 53, 54, 55, 56, 57,
+        59, 60, 61, 62, 64, 65, 66, 67, 68, 69, 71, 71, 71, 71, 39, 39, 39, 38,
+        38, 40, 42, 45, 48, 49, 50, 52, 53, 54, 56, 57, 58, 59, 60, 61, 63, 64,
+        65, 66, 67, 68, 69, 70, 71, 71, 71, 71, 41, 41, 40, 40, 39, 41, 44, 46,
+        49, 50, 52, 53, 55, 56, 57, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69,
+        70, 71, 72, 72, 72, 72, 43, 43, 42, 41, 40, 42, 45, 47, 50, 52, 53, 55,
+        57, 58, 59, 61, 62, 63, 64, 65, 66, 67, 68, 68, 69, 70, 71, 72, 73, 73,
+        73, 73, 45, 44, 43, 43, 42, 44, 46, 48, 51, 53, 54, 56, 58, 59, 61, 62,
+        63, 64, 65, 66, 67, 68, 69, 69, 70, 71, 72, 73, 74, 74, 74, 74, 47, 46,
+        45, 44, 44, 45, 48, 50, 52, 54, 56, 57, 59, 61, 62, 63, 64, 65, 66, 67,
+        68, 69, 70, 71, 71, 72, 73, 74, 74, 74, 74, 74, 49, 48, 47, 46, 45, 47,
+        49, 51, 53, 55, 57, 59, 61, 62, 63, 64, 66, 66, 67, 68, 69, 70, 71, 72,
+        72, 73, 74, 74, 75, 75, 75, 75, 52, 51, 49, 48, 47, 49, 51, 53, 55, 56,
+        58, 60, 62, 63, 64, 66, 67, 68, 69, 70, 71, 71, 72, 73, 73, 74, 75, 75,
+        76, 76, 76, 76, 54, 52, 51, 50, 49, 51, 52, 54, 56, 57, 59, 61, 63, 64,
+        65, 66, 68, 69, 70, 70, 71, 72, 73, 73, 74, 75, 75, 76, 76, 76, 76, 76,
+        55, 54, 53, 52, 51, 52, 54, 55, 57, 59, 60, 62, 64, 65, 66, 67, 69, 70,
+        70, 71, 72, 73, 73, 74, 75, 75, 76, 77, 77, 77, 77, 77, 57, 56, 55, 54,
+        53, 54, 55, 57, 58, 60, 61, 63, 65, 66, 67, 68, 70, 70, 71, 72, 73, 74,
+        74, 75, 76, 76, 77, 77, 78, 78, 78, 78, 60, 58, 57, 56, 55, 56, 57, 59,
+        60, 61, 63, 64, 66, 67, 68, 69, 71, 71, 72, 73, 74, 75, 75, 76, 76, 77,
+        77, 78, 78, 78, 78, 78, 61, 60, 59, 57, 56, 57, 59, 60, 61, 62, 64, 65,
+        67, 68, 69, 70, 71, 72, 73, 74, 75, 75, 76, 76, 77, 77, 78, 78, 79, 79,
+        79, 79, 63, 62, 60, 59, 58, 59, 60, 61, 62, 64, 65, 66, 68, 69, 70, 71,
+        72, 73, 73, 74, 75, 76, 76, 77, 77, 78, 78, 79, 79, 79, 79, 79, 65, 63,
+        62, 61, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 71, 72, 73, 73, 74, 75,
+        76, 76, 77, 77, 78, 78, 79, 79, 80, 80, 80, 80, 66, 65, 64, 63, 62, 62,
+        63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 76, 77, 77, 78,
+        79, 79, 79, 80, 80, 80, 80, 80, 68, 66, 65, 64, 63, 64, 65, 65, 66, 67,
+        68, 69, 70, 71, 72, 73, 74, 75, 75, 76, 77, 77, 78, 78, 79, 79, 80, 80,
+        81, 81, 81, 81, 69, 68, 67, 65, 64, 65, 66, 67, 67, 68, 69, 70, 71, 72,
+        73, 74, 75, 75, 76, 77, 77, 78, 78, 79, 79, 80, 80, 81, 81, 81, 81, 81,
+        70, 69, 68, 67, 66, 66, 67, 68, 69, 69, 70, 71, 72, 73, 74, 74, 75, 76,
+        77, 77, 78, 78, 79, 79, 80, 80, 81, 81, 81, 81, 81, 81, 72, 70, 69, 68,
+        67, 68, 69, 69, 70, 71, 71, 72, 73, 74, 74, 75, 76, 76, 77, 78, 78, 79,
+        79, 80, 80, 81, 81, 81, 82, 82, 82, 82, 72, 70, 69, 68, 67, 68, 69, 69,
+        70, 71, 71, 72, 73, 74, 74, 75, 76, 76, 77, 78, 78, 79, 79, 80, 80, 81,
+        81, 81, 82, 82, 82, 82, 72, 70, 69, 68, 67, 68, 69, 69, 70, 71, 71, 72,
+        73, 74, 74, 75, 76, 76, 77, 78, 78, 79, 79, 80, 80, 81, 81, 81, 82, 82,
+        82, 82, 72, 70, 69, 68, 67, 68, 69, 69, 70, 71, 71, 72, 73, 74, 74, 75,
+        76, 76, 77, 78, 78, 79, 79, 80, 80, 81, 81, 81, 82, 82, 82, 82 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 86, 90, 101, 86, 95, 98, 105, 90, 98, 110, 117, 101, 105, 117, 125,
+        /* Size 8 */
+        64, 55, 81, 84, 87, 92, 98, 104, 55, 71, 82, 77, 79, 83, 89, 95, 81, 82,
+        89, 88, 88, 90, 94, 99, 84, 77, 88, 94, 96, 98, 101, 104, 87, 79, 88,
+        96, 101, 104, 107, 110, 92, 83, 90, 98, 104, 109, 112, 115, 98, 89, 94,
+        101, 107, 112, 116, 118, 104, 95, 99, 104, 110, 115, 118, 121,
+        /* Size 16 */
+        64, 59, 55, 66, 81, 82, 84, 85, 87, 90, 92, 95, 98, 101, 104, 104, 59,
+        61, 62, 70, 82, 81, 80, 82, 83, 85, 87, 90, 93, 96, 99, 99, 55, 62, 71,
+        76, 82, 80, 77, 78, 79, 81, 83, 86, 89, 92, 95, 95, 66, 70, 76, 80, 86,
+        84, 82, 83, 83, 85, 86, 89, 91, 94, 97, 97, 81, 82, 82, 86, 89, 89, 88,
+        88, 88, 89, 90, 92, 94, 96, 99, 99, 82, 81, 80, 84, 89, 90, 91, 91, 92,
+        93, 94, 96, 97, 99, 101, 101, 84, 80, 77, 82, 88, 91, 94, 95, 96, 97,
+        98, 99, 101, 102, 104, 104, 85, 82, 78, 83, 88, 91, 95, 97, 98, 100,
+        101, 102, 104, 105, 107, 107, 87, 83, 79, 83, 88, 92, 96, 98, 101, 103,
+        104, 106, 107, 108, 110, 110, 90, 85, 81, 85, 89, 93, 97, 100, 103, 105,
+        107, 108, 109, 111, 112, 112, 92, 87, 83, 86, 90, 94, 98, 101, 104, 107,
+        109, 110, 112, 113, 115, 115, 95, 90, 86, 89, 92, 96, 99, 102, 106, 108,
+        110, 112, 114, 115, 116, 116, 98, 93, 89, 91, 94, 97, 101, 104, 107,
+        109, 112, 114, 116, 117, 118, 118, 101, 96, 92, 94, 96, 99, 102, 105,
+        108, 111, 113, 115, 117, 118, 120, 120, 104, 99, 95, 97, 99, 101, 104,
+        107, 110, 112, 115, 116, 118, 120, 121, 121, 104, 99, 95, 97, 99, 101,
+        104, 107, 110, 112, 115, 116, 118, 120, 121, 121,
+        /* Size 32 */
+        64, 62, 59, 57, 55, 60, 66, 73, 81, 82, 82, 83, 84, 85, 85, 86, 87, 88,
+        90, 91, 92, 94, 95, 97, 98, 100, 101, 102, 104, 104, 104, 104, 62, 61,
+        60, 59, 58, 63, 68, 74, 81, 82, 82, 82, 82, 83, 83, 84, 85, 86, 87, 88,
+        90, 91, 93, 94, 96, 97, 98, 100, 101, 101, 101, 101, 59, 60, 61, 61, 62,
+        66, 70, 76, 82, 81, 81, 81, 80, 81, 82, 82, 83, 84, 85, 86, 87, 89, 90,
+        92, 93, 95, 96, 98, 99, 99, 99, 99, 57, 59, 61, 64, 66, 69, 73, 77, 82,
+        81, 80, 80, 79, 79, 80, 80, 81, 82, 83, 84, 85, 87, 88, 89, 91, 92, 94,
+        95, 97, 97, 97, 97, 55, 58, 62, 66, 71, 73, 76, 79, 82, 81, 80, 78, 77,
+        78, 78, 78, 79, 80, 81, 82, 83, 84, 86, 87, 89, 90, 92, 93, 95, 95, 95,
+        95, 60, 63, 66, 69, 73, 76, 78, 81, 84, 83, 82, 81, 80, 80, 80, 81, 81,
+        82, 83, 84, 85, 86, 87, 89, 90, 91, 93, 94, 96, 96, 96, 96, 66, 68, 70,
+        73, 76, 78, 80, 83, 86, 85, 84, 83, 82, 82, 83, 83, 83, 84, 85, 86, 86,
+        88, 89, 90, 91, 93, 94, 95, 97, 97, 97, 97, 73, 74, 76, 77, 79, 81, 83,
+        85, 87, 87, 86, 86, 85, 85, 85, 85, 85, 86, 87, 87, 88, 89, 90, 92, 93,
+        94, 95, 96, 98, 98, 98, 98, 81, 81, 82, 82, 82, 84, 86, 87, 89, 89, 89,
+        88, 88, 88, 88, 88, 88, 88, 89, 89, 90, 91, 92, 93, 94, 95, 96, 98, 99,
+        99, 99, 99, 82, 82, 81, 81, 81, 83, 85, 87, 89, 89, 89, 89, 89, 89, 89,
+        89, 90, 90, 91, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 100, 100, 100,
+        82, 82, 81, 80, 80, 82, 84, 86, 89, 89, 90, 90, 91, 91, 91, 91, 92, 92,
+        93, 93, 94, 95, 96, 96, 97, 98, 99, 100, 101, 101, 101, 101, 83, 82, 81,
+        80, 78, 81, 83, 86, 88, 89, 90, 91, 92, 92, 93, 93, 94, 94, 95, 95, 96,
+        97, 97, 98, 99, 100, 101, 102, 103, 103, 103, 103, 84, 82, 80, 79, 77,
+        80, 82, 85, 88, 89, 91, 92, 94, 94, 95, 95, 96, 96, 97, 97, 98, 99, 99,
+        100, 101, 102, 102, 103, 104, 104, 104, 104, 85, 83, 81, 79, 78, 80, 82,
+        85, 88, 89, 91, 92, 94, 95, 96, 96, 97, 98, 98, 99, 99, 100, 101, 102,
+        102, 103, 104, 105, 106, 106, 106, 106, 85, 83, 82, 80, 78, 80, 83, 85,
+        88, 89, 91, 93, 95, 96, 97, 98, 98, 99, 100, 100, 101, 102, 102, 103,
+        104, 105, 105, 106, 107, 107, 107, 107, 86, 84, 82, 80, 78, 81, 83, 85,
+        88, 89, 91, 93, 95, 96, 98, 99, 100, 100, 101, 102, 103, 103, 104, 105,
+        105, 106, 107, 108, 108, 108, 108, 108, 87, 85, 83, 81, 79, 81, 83, 85,
+        88, 90, 92, 94, 96, 97, 98, 100, 101, 102, 103, 104, 104, 105, 106, 106,
+        107, 108, 108, 109, 110, 110, 110, 110, 88, 86, 84, 82, 80, 82, 84, 86,
+        88, 90, 92, 94, 96, 98, 99, 100, 102, 103, 104, 105, 105, 106, 107, 108,
+        108, 109, 110, 110, 111, 111, 111, 111, 90, 87, 85, 83, 81, 83, 85, 87,
+        89, 91, 93, 95, 97, 98, 100, 101, 103, 104, 105, 106, 107, 107, 108,
+        109, 109, 110, 111, 111, 112, 112, 112, 112, 91, 88, 86, 84, 82, 84, 86,
+        87, 89, 91, 93, 95, 97, 99, 100, 102, 104, 105, 106, 107, 108, 108, 109,
+        110, 111, 111, 112, 113, 113, 113, 113, 113, 92, 90, 87, 85, 83, 85, 86,
+        88, 90, 92, 94, 96, 98, 99, 101, 103, 104, 105, 107, 108, 109, 110, 110,
+        111, 112, 113, 113, 114, 115, 115, 115, 115, 94, 91, 89, 87, 84, 86, 88,
+        89, 91, 93, 95, 97, 99, 100, 102, 103, 105, 106, 107, 108, 110, 110,
+        111, 112, 113, 114, 114, 115, 116, 116, 116, 116, 95, 93, 90, 88, 86,
+        87, 89, 90, 92, 94, 96, 97, 99, 101, 102, 104, 106, 107, 108, 109, 110,
+        111, 112, 113, 114, 114, 115, 116, 116, 116, 116, 116, 97, 94, 92, 89,
+        87, 89, 90, 92, 93, 95, 96, 98, 100, 102, 103, 105, 106, 108, 109, 110,
+        111, 112, 113, 114, 115, 115, 116, 117, 117, 117, 117, 117, 98, 96, 93,
+        91, 89, 90, 91, 93, 94, 96, 97, 99, 101, 102, 104, 105, 107, 108, 109,
+        111, 112, 113, 114, 115, 116, 116, 117, 118, 118, 118, 118, 118, 100,
+        97, 95, 92, 90, 91, 93, 94, 95, 97, 98, 100, 102, 103, 105, 106, 108,
+        109, 110, 111, 113, 114, 114, 115, 116, 117, 118, 118, 119, 119, 119,
+        119, 101, 98, 96, 94, 92, 93, 94, 95, 96, 98, 99, 101, 102, 104, 105,
+        107, 108, 110, 111, 112, 113, 114, 115, 116, 117, 118, 118, 119, 120,
+        120, 120, 120, 102, 100, 98, 95, 93, 94, 95, 96, 98, 99, 100, 102, 103,
+        105, 106, 108, 109, 110, 111, 113, 114, 115, 116, 117, 118, 118, 119,
+        120, 121, 121, 121, 121, 104, 101, 99, 97, 95, 96, 97, 98, 99, 100, 101,
+        103, 104, 106, 107, 108, 110, 111, 112, 113, 115, 116, 116, 117, 118,
+        119, 120, 121, 121, 121, 121, 121, 104, 101, 99, 97, 95, 96, 97, 98, 99,
+        100, 101, 103, 104, 106, 107, 108, 110, 111, 112, 113, 115, 116, 116,
+        117, 118, 119, 120, 121, 121, 121, 121, 121, 104, 101, 99, 97, 95, 96,
+        97, 98, 99, 100, 101, 103, 104, 106, 107, 108, 110, 111, 112, 113, 115,
+        116, 116, 117, 118, 119, 120, 121, 121, 121, 121, 121, 104, 101, 99, 97,
+        95, 96, 97, 98, 99, 100, 101, 103, 104, 106, 107, 108, 110, 111, 112,
+        113, 115, 116, 116, 117, 118, 119, 120, 121, 121, 121, 121, 121 },
+      { /* Intra matrices */
+        /* Size 4 */
+        40, 54, 57, 65, 54, 60, 62, 67, 57, 62, 71, 75, 65, 67, 75, 81,
+        /* Size 8 */
+        42, 36, 54, 56, 58, 62, 66, 70, 36, 47, 54, 51, 52, 55, 59, 64, 54, 54,
+        60, 59, 58, 60, 63, 66, 56, 51, 59, 63, 64, 66, 68, 70, 58, 52, 58, 64,
+        68, 70, 72, 74, 62, 55, 60, 66, 70, 74, 76, 78, 66, 59, 63, 68, 72, 76,
+        79, 81, 70, 64, 66, 70, 74, 78, 81, 83,
+        /* Size 16 */
+        41, 38, 35, 43, 53, 54, 55, 56, 57, 59, 61, 63, 65, 67, 69, 69, 38, 39,
+        40, 46, 53, 53, 53, 53, 54, 56, 57, 59, 62, 64, 66, 66, 35, 40, 46, 49,
+        54, 52, 51, 51, 51, 53, 54, 56, 58, 60, 63, 63, 43, 46, 49, 53, 56, 55,
+        54, 54, 54, 56, 57, 58, 60, 62, 64, 64, 53, 53, 54, 56, 59, 58, 58, 58,
+        58, 58, 59, 61, 62, 64, 65, 65, 54, 53, 52, 55, 58, 59, 60, 60, 60, 61,
+        62, 63, 64, 66, 67, 67, 55, 53, 51, 54, 58, 60, 62, 63, 63, 64, 65, 66,
+        67, 68, 69, 69, 56, 53, 51, 54, 58, 60, 63, 64, 65, 66, 67, 68, 69, 70,
+        71, 71, 57, 54, 51, 54, 58, 60, 63, 65, 67, 68, 69, 70, 71, 72, 73, 73,
+        59, 56, 53, 56, 58, 61, 64, 66, 68, 70, 71, 72, 73, 74, 75, 75, 61, 57,
+        54, 57, 59, 62, 65, 67, 69, 71, 73, 74, 75, 76, 77, 77, 63, 59, 56, 58,
+        61, 63, 66, 68, 70, 72, 74, 75, 76, 77, 78, 78, 65, 62, 58, 60, 62, 64,
+        67, 69, 71, 73, 75, 76, 78, 79, 80, 80, 67, 64, 60, 62, 64, 66, 68, 70,
+        72, 74, 76, 77, 79, 80, 81, 81, 69, 66, 63, 64, 65, 67, 69, 71, 73, 75,
+        77, 78, 80, 81, 82, 82, 69, 66, 63, 64, 65, 67, 69, 71, 73, 75, 77, 78,
+        80, 81, 82, 82,
+        /* Size 32 */
+        41, 39, 38, 37, 35, 38, 42, 47, 53, 53, 54, 54, 55, 55, 56, 56, 57, 58,
+        59, 59, 60, 61, 62, 63, 65, 65, 66, 68, 69, 69, 69, 69, 39, 39, 38, 38,
+        37, 40, 44, 48, 53, 53, 53, 53, 53, 54, 54, 55, 55, 56, 57, 58, 59, 60,
+        61, 62, 63, 64, 65, 66, 67, 67, 67, 67, 38, 38, 39, 39, 40, 42, 45, 49,
+        53, 53, 53, 52, 52, 53, 53, 53, 54, 55, 55, 56, 57, 58, 59, 60, 61, 62,
+        63, 64, 65, 65, 65, 65, 37, 38, 39, 41, 42, 45, 47, 50, 53, 53, 52, 52,
+        51, 51, 52, 52, 52, 53, 54, 55, 56, 56, 57, 58, 59, 60, 61, 63, 64, 64,
+        64, 64, 35, 37, 40, 42, 46, 47, 49, 51, 53, 52, 52, 51, 50, 50, 51, 51,
+        51, 52, 53, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 62, 62, 62, 38, 40,
+        42, 45, 47, 49, 51, 52, 54, 54, 53, 52, 52, 52, 52, 52, 52, 53, 54, 54,
+        55, 56, 57, 58, 59, 60, 61, 62, 63, 63, 63, 63, 42, 44, 45, 47, 49, 51,
+        52, 54, 56, 55, 55, 54, 53, 54, 54, 54, 54, 55, 55, 56, 56, 57, 58, 59,
+        60, 61, 62, 63, 64, 64, 64, 64, 47, 48, 49, 50, 51, 52, 54, 55, 57, 57,
+        56, 56, 55, 55, 55, 55, 55, 56, 57, 57, 58, 58, 59, 60, 61, 62, 62, 63,
+        64, 64, 64, 64, 53, 53, 53, 53, 53, 54, 56, 57, 58, 58, 58, 58, 57, 57,
+        57, 57, 57, 58, 58, 58, 59, 60, 60, 61, 62, 62, 63, 64, 65, 65, 65, 65,
+        53, 53, 53, 53, 52, 54, 55, 57, 58, 58, 58, 58, 58, 58, 58, 58, 59, 59,
+        59, 60, 60, 61, 61, 62, 63, 64, 64, 65, 66, 66, 66, 66, 54, 53, 53, 52,
+        52, 53, 55, 56, 58, 58, 59, 59, 59, 59, 60, 60, 60, 60, 61, 61, 61, 62,
+        63, 63, 64, 65, 65, 66, 67, 67, 67, 67, 54, 53, 52, 52, 51, 52, 54, 56,
+        58, 58, 59, 60, 60, 61, 61, 61, 61, 62, 62, 63, 63, 63, 64, 65, 65, 66,
+        66, 67, 68, 68, 68, 68, 55, 53, 52, 51, 50, 52, 53, 55, 57, 58, 59, 60,
+        61, 62, 62, 63, 63, 63, 64, 64, 64, 65, 65, 66, 66, 67, 68, 68, 69, 69,
+        69, 69, 55, 54, 53, 51, 50, 52, 54, 55, 57, 58, 59, 61, 62, 62, 63, 63,
+        64, 64, 65, 65, 65, 66, 66, 67, 67, 68, 69, 69, 70, 70, 70, 70, 56, 54,
+        53, 52, 51, 52, 54, 55, 57, 58, 60, 61, 62, 63, 63, 64, 65, 65, 66, 66,
+        67, 67, 68, 68, 69, 69, 70, 70, 71, 71, 71, 71, 56, 55, 53, 52, 51, 52,
+        54, 55, 57, 58, 60, 61, 63, 63, 64, 65, 66, 66, 67, 67, 68, 68, 69, 69,
+        70, 70, 71, 71, 72, 72, 72, 72, 57, 55, 54, 52, 51, 52, 54, 55, 57, 59,
+        60, 61, 63, 64, 65, 66, 67, 67, 68, 68, 69, 69, 70, 70, 71, 71, 72, 72,
+        73, 73, 73, 73, 58, 56, 55, 53, 52, 53, 55, 56, 58, 59, 60, 62, 63, 64,
+        65, 66, 67, 68, 68, 69, 70, 70, 71, 71, 72, 72, 73, 73, 74, 74, 74, 74,
+        59, 57, 55, 54, 53, 54, 55, 57, 58, 59, 61, 62, 64, 65, 66, 67, 68, 68,
+        69, 70, 70, 71, 71, 72, 73, 73, 73, 74, 74, 74, 74, 74, 59, 58, 56, 55,
+        53, 54, 56, 57, 58, 60, 61, 63, 64, 65, 66, 67, 68, 69, 70, 71, 71, 72,
+        72, 73, 73, 74, 74, 75, 75, 75, 75, 75, 60, 59, 57, 56, 54, 55, 56, 58,
+        59, 60, 61, 63, 64, 65, 67, 68, 69, 70, 70, 71, 72, 73, 73, 74, 74, 75,
+        75, 76, 76, 76, 76, 76, 61, 60, 58, 56, 55, 56, 57, 58, 60, 61, 62, 63,
+        65, 66, 67, 68, 69, 70, 71, 72, 73, 73, 74, 74, 75, 75, 76, 76, 77, 77,
+        77, 77, 62, 61, 59, 57, 56, 57, 58, 59, 60, 61, 63, 64, 65, 66, 68, 69,
+        70, 71, 71, 72, 73, 74, 74, 75, 76, 76, 77, 77, 78, 78, 78, 78, 63, 62,
+        60, 58, 57, 58, 59, 60, 61, 62, 63, 65, 66, 67, 68, 69, 70, 71, 72, 73,
+        74, 74, 75, 76, 76, 77, 77, 78, 78, 78, 78, 78, 65, 63, 61, 59, 58, 59,
+        60, 61, 62, 63, 64, 65, 66, 67, 69, 70, 71, 72, 73, 73, 74, 75, 76, 76,
+        77, 77, 78, 78, 79, 79, 79, 79, 65, 64, 62, 60, 59, 60, 61, 62, 62, 64,
+        65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 75, 76, 77, 77, 78, 78, 79,
+        79, 79, 79, 79, 66, 65, 63, 61, 60, 61, 62, 62, 63, 64, 65, 66, 68, 69,
+        70, 71, 72, 73, 73, 74, 75, 76, 77, 77, 78, 78, 79, 79, 80, 80, 80, 80,
+        68, 66, 64, 63, 61, 62, 63, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73,
+        74, 75, 76, 76, 77, 78, 78, 79, 79, 80, 81, 81, 81, 81, 69, 67, 65, 64,
+        62, 63, 64, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 74, 75, 76, 77,
+        78, 78, 79, 79, 80, 81, 81, 81, 81, 81, 69, 67, 65, 64, 62, 63, 64, 64,
+        65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 74, 75, 76, 77, 78, 78, 79, 79,
+        80, 81, 81, 81, 81, 81, 69, 67, 65, 64, 62, 63, 64, 64, 65, 66, 67, 68,
+        69, 70, 71, 72, 73, 74, 74, 75, 76, 77, 78, 78, 79, 79, 80, 81, 81, 81,
+        81, 81, 69, 67, 65, 64, 62, 63, 64, 64, 65, 66, 67, 68, 69, 70, 71, 72,
+        73, 74, 74, 75, 76, 77, 78, 78, 79, 79, 80, 81, 81, 81, 81, 81 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 67, 88, 105, 67, 84, 97, 108, 88, 97, 109, 116, 105, 108, 116, 120,
+        /* Size 8 */
+        64, 54, 57, 66, 77, 86, 94, 99, 54, 59, 57, 63, 72, 81, 88, 95, 57, 57,
+        69, 75, 80, 87, 92, 97, 66, 63, 75, 83, 89, 93, 97, 101, 77, 72, 80, 89,
+        94, 98, 101, 104, 86, 81, 87, 93, 98, 102, 104, 106, 94, 88, 92, 97,
+        101, 104, 106, 108, 99, 95, 97, 101, 104, 106, 108, 109,
+        /* Size 16 */
+        64, 59, 54, 55, 57, 61, 66, 71, 77, 81, 86, 90, 94, 96, 99, 99, 59, 57,
+        56, 57, 57, 61, 65, 69, 74, 79, 83, 87, 91, 94, 97, 97, 54, 56, 59, 58,
+        57, 60, 63, 67, 72, 76, 81, 84, 88, 91, 95, 95, 55, 57, 58, 60, 63, 65,
+        68, 72, 76, 79, 83, 87, 90, 93, 96, 96, 57, 57, 57, 63, 69, 72, 75, 78,
+        80, 83, 87, 89, 92, 95, 97, 97, 61, 61, 60, 65, 72, 75, 79, 82, 84, 87,
+        90, 92, 95, 97, 99, 99, 66, 65, 63, 68, 75, 79, 83, 86, 89, 91, 93, 95,
+        97, 99, 101, 101, 71, 69, 67, 72, 78, 82, 86, 89, 91, 93, 95, 97, 99,
+        100, 102, 102, 77, 74, 72, 76, 80, 84, 89, 91, 94, 96, 98, 100, 101,
+        102, 104, 104, 81, 79, 76, 79, 83, 87, 91, 93, 96, 98, 100, 101, 103,
+        104, 105, 105, 86, 83, 81, 83, 87, 90, 93, 95, 98, 100, 102, 103, 104,
+        105, 106, 106, 90, 87, 84, 87, 89, 92, 95, 97, 100, 101, 103, 104, 105,
+        106, 107, 107, 94, 91, 88, 90, 92, 95, 97, 99, 101, 103, 104, 105, 106,
+        107, 108, 108, 96, 94, 91, 93, 95, 97, 99, 100, 102, 104, 105, 106, 107,
+        108, 109, 109, 99, 97, 95, 96, 97, 99, 101, 102, 104, 105, 106, 107,
+        108, 109, 109, 109, 99, 97, 95, 96, 97, 99, 101, 102, 104, 105, 106,
+        107, 108, 109, 109, 109,
+        /* Size 32 */
+        64, 61, 59, 56, 54, 55, 55, 56, 57, 59, 61, 64, 66, 69, 71, 74, 77, 79,
+        81, 84, 86, 88, 90, 92, 94, 95, 96, 98, 99, 99, 99, 99, 61, 60, 58, 57,
+        55, 56, 56, 56, 57, 59, 61, 63, 66, 68, 70, 73, 76, 78, 80, 82, 85, 87,
+        88, 90, 92, 94, 95, 97, 98, 98, 98, 98, 59, 58, 57, 57, 56, 56, 57, 57,
+        57, 59, 61, 63, 65, 67, 69, 72, 74, 76, 79, 81, 83, 85, 87, 89, 91, 92,
+        94, 95, 97, 97, 97, 97, 56, 57, 57, 57, 58, 57, 57, 57, 57, 59, 60, 62,
+        64, 66, 68, 70, 73, 75, 77, 79, 82, 84, 86, 88, 90, 91, 93, 94, 96, 96,
+        96, 96, 54, 55, 56, 58, 59, 58, 58, 58, 57, 59, 60, 61, 63, 65, 67, 69,
+        72, 74, 76, 78, 81, 82, 84, 86, 88, 90, 91, 93, 95, 95, 95, 95, 55, 56,
+        56, 57, 58, 59, 59, 59, 60, 61, 63, 64, 65, 67, 69, 71, 74, 76, 78, 80,
+        82, 84, 86, 87, 89, 91, 92, 94, 95, 95, 95, 95, 55, 56, 57, 57, 58, 59,
+        60, 61, 63, 64, 65, 67, 68, 70, 72, 74, 76, 78, 79, 81, 83, 85, 87, 89,
+        90, 92, 93, 95, 96, 96, 96, 96, 56, 56, 57, 57, 58, 59, 61, 64, 66, 67,
+        68, 70, 71, 73, 75, 76, 78, 80, 81, 83, 85, 86, 88, 90, 91, 93, 94, 95,
+        97, 97, 97, 97, 57, 57, 57, 57, 57, 60, 63, 66, 69, 71, 72, 73, 75, 76,
+        78, 79, 80, 82, 83, 85, 87, 88, 89, 91, 92, 94, 95, 96, 97, 97, 97, 97,
+        59, 59, 59, 59, 59, 61, 64, 67, 71, 72, 74, 75, 77, 78, 79, 81, 82, 84,
+        85, 87, 88, 89, 91, 92, 93, 95, 96, 97, 98, 98, 98, 98, 61, 61, 61, 60,
+        60, 63, 65, 68, 72, 74, 75, 77, 79, 80, 82, 83, 84, 86, 87, 88, 90, 91,
+        92, 93, 95, 96, 97, 98, 99, 99, 99, 99, 64, 63, 63, 62, 61, 64, 67, 70,
+        73, 75, 77, 79, 81, 82, 84, 85, 86, 88, 89, 90, 91, 92, 93, 95, 96, 97,
+        98, 99, 100, 100, 100, 100, 66, 66, 65, 64, 63, 65, 68, 71, 75, 77, 79,
+        81, 83, 85, 86, 87, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100,
+        101, 101, 101, 101, 69, 68, 67, 66, 65, 67, 70, 73, 76, 78, 80, 82, 85,
+        86, 87, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 100, 101, 101,
+        101, 101, 71, 70, 69, 68, 67, 69, 72, 75, 78, 79, 82, 84, 86, 87, 89,
+        90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 100, 101, 102, 102, 102,
+        102, 74, 73, 72, 70, 69, 71, 74, 76, 79, 81, 83, 85, 87, 89, 90, 91, 93,
+        94, 95, 96, 97, 98, 98, 99, 100, 101, 101, 102, 103, 103, 103, 103, 77,
+        76, 74, 73, 72, 74, 76, 78, 80, 82, 84, 86, 89, 90, 91, 93, 94, 95, 96,
+        97, 98, 99, 100, 100, 101, 102, 102, 103, 104, 104, 104, 104, 79, 78,
+        76, 75, 74, 76, 78, 80, 82, 84, 86, 88, 90, 91, 92, 94, 95, 96, 97, 98,
+        99, 100, 100, 101, 102, 102, 103, 104, 104, 104, 104, 104, 81, 80, 79,
+        77, 76, 78, 79, 81, 83, 85, 87, 89, 91, 92, 93, 95, 96, 97, 98, 99, 100,
+        101, 101, 102, 103, 103, 104, 104, 105, 105, 105, 105, 84, 82, 81, 79,
+        78, 80, 81, 83, 85, 87, 88, 90, 92, 93, 94, 96, 97, 98, 99, 100, 101,
+        101, 102, 103, 103, 104, 104, 105, 105, 105, 105, 105, 86, 85, 83, 82,
+        81, 82, 83, 85, 87, 88, 90, 91, 93, 94, 95, 97, 98, 99, 100, 101, 102,
+        102, 103, 103, 104, 105, 105, 106, 106, 106, 106, 106, 88, 87, 85, 84,
+        82, 84, 85, 86, 88, 89, 91, 92, 94, 95, 96, 98, 99, 100, 101, 101, 102,
+        103, 103, 104, 105, 105, 106, 106, 107, 107, 107, 107, 90, 88, 87, 86,
+        84, 86, 87, 88, 89, 91, 92, 93, 95, 96, 97, 98, 100, 100, 101, 102, 103,
+        103, 104, 105, 105, 106, 106, 107, 107, 107, 107, 107, 92, 90, 89, 88,
+        86, 87, 89, 90, 91, 92, 93, 95, 96, 97, 98, 99, 100, 101, 102, 103, 103,
+        104, 105, 105, 106, 106, 107, 107, 107, 107, 107, 107, 94, 92, 91, 90,
+        88, 89, 90, 91, 92, 93, 95, 96, 97, 98, 99, 100, 101, 102, 103, 103,
+        104, 105, 105, 106, 106, 107, 107, 108, 108, 108, 108, 108, 95, 94, 92,
+        91, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 102, 103,
+        104, 105, 105, 106, 106, 107, 107, 107, 108, 108, 108, 108, 108, 96, 95,
+        94, 93, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 100, 101, 102, 103,
+        104, 104, 105, 106, 106, 107, 107, 107, 108, 108, 109, 109, 109, 109,
+        98, 97, 95, 94, 93, 94, 95, 95, 96, 97, 98, 99, 100, 100, 101, 102, 103,
+        104, 104, 105, 106, 106, 107, 107, 108, 108, 108, 109, 109, 109, 109,
+        109, 99, 98, 97, 96, 95, 95, 96, 97, 97, 98, 99, 100, 101, 101, 102,
+        103, 104, 104, 105, 105, 106, 107, 107, 107, 108, 108, 109, 109, 109,
+        109, 109, 109, 99, 98, 97, 96, 95, 95, 96, 97, 97, 98, 99, 100, 101,
+        101, 102, 103, 104, 104, 105, 105, 106, 107, 107, 107, 108, 108, 109,
+        109, 109, 109, 109, 109, 99, 98, 97, 96, 95, 95, 96, 97, 97, 98, 99,
+        100, 101, 101, 102, 103, 104, 104, 105, 105, 106, 107, 107, 107, 108,
+        108, 109, 109, 109, 109, 109, 109, 99, 98, 97, 96, 95, 95, 96, 97, 97,
+        98, 99, 100, 101, 101, 102, 103, 104, 104, 105, 105, 106, 107, 107, 107,
+        108, 108, 109, 109, 109, 109, 109, 109 },
+      { /* Intra matrices */
+        /* Size 4 */
+        41, 43, 57, 69, 43, 54, 63, 71, 57, 63, 72, 77, 69, 71, 77, 80,
+        /* Size 8 */
+        46, 38, 40, 47, 55, 63, 68, 73, 38, 42, 41, 45, 51, 58, 64, 69, 40, 41,
+        50, 54, 58, 63, 67, 71, 47, 45, 54, 60, 64, 68, 71, 74, 55, 51, 58, 64,
+        69, 72, 74, 76, 63, 58, 63, 68, 72, 75, 77, 78, 68, 64, 67, 71, 74, 77,
+        78, 80, 73, 69, 71, 74, 76, 78, 80, 81,
+        /* Size 16 */
+        45, 41, 38, 38, 39, 43, 47, 50, 54, 58, 61, 64, 67, 69, 71, 71, 41, 40,
+        39, 39, 40, 42, 45, 49, 52, 56, 59, 62, 65, 67, 70, 70, 38, 39, 41, 40,
+        40, 42, 44, 47, 50, 54, 57, 60, 63, 65, 68, 68, 38, 39, 40, 42, 44, 46,
+        48, 51, 54, 56, 59, 62, 65, 67, 69, 69, 39, 40, 40, 44, 49, 51, 53, 55,
+        57, 59, 62, 64, 66, 68, 70, 70, 43, 42, 42, 46, 51, 53, 56, 58, 60, 62,
+        64, 66, 68, 69, 71, 71, 47, 45, 44, 48, 53, 56, 59, 61, 63, 65, 67, 68,
+        70, 71, 72, 72, 50, 49, 47, 51, 55, 58, 61, 63, 65, 67, 69, 70, 71, 72,
+        74, 74, 54, 52, 50, 54, 57, 60, 63, 65, 68, 69, 71, 72, 73, 74, 75, 75,
+        58, 56, 54, 56, 59, 62, 65, 67, 69, 70, 72, 73, 74, 75, 76, 76, 61, 59,
+        57, 59, 62, 64, 67, 69, 71, 72, 73, 74, 75, 76, 77, 77, 64, 62, 60, 62,
+        64, 66, 68, 70, 72, 73, 74, 75, 76, 77, 78, 78, 67, 65, 63, 65, 66, 68,
+        70, 71, 73, 74, 75, 76, 77, 78, 78, 78, 69, 67, 65, 67, 68, 69, 71, 72,
+        74, 75, 76, 77, 78, 78, 79, 79, 71, 70, 68, 69, 70, 71, 72, 74, 75, 76,
+        77, 78, 78, 79, 79, 79, 71, 70, 68, 69, 70, 71, 72, 74, 75, 76, 77, 78,
+        78, 79, 79, 79,
+        /* Size 32 */
+        44, 42, 40, 39, 37, 38, 38, 39, 39, 41, 42, 44, 46, 48, 50, 52, 54, 56,
+        57, 59, 61, 62, 64, 65, 66, 68, 69, 70, 71, 71, 71, 71, 42, 41, 40, 39,
+        38, 38, 39, 39, 39, 41, 42, 44, 45, 47, 49, 51, 53, 54, 56, 58, 60, 61,
+        62, 64, 65, 67, 68, 69, 70, 70, 70, 70, 40, 40, 40, 39, 39, 39, 39, 39,
+        39, 41, 42, 43, 45, 46, 48, 50, 52, 53, 55, 57, 59, 60, 61, 63, 64, 66,
+        67, 68, 69, 69, 69, 69, 39, 39, 39, 39, 40, 40, 40, 39, 39, 40, 42, 43,
+        44, 46, 47, 49, 51, 52, 54, 56, 58, 59, 60, 62, 63, 65, 66, 67, 68, 68,
+        68, 68, 37, 38, 39, 40, 41, 40, 40, 40, 39, 40, 41, 42, 44, 45, 47, 48,
+        50, 51, 53, 55, 57, 58, 59, 61, 63, 64, 65, 66, 67, 67, 67, 67, 38, 38,
+        39, 40, 40, 41, 41, 41, 41, 42, 43, 44, 45, 47, 48, 50, 51, 53, 54, 56,
+        58, 59, 60, 62, 63, 64, 65, 67, 68, 68, 68, 68, 38, 39, 39, 40, 40, 41,
+        42, 43, 43, 44, 45, 46, 48, 49, 50, 52, 53, 54, 56, 57, 59, 60, 61, 63,
+        64, 65, 66, 67, 68, 68, 68, 68, 39, 39, 39, 39, 40, 41, 43, 44, 46, 47,
+        48, 49, 50, 51, 52, 53, 55, 56, 57, 59, 60, 61, 62, 63, 65, 66, 67, 68,
+        69, 69, 69, 69, 39, 39, 39, 39, 39, 41, 43, 46, 48, 49, 50, 51, 52, 53,
+        54, 55, 56, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 69, 69, 69,
+        41, 41, 41, 40, 40, 42, 44, 47, 49, 50, 51, 53, 54, 55, 56, 57, 58, 59,
+        60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 70, 70, 70, 42, 42, 42, 42,
+        41, 43, 45, 48, 50, 51, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64,
+        65, 66, 67, 68, 69, 70, 71, 71, 71, 71, 44, 44, 43, 43, 42, 44, 46, 49,
+        51, 53, 54, 55, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 66, 67, 68, 69,
+        70, 70, 71, 71, 71, 71, 46, 45, 45, 44, 44, 45, 48, 50, 52, 54, 55, 57,
+        59, 60, 61, 62, 63, 63, 64, 65, 66, 67, 67, 68, 69, 70, 70, 71, 72, 72,
+        72, 72, 48, 47, 46, 46, 45, 47, 49, 51, 53, 55, 56, 58, 60, 61, 62, 63,
+        64, 64, 65, 66, 67, 68, 68, 69, 70, 70, 71, 72, 72, 72, 72, 72, 50, 49,
+        48, 47, 47, 48, 50, 52, 54, 56, 57, 59, 61, 62, 63, 64, 65, 66, 66, 67,
+        68, 69, 69, 70, 71, 71, 72, 72, 73, 73, 73, 73, 52, 51, 50, 49, 48, 50,
+        52, 53, 55, 57, 58, 60, 62, 63, 64, 65, 66, 67, 67, 68, 69, 69, 70, 71,
+        71, 72, 72, 73, 74, 74, 74, 74, 54, 53, 52, 51, 50, 51, 53, 55, 56, 58,
+        59, 61, 63, 64, 65, 66, 67, 68, 68, 69, 70, 70, 71, 72, 72, 73, 73, 74,
+        74, 74, 74, 74, 56, 54, 53, 52, 51, 53, 54, 56, 58, 59, 60, 62, 63, 64,
+        66, 67, 68, 68, 69, 70, 71, 71, 72, 72, 73, 73, 74, 74, 75, 75, 75, 75,
+        57, 56, 55, 54, 53, 54, 56, 57, 59, 60, 61, 63, 64, 65, 66, 67, 68, 69,
+        70, 71, 71, 72, 72, 73, 73, 74, 74, 75, 75, 75, 75, 75, 59, 58, 57, 56,
+        55, 56, 57, 59, 60, 61, 62, 64, 65, 66, 67, 68, 69, 70, 71, 71, 72, 72,
+        73, 73, 74, 74, 75, 75, 76, 76, 76, 76, 61, 60, 59, 58, 57, 58, 59, 60,
+        61, 62, 63, 65, 66, 67, 68, 69, 70, 71, 71, 72, 73, 73, 74, 74, 75, 75,
+        75, 76, 76, 76, 76, 76, 62, 61, 60, 59, 58, 59, 60, 61, 62, 63, 64, 66,
+        67, 68, 69, 69, 70, 71, 72, 72, 73, 74, 74, 75, 75, 75, 76, 76, 76, 76,
+        76, 76, 64, 62, 61, 60, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70,
+        71, 72, 72, 73, 74, 74, 75, 75, 75, 76, 76, 76, 77, 77, 77, 77, 65, 64,
+        63, 62, 61, 62, 63, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 72, 73, 73,
+        74, 75, 75, 75, 76, 76, 77, 77, 77, 77, 77, 77, 66, 65, 64, 63, 63, 63,
+        64, 65, 65, 66, 67, 68, 69, 70, 71, 71, 72, 73, 73, 74, 75, 75, 75, 76,
+        76, 77, 77, 77, 78, 78, 78, 78, 68, 67, 66, 65, 64, 64, 65, 66, 66, 67,
+        68, 69, 70, 70, 71, 72, 73, 73, 74, 74, 75, 75, 76, 76, 77, 77, 77, 78,
+        78, 78, 78, 78, 69, 68, 67, 66, 65, 65, 66, 67, 67, 68, 69, 70, 70, 71,
+        72, 72, 73, 74, 74, 75, 75, 76, 76, 77, 77, 77, 77, 78, 78, 78, 78, 78,
+        70, 69, 68, 67, 66, 67, 67, 68, 68, 69, 70, 70, 71, 72, 72, 73, 74, 74,
+        75, 75, 76, 76, 76, 77, 77, 78, 78, 78, 78, 78, 78, 78, 71, 70, 69, 68,
+        67, 68, 68, 69, 69, 70, 71, 71, 72, 72, 73, 74, 74, 75, 75, 76, 76, 76,
+        77, 77, 78, 78, 78, 78, 79, 79, 79, 79, 71, 70, 69, 68, 67, 68, 68, 69,
+        69, 70, 71, 71, 72, 72, 73, 74, 74, 75, 75, 76, 76, 76, 77, 77, 78, 78,
+        78, 78, 79, 79, 79, 79, 71, 70, 69, 68, 67, 68, 68, 69, 69, 70, 71, 71,
+        72, 72, 73, 74, 74, 75, 75, 76, 76, 76, 77, 77, 78, 78, 78, 78, 79, 79,
+        79, 79, 71, 70, 69, 68, 67, 68, 68, 69, 69, 70, 71, 71, 72, 72, 73, 74,
+        74, 75, 75, 76, 76, 76, 77, 77, 78, 78, 78, 78, 79, 79, 79, 79 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 83, 86, 96, 83, 90, 93, 98, 86, 93, 103, 108, 96, 98, 108, 114,
+        /* Size 8 */
+        64, 56, 79, 81, 84, 88, 93, 97, 56, 70, 80, 76, 77, 81, 85, 90, 79, 80,
+        86, 84, 84, 86, 90, 93, 81, 76, 84, 89, 91, 93, 95, 98, 84, 77, 84, 91,
+        95, 98, 100, 102, 88, 81, 86, 93, 98, 101, 104, 106, 93, 85, 90, 95,
+        100, 104, 106, 108, 97, 90, 93, 98, 102, 106, 108, 111,
+        /* Size 16 */
+        64, 60, 56, 66, 79, 80, 81, 82, 84, 86, 88, 90, 93, 95, 97, 97, 60, 61,
+        62, 70, 79, 79, 78, 79, 80, 82, 84, 86, 89, 91, 94, 94, 56, 62, 70, 74,
+        80, 78, 76, 76, 77, 79, 81, 83, 85, 88, 90, 90, 66, 70, 74, 78, 83, 81,
+        80, 80, 80, 82, 83, 85, 87, 89, 92, 92, 79, 79, 80, 83, 86, 85, 84, 84,
+        84, 85, 86, 88, 90, 91, 93, 93, 80, 79, 78, 81, 85, 86, 87, 87, 88, 88,
+        89, 91, 92, 94, 95, 95, 81, 78, 76, 80, 84, 87, 89, 90, 91, 92, 93, 94,
+        95, 96, 98, 98, 82, 79, 76, 80, 84, 87, 90, 92, 93, 94, 95, 96, 97, 98,
+        100, 100, 84, 80, 77, 80, 84, 88, 91, 93, 95, 96, 98, 99, 100, 101, 102,
+        102, 86, 82, 79, 82, 85, 88, 92, 94, 96, 98, 99, 101, 102, 103, 104,
+        104, 88, 84, 81, 83, 86, 89, 93, 95, 98, 99, 101, 102, 104, 105, 106,
+        106, 90, 86, 83, 85, 88, 91, 94, 96, 99, 101, 102, 104, 105, 106, 107,
+        107, 93, 89, 85, 87, 90, 92, 95, 97, 100, 102, 104, 105, 106, 107, 108,
+        108, 95, 91, 88, 89, 91, 94, 96, 98, 101, 103, 105, 106, 107, 108, 109,
+        109, 97, 94, 90, 92, 93, 95, 98, 100, 102, 104, 106, 107, 108, 109, 111,
+        111, 97, 94, 90, 92, 93, 95, 98, 100, 102, 104, 106, 107, 108, 109, 111,
+        111,
+        /* Size 32 */
+        64, 62, 60, 58, 56, 60, 66, 72, 79, 79, 80, 81, 81, 82, 82, 83, 84, 85,
+        86, 87, 88, 89, 90, 92, 93, 94, 95, 96, 97, 97, 97, 97, 62, 61, 60, 60,
+        59, 63, 68, 73, 79, 79, 79, 80, 80, 80, 81, 81, 82, 83, 84, 85, 86, 87,
+        88, 90, 91, 92, 93, 94, 95, 95, 95, 95, 60, 60, 61, 62, 62, 66, 70, 74,
+        79, 79, 79, 79, 78, 79, 79, 80, 80, 81, 82, 83, 84, 85, 86, 88, 89, 90,
+        91, 92, 94, 94, 94, 94, 58, 60, 62, 64, 66, 69, 72, 75, 79, 79, 78, 78,
+        77, 77, 78, 78, 78, 79, 80, 81, 82, 83, 85, 86, 87, 88, 89, 91, 92, 92,
+        92, 92, 56, 59, 62, 66, 70, 72, 74, 77, 80, 79, 78, 77, 76, 76, 76, 77,
+        77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 88, 89, 90, 90, 90, 90, 60, 63,
+        66, 69, 72, 74, 76, 79, 81, 80, 79, 78, 78, 78, 78, 78, 79, 79, 80, 81,
+        82, 83, 84, 85, 86, 87, 89, 90, 91, 91, 91, 91, 66, 68, 70, 72, 74, 76,
+        78, 80, 83, 82, 81, 80, 80, 80, 80, 80, 80, 81, 82, 83, 83, 84, 85, 86,
+        87, 88, 89, 91, 92, 92, 92, 92, 72, 73, 74, 75, 77, 79, 80, 82, 84, 84,
+        83, 83, 82, 82, 82, 82, 82, 83, 83, 84, 85, 86, 87, 87, 88, 89, 90, 91,
+        92, 92, 92, 92, 79, 79, 79, 79, 80, 81, 83, 84, 86, 85, 85, 85, 84, 84,
+        84, 84, 84, 85, 85, 86, 86, 87, 88, 89, 90, 90, 91, 92, 93, 93, 93, 93,
+        79, 79, 79, 79, 79, 80, 82, 84, 85, 85, 85, 85, 86, 86, 86, 86, 86, 86,
+        87, 87, 88, 88, 89, 90, 91, 92, 93, 93, 94, 94, 94, 94, 80, 79, 79, 78,
+        78, 79, 81, 83, 85, 85, 86, 86, 87, 87, 87, 87, 88, 88, 88, 89, 89, 90,
+        91, 91, 92, 93, 94, 95, 95, 95, 95, 95, 81, 80, 79, 78, 77, 78, 80, 83,
+        85, 85, 86, 87, 88, 88, 89, 89, 89, 90, 90, 91, 91, 92, 92, 93, 93, 94,
+        95, 96, 96, 96, 96, 96, 81, 80, 78, 77, 76, 78, 80, 82, 84, 86, 87, 88,
+        89, 90, 90, 91, 91, 91, 92, 92, 93, 93, 94, 94, 95, 96, 96, 97, 98, 98,
+        98, 98, 82, 80, 79, 77, 76, 78, 80, 82, 84, 86, 87, 88, 90, 90, 91, 91,
+        92, 92, 93, 93, 94, 94, 95, 95, 96, 97, 97, 98, 99, 99, 99, 99, 82, 81,
+        79, 78, 76, 78, 80, 82, 84, 86, 87, 89, 90, 91, 92, 92, 93, 94, 94, 95,
+        95, 96, 96, 97, 97, 98, 98, 99, 100, 100, 100, 100, 83, 81, 80, 78, 77,
+        78, 80, 82, 84, 86, 87, 89, 91, 91, 92, 93, 94, 95, 95, 96, 96, 97, 97,
+        98, 99, 99, 100, 100, 101, 101, 101, 101, 84, 82, 80, 78, 77, 79, 80,
+        82, 84, 86, 88, 89, 91, 92, 93, 94, 95, 96, 96, 97, 98, 98, 99, 99, 100,
+        100, 101, 101, 102, 102, 102, 102, 85, 83, 81, 79, 78, 79, 81, 83, 85,
+        86, 88, 90, 91, 92, 94, 95, 96, 96, 97, 98, 99, 99, 100, 100, 101, 101,
+        102, 102, 103, 103, 103, 103, 86, 84, 82, 80, 79, 80, 82, 83, 85, 87,
+        88, 90, 92, 93, 94, 95, 96, 97, 98, 99, 99, 100, 101, 101, 102, 102,
+        103, 103, 104, 104, 104, 104, 87, 85, 83, 81, 80, 81, 83, 84, 86, 87,
+        89, 91, 92, 93, 95, 96, 97, 98, 99, 99, 100, 101, 101, 102, 103, 103,
+        104, 104, 105, 105, 105, 105, 88, 86, 84, 82, 81, 82, 83, 85, 86, 88,
+        89, 91, 93, 94, 95, 96, 98, 99, 99, 100, 101, 102, 102, 103, 104, 104,
+        105, 105, 106, 106, 106, 106, 89, 87, 85, 83, 82, 83, 84, 86, 87, 88,
+        90, 92, 93, 94, 96, 97, 98, 99, 100, 101, 102, 102, 103, 104, 104, 105,
+        105, 106, 106, 106, 106, 106, 90, 88, 86, 85, 83, 84, 85, 87, 88, 89,
+        91, 92, 94, 95, 96, 97, 99, 100, 101, 101, 102, 103, 104, 104, 105, 105,
+        106, 106, 107, 107, 107, 107, 92, 90, 88, 86, 84, 85, 86, 87, 89, 90,
+        91, 93, 94, 95, 97, 98, 99, 100, 101, 102, 103, 104, 104, 105, 106, 106,
+        107, 107, 108, 108, 108, 108, 93, 91, 89, 87, 85, 86, 87, 88, 90, 91,
+        92, 93, 95, 96, 97, 99, 100, 101, 102, 103, 104, 104, 105, 106, 106,
+        107, 107, 108, 108, 108, 108, 108, 94, 92, 90, 88, 86, 87, 88, 89, 90,
+        92, 93, 94, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 105, 106, 107,
+        107, 108, 108, 109, 109, 109, 109, 95, 93, 91, 89, 88, 89, 89, 90, 91,
+        93, 94, 95, 96, 97, 98, 100, 101, 102, 103, 104, 105, 105, 106, 107,
+        107, 108, 108, 109, 109, 109, 109, 109, 96, 94, 92, 91, 89, 90, 91, 91,
+        92, 93, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 106, 107,
+        108, 108, 109, 109, 110, 110, 110, 110, 97, 95, 94, 92, 90, 91, 92, 92,
+        93, 94, 95, 96, 98, 99, 100, 101, 102, 103, 104, 105, 106, 106, 107,
+        108, 108, 109, 109, 110, 111, 111, 111, 111, 97, 95, 94, 92, 90, 91, 92,
+        92, 93, 94, 95, 96, 98, 99, 100, 101, 102, 103, 104, 105, 106, 106, 107,
+        108, 108, 109, 109, 110, 111, 111, 111, 111, 97, 95, 94, 92, 90, 91, 92,
+        92, 93, 94, 95, 96, 98, 99, 100, 101, 102, 103, 104, 105, 106, 106, 107,
+        108, 108, 109, 109, 110, 111, 111, 111, 111, 97, 95, 94, 92, 90, 91, 92,
+        92, 93, 94, 95, 96, 98, 99, 100, 101, 102, 103, 104, 105, 106, 106, 107,
+        108, 108, 109, 109, 110, 111, 111, 111, 111 },
+      { /* Intra matrices */
+        /* Size 4 */
+        42, 56, 58, 65, 56, 61, 63, 67, 58, 63, 70, 74, 65, 67, 74, 78,
+        /* Size 8 */
+        45, 39, 55, 57, 59, 62, 66, 69, 39, 49, 56, 53, 54, 57, 60, 64, 55, 56,
+        61, 60, 60, 61, 63, 66, 57, 53, 60, 63, 65, 66, 68, 70, 59, 54, 60, 65,
+        68, 70, 71, 73, 62, 57, 61, 66, 70, 72, 74, 76, 66, 60, 63, 68, 71, 74,
+        76, 78, 69, 64, 66, 70, 73, 76, 78, 80,
+        /* Size 16 */
+        44, 41, 38, 45, 55, 56, 56, 57, 58, 60, 62, 63, 65, 67, 69, 69, 41, 42,
+        43, 48, 55, 55, 54, 55, 56, 57, 59, 60, 62, 64, 66, 66, 38, 43, 48, 52,
+        55, 54, 52, 53, 53, 55, 56, 58, 59, 61, 63, 63, 45, 48, 52, 54, 57, 56,
+        55, 56, 56, 57, 58, 60, 61, 63, 64, 64, 55, 55, 55, 57, 60, 59, 59, 59,
+        59, 59, 60, 61, 63, 64, 66, 66, 56, 55, 54, 56, 59, 60, 61, 61, 61, 62,
+        63, 64, 65, 66, 67, 67, 56, 54, 52, 55, 59, 61, 62, 63, 64, 64, 65, 66,
+        67, 68, 69, 69, 57, 55, 53, 56, 59, 61, 63, 64, 65, 66, 67, 68, 69, 69,
+        70, 70, 58, 56, 53, 56, 59, 61, 64, 65, 67, 68, 69, 70, 70, 71, 72, 72,
+        60, 57, 55, 57, 59, 62, 64, 66, 68, 69, 70, 71, 72, 73, 73, 73, 62, 59,
+        56, 58, 60, 63, 65, 67, 69, 70, 71, 72, 73, 74, 75, 75, 63, 60, 58, 60,
+        61, 64, 66, 68, 70, 71, 72, 73, 74, 75, 76, 76, 65, 62, 59, 61, 63, 65,
+        67, 69, 70, 72, 73, 74, 75, 76, 77, 77, 67, 64, 61, 63, 64, 66, 68, 69,
+        71, 73, 74, 75, 76, 77, 78, 78, 69, 66, 63, 64, 66, 67, 69, 70, 72, 73,
+        75, 76, 77, 78, 79, 79, 69, 66, 63, 64, 66, 67, 69, 70, 72, 73, 75, 76,
+        77, 78, 79, 79,
+        /* Size 32 */
+        44, 42, 41, 39, 38, 41, 45, 49, 54, 55, 55, 56, 56, 57, 57, 58, 58, 59,
+        60, 60, 61, 62, 63, 64, 65, 66, 66, 67, 68, 68, 68, 68, 42, 42, 41, 41,
+        40, 43, 46, 50, 55, 55, 55, 55, 55, 55, 56, 56, 57, 57, 58, 59, 60, 61,
+        61, 62, 63, 64, 65, 66, 67, 67, 67, 67, 41, 41, 42, 42, 42, 45, 48, 51,
+        55, 55, 54, 54, 54, 54, 55, 55, 55, 56, 57, 58, 58, 59, 60, 61, 62, 63,
+        63, 64, 65, 65, 65, 65, 39, 41, 42, 43, 45, 47, 49, 52, 55, 54, 54, 53,
+        53, 53, 54, 54, 54, 55, 55, 56, 57, 58, 59, 59, 60, 61, 62, 63, 64, 64,
+        64, 64, 38, 40, 42, 45, 48, 49, 51, 53, 55, 54, 53, 53, 52, 52, 53, 53,
+        53, 54, 54, 55, 56, 56, 57, 58, 59, 60, 61, 62, 63, 63, 63, 63, 41, 43,
+        45, 47, 49, 51, 53, 54, 56, 55, 55, 54, 54, 54, 54, 54, 54, 55, 55, 56,
+        57, 57, 58, 59, 60, 61, 62, 62, 63, 63, 63, 63, 45, 46, 48, 49, 51, 53,
+        54, 56, 57, 57, 56, 56, 55, 55, 55, 55, 56, 56, 57, 57, 58, 58, 59, 60,
+        61, 61, 62, 63, 64, 64, 64, 64, 49, 50, 51, 52, 53, 54, 56, 57, 58, 58,
+        57, 57, 57, 57, 57, 57, 57, 57, 58, 58, 59, 59, 60, 61, 61, 62, 63, 64,
+        64, 64, 64, 64, 54, 55, 55, 55, 55, 56, 57, 58, 59, 59, 59, 59, 58, 58,
+        58, 58, 58, 59, 59, 59, 60, 60, 61, 62, 62, 63, 64, 64, 65, 65, 65, 65,
+        55, 55, 55, 54, 54, 55, 57, 58, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60,
+        60, 61, 61, 62, 62, 63, 63, 64, 65, 65, 66, 66, 66, 66, 55, 55, 54, 54,
+        53, 55, 56, 57, 59, 59, 60, 60, 60, 60, 60, 61, 61, 61, 61, 62, 62, 63,
+        63, 64, 64, 65, 65, 66, 67, 67, 67, 67, 56, 55, 54, 53, 53, 54, 56, 57,
+        59, 59, 60, 60, 61, 61, 62, 62, 62, 62, 63, 63, 63, 64, 64, 65, 65, 66,
+        66, 67, 67, 67, 67, 67, 56, 55, 54, 53, 52, 54, 55, 57, 58, 59, 60, 61,
+        62, 62, 63, 63, 63, 64, 64, 64, 65, 65, 65, 66, 66, 67, 67, 68, 68, 68,
+        68, 68, 57, 55, 54, 53, 52, 54, 55, 57, 58, 59, 60, 61, 62, 63, 63, 64,
+        64, 64, 65, 65, 66, 66, 66, 67, 67, 68, 68, 69, 69, 69, 69, 69, 57, 56,
+        55, 54, 53, 54, 55, 57, 58, 59, 60, 62, 63, 63, 64, 64, 65, 65, 66, 66,
+        66, 67, 67, 68, 68, 69, 69, 69, 70, 70, 70, 70, 58, 56, 55, 54, 53, 54,
+        55, 57, 58, 59, 61, 62, 63, 64, 64, 65, 66, 66, 67, 67, 67, 68, 68, 69,
+        69, 69, 70, 70, 71, 71, 71, 71, 58, 57, 55, 54, 53, 54, 56, 57, 58, 60,
+        61, 62, 63, 64, 65, 66, 66, 67, 67, 68, 68, 69, 69, 70, 70, 70, 71, 71,
+        72, 72, 72, 72, 59, 57, 56, 55, 54, 55, 56, 57, 59, 60, 61, 62, 64, 64,
+        65, 66, 67, 67, 68, 68, 69, 69, 70, 70, 71, 71, 71, 72, 72, 72, 72, 72,
+        60, 58, 57, 55, 54, 55, 57, 58, 59, 60, 61, 63, 64, 65, 66, 67, 67, 68,
+        69, 69, 70, 70, 71, 71, 71, 72, 72, 73, 73, 73, 73, 73, 60, 59, 58, 56,
+        55, 56, 57, 58, 59, 61, 62, 63, 64, 65, 66, 67, 68, 68, 69, 70, 70, 71,
+        71, 72, 72, 72, 73, 73, 74, 74, 74, 74, 61, 60, 58, 57, 56, 57, 58, 59,
+        60, 61, 62, 63, 65, 66, 66, 67, 68, 69, 70, 70, 71, 71, 72, 72, 73, 73,
+        74, 74, 74, 74, 74, 74, 62, 61, 59, 58, 56, 57, 58, 59, 60, 62, 63, 64,
+        65, 66, 67, 68, 69, 69, 70, 71, 71, 72, 72, 73, 73, 74, 74, 74, 75, 75,
+        75, 75, 63, 61, 60, 59, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68,
+        69, 70, 71, 71, 72, 72, 73, 73, 74, 74, 75, 75, 75, 75, 75, 75, 64, 62,
+        61, 59, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 70, 71, 72,
+        72, 73, 73, 74, 74, 75, 75, 76, 76, 76, 76, 76, 65, 63, 62, 60, 59, 60,
+        61, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 71, 72, 73, 73, 74, 74,
+        75, 75, 76, 76, 76, 76, 76, 76, 66, 64, 63, 61, 60, 61, 61, 62, 63, 64,
+        65, 66, 67, 68, 69, 69, 70, 71, 72, 72, 73, 74, 74, 75, 75, 76, 76, 76,
+        77, 77, 77, 77, 66, 65, 63, 62, 61, 62, 62, 63, 64, 65, 65, 66, 67, 68,
+        69, 70, 71, 71, 72, 73, 74, 74, 75, 75, 76, 76, 76, 77, 77, 77, 77, 77,
+        67, 66, 64, 63, 62, 62, 63, 64, 64, 65, 66, 67, 68, 69, 69, 70, 71, 72,
+        73, 73, 74, 74, 75, 76, 76, 76, 77, 77, 78, 78, 78, 78, 68, 67, 65, 64,
+        63, 63, 64, 64, 65, 66, 67, 67, 68, 69, 70, 71, 72, 72, 73, 74, 74, 75,
+        75, 76, 76, 77, 77, 78, 78, 78, 78, 78, 68, 67, 65, 64, 63, 63, 64, 64,
+        65, 66, 67, 67, 68, 69, 70, 71, 72, 72, 73, 74, 74, 75, 75, 76, 76, 77,
+        77, 78, 78, 78, 78, 78, 68, 67, 65, 64, 63, 63, 64, 64, 65, 66, 67, 67,
+        68, 69, 70, 71, 72, 72, 73, 74, 74, 75, 75, 76, 76, 77, 77, 78, 78, 78,
+        78, 78, 68, 67, 65, 64, 63, 63, 64, 64, 65, 66, 67, 67, 68, 69, 70, 71,
+        72, 72, 73, 74, 74, 75, 75, 76, 76, 77, 77, 78, 78, 78, 78, 78 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 67, 84, 97, 67, 81, 91, 99, 84, 91, 100, 105, 97, 99, 105, 108,
+        /* Size 8 */
+        64, 55, 58, 66, 75, 82, 88, 92, 55, 59, 58, 63, 70, 78, 84, 89, 58, 58,
+        68, 73, 78, 82, 87, 91, 66, 63, 73, 80, 84, 87, 90, 93, 75, 70, 78, 84,
+        88, 91, 93, 95, 82, 78, 82, 87, 91, 94, 96, 97, 88, 84, 87, 90, 93, 96,
+        97, 98, 92, 89, 91, 93, 95, 97, 98, 99,
+        /* Size 16 */
+        64, 59, 55, 56, 58, 62, 66, 70, 75, 78, 82, 85, 88, 90, 92, 92, 59, 58,
+        57, 58, 58, 61, 65, 68, 73, 76, 80, 83, 86, 88, 90, 90, 55, 57, 59, 59,
+        58, 60, 63, 67, 70, 74, 78, 81, 84, 86, 89, 89, 56, 58, 59, 61, 63, 65,
+        68, 71, 74, 77, 80, 83, 85, 87, 90, 90, 58, 58, 58, 63, 68, 71, 73, 75,
+        78, 80, 82, 85, 87, 89, 91, 91, 62, 61, 60, 65, 71, 73, 76, 78, 81, 83,
+        85, 87, 89, 90, 92, 92, 66, 65, 63, 68, 73, 76, 80, 82, 84, 86, 87, 89,
+        90, 92, 93, 93, 70, 68, 67, 71, 75, 78, 82, 84, 86, 88, 89, 91, 92, 93,
+        94, 94, 75, 73, 70, 74, 78, 81, 84, 86, 88, 90, 91, 92, 93, 94, 95, 95,
+        78, 76, 74, 77, 80, 83, 86, 88, 90, 91, 92, 93, 94, 95, 96, 96, 82, 80,
+        78, 80, 82, 85, 87, 89, 91, 92, 94, 95, 96, 96, 97, 97, 85, 83, 81, 83,
+        85, 87, 89, 91, 92, 93, 95, 96, 96, 97, 98, 98, 88, 86, 84, 85, 87, 89,
+        90, 92, 93, 94, 96, 96, 97, 98, 98, 98, 90, 88, 86, 87, 89, 90, 92, 93,
+        94, 95, 96, 97, 98, 98, 99, 99, 92, 90, 89, 90, 91, 92, 93, 94, 95, 96,
+        97, 98, 98, 99, 99, 99, 92, 90, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98,
+        98, 99, 99, 99,
+        /* Size 32 */
+        64, 62, 59, 57, 55, 56, 56, 57, 58, 60, 62, 64, 66, 68, 70, 72, 75, 77,
+        78, 80, 82, 84, 85, 86, 88, 89, 90, 91, 92, 92, 92, 92, 62, 60, 59, 58,
+        56, 57, 57, 57, 58, 60, 61, 63, 65, 67, 69, 71, 74, 75, 77, 79, 81, 82,
+        84, 85, 87, 88, 89, 90, 91, 91, 91, 91, 59, 59, 58, 58, 57, 57, 58, 58,
+        58, 59, 61, 63, 65, 66, 68, 70, 73, 74, 76, 78, 80, 81, 83, 84, 86, 87,
+        88, 89, 90, 90, 90, 90, 57, 58, 58, 58, 58, 58, 58, 58, 58, 59, 61, 62,
+        64, 66, 67, 69, 72, 73, 75, 77, 79, 80, 82, 83, 85, 86, 87, 88, 90, 90,
+        90, 90, 55, 56, 57, 58, 59, 59, 59, 58, 58, 59, 60, 62, 63, 65, 67, 68,
+        70, 72, 74, 76, 78, 79, 81, 82, 84, 85, 86, 87, 89, 89, 89, 89, 56, 57,
+        57, 58, 59, 59, 60, 60, 60, 62, 63, 64, 65, 67, 69, 70, 72, 74, 75, 77,
+        79, 80, 82, 83, 85, 86, 87, 88, 89, 89, 89, 89, 56, 57, 58, 58, 59, 60,
+        61, 62, 63, 64, 65, 66, 68, 69, 71, 72, 74, 75, 77, 78, 80, 81, 83, 84,
+        85, 86, 87, 89, 90, 90, 90, 90, 57, 57, 58, 58, 58, 60, 62, 64, 66, 67,
+        68, 69, 70, 72, 73, 74, 76, 77, 78, 80, 81, 82, 84, 85, 86, 87, 88, 89,
+        90, 90, 90, 90, 58, 58, 58, 58, 58, 60, 63, 66, 68, 70, 71, 72, 73, 74,
+        75, 76, 78, 79, 80, 81, 82, 83, 85, 86, 87, 88, 89, 90, 91, 91, 91, 91,
+        60, 60, 59, 59, 59, 62, 64, 67, 70, 71, 72, 73, 75, 76, 77, 78, 79, 80,
+        81, 82, 84, 85, 86, 87, 88, 89, 89, 90, 91, 91, 91, 91, 62, 61, 61, 61,
+        60, 63, 65, 68, 71, 72, 73, 75, 76, 77, 78, 80, 81, 82, 83, 84, 85, 86,
+        87, 88, 89, 89, 90, 91, 92, 92, 92, 92, 64, 63, 63, 62, 62, 64, 66, 69,
+        72, 73, 75, 76, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 89, 90,
+        91, 92, 92, 92, 92, 92, 66, 65, 65, 64, 63, 65, 68, 70, 73, 75, 76, 78,
+        80, 81, 82, 83, 84, 85, 86, 87, 87, 88, 89, 90, 90, 91, 92, 92, 93, 93,
+        93, 93, 68, 67, 66, 66, 65, 67, 69, 72, 74, 76, 77, 79, 81, 82, 83, 84,
+        85, 86, 87, 87, 88, 89, 90, 90, 91, 92, 92, 93, 94, 94, 94, 94, 70, 69,
+        68, 67, 67, 69, 71, 73, 75, 77, 78, 80, 82, 83, 84, 85, 86, 87, 88, 88,
+        89, 90, 91, 91, 92, 92, 93, 94, 94, 94, 94, 94, 72, 71, 70, 69, 68, 70,
+        72, 74, 76, 78, 80, 81, 83, 84, 85, 86, 87, 88, 89, 89, 90, 91, 91, 92,
+        93, 93, 94, 94, 95, 95, 95, 95, 75, 74, 73, 72, 70, 72, 74, 76, 78, 79,
+        81, 82, 84, 85, 86, 87, 88, 89, 90, 90, 91, 92, 92, 93, 93, 94, 94, 95,
+        95, 95, 95, 95, 77, 75, 74, 73, 72, 74, 75, 77, 79, 80, 82, 83, 85, 86,
+        87, 88, 89, 90, 90, 91, 92, 92, 93, 93, 94, 94, 95, 95, 96, 96, 96, 96,
+        78, 77, 76, 75, 74, 75, 77, 78, 80, 81, 83, 84, 86, 87, 88, 89, 90, 90,
+        91, 92, 92, 93, 93, 94, 94, 95, 95, 96, 96, 96, 96, 96, 80, 79, 78, 77,
+        76, 77, 78, 80, 81, 82, 84, 85, 87, 87, 88, 89, 90, 91, 92, 92, 93, 94,
+        94, 95, 95, 95, 96, 96, 97, 97, 97, 97, 82, 81, 80, 79, 78, 79, 80, 81,
+        82, 84, 85, 86, 87, 88, 89, 90, 91, 92, 92, 93, 94, 94, 95, 95, 96, 96,
+        96, 97, 97, 97, 97, 97, 84, 82, 81, 80, 79, 80, 81, 82, 83, 85, 86, 87,
+        88, 89, 90, 91, 92, 92, 93, 94, 94, 95, 95, 96, 96, 96, 97, 97, 97, 97,
+        97, 97, 85, 84, 83, 82, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 91,
+        92, 93, 93, 94, 95, 95, 96, 96, 96, 97, 97, 97, 98, 98, 98, 98, 86, 85,
+        84, 83, 82, 83, 84, 85, 86, 87, 88, 89, 90, 90, 91, 92, 93, 93, 94, 95,
+        95, 96, 96, 96, 97, 97, 97, 98, 98, 98, 98, 98, 88, 87, 86, 85, 84, 85,
+        85, 86, 87, 88, 89, 89, 90, 91, 92, 93, 93, 94, 94, 95, 96, 96, 96, 97,
+        97, 97, 98, 98, 98, 98, 98, 98, 89, 88, 87, 86, 85, 86, 86, 87, 88, 89,
+        89, 90, 91, 92, 92, 93, 94, 94, 95, 95, 96, 96, 97, 97, 97, 98, 98, 98,
+        99, 99, 99, 99, 90, 89, 88, 87, 86, 87, 87, 88, 89, 89, 90, 91, 92, 92,
+        93, 94, 94, 95, 95, 96, 96, 97, 97, 97, 98, 98, 98, 99, 99, 99, 99, 99,
+        91, 90, 89, 88, 87, 88, 89, 89, 90, 90, 91, 92, 92, 93, 94, 94, 95, 95,
+        96, 96, 97, 97, 97, 98, 98, 98, 99, 99, 99, 99, 99, 99, 92, 91, 90, 90,
+        89, 89, 90, 90, 91, 91, 92, 92, 93, 94, 94, 95, 95, 96, 96, 97, 97, 97,
+        98, 98, 98, 99, 99, 99, 99, 99, 99, 99, 92, 91, 90, 90, 89, 89, 90, 90,
+        91, 91, 92, 92, 93, 94, 94, 95, 95, 96, 96, 97, 97, 97, 98, 98, 98, 99,
+        99, 99, 99, 99, 99, 99, 92, 91, 90, 90, 89, 89, 90, 90, 91, 91, 92, 92,
+        93, 94, 94, 95, 95, 96, 96, 97, 97, 97, 98, 98, 98, 99, 99, 99, 99, 99,
+        99, 99, 92, 91, 90, 90, 89, 89, 90, 90, 91, 91, 92, 92, 93, 94, 94, 95,
+        95, 96, 96, 97, 97, 97, 98, 98, 98, 99, 99, 99, 99, 99, 99, 99 },
+      { /* Intra matrices */
+        /* Size 4 */
+        44, 46, 58, 68, 46, 56, 64, 70, 58, 64, 71, 74, 68, 70, 74, 77,
+        /* Size 8 */
+        49, 42, 44, 50, 57, 63, 68, 71, 42, 45, 44, 48, 54, 60, 65, 69, 44, 44,
+        52, 56, 60, 63, 67, 70, 50, 48, 56, 61, 65, 68, 70, 72, 57, 54, 60, 65,
+        68, 71, 73, 74, 63, 60, 63, 68, 71, 73, 74, 76, 68, 65, 67, 70, 73, 74,
+        76, 77, 71, 69, 70, 72, 74, 76, 77, 78,
+        /* Size 16 */
+        48, 44, 41, 42, 43, 46, 49, 53, 56, 59, 62, 65, 67, 69, 70, 70, 44, 43,
+        43, 43, 43, 45, 48, 51, 55, 57, 61, 63, 65, 67, 69, 69, 41, 43, 44, 44,
+        43, 45, 47, 50, 53, 56, 59, 61, 64, 66, 68, 68, 42, 43, 44, 45, 47, 49,
+        51, 53, 56, 58, 61, 63, 65, 67, 68, 68, 43, 43, 43, 47, 51, 53, 55, 57,
+        59, 61, 63, 64, 66, 68, 69, 69, 46, 45, 45, 49, 53, 55, 58, 59, 61, 63,
+        64, 66, 68, 69, 70, 70, 49, 48, 47, 51, 55, 58, 61, 62, 64, 65, 67, 68,
+        69, 70, 71, 71, 53, 51, 50, 53, 57, 59, 62, 64, 66, 67, 68, 69, 70, 71,
+        72, 72, 56, 55, 53, 56, 59, 61, 64, 66, 67, 69, 70, 71, 72, 72, 73, 73,
+        59, 57, 56, 58, 61, 63, 65, 67, 69, 70, 71, 72, 72, 73, 74, 74, 62, 61,
+        59, 61, 63, 64, 67, 68, 70, 71, 72, 73, 73, 74, 75, 75, 65, 63, 61, 63,
+        64, 66, 68, 69, 71, 72, 73, 73, 74, 75, 75, 75, 67, 65, 64, 65, 66, 68,
+        69, 70, 72, 72, 73, 74, 75, 75, 76, 76, 69, 67, 66, 67, 68, 69, 70, 71,
+        72, 73, 74, 75, 75, 76, 76, 76, 70, 69, 68, 68, 69, 70, 71, 72, 73, 74,
+        75, 75, 76, 76, 76, 76, 70, 69, 68, 68, 69, 70, 71, 72, 73, 74, 75, 75,
+        76, 76, 76, 76,
+        /* Size 32 */
+        47, 46, 44, 42, 41, 41, 42, 42, 43, 44, 46, 47, 49, 51, 52, 54, 56, 57,
+        59, 60, 62, 63, 64, 65, 66, 67, 68, 69, 70, 70, 70, 70, 46, 44, 43, 42,
+        41, 42, 42, 42, 43, 44, 45, 47, 48, 50, 52, 53, 55, 56, 58, 59, 61, 62,
+        63, 64, 66, 66, 67, 68, 69, 69, 69, 69, 44, 43, 43, 43, 42, 42, 42, 43,
+        43, 44, 45, 46, 48, 49, 51, 52, 54, 56, 57, 58, 60, 61, 62, 64, 65, 66,
+        67, 68, 68, 68, 68, 68, 42, 42, 43, 43, 43, 43, 43, 43, 43, 44, 45, 46,
+        47, 49, 50, 52, 53, 55, 56, 58, 59, 60, 61, 63, 64, 65, 66, 67, 68, 68,
+        68, 68, 41, 41, 42, 43, 44, 44, 43, 43, 43, 44, 45, 46, 47, 48, 49, 51,
+        53, 54, 55, 57, 58, 59, 61, 62, 63, 64, 65, 66, 67, 67, 67, 67, 41, 42,
+        42, 43, 44, 44, 44, 44, 45, 46, 46, 47, 48, 50, 51, 52, 54, 55, 56, 58,
+        59, 60, 61, 63, 64, 65, 66, 67, 68, 68, 68, 68, 42, 42, 42, 43, 43, 44,
+        45, 46, 47, 47, 48, 49, 50, 51, 53, 54, 55, 56, 58, 59, 60, 61, 62, 63,
+        64, 65, 66, 67, 68, 68, 68, 68, 42, 42, 43, 43, 43, 44, 46, 47, 49, 50,
+        50, 51, 52, 53, 54, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 67,
+        68, 68, 68, 68, 43, 43, 43, 43, 43, 45, 47, 49, 51, 52, 53, 54, 55, 55,
+        56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 66, 67, 68, 69, 69, 69, 69,
+        44, 44, 44, 44, 44, 46, 47, 50, 52, 53, 54, 55, 56, 57, 58, 58, 59, 60,
+        61, 62, 63, 64, 65, 65, 66, 67, 68, 68, 69, 69, 69, 69, 46, 45, 45, 45,
+        45, 46, 48, 50, 53, 54, 55, 56, 57, 58, 59, 60, 61, 61, 62, 63, 64, 65,
+        65, 66, 67, 68, 68, 69, 70, 70, 70, 70, 47, 47, 46, 46, 46, 47, 49, 51,
+        54, 55, 56, 57, 59, 59, 60, 61, 62, 63, 63, 64, 65, 66, 66, 67, 68, 68,
+        69, 70, 70, 70, 70, 70, 49, 48, 48, 47, 47, 48, 50, 52, 55, 56, 57, 59,
+        60, 61, 62, 63, 63, 64, 65, 65, 66, 67, 67, 68, 69, 69, 70, 70, 71, 71,
+        71, 71, 51, 50, 49, 49, 48, 50, 51, 53, 55, 57, 58, 59, 61, 62, 63, 63,
+        64, 65, 65, 66, 67, 67, 68, 69, 69, 70, 70, 71, 71, 71, 71, 71, 52, 52,
+        51, 50, 49, 51, 53, 54, 56, 58, 59, 60, 62, 63, 63, 64, 65, 66, 66, 67,
+        68, 68, 69, 69, 70, 70, 71, 71, 72, 72, 72, 72, 54, 53, 52, 52, 51, 52,
+        54, 56, 57, 58, 60, 61, 63, 63, 64, 65, 66, 67, 67, 68, 68, 69, 69, 70,
+        70, 71, 71, 72, 72, 72, 72, 72, 56, 55, 54, 53, 53, 54, 55, 57, 58, 59,
+        61, 62, 63, 64, 65, 66, 67, 67, 68, 69, 69, 70, 70, 71, 71, 71, 72, 72,
+        72, 72, 72, 72, 57, 56, 56, 55, 54, 55, 56, 58, 59, 60, 61, 63, 64, 65,
+        66, 67, 67, 68, 69, 69, 70, 70, 71, 71, 71, 72, 72, 72, 73, 73, 73, 73,
+        59, 58, 57, 56, 55, 56, 58, 59, 60, 61, 62, 63, 65, 65, 66, 67, 68, 69,
+        69, 70, 70, 71, 71, 71, 72, 72, 73, 73, 73, 73, 73, 73, 60, 59, 58, 58,
+        57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 69, 70, 70, 71, 71,
+        72, 72, 72, 73, 73, 73, 74, 74, 74, 74, 62, 61, 60, 59, 58, 59, 60, 61,
+        62, 63, 64, 65, 66, 67, 68, 68, 69, 70, 70, 71, 71, 72, 72, 72, 73, 73,
+        73, 74, 74, 74, 74, 74, 63, 62, 61, 60, 59, 60, 61, 62, 63, 64, 65, 66,
+        67, 67, 68, 69, 70, 70, 71, 71, 72, 72, 72, 73, 73, 73, 74, 74, 74, 74,
+        74, 74, 64, 63, 62, 61, 61, 61, 62, 63, 64, 65, 65, 66, 67, 68, 69, 69,
+        70, 71, 71, 72, 72, 72, 73, 73, 73, 74, 74, 74, 75, 75, 75, 75, 65, 64,
+        64, 63, 62, 63, 63, 64, 65, 65, 66, 67, 68, 69, 69, 70, 71, 71, 71, 72,
+        72, 73, 73, 73, 74, 74, 74, 75, 75, 75, 75, 75, 66, 66, 65, 64, 63, 64,
+        64, 65, 66, 66, 67, 68, 69, 69, 70, 70, 71, 71, 72, 72, 73, 73, 73, 74,
+        74, 74, 75, 75, 75, 75, 75, 75, 67, 66, 66, 65, 64, 65, 65, 66, 66, 67,
+        68, 68, 69, 70, 70, 71, 71, 72, 72, 73, 73, 73, 74, 74, 74, 75, 75, 75,
+        75, 75, 75, 75, 68, 67, 67, 66, 65, 66, 66, 67, 67, 68, 68, 69, 70, 70,
+        71, 71, 72, 72, 73, 73, 73, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 75,
+        69, 68, 68, 67, 66, 67, 67, 67, 68, 68, 69, 70, 70, 71, 71, 72, 72, 72,
+        73, 73, 74, 74, 74, 75, 75, 75, 75, 75, 76, 76, 76, 76, 70, 69, 68, 68,
+        67, 68, 68, 68, 69, 69, 70, 70, 71, 71, 72, 72, 72, 73, 73, 74, 74, 74,
+        75, 75, 75, 75, 75, 76, 76, 76, 76, 76, 70, 69, 68, 68, 67, 68, 68, 68,
+        69, 69, 70, 70, 71, 71, 72, 72, 72, 73, 73, 74, 74, 74, 75, 75, 75, 75,
+        75, 76, 76, 76, 76, 76, 70, 69, 68, 68, 67, 68, 68, 68, 69, 69, 70, 70,
+        71, 71, 72, 72, 72, 73, 73, 74, 74, 74, 75, 75, 75, 75, 75, 76, 76, 76,
+        76, 76, 70, 69, 68, 68, 67, 68, 68, 68, 69, 69, 70, 70, 71, 71, 72, 72,
+        72, 73, 73, 74, 74, 74, 75, 75, 75, 75, 75, 76, 76, 76, 76, 76 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 80, 83, 90, 80, 86, 88, 92, 83, 88, 96, 99, 90, 92, 99, 104,
+        /* Size 8 */
+        64, 57, 77, 78, 80, 84, 88, 91, 57, 69, 77, 74, 75, 78, 82, 86, 77, 77,
+        82, 81, 81, 82, 85, 88, 78, 74, 81, 85, 86, 87, 89, 91, 80, 75, 81, 86,
+        89, 91, 93, 95, 84, 78, 82, 87, 91, 94, 96, 97, 88, 82, 85, 89, 93, 96,
+        98, 99, 91, 86, 88, 91, 95, 97, 99, 101,
+        /* Size 16 */
+        64, 60, 57, 65, 77, 77, 78, 79, 80, 82, 84, 86, 88, 89, 91, 91, 60, 61,
+        62, 69, 77, 76, 76, 77, 78, 79, 81, 83, 85, 86, 88, 88, 57, 62, 69, 73,
+        77, 75, 74, 74, 75, 76, 78, 80, 82, 84, 86, 86, 65, 69, 73, 76, 80, 78,
+        77, 78, 78, 79, 80, 82, 83, 85, 87, 87, 77, 77, 77, 80, 82, 82, 81, 81,
+        81, 82, 82, 84, 85, 87, 88, 88, 77, 76, 75, 78, 82, 82, 83, 83, 83, 84,
+        85, 86, 87, 88, 90, 90, 78, 76, 74, 77, 81, 83, 85, 85, 86, 87, 87, 88,
+        89, 90, 91, 91, 79, 77, 74, 78, 81, 83, 85, 87, 88, 89, 89, 90, 91, 92,
+        93, 93, 80, 78, 75, 78, 81, 83, 86, 88, 89, 90, 91, 92, 93, 94, 95, 95,
+        82, 79, 76, 79, 82, 84, 87, 89, 90, 92, 93, 94, 94, 95, 96, 96, 84, 81,
+        78, 80, 82, 85, 87, 89, 91, 93, 94, 95, 96, 97, 97, 97, 86, 83, 80, 82,
+        84, 86, 88, 90, 92, 94, 95, 96, 97, 98, 98, 98, 88, 85, 82, 83, 85, 87,
+        89, 91, 93, 94, 96, 97, 98, 99, 99, 99, 89, 86, 84, 85, 87, 88, 90, 92,
+        94, 95, 97, 98, 99, 99, 100, 100, 91, 88, 86, 87, 88, 90, 91, 93, 95,
+        96, 97, 98, 99, 100, 101, 101, 91, 88, 86, 87, 88, 90, 91, 93, 95, 96,
+        97, 98, 99, 100, 101, 101,
+        /* Size 32 */
+        64, 62, 60, 59, 57, 61, 65, 70, 77, 77, 77, 78, 78, 79, 79, 80, 80, 81,
+        82, 83, 84, 85, 86, 87, 88, 88, 89, 90, 91, 91, 91, 91, 62, 61, 61, 60,
+        60, 63, 67, 72, 77, 77, 77, 77, 77, 78, 78, 79, 79, 80, 81, 81, 82, 83,
+        84, 85, 86, 87, 88, 89, 90, 90, 90, 90, 60, 61, 61, 62, 62, 65, 69, 73,
+        77, 77, 76, 76, 76, 76, 77, 77, 78, 78, 79, 80, 81, 82, 83, 84, 85, 85,
+        86, 87, 88, 88, 88, 88, 59, 60, 62, 64, 66, 68, 71, 74, 77, 76, 76, 75,
+        75, 75, 76, 76, 76, 77, 78, 79, 79, 80, 81, 82, 83, 84, 85, 86, 87, 87,
+        87, 87, 57, 60, 62, 66, 69, 71, 73, 75, 77, 76, 75, 75, 74, 74, 74, 75,
+        75, 76, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 86, 86, 86, 61, 63,
+        65, 68, 71, 73, 74, 76, 78, 78, 77, 76, 76, 76, 76, 76, 76, 77, 78, 78,
+        79, 80, 81, 82, 82, 83, 84, 85, 86, 86, 86, 86, 65, 67, 69, 71, 73, 74,
+        76, 78, 80, 79, 78, 78, 77, 77, 78, 78, 78, 78, 79, 80, 80, 81, 82, 82,
+        83, 84, 85, 86, 87, 87, 87, 87, 70, 72, 73, 74, 75, 76, 78, 79, 81, 80,
+        80, 80, 79, 79, 79, 79, 79, 80, 80, 81, 81, 82, 83, 83, 84, 85, 86, 87,
+        87, 87, 87, 87, 77, 77, 77, 77, 77, 78, 80, 81, 82, 82, 82, 81, 81, 81,
+        81, 81, 81, 81, 82, 82, 82, 83, 84, 84, 85, 86, 87, 87, 88, 88, 88, 88,
+        77, 77, 77, 76, 76, 78, 79, 80, 82, 82, 82, 82, 82, 82, 82, 82, 82, 83,
+        83, 83, 84, 84, 85, 85, 86, 87, 87, 88, 89, 89, 89, 89, 77, 77, 76, 76,
+        75, 77, 78, 80, 82, 82, 82, 83, 83, 83, 83, 83, 83, 84, 84, 85, 85, 85,
+        86, 87, 87, 88, 88, 89, 90, 90, 90, 90, 78, 77, 76, 75, 75, 76, 78, 80,
+        81, 82, 83, 83, 84, 84, 84, 85, 85, 85, 86, 86, 86, 87, 87, 88, 88, 89,
+        89, 90, 90, 90, 90, 90, 78, 77, 76, 75, 74, 76, 77, 79, 81, 82, 83, 84,
+        85, 85, 85, 86, 86, 87, 87, 87, 87, 88, 88, 89, 89, 90, 90, 91, 91, 91,
+        91, 91, 79, 78, 76, 75, 74, 76, 77, 79, 81, 82, 83, 84, 85, 86, 86, 87,
+        87, 87, 88, 88, 88, 89, 89, 90, 90, 91, 91, 92, 92, 92, 92, 92, 79, 78,
+        77, 76, 74, 76, 78, 79, 81, 82, 83, 84, 85, 86, 87, 87, 88, 88, 89, 89,
+        89, 90, 90, 91, 91, 92, 92, 92, 93, 93, 93, 93, 80, 79, 77, 76, 75, 76,
+        78, 79, 81, 82, 83, 85, 86, 87, 87, 88, 89, 89, 89, 90, 90, 91, 91, 92,
+        92, 92, 93, 93, 94, 94, 94, 94, 80, 79, 78, 76, 75, 76, 78, 79, 81, 82,
+        83, 85, 86, 87, 88, 89, 89, 90, 90, 91, 91, 92, 92, 93, 93, 93, 94, 94,
+        95, 95, 95, 95, 81, 80, 78, 77, 76, 77, 78, 80, 81, 83, 84, 85, 87, 87,
+        88, 89, 90, 90, 91, 91, 92, 92, 93, 93, 94, 94, 94, 95, 95, 95, 95, 95,
+        82, 81, 79, 78, 76, 78, 79, 80, 82, 83, 84, 86, 87, 88, 89, 89, 90, 91,
+        92, 92, 93, 93, 94, 94, 94, 95, 95, 96, 96, 96, 96, 96, 83, 81, 80, 79,
+        77, 78, 80, 81, 82, 83, 85, 86, 87, 88, 89, 90, 91, 91, 92, 93, 93, 94,
+        94, 95, 95, 95, 96, 96, 97, 97, 97, 97, 84, 82, 81, 79, 78, 79, 80, 81,
+        82, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 93, 94, 94, 95, 95, 96, 96,
+        97, 97, 97, 97, 97, 97, 85, 83, 82, 80, 79, 80, 81, 82, 83, 84, 85, 87,
+        88, 89, 90, 91, 92, 92, 93, 94, 94, 95, 95, 96, 96, 97, 97, 97, 98, 98,
+        98, 98, 86, 84, 83, 81, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91,
+        92, 93, 94, 94, 95, 95, 96, 96, 97, 97, 98, 98, 98, 98, 98, 98, 87, 85,
+        84, 82, 81, 82, 82, 83, 84, 85, 87, 88, 89, 90, 91, 92, 93, 93, 94, 95,
+        95, 96, 96, 97, 97, 98, 98, 98, 99, 99, 99, 99, 88, 86, 85, 83, 82, 82,
+        83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 94, 95, 96, 96, 97, 97,
+        98, 98, 99, 99, 99, 99, 99, 99, 88, 87, 85, 84, 83, 83, 84, 85, 86, 87,
+        88, 89, 90, 91, 92, 92, 93, 94, 95, 95, 96, 97, 97, 98, 98, 99, 99, 99,
+        100, 100, 100, 100, 89, 88, 86, 85, 84, 84, 85, 86, 87, 87, 88, 89, 90,
+        91, 92, 93, 94, 94, 95, 96, 97, 97, 98, 98, 99, 99, 99, 100, 100, 100,
+        100, 100, 90, 89, 87, 86, 85, 85, 86, 87, 87, 88, 89, 90, 91, 92, 92,
+        93, 94, 95, 96, 96, 97, 97, 98, 98, 99, 99, 100, 100, 101, 101, 101,
+        101, 91, 90, 88, 87, 86, 86, 87, 87, 88, 89, 90, 90, 91, 92, 93, 94, 95,
+        95, 96, 97, 97, 98, 98, 99, 99, 100, 100, 101, 101, 101, 101, 101, 91,
+        90, 88, 87, 86, 86, 87, 87, 88, 89, 90, 90, 91, 92, 93, 94, 95, 95, 96,
+        97, 97, 98, 98, 99, 99, 100, 100, 101, 101, 101, 101, 101, 91, 90, 88,
+        87, 86, 86, 87, 87, 88, 89, 90, 90, 91, 92, 93, 94, 95, 95, 96, 97, 97,
+        98, 98, 99, 99, 100, 100, 101, 101, 101, 101, 101, 91, 90, 88, 87, 86,
+        86, 87, 87, 88, 89, 90, 90, 91, 92, 93, 94, 95, 95, 96, 97, 97, 98, 98,
+        99, 99, 100, 100, 101, 101, 101, 101, 101 },
+      { /* Intra matrices */
+        /* Size 4 */
+        45, 57, 59, 65, 57, 62, 63, 67, 59, 63, 69, 72, 65, 67, 72, 76,
+        /* Size 8 */
+        47, 42, 57, 59, 60, 63, 66, 69, 42, 51, 57, 55, 56, 58, 61, 64, 57, 57,
+        61, 61, 61, 62, 64, 66, 59, 55, 61, 64, 65, 66, 67, 69, 60, 56, 61, 65,
+        67, 69, 70, 71, 63, 58, 62, 66, 69, 71, 72, 74, 66, 61, 64, 67, 70, 72,
+        74, 75, 69, 64, 66, 69, 71, 74, 75, 77,
+        /* Size 16 */
+        47, 44, 41, 48, 56, 57, 58, 59, 60, 61, 62, 64, 65, 67, 68, 68, 44, 45,
+        46, 50, 57, 56, 56, 57, 57, 58, 60, 61, 63, 64, 66, 66, 41, 46, 51, 54,
+        57, 56, 54, 55, 55, 56, 57, 59, 60, 62, 63, 63, 48, 50, 54, 56, 59, 58,
+        57, 57, 57, 58, 59, 60, 62, 63, 64, 64, 56, 57, 57, 59, 61, 60, 60, 60,
+        60, 60, 61, 62, 63, 64, 65, 65, 57, 56, 56, 58, 60, 61, 61, 62, 62, 62,
+        63, 64, 65, 66, 67, 67, 58, 56, 54, 57, 60, 61, 63, 63, 64, 65, 65, 66,
+        66, 67, 68, 68, 59, 57, 55, 57, 60, 62, 63, 64, 65, 66, 67, 67, 68, 69,
+        69, 69, 60, 57, 55, 57, 60, 62, 64, 65, 67, 67, 68, 69, 69, 70, 71, 71,
+        61, 58, 56, 58, 60, 62, 65, 66, 67, 68, 69, 70, 71, 71, 72, 72, 62, 60,
+        57, 59, 61, 63, 65, 67, 68, 69, 70, 71, 72, 72, 73, 73, 64, 61, 59, 60,
+        62, 64, 66, 67, 69, 70, 71, 72, 73, 73, 74, 74, 65, 63, 60, 62, 63, 65,
+        66, 68, 69, 71, 72, 73, 73, 74, 75, 75, 67, 64, 62, 63, 64, 66, 67, 69,
+        70, 71, 72, 73, 74, 75, 75, 75, 68, 66, 63, 64, 65, 67, 68, 69, 71, 72,
+        73, 74, 75, 75, 76, 76, 68, 66, 63, 64, 65, 67, 68, 69, 71, 72, 73, 74,
+        75, 75, 76, 76,
+        /* Size 32 */
+        46, 45, 44, 42, 41, 44, 47, 51, 56, 56, 57, 57, 58, 58, 58, 59, 59, 60,
+        61, 61, 62, 63, 63, 64, 65, 65, 66, 67, 68, 68, 68, 68, 45, 45, 44, 44,
+        43, 46, 49, 52, 56, 56, 56, 57, 57, 57, 57, 58, 58, 59, 59, 60, 61, 61,
+        62, 63, 64, 64, 65, 66, 66, 66, 66, 66, 44, 44, 44, 45, 45, 48, 50, 53,
+        56, 56, 56, 56, 56, 56, 56, 57, 57, 58, 58, 59, 59, 60, 61, 62, 62, 63,
+        64, 65, 65, 65, 65, 65, 42, 44, 45, 46, 48, 50, 52, 54, 56, 56, 56, 55,
+        55, 55, 55, 56, 56, 56, 57, 58, 58, 59, 60, 60, 61, 62, 63, 63, 64, 64,
+        64, 64, 41, 43, 45, 48, 50, 52, 53, 55, 57, 56, 55, 55, 54, 54, 54, 55,
+        55, 55, 56, 57, 57, 58, 59, 59, 60, 61, 62, 62, 63, 63, 63, 63, 44, 46,
+        48, 50, 52, 53, 54, 56, 57, 57, 56, 56, 55, 56, 56, 56, 56, 56, 57, 57,
+        58, 59, 59, 60, 61, 61, 62, 63, 64, 64, 64, 64, 47, 49, 50, 52, 53, 54,
+        56, 57, 58, 58, 58, 57, 57, 57, 57, 57, 57, 58, 58, 58, 59, 60, 60, 61,
+        61, 62, 63, 63, 64, 64, 64, 64, 51, 52, 53, 54, 55, 56, 57, 58, 59, 59,
+        59, 58, 58, 58, 58, 58, 58, 59, 59, 59, 60, 60, 61, 62, 62, 63, 63, 64,
+        65, 65, 65, 65, 56, 56, 56, 56, 57, 57, 58, 59, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 61, 61, 62, 62, 63, 63, 64, 65, 65, 65, 65, 65,
+        56, 56, 56, 56, 56, 57, 58, 59, 60, 60, 60, 60, 60, 60, 60, 60, 61, 61,
+        61, 61, 62, 62, 63, 63, 64, 64, 65, 65, 66, 66, 66, 66, 57, 56, 56, 56,
+        55, 56, 58, 59, 60, 60, 61, 61, 61, 61, 61, 61, 62, 62, 62, 62, 63, 63,
+        64, 64, 64, 65, 65, 66, 66, 66, 66, 66, 57, 57, 56, 55, 55, 56, 57, 58,
+        60, 60, 61, 61, 62, 62, 62, 62, 63, 63, 63, 63, 64, 64, 64, 65, 65, 66,
+        66, 67, 67, 67, 67, 67, 58, 57, 56, 55, 54, 55, 57, 58, 60, 60, 61, 62,
+        63, 63, 63, 63, 64, 64, 64, 64, 65, 65, 65, 66, 66, 66, 67, 67, 68, 68,
+        68, 68, 58, 57, 56, 55, 54, 56, 57, 58, 60, 60, 61, 62, 63, 63, 64, 64,
+        64, 65, 65, 65, 65, 66, 66, 66, 67, 67, 68, 68, 68, 68, 68, 68, 58, 57,
+        56, 55, 54, 56, 57, 58, 60, 60, 61, 62, 63, 64, 64, 64, 65, 65, 66, 66,
+        66, 67, 67, 67, 68, 68, 68, 69, 69, 69, 69, 69, 59, 58, 57, 56, 55, 56,
+        57, 58, 60, 60, 61, 62, 63, 64, 64, 65, 66, 66, 66, 67, 67, 67, 68, 68,
+        68, 69, 69, 69, 70, 70, 70, 70, 59, 58, 57, 56, 55, 56, 57, 58, 60, 61,
+        62, 63, 64, 64, 65, 66, 66, 67, 67, 67, 68, 68, 68, 69, 69, 69, 70, 70,
+        70, 70, 70, 70, 60, 59, 58, 56, 55, 56, 58, 59, 60, 61, 62, 63, 64, 65,
+        65, 66, 67, 67, 67, 68, 68, 69, 69, 69, 70, 70, 70, 71, 71, 71, 71, 71,
+        61, 59, 58, 57, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 66, 67, 67,
+        68, 68, 69, 69, 69, 70, 70, 70, 71, 71, 71, 71, 71, 71, 61, 60, 59, 58,
+        57, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 67, 68, 68, 69, 69, 70,
+        70, 70, 71, 71, 71, 72, 72, 72, 72, 72, 62, 61, 59, 58, 57, 58, 59, 60,
+        61, 62, 63, 64, 65, 65, 66, 67, 68, 68, 69, 69, 70, 70, 71, 71, 71, 72,
+        72, 72, 72, 72, 72, 72, 63, 61, 60, 59, 58, 59, 60, 60, 61, 62, 63, 64,
+        65, 66, 67, 67, 68, 69, 69, 70, 70, 71, 71, 71, 72, 72, 72, 73, 73, 73,
+        73, 73, 63, 62, 61, 60, 59, 59, 60, 61, 62, 63, 64, 64, 65, 66, 67, 68,
+        68, 69, 69, 70, 71, 71, 71, 72, 72, 72, 73, 73, 73, 73, 73, 73, 64, 63,
+        62, 60, 59, 60, 61, 62, 62, 63, 64, 65, 66, 66, 67, 68, 69, 69, 70, 70,
+        71, 71, 72, 72, 73, 73, 73, 73, 74, 74, 74, 74, 65, 64, 62, 61, 60, 61,
+        61, 62, 63, 64, 64, 65, 66, 67, 68, 68, 69, 70, 70, 71, 71, 72, 72, 73,
+        73, 73, 74, 74, 74, 74, 74, 74, 65, 64, 63, 62, 61, 61, 62, 63, 63, 64,
+        65, 66, 66, 67, 68, 69, 69, 70, 70, 71, 72, 72, 72, 73, 73, 74, 74, 74,
+        74, 74, 74, 74, 66, 65, 64, 63, 62, 62, 63, 63, 64, 65, 65, 66, 67, 68,
+        68, 69, 70, 70, 71, 71, 72, 72, 73, 73, 74, 74, 74, 74, 75, 75, 75, 75,
+        67, 66, 65, 63, 62, 63, 63, 64, 65, 65, 66, 67, 67, 68, 69, 69, 70, 71,
+        71, 72, 72, 73, 73, 73, 74, 74, 74, 75, 75, 75, 75, 75, 68, 66, 65, 64,
+        63, 64, 64, 65, 65, 66, 66, 67, 68, 68, 69, 70, 70, 71, 71, 72, 72, 73,
+        73, 74, 74, 74, 75, 75, 75, 75, 75, 75, 68, 66, 65, 64, 63, 64, 64, 65,
+        65, 66, 66, 67, 68, 68, 69, 70, 70, 71, 71, 72, 72, 73, 73, 74, 74, 74,
+        75, 75, 75, 75, 75, 75, 68, 66, 65, 64, 63, 64, 64, 65, 65, 66, 66, 67,
+        68, 68, 69, 70, 70, 71, 71, 72, 72, 73, 73, 74, 74, 74, 75, 75, 75, 75,
+        75, 75, 68, 66, 65, 64, 63, 64, 64, 65, 65, 66, 66, 67, 68, 68, 69, 70,
+        70, 71, 71, 72, 72, 73, 73, 74, 74, 74, 75, 75, 75, 75, 75, 75 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 67, 80, 89, 67, 77, 85, 91, 80, 85, 92, 95, 89, 91, 95, 97,
+        /* Size 8 */
+        64, 57, 59, 66, 73, 78, 82, 85, 57, 60, 59, 63, 69, 75, 80, 83, 59, 59,
+        68, 71, 75, 78, 82, 84, 66, 63, 71, 77, 80, 82, 84, 86, 73, 69, 75, 80,
+        83, 85, 86, 88, 78, 75, 78, 82, 85, 87, 88, 89, 82, 80, 82, 84, 86, 88,
+        89, 90, 85, 83, 84, 86, 88, 89, 90, 91,
+        /* Size 16 */
+        64, 60, 57, 58, 59, 62, 66, 69, 73, 75, 78, 80, 82, 84, 85, 85, 60, 59,
+        58, 59, 59, 62, 64, 68, 71, 74, 77, 79, 81, 83, 84, 84, 57, 58, 60, 60,
+        59, 61, 63, 66, 69, 72, 75, 77, 80, 81, 83, 83, 58, 59, 60, 61, 63, 65,
+        67, 69, 72, 74, 77, 79, 81, 82, 84, 84, 59, 59, 59, 63, 68, 69, 71, 73,
+        75, 77, 78, 80, 82, 83, 84, 84, 62, 62, 61, 65, 69, 72, 74, 75, 77, 79,
+        80, 82, 83, 84, 85, 85, 66, 64, 63, 67, 71, 74, 77, 78, 80, 81, 82, 83,
+        84, 85, 86, 86, 69, 68, 66, 69, 73, 75, 78, 80, 81, 82, 83, 84, 85, 86,
+        87, 87, 73, 71, 69, 72, 75, 77, 80, 81, 83, 84, 85, 86, 86, 87, 88, 88,
+        75, 74, 72, 74, 77, 79, 81, 82, 84, 85, 86, 86, 87, 88, 88, 88, 78, 77,
+        75, 77, 78, 80, 82, 83, 85, 86, 87, 87, 88, 88, 89, 89, 80, 79, 77, 79,
+        80, 82, 83, 84, 86, 86, 87, 88, 89, 89, 89, 89, 82, 81, 80, 81, 82, 83,
+        84, 85, 86, 87, 88, 89, 89, 89, 90, 90, 84, 83, 81, 82, 83, 84, 85, 86,
+        87, 88, 88, 89, 89, 90, 90, 90, 85, 84, 83, 84, 84, 85, 86, 87, 88, 88,
+        89, 89, 90, 90, 91, 91, 85, 84, 83, 84, 84, 85, 86, 87, 88, 88, 89, 89,
+        90, 90, 91, 91,
+        /* Size 32 */
+        64, 62, 60, 58, 57, 57, 58, 58, 59, 60, 62, 64, 66, 67, 69, 71, 73, 74,
+        75, 77, 78, 79, 80, 81, 82, 83, 84, 85, 85, 85, 85, 85, 62, 61, 60, 59,
+        57, 58, 58, 58, 59, 60, 62, 63, 65, 67, 68, 70, 72, 73, 75, 76, 77, 78,
+        79, 81, 82, 82, 83, 84, 85, 85, 85, 85, 60, 60, 59, 59, 58, 58, 59, 59,
+        59, 60, 62, 63, 64, 66, 68, 69, 71, 72, 74, 75, 77, 78, 79, 80, 81, 82,
+        83, 83, 84, 84, 84, 84, 58, 59, 59, 59, 59, 59, 59, 59, 59, 60, 61, 63,
+        64, 65, 67, 68, 70, 71, 73, 74, 76, 77, 78, 79, 80, 81, 82, 83, 84, 84,
+        84, 84, 57, 57, 58, 59, 60, 60, 60, 59, 59, 60, 61, 62, 63, 65, 66, 68,
+        69, 71, 72, 73, 75, 76, 77, 78, 80, 80, 81, 82, 83, 83, 83, 83, 57, 58,
+        58, 59, 60, 60, 60, 61, 61, 62, 63, 64, 65, 66, 68, 69, 71, 72, 73, 74,
+        76, 77, 78, 79, 80, 81, 82, 83, 83, 83, 83, 83, 58, 58, 59, 59, 60, 60,
+        61, 62, 63, 64, 65, 66, 67, 68, 69, 71, 72, 73, 74, 75, 77, 78, 79, 80,
+        81, 81, 82, 83, 84, 84, 84, 84, 58, 58, 59, 59, 59, 61, 62, 64, 65, 66,
+        67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 78, 78, 79, 80, 81, 82, 83, 83,
+        84, 84, 84, 84, 59, 59, 59, 59, 59, 61, 63, 65, 68, 69, 69, 70, 71, 72,
+        73, 74, 75, 76, 77, 77, 78, 79, 80, 81, 82, 82, 83, 84, 84, 84, 84, 84,
+        60, 60, 60, 60, 60, 62, 64, 66, 69, 69, 70, 71, 73, 73, 74, 75, 76, 77,
+        78, 78, 79, 80, 81, 82, 82, 83, 84, 84, 85, 85, 85, 85, 62, 62, 62, 61,
+        61, 63, 65, 67, 69, 70, 72, 73, 74, 75, 75, 76, 77, 78, 79, 79, 80, 81,
+        82, 82, 83, 84, 84, 85, 85, 85, 85, 85, 64, 63, 63, 63, 62, 64, 66, 68,
+        70, 71, 73, 74, 75, 76, 77, 78, 78, 79, 80, 80, 81, 82, 82, 83, 84, 84,
+        85, 85, 86, 86, 86, 86, 66, 65, 64, 64, 63, 65, 67, 69, 71, 73, 74, 75,
+        77, 77, 78, 79, 80, 80, 81, 81, 82, 83, 83, 84, 84, 85, 85, 86, 86, 86,
+        86, 86, 67, 67, 66, 65, 65, 66, 68, 70, 72, 73, 75, 76, 77, 78, 79, 80,
+        80, 81, 82, 82, 83, 83, 84, 84, 85, 85, 86, 86, 86, 86, 86, 86, 69, 68,
+        68, 67, 66, 68, 69, 71, 73, 74, 75, 77, 78, 79, 80, 80, 81, 82, 82, 83,
+        83, 84, 84, 85, 85, 86, 86, 86, 87, 87, 87, 87, 71, 70, 69, 68, 68, 69,
+        71, 72, 74, 75, 76, 78, 79, 80, 80, 81, 82, 83, 83, 84, 84, 85, 85, 85,
+        86, 86, 87, 87, 87, 87, 87, 87, 73, 72, 71, 70, 69, 71, 72, 73, 75, 76,
+        77, 78, 80, 80, 81, 82, 83, 83, 84, 84, 85, 85, 86, 86, 86, 87, 87, 87,
+        88, 88, 88, 88, 74, 73, 72, 71, 71, 72, 73, 74, 76, 77, 78, 79, 80, 81,
+        82, 83, 83, 84, 84, 85, 85, 86, 86, 86, 87, 87, 87, 88, 88, 88, 88, 88,
+        75, 75, 74, 73, 72, 73, 74, 75, 77, 78, 79, 80, 81, 82, 82, 83, 84, 84,
+        85, 85, 86, 86, 86, 87, 87, 87, 88, 88, 88, 88, 88, 88, 77, 76, 75, 74,
+        73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 84, 85, 85, 86, 86, 87,
+        87, 87, 88, 88, 88, 88, 89, 89, 89, 89, 78, 77, 77, 76, 75, 76, 77, 78,
+        78, 79, 80, 81, 82, 83, 83, 84, 85, 85, 86, 86, 87, 87, 87, 88, 88, 88,
+        88, 89, 89, 89, 89, 89, 79, 78, 78, 77, 76, 77, 78, 78, 79, 80, 81, 82,
+        83, 83, 84, 85, 85, 86, 86, 87, 87, 87, 88, 88, 88, 88, 89, 89, 89, 89,
+        89, 89, 80, 79, 79, 78, 77, 78, 79, 79, 80, 81, 82, 82, 83, 84, 84, 85,
+        86, 86, 86, 87, 87, 88, 88, 88, 89, 89, 89, 89, 89, 89, 89, 89, 81, 81,
+        80, 79, 78, 79, 80, 80, 81, 82, 82, 83, 84, 84, 85, 85, 86, 86, 87, 87,
+        88, 88, 88, 88, 89, 89, 89, 89, 90, 90, 90, 90, 82, 82, 81, 80, 80, 80,
+        81, 81, 82, 82, 83, 84, 84, 85, 85, 86, 86, 87, 87, 88, 88, 88, 89, 89,
+        89, 89, 89, 90, 90, 90, 90, 90, 83, 82, 82, 81, 80, 81, 81, 82, 82, 83,
+        84, 84, 85, 85, 86, 86, 87, 87, 87, 88, 88, 88, 89, 89, 89, 89, 90, 90,
+        90, 90, 90, 90, 84, 83, 83, 82, 81, 82, 82, 83, 83, 84, 84, 85, 85, 86,
+        86, 87, 87, 87, 88, 88, 88, 89, 89, 89, 89, 90, 90, 90, 90, 90, 90, 90,
+        85, 84, 83, 83, 82, 83, 83, 83, 84, 84, 85, 85, 86, 86, 86, 87, 87, 88,
+        88, 88, 89, 89, 89, 89, 90, 90, 90, 90, 90, 90, 90, 90, 85, 85, 84, 84,
+        83, 83, 84, 84, 84, 85, 85, 86, 86, 86, 87, 87, 88, 88, 88, 89, 89, 89,
+        89, 90, 90, 90, 90, 90, 91, 91, 91, 91, 85, 85, 84, 84, 83, 83, 84, 84,
+        84, 85, 85, 86, 86, 86, 87, 87, 88, 88, 88, 89, 89, 89, 89, 90, 90, 90,
+        90, 90, 91, 91, 91, 91, 85, 85, 84, 84, 83, 83, 84, 84, 84, 85, 85, 86,
+        86, 86, 87, 87, 88, 88, 88, 89, 89, 89, 89, 90, 90, 90, 90, 90, 91, 91,
+        91, 91, 85, 85, 84, 84, 83, 83, 84, 84, 84, 85, 85, 86, 86, 86, 87, 87,
+        88, 88, 88, 89, 89, 89, 89, 90, 90, 90, 90, 90, 91, 91, 91, 91 },
+      { /* Intra matrices */
+        /* Size 4 */
+        48, 50, 60, 68, 50, 58, 64, 69, 60, 64, 70, 72, 68, 69, 72, 74,
+        /* Size 8 */
+        52, 45, 47, 53, 59, 64, 68, 70, 45, 48, 47, 51, 56, 61, 65, 68, 47, 47,
+        55, 58, 61, 64, 67, 69, 53, 51, 58, 62, 65, 67, 69, 71, 59, 56, 61, 65,
+        68, 70, 71, 72, 64, 61, 64, 67, 70, 71, 72, 73, 68, 65, 67, 69, 71, 72,
+        73, 74, 70, 68, 69, 71, 72, 73, 74, 75,
+        /* Size 16 */
+        51, 48, 45, 46, 47, 49, 52, 55, 58, 61, 63, 65, 67, 68, 69, 69, 48, 47,
+        46, 46, 47, 49, 51, 54, 57, 59, 62, 63, 65, 67, 68, 68, 45, 46, 48, 47,
+        47, 48, 50, 53, 55, 58, 60, 62, 64, 66, 67, 67, 46, 46, 47, 49, 50, 52,
+        53, 55, 58, 60, 62, 63, 65, 66, 68, 68, 47, 47, 47, 50, 54, 56, 57, 59,
+        60, 62, 63, 65, 66, 67, 68, 68, 49, 49, 48, 52, 56, 57, 59, 61, 62, 63,
+        65, 66, 67, 68, 69, 69, 52, 51, 50, 53, 57, 59, 62, 63, 64, 65, 66, 67,
+        68, 69, 70, 70, 55, 54, 53, 55, 59, 61, 63, 64, 66, 67, 68, 68, 69, 70,
+        71, 71, 58, 57, 55, 58, 60, 62, 64, 66, 67, 68, 69, 69, 70, 71, 71, 71,
+        61, 59, 58, 60, 62, 63, 65, 67, 68, 69, 70, 70, 71, 71, 72, 72, 63, 62,
+        60, 62, 63, 65, 66, 68, 69, 70, 70, 71, 71, 72, 72, 72, 65, 63, 62, 63,
+        65, 66, 67, 68, 69, 70, 71, 71, 72, 72, 73, 73, 67, 65, 64, 65, 66, 67,
+        68, 69, 70, 71, 71, 72, 72, 73, 73, 73, 68, 67, 66, 66, 67, 68, 69, 70,
+        71, 71, 72, 72, 73, 73, 73, 73, 69, 68, 67, 68, 68, 69, 70, 71, 71, 72,
+        72, 73, 73, 73, 74, 74, 69, 68, 67, 68, 68, 69, 70, 71, 71, 72, 72, 73,
+        73, 73, 74, 74,
+        /* Size 32 */
+        51, 49, 47, 46, 45, 45, 45, 46, 46, 48, 49, 50, 52, 53, 55, 56, 58, 59,
+        60, 61, 63, 64, 64, 65, 66, 67, 68, 68, 69, 69, 69, 69, 49, 48, 47, 46,
+        45, 46, 46, 46, 46, 48, 49, 50, 52, 53, 54, 56, 57, 58, 59, 61, 62, 63,
+        64, 65, 66, 66, 67, 68, 68, 68, 68, 68, 47, 47, 47, 46, 46, 46, 46, 46,
+        46, 47, 49, 50, 51, 52, 54, 55, 56, 58, 59, 60, 61, 62, 63, 64, 65, 66,
+        66, 67, 68, 68, 68, 68, 46, 46, 46, 47, 47, 47, 47, 47, 46, 47, 48, 49,
+        50, 52, 53, 54, 56, 57, 58, 59, 61, 61, 62, 63, 64, 65, 66, 67, 67, 67,
+        67, 67, 45, 45, 46, 47, 47, 47, 47, 47, 47, 47, 48, 49, 50, 51, 52, 54,
+        55, 56, 57, 59, 60, 61, 62, 63, 64, 65, 65, 66, 67, 67, 67, 67, 45, 46,
+        46, 47, 47, 47, 48, 48, 48, 49, 50, 51, 51, 53, 54, 55, 56, 57, 58, 59,
+        61, 61, 62, 63, 64, 65, 66, 66, 67, 67, 67, 67, 45, 46, 46, 47, 47, 48,
+        48, 49, 50, 51, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64,
+        65, 65, 66, 67, 67, 67, 67, 67, 46, 46, 46, 47, 47, 48, 49, 50, 52, 52,
+        53, 54, 55, 56, 57, 58, 58, 59, 60, 61, 62, 63, 64, 64, 65, 66, 66, 67,
+        68, 68, 68, 68, 46, 46, 46, 46, 47, 48, 50, 52, 54, 54, 55, 56, 57, 57,
+        58, 59, 60, 60, 61, 62, 63, 64, 64, 65, 66, 66, 67, 67, 68, 68, 68, 68,
+        48, 48, 47, 47, 47, 49, 51, 52, 54, 55, 56, 57, 58, 59, 59, 60, 61, 61,
+        62, 63, 64, 64, 65, 66, 66, 67, 67, 68, 68, 68, 68, 68, 49, 49, 49, 48,
+        48, 50, 51, 53, 55, 56, 57, 58, 59, 60, 60, 61, 62, 62, 63, 64, 64, 65,
+        66, 66, 67, 67, 68, 68, 69, 69, 69, 69, 50, 50, 50, 49, 49, 51, 52, 54,
+        56, 57, 58, 59, 60, 61, 61, 62, 63, 63, 64, 65, 65, 66, 66, 67, 67, 68,
+        68, 69, 69, 69, 69, 69, 52, 52, 51, 50, 50, 51, 53, 55, 57, 58, 59, 60,
+        61, 62, 63, 63, 64, 64, 65, 65, 66, 66, 67, 67, 68, 68, 69, 69, 69, 69,
+        69, 69, 53, 53, 52, 52, 51, 53, 54, 56, 57, 59, 60, 61, 62, 63, 63, 64,
+        65, 65, 66, 66, 67, 67, 67, 68, 68, 69, 69, 69, 70, 70, 70, 70, 55, 54,
+        54, 53, 52, 54, 55, 57, 58, 59, 60, 61, 63, 63, 64, 65, 65, 66, 66, 67,
+        67, 68, 68, 68, 69, 69, 69, 70, 70, 70, 70, 70, 56, 56, 55, 54, 54, 55,
+        56, 58, 59, 60, 61, 62, 63, 64, 65, 65, 66, 66, 67, 67, 68, 68, 68, 69,
+        69, 70, 70, 70, 70, 70, 70, 70, 58, 57, 56, 56, 55, 56, 57, 58, 60, 61,
+        62, 63, 64, 65, 65, 66, 67, 67, 67, 68, 68, 69, 69, 69, 70, 70, 70, 71,
+        71, 71, 71, 71, 59, 58, 58, 57, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65,
+        66, 66, 67, 67, 68, 68, 69, 69, 69, 70, 70, 70, 71, 71, 71, 71, 71, 71,
+        60, 59, 59, 58, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 66, 67, 67, 68,
+        68, 69, 69, 69, 70, 70, 70, 71, 71, 71, 71, 71, 71, 71, 61, 61, 60, 59,
+        59, 59, 60, 61, 62, 63, 64, 65, 65, 66, 67, 67, 68, 68, 69, 69, 70, 70,
+        70, 70, 71, 71, 71, 71, 72, 72, 72, 72, 63, 62, 61, 61, 60, 61, 61, 62,
+        63, 64, 64, 65, 66, 67, 67, 68, 68, 69, 69, 70, 70, 70, 70, 71, 71, 71,
+        71, 72, 72, 72, 72, 72, 64, 63, 62, 61, 61, 61, 62, 63, 64, 64, 65, 66,
+        66, 67, 68, 68, 69, 69, 69, 70, 70, 70, 71, 71, 71, 71, 72, 72, 72, 72,
+        72, 72, 64, 64, 63, 62, 62, 62, 63, 64, 64, 65, 66, 66, 67, 67, 68, 68,
+        69, 69, 70, 70, 70, 71, 71, 71, 72, 72, 72, 72, 72, 72, 72, 72, 65, 65,
+        64, 63, 63, 63, 64, 64, 65, 66, 66, 67, 67, 68, 68, 69, 69, 70, 70, 70,
+        71, 71, 71, 72, 72, 72, 72, 72, 72, 72, 72, 72, 66, 66, 65, 64, 64, 64,
+        65, 65, 66, 66, 67, 67, 68, 68, 69, 69, 70, 70, 70, 71, 71, 71, 72, 72,
+        72, 72, 72, 73, 73, 73, 73, 73, 67, 66, 66, 65, 65, 65, 65, 66, 66, 67,
+        67, 68, 68, 69, 69, 70, 70, 70, 71, 71, 71, 71, 72, 72, 72, 72, 73, 73,
+        73, 73, 73, 73, 68, 67, 66, 66, 65, 66, 66, 66, 67, 67, 68, 68, 69, 69,
+        69, 70, 70, 71, 71, 71, 71, 72, 72, 72, 72, 73, 73, 73, 73, 73, 73, 73,
+        68, 68, 67, 67, 66, 66, 67, 67, 67, 68, 68, 69, 69, 69, 70, 70, 71, 71,
+        71, 71, 72, 72, 72, 72, 73, 73, 73, 73, 73, 73, 73, 73, 69, 68, 68, 67,
+        67, 67, 67, 68, 68, 68, 69, 69, 69, 70, 70, 70, 71, 71, 71, 72, 72, 72,
+        72, 72, 73, 73, 73, 73, 73, 73, 73, 73, 69, 68, 68, 67, 67, 67, 67, 68,
+        68, 68, 69, 69, 69, 70, 70, 70, 71, 71, 71, 72, 72, 72, 72, 72, 73, 73,
+        73, 73, 73, 73, 73, 73, 69, 68, 68, 67, 67, 67, 67, 68, 68, 68, 69, 69,
+        69, 70, 70, 70, 71, 71, 71, 72, 72, 72, 72, 72, 73, 73, 73, 73, 73, 73,
+        73, 73, 69, 68, 68, 67, 67, 67, 67, 68, 68, 68, 69, 69, 69, 70, 70, 70,
+        71, 71, 71, 72, 72, 72, 72, 72, 73, 73, 73, 73, 73, 73, 73, 73 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 77, 79, 85, 77, 81, 83, 86, 79, 83, 89, 91, 85, 86, 91, 94,
+        /* Size 8 */
+        64, 58, 74, 76, 77, 80, 83, 85, 58, 68, 75, 72, 73, 75, 78, 81, 74, 75,
+        78, 78, 78, 79, 81, 83, 76, 72, 78, 80, 82, 83, 84, 85, 77, 73, 78, 82,
+        84, 85, 87, 88, 80, 75, 79, 83, 85, 87, 89, 90, 83, 78, 81, 84, 87, 89,
+        90, 91, 85, 81, 83, 85, 88, 90, 91, 92,
+        /* Size 16 */
+        64, 61, 58, 65, 74, 75, 76, 76, 77, 78, 80, 81, 83, 84, 85, 85, 61, 62,
+        63, 68, 74, 74, 74, 74, 75, 76, 77, 79, 80, 82, 83, 83, 58, 63, 68, 71,
+        75, 73, 72, 72, 73, 74, 75, 77, 78, 80, 81, 81, 65, 68, 71, 74, 76, 76,
+        75, 75, 75, 76, 77, 78, 79, 81, 82, 82, 74, 74, 75, 76, 78, 78, 78, 78,
+        78, 78, 79, 80, 81, 82, 83, 83, 75, 74, 73, 76, 78, 79, 79, 79, 79, 80,
+        81, 81, 82, 83, 84, 84, 76, 74, 72, 75, 78, 79, 80, 81, 82, 82, 83, 83,
+        84, 85, 85, 85, 76, 74, 72, 75, 78, 79, 81, 82, 83, 83, 84, 85, 85, 86,
+        86, 86, 77, 75, 73, 75, 78, 79, 82, 83, 84, 85, 85, 86, 87, 87, 88, 88,
+        78, 76, 74, 76, 78, 80, 82, 83, 85, 85, 86, 87, 88, 88, 89, 89, 80, 77,
+        75, 77, 79, 81, 83, 84, 85, 86, 87, 88, 89, 89, 90, 90, 81, 79, 77, 78,
+        80, 81, 83, 85, 86, 87, 88, 89, 89, 90, 90, 90, 83, 80, 78, 79, 81, 82,
+        84, 85, 87, 88, 89, 89, 90, 91, 91, 91, 84, 82, 80, 81, 82, 83, 85, 86,
+        87, 88, 89, 90, 91, 91, 92, 92, 85, 83, 81, 82, 83, 84, 85, 86, 88, 89,
+        90, 90, 91, 92, 92, 92, 85, 83, 81, 82, 83, 84, 85, 86, 88, 89, 90, 90,
+        91, 92, 92, 92,
+        /* Size 32 */
+        64, 62, 61, 59, 58, 61, 65, 69, 74, 74, 75, 75, 76, 76, 76, 77, 77, 78,
+        78, 79, 80, 80, 81, 82, 83, 83, 84, 85, 85, 85, 85, 85, 62, 62, 61, 61,
+        60, 63, 67, 70, 74, 74, 74, 75, 75, 75, 75, 76, 76, 77, 77, 78, 79, 79,
+        80, 81, 81, 82, 83, 83, 84, 84, 84, 84, 61, 61, 62, 62, 63, 65, 68, 71,
+        74, 74, 74, 74, 74, 74, 74, 75, 75, 76, 76, 77, 77, 78, 79, 80, 80, 81,
+        82, 82, 83, 83, 83, 83, 59, 61, 62, 64, 65, 67, 70, 72, 74, 74, 74, 73,
+        73, 73, 73, 74, 74, 74, 75, 76, 76, 77, 78, 78, 79, 80, 81, 81, 82, 82,
+        82, 82, 58, 60, 63, 65, 68, 70, 71, 73, 75, 74, 73, 73, 72, 72, 72, 73,
+        73, 73, 74, 75, 75, 76, 77, 77, 78, 79, 80, 80, 81, 81, 81, 81, 61, 63,
+        65, 67, 70, 71, 72, 74, 76, 75, 74, 74, 73, 74, 74, 74, 74, 74, 75, 76,
+        76, 77, 77, 78, 79, 79, 80, 81, 81, 81, 81, 81, 65, 67, 68, 70, 71, 72,
+        74, 75, 76, 76, 76, 75, 75, 75, 75, 75, 75, 76, 76, 76, 77, 78, 78, 79,
+        79, 80, 81, 81, 82, 82, 82, 82, 69, 70, 71, 72, 73, 74, 75, 76, 77, 77,
+        77, 76, 76, 76, 76, 76, 76, 77, 77, 77, 78, 78, 79, 79, 80, 81, 81, 82,
+        82, 82, 82, 82, 74, 74, 74, 74, 75, 76, 76, 77, 78, 78, 78, 78, 78, 78,
+        78, 78, 78, 78, 78, 78, 79, 79, 80, 80, 81, 81, 82, 82, 83, 83, 83, 83,
+        74, 74, 74, 74, 74, 75, 76, 77, 78, 78, 78, 78, 78, 78, 78, 78, 78, 79,
+        79, 79, 80, 80, 81, 81, 81, 82, 82, 83, 83, 83, 83, 83, 75, 74, 74, 74,
+        73, 74, 76, 77, 78, 78, 79, 79, 79, 79, 79, 79, 79, 80, 80, 80, 81, 81,
+        81, 82, 82, 83, 83, 84, 84, 84, 84, 84, 75, 75, 74, 73, 73, 74, 75, 76,
+        78, 78, 79, 79, 80, 80, 80, 80, 81, 81, 81, 81, 82, 82, 82, 83, 83, 83,
+        84, 84, 85, 85, 85, 85, 76, 75, 74, 73, 72, 73, 75, 76, 78, 78, 79, 80,
+        80, 81, 81, 81, 82, 82, 82, 82, 83, 83, 83, 83, 84, 84, 85, 85, 85, 85,
+        85, 85, 76, 75, 74, 73, 72, 74, 75, 76, 78, 78, 79, 80, 81, 81, 81, 82,
+        82, 82, 83, 83, 83, 84, 84, 84, 84, 85, 85, 86, 86, 86, 86, 86, 76, 75,
+        74, 73, 72, 74, 75, 76, 78, 78, 79, 80, 81, 81, 82, 82, 83, 83, 83, 84,
+        84, 84, 85, 85, 85, 85, 86, 86, 86, 86, 86, 86, 77, 76, 75, 74, 73, 74,
+        75, 76, 78, 78, 79, 80, 81, 82, 82, 83, 83, 84, 84, 84, 85, 85, 85, 86,
+        86, 86, 86, 87, 87, 87, 87, 87, 77, 76, 75, 74, 73, 74, 75, 76, 78, 78,
+        79, 81, 82, 82, 83, 83, 84, 84, 85, 85, 85, 86, 86, 86, 87, 87, 87, 87,
+        88, 88, 88, 88, 78, 77, 76, 74, 73, 74, 76, 77, 78, 79, 80, 81, 82, 82,
+        83, 84, 84, 85, 85, 85, 86, 86, 86, 87, 87, 87, 88, 88, 88, 88, 88, 88,
+        78, 77, 76, 75, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 83, 84, 85, 85,
+        85, 86, 86, 87, 87, 87, 88, 88, 88, 88, 89, 89, 89, 89, 79, 78, 77, 76,
+        75, 76, 76, 77, 78, 79, 80, 81, 82, 83, 84, 84, 85, 85, 86, 86, 87, 87,
+        87, 88, 88, 88, 89, 89, 89, 89, 89, 89, 80, 79, 77, 76, 75, 76, 77, 78,
+        79, 80, 81, 82, 83, 83, 84, 85, 85, 86, 86, 87, 87, 88, 88, 88, 89, 89,
+        89, 89, 90, 90, 90, 90, 80, 79, 78, 77, 76, 77, 78, 78, 79, 80, 81, 82,
+        83, 84, 84, 85, 86, 86, 87, 87, 88, 88, 88, 89, 89, 89, 89, 90, 90, 90,
+        90, 90, 81, 80, 79, 78, 77, 77, 78, 79, 80, 81, 81, 82, 83, 84, 85, 85,
+        86, 86, 87, 87, 88, 88, 89, 89, 89, 90, 90, 90, 90, 90, 90, 90, 82, 81,
+        80, 78, 77, 78, 79, 79, 80, 81, 82, 83, 83, 84, 85, 86, 86, 87, 87, 88,
+        88, 89, 89, 89, 90, 90, 90, 90, 91, 91, 91, 91, 83, 81, 80, 79, 78, 79,
+        79, 80, 81, 81, 82, 83, 84, 84, 85, 86, 87, 87, 88, 88, 89, 89, 89, 90,
+        90, 90, 91, 91, 91, 91, 91, 91, 83, 82, 81, 80, 79, 79, 80, 81, 81, 82,
+        83, 83, 84, 85, 85, 86, 87, 87, 88, 88, 89, 89, 90, 90, 90, 91, 91, 91,
+        91, 91, 91, 91, 84, 83, 82, 81, 80, 80, 81, 81, 82, 82, 83, 84, 85, 85,
+        86, 86, 87, 88, 88, 89, 89, 89, 90, 90, 91, 91, 91, 91, 92, 92, 92, 92,
+        85, 83, 82, 81, 80, 81, 81, 82, 82, 83, 84, 84, 85, 86, 86, 87, 87, 88,
+        88, 89, 89, 90, 90, 90, 91, 91, 91, 92, 92, 92, 92, 92, 85, 84, 83, 82,
+        81, 81, 82, 82, 83, 83, 84, 85, 85, 86, 86, 87, 88, 88, 89, 89, 90, 90,
+        90, 91, 91, 91, 92, 92, 92, 92, 92, 92, 85, 84, 83, 82, 81, 81, 82, 82,
+        83, 83, 84, 85, 85, 86, 86, 87, 88, 88, 89, 89, 90, 90, 90, 91, 91, 91,
+        92, 92, 92, 92, 92, 92, 85, 84, 83, 82, 81, 81, 82, 82, 83, 83, 84, 85,
+        85, 86, 86, 87, 88, 88, 89, 89, 90, 90, 90, 91, 91, 91, 92, 92, 92, 92,
+        92, 92, 85, 84, 83, 82, 81, 81, 82, 82, 83, 83, 84, 85, 85, 86, 86, 87,
+        88, 88, 89, 89, 90, 90, 90, 91, 91, 91, 92, 92, 92, 92, 92, 92 },
+      { /* Intra matrices */
+        /* Size 4 */
+        48, 59, 60, 65, 59, 62, 64, 66, 60, 64, 68, 70, 65, 66, 70, 73,
+        /* Size 8 */
+        50, 45, 59, 60, 61, 63, 66, 68, 45, 54, 59, 57, 57, 59, 62, 64, 59, 59,
+        62, 61, 61, 62, 64, 66, 60, 57, 61, 64, 65, 66, 67, 68, 61, 57, 61, 65,
+        67, 68, 69, 70, 63, 59, 62, 66, 68, 70, 71, 72, 66, 62, 64, 67, 69, 71,
+        72, 73, 68, 64, 66, 68, 70, 72, 73, 74,
+        /* Size 16 */
+        50, 47, 45, 51, 58, 59, 59, 60, 61, 62, 63, 64, 65, 66, 67, 67, 47, 48,
+        49, 53, 58, 58, 58, 58, 59, 60, 61, 62, 63, 64, 65, 65, 45, 49, 53, 56,
+        58, 57, 56, 57, 57, 58, 59, 60, 61, 63, 64, 64, 51, 53, 56, 58, 60, 59,
+        59, 59, 59, 60, 60, 61, 62, 63, 65, 65, 58, 58, 58, 60, 62, 61, 61, 61,
+        61, 61, 62, 63, 63, 64, 65, 65, 59, 58, 57, 59, 61, 62, 62, 62, 62, 63,
+        63, 64, 65, 66, 66, 66, 59, 58, 56, 59, 61, 62, 63, 64, 64, 65, 65, 66,
+        66, 67, 67, 67, 60, 58, 57, 59, 61, 62, 64, 64, 65, 66, 66, 67, 67, 68,
+        68, 68, 61, 59, 57, 59, 61, 62, 64, 65, 66, 67, 67, 68, 68, 69, 69, 69,
+        62, 60, 58, 60, 61, 63, 65, 66, 67, 68, 68, 69, 69, 70, 70, 70, 63, 61,
+        59, 60, 62, 63, 65, 66, 67, 68, 69, 70, 70, 71, 71, 71, 64, 62, 60, 61,
+        63, 64, 66, 67, 68, 69, 70, 70, 71, 71, 72, 72, 65, 63, 61, 62, 63, 65,
+        66, 67, 68, 69, 70, 71, 71, 72, 72, 72, 66, 64, 63, 63, 64, 66, 67, 68,
+        69, 70, 71, 71, 72, 72, 73, 73, 67, 65, 64, 65, 65, 66, 67, 68, 69, 70,
+        71, 72, 72, 73, 73, 73, 67, 65, 64, 65, 65, 66, 67, 68, 69, 70, 71, 72,
+        72, 73, 73, 73,
+        /* Size 32 */
+        49, 48, 47, 46, 45, 47, 50, 54, 58, 58, 58, 59, 59, 59, 60, 60, 60, 61,
+        61, 62, 62, 63, 64, 64, 65, 65, 66, 66, 67, 67, 67, 67, 48, 48, 47, 47,
+        46, 49, 52, 55, 58, 58, 58, 58, 58, 58, 59, 59, 59, 60, 60, 61, 61, 62,
+        63, 63, 64, 64, 65, 65, 66, 66, 66, 66, 47, 47, 48, 48, 48, 50, 53, 55,
+        58, 58, 58, 58, 57, 58, 58, 58, 58, 59, 59, 60, 61, 61, 62, 62, 63, 63,
+        64, 65, 65, 65, 65, 65, 46, 47, 48, 49, 51, 52, 54, 56, 58, 58, 57, 57,
+        57, 57, 57, 57, 58, 58, 59, 59, 60, 60, 61, 61, 62, 63, 63, 64, 64, 64,
+        64, 64, 45, 46, 48, 51, 53, 54, 55, 57, 58, 58, 57, 57, 56, 56, 56, 57,
+        57, 57, 58, 58, 59, 59, 60, 60, 61, 62, 62, 63, 63, 63, 63, 63, 47, 49,
+        50, 52, 54, 55, 56, 58, 59, 58, 58, 58, 57, 57, 57, 57, 58, 58, 58, 59,
+        59, 60, 60, 61, 62, 62, 63, 63, 64, 64, 64, 64, 50, 52, 53, 54, 55, 56,
+        57, 59, 60, 59, 59, 59, 58, 58, 58, 58, 59, 59, 59, 60, 60, 61, 61, 62,
+        62, 63, 63, 64, 64, 64, 64, 64, 54, 55, 55, 56, 57, 58, 59, 60, 60, 60,
+        60, 60, 59, 59, 59, 60, 60, 60, 60, 60, 61, 61, 62, 62, 63, 63, 64, 64,
+        65, 65, 65, 65, 58, 58, 58, 58, 58, 59, 60, 60, 61, 61, 61, 61, 61, 61,
+        61, 61, 61, 61, 61, 61, 62, 62, 62, 63, 63, 64, 64, 65, 65, 65, 65, 65,
+        58, 58, 58, 58, 58, 58, 59, 60, 61, 61, 61, 61, 61, 61, 61, 61, 61, 62,
+        62, 62, 62, 63, 63, 63, 64, 64, 65, 65, 66, 66, 66, 66, 58, 58, 58, 57,
+        57, 58, 59, 60, 61, 61, 61, 62, 62, 62, 62, 62, 62, 62, 63, 63, 63, 63,
+        64, 64, 64, 65, 65, 66, 66, 66, 66, 66, 59, 58, 58, 57, 57, 58, 59, 60,
+        61, 61, 62, 62, 62, 63, 63, 63, 63, 63, 63, 64, 64, 64, 65, 65, 65, 65,
+        66, 66, 67, 67, 67, 67, 59, 58, 57, 57, 56, 57, 58, 59, 61, 61, 62, 62,
+        63, 63, 63, 64, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67,
+        67, 67, 59, 58, 58, 57, 56, 57, 58, 59, 61, 61, 62, 63, 63, 64, 64, 64,
+        64, 65, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 68, 68, 68, 68, 60, 59,
+        58, 57, 56, 57, 58, 59, 61, 61, 62, 63, 63, 64, 64, 65, 65, 65, 65, 66,
+        66, 66, 66, 67, 67, 67, 67, 68, 68, 68, 68, 68, 60, 59, 58, 57, 57, 57,
+        58, 60, 61, 61, 62, 63, 64, 64, 65, 65, 65, 66, 66, 66, 67, 67, 67, 67,
+        68, 68, 68, 68, 69, 69, 69, 69, 60, 59, 58, 58, 57, 58, 59, 60, 61, 61,
+        62, 63, 64, 64, 65, 65, 66, 66, 67, 67, 67, 67, 68, 68, 68, 68, 69, 69,
+        69, 69, 69, 69, 61, 60, 59, 58, 57, 58, 59, 60, 61, 62, 62, 63, 64, 65,
+        65, 66, 66, 67, 67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69,
+        61, 60, 59, 59, 58, 58, 59, 60, 61, 62, 63, 63, 64, 65, 65, 66, 67, 67,
+        67, 68, 68, 68, 68, 69, 69, 69, 69, 70, 70, 70, 70, 70, 62, 61, 60, 59,
+        58, 59, 60, 60, 61, 62, 63, 64, 65, 65, 66, 66, 67, 67, 68, 68, 68, 69,
+        69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 62, 61, 61, 60, 59, 59, 60, 61,
+        62, 62, 63, 64, 65, 65, 66, 67, 67, 68, 68, 68, 69, 69, 69, 70, 70, 70,
+        70, 70, 71, 71, 71, 71, 63, 62, 61, 60, 59, 60, 61, 61, 62, 63, 63, 64,
+        65, 66, 66, 67, 67, 68, 68, 69, 69, 69, 70, 70, 70, 70, 71, 71, 71, 71,
+        71, 71, 64, 63, 62, 61, 60, 60, 61, 62, 62, 63, 64, 65, 65, 66, 66, 67,
+        68, 68, 68, 69, 69, 70, 70, 70, 70, 71, 71, 71, 71, 71, 71, 71, 64, 63,
+        62, 61, 60, 61, 62, 62, 63, 63, 64, 65, 66, 66, 67, 67, 68, 68, 69, 69,
+        70, 70, 70, 70, 71, 71, 71, 71, 72, 72, 72, 72, 65, 64, 63, 62, 61, 62,
+        62, 63, 63, 64, 64, 65, 66, 66, 67, 68, 68, 69, 69, 69, 70, 70, 70, 71,
+        71, 71, 71, 72, 72, 72, 72, 72, 65, 64, 63, 63, 62, 62, 63, 63, 64, 64,
+        65, 65, 66, 67, 67, 68, 68, 69, 69, 70, 70, 70, 71, 71, 71, 71, 72, 72,
+        72, 72, 72, 72, 66, 65, 64, 63, 62, 63, 63, 64, 64, 65, 65, 66, 66, 67,
+        67, 68, 69, 69, 69, 70, 70, 71, 71, 71, 71, 72, 72, 72, 72, 72, 72, 72,
+        66, 65, 65, 64, 63, 63, 64, 64, 65, 65, 66, 66, 67, 67, 68, 68, 69, 69,
+        70, 70, 70, 71, 71, 71, 72, 72, 72, 72, 73, 73, 73, 73, 67, 66, 65, 64,
+        63, 64, 64, 65, 65, 66, 66, 67, 67, 68, 68, 69, 69, 69, 70, 70, 71, 71,
+        71, 72, 72, 72, 72, 73, 73, 73, 73, 73, 67, 66, 65, 64, 63, 64, 64, 65,
+        65, 66, 66, 67, 67, 68, 68, 69, 69, 69, 70, 70, 71, 71, 71, 72, 72, 72,
+        72, 73, 73, 73, 73, 73, 67, 66, 65, 64, 63, 64, 64, 65, 65, 66, 66, 67,
+        67, 68, 68, 69, 69, 69, 70, 70, 71, 71, 71, 72, 72, 72, 72, 73, 73, 73,
+        73, 73, 67, 66, 65, 64, 63, 64, 64, 65, 65, 66, 66, 67, 67, 68, 68, 69,
+        69, 69, 70, 70, 71, 71, 71, 72, 72, 72, 72, 73, 73, 73, 73, 73 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 66, 76, 82, 66, 74, 79, 83, 76, 79, 84, 86, 82, 83, 86, 87,
+        /* Size 8 */
+        64, 58, 60, 65, 71, 75, 77, 79, 58, 61, 60, 63, 68, 72, 75, 78, 60, 60,
+        67, 69, 72, 75, 77, 79, 65, 63, 69, 73, 75, 77, 79, 80, 71, 68, 72, 75,
+        78, 79, 80, 81, 75, 72, 75, 77, 79, 80, 81, 82, 77, 75, 77, 79, 80, 81,
+        82, 82, 79, 78, 79, 80, 81, 82, 82, 83,
+        /* Size 16 */
+        64, 61, 58, 59, 60, 62, 65, 68, 71, 72, 75, 76, 77, 78, 79, 79, 61, 60,
+        60, 60, 60, 62, 64, 67, 69, 71, 73, 75, 76, 77, 79, 79, 58, 60, 61, 61,
+        60, 62, 63, 66, 68, 70, 72, 74, 75, 77, 78, 78, 59, 60, 61, 62, 63, 65,
+        66, 68, 70, 72, 73, 75, 76, 77, 78, 78, 60, 60, 60, 63, 67, 68, 69, 71,
+        72, 73, 75, 76, 77, 78, 79, 79, 62, 62, 62, 65, 68, 70, 71, 73, 74, 75,
+        76, 77, 78, 79, 79, 79, 65, 64, 63, 66, 69, 71, 73, 74, 75, 76, 77, 78,
+        79, 79, 80, 80, 68, 67, 66, 68, 71, 73, 74, 75, 77, 77, 78, 79, 79, 80,
+        80, 80, 71, 69, 68, 70, 72, 74, 75, 77, 78, 78, 79, 80, 80, 80, 81, 81,
+        72, 71, 70, 72, 73, 75, 76, 77, 78, 79, 80, 80, 81, 81, 81, 81, 75, 73,
+        72, 73, 75, 76, 77, 78, 79, 80, 80, 81, 81, 81, 82, 82, 76, 75, 74, 75,
+        76, 77, 78, 79, 80, 80, 81, 81, 81, 82, 82, 82, 77, 76, 75, 76, 77, 78,
+        79, 79, 80, 81, 81, 81, 82, 82, 82, 82, 78, 77, 77, 77, 78, 79, 79, 80,
+        80, 81, 81, 82, 82, 82, 83, 83, 79, 79, 78, 78, 79, 79, 80, 80, 81, 81,
+        82, 82, 82, 83, 83, 83, 79, 79, 78, 78, 79, 79, 80, 80, 81, 81, 82, 82,
+        82, 83, 83, 83,
+        /* Size 32 */
+        64, 62, 61, 60, 58, 59, 59, 59, 60, 61, 62, 64, 65, 67, 68, 69, 71, 71,
+        72, 73, 75, 75, 76, 77, 77, 78, 78, 79, 79, 79, 79, 79, 62, 61, 61, 60,
+        59, 59, 59, 60, 60, 61, 62, 64, 65, 66, 67, 69, 70, 71, 72, 73, 74, 75,
+        75, 76, 77, 77, 78, 78, 79, 79, 79, 79, 61, 61, 60, 60, 60, 60, 60, 60,
+        60, 61, 62, 63, 64, 65, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 76, 77,
+        77, 78, 79, 79, 79, 79, 60, 60, 60, 60, 60, 60, 60, 60, 60, 61, 62, 63,
+        64, 65, 66, 67, 69, 70, 71, 72, 73, 73, 74, 75, 76, 76, 77, 78, 78, 78,
+        78, 78, 58, 59, 60, 60, 61, 61, 61, 60, 60, 61, 62, 63, 63, 64, 66, 67,
+        68, 69, 70, 71, 72, 73, 74, 75, 75, 76, 77, 77, 78, 78, 78, 78, 59, 59,
+        60, 60, 61, 61, 61, 61, 62, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72,
+        73, 73, 74, 75, 76, 76, 77, 77, 78, 78, 78, 78, 59, 59, 60, 60, 61, 61,
+        62, 63, 63, 64, 65, 66, 66, 67, 68, 69, 70, 71, 72, 72, 73, 74, 75, 75,
+        76, 77, 77, 78, 78, 78, 78, 78, 59, 60, 60, 60, 60, 61, 63, 64, 65, 66,
+        66, 67, 68, 69, 69, 70, 71, 72, 72, 73, 74, 75, 75, 76, 77, 77, 77, 78,
+        78, 78, 78, 78, 60, 60, 60, 60, 60, 62, 63, 65, 67, 67, 68, 69, 69, 70,
+        71, 71, 72, 73, 73, 74, 75, 75, 76, 76, 77, 77, 78, 78, 79, 79, 79, 79,
+        61, 61, 61, 61, 61, 62, 64, 66, 67, 68, 69, 70, 70, 71, 72, 72, 73, 73,
+        74, 75, 75, 76, 76, 77, 77, 78, 78, 79, 79, 79, 79, 79, 62, 62, 62, 62,
+        62, 63, 65, 66, 68, 69, 70, 71, 71, 72, 73, 73, 74, 74, 75, 75, 76, 76,
+        77, 77, 78, 78, 79, 79, 79, 79, 79, 79, 64, 64, 63, 63, 63, 64, 66, 67,
+        69, 70, 71, 71, 72, 73, 73, 74, 75, 75, 76, 76, 76, 77, 77, 78, 78, 79,
+        79, 79, 80, 80, 80, 80, 65, 65, 64, 64, 63, 65, 66, 68, 69, 70, 71, 72,
+        73, 74, 74, 75, 75, 76, 76, 77, 77, 78, 78, 78, 79, 79, 79, 80, 80, 80,
+        80, 80, 67, 66, 65, 65, 64, 66, 67, 69, 70, 71, 72, 73, 74, 74, 75, 75,
+        76, 76, 77, 77, 78, 78, 78, 79, 79, 79, 80, 80, 80, 80, 80, 80, 68, 67,
+        67, 66, 66, 67, 68, 69, 71, 72, 73, 73, 74, 75, 75, 76, 77, 77, 77, 78,
+        78, 78, 79, 79, 79, 80, 80, 80, 80, 80, 80, 80, 69, 69, 68, 67, 67, 68,
+        69, 70, 71, 72, 73, 74, 75, 75, 76, 77, 77, 77, 78, 78, 79, 79, 79, 79,
+        80, 80, 80, 80, 81, 81, 81, 81, 71, 70, 69, 69, 68, 69, 70, 71, 72, 73,
+        74, 75, 75, 76, 77, 77, 78, 78, 78, 79, 79, 79, 80, 80, 80, 80, 80, 81,
+        81, 81, 81, 81, 71, 71, 70, 70, 69, 70, 71, 72, 73, 73, 74, 75, 76, 76,
+        77, 77, 78, 78, 79, 79, 79, 80, 80, 80, 80, 80, 81, 81, 81, 81, 81, 81,
+        72, 72, 71, 71, 70, 71, 72, 72, 73, 74, 75, 76, 76, 77, 77, 78, 78, 79,
+        79, 79, 80, 80, 80, 80, 81, 81, 81, 81, 81, 81, 81, 81, 73, 73, 72, 72,
+        71, 72, 72, 73, 74, 75, 75, 76, 77, 77, 78, 78, 79, 79, 79, 80, 80, 80,
+        80, 81, 81, 81, 81, 81, 82, 82, 82, 82, 75, 74, 73, 73, 72, 73, 73, 74,
+        75, 75, 76, 76, 77, 78, 78, 79, 79, 79, 80, 80, 80, 80, 81, 81, 81, 81,
+        81, 82, 82, 82, 82, 82, 75, 75, 74, 73, 73, 73, 74, 75, 75, 76, 76, 77,
+        78, 78, 78, 79, 79, 80, 80, 80, 80, 81, 81, 81, 81, 81, 82, 82, 82, 82,
+        82, 82, 76, 75, 75, 74, 74, 74, 75, 75, 76, 76, 77, 77, 78, 78, 79, 79,
+        80, 80, 80, 80, 81, 81, 81, 81, 81, 82, 82, 82, 82, 82, 82, 82, 77, 76,
+        76, 75, 75, 75, 75, 76, 76, 77, 77, 78, 78, 79, 79, 79, 80, 80, 80, 81,
+        81, 81, 81, 81, 82, 82, 82, 82, 82, 82, 82, 82, 77, 77, 76, 76, 75, 76,
+        76, 77, 77, 77, 78, 78, 79, 79, 79, 80, 80, 80, 81, 81, 81, 81, 81, 82,
+        82, 82, 82, 82, 82, 82, 82, 82, 78, 77, 77, 76, 76, 76, 77, 77, 77, 78,
+        78, 79, 79, 79, 80, 80, 80, 80, 81, 81, 81, 81, 82, 82, 82, 82, 82, 82,
+        82, 82, 82, 82, 78, 78, 77, 77, 77, 77, 77, 77, 78, 78, 79, 79, 79, 80,
+        80, 80, 80, 81, 81, 81, 81, 82, 82, 82, 82, 82, 82, 82, 83, 83, 83, 83,
+        79, 78, 78, 78, 77, 77, 78, 78, 78, 79, 79, 79, 80, 80, 80, 80, 81, 81,
+        81, 81, 82, 82, 82, 82, 82, 82, 82, 83, 83, 83, 83, 83, 79, 79, 79, 78,
+        78, 78, 78, 78, 79, 79, 79, 80, 80, 80, 80, 81, 81, 81, 81, 82, 82, 82,
+        82, 82, 82, 82, 83, 83, 83, 83, 83, 83, 79, 79, 79, 78, 78, 78, 78, 78,
+        79, 79, 79, 80, 80, 80, 80, 81, 81, 81, 81, 82, 82, 82, 82, 82, 82, 82,
+        83, 83, 83, 83, 83, 83, 79, 79, 79, 78, 78, 78, 78, 78, 79, 79, 79, 80,
+        80, 80, 80, 81, 81, 81, 81, 82, 82, 82, 82, 82, 82, 82, 83, 83, 83, 83,
+        83, 83, 79, 79, 79, 78, 78, 78, 78, 78, 79, 79, 79, 80, 80, 80, 80, 81,
+        81, 81, 81, 82, 82, 82, 82, 82, 82, 82, 83, 83, 83, 83, 83, 83 },
+      { /* Intra matrices */
+        /* Size 4 */
+        51, 53, 61, 67, 53, 60, 65, 68, 61, 65, 68, 70, 67, 68, 70, 71,
+        /* Size 8 */
+        55, 49, 51, 56, 61, 64, 67, 69, 49, 52, 51, 54, 58, 62, 65, 67, 51, 51,
+        57, 60, 62, 64, 66, 68, 56, 54, 60, 63, 65, 67, 68, 69, 61, 58, 62, 65,
+        67, 68, 69, 70, 64, 62, 64, 67, 68, 70, 70, 71, 67, 65, 66, 68, 69, 70,
+        71, 71, 69, 67, 68, 69, 70, 71, 71, 72,
+        /* Size 16 */
+        54, 51, 49, 50, 50, 53, 55, 58, 60, 62, 64, 65, 66, 67, 68, 68, 51, 51,
+        50, 50, 51, 52, 54, 57, 59, 61, 63, 64, 65, 66, 67, 67, 49, 50, 51, 51,
+        51, 52, 54, 56, 58, 60, 61, 63, 64, 65, 67, 67, 50, 50, 51, 52, 53, 55,
+        56, 58, 59, 61, 63, 64, 65, 66, 67, 67, 50, 51, 51, 53, 57, 58, 59, 60,
+        61, 63, 64, 65, 66, 67, 67, 67, 53, 52, 52, 55, 58, 59, 61, 62, 63, 64,
+        65, 66, 67, 67, 68, 68, 55, 54, 54, 56, 59, 61, 63, 63, 64, 65, 66, 67,
+        67, 68, 69, 69, 58, 57, 56, 58, 60, 62, 63, 64, 65, 66, 67, 67, 68, 68,
+        69, 69, 60, 59, 58, 59, 61, 63, 64, 65, 66, 67, 68, 68, 69, 69, 69, 69,
+        62, 61, 60, 61, 63, 64, 65, 66, 67, 68, 68, 69, 69, 69, 70, 70, 64, 63,
+        61, 63, 64, 65, 66, 67, 68, 68, 69, 69, 70, 70, 70, 70, 65, 64, 63, 64,
+        65, 66, 67, 67, 68, 69, 69, 70, 70, 70, 70, 70, 66, 65, 64, 65, 66, 67,
+        67, 68, 69, 69, 70, 70, 70, 71, 71, 71, 67, 66, 65, 66, 67, 67, 68, 68,
+        69, 69, 70, 70, 71, 71, 71, 71, 68, 67, 67, 67, 67, 68, 69, 69, 69, 70,
+        70, 70, 71, 71, 71, 71, 68, 67, 67, 67, 67, 68, 69, 69, 69, 70, 70, 70,
+        71, 71, 71, 71,
+        /* Size 32 */
+        54, 52, 51, 50, 49, 49, 49, 50, 50, 51, 53, 54, 55, 56, 57, 58, 60, 61,
+        61, 62, 63, 64, 65, 65, 66, 66, 67, 67, 68, 68, 68, 68, 52, 52, 51, 50,
+        49, 50, 50, 50, 50, 51, 52, 53, 55, 56, 57, 58, 59, 60, 61, 62, 63, 63,
+        64, 65, 65, 66, 66, 67, 67, 67, 67, 67, 51, 51, 51, 50, 50, 50, 50, 50,
+        50, 51, 52, 53, 54, 55, 56, 57, 59, 59, 60, 61, 62, 63, 64, 64, 65, 65,
+        66, 67, 67, 67, 67, 67, 50, 50, 50, 50, 51, 51, 50, 50, 50, 51, 52, 53,
+        54, 55, 56, 57, 58, 59, 60, 61, 62, 62, 63, 64, 65, 65, 66, 66, 67, 67,
+        67, 67, 49, 49, 50, 51, 51, 51, 51, 51, 50, 51, 52, 53, 53, 54, 55, 56,
+        57, 58, 59, 60, 61, 62, 63, 63, 64, 65, 65, 66, 66, 66, 66, 66, 49, 50,
+        50, 51, 51, 51, 51, 52, 52, 52, 53, 54, 55, 55, 56, 57, 58, 59, 60, 61,
+        62, 62, 63, 64, 64, 65, 65, 66, 67, 67, 67, 67, 49, 50, 50, 50, 51, 51,
+        52, 53, 53, 54, 55, 55, 56, 57, 58, 58, 59, 60, 61, 61, 62, 63, 63, 64,
+        65, 65, 66, 66, 67, 67, 67, 67, 50, 50, 50, 50, 51, 52, 53, 54, 55, 55,
+        56, 57, 57, 58, 59, 59, 60, 61, 61, 62, 63, 63, 64, 65, 65, 66, 66, 66,
+        67, 67, 67, 67, 50, 50, 50, 50, 50, 52, 53, 55, 56, 57, 58, 58, 59, 59,
+        60, 60, 61, 62, 62, 63, 63, 64, 64, 65, 65, 66, 66, 67, 67, 67, 67, 67,
+        51, 51, 51, 51, 51, 52, 54, 55, 57, 58, 58, 59, 60, 60, 61, 61, 62, 62,
+        63, 63, 64, 64, 65, 65, 66, 66, 67, 67, 67, 67, 67, 67, 53, 52, 52, 52,
+        52, 53, 55, 56, 58, 58, 59, 60, 60, 61, 62, 62, 63, 63, 64, 64, 65, 65,
+        65, 66, 66, 67, 67, 67, 68, 68, 68, 68, 54, 53, 53, 53, 53, 54, 55, 57,
+        58, 59, 60, 61, 61, 62, 62, 63, 63, 64, 64, 65, 65, 65, 66, 66, 67, 67,
+        67, 68, 68, 68, 68, 68, 55, 55, 54, 54, 53, 55, 56, 57, 59, 60, 60, 61,
+        62, 63, 63, 64, 64, 65, 65, 65, 66, 66, 66, 67, 67, 67, 68, 68, 68, 68,
+        68, 68, 56, 56, 55, 55, 54, 55, 57, 58, 59, 60, 61, 62, 63, 63, 64, 64,
+        65, 65, 65, 66, 66, 66, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 57, 57,
+        56, 56, 55, 56, 58, 59, 60, 61, 62, 62, 63, 64, 64, 65, 65, 65, 66, 66,
+        67, 67, 67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 58, 58, 57, 57, 56, 57,
+        58, 59, 60, 61, 62, 63, 64, 64, 65, 65, 66, 66, 66, 67, 67, 67, 67, 68,
+        68, 68, 68, 69, 69, 69, 69, 69, 60, 59, 59, 58, 57, 58, 59, 60, 61, 62,
+        63, 63, 64, 65, 65, 66, 66, 66, 67, 67, 67, 68, 68, 68, 68, 69, 69, 69,
+        69, 69, 69, 69, 61, 60, 59, 59, 58, 59, 60, 61, 62, 62, 63, 64, 65, 65,
+        65, 66, 66, 67, 67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69,
+        61, 61, 60, 60, 59, 60, 61, 61, 62, 63, 64, 64, 65, 65, 66, 66, 67, 67,
+        67, 68, 68, 68, 68, 69, 69, 69, 69, 69, 70, 70, 70, 70, 62, 62, 61, 61,
+        60, 61, 61, 62, 63, 63, 64, 65, 65, 66, 66, 67, 67, 67, 68, 68, 68, 68,
+        69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 63, 63, 62, 62, 61, 62, 62, 63,
+        63, 64, 65, 65, 66, 66, 67, 67, 67, 68, 68, 68, 69, 69, 69, 69, 69, 69,
+        70, 70, 70, 70, 70, 70, 64, 63, 63, 62, 62, 62, 63, 63, 64, 64, 65, 65,
+        66, 66, 67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70,
+        70, 70, 65, 64, 64, 63, 63, 63, 63, 64, 64, 65, 65, 66, 66, 67, 67, 67,
+        68, 68, 68, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70, 65, 65,
+        64, 64, 63, 64, 64, 65, 65, 65, 66, 66, 67, 67, 67, 68, 68, 68, 69, 69,
+        69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70, 70, 66, 65, 65, 65, 64, 64,
+        65, 65, 65, 66, 66, 67, 67, 67, 68, 68, 68, 69, 69, 69, 69, 69, 70, 70,
+        70, 70, 70, 70, 70, 70, 70, 70, 66, 66, 65, 65, 65, 65, 65, 66, 66, 66,
+        67, 67, 67, 68, 68, 68, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70,
+        71, 71, 71, 71, 67, 66, 66, 66, 65, 65, 66, 66, 66, 67, 67, 67, 68, 68,
+        68, 68, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 71, 71, 71, 71, 71,
+        67, 67, 67, 66, 66, 66, 66, 66, 67, 67, 67, 68, 68, 68, 68, 69, 69, 69,
+        69, 70, 70, 70, 70, 70, 70, 70, 71, 71, 71, 71, 71, 71, 68, 67, 67, 67,
+        66, 67, 67, 67, 67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 70, 70, 70, 70,
+        70, 70, 70, 71, 71, 71, 71, 71, 71, 71, 68, 67, 67, 67, 66, 67, 67, 67,
+        67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 71,
+        71, 71, 71, 71, 71, 71, 68, 67, 67, 67, 66, 67, 67, 67, 67, 67, 68, 68,
+        68, 68, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 71, 71, 71, 71, 71,
+        71, 71, 68, 67, 67, 67, 66, 67, 67, 67, 67, 67, 68, 68, 68, 68, 69, 69,
+        69, 69, 70, 70, 70, 70, 70, 70, 70, 71, 71, 71, 71, 71, 71, 71 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 74, 75, 79, 74, 77, 78, 80, 75, 78, 82, 84, 79, 80, 84, 86,
+        /* Size 8 */
+        64, 59, 72, 73, 74, 76, 78, 80, 59, 67, 72, 70, 71, 72, 75, 77, 72, 72,
+        75, 74, 74, 75, 76, 78, 73, 70, 74, 76, 77, 78, 79, 80, 74, 71, 74, 77,
+        79, 80, 80, 81, 76, 72, 75, 78, 80, 81, 82, 82, 78, 75, 76, 79, 80, 82,
+        83, 83, 80, 77, 78, 80, 81, 82, 83, 84,
+        /* Size 16 */
+        64, 62, 59, 65, 72, 72, 73, 73, 74, 75, 76, 77, 78, 79, 80, 80, 62, 62,
+        63, 67, 72, 72, 71, 72, 72, 73, 74, 75, 76, 77, 78, 78, 59, 63, 67, 70,
+        72, 71, 70, 70, 71, 72, 72, 74, 75, 76, 77, 77, 65, 67, 70, 71, 73, 73,
+        72, 72, 72, 73, 74, 75, 75, 76, 77, 77, 72, 72, 72, 73, 75, 74, 74, 74,
+        74, 75, 75, 76, 76, 77, 78, 78, 72, 72, 71, 73, 74, 75, 75, 75, 76, 76,
+        76, 77, 77, 78, 79, 79, 73, 71, 70, 72, 74, 75, 76, 77, 77, 77, 78, 78,
+        79, 79, 80, 80, 73, 72, 70, 72, 74, 75, 77, 77, 78, 78, 79, 79, 79, 80,
+        80, 80, 74, 72, 71, 72, 74, 76, 77, 78, 79, 79, 80, 80, 80, 81, 81, 81,
+        75, 73, 72, 73, 75, 76, 77, 78, 79, 80, 80, 81, 81, 81, 82, 82, 76, 74,
+        72, 74, 75, 76, 78, 79, 80, 80, 81, 81, 82, 82, 82, 82, 77, 75, 74, 75,
+        76, 77, 78, 79, 80, 81, 81, 82, 82, 83, 83, 83, 78, 76, 75, 75, 76, 77,
+        79, 79, 80, 81, 82, 82, 83, 83, 83, 83, 79, 77, 76, 76, 77, 78, 79, 80,
+        81, 81, 82, 83, 83, 83, 84, 84, 80, 78, 77, 77, 78, 79, 80, 80, 81, 82,
+        82, 83, 83, 84, 84, 84, 80, 78, 77, 77, 78, 79, 80, 80, 81, 82, 82, 83,
+        83, 84, 84, 84,
+        /* Size 32 */
+        64, 63, 62, 60, 59, 62, 65, 68, 72, 72, 72, 72, 73, 73, 73, 74, 74, 74,
+        75, 75, 76, 76, 77, 77, 78, 78, 79, 79, 80, 80, 80, 80, 63, 62, 62, 61,
+        61, 63, 66, 69, 72, 72, 72, 72, 72, 72, 73, 73, 73, 74, 74, 74, 75, 75,
+        76, 76, 77, 77, 78, 78, 79, 79, 79, 79, 62, 62, 62, 63, 63, 65, 67, 69,
+        72, 72, 72, 72, 71, 72, 72, 72, 72, 73, 73, 74, 74, 75, 75, 76, 76, 77,
+        77, 78, 78, 78, 78, 78, 60, 61, 63, 64, 65, 67, 68, 70, 72, 72, 71, 71,
+        71, 71, 71, 71, 71, 72, 72, 73, 73, 74, 74, 75, 75, 76, 76, 77, 77, 77,
+        77, 77, 59, 61, 63, 65, 67, 68, 70, 71, 72, 72, 71, 71, 70, 70, 70, 71,
+        71, 71, 72, 72, 72, 73, 74, 74, 75, 75, 76, 76, 77, 77, 77, 77, 62, 63,
+        65, 67, 68, 69, 70, 72, 73, 72, 72, 72, 71, 71, 71, 71, 72, 72, 72, 73,
+        73, 74, 74, 75, 75, 75, 76, 76, 77, 77, 77, 77, 65, 66, 67, 68, 70, 70,
+        71, 72, 73, 73, 73, 72, 72, 72, 72, 72, 72, 73, 73, 73, 74, 74, 75, 75,
+        75, 76, 76, 77, 77, 77, 77, 77, 68, 69, 69, 70, 71, 72, 72, 73, 74, 74,
+        74, 73, 73, 73, 73, 73, 73, 74, 74, 74, 74, 75, 75, 76, 76, 76, 77, 77,
+        78, 78, 78, 78, 72, 72, 72, 72, 72, 73, 73, 74, 75, 75, 74, 74, 74, 74,
+        74, 74, 74, 74, 75, 75, 75, 75, 76, 76, 76, 77, 77, 78, 78, 78, 78, 78,
+        72, 72, 72, 72, 72, 72, 73, 74, 75, 75, 75, 75, 75, 75, 75, 75, 75, 75,
+        75, 75, 76, 76, 76, 77, 77, 77, 78, 78, 78, 78, 78, 78, 72, 72, 72, 71,
+        71, 72, 73, 74, 74, 75, 75, 75, 75, 75, 75, 75, 76, 76, 76, 76, 76, 77,
+        77, 77, 77, 78, 78, 78, 79, 79, 79, 79, 72, 72, 72, 71, 71, 72, 72, 73,
+        74, 75, 75, 75, 76, 76, 76, 76, 76, 76, 77, 77, 77, 77, 77, 78, 78, 78,
+        79, 79, 79, 79, 79, 79, 73, 72, 71, 71, 70, 71, 72, 73, 74, 75, 75, 76,
+        76, 76, 77, 77, 77, 77, 77, 78, 78, 78, 78, 78, 79, 79, 79, 79, 80, 80,
+        80, 80, 73, 72, 72, 71, 70, 71, 72, 73, 74, 75, 75, 76, 76, 77, 77, 77,
+        77, 78, 78, 78, 78, 78, 79, 79, 79, 79, 80, 80, 80, 80, 80, 80, 73, 73,
+        72, 71, 70, 71, 72, 73, 74, 75, 75, 76, 77, 77, 77, 78, 78, 78, 78, 78,
+        79, 79, 79, 79, 79, 80, 80, 80, 80, 80, 80, 80, 74, 73, 72, 71, 71, 71,
+        72, 73, 74, 75, 75, 76, 77, 77, 78, 78, 78, 78, 79, 79, 79, 79, 80, 80,
+        80, 80, 80, 81, 81, 81, 81, 81, 74, 73, 72, 71, 71, 72, 72, 73, 74, 75,
+        76, 76, 77, 77, 78, 78, 79, 79, 79, 79, 80, 80, 80, 80, 80, 81, 81, 81,
+        81, 81, 81, 81, 74, 74, 73, 72, 71, 72, 73, 74, 74, 75, 76, 76, 77, 78,
+        78, 78, 79, 79, 79, 80, 80, 80, 80, 81, 81, 81, 81, 81, 82, 82, 82, 82,
+        75, 74, 73, 72, 72, 72, 73, 74, 75, 75, 76, 77, 77, 78, 78, 79, 79, 79,
+        80, 80, 80, 80, 81, 81, 81, 81, 81, 82, 82, 82, 82, 82, 75, 74, 74, 73,
+        72, 73, 73, 74, 75, 75, 76, 77, 78, 78, 78, 79, 79, 80, 80, 80, 81, 81,
+        81, 81, 81, 82, 82, 82, 82, 82, 82, 82, 76, 75, 74, 73, 72, 73, 74, 74,
+        75, 76, 76, 77, 78, 78, 79, 79, 80, 80, 80, 81, 81, 81, 81, 82, 82, 82,
+        82, 82, 82, 82, 82, 82, 76, 75, 75, 74, 73, 74, 74, 75, 75, 76, 77, 77,
+        78, 78, 79, 79, 80, 80, 80, 81, 81, 81, 82, 82, 82, 82, 82, 83, 83, 83,
+        83, 83, 77, 76, 75, 74, 74, 74, 75, 75, 76, 76, 77, 77, 78, 79, 79, 80,
+        80, 80, 81, 81, 81, 82, 82, 82, 82, 82, 83, 83, 83, 83, 83, 83, 77, 76,
+        76, 75, 74, 75, 75, 76, 76, 77, 77, 78, 78, 79, 79, 80, 80, 81, 81, 81,
+        82, 82, 82, 82, 83, 83, 83, 83, 83, 83, 83, 83, 78, 77, 76, 75, 75, 75,
+        75, 76, 76, 77, 77, 78, 79, 79, 79, 80, 80, 81, 81, 81, 82, 82, 82, 83,
+        83, 83, 83, 83, 83, 83, 83, 83, 78, 77, 77, 76, 75, 75, 76, 76, 77, 77,
+        78, 78, 79, 79, 80, 80, 81, 81, 81, 82, 82, 82, 82, 83, 83, 83, 83, 83,
+        84, 84, 84, 84, 79, 78, 77, 76, 76, 76, 76, 77, 77, 78, 78, 79, 79, 80,
+        80, 80, 81, 81, 81, 82, 82, 82, 83, 83, 83, 83, 83, 84, 84, 84, 84, 84,
+        79, 78, 78, 77, 76, 76, 77, 77, 78, 78, 78, 79, 79, 80, 80, 81, 81, 81,
+        82, 82, 82, 83, 83, 83, 83, 83, 84, 84, 84, 84, 84, 84, 80, 79, 78, 77,
+        77, 77, 77, 78, 78, 78, 79, 79, 80, 80, 80, 81, 81, 82, 82, 82, 82, 83,
+        83, 83, 83, 84, 84, 84, 84, 84, 84, 84, 80, 79, 78, 77, 77, 77, 77, 78,
+        78, 78, 79, 79, 80, 80, 80, 81, 81, 82, 82, 82, 82, 83, 83, 83, 83, 84,
+        84, 84, 84, 84, 84, 84, 80, 79, 78, 77, 77, 77, 77, 78, 78, 78, 79, 79,
+        80, 80, 80, 81, 81, 82, 82, 82, 82, 83, 83, 83, 83, 84, 84, 84, 84, 84,
+        84, 84, 80, 79, 78, 77, 77, 77, 77, 78, 78, 78, 79, 79, 80, 80, 80, 81,
+        81, 82, 82, 82, 82, 83, 83, 83, 83, 84, 84, 84, 84, 84, 84, 84 },
+      { /* Intra matrices */
+        /* Size 4 */
+        52, 60, 61, 65, 60, 63, 64, 66, 61, 64, 67, 69, 65, 66, 69, 71,
+        /* Size 8 */
+        53, 49, 60, 61, 62, 64, 65, 67, 49, 56, 60, 59, 59, 61, 63, 64, 60, 60,
+        63, 62, 62, 63, 64, 66, 61, 59, 62, 64, 65, 65, 66, 67, 62, 59, 62, 65,
+        66, 67, 68, 69, 64, 61, 63, 65, 67, 68, 69, 70, 65, 63, 64, 66, 68, 69,
+        70, 71, 67, 64, 66, 67, 69, 70, 71, 71,
+        /* Size 16 */
+        53, 51, 49, 54, 60, 60, 61, 61, 62, 62, 63, 64, 65, 66, 67, 67, 51, 51,
+        52, 56, 60, 60, 59, 60, 60, 61, 62, 63, 64, 64, 65, 65, 49, 52, 56, 58,
+        60, 59, 58, 58, 59, 59, 60, 61, 62, 63, 64, 64, 54, 56, 58, 59, 61, 61,
+        60, 60, 60, 61, 61, 62, 63, 64, 65, 65, 60, 60, 60, 61, 62, 62, 62, 62,
+        62, 62, 63, 63, 64, 64, 65, 65, 60, 60, 59, 61, 62, 62, 63, 63, 63, 63,
+        64, 64, 65, 65, 66, 66, 61, 59, 58, 60, 62, 63, 64, 64, 64, 65, 65, 65,
+        66, 66, 67, 67, 61, 60, 58, 60, 62, 63, 64, 64, 65, 65, 66, 66, 67, 67,
+        67, 67, 62, 60, 59, 60, 62, 63, 64, 65, 66, 66, 67, 67, 67, 68, 68, 68,
+        62, 61, 59, 61, 62, 63, 65, 65, 66, 67, 67, 68, 68, 68, 69, 69, 63, 62,
+        60, 61, 63, 64, 65, 66, 67, 67, 68, 68, 69, 69, 69, 69, 64, 63, 61, 62,
+        63, 64, 65, 66, 67, 68, 68, 69, 69, 69, 70, 70, 65, 64, 62, 63, 64, 65,
+        66, 67, 67, 68, 69, 69, 69, 70, 70, 70, 66, 64, 63, 64, 64, 65, 66, 67,
+        68, 68, 69, 69, 70, 70, 70, 70, 67, 65, 64, 65, 65, 66, 67, 67, 68, 69,
+        69, 70, 70, 70, 71, 71, 67, 65, 64, 65, 65, 66, 67, 67, 68, 69, 69, 70,
+        70, 70, 71, 71,
+        /* Size 32 */
+        53, 52, 51, 50, 49, 51, 53, 56, 59, 60, 60, 60, 60, 61, 61, 61, 61, 62,
+        62, 63, 63, 63, 64, 64, 65, 65, 66, 66, 66, 66, 66, 66, 52, 51, 51, 50,
+        50, 52, 54, 57, 59, 60, 60, 60, 60, 60, 60, 60, 61, 61, 61, 62, 62, 63,
+        63, 64, 64, 64, 65, 65, 66, 66, 66, 66, 51, 51, 51, 51, 52, 54, 55, 57,
+        60, 59, 59, 59, 59, 59, 60, 60, 60, 60, 61, 61, 62, 62, 62, 63, 63, 64,
+        64, 65, 65, 65, 65, 65, 50, 50, 51, 52, 54, 55, 56, 58, 60, 59, 59, 59,
+        59, 59, 59, 59, 59, 60, 60, 60, 61, 61, 62, 62, 63, 63, 63, 64, 64, 64,
+        64, 64, 49, 50, 52, 54, 55, 56, 58, 59, 60, 59, 59, 58, 58, 58, 58, 58,
+        59, 59, 59, 60, 60, 61, 61, 61, 62, 62, 63, 63, 64, 64, 64, 64, 51, 52,
+        54, 55, 56, 57, 58, 59, 60, 60, 60, 59, 59, 59, 59, 59, 59, 60, 60, 60,
+        61, 61, 61, 62, 62, 63, 63, 64, 64, 64, 64, 64, 53, 54, 55, 56, 58, 58,
+        59, 60, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 61, 61, 61, 62, 62, 62,
+        63, 63, 64, 64, 64, 64, 64, 64, 56, 57, 57, 58, 59, 59, 60, 61, 61, 61,
+        61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 62, 62, 62, 63, 63, 63, 64, 64,
+        65, 65, 65, 65, 59, 59, 60, 60, 60, 60, 61, 61, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 63, 63, 63, 64, 64, 64, 65, 65, 65, 65, 65,
+        60, 60, 59, 59, 59, 60, 61, 61, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62,
+        63, 63, 63, 63, 63, 64, 64, 64, 65, 65, 65, 65, 65, 65, 60, 60, 59, 59,
+        59, 60, 60, 61, 62, 62, 62, 62, 62, 63, 63, 63, 63, 63, 63, 63, 63, 64,
+        64, 64, 64, 65, 65, 65, 66, 66, 66, 66, 60, 60, 59, 59, 58, 59, 60, 61,
+        62, 62, 62, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 65, 65, 65, 65,
+        65, 66, 66, 66, 66, 66, 60, 60, 59, 59, 58, 59, 60, 61, 62, 62, 62, 63,
+        63, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66,
+        66, 66, 61, 60, 59, 59, 58, 59, 60, 61, 62, 62, 63, 63, 64, 64, 64, 64,
+        64, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 61, 60,
+        60, 59, 58, 59, 60, 61, 62, 62, 63, 63, 64, 64, 64, 65, 65, 65, 65, 65,
+        66, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 61, 60, 60, 59, 58, 59,
+        60, 61, 62, 62, 63, 63, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 66, 67,
+        67, 67, 67, 67, 67, 67, 67, 67, 61, 61, 60, 59, 59, 59, 60, 61, 62, 62,
+        63, 63, 64, 64, 65, 65, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 68,
+        68, 68, 68, 68, 62, 61, 60, 60, 59, 60, 60, 61, 62, 62, 63, 64, 64, 65,
+        65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68,
+        62, 61, 61, 60, 59, 60, 61, 61, 62, 63, 63, 64, 64, 65, 65, 66, 66, 66,
+        66, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 63, 62, 61, 60,
+        60, 60, 61, 61, 62, 63, 63, 64, 65, 65, 65, 66, 66, 66, 67, 67, 67, 67,
+        68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 63, 62, 62, 61, 60, 61, 61, 62,
+        62, 63, 63, 64, 65, 65, 66, 66, 66, 67, 67, 67, 68, 68, 68, 68, 68, 68,
+        69, 69, 69, 69, 69, 69, 63, 63, 62, 61, 61, 61, 62, 62, 63, 63, 64, 64,
+        65, 65, 66, 66, 67, 67, 67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69,
+        69, 69, 64, 63, 62, 62, 61, 61, 62, 62, 63, 63, 64, 65, 65, 65, 66, 66,
+        67, 67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 64, 64,
+        63, 62, 61, 62, 62, 63, 63, 64, 64, 65, 65, 66, 66, 67, 67, 67, 68, 68,
+        68, 68, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 65, 64, 63, 63, 62, 62,
+        63, 63, 64, 64, 64, 65, 65, 66, 66, 67, 67, 67, 68, 68, 68, 69, 69, 69,
+        69, 69, 70, 70, 70, 70, 70, 70, 65, 64, 64, 63, 62, 63, 63, 63, 64, 64,
+        65, 65, 66, 66, 66, 67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 70, 70, 70,
+        70, 70, 70, 70, 66, 65, 64, 63, 63, 63, 64, 64, 64, 65, 65, 65, 66, 66,
+        67, 67, 67, 68, 68, 68, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70,
+        66, 65, 65, 64, 63, 64, 64, 64, 65, 65, 65, 66, 66, 67, 67, 67, 68, 68,
+        68, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70, 66, 66, 65, 64,
+        64, 64, 64, 65, 65, 65, 66, 66, 66, 67, 67, 67, 68, 68, 68, 69, 69, 69,
+        69, 70, 70, 70, 70, 70, 70, 70, 70, 70, 66, 66, 65, 64, 64, 64, 64, 65,
+        65, 65, 66, 66, 66, 67, 67, 67, 68, 68, 68, 69, 69, 69, 69, 70, 70, 70,
+        70, 70, 70, 70, 70, 70, 66, 66, 65, 64, 64, 64, 64, 65, 65, 65, 66, 66,
+        66, 67, 67, 67, 68, 68, 68, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70,
+        70, 70, 66, 66, 65, 64, 64, 64, 64, 65, 65, 65, 66, 66, 66, 67, 67, 67,
+        68, 68, 68, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70, 70 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 65, 72, 76, 65, 71, 74, 76, 72, 74, 77, 78, 76, 76, 78, 79,
+        /* Size 8 */
+        64, 60, 61, 65, 68, 71, 73, 74, 60, 62, 61, 64, 67, 69, 71, 73, 61, 61,
+        66, 68, 69, 71, 72, 73, 65, 64, 68, 70, 71, 72, 73, 74, 68, 67, 69, 71,
+        73, 74, 74, 75, 71, 69, 71, 72, 74, 74, 75, 75, 73, 71, 72, 73, 74, 75,
+        75, 76, 74, 73, 73, 74, 75, 75, 76, 76,
+        /* Size 16 */
+        64, 62, 60, 60, 61, 63, 65, 67, 68, 70, 71, 72, 73, 73, 74, 74, 62, 61,
+        61, 61, 61, 63, 64, 66, 68, 69, 70, 71, 72, 73, 73, 73, 60, 61, 62, 62,
+        61, 62, 64, 65, 67, 68, 69, 70, 71, 72, 73, 73, 60, 61, 62, 63, 64, 65,
+        66, 67, 68, 69, 70, 71, 72, 73, 73, 73, 61, 61, 61, 64, 66, 67, 68, 68,
+        69, 70, 71, 72, 72, 73, 73, 73, 63, 63, 62, 65, 67, 68, 69, 70, 70, 71,
+        72, 72, 73, 73, 74, 74, 65, 64, 64, 66, 68, 69, 70, 71, 71, 72, 72, 73,
+        73, 74, 74, 74, 67, 66, 65, 67, 68, 70, 71, 71, 72, 73, 73, 73, 74, 74,
+        74, 74, 68, 68, 67, 68, 69, 70, 71, 72, 73, 73, 74, 74, 74, 74, 75, 75,
+        70, 69, 68, 69, 70, 71, 72, 73, 73, 74, 74, 74, 75, 75, 75, 75, 71, 70,
+        69, 70, 71, 72, 72, 73, 74, 74, 74, 75, 75, 75, 75, 75, 72, 71, 70, 71,
+        72, 72, 73, 73, 74, 74, 75, 75, 75, 75, 75, 75, 73, 72, 71, 72, 72, 73,
+        73, 74, 74, 75, 75, 75, 75, 75, 76, 76, 73, 73, 72, 73, 73, 73, 74, 74,
+        74, 75, 75, 75, 75, 76, 76, 76, 74, 73, 73, 73, 73, 74, 74, 74, 75, 75,
+        75, 75, 76, 76, 76, 76, 74, 73, 73, 73, 73, 74, 74, 74, 75, 75, 75, 75,
+        76, 76, 76, 76,
+        /* Size 32 */
+        64, 63, 62, 61, 60, 60, 60, 61, 61, 62, 63, 64, 65, 66, 67, 67, 68, 69,
+        70, 70, 71, 71, 72, 72, 73, 73, 73, 74, 74, 74, 74, 74, 63, 62, 62, 61,
+        60, 61, 61, 61, 61, 62, 63, 64, 65, 65, 66, 67, 68, 69, 69, 70, 70, 71,
+        71, 72, 72, 73, 73, 73, 74, 74, 74, 74, 62, 62, 61, 61, 61, 61, 61, 61,
+        61, 62, 63, 63, 64, 65, 66, 67, 68, 68, 69, 69, 70, 71, 71, 72, 72, 72,
+        73, 73, 73, 73, 73, 73, 61, 61, 61, 61, 61, 61, 61, 61, 61, 62, 63, 63,
+        64, 65, 65, 66, 67, 68, 68, 69, 70, 70, 71, 71, 72, 72, 72, 73, 73, 73,
+        73, 73, 60, 60, 61, 61, 62, 62, 62, 61, 61, 62, 62, 63, 64, 64, 65, 66,
+        67, 67, 68, 69, 69, 70, 70, 71, 71, 72, 72, 73, 73, 73, 73, 73, 60, 61,
+        61, 61, 62, 62, 62, 62, 62, 63, 63, 64, 65, 65, 66, 67, 67, 68, 69, 69,
+        70, 70, 71, 71, 72, 72, 72, 73, 73, 73, 73, 73, 60, 61, 61, 61, 62, 62,
+        63, 63, 64, 64, 65, 65, 66, 66, 67, 67, 68, 69, 69, 70, 70, 71, 71, 71,
+        72, 72, 73, 73, 73, 73, 73, 73, 61, 61, 61, 61, 61, 62, 63, 64, 65, 65,
+        66, 66, 67, 67, 68, 68, 69, 69, 70, 70, 71, 71, 71, 72, 72, 72, 73, 73,
+        73, 73, 73, 73, 61, 61, 61, 61, 61, 62, 64, 65, 66, 66, 67, 67, 68, 68,
+        68, 69, 69, 70, 70, 71, 71, 71, 72, 72, 72, 73, 73, 73, 73, 73, 73, 73,
+        62, 62, 62, 62, 62, 63, 64, 65, 66, 67, 67, 68, 68, 69, 69, 69, 70, 70,
+        71, 71, 71, 72, 72, 72, 73, 73, 73, 73, 74, 74, 74, 74, 63, 63, 63, 63,
+        62, 63, 65, 66, 67, 67, 68, 68, 69, 69, 70, 70, 70, 71, 71, 71, 72, 72,
+        72, 73, 73, 73, 73, 74, 74, 74, 74, 74, 64, 64, 63, 63, 63, 64, 65, 66,
+        67, 68, 68, 69, 69, 70, 70, 71, 71, 71, 72, 72, 72, 72, 73, 73, 73, 73,
+        74, 74, 74, 74, 74, 74, 65, 65, 64, 64, 64, 65, 66, 67, 68, 68, 69, 69,
+        70, 70, 71, 71, 71, 72, 72, 72, 72, 73, 73, 73, 73, 74, 74, 74, 74, 74,
+        74, 74, 66, 65, 65, 65, 64, 65, 66, 67, 68, 69, 69, 70, 70, 71, 71, 71,
+        72, 72, 72, 73, 73, 73, 73, 73, 74, 74, 74, 74, 74, 74, 74, 74, 67, 66,
+        66, 65, 65, 66, 67, 68, 68, 69, 70, 70, 71, 71, 71, 72, 72, 72, 73, 73,
+        73, 73, 73, 74, 74, 74, 74, 74, 74, 74, 74, 74, 67, 67, 67, 66, 66, 67,
+        67, 68, 69, 69, 70, 71, 71, 71, 72, 72, 72, 73, 73, 73, 73, 74, 74, 74,
+        74, 74, 74, 74, 75, 75, 75, 75, 68, 68, 68, 67, 67, 67, 68, 69, 69, 70,
+        70, 71, 71, 72, 72, 72, 73, 73, 73, 73, 74, 74, 74, 74, 74, 74, 74, 75,
+        75, 75, 75, 75, 69, 69, 68, 68, 67, 68, 69, 69, 70, 70, 71, 71, 72, 72,
+        72, 73, 73, 73, 73, 74, 74, 74, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75,
+        70, 69, 69, 68, 68, 69, 69, 70, 70, 71, 71, 72, 72, 72, 73, 73, 73, 73,
+        74, 74, 74, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 75, 70, 70, 69, 69,
+        69, 69, 70, 70, 71, 71, 71, 72, 72, 73, 73, 73, 73, 74, 74, 74, 74, 74,
+        74, 75, 75, 75, 75, 75, 75, 75, 75, 75, 71, 70, 70, 70, 69, 70, 70, 71,
+        71, 71, 72, 72, 72, 73, 73, 73, 74, 74, 74, 74, 74, 74, 75, 75, 75, 75,
+        75, 75, 75, 75, 75, 75, 71, 71, 71, 70, 70, 70, 71, 71, 71, 72, 72, 72,
+        73, 73, 73, 74, 74, 74, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 75, 75,
+        75, 75, 72, 71, 71, 71, 70, 71, 71, 71, 72, 72, 72, 73, 73, 73, 73, 74,
+        74, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 75, 75, 75, 75, 75, 72, 72,
+        72, 71, 71, 71, 71, 72, 72, 72, 73, 73, 73, 73, 74, 74, 74, 74, 74, 75,
+        75, 75, 75, 75, 75, 75, 75, 75, 75, 75, 75, 75, 73, 72, 72, 72, 71, 72,
+        72, 72, 72, 73, 73, 73, 73, 74, 74, 74, 74, 74, 75, 75, 75, 75, 75, 75,
+        75, 75, 75, 75, 76, 76, 76, 76, 73, 73, 72, 72, 72, 72, 72, 72, 73, 73,
+        73, 73, 74, 74, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 75, 75, 75, 76,
+        76, 76, 76, 76, 73, 73, 73, 72, 72, 72, 73, 73, 73, 73, 73, 74, 74, 74,
+        74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 75, 75, 76, 76, 76, 76, 76, 76,
+        74, 73, 73, 73, 73, 73, 73, 73, 73, 73, 74, 74, 74, 74, 74, 74, 75, 75,
+        75, 75, 75, 75, 75, 75, 75, 76, 76, 76, 76, 76, 76, 76, 74, 74, 73, 73,
+        73, 73, 73, 73, 73, 74, 74, 74, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75,
+        75, 75, 76, 76, 76, 76, 76, 76, 76, 76, 74, 74, 73, 73, 73, 73, 73, 73,
+        73, 74, 74, 74, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 75, 75, 76, 76,
+        76, 76, 76, 76, 76, 76, 74, 74, 73, 73, 73, 73, 73, 73, 73, 74, 74, 74,
+        74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 75, 75, 76, 76, 76, 76, 76, 76,
+        76, 76, 74, 74, 73, 73, 73, 73, 73, 73, 73, 74, 74, 74, 74, 74, 74, 75,
+        75, 75, 75, 75, 75, 75, 75, 75, 76, 76, 76, 76, 76, 76, 76, 76 },
+      { /* Intra matrices */
+        /* Size 4 */
+        55, 57, 62, 66, 57, 61, 65, 67, 62, 65, 67, 68, 66, 67, 68, 69,
+        /* Size 8 */
+        58, 54, 55, 59, 62, 64, 66, 67, 54, 56, 55, 57, 60, 63, 65, 66, 55, 55,
+        60, 61, 63, 64, 66, 67, 59, 57, 61, 64, 65, 66, 67, 68, 62, 60, 63, 65,
+        66, 67, 68, 68, 64, 63, 64, 66, 67, 68, 68, 69, 66, 65, 66, 67, 68, 68,
+        69, 69, 67, 66, 67, 68, 68, 69, 69, 69,
+        /* Size 16 */
+        57, 55, 53, 54, 55, 56, 58, 60, 62, 63, 64, 65, 66, 66, 67, 67, 55, 55,
+        54, 55, 55, 56, 58, 59, 61, 62, 63, 64, 65, 66, 66, 66, 53, 54, 55, 55,
+        55, 56, 57, 58, 60, 61, 63, 63, 64, 65, 66, 66, 54, 55, 55, 56, 57, 58,
+        59, 60, 61, 62, 63, 64, 65, 66, 66, 66, 55, 55, 55, 57, 59, 60, 61, 62,
+        62, 63, 64, 65, 65, 66, 66, 66, 56, 56, 56, 58, 60, 61, 62, 63, 63, 64,
+        65, 65, 66, 66, 67, 67, 58, 58, 57, 59, 61, 62, 63, 64, 65, 65, 66, 66,
+        66, 67, 67, 67, 60, 59, 58, 60, 62, 63, 64, 65, 65, 66, 66, 66, 67, 67,
+        67, 67, 62, 61, 60, 61, 62, 63, 65, 65, 66, 66, 67, 67, 67, 67, 68, 68,
+        63, 62, 61, 62, 63, 64, 65, 66, 66, 67, 67, 67, 67, 68, 68, 68, 64, 63,
+        63, 63, 64, 65, 66, 66, 67, 67, 67, 68, 68, 68, 68, 68, 65, 64, 63, 64,
+        65, 65, 66, 66, 67, 67, 68, 68, 68, 68, 68, 68, 66, 65, 64, 65, 65, 66,
+        66, 67, 67, 67, 68, 68, 68, 68, 68, 68, 66, 66, 65, 66, 66, 66, 67, 67,
+        67, 68, 68, 68, 68, 68, 69, 69, 67, 66, 66, 66, 66, 67, 67, 67, 68, 68,
+        68, 68, 68, 69, 69, 69, 67, 66, 66, 66, 66, 67, 67, 67, 68, 68, 68, 68,
+        68, 69, 69, 69,
+        /* Size 32 */
+        57, 56, 55, 54, 53, 54, 54, 54, 54, 55, 56, 57, 58, 59, 60, 60, 61, 62,
+        63, 63, 64, 64, 65, 65, 65, 66, 66, 66, 67, 67, 67, 67, 56, 56, 55, 54,
+        54, 54, 54, 54, 54, 55, 56, 57, 58, 58, 59, 60, 61, 62, 62, 63, 63, 64,
+        64, 65, 65, 65, 66, 66, 66, 66, 66, 66, 55, 55, 55, 54, 54, 54, 54, 54,
+        55, 55, 56, 57, 57, 58, 59, 60, 61, 61, 62, 62, 63, 63, 64, 64, 65, 65,
+        65, 66, 66, 66, 66, 66, 54, 54, 54, 55, 55, 55, 55, 55, 55, 55, 56, 56,
+        57, 58, 59, 59, 60, 61, 61, 62, 63, 63, 64, 64, 65, 65, 65, 66, 66, 66,
+        66, 66, 53, 54, 54, 55, 55, 55, 55, 55, 55, 55, 56, 56, 57, 57, 58, 59,
+        60, 60, 61, 62, 62, 63, 63, 64, 64, 65, 65, 65, 66, 66, 66, 66, 54, 54,
+        54, 55, 55, 55, 55, 56, 56, 56, 57, 57, 58, 58, 59, 60, 60, 61, 61, 62,
+        63, 63, 64, 64, 64, 65, 65, 65, 66, 66, 66, 66, 54, 54, 54, 55, 55, 55,
+        56, 56, 57, 57, 58, 58, 59, 59, 60, 60, 61, 61, 62, 63, 63, 63, 64, 64,
+        65, 65, 65, 66, 66, 66, 66, 66, 54, 54, 54, 55, 55, 56, 56, 57, 58, 58,
+        59, 59, 60, 60, 61, 61, 62, 62, 62, 63, 63, 64, 64, 65, 65, 65, 65, 66,
+        66, 66, 66, 66, 54, 54, 55, 55, 55, 56, 57, 58, 59, 59, 60, 60, 61, 61,
+        61, 62, 62, 63, 63, 63, 64, 64, 64, 65, 65, 65, 66, 66, 66, 66, 66, 66,
+        55, 55, 55, 55, 55, 56, 57, 58, 59, 60, 60, 61, 61, 62, 62, 62, 63, 63,
+        63, 64, 64, 64, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 56, 56, 56, 56,
+        56, 57, 58, 59, 60, 60, 61, 61, 62, 62, 63, 63, 63, 64, 64, 64, 65, 65,
+        65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 57, 57, 57, 56, 56, 57, 58, 59,
+        60, 61, 61, 62, 62, 63, 63, 63, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66,
+        66, 66, 67, 67, 67, 67, 58, 58, 57, 57, 57, 58, 59, 60, 61, 61, 62, 62,
+        63, 63, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 67, 67, 67,
+        67, 67, 59, 58, 58, 58, 57, 58, 59, 60, 61, 62, 62, 63, 63, 64, 64, 64,
+        65, 65, 65, 65, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 60, 59,
+        59, 59, 58, 59, 60, 61, 61, 62, 63, 63, 64, 64, 64, 65, 65, 65, 65, 66,
+        66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 67, 60, 60, 60, 59, 59, 60,
+        60, 61, 62, 62, 63, 63, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 66, 67,
+        67, 67, 67, 67, 67, 67, 67, 67, 61, 61, 61, 60, 60, 60, 61, 62, 62, 63,
+        63, 64, 64, 65, 65, 65, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67,
+        67, 67, 67, 67, 62, 62, 61, 61, 60, 61, 61, 62, 63, 63, 64, 64, 65, 65,
+        65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68,
+        63, 62, 62, 61, 61, 61, 62, 62, 63, 63, 64, 64, 65, 65, 65, 66, 66, 66,
+        66, 67, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 63, 63, 62, 62,
+        62, 62, 63, 63, 63, 64, 64, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67,
+        67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 64, 63, 63, 63, 62, 63, 63, 63,
+        64, 64, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 68, 68,
+        68, 68, 68, 68, 68, 68, 64, 64, 63, 63, 63, 63, 63, 64, 64, 64, 65, 65,
+        66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 65, 64, 64, 64, 63, 64, 64, 64, 64, 65, 65, 65, 66, 66, 66, 66,
+        67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 65, 65,
+        64, 64, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67,
+        67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 65, 65, 65, 65, 64, 64,
+        65, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68,
+        68, 68, 68, 68, 68, 68, 68, 68, 66, 65, 65, 65, 65, 65, 65, 65, 65, 66,
+        66, 66, 66, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 68, 68, 66, 66, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 67,
+        67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68,
+        66, 66, 66, 66, 65, 65, 66, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67,
+        68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 67, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68,
+        68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 67, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 68, 68, 68, 68, 67, 66, 66, 66, 66, 66, 66, 66, 66, 66, 67, 67,
+        67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 67, 66, 66, 66, 66, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67,
+        67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 71, 72, 74, 71, 73, 73, 75, 72, 73, 76, 77, 74, 75, 77, 78,
+        /* Size 8 */
+        64, 61, 69, 70, 71, 72, 73, 74, 61, 66, 69, 68, 69, 70, 71, 72, 69, 69,
+        71, 71, 71, 71, 72, 73, 70, 68, 71, 72, 73, 73, 74, 74, 71, 69, 71, 73,
+        74, 74, 75, 75, 72, 70, 71, 73, 74, 75, 75, 76, 73, 71, 72, 74, 75, 75,
+        76, 76, 74, 72, 73, 74, 75, 76, 76, 77,
+        /* Size 16 */
+        64, 62, 61, 65, 69, 70, 70, 70, 71, 71, 72, 72, 73, 74, 74, 74, 62, 63,
+        63, 66, 69, 69, 69, 69, 70, 70, 71, 71, 72, 73, 73, 73, 61, 63, 66, 68,
+        69, 69, 68, 68, 69, 69, 70, 70, 71, 72, 72, 72, 65, 66, 68, 69, 70, 70,
+        69, 70, 70, 70, 70, 71, 72, 72, 73, 73, 69, 69, 69, 70, 71, 71, 71, 71,
+        71, 71, 71, 72, 72, 73, 73, 73, 70, 69, 69, 70, 71, 71, 71, 72, 72, 72,
+        72, 72, 73, 73, 74, 74, 70, 69, 68, 69, 71, 71, 72, 72, 73, 73, 73, 73,
+        74, 74, 74, 74, 70, 69, 68, 70, 71, 72, 72, 73, 73, 73, 74, 74, 74, 74,
+        75, 75, 71, 70, 69, 70, 71, 72, 73, 73, 74, 74, 74, 74, 75, 75, 75, 75,
+        71, 70, 69, 70, 71, 72, 73, 73, 74, 74, 75, 75, 75, 75, 76, 76, 72, 71,
+        70, 70, 71, 72, 73, 74, 74, 75, 75, 75, 75, 76, 76, 76, 72, 71, 70, 71,
+        72, 72, 73, 74, 74, 75, 75, 75, 76, 76, 76, 76, 73, 72, 71, 72, 72, 73,
+        74, 74, 75, 75, 75, 76, 76, 76, 76, 76, 74, 73, 72, 72, 73, 73, 74, 74,
+        75, 75, 76, 76, 76, 76, 77, 77, 74, 73, 72, 73, 73, 74, 74, 75, 75, 76,
+        76, 76, 76, 77, 77, 77, 74, 73, 72, 73, 73, 74, 74, 75, 75, 76, 76, 76,
+        76, 77, 77, 77,
+        /* Size 32 */
+        64, 63, 62, 61, 61, 63, 65, 67, 69, 69, 70, 70, 70, 70, 70, 70, 71, 71,
+        71, 72, 72, 72, 72, 73, 73, 73, 74, 74, 74, 74, 74, 74, 63, 63, 62, 62,
+        62, 64, 65, 67, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 71, 71, 71, 72,
+        72, 72, 73, 73, 73, 73, 74, 74, 74, 74, 62, 62, 63, 63, 63, 65, 66, 68,
+        69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 71, 71, 71, 72, 72, 72,
+        73, 73, 73, 73, 73, 73, 61, 62, 63, 64, 65, 66, 67, 68, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 70, 70, 70, 71, 71, 71, 72, 72, 72, 72, 73, 73,
+        73, 73, 61, 62, 63, 65, 66, 67, 68, 69, 69, 69, 69, 68, 68, 68, 68, 68,
+        69, 69, 69, 69, 70, 70, 70, 71, 71, 71, 72, 72, 72, 72, 72, 72, 63, 64,
+        65, 66, 67, 68, 68, 69, 70, 70, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70,
+        70, 70, 71, 71, 71, 72, 72, 72, 73, 73, 73, 73, 65, 65, 66, 67, 68, 68,
+        69, 70, 70, 70, 70, 70, 69, 70, 70, 70, 70, 70, 70, 70, 70, 71, 71, 71,
+        72, 72, 72, 72, 73, 73, 73, 73, 67, 67, 68, 68, 69, 69, 70, 70, 71, 71,
+        70, 70, 70, 70, 70, 70, 70, 70, 71, 71, 71, 71, 71, 72, 72, 72, 72, 73,
+        73, 73, 73, 73, 69, 69, 69, 69, 69, 70, 70, 71, 71, 71, 71, 71, 71, 71,
+        71, 71, 71, 71, 71, 71, 71, 72, 72, 72, 72, 72, 73, 73, 73, 73, 73, 73,
+        69, 69, 69, 69, 69, 70, 70, 71, 71, 71, 71, 71, 71, 71, 71, 71, 71, 71,
+        71, 72, 72, 72, 72, 72, 73, 73, 73, 73, 73, 73, 73, 73, 70, 69, 69, 69,
+        69, 69, 70, 70, 71, 71, 71, 71, 71, 72, 72, 72, 72, 72, 72, 72, 72, 72,
+        72, 73, 73, 73, 73, 73, 74, 74, 74, 74, 70, 69, 69, 69, 68, 69, 70, 70,
+        71, 71, 71, 72, 72, 72, 72, 72, 72, 72, 72, 72, 73, 73, 73, 73, 73, 73,
+        74, 74, 74, 74, 74, 74, 70, 69, 69, 69, 68, 69, 69, 70, 71, 71, 71, 72,
+        72, 72, 72, 72, 73, 73, 73, 73, 73, 73, 73, 73, 74, 74, 74, 74, 74, 74,
+        74, 74, 70, 70, 69, 69, 68, 69, 70, 70, 71, 71, 72, 72, 72, 72, 73, 73,
+        73, 73, 73, 73, 73, 73, 74, 74, 74, 74, 74, 74, 74, 74, 74, 74, 70, 70,
+        69, 69, 68, 69, 70, 70, 71, 71, 72, 72, 72, 73, 73, 73, 73, 73, 73, 73,
+        74, 74, 74, 74, 74, 74, 74, 75, 75, 75, 75, 75, 70, 70, 69, 69, 68, 69,
+        70, 70, 71, 71, 72, 72, 72, 73, 73, 73, 73, 73, 74, 74, 74, 74, 74, 74,
+        74, 74, 75, 75, 75, 75, 75, 75, 71, 70, 70, 69, 69, 69, 70, 70, 71, 71,
+        72, 72, 73, 73, 73, 73, 74, 74, 74, 74, 74, 74, 74, 75, 75, 75, 75, 75,
+        75, 75, 75, 75, 71, 70, 70, 69, 69, 69, 70, 70, 71, 71, 72, 72, 73, 73,
+        73, 73, 74, 74, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 75, 75, 75, 75,
+        71, 71, 70, 70, 69, 70, 70, 71, 71, 71, 72, 72, 73, 73, 73, 74, 74, 74,
+        74, 74, 75, 75, 75, 75, 75, 75, 75, 75, 76, 76, 76, 76, 72, 71, 70, 70,
+        69, 70, 70, 71, 71, 72, 72, 72, 73, 73, 73, 74, 74, 74, 74, 75, 75, 75,
+        75, 75, 75, 75, 75, 76, 76, 76, 76, 76, 72, 71, 71, 70, 70, 70, 70, 71,
+        71, 72, 72, 73, 73, 73, 74, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 76,
+        76, 76, 76, 76, 76, 76, 72, 72, 71, 71, 70, 70, 71, 71, 72, 72, 72, 73,
+        73, 73, 74, 74, 74, 75, 75, 75, 75, 75, 75, 75, 76, 76, 76, 76, 76, 76,
+        76, 76, 72, 72, 71, 71, 70, 71, 71, 71, 72, 72, 72, 73, 73, 74, 74, 74,
+        74, 75, 75, 75, 75, 75, 75, 76, 76, 76, 76, 76, 76, 76, 76, 76, 73, 72,
+        72, 71, 71, 71, 71, 72, 72, 72, 73, 73, 73, 74, 74, 74, 75, 75, 75, 75,
+        75, 75, 76, 76, 76, 76, 76, 76, 76, 76, 76, 76, 73, 73, 72, 72, 71, 71,
+        72, 72, 72, 73, 73, 73, 74, 74, 74, 74, 75, 75, 75, 75, 75, 76, 76, 76,
+        76, 76, 76, 76, 76, 76, 76, 76, 73, 73, 72, 72, 71, 72, 72, 72, 72, 73,
+        73, 73, 74, 74, 74, 74, 75, 75, 75, 75, 76, 76, 76, 76, 76, 76, 76, 76,
+        77, 77, 77, 77, 74, 73, 73, 72, 72, 72, 72, 72, 73, 73, 73, 74, 74, 74,
+        74, 75, 75, 75, 75, 75, 76, 76, 76, 76, 76, 76, 76, 77, 77, 77, 77, 77,
+        74, 73, 73, 72, 72, 72, 72, 73, 73, 73, 73, 74, 74, 74, 75, 75, 75, 75,
+        75, 76, 76, 76, 76, 76, 76, 76, 77, 77, 77, 77, 77, 77, 74, 74, 73, 73,
+        72, 73, 73, 73, 73, 73, 74, 74, 74, 74, 75, 75, 75, 75, 76, 76, 76, 76,
+        76, 76, 76, 77, 77, 77, 77, 77, 77, 77, 74, 74, 73, 73, 72, 73, 73, 73,
+        73, 73, 74, 74, 74, 74, 75, 75, 75, 75, 76, 76, 76, 76, 76, 76, 76, 77,
+        77, 77, 77, 77, 77, 77, 74, 74, 73, 73, 72, 73, 73, 73, 73, 73, 74, 74,
+        74, 74, 75, 75, 75, 75, 76, 76, 76, 76, 76, 76, 76, 77, 77, 77, 77, 77,
+        77, 77, 74, 74, 73, 73, 72, 73, 73, 73, 73, 73, 74, 74, 74, 74, 75, 75,
+        75, 75, 76, 76, 76, 76, 76, 76, 76, 77, 77, 77, 77, 77, 77, 77 },
+      { /* Intra matrices */
+        /* Size 4 */
+        55, 61, 62, 65, 61, 63, 64, 65, 62, 64, 66, 67, 65, 65, 67, 68,
+        /* Size 8 */
+        57, 53, 61, 62, 63, 64, 65, 66, 53, 59, 62, 60, 61, 62, 63, 64, 61, 62,
+        63, 63, 63, 63, 64, 65, 62, 60, 63, 64, 65, 65, 66, 66, 63, 61, 63, 65,
+        66, 66, 67, 67, 64, 62, 63, 65, 66, 67, 67, 68, 65, 63, 64, 66, 67, 67,
+        68, 68, 66, 64, 65, 66, 67, 68, 68, 69,
+        /* Size 16 */
+        56, 55, 53, 57, 61, 61, 62, 62, 62, 63, 64, 64, 65, 65, 66, 66, 55, 55,
+        56, 58, 61, 61, 61, 61, 61, 62, 63, 63, 64, 64, 65, 65, 53, 56, 58, 60,
+        61, 61, 60, 60, 61, 61, 62, 62, 63, 63, 64, 64, 57, 58, 60, 61, 62, 62,
+        61, 61, 62, 62, 62, 63, 63, 64, 64, 64, 61, 61, 61, 62, 63, 63, 63, 63,
+        63, 63, 63, 64, 64, 64, 65, 65, 61, 61, 61, 62, 63, 63, 63, 63, 63, 64,
+        64, 64, 65, 65, 65, 65, 62, 61, 60, 61, 63, 63, 64, 64, 64, 64, 65, 65,
+        65, 65, 66, 66, 62, 61, 60, 61, 63, 63, 64, 64, 65, 65, 65, 65, 66, 66,
+        66, 66, 62, 61, 61, 62, 63, 63, 64, 65, 65, 66, 66, 66, 66, 66, 67, 67,
+        63, 62, 61, 62, 63, 64, 64, 65, 66, 66, 66, 66, 67, 67, 67, 67, 64, 63,
+        62, 62, 63, 64, 65, 65, 66, 66, 67, 67, 67, 67, 67, 67, 64, 63, 62, 63,
+        64, 64, 65, 65, 66, 66, 67, 67, 67, 68, 68, 68, 65, 64, 63, 63, 64, 65,
+        65, 66, 66, 67, 67, 67, 68, 68, 68, 68, 65, 64, 63, 64, 64, 65, 65, 66,
+        66, 67, 67, 68, 68, 68, 68, 68, 66, 65, 64, 64, 65, 65, 66, 66, 67, 67,
+        67, 68, 68, 68, 68, 68, 66, 65, 64, 64, 65, 65, 66, 66, 67, 67, 67, 68,
+        68, 68, 68, 68,
+        /* Size 32 */
+        56, 55, 55, 54, 53, 55, 57, 59, 61, 61, 61, 61, 62, 62, 62, 62, 62, 63,
+        63, 63, 63, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 55, 55, 55, 54,
+        54, 56, 57, 59, 61, 61, 61, 61, 61, 61, 62, 62, 62, 62, 62, 63, 63, 63,
+        63, 64, 64, 64, 65, 65, 65, 65, 65, 65, 55, 55, 55, 55, 55, 57, 58, 60,
+        61, 61, 61, 61, 61, 61, 61, 61, 61, 62, 62, 62, 62, 63, 63, 63, 64, 64,
+        64, 64, 65, 65, 65, 65, 54, 54, 55, 56, 57, 58, 59, 60, 61, 61, 61, 61,
+        60, 61, 61, 61, 61, 61, 61, 62, 62, 62, 63, 63, 63, 63, 64, 64, 64, 64,
+        64, 64, 53, 54, 55, 57, 58, 59, 60, 60, 61, 61, 61, 60, 60, 60, 60, 60,
+        60, 61, 61, 61, 61, 62, 62, 62, 63, 63, 63, 64, 64, 64, 64, 64, 55, 56,
+        57, 58, 59, 60, 60, 61, 62, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 62,
+        62, 62, 62, 63, 63, 63, 64, 64, 64, 64, 64, 64, 57, 57, 58, 59, 60, 60,
+        61, 61, 62, 62, 62, 61, 61, 61, 61, 61, 61, 62, 62, 62, 62, 62, 63, 63,
+        63, 64, 64, 64, 64, 64, 64, 64, 59, 59, 60, 60, 60, 61, 61, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 63, 63, 63, 63, 64, 64, 64, 64,
+        64, 64, 64, 64, 61, 61, 61, 61, 61, 62, 62, 62, 63, 63, 63, 63, 62, 62,
+        62, 62, 62, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 65, 65, 65, 65,
+        61, 61, 61, 61, 61, 61, 62, 62, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 63, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 61, 61, 61, 61,
+        61, 61, 62, 62, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64,
+        64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 61, 61, 61, 61, 60, 61, 61, 62,
+        63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 62, 61, 61, 60, 60, 61, 61, 62, 62, 63, 63, 63,
+        64, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66,
+        66, 66, 62, 61, 61, 61, 60, 61, 61, 62, 62, 63, 63, 63, 64, 64, 64, 64,
+        64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 62, 62,
+        61, 61, 60, 61, 61, 62, 62, 63, 63, 64, 64, 64, 64, 64, 65, 65, 65, 65,
+        65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 62, 62, 61, 61, 60, 61,
+        61, 62, 62, 63, 63, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 66, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 62, 62, 61, 61, 60, 61, 61, 62, 62, 63,
+        63, 64, 64, 64, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        67, 67, 67, 67, 63, 62, 62, 61, 61, 61, 62, 62, 63, 63, 63, 64, 64, 64,
+        65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67,
+        63, 62, 62, 61, 61, 61, 62, 62, 63, 63, 64, 64, 64, 65, 65, 65, 65, 66,
+        66, 66, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 63, 63, 62, 62,
+        61, 62, 62, 62, 63, 63, 64, 64, 64, 65, 65, 65, 66, 66, 66, 66, 66, 66,
+        66, 67, 67, 67, 67, 67, 67, 67, 67, 67, 63, 63, 62, 62, 61, 62, 62, 63,
+        63, 63, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67,
+        67, 67, 67, 67, 67, 67, 64, 63, 63, 62, 62, 62, 62, 63, 63, 64, 64, 64,
+        65, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67,
+        67, 67, 64, 63, 63, 63, 62, 62, 63, 63, 63, 64, 64, 64, 65, 65, 65, 66,
+        66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 64, 64,
+        63, 63, 62, 63, 63, 63, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 67,
+        67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 65, 64, 64, 63, 63, 63,
+        63, 64, 64, 64, 64, 65, 65, 65, 66, 66, 66, 66, 66, 67, 67, 67, 67, 67,
+        67, 68, 68, 68, 68, 68, 68, 68, 65, 64, 64, 63, 63, 63, 64, 64, 64, 64,
+        65, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68,
+        68, 68, 68, 68, 65, 65, 64, 64, 63, 64, 64, 64, 64, 65, 65, 65, 65, 66,
+        66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68,
+        65, 65, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 67,
+        67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 66, 65, 65, 64,
+        64, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67,
+        68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 66, 65, 65, 64, 64, 64, 64, 64,
+        65, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68,
+        68, 68, 68, 68, 68, 68, 66, 65, 65, 64, 64, 64, 64, 64, 65, 65, 65, 65,
+        66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 66, 65, 65, 64, 64, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66,
+        67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 65, 68, 70, 65, 67, 69, 70, 68, 69, 70, 71, 70, 70, 71, 71,
+        /* Size 8 */
+        64, 62, 62, 64, 66, 67, 68, 69, 62, 63, 63, 64, 65, 67, 68, 68, 62, 63,
+        65, 66, 67, 67, 68, 69, 64, 64, 66, 67, 68, 68, 69, 69, 66, 65, 67, 68,
+        68, 69, 69, 69, 67, 67, 67, 68, 69, 69, 69, 69, 68, 68, 68, 69, 69, 69,
+        69, 69, 69, 68, 69, 69, 69, 69, 69, 70,
+        /* Size 16 */
+        64, 63, 62, 62, 62, 63, 64, 65, 66, 67, 67, 68, 68, 68, 69, 69, 63, 63,
+        62, 62, 63, 63, 64, 65, 66, 66, 67, 67, 68, 68, 69, 69, 62, 62, 63, 63,
+        63, 63, 64, 65, 65, 66, 67, 67, 68, 68, 68, 68, 62, 62, 63, 63, 64, 64,
+        65, 65, 66, 67, 67, 67, 68, 68, 68, 68, 62, 63, 63, 64, 65, 65, 66, 66,
+        67, 67, 67, 68, 68, 68, 69, 69, 63, 63, 63, 64, 65, 66, 66, 67, 67, 67,
+        68, 68, 68, 68, 69, 69, 64, 64, 64, 65, 66, 66, 67, 67, 68, 68, 68, 68,
+        69, 69, 69, 69, 65, 65, 65, 65, 66, 67, 67, 68, 68, 68, 68, 69, 69, 69,
+        69, 69, 66, 66, 65, 66, 67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69,
+        67, 66, 66, 67, 67, 67, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 67, 67,
+        67, 67, 67, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 68, 67, 67, 67,
+        68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 68, 68, 68, 68, 68, 68,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 68, 68, 68, 68, 68, 68, 69, 69,
+        69, 69, 69, 69, 69, 69, 70, 70, 69, 69, 68, 68, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 70, 70, 70, 69, 69, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 70, 70, 70,
+        /* Size 32 */
+        64, 63, 63, 62, 62, 62, 62, 62, 62, 63, 63, 64, 64, 65, 65, 66, 66, 66,
+        67, 67, 67, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 63, 63, 63, 62,
+        62, 62, 62, 62, 62, 63, 63, 64, 64, 65, 65, 66, 66, 66, 67, 67, 67, 67,
+        68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 63, 63, 63, 62, 62, 62, 62, 62,
+        63, 63, 63, 64, 64, 65, 65, 65, 66, 66, 66, 67, 67, 67, 67, 68, 68, 68,
+        68, 68, 69, 69, 69, 69, 62, 62, 62, 63, 63, 63, 63, 63, 63, 63, 63, 64,
+        64, 64, 65, 65, 66, 66, 66, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 62, 62, 62, 63, 63, 63, 63, 63, 63, 63, 63, 63, 64, 64, 65, 65,
+        65, 66, 66, 66, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 62, 62,
+        62, 63, 63, 63, 63, 63, 63, 63, 64, 64, 64, 65, 65, 65, 66, 66, 66, 67,
+        67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 62, 62, 62, 63, 63, 63,
+        63, 63, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66, 67, 67, 67, 67, 67, 68,
+        68, 68, 68, 68, 68, 68, 68, 68, 62, 62, 62, 63, 63, 63, 63, 64, 64, 65,
+        65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68,
+        68, 68, 68, 68, 62, 62, 63, 63, 63, 63, 64, 64, 65, 65, 65, 66, 66, 66,
+        66, 66, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69,
+        63, 63, 63, 63, 63, 63, 64, 65, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67,
+        67, 67, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 63, 63, 63, 63,
+        63, 64, 64, 65, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67, 67, 68, 68, 68,
+        68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 64, 64, 64, 64, 63, 64, 65, 65,
+        66, 66, 66, 66, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68,
+        69, 69, 69, 69, 69, 69, 64, 64, 64, 64, 64, 64, 65, 65, 66, 66, 66, 67,
+        67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69,
+        69, 69, 65, 65, 65, 64, 64, 65, 65, 66, 66, 66, 67, 67, 67, 67, 67, 68,
+        68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 65, 65,
+        65, 65, 65, 65, 65, 66, 66, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68,
+        68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 66, 66, 65, 65, 65, 65,
+        66, 66, 66, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 66, 66, 66, 66, 65, 66, 66, 66, 67, 67,
+        67, 67, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 69, 66, 66, 66, 66, 66, 66, 66, 67, 67, 67, 67, 68, 68, 68,
+        68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        67, 67, 66, 66, 66, 66, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 67, 67, 67, 67,
+        66, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 67, 67, 67, 67, 67, 67, 67, 67,
+        67, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 68, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68,
+        68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 68, 68, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 68, 68,
+        68, 68, 67, 68, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 68, 68, 68, 68, 68, 68,
+        68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        70, 70, 70, 70, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70,
+        69, 68, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 69, 69, 69, 68,
+        68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 69, 69, 69, 68, 68, 68, 68, 68,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 70,
+        70, 70, 70, 70, 70, 70, 69, 69, 69, 68, 68, 68, 68, 68, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70,
+        70, 70, 69, 69, 69, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70 },
+      { /* Intra matrices */
+        /* Size 4 */
+        59, 60, 63, 65, 60, 63, 64, 65, 63, 64, 66, 66, 65, 65, 66, 66,
+        /* Size 8 */
+        61, 59, 59, 61, 63, 64, 65, 66, 59, 60, 59, 61, 62, 64, 65, 65, 59, 59,
+        62, 63, 64, 64, 65, 65, 61, 61, 63, 64, 65, 65, 65, 66, 63, 62, 64, 65,
+        65, 66, 66, 66, 64, 64, 64, 65, 66, 66, 66, 66, 65, 65, 65, 65, 66, 66,
+        66, 66, 66, 65, 65, 66, 66, 66, 66, 67,
+        /* Size 16 */
+        61, 60, 58, 59, 59, 60, 61, 62, 63, 63, 64, 64, 65, 65, 65, 65, 60, 59,
+        59, 59, 59, 60, 61, 62, 62, 63, 64, 64, 65, 65, 65, 65, 58, 59, 60, 59,
+        59, 60, 60, 61, 62, 63, 63, 64, 64, 65, 65, 65, 59, 59, 59, 60, 60, 61,
+        61, 62, 63, 63, 64, 64, 65, 65, 65, 65, 59, 59, 59, 60, 62, 62, 63, 63,
+        63, 64, 64, 64, 65, 65, 65, 65, 60, 60, 60, 61, 62, 63, 63, 63, 64, 64,
+        64, 65, 65, 65, 65, 65, 61, 61, 60, 61, 63, 63, 64, 64, 64, 65, 65, 65,
+        65, 65, 66, 66, 62, 62, 61, 62, 63, 63, 64, 64, 65, 65, 65, 65, 65, 66,
+        66, 66, 63, 62, 62, 63, 63, 64, 64, 65, 65, 65, 65, 65, 66, 66, 66, 66,
+        63, 63, 63, 63, 64, 64, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 64, 64,
+        63, 64, 64, 64, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 64, 64, 64, 64,
+        64, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 65, 65, 64, 65, 65, 65,
+        65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 65, 65, 65, 65, 65, 65, 65, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66,
+        66, 66, 66, 66,
+        /* Size 32 */
+        61, 60, 59, 59, 58, 59, 59, 59, 59, 60, 60, 61, 61, 61, 62, 62, 63, 63,
+        63, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 60, 60, 59, 59,
+        59, 59, 59, 59, 59, 59, 60, 60, 61, 61, 62, 62, 63, 63, 63, 63, 64, 64,
+        64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 60, 60, 61, 61, 62, 62, 62, 63, 63, 63, 64, 64, 64, 64, 65, 65,
+        65, 65, 65, 65, 65, 65, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60,
+        61, 61, 61, 62, 62, 62, 63, 63, 63, 64, 64, 64, 64, 65, 65, 65, 65, 65,
+        65, 65, 58, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 60, 61, 61, 62,
+        62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 59, 59,
+        59, 59, 59, 59, 60, 60, 60, 60, 60, 61, 61, 61, 62, 62, 62, 63, 63, 63,
+        63, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 59, 59, 59, 59, 59, 60,
+        60, 60, 60, 61, 61, 61, 61, 62, 62, 62, 63, 63, 63, 63, 64, 64, 64, 64,
+        64, 65, 65, 65, 65, 65, 65, 65, 59, 59, 59, 59, 59, 60, 60, 60, 61, 61,
+        61, 62, 62, 62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64, 65, 65, 65, 65,
+        65, 65, 65, 65, 59, 59, 59, 59, 59, 60, 60, 61, 62, 62, 62, 62, 62, 63,
+        63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65,
+        60, 59, 59, 59, 59, 60, 61, 61, 62, 62, 62, 62, 63, 63, 63, 63, 63, 64,
+        64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 60, 60, 60, 60,
+        60, 60, 61, 61, 62, 62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64,
+        65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 61, 60, 60, 60, 60, 61, 61, 62,
+        62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 61, 61, 61, 61, 60, 61, 61, 62, 62, 63, 63, 63,
+        64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 65, 61, 61, 61, 61, 61, 61, 62, 62, 63, 63, 63, 63, 64, 64, 64, 64,
+        64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 62, 62,
+        62, 61, 61, 62, 62, 62, 63, 63, 63, 64, 64, 64, 64, 64, 65, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 62, 62, 62, 62, 62, 62,
+        62, 63, 63, 63, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 65, 66, 66, 66, 66, 66, 66, 63, 63, 62, 62, 62, 62, 63, 63, 63, 63,
+        64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66,
+        66, 66, 66, 66, 63, 63, 63, 62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 65,
+        65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        63, 63, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65,
+        65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 64, 63, 63, 63,
+        63, 63, 63, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 64, 64, 64, 63, 63, 63, 64, 64,
+        64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 64, 64, 64, 64, 63, 64, 64, 64, 64, 64, 64, 65,
+        65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65,
+        65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 65, 64,
+        64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 65, 65, 65, 64, 64, 64,
+        64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 65, 65, 65, 65, 64, 64, 65, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 65, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 67, 68, 69, 67, 68, 69, 69, 68, 69, 70, 70, 69, 69, 70, 71,
+        /* Size 8 */
+        64, 62, 67, 67, 67, 68, 68, 69, 62, 65, 67, 66, 66, 67, 68, 68, 67, 67,
+        68, 67, 67, 68, 68, 69, 67, 66, 67, 68, 68, 68, 69, 69, 67, 66, 67, 68,
+        69, 69, 69, 69, 68, 67, 68, 68, 69, 69, 70, 70, 68, 68, 68, 69, 69, 70,
+        70, 70, 69, 68, 69, 69, 69, 70, 70, 70,
+        /* Size 16 */
+        64, 63, 62, 64, 67, 67, 67, 67, 67, 68, 68, 68, 68, 69, 69, 69, 63, 63,
+        64, 65, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 69, 69, 62, 64, 65, 66,
+        67, 66, 66, 66, 66, 67, 67, 67, 68, 68, 68, 68, 64, 65, 66, 67, 67, 67,
+        67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 67, 67, 67, 67, 68, 67, 67, 67,
+        67, 68, 68, 68, 68, 68, 69, 69, 67, 67, 66, 67, 67, 68, 68, 68, 68, 68,
+        68, 68, 68, 69, 69, 69, 67, 67, 66, 67, 67, 68, 68, 68, 68, 68, 68, 69,
+        69, 69, 69, 69, 67, 67, 66, 67, 67, 68, 68, 68, 68, 69, 69, 69, 69, 69,
+        69, 69, 67, 67, 66, 67, 67, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69,
+        68, 67, 67, 67, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 70, 70, 68, 67,
+        67, 67, 68, 68, 68, 69, 69, 69, 69, 69, 70, 70, 70, 70, 68, 68, 67, 68,
+        68, 68, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 68, 68, 68, 68, 68, 68,
+        69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 69, 68, 68, 68, 68, 69, 69, 69,
+        69, 69, 70, 70, 70, 70, 70, 70, 69, 69, 68, 68, 69, 69, 69, 69, 69, 70,
+        70, 70, 70, 70, 70, 70, 69, 69, 68, 68, 69, 69, 69, 69, 69, 70, 70, 70,
+        70, 70, 70, 70,
+        /* Size 32 */
+        64, 64, 63, 63, 62, 63, 64, 65, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67,
+        68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 64, 63, 63, 63,
+        63, 64, 65, 66, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 68, 68,
+        68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 63, 63, 63, 63, 64, 64, 65, 66,
+        67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68,
+        68, 68, 69, 69, 69, 69, 63, 63, 63, 64, 64, 65, 66, 66, 67, 67, 67, 66,
+        66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 62, 63, 64, 64, 65, 66, 66, 66, 67, 67, 66, 66, 66, 66, 66, 66,
+        66, 66, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 63, 64,
+        64, 65, 66, 66, 66, 67, 67, 67, 67, 67, 66, 66, 67, 67, 67, 67, 67, 67,
+        67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 64, 65, 65, 66, 66, 66,
+        67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 68, 68,
+        68, 68, 68, 68, 68, 68, 68, 68, 65, 66, 66, 66, 66, 67, 67, 67, 67, 67,
+        67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 68, 68, 67, 67, 67, 67, 67, 67, 67, 67, 68, 68, 67, 67, 67, 67,
+        67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69,
+        67, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 67, 67, 67, 67,
+        66, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 67, 67, 67, 66, 66, 67, 67, 67,
+        67, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 69, 69,
+        69, 69, 69, 69, 69, 69, 67, 67, 67, 66, 66, 66, 67, 67, 67, 68, 68, 68,
+        68, 68, 68, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 67, 67, 67, 66, 66, 66, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68,
+        68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 67, 67,
+        67, 66, 66, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 67, 67, 67, 67, 66, 67,
+        67, 67, 67, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 67, 67, 67, 67, 66, 67, 67, 67, 67, 68,
+        68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 69, 67, 67, 67, 67, 66, 67, 67, 67, 67, 68, 68, 68, 68, 68,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        68, 67, 67, 67, 67, 67, 67, 67, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 68, 67, 67, 67,
+        67, 67, 67, 67, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 68, 68, 67, 67, 67, 67, 67, 67,
+        68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70,
+        70, 70, 70, 70, 70, 70, 68, 68, 68, 67, 67, 67, 67, 68, 68, 68, 68, 68,
+        68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70,
+        70, 70, 68, 68, 68, 67, 67, 67, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 68, 68,
+        68, 68, 67, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69,
+        70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 68, 68, 68, 68, 68, 68,
+        68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70,
+        70, 70, 70, 70, 70, 70, 70, 70, 69, 68, 68, 68, 68, 68, 68, 68, 68, 68,
+        68, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70, 70,
+        70, 70, 70, 70, 69, 68, 68, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70,
+        69, 69, 68, 68, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69,
+        70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 69, 69, 69, 68,
+        68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70,
+        70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 69, 69, 69, 68, 68, 68, 68, 68,
+        69, 69, 69, 69, 69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70,
+        70, 70, 70, 70, 70, 70, 69, 69, 69, 68, 68, 68, 68, 68, 69, 69, 69, 69,
+        69, 69, 69, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70,
+        70, 70, 69, 69, 69, 68, 68, 68, 68, 68, 69, 69, 69, 69, 69, 69, 69, 69,
+        69, 69, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70, 70 },
+      { /* Intra matrices */
+        /* Size 4 */
+        59, 63, 63, 64, 63, 64, 64, 65, 63, 64, 65, 66, 64, 65, 66, 66,
+        /* Size 8 */
+        60, 58, 63, 63, 63, 64, 65, 65, 58, 61, 63, 62, 62, 63, 64, 64, 63, 63,
+        64, 64, 64, 64, 64, 65, 63, 62, 64, 64, 64, 65, 65, 65, 63, 62, 64, 64,
+        65, 65, 65, 66, 64, 63, 64, 65, 65, 65, 66, 66, 65, 64, 64, 65, 65, 66,
+        66, 66, 65, 64, 65, 65, 66, 66, 66, 66,
+        /* Size 16 */
+        60, 59, 58, 60, 63, 63, 63, 63, 63, 64, 64, 64, 64, 65, 65, 65, 59, 59,
+        60, 61, 63, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 58, 60, 61, 62,
+        63, 62, 62, 62, 62, 63, 63, 63, 63, 64, 64, 64, 60, 61, 62, 62, 63, 63,
+        63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 63, 63, 63, 63, 64, 63, 63, 63,
+        63, 63, 64, 64, 64, 64, 64, 64, 63, 63, 62, 63, 63, 64, 64, 64, 64, 64,
+        64, 64, 64, 65, 65, 65, 63, 63, 62, 63, 63, 64, 64, 64, 64, 64, 64, 65,
+        65, 65, 65, 65, 63, 63, 62, 63, 63, 64, 64, 64, 64, 65, 65, 65, 65, 65,
+        65, 65, 63, 63, 62, 63, 63, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65,
+        64, 63, 63, 63, 63, 64, 64, 65, 65, 65, 65, 65, 65, 65, 66, 66, 64, 63,
+        63, 63, 64, 64, 64, 65, 65, 65, 65, 65, 66, 66, 66, 66, 64, 64, 63, 63,
+        64, 64, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 64, 64, 63, 64, 64, 64,
+        65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 65, 64, 64, 64, 64, 65, 65, 65,
+        65, 65, 66, 66, 66, 66, 66, 66, 65, 64, 64, 64, 64, 65, 65, 65, 65, 66,
+        66, 66, 66, 66, 66, 66, 65, 64, 64, 64, 64, 65, 65, 65, 65, 66, 66, 66,
+        66, 66, 66, 66,
+        /* Size 32 */
+        60, 59, 59, 59, 58, 59, 60, 61, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 59, 59, 59, 59,
+        59, 60, 61, 62, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 64, 64,
+        64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 59, 59, 59, 59, 60, 60, 61, 62,
+        63, 63, 62, 62, 62, 62, 63, 63, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 59, 59, 59, 60, 60, 61, 61, 62, 63, 62, 62, 62,
+        62, 62, 62, 62, 62, 63, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64,
+        64, 64, 58, 59, 60, 60, 61, 61, 62, 62, 63, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 59, 60,
+        60, 61, 61, 62, 62, 62, 63, 63, 63, 62, 62, 62, 62, 62, 62, 63, 63, 63,
+        63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 60, 61, 61, 61, 62, 62,
+        62, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 61, 62, 62, 62, 62, 62, 63, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        63, 63, 63, 62, 62, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 63, 63, 62, 62,
+        62, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 63, 63, 62, 62, 62, 62, 63, 63,
+        63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        65, 65, 65, 65, 65, 65, 63, 63, 62, 62, 62, 62, 63, 63, 63, 63, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65,
+        65, 65, 63, 63, 62, 62, 62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 63, 63,
+        63, 62, 62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 65,
+        65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 63, 63, 63, 62, 62, 62,
+        63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 65, 65, 63, 63, 63, 62, 62, 62, 63, 63, 63, 63,
+        64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 65, 65, 65, 63, 63, 63, 63, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64,
+        64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65,
+        63, 63, 63, 63, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 64, 63, 63, 63,
+        63, 63, 63, 63, 63, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 64, 64, 63, 63, 63, 63, 63, 63,
+        64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65,
+        66, 66, 66, 66, 66, 66, 64, 64, 63, 63, 63, 63, 63, 63, 64, 64, 64, 64,
+        64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66,
+        66, 66, 64, 64, 64, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 64, 64,
+        64, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65,
+        65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 64, 64, 64, 64, 63, 64,
+        64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 66, 66, 65, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        65, 64, 64, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 65, 65, 64, 64,
+        64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 65, 65, 64, 64, 64, 64, 64, 64,
+        64, 64, 65, 65, 65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 65, 65, 64, 64, 64, 64, 64, 64, 64, 64, 65, 65,
+        65, 65, 65, 65, 65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 65, 65, 64, 64, 64, 64, 64, 64, 64, 64, 65, 65, 65, 65, 65, 65,
+        65, 65, 65, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 8 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 16 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64,
+        /* Size 32 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 },
+      { /* Intra matrices */
+        /* Size 4 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 8 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 16 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64,
+        /* Size 32 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 8 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 16 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64,
+        /* Size 32 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 },
+      { /* Intra matrices */
+        /* Size 4 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 8 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 16 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64,
+        /* Size 32 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 } } }
+};
+
+static uint16_t wt_matrix_ref[NUM_QM_LEVELS][2][2][4 * 4 + 8 * 8 + 16 * 16 +
+                                                   32 * 32] = {
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 58, 33, 19, 58, 37, 25, 17, 33, 25, 16, 12, 19, 17, 12, 10,
+        /* Size 8 */
+        64, 87, 80, 59, 42, 31, 24, 19, 87, 75, 79, 66, 50, 38, 29, 23, 80, 79,
+        54, 46, 38, 31, 25, 20, 59, 66, 46, 34, 29, 24, 21, 18, 42, 50, 38, 29,
+        23, 20, 17, 15, 31, 38, 31, 24, 20, 17, 15, 13, 24, 29, 25, 21, 17, 15,
+        13, 12, 19, 23, 20, 18, 15, 13, 12, 11,
+        /* Size 16 */
+        64, 76, 87, 84, 80, 70, 59, 51, 42, 37, 31, 27, 24, 21, 19, 19, 76, 79,
+        81, 81, 80, 71, 63, 55, 46, 40, 34, 30, 26, 23, 21, 21, 87, 81, 75, 77,
+        79, 73, 66, 58, 50, 44, 38, 33, 29, 26, 23, 23, 84, 81, 77, 72, 67, 61,
+        56, 50, 44, 39, 34, 31, 27, 24, 21, 21, 80, 80, 79, 67, 54, 50, 46, 42,
+        38, 34, 31, 28, 25, 23, 20, 20, 70, 71, 73, 61, 50, 45, 40, 37, 33, 30,
+        28, 25, 23, 21, 19, 19, 59, 63, 66, 56, 46, 40, 34, 31, 29, 26, 24, 22,
+        21, 19, 18, 18, 51, 55, 58, 50, 42, 37, 31, 29, 26, 24, 22, 20, 19, 18,
+        16, 16, 42, 46, 50, 44, 38, 33, 29, 26, 23, 21, 20, 18, 17, 16, 15, 15,
+        37, 40, 44, 39, 34, 30, 26, 24, 21, 20, 18, 17, 16, 15, 14, 14, 31, 34,
+        38, 34, 31, 28, 24, 22, 20, 18, 17, 16, 15, 14, 13, 13, 27, 30, 33, 31,
+        28, 25, 22, 20, 18, 17, 16, 15, 14, 13, 13, 13, 24, 26, 29, 27, 25, 23,
+        21, 19, 17, 16, 15, 14, 13, 13, 12, 12, 21, 23, 26, 24, 23, 21, 19, 18,
+        16, 15, 14, 13, 13, 12, 12, 12, 19, 21, 23, 21, 20, 19, 18, 16, 15, 14,
+        13, 13, 12, 12, 11, 11, 19, 21, 23, 21, 20, 19, 18, 16, 15, 14, 13, 13,
+        12, 12, 11, 11,
+        /* Size 32 */
+        64, 70, 76, 82, 87, 86, 84, 82, 80, 75, 70, 65, 59, 55, 51, 47, 42, 40,
+        37, 34, 31, 29, 27, 26, 24, 22, 21, 20, 19, 19, 19, 19, 70, 74, 77, 81,
+        84, 83, 82, 81, 80, 75, 71, 66, 61, 57, 53, 49, 44, 41, 39, 36, 33, 31,
+        29, 27, 25, 24, 22, 21, 20, 20, 20, 20, 76, 77, 79, 80, 81, 81, 81, 80,
+        80, 75, 71, 67, 63, 59, 55, 51, 46, 43, 40, 37, 34, 32, 30, 28, 26, 25,
+        23, 22, 21, 21, 21, 21, 82, 81, 80, 79, 78, 79, 79, 79, 79, 76, 72, 68,
+        65, 61, 56, 52, 48, 45, 42, 39, 36, 34, 32, 30, 27, 26, 25, 23, 22, 22,
+        22, 22, 87, 84, 81, 78, 75, 76, 77, 78, 79, 76, 73, 70, 66, 62, 58, 54,
+        50, 47, 44, 41, 38, 36, 33, 31, 29, 27, 26, 24, 23, 23, 23, 23, 86, 83,
+        81, 79, 76, 75, 75, 74, 73, 70, 67, 64, 61, 58, 54, 51, 47, 44, 42, 39,
+        36, 34, 32, 30, 28, 26, 25, 23, 22, 22, 22, 22, 84, 82, 81, 79, 77, 75,
+        72, 69, 67, 64, 61, 59, 56, 53, 50, 47, 44, 42, 39, 37, 34, 32, 31, 29,
+        27, 25, 24, 23, 21, 21, 21, 21, 82, 81, 80, 79, 78, 74, 69, 65, 61, 58,
+        56, 53, 51, 48, 46, 44, 41, 39, 37, 35, 33, 31, 29, 28, 26, 25, 23, 22,
+        21, 21, 21, 21, 80, 80, 80, 79, 79, 73, 67, 61, 54, 52, 50, 48, 46, 44,
+        42, 40, 38, 36, 34, 33, 31, 29, 28, 26, 25, 24, 23, 22, 20, 20, 20, 20,
+        75, 75, 75, 76, 76, 70, 64, 58, 52, 50, 47, 45, 43, 41, 39, 37, 36, 34,
+        32, 31, 29, 28, 27, 25, 24, 23, 22, 21, 20, 20, 20, 20, 70, 71, 71, 72,
+        73, 67, 61, 56, 50, 47, 45, 42, 40, 38, 37, 35, 33, 32, 30, 29, 28, 26,
+        25, 24, 23, 22, 21, 20, 19, 19, 19, 19, 65, 66, 67, 68, 70, 64, 59, 53,
+        48, 45, 42, 40, 37, 36, 34, 32, 31, 30, 28, 27, 26, 25, 24, 23, 22, 21,
+        20, 19, 18, 18, 18, 18, 59, 61, 63, 65, 66, 61, 56, 51, 46, 43, 40, 37,
+        34, 33, 31, 30, 29, 27, 26, 25, 24, 23, 22, 22, 21, 20, 19, 18, 18, 18,
+        18, 18, 55, 57, 59, 61, 62, 58, 53, 48, 44, 41, 38, 36, 33, 31, 30, 29,
+        27, 26, 25, 24, 23, 22, 21, 21, 20, 19, 18, 18, 17, 17, 17, 17, 51, 53,
+        55, 56, 58, 54, 50, 46, 42, 39, 37, 34, 31, 30, 29, 27, 26, 25, 24, 23,
+        22, 21, 20, 20, 19, 18, 18, 17, 16, 16, 16, 16, 47, 49, 51, 52, 54, 51,
+        47, 44, 40, 37, 35, 32, 30, 29, 27, 26, 24, 24, 23, 22, 21, 20, 19, 19,
+        18, 18, 17, 16, 16, 16, 16, 16, 42, 44, 46, 48, 50, 47, 44, 41, 38, 36,
+        33, 31, 29, 27, 26, 24, 23, 22, 21, 21, 20, 19, 18, 18, 17, 17, 16, 16,
+        15, 15, 15, 15, 40, 41, 43, 45, 47, 44, 42, 39, 36, 34, 32, 30, 27, 26,
+        25, 24, 22, 21, 21, 20, 19, 18, 18, 17, 17, 16, 16, 15, 15, 15, 15, 15,
+        37, 39, 40, 42, 44, 42, 39, 37, 34, 32, 30, 28, 26, 25, 24, 23, 21, 21,
+        20, 19, 18, 18, 17, 17, 16, 16, 15, 15, 14, 14, 14, 14, 34, 36, 37, 39,
+        41, 39, 37, 35, 33, 31, 29, 27, 25, 24, 23, 22, 21, 20, 19, 18, 18, 17,
+        16, 16, 15, 15, 15, 14, 14, 14, 14, 14, 31, 33, 34, 36, 38, 36, 34, 33,
+        31, 29, 28, 26, 24, 23, 22, 21, 20, 19, 18, 18, 17, 16, 16, 15, 15, 15,
+        14, 14, 13, 13, 13, 13, 29, 31, 32, 34, 36, 34, 32, 31, 29, 28, 26, 25,
+        23, 22, 21, 20, 19, 18, 18, 17, 16, 16, 15, 15, 14, 14, 14, 13, 13, 13,
+        13, 13, 27, 29, 30, 32, 33, 32, 31, 29, 28, 27, 25, 24, 22, 21, 20, 19,
+        18, 18, 17, 16, 16, 15, 15, 14, 14, 14, 13, 13, 13, 13, 13, 13, 26, 27,
+        28, 30, 31, 30, 29, 28, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 17, 16,
+        15, 15, 14, 14, 14, 13, 13, 13, 12, 12, 12, 12, 24, 25, 26, 27, 29, 28,
+        27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 17, 16, 15, 15, 14, 14, 14,
+        13, 13, 13, 12, 12, 12, 12, 12, 22, 24, 25, 26, 27, 26, 25, 25, 24, 23,
+        22, 21, 20, 19, 18, 18, 17, 16, 16, 15, 15, 14, 14, 13, 13, 13, 12, 12,
+        12, 12, 12, 12, 21, 22, 23, 25, 26, 25, 24, 23, 23, 22, 21, 20, 19, 18,
+        18, 17, 16, 16, 15, 15, 14, 14, 13, 13, 13, 12, 12, 12, 12, 12, 12, 12,
+        20, 21, 22, 23, 24, 23, 23, 22, 22, 21, 20, 19, 18, 18, 17, 16, 16, 15,
+        15, 14, 14, 13, 13, 13, 12, 12, 12, 12, 11, 11, 11, 11, 19, 20, 21, 22,
+        23, 22, 21, 21, 20, 20, 19, 18, 18, 17, 16, 16, 15, 15, 14, 14, 13, 13,
+        13, 12, 12, 12, 12, 11, 11, 11, 11, 11, 19, 20, 21, 22, 23, 22, 21, 21,
+        20, 20, 19, 18, 18, 17, 16, 16, 15, 15, 14, 14, 13, 13, 13, 12, 12, 12,
+        12, 11, 11, 11, 11, 11, 19, 20, 21, 22, 23, 22, 21, 21, 20, 20, 19, 18,
+        18, 17, 16, 16, 15, 15, 14, 14, 13, 13, 13, 12, 12, 12, 12, 11, 11, 11,
+        11, 11, 19, 20, 21, 22, 23, 22, 21, 21, 20, 20, 19, 18, 18, 17, 16, 16,
+        15, 15, 14, 14, 13, 13, 13, 12, 12, 12, 12, 11, 11, 11, 11, 11 },
+      { /* Intra matrices */
+        /* Size 4 */
+        249, 225, 124, 69, 225, 139, 91, 60, 124, 91, 57, 42, 69, 60, 42, 32,
+        /* Size 8 */
+        206, 285, 261, 191, 134, 96, 71, 55, 285, 245, 257, 214, 161, 118, 88,
+        68, 261, 257, 174, 145, 119, 95, 75, 60, 191, 214, 145, 107, 87, 73, 61,
+        51, 134, 161, 119, 87, 69, 58, 50, 43, 96, 118, 95, 73, 58, 48, 42, 37,
+        71, 88, 75, 61, 50, 42, 36, 32, 55, 68, 60, 51, 43, 37, 32, 29,
+        /* Size 16 */
+        217, 259, 300, 287, 275, 237, 200, 171, 141, 121, 101, 88, 75, 66, 57,
+        57, 259, 269, 279, 275, 272, 243, 213, 184, 155, 134, 113, 98, 84, 74,
+        64, 64, 300, 279, 257, 264, 270, 248, 226, 197, 169, 147, 124, 109, 93,
+        82, 71, 71, 287, 275, 264, 245, 227, 208, 189, 168, 147, 130, 112, 99,
+        86, 77, 67, 67, 275, 272, 270, 227, 183, 168, 152, 139, 125, 113, 100,
+        90, 79, 71, 63, 63, 237, 243, 248, 208, 168, 150, 132, 120, 109, 98, 88,
+        80, 72, 65, 58, 58, 200, 213, 226, 189, 152, 132, 113, 102, 92, 84, 77,
+        70, 64, 59, 54, 54, 171, 184, 197, 168, 139, 120, 102, 92, 82, 75, 69,
+        63, 58, 54, 49, 49, 141, 155, 169, 147, 125, 109, 92, 82, 73, 67, 61,
+        56, 52, 49, 45, 45, 121, 134, 147, 130, 113, 98, 84, 75, 67, 61, 56, 52,
+        48, 45, 42, 42, 101, 113, 124, 112, 100, 88, 77, 69, 61, 56, 50, 47, 44,
+        41, 39, 39, 88, 98, 109, 99, 90, 80, 70, 63, 56, 52, 47, 44, 41, 39, 36,
+        36, 75, 84, 93, 86, 79, 72, 64, 58, 52, 48, 44, 41, 38, 36, 34, 34, 66,
+        74, 82, 77, 71, 65, 59, 54, 49, 45, 41, 39, 36, 34, 32, 32, 57, 64, 71,
+        67, 63, 58, 54, 49, 45, 42, 39, 36, 34, 32, 31, 31, 57, 64, 71, 67, 63,
+        58, 54, 49, 45, 42, 39, 36, 34, 32, 31, 31,
+        /* Size 32 */
+        223, 244, 265, 287, 308, 301, 295, 288, 282, 263, 244, 225, 206, 190,
+        175, 160, 145, 134, 124, 114, 104, 97, 90, 83, 77, 72, 68, 63, 59, 59,
+        59, 59, 244, 257, 271, 284, 297, 293, 289, 285, 281, 264, 246, 229, 212,
+        197, 182, 167, 152, 141, 131, 120, 110, 103, 95, 88, 81, 77, 72, 67, 62,
+        62, 62, 62, 265, 271, 276, 281, 286, 284, 283, 281, 280, 264, 249, 234,
+        219, 204, 189, 174, 159, 148, 137, 127, 116, 108, 101, 93, 86, 81, 76,
+        71, 66, 66, 66, 66, 287, 284, 281, 278, 275, 276, 277, 278, 278, 265,
+        252, 238, 225, 210, 196, 181, 166, 155, 144, 133, 122, 114, 106, 98, 91,
+        85, 80, 75, 69, 69, 69, 69, 308, 297, 286, 275, 264, 267, 271, 274, 277,
+        266, 254, 243, 231, 217, 203, 188, 174, 162, 151, 139, 128, 120, 111,
+        103, 95, 90, 84, 78, 73, 73, 73, 73, 301, 293, 284, 276, 267, 264, 261,
+        258, 255, 244, 234, 223, 213, 200, 187, 175, 162, 152, 142, 132, 121,
+        114, 107, 99, 92, 87, 81, 76, 71, 71, 71, 71, 295, 289, 283, 277, 271,
+        261, 252, 242, 233, 223, 213, 203, 194, 183, 172, 162, 151, 142, 133,
+        124, 115, 108, 102, 95, 88, 83, 79, 74, 69, 69, 69, 69, 288, 285, 281,
+        278, 274, 258, 242, 226, 210, 201, 193, 184, 175, 166, 157, 149, 140,
+        132, 124, 117, 109, 103, 97, 91, 85, 80, 76, 71, 67, 67, 67, 67, 282,
+        281, 280, 278, 277, 255, 233, 210, 188, 180, 172, 164, 156, 149, 142,
+        135, 129, 122, 116, 109, 103, 97, 92, 87, 81, 77, 73, 69, 65, 65, 65,
+        65, 263, 264, 264, 265, 266, 244, 223, 201, 180, 171, 163, 154, 146,
+        139, 133, 126, 120, 114, 108, 103, 97, 92, 87, 82, 77, 74, 70, 66, 62,
+        62, 62, 62, 244, 246, 249, 252, 254, 234, 213, 193, 172, 163, 154, 145,
+        136, 130, 124, 118, 111, 106, 101, 96, 91, 86, 82, 78, 73, 70, 67, 63,
+        60, 60, 60, 60, 225, 229, 234, 238, 243, 223, 203, 184, 164, 154, 145,
+        135, 126, 120, 114, 109, 103, 98, 94, 89, 85, 81, 77, 73, 70, 67, 64,
+        61, 57, 57, 57, 57, 206, 212, 219, 225, 231, 213, 194, 175, 156, 146,
+        136, 126, 116, 110, 105, 100, 94, 90, 87, 83, 79, 76, 72, 69, 66, 63,
+        60, 58, 55, 55, 55, 55, 190, 197, 204, 210, 217, 200, 183, 166, 149,
+        139, 130, 120, 110, 105, 100, 95, 89, 86, 82, 78, 75, 72, 69, 66, 63,
+        60, 58, 55, 53, 53, 53, 53, 175, 182, 189, 196, 203, 187, 172, 157, 142,
+        133, 124, 114, 105, 100, 95, 90, 84, 81, 77, 74, 71, 68, 65, 62, 60, 57,
+        55, 53, 51, 51, 51, 51, 160, 167, 174, 181, 188, 175, 162, 149, 135,
+        126, 118, 109, 100, 95, 90, 84, 79, 76, 73, 70, 66, 64, 61, 59, 57, 55,
+        53, 51, 49, 49, 49, 49, 145, 152, 159, 166, 174, 162, 151, 140, 129,
+        120, 111, 103, 94, 89, 84, 79, 74, 71, 68, 65, 62, 60, 58, 56, 53, 52,
+        50, 48, 46, 46, 46, 46, 134, 141, 148, 155, 162, 152, 142, 132, 122,
+        114, 106, 98, 90, 86, 81, 76, 71, 68, 66, 63, 60, 58, 55, 53, 51, 50,
+        48, 46, 45, 45, 45, 45, 124, 131, 137, 144, 151, 142, 133, 124, 116,
+        108, 101, 94, 87, 82, 77, 73, 68, 66, 63, 60, 57, 55, 53, 51, 49, 48,
+        46, 45, 43, 43, 43, 43, 114, 120, 127, 133, 139, 132, 124, 117, 109,
+        103, 96, 89, 83, 78, 74, 70, 65, 63, 60, 57, 54, 53, 51, 49, 47, 46, 44,
+        43, 41, 41, 41, 41, 104, 110, 116, 122, 128, 121, 115, 109, 103, 97, 91,
+        85, 79, 75, 71, 66, 62, 60, 57, 54, 52, 50, 48, 47, 45, 44, 42, 41, 40,
+        40, 40, 40, 97, 103, 108, 114, 120, 114, 108, 103, 97, 92, 86, 81, 76,
+        72, 68, 64, 60, 58, 55, 53, 50, 48, 47, 45, 43, 42, 41, 40, 38, 38, 38,
+        38, 90, 95, 101, 106, 111, 107, 102, 97, 92, 87, 82, 77, 72, 69, 65, 61,
+        58, 55, 53, 51, 48, 47, 45, 44, 42, 41, 40, 38, 37, 37, 37, 37, 83, 88,
+        93, 98, 103, 99, 95, 91, 87, 82, 78, 73, 69, 66, 62, 59, 56, 53, 51, 49,
+        47, 45, 44, 42, 40, 39, 38, 37, 36, 36, 36, 36, 77, 81, 86, 91, 95, 92,
+        88, 85, 81, 77, 73, 70, 66, 63, 60, 57, 53, 51, 49, 47, 45, 43, 42, 40,
+        39, 38, 37, 36, 35, 35, 35, 35, 72, 77, 81, 85, 90, 87, 83, 80, 77, 74,
+        70, 67, 63, 60, 57, 55, 52, 50, 48, 46, 44, 42, 41, 39, 38, 37, 36, 35,
+        34, 34, 34, 34, 68, 72, 76, 80, 84, 81, 79, 76, 73, 70, 67, 64, 60, 58,
+        55, 53, 50, 48, 46, 44, 42, 41, 40, 38, 37, 36, 35, 34, 33, 33, 33, 33,
+        63, 67, 71, 75, 78, 76, 74, 71, 69, 66, 63, 61, 58, 55, 53, 51, 48, 46,
+        45, 43, 41, 40, 38, 37, 36, 35, 34, 33, 32, 32, 32, 32, 59, 62, 66, 69,
+        73, 71, 69, 67, 65, 62, 60, 57, 55, 53, 51, 49, 46, 45, 43, 41, 40, 38,
+        37, 36, 35, 34, 33, 32, 31, 31, 31, 31, 59, 62, 66, 69, 73, 71, 69, 67,
+        65, 62, 60, 57, 55, 53, 51, 49, 46, 45, 43, 41, 40, 38, 37, 36, 35, 34,
+        33, 32, 31, 31, 31, 31, 59, 62, 66, 69, 73, 71, 69, 67, 65, 62, 60, 57,
+        55, 53, 51, 49, 46, 45, 43, 41, 40, 38, 37, 36, 35, 34, 33, 32, 31, 31,
+        31, 31, 59, 62, 66, 69, 73, 71, 69, 67, 65, 62, 60, 57, 55, 53, 51, 49,
+        46, 45, 43, 41, 40, 38, 37, 36, 35, 34, 33, 32, 31, 31, 31, 31 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 39, 35, 27, 39, 31, 29, 25, 35, 29, 21, 18, 27, 25, 18, 15,
+        /* Size 8 */
+        64, 81, 42, 39, 36, 32, 28, 24, 81, 54, 41, 46, 44, 40, 35, 30, 42, 41,
+        34, 36, 36, 34, 31, 27, 39, 46, 36, 31, 29, 28, 26, 24, 36, 44, 36, 29,
+        26, 24, 22, 20, 32, 40, 34, 28, 24, 21, 19, 18, 28, 35, 31, 26, 22, 19,
+        17, 16, 24, 30, 27, 24, 20, 18, 16, 15,
+        /* Size 16 */
+        64, 72, 81, 61, 42, 41, 39, 38, 36, 34, 32, 30, 28, 26, 24, 24, 72, 70,
+        67, 54, 42, 42, 43, 42, 40, 38, 36, 34, 31, 29, 27, 27, 81, 67, 54, 48,
+        41, 44, 46, 45, 44, 42, 40, 37, 35, 32, 30, 30, 61, 54, 48, 43, 38, 39,
+        41, 40, 40, 39, 37, 35, 33, 31, 29, 29, 42, 42, 41, 38, 34, 35, 36, 36,
+        36, 35, 34, 32, 31, 29, 27, 27, 41, 42, 44, 39, 35, 34, 33, 33, 33, 32,
+        31, 30, 28, 27, 25, 25, 39, 43, 46, 41, 36, 33, 31, 30, 29, 29, 28, 27,
+        26, 25, 24, 24, 38, 42, 45, 40, 36, 33, 30, 29, 27, 27, 26, 25, 24, 23,
+        22, 22, 36, 40, 44, 40, 36, 33, 29, 27, 26, 25, 24, 23, 22, 21, 20, 20,
+        34, 38, 42, 39, 35, 32, 29, 27, 25, 23, 22, 21, 21, 20, 19, 19, 32, 36,
+        40, 37, 34, 31, 28, 26, 24, 22, 21, 20, 19, 19, 18, 18, 30, 34, 37, 35,
+        32, 30, 27, 25, 23, 21, 20, 19, 18, 18, 17, 17, 28, 31, 35, 33, 31, 28,
+        26, 24, 22, 21, 19, 18, 17, 17, 16, 16, 26, 29, 32, 31, 29, 27, 25, 23,
+        21, 20, 19, 18, 17, 16, 15, 15, 24, 27, 30, 29, 27, 25, 24, 22, 20, 19,
+        18, 17, 16, 15, 15, 15, 24, 27, 30, 29, 27, 25, 24, 22, 20, 19, 18, 17,
+        16, 15, 15, 15,
+        /* Size 32 */
+        64, 68, 72, 76, 81, 71, 61, 52, 42, 41, 41, 40, 39, 39, 38, 37, 36, 35,
+        34, 33, 32, 31, 30, 29, 28, 27, 26, 25, 24, 24, 24, 24, 68, 70, 71, 73,
+        74, 66, 58, 50, 42, 42, 41, 41, 41, 40, 40, 39, 38, 37, 36, 35, 34, 33,
+        32, 31, 29, 28, 27, 26, 25, 25, 25, 25, 72, 71, 70, 69, 67, 61, 54, 48,
+        42, 42, 42, 42, 43, 42, 42, 41, 40, 39, 38, 37, 36, 35, 34, 32, 31, 30,
+        29, 28, 27, 27, 27, 27, 76, 73, 69, 65, 61, 56, 51, 46, 41, 42, 43, 44,
+        44, 44, 43, 43, 42, 41, 40, 39, 38, 37, 36, 34, 33, 32, 31, 30, 29, 29,
+        29, 29, 81, 74, 67, 61, 54, 51, 48, 44, 41, 42, 44, 45, 46, 46, 45, 45,
+        44, 43, 42, 41, 40, 39, 37, 36, 35, 34, 32, 31, 30, 30, 30, 30, 71, 66,
+        61, 56, 51, 48, 45, 42, 39, 40, 41, 42, 43, 43, 43, 43, 42, 41, 40, 39,
+        38, 37, 36, 35, 34, 33, 32, 30, 29, 29, 29, 29, 61, 58, 54, 51, 48, 45,
+        43, 40, 38, 38, 39, 40, 41, 41, 40, 40, 40, 39, 39, 38, 37, 36, 35, 34,
+        33, 32, 31, 30, 29, 29, 29, 29, 52, 50, 48, 46, 44, 42, 40, 38, 36, 37,
+        37, 38, 38, 38, 38, 38, 38, 37, 37, 36, 35, 34, 33, 33, 32, 31, 30, 29,
+        28, 28, 28, 28, 42, 42, 42, 41, 41, 39, 38, 36, 34, 35, 35, 35, 36, 36,
+        36, 36, 36, 35, 35, 34, 34, 33, 32, 31, 31, 30, 29, 28, 27, 27, 27, 27,
+        41, 42, 42, 42, 42, 40, 38, 37, 35, 35, 35, 35, 35, 34, 34, 34, 34, 34,
+        33, 33, 32, 32, 31, 30, 29, 29, 28, 27, 26, 26, 26, 26, 41, 41, 42, 43,
+        44, 41, 39, 37, 35, 35, 34, 34, 33, 33, 33, 33, 33, 32, 32, 31, 31, 30,
+        30, 29, 28, 28, 27, 26, 25, 25, 25, 25, 40, 41, 42, 44, 45, 42, 40, 38,
+        35, 35, 34, 33, 32, 32, 32, 31, 31, 30, 30, 30, 29, 29, 28, 28, 27, 26,
+        26, 25, 25, 25, 25, 25, 39, 41, 43, 44, 46, 43, 41, 38, 36, 35, 33, 32,
+        31, 31, 30, 30, 29, 29, 29, 28, 28, 27, 27, 26, 26, 25, 25, 24, 24, 24,
+        24, 24, 39, 40, 42, 44, 46, 43, 41, 38, 36, 34, 33, 32, 31, 30, 29, 29,
+        28, 28, 28, 27, 27, 26, 26, 25, 25, 24, 24, 23, 23, 23, 23, 23, 38, 40,
+        42, 43, 45, 43, 40, 38, 36, 34, 33, 32, 30, 29, 29, 28, 27, 27, 27, 26,
+        26, 25, 25, 24, 24, 23, 23, 23, 22, 22, 22, 22, 37, 39, 41, 43, 45, 43,
+        40, 38, 36, 34, 33, 31, 30, 29, 28, 27, 27, 26, 26, 25, 25, 24, 24, 23,
+        23, 23, 22, 22, 21, 21, 21, 21, 36, 38, 40, 42, 44, 42, 40, 38, 36, 34,
+        33, 31, 29, 28, 27, 27, 26, 25, 25, 24, 24, 23, 23, 22, 22, 22, 21, 21,
+        20, 20, 20, 20, 35, 37, 39, 41, 43, 41, 39, 37, 35, 34, 32, 30, 29, 28,
+        27, 26, 25, 25, 24, 23, 23, 23, 22, 22, 21, 21, 21, 20, 20, 20, 20, 20,
+        34, 36, 38, 40, 42, 40, 39, 37, 35, 33, 32, 30, 29, 28, 27, 26, 25, 24,
+        23, 23, 22, 22, 21, 21, 21, 20, 20, 20, 19, 19, 19, 19, 33, 35, 37, 39,
+        41, 39, 38, 36, 34, 33, 31, 30, 28, 27, 26, 25, 24, 23, 23, 22, 22, 21,
+        21, 20, 20, 20, 19, 19, 19, 19, 19, 19, 32, 34, 36, 38, 40, 38, 37, 35,
+        34, 32, 31, 29, 28, 27, 26, 25, 24, 23, 22, 22, 21, 21, 20, 20, 19, 19,
+        19, 18, 18, 18, 18, 18, 31, 33, 35, 37, 39, 37, 36, 34, 33, 32, 30, 29,
+        27, 26, 25, 24, 23, 23, 22, 21, 21, 20, 20, 19, 19, 18, 18, 18, 17, 17,
+        17, 17, 30, 32, 34, 36, 37, 36, 35, 33, 32, 31, 30, 28, 27, 26, 25, 24,
+        23, 22, 21, 21, 20, 20, 19, 19, 18, 18, 18, 17, 17, 17, 17, 17, 29, 31,
+        32, 34, 36, 35, 34, 33, 31, 30, 29, 28, 26, 25, 24, 23, 22, 22, 21, 20,
+        20, 19, 19, 18, 18, 17, 17, 17, 16, 16, 16, 16, 28, 29, 31, 33, 35, 34,
+        33, 32, 31, 29, 28, 27, 26, 25, 24, 23, 22, 21, 21, 20, 19, 19, 18, 18,
+        17, 17, 17, 16, 16, 16, 16, 16, 27, 28, 30, 32, 34, 33, 32, 31, 30, 29,
+        28, 26, 25, 24, 23, 23, 22, 21, 20, 20, 19, 18, 18, 17, 17, 17, 16, 16,
+        16, 16, 16, 16, 26, 27, 29, 31, 32, 32, 31, 30, 29, 28, 27, 26, 25, 24,
+        23, 22, 21, 21, 20, 19, 19, 18, 18, 17, 17, 16, 16, 16, 15, 15, 15, 15,
+        25, 26, 28, 30, 31, 30, 30, 29, 28, 27, 26, 25, 24, 23, 23, 22, 21, 20,
+        20, 19, 18, 18, 17, 17, 16, 16, 16, 15, 15, 15, 15, 15, 24, 25, 27, 29,
+        30, 29, 29, 28, 27, 26, 25, 25, 24, 23, 22, 21, 20, 20, 19, 19, 18, 17,
+        17, 16, 16, 16, 15, 15, 15, 15, 15, 15, 24, 25, 27, 29, 30, 29, 29, 28,
+        27, 26, 25, 25, 24, 23, 22, 21, 20, 20, 19, 19, 18, 17, 17, 16, 16, 16,
+        15, 15, 15, 15, 15, 15, 24, 25, 27, 29, 30, 29, 29, 28, 27, 26, 25, 25,
+        24, 23, 22, 21, 20, 20, 19, 19, 18, 17, 17, 16, 16, 16, 15, 15, 15, 15,
+        15, 15, 24, 25, 27, 29, 30, 29, 29, 28, 27, 26, 25, 25, 24, 23, 22, 21,
+        20, 20, 19, 19, 18, 17, 17, 16, 16, 16, 15, 15, 15, 15, 15, 15 },
+      { /* Intra matrices */
+        /* Size 4 */
+        175, 103, 93, 70, 103, 83, 76, 64, 93, 76, 55, 46, 70, 64, 46, 36,
+        /* Size 8 */
+        162, 205, 104, 97, 90, 78, 67, 57, 205, 136, 102, 115, 111, 99, 86, 73,
+        104, 102, 84, 88, 88, 83, 75, 66, 97, 115, 88, 75, 71, 67, 62, 56, 90,
+        111, 88, 71, 62, 56, 52, 48, 78, 99, 83, 67, 56, 49, 45, 41, 67, 86, 75,
+        62, 52, 45, 40, 36, 57, 73, 66, 56, 48, 41, 36, 33,
+        /* Size 16 */
+        168, 190, 213, 160, 108, 104, 101, 97, 93, 87, 81, 75, 69, 64, 59, 59,
+        190, 184, 177, 142, 107, 108, 110, 107, 104, 98, 92, 85, 79, 73, 67, 67,
+        213, 177, 141, 123, 106, 112, 119, 117, 115, 109, 103, 96, 89, 82, 76,
+        76, 160, 142, 123, 110, 96, 101, 105, 104, 103, 99, 94, 89, 83, 77, 72,
+        72, 108, 107, 106, 96, 87, 89, 91, 91, 91, 89, 86, 82, 77, 73, 68, 68,
+        104, 108, 112, 101, 89, 87, 85, 84, 82, 80, 78, 74, 71, 67, 63, 63, 101,
+        110, 119, 105, 91, 85, 78, 76, 74, 72, 70, 67, 64, 61, 58, 58, 97, 107,
+        117, 104, 91, 84, 76, 72, 69, 66, 64, 62, 59, 57, 54, 54, 93, 104, 115,
+        103, 91, 82, 74, 69, 64, 61, 58, 56, 54, 52, 50, 50, 87, 98, 109, 99,
+        89, 80, 72, 66, 61, 58, 55, 52, 50, 48, 46, 46, 81, 92, 103, 94, 86, 78,
+        70, 64, 58, 55, 51, 49, 46, 45, 43, 43, 75, 85, 96, 89, 82, 74, 67, 62,
+        56, 52, 49, 46, 44, 42, 40, 40, 69, 79, 89, 83, 77, 71, 64, 59, 54, 50,
+        46, 44, 41, 40, 38, 38, 64, 73, 82, 77, 73, 67, 61, 57, 52, 48, 45, 42,
+        40, 38, 36, 36, 59, 67, 76, 72, 68, 63, 58, 54, 50, 46, 43, 40, 38, 36,
+        34, 34, 59, 67, 76, 72, 68, 63, 58, 54, 50, 46, 43, 40, 38, 36, 34, 34,
+        /* Size 32 */
+        171, 182, 194, 205, 217, 190, 163, 137, 110, 108, 106, 104, 103, 101,
+        99, 97, 95, 92, 89, 86, 83, 80, 77, 74, 71, 68, 65, 63, 60, 60, 60, 60,
+        182, 186, 190, 194, 198, 176, 154, 132, 109, 109, 108, 108, 107, 105,
+        104, 102, 100, 97, 94, 91, 88, 85, 82, 79, 76, 73, 70, 67, 64, 64, 64,
+        64, 194, 190, 187, 184, 180, 162, 144, 127, 109, 110, 110, 111, 112,
+        110, 109, 107, 106, 103, 100, 97, 94, 90, 87, 84, 80, 78, 75, 72, 69,
+        69, 69, 69, 205, 194, 184, 173, 162, 148, 135, 122, 108, 110, 112, 114,
+        116, 115, 114, 113, 111, 108, 105, 102, 99, 96, 92, 89, 85, 82, 79, 76,
+        73, 73, 73, 73, 217, 198, 180, 162, 144, 135, 126, 117, 108, 111, 114,
+        118, 121, 120, 119, 118, 117, 114, 111, 108, 104, 101, 97, 94, 90, 87,
+        84, 80, 77, 77, 77, 77, 190, 176, 162, 148, 135, 127, 119, 111, 103,
+        106, 108, 111, 114, 113, 112, 112, 111, 108, 106, 103, 100, 97, 94, 91,
+        87, 84, 81, 78, 75, 75, 75, 75, 163, 154, 144, 135, 126, 119, 112, 105,
+        98, 100, 103, 105, 107, 106, 106, 105, 105, 103, 101, 98, 96, 93, 90,
+        87, 85, 82, 79, 76, 73, 73, 73, 73, 137, 132, 127, 122, 117, 111, 105,
+        99, 94, 95, 97, 98, 100, 100, 99, 99, 99, 97, 95, 94, 92, 89, 87, 84,
+        82, 79, 76, 74, 71, 71, 71, 71, 110, 109, 109, 108, 108, 103, 98, 94,
+        89, 90, 91, 92, 93, 93, 93, 93, 93, 92, 90, 89, 87, 85, 83, 81, 79, 76,
+        74, 72, 69, 69, 69, 69, 108, 109, 110, 110, 111, 106, 100, 95, 90, 90,
+        90, 90, 89, 89, 89, 89, 89, 87, 86, 85, 83, 81, 79, 77, 75, 73, 71, 69,
+        67, 67, 67, 67, 106, 108, 110, 112, 114, 108, 103, 97, 91, 90, 89, 87,
+        86, 86, 85, 85, 84, 83, 82, 80, 79, 77, 76, 74, 72, 70, 68, 66, 64, 64,
+        64, 64, 104, 108, 111, 114, 118, 111, 105, 98, 92, 90, 87, 85, 83, 82,
+        81, 80, 79, 78, 77, 76, 75, 74, 72, 70, 69, 67, 65, 64, 62, 62, 62, 62,
+        103, 107, 112, 116, 121, 114, 107, 100, 93, 89, 86, 83, 80, 78, 77, 76,
+        75, 74, 73, 72, 71, 70, 68, 67, 66, 64, 63, 61, 60, 60, 60, 60, 101,
+        105, 110, 115, 120, 113, 106, 100, 93, 89, 86, 82, 78, 77, 75, 74, 72,
+        71, 70, 69, 68, 67, 65, 64, 63, 62, 60, 59, 57, 57, 57, 57, 99, 104,
+        109, 114, 119, 112, 106, 99, 93, 89, 85, 81, 77, 75, 74, 72, 70, 69, 68,
+        66, 65, 64, 63, 61, 60, 59, 58, 56, 55, 55, 55, 55, 97, 102, 107, 113,
+        118, 112, 105, 99, 93, 89, 85, 80, 76, 74, 72, 70, 67, 66, 65, 64, 62,
+        61, 60, 59, 58, 56, 55, 54, 53, 53, 53, 53, 95, 100, 106, 111, 117, 111,
+        105, 99, 93, 89, 84, 79, 75, 72, 70, 67, 65, 64, 62, 61, 59, 58, 57, 56,
+        55, 54, 53, 52, 51, 51, 51, 51, 92, 97, 103, 108, 114, 108, 103, 97, 92,
+        87, 83, 78, 74, 71, 69, 66, 64, 62, 61, 59, 57, 56, 55, 54, 53, 52, 51,
+        50, 49, 49, 49, 49, 89, 94, 100, 105, 111, 106, 101, 95, 90, 86, 82, 77,
+        73, 70, 68, 65, 62, 61, 59, 57, 56, 55, 53, 52, 51, 50, 49, 48, 47, 47,
+        47, 47, 86, 91, 97, 102, 108, 103, 98, 94, 89, 85, 80, 76, 72, 69, 66,
+        64, 61, 59, 57, 56, 54, 53, 52, 50, 49, 48, 47, 46, 45, 45, 45, 45, 83,
+        88, 94, 99, 104, 100, 96, 92, 87, 83, 79, 75, 71, 68, 65, 62, 59, 57,
+        56, 54, 52, 51, 50, 48, 47, 46, 45, 44, 44, 44, 44, 44, 80, 85, 90, 96,
+        101, 97, 93, 89, 85, 81, 77, 74, 70, 67, 64, 61, 58, 56, 55, 53, 51, 50,
+        48, 47, 46, 45, 44, 43, 42, 42, 42, 42, 77, 82, 87, 92, 97, 94, 90, 87,
+        83, 79, 76, 72, 68, 65, 63, 60, 57, 55, 53, 52, 50, 48, 47, 46, 45, 44,
+        43, 42, 41, 41, 41, 41, 74, 79, 84, 89, 94, 91, 87, 84, 81, 77, 74, 70,
+        67, 64, 61, 59, 56, 54, 52, 50, 48, 47, 46, 45, 43, 42, 42, 41, 40, 40,
+        40, 40, 71, 76, 80, 85, 90, 87, 85, 82, 79, 75, 72, 69, 66, 63, 60, 58,
+        55, 53, 51, 49, 47, 46, 45, 43, 42, 41, 40, 39, 38, 38, 38, 38, 68, 73,
+        78, 82, 87, 84, 82, 79, 76, 73, 70, 67, 64, 62, 59, 56, 54, 52, 50, 48,
+        46, 45, 44, 42, 41, 40, 39, 38, 37, 37, 37, 37, 65, 70, 75, 79, 84, 81,
+        79, 76, 74, 71, 68, 65, 63, 60, 58, 55, 53, 51, 49, 47, 45, 44, 43, 42,
+        40, 39, 38, 37, 37, 37, 37, 37, 63, 67, 72, 76, 80, 78, 76, 74, 72, 69,
+        66, 64, 61, 59, 56, 54, 52, 50, 48, 46, 44, 43, 42, 41, 39, 38, 37, 37,
+        36, 36, 36, 36, 60, 64, 69, 73, 77, 75, 73, 71, 69, 67, 64, 62, 60, 57,
+        55, 53, 51, 49, 47, 45, 44, 42, 41, 40, 38, 37, 37, 36, 35, 35, 35, 35,
+        60, 64, 69, 73, 77, 75, 73, 71, 69, 67, 64, 62, 60, 57, 55, 53, 51, 49,
+        47, 45, 44, 42, 41, 40, 38, 37, 37, 36, 35, 35, 35, 35, 60, 64, 69, 73,
+        77, 75, 73, 71, 69, 67, 64, 62, 60, 57, 55, 53, 51, 49, 47, 45, 44, 42,
+        41, 40, 38, 37, 37, 36, 35, 35, 35, 35, 60, 64, 69, 73, 77, 75, 73, 71,
+        69, 67, 64, 62, 60, 57, 55, 53, 51, 49, 47, 45, 44, 42, 41, 40, 38, 37,
+        37, 36, 35, 35, 35, 35 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 58, 34, 21, 58, 38, 26, 19, 34, 26, 18, 14, 21, 19, 14, 12,
+        /* Size 8 */
+        64, 86, 80, 59, 43, 33, 25, 21, 86, 75, 78, 66, 51, 39, 30, 24, 80, 78,
+        55, 46, 39, 32, 27, 22, 59, 66, 46, 36, 30, 26, 23, 20, 43, 51, 39, 30,
+        25, 22, 19, 17, 33, 39, 32, 26, 22, 19, 17, 16, 25, 30, 27, 23, 19, 17,
+        16, 14, 21, 24, 22, 20, 17, 16, 14, 14,
+        /* Size 16 */
+        64, 75, 86, 83, 80, 70, 59, 51, 43, 38, 33, 29, 25, 23, 21, 21, 75, 78,
+        81, 80, 79, 71, 63, 55, 47, 41, 36, 32, 28, 25, 23, 23, 86, 81, 75, 77,
+        78, 72, 66, 59, 51, 45, 39, 35, 30, 27, 24, 24, 83, 80, 77, 72, 67, 61,
+        56, 51, 45, 40, 36, 32, 29, 26, 23, 23, 80, 79, 78, 67, 55, 51, 46, 43,
+        39, 36, 32, 29, 27, 24, 22, 22, 70, 71, 72, 61, 51, 46, 41, 38, 35, 32,
+        29, 27, 25, 23, 21, 21, 59, 63, 66, 56, 46, 41, 36, 33, 30, 28, 26, 24,
+        23, 21, 20, 20, 51, 55, 59, 51, 43, 38, 33, 30, 28, 26, 24, 22, 21, 20,
+        19, 19, 43, 47, 51, 45, 39, 35, 30, 28, 25, 23, 22, 21, 19, 18, 17, 17,
+        38, 41, 45, 40, 36, 32, 28, 26, 23, 22, 20, 19, 18, 17, 17, 17, 33, 36,
+        39, 36, 32, 29, 26, 24, 22, 20, 19, 18, 17, 16, 16, 16, 29, 32, 35, 32,
+        29, 27, 24, 22, 21, 19, 18, 17, 16, 16, 15, 15, 25, 28, 30, 29, 27, 25,
+        23, 21, 19, 18, 17, 16, 16, 15, 14, 14, 23, 25, 27, 26, 24, 23, 21, 20,
+        18, 17, 16, 16, 15, 14, 14, 14, 21, 23, 24, 23, 22, 21, 20, 19, 17, 17,
+        16, 15, 14, 14, 14, 14, 21, 23, 24, 23, 22, 21, 20, 19, 17, 17, 16, 15,
+        14, 14, 14, 14,
+        /* Size 32 */
+        64, 70, 75, 81, 86, 85, 83, 81, 80, 75, 70, 64, 59, 55, 51, 47, 43, 41,
+        38, 35, 33, 31, 29, 27, 25, 24, 23, 22, 21, 21, 21, 21, 70, 73, 77, 80,
+        84, 82, 81, 80, 79, 75, 70, 66, 61, 57, 53, 49, 45, 43, 40, 37, 34, 32,
+        30, 29, 27, 25, 24, 23, 22, 22, 22, 22, 75, 77, 78, 79, 81, 80, 80, 79,
+        79, 75, 71, 67, 63, 59, 55, 51, 47, 44, 41, 39, 36, 34, 32, 30, 28, 27,
+        25, 24, 23, 23, 23, 23, 81, 80, 79, 78, 78, 78, 78, 78, 79, 75, 72, 68,
+        65, 61, 57, 53, 49, 46, 43, 40, 37, 35, 33, 31, 29, 28, 26, 25, 24, 24,
+        24, 24, 86, 84, 81, 78, 75, 76, 77, 77, 78, 75, 72, 69, 66, 62, 59, 55,
+        51, 48, 45, 42, 39, 37, 35, 33, 30, 29, 27, 26, 24, 24, 24, 24, 85, 82,
+        80, 78, 76, 75, 74, 73, 72, 70, 67, 64, 61, 58, 55, 51, 48, 45, 43, 40,
+        37, 35, 33, 31, 29, 28, 27, 25, 24, 24, 24, 24, 83, 81, 80, 78, 77, 74,
+        72, 69, 67, 64, 61, 59, 56, 54, 51, 48, 45, 43, 40, 38, 36, 34, 32, 30,
+        29, 27, 26, 25, 23, 23, 23, 23, 81, 80, 79, 78, 77, 73, 69, 65, 61, 58,
+        56, 54, 51, 49, 47, 44, 42, 40, 38, 36, 34, 32, 31, 29, 28, 26, 25, 24,
+        23, 23, 23, 23, 80, 79, 79, 79, 78, 72, 67, 61, 55, 53, 51, 48, 46, 45,
+        43, 41, 39, 37, 36, 34, 32, 31, 29, 28, 27, 26, 24, 23, 22, 22, 22, 22,
+        75, 75, 75, 75, 75, 70, 64, 58, 53, 50, 48, 46, 44, 42, 40, 39, 37, 35,
+        34, 32, 31, 29, 28, 27, 26, 25, 24, 23, 22, 22, 22, 22, 70, 70, 71, 72,
+        72, 67, 61, 56, 51, 48, 46, 43, 41, 39, 38, 36, 35, 33, 32, 31, 29, 28,
+        27, 26, 25, 24, 23, 22, 21, 21, 21, 21, 64, 66, 67, 68, 69, 64, 59, 54,
+        48, 46, 43, 41, 38, 37, 35, 34, 32, 31, 30, 29, 28, 27, 26, 25, 24, 23,
+        22, 21, 20, 20, 20, 20, 59, 61, 63, 65, 66, 61, 56, 51, 46, 44, 41, 38,
+        36, 34, 33, 32, 30, 29, 28, 27, 26, 25, 24, 23, 23, 22, 21, 20, 20, 20,
+        20, 20, 55, 57, 59, 61, 62, 58, 54, 49, 45, 42, 39, 37, 34, 33, 32, 30,
+        29, 28, 27, 26, 25, 24, 23, 23, 22, 21, 20, 20, 19, 19, 19, 19, 51, 53,
+        55, 57, 59, 55, 51, 47, 43, 40, 38, 35, 33, 32, 30, 29, 28, 27, 26, 25,
+        24, 23, 22, 22, 21, 20, 20, 19, 19, 19, 19, 19, 47, 49, 51, 53, 55, 51,
+        48, 44, 41, 39, 36, 34, 32, 30, 29, 28, 26, 25, 24, 24, 23, 22, 21, 21,
+        20, 20, 19, 19, 18, 18, 18, 18, 43, 45, 47, 49, 51, 48, 45, 42, 39, 37,
+        35, 32, 30, 29, 28, 26, 25, 24, 23, 22, 22, 21, 21, 20, 19, 19, 18, 18,
+        17, 17, 17, 17, 41, 43, 44, 46, 48, 45, 43, 40, 37, 35, 33, 31, 29, 28,
+        27, 25, 24, 23, 23, 22, 21, 20, 20, 19, 19, 18, 18, 17, 17, 17, 17, 17,
+        38, 40, 41, 43, 45, 43, 40, 38, 36, 34, 32, 30, 28, 27, 26, 24, 23, 23,
+        22, 21, 20, 20, 19, 19, 18, 18, 17, 17, 17, 17, 17, 17, 35, 37, 39, 40,
+        42, 40, 38, 36, 34, 32, 31, 29, 27, 26, 25, 24, 22, 22, 21, 20, 20, 19,
+        19, 18, 18, 17, 17, 17, 16, 16, 16, 16, 33, 34, 36, 37, 39, 37, 36, 34,
+        32, 31, 29, 28, 26, 25, 24, 23, 22, 21, 20, 20, 19, 18, 18, 18, 17, 17,
+        16, 16, 16, 16, 16, 16, 31, 32, 34, 35, 37, 35, 34, 32, 31, 29, 28, 27,
+        25, 24, 23, 22, 21, 20, 20, 19, 18, 18, 18, 17, 17, 16, 16, 16, 15, 15,
+        15, 15, 29, 30, 32, 33, 35, 33, 32, 31, 29, 28, 27, 26, 24, 23, 22, 21,
+        21, 20, 19, 19, 18, 18, 17, 17, 16, 16, 16, 15, 15, 15, 15, 15, 27, 29,
+        30, 31, 33, 31, 30, 29, 28, 27, 26, 25, 23, 23, 22, 21, 20, 19, 19, 18,
+        18, 17, 17, 16, 16, 16, 15, 15, 15, 15, 15, 15, 25, 27, 28, 29, 30, 29,
+        29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 19, 18, 18, 17, 17, 16, 16,
+        16, 15, 15, 15, 14, 14, 14, 14, 24, 25, 27, 28, 29, 28, 27, 26, 26, 25,
+        24, 23, 22, 21, 20, 20, 19, 18, 18, 17, 17, 16, 16, 16, 15, 15, 15, 14,
+        14, 14, 14, 14, 23, 24, 25, 26, 27, 27, 26, 25, 24, 24, 23, 22, 21, 20,
+        20, 19, 18, 18, 17, 17, 16, 16, 16, 15, 15, 15, 14, 14, 14, 14, 14, 14,
+        22, 23, 24, 25, 26, 25, 25, 24, 23, 23, 22, 21, 20, 20, 19, 19, 18, 17,
+        17, 17, 16, 16, 15, 15, 15, 14, 14, 14, 14, 14, 14, 14, 21, 22, 23, 24,
+        24, 24, 23, 23, 22, 22, 21, 20, 20, 19, 19, 18, 17, 17, 17, 16, 16, 15,
+        15, 15, 14, 14, 14, 14, 14, 14, 14, 14, 21, 22, 23, 24, 24, 24, 23, 23,
+        22, 22, 21, 20, 20, 19, 19, 18, 17, 17, 17, 16, 16, 15, 15, 15, 14, 14,
+        14, 14, 14, 14, 14, 14, 21, 22, 23, 24, 24, 24, 23, 23, 22, 22, 21, 20,
+        20, 19, 19, 18, 17, 17, 17, 16, 16, 15, 15, 15, 14, 14, 14, 14, 14, 14,
+        14, 14, 21, 22, 23, 24, 24, 24, 23, 23, 22, 22, 21, 20, 20, 19, 19, 18,
+        17, 17, 17, 16, 16, 15, 15, 15, 14, 14, 14, 14, 14, 14, 14, 14 },
+      { /* Intra matrices */
+        /* Size 4 */
+        217, 196, 112, 65, 196, 124, 84, 58, 112, 84, 55, 43, 65, 58, 43, 34,
+        /* Size 8 */
+        180, 246, 226, 167, 120, 88, 67, 53, 246, 212, 222, 187, 142, 107, 81,
+        64, 226, 222, 153, 128, 107, 87, 71, 58, 167, 187, 128, 97, 81, 69, 59,
+        50, 120, 142, 107, 81, 65, 56, 49, 44, 88, 107, 87, 69, 56, 48, 43, 39,
+        67, 81, 71, 59, 49, 43, 38, 35, 53, 64, 58, 50, 44, 39, 35, 32,
+        /* Size 16 */
+        188, 223, 257, 246, 236, 205, 174, 150, 125, 108, 92, 81, 70, 63, 56,
+        56, 223, 231, 239, 237, 234, 209, 185, 161, 137, 119, 102, 90, 78, 70,
+        62, 62, 257, 239, 221, 227, 232, 214, 195, 172, 148, 130, 111, 98, 85,
+        76, 67, 67, 246, 237, 227, 211, 196, 180, 165, 148, 130, 116, 101, 90,
+        80, 72, 64, 64, 236, 234, 232, 196, 160, 147, 134, 123, 112, 102, 91,
+        83, 74, 67, 61, 61, 205, 209, 214, 180, 147, 133, 118, 108, 98, 90, 82,
+        75, 68, 62, 57, 57, 174, 185, 195, 165, 134, 118, 102, 93, 84, 78, 72,
+        67, 61, 57, 53, 53, 150, 161, 172, 148, 123, 108, 93, 85, 76, 71, 65,
+        61, 56, 53, 49, 49, 125, 137, 148, 130, 112, 98, 84, 76, 68, 64, 59, 55,
+        51, 49, 46, 46, 108, 119, 130, 116, 102, 90, 78, 71, 64, 59, 54, 51, 48,
+        46, 43, 43, 92, 102, 111, 101, 91, 82, 72, 65, 59, 54, 50, 47, 45, 42,
+        40, 40, 81, 90, 98, 90, 83, 75, 67, 61, 55, 51, 47, 45, 42, 40, 38, 38,
+        70, 78, 85, 80, 74, 68, 61, 56, 51, 48, 45, 42, 40, 38, 36, 36, 63, 70,
+        76, 72, 67, 62, 57, 53, 49, 46, 42, 40, 38, 37, 35, 35, 56, 62, 67, 64,
+        61, 57, 53, 49, 46, 43, 40, 38, 36, 35, 34, 34, 56, 62, 67, 64, 61, 57,
+        53, 49, 46, 43, 40, 38, 36, 35, 34, 34,
+        /* Size 32 */
+        193, 210, 228, 245, 263, 258, 252, 247, 241, 226, 210, 194, 178, 166,
+        153, 141, 128, 120, 111, 103, 94, 89, 83, 77, 72, 68, 65, 61, 57, 57,
+        57, 57, 210, 221, 232, 243, 254, 250, 247, 244, 240, 226, 212, 198, 184,
+        171, 159, 146, 134, 125, 117, 108, 99, 93, 87, 82, 76, 72, 68, 64, 60,
+        60, 60, 60, 228, 232, 236, 240, 245, 243, 242, 241, 239, 227, 214, 202,
+        189, 177, 165, 152, 140, 131, 122, 113, 104, 98, 92, 86, 80, 75, 71, 67,
+        63, 63, 63, 63, 245, 243, 240, 238, 236, 236, 237, 238, 238, 227, 216,
+        205, 194, 182, 170, 158, 146, 137, 127, 118, 109, 103, 96, 90, 83, 79,
+        75, 70, 66, 66, 66, 66, 263, 254, 245, 236, 227, 229, 232, 235, 238,
+        228, 219, 209, 200, 188, 176, 164, 152, 142, 133, 123, 114, 107, 101,
+        94, 87, 83, 78, 73, 69, 69, 69, 69, 258, 250, 243, 236, 229, 227, 224,
+        222, 219, 210, 202, 193, 184, 174, 163, 153, 143, 134, 126, 117, 109,
+        103, 97, 90, 84, 80, 76, 71, 67, 67, 67, 67, 252, 247, 242, 237, 232,
+        224, 216, 208, 201, 193, 185, 177, 169, 160, 151, 142, 133, 126, 118,
+        111, 104, 98, 93, 87, 81, 77, 73, 69, 65, 65, 65, 65, 247, 244, 241,
+        238, 235, 222, 208, 195, 182, 175, 168, 160, 153, 146, 139, 131, 124,
+        118, 111, 105, 98, 93, 88, 83, 78, 75, 71, 67, 64, 64, 64, 64, 241, 240,
+        239, 238, 238, 219, 201, 182, 164, 157, 151, 144, 137, 132, 126, 120,
+        115, 109, 104, 99, 93, 89, 84, 80, 76, 72, 69, 65, 62, 62, 62, 62, 226,
+        226, 227, 227, 228, 210, 193, 175, 157, 150, 143, 136, 129, 124, 118,
+        113, 108, 103, 98, 93, 88, 84, 80, 76, 72, 69, 66, 63, 60, 60, 60, 60,
+        210, 212, 214, 216, 219, 202, 185, 168, 151, 143, 136, 128, 121, 116,
+        111, 106, 101, 96, 92, 88, 83, 80, 76, 73, 69, 66, 64, 61, 58, 58, 58,
+        58, 194, 198, 202, 205, 209, 193, 177, 160, 144, 136, 128, 120, 112,
+        108, 103, 98, 94, 90, 86, 82, 79, 75, 72, 69, 66, 64, 61, 59, 56, 56,
+        56, 56, 178, 184, 189, 194, 200, 184, 169, 153, 137, 129, 121, 112, 104,
+        100, 95, 91, 86, 83, 80, 77, 74, 71, 68, 66, 63, 61, 58, 56, 54, 54, 54,
+        54, 166, 171, 177, 182, 188, 174, 160, 146, 132, 124, 116, 108, 100, 95,
+        91, 87, 82, 79, 76, 73, 70, 68, 65, 63, 60, 58, 56, 54, 52, 52, 52, 52,
+        153, 159, 165, 170, 176, 163, 151, 139, 126, 118, 111, 103, 95, 91, 87,
+        82, 78, 75, 73, 70, 67, 65, 62, 60, 58, 56, 54, 52, 50, 50, 50, 50, 141,
+        146, 152, 158, 164, 153, 142, 131, 120, 113, 106, 98, 91, 87, 82, 78,
+        74, 71, 69, 66, 63, 61, 59, 57, 55, 54, 52, 50, 49, 49, 49, 49, 128,
+        134, 140, 146, 152, 143, 133, 124, 115, 108, 101, 94, 86, 82, 78, 74,
+        70, 68, 65, 63, 60, 58, 56, 54, 53, 51, 50, 48, 47, 47, 47, 47, 120,
+        125, 131, 137, 142, 134, 126, 118, 109, 103, 96, 90, 83, 79, 75, 71, 68,
+        65, 63, 60, 58, 56, 54, 53, 51, 50, 48, 47, 45, 45, 45, 45, 111, 117,
+        122, 127, 133, 126, 118, 111, 104, 98, 92, 86, 80, 76, 73, 69, 65, 63,
+        60, 58, 56, 54, 52, 51, 49, 48, 47, 45, 44, 44, 44, 44, 103, 108, 113,
+        118, 123, 117, 111, 105, 99, 93, 88, 82, 77, 73, 70, 66, 63, 60, 58, 56,
+        53, 52, 50, 49, 47, 46, 45, 44, 43, 43, 43, 43, 94, 99, 104, 109, 114,
+        109, 104, 98, 93, 88, 83, 79, 74, 70, 67, 63, 60, 58, 56, 53, 51, 50,
+        48, 47, 46, 44, 43, 42, 41, 41, 41, 41, 89, 93, 98, 103, 107, 103, 98,
+        93, 89, 84, 80, 75, 71, 68, 65, 61, 58, 56, 54, 52, 50, 49, 47, 46, 44,
+        43, 42, 41, 40, 40, 40, 40, 83, 87, 92, 96, 101, 97, 93, 88, 84, 80, 76,
+        72, 68, 65, 62, 59, 56, 54, 52, 50, 48, 47, 46, 44, 43, 42, 41, 40, 39,
+        39, 39, 39, 77, 82, 86, 90, 94, 90, 87, 83, 80, 76, 73, 69, 66, 63, 60,
+        57, 54, 53, 51, 49, 47, 46, 44, 43, 42, 41, 40, 39, 38, 38, 38, 38, 72,
+        76, 80, 83, 87, 84, 81, 78, 76, 72, 69, 66, 63, 60, 58, 55, 53, 51, 49,
+        47, 46, 44, 43, 42, 41, 40, 39, 38, 37, 37, 37, 37, 68, 72, 75, 79, 83,
+        80, 77, 75, 72, 69, 66, 64, 61, 58, 56, 54, 51, 50, 48, 46, 44, 43, 42,
+        41, 40, 39, 38, 37, 37, 37, 37, 37, 65, 68, 71, 75, 78, 76, 73, 71, 69,
+        66, 64, 61, 58, 56, 54, 52, 50, 48, 47, 45, 43, 42, 41, 40, 39, 38, 37,
+        37, 36, 36, 36, 36, 61, 64, 67, 70, 73, 71, 69, 67, 65, 63, 61, 59, 56,
+        54, 52, 50, 48, 47, 45, 44, 42, 41, 40, 39, 38, 37, 37, 36, 35, 35, 35,
+        35, 57, 60, 63, 66, 69, 67, 65, 64, 62, 60, 58, 56, 54, 52, 50, 49, 47,
+        45, 44, 43, 41, 40, 39, 38, 37, 37, 36, 35, 34, 34, 34, 34, 57, 60, 63,
+        66, 69, 67, 65, 64, 62, 60, 58, 56, 54, 52, 50, 49, 47, 45, 44, 43, 41,
+        40, 39, 38, 37, 37, 36, 35, 34, 34, 34, 34, 57, 60, 63, 66, 69, 67, 65,
+        64, 62, 60, 58, 56, 54, 52, 50, 49, 47, 45, 44, 43, 41, 40, 39, 38, 37,
+        37, 36, 35, 34, 34, 34, 34, 57, 60, 63, 66, 69, 67, 65, 64, 62, 60, 58,
+        56, 54, 52, 50, 49, 47, 45, 44, 43, 41, 40, 39, 38, 37, 37, 36, 35, 34,
+        34, 34, 34 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 39, 36, 28, 39, 32, 30, 26, 36, 30, 23, 20, 28, 26, 20, 16,
+        /* Size 8 */
+        64, 80, 43, 40, 37, 33, 29, 25, 80, 55, 42, 47, 45, 41, 36, 31, 43, 42,
+        35, 37, 37, 35, 32, 29, 40, 47, 37, 32, 31, 29, 27, 25, 37, 45, 37, 31,
+        27, 25, 24, 22, 33, 41, 35, 29, 25, 23, 21, 20, 29, 36, 32, 27, 24, 21,
+        19, 18, 25, 31, 29, 25, 22, 20, 18, 17,
+        /* Size 16 */
+        64, 72, 80, 61, 43, 41, 40, 39, 37, 35, 33, 31, 29, 27, 25, 25, 72, 70,
+        67, 55, 42, 43, 43, 42, 41, 39, 37, 35, 32, 30, 28, 28, 80, 67, 55, 48,
+        42, 44, 47, 46, 45, 43, 41, 38, 36, 34, 31, 31, 61, 55, 48, 43, 39, 40,
+        42, 41, 41, 39, 38, 36, 34, 32, 30, 30, 43, 42, 42, 39, 35, 36, 37, 37,
+        37, 36, 35, 33, 32, 30, 29, 29, 41, 43, 44, 40, 36, 35, 34, 34, 34, 33,
+        32, 31, 30, 28, 27, 27, 40, 43, 47, 42, 37, 34, 32, 31, 31, 30, 29, 28,
+        27, 26, 25, 25, 39, 42, 46, 41, 37, 34, 31, 30, 29, 28, 27, 26, 25, 25,
+        24, 24, 37, 41, 45, 41, 37, 34, 31, 29, 27, 26, 25, 24, 24, 23, 22, 22,
+        35, 39, 43, 39, 36, 33, 30, 28, 26, 25, 24, 23, 22, 22, 21, 21, 33, 37,
+        41, 38, 35, 32, 29, 27, 25, 24, 23, 22, 21, 20, 20, 20, 31, 35, 38, 36,
+        33, 31, 28, 26, 24, 23, 22, 21, 20, 19, 19, 19, 29, 32, 36, 34, 32, 30,
+        27, 25, 24, 22, 21, 20, 19, 18, 18, 18, 27, 30, 34, 32, 30, 28, 26, 25,
+        23, 22, 20, 19, 18, 18, 17, 17, 25, 28, 31, 30, 29, 27, 25, 24, 22, 21,
+        20, 19, 18, 17, 17, 17, 25, 28, 31, 30, 29, 27, 25, 24, 22, 21, 20, 19,
+        18, 17, 17, 17,
+        /* Size 32 */
+        64, 68, 72, 76, 80, 71, 61, 52, 43, 42, 41, 41, 40, 39, 39, 38, 37, 36,
+        35, 34, 33, 32, 31, 30, 29, 28, 27, 26, 25, 25, 25, 25, 68, 69, 71, 72,
+        74, 66, 58, 50, 43, 42, 42, 42, 42, 41, 41, 40, 39, 38, 37, 36, 35, 34,
+        33, 32, 31, 30, 29, 28, 27, 27, 27, 27, 72, 71, 70, 68, 67, 61, 55, 49,
+        42, 43, 43, 43, 43, 43, 42, 42, 41, 40, 39, 38, 37, 36, 35, 34, 32, 31,
+        30, 29, 28, 28, 28, 28, 76, 72, 68, 65, 61, 56, 52, 47, 42, 43, 44, 44,
+        45, 45, 44, 44, 43, 42, 41, 40, 39, 38, 37, 35, 34, 33, 32, 31, 30, 30,
+        30, 30, 80, 74, 67, 61, 55, 51, 48, 45, 42, 43, 44, 45, 47, 46, 46, 46,
+        45, 44, 43, 42, 41, 40, 38, 37, 36, 35, 34, 33, 31, 31, 31, 31, 71, 66,
+        61, 56, 51, 49, 46, 43, 40, 41, 42, 43, 44, 44, 44, 43, 43, 42, 41, 40,
+        39, 38, 37, 36, 35, 34, 33, 32, 31, 31, 31, 31, 61, 58, 55, 52, 48, 46,
+        43, 41, 39, 39, 40, 41, 42, 42, 41, 41, 41, 40, 39, 39, 38, 37, 36, 35,
+        34, 33, 32, 31, 30, 30, 30, 30, 52, 50, 49, 47, 45, 43, 41, 39, 37, 38,
+        38, 39, 39, 39, 39, 39, 39, 38, 38, 37, 36, 36, 35, 34, 33, 32, 31, 30,
+        29, 29, 29, 29, 43, 43, 42, 42, 42, 40, 39, 37, 35, 36, 36, 36, 37, 37,
+        37, 37, 37, 36, 36, 35, 35, 34, 33, 33, 32, 31, 30, 29, 29, 29, 29, 29,
+        42, 42, 43, 43, 43, 41, 39, 38, 36, 36, 36, 36, 36, 36, 35, 35, 35, 35,
+        34, 34, 33, 33, 32, 31, 31, 30, 29, 28, 28, 28, 28, 28, 41, 42, 43, 44,
+        44, 42, 40, 38, 36, 36, 35, 35, 34, 34, 34, 34, 34, 33, 33, 32, 32, 31,
+        31, 30, 30, 29, 28, 28, 27, 27, 27, 27, 41, 42, 43, 44, 45, 43, 41, 39,
+        36, 36, 35, 34, 33, 33, 33, 32, 32, 32, 31, 31, 31, 30, 30, 29, 28, 28,
+        27, 27, 26, 26, 26, 26, 40, 42, 43, 45, 47, 44, 42, 39, 37, 36, 34, 33,
+        32, 32, 31, 31, 31, 30, 30, 30, 29, 29, 28, 28, 27, 27, 26, 26, 25, 25,
+        25, 25, 39, 41, 43, 45, 46, 44, 42, 39, 37, 36, 34, 33, 32, 31, 31, 30,
+        30, 29, 29, 29, 28, 28, 27, 27, 26, 26, 25, 25, 24, 24, 24, 24, 39, 41,
+        42, 44, 46, 44, 41, 39, 37, 35, 34, 33, 31, 31, 30, 29, 29, 28, 28, 28,
+        27, 27, 26, 26, 25, 25, 25, 24, 24, 24, 24, 24, 38, 40, 42, 44, 46, 43,
+        41, 39, 37, 35, 34, 32, 31, 30, 29, 29, 28, 27, 27, 27, 26, 26, 25, 25,
+        25, 24, 24, 23, 23, 23, 23, 23, 37, 39, 41, 43, 45, 43, 41, 39, 37, 35,
+        34, 32, 31, 30, 29, 28, 27, 27, 26, 26, 25, 25, 24, 24, 24, 23, 23, 22,
+        22, 22, 22, 22, 36, 38, 40, 42, 44, 42, 40, 38, 36, 35, 33, 32, 30, 29,
+        28, 27, 27, 26, 26, 25, 24, 24, 24, 23, 23, 23, 22, 22, 21, 21, 21, 21,
+        35, 37, 39, 41, 43, 41, 39, 38, 36, 34, 33, 31, 30, 29, 28, 27, 26, 26,
+        25, 24, 24, 23, 23, 23, 22, 22, 22, 21, 21, 21, 21, 21, 34, 36, 38, 40,
+        42, 40, 39, 37, 35, 34, 32, 31, 30, 29, 28, 27, 26, 25, 24, 24, 23, 23,
+        22, 22, 22, 21, 21, 21, 20, 20, 20, 20, 33, 35, 37, 39, 41, 39, 38, 36,
+        35, 33, 32, 31, 29, 28, 27, 26, 25, 24, 24, 23, 23, 22, 22, 21, 21, 21,
+        20, 20, 20, 20, 20, 20, 32, 34, 36, 38, 40, 38, 37, 36, 34, 33, 31, 30,
+        29, 28, 27, 26, 25, 24, 23, 23, 22, 22, 21, 21, 20, 20, 20, 19, 19, 19,
+        19, 19, 31, 33, 35, 37, 38, 37, 36, 35, 33, 32, 31, 30, 28, 27, 26, 25,
+        24, 24, 23, 22, 22, 21, 21, 20, 20, 20, 19, 19, 19, 19, 19, 19, 30, 32,
+        34, 35, 37, 36, 35, 34, 33, 31, 30, 29, 28, 27, 26, 25, 24, 23, 23, 22,
+        21, 21, 20, 20, 20, 19, 19, 19, 18, 18, 18, 18, 29, 31, 32, 34, 36, 35,
+        34, 33, 32, 31, 30, 28, 27, 26, 25, 25, 24, 23, 22, 22, 21, 20, 20, 20,
+        19, 19, 18, 18, 18, 18, 18, 18, 28, 30, 31, 33, 35, 34, 33, 32, 31, 30,
+        29, 28, 27, 26, 25, 24, 23, 23, 22, 21, 21, 20, 20, 19, 19, 18, 18, 18,
+        18, 18, 18, 18, 27, 29, 30, 32, 34, 33, 32, 31, 30, 29, 28, 27, 26, 25,
+        25, 24, 23, 22, 22, 21, 20, 20, 19, 19, 18, 18, 18, 18, 17, 17, 17, 17,
+        26, 28, 29, 31, 33, 32, 31, 30, 29, 28, 28, 27, 26, 25, 24, 23, 22, 22,
+        21, 21, 20, 19, 19, 19, 18, 18, 18, 17, 17, 17, 17, 17, 25, 27, 28, 30,
+        31, 31, 30, 29, 29, 28, 27, 26, 25, 24, 24, 23, 22, 21, 21, 20, 20, 19,
+        19, 18, 18, 18, 17, 17, 17, 17, 17, 17, 25, 27, 28, 30, 31, 31, 30, 29,
+        29, 28, 27, 26, 25, 24, 24, 23, 22, 21, 21, 20, 20, 19, 19, 18, 18, 18,
+        17, 17, 17, 17, 17, 17, 25, 27, 28, 30, 31, 31, 30, 29, 29, 28, 27, 26,
+        25, 24, 24, 23, 22, 21, 21, 20, 20, 19, 19, 18, 18, 18, 17, 17, 17, 17,
+        17, 17, 25, 27, 28, 30, 31, 31, 30, 29, 29, 28, 27, 26, 25, 24, 24, 23,
+        22, 21, 21, 20, 20, 19, 19, 18, 18, 18, 17, 17, 17, 17, 17, 17 },
+      { /* Intra matrices */
+        /* Size 4 */
+        162, 98, 89, 68, 98, 80, 74, 63, 89, 74, 55, 47, 68, 63, 47, 38,
+        /* Size 8 */
+        151, 190, 99, 93, 86, 76, 66, 57, 190, 128, 97, 108, 105, 94, 82, 71,
+        99, 97, 81, 84, 85, 80, 72, 64, 93, 108, 84, 73, 69, 66, 61, 56, 86,
+        105, 85, 69, 61, 56, 52, 49, 76, 94, 80, 66, 56, 50, 46, 43, 66, 82, 72,
+        61, 52, 46, 41, 38, 57, 71, 64, 56, 49, 43, 38, 35,
+        /* Size 16 */
+        156, 176, 196, 149, 102, 99, 96, 92, 89, 83, 78, 73, 68, 63, 58, 58,
+        176, 170, 164, 133, 101, 103, 104, 101, 99, 93, 88, 82, 76, 71, 66, 66,
+        196, 164, 132, 116, 100, 106, 112, 110, 108, 103, 97, 91, 85, 79, 73,
+        73, 149, 133, 116, 104, 92, 96, 99, 99, 98, 94, 90, 85, 80, 75, 70, 70,
+        102, 101, 100, 92, 84, 85, 87, 87, 87, 85, 82, 79, 75, 71, 67, 67, 99,
+        103, 106, 96, 85, 83, 81, 80, 79, 77, 75, 72, 69, 66, 62, 62, 96, 104,
+        112, 99, 87, 81, 76, 74, 71, 70, 68, 66, 63, 61, 58, 58, 92, 101, 110,
+        99, 87, 80, 74, 70, 67, 65, 63, 61, 59, 56, 54, 54, 89, 99, 108, 98, 87,
+        79, 71, 67, 63, 60, 58, 56, 54, 52, 50, 50, 83, 93, 103, 94, 85, 77, 70,
+        65, 60, 57, 55, 53, 51, 49, 47, 47, 78, 88, 97, 90, 82, 75, 68, 63, 58,
+        55, 51, 49, 47, 46, 44, 44, 73, 82, 91, 85, 79, 72, 66, 61, 56, 53, 49,
+        47, 45, 43, 42, 42, 68, 76, 85, 80, 75, 69, 63, 59, 54, 51, 47, 45, 43,
+        41, 39, 39, 63, 71, 79, 75, 71, 66, 61, 56, 52, 49, 46, 43, 41, 39, 38,
+        38, 58, 66, 73, 70, 67, 62, 58, 54, 50, 47, 44, 42, 39, 38, 36, 36, 58,
+        66, 73, 70, 67, 62, 58, 54, 50, 47, 44, 42, 39, 38, 36, 36,
+        /* Size 32 */
+        158, 169, 179, 189, 199, 175, 152, 128, 104, 102, 101, 99, 97, 96, 94,
+        92, 90, 88, 85, 82, 79, 77, 74, 71, 69, 66, 64, 62, 59, 59, 59, 59, 169,
+        172, 176, 179, 183, 163, 143, 123, 103, 103, 102, 102, 101, 100, 98, 97,
+        95, 93, 90, 87, 84, 82, 79, 76, 73, 71, 68, 66, 63, 63, 63, 63, 179,
+        176, 173, 170, 167, 151, 135, 119, 103, 104, 104, 105, 106, 104, 103,
+        102, 100, 97, 95, 92, 89, 86, 83, 81, 78, 75, 72, 70, 67, 67, 67, 67,
+        189, 179, 170, 160, 150, 138, 126, 114, 102, 104, 106, 108, 110, 109,
+        107, 106, 105, 102, 100, 97, 94, 91, 88, 85, 82, 79, 76, 74, 71, 71, 71,
+        71, 199, 183, 167, 150, 134, 126, 118, 110, 102, 105, 108, 111, 114,
+        113, 112, 111, 110, 107, 105, 102, 99, 96, 93, 90, 86, 84, 81, 78, 75,
+        75, 75, 75, 175, 163, 151, 138, 126, 119, 112, 105, 98, 100, 103, 105,
+        107, 107, 106, 106, 105, 102, 100, 98, 95, 92, 90, 87, 84, 81, 78, 76,
+        73, 73, 73, 73, 152, 143, 135, 126, 118, 112, 106, 100, 94, 95, 97, 99,
+        101, 101, 100, 100, 100, 98, 96, 94, 91, 89, 86, 84, 81, 79, 76, 74, 71,
+        71, 71, 71, 128, 123, 119, 114, 110, 105, 100, 95, 89, 91, 92, 94, 95,
+        95, 95, 94, 94, 93, 91, 89, 88, 85, 83, 81, 79, 76, 74, 72, 69, 69, 69,
+        69, 104, 103, 103, 102, 102, 98, 94, 89, 85, 86, 87, 88, 89, 89, 89, 89,
+        89, 88, 86, 85, 84, 82, 80, 78, 76, 74, 72, 70, 68, 68, 68, 68, 102,
+        103, 104, 104, 105, 100, 95, 91, 86, 86, 86, 86, 86, 85, 85, 85, 85, 84,
+        83, 81, 80, 78, 77, 75, 73, 71, 69, 67, 66, 66, 66, 66, 101, 102, 104,
+        106, 108, 103, 97, 92, 87, 86, 85, 84, 83, 82, 82, 81, 81, 80, 79, 78,
+        77, 75, 73, 72, 70, 69, 67, 65, 63, 63, 63, 63, 99, 102, 105, 108, 111,
+        105, 99, 94, 88, 86, 84, 82, 80, 79, 78, 78, 77, 76, 75, 74, 73, 71, 70,
+        69, 67, 66, 64, 63, 61, 61, 61, 61, 97, 101, 106, 110, 114, 107, 101,
+        95, 89, 86, 83, 80, 77, 76, 75, 74, 73, 72, 71, 70, 69, 68, 67, 66, 64,
+        63, 62, 60, 59, 59, 59, 59, 96, 100, 104, 109, 113, 107, 101, 95, 89,
+        85, 82, 79, 76, 74, 73, 72, 70, 69, 68, 68, 67, 65, 64, 63, 62, 61, 60,
+        58, 57, 57, 57, 57, 94, 98, 103, 107, 112, 106, 100, 95, 89, 85, 82, 78,
+        75, 73, 72, 70, 68, 67, 66, 65, 64, 63, 62, 61, 60, 58, 57, 56, 55, 55,
+        55, 55, 92, 97, 102, 106, 111, 106, 100, 94, 89, 85, 81, 78, 74, 72, 70,
+        68, 66, 65, 64, 63, 61, 60, 59, 58, 57, 56, 55, 54, 53, 53, 53, 53, 90,
+        95, 100, 105, 110, 105, 100, 94, 89, 85, 81, 77, 73, 70, 68, 66, 64, 63,
+        61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 51, 51, 51, 88, 93, 97, 102,
+        107, 102, 98, 93, 88, 84, 80, 76, 72, 69, 67, 65, 63, 61, 60, 59, 57,
+        56, 55, 54, 53, 52, 51, 50, 49, 49, 49, 49, 85, 90, 95, 100, 105, 100,
+        96, 91, 86, 83, 79, 75, 71, 68, 66, 64, 61, 60, 58, 57, 56, 55, 53, 52,
+        51, 51, 50, 49, 48, 48, 48, 48, 82, 87, 92, 97, 102, 98, 94, 89, 85, 81,
+        78, 74, 70, 68, 65, 63, 60, 59, 57, 55, 54, 53, 52, 51, 50, 49, 48, 47,
+        46, 46, 46, 46, 79, 84, 89, 94, 99, 95, 91, 88, 84, 80, 77, 73, 69, 67,
+        64, 61, 59, 57, 56, 54, 52, 51, 50, 49, 48, 47, 46, 46, 45, 45, 45, 45,
+        77, 82, 86, 91, 96, 92, 89, 85, 82, 78, 75, 71, 68, 65, 63, 60, 58, 56,
+        55, 53, 51, 50, 49, 48, 47, 46, 45, 44, 44, 44, 44, 44, 74, 79, 83, 88,
+        93, 90, 86, 83, 80, 77, 73, 70, 67, 64, 62, 59, 57, 55, 53, 52, 50, 49,
+        48, 47, 46, 45, 44, 43, 42, 42, 42, 42, 71, 76, 81, 85, 90, 87, 84, 81,
+        78, 75, 72, 69, 66, 63, 61, 58, 56, 54, 52, 51, 49, 48, 47, 46, 45, 44,
+        43, 42, 41, 41, 41, 41, 69, 73, 78, 82, 86, 84, 81, 79, 76, 73, 70, 67,
+        64, 62, 60, 57, 55, 53, 51, 50, 48, 47, 46, 45, 43, 43, 42, 41, 40, 40,
+        40, 40, 66, 71, 75, 79, 84, 81, 79, 76, 74, 71, 69, 66, 63, 61, 58, 56,
+        54, 52, 51, 49, 47, 46, 45, 44, 43, 42, 41, 40, 39, 39, 39, 39, 64, 68,
+        72, 76, 81, 78, 76, 74, 72, 69, 67, 64, 62, 60, 57, 55, 53, 51, 50, 48,
+        46, 45, 44, 43, 42, 41, 40, 39, 38, 38, 38, 38, 62, 66, 70, 74, 78, 76,
+        74, 72, 70, 67, 65, 63, 60, 58, 56, 54, 52, 50, 49, 47, 46, 44, 43, 42,
+        41, 40, 39, 38, 38, 38, 38, 38, 59, 63, 67, 71, 75, 73, 71, 69, 68, 66,
+        63, 61, 59, 57, 55, 53, 51, 49, 48, 46, 45, 44, 42, 41, 40, 39, 38, 38,
+        37, 37, 37, 37, 59, 63, 67, 71, 75, 73, 71, 69, 68, 66, 63, 61, 59, 57,
+        55, 53, 51, 49, 48, 46, 45, 44, 42, 41, 40, 39, 38, 38, 37, 37, 37, 37,
+        59, 63, 67, 71, 75, 73, 71, 69, 68, 66, 63, 61, 59, 57, 55, 53, 51, 49,
+        48, 46, 45, 44, 42, 41, 40, 39, 38, 38, 37, 37, 37, 37, 59, 63, 67, 71,
+        75, 73, 71, 69, 68, 66, 63, 61, 59, 57, 55, 53, 51, 49, 48, 46, 45, 44,
+        42, 41, 40, 39, 38, 38, 37, 37, 37, 37 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 58, 35, 23, 58, 39, 28, 21, 35, 28, 20, 16, 23, 21, 16, 14,
+        /* Size 8 */
+        64, 85, 79, 60, 44, 34, 27, 23, 85, 74, 78, 66, 52, 40, 32, 26, 79, 78,
+        55, 47, 40, 34, 29, 24, 60, 66, 47, 37, 32, 28, 25, 22, 44, 52, 40, 32,
+        27, 24, 22, 20, 34, 40, 34, 28, 24, 21, 19, 18, 27, 32, 29, 25, 22, 19,
+        18, 17, 23, 26, 24, 22, 20, 18, 17, 16,
+        /* Size 16 */
+        64, 75, 85, 82, 79, 69, 60, 52, 44, 39, 34, 31, 27, 25, 23, 23, 75, 77,
+        80, 79, 78, 71, 63, 55, 48, 43, 37, 33, 30, 27, 25, 25, 85, 80, 74, 76,
+        78, 72, 66, 59, 52, 46, 40, 36, 32, 29, 26, 26, 82, 79, 76, 71, 66, 62,
+        57, 51, 46, 42, 37, 34, 30, 28, 25, 25, 79, 78, 78, 66, 55, 51, 47, 44,
+        40, 37, 34, 31, 29, 26, 24, 24, 69, 71, 72, 62, 51, 47, 42, 39, 36, 33,
+        31, 29, 27, 25, 23, 23, 60, 63, 66, 57, 47, 42, 37, 34, 32, 30, 28, 26,
+        25, 23, 22, 22, 52, 55, 59, 51, 44, 39, 34, 32, 29, 28, 26, 24, 23, 22,
+        21, 21, 44, 48, 52, 46, 40, 36, 32, 29, 27, 25, 24, 23, 22, 21, 20, 20,
+        39, 43, 46, 42, 37, 33, 30, 28, 25, 24, 22, 21, 20, 20, 19, 19, 34, 37,
+        40, 37, 34, 31, 28, 26, 24, 22, 21, 20, 19, 19, 18, 18, 31, 33, 36, 34,
+        31, 29, 26, 24, 23, 21, 20, 19, 19, 18, 17, 17, 27, 30, 32, 30, 29, 27,
+        25, 23, 22, 20, 19, 19, 18, 17, 17, 17, 25, 27, 29, 28, 26, 25, 23, 22,
+        21, 20, 19, 18, 17, 17, 16, 16, 23, 25, 26, 25, 24, 23, 22, 21, 20, 19,
+        18, 17, 17, 16, 16, 16, 23, 25, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17,
+        17, 16, 16, 16,
+        /* Size 32 */
+        64, 69, 75, 80, 85, 84, 82, 80, 79, 74, 69, 64, 60, 56, 52, 48, 44, 42,
+        39, 37, 34, 32, 31, 29, 27, 26, 25, 24, 23, 23, 23, 23, 69, 73, 76, 79,
+        83, 82, 81, 80, 78, 74, 70, 66, 61, 58, 54, 50, 46, 44, 41, 38, 36, 34,
+        32, 30, 29, 27, 26, 25, 24, 24, 24, 24, 75, 76, 77, 79, 80, 79, 79, 79,
+        78, 74, 71, 67, 63, 59, 55, 52, 48, 45, 43, 40, 37, 35, 33, 32, 30, 28,
+        27, 26, 25, 25, 25, 25, 80, 79, 79, 78, 77, 77, 77, 78, 78, 75, 71, 68,
+        65, 61, 57, 54, 50, 47, 44, 41, 39, 37, 35, 33, 31, 30, 28, 27, 26, 26,
+        26, 26, 85, 83, 80, 77, 74, 75, 76, 77, 78, 75, 72, 69, 66, 63, 59, 55,
+        52, 49, 46, 43, 40, 38, 36, 34, 32, 31, 29, 28, 26, 26, 26, 26, 84, 82,
+        79, 77, 75, 74, 74, 73, 72, 69, 67, 64, 61, 58, 55, 52, 49, 46, 44, 41,
+        39, 37, 35, 33, 31, 30, 29, 27, 26, 26, 26, 26, 82, 81, 79, 77, 76, 74,
+        71, 69, 66, 64, 62, 59, 57, 54, 51, 49, 46, 44, 42, 39, 37, 35, 34, 32,
+        30, 29, 28, 27, 25, 25, 25, 25, 80, 80, 79, 78, 77, 73, 69, 65, 61, 59,
+        56, 54, 52, 50, 48, 45, 43, 41, 39, 37, 35, 34, 32, 31, 29, 28, 27, 26,
+        25, 25, 25, 25, 79, 78, 78, 78, 78, 72, 66, 61, 55, 53, 51, 49, 47, 46,
+        44, 42, 40, 39, 37, 36, 34, 33, 31, 30, 29, 27, 26, 25, 24, 24, 24, 24,
+        74, 74, 74, 75, 75, 69, 64, 59, 53, 51, 49, 47, 45, 43, 41, 40, 38, 37,
+        35, 34, 32, 31, 30, 29, 28, 27, 26, 25, 24, 24, 24, 24, 69, 70, 71, 71,
+        72, 67, 62, 56, 51, 49, 47, 44, 42, 41, 39, 38, 36, 35, 33, 32, 31, 30,
+        29, 28, 27, 26, 25, 24, 23, 23, 23, 23, 64, 66, 67, 68, 69, 64, 59, 54,
+        49, 47, 44, 42, 40, 38, 37, 35, 34, 33, 32, 31, 29, 28, 28, 27, 26, 25,
+        24, 23, 23, 23, 23, 23, 60, 61, 63, 65, 66, 61, 57, 52, 47, 45, 42, 40,
+        37, 36, 34, 33, 32, 31, 30, 29, 28, 27, 26, 25, 25, 24, 23, 23, 22, 22,
+        22, 22, 56, 58, 59, 61, 63, 58, 54, 50, 46, 43, 41, 38, 36, 34, 33, 32,
+        31, 30, 29, 28, 27, 26, 25, 25, 24, 23, 23, 22, 21, 21, 21, 21, 52, 54,
+        55, 57, 59, 55, 51, 48, 44, 41, 39, 37, 34, 33, 32, 31, 29, 28, 28, 27,
+        26, 25, 24, 24, 23, 23, 22, 21, 21, 21, 21, 21, 48, 50, 52, 54, 55, 52,
+        49, 45, 42, 40, 38, 35, 33, 32, 31, 29, 28, 27, 26, 26, 25, 24, 24, 23,
+        22, 22, 21, 21, 20, 20, 20, 20, 44, 46, 48, 50, 52, 49, 46, 43, 40, 38,
+        36, 34, 32, 31, 29, 28, 27, 26, 25, 25, 24, 23, 23, 22, 22, 21, 21, 20,
+        20, 20, 20, 20, 42, 44, 45, 47, 49, 46, 44, 41, 39, 37, 35, 33, 31, 30,
+        28, 27, 26, 25, 25, 24, 23, 23, 22, 22, 21, 21, 20, 20, 19, 19, 19, 19,
+        39, 41, 43, 44, 46, 44, 42, 39, 37, 35, 33, 32, 30, 29, 28, 26, 25, 25,
+        24, 23, 22, 22, 21, 21, 20, 20, 20, 19, 19, 19, 19, 19, 37, 38, 40, 41,
+        43, 41, 39, 37, 36, 34, 32, 31, 29, 28, 27, 26, 25, 24, 23, 22, 22, 21,
+        21, 20, 20, 20, 19, 19, 19, 19, 19, 19, 34, 36, 37, 39, 40, 39, 37, 35,
+        34, 32, 31, 29, 28, 27, 26, 25, 24, 23, 22, 22, 21, 21, 20, 20, 19, 19,
+        19, 18, 18, 18, 18, 18, 32, 34, 35, 37, 38, 37, 35, 34, 33, 31, 30, 28,
+        27, 26, 25, 24, 23, 23, 22, 21, 21, 20, 20, 19, 19, 19, 18, 18, 18, 18,
+        18, 18, 31, 32, 33, 35, 36, 35, 34, 32, 31, 30, 29, 28, 26, 25, 24, 24,
+        23, 22, 21, 21, 20, 20, 19, 19, 19, 18, 18, 18, 17, 17, 17, 17, 29, 30,
+        32, 33, 34, 33, 32, 31, 30, 29, 28, 27, 25, 25, 24, 23, 22, 22, 21, 20,
+        20, 19, 19, 19, 18, 18, 18, 17, 17, 17, 17, 17, 27, 29, 30, 31, 32, 31,
+        30, 29, 29, 28, 27, 26, 25, 24, 23, 22, 22, 21, 20, 20, 19, 19, 19, 18,
+        18, 18, 17, 17, 17, 17, 17, 17, 26, 27, 28, 30, 31, 30, 29, 28, 27, 27,
+        26, 25, 24, 23, 23, 22, 21, 21, 20, 20, 19, 19, 18, 18, 18, 17, 17, 17,
+        17, 17, 17, 17, 25, 26, 27, 28, 29, 29, 28, 27, 26, 26, 25, 24, 23, 23,
+        22, 21, 21, 20, 20, 19, 19, 18, 18, 18, 17, 17, 17, 17, 16, 16, 16, 16,
+        24, 25, 26, 27, 28, 27, 27, 26, 25, 25, 24, 23, 23, 22, 21, 21, 20, 20,
+        19, 19, 18, 18, 18, 17, 17, 17, 17, 16, 16, 16, 16, 16, 23, 24, 25, 26,
+        26, 26, 25, 25, 24, 24, 23, 23, 22, 21, 21, 20, 20, 19, 19, 19, 18, 18,
+        17, 17, 17, 17, 16, 16, 16, 16, 16, 16, 23, 24, 25, 26, 26, 26, 25, 25,
+        24, 24, 23, 23, 22, 21, 21, 20, 20, 19, 19, 19, 18, 18, 17, 17, 17, 17,
+        16, 16, 16, 16, 16, 16, 23, 24, 25, 26, 26, 26, 25, 25, 24, 24, 23, 23,
+        22, 21, 21, 20, 20, 19, 19, 19, 18, 18, 17, 17, 17, 17, 16, 16, 16, 16,
+        16, 16, 23, 24, 25, 26, 26, 26, 25, 25, 24, 24, 23, 23, 22, 21, 21, 20,
+        20, 19, 19, 19, 18, 18, 17, 17, 17, 17, 16, 16, 16, 16, 16, 16 },
+      { /* Intra matrices */
+        /* Size 4 */
+        192, 174, 103, 63, 174, 113, 79, 57, 103, 79, 55, 44, 63, 57, 44, 37,
+        /* Size 8 */
+        160, 216, 199, 149, 109, 82, 65, 53, 216, 187, 195, 166, 128, 98, 77,
+        62, 199, 195, 137, 116, 98, 81, 67, 57, 149, 166, 116, 90, 76, 66, 57,
+        50, 109, 128, 98, 76, 63, 55, 49, 45, 82, 98, 81, 66, 55, 48, 44, 40,
+        65, 77, 67, 57, 49, 44, 40, 37, 53, 62, 57, 50, 45, 40, 37, 35,
+        /* Size 16 */
+        167, 195, 224, 216, 207, 181, 155, 134, 113, 99, 85, 76, 67, 61, 55, 55,
+        195, 202, 209, 207, 205, 184, 164, 143, 123, 108, 94, 84, 74, 67, 60,
+        60, 224, 209, 194, 199, 203, 188, 172, 153, 133, 117, 102, 91, 80, 72,
+        65, 65, 216, 207, 199, 186, 173, 160, 147, 132, 118, 106, 93, 84, 75,
+        68, 62, 62, 207, 205, 203, 173, 143, 132, 121, 112, 102, 94, 85, 78, 70,
+        65, 59, 59, 181, 184, 188, 160, 132, 120, 107, 99, 91, 84, 77, 71, 65,
+        60, 56, 56, 155, 164, 172, 147, 121, 107, 94, 86, 79, 74, 69, 64, 60,
+        56, 53, 53, 134, 143, 153, 132, 112, 99, 86, 79, 72, 68, 63, 59, 56, 53,
+        50, 50, 113, 123, 133, 118, 102, 91, 79, 72, 66, 62, 57, 54, 51, 49, 47,
+        47, 99, 108, 117, 106, 94, 84, 74, 68, 62, 58, 54, 51, 48, 46, 44, 44,
+        85, 94, 102, 93, 85, 77, 69, 63, 57, 54, 50, 48, 46, 44, 42, 42, 76, 84,
+        91, 84, 78, 71, 64, 59, 54, 51, 48, 46, 44, 42, 40, 40, 67, 74, 80, 75,
+        70, 65, 60, 56, 51, 48, 46, 44, 42, 40, 39, 39, 61, 67, 72, 68, 65, 60,
+        56, 53, 49, 46, 44, 42, 40, 39, 38, 38, 55, 60, 65, 62, 59, 56, 53, 50,
+        47, 44, 42, 40, 39, 38, 36, 36, 55, 60, 65, 62, 59, 56, 53, 50, 47, 44,
+        42, 40, 39, 38, 36, 36,
+        /* Size 32 */
+        170, 185, 200, 214, 229, 225, 220, 215, 211, 198, 185, 171, 158, 147,
+        137, 126, 116, 109, 101, 94, 87, 83, 78, 73, 69, 66, 62, 59, 56, 56, 56,
+        56, 185, 194, 203, 212, 221, 219, 216, 213, 210, 198, 186, 174, 163,
+        152, 142, 131, 121, 113, 106, 99, 91, 87, 82, 77, 72, 69, 65, 62, 59,
+        59, 59, 59, 200, 203, 207, 210, 214, 213, 212, 210, 209, 199, 188, 178,
+        167, 157, 146, 136, 126, 118, 111, 103, 96, 90, 85, 80, 75, 72, 68, 65,
+        61, 61, 61, 61, 214, 212, 210, 208, 206, 207, 207, 208, 209, 199, 190,
+        181, 172, 161, 151, 141, 131, 123, 115, 108, 100, 94, 89, 84, 78, 75,
+        71, 67, 64, 64, 64, 64, 229, 221, 214, 206, 199, 201, 203, 205, 208,
+        200, 192, 184, 176, 166, 156, 146, 136, 128, 120, 112, 104, 98, 93, 87,
+        81, 78, 74, 70, 66, 66, 66, 66, 225, 219, 213, 207, 201, 199, 197, 194,
+        192, 185, 178, 170, 163, 154, 145, 137, 128, 121, 114, 107, 100, 95, 89,
+        84, 79, 75, 72, 68, 65, 65, 65, 65, 220, 216, 212, 207, 203, 197, 190,
+        183, 177, 170, 163, 157, 150, 142, 135, 128, 120, 114, 108, 102, 95, 91,
+        86, 81, 77, 73, 70, 67, 63, 63, 63, 63, 215, 213, 210, 208, 205, 194,
+        183, 172, 161, 155, 149, 143, 137, 131, 125, 118, 112, 107, 102, 96, 91,
+        87, 83, 78, 74, 71, 68, 65, 62, 62, 62, 62, 211, 210, 209, 209, 208,
+        192, 177, 161, 146, 140, 135, 129, 124, 119, 114, 109, 105, 100, 96, 91,
+        87, 83, 79, 75, 72, 69, 66, 63, 60, 60, 60, 60, 198, 198, 199, 199, 200,
+        185, 170, 155, 140, 134, 128, 123, 117, 112, 108, 103, 99, 95, 91, 87,
+        82, 79, 76, 72, 69, 66, 64, 61, 59, 59, 59, 59, 185, 186, 188, 190, 192,
+        178, 163, 149, 135, 128, 122, 116, 110, 105, 101, 97, 93, 89, 86, 82,
+        78, 75, 72, 69, 66, 64, 62, 59, 57, 57, 57, 57, 171, 174, 178, 181, 184,
+        170, 157, 143, 129, 123, 116, 109, 103, 99, 95, 91, 87, 84, 81, 77, 74,
+        72, 69, 66, 64, 62, 60, 57, 55, 55, 55, 55, 158, 163, 167, 172, 176,
+        163, 150, 137, 124, 117, 110, 103, 96, 92, 88, 85, 81, 78, 75, 73, 70,
+        68, 66, 63, 61, 59, 57, 55, 54, 54, 54, 54, 147, 152, 157, 161, 166,
+        154, 142, 131, 119, 112, 105, 99, 92, 88, 85, 81, 77, 75, 72, 70, 67,
+        65, 63, 61, 59, 57, 55, 54, 52, 52, 52, 52, 137, 142, 146, 151, 156,
+        145, 135, 125, 114, 108, 101, 95, 88, 85, 81, 78, 74, 72, 69, 67, 64,
+        62, 61, 59, 57, 55, 54, 52, 51, 51, 51, 51, 126, 131, 136, 141, 146,
+        137, 128, 118, 109, 103, 97, 91, 85, 81, 78, 74, 71, 68, 66, 64, 61, 60,
+        58, 56, 55, 53, 52, 50, 49, 49, 49, 49, 116, 121, 126, 131, 136, 128,
+        120, 112, 105, 99, 93, 87, 81, 77, 74, 71, 67, 65, 63, 61, 59, 57, 56,
+        54, 52, 51, 50, 49, 48, 48, 48, 48, 109, 113, 118, 123, 128, 121, 114,
+        107, 100, 95, 89, 84, 78, 75, 72, 68, 65, 63, 61, 59, 57, 55, 54, 52,
+        51, 50, 49, 48, 46, 46, 46, 46, 101, 106, 111, 115, 120, 114, 108, 102,
+        96, 91, 86, 81, 75, 72, 69, 66, 63, 61, 59, 57, 55, 54, 52, 51, 49, 48,
+        47, 46, 45, 45, 45, 45, 94, 99, 103, 108, 112, 107, 102, 96, 91, 87, 82,
+        77, 73, 70, 67, 64, 61, 59, 57, 55, 53, 52, 51, 49, 48, 47, 46, 45, 44,
+        44, 44, 44, 87, 91, 96, 100, 104, 100, 95, 91, 87, 82, 78, 74, 70, 67,
+        64, 61, 59, 57, 55, 53, 51, 50, 49, 48, 47, 46, 45, 44, 43, 43, 43, 43,
+        83, 87, 90, 94, 98, 95, 91, 87, 83, 79, 75, 72, 68, 65, 62, 60, 57, 55,
+        54, 52, 50, 49, 48, 47, 45, 45, 44, 43, 42, 42, 42, 42, 78, 82, 85, 89,
+        93, 89, 86, 83, 79, 76, 72, 69, 66, 63, 61, 58, 56, 54, 52, 51, 49, 48,
+        47, 46, 44, 44, 43, 42, 41, 41, 41, 41, 73, 77, 80, 84, 87, 84, 81, 78,
+        75, 72, 69, 66, 63, 61, 59, 56, 54, 52, 51, 49, 48, 47, 46, 45, 43, 43,
+        42, 41, 40, 40, 40, 40, 69, 72, 75, 78, 81, 79, 77, 74, 72, 69, 66, 64,
+        61, 59, 57, 55, 52, 51, 49, 48, 47, 45, 44, 43, 42, 42, 41, 40, 40, 40,
+        40, 40, 66, 69, 72, 75, 78, 75, 73, 71, 69, 66, 64, 62, 59, 57, 55, 53,
+        51, 50, 48, 47, 46, 45, 44, 43, 42, 41, 40, 40, 39, 39, 39, 39, 62, 65,
+        68, 71, 74, 72, 70, 68, 66, 64, 62, 60, 57, 55, 54, 52, 50, 49, 47, 46,
+        45, 44, 43, 42, 41, 40, 40, 39, 38, 38, 38, 38, 59, 62, 65, 67, 70, 68,
+        67, 65, 63, 61, 59, 57, 55, 54, 52, 50, 49, 48, 46, 45, 44, 43, 42, 41,
+        40, 40, 39, 38, 38, 38, 38, 38, 56, 59, 61, 64, 66, 65, 63, 62, 60, 59,
+        57, 55, 54, 52, 51, 49, 48, 46, 45, 44, 43, 42, 41, 40, 40, 39, 38, 38,
+        37, 37, 37, 37, 56, 59, 61, 64, 66, 65, 63, 62, 60, 59, 57, 55, 54, 52,
+        51, 49, 48, 46, 45, 44, 43, 42, 41, 40, 40, 39, 38, 38, 37, 37, 37, 37,
+        56, 59, 61, 64, 66, 65, 63, 62, 60, 59, 57, 55, 54, 52, 51, 49, 48, 46,
+        45, 44, 43, 42, 41, 40, 40, 39, 38, 38, 37, 37, 37, 37, 56, 59, 61, 64,
+        66, 65, 63, 62, 60, 59, 57, 55, 54, 52, 51, 49, 48, 46, 45, 44, 43, 42,
+        41, 40, 40, 39, 38, 38, 37, 37, 37, 37 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 40, 37, 29, 40, 34, 32, 27, 37, 32, 25, 22, 29, 27, 22, 18,
+        /* Size 8 */
+        64, 79, 44, 41, 39, 35, 31, 27, 79, 55, 43, 47, 46, 42, 37, 33, 44, 43,
+        37, 38, 38, 36, 33, 30, 41, 47, 38, 34, 32, 31, 29, 27, 39, 46, 38, 32,
+        29, 27, 25, 24, 35, 42, 36, 31, 27, 24, 23, 21, 31, 37, 33, 29, 25, 23,
+        21, 20, 27, 33, 30, 27, 24, 21, 20, 19,
+        /* Size 16 */
+        64, 72, 79, 62, 44, 42, 41, 40, 39, 37, 35, 33, 31, 29, 27, 27, 72, 69,
+        67, 55, 43, 44, 44, 43, 42, 40, 38, 36, 34, 32, 30, 30, 79, 67, 55, 49,
+        43, 45, 47, 47, 46, 44, 42, 39, 37, 35, 33, 33, 62, 55, 49, 44, 40, 41,
+        43, 42, 42, 41, 39, 37, 35, 33, 31, 31, 44, 43, 43, 40, 37, 37, 38, 38,
+        38, 37, 36, 35, 33, 32, 30, 30, 42, 44, 45, 41, 37, 37, 36, 35, 35, 34,
+        33, 32, 31, 30, 28, 28, 41, 44, 47, 43, 38, 36, 34, 33, 32, 31, 31, 30,
+        29, 28, 27, 27, 40, 43, 47, 42, 38, 35, 33, 32, 30, 29, 29, 28, 27, 26,
+        25, 25, 39, 42, 46, 42, 38, 35, 32, 30, 29, 28, 27, 26, 25, 25, 24, 24,
+        37, 40, 44, 41, 37, 34, 31, 29, 28, 27, 26, 25, 24, 23, 23, 23, 35, 38,
+        42, 39, 36, 33, 31, 29, 27, 26, 24, 24, 23, 22, 21, 21, 33, 36, 39, 37,
+        35, 32, 30, 28, 26, 25, 24, 23, 22, 21, 21, 21, 31, 34, 37, 35, 33, 31,
+        29, 27, 25, 24, 23, 22, 21, 20, 20, 20, 29, 32, 35, 33, 32, 30, 28, 26,
+        25, 23, 22, 21, 20, 20, 19, 19, 27, 30, 33, 31, 30, 28, 27, 25, 24, 23,
+        21, 21, 20, 19, 19, 19, 27, 30, 33, 31, 30, 28, 27, 25, 24, 23, 21, 21,
+        20, 19, 19, 19,
+        /* Size 32 */
+        64, 68, 72, 75, 79, 70, 62, 53, 44, 43, 42, 42, 41, 41, 40, 39, 39, 38,
+        37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 27, 27, 27, 68, 69, 71, 72,
+        73, 66, 58, 51, 43, 43, 43, 43, 43, 42, 42, 41, 40, 39, 38, 37, 36, 35,
+        34, 33, 32, 31, 30, 29, 28, 28, 28, 28, 72, 71, 69, 68, 67, 61, 55, 49,
+        43, 44, 44, 44, 44, 44, 43, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 33,
+        32, 31, 30, 30, 30, 30, 75, 72, 68, 65, 61, 57, 52, 48, 43, 44, 44, 45,
+        46, 45, 45, 45, 44, 43, 42, 41, 40, 39, 38, 37, 35, 34, 33, 32, 31, 31,
+        31, 31, 79, 73, 67, 61, 55, 52, 49, 46, 43, 44, 45, 46, 47, 47, 47, 46,
+        46, 45, 44, 43, 42, 41, 39, 38, 37, 36, 35, 34, 33, 33, 33, 33, 70, 66,
+        61, 57, 52, 49, 47, 44, 41, 42, 43, 44, 45, 45, 44, 44, 44, 43, 42, 41,
+        40, 39, 38, 37, 36, 35, 34, 33, 32, 32, 32, 32, 62, 58, 55, 52, 49, 47,
+        44, 42, 40, 40, 41, 42, 43, 42, 42, 42, 42, 41, 41, 40, 39, 38, 37, 36,
+        35, 34, 33, 32, 31, 31, 31, 31, 53, 51, 49, 48, 46, 44, 42, 40, 38, 39,
+        39, 40, 40, 40, 40, 40, 40, 39, 39, 38, 38, 37, 36, 35, 34, 33, 32, 32,
+        31, 31, 31, 31, 44, 43, 43, 43, 43, 41, 40, 38, 37, 37, 37, 38, 38, 38,
+        38, 38, 38, 38, 37, 37, 36, 35, 35, 34, 33, 32, 32, 31, 30, 30, 30, 30,
+        43, 43, 44, 44, 44, 42, 40, 39, 37, 37, 37, 37, 37, 37, 37, 37, 37, 36,
+        36, 35, 35, 34, 33, 33, 32, 31, 31, 30, 29, 29, 29, 29, 42, 43, 44, 44,
+        45, 43, 41, 39, 37, 37, 37, 36, 36, 36, 35, 35, 35, 35, 34, 34, 33, 33,
+        32, 32, 31, 30, 30, 29, 28, 28, 28, 28, 42, 43, 44, 45, 46, 44, 42, 40,
+        38, 37, 36, 35, 35, 34, 34, 34, 33, 33, 33, 32, 32, 31, 31, 30, 30, 29,
+        29, 28, 28, 28, 28, 28, 41, 43, 44, 46, 47, 45, 43, 40, 38, 37, 36, 35,
+        34, 33, 33, 32, 32, 32, 31, 31, 31, 30, 30, 29, 29, 28, 28, 27, 27, 27,
+        27, 27, 41, 42, 44, 45, 47, 45, 42, 40, 38, 37, 36, 34, 33, 33, 32, 32,
+        31, 31, 30, 30, 30, 29, 29, 28, 28, 27, 27, 27, 26, 26, 26, 26, 40, 42,
+        43, 45, 47, 44, 42, 40, 38, 37, 35, 34, 33, 32, 32, 31, 30, 30, 29, 29,
+        29, 28, 28, 27, 27, 27, 26, 26, 25, 25, 25, 25, 39, 41, 43, 45, 46, 44,
+        42, 40, 38, 37, 35, 34, 32, 32, 31, 30, 29, 29, 29, 28, 28, 27, 27, 27,
+        26, 26, 25, 25, 25, 25, 25, 25, 39, 40, 42, 44, 46, 44, 42, 40, 38, 37,
+        35, 33, 32, 31, 30, 29, 29, 28, 28, 27, 27, 26, 26, 26, 25, 25, 25, 24,
+        24, 24, 24, 24, 38, 39, 41, 43, 45, 43, 41, 39, 38, 36, 35, 33, 32, 31,
+        30, 29, 28, 28, 27, 27, 26, 26, 25, 25, 25, 24, 24, 24, 23, 23, 23, 23,
+        37, 38, 40, 42, 44, 42, 41, 39, 37, 36, 34, 33, 31, 30, 29, 29, 28, 27,
+        27, 26, 26, 25, 25, 24, 24, 24, 23, 23, 23, 23, 23, 23, 36, 37, 39, 41,
+        43, 41, 40, 38, 37, 35, 34, 32, 31, 30, 29, 28, 27, 27, 26, 26, 25, 25,
+        24, 24, 23, 23, 23, 22, 22, 22, 22, 22, 35, 36, 38, 40, 42, 40, 39, 38,
+        36, 35, 33, 32, 31, 30, 29, 28, 27, 26, 26, 25, 24, 24, 24, 23, 23, 22,
+        22, 22, 21, 21, 21, 21, 34, 35, 37, 39, 41, 39, 38, 37, 35, 34, 33, 31,
+        30, 29, 28, 27, 26, 26, 25, 25, 24, 24, 23, 23, 22, 22, 22, 21, 21, 21,
+        21, 21, 33, 34, 36, 38, 39, 38, 37, 36, 35, 33, 32, 31, 30, 29, 28, 27,
+        26, 25, 25, 24, 24, 23, 23, 22, 22, 22, 21, 21, 21, 21, 21, 21, 32, 33,
+        35, 37, 38, 37, 36, 35, 34, 33, 32, 30, 29, 28, 27, 27, 26, 25, 24, 24,
+        23, 23, 22, 22, 21, 21, 21, 21, 20, 20, 20, 20, 31, 32, 34, 35, 37, 36,
+        35, 34, 33, 32, 31, 30, 29, 28, 27, 26, 25, 25, 24, 23, 23, 22, 22, 21,
+        21, 21, 20, 20, 20, 20, 20, 20, 30, 31, 33, 34, 36, 35, 34, 33, 32, 31,
+        30, 29, 28, 27, 27, 26, 25, 24, 24, 23, 22, 22, 22, 21, 21, 20, 20, 20,
+        19, 19, 19, 19, 29, 30, 32, 33, 35, 34, 33, 32, 32, 31, 30, 29, 28, 27,
+        26, 25, 25, 24, 23, 23, 22, 22, 21, 21, 20, 20, 20, 19, 19, 19, 19, 19,
+        28, 29, 31, 32, 34, 33, 32, 32, 31, 30, 29, 28, 27, 27, 26, 25, 24, 24,
+        23, 22, 22, 21, 21, 21, 20, 20, 19, 19, 19, 19, 19, 19, 27, 28, 30, 31,
+        33, 32, 31, 31, 30, 29, 28, 28, 27, 26, 25, 25, 24, 23, 23, 22, 21, 21,
+        21, 20, 20, 19, 19, 19, 19, 19, 19, 19, 27, 28, 30, 31, 33, 32, 31, 31,
+        30, 29, 28, 28, 27, 26, 25, 25, 24, 23, 23, 22, 21, 21, 21, 20, 20, 19,
+        19, 19, 19, 19, 19, 19, 27, 28, 30, 31, 33, 32, 31, 31, 30, 29, 28, 28,
+        27, 26, 25, 25, 24, 23, 23, 22, 21, 21, 21, 20, 20, 19, 19, 19, 19, 19,
+        19, 19, 27, 28, 30, 31, 33, 32, 31, 31, 30, 29, 28, 28, 27, 26, 25, 25,
+        24, 23, 23, 22, 21, 21, 21, 20, 20, 19, 19, 19, 19, 19, 19, 19 },
+      { /* Intra matrices */
+        /* Size 4 */
+        152, 94, 86, 67, 94, 77, 72, 62, 86, 72, 55, 48, 67, 62, 48, 40,
+        /* Size 8 */
+        141, 176, 95, 89, 83, 74, 65, 57, 176, 120, 93, 103, 100, 90, 80, 70,
+        95, 93, 79, 81, 82, 77, 71, 64, 89, 103, 81, 71, 68, 65, 61, 56, 83,
+        100, 82, 68, 60, 56, 53, 49, 74, 90, 77, 65, 56, 50, 47, 44, 65, 80, 71,
+        61, 53, 47, 43, 40, 57, 70, 64, 56, 49, 44, 40, 37,
+        /* Size 16 */
+        145, 163, 181, 139, 97, 94, 91, 88, 85, 81, 76, 71, 66, 62, 58, 58, 163,
+        158, 152, 124, 96, 98, 99, 96, 94, 89, 84, 79, 74, 70, 65, 65, 181, 152,
+        124, 110, 95, 101, 106, 104, 103, 98, 93, 87, 82, 77, 72, 72, 139, 124,
+        110, 99, 88, 91, 95, 94, 93, 90, 86, 82, 77, 73, 69, 69, 97, 96, 95, 88,
+        81, 82, 84, 84, 84, 82, 80, 76, 73, 69, 65, 65, 94, 98, 101, 91, 82, 80,
+        79, 78, 77, 75, 73, 70, 68, 65, 62, 62, 91, 99, 106, 95, 84, 79, 74, 72,
+        70, 68, 67, 65, 63, 60, 58, 58, 88, 96, 104, 94, 84, 78, 72, 69, 66, 64,
+        62, 60, 58, 56, 54, 54, 85, 94, 103, 93, 84, 77, 70, 66, 62, 60, 58, 56,
+        54, 52, 51, 51, 81, 89, 98, 90, 82, 75, 68, 64, 60, 57, 55, 53, 51, 50,
+        48, 48, 76, 84, 93, 86, 80, 73, 67, 62, 58, 55, 52, 50, 48, 47, 45, 45,
+        71, 79, 87, 82, 76, 70, 65, 60, 56, 53, 50, 48, 46, 45, 43, 43, 66, 74,
+        82, 77, 73, 68, 63, 58, 54, 51, 48, 46, 44, 43, 41, 41, 62, 70, 77, 73,
+        69, 65, 60, 56, 52, 50, 47, 45, 43, 41, 40, 40, 58, 65, 72, 69, 65, 62,
+        58, 54, 51, 48, 45, 43, 41, 40, 38, 38, 58, 65, 72, 69, 65, 62, 58, 54,
+        51, 48, 45, 43, 41, 40, 38, 38,
+        /* Size 32 */
+        147, 156, 165, 175, 184, 163, 141, 120, 99, 97, 96, 94, 93, 91, 90, 88,
+        87, 84, 82, 79, 77, 75, 72, 70, 67, 65, 63, 61, 59, 59, 59, 59, 156,
+        160, 163, 166, 169, 151, 134, 116, 98, 98, 97, 97, 97, 95, 94, 92, 91,
+        89, 86, 84, 81, 79, 76, 74, 71, 69, 67, 65, 62, 62, 62, 62, 165, 163,
+        160, 157, 155, 140, 126, 112, 98, 98, 99, 100, 100, 99, 98, 97, 95, 93,
+        91, 88, 86, 83, 80, 78, 75, 73, 71, 68, 66, 66, 66, 66, 175, 166, 157,
+        149, 140, 129, 119, 108, 97, 99, 101, 102, 104, 103, 102, 101, 100, 97,
+        95, 93, 90, 87, 85, 82, 79, 77, 74, 72, 69, 69, 69, 69, 184, 169, 155,
+        140, 126, 118, 111, 104, 97, 100, 102, 105, 108, 107, 106, 105, 104,
+        102, 99, 97, 94, 92, 89, 86, 83, 81, 78, 75, 73, 73, 73, 73, 163, 151,
+        140, 129, 118, 112, 106, 99, 93, 95, 98, 100, 102, 101, 101, 100, 100,
+        97, 95, 93, 91, 88, 86, 83, 81, 78, 76, 74, 71, 71, 71, 71, 141, 134,
+        126, 119, 111, 106, 100, 95, 89, 91, 93, 95, 96, 96, 96, 95, 95, 93, 91,
+        89, 88, 85, 83, 81, 79, 76, 74, 72, 70, 70, 70, 70, 120, 116, 112, 108,
+        104, 99, 95, 90, 86, 87, 88, 89, 91, 91, 90, 90, 90, 89, 87, 86, 84, 82,
+        80, 78, 76, 74, 72, 70, 68, 68, 68, 68, 99, 98, 98, 97, 97, 93, 89, 86,
+        82, 83, 84, 84, 85, 85, 85, 85, 85, 84, 83, 82, 81, 79, 77, 76, 74, 72,
+        70, 68, 66, 66, 66, 66, 97, 98, 98, 99, 100, 95, 91, 87, 83, 83, 83, 83,
+        82, 82, 82, 82, 82, 81, 80, 79, 78, 76, 74, 73, 71, 70, 68, 66, 64, 64,
+        64, 64, 96, 97, 99, 101, 102, 98, 93, 88, 84, 83, 82, 81, 80, 79, 79,
+        79, 78, 77, 76, 75, 74, 73, 71, 70, 69, 67, 66, 64, 63, 63, 63, 63, 94,
+        97, 100, 102, 105, 100, 95, 89, 84, 83, 81, 79, 77, 77, 76, 75, 75, 74,
+        73, 72, 71, 70, 69, 67, 66, 65, 63, 62, 61, 61, 61, 61, 93, 97, 100,
+        104, 108, 102, 96, 91, 85, 82, 80, 77, 75, 74, 73, 72, 71, 70, 69, 68,
+        68, 67, 66, 65, 64, 62, 61, 60, 59, 59, 59, 59, 91, 95, 99, 103, 107,
+        101, 96, 91, 85, 82, 79, 77, 74, 72, 71, 70, 69, 68, 67, 66, 65, 64, 63,
+        62, 61, 60, 59, 58, 57, 57, 57, 57, 90, 94, 98, 102, 106, 101, 96, 90,
+        85, 82, 79, 76, 73, 71, 70, 68, 67, 66, 65, 64, 63, 62, 61, 60, 59, 58,
+        57, 56, 55, 55, 55, 55, 88, 92, 97, 101, 105, 100, 95, 90, 85, 82, 79,
+        75, 72, 70, 68, 67, 65, 64, 63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 53,
+        53, 53, 53, 87, 91, 95, 100, 104, 100, 95, 90, 85, 82, 78, 75, 71, 69,
+        67, 65, 63, 62, 61, 60, 58, 58, 57, 56, 55, 54, 53, 52, 52, 52, 52, 52,
+        84, 89, 93, 97, 102, 97, 93, 89, 84, 81, 77, 74, 70, 68, 66, 64, 62, 61,
+        59, 58, 57, 56, 55, 54, 53, 53, 52, 51, 50, 50, 50, 50, 82, 86, 91, 95,
+        99, 95, 91, 87, 83, 80, 76, 73, 69, 67, 65, 63, 61, 59, 58, 57, 56, 55,
+        54, 53, 52, 51, 50, 50, 49, 49, 49, 49, 79, 84, 88, 93, 97, 93, 89, 86,
+        82, 79, 75, 72, 68, 66, 64, 62, 60, 58, 57, 56, 54, 53, 52, 51, 50, 50,
+        49, 48, 47, 47, 47, 47, 77, 81, 86, 90, 94, 91, 88, 84, 81, 78, 74, 71,
+        68, 65, 63, 61, 58, 57, 56, 54, 53, 52, 51, 50, 49, 48, 47, 47, 46, 46,
+        46, 46, 75, 79, 83, 87, 92, 88, 85, 82, 79, 76, 73, 70, 67, 64, 62, 60,
+        58, 56, 55, 53, 52, 51, 50, 49, 48, 47, 46, 46, 45, 45, 45, 45, 72, 76,
+        80, 85, 89, 86, 83, 80, 77, 74, 71, 69, 66, 63, 61, 59, 57, 55, 54, 52,
+        51, 50, 49, 48, 47, 46, 45, 45, 44, 44, 44, 44, 70, 74, 78, 82, 86, 83,
+        81, 78, 76, 73, 70, 67, 65, 62, 60, 58, 56, 54, 53, 51, 50, 49, 48, 47,
+        46, 45, 44, 44, 43, 43, 43, 43, 67, 71, 75, 79, 83, 81, 79, 76, 74, 71,
+        69, 66, 64, 61, 59, 57, 55, 53, 52, 50, 49, 48, 47, 46, 45, 44, 43, 43,
+        42, 42, 42, 42, 65, 69, 73, 77, 81, 78, 76, 74, 72, 70, 67, 65, 62, 60,
+        58, 56, 54, 53, 51, 50, 48, 47, 46, 45, 44, 43, 43, 42, 41, 41, 41, 41,
+        63, 67, 71, 74, 78, 76, 74, 72, 70, 68, 66, 63, 61, 59, 57, 55, 53, 52,
+        50, 49, 47, 46, 45, 44, 43, 43, 42, 41, 40, 40, 40, 40, 61, 65, 68, 72,
+        75, 74, 72, 70, 68, 66, 64, 62, 60, 58, 56, 54, 52, 51, 50, 48, 47, 46,
+        45, 44, 43, 42, 41, 40, 40, 40, 40, 40, 59, 62, 66, 69, 73, 71, 70, 68,
+        66, 64, 63, 61, 59, 57, 55, 53, 52, 50, 49, 47, 46, 45, 44, 43, 42, 41,
+        40, 40, 39, 39, 39, 39, 59, 62, 66, 69, 73, 71, 70, 68, 66, 64, 63, 61,
+        59, 57, 55, 53, 52, 50, 49, 47, 46, 45, 44, 43, 42, 41, 40, 40, 39, 39,
+        39, 39, 59, 62, 66, 69, 73, 71, 70, 68, 66, 64, 63, 61, 59, 57, 55, 53,
+        52, 50, 49, 47, 46, 45, 44, 43, 42, 41, 40, 40, 39, 39, 39, 39, 59, 62,
+        66, 69, 73, 71, 70, 68, 66, 64, 63, 61, 59, 57, 55, 53, 52, 50, 49, 47,
+        46, 45, 44, 43, 42, 41, 40, 40, 39, 39, 39, 39 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 59, 37, 24, 59, 40, 29, 23, 37, 29, 22, 19, 24, 23, 19, 16,
+        /* Size 8 */
+        64, 84, 78, 60, 45, 36, 29, 25, 84, 74, 77, 66, 52, 41, 34, 28, 78, 77,
+        56, 48, 42, 36, 30, 27, 60, 66, 48, 39, 34, 30, 27, 24, 45, 52, 42, 34,
+        29, 26, 24, 22, 36, 41, 36, 30, 26, 23, 22, 21, 29, 34, 30, 27, 24, 22,
+        20, 19, 25, 28, 27, 24, 22, 21, 19, 19,
+        /* Size 16 */
+        64, 74, 84, 81, 78, 69, 60, 53, 45, 41, 36, 33, 29, 27, 25, 25, 74, 77,
+        79, 78, 77, 70, 63, 56, 49, 44, 39, 35, 32, 29, 27, 27, 84, 79, 74, 75,
+        77, 71, 66, 59, 52, 47, 41, 38, 34, 31, 28, 28, 81, 78, 75, 71, 66, 62,
+        57, 52, 47, 43, 38, 35, 32, 30, 28, 28, 78, 77, 77, 66, 56, 52, 48, 45,
+        42, 39, 36, 33, 30, 29, 27, 27, 69, 70, 71, 62, 52, 48, 43, 40, 38, 35,
+        33, 31, 29, 27, 25, 25, 60, 63, 66, 57, 48, 43, 39, 36, 34, 32, 30, 28,
+        27, 26, 24, 24, 53, 56, 59, 52, 45, 40, 36, 34, 31, 30, 28, 27, 25, 24,
+        23, 23, 45, 49, 52, 47, 42, 38, 34, 31, 29, 27, 26, 25, 24, 23, 22, 22,
+        41, 44, 47, 43, 39, 35, 32, 30, 27, 26, 25, 24, 23, 22, 21, 21, 36, 39,
+        41, 38, 36, 33, 30, 28, 26, 25, 23, 23, 22, 21, 21, 21, 33, 35, 38, 35,
+        33, 31, 28, 27, 25, 24, 23, 22, 21, 21, 20, 20, 29, 32, 34, 32, 30, 29,
+        27, 25, 24, 23, 22, 21, 20, 20, 19, 19, 27, 29, 31, 30, 29, 27, 26, 24,
+        23, 22, 21, 21, 20, 20, 19, 19, 25, 27, 28, 28, 27, 25, 24, 23, 22, 21,
+        21, 20, 19, 19, 19, 19, 25, 27, 28, 28, 27, 25, 24, 23, 22, 21, 21, 20,
+        19, 19, 19, 19,
+        /* Size 32 */
+        64, 69, 74, 79, 84, 83, 81, 80, 78, 73, 69, 64, 60, 56, 53, 49, 45, 43,
+        41, 38, 36, 34, 33, 31, 29, 28, 27, 26, 25, 25, 25, 25, 69, 72, 75, 78,
+        82, 81, 80, 79, 78, 74, 70, 66, 61, 58, 54, 51, 47, 45, 42, 40, 37, 35,
+        34, 32, 30, 29, 28, 27, 26, 26, 26, 26, 74, 75, 77, 78, 79, 79, 78, 78,
+        77, 74, 70, 67, 63, 59, 56, 52, 49, 46, 44, 41, 39, 37, 35, 33, 32, 30,
+        29, 28, 27, 27, 27, 27, 79, 78, 78, 77, 76, 77, 77, 77, 77, 74, 71, 68,
+        65, 61, 58, 54, 51, 48, 45, 43, 40, 38, 36, 35, 33, 31, 30, 29, 28, 28,
+        28, 28, 84, 82, 79, 76, 74, 75, 75, 76, 77, 74, 71, 69, 66, 63, 59, 56,
+        52, 50, 47, 44, 41, 40, 38, 36, 34, 32, 31, 30, 28, 28, 28, 28, 83, 81,
+        79, 77, 75, 74, 73, 72, 72, 69, 67, 64, 62, 59, 56, 53, 50, 47, 45, 42,
+        40, 38, 36, 35, 33, 32, 30, 29, 28, 28, 28, 28, 81, 80, 78, 77, 75, 73,
+        71, 69, 66, 64, 62, 59, 57, 55, 52, 50, 47, 45, 43, 41, 38, 37, 35, 34,
+        32, 31, 30, 29, 28, 28, 28, 28, 80, 79, 78, 77, 76, 72, 69, 65, 61, 59,
+        57, 55, 53, 51, 48, 46, 44, 43, 41, 39, 37, 36, 34, 33, 31, 30, 29, 28,
+        27, 27, 27, 27, 78, 78, 77, 77, 77, 72, 66, 61, 56, 54, 52, 50, 48, 47,
+        45, 43, 42, 40, 39, 37, 36, 34, 33, 32, 30, 29, 29, 28, 27, 27, 27, 27,
+        73, 74, 74, 74, 74, 69, 64, 59, 54, 52, 50, 48, 46, 44, 43, 41, 40, 38,
+        37, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, 26, 26, 26, 69, 70, 70, 71,
+        71, 67, 62, 57, 52, 50, 48, 46, 43, 42, 40, 39, 38, 36, 35, 34, 33, 32,
+        31, 30, 29, 28, 27, 26, 25, 25, 25, 25, 64, 66, 67, 68, 69, 64, 59, 55,
+        50, 48, 46, 43, 41, 40, 38, 37, 36, 35, 33, 32, 31, 30, 30, 29, 28, 27,
+        26, 26, 25, 25, 25, 25, 60, 61, 63, 65, 66, 62, 57, 53, 48, 46, 43, 41,
+        39, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 28, 27, 26, 26, 25, 24, 24,
+        24, 24, 56, 58, 59, 61, 63, 59, 55, 51, 47, 44, 42, 40, 37, 36, 35, 34,
+        32, 32, 31, 30, 29, 28, 27, 27, 26, 25, 25, 24, 24, 24, 24, 24, 53, 54,
+        56, 58, 59, 56, 52, 48, 45, 43, 40, 38, 36, 35, 34, 32, 31, 30, 30, 29,
+        28, 27, 27, 26, 25, 25, 24, 24, 23, 23, 23, 23, 49, 51, 52, 54, 56, 53,
+        50, 46, 43, 41, 39, 37, 35, 34, 32, 31, 30, 29, 28, 28, 27, 26, 26, 25,
+        25, 24, 24, 23, 23, 23, 23, 23, 45, 47, 49, 51, 52, 50, 47, 44, 42, 40,
+        38, 36, 34, 32, 31, 30, 29, 28, 27, 27, 26, 25, 25, 24, 24, 23, 23, 23,
+        22, 22, 22, 22, 43, 45, 46, 48, 50, 47, 45, 43, 40, 38, 36, 35, 33, 32,
+        30, 29, 28, 27, 27, 26, 25, 25, 24, 24, 23, 23, 23, 22, 22, 22, 22, 22,
+        41, 42, 44, 45, 47, 45, 43, 41, 39, 37, 35, 33, 32, 31, 30, 28, 27, 27,
+        26, 25, 25, 24, 24, 23, 23, 22, 22, 22, 21, 21, 21, 21, 38, 40, 41, 43,
+        44, 42, 41, 39, 37, 35, 34, 32, 31, 30, 29, 28, 27, 26, 25, 25, 24, 24,
+        23, 23, 22, 22, 22, 21, 21, 21, 21, 21, 36, 37, 39, 40, 41, 40, 38, 37,
+        36, 34, 33, 31, 30, 29, 28, 27, 26, 25, 25, 24, 23, 23, 23, 22, 22, 22,
+        21, 21, 21, 21, 21, 21, 34, 35, 37, 38, 40, 38, 37, 36, 34, 33, 32, 30,
+        29, 28, 27, 26, 25, 25, 24, 24, 23, 23, 22, 22, 21, 21, 21, 21, 20, 20,
+        20, 20, 33, 34, 35, 36, 38, 36, 35, 34, 33, 32, 31, 30, 28, 27, 27, 26,
+        25, 24, 24, 23, 23, 22, 22, 22, 21, 21, 21, 20, 20, 20, 20, 20, 31, 32,
+        33, 35, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, 25, 24, 24, 23, 23,
+        22, 22, 22, 21, 21, 21, 20, 20, 20, 20, 20, 20, 29, 30, 32, 33, 34, 33,
+        32, 31, 30, 30, 29, 28, 27, 26, 25, 25, 24, 23, 23, 22, 22, 21, 21, 21,
+        20, 20, 20, 20, 19, 19, 19, 19, 28, 29, 30, 31, 32, 32, 31, 30, 29, 29,
+        28, 27, 26, 25, 25, 24, 23, 23, 22, 22, 22, 21, 21, 21, 20, 20, 20, 19,
+        19, 19, 19, 19, 27, 28, 29, 30, 31, 30, 30, 29, 29, 28, 27, 26, 26, 25,
+        24, 24, 23, 23, 22, 22, 21, 21, 21, 20, 20, 20, 20, 19, 19, 19, 19, 19,
+        26, 27, 28, 29, 30, 29, 29, 28, 28, 27, 26, 26, 25, 24, 24, 23, 23, 22,
+        22, 21, 21, 21, 20, 20, 20, 19, 19, 19, 19, 19, 19, 19, 25, 26, 27, 28,
+        28, 28, 28, 27, 27, 26, 25, 25, 24, 24, 23, 23, 22, 22, 21, 21, 21, 20,
+        20, 20, 19, 19, 19, 19, 19, 19, 19, 19, 25, 26, 27, 28, 28, 28, 28, 27,
+        27, 26, 25, 25, 24, 24, 23, 23, 22, 22, 21, 21, 21, 20, 20, 20, 19, 19,
+        19, 19, 19, 19, 19, 19, 25, 26, 27, 28, 28, 28, 28, 27, 27, 26, 25, 25,
+        24, 24, 23, 23, 22, 22, 21, 21, 21, 20, 20, 20, 19, 19, 19, 19, 19, 19,
+        19, 19, 25, 26, 27, 28, 28, 28, 28, 27, 27, 26, 25, 25, 24, 24, 23, 23,
+        22, 22, 21, 21, 21, 20, 20, 20, 19, 19, 19, 19, 19, 19, 19, 19 },
+      { /* Intra matrices */
+        /* Size 4 */
+        171, 157, 95, 62, 157, 104, 75, 57, 95, 75, 54, 45, 62, 57, 45, 39,
+        /* Size 8 */
+        144, 192, 177, 135, 101, 78, 63, 53, 192, 167, 174, 149, 117, 91, 73,
+        61, 177, 174, 125, 107, 92, 77, 65, 56, 135, 149, 107, 84, 73, 64, 57,
+        51, 101, 117, 92, 73, 62, 55, 50, 46, 78, 91, 77, 64, 55, 49, 45, 42,
+        63, 73, 65, 57, 50, 45, 42, 39, 53, 61, 56, 51, 46, 42, 39, 38,
+        /* Size 16 */
+        149, 174, 199, 191, 184, 162, 140, 122, 104, 92, 81, 73, 65, 60, 55, 55,
+        174, 180, 186, 184, 182, 165, 147, 130, 113, 100, 88, 79, 70, 65, 59,
+        59, 199, 186, 173, 177, 181, 168, 154, 138, 121, 108, 95, 85, 76, 69,
+        63, 63, 191, 184, 177, 166, 155, 144, 133, 120, 108, 98, 87, 80, 72, 66,
+        61, 61, 184, 182, 181, 155, 129, 120, 111, 103, 95, 88, 80, 74, 68, 63,
+        58, 58, 162, 165, 168, 144, 120, 110, 99, 92, 85, 79, 73, 68, 63, 59,
+        55, 55, 140, 147, 154, 133, 111, 99, 88, 81, 75, 71, 66, 63, 59, 56, 53,
+        53, 122, 130, 138, 120, 103, 92, 81, 75, 70, 66, 62, 58, 55, 53, 50, 50,
+        104, 113, 121, 108, 95, 85, 75, 70, 64, 60, 57, 54, 52, 50, 48, 48, 92,
+        100, 108, 98, 88, 79, 71, 66, 60, 57, 54, 51, 49, 47, 46, 46, 81, 88,
+        95, 87, 80, 73, 66, 62, 57, 54, 51, 49, 47, 45, 44, 44, 73, 79, 85, 80,
+        74, 68, 63, 58, 54, 51, 49, 47, 45, 44, 42, 42, 65, 70, 76, 72, 68, 63,
+        59, 55, 52, 49, 47, 45, 43, 42, 41, 41, 60, 65, 69, 66, 63, 59, 56, 53,
+        50, 47, 45, 44, 42, 41, 40, 40, 55, 59, 63, 61, 58, 55, 53, 50, 48, 46,
+        44, 42, 41, 40, 39, 39, 55, 59, 63, 61, 58, 55, 53, 50, 48, 46, 44, 42,
+        41, 40, 39, 39,
+        /* Size 32 */
+        152, 165, 177, 190, 202, 198, 195, 191, 187, 176, 165, 153, 142, 133,
+        124, 115, 106, 100, 94, 88, 82, 78, 74, 70, 66, 64, 61, 58, 56, 56, 56,
+        56, 165, 173, 180, 188, 196, 193, 191, 189, 186, 176, 166, 156, 146,
+        137, 128, 119, 111, 104, 98, 92, 86, 82, 77, 73, 69, 66, 63, 61, 58, 58,
+        58, 58, 177, 180, 183, 186, 189, 188, 187, 187, 186, 177, 168, 159, 150,
+        141, 132, 124, 115, 108, 102, 96, 89, 85, 80, 76, 72, 69, 66, 63, 60,
+        60, 60, 60, 190, 188, 186, 185, 183, 183, 184, 184, 185, 177, 169, 161,
+        154, 145, 136, 128, 119, 112, 106, 99, 93, 88, 84, 79, 74, 71, 68, 65,
+        62, 62, 62, 62, 202, 196, 189, 183, 176, 178, 180, 182, 184, 177, 171,
+        164, 157, 149, 140, 132, 123, 117, 110, 103, 96, 92, 87, 82, 77, 74, 71,
+        67, 64, 64, 64, 64, 198, 193, 188, 183, 178, 177, 175, 173, 171, 165,
+        159, 152, 146, 139, 131, 124, 117, 111, 105, 99, 93, 88, 84, 80, 75, 72,
+        69, 66, 63, 63, 63, 63, 195, 191, 187, 184, 180, 175, 169, 164, 158,
+        152, 147, 141, 135, 129, 123, 116, 110, 105, 99, 94, 89, 85, 81, 77, 73,
+        70, 67, 65, 62, 62, 62, 62, 191, 189, 187, 184, 182, 173, 164, 154, 145,
+        140, 134, 129, 124, 119, 114, 109, 103, 99, 94, 90, 85, 82, 78, 75, 71,
+        68, 66, 63, 61, 61, 61, 61, 187, 186, 186, 185, 184, 171, 158, 145, 132,
+        127, 122, 118, 113, 109, 105, 101, 97, 93, 89, 85, 82, 78, 75, 72, 69,
+        67, 64, 62, 59, 59, 59, 59, 176, 176, 177, 177, 177, 165, 152, 140, 127,
+        122, 117, 112, 107, 103, 99, 96, 92, 88, 85, 81, 78, 75, 72, 70, 67, 64,
+        62, 60, 58, 58, 58, 58, 165, 166, 168, 169, 171, 159, 147, 134, 122,
+        117, 112, 106, 101, 97, 94, 90, 87, 84, 81, 78, 75, 72, 69, 67, 64, 62,
+        60, 58, 56, 56, 56, 56, 153, 156, 159, 161, 164, 152, 141, 129, 118,
+        112, 106, 101, 95, 92, 88, 85, 82, 79, 76, 74, 71, 69, 67, 64, 62, 60,
+        59, 57, 55, 55, 55, 55, 142, 146, 150, 154, 157, 146, 135, 124, 113,
+        107, 101, 95, 89, 86, 83, 80, 77, 74, 72, 70, 68, 66, 64, 62, 60, 58,
+        57, 55, 54, 54, 54, 54, 133, 137, 141, 145, 149, 139, 129, 119, 109,
+        103, 97, 92, 86, 83, 80, 77, 74, 72, 69, 67, 65, 63, 62, 60, 58, 57, 55,
+        54, 52, 52, 52, 52, 124, 128, 132, 136, 140, 131, 123, 114, 105, 99, 94,
+        88, 83, 80, 77, 74, 71, 69, 67, 65, 63, 61, 59, 58, 56, 55, 54, 52, 51,
+        51, 51, 51, 115, 119, 124, 128, 132, 124, 116, 109, 101, 96, 90, 85, 80,
+        77, 74, 71, 68, 66, 64, 62, 60, 59, 57, 56, 54, 53, 52, 51, 50, 50, 50,
+        50, 106, 111, 115, 119, 123, 117, 110, 103, 97, 92, 87, 82, 77, 74, 71,
+        68, 65, 63, 61, 60, 58, 57, 55, 54, 53, 52, 51, 50, 48, 48, 48, 48, 100,
+        104, 108, 112, 117, 111, 105, 99, 93, 88, 84, 79, 74, 72, 69, 66, 63,
+        61, 60, 58, 56, 55, 54, 53, 51, 50, 49, 48, 47, 47, 47, 47, 94, 98, 102,
+        106, 110, 105, 99, 94, 89, 85, 81, 76, 72, 69, 67, 64, 61, 60, 58, 56,
+        55, 54, 52, 51, 50, 49, 48, 47, 46, 46, 46, 46, 88, 92, 96, 99, 103, 99,
+        94, 90, 85, 81, 78, 74, 70, 67, 65, 62, 60, 58, 56, 55, 53, 52, 51, 50,
+        49, 48, 47, 46, 46, 46, 46, 46, 82, 86, 89, 93, 96, 93, 89, 85, 82, 78,
+        75, 71, 68, 65, 63, 60, 58, 56, 55, 53, 52, 51, 50, 49, 48, 47, 46, 45,
+        45, 45, 45, 45, 78, 82, 85, 88, 92, 88, 85, 82, 78, 75, 72, 69, 66, 63,
+        61, 59, 57, 55, 54, 52, 51, 50, 49, 48, 47, 46, 45, 45, 44, 44, 44, 44,
+        74, 77, 80, 84, 87, 84, 81, 78, 75, 72, 69, 67, 64, 62, 59, 57, 55, 54,
+        52, 51, 50, 49, 48, 47, 46, 45, 44, 44, 43, 43, 43, 43, 70, 73, 76, 79,
+        82, 80, 77, 75, 72, 70, 67, 64, 62, 60, 58, 56, 54, 53, 51, 50, 49, 48,
+        47, 46, 45, 44, 44, 43, 42, 42, 42, 42, 66, 69, 72, 74, 77, 75, 73, 71,
+        69, 67, 64, 62, 60, 58, 56, 54, 53, 51, 50, 49, 48, 47, 46, 45, 44, 44,
+        43, 42, 42, 42, 42, 42, 64, 66, 69, 71, 74, 72, 70, 68, 67, 64, 62, 60,
+        58, 57, 55, 53, 52, 50, 49, 48, 47, 46, 45, 44, 44, 43, 42, 42, 41, 41,
+        41, 41, 61, 63, 66, 68, 71, 69, 67, 66, 64, 62, 60, 59, 57, 55, 54, 52,
+        51, 49, 48, 47, 46, 45, 44, 44, 43, 42, 42, 41, 41, 41, 41, 41, 58, 61,
+        63, 65, 67, 66, 65, 63, 62, 60, 58, 57, 55, 54, 52, 51, 50, 48, 47, 46,
+        45, 45, 44, 43, 42, 42, 41, 41, 40, 40, 40, 40, 56, 58, 60, 62, 64, 63,
+        62, 61, 59, 58, 56, 55, 54, 52, 51, 50, 48, 47, 46, 46, 45, 44, 43, 42,
+        42, 41, 41, 40, 40, 40, 40, 40, 56, 58, 60, 62, 64, 63, 62, 61, 59, 58,
+        56, 55, 54, 52, 51, 50, 48, 47, 46, 46, 45, 44, 43, 42, 42, 41, 41, 40,
+        40, 40, 40, 40, 56, 58, 60, 62, 64, 63, 62, 61, 59, 58, 56, 55, 54, 52,
+        51, 50, 48, 47, 46, 46, 45, 44, 43, 42, 42, 41, 41, 40, 40, 40, 40, 40,
+        56, 58, 60, 62, 64, 63, 62, 61, 59, 58, 56, 55, 54, 52, 51, 50, 48, 47,
+        46, 46, 45, 44, 43, 42, 42, 41, 41, 40, 40, 40, 40, 40 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 41, 38, 31, 41, 35, 33, 29, 38, 33, 26, 23, 31, 29, 23, 20,
+        /* Size 8 */
+        64, 79, 45, 42, 40, 36, 32, 29, 79, 55, 44, 48, 47, 43, 38, 34, 45, 44,
+        38, 39, 39, 37, 35, 32, 42, 48, 39, 35, 33, 32, 31, 29, 40, 47, 39, 33,
+        30, 28, 27, 26, 36, 43, 37, 32, 28, 26, 25, 23, 32, 38, 35, 31, 27, 25,
+        23, 22, 29, 34, 32, 29, 26, 23, 22, 21,
+        /* Size 16 */
+        64, 71, 79, 62, 45, 43, 42, 41, 40, 38, 36, 34, 32, 30, 29, 29, 71, 69,
+        67, 56, 44, 45, 45, 44, 43, 41, 39, 37, 35, 33, 31, 31, 79, 67, 55, 50,
+        44, 46, 48, 47, 47, 45, 43, 41, 38, 36, 34, 34, 62, 56, 50, 45, 41, 42,
+        44, 43, 43, 42, 40, 38, 37, 35, 33, 33, 45, 44, 44, 41, 38, 39, 39, 39,
+        39, 38, 37, 36, 35, 33, 32, 32, 43, 45, 46, 42, 39, 38, 37, 37, 36, 36,
+        35, 34, 33, 31, 30, 30, 42, 45, 48, 44, 39, 37, 35, 34, 33, 33, 32, 31,
+        31, 30, 29, 29, 41, 44, 47, 43, 39, 37, 34, 33, 32, 31, 30, 30, 29, 28,
+        27, 27, 40, 43, 47, 43, 39, 36, 33, 32, 30, 29, 28, 28, 27, 26, 26, 26,
+        38, 41, 45, 42, 38, 36, 33, 31, 29, 28, 27, 27, 26, 25, 25, 25, 36, 39,
+        43, 40, 37, 35, 32, 30, 28, 27, 26, 25, 25, 24, 23, 23, 34, 37, 41, 38,
+        36, 34, 31, 30, 28, 27, 25, 25, 24, 23, 23, 23, 32, 35, 38, 37, 35, 33,
+        31, 29, 27, 26, 25, 24, 23, 22, 22, 22, 30, 33, 36, 35, 33, 31, 30, 28,
+        26, 25, 24, 23, 22, 22, 21, 21, 29, 31, 34, 33, 32, 30, 29, 27, 26, 25,
+        23, 23, 22, 21, 21, 21, 29, 31, 34, 33, 32, 30, 29, 27, 26, 25, 23, 23,
+        22, 21, 21, 21,
+        /* Size 32 */
+        64, 68, 71, 75, 79, 70, 62, 53, 45, 44, 43, 43, 42, 42, 41, 40, 40, 39,
+        38, 37, 36, 35, 34, 33, 32, 31, 30, 30, 29, 29, 29, 29, 68, 69, 70, 72,
+        73, 66, 59, 52, 44, 44, 44, 44, 44, 43, 43, 42, 42, 41, 40, 39, 38, 37,
+        36, 35, 34, 33, 32, 31, 30, 30, 30, 30, 71, 70, 69, 68, 67, 61, 56, 50,
+        44, 44, 45, 45, 45, 45, 44, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34,
+        33, 32, 31, 31, 31, 31, 75, 72, 68, 65, 61, 57, 53, 48, 44, 45, 45, 46,
+        47, 46, 46, 45, 45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 33,
+        33, 33, 79, 73, 67, 61, 55, 52, 50, 47, 44, 45, 46, 47, 48, 48, 47, 47,
+        47, 46, 45, 44, 43, 42, 41, 40, 38, 37, 36, 35, 34, 34, 34, 34, 70, 66,
+        61, 57, 52, 50, 47, 45, 42, 43, 44, 45, 46, 46, 45, 45, 45, 44, 43, 42,
+        42, 41, 39, 38, 37, 36, 36, 35, 34, 34, 34, 34, 62, 59, 56, 53, 50, 47,
+        45, 43, 41, 42, 42, 43, 44, 43, 43, 43, 43, 42, 42, 41, 40, 39, 38, 37,
+        37, 36, 35, 34, 33, 33, 33, 33, 53, 52, 50, 48, 47, 45, 43, 41, 39, 40,
+        40, 41, 41, 41, 41, 41, 41, 41, 40, 39, 39, 38, 37, 36, 36, 35, 34, 33,
+        32, 32, 32, 32, 45, 44, 44, 44, 44, 42, 41, 39, 38, 38, 39, 39, 39, 39,
+        39, 39, 39, 39, 38, 38, 37, 37, 36, 35, 35, 34, 33, 32, 32, 32, 32, 32,
+        44, 44, 44, 45, 45, 43, 42, 40, 38, 38, 38, 38, 38, 38, 38, 38, 38, 37,
+        37, 37, 36, 36, 35, 34, 34, 33, 32, 32, 31, 31, 31, 31, 43, 44, 45, 45,
+        46, 44, 42, 40, 39, 38, 38, 37, 37, 37, 37, 37, 36, 36, 36, 35, 35, 34,
+        34, 33, 33, 32, 31, 31, 30, 30, 30, 30, 43, 44, 45, 46, 47, 45, 43, 41,
+        39, 38, 37, 37, 36, 36, 35, 35, 35, 35, 34, 34, 34, 33, 33, 32, 32, 31,
+        30, 30, 29, 29, 29, 29, 42, 44, 45, 47, 48, 46, 44, 41, 39, 38, 37, 36,
+        35, 35, 34, 34, 33, 33, 33, 33, 32, 32, 31, 31, 31, 30, 30, 29, 29, 29,
+        29, 29, 42, 43, 45, 46, 48, 46, 43, 41, 39, 38, 37, 36, 35, 34, 34, 33,
+        33, 32, 32, 32, 31, 31, 30, 30, 30, 29, 29, 28, 28, 28, 28, 28, 41, 43,
+        44, 46, 47, 45, 43, 41, 39, 38, 37, 35, 34, 34, 33, 32, 32, 31, 31, 31,
+        30, 30, 30, 29, 29, 28, 28, 28, 27, 27, 27, 27, 40, 42, 44, 45, 47, 45,
+        43, 41, 39, 38, 37, 35, 34, 33, 32, 32, 31, 31, 30, 30, 29, 29, 29, 28,
+        28, 28, 27, 27, 26, 26, 26, 26, 40, 42, 43, 45, 47, 45, 43, 41, 39, 38,
+        36, 35, 33, 33, 32, 31, 30, 30, 29, 29, 28, 28, 28, 27, 27, 27, 26, 26,
+        26, 26, 26, 26, 39, 41, 42, 44, 46, 44, 42, 41, 39, 37, 36, 35, 33, 32,
+        31, 31, 30, 29, 29, 28, 28, 28, 27, 27, 26, 26, 26, 25, 25, 25, 25, 25,
+        38, 40, 41, 43, 45, 43, 42, 40, 38, 37, 36, 34, 33, 32, 31, 30, 29, 29,
+        28, 28, 27, 27, 27, 26, 26, 26, 25, 25, 25, 25, 25, 25, 37, 39, 40, 42,
+        44, 42, 41, 39, 38, 37, 35, 34, 33, 32, 31, 30, 29, 28, 28, 27, 27, 26,
+        26, 26, 25, 25, 25, 24, 24, 24, 24, 24, 36, 38, 39, 41, 43, 42, 40, 39,
+        37, 36, 35, 34, 32, 31, 30, 29, 28, 28, 27, 27, 26, 26, 25, 25, 25, 24,
+        24, 24, 23, 23, 23, 23, 35, 37, 38, 40, 42, 41, 39, 38, 37, 36, 34, 33,
+        32, 31, 30, 29, 28, 28, 27, 26, 26, 25, 25, 25, 24, 24, 24, 23, 23, 23,
+        23, 23, 34, 36, 37, 39, 41, 39, 38, 37, 36, 35, 34, 33, 31, 30, 30, 29,
+        28, 27, 27, 26, 25, 25, 25, 24, 24, 24, 23, 23, 23, 23, 23, 23, 33, 35,
+        36, 38, 40, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 27, 26, 26,
+        25, 25, 24, 24, 23, 23, 23, 23, 22, 22, 22, 22, 32, 34, 35, 37, 38, 37,
+        37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, 26, 25, 25, 24, 24, 23,
+        23, 23, 22, 22, 22, 22, 22, 22, 31, 33, 34, 36, 37, 36, 36, 35, 34, 33,
+        32, 31, 30, 29, 28, 28, 27, 26, 26, 25, 24, 24, 24, 23, 23, 22, 22, 22,
+        22, 22, 22, 22, 30, 32, 33, 35, 36, 36, 35, 34, 33, 32, 31, 30, 30, 29,
+        28, 27, 26, 26, 25, 25, 24, 24, 23, 23, 22, 22, 22, 22, 21, 21, 21, 21,
+        30, 31, 32, 34, 35, 35, 34, 33, 32, 32, 31, 30, 29, 28, 28, 27, 26, 25,
+        25, 24, 24, 23, 23, 23, 22, 22, 22, 21, 21, 21, 21, 21, 29, 30, 31, 33,
+        34, 34, 33, 32, 32, 31, 30, 29, 29, 28, 27, 26, 26, 25, 25, 24, 23, 23,
+        23, 22, 22, 22, 21, 21, 21, 21, 21, 21, 29, 30, 31, 33, 34, 34, 33, 32,
+        32, 31, 30, 29, 29, 28, 27, 26, 26, 25, 25, 24, 23, 23, 23, 22, 22, 22,
+        21, 21, 21, 21, 21, 21, 29, 30, 31, 33, 34, 34, 33, 32, 32, 31, 30, 29,
+        29, 28, 27, 26, 26, 25, 25, 24, 23, 23, 23, 22, 22, 22, 21, 21, 21, 21,
+        21, 21, 29, 30, 31, 33, 34, 34, 33, 32, 32, 31, 30, 29, 29, 28, 27, 26,
+        26, 25, 25, 24, 23, 23, 23, 22, 22, 22, 21, 21, 21, 21, 21, 21 },
+      { /* Intra matrices */
+        /* Size 4 */
+        142, 90, 83, 66, 90, 75, 71, 62, 83, 71, 55, 49, 66, 62, 49, 41,
+        /* Size 8 */
+        132, 163, 91, 85, 80, 72, 64, 57, 163, 113, 89, 98, 95, 87, 77, 68, 91,
+        89, 76, 79, 79, 75, 69, 63, 85, 98, 79, 70, 67, 64, 60, 56, 80, 95, 79,
+        67, 60, 56, 53, 50, 72, 87, 75, 64, 56, 51, 48, 45, 64, 77, 69, 60, 53,
+        48, 44, 42, 57, 68, 63, 56, 50, 45, 42, 39,
+        /* Size 16 */
+        136, 152, 168, 130, 93, 90, 88, 85, 82, 78, 74, 70, 65, 62, 58, 58, 152,
+        147, 142, 117, 92, 93, 94, 92, 90, 86, 81, 77, 72, 68, 64, 64, 168, 142,
+        116, 104, 91, 96, 101, 99, 98, 93, 89, 84, 79, 75, 70, 70, 130, 117,
+        104, 94, 85, 88, 91, 90, 90, 86, 83, 79, 75, 71, 67, 67, 93, 92, 91, 85,
+        78, 80, 81, 81, 81, 79, 77, 74, 71, 68, 65, 65, 90, 93, 96, 88, 80, 78,
+        76, 76, 75, 73, 71, 69, 67, 64, 61, 61, 88, 94, 101, 91, 81, 76, 72, 70,
+        68, 67, 66, 64, 62, 60, 58, 58, 85, 92, 99, 90, 81, 76, 70, 68, 65, 63,
+        62, 60, 58, 56, 55, 55, 82, 90, 98, 90, 81, 75, 68, 65, 61, 60, 58, 56,
+        54, 53, 51, 51, 78, 86, 93, 86, 79, 73, 67, 63, 60, 57, 55, 53, 52, 50,
+        49, 49, 74, 81, 89, 83, 77, 71, 66, 62, 58, 55, 52, 51, 49, 48, 47, 47,
+        70, 77, 84, 79, 74, 69, 64, 60, 56, 53, 51, 49, 47, 46, 45, 45, 65, 72,
+        79, 75, 71, 67, 62, 58, 54, 52, 49, 47, 46, 44, 43, 43, 62, 68, 75, 71,
+        68, 64, 60, 56, 53, 50, 48, 46, 44, 43, 42, 42, 58, 64, 70, 67, 65, 61,
+        58, 55, 51, 49, 47, 45, 43, 42, 40, 40, 58, 64, 70, 67, 65, 61, 58, 55,
+        51, 49, 47, 45, 43, 42, 40, 40,
+        /* Size 32 */
+        137, 146, 154, 162, 170, 151, 132, 113, 94, 93, 92, 90, 89, 88, 86, 85,
+        83, 81, 79, 77, 75, 73, 71, 68, 66, 64, 63, 61, 59, 59, 59, 59, 146,
+        148, 151, 154, 157, 141, 125, 110, 94, 93, 93, 93, 92, 91, 90, 89, 87,
+        85, 83, 81, 79, 76, 74, 72, 70, 68, 66, 64, 62, 62, 62, 62, 154, 151,
+        149, 146, 144, 131, 119, 106, 93, 94, 94, 95, 96, 94, 93, 92, 91, 89,
+        87, 85, 83, 80, 78, 76, 73, 71, 69, 67, 65, 65, 65, 65, 162, 154, 146,
+        139, 131, 122, 112, 103, 93, 94, 96, 97, 99, 98, 97, 96, 95, 93, 91, 89,
+        86, 84, 82, 79, 77, 75, 72, 70, 68, 68, 68, 68, 170, 157, 144, 131, 118,
+        112, 105, 99, 93, 95, 97, 100, 102, 101, 101, 100, 99, 97, 95, 93, 90,
+        88, 85, 83, 80, 78, 76, 73, 71, 71, 71, 71, 151, 141, 131, 122, 112,
+        106, 100, 95, 89, 91, 93, 95, 97, 97, 96, 95, 95, 93, 91, 89, 87, 85,
+        83, 81, 78, 76, 74, 72, 70, 70, 70, 70, 132, 125, 119, 112, 105, 100,
+        96, 91, 86, 87, 89, 91, 92, 92, 91, 91, 91, 89, 88, 86, 84, 82, 80, 78,
+        76, 74, 72, 70, 68, 68, 68, 68, 113, 110, 106, 103, 99, 95, 91, 87, 83,
+        84, 85, 86, 87, 87, 87, 87, 87, 85, 84, 83, 81, 80, 78, 76, 74, 72, 70,
+        69, 67, 67, 67, 67, 94, 94, 93, 93, 93, 89, 86, 83, 79, 80, 81, 81, 82,
+        82, 82, 82, 82, 81, 80, 79, 78, 77, 75, 74, 72, 70, 69, 67, 65, 65, 65,
+        65, 93, 93, 94, 94, 95, 91, 87, 84, 80, 80, 80, 80, 80, 80, 79, 79, 79,
+        78, 77, 76, 75, 74, 73, 71, 70, 68, 67, 65, 64, 64, 64, 64, 92, 93, 94,
+        96, 97, 93, 89, 85, 81, 80, 79, 78, 77, 77, 77, 76, 76, 75, 74, 73, 72,
+        71, 70, 69, 67, 66, 65, 63, 62, 62, 62, 62, 90, 93, 95, 97, 100, 95, 91,
+        86, 81, 80, 78, 77, 75, 74, 74, 73, 73, 72, 71, 70, 69, 68, 67, 66, 65,
+        64, 63, 61, 60, 60, 60, 60, 89, 92, 96, 99, 102, 97, 92, 87, 82, 80, 77,
+        75, 73, 72, 71, 70, 69, 69, 68, 67, 67, 66, 65, 64, 63, 62, 61, 60, 59,
+        59, 59, 59, 88, 91, 94, 98, 101, 97, 92, 87, 82, 80, 77, 74, 72, 71, 70,
+        69, 68, 67, 66, 65, 64, 64, 63, 62, 61, 60, 59, 58, 57, 57, 57, 57, 86,
+        90, 93, 97, 101, 96, 91, 87, 82, 79, 77, 74, 71, 70, 68, 67, 66, 65, 64,
+        63, 62, 62, 61, 60, 59, 58, 57, 56, 55, 55, 55, 55, 85, 89, 92, 96, 100,
+        95, 91, 87, 82, 79, 76, 73, 70, 69, 67, 66, 64, 63, 62, 61, 60, 60, 59,
+        58, 57, 56, 55, 55, 54, 54, 54, 54, 83, 87, 91, 95, 99, 95, 91, 87, 82,
+        79, 76, 73, 69, 68, 66, 64, 62, 61, 60, 59, 58, 58, 57, 56, 55, 54, 54,
+        53, 52, 52, 52, 52, 81, 85, 89, 93, 97, 93, 89, 85, 81, 78, 75, 72, 69,
+        67, 65, 63, 61, 60, 59, 58, 57, 56, 55, 55, 54, 53, 52, 52, 51, 51, 51,
+        51, 79, 83, 87, 91, 95, 91, 88, 84, 80, 77, 74, 71, 68, 66, 64, 62, 60,
+        59, 58, 57, 56, 55, 54, 53, 53, 52, 51, 50, 50, 50, 50, 50, 77, 81, 85,
+        89, 93, 89, 86, 83, 79, 76, 73, 70, 67, 65, 63, 61, 59, 58, 57, 56, 54,
+        54, 53, 52, 51, 50, 50, 49, 48, 48, 48, 48, 75, 79, 83, 86, 90, 87, 84,
+        81, 78, 75, 72, 69, 67, 64, 62, 60, 58, 57, 56, 54, 53, 52, 52, 51, 50,
+        49, 48, 48, 47, 47, 47, 47, 73, 76, 80, 84, 88, 85, 82, 80, 77, 74, 71,
+        68, 66, 64, 62, 60, 58, 56, 55, 54, 52, 52, 51, 50, 49, 48, 48, 47, 46,
+        46, 46, 46, 71, 74, 78, 82, 85, 83, 80, 78, 75, 73, 70, 67, 65, 63, 61,
+        59, 57, 55, 54, 53, 52, 51, 50, 49, 48, 47, 47, 46, 45, 45, 45, 45, 68,
+        72, 76, 79, 83, 81, 78, 76, 74, 71, 69, 66, 64, 62, 60, 58, 56, 55, 53,
+        52, 51, 50, 49, 48, 47, 46, 46, 45, 44, 44, 44, 44, 66, 70, 73, 77, 80,
+        78, 76, 74, 72, 70, 67, 65, 63, 61, 59, 57, 55, 54, 53, 51, 50, 49, 48,
+        47, 46, 45, 45, 44, 44, 44, 44, 44, 64, 68, 71, 75, 78, 76, 74, 72, 70,
+        68, 66, 64, 62, 60, 58, 56, 54, 53, 52, 50, 49, 48, 47, 46, 45, 45, 44,
+        44, 43, 43, 43, 43, 63, 66, 69, 72, 76, 74, 72, 70, 69, 67, 65, 63, 61,
+        59, 57, 55, 54, 52, 51, 50, 48, 48, 47, 46, 45, 44, 44, 43, 42, 42, 42,
+        42, 61, 64, 67, 70, 73, 72, 70, 69, 67, 65, 63, 61, 60, 58, 56, 55, 53,
+        52, 50, 49, 48, 47, 46, 45, 44, 44, 43, 42, 42, 42, 42, 42, 59, 62, 65,
+        68, 71, 70, 68, 67, 65, 64, 62, 60, 59, 57, 55, 54, 52, 51, 50, 48, 47,
+        46, 45, 44, 44, 43, 42, 42, 41, 41, 41, 41, 59, 62, 65, 68, 71, 70, 68,
+        67, 65, 64, 62, 60, 59, 57, 55, 54, 52, 51, 50, 48, 47, 46, 45, 44, 44,
+        43, 42, 42, 41, 41, 41, 41, 59, 62, 65, 68, 71, 70, 68, 67, 65, 64, 62,
+        60, 59, 57, 55, 54, 52, 51, 50, 48, 47, 46, 45, 44, 44, 43, 42, 42, 41,
+        41, 41, 41, 59, 62, 65, 68, 71, 70, 68, 67, 65, 64, 62, 60, 59, 57, 55,
+        54, 52, 51, 50, 48, 47, 46, 45, 44, 44, 43, 42, 42, 41, 41, 41,
+        41 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 59, 38, 26, 59, 41, 31, 25, 38, 31, 24, 21, 26, 25, 21, 19,
+        /* Size 8 */
+        64, 83, 77, 60, 47, 37, 31, 28, 83, 73, 76, 66, 53, 43, 36, 31, 77, 76,
+        56, 49, 43, 37, 32, 29, 60, 66, 49, 40, 35, 32, 29, 27, 47, 53, 43, 35,
+        31, 28, 26, 25, 37, 43, 37, 32, 28, 26, 24, 23, 31, 36, 32, 29, 26, 24,
+        23, 22, 28, 31, 29, 27, 25, 23, 22, 21,
+        /* Size 16 */
+        64, 73, 83, 80, 77, 69, 60, 53, 47, 42, 37, 34, 31, 30, 28, 28, 73, 76,
+        78, 77, 77, 70, 63, 56, 50, 45, 40, 37, 34, 31, 29, 29, 83, 78, 73, 75,
+        76, 71, 66, 59, 53, 48, 43, 39, 36, 33, 31, 31, 80, 77, 75, 70, 66, 62,
+        58, 53, 48, 44, 40, 37, 34, 32, 30, 30, 77, 77, 76, 66, 56, 53, 49, 46,
+        43, 40, 37, 35, 32, 31, 29, 29, 69, 70, 71, 62, 53, 49, 45, 42, 39, 37,
+        35, 33, 31, 29, 28, 28, 60, 63, 66, 58, 49, 45, 40, 38, 35, 34, 32, 31,
+        29, 28, 27, 27, 53, 56, 59, 53, 46, 42, 38, 35, 33, 32, 30, 29, 28, 27,
+        26, 26, 47, 50, 53, 48, 43, 39, 35, 33, 31, 30, 28, 27, 26, 26, 25, 25,
+        42, 45, 48, 44, 40, 37, 34, 32, 30, 28, 27, 26, 25, 25, 24, 24, 37, 40,
+        43, 40, 37, 35, 32, 30, 28, 27, 26, 25, 24, 24, 23, 23, 34, 37, 39, 37,
+        35, 33, 31, 29, 27, 26, 25, 24, 24, 23, 23, 23, 31, 34, 36, 34, 32, 31,
+        29, 28, 26, 25, 24, 24, 23, 23, 22, 22, 30, 31, 33, 32, 31, 29, 28, 27,
+        26, 25, 24, 23, 23, 22, 22, 22, 28, 29, 31, 30, 29, 28, 27, 26, 25, 24,
+        23, 23, 22, 22, 21, 21, 28, 29, 31, 30, 29, 28, 27, 26, 25, 24, 23, 23,
+        22, 22, 21, 21,
+        /* Size 32 */
+        64, 69, 73, 78, 83, 81, 80, 79, 77, 73, 69, 64, 60, 57, 53, 50, 47, 44,
+        42, 40, 37, 36, 34, 33, 31, 30, 30, 29, 28, 28, 28, 28, 69, 72, 75, 78,
+        80, 80, 79, 78, 77, 73, 69, 65, 62, 58, 55, 52, 48, 46, 44, 41, 39, 37,
+        36, 34, 33, 31, 30, 29, 28, 28, 28, 28, 73, 75, 76, 77, 78, 78, 77, 77,
+        77, 73, 70, 66, 63, 60, 56, 53, 50, 47, 45, 43, 40, 38, 37, 35, 34, 32,
+        31, 30, 29, 29, 29, 29, 78, 78, 77, 76, 76, 76, 76, 76, 76, 73, 70, 67,
+        64, 61, 58, 55, 51, 49, 46, 44, 41, 40, 38, 36, 35, 33, 32, 31, 30, 30,
+        30, 30, 83, 80, 78, 76, 73, 74, 75, 75, 76, 74, 71, 68, 66, 63, 59, 56,
+        53, 50, 48, 45, 43, 41, 39, 37, 36, 34, 33, 32, 31, 31, 31, 31, 81, 80,
+        78, 76, 74, 73, 72, 72, 71, 69, 66, 64, 62, 59, 56, 53, 51, 48, 46, 44,
+        41, 40, 38, 36, 35, 34, 33, 31, 30, 30, 30, 30, 80, 79, 77, 76, 75, 72,
+        70, 68, 66, 64, 62, 60, 58, 55, 53, 50, 48, 46, 44, 42, 40, 39, 37, 36,
+        34, 33, 32, 31, 30, 30, 30, 30, 79, 78, 77, 76, 75, 72, 68, 65, 61, 59,
+        57, 55, 53, 51, 49, 47, 46, 44, 42, 40, 39, 37, 36, 35, 33, 32, 31, 30,
+        29, 29, 29, 29, 77, 77, 77, 76, 76, 71, 66, 61, 56, 54, 53, 51, 49, 48,
+        46, 45, 43, 42, 40, 39, 37, 36, 35, 34, 32, 32, 31, 30, 29, 29, 29, 29,
+        73, 73, 73, 73, 74, 69, 64, 59, 54, 53, 51, 49, 47, 45, 44, 43, 41, 40,
+        39, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 28, 28, 28, 69, 69, 70, 70,
+        71, 66, 62, 57, 53, 51, 49, 47, 45, 43, 42, 41, 39, 38, 37, 36, 35, 34,
+        33, 32, 31, 30, 29, 29, 28, 28, 28, 28, 64, 65, 66, 67, 68, 64, 60, 55,
+        51, 49, 47, 45, 42, 41, 40, 39, 37, 36, 35, 34, 33, 32, 32, 31, 30, 29,
+        29, 28, 27, 27, 27, 27, 60, 62, 63, 64, 66, 62, 58, 53, 49, 47, 45, 42,
+        40, 39, 38, 37, 35, 35, 34, 33, 32, 31, 31, 30, 29, 28, 28, 27, 27, 27,
+        27, 27, 57, 58, 60, 61, 63, 59, 55, 51, 48, 45, 43, 41, 39, 38, 37, 35,
+        34, 33, 33, 32, 31, 30, 30, 29, 28, 28, 27, 27, 26, 26, 26, 26, 53, 55,
+        56, 58, 59, 56, 53, 49, 46, 44, 42, 40, 38, 37, 35, 34, 33, 32, 32, 31,
+        30, 30, 29, 28, 28, 27, 27, 26, 26, 26, 26, 26, 50, 52, 53, 55, 56, 53,
+        50, 47, 45, 43, 41, 39, 37, 35, 34, 33, 32, 31, 31, 30, 29, 29, 28, 28,
+        27, 27, 26, 26, 25, 25, 25, 25, 47, 48, 50, 51, 53, 51, 48, 46, 43, 41,
+        39, 37, 35, 34, 33, 32, 31, 30, 30, 29, 28, 28, 27, 27, 26, 26, 26, 25,
+        25, 25, 25, 25, 44, 46, 47, 49, 50, 48, 46, 44, 42, 40, 38, 36, 35, 33,
+        32, 31, 30, 30, 29, 28, 28, 27, 27, 26, 26, 25, 25, 25, 24, 24, 24, 24,
+        42, 44, 45, 46, 48, 46, 44, 42, 40, 39, 37, 35, 34, 33, 32, 31, 30, 29,
+        28, 28, 27, 27, 26, 26, 25, 25, 25, 24, 24, 24, 24, 24, 40, 41, 43, 44,
+        45, 44, 42, 40, 39, 37, 36, 34, 33, 32, 31, 30, 29, 28, 28, 27, 27, 26,
+        26, 25, 25, 25, 24, 24, 24, 24, 24, 24, 37, 39, 40, 41, 43, 41, 40, 39,
+        37, 36, 35, 33, 32, 31, 30, 29, 28, 28, 27, 27, 26, 26, 25, 25, 24, 24,
+        24, 24, 23, 23, 23, 23, 36, 37, 38, 40, 41, 40, 39, 37, 36, 35, 34, 32,
+        31, 30, 30, 29, 28, 27, 27, 26, 26, 25, 25, 24, 24, 24, 24, 23, 23, 23,
+        23, 23, 34, 36, 37, 38, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28,
+        27, 27, 26, 26, 25, 25, 24, 24, 24, 23, 23, 23, 23, 23, 23, 23, 33, 34,
+        35, 36, 37, 36, 36, 35, 34, 33, 32, 31, 30, 29, 28, 28, 27, 26, 26, 25,
+        25, 24, 24, 24, 23, 23, 23, 23, 22, 22, 22, 22, 31, 33, 34, 35, 36, 35,
+        34, 33, 32, 32, 31, 30, 29, 28, 28, 27, 26, 26, 25, 25, 24, 24, 24, 23,
+        23, 23, 23, 22, 22, 22, 22, 22, 30, 31, 32, 33, 34, 34, 33, 32, 32, 31,
+        30, 29, 28, 28, 27, 27, 26, 25, 25, 25, 24, 24, 23, 23, 23, 23, 22, 22,
+        22, 22, 22, 22, 30, 30, 31, 32, 33, 33, 32, 31, 31, 30, 29, 29, 28, 27,
+        27, 26, 26, 25, 25, 24, 24, 24, 23, 23, 23, 22, 22, 22, 22, 22, 22, 22,
+        29, 29, 30, 31, 32, 31, 31, 30, 30, 29, 29, 28, 27, 27, 26, 26, 25, 25,
+        24, 24, 24, 23, 23, 23, 22, 22, 22, 22, 22, 22, 22, 22, 28, 28, 29, 30,
+        31, 30, 30, 29, 29, 28, 28, 27, 27, 26, 26, 25, 25, 24, 24, 24, 23, 23,
+        23, 22, 22, 22, 22, 22, 21, 21, 21, 21, 28, 28, 29, 30, 31, 30, 30, 29,
+        29, 28, 28, 27, 27, 26, 26, 25, 25, 24, 24, 24, 23, 23, 23, 22, 22, 22,
+        22, 22, 21, 21, 21, 21, 28, 28, 29, 30, 31, 30, 30, 29, 29, 28, 28, 27,
+        27, 26, 26, 25, 25, 24, 24, 24, 23, 23, 23, 22, 22, 22, 22, 22, 21, 21,
+        21, 21, 28, 28, 29, 30, 31, 30, 30, 29, 29, 28, 28, 27, 27, 26, 26, 25,
+        25, 24, 24, 24, 23, 23, 23, 22, 22, 22, 22, 22, 21, 21, 21, 21 },
+      { /* Intra matrices */
+        /* Size 4 */
+        155, 142, 90, 61, 142, 97, 72, 56, 90, 72, 54, 46, 61, 56, 46, 41,
+        /* Size 8 */
+        131, 172, 159, 123, 94, 74, 62, 53, 172, 151, 157, 135, 108, 86, 70, 60,
+        159, 157, 115, 99, 86, 74, 64, 56, 123, 135, 99, 80, 70, 63, 56, 51, 94,
+        108, 86, 70, 61, 55, 50, 47, 74, 86, 74, 63, 55, 50, 46, 44, 62, 70, 64,
+        56, 50, 46, 44, 42, 53, 60, 56, 51, 47, 44, 42, 40,
+        /* Size 16 */
+        136, 157, 177, 171, 165, 146, 127, 112, 97, 87, 77, 70, 64, 59, 55, 55,
+        157, 162, 167, 165, 163, 148, 133, 119, 104, 93, 83, 75, 68, 63, 58, 58,
+        177, 167, 156, 159, 162, 151, 140, 126, 111, 100, 89, 81, 73, 67, 62,
+        62, 171, 165, 159, 150, 140, 131, 121, 111, 100, 91, 83, 76, 69, 65, 60,
+        60, 165, 163, 162, 140, 118, 110, 103, 96, 89, 83, 76, 71, 66, 62, 58,
+        58, 146, 148, 151, 131, 110, 102, 93, 87, 81, 76, 71, 66, 62, 59, 55,
+        55, 127, 133, 140, 121, 103, 93, 83, 78, 72, 68, 65, 61, 58, 56, 53, 53,
+        112, 119, 126, 111, 96, 87, 78, 72, 67, 64, 61, 58, 55, 53, 51, 51, 97,
+        104, 111, 100, 89, 81, 72, 67, 63, 60, 57, 54, 52, 50, 49, 49, 87, 93,
+        100, 91, 83, 76, 68, 64, 60, 57, 54, 52, 50, 49, 47, 47, 77, 83, 89, 83,
+        76, 71, 65, 61, 57, 54, 51, 50, 48, 47, 45, 45, 70, 75, 81, 76, 71, 66,
+        61, 58, 54, 52, 50, 48, 46, 45, 44, 44, 64, 68, 73, 69, 66, 62, 58, 55,
+        52, 50, 48, 46, 45, 44, 43, 43, 59, 63, 67, 65, 62, 59, 56, 53, 50, 49,
+        47, 45, 44, 43, 42, 42, 55, 58, 62, 60, 58, 55, 53, 51, 49, 47, 45, 44,
+        43, 42, 41, 41, 55, 58, 62, 60, 58, 55, 53, 51, 49, 47, 45, 44, 43, 42,
+        41, 41,
+        /* Size 32 */
+        138, 148, 159, 170, 180, 177, 174, 171, 167, 158, 148, 139, 129, 122,
+        114, 106, 99, 94, 88, 83, 78, 75, 71, 68, 65, 62, 60, 58, 56, 56, 56,
+        56, 148, 155, 162, 168, 175, 173, 171, 169, 167, 158, 150, 141, 132,
+        125, 117, 110, 102, 97, 92, 86, 81, 78, 74, 71, 67, 65, 62, 60, 58, 58,
+        58, 58, 159, 162, 164, 167, 169, 169, 168, 167, 166, 159, 151, 143, 136,
+        128, 121, 113, 106, 100, 95, 90, 84, 80, 77, 73, 69, 67, 64, 62, 59, 59,
+        59, 59, 170, 168, 167, 165, 164, 164, 165, 165, 166, 159, 152, 146, 139,
+        132, 124, 117, 110, 104, 98, 93, 87, 83, 79, 76, 72, 69, 66, 64, 61, 61,
+        61, 61, 180, 175, 169, 164, 158, 160, 162, 163, 165, 159, 154, 148, 142,
+        135, 128, 120, 113, 107, 102, 96, 90, 86, 82, 78, 74, 71, 68, 66, 63,
+        63, 63, 63, 177, 173, 169, 164, 160, 159, 157, 155, 154, 149, 143, 138,
+        133, 126, 120, 114, 108, 102, 97, 92, 87, 83, 80, 76, 72, 70, 67, 64,
+        62, 62, 62, 62, 174, 171, 168, 165, 162, 157, 152, 147, 143, 138, 133,
+        128, 123, 118, 113, 107, 102, 97, 93, 88, 84, 81, 77, 74, 70, 68, 66,
+        63, 61, 61, 61, 61, 171, 169, 167, 165, 163, 155, 147, 139, 132, 127,
+        123, 118, 114, 109, 105, 101, 96, 92, 89, 85, 81, 78, 75, 72, 69, 66,
+        64, 62, 60, 60, 60, 60, 167, 167, 166, 166, 165, 154, 143, 132, 120,
+        116, 112, 108, 104, 101, 98, 94, 91, 87, 84, 81, 78, 75, 72, 70, 67, 65,
+        63, 61, 59, 59, 59, 59, 158, 158, 159, 159, 159, 149, 138, 127, 116,
+        112, 108, 104, 99, 96, 93, 90, 86, 83, 81, 78, 75, 72, 70, 67, 65, 63,
+        61, 59, 57, 57, 57, 57, 148, 150, 151, 152, 154, 143, 133, 123, 112,
+        108, 103, 99, 94, 91, 88, 85, 82, 79, 77, 74, 72, 70, 67, 65, 63, 61,
+        60, 58, 56, 56, 56, 56, 139, 141, 143, 146, 148, 138, 128, 118, 108,
+        104, 99, 94, 89, 86, 83, 81, 78, 76, 73, 71, 69, 67, 65, 63, 61, 60, 58,
+        57, 55, 55, 55, 55, 129, 132, 136, 139, 142, 133, 123, 114, 104, 99, 94,
+        89, 84, 81, 79, 76, 74, 72, 70, 68, 66, 64, 62, 61, 59, 58, 57, 55, 54,
+        54, 54, 54, 122, 125, 128, 132, 135, 126, 118, 109, 101, 96, 91, 86, 81,
+        79, 76, 74, 71, 69, 67, 66, 64, 62, 61, 59, 58, 56, 55, 54, 53, 53, 53,
+        53, 114, 117, 121, 124, 128, 120, 113, 105, 98, 93, 88, 83, 79, 76, 74,
+        71, 69, 67, 65, 63, 62, 60, 59, 57, 56, 55, 54, 53, 52, 52, 52, 52, 106,
+        110, 113, 117, 120, 114, 107, 101, 94, 90, 85, 81, 76, 74, 71, 69, 66,
+        64, 63, 61, 60, 58, 57, 56, 55, 54, 53, 52, 51, 51, 51, 51, 99, 102,
+        106, 110, 113, 108, 102, 96, 91, 86, 82, 78, 74, 71, 69, 66, 64, 62, 61,
+        59, 57, 56, 55, 54, 53, 52, 51, 50, 49, 49, 49, 49, 94, 97, 100, 104,
+        107, 102, 97, 92, 87, 83, 79, 76, 72, 69, 67, 64, 62, 61, 59, 58, 56,
+        55, 54, 53, 52, 51, 50, 49, 49, 49, 49, 49, 88, 92, 95, 98, 102, 97, 93,
+        89, 84, 81, 77, 73, 70, 67, 65, 63, 61, 59, 58, 56, 55, 54, 53, 52, 51,
+        50, 49, 49, 48, 48, 48, 48, 83, 86, 90, 93, 96, 92, 88, 85, 81, 78, 74,
+        71, 68, 66, 63, 61, 59, 58, 56, 55, 54, 53, 52, 51, 50, 49, 48, 48, 47,
+        47, 47, 47, 78, 81, 84, 87, 90, 87, 84, 81, 78, 75, 72, 69, 66, 64, 62,
+        60, 57, 56, 55, 54, 52, 51, 50, 50, 49, 48, 47, 47, 46, 46, 46, 46, 75,
+        78, 80, 83, 86, 83, 81, 78, 75, 72, 70, 67, 64, 62, 60, 58, 56, 55, 54,
+        53, 51, 51, 50, 49, 48, 47, 47, 46, 46, 46, 46, 46, 71, 74, 77, 79, 82,
+        80, 77, 75, 72, 70, 67, 65, 62, 61, 59, 57, 55, 54, 53, 52, 50, 50, 49,
+        48, 47, 47, 46, 45, 45, 45, 45, 45, 68, 71, 73, 76, 78, 76, 74, 72, 70,
+        67, 65, 63, 61, 59, 57, 56, 54, 53, 52, 51, 50, 49, 48, 47, 47, 46, 45,
+        45, 44, 44, 44, 44, 65, 67, 69, 72, 74, 72, 70, 69, 67, 65, 63, 61, 59,
+        58, 56, 55, 53, 52, 51, 50, 49, 48, 47, 47, 46, 45, 45, 44, 44, 44, 44,
+        44, 62, 65, 67, 69, 71, 70, 68, 66, 65, 63, 61, 60, 58, 56, 55, 54, 52,
+        51, 50, 49, 48, 47, 47, 46, 45, 45, 44, 44, 43, 43, 43, 43, 60, 62, 64,
+        66, 68, 67, 66, 64, 63, 61, 60, 58, 57, 55, 54, 53, 51, 50, 49, 48, 47,
+        47, 46, 45, 45, 44, 44, 43, 43, 43, 43, 43, 58, 60, 62, 64, 66, 64, 63,
+        62, 61, 59, 58, 57, 55, 54, 53, 52, 50, 49, 49, 48, 47, 46, 45, 45, 44,
+        44, 43, 43, 42, 42, 42, 42, 56, 58, 59, 61, 63, 62, 61, 60, 59, 57, 56,
+        55, 54, 53, 52, 51, 49, 49, 48, 47, 46, 46, 45, 44, 44, 43, 43, 42, 42,
+        42, 42, 42, 56, 58, 59, 61, 63, 62, 61, 60, 59, 57, 56, 55, 54, 53, 52,
+        51, 49, 49, 48, 47, 46, 46, 45, 44, 44, 43, 43, 42, 42, 42, 42, 42, 56,
+        58, 59, 61, 63, 62, 61, 60, 59, 57, 56, 55, 54, 53, 52, 51, 49, 49, 48,
+        47, 46, 46, 45, 44, 44, 43, 43, 42, 42, 42, 42, 42, 56, 58, 59, 61, 63,
+        62, 61, 60, 59, 57, 56, 55, 54, 53, 52, 51, 49, 49, 48, 47, 46, 46, 45,
+        44, 44, 43, 43, 42, 42, 42, 42, 42 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 42, 40, 32, 42, 36, 34, 31, 40, 34, 28, 25, 32, 31, 25, 22,
+        /* Size 8 */
+        64, 78, 46, 43, 41, 37, 34, 31, 78, 56, 45, 49, 48, 44, 40, 36, 46, 45,
+        39, 40, 41, 39, 36, 33, 43, 49, 40, 36, 35, 34, 32, 30, 41, 48, 41, 35,
+        32, 30, 29, 28, 37, 44, 39, 34, 30, 28, 27, 26, 34, 40, 36, 32, 29, 27,
+        25, 24, 31, 36, 33, 30, 28, 26, 24, 23,
+        /* Size 16 */
+        64, 71, 78, 62, 46, 45, 43, 42, 41, 39, 37, 36, 34, 32, 31, 31, 71, 69,
+        67, 56, 45, 46, 46, 45, 44, 43, 41, 39, 37, 35, 33, 33, 78, 67, 56, 50,
+        45, 47, 49, 48, 48, 46, 44, 42, 40, 38, 36, 36, 62, 56, 50, 46, 42, 43,
+        45, 44, 44, 43, 41, 40, 38, 36, 35, 35, 46, 45, 45, 42, 39, 40, 40, 40,
+        41, 40, 39, 38, 36, 35, 33, 33, 45, 46, 47, 43, 40, 39, 38, 38, 38, 37,
+        36, 35, 34, 33, 32, 32, 43, 46, 49, 45, 40, 38, 36, 36, 35, 34, 34, 33,
+        32, 31, 30, 30, 42, 45, 48, 44, 40, 38, 36, 35, 34, 33, 32, 31, 31, 30,
+        29, 29, 41, 44, 48, 44, 41, 38, 35, 34, 32, 31, 30, 30, 29, 28, 28, 28,
+        39, 43, 46, 43, 40, 37, 34, 33, 31, 30, 29, 29, 28, 27, 27, 27, 37, 41,
+        44, 41, 39, 36, 34, 32, 30, 29, 28, 27, 27, 26, 26, 26, 36, 39, 42, 40,
+        38, 35, 33, 31, 30, 29, 27, 27, 26, 25, 25, 25, 34, 37, 40, 38, 36, 34,
+        32, 31, 29, 28, 27, 26, 25, 25, 24, 24, 32, 35, 38, 36, 35, 33, 31, 30,
+        28, 27, 26, 25, 25, 24, 23, 23, 31, 33, 36, 35, 33, 32, 30, 29, 28, 27,
+        26, 25, 24, 23, 23, 23, 31, 33, 36, 35, 33, 32, 30, 29, 28, 27, 26, 25,
+        24, 23, 23, 23,
+        /* Size 32 */
+        64, 67, 71, 74, 78, 70, 62, 54, 46, 45, 45, 44, 43, 43, 42, 42, 41, 40,
+        39, 38, 37, 36, 36, 35, 34, 33, 32, 31, 31, 31, 31, 31, 67, 69, 70, 71,
+        72, 66, 59, 52, 45, 45, 45, 45, 45, 44, 44, 43, 43, 42, 41, 40, 39, 38,
+        37, 36, 35, 34, 34, 33, 32, 32, 32, 32, 71, 70, 69, 68, 67, 61, 56, 51,
+        45, 46, 46, 46, 46, 46, 45, 45, 44, 43, 43, 42, 41, 40, 39, 38, 37, 36,
+        35, 34, 33, 33, 33, 33, 74, 71, 68, 65, 61, 57, 53, 49, 45, 46, 46, 47,
+        48, 47, 47, 46, 46, 45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 34,
+        34, 34, 78, 72, 67, 61, 56, 53, 50, 48, 45, 46, 47, 48, 49, 49, 48, 48,
+        48, 47, 46, 45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 36, 36, 36, 70, 66,
+        61, 57, 53, 51, 48, 46, 44, 44, 45, 46, 47, 47, 46, 46, 46, 45, 44, 44,
+        43, 42, 41, 40, 39, 38, 37, 36, 35, 35, 35, 35, 62, 59, 56, 53, 50, 48,
+        46, 44, 42, 43, 43, 44, 45, 45, 44, 44, 44, 43, 43, 42, 41, 41, 40, 39,
+        38, 37, 36, 35, 35, 35, 35, 35, 54, 52, 51, 49, 48, 46, 44, 42, 41, 41,
+        42, 42, 43, 43, 42, 42, 42, 42, 41, 41, 40, 39, 39, 38, 37, 36, 36, 35,
+        34, 34, 34, 34, 46, 45, 45, 45, 45, 44, 42, 41, 39, 40, 40, 40, 40, 40,
+        40, 41, 41, 40, 40, 39, 39, 38, 38, 37, 36, 35, 35, 34, 33, 33, 33, 33,
+        45, 45, 46, 46, 46, 44, 43, 41, 40, 40, 40, 39, 39, 39, 39, 39, 39, 39,
+        38, 38, 38, 37, 36, 36, 35, 35, 34, 33, 33, 33, 33, 33, 45, 45, 46, 46,
+        47, 45, 43, 42, 40, 40, 39, 39, 38, 38, 38, 38, 38, 37, 37, 37, 36, 36,
+        35, 35, 34, 34, 33, 32, 32, 32, 32, 32, 44, 45, 46, 47, 48, 46, 44, 42,
+        40, 39, 39, 38, 37, 37, 37, 37, 36, 36, 36, 35, 35, 35, 34, 34, 33, 33,
+        32, 32, 31, 31, 31, 31, 43, 45, 46, 48, 49, 47, 45, 43, 40, 39, 38, 37,
+        36, 36, 36, 35, 35, 35, 34, 34, 34, 33, 33, 33, 32, 32, 31, 31, 30, 30,
+        30, 30, 43, 44, 46, 47, 49, 47, 45, 43, 40, 39, 38, 37, 36, 36, 35, 35,
+        34, 34, 34, 33, 33, 33, 32, 32, 31, 31, 31, 30, 30, 30, 30, 30, 42, 44,
+        45, 47, 48, 46, 44, 42, 40, 39, 38, 37, 36, 35, 35, 34, 34, 33, 33, 32,
+        32, 32, 31, 31, 31, 30, 30, 29, 29, 29, 29, 29, 42, 43, 45, 46, 48, 46,
+        44, 42, 41, 39, 38, 37, 35, 35, 34, 33, 33, 32, 32, 32, 31, 31, 31, 30,
+        30, 29, 29, 29, 28, 28, 28, 28, 41, 43, 44, 46, 48, 46, 44, 42, 41, 39,
+        38, 36, 35, 34, 34, 33, 32, 32, 31, 31, 30, 30, 30, 29, 29, 29, 28, 28,
+        28, 28, 28, 28, 40, 42, 43, 45, 47, 45, 43, 42, 40, 39, 37, 36, 35, 34,
+        33, 32, 32, 31, 31, 30, 30, 29, 29, 29, 28, 28, 28, 28, 27, 27, 27, 27,
+        39, 41, 43, 44, 46, 44, 43, 41, 40, 38, 37, 36, 34, 34, 33, 32, 31, 31,
+        30, 30, 29, 29, 29, 28, 28, 28, 27, 27, 27, 27, 27, 27, 38, 40, 42, 43,
+        45, 44, 42, 41, 39, 38, 37, 35, 34, 33, 32, 32, 31, 30, 30, 29, 29, 28,
+        28, 28, 27, 27, 27, 26, 26, 26, 26, 26, 37, 39, 41, 42, 44, 43, 41, 40,
+        39, 38, 36, 35, 34, 33, 32, 31, 30, 30, 29, 29, 28, 28, 27, 27, 27, 26,
+        26, 26, 26, 26, 26, 26, 36, 38, 40, 41, 43, 42, 41, 39, 38, 37, 36, 35,
+        33, 33, 32, 31, 30, 29, 29, 28, 28, 27, 27, 27, 26, 26, 26, 25, 25, 25,
+        25, 25, 36, 37, 39, 40, 42, 41, 40, 39, 38, 36, 35, 34, 33, 32, 31, 31,
+        30, 29, 29, 28, 27, 27, 27, 26, 26, 26, 25, 25, 25, 25, 25, 25, 35, 36,
+        38, 39, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 29, 28, 28,
+        27, 27, 26, 26, 26, 25, 25, 25, 24, 24, 24, 24, 34, 35, 37, 38, 40, 39,
+        38, 37, 36, 35, 34, 33, 32, 31, 31, 30, 29, 28, 28, 27, 27, 26, 26, 26,
+        25, 25, 25, 24, 24, 24, 24, 24, 33, 34, 36, 37, 39, 38, 37, 36, 35, 35,
+        34, 33, 32, 31, 30, 29, 29, 28, 28, 27, 26, 26, 26, 25, 25, 25, 24, 24,
+        24, 24, 24, 24, 32, 34, 35, 36, 38, 37, 36, 36, 35, 34, 33, 32, 31, 31,
+        30, 29, 28, 28, 27, 27, 26, 26, 25, 25, 25, 24, 24, 24, 23, 23, 23, 23,
+        31, 33, 34, 35, 37, 36, 35, 35, 34, 33, 32, 32, 31, 30, 29, 29, 28, 28,
+        27, 26, 26, 25, 25, 25, 24, 24, 24, 23, 23, 23, 23, 23, 31, 32, 33, 34,
+        36, 35, 35, 34, 33, 33, 32, 31, 30, 30, 29, 28, 28, 27, 27, 26, 26, 25,
+        25, 24, 24, 24, 23, 23, 23, 23, 23, 23, 31, 32, 33, 34, 36, 35, 35, 34,
+        33, 33, 32, 31, 30, 30, 29, 28, 28, 27, 27, 26, 26, 25, 25, 24, 24, 24,
+        23, 23, 23, 23, 23, 23, 31, 32, 33, 34, 36, 35, 35, 34, 33, 33, 32, 31,
+        30, 30, 29, 28, 28, 27, 27, 26, 26, 25, 25, 24, 24, 24, 23, 23, 23, 23,
+        23, 23, 31, 32, 33, 34, 36, 35, 35, 34, 33, 33, 32, 31, 30, 30, 29, 28,
+        28, 27, 27, 26, 26, 25, 25, 24, 24, 24, 23, 23, 23, 23, 23, 23 },
+      { /* Intra matrices */
+        /* Size 4 */
+        133, 87, 80, 65, 87, 73, 69, 61, 80, 69, 56, 50, 65, 61, 50, 43,
+        /* Size 8 */
+        124, 152, 87, 83, 78, 70, 63, 57, 152, 107, 86, 94, 91, 84, 75, 67, 87,
+        86, 74, 77, 77, 73, 68, 62, 83, 94, 77, 69, 66, 63, 60, 56, 78, 91, 77,
+        66, 60, 56, 54, 51, 70, 84, 73, 63, 56, 52, 49, 47, 63, 75, 68, 60, 54,
+        49, 46, 44, 57, 67, 62, 56, 51, 47, 44, 41,
+        /* Size 16 */
+        127, 141, 155, 122, 89, 87, 84, 82, 80, 76, 72, 68, 65, 61, 58, 58, 141,
+        137, 133, 111, 88, 89, 90, 88, 87, 83, 79, 75, 71, 67, 63, 63, 155, 133,
+        110, 99, 88, 92, 96, 95, 93, 90, 86, 81, 77, 73, 69, 69, 122, 111, 99,
+        90, 82, 85, 87, 87, 86, 83, 80, 77, 73, 70, 66, 66, 89, 88, 88, 82, 76,
+        77, 78, 79, 79, 77, 75, 72, 70, 67, 64, 64, 87, 89, 92, 85, 77, 76, 74,
+        74, 73, 71, 70, 68, 66, 63, 61, 61, 84, 90, 96, 87, 78, 74, 70, 69, 67,
+        66, 65, 63, 62, 60, 58, 58, 82, 88, 95, 87, 79, 74, 69, 67, 64, 63, 61,
+        60, 58, 57, 55, 55, 80, 87, 93, 86, 79, 73, 67, 64, 61, 59, 58, 56, 55,
+        54, 52, 52, 76, 83, 90, 83, 77, 71, 66, 63, 59, 57, 55, 54, 53, 51, 50,
+        50, 72, 79, 86, 80, 75, 70, 65, 61, 58, 55, 53, 52, 50, 49, 48, 48, 68,
+        75, 81, 77, 72, 68, 63, 60, 56, 54, 52, 50, 49, 47, 46, 46, 65, 71, 77,
+        73, 70, 66, 62, 58, 55, 53, 50, 49, 47, 46, 45, 45, 61, 67, 73, 70, 67,
+        63, 60, 57, 54, 51, 49, 47, 46, 45, 43, 43, 58, 63, 69, 66, 64, 61, 58,
+        55, 52, 50, 48, 46, 45, 43, 42, 42, 58, 63, 69, 66, 64, 61, 58, 55, 52,
+        50, 48, 46, 45, 43, 42, 42,
+        /* Size 32 */
+        129, 136, 143, 150, 157, 141, 124, 107, 90, 89, 88, 87, 86, 84, 83, 82,
+        81, 79, 77, 75, 73, 71, 69, 67, 65, 64, 62, 60, 59, 59, 59, 59, 136,
+        138, 141, 143, 146, 132, 118, 104, 90, 90, 89, 89, 88, 87, 86, 85, 84,
+        82, 80, 78, 76, 74, 72, 71, 69, 67, 65, 63, 62, 62, 62, 62, 143, 141,
+        139, 137, 134, 123, 112, 101, 89, 90, 90, 91, 91, 90, 89, 89, 88, 86,
+        84, 82, 80, 78, 76, 74, 72, 70, 68, 66, 64, 64, 64, 64, 150, 143, 137,
+        130, 123, 114, 106, 98, 89, 90, 92, 93, 94, 93, 93, 92, 91, 89, 87, 85,
+        83, 81, 79, 77, 75, 73, 71, 69, 67, 67, 67, 67, 157, 146, 134, 123, 111,
+        106, 100, 94, 89, 91, 93, 95, 97, 97, 96, 95, 95, 93, 91, 89, 87, 85,
+        82, 80, 78, 76, 74, 72, 70, 70, 70, 70, 141, 132, 123, 114, 106, 101,
+        96, 91, 86, 88, 89, 91, 93, 92, 92, 91, 91, 89, 87, 86, 84, 82, 80, 78,
+        76, 74, 72, 70, 68, 68, 68, 68, 124, 118, 112, 106, 100, 96, 91, 87, 83,
+        84, 86, 87, 88, 88, 88, 87, 87, 86, 84, 83, 81, 80, 78, 76, 74, 72, 71,
+        69, 67, 67, 67, 67, 107, 104, 101, 98, 94, 91, 87, 84, 80, 81, 82, 83,
+        84, 84, 84, 83, 83, 82, 81, 80, 79, 77, 76, 74, 72, 71, 69, 68, 66, 66,
+        66, 66, 90, 90, 89, 89, 89, 86, 83, 80, 77, 78, 78, 79, 79, 79, 80, 80,
+        80, 79, 78, 77, 76, 75, 73, 72, 71, 69, 68, 66, 65, 65, 65, 65, 89, 90,
+        90, 90, 91, 88, 84, 81, 78, 78, 77, 77, 77, 77, 77, 77, 77, 76, 75, 74,
+        73, 72, 71, 70, 69, 67, 66, 64, 63, 63, 63, 63, 88, 89, 90, 92, 93, 89,
+        86, 82, 78, 77, 77, 76, 75, 75, 75, 74, 74, 73, 72, 72, 71, 70, 69, 68,
+        66, 65, 64, 63, 62, 62, 62, 62, 87, 89, 91, 93, 95, 91, 87, 83, 79, 77,
+        76, 75, 73, 73, 72, 72, 71, 70, 70, 69, 68, 67, 66, 65, 64, 63, 62, 61,
+        60, 60, 60, 60, 86, 88, 91, 94, 97, 93, 88, 84, 79, 77, 75, 73, 71, 70,
+        70, 69, 68, 68, 67, 66, 66, 65, 64, 63, 62, 61, 60, 59, 59, 59, 59, 59,
+        84, 87, 90, 93, 97, 92, 88, 84, 79, 77, 75, 73, 70, 69, 68, 68, 67, 66,
+        65, 65, 64, 63, 62, 61, 61, 60, 59, 58, 57, 57, 57, 57, 83, 86, 89, 93,
+        96, 92, 88, 84, 80, 77, 75, 72, 70, 68, 67, 66, 65, 64, 64, 63, 62, 61,
+        60, 60, 59, 58, 57, 57, 56, 56, 56, 56, 82, 85, 89, 92, 95, 91, 87, 83,
+        80, 77, 74, 72, 69, 68, 66, 65, 63, 63, 62, 61, 60, 59, 59, 58, 57, 57,
+        56, 55, 54, 54, 54, 54, 81, 84, 88, 91, 95, 91, 87, 83, 80, 77, 74, 71,
+        68, 67, 65, 63, 62, 61, 60, 59, 58, 58, 57, 56, 56, 55, 54, 54, 53, 53,
+        53, 53, 79, 82, 86, 89, 93, 89, 86, 82, 79, 76, 73, 70, 68, 66, 64, 63,
+        61, 60, 59, 58, 57, 57, 56, 55, 54, 54, 53, 52, 52, 52, 52, 52, 77, 80,
+        84, 87, 91, 87, 84, 81, 78, 75, 72, 70, 67, 65, 64, 62, 60, 59, 58, 57,
+        56, 55, 55, 54, 53, 53, 52, 51, 51, 51, 51, 51, 75, 78, 82, 85, 89, 86,
+        83, 80, 77, 74, 72, 69, 66, 65, 63, 61, 59, 58, 57, 56, 55, 54, 53, 53,
+        52, 51, 51, 50, 50, 50, 50, 50, 73, 76, 80, 83, 87, 84, 81, 79, 76, 73,
+        71, 68, 66, 64, 62, 60, 58, 57, 56, 55, 54, 53, 52, 52, 51, 50, 50, 49,
+        48, 48, 48, 48, 71, 74, 78, 81, 85, 82, 80, 77, 75, 72, 70, 67, 65, 63,
+        61, 59, 58, 57, 55, 54, 53, 52, 52, 51, 50, 49, 49, 48, 48, 48, 48, 48,
+        69, 72, 76, 79, 82, 80, 78, 76, 73, 71, 69, 66, 64, 62, 60, 59, 57, 56,
+        55, 53, 52, 52, 51, 50, 49, 49, 48, 47, 47, 47, 47, 47, 67, 71, 74, 77,
+        80, 78, 76, 74, 72, 70, 68, 65, 63, 61, 60, 58, 56, 55, 54, 53, 52, 51,
+        50, 49, 48, 48, 47, 47, 46, 46, 46, 46, 65, 69, 72, 75, 78, 76, 74, 72,
+        71, 69, 66, 64, 62, 61, 59, 57, 56, 54, 53, 52, 51, 50, 49, 48, 48, 47,
+        46, 46, 45, 45, 45, 45, 64, 67, 70, 73, 76, 74, 72, 71, 69, 67, 65, 63,
+        61, 60, 58, 57, 55, 54, 53, 51, 50, 49, 49, 48, 47, 46, 46, 45, 45, 45,
+        45, 45, 62, 65, 68, 71, 74, 72, 71, 69, 68, 66, 64, 62, 60, 59, 57, 56,
+        54, 53, 52, 51, 50, 49, 48, 47, 46, 46, 45, 45, 44, 44, 44, 44, 60, 63,
+        66, 69, 72, 70, 69, 68, 66, 64, 63, 61, 59, 58, 57, 55, 54, 52, 51, 50,
+        49, 48, 47, 47, 46, 45, 45, 44, 43, 43, 43, 43, 59, 62, 64, 67, 70, 68,
+        67, 66, 65, 63, 62, 60, 59, 57, 56, 54, 53, 52, 51, 50, 48, 48, 47, 46,
+        45, 45, 44, 43, 43, 43, 43, 43, 59, 62, 64, 67, 70, 68, 67, 66, 65, 63,
+        62, 60, 59, 57, 56, 54, 53, 52, 51, 50, 48, 48, 47, 46, 45, 45, 44, 43,
+        43, 43, 43, 43, 59, 62, 64, 67, 70, 68, 67, 66, 65, 63, 62, 60, 59, 57,
+        56, 54, 53, 52, 51, 50, 48, 48, 47, 46, 45, 45, 44, 43, 43, 43, 43, 43,
+        59, 62, 64, 67, 70, 68, 67, 66, 65, 63, 62, 60, 59, 57, 56, 54, 53, 52,
+        51, 50, 48, 48, 47, 46, 45, 45, 44, 43, 43, 43, 43, 43 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 59, 40, 29, 59, 42, 33, 27, 40, 33, 26, 23, 29, 27, 23, 21,
+        /* Size 8 */
+        64, 82, 76, 60, 48, 39, 34, 30, 82, 73, 75, 66, 54, 44, 38, 33, 76, 75,
+        57, 50, 44, 39, 35, 31, 60, 66, 50, 42, 37, 34, 31, 29, 48, 54, 44, 37,
+        33, 31, 29, 27, 39, 44, 39, 34, 31, 29, 27, 26, 34, 38, 35, 31, 29, 27,
+        26, 25, 30, 33, 31, 29, 27, 26, 25, 24,
+        /* Size 16 */
+        64, 73, 82, 79, 76, 68, 60, 54, 48, 44, 39, 36, 34, 32, 30, 30, 73, 75,
+        77, 76, 76, 69, 63, 57, 51, 46, 42, 39, 36, 34, 31, 31, 82, 77, 73, 74,
+        75, 71, 66, 60, 54, 49, 44, 41, 38, 35, 33, 33, 79, 76, 74, 70, 66, 62,
+        58, 54, 49, 45, 42, 39, 36, 34, 32, 32, 76, 76, 75, 66, 57, 53, 50, 47,
+        44, 42, 39, 37, 35, 33, 31, 31, 68, 69, 71, 62, 53, 50, 46, 43, 41, 39,
+        37, 35, 33, 32, 30, 30, 60, 63, 66, 58, 50, 46, 42, 40, 37, 36, 34, 33,
+        31, 30, 29, 29, 54, 57, 60, 54, 47, 43, 40, 37, 35, 34, 32, 31, 30, 29,
+        28, 28, 48, 51, 54, 49, 44, 41, 37, 35, 33, 32, 31, 30, 29, 28, 27, 27,
+        44, 46, 49, 45, 42, 39, 36, 34, 32, 31, 30, 29, 28, 27, 27, 27, 39, 42,
+        44, 42, 39, 37, 34, 32, 31, 30, 29, 28, 27, 27, 26, 26, 36, 39, 41, 39,
+        37, 35, 33, 31, 30, 29, 28, 27, 26, 26, 25, 25, 34, 36, 38, 36, 35, 33,
+        31, 30, 29, 28, 27, 26, 26, 25, 25, 25, 32, 34, 35, 34, 33, 32, 30, 29,
+        28, 27, 27, 26, 25, 25, 25, 25, 30, 31, 33, 32, 31, 30, 29, 28, 27, 27,
+        26, 25, 25, 25, 24, 24, 30, 31, 33, 32, 31, 30, 29, 28, 27, 27, 26, 25,
+        25, 25, 24, 24,
+        /* Size 32 */
+        64, 68, 73, 77, 82, 80, 79, 78, 76, 72, 68, 64, 60, 57, 54, 51, 48, 46,
+        44, 41, 39, 38, 36, 35, 34, 33, 32, 31, 30, 30, 30, 30, 68, 71, 74, 77,
+        79, 79, 78, 77, 76, 72, 69, 65, 62, 59, 56, 52, 49, 47, 45, 43, 41, 39,
+        38, 36, 35, 34, 33, 32, 31, 31, 31, 31, 73, 74, 75, 76, 77, 77, 76, 76,
+        76, 73, 69, 66, 63, 60, 57, 54, 51, 49, 46, 44, 42, 40, 39, 37, 36, 35,
+        34, 32, 31, 31, 31, 31, 77, 77, 76, 75, 75, 75, 75, 75, 76, 73, 70, 67,
+        64, 61, 58, 55, 52, 50, 48, 45, 43, 41, 40, 38, 37, 35, 34, 33, 32, 32,
+        32, 32, 82, 79, 77, 75, 73, 73, 74, 75, 75, 73, 71, 68, 66, 63, 60, 57,
+        54, 51, 49, 47, 44, 43, 41, 39, 38, 36, 35, 34, 33, 33, 33, 33, 80, 79,
+        77, 75, 73, 73, 72, 71, 71, 68, 66, 64, 62, 59, 57, 54, 51, 49, 47, 45,
+        43, 41, 40, 38, 37, 36, 35, 34, 32, 32, 32, 32, 79, 78, 76, 75, 74, 72,
+        70, 68, 66, 64, 62, 60, 58, 56, 54, 51, 49, 47, 45, 44, 42, 40, 39, 37,
+        36, 35, 34, 33, 32, 32, 32, 32, 78, 77, 76, 75, 75, 71, 68, 65, 61, 60,
+        58, 56, 54, 52, 50, 49, 47, 45, 44, 42, 40, 39, 38, 37, 35, 34, 33, 33,
+        32, 32, 32, 32, 76, 76, 76, 76, 75, 71, 66, 61, 57, 55, 53, 52, 50, 49,
+        47, 46, 44, 43, 42, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 31, 31, 31,
+        72, 72, 73, 73, 73, 68, 64, 60, 55, 53, 52, 50, 48, 47, 45, 44, 43, 41,
+        40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 31, 31, 31, 31, 68, 69, 69, 70,
+        71, 66, 62, 58, 53, 52, 50, 48, 46, 45, 43, 42, 41, 40, 39, 38, 37, 36,
+        35, 34, 33, 32, 32, 31, 30, 30, 30, 30, 64, 65, 66, 67, 68, 64, 60, 56,
+        52, 50, 48, 46, 44, 43, 41, 40, 39, 38, 37, 36, 35, 35, 34, 33, 32, 32,
+        31, 30, 30, 30, 30, 30, 60, 62, 63, 64, 66, 62, 58, 54, 50, 48, 46, 44,
+        42, 41, 40, 38, 37, 37, 36, 35, 34, 33, 33, 32, 31, 31, 30, 30, 29, 29,
+        29, 29, 57, 59, 60, 61, 63, 59, 56, 52, 49, 47, 45, 43, 41, 40, 38, 37,
+        36, 36, 35, 34, 33, 33, 32, 31, 31, 30, 30, 29, 29, 29, 29, 29, 54, 56,
+        57, 58, 60, 57, 54, 50, 47, 45, 43, 41, 40, 38, 37, 36, 35, 35, 34, 33,
+        32, 32, 31, 31, 30, 30, 29, 29, 28, 28, 28, 28, 51, 52, 54, 55, 57, 54,
+        51, 49, 46, 44, 42, 40, 38, 37, 36, 35, 34, 34, 33, 32, 32, 31, 31, 30,
+        30, 29, 29, 28, 28, 28, 28, 28, 48, 49, 51, 52, 54, 51, 49, 47, 44, 43,
+        41, 39, 37, 36, 35, 34, 33, 33, 32, 31, 31, 30, 30, 29, 29, 28, 28, 28,
+        27, 27, 27, 27, 46, 47, 49, 50, 51, 49, 47, 45, 43, 41, 40, 38, 37, 36,
+        35, 34, 33, 32, 31, 31, 30, 30, 29, 29, 28, 28, 28, 27, 27, 27, 27, 27,
+        44, 45, 46, 48, 49, 47, 45, 44, 42, 40, 39, 37, 36, 35, 34, 33, 32, 31,
+        31, 30, 30, 29, 29, 28, 28, 28, 27, 27, 27, 27, 27, 27, 41, 43, 44, 45,
+        47, 45, 44, 42, 40, 39, 38, 36, 35, 34, 33, 32, 31, 31, 30, 30, 29, 29,
+        28, 28, 28, 27, 27, 27, 26, 26, 26, 26, 39, 41, 42, 43, 44, 43, 42, 40,
+        39, 38, 37, 35, 34, 33, 32, 32, 31, 30, 30, 29, 29, 28, 28, 27, 27, 27,
+        27, 26, 26, 26, 26, 26, 38, 39, 40, 41, 43, 41, 40, 39, 38, 37, 36, 35,
+        33, 33, 32, 31, 30, 30, 29, 29, 28, 28, 27, 27, 27, 27, 26, 26, 26, 26,
+        26, 26, 36, 38, 39, 40, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 31,
+        30, 29, 29, 28, 28, 27, 27, 27, 26, 26, 26, 26, 25, 25, 25, 25, 35, 36,
+        37, 38, 39, 38, 37, 37, 36, 35, 34, 33, 32, 31, 31, 30, 29, 29, 28, 28,
+        27, 27, 27, 26, 26, 26, 26, 25, 25, 25, 25, 25, 34, 35, 36, 37, 38, 37,
+        36, 35, 35, 34, 33, 32, 31, 31, 30, 30, 29, 28, 28, 28, 27, 27, 26, 26,
+        26, 26, 25, 25, 25, 25, 25, 25, 33, 34, 35, 35, 36, 36, 35, 34, 34, 33,
+        32, 32, 31, 30, 30, 29, 28, 28, 28, 27, 27, 27, 26, 26, 26, 25, 25, 25,
+        25, 25, 25, 25, 32, 33, 34, 34, 35, 35, 34, 33, 33, 32, 32, 31, 30, 30,
+        29, 29, 28, 28, 27, 27, 27, 26, 26, 26, 25, 25, 25, 25, 25, 25, 25, 25,
+        31, 32, 32, 33, 34, 34, 33, 33, 32, 31, 31, 30, 30, 29, 29, 28, 28, 27,
+        27, 27, 26, 26, 26, 25, 25, 25, 25, 25, 24, 24, 24, 24, 30, 31, 31, 32,
+        33, 32, 32, 32, 31, 31, 30, 30, 29, 29, 28, 28, 27, 27, 27, 26, 26, 26,
+        25, 25, 25, 25, 25, 24, 24, 24, 24, 24, 30, 31, 31, 32, 33, 32, 32, 32,
+        31, 31, 30, 30, 29, 29, 28, 28, 27, 27, 27, 26, 26, 26, 25, 25, 25, 25,
+        25, 24, 24, 24, 24, 24, 30, 31, 31, 32, 33, 32, 32, 32, 31, 31, 30, 30,
+        29, 29, 28, 28, 27, 27, 27, 26, 26, 26, 25, 25, 25, 25, 25, 24, 24, 24,
+        24, 24, 30, 31, 31, 32, 33, 32, 32, 32, 31, 31, 30, 30, 29, 29, 28, 28,
+        27, 27, 27, 26, 26, 26, 25, 25, 25, 25, 25, 24, 24, 24, 24, 24 },
+      { /* Intra matrices */
+        /* Size 4 */
+        141, 130, 85, 60, 130, 91, 70, 56, 85, 70, 55, 48, 60, 56, 48, 43,
+        /* Size 8 */
+        120, 155, 144, 113, 88, 72, 61, 53, 155, 137, 143, 124, 100, 82, 68, 59,
+        144, 143, 106, 93, 82, 71, 63, 56, 113, 124, 93, 77, 68, 62, 56, 52, 88,
+        100, 82, 68, 60, 55, 51, 48, 72, 82, 71, 62, 55, 51, 48, 46, 61, 68, 63,
+        56, 51, 48, 45, 44, 53, 59, 56, 52, 48, 46, 44, 42,
+        /* Size 16 */
+        124, 142, 160, 154, 149, 133, 117, 104, 91, 82, 74, 68, 63, 59, 55, 55,
+        142, 146, 150, 149, 148, 135, 122, 110, 97, 88, 79, 73, 66, 62, 58, 58,
+        160, 150, 141, 144, 147, 137, 128, 115, 103, 94, 84, 77, 70, 66, 61, 61,
+        154, 149, 144, 136, 128, 120, 112, 103, 94, 86, 79, 73, 67, 63, 59, 59,
+        149, 148, 147, 128, 109, 103, 96, 90, 84, 79, 73, 69, 64, 61, 58, 58,
+        133, 135, 137, 120, 103, 95, 87, 82, 77, 73, 68, 65, 61, 58, 56, 56,
+        117, 122, 128, 112, 96, 87, 79, 74, 70, 67, 63, 61, 58, 56, 53, 53, 104,
+        110, 115, 103, 90, 82, 74, 70, 66, 63, 60, 58, 55, 53, 52, 52, 91, 97,
+        103, 94, 84, 77, 70, 66, 62, 59, 57, 55, 53, 51, 50, 50, 82, 88, 94, 86,
+        79, 73, 67, 63, 59, 57, 54, 53, 51, 50, 48, 48, 74, 79, 84, 79, 73, 68,
+        63, 60, 57, 54, 52, 51, 49, 48, 47, 47, 68, 73, 77, 73, 69, 65, 61, 58,
+        55, 53, 51, 49, 48, 47, 46, 46, 63, 66, 70, 67, 64, 61, 58, 55, 53, 51,
+        49, 48, 47, 46, 45, 45, 59, 62, 66, 63, 61, 58, 56, 53, 51, 50, 48, 47,
+        46, 45, 44, 44, 55, 58, 61, 59, 58, 56, 53, 52, 50, 48, 47, 46, 45, 44,
+        44, 44, 55, 58, 61, 59, 58, 56, 53, 52, 50, 48, 47, 46, 45, 44, 44, 44,
+        /* Size 32 */
+        126, 135, 144, 153, 162, 159, 157, 154, 151, 143, 135, 127, 118, 112,
+        105, 99, 92, 88, 84, 79, 75, 72, 69, 66, 63, 62, 60, 58, 56, 56, 56, 56,
+        135, 141, 146, 152, 157, 156, 154, 152, 151, 143, 136, 129, 121, 115,
+        108, 102, 96, 91, 87, 82, 78, 75, 71, 68, 65, 63, 61, 59, 57, 57, 57,
+        57, 144, 146, 148, 151, 153, 152, 151, 151, 150, 144, 137, 130, 124,
+        118, 111, 105, 99, 94, 89, 85, 80, 77, 74, 71, 67, 65, 63, 61, 59, 59,
+        59, 59, 153, 152, 151, 149, 148, 148, 149, 149, 150, 144, 138, 132, 127,
+        120, 114, 108, 102, 97, 92, 87, 83, 79, 76, 73, 69, 67, 65, 63, 60, 60,
+        60, 60, 162, 157, 153, 148, 143, 145, 146, 148, 149, 144, 139, 134, 130,
+        123, 117, 111, 105, 100, 95, 90, 85, 82, 78, 75, 71, 69, 67, 64, 62, 62,
+        62, 62, 159, 156, 152, 148, 145, 143, 142, 141, 139, 135, 130, 126, 121,
+        116, 111, 105, 100, 96, 91, 87, 83, 79, 76, 73, 70, 68, 65, 63, 61, 61,
+        61, 61, 157, 154, 151, 149, 146, 142, 138, 134, 130, 126, 122, 118, 113,
+        109, 104, 100, 95, 91, 88, 84, 80, 77, 74, 71, 68, 66, 64, 62, 60, 60,
+        60, 60, 154, 152, 151, 149, 148, 141, 134, 127, 120, 117, 113, 109, 105,
+        102, 98, 94, 90, 87, 84, 81, 77, 75, 72, 69, 67, 65, 63, 61, 59, 59, 59,
+        59, 151, 151, 150, 150, 149, 139, 130, 120, 111, 108, 104, 101, 97, 94,
+        91, 89, 86, 83, 80, 77, 75, 72, 70, 68, 65, 64, 62, 60, 58, 58, 58, 58,
+        143, 143, 144, 144, 144, 135, 126, 117, 108, 104, 100, 97, 93, 90, 87,
+        85, 82, 79, 77, 74, 72, 70, 68, 66, 64, 62, 61, 59, 57, 57, 57, 57, 135,
+        136, 137, 138, 139, 130, 122, 113, 104, 100, 96, 93, 89, 86, 84, 81, 78,
+        76, 74, 72, 69, 68, 66, 64, 62, 61, 59, 58, 56, 56, 56, 56, 127, 129,
+        130, 132, 134, 126, 118, 109, 101, 97, 93, 88, 84, 82, 80, 77, 75, 73,
+        71, 69, 67, 65, 64, 62, 60, 59, 58, 57, 55, 55, 55, 55, 118, 121, 124,
+        127, 130, 121, 113, 105, 97, 93, 89, 84, 80, 78, 76, 73, 71, 69, 68, 66,
+        64, 63, 62, 60, 59, 58, 57, 55, 54, 54, 54, 54, 112, 115, 118, 120, 123,
+        116, 109, 102, 94, 90, 86, 82, 78, 76, 73, 71, 69, 67, 66, 64, 63, 61,
+        60, 59, 58, 56, 55, 54, 53, 53, 53, 53, 105, 108, 111, 114, 117, 111,
+        104, 98, 91, 87, 84, 80, 76, 73, 71, 69, 67, 65, 64, 62, 61, 60, 59, 57,
+        56, 55, 54, 53, 52, 52, 52, 52, 99, 102, 105, 108, 111, 105, 100, 94,
+        89, 85, 81, 77, 73, 71, 69, 67, 65, 63, 62, 60, 59, 58, 57, 56, 55, 54,
+        53, 52, 51, 51, 51, 51, 92, 96, 99, 102, 105, 100, 95, 90, 86, 82, 78,
+        75, 71, 69, 67, 65, 63, 61, 60, 59, 57, 56, 55, 55, 54, 53, 52, 51, 51,
+        51, 51, 51, 88, 91, 94, 97, 100, 96, 91, 87, 83, 79, 76, 73, 69, 67, 65,
+        63, 61, 60, 59, 57, 56, 55, 54, 54, 53, 52, 51, 51, 50, 50, 50, 50, 84,
+        87, 89, 92, 95, 91, 88, 84, 80, 77, 74, 71, 68, 66, 64, 62, 60, 59, 58,
+        56, 55, 54, 53, 53, 52, 51, 50, 50, 49, 49, 49, 49, 79, 82, 85, 87, 90,
+        87, 84, 81, 77, 74, 72, 69, 66, 64, 62, 60, 59, 57, 56, 55, 54, 53, 52,
+        52, 51, 50, 50, 49, 48, 48, 48, 48, 75, 78, 80, 83, 85, 83, 80, 77, 75,
+        72, 69, 67, 64, 63, 61, 59, 57, 56, 55, 54, 53, 52, 51, 51, 50, 49, 49,
+        48, 48, 48, 48, 48, 72, 75, 77, 79, 82, 79, 77, 75, 72, 70, 68, 65, 63,
+        61, 60, 58, 56, 55, 54, 53, 52, 51, 51, 50, 49, 49, 48, 48, 47, 47, 47,
+        47, 69, 71, 74, 76, 78, 76, 74, 72, 70, 68, 66, 64, 62, 60, 59, 57, 55,
+        54, 53, 52, 51, 51, 50, 49, 49, 48, 48, 47, 47, 47, 47, 47, 66, 68, 71,
+        73, 75, 73, 71, 69, 68, 66, 64, 62, 60, 59, 57, 56, 55, 54, 53, 52, 51,
+        50, 49, 49, 48, 48, 47, 47, 46, 46, 46, 46, 63, 65, 67, 69, 71, 70, 68,
+        67, 65, 64, 62, 60, 59, 58, 56, 55, 54, 53, 52, 51, 50, 49, 49, 48, 47,
+        47, 47, 46, 46, 46, 46, 46, 62, 63, 65, 67, 69, 68, 66, 65, 64, 62, 61,
+        59, 58, 56, 55, 54, 53, 52, 51, 50, 49, 49, 48, 48, 47, 47, 46, 46, 45,
+        45, 45, 45, 60, 61, 63, 65, 67, 65, 64, 63, 62, 61, 59, 58, 57, 55, 54,
+        53, 52, 51, 50, 50, 49, 48, 48, 47, 47, 46, 46, 45, 45, 45, 45, 45, 58,
+        59, 61, 63, 64, 63, 62, 61, 60, 59, 58, 57, 55, 54, 53, 52, 51, 51, 50,
+        49, 48, 48, 47, 47, 46, 46, 45, 45, 45, 45, 45, 45, 56, 57, 59, 60, 62,
+        61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 51, 50, 49, 48, 48, 47, 47,
+        46, 46, 45, 45, 45, 44, 44, 44, 44, 56, 57, 59, 60, 62, 61, 60, 59, 58,
+        57, 56, 55, 54, 53, 52, 51, 51, 50, 49, 48, 48, 47, 47, 46, 46, 45, 45,
+        45, 44, 44, 44, 44, 56, 57, 59, 60, 62, 61, 60, 59, 58, 57, 56, 55, 54,
+        53, 52, 51, 51, 50, 49, 48, 48, 47, 47, 46, 46, 45, 45, 45, 44, 44, 44,
+        44, 56, 57, 59, 60, 62, 61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 51,
+        50, 49, 48, 48, 47, 47, 46, 46, 45, 45, 45, 44, 44, 44, 44 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 44, 41, 34, 44, 38, 36, 33, 41, 36, 30, 27, 34, 33, 27, 25,
+        /* Size 8 */
+        64, 77, 47, 45, 42, 39, 36, 33, 77, 56, 46, 50, 49, 45, 41, 37, 47, 46,
+        41, 42, 42, 40, 38, 35, 45, 50, 42, 38, 37, 36, 34, 32, 42, 49, 42, 37,
+        34, 32, 31, 30, 39, 45, 40, 36, 32, 30, 29, 28, 36, 41, 38, 34, 31, 29,
+        27, 26, 33, 37, 35, 32, 30, 28, 26, 25,
+        /* Size 16 */
+        64, 71, 77, 62, 47, 46, 45, 43, 42, 41, 39, 37, 36, 34, 33, 33, 71, 69,
+        67, 57, 46, 47, 47, 46, 46, 44, 42, 40, 38, 37, 35, 35, 77, 67, 56, 51,
+        46, 48, 50, 49, 49, 47, 45, 43, 41, 39, 37, 37, 62, 57, 51, 47, 43, 45,
+        46, 46, 45, 44, 43, 41, 40, 38, 36, 36, 47, 46, 46, 43, 41, 41, 42, 42,
+        42, 41, 40, 39, 38, 37, 35, 35, 46, 47, 48, 45, 41, 41, 40, 40, 39, 39,
+        38, 37, 36, 35, 34, 34, 45, 47, 50, 46, 42, 40, 38, 37, 37, 36, 36, 35,
+        34, 33, 32, 32, 43, 46, 49, 46, 42, 40, 37, 36, 35, 35, 34, 33, 33, 32,
+        31, 31, 42, 46, 49, 45, 42, 39, 37, 35, 34, 33, 32, 32, 31, 30, 30, 30,
+        41, 44, 47, 44, 41, 39, 36, 35, 33, 32, 31, 31, 30, 29, 29, 29, 39, 42,
+        45, 43, 40, 38, 36, 34, 32, 31, 30, 30, 29, 28, 28, 28, 37, 40, 43, 41,
+        39, 37, 35, 33, 32, 31, 30, 29, 28, 28, 27, 27, 36, 38, 41, 40, 38, 36,
+        34, 33, 31, 30, 29, 28, 27, 27, 26, 26, 34, 37, 39, 38, 37, 35, 33, 32,
+        30, 29, 28, 28, 27, 26, 26, 26, 33, 35, 37, 36, 35, 34, 32, 31, 30, 29,
+        28, 27, 26, 26, 25, 25, 33, 35, 37, 36, 35, 34, 32, 31, 30, 29, 28, 27,
+        26, 26, 25, 25,
+        /* Size 32 */
+        64, 67, 71, 74, 77, 69, 62, 54, 47, 46, 46, 45, 45, 44, 43, 43, 42, 42,
+        41, 40, 39, 38, 37, 36, 36, 35, 34, 33, 33, 33, 33, 33, 67, 68, 70, 71,
+        72, 66, 59, 53, 47, 46, 46, 46, 46, 45, 45, 44, 44, 43, 42, 41, 40, 40,
+        39, 38, 37, 36, 35, 35, 34, 34, 34, 34, 71, 70, 69, 68, 67, 62, 57, 51,
+        46, 47, 47, 47, 47, 47, 46, 46, 46, 45, 44, 43, 42, 41, 40, 39, 38, 38,
+        37, 36, 35, 35, 35, 35, 74, 71, 68, 65, 61, 58, 54, 50, 46, 47, 47, 48,
+        49, 48, 48, 47, 47, 46, 45, 44, 44, 43, 42, 41, 40, 39, 38, 37, 36, 36,
+        36, 36, 77, 72, 67, 61, 56, 54, 51, 49, 46, 47, 48, 49, 50, 50, 49, 49,
+        49, 48, 47, 46, 45, 44, 43, 42, 41, 40, 39, 38, 37, 37, 37, 37, 69, 66,
+        62, 58, 54, 51, 49, 47, 45, 46, 46, 47, 48, 48, 47, 47, 47, 46, 45, 45,
+        44, 43, 42, 41, 40, 39, 39, 38, 37, 37, 37, 37, 62, 59, 57, 54, 51, 49,
+        47, 45, 43, 44, 45, 45, 46, 46, 46, 45, 45, 45, 44, 43, 43, 42, 41, 40,
+        40, 39, 38, 37, 36, 36, 36, 36, 54, 53, 51, 50, 49, 47, 45, 44, 42, 43,
+        43, 43, 44, 44, 44, 44, 44, 43, 43, 42, 42, 41, 40, 39, 39, 38, 37, 36,
+        36, 36, 36, 36, 47, 47, 46, 46, 46, 45, 43, 42, 41, 41, 41, 42, 42, 42,
+        42, 42, 42, 42, 41, 41, 40, 40, 39, 38, 38, 37, 37, 36, 35, 35, 35, 35,
+        46, 46, 47, 47, 47, 46, 44, 43, 41, 41, 41, 41, 41, 41, 41, 41, 41, 40,
+        40, 40, 39, 39, 38, 37, 37, 36, 36, 35, 34, 34, 34, 34, 46, 46, 47, 47,
+        48, 46, 45, 43, 41, 41, 41, 40, 40, 40, 40, 40, 39, 39, 39, 38, 38, 37,
+        37, 36, 36, 35, 35, 34, 34, 34, 34, 34, 45, 46, 47, 48, 49, 47, 45, 43,
+        42, 41, 40, 40, 39, 39, 39, 38, 38, 38, 37, 37, 37, 36, 36, 36, 35, 35,
+        34, 34, 33, 33, 33, 33, 45, 46, 47, 49, 50, 48, 46, 44, 42, 41, 40, 39,
+        38, 38, 37, 37, 37, 36, 36, 36, 36, 35, 35, 35, 34, 34, 33, 33, 32, 32,
+        32, 32, 44, 45, 47, 48, 50, 48, 46, 44, 42, 41, 40, 39, 38, 37, 37, 36,
+        36, 36, 35, 35, 35, 34, 34, 34, 33, 33, 33, 32, 32, 32, 32, 32, 43, 45,
+        46, 48, 49, 47, 46, 44, 42, 41, 40, 39, 37, 37, 36, 36, 35, 35, 35, 34,
+        34, 34, 33, 33, 33, 32, 32, 32, 31, 31, 31, 31, 43, 44, 46, 47, 49, 47,
+        45, 44, 42, 41, 40, 38, 37, 36, 36, 35, 35, 34, 34, 34, 33, 33, 33, 32,
+        32, 32, 31, 31, 31, 31, 31, 31, 42, 44, 46, 47, 49, 47, 45, 44, 42, 41,
+        39, 38, 37, 36, 35, 35, 34, 34, 33, 33, 32, 32, 32, 31, 31, 31, 30, 30,
+        30, 30, 30, 30, 42, 43, 45, 46, 48, 46, 45, 43, 42, 40, 39, 38, 36, 36,
+        35, 34, 34, 33, 33, 32, 32, 32, 31, 31, 31, 30, 30, 30, 29, 29, 29, 29,
+        41, 42, 44, 45, 47, 45, 44, 43, 41, 40, 39, 37, 36, 35, 35, 34, 33, 33,
+        32, 32, 31, 31, 31, 30, 30, 30, 29, 29, 29, 29, 29, 29, 40, 41, 43, 44,
+        46, 45, 43, 42, 41, 40, 38, 37, 36, 35, 34, 34, 33, 32, 32, 31, 31, 30,
+        30, 30, 29, 29, 29, 29, 28, 28, 28, 28, 39, 40, 42, 44, 45, 44, 43, 42,
+        40, 39, 38, 37, 36, 35, 34, 33, 32, 32, 31, 31, 30, 30, 30, 29, 29, 29,
+        28, 28, 28, 28, 28, 28, 38, 40, 41, 43, 44, 43, 42, 41, 40, 39, 37, 36,
+        35, 34, 34, 33, 32, 32, 31, 30, 30, 30, 29, 29, 29, 28, 28, 28, 28, 28,
+        28, 28, 37, 39, 40, 42, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 33,
+        32, 31, 31, 30, 30, 29, 29, 29, 28, 28, 28, 27, 27, 27, 27, 27, 36, 38,
+        39, 41, 42, 41, 40, 39, 38, 37, 36, 36, 35, 34, 33, 32, 31, 31, 30, 30,
+        29, 29, 29, 28, 28, 28, 27, 27, 27, 27, 27, 27, 36, 37, 38, 40, 41, 40,
+        40, 39, 38, 37, 36, 35, 34, 33, 33, 32, 31, 31, 30, 29, 29, 29, 28, 28,
+        27, 27, 27, 27, 26, 26, 26, 26, 35, 36, 38, 39, 40, 39, 39, 38, 37, 36,
+        35, 35, 34, 33, 32, 32, 31, 30, 30, 29, 29, 28, 28, 28, 27, 27, 27, 26,
+        26, 26, 26, 26, 34, 35, 37, 38, 39, 39, 38, 37, 37, 36, 35, 34, 33, 33,
+        32, 31, 30, 30, 29, 29, 28, 28, 28, 27, 27, 27, 26, 26, 26, 26, 26, 26,
+        33, 35, 36, 37, 38, 38, 37, 36, 36, 35, 34, 34, 33, 32, 32, 31, 30, 30,
+        29, 29, 28, 28, 27, 27, 27, 26, 26, 26, 26, 26, 26, 26, 33, 34, 35, 36,
+        37, 37, 36, 36, 35, 34, 34, 33, 32, 32, 31, 31, 30, 29, 29, 28, 28, 28,
+        27, 27, 26, 26, 26, 26, 25, 25, 25, 25, 33, 34, 35, 36, 37, 37, 36, 36,
+        35, 34, 34, 33, 32, 32, 31, 31, 30, 29, 29, 28, 28, 28, 27, 27, 26, 26,
+        26, 26, 25, 25, 25, 25, 33, 34, 35, 36, 37, 37, 36, 36, 35, 34, 34, 33,
+        32, 32, 31, 31, 30, 29, 29, 28, 28, 28, 27, 27, 26, 26, 26, 26, 25, 25,
+        25, 25, 33, 34, 35, 36, 37, 37, 36, 36, 35, 34, 34, 33, 32, 32, 31, 31,
+        30, 29, 29, 28, 28, 28, 27, 27, 26, 26, 26, 26, 25, 25, 25, 25 },
+      { /* Intra matrices */
+        /* Size 4 */
+        125, 84, 78, 64, 84, 72, 68, 61, 78, 68, 56, 51, 64, 61, 51, 45,
+        /* Size 8 */
+        117, 141, 84, 80, 76, 69, 63, 57, 141, 102, 83, 90, 88, 81, 73, 66, 84,
+        83, 73, 75, 75, 72, 67, 62, 80, 90, 75, 68, 65, 63, 60, 57, 76, 88, 75,
+        65, 60, 57, 54, 52, 69, 81, 72, 63, 57, 53, 50, 48, 63, 73, 67, 60, 54,
+        50, 47, 45, 57, 66, 62, 57, 52, 48, 45, 43,
+        /* Size 16 */
+        119, 132, 144, 115, 86, 84, 82, 79, 77, 74, 71, 67, 64, 61, 58, 58, 132,
+        128, 124, 105, 85, 86, 87, 85, 83, 80, 77, 73, 69, 66, 63, 63, 144, 124,
+        104, 94, 84, 88, 92, 91, 89, 86, 83, 79, 75, 71, 68, 68, 115, 105, 94,
+        87, 79, 82, 84, 83, 83, 80, 78, 75, 72, 69, 65, 65, 86, 85, 84, 79, 74,
+        75, 76, 76, 76, 75, 73, 71, 68, 66, 63, 63, 84, 86, 88, 82, 75, 74, 73,
+        72, 71, 70, 69, 67, 65, 63, 61, 61, 82, 87, 92, 84, 76, 73, 69, 68, 66,
+        65, 64, 63, 61, 60, 58, 58, 79, 85, 91, 83, 76, 72, 68, 66, 64, 62, 61,
+        60, 58, 57, 56, 56, 77, 83, 89, 83, 76, 71, 66, 64, 61, 59, 58, 57, 55,
+        54, 53, 53, 74, 80, 86, 80, 75, 70, 65, 62, 59, 58, 56, 55, 53, 52, 51,
+        51, 71, 77, 83, 78, 73, 69, 64, 61, 58, 56, 54, 53, 51, 50, 49, 49, 67,
+        73, 79, 75, 71, 67, 63, 60, 57, 55, 53, 51, 50, 49, 48, 48, 64, 69, 75,
+        72, 68, 65, 61, 58, 55, 53, 51, 50, 48, 47, 46, 46, 61, 66, 71, 69, 66,
+        63, 60, 57, 54, 52, 50, 49, 47, 46, 45, 45, 58, 63, 68, 65, 63, 61, 58,
+        56, 53, 51, 49, 48, 46, 45, 44, 44, 58, 63, 68, 65, 63, 61, 58, 56, 53,
+        51, 49, 48, 46, 45, 44, 44,
+        /* Size 32 */
+        120, 127, 133, 140, 146, 131, 116, 101, 87, 86, 85, 84, 82, 81, 80, 79,
+        78, 76, 75, 73, 71, 70, 68, 66, 65, 63, 62, 60, 59, 59, 59, 59, 127,
+        129, 131, 134, 136, 123, 111, 99, 86, 86, 86, 85, 85, 84, 83, 82, 81,
+        80, 78, 76, 74, 73, 71, 69, 67, 66, 64, 63, 61, 61, 61, 61, 133, 131,
+        129, 127, 126, 116, 106, 96, 86, 86, 87, 87, 88, 87, 86, 85, 84, 83, 81,
+        79, 77, 76, 74, 72, 70, 69, 67, 65, 64, 64, 64, 64, 140, 134, 127, 121,
+        115, 108, 101, 93, 86, 87, 88, 89, 90, 89, 89, 88, 87, 86, 84, 82, 81,
+        79, 77, 75, 73, 71, 70, 68, 66, 66, 66, 66, 146, 136, 126, 115, 105,
+        100, 95, 90, 85, 87, 89, 91, 93, 92, 92, 91, 90, 89, 87, 85, 84, 82, 80,
+        78, 76, 74, 72, 70, 68, 68, 68, 68, 131, 123, 116, 108, 100, 96, 92, 87,
+        83, 84, 86, 87, 89, 88, 88, 88, 87, 86, 84, 83, 81, 79, 78, 76, 74, 72,
+        71, 69, 67, 67, 67, 67, 116, 111, 106, 101, 95, 92, 88, 84, 80, 81, 83,
+        84, 85, 85, 84, 84, 84, 83, 81, 80, 79, 77, 76, 74, 72, 71, 69, 68, 66,
+        66, 66, 66, 101, 99, 96, 93, 90, 87, 84, 81, 78, 78, 79, 80, 81, 81, 81,
+        81, 81, 80, 79, 78, 76, 75, 74, 72, 71, 69, 68, 67, 65, 65, 65, 65, 87,
+        86, 86, 86, 85, 83, 80, 78, 75, 75, 76, 77, 77, 77, 77, 77, 77, 76, 76,
+        75, 74, 73, 72, 70, 69, 68, 67, 65, 64, 64, 64, 64, 86, 86, 86, 87, 87,
+        84, 81, 78, 75, 75, 75, 75, 75, 75, 75, 75, 75, 74, 73, 73, 72, 71, 70,
+        69, 67, 66, 65, 64, 63, 63, 63, 63, 85, 86, 87, 88, 89, 86, 83, 79, 76,
+        75, 75, 74, 73, 73, 73, 73, 72, 72, 71, 70, 70, 69, 68, 67, 66, 65, 63,
+        62, 61, 61, 61, 61, 84, 85, 87, 89, 91, 87, 84, 80, 77, 75, 74, 73, 72,
+        71, 71, 70, 70, 69, 68, 68, 67, 66, 66, 65, 64, 63, 62, 61, 60, 60, 60,
+        60, 82, 85, 88, 90, 93, 89, 85, 81, 77, 75, 73, 72, 70, 69, 68, 68, 67,
+        67, 66, 65, 65, 64, 63, 63, 62, 61, 60, 59, 59, 59, 59, 59, 81, 84, 87,
+        89, 92, 88, 85, 81, 77, 75, 73, 71, 69, 68, 67, 67, 66, 65, 65, 64, 63,
+        63, 62, 61, 61, 60, 59, 58, 57, 57, 57, 57, 80, 83, 86, 89, 92, 88, 84,
+        81, 77, 75, 73, 71, 68, 67, 66, 65, 64, 64, 63, 62, 62, 61, 60, 60, 59,
+        58, 58, 57, 56, 56, 56, 56, 79, 82, 85, 88, 91, 88, 84, 81, 77, 75, 73,
+        70, 68, 67, 65, 64, 63, 62, 62, 61, 60, 59, 59, 58, 58, 57, 56, 56, 55,
+        55, 55, 55, 78, 81, 84, 87, 90, 87, 84, 81, 77, 75, 72, 70, 67, 66, 64,
+        63, 62, 61, 60, 59, 58, 58, 57, 57, 56, 55, 55, 54, 54, 54, 54, 54, 76,
+        80, 83, 86, 89, 86, 83, 80, 76, 74, 72, 69, 67, 65, 64, 62, 61, 60, 59,
+        58, 57, 57, 56, 56, 55, 54, 54, 53, 53, 53, 53, 53, 75, 78, 81, 84, 87,
+        84, 81, 79, 76, 73, 71, 68, 66, 65, 63, 62, 60, 59, 58, 57, 56, 56, 55,
+        55, 54, 53, 53, 52, 52, 52, 52, 52, 73, 76, 79, 82, 85, 83, 80, 78, 75,
+        73, 70, 68, 65, 64, 62, 61, 59, 58, 57, 56, 55, 55, 54, 54, 53, 52, 52,
+        51, 51, 51, 51, 51, 71, 74, 77, 81, 84, 81, 79, 76, 74, 72, 70, 67, 65,
+        63, 62, 60, 58, 57, 56, 55, 54, 54, 53, 52, 52, 51, 51, 50, 50, 50, 50,
+        50, 70, 73, 76, 79, 82, 79, 77, 75, 73, 71, 69, 66, 64, 63, 61, 59, 58,
+        57, 56, 55, 54, 53, 52, 52, 51, 51, 50, 50, 49, 49, 49, 49, 68, 71, 74,
+        77, 80, 78, 76, 74, 72, 70, 68, 66, 63, 62, 60, 59, 57, 56, 55, 54, 53,
+        52, 52, 51, 50, 50, 49, 49, 48, 48, 48, 48, 66, 69, 72, 75, 78, 76, 74,
+        72, 70, 69, 67, 65, 63, 61, 60, 58, 57, 56, 55, 54, 52, 52, 51, 50, 50,
+        49, 49, 48, 48, 48, 48, 48, 65, 67, 70, 73, 76, 74, 72, 71, 69, 67, 66,
+        64, 62, 61, 59, 58, 56, 55, 54, 53, 52, 51, 50, 50, 49, 48, 48, 47, 47,
+        47, 47, 47, 63, 66, 69, 71, 74, 72, 71, 69, 68, 66, 65, 63, 61, 60, 58,
+        57, 55, 54, 53, 52, 51, 51, 50, 49, 48, 48, 47, 47, 46, 46, 46, 46, 62,
+        64, 67, 70, 72, 71, 69, 68, 67, 65, 63, 62, 60, 59, 58, 56, 55, 54, 53,
+        52, 51, 50, 49, 49, 48, 47, 47, 46, 46, 46, 46, 46, 60, 63, 65, 68, 70,
+        69, 68, 67, 65, 64, 62, 61, 59, 58, 57, 56, 54, 53, 52, 51, 50, 50, 49,
+        48, 47, 47, 46, 46, 45, 45, 45, 45, 59, 61, 64, 66, 68, 67, 66, 65, 64,
+        63, 61, 60, 59, 57, 56, 55, 54, 53, 52, 51, 50, 49, 48, 48, 47, 46, 46,
+        45, 45, 45, 45, 45, 59, 61, 64, 66, 68, 67, 66, 65, 64, 63, 61, 60, 59,
+        57, 56, 55, 54, 53, 52, 51, 50, 49, 48, 48, 47, 46, 46, 45, 45, 45, 45,
+        45, 59, 61, 64, 66, 68, 67, 66, 65, 64, 63, 61, 60, 59, 57, 56, 55, 54,
+        53, 52, 51, 50, 49, 48, 48, 47, 46, 46, 45, 45, 45, 45, 45, 59, 61, 64,
+        66, 68, 67, 66, 65, 64, 63, 61, 60, 59, 57, 56, 55, 54, 53, 52, 51, 50,
+        49, 48, 48, 47, 46, 46, 45, 45, 45, 45, 45 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 60, 41, 31, 60, 44, 35, 29, 41, 35, 29, 26, 31, 29, 26, 24,
+        /* Size 8 */
+        64, 80, 75, 61, 49, 41, 36, 33, 80, 72, 74, 66, 55, 46, 40, 35, 75, 74,
+        57, 51, 46, 41, 37, 34, 61, 66, 51, 43, 39, 36, 34, 32, 49, 55, 46, 39,
+        36, 33, 32, 30, 41, 46, 41, 36, 33, 31, 30, 29, 36, 40, 37, 34, 32, 30,
+        29, 28, 33, 35, 34, 32, 30, 29, 28, 27,
+        /* Size 16 */
+        64, 72, 80, 78, 75, 68, 61, 55, 49, 45, 41, 39, 36, 34, 33, 33, 72, 74,
+        76, 75, 75, 69, 63, 57, 52, 48, 43, 41, 38, 36, 34, 34, 80, 76, 72, 73,
+        74, 70, 66, 60, 55, 50, 46, 43, 40, 37, 35, 35, 78, 75, 73, 69, 66, 62,
+        58, 54, 50, 47, 43, 41, 38, 36, 34, 34, 75, 75, 74, 66, 57, 54, 51, 49,
+        46, 43, 41, 39, 37, 35, 34, 34, 68, 69, 70, 62, 54, 51, 47, 45, 43, 41,
+        39, 37, 35, 34, 33, 33, 61, 63, 66, 58, 51, 47, 43, 41, 39, 38, 36, 35,
+        34, 33, 32, 32, 55, 57, 60, 54, 49, 45, 41, 39, 37, 36, 35, 34, 33, 32,
+        31, 31, 49, 52, 55, 50, 46, 43, 39, 37, 36, 34, 33, 32, 32, 31, 30, 30,
+        45, 48, 50, 47, 43, 41, 38, 36, 34, 33, 32, 31, 31, 30, 30, 30, 41, 43,
+        46, 43, 41, 39, 36, 35, 33, 32, 31, 31, 30, 29, 29, 29, 39, 41, 43, 41,
+        39, 37, 35, 34, 32, 31, 31, 30, 29, 29, 28, 28, 36, 38, 40, 38, 37, 35,
+        34, 33, 32, 31, 30, 29, 29, 28, 28, 28, 34, 36, 37, 36, 35, 34, 33, 32,
+        31, 30, 29, 29, 28, 28, 28, 28, 33, 34, 35, 34, 34, 33, 32, 31, 30, 30,
+        29, 28, 28, 28, 27, 27, 33, 34, 35, 34, 34, 33, 32, 31, 30, 30, 29, 28,
+        28, 28, 27, 27,
+        /* Size 32 */
+        64, 68, 72, 76, 80, 79, 78, 77, 75, 72, 68, 64, 61, 58, 55, 52, 49, 47,
+        45, 43, 41, 40, 39, 37, 36, 35, 34, 33, 33, 33, 33, 33, 68, 71, 73, 76,
+        78, 77, 77, 76, 75, 72, 69, 65, 62, 59, 56, 53, 50, 48, 46, 44, 42, 41,
+        40, 38, 37, 36, 35, 34, 33, 33, 33, 33, 72, 73, 74, 75, 76, 76, 75, 75,
+        75, 72, 69, 66, 63, 60, 57, 55, 52, 50, 48, 46, 43, 42, 41, 39, 38, 37,
+        36, 35, 34, 34, 34, 34, 76, 76, 75, 75, 74, 74, 74, 74, 75, 72, 70, 67,
+        64, 62, 59, 56, 53, 51, 49, 47, 45, 43, 42, 40, 39, 38, 37, 36, 35, 35,
+        35, 35, 80, 78, 76, 74, 72, 72, 73, 74, 74, 72, 70, 68, 66, 63, 60, 57,
+        55, 52, 50, 48, 46, 44, 43, 41, 40, 38, 37, 36, 35, 35, 35, 35, 79, 77,
+        76, 74, 72, 72, 71, 71, 70, 68, 66, 64, 62, 60, 57, 55, 52, 50, 48, 47,
+        45, 43, 42, 40, 39, 38, 37, 36, 35, 35, 35, 35, 78, 77, 75, 74, 73, 71,
+        69, 68, 66, 64, 62, 60, 58, 56, 54, 52, 50, 49, 47, 45, 43, 42, 41, 39,
+        38, 37, 36, 35, 34, 34, 34, 34, 77, 76, 75, 74, 74, 71, 68, 65, 62, 60,
+        58, 56, 55, 53, 51, 50, 48, 47, 45, 44, 42, 41, 40, 39, 38, 37, 36, 35,
+        34, 34, 34, 34, 75, 75, 75, 75, 74, 70, 66, 62, 57, 56, 54, 53, 51, 50,
+        49, 47, 46, 45, 43, 42, 41, 40, 39, 38, 37, 36, 35, 35, 34, 34, 34, 34,
+        72, 72, 72, 72, 72, 68, 64, 60, 56, 54, 53, 51, 49, 48, 47, 46, 44, 43,
+        42, 41, 40, 39, 38, 37, 36, 35, 35, 34, 33, 33, 33, 33, 68, 69, 69, 70,
+        70, 66, 62, 58, 54, 53, 51, 49, 47, 46, 45, 44, 43, 42, 41, 40, 39, 38,
+        37, 36, 35, 35, 34, 33, 33, 33, 33, 33, 64, 65, 66, 67, 68, 64, 60, 56,
+        53, 51, 49, 47, 45, 44, 43, 42, 41, 40, 39, 38, 38, 37, 36, 35, 35, 34,
+        33, 33, 32, 32, 32, 32, 61, 62, 63, 64, 66, 62, 58, 55, 51, 49, 47, 45,
+        43, 42, 41, 40, 39, 39, 38, 37, 36, 36, 35, 35, 34, 33, 33, 32, 32, 32,
+        32, 32, 58, 59, 60, 62, 63, 60, 56, 53, 50, 48, 46, 44, 42, 41, 40, 39,
+        38, 38, 37, 36, 36, 35, 34, 34, 33, 33, 32, 32, 31, 31, 31, 31, 55, 56,
+        57, 59, 60, 57, 54, 51, 49, 47, 45, 43, 41, 40, 39, 38, 37, 37, 36, 35,
+        35, 34, 34, 33, 33, 32, 32, 31, 31, 31, 31, 31, 52, 53, 55, 56, 57, 55,
+        52, 50, 47, 46, 44, 42, 40, 39, 38, 37, 37, 36, 35, 35, 34, 34, 33, 33,
+        32, 32, 31, 31, 31, 31, 31, 31, 49, 50, 52, 53, 55, 52, 50, 48, 46, 44,
+        43, 41, 39, 38, 37, 37, 36, 35, 34, 34, 33, 33, 32, 32, 32, 31, 31, 31,
+        30, 30, 30, 30, 47, 48, 50, 51, 52, 50, 49, 47, 45, 43, 42, 40, 39, 38,
+        37, 36, 35, 34, 34, 33, 33, 32, 32, 32, 31, 31, 31, 30, 30, 30, 30, 30,
+        45, 46, 48, 49, 50, 48, 47, 45, 43, 42, 41, 39, 38, 37, 36, 35, 34, 34,
+        33, 33, 32, 32, 31, 31, 31, 30, 30, 30, 30, 30, 30, 30, 43, 44, 46, 47,
+        48, 47, 45, 44, 42, 41, 40, 38, 37, 36, 35, 35, 34, 33, 33, 32, 32, 31,
+        31, 31, 30, 30, 30, 29, 29, 29, 29, 29, 41, 42, 43, 45, 46, 45, 43, 42,
+        41, 40, 39, 38, 36, 36, 35, 34, 33, 33, 32, 32, 31, 31, 31, 30, 30, 30,
+        29, 29, 29, 29, 29, 29, 40, 41, 42, 43, 44, 43, 42, 41, 40, 39, 38, 37,
+        36, 35, 34, 34, 33, 32, 32, 31, 31, 31, 30, 30, 30, 29, 29, 29, 29, 29,
+        29, 29, 39, 40, 41, 42, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 34, 33,
+        32, 32, 31, 31, 31, 30, 30, 30, 29, 29, 29, 29, 28, 28, 28, 28, 37, 38,
+        39, 40, 41, 40, 39, 39, 38, 37, 36, 35, 35, 34, 33, 33, 32, 32, 31, 31,
+        30, 30, 30, 29, 29, 29, 29, 28, 28, 28, 28, 28, 36, 37, 38, 39, 40, 39,
+        38, 38, 37, 36, 35, 35, 34, 33, 33, 32, 32, 31, 31, 30, 30, 30, 29, 29,
+        29, 29, 28, 28, 28, 28, 28, 28, 35, 36, 37, 38, 38, 38, 37, 37, 36, 35,
+        35, 34, 33, 33, 32, 32, 31, 31, 30, 30, 30, 29, 29, 29, 29, 28, 28, 28,
+        28, 28, 28, 28, 34, 35, 36, 37, 37, 37, 36, 36, 35, 35, 34, 33, 33, 32,
+        32, 31, 31, 31, 30, 30, 29, 29, 29, 29, 28, 28, 28, 28, 28, 28, 28, 28,
+        33, 34, 35, 36, 36, 36, 35, 35, 35, 34, 33, 33, 32, 32, 31, 31, 31, 30,
+        30, 29, 29, 29, 29, 28, 28, 28, 28, 28, 27, 27, 27, 27, 33, 33, 34, 35,
+        35, 35, 34, 34, 34, 33, 33, 32, 32, 31, 31, 31, 30, 30, 30, 29, 29, 29,
+        28, 28, 28, 28, 28, 27, 27, 27, 27, 27, 33, 33, 34, 35, 35, 35, 34, 34,
+        34, 33, 33, 32, 32, 31, 31, 31, 30, 30, 30, 29, 29, 29, 28, 28, 28, 28,
+        28, 27, 27, 27, 27, 27, 33, 33, 34, 35, 35, 35, 34, 34, 34, 33, 33, 32,
+        32, 31, 31, 31, 30, 30, 30, 29, 29, 29, 28, 28, 28, 28, 28, 27, 27, 27,
+        27, 27, 33, 33, 34, 35, 35, 35, 34, 34, 34, 33, 33, 32, 32, 31, 31, 31,
+        30, 30, 30, 29, 29, 29, 28, 28, 28, 28, 28, 27, 27, 27, 27, 27 },
+      { /* Intra matrices */
+        /* Size 4 */
+        129, 119, 81, 60, 119, 86, 68, 56, 81, 68, 55, 49, 60, 56, 49, 45,
+        /* Size 8 */
+        111, 141, 132, 105, 84, 70, 60, 54, 141, 125, 130, 114, 94, 78, 67, 59,
+        132, 130, 99, 88, 78, 69, 62, 56, 105, 114, 88, 74, 66, 61, 56, 53, 84,
+        94, 78, 66, 59, 55, 52, 50, 70, 78, 69, 61, 55, 52, 49, 47, 60, 67, 62,
+        56, 52, 49, 47, 46, 54, 59, 56, 53, 50, 47, 46, 44,
+        /* Size 16 */
+        114, 129, 144, 140, 135, 122, 108, 97, 86, 79, 71, 67, 62, 59, 55, 55,
+        129, 133, 137, 135, 134, 123, 113, 102, 91, 84, 76, 70, 65, 62, 58, 58,
+        144, 137, 129, 131, 134, 125, 117, 107, 96, 88, 80, 74, 68, 64, 60, 60,
+        140, 135, 131, 124, 118, 111, 104, 96, 88, 82, 76, 71, 66, 62, 59, 59,
+        135, 134, 134, 118, 102, 96, 90, 85, 80, 76, 71, 67, 63, 60, 58, 58,
+        122, 123, 125, 111, 96, 89, 83, 79, 74, 71, 67, 64, 61, 58, 56, 56, 108,
+        113, 117, 104, 90, 83, 76, 72, 68, 65, 63, 60, 58, 56, 54, 54, 97, 102,
+        107, 96, 85, 79, 72, 68, 65, 62, 60, 58, 56, 54, 53, 53, 86, 91, 96, 88,
+        80, 74, 68, 65, 61, 59, 57, 55, 54, 52, 51, 51, 79, 84, 88, 82, 76, 71,
+        65, 62, 59, 57, 55, 53, 52, 51, 50, 50, 71, 76, 80, 76, 71, 67, 63, 60,
+        57, 55, 53, 52, 50, 50, 49, 49, 67, 70, 74, 71, 67, 64, 60, 58, 55, 53,
+        52, 51, 49, 49, 48, 48, 62, 65, 68, 66, 63, 61, 58, 56, 54, 52, 50, 49,
+        48, 48, 47, 47, 59, 62, 64, 62, 60, 58, 56, 54, 52, 51, 50, 49, 48, 47,
+        46, 46, 55, 58, 60, 59, 58, 56, 54, 53, 51, 50, 49, 48, 47, 46, 46, 46,
+        55, 58, 60, 59, 58, 56, 54, 53, 51, 50, 49, 48, 47, 46, 46, 46,
+        /* Size 32 */
+        116, 123, 131, 139, 146, 144, 142, 139, 137, 130, 123, 116, 109, 104,
+        98, 93, 87, 84, 80, 76, 72, 70, 68, 65, 63, 61, 59, 58, 56, 56, 56, 56,
+        123, 128, 133, 138, 142, 141, 140, 138, 137, 130, 124, 118, 112, 106,
+        101, 95, 90, 86, 82, 78, 75, 72, 69, 67, 64, 63, 61, 59, 57, 57, 57, 57,
+        131, 133, 135, 137, 138, 138, 137, 137, 136, 131, 125, 120, 114, 109,
+        103, 98, 93, 89, 85, 81, 77, 74, 71, 69, 66, 64, 62, 61, 59, 59, 59, 59,
+        139, 138, 137, 136, 134, 135, 135, 135, 136, 131, 126, 121, 116, 111,
+        106, 100, 95, 91, 87, 83, 79, 76, 73, 71, 68, 66, 64, 62, 60, 60, 60,
+        60, 146, 142, 138, 134, 130, 132, 133, 134, 135, 131, 127, 123, 119,
+        113, 108, 103, 98, 94, 89, 85, 81, 78, 75, 72, 69, 67, 65, 63, 61, 61,
+        61, 61, 144, 141, 138, 135, 132, 131, 129, 128, 127, 123, 120, 116, 112,
+        107, 103, 98, 94, 90, 86, 83, 79, 76, 73, 71, 68, 66, 64, 62, 61, 61,
+        61, 61, 142, 140, 137, 135, 133, 129, 126, 123, 119, 116, 112, 109, 105,
+        101, 97, 93, 90, 86, 83, 80, 77, 74, 72, 69, 67, 65, 63, 62, 60, 60, 60,
+        60, 139, 138, 137, 135, 134, 128, 123, 117, 111, 108, 105, 101, 98, 95,
+        92, 89, 86, 83, 80, 77, 74, 72, 70, 68, 66, 64, 62, 61, 59, 59, 59, 59,
+        137, 137, 136, 136, 135, 127, 119, 111, 103, 100, 97, 94, 91, 89, 86,
+        84, 81, 79, 77, 74, 72, 70, 68, 66, 64, 63, 61, 60, 58, 58, 58, 58, 130,
+        130, 131, 131, 131, 123, 116, 108, 100, 97, 94, 91, 88, 85, 83, 81, 78,
+        76, 74, 72, 70, 68, 66, 65, 63, 62, 60, 59, 57, 57, 57, 57, 123, 124,
+        125, 126, 127, 120, 112, 105, 97, 94, 91, 87, 84, 82, 80, 77, 75, 73,
+        71, 70, 68, 66, 65, 63, 61, 60, 59, 58, 57, 57, 57, 57, 116, 118, 120,
+        121, 123, 116, 109, 101, 94, 91, 87, 84, 80, 78, 76, 74, 72, 71, 69, 67,
+        66, 64, 63, 61, 60, 59, 58, 57, 56, 56, 56, 56, 109, 112, 114, 116, 119,
+        112, 105, 98, 91, 88, 84, 80, 77, 75, 73, 71, 69, 68, 66, 65, 63, 62,
+        61, 60, 59, 58, 57, 56, 55, 55, 55, 55, 104, 106, 109, 111, 113, 107,
+        101, 95, 89, 85, 82, 78, 75, 73, 71, 69, 67, 66, 65, 63, 62, 61, 60, 59,
+        58, 57, 56, 55, 54, 54, 54, 54, 98, 101, 103, 106, 108, 103, 97, 92, 86,
+        83, 80, 76, 73, 71, 69, 67, 65, 64, 63, 62, 60, 59, 58, 57, 56, 56, 55,
+        54, 53, 53, 53, 53, 93, 95, 98, 100, 103, 98, 93, 89, 84, 81, 77, 74,
+        71, 69, 67, 65, 64, 62, 61, 60, 59, 58, 57, 56, 55, 55, 54, 53, 52, 52,
+        52, 52, 87, 90, 93, 95, 98, 94, 90, 86, 81, 78, 75, 72, 69, 67, 65, 64,
+        62, 61, 60, 59, 57, 57, 56, 55, 54, 54, 53, 52, 52, 52, 52, 52, 84, 86,
+        89, 91, 94, 90, 86, 83, 79, 76, 73, 71, 68, 66, 64, 62, 61, 60, 59, 58,
+        56, 56, 55, 54, 53, 53, 52, 52, 51, 51, 51, 51, 80, 82, 85, 87, 89, 86,
+        83, 80, 77, 74, 71, 69, 66, 65, 63, 61, 60, 59, 58, 57, 56, 55, 54, 53,
+        53, 52, 52, 51, 50, 50, 50, 50, 76, 78, 81, 83, 85, 83, 80, 77, 74, 72,
+        70, 67, 65, 63, 62, 60, 59, 58, 57, 56, 55, 54, 53, 53, 52, 51, 51, 50,
+        50, 50, 50, 50, 72, 75, 77, 79, 81, 79, 77, 74, 72, 70, 68, 66, 63, 62,
+        60, 59, 57, 56, 56, 55, 54, 53, 52, 52, 51, 51, 50, 50, 49, 49, 49, 49,
+        70, 72, 74, 76, 78, 76, 74, 72, 70, 68, 66, 64, 62, 61, 59, 58, 57, 56,
+        55, 54, 53, 52, 52, 51, 51, 50, 50, 49, 49, 49, 49, 49, 68, 69, 71, 73,
+        75, 73, 72, 70, 68, 66, 65, 63, 61, 60, 58, 57, 56, 55, 54, 53, 52, 52,
+        51, 51, 50, 50, 49, 49, 48, 48, 48, 48, 65, 67, 69, 71, 72, 71, 69, 68,
+        66, 65, 63, 61, 60, 59, 57, 56, 55, 54, 53, 53, 52, 51, 51, 50, 50, 49,
+        49, 48, 48, 48, 48, 48, 63, 64, 66, 68, 69, 68, 67, 66, 64, 63, 61, 60,
+        59, 58, 56, 55, 54, 53, 53, 52, 51, 51, 50, 50, 49, 49, 48, 48, 47, 47,
+        47, 47, 61, 63, 64, 66, 67, 66, 65, 64, 63, 62, 60, 59, 58, 57, 56, 55,
+        54, 53, 52, 51, 51, 50, 50, 49, 49, 48, 48, 48, 47, 47, 47, 47, 59, 61,
+        62, 64, 65, 64, 63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 52, 51,
+        50, 50, 49, 49, 48, 48, 48, 47, 47, 47, 47, 47, 58, 59, 61, 62, 63, 62,
+        62, 61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 52, 51, 50, 50, 49, 49, 48,
+        48, 48, 47, 47, 47, 47, 47, 47, 56, 57, 59, 60, 61, 61, 60, 59, 58, 57,
+        57, 56, 55, 54, 53, 52, 52, 51, 50, 50, 49, 49, 48, 48, 47, 47, 47, 47,
+        46, 46, 46, 46, 56, 57, 59, 60, 61, 61, 60, 59, 58, 57, 57, 56, 55, 54,
+        53, 52, 52, 51, 50, 50, 49, 49, 48, 48, 47, 47, 47, 47, 46, 46, 46, 46,
+        56, 57, 59, 60, 61, 61, 60, 59, 58, 57, 57, 56, 55, 54, 53, 52, 52, 51,
+        50, 50, 49, 49, 48, 48, 47, 47, 47, 47, 46, 46, 46, 46, 56, 57, 59, 60,
+        61, 61, 60, 59, 58, 57, 57, 56, 55, 54, 53, 52, 52, 51, 50, 50, 49, 49,
+        48, 48, 47, 47, 47, 47, 46, 46, 46, 46 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 45, 42, 36, 45, 40, 38, 35, 42, 38, 32, 30, 36, 35, 30, 27,
+        /* Size 8 */
+        64, 76, 48, 46, 44, 41, 37, 35, 76, 57, 47, 51, 50, 46, 43, 39, 48, 47,
+        42, 43, 43, 42, 40, 37, 46, 51, 43, 40, 39, 38, 36, 35, 44, 50, 43, 39,
+        36, 34, 33, 32, 41, 46, 42, 38, 34, 33, 31, 30, 37, 43, 40, 36, 33, 31,
+        30, 29, 35, 39, 37, 35, 32, 30, 29, 28,
+        /* Size 16 */
+        64, 70, 76, 62, 48, 47, 46, 45, 44, 42, 41, 39, 37, 36, 35, 35, 70, 68,
+        66, 57, 48, 48, 48, 48, 47, 45, 44, 42, 40, 39, 37, 37, 76, 66, 57, 52,
+        47, 49, 51, 50, 50, 48, 46, 45, 43, 41, 39, 39, 62, 57, 52, 48, 45, 46,
+        47, 47, 47, 45, 44, 43, 41, 40, 38, 38, 48, 48, 47, 45, 42, 43, 43, 43,
+        43, 43, 42, 41, 40, 38, 37, 37, 47, 48, 49, 46, 43, 42, 42, 41, 41, 40,
+        40, 39, 38, 37, 36, 36, 46, 48, 51, 47, 43, 42, 40, 39, 39, 38, 38, 37,
+        36, 35, 35, 35, 45, 48, 50, 47, 43, 41, 39, 38, 37, 37, 36, 35, 35, 34,
+        33, 33, 44, 47, 50, 47, 43, 41, 39, 37, 36, 35, 34, 34, 33, 33, 32, 32,
+        42, 45, 48, 45, 43, 40, 38, 37, 35, 34, 34, 33, 32, 32, 31, 31, 41, 44,
+        46, 44, 42, 40, 38, 36, 34, 34, 33, 32, 31, 31, 30, 30, 39, 42, 45, 43,
+        41, 39, 37, 35, 34, 33, 32, 31, 31, 30, 30, 30, 37, 40, 43, 41, 40, 38,
+        36, 35, 33, 32, 31, 31, 30, 29, 29, 29, 36, 39, 41, 40, 38, 37, 35, 34,
+        33, 32, 31, 30, 29, 29, 28, 28, 35, 37, 39, 38, 37, 36, 35, 33, 32, 31,
+        30, 30, 29, 28, 28, 28, 35, 37, 39, 38, 37, 36, 35, 33, 32, 31, 30, 30,
+        29, 28, 28, 28,
+        /* Size 32 */
+        64, 67, 70, 73, 76, 69, 62, 55, 48, 47, 47, 46, 46, 45, 45, 44, 44, 43,
+        42, 41, 41, 40, 39, 38, 37, 37, 36, 35, 35, 35, 35, 35, 67, 68, 69, 70,
+        71, 65, 60, 54, 48, 48, 47, 47, 47, 47, 46, 46, 45, 44, 44, 43, 42, 41,
+        40, 40, 39, 38, 37, 37, 36, 36, 36, 36, 70, 69, 68, 67, 66, 62, 57, 52,
+        48, 48, 48, 48, 48, 48, 48, 47, 47, 46, 45, 44, 44, 43, 42, 41, 40, 39,
+        39, 38, 37, 37, 37, 37, 73, 70, 67, 65, 62, 58, 55, 51, 47, 48, 49, 49,
+        50, 49, 49, 49, 48, 47, 47, 46, 45, 44, 43, 42, 41, 41, 40, 39, 38, 38,
+        38, 38, 76, 71, 66, 62, 57, 54, 52, 50, 47, 48, 49, 50, 51, 51, 50, 50,
+        50, 49, 48, 47, 46, 46, 45, 44, 43, 42, 41, 40, 39, 39, 39, 39, 69, 65,
+        62, 58, 54, 52, 50, 48, 46, 47, 47, 48, 49, 49, 49, 48, 48, 47, 47, 46,
+        45, 44, 44, 43, 42, 41, 40, 39, 39, 39, 39, 39, 62, 60, 57, 55, 52, 50,
+        48, 47, 45, 45, 46, 46, 47, 47, 47, 47, 47, 46, 45, 45, 44, 43, 43, 42,
+        41, 40, 40, 39, 38, 38, 38, 38, 55, 54, 52, 51, 50, 48, 47, 45, 44, 44,
+        44, 45, 45, 45, 45, 45, 45, 45, 44, 44, 43, 42, 42, 41, 40, 40, 39, 38,
+        38, 38, 38, 38, 48, 48, 48, 47, 47, 46, 45, 44, 42, 43, 43, 43, 43, 43,
+        43, 43, 43, 43, 43, 42, 42, 41, 41, 40, 40, 39, 38, 38, 37, 37, 37, 37,
+        47, 48, 48, 48, 48, 47, 45, 44, 43, 43, 43, 42, 42, 42, 42, 42, 42, 42,
+        42, 41, 41, 40, 40, 39, 39, 38, 38, 37, 36, 36, 36, 36, 47, 47, 48, 49,
+        49, 47, 46, 44, 43, 43, 42, 42, 42, 41, 41, 41, 41, 41, 40, 40, 40, 39,
+        39, 38, 38, 37, 37, 36, 36, 36, 36, 36, 46, 47, 48, 49, 50, 48, 46, 45,
+        43, 42, 42, 41, 41, 40, 40, 40, 40, 40, 39, 39, 39, 38, 38, 37, 37, 37,
+        36, 36, 35, 35, 35, 35, 46, 47, 48, 50, 51, 49, 47, 45, 43, 42, 42, 41,
+        40, 40, 39, 39, 39, 38, 38, 38, 38, 37, 37, 37, 36, 36, 35, 35, 35, 35,
+        35, 35, 45, 47, 48, 49, 51, 49, 47, 45, 43, 42, 41, 40, 40, 39, 39, 38,
+        38, 38, 37, 37, 37, 36, 36, 36, 35, 35, 35, 34, 34, 34, 34, 34, 45, 46,
+        48, 49, 50, 49, 47, 45, 43, 42, 41, 40, 39, 39, 38, 38, 37, 37, 37, 36,
+        36, 36, 35, 35, 35, 34, 34, 34, 33, 33, 33, 33, 44, 46, 47, 49, 50, 48,
+        47, 45, 43, 42, 41, 40, 39, 38, 38, 37, 37, 36, 36, 36, 35, 35, 35, 34,
+        34, 34, 33, 33, 33, 33, 33, 33, 44, 45, 47, 48, 50, 48, 47, 45, 43, 42,
+        41, 40, 39, 38, 37, 37, 36, 36, 35, 35, 34, 34, 34, 34, 33, 33, 33, 32,
+        32, 32, 32, 32, 43, 44, 46, 47, 49, 47, 46, 45, 43, 42, 41, 40, 38, 38,
+        37, 36, 36, 35, 35, 34, 34, 34, 33, 33, 33, 33, 32, 32, 32, 32, 32, 32,
+        42, 44, 45, 47, 48, 47, 45, 44, 43, 42, 40, 39, 38, 37, 37, 36, 35, 35,
+        34, 34, 34, 33, 33, 33, 32, 32, 32, 32, 31, 31, 31, 31, 41, 43, 44, 46,
+        47, 46, 45, 44, 42, 41, 40, 39, 38, 37, 36, 36, 35, 34, 34, 33, 33, 33,
+        32, 32, 32, 32, 31, 31, 31, 31, 31, 31, 41, 42, 44, 45, 46, 45, 44, 43,
+        42, 41, 40, 39, 38, 37, 36, 35, 34, 34, 34, 33, 33, 32, 32, 32, 31, 31,
+        31, 31, 30, 30, 30, 30, 40, 41, 43, 44, 46, 44, 43, 42, 41, 40, 39, 38,
+        37, 36, 36, 35, 34, 34, 33, 33, 32, 32, 32, 31, 31, 31, 30, 30, 30, 30,
+        30, 30, 39, 40, 42, 43, 45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 35,
+        34, 33, 33, 32, 32, 32, 31, 31, 31, 30, 30, 30, 30, 30, 30, 30, 38, 40,
+        41, 42, 44, 43, 42, 41, 40, 39, 38, 37, 37, 36, 35, 34, 34, 33, 33, 32,
+        32, 31, 31, 31, 30, 30, 30, 30, 29, 29, 29, 29, 37, 39, 40, 41, 43, 42,
+        41, 40, 40, 39, 38, 37, 36, 35, 35, 34, 33, 33, 32, 32, 31, 31, 31, 30,
+        30, 30, 29, 29, 29, 29, 29, 29, 37, 38, 39, 41, 42, 41, 40, 40, 39, 38,
+        37, 37, 36, 35, 34, 34, 33, 33, 32, 32, 31, 31, 30, 30, 30, 29, 29, 29,
+        29, 29, 29, 29, 36, 37, 39, 40, 41, 40, 40, 39, 38, 38, 37, 36, 35, 35,
+        34, 33, 33, 32, 32, 31, 31, 30, 30, 30, 29, 29, 29, 29, 28, 28, 28, 28,
+        35, 37, 38, 39, 40, 39, 39, 38, 38, 37, 36, 36, 35, 34, 34, 33, 32, 32,
+        32, 31, 31, 30, 30, 30, 29, 29, 29, 28, 28, 28, 28, 28, 35, 36, 37, 38,
+        39, 39, 38, 38, 37, 36, 36, 35, 35, 34, 33, 33, 32, 32, 31, 31, 30, 30,
+        30, 29, 29, 29, 28, 28, 28, 28, 28, 28, 35, 36, 37, 38, 39, 39, 38, 38,
+        37, 36, 36, 35, 35, 34, 33, 33, 32, 32, 31, 31, 30, 30, 30, 29, 29, 29,
+        28, 28, 28, 28, 28, 28, 35, 36, 37, 38, 39, 39, 38, 38, 37, 36, 36, 35,
+        35, 34, 33, 33, 32, 32, 31, 31, 30, 30, 30, 29, 29, 29, 28, 28, 28, 28,
+        28, 28, 35, 36, 37, 38, 39, 39, 38, 38, 37, 36, 36, 35, 35, 34, 33, 33,
+        32, 32, 31, 31, 30, 30, 30, 29, 29, 29, 28, 28, 28, 28, 28, 28 },
+      { /* Intra matrices */
+        /* Size 4 */
+        117, 81, 76, 64, 81, 71, 67, 61, 76, 67, 57, 52, 64, 61, 52, 47,
+        /* Size 8 */
+        110, 131, 81, 77, 74, 68, 62, 57, 131, 97, 80, 86, 84, 78, 72, 65, 81,
+        80, 71, 73, 73, 70, 66, 62, 77, 86, 73, 67, 64, 63, 60, 57, 74, 84, 73,
+        64, 60, 57, 55, 53, 68, 78, 70, 63, 57, 54, 51, 50, 62, 72, 66, 60, 55,
+        51, 49, 47, 57, 65, 62, 57, 53, 50, 47, 45,
+        /* Size 16 */
+        112, 123, 134, 108, 83, 81, 79, 77, 75, 72, 69, 66, 64, 61, 59, 59, 123,
+        120, 116, 99, 82, 83, 83, 82, 81, 78, 75, 71, 68, 65, 63, 63, 134, 116,
+        99, 90, 81, 85, 88, 87, 86, 83, 80, 77, 73, 70, 67, 67, 108, 99, 90, 84,
+        77, 79, 81, 81, 80, 78, 76, 73, 70, 68, 65, 65, 83, 82, 81, 77, 72, 73,
+        74, 74, 74, 73, 72, 70, 67, 65, 63, 63, 81, 83, 85, 79, 73, 72, 71, 71,
+        70, 69, 68, 66, 64, 62, 61, 61, 79, 83, 88, 81, 74, 71, 68, 67, 66, 65,
+        64, 62, 61, 60, 58, 58, 77, 82, 87, 81, 74, 71, 67, 65, 63, 62, 61, 60,
+        59, 57, 56, 56, 75, 81, 86, 80, 74, 70, 66, 63, 61, 59, 58, 57, 56, 55,
+        54, 54, 72, 78, 83, 78, 73, 69, 65, 62, 59, 58, 56, 55, 54, 53, 52, 52,
+        69, 75, 80, 76, 72, 68, 64, 61, 58, 56, 55, 54, 52, 51, 51, 51, 66, 71,
+        77, 73, 70, 66, 62, 60, 57, 55, 54, 52, 51, 50, 49, 49, 64, 68, 73, 70,
+        67, 64, 61, 59, 56, 54, 52, 51, 50, 49, 48, 48, 61, 65, 70, 68, 65, 62,
+        60, 57, 55, 53, 51, 50, 49, 48, 47, 47, 59, 63, 67, 65, 63, 61, 58, 56,
+        54, 52, 51, 49, 48, 47, 46, 46, 59, 63, 67, 65, 63, 61, 58, 56, 54, 52,
+        51, 49, 48, 47, 46, 46,
+        /* Size 32 */
+        113, 119, 124, 130, 135, 122, 109, 96, 83, 82, 82, 81, 80, 79, 78, 77,
+        76, 74, 73, 71, 70, 69, 67, 66, 64, 63, 62, 60, 59, 59, 59, 59, 119,
+        121, 123, 124, 126, 116, 105, 94, 83, 83, 83, 82, 82, 81, 80, 79, 79,
+        77, 76, 74, 73, 71, 70, 68, 67, 65, 64, 63, 61, 61, 61, 61, 124, 123,
+        121, 119, 118, 109, 100, 91, 83, 83, 84, 84, 84, 84, 83, 82, 81, 80, 78,
+        77, 75, 74, 72, 71, 69, 68, 66, 65, 63, 63, 63, 63, 130, 124, 119, 114,
+        109, 102, 96, 89, 83, 84, 84, 85, 86, 86, 85, 85, 84, 83, 81, 80, 78,
+        76, 75, 73, 71, 70, 68, 67, 65, 65, 65, 65, 135, 126, 118, 109, 100, 95,
+        91, 87, 82, 84, 85, 87, 89, 88, 88, 87, 87, 85, 84, 82, 81, 79, 77, 76,
+        74, 72, 71, 69, 67, 67, 67, 67, 122, 116, 109, 102, 95, 92, 88, 84, 80,
+        81, 83, 84, 85, 85, 85, 84, 84, 83, 81, 80, 79, 77, 76, 74, 72, 71, 69,
+        68, 66, 66, 66, 66, 109, 105, 100, 96, 91, 88, 84, 81, 78, 79, 80, 81,
+        82, 82, 81, 81, 81, 80, 79, 78, 77, 75, 74, 72, 71, 70, 68, 67, 65, 65,
+        65, 65, 96, 94, 91, 89, 87, 84, 81, 78, 75, 76, 77, 78, 78, 78, 78, 78,
+        78, 77, 76, 75, 74, 73, 72, 71, 70, 68, 67, 66, 65, 65, 65, 65, 83, 83,
+        83, 83, 82, 80, 78, 75, 73, 74, 74, 75, 75, 75, 75, 75, 75, 74, 74, 73,
+        72, 71, 70, 69, 68, 67, 66, 65, 64, 64, 64, 64, 82, 83, 83, 84, 84, 81,
+        79, 76, 74, 74, 73, 73, 73, 73, 73, 73, 73, 72, 72, 71, 70, 69, 68, 68,
+        67, 66, 64, 63, 62, 62, 62, 62, 82, 83, 84, 84, 85, 83, 80, 77, 74, 73,
+        73, 72, 72, 72, 71, 71, 71, 70, 70, 69, 68, 68, 67, 66, 65, 64, 63, 62,
+        61, 61, 61, 61, 81, 82, 84, 85, 87, 84, 81, 78, 75, 73, 72, 71, 70, 70,
+        69, 69, 69, 68, 67, 67, 66, 66, 65, 64, 63, 63, 62, 61, 60, 60, 60, 60,
+        80, 82, 84, 86, 89, 85, 82, 78, 75, 73, 72, 70, 69, 68, 67, 67, 66, 66,
+        65, 65, 64, 64, 63, 62, 62, 61, 60, 60, 59, 59, 59, 59, 79, 81, 84, 86,
+        88, 85, 82, 78, 75, 73, 72, 70, 68, 67, 67, 66, 65, 65, 64, 63, 63, 62,
+        62, 61, 60, 60, 59, 58, 58, 58, 58, 58, 78, 80, 83, 85, 88, 85, 81, 78,
+        75, 73, 71, 69, 67, 67, 66, 65, 64, 63, 63, 62, 62, 61, 60, 60, 59, 59,
+        58, 57, 57, 57, 57, 57, 77, 79, 82, 85, 87, 84, 81, 78, 75, 73, 71, 69,
+        67, 66, 65, 64, 63, 62, 61, 61, 60, 60, 59, 58, 58, 57, 57, 56, 56, 56,
+        56, 56, 76, 79, 81, 84, 87, 84, 81, 78, 75, 73, 71, 69, 66, 65, 64, 63,
+        61, 61, 60, 59, 59, 58, 58, 57, 57, 56, 56, 55, 54, 54, 54, 54, 74, 77,
+        80, 83, 85, 83, 80, 77, 74, 72, 70, 68, 66, 65, 63, 62, 61, 60, 59, 59,
+        58, 57, 57, 56, 56, 55, 55, 54, 54, 54, 54, 54, 73, 76, 78, 81, 84, 81,
+        79, 76, 74, 72, 70, 67, 65, 64, 63, 61, 60, 59, 59, 58, 57, 56, 56, 55,
+        55, 54, 54, 53, 53, 53, 53, 53, 71, 74, 77, 80, 82, 80, 78, 75, 73, 71,
+        69, 67, 65, 63, 62, 61, 59, 59, 58, 57, 56, 56, 55, 54, 54, 53, 53, 52,
+        52, 52, 52, 52, 70, 73, 75, 78, 81, 79, 77, 74, 72, 70, 68, 66, 64, 63,
+        62, 60, 59, 58, 57, 56, 55, 55, 54, 53, 53, 52, 52, 51, 51, 51, 51, 51,
+        69, 71, 74, 76, 79, 77, 75, 73, 71, 69, 68, 66, 64, 62, 61, 60, 58, 57,
+        56, 56, 55, 54, 53, 53, 52, 52, 51, 51, 50, 50, 50, 50, 67, 70, 72, 75,
+        77, 76, 74, 72, 70, 68, 67, 65, 63, 62, 60, 59, 58, 57, 56, 55, 54, 53,
+        53, 52, 52, 51, 51, 50, 50, 50, 50, 50, 66, 68, 71, 73, 76, 74, 72, 71,
+        69, 68, 66, 64, 62, 61, 60, 58, 57, 56, 55, 54, 53, 53, 52, 52, 51, 51,
+        50, 50, 49, 49, 49, 49, 64, 67, 69, 71, 74, 72, 71, 70, 68, 67, 65, 63,
+        62, 60, 59, 58, 57, 56, 55, 54, 53, 52, 52, 51, 50, 50, 49, 49, 49, 49,
+        49, 49, 63, 65, 68, 70, 72, 71, 70, 68, 67, 66, 64, 63, 61, 60, 59, 57,
+        56, 55, 54, 53, 52, 52, 51, 51, 50, 49, 49, 49, 48, 48, 48, 48, 62, 64,
+        66, 68, 71, 69, 68, 67, 66, 64, 63, 62, 60, 59, 58, 57, 56, 55, 54, 53,
+        52, 51, 51, 50, 49, 49, 49, 48, 48, 48, 48, 48, 60, 63, 65, 67, 69, 68,
+        67, 66, 65, 63, 62, 61, 60, 58, 57, 56, 55, 54, 53, 52, 51, 51, 50, 50,
+        49, 49, 48, 48, 47, 47, 47, 47, 59, 61, 63, 65, 67, 66, 65, 65, 64, 62,
+        61, 60, 59, 58, 57, 56, 54, 54, 53, 52, 51, 50, 50, 49, 49, 48, 48, 47,
+        47, 47, 47, 47, 59, 61, 63, 65, 67, 66, 65, 65, 64, 62, 61, 60, 59, 58,
+        57, 56, 54, 54, 53, 52, 51, 50, 50, 49, 49, 48, 48, 47, 47, 47, 47, 47,
+        59, 61, 63, 65, 67, 66, 65, 65, 64, 62, 61, 60, 59, 58, 57, 56, 54, 54,
+        53, 52, 51, 50, 50, 49, 49, 48, 48, 47, 47, 47, 47, 47, 59, 61, 63, 65,
+        67, 66, 65, 65, 64, 62, 61, 60, 59, 58, 57, 56, 54, 54, 53, 52, 51, 50,
+        50, 49, 49, 48, 48, 47, 47, 47, 47, 47 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 60, 43, 33, 60, 45, 37, 32, 43, 37, 31, 29, 33, 32, 29, 27,
+        /* Size 8 */
+        64, 79, 74, 61, 50, 43, 38, 35, 79, 71, 73, 66, 55, 47, 42, 38, 74, 73,
+        58, 52, 48, 43, 39, 36, 61, 66, 52, 45, 42, 39, 37, 35, 50, 55, 48, 42,
+        38, 36, 34, 33, 43, 47, 43, 39, 36, 34, 33, 32, 38, 42, 39, 37, 34, 33,
+        32, 31, 35, 38, 36, 35, 33, 32, 31, 31,
+        /* Size 16 */
+        64, 71, 79, 77, 74, 68, 61, 56, 50, 47, 43, 41, 38, 37, 35, 35, 71, 73,
+        75, 74, 74, 69, 63, 58, 53, 49, 45, 43, 40, 38, 37, 37, 79, 75, 71, 72,
+        73, 70, 66, 60, 55, 51, 47, 45, 42, 40, 38, 38, 77, 74, 72, 69, 66, 62,
+        59, 55, 51, 48, 45, 43, 40, 39, 37, 37, 74, 74, 73, 66, 58, 55, 52, 50,
+        48, 45, 43, 41, 39, 38, 36, 36, 68, 69, 70, 62, 55, 52, 49, 47, 45, 43,
+        41, 39, 38, 37, 36, 36, 61, 63, 66, 59, 52, 49, 45, 43, 42, 40, 39, 38,
+        37, 36, 35, 35, 56, 58, 60, 55, 50, 47, 43, 42, 40, 39, 37, 36, 35, 35,
+        34, 34, 50, 53, 55, 51, 48, 45, 42, 40, 38, 37, 36, 35, 34, 34, 33, 33,
+        47, 49, 51, 48, 45, 43, 40, 39, 37, 36, 35, 34, 34, 33, 33, 33, 43, 45,
+        47, 45, 43, 41, 39, 37, 36, 35, 34, 33, 33, 32, 32, 32, 41, 43, 45, 43,
+        41, 39, 38, 36, 35, 34, 33, 33, 32, 32, 32, 32, 38, 40, 42, 40, 39, 38,
+        37, 35, 34, 34, 33, 32, 32, 31, 31, 31, 37, 38, 40, 39, 38, 37, 36, 35,
+        34, 33, 32, 32, 31, 31, 31, 31, 35, 37, 38, 37, 36, 36, 35, 34, 33, 33,
+        32, 32, 31, 31, 31, 31, 35, 37, 38, 37, 36, 36, 35, 34, 33, 33, 32, 32,
+        31, 31, 31, 31,
+        /* Size 32 */
+        64, 68, 71, 75, 79, 78, 77, 75, 74, 71, 68, 64, 61, 58, 56, 53, 50, 49,
+        47, 45, 43, 42, 41, 40, 38, 38, 37, 36, 35, 35, 35, 35, 68, 70, 72, 75,
+        77, 76, 76, 75, 74, 71, 68, 65, 62, 59, 57, 54, 52, 50, 48, 46, 44, 43,
+        42, 40, 39, 38, 38, 37, 36, 36, 36, 36, 71, 72, 73, 74, 75, 75, 74, 74,
+        74, 71, 69, 66, 63, 61, 58, 55, 53, 51, 49, 47, 45, 44, 43, 41, 40, 39,
+        38, 37, 37, 37, 37, 37, 75, 75, 74, 74, 73, 73, 73, 74, 74, 71, 69, 67,
+        64, 62, 59, 57, 54, 52, 50, 48, 46, 45, 44, 42, 41, 40, 39, 38, 37, 37,
+        37, 37, 79, 77, 75, 73, 71, 72, 72, 73, 73, 72, 70, 68, 66, 63, 60, 58,
+        55, 53, 51, 49, 47, 46, 45, 43, 42, 41, 40, 39, 38, 38, 38, 38, 78, 76,
+        75, 73, 72, 71, 71, 70, 70, 68, 66, 64, 62, 60, 58, 56, 53, 52, 50, 48,
+        46, 45, 44, 42, 41, 40, 39, 38, 37, 37, 37, 37, 77, 76, 74, 73, 72, 71,
+        69, 67, 66, 64, 62, 61, 59, 57, 55, 53, 51, 50, 48, 47, 45, 44, 43, 42,
+        40, 40, 39, 38, 37, 37, 37, 37, 75, 75, 74, 74, 73, 70, 67, 65, 62, 60,
+        59, 57, 56, 54, 53, 51, 49, 48, 47, 45, 44, 43, 42, 41, 40, 39, 38, 37,
+        37, 37, 37, 37, 74, 74, 74, 74, 73, 70, 66, 62, 58, 56, 55, 54, 52, 51,
+        50, 49, 48, 46, 45, 44, 43, 42, 41, 40, 39, 39, 38, 37, 36, 36, 36, 36,
+        71, 71, 71, 71, 72, 68, 64, 60, 56, 55, 54, 52, 51, 49, 48, 47, 46, 45,
+        44, 43, 42, 41, 40, 39, 39, 38, 37, 37, 36, 36, 36, 36, 68, 68, 69, 69,
+        70, 66, 62, 59, 55, 54, 52, 50, 49, 48, 47, 46, 45, 44, 43, 42, 41, 40,
+        39, 39, 38, 37, 37, 36, 36, 36, 36, 36, 64, 65, 66, 67, 68, 64, 61, 57,
+        54, 52, 50, 49, 47, 46, 45, 44, 43, 42, 41, 41, 40, 39, 39, 38, 37, 37,
+        36, 36, 35, 35, 35, 35, 61, 62, 63, 64, 66, 62, 59, 56, 52, 51, 49, 47,
+        45, 44, 43, 42, 42, 41, 40, 39, 39, 38, 38, 37, 37, 36, 36, 35, 35, 35,
+        35, 35, 58, 59, 61, 62, 63, 60, 57, 54, 51, 49, 48, 46, 44, 43, 42, 42,
+        41, 40, 39, 39, 38, 38, 37, 37, 36, 36, 35, 35, 34, 34, 34, 34, 56, 57,
+        58, 59, 60, 58, 55, 53, 50, 48, 47, 45, 43, 42, 42, 41, 40, 39, 39, 38,
+        37, 37, 36, 36, 35, 35, 35, 34, 34, 34, 34, 34, 53, 54, 55, 57, 58, 56,
+        53, 51, 49, 47, 46, 44, 42, 42, 41, 40, 39, 38, 38, 37, 37, 36, 36, 35,
+        35, 35, 34, 34, 34, 34, 34, 34, 50, 52, 53, 54, 55, 53, 51, 49, 48, 46,
+        45, 43, 42, 41, 40, 39, 38, 38, 37, 36, 36, 36, 35, 35, 34, 34, 34, 33,
+        33, 33, 33, 33, 49, 50, 51, 52, 53, 52, 50, 48, 46, 45, 44, 42, 41, 40,
+        39, 38, 38, 37, 36, 36, 35, 35, 35, 34, 34, 34, 33, 33, 33, 33, 33, 33,
+        47, 48, 49, 50, 51, 50, 48, 47, 45, 44, 43, 41, 40, 39, 39, 38, 37, 36,
+        36, 36, 35, 35, 34, 34, 34, 33, 33, 33, 33, 33, 33, 33, 45, 46, 47, 48,
+        49, 48, 47, 45, 44, 43, 42, 41, 39, 39, 38, 37, 36, 36, 36, 35, 35, 34,
+        34, 34, 33, 33, 33, 33, 32, 32, 32, 32, 43, 44, 45, 46, 47, 46, 45, 44,
+        43, 42, 41, 40, 39, 38, 37, 37, 36, 35, 35, 35, 34, 34, 33, 33, 33, 33,
+        32, 32, 32, 32, 32, 32, 42, 43, 44, 45, 46, 45, 44, 43, 42, 41, 40, 39,
+        38, 38, 37, 36, 36, 35, 35, 34, 34, 33, 33, 33, 33, 32, 32, 32, 32, 32,
+        32, 32, 41, 42, 43, 44, 45, 44, 43, 42, 41, 40, 39, 39, 38, 37, 36, 36,
+        35, 35, 34, 34, 33, 33, 33, 33, 32, 32, 32, 32, 32, 32, 32, 32, 40, 40,
+        41, 42, 43, 42, 42, 41, 40, 39, 39, 38, 37, 37, 36, 35, 35, 34, 34, 34,
+        33, 33, 33, 32, 32, 32, 32, 32, 31, 31, 31, 31, 38, 39, 40, 41, 42, 41,
+        40, 40, 39, 39, 38, 37, 37, 36, 35, 35, 34, 34, 34, 33, 33, 33, 32, 32,
+        32, 32, 31, 31, 31, 31, 31, 31, 38, 38, 39, 40, 41, 40, 40, 39, 39, 38,
+        37, 37, 36, 36, 35, 35, 34, 34, 33, 33, 33, 32, 32, 32, 32, 31, 31, 31,
+        31, 31, 31, 31, 37, 38, 38, 39, 40, 39, 39, 38, 38, 37, 37, 36, 36, 35,
+        35, 34, 34, 33, 33, 33, 32, 32, 32, 32, 31, 31, 31, 31, 31, 31, 31, 31,
+        36, 37, 37, 38, 39, 38, 38, 37, 37, 37, 36, 36, 35, 35, 34, 34, 33, 33,
+        33, 33, 32, 32, 32, 32, 31, 31, 31, 31, 31, 31, 31, 31, 35, 36, 37, 37,
+        38, 37, 37, 37, 36, 36, 36, 35, 35, 34, 34, 34, 33, 33, 33, 32, 32, 32,
+        32, 31, 31, 31, 31, 31, 31, 31, 31, 31, 35, 36, 37, 37, 38, 37, 37, 37,
+        36, 36, 36, 35, 35, 34, 34, 34, 33, 33, 33, 32, 32, 32, 32, 31, 31, 31,
+        31, 31, 31, 31, 31, 31, 35, 36, 37, 37, 38, 37, 37, 37, 36, 36, 36, 35,
+        35, 34, 34, 34, 33, 33, 33, 32, 32, 32, 32, 31, 31, 31, 31, 31, 31, 31,
+        31, 31, 35, 36, 37, 37, 38, 37, 37, 37, 36, 36, 36, 35, 35, 34, 34, 34,
+        33, 33, 33, 32, 32, 32, 32, 31, 31, 31, 31, 31, 31, 31, 31, 31 },
+      { /* Intra matrices */
+        /* Size 4 */
+        118, 110, 77, 59, 110, 82, 67, 57, 77, 67, 56, 51, 59, 57, 51, 47,
+        /* Size 8 */
+        103, 128, 121, 98, 80, 68, 60, 55, 128, 115, 119, 106, 89, 75, 65, 59,
+        121, 119, 93, 83, 75, 68, 61, 56, 98, 106, 83, 71, 65, 61, 57, 54, 80,
+        89, 75, 65, 59, 56, 53, 51, 68, 75, 68, 61, 56, 53, 51, 49, 60, 65, 61,
+        57, 53, 51, 49, 48, 55, 59, 56, 54, 51, 49, 48, 47,
+        /* Size 16 */
+        106, 118, 131, 127, 123, 112, 100, 91, 82, 76, 69, 65, 61, 59, 56, 56,
+        118, 122, 125, 124, 123, 113, 104, 95, 86, 80, 73, 69, 64, 61, 58, 58,
+        131, 125, 118, 120, 122, 115, 108, 99, 91, 84, 77, 72, 67, 64, 60, 60,
+        127, 124, 120, 114, 108, 103, 97, 90, 84, 78, 73, 69, 65, 62, 59, 59,
+        123, 123, 122, 108, 95, 90, 85, 81, 77, 73, 69, 66, 63, 60, 58, 58, 112,
+        113, 115, 103, 90, 85, 79, 76, 72, 69, 66, 63, 60, 58, 56, 56, 100, 104,
+        108, 97, 85, 79, 73, 70, 67, 64, 62, 60, 58, 56, 55, 55, 91, 95, 99, 90,
+        81, 76, 70, 67, 64, 62, 60, 58, 56, 55, 54, 54, 82, 86, 91, 84, 77, 72,
+        67, 64, 61, 59, 57, 56, 54, 53, 52, 52, 76, 80, 84, 78, 73, 69, 64, 62,
+        59, 57, 55, 54, 53, 52, 51, 51, 69, 73, 77, 73, 69, 66, 62, 60, 57, 55,
+        54, 53, 52, 51, 50, 50, 65, 69, 72, 69, 66, 63, 60, 58, 56, 54, 53, 52,
+        51, 50, 49, 49, 61, 64, 67, 65, 63, 60, 58, 56, 54, 53, 52, 51, 50, 49,
+        49, 49, 59, 61, 64, 62, 60, 58, 56, 55, 53, 52, 51, 50, 49, 49, 48, 48,
+        56, 58, 60, 59, 58, 56, 55, 54, 52, 51, 50, 49, 49, 48, 48, 48, 56, 58,
+        60, 59, 58, 56, 55, 54, 52, 51, 50, 49, 49, 48, 48, 48,
+        /* Size 32 */
+        107, 113, 120, 126, 133, 131, 129, 127, 125, 119, 113, 107, 102, 97, 92,
+        88, 83, 80, 77, 73, 70, 68, 66, 64, 62, 61, 59, 58, 57, 57, 57, 57, 113,
+        117, 121, 125, 129, 128, 127, 126, 125, 119, 114, 109, 104, 99, 94, 90,
+        85, 82, 79, 75, 72, 70, 68, 66, 64, 62, 61, 59, 58, 58, 58, 58, 120,
+        121, 123, 125, 126, 126, 125, 125, 124, 119, 115, 110, 105, 101, 96, 92,
+        87, 84, 81, 77, 74, 72, 69, 67, 65, 63, 62, 60, 59, 59, 59, 59, 126,
+        125, 125, 124, 123, 123, 123, 124, 124, 120, 116, 112, 107, 103, 98, 94,
+        90, 86, 83, 79, 76, 73, 71, 69, 66, 65, 63, 61, 60, 60, 60, 60, 133,
+        129, 126, 123, 119, 120, 121, 122, 123, 120, 116, 113, 109, 105, 101,
+        96, 92, 88, 85, 81, 78, 75, 73, 70, 68, 66, 64, 63, 61, 61, 61, 61, 131,
+        128, 126, 123, 120, 119, 118, 118, 117, 113, 110, 107, 104, 100, 96, 92,
+        88, 85, 82, 79, 76, 74, 71, 69, 67, 65, 64, 62, 60, 60, 60, 60, 129,
+        127, 125, 123, 121, 118, 116, 113, 110, 107, 104, 101, 98, 95, 91, 88,
+        85, 82, 79, 77, 74, 72, 70, 68, 66, 64, 63, 61, 60, 60, 60, 60, 127,
+        126, 125, 124, 122, 118, 113, 108, 103, 100, 98, 95, 92, 89, 87, 84, 81,
+        79, 77, 74, 72, 70, 68, 66, 65, 63, 62, 60, 59, 59, 59, 59, 125, 125,
+        124, 124, 123, 117, 110, 103, 96, 94, 91, 89, 86, 84, 82, 80, 78, 76,
+        74, 72, 70, 68, 67, 65, 63, 62, 61, 60, 58, 58, 58, 58, 119, 119, 119,
+        120, 120, 113, 107, 100, 94, 91, 88, 86, 83, 81, 79, 77, 75, 74, 72, 70,
+        68, 67, 65, 64, 62, 61, 60, 59, 58, 58, 58, 58, 113, 114, 115, 116, 116,
+        110, 104, 98, 91, 88, 86, 83, 80, 78, 76, 75, 73, 71, 70, 68, 66, 65,
+        64, 62, 61, 60, 59, 58, 57, 57, 57, 57, 107, 109, 110, 112, 113, 107,
+        101, 95, 89, 86, 83, 80, 77, 75, 74, 72, 70, 69, 67, 66, 65, 63, 62, 61,
+        60, 59, 58, 57, 56, 56, 56, 56, 102, 104, 105, 107, 109, 104, 98, 92,
+        86, 83, 80, 77, 74, 72, 71, 69, 67, 66, 65, 64, 63, 62, 61, 60, 59, 58,
+        57, 56, 55, 55, 55, 55, 97, 99, 101, 103, 105, 100, 95, 89, 84, 81, 78,
+        75, 72, 71, 69, 68, 66, 65, 64, 63, 61, 61, 60, 59, 58, 57, 56, 56, 55,
+        55, 55, 55, 92, 94, 96, 98, 101, 96, 91, 87, 82, 79, 76, 74, 71, 69, 68,
+        66, 64, 63, 62, 61, 60, 59, 59, 58, 57, 56, 56, 55, 54, 54, 54, 54, 88,
+        90, 92, 94, 96, 92, 88, 84, 80, 77, 75, 72, 69, 68, 66, 64, 63, 62, 61,
+        60, 59, 58, 57, 57, 56, 55, 55, 54, 53, 53, 53, 53, 83, 85, 87, 90, 92,
+        88, 85, 81, 78, 75, 73, 70, 67, 66, 64, 63, 61, 60, 60, 59, 58, 57, 56,
+        56, 55, 54, 54, 53, 53, 53, 53, 53, 80, 82, 84, 86, 88, 85, 82, 79, 76,
+        74, 71, 69, 66, 65, 63, 62, 60, 60, 59, 58, 57, 56, 56, 55, 54, 54, 53,
+        53, 52, 52, 52, 52, 77, 79, 81, 83, 85, 82, 79, 77, 74, 72, 70, 67, 65,
+        64, 62, 61, 60, 59, 58, 57, 56, 55, 55, 54, 54, 53, 53, 52, 52, 52, 52,
+        52, 73, 75, 77, 79, 81, 79, 77, 74, 72, 70, 68, 66, 64, 63, 61, 60, 59,
+        58, 57, 56, 55, 55, 54, 54, 53, 53, 52, 52, 51, 51, 51, 51, 70, 72, 74,
+        76, 78, 76, 74, 72, 70, 68, 66, 65, 63, 61, 60, 59, 58, 57, 56, 55, 54,
+        54, 53, 53, 52, 52, 52, 51, 51, 51, 51, 51, 68, 70, 72, 73, 75, 74, 72,
+        70, 68, 67, 65, 63, 62, 61, 59, 58, 57, 56, 55, 55, 54, 53, 53, 52, 52,
+        52, 51, 51, 50, 50, 50, 50, 66, 68, 69, 71, 73, 71, 70, 68, 67, 65, 64,
+        62, 61, 60, 59, 57, 56, 56, 55, 54, 53, 53, 52, 52, 51, 51, 51, 50, 50,
+        50, 50, 50, 64, 66, 67, 69, 70, 69, 68, 66, 65, 64, 62, 61, 60, 59, 58,
+        57, 56, 55, 54, 54, 53, 52, 52, 51, 51, 51, 50, 50, 50, 50, 50, 50, 62,
+        64, 65, 66, 68, 67, 66, 65, 63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 54,
+        53, 52, 52, 51, 51, 51, 50, 50, 50, 49, 49, 49, 49, 61, 62, 63, 65, 66,
+        65, 64, 63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 54, 53, 53, 52, 52, 51,
+        51, 50, 50, 50, 49, 49, 49, 49, 49, 59, 61, 62, 63, 64, 64, 63, 62, 61,
+        60, 59, 58, 57, 56, 56, 55, 54, 53, 53, 52, 52, 51, 51, 50, 50, 50, 49,
+        49, 49, 49, 49, 49, 58, 59, 60, 61, 63, 62, 61, 60, 60, 59, 58, 57, 56,
+        56, 55, 54, 53, 53, 52, 52, 51, 51, 50, 50, 50, 49, 49, 49, 49, 49, 49,
+        49, 57, 58, 59, 60, 61, 60, 60, 59, 58, 58, 57, 56, 55, 55, 54, 53, 53,
+        52, 52, 51, 51, 50, 50, 50, 49, 49, 49, 49, 48, 48, 48, 48, 57, 58, 59,
+        60, 61, 60, 60, 59, 58, 58, 57, 56, 55, 55, 54, 53, 53, 52, 52, 51, 51,
+        50, 50, 50, 49, 49, 49, 49, 48, 48, 48, 48, 57, 58, 59, 60, 61, 60, 60,
+        59, 58, 58, 57, 56, 55, 55, 54, 53, 53, 52, 52, 51, 51, 50, 50, 50, 49,
+        49, 49, 49, 48, 48, 48, 48, 57, 58, 59, 60, 61, 60, 60, 59, 58, 58, 57,
+        56, 55, 55, 54, 53, 53, 52, 52, 51, 51, 50, 50, 50, 49, 49, 49, 49, 48,
+        48, 48, 48 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 46, 44, 38, 46, 41, 40, 37, 44, 40, 35, 32, 38, 37, 32, 30,
+        /* Size 8 */
+        64, 75, 49, 47, 45, 42, 40, 37, 75, 57, 49, 52, 51, 48, 44, 41, 49, 49,
+        44, 45, 45, 44, 42, 39, 47, 52, 45, 42, 41, 40, 38, 37, 45, 51, 45, 41,
+        38, 37, 36, 35, 42, 48, 44, 40, 37, 35, 34, 33, 40, 44, 42, 38, 36, 34,
+        33, 32, 37, 41, 39, 37, 35, 33, 32, 31,
+        /* Size 16 */
+        64, 70, 75, 62, 49, 48, 47, 46, 45, 44, 42, 41, 40, 38, 37, 37, 70, 68,
+        66, 58, 49, 49, 50, 49, 48, 47, 45, 44, 42, 40, 39, 39, 75, 66, 57, 53,
+        49, 50, 52, 51, 51, 49, 48, 46, 44, 43, 41, 41, 62, 58, 53, 50, 46, 47,
+        48, 48, 48, 47, 46, 44, 43, 42, 40, 40, 49, 49, 49, 46, 44, 44, 45, 45,
+        45, 44, 44, 43, 42, 40, 39, 39, 48, 49, 50, 47, 44, 44, 43, 43, 43, 42,
+        42, 41, 40, 39, 38, 38, 47, 50, 52, 48, 45, 43, 42, 41, 41, 40, 40, 39,
+        38, 38, 37, 37, 46, 49, 51, 48, 45, 43, 41, 40, 39, 39, 38, 38, 37, 36,
+        36, 36, 45, 48, 51, 48, 45, 43, 41, 39, 38, 37, 37, 36, 36, 35, 35, 35,
+        44, 47, 49, 47, 44, 42, 40, 39, 37, 37, 36, 35, 35, 34, 34, 34, 42, 45,
+        48, 46, 44, 42, 40, 38, 37, 36, 35, 34, 34, 33, 33, 33, 41, 44, 46, 44,
+        43, 41, 39, 38, 36, 35, 34, 34, 33, 33, 32, 32, 40, 42, 44, 43, 42, 40,
+        38, 37, 36, 35, 34, 33, 33, 32, 32, 32, 38, 40, 43, 42, 40, 39, 38, 36,
+        35, 34, 33, 33, 32, 32, 31, 31, 37, 39, 41, 40, 39, 38, 37, 36, 35, 34,
+        33, 32, 32, 31, 31, 31, 37, 39, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32,
+        32, 31, 31, 31,
+        /* Size 32 */
+        64, 67, 70, 72, 75, 69, 62, 56, 49, 49, 48, 48, 47, 47, 46, 46, 45, 45,
+        44, 43, 42, 42, 41, 40, 40, 39, 38, 38, 37, 37, 37, 37, 67, 68, 69, 70,
+        71, 65, 60, 54, 49, 49, 49, 49, 48, 48, 48, 47, 47, 46, 45, 45, 44, 43,
+        42, 41, 41, 40, 39, 39, 38, 38, 38, 38, 70, 69, 68, 67, 66, 62, 58, 53,
+        49, 49, 49, 49, 50, 49, 49, 48, 48, 47, 47, 46, 45, 44, 44, 43, 42, 41,
+        40, 40, 39, 39, 39, 39, 72, 70, 67, 64, 62, 59, 55, 52, 49, 49, 50, 50,
+        51, 50, 50, 50, 49, 49, 48, 47, 46, 46, 45, 44, 43, 42, 42, 41, 40, 40,
+        40, 40, 75, 71, 66, 62, 57, 55, 53, 51, 49, 49, 50, 51, 52, 52, 51, 51,
+        51, 50, 49, 49, 48, 47, 46, 45, 44, 44, 43, 42, 41, 41, 41, 41, 69, 65,
+        62, 59, 55, 53, 51, 49, 47, 48, 49, 49, 50, 50, 50, 50, 49, 49, 48, 47,
+        47, 46, 45, 44, 44, 43, 42, 41, 41, 41, 41, 41, 62, 60, 58, 55, 53, 51,
+        50, 48, 46, 47, 47, 48, 48, 48, 48, 48, 48, 47, 47, 46, 46, 45, 44, 44,
+        43, 42, 42, 41, 40, 40, 40, 40, 56, 54, 53, 52, 51, 49, 48, 47, 45, 46,
+        46, 46, 47, 47, 47, 47, 46, 46, 46, 45, 45, 44, 43, 43, 42, 42, 41, 40,
+        40, 40, 40, 40, 49, 49, 49, 49, 49, 47, 46, 45, 44, 44, 44, 45, 45, 45,
+        45, 45, 45, 45, 44, 44, 44, 43, 43, 42, 42, 41, 40, 40, 39, 39, 39, 39,
+        49, 49, 49, 49, 49, 48, 47, 46, 44, 44, 44, 44, 44, 44, 44, 44, 44, 44,
+        43, 43, 43, 42, 42, 41, 41, 40, 40, 39, 39, 39, 39, 39, 48, 49, 49, 50,
+        50, 49, 47, 46, 44, 44, 44, 44, 43, 43, 43, 43, 43, 43, 42, 42, 42, 41,
+        41, 40, 40, 39, 39, 38, 38, 38, 38, 38, 48, 49, 49, 50, 51, 49, 48, 46,
+        45, 44, 44, 43, 43, 42, 42, 42, 42, 41, 41, 41, 41, 40, 40, 39, 39, 39,
+        38, 38, 37, 37, 37, 37, 47, 48, 50, 51, 52, 50, 48, 47, 45, 44, 43, 43,
+        42, 41, 41, 41, 41, 40, 40, 40, 40, 39, 39, 39, 38, 38, 38, 37, 37, 37,
+        37, 37, 47, 48, 49, 50, 52, 50, 48, 47, 45, 44, 43, 42, 41, 41, 41, 40,
+        40, 40, 39, 39, 39, 39, 38, 38, 38, 37, 37, 37, 36, 36, 36, 36, 46, 48,
+        49, 50, 51, 50, 48, 47, 45, 44, 43, 42, 41, 41, 40, 40, 39, 39, 39, 38,
+        38, 38, 38, 37, 37, 37, 36, 36, 36, 36, 36, 36, 46, 47, 48, 50, 51, 50,
+        48, 47, 45, 44, 43, 42, 41, 40, 40, 39, 39, 38, 38, 38, 37, 37, 37, 37,
+        36, 36, 36, 35, 35, 35, 35, 35, 45, 47, 48, 49, 51, 49, 48, 46, 45, 44,
+        43, 42, 41, 40, 39, 39, 38, 38, 37, 37, 37, 37, 36, 36, 36, 35, 35, 35,
+        35, 35, 35, 35, 45, 46, 47, 49, 50, 49, 47, 46, 45, 44, 43, 41, 40, 40,
+        39, 38, 38, 37, 37, 37, 36, 36, 36, 36, 35, 35, 35, 34, 34, 34, 34, 34,
+        44, 45, 47, 48, 49, 48, 47, 46, 44, 43, 42, 41, 40, 39, 39, 38, 37, 37,
+        37, 36, 36, 36, 35, 35, 35, 35, 34, 34, 34, 34, 34, 34, 43, 45, 46, 47,
+        49, 47, 46, 45, 44, 43, 42, 41, 40, 39, 38, 38, 37, 37, 36, 36, 35, 35,
+        35, 35, 34, 34, 34, 34, 33, 33, 33, 33, 42, 44, 45, 46, 48, 47, 46, 45,
+        44, 43, 42, 41, 40, 39, 38, 37, 37, 36, 36, 35, 35, 35, 34, 34, 34, 34,
+        33, 33, 33, 33, 33, 33, 42, 43, 44, 46, 47, 46, 45, 44, 43, 42, 41, 40,
+        39, 39, 38, 37, 37, 36, 36, 35, 35, 34, 34, 34, 34, 33, 33, 33, 33, 33,
+        33, 33, 41, 42, 44, 45, 46, 45, 44, 43, 43, 42, 41, 40, 39, 38, 38, 37,
+        36, 36, 35, 35, 34, 34, 34, 34, 33, 33, 33, 33, 32, 32, 32, 32, 40, 41,
+        43, 44, 45, 44, 44, 43, 42, 41, 40, 39, 39, 38, 37, 37, 36, 36, 35, 35,
+        34, 34, 34, 33, 33, 33, 32, 32, 32, 32, 32, 32, 40, 41, 42, 43, 44, 44,
+        43, 42, 42, 41, 40, 39, 38, 38, 37, 36, 36, 35, 35, 34, 34, 34, 33, 33,
+        33, 32, 32, 32, 32, 32, 32, 32, 39, 40, 41, 42, 44, 43, 42, 42, 41, 40,
+        39, 39, 38, 37, 37, 36, 35, 35, 35, 34, 34, 33, 33, 33, 32, 32, 32, 32,
+        31, 31, 31, 31, 38, 39, 40, 42, 43, 42, 42, 41, 40, 40, 39, 38, 38, 37,
+        36, 36, 35, 35, 34, 34, 33, 33, 33, 32, 32, 32, 32, 31, 31, 31, 31, 31,
+        38, 39, 40, 41, 42, 41, 41, 40, 40, 39, 38, 38, 37, 37, 36, 35, 35, 34,
+        34, 34, 33, 33, 33, 32, 32, 32, 31, 31, 31, 31, 31, 31, 37, 38, 39, 40,
+        41, 41, 40, 40, 39, 39, 38, 37, 37, 36, 36, 35, 35, 34, 34, 33, 33, 33,
+        32, 32, 32, 31, 31, 31, 31, 31, 31, 31, 37, 38, 39, 40, 41, 41, 40, 40,
+        39, 39, 38, 37, 37, 36, 36, 35, 35, 34, 34, 33, 33, 33, 32, 32, 32, 31,
+        31, 31, 31, 31, 31, 31, 37, 38, 39, 40, 41, 41, 40, 40, 39, 39, 38, 37,
+        37, 36, 36, 35, 35, 34, 34, 33, 33, 33, 32, 32, 32, 31, 31, 31, 31, 31,
+        31, 31, 37, 38, 39, 40, 41, 41, 40, 40, 39, 39, 38, 37, 37, 36, 36, 35,
+        35, 34, 34, 33, 33, 33, 32, 32, 32, 31, 31, 31, 31, 31, 31, 31 },
+      { /* Intra matrices */
+        /* Size 4 */
+        110, 78, 74, 64, 78, 69, 66, 61, 74, 66, 57, 53, 64, 61, 53, 49,
+        /* Size 8 */
+        103, 122, 78, 75, 72, 67, 62, 58, 122, 92, 77, 83, 81, 76, 70, 65, 78,
+        77, 70, 71, 71, 69, 66, 62, 75, 83, 71, 66, 64, 62, 60, 58, 72, 81, 71,
+        64, 60, 58, 56, 54, 67, 76, 69, 62, 58, 55, 53, 51, 62, 70, 66, 60, 56,
+        53, 50, 49, 58, 65, 62, 58, 54, 51, 49, 47,
+        /* Size 16 */
+        105, 115, 124, 102, 80, 78, 77, 75, 73, 71, 68, 66, 63, 61, 59, 59, 115,
+        112, 109, 94, 79, 80, 80, 79, 78, 75, 73, 70, 67, 65, 62, 62, 124, 109,
+        94, 86, 79, 82, 84, 83, 83, 80, 77, 74, 72, 69, 66, 66, 102, 94, 86, 81,
+        75, 77, 78, 78, 78, 76, 74, 71, 69, 67, 64, 64, 80, 79, 79, 75, 71, 72,
+        73, 73, 73, 71, 70, 68, 67, 65, 63, 63, 78, 80, 82, 77, 72, 71, 70, 69,
+        69, 68, 67, 65, 64, 62, 61, 61, 77, 80, 84, 78, 73, 70, 67, 66, 65, 64,
+        63, 62, 61, 60, 59, 59, 75, 79, 83, 78, 73, 69, 66, 64, 63, 62, 61, 60,
+        59, 58, 57, 57, 73, 78, 83, 78, 73, 69, 65, 63, 61, 60, 59, 58, 57, 56,
+        55, 55, 71, 75, 80, 76, 71, 68, 64, 62, 60, 58, 57, 56, 55, 54, 53, 53,
+        68, 73, 77, 74, 70, 67, 63, 61, 59, 57, 55, 54, 53, 53, 52, 52, 66, 70,
+        74, 71, 68, 65, 62, 60, 58, 56, 54, 53, 52, 52, 51, 51, 63, 67, 72, 69,
+        67, 64, 61, 59, 57, 55, 53, 52, 51, 51, 50, 50, 61, 65, 69, 67, 65, 62,
+        60, 58, 56, 54, 53, 52, 51, 50, 49, 49, 59, 62, 66, 64, 63, 61, 59, 57,
+        55, 53, 52, 51, 50, 49, 48, 48, 59, 62, 66, 64, 63, 61, 59, 57, 55, 53,
+        52, 51, 50, 49, 48, 48,
+        /* Size 32 */
+        106, 111, 116, 121, 126, 114, 103, 92, 80, 80, 79, 78, 77, 76, 76, 75,
+        74, 73, 71, 70, 69, 68, 66, 65, 64, 63, 62, 60, 59, 59, 59, 59, 111,
+        113, 114, 116, 118, 108, 99, 90, 80, 80, 80, 79, 79, 79, 78, 77, 76, 75,
+        74, 72, 71, 70, 69, 67, 66, 65, 64, 62, 61, 61, 61, 61, 116, 114, 113,
+        112, 110, 103, 95, 87, 80, 80, 81, 81, 81, 81, 80, 79, 79, 77, 76, 75,
+        73, 72, 71, 69, 68, 67, 65, 64, 63, 63, 63, 63, 121, 116, 112, 107, 102,
+        97, 91, 85, 80, 81, 81, 82, 83, 83, 82, 82, 81, 80, 78, 77, 76, 74, 73,
+        71, 70, 69, 67, 66, 65, 65, 65, 65, 126, 118, 110, 102, 95, 91, 87, 83,
+        79, 81, 82, 84, 85, 85, 84, 84, 83, 82, 81, 79, 78, 77, 75, 74, 72, 71,
+        69, 68, 67, 67, 67, 67, 114, 108, 103, 97, 91, 87, 84, 81, 77, 79, 80,
+        81, 82, 82, 81, 81, 81, 80, 79, 77, 76, 75, 74, 72, 71, 70, 68, 67, 66,
+        66, 66, 66, 103, 99, 95, 91, 87, 84, 81, 78, 75, 76, 77, 78, 79, 79, 79,
+        79, 78, 77, 76, 75, 75, 73, 72, 71, 70, 68, 67, 66, 65, 65, 65, 65, 92,
+        90, 87, 85, 83, 81, 78, 76, 74, 74, 75, 75, 76, 76, 76, 76, 76, 75, 74,
+        73, 73, 72, 71, 70, 68, 67, 66, 65, 64, 64, 64, 64, 80, 80, 80, 80, 79,
+        77, 75, 74, 72, 72, 72, 73, 73, 73, 73, 73, 73, 73, 72, 72, 71, 70, 69,
+        68, 67, 66, 65, 64, 63, 63, 63, 63, 80, 80, 80, 81, 81, 79, 76, 74, 72,
+        72, 72, 72, 72, 72, 72, 71, 71, 71, 70, 70, 69, 68, 68, 67, 66, 65, 64,
+        63, 62, 62, 62, 62, 79, 80, 81, 81, 82, 80, 77, 75, 72, 72, 71, 71, 70,
+        70, 70, 70, 69, 69, 68, 68, 67, 67, 66, 65, 64, 64, 63, 62, 61, 61, 61,
+        61, 78, 79, 81, 82, 84, 81, 78, 75, 73, 72, 71, 70, 69, 69, 68, 68, 68,
+        67, 67, 66, 66, 65, 64, 64, 63, 62, 62, 61, 60, 60, 60, 60, 77, 79, 81,
+        83, 85, 82, 79, 76, 73, 72, 70, 69, 68, 67, 67, 66, 66, 65, 65, 64, 64,
+        63, 63, 62, 62, 61, 60, 60, 59, 59, 59, 59, 76, 79, 81, 83, 85, 82, 79,
+        76, 73, 72, 70, 69, 67, 66, 66, 65, 65, 64, 64, 63, 63, 62, 62, 61, 61,
+        60, 59, 59, 58, 58, 58, 58, 76, 78, 80, 82, 84, 81, 79, 76, 73, 72, 70,
+        68, 67, 66, 65, 64, 64, 63, 62, 62, 61, 61, 60, 60, 59, 59, 58, 58, 57,
+        57, 57, 57, 75, 77, 79, 82, 84, 81, 79, 76, 73, 71, 70, 68, 66, 65, 64,
+        63, 62, 62, 61, 61, 60, 60, 59, 59, 58, 58, 57, 57, 56, 56, 56, 56, 74,
+        76, 79, 81, 83, 81, 78, 76, 73, 71, 69, 68, 66, 65, 64, 62, 61, 61, 60,
+        60, 59, 59, 58, 58, 57, 57, 56, 56, 55, 55, 55, 55, 73, 75, 77, 80, 82,
+        80, 77, 75, 73, 71, 69, 67, 65, 64, 63, 62, 61, 60, 60, 59, 58, 58, 57,
+        57, 56, 56, 55, 55, 55, 55, 55, 55, 71, 74, 76, 78, 81, 79, 76, 74, 72,
+        70, 68, 67, 65, 64, 62, 61, 60, 60, 59, 58, 58, 57, 57, 56, 56, 55, 55,
+        54, 54, 54, 54, 54, 70, 72, 75, 77, 79, 77, 75, 73, 72, 70, 68, 66, 64,
+        63, 62, 61, 60, 59, 58, 57, 57, 56, 56, 55, 55, 54, 54, 54, 53, 53, 53,
+        53, 69, 71, 73, 76, 78, 76, 75, 73, 71, 69, 67, 66, 64, 63, 61, 60, 59,
+        58, 58, 57, 56, 55, 55, 54, 54, 54, 53, 53, 52, 52, 52, 52, 68, 70, 72,
+        74, 77, 75, 73, 72, 70, 68, 67, 65, 63, 62, 61, 60, 59, 58, 57, 56, 55,
+        55, 54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 66, 69, 71, 73, 75, 74, 72,
+        71, 69, 68, 66, 64, 63, 62, 60, 59, 58, 57, 57, 56, 55, 54, 54, 53, 53,
+        52, 52, 52, 51, 51, 51, 51, 65, 67, 69, 71, 74, 72, 71, 70, 68, 67, 65,
+        64, 62, 61, 60, 59, 58, 57, 56, 55, 54, 54, 53, 53, 52, 52, 52, 51, 51,
+        51, 51, 51, 64, 66, 68, 70, 72, 71, 70, 68, 67, 66, 64, 63, 62, 61, 59,
+        58, 57, 56, 56, 55, 54, 53, 53, 52, 52, 51, 51, 51, 50, 50, 50, 50, 63,
+        65, 67, 69, 71, 70, 68, 67, 66, 65, 64, 62, 61, 60, 59, 58, 57, 56, 55,
+        54, 54, 53, 52, 52, 51, 51, 51, 50, 50, 50, 50, 50, 62, 64, 65, 67, 69,
+        68, 67, 66, 65, 64, 63, 62, 60, 59, 58, 57, 56, 55, 55, 54, 53, 53, 52,
+        52, 51, 51, 50, 50, 49, 49, 49, 49, 60, 62, 64, 66, 68, 67, 66, 65, 64,
+        63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 54, 53, 52, 52, 51, 51, 50, 50,
+        49, 49, 49, 49, 49, 59, 61, 63, 65, 67, 66, 65, 64, 63, 62, 61, 60, 59,
+        58, 57, 56, 55, 55, 54, 53, 52, 52, 51, 51, 50, 50, 49, 49, 49, 49, 49,
+        49, 59, 61, 63, 65, 67, 66, 65, 64, 63, 62, 61, 60, 59, 58, 57, 56, 55,
+        55, 54, 53, 52, 52, 51, 51, 50, 50, 49, 49, 49, 49, 49, 49, 59, 61, 63,
+        65, 67, 66, 65, 64, 63, 62, 61, 60, 59, 58, 57, 56, 55, 55, 54, 53, 52,
+        52, 51, 51, 50, 50, 49, 49, 49, 49, 49, 49, 59, 61, 63, 65, 67, 66, 65,
+        64, 63, 62, 61, 60, 59, 58, 57, 56, 55, 55, 54, 53, 52, 52, 51, 51, 50,
+        50, 49, 49, 49, 49, 49, 49 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 60, 45, 36, 60, 47, 40, 35, 45, 40, 34, 32, 36, 35, 32, 31,
+        /* Size 8 */
+        64, 77, 73, 61, 52, 45, 41, 38, 77, 70, 73, 65, 56, 49, 44, 40, 73, 73,
+        58, 53, 49, 45, 42, 39, 61, 65, 53, 47, 44, 41, 39, 38, 52, 56, 49, 44,
+        41, 39, 37, 36, 45, 49, 45, 41, 39, 37, 36, 35, 41, 44, 42, 39, 37, 36,
+        35, 34, 38, 40, 39, 38, 36, 35, 34, 34,
+        /* Size 16 */
+        64, 71, 77, 75, 73, 67, 61, 56, 52, 48, 45, 43, 41, 40, 38, 38, 71, 72,
+        74, 73, 73, 68, 63, 59, 54, 51, 47, 45, 42, 41, 39, 39, 77, 74, 70, 71,
+        73, 69, 65, 61, 56, 53, 49, 46, 44, 42, 40, 40, 75, 73, 71, 69, 66, 62,
+        59, 56, 53, 50, 47, 45, 43, 41, 40, 40, 73, 73, 73, 66, 58, 56, 53, 51,
+        49, 47, 45, 43, 42, 40, 39, 39, 67, 68, 69, 62, 56, 53, 50, 48, 46, 45,
+        43, 42, 41, 39, 38, 38, 61, 63, 65, 59, 53, 50, 47, 45, 44, 43, 41, 40,
+        39, 38, 38, 38, 56, 59, 61, 56, 51, 48, 45, 44, 42, 41, 40, 39, 38, 38,
+        37, 37, 52, 54, 56, 53, 49, 46, 44, 42, 41, 40, 39, 38, 37, 37, 36, 36,
+        48, 51, 53, 50, 47, 45, 43, 41, 40, 39, 38, 37, 37, 36, 36, 36, 45, 47,
+        49, 47, 45, 43, 41, 40, 39, 38, 37, 37, 36, 36, 35, 35, 43, 45, 46, 45,
+        43, 42, 40, 39, 38, 37, 37, 36, 36, 35, 35, 35, 41, 42, 44, 43, 42, 41,
+        39, 38, 37, 37, 36, 36, 35, 35, 34, 34, 40, 41, 42, 41, 40, 39, 38, 38,
+        37, 36, 36, 35, 35, 34, 34, 34, 38, 39, 40, 40, 39, 38, 38, 37, 36, 36,
+        35, 35, 34, 34, 34, 34, 38, 39, 40, 40, 39, 38, 38, 37, 36, 36, 35, 35,
+        34, 34, 34, 34,
+        /* Size 32 */
+        64, 67, 71, 74, 77, 76, 75, 74, 73, 70, 67, 64, 61, 59, 56, 54, 52, 50,
+        48, 47, 45, 44, 43, 42, 41, 40, 40, 39, 38, 38, 38, 38, 67, 69, 71, 74,
+        76, 75, 74, 74, 73, 70, 68, 65, 62, 60, 58, 55, 53, 51, 50, 48, 46, 45,
+        44, 43, 42, 41, 40, 40, 39, 39, 39, 39, 71, 71, 72, 73, 74, 74, 73, 73,
+        73, 71, 68, 66, 63, 61, 59, 56, 54, 52, 51, 49, 47, 46, 45, 44, 42, 42,
+        41, 40, 39, 39, 39, 39, 74, 74, 73, 73, 72, 72, 72, 73, 73, 71, 69, 66,
+        64, 62, 60, 57, 55, 53, 52, 50, 48, 47, 46, 44, 43, 42, 42, 41, 40, 40,
+        40, 40, 77, 76, 74, 72, 70, 71, 71, 72, 73, 71, 69, 67, 65, 63, 61, 59,
+        56, 54, 53, 51, 49, 48, 46, 45, 44, 43, 42, 41, 40, 40, 40, 40, 76, 75,
+        74, 72, 71, 70, 70, 70, 69, 67, 66, 64, 62, 60, 58, 56, 54, 53, 51, 50,
+        48, 47, 46, 45, 43, 43, 42, 41, 40, 40, 40, 40, 75, 74, 73, 72, 71, 70,
+        69, 67, 66, 64, 62, 61, 59, 58, 56, 54, 53, 51, 50, 48, 47, 46, 45, 44,
+        43, 42, 41, 41, 40, 40, 40, 40, 74, 74, 73, 73, 72, 70, 67, 65, 62, 61,
+        59, 58, 56, 55, 54, 52, 51, 50, 49, 47, 46, 45, 44, 43, 42, 42, 41, 40,
+        39, 39, 39, 39, 73, 73, 73, 73, 73, 69, 66, 62, 58, 57, 56, 55, 53, 52,
+        51, 50, 49, 48, 47, 46, 45, 44, 43, 43, 42, 41, 40, 40, 39, 39, 39, 39,
+        70, 70, 71, 71, 71, 67, 64, 61, 57, 56, 55, 53, 52, 51, 50, 49, 48, 47,
+        46, 45, 44, 43, 43, 42, 41, 41, 40, 39, 39, 39, 39, 39, 67, 68, 68, 69,
+        69, 66, 62, 59, 56, 55, 53, 52, 50, 49, 48, 47, 46, 46, 45, 44, 43, 43,
+        42, 41, 41, 40, 39, 39, 38, 38, 38, 38, 64, 65, 66, 66, 67, 64, 61, 58,
+        55, 53, 52, 50, 49, 48, 47, 46, 45, 44, 44, 43, 42, 42, 41, 41, 40, 39,
+        39, 38, 38, 38, 38, 38, 61, 62, 63, 64, 65, 62, 59, 56, 53, 52, 50, 49,
+        47, 46, 45, 45, 44, 43, 43, 42, 41, 41, 40, 40, 39, 39, 38, 38, 38, 38,
+        38, 38, 59, 60, 61, 62, 63, 60, 58, 55, 52, 51, 49, 48, 46, 45, 45, 44,
+        43, 42, 42, 41, 41, 40, 40, 39, 39, 38, 38, 38, 37, 37, 37, 37, 56, 58,
+        59, 60, 61, 58, 56, 54, 51, 50, 48, 47, 45, 45, 44, 43, 42, 42, 41, 41,
+        40, 40, 39, 39, 38, 38, 38, 37, 37, 37, 37, 37, 54, 55, 56, 57, 59, 56,
+        54, 52, 50, 49, 47, 46, 45, 44, 43, 42, 41, 41, 40, 40, 39, 39, 39, 38,
+        38, 38, 37, 37, 37, 37, 37, 37, 52, 53, 54, 55, 56, 54, 53, 51, 49, 48,
+        46, 45, 44, 43, 42, 41, 41, 40, 40, 39, 39, 38, 38, 38, 37, 37, 37, 37,
+        36, 36, 36, 36, 50, 51, 52, 53, 54, 53, 51, 50, 48, 47, 46, 44, 43, 42,
+        42, 41, 40, 40, 39, 39, 38, 38, 38, 37, 37, 37, 37, 36, 36, 36, 36, 36,
+        48, 50, 51, 52, 53, 51, 50, 49, 47, 46, 45, 44, 43, 42, 41, 40, 40, 39,
+        39, 38, 38, 38, 37, 37, 37, 36, 36, 36, 36, 36, 36, 36, 47, 48, 49, 50,
+        51, 50, 48, 47, 46, 45, 44, 43, 42, 41, 41, 40, 39, 39, 38, 38, 38, 37,
+        37, 37, 36, 36, 36, 36, 35, 35, 35, 35, 45, 46, 47, 48, 49, 48, 47, 46,
+        45, 44, 43, 42, 41, 41, 40, 39, 39, 38, 38, 38, 37, 37, 37, 36, 36, 36,
+        36, 35, 35, 35, 35, 35, 44, 45, 46, 47, 48, 47, 46, 45, 44, 43, 43, 42,
+        41, 40, 40, 39, 38, 38, 38, 37, 37, 37, 36, 36, 36, 36, 35, 35, 35, 35,
+        35, 35, 43, 44, 45, 46, 46, 46, 45, 44, 43, 43, 42, 41, 40, 40, 39, 39,
+        38, 38, 37, 37, 37, 36, 36, 36, 36, 35, 35, 35, 35, 35, 35, 35, 42, 43,
+        44, 44, 45, 45, 44, 43, 43, 42, 41, 41, 40, 39, 39, 38, 38, 37, 37, 37,
+        36, 36, 36, 36, 35, 35, 35, 35, 35, 35, 35, 35, 41, 42, 42, 43, 44, 43,
+        43, 42, 42, 41, 41, 40, 39, 39, 38, 38, 37, 37, 37, 36, 36, 36, 36, 35,
+        35, 35, 35, 35, 34, 34, 34, 34, 40, 41, 42, 42, 43, 43, 42, 42, 41, 41,
+        40, 39, 39, 38, 38, 38, 37, 37, 36, 36, 36, 36, 35, 35, 35, 35, 35, 34,
+        34, 34, 34, 34, 40, 40, 41, 42, 42, 42, 41, 41, 40, 40, 39, 39, 38, 38,
+        38, 37, 37, 37, 36, 36, 36, 35, 35, 35, 35, 35, 34, 34, 34, 34, 34, 34,
+        39, 40, 40, 41, 41, 41, 41, 40, 40, 39, 39, 38, 38, 38, 37, 37, 37, 36,
+        36, 36, 35, 35, 35, 35, 35, 34, 34, 34, 34, 34, 34, 34, 38, 39, 39, 40,
+        40, 40, 40, 39, 39, 39, 38, 38, 38, 37, 37, 37, 36, 36, 36, 35, 35, 35,
+        35, 35, 34, 34, 34, 34, 34, 34, 34, 34, 38, 39, 39, 40, 40, 40, 40, 39,
+        39, 39, 38, 38, 38, 37, 37, 37, 36, 36, 36, 35, 35, 35, 35, 35, 34, 34,
+        34, 34, 34, 34, 34, 34, 38, 39, 39, 40, 40, 40, 40, 39, 39, 39, 38, 38,
+        38, 37, 37, 37, 36, 36, 36, 35, 35, 35, 35, 35, 34, 34, 34, 34, 34, 34,
+        34, 34, 38, 39, 39, 40, 40, 40, 40, 39, 39, 39, 38, 38, 38, 37, 37, 37,
+        36, 36, 36, 35, 35, 35, 35, 35, 34, 34, 34, 34, 34, 34, 34, 34 },
+      { /* Intra matrices */
+        /* Size 4 */
+        109, 102, 75, 59, 102, 79, 66, 57, 75, 66, 56, 52, 59, 57, 52, 49,
+        /* Size 8 */
+        96, 117, 111, 92, 77, 67, 60, 55, 117, 106, 110, 98, 84, 73, 64, 59,
+        111, 110, 87, 80, 73, 66, 61, 57, 92, 98, 80, 69, 64, 60, 57, 55, 77,
+        84, 73, 64, 59, 56, 54, 52, 67, 73, 66, 60, 56, 54, 52, 51, 60, 64, 61,
+        57, 54, 52, 51, 49, 55, 59, 57, 55, 52, 51, 49, 49,
+        /* Size 16 */
+        98, 109, 120, 116, 113, 103, 94, 86, 78, 73, 68, 65, 61, 59, 57, 57,
+        109, 111, 114, 113, 112, 105, 97, 89, 82, 76, 71, 67, 63, 61, 58, 58,
+        120, 114, 108, 110, 112, 106, 100, 93, 86, 80, 74, 70, 66, 63, 60, 60,
+        116, 113, 110, 105, 101, 96, 91, 85, 80, 75, 71, 67, 64, 62, 59, 59,
+        113, 112, 112, 101, 89, 85, 81, 78, 74, 71, 68, 65, 62, 60, 58, 58, 103,
+        105, 106, 96, 85, 81, 76, 73, 70, 67, 65, 62, 60, 59, 57, 57, 94, 97,
+        100, 91, 81, 76, 71, 68, 66, 64, 62, 60, 58, 57, 56, 56, 86, 89, 93, 85,
+        78, 73, 68, 66, 63, 61, 60, 58, 57, 56, 55, 55, 78, 82, 86, 80, 74, 70,
+        66, 63, 61, 59, 57, 56, 55, 54, 53, 53, 73, 76, 80, 75, 71, 67, 64, 61,
+        59, 58, 56, 55, 54, 53, 53, 53, 68, 71, 74, 71, 68, 65, 62, 60, 57, 56,
+        55, 54, 53, 52, 52, 52, 65, 67, 70, 67, 65, 62, 60, 58, 56, 55, 54, 53,
+        52, 52, 51, 51, 61, 63, 66, 64, 62, 60, 58, 57, 55, 54, 53, 52, 52, 51,
+        51, 51, 59, 61, 63, 62, 60, 59, 57, 56, 54, 53, 52, 52, 51, 51, 50, 50,
+        57, 58, 60, 59, 58, 57, 56, 55, 53, 53, 52, 51, 51, 50, 50, 50, 57, 58,
+        60, 59, 58, 57, 56, 55, 53, 53, 52, 51, 51, 50, 50, 50,
+        /* Size 32 */
+        99, 105, 110, 115, 121, 119, 118, 116, 114, 109, 104, 100, 95, 91, 87,
+        83, 79, 76, 74, 71, 69, 67, 65, 63, 62, 61, 59, 58, 57, 57, 57, 57, 105,
+        108, 111, 115, 118, 117, 116, 115, 114, 110, 105, 101, 96, 93, 89, 85,
+        81, 78, 76, 73, 70, 68, 67, 65, 63, 62, 61, 59, 58, 58, 58, 58, 110,
+        111, 113, 114, 115, 115, 114, 114, 114, 110, 106, 102, 98, 94, 90, 87,
+        83, 80, 77, 74, 72, 70, 68, 66, 64, 63, 62, 60, 59, 59, 59, 59, 115,
+        115, 114, 113, 112, 113, 113, 113, 113, 110, 106, 103, 100, 96, 92, 88,
+        85, 82, 79, 76, 73, 71, 69, 67, 65, 64, 63, 61, 60, 60, 60, 60, 121,
+        118, 115, 112, 110, 110, 111, 112, 113, 110, 107, 104, 101, 98, 94, 90,
+        87, 84, 81, 78, 75, 73, 71, 69, 67, 65, 64, 62, 61, 61, 61, 61, 119,
+        117, 115, 113, 110, 110, 109, 108, 107, 105, 102, 99, 96, 93, 90, 87,
+        84, 81, 78, 76, 73, 71, 69, 68, 66, 64, 63, 62, 60, 60, 60, 60, 118,
+        116, 114, 113, 111, 109, 106, 104, 102, 99, 97, 94, 92, 89, 86, 83, 81,
+        78, 76, 74, 72, 70, 68, 66, 65, 63, 62, 61, 60, 60, 60, 60, 116, 115,
+        114, 113, 112, 108, 104, 100, 96, 94, 91, 89, 87, 85, 82, 80, 78, 76,
+        74, 72, 70, 68, 67, 65, 64, 63, 62, 60, 59, 59, 59, 59, 114, 114, 114,
+        113, 113, 107, 102, 96, 90, 88, 86, 84, 82, 80, 79, 77, 75, 73, 72, 70,
+        68, 67, 66, 64, 63, 62, 61, 60, 59, 59, 59, 59, 109, 110, 110, 110, 110,
+        105, 99, 94, 88, 86, 84, 82, 79, 78, 76, 74, 73, 71, 70, 68, 67, 66, 64,
+        63, 62, 61, 60, 59, 58, 58, 58, 58, 104, 105, 106, 106, 107, 102, 97,
+        91, 86, 84, 81, 79, 77, 75, 74, 72, 71, 69, 68, 67, 65, 64, 63, 62, 61,
+        60, 59, 58, 57, 57, 57, 57, 100, 101, 102, 103, 104, 99, 94, 89, 84, 82,
+        79, 77, 74, 73, 71, 70, 68, 67, 66, 65, 64, 63, 62, 61, 60, 59, 58, 58,
+        57, 57, 57, 57, 95, 96, 98, 100, 101, 96, 92, 87, 82, 79, 77, 74, 72,
+        70, 69, 68, 66, 65, 64, 63, 62, 61, 61, 60, 59, 58, 58, 57, 56, 56, 56,
+        56, 91, 93, 94, 96, 98, 93, 89, 85, 80, 78, 75, 73, 70, 69, 68, 66, 65,
+        64, 63, 62, 61, 60, 60, 59, 58, 58, 57, 56, 56, 56, 56, 56, 87, 89, 90,
+        92, 94, 90, 86, 82, 79, 76, 74, 71, 69, 68, 66, 65, 64, 63, 62, 61, 60,
+        59, 59, 58, 57, 57, 56, 56, 55, 55, 55, 55, 83, 85, 87, 88, 90, 87, 83,
+        80, 77, 74, 72, 70, 68, 66, 65, 64, 62, 62, 61, 60, 59, 59, 58, 57, 57,
+        56, 56, 55, 55, 55, 55, 55, 79, 81, 83, 85, 87, 84, 81, 78, 75, 73, 71,
+        68, 66, 65, 64, 62, 61, 60, 60, 59, 58, 58, 57, 56, 56, 55, 55, 54, 54,
+        54, 54, 54, 76, 78, 80, 82, 84, 81, 78, 76, 73, 71, 69, 67, 65, 64, 63,
+        62, 60, 60, 59, 58, 57, 57, 56, 56, 55, 55, 54, 54, 54, 54, 54, 54, 74,
+        76, 77, 79, 81, 78, 76, 74, 72, 70, 68, 66, 64, 63, 62, 61, 60, 59, 58,
+        57, 57, 56, 56, 55, 55, 54, 54, 54, 53, 53, 53, 53, 71, 73, 74, 76, 78,
+        76, 74, 72, 70, 68, 67, 65, 63, 62, 61, 60, 59, 58, 57, 57, 56, 56, 55,
+        55, 54, 54, 53, 53, 53, 53, 53, 53, 69, 70, 72, 73, 75, 73, 72, 70, 68,
+        67, 65, 64, 62, 61, 60, 59, 58, 57, 57, 56, 55, 55, 55, 54, 54, 53, 53,
+        53, 52, 52, 52, 52, 67, 68, 70, 71, 73, 71, 70, 68, 67, 66, 64, 63, 61,
+        60, 59, 59, 58, 57, 56, 56, 55, 55, 54, 54, 53, 53, 53, 52, 52, 52, 52,
+        52, 65, 67, 68, 69, 71, 69, 68, 67, 66, 64, 63, 62, 61, 60, 59, 58, 57,
+        56, 56, 55, 55, 54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 52, 63, 65, 66,
+        67, 69, 68, 66, 65, 64, 63, 62, 61, 60, 59, 58, 57, 56, 56, 55, 55, 54,
+        54, 53, 53, 52, 52, 52, 52, 51, 51, 51, 51, 62, 63, 64, 65, 67, 66, 65,
+        64, 63, 62, 61, 60, 59, 58, 57, 57, 56, 55, 55, 54, 54, 53, 53, 52, 52,
+        52, 52, 51, 51, 51, 51, 51, 61, 62, 63, 64, 65, 64, 63, 63, 62, 61, 60,
+        59, 58, 58, 57, 56, 55, 55, 54, 54, 53, 53, 53, 52, 52, 52, 51, 51, 51,
+        51, 51, 51, 59, 61, 62, 63, 64, 63, 62, 62, 61, 60, 59, 58, 58, 57, 56,
+        56, 55, 54, 54, 53, 53, 53, 52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 58,
+        59, 60, 61, 62, 62, 61, 60, 60, 59, 58, 58, 57, 56, 56, 55, 54, 54, 54,
+        53, 53, 52, 52, 52, 51, 51, 51, 51, 50, 50, 50, 50, 57, 58, 59, 60, 61,
+        60, 60, 59, 59, 58, 57, 57, 56, 56, 55, 55, 54, 54, 53, 53, 52, 52, 52,
+        51, 51, 51, 51, 50, 50, 50, 50, 50, 57, 58, 59, 60, 61, 60, 60, 59, 59,
+        58, 57, 57, 56, 56, 55, 55, 54, 54, 53, 53, 52, 52, 52, 51, 51, 51, 51,
+        50, 50, 50, 50, 50, 57, 58, 59, 60, 61, 60, 60, 59, 59, 58, 57, 57, 56,
+        56, 55, 55, 54, 54, 53, 53, 52, 52, 52, 51, 51, 51, 51, 50, 50, 50, 50,
+        50, 57, 58, 59, 60, 61, 60, 60, 59, 59, 58, 57, 57, 56, 56, 55, 55, 54,
+        54, 53, 53, 52, 52, 52, 51, 51, 51, 51, 50, 50, 50, 50, 50 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 48, 46, 40, 48, 43, 42, 39, 46, 42, 37, 35, 40, 39, 35, 33,
+        /* Size 8 */
+        64, 74, 50, 49, 47, 44, 42, 39, 74, 58, 50, 53, 52, 49, 46, 43, 50, 50,
+        46, 47, 47, 45, 44, 41, 49, 53, 47, 44, 43, 42, 41, 39, 47, 52, 47, 43,
+        40, 39, 38, 37, 44, 49, 45, 42, 39, 38, 37, 36, 42, 46, 44, 41, 38, 37,
+        35, 35, 39, 43, 41, 39, 37, 36, 35, 34,
+        /* Size 16 */
+        64, 69, 74, 62, 50, 50, 49, 48, 47, 46, 44, 43, 42, 41, 39, 39, 69, 68,
+        66, 58, 50, 51, 51, 50, 50, 48, 47, 45, 44, 43, 41, 41, 74, 66, 58, 54,
+        50, 51, 53, 52, 52, 51, 49, 48, 46, 45, 43, 43, 62, 58, 54, 51, 48, 49,
+        50, 50, 49, 48, 47, 46, 45, 44, 42, 42, 50, 50, 50, 48, 46, 46, 47, 47,
+        47, 46, 45, 45, 44, 43, 41, 41, 50, 51, 51, 49, 46, 46, 45, 45, 45, 44,
+        44, 43, 42, 41, 40, 40, 49, 51, 53, 50, 47, 45, 44, 43, 43, 42, 42, 41,
+        41, 40, 39, 39, 48, 50, 52, 50, 47, 45, 43, 42, 42, 41, 41, 40, 39, 39,
+        38, 38, 47, 50, 52, 49, 47, 45, 43, 42, 40, 40, 39, 39, 38, 38, 37, 37,
+        46, 48, 51, 48, 46, 44, 42, 41, 40, 39, 38, 38, 37, 37, 37, 37, 44, 47,
+        49, 47, 45, 44, 42, 41, 39, 38, 38, 37, 37, 36, 36, 36, 43, 45, 48, 46,
+        45, 43, 41, 40, 39, 38, 37, 37, 36, 36, 35, 35, 42, 44, 46, 45, 44, 42,
+        41, 39, 38, 37, 37, 36, 35, 35, 35, 35, 41, 43, 45, 44, 43, 41, 40, 39,
+        38, 37, 36, 36, 35, 35, 34, 34, 39, 41, 43, 42, 41, 40, 39, 38, 37, 37,
+        36, 35, 35, 34, 34, 34, 39, 41, 43, 42, 41, 40, 39, 38, 37, 37, 36, 35,
+        35, 34, 34, 34,
+        /* Size 32 */
+        64, 67, 69, 72, 74, 68, 62, 56, 50, 50, 50, 49, 49, 48, 48, 48, 47, 46,
+        46, 45, 44, 44, 43, 42, 42, 41, 41, 40, 39, 39, 39, 39, 67, 67, 68, 69,
+        70, 65, 60, 55, 50, 50, 50, 50, 50, 49, 49, 49, 48, 48, 47, 46, 46, 45,
+        44, 44, 43, 42, 42, 41, 40, 40, 40, 40, 69, 68, 68, 67, 66, 62, 58, 54,
+        50, 50, 51, 51, 51, 51, 50, 50, 50, 49, 48, 48, 47, 46, 45, 45, 44, 43,
+        43, 42, 41, 41, 41, 41, 72, 69, 67, 64, 62, 59, 56, 53, 50, 51, 51, 51,
+        52, 52, 51, 51, 51, 50, 49, 49, 48, 47, 47, 46, 45, 44, 44, 43, 42, 42,
+        42, 42, 74, 70, 66, 62, 58, 56, 54, 52, 50, 51, 51, 52, 53, 53, 52, 52,
+        52, 51, 51, 50, 49, 48, 48, 47, 46, 45, 45, 44, 43, 43, 43, 43, 68, 65,
+        62, 59, 56, 54, 52, 51, 49, 50, 50, 51, 51, 51, 51, 51, 51, 50, 50, 49,
+        48, 48, 47, 46, 45, 45, 44, 43, 43, 43, 43, 43, 62, 60, 58, 56, 54, 52,
+        51, 49, 48, 48, 49, 49, 50, 50, 50, 49, 49, 49, 48, 48, 47, 47, 46, 45,
+        45, 44, 44, 43, 42, 42, 42, 42, 56, 55, 54, 53, 52, 51, 49, 48, 47, 47,
+        48, 48, 48, 48, 48, 48, 48, 48, 47, 47, 46, 46, 45, 45, 44, 44, 43, 42,
+        42, 42, 42, 42, 50, 50, 50, 50, 50, 49, 48, 47, 46, 46, 46, 46, 47, 47,
+        47, 47, 47, 46, 46, 46, 45, 45, 45, 44, 44, 43, 43, 42, 41, 41, 41, 41,
+        50, 50, 50, 51, 51, 50, 48, 47, 46, 46, 46, 46, 46, 46, 46, 46, 46, 45,
+        45, 45, 45, 44, 44, 43, 43, 42, 42, 41, 41, 41, 41, 41, 50, 50, 51, 51,
+        51, 50, 49, 48, 46, 46, 46, 45, 45, 45, 45, 45, 45, 44, 44, 44, 44, 43,
+        43, 42, 42, 42, 41, 41, 40, 40, 40, 40, 49, 50, 51, 51, 52, 51, 49, 48,
+        46, 46, 45, 45, 44, 44, 44, 44, 44, 43, 43, 43, 43, 42, 42, 42, 41, 41,
+        41, 40, 40, 40, 40, 40, 49, 50, 51, 52, 53, 51, 50, 48, 47, 46, 45, 44,
+        44, 43, 43, 43, 43, 42, 42, 42, 42, 42, 41, 41, 41, 40, 40, 40, 39, 39,
+        39, 39, 48, 49, 51, 52, 53, 51, 50, 48, 47, 46, 45, 44, 43, 43, 43, 42,
+        42, 42, 42, 41, 41, 41, 41, 40, 40, 40, 39, 39, 39, 39, 39, 39, 48, 49,
+        50, 51, 52, 51, 50, 48, 47, 46, 45, 44, 43, 43, 42, 42, 42, 41, 41, 41,
+        41, 40, 40, 40, 39, 39, 39, 39, 38, 38, 38, 38, 48, 49, 50, 51, 52, 51,
+        49, 48, 47, 46, 45, 44, 43, 42, 42, 42, 41, 41, 40, 40, 40, 40, 39, 39,
+        39, 39, 38, 38, 38, 38, 38, 38, 47, 48, 50, 51, 52, 51, 49, 48, 47, 46,
+        45, 44, 43, 42, 42, 41, 40, 40, 40, 40, 39, 39, 39, 39, 38, 38, 38, 38,
+        37, 37, 37, 37, 46, 48, 49, 50, 51, 50, 49, 48, 46, 45, 44, 43, 42, 42,
+        41, 41, 40, 40, 40, 39, 39, 39, 38, 38, 38, 38, 37, 37, 37, 37, 37, 37,
+        46, 47, 48, 49, 51, 50, 48, 47, 46, 45, 44, 43, 42, 42, 41, 40, 40, 40,
+        39, 39, 38, 38, 38, 38, 37, 37, 37, 37, 37, 37, 37, 37, 45, 46, 48, 49,
+        50, 49, 48, 47, 46, 45, 44, 43, 42, 41, 41, 40, 40, 39, 39, 38, 38, 38,
+        38, 37, 37, 37, 37, 36, 36, 36, 36, 36, 44, 46, 47, 48, 49, 48, 47, 46,
+        45, 45, 44, 43, 42, 41, 41, 40, 39, 39, 38, 38, 38, 37, 37, 37, 37, 36,
+        36, 36, 36, 36, 36, 36, 44, 45, 46, 47, 48, 48, 47, 46, 45, 44, 43, 42,
+        42, 41, 40, 40, 39, 39, 38, 38, 37, 37, 37, 37, 36, 36, 36, 36, 35, 35,
+        35, 35, 43, 44, 45, 47, 48, 47, 46, 45, 45, 44, 43, 42, 41, 41, 40, 39,
+        39, 38, 38, 38, 37, 37, 37, 36, 36, 36, 36, 35, 35, 35, 35, 35, 42, 44,
+        45, 46, 47, 46, 45, 45, 44, 43, 42, 42, 41, 40, 40, 39, 39, 38, 38, 37,
+        37, 37, 36, 36, 36, 36, 35, 35, 35, 35, 35, 35, 42, 43, 44, 45, 46, 45,
+        45, 44, 44, 43, 42, 41, 41, 40, 39, 39, 38, 38, 37, 37, 37, 36, 36, 36,
+        35, 35, 35, 35, 35, 35, 35, 35, 41, 42, 43, 44, 45, 45, 44, 44, 43, 42,
+        42, 41, 40, 40, 39, 39, 38, 38, 37, 37, 36, 36, 36, 36, 35, 35, 35, 35,
+        34, 34, 34, 34, 41, 42, 43, 44, 45, 44, 44, 43, 43, 42, 41, 41, 40, 39,
+        39, 38, 38, 37, 37, 37, 36, 36, 36, 35, 35, 35, 35, 34, 34, 34, 34, 34,
+        40, 41, 42, 43, 44, 43, 43, 42, 42, 41, 41, 40, 40, 39, 39, 38, 38, 37,
+        37, 36, 36, 36, 35, 35, 35, 35, 34, 34, 34, 34, 34, 34, 39, 40, 41, 42,
+        43, 43, 42, 42, 41, 41, 40, 40, 39, 39, 38, 38, 37, 37, 37, 36, 36, 35,
+        35, 35, 35, 34, 34, 34, 34, 34, 34, 34, 39, 40, 41, 42, 43, 43, 42, 42,
+        41, 41, 40, 40, 39, 39, 38, 38, 37, 37, 37, 36, 36, 35, 35, 35, 35, 34,
+        34, 34, 34, 34, 34, 34, 39, 40, 41, 42, 43, 43, 42, 42, 41, 41, 40, 40,
+        39, 39, 38, 38, 37, 37, 37, 36, 36, 35, 35, 35, 35, 34, 34, 34, 34, 34,
+        34, 34, 39, 40, 41, 42, 43, 43, 42, 42, 41, 41, 40, 40, 39, 39, 38, 38,
+        37, 37, 37, 36, 36, 35, 35, 35, 35, 34, 34, 34, 34, 34, 34, 34 },
+      { /* Intra matrices */
+        /* Size 4 */
+        103, 76, 72, 63, 76, 68, 66, 61, 72, 66, 58, 54, 63, 61, 54, 51,
+        /* Size 8 */
+        98, 114, 76, 73, 71, 66, 62, 58, 114, 88, 75, 80, 78, 74, 69, 64, 76,
+        75, 69, 70, 70, 68, 65, 62, 73, 80, 70, 65, 64, 62, 60, 58, 71, 78, 70,
+        64, 60, 58, 57, 55, 66, 74, 68, 62, 58, 56, 54, 53, 62, 69, 65, 60, 57,
+        54, 52, 51, 58, 64, 62, 58, 55, 53, 51, 49,
+        /* Size 16 */
+        99, 107, 115, 96, 77, 76, 74, 73, 72, 69, 67, 65, 63, 61, 59, 59, 107,
+        105, 102, 90, 77, 77, 78, 77, 76, 73, 71, 69, 67, 64, 62, 62, 115, 102,
+        89, 83, 76, 79, 81, 80, 80, 77, 75, 73, 70, 68, 65, 65, 96, 90, 83, 78,
+        73, 74, 76, 76, 75, 74, 72, 70, 68, 66, 64, 64, 77, 77, 76, 73, 70, 70,
+        71, 71, 71, 70, 69, 68, 66, 64, 63, 63, 76, 77, 79, 74, 70, 69, 69, 68,
+        68, 67, 66, 65, 64, 62, 61, 61, 74, 78, 81, 76, 71, 69, 66, 65, 65, 64,
+        63, 62, 61, 60, 59, 59, 73, 77, 80, 76, 71, 68, 65, 64, 63, 62, 61, 60,
+        59, 58, 57, 57, 72, 76, 80, 75, 71, 68, 65, 63, 61, 60, 59, 58, 57, 57,
+        56, 56, 69, 73, 77, 74, 70, 67, 64, 62, 60, 59, 58, 57, 56, 55, 55, 55,
+        67, 71, 75, 72, 69, 66, 63, 61, 59, 58, 56, 56, 55, 54, 53, 53, 65, 69,
+        73, 70, 68, 65, 62, 60, 58, 57, 56, 55, 54, 53, 52, 52, 63, 67, 70, 68,
+        66, 64, 61, 59, 57, 56, 55, 54, 53, 52, 51, 51, 61, 64, 68, 66, 64, 62,
+        60, 58, 57, 55, 54, 53, 52, 51, 51, 51, 59, 62, 65, 64, 63, 61, 59, 57,
+        56, 55, 53, 52, 51, 51, 50, 50, 59, 62, 65, 64, 63, 61, 59, 57, 56, 55,
+        53, 52, 51, 51, 50, 50,
+        /* Size 32 */
+        100, 104, 108, 112, 116, 107, 97, 87, 78, 77, 76, 76, 75, 74, 74, 73,
+        72, 71, 70, 69, 68, 67, 66, 65, 63, 63, 62, 61, 60, 60, 60, 60, 104,
+        105, 107, 108, 110, 102, 94, 86, 78, 77, 77, 77, 77, 76, 75, 75, 74, 73,
+        72, 71, 70, 69, 68, 66, 65, 64, 63, 62, 61, 61, 61, 61, 108, 107, 106,
+        104, 103, 97, 90, 84, 77, 78, 78, 78, 78, 78, 77, 77, 76, 75, 74, 73,
+        72, 71, 69, 68, 67, 66, 65, 64, 63, 63, 63, 63, 112, 108, 104, 100, 96,
+        92, 87, 82, 77, 78, 79, 79, 80, 80, 79, 79, 78, 77, 76, 75, 74, 73, 71,
+        70, 69, 68, 67, 65, 64, 64, 64, 64, 116, 110, 103, 96, 90, 87, 83, 80,
+        77, 78, 79, 81, 82, 81, 81, 81, 80, 79, 78, 77, 76, 74, 73, 72, 71, 69,
+        68, 67, 66, 66, 66, 66, 107, 102, 97, 92, 87, 84, 81, 78, 75, 76, 77,
+        78, 79, 79, 79, 78, 78, 77, 76, 75, 74, 73, 72, 71, 70, 68, 67, 66, 65,
+        65, 65, 65, 97, 94, 90, 87, 83, 81, 78, 76, 74, 74, 75, 76, 77, 76, 76,
+        76, 76, 75, 74, 73, 73, 72, 71, 70, 69, 68, 67, 65, 64, 64, 64, 64, 87,
+        86, 84, 82, 80, 78, 76, 74, 72, 72, 73, 73, 74, 74, 74, 74, 74, 73, 72,
+        72, 71, 70, 69, 68, 67, 67, 66, 65, 64, 64, 64, 64, 78, 78, 77, 77, 77,
+        75, 74, 72, 70, 70, 71, 71, 72, 72, 72, 72, 72, 71, 71, 70, 70, 69, 68,
+        67, 66, 66, 65, 64, 63, 63, 63, 63, 77, 77, 78, 78, 78, 76, 74, 72, 70,
+        70, 70, 70, 70, 70, 70, 70, 70, 70, 69, 69, 68, 67, 67, 66, 65, 64, 64,
+        63, 62, 62, 62, 62, 76, 77, 78, 79, 79, 77, 75, 73, 71, 70, 70, 70, 69,
+        69, 69, 69, 68, 68, 67, 67, 67, 66, 65, 65, 64, 63, 63, 62, 61, 61, 61,
+        61, 76, 77, 78, 79, 81, 78, 76, 73, 71, 70, 70, 69, 68, 68, 67, 67, 67,
+        66, 66, 66, 65, 65, 64, 63, 63, 62, 62, 61, 60, 60, 60, 60, 75, 77, 78,
+        80, 82, 79, 77, 74, 72, 70, 69, 68, 67, 66, 66, 65, 65, 65, 64, 64, 64,
+        63, 63, 62, 62, 61, 61, 60, 60, 60, 60, 60, 74, 76, 78, 80, 81, 79, 76,
+        74, 72, 70, 69, 68, 66, 66, 65, 65, 64, 64, 63, 63, 63, 62, 62, 61, 61,
+        60, 60, 59, 59, 59, 59, 59, 74, 75, 77, 79, 81, 79, 76, 74, 72, 70, 69,
+        67, 66, 65, 65, 64, 63, 63, 62, 62, 62, 61, 61, 60, 60, 59, 59, 58, 58,
+        58, 58, 58, 73, 75, 77, 79, 81, 78, 76, 74, 72, 70, 69, 67, 65, 65, 64,
+        63, 62, 62, 61, 61, 60, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 57, 72,
+        74, 76, 78, 80, 78, 76, 74, 72, 70, 68, 67, 65, 64, 63, 62, 61, 61, 60,
+        60, 59, 59, 59, 58, 58, 57, 57, 57, 56, 56, 56, 56, 71, 73, 75, 77, 79,
+        77, 75, 73, 71, 70, 68, 66, 65, 64, 63, 62, 61, 60, 60, 59, 59, 58, 58,
+        58, 57, 57, 56, 56, 56, 56, 56, 56, 70, 72, 74, 76, 78, 76, 74, 72, 71,
+        69, 67, 66, 64, 63, 62, 61, 60, 60, 59, 59, 58, 58, 57, 57, 56, 56, 56,
+        55, 55, 55, 55, 55, 69, 71, 73, 75, 77, 75, 73, 72, 70, 69, 67, 66, 64,
+        63, 62, 61, 60, 59, 59, 58, 57, 57, 57, 56, 56, 55, 55, 55, 54, 54, 54,
+        54, 68, 70, 72, 74, 76, 74, 73, 71, 70, 68, 67, 65, 64, 63, 62, 60, 59,
+        59, 58, 57, 57, 56, 56, 56, 55, 55, 54, 54, 54, 54, 54, 54, 67, 69, 71,
+        73, 74, 73, 72, 70, 69, 67, 66, 65, 63, 62, 61, 60, 59, 58, 58, 57, 56,
+        56, 56, 55, 55, 54, 54, 54, 53, 53, 53, 53, 66, 68, 69, 71, 73, 72, 71,
+        69, 68, 67, 65, 64, 63, 62, 61, 60, 59, 58, 57, 57, 56, 56, 55, 55, 54,
+        54, 53, 53, 53, 53, 53, 53, 65, 66, 68, 70, 72, 71, 70, 68, 67, 66, 65,
+        63, 62, 61, 60, 59, 58, 58, 57, 56, 56, 55, 55, 54, 54, 53, 53, 53, 52,
+        52, 52, 52, 63, 65, 67, 69, 71, 70, 69, 67, 66, 65, 64, 63, 62, 61, 60,
+        59, 58, 57, 56, 56, 55, 55, 54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 63,
+        64, 66, 68, 69, 68, 68, 67, 66, 64, 63, 62, 61, 60, 59, 58, 57, 57, 56,
+        55, 55, 54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 52, 62, 63, 65, 67, 68,
+        67, 67, 66, 65, 64, 63, 62, 61, 60, 59, 58, 57, 56, 56, 55, 54, 54, 53,
+        53, 53, 52, 52, 52, 51, 51, 51, 51, 61, 62, 64, 65, 67, 66, 65, 65, 64,
+        63, 62, 61, 60, 59, 58, 58, 57, 56, 55, 55, 54, 54, 53, 53, 52, 52, 52,
+        51, 51, 51, 51, 51, 60, 61, 63, 64, 66, 65, 64, 64, 63, 62, 61, 60, 60,
+        59, 58, 57, 56, 56, 55, 54, 54, 53, 53, 52, 52, 52, 51, 51, 51, 51, 51,
+        51, 60, 61, 63, 64, 66, 65, 64, 64, 63, 62, 61, 60, 60, 59, 58, 57, 56,
+        56, 55, 54, 54, 53, 53, 52, 52, 52, 51, 51, 51, 51, 51, 51, 60, 61, 63,
+        64, 66, 65, 64, 64, 63, 62, 61, 60, 60, 59, 58, 57, 56, 56, 55, 54, 54,
+        53, 53, 52, 52, 52, 51, 51, 51, 51, 51, 51, 60, 61, 63, 64, 66, 65, 64,
+        64, 63, 62, 61, 60, 60, 59, 58, 57, 56, 56, 55, 54, 54, 53, 53, 52, 52,
+        52, 51, 51, 51, 51, 51, 51 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 61, 47, 39, 61, 49, 42, 38, 47, 42, 38, 35, 39, 38, 35, 34,
+        /* Size 8 */
+        64, 76, 72, 62, 53, 47, 44, 41, 76, 70, 72, 65, 57, 51, 46, 43, 72, 72,
+        59, 55, 51, 47, 44, 42, 62, 65, 55, 49, 46, 44, 42, 41, 53, 57, 51, 46,
+        43, 42, 41, 40, 47, 51, 47, 44, 42, 40, 39, 39, 44, 46, 44, 42, 41, 39,
+        39, 38, 41, 43, 42, 41, 40, 39, 38, 37,
+        /* Size 16 */
+        64, 70, 76, 74, 72, 67, 62, 57, 53, 50, 47, 46, 44, 43, 41, 41, 70, 71,
+        73, 72, 72, 68, 63, 59, 55, 52, 49, 47, 45, 44, 42, 42, 76, 73, 70, 71,
+        72, 68, 65, 61, 57, 54, 51, 49, 46, 45, 43, 43, 74, 72, 71, 68, 65, 63,
+        60, 57, 54, 52, 49, 47, 45, 44, 43, 43, 72, 72, 72, 65, 59, 57, 55, 53,
+        51, 49, 47, 46, 44, 43, 42, 42, 67, 68, 68, 63, 57, 54, 52, 50, 49, 47,
+        46, 44, 43, 42, 41, 41, 62, 63, 65, 60, 55, 52, 49, 48, 46, 45, 44, 43,
+        42, 41, 41, 41, 57, 59, 61, 57, 53, 50, 48, 46, 45, 44, 43, 42, 41, 41,
+        40, 40, 53, 55, 57, 54, 51, 49, 46, 45, 43, 43, 42, 41, 41, 40, 40, 40,
+        50, 52, 54, 52, 49, 47, 45, 44, 43, 42, 41, 40, 40, 40, 39, 39, 47, 49,
+        51, 49, 47, 46, 44, 43, 42, 41, 40, 40, 39, 39, 39, 39, 46, 47, 49, 47,
+        46, 44, 43, 42, 41, 40, 40, 39, 39, 39, 38, 38, 44, 45, 46, 45, 44, 43,
+        42, 41, 41, 40, 39, 39, 39, 38, 38, 38, 43, 44, 45, 44, 43, 42, 41, 41,
+        40, 40, 39, 39, 38, 38, 38, 38, 41, 42, 43, 43, 42, 41, 41, 40, 40, 39,
+        39, 38, 38, 38, 37, 37, 41, 42, 43, 43, 42, 41, 41, 40, 40, 39, 39, 38,
+        38, 38, 37, 37,
+        /* Size 32 */
+        64, 67, 70, 73, 76, 75, 74, 73, 72, 70, 67, 64, 62, 60, 57, 55, 53, 52,
+        50, 49, 47, 47, 46, 45, 44, 43, 43, 42, 41, 41, 41, 41, 67, 69, 71, 72,
+        74, 74, 73, 73, 72, 70, 67, 65, 63, 60, 58, 56, 54, 53, 51, 50, 48, 47,
+        46, 45, 44, 44, 43, 42, 42, 42, 42, 42, 70, 71, 71, 72, 73, 73, 72, 72,
+        72, 70, 68, 66, 63, 61, 59, 57, 55, 54, 52, 51, 49, 48, 47, 46, 45, 44,
+        44, 43, 42, 42, 42, 42, 73, 72, 72, 72, 71, 71, 71, 72, 72, 70, 68, 66,
+        64, 62, 60, 58, 56, 55, 53, 52, 50, 49, 48, 47, 46, 45, 44, 43, 43, 43,
+        43, 43, 76, 74, 73, 71, 70, 70, 71, 71, 72, 70, 68, 67, 65, 63, 61, 59,
+        57, 56, 54, 52, 51, 50, 49, 47, 46, 46, 45, 44, 43, 43, 43, 43, 75, 74,
+        73, 71, 70, 70, 69, 69, 68, 67, 66, 64, 63, 61, 59, 57, 56, 54, 53, 51,
+        50, 49, 48, 47, 46, 45, 44, 44, 43, 43, 43, 43, 74, 73, 72, 71, 71, 69,
+        68, 67, 65, 64, 63, 61, 60, 58, 57, 56, 54, 53, 52, 50, 49, 48, 47, 46,
+        45, 45, 44, 43, 43, 43, 43, 43, 73, 73, 72, 72, 71, 69, 67, 64, 62, 61,
+        60, 59, 57, 56, 55, 54, 52, 51, 50, 49, 48, 47, 47, 46, 45, 44, 44, 43,
+        42, 42, 42, 42, 72, 72, 72, 72, 72, 68, 65, 62, 59, 58, 57, 56, 55, 54,
+        53, 52, 51, 50, 49, 48, 47, 47, 46, 45, 44, 44, 43, 43, 42, 42, 42, 42,
+        70, 70, 70, 70, 70, 67, 64, 61, 58, 57, 56, 55, 53, 52, 52, 51, 50, 49,
+        48, 47, 47, 46, 45, 45, 44, 43, 43, 42, 42, 42, 42, 42, 67, 67, 68, 68,
+        68, 66, 63, 60, 57, 56, 54, 53, 52, 51, 50, 49, 49, 48, 47, 46, 46, 45,
+        44, 44, 43, 43, 42, 42, 41, 41, 41, 41, 64, 65, 66, 66, 67, 64, 61, 59,
+        56, 55, 53, 52, 51, 50, 49, 48, 47, 47, 46, 45, 45, 44, 44, 43, 43, 42,
+        42, 42, 41, 41, 41, 41, 62, 63, 63, 64, 65, 63, 60, 57, 55, 53, 52, 51,
+        49, 48, 48, 47, 46, 46, 45, 45, 44, 44, 43, 43, 42, 42, 41, 41, 41, 41,
+        41, 41, 60, 60, 61, 62, 63, 61, 58, 56, 54, 52, 51, 50, 48, 48, 47, 46,
+        46, 45, 44, 44, 43, 43, 43, 42, 42, 41, 41, 41, 40, 40, 40, 40, 57, 58,
+        59, 60, 61, 59, 57, 55, 53, 52, 50, 49, 48, 47, 46, 46, 45, 44, 44, 43,
+        43, 43, 42, 42, 41, 41, 41, 40, 40, 40, 40, 40, 55, 56, 57, 58, 59, 57,
+        56, 54, 52, 51, 49, 48, 47, 46, 46, 45, 44, 44, 43, 43, 42, 42, 42, 41,
+        41, 41, 40, 40, 40, 40, 40, 40, 53, 54, 55, 56, 57, 56, 54, 52, 51, 50,
+        49, 47, 46, 46, 45, 44, 43, 43, 43, 42, 42, 41, 41, 41, 41, 40, 40, 40,
+        40, 40, 40, 40, 52, 53, 54, 55, 56, 54, 53, 51, 50, 49, 48, 47, 46, 45,
+        44, 44, 43, 43, 42, 42, 41, 41, 41, 41, 40, 40, 40, 40, 39, 39, 39, 39,
+        50, 51, 52, 53, 54, 53, 52, 50, 49, 48, 47, 46, 45, 44, 44, 43, 43, 42,
+        42, 41, 41, 41, 40, 40, 40, 40, 40, 39, 39, 39, 39, 39, 49, 50, 51, 52,
+        52, 51, 50, 49, 48, 47, 46, 45, 45, 44, 43, 43, 42, 42, 41, 41, 41, 40,
+        40, 40, 40, 39, 39, 39, 39, 39, 39, 39, 47, 48, 49, 50, 51, 50, 49, 48,
+        47, 47, 46, 45, 44, 43, 43, 42, 42, 41, 41, 41, 40, 40, 40, 40, 39, 39,
+        39, 39, 39, 39, 39, 39, 47, 47, 48, 49, 50, 49, 48, 47, 47, 46, 45, 44,
+        44, 43, 43, 42, 41, 41, 41, 40, 40, 40, 40, 39, 39, 39, 39, 39, 38, 38,
+        38, 38, 46, 46, 47, 48, 49, 48, 47, 47, 46, 45, 44, 44, 43, 43, 42, 42,
+        41, 41, 40, 40, 40, 40, 39, 39, 39, 39, 39, 38, 38, 38, 38, 38, 45, 45,
+        46, 47, 47, 47, 46, 46, 45, 45, 44, 43, 43, 42, 42, 41, 41, 41, 40, 40,
+        40, 39, 39, 39, 39, 39, 38, 38, 38, 38, 38, 38, 44, 44, 45, 46, 46, 46,
+        45, 45, 44, 44, 43, 43, 42, 42, 41, 41, 41, 40, 40, 40, 39, 39, 39, 39,
+        39, 38, 38, 38, 38, 38, 38, 38, 43, 44, 44, 45, 46, 45, 45, 44, 44, 43,
+        43, 42, 42, 41, 41, 41, 40, 40, 40, 39, 39, 39, 39, 39, 38, 38, 38, 38,
+        38, 38, 38, 38, 43, 43, 44, 44, 45, 44, 44, 44, 43, 43, 42, 42, 41, 41,
+        41, 40, 40, 40, 40, 39, 39, 39, 39, 38, 38, 38, 38, 38, 38, 38, 38, 38,
+        42, 42, 43, 43, 44, 44, 43, 43, 43, 42, 42, 42, 41, 41, 40, 40, 40, 40,
+        39, 39, 39, 39, 38, 38, 38, 38, 38, 38, 38, 38, 38, 38, 41, 42, 42, 43,
+        43, 43, 43, 42, 42, 42, 41, 41, 41, 40, 40, 40, 40, 39, 39, 39, 39, 38,
+        38, 38, 38, 38, 38, 38, 37, 37, 37, 37, 41, 42, 42, 43, 43, 43, 43, 42,
+        42, 42, 41, 41, 41, 40, 40, 40, 40, 39, 39, 39, 39, 38, 38, 38, 38, 38,
+        38, 38, 37, 37, 37, 37, 41, 42, 42, 43, 43, 43, 43, 42, 42, 42, 41, 41,
+        41, 40, 40, 40, 40, 39, 39, 39, 39, 38, 38, 38, 38, 38, 38, 38, 37, 37,
+        37, 37, 41, 42, 42, 43, 43, 43, 43, 42, 42, 42, 41, 41, 41, 40, 40, 40,
+        40, 39, 39, 39, 39, 38, 38, 38, 38, 38, 38, 38, 37, 37, 37, 37 },
+      { /* Intra matrices */
+        /* Size 4 */
+        100, 95, 72, 60, 95, 76, 65, 58, 72, 65, 57, 54, 60, 58, 54, 51,
+        /* Size 8 */
+        90, 107, 102, 86, 74, 65, 60, 56, 107, 98, 101, 92, 80, 70, 64, 59, 102,
+        101, 83, 76, 71, 65, 61, 58, 86, 92, 76, 68, 64, 60, 58, 56, 74, 80, 71,
+        64, 60, 57, 55, 54, 65, 70, 65, 60, 57, 55, 53, 52, 60, 64, 61, 58, 55,
+        53, 52, 51, 56, 59, 58, 56, 54, 52, 51, 51,
+        /* Size 16 */
+        92, 100, 109, 107, 104, 96, 88, 82, 75, 71, 67, 64, 61, 59, 57, 57, 100,
+        102, 105, 104, 103, 97, 91, 84, 78, 74, 69, 66, 63, 61, 59, 59, 109,
+        105, 100, 101, 103, 98, 93, 87, 81, 76, 72, 68, 65, 63, 60, 60, 107,
+        104, 101, 97, 94, 89, 85, 81, 77, 73, 69, 66, 63, 61, 59, 59, 104, 103,
+        103, 94, 84, 81, 78, 75, 72, 69, 66, 64, 62, 60, 59, 59, 96, 97, 98, 89,
+        81, 77, 73, 71, 68, 66, 64, 62, 60, 59, 58, 58, 88, 91, 93, 85, 78, 73,
+        69, 67, 65, 63, 62, 60, 59, 58, 57, 57, 82, 84, 87, 81, 75, 71, 67, 65,
+        63, 61, 60, 59, 58, 57, 56, 56, 75, 78, 81, 77, 72, 68, 65, 63, 61, 59,
+        58, 57, 56, 55, 55, 55, 71, 74, 76, 73, 69, 66, 63, 61, 59, 58, 57, 56,
+        55, 55, 54, 54, 67, 69, 72, 69, 66, 64, 62, 60, 58, 57, 56, 55, 54, 54,
+        53, 53, 64, 66, 68, 66, 64, 62, 60, 59, 57, 56, 55, 54, 54, 53, 53, 53,
+        61, 63, 65, 63, 62, 60, 59, 58, 56, 55, 54, 54, 53, 53, 52, 52, 59, 61,
+        63, 61, 60, 59, 58, 57, 55, 55, 54, 53, 53, 52, 52, 52, 57, 59, 60, 59,
+        59, 58, 57, 56, 55, 54, 53, 53, 52, 52, 52, 52, 57, 59, 60, 59, 59, 58,
+        57, 56, 55, 54, 53, 53, 52, 52, 52, 52,
+        /* Size 32 */
+        92, 97, 101, 106, 110, 109, 107, 106, 105, 101, 97, 93, 89, 86, 82, 79,
+        76, 74, 72, 69, 67, 66, 64, 63, 62, 61, 60, 59, 58, 58, 58, 58, 97, 100,
+        102, 105, 108, 107, 106, 105, 104, 101, 97, 94, 90, 87, 84, 81, 77, 75,
+        73, 71, 69, 67, 66, 64, 63, 62, 61, 60, 59, 59, 59, 59, 101, 102, 103,
+        105, 106, 105, 105, 105, 104, 101, 98, 95, 91, 88, 85, 82, 79, 77, 74,
+        72, 70, 68, 67, 65, 64, 63, 61, 60, 59, 59, 59, 59, 106, 105, 105, 104,
+        103, 103, 104, 104, 104, 101, 98, 96, 93, 90, 87, 84, 80, 78, 76, 73,
+        71, 69, 68, 66, 65, 63, 62, 61, 60, 60, 60, 60, 110, 108, 106, 103, 101,
+        102, 102, 103, 104, 101, 99, 97, 94, 91, 88, 85, 82, 80, 77, 75, 72, 71,
+        69, 67, 66, 64, 63, 62, 61, 61, 61, 61, 109, 107, 105, 103, 102, 101,
+        100, 100, 99, 97, 95, 92, 90, 88, 85, 82, 80, 77, 75, 73, 71, 69, 68,
+        66, 65, 64, 63, 61, 60, 60, 60, 60, 107, 106, 105, 104, 102, 100, 98,
+        96, 94, 92, 90, 88, 86, 84, 82, 79, 77, 75, 73, 72, 70, 68, 67, 65, 64,
+        63, 62, 61, 60, 60, 60, 60, 106, 105, 105, 104, 103, 100, 96, 93, 90,
+        88, 86, 84, 82, 80, 79, 77, 75, 73, 72, 70, 68, 67, 66, 65, 63, 62, 61,
+        60, 60, 60, 60, 60, 105, 104, 104, 104, 104, 99, 94, 90, 85, 83, 82, 80,
+        78, 77, 75, 74, 73, 71, 70, 68, 67, 66, 65, 64, 63, 62, 61, 60, 59, 59,
+        59, 59, 101, 101, 101, 101, 101, 97, 92, 88, 83, 82, 80, 78, 76, 75, 73,
+        72, 71, 69, 68, 67, 66, 65, 64, 63, 62, 61, 60, 59, 59, 59, 59, 59, 97,
+        97, 98, 98, 99, 95, 90, 86, 82, 80, 78, 76, 74, 73, 71, 70, 69, 68, 67,
+        66, 65, 64, 63, 62, 61, 60, 60, 59, 58, 58, 58, 58, 93, 94, 95, 96, 97,
+        92, 88, 84, 80, 78, 76, 74, 72, 71, 70, 68, 67, 66, 65, 64, 63, 63, 62,
+        61, 60, 59, 59, 58, 58, 58, 58, 58, 89, 90, 91, 93, 94, 90, 86, 82, 78,
+        76, 74, 72, 70, 69, 68, 66, 65, 65, 64, 63, 62, 61, 61, 60, 59, 59, 58,
+        58, 57, 57, 57, 57, 86, 87, 88, 90, 91, 88, 84, 80, 77, 75, 73, 71, 69,
+        68, 66, 65, 64, 64, 63, 62, 61, 61, 60, 59, 59, 58, 58, 57, 57, 57, 57,
+        57, 82, 84, 85, 87, 88, 85, 82, 79, 75, 73, 71, 70, 68, 66, 65, 64, 63,
+        63, 62, 61, 60, 60, 59, 59, 58, 58, 57, 57, 56, 56, 56, 56, 79, 81, 82,
+        84, 85, 82, 79, 77, 74, 72, 70, 68, 66, 65, 64, 63, 62, 62, 61, 60, 59,
+        59, 58, 58, 57, 57, 57, 56, 56, 56, 56, 56, 76, 77, 79, 80, 82, 80, 77,
+        75, 73, 71, 69, 67, 65, 64, 63, 62, 61, 61, 60, 59, 59, 58, 58, 57, 57,
+        56, 56, 56, 55, 55, 55, 55, 74, 75, 77, 78, 80, 77, 75, 73, 71, 69, 68,
+        66, 65, 64, 63, 62, 61, 60, 59, 59, 58, 58, 57, 57, 56, 56, 56, 55, 55,
+        55, 55, 55, 72, 73, 74, 76, 77, 75, 73, 72, 70, 68, 67, 65, 64, 63, 62,
+        61, 60, 59, 59, 58, 57, 57, 57, 56, 56, 55, 55, 55, 55, 55, 55, 55, 69,
+        71, 72, 73, 75, 73, 72, 70, 68, 67, 66, 64, 63, 62, 61, 60, 59, 59, 58,
+        58, 57, 57, 56, 56, 55, 55, 55, 54, 54, 54, 54, 54, 67, 69, 70, 71, 72,
+        71, 70, 68, 67, 66, 65, 63, 62, 61, 60, 59, 59, 58, 57, 57, 56, 56, 56,
+        55, 55, 55, 54, 54, 54, 54, 54, 54, 66, 67, 68, 69, 71, 69, 68, 67, 66,
+        65, 64, 63, 61, 61, 60, 59, 58, 58, 57, 57, 56, 56, 55, 55, 55, 54, 54,
+        54, 54, 54, 54, 54, 64, 66, 67, 68, 69, 68, 67, 66, 65, 64, 63, 62, 61,
+        60, 59, 58, 58, 57, 57, 56, 56, 55, 55, 55, 54, 54, 54, 54, 53, 53, 53,
+        53, 63, 64, 65, 66, 67, 66, 65, 65, 64, 63, 62, 61, 60, 59, 59, 58, 57,
+        57, 56, 56, 55, 55, 55, 54, 54, 54, 54, 53, 53, 53, 53, 53, 62, 63, 64,
+        65, 66, 65, 64, 63, 63, 62, 61, 60, 59, 59, 58, 57, 57, 56, 56, 55, 55,
+        55, 54, 54, 54, 53, 53, 53, 53, 53, 53, 53, 61, 62, 63, 63, 64, 64, 63,
+        62, 62, 61, 60, 59, 59, 58, 58, 57, 56, 56, 55, 55, 55, 54, 54, 54, 53,
+        53, 53, 53, 53, 53, 53, 53, 60, 61, 61, 62, 63, 63, 62, 61, 61, 60, 60,
+        59, 58, 58, 57, 57, 56, 56, 55, 55, 54, 54, 54, 54, 53, 53, 53, 53, 52,
+        52, 52, 52, 59, 60, 60, 61, 62, 61, 61, 60, 60, 59, 59, 58, 58, 57, 57,
+        56, 56, 55, 55, 54, 54, 54, 54, 53, 53, 53, 53, 52, 52, 52, 52, 52, 58,
+        59, 59, 60, 61, 60, 60, 60, 59, 59, 58, 58, 57, 57, 56, 56, 55, 55, 55,
+        54, 54, 54, 53, 53, 53, 53, 52, 52, 52, 52, 52, 52, 58, 59, 59, 60, 61,
+        60, 60, 60, 59, 59, 58, 58, 57, 57, 56, 56, 55, 55, 55, 54, 54, 54, 53,
+        53, 53, 53, 52, 52, 52, 52, 52, 52, 58, 59, 59, 60, 61, 60, 60, 60, 59,
+        59, 58, 58, 57, 57, 56, 56, 55, 55, 55, 54, 54, 54, 53, 53, 53, 53, 52,
+        52, 52, 52, 52, 52, 58, 59, 59, 60, 61, 60, 60, 60, 59, 59, 58, 58, 57,
+        57, 56, 56, 55, 55, 55, 54, 54, 54, 53, 53, 53, 53, 52, 52, 52, 52, 52,
+        52 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 50, 48, 43, 50, 45, 44, 42, 48, 44, 40, 38, 43, 42, 38, 36,
+        /* Size 8 */
+        64, 73, 52, 50, 49, 47, 44, 42, 73, 59, 51, 54, 53, 51, 48, 45, 52, 51,
+        48, 49, 49, 47, 46, 44, 50, 54, 49, 46, 45, 44, 43, 42, 49, 53, 49, 45,
+        43, 42, 41, 40, 47, 51, 47, 44, 42, 40, 40, 39, 44, 48, 46, 43, 41, 40,
+        39, 38, 42, 45, 44, 42, 40, 39, 38, 37,
+        /* Size 16 */
+        64, 69, 73, 63, 52, 51, 50, 50, 49, 48, 47, 45, 44, 43, 42, 42, 69, 67,
+        66, 59, 52, 52, 52, 52, 51, 50, 49, 47, 46, 45, 44, 44, 73, 66, 59, 55,
+        51, 53, 54, 54, 53, 52, 51, 49, 48, 47, 45, 45, 63, 59, 55, 52, 50, 50,
+        51, 51, 51, 50, 49, 48, 47, 46, 45, 45, 52, 52, 51, 50, 48, 48, 49, 49,
+        49, 48, 47, 47, 46, 45, 44, 44, 51, 52, 53, 50, 48, 48, 47, 47, 47, 46,
+        46, 45, 44, 44, 43, 43, 50, 52, 54, 51, 49, 47, 46, 45, 45, 45, 44, 44,
+        43, 43, 42, 42, 50, 52, 54, 51, 49, 47, 45, 45, 44, 44, 43, 43, 42, 42,
+        41, 41, 49, 51, 53, 51, 49, 47, 45, 44, 43, 42, 42, 41, 41, 41, 40, 40,
+        48, 50, 52, 50, 48, 46, 45, 44, 42, 42, 41, 41, 40, 40, 39, 39, 47, 49,
+        51, 49, 47, 46, 44, 43, 42, 41, 40, 40, 40, 39, 39, 39, 45, 47, 49, 48,
+        47, 45, 44, 43, 41, 41, 40, 40, 39, 39, 38, 38, 44, 46, 48, 47, 46, 44,
+        43, 42, 41, 40, 40, 39, 39, 38, 38, 38, 43, 45, 47, 46, 45, 44, 43, 42,
+        41, 40, 39, 39, 38, 38, 37, 37, 42, 44, 45, 45, 44, 43, 42, 41, 40, 39,
+        39, 38, 38, 37, 37, 37, 42, 44, 45, 45, 44, 43, 42, 41, 40, 39, 39, 38,
+        38, 37, 37, 37,
+        /* Size 32 */
+        64, 66, 69, 71, 73, 68, 63, 57, 52, 52, 51, 51, 50, 50, 50, 49, 49, 48,
+        48, 47, 47, 46, 45, 45, 44, 44, 43, 43, 42, 42, 42, 42, 66, 67, 68, 69,
+        69, 65, 61, 56, 52, 52, 52, 52, 51, 51, 51, 50, 50, 49, 49, 48, 48, 47,
+        46, 46, 45, 45, 44, 43, 43, 43, 43, 43, 69, 68, 67, 67, 66, 62, 59, 55,
+        52, 52, 52, 52, 52, 52, 52, 51, 51, 51, 50, 49, 49, 48, 47, 47, 46, 46,
+        45, 44, 44, 44, 44, 44, 71, 69, 67, 64, 62, 60, 57, 54, 52, 52, 52, 53,
+        53, 53, 53, 52, 52, 52, 51, 50, 50, 49, 48, 48, 47, 46, 46, 45, 45, 45,
+        45, 45, 73, 69, 66, 62, 59, 57, 55, 53, 51, 52, 53, 53, 54, 54, 54, 54,
+        53, 53, 52, 51, 51, 50, 49, 49, 48, 47, 47, 46, 45, 45, 45, 45, 68, 65,
+        62, 60, 57, 55, 54, 52, 51, 51, 52, 52, 53, 53, 52, 52, 52, 52, 51, 51,
+        50, 49, 49, 48, 47, 47, 46, 46, 45, 45, 45, 45, 63, 61, 59, 57, 55, 54,
+        52, 51, 50, 50, 50, 51, 51, 51, 51, 51, 51, 51, 50, 50, 49, 49, 48, 47,
+        47, 46, 46, 45, 45, 45, 45, 45, 57, 56, 55, 54, 53, 52, 51, 50, 49, 49,
+        49, 50, 50, 50, 50, 50, 50, 49, 49, 49, 48, 48, 47, 47, 46, 46, 45, 45,
+        44, 44, 44, 44, 52, 52, 52, 52, 51, 51, 50, 49, 48, 48, 48, 48, 49, 49,
+        49, 49, 49, 48, 48, 48, 47, 47, 47, 46, 46, 45, 45, 44, 44, 44, 44, 44,
+        52, 52, 52, 52, 52, 51, 50, 49, 48, 48, 48, 48, 48, 48, 48, 48, 48, 47,
+        47, 47, 47, 46, 46, 46, 45, 45, 44, 44, 43, 43, 43, 43, 51, 52, 52, 52,
+        53, 52, 50, 49, 48, 48, 48, 47, 47, 47, 47, 47, 47, 47, 46, 46, 46, 46,
+        45, 45, 44, 44, 44, 43, 43, 43, 43, 43, 51, 52, 52, 53, 53, 52, 51, 50,
+        48, 48, 47, 47, 47, 46, 46, 46, 46, 46, 45, 45, 45, 45, 44, 44, 44, 43,
+        43, 43, 42, 42, 42, 42, 50, 51, 52, 53, 54, 53, 51, 50, 49, 48, 47, 47,
+        46, 46, 45, 45, 45, 45, 45, 44, 44, 44, 44, 43, 43, 43, 43, 42, 42, 42,
+        42, 42, 50, 51, 52, 53, 54, 53, 51, 50, 49, 48, 47, 46, 46, 45, 45, 45,
+        45, 44, 44, 44, 44, 43, 43, 43, 43, 42, 42, 42, 42, 42, 42, 42, 50, 51,
+        52, 53, 54, 52, 51, 50, 49, 48, 47, 46, 45, 45, 45, 44, 44, 44, 44, 43,
+        43, 43, 43, 42, 42, 42, 42, 41, 41, 41, 41, 41, 49, 50, 51, 52, 54, 52,
+        51, 50, 49, 48, 47, 46, 45, 45, 44, 44, 44, 43, 43, 43, 42, 42, 42, 42,
+        42, 41, 41, 41, 41, 41, 41, 41, 49, 50, 51, 52, 53, 52, 51, 50, 49, 48,
+        47, 46, 45, 45, 44, 44, 43, 43, 42, 42, 42, 42, 41, 41, 41, 41, 41, 40,
+        40, 40, 40, 40, 48, 49, 51, 52, 53, 52, 51, 49, 48, 47, 47, 46, 45, 44,
+        44, 43, 43, 42, 42, 42, 42, 41, 41, 41, 41, 40, 40, 40, 40, 40, 40, 40,
+        48, 49, 50, 51, 52, 51, 50, 49, 48, 47, 46, 45, 45, 44, 44, 43, 42, 42,
+        42, 42, 41, 41, 41, 41, 40, 40, 40, 40, 39, 39, 39, 39, 47, 48, 49, 50,
+        51, 51, 50, 49, 48, 47, 46, 45, 44, 44, 43, 43, 42, 42, 42, 41, 41, 41,
+        40, 40, 40, 40, 40, 39, 39, 39, 39, 39, 47, 48, 49, 50, 51, 50, 49, 48,
+        47, 47, 46, 45, 44, 44, 43, 42, 42, 42, 41, 41, 40, 40, 40, 40, 40, 39,
+        39, 39, 39, 39, 39, 39, 46, 47, 48, 49, 50, 49, 49, 48, 47, 46, 46, 45,
+        44, 43, 43, 42, 42, 41, 41, 41, 40, 40, 40, 40, 39, 39, 39, 39, 39, 39,
+        39, 39, 45, 46, 47, 48, 49, 49, 48, 47, 47, 46, 45, 44, 44, 43, 43, 42,
+        41, 41, 41, 40, 40, 40, 40, 39, 39, 39, 39, 38, 38, 38, 38, 38, 45, 46,
+        47, 48, 49, 48, 47, 47, 46, 46, 45, 44, 43, 43, 42, 42, 41, 41, 41, 40,
+        40, 40, 39, 39, 39, 39, 38, 38, 38, 38, 38, 38, 44, 45, 46, 47, 48, 47,
+        47, 46, 46, 45, 44, 44, 43, 43, 42, 42, 41, 41, 40, 40, 40, 39, 39, 39,
+        39, 38, 38, 38, 38, 38, 38, 38, 44, 45, 46, 46, 47, 47, 46, 46, 45, 45,
+        44, 43, 43, 42, 42, 41, 41, 40, 40, 40, 39, 39, 39, 39, 38, 38, 38, 38,
+        38, 38, 38, 38, 43, 44, 45, 46, 47, 46, 46, 45, 45, 44, 44, 43, 43, 42,
+        42, 41, 41, 40, 40, 40, 39, 39, 39, 38, 38, 38, 38, 38, 37, 37, 37, 37,
+        43, 43, 44, 45, 46, 46, 45, 45, 44, 44, 43, 43, 42, 42, 41, 41, 40, 40,
+        40, 39, 39, 39, 38, 38, 38, 38, 38, 37, 37, 37, 37, 37, 42, 43, 44, 45,
+        45, 45, 45, 44, 44, 43, 43, 42, 42, 42, 41, 41, 40, 40, 39, 39, 39, 39,
+        38, 38, 38, 38, 37, 37, 37, 37, 37, 37, 42, 43, 44, 45, 45, 45, 45, 44,
+        44, 43, 43, 42, 42, 42, 41, 41, 40, 40, 39, 39, 39, 39, 38, 38, 38, 38,
+        37, 37, 37, 37, 37, 37, 42, 43, 44, 45, 45, 45, 45, 44, 44, 43, 43, 42,
+        42, 42, 41, 41, 40, 40, 39, 39, 39, 39, 38, 38, 38, 38, 37, 37, 37, 37,
+        37, 37, 42, 43, 44, 45, 45, 45, 45, 44, 44, 43, 43, 42, 42, 42, 41, 41,
+        40, 40, 39, 39, 39, 39, 38, 38, 38, 38, 37, 37, 37, 37, 37, 37 },
+      { /* Intra matrices */
+        /* Size 4 */
+        97, 74, 71, 63, 74, 67, 65, 61, 71, 65, 58, 56, 63, 61, 56, 52,
+        /* Size 8 */
+        92, 106, 74, 72, 69, 66, 62, 59, 106, 84, 73, 77, 76, 72, 68, 64, 74,
+        73, 68, 69, 69, 67, 65, 62, 72, 77, 69, 65, 63, 62, 61, 59, 69, 76, 69,
+        63, 60, 59, 57, 56, 66, 72, 67, 62, 59, 57, 55, 54, 62, 68, 65, 61, 57,
+        55, 54, 53, 59, 64, 62, 59, 56, 54, 53, 51,
+        /* Size 16 */
+        93, 100, 107, 91, 75, 74, 73, 71, 70, 68, 67, 65, 63, 61, 60, 60, 100,
+        98, 96, 85, 74, 75, 75, 74, 74, 72, 70, 68, 66, 64, 62, 62, 107, 96, 85,
+        79, 74, 76, 78, 77, 77, 75, 73, 71, 69, 67, 65, 65, 91, 85, 79, 75, 71,
+        73, 74, 74, 73, 72, 71, 69, 67, 65, 64, 64, 75, 74, 74, 71, 68, 69, 70,
+        70, 70, 69, 68, 67, 65, 64, 63, 63, 74, 75, 76, 73, 69, 68, 68, 67, 67,
+        66, 65, 64, 63, 62, 61, 61, 73, 75, 78, 74, 70, 68, 66, 65, 64, 64, 63,
+        62, 61, 60, 60, 60, 71, 74, 77, 74, 70, 67, 65, 64, 63, 62, 61, 61, 60,
+        59, 58, 58, 70, 74, 77, 73, 70, 67, 64, 63, 61, 60, 59, 59, 58, 58, 57,
+        57, 68, 72, 75, 72, 69, 66, 64, 62, 60, 59, 58, 58, 57, 56, 56, 56, 67,
+        70, 73, 71, 68, 65, 63, 61, 59, 58, 57, 57, 56, 55, 55, 55, 65, 68, 71,
+        69, 67, 64, 62, 61, 59, 58, 57, 56, 55, 55, 54, 54, 63, 66, 69, 67, 65,
+        63, 61, 60, 58, 57, 56, 55, 54, 54, 53, 53, 61, 64, 67, 65, 64, 62, 60,
+        59, 58, 56, 55, 55, 54, 53, 53, 53, 60, 62, 65, 64, 63, 61, 60, 58, 57,
+        56, 55, 54, 53, 53, 52, 52, 60, 62, 65, 64, 63, 61, 60, 58, 57, 56, 55,
+        54, 53, 53, 52, 52,
+        /* Size 32 */
+        94, 97, 101, 104, 108, 100, 91, 83, 75, 75, 74, 74, 73, 72, 72, 71, 71,
+        70, 69, 68, 67, 66, 65, 64, 63, 63, 62, 61, 60, 60, 60, 60, 97, 98, 100,
+        101, 102, 95, 89, 82, 75, 75, 75, 75, 74, 74, 73, 73, 72, 71, 70, 70,
+        69, 68, 67, 66, 65, 64, 63, 62, 61, 61, 61, 61, 101, 100, 99, 98, 97,
+        91, 86, 80, 75, 75, 75, 76, 76, 75, 75, 74, 74, 73, 72, 71, 70, 69, 68,
+        67, 66, 65, 65, 64, 63, 63, 63, 63, 104, 101, 98, 94, 91, 87, 83, 79,
+        75, 75, 76, 77, 77, 77, 76, 76, 76, 75, 74, 73, 72, 71, 70, 69, 68, 67,
+        66, 65, 64, 64, 64, 64, 108, 102, 97, 91, 85, 83, 80, 77, 75, 76, 77,
+        78, 79, 78, 78, 78, 77, 76, 75, 75, 74, 73, 71, 70, 69, 68, 67, 66, 65,
+        65, 65, 65, 100, 95, 91, 87, 83, 80, 78, 76, 73, 74, 75, 76, 76, 76, 76,
+        76, 76, 75, 74, 73, 72, 71, 70, 69, 68, 68, 67, 66, 65, 65, 65, 65, 91,
+        89, 86, 83, 80, 78, 76, 74, 72, 72, 73, 74, 74, 74, 74, 74, 74, 73, 72,
+        72, 71, 70, 69, 68, 68, 67, 66, 65, 64, 64, 64, 64, 83, 82, 80, 79, 77,
+        76, 74, 72, 70, 71, 71, 72, 72, 72, 72, 72, 72, 71, 71, 70, 70, 69, 68,
+        67, 67, 66, 65, 64, 64, 64, 64, 64, 75, 75, 75, 75, 75, 73, 72, 70, 69,
+        69, 69, 70, 70, 70, 70, 70, 70, 70, 69, 69, 68, 68, 67, 66, 66, 65, 64,
+        64, 63, 63, 63, 63, 75, 75, 75, 75, 76, 74, 72, 71, 69, 69, 69, 69, 69,
+        69, 69, 69, 69, 68, 68, 68, 67, 67, 66, 65, 65, 64, 63, 63, 62, 62, 62,
+        62, 74, 75, 75, 76, 77, 75, 73, 71, 69, 69, 69, 68, 68, 68, 68, 68, 67,
+        67, 67, 66, 66, 65, 65, 64, 64, 63, 63, 62, 61, 61, 61, 61, 74, 75, 76,
+        77, 78, 76, 74, 72, 70, 69, 68, 68, 67, 67, 67, 66, 66, 66, 65, 65, 65,
+        64, 64, 63, 63, 62, 62, 61, 61, 61, 61, 61, 73, 74, 76, 77, 79, 76, 74,
+        72, 70, 69, 68, 67, 66, 66, 65, 65, 65, 64, 64, 64, 63, 63, 63, 62, 62,
+        61, 61, 60, 60, 60, 60, 60, 72, 74, 75, 77, 78, 76, 74, 72, 70, 69, 68,
+        67, 66, 65, 65, 64, 64, 64, 63, 63, 63, 62, 62, 61, 61, 61, 60, 60, 59,
+        59, 59, 59, 72, 73, 75, 76, 78, 76, 74, 72, 70, 69, 68, 67, 65, 65, 64,
+        64, 63, 63, 62, 62, 62, 61, 61, 61, 60, 60, 59, 59, 59, 59, 59, 59, 71,
+        73, 74, 76, 78, 76, 74, 72, 70, 69, 68, 66, 65, 64, 64, 63, 62, 62, 62,
+        61, 61, 60, 60, 60, 59, 59, 59, 58, 58, 58, 58, 58, 71, 72, 74, 76, 77,
+        76, 74, 72, 70, 69, 67, 66, 65, 64, 63, 62, 62, 61, 61, 60, 60, 60, 59,
+        59, 59, 58, 58, 58, 57, 57, 57, 57, 70, 71, 73, 75, 76, 75, 73, 71, 70,
+        68, 67, 66, 64, 64, 63, 62, 61, 61, 60, 60, 59, 59, 59, 58, 58, 58, 57,
+        57, 57, 57, 57, 57, 69, 70, 72, 74, 75, 74, 72, 71, 69, 68, 67, 65, 64,
+        63, 62, 62, 61, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56,
+        56, 68, 70, 71, 73, 75, 73, 72, 70, 69, 68, 66, 65, 64, 63, 62, 61, 60,
+        60, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56, 56, 56, 67, 69, 70,
+        72, 74, 72, 71, 70, 68, 67, 66, 65, 63, 63, 62, 61, 60, 59, 59, 58, 58,
+        57, 57, 57, 56, 56, 56, 55, 55, 55, 55, 55, 66, 68, 69, 71, 73, 71, 70,
+        69, 68, 67, 65, 64, 63, 62, 61, 60, 60, 59, 58, 58, 57, 57, 57, 56, 56,
+        56, 55, 55, 55, 55, 55, 55, 65, 67, 68, 70, 71, 70, 69, 68, 67, 66, 65,
+        64, 63, 62, 61, 60, 59, 59, 58, 58, 57, 57, 56, 56, 55, 55, 55, 55, 54,
+        54, 54, 54, 64, 66, 67, 69, 70, 69, 68, 67, 66, 65, 64, 63, 62, 61, 61,
+        60, 59, 58, 58, 57, 57, 56, 56, 55, 55, 55, 55, 54, 54, 54, 54, 54, 63,
+        65, 66, 68, 69, 68, 68, 67, 66, 65, 64, 63, 62, 61, 60, 59, 59, 58, 57,
+        57, 56, 56, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 63, 64, 65, 67, 68,
+        68, 67, 66, 65, 64, 63, 62, 61, 61, 60, 59, 58, 58, 57, 57, 56, 56, 55,
+        55, 54, 54, 54, 54, 53, 53, 53, 53, 62, 63, 65, 66, 67, 67, 66, 65, 64,
+        63, 63, 62, 61, 60, 59, 59, 58, 57, 57, 56, 56, 55, 55, 55, 54, 54, 54,
+        53, 53, 53, 53, 53, 61, 62, 64, 65, 66, 66, 65, 64, 64, 63, 62, 61, 60,
+        60, 59, 58, 58, 57, 56, 56, 55, 55, 55, 54, 54, 54, 53, 53, 53, 53, 53,
+        53, 60, 61, 63, 64, 65, 65, 64, 64, 63, 62, 61, 61, 60, 59, 59, 58, 57,
+        57, 56, 56, 55, 55, 54, 54, 54, 53, 53, 53, 52, 52, 52, 52, 60, 61, 63,
+        64, 65, 65, 64, 64, 63, 62, 61, 61, 60, 59, 59, 58, 57, 57, 56, 56, 55,
+        55, 54, 54, 54, 53, 53, 53, 52, 52, 52, 52, 60, 61, 63, 64, 65, 65, 64,
+        64, 63, 62, 61, 61, 60, 59, 59, 58, 57, 57, 56, 56, 55, 55, 54, 54, 54,
+        53, 53, 53, 52, 52, 52, 52, 60, 61, 63, 64, 65, 65, 64, 64, 63, 62, 61,
+        61, 60, 59, 59, 58, 57, 57, 56, 56, 55, 55, 54, 54, 54, 53, 53, 53, 52,
+        52, 52, 52 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 61, 49, 42, 61, 51, 45, 41, 49, 45, 41, 39, 42, 41, 39, 38,
+        /* Size 8 */
+        64, 74, 71, 62, 55, 50, 47, 45, 74, 69, 70, 65, 58, 53, 49, 46, 71, 70,
+        60, 56, 53, 50, 47, 45, 62, 65, 56, 51, 49, 47, 45, 44, 55, 58, 53, 49,
+        46, 45, 44, 43, 50, 53, 50, 47, 45, 44, 43, 42, 47, 49, 47, 45, 44, 43,
+        42, 42, 45, 46, 45, 44, 43, 42, 42, 41,
+        /* Size 16 */
+        64, 69, 74, 73, 71, 66, 62, 58, 55, 52, 50, 48, 47, 46, 45, 45, 69, 70,
+        71, 71, 71, 67, 63, 60, 56, 54, 51, 49, 48, 47, 45, 45, 74, 71, 69, 70,
+        70, 68, 65, 62, 58, 55, 53, 51, 49, 48, 46, 46, 73, 71, 70, 67, 65, 63,
+        61, 58, 55, 53, 51, 50, 48, 47, 46, 46, 71, 71, 70, 65, 60, 58, 56, 54,
+        53, 51, 50, 48, 47, 46, 45, 45, 66, 67, 68, 63, 58, 56, 54, 52, 51, 50,
+        48, 47, 46, 45, 45, 45, 62, 63, 65, 61, 56, 54, 51, 50, 49, 48, 47, 46,
+        45, 45, 44, 44, 58, 60, 62, 58, 54, 52, 50, 49, 48, 47, 46, 45, 45, 44,
+        44, 44, 55, 56, 58, 55, 53, 51, 49, 48, 46, 46, 45, 44, 44, 43, 43, 43,
+        52, 54, 55, 53, 51, 50, 48, 47, 46, 45, 44, 44, 43, 43, 43, 43, 50, 51,
+        53, 51, 50, 48, 47, 46, 45, 44, 44, 43, 43, 43, 42, 42, 48, 49, 51, 50,
+        48, 47, 46, 45, 44, 44, 43, 43, 42, 42, 42, 42, 47, 48, 49, 48, 47, 46,
+        45, 45, 44, 43, 43, 42, 42, 42, 42, 42, 46, 47, 48, 47, 46, 45, 45, 44,
+        43, 43, 43, 42, 42, 42, 41, 41, 45, 45, 46, 46, 45, 45, 44, 44, 43, 43,
+        42, 42, 42, 41, 41, 41, 45, 45, 46, 46, 45, 45, 44, 44, 43, 43, 42, 42,
+        42, 41, 41, 41,
+        /* Size 32 */
+        64, 67, 69, 72, 74, 73, 73, 72, 71, 69, 66, 64, 62, 60, 58, 57, 55, 53,
+        52, 51, 50, 49, 48, 47, 47, 46, 46, 45, 45, 45, 45, 45, 67, 68, 70, 71,
+        73, 72, 72, 71, 71, 69, 67, 65, 63, 61, 59, 57, 56, 54, 53, 52, 51, 50,
+        49, 48, 47, 47, 46, 45, 45, 45, 45, 45, 69, 70, 70, 71, 71, 71, 71, 71,
+        71, 69, 67, 65, 63, 62, 60, 58, 56, 55, 54, 53, 51, 50, 49, 49, 48, 47,
+        47, 46, 45, 45, 45, 45, 72, 71, 71, 71, 70, 70, 70, 70, 71, 69, 67, 66,
+        64, 63, 61, 59, 57, 56, 55, 53, 52, 51, 50, 49, 48, 48, 47, 46, 46, 46,
+        46, 46, 74, 73, 71, 70, 69, 69, 70, 70, 70, 69, 68, 66, 65, 63, 62, 60,
+        58, 57, 55, 54, 53, 52, 51, 50, 49, 48, 48, 47, 46, 46, 46, 46, 73, 72,
+        71, 70, 69, 69, 69, 68, 68, 67, 65, 64, 63, 61, 60, 58, 57, 56, 54, 53,
+        52, 51, 50, 49, 48, 48, 47, 47, 46, 46, 46, 46, 73, 72, 71, 70, 70, 69,
+        67, 66, 65, 64, 63, 62, 61, 59, 58, 57, 55, 54, 53, 52, 51, 50, 50, 49,
+        48, 47, 47, 46, 46, 46, 46, 46, 72, 71, 71, 70, 70, 68, 66, 64, 62, 61,
+        60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 50, 50, 49, 48, 48, 47, 47, 46,
+        45, 45, 45, 45, 71, 71, 71, 71, 70, 68, 65, 62, 60, 59, 58, 57, 56, 55,
+        54, 54, 53, 52, 51, 50, 50, 49, 48, 48, 47, 47, 46, 46, 45, 45, 45, 45,
+        69, 69, 69, 69, 69, 67, 64, 61, 59, 58, 57, 56, 55, 54, 53, 53, 52, 51,
+        50, 50, 49, 48, 48, 47, 47, 46, 46, 45, 45, 45, 45, 45, 66, 67, 67, 67,
+        68, 65, 63, 60, 58, 57, 56, 55, 54, 53, 52, 51, 51, 50, 50, 49, 48, 48,
+        47, 47, 46, 46, 45, 45, 45, 45, 45, 45, 64, 65, 65, 66, 66, 64, 62, 59,
+        57, 56, 55, 54, 52, 52, 51, 50, 50, 49, 49, 48, 48, 47, 47, 46, 46, 45,
+        45, 45, 44, 44, 44, 44, 62, 63, 63, 64, 65, 63, 61, 58, 56, 55, 54, 52,
+        51, 51, 50, 49, 49, 48, 48, 47, 47, 46, 46, 46, 45, 45, 45, 44, 44, 44,
+        44, 44, 60, 61, 62, 63, 63, 61, 59, 57, 55, 54, 53, 52, 51, 50, 49, 49,
+        48, 48, 47, 47, 46, 46, 46, 45, 45, 45, 44, 44, 44, 44, 44, 44, 58, 59,
+        60, 61, 62, 60, 58, 56, 54, 53, 52, 51, 50, 49, 49, 48, 48, 47, 47, 46,
+        46, 46, 45, 45, 45, 44, 44, 44, 44, 44, 44, 44, 57, 57, 58, 59, 60, 58,
+        57, 55, 54, 53, 51, 50, 49, 49, 48, 48, 47, 47, 46, 46, 45, 45, 45, 45,
+        44, 44, 44, 44, 43, 43, 43, 43, 55, 56, 56, 57, 58, 57, 55, 54, 53, 52,
+        51, 50, 49, 48, 48, 47, 46, 46, 46, 45, 45, 45, 44, 44, 44, 44, 43, 43,
+        43, 43, 43, 43, 53, 54, 55, 56, 57, 56, 54, 53, 52, 51, 50, 49, 48, 48,
+        47, 47, 46, 46, 45, 45, 45, 44, 44, 44, 44, 43, 43, 43, 43, 43, 43, 43,
+        52, 53, 54, 55, 55, 54, 53, 52, 51, 50, 50, 49, 48, 47, 47, 46, 46, 45,
+        45, 45, 44, 44, 44, 44, 43, 43, 43, 43, 43, 43, 43, 43, 51, 52, 53, 53,
+        54, 53, 52, 51, 50, 50, 49, 48, 47, 47, 46, 46, 45, 45, 45, 44, 44, 44,
+        44, 43, 43, 43, 43, 43, 42, 42, 42, 42, 50, 51, 51, 52, 53, 52, 51, 50,
+        50, 49, 48, 48, 47, 46, 46, 45, 45, 45, 44, 44, 44, 43, 43, 43, 43, 43,
+        43, 42, 42, 42, 42, 42, 49, 50, 50, 51, 52, 51, 50, 50, 49, 48, 48, 47,
+        46, 46, 46, 45, 45, 44, 44, 44, 43, 43, 43, 43, 43, 43, 42, 42, 42, 42,
+        42, 42, 48, 49, 49, 50, 51, 50, 50, 49, 48, 48, 47, 47, 46, 46, 45, 45,
+        44, 44, 44, 44, 43, 43, 43, 43, 42, 42, 42, 42, 42, 42, 42, 42, 47, 48,
+        49, 49, 50, 49, 49, 48, 48, 47, 47, 46, 46, 45, 45, 45, 44, 44, 44, 43,
+        43, 43, 43, 43, 42, 42, 42, 42, 42, 42, 42, 42, 47, 47, 48, 48, 49, 48,
+        48, 48, 47, 47, 46, 46, 45, 45, 45, 44, 44, 44, 43, 43, 43, 43, 42, 42,
+        42, 42, 42, 42, 42, 42, 42, 42, 46, 47, 47, 48, 48, 48, 47, 47, 47, 46,
+        46, 45, 45, 45, 44, 44, 44, 43, 43, 43, 43, 43, 42, 42, 42, 42, 42, 42,
+        42, 42, 42, 42, 46, 46, 47, 47, 48, 47, 47, 47, 46, 46, 45, 45, 45, 44,
+        44, 44, 43, 43, 43, 43, 43, 42, 42, 42, 42, 42, 42, 42, 41, 41, 41, 41,
+        45, 45, 46, 46, 47, 47, 46, 46, 46, 45, 45, 45, 44, 44, 44, 44, 43, 43,
+        43, 43, 42, 42, 42, 42, 42, 42, 42, 41, 41, 41, 41, 41, 45, 45, 45, 46,
+        46, 46, 46, 45, 45, 45, 45, 44, 44, 44, 44, 43, 43, 43, 43, 42, 42, 42,
+        42, 42, 42, 42, 41, 41, 41, 41, 41, 41, 45, 45, 45, 46, 46, 46, 46, 45,
+        45, 45, 45, 44, 44, 44, 44, 43, 43, 43, 43, 42, 42, 42, 42, 42, 42, 42,
+        41, 41, 41, 41, 41, 41, 45, 45, 45, 46, 46, 46, 46, 45, 45, 45, 45, 44,
+        44, 44, 44, 43, 43, 43, 43, 42, 42, 42, 42, 42, 42, 42, 41, 41, 41, 41,
+        41, 41, 45, 45, 45, 46, 46, 46, 46, 45, 45, 45, 45, 44, 44, 44, 44, 43,
+        43, 43, 43, 42, 42, 42, 42, 42, 42, 42, 41, 41, 41, 41, 41, 41 },
+      { /* Intra matrices */
+        /* Size 4 */
+        93, 88, 70, 60, 88, 73, 64, 58, 70, 64, 58, 55, 60, 58, 55, 53,
+        /* Size 8 */
+        84, 98, 94, 82, 71, 65, 60, 57, 98, 91, 93, 86, 76, 69, 63, 60, 94, 93,
+        79, 73, 69, 65, 61, 58, 82, 86, 73, 67, 63, 61, 58, 57, 71, 76, 69, 63,
+        60, 58, 56, 55, 65, 69, 65, 61, 58, 56, 55, 54, 60, 63, 61, 58, 56, 55,
+        54, 53, 57, 60, 58, 57, 55, 54, 53, 53,
+        /* Size 16 */
+        86, 93, 100, 98, 96, 89, 83, 78, 73, 69, 66, 63, 61, 60, 58, 58, 93, 94,
+        96, 96, 95, 90, 85, 80, 75, 71, 68, 65, 63, 61, 59, 59, 100, 96, 93, 94,
+        95, 91, 87, 82, 77, 74, 70, 67, 64, 62, 61, 61, 98, 96, 94, 90, 87, 84,
+        81, 77, 74, 71, 68, 65, 63, 61, 60, 60, 96, 95, 95, 87, 80, 77, 74, 72,
+        70, 68, 66, 64, 62, 61, 59, 59, 89, 90, 91, 84, 77, 74, 71, 69, 67, 65,
+        64, 62, 61, 59, 58, 58, 83, 85, 87, 81, 74, 71, 68, 66, 64, 63, 62, 60,
+        59, 58, 58, 58, 78, 80, 82, 77, 72, 69, 66, 64, 62, 61, 60, 59, 58, 58,
+        57, 57, 73, 75, 77, 74, 70, 67, 64, 62, 61, 60, 59, 58, 57, 57, 56, 56,
+        69, 71, 74, 71, 68, 65, 63, 61, 60, 59, 58, 57, 57, 56, 56, 56, 66, 68,
+        70, 68, 66, 64, 62, 60, 59, 58, 57, 56, 56, 55, 55, 55, 63, 65, 67, 65,
+        64, 62, 60, 59, 58, 57, 56, 56, 55, 55, 55, 55, 61, 63, 64, 63, 62, 61,
+        59, 58, 57, 57, 56, 55, 55, 55, 54, 54, 60, 61, 62, 61, 61, 59, 58, 58,
+        57, 56, 55, 55, 55, 54, 54, 54, 58, 59, 61, 60, 59, 58, 58, 57, 56, 56,
+        55, 55, 54, 54, 54, 54, 58, 59, 61, 60, 59, 58, 58, 57, 56, 56, 55, 55,
+        54, 54, 54, 54,
+        /* Size 32 */
+        86, 90, 93, 97, 101, 100, 98, 97, 96, 93, 90, 87, 83, 81, 78, 76, 73,
+        71, 70, 68, 66, 65, 64, 63, 62, 61, 60, 59, 59, 59, 59, 59, 90, 92, 94,
+        97, 99, 98, 97, 97, 96, 93, 90, 87, 84, 82, 79, 77, 74, 73, 71, 69, 67,
+        66, 65, 64, 62, 62, 61, 60, 59, 59, 59, 59, 93, 94, 95, 96, 97, 97, 96,
+        96, 96, 93, 91, 88, 86, 83, 81, 78, 76, 74, 72, 70, 68, 67, 66, 64, 63,
+        62, 62, 61, 60, 60, 60, 60, 97, 97, 96, 96, 95, 95, 95, 96, 96, 93, 91,
+        89, 87, 84, 82, 79, 77, 75, 73, 71, 69, 68, 67, 65, 64, 63, 62, 61, 60,
+        60, 60, 60, 101, 99, 97, 95, 93, 94, 94, 95, 95, 94, 92, 90, 88, 85, 83,
+        80, 78, 76, 74, 72, 70, 69, 68, 66, 65, 64, 63, 62, 61, 61, 61, 61, 100,
+        98, 97, 95, 94, 93, 93, 92, 92, 90, 88, 86, 85, 82, 80, 78, 76, 74, 73,
+        71, 69, 68, 67, 65, 64, 63, 62, 62, 61, 61, 61, 61, 98, 97, 96, 95, 94,
+        93, 91, 90, 88, 86, 85, 83, 81, 80, 78, 76, 74, 73, 71, 70, 68, 67, 66,
+        65, 64, 63, 62, 61, 60, 60, 60, 60, 97, 97, 96, 96, 95, 92, 90, 87, 84,
+        83, 81, 80, 78, 77, 75, 74, 72, 71, 70, 68, 67, 66, 65, 64, 63, 62, 61,
+        61, 60, 60, 60, 60, 96, 96, 96, 96, 95, 92, 88, 84, 80, 79, 78, 76, 75,
+        74, 73, 72, 70, 69, 68, 67, 66, 65, 64, 63, 62, 62, 61, 60, 60, 60, 60,
+        60, 93, 93, 93, 93, 94, 90, 86, 83, 79, 78, 76, 75, 73, 72, 71, 70, 69,
+        68, 67, 66, 65, 64, 63, 63, 62, 61, 60, 60, 59, 59, 59, 59, 90, 90, 91,
+        91, 92, 88, 85, 81, 78, 76, 75, 73, 72, 71, 70, 69, 67, 67, 66, 65, 64,
+        63, 63, 62, 61, 61, 60, 59, 59, 59, 59, 59, 87, 87, 88, 89, 90, 86, 83,
+        80, 76, 75, 73, 72, 70, 69, 68, 67, 66, 65, 65, 64, 63, 62, 62, 61, 60,
+        60, 59, 59, 58, 58, 58, 58, 83, 84, 86, 87, 88, 85, 81, 78, 75, 73, 72,
+        70, 68, 67, 66, 66, 65, 64, 63, 63, 62, 61, 61, 60, 60, 59, 59, 58, 58,
+        58, 58, 58, 81, 82, 83, 84, 85, 82, 80, 77, 74, 72, 71, 69, 67, 66, 66,
+        65, 64, 63, 63, 62, 61, 61, 60, 60, 59, 59, 58, 58, 58, 58, 58, 58, 78,
+        79, 81, 82, 83, 80, 78, 75, 73, 71, 70, 68, 66, 66, 65, 64, 63, 62, 62,
+        61, 61, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 57, 76, 77, 78, 79, 80,
+        78, 76, 74, 72, 70, 69, 67, 66, 65, 64, 63, 62, 62, 61, 60, 60, 59, 59,
+        59, 58, 58, 58, 57, 57, 57, 57, 57, 73, 74, 76, 77, 78, 76, 74, 72, 70,
+        69, 67, 66, 65, 64, 63, 62, 61, 61, 60, 60, 59, 59, 58, 58, 58, 57, 57,
+        57, 57, 57, 57, 57, 71, 73, 74, 75, 76, 74, 73, 71, 69, 68, 67, 65, 64,
+        63, 62, 62, 61, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56,
+        56, 70, 71, 72, 73, 74, 73, 71, 70, 68, 67, 66, 65, 63, 63, 62, 61, 60,
+        60, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56, 56, 56, 68, 69, 70,
+        71, 72, 71, 70, 68, 67, 66, 65, 64, 63, 62, 61, 60, 60, 59, 59, 58, 58,
+        58, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 66, 67, 68, 69, 70, 69, 68,
+        67, 66, 65, 64, 63, 62, 61, 61, 60, 59, 59, 58, 58, 57, 57, 57, 57, 56,
+        56, 56, 56, 55, 55, 55, 55, 65, 66, 67, 68, 69, 68, 67, 66, 65, 64, 63,
+        62, 61, 61, 60, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56, 55, 55,
+        55, 55, 55, 64, 65, 66, 67, 68, 67, 66, 65, 64, 63, 63, 62, 61, 60, 60,
+        59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 63,
+        64, 64, 65, 66, 65, 65, 64, 63, 63, 62, 61, 60, 60, 59, 59, 58, 58, 57,
+        57, 57, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 62, 62, 63, 64, 65,
+        64, 64, 63, 62, 62, 61, 60, 60, 59, 59, 58, 58, 57, 57, 57, 56, 56, 56,
+        56, 55, 55, 55, 55, 55, 55, 55, 55, 61, 62, 62, 63, 64, 63, 63, 62, 62,
+        61, 61, 60, 59, 59, 58, 58, 57, 57, 57, 56, 56, 56, 56, 55, 55, 55, 55,
+        55, 54, 54, 54, 54, 60, 61, 62, 62, 63, 62, 62, 61, 61, 60, 60, 59, 59,
+        58, 58, 58, 57, 57, 56, 56, 56, 56, 55, 55, 55, 55, 55, 54, 54, 54, 54,
+        54, 59, 60, 61, 61, 62, 62, 61, 61, 60, 60, 59, 59, 58, 58, 58, 57, 57,
+        57, 56, 56, 56, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 59, 59, 60,
+        60, 61, 61, 60, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 55,
+        55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 59, 59, 60, 60, 61, 61, 60,
+        60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 55, 55, 55, 55, 55,
+        54, 54, 54, 54, 54, 54, 54, 59, 59, 60, 60, 61, 61, 60, 60, 60, 59, 59,
+        58, 58, 58, 57, 57, 57, 56, 56, 56, 55, 55, 55, 55, 55, 54, 54, 54, 54,
+        54, 54, 54, 59, 59, 60, 60, 61, 61, 60, 60, 60, 59, 59, 58, 58, 58, 57,
+        57, 57, 56, 56, 56, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 51, 50, 45, 51, 48, 47, 44, 50, 47, 43, 41, 45, 44, 41, 40,
+        /* Size 8 */
+        64, 72, 54, 52, 51, 49, 47, 45, 72, 59, 53, 55, 55, 53, 50, 48, 54, 53,
+        50, 51, 51, 50, 48, 47, 52, 55, 51, 48, 48, 47, 46, 45, 51, 55, 51, 48,
+        46, 45, 44, 43, 49, 53, 50, 47, 45, 44, 43, 42, 47, 50, 48, 46, 44, 43,
+        42, 41, 45, 48, 47, 45, 43, 42, 41, 41,
+        /* Size 16 */
+        64, 68, 72, 63, 54, 53, 52, 52, 51, 50, 49, 48, 47, 46, 45, 45, 68, 67,
+        66, 59, 53, 54, 54, 53, 53, 52, 51, 50, 48, 47, 46, 46, 72, 66, 59, 56,
+        53, 54, 55, 55, 55, 54, 53, 51, 50, 49, 48, 48, 63, 59, 56, 54, 52, 52,
+        53, 53, 53, 52, 51, 50, 49, 48, 47, 47, 54, 53, 53, 52, 50, 50, 51, 51,
+        51, 50, 50, 49, 48, 47, 47, 47, 53, 54, 54, 52, 50, 50, 49, 49, 49, 49,
+        48, 48, 47, 46, 46, 46, 52, 54, 55, 53, 51, 49, 48, 48, 48, 47, 47, 46,
+        46, 45, 45, 45, 52, 53, 55, 53, 51, 49, 48, 47, 47, 46, 46, 45, 45, 45,
+        44, 44, 51, 53, 55, 53, 51, 49, 48, 47, 46, 45, 45, 44, 44, 44, 43, 43,
+        50, 52, 54, 52, 50, 49, 47, 46, 45, 45, 44, 44, 43, 43, 43, 43, 49, 51,
+        53, 51, 50, 48, 47, 46, 45, 44, 44, 43, 43, 42, 42, 42, 48, 50, 51, 50,
+        49, 48, 46, 45, 44, 44, 43, 43, 42, 42, 42, 42, 47, 48, 50, 49, 48, 47,
+        46, 45, 44, 43, 43, 42, 42, 42, 41, 41, 46, 47, 49, 48, 47, 46, 45, 45,
+        44, 43, 42, 42, 42, 41, 41, 41, 45, 46, 48, 47, 47, 46, 45, 44, 43, 43,
+        42, 42, 41, 41, 41, 41, 45, 46, 48, 47, 47, 46, 45, 44, 43, 43, 42, 42,
+        41, 41, 41, 41,
+        /* Size 32 */
+        64, 66, 68, 70, 72, 67, 63, 58, 54, 53, 53, 53, 52, 52, 52, 51, 51, 50,
+        50, 49, 49, 48, 48, 47, 47, 46, 46, 45, 45, 45, 45, 45, 66, 67, 67, 68,
+        69, 65, 61, 57, 53, 53, 53, 53, 53, 53, 52, 52, 52, 51, 51, 50, 50, 49,
+        49, 48, 48, 47, 47, 46, 46, 46, 46, 46, 68, 67, 67, 66, 66, 63, 59, 56,
+        53, 53, 54, 54, 54, 54, 53, 53, 53, 52, 52, 51, 51, 50, 50, 49, 48, 48,
+        47, 47, 46, 46, 46, 46, 70, 68, 66, 64, 62, 60, 58, 56, 53, 54, 54, 54,
+        55, 54, 54, 54, 54, 53, 53, 52, 52, 51, 50, 50, 49, 49, 48, 48, 47, 47,
+        47, 47, 72, 69, 66, 62, 59, 58, 56, 55, 53, 54, 54, 55, 55, 55, 55, 55,
+        55, 54, 54, 53, 53, 52, 51, 51, 50, 50, 49, 48, 48, 48, 48, 48, 67, 65,
+        63, 60, 58, 56, 55, 54, 52, 53, 53, 54, 54, 54, 54, 54, 54, 53, 53, 52,
+        52, 51, 51, 50, 50, 49, 49, 48, 48, 48, 48, 48, 63, 61, 59, 58, 56, 55,
+        54, 53, 52, 52, 52, 53, 53, 53, 53, 53, 53, 52, 52, 52, 51, 51, 50, 50,
+        49, 49, 48, 48, 47, 47, 47, 47, 58, 57, 56, 56, 55, 54, 53, 52, 51, 51,
+        51, 52, 52, 52, 52, 52, 52, 51, 51, 51, 50, 50, 50, 49, 49, 48, 48, 47,
+        47, 47, 47, 47, 54, 53, 53, 53, 53, 52, 52, 51, 50, 50, 50, 50, 51, 51,
+        51, 51, 51, 50, 50, 50, 50, 49, 49, 49, 48, 48, 47, 47, 47, 47, 47, 47,
+        53, 53, 53, 54, 54, 53, 52, 51, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50,
+        49, 49, 49, 49, 48, 48, 48, 47, 47, 46, 46, 46, 46, 46, 53, 53, 54, 54,
+        54, 53, 52, 51, 50, 50, 50, 50, 49, 49, 49, 49, 49, 49, 49, 48, 48, 48,
+        48, 47, 47, 47, 46, 46, 46, 46, 46, 46, 53, 53, 54, 54, 55, 54, 53, 52,
+        50, 50, 50, 49, 49, 49, 49, 48, 48, 48, 48, 48, 48, 47, 47, 47, 46, 46,
+        46, 46, 45, 45, 45, 45, 52, 53, 54, 55, 55, 54, 53, 52, 51, 50, 49, 49,
+        48, 48, 48, 48, 48, 47, 47, 47, 47, 47, 46, 46, 46, 46, 45, 45, 45, 45,
+        45, 45, 52, 53, 54, 54, 55, 54, 53, 52, 51, 50, 49, 49, 48, 48, 48, 47,
+        47, 47, 47, 47, 46, 46, 46, 46, 45, 45, 45, 45, 44, 44, 44, 44, 52, 52,
+        53, 54, 55, 54, 53, 52, 51, 50, 49, 49, 48, 48, 47, 47, 47, 46, 46, 46,
+        46, 46, 45, 45, 45, 45, 45, 44, 44, 44, 44, 44, 51, 52, 53, 54, 55, 54,
+        53, 52, 51, 50, 49, 48, 48, 47, 47, 47, 46, 46, 46, 46, 45, 45, 45, 45,
+        45, 44, 44, 44, 44, 44, 44, 44, 51, 52, 53, 54, 55, 54, 53, 52, 51, 50,
+        49, 48, 48, 47, 47, 46, 46, 46, 45, 45, 45, 45, 44, 44, 44, 44, 44, 44,
+        43, 43, 43, 43, 50, 51, 52, 53, 54, 53, 52, 51, 50, 50, 49, 48, 47, 47,
+        46, 46, 46, 45, 45, 45, 45, 44, 44, 44, 44, 44, 43, 43, 43, 43, 43, 43,
+        50, 51, 52, 53, 54, 53, 52, 51, 50, 49, 49, 48, 47, 47, 46, 46, 45, 45,
+        45, 44, 44, 44, 44, 44, 43, 43, 43, 43, 43, 43, 43, 43, 49, 50, 51, 52,
+        53, 52, 52, 51, 50, 49, 48, 48, 47, 47, 46, 46, 45, 45, 44, 44, 44, 44,
+        43, 43, 43, 43, 43, 43, 42, 42, 42, 42, 49, 50, 51, 52, 53, 52, 51, 50,
+        50, 49, 48, 48, 47, 46, 46, 45, 45, 45, 44, 44, 44, 43, 43, 43, 43, 43,
+        42, 42, 42, 42, 42, 42, 48, 49, 50, 51, 52, 51, 51, 50, 49, 49, 48, 47,
+        47, 46, 46, 45, 45, 44, 44, 44, 43, 43, 43, 43, 43, 42, 42, 42, 42, 42,
+        42, 42, 48, 49, 50, 50, 51, 51, 50, 50, 49, 48, 48, 47, 46, 46, 45, 45,
+        44, 44, 44, 43, 43, 43, 43, 43, 42, 42, 42, 42, 42, 42, 42, 42, 47, 48,
+        49, 50, 51, 50, 50, 49, 49, 48, 47, 47, 46, 46, 45, 45, 44, 44, 44, 43,
+        43, 43, 43, 42, 42, 42, 42, 42, 41, 41, 41, 41, 47, 48, 48, 49, 50, 50,
+        49, 49, 48, 48, 47, 46, 46, 45, 45, 45, 44, 44, 43, 43, 43, 43, 42, 42,
+        42, 42, 42, 41, 41, 41, 41, 41, 46, 47, 48, 49, 50, 49, 49, 48, 48, 47,
+        47, 46, 46, 45, 45, 44, 44, 44, 43, 43, 43, 42, 42, 42, 42, 42, 41, 41,
+        41, 41, 41, 41, 46, 47, 47, 48, 49, 49, 48, 48, 47, 47, 46, 46, 45, 45,
+        45, 44, 44, 43, 43, 43, 42, 42, 42, 42, 42, 41, 41, 41, 41, 41, 41, 41,
+        45, 46, 47, 48, 48, 48, 48, 47, 47, 46, 46, 46, 45, 45, 44, 44, 44, 43,
+        43, 43, 42, 42, 42, 42, 41, 41, 41, 41, 41, 41, 41, 41, 45, 46, 46, 47,
+        48, 48, 47, 47, 47, 46, 46, 45, 45, 44, 44, 44, 43, 43, 43, 42, 42, 42,
+        42, 41, 41, 41, 41, 41, 41, 41, 41, 41, 45, 46, 46, 47, 48, 48, 47, 47,
+        47, 46, 46, 45, 45, 44, 44, 44, 43, 43, 43, 42, 42, 42, 42, 41, 41, 41,
+        41, 41, 41, 41, 41, 41, 45, 46, 46, 47, 48, 48, 47, 47, 47, 46, 46, 45,
+        45, 44, 44, 44, 43, 43, 43, 42, 42, 42, 42, 41, 41, 41, 41, 41, 41, 41,
+        41, 41, 45, 46, 46, 47, 48, 48, 47, 47, 47, 46, 46, 45, 45, 44, 44, 44,
+        43, 43, 43, 42, 42, 42, 42, 41, 41, 41, 41, 41, 41, 41, 41, 41 },
+      { /* Intra matrices */
+        /* Size 4 */
+        90, 72, 69, 63, 72, 66, 65, 62, 69, 65, 59, 57, 63, 62, 57, 54,
+        /* Size 8 */
+        87, 98, 72, 70, 68, 65, 62, 60, 98, 80, 71, 74, 74, 70, 67, 64, 72, 71,
+        67, 68, 68, 66, 64, 62, 70, 74, 68, 64, 63, 62, 61, 59, 68, 74, 68, 63,
+        61, 59, 58, 57, 65, 70, 66, 62, 59, 58, 57, 56, 62, 67, 64, 61, 58, 57,
+        55, 54, 60, 64, 62, 59, 57, 56, 54, 53,
+        /* Size 16 */
+        88, 93, 99, 86, 73, 72, 71, 70, 69, 67, 66, 64, 63, 62, 60, 60, 93, 92,
+        90, 81, 72, 73, 73, 72, 72, 70, 69, 67, 65, 64, 62, 62, 99, 90, 81, 76,
+        72, 74, 75, 75, 74, 73, 71, 69, 68, 66, 65, 65, 86, 81, 76, 73, 70, 71,
+        72, 72, 71, 70, 69, 68, 66, 65, 64, 64, 73, 72, 72, 70, 67, 68, 68, 68,
+        68, 68, 67, 66, 65, 64, 63, 63, 72, 73, 74, 71, 68, 67, 67, 66, 66, 66,
+        65, 64, 63, 62, 61, 61, 71, 73, 75, 72, 68, 67, 65, 65, 64, 63, 63, 62,
+        62, 61, 60, 60, 70, 72, 75, 72, 68, 66, 65, 64, 63, 62, 62, 61, 60, 60,
+        59, 59, 69, 72, 74, 71, 68, 66, 64, 63, 61, 61, 60, 60, 59, 58, 58, 58,
+        67, 70, 73, 70, 68, 66, 63, 62, 61, 60, 59, 59, 58, 58, 57, 57, 66, 69,
+        71, 69, 67, 65, 63, 62, 60, 59, 58, 58, 57, 57, 56, 56, 64, 67, 69, 68,
+        66, 64, 62, 61, 60, 59, 58, 57, 56, 56, 56, 56, 63, 65, 68, 66, 65, 63,
+        62, 60, 59, 58, 57, 56, 56, 55, 55, 55, 62, 64, 66, 65, 64, 62, 61, 60,
+        58, 58, 57, 56, 55, 55, 54, 54, 60, 62, 65, 64, 63, 61, 60, 59, 58, 57,
+        56, 56, 55, 54, 54, 54, 60, 62, 65, 64, 63, 61, 60, 59, 58, 57, 56, 56,
+        55, 54, 54, 54,
+        /* Size 32 */
+        88, 91, 94, 97, 100, 93, 86, 80, 73, 73, 72, 72, 71, 71, 70, 70, 69, 68,
+        68, 67, 66, 65, 65, 64, 63, 63, 62, 61, 61, 61, 61, 61, 91, 92, 93, 94,
+        95, 89, 84, 78, 73, 73, 73, 72, 72, 72, 71, 71, 71, 70, 69, 68, 68, 67,
+        66, 65, 64, 64, 63, 62, 62, 62, 62, 62, 94, 93, 92, 91, 90, 86, 82, 77,
+        73, 73, 73, 73, 73, 73, 73, 72, 72, 71, 70, 70, 69, 68, 67, 66, 66, 65,
+        64, 63, 63, 63, 63, 63, 97, 94, 91, 89, 86, 83, 79, 76, 73, 73, 74, 74,
+        75, 74, 74, 74, 73, 73, 72, 71, 70, 69, 69, 68, 67, 66, 65, 65, 64, 64,
+        64, 64, 100, 95, 90, 86, 81, 79, 77, 75, 72, 73, 74, 75, 76, 75, 75, 75,
+        75, 74, 73, 72, 72, 71, 70, 69, 68, 67, 66, 66, 65, 65, 65, 65, 93, 89,
+        86, 83, 79, 77, 75, 73, 71, 72, 73, 73, 74, 74, 74, 73, 73, 73, 72, 71,
+        71, 70, 69, 68, 67, 67, 66, 65, 64, 64, 64, 64, 86, 84, 82, 79, 77, 75,
+        73, 72, 70, 71, 71, 72, 72, 72, 72, 72, 72, 71, 71, 70, 70, 69, 68, 67,
+        67, 66, 65, 65, 64, 64, 64, 64, 80, 78, 77, 76, 75, 73, 72, 70, 69, 69,
+        70, 70, 70, 70, 70, 70, 70, 70, 69, 69, 68, 68, 67, 67, 66, 65, 65, 64,
+        63, 63, 63, 63, 73, 73, 73, 73, 72, 71, 70, 69, 68, 68, 68, 68, 69, 69,
+        69, 69, 69, 68, 68, 68, 67, 67, 66, 66, 65, 65, 64, 63, 63, 63, 63, 63,
+        73, 73, 73, 73, 73, 72, 71, 69, 68, 68, 68, 68, 68, 68, 68, 68, 68, 67,
+        67, 67, 66, 66, 65, 65, 64, 64, 63, 63, 62, 62, 62, 62, 72, 73, 73, 74,
+        74, 73, 71, 70, 68, 68, 68, 67, 67, 67, 67, 67, 67, 66, 66, 66, 65, 65,
+        64, 64, 64, 63, 63, 62, 62, 62, 62, 62, 72, 72, 73, 74, 75, 73, 72, 70,
+        68, 68, 67, 67, 66, 66, 66, 66, 65, 65, 65, 65, 64, 64, 64, 63, 63, 62,
+        62, 62, 61, 61, 61, 61, 71, 72, 73, 75, 76, 74, 72, 70, 69, 68, 67, 66,
+        65, 65, 65, 65, 64, 64, 64, 64, 63, 63, 63, 62, 62, 62, 61, 61, 60, 60,
+        60, 60, 71, 72, 73, 74, 75, 74, 72, 70, 69, 68, 67, 66, 65, 65, 64, 64,
+        64, 63, 63, 63, 63, 62, 62, 62, 61, 61, 61, 60, 60, 60, 60, 60, 70, 71,
+        73, 74, 75, 74, 72, 70, 69, 68, 67, 66, 65, 64, 64, 64, 63, 63, 62, 62,
+        62, 62, 61, 61, 61, 60, 60, 60, 59, 59, 59, 59, 70, 71, 72, 74, 75, 73,
+        72, 70, 69, 68, 67, 66, 65, 64, 64, 63, 62, 62, 62, 61, 61, 61, 61, 60,
+        60, 60, 59, 59, 59, 59, 59, 59, 69, 71, 72, 73, 75, 73, 72, 70, 69, 68,
+        67, 65, 64, 64, 63, 62, 62, 61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59,
+        58, 58, 58, 58, 68, 70, 71, 73, 74, 73, 71, 70, 68, 67, 66, 65, 64, 63,
+        63, 62, 61, 61, 61, 60, 60, 60, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58,
+        68, 69, 70, 72, 73, 72, 71, 69, 68, 67, 66, 65, 64, 63, 62, 62, 61, 61,
+        60, 60, 60, 59, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 67, 68, 70, 71,
+        72, 71, 70, 69, 68, 67, 66, 65, 64, 63, 62, 61, 61, 60, 60, 60, 59, 59,
+        58, 58, 58, 58, 57, 57, 57, 57, 57, 57, 66, 68, 69, 70, 72, 71, 70, 68,
+        67, 66, 65, 64, 63, 63, 62, 61, 60, 60, 60, 59, 59, 58, 58, 58, 57, 57,
+        57, 57, 57, 57, 57, 57, 65, 67, 68, 69, 71, 70, 69, 68, 67, 66, 65, 64,
+        63, 62, 62, 61, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56,
+        56, 56, 65, 66, 67, 69, 70, 69, 68, 67, 66, 65, 64, 64, 63, 62, 61, 61,
+        60, 59, 59, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 64, 65,
+        66, 68, 69, 68, 67, 67, 66, 65, 64, 63, 62, 62, 61, 60, 60, 59, 59, 58,
+        58, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 56, 63, 64, 66, 67, 68, 67,
+        67, 66, 65, 64, 64, 63, 62, 61, 61, 60, 59, 59, 58, 58, 57, 57, 57, 56,
+        56, 56, 56, 55, 55, 55, 55, 55, 63, 64, 65, 66, 67, 67, 66, 65, 65, 64,
+        63, 62, 62, 61, 60, 60, 59, 59, 58, 58, 57, 57, 57, 56, 56, 56, 55, 55,
+        55, 55, 55, 55, 62, 63, 64, 65, 66, 66, 65, 65, 64, 63, 63, 62, 61, 61,
+        60, 59, 59, 58, 58, 57, 57, 57, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55,
+        61, 62, 63, 65, 66, 65, 65, 64, 63, 63, 62, 62, 61, 60, 60, 59, 59, 58,
+        58, 57, 57, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 61, 62, 63, 64,
+        65, 64, 64, 63, 63, 62, 62, 61, 60, 60, 59, 59, 58, 58, 57, 57, 57, 56,
+        56, 56, 55, 55, 55, 55, 54, 54, 54, 54, 61, 62, 63, 64, 65, 64, 64, 63,
+        63, 62, 62, 61, 60, 60, 59, 59, 58, 58, 57, 57, 57, 56, 56, 56, 55, 55,
+        55, 55, 54, 54, 54, 54, 61, 62, 63, 64, 65, 64, 64, 63, 63, 62, 62, 61,
+        60, 60, 59, 59, 58, 58, 57, 57, 57, 56, 56, 56, 55, 55, 55, 55, 54, 54,
+        54, 54, 61, 62, 63, 64, 65, 64, 64, 63, 63, 62, 62, 61, 60, 60, 59, 59,
+        58, 58, 57, 57, 57, 56, 56, 56, 55, 55, 55, 55, 54, 54, 54, 54 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 62, 51, 46, 62, 53, 48, 45, 51, 48, 45, 43, 46, 45, 43, 42,
+        /* Size 8 */
+        64, 72, 70, 62, 56, 52, 50, 48, 72, 68, 69, 65, 59, 55, 52, 49, 70, 69,
+        61, 57, 55, 52, 50, 49, 62, 65, 57, 53, 51, 50, 49, 48, 56, 59, 55, 51,
+        49, 48, 47, 47, 52, 55, 52, 50, 48, 47, 47, 46, 50, 52, 50, 49, 47, 47,
+        46, 46, 48, 49, 49, 48, 47, 46, 46, 45,
+        /* Size 16 */
+        64, 68, 72, 71, 70, 66, 62, 59, 56, 54, 52, 51, 50, 49, 48, 48, 68, 69,
+        70, 70, 70, 67, 64, 61, 58, 56, 54, 52, 51, 50, 49, 49, 72, 70, 68, 69,
+        69, 67, 65, 62, 59, 57, 55, 53, 52, 50, 49, 49, 71, 70, 69, 67, 65, 63,
+        61, 59, 57, 55, 53, 52, 51, 50, 49, 49, 70, 70, 69, 65, 61, 59, 57, 56,
+        55, 53, 52, 51, 50, 49, 49, 49, 66, 67, 67, 63, 59, 57, 55, 54, 53, 52,
+        51, 50, 49, 49, 48, 48, 62, 64, 65, 61, 57, 55, 53, 52, 51, 51, 50, 49,
+        49, 48, 48, 48, 59, 61, 62, 59, 56, 54, 52, 51, 50, 50, 49, 49, 48, 48,
+        47, 47, 56, 58, 59, 57, 55, 53, 51, 50, 49, 49, 48, 48, 47, 47, 47, 47,
+        54, 56, 57, 55, 53, 52, 51, 50, 49, 48, 48, 47, 47, 47, 46, 46, 52, 54,
+        55, 53, 52, 51, 50, 49, 48, 48, 47, 47, 47, 46, 46, 46, 51, 52, 53, 52,
+        51, 50, 49, 49, 48, 47, 47, 47, 46, 46, 46, 46, 50, 51, 52, 51, 50, 49,
+        49, 48, 47, 47, 47, 46, 46, 46, 46, 46, 49, 50, 50, 50, 49, 49, 48, 48,
+        47, 47, 46, 46, 46, 46, 45, 45, 48, 49, 49, 49, 49, 48, 48, 47, 47, 46,
+        46, 46, 46, 45, 45, 45, 48, 49, 49, 49, 49, 48, 48, 47, 47, 46, 46, 46,
+        46, 45, 45, 45,
+        /* Size 32 */
+        64, 66, 68, 70, 72, 72, 71, 70, 70, 68, 66, 64, 62, 61, 59, 58, 56, 55,
+        54, 53, 52, 52, 51, 50, 50, 49, 49, 48, 48, 48, 48, 48, 66, 67, 69, 70,
+        71, 71, 70, 70, 70, 68, 66, 65, 63, 61, 60, 59, 57, 56, 55, 54, 53, 52,
+        52, 51, 50, 50, 49, 49, 48, 48, 48, 48, 68, 69, 69, 70, 70, 70, 70, 70,
+        70, 68, 67, 65, 64, 62, 61, 59, 58, 57, 56, 55, 54, 53, 52, 51, 51, 50,
+        50, 49, 49, 49, 49, 49, 70, 70, 70, 69, 69, 69, 69, 69, 69, 68, 67, 66,
+        64, 63, 61, 60, 58, 57, 56, 55, 54, 53, 53, 52, 51, 51, 50, 49, 49, 49,
+        49, 49, 72, 71, 70, 69, 68, 68, 69, 69, 69, 68, 67, 66, 65, 63, 62, 61,
+        59, 58, 57, 56, 55, 54, 53, 52, 52, 51, 50, 50, 49, 49, 49, 49, 72, 71,
+        70, 69, 68, 68, 68, 67, 67, 66, 65, 64, 63, 62, 61, 59, 58, 57, 56, 55,
+        54, 53, 53, 52, 51, 51, 50, 50, 49, 49, 49, 49, 71, 70, 70, 69, 69, 68,
+        67, 66, 65, 64, 63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 53, 53, 52, 51,
+        51, 50, 50, 49, 49, 49, 49, 49, 70, 70, 70, 69, 69, 67, 66, 64, 63, 62,
+        61, 60, 59, 58, 58, 57, 56, 55, 54, 54, 53, 52, 52, 51, 50, 50, 50, 49,
+        49, 49, 49, 49, 70, 70, 70, 69, 69, 67, 65, 63, 61, 60, 59, 58, 57, 57,
+        56, 55, 55, 54, 53, 53, 52, 52, 51, 51, 50, 50, 49, 49, 49, 49, 49, 49,
+        68, 68, 68, 68, 68, 66, 64, 62, 60, 59, 58, 57, 56, 56, 55, 55, 54, 53,
+        53, 52, 52, 51, 51, 50, 50, 49, 49, 49, 48, 48, 48, 48, 66, 66, 67, 67,
+        67, 65, 63, 61, 59, 58, 57, 56, 55, 55, 54, 54, 53, 53, 52, 52, 51, 51,
+        50, 50, 49, 49, 49, 48, 48, 48, 48, 48, 64, 65, 65, 66, 66, 64, 62, 60,
+        58, 57, 56, 55, 54, 54, 53, 53, 52, 52, 51, 51, 50, 50, 50, 49, 49, 49,
+        48, 48, 48, 48, 48, 48, 62, 63, 64, 64, 65, 63, 61, 59, 57, 56, 55, 54,
+        53, 53, 52, 52, 51, 51, 51, 50, 50, 50, 49, 49, 49, 48, 48, 48, 48, 48,
+        48, 48, 61, 61, 62, 63, 63, 62, 60, 58, 57, 56, 55, 54, 53, 52, 52, 51,
+        51, 51, 50, 50, 49, 49, 49, 49, 48, 48, 48, 48, 47, 47, 47, 47, 59, 60,
+        61, 61, 62, 61, 59, 58, 56, 55, 54, 53, 52, 52, 51, 51, 50, 50, 50, 49,
+        49, 49, 49, 48, 48, 48, 48, 47, 47, 47, 47, 47, 58, 59, 59, 60, 61, 59,
+        58, 57, 55, 55, 54, 53, 52, 51, 51, 50, 50, 50, 49, 49, 49, 48, 48, 48,
+        48, 48, 47, 47, 47, 47, 47, 47, 56, 57, 58, 58, 59, 58, 57, 56, 55, 54,
+        53, 52, 51, 51, 50, 50, 49, 49, 49, 49, 48, 48, 48, 48, 47, 47, 47, 47,
+        47, 47, 47, 47, 55, 56, 57, 57, 58, 57, 56, 55, 54, 53, 53, 52, 51, 51,
+        50, 50, 49, 49, 49, 48, 48, 48, 48, 47, 47, 47, 47, 47, 47, 47, 47, 47,
+        54, 55, 56, 56, 57, 56, 55, 54, 53, 53, 52, 51, 51, 50, 50, 49, 49, 49,
+        48, 48, 48, 48, 47, 47, 47, 47, 47, 47, 46, 46, 46, 46, 53, 54, 55, 55,
+        56, 55, 54, 54, 53, 52, 52, 51, 50, 50, 49, 49, 49, 48, 48, 48, 48, 47,
+        47, 47, 47, 47, 47, 46, 46, 46, 46, 46, 52, 53, 54, 54, 55, 54, 53, 53,
+        52, 52, 51, 50, 50, 49, 49, 49, 48, 48, 48, 48, 47, 47, 47, 47, 47, 46,
+        46, 46, 46, 46, 46, 46, 52, 52, 53, 53, 54, 53, 53, 52, 52, 51, 51, 50,
+        50, 49, 49, 48, 48, 48, 48, 47, 47, 47, 47, 47, 46, 46, 46, 46, 46, 46,
+        46, 46, 51, 52, 52, 53, 53, 53, 52, 52, 51, 51, 50, 50, 49, 49, 49, 48,
+        48, 48, 47, 47, 47, 47, 47, 46, 46, 46, 46, 46, 46, 46, 46, 46, 50, 51,
+        51, 52, 52, 52, 51, 51, 51, 50, 50, 49, 49, 49, 48, 48, 48, 47, 47, 47,
+        47, 47, 46, 46, 46, 46, 46, 46, 46, 46, 46, 46, 50, 50, 51, 51, 52, 51,
+        51, 50, 50, 50, 49, 49, 49, 48, 48, 48, 47, 47, 47, 47, 47, 46, 46, 46,
+        46, 46, 46, 46, 46, 46, 46, 46, 49, 50, 50, 51, 51, 51, 50, 50, 50, 49,
+        49, 49, 48, 48, 48, 48, 47, 47, 47, 47, 46, 46, 46, 46, 46, 46, 46, 46,
+        46, 46, 46, 46, 49, 49, 50, 50, 50, 50, 50, 50, 49, 49, 49, 48, 48, 48,
+        48, 47, 47, 47, 47, 47, 46, 46, 46, 46, 46, 46, 46, 46, 45, 45, 45, 45,
+        48, 49, 49, 49, 50, 50, 49, 49, 49, 49, 48, 48, 48, 48, 47, 47, 47, 47,
+        47, 46, 46, 46, 46, 46, 46, 46, 46, 45, 45, 45, 45, 45, 48, 48, 49, 49,
+        49, 49, 49, 49, 49, 48, 48, 48, 48, 47, 47, 47, 47, 47, 46, 46, 46, 46,
+        46, 46, 46, 46, 45, 45, 45, 45, 45, 45, 48, 48, 49, 49, 49, 49, 49, 49,
+        49, 48, 48, 48, 48, 47, 47, 47, 47, 47, 46, 46, 46, 46, 46, 46, 46, 46,
+        45, 45, 45, 45, 45, 45, 48, 48, 49, 49, 49, 49, 49, 49, 49, 48, 48, 48,
+        48, 47, 47, 47, 47, 47, 46, 46, 46, 46, 46, 46, 46, 46, 45, 45, 45, 45,
+        45, 45, 48, 48, 49, 49, 49, 49, 49, 49, 49, 48, 48, 48, 48, 47, 47, 47,
+        47, 47, 46, 46, 46, 46, 46, 46, 46, 46, 45, 45, 45, 45, 45, 45 },
+      { /* Intra matrices */
+        /* Size 4 */
+        86, 83, 68, 60, 83, 70, 64, 59, 68, 64, 59, 57, 60, 59, 57, 55,
+        /* Size 8 */
+        79, 90, 87, 77, 69, 64, 61, 58, 90, 85, 86, 80, 73, 67, 63, 60, 87, 86,
+        75, 71, 67, 64, 61, 59, 77, 80, 71, 66, 63, 61, 59, 58, 69, 73, 67, 63,
+        60, 59, 58, 57, 64, 67, 64, 61, 59, 57, 57, 56, 61, 63, 61, 59, 58, 57,
+        56, 55, 58, 60, 59, 58, 57, 56, 55, 55,
+        /* Size 16 */
+        80, 86, 91, 90, 88, 83, 78, 74, 70, 68, 65, 63, 61, 60, 59, 59, 86, 87,
+        89, 88, 88, 84, 80, 76, 72, 69, 66, 65, 63, 61, 60, 60, 91, 89, 86, 87,
+        87, 84, 82, 78, 74, 71, 68, 66, 64, 62, 61, 61, 90, 88, 87, 84, 82, 79,
+        77, 74, 71, 69, 66, 65, 63, 62, 60, 60, 88, 88, 87, 82, 76, 74, 72, 70,
+        68, 66, 65, 63, 62, 61, 60, 60, 83, 84, 84, 79, 74, 71, 69, 68, 66, 65,
+        63, 62, 61, 60, 59, 59, 78, 80, 82, 77, 72, 69, 66, 65, 64, 63, 62, 61,
+        60, 59, 59, 59, 74, 76, 78, 74, 70, 68, 65, 64, 62, 62, 61, 60, 59, 59,
+        58, 58, 70, 72, 74, 71, 68, 66, 64, 62, 61, 60, 60, 59, 58, 58, 57, 57,
+        68, 69, 71, 69, 66, 65, 63, 62, 60, 60, 59, 58, 58, 57, 57, 57, 65, 66,
+        68, 66, 65, 63, 62, 61, 60, 59, 58, 58, 57, 57, 57, 57, 63, 65, 66, 65,
+        63, 62, 61, 60, 59, 58, 58, 57, 57, 57, 56, 56, 61, 63, 64, 63, 62, 61,
+        60, 59, 58, 58, 57, 57, 57, 56, 56, 56, 60, 61, 62, 62, 61, 60, 59, 59,
+        58, 57, 57, 57, 56, 56, 56, 56, 59, 60, 61, 60, 60, 59, 59, 58, 57, 57,
+        57, 56, 56, 56, 56, 56, 59, 60, 61, 60, 60, 59, 59, 58, 57, 57, 57, 56,
+        56, 56, 56, 56,
+        /* Size 32 */
+        81, 84, 86, 89, 92, 91, 90, 89, 89, 86, 84, 81, 79, 77, 75, 73, 71, 69,
+        68, 67, 65, 64, 64, 63, 62, 61, 61, 60, 59, 59, 59, 59, 84, 85, 87, 89,
+        91, 90, 89, 89, 88, 86, 84, 82, 79, 78, 76, 74, 72, 70, 69, 67, 66, 65,
+        64, 63, 62, 62, 61, 61, 60, 60, 60, 60, 86, 87, 88, 88, 89, 89, 89, 88,
+        88, 86, 84, 82, 80, 78, 76, 75, 73, 71, 70, 68, 67, 66, 65, 64, 63, 62,
+        62, 61, 60, 60, 60, 60, 89, 89, 88, 88, 88, 88, 88, 88, 88, 86, 85, 83,
+        81, 79, 77, 75, 74, 72, 71, 69, 68, 67, 66, 65, 64, 63, 62, 62, 61, 61,
+        61, 61, 92, 91, 89, 88, 86, 87, 87, 88, 88, 86, 85, 83, 82, 80, 78, 76,
+        74, 73, 71, 70, 68, 67, 66, 65, 64, 64, 63, 62, 61, 61, 61, 61, 91, 90,
+        89, 88, 87, 86, 86, 85, 85, 84, 82, 81, 80, 78, 76, 75, 73, 72, 70, 69,
+        68, 67, 66, 65, 64, 63, 62, 62, 61, 61, 61, 61, 90, 89, 89, 88, 87, 86,
+        85, 83, 82, 81, 80, 78, 77, 76, 74, 73, 72, 70, 69, 68, 67, 66, 65, 64,
+        63, 63, 62, 61, 61, 61, 61, 61, 89, 89, 88, 88, 88, 85, 83, 81, 79, 78,
+        77, 76, 75, 73, 72, 71, 70, 69, 68, 67, 66, 65, 64, 64, 63, 62, 62, 61,
+        61, 61, 61, 61, 89, 88, 88, 88, 88, 85, 82, 79, 76, 75, 74, 73, 72, 71,
+        70, 69, 69, 68, 67, 66, 65, 64, 64, 63, 62, 62, 61, 61, 60, 60, 60, 60,
+        86, 86, 86, 86, 86, 84, 81, 78, 75, 74, 73, 72, 71, 70, 69, 68, 67, 67,
+        66, 65, 64, 64, 63, 63, 62, 61, 61, 60, 60, 60, 60, 60, 84, 84, 84, 85,
+        85, 82, 80, 77, 74, 73, 72, 71, 70, 69, 68, 67, 66, 66, 65, 64, 64, 63,
+        63, 62, 61, 61, 61, 60, 60, 60, 60, 60, 81, 82, 82, 83, 83, 81, 78, 76,
+        73, 72, 71, 69, 68, 67, 67, 66, 65, 65, 64, 63, 63, 62, 62, 61, 61, 60,
+        60, 60, 59, 59, 59, 59, 79, 79, 80, 81, 82, 80, 77, 75, 72, 71, 70, 68,
+        67, 66, 65, 65, 64, 64, 63, 63, 62, 62, 61, 61, 60, 60, 60, 59, 59, 59,
+        59, 59, 77, 78, 78, 79, 80, 78, 76, 73, 71, 70, 69, 67, 66, 66, 65, 64,
+        63, 63, 63, 62, 62, 61, 61, 60, 60, 60, 59, 59, 59, 59, 59, 59, 75, 76,
+        76, 77, 78, 76, 74, 72, 70, 69, 68, 67, 65, 65, 64, 63, 63, 62, 62, 61,
+        61, 61, 60, 60, 60, 59, 59, 59, 58, 58, 58, 58, 73, 74, 75, 75, 76, 75,
+        73, 71, 69, 68, 67, 66, 65, 64, 63, 63, 62, 62, 61, 61, 60, 60, 60, 60,
+        59, 59, 59, 58, 58, 58, 58, 58, 71, 72, 73, 74, 74, 73, 72, 70, 69, 67,
+        66, 65, 64, 63, 63, 62, 62, 61, 61, 60, 60, 60, 59, 59, 59, 59, 58, 58,
+        58, 58, 58, 58, 69, 70, 71, 72, 73, 72, 70, 69, 68, 67, 66, 65, 64, 63,
+        62, 62, 61, 61, 60, 60, 60, 59, 59, 59, 58, 58, 58, 58, 58, 58, 58, 58,
+        68, 69, 70, 71, 71, 70, 69, 68, 67, 66, 65, 64, 63, 63, 62, 61, 61, 60,
+        60, 60, 59, 59, 59, 58, 58, 58, 58, 58, 57, 57, 57, 57, 67, 67, 68, 69,
+        70, 69, 68, 67, 66, 65, 64, 63, 63, 62, 61, 61, 60, 60, 60, 59, 59, 59,
+        58, 58, 58, 58, 58, 57, 57, 57, 57, 57, 65, 66, 67, 68, 68, 68, 67, 66,
+        65, 64, 64, 63, 62, 62, 61, 60, 60, 60, 59, 59, 59, 58, 58, 58, 58, 57,
+        57, 57, 57, 57, 57, 57, 64, 65, 66, 67, 67, 67, 66, 65, 64, 64, 63, 62,
+        62, 61, 61, 60, 60, 59, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 57, 57,
+        57, 57, 64, 64, 65, 66, 66, 66, 65, 64, 64, 63, 63, 62, 61, 61, 60, 60,
+        59, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 57, 57, 57, 57, 57, 63, 63,
+        64, 65, 65, 65, 64, 64, 63, 63, 62, 61, 61, 60, 60, 60, 59, 59, 58, 58,
+        58, 58, 57, 57, 57, 57, 57, 57, 57, 57, 57, 57, 62, 62, 63, 64, 64, 64,
+        63, 63, 62, 62, 61, 61, 60, 60, 60, 59, 59, 58, 58, 58, 58, 57, 57, 57,
+        57, 57, 57, 56, 56, 56, 56, 56, 61, 62, 62, 63, 64, 63, 63, 62, 62, 61,
+        61, 60, 60, 60, 59, 59, 59, 58, 58, 58, 57, 57, 57, 57, 57, 57, 56, 56,
+        56, 56, 56, 56, 61, 61, 62, 62, 63, 62, 62, 62, 61, 61, 61, 60, 60, 59,
+        59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56,
+        60, 61, 61, 62, 62, 62, 61, 61, 61, 60, 60, 60, 59, 59, 59, 58, 58, 58,
+        58, 57, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 56, 59, 60, 60, 61,
+        61, 61, 61, 61, 60, 60, 60, 59, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57,
+        57, 57, 56, 56, 56, 56, 56, 56, 56, 56, 59, 60, 60, 61, 61, 61, 61, 61,
+        60, 60, 60, 59, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 57, 57, 56, 56,
+        56, 56, 56, 56, 56, 56, 59, 60, 60, 61, 61, 61, 61, 61, 60, 60, 60, 59,
+        59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56,
+        56, 56, 59, 60, 60, 61, 61, 61, 61, 61, 60, 60, 60, 59, 59, 59, 58, 58,
+        58, 58, 57, 57, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 56 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 53, 52, 48, 53, 50, 49, 48, 52, 49, 46, 45, 48, 48, 45, 43,
+        /* Size 8 */
+        64, 71, 55, 54, 53, 51, 50, 48, 71, 60, 55, 57, 56, 54, 52, 51, 55, 55,
+        52, 53, 53, 52, 51, 49, 54, 57, 53, 51, 50, 50, 49, 48, 53, 56, 53, 50,
+        49, 48, 47, 47, 51, 54, 52, 50, 48, 47, 46, 46, 50, 52, 51, 49, 47, 46,
+        46, 45, 48, 51, 49, 48, 47, 46, 45, 44,
+        /* Size 16 */
+        64, 67, 71, 63, 55, 55, 54, 54, 53, 52, 51, 50, 50, 49, 48, 48, 67, 66,
+        65, 60, 55, 55, 56, 55, 55, 54, 53, 52, 51, 50, 49, 49, 71, 65, 60, 58,
+        55, 56, 57, 57, 56, 55, 54, 53, 52, 51, 51, 51, 63, 60, 58, 56, 54, 54,
+        55, 55, 55, 54, 53, 52, 52, 51, 50, 50, 55, 55, 55, 54, 52, 53, 53, 53,
+        53, 52, 52, 51, 51, 50, 49, 49, 55, 55, 56, 54, 53, 52, 52, 52, 52, 51,
+        51, 50, 50, 49, 49, 49, 54, 56, 57, 55, 53, 52, 51, 51, 50, 50, 50, 49,
+        49, 48, 48, 48, 54, 55, 57, 55, 53, 52, 51, 50, 50, 49, 49, 48, 48, 48,
+        47, 47, 53, 55, 56, 55, 53, 52, 50, 50, 49, 48, 48, 48, 47, 47, 47, 47,
+        52, 54, 55, 54, 52, 51, 50, 49, 48, 48, 47, 47, 47, 47, 46, 46, 51, 53,
+        54, 53, 52, 51, 50, 49, 48, 47, 47, 47, 46, 46, 46, 46, 50, 52, 53, 52,
+        51, 50, 49, 48, 48, 47, 47, 46, 46, 46, 45, 45, 50, 51, 52, 52, 51, 50,
+        49, 48, 47, 47, 46, 46, 46, 45, 45, 45, 49, 50, 51, 51, 50, 49, 48, 48,
+        47, 47, 46, 46, 45, 45, 45, 45, 48, 49, 51, 50, 49, 49, 48, 47, 47, 46,
+        46, 45, 45, 45, 44, 44, 48, 49, 51, 50, 49, 49, 48, 47, 47, 46, 46, 45,
+        45, 45, 44, 44,
+        /* Size 32 */
+        64, 66, 67, 69, 71, 67, 63, 59, 55, 55, 55, 54, 54, 54, 54, 53, 53, 53,
+        52, 52, 51, 51, 50, 50, 50, 49, 49, 48, 48, 48, 48, 48, 66, 66, 67, 67,
+        68, 65, 62, 58, 55, 55, 55, 55, 55, 55, 54, 54, 54, 53, 53, 53, 52, 52,
+        51, 51, 50, 50, 50, 49, 49, 49, 49, 49, 67, 67, 66, 66, 65, 63, 60, 58,
+        55, 55, 55, 55, 56, 55, 55, 55, 55, 54, 54, 53, 53, 52, 52, 51, 51, 51,
+        50, 50, 49, 49, 49, 49, 69, 67, 66, 64, 63, 61, 59, 57, 55, 55, 56, 56,
+        56, 56, 56, 56, 55, 55, 55, 54, 54, 53, 53, 52, 52, 51, 51, 50, 50, 50,
+        50, 50, 71, 68, 65, 63, 60, 59, 58, 56, 55, 55, 56, 56, 57, 57, 57, 56,
+        56, 56, 55, 55, 54, 54, 53, 53, 52, 52, 51, 51, 51, 51, 51, 51, 67, 65,
+        63, 61, 59, 58, 57, 55, 54, 55, 55, 55, 56, 56, 56, 56, 55, 55, 55, 54,
+        54, 53, 53, 52, 52, 52, 51, 51, 50, 50, 50, 50, 63, 62, 60, 59, 58, 57,
+        56, 55, 54, 54, 54, 54, 55, 55, 55, 55, 55, 54, 54, 54, 53, 53, 52, 52,
+        52, 51, 51, 50, 50, 50, 50, 50, 59, 58, 58, 57, 56, 55, 55, 54, 53, 53,
+        53, 54, 54, 54, 54, 54, 54, 53, 53, 53, 53, 52, 52, 52, 51, 51, 50, 50,
+        50, 50, 50, 50, 55, 55, 55, 55, 55, 54, 54, 53, 52, 52, 53, 53, 53, 53,
+        53, 53, 53, 53, 52, 52, 52, 52, 51, 51, 51, 50, 50, 50, 49, 49, 49, 49,
+        55, 55, 55, 55, 55, 55, 54, 53, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
+        52, 52, 51, 51, 51, 51, 50, 50, 50, 49, 49, 49, 49, 49, 55, 55, 55, 56,
+        56, 55, 54, 53, 53, 52, 52, 52, 52, 52, 52, 52, 52, 51, 51, 51, 51, 51,
+        50, 50, 50, 50, 49, 49, 49, 49, 49, 49, 54, 55, 55, 56, 56, 55, 54, 54,
+        53, 52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 50, 50, 50, 50, 50, 49, 49,
+        49, 49, 48, 48, 48, 48, 54, 55, 56, 56, 57, 56, 55, 54, 53, 52, 52, 51,
+        51, 51, 51, 50, 50, 50, 50, 50, 50, 49, 49, 49, 49, 49, 48, 48, 48, 48,
+        48, 48, 54, 55, 55, 56, 57, 56, 55, 54, 53, 52, 52, 51, 51, 51, 50, 50,
+        50, 50, 50, 49, 49, 49, 49, 49, 49, 48, 48, 48, 48, 48, 48, 48, 54, 54,
+        55, 56, 57, 56, 55, 54, 53, 52, 52, 51, 51, 50, 50, 50, 50, 49, 49, 49,
+        49, 49, 48, 48, 48, 48, 48, 48, 47, 47, 47, 47, 53, 54, 55, 56, 56, 56,
+        55, 54, 53, 52, 52, 51, 50, 50, 50, 49, 49, 49, 49, 49, 48, 48, 48, 48,
+        48, 48, 47, 47, 47, 47, 47, 47, 53, 54, 55, 55, 56, 55, 55, 54, 53, 52,
+        52, 51, 50, 50, 50, 49, 49, 49, 48, 48, 48, 48, 48, 48, 47, 47, 47, 47,
+        47, 47, 47, 47, 53, 53, 54, 55, 56, 55, 54, 53, 53, 52, 51, 51, 50, 50,
+        49, 49, 49, 48, 48, 48, 48, 48, 47, 47, 47, 47, 47, 47, 46, 46, 46, 46,
+        52, 53, 54, 55, 55, 55, 54, 53, 52, 52, 51, 51, 50, 50, 49, 49, 48, 48,
+        48, 48, 47, 47, 47, 47, 47, 47, 47, 46, 46, 46, 46, 46, 52, 53, 53, 54,
+        55, 54, 54, 53, 52, 52, 51, 50, 50, 49, 49, 49, 48, 48, 48, 47, 47, 47,
+        47, 47, 47, 46, 46, 46, 46, 46, 46, 46, 51, 52, 53, 54, 54, 54, 53, 53,
+        52, 51, 51, 50, 50, 49, 49, 48, 48, 48, 47, 47, 47, 47, 47, 46, 46, 46,
+        46, 46, 46, 46, 46, 46, 51, 52, 52, 53, 54, 53, 53, 52, 52, 51, 51, 50,
+        49, 49, 49, 48, 48, 48, 47, 47, 47, 47, 46, 46, 46, 46, 46, 46, 46, 46,
+        46, 46, 50, 51, 52, 53, 53, 53, 52, 52, 51, 51, 50, 50, 49, 49, 48, 48,
+        48, 47, 47, 47, 47, 46, 46, 46, 46, 46, 46, 45, 45, 45, 45, 45, 50, 51,
+        51, 52, 53, 52, 52, 52, 51, 51, 50, 50, 49, 49, 48, 48, 48, 47, 47, 47,
+        46, 46, 46, 46, 46, 46, 45, 45, 45, 45, 45, 45, 50, 50, 51, 52, 52, 52,
+        52, 51, 51, 50, 50, 49, 49, 49, 48, 48, 47, 47, 47, 47, 46, 46, 46, 46,
+        46, 45, 45, 45, 45, 45, 45, 45, 49, 50, 51, 51, 52, 52, 51, 51, 50, 50,
+        50, 49, 49, 48, 48, 48, 47, 47, 47, 46, 46, 46, 46, 46, 45, 45, 45, 45,
+        45, 45, 45, 45, 49, 50, 50, 51, 51, 51, 51, 50, 50, 50, 49, 49, 48, 48,
+        48, 47, 47, 47, 47, 46, 46, 46, 46, 45, 45, 45, 45, 45, 45, 45, 45, 45,
+        48, 49, 50, 50, 51, 51, 50, 50, 50, 49, 49, 49, 48, 48, 48, 47, 47, 47,
+        46, 46, 46, 46, 45, 45, 45, 45, 45, 45, 45, 45, 45, 45, 48, 49, 49, 50,
+        51, 50, 50, 50, 49, 49, 49, 48, 48, 48, 47, 47, 47, 46, 46, 46, 46, 46,
+        45, 45, 45, 45, 45, 45, 44, 44, 44, 44, 48, 49, 49, 50, 51, 50, 50, 50,
+        49, 49, 49, 48, 48, 48, 47, 47, 47, 46, 46, 46, 46, 46, 45, 45, 45, 45,
+        45, 45, 44, 44, 44, 44, 48, 49, 49, 50, 51, 50, 50, 50, 49, 49, 49, 48,
+        48, 48, 47, 47, 47, 46, 46, 46, 46, 46, 45, 45, 45, 45, 45, 45, 44, 44,
+        44, 44, 48, 49, 49, 50, 51, 50, 50, 50, 49, 49, 49, 48, 48, 48, 47, 47,
+        47, 46, 46, 46, 46, 46, 45, 45, 45, 45, 45, 45, 44, 44, 44, 44 },
+      { /* Intra matrices */
+        /* Size 4 */
+        85, 70, 68, 63, 70, 66, 64, 62, 68, 64, 60, 58, 63, 62, 58, 56,
+        /* Size 8 */
+        82, 91, 70, 69, 67, 65, 62, 60, 91, 76, 70, 72, 71, 69, 66, 64, 70, 70,
+        66, 67, 67, 66, 64, 62, 69, 72, 67, 64, 63, 62, 61, 60, 67, 71, 67, 63,
+        61, 60, 59, 59, 65, 69, 66, 62, 60, 59, 58, 57, 62, 66, 64, 61, 59, 58,
+        57, 56, 60, 64, 62, 60, 59, 57, 56, 55,
+        /* Size 16 */
+        82, 87, 91, 81, 71, 70, 69, 68, 68, 66, 65, 64, 63, 62, 61, 61, 87, 86,
+        84, 77, 70, 71, 71, 70, 70, 69, 67, 66, 65, 64, 63, 63, 91, 84, 77, 74,
+        70, 71, 73, 72, 72, 71, 70, 68, 67, 65, 64, 64, 81, 77, 74, 71, 68, 69,
+        70, 70, 70, 69, 68, 67, 66, 65, 63, 63, 71, 70, 70, 68, 67, 67, 67, 67,
+        67, 67, 66, 65, 65, 64, 63, 63, 70, 71, 71, 69, 67, 66, 66, 66, 66, 65,
+        65, 64, 63, 62, 62, 62, 69, 71, 73, 70, 67, 66, 65, 64, 64, 63, 63, 62,
+        62, 61, 61, 61, 68, 70, 72, 70, 67, 66, 64, 64, 63, 62, 62, 61, 61, 60,
+        60, 60, 68, 70, 72, 70, 67, 66, 64, 63, 62, 61, 61, 60, 60, 59, 59, 59,
+        66, 69, 71, 69, 67, 65, 63, 62, 61, 61, 60, 60, 59, 59, 58, 58, 65, 67,
+        70, 68, 66, 65, 63, 62, 61, 60, 59, 59, 58, 58, 58, 58, 64, 66, 68, 67,
+        65, 64, 62, 61, 60, 60, 59, 58, 58, 58, 57, 57, 63, 65, 67, 66, 65, 63,
+        62, 61, 60, 59, 58, 58, 57, 57, 57, 57, 62, 64, 65, 65, 64, 62, 61, 60,
+        59, 59, 58, 58, 57, 57, 56, 56, 61, 63, 64, 63, 63, 62, 61, 60, 59, 58,
+        58, 57, 57, 56, 56, 56, 61, 63, 64, 63, 63, 62, 61, 60, 59, 58, 58, 57,
+        57, 56, 56, 56,
+        /* Size 32 */
+        83, 85, 87, 90, 92, 87, 81, 76, 71, 71, 70, 70, 69, 69, 69, 68, 68, 67,
+        67, 66, 66, 65, 64, 64, 63, 63, 62, 62, 61, 61, 61, 61, 85, 86, 87, 87,
+        88, 84, 79, 75, 71, 71, 71, 70, 70, 70, 70, 69, 69, 68, 68, 67, 67, 66,
+        65, 65, 64, 64, 63, 63, 62, 62, 62, 62, 87, 87, 86, 85, 85, 81, 78, 74,
+        71, 71, 71, 71, 71, 71, 71, 70, 70, 69, 69, 68, 68, 67, 66, 66, 65, 65,
+        64, 63, 63, 63, 63, 63, 90, 87, 85, 83, 81, 78, 76, 73, 71, 71, 71, 72,
+        72, 72, 72, 71, 71, 71, 70, 69, 69, 68, 67, 67, 66, 66, 65, 64, 64, 64,
+        64, 64, 92, 88, 85, 81, 77, 76, 74, 72, 70, 71, 72, 72, 73, 73, 73, 72,
+        72, 72, 71, 70, 70, 69, 68, 68, 67, 66, 66, 65, 65, 65, 65, 65, 87, 84,
+        81, 78, 76, 74, 73, 71, 70, 70, 71, 71, 72, 72, 71, 71, 71, 71, 70, 70,
+        69, 68, 68, 67, 67, 66, 65, 65, 64, 64, 64, 64, 81, 79, 78, 76, 74, 73,
+        71, 70, 69, 69, 69, 70, 70, 70, 70, 70, 70, 70, 69, 69, 68, 68, 67, 66,
+        66, 65, 65, 64, 64, 64, 64, 64, 76, 75, 74, 73, 72, 71, 70, 69, 68, 68,
+        68, 69, 69, 69, 69, 69, 69, 68, 68, 68, 67, 67, 66, 66, 65, 65, 64, 64,
+        63, 63, 63, 63, 71, 71, 71, 71, 70, 70, 69, 68, 67, 67, 67, 67, 68, 68,
+        68, 68, 68, 67, 67, 67, 67, 66, 66, 65, 65, 64, 64, 63, 63, 63, 63, 63,
+        71, 71, 71, 71, 71, 70, 69, 68, 67, 67, 67, 67, 67, 67, 67, 67, 67, 66,
+        66, 66, 66, 65, 65, 65, 64, 64, 63, 63, 62, 62, 62, 62, 70, 71, 71, 71,
+        72, 71, 69, 68, 67, 67, 67, 66, 66, 66, 66, 66, 66, 66, 65, 65, 65, 65,
+        64, 64, 64, 63, 63, 62, 62, 62, 62, 62, 70, 70, 71, 72, 72, 71, 70, 69,
+        67, 67, 66, 66, 66, 65, 65, 65, 65, 65, 65, 64, 64, 64, 63, 63, 63, 63,
+        62, 62, 62, 62, 62, 62, 69, 70, 71, 72, 73, 72, 70, 69, 68, 67, 66, 66,
+        65, 65, 65, 64, 64, 64, 64, 63, 63, 63, 63, 63, 62, 62, 62, 61, 61, 61,
+        61, 61, 69, 70, 71, 72, 73, 72, 70, 69, 68, 67, 66, 65, 65, 64, 64, 64,
+        64, 63, 63, 63, 63, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 69, 70,
+        71, 72, 73, 71, 70, 69, 68, 67, 66, 65, 65, 64, 64, 63, 63, 63, 63, 62,
+        62, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 60, 68, 69, 70, 71, 72, 71,
+        70, 69, 68, 67, 66, 65, 64, 64, 63, 63, 63, 62, 62, 62, 62, 61, 61, 61,
+        61, 60, 60, 60, 60, 60, 60, 60, 68, 69, 70, 71, 72, 71, 70, 69, 68, 67,
+        66, 65, 64, 64, 63, 63, 62, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 60,
+        59, 59, 59, 59, 67, 68, 69, 71, 72, 71, 70, 68, 67, 66, 66, 65, 64, 63,
+        63, 62, 62, 62, 61, 61, 61, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59,
+        67, 68, 69, 70, 71, 70, 69, 68, 67, 66, 65, 65, 64, 63, 63, 62, 62, 61,
+        61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 66, 67, 68, 69,
+        70, 70, 69, 68, 67, 66, 65, 64, 63, 63, 62, 62, 61, 61, 61, 60, 60, 60,
+        59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 66, 67, 68, 69, 70, 69, 68, 67,
+        67, 66, 65, 64, 63, 63, 62, 62, 61, 61, 60, 60, 60, 59, 59, 59, 59, 58,
+        58, 58, 58, 58, 58, 58, 65, 66, 67, 68, 69, 68, 68, 67, 66, 65, 65, 64,
+        63, 62, 62, 61, 61, 60, 60, 60, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58,
+        58, 58, 64, 65, 66, 67, 68, 68, 67, 66, 66, 65, 64, 63, 63, 62, 62, 61,
+        61, 60, 60, 59, 59, 59, 59, 58, 58, 58, 58, 58, 57, 57, 57, 57, 64, 65,
+        66, 67, 68, 67, 66, 66, 65, 65, 64, 63, 63, 62, 61, 61, 60, 60, 60, 59,
+        59, 59, 58, 58, 58, 58, 58, 57, 57, 57, 57, 57, 63, 64, 65, 66, 67, 67,
+        66, 65, 65, 64, 64, 63, 62, 62, 61, 61, 60, 60, 59, 59, 59, 58, 58, 58,
+        58, 57, 57, 57, 57, 57, 57, 57, 63, 64, 65, 66, 66, 66, 65, 65, 64, 64,
+        63, 63, 62, 61, 61, 60, 60, 60, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57,
+        57, 57, 57, 57, 62, 63, 64, 65, 66, 65, 65, 64, 64, 63, 63, 62, 62, 61,
+        61, 60, 60, 59, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 57, 57, 57, 57,
+        62, 63, 63, 64, 65, 65, 64, 64, 63, 63, 62, 62, 61, 61, 60, 60, 60, 59,
+        59, 58, 58, 58, 58, 57, 57, 57, 57, 57, 56, 56, 56, 56, 61, 62, 63, 64,
+        65, 64, 64, 63, 63, 62, 62, 62, 61, 61, 60, 60, 59, 59, 59, 58, 58, 58,
+        57, 57, 57, 57, 57, 56, 56, 56, 56, 56, 61, 62, 63, 64, 65, 64, 64, 63,
+        63, 62, 62, 62, 61, 61, 60, 60, 59, 59, 59, 58, 58, 58, 57, 57, 57, 57,
+        57, 56, 56, 56, 56, 56, 61, 62, 63, 64, 65, 64, 64, 63, 63, 62, 62, 62,
+        61, 61, 60, 60, 59, 59, 59, 58, 58, 58, 57, 57, 57, 57, 57, 56, 56, 56,
+        56, 56, 61, 62, 63, 64, 65, 64, 64, 63, 63, 62, 62, 62, 61, 61, 60, 60,
+        59, 59, 59, 58, 58, 58, 57, 57, 57, 57, 57, 56, 56, 56, 56, 56 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 62, 54, 50, 62, 55, 52, 49, 54, 52, 49, 48, 50, 49, 48, 47,
+        /* Size 8 */
+        64, 70, 68, 63, 58, 55, 53, 52, 70, 67, 68, 65, 60, 57, 54, 53, 68, 68,
+        61, 59, 57, 55, 53, 52, 63, 65, 59, 56, 54, 53, 52, 51, 58, 60, 57, 54,
+        53, 52, 51, 51, 55, 57, 55, 53, 52, 51, 51, 50, 53, 54, 53, 52, 51, 51,
+        50, 50, 52, 53, 52, 51, 51, 50, 50, 50,
+        /* Size 16 */
+        64, 67, 70, 69, 68, 66, 63, 60, 58, 57, 55, 54, 53, 52, 52, 52, 67, 68,
+        69, 69, 68, 66, 64, 61, 59, 58, 56, 55, 54, 53, 52, 52, 70, 69, 67, 68,
+        68, 66, 65, 62, 60, 59, 57, 56, 54, 53, 53, 53, 69, 69, 68, 66, 65, 63,
+        62, 60, 59, 57, 56, 55, 54, 53, 52, 52, 68, 68, 68, 65, 61, 60, 59, 58,
+        57, 56, 55, 54, 53, 53, 52, 52, 66, 66, 66, 63, 60, 59, 57, 56, 56, 55,
+        54, 53, 53, 52, 52, 52, 63, 64, 65, 62, 59, 57, 56, 55, 54, 54, 53, 53,
+        52, 52, 51, 51, 60, 61, 62, 60, 58, 56, 55, 54, 54, 53, 52, 52, 52, 51,
+        51, 51, 58, 59, 60, 59, 57, 56, 54, 54, 53, 52, 52, 52, 51, 51, 51, 51,
+        57, 58, 59, 57, 56, 55, 54, 53, 52, 52, 51, 51, 51, 51, 50, 50, 55, 56,
+        57, 56, 55, 54, 53, 52, 52, 51, 51, 51, 51, 50, 50, 50, 54, 55, 56, 55,
+        54, 53, 53, 52, 52, 51, 51, 51, 50, 50, 50, 50, 53, 54, 54, 54, 53, 53,
+        52, 52, 51, 51, 51, 50, 50, 50, 50, 50, 52, 53, 53, 53, 53, 52, 52, 51,
+        51, 51, 50, 50, 50, 50, 50, 50, 52, 52, 53, 52, 52, 52, 51, 51, 51, 50,
+        50, 50, 50, 50, 50, 50, 52, 52, 53, 52, 52, 52, 51, 51, 51, 50, 50, 50,
+        50, 50, 50, 50,
+        /* Size 32 */
+        64, 66, 67, 69, 70, 70, 69, 69, 68, 67, 66, 64, 63, 62, 60, 59, 58, 57,
+        57, 56, 55, 54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 52, 66, 67, 68, 69,
+        70, 69, 69, 69, 68, 67, 66, 64, 63, 62, 61, 60, 59, 58, 57, 56, 55, 55,
+        54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 67, 68, 68, 68, 69, 69, 69, 68,
+        68, 67, 66, 65, 64, 63, 61, 60, 59, 58, 58, 57, 56, 55, 55, 54, 54, 53,
+        53, 52, 52, 52, 52, 52, 69, 69, 68, 68, 68, 68, 68, 68, 68, 67, 66, 65,
+        64, 63, 62, 61, 60, 59, 58, 57, 56, 56, 55, 55, 54, 54, 53, 53, 52, 52,
+        52, 52, 70, 70, 69, 68, 67, 67, 68, 68, 68, 67, 66, 66, 65, 64, 62, 61,
+        60, 59, 59, 58, 57, 56, 56, 55, 54, 54, 53, 53, 53, 53, 53, 53, 70, 69,
+        69, 68, 67, 67, 67, 67, 66, 66, 65, 64, 63, 62, 61, 60, 59, 59, 58, 57,
+        56, 56, 55, 55, 54, 54, 53, 53, 52, 52, 52, 52, 69, 69, 69, 68, 68, 67,
+        66, 65, 65, 64, 63, 63, 62, 61, 60, 59, 59, 58, 57, 57, 56, 55, 55, 54,
+        54, 53, 53, 53, 52, 52, 52, 52, 69, 69, 68, 68, 68, 67, 65, 64, 63, 62,
+        62, 61, 60, 60, 59, 58, 58, 57, 57, 56, 55, 55, 54, 54, 54, 53, 53, 53,
+        52, 52, 52, 52, 68, 68, 68, 68, 68, 66, 65, 63, 61, 61, 60, 60, 59, 58,
+        58, 57, 57, 56, 56, 55, 55, 54, 54, 54, 53, 53, 53, 52, 52, 52, 52, 52,
+        67, 67, 67, 67, 67, 66, 64, 62, 61, 60, 59, 59, 58, 58, 57, 57, 56, 56,
+        55, 55, 54, 54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 52, 66, 66, 66, 66,
+        66, 65, 63, 62, 60, 59, 59, 58, 57, 57, 56, 56, 56, 55, 55, 54, 54, 54,
+        53, 53, 53, 52, 52, 52, 52, 52, 52, 52, 64, 64, 65, 65, 66, 64, 63, 61,
+        60, 59, 58, 57, 57, 56, 56, 55, 55, 55, 54, 54, 54, 53, 53, 53, 52, 52,
+        52, 52, 51, 51, 51, 51, 63, 63, 64, 64, 65, 63, 62, 60, 59, 58, 57, 57,
+        56, 55, 55, 55, 54, 54, 54, 53, 53, 53, 53, 52, 52, 52, 52, 51, 51, 51,
+        51, 51, 62, 62, 63, 63, 64, 62, 61, 60, 58, 58, 57, 56, 55, 55, 55, 54,
+        54, 54, 53, 53, 53, 53, 52, 52, 52, 52, 52, 51, 51, 51, 51, 51, 60, 61,
+        61, 62, 62, 61, 60, 59, 58, 57, 56, 56, 55, 55, 54, 54, 54, 53, 53, 53,
+        52, 52, 52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 59, 60, 60, 61, 61, 60,
+        59, 58, 57, 57, 56, 55, 55, 54, 54, 54, 53, 53, 53, 52, 52, 52, 52, 52,
+        51, 51, 51, 51, 51, 51, 51, 51, 58, 59, 59, 60, 60, 59, 59, 58, 57, 56,
+        56, 55, 54, 54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 51, 51, 51, 51, 51,
+        51, 51, 51, 51, 57, 58, 58, 59, 59, 59, 58, 57, 56, 56, 55, 55, 54, 54,
+        53, 53, 53, 52, 52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 51, 51, 51, 51,
+        57, 57, 58, 58, 59, 58, 57, 57, 56, 55, 55, 54, 54, 53, 53, 53, 52, 52,
+        52, 52, 51, 51, 51, 51, 51, 51, 51, 50, 50, 50, 50, 50, 56, 56, 57, 57,
+        58, 57, 57, 56, 55, 55, 54, 54, 53, 53, 53, 52, 52, 52, 52, 51, 51, 51,
+        51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 55, 55, 56, 56, 57, 56, 56, 55,
+        55, 54, 54, 54, 53, 53, 52, 52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 50,
+        50, 50, 50, 50, 50, 50, 54, 55, 55, 56, 56, 56, 55, 55, 54, 54, 54, 53,
+        53, 53, 52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50,
+        50, 50, 54, 54, 55, 55, 56, 55, 55, 54, 54, 54, 53, 53, 53, 52, 52, 52,
+        52, 51, 51, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 50, 50, 50, 53, 54,
+        54, 55, 55, 55, 54, 54, 54, 53, 53, 53, 52, 52, 52, 52, 51, 51, 51, 51,
+        51, 51, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 53, 53, 54, 54, 54, 54,
+        54, 54, 53, 53, 53, 52, 52, 52, 52, 51, 51, 51, 51, 51, 51, 50, 50, 50,
+        50, 50, 50, 50, 50, 50, 50, 50, 53, 53, 53, 54, 54, 54, 53, 53, 53, 53,
+        52, 52, 52, 52, 51, 51, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 50, 50,
+        50, 50, 50, 50, 52, 53, 53, 53, 53, 53, 53, 53, 53, 52, 52, 52, 52, 52,
+        51, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50,
+        52, 52, 52, 53, 53, 53, 53, 53, 52, 52, 52, 52, 51, 51, 51, 51, 51, 51,
+        50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 52, 52, 52, 52,
+        53, 52, 52, 52, 52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 50, 50, 50, 50,
+        50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 52, 52, 52, 52, 53, 52, 52, 52,
+        52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 50, 50,
+        50, 50, 50, 50, 50, 50, 52, 52, 52, 52, 53, 52, 52, 52, 52, 52, 52, 51,
+        51, 51, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50,
+        50, 50, 52, 52, 52, 52, 53, 52, 52, 52, 52, 52, 52, 51, 51, 51, 51, 51,
+        51, 51, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50 },
+      { /* Intra matrices */
+        /* Size 4 */
+        80, 77, 67, 61, 77, 68, 63, 60, 67, 63, 60, 58, 61, 60, 58, 57,
+        /* Size 8 */
+        75, 83, 80, 73, 68, 64, 61, 60, 83, 79, 80, 76, 70, 66, 63, 61, 80, 80,
+        72, 69, 66, 64, 62, 60, 73, 76, 69, 65, 63, 61, 60, 59, 68, 70, 66, 63,
+        61, 60, 59, 58, 64, 66, 64, 61, 60, 59, 58, 58, 61, 63, 62, 60, 59, 58,
+        58, 57, 60, 61, 60, 59, 58, 58, 57, 57,
+        /* Size 16 */
+        76, 80, 84, 82, 81, 78, 74, 71, 68, 66, 64, 63, 62, 61, 60, 60, 80, 81,
+        82, 81, 81, 78, 75, 72, 70, 68, 66, 64, 63, 62, 61, 61, 84, 82, 80, 80,
+        81, 79, 76, 74, 71, 69, 67, 65, 64, 63, 61, 61, 82, 81, 80, 78, 77, 75,
+        73, 71, 69, 67, 65, 64, 63, 62, 61, 61, 81, 81, 81, 77, 72, 71, 69, 68,
+        67, 66, 64, 63, 62, 61, 61, 61, 78, 78, 79, 75, 71, 69, 67, 66, 65, 64,
+        63, 62, 62, 61, 60, 60, 74, 75, 76, 73, 69, 67, 66, 65, 64, 63, 62, 61,
+        61, 60, 60, 60, 71, 72, 74, 71, 68, 66, 65, 64, 63, 62, 61, 61, 60, 60,
+        59, 59, 68, 70, 71, 69, 67, 65, 64, 63, 62, 61, 60, 60, 60, 59, 59, 59,
+        66, 68, 69, 67, 66, 64, 63, 62, 61, 61, 60, 60, 59, 59, 59, 59, 64, 66,
+        67, 65, 64, 63, 62, 61, 60, 60, 59, 59, 59, 59, 58, 58, 63, 64, 65, 64,
+        63, 62, 61, 61, 60, 60, 59, 59, 59, 58, 58, 58, 62, 63, 64, 63, 62, 62,
+        61, 60, 60, 59, 59, 59, 58, 58, 58, 58, 61, 62, 63, 62, 61, 61, 60, 60,
+        59, 59, 59, 58, 58, 58, 58, 58, 60, 61, 61, 61, 61, 60, 60, 59, 59, 59,
+        58, 58, 58, 58, 58, 58, 60, 61, 61, 61, 61, 60, 60, 59, 59, 59, 58, 58,
+        58, 58, 58, 58,
+        /* Size 32 */
+        76, 78, 80, 82, 84, 83, 83, 82, 82, 80, 78, 76, 74, 73, 71, 70, 69, 68,
+        67, 66, 65, 64, 63, 63, 62, 62, 61, 61, 60, 60, 60, 60, 78, 79, 81, 82,
+        83, 83, 82, 82, 82, 80, 78, 77, 75, 74, 72, 71, 69, 68, 67, 66, 65, 65,
+        64, 63, 63, 62, 62, 61, 61, 61, 61, 61, 80, 81, 81, 82, 82, 82, 82, 82,
+        81, 80, 79, 77, 76, 74, 73, 71, 70, 69, 68, 67, 66, 65, 64, 64, 63, 63,
+        62, 62, 61, 61, 61, 61, 82, 82, 82, 81, 81, 81, 81, 81, 81, 80, 79, 77,
+        76, 75, 73, 72, 71, 70, 69, 67, 66, 66, 65, 64, 63, 63, 62, 62, 61, 61,
+        61, 61, 84, 83, 82, 81, 80, 80, 81, 81, 81, 80, 79, 78, 77, 75, 74, 73,
+        71, 70, 69, 68, 67, 66, 65, 65, 64, 63, 63, 62, 62, 62, 62, 62, 83, 83,
+        82, 81, 80, 80, 80, 79, 79, 78, 77, 76, 75, 74, 73, 71, 70, 69, 68, 67,
+        66, 66, 65, 64, 64, 63, 63, 62, 62, 62, 62, 62, 83, 82, 82, 81, 81, 80,
+        79, 78, 77, 76, 75, 74, 73, 72, 71, 70, 69, 68, 67, 67, 66, 65, 65, 64,
+        63, 63, 62, 62, 61, 61, 61, 61, 82, 82, 82, 81, 81, 79, 78, 76, 75, 74,
+        73, 72, 71, 71, 70, 69, 68, 67, 67, 66, 65, 65, 64, 63, 63, 62, 62, 62,
+        61, 61, 61, 61, 82, 82, 81, 81, 81, 79, 77, 75, 73, 72, 71, 70, 70, 69,
+        68, 68, 67, 66, 66, 65, 65, 64, 64, 63, 63, 62, 62, 61, 61, 61, 61, 61,
+        80, 80, 80, 80, 80, 78, 76, 74, 72, 71, 70, 70, 69, 68, 67, 67, 66, 66,
+        65, 65, 64, 64, 63, 63, 62, 62, 61, 61, 61, 61, 61, 61, 78, 78, 79, 79,
+        79, 77, 75, 73, 71, 70, 69, 69, 68, 67, 67, 66, 65, 65, 64, 64, 63, 63,
+        63, 62, 62, 62, 61, 61, 61, 61, 61, 61, 76, 77, 77, 77, 78, 76, 74, 72,
+        70, 70, 69, 68, 67, 66, 66, 65, 65, 64, 64, 63, 63, 63, 62, 62, 61, 61,
+        61, 61, 60, 60, 60, 60, 74, 75, 76, 76, 77, 75, 73, 71, 70, 69, 68, 67,
+        66, 65, 65, 64, 64, 63, 63, 63, 62, 62, 62, 61, 61, 61, 61, 60, 60, 60,
+        60, 60, 73, 74, 74, 75, 75, 74, 72, 71, 69, 68, 67, 66, 65, 65, 64, 64,
+        63, 63, 63, 62, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 71, 72,
+        73, 73, 74, 73, 71, 70, 68, 67, 67, 66, 65, 64, 64, 63, 63, 63, 62, 62,
+        62, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 70, 71, 71, 72, 73, 71,
+        70, 69, 68, 67, 66, 65, 64, 64, 63, 63, 62, 62, 62, 61, 61, 61, 61, 60,
+        60, 60, 60, 60, 59, 59, 59, 59, 69, 69, 70, 71, 71, 70, 69, 68, 67, 66,
+        65, 65, 64, 63, 63, 62, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 60, 59,
+        59, 59, 59, 59, 68, 68, 69, 70, 70, 69, 68, 67, 66, 66, 65, 64, 63, 63,
+        63, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59,
+        67, 67, 68, 69, 69, 68, 67, 67, 66, 65, 64, 64, 63, 63, 62, 62, 61, 61,
+        61, 61, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 66, 66, 67, 67,
+        68, 67, 67, 66, 65, 65, 64, 63, 63, 62, 62, 61, 61, 61, 61, 60, 60, 60,
+        60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 65, 65, 66, 66, 67, 66, 66, 65,
+        65, 64, 63, 63, 62, 62, 62, 61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 64, 65, 65, 66, 66, 66, 65, 65, 64, 64, 63, 63,
+        62, 62, 61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 58, 58,
+        58, 58, 63, 64, 64, 65, 65, 65, 65, 64, 64, 63, 63, 62, 62, 61, 61, 61,
+        60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 63, 63,
+        64, 64, 65, 64, 64, 63, 63, 63, 62, 62, 61, 61, 61, 60, 60, 60, 60, 59,
+        59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58, 62, 63, 63, 63, 64, 64,
+        63, 63, 63, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59,
+        59, 58, 58, 58, 58, 58, 58, 58, 62, 62, 63, 63, 63, 63, 63, 62, 62, 62,
+        62, 61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58,
+        58, 58, 58, 58, 61, 62, 62, 62, 63, 63, 62, 62, 62, 61, 61, 61, 61, 60,
+        60, 60, 60, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58, 58, 58, 58,
+        61, 61, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 59, 59,
+        59, 59, 59, 59, 58, 58, 58, 58, 58, 58, 58, 58, 58, 58, 60, 61, 61, 61,
+        62, 62, 61, 61, 61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 58,
+        58, 58, 58, 58, 58, 58, 58, 58, 58, 58, 60, 61, 61, 61, 62, 62, 61, 61,
+        61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58,
+        58, 58, 58, 58, 58, 58, 60, 61, 61, 61, 62, 62, 61, 61, 61, 61, 61, 60,
+        60, 60, 60, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58, 58, 58, 58,
+        58, 58, 60, 61, 61, 61, 62, 62, 61, 61, 61, 61, 61, 60, 60, 60, 60, 59,
+        59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58, 58, 58, 58, 58, 58 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 56, 54, 52, 56, 53, 52, 51, 54, 52, 50, 49, 52, 51, 49, 48,
+        /* Size 8 */
+        64, 69, 57, 56, 55, 54, 53, 52, 69, 61, 57, 58, 58, 57, 55, 53, 57, 57,
+        55, 55, 55, 55, 54, 53, 56, 58, 55, 54, 53, 53, 52, 51, 55, 58, 55, 53,
+        52, 51, 51, 50, 54, 57, 55, 53, 51, 51, 50, 50, 53, 55, 54, 52, 51, 50,
+        49, 49, 52, 53, 53, 51, 50, 50, 49, 49,
+        /* Size 16 */
+        64, 67, 69, 63, 57, 57, 56, 56, 55, 55, 54, 53, 53, 52, 52, 52, 67, 66,
+        65, 61, 57, 57, 57, 57, 57, 56, 55, 55, 54, 53, 52, 52, 69, 65, 61, 59,
+        57, 58, 58, 58, 58, 57, 57, 56, 55, 54, 53, 53, 63, 61, 59, 57, 56, 56,
+        57, 57, 57, 56, 56, 55, 54, 54, 53, 53, 57, 57, 57, 56, 55, 55, 55, 55,
+        55, 55, 55, 54, 54, 53, 53, 53, 57, 57, 58, 56, 55, 55, 54, 54, 54, 54,
+        54, 53, 53, 52, 52, 52, 56, 57, 58, 57, 55, 54, 54, 53, 53, 53, 53, 52,
+        52, 52, 51, 51, 56, 57, 58, 57, 55, 54, 53, 53, 53, 52, 52, 52, 52, 51,
+        51, 51, 55, 57, 58, 57, 55, 54, 53, 53, 52, 52, 51, 51, 51, 51, 50, 50,
+        55, 56, 57, 56, 55, 54, 53, 52, 52, 51, 51, 51, 51, 50, 50, 50, 54, 55,
+        57, 56, 55, 54, 53, 52, 51, 51, 51, 50, 50, 50, 50, 50, 53, 55, 56, 55,
+        54, 53, 52, 52, 51, 51, 50, 50, 50, 50, 49, 49, 53, 54, 55, 54, 54, 53,
+        52, 52, 51, 51, 50, 50, 49, 49, 49, 49, 52, 53, 54, 54, 53, 52, 52, 51,
+        51, 50, 50, 50, 49, 49, 49, 49, 52, 52, 53, 53, 53, 52, 51, 51, 50, 50,
+        50, 49, 49, 49, 49, 49, 52, 52, 53, 53, 53, 52, 51, 51, 50, 50, 50, 49,
+        49, 49, 49, 49,
+        /* Size 32 */
+        64, 65, 67, 68, 69, 66, 63, 60, 57, 57, 57, 57, 56, 56, 56, 56, 55, 55,
+        55, 54, 54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 52, 52, 65, 66, 66, 67,
+        67, 65, 62, 60, 57, 57, 57, 57, 57, 57, 56, 56, 56, 56, 55, 55, 55, 54,
+        54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 67, 66, 66, 65, 65, 63, 61, 59,
+        57, 57, 57, 57, 57, 57, 57, 57, 57, 56, 56, 56, 55, 55, 55, 54, 54, 53,
+        53, 53, 52, 52, 52, 52, 68, 67, 65, 64, 63, 61, 60, 58, 57, 57, 57, 58,
+        58, 58, 58, 57, 57, 57, 57, 56, 56, 56, 55, 55, 54, 54, 54, 53, 53, 53,
+        53, 53, 69, 67, 65, 63, 61, 60, 59, 58, 57, 57, 58, 58, 58, 58, 58, 58,
+        58, 58, 57, 57, 57, 56, 56, 55, 55, 55, 54, 54, 53, 53, 53, 53, 66, 65,
+        63, 61, 60, 59, 58, 57, 56, 57, 57, 57, 58, 58, 57, 57, 57, 57, 57, 56,
+        56, 56, 55, 55, 55, 54, 54, 54, 53, 53, 53, 53, 63, 62, 61, 60, 59, 58,
+        57, 57, 56, 56, 56, 57, 57, 57, 57, 57, 57, 56, 56, 56, 56, 55, 55, 55,
+        54, 54, 54, 53, 53, 53, 53, 53, 60, 60, 59, 58, 58, 57, 57, 56, 55, 55,
+        56, 56, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 54, 54, 54, 53, 53,
+        53, 53, 53, 53, 57, 57, 57, 57, 57, 56, 56, 55, 55, 55, 55, 55, 55, 55,
+        55, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 53, 53, 53, 53, 53, 53, 53,
+        57, 57, 57, 57, 57, 57, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55,
+        54, 54, 54, 54, 54, 53, 53, 53, 53, 53, 52, 52, 52, 52, 57, 57, 57, 57,
+        58, 57, 56, 56, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 54, 53,
+        53, 53, 53, 53, 52, 52, 52, 52, 52, 52, 57, 57, 57, 58, 58, 57, 57, 56,
+        55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 53, 53, 53, 53, 53, 53, 53, 52,
+        52, 52, 52, 52, 52, 52, 56, 57, 57, 58, 58, 58, 57, 56, 55, 55, 54, 54,
+        54, 54, 53, 53, 53, 53, 53, 53, 53, 53, 52, 52, 52, 52, 52, 52, 51, 51,
+        51, 51, 56, 57, 57, 58, 58, 58, 57, 56, 55, 55, 54, 54, 54, 53, 53, 53,
+        53, 53, 53, 53, 52, 52, 52, 52, 52, 52, 52, 51, 51, 51, 51, 51, 56, 56,
+        57, 58, 58, 57, 57, 56, 55, 55, 54, 54, 53, 53, 53, 53, 53, 52, 52, 52,
+        52, 52, 52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 56, 56, 57, 57, 58, 57,
+        57, 56, 55, 55, 54, 54, 53, 53, 53, 53, 52, 52, 52, 52, 52, 52, 51, 51,
+        51, 51, 51, 51, 51, 51, 51, 51, 55, 56, 57, 57, 58, 57, 57, 56, 55, 55,
+        54, 54, 53, 53, 53, 52, 52, 52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 51,
+        50, 50, 50, 50, 55, 56, 56, 57, 58, 57, 56, 56, 55, 55, 54, 54, 53, 53,
+        52, 52, 52, 52, 52, 51, 51, 51, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50,
+        55, 55, 56, 57, 57, 57, 56, 55, 55, 54, 54, 53, 53, 53, 52, 52, 52, 52,
+        51, 51, 51, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 50, 54, 55, 56, 56,
+        57, 56, 56, 55, 55, 54, 54, 53, 53, 53, 52, 52, 52, 51, 51, 51, 51, 51,
+        51, 50, 50, 50, 50, 50, 50, 50, 50, 50, 54, 55, 55, 56, 57, 56, 56, 55,
+        55, 54, 54, 53, 53, 52, 52, 52, 51, 51, 51, 51, 51, 50, 50, 50, 50, 50,
+        50, 50, 50, 50, 50, 50, 54, 54, 55, 56, 56, 56, 55, 55, 54, 54, 53, 53,
+        53, 52, 52, 52, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 50, 50, 50, 50,
+        50, 50, 53, 54, 55, 55, 56, 55, 55, 55, 54, 54, 53, 53, 52, 52, 52, 51,
+        51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 50, 49, 49, 49, 49, 49, 53, 54,
+        54, 55, 55, 55, 55, 54, 54, 53, 53, 53, 52, 52, 52, 51, 51, 51, 51, 50,
+        50, 50, 50, 50, 50, 50, 49, 49, 49, 49, 49, 49, 53, 53, 54, 54, 55, 55,
+        54, 54, 54, 53, 53, 53, 52, 52, 52, 51, 51, 51, 51, 50, 50, 50, 50, 50,
+        49, 49, 49, 49, 49, 49, 49, 49, 52, 53, 53, 54, 55, 54, 54, 54, 53, 53,
+        53, 52, 52, 52, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 49, 49, 49, 49,
+        49, 49, 49, 49, 52, 53, 53, 54, 54, 54, 54, 53, 53, 53, 52, 52, 52, 52,
+        51, 51, 51, 50, 50, 50, 50, 50, 50, 49, 49, 49, 49, 49, 49, 49, 49, 49,
+        52, 52, 53, 53, 54, 54, 53, 53, 53, 53, 52, 52, 52, 51, 51, 51, 51, 50,
+        50, 50, 50, 50, 49, 49, 49, 49, 49, 49, 49, 49, 49, 49, 52, 52, 52, 53,
+        53, 53, 53, 53, 53, 52, 52, 52, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50,
+        49, 49, 49, 49, 49, 49, 49, 49, 49, 49, 52, 52, 52, 53, 53, 53, 53, 53,
+        53, 52, 52, 52, 51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 49, 49, 49, 49,
+        49, 49, 49, 49, 49, 49, 52, 52, 52, 53, 53, 53, 53, 53, 53, 52, 52, 52,
+        51, 51, 51, 51, 50, 50, 50, 50, 50, 50, 49, 49, 49, 49, 49, 49, 49, 49,
+        49, 49, 52, 52, 52, 53, 53, 53, 53, 53, 53, 52, 52, 52, 51, 51, 51, 51,
+        50, 50, 50, 50, 50, 50, 49, 49, 49, 49, 49, 49, 49, 49, 49, 49 },
+      { /* Intra matrices */
+        /* Size 4 */
+        79, 68, 67, 63, 68, 65, 64, 62, 67, 64, 61, 60, 63, 62, 60, 58,
+        /* Size 8 */
+        77, 84, 68, 67, 66, 64, 63, 61, 84, 73, 68, 70, 69, 67, 65, 64, 68, 68,
+        65, 66, 66, 65, 64, 62, 67, 70, 66, 64, 63, 63, 62, 61, 66, 69, 66, 63,
+        62, 61, 60, 60, 64, 67, 65, 63, 61, 60, 59, 59, 63, 65, 64, 62, 60, 59,
+        59, 58, 61, 64, 62, 61, 60, 59, 58, 58,
+        /* Size 16 */
+        77, 81, 84, 76, 69, 68, 68, 67, 67, 66, 65, 64, 63, 62, 62, 62, 81, 80,
+        79, 74, 69, 69, 69, 69, 68, 67, 66, 65, 64, 64, 63, 63, 84, 79, 74, 71,
+        68, 69, 70, 70, 70, 69, 68, 67, 66, 65, 64, 64, 76, 74, 71, 69, 67, 68,
+        68, 68, 68, 67, 67, 66, 65, 64, 63, 63, 69, 69, 68, 67, 66, 66, 66, 66,
+        66, 66, 66, 65, 64, 64, 63, 63, 68, 69, 69, 68, 66, 66, 65, 65, 65, 65,
+        64, 64, 63, 63, 62, 62, 68, 69, 70, 68, 66, 65, 64, 64, 64, 63, 63, 63,
+        62, 62, 61, 61, 67, 69, 70, 68, 66, 65, 64, 64, 63, 63, 62, 62, 62, 61,
+        61, 61, 67, 68, 70, 68, 66, 65, 64, 63, 62, 62, 61, 61, 61, 61, 60, 60,
+        66, 67, 69, 67, 66, 65, 63, 63, 62, 61, 61, 61, 60, 60, 60, 60, 65, 66,
+        68, 67, 66, 64, 63, 62, 61, 61, 60, 60, 60, 59, 59, 59, 64, 65, 67, 66,
+        65, 64, 63, 62, 61, 61, 60, 60, 59, 59, 59, 59, 63, 64, 66, 65, 64, 63,
+        62, 62, 61, 60, 60, 59, 59, 59, 58, 58, 62, 64, 65, 64, 64, 63, 62, 61,
+        61, 60, 59, 59, 59, 58, 58, 58, 62, 63, 64, 63, 63, 62, 61, 61, 60, 60,
+        59, 59, 58, 58, 58, 58, 62, 63, 64, 63, 63, 62, 61, 61, 60, 60, 59, 59,
+        58, 58, 58, 58,
+        /* Size 32 */
+        78, 79, 81, 83, 84, 81, 77, 73, 69, 69, 68, 68, 68, 68, 67, 67, 67, 66,
+        66, 65, 65, 65, 64, 64, 63, 63, 63, 62, 62, 62, 62, 62, 79, 80, 81, 81,
+        82, 79, 75, 72, 69, 69, 69, 69, 69, 68, 68, 68, 68, 67, 67, 66, 66, 65,
+        65, 64, 64, 64, 63, 63, 62, 62, 62, 62, 81, 81, 80, 80, 79, 77, 74, 71,
+        69, 69, 69, 69, 69, 69, 69, 69, 68, 68, 67, 67, 67, 66, 66, 65, 65, 64,
+        64, 63, 63, 63, 63, 63, 83, 81, 80, 78, 76, 75, 73, 71, 69, 69, 69, 70,
+        70, 70, 70, 69, 69, 69, 68, 68, 67, 67, 66, 66, 65, 65, 65, 64, 64, 64,
+        64, 64, 84, 82, 79, 76, 74, 73, 71, 70, 69, 69, 70, 70, 71, 70, 70, 70,
+        70, 70, 69, 69, 68, 68, 67, 67, 66, 66, 65, 65, 64, 64, 64, 64, 81, 79,
+        77, 75, 73, 71, 70, 69, 68, 68, 69, 69, 70, 69, 69, 69, 69, 69, 68, 68,
+        68, 67, 67, 66, 66, 65, 65, 64, 64, 64, 64, 64, 77, 75, 74, 73, 71, 70,
+        69, 68, 67, 68, 68, 68, 69, 68, 68, 68, 68, 68, 68, 67, 67, 67, 66, 66,
+        65, 65, 64, 64, 64, 64, 64, 64, 73, 72, 71, 71, 70, 69, 68, 67, 67, 67,
+        67, 67, 68, 67, 67, 67, 67, 67, 67, 67, 66, 66, 66, 65, 65, 65, 64, 64,
+        63, 63, 63, 63, 69, 69, 69, 69, 69, 68, 67, 67, 66, 66, 66, 66, 66, 67,
+        67, 67, 67, 66, 66, 66, 66, 65, 65, 65, 64, 64, 64, 63, 63, 63, 63, 63,
+        69, 69, 69, 69, 69, 68, 68, 67, 66, 66, 66, 66, 66, 66, 66, 66, 66, 66,
+        66, 65, 65, 65, 65, 64, 64, 64, 63, 63, 63, 63, 63, 63, 68, 69, 69, 69,
+        70, 69, 68, 67, 66, 66, 66, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 64,
+        64, 64, 64, 63, 63, 63, 62, 62, 62, 62, 68, 69, 69, 70, 70, 69, 68, 67,
+        66, 66, 66, 65, 65, 65, 65, 65, 65, 64, 64, 64, 64, 64, 63, 63, 63, 63,
+        63, 62, 62, 62, 62, 62, 68, 69, 69, 70, 71, 70, 69, 68, 66, 66, 66, 65,
+        65, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62,
+        62, 62, 68, 68, 69, 70, 70, 69, 68, 67, 67, 66, 65, 65, 64, 64, 64, 64,
+        64, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61, 67, 68,
+        69, 70, 70, 69, 68, 67, 67, 66, 65, 65, 64, 64, 64, 63, 63, 63, 63, 63,
+        62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 67, 68, 69, 69, 70, 69,
+        68, 67, 67, 66, 65, 65, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62,
+        61, 61, 61, 61, 61, 61, 61, 61, 67, 68, 68, 69, 70, 69, 68, 67, 67, 66,
+        65, 65, 64, 64, 63, 63, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61,
+        60, 60, 60, 60, 66, 67, 68, 69, 70, 69, 68, 67, 66, 66, 65, 64, 64, 63,
+        63, 63, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60,
+        66, 67, 67, 68, 69, 68, 68, 67, 66, 66, 65, 64, 64, 63, 63, 62, 62, 62,
+        62, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 65, 66, 67, 68,
+        69, 68, 67, 67, 66, 65, 65, 64, 63, 63, 63, 62, 62, 62, 61, 61, 61, 61,
+        61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 65, 66, 67, 67, 68, 68, 67, 66,
+        66, 65, 65, 64, 63, 63, 62, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 60,
+        60, 60, 59, 59, 59, 59, 65, 65, 66, 67, 68, 67, 67, 66, 65, 65, 64, 64,
+        63, 63, 62, 62, 62, 61, 61, 61, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59,
+        59, 59, 64, 65, 66, 66, 67, 67, 66, 66, 65, 65, 64, 63, 63, 63, 62, 62,
+        61, 61, 61, 61, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 64, 64,
+        65, 66, 67, 66, 66, 65, 65, 64, 64, 63, 63, 62, 62, 62, 61, 61, 61, 60,
+        60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 63, 64, 65, 65, 66, 66,
+        65, 65, 64, 64, 64, 63, 63, 62, 62, 61, 61, 61, 60, 60, 60, 60, 60, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 63, 64, 64, 65, 66, 65, 65, 65, 64, 64,
+        63, 63, 62, 62, 62, 61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 63, 63, 64, 65, 65, 65, 64, 64, 64, 63, 63, 63, 62, 62,
+        61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58,
+        62, 63, 63, 64, 65, 64, 64, 64, 63, 63, 63, 62, 62, 62, 61, 61, 61, 60,
+        60, 60, 60, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 62, 62, 63, 64,
+        64, 64, 64, 63, 63, 63, 62, 62, 62, 61, 61, 61, 60, 60, 60, 60, 59, 59,
+        59, 59, 59, 59, 58, 58, 58, 58, 58, 58, 62, 62, 63, 64, 64, 64, 64, 63,
+        63, 63, 62, 62, 62, 61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59,
+        58, 58, 58, 58, 58, 58, 62, 62, 63, 64, 64, 64, 64, 63, 63, 63, 62, 62,
+        62, 61, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58,
+        58, 58, 62, 62, 63, 64, 64, 64, 64, 63, 63, 63, 62, 62, 62, 61, 61, 61,
+        60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 63, 57, 54, 63, 58, 55, 54, 57, 55, 53, 53, 54, 54, 53, 52,
+        /* Size 8 */
+        64, 68, 67, 63, 60, 58, 56, 55, 68, 66, 67, 64, 61, 59, 57, 56, 67, 67,
+        62, 61, 59, 58, 57, 56, 63, 64, 61, 58, 57, 56, 56, 55, 60, 61, 59, 57,
+        56, 56, 55, 55, 58, 59, 58, 56, 56, 55, 55, 54, 56, 57, 57, 56, 55, 55,
+        54, 54, 55, 56, 56, 55, 55, 54, 54, 54,
+        /* Size 16 */
+        64, 66, 68, 68, 67, 65, 63, 62, 60, 59, 58, 57, 56, 56, 55, 55, 66, 67,
+        67, 67, 67, 65, 64, 62, 61, 60, 58, 58, 57, 56, 56, 56, 68, 67, 66, 66,
+        67, 66, 64, 63, 61, 60, 59, 58, 57, 57, 56, 56, 68, 67, 66, 65, 65, 63,
+        62, 61, 60, 59, 58, 58, 57, 56, 56, 56, 67, 67, 67, 65, 62, 61, 61, 60,
+        59, 58, 58, 57, 57, 56, 56, 56, 65, 65, 66, 63, 61, 60, 59, 59, 58, 58,
+        57, 57, 56, 56, 56, 56, 63, 64, 64, 62, 61, 59, 58, 58, 57, 57, 56, 56,
+        56, 56, 55, 55, 62, 62, 63, 61, 60, 59, 58, 57, 57, 56, 56, 56, 55, 55,
+        55, 55, 60, 61, 61, 60, 59, 58, 57, 57, 56, 56, 56, 55, 55, 55, 55, 55,
+        59, 60, 60, 59, 58, 58, 57, 56, 56, 56, 55, 55, 55, 55, 55, 55, 58, 58,
+        59, 58, 58, 57, 56, 56, 56, 55, 55, 55, 55, 55, 54, 54, 57, 58, 58, 58,
+        57, 57, 56, 56, 55, 55, 55, 55, 55, 54, 54, 54, 56, 57, 57, 57, 57, 56,
+        56, 55, 55, 55, 55, 55, 54, 54, 54, 54, 56, 56, 57, 56, 56, 56, 56, 55,
+        55, 55, 55, 54, 54, 54, 54, 54, 55, 56, 56, 56, 56, 56, 55, 55, 55, 55,
+        54, 54, 54, 54, 54, 54, 55, 56, 56, 56, 56, 56, 55, 55, 55, 55, 54, 54,
+        54, 54, 54, 54,
+        /* Size 32 */
+        64, 65, 66, 67, 68, 68, 68, 67, 67, 66, 65, 64, 63, 62, 62, 61, 60, 59,
+        59, 58, 58, 57, 57, 57, 56, 56, 56, 56, 55, 55, 55, 55, 65, 66, 66, 67,
+        68, 68, 67, 67, 67, 66, 65, 64, 63, 63, 62, 61, 60, 60, 59, 59, 58, 58,
+        57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 66, 66, 67, 67, 67, 67, 67, 67,
+        67, 66, 65, 65, 64, 63, 62, 61, 61, 60, 60, 59, 58, 58, 58, 57, 57, 57,
+        56, 56, 56, 56, 56, 56, 67, 67, 67, 67, 67, 67, 67, 67, 67, 66, 66, 65,
+        64, 63, 63, 62, 61, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56,
+        56, 56, 68, 68, 67, 67, 66, 66, 66, 67, 67, 66, 66, 65, 64, 64, 63, 62,
+        61, 61, 60, 60, 59, 59, 58, 58, 57, 57, 57, 56, 56, 56, 56, 56, 68, 68,
+        67, 67, 66, 66, 66, 66, 66, 65, 65, 64, 63, 63, 62, 62, 61, 60, 60, 59,
+        59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56, 56, 68, 67, 67, 67, 66, 66,
+        65, 65, 65, 64, 63, 63, 62, 62, 61, 61, 60, 60, 59, 59, 58, 58, 58, 57,
+        57, 57, 56, 56, 56, 56, 56, 56, 67, 67, 67, 67, 67, 66, 65, 64, 63, 63,
+        62, 62, 62, 61, 61, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 57, 56, 56,
+        56, 56, 56, 56, 67, 67, 67, 67, 67, 66, 65, 63, 62, 62, 61, 61, 61, 60,
+        60, 59, 59, 59, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56,
+        66, 66, 66, 66, 66, 65, 64, 63, 62, 61, 61, 60, 60, 60, 59, 59, 59, 58,
+        58, 58, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 56, 65, 65, 65, 66,
+        66, 65, 63, 62, 61, 61, 60, 60, 59, 59, 59, 59, 58, 58, 58, 57, 57, 57,
+        57, 56, 56, 56, 56, 56, 56, 56, 56, 56, 64, 64, 65, 65, 65, 64, 63, 62,
+        61, 60, 60, 59, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56, 56,
+        56, 56, 55, 55, 55, 55, 63, 63, 64, 64, 64, 63, 62, 62, 61, 60, 59, 59,
+        58, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 55, 55, 55,
+        55, 55, 62, 63, 63, 63, 64, 63, 62, 61, 60, 60, 59, 59, 58, 58, 58, 57,
+        57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 62, 62,
+        62, 63, 63, 62, 61, 61, 60, 59, 59, 58, 58, 58, 57, 57, 57, 57, 56, 56,
+        56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 61, 61, 61, 62, 62, 62,
+        61, 60, 59, 59, 59, 58, 58, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 55,
+        55, 55, 55, 55, 55, 55, 55, 55, 60, 60, 61, 61, 61, 61, 60, 60, 59, 59,
+        58, 58, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55,
+        55, 55, 55, 55, 59, 60, 60, 60, 61, 60, 60, 59, 59, 58, 58, 58, 57, 57,
+        57, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55,
+        59, 59, 60, 60, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56,
+        56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 58, 59, 59, 59,
+        60, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56, 56, 56, 55, 55, 55,
+        55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 58, 58, 58, 59, 59, 59, 58, 58,
+        58, 57, 57, 57, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55,
+        55, 55, 54, 54, 54, 54, 57, 58, 58, 58, 59, 58, 58, 58, 57, 57, 57, 57,
+        56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 54, 54, 54,
+        54, 54, 57, 57, 58, 58, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56, 56, 56,
+        55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 57, 57,
+        57, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55,
+        55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 56, 57, 57, 57, 57, 57,
+        57, 57, 57, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 54,
+        54, 54, 54, 54, 54, 54, 54, 54, 56, 56, 57, 57, 57, 57, 57, 57, 56, 56,
+        56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54,
+        54, 54, 54, 54, 56, 56, 56, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 55,
+        55, 55, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54,
+        56, 56, 56, 56, 56, 56, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55,
+        55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 55, 56, 56, 56,
+        56, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 54, 54,
+        54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 55, 56, 56, 56, 56, 56, 56, 56,
+        56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54,
+        54, 54, 54, 54, 54, 54, 55, 56, 56, 56, 56, 56, 56, 56, 56, 56, 56, 55,
+        55, 55, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54,
+        54, 54, 55, 56, 56, 56, 56, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55,
+        55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54 },
+      { /* Intra matrices */
+        /* Size 4 */
+        74, 72, 66, 62, 72, 67, 63, 61, 66, 63, 61, 60, 62, 61, 60, 59,
+        /* Size 8 */
+        71, 76, 75, 70, 66, 64, 62, 61, 76, 73, 74, 71, 68, 65, 63, 62, 75, 74,
+        69, 67, 65, 64, 62, 61, 70, 71, 67, 64, 63, 62, 61, 61, 66, 68, 65, 63,
+        62, 61, 61, 60, 64, 65, 64, 62, 61, 60, 60, 60, 62, 63, 62, 61, 61, 60,
+        60, 59, 61, 62, 61, 61, 60, 60, 59, 59,
+        /* Size 16 */
+        71, 74, 77, 76, 75, 73, 70, 68, 67, 65, 64, 63, 62, 62, 61, 61, 74, 75,
+        75, 75, 75, 73, 71, 69, 67, 66, 65, 64, 63, 62, 62, 62, 77, 75, 74, 74,
+        75, 73, 72, 70, 68, 67, 66, 65, 64, 63, 62, 62, 76, 75, 74, 73, 72, 71,
+        70, 68, 67, 66, 65, 64, 63, 63, 62, 62, 75, 75, 75, 72, 69, 68, 67, 66,
+        66, 65, 64, 63, 63, 62, 62, 62, 73, 73, 73, 71, 68, 67, 66, 65, 65, 64,
+        63, 63, 62, 62, 61, 61, 70, 71, 72, 70, 67, 66, 65, 64, 63, 63, 63, 62,
+        62, 61, 61, 61, 68, 69, 70, 68, 66, 65, 64, 64, 63, 62, 62, 62, 61, 61,
+        61, 61, 67, 67, 68, 67, 66, 65, 63, 63, 62, 62, 62, 61, 61, 61, 61, 61,
+        65, 66, 67, 66, 65, 64, 63, 62, 62, 62, 61, 61, 61, 61, 60, 60, 64, 65,
+        66, 65, 64, 63, 63, 62, 62, 61, 61, 61, 60, 60, 60, 60, 63, 64, 65, 64,
+        63, 63, 62, 62, 61, 61, 61, 60, 60, 60, 60, 60, 62, 63, 64, 63, 63, 62,
+        62, 61, 61, 61, 60, 60, 60, 60, 60, 60, 62, 62, 63, 63, 62, 62, 61, 61,
+        61, 61, 60, 60, 60, 60, 60, 60, 61, 62, 62, 62, 62, 61, 61, 61, 61, 60,
+        60, 60, 60, 60, 60, 60, 61, 62, 62, 62, 62, 61, 61, 61, 61, 60, 60, 60,
+        60, 60, 60, 60,
+        /* Size 32 */
+        72, 73, 74, 76, 77, 76, 76, 76, 75, 74, 73, 72, 71, 70, 69, 68, 67, 66,
+        66, 65, 64, 64, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 73, 74, 75, 75,
+        76, 76, 76, 75, 75, 74, 73, 72, 71, 70, 69, 68, 67, 67, 66, 65, 65, 64,
+        64, 63, 63, 63, 62, 62, 62, 62, 62, 62, 74, 75, 75, 75, 76, 75, 75, 75,
+        75, 74, 73, 72, 71, 70, 70, 69, 68, 67, 66, 66, 65, 65, 64, 64, 63, 63,
+        63, 62, 62, 62, 62, 62, 76, 75, 75, 75, 75, 75, 75, 75, 75, 74, 73, 73,
+        72, 71, 70, 69, 68, 67, 67, 66, 65, 65, 64, 64, 63, 63, 63, 62, 62, 62,
+        62, 62, 77, 76, 76, 75, 74, 74, 75, 75, 75, 74, 74, 73, 72, 71, 70, 69,
+        69, 68, 67, 66, 66, 65, 65, 64, 64, 63, 63, 63, 62, 62, 62, 62, 76, 76,
+        75, 75, 74, 74, 74, 74, 74, 73, 72, 72, 71, 70, 69, 69, 68, 67, 67, 66,
+        65, 65, 64, 64, 64, 63, 63, 63, 62, 62, 62, 62, 76, 76, 75, 75, 75, 74,
+        73, 73, 72, 72, 71, 70, 70, 69, 69, 68, 67, 67, 66, 66, 65, 65, 64, 64,
+        63, 63, 63, 62, 62, 62, 62, 62, 76, 75, 75, 75, 75, 74, 73, 72, 71, 70,
+        70, 69, 69, 68, 68, 67, 66, 66, 66, 65, 65, 64, 64, 63, 63, 63, 63, 62,
+        62, 62, 62, 62, 75, 75, 75, 75, 75, 74, 72, 71, 69, 69, 68, 68, 67, 67,
+        67, 66, 66, 65, 65, 65, 64, 64, 64, 63, 63, 63, 62, 62, 62, 62, 62, 62,
+        74, 74, 74, 74, 74, 73, 72, 70, 69, 68, 68, 67, 67, 66, 66, 66, 65, 65,
+        65, 64, 64, 64, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 73, 73, 73, 73,
+        74, 72, 71, 70, 68, 68, 67, 67, 66, 66, 65, 65, 65, 64, 64, 64, 63, 63,
+        63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 72, 72, 72, 73, 73, 72, 70, 69,
+        68, 67, 67, 66, 66, 65, 65, 65, 64, 64, 64, 63, 63, 63, 63, 62, 62, 62,
+        62, 62, 61, 61, 61, 61, 71, 71, 71, 72, 72, 71, 70, 69, 67, 67, 66, 66,
+        65, 65, 64, 64, 64, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61,
+        61, 61, 70, 70, 70, 71, 71, 70, 69, 68, 67, 66, 66, 65, 65, 64, 64, 64,
+        63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 69, 69,
+        70, 70, 70, 69, 69, 68, 67, 66, 65, 65, 64, 64, 64, 63, 63, 63, 63, 62,
+        62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 68, 68, 69, 69, 69, 69,
+        68, 67, 66, 66, 65, 65, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62,
+        61, 61, 61, 61, 61, 61, 61, 61, 67, 67, 68, 68, 69, 68, 67, 66, 66, 65,
+        65, 64, 64, 63, 63, 63, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61,
+        61, 61, 61, 61, 66, 67, 67, 67, 68, 67, 67, 66, 65, 65, 64, 64, 63, 63,
+        63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61,
+        66, 66, 66, 67, 67, 67, 66, 66, 65, 65, 64, 64, 63, 63, 63, 62, 62, 62,
+        62, 62, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 65, 65, 66, 66,
+        66, 66, 66, 65, 65, 64, 64, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61,
+        61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 64, 65, 65, 65, 66, 65, 65, 65,
+        64, 64, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 61,
+        60, 60, 60, 60, 60, 60, 64, 64, 65, 65, 65, 65, 65, 64, 64, 64, 63, 63,
+        63, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60,
+        60, 60, 63, 64, 64, 64, 65, 64, 64, 64, 64, 63, 63, 63, 62, 62, 62, 62,
+        61, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 63, 63,
+        64, 64, 64, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61,
+        61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 63, 63, 63, 63, 64, 64,
+        63, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 60, 60,
+        60, 60, 60, 60, 60, 60, 60, 60, 62, 63, 63, 63, 63, 63, 63, 63, 63, 62,
+        62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 62, 62, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 61,
+        61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        62, 62, 62, 62, 63, 63, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61,
+        61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 62, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 62, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 61,
+        61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61,
+        61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 58, 57, 55, 58, 56, 56, 55, 57, 56, 54, 53, 55, 55, 53, 53,
+        /* Size 8 */
+        64, 68, 59, 59, 58, 57, 56, 55, 68, 62, 59, 60, 60, 59, 58, 57, 59, 59,
+        58, 58, 58, 57, 57, 56, 59, 60, 58, 57, 56, 56, 56, 55, 58, 60, 58, 56,
+        56, 55, 55, 55, 57, 59, 57, 56, 55, 55, 54, 54, 56, 58, 57, 56, 55, 54,
+        54, 54, 55, 57, 56, 55, 55, 54, 54, 53,
+        /* Size 16 */
+        64, 66, 68, 63, 59, 59, 59, 58, 58, 58, 57, 57, 56, 56, 55, 55, 66, 65,
+        65, 62, 59, 59, 59, 59, 59, 58, 58, 57, 57, 56, 56, 56, 68, 65, 62, 60,
+        59, 60, 60, 60, 60, 59, 59, 58, 58, 57, 57, 57, 63, 62, 60, 59, 58, 59,
+        59, 59, 59, 58, 58, 58, 57, 57, 56, 56, 59, 59, 59, 58, 58, 58, 58, 58,
+        58, 58, 57, 57, 57, 56, 56, 56, 59, 59, 60, 59, 58, 58, 57, 57, 57, 57,
+        57, 57, 56, 56, 56, 56, 59, 59, 60, 59, 58, 57, 57, 57, 56, 56, 56, 56,
+        56, 55, 55, 55, 58, 59, 60, 59, 58, 57, 57, 56, 56, 56, 56, 55, 55, 55,
+        55, 55, 58, 59, 60, 59, 58, 57, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55,
+        58, 58, 59, 58, 58, 57, 56, 56, 55, 55, 55, 55, 55, 54, 54, 54, 57, 58,
+        59, 58, 57, 57, 56, 56, 55, 55, 55, 54, 54, 54, 54, 54, 57, 57, 58, 58,
+        57, 57, 56, 55, 55, 55, 54, 54, 54, 54, 54, 54, 56, 57, 58, 57, 57, 56,
+        56, 55, 55, 55, 54, 54, 54, 54, 54, 54, 56, 56, 57, 57, 56, 56, 55, 55,
+        55, 54, 54, 54, 54, 54, 53, 53, 55, 56, 57, 56, 56, 56, 55, 55, 55, 54,
+        54, 54, 54, 53, 53, 53, 55, 56, 57, 56, 56, 56, 55, 55, 55, 54, 54, 54,
+        54, 53, 53, 53,
+        /* Size 32 */
+        64, 65, 66, 67, 68, 66, 63, 61, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58,
+        58, 57, 57, 57, 57, 56, 56, 56, 56, 55, 55, 55, 55, 55, 65, 65, 66, 66,
+        66, 64, 63, 61, 59, 59, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 57, 57,
+        57, 57, 56, 56, 56, 56, 56, 56, 56, 56, 66, 66, 65, 65, 65, 63, 62, 61,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57,
+        56, 56, 56, 56, 56, 56, 67, 66, 65, 64, 63, 62, 61, 60, 59, 59, 59, 60,
+        60, 60, 60, 59, 59, 59, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 56, 56,
+        56, 56, 68, 66, 65, 63, 62, 61, 60, 60, 59, 59, 60, 60, 60, 60, 60, 60,
+        60, 60, 59, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 57, 57, 57, 66, 64,
+        63, 62, 61, 61, 60, 59, 59, 59, 59, 59, 60, 59, 59, 59, 59, 59, 59, 59,
+        58, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56, 56, 63, 63, 62, 61, 60, 60,
+        59, 59, 58, 58, 59, 59, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 57,
+        57, 57, 57, 57, 56, 56, 56, 56, 61, 61, 61, 60, 60, 59, 59, 58, 58, 58,
+        58, 58, 58, 58, 58, 58, 58, 58, 58, 58, 58, 58, 57, 57, 57, 57, 57, 56,
+        56, 56, 56, 56, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58, 58, 58,
+        58, 58, 58, 58, 58, 58, 57, 57, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56,
+        59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58, 58, 58, 58, 58, 58, 57,
+        57, 57, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 56, 59, 59, 59, 59,
+        60, 59, 59, 58, 58, 58, 58, 57, 57, 57, 57, 57, 57, 57, 57, 57, 57, 57,
+        57, 56, 56, 56, 56, 56, 56, 56, 56, 56, 59, 59, 59, 60, 60, 59, 59, 58,
+        58, 58, 57, 57, 57, 57, 57, 57, 57, 57, 57, 57, 56, 56, 56, 56, 56, 56,
+        56, 56, 55, 55, 55, 55, 59, 59, 59, 60, 60, 60, 59, 58, 58, 58, 57, 57,
+        57, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55,
+        55, 55, 58, 59, 59, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 57, 56, 56,
+        56, 56, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 58, 59,
+        59, 60, 60, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56,
+        56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 58, 59, 59, 59, 60, 59,
+        59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55,
+        55, 55, 55, 55, 55, 55, 55, 55, 58, 58, 59, 59, 60, 59, 59, 58, 58, 58,
+        57, 57, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55,
+        55, 55, 55, 55, 58, 58, 59, 59, 60, 59, 59, 58, 58, 57, 57, 57, 56, 56,
+        56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54,
+        58, 58, 58, 59, 59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56, 55, 55,
+        55, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 57, 58, 58, 59,
+        59, 59, 58, 58, 58, 57, 57, 57, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55,
+        55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 57, 57, 58, 58, 59, 58, 58, 58,
+        57, 57, 57, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54,
+        54, 54, 54, 54, 54, 54, 57, 57, 58, 58, 58, 58, 58, 58, 57, 57, 57, 56,
+        56, 56, 56, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 54,
+        54, 54, 57, 57, 57, 58, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56, 55, 55,
+        55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 56, 57,
+        57, 58, 58, 58, 57, 57, 57, 57, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55,
+        54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54, 56, 56, 57, 57, 58, 57,
+        57, 57, 57, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54,
+        54, 54, 54, 54, 54, 54, 54, 54, 56, 56, 57, 57, 57, 57, 57, 57, 57, 56,
+        56, 56, 56, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 54, 54,
+        53, 53, 53, 53, 56, 56, 56, 57, 57, 57, 57, 57, 56, 56, 56, 56, 55, 55,
+        55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 54, 53, 53, 53, 53, 53,
+        55, 56, 56, 57, 57, 57, 57, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 54,
+        54, 54, 54, 54, 54, 54, 54, 54, 53, 53, 53, 53, 53, 53, 55, 56, 56, 56,
+        57, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54,
+        54, 54, 54, 53, 53, 53, 53, 53, 53, 53, 55, 56, 56, 56, 57, 56, 56, 56,
+        56, 56, 56, 55, 55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 53,
+        53, 53, 53, 53, 53, 53, 55, 56, 56, 56, 57, 56, 56, 56, 56, 56, 56, 55,
+        55, 55, 55, 55, 55, 54, 54, 54, 54, 54, 54, 54, 54, 53, 53, 53, 53, 53,
+        53, 53, 55, 56, 56, 56, 57, 56, 56, 56, 56, 56, 56, 55, 55, 55, 55, 55,
+        55, 54, 54, 54, 54, 54, 54, 54, 54, 53, 53, 53, 53, 53, 53, 53 },
+      { /* Intra matrices */
+        /* Size 4 */
+        74, 67, 66, 63, 67, 65, 64, 63, 66, 64, 62, 61, 63, 63, 61, 60,
+        /* Size 8 */
+        72, 77, 67, 66, 65, 64, 63, 62, 77, 70, 66, 68, 67, 66, 65, 64, 67, 66,
+        65, 65, 65, 65, 64, 63, 66, 68, 65, 64, 63, 63, 63, 62, 65, 67, 65, 63,
+        62, 62, 62, 61, 64, 66, 65, 63, 62, 61, 61, 60, 63, 65, 64, 63, 62, 61,
+        60, 60, 62, 64, 63, 62, 61, 60, 60, 60,
+        /* Size 16 */
+        73, 75, 77, 72, 67, 67, 66, 66, 66, 65, 64, 64, 63, 63, 62, 62, 75, 74,
+        74, 70, 67, 67, 67, 67, 67, 66, 65, 65, 64, 64, 63, 63, 77, 74, 70, 69,
+        67, 67, 68, 68, 68, 67, 67, 66, 65, 65, 64, 64, 72, 70, 69, 67, 66, 66,
+        67, 67, 67, 66, 66, 65, 65, 64, 64, 64, 67, 67, 67, 66, 65, 65, 65, 65,
+        65, 65, 65, 64, 64, 64, 63, 63, 67, 67, 67, 66, 65, 65, 65, 65, 65, 64,
+        64, 64, 63, 63, 63, 63, 66, 67, 68, 67, 65, 65, 64, 64, 64, 64, 63, 63,
+        63, 63, 62, 62, 66, 67, 68, 67, 65, 65, 64, 64, 63, 63, 63, 63, 62, 62,
+        62, 62, 66, 67, 68, 67, 65, 65, 64, 63, 63, 62, 62, 62, 62, 62, 61, 61,
+        65, 66, 67, 66, 65, 64, 64, 63, 62, 62, 62, 62, 61, 61, 61, 61, 64, 65,
+        67, 66, 65, 64, 63, 63, 62, 62, 62, 61, 61, 61, 61, 61, 64, 65, 66, 65,
+        64, 64, 63, 63, 62, 62, 61, 61, 61, 61, 60, 60, 63, 64, 65, 65, 64, 63,
+        63, 62, 62, 61, 61, 61, 61, 60, 60, 60, 63, 64, 65, 64, 64, 63, 63, 62,
+        62, 61, 61, 61, 60, 60, 60, 60, 62, 63, 64, 64, 63, 63, 62, 62, 61, 61,
+        61, 60, 60, 60, 60, 60, 62, 63, 64, 64, 63, 63, 62, 62, 61, 61, 61, 60,
+        60, 60, 60, 60,
+        /* Size 32 */
+        73, 74, 75, 76, 77, 75, 72, 70, 67, 67, 67, 67, 66, 66, 66, 66, 66, 65,
+        65, 65, 65, 64, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 74, 74, 75, 75,
+        76, 73, 71, 69, 67, 67, 67, 67, 67, 67, 67, 66, 66, 66, 66, 65, 65, 65,
+        65, 64, 64, 64, 63, 63, 63, 63, 63, 63, 75, 75, 74, 74, 74, 72, 70, 69,
+        67, 67, 67, 67, 67, 67, 67, 67, 67, 66, 66, 66, 66, 65, 65, 65, 64, 64,
+        64, 64, 63, 63, 63, 63, 76, 75, 74, 73, 72, 71, 70, 68, 67, 67, 67, 68,
+        68, 68, 68, 67, 67, 67, 67, 66, 66, 66, 65, 65, 65, 65, 64, 64, 64, 64,
+        64, 64, 77, 76, 74, 72, 70, 70, 69, 68, 67, 67, 68, 68, 68, 68, 68, 68,
+        68, 68, 67, 67, 67, 66, 66, 66, 65, 65, 65, 64, 64, 64, 64, 64, 75, 73,
+        72, 71, 70, 69, 68, 67, 67, 67, 67, 67, 68, 67, 67, 67, 67, 67, 67, 67,
+        66, 66, 66, 65, 65, 65, 64, 64, 64, 64, 64, 64, 72, 71, 70, 70, 69, 68,
+        67, 67, 66, 66, 66, 67, 67, 67, 67, 67, 67, 67, 66, 66, 66, 66, 65, 65,
+        65, 65, 64, 64, 64, 64, 64, 64, 70, 69, 69, 68, 68, 67, 67, 66, 66, 66,
+        66, 66, 66, 66, 66, 66, 66, 66, 66, 66, 65, 65, 65, 65, 64, 64, 64, 64,
+        64, 64, 64, 64, 67, 67, 67, 67, 67, 67, 66, 66, 65, 65, 65, 65, 66, 66,
+        66, 66, 66, 65, 65, 65, 65, 65, 65, 64, 64, 64, 64, 64, 63, 63, 63, 63,
+        67, 67, 67, 67, 67, 67, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 65, 65, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 67, 67, 67, 67,
+        68, 67, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 64, 64, 64, 64,
+        64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 67, 67, 67, 68, 68, 67, 67, 66,
+        65, 65, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 66, 67, 67, 68, 68, 68, 67, 66, 66, 65, 65, 65,
+        64, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62,
+        62, 62, 66, 67, 67, 68, 68, 67, 67, 66, 66, 65, 65, 65, 64, 64, 64, 64,
+        64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 66, 67,
+        67, 68, 68, 67, 67, 66, 66, 65, 65, 64, 64, 64, 64, 64, 63, 63, 63, 63,
+        63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 66, 66, 67, 67, 68, 67,
+        67, 66, 66, 65, 65, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 66, 66, 67, 67, 68, 67, 67, 66, 66, 65,
+        65, 64, 64, 64, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 62, 65, 66, 66, 67, 68, 67, 67, 66, 65, 65, 65, 64, 64, 64,
+        63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61,
+        65, 66, 66, 67, 67, 67, 66, 66, 65, 65, 64, 64, 64, 63, 63, 63, 63, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 65, 65, 66, 66,
+        67, 67, 66, 66, 65, 65, 64, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 62,
+        62, 62, 61, 61, 61, 61, 61, 61, 61, 61, 65, 65, 66, 66, 67, 66, 66, 65,
+        65, 65, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61,
+        61, 61, 61, 61, 61, 61, 64, 65, 65, 66, 66, 66, 66, 65, 65, 64, 64, 64,
+        63, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 61, 61,
+        61, 61, 64, 65, 65, 65, 66, 66, 65, 65, 65, 64, 64, 64, 63, 63, 63, 62,
+        62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 64, 64,
+        65, 65, 66, 65, 65, 65, 64, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 62,
+        61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 63, 64, 64, 65, 65, 65,
+        65, 64, 64, 64, 64, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61,
+        61, 61, 61, 60, 60, 60, 60, 60, 63, 64, 64, 65, 65, 65, 65, 64, 64, 64,
+        63, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 60, 60,
+        60, 60, 60, 60, 63, 63, 64, 64, 65, 64, 64, 64, 64, 64, 63, 63, 63, 62,
+        62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60,
+        63, 63, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 61,
+        61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 62, 63, 63, 64,
+        64, 64, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61,
+        61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 62, 63, 63, 64, 64, 64, 64, 64,
+        63, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 60, 60,
+        60, 60, 60, 60, 60, 60, 62, 63, 63, 64, 64, 64, 64, 64, 63, 63, 63, 63,
+        62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60,
+        60, 60, 62, 63, 63, 64, 64, 64, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62,
+        62, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 63, 60, 59, 63, 61, 59, 59, 60, 59, 58, 58, 59, 59, 58, 58,
+        /* Size 8 */
+        64, 66, 66, 64, 62, 61, 60, 60, 66, 65, 65, 64, 63, 61, 61, 60, 66, 65,
+        63, 62, 61, 61, 60, 60, 64, 64, 62, 61, 61, 60, 60, 59, 62, 63, 61, 61,
+        60, 60, 59, 59, 61, 61, 61, 60, 60, 59, 59, 59, 60, 61, 60, 60, 59, 59,
+        59, 59, 60, 60, 60, 59, 59, 59, 59, 59,
+        /* Size 16 */
+        64, 65, 66, 66, 66, 65, 64, 63, 62, 61, 61, 60, 60, 60, 60, 60, 65, 65,
+        66, 66, 66, 65, 64, 63, 62, 62, 61, 61, 60, 60, 60, 60, 66, 66, 65, 65,
+        65, 65, 64, 63, 63, 62, 61, 61, 61, 60, 60, 60, 66, 66, 65, 65, 64, 64,
+        63, 63, 62, 62, 61, 61, 60, 60, 60, 60, 66, 66, 65, 64, 63, 63, 62, 62,
+        61, 61, 61, 60, 60, 60, 60, 60, 65, 65, 65, 64, 63, 62, 62, 61, 61, 61,
+        60, 60, 60, 60, 60, 60, 64, 64, 64, 63, 62, 62, 61, 61, 61, 60, 60, 60,
+        60, 60, 59, 59, 63, 63, 63, 63, 62, 61, 61, 61, 60, 60, 60, 60, 60, 59,
+        59, 59, 62, 62, 63, 62, 61, 61, 61, 60, 60, 60, 60, 60, 59, 59, 59, 59,
+        61, 62, 62, 62, 61, 61, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 61, 61,
+        61, 61, 61, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 60, 61, 61, 61,
+        60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 60, 60, 61, 60, 60, 60,
+        60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 60, 60, 60, 60, 60, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59,
+        /* Size 32 */
+        64, 65, 65, 66, 66, 66, 66, 66, 66, 65, 65, 64, 64, 63, 63, 62, 62, 62,
+        61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 65, 65, 65, 66,
+        66, 66, 66, 66, 66, 65, 65, 64, 64, 63, 63, 62, 62, 62, 62, 61, 61, 61,
+        61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 65, 65, 65, 66, 66, 66, 66, 66,
+        66, 65, 65, 64, 64, 63, 63, 63, 62, 62, 62, 61, 61, 61, 61, 61, 60, 60,
+        60, 60, 60, 60, 60, 60, 66, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 64,
+        64, 64, 63, 63, 62, 62, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60,
+        60, 60, 66, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 65, 64, 64, 63, 63,
+        63, 62, 62, 62, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 66, 66,
+        66, 65, 65, 65, 65, 65, 65, 65, 64, 64, 64, 63, 63, 63, 62, 62, 62, 62,
+        61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 66, 66, 66, 65, 65, 65,
+        65, 65, 64, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 61, 61, 61, 61, 61,
+        60, 60, 60, 60, 60, 60, 60, 60, 66, 66, 66, 65, 65, 65, 65, 64, 64, 63,
+        63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 66, 66, 66, 65, 65, 65, 64, 64, 63, 63, 63, 62, 62, 62,
+        62, 62, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        65, 65, 65, 65, 65, 65, 64, 63, 63, 63, 62, 62, 62, 62, 62, 61, 61, 61,
+        61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 65, 65, 65, 65,
+        65, 64, 64, 63, 63, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 60, 60,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 64, 64, 64, 64, 65, 64, 63, 63,
+        62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 64, 64, 64, 64, 64, 64, 63, 63, 62, 62, 62, 61,
+        61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59,
+        59, 59, 63, 63, 63, 64, 64, 63, 63, 62, 62, 62, 61, 61, 61, 61, 61, 61,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 63, 63,
+        63, 63, 63, 63, 63, 62, 62, 62, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 62, 62, 63, 63, 63, 63,
+        62, 62, 62, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        60, 59, 59, 59, 59, 59, 59, 59, 62, 62, 62, 62, 63, 62, 62, 62, 61, 61,
+        61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59,
+        59, 59, 59, 59, 62, 62, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 60, 60,
+        60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        61, 62, 62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 61, 61, 61, 62,
+        62, 62, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 61, 61, 61, 61, 61, 61, 61, 61,
+        61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 61, 61, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 60, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60,
+        60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60,
+        61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 60, 60, 61, 60,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59 },
+      { /* Intra matrices */
+        /* Size 4 */
+        69, 68, 65, 63, 68, 65, 64, 63, 65, 64, 62, 62, 63, 63, 62, 62,
+        /* Size 8 */
+        67, 70, 69, 67, 65, 64, 63, 62, 70, 69, 69, 68, 66, 64, 63, 63, 69, 69,
+        66, 65, 64, 64, 63, 63, 67, 68, 65, 64, 63, 63, 63, 62, 65, 66, 64, 63,
+        63, 62, 62, 62, 64, 64, 64, 63, 62, 62, 62, 62, 63, 63, 63, 63, 62, 62,
+        62, 62, 62, 63, 63, 62, 62, 62, 62, 62,
+        /* Size 16 */
+        68, 69, 70, 70, 69, 68, 67, 66, 65, 65, 64, 64, 63, 63, 63, 63, 69, 69,
+        69, 69, 69, 68, 67, 66, 66, 65, 64, 64, 63, 63, 63, 63, 70, 69, 69, 69,
+        69, 68, 68, 67, 66, 65, 65, 64, 64, 63, 63, 63, 70, 69, 69, 68, 68, 67,
+        67, 66, 65, 65, 64, 64, 63, 63, 63, 63, 69, 69, 69, 68, 66, 66, 65, 65,
+        65, 64, 64, 64, 63, 63, 63, 63, 68, 68, 68, 67, 66, 65, 65, 65, 64, 64,
+        64, 63, 63, 63, 63, 63, 67, 67, 68, 67, 65, 65, 64, 64, 64, 63, 63, 63,
+        63, 63, 62, 62, 66, 66, 67, 66, 65, 65, 64, 64, 63, 63, 63, 63, 63, 62,
+        62, 62, 65, 66, 66, 65, 65, 64, 64, 63, 63, 63, 63, 63, 62, 62, 62, 62,
+        65, 65, 65, 65, 64, 64, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 64, 64,
+        65, 64, 64, 64, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 64, 64, 64, 64,
+        64, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 63, 63, 64, 63, 63, 63,
+        63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 63, 63, 63, 63, 63, 63, 63, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 62,
+        /* Size 32 */
+        68, 68, 69, 70, 70, 70, 70, 70, 69, 69, 68, 68, 67, 67, 66, 66, 65, 65,
+        65, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 68, 69, 69, 69,
+        70, 70, 70, 69, 69, 69, 68, 68, 67, 67, 66, 66, 65, 65, 65, 65, 64, 64,
+        64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 69, 69, 69, 69, 70, 69, 69, 69,
+        69, 69, 68, 68, 67, 67, 67, 66, 66, 65, 65, 65, 64, 64, 64, 64, 63, 63,
+        63, 63, 63, 63, 63, 63, 70, 69, 69, 69, 69, 69, 69, 69, 69, 69, 68, 68,
+        68, 67, 67, 66, 66, 66, 65, 65, 65, 64, 64, 64, 64, 63, 63, 63, 63, 63,
+        63, 63, 70, 70, 70, 69, 69, 69, 69, 69, 69, 69, 69, 68, 68, 67, 67, 67,
+        66, 66, 65, 65, 65, 65, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 70, 70,
+        69, 69, 69, 69, 69, 69, 69, 68, 68, 68, 67, 67, 67, 66, 66, 65, 65, 65,
+        65, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 70, 70, 69, 69, 69, 69,
+        68, 68, 68, 68, 67, 67, 67, 66, 66, 66, 65, 65, 65, 65, 64, 64, 64, 64,
+        64, 63, 63, 63, 63, 63, 63, 63, 70, 69, 69, 69, 69, 69, 68, 68, 67, 67,
+        67, 66, 66, 66, 66, 65, 65, 65, 65, 64, 64, 64, 64, 64, 63, 63, 63, 63,
+        63, 63, 63, 63, 69, 69, 69, 69, 69, 69, 68, 67, 67, 66, 66, 66, 66, 65,
+        65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63,
+        69, 69, 69, 69, 69, 68, 68, 67, 66, 66, 66, 66, 65, 65, 65, 65, 65, 64,
+        64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 68, 68, 68, 68,
+        69, 68, 67, 67, 66, 66, 66, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 64,
+        63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 68, 68, 68, 68, 68, 68, 67, 66,
+        66, 66, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 67, 67, 67, 68, 68, 67, 67, 66, 66, 65, 65, 65,
+        64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 67, 67, 67, 67, 67, 67, 66, 66, 65, 65, 65, 65, 64, 64, 64, 64,
+        64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 66, 66,
+        67, 67, 67, 67, 66, 66, 65, 65, 65, 64, 64, 64, 64, 64, 63, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 66, 66, 66, 66, 67, 66,
+        66, 65, 65, 65, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 62, 62, 62, 62, 62, 62, 65, 65, 66, 66, 66, 66, 65, 65, 65, 65,
+        64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62,
+        62, 62, 62, 62, 65, 65, 65, 66, 66, 65, 65, 65, 65, 64, 64, 64, 64, 63,
+        63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62,
+        65, 65, 65, 65, 65, 65, 65, 65, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63,
+        63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 64, 65, 65, 65,
+        65, 65, 65, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 64, 64, 64, 65, 65, 65, 64, 64,
+        64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 64, 64, 64, 64, 65, 64, 64, 64, 64, 64, 64, 63,
+        63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63,
+        63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 63, 64,
+        64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 63, 63, 63, 64, 64, 64,
+        64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 63, 63, 63, 63, 64, 64, 63, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 62, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62,
+        63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 63, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 61, 60, 59, 61, 60, 60, 59, 60, 60, 59, 58, 59, 59, 58, 58,
+        /* Size 8 */
+        64, 66, 61, 61, 61, 60, 60, 59, 66, 63, 61, 62, 62, 61, 61, 60, 61, 61,
+        61, 61, 61, 61, 60, 60, 61, 62, 61, 60, 60, 60, 60, 59, 61, 62, 61, 60,
+        60, 59, 59, 59, 60, 61, 61, 60, 59, 59, 59, 59, 60, 61, 60, 60, 59, 59,
+        59, 59, 59, 60, 60, 59, 59, 59, 59, 58,
+        /* Size 16 */
+        64, 65, 66, 64, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 59, 59, 65, 65,
+        64, 63, 61, 61, 62, 61, 61, 61, 61, 61, 60, 60, 60, 60, 66, 64, 63, 62,
+        61, 62, 62, 62, 62, 62, 61, 61, 61, 60, 60, 60, 64, 63, 62, 62, 61, 61,
+        61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 61, 61, 61, 61, 61, 61, 61, 61,
+        61, 61, 61, 60, 60, 60, 60, 60, 61, 61, 62, 61, 61, 61, 60, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 61, 62, 62, 61, 61, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 59, 59, 61, 61, 62, 61, 61, 60, 60, 60, 60, 60, 60, 60, 59, 59,
+        59, 59, 61, 61, 62, 61, 61, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59,
+        61, 61, 62, 61, 61, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 60, 61,
+        61, 61, 61, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 60, 61, 61, 61,
+        60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 61, 60, 60, 60,
+        60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 60, 60, 60, 60, 60, 59,
+        59, 59, 59, 59, 59, 59, 58, 58, 59, 60, 60, 60, 60, 60, 59, 59, 59, 59,
+        59, 59, 59, 58, 58, 58, 59, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59,
+        59, 58, 58, 58,
+        /* Size 32 */
+        64, 64, 65, 65, 66, 65, 64, 63, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61,
+        61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 64, 65, 65, 65,
+        65, 64, 63, 62, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 60,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 65, 65, 65, 65, 64, 64, 63, 62,
+        61, 61, 61, 62, 62, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 65, 65, 65, 64, 64, 63, 63, 62, 61, 61, 62, 62,
+        62, 62, 62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60,
+        60, 60, 66, 65, 64, 64, 63, 63, 62, 62, 61, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 65, 64,
+        64, 63, 63, 62, 62, 62, 61, 61, 61, 62, 62, 62, 62, 62, 62, 61, 61, 61,
+        61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 64, 63, 63, 63, 62, 62,
+        62, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61,
+        60, 60, 60, 60, 60, 60, 60, 60, 63, 62, 62, 62, 62, 62, 61, 61, 61, 61,
+        61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61,
+        61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        61, 61, 61, 61, 62, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61, 61,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 61, 61, 61, 62,
+        62, 61, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 61, 61, 62, 62, 62, 62, 61, 61,
+        61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 59, 59, 59, 59, 61, 61, 62, 62, 62, 62, 61, 61, 61, 61, 60, 60,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59,
+        59, 59, 61, 61, 61, 62, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 61, 61,
+        61, 62, 62, 62, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 61, 61, 61, 62, 62, 62,
+        61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 61, 61, 61, 62, 62, 62, 61, 61, 61, 61,
+        60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 61, 61, 61, 61, 62, 61, 61, 61, 61, 61, 60, 60, 60, 60,
+        60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        61, 61, 61, 61, 62, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 61, 61, 61,
+        61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 61, 61, 61, 61, 61, 61, 61,
+        61, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 60, 60, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60,
+        60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 60, 60, 61, 61, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60,
+        60, 61, 61, 61, 61, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 60, 60, 61, 61,
+        60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 60, 60, 60, 60, 61, 60, 60, 60, 60, 60,
+        60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        58, 58, 58, 58, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58,
+        60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58, 59, 60, 60, 60,
+        60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59,
+        59, 59, 59, 58, 58, 58, 58, 58, 58, 58, 59, 60, 60, 60, 60, 60, 60, 60,
+        60, 60, 60, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 58,
+        58, 58, 58, 58, 58, 58, 59, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58,
+        58, 58, 59, 60, 60, 60, 60, 60, 60, 60, 60, 60, 60, 59, 59, 59, 59, 59,
+        59, 59, 59, 59, 59, 59, 59, 59, 59, 58, 58, 58, 58, 58, 58, 58 },
+      { /* Intra matrices */
+        /* Size 4 */
+        69, 65, 65, 64, 65, 64, 64, 63, 65, 64, 63, 62, 64, 63, 62, 62,
+        /* Size 8 */
+        68, 70, 65, 65, 65, 64, 63, 63, 70, 67, 65, 66, 66, 65, 64, 64, 65, 65,
+        64, 64, 65, 64, 64, 63, 65, 66, 64, 64, 64, 63, 63, 63, 65, 66, 65, 64,
+        63, 63, 63, 63, 64, 65, 64, 63, 63, 63, 62, 62, 63, 64, 64, 63, 63, 62,
+        62, 62, 63, 64, 63, 63, 63, 62, 62, 62,
+        /* Size 16 */
+        68, 69, 70, 68, 65, 65, 65, 65, 65, 64, 64, 64, 64, 63, 63, 63, 69, 69,
+        69, 67, 65, 65, 66, 65, 65, 65, 65, 64, 64, 64, 64, 64, 70, 69, 67, 66,
+        65, 66, 66, 66, 66, 65, 65, 65, 65, 64, 64, 64, 68, 67, 66, 66, 65, 65,
+        65, 65, 65, 65, 65, 65, 64, 64, 64, 64, 65, 65, 65, 65, 64, 65, 65, 65,
+        65, 65, 64, 64, 64, 64, 64, 64, 65, 65, 66, 65, 65, 64, 64, 64, 64, 64,
+        64, 64, 64, 63, 63, 63, 65, 66, 66, 65, 65, 64, 64, 64, 64, 64, 64, 63,
+        63, 63, 63, 63, 65, 65, 66, 65, 65, 64, 64, 64, 64, 63, 63, 63, 63, 63,
+        63, 63, 65, 65, 66, 65, 65, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63,
+        64, 65, 65, 65, 65, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 64, 65,
+        65, 65, 64, 64, 64, 63, 63, 63, 63, 63, 63, 62, 62, 62, 64, 64, 65, 65,
+        64, 64, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 64, 64, 65, 64, 64, 64,
+        63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 63, 64, 64, 64, 64, 63, 63, 63,
+        63, 63, 62, 62, 62, 62, 62, 62, 63, 64, 64, 64, 64, 63, 63, 63, 63, 63,
+        62, 62, 62, 62, 62, 62, 63, 64, 64, 64, 64, 63, 63, 63, 63, 63, 62, 62,
+        62, 62, 62, 62,
+        /* Size 32 */
+        68, 69, 69, 70, 71, 69, 68, 67, 66, 65, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 69, 69, 69, 69,
+        70, 69, 68, 67, 66, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 64, 64,
+        64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 69, 69, 69, 69, 69, 68, 67, 66,
+        65, 66, 66, 66, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 70, 69, 69, 68, 68, 67, 67, 66, 65, 66, 66, 66,
+        66, 66, 66, 66, 66, 65, 65, 65, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64,
+        64, 64, 71, 70, 69, 68, 67, 67, 66, 66, 65, 66, 66, 66, 66, 66, 66, 66,
+        66, 66, 66, 65, 65, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 64, 69, 69,
+        68, 67, 67, 66, 66, 66, 65, 65, 65, 66, 66, 66, 66, 66, 66, 65, 65, 65,
+        65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 64, 64, 68, 68, 67, 67, 66, 66,
+        66, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 67, 67, 66, 66, 66, 66, 65, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65,
+        65, 65, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        65, 65, 66, 66, 66, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 65, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 65, 65, 66, 66,
+        66, 65, 65, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 65, 65, 66, 66, 66, 66, 65, 65,
+        65, 65, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        63, 63, 63, 63, 63, 63, 65, 65, 66, 66, 66, 66, 65, 65, 65, 65, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63,
+        63, 63, 65, 65, 66, 66, 66, 66, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 65, 65,
+        65, 66, 66, 66, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 64, 64, 64, 63,
+        63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 65, 65, 65, 66, 66, 66,
+        65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 63, 63, 65, 65, 65, 66, 66, 66, 65, 65, 65, 65,
+        64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 63, 63, 65, 65, 65, 65, 66, 65, 65, 65, 65, 64, 64, 64, 64, 64,
+        64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63,
+        65, 65, 65, 65, 66, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 64, 65, 65, 65,
+        65, 65, 65, 65, 65, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 64, 64, 65, 65, 65, 65, 65, 65,
+        64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 62, 62, 62, 62, 62, 64, 64, 65, 65, 65, 65, 65, 65, 64, 64, 64, 64,
+        64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62,
+        62, 62, 64, 64, 64, 65, 65, 65, 65, 64, 64, 64, 64, 64, 64, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 64, 64,
+        64, 65, 65, 65, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 64, 64, 64, 64, 65, 64,
+        64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 62, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63,
+        63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62,
+        63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63,
+        63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 63, 63, 64, 64,
+        64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 63, 63, 64, 64, 64, 64, 64, 64,
+        64, 64, 63, 63, 63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 62, 62, 62, 62, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 63, 63,
+        63, 63, 63, 63, 63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62,
+        62, 62, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 63, 63,
+        63, 63, 63, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62, 62 } } },
+  { {   /* Luma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 8 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 16 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64,
+        /* Size 32 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 },
+      { /* Intra matrices */
+        /* Size 4 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 8 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 16 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64,
+        /* Size 32 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 } },
+    {   /* Chroma matrices */
+      { /* Inter matrices */
+        /* Size 4 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 8 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 16 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64,
+        /* Size 32 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 },
+      { /* Intra matrices */
+        /* Size 4 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 8 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        /* Size 16 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64,
+        /* Size 32 */
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+        64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64 } } }
+};
+
+#endif
+
+#if CONFIG_PVQ || CONFIG_DAALA_DIST
+/* Quantization matrices for 8x8. For other block sizes, we currently just do
+   resampling. */
+/* Flat quantization, i.e. optimize for PSNR. */
+const int OD_QM8_Q4_FLAT[] = { 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
+                               16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
+                               16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
+                               16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
+                               16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
+                               16, 16, 16, 16, 16, 16, 16, 16, 16 };
+#if 0
+/* M1: MPEG2 matrix for inter (which has a dead zone). */
+const int OD_QM8_Q4[] = {
+  16, 17, 18, 19, 20, 21, 22, 23,
+  17, 18, 19, 20, 21, 22, 23, 24,
+  18, 19, 20, 21, 22, 23, 24, 25,
+  19, 20, 21, 22, 23, 24, 26, 27,
+  20, 21, 22, 23, 25, 26, 27, 28,
+  21, 22, 23, 24, 26, 27, 28, 30,
+  22, 23, 24, 26, 27, 28, 30, 31,
+  23, 24, 25, 27, 28, 30, 31, 33};
+#endif
+#if 0
+/* M2: MPEG2 matrix for intra (no dead zone). */
+const int OD_QM8_Q4[] = {
+  16, 16, 19, 22, 22, 26, 26, 27,
+  16, 16, 22, 22, 26, 27, 27, 29,
+  19, 22, 26, 26, 27, 29, 29, 35,
+  22, 24, 27, 27, 29, 32, 34, 38,
+  26, 27, 29, 29, 32, 35, 38, 46,
+  27, 29, 34, 34, 35, 40, 46, 56,
+  29, 34, 34, 37, 40, 48, 56, 69,
+  34, 37, 38, 40, 48, 58, 69, 83
+};
+#endif
+#if 0
+/* M3: Taken from dump_psnrhvs. */
+const int OD_QM8_Q4[] = {
+  16, 16, 17, 20, 24, 29, 36, 42,
+  16, 17, 17, 19, 22, 26, 31, 37,
+  17, 17, 21, 23, 26, 30, 34, 40,
+  20, 19, 23, 28, 31, 35, 39, 45,
+  24, 22, 26, 31, 36, 41, 46, 51,
+  29, 26, 30, 35, 41, 47, 52, 58,
+  36, 31, 34, 39, 46, 52, 59, 66,
+  42, 37, 40, 45, 51, 58, 66, 73
+};
+#endif
+#if 1
+/* M4: a compromise equal to .5*(M3 + .5*(M2+transpose(M2))) */
+const int OD_QM8_Q4_HVS[] = { 16, 16, 18, 21, 24, 28, 32, 36, 16, 17, 20,
+                              21, 24, 27, 31, 35, 18, 20, 24, 25, 27, 31,
+                              33, 38, 21, 21, 25, 28, 30, 34, 37, 42, 24,
+                              24, 27, 30, 34, 38, 43, 49, 28, 27, 31, 34,
+                              38, 44, 50, 58, 32, 31, 33, 37, 43, 50, 58,
+                              68, 36, 35, 38, 42, 49, 58, 68, 78 };
+#endif
+#endif
diff --git a/third_party/aom/av1/common/quant_common.h b/third_party/aom/av1/common/quant_common.h
new file mode 100644
index 0000000000..3f442427de
--- /dev/null
+++ b/third_party/aom/av1/common/quant_common.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_QUANT_COMMON_H_
+#define AV1_COMMON_QUANT_COMMON_H_
+
+#include "aom/aom_codec.h"
+#include "av1/common/seg_common.h"
+#include "av1/common/enums.h"
+#include "av1/common/entropy.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MINQ 0
+#define MAXQ 255
+#define QINDEX_RANGE (MAXQ - MINQ + 1)
+#define QINDEX_BITS 8
+#if CONFIG_AOM_QM
+// Total number of QM sets stored
+#define QM_LEVEL_BITS 4
+#define NUM_QM_LEVELS (1 << QM_LEVEL_BITS)
+/* Offset into the list of QMs. Actual number of levels used is
+   (NUM_QM_LEVELS-AOM_QM_OFFSET)
+   Lower value of AOM_QM_OFFSET implies more heavily weighted matrices.*/
+#define DEFAULT_QM_FIRST (NUM_QM_LEVELS / 2)
+#define DEFAULT_QM_LAST (NUM_QM_LEVELS - 1)
+#endif
+
+struct AV1Common;
+
+int16_t av1_dc_quant(int qindex, int delta, aom_bit_depth_t bit_depth);
+int16_t av1_ac_quant(int qindex, int delta, aom_bit_depth_t bit_depth);
+int16_t av1_qindex_from_ac(int ac, aom_bit_depth_t bit_depth);
+
+int av1_get_qindex(const struct segmentation *seg, int segment_id,
+                   int base_qindex);
+#if CONFIG_AOM_QM
+// Reduce the large number of quantizers to a smaller number of levels for which
+// different matrices may be defined
+static INLINE int aom_get_qmlevel(int qindex, int first, int last) {
+  int qmlevel = (qindex * (last + 1 - first) + QINDEX_RANGE / 2) / QINDEX_RANGE;
+  qmlevel = AOMMIN(qmlevel + first, NUM_QM_LEVELS - 1);
+  return qmlevel;
+}
+void aom_qm_init(struct AV1Common *cm);
+qm_val_t *aom_iqmatrix(struct AV1Common *cm, int qindex, int comp,
+                       int log2sizem2, int is_intra);
+qm_val_t *aom_qmatrix(struct AV1Common *cm, int qindex, int comp,
+                      int log2sizem2, int is_intra);
+#endif
+
+#if CONFIG_NEW_QUANT
+
+#define QUANT_PROFILES 4
+#define QUANT_RANGES 2
+#define NUQ_KNOTS 3
+
+typedef tran_low_t dequant_val_type_nuq[NUQ_KNOTS + 1];
+typedef tran_low_t cuml_bins_type_nuq[NUQ_KNOTS];
+void av1_get_dequant_val_nuq(int q, int band, tran_low_t *dq,
+                             tran_low_t *cuml_bins, int dq_off_index);
+tran_low_t av1_dequant_abscoeff_nuq(int v, int q, const tran_low_t *dq);
+tran_low_t av1_dequant_coeff_nuq(int v, int q, const tran_low_t *dq);
+
+static INLINE int qindex_to_qrange(int qindex) {
+  return (qindex < 140 ? 1 : 0);
+}
+
+static INLINE int get_dq_profile_from_ctx(int qindex, int q_ctx, int is_inter,
+                                          PLANE_TYPE plane_type) {
+  // intra/inter, Y/UV, ctx, qrange
+  static const int
+      def_dq_profile_lookup[REF_TYPES][PLANE_TYPES][COEFF_CONTEXTS0]
+                           [QUANT_RANGES] = {
+                             {
+                                 // intra
+                                 { { 2, 1 }, { 2, 1 }, { 2, 1 } },  // Y
+                                 { { 3, 1 }, { 3, 1 }, { 3, 1 } },  // UV
+                             },
+                             {
+                                 // inter
+                                 { { 3, 1 }, { 2, 1 }, { 2, 1 } },  // Y
+                                 { { 3, 1 }, { 3, 1 }, { 3, 1 } },  // UV
+                             },
+                           };
+  if (!qindex) return 0;  // lossless
+  return def_dq_profile_lookup[is_inter][plane_type][q_ctx]
+                              [qindex_to_qrange(qindex)];
+}
+#endif  // CONFIG_NEW_QUANT
+
+#if CONFIG_PVQ || CONFIG_DAALA_DIST
+extern const int OD_QM8_Q4_FLAT[];
+extern const int OD_QM8_Q4_HVS[];
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_QUANT_COMMON_H_
diff --git a/third_party/aom/av1/common/reconinter.c b/third_party/aom/av1/common/reconinter.c
new file mode 100644
index 0000000000..ed7065757b
--- /dev/null
+++ b/third_party/aom/av1/common/reconinter.c
@@ -0,0 +1,3083 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "./aom_scale_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/blend.h"
+
+#include "av1/common/blockd.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+#if CONFIG_MOTION_VAR
+#include "av1/common/onyxc_int.h"
+#endif  // CONFIG_MOTION_VAR
+
+#if CONFIG_EXT_INTER
+
+#define NSMOOTHERS 1
+
+// [smoother][negative][direction]
+DECLARE_ALIGNED(16, static uint8_t,
+                wedge_mask_obl[NSMOOTHERS][2][WEDGE_DIRECTIONS]
+                              [MASK_MASTER_SIZE * MASK_MASTER_SIZE]);
+
+DECLARE_ALIGNED(16, static uint8_t,
+                wedge_signflip_lookup[BLOCK_SIZES][MAX_WEDGE_TYPES]);
+
+// 3 * MAX_WEDGE_SQUARE is an easy to compute and fairly tight upper bound
+// on the sum of all mask sizes up to an including MAX_WEDGE_SQUARE.
+DECLARE_ALIGNED(16, static uint8_t,
+                wedge_mask_buf[2 * MAX_WEDGE_TYPES * 3 * MAX_WEDGE_SQUARE]);
+
+static wedge_masks_type wedge_masks[BLOCK_SIZES][2];
+
+// Some unused wedge codebooks left temporarily to facilitate experiments.
+// To be removed when settled.
+/*
+static wedge_code_type wedge_codebook_8_hgtw[8] = {
+  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
+  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
+  { WEDGE_OBLIQUE27, 4, 2 },  { WEDGE_OBLIQUE27, 4, 6 },
+  { WEDGE_OBLIQUE153, 4, 2 }, { WEDGE_OBLIQUE153, 4, 6 },
+};
+
+static wedge_code_type wedge_codebook_8_hltw[8] = {
+  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
+  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
+  { WEDGE_OBLIQUE63, 2, 4 },  { WEDGE_OBLIQUE63, 6, 4 },
+  { WEDGE_OBLIQUE117, 2, 4 }, { WEDGE_OBLIQUE117, 6, 4 },
+};
+
+static wedge_code_type wedge_codebook_8_heqw[8] = {
+  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
+  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
+  { WEDGE_HORIZONTAL, 4, 2 }, { WEDGE_HORIZONTAL, 4, 6 },
+  { WEDGE_VERTICAL, 2, 4 },   { WEDGE_VERTICAL, 6, 4 },
+};
+
+static const wedge_code_type wedge_codebook_32_hgtw[32] = {
+  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
+  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
+  { WEDGE_HORIZONTAL, 4, 2 }, { WEDGE_HORIZONTAL, 4, 4 },
+  { WEDGE_HORIZONTAL, 4, 6 }, { WEDGE_VERTICAL, 4, 4 },
+  { WEDGE_OBLIQUE27, 4, 1 },  { WEDGE_OBLIQUE27, 4, 2 },
+  { WEDGE_OBLIQUE27, 4, 3 },  { WEDGE_OBLIQUE27, 4, 5 },
+  { WEDGE_OBLIQUE27, 4, 6 },  { WEDGE_OBLIQUE27, 4, 7 },
+  { WEDGE_OBLIQUE153, 4, 1 }, { WEDGE_OBLIQUE153, 4, 2 },
+  { WEDGE_OBLIQUE153, 4, 3 }, { WEDGE_OBLIQUE153, 4, 5 },
+  { WEDGE_OBLIQUE153, 4, 6 }, { WEDGE_OBLIQUE153, 4, 7 },
+  { WEDGE_OBLIQUE63, 1, 4 },  { WEDGE_OBLIQUE63, 2, 4 },
+  { WEDGE_OBLIQUE63, 3, 4 },  { WEDGE_OBLIQUE63, 5, 4 },
+  { WEDGE_OBLIQUE63, 6, 4 },  { WEDGE_OBLIQUE63, 7, 4 },
+  { WEDGE_OBLIQUE117, 1, 4 }, { WEDGE_OBLIQUE117, 2, 4 },
+  { WEDGE_OBLIQUE117, 3, 4 }, { WEDGE_OBLIQUE117, 5, 4 },
+  { WEDGE_OBLIQUE117, 6, 4 }, { WEDGE_OBLIQUE117, 7, 4 },
+};
+
+static const wedge_code_type wedge_codebook_32_hltw[32] = {
+  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
+  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
+  { WEDGE_VERTICAL, 2, 4 },   { WEDGE_VERTICAL, 4, 4 },
+  { WEDGE_VERTICAL, 6, 4 },   { WEDGE_HORIZONTAL, 4, 4 },
+  { WEDGE_OBLIQUE27, 4, 1 },  { WEDGE_OBLIQUE27, 4, 2 },
+  { WEDGE_OBLIQUE27, 4, 3 },  { WEDGE_OBLIQUE27, 4, 5 },
+  { WEDGE_OBLIQUE27, 4, 6 },  { WEDGE_OBLIQUE27, 4, 7 },
+  { WEDGE_OBLIQUE153, 4, 1 }, { WEDGE_OBLIQUE153, 4, 2 },
+  { WEDGE_OBLIQUE153, 4, 3 }, { WEDGE_OBLIQUE153, 4, 5 },
+  { WEDGE_OBLIQUE153, 4, 6 }, { WEDGE_OBLIQUE153, 4, 7 },
+  { WEDGE_OBLIQUE63, 1, 4 },  { WEDGE_OBLIQUE63, 2, 4 },
+  { WEDGE_OBLIQUE63, 3, 4 },  { WEDGE_OBLIQUE63, 5, 4 },
+  { WEDGE_OBLIQUE63, 6, 4 },  { WEDGE_OBLIQUE63, 7, 4 },
+  { WEDGE_OBLIQUE117, 1, 4 }, { WEDGE_OBLIQUE117, 2, 4 },
+  { WEDGE_OBLIQUE117, 3, 4 }, { WEDGE_OBLIQUE117, 5, 4 },
+  { WEDGE_OBLIQUE117, 6, 4 }, { WEDGE_OBLIQUE117, 7, 4 },
+};
+
+static const wedge_code_type wedge_codebook_32_heqw[32] = {
+  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
+  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
+  { WEDGE_HORIZONTAL, 4, 2 }, { WEDGE_HORIZONTAL, 4, 6 },
+  { WEDGE_VERTICAL, 2, 4 },   { WEDGE_VERTICAL, 6, 4 },
+  { WEDGE_OBLIQUE27, 4, 1 },  { WEDGE_OBLIQUE27, 4, 2 },
+  { WEDGE_OBLIQUE27, 4, 3 },  { WEDGE_OBLIQUE27, 4, 5 },
+  { WEDGE_OBLIQUE27, 4, 6 },  { WEDGE_OBLIQUE27, 4, 7 },
+  { WEDGE_OBLIQUE153, 4, 1 }, { WEDGE_OBLIQUE153, 4, 2 },
+  { WEDGE_OBLIQUE153, 4, 3 }, { WEDGE_OBLIQUE153, 4, 5 },
+  { WEDGE_OBLIQUE153, 4, 6 }, { WEDGE_OBLIQUE153, 4, 7 },
+  { WEDGE_OBLIQUE63, 1, 4 },  { WEDGE_OBLIQUE63, 2, 4 },
+  { WEDGE_OBLIQUE63, 3, 4 },  { WEDGE_OBLIQUE63, 5, 4 },
+  { WEDGE_OBLIQUE63, 6, 4 },  { WEDGE_OBLIQUE63, 7, 4 },
+  { WEDGE_OBLIQUE117, 1, 4 }, { WEDGE_OBLIQUE117, 2, 4 },
+  { WEDGE_OBLIQUE117, 3, 4 }, { WEDGE_OBLIQUE117, 5, 4 },
+  { WEDGE_OBLIQUE117, 6, 4 }, { WEDGE_OBLIQUE117, 7, 4 },
+};
+*/
+
+static const wedge_code_type wedge_codebook_16_hgtw[16] = {
+  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
+  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
+  { WEDGE_HORIZONTAL, 4, 2 }, { WEDGE_HORIZONTAL, 4, 4 },
+  { WEDGE_HORIZONTAL, 4, 6 }, { WEDGE_VERTICAL, 4, 4 },
+  { WEDGE_OBLIQUE27, 4, 2 },  { WEDGE_OBLIQUE27, 4, 6 },
+  { WEDGE_OBLIQUE153, 4, 2 }, { WEDGE_OBLIQUE153, 4, 6 },
+  { WEDGE_OBLIQUE63, 2, 4 },  { WEDGE_OBLIQUE63, 6, 4 },
+  { WEDGE_OBLIQUE117, 2, 4 }, { WEDGE_OBLIQUE117, 6, 4 },
+};
+
+static const wedge_code_type wedge_codebook_16_hltw[16] = {
+  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
+  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
+  { WEDGE_VERTICAL, 2, 4 },   { WEDGE_VERTICAL, 4, 4 },
+  { WEDGE_VERTICAL, 6, 4 },   { WEDGE_HORIZONTAL, 4, 4 },
+  { WEDGE_OBLIQUE27, 4, 2 },  { WEDGE_OBLIQUE27, 4, 6 },
+  { WEDGE_OBLIQUE153, 4, 2 }, { WEDGE_OBLIQUE153, 4, 6 },
+  { WEDGE_OBLIQUE63, 2, 4 },  { WEDGE_OBLIQUE63, 6, 4 },
+  { WEDGE_OBLIQUE117, 2, 4 }, { WEDGE_OBLIQUE117, 6, 4 },
+};
+
+static const wedge_code_type wedge_codebook_16_heqw[16] = {
+  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
+  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
+  { WEDGE_HORIZONTAL, 4, 2 }, { WEDGE_HORIZONTAL, 4, 6 },
+  { WEDGE_VERTICAL, 2, 4 },   { WEDGE_VERTICAL, 6, 4 },
+  { WEDGE_OBLIQUE27, 4, 2 },  { WEDGE_OBLIQUE27, 4, 6 },
+  { WEDGE_OBLIQUE153, 4, 2 }, { WEDGE_OBLIQUE153, 4, 6 },
+  { WEDGE_OBLIQUE63, 2, 4 },  { WEDGE_OBLIQUE63, 6, 4 },
+  { WEDGE_OBLIQUE117, 2, 4 }, { WEDGE_OBLIQUE117, 6, 4 },
+};
+
+const wedge_params_type wedge_params_lookup[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  { 0, NULL, NULL, 0, NULL },
+  { 0, NULL, NULL, 0, NULL },
+  { 0, NULL, NULL, 0, NULL },
+#endif  // CONFIG_CB4X4
+  { 0, NULL, NULL, 0, NULL },
+  { 0, NULL, NULL, 0, NULL },
+  { 0, NULL, NULL, 0, NULL },
+#if CONFIG_WEDGE
+  { 4, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_8X8], 0,
+    wedge_masks[BLOCK_8X8] },
+  { 4, wedge_codebook_16_hgtw, wedge_signflip_lookup[BLOCK_8X16], 0,
+    wedge_masks[BLOCK_8X16] },
+  { 4, wedge_codebook_16_hltw, wedge_signflip_lookup[BLOCK_16X8], 0,
+    wedge_masks[BLOCK_16X8] },
+  { 4, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_16X16], 0,
+    wedge_masks[BLOCK_16X16] },
+  { 4, wedge_codebook_16_hgtw, wedge_signflip_lookup[BLOCK_16X32], 0,
+    wedge_masks[BLOCK_16X32] },
+  { 4, wedge_codebook_16_hltw, wedge_signflip_lookup[BLOCK_32X16], 0,
+    wedge_masks[BLOCK_32X16] },
+  { 4, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_32X32], 0,
+    wedge_masks[BLOCK_32X32] },
+  { 0, wedge_codebook_16_hgtw, wedge_signflip_lookup[BLOCK_32X64], 0,
+    wedge_masks[BLOCK_32X64] },
+  { 0, wedge_codebook_16_hltw, wedge_signflip_lookup[BLOCK_64X32], 0,
+    wedge_masks[BLOCK_64X32] },
+  { 0, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_64X64], 0,
+    wedge_masks[BLOCK_64X64] },
+#else
+  { 0, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_8X8], 0,
+    wedge_masks[BLOCK_8X8] },
+  { 0, wedge_codebook_16_hgtw, wedge_signflip_lookup[BLOCK_8X16], 0,
+    wedge_masks[BLOCK_8X16] },
+  { 0, wedge_codebook_16_hltw, wedge_signflip_lookup[BLOCK_16X8], 0,
+    wedge_masks[BLOCK_16X8] },
+  { 0, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_16X16], 0,
+    wedge_masks[BLOCK_16X16] },
+  { 0, wedge_codebook_16_hgtw, wedge_signflip_lookup[BLOCK_16X32], 0,
+    wedge_masks[BLOCK_16X32] },
+  { 0, wedge_codebook_16_hltw, wedge_signflip_lookup[BLOCK_32X16], 0,
+    wedge_masks[BLOCK_32X16] },
+  { 0, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_32X32], 0,
+    wedge_masks[BLOCK_32X32] },
+  { 0, wedge_codebook_16_hgtw, wedge_signflip_lookup[BLOCK_32X64], 0,
+    wedge_masks[BLOCK_32X64] },
+  { 0, wedge_codebook_16_hltw, wedge_signflip_lookup[BLOCK_64X32], 0,
+    wedge_masks[BLOCK_64X32] },
+  { 0, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_64X64], 0,
+    wedge_masks[BLOCK_64X64] },
+#endif  // CONFIG_WEDGE
+#if CONFIG_EXT_PARTITION
+  { 0, NULL, NULL, 0, NULL },
+  { 0, NULL, NULL, 0, NULL },
+  { 0, NULL, NULL, 0, NULL },
+#endif  // CONFIG_EXT_PARTITION
+};
+
+static const uint8_t *get_wedge_mask_inplace(int wedge_index, int neg,
+                                             BLOCK_SIZE sb_type) {
+  const uint8_t *master;
+  const int bh = block_size_high[sb_type];
+  const int bw = block_size_wide[sb_type];
+  const wedge_code_type *a =
+      wedge_params_lookup[sb_type].codebook + wedge_index;
+  const int smoother = wedge_params_lookup[sb_type].smoother;
+  int woff, hoff;
+  const uint8_t wsignflip = wedge_params_lookup[sb_type].signflip[wedge_index];
+
+  assert(wedge_index >= 0 &&
+         wedge_index < (1 << get_wedge_bits_lookup(sb_type)));
+  woff = (a->x_offset * bw) >> 3;
+  hoff = (a->y_offset * bh) >> 3;
+  master = wedge_mask_obl[smoother][neg ^ wsignflip][a->direction] +
+           MASK_MASTER_STRIDE * (MASK_MASTER_SIZE / 2 - hoff) +
+           MASK_MASTER_SIZE / 2 - woff;
+  return master;
+}
+
+const uint8_t *av1_get_soft_mask(int wedge_index, int wedge_sign,
+                                 BLOCK_SIZE sb_type, int offset_x,
+                                 int offset_y) {
+  const uint8_t *mask =
+      get_wedge_mask_inplace(wedge_index, wedge_sign, sb_type);
+  if (mask) mask -= (offset_x + offset_y * MASK_MASTER_STRIDE);
+  return mask;
+}
+
+#if CONFIG_COMPOUND_SEGMENT
+static uint8_t *invert_mask(uint8_t *mask_inv_buffer, const uint8_t *const mask,
+                            int h, int w, int stride) {
+  int i, j;
+
+  for (i = 0; i < h; ++i)
+    for (j = 0; j < w; ++j) {
+      mask_inv_buffer[i * stride + j] =
+          AOM_BLEND_A64_MAX_ALPHA - mask[i * stride + j];
+    }
+  return mask_inv_buffer;
+}
+#endif  // CONFIG_COMPOUND_SEGMENT
+
+const uint8_t *av1_get_compound_type_mask_inverse(
+    const INTERINTER_COMPOUND_DATA *const comp_data,
+#if CONFIG_COMPOUND_SEGMENT
+    uint8_t *mask_buffer, int h, int w, int stride,
+#endif
+    BLOCK_SIZE sb_type) {
+  assert(is_masked_compound_type(comp_data->interinter_compound_type));
+  (void)sb_type;
+  switch (comp_data->interinter_compound_type) {
+#if CONFIG_WEDGE
+    case COMPOUND_WEDGE:
+      return av1_get_contiguous_soft_mask(comp_data->wedge_index,
+                                          !comp_data->wedge_sign, sb_type);
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+    case COMPOUND_SEG:
+      return invert_mask(mask_buffer, comp_data->seg_mask, h, w, stride);
+#endif  // CONFIG_COMPOUND_SEGMENT
+    default: assert(0); return NULL;
+  }
+}
+
+const uint8_t *av1_get_compound_type_mask(
+    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE sb_type) {
+  assert(is_masked_compound_type(comp_data->interinter_compound_type));
+  (void)sb_type;
+  switch (comp_data->interinter_compound_type) {
+#if CONFIG_WEDGE
+    case COMPOUND_WEDGE:
+      return av1_get_contiguous_soft_mask(comp_data->wedge_index,
+                                          comp_data->wedge_sign, sb_type);
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+    case COMPOUND_SEG: return comp_data->seg_mask;
+#endif  // CONFIG_COMPOUND_SEGMENT
+    default: assert(0); return NULL;
+  }
+}
+
+#if CONFIG_COMPOUND_SEGMENT
+#if COMPOUND_SEGMENT_TYPE == 0
+static void uniform_mask(uint8_t *mask, int which_inverse, BLOCK_SIZE sb_type,
+                         int h, int w, int mask_val) {
+  int i, j;
+  int block_stride = block_size_wide[sb_type];
+  for (i = 0; i < h; ++i)
+    for (j = 0; j < w; ++j) {
+      mask[i * block_stride + j] =
+          which_inverse ? AOM_BLEND_A64_MAX_ALPHA - mask_val : mask_val;
+    }
+}
+
+void build_compound_seg_mask(uint8_t *mask, SEG_MASK_TYPE mask_type,
+                             const uint8_t *src0, int src0_stride,
+                             const uint8_t *src1, int src1_stride,
+                             BLOCK_SIZE sb_type, int h, int w) {
+  (void)src0;
+  (void)src1;
+  (void)src0_stride;
+  (void)src1_stride;
+  switch (mask_type) {
+    case UNIFORM_45: uniform_mask(mask, 0, sb_type, h, w, 45); break;
+    case UNIFORM_45_INV: uniform_mask(mask, 1, sb_type, h, w, 45); break;
+    default: assert(0);
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void build_compound_seg_mask_highbd(uint8_t *mask, SEG_MASK_TYPE mask_type,
+                                    const uint8_t *src0, int src0_stride,
+                                    const uint8_t *src1, int src1_stride,
+                                    BLOCK_SIZE sb_type, int h, int w, int bd) {
+  (void)src0;
+  (void)src1;
+  (void)src0_stride;
+  (void)src1_stride;
+  (void)bd;
+  switch (mask_type) {
+    case UNIFORM_45: uniform_mask(mask, 0, sb_type, h, w, 45); break;
+    case UNIFORM_45_INV: uniform_mask(mask, 1, sb_type, h, w, 45); break;
+    default: assert(0);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#elif COMPOUND_SEGMENT_TYPE == 1
+#define DIFF_FACTOR 16
+static void diffwtd_mask(uint8_t *mask, int which_inverse, int mask_base,
+                         const uint8_t *src0, int src0_stride,
+                         const uint8_t *src1, int src1_stride,
+                         BLOCK_SIZE sb_type, int h, int w) {
+  int i, j, m, diff;
+  int block_stride = block_size_wide[sb_type];
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      diff =
+          abs((int)src0[i * src0_stride + j] - (int)src1[i * src1_stride + j]);
+      m = clamp(mask_base + (diff / DIFF_FACTOR), 0, AOM_BLEND_A64_MAX_ALPHA);
+      mask[i * block_stride + j] =
+          which_inverse ? AOM_BLEND_A64_MAX_ALPHA - m : m;
+    }
+  }
+}
+
+void build_compound_seg_mask(uint8_t *mask, SEG_MASK_TYPE mask_type,
+                             const uint8_t *src0, int src0_stride,
+                             const uint8_t *src1, int src1_stride,
+                             BLOCK_SIZE sb_type, int h, int w) {
+  switch (mask_type) {
+    case DIFFWTD_42:
+      diffwtd_mask(mask, 0, 42, src0, src0_stride, src1, src1_stride, sb_type,
+                   h, w);
+      break;
+    case DIFFWTD_42_INV:
+      diffwtd_mask(mask, 1, 42, src0, src0_stride, src1, src1_stride, sb_type,
+                   h, w);
+      break;
+    default: assert(0);
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void diffwtd_mask_highbd(uint8_t *mask, int which_inverse, int mask_base,
+                                const uint16_t *src0, int src0_stride,
+                                const uint16_t *src1, int src1_stride,
+                                BLOCK_SIZE sb_type, int h, int w, int bd) {
+  int i, j, m, diff;
+  int block_stride = block_size_wide[sb_type];
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; ++j) {
+      diff = abs((int)src0[i * src0_stride + j] -
+                 (int)src1[i * src1_stride + j]) >>
+             (bd - 8);
+      m = clamp(mask_base + (diff / DIFF_FACTOR), 0, AOM_BLEND_A64_MAX_ALPHA);
+      mask[i * block_stride + j] =
+          which_inverse ? AOM_BLEND_A64_MAX_ALPHA - m : m;
+    }
+  }
+}
+
+void build_compound_seg_mask_highbd(uint8_t *mask, SEG_MASK_TYPE mask_type,
+                                    const uint8_t *src0, int src0_stride,
+                                    const uint8_t *src1, int src1_stride,
+                                    BLOCK_SIZE sb_type, int h, int w, int bd) {
+  switch (mask_type) {
+    case DIFFWTD_42:
+      diffwtd_mask_highbd(mask, 0, 42, CONVERT_TO_SHORTPTR(src0), src0_stride,
+                          CONVERT_TO_SHORTPTR(src1), src1_stride, sb_type, h, w,
+                          bd);
+      break;
+    case DIFFWTD_42_INV:
+      diffwtd_mask_highbd(mask, 1, 42, CONVERT_TO_SHORTPTR(src0), src0_stride,
+                          CONVERT_TO_SHORTPTR(src1), src1_stride, sb_type, h, w,
+                          bd);
+      break;
+    default: assert(0);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // COMPOUND_SEGMENT_TYPE
+#endif  // CONFIG_COMPOUND_SEGMENT
+
+#if MASK_MASTER_SIZE == 64
+static const uint8_t wedge_master_oblique_odd[NSMOOTHERS][MASK_MASTER_SIZE] = {
+  {
+      0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
+      0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  1,  2,  6,  18,
+      37, 53, 60, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+      64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+  }
+};
+static const uint8_t wedge_master_oblique_even[NSMOOTHERS][MASK_MASTER_SIZE] = {
+  {
+      0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
+      0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  1,  4,  11, 27,
+      46, 58, 62, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+      64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+  }
+};
+static const uint8_t wedge_master_vertical[NSMOOTHERS][MASK_MASTER_SIZE] = { {
+    0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
+    0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  2,  7,  21,
+    43, 57, 62, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+    64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+} };
+
+static void shift_copy(const uint8_t *src, uint8_t *dst, int shift, int width) {
+  if (shift >= 0) {
+    memcpy(dst + shift, src, width - shift);
+    memset(dst, src[0], shift);
+  } else {
+    shift = -shift;
+    memcpy(dst, src + shift, width - shift);
+    memset(dst + width - shift, src[width - 1], shift);
+  }
+}
+#else
+static const double smoother_param[NSMOOTHERS] = { 2.83 };
+#endif  // MASK_MASTER_SIZE == 64
+
+static void init_wedge_master_masks() {
+  int i, j, s;
+  const int w = MASK_MASTER_SIZE;
+  const int h = MASK_MASTER_SIZE;
+  const int stride = MASK_MASTER_STRIDE;
+  for (s = 0; s < NSMOOTHERS; s++) {
+#if MASK_MASTER_SIZE == 64
+    // Generate prototype by shifting the masters
+    int shift = h / 4;
+    for (i = 0; i < h; i += 2) {
+      shift_copy(wedge_master_oblique_even[s],
+                 &wedge_mask_obl[s][1][WEDGE_OBLIQUE63][i * stride], shift,
+                 MASK_MASTER_SIZE);
+      shift--;
+      shift_copy(wedge_master_oblique_odd[s],
+                 &wedge_mask_obl[s][1][WEDGE_OBLIQUE63][(i + 1) * stride],
+                 shift, MASK_MASTER_SIZE);
+      memcpy(&wedge_mask_obl[s][1][WEDGE_VERTICAL][i * stride],
+             wedge_master_vertical[s],
+             MASK_MASTER_SIZE * sizeof(wedge_master_vertical[s][0]));
+      memcpy(&wedge_mask_obl[s][1][WEDGE_VERTICAL][(i + 1) * stride],
+             wedge_master_vertical[s],
+             MASK_MASTER_SIZE * sizeof(wedge_master_vertical[s][0]));
+    }
+#else
+    const int a[2] = { 2, 1 };
+    const double asqrt = sqrt(a[0] * a[0] + a[1] * a[1]);
+    for (i = 0; i < h; i++) {
+      for (j = 0; j < w; ++j) {
+        int x = (2 * j + 1 - w);
+        int y = (2 * i + 1 - h);
+        double d = (a[0] * x + a[1] * y) / asqrt;
+        const int msk = (int)rint((1.0 + tanh(d / smoother_param[s])) * 32);
+        wedge_mask_obl[s][1][WEDGE_OBLIQUE63][i * stride + j] = msk;
+        const int mskx = (int)rint((1.0 + tanh(x / smoother_param[s])) * 32);
+        wedge_mask_obl[s][1][WEDGE_VERTICAL][i * stride + j] = mskx;
+      }
+    }
+#endif  // MASK_MASTER_SIZE == 64
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; ++j) {
+        const int msk = wedge_mask_obl[s][1][WEDGE_OBLIQUE63][i * stride + j];
+        wedge_mask_obl[s][1][WEDGE_OBLIQUE27][j * stride + i] = msk;
+        wedge_mask_obl[s][1][WEDGE_OBLIQUE117][i * stride + w - 1 - j] =
+            wedge_mask_obl[s][1][WEDGE_OBLIQUE153][(w - 1 - j) * stride + i] =
+                (1 << WEDGE_WEIGHT_BITS) - msk;
+        wedge_mask_obl[s][0][WEDGE_OBLIQUE63][i * stride + j] =
+            wedge_mask_obl[s][0][WEDGE_OBLIQUE27][j * stride + i] =
+                (1 << WEDGE_WEIGHT_BITS) - msk;
+        wedge_mask_obl[s][0][WEDGE_OBLIQUE117][i * stride + w - 1 - j] =
+            wedge_mask_obl[s][0][WEDGE_OBLIQUE153][(w - 1 - j) * stride + i] =
+                msk;
+        const int mskx = wedge_mask_obl[s][1][WEDGE_VERTICAL][i * stride + j];
+        wedge_mask_obl[s][1][WEDGE_HORIZONTAL][j * stride + i] = mskx;
+        wedge_mask_obl[s][0][WEDGE_VERTICAL][i * stride + j] =
+            wedge_mask_obl[s][0][WEDGE_HORIZONTAL][j * stride + i] =
+                (1 << WEDGE_WEIGHT_BITS) - mskx;
+      }
+    }
+  }
+}
+
+// If the signs for the wedges for various blocksizes are
+// inconsistent flip the sign flag. Do it only once for every
+// wedge codebook.
+static void init_wedge_signs() {
+  BLOCK_SIZE sb_type;
+  memset(wedge_signflip_lookup, 0, sizeof(wedge_signflip_lookup));
+  for (sb_type = BLOCK_4X4; sb_type < BLOCK_SIZES; ++sb_type) {
+    const int bw = block_size_wide[sb_type];
+    const int bh = block_size_high[sb_type];
+    const wedge_params_type wedge_params = wedge_params_lookup[sb_type];
+    const int wbits = wedge_params.bits;
+    const int wtypes = 1 << wbits;
+    int i, w;
+    if (wbits == 0) continue;
+    for (w = 0; w < wtypes; ++w) {
+      const uint8_t *mask = get_wedge_mask_inplace(w, 0, sb_type);
+      int sum = 0;
+      for (i = 0; i < bw; ++i) sum += mask[i];
+      for (i = 0; i < bh; ++i) sum += mask[i * MASK_MASTER_STRIDE];
+      sum = (sum + (bw + bh) / 2) / (bw + bh);
+      wedge_params.signflip[w] = (sum < 32);
+    }
+  }
+}
+
+static void init_wedge_masks() {
+  uint8_t *dst = wedge_mask_buf;
+  BLOCK_SIZE bsize;
+  memset(wedge_masks, 0, sizeof(wedge_masks));
+  for (bsize = BLOCK_4X4; bsize < BLOCK_SIZES; ++bsize) {
+    const uint8_t *mask;
+    const int bw = block_size_wide[bsize];
+    const int bh = block_size_high[bsize];
+    const wedge_params_type *wedge_params = &wedge_params_lookup[bsize];
+    const int wbits = wedge_params->bits;
+    const int wtypes = 1 << wbits;
+    int w;
+    if (wbits == 0) continue;
+    for (w = 0; w < wtypes; ++w) {
+      mask = get_wedge_mask_inplace(w, 0, bsize);
+      aom_convolve_copy(mask, MASK_MASTER_STRIDE, dst, bw, NULL, 0, NULL, 0, bw,
+                        bh);
+      wedge_params->masks[0][w] = dst;
+      dst += bw * bh;
+
+      mask = get_wedge_mask_inplace(w, 1, bsize);
+      aom_convolve_copy(mask, MASK_MASTER_STRIDE, dst, bw, NULL, 0, NULL, 0, bw,
+                        bh);
+      wedge_params->masks[1][w] = dst;
+      dst += bw * bh;
+    }
+    assert(sizeof(wedge_mask_buf) >= (size_t)(dst - wedge_mask_buf));
+  }
+}
+
+// Equation of line: f(x, y) = a[0]*(x - a[2]*w/8) + a[1]*(y - a[3]*h/8) = 0
+void av1_init_wedge_masks() {
+  init_wedge_master_masks();
+  init_wedge_signs();
+  init_wedge_masks();
+}
+
+#if CONFIG_SUPERTX
+static void build_masked_compound_wedge_extend(
+    uint8_t *dst, int dst_stride, const uint8_t *src0, int src0_stride,
+    const uint8_t *src1, int src1_stride,
+    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE sb_type,
+    int wedge_offset_x, int wedge_offset_y, int h, int w) {
+  const int subh = (2 << b_height_log2_lookup[sb_type]) == h;
+  const int subw = (2 << b_width_log2_lookup[sb_type]) == w;
+  const uint8_t *mask;
+  size_t mask_stride;
+  switch (comp_data->interinter_compound_type) {
+    case COMPOUND_WEDGE:
+      mask = av1_get_soft_mask(comp_data->wedge_index, comp_data->wedge_sign,
+                               sb_type, wedge_offset_x, wedge_offset_y);
+      mask_stride = MASK_MASTER_STRIDE;
+      break;
+#if CONFIG_COMPOUND_SEGMENT
+    case COMPOUND_SEG:
+      mask = comp_data->seg_mask;
+      mask_stride = block_size_wide[sb_type];
+      break;
+#endif
+    default: assert(0); return;
+  }
+  aom_blend_a64_mask(dst, dst_stride, src0, src0_stride, src1, src1_stride,
+                     mask, (int)mask_stride, h, w, subh, subw);
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void build_masked_compound_wedge_extend_highbd(
+    uint8_t *dst_8, int dst_stride, const uint8_t *src0_8, int src0_stride,
+    const uint8_t *src1_8, int src1_stride,
+    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE sb_type,
+    int wedge_offset_x, int wedge_offset_y, int h, int w, int bd) {
+  const int subh = (2 << b_height_log2_lookup[sb_type]) == h;
+  const int subw = (2 << b_width_log2_lookup[sb_type]) == w;
+  const uint8_t *mask;
+  size_t mask_stride;
+  switch (comp_data->interinter_compound_type) {
+    case COMPOUND_WEDGE:
+      mask = av1_get_soft_mask(comp_data->wedge_index, comp_data->wedge_sign,
+                               sb_type, wedge_offset_x, wedge_offset_y);
+      mask_stride = MASK_MASTER_STRIDE;
+      break;
+#if CONFIG_COMPOUND_SEGMENT
+    case COMPOUND_SEG:
+      mask = comp_data->seg_mask;
+      mask_stride = block_size_wide[sb_type];
+      break;
+#endif
+    default: assert(0); return;
+  }
+  aom_highbd_blend_a64_mask(dst_8, dst_stride, src0_8, src0_stride, src1_8,
+                            src1_stride, mask, (int)mask_stride, h, w, subh,
+                            subw, bd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+#else
+static void build_masked_compound(
+    uint8_t *dst, int dst_stride, const uint8_t *src0, int src0_stride,
+    const uint8_t *src1, int src1_stride,
+    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE sb_type, int h,
+    int w) {
+  // Derive subsampling from h and w passed in. May be refactored to
+  // pass in subsampling factors directly.
+  const int subh = (2 << b_height_log2_lookup[sb_type]) == h;
+  const int subw = (2 << b_width_log2_lookup[sb_type]) == w;
+  const uint8_t *mask = av1_get_compound_type_mask(comp_data, sb_type);
+  aom_blend_a64_mask(dst, dst_stride, src0, src0_stride, src1, src1_stride,
+                     mask, block_size_wide[sb_type], h, w, subh, subw);
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void build_masked_compound_highbd(
+    uint8_t *dst_8, int dst_stride, const uint8_t *src0_8, int src0_stride,
+    const uint8_t *src1_8, int src1_stride,
+    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE sb_type, int h,
+    int w, int bd) {
+  // Derive subsampling from h and w passed in. May be refactored to
+  // pass in subsampling factors directly.
+  const int subh = (2 << b_height_log2_lookup[sb_type]) == h;
+  const int subw = (2 << b_width_log2_lookup[sb_type]) == w;
+  const uint8_t *mask = av1_get_compound_type_mask(comp_data, sb_type);
+  // const uint8_t *mask =
+  //     av1_get_contiguous_soft_mask(wedge_index, wedge_sign, sb_type);
+  aom_highbd_blend_a64_mask(dst_8, dst_stride, src0_8, src0_stride, src1_8,
+                            src1_stride, mask, block_size_wide[sb_type], h, w,
+                            subh, subw, bd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_SUPERTX
+
+void av1_make_masked_inter_predictor(const uint8_t *pre, int pre_stride,
+                                     uint8_t *dst, int dst_stride,
+                                     const int subpel_x, const int subpel_y,
+                                     const struct scale_factors *sf, int w,
+                                     int h,
+#if CONFIG_DUAL_FILTER
+                                     const InterpFilter *interp_filter,
+#else
+                                     const InterpFilter interp_filter,
+#endif
+                                     int xs, int ys,
+#if CONFIG_SUPERTX
+                                     int wedge_offset_x, int wedge_offset_y,
+#endif  // CONFIG_SUPERTX
+                                     int plane,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     const WarpTypesAllowed *warp_types,
+                                     int p_col, int p_row, int ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     MACROBLOCKD *xd) {
+  MODE_INFO *mi = xd->mi[0];
+  const INTERINTER_COMPOUND_DATA comp_data = {
+#if CONFIG_WEDGE
+    mi->mbmi.wedge_index,
+    mi->mbmi.wedge_sign,
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+    mi->mbmi.mask_type,
+    xd->seg_mask,
+#endif  // CONFIG_COMPOUND_SEGMENT
+    mi->mbmi.interinter_compound_type
+  };
+// The prediction filter types used here should be those for
+// the second reference block.
+#if CONFIG_DUAL_FILTER
+  InterpFilter tmp_ipf[4] = {
+    interp_filter[2], interp_filter[3], interp_filter[2], interp_filter[3],
+  };
+#else
+  InterpFilter tmp_ipf = interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+  ConvolveParams conv_params = get_conv_params(0, plane);
+
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint8_t, tmp_dst_[2 * MAX_SB_SQUARE]);
+  uint8_t *tmp_dst = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+                         ? CONVERT_TO_BYTEPTR(tmp_dst_)
+                         : tmp_dst_;
+  av1_make_inter_predictor(pre, pre_stride, tmp_dst, MAX_SB_SIZE, subpel_x,
+                           subpel_y, sf, w, h, &conv_params, tmp_ipf,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                           warp_types, p_col, p_row, plane, ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+                           0, 0,
+#endif
+                           xs, ys, xd);
+#if CONFIG_COMPOUND_SEGMENT
+  if (!plane && comp_data.interinter_compound_type == COMPOUND_SEG) {
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+      build_compound_seg_mask_highbd(comp_data.seg_mask, comp_data.mask_type,
+                                     dst, dst_stride, tmp_dst, MAX_SB_SIZE,
+                                     mi->mbmi.sb_type, h, w, xd->bd);
+    else
+      build_compound_seg_mask(comp_data.seg_mask, comp_data.mask_type, dst,
+                              dst_stride, tmp_dst, MAX_SB_SIZE,
+                              mi->mbmi.sb_type, h, w);
+  }
+#endif  // CONFIG_COMPOUND_SEGMENT
+
+#if CONFIG_SUPERTX
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+    build_masked_compound_wedge_extend_highbd(
+        dst, dst_stride, dst, dst_stride, tmp_dst, MAX_SB_SIZE, &comp_data,
+        mi->mbmi.sb_type, wedge_offset_x, wedge_offset_y, h, w, xd->bd);
+  else
+    build_masked_compound_wedge_extend(
+        dst, dst_stride, dst, dst_stride, tmp_dst, MAX_SB_SIZE, &comp_data,
+        mi->mbmi.sb_type, wedge_offset_x, wedge_offset_y, h, w);
+#else
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+    build_masked_compound_highbd(dst, dst_stride, dst, dst_stride, tmp_dst,
+                                 MAX_SB_SIZE, &comp_data, mi->mbmi.sb_type, h,
+                                 w, xd->bd);
+  else
+    build_masked_compound(dst, dst_stride, dst, dst_stride, tmp_dst,
+                          MAX_SB_SIZE, &comp_data, mi->mbmi.sb_type, h, w);
+#endif  // CONFIG_SUPERTX
+
+#else  // CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint8_t, tmp_dst[MAX_SB_SQUARE]);
+  av1_make_inter_predictor(pre, pre_stride, tmp_dst, MAX_SB_SIZE, subpel_x,
+                           subpel_y, sf, w, h, &conv_params, tmp_ipf,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                           warp_types, p_col, p_row, plane, ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+                           0, 0,
+#endif
+                           xs, ys, xd);
+#if CONFIG_COMPOUND_SEGMENT
+  if (!plane && comp_data.interinter_compound_type == COMPOUND_SEG)
+    build_compound_seg_mask(comp_data.seg_mask, comp_data.mask_type, dst,
+                            dst_stride, tmp_dst, MAX_SB_SIZE, mi->mbmi.sb_type,
+                            h, w);
+#endif  // CONFIG_COMPOUND_SEGMENT
+#if CONFIG_SUPERTX
+  build_masked_compound_wedge_extend(dst, dst_stride, dst, dst_stride, tmp_dst,
+                                     MAX_SB_SIZE, &comp_data, mi->mbmi.sb_type,
+                                     wedge_offset_x, wedge_offset_y, h, w);
+#else
+  build_masked_compound(dst, dst_stride, dst, dst_stride, tmp_dst, MAX_SB_SIZE,
+                        &comp_data, mi->mbmi.sb_type, h, w);
+#endif  // CONFIG_SUPERTX
+#endif  // CONFIG_HIGHBITDEPTH
+#if CONFIG_COMPOUND_SEGMENT
+  (void)plane;
+#endif  // CONFIG_COMPOUND_SEGMENT
+}
+#endif  // CONFIG_EXT_INTER
+
+// TODO(sarahparker) av1_highbd_build_inter_predictor and
+// av1_build_inter_predictor should be combined with
+// av1_make_inter_predictor
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_build_inter_predictor(
+    const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride,
+    const MV *src_mv, const struct scale_factors *sf, int w, int h, int ref,
+#if CONFIG_DUAL_FILTER
+    const InterpFilter *interp_filter,
+#else
+    const InterpFilter interp_filter,
+#endif
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+    const WarpTypesAllowed *warp_types, int p_col, int p_row,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+    int plane, enum mv_precision precision, int x, int y,
+    const MACROBLOCKD *xd) {
+  const int is_q4 = precision == MV_PRECISION_Q4;
+  const MV mv_q4 = { is_q4 ? src_mv->row : src_mv->row * 2,
+                     is_q4 ? src_mv->col : src_mv->col * 2 };
+  MV32 mv = av1_scale_mv(&mv_q4, x, y, sf);
+  const int subpel_x = mv.col & SUBPEL_MASK;
+  const int subpel_y = mv.row & SUBPEL_MASK;
+  ConvolveParams conv_params = get_conv_params(ref, plane);
+
+  src += (mv.row >> SUBPEL_BITS) * src_stride + (mv.col >> SUBPEL_BITS);
+
+  av1_make_inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
+                           sf, w, h, &conv_params, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                           warp_types, p_col, p_row, plane, ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+                           0, 0,
+#endif
+                           sf->x_step_q4, sf->y_step_q4, xd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_build_inter_predictor(const uint8_t *src, int src_stride, uint8_t *dst,
+                               int dst_stride, const MV *src_mv,
+                               const struct scale_factors *sf, int w, int h,
+                               ConvolveParams *conv_params,
+#if CONFIG_DUAL_FILTER
+                               const InterpFilter *interp_filter,
+#else
+                               const InterpFilter interp_filter,
+#endif
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                               const WarpTypesAllowed *warp_types, int p_col,
+                               int p_row, int plane, int ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                               enum mv_precision precision, int x, int y,
+                               const MACROBLOCKD *xd) {
+  const int is_q4 = precision == MV_PRECISION_Q4;
+  const MV mv_q4 = { is_q4 ? src_mv->row : src_mv->row * 2,
+                     is_q4 ? src_mv->col : src_mv->col * 2 };
+  MV32 mv = av1_scale_mv(&mv_q4, x, y, sf);
+  const int subpel_x = mv.col & SUBPEL_MASK;
+  const int subpel_y = mv.row & SUBPEL_MASK;
+
+  src += (mv.row >> SUBPEL_BITS) * src_stride + (mv.col >> SUBPEL_BITS);
+
+  av1_make_inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
+                           sf, w, h, conv_params, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                           warp_types, p_col, p_row, plane, ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+                           0, 0,
+#endif
+                           sf->x_step_q4, sf->y_step_q4, xd);
+}
+
+typedef struct SubpelParams {
+  int xs;
+  int ys;
+  int subpel_x;
+  int subpel_y;
+} SubpelParams;
+
+void build_inter_predictors(MACROBLOCKD *xd, int plane,
+#if CONFIG_MOTION_VAR
+                            int mi_col_offset, int mi_row_offset,
+#endif  // CONFIG_MOTION_VAR
+                            int block, int bw, int bh, int x, int y, int w,
+                            int h,
+#if CONFIG_SUPERTX && CONFIG_EXT_INTER
+                            int wedge_offset_x, int wedge_offset_y,
+#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
+                            int mi_x, int mi_y) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+#if CONFIG_MOTION_VAR
+  const MODE_INFO *mi = xd->mi[mi_col_offset + xd->mi_stride * mi_row_offset];
+#if !CONFIG_CB4X4 || CONFIG_SUB8X8_MC
+  const int build_for_obmc = !(mi_col_offset == 0 && mi_row_offset == 0);
+#endif  // !CONFIG_CB4X4 || CONFIG_SUB8X8_MC
+#else
+  const MODE_INFO *mi = xd->mi[0];
+#endif  // CONFIG_MOTION_VAR
+  const int is_compound = has_second_ref(&mi->mbmi);
+  int ref;
+#if CONFIG_INTRABC
+  const int is_intrabc = is_intrabc_block(&mi->mbmi);
+  struct scale_factors sf_identity;
+#if CONFIG_HIGHBITDEPTH
+  av1_setup_scale_factors_for_frame(
+      &sf_identity, 64, 64, 64, 64,
+      xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH);
+#else
+  av1_setup_scale_factors_for_frame(&sf_identity, 64, 64, 64, 64);
+#endif  // CONFIG_HIGHBITDEPTH
+  assert(IMPLIES(is_intrabc, !is_compound));
+#endif  // CONFIG_INTRABC
+#if CONFIG_GLOBAL_MOTION
+  int is_global[2];
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    WarpedMotionParams *const wm = &xd->global_motion[mi->mbmi.ref_frame[ref]];
+    is_global[ref] = is_global_mv_block(mi, block, wm->wmtype);
+  }
+#endif  // CONFIG_GLOBAL_MOTION
+
+#if CONFIG_CB4X4
+  (void)block;
+#endif
+
+#if CONFIG_SUB8X8_MC
+#if CONFIG_MOTION_VAR
+  if (mi->mbmi.sb_type < BLOCK_8X8 && plane > 0 && !build_for_obmc) {
+#else
+  if (mi->mbmi.sb_type < BLOCK_8X8 && plane > 0) {
+#endif  // CONFIG_MOTION_VAR
+    // block size in log2
+    const int b4_wl = b_width_log2_lookup[mi->mbmi.sb_type];
+    const int b4_hl = b_height_log2_lookup[mi->mbmi.sb_type];
+    const int b8_sl = b_width_log2_lookup[BLOCK_8X8];
+
+    // block size
+    const int b4_w = 1 << b4_wl;
+    const int b4_h = 1 << b4_hl;
+    const int b8_s = 1 << b8_sl;
+    int idx, idy;
+
+    const int x_base = x;
+    const int y_base = y;
+
+    // processing unit size
+    const int x_step = w >> (b8_sl - b4_wl);
+    const int y_step = h >> (b8_sl - b4_hl);
+
+    for (idy = 0; idy < b8_s; idy += b4_h) {
+      for (idx = 0; idx < b8_s; idx += b4_w) {
+        const int chr_idx = (idy * 2) + idx;
+        for (ref = 0; ref < 1 + is_compound; ++ref) {
+          struct buf_2d *const dst_buf = &pd->dst;
+#if CONFIG_INTRABC
+          const struct scale_factors *const sf =
+              is_intrabc ? &sf_identity : &xd->block_refs[ref]->sf;
+          struct buf_2d *const pre_buf = is_intrabc ? dst_buf : &pd->pre[ref];
+#else
+          const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
+          struct buf_2d *const pre_buf = &pd->pre[ref];
+#endif  // CONFIG_INTRABC
+          uint8_t *dst = dst_buf->buf;
+          const MV mv = mi->bmi[chr_idx].as_mv[ref].as_mv;
+          const MV mv_q4 = clamp_mv_to_umv_border_sb(
+              xd, &mv, bw, bh, pd->subsampling_x, pd->subsampling_y);
+          uint8_t *pre;
+          MV32 scaled_mv;
+          int xs, ys, subpel_x, subpel_y;
+          const int is_scaled = av1_is_scaled(sf);
+          ConvolveParams conv_params = get_conv_params(ref, plane);
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+          WarpTypesAllowed warp_types;
+#if CONFIG_GLOBAL_MOTION
+          warp_types.global_warp_allowed = is_global[ref];
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_WARPED_MOTION
+          warp_types.local_warp_allowed = mi->mbmi.motion_mode == WARPED_CAUSAL;
+#endif  // CONFIG_WARPED_MOTION
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+
+          x = x_base + idx * x_step;
+          y = y_base + idy * y_step;
+
+          dst += dst_buf->stride * y + x;
+
+          if (is_scaled) {
+            pre =
+                pre_buf->buf + scaled_buffer_offset(x, y, pre_buf->stride, sf);
+            scaled_mv = av1_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
+            xs = sf->x_step_q4;
+            ys = sf->y_step_q4;
+          } else {
+            pre = pre_buf->buf + y * pre_buf->stride + x;
+            scaled_mv.row = mv_q4.row;
+            scaled_mv.col = mv_q4.col;
+            xs = ys = 16;
+          }
+
+          subpel_x = scaled_mv.col & SUBPEL_MASK;
+          subpel_y = scaled_mv.row & SUBPEL_MASK;
+          pre += (scaled_mv.row >> SUBPEL_BITS) * pre_buf->stride +
+                 (scaled_mv.col >> SUBPEL_BITS);
+
+#if CONFIG_EXT_INTER
+          if (ref && is_masked_compound_type(mi->mbmi.interinter_compound_type))
+            av1_make_masked_inter_predictor(
+                pre, pre_buf->stride, dst, dst_buf->stride, subpel_x, subpel_y,
+                sf, w, h, mi->mbmi.interp_filter, xs, ys,
+#if CONFIG_SUPERTX
+                wedge_offset_x, wedge_offset_y,
+#endif  // CONFIG_SUPERTX
+                plane,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                &warp_types, (mi_x >> pd->subsampling_x) + x,
+                (mi_y >> pd->subsampling_y) + y, ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                xd);
+          else
+#endif  // CONFIG_EXT_INTER
+            av1_make_inter_predictor(
+                pre, pre_buf->stride, dst, dst_buf->stride, subpel_x, subpel_y,
+                sf, x_step, y_step, &conv_params, mi->mbmi.interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                &warp_types, (mi_x >> pd->subsampling_x) + x,
+                (mi_y >> pd->subsampling_y) + y, plane, ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+                mi_col_offset, mi_row_offset,
+#endif
+                xs, ys, xd);
+        }
+      }
+    }
+    return;
+  }
+#endif
+
+  {
+    struct buf_2d *const dst_buf = &pd->dst;
+    uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
+    uint8_t *pre[2];
+    MV32 scaled_mv[2];
+    SubpelParams subpel_params[2];
+#if CONFIG_CONVOLVE_ROUND
+    DECLARE_ALIGNED(16, int32_t, tmp_dst[MAX_SB_SIZE * MAX_SB_SIZE]);
+    av1_zero(tmp_dst);
+#endif  // CONFIG_CONVOLVE_ROUND
+
+    for (ref = 0; ref < 1 + is_compound; ++ref) {
+#if CONFIG_INTRABC
+      const struct scale_factors *const sf =
+          is_intrabc ? &sf_identity : &xd->block_refs[ref]->sf;
+      struct buf_2d *const pre_buf = is_intrabc ? dst_buf : &pd->pre[ref];
+#else
+      const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
+      struct buf_2d *const pre_buf = &pd->pre[ref];
+#endif  // CONFIG_INTRABC
+#if CONFIG_CB4X4
+      const MV mv = mi->mbmi.mv[ref].as_mv;
+#else
+      const MV mv =
+#if CONFIG_MOTION_VAR
+          (mi->mbmi.sb_type < BLOCK_8X8 && !build_for_obmc)
+              ?
+#else
+          mi->mbmi.sb_type < BLOCK_8X8 ?
+#endif
+              average_split_mvs(pd, mi, ref, block)
+              : mi->mbmi.mv[ref].as_mv;
+#endif
+
+      // TODO(jkoleszar): This clamping is done in the incorrect place for the
+      // scaling case. It needs to be done on the scaled MV, not the pre-scaling
+      // MV. Note however that it performs the subsampling aware scaling so
+      // that the result is always q4.
+      // mv_precision precision is MV_PRECISION_Q4.
+      const MV mv_q4 = clamp_mv_to_umv_border_sb(
+          xd, &mv, bw, bh, pd->subsampling_x, pd->subsampling_y);
+
+      const int is_scaled = av1_is_scaled(sf);
+      if (is_scaled) {
+        pre[ref] =
+            pre_buf->buf + scaled_buffer_offset(x, y, pre_buf->stride, sf);
+        scaled_mv[ref] = av1_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
+        subpel_params[ref].xs = sf->x_step_q4;
+        subpel_params[ref].ys = sf->y_step_q4;
+      } else {
+        pre[ref] = pre_buf->buf + (y * pre_buf->stride + x);
+        scaled_mv[ref].row = mv_q4.row;
+        scaled_mv[ref].col = mv_q4.col;
+        subpel_params[ref].xs = 16;
+        subpel_params[ref].ys = 16;
+      }
+
+      subpel_params[ref].subpel_x = scaled_mv[ref].col & SUBPEL_MASK;
+      subpel_params[ref].subpel_y = scaled_mv[ref].row & SUBPEL_MASK;
+      pre[ref] += (scaled_mv[ref].row >> SUBPEL_BITS) * pre_buf->stride +
+                  (scaled_mv[ref].col >> SUBPEL_BITS);
+    }
+
+#if CONFIG_CONVOLVE_ROUND
+    ConvolveParams conv_params =
+        get_conv_params_no_round(ref, plane, tmp_dst, MAX_SB_SIZE);
+#else
+    ConvolveParams conv_params = get_conv_params(ref, plane);
+#endif  // CONFIG_CONVOLVE_ROUND
+    for (ref = 0; ref < 1 + is_compound; ++ref) {
+#if CONFIG_INTRABC
+      const struct scale_factors *const sf =
+          is_intrabc ? &sf_identity : &xd->block_refs[ref]->sf;
+      struct buf_2d *const pre_buf = is_intrabc ? dst_buf : &pd->pre[ref];
+#else
+      const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
+      struct buf_2d *const pre_buf = &pd->pre[ref];
+#endif  // CONFIG_INTRABC
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+      WarpTypesAllowed warp_types;
+#if CONFIG_GLOBAL_MOTION
+      warp_types.global_warp_allowed = is_global[ref];
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_WARPED_MOTION
+      warp_types.local_warp_allowed = mi->mbmi.motion_mode == WARPED_CAUSAL;
+#endif  // CONFIG_WARPED_MOTION
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+      conv_params.ref = ref;
+#if CONFIG_EXT_INTER
+      if (ref && is_masked_compound_type(mi->mbmi.interinter_compound_type))
+        av1_make_masked_inter_predictor(
+            pre[ref], pre_buf->stride, dst, dst_buf->stride,
+            subpel_params[ref].subpel_x, subpel_params[ref].subpel_y, sf, w, h,
+            mi->mbmi.interp_filter, subpel_params[ref].xs,
+            subpel_params[ref].ys,
+#if CONFIG_SUPERTX
+            wedge_offset_x, wedge_offset_y,
+#endif  // CONFIG_SUPERTX
+            plane,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+            &warp_types, (mi_x >> pd->subsampling_x) + x,
+            (mi_y >> pd->subsampling_y) + y, ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+            xd);
+      else
+#endif  // CONFIG_EXT_INTER
+        av1_make_inter_predictor(
+            pre[ref], pre_buf->stride, dst, dst_buf->stride,
+            subpel_params[ref].subpel_x, subpel_params[ref].subpel_y, sf, w, h,
+            &conv_params, mi->mbmi.interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+            &warp_types, (mi_x >> pd->subsampling_x) + x,
+            (mi_y >> pd->subsampling_y) + y, plane, ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+            mi_col_offset, mi_row_offset,
+#endif
+            subpel_params[ref].xs, subpel_params[ref].ys, xd);
+    }
+
+#if CONFIG_CONVOLVE_ROUND
+// TODO(angiebird): This part needs optimization
+#if CONFIG_HIGHBITDEPTH
+    if (!(xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH))
+#endif  // CONFIG_HIGHBITDEPTH
+      av1_convolve_rounding(tmp_dst, MAX_SB_SIZE, dst, dst_buf->stride, w, h,
+                            FILTER_BITS * 2 + is_compound -
+                                conv_params.round_0 - conv_params.round_1);
+#endif  // CONFIG_CONVOLVE_ROUND
+  }
+}
+
+void av1_build_inter_predictor_sub8x8(MACROBLOCKD *xd, int plane, int i, int ir,
+                                      int ic, int mi_row, int mi_col) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  MODE_INFO *const mi = xd->mi[0];
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
+  const int width = block_size_wide[plane_bsize];
+  const int height = block_size_high[plane_bsize];
+  uint8_t *const dst = &pd->dst.buf[(ir * pd->dst.stride + ic) << 2];
+  int ref;
+  const int is_compound = has_second_ref(&mi->mbmi);
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+  WarpTypesAllowed warp_types;
+  const int p_col = ((mi_col * MI_SIZE) >> pd->subsampling_x) + 4 * ic;
+  const int p_row = ((mi_row * MI_SIZE) >> pd->subsampling_y) + 4 * ir;
+#if CONFIG_GLOBAL_MOTION
+  int is_global[2];
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    WarpedMotionParams *const wm = &xd->global_motion[mi->mbmi.ref_frame[ref]];
+    is_global[ref] = is_global_mv_block(mi, i, wm->wmtype);
+  }
+#endif  // CONFIG_GLOBAL_MOTION
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    ConvolveParams conv_params = get_conv_params(ref, plane);
+    const uint8_t *pre =
+        &pd->pre[ref].buf[(ir * pd->pre[ref].stride + ic) << 2];
+#if CONFIG_GLOBAL_MOTION
+    warp_types.global_warp_allowed = is_global[ref];
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_WARPED_MOTION
+    warp_types.local_warp_allowed = mi->mbmi.motion_mode == WARPED_CAUSAL;
+#endif  // CONFIG_WARPED_MOTION
+
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+      av1_highbd_build_inter_predictor(
+          pre, pd->pre[ref].stride, dst, pd->dst.stride,
+          &mi->bmi[i].as_mv[ref].as_mv, &xd->block_refs[ref]->sf, width, height,
+          ref, mi->mbmi.interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+          &warp_types, p_col, p_row,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+          plane, MV_PRECISION_Q3, mi_col * MI_SIZE + 4 * ic,
+          mi_row * MI_SIZE + 4 * ir, xd);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      av1_build_inter_predictor(pre, pd->pre[ref].stride, dst, pd->dst.stride,
+                                &mi->bmi[i].as_mv[ref].as_mv,
+                                &xd->block_refs[ref]->sf, width, height,
+                                &conv_params, mi->mbmi.interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                &warp_types, p_col, p_row, plane, ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                MV_PRECISION_Q3, mi_col * MI_SIZE + 4 * ic,
+                                mi_row * MI_SIZE + 4 * ir, xd);
+  }
+}
+
+static void build_inter_predictors_for_planes(MACROBLOCKD *xd, BLOCK_SIZE bsize,
+                                              int mi_row, int mi_col,
+                                              int plane_from, int plane_to) {
+  int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+  for (plane = plane_from; plane <= plane_to; ++plane) {
+    const struct macroblockd_plane *pd = &xd->plane[plane];
+    const int bw = pd->width;
+    const int bh = pd->height;
+
+#if CONFIG_CB4X4
+    if (!is_chroma_reference(mi_row, mi_col, bsize, pd->subsampling_x,
+                             pd->subsampling_y))
+      continue;
+#endif
+
+    if (xd->mi[0]->mbmi.sb_type < BLOCK_8X8 && !unify_bsize) {
+      const PARTITION_TYPE bp = bsize - xd->mi[0]->mbmi.sb_type;
+      const int have_vsplit = bp != PARTITION_HORZ;
+      const int have_hsplit = bp != PARTITION_VERT;
+      const int num_4x4_w = 2 >> ((!have_vsplit) | pd->subsampling_x);
+      const int num_4x4_h = 2 >> ((!have_hsplit) | pd->subsampling_y);
+      const int pw = 8 >> (have_vsplit | pd->subsampling_x);
+      const int ph = 8 >> (have_hsplit | pd->subsampling_y);
+      int x, y;
+      assert(bp != PARTITION_NONE && bp < PARTITION_TYPES);
+      assert(bsize == BLOCK_8X8);
+      assert(pw * num_4x4_w == bw && ph * num_4x4_h == bh);
+      for (y = 0; y < num_4x4_h; ++y)
+        for (x = 0; x < num_4x4_w; ++x)
+          build_inter_predictors(xd, plane,
+#if CONFIG_MOTION_VAR
+                                 0, 0,
+#endif  // CONFIG_MOTION_VAR
+                                 y * 2 + x, bw, bh, 4 * x, 4 * y, pw, ph,
+#if CONFIG_SUPERTX && CONFIG_EXT_INTER
+                                 0, 0,
+#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
+                                 mi_x, mi_y);
+    } else {
+      build_inter_predictors(xd, plane,
+#if CONFIG_MOTION_VAR
+                             0, 0,
+#endif  // CONFIG_MOTION_VAR
+                             0, bw, bh, 0, 0, bw, bh,
+#if CONFIG_SUPERTX && CONFIG_EXT_INTER
+                             0, 0,
+#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
+                             mi_x, mi_y);
+    }
+  }
+}
+
+void av1_build_inter_predictors_sby(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                    BUFFER_SET *ctx, BLOCK_SIZE bsize) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, 0, 0);
+#if CONFIG_EXT_INTER
+  if (is_interintra_pred(&xd->mi[0]->mbmi)) {
+    BUFFER_SET default_ctx = { { xd->plane[0].dst.buf, NULL, NULL },
+                               { xd->plane[0].dst.stride, 0, 0 } };
+    if (!ctx) ctx = &default_ctx;
+    av1_build_interintra_predictors_sby(xd, xd->plane[0].dst.buf,
+                                        xd->plane[0].dst.stride, ctx, bsize);
+  }
+#else
+  (void)ctx;
+#endif  // CONFIG_EXT_INTER
+}
+
+void av1_build_inter_predictors_sbuv(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                     BUFFER_SET *ctx, BLOCK_SIZE bsize) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, 1,
+                                    MAX_MB_PLANE - 1);
+#if CONFIG_EXT_INTER
+  if (is_interintra_pred(&xd->mi[0]->mbmi)) {
+    BUFFER_SET default_ctx = {
+      { NULL, xd->plane[1].dst.buf, xd->plane[2].dst.buf },
+      { 0, xd->plane[1].dst.stride, xd->plane[2].dst.stride }
+    };
+    if (!ctx) ctx = &default_ctx;
+    av1_build_interintra_predictors_sbuv(
+        xd, xd->plane[1].dst.buf, xd->plane[2].dst.buf, xd->plane[1].dst.stride,
+        xd->plane[2].dst.stride, ctx, bsize);
+  }
+#else
+  (void)ctx;
+#endif  // CONFIG_EXT_INTER
+}
+
+// TODO(afergs): Check if ctx can be made constant
+void av1_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                   BUFFER_SET *ctx, BLOCK_SIZE bsize) {
+  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, 0,
+                                    MAX_MB_PLANE - 1);
+#if CONFIG_EXT_INTER
+  if (is_interintra_pred(&xd->mi[0]->mbmi)) {
+    BUFFER_SET default_ctx = {
+      { xd->plane[0].dst.buf, xd->plane[1].dst.buf, xd->plane[2].dst.buf },
+      { xd->plane[0].dst.stride, xd->plane[1].dst.stride,
+        xd->plane[2].dst.stride }
+    };
+    if (!ctx) ctx = &default_ctx;
+    av1_build_interintra_predictors(
+        xd, xd->plane[0].dst.buf, xd->plane[1].dst.buf, xd->plane[2].dst.buf,
+        xd->plane[0].dst.stride, xd->plane[1].dst.stride,
+        xd->plane[2].dst.stride, ctx, bsize);
+  }
+#else
+  (void)ctx;
+#endif  // CONFIG_EXT_INTER
+}
+
+void av1_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
+                          BLOCK_SIZE bsize, const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col) {
+  uint8_t *const buffers[MAX_MB_PLANE] = { src->y_buffer, src->u_buffer,
+                                           src->v_buffer };
+  const int widths[MAX_MB_PLANE] = { src->y_crop_width, src->uv_crop_width,
+                                     src->uv_crop_width };
+  const int heights[MAX_MB_PLANE] = { src->y_crop_height, src->uv_crop_height,
+                                      src->uv_crop_height };
+  const int strides[MAX_MB_PLANE] = { src->y_stride, src->uv_stride,
+                                      src->uv_stride };
+  int i;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    struct macroblockd_plane *const pd = &planes[i];
+    setup_pred_plane(&pd->dst, bsize, buffers[i], widths[i], heights[i],
+                     strides[i], mi_row, mi_col, NULL, pd->subsampling_x,
+                     pd->subsampling_y);
+  }
+}
+
+void av1_setup_pre_planes(MACROBLOCKD *xd, int idx,
+                          const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
+                          const struct scale_factors *sf) {
+  if (src != NULL) {
+    int i;
+    uint8_t *const buffers[MAX_MB_PLANE] = { src->y_buffer, src->u_buffer,
+                                             src->v_buffer };
+    const int widths[MAX_MB_PLANE] = { src->y_crop_width, src->uv_crop_width,
+                                       src->uv_crop_width };
+    const int heights[MAX_MB_PLANE] = { src->y_crop_height, src->uv_crop_height,
+                                        src->uv_crop_height };
+    const int strides[MAX_MB_PLANE] = { src->y_stride, src->uv_stride,
+                                        src->uv_stride };
+    for (i = 0; i < MAX_MB_PLANE; ++i) {
+      struct macroblockd_plane *const pd = &xd->plane[i];
+      setup_pred_plane(&pd->pre[idx], xd->mi[0]->mbmi.sb_type, buffers[i],
+                       widths[i], heights[i], strides[i], mi_row, mi_col, sf,
+                       pd->subsampling_x, pd->subsampling_y);
+    }
+  }
+}
+
+#if CONFIG_SUPERTX
+#if CONFIG_CB4X4
+static const uint8_t mask_4[4] = { 64, 52, 12, 0 };
+static const uint8_t mask_4_uv[4] = { 64, 52, 12, 0 };
+#endif  // CONFIG_CB4X4
+static const uint8_t mask_8[8] = { 64, 64, 62, 52, 12, 2, 0, 0 };
+
+static const uint8_t mask_16[16] = { 63, 62, 60, 58, 55, 50, 43, 36,
+                                     28, 21, 14, 9,  6,  4,  2,  1 };
+
+static const uint8_t mask_32[32] = { 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 63,
+                                     61, 57, 52, 45, 36, 28, 19, 12, 7,  3,  1,
+                                     0,  0,  0,  0,  0,  0,  0,  0,  0,  0 };
+
+static const uint8_t mask_8_uv[8] = { 64, 64, 62, 52, 12, 2, 0, 0 };
+
+static const uint8_t mask_16_uv[16] = { 64, 64, 64, 64, 61, 53, 45, 36,
+                                        28, 19, 11, 3,  0,  0,  0,  0 };
+
+static const uint8_t mask_32_uv[32] = { 64, 64, 64, 64, 64, 64, 64, 64,
+                                        64, 64, 64, 64, 60, 54, 46, 36,
+                                        28, 18, 10, 4,  0,  0,  0,  0,
+                                        0,  0,  0,  0,  0,  0,  0,  0 };
+
+static const uint8_t *get_supertx_mask(int length, int plane) {
+  switch (length) {
+#if CONFIG_CB4X4
+    case 4: return plane ? mask_4_uv : mask_4;
+#endif  // CONFIG_CB4X4
+    case 8: return plane ? mask_8_uv : mask_8;
+    case 16: return plane ? mask_16_uv : mask_16;
+    case 32: return plane ? mask_32_uv : mask_32;
+    default: assert(0);
+  }
+  return NULL;
+}
+
+void av1_build_masked_inter_predictor_complex(
+    MACROBLOCKD *xd, uint8_t *dst, int dst_stride, const uint8_t *pre,
+    int pre_stride, int mi_row, int mi_col, int mi_row_ori, int mi_col_ori,
+    BLOCK_SIZE bsize, BLOCK_SIZE top_bsize, PARTITION_TYPE partition,
+    int plane) {
+  const struct macroblockd_plane *pd = &xd->plane[plane];
+  const int ssx = pd->subsampling_x;
+  const int ssy = pd->subsampling_y;
+  const int top_w = block_size_wide[top_bsize] >> ssx;
+  const int top_h = block_size_high[top_bsize] >> ssy;
+  const int w = block_size_wide[bsize] >> ssx;
+  const int h = block_size_high[bsize] >> ssy;
+  const int w_offset = ((mi_col - mi_col_ori) * MI_SIZE) >> ssx;
+  const int h_offset = ((mi_row - mi_row_ori) * MI_SIZE) >> ssy;
+
+  int w_remain, h_remain;
+
+#if CONFIG_HIGHBITDEPTH
+  const int is_hdb = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0;
+#endif  // CONFIG_HIGHBITDEPTH
+
+  assert(bsize <= BLOCK_32X32);
+  assert(IMPLIES(plane == 0, ssx == 0));
+  assert(IMPLIES(plane == 0, ssy == 0));
+
+  switch (partition) {
+    case PARTITION_HORZ: {
+      const uint8_t *const mask = get_supertx_mask(h, ssy);
+
+      w_remain = top_w;
+      h_remain = top_h - h_offset - h;
+      dst += h_offset * dst_stride;
+      pre += h_offset * pre_stride;
+
+#if CONFIG_HIGHBITDEPTH
+      if (is_hdb)
+        aom_highbd_blend_a64_vmask(dst, dst_stride, dst, dst_stride, pre,
+                                   pre_stride, mask, h, top_w, xd->bd);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        aom_blend_a64_vmask(dst, dst_stride, dst, dst_stride, pre, pre_stride,
+                            mask, h, top_w);
+
+      dst += h * dst_stride;
+      pre += h * pre_stride;
+      break;
+    }
+    case PARTITION_VERT: {
+      const uint8_t *const mask = get_supertx_mask(w, ssx);
+
+      w_remain = top_w - w_offset - w;
+      h_remain = top_h;
+      dst += w_offset;
+      pre += w_offset;
+
+#if CONFIG_HIGHBITDEPTH
+      if (is_hdb)
+        aom_highbd_blend_a64_hmask(dst, dst_stride, dst, dst_stride, pre,
+                                   pre_stride, mask, top_h, w, xd->bd);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        aom_blend_a64_hmask(dst, dst_stride, dst, dst_stride, pre, pre_stride,
+                            mask, top_h, w);
+
+      dst += w;
+      pre += w;
+      break;
+    }
+    default: {
+      assert(0);
+      return;
+    }
+  }
+
+  if (w_remain == 0 || h_remain == 0) {
+    return;
+  }
+
+#if CONFIG_HIGHBITDEPTH
+  if (is_hdb) {
+    dst = (uint8_t *)CONVERT_TO_SHORTPTR(dst);
+    pre = (const uint8_t *)CONVERT_TO_SHORTPTR(pre);
+    dst_stride *= 2;
+    pre_stride *= 2;
+    w_remain *= 2;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  do {
+    memcpy(dst, pre, w_remain * sizeof(uint8_t));
+    dst += dst_stride;
+    pre += pre_stride;
+  } while (--h_remain);
+}
+
+void av1_build_inter_predictors_sb_sub8x8_extend(MACROBLOCKD *xd,
+#if CONFIG_EXT_INTER
+                                                 int mi_row_ori, int mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                                                 int mi_row, int mi_col,
+                                                 BLOCK_SIZE bsize, int block) {
+  // Prediction function used in supertx:
+  // Use the mv at current block (which is less than 8x8)
+  // to get prediction of a block located at (mi_row, mi_col) at size of bsize
+  // bsize can be larger than 8x8.
+  // block (0-3): the sub8x8 location of current block
+  int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
+#if CONFIG_EXT_INTER
+  const int wedge_offset_x = (mi_col_ori - mi_col) * MI_SIZE;
+  const int wedge_offset_y = (mi_row_ori - mi_row) * MI_SIZE;
+#endif  // CONFIG_EXT_INTER
+
+  // For sub8x8 uv:
+  // Skip uv prediction in supertx except the first block (block = 0)
+  int max_plane = block ? 1 : MAX_MB_PLANE;
+
+  for (plane = 0; plane < max_plane; plane++) {
+    const BLOCK_SIZE plane_bsize =
+        get_plane_block_size(bsize, &xd->plane[plane]);
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+    const int bw = 4 * num_4x4_w;
+    const int bh = 4 * num_4x4_h;
+
+    build_inter_predictors(xd, plane,
+#if CONFIG_MOTION_VAR
+                           0, 0,
+#endif  // CONFIG_MOTION_VAR
+                           block, bw, bh, 0, 0, bw, bh,
+#if CONFIG_EXT_INTER
+                           wedge_offset_x, wedge_offset_y,
+#endif  // CONFIG_EXT_INTER
+                           mi_x, mi_y);
+  }
+#if CONFIG_EXT_INTER
+  if (is_interintra_pred(&xd->mi[0]->mbmi)) {
+    BUFFER_SET ctx = { { xd->plane[0].dst.buf, xd->plane[1].dst.buf,
+                         xd->plane[2].dst.buf },
+                       { xd->plane[0].dst.stride, xd->plane[1].dst.stride,
+                         xd->plane[2].dst.stride } };
+    av1_build_interintra_predictors(
+        xd, xd->plane[0].dst.buf, xd->plane[1].dst.buf, xd->plane[2].dst.buf,
+        xd->plane[0].dst.stride, xd->plane[1].dst.stride,
+        xd->plane[2].dst.stride, &ctx, bsize);
+  }
+#endif  // CONFIG_EXT_INTER
+}
+
+void av1_build_inter_predictors_sb_extend(MACROBLOCKD *xd,
+#if CONFIG_EXT_INTER
+                                          int mi_row_ori, int mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                                          int mi_row, int mi_col,
+                                          BLOCK_SIZE bsize) {
+  int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
+#if CONFIG_EXT_INTER
+  const int wedge_offset_x = (mi_col_ori - mi_col) * MI_SIZE;
+  const int wedge_offset_y = (mi_row_ori - mi_row) * MI_SIZE;
+#endif  // CONFIG_EXT_INTER
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    const BLOCK_SIZE plane_bsize =
+        get_plane_block_size(bsize, &xd->plane[plane]);
+    const int bw = block_size_wide[plane_bsize];
+    const int bh = block_size_high[plane_bsize];
+
+    build_inter_predictors(xd, plane,
+#if CONFIG_MOTION_VAR
+                           0, 0,
+#endif  // CONFIG_MOTION_VAR
+                           0, bw, bh, 0, 0, bw, bh,
+#if CONFIG_EXT_INTER
+                           wedge_offset_x, wedge_offset_y,
+#endif  // CONFIG_EXT_INTER
+                           mi_x, mi_y);
+  }
+}
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_MOTION_VAR
+// obmc_mask_N[overlap_position]
+static const uint8_t obmc_mask_1[1] = { 64 };
+
+static const uint8_t obmc_mask_2[2] = { 45, 64 };
+
+static const uint8_t obmc_mask_4[4] = { 39, 50, 59, 64 };
+
+static const uint8_t obmc_mask_8[8] = { 36, 42, 48, 53, 57, 61, 64, 64 };
+
+static const uint8_t obmc_mask_16[16] = { 34, 37, 40, 43, 46, 49, 52, 54,
+                                          56, 58, 60, 61, 64, 64, 64, 64 };
+
+static const uint8_t obmc_mask_32[32] = { 33, 35, 36, 38, 40, 41, 43, 44,
+                                          45, 47, 48, 50, 51, 52, 53, 55,
+                                          56, 57, 58, 59, 60, 60, 61, 62,
+                                          64, 64, 64, 64, 64, 64, 64, 64 };
+
+#if CONFIG_EXT_PARTITION
+static const uint8_t obmc_mask_64[64] = {
+  33, 34, 35, 35, 36, 37, 38, 39, 40, 40, 41, 42, 43, 44, 44, 44,
+  45, 46, 47, 47, 48, 49, 50, 51, 51, 51, 52, 52, 53, 54, 55, 56,
+  56, 56, 57, 57, 58, 58, 59, 60, 60, 60, 60, 60, 61, 62, 62, 62,
+  62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
+};
+#endif  // CONFIG_EXT_PARTITION
+
+const uint8_t *av1_get_obmc_mask(int length) {
+  switch (length) {
+    case 1: return obmc_mask_1;
+    case 2: return obmc_mask_2;
+    case 4: return obmc_mask_4;
+    case 8: return obmc_mask_8;
+    case 16: return obmc_mask_16;
+    case 32: return obmc_mask_32;
+#if CONFIG_EXT_PARTITION
+    case 64: return obmc_mask_64;
+#endif  // CONFIG_EXT_PARTITION
+    default: assert(0); return NULL;
+  }
+}
+
+#if CONFIG_NCOBMC
+// obmc_mask_flipN[overlap_position]
+static const uint8_t obmc_mask_flip1[1] = { 55 };
+
+static const uint8_t obmc_mask_flip2[2] = { 62, 45 };
+
+static const uint8_t obmc_mask_flip4[4] = { 64, 59, 50, 39 };
+
+static const uint8_t obmc_mask_flip8[8] = { 64, 63, 61, 57, 53, 48, 42, 36 };
+
+static const uint8_t obmc_mask_flip16[16] = { 64, 64, 64, 63, 61, 60, 58, 56,
+                                              54, 52, 49, 46, 43, 40, 37, 34 };
+
+static const uint8_t obmc_mask_flip32[32] = { 64, 64, 64, 64, 64, 63, 63, 62,
+                                              62, 61, 60, 60, 59, 58, 57, 56,
+                                              55, 53, 52, 51, 50, 48, 47, 45,
+                                              44, 43, 41, 40, 38, 36, 35, 33 };
+
+#if CONFIG_EXT_PARTITION
+static const uint8_t obmc_mask_flip64[64] = {
+  64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 63, 63, 63, 63, 62, 62,
+  62, 62, 62, 61, 60, 60, 60, 60, 60, 59, 58, 58, 57, 57, 56, 56,
+  56, 55, 54, 53, 52, 52, 51, 51, 51, 50, 49, 48, 47, 47, 46, 45,
+  44, 44, 44, 43, 42, 41, 40, 40, 39, 38, 37, 36, 35, 35, 34, 33,
+};
+#endif  // CONFIG_EXT_PARTITION
+
+const uint8_t *av1_get_obmc_mask_flipped(int length) {
+  switch (length) {
+    case 1: return obmc_mask_flip1;
+    case 2: return obmc_mask_flip2;
+    case 4: return obmc_mask_flip4;
+    case 8: return obmc_mask_flip8;
+    case 16: return obmc_mask_flip16;
+    case 32: return obmc_mask_flip32;
+#if CONFIG_EXT_PARTITION
+    case 64: return obmc_mask_flip64;
+#endif  // CONFIG_EXT_PARTITION
+    default: assert(0); return NULL;
+  }
+}
+#endif  // CONFIG_NCOBMC
+
+void av1_count_overlappable_neighbors(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                      int mi_row, int mi_col) {
+  int i, mi_step;
+
+  xd->mi[0]->mbmi.overlappable_neighbors[0] = 0;
+  xd->mi[0]->mbmi.overlappable_neighbors[1] = 0;
+
+  if (xd->up_available) {
+    const int ilimit = AOMMIN(xd->n8_w, cm->mi_cols - mi_col);
+    for (i = 0; i < ilimit; i += mi_step) {
+      int mi_row_offset = -1;
+      int mi_col_offset = i;
+      MODE_INFO *above_mi =
+          xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+      MB_MODE_INFO *above_mbmi = &above_mi->mbmi;
+
+      mi_step = AOMMIN(xd->n8_w, mi_size_wide[above_mbmi->sb_type]);
+
+      if (is_neighbor_overlappable(above_mbmi))
+        xd->mi[0]->mbmi.overlappable_neighbors[0]++;
+    }
+  }
+
+  if (xd->left_available) {
+    const int ilimit = AOMMIN(xd->n8_h, cm->mi_rows - mi_row);
+    for (i = 0; i < ilimit; i += mi_step) {
+      int mi_row_offset = i;
+      int mi_col_offset = -1;
+      MODE_INFO *left_mi =
+          xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+      MB_MODE_INFO *left_mbmi = &left_mi->mbmi;
+
+      mi_step = AOMMIN(xd->n8_h, mi_size_high[left_mbmi->sb_type]);
+
+      if (is_neighbor_overlappable(left_mbmi))
+        xd->mi[0]->mbmi.overlappable_neighbors[1]++;
+    }
+  }
+}
+
+// HW does not support < 4x4 prediction. To limit the bandwidth requirement, for
+// small blocks, only blend with neighbors from one side. If block-size of
+// current plane is 4x4 or 8x4, the above neighbor (dir = 0) will be skipped. If
+// it is 4x8, the left neighbor (dir = 1) will be skipped.
+#define DISABLE_CHROMA_U8X8_OBMC 0  // 0: one-sided obmc; 1: disable
+
+int skip_u4x4_pred_in_obmc(BLOCK_SIZE bsize, const struct macroblockd_plane *pd,
+                           int dir) {
+  assert(is_motion_variation_allowed_bsize(bsize));
+
+  BLOCK_SIZE bsize_plane =
+      ss_size_lookup[bsize][pd->subsampling_x][pd->subsampling_y];
+#if CONFIG_CB4X4
+  if (bsize_plane < BLOCK_4X4) return 1;
+#endif
+  switch (bsize_plane) {
+#if DISABLE_CHROMA_U8X8_OBMC
+    case BLOCK_4X4:
+    case BLOCK_8X4:
+    case BLOCK_4X8: return 1; break;
+#else
+    case BLOCK_4X4:
+    case BLOCK_8X4:
+    case BLOCK_4X8: return dir == 1; break;
+#endif
+    default: return 0;
+  }
+}
+
+// This function combines motion compensated predictions that is generated by
+// top/left neighboring blocks' inter predictors with the regular inter
+// prediction. We assume the original prediction (bmc) is stored in
+// xd->plane[].dst.buf
+void av1_build_obmc_inter_prediction(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                     int mi_row, int mi_col,
+                                     uint8_t *above[MAX_MB_PLANE],
+                                     int above_stride[MAX_MB_PLANE],
+                                     uint8_t *left[MAX_MB_PLANE],
+                                     int left_stride[MAX_MB_PLANE]) {
+  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  int plane, i;
+#if CONFIG_HIGHBITDEPTH
+  const int is_hbd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0;
+#endif  // CONFIG_HIGHBITDEPTH
+
+  // handle above row
+  if (xd->up_available) {
+    const int overlap = num_4x4_blocks_high_lookup[bsize] * 2;
+    const int miw = AOMMIN(xd->n8_w, cm->mi_cols - mi_col);
+    const int mi_row_offset = -1;
+    const int neighbor_limit = max_neighbor_obmc[b_width_log2_lookup[bsize]];
+    int neighbor_count = 0;
+
+    assert(miw > 0);
+
+    i = 0;
+    do {  // for each mi in the above row
+      const int mi_col_offset = i;
+      const MB_MODE_INFO *const above_mbmi =
+          &xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride]->mbmi;
+      const BLOCK_SIZE a_bsize = above_mbmi->sb_type;
+      const int mi_step = AOMMIN(xd->n8_w, mi_size_wide[a_bsize]);
+
+      if (is_neighbor_overlappable(above_mbmi)) {
+        neighbor_count++;
+        if (neighbor_count > neighbor_limit) break;
+        for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+          const struct macroblockd_plane *pd = &xd->plane[plane];
+          const int bw = (mi_step * MI_SIZE) >> pd->subsampling_x;
+          const int bh = overlap >> pd->subsampling_y;
+
+          if (skip_u4x4_pred_in_obmc(bsize, pd, 0)) continue;
+
+          const int dst_stride = pd->dst.stride;
+          uint8_t *const dst = &pd->dst.buf[(i * MI_SIZE) >> pd->subsampling_x];
+          const int tmp_stride = above_stride[plane];
+          const uint8_t *const tmp =
+              &above[plane][(i * MI_SIZE) >> pd->subsampling_x];
+          const uint8_t *const mask = av1_get_obmc_mask(bh);
+
+#if CONFIG_HIGHBITDEPTH
+          if (is_hbd)
+            aom_highbd_blend_a64_vmask(dst, dst_stride, dst, dst_stride, tmp,
+                                       tmp_stride, mask, bh, bw, xd->bd);
+          else
+#endif  // CONFIG_HIGHBITDEPTH
+            aom_blend_a64_vmask(dst, dst_stride, dst, dst_stride, tmp,
+                                tmp_stride, mask, bh, bw);
+        }
+      }
+      i += mi_step;
+    } while (i < miw);
+  }
+
+  // handle left column
+  if (xd->left_available) {
+    const int overlap = num_4x4_blocks_wide_lookup[bsize] * 2;
+    const int mih = AOMMIN(xd->n8_h, cm->mi_rows - mi_row);
+    const int mi_col_offset = -1;
+    const int neighbor_limit = max_neighbor_obmc[b_height_log2_lookup[bsize]];
+    int neighbor_count = 0;
+
+    assert(mih > 0);
+
+    i = 0;
+    do {  // for each mi in the left column
+      const int mi_row_offset = i;
+      const MB_MODE_INFO *const left_mbmi =
+          &xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride]->mbmi;
+      const BLOCK_SIZE l_bsize = left_mbmi->sb_type;
+      const int mi_step = AOMMIN(xd->n8_h, mi_size_high[l_bsize]);
+
+      if (is_neighbor_overlappable(left_mbmi)) {
+        neighbor_count++;
+        if (neighbor_count > neighbor_limit) break;
+        for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+          const struct macroblockd_plane *pd = &xd->plane[plane];
+          const int bw = overlap >> pd->subsampling_x;
+          const int bh = (mi_step * MI_SIZE) >> pd->subsampling_y;
+
+          if (skip_u4x4_pred_in_obmc(bsize, pd, 1)) continue;
+
+          const int dst_stride = pd->dst.stride;
+          uint8_t *const dst =
+              &pd->dst.buf[(i * MI_SIZE * dst_stride) >> pd->subsampling_y];
+          const int tmp_stride = left_stride[plane];
+          const uint8_t *const tmp =
+              &left[plane][(i * MI_SIZE * tmp_stride) >> pd->subsampling_y];
+          const uint8_t *const mask = av1_get_obmc_mask(bw);
+
+#if CONFIG_HIGHBITDEPTH
+          if (is_hbd)
+            aom_highbd_blend_a64_hmask(dst, dst_stride, dst, dst_stride, tmp,
+                                       tmp_stride, mask, bh, bw, xd->bd);
+          else
+#endif  // CONFIG_HIGHBITDEPTH
+            aom_blend_a64_hmask(dst, dst_stride, dst, dst_stride, tmp,
+                                tmp_stride, mask, bh, bw);
+        }
+      }
+      i += mi_step;
+    } while (i < mih);
+  }
+}
+
+void modify_neighbor_predictor_for_obmc(MB_MODE_INFO *mbmi) {
+#if CONFIG_EXT_INTER
+  if (is_interintra_pred(mbmi)) {
+    mbmi->ref_frame[1] = NONE_FRAME;
+  } else if (has_second_ref(mbmi) &&
+             is_masked_compound_type(mbmi->interinter_compound_type)) {
+    mbmi->interinter_compound_type = COMPOUND_AVERAGE;
+    mbmi->ref_frame[1] = NONE_FRAME;
+  }
+#endif  // CONFIG_EXT_INTER
+  if (has_second_ref(mbmi)) mbmi->ref_frame[1] = NONE_FRAME;
+  return;
+}
+
+void av1_build_prediction_by_above_preds(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                         int mi_row, int mi_col,
+                                         uint8_t *tmp_buf[MAX_MB_PLANE],
+                                         int tmp_width[MAX_MB_PLANE],
+                                         int tmp_height[MAX_MB_PLANE],
+                                         int tmp_stride[MAX_MB_PLANE]) {
+  const TileInfo *const tile = &xd->tile;
+  BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  int i, j, mi_step, ref;
+  const int ilimit = AOMMIN(xd->n8_w, cm->mi_cols - mi_col);
+  int mb_to_right_edge_base = xd->mb_to_right_edge;
+  const int neighbor_limit = max_neighbor_obmc[b_width_log2_lookup[bsize]];
+  int neighbor_count = 0;
+
+  if (mi_row <= tile->mi_row_start) return;
+
+  xd->mb_to_bottom_edge += xd->n8_h * 32;
+  for (i = 0; i < ilimit; i += mi_step) {
+    int mi_row_offset = -1;
+    int mi_col_offset = i;
+    int mi_x, mi_y, bw, bh;
+    MODE_INFO *above_mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+    MB_MODE_INFO *above_mbmi = &above_mi->mbmi;
+    const BLOCK_SIZE a_bsize = above_mbmi->sb_type;
+    MB_MODE_INFO backup_mbmi;
+
+    mi_step = AOMMIN(xd->n8_w, mi_size_wide[a_bsize]);
+
+    if (!is_neighbor_overlappable(above_mbmi)) continue;
+
+    neighbor_count++;
+    if (neighbor_count > neighbor_limit) break;
+
+    backup_mbmi = *above_mbmi;
+    modify_neighbor_predictor_for_obmc(above_mbmi);
+
+    for (j = 0; j < MAX_MB_PLANE; ++j) {
+      struct macroblockd_plane *const pd = &xd->plane[j];
+      setup_pred_plane(&pd->dst, AOMMAX(a_bsize, BLOCK_8X8), tmp_buf[j],
+                       tmp_width[j], tmp_height[j], tmp_stride[j], 0, i, NULL,
+                       pd->subsampling_x, pd->subsampling_y);
+    }
+    for (ref = 0; ref < 1 + has_second_ref(above_mbmi); ++ref) {
+      const MV_REFERENCE_FRAME frame = above_mbmi->ref_frame[ref];
+      const RefBuffer *const ref_buf = &cm->frame_refs[frame - LAST_FRAME];
+
+      xd->block_refs[ref] = ref_buf;
+      if ((!av1_is_valid_scale(&ref_buf->sf)))
+        aom_internal_error(xd->error_info, AOM_CODEC_UNSUP_BITSTREAM,
+                           "Reference frame has invalid dimensions");
+      av1_setup_pre_planes(xd, ref, ref_buf->buf, mi_row, mi_col + i,
+                           &ref_buf->sf);
+    }
+
+    xd->mb_to_left_edge = -(((mi_col + i) * MI_SIZE) * 8);
+    xd->mb_to_right_edge =
+        mb_to_right_edge_base + (xd->n8_w - i - mi_step) * 64;
+    mi_x = (mi_col + i) << MI_SIZE_LOG2;
+    mi_y = mi_row << MI_SIZE_LOG2;
+
+    for (j = 0; j < MAX_MB_PLANE; ++j) {
+      const struct macroblockd_plane *pd = &xd->plane[j];
+      bw = (mi_step * MI_SIZE) >> pd->subsampling_x;
+      bh = AOMMAX((num_4x4_blocks_high_lookup[bsize] * 2) >> pd->subsampling_y,
+                  4);
+
+      if (skip_u4x4_pred_in_obmc(bsize, pd, 0)) continue;
+      build_inter_predictors(xd, j, mi_col_offset, mi_row_offset, 0, bw, bh, 0,
+                             0, bw, bh,
+#if CONFIG_SUPERTX && CONFIG_EXT_INTER
+                             0, 0,
+#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
+                             mi_x, mi_y);
+    }
+    *above_mbmi = backup_mbmi;
+  }
+  xd->mb_to_left_edge = -((mi_col * MI_SIZE) * 8);
+  xd->mb_to_right_edge = mb_to_right_edge_base;
+  xd->mb_to_bottom_edge -= xd->n8_h * 32;
+}
+
+void av1_build_prediction_by_left_preds(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                        int mi_row, int mi_col,
+                                        uint8_t *tmp_buf[MAX_MB_PLANE],
+                                        int tmp_width[MAX_MB_PLANE],
+                                        int tmp_height[MAX_MB_PLANE],
+                                        int tmp_stride[MAX_MB_PLANE]) {
+  const TileInfo *const tile = &xd->tile;
+  BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  int i, j, mi_step, ref;
+  const int ilimit = AOMMIN(xd->n8_h, cm->mi_rows - mi_row);
+  int mb_to_bottom_edge_base = xd->mb_to_bottom_edge;
+  const int neighbor_limit = max_neighbor_obmc[b_height_log2_lookup[bsize]];
+  int neighbor_count = 0;
+
+  if (mi_col == 0 || (mi_col - 1 < tile->mi_col_start)) return;
+
+  xd->mb_to_right_edge += xd->n8_w * 32;
+  for (i = 0; i < ilimit; i += mi_step) {
+    int mi_row_offset = i;
+    int mi_col_offset = -1;
+    int mi_x, mi_y, bw, bh;
+    MODE_INFO *left_mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+    MB_MODE_INFO *left_mbmi = &left_mi->mbmi;
+    const BLOCK_SIZE l_bsize = left_mbmi->sb_type;
+    MB_MODE_INFO backup_mbmi;
+
+    mi_step = AOMMIN(xd->n8_h, mi_size_high[l_bsize]);
+
+    if (!is_neighbor_overlappable(left_mbmi)) continue;
+
+    neighbor_count++;
+    if (neighbor_count > neighbor_limit) break;
+
+    backup_mbmi = *left_mbmi;
+    modify_neighbor_predictor_for_obmc(left_mbmi);
+
+    for (j = 0; j < MAX_MB_PLANE; ++j) {
+      struct macroblockd_plane *const pd = &xd->plane[j];
+      setup_pred_plane(&pd->dst, AOMMAX(l_bsize, BLOCK_8X8), tmp_buf[j],
+                       tmp_width[j], tmp_height[j], tmp_stride[j], i, 0, NULL,
+                       pd->subsampling_x, pd->subsampling_y);
+    }
+    for (ref = 0; ref < 1 + has_second_ref(left_mbmi); ++ref) {
+      const MV_REFERENCE_FRAME frame = left_mbmi->ref_frame[ref];
+      const RefBuffer *const ref_buf = &cm->frame_refs[frame - LAST_FRAME];
+
+      xd->block_refs[ref] = ref_buf;
+      if ((!av1_is_valid_scale(&ref_buf->sf)))
+        aom_internal_error(xd->error_info, AOM_CODEC_UNSUP_BITSTREAM,
+                           "Reference frame has invalid dimensions");
+      av1_setup_pre_planes(xd, ref, ref_buf->buf, mi_row + i, mi_col,
+                           &ref_buf->sf);
+    }
+
+    xd->mb_to_top_edge = -(((mi_row + i) * MI_SIZE) * 8);
+    xd->mb_to_bottom_edge =
+        mb_to_bottom_edge_base + (xd->n8_h - i - mi_step) * 64;
+    mi_x = mi_col << MI_SIZE_LOG2;
+    mi_y = (mi_row + i) << MI_SIZE_LOG2;
+
+    for (j = 0; j < MAX_MB_PLANE; ++j) {
+      const struct macroblockd_plane *pd = &xd->plane[j];
+      bw = AOMMAX((num_4x4_blocks_wide_lookup[bsize] * 2) >> pd->subsampling_x,
+                  4);
+      bh = (mi_step << MI_SIZE_LOG2) >> pd->subsampling_y;
+
+      if (skip_u4x4_pred_in_obmc(bsize, pd, 1)) continue;
+      build_inter_predictors(xd, j, mi_col_offset, mi_row_offset, 0, bw, bh, 0,
+                             0, bw, bh,
+#if CONFIG_SUPERTX && CONFIG_EXT_INTER
+                             0, 0,
+#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
+                             mi_x, mi_y);
+    }
+    *left_mbmi = backup_mbmi;
+  }
+  xd->mb_to_top_edge = -((mi_row * MI_SIZE) * 8);
+  xd->mb_to_bottom_edge = mb_to_bottom_edge_base;
+  xd->mb_to_right_edge -= xd->n8_w * 32;
+}
+
+void av1_build_obmc_inter_predictors_sb(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                        int mi_row, int mi_col) {
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf1[2 * MAX_MB_PLANE * MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf2[2 * MAX_MB_PLANE * MAX_SB_SQUARE]);
+#else
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf1[MAX_MB_PLANE * MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf2[MAX_MB_PLANE * MAX_SB_SQUARE]);
+#endif  // CONFIG_HIGHBITDEPTH
+  uint8_t *dst_buf1[MAX_MB_PLANE], *dst_buf2[MAX_MB_PLANE];
+  int dst_stride1[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_stride2[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_width1[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_width2[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_height1[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_height2[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    int len = sizeof(uint16_t);
+    dst_buf1[0] = CONVERT_TO_BYTEPTR(tmp_buf1);
+    dst_buf1[1] = CONVERT_TO_BYTEPTR(tmp_buf1 + MAX_SB_SQUARE * len);
+    dst_buf1[2] = CONVERT_TO_BYTEPTR(tmp_buf1 + MAX_SB_SQUARE * 2 * len);
+    dst_buf2[0] = CONVERT_TO_BYTEPTR(tmp_buf2);
+    dst_buf2[1] = CONVERT_TO_BYTEPTR(tmp_buf2 + MAX_SB_SQUARE * len);
+    dst_buf2[2] = CONVERT_TO_BYTEPTR(tmp_buf2 + MAX_SB_SQUARE * 2 * len);
+  } else {
+#endif  // CONFIG_HIGHBITDEPTH
+    dst_buf1[0] = tmp_buf1;
+    dst_buf1[1] = tmp_buf1 + MAX_SB_SQUARE;
+    dst_buf1[2] = tmp_buf1 + MAX_SB_SQUARE * 2;
+    dst_buf2[0] = tmp_buf2;
+    dst_buf2[1] = tmp_buf2 + MAX_SB_SQUARE;
+    dst_buf2[2] = tmp_buf2 + MAX_SB_SQUARE * 2;
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  av1_build_prediction_by_above_preds(cm, xd, mi_row, mi_col, dst_buf1,
+                                      dst_width1, dst_height1, dst_stride1);
+  av1_build_prediction_by_left_preds(cm, xd, mi_row, mi_col, dst_buf2,
+                                     dst_width2, dst_height2, dst_stride2);
+  av1_setup_dst_planes(xd->plane, xd->mi[0]->mbmi.sb_type,
+                       get_frame_new_buffer(cm), mi_row, mi_col);
+  av1_build_obmc_inter_prediction(cm, xd, mi_row, mi_col, dst_buf1, dst_stride1,
+                                  dst_buf2, dst_stride2);
+}
+
+#if CONFIG_NCOBMC
+void av1_build_prediction_by_bottom_preds(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                          int mi_row, int mi_col,
+                                          uint8_t *tmp_buf[MAX_MB_PLANE],
+                                          int tmp_width[MAX_MB_PLANE],
+                                          int tmp_height[MAX_MB_PLANE],
+                                          int tmp_stride[MAX_MB_PLANE]) {
+  const TileInfo *const tile = &xd->tile;
+  BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  int i, j, mi_step, ref;
+  const int ilimit = AOMMIN(xd->n8_w, cm->mi_cols - mi_col);
+  int mb_to_right_edge_base = xd->mb_to_right_edge;
+
+  if (mi_row + xd->n8_h >= tile->mi_row_end ||
+      (mi_row + xd->n8_h) % MI_SIZE == 0 || (mi_row + xd->n8_h) >= cm->mi_rows)
+    return;
+  assert(bsize >= BLOCK_8X8);
+
+  xd->mb_to_top_edge -= xd->n8_h * 32;
+  for (i = 0; i < ilimit; i += mi_step) {
+    int mi_row_offset = xd->n8_h;
+    int mi_col_offset = i;
+    int mi_x, mi_y, bw, bh;
+    MODE_INFO *mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+    MB_MODE_INFO *mbmi = &mi->mbmi;
+#if CONFIG_EXT_INTER
+    MB_MODE_INFO backup_mbmi;
+#endif  // CONFIG_EXT_INTER
+
+    mi_step = AOMMIN(xd->n8_w, mi_size_wide[mbmi->sb_type]);
+
+    if (!is_neighbor_overlappable(mbmi)) continue;
+
+#if CONFIG_EXT_INTER
+    backup_mbmi = *mbmi;
+    modify_neighbor_predictor_for_obmc(mbmi);
+#endif  // CONFIG_EXT_INTER
+
+    for (j = 0; j < MAX_MB_PLANE; ++j) {
+      struct macroblockd_plane *const pd = &xd->plane[j];
+      setup_pred_plane(&pd->dst, AOMMAX(mbmi->sb_type, BLOCK_8X8), tmp_buf[j],
+                       tmp_width[j], tmp_height[j], tmp_stride[j],
+                       (xd->n8_h >> 1), i, NULL, pd->subsampling_x,
+                       pd->subsampling_y);
+    }
+    for (ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) {
+      const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
+      const RefBuffer *const ref_buf = &cm->frame_refs[frame - LAST_FRAME];
+
+      xd->block_refs[ref] = ref_buf;
+      if ((!av1_is_valid_scale(&ref_buf->sf)))
+        aom_internal_error(xd->error_info, AOM_CODEC_UNSUP_BITSTREAM,
+                           "Reference frame has invalid dimensions");
+      av1_setup_pre_planes(xd, ref, ref_buf->buf, mi_row + (xd->n8_h >> 1),
+                           mi_col + i, &ref_buf->sf);
+    }
+
+    xd->mb_to_left_edge = -(((mi_col + i) * MI_SIZE) * 8);
+    xd->mb_to_right_edge =
+        mb_to_right_edge_base + (xd->n8_w - i - mi_step) * 64;
+    mi_x = (mi_col + i) << MI_SIZE_LOG2;
+    mi_y = (mi_row << MI_SIZE_LOG2) + xd->n8_h * 4;
+
+    for (j = 0; j < MAX_MB_PLANE; ++j) {
+      const struct macroblockd_plane *pd = &xd->plane[j];
+      bw = (mi_step << MI_SIZE_LOG2) >> pd->subsampling_x;
+      bh = (num_4x4_blocks_high_lookup[bsize] << 1) >> pd->subsampling_y;
+
+      if (mbmi->sb_type < BLOCK_8X8 && !CONFIG_CB4X4) {
+        const PARTITION_TYPE bp = BLOCK_8X8 - mbmi->sb_type;
+        const int have_vsplit = bp != PARTITION_HORZ;
+        const int have_hsplit = bp != PARTITION_VERT;
+        const int num_4x4_w = 2 >> (!have_vsplit);
+        const int num_4x4_h = 2 >> (!have_hsplit);
+        const int pw = 8 >> (have_vsplit + pd->subsampling_x);
+        int x, y;
+
+        for (y = 0; y < num_4x4_h; ++y)
+          for (x = 0; x < num_4x4_w; ++x) {
+            if ((bp == PARTITION_HORZ || bp == PARTITION_SPLIT) && y != 0)
+              continue;
+
+            build_inter_predictors(
+                xd, j, mi_col_offset, mi_row_offset, y * 2 + x, bw, bh,
+                (4 * x) >> pd->subsampling_x,
+                xd->n8_h == 1 ? (4 >> pd->subsampling_y) : 0, pw, bh,
+#if CONFIG_SUPERTX && CONFIG_EXT_INTER
+                0, 0,
+#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
+                mi_x, mi_y);
+          }
+      } else {
+        build_inter_predictors(xd, j, mi_col_offset, mi_row_offset, 0, bw, bh,
+                               0, xd->n8_h == 1 ? (4 >> pd->subsampling_y) : 0,
+                               bw, bh,
+#if CONFIG_SUPERTX && CONFIG_EXT_INTER
+                               0, 0,
+#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
+                               mi_x, mi_y);
+      }
+    }
+#if CONFIG_EXT_INTER
+    *mbmi = backup_mbmi;
+#endif  // CONFIG_EXT_INTER
+  }
+  xd->mb_to_left_edge = -((mi_col * MI_SIZE) * 8);
+  xd->mb_to_right_edge = mb_to_right_edge_base;
+  xd->mb_to_top_edge += xd->n8_h * 32;
+}
+
+void av1_build_prediction_by_right_preds(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                         int mi_row, int mi_col,
+                                         uint8_t *tmp_buf[MAX_MB_PLANE],
+                                         int tmp_width[MAX_MB_PLANE],
+                                         int tmp_height[MAX_MB_PLANE],
+                                         const int tmp_stride[MAX_MB_PLANE]) {
+  const TileInfo *const tile = &xd->tile;
+  BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  int i, j, mi_step, ref;
+  const int ilimit = AOMMIN(xd->n8_h, cm->mi_rows - mi_row);
+  int mb_to_bottom_edge_base = xd->mb_to_bottom_edge;
+
+  if (mi_col + xd->n8_w >= tile->mi_col_end ||
+      (mi_col + xd->n8_w) % MI_SIZE == 0 || (mi_col + xd->n8_w) >= cm->mi_cols)
+    return;
+
+  xd->mb_to_left_edge -= xd->n8_w * 32;
+  for (i = 0; i < ilimit; i += mi_step) {
+    int mi_row_offset = i;
+    int mi_col_offset = xd->n8_w;
+    int mi_x, mi_y, bw, bh;
+    MODE_INFO *mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+    MB_MODE_INFO *mbmi = &mi->mbmi;
+#if CONFIG_EXT_INTER
+    MB_MODE_INFO backup_mbmi;
+#endif  // CONFIG_EXT_INTER
+
+    mi_step = AOMMIN(xd->n8_h, mi_size_high[mbmi->sb_type]);
+
+    if (!is_neighbor_overlappable(mbmi)) continue;
+
+#if CONFIG_EXT_INTER
+    backup_mbmi = *mbmi;
+    modify_neighbor_predictor_for_obmc(mbmi);
+#endif  // CONFIG_EXT_INTER
+
+    for (j = 0; j < MAX_MB_PLANE; ++j) {
+      struct macroblockd_plane *const pd = &xd->plane[j];
+      setup_pred_plane(&pd->dst, AOMMAX(mbmi->sb_type, BLOCK_8X8), tmp_buf[j],
+                       tmp_width[j], tmp_height[j], tmp_stride[j], i,
+                       xd->n8_w >> 1, NULL, pd->subsampling_x,
+                       pd->subsampling_y);
+    }
+    for (ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) {
+      const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
+      const RefBuffer *const ref_buf = &cm->frame_refs[frame - LAST_FRAME];
+
+      xd->block_refs[ref] = ref_buf;
+      if ((!av1_is_valid_scale(&ref_buf->sf)))
+        aom_internal_error(xd->error_info, AOM_CODEC_UNSUP_BITSTREAM,
+                           "Reference frame has invalid dimensions");
+      av1_setup_pre_planes(xd, ref, ref_buf->buf, mi_row + i,
+                           mi_col + (xd->n8_w >> 1), &ref_buf->sf);
+    }
+
+    xd->mb_to_top_edge = -(((mi_row + i) * MI_SIZE) * 8);
+    xd->mb_to_bottom_edge =
+        mb_to_bottom_edge_base + (xd->n8_h - i - mi_step) * 64;
+    mi_x = (mi_col << MI_SIZE_LOG2) + xd->n8_w * 4;
+    mi_y = (mi_row + i) << MI_SIZE_LOG2;
+
+    for (j = 0; j < MAX_MB_PLANE; ++j) {
+      const struct macroblockd_plane *pd = &xd->plane[j];
+      bw = (num_4x4_blocks_wide_lookup[bsize] << 1) >> pd->subsampling_x;
+      bh = (mi_step << MI_SIZE_LOG2) >> pd->subsampling_y;
+
+      if (mbmi->sb_type < BLOCK_8X8 && !CONFIG_CB4X4) {
+        const PARTITION_TYPE bp = BLOCK_8X8 - mbmi->sb_type;
+        const int have_vsplit = bp != PARTITION_HORZ;
+        const int have_hsplit = bp != PARTITION_VERT;
+        const int num_4x4_w = 2 >> (!have_vsplit);
+        const int num_4x4_h = 2 >> (!have_hsplit);
+        const int ph = 8 >> (have_hsplit + pd->subsampling_y);
+        int x, y;
+
+        for (y = 0; y < num_4x4_h; ++y)
+          for (x = 0; x < num_4x4_w; ++x) {
+            if ((bp == PARTITION_VERT || bp == PARTITION_SPLIT) && x != 0)
+              continue;
+
+            build_inter_predictors(xd, j, mi_col_offset, mi_row_offset,
+                                   y * 2 + x, bw, bh,
+                                   xd->n8_w == 1 ? 4 >> pd->subsampling_x : 0,
+                                   (4 * y) >> pd->subsampling_y, bw, ph,
+#if CONFIG_SUPERTX && CONFIG_EXT_INTER
+                                   0, 0,
+#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
+                                   mi_x, mi_y);
+          }
+      } else {
+        build_inter_predictors(xd, j, mi_col_offset, mi_row_offset, 0, bw, bh,
+                               xd->n8_w == 1 ? 4 >> pd->subsampling_x : 0, 0,
+                               bw, bh,
+#if CONFIG_SUPERTX && CONFIG_EXT_INTER
+                               0, 0,
+#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
+                               mi_x, mi_y);
+      }
+    }
+#if CONFIG_EXT_INTER
+    *mbmi = backup_mbmi;
+#endif  // CONFIG_EXT_INTER
+  }
+  xd->mb_to_top_edge = -((mi_row * MI_SIZE) * 8);
+  xd->mb_to_bottom_edge = mb_to_bottom_edge_base;
+  xd->mb_to_left_edge += xd->n8_w * 32;
+}
+
+// This function combines motion compensated predictions that is generated by
+// bottom/right neighboring blocks' inter predictors with prediction in dst
+// buffer.
+void av1_merge_dst_bottom_right_preds(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                      int mi_row, int mi_col,
+                                      uint8_t *bottom[MAX_MB_PLANE],
+                                      const int bottom_stride[MAX_MB_PLANE],
+                                      uint8_t *right[MAX_MB_PLANE],
+                                      const int right_stride[MAX_MB_PLANE]) {
+  const TileInfo *const tile = &xd->tile;
+  BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  int plane, i, mi_step;
+  const int bottom_available = mi_row + xd->n8_h < tile->mi_row_end &&
+                               (mi_row + xd->n8_h) % MI_SIZE != 0 &&
+                               (mi_row + xd->n8_h) < cm->mi_rows;
+#if CONFIG_HIGHBITDEPTH
+  int is_hbd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0;
+#endif  // CONFIG_HIGHBITDEPTH
+
+  // handle bottom row
+  for (i = 0; bottom_available && i < AOMMIN(xd->n8_w, cm->mi_cols - mi_col);
+       i += mi_step) {
+    int mi_row_offset = xd->n8_h;
+    int mi_col_offset = i;
+    MODE_INFO *mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+    MB_MODE_INFO *mbmi = &mi->mbmi;
+    int overlap;
+
+    mi_step = AOMMIN(xd->n8_w, mi_size_wide[mbmi->sb_type]);
+
+    if (!is_neighbor_overlappable(mbmi)) continue;
+
+    overlap = num_4x4_blocks_high_lookup[bsize] << 1;
+
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+      const struct macroblockd_plane *pd = &xd->plane[plane];
+      const int bw = (mi_step * MI_SIZE) >> pd->subsampling_x;
+      const int bh = overlap >> pd->subsampling_y;
+      const int dst_stride = pd->dst.stride;
+      uint8_t *dst =
+          &pd->dst.buf[((i * MI_SIZE) >> pd->subsampling_x) +
+                       (((xd->n8_h * MI_SIZE - overlap) * dst_stride) >>
+                        pd->subsampling_y)];
+      const int tmp_stride = bottom_stride[plane];
+      const uint8_t *const tmp =
+          &bottom[plane][((i * MI_SIZE) >> pd->subsampling_x) +
+                         (((xd->n8_h * MI_SIZE - overlap) * tmp_stride) >>
+                          pd->subsampling_y)];
+      const uint8_t *const mask = av1_get_obmc_mask_flipped(bh);
+
+#if CONFIG_HIGHBITDEPTH
+      if (is_hbd)
+        aom_highbd_blend_a64_vmask(dst, dst_stride, dst, dst_stride, tmp,
+                                   tmp_stride, mask, bh, bw, xd->bd);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        aom_blend_a64_vmask(dst, dst_stride, dst, dst_stride, tmp, tmp_stride,
+                            mask, bh, bw);
+    }
+  }  // each mi in the bottom row
+
+  // handle right column
+  if (mi_col + xd->n8_w >= tile->mi_col_end ||
+      (mi_col + xd->n8_w) % MI_SIZE == 0 || (mi_col + xd->n8_w) >= cm->mi_cols)
+    return;
+
+  for (i = 0; i < AOMMIN(xd->n8_h, cm->mi_rows - mi_row); i += mi_step) {
+    int mi_row_offset = i;
+    int mi_col_offset = xd->n8_w;
+    int overlap;
+    MODE_INFO *mi = xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride];
+    MB_MODE_INFO *mbmi = &mi->mbmi;
+
+    mi_step = AOMMIN(xd->n8_h, mi_size_high[mbmi->sb_type]);
+
+    if (!is_neighbor_overlappable(mbmi)) continue;
+
+    overlap = num_4x4_blocks_wide_lookup[bsize] << 1;
+
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+      const struct macroblockd_plane *pd = &xd->plane[plane];
+      const int bw = overlap >> pd->subsampling_x;
+      const int bh = (mi_step * MI_SIZE) >> pd->subsampling_y;
+      const int dst_stride = pd->dst.stride;
+      uint8_t *dst =
+          &pd->dst.buf[((i * MI_SIZE * dst_stride) >> pd->subsampling_y) +
+                       ((xd->n8_w * MI_SIZE - overlap) >> pd->subsampling_x)];
+      const int tmp_stride = right_stride[plane];
+      const uint8_t *const tmp =
+          &right[plane][((i * MI_SIZE * tmp_stride) >> pd->subsampling_y) +
+                        ((xd->n8_w * MI_SIZE - overlap) >> pd->subsampling_x)];
+      const uint8_t *const mask = av1_get_obmc_mask_flipped(bw);
+
+#if CONFIG_HIGHBITDEPTH
+      if (is_hbd)
+        aom_highbd_blend_a64_hmask(dst, dst_stride, dst, dst_stride, tmp,
+                                   tmp_stride, mask, bh, bw, xd->bd);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        aom_blend_a64_hmask(dst, dst_stride, dst, dst_stride, tmp, tmp_stride,
+                            mask, bh, bw);
+    }
+  }  // each mi in the right column
+}
+
+// This function generates 4 sided obmc. (1) Prediction blocks generated by
+// bottom and right motion vectors are calculated. (2) Combine them with the
+// original prediction block (which should be pre-stored in xd->plane[].dst.buf
+// before calling this function). The results is updated in xd->plane[].dst.buf
+// (3) Call causal obmc prediction function, which will generate left and above
+// preds, and then merge them and xd->plane[].dst.buf.
+void av1_build_ncobmc_inter_predictors_sb(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                          int mi_row, int mi_col) {
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf1[2 * MAX_MB_PLANE * MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf2[2 * MAX_MB_PLANE * MAX_SB_SQUARE]);
+#else
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf1[MAX_MB_PLANE * MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf2[MAX_MB_PLANE * MAX_SB_SQUARE]);
+#endif  // CONFIG_HIGHBITDEPTH
+  uint8_t *dst_buf1[MAX_MB_PLANE], *dst_buf2[MAX_MB_PLANE];
+  int dst_stride1[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_stride2[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_width1[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_width2[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_height1[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_height2[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    int len = sizeof(uint16_t);
+    dst_buf1[0] = CONVERT_TO_BYTEPTR(tmp_buf1);
+    dst_buf1[1] = CONVERT_TO_BYTEPTR(tmp_buf1 + MAX_SB_SQUARE * len);
+    dst_buf1[2] = CONVERT_TO_BYTEPTR(tmp_buf1 + MAX_SB_SQUARE * 2 * len);
+    dst_buf2[0] = CONVERT_TO_BYTEPTR(tmp_buf2);
+    dst_buf2[1] = CONVERT_TO_BYTEPTR(tmp_buf2 + MAX_SB_SQUARE * len);
+    dst_buf2[2] = CONVERT_TO_BYTEPTR(tmp_buf2 + MAX_SB_SQUARE * 2 * len);
+  } else {
+#endif  // CONFIG_HIGHBITDEPTH
+    dst_buf1[0] = tmp_buf1;
+    dst_buf1[1] = tmp_buf1 + MAX_SB_SQUARE;
+    dst_buf1[2] = tmp_buf1 + MAX_SB_SQUARE * 2;
+    dst_buf2[0] = tmp_buf2;
+    dst_buf2[1] = tmp_buf2 + MAX_SB_SQUARE;
+    dst_buf2[2] = tmp_buf2 + MAX_SB_SQUARE * 2;
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  av1_build_prediction_by_bottom_preds(cm, xd, mi_row, mi_col, dst_buf1,
+                                       dst_width1, dst_height1, dst_stride1);
+  av1_build_prediction_by_right_preds(cm, xd, mi_row, mi_col, dst_buf2,
+                                      dst_width2, dst_height2, dst_stride2);
+  av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                       mi_col);
+  av1_merge_dst_bottom_right_preds(cm, xd, mi_row, mi_col, dst_buf1,
+                                   dst_stride1, dst_buf2, dst_stride2);
+  av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                       mi_col);
+  av1_build_obmc_inter_predictors_sb(cm, xd, mi_row, mi_col);
+  av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                       mi_col);
+}
+#endif  // CONFIG_NCOBMC
+#endif  // CONFIG_MOTION_VAR
+
+#if CONFIG_EXT_INTER
+/* clang-format off */
+#if CONFIG_EXT_PARTITION
+static const int ii_weights1d[MAX_SB_SIZE] = {
+  26, 25, 24, 24, 23, 23, 22, 22, 21, 21, 20, 20, 19, 19, 18, 18, 17, 17, 17,
+  16, 16, 16, 15, 15, 15, 15, 14, 14, 14, 13, 13, 13, 13, 13, 12, 12, 12, 12,
+  12, 11, 11, 11, 11, 11, 11, 10, 10, 10, 10, 10, 10, 10, 10, 10,  9,  9,  9,
+  9,  9,  9,  9,  9,  9,  9,  9,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  7,  7,  7,  7,  7,  7,  7,  7,  7,
+  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,
+  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7,  7
+};
+static int ii_size_scales[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+    32, 32, 32,
+#endif
+    32, 16, 16, 16, 8, 8, 8, 4,
+    4,  4,  2,  2,  2, 1, 1, 1,
+};
+#else
+static const int ii_weights1d[MAX_SB_SIZE] = {
+  26, 25, 24, 24, 23, 23, 22, 22, 21, 21, 20, 20, 19, 19, 18, 18,
+  17, 17, 17, 16, 16, 16, 15, 15, 15, 15, 14, 14, 14, 13, 13, 13,
+  13, 13, 12, 12, 12, 12, 12, 11, 11, 11, 11, 11, 11, 10, 10, 10,
+  10, 10, 10, 10, 10, 10,  9,  9,  9,  9,  9,  9,  9,  9,  9,  9
+};
+static int ii_size_scales[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+    16, 16, 16,
+#endif
+    16, 8, 8, 8, 4, 4, 4,
+    2,  2, 2, 1, 1, 1,
+};
+/* clang-format on */
+#endif  // CONFIG_EXT_PARTITION
+
+static void combine_interintra(INTERINTRA_MODE mode, int use_wedge_interintra,
+                               int wedge_index, int wedge_sign,
+                               BLOCK_SIZE bsize, BLOCK_SIZE plane_bsize,
+                               uint8_t *comppred, int compstride,
+                               const uint8_t *interpred, int interstride,
+                               const uint8_t *intrapred, int intrastride) {
+  const int bw = block_size_wide[plane_bsize];
+  const int bh = block_size_high[plane_bsize];
+  const int size_scale = ii_size_scales[plane_bsize];
+  int i, j;
+
+  if (use_wedge_interintra) {
+    if (is_interintra_wedge_used(bsize)) {
+      const uint8_t *mask =
+          av1_get_contiguous_soft_mask(wedge_index, wedge_sign, bsize);
+      const int subw = 2 * num_4x4_blocks_wide_lookup[bsize] == bw;
+      const int subh = 2 * num_4x4_blocks_high_lookup[bsize] == bh;
+      aom_blend_a64_mask(comppred, compstride, intrapred, intrastride,
+                         interpred, interstride, mask, block_size_wide[bsize],
+                         bh, bw, subh, subw);
+    }
+    return;
+  }
+
+  switch (mode) {
+    case II_V_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale = ii_weights1d[i * size_scale];
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_H_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale = ii_weights1d[j * size_scale];
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_D63_PRED:
+    case II_D117_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale = (ii_weights1d[i * size_scale] * 3 +
+                       ii_weights1d[j * size_scale]) >>
+                      2;
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_D207_PRED:
+    case II_D153_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale = (ii_weights1d[j * size_scale] * 3 +
+                       ii_weights1d[i * size_scale]) >>
+                      2;
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_D135_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale = ii_weights1d[(i < j ? i : j) * size_scale];
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_D45_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale =
+              (ii_weights1d[i * size_scale] + ii_weights1d[j * size_scale]) >>
+              1;
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_TM_PRED:
+    case II_DC_PRED:
+    default:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          comppred[i * compstride + j] = AOM_BLEND_AVG(
+              intrapred[i * intrastride + j], interpred[i * interstride + j]);
+        }
+      }
+      break;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void combine_interintra_highbd(
+    INTERINTRA_MODE mode, int use_wedge_interintra, int wedge_index,
+    int wedge_sign, BLOCK_SIZE bsize, BLOCK_SIZE plane_bsize,
+    uint8_t *comppred8, int compstride, const uint8_t *interpred8,
+    int interstride, const uint8_t *intrapred8, int intrastride, int bd) {
+  const int bw = block_size_wide[plane_bsize];
+  const int bh = block_size_high[plane_bsize];
+  const int size_scale = ii_size_scales[plane_bsize];
+  int i, j;
+
+  uint16_t *comppred = CONVERT_TO_SHORTPTR(comppred8);
+  const uint16_t *interpred = CONVERT_TO_SHORTPTR(interpred8);
+  const uint16_t *intrapred = CONVERT_TO_SHORTPTR(intrapred8);
+
+  if (use_wedge_interintra) {
+    if (is_interintra_wedge_used(bsize)) {
+      const uint8_t *mask =
+          av1_get_contiguous_soft_mask(wedge_index, wedge_sign, bsize);
+      const int subh = 2 * num_4x4_blocks_high_lookup[bsize] == bh;
+      const int subw = 2 * num_4x4_blocks_wide_lookup[bsize] == bw;
+      aom_highbd_blend_a64_mask(comppred8, compstride, intrapred8, intrastride,
+                                interpred8, interstride, mask, bw, bh, bw, subh,
+                                subw, bd);
+    }
+    return;
+  }
+
+  switch (mode) {
+    case II_V_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale = ii_weights1d[i * size_scale];
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_H_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale = ii_weights1d[j * size_scale];
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_D63_PRED:
+    case II_D117_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale = (ii_weights1d[i * size_scale] * 3 +
+                       ii_weights1d[j * size_scale]) >>
+                      2;
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_D207_PRED:
+    case II_D153_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale = (ii_weights1d[j * size_scale] * 3 +
+                       ii_weights1d[i * size_scale]) >>
+                      2;
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_D135_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale = ii_weights1d[(i < j ? i : j) * size_scale];
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_D45_PRED:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          int scale =
+              (ii_weights1d[i * size_scale] + ii_weights1d[j * size_scale]) >>
+              1;
+          comppred[i * compstride + j] =
+              AOM_BLEND_A64(scale, intrapred[i * intrastride + j],
+                            interpred[i * interstride + j]);
+        }
+      }
+      break;
+
+    case II_TM_PRED:
+    case II_DC_PRED:
+    default:
+      for (i = 0; i < bh; ++i) {
+        for (j = 0; j < bw; ++j) {
+          comppred[i * compstride + j] = AOM_BLEND_AVG(
+              interpred[i * interstride + j], intrapred[i * intrastride + j]);
+        }
+      }
+      break;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_build_intra_predictors_for_interintra(MACROBLOCKD *xd,
+                                               BLOCK_SIZE bsize, int plane,
+                                               BUFFER_SET *ctx, uint8_t *dst,
+                                               int dst_stride) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, &xd->plane[plane]);
+  PREDICTION_MODE mode =
+      interintra_to_intra_mode[xd->mi[0]->mbmi.interintra_mode];
+
+  av1_predict_intra_block(xd, pd->width, pd->height, plane_bsize, mode,
+                          ctx->plane[plane], ctx->stride[plane], dst,
+                          dst_stride, 0, 0, plane);
+}
+
+void av1_combine_interintra(MACROBLOCKD *xd, BLOCK_SIZE bsize, int plane,
+                            const uint8_t *inter_pred, int inter_stride,
+                            const uint8_t *intra_pred, int intra_stride) {
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, &xd->plane[plane]);
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    combine_interintra_highbd(
+        xd->mi[0]->mbmi.interintra_mode, xd->mi[0]->mbmi.use_wedge_interintra,
+        xd->mi[0]->mbmi.interintra_wedge_index,
+        xd->mi[0]->mbmi.interintra_wedge_sign, bsize, plane_bsize,
+        xd->plane[plane].dst.buf, xd->plane[plane].dst.stride, inter_pred,
+        inter_stride, intra_pred, intra_stride, xd->bd);
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  combine_interintra(xd->mi[0]->mbmi.interintra_mode,
+                     xd->mi[0]->mbmi.use_wedge_interintra,
+                     xd->mi[0]->mbmi.interintra_wedge_index,
+                     xd->mi[0]->mbmi.interintra_wedge_sign, bsize, plane_bsize,
+                     xd->plane[plane].dst.buf, xd->plane[plane].dst.stride,
+                     inter_pred, inter_stride, intra_pred, intra_stride);
+}
+
+void av1_build_interintra_predictors_sby(MACROBLOCKD *xd, uint8_t *ypred,
+                                         int ystride, BUFFER_SET *ctx,
+                                         BLOCK_SIZE bsize) {
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    DECLARE_ALIGNED(16, uint16_t, intrapredictor[MAX_SB_SQUARE]);
+    av1_build_intra_predictors_for_interintra(
+        xd, bsize, 0, ctx, CONVERT_TO_BYTEPTR(intrapredictor), MAX_SB_SIZE);
+    av1_combine_interintra(xd, bsize, 0, ypred, ystride,
+                           CONVERT_TO_BYTEPTR(intrapredictor), MAX_SB_SIZE);
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  {
+    DECLARE_ALIGNED(16, uint8_t, intrapredictor[MAX_SB_SQUARE]);
+    av1_build_intra_predictors_for_interintra(xd, bsize, 0, ctx, intrapredictor,
+                                              MAX_SB_SIZE);
+    av1_combine_interintra(xd, bsize, 0, ypred, ystride, intrapredictor,
+                           MAX_SB_SIZE);
+  }
+}
+
+void av1_build_interintra_predictors_sbc(MACROBLOCKD *xd, uint8_t *upred,
+                                         int ustride, BUFFER_SET *ctx,
+                                         int plane, BLOCK_SIZE bsize) {
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    DECLARE_ALIGNED(16, uint16_t, uintrapredictor[MAX_SB_SQUARE]);
+    av1_build_intra_predictors_for_interintra(
+        xd, bsize, plane, ctx, CONVERT_TO_BYTEPTR(uintrapredictor),
+        MAX_SB_SIZE);
+    av1_combine_interintra(xd, bsize, plane, upred, ustride,
+                           CONVERT_TO_BYTEPTR(uintrapredictor), MAX_SB_SIZE);
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  {
+    DECLARE_ALIGNED(16, uint8_t, uintrapredictor[MAX_SB_SQUARE]);
+    av1_build_intra_predictors_for_interintra(xd, bsize, plane, ctx,
+                                              uintrapredictor, MAX_SB_SIZE);
+    av1_combine_interintra(xd, bsize, plane, upred, ustride, uintrapredictor,
+                           MAX_SB_SIZE);
+  }
+}
+
+void av1_build_interintra_predictors_sbuv(MACROBLOCKD *xd, uint8_t *upred,
+                                          uint8_t *vpred, int ustride,
+                                          int vstride, BUFFER_SET *ctx,
+                                          BLOCK_SIZE bsize) {
+  av1_build_interintra_predictors_sbc(xd, upred, ustride, ctx, 1, bsize);
+  av1_build_interintra_predictors_sbc(xd, vpred, vstride, ctx, 2, bsize);
+}
+
+void av1_build_interintra_predictors(MACROBLOCKD *xd, uint8_t *ypred,
+                                     uint8_t *upred, uint8_t *vpred,
+                                     int ystride, int ustride, int vstride,
+                                     BUFFER_SET *ctx, BLOCK_SIZE bsize) {
+  av1_build_interintra_predictors_sby(xd, ypred, ystride, ctx, bsize);
+  av1_build_interintra_predictors_sbuv(xd, upred, vpred, ustride, vstride, ctx,
+                                       bsize);
+}
+
+// Builds the inter-predictor for the single ref case
+// for use in the encoder to search the wedges efficiently.
+static void build_inter_predictors_single_buf(MACROBLOCKD *xd, int plane,
+                                              int block, int bw, int bh, int x,
+                                              int y, int w, int h, int mi_x,
+                                              int mi_y, int ref,
+                                              uint8_t *const ext_dst,
+                                              int ext_dst_stride) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const MODE_INFO *mi = xd->mi[0];
+
+  const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
+  struct buf_2d *const pre_buf = &pd->pre[ref];
+#if CONFIG_HIGHBITDEPTH
+  uint8_t *const dst =
+      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH ? CONVERT_TO_BYTEPTR(ext_dst)
+                                                   : ext_dst) +
+      ext_dst_stride * y + x;
+#else
+  uint8_t *const dst = ext_dst + ext_dst_stride * y + x;
+#endif
+  const MV mv = mi->mbmi.sb_type < BLOCK_8X8
+                    ? average_split_mvs(pd, mi, ref, block)
+                    : mi->mbmi.mv[ref].as_mv;
+
+  // TODO(jkoleszar): This clamping is done in the incorrect place for the
+  // scaling case. It needs to be done on the scaled MV, not the pre-scaling
+  // MV. Note however that it performs the subsampling aware scaling so
+  // that the result is always q4.
+  // mv_precision precision is MV_PRECISION_Q4.
+  const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, &mv, bw, bh, pd->subsampling_x,
+                                             pd->subsampling_y);
+
+  uint8_t *pre;
+  MV32 scaled_mv;
+  int xs, ys, subpel_x, subpel_y;
+  const int is_scaled = av1_is_scaled(sf);
+  ConvolveParams conv_params = get_conv_params(0, plane);
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+  WarpTypesAllowed warp_types;
+#if CONFIG_GLOBAL_MOTION
+  WarpedMotionParams *const wm = &xd->global_motion[mi->mbmi.ref_frame[ref]];
+  warp_types.global_warp_allowed = is_global_mv_block(mi, block, wm->wmtype);
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_WARPED_MOTION
+  warp_types.local_warp_allowed = mi->mbmi.motion_mode == WARPED_CAUSAL;
+#endif  // CONFIG_WARPED_MOTION
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+
+  if (is_scaled) {
+    pre = pre_buf->buf + scaled_buffer_offset(x, y, pre_buf->stride, sf);
+    scaled_mv = av1_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
+    xs = sf->x_step_q4;
+    ys = sf->y_step_q4;
+  } else {
+    pre = pre_buf->buf + (y * pre_buf->stride + x);
+    scaled_mv.row = mv_q4.row;
+    scaled_mv.col = mv_q4.col;
+    xs = ys = 16;
+  }
+
+  subpel_x = scaled_mv.col & SUBPEL_MASK;
+  subpel_y = scaled_mv.row & SUBPEL_MASK;
+  pre += (scaled_mv.row >> SUBPEL_BITS) * pre_buf->stride +
+         (scaled_mv.col >> SUBPEL_BITS);
+
+  av1_make_inter_predictor(pre, pre_buf->stride, dst, ext_dst_stride, subpel_x,
+                           subpel_y, sf, w, h, &conv_params,
+                           mi->mbmi.interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                           &warp_types, (mi_x >> pd->subsampling_x) + x,
+                           (mi_y >> pd->subsampling_y) + y, plane, ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+                           0, 0,
+#endif
+                           xs, ys, xd);
+}
+
+void av1_build_inter_predictors_for_planes_single_buf(
+    MACROBLOCKD *xd, BLOCK_SIZE bsize, int plane_from, int plane_to, int mi_row,
+    int mi_col, int ref, uint8_t *ext_dst[3], int ext_dst_stride[3]) {
+  int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
+  for (plane = plane_from; plane <= plane_to; ++plane) {
+    const BLOCK_SIZE plane_bsize =
+        get_plane_block_size(bsize, &xd->plane[plane]);
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+    const int bw = block_size_wide[plane_bsize];
+    const int bh = block_size_high[plane_bsize];
+
+    if (xd->mi[0]->mbmi.sb_type < BLOCK_8X8 && !CONFIG_CB4X4) {
+      int x, y;
+      assert(bsize == BLOCK_8X8);
+      for (y = 0; y < num_4x4_h; ++y)
+        for (x = 0; x < num_4x4_w; ++x)
+          build_inter_predictors_single_buf(
+              xd, plane, y * 2 + x, bw, bh, 4 * x, 4 * y, 4, 4, mi_x, mi_y, ref,
+              ext_dst[plane], ext_dst_stride[plane]);
+    } else {
+      build_inter_predictors_single_buf(xd, plane, 0, bw, bh, 0, 0, bw, bh,
+                                        mi_x, mi_y, ref, ext_dst[plane],
+                                        ext_dst_stride[plane]);
+    }
+  }
+}
+
+static void build_wedge_inter_predictor_from_buf(
+    MACROBLOCKD *xd, int plane, int x, int y, int w, int h,
+#if CONFIG_SUPERTX
+    int wedge_offset_x, int wedge_offset_y,
+#endif  // CONFIG_SUPERTX
+    uint8_t *ext_dst0, int ext_dst_stride0, uint8_t *ext_dst1,
+    int ext_dst_stride1) {
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int is_compound = has_second_ref(mbmi);
+  MACROBLOCKD_PLANE *const pd = &xd->plane[plane];
+  struct buf_2d *const dst_buf = &pd->dst;
+  uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
+  const INTERINTER_COMPOUND_DATA comp_data = {
+#if CONFIG_WEDGE
+    mbmi->wedge_index,
+    mbmi->wedge_sign,
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+    mbmi->mask_type,
+    xd->seg_mask,
+#endif  // CONFIG_COMPOUND_SEGMENT
+    mbmi->interinter_compound_type
+  };
+
+  if (is_compound && is_masked_compound_type(mbmi->interinter_compound_type)) {
+#if CONFIG_COMPOUND_SEGMENT
+    if (!plane && comp_data.interinter_compound_type == COMPOUND_SEG) {
+#if CONFIG_HIGHBITDEPTH
+      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+        build_compound_seg_mask_highbd(
+            comp_data.seg_mask, comp_data.mask_type,
+            CONVERT_TO_BYTEPTR(ext_dst0), ext_dst_stride0,
+            CONVERT_TO_BYTEPTR(ext_dst1), ext_dst_stride1, mbmi->sb_type, h, w,
+            xd->bd);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        build_compound_seg_mask(comp_data.seg_mask, comp_data.mask_type,
+                                ext_dst0, ext_dst_stride0, ext_dst1,
+                                ext_dst_stride1, mbmi->sb_type, h, w);
+    }
+#endif  // CONFIG_COMPOUND_SEGMENT
+
+#if CONFIG_SUPERTX
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+      build_masked_compound_wedge_extend_highbd(
+          dst, dst_buf->stride, CONVERT_TO_BYTEPTR(ext_dst0), ext_dst_stride0,
+          CONVERT_TO_BYTEPTR(ext_dst1), ext_dst_stride1, &comp_data,
+          mbmi->sb_type, wedge_offset_x, wedge_offset_y, h, w, xd->bd);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      build_masked_compound_wedge_extend(
+          dst, dst_buf->stride, ext_dst0, ext_dst_stride0, ext_dst1,
+          ext_dst_stride1, &comp_data, mbmi->sb_type, wedge_offset_x,
+          wedge_offset_y, h, w);
+#else
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+      build_masked_compound_highbd(
+          dst, dst_buf->stride, CONVERT_TO_BYTEPTR(ext_dst0), ext_dst_stride0,
+          CONVERT_TO_BYTEPTR(ext_dst1), ext_dst_stride1, &comp_data,
+          mbmi->sb_type, h, w, xd->bd);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      build_masked_compound(dst, dst_buf->stride, ext_dst0, ext_dst_stride0,
+                            ext_dst1, ext_dst_stride1, &comp_data,
+                            mbmi->sb_type, h, w);
+#endif  // CONFIG_SUPERTX
+  } else {
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+      aom_highbd_convolve_copy(CONVERT_TO_BYTEPTR(ext_dst0), ext_dst_stride0,
+                               dst, dst_buf->stride, NULL, 0, NULL, 0, w, h,
+                               xd->bd);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      aom_convolve_copy(ext_dst0, ext_dst_stride0, dst, dst_buf->stride, NULL,
+                        0, NULL, 0, w, h);
+  }
+}
+
+void av1_build_wedge_inter_predictor_from_buf(
+    MACROBLOCKD *xd, BLOCK_SIZE bsize, int plane_from, int plane_to,
+#if CONFIG_SUPERTX
+    int wedge_offset_x, int wedge_offset_y,
+#endif  // CONFIG_SUPERTX
+    uint8_t *ext_dst0[3], int ext_dst_stride0[3], uint8_t *ext_dst1[3],
+    int ext_dst_stride1[3]) {
+  int plane;
+  for (plane = plane_from; plane <= plane_to; ++plane) {
+    const BLOCK_SIZE plane_bsize =
+        get_plane_block_size(bsize, &xd->plane[plane]);
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+
+    if (xd->mi[0]->mbmi.sb_type < BLOCK_8X8 && !CONFIG_CB4X4) {
+      int x, y;
+      assert(bsize == BLOCK_8X8);
+      for (y = 0; y < num_4x4_h; ++y)
+        for (x = 0; x < num_4x4_w; ++x)
+          build_wedge_inter_predictor_from_buf(
+              xd, plane, 4 * x, 4 * y, 4, 4,
+#if CONFIG_SUPERTX
+              wedge_offset_x, wedge_offset_y,
+#endif  // CONFIG_SUPERTX
+              ext_dst0[plane], ext_dst_stride0[plane], ext_dst1[plane],
+              ext_dst_stride1[plane]);
+    } else {
+      const int bw = block_size_wide[plane_bsize];
+      const int bh = block_size_high[plane_bsize];
+      build_wedge_inter_predictor_from_buf(
+          xd, plane, 0, 0, bw, bh,
+#if CONFIG_SUPERTX
+          wedge_offset_x, wedge_offset_y,
+#endif  // CONFIG_SUPERTX
+          ext_dst0[plane], ext_dst_stride0[plane], ext_dst1[plane],
+          ext_dst_stride1[plane]);
+    }
+  }
+}
+#endif  // CONFIG_EXT_INTER
diff --git a/third_party/aom/av1/common/reconinter.h b/third_party/aom/av1/common/reconinter.h
new file mode 100644
index 0000000000..10933a751c
--- /dev/null
+++ b/third_party/aom/av1/common/reconinter.h
@@ -0,0 +1,828 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_RECONINTER_H_
+#define AV1_COMMON_RECONINTER_H_
+
+#include "av1/common/filter.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/convolve.h"
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#include "av1/common/warped_motion.h"
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#include "aom/aom_integer.h"
+
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+#define WARP_WM_NEIGHBORS_WITH_OBMC 0
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+
+#if CONFIG_MOTION_VAR && CONFIG_GLOBAL_MOTION
+#define WARP_GM_NEIGHBORS_WITH_OBMC 0
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE void inter_predictor(const uint8_t *src, int src_stride,
+                                   uint8_t *dst, int dst_stride,
+                                   const int subpel_x, const int subpel_y,
+                                   const struct scale_factors *sf, int w, int h,
+                                   ConvolveParams *conv_params,
+#if CONFIG_DUAL_FILTER
+                                   const InterpFilter *interp_filter,
+#else
+                                   const InterpFilter interp_filter,
+#endif
+                                   int xs, int ys) {
+#if CONFIG_DUAL_FILTER
+  InterpFilter filter_x = av1_get_plane_interp_filter(
+      interp_filter[1 + 2 * conv_params->ref], conv_params->plane);
+  InterpFilter filter_y = av1_get_plane_interp_filter(
+      interp_filter[0 + 2 * conv_params->ref], conv_params->plane);
+  InterpFilterParams interp_filter_params_x =
+      av1_get_interp_filter_params(filter_x);
+  InterpFilterParams interp_filter_params_y =
+      av1_get_interp_filter_params(filter_y);
+#else
+  InterpFilterParams interp_filter_params =
+      av1_get_interp_filter_params(interp_filter);
+#endif
+
+  assert(sf);
+#if CONFIG_DUAL_FILTER
+  if (interp_filter_params_x.taps == SUBPEL_TAPS &&
+      interp_filter_params_y.taps == SUBPEL_TAPS && w > 2 && h > 2 &&
+      conv_params->round == CONVOLVE_OPT_ROUND && xs == 16 && ys == 16) {
+    const int16_t *kernel_x =
+        av1_get_interp_filter_subpel_kernel(interp_filter_params_x, subpel_x);
+    const int16_t *kernel_y =
+        av1_get_interp_filter_subpel_kernel(interp_filter_params_y, subpel_y);
+#else
+  if (interp_filter_params.taps == SUBPEL_TAPS && w > 2 && h > 2 &&
+      conv_params->round == CONVOLVE_OPT_ROUND && xs == 16 && ys == 16) {
+    const int16_t *kernel_x =
+        av1_get_interp_filter_subpel_kernel(interp_filter_params, subpel_x);
+    const int16_t *kernel_y =
+        av1_get_interp_filter_subpel_kernel(interp_filter_params, subpel_y);
+#endif
+    sf->predict[subpel_x != 0][subpel_y != 0][conv_params->ref](
+        src, src_stride, dst, dst_stride, kernel_x, xs, kernel_y, ys, w, h);
+  } else {
+// ref_idx > 0 means this is the second reference frame
+// first reference frame's prediction result is already in dst
+// therefore we need to average the first and second results
+#if CONFIG_CONVOLVE_ROUND
+    if (conv_params->round == CONVOLVE_OPT_NO_ROUND && xs == 16 && ys == 16)
+      av1_convolve_2d_facade(src, src_stride, dst, dst_stride, w, h,
+#if CONFIG_DUAL_FILTER
+                             interp_filter,
+#else
+                             &interp_filter,
+#endif
+                             subpel_x, xs, subpel_y, ys, conv_params);
+    else
+#endif
+    {
+      if (xs == 16 && ys == 16) {
+        av1_convolve(src, src_stride, dst, dst_stride, w, h, interp_filter,
+                     subpel_x, xs, subpel_y, ys, conv_params);
+      } else {
+        // If xs == 16 || ys == 16 scaling is happening and the SSE2
+        // instructions don't support scaling; use the C versions to be safe.
+        av1_convolve_c(src, src_stride, dst, dst_stride, w, h, interp_filter,
+                       subpel_x, xs, subpel_y, ys, conv_params);
+      }
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE void highbd_inter_predictor(const uint8_t *src, int src_stride,
+                                          uint8_t *dst, int dst_stride,
+                                          const int subpel_x,
+                                          const int subpel_y,
+                                          const struct scale_factors *sf, int w,
+                                          int h, int ref,
+#if CONFIG_DUAL_FILTER
+                                          const InterpFilter *interp_filter,
+#else
+                                          const InterpFilter interp_filter,
+#endif
+                                          int xs, int ys, int bd) {
+#if CONFIG_DUAL_FILTER
+  InterpFilterParams interp_filter_params_x =
+      av1_get_interp_filter_params(interp_filter[1 + 2 * ref]);
+  InterpFilterParams interp_filter_params_y =
+      av1_get_interp_filter_params(interp_filter[0 + 2 * ref]);
+#else
+  InterpFilterParams interp_filter_params =
+      av1_get_interp_filter_params(interp_filter);
+#endif
+
+#if CONFIG_DUAL_FILTER
+  if (interp_filter_params_x.taps == SUBPEL_TAPS &&
+      interp_filter_params_y.taps == SUBPEL_TAPS && w > 2 && h > 2) {
+    const int16_t *kernel_x =
+        av1_get_interp_filter_subpel_kernel(interp_filter_params_x, subpel_x);
+    const int16_t *kernel_y =
+        av1_get_interp_filter_subpel_kernel(interp_filter_params_y, subpel_y);
+#else
+  if (interp_filter_params.taps == SUBPEL_TAPS && w > 2 && h > 2) {
+    const int16_t *kernel_x =
+        av1_get_interp_filter_subpel_kernel(interp_filter_params, subpel_x);
+    const int16_t *kernel_y =
+        av1_get_interp_filter_subpel_kernel(interp_filter_params, subpel_y);
+#endif  // CONFIG_DUAL_FILTER
+    sf->highbd_predict[subpel_x != 0][subpel_y != 0][ref](
+        src, src_stride, dst, dst_stride, kernel_x, xs, kernel_y, ys, w, h, bd);
+  } else {
+    // ref > 0 means this is the second reference frame
+    // first reference frame's prediction result is already in dst
+    // therefore we need to average the first and second results
+    int avg = ref > 0;
+    av1_highbd_convolve(src, src_stride, dst, dst_stride, w, h, interp_filter,
+                        subpel_x, xs, subpel_y, ys, avg, bd);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_EXT_INTER
+// Set to (1 << 5) if the 32-ary codebooks are used for any bock size
+#define MAX_WEDGE_TYPES (1 << 4)
+
+#define MAX_WEDGE_SIZE_LOG2 5  // 32x32
+#define MAX_WEDGE_SIZE (1 << MAX_WEDGE_SIZE_LOG2)
+#define MAX_WEDGE_SQUARE (MAX_WEDGE_SIZE * MAX_WEDGE_SIZE)
+
+#define WEDGE_WEIGHT_BITS 6
+
+#define WEDGE_NONE -1
+
+// Angles are with respect to horizontal anti-clockwise
+typedef enum {
+  WEDGE_HORIZONTAL = 0,
+  WEDGE_VERTICAL = 1,
+  WEDGE_OBLIQUE27 = 2,
+  WEDGE_OBLIQUE63 = 3,
+  WEDGE_OBLIQUE117 = 4,
+  WEDGE_OBLIQUE153 = 5,
+  WEDGE_DIRECTIONS
+} WedgeDirectionType;
+
+// 3-tuple: {direction, x_offset, y_offset}
+typedef struct {
+  WedgeDirectionType direction;
+  int x_offset;
+  int y_offset;
+} wedge_code_type;
+
+typedef uint8_t *wedge_masks_type[MAX_WEDGE_TYPES];
+
+typedef struct {
+  int bits;
+  const wedge_code_type *codebook;
+  uint8_t *signflip;
+  int smoother;
+  wedge_masks_type *masks;
+} wedge_params_type;
+
+extern const wedge_params_type wedge_params_lookup[BLOCK_SIZES];
+
+static INLINE int is_interinter_compound_used(COMPOUND_TYPE type,
+                                              BLOCK_SIZE sb_type) {
+  (void)sb_type;
+  switch (type) {
+    case COMPOUND_AVERAGE: return 1;
+#if CONFIG_WEDGE
+    case COMPOUND_WEDGE: return wedge_params_lookup[sb_type].bits > 0;
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+    case COMPOUND_SEG: return sb_type >= BLOCK_8X8;
+#endif  // CONFIG_COMPOUND_SEGMENT
+    default: assert(0); return 0;
+  }
+}
+
+static INLINE int is_any_masked_compound_used(BLOCK_SIZE sb_type) {
+  COMPOUND_TYPE comp_type;
+  for (comp_type = 0; comp_type < COMPOUND_TYPES; comp_type++) {
+    if (is_masked_compound_type(comp_type) &&
+        is_interinter_compound_used(comp_type, sb_type))
+      return 1;
+  }
+  return 0;
+}
+
+static INLINE int get_wedge_bits_lookup(BLOCK_SIZE sb_type) {
+  return wedge_params_lookup[sb_type].bits;
+}
+
+static INLINE int get_interinter_wedge_bits(BLOCK_SIZE sb_type) {
+  const int wbits = wedge_params_lookup[sb_type].bits;
+  return (wbits > 0) ? wbits + 1 : 0;
+}
+
+static INLINE int is_interintra_wedge_used(BLOCK_SIZE sb_type) {
+  (void)sb_type;
+  return wedge_params_lookup[sb_type].bits > 0;
+}
+
+static INLINE int get_interintra_wedge_bits(BLOCK_SIZE sb_type) {
+  return wedge_params_lookup[sb_type].bits;
+}
+
+#if CONFIG_COMPOUND_SEGMENT
+void build_compound_seg_mask(uint8_t *mask, SEG_MASK_TYPE mask_type,
+                             const uint8_t *src0, int src0_stride,
+                             const uint8_t *src1, int src1_stride,
+                             BLOCK_SIZE sb_type, int h, int w);
+#if CONFIG_HIGHBITDEPTH
+void build_compound_seg_mask_highbd(uint8_t *mask, SEG_MASK_TYPE mask_type,
+                                    const uint8_t *src0, int src0_stride,
+                                    const uint8_t *src1, int src1_stride,
+                                    BLOCK_SIZE sb_type, int h, int w, int bd);
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_COMPOUND_SEGMENT
+#endif  // CONFIG_EXT_INTER
+
+void build_inter_predictors(MACROBLOCKD *xd, int plane,
+#if CONFIG_MOTION_VAR
+                            int mi_col_offset, int mi_row_offset,
+#endif  // CONFIG_MOTION_VAR
+                            int block, int bw, int bh, int x, int y, int w,
+                            int h,
+#if CONFIG_SUPERTX && CONFIG_EXT_INTER
+                            int wedge_offset_x, int wedge_offset_y,
+#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
+                            int mi_x, int mi_y);
+
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+// This function will determine whether or not to create a warped
+// prediction and return the appropriate motion model depending
+// on the configuration. Behavior will change with different
+// combinations of GLOBAL_MOTION, WARPED_MOTION and MOTION_VAR.
+static INLINE int allow_warp(const MODE_INFO *const mi,
+                             const WarpTypesAllowed *const warp_types,
+#if CONFIG_GLOBAL_MOTION
+                             const WarpedMotionParams *const gm_params,
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_MOTION_VAR
+                             int mi_col_offset, int mi_row_offset,
+#endif  // CONFIG_MOTION_VAR
+                             WarpedMotionParams *final_warp_params) {
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  set_default_warp_params(final_warp_params);
+
+// Only global motion configured
+#if CONFIG_GLOBAL_MOTION && !CONFIG_WARPED_MOTION && !CONFIG_MOTION_VAR
+  (void)mbmi;
+  if (warp_types->global_warp_allowed) {
+    memcpy(final_warp_params, gm_params, sizeof(*final_warp_params));
+    return 1;
+  }
+#endif  // CONFIG_GLOBAL_MOTION && !CONFIG_WARPED_MOTION && !CONFIG_MOTION_VAR
+
+// Only warped motion configured
+#if CONFIG_WARPED_MOTION && !CONFIG_GLOBAL_MOTION && !CONFIG_MOTION_VAR
+  if (warp_types->local_warp_allowed) {
+    memcpy(final_warp_params, &mbmi->wm_params[0], sizeof(*final_warp_params));
+    return 1;
+  }
+#endif  // CONFIG_WARPED_MOTION && !CONFIG_GLOBAL_MOTION && !CONFIG_MOTION_VAR
+
+// Warped and global motion configured
+#if CONFIG_GLOBAL_MOTION && CONFIG_WARPED_MOTION && !CONFIG_MOTION_VAR
+  // When both are enabled, warped will take priority. The global parameters
+  // will only be used to compute projection samples to find the warped model.
+  // Note that, if SEPARATE_GLOBAL_MOTION is enabled and a block chooses
+  // global, it will not be possible to select WARPED_CAUSAL.
+  if (warp_types->local_warp_allowed) {
+    memcpy(final_warp_params, &mbmi->wm_params[0], sizeof(*final_warp_params));
+    return 1;
+  } else if (warp_types->global_warp_allowed) {
+    memcpy(final_warp_params, gm_params, sizeof(*final_warp_params));
+    return 1;
+  }
+#endif  // CONFIG_GLOBAL_MOTION && CONFIG_WARPED_MOTION && !CONFIG_MOTION_VAR
+
+// Motion var and global motion configured
+#if CONFIG_GLOBAL_MOTION && CONFIG_MOTION_VAR && !CONFIG_WARPED_MOTION
+  // We warp if either case is true:
+  //   1.) We are predicting a block which uses global motion
+  //   2.) We are predicting a neighboring block of a block using OBMC,
+  //       the neighboring block uses global motion, and we have enabled
+  //       WARP_GM_NEIGHBORS_WITH_OBMC
+  const int build_for_obmc = !(mi_col_offset == 0 && mi_row_offset == 0);
+  (void)mbmi;
+  if (warp_types->global_warp_allowed &&
+      (WARP_GM_NEIGHBORS_WITH_OBMC || !build_for_obmc)) {
+    memcpy(final_warp_params, gm_params, sizeof(*final_warp_params));
+    return 1;
+  }
+#endif  // CONFIG_GLOBAL_MOTION && CONFIG_MOTION_VAR && !CONFIG_WARPED_MOTION
+
+// Motion var and warped motion configured
+#if CONFIG_WARPED_MOTION && CONFIG_MOTION_VAR && !CONFIG_GLOBAL_MOTION
+  // We warp if either case is true:
+  //   1.) We are predicting a block with motion mode WARPED_CAUSAL
+  //   2.) We are predicting a neighboring block of a block using OBMC,
+  //       the neighboring block has mode WARPED_CAUSAL, and we have enabled
+  //       WARP_WM_NEIGHBORS_WITH_OBMC
+  const int build_for_obmc = !(mi_col_offset == 0 && mi_row_offset == 0);
+  if (warp_types->local_warp_allowed) {
+    if ((build_for_obmc && WARP_WM_NEIGHBORS_WITH_OBMC) || (!build_for_obmc)) {
+      memcpy(final_warp_params, &mbmi->wm_params[0],
+             sizeof(*final_warp_params));
+      return 1;
+    }
+  }
+#endif  // CONFIG_WARPED_MOTION && CONFIG_MOTION_VAR && !CONFIG_GLOBAL_MOTION
+
+// Motion var, warped motion and global motion all configured
+#if CONFIG_WARPED_MOTION && CONFIG_MOTION_VAR && CONFIG_GLOBAL_MOTION
+  const int build_for_obmc = !(mi_col_offset == 0 && mi_row_offset == 0);
+  if (warp_types->local_warp_allowed) {
+    if ((build_for_obmc && WARP_WM_NEIGHBORS_WITH_OBMC) || (!build_for_obmc)) {
+      memcpy(final_warp_params, &mbmi->wm_params[0],
+             sizeof(*final_warp_params));
+      return 1;
+    }
+  } else if (warp_types->global_warp_allowed &&
+             (WARP_GM_NEIGHBORS_WITH_OBMC || !build_for_obmc)) {
+    memcpy(final_warp_params, gm_params, sizeof(*final_warp_params));
+    return 1;
+  }
+#endif  // CONFIG_WARPED_MOTION && CONFIG_MOTION_VAR && CONFIG_GLOBAL_MOTION
+
+  return 0;
+}
+#endif  // CONFIG_GLOBAL_MOTION ||CONFIG_WARPED_MOTION
+
+static INLINE void av1_make_inter_predictor(
+    const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride,
+    const int subpel_x, const int subpel_y, const struct scale_factors *sf,
+    int w, int h, ConvolveParams *conv_params,
+#if CONFIG_DUAL_FILTER
+    const InterpFilter *interp_filter,
+#else
+    const InterpFilter interp_filter,
+#endif
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+    const WarpTypesAllowed *warp_types, int p_col, int p_row, int plane,
+    int ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+    int mi_col_offset, int mi_row_offset,
+#endif
+    int xs, int ys, const MACROBLOCKD *xd) {
+  (void)xd;
+
+#if CONFIG_MOTION_VAR
+  const MODE_INFO *mi = xd->mi[mi_col_offset + xd->mi_stride * mi_row_offset];
+#else
+  const MODE_INFO *mi = xd->mi[0];
+  (void)mi;
+#endif  // CONFIG_MOTION_VAR
+
+// Make sure the selected motion mode is valid for this configuration
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  assert_motion_mode_valid(mi->mbmi.motion_mode,
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+                           0, xd->global_motion,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+                           mi);
+#endif  // CONFIG MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_WARPED_MOTION || CONFIG_GLOBAL_MOTION
+  WarpedMotionParams final_warp_params;
+  const int do_warp = allow_warp(mi, warp_types,
+#if CONFIG_GLOBAL_MOTION
+                                 &xd->global_motion[mi->mbmi.ref_frame[ref]],
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_MOTION_VAR
+                                 mi_col_offset, mi_row_offset,
+#endif  // CONFIG_MOTION_VAR
+                                 &final_warp_params);
+  if (do_warp) {
+    const struct macroblockd_plane *const pd = &xd->plane[plane];
+    const struct buf_2d *const pre_buf = &pd->pre[ref];
+    av1_warp_plane(&final_warp_params,
+#if CONFIG_HIGHBITDEPTH
+                   xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH, xd->bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                   pre_buf->buf0, pre_buf->width, pre_buf->height,
+                   pre_buf->stride, dst, p_col, p_row, w, h, dst_stride,
+                   pd->subsampling_x, pd->subsampling_y, xs, ys, ref);
+    return;
+  }
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    highbd_inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
+                           sf, w, h, conv_params->ref, interp_filter, xs, ys,
+                           xd->bd);
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y, sf, w,
+                  h, conv_params, interp_filter, xs, ys);
+}
+
+#if CONFIG_EXT_INTER
+void av1_make_masked_inter_predictor(const uint8_t *pre, int pre_stride,
+                                     uint8_t *dst, int dst_stride,
+                                     const int subpel_x, const int subpel_y,
+                                     const struct scale_factors *sf, int w,
+                                     int h,
+#if CONFIG_DUAL_FILTER
+                                     const InterpFilter *interp_filter,
+#else
+                                     const InterpFilter interp_filter,
+#endif
+                                     int xs, int ys,
+#if CONFIG_SUPERTX
+                                     int wedge_offset_x, int wedge_offset_y,
+#endif  // CONFIG_SUPERTX
+                                     int plane,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     const WarpTypesAllowed *warp_types,
+                                     int p_col, int p_row, int ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     MACROBLOCKD *xd);
+#endif  // CONFIG_EXT_INTER
+
+static INLINE int round_mv_comp_q4(int value) {
+  return (value < 0 ? value - 2 : value + 2) / 4;
+}
+
+static MV mi_mv_pred_q4(const MODE_INFO *mi, int idx) {
+  MV res = {
+    round_mv_comp_q4(
+        mi->bmi[0].as_mv[idx].as_mv.row + mi->bmi[1].as_mv[idx].as_mv.row +
+        mi->bmi[2].as_mv[idx].as_mv.row + mi->bmi[3].as_mv[idx].as_mv.row),
+    round_mv_comp_q4(
+        mi->bmi[0].as_mv[idx].as_mv.col + mi->bmi[1].as_mv[idx].as_mv.col +
+        mi->bmi[2].as_mv[idx].as_mv.col + mi->bmi[3].as_mv[idx].as_mv.col)
+  };
+  return res;
+}
+
+static INLINE int round_mv_comp_q2(int value) {
+  return (value < 0 ? value - 1 : value + 1) / 2;
+}
+
+static MV mi_mv_pred_q2(const MODE_INFO *mi, int idx, int block0, int block1) {
+  MV res = { round_mv_comp_q2(mi->bmi[block0].as_mv[idx].as_mv.row +
+                              mi->bmi[block1].as_mv[idx].as_mv.row),
+             round_mv_comp_q2(mi->bmi[block0].as_mv[idx].as_mv.col +
+                              mi->bmi[block1].as_mv[idx].as_mv.col) };
+  return res;
+}
+
+// TODO(jkoleszar): yet another mv clamping function :-(
+static INLINE MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd,
+                                           const MV *src_mv, int bw, int bh,
+                                           int ss_x, int ss_y) {
+  // If the MV points so far into the UMV border that no visible pixels
+  // are used for reconstruction, the subpel part of the MV can be
+  // discarded and the MV limited to 16 pixels with equivalent results.
+  const int spel_left = (AOM_INTERP_EXTEND + bw) << SUBPEL_BITS;
+  const int spel_right = spel_left - SUBPEL_SHIFTS;
+  const int spel_top = (AOM_INTERP_EXTEND + bh) << SUBPEL_BITS;
+  const int spel_bottom = spel_top - SUBPEL_SHIFTS;
+  MV clamped_mv = { src_mv->row * (1 << (1 - ss_y)),
+                    src_mv->col * (1 << (1 - ss_x)) };
+  assert(ss_x <= 1);
+  assert(ss_y <= 1);
+
+  clamp_mv(&clamped_mv, xd->mb_to_left_edge * (1 << (1 - ss_x)) - spel_left,
+           xd->mb_to_right_edge * (1 << (1 - ss_x)) + spel_right,
+           xd->mb_to_top_edge * (1 << (1 - ss_y)) - spel_top,
+           xd->mb_to_bottom_edge * (1 << (1 - ss_y)) + spel_bottom);
+
+  return clamped_mv;
+}
+
+static INLINE MV average_split_mvs(const struct macroblockd_plane *pd,
+                                   const MODE_INFO *mi, int ref, int block) {
+  const int ss_idx = ((pd->subsampling_x > 0) << 1) | (pd->subsampling_y > 0);
+  MV res = { 0, 0 };
+  switch (ss_idx) {
+    case 0: res = mi->bmi[block].as_mv[ref].as_mv; break;
+    case 1: res = mi_mv_pred_q2(mi, ref, block, block + 2); break;
+    case 2: res = mi_mv_pred_q2(mi, ref, block, block + 1); break;
+    case 3: res = mi_mv_pred_q4(mi, ref); break;
+    default: assert(ss_idx <= 3 && ss_idx >= 0);
+  }
+  return res;
+}
+
+void av1_build_inter_predictor_sub8x8(MACROBLOCKD *xd, int plane, int i, int ir,
+                                      int ic, int mi_row, int mi_col);
+
+void av1_build_inter_predictors_sby(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                    BUFFER_SET *ctx, BLOCK_SIZE bsize);
+
+void av1_build_inter_predictors_sbuv(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                     BUFFER_SET *ctx, BLOCK_SIZE bsize);
+
+void av1_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                   BUFFER_SET *ctx, BLOCK_SIZE bsize);
+
+#if CONFIG_SUPERTX
+void av1_build_inter_predictors_sb_sub8x8_extend(MACROBLOCKD *xd,
+#if CONFIG_EXT_INTER
+                                                 int mi_row_ori, int mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                                                 int mi_row, int mi_col,
+                                                 BLOCK_SIZE bsize, int block);
+
+void av1_build_inter_predictors_sb_extend(MACROBLOCKD *xd,
+#if CONFIG_EXT_INTER
+                                          int mi_row_ori, int mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                                          int mi_row, int mi_col,
+                                          BLOCK_SIZE bsize);
+struct macroblockd_plane;
+void av1_build_masked_inter_predictor_complex(
+    MACROBLOCKD *xd, uint8_t *dst, int dst_stride, const uint8_t *pre,
+    int pre_stride, int mi_row, int mi_col, int mi_row_ori, int mi_col_ori,
+    BLOCK_SIZE bsize, BLOCK_SIZE top_bsize, PARTITION_TYPE partition,
+    int plane);
+#endif  // CONFIG_SUPERTX
+
+void av1_build_inter_predictor(const uint8_t *src, int src_stride, uint8_t *dst,
+                               int dst_stride, const MV *src_mv,
+                               const struct scale_factors *sf, int w, int h,
+                               ConvolveParams *conv_params,
+#if CONFIG_DUAL_FILTER
+                               const InterpFilter *interp_filter,
+#else
+                               const InterpFilter interp_filter,
+#endif
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                               const WarpTypesAllowed *warp_types, int p_col,
+                               int p_row, int plane, int ref,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                               enum mv_precision precision, int x, int y,
+                               const MACROBLOCKD *xd);
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_build_inter_predictor(
+    const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride,
+    const MV *mv_q3, const struct scale_factors *sf, int w, int h, int do_avg,
+#if CONFIG_DUAL_FILTER
+    const InterpFilter *interp_filter,
+#else
+    const InterpFilter interp_filter,
+#endif
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+    const WarpTypesAllowed *warp_types, int p_col, int p_row,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+    int plane, enum mv_precision precision, int x, int y,
+    const MACROBLOCKD *xd);
+#endif
+
+static INLINE int scaled_buffer_offset(int x_offset, int y_offset, int stride,
+                                       const struct scale_factors *sf) {
+  const int x = sf ? sf->scale_value_x(x_offset, sf) : x_offset;
+  const int y = sf ? sf->scale_value_y(y_offset, sf) : y_offset;
+  return y * stride + x;
+}
+
+static INLINE void setup_pred_plane(struct buf_2d *dst, BLOCK_SIZE bsize,
+                                    uint8_t *src, int width, int height,
+                                    int stride, int mi_row, int mi_col,
+                                    const struct scale_factors *scale,
+                                    int subsampling_x, int subsampling_y) {
+#if CONFIG_CHROMA_SUB8X8
+  if (bsize < BLOCK_8X8) {
+    // Offset the buffer pointer
+    if (subsampling_y && (mi_row & 0x01)) mi_row -= 1;
+    if (subsampling_x && (mi_col & 0x01)) mi_col -= 1;
+  }
+#else
+  (void)bsize;
+#endif
+
+  const int x = (MI_SIZE * mi_col) >> subsampling_x;
+  const int y = (MI_SIZE * mi_row) >> subsampling_y;
+  dst->buf = src + scaled_buffer_offset(x, y, stride, scale);
+  dst->buf0 = src;
+  dst->width = width;
+  dst->height = height;
+  dst->stride = stride;
+}
+
+void av1_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
+                          BLOCK_SIZE bsize, const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col);
+
+void av1_setup_pre_planes(MACROBLOCKD *xd, int idx,
+                          const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
+                          const struct scale_factors *sf);
+
+// Detect if the block have sub-pixel level motion vectors
+// per component.
+#define CHECK_SUBPEL 0
+static INLINE int has_subpel_mv_component(const MODE_INFO *const mi,
+                                          const MACROBLOCKD *const xd,
+                                          int dir) {
+#if CHECK_SUBPEL
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  int plane;
+  int ref = (dir >> 1);
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+  if (bsize >= BLOCK_8X8 || unify_bsize) {
+    if (dir & 0x01) {
+      if (mbmi->mv[ref].as_mv.col & SUBPEL_MASK) return 1;
+    } else {
+      if (mbmi->mv[ref].as_mv.row & SUBPEL_MASK) return 1;
+    }
+  } else {
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+      const PARTITION_TYPE bp = BLOCK_8X8 - bsize;
+      const struct macroblockd_plane *const pd = &xd->plane[plane];
+      const int have_vsplit = bp != PARTITION_HORZ;
+      const int have_hsplit = bp != PARTITION_VERT;
+      const int num_4x4_w = 2 >> ((!have_vsplit) | pd->subsampling_x);
+      const int num_4x4_h = 2 >> ((!have_hsplit) | pd->subsampling_y);
+
+      int x, y;
+      for (y = 0; y < num_4x4_h; ++y) {
+        for (x = 0; x < num_4x4_w; ++x) {
+          const MV mv = average_split_mvs(pd, mi, ref, y * 2 + x);
+          if (dir & 0x01) {
+            if (mv.col & SUBPEL_MASK) return 1;
+          } else {
+            if (mv.row & SUBPEL_MASK) return 1;
+          }
+        }
+      }
+    }
+  }
+
+  return 0;
+#else
+  (void)mi;
+  (void)xd;
+  (void)dir;
+  return 1;
+#endif
+}
+
+static INLINE void set_default_interp_filters(
+    MB_MODE_INFO *const mbmi, InterpFilter frame_interp_filter) {
+#if CONFIG_DUAL_FILTER
+  int dir;
+  for (dir = 0; dir < 4; ++dir)
+    mbmi->interp_filter[dir] = frame_interp_filter == SWITCHABLE
+                                   ? EIGHTTAP_REGULAR
+                                   : frame_interp_filter;
+#else
+  mbmi->interp_filter = frame_interp_filter == SWITCHABLE ? EIGHTTAP_REGULAR
+                                                          : frame_interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+}
+
+static INLINE int av1_is_interp_needed(const MACROBLOCKD *const xd) {
+  (void)xd;
+#if CONFIG_WARPED_MOTION
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  if (mbmi->motion_mode == WARPED_CAUSAL) return 0;
+#endif  // CONFIG_WARPED_MOTION
+#if CONFIG_GLOBAL_MOTION
+  if (is_nontrans_global_motion(xd)) return 0;
+#endif  // CONFIG_GLOBAL_MOTION
+  return 1;
+}
+
+static INLINE int av1_is_interp_search_needed(const MACROBLOCKD *const xd) {
+  MODE_INFO *const mi = xd->mi[0];
+  const int is_compound = has_second_ref(&mi->mbmi);
+  int ref;
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    int row_col;
+    for (row_col = 0; row_col < 2; ++row_col) {
+      const int dir = (ref << 1) + row_col;
+      if (has_subpel_mv_component(mi, xd, dir)) {
+        return 1;
+      }
+    }
+  }
+  return 0;
+}
+
+#if CONFIG_MOTION_VAR
+const uint8_t *av1_get_obmc_mask(int length);
+void av1_count_overlappable_neighbors(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                      int mi_row, int mi_col);
+void av1_build_obmc_inter_prediction(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                     int mi_row, int mi_col,
+                                     uint8_t *above[MAX_MB_PLANE],
+                                     int above_stride[MAX_MB_PLANE],
+                                     uint8_t *left[MAX_MB_PLANE],
+                                     int left_stride[MAX_MB_PLANE]);
+void av1_build_prediction_by_above_preds(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                         int mi_row, int mi_col,
+                                         uint8_t *tmp_buf[MAX_MB_PLANE],
+                                         int tmp_width[MAX_MB_PLANE],
+                                         int tmp_height[MAX_MB_PLANE],
+                                         int tmp_stride[MAX_MB_PLANE]);
+void av1_build_prediction_by_left_preds(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                        int mi_row, int mi_col,
+                                        uint8_t *tmp_buf[MAX_MB_PLANE],
+                                        int tmp_width[MAX_MB_PLANE],
+                                        int tmp_height[MAX_MB_PLANE],
+                                        int tmp_stride[MAX_MB_PLANE]);
+void av1_build_obmc_inter_predictors_sb(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                        int mi_row, int mi_col);
+#if CONFIG_NCOBMC
+void av1_build_ncobmc_inter_predictors_sb(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                          int mi_row, int mi_col);
+#endif
+#endif  // CONFIG_MOTION_VAR
+
+#if CONFIG_EXT_INTER
+#define MASK_MASTER_SIZE ((MAX_WEDGE_SIZE) << 1)
+#define MASK_MASTER_STRIDE (MASK_MASTER_SIZE)
+
+void av1_init_wedge_masks();
+
+static INLINE const uint8_t *av1_get_contiguous_soft_mask(int wedge_index,
+                                                          int wedge_sign,
+                                                          BLOCK_SIZE sb_type) {
+  return wedge_params_lookup[sb_type].masks[wedge_sign][wedge_index];
+}
+
+const uint8_t *av1_get_soft_mask(int wedge_index, int wedge_sign,
+                                 BLOCK_SIZE sb_type, int wedge_offset_x,
+                                 int wedge_offset_y);
+
+const uint8_t *av1_get_compound_type_mask_inverse(
+    const INTERINTER_COMPOUND_DATA *const comp_data,
+#if CONFIG_COMPOUND_SEGMENT
+    uint8_t *mask_buffer, int h, int w, int stride,
+#endif
+    BLOCK_SIZE sb_type);
+
+const uint8_t *av1_get_compound_type_mask(
+    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE sb_type);
+
+void av1_build_interintra_predictors(MACROBLOCKD *xd, uint8_t *ypred,
+                                     uint8_t *upred, uint8_t *vpred,
+                                     int ystride, int ustride, int vstride,
+                                     BUFFER_SET *ctx, BLOCK_SIZE bsize);
+void av1_build_interintra_predictors_sby(MACROBLOCKD *xd, uint8_t *ypred,
+                                         int ystride, BUFFER_SET *ctx,
+                                         BLOCK_SIZE bsize);
+void av1_build_interintra_predictors_sbc(MACROBLOCKD *xd, uint8_t *upred,
+                                         int ustride, BUFFER_SET *ctx,
+                                         int plane, BLOCK_SIZE bsize);
+void av1_build_interintra_predictors_sbuv(MACROBLOCKD *xd, uint8_t *upred,
+                                          uint8_t *vpred, int ustride,
+                                          int vstride, BUFFER_SET *ctx,
+                                          BLOCK_SIZE bsize);
+
+void av1_build_intra_predictors_for_interintra(MACROBLOCKD *xd,
+                                               BLOCK_SIZE bsize, int plane,
+                                               BUFFER_SET *ctx,
+                                               uint8_t *intra_pred,
+                                               int intra_stride);
+void av1_combine_interintra(MACROBLOCKD *xd, BLOCK_SIZE bsize, int plane,
+                            const uint8_t *inter_pred, int inter_stride,
+                            const uint8_t *intra_pred, int intra_stride);
+
+// Encoder only
+void av1_build_inter_predictors_for_planes_single_buf(
+    MACROBLOCKD *xd, BLOCK_SIZE bsize, int plane_from, int plane_to, int mi_row,
+    int mi_col, int ref, uint8_t *ext_dst[3], int ext_dst_stride[3]);
+void av1_build_wedge_inter_predictor_from_buf(
+    MACROBLOCKD *xd, BLOCK_SIZE bsize, int plane_from, int plane_to,
+#if CONFIG_SUPERTX
+    int wedge_offset_x, int wedge_offset_y,
+#endif  // CONFIG_SUPERTX
+    uint8_t *ext_dst0[3], int ext_dst_stride0[3], uint8_t *ext_dst1[3],
+    int ext_dst_stride1[3]);
+#endif  // CONFIG_EXT_INTER
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_RECONINTER_H_
diff --git a/third_party/aom/av1/common/reconintra.c b/third_party/aom/av1/common/reconintra.c
new file mode 100644
index 0000000000..6e0ff52ce6
--- /dev/null
+++ b/third_party/aom/av1/common/reconintra.c
@@ -0,0 +1,2467 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/system_state.h"
+
+#if CONFIG_HIGHBITDEPTH
+#include "aom_dsp/aom_dsp_common.h"
+#endif  // CONFIG_HIGHBITDEPTH
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/aom_once.h"
+#include "av1/common/reconintra.h"
+#include "av1/common/onyxc_int.h"
+#if CONFIG_CFL
+#include "av1/common/cfl.h"
+#endif
+
+enum {
+  NEED_LEFT = 1 << 1,
+  NEED_ABOVE = 1 << 2,
+  NEED_ABOVERIGHT = 1 << 3,
+  NEED_ABOVELEFT = 1 << 4,
+  NEED_BOTTOMLEFT = 1 << 5,
+};
+
+static const uint8_t extend_modes[INTRA_MODES] = {
+  NEED_ABOVE | NEED_LEFT,                   // DC
+  NEED_ABOVE,                               // V
+  NEED_LEFT,                                // H
+  NEED_ABOVE | NEED_ABOVERIGHT,             // D45
+  NEED_LEFT | NEED_ABOVE | NEED_ABOVELEFT,  // D135
+  NEED_LEFT | NEED_ABOVE | NEED_ABOVELEFT,  // D117
+  NEED_LEFT | NEED_ABOVE | NEED_ABOVELEFT,  // D153
+  NEED_LEFT | NEED_BOTTOMLEFT,              // D207
+  NEED_ABOVE | NEED_ABOVERIGHT,             // D63
+#if CONFIG_ALT_INTRA
+  NEED_LEFT | NEED_ABOVE,                   // SMOOTH
+#endif                                      // CONFIG_ALT_INTRA
+  NEED_LEFT | NEED_ABOVE | NEED_ABOVELEFT,  // TM
+};
+
+static const uint16_t orders_128x128[1] = { 0 };
+static const uint16_t orders_128x64[2] = { 0, 1 };
+static const uint16_t orders_64x128[2] = { 0, 1 };
+static const uint16_t orders_64x64[4] = {
+  0, 1, 2, 3,
+};
+static const uint16_t orders_64x32[8] = {
+  0, 2, 1, 3, 4, 6, 5, 7,
+};
+static const uint16_t orders_32x64[8] = {
+  0, 1, 2, 3, 4, 5, 6, 7,
+};
+static const uint16_t orders_32x32[16] = {
+  0, 1, 4, 5, 2, 3, 6, 7, 8, 9, 12, 13, 10, 11, 14, 15,
+};
+static const uint16_t orders_32x16[32] = {
+  0,  2,  8,  10, 1,  3,  9,  11, 4,  6,  12, 14, 5,  7,  13, 15,
+  16, 18, 24, 26, 17, 19, 25, 27, 20, 22, 28, 30, 21, 23, 29, 31,
+};
+static const uint16_t orders_16x32[32] = {
+  0,  1,  2,  3,  8,  9,  10, 11, 4,  5,  6,  7,  12, 13, 14, 15,
+  16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31,
+};
+static const uint16_t orders_16x16[64] = {
+  0,  1,  4,  5,  16, 17, 20, 21, 2,  3,  6,  7,  18, 19, 22, 23,
+  8,  9,  12, 13, 24, 25, 28, 29, 10, 11, 14, 15, 26, 27, 30, 31,
+  32, 33, 36, 37, 48, 49, 52, 53, 34, 35, 38, 39, 50, 51, 54, 55,
+  40, 41, 44, 45, 56, 57, 60, 61, 42, 43, 46, 47, 58, 59, 62, 63,
+};
+
+#if CONFIG_CB4X4 || CONFIG_EXT_PARTITION
+static const uint16_t orders_16x8[128] = {
+  0,  2,  8,  10, 32,  34,  40,  42,  1,  3,  9,  11, 33,  35,  41,  43,
+  4,  6,  12, 14, 36,  38,  44,  46,  5,  7,  13, 15, 37,  39,  45,  47,
+  16, 18, 24, 26, 48,  50,  56,  58,  17, 19, 25, 27, 49,  51,  57,  59,
+  20, 22, 28, 30, 52,  54,  60,  62,  21, 23, 29, 31, 53,  55,  61,  63,
+  64, 66, 72, 74, 96,  98,  104, 106, 65, 67, 73, 75, 97,  99,  105, 107,
+  68, 70, 76, 78, 100, 102, 108, 110, 69, 71, 77, 79, 101, 103, 109, 111,
+  80, 82, 88, 90, 112, 114, 120, 122, 81, 83, 89, 91, 113, 115, 121, 123,
+  84, 86, 92, 94, 116, 118, 124, 126, 85, 87, 93, 95, 117, 119, 125, 127,
+};
+static const uint16_t orders_8x16[128] = {
+  0,  1,  2,  3,  8,  9,  10, 11, 32,  33,  34,  35,  40,  41,  42,  43,
+  4,  5,  6,  7,  12, 13, 14, 15, 36,  37,  38,  39,  44,  45,  46,  47,
+  16, 17, 18, 19, 24, 25, 26, 27, 48,  49,  50,  51,  56,  57,  58,  59,
+  20, 21, 22, 23, 28, 29, 30, 31, 52,  53,  54,  55,  60,  61,  62,  63,
+  64, 65, 66, 67, 72, 73, 74, 75, 96,  97,  98,  99,  104, 105, 106, 107,
+  68, 69, 70, 71, 76, 77, 78, 79, 100, 101, 102, 103, 108, 109, 110, 111,
+  80, 81, 82, 83, 88, 89, 90, 91, 112, 113, 114, 115, 120, 121, 122, 123,
+  84, 85, 86, 87, 92, 93, 94, 95, 116, 117, 118, 119, 124, 125, 126, 127,
+};
+static const uint16_t orders_8x8[256] = {
+  0,   1,   4,   5,   16,  17,  20,  21,  64,  65,  68,  69,  80,  81,  84,
+  85,  2,   3,   6,   7,   18,  19,  22,  23,  66,  67,  70,  71,  82,  83,
+  86,  87,  8,   9,   12,  13,  24,  25,  28,  29,  72,  73,  76,  77,  88,
+  89,  92,  93,  10,  11,  14,  15,  26,  27,  30,  31,  74,  75,  78,  79,
+  90,  91,  94,  95,  32,  33,  36,  37,  48,  49,  52,  53,  96,  97,  100,
+  101, 112, 113, 116, 117, 34,  35,  38,  39,  50,  51,  54,  55,  98,  99,
+  102, 103, 114, 115, 118, 119, 40,  41,  44,  45,  56,  57,  60,  61,  104,
+  105, 108, 109, 120, 121, 124, 125, 42,  43,  46,  47,  58,  59,  62,  63,
+  106, 107, 110, 111, 122, 123, 126, 127, 128, 129, 132, 133, 144, 145, 148,
+  149, 192, 193, 196, 197, 208, 209, 212, 213, 130, 131, 134, 135, 146, 147,
+  150, 151, 194, 195, 198, 199, 210, 211, 214, 215, 136, 137, 140, 141, 152,
+  153, 156, 157, 200, 201, 204, 205, 216, 217, 220, 221, 138, 139, 142, 143,
+  154, 155, 158, 159, 202, 203, 206, 207, 218, 219, 222, 223, 160, 161, 164,
+  165, 176, 177, 180, 181, 224, 225, 228, 229, 240, 241, 244, 245, 162, 163,
+  166, 167, 178, 179, 182, 183, 226, 227, 230, 231, 242, 243, 246, 247, 168,
+  169, 172, 173, 184, 185, 188, 189, 232, 233, 236, 237, 248, 249, 252, 253,
+  170, 171, 174, 175, 186, 187, 190, 191, 234, 235, 238, 239, 250, 251, 254,
+  255,
+};
+
+#if CONFIG_CB4X4 && CONFIG_EXT_PARTITION
+static const uint16_t orders_4x8[512] = {
+  0,   1,   2,   3,   8,   9,   10,  11,  32,  33,  34,  35,  40,  41,  42,
+  43,  128, 129, 130, 131, 136, 137, 138, 139, 160, 161, 162, 163, 168, 169,
+  170, 171, 4,   5,   6,   7,   12,  13,  14,  15,  36,  37,  38,  39,  44,
+  45,  46,  47,  132, 133, 134, 135, 140, 141, 142, 143, 164, 165, 166, 167,
+  172, 173, 174, 175, 16,  17,  18,  19,  24,  25,  26,  27,  48,  49,  50,
+  51,  56,  57,  58,  59,  144, 145, 146, 147, 152, 153, 154, 155, 176, 177,
+  178, 179, 184, 185, 186, 187, 20,  21,  22,  23,  28,  29,  30,  31,  52,
+  53,  54,  55,  60,  61,  62,  63,  148, 149, 150, 151, 156, 157, 158, 159,
+  180, 181, 182, 183, 188, 189, 190, 191, 64,  65,  66,  67,  72,  73,  74,
+  75,  96,  97,  98,  99,  104, 105, 106, 107, 192, 193, 194, 195, 200, 201,
+  202, 203, 224, 225, 226, 227, 232, 233, 234, 235, 68,  69,  70,  71,  76,
+  77,  78,  79,  100, 101, 102, 103, 108, 109, 110, 111, 196, 197, 198, 199,
+  204, 205, 206, 207, 228, 229, 230, 231, 236, 237, 238, 239, 80,  81,  82,
+  83,  88,  89,  90,  91,  112, 113, 114, 115, 120, 121, 122, 123, 208, 209,
+  210, 211, 216, 217, 218, 219, 240, 241, 242, 243, 248, 249, 250, 251, 84,
+  85,  86,  87,  92,  93,  94,  95,  116, 117, 118, 119, 124, 125, 126, 127,
+  212, 213, 214, 215, 220, 221, 222, 223, 244, 245, 246, 247, 252, 253, 254,
+  255, 256, 257, 258, 259, 264, 265, 266, 267, 288, 289, 290, 291, 296, 297,
+  298, 299, 384, 385, 386, 387, 392, 393, 394, 395, 416, 417, 418, 419, 424,
+  425, 426, 427, 260, 261, 262, 263, 268, 269, 270, 271, 292, 293, 294, 295,
+  300, 301, 302, 303, 388, 389, 390, 391, 396, 397, 398, 399, 420, 421, 422,
+  423, 428, 429, 430, 431, 272, 273, 274, 275, 280, 281, 282, 283, 304, 305,
+  306, 307, 312, 313, 314, 315, 400, 401, 402, 403, 408, 409, 410, 411, 432,
+  433, 434, 435, 440, 441, 442, 443, 276, 277, 278, 279, 284, 285, 286, 287,
+  308, 309, 310, 311, 316, 317, 318, 319, 404, 405, 406, 407, 412, 413, 414,
+  415, 436, 437, 438, 439, 444, 445, 446, 447, 320, 321, 322, 323, 328, 329,
+  330, 331, 352, 353, 354, 355, 360, 361, 362, 363, 448, 449, 450, 451, 456,
+  457, 458, 459, 480, 481, 482, 483, 488, 489, 490, 491, 324, 325, 326, 327,
+  332, 333, 334, 335, 356, 357, 358, 359, 364, 365, 366, 367, 452, 453, 454,
+  455, 460, 461, 462, 463, 484, 485, 486, 487, 492, 493, 494, 495, 336, 337,
+  338, 339, 344, 345, 346, 347, 368, 369, 370, 371, 376, 377, 378, 379, 464,
+  465, 466, 467, 472, 473, 474, 475, 496, 497, 498, 499, 504, 505, 506, 507,
+  340, 341, 342, 343, 348, 349, 350, 351, 372, 373, 374, 375, 380, 381, 382,
+  383, 468, 469, 470, 471, 476, 477, 478, 479, 500, 501, 502, 503, 508, 509,
+  510, 511,
+};
+
+static const uint16_t orders_8x4[512] = {
+  0,   2,   8,   10,  32,  34,  40,  42,  128, 130, 136, 138, 160, 162, 168,
+  170, 1,   3,   9,   11,  33,  35,  41,  43,  129, 131, 137, 139, 161, 163,
+  169, 171, 4,   6,   12,  14,  36,  38,  44,  46,  132, 134, 140, 142, 164,
+  166, 172, 174, 5,   7,   13,  15,  37,  39,  45,  47,  133, 135, 141, 143,
+  165, 167, 173, 175, 16,  18,  24,  26,  48,  50,  56,  58,  144, 146, 152,
+  154, 176, 178, 184, 186, 17,  19,  25,  27,  49,  51,  57,  59,  145, 147,
+  153, 155, 177, 179, 185, 187, 20,  22,  28,  30,  52,  54,  60,  62,  148,
+  150, 156, 158, 180, 182, 188, 190, 21,  23,  29,  31,  53,  55,  61,  63,
+  149, 151, 157, 159, 181, 183, 189, 191, 64,  66,  72,  74,  96,  98,  104,
+  106, 192, 194, 200, 202, 224, 226, 232, 234, 65,  67,  73,  75,  97,  99,
+  105, 107, 193, 195, 201, 203, 225, 227, 233, 235, 68,  70,  76,  78,  100,
+  102, 108, 110, 196, 198, 204, 206, 228, 230, 236, 238, 69,  71,  77,  79,
+  101, 103, 109, 111, 197, 199, 205, 207, 229, 231, 237, 239, 80,  82,  88,
+  90,  112, 114, 120, 122, 208, 210, 216, 218, 240, 242, 248, 250, 81,  83,
+  89,  91,  113, 115, 121, 123, 209, 211, 217, 219, 241, 243, 249, 251, 84,
+  86,  92,  94,  116, 118, 124, 126, 212, 214, 220, 222, 244, 246, 252, 254,
+  85,  87,  93,  95,  117, 119, 125, 127, 213, 215, 221, 223, 245, 247, 253,
+  255, 256, 258, 264, 266, 288, 290, 296, 298, 384, 386, 392, 394, 416, 418,
+  424, 426, 257, 259, 265, 267, 289, 291, 297, 299, 385, 387, 393, 395, 417,
+  419, 425, 427, 260, 262, 268, 270, 292, 294, 300, 302, 388, 390, 396, 398,
+  420, 422, 428, 430, 261, 263, 269, 271, 293, 295, 301, 303, 389, 391, 397,
+  399, 421, 423, 429, 431, 272, 274, 280, 282, 304, 306, 312, 314, 400, 402,
+  408, 410, 432, 434, 440, 442, 273, 275, 281, 283, 305, 307, 313, 315, 401,
+  403, 409, 411, 433, 435, 441, 443, 276, 278, 284, 286, 308, 310, 316, 318,
+  404, 406, 412, 414, 436, 438, 444, 446, 277, 279, 285, 287, 309, 311, 317,
+  319, 405, 407, 413, 415, 437, 439, 445, 447, 320, 322, 328, 330, 352, 354,
+  360, 362, 448, 450, 456, 458, 480, 482, 488, 490, 321, 323, 329, 331, 353,
+  355, 361, 363, 449, 451, 457, 459, 481, 483, 489, 491, 324, 326, 332, 334,
+  356, 358, 364, 366, 452, 454, 460, 462, 484, 486, 492, 494, 325, 327, 333,
+  335, 357, 359, 365, 367, 453, 455, 461, 463, 485, 487, 493, 495, 336, 338,
+  344, 346, 368, 370, 376, 378, 464, 466, 472, 474, 496, 498, 504, 506, 337,
+  339, 345, 347, 369, 371, 377, 379, 465, 467, 473, 475, 497, 499, 505, 507,
+  340, 342, 348, 350, 372, 374, 380, 382, 468, 470, 476, 478, 500, 502, 508,
+  510, 341, 343, 349, 351, 373, 375, 381, 383, 469, 471, 477, 479, 501, 503,
+  509, 511,
+};
+
+static const uint16_t orders_4x4[1024] = {
+  0,    1,    4,    5,    16,   17,   20,   21,   64,   65,   68,   69,   80,
+  81,   84,   85,   256,  257,  260,  261,  272,  273,  276,  277,  320,  321,
+  324,  325,  336,  337,  340,  341,  2,    3,    6,    7,    18,   19,   22,
+  23,   66,   67,   70,   71,   82,   83,   86,   87,   258,  259,  262,  263,
+  274,  275,  278,  279,  322,  323,  326,  327,  338,  339,  342,  343,  8,
+  9,    12,   13,   24,   25,   28,   29,   72,   73,   76,   77,   88,   89,
+  92,   93,   264,  265,  268,  269,  280,  281,  284,  285,  328,  329,  332,
+  333,  344,  345,  348,  349,  10,   11,   14,   15,   26,   27,   30,   31,
+  74,   75,   78,   79,   90,   91,   94,   95,   266,  267,  270,  271,  282,
+  283,  286,  287,  330,  331,  334,  335,  346,  347,  350,  351,  32,   33,
+  36,   37,   48,   49,   52,   53,   96,   97,   100,  101,  112,  113,  116,
+  117,  288,  289,  292,  293,  304,  305,  308,  309,  352,  353,  356,  357,
+  368,  369,  372,  373,  34,   35,   38,   39,   50,   51,   54,   55,   98,
+  99,   102,  103,  114,  115,  118,  119,  290,  291,  294,  295,  306,  307,
+  310,  311,  354,  355,  358,  359,  370,  371,  374,  375,  40,   41,   44,
+  45,   56,   57,   60,   61,   104,  105,  108,  109,  120,  121,  124,  125,
+  296,  297,  300,  301,  312,  313,  316,  317,  360,  361,  364,  365,  376,
+  377,  380,  381,  42,   43,   46,   47,   58,   59,   62,   63,   106,  107,
+  110,  111,  122,  123,  126,  127,  298,  299,  302,  303,  314,  315,  318,
+  319,  362,  363,  366,  367,  378,  379,  382,  383,  128,  129,  132,  133,
+  144,  145,  148,  149,  192,  193,  196,  197,  208,  209,  212,  213,  384,
+  385,  388,  389,  400,  401,  404,  405,  448,  449,  452,  453,  464,  465,
+  468,  469,  130,  131,  134,  135,  146,  147,  150,  151,  194,  195,  198,
+  199,  210,  211,  214,  215,  386,  387,  390,  391,  402,  403,  406,  407,
+  450,  451,  454,  455,  466,  467,  470,  471,  136,  137,  140,  141,  152,
+  153,  156,  157,  200,  201,  204,  205,  216,  217,  220,  221,  392,  393,
+  396,  397,  408,  409,  412,  413,  456,  457,  460,  461,  472,  473,  476,
+  477,  138,  139,  142,  143,  154,  155,  158,  159,  202,  203,  206,  207,
+  218,  219,  222,  223,  394,  395,  398,  399,  410,  411,  414,  415,  458,
+  459,  462,  463,  474,  475,  478,  479,  160,  161,  164,  165,  176,  177,
+  180,  181,  224,  225,  228,  229,  240,  241,  244,  245,  416,  417,  420,
+  421,  432,  433,  436,  437,  480,  481,  484,  485,  496,  497,  500,  501,
+  162,  163,  166,  167,  178,  179,  182,  183,  226,  227,  230,  231,  242,
+  243,  246,  247,  418,  419,  422,  423,  434,  435,  438,  439,  482,  483,
+  486,  487,  498,  499,  502,  503,  168,  169,  172,  173,  184,  185,  188,
+  189,  232,  233,  236,  237,  248,  249,  252,  253,  424,  425,  428,  429,
+  440,  441,  444,  445,  488,  489,  492,  493,  504,  505,  508,  509,  170,
+  171,  174,  175,  186,  187,  190,  191,  234,  235,  238,  239,  250,  251,
+  254,  255,  426,  427,  430,  431,  442,  443,  446,  447,  490,  491,  494,
+  495,  506,  507,  510,  511,  512,  513,  516,  517,  528,  529,  532,  533,
+  576,  577,  580,  581,  592,  593,  596,  597,  768,  769,  772,  773,  784,
+  785,  788,  789,  832,  833,  836,  837,  848,  849,  852,  853,  514,  515,
+  518,  519,  530,  531,  534,  535,  578,  579,  582,  583,  594,  595,  598,
+  599,  770,  771,  774,  775,  786,  787,  790,  791,  834,  835,  838,  839,
+  850,  851,  854,  855,  520,  521,  524,  525,  536,  537,  540,  541,  584,
+  585,  588,  589,  600,  601,  604,  605,  776,  777,  780,  781,  792,  793,
+  796,  797,  840,  841,  844,  845,  856,  857,  860,  861,  522,  523,  526,
+  527,  538,  539,  542,  543,  586,  587,  590,  591,  602,  603,  606,  607,
+  778,  779,  782,  783,  794,  795,  798,  799,  842,  843,  846,  847,  858,
+  859,  862,  863,  544,  545,  548,  549,  560,  561,  564,  565,  608,  609,
+  612,  613,  624,  625,  628,  629,  800,  801,  804,  805,  816,  817,  820,
+  821,  864,  865,  868,  869,  880,  881,  884,  885,  546,  547,  550,  551,
+  562,  563,  566,  567,  610,  611,  614,  615,  626,  627,  630,  631,  802,
+  803,  806,  807,  818,  819,  822,  823,  866,  867,  870,  871,  882,  883,
+  886,  887,  552,  553,  556,  557,  568,  569,  572,  573,  616,  617,  620,
+  621,  632,  633,  636,  637,  808,  809,  812,  813,  824,  825,  828,  829,
+  872,  873,  876,  877,  888,  889,  892,  893,  554,  555,  558,  559,  570,
+  571,  574,  575,  618,  619,  622,  623,  634,  635,  638,  639,  810,  811,
+  814,  815,  826,  827,  830,  831,  874,  875,  878,  879,  890,  891,  894,
+  895,  640,  641,  644,  645,  656,  657,  660,  661,  704,  705,  708,  709,
+  720,  721,  724,  725,  896,  897,  900,  901,  912,  913,  916,  917,  960,
+  961,  964,  965,  976,  977,  980,  981,  642,  643,  646,  647,  658,  659,
+  662,  663,  706,  707,  710,  711,  722,  723,  726,  727,  898,  899,  902,
+  903,  914,  915,  918,  919,  962,  963,  966,  967,  978,  979,  982,  983,
+  648,  649,  652,  653,  664,  665,  668,  669,  712,  713,  716,  717,  728,
+  729,  732,  733,  904,  905,  908,  909,  920,  921,  924,  925,  968,  969,
+  972,  973,  984,  985,  988,  989,  650,  651,  654,  655,  666,  667,  670,
+  671,  714,  715,  718,  719,  730,  731,  734,  735,  906,  907,  910,  911,
+  922,  923,  926,  927,  970,  971,  974,  975,  986,  987,  990,  991,  672,
+  673,  676,  677,  688,  689,  692,  693,  736,  737,  740,  741,  752,  753,
+  756,  757,  928,  929,  932,  933,  944,  945,  948,  949,  992,  993,  996,
+  997,  1008, 1009, 1012, 1013, 674,  675,  678,  679,  690,  691,  694,  695,
+  738,  739,  742,  743,  754,  755,  758,  759,  930,  931,  934,  935,  946,
+  947,  950,  951,  994,  995,  998,  999,  1010, 1011, 1014, 1015, 680,  681,
+  684,  685,  696,  697,  700,  701,  744,  745,  748,  749,  760,  761,  764,
+  765,  936,  937,  940,  941,  952,  953,  956,  957,  1000, 1001, 1004, 1005,
+  1016, 1017, 1020, 1021, 682,  683,  686,  687,  698,  699,  702,  703,  746,
+  747,  750,  751,  762,  763,  766,  767,  938,  939,  942,  943,  954,  955,
+  958,  959,  1002, 1003, 1006, 1007, 1018, 1019, 1022, 1023,
+};
+#endif
+#endif  // CONFIG_CB4X4 || CONFIG_EXT_PARTITION
+
+#if CONFIG_EXT_PARTITION
+/* clang-format off */
+static const uint16_t *const orders[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  // 2X2,         2X4,            4X2
+  orders_4x4,     orders_4x4,     orders_4x4,
+  //                              4X4
+                                  orders_4x4,
+  // 4X8,         8X4,            8X8
+  orders_4x8,     orders_8x4,     orders_8x8,
+#else
+  //                              4X4
+                                  orders_8x8,
+  // 4X8,         8X4,            8X8
+  orders_8x8,     orders_8x8,     orders_8x8,
+#endif
+  // 8X16,        16X8,           16X16
+  orders_8x16,    orders_16x8,    orders_16x16,
+  // 16X32,       32X16,          32X32
+  orders_16x32,   orders_32x16,   orders_32x32,
+  // 32X64,       64X32,          64X64
+  orders_32x64,   orders_64x32,   orders_64x64,
+  // 64x128,      128x64,         128x128
+  orders_64x128,  orders_128x64,  orders_128x128
+};
+/* clang-format on */
+#else
+/* clang-format off */
+static const uint16_t *const orders[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  // 2X2,         2X4,            4X2
+  orders_8x8,     orders_8x8,     orders_8x8,
+  //                              4X4
+                                  orders_8x8,
+  // 4X8,         8X4,            8X8
+  orders_8x16,    orders_16x8,    orders_16x16,
+#else
+  //                              4X4
+                                  orders_16x16,
+  // 4X8,         8X4,            8X8
+  orders_16x16,   orders_16x16,   orders_16x16,
+#endif
+  // 8X16,        16X8,           16X16
+  orders_16x32,   orders_32x16,   orders_32x32,
+  // 16X32,       32X16,          32X32
+  orders_32x64,   orders_64x32,   orders_64x64,
+  // 32X64,       64X32,          64X64
+  orders_64x128,  orders_128x64,  orders_128x128
+};
+/* clang-format on */
+#endif  // CONFIG_EXT_PARTITION
+
+#if CONFIG_EXT_PARTITION_TYPES
+static const uint16_t orders_verta_64x64[4] = {
+  0, 2, 1, 2,
+};
+static const uint16_t orders_verta_32x32[16] = {
+  0, 2, 4, 6, 1, 2, 5, 6, 8, 10, 12, 14, 9, 10, 13, 14,
+};
+static const uint16_t orders_verta_16x16[64] = {
+  0,  2,  4,  6,  16, 18, 20, 22, 1,  2,  5,  6,  17, 18, 21, 22,
+  8,  10, 12, 14, 24, 26, 28, 30, 9,  10, 13, 14, 25, 26, 29, 30,
+  32, 34, 36, 38, 48, 50, 52, 54, 33, 34, 37, 38, 49, 50, 53, 54,
+  40, 42, 44, 46, 56, 58, 60, 62, 41, 42, 45, 46, 57, 58, 61, 62,
+};
+#if CONFIG_EXT_PARTITION || CONFIG_CB4X4
+static const uint16_t orders_verta_8x8[256] = {
+  0,   2,   4,   6,   16,  18,  20,  22,  64,  66,  68,  70,  80,  82,  84,
+  86,  1,   2,   5,   6,   17,  18,  21,  22,  65,  66,  69,  70,  81,  82,
+  85,  86,  8,   10,  12,  14,  24,  26,  28,  30,  72,  74,  76,  78,  88,
+  90,  92,  94,  9,   10,  13,  14,  25,  26,  29,  30,  73,  74,  77,  78,
+  89,  90,  93,  94,  32,  34,  36,  38,  48,  50,  52,  54,  96,  98,  100,
+  102, 112, 114, 116, 118, 33,  34,  37,  38,  49,  50,  53,  54,  97,  98,
+  101, 102, 113, 114, 117, 118, 40,  42,  44,  46,  56,  58,  60,  62,  104,
+  106, 108, 110, 120, 122, 124, 126, 41,  42,  45,  46,  57,  58,  61,  62,
+  105, 106, 109, 110, 121, 122, 125, 126, 128, 130, 132, 134, 144, 146, 148,
+  150, 192, 194, 196, 198, 208, 210, 212, 214, 129, 130, 133, 134, 145, 146,
+  149, 150, 193, 194, 197, 198, 209, 210, 213, 214, 136, 138, 140, 142, 152,
+  154, 156, 158, 200, 202, 204, 206, 216, 218, 220, 222, 137, 138, 141, 142,
+  153, 154, 157, 158, 201, 202, 205, 206, 217, 218, 221, 222, 160, 162, 164,
+  166, 176, 178, 180, 182, 224, 226, 228, 230, 240, 242, 244, 246, 161, 162,
+  165, 166, 177, 178, 181, 182, 225, 226, 229, 230, 241, 242, 245, 246, 168,
+  170, 172, 174, 184, 186, 188, 190, 232, 234, 236, 238, 248, 250, 252, 254,
+  169, 170, 173, 174, 185, 186, 189, 190, 233, 234, 237, 238, 249, 250, 253,
+  254,
+};
+#endif  // CONFIG_EXT_PARTITION || CONFIG_CB4X4
+
+#if CONFIG_EXT_PARTITION
+/* clang-format off */
+static const uint16_t *const orders_verta[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  // 2X2,           2X4,              4X2
+  orders_4x4,       orders_4x4,       orders_4x4,
+#endif
+  //                                  4X4
+                                      orders_verta_8x8,
+  // 4X8,           8X4,              8X8
+  orders_verta_8x8, orders_verta_8x8, orders_verta_8x8,
+  // 8X16,          16X8,             16X16
+  orders_8x16,      orders_16x8,      orders_verta_16x16,
+  // 16X32,         32X16,            32X32
+  orders_16x32,     orders_32x16,     orders_verta_32x32,
+  // 32X64,         64X32,            64X64
+  orders_32x64,     orders_64x32,     orders_verta_64x64,
+  // 64x128,        128x64,           128x128
+  orders_64x128,    orders_128x64,    orders_128x128
+};
+/* clang-format on */
+#else
+/* clang-format off */
+static const uint16_t *const orders_verta[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  // 2X2,             2X4,                4X2
+  orders_verta_8x8,   orders_verta_8x8,   orders_verta_8x8,
+  //                                      4X4
+                                          orders_verta_8x8,
+  // 4X8,             8X4,                8X8
+  orders_verta_8x8,   orders_verta_8x8,   orders_verta_16x16,
+#else
+  //                                      4X4
+                                          orders_verta_16x16,
+  // 4X8,             8X4,                8X8
+  orders_verta_16x16, orders_verta_16x16, orders_verta_16x16,
+#endif
+  // 8X16,            16X8,               16X16
+  orders_16x32,       orders_32x16,       orders_verta_32x32,
+  // 16X32,           32X16,              32X32
+  orders_32x64,       orders_64x32,       orders_verta_64x64,
+  // 32X64,           64X32,              64X64
+  orders_64x128,      orders_128x64,      orders_128x128
+};
+/* clang-format on */
+#endif  // CONFIG_EXT_PARTITION
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+static int has_top_right(BLOCK_SIZE bsize, int mi_row, int mi_col,
+                         int top_available, int right_available,
+#if CONFIG_EXT_PARTITION_TYPES
+                         PARTITION_TYPE partition,
+#endif
+                         TX_SIZE txsz, int row_off, int col_off, int ss_x) {
+  if (!top_available || !right_available) return 0;
+
+#if !CONFIG_CB4X4
+  // TODO(bshacklett, huisu): Currently the RD loop traverses 4X8 blocks in
+  // inverted N order while in the bitstream the subblocks are stored in Z
+  // order. This discrepancy makes this function incorrect when considering 4X8
+  // blocks in the RD loop, so we disable the extended right edge for these
+  // blocks. The correct solution is to change the bitstream to store these
+  // blocks in inverted N order, and then update this function appropriately.
+  if (bsize == BLOCK_4X8 && row_off == 1) return 0;
+#endif
+
+  const int bw_unit = block_size_wide[bsize] >> tx_size_wide_log2[0];
+  const int plane_bw_unit = AOMMAX(bw_unit >> ss_x, 1);
+  const int top_right_count_unit = tx_size_wide_unit[txsz];
+
+  // Special handling for block sizes 4x8 and 4x4.
+  if (ss_x == 0 && bw_unit < 2 && col_off == 0) return 1;
+
+  if (row_off > 0) {  // Just need to check if enough pixels on the right.
+    return col_off + top_right_count_unit < plane_bw_unit;
+  } else {
+    // All top-right pixels are in the block above, which is already available.
+    if (col_off + top_right_count_unit < plane_bw_unit) return 1;
+
+    const int bw_in_mi_log2 = mi_width_log2_lookup[bsize];
+    const int bh_in_mi_log2 = mi_height_log2_lookup[bsize];
+    const int blk_row_in_sb = (mi_row & MAX_MIB_MASK) >> bh_in_mi_log2;
+    const int blk_col_in_sb = (mi_col & MAX_MIB_MASK) >> bw_in_mi_log2;
+
+    // Top row of superblock: so top-right pixels are in the top and/or
+    // top-right superblocks, both of which are already available.
+    if (blk_row_in_sb == 0) return 1;
+
+    // Rightmost column of superblock (and not the top row): so top-right pixels
+    // fall in the right superblock, which is not available yet.
+    if (((blk_col_in_sb + 1) << bw_in_mi_log2) >= MAX_MIB_SIZE) return 0;
+
+    // General case (neither top row nor rightmost column): check if the
+    // top-right block is coded before the current block.
+    const uint16_t *const order =
+#if CONFIG_EXT_PARTITION_TYPES
+        (partition == PARTITION_VERT_A) ? orders_verta[bsize] :
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                                        orders[bsize];
+    const int this_blk_index =
+        ((blk_row_in_sb + 0) << (MAX_MIB_SIZE_LOG2 - bw_in_mi_log2)) +
+        blk_col_in_sb + 0;
+    const uint16_t this_blk_order = order[this_blk_index];
+    const int tr_blk_index =
+        ((blk_row_in_sb - 1) << (MAX_MIB_SIZE_LOG2 - bw_in_mi_log2)) +
+        blk_col_in_sb + 1;
+    const uint16_t tr_blk_order = order[tr_blk_index];
+    return tr_blk_order < this_blk_order;
+  }
+}
+
+static int has_bottom_left(BLOCK_SIZE bsize, int mi_row, int mi_col,
+                           int bottom_available, int left_available,
+                           TX_SIZE txsz, int row_off, int col_off, int ss_y) {
+  if (!bottom_available || !left_available) return 0;
+
+  if (col_off > 0) {
+    // Bottom-left pixels are in the bottom-left block, which is not available.
+    return 0;
+  } else {
+    const int bh_unit = block_size_high[bsize] >> tx_size_high_log2[0];
+    const int plane_bh_unit = AOMMAX(bh_unit >> ss_y, 1);
+    const int bottom_left_count_unit = tx_size_high_unit[txsz];
+
+#if !CONFIG_CB4X4
+    // Special handling for block sizes 8x4 and 4x4.
+    if (ss_y == 0 && bh_unit < 2 && row_off == 0) return 1;
+#endif
+
+    // All bottom-left pixels are in the left block, which is already available.
+    if (row_off + bottom_left_count_unit < plane_bh_unit) return 1;
+
+    const int bw_in_mi_log2 = mi_width_log2_lookup[bsize];
+    const int bh_in_mi_log2 = mi_height_log2_lookup[bsize];
+    const int blk_row_in_sb = (mi_row & MAX_MIB_MASK) >> bh_in_mi_log2;
+    const int blk_col_in_sb = (mi_col & MAX_MIB_MASK) >> bw_in_mi_log2;
+
+    // Leftmost column of superblock: so bottom-left pixels maybe in the left
+    // and/or bottom-left superblocks. But only the left superblock is
+    // available, so check if all required pixels fall in that superblock.
+    if (blk_col_in_sb == 0) {
+      const int blk_start_row_off = blk_row_in_sb << (bh_in_mi_log2 + !ss_y);
+      const int row_off_in_sb = blk_start_row_off + row_off;
+      const int sb_height_unit = MAX_MIB_SIZE << !ss_y;
+      return row_off_in_sb + bottom_left_count_unit < sb_height_unit;
+    }
+
+    // Bottom row of superblock (and not the leftmost column): so bottom-left
+    // pixels fall in the bottom superblock, which is not available yet.
+    if (((blk_row_in_sb + 1) << bh_in_mi_log2) >= MAX_MIB_SIZE) return 0;
+
+    // General case (neither leftmost column nor bottom row): check if the
+    // bottom-left block is coded before the current block.
+    const uint16_t *const order = orders[bsize];
+    const int this_blk_index =
+        ((blk_row_in_sb + 0) << (MAX_MIB_SIZE_LOG2 - bw_in_mi_log2)) +
+        blk_col_in_sb + 0;
+    const uint16_t this_blk_order = order[this_blk_index];
+    const int bl_blk_index =
+        ((blk_row_in_sb + 1) << (MAX_MIB_SIZE_LOG2 - bw_in_mi_log2)) +
+        blk_col_in_sb - 1;
+    const uint16_t bl_blk_order = order[bl_blk_index];
+    return bl_blk_order < this_blk_order;
+  }
+}
+
+typedef void (*intra_pred_fn)(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left);
+
+static intra_pred_fn pred[INTRA_MODES][TX_SIZES];
+static intra_pred_fn dc_pred[2][2][TX_SIZES];
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*intra_high_pred_fn)(uint16_t *dst, ptrdiff_t stride,
+                                   const uint16_t *above, const uint16_t *left,
+                                   int bd);
+static intra_high_pred_fn pred_high[INTRA_MODES][TX_SIZES];
+static intra_high_pred_fn dc_pred_high[2][2][TX_SIZES];
+#endif  // CONFIG_HIGHBITDEPTH
+
+static void av1_init_intra_predictors_internal(void) {
+#if CONFIG_TX64X64
+#define INIT_NO_4X4(p, type)                  \
+  p[TX_8X8] = aom_##type##_predictor_8x8;     \
+  p[TX_16X16] = aom_##type##_predictor_16x16; \
+  p[TX_32X32] = aom_##type##_predictor_32x32; \
+  p[TX_64X64] = aom_##type##_predictor_64x64
+#else
+#define INIT_NO_4X4(p, type)                  \
+  p[TX_8X8] = aom_##type##_predictor_8x8;     \
+  p[TX_16X16] = aom_##type##_predictor_16x16; \
+  p[TX_32X32] = aom_##type##_predictor_32x32
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_CB4X4
+#define INIT_ALL_SIZES(p, type)           \
+  p[TX_2X2] = aom_##type##_predictor_2x2; \
+  p[TX_4X4] = aom_##type##_predictor_4x4; \
+  INIT_NO_4X4(p, type)
+#else
+#define INIT_ALL_SIZES(p, type)           \
+  p[TX_4X4] = aom_##type##_predictor_4x4; \
+  INIT_NO_4X4(p, type)
+#endif
+
+  INIT_ALL_SIZES(pred[V_PRED], v);
+  INIT_ALL_SIZES(pred[H_PRED], h);
+  INIT_ALL_SIZES(pred[D207_PRED], d207e);
+  INIT_ALL_SIZES(pred[D45_PRED], d45e);
+  INIT_ALL_SIZES(pred[D63_PRED], d63e);
+  INIT_ALL_SIZES(pred[D117_PRED], d117);
+  INIT_ALL_SIZES(pred[D135_PRED], d135);
+  INIT_ALL_SIZES(pred[D153_PRED], d153);
+
+#if CONFIG_ALT_INTRA
+  INIT_ALL_SIZES(pred[TM_PRED], paeth);
+  INIT_ALL_SIZES(pred[SMOOTH_PRED], smooth);
+#else
+  INIT_ALL_SIZES(pred[TM_PRED], tm);
+#endif  // CONFIG_ALT_INTRA
+
+  INIT_ALL_SIZES(dc_pred[0][0], dc_128);
+  INIT_ALL_SIZES(dc_pred[0][1], dc_top);
+  INIT_ALL_SIZES(dc_pred[1][0], dc_left);
+  INIT_ALL_SIZES(dc_pred[1][1], dc);
+
+#if CONFIG_HIGHBITDEPTH
+  INIT_ALL_SIZES(pred_high[V_PRED], highbd_v);
+  INIT_ALL_SIZES(pred_high[H_PRED], highbd_h);
+  INIT_ALL_SIZES(pred_high[D207_PRED], highbd_d207e);
+  INIT_ALL_SIZES(pred_high[D45_PRED], highbd_d45e);
+  INIT_ALL_SIZES(pred_high[D63_PRED], highbd_d63e);
+  INIT_ALL_SIZES(pred_high[D117_PRED], highbd_d117);
+  INIT_ALL_SIZES(pred_high[D135_PRED], highbd_d135);
+  INIT_ALL_SIZES(pred_high[D153_PRED], highbd_d153);
+
+#if CONFIG_ALT_INTRA
+  INIT_ALL_SIZES(pred_high[TM_PRED], highbd_paeth);
+  INIT_ALL_SIZES(pred_high[SMOOTH_PRED], highbd_smooth);
+#else
+  INIT_ALL_SIZES(pred_high[TM_PRED], highbd_tm);
+#endif  // CONFIG_ALT_INTRA
+
+  INIT_ALL_SIZES(dc_pred_high[0][0], highbd_dc_128);
+  INIT_ALL_SIZES(dc_pred_high[0][1], highbd_dc_top);
+  INIT_ALL_SIZES(dc_pred_high[1][0], highbd_dc_left);
+  INIT_ALL_SIZES(dc_pred_high[1][1], highbd_dc);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#undef intra_pred_allsizes
+}
+
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+static int intra_subpel_interp(int base, int shift, const uint8_t *ref,
+                               int ref_start_idx, int ref_end_idx,
+                               INTRA_FILTER filter_type) {
+  int val, k, idx, filter_idx = 0;
+  const int16_t *filter = NULL;
+
+  if (filter_type == INTRA_FILTER_LINEAR) {
+    val = ref[base] * (256 - shift) + ref[base + 1] * shift;
+    val = ROUND_POWER_OF_TWO(val, 8);
+  } else {
+    filter_idx = ROUND_POWER_OF_TWO(shift, 8 - SUBPEL_BITS);
+    filter = av1_intra_filter_kernels[filter_type][filter_idx];
+
+    if (filter_idx < (1 << SUBPEL_BITS)) {
+      val = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) {
+        idx = base + 1 - (SUBPEL_TAPS / 2) + k;
+        idx = AOMMAX(AOMMIN(idx, ref_end_idx), ref_start_idx);
+        val += ref[idx] * filter[k];
+      }
+      val = ROUND_POWER_OF_TWO(val, FILTER_BITS);
+    } else {
+      val = ref[base + 1];
+    }
+  }
+
+  return val;
+}
+#endif  // CONFIG_INTRA_INTERP
+
+// Directional prediction, zone 1: 0 < angle < 90
+static void dr_prediction_z1(uint8_t *dst, ptrdiff_t stride, int bs,
+                             const uint8_t *above, const uint8_t *left,
+#if CONFIG_INTRA_INTERP
+                             INTRA_FILTER filter_type,
+#endif  // CONFIG_INTRA_INTERP
+                             int dx, int dy) {
+  int r, c, x, base, shift, val;
+
+  (void)left;
+  (void)dy;
+  assert(dy == 1);
+  assert(dx > 0);
+
+#if CONFIG_INTRA_INTERP
+  if (filter_type != INTRA_FILTER_LINEAR) {
+    const int pad_size = SUBPEL_TAPS >> 1;
+    int len;
+    DECLARE_ALIGNED(16, uint8_t, buf[SUBPEL_SHIFTS][MAX_SB_SIZE]);
+    DECLARE_ALIGNED(16, uint8_t, src[MAX_SB_SIZE + SUBPEL_TAPS]);
+    uint8_t flags[SUBPEL_SHIFTS];
+
+    memset(flags, 0, SUBPEL_SHIFTS * sizeof(flags[0]));
+    memset(src, above[0], pad_size * sizeof(above[0]));
+    memcpy(src + pad_size, above, 2 * bs * sizeof(above[0]));
+    memset(src + pad_size + 2 * bs, above[2 * bs - 1],
+           pad_size * sizeof(above[0]));
+    flags[0] = 1;
+    x = dx;
+    for (r = 0; r < bs; ++r, dst += stride, x += dx) {
+      base = x >> 8;
+      shift = x & 0xFF;
+      shift = ROUND_POWER_OF_TWO(shift, 8 - SUBPEL_BITS);
+      if (shift == SUBPEL_SHIFTS) {
+        base += 1;
+        shift = 0;
+      }
+      len = AOMMIN(bs, 2 * bs - 1 - base);
+      if (len <= 0) {
+        int i;
+        for (i = r; i < bs; ++i) {
+          memset(dst, above[2 * bs - 1], bs * sizeof(dst[0]));
+          dst += stride;
+        }
+        return;
+      }
+
+      if (len <= (bs >> 1) && !flags[shift]) {
+        base = x >> 8;
+        shift = x & 0xFF;
+        for (c = 0; c < len; ++c) {
+          val = intra_subpel_interp(base, shift, above, 0, 2 * bs - 1,
+                                    filter_type);
+          dst[c] = clip_pixel(val);
+          ++base;
+        }
+      } else {
+        if (!flags[shift]) {
+          const int16_t *filter = av1_intra_filter_kernels[filter_type][shift];
+          aom_convolve8_horiz(src + pad_size, 2 * bs, buf[shift], 2 * bs,
+                              filter, 16, NULL, 16, 2 * bs,
+                              2 * bs < 16 ? 2 : 1);
+          flags[shift] = 1;
+        }
+        memcpy(dst, shift == 0 ? src + pad_size + base : &buf[shift][base],
+               len * sizeof(dst[0]));
+      }
+
+      if (len < bs)
+        memset(dst + len, above[2 * bs - 1], (bs - len) * sizeof(dst[0]));
+    }
+    return;
+  }
+#endif  // CONFIG_INTRA_INTERP
+
+  x = dx;
+  for (r = 0; r < bs; ++r, dst += stride, x += dx) {
+    base = x >> 8;
+    shift = x & 0xFF;
+
+    if (base >= 2 * bs - 1) {
+      int i;
+      for (i = r; i < bs; ++i) {
+        memset(dst, above[2 * bs - 1], bs * sizeof(dst[0]));
+        dst += stride;
+      }
+      return;
+    }
+
+    for (c = 0; c < bs; ++c, ++base) {
+      if (base < 2 * bs - 1) {
+        val = above[base] * (256 - shift) + above[base + 1] * shift;
+        val = ROUND_POWER_OF_TWO(val, 8);
+        dst[c] = clip_pixel(val);
+      } else {
+        dst[c] = above[2 * bs - 1];
+      }
+    }
+  }
+}
+
+// Directional prediction, zone 2: 90 < angle < 180
+static void dr_prediction_z2(uint8_t *dst, ptrdiff_t stride, int bs,
+                             const uint8_t *above, const uint8_t *left,
+#if CONFIG_INTRA_INTERP
+                             INTRA_FILTER filter_type,
+#endif  // CONFIG_INTRA_INTERP
+                             int dx, int dy) {
+  int r, c, x, y, shift1, shift2, val, base1, base2;
+
+  assert(dx > 0);
+  assert(dy > 0);
+
+  x = -dx;
+  for (r = 0; r < bs; ++r, x -= dx, dst += stride) {
+    base1 = x >> 8;
+    y = (r << 8) - dy;
+    for (c = 0; c < bs; ++c, ++base1, y -= dy) {
+      if (base1 >= -1) {
+        shift1 = x & 0xFF;
+#if CONFIG_INTRA_INTERP
+        val =
+            intra_subpel_interp(base1, shift1, above, -1, bs - 1, filter_type);
+#else
+        val = above[base1] * (256 - shift1) + above[base1 + 1] * shift1;
+        val = ROUND_POWER_OF_TWO(val, 8);
+#endif  // CONFIG_INTRA_INTERP
+      } else {
+        base2 = y >> 8;
+        shift2 = y & 0xFF;
+#if CONFIG_INTRA_INTERP
+        val = intra_subpel_interp(base2, shift2, left, -1, bs - 1, filter_type);
+#else
+        val = left[base2] * (256 - shift2) + left[base2 + 1] * shift2;
+        val = ROUND_POWER_OF_TWO(val, 8);
+#endif  // CONFIG_INTRA_INTERP
+      }
+      dst[c] = clip_pixel(val);
+    }
+  }
+}
+
+// Directional prediction, zone 3: 180 < angle < 270
+static void dr_prediction_z3(uint8_t *dst, ptrdiff_t stride, int bs,
+                             const uint8_t *above, const uint8_t *left,
+#if CONFIG_INTRA_INTERP
+                             INTRA_FILTER filter_type,
+#endif  // CONFIG_INTRA_INTERP
+                             int dx, int dy) {
+  int r, c, y, base, shift, val;
+
+  (void)above;
+  (void)dx;
+
+  assert(dx == 1);
+  assert(dy > 0);
+
+#if CONFIG_INTRA_INTERP
+  if (filter_type != INTRA_FILTER_LINEAR) {
+    const int pad_size = SUBPEL_TAPS >> 1;
+    int len, i;
+    DECLARE_ALIGNED(16, uint8_t, buf[MAX_SB_SIZE][4 * SUBPEL_SHIFTS]);
+    DECLARE_ALIGNED(16, uint8_t, src[(MAX_SB_SIZE + SUBPEL_TAPS) * 4]);
+    uint8_t flags[SUBPEL_SHIFTS];
+
+    memset(flags, 0, SUBPEL_SHIFTS * sizeof(flags[0]));
+    for (i = 0; i < pad_size; ++i) src[4 * i] = left[0];
+    for (i = 0; i < 2 * bs; ++i) src[4 * (i + pad_size)] = left[i];
+    for (i = 0; i < pad_size; ++i)
+      src[4 * (i + 2 * bs + pad_size)] = left[2 * bs - 1];
+    flags[0] = 1;
+    y = dy;
+    for (c = 0; c < bs; ++c, y += dy) {
+      base = y >> 8;
+      shift = y & 0xFF;
+      shift = ROUND_POWER_OF_TWO(shift, 8 - SUBPEL_BITS);
+      if (shift == SUBPEL_SHIFTS) {
+        base += 1;
+        shift = 0;
+      }
+      len = AOMMIN(bs, 2 * bs - 1 - base);
+
+      if (len <= 0) {
+        for (r = 0; r < bs; ++r) {
+          dst[r * stride + c] = left[2 * bs - 1];
+        }
+        continue;
+      }
+
+      if (len <= (bs >> 1) && !flags[shift]) {
+        base = y >> 8;
+        shift = y & 0xFF;
+        for (r = 0; r < len; ++r) {
+          val = intra_subpel_interp(base, shift, left, 0, 2 * bs - 1,
+                                    filter_type);
+          dst[r * stride + c] = clip_pixel(val);
+          ++base;
+        }
+      } else {
+        if (!flags[shift]) {
+          const int16_t *filter = av1_intra_filter_kernels[filter_type][shift];
+          aom_convolve8_vert(src + 4 * pad_size, 4, buf[0] + 4 * shift,
+                             4 * SUBPEL_SHIFTS, NULL, 16, filter, 16,
+                             2 * bs < 16 ? 4 : 4, 2 * bs);
+          flags[shift] = 1;
+        }
+
+        if (shift == 0) {
+          for (r = 0; r < len; ++r) {
+            dst[r * stride + c] = left[r + base];
+          }
+        } else {
+          for (r = 0; r < len; ++r) {
+            dst[r * stride + c] = buf[r + base][4 * shift];
+          }
+        }
+      }
+
+      if (len < bs) {
+        for (r = len; r < bs; ++r) {
+          dst[r * stride + c] = left[2 * bs - 1];
+        }
+      }
+    }
+    return;
+  }
+#endif  // CONFIG_INTRA_INTERP
+
+  y = dy;
+  for (c = 0; c < bs; ++c, y += dy) {
+    base = y >> 8;
+    shift = y & 0xFF;
+
+    for (r = 0; r < bs; ++r, ++base) {
+      if (base < 2 * bs - 1) {
+        val = left[base] * (256 - shift) + left[base + 1] * shift;
+        val = ROUND_POWER_OF_TWO(val, 8);
+        dst[r * stride + c] = clip_pixel(val);
+      } else {
+        for (; r < bs; ++r) dst[r * stride + c] = left[2 * bs - 1];
+        break;
+      }
+    }
+  }
+}
+
+// Get the shift (up-scaled by 256) in X w.r.t a unit change in Y.
+// If angle > 0 && angle < 90, dx = -((int)(256 / t));
+// If angle > 90 && angle < 180, dx = (int)(256 / t);
+// If angle > 180 && angle < 270, dx = 1;
+static INLINE int get_dx(int angle) {
+  if (angle > 0 && angle < 90) {
+    return dr_intra_derivative[angle];
+  } else if (angle > 90 && angle < 180) {
+    return dr_intra_derivative[180 - angle];
+  } else {
+    // In this case, we are not really going to use dx. We may return any value.
+    return 1;
+  }
+}
+
+// Get the shift (up-scaled by 256) in Y w.r.t a unit change in X.
+// If angle > 0 && angle < 90, dy = 1;
+// If angle > 90 && angle < 180, dy = (int)(256 * t);
+// If angle > 180 && angle < 270, dy = -((int)(256 * t));
+static INLINE int get_dy(int angle) {
+  if (angle > 90 && angle < 180) {
+    return dr_intra_derivative[angle - 90];
+  } else if (angle > 180 && angle < 270) {
+    return dr_intra_derivative[270 - angle];
+  } else {
+    // In this case, we are not really going to use dy. We may return any value.
+    return 1;
+  }
+}
+
+static void dr_predictor(uint8_t *dst, ptrdiff_t stride, TX_SIZE tx_size,
+                         const uint8_t *above, const uint8_t *left,
+#if CONFIG_INTRA_INTERP
+                         INTRA_FILTER filter_type,
+#endif  // CONFIG_INTRA_INTERP
+                         int angle) {
+  const int dx = get_dx(angle);
+  const int dy = get_dy(angle);
+  const int bs = tx_size_wide[tx_size];
+  assert(angle > 0 && angle < 270);
+
+  if (angle > 0 && angle < 90) {
+    dr_prediction_z1(dst, stride, bs, above, left,
+#if CONFIG_INTRA_INTERP
+                     filter_type,
+#endif  // CONFIG_INTRA_INTERP
+                     dx, dy);
+  } else if (angle > 90 && angle < 180) {
+    dr_prediction_z2(dst, stride, bs, above, left,
+#if CONFIG_INTRA_INTERP
+                     filter_type,
+#endif  // CONFIG_INTRA_INTERP
+                     dx, dy);
+  } else if (angle > 180 && angle < 270) {
+    dr_prediction_z3(dst, stride, bs, above, left,
+#if CONFIG_INTRA_INTERP
+                     filter_type,
+#endif  // CONFIG_INTRA_INTERP
+                     dx, dy);
+  } else if (angle == 90) {
+    pred[V_PRED][tx_size](dst, stride, above, left);
+  } else if (angle == 180) {
+    pred[H_PRED][tx_size](dst, stride, above, left);
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_INTRA_INTERP
+static int highbd_intra_subpel_interp(int base, int shift, const uint16_t *ref,
+                                      int ref_start_idx, int ref_end_idx,
+                                      INTRA_FILTER filter_type) {
+  int val, k, idx, filter_idx = 0;
+  const int16_t *filter = NULL;
+
+  if (filter_type == INTRA_FILTER_LINEAR) {
+    val = ref[base] * (256 - shift) + ref[base + 1] * shift;
+    val = ROUND_POWER_OF_TWO(val, 8);
+  } else {
+    filter_idx = ROUND_POWER_OF_TWO(shift, 8 - SUBPEL_BITS);
+    filter = av1_intra_filter_kernels[filter_type][filter_idx];
+
+    if (filter_idx < (1 << SUBPEL_BITS)) {
+      val = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k) {
+        idx = base + 1 - (SUBPEL_TAPS / 2) + k;
+        idx = AOMMAX(AOMMIN(idx, ref_end_idx), ref_start_idx);
+        val += ref[idx] * filter[k];
+      }
+      val = ROUND_POWER_OF_TWO(val, FILTER_BITS);
+    } else {
+      val = ref[base + 1];
+    }
+  }
+
+  return val;
+}
+#endif  // CONFIG_INTRA_INTERP
+
+// Directional prediction, zone 1: 0 < angle < 90
+static void highbd_dr_prediction_z1(uint16_t *dst, ptrdiff_t stride, int bs,
+                                    const uint16_t *above, const uint16_t *left,
+#if CONFIG_INTRA_INTERP
+                                    INTRA_FILTER filter_type,
+#endif  // CONFIG_INTRA_INTERP
+                                    int dx, int dy, int bd) {
+  int r, c, x, base, shift, val;
+
+  (void)left;
+  (void)dy;
+  assert(dy == 1);
+  assert(dx > 0);
+
+  x = dx;
+  for (r = 0; r < bs; ++r, dst += stride, x += dx) {
+    base = x >> 8;
+    shift = x & 0xFF;
+
+    if (base >= 2 * bs - 1) {
+      int i;
+      for (i = r; i < bs; ++i) {
+        aom_memset16(dst, above[2 * bs - 1], bs);
+        dst += stride;
+      }
+      return;
+    }
+
+    for (c = 0; c < bs; ++c, ++base) {
+      if (base < 2 * bs - 1) {
+#if CONFIG_INTRA_INTERP
+        val = highbd_intra_subpel_interp(base, shift, above, 0, 2 * bs - 1,
+                                         filter_type);
+#else
+        val = above[base] * (256 - shift) + above[base + 1] * shift;
+        val = ROUND_POWER_OF_TWO(val, 8);
+#endif  // CONFIG_INTRA_INTERP
+        dst[c] = clip_pixel_highbd(val, bd);
+      } else {
+        dst[c] = above[2 * bs - 1];
+      }
+    }
+  }
+}
+
+// Directional prediction, zone 2: 90 < angle < 180
+static void highbd_dr_prediction_z2(uint16_t *dst, ptrdiff_t stride, int bs,
+                                    const uint16_t *above, const uint16_t *left,
+#if CONFIG_INTRA_INTERP
+                                    INTRA_FILTER filter_type,
+#endif  // CONFIG_INTRA_INTERP
+                                    int dx, int dy, int bd) {
+  int r, c, x, y, shift, val, base;
+
+  assert(dx > 0);
+  assert(dy > 0);
+
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      y = r + 1;
+      x = (c << 8) - y * dx;
+      base = x >> 8;
+      if (base >= -1) {
+        shift = x & 0xFF;
+#if CONFIG_INTRA_INTERP
+        val = highbd_intra_subpel_interp(base, shift, above, -1, bs - 1,
+                                         filter_type);
+#else
+        val = above[base] * (256 - shift) + above[base + 1] * shift;
+        val = ROUND_POWER_OF_TWO(val, 8);
+#endif  // CONFIG_INTRA_INTERP
+      } else {
+        x = c + 1;
+        y = (r << 8) - x * dy;
+        base = y >> 8;
+        shift = y & 0xFF;
+#if CONFIG_INTRA_INTERP
+        val = highbd_intra_subpel_interp(base, shift, left, -1, bs - 1,
+                                         filter_type);
+#else
+        val = left[base] * (256 - shift) + left[base + 1] * shift;
+        val = ROUND_POWER_OF_TWO(val, 8);
+#endif  // CONFIG_INTRA_INTERP
+      }
+      dst[c] = clip_pixel_highbd(val, bd);
+    }
+    dst += stride;
+  }
+}
+
+// Directional prediction, zone 3: 180 < angle < 270
+static void highbd_dr_prediction_z3(uint16_t *dst, ptrdiff_t stride, int bs,
+                                    const uint16_t *above, const uint16_t *left,
+#if CONFIG_INTRA_INTERP
+                                    INTRA_FILTER filter_type,
+#endif  // CONFIG_INTRA_INTERP
+                                    int dx, int dy, int bd) {
+  int r, c, y, base, shift, val;
+
+  (void)above;
+  (void)dx;
+  assert(dx == 1);
+  assert(dy > 0);
+
+  y = dy;
+  for (c = 0; c < bs; ++c, y += dy) {
+    base = y >> 8;
+    shift = y & 0xFF;
+
+    for (r = 0; r < bs; ++r, ++base) {
+      if (base < 2 * bs - 1) {
+#if CONFIG_INTRA_INTERP
+        val = highbd_intra_subpel_interp(base, shift, left, 0, 2 * bs - 1,
+                                         filter_type);
+#else
+        val = left[base] * (256 - shift) + left[base + 1] * shift;
+        val = ROUND_POWER_OF_TWO(val, 8);
+#endif  // CONFIG_INTRA_INTERP
+        dst[r * stride + c] = clip_pixel_highbd(val, bd);
+      } else {
+        for (; r < bs; ++r) dst[r * stride + c] = left[2 * bs - 1];
+        break;
+      }
+    }
+  }
+}
+
+static INLINE void highbd_v_predictor(uint16_t *dst, ptrdiff_t stride, int bs,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  int r;
+  (void)left;
+  (void)bd;
+  for (r = 0; r < bs; r++) {
+    memcpy(dst, above, bs * sizeof(uint16_t));
+    dst += stride;
+  }
+}
+
+static INLINE void highbd_h_predictor(uint16_t *dst, ptrdiff_t stride, int bs,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  int r;
+  (void)above;
+  (void)bd;
+  for (r = 0; r < bs; r++) {
+    aom_memset16(dst, left[r], bs);
+    dst += stride;
+  }
+}
+
+static void highbd_dr_predictor(uint16_t *dst, ptrdiff_t stride, int bs,
+                                const uint16_t *above, const uint16_t *left,
+#if CONFIG_INTRA_INTERP
+                                INTRA_FILTER filter,
+#endif  // CONFIG_INTRA_INTERP
+                                int angle, int bd) {
+  const int dx = get_dx(angle);
+  const int dy = get_dy(angle);
+  assert(angle > 0 && angle < 270);
+
+  if (angle > 0 && angle < 90) {
+    highbd_dr_prediction_z1(dst, stride, bs, above, left,
+#if CONFIG_INTRA_INTERP
+                            filter,
+#endif  // CONFIG_INTRA_INTERP
+                            dx, dy, bd);
+  } else if (angle > 90 && angle < 180) {
+    highbd_dr_prediction_z2(dst, stride, bs, above, left,
+#if CONFIG_INTRA_INTERP
+                            filter,
+#endif  // CONFIG_INTRA_INTERP
+                            dx, dy, bd);
+  } else if (angle > 180 && angle < 270) {
+    highbd_dr_prediction_z3(dst, stride, bs, above, left,
+#if CONFIG_INTRA_INTERP
+                            filter,
+#endif  // CONFIG_INTRA_INTERP
+                            dx, dy, bd);
+  } else if (angle == 90) {
+    highbd_v_predictor(dst, stride, bs, above, left, bd);
+  } else if (angle == 180) {
+    highbd_h_predictor(dst, stride, bs, above, left, bd);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_EXT_INTRA
+
+#if CONFIG_FILTER_INTRA
+#if USE_3TAP_INTRA_FILTER
+int av1_filter_intra_taps_3[TX_SIZES][INTRA_MODES][3] = {
+#if CONFIG_CB4X4
+  {
+      { 697, 836, -509 },
+      { 993, 513, -482 },
+      { 381, 984, -341 },
+      { 642, 1169, -787 },
+      { 590, 553, -119 },
+      { 762, 385, -123 },
+      { 358, 687, -21 },
+      { 411, 1083, -470 },
+      { 912, 814, -702 },
+      { 883, 902, 761 },
+  },
+#endif
+  {
+      { 697, 836, -509 },
+      { 993, 513, -482 },
+      { 381, 984, -341 },
+      { 642, 1169, -787 },
+      { 590, 553, -119 },
+      { 762, 385, -123 },
+      { 358, 687, -21 },
+      { 411, 1083, -470 },
+      { 912, 814, -702 },
+      { 883, 902, 761 },
+  },
+  {
+      { 659, 816, -451 },
+      { 980, 625, -581 },
+      { 558, 962, -496 },
+      { 681, 888, -545 },
+      { 591, 613, 180 },
+      { 778, 399, -153 },
+      { 495, 641, -112 },
+      { 671, 937, -584 },
+      { 745, 940, -661 },
+      { 839, 911, -726 },
+  },
+  {
+      { 539, 927, -442 },
+      { 1003, 714, -693 },
+      { 349, 1271, -596 },
+      { 820, 764, -560 },
+      { 524, 816, -316 },
+      { 780, 681, -437 },
+      { 586, 795, -357 },
+      { 551, 1135, -663 },
+      { 593, 1061, -630 },
+      { 974, 970, -920 },
+  },
+  {
+      { 595, 919, -490 },
+      { 945, 668, -579 },
+      { 495, 962, -433 },
+      { 385, 1551, -912 },
+      { 455, 554, 15 },
+      { 852, 478, -306 },
+      { 177, 760, -87 },
+      { -65, 1611, -522 },
+      { 815, 894, -685 },
+      { 846, 1010, -832 },
+  },
+#if CONFIG_TX64X64
+  {
+      { 595, 919, -490 },
+      { 945, 668, -579 },
+      { 495, 962, -433 },
+      { 385, 1551, -912 },
+      { 455, 554, 15 },
+      { 852, 478, -306 },
+      { 177, 760, -87 },
+      { -65, 1611, -522 },
+      { 815, 894, -685 },
+      { 846, 1010, -832 },
+  },
+#endif  // CONFIG_TX64X64
+};
+#else
+int av1_filter_intra_taps_4[TX_SIZES][INTRA_MODES][4] = {
+#if CONFIG_CB4X4
+  {
+      { 735, 881, -537, -54 },
+      { 1005, 519, -488, -11 },
+      { 383, 990, -343, -6 },
+      { 442, 805, -542, 319 },
+      { 658, 616, -133, -116 },
+      { 875, 442, -141, -151 },
+      { 386, 741, -23, -80 },
+      { 390, 1027, -446, 51 },
+      { 679, 606, -523, 262 },
+      { 903, 922, -778, -23 },
+  },
+#endif
+  {
+      { 735, 881, -537, -54 },
+      { 1005, 519, -488, -11 },
+      { 383, 990, -343, -6 },
+      { 442, 805, -542, 319 },
+      { 658, 616, -133, -116 },
+      { 875, 442, -141, -151 },
+      { 386, 741, -23, -80 },
+      { 390, 1027, -446, 51 },
+      { 679, 606, -523, 262 },
+      { 903, 922, -778, -23 },
+  },
+  {
+      { 648, 803, -444, 16 },
+      { 972, 620, -576, 7 },
+      { 561, 967, -499, -5 },
+      { 585, 762, -468, 144 },
+      { 596, 619, -182, -9 },
+      { 895, 459, -176, -153 },
+      { 557, 722, -126, -129 },
+      { 601, 839, -523, 105 },
+      { 562, 709, -499, 251 },
+      { 803, 872, -695, 43 },
+  },
+  {
+      { 423, 728, -347, 111 },
+      { 963, 685, -665, 23 },
+      { 281, 1024, -480, 216 },
+      { 640, 596, -437, 78 },
+      { 429, 669, -259, 99 },
+      { 740, 646, -415, 23 },
+      { 568, 771, -346, 40 },
+      { 404, 833, -486, 209 },
+      { 398, 712, -423, 307 },
+      { 939, 935, -887, 17 },
+  },
+  {
+      { 477, 737, -393, 150 },
+      { 881, 630, -546, 67 },
+      { 506, 984, -443, -20 },
+      { 114, 459, -270, 528 },
+      { 433, 528, 14, 3 },
+      { 837, 470, -301, -30 },
+      { 181, 777, 89, -107 },
+      { -29, 716, -232, 259 },
+      { 589, 646, -495, 255 },
+      { 740, 884, -728, 77 },
+  },
+#if CONFIG_TX64X64
+  {
+      { 477, 737, -393, 150 },
+      { 881, 630, -546, 67 },
+      { 506, 984, -443, -20 },
+      { 114, 459, -270, 528 },
+      { 433, 528, 14, 3 },
+      { 837, 470, -301, -30 },
+      { 181, 777, 89, -107 },
+      { -29, 716, -232, 259 },
+      { 589, 646, -495, 255 },
+      { 740, 884, -728, 77 },
+  },
+#endif  // CONFIG_TX64X64
+};
+#endif
+
+static INLINE TX_SIZE get_txsize_from_blocklen(int bs) {
+  switch (bs) {
+    case 4: return TX_4X4;
+    case 8: return TX_8X8;
+    case 16: return TX_16X16;
+    case 32: return TX_32X32;
+#if CONFIG_TX64X64
+    case 64: return TX_64X64;
+#endif  // CONFIG_TX64X64
+    default: assert(0); return TX_INVALID;
+  }
+}
+
+#if USE_3TAP_INTRA_FILTER
+static void filter_intra_predictors_3tap(uint8_t *dst, ptrdiff_t stride, int bs,
+                                         const uint8_t *above,
+                                         const uint8_t *left, int mode) {
+  int k, r, c;
+  int mean, ipred;
+#if CONFIG_TX64X64
+  int buffer[65][65];
+#else
+  int buffer[33][33];
+#endif  // CONFIG_TX64X64
+  const TX_SIZE tx_size = get_txsize_from_blocklen(bs);
+  const int c0 = av1_filter_intra_taps_3[tx_size][mode][0];
+  const int c1 = av1_filter_intra_taps_3[tx_size][mode][1];
+  const int c2 = av1_filter_intra_taps_3[tx_size][mode][2];
+
+  k = 0;
+  mean = 0;
+  while (k < bs) {
+    mean = mean + (int)left[k];
+    mean = mean + (int)above[k];
+    k++;
+  }
+  mean = (mean + bs) / (2 * bs);
+
+  for (r = 0; r < bs; ++r) buffer[r + 1][0] = (int)left[r] - mean;
+
+  for (c = 0; c < bs + 1; ++c) buffer[0][c] = (int)above[c - 1] - mean;
+
+  for (r = 1; r < bs + 1; ++r)
+    for (c = 1; c < bs + 1; ++c) {
+      ipred = c0 * buffer[r - 1][c] + c1 * buffer[r][c - 1] +
+              c2 * buffer[r - 1][c - 1];
+      buffer[r][c] = ROUND_POWER_OF_TWO_SIGNED(ipred, FILTER_INTRA_PREC_BITS);
+    }
+
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      ipred = buffer[r + 1][c + 1] + mean;
+      dst[c] = clip_pixel(ipred);
+    }
+    dst += stride;
+  }
+}
+#else
+static void filter_intra_predictors_4tap(uint8_t *dst, ptrdiff_t stride, int bs,
+                                         const uint8_t *above,
+                                         const uint8_t *left, int mode) {
+  int k, r, c;
+  int mean, ipred;
+#if CONFIG_TX64X64
+  int buffer[65][129];
+#else
+  int buffer[33][65];
+#endif  // CONFIG_TX64X64
+  const TX_SIZE tx_size = get_txsize_from_blocklen(bs);
+  const int c0 = av1_filter_intra_taps_4[tx_size][mode][0];
+  const int c1 = av1_filter_intra_taps_4[tx_size][mode][1];
+  const int c2 = av1_filter_intra_taps_4[tx_size][mode][2];
+  const int c3 = av1_filter_intra_taps_4[tx_size][mode][3];
+
+  k = 0;
+  mean = 0;
+  while (k < bs) {
+    mean = mean + (int)left[k];
+    mean = mean + (int)above[k];
+    k++;
+  }
+  mean = (mean + bs) / (2 * bs);
+
+  for (r = 0; r < bs; ++r) buffer[r + 1][0] = (int)left[r] - mean;
+
+  for (c = 0; c < 2 * bs + 1; ++c) buffer[0][c] = (int)above[c - 1] - mean;
+
+  for (r = 1; r < bs + 1; ++r)
+    for (c = 1; c < 2 * bs + 1 - r; ++c) {
+      ipred = c0 * buffer[r - 1][c] + c1 * buffer[r][c - 1] +
+              c2 * buffer[r - 1][c - 1] + c3 * buffer[r - 1][c + 1];
+      buffer[r][c] = ROUND_POWER_OF_TWO_SIGNED(ipred, FILTER_INTRA_PREC_BITS);
+    }
+
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      ipred = buffer[r + 1][c + 1] + mean;
+      dst[c] = clip_pixel(ipred);
+    }
+    dst += stride;
+  }
+}
+#endif
+
+void av1_dc_filter_predictor_c(uint8_t *dst, ptrdiff_t stride, int bs,
+                               const uint8_t *above, const uint8_t *left) {
+#if USE_3TAP_INTRA_FILTER
+  filter_intra_predictors_3tap(dst, stride, bs, above, left, DC_PRED);
+#else
+  filter_intra_predictors_4tap(dst, stride, bs, above, left, DC_PRED);
+#endif
+}
+
+void av1_v_filter_predictor_c(uint8_t *dst, ptrdiff_t stride, int bs,
+                              const uint8_t *above, const uint8_t *left) {
+#if USE_3TAP_INTRA_FILTER
+  filter_intra_predictors_3tap(dst, stride, bs, above, left, V_PRED);
+#else
+  filter_intra_predictors_4tap(dst, stride, bs, above, left, V_PRED);
+#endif
+}
+
+void av1_h_filter_predictor_c(uint8_t *dst, ptrdiff_t stride, int bs,
+                              const uint8_t *above, const uint8_t *left) {
+#if USE_3TAP_INTRA_FILTER
+  filter_intra_predictors_3tap(dst, stride, bs, above, left, H_PRED);
+#else
+  filter_intra_predictors_4tap(dst, stride, bs, above, left, H_PRED);
+#endif
+}
+
+void av1_d45_filter_predictor_c(uint8_t *dst, ptrdiff_t stride, int bs,
+                                const uint8_t *above, const uint8_t *left) {
+#if USE_3TAP_INTRA_FILTER
+  filter_intra_predictors_3tap(dst, stride, bs, above, left, D45_PRED);
+#else
+  filter_intra_predictors_4tap(dst, stride, bs, above, left, D45_PRED);
+#endif
+}
+
+void av1_d135_filter_predictor_c(uint8_t *dst, ptrdiff_t stride, int bs,
+                                 const uint8_t *above, const uint8_t *left) {
+#if USE_3TAP_INTRA_FILTER
+  filter_intra_predictors_3tap(dst, stride, bs, above, left, D135_PRED);
+#else
+  filter_intra_predictors_4tap(dst, stride, bs, above, left, D135_PRED);
+#endif
+}
+
+void av1_d117_filter_predictor_c(uint8_t *dst, ptrdiff_t stride, int bs,
+                                 const uint8_t *above, const uint8_t *left) {
+#if USE_3TAP_INTRA_FILTER
+  filter_intra_predictors_3tap(dst, stride, bs, above, left, D117_PRED);
+#else
+  filter_intra_predictors_4tap(dst, stride, bs, above, left, D117_PRED);
+#endif
+}
+
+void av1_d153_filter_predictor_c(uint8_t *dst, ptrdiff_t stride, int bs,
+                                 const uint8_t *above, const uint8_t *left) {
+#if USE_3TAP_INTRA_FILTER
+  filter_intra_predictors_3tap(dst, stride, bs, above, left, D153_PRED);
+#else
+  filter_intra_predictors_4tap(dst, stride, bs, above, left, D153_PRED);
+#endif
+}
+
+void av1_d207_filter_predictor_c(uint8_t *dst, ptrdiff_t stride, int bs,
+                                 const uint8_t *above, const uint8_t *left) {
+#if USE_3TAP_INTRA_FILTER
+  filter_intra_predictors_3tap(dst, stride, bs, above, left, D207_PRED);
+#else
+  filter_intra_predictors_4tap(dst, stride, bs, above, left, D207_PRED);
+#endif
+}
+
+void av1_d63_filter_predictor_c(uint8_t *dst, ptrdiff_t stride, int bs,
+                                const uint8_t *above, const uint8_t *left) {
+#if USE_3TAP_INTRA_FILTER
+  filter_intra_predictors_3tap(dst, stride, bs, above, left, D63_PRED);
+#else
+  filter_intra_predictors_4tap(dst, stride, bs, above, left, D63_PRED);
+#endif
+}
+
+void av1_tm_filter_predictor_c(uint8_t *dst, ptrdiff_t stride, int bs,
+                               const uint8_t *above, const uint8_t *left) {
+#if USE_3TAP_INTRA_FILTER
+  filter_intra_predictors_3tap(dst, stride, bs, above, left, TM_PRED);
+#else
+  filter_intra_predictors_4tap(dst, stride, bs, above, left, TM_PRED);
+#endif
+}
+
+static void filter_intra_predictors(FILTER_INTRA_MODE mode, uint8_t *dst,
+                                    ptrdiff_t stride, int bs,
+                                    const uint8_t *above, const uint8_t *left) {
+  switch (mode) {
+    case FILTER_DC_PRED:
+      av1_dc_filter_predictor(dst, stride, bs, above, left);
+      break;
+    case FILTER_V_PRED:
+      av1_v_filter_predictor(dst, stride, bs, above, left);
+      break;
+    case FILTER_H_PRED:
+      av1_h_filter_predictor(dst, stride, bs, above, left);
+      break;
+    case FILTER_D45_PRED:
+      av1_d45_filter_predictor(dst, stride, bs, above, left);
+      break;
+    case FILTER_D135_PRED:
+      av1_d135_filter_predictor(dst, stride, bs, above, left);
+      break;
+    case FILTER_D117_PRED:
+      av1_d117_filter_predictor(dst, stride, bs, above, left);
+      break;
+    case FILTER_D153_PRED:
+      av1_d153_filter_predictor(dst, stride, bs, above, left);
+      break;
+    case FILTER_D207_PRED:
+      av1_d207_filter_predictor(dst, stride, bs, above, left);
+      break;
+    case FILTER_D63_PRED:
+      av1_d63_filter_predictor(dst, stride, bs, above, left);
+      break;
+    case FILTER_TM_PRED:
+      av1_tm_filter_predictor(dst, stride, bs, above, left);
+      break;
+    default: assert(0);
+  }
+}
+#if CONFIG_HIGHBITDEPTH
+#if USE_3TAP_INTRA_FILTER
+static void highbd_filter_intra_predictors_3tap(uint16_t *dst, ptrdiff_t stride,
+                                                int bs, const uint16_t *above,
+                                                const uint16_t *left, int mode,
+                                                int bd) {
+  int k, r, c;
+  int mean, ipred;
+#if CONFIG_TX64X64
+  int preds[65][65];
+#else
+  int preds[33][33];
+#endif  // CONFIG_TX64X64
+  const TX_SIZE tx_size = get_txsize_from_blocklen(bs);
+  const int c0 = av1_filter_intra_taps_3[tx_size][mode][0];
+  const int c1 = av1_filter_intra_taps_3[tx_size][mode][1];
+  const int c2 = av1_filter_intra_taps_3[tx_size][mode][2];
+
+  k = 0;
+  mean = 0;
+  while (k < bs) {
+    mean = mean + (int)left[k];
+    mean = mean + (int)above[k];
+    k++;
+  }
+  mean = (mean + bs) / (2 * bs);
+
+  for (r = 0; r < bs; ++r) preds[r + 1][0] = (int)left[r] - mean;
+
+  for (c = 0; c < bs + 1; ++c) preds[0][c] = (int)above[c - 1] - mean;
+
+  for (r = 1; r < bs + 1; ++r)
+    for (c = 1; c < bs + 1; ++c) {
+      ipred = c0 * preds[r - 1][c] + c1 * preds[r][c - 1] +
+              c2 * preds[r - 1][c - 1];
+      preds[r][c] = ROUND_POWER_OF_TWO_SIGNED(ipred, FILTER_INTRA_PREC_BITS);
+    }
+
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      ipred = preds[r + 1][c + 1] + mean;
+      dst[c] = clip_pixel_highbd(ipred, bd);
+    }
+    dst += stride;
+  }
+}
+#else
+static void highbd_filter_intra_predictors_4tap(uint16_t *dst, ptrdiff_t stride,
+                                                int bs, const uint16_t *above,
+                                                const uint16_t *left, int mode,
+                                                int bd) {
+  int k, r, c;
+  int mean, ipred;
+#if CONFIG_TX64X64
+  int preds[65][129];
+#else
+  int preds[33][65];
+#endif  // CONFIG_TX64X64
+  const TX_SIZE tx_size = get_txsize_from_blocklen(bs);
+  const int c0 = av1_filter_intra_taps_4[tx_size][mode][0];
+  const int c1 = av1_filter_intra_taps_4[tx_size][mode][1];
+  const int c2 = av1_filter_intra_taps_4[tx_size][mode][2];
+  const int c3 = av1_filter_intra_taps_4[tx_size][mode][3];
+
+  k = 0;
+  mean = 0;
+  while (k < bs) {
+    mean = mean + (int)left[k];
+    mean = mean + (int)above[k];
+    k++;
+  }
+  mean = (mean + bs) / (2 * bs);
+
+  for (r = 0; r < bs; ++r) preds[r + 1][0] = (int)left[r] - mean;
+
+  for (c = 0; c < 2 * bs + 1; ++c) preds[0][c] = (int)above[c - 1] - mean;
+
+  for (r = 1; r < bs + 1; ++r)
+    for (c = 1; c < 2 * bs + 1 - r; ++c) {
+      ipred = c0 * preds[r - 1][c] + c1 * preds[r][c - 1] +
+              c2 * preds[r - 1][c - 1] + c3 * preds[r - 1][c + 1];
+      preds[r][c] = ROUND_POWER_OF_TWO_SIGNED(ipred, FILTER_INTRA_PREC_BITS);
+    }
+
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      ipred = preds[r + 1][c + 1] + mean;
+      dst[c] = clip_pixel_highbd(ipred, bd);
+    }
+    dst += stride;
+  }
+}
+#endif
+
+void av1_highbd_dc_filter_predictor_c(uint16_t *dst, ptrdiff_t stride, int bs,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+#if USE_3TAP_INTRA_FILTER
+  highbd_filter_intra_predictors_3tap(dst, stride, bs, above, left, DC_PRED,
+                                      bd);
+#else
+  highbd_filter_intra_predictors_4tap(dst, stride, bs, above, left, DC_PRED,
+                                      bd);
+#endif
+}
+
+void av1_highbd_v_filter_predictor_c(uint16_t *dst, ptrdiff_t stride, int bs,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+#if USE_3TAP_INTRA_FILTER
+  highbd_filter_intra_predictors_3tap(dst, stride, bs, above, left, V_PRED, bd);
+#else
+  highbd_filter_intra_predictors_4tap(dst, stride, bs, above, left, V_PRED, bd);
+#endif
+}
+
+void av1_highbd_h_filter_predictor_c(uint16_t *dst, ptrdiff_t stride, int bs,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+#if USE_3TAP_INTRA_FILTER
+  highbd_filter_intra_predictors_3tap(dst, stride, bs, above, left, H_PRED, bd);
+#else
+  highbd_filter_intra_predictors_4tap(dst, stride, bs, above, left, H_PRED, bd);
+#endif
+}
+
+void av1_highbd_d45_filter_predictor_c(uint16_t *dst, ptrdiff_t stride, int bs,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+#if USE_3TAP_INTRA_FILTER
+  highbd_filter_intra_predictors_3tap(dst, stride, bs, above, left, D45_PRED,
+                                      bd);
+#else
+  highbd_filter_intra_predictors_4tap(dst, stride, bs, above, left, D45_PRED,
+                                      bd);
+#endif
+}
+
+void av1_highbd_d135_filter_predictor_c(uint16_t *dst, ptrdiff_t stride, int bs,
+                                        const uint16_t *above,
+                                        const uint16_t *left, int bd) {
+#if USE_3TAP_INTRA_FILTER
+  highbd_filter_intra_predictors_3tap(dst, stride, bs, above, left, D135_PRED,
+                                      bd);
+#else
+  highbd_filter_intra_predictors_4tap(dst, stride, bs, above, left, D135_PRED,
+                                      bd);
+#endif
+}
+
+void av1_highbd_d117_filter_predictor_c(uint16_t *dst, ptrdiff_t stride, int bs,
+                                        const uint16_t *above,
+                                        const uint16_t *left, int bd) {
+#if USE_3TAP_INTRA_FILTER
+  highbd_filter_intra_predictors_3tap(dst, stride, bs, above, left, D117_PRED,
+                                      bd);
+#else
+  highbd_filter_intra_predictors_4tap(dst, stride, bs, above, left, D117_PRED,
+                                      bd);
+#endif
+}
+
+void av1_highbd_d153_filter_predictor_c(uint16_t *dst, ptrdiff_t stride, int bs,
+                                        const uint16_t *above,
+                                        const uint16_t *left, int bd) {
+#if USE_3TAP_INTRA_FILTER
+  highbd_filter_intra_predictors_3tap(dst, stride, bs, above, left, D153_PRED,
+                                      bd);
+#else
+  highbd_filter_intra_predictors_4tap(dst, stride, bs, above, left, D153_PRED,
+                                      bd);
+#endif
+}
+
+void av1_highbd_d207_filter_predictor_c(uint16_t *dst, ptrdiff_t stride, int bs,
+                                        const uint16_t *above,
+                                        const uint16_t *left, int bd) {
+#if USE_3TAP_INTRA_FILTER
+  highbd_filter_intra_predictors_3tap(dst, stride, bs, above, left, D207_PRED,
+                                      bd);
+#else
+  highbd_filter_intra_predictors_4tap(dst, stride, bs, above, left, D207_PRED,
+                                      bd);
+#endif
+}
+
+void av1_highbd_d63_filter_predictor_c(uint16_t *dst, ptrdiff_t stride, int bs,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+#if USE_3TAP_INTRA_FILTER
+  highbd_filter_intra_predictors_3tap(dst, stride, bs, above, left, D63_PRED,
+                                      bd);
+#else
+  highbd_filter_intra_predictors_4tap(dst, stride, bs, above, left, D63_PRED,
+                                      bd);
+#endif
+}
+
+void av1_highbd_tm_filter_predictor_c(uint16_t *dst, ptrdiff_t stride, int bs,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+#if USE_3TAP_INTRA_FILTER
+  highbd_filter_intra_predictors_3tap(dst, stride, bs, above, left, TM_PRED,
+                                      bd);
+#else
+  highbd_filter_intra_predictors_4tap(dst, stride, bs, above, left, TM_PRED,
+                                      bd);
+#endif
+}
+
+static void highbd_filter_intra_predictors(FILTER_INTRA_MODE mode,
+                                           uint16_t *dst, ptrdiff_t stride,
+                                           int bs, const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  switch (mode) {
+    case FILTER_DC_PRED:
+      av1_highbd_dc_filter_predictor(dst, stride, bs, above, left, bd);
+      break;
+    case FILTER_V_PRED:
+      av1_highbd_v_filter_predictor(dst, stride, bs, above, left, bd);
+      break;
+    case FILTER_H_PRED:
+      av1_highbd_h_filter_predictor(dst, stride, bs, above, left, bd);
+      break;
+    case FILTER_D45_PRED:
+      av1_highbd_d45_filter_predictor(dst, stride, bs, above, left, bd);
+      break;
+    case FILTER_D135_PRED:
+      av1_highbd_d135_filter_predictor(dst, stride, bs, above, left, bd);
+      break;
+    case FILTER_D117_PRED:
+      av1_highbd_d117_filter_predictor(dst, stride, bs, above, left, bd);
+      break;
+    case FILTER_D153_PRED:
+      av1_highbd_d153_filter_predictor(dst, stride, bs, above, left, bd);
+      break;
+    case FILTER_D207_PRED:
+      av1_highbd_d207_filter_predictor(dst, stride, bs, above, left, bd);
+      break;
+    case FILTER_D63_PRED:
+      av1_highbd_d63_filter_predictor(dst, stride, bs, above, left, bd);
+      break;
+    case FILTER_TM_PRED:
+      av1_highbd_tm_filter_predictor(dst, stride, bs, above, left, bd);
+      break;
+    default: assert(0);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_HIGHBITDEPTH
+static void build_intra_predictors_high(
+    const MACROBLOCKD *xd, const uint8_t *ref8, int ref_stride, uint8_t *dst8,
+    int dst_stride, PREDICTION_MODE mode, TX_SIZE tx_size, int n_top_px,
+    int n_topright_px, int n_left_px, int n_bottomleft_px, int plane) {
+  int i;
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  DECLARE_ALIGNED(16, uint16_t, left_data[MAX_TX_SIZE * 2 + 16]);
+  DECLARE_ALIGNED(16, uint16_t, above_data[MAX_TX_SIZE * 2 + 16]);
+  uint16_t *above_row = above_data + 16;
+  uint16_t *left_col = left_data + 16;
+  const uint16_t *const_above_row = above_row;
+  const int bs = tx_size_wide[tx_size];
+  int need_left = extend_modes[mode] & NEED_LEFT;
+  int need_above = extend_modes[mode] & NEED_ABOVE;
+  int need_above_left = extend_modes[mode] & NEED_ABOVELEFT;
+  const uint16_t *above_ref = ref - ref_stride;
+#if CONFIG_EXT_INTRA
+  int p_angle = 0;
+  const int is_dr_mode = av1_is_directional_mode(mode, xd->mi[0]->mbmi.sb_type);
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  const FILTER_INTRA_MODE_INFO *filter_intra_mode_info =
+      &xd->mi[0]->mbmi.filter_intra_mode_info;
+  const FILTER_INTRA_MODE filter_intra_mode =
+      filter_intra_mode_info->filter_intra_mode[plane != 0];
+#endif  // CONFIG_FILTER_INTRA
+  int base = 128 << (xd->bd - 8);
+  assert(tx_size_wide[tx_size] == tx_size_high[tx_size]);
+
+  // base-1 base-1 base-1 .. base-1 base-1 base-1 base-1 base-1 base-1
+  // base+1   A      B  ..     Y      Z
+  // base+1   C      D  ..     W      X
+  // base+1   E      F  ..     U      V
+  // base+1   G      H  ..     S      T      T      T      T      T
+  aom_memset16(left_data, base + 1, sizeof(left_data) / sizeof(*left_data));
+
+#if CONFIG_EXT_INTRA
+  if (is_dr_mode) {
+    p_angle = mode_to_angle_map[mode] +
+              xd->mi[0]->mbmi.angle_delta[plane != 0] * ANGLE_STEP;
+    if (p_angle <= 90)
+      need_above = 1, need_left = 0, need_above_left = 1;
+    else if (p_angle < 180)
+      need_above = 1, need_left = 1, need_above_left = 1;
+    else
+      need_above = 0, need_left = 1, need_above_left = 1;
+  }
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  if (filter_intra_mode_info->use_filter_intra_mode[plane != 0])
+    need_left = need_above = need_above_left = 1;
+#endif  // CONFIG_FILTER_INTRA
+
+  (void)plane;
+  assert(n_top_px >= 0);
+  assert(n_topright_px >= 0);
+  assert(n_left_px >= 0);
+  assert(n_bottomleft_px >= 0);
+
+  if ((!need_above && n_left_px == 0) || (!need_left && n_top_px == 0)) {
+    const int val = need_left ? base + 1 : base - 1;
+    for (i = 0; i < bs; ++i) {
+      aom_memset16(dst, val, bs);
+      dst += dst_stride;
+    }
+    return;
+  }
+
+  // NEED_LEFT
+  if (need_left) {
+#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA
+    int need_bottom = !!(extend_modes[mode] & NEED_BOTTOMLEFT);
+#if CONFIG_FILTER_INTRA
+    if (filter_intra_mode_info->use_filter_intra_mode[plane != 0])
+      need_bottom = 0;
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_EXT_INTRA
+    if (is_dr_mode) need_bottom = p_angle > 180;
+#endif  // CONFIG_EXT_INTRA
+#else
+    const int need_bottom = !!(extend_modes[mode] & NEED_BOTTOMLEFT);
+#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA
+    i = 0;
+    if (n_left_px > 0) {
+      for (; i < n_left_px; i++) left_col[i] = ref[i * ref_stride - 1];
+      if (need_bottom && n_bottomleft_px > 0) {
+        assert(i == bs);
+        for (; i < bs + n_bottomleft_px; i++)
+          left_col[i] = ref[i * ref_stride - 1];
+      }
+      if (i < (bs << need_bottom))
+        aom_memset16(&left_col[i], left_col[i - 1], (bs << need_bottom) - i);
+    } else {
+      aom_memset16(left_col, base + 1, bs << need_bottom);
+    }
+  }
+
+  // NEED_ABOVE
+  if (need_above) {
+#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA
+    int need_right = !!(extend_modes[mode] & NEED_ABOVERIGHT);
+#if CONFIG_FILTER_INTRA
+    if (filter_intra_mode_info->use_filter_intra_mode[plane != 0])
+      need_right = 1;
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_EXT_INTRA
+    if (is_dr_mode) need_right = p_angle < 90;
+#endif  // CONFIG_EXT_INTRA
+#else
+    const int need_right = !!(extend_modes[mode] & NEED_ABOVERIGHT);
+#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA
+    if (n_top_px > 0) {
+      memcpy(above_row, above_ref, n_top_px * sizeof(above_ref[0]));
+      i = n_top_px;
+      if (need_right && n_topright_px > 0) {
+        assert(n_top_px == bs);
+        memcpy(above_row + bs, above_ref + bs,
+               n_topright_px * sizeof(above_ref[0]));
+        i += n_topright_px;
+      }
+      if (i < (bs << need_right))
+        aom_memset16(&above_row[i], above_row[i - 1], (bs << need_right) - i);
+    } else {
+      aom_memset16(above_row, base - 1, bs << need_right);
+    }
+  }
+
+  if (need_above_left) {
+    above_row[-1] =
+        n_top_px > 0 ? (n_left_px > 0 ? above_ref[-1] : base + 1) : base - 1;
+    left_col[-1] = above_row[-1];
+  }
+
+#if CONFIG_FILTER_INTRA
+  if (filter_intra_mode_info->use_filter_intra_mode[plane != 0]) {
+    highbd_filter_intra_predictors(filter_intra_mode, dst, dst_stride, bs,
+                                   const_above_row, left_col, xd->bd);
+    return;
+  }
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_EXT_INTRA
+  if (is_dr_mode) {
+#if CONFIG_INTRA_INTERP
+    INTRA_FILTER filter = INTRA_FILTER_LINEAR;
+    if (plane == 0 && av1_is_intra_filter_switchable(p_angle))
+      filter = xd->mi[0]->mbmi.intra_filter;
+#endif  // CONFIG_INTRA_INTERP
+    highbd_dr_predictor(dst, dst_stride, bs, const_above_row, left_col,
+#if CONFIG_INTRA_INTERP
+                        filter,
+#endif  // CONFIG_INTRA_INTERP
+                        p_angle, xd->bd);
+    return;
+  }
+#endif  // CONFIG_EXT_INTRA
+
+  // predict
+  if (mode == DC_PRED) {
+    dc_pred_high[n_left_px > 0][n_top_px > 0][tx_size](
+        dst, dst_stride, const_above_row, left_col, xd->bd);
+  } else {
+    pred_high[mode][tx_size](dst, dst_stride, const_above_row, left_col,
+                             xd->bd);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static void build_intra_predictors(const MACROBLOCKD *xd, const uint8_t *ref,
+                                   int ref_stride, uint8_t *dst, int dst_stride,
+                                   PREDICTION_MODE mode, TX_SIZE tx_size,
+                                   int n_top_px, int n_topright_px,
+                                   int n_left_px, int n_bottomleft_px,
+                                   int plane) {
+  int i;
+  const uint8_t *above_ref = ref - ref_stride;
+  DECLARE_ALIGNED(16, uint8_t, left_data[MAX_TX_SIZE * 2 + 16]);
+  DECLARE_ALIGNED(16, uint8_t, above_data[MAX_TX_SIZE * 2 + 16]);
+  uint8_t *above_row = above_data + 16;
+  uint8_t *left_col = left_data + 16;
+  const uint8_t *const_above_row = above_row;
+  const int bs = tx_size_wide[tx_size];
+  int need_left = extend_modes[mode] & NEED_LEFT;
+  int need_above = extend_modes[mode] & NEED_ABOVE;
+  int need_above_left = extend_modes[mode] & NEED_ABOVELEFT;
+#if CONFIG_EXT_INTRA
+  int p_angle = 0;
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int is_dr_mode = av1_is_directional_mode(mode, mbmi->sb_type);
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  const FILTER_INTRA_MODE_INFO *filter_intra_mode_info =
+      &xd->mi[0]->mbmi.filter_intra_mode_info;
+  const FILTER_INTRA_MODE filter_intra_mode =
+      filter_intra_mode_info->filter_intra_mode[plane != 0];
+#endif  // CONFIG_FILTER_INTRA
+  assert(tx_size_wide[tx_size] == tx_size_high[tx_size]);
+
+  // 127 127 127 .. 127 127 127 127 127 127
+  // 129  A   B  ..  Y   Z
+  // 129  C   D  ..  W   X
+  // 129  E   F  ..  U   V
+  // 129  G   H  ..  S   T   T   T   T   T
+  // ..
+  memset(left_data, 129, sizeof(left_data));
+
+#if CONFIG_EXT_INTRA
+  if (is_dr_mode) {
+    p_angle = mode_to_angle_map[mode] +
+              xd->mi[0]->mbmi.angle_delta[plane != 0] * ANGLE_STEP;
+    if (p_angle <= 90)
+      need_above = 1, need_left = 0, need_above_left = 1;
+    else if (p_angle < 180)
+      need_above = 1, need_left = 1, need_above_left = 1;
+    else
+      need_above = 0, need_left = 1, need_above_left = 1;
+  }
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  if (filter_intra_mode_info->use_filter_intra_mode[plane != 0])
+    need_left = need_above = need_above_left = 1;
+#endif  // CONFIG_FILTER_INTRA
+
+  (void)xd;
+  (void)plane;
+  assert(n_top_px >= 0);
+  assert(n_topright_px >= 0);
+  assert(n_left_px >= 0);
+  assert(n_bottomleft_px >= 0);
+
+  if ((!need_above && n_left_px == 0) || (!need_left && n_top_px == 0)) {
+    const int val = need_left ? 129 : 127;
+    for (i = 0; i < bs; ++i) {
+      memset(dst, val, bs);
+      dst += dst_stride;
+    }
+    return;
+  }
+
+  // NEED_LEFT
+  if (need_left) {
+#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA
+    int need_bottom = !!(extend_modes[mode] & NEED_BOTTOMLEFT);
+#if CONFIG_FILTER_INTRA
+    if (filter_intra_mode_info->use_filter_intra_mode[plane != 0])
+      need_bottom = 0;
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_EXT_INTRA
+    if (is_dr_mode) need_bottom = p_angle > 180;
+#endif  // CONFIG_EXT_INTRA
+#else
+    const int need_bottom = !!(extend_modes[mode] & NEED_BOTTOMLEFT);
+#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA
+    i = 0;
+    if (n_left_px > 0) {
+      for (; i < n_left_px; i++) left_col[i] = ref[i * ref_stride - 1];
+      if (need_bottom && n_bottomleft_px > 0) {
+        assert(i == bs);
+        for (; i < bs + n_bottomleft_px; i++)
+          left_col[i] = ref[i * ref_stride - 1];
+      }
+      if (i < (bs << need_bottom))
+        memset(&left_col[i], left_col[i - 1], (bs << need_bottom) - i);
+    } else {
+      memset(left_col, 129, bs << need_bottom);
+    }
+  }
+
+  // NEED_ABOVE
+  if (need_above) {
+#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA
+    int need_right = !!(extend_modes[mode] & NEED_ABOVERIGHT);
+#if CONFIG_FILTER_INTRA
+    if (filter_intra_mode_info->use_filter_intra_mode[plane != 0])
+      need_right = 1;
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_EXT_INTRA
+    if (is_dr_mode) need_right = p_angle < 90;
+#endif  // CONFIG_EXT_INTRA
+#else
+    const int need_right = !!(extend_modes[mode] & NEED_ABOVERIGHT);
+#endif  // CONFIG_EXT_INTRA || CONFIG_FITLER_INTRA
+    if (n_top_px > 0) {
+      memcpy(above_row, above_ref, n_top_px);
+      i = n_top_px;
+      if (need_right && n_topright_px > 0) {
+        assert(n_top_px == bs);
+        memcpy(above_row + bs, above_ref + bs, n_topright_px);
+        i += n_topright_px;
+      }
+      if (i < (bs << need_right))
+        memset(&above_row[i], above_row[i - 1], (bs << need_right) - i);
+    } else {
+      memset(above_row, 127, bs << need_right);
+    }
+  }
+
+  if (need_above_left) {
+    above_row[-1] = n_top_px > 0 ? (n_left_px > 0 ? above_ref[-1] : 129) : 127;
+    left_col[-1] = above_row[-1];
+  }
+
+#if CONFIG_FILTER_INTRA
+  if (filter_intra_mode_info->use_filter_intra_mode[plane != 0]) {
+    filter_intra_predictors(filter_intra_mode, dst, dst_stride, bs,
+                            const_above_row, left_col);
+    return;
+  }
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_EXT_INTRA
+  if (is_dr_mode) {
+#if CONFIG_INTRA_INTERP
+    INTRA_FILTER filter = INTRA_FILTER_LINEAR;
+    if (plane == 0 && av1_is_intra_filter_switchable(p_angle))
+      filter = xd->mi[0]->mbmi.intra_filter;
+#endif  // CONFIG_INTRA_INTERP
+    dr_predictor(dst, dst_stride, tx_size, const_above_row, left_col,
+#if CONFIG_INTRA_INTERP
+                 filter,
+#endif  // CONFIG_INTRA_INTERP
+                 p_angle);
+    return;
+  }
+#endif  // CONFIG_EXT_INTRA
+
+  // predict
+  if (mode == DC_PRED) {
+#if CONFIG_CFL
+    // CFL predict its own DC_PRED for Chromatic planes
+    if (plane == AOM_PLANE_Y) {
+#endif
+      dc_pred[n_left_px > 0][n_top_px > 0][tx_size](dst, dst_stride,
+                                                    const_above_row, left_col);
+#if CONFIG_CFL
+    }
+#endif
+
+  } else {
+    pred[mode][tx_size](dst, dst_stride, const_above_row, left_col);
+  }
+}
+
+static void predict_square_intra_block(const MACROBLOCKD *xd, int wpx, int hpx,
+                                       TX_SIZE tx_size, PREDICTION_MODE mode,
+                                       const uint8_t *ref, int ref_stride,
+                                       uint8_t *dst, int dst_stride,
+                                       int col_off, int row_off, int plane) {
+  BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int txw = tx_size_wide_unit[tx_size];
+#if CONFIG_CB4X4 && CONFIG_CHROMA_SUB8X8
+  const int have_top = row_off || (pd->subsampling_y ? xd->chroma_up_available
+                                                     : xd->up_available);
+  const int have_left =
+      col_off ||
+      (pd->subsampling_x ? xd->chroma_left_available : xd->left_available);
+#else
+  const int have_top = row_off || xd->up_available;
+  const int have_left = col_off || xd->left_available;
+#endif
+  const int x = col_off << tx_size_wide_log2[0];
+  const int y = row_off << tx_size_high_log2[0];
+  const int mi_row = -xd->mb_to_top_edge >> (3 + MI_SIZE_LOG2);
+  const int mi_col = -xd->mb_to_left_edge >> (3 + MI_SIZE_LOG2);
+  const int txwpx = tx_size_wide[tx_size];
+  const int txhpx = tx_size_high[tx_size];
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+  const int xr_chr_offset = (pd->subsampling_x && bsize < BLOCK_8X8) ? 2 : 0;
+  const int yd_chr_offset = (pd->subsampling_y && bsize < BLOCK_8X8) ? 2 : 0;
+#else
+  const int xr_chr_offset = 0;
+  const int yd_chr_offset = 0;
+#endif
+
+  // Distance between the right edge of this prediction block to
+  // the frame right edge
+  const int xr = (xd->mb_to_right_edge >> (3 + pd->subsampling_x)) +
+                 (wpx - x - txwpx) - xr_chr_offset;
+  // Distance between the bottom edge of this prediction block to
+  // the frame bottom edge
+  const int yd = (xd->mb_to_bottom_edge >> (3 + pd->subsampling_y)) +
+                 (hpx - y - txhpx) - yd_chr_offset;
+  const int right_available =
+      (mi_col + ((col_off + txw) >> (1 - pd->subsampling_x))) <
+      xd->tile.mi_col_end;
+  const int bottom_available = (yd > 0);
+#if CONFIG_EXT_PARTITION_TYPES
+  const PARTITION_TYPE partition = xd->mi[0]->mbmi.partition;
+#endif
+
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+  // force 4x4 chroma component block size.
+  bsize = scale_chroma_bsize(bsize, pd->subsampling_x, pd->subsampling_y);
+#endif
+
+  const int have_top_right =
+      has_top_right(bsize, mi_row, mi_col, have_top, right_available,
+#if CONFIG_EXT_PARTITION_TYPES
+                    partition,
+#endif
+                    tx_size, row_off, col_off, pd->subsampling_x);
+  const int have_bottom_left =
+      has_bottom_left(bsize, mi_row, mi_col, bottom_available, have_left,
+                      tx_size, row_off, col_off, pd->subsampling_y);
+  assert(txwpx == txhpx);
+
+#if CONFIG_PALETTE
+  if (xd->mi[0]->mbmi.palette_mode_info.palette_size[plane != 0] > 0) {
+    const int bs = tx_size_wide[tx_size];
+    const int stride = wpx;
+    int r, c;
+    const uint8_t *const map = xd->plane[plane != 0].color_index_map;
+#if CONFIG_HIGHBITDEPTH
+    uint16_t *palette = xd->mi[0]->mbmi.palette_mode_info.palette_colors +
+                        plane * PALETTE_MAX_SIZE;
+#else
+    uint8_t *palette = xd->mi[0]->mbmi.palette_mode_info.palette_colors +
+                       plane * PALETTE_MAX_SIZE;
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
+      for (r = 0; r < bs; ++r)
+        for (c = 0; c < bs; ++c)
+          dst16[r * dst_stride + c] = palette[map[(r + y) * stride + c + x]];
+    } else {
+      for (r = 0; r < bs; ++r)
+        for (c = 0; c < bs; ++c)
+          dst[r * dst_stride + c] =
+              (uint8_t)(palette[map[(r + y) * stride + c + x]]);
+    }
+#else
+    for (r = 0; r < bs; ++r)
+      for (c = 0; c < bs; ++c)
+        dst[r * dst_stride + c] = palette[map[(r + y) * stride + c + x]];
+#endif  // CONFIG_HIGHBITDEPTH
+    return;
+  }
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    build_intra_predictors_high(
+        xd, ref, ref_stride, dst, dst_stride, mode, tx_size,
+        have_top ? AOMMIN(txwpx, xr + txwpx) : 0,
+        have_top_right ? AOMMIN(txwpx, xr) : 0,
+        have_left ? AOMMIN(txhpx, yd + txhpx) : 0,
+        have_bottom_left ? AOMMIN(txhpx, yd) : 0, plane);
+    return;
+  }
+#endif
+  build_intra_predictors(xd, ref, ref_stride, dst, dst_stride, mode, tx_size,
+                         have_top ? AOMMIN(txwpx, xr + txwpx) : 0,
+                         have_top_right ? AOMMIN(txwpx, xr) : 0,
+                         have_left ? AOMMIN(txhpx, yd + txhpx) : 0,
+                         have_bottom_left ? AOMMIN(txhpx, yd) : 0, plane);
+}
+
+void av1_predict_intra_block_facade(MACROBLOCKD *xd, int plane, int block_idx,
+                                    int blk_col, int blk_row, TX_SIZE tx_size) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int dst_stride = pd->dst.stride;
+  uint8_t *dst =
+      &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int block_raster_idx =
+      av1_block_index_to_raster_order(tx_size, block_idx);
+  const PREDICTION_MODE mode =
+      (plane == 0) ? get_y_mode(xd->mi[0], block_raster_idx) : mbmi->uv_mode;
+  av1_predict_intra_block(xd, pd->width, pd->height, txsize_to_bsize[tx_size],
+                          mode, dst, dst_stride, dst, dst_stride, blk_col,
+                          blk_row, plane);
+#if CONFIG_CFL
+  if (plane != AOM_PLANE_Y && mbmi->uv_mode == DC_PRED) {
+    if (plane == AOM_PLANE_U && blk_col == 0 && blk_row == 0) {
+      // Compute the block-level DC_PRED for both chromatic planes prior to
+      // processing the first chromatic plane in order to compute alpha_cb and
+      // alpha_cr. Note: This is not required on the decoder side because alpha
+      // is signaled.
+      cfl_dc_pred(xd, get_plane_block_size(block_idx, pd), tx_size);
+    }
+    cfl_predict_block(xd->cfl, dst, pd->dst.stride, blk_row, blk_col, tx_size,
+                      xd->cfl->dc_pred[plane - 1]);
+  }
+#endif
+}
+
+void av1_predict_intra_block(const MACROBLOCKD *xd, int wpx, int hpx,
+                             BLOCK_SIZE bsize, PREDICTION_MODE mode,
+                             const uint8_t *ref, int ref_stride, uint8_t *dst,
+                             int dst_stride, int col_off, int row_off,
+                             int plane) {
+  const int block_width = block_size_wide[bsize];
+  const int block_height = block_size_high[bsize];
+  TX_SIZE tx_size = max_txsize_lookup[bsize];
+  assert(tx_size < TX_SIZES);
+  if (block_width == block_height) {
+    predict_square_intra_block(xd, wpx, hpx, tx_size, mode, ref, ref_stride,
+                               dst, dst_stride, col_off, row_off, plane);
+  } else {
+#if (CONFIG_RECT_TX && (CONFIG_VAR_TX || CONFIG_EXT_TX)) || (CONFIG_EXT_INTER)
+#if CONFIG_HIGHBITDEPTH
+    uint16_t tmp16[MAX_SB_SIZE];
+#endif
+    uint8_t tmp[MAX_SB_SIZE];
+    assert((block_width == wpx && block_height == hpx) ||
+           (block_width == (wpx >> 1) && block_height == hpx) ||
+           (block_width == wpx && block_height == (hpx >> 1)));
+
+    if (block_width < block_height) {
+      assert(block_height == (block_width << 1));
+      // Predict the top square sub-block.
+      predict_square_intra_block(xd, wpx, hpx, tx_size, mode, ref, ref_stride,
+                                 dst, dst_stride, col_off, row_off, plane);
+      {
+        const int half_block_height = block_height >> 1;
+        const int half_block_height_unit =
+            half_block_height >> tx_size_wide_log2[0];
+        // Cast away const to modify 'ref' temporarily; will be restored later.
+        uint8_t *src_2 = (uint8_t *)ref + half_block_height * ref_stride;
+        uint8_t *dst_2 = dst + half_block_height * dst_stride;
+        const int row_off_2 = row_off + half_block_height_unit;
+        // Save the last row of top square sub-block as 'above' row for bottom
+        // square sub-block.
+        if (src_2 != dst_2 || ref_stride != dst_stride) {
+#if CONFIG_HIGHBITDEPTH
+          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            uint16_t *src_2_16 = CONVERT_TO_SHORTPTR(src_2);
+            uint16_t *dst_2_16 = CONVERT_TO_SHORTPTR(dst_2);
+            memcpy(tmp16, src_2_16 - ref_stride,
+                   block_width * sizeof(*src_2_16));
+            memcpy(src_2_16 - ref_stride, dst_2_16 - dst_stride,
+                   block_width * sizeof(*src_2_16));
+          } else {
+#endif  // CONFIG_HIGHBITDEPTH
+            memcpy(tmp, src_2 - ref_stride, block_width * sizeof(*src_2));
+            memcpy(src_2 - ref_stride, dst_2 - dst_stride,
+                   block_width * sizeof(*src_2));
+#if CONFIG_HIGHBITDEPTH
+          }
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+        // Predict the bottom square sub-block.
+        predict_square_intra_block(xd, wpx, hpx, tx_size, mode, src_2,
+                                   ref_stride, dst_2, dst_stride, col_off,
+                                   row_off_2, plane);
+        // Restore the last row of top square sub-block.
+        if (src_2 != dst_2 || ref_stride != dst_stride) {
+#if CONFIG_HIGHBITDEPTH
+          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            uint16_t *src_2_16 = CONVERT_TO_SHORTPTR(src_2);
+            memcpy(src_2_16 - ref_stride, tmp16,
+                   block_width * sizeof(*src_2_16));
+          } else {
+#endif  // CONFIG_HIGHBITDEPTH
+            memcpy(src_2 - ref_stride, tmp, block_width * sizeof(*src_2));
+#if CONFIG_HIGHBITDEPTH
+          }
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+    } else {  // block_width > block_height
+      assert(block_width == (block_height << 1));
+      // Predict the left square sub-block
+      predict_square_intra_block(xd, wpx, hpx, tx_size, mode, ref, ref_stride,
+                                 dst, dst_stride, col_off, row_off, plane);
+      {
+        int i;
+        const int half_block_width = block_width >> 1;
+        const int half_block_width_unit =
+            half_block_width >> tx_size_wide_log2[0];
+        // Cast away const to modify 'ref' temporarily; will be restored later.
+        uint8_t *src_2 = (uint8_t *)ref + half_block_width;
+        uint8_t *dst_2 = dst + half_block_width;
+        const int col_off_2 = col_off + half_block_width_unit;
+        // Save the last column of left square sub-block as 'left' column for
+        // right square sub-block.
+        const int save_src = src_2 != dst_2 || ref_stride != dst_stride;
+        if (save_src) {
+#if CONFIG_HIGHBITDEPTH
+          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            uint16_t *src_2_16 = CONVERT_TO_SHORTPTR(src_2);
+            uint16_t *dst_2_16 = CONVERT_TO_SHORTPTR(dst_2);
+            for (i = 0; i < block_height; ++i) {
+              tmp16[i] = src_2_16[i * ref_stride - 1];
+              src_2_16[i * ref_stride - 1] = dst_2_16[i * dst_stride - 1];
+            }
+          } else {
+#endif  // CONFIG_HIGHBITDEPTH
+            for (i = 0; i < block_height; ++i) {
+              tmp[i] = src_2[i * ref_stride - 1];
+              src_2[i * ref_stride - 1] = dst_2[i * dst_stride - 1];
+            }
+#if CONFIG_HIGHBITDEPTH
+          }
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+        // Predict the right square sub-block.
+        predict_square_intra_block(xd, wpx, hpx, tx_size, mode, src_2,
+                                   ref_stride, dst_2, dst_stride, col_off_2,
+                                   row_off, plane);
+        // Restore the last column of left square sub-block.
+        if (save_src) {
+#if CONFIG_HIGHBITDEPTH
+          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            uint16_t *src_2_16 = CONVERT_TO_SHORTPTR(src_2);
+            for (i = 0; i < block_height; ++i) {
+              src_2_16[i * ref_stride - 1] = tmp16[i];
+            }
+          } else {
+#endif  // CONFIG_HIGHBITDEPTH
+            for (i = 0; i < block_height; ++i) {
+              src_2[i * ref_stride - 1] = tmp[i];
+            }
+#if CONFIG_HIGHBITDEPTH
+          }
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+    }
+#else
+    assert(0);
+#endif  // (CONFIG_RECT_TX && (CONFIG_VAR_TX || CONFIG_EXT_TX)) ||
+        // (CONFIG_EXT_INTER)
+  }
+}
+
+void av1_init_intra_predictors(void) {
+  once(av1_init_intra_predictors_internal);
+}
diff --git a/third_party/aom/av1/common/reconintra.h b/third_party/aom/av1/common/reconintra.h
new file mode 100644
index 0000000000..7ee0c495e0
--- /dev/null
+++ b/third_party/aom/av1/common/reconintra.h
@@ -0,0 +1,67 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_RECONINTRA_H_
+#define AV1_COMMON_RECONINTRA_H_
+
+#include "aom/aom_integer.h"
+#include "av1/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_init_intra_predictors(void);
+void av1_predict_intra_block_facade(MACROBLOCKD *xd, int plane, int block_idx,
+                                    int blk_col, int blk_row, TX_SIZE tx_size);
+void av1_predict_intra_block(const MACROBLOCKD *xd, int bw, int bh,
+                             BLOCK_SIZE bsize, PREDICTION_MODE mode,
+                             const uint8_t *ref, int ref_stride, uint8_t *dst,
+                             int dst_stride, int aoff, int loff, int plane);
+
+#if CONFIG_EXT_INTER
+// Mapping of interintra to intra mode for use in the intra component
+static const PREDICTION_MODE interintra_to_intra_mode[INTERINTRA_MODES] = {
+  DC_PRED,   V_PRED,    H_PRED,    D45_PRED, D135_PRED,
+  D117_PRED, D153_PRED, D207_PRED, D63_PRED, TM_PRED
+};
+
+// Mapping of intra mode to the interintra mode
+static const INTERINTRA_MODE intra_to_interintra_mode[INTRA_MODES] = {
+  II_DC_PRED,   II_V_PRED,    II_H_PRED,    II_D45_PRED, II_D135_PRED,
+  II_D117_PRED, II_D153_PRED, II_D207_PRED, II_D63_PRED,
+#if CONFIG_ALT_INTRA
+  II_DC_PRED,  // Note: Filler value, as there's no II_SMOOTH_PRED.
+#endif         // CONFIG_ALT_INTRA
+  II_TM_PRED
+};
+#endif  // CONFIG_EXT_INTER
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#if CONFIG_FILTER_INTRA
+#define FILTER_INTRA_PREC_BITS 10
+extern int av1_filter_intra_taps_4[TX_SIZES][INTRA_MODES][4];
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_EXT_INTRA
+static INLINE int av1_is_directional_mode(PREDICTION_MODE mode,
+                                          BLOCK_SIZE bsize) {
+  return mode != DC_PRED && mode != TM_PRED &&
+#if CONFIG_ALT_INTRA
+         mode != SMOOTH_PRED &&
+#endif  // CONFIG_ALT_INTRA
+         bsize >= BLOCK_8X8;
+}
+#endif  // CONFIG_EXT_INTRA
+
+#endif  // AV1_COMMON_RECONINTRA_H_
diff --git a/third_party/aom/av1/common/resize.c b/third_party/aom/av1/common/resize.c
new file mode 100644
index 0000000000..8c0d3aa09b
--- /dev/null
+++ b/third_party/aom/av1/common/resize.c
@@ -0,0 +1,821 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "./aom_config.h"
+#if CONFIG_HIGHBITDEPTH
+#include "aom_dsp/aom_dsp_common.h"
+#endif  // CONFIG_HIGHBITDEPTH
+#include "aom_ports/mem.h"
+#include "av1/common/common.h"
+#include "av1/common/resize.h"
+
+#define FILTER_BITS 7
+
+#define INTERP_TAPS 8
+#define SUBPEL_BITS 5
+#define SUBPEL_MASK ((1 << SUBPEL_BITS) - 1)
+#define INTERP_PRECISION_BITS 32
+
+typedef int16_t interp_kernel[INTERP_TAPS];
+
+// Filters for interpolation (0.5-band) - note this also filters integer pels.
+static const interp_kernel filteredinterp_filters500[(1 << SUBPEL_BITS)] = {
+  { -3, 0, 35, 64, 35, 0, -3, 0 },    { -3, -1, 34, 64, 36, 1, -3, 0 },
+  { -3, -1, 32, 64, 38, 1, -3, 0 },   { -2, -2, 31, 63, 39, 2, -3, 0 },
+  { -2, -2, 29, 63, 41, 2, -3, 0 },   { -2, -2, 28, 63, 42, 3, -4, 0 },
+  { -2, -3, 27, 63, 43, 4, -4, 0 },   { -2, -3, 25, 62, 45, 5, -4, 0 },
+  { -2, -3, 24, 62, 46, 5, -4, 0 },   { -2, -3, 23, 61, 47, 6, -4, 0 },
+  { -2, -3, 21, 60, 49, 7, -4, 0 },   { -1, -4, 20, 60, 50, 8, -4, -1 },
+  { -1, -4, 19, 59, 51, 9, -4, -1 },  { -1, -4, 17, 58, 52, 10, -4, 0 },
+  { -1, -4, 16, 57, 53, 12, -4, -1 }, { -1, -4, 15, 56, 54, 13, -4, -1 },
+  { -1, -4, 14, 55, 55, 14, -4, -1 }, { -1, -4, 13, 54, 56, 15, -4, -1 },
+  { -1, -4, 12, 53, 57, 16, -4, -1 }, { 0, -4, 10, 52, 58, 17, -4, -1 },
+  { -1, -4, 9, 51, 59, 19, -4, -1 },  { -1, -4, 8, 50, 60, 20, -4, -1 },
+  { 0, -4, 7, 49, 60, 21, -3, -2 },   { 0, -4, 6, 47, 61, 23, -3, -2 },
+  { 0, -4, 5, 46, 62, 24, -3, -2 },   { 0, -4, 5, 45, 62, 25, -3, -2 },
+  { 0, -4, 4, 43, 63, 27, -3, -2 },   { 0, -4, 3, 42, 63, 28, -2, -2 },
+  { 0, -3, 2, 41, 63, 29, -2, -2 },   { 0, -3, 2, 39, 63, 31, -2, -2 },
+  { 0, -3, 1, 38, 64, 32, -1, -3 },   { 0, -3, 1, 36, 64, 34, -1, -3 }
+};
+
+// Filters for interpolation (0.625-band) - note this also filters integer pels.
+static const interp_kernel filteredinterp_filters625[(1 << SUBPEL_BITS)] = {
+  { -1, -8, 33, 80, 33, -8, -1, 0 }, { -1, -8, 30, 80, 35, -8, -1, 1 },
+  { -1, -8, 28, 80, 37, -7, -2, 1 }, { 0, -8, 26, 79, 39, -7, -2, 1 },
+  { 0, -8, 24, 79, 41, -7, -2, 1 },  { 0, -8, 22, 78, 43, -6, -2, 1 },
+  { 0, -8, 20, 78, 45, -5, -3, 1 },  { 0, -8, 18, 77, 48, -5, -3, 1 },
+  { 0, -8, 16, 76, 50, -4, -3, 1 },  { 0, -8, 15, 75, 52, -3, -4, 1 },
+  { 0, -7, 13, 74, 54, -3, -4, 1 },  { 0, -7, 11, 73, 56, -2, -4, 1 },
+  { 0, -7, 10, 71, 58, -1, -4, 1 },  { 1, -7, 8, 70, 60, 0, -5, 1 },
+  { 1, -6, 6, 68, 62, 1, -5, 1 },    { 1, -6, 5, 67, 63, 2, -5, 1 },
+  { 1, -6, 4, 65, 65, 4, -6, 1 },    { 1, -5, 2, 63, 67, 5, -6, 1 },
+  { 1, -5, 1, 62, 68, 6, -6, 1 },    { 1, -5, 0, 60, 70, 8, -7, 1 },
+  { 1, -4, -1, 58, 71, 10, -7, 0 },  { 1, -4, -2, 56, 73, 11, -7, 0 },
+  { 1, -4, -3, 54, 74, 13, -7, 0 },  { 1, -4, -3, 52, 75, 15, -8, 0 },
+  { 1, -3, -4, 50, 76, 16, -8, 0 },  { 1, -3, -5, 48, 77, 18, -8, 0 },
+  { 1, -3, -5, 45, 78, 20, -8, 0 },  { 1, -2, -6, 43, 78, 22, -8, 0 },
+  { 1, -2, -7, 41, 79, 24, -8, 0 },  { 1, -2, -7, 39, 79, 26, -8, 0 },
+  { 1, -2, -7, 37, 80, 28, -8, -1 }, { 1, -1, -8, 35, 80, 30, -8, -1 },
+};
+
+// Filters for interpolation (0.75-band) - note this also filters integer pels.
+static const interp_kernel filteredinterp_filters750[(1 << SUBPEL_BITS)] = {
+  { 2, -11, 25, 96, 25, -11, 2, 0 }, { 2, -11, 22, 96, 28, -11, 2, 0 },
+  { 2, -10, 19, 95, 31, -11, 2, 0 }, { 2, -10, 17, 95, 34, -12, 2, 0 },
+  { 2, -9, 14, 94, 37, -12, 2, 0 },  { 2, -8, 12, 93, 40, -12, 1, 0 },
+  { 2, -8, 9, 92, 43, -12, 1, 1 },   { 2, -7, 7, 91, 46, -12, 1, 0 },
+  { 2, -7, 5, 90, 49, -12, 1, 0 },   { 2, -6, 3, 88, 52, -12, 0, 1 },
+  { 2, -5, 1, 86, 55, -12, 0, 1 },   { 2, -5, -1, 84, 58, -11, 0, 1 },
+  { 2, -4, -2, 82, 61, -11, -1, 1 }, { 2, -4, -4, 80, 64, -10, -1, 1 },
+  { 1, -3, -5, 77, 67, -9, -1, 1 },  { 1, -3, -6, 75, 70, -8, -2, 1 },
+  { 1, -2, -7, 72, 72, -7, -2, 1 },  { 1, -2, -8, 70, 75, -6, -3, 1 },
+  { 1, -1, -9, 67, 77, -5, -3, 1 },  { 1, -1, -10, 64, 80, -4, -4, 2 },
+  { 1, -1, -11, 61, 82, -2, -4, 2 }, { 1, 0, -11, 58, 84, -1, -5, 2 },
+  { 1, 0, -12, 55, 86, 1, -5, 2 },   { 1, 0, -12, 52, 88, 3, -6, 2 },
+  { 0, 1, -12, 49, 90, 5, -7, 2 },   { 0, 1, -12, 46, 91, 7, -7, 2 },
+  { 1, 1, -12, 43, 92, 9, -8, 2 },   { 0, 1, -12, 40, 93, 12, -8, 2 },
+  { 0, 2, -12, 37, 94, 14, -9, 2 },  { 0, 2, -12, 34, 95, 17, -10, 2 },
+  { 0, 2, -11, 31, 95, 19, -10, 2 }, { 0, 2, -11, 28, 96, 22, -11, 2 }
+};
+
+// Filters for interpolation (0.875-band) - note this also filters integer pels.
+static const interp_kernel filteredinterp_filters875[(1 << SUBPEL_BITS)] = {
+  { 3, -8, 13, 112, 13, -8, 3, 0 },   { 3, -7, 10, 112, 17, -9, 3, -1 },
+  { 2, -6, 7, 111, 21, -9, 3, -1 },   { 2, -5, 4, 111, 24, -10, 3, -1 },
+  { 2, -4, 1, 110, 28, -11, 3, -1 },  { 1, -3, -1, 108, 32, -12, 4, -1 },
+  { 1, -2, -3, 106, 36, -13, 4, -1 }, { 1, -1, -6, 105, 40, -14, 4, -1 },
+  { 1, -1, -7, 102, 44, -14, 4, -1 }, { 1, 0, -9, 100, 48, -15, 4, -1 },
+  { 1, 1, -11, 97, 53, -16, 4, -1 },  { 0, 1, -12, 95, 57, -16, 4, -1 },
+  { 0, 2, -13, 91, 61, -16, 4, -1 },  { 0, 2, -14, 88, 65, -16, 4, -1 },
+  { 0, 3, -15, 84, 69, -17, 4, 0 },   { 0, 3, -16, 81, 73, -16, 3, 0 },
+  { 0, 3, -16, 77, 77, -16, 3, 0 },   { 0, 3, -16, 73, 81, -16, 3, 0 },
+  { 0, 4, -17, 69, 84, -15, 3, 0 },   { -1, 4, -16, 65, 88, -14, 2, 0 },
+  { -1, 4, -16, 61, 91, -13, 2, 0 },  { -1, 4, -16, 57, 95, -12, 1, 0 },
+  { -1, 4, -16, 53, 97, -11, 1, 1 },  { -1, 4, -15, 48, 100, -9, 0, 1 },
+  { -1, 4, -14, 44, 102, -7, -1, 1 }, { -1, 4, -14, 40, 105, -6, -1, 1 },
+  { -1, 4, -13, 36, 106, -3, -2, 1 }, { -1, 4, -12, 32, 108, -1, -3, 1 },
+  { -1, 3, -11, 28, 110, 1, -4, 2 },  { -1, 3, -10, 24, 111, 4, -5, 2 },
+  { -1, 3, -9, 21, 111, 7, -6, 2 },   { -1, 3, -9, 17, 112, 10, -7, 3 }
+};
+
+// Filters for interpolation (full-band) - no filtering for integer pixels
+static const interp_kernel filteredinterp_filters1000[(1 << SUBPEL_BITS)] = {
+  { 0, 0, 0, 128, 0, 0, 0, 0 },        { 0, 1, -3, 128, 3, -1, 0, 0 },
+  { -1, 2, -6, 127, 7, -2, 1, 0 },     { -1, 3, -9, 126, 12, -4, 1, 0 },
+  { -1, 4, -12, 125, 16, -5, 1, 0 },   { -1, 4, -14, 123, 20, -6, 2, 0 },
+  { -1, 5, -15, 120, 25, -8, 2, 0 },   { -1, 5, -17, 118, 30, -9, 3, -1 },
+  { -1, 6, -18, 114, 35, -10, 3, -1 }, { -1, 6, -19, 111, 41, -12, 3, -1 },
+  { -1, 6, -20, 107, 46, -13, 4, -1 }, { -1, 6, -21, 103, 52, -14, 4, -1 },
+  { -1, 6, -21, 99, 57, -16, 5, -1 },  { -1, 6, -21, 94, 63, -17, 5, -1 },
+  { -1, 6, -20, 89, 68, -18, 5, -1 },  { -1, 6, -20, 84, 73, -19, 6, -1 },
+  { -1, 6, -20, 79, 79, -20, 6, -1 },  { -1, 6, -19, 73, 84, -20, 6, -1 },
+  { -1, 5, -18, 68, 89, -20, 6, -1 },  { -1, 5, -17, 63, 94, -21, 6, -1 },
+  { -1, 5, -16, 57, 99, -21, 6, -1 },  { -1, 4, -14, 52, 103, -21, 6, -1 },
+  { -1, 4, -13, 46, 107, -20, 6, -1 }, { -1, 3, -12, 41, 111, -19, 6, -1 },
+  { -1, 3, -10, 35, 114, -18, 6, -1 }, { -1, 3, -9, 30, 118, -17, 5, -1 },
+  { 0, 2, -8, 25, 120, -15, 5, -1 },   { 0, 2, -6, 20, 123, -14, 4, -1 },
+  { 0, 1, -5, 16, 125, -12, 4, -1 },   { 0, 1, -4, 12, 126, -9, 3, -1 },
+  { 0, 1, -2, 7, 127, -6, 2, -1 },     { 0, 0, -1, 3, 128, -3, 1, 0 }
+};
+
+// Filters for factor of 2 downsampling.
+static const int16_t av1_down2_symeven_half_filter[] = { 56, 12, -3, -1 };
+static const int16_t av1_down2_symodd_half_filter[] = { 64, 35, 0, -3 };
+
+static const interp_kernel *choose_interp_filter(int inlength, int outlength) {
+  int outlength16 = outlength * 16;
+  if (outlength16 >= inlength * 16)
+    return filteredinterp_filters1000;
+  else if (outlength16 >= inlength * 13)
+    return filteredinterp_filters875;
+  else if (outlength16 >= inlength * 11)
+    return filteredinterp_filters750;
+  else if (outlength16 >= inlength * 9)
+    return filteredinterp_filters625;
+  else
+    return filteredinterp_filters500;
+}
+
+static void interpolate(const uint8_t *const input, int inlength,
+                        uint8_t *output, int outlength) {
+  const int64_t delta =
+      (((uint64_t)inlength << 32) + outlength / 2) / outlength;
+  const int64_t offset =
+      inlength > outlength
+          ? (((int64_t)(inlength - outlength) << 31) + outlength / 2) /
+                outlength
+          : -(((int64_t)(outlength - inlength) << 31) + outlength / 2) /
+                outlength;
+  uint8_t *optr = output;
+  int x, x1, x2, sum, k, int_pel, sub_pel;
+  int64_t y;
+
+  const interp_kernel *interp_filters =
+      choose_interp_filter(inlength, outlength);
+
+  x = 0;
+  y = offset;
+  while ((y >> INTERP_PRECISION_BITS) < (INTERP_TAPS / 2 - 1)) {
+    x++;
+    y += delta;
+  }
+  x1 = x;
+  x = outlength - 1;
+  y = delta * x + offset;
+  while ((y >> INTERP_PRECISION_BITS) + (int64_t)(INTERP_TAPS / 2) >=
+         inlength) {
+    x--;
+    y -= delta;
+  }
+  x2 = x;
+  if (x1 > x2) {
+    for (x = 0, y = offset; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k) {
+        const int pk = int_pel - INTERP_TAPS / 2 + 1 + k;
+        sum += filter[k] *
+               input[(pk < 0 ? 0 : (pk >= inlength ? inlength - 1 : pk))];
+      }
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+  } else {
+    // Initial part.
+    for (x = 0, y = offset; x < x1; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k < 0
+                                      ? 0
+                                      : int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+    // Middle part.
+    for (; x <= x2; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[int_pel - INTERP_TAPS / 2 + 1 + k];
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+    // End part.
+    for (; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k >= inlength
+                                      ? inlength - 1
+                                      : int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+  }
+}
+
+static void down2_symeven(const uint8_t *const input, int length,
+                          uint8_t *output) {
+  // Actual filter len = 2 * filter_len_half.
+  const int16_t *filter = av1_down2_symeven_half_filter;
+  const int filter_len_half = sizeof(av1_down2_symeven_half_filter) / 2;
+  int i, j;
+  uint8_t *optr = output;
+  int l1 = filter_len_half;
+  int l2 = (length - filter_len_half);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+               filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+               filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  }
+}
+
+static void down2_symodd(const uint8_t *const input, int length,
+                         uint8_t *output) {
+  // Actual filter len = 2 * filter_len_half - 1.
+  const int16_t *filter = av1_down2_symodd_half_filter;
+  const int filter_len_half = sizeof(av1_down2_symodd_half_filter) / 2;
+  int i, j;
+  uint8_t *optr = output;
+  int l1 = filter_len_half - 1;
+  int l2 = (length - filter_len_half + 1);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + j >= length ? length - 1 : i + j)]) *
+               filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[(i + j >= length ? length - 1 : i + j)]) *
+               filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel(sum);
+    }
+  }
+}
+
+static int get_down2_length(int length, int steps) {
+  int s;
+  for (s = 0; s < steps; ++s) length = (length + 1) >> 1;
+  return length;
+}
+
+static int get_down2_steps(int in_length, int out_length) {
+  int steps = 0;
+  int proj_in_length;
+  while ((proj_in_length = get_down2_length(in_length, 1)) >= out_length) {
+    ++steps;
+    in_length = proj_in_length;
+  }
+  return steps;
+}
+
+static void resize_multistep(const uint8_t *const input, int length,
+                             uint8_t *output, int olength, uint8_t *otmp) {
+  int steps;
+  if (length == olength) {
+    memcpy(output, input, sizeof(output[0]) * length);
+    return;
+  }
+  steps = get_down2_steps(length, olength);
+
+  if (steps > 0) {
+    int s;
+    uint8_t *out = NULL;
+    uint8_t *otmp2;
+    int filteredlength = length;
+
+    assert(otmp != NULL);
+    otmp2 = otmp + get_down2_length(length, 1);
+    for (s = 0; s < steps; ++s) {
+      const int proj_filteredlength = get_down2_length(filteredlength, 1);
+      const uint8_t *const in = (s == 0 ? input : out);
+      if (s == steps - 1 && proj_filteredlength == olength)
+        out = output;
+      else
+        out = (s & 1 ? otmp2 : otmp);
+      if (filteredlength & 1)
+        down2_symodd(in, filteredlength, out);
+      else
+        down2_symeven(in, filteredlength, out);
+      filteredlength = proj_filteredlength;
+    }
+    if (filteredlength != olength) {
+      interpolate(out, filteredlength, output, olength);
+    }
+  } else {
+    interpolate(input, length, output, olength);
+  }
+}
+
+static void fill_col_to_arr(uint8_t *img, int stride, int len, uint8_t *arr) {
+  int i;
+  uint8_t *iptr = img;
+  uint8_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *aptr++ = *iptr;
+  }
+}
+
+static void fill_arr_to_col(uint8_t *img, int stride, int len, uint8_t *arr) {
+  int i;
+  uint8_t *iptr = img;
+  uint8_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *iptr = *aptr++;
+  }
+}
+
+void av1_resize_plane(const uint8_t *const input, int height, int width,
+                      int in_stride, uint8_t *output, int height2, int width2,
+                      int out_stride) {
+  int i;
+  uint8_t *intbuf = (uint8_t *)malloc(sizeof(uint8_t) * width2 * height);
+  uint8_t *tmpbuf =
+      (uint8_t *)malloc(sizeof(uint8_t) * (width < height ? height : width));
+  uint8_t *arrbuf = (uint8_t *)malloc(sizeof(uint8_t) * height);
+  uint8_t *arrbuf2 = (uint8_t *)malloc(sizeof(uint8_t) * height2);
+  if (intbuf == NULL || tmpbuf == NULL || arrbuf == NULL || arrbuf2 == NULL)
+    goto Error;
+  assert(width > 0);
+  assert(height > 0);
+  assert(width2 > 0);
+  assert(height2 > 0);
+  for (i = 0; i < height; ++i)
+    resize_multistep(input + in_stride * i, width, intbuf + width2 * i, width2,
+                     tmpbuf);
+  for (i = 0; i < width2; ++i) {
+    fill_col_to_arr(intbuf + i, width2, height, arrbuf);
+    resize_multistep(arrbuf, height, arrbuf2, height2, tmpbuf);
+    fill_arr_to_col(output + i, out_stride, height2, arrbuf2);
+  }
+
+Error:
+  free(intbuf);
+  free(tmpbuf);
+  free(arrbuf);
+  free(arrbuf2);
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_interpolate(const uint16_t *const input, int inlength,
+                               uint16_t *output, int outlength, int bd) {
+  const int64_t delta =
+      (((uint64_t)inlength << 32) + outlength / 2) / outlength;
+  const int64_t offset =
+      inlength > outlength
+          ? (((int64_t)(inlength - outlength) << 31) + outlength / 2) /
+                outlength
+          : -(((int64_t)(outlength - inlength) << 31) + outlength / 2) /
+                outlength;
+  uint16_t *optr = output;
+  int x, x1, x2, sum, k, int_pel, sub_pel;
+  int64_t y;
+
+  const interp_kernel *interp_filters =
+      choose_interp_filter(inlength, outlength);
+
+  x = 0;
+  y = offset;
+  while ((y >> INTERP_PRECISION_BITS) < (INTERP_TAPS / 2 - 1)) {
+    x++;
+    y += delta;
+  }
+  x1 = x;
+  x = outlength - 1;
+  y = delta * x + offset;
+  while ((y >> INTERP_PRECISION_BITS) + (int64_t)(INTERP_TAPS / 2) >=
+         inlength) {
+    x--;
+    y -= delta;
+  }
+  x2 = x;
+  if (x1 > x2) {
+    for (x = 0, y = offset; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k) {
+        const int pk = int_pel - INTERP_TAPS / 2 + 1 + k;
+        sum += filter[k] *
+               input[(pk < 0 ? 0 : (pk >= inlength ? inlength - 1 : pk))];
+      }
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+  } else {
+    // Initial part.
+    for (x = 0, y = offset; x < x1; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k < 0
+                                      ? 0
+                                      : int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+    // Middle part.
+    for (; x <= x2; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[int_pel - INTERP_TAPS / 2 + 1 + k];
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+    // End part.
+    for (; x < outlength; ++x, y += delta) {
+      const int16_t *filter;
+      int_pel = y >> INTERP_PRECISION_BITS;
+      sub_pel = (y >> (INTERP_PRECISION_BITS - SUBPEL_BITS)) & SUBPEL_MASK;
+      filter = interp_filters[sub_pel];
+      sum = 0;
+      for (k = 0; k < INTERP_TAPS; ++k)
+        sum += filter[k] * input[(int_pel - INTERP_TAPS / 2 + 1 + k >= inlength
+                                      ? inlength - 1
+                                      : int_pel - INTERP_TAPS / 2 + 1 + k)];
+      *optr++ = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+    }
+  }
+}
+
+static void highbd_down2_symeven(const uint16_t *const input, int length,
+                                 uint16_t *output, int bd) {
+  // Actual filter len = 2 * filter_len_half.
+  static const int16_t *filter = av1_down2_symeven_half_filter;
+  const int filter_len_half = sizeof(av1_down2_symeven_half_filter) / 2;
+  int i, j;
+  uint16_t *optr = output;
+  int l1 = filter_len_half;
+  int l2 = (length - filter_len_half);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+               filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + 1 + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1));
+      for (j = 0; j < filter_len_half; ++j) {
+        sum += (input[i - j] +
+                input[(i + 1 + j >= length ? length - 1 : i + 1 + j)]) *
+               filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  }
+}
+
+static void highbd_down2_symodd(const uint16_t *const input, int length,
+                                uint16_t *output, int bd) {
+  // Actual filter len = 2 * filter_len_half - 1.
+  static const int16_t *filter = av1_down2_symodd_half_filter;
+  const int filter_len_half = sizeof(av1_down2_symodd_half_filter) / 2;
+  int i, j;
+  uint16_t *optr = output;
+  int l1 = filter_len_half - 1;
+  int l2 = (length - filter_len_half + 1);
+  l1 += (l1 & 1);
+  l2 += (l2 & 1);
+  if (l1 > l2) {
+    // Short input length.
+    for (i = 0; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] +
+                input[(i + j >= length ? length - 1 : i + j)]) *
+               filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  } else {
+    // Initial part.
+    for (i = 0; i < l1; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[(i - j < 0 ? 0 : i - j)] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // Middle part.
+    for (; i < l2; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[i + j]) * filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+    // End part.
+    for (; i < length; i += 2) {
+      int sum = (1 << (FILTER_BITS - 1)) + input[i] * filter[0];
+      for (j = 1; j < filter_len_half; ++j) {
+        sum += (input[i - j] + input[(i + j >= length ? length - 1 : i + j)]) *
+               filter[j];
+      }
+      sum >>= FILTER_BITS;
+      *optr++ = clip_pixel_highbd(sum, bd);
+    }
+  }
+}
+
+static void highbd_resize_multistep(const uint16_t *const input, int length,
+                                    uint16_t *output, int olength,
+                                    uint16_t *otmp, int bd) {
+  int steps;
+  if (length == olength) {
+    memcpy(output, input, sizeof(output[0]) * length);
+    return;
+  }
+  steps = get_down2_steps(length, olength);
+
+  if (steps > 0) {
+    int s;
+    uint16_t *out = NULL;
+    uint16_t *otmp2;
+    int filteredlength = length;
+
+    assert(otmp != NULL);
+    otmp2 = otmp + get_down2_length(length, 1);
+    for (s = 0; s < steps; ++s) {
+      const int proj_filteredlength = get_down2_length(filteredlength, 1);
+      const uint16_t *const in = (s == 0 ? input : out);
+      if (s == steps - 1 && proj_filteredlength == olength)
+        out = output;
+      else
+        out = (s & 1 ? otmp2 : otmp);
+      if (filteredlength & 1)
+        highbd_down2_symodd(in, filteredlength, out, bd);
+      else
+        highbd_down2_symeven(in, filteredlength, out, bd);
+      filteredlength = proj_filteredlength;
+    }
+    if (filteredlength != olength) {
+      highbd_interpolate(out, filteredlength, output, olength, bd);
+    }
+  } else {
+    highbd_interpolate(input, length, output, olength, bd);
+  }
+}
+
+static void highbd_fill_col_to_arr(uint16_t *img, int stride, int len,
+                                   uint16_t *arr) {
+  int i;
+  uint16_t *iptr = img;
+  uint16_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *aptr++ = *iptr;
+  }
+}
+
+static void highbd_fill_arr_to_col(uint16_t *img, int stride, int len,
+                                   uint16_t *arr) {
+  int i;
+  uint16_t *iptr = img;
+  uint16_t *aptr = arr;
+  for (i = 0; i < len; ++i, iptr += stride) {
+    *iptr = *aptr++;
+  }
+}
+
+void av1_highbd_resize_plane(const uint8_t *const input, int height, int width,
+                             int in_stride, uint8_t *output, int height2,
+                             int width2, int out_stride, int bd) {
+  int i;
+  uint16_t *intbuf = (uint16_t *)malloc(sizeof(uint16_t) * width2 * height);
+  uint16_t *tmpbuf =
+      (uint16_t *)malloc(sizeof(uint16_t) * (width < height ? height : width));
+  uint16_t *arrbuf = (uint16_t *)malloc(sizeof(uint16_t) * height);
+  uint16_t *arrbuf2 = (uint16_t *)malloc(sizeof(uint16_t) * height2);
+  if (intbuf == NULL || tmpbuf == NULL || arrbuf == NULL || arrbuf2 == NULL)
+    goto Error;
+  for (i = 0; i < height; ++i) {
+    highbd_resize_multistep(CONVERT_TO_SHORTPTR(input + in_stride * i), width,
+                            intbuf + width2 * i, width2, tmpbuf, bd);
+  }
+  for (i = 0; i < width2; ++i) {
+    highbd_fill_col_to_arr(intbuf + i, width2, height, arrbuf);
+    highbd_resize_multistep(arrbuf, height, arrbuf2, height2, tmpbuf, bd);
+    highbd_fill_arr_to_col(CONVERT_TO_SHORTPTR(output + i), out_stride, height2,
+                           arrbuf2);
+  }
+
+Error:
+  free(intbuf);
+  free(tmpbuf);
+  free(arrbuf);
+  free(arrbuf2);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_resize_frame420(const uint8_t *const y, int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride, int height, int width, uint8_t *oy,
+                         int oy_stride, uint8_t *ou, uint8_t *ov,
+                         int ouv_stride, int oheight, int owidth) {
+  av1_resize_plane(y, height, width, y_stride, oy, oheight, owidth, oy_stride);
+  av1_resize_plane(u, height / 2, width / 2, uv_stride, ou, oheight / 2,
+                   owidth / 2, ouv_stride);
+  av1_resize_plane(v, height / 2, width / 2, uv_stride, ov, oheight / 2,
+                   owidth / 2, ouv_stride);
+}
+
+void av1_resize_frame422(const uint8_t *const y, int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride, int height, int width, uint8_t *oy,
+                         int oy_stride, uint8_t *ou, uint8_t *ov,
+                         int ouv_stride, int oheight, int owidth) {
+  av1_resize_plane(y, height, width, y_stride, oy, oheight, owidth, oy_stride);
+  av1_resize_plane(u, height, width / 2, uv_stride, ou, oheight, owidth / 2,
+                   ouv_stride);
+  av1_resize_plane(v, height, width / 2, uv_stride, ov, oheight, owidth / 2,
+                   ouv_stride);
+}
+
+void av1_resize_frame444(const uint8_t *const y, int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride, int height, int width, uint8_t *oy,
+                         int oy_stride, uint8_t *ou, uint8_t *ov,
+                         int ouv_stride, int oheight, int owidth) {
+  av1_resize_plane(y, height, width, y_stride, oy, oheight, owidth, oy_stride);
+  av1_resize_plane(u, height, width, uv_stride, ou, oheight, owidth,
+                   ouv_stride);
+  av1_resize_plane(v, height, width, uv_stride, ov, oheight, owidth,
+                   ouv_stride);
+}
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_resize_frame420(const uint8_t *const y, int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride, int height, int width,
+                                uint8_t *oy, int oy_stride, uint8_t *ou,
+                                uint8_t *ov, int ouv_stride, int oheight,
+                                int owidth, int bd) {
+  av1_highbd_resize_plane(y, height, width, y_stride, oy, oheight, owidth,
+                          oy_stride, bd);
+  av1_highbd_resize_plane(u, height / 2, width / 2, uv_stride, ou, oheight / 2,
+                          owidth / 2, ouv_stride, bd);
+  av1_highbd_resize_plane(v, height / 2, width / 2, uv_stride, ov, oheight / 2,
+                          owidth / 2, ouv_stride, bd);
+}
+
+void av1_highbd_resize_frame422(const uint8_t *const y, int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride, int height, int width,
+                                uint8_t *oy, int oy_stride, uint8_t *ou,
+                                uint8_t *ov, int ouv_stride, int oheight,
+                                int owidth, int bd) {
+  av1_highbd_resize_plane(y, height, width, y_stride, oy, oheight, owidth,
+                          oy_stride, bd);
+  av1_highbd_resize_plane(u, height, width / 2, uv_stride, ou, oheight,
+                          owidth / 2, ouv_stride, bd);
+  av1_highbd_resize_plane(v, height, width / 2, uv_stride, ov, oheight,
+                          owidth / 2, ouv_stride, bd);
+}
+
+void av1_highbd_resize_frame444(const uint8_t *const y, int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride, int height, int width,
+                                uint8_t *oy, int oy_stride, uint8_t *ou,
+                                uint8_t *ov, int ouv_stride, int oheight,
+                                int owidth, int bd) {
+  av1_highbd_resize_plane(y, height, width, y_stride, oy, oheight, owidth,
+                          oy_stride, bd);
+  av1_highbd_resize_plane(u, height, width, uv_stride, ou, oheight, owidth,
+                          ouv_stride, bd);
+  av1_highbd_resize_plane(v, height, width, uv_stride, ov, oheight, owidth,
+                          ouv_stride, bd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/av1/common/resize.h b/third_party/aom/av1/common/resize.h
new file mode 100644
index 0000000000..959cda969a
--- /dev/null
+++ b/third_party/aom/av1/common/resize.h
@@ -0,0 +1,69 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_RESIZE_H_
+#define AV1_ENCODER_RESIZE_H_
+
+#include <stdio.h>
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_resize_plane(const uint8_t *const input, int height, int width,
+                      int in_stride, uint8_t *output, int height2, int width2,
+                      int out_stride);
+void av1_resize_frame420(const uint8_t *const y, int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride, int height, int width, uint8_t *oy,
+                         int oy_stride, uint8_t *ou, uint8_t *ov,
+                         int ouv_stride, int oheight, int owidth);
+void av1_resize_frame422(const uint8_t *const y, int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride, int height, int width, uint8_t *oy,
+                         int oy_stride, uint8_t *ou, uint8_t *ov,
+                         int ouv_stride, int oheight, int owidth);
+void av1_resize_frame444(const uint8_t *const y, int y_stride,
+                         const uint8_t *const u, const uint8_t *const v,
+                         int uv_stride, int height, int width, uint8_t *oy,
+                         int oy_stride, uint8_t *ou, uint8_t *ov,
+                         int ouv_stride, int oheight, int owidth);
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_resize_plane(const uint8_t *const input, int height, int width,
+                             int in_stride, uint8_t *output, int height2,
+                             int width2, int out_stride, int bd);
+void av1_highbd_resize_frame420(const uint8_t *const y, int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride, int height, int width,
+                                uint8_t *oy, int oy_stride, uint8_t *ou,
+                                uint8_t *ov, int ouv_stride, int oheight,
+                                int owidth, int bd);
+void av1_highbd_resize_frame422(const uint8_t *const y, int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride, int height, int width,
+                                uint8_t *oy, int oy_stride, uint8_t *ou,
+                                uint8_t *ov, int ouv_stride, int oheight,
+                                int owidth, int bd);
+void av1_highbd_resize_frame444(const uint8_t *const y, int y_stride,
+                                const uint8_t *const u, const uint8_t *const v,
+                                int uv_stride, int height, int width,
+                                uint8_t *oy, int oy_stride, uint8_t *ou,
+                                uint8_t *ov, int ouv_stride, int oheight,
+                                int owidth, int bd);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_RESIZE_H_
diff --git a/third_party/aom/av1/common/restoration.c b/third_party/aom/av1/common/restoration.c
new file mode 100644
index 0000000000..b7ed9f98bc
--- /dev/null
+++ b/third_party/aom/av1/common/restoration.c
@@ -0,0 +1,1401 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ *
+ */
+
+#include <math.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "./aom_scale_rtcd.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/restoration.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+const sgr_params_type sgr_params[SGRPROJ_PARAMS] = {
+#if USE_HIGHPASS_IN_SGRPROJ
+  // corner, edge, r2, eps2
+  { -1, 2, 1, 1 }, { -1, 2, 1, 2 }, { -1, 2, 1, 3 }, { -1, 2, 1, 4 },
+  { -1, 2, 1, 5 }, { -2, 3, 1, 2 }, { -2, 3, 1, 3 }, { -2, 3, 1, 4 },
+  { -2, 3, 1, 5 }, { -2, 3, 1, 6 }, { -3, 4, 1, 3 }, { -3, 4, 1, 4 },
+  { -3, 4, 1, 5 }, { -3, 4, 1, 6 }, { -3, 4, 1, 7 }, { -3, 4, 1, 8 }
+#else
+  // r1, eps1, r2, eps2
+  { 2, 12, 1, 4 },  { 2, 15, 1, 6 },  { 2, 18, 1, 8 },  { 2, 20, 1, 9 },
+  { 2, 22, 1, 10 }, { 2, 25, 1, 11 }, { 2, 35, 1, 12 }, { 2, 45, 1, 13 },
+  { 2, 55, 1, 14 }, { 2, 65, 1, 15 }, { 2, 75, 1, 16 }, { 3, 30, 1, 10 },
+  { 3, 50, 1, 12 }, { 3, 50, 2, 25 }, { 3, 60, 2, 35 }, { 3, 70, 2, 45 },
+#endif
+};
+
+typedef void (*restore_func_type)(uint8_t *data8, int width, int height,
+                                  int stride, RestorationInternal *rst,
+                                  uint8_t *dst8, int dst_stride);
+#if CONFIG_HIGHBITDEPTH
+typedef void (*restore_func_highbd_type)(uint8_t *data8, int width, int height,
+                                         int stride, RestorationInternal *rst,
+                                         int bit_depth, uint8_t *dst8,
+                                         int dst_stride);
+#endif  // CONFIG_HIGHBITDEPTH
+
+int av1_alloc_restoration_struct(AV1_COMMON *cm, RestorationInfo *rst_info,
+                                 int width, int height) {
+  const int ntiles = av1_get_rest_ntiles(
+      width, height, rst_info->restoration_tilesize, NULL, NULL, NULL, NULL);
+  aom_free(rst_info->restoration_type);
+  CHECK_MEM_ERROR(cm, rst_info->restoration_type,
+                  (RestorationType *)aom_malloc(
+                      sizeof(*rst_info->restoration_type) * ntiles));
+  aom_free(rst_info->wiener_info);
+  CHECK_MEM_ERROR(
+      cm, rst_info->wiener_info,
+      (WienerInfo *)aom_memalign(16, sizeof(*rst_info->wiener_info) * ntiles));
+  memset(rst_info->wiener_info, 0, sizeof(*rst_info->wiener_info) * ntiles);
+  aom_free(rst_info->sgrproj_info);
+  CHECK_MEM_ERROR(
+      cm, rst_info->sgrproj_info,
+      (SgrprojInfo *)aom_malloc(sizeof(*rst_info->sgrproj_info) * ntiles));
+  return ntiles;
+}
+
+void av1_free_restoration_struct(RestorationInfo *rst_info) {
+  aom_free(rst_info->restoration_type);
+  rst_info->restoration_type = NULL;
+  aom_free(rst_info->wiener_info);
+  rst_info->wiener_info = NULL;
+  aom_free(rst_info->sgrproj_info);
+  rst_info->sgrproj_info = NULL;
+}
+
+#define MAX_RADIUS 3  // Only 1, 2, 3 allowed
+#define MAX_EPS 80    // Max value of eps
+#define MAX_NELEM ((2 * MAX_RADIUS + 1) * (2 * MAX_RADIUS + 1))
+#define SGRPROJ_MTABLE_BITS 20
+#define SGRPROJ_RECIP_BITS 12
+
+// TODO(debargha): This table can be substantially reduced since only a few
+// values are actually used.
+int sgrproj_mtable[MAX_EPS][MAX_NELEM];
+
+static void GenSgrprojVtable() {
+  int e, n;
+  for (e = 1; e <= MAX_EPS; ++e)
+    for (n = 1; n <= MAX_NELEM; ++n) {
+      const int n2e = n * n * e;
+      sgrproj_mtable[e - 1][n - 1] =
+          (((1 << SGRPROJ_MTABLE_BITS) + n2e / 2) / n2e);
+    }
+}
+
+void av1_loop_restoration_precal() { GenSgrprojVtable(); }
+
+static void loop_restoration_init(RestorationInternal *rst, int kf) {
+  rst->keyframe = kf;
+}
+
+void extend_frame(uint8_t *data, int width, int height, int stride) {
+  uint8_t *data_p;
+  int i;
+  for (i = 0; i < height; ++i) {
+    data_p = data + i * stride;
+    memset(data_p - WIENER_HALFWIN, data_p[0], WIENER_HALFWIN);
+    memset(data_p + width, data_p[width - 1], WIENER_HALFWIN);
+  }
+  data_p = data - WIENER_HALFWIN;
+  for (i = -WIENER_HALFWIN; i < 0; ++i) {
+    memcpy(data_p + i * stride, data_p, width + 2 * WIENER_HALFWIN);
+  }
+  for (i = height; i < height + WIENER_HALFWIN; ++i) {
+    memcpy(data_p + i * stride, data_p + (height - 1) * stride,
+           width + 2 * WIENER_HALFWIN);
+  }
+}
+
+static void loop_copy_tile(uint8_t *data, int tile_idx, int subtile_idx,
+                           int subtile_bits, int width, int height, int stride,
+                           RestorationInternal *rst, uint8_t *dst,
+                           int dst_stride) {
+  const int tile_width = rst->tile_width;
+  const int tile_height = rst->tile_height;
+  int i;
+  int h_start, h_end, v_start, v_end;
+  av1_get_rest_tile_limits(tile_idx, subtile_idx, subtile_bits, rst->nhtiles,
+                           rst->nvtiles, tile_width, tile_height, width, height,
+                           0, 0, &h_start, &h_end, &v_start, &v_end);
+  for (i = v_start; i < v_end; ++i)
+    memcpy(dst + i * dst_stride + h_start, data + i * stride + h_start,
+           h_end - h_start);
+}
+
+static void loop_wiener_filter_tile(uint8_t *data, int tile_idx, int width,
+                                    int height, int stride,
+                                    RestorationInternal *rst, uint8_t *dst,
+                                    int dst_stride) {
+  const int tile_width = rst->tile_width;
+  const int tile_height = rst->tile_height;
+  int i, j;
+  int h_start, h_end, v_start, v_end;
+  if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
+    loop_copy_tile(data, tile_idx, 0, 0, width, height, stride, rst, dst,
+                   dst_stride);
+    return;
+  }
+  av1_get_rest_tile_limits(tile_idx, 0, 0, rst->nhtiles, rst->nvtiles,
+                           tile_width, tile_height, width, height, 0, 0,
+                           &h_start, &h_end, &v_start, &v_end);
+  // Convolve the whole tile (done in blocks here to match the requirements
+  // of the vectorized convolve functions, but the result is equivalent)
+  for (i = v_start; i < v_end; i += MAX_SB_SIZE)
+    for (j = h_start; j < h_end; j += MAX_SB_SIZE) {
+      int w = AOMMIN(MAX_SB_SIZE, (h_end - j + 15) & ~15);
+      int h = AOMMIN(MAX_SB_SIZE, (v_end - i + 15) & ~15);
+      const uint8_t *data_p = data + i * stride + j;
+      uint8_t *dst_p = dst + i * dst_stride + j;
+      aom_convolve8_add_src(data_p, stride, dst_p, dst_stride,
+                            rst->rsi->wiener_info[tile_idx].hfilter, 16,
+                            rst->rsi->wiener_info[tile_idx].vfilter, 16, w, h);
+    }
+}
+
+static void loop_wiener_filter(uint8_t *data, int width, int height, int stride,
+                               RestorationInternal *rst, uint8_t *dst,
+                               int dst_stride) {
+  int tile_idx;
+  extend_frame(data, width, height, stride);
+  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
+    loop_wiener_filter_tile(data, tile_idx, width, height, stride, rst, dst,
+                            dst_stride);
+  }
+}
+
+/* Calculate windowed sums (if sqr=0) or sums of squares (if sqr=1)
+   over the input. The window is of size (2r + 1)x(2r + 1), and we
+   specialize to r = 1, 2, 3. A default function is used for r > 3.
+
+   Each loop follows the same format: We keep a window's worth of input
+   in individual variables and select data out of that as appropriate.
+*/
+static void boxsum1(int32_t *src, int width, int height, int src_stride,
+                    int sqr, int32_t *dst, int dst_stride) {
+  int i, j, a, b, c;
+
+  // Vertical sum over 3-pixel regions, from src into dst.
+  if (!sqr) {
+    for (j = 0; j < width; ++j) {
+      a = src[j];
+      b = src[src_stride + j];
+      c = src[2 * src_stride + j];
+
+      dst[j] = a + b;
+      for (i = 1; i < height - 2; ++i) {
+        // Loop invariant: At the start of each iteration,
+        // a = src[(i - 1) * src_stride + j]
+        // b = src[(i    ) * src_stride + j]
+        // c = src[(i + 1) * src_stride + j]
+        dst[i * dst_stride + j] = a + b + c;
+        a = b;
+        b = c;
+        c = src[(i + 2) * src_stride + j];
+      }
+      dst[i * dst_stride + j] = a + b + c;
+      dst[(i + 1) * dst_stride + j] = b + c;
+    }
+  } else {
+    for (j = 0; j < width; ++j) {
+      a = src[j] * src[j];
+      b = src[src_stride + j] * src[src_stride + j];
+      c = src[2 * src_stride + j] * src[2 * src_stride + j];
+
+      dst[j] = a + b;
+      for (i = 1; i < height - 2; ++i) {
+        dst[i * dst_stride + j] = a + b + c;
+        a = b;
+        b = c;
+        c = src[(i + 2) * src_stride + j] * src[(i + 2) * src_stride + j];
+      }
+      dst[i * dst_stride + j] = a + b + c;
+      dst[(i + 1) * dst_stride + j] = b + c;
+    }
+  }
+
+  // Horizontal sum over 3-pixel regions of dst
+  for (i = 0; i < height; ++i) {
+    a = dst[i * dst_stride];
+    b = dst[i * dst_stride + 1];
+    c = dst[i * dst_stride + 2];
+
+    dst[i * dst_stride] = a + b;
+    for (j = 1; j < width - 2; ++j) {
+      // Loop invariant: At the start of each iteration,
+      // a = src[i * src_stride + (j - 1)]
+      // b = src[i * src_stride + (j    )]
+      // c = src[i * src_stride + (j + 1)]
+      dst[i * dst_stride + j] = a + b + c;
+      a = b;
+      b = c;
+      c = dst[i * dst_stride + (j + 2)];
+    }
+    dst[i * dst_stride + j] = a + b + c;
+    dst[i * dst_stride + (j + 1)] = b + c;
+  }
+}
+
+static void boxsum2(int32_t *src, int width, int height, int src_stride,
+                    int sqr, int32_t *dst, int dst_stride) {
+  int i, j, a, b, c, d, e;
+
+  // Vertical sum over 5-pixel regions, from src into dst.
+  if (!sqr) {
+    for (j = 0; j < width; ++j) {
+      a = src[j];
+      b = src[src_stride + j];
+      c = src[2 * src_stride + j];
+      d = src[3 * src_stride + j];
+      e = src[4 * src_stride + j];
+
+      dst[j] = a + b + c;
+      dst[dst_stride + j] = a + b + c + d;
+      for (i = 2; i < height - 3; ++i) {
+        // Loop invariant: At the start of each iteration,
+        // a = src[(i - 2) * src_stride + j]
+        // b = src[(i - 1) * src_stride + j]
+        // c = src[(i    ) * src_stride + j]
+        // d = src[(i + 1) * src_stride + j]
+        // e = src[(i + 2) * src_stride + j]
+        dst[i * dst_stride + j] = a + b + c + d + e;
+        a = b;
+        b = c;
+        c = d;
+        d = e;
+        e = src[(i + 3) * src_stride + j];
+      }
+      dst[i * dst_stride + j] = a + b + c + d + e;
+      dst[(i + 1) * dst_stride + j] = b + c + d + e;
+      dst[(i + 2) * dst_stride + j] = c + d + e;
+    }
+  } else {
+    for (j = 0; j < width; ++j) {
+      a = src[j] * src[j];
+      b = src[src_stride + j] * src[src_stride + j];
+      c = src[2 * src_stride + j] * src[2 * src_stride + j];
+      d = src[3 * src_stride + j] * src[3 * src_stride + j];
+      e = src[4 * src_stride + j] * src[4 * src_stride + j];
+
+      dst[j] = a + b + c;
+      dst[dst_stride + j] = a + b + c + d;
+      for (i = 2; i < height - 3; ++i) {
+        dst[i * dst_stride + j] = a + b + c + d + e;
+        a = b;
+        b = c;
+        c = d;
+        d = e;
+        e = src[(i + 3) * src_stride + j] * src[(i + 3) * src_stride + j];
+      }
+      dst[i * dst_stride + j] = a + b + c + d + e;
+      dst[(i + 1) * dst_stride + j] = b + c + d + e;
+      dst[(i + 2) * dst_stride + j] = c + d + e;
+    }
+  }
+
+  // Horizontal sum over 5-pixel regions of dst
+  for (i = 0; i < height; ++i) {
+    a = dst[i * dst_stride];
+    b = dst[i * dst_stride + 1];
+    c = dst[i * dst_stride + 2];
+    d = dst[i * dst_stride + 3];
+    e = dst[i * dst_stride + 4];
+
+    dst[i * dst_stride] = a + b + c;
+    dst[i * dst_stride + 1] = a + b + c + d;
+    for (j = 2; j < width - 3; ++j) {
+      // Loop invariant: At the start of each iteration,
+      // a = src[i * src_stride + (j - 2)]
+      // b = src[i * src_stride + (j - 1)]
+      // c = src[i * src_stride + (j    )]
+      // d = src[i * src_stride + (j + 1)]
+      // e = src[i * src_stride + (j + 2)]
+      dst[i * dst_stride + j] = a + b + c + d + e;
+      a = b;
+      b = c;
+      c = d;
+      d = e;
+      e = dst[i * dst_stride + (j + 3)];
+    }
+    dst[i * dst_stride + j] = a + b + c + d + e;
+    dst[i * dst_stride + (j + 1)] = b + c + d + e;
+    dst[i * dst_stride + (j + 2)] = c + d + e;
+  }
+}
+
+static void boxsum3(int32_t *src, int width, int height, int src_stride,
+                    int sqr, int32_t *dst, int dst_stride) {
+  int i, j, a, b, c, d, e, f, g;
+
+  // Vertical sum over 7-pixel regions, from src into dst.
+  if (!sqr) {
+    for (j = 0; j < width; ++j) {
+      a = src[j];
+      b = src[1 * src_stride + j];
+      c = src[2 * src_stride + j];
+      d = src[3 * src_stride + j];
+      e = src[4 * src_stride + j];
+      f = src[5 * src_stride + j];
+      g = src[6 * src_stride + j];
+
+      dst[j] = a + b + c + d;
+      dst[dst_stride + j] = a + b + c + d + e;
+      dst[2 * dst_stride + j] = a + b + c + d + e + f;
+      for (i = 3; i < height - 4; ++i) {
+        dst[i * dst_stride + j] = a + b + c + d + e + f + g;
+        a = b;
+        b = c;
+        c = d;
+        d = e;
+        e = f;
+        f = g;
+        g = src[(i + 4) * src_stride + j];
+      }
+      dst[i * dst_stride + j] = a + b + c + d + e + f + g;
+      dst[(i + 1) * dst_stride + j] = b + c + d + e + f + g;
+      dst[(i + 2) * dst_stride + j] = c + d + e + f + g;
+      dst[(i + 3) * dst_stride + j] = d + e + f + g;
+    }
+  } else {
+    for (j = 0; j < width; ++j) {
+      a = src[j] * src[j];
+      b = src[1 * src_stride + j] * src[1 * src_stride + j];
+      c = src[2 * src_stride + j] * src[2 * src_stride + j];
+      d = src[3 * src_stride + j] * src[3 * src_stride + j];
+      e = src[4 * src_stride + j] * src[4 * src_stride + j];
+      f = src[5 * src_stride + j] * src[5 * src_stride + j];
+      g = src[6 * src_stride + j] * src[6 * src_stride + j];
+
+      dst[j] = a + b + c + d;
+      dst[dst_stride + j] = a + b + c + d + e;
+      dst[2 * dst_stride + j] = a + b + c + d + e + f;
+      for (i = 3; i < height - 4; ++i) {
+        dst[i * dst_stride + j] = a + b + c + d + e + f + g;
+        a = b;
+        b = c;
+        c = d;
+        d = e;
+        e = f;
+        f = g;
+        g = src[(i + 4) * src_stride + j] * src[(i + 4) * src_stride + j];
+      }
+      dst[i * dst_stride + j] = a + b + c + d + e + f + g;
+      dst[(i + 1) * dst_stride + j] = b + c + d + e + f + g;
+      dst[(i + 2) * dst_stride + j] = c + d + e + f + g;
+      dst[(i + 3) * dst_stride + j] = d + e + f + g;
+    }
+  }
+
+  // Horizontal sum over 7-pixel regions of dst
+  for (i = 0; i < height; ++i) {
+    a = dst[i * dst_stride];
+    b = dst[i * dst_stride + 1];
+    c = dst[i * dst_stride + 2];
+    d = dst[i * dst_stride + 3];
+    e = dst[i * dst_stride + 4];
+    f = dst[i * dst_stride + 5];
+    g = dst[i * dst_stride + 6];
+
+    dst[i * dst_stride] = a + b + c + d;
+    dst[i * dst_stride + 1] = a + b + c + d + e;
+    dst[i * dst_stride + 2] = a + b + c + d + e + f;
+    for (j = 3; j < width - 4; ++j) {
+      dst[i * dst_stride + j] = a + b + c + d + e + f + g;
+      a = b;
+      b = c;
+      c = d;
+      d = e;
+      e = f;
+      f = g;
+      g = dst[i * dst_stride + (j + 4)];
+    }
+    dst[i * dst_stride + j] = a + b + c + d + e + f + g;
+    dst[i * dst_stride + (j + 1)] = b + c + d + e + f + g;
+    dst[i * dst_stride + (j + 2)] = c + d + e + f + g;
+    dst[i * dst_stride + (j + 3)] = d + e + f + g;
+  }
+}
+
+// Generic version for any r. To be removed after experiments are done.
+static void boxsumr(int32_t *src, int width, int height, int src_stride, int r,
+                    int sqr, int32_t *dst, int dst_stride) {
+  int32_t *tmp = aom_malloc(width * height * sizeof(*tmp));
+  int tmp_stride = width;
+  int i, j;
+  if (sqr) {
+    for (j = 0; j < width; ++j) tmp[j] = src[j] * src[j];
+    for (j = 0; j < width; ++j)
+      for (i = 1; i < height; ++i)
+        tmp[i * tmp_stride + j] =
+            tmp[(i - 1) * tmp_stride + j] +
+            src[i * src_stride + j] * src[i * src_stride + j];
+  } else {
+    memcpy(tmp, src, sizeof(*tmp) * width);
+    for (j = 0; j < width; ++j)
+      for (i = 1; i < height; ++i)
+        tmp[i * tmp_stride + j] =
+            tmp[(i - 1) * tmp_stride + j] + src[i * src_stride + j];
+  }
+  for (i = 0; i <= r; ++i)
+    memcpy(&dst[i * dst_stride], &tmp[(i + r) * tmp_stride],
+           sizeof(*tmp) * width);
+  for (i = r + 1; i < height - r; ++i)
+    for (j = 0; j < width; ++j)
+      dst[i * dst_stride + j] =
+          tmp[(i + r) * tmp_stride + j] - tmp[(i - r - 1) * tmp_stride + j];
+  for (i = height - r; i < height; ++i)
+    for (j = 0; j < width; ++j)
+      dst[i * dst_stride + j] = tmp[(height - 1) * tmp_stride + j] -
+                                tmp[(i - r - 1) * tmp_stride + j];
+
+  for (i = 0; i < height; ++i) tmp[i * tmp_stride] = dst[i * dst_stride];
+  for (i = 0; i < height; ++i)
+    for (j = 1; j < width; ++j)
+      tmp[i * tmp_stride + j] =
+          tmp[i * tmp_stride + j - 1] + dst[i * src_stride + j];
+
+  for (j = 0; j <= r; ++j)
+    for (i = 0; i < height; ++i)
+      dst[i * dst_stride + j] = tmp[i * tmp_stride + j + r];
+  for (j = r + 1; j < width - r; ++j)
+    for (i = 0; i < height; ++i)
+      dst[i * dst_stride + j] =
+          tmp[i * tmp_stride + j + r] - tmp[i * tmp_stride + j - r - 1];
+  for (j = width - r; j < width; ++j)
+    for (i = 0; i < height; ++i)
+      dst[i * dst_stride + j] =
+          tmp[i * tmp_stride + width - 1] - tmp[i * tmp_stride + j - r - 1];
+  aom_free(tmp);
+}
+
+static void boxsum(int32_t *src, int width, int height, int src_stride, int r,
+                   int sqr, int32_t *dst, int dst_stride) {
+  if (r == 1)
+    boxsum1(src, width, height, src_stride, sqr, dst, dst_stride);
+  else if (r == 2)
+    boxsum2(src, width, height, src_stride, sqr, dst, dst_stride);
+  else if (r == 3)
+    boxsum3(src, width, height, src_stride, sqr, dst, dst_stride);
+  else
+    boxsumr(src, width, height, src_stride, r, sqr, dst, dst_stride);
+}
+
+static void boxnum(int width, int height, int r, int8_t *num, int num_stride) {
+  int i, j;
+  for (i = 0; i <= r; ++i) {
+    for (j = 0; j <= r; ++j) {
+      num[i * num_stride + j] = (r + 1 + i) * (r + 1 + j);
+      num[i * num_stride + (width - 1 - j)] = num[i * num_stride + j];
+      num[(height - 1 - i) * num_stride + j] = num[i * num_stride + j];
+      num[(height - 1 - i) * num_stride + (width - 1 - j)] =
+          num[i * num_stride + j];
+    }
+  }
+  for (j = 0; j <= r; ++j) {
+    const int val = (2 * r + 1) * (r + 1 + j);
+    for (i = r + 1; i < height - r; ++i) {
+      num[i * num_stride + j] = val;
+      num[i * num_stride + (width - 1 - j)] = val;
+    }
+  }
+  for (i = 0; i <= r; ++i) {
+    const int val = (2 * r + 1) * (r + 1 + i);
+    for (j = r + 1; j < width - r; ++j) {
+      num[i * num_stride + j] = val;
+      num[(height - 1 - i) * num_stride + j] = val;
+    }
+  }
+  for (i = r + 1; i < height - r; ++i) {
+    for (j = r + 1; j < width - r; ++j) {
+      num[i * num_stride + j] = (2 * r + 1) * (2 * r + 1);
+    }
+  }
+}
+
+void decode_xq(int *xqd, int *xq) {
+  xq[0] = xqd[0];
+  xq[1] = (1 << SGRPROJ_PRJ_BITS) - xq[0] - xqd[1];
+}
+
+const int32_t x_by_xplus1[256] = {
+  0,   128, 171, 192, 205, 213, 219, 224, 228, 230, 233, 235, 236, 238, 239,
+  240, 241, 242, 243, 243, 244, 244, 245, 245, 246, 246, 247, 247, 247, 247,
+  248, 248, 248, 248, 249, 249, 249, 249, 249, 250, 250, 250, 250, 250, 250,
+  250, 251, 251, 251, 251, 251, 251, 251, 251, 251, 251, 252, 252, 252, 252,
+  252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 253, 253,
+  253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253,
+  253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 254, 254, 254,
+  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
+  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
+  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
+  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
+  254, 254, 254, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
+  256,
+};
+
+const int32_t one_by_x[MAX_NELEM] = {
+  4096, 2048, 1365, 1024, 819, 683, 585, 512, 455, 410, 372, 341, 315,
+  293,  273,  256,  241,  228, 216, 205, 195, 186, 178, 171, 164, 158,
+  152,  146,  141,  137,  132, 128, 124, 120, 117, 114, 111, 108, 105,
+  102,  100,  98,   95,   93,  91,  89,  87,  85,  84
+};
+
+static void av1_selfguided_restoration_internal(int32_t *dgd, int width,
+                                                int height, int stride,
+                                                int bit_depth, int r, int eps,
+                                                int32_t *tmpbuf) {
+  int32_t *A = tmpbuf;
+  int32_t *B = A + SGRPROJ_OUTBUF_SIZE;
+  int8_t num[RESTORATION_TILEPELS_MAX];
+  int i, j;
+  // Adjusting the stride of A and B here appears to avoid bad cache effects,
+  // leading to a significant speed improvement.
+  // We also align the stride to a multiple of 16 bytes, for consistency
+  // with the SIMD version of this function.
+  int buf_stride = ((width + 3) & ~3) + 16;
+
+  // Don't filter tiles with dimensions < 5 on any axis
+  if ((width < 5) || (height < 5)) return;
+
+  boxsum(dgd, width, height, stride, r, 0, B, buf_stride);
+  boxsum(dgd, width, height, stride, r, 1, A, buf_stride);
+  boxnum(width, height, r, num, width);
+  assert(r <= 3);
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      const int k = i * buf_stride + j;
+      const int n = num[i * width + j];
+
+      // a < 2^16 * n < 2^22 regardless of bit depth
+      uint32_t a = ROUND_POWER_OF_TWO(A[k], 2 * (bit_depth - 8));
+      // b < 2^8 * n < 2^14 regardless of bit depth
+      uint32_t b = ROUND_POWER_OF_TWO(B[k], bit_depth - 8);
+
+      // Each term in calculating p = a * n - b * b is < 2^16 * n^2 < 2^28,
+      // and p itself satisfies p < 2^14 * n^2 < 2^26.
+      // Note: Sometimes, in high bit depth, we can end up with a*n < b*b.
+      // This is an artefact of rounding, and can only happen if all pixels
+      // are (almost) identical, so in this case we saturate to p=0.
+      uint32_t p = (a * n < b * b) ? 0 : a * n - b * b;
+      uint32_t s = sgrproj_mtable[eps - 1][n - 1];
+
+      // p * s < (2^14 * n^2) * round(2^20 / n^2 eps) < 2^34 / eps < 2^32
+      // as long as eps >= 4. So p * s fits into a uint32_t, and z < 2^12
+      // (this holds even after accounting for the rounding in s)
+      const uint32_t z = ROUND_POWER_OF_TWO(p * s, SGRPROJ_MTABLE_BITS);
+
+      A[k] = x_by_xplus1[AOMMIN(z, 255)];  // < 2^8
+
+      // SGRPROJ_SGR - A[k] < 2^8, B[k] < 2^(bit_depth) * n,
+      // one_by_x[n - 1] = round(2^12 / n)
+      // => the product here is < 2^(20 + bit_depth) <= 2^32,
+      // and B[k] is set to a value < 2^(8 + bit depth)
+      B[k] = (int32_t)ROUND_POWER_OF_TWO((uint32_t)(SGRPROJ_SGR - A[k]) *
+                                             (uint32_t)B[k] *
+                                             (uint32_t)one_by_x[n - 1],
+                                         SGRPROJ_RECIP_BITS);
+    }
+  }
+  i = 0;
+  j = 0;
+  {
+    const int k = i * buf_stride + j;
+    const int l = i * stride + j;
+    const int nb = 3;
+    const int32_t a =
+        3 * A[k] + 2 * A[k + 1] + 2 * A[k + buf_stride] + A[k + buf_stride + 1];
+    const int32_t b =
+        3 * B[k] + 2 * B[k + 1] + 2 * B[k + buf_stride] + B[k + buf_stride + 1];
+    const int32_t v = a * dgd[l] + b;
+    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+  }
+  i = 0;
+  j = width - 1;
+  {
+    const int k = i * buf_stride + j;
+    const int l = i * stride + j;
+    const int nb = 3;
+    const int32_t a =
+        3 * A[k] + 2 * A[k - 1] + 2 * A[k + buf_stride] + A[k + buf_stride - 1];
+    const int32_t b =
+        3 * B[k] + 2 * B[k - 1] + 2 * B[k + buf_stride] + B[k + buf_stride - 1];
+    const int32_t v = a * dgd[l] + b;
+    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+  }
+  i = height - 1;
+  j = 0;
+  {
+    const int k = i * buf_stride + j;
+    const int l = i * stride + j;
+    const int nb = 3;
+    const int32_t a =
+        3 * A[k] + 2 * A[k + 1] + 2 * A[k - buf_stride] + A[k - buf_stride + 1];
+    const int32_t b =
+        3 * B[k] + 2 * B[k + 1] + 2 * B[k - buf_stride] + B[k - buf_stride + 1];
+    const int32_t v = a * dgd[l] + b;
+    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+  }
+  i = height - 1;
+  j = width - 1;
+  {
+    const int k = i * buf_stride + j;
+    const int l = i * stride + j;
+    const int nb = 3;
+    const int32_t a =
+        3 * A[k] + 2 * A[k - 1] + 2 * A[k - buf_stride] + A[k - buf_stride - 1];
+    const int32_t b =
+        3 * B[k] + 2 * B[k - 1] + 2 * B[k - buf_stride] + B[k - buf_stride - 1];
+    const int32_t v = a * dgd[l] + b;
+    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+  }
+  i = 0;
+  for (j = 1; j < width - 1; ++j) {
+    const int k = i * buf_stride + j;
+    const int l = i * stride + j;
+    const int nb = 3;
+    const int32_t a = A[k] + 2 * (A[k - 1] + A[k + 1]) + A[k + buf_stride] +
+                      A[k + buf_stride - 1] + A[k + buf_stride + 1];
+    const int32_t b = B[k] + 2 * (B[k - 1] + B[k + 1]) + B[k + buf_stride] +
+                      B[k + buf_stride - 1] + B[k + buf_stride + 1];
+    const int32_t v = a * dgd[l] + b;
+    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+  }
+  i = height - 1;
+  for (j = 1; j < width - 1; ++j) {
+    const int k = i * buf_stride + j;
+    const int l = i * stride + j;
+    const int nb = 3;
+    const int32_t a = A[k] + 2 * (A[k - 1] + A[k + 1]) + A[k - buf_stride] +
+                      A[k - buf_stride - 1] + A[k - buf_stride + 1];
+    const int32_t b = B[k] + 2 * (B[k - 1] + B[k + 1]) + B[k - buf_stride] +
+                      B[k - buf_stride - 1] + B[k - buf_stride + 1];
+    const int32_t v = a * dgd[l] + b;
+    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+  }
+  j = 0;
+  for (i = 1; i < height - 1; ++i) {
+    const int k = i * buf_stride + j;
+    const int l = i * stride + j;
+    const int nb = 3;
+    const int32_t a = A[k] + 2 * (A[k - buf_stride] + A[k + buf_stride]) +
+                      A[k + 1] + A[k - buf_stride + 1] + A[k + buf_stride + 1];
+    const int32_t b = B[k] + 2 * (B[k - buf_stride] + B[k + buf_stride]) +
+                      B[k + 1] + B[k - buf_stride + 1] + B[k + buf_stride + 1];
+    const int32_t v = a * dgd[l] + b;
+    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+  }
+  j = width - 1;
+  for (i = 1; i < height - 1; ++i) {
+    const int k = i * buf_stride + j;
+    const int l = i * stride + j;
+    const int nb = 3;
+    const int32_t a = A[k] + 2 * (A[k - buf_stride] + A[k + buf_stride]) +
+                      A[k - 1] + A[k - buf_stride - 1] + A[k + buf_stride - 1];
+    const int32_t b = B[k] + 2 * (B[k - buf_stride] + B[k + buf_stride]) +
+                      B[k - 1] + B[k - buf_stride - 1] + B[k + buf_stride - 1];
+    const int32_t v = a * dgd[l] + b;
+    dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+  }
+  for (i = 1; i < height - 1; ++i) {
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int nb = 5;
+      const int32_t a =
+          (A[k] + A[k - 1] + A[k + 1] + A[k - buf_stride] + A[k + buf_stride]) *
+              4 +
+          (A[k - 1 - buf_stride] + A[k - 1 + buf_stride] +
+           A[k + 1 - buf_stride] + A[k + 1 + buf_stride]) *
+              3;
+      const int32_t b =
+          (B[k] + B[k - 1] + B[k + 1] + B[k - buf_stride] + B[k + buf_stride]) *
+              4 +
+          (B[k - 1 - buf_stride] + B[k - 1 + buf_stride] +
+           B[k + 1 - buf_stride] + B[k + 1 + buf_stride]) *
+              3;
+      const int32_t v = a * dgd[l] + b;
+      dgd[l] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+  }
+}
+
+void av1_selfguided_restoration_c(uint8_t *dgd, int width, int height,
+                                  int stride, int32_t *dst, int dst_stride,
+                                  int r, int eps, int32_t *tmpbuf) {
+  int i, j;
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      dst[i * dst_stride + j] = dgd[i * stride + j];
+    }
+  }
+  av1_selfguided_restoration_internal(dst, width, height, dst_stride, 8, r, eps,
+                                      tmpbuf);
+}
+
+void av1_highpass_filter_c(uint8_t *dgd, int width, int height, int stride,
+                           int32_t *dst, int dst_stride, int corner, int edge) {
+  int i, j;
+  const int center = (1 << SGRPROJ_RST_BITS) - 4 * (corner + edge);
+
+  i = 0;
+  j = 0;
+  {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] =
+        center * dgd[k] + edge * (dgd[k + 1] + dgd[k + stride] + dgd[k] * 2) +
+        corner * (dgd[k + stride + 1] + dgd[k + 1] + dgd[k + stride] + dgd[k]);
+  }
+  i = 0;
+  j = width - 1;
+  {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] =
+        center * dgd[k] + edge * (dgd[k - 1] + dgd[k + stride] + dgd[k] * 2) +
+        corner * (dgd[k + stride - 1] + dgd[k - 1] + dgd[k + stride] + dgd[k]);
+  }
+  i = height - 1;
+  j = 0;
+  {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] =
+        center * dgd[k] + edge * (dgd[k + 1] + dgd[k - stride] + dgd[k] * 2) +
+        corner * (dgd[k - stride + 1] + dgd[k + 1] + dgd[k - stride] + dgd[k]);
+  }
+  i = height - 1;
+  j = width - 1;
+  {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] =
+        center * dgd[k] + edge * (dgd[k - 1] + dgd[k - stride] + dgd[k] * 2) +
+        corner * (dgd[k - stride - 1] + dgd[k - 1] + dgd[k - stride] + dgd[k]);
+  }
+  i = 0;
+  for (j = 1; j < width - 1; ++j) {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] = center * dgd[k] +
+             edge * (dgd[k - 1] + dgd[k + stride] + dgd[k + 1] + dgd[k]) +
+             corner * (dgd[k + stride - 1] + dgd[k + stride + 1] + dgd[k - 1] +
+                       dgd[k + 1]);
+  }
+  i = height - 1;
+  for (j = 1; j < width - 1; ++j) {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] = center * dgd[k] +
+             edge * (dgd[k - 1] + dgd[k - stride] + dgd[k + 1] + dgd[k]) +
+             corner * (dgd[k - stride - 1] + dgd[k - stride + 1] + dgd[k - 1] +
+                       dgd[k + 1]);
+  }
+  j = 0;
+  for (i = 1; i < height - 1; ++i) {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] = center * dgd[k] +
+             edge * (dgd[k - stride] + dgd[k + 1] + dgd[k + stride] + dgd[k]) +
+             corner * (dgd[k + stride + 1] + dgd[k - stride + 1] +
+                       dgd[k - stride] + dgd[k + stride]);
+  }
+  j = width - 1;
+  for (i = 1; i < height - 1; ++i) {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] = center * dgd[k] +
+             edge * (dgd[k - stride] + dgd[k - 1] + dgd[k + stride] + dgd[k]) +
+             corner * (dgd[k + stride - 1] + dgd[k - stride - 1] +
+                       dgd[k - stride] + dgd[k + stride]);
+  }
+  for (i = 1; i < height - 1; ++i) {
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] +
+          edge * (dgd[k - stride] + dgd[k - 1] + dgd[k + stride] + dgd[k + 1]) +
+          corner * (dgd[k + stride - 1] + dgd[k - stride - 1] +
+                    dgd[k - stride + 1] + dgd[k + stride + 1]);
+    }
+  }
+}
+
+void apply_selfguided_restoration_c(uint8_t *dat, int width, int height,
+                                    int stride, int eps, int *xqd, uint8_t *dst,
+                                    int dst_stride, int32_t *tmpbuf) {
+  int xq[2];
+  int32_t *flt1 = tmpbuf;
+  int32_t *flt2 = flt1 + RESTORATION_TILEPELS_MAX;
+  int32_t *tmpbuf2 = flt2 + RESTORATION_TILEPELS_MAX;
+  int i, j;
+  assert(width * height <= RESTORATION_TILEPELS_MAX);
+#if USE_HIGHPASS_IN_SGRPROJ
+  av1_highpass_filter_c(dat, width, height, stride, flt1, width,
+                        sgr_params[eps].corner, sgr_params[eps].edge);
+#else
+  av1_selfguided_restoration_c(dat, width, height, stride, flt1, width,
+                               sgr_params[eps].r1, sgr_params[eps].e1, tmpbuf2);
+#endif  // USE_HIGHPASS_IN_SGRPROJ
+  av1_selfguided_restoration_c(dat, width, height, stride, flt2, width,
+                               sgr_params[eps].r2, sgr_params[eps].e2, tmpbuf2);
+  decode_xq(xqd, xq);
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      const int k = i * width + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int32_t u = ((int32_t)dat[l] << SGRPROJ_RST_BITS);
+      const int32_t f1 = (int32_t)flt1[k] - u;
+      const int32_t f2 = (int32_t)flt2[k] - u;
+      const int32_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
+      const int16_t w =
+          (int16_t)ROUND_POWER_OF_TWO(v, SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+      dst[m] = clip_pixel(w);
+    }
+  }
+}
+
+static void loop_sgrproj_filter_tile(uint8_t *data, int tile_idx, int width,
+                                     int height, int stride,
+                                     RestorationInternal *rst, uint8_t *dst,
+                                     int dst_stride) {
+  const int tile_width = rst->tile_width;
+  const int tile_height = rst->tile_height;
+  int h_start, h_end, v_start, v_end;
+  uint8_t *data_p, *dst_p;
+
+  if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
+    loop_copy_tile(data, tile_idx, 0, 0, width, height, stride, rst, dst,
+                   dst_stride);
+    return;
+  }
+  av1_get_rest_tile_limits(tile_idx, 0, 0, rst->nhtiles, rst->nvtiles,
+                           tile_width, tile_height, width, height, 0, 0,
+                           &h_start, &h_end, &v_start, &v_end);
+  data_p = data + h_start + v_start * stride;
+  dst_p = dst + h_start + v_start * dst_stride;
+  apply_selfguided_restoration(data_p, h_end - h_start, v_end - v_start, stride,
+                               rst->rsi->sgrproj_info[tile_idx].ep,
+                               rst->rsi->sgrproj_info[tile_idx].xqd, dst_p,
+                               dst_stride, rst->tmpbuf);
+}
+
+static void loop_sgrproj_filter(uint8_t *data, int width, int height,
+                                int stride, RestorationInternal *rst,
+                                uint8_t *dst, int dst_stride) {
+  int tile_idx;
+  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
+    loop_sgrproj_filter_tile(data, tile_idx, width, height, stride, rst, dst,
+                             dst_stride);
+  }
+}
+
+static void loop_switchable_filter(uint8_t *data, int width, int height,
+                                   int stride, RestorationInternal *rst,
+                                   uint8_t *dst, int dst_stride) {
+  int tile_idx;
+  extend_frame(data, width, height, stride);
+  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
+    if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
+      loop_copy_tile(data, tile_idx, 0, 0, width, height, stride, rst, dst,
+                     dst_stride);
+    } else if (rst->rsi->restoration_type[tile_idx] == RESTORE_WIENER) {
+      loop_wiener_filter_tile(data, tile_idx, width, height, stride, rst, dst,
+                              dst_stride);
+    } else if (rst->rsi->restoration_type[tile_idx] == RESTORE_SGRPROJ) {
+      loop_sgrproj_filter_tile(data, tile_idx, width, height, stride, rst, dst,
+                               dst_stride);
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void extend_frame_highbd(uint16_t *data, int width, int height, int stride) {
+  uint16_t *data_p;
+  int i, j;
+  for (i = 0; i < height; ++i) {
+    data_p = data + i * stride;
+    for (j = -WIENER_HALFWIN; j < 0; ++j) data_p[j] = data_p[0];
+    for (j = width; j < width + WIENER_HALFWIN; ++j)
+      data_p[j] = data_p[width - 1];
+  }
+  data_p = data - WIENER_HALFWIN;
+  for (i = -WIENER_HALFWIN; i < 0; ++i) {
+    memcpy(data_p + i * stride, data_p,
+           (width + 2 * WIENER_HALFWIN) * sizeof(uint16_t));
+  }
+  for (i = height; i < height + WIENER_HALFWIN; ++i) {
+    memcpy(data_p + i * stride, data_p + (height - 1) * stride,
+           (width + 2 * WIENER_HALFWIN) * sizeof(uint16_t));
+  }
+}
+
+static void loop_copy_tile_highbd(uint16_t *data, int tile_idx, int subtile_idx,
+                                  int subtile_bits, int width, int height,
+                                  int stride, RestorationInternal *rst,
+                                  uint16_t *dst, int dst_stride) {
+  const int tile_width = rst->tile_width;
+  const int tile_height = rst->tile_height;
+  int i;
+  int h_start, h_end, v_start, v_end;
+  av1_get_rest_tile_limits(tile_idx, subtile_idx, subtile_bits, rst->nhtiles,
+                           rst->nvtiles, tile_width, tile_height, width, height,
+                           0, 0, &h_start, &h_end, &v_start, &v_end);
+  for (i = v_start; i < v_end; ++i)
+    memcpy(dst + i * dst_stride + h_start, data + i * stride + h_start,
+           (h_end - h_start) * sizeof(*dst));
+}
+
+static void loop_wiener_filter_tile_highbd(uint16_t *data, int tile_idx,
+                                           int width, int height, int stride,
+                                           RestorationInternal *rst,
+                                           int bit_depth, uint16_t *dst,
+                                           int dst_stride) {
+  const int tile_width = rst->tile_width;
+  const int tile_height = rst->tile_height;
+  int h_start, h_end, v_start, v_end;
+  int i, j;
+
+  if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
+    loop_copy_tile_highbd(data, tile_idx, 0, 0, width, height, stride, rst, dst,
+                          dst_stride);
+    return;
+  }
+  av1_get_rest_tile_limits(tile_idx, 0, 0, rst->nhtiles, rst->nvtiles,
+                           tile_width, tile_height, width, height, 0, 0,
+                           &h_start, &h_end, &v_start, &v_end);
+  // Convolve the whole tile (done in blocks here to match the requirements
+  // of the vectorized convolve functions, but the result is equivalent)
+  for (i = v_start; i < v_end; i += MAX_SB_SIZE)
+    for (j = h_start; j < h_end; j += MAX_SB_SIZE) {
+      int w = AOMMIN(MAX_SB_SIZE, (h_end - j + 15) & ~15);
+      int h = AOMMIN(MAX_SB_SIZE, (v_end - i + 15) & ~15);
+      const uint16_t *data_p = data + i * stride + j;
+      uint16_t *dst_p = dst + i * dst_stride + j;
+      aom_highbd_convolve8_add_src(
+          CONVERT_TO_BYTEPTR(data_p), stride, CONVERT_TO_BYTEPTR(dst_p),
+          dst_stride, rst->rsi->wiener_info[tile_idx].hfilter, 16,
+          rst->rsi->wiener_info[tile_idx].vfilter, 16, w, h, bit_depth);
+    }
+}
+
+static void loop_wiener_filter_highbd(uint8_t *data8, int width, int height,
+                                      int stride, RestorationInternal *rst,
+                                      int bit_depth, uint8_t *dst8,
+                                      int dst_stride) {
+  uint16_t *data = CONVERT_TO_SHORTPTR(data8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  int tile_idx;
+  extend_frame_highbd(data, width, height, stride);
+  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
+    loop_wiener_filter_tile_highbd(data, tile_idx, width, height, stride, rst,
+                                   bit_depth, dst, dst_stride);
+  }
+}
+
+void av1_selfguided_restoration_highbd_c(uint16_t *dgd, int width, int height,
+                                         int stride, int32_t *dst,
+                                         int dst_stride, int bit_depth, int r,
+                                         int eps, int32_t *tmpbuf) {
+  int i, j;
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      dst[i * dst_stride + j] = dgd[i * stride + j];
+    }
+  }
+  av1_selfguided_restoration_internal(dst, width, height, dst_stride, bit_depth,
+                                      r, eps, tmpbuf);
+}
+
+void av1_highpass_filter_highbd_c(uint16_t *dgd, int width, int height,
+                                  int stride, int32_t *dst, int dst_stride,
+                                  int corner, int edge) {
+  int i, j;
+  const int center = (1 << SGRPROJ_RST_BITS) - 4 * (corner + edge);
+
+  i = 0;
+  j = 0;
+  {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] =
+        center * dgd[k] + edge * (dgd[k + 1] + dgd[k + stride] + dgd[k] * 2) +
+        corner * (dgd[k + stride + 1] + dgd[k + 1] + dgd[k + stride] + dgd[k]);
+  }
+  i = 0;
+  j = width - 1;
+  {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] =
+        center * dgd[k] + edge * (dgd[k - 1] + dgd[k + stride] + dgd[k] * 2) +
+        corner * (dgd[k + stride - 1] + dgd[k - 1] + dgd[k + stride] + dgd[k]);
+  }
+  i = height - 1;
+  j = 0;
+  {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] =
+        center * dgd[k] + edge * (dgd[k + 1] + dgd[k - stride] + dgd[k] * 2) +
+        corner * (dgd[k - stride + 1] + dgd[k + 1] + dgd[k - stride] + dgd[k]);
+  }
+  i = height - 1;
+  j = width - 1;
+  {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] =
+        center * dgd[k] + edge * (dgd[k - 1] + dgd[k - stride] + dgd[k] * 2) +
+        corner * (dgd[k - stride - 1] + dgd[k - 1] + dgd[k - stride] + dgd[k]);
+  }
+  i = 0;
+  for (j = 1; j < width - 1; ++j) {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] = center * dgd[k] +
+             edge * (dgd[k - 1] + dgd[k + stride] + dgd[k + 1] + dgd[k]) +
+             corner * (dgd[k + stride - 1] + dgd[k + stride + 1] + dgd[k - 1] +
+                       dgd[k + 1]);
+  }
+  i = height - 1;
+  for (j = 1; j < width - 1; ++j) {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] = center * dgd[k] +
+             edge * (dgd[k - 1] + dgd[k - stride] + dgd[k + 1] + dgd[k]) +
+             corner * (dgd[k - stride - 1] + dgd[k - stride + 1] + dgd[k - 1] +
+                       dgd[k + 1]);
+  }
+  j = 0;
+  for (i = 1; i < height - 1; ++i) {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] = center * dgd[k] +
+             edge * (dgd[k - stride] + dgd[k + 1] + dgd[k + stride] + dgd[k]) +
+             corner * (dgd[k + stride + 1] + dgd[k - stride + 1] +
+                       dgd[k - stride] + dgd[k + stride]);
+  }
+  j = width - 1;
+  for (i = 1; i < height - 1; ++i) {
+    const int k = i * stride + j;
+    const int l = i * dst_stride + j;
+    dst[l] = center * dgd[k] +
+             edge * (dgd[k - stride] + dgd[k - 1] + dgd[k + stride] + dgd[k]) +
+             corner * (dgd[k + stride - 1] + dgd[k - stride - 1] +
+                       dgd[k - stride] + dgd[k + stride]);
+  }
+  for (i = 1; i < height - 1; ++i) {
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] +
+          edge * (dgd[k - stride] + dgd[k - 1] + dgd[k + stride] + dgd[k + 1]) +
+          corner * (dgd[k + stride - 1] + dgd[k - stride - 1] +
+                    dgd[k - stride + 1] + dgd[k + stride + 1]);
+    }
+  }
+}
+
+void apply_selfguided_restoration_highbd_c(uint16_t *dat, int width, int height,
+                                           int stride, int bit_depth, int eps,
+                                           int *xqd, uint16_t *dst,
+                                           int dst_stride, int32_t *tmpbuf) {
+  int xq[2];
+  int32_t *flt1 = tmpbuf;
+  int32_t *flt2 = flt1 + RESTORATION_TILEPELS_MAX;
+  int32_t *tmpbuf2 = flt2 + RESTORATION_TILEPELS_MAX;
+  int i, j;
+  assert(width * height <= RESTORATION_TILEPELS_MAX);
+#if USE_HIGHPASS_IN_SGRPROJ
+  av1_highpass_filter_highbd_c(dat, width, height, stride, flt1, width,
+                               sgr_params[eps].corner, sgr_params[eps].edge);
+#else
+  av1_selfguided_restoration_highbd_c(dat, width, height, stride, flt1, width,
+                                      bit_depth, sgr_params[eps].r1,
+                                      sgr_params[eps].e1, tmpbuf2);
+#endif  // USE_HIGHPASS_IN_SGRPROJ
+  av1_selfguided_restoration_highbd_c(dat, width, height, stride, flt2, width,
+                                      bit_depth, sgr_params[eps].r2,
+                                      sgr_params[eps].e2, tmpbuf2);
+  decode_xq(xqd, xq);
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < width; ++j) {
+      const int k = i * width + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int32_t u = ((int32_t)dat[l] << SGRPROJ_RST_BITS);
+      const int32_t f1 = (int32_t)flt1[k] - u;
+      const int32_t f2 = (int32_t)flt2[k] - u;
+      const int32_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
+      const int16_t w =
+          (int16_t)ROUND_POWER_OF_TWO(v, SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+      dst[m] = (uint16_t)clip_pixel_highbd(w, bit_depth);
+    }
+  }
+}
+
+static void loop_sgrproj_filter_tile_highbd(uint16_t *data, int tile_idx,
+                                            int width, int height, int stride,
+                                            RestorationInternal *rst,
+                                            int bit_depth, uint16_t *dst,
+                                            int dst_stride) {
+  const int tile_width = rst->tile_width;
+  const int tile_height = rst->tile_height;
+  int h_start, h_end, v_start, v_end;
+  uint16_t *data_p, *dst_p;
+
+  if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
+    loop_copy_tile_highbd(data, tile_idx, 0, 0, width, height, stride, rst, dst,
+                          dst_stride);
+    return;
+  }
+  av1_get_rest_tile_limits(tile_idx, 0, 0, rst->nhtiles, rst->nvtiles,
+                           tile_width, tile_height, width, height, 0, 0,
+                           &h_start, &h_end, &v_start, &v_end);
+  data_p = data + h_start + v_start * stride;
+  dst_p = dst + h_start + v_start * dst_stride;
+  apply_selfguided_restoration_highbd(
+      data_p, h_end - h_start, v_end - v_start, stride, bit_depth,
+      rst->rsi->sgrproj_info[tile_idx].ep, rst->rsi->sgrproj_info[tile_idx].xqd,
+      dst_p, dst_stride, rst->tmpbuf);
+}
+
+static void loop_sgrproj_filter_highbd(uint8_t *data8, int width, int height,
+                                       int stride, RestorationInternal *rst,
+                                       int bit_depth, uint8_t *dst8,
+                                       int dst_stride) {
+  int tile_idx;
+  uint16_t *data = CONVERT_TO_SHORTPTR(data8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
+    loop_sgrproj_filter_tile_highbd(data, tile_idx, width, height, stride, rst,
+                                    bit_depth, dst, dst_stride);
+  }
+}
+
+static void loop_switchable_filter_highbd(uint8_t *data8, int width, int height,
+                                          int stride, RestorationInternal *rst,
+                                          int bit_depth, uint8_t *dst8,
+                                          int dst_stride) {
+  uint16_t *data = CONVERT_TO_SHORTPTR(data8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  int tile_idx;
+  extend_frame_highbd(data, width, height, stride);
+  for (tile_idx = 0; tile_idx < rst->ntiles; ++tile_idx) {
+    if (rst->rsi->restoration_type[tile_idx] == RESTORE_NONE) {
+      loop_copy_tile_highbd(data, tile_idx, 0, 0, width, height, stride, rst,
+                            dst, dst_stride);
+    } else if (rst->rsi->restoration_type[tile_idx] == RESTORE_WIENER) {
+      loop_wiener_filter_tile_highbd(data, tile_idx, width, height, stride, rst,
+                                     bit_depth, dst, dst_stride);
+    } else if (rst->rsi->restoration_type[tile_idx] == RESTORE_SGRPROJ) {
+      loop_sgrproj_filter_tile_highbd(data, tile_idx, width, height, stride,
+                                      rst, bit_depth, dst, dst_stride);
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static void loop_restoration_rows(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
+                                  int start_mi_row, int end_mi_row,
+                                  int components_pattern, RestorationInfo *rsi,
+                                  YV12_BUFFER_CONFIG *dst) {
+  const int ywidth = frame->y_crop_width;
+  const int ystride = frame->y_stride;
+  const int uvwidth = frame->uv_crop_width;
+  const int uvstride = frame->uv_stride;
+  const int ystart = start_mi_row << MI_SIZE_LOG2;
+  const int uvstart = ystart >> cm->subsampling_y;
+  int yend = end_mi_row << MI_SIZE_LOG2;
+  int uvend = yend >> cm->subsampling_y;
+  restore_func_type restore_funcs[RESTORE_TYPES] = {
+    NULL, loop_wiener_filter, loop_sgrproj_filter, loop_switchable_filter
+  };
+#if CONFIG_HIGHBITDEPTH
+  restore_func_highbd_type restore_funcs_highbd[RESTORE_TYPES] = {
+    NULL, loop_wiener_filter_highbd, loop_sgrproj_filter_highbd,
+    loop_switchable_filter_highbd
+  };
+#endif  // CONFIG_HIGHBITDEPTH
+  restore_func_type restore_func;
+#if CONFIG_HIGHBITDEPTH
+  restore_func_highbd_type restore_func_highbd;
+#endif  // CONFIG_HIGHBITDEPTH
+  YV12_BUFFER_CONFIG dst_;
+
+  yend = AOMMIN(yend, cm->height);
+  uvend = AOMMIN(uvend, cm->subsampling_y ? (cm->height + 1) >> 1 : cm->height);
+
+  if (components_pattern == (1 << AOM_PLANE_Y)) {
+    // Only y
+    if (rsi[0].frame_restoration_type == RESTORE_NONE) {
+      if (dst) aom_yv12_copy_y(frame, dst);
+      return;
+    }
+  } else if (components_pattern == (1 << AOM_PLANE_U)) {
+    // Only U
+    if (rsi[1].frame_restoration_type == RESTORE_NONE) {
+      if (dst) aom_yv12_copy_u(frame, dst);
+      return;
+    }
+  } else if (components_pattern == (1 << AOM_PLANE_V)) {
+    // Only V
+    if (rsi[2].frame_restoration_type == RESTORE_NONE) {
+      if (dst) aom_yv12_copy_v(frame, dst);
+      return;
+    }
+  } else if (components_pattern ==
+             ((1 << AOM_PLANE_Y) | (1 << AOM_PLANE_U) | (1 << AOM_PLANE_V))) {
+    // All components
+    if (rsi[0].frame_restoration_type == RESTORE_NONE &&
+        rsi[1].frame_restoration_type == RESTORE_NONE &&
+        rsi[2].frame_restoration_type == RESTORE_NONE) {
+      if (dst) aom_yv12_copy_frame(frame, dst);
+      return;
+    }
+  }
+
+  if (!dst) {
+    dst = &dst_;
+    memset(dst, 0, sizeof(YV12_BUFFER_CONFIG));
+    if (aom_realloc_frame_buffer(
+            dst, cm->width, cm->height, cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+            cm->use_highbitdepth,
+#endif
+            AOM_BORDER_IN_PIXELS, cm->byte_alignment, NULL, NULL, NULL) < 0)
+      aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                         "Failed to allocate restoration dst buffer");
+  }
+
+  if ((components_pattern >> AOM_PLANE_Y) & 1) {
+    if (rsi[0].frame_restoration_type != RESTORE_NONE) {
+      cm->rst_internal.ntiles = av1_get_rest_ntiles(
+          cm->width, cm->height, cm->rst_info[AOM_PLANE_Y].restoration_tilesize,
+          &cm->rst_internal.tile_width, &cm->rst_internal.tile_height,
+          &cm->rst_internal.nhtiles, &cm->rst_internal.nvtiles);
+      cm->rst_internal.rsi = &rsi[0];
+      restore_func =
+          restore_funcs[cm->rst_internal.rsi->frame_restoration_type];
+#if CONFIG_HIGHBITDEPTH
+      restore_func_highbd =
+          restore_funcs_highbd[cm->rst_internal.rsi->frame_restoration_type];
+      if (cm->use_highbitdepth)
+        restore_func_highbd(
+            frame->y_buffer + ystart * ystride, ywidth, yend - ystart, ystride,
+            &cm->rst_internal, cm->bit_depth,
+            dst->y_buffer + ystart * dst->y_stride, dst->y_stride);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        restore_func(frame->y_buffer + ystart * ystride, ywidth, yend - ystart,
+                     ystride, &cm->rst_internal,
+                     dst->y_buffer + ystart * dst->y_stride, dst->y_stride);
+    } else {
+      aom_yv12_copy_y(frame, dst);
+    }
+  }
+
+  if ((components_pattern >> AOM_PLANE_U) & 1) {
+    if (rsi[AOM_PLANE_U].frame_restoration_type != RESTORE_NONE) {
+      cm->rst_internal.ntiles = av1_get_rest_ntiles(
+          ROUND_POWER_OF_TWO(cm->width, cm->subsampling_x),
+          ROUND_POWER_OF_TWO(cm->height, cm->subsampling_y),
+          cm->rst_info[AOM_PLANE_U].restoration_tilesize,
+          &cm->rst_internal.tile_width, &cm->rst_internal.tile_height,
+          &cm->rst_internal.nhtiles, &cm->rst_internal.nvtiles);
+      cm->rst_internal.rsi = &rsi[AOM_PLANE_U];
+      restore_func =
+          restore_funcs[cm->rst_internal.rsi->frame_restoration_type];
+#if CONFIG_HIGHBITDEPTH
+      restore_func_highbd =
+          restore_funcs_highbd[cm->rst_internal.rsi->frame_restoration_type];
+      if (cm->use_highbitdepth)
+        restore_func_highbd(
+            frame->u_buffer + uvstart * uvstride, uvwidth, uvend - uvstart,
+            uvstride, &cm->rst_internal, cm->bit_depth,
+            dst->u_buffer + uvstart * dst->uv_stride, dst->uv_stride);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        restore_func(frame->u_buffer + uvstart * uvstride, uvwidth,
+                     uvend - uvstart, uvstride, &cm->rst_internal,
+                     dst->u_buffer + uvstart * dst->uv_stride, dst->uv_stride);
+    } else {
+      aom_yv12_copy_u(frame, dst);
+    }
+  }
+
+  if ((components_pattern >> AOM_PLANE_V) & 1) {
+    if (rsi[AOM_PLANE_V].frame_restoration_type != RESTORE_NONE) {
+      cm->rst_internal.ntiles = av1_get_rest_ntiles(
+          ROUND_POWER_OF_TWO(cm->width, cm->subsampling_x),
+          ROUND_POWER_OF_TWO(cm->height, cm->subsampling_y),
+          cm->rst_info[AOM_PLANE_V].restoration_tilesize,
+          &cm->rst_internal.tile_width, &cm->rst_internal.tile_height,
+          &cm->rst_internal.nhtiles, &cm->rst_internal.nvtiles);
+      cm->rst_internal.rsi = &rsi[AOM_PLANE_V];
+      restore_func =
+          restore_funcs[cm->rst_internal.rsi->frame_restoration_type];
+#if CONFIG_HIGHBITDEPTH
+      restore_func_highbd =
+          restore_funcs_highbd[cm->rst_internal.rsi->frame_restoration_type];
+      if (cm->use_highbitdepth)
+        restore_func_highbd(
+            frame->v_buffer + uvstart * uvstride, uvwidth, uvend - uvstart,
+            uvstride, &cm->rst_internal, cm->bit_depth,
+            dst->v_buffer + uvstart * dst->uv_stride, dst->uv_stride);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        restore_func(frame->v_buffer + uvstart * uvstride, uvwidth,
+                     uvend - uvstart, uvstride, &cm->rst_internal,
+                     dst->v_buffer + uvstart * dst->uv_stride, dst->uv_stride);
+    } else {
+      aom_yv12_copy_v(frame, dst);
+    }
+  }
+
+  if (dst == &dst_) {
+    if ((components_pattern >> AOM_PLANE_Y) & 1) aom_yv12_copy_y(dst, frame);
+    if ((components_pattern >> AOM_PLANE_U) & 1) aom_yv12_copy_u(dst, frame);
+    if ((components_pattern >> AOM_PLANE_V) & 1) aom_yv12_copy_v(dst, frame);
+    aom_free_frame_buffer(dst);
+  }
+}
+
+void av1_loop_restoration_frame(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
+                                RestorationInfo *rsi, int components_pattern,
+                                int partial_frame, YV12_BUFFER_CONFIG *dst) {
+  int start_mi_row, end_mi_row, mi_rows_to_filter;
+  start_mi_row = 0;
+  mi_rows_to_filter = cm->mi_rows;
+  if (partial_frame && cm->mi_rows > 8) {
+    start_mi_row = cm->mi_rows >> 1;
+    start_mi_row &= 0xfffffff8;
+    mi_rows_to_filter = AOMMAX(cm->mi_rows / 8, 8);
+  }
+  end_mi_row = start_mi_row + mi_rows_to_filter;
+  loop_restoration_init(&cm->rst_internal, cm->frame_type == KEY_FRAME);
+  loop_restoration_rows(frame, cm, start_mi_row, end_mi_row, components_pattern,
+                        rsi, dst);
+}
diff --git a/third_party/aom/av1/common/restoration.h b/third_party/aom/av1/common/restoration.h
new file mode 100644
index 0000000000..866f78b792
--- /dev/null
+++ b/third_party/aom/av1/common/restoration.h
@@ -0,0 +1,257 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_RESTORATION_H_
+#define AV1_COMMON_RESTORATION_H_
+
+#include "aom_ports/mem.h"
+#include "./aom_config.h"
+
+#include "av1/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define CLIP(x, lo, hi) ((x) < (lo) ? (lo) : (x) > (hi) ? (hi) : (x))
+#define RINT(x) ((x) < 0 ? (int)((x)-0.5) : (int)((x) + 0.5))
+
+#define RESTORATION_TILESIZE_MAX 256
+#define RESTORATION_TILEPELS_MAX \
+  (RESTORATION_TILESIZE_MAX * RESTORATION_TILESIZE_MAX * 9 / 4)
+
+// 4 32-bit buffers needed for the filter:
+// 2 for the restored versions of the frame and
+// 2 for each restoration operation
+#define SGRPROJ_OUTBUF_SIZE \
+  ((RESTORATION_TILESIZE_MAX * 3 / 2) * (RESTORATION_TILESIZE_MAX * 3 / 2 + 16))
+#define SGRPROJ_TMPBUF_SIZE                         \
+  (RESTORATION_TILEPELS_MAX * 2 * sizeof(int32_t) + \
+   SGRPROJ_OUTBUF_SIZE * 2 * sizeof(int32_t))
+#define SGRPROJ_EXTBUF_SIZE (0)
+#define SGRPROJ_PARAMS_BITS 4
+#define SGRPROJ_PARAMS (1 << SGRPROJ_PARAMS_BITS)
+#define USE_HIGHPASS_IN_SGRPROJ 0
+
+// Precision bits for projection
+#define SGRPROJ_PRJ_BITS 7
+// Restoration precision bits generated higher than source before projection
+#define SGRPROJ_RST_BITS 4
+// Internal precision bits for core selfguided_restoration
+#define SGRPROJ_SGR_BITS 8
+#define SGRPROJ_SGR (1 << SGRPROJ_SGR_BITS)
+
+#if USE_HIGHPASS_IN_SGRPROJ
+#define SGRPROJ_PRJ_MIN0 (-(1 << SGRPROJ_PRJ_BITS) / 8)
+#define SGRPROJ_PRJ_MAX0 (SGRPROJ_PRJ_MIN0 + (1 << SGRPROJ_PRJ_BITS) - 1)
+#define SGRPROJ_PRJ_MIN1 (-(1 << SGRPROJ_PRJ_BITS) / 2)
+#define SGRPROJ_PRJ_MAX1 (SGRPROJ_PRJ_MIN1 + (1 << SGRPROJ_PRJ_BITS) - 1)
+#else
+#define SGRPROJ_PRJ_MIN0 (-(1 << SGRPROJ_PRJ_BITS) * 3 / 4)
+#define SGRPROJ_PRJ_MAX0 (SGRPROJ_PRJ_MIN0 + (1 << SGRPROJ_PRJ_BITS) - 1)
+#define SGRPROJ_PRJ_MIN1 (-(1 << SGRPROJ_PRJ_BITS) / 4)
+#define SGRPROJ_PRJ_MAX1 (SGRPROJ_PRJ_MIN1 + (1 << SGRPROJ_PRJ_BITS) - 1)
+#endif  // USE_HIGHPASS_IN_SGRPROJ
+
+#define SGRPROJ_PRJ_SUBEXP_K 4
+
+#define SGRPROJ_BITS (SGRPROJ_PRJ_BITS * 2 + SGRPROJ_PARAMS_BITS)
+
+#define MAX_RADIUS 3  // Only 1, 2, 3 allowed
+#define MAX_EPS 80    // Max value of eps
+#define MAX_NELEM ((2 * MAX_RADIUS + 1) * (2 * MAX_RADIUS + 1))
+#define SGRPROJ_MTABLE_BITS 20
+#define SGRPROJ_RECIP_BITS 12
+
+#define WIENER_HALFWIN 3
+#define WIENER_HALFWIN1 (WIENER_HALFWIN + 1)
+#define WIENER_WIN (2 * WIENER_HALFWIN + 1)
+#define WIENER_WIN2 ((WIENER_WIN) * (WIENER_WIN))
+#define WIENER_TMPBUF_SIZE (0)
+#define WIENER_EXTBUF_SIZE (0)
+
+#define WIENER_FILT_PREC_BITS 7
+#define WIENER_FILT_STEP (1 << WIENER_FILT_PREC_BITS)
+
+// Central values for the taps
+#define WIENER_FILT_TAP0_MIDV (3)
+#define WIENER_FILT_TAP1_MIDV (-7)
+#define WIENER_FILT_TAP2_MIDV (15)
+
+#define WIENER_FILT_TAP0_BITS 4
+#define WIENER_FILT_TAP1_BITS 5
+#define WIENER_FILT_TAP2_BITS 6
+
+#define WIENER_FILT_BITS \
+  ((WIENER_FILT_TAP0_BITS + WIENER_FILT_TAP1_BITS + WIENER_FILT_TAP2_BITS) * 2)
+
+#define WIENER_FILT_TAP0_MINV \
+  (WIENER_FILT_TAP0_MIDV - (1 << WIENER_FILT_TAP0_BITS) / 2)
+#define WIENER_FILT_TAP1_MINV \
+  (WIENER_FILT_TAP1_MIDV - (1 << WIENER_FILT_TAP1_BITS) / 2)
+#define WIENER_FILT_TAP2_MINV \
+  (WIENER_FILT_TAP2_MIDV - (1 << WIENER_FILT_TAP2_BITS) / 2)
+
+#define WIENER_FILT_TAP0_MAXV \
+  (WIENER_FILT_TAP0_MIDV - 1 + (1 << WIENER_FILT_TAP0_BITS) / 2)
+#define WIENER_FILT_TAP1_MAXV \
+  (WIENER_FILT_TAP1_MIDV - 1 + (1 << WIENER_FILT_TAP1_BITS) / 2)
+#define WIENER_FILT_TAP2_MAXV \
+  (WIENER_FILT_TAP2_MIDV - 1 + (1 << WIENER_FILT_TAP2_BITS) / 2)
+
+#define WIENER_FILT_TAP0_SUBEXP_K 1
+#define WIENER_FILT_TAP1_SUBEXP_K 2
+#define WIENER_FILT_TAP2_SUBEXP_K 3
+
+// Max of SGRPROJ_TMPBUF_SIZE, DOMAINTXFMRF_TMPBUF_SIZE, WIENER_TMPBUF_SIZE
+#define RESTORATION_TMPBUF_SIZE (SGRPROJ_TMPBUF_SIZE)
+
+// Max of SGRPROJ_EXTBUF_SIZE, WIENER_EXTBUF_SIZE
+#define RESTORATION_EXTBUF_SIZE (WIENER_EXTBUF_SIZE)
+
+// Check the assumptions of the existing code
+#if SUBPEL_TAPS != WIENER_WIN + 1
+#error "Wiener filter currently only works if SUBPEL_TAPS == WIENER_WIN + 1"
+#endif
+#if WIENER_FILT_PREC_BITS != 7
+#error "Wiener filter currently only works if WIENER_FILT_PREC_BITS == 7"
+#endif
+typedef struct {
+  DECLARE_ALIGNED(16, InterpKernel, vfilter);
+  DECLARE_ALIGNED(16, InterpKernel, hfilter);
+} WienerInfo;
+
+typedef struct {
+#if USE_HIGHPASS_IN_SGRPROJ
+  int corner;
+  int edge;
+#else
+  int r1;
+  int e1;
+#endif  // USE_HIGHPASS_IN_SGRPROJ
+  int r2;
+  int e2;
+} sgr_params_type;
+
+typedef struct {
+  int ep;
+  int xqd[2];
+} SgrprojInfo;
+
+typedef struct {
+  int restoration_tilesize;
+  RestorationType frame_restoration_type;
+  RestorationType *restoration_type;
+  // Wiener filter
+  WienerInfo *wiener_info;
+  // Selfguided proj filter
+  SgrprojInfo *sgrproj_info;
+} RestorationInfo;
+
+typedef struct {
+  RestorationInfo *rsi;
+  int keyframe;
+  int ntiles;
+  int tile_width, tile_height;
+  int nhtiles, nvtiles;
+  int32_t *tmpbuf;
+} RestorationInternal;
+
+static INLINE void set_default_sgrproj(SgrprojInfo *sgrproj_info) {
+  sgrproj_info->xqd[0] = (SGRPROJ_PRJ_MIN0 + SGRPROJ_PRJ_MAX0) / 2;
+  sgrproj_info->xqd[1] = (SGRPROJ_PRJ_MIN1 + SGRPROJ_PRJ_MAX1) / 2;
+}
+
+static INLINE void set_default_wiener(WienerInfo *wiener_info) {
+  wiener_info->vfilter[0] = wiener_info->hfilter[0] = WIENER_FILT_TAP0_MIDV;
+  wiener_info->vfilter[1] = wiener_info->hfilter[1] = WIENER_FILT_TAP1_MIDV;
+  wiener_info->vfilter[2] = wiener_info->hfilter[2] = WIENER_FILT_TAP2_MIDV;
+  wiener_info->vfilter[WIENER_HALFWIN] = wiener_info->hfilter[WIENER_HALFWIN] =
+      -2 *
+      (WIENER_FILT_TAP2_MIDV + WIENER_FILT_TAP1_MIDV + WIENER_FILT_TAP0_MIDV);
+  wiener_info->vfilter[4] = wiener_info->hfilter[4] = WIENER_FILT_TAP2_MIDV;
+  wiener_info->vfilter[5] = wiener_info->hfilter[5] = WIENER_FILT_TAP1_MIDV;
+  wiener_info->vfilter[6] = wiener_info->hfilter[6] = WIENER_FILT_TAP0_MIDV;
+}
+
+static INLINE int av1_get_rest_ntiles(int width, int height, int tilesize,
+                                      int *tile_width, int *tile_height,
+                                      int *nhtiles, int *nvtiles) {
+  int nhtiles_, nvtiles_;
+  int tile_width_, tile_height_;
+  tile_width_ = (tilesize < 0) ? width : AOMMIN(tilesize, width);
+  tile_height_ = (tilesize < 0) ? height : AOMMIN(tilesize, height);
+  nhtiles_ = (width + (tile_width_ >> 1)) / tile_width_;
+  nvtiles_ = (height + (tile_height_ >> 1)) / tile_height_;
+  if (tile_width) *tile_width = tile_width_;
+  if (tile_height) *tile_height = tile_height_;
+  if (nhtiles) *nhtiles = nhtiles_;
+  if (nvtiles) *nvtiles = nvtiles_;
+  return (nhtiles_ * nvtiles_);
+}
+
+static INLINE void av1_get_rest_tile_limits(
+    int tile_idx, int subtile_idx, int subtile_bits, int nhtiles, int nvtiles,
+    int tile_width, int tile_height, int im_width, int im_height, int clamp_h,
+    int clamp_v, int *h_start, int *h_end, int *v_start, int *v_end) {
+  const int htile_idx = tile_idx % nhtiles;
+  const int vtile_idx = tile_idx / nhtiles;
+  *h_start = htile_idx * tile_width;
+  *v_start = vtile_idx * tile_height;
+  *h_end = (htile_idx < nhtiles - 1) ? *h_start + tile_width : im_width;
+  *v_end = (vtile_idx < nvtiles - 1) ? *v_start + tile_height : im_height;
+  if (subtile_bits) {
+    const int num_subtiles_1d = (1 << subtile_bits);
+    const int subtile_width = (*h_end - *h_start) >> subtile_bits;
+    const int subtile_height = (*v_end - *v_start) >> subtile_bits;
+    const int subtile_idx_h = subtile_idx & (num_subtiles_1d - 1);
+    const int subtile_idx_v = subtile_idx >> subtile_bits;
+    *h_start += subtile_idx_h * subtile_width;
+    *v_start += subtile_idx_v * subtile_height;
+    *h_end = subtile_idx_h == num_subtiles_1d - 1 ? *h_end
+                                                  : *h_start + subtile_width;
+    *v_end = subtile_idx_v == num_subtiles_1d - 1 ? *v_end
+                                                  : *v_start + subtile_height;
+  }
+  if (clamp_h) {
+    *h_start = AOMMAX(*h_start, clamp_h);
+    *h_end = AOMMIN(*h_end, im_width - clamp_h);
+  }
+  if (clamp_v) {
+    *v_start = AOMMAX(*v_start, clamp_v);
+    *v_end = AOMMIN(*v_end, im_height - clamp_v);
+  }
+}
+
+extern const sgr_params_type sgr_params[SGRPROJ_PARAMS];
+extern int sgrproj_mtable[MAX_EPS][MAX_NELEM];
+extern const int32_t x_by_xplus1[256];
+extern const int32_t one_by_x[MAX_NELEM];
+
+int av1_alloc_restoration_struct(struct AV1Common *cm,
+                                 RestorationInfo *rst_info, int width,
+                                 int height);
+void av1_free_restoration_struct(RestorationInfo *rst_info);
+
+void extend_frame(uint8_t *data, int width, int height, int stride);
+#if CONFIG_HIGHBITDEPTH
+void extend_frame_highbd(uint16_t *data, int width, int height, int stride);
+#endif  // CONFIG_HIGHBITDEPTH
+void decode_xq(int *xqd, int *xq);
+void av1_loop_restoration_frame(YV12_BUFFER_CONFIG *frame, struct AV1Common *cm,
+                                RestorationInfo *rsi, int components_pattern,
+                                int partial_frame, YV12_BUFFER_CONFIG *dst);
+void av1_loop_restoration_precal();
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_RESTORATION_H_
diff --git a/third_party/aom/av1/common/scale.c b/third_party/aom/av1/common/scale.c
new file mode 100644
index 0000000000..76beaa2bdb
--- /dev/null
+++ b/third_party/aom/av1/common/scale.c
@@ -0,0 +1,164 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_dsp_rtcd.h"
+#include "av1/common/filter.h"
+#include "av1/common/scale.h"
+#include "aom_dsp/aom_filter.h"
+
+static INLINE int scaled_x(int val, const struct scale_factors *sf) {
+  return (int)((int64_t)val * sf->x_scale_fp >> REF_SCALE_SHIFT);
+}
+
+static INLINE int scaled_y(int val, const struct scale_factors *sf) {
+  return (int)((int64_t)val * sf->y_scale_fp >> REF_SCALE_SHIFT);
+}
+
+static int unscaled_value(int val, const struct scale_factors *sf) {
+  (void)sf;
+  return val;
+}
+
+static int get_fixed_point_scale_factor(int other_size, int this_size) {
+  // Calculate scaling factor once for each reference frame
+  // and use fixed point scaling factors in decoding and encoding routines.
+  // Hardware implementations can calculate scale factor in device driver
+  // and use multiplication and shifting on hardware instead of division.
+  return (other_size << REF_SCALE_SHIFT) / this_size;
+}
+
+MV32 av1_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf) {
+  const int x_off_q4 = scaled_x(x << SUBPEL_BITS, sf) & SUBPEL_MASK;
+  const int y_off_q4 = scaled_y(y << SUBPEL_BITS, sf) & SUBPEL_MASK;
+  const MV32 res = { scaled_y(mv->row, sf) + y_off_q4,
+                     scaled_x(mv->col, sf) + x_off_q4 };
+  return res;
+}
+
+#if CONFIG_HIGHBITDEPTH
+void av1_setup_scale_factors_for_frame(struct scale_factors *sf, int other_w,
+                                       int other_h, int this_w, int this_h,
+                                       int use_highbd) {
+#else
+void av1_setup_scale_factors_for_frame(struct scale_factors *sf, int other_w,
+                                       int other_h, int this_w, int this_h) {
+#endif
+  if (!valid_ref_frame_size(other_w, other_h, this_w, this_h)) {
+    sf->x_scale_fp = REF_INVALID_SCALE;
+    sf->y_scale_fp = REF_INVALID_SCALE;
+    return;
+  }
+
+  sf->x_scale_fp = get_fixed_point_scale_factor(other_w, this_w);
+  sf->y_scale_fp = get_fixed_point_scale_factor(other_h, this_h);
+  sf->x_step_q4 = scaled_x(16, sf);
+  sf->y_step_q4 = scaled_y(16, sf);
+
+  if (av1_is_scaled(sf)) {
+    sf->scale_value_x = scaled_x;
+    sf->scale_value_y = scaled_y;
+  } else {
+    sf->scale_value_x = unscaled_value;
+    sf->scale_value_y = unscaled_value;
+  }
+
+  // TODO(agrange): Investigate the best choice of functions to use here
+  // for EIGHTTAP_SMOOTH. Since it is not interpolating, need to choose what
+  // to do at full-pel offsets. The current selection, where the filter is
+  // applied in one direction only, and not at all for 0,0, seems to give the
+  // best quality, but it may be worth trying an additional mode that does
+  // do the filtering on full-pel.
+  if (sf->x_step_q4 == 16) {
+    if (sf->y_step_q4 == 16) {
+      // No scaling in either direction.
+      sf->predict[0][0][0] = aom_convolve_copy;
+      sf->predict[0][0][1] = aom_convolve_avg;
+      sf->predict[0][1][0] = aom_convolve8_vert;
+      sf->predict[0][1][1] = aom_convolve8_avg_vert;
+      sf->predict[1][0][0] = aom_convolve8_horiz;
+      sf->predict[1][0][1] = aom_convolve8_avg_horiz;
+    } else {
+      // No scaling in x direction. Must always scale in the y direction.
+      sf->predict[0][0][0] = aom_convolve8_vert;
+      sf->predict[0][0][1] = aom_convolve8_avg_vert;
+      sf->predict[0][1][0] = aom_convolve8_vert;
+      sf->predict[0][1][1] = aom_convolve8_avg_vert;
+      sf->predict[1][0][0] = aom_convolve8;
+      sf->predict[1][0][1] = aom_convolve8_avg;
+    }
+  } else {
+    if (sf->y_step_q4 == 16) {
+      // No scaling in the y direction. Must always scale in the x direction.
+      sf->predict[0][0][0] = aom_convolve8_horiz;
+      sf->predict[0][0][1] = aom_convolve8_avg_horiz;
+      sf->predict[0][1][0] = aom_convolve8;
+      sf->predict[0][1][1] = aom_convolve8_avg;
+      sf->predict[1][0][0] = aom_convolve8_horiz;
+      sf->predict[1][0][1] = aom_convolve8_avg_horiz;
+    } else {
+      // Must always scale in both directions.
+      sf->predict[0][0][0] = aom_convolve8;
+      sf->predict[0][0][1] = aom_convolve8_avg;
+      sf->predict[0][1][0] = aom_convolve8;
+      sf->predict[0][1][1] = aom_convolve8_avg;
+      sf->predict[1][0][0] = aom_convolve8;
+      sf->predict[1][0][1] = aom_convolve8_avg;
+    }
+  }
+  // 2D subpel motion always gets filtered in both directions
+  sf->predict[1][1][0] = aom_convolve8;
+  sf->predict[1][1][1] = aom_convolve8_avg;
+
+#if CONFIG_HIGHBITDEPTH
+  if (use_highbd) {
+    if (sf->x_step_q4 == 16) {
+      if (sf->y_step_q4 == 16) {
+        // No scaling in either direction.
+        sf->highbd_predict[0][0][0] = aom_highbd_convolve_copy;
+        sf->highbd_predict[0][0][1] = aom_highbd_convolve_avg;
+        sf->highbd_predict[0][1][0] = aom_highbd_convolve8_vert;
+        sf->highbd_predict[0][1][1] = aom_highbd_convolve8_avg_vert;
+        sf->highbd_predict[1][0][0] = aom_highbd_convolve8_horiz;
+        sf->highbd_predict[1][0][1] = aom_highbd_convolve8_avg_horiz;
+      } else {
+        // No scaling in x direction. Must always scale in the y direction.
+        sf->highbd_predict[0][0][0] = aom_highbd_convolve8_vert;
+        sf->highbd_predict[0][0][1] = aom_highbd_convolve8_avg_vert;
+        sf->highbd_predict[0][1][0] = aom_highbd_convolve8_vert;
+        sf->highbd_predict[0][1][1] = aom_highbd_convolve8_avg_vert;
+        sf->highbd_predict[1][0][0] = aom_highbd_convolve8;
+        sf->highbd_predict[1][0][1] = aom_highbd_convolve8_avg;
+      }
+    } else {
+      if (sf->y_step_q4 == 16) {
+        // No scaling in the y direction. Must always scale in the x direction.
+        sf->highbd_predict[0][0][0] = aom_highbd_convolve8_horiz;
+        sf->highbd_predict[0][0][1] = aom_highbd_convolve8_avg_horiz;
+        sf->highbd_predict[0][1][0] = aom_highbd_convolve8;
+        sf->highbd_predict[0][1][1] = aom_highbd_convolve8_avg;
+        sf->highbd_predict[1][0][0] = aom_highbd_convolve8_horiz;
+        sf->highbd_predict[1][0][1] = aom_highbd_convolve8_avg_horiz;
+      } else {
+        // Must always scale in both directions.
+        sf->highbd_predict[0][0][0] = aom_highbd_convolve8;
+        sf->highbd_predict[0][0][1] = aom_highbd_convolve8_avg;
+        sf->highbd_predict[0][1][0] = aom_highbd_convolve8;
+        sf->highbd_predict[0][1][1] = aom_highbd_convolve8_avg;
+        sf->highbd_predict[1][0][0] = aom_highbd_convolve8;
+        sf->highbd_predict[1][0][1] = aom_highbd_convolve8_avg;
+      }
+    }
+    // 2D subpel motion always gets filtered in both directions.
+    sf->highbd_predict[1][1][0] = aom_highbd_convolve8;
+    sf->highbd_predict[1][1][1] = aom_highbd_convolve8_avg;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+}
diff --git a/third_party/aom/av1/common/scale.h b/third_party/aom/av1/common/scale.h
new file mode 100644
index 0000000000..ea81efab0e
--- /dev/null
+++ b/third_party/aom/av1/common/scale.h
@@ -0,0 +1,72 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_SCALE_H_
+#define AV1_COMMON_SCALE_H_
+
+#include "av1/common/mv.h"
+#include "aom_dsp/aom_convolve.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define REF_SCALE_SHIFT 14
+#define REF_NO_SCALE (1 << REF_SCALE_SHIFT)
+#define REF_INVALID_SCALE -1
+
+struct scale_factors {
+  int x_scale_fp;  // horizontal fixed point scale factor
+  int y_scale_fp;  // vertical fixed point scale factor
+  int x_step_q4;
+  int y_step_q4;
+
+  int (*scale_value_x)(int val, const struct scale_factors *sf);
+  int (*scale_value_y)(int val, const struct scale_factors *sf);
+
+  convolve_fn_t predict[2][2][2];  // horiz, vert, avg
+#if CONFIG_HIGHBITDEPTH
+  highbd_convolve_fn_t highbd_predict[2][2][2];  // horiz, vert, avg
+#endif                                           // CONFIG_HIGHBITDEPTH
+};
+
+MV32 av1_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf);
+
+#if CONFIG_HIGHBITDEPTH
+void av1_setup_scale_factors_for_frame(struct scale_factors *sf, int other_w,
+                                       int other_h, int this_w, int this_h,
+                                       int use_high);
+#else
+void av1_setup_scale_factors_for_frame(struct scale_factors *sf, int other_w,
+                                       int other_h, int this_w, int this_h);
+#endif  // CONFIG_HIGHBITDEPTH
+
+static INLINE int av1_is_valid_scale(const struct scale_factors *sf) {
+  return sf->x_scale_fp != REF_INVALID_SCALE &&
+         sf->y_scale_fp != REF_INVALID_SCALE;
+}
+
+static INLINE int av1_is_scaled(const struct scale_factors *sf) {
+  return av1_is_valid_scale(sf) &&
+         (sf->x_scale_fp != REF_NO_SCALE || sf->y_scale_fp != REF_NO_SCALE);
+}
+
+static INLINE int valid_ref_frame_size(int ref_width, int ref_height,
+                                       int this_width, int this_height) {
+  return 2 * this_width >= ref_width && 2 * this_height >= ref_height &&
+         this_width <= 16 * ref_width && this_height <= 16 * ref_height;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_SCALE_H_
diff --git a/third_party/aom/av1/common/scan.c b/third_party/aom/av1/common/scan.c
new file mode 100644
index 0000000000..9ad6c0b2f6
--- /dev/null
+++ b/third_party/aom/av1/common/scan.c
@@ -0,0 +1,6853 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/common_data.h"
+#include "av1/common/scan.h"
+
+#if CONFIG_CB4X4
+DECLARE_ALIGNED(16, static const int16_t, default_scan_2x2[4]) = {
+  0, 1, 2, 3,
+};
+#endif
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_4x4[16]) = {
+  0, 4, 1, 5, 8, 2, 12, 9, 3, 6, 13, 10, 7, 14, 11, 15,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_4x4[16]) = {
+  0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_4x4[16]) = {
+  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, col_scan_4x4[16]) = {
+  0, 4, 8, 1, 12, 5, 9, 2, 13, 6, 10, 3, 7, 14, 11, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, row_scan_4x4[16]) = {
+  0, 1, 4, 2, 5, 3, 6, 8, 9, 7, 12, 10, 13, 11, 14, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_4x8[32]) = {
+  0,  1,  4,  5,  2,  8,  6,  9,  10, 3,  12, 7,  13, 11, 14, 16,
+  17, 15, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_4x8[32]) = {
+  0, 4, 8,  12, 16, 20, 24, 28, 1, 5, 9,  13, 17, 21, 25, 29,
+  2, 6, 10, 14, 18, 22, 26, 30, 3, 7, 11, 15, 19, 23, 27, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_4x8[32]) = {
+  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_8x4[32]) = {
+  0,  1,  8,  9, 2,  16, 10, 17, 18, 3,  24, 11, 25, 19, 26, 4,
+  12, 27, 20, 5, 28, 13, 21, 29, 6,  14, 22, 30, 7,  15, 23, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_8x4[32]) = {
+  0, 8,  16, 24, 1, 9,  17, 25, 2, 10, 18, 26, 3, 11, 19, 27,
+  4, 12, 20, 28, 5, 13, 21, 29, 6, 14, 22, 30, 7, 15, 23, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_8x4[32]) = {
+  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_4x16[64]) = {
+  0,  1,  4,  2,  5,  8,  3,  6,  9,  12, 7,  10, 13, 16, 11, 14,
+  17, 20, 15, 18, 21, 24, 19, 22, 25, 28, 23, 26, 29, 32, 27, 30,
+  33, 36, 31, 34, 37, 40, 35, 38, 41, 44, 39, 42, 45, 48, 43, 46,
+  49, 52, 47, 50, 53, 56, 51, 54, 57, 60, 55, 58, 61, 59, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_16x4[64]) = {
+  0,  1,  16, 2,  17, 32, 3,  18, 33, 48, 4,  19, 34, 49, 5,  20,
+  35, 50, 6,  21, 36, 51, 7,  22, 37, 52, 8,  23, 38, 53, 9,  24,
+  39, 54, 10, 25, 40, 55, 11, 26, 41, 56, 12, 27, 42, 57, 13, 28,
+  43, 58, 14, 29, 44, 59, 15, 30, 45, 60, 31, 46, 61, 47, 62, 63,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_4x16[64]) = {
+  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+  32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+  48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_16x4[64]) = {
+  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+  32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+  48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_4x16[64]) = {
+  0, 4, 8,  12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 60,
+  1, 5, 9,  13, 17, 21, 25, 29, 33, 37, 41, 45, 49, 53, 57, 61,
+  2, 6, 10, 14, 18, 22, 26, 30, 34, 38, 42, 46, 50, 54, 58, 62,
+  3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47, 51, 55, 59, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_16x4[64]) = {
+  0,  16, 32, 48, 1,  17, 33, 49, 2,  18, 34, 50, 3,  19, 35, 51,
+  4,  20, 36, 52, 5,  21, 37, 53, 6,  22, 38, 54, 7,  23, 39, 55,
+  8,  24, 40, 56, 9,  25, 41, 57, 10, 26, 42, 58, 11, 27, 43, 59,
+  12, 28, 44, 60, 13, 29, 45, 61, 14, 30, 46, 62, 15, 31, 47, 63,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_8x32[256]) = {
+  0,   1,   8,   2,   9,   16,  3,   10,  17,  24,  4,   11,  18,  25,  32,
+  5,   12,  19,  26,  33,  40,  6,   13,  20,  27,  34,  41,  48,  7,   14,
+  21,  28,  35,  42,  49,  56,  15,  22,  29,  36,  43,  50,  57,  64,  23,
+  30,  37,  44,  51,  58,  65,  72,  31,  38,  45,  52,  59,  66,  73,  80,
+  39,  46,  53,  60,  67,  74,  81,  88,  47,  54,  61,  68,  75,  82,  89,
+  96,  55,  62,  69,  76,  83,  90,  97,  104, 63,  70,  77,  84,  91,  98,
+  105, 112, 71,  78,  85,  92,  99,  106, 113, 120, 79,  86,  93,  100, 107,
+  114, 121, 128, 87,  94,  101, 108, 115, 122, 129, 136, 95,  102, 109, 116,
+  123, 130, 137, 144, 103, 110, 117, 124, 131, 138, 145, 152, 111, 118, 125,
+  132, 139, 146, 153, 160, 119, 126, 133, 140, 147, 154, 161, 168, 127, 134,
+  141, 148, 155, 162, 169, 176, 135, 142, 149, 156, 163, 170, 177, 184, 143,
+  150, 157, 164, 171, 178, 185, 192, 151, 158, 165, 172, 179, 186, 193, 200,
+  159, 166, 173, 180, 187, 194, 201, 208, 167, 174, 181, 188, 195, 202, 209,
+  216, 175, 182, 189, 196, 203, 210, 217, 224, 183, 190, 197, 204, 211, 218,
+  225, 232, 191, 198, 205, 212, 219, 226, 233, 240, 199, 206, 213, 220, 227,
+  234, 241, 248, 207, 214, 221, 228, 235, 242, 249, 215, 222, 229, 236, 243,
+  250, 223, 230, 237, 244, 251, 231, 238, 245, 252, 239, 246, 253, 247, 254,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_32x8[256]) = {
+  0,   1,   32,  2,   33,  64,  3,   34,  65,  96,  4,   35,  66,  97,  128,
+  5,   36,  67,  98,  129, 160, 6,   37,  68,  99,  130, 161, 192, 7,   38,
+  69,  100, 131, 162, 193, 224, 8,   39,  70,  101, 132, 163, 194, 225, 9,
+  40,  71,  102, 133, 164, 195, 226, 10,  41,  72,  103, 134, 165, 196, 227,
+  11,  42,  73,  104, 135, 166, 197, 228, 12,  43,  74,  105, 136, 167, 198,
+  229, 13,  44,  75,  106, 137, 168, 199, 230, 14,  45,  76,  107, 138, 169,
+  200, 231, 15,  46,  77,  108, 139, 170, 201, 232, 16,  47,  78,  109, 140,
+  171, 202, 233, 17,  48,  79,  110, 141, 172, 203, 234, 18,  49,  80,  111,
+  142, 173, 204, 235, 19,  50,  81,  112, 143, 174, 205, 236, 20,  51,  82,
+  113, 144, 175, 206, 237, 21,  52,  83,  114, 145, 176, 207, 238, 22,  53,
+  84,  115, 146, 177, 208, 239, 23,  54,  85,  116, 147, 178, 209, 240, 24,
+  55,  86,  117, 148, 179, 210, 241, 25,  56,  87,  118, 149, 180, 211, 242,
+  26,  57,  88,  119, 150, 181, 212, 243, 27,  58,  89,  120, 151, 182, 213,
+  244, 28,  59,  90,  121, 152, 183, 214, 245, 29,  60,  91,  122, 153, 184,
+  215, 246, 30,  61,  92,  123, 154, 185, 216, 247, 31,  62,  93,  124, 155,
+  186, 217, 248, 63,  94,  125, 156, 187, 218, 249, 95,  126, 157, 188, 219,
+  250, 127, 158, 189, 220, 251, 159, 190, 221, 252, 191, 222, 253, 223, 254,
+  255,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_8x32[256]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134,
+  135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149,
+  150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164,
+  165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
+  180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194,
+  195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209,
+  210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224,
+  225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
+  240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_32x8[256]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134,
+  135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149,
+  150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164,
+  165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
+  180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194,
+  195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209,
+  210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224,
+  225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
+  240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_8x32[256]) = {
+  0,   8,   16,  24,  32,  40,  48,  56,  64,  72,  80,  88,  96,  104, 112,
+  120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232,
+  240, 248, 1,   9,   17,  25,  33,  41,  49,  57,  65,  73,  81,  89,  97,
+  105, 113, 121, 129, 137, 145, 153, 161, 169, 177, 185, 193, 201, 209, 217,
+  225, 233, 241, 249, 2,   10,  18,  26,  34,  42,  50,  58,  66,  74,  82,
+  90,  98,  106, 114, 122, 130, 138, 146, 154, 162, 170, 178, 186, 194, 202,
+  210, 218, 226, 234, 242, 250, 3,   11,  19,  27,  35,  43,  51,  59,  67,
+  75,  83,  91,  99,  107, 115, 123, 131, 139, 147, 155, 163, 171, 179, 187,
+  195, 203, 211, 219, 227, 235, 243, 251, 4,   12,  20,  28,  36,  44,  52,
+  60,  68,  76,  84,  92,  100, 108, 116, 124, 132, 140, 148, 156, 164, 172,
+  180, 188, 196, 204, 212, 220, 228, 236, 244, 252, 5,   13,  21,  29,  37,
+  45,  53,  61,  69,  77,  85,  93,  101, 109, 117, 125, 133, 141, 149, 157,
+  165, 173, 181, 189, 197, 205, 213, 221, 229, 237, 245, 253, 6,   14,  22,
+  30,  38,  46,  54,  62,  70,  78,  86,  94,  102, 110, 118, 126, 134, 142,
+  150, 158, 166, 174, 182, 190, 198, 206, 214, 222, 230, 238, 246, 254, 7,
+  15,  23,  31,  39,  47,  55,  63,  71,  79,  87,  95,  103, 111, 119, 127,
+  135, 143, 151, 159, 167, 175, 183, 191, 199, 207, 215, 223, 231, 239, 247,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_32x8[256]) = {
+  0,  32, 64, 96,  128, 160, 192, 224, 1,  33, 65, 97,  129, 161, 193, 225,
+  2,  34, 66, 98,  130, 162, 194, 226, 3,  35, 67, 99,  131, 163, 195, 227,
+  4,  36, 68, 100, 132, 164, 196, 228, 5,  37, 69, 101, 133, 165, 197, 229,
+  6,  38, 70, 102, 134, 166, 198, 230, 7,  39, 71, 103, 135, 167, 199, 231,
+  8,  40, 72, 104, 136, 168, 200, 232, 9,  41, 73, 105, 137, 169, 201, 233,
+  10, 42, 74, 106, 138, 170, 202, 234, 11, 43, 75, 107, 139, 171, 203, 235,
+  12, 44, 76, 108, 140, 172, 204, 236, 13, 45, 77, 109, 141, 173, 205, 237,
+  14, 46, 78, 110, 142, 174, 206, 238, 15, 47, 79, 111, 143, 175, 207, 239,
+  16, 48, 80, 112, 144, 176, 208, 240, 17, 49, 81, 113, 145, 177, 209, 241,
+  18, 50, 82, 114, 146, 178, 210, 242, 19, 51, 83, 115, 147, 179, 211, 243,
+  20, 52, 84, 116, 148, 180, 212, 244, 21, 53, 85, 117, 149, 181, 213, 245,
+  22, 54, 86, 118, 150, 182, 214, 246, 23, 55, 87, 119, 151, 183, 215, 247,
+  24, 56, 88, 120, 152, 184, 216, 248, 25, 57, 89, 121, 153, 185, 217, 249,
+  26, 58, 90, 122, 154, 186, 218, 250, 27, 59, 91, 123, 155, 187, 219, 251,
+  28, 60, 92, 124, 156, 188, 220, 252, 29, 61, 93, 125, 157, 189, 221, 253,
+  30, 62, 94, 126, 158, 190, 222, 254, 31, 63, 95, 127, 159, 191, 223, 255,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_8x8[64]) = {
+  0,  8,  1,  16, 9,  2,  17, 24, 10, 3,  18, 25, 32, 11, 4,  26,
+  33, 19, 40, 12, 34, 27, 5,  41, 20, 48, 13, 35, 42, 28, 21, 6,
+  49, 56, 36, 43, 29, 7,  14, 50, 57, 44, 22, 37, 15, 51, 58, 30,
+  45, 23, 52, 59, 38, 31, 60, 53, 46, 39, 61, 54, 47, 62, 55, 63,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_8x8[64]) = {
+  0, 8,  16, 24, 32, 40, 48, 56, 1, 9,  17, 25, 33, 41, 49, 57,
+  2, 10, 18, 26, 34, 42, 50, 58, 3, 11, 19, 27, 35, 43, 51, 59,
+  4, 12, 20, 28, 36, 44, 52, 60, 5, 13, 21, 29, 37, 45, 53, 61,
+  6, 14, 22, 30, 38, 46, 54, 62, 7, 15, 23, 31, 39, 47, 55, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_8x8[64]) = {
+  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+  32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+  48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, col_scan_8x8[64]) = {
+  0,  8,  16, 1,  24, 9,  32, 17, 2,  40, 25, 10, 33, 18, 48, 3,
+  26, 41, 11, 56, 19, 34, 4,  49, 27, 42, 12, 35, 20, 57, 50, 28,
+  5,  43, 13, 36, 58, 51, 21, 44, 6,  29, 59, 37, 14, 52, 22, 7,
+  45, 60, 30, 15, 38, 53, 23, 46, 31, 61, 39, 54, 47, 62, 55, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, row_scan_8x8[64]) = {
+  0,  1,  2,  8,  9,  3,  16, 10, 4,  17, 11, 24, 5,  18, 25, 12,
+  19, 26, 32, 6,  13, 20, 33, 27, 7,  34, 40, 21, 28, 41, 14, 35,
+  48, 42, 29, 36, 49, 22, 43, 15, 56, 37, 50, 44, 30, 57, 23, 51,
+  58, 45, 38, 52, 31, 59, 53, 46, 60, 39, 61, 47, 54, 55, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_8x16[128]) = {
+  0,   1,   8,   2,   9,   16,  3,   10,  17,  24,  4,   11,  18,  25,  32,
+  5,   12,  19,  26,  33,  40,  6,   13,  20,  27,  34,  41,  48,  7,   14,
+  21,  28,  35,  42,  49,  56,  15,  22,  29,  36,  43,  50,  57,  64,  23,
+  30,  37,  44,  51,  58,  65,  72,  31,  38,  45,  52,  59,  66,  73,  80,
+  39,  46,  53,  60,  67,  74,  81,  88,  47,  54,  61,  68,  75,  82,  89,
+  96,  55,  62,  69,  76,  83,  90,  97,  104, 63,  70,  77,  84,  91,  98,
+  105, 112, 71,  78,  85,  92,  99,  106, 113, 120, 79,  86,  93,  100, 107,
+  114, 121, 87,  94,  101, 108, 115, 122, 95,  102, 109, 116, 123, 103, 110,
+  117, 124, 111, 118, 125, 119, 126, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_16x8[128]) = {
+  0,   1,  16,  2,   17,  32,  3,  18, 33,  48,  4,   19,  34,  49,  64,  5,
+  20,  35, 50,  65,  80,  6,   21, 36, 51,  66,  81,  96,  7,   22,  37,  52,
+  67,  82, 97,  112, 8,   23,  38, 53, 68,  83,  98,  113, 9,   24,  39,  54,
+  69,  84, 99,  114, 10,  25,  40, 55, 70,  85,  100, 115, 11,  26,  41,  56,
+  71,  86, 101, 116, 12,  27,  42, 57, 72,  87,  102, 117, 13,  28,  43,  58,
+  73,  88, 103, 118, 14,  29,  44, 59, 74,  89,  104, 119, 15,  30,  45,  60,
+  75,  90, 105, 120, 31,  46,  61, 76, 91,  106, 121, 47,  62,  77,  92,  107,
+  122, 63, 78,  93,  108, 123, 79, 94, 109, 124, 95,  110, 125, 111, 126, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_8x16[128]) = {
+  0, 8,  16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96,  104, 112, 120,
+  1, 9,  17, 25, 33, 41, 49, 57, 65, 73, 81, 89, 97,  105, 113, 121,
+  2, 10, 18, 26, 34, 42, 50, 58, 66, 74, 82, 90, 98,  106, 114, 122,
+  3, 11, 19, 27, 35, 43, 51, 59, 67, 75, 83, 91, 99,  107, 115, 123,
+  4, 12, 20, 28, 36, 44, 52, 60, 68, 76, 84, 92, 100, 108, 116, 124,
+  5, 13, 21, 29, 37, 45, 53, 61, 69, 77, 85, 93, 101, 109, 117, 125,
+  6, 14, 22, 30, 38, 46, 54, 62, 70, 78, 86, 94, 102, 110, 118, 126,
+  7, 15, 23, 31, 39, 47, 55, 63, 71, 79, 87, 95, 103, 111, 119, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_16x8[128]) = {
+  0,  16, 32, 48, 64, 80, 96,  112, 1,  17, 33, 49, 65, 81, 97,  113,
+  2,  18, 34, 50, 66, 82, 98,  114, 3,  19, 35, 51, 67, 83, 99,  115,
+  4,  20, 36, 52, 68, 84, 100, 116, 5,  21, 37, 53, 69, 85, 101, 117,
+  6,  22, 38, 54, 70, 86, 102, 118, 7,  23, 39, 55, 71, 87, 103, 119,
+  8,  24, 40, 56, 72, 88, 104, 120, 9,  25, 41, 57, 73, 89, 105, 121,
+  10, 26, 42, 58, 74, 90, 106, 122, 11, 27, 43, 59, 75, 91, 107, 123,
+  12, 28, 44, 60, 76, 92, 108, 124, 13, 29, 45, 61, 77, 93, 109, 125,
+  14, 30, 46, 62, 78, 94, 110, 126, 15, 31, 47, 63, 79, 95, 111, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_8x16[128]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_16x8[128]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_16x32[512]) = {
+  0,   1,   16,  2,   17,  32,  3,   18,  33,  48,  4,   19,  34,  49,  64,
+  5,   20,  35,  50,  65,  80,  6,   21,  36,  51,  66,  81,  96,  7,   22,
+  37,  52,  67,  82,  97,  112, 8,   23,  38,  53,  68,  83,  98,  113, 128,
+  9,   24,  39,  54,  69,  84,  99,  114, 129, 144, 10,  25,  40,  55,  70,
+  85,  100, 115, 130, 145, 160, 11,  26,  41,  56,  71,  86,  101, 116, 131,
+  146, 161, 176, 12,  27,  42,  57,  72,  87,  102, 117, 132, 147, 162, 177,
+  192, 13,  28,  43,  58,  73,  88,  103, 118, 133, 148, 163, 178, 193, 208,
+  14,  29,  44,  59,  74,  89,  104, 119, 134, 149, 164, 179, 194, 209, 224,
+  15,  30,  45,  60,  75,  90,  105, 120, 135, 150, 165, 180, 195, 210, 225,
+  240, 31,  46,  61,  76,  91,  106, 121, 136, 151, 166, 181, 196, 211, 226,
+  241, 256, 47,  62,  77,  92,  107, 122, 137, 152, 167, 182, 197, 212, 227,
+  242, 257, 272, 63,  78,  93,  108, 123, 138, 153, 168, 183, 198, 213, 228,
+  243, 258, 273, 288, 79,  94,  109, 124, 139, 154, 169, 184, 199, 214, 229,
+  244, 259, 274, 289, 304, 95,  110, 125, 140, 155, 170, 185, 200, 215, 230,
+  245, 260, 275, 290, 305, 320, 111, 126, 141, 156, 171, 186, 201, 216, 231,
+  246, 261, 276, 291, 306, 321, 336, 127, 142, 157, 172, 187, 202, 217, 232,
+  247, 262, 277, 292, 307, 322, 337, 352, 143, 158, 173, 188, 203, 218, 233,
+  248, 263, 278, 293, 308, 323, 338, 353, 368, 159, 174, 189, 204, 219, 234,
+  249, 264, 279, 294, 309, 324, 339, 354, 369, 384, 175, 190, 205, 220, 235,
+  250, 265, 280, 295, 310, 325, 340, 355, 370, 385, 400, 191, 206, 221, 236,
+  251, 266, 281, 296, 311, 326, 341, 356, 371, 386, 401, 416, 207, 222, 237,
+  252, 267, 282, 297, 312, 327, 342, 357, 372, 387, 402, 417, 432, 223, 238,
+  253, 268, 283, 298, 313, 328, 343, 358, 373, 388, 403, 418, 433, 448, 239,
+  254, 269, 284, 299, 314, 329, 344, 359, 374, 389, 404, 419, 434, 449, 464,
+  255, 270, 285, 300, 315, 330, 345, 360, 375, 390, 405, 420, 435, 450, 465,
+  480, 271, 286, 301, 316, 331, 346, 361, 376, 391, 406, 421, 436, 451, 466,
+  481, 496, 287, 302, 317, 332, 347, 362, 377, 392, 407, 422, 437, 452, 467,
+  482, 497, 303, 318, 333, 348, 363, 378, 393, 408, 423, 438, 453, 468, 483,
+  498, 319, 334, 349, 364, 379, 394, 409, 424, 439, 454, 469, 484, 499, 335,
+  350, 365, 380, 395, 410, 425, 440, 455, 470, 485, 500, 351, 366, 381, 396,
+  411, 426, 441, 456, 471, 486, 501, 367, 382, 397, 412, 427, 442, 457, 472,
+  487, 502, 383, 398, 413, 428, 443, 458, 473, 488, 503, 399, 414, 429, 444,
+  459, 474, 489, 504, 415, 430, 445, 460, 475, 490, 505, 431, 446, 461, 476,
+  491, 506, 447, 462, 477, 492, 507, 463, 478, 493, 508, 479, 494, 509, 495,
+  510, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_32x16[512]) = {
+  0,   1,   32,  2,   33,  64,  3,   34,  65,  96,  4,   35,  66,  97,  128,
+  5,   36,  67,  98,  129, 160, 6,   37,  68,  99,  130, 161, 192, 7,   38,
+  69,  100, 131, 162, 193, 224, 8,   39,  70,  101, 132, 163, 194, 225, 256,
+  9,   40,  71,  102, 133, 164, 195, 226, 257, 288, 10,  41,  72,  103, 134,
+  165, 196, 227, 258, 289, 320, 11,  42,  73,  104, 135, 166, 197, 228, 259,
+  290, 321, 352, 12,  43,  74,  105, 136, 167, 198, 229, 260, 291, 322, 353,
+  384, 13,  44,  75,  106, 137, 168, 199, 230, 261, 292, 323, 354, 385, 416,
+  14,  45,  76,  107, 138, 169, 200, 231, 262, 293, 324, 355, 386, 417, 448,
+  15,  46,  77,  108, 139, 170, 201, 232, 263, 294, 325, 356, 387, 418, 449,
+  480, 16,  47,  78,  109, 140, 171, 202, 233, 264, 295, 326, 357, 388, 419,
+  450, 481, 17,  48,  79,  110, 141, 172, 203, 234, 265, 296, 327, 358, 389,
+  420, 451, 482, 18,  49,  80,  111, 142, 173, 204, 235, 266, 297, 328, 359,
+  390, 421, 452, 483, 19,  50,  81,  112, 143, 174, 205, 236, 267, 298, 329,
+  360, 391, 422, 453, 484, 20,  51,  82,  113, 144, 175, 206, 237, 268, 299,
+  330, 361, 392, 423, 454, 485, 21,  52,  83,  114, 145, 176, 207, 238, 269,
+  300, 331, 362, 393, 424, 455, 486, 22,  53,  84,  115, 146, 177, 208, 239,
+  270, 301, 332, 363, 394, 425, 456, 487, 23,  54,  85,  116, 147, 178, 209,
+  240, 271, 302, 333, 364, 395, 426, 457, 488, 24,  55,  86,  117, 148, 179,
+  210, 241, 272, 303, 334, 365, 396, 427, 458, 489, 25,  56,  87,  118, 149,
+  180, 211, 242, 273, 304, 335, 366, 397, 428, 459, 490, 26,  57,  88,  119,
+  150, 181, 212, 243, 274, 305, 336, 367, 398, 429, 460, 491, 27,  58,  89,
+  120, 151, 182, 213, 244, 275, 306, 337, 368, 399, 430, 461, 492, 28,  59,
+  90,  121, 152, 183, 214, 245, 276, 307, 338, 369, 400, 431, 462, 493, 29,
+  60,  91,  122, 153, 184, 215, 246, 277, 308, 339, 370, 401, 432, 463, 494,
+  30,  61,  92,  123, 154, 185, 216, 247, 278, 309, 340, 371, 402, 433, 464,
+  495, 31,  62,  93,  124, 155, 186, 217, 248, 279, 310, 341, 372, 403, 434,
+  465, 496, 63,  94,  125, 156, 187, 218, 249, 280, 311, 342, 373, 404, 435,
+  466, 497, 95,  126, 157, 188, 219, 250, 281, 312, 343, 374, 405, 436, 467,
+  498, 127, 158, 189, 220, 251, 282, 313, 344, 375, 406, 437, 468, 499, 159,
+  190, 221, 252, 283, 314, 345, 376, 407, 438, 469, 500, 191, 222, 253, 284,
+  315, 346, 377, 408, 439, 470, 501, 223, 254, 285, 316, 347, 378, 409, 440,
+  471, 502, 255, 286, 317, 348, 379, 410, 441, 472, 503, 287, 318, 349, 380,
+  411, 442, 473, 504, 319, 350, 381, 412, 443, 474, 505, 351, 382, 413, 444,
+  475, 506, 383, 414, 445, 476, 507, 415, 446, 477, 508, 447, 478, 509, 479,
+  510, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_16x32[512]) = {
+  0,   16,  32,  48,  64,  80,  96,  112, 128, 144, 160, 176, 192, 208, 224,
+  240, 256, 272, 288, 304, 320, 336, 352, 368, 384, 400, 416, 432, 448, 464,
+  480, 496, 1,   17,  33,  49,  65,  81,  97,  113, 129, 145, 161, 177, 193,
+  209, 225, 241, 257, 273, 289, 305, 321, 337, 353, 369, 385, 401, 417, 433,
+  449, 465, 481, 497, 2,   18,  34,  50,  66,  82,  98,  114, 130, 146, 162,
+  178, 194, 210, 226, 242, 258, 274, 290, 306, 322, 338, 354, 370, 386, 402,
+  418, 434, 450, 466, 482, 498, 3,   19,  35,  51,  67,  83,  99,  115, 131,
+  147, 163, 179, 195, 211, 227, 243, 259, 275, 291, 307, 323, 339, 355, 371,
+  387, 403, 419, 435, 451, 467, 483, 499, 4,   20,  36,  52,  68,  84,  100,
+  116, 132, 148, 164, 180, 196, 212, 228, 244, 260, 276, 292, 308, 324, 340,
+  356, 372, 388, 404, 420, 436, 452, 468, 484, 500, 5,   21,  37,  53,  69,
+  85,  101, 117, 133, 149, 165, 181, 197, 213, 229, 245, 261, 277, 293, 309,
+  325, 341, 357, 373, 389, 405, 421, 437, 453, 469, 485, 501, 6,   22,  38,
+  54,  70,  86,  102, 118, 134, 150, 166, 182, 198, 214, 230, 246, 262, 278,
+  294, 310, 326, 342, 358, 374, 390, 406, 422, 438, 454, 470, 486, 502, 7,
+  23,  39,  55,  71,  87,  103, 119, 135, 151, 167, 183, 199, 215, 231, 247,
+  263, 279, 295, 311, 327, 343, 359, 375, 391, 407, 423, 439, 455, 471, 487,
+  503, 8,   24,  40,  56,  72,  88,  104, 120, 136, 152, 168, 184, 200, 216,
+  232, 248, 264, 280, 296, 312, 328, 344, 360, 376, 392, 408, 424, 440, 456,
+  472, 488, 504, 9,   25,  41,  57,  73,  89,  105, 121, 137, 153, 169, 185,
+  201, 217, 233, 249, 265, 281, 297, 313, 329, 345, 361, 377, 393, 409, 425,
+  441, 457, 473, 489, 505, 10,  26,  42,  58,  74,  90,  106, 122, 138, 154,
+  170, 186, 202, 218, 234, 250, 266, 282, 298, 314, 330, 346, 362, 378, 394,
+  410, 426, 442, 458, 474, 490, 506, 11,  27,  43,  59,  75,  91,  107, 123,
+  139, 155, 171, 187, 203, 219, 235, 251, 267, 283, 299, 315, 331, 347, 363,
+  379, 395, 411, 427, 443, 459, 475, 491, 507, 12,  28,  44,  60,  76,  92,
+  108, 124, 140, 156, 172, 188, 204, 220, 236, 252, 268, 284, 300, 316, 332,
+  348, 364, 380, 396, 412, 428, 444, 460, 476, 492, 508, 13,  29,  45,  61,
+  77,  93,  109, 125, 141, 157, 173, 189, 205, 221, 237, 253, 269, 285, 301,
+  317, 333, 349, 365, 381, 397, 413, 429, 445, 461, 477, 493, 509, 14,  30,
+  46,  62,  78,  94,  110, 126, 142, 158, 174, 190, 206, 222, 238, 254, 270,
+  286, 302, 318, 334, 350, 366, 382, 398, 414, 430, 446, 462, 478, 494, 510,
+  15,  31,  47,  63,  79,  95,  111, 127, 143, 159, 175, 191, 207, 223, 239,
+  255, 271, 287, 303, 319, 335, 351, 367, 383, 399, 415, 431, 447, 463, 479,
+  495, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_32x16[512]) = {
+  0,  32, 64, 96,  128, 160, 192, 224, 256, 288, 320, 352, 384, 416, 448, 480,
+  1,  33, 65, 97,  129, 161, 193, 225, 257, 289, 321, 353, 385, 417, 449, 481,
+  2,  34, 66, 98,  130, 162, 194, 226, 258, 290, 322, 354, 386, 418, 450, 482,
+  3,  35, 67, 99,  131, 163, 195, 227, 259, 291, 323, 355, 387, 419, 451, 483,
+  4,  36, 68, 100, 132, 164, 196, 228, 260, 292, 324, 356, 388, 420, 452, 484,
+  5,  37, 69, 101, 133, 165, 197, 229, 261, 293, 325, 357, 389, 421, 453, 485,
+  6,  38, 70, 102, 134, 166, 198, 230, 262, 294, 326, 358, 390, 422, 454, 486,
+  7,  39, 71, 103, 135, 167, 199, 231, 263, 295, 327, 359, 391, 423, 455, 487,
+  8,  40, 72, 104, 136, 168, 200, 232, 264, 296, 328, 360, 392, 424, 456, 488,
+  9,  41, 73, 105, 137, 169, 201, 233, 265, 297, 329, 361, 393, 425, 457, 489,
+  10, 42, 74, 106, 138, 170, 202, 234, 266, 298, 330, 362, 394, 426, 458, 490,
+  11, 43, 75, 107, 139, 171, 203, 235, 267, 299, 331, 363, 395, 427, 459, 491,
+  12, 44, 76, 108, 140, 172, 204, 236, 268, 300, 332, 364, 396, 428, 460, 492,
+  13, 45, 77, 109, 141, 173, 205, 237, 269, 301, 333, 365, 397, 429, 461, 493,
+  14, 46, 78, 110, 142, 174, 206, 238, 270, 302, 334, 366, 398, 430, 462, 494,
+  15, 47, 79, 111, 143, 175, 207, 239, 271, 303, 335, 367, 399, 431, 463, 495,
+  16, 48, 80, 112, 144, 176, 208, 240, 272, 304, 336, 368, 400, 432, 464, 496,
+  17, 49, 81, 113, 145, 177, 209, 241, 273, 305, 337, 369, 401, 433, 465, 497,
+  18, 50, 82, 114, 146, 178, 210, 242, 274, 306, 338, 370, 402, 434, 466, 498,
+  19, 51, 83, 115, 147, 179, 211, 243, 275, 307, 339, 371, 403, 435, 467, 499,
+  20, 52, 84, 116, 148, 180, 212, 244, 276, 308, 340, 372, 404, 436, 468, 500,
+  21, 53, 85, 117, 149, 181, 213, 245, 277, 309, 341, 373, 405, 437, 469, 501,
+  22, 54, 86, 118, 150, 182, 214, 246, 278, 310, 342, 374, 406, 438, 470, 502,
+  23, 55, 87, 119, 151, 183, 215, 247, 279, 311, 343, 375, 407, 439, 471, 503,
+  24, 56, 88, 120, 152, 184, 216, 248, 280, 312, 344, 376, 408, 440, 472, 504,
+  25, 57, 89, 121, 153, 185, 217, 249, 281, 313, 345, 377, 409, 441, 473, 505,
+  26, 58, 90, 122, 154, 186, 218, 250, 282, 314, 346, 378, 410, 442, 474, 506,
+  27, 59, 91, 123, 155, 187, 219, 251, 283, 315, 347, 379, 411, 443, 475, 507,
+  28, 60, 92, 124, 156, 188, 220, 252, 284, 316, 348, 380, 412, 444, 476, 508,
+  29, 61, 93, 125, 157, 189, 221, 253, 285, 317, 349, 381, 413, 445, 477, 509,
+  30, 62, 94, 126, 158, 190, 222, 254, 286, 318, 350, 382, 414, 446, 478, 510,
+  31, 63, 95, 127, 159, 191, 223, 255, 287, 319, 351, 383, 415, 447, 479, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_16x32[512]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134,
+  135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149,
+  150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164,
+  165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
+  180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194,
+  195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209,
+  210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224,
+  225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
+  240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254,
+  255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269,
+  270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284,
+  285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299,
+  300, 301, 302, 303, 304, 305, 306, 307, 308, 309, 310, 311, 312, 313, 314,
+  315, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329,
+  330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 343, 344,
+  345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359,
+  360, 361, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 373, 374,
+  375, 376, 377, 378, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389,
+  390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404,
+  405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419,
+  420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431, 432, 433, 434,
+  435, 436, 437, 438, 439, 440, 441, 442, 443, 444, 445, 446, 447, 448, 449,
+  450, 451, 452, 453, 454, 455, 456, 457, 458, 459, 460, 461, 462, 463, 464,
+  465, 466, 467, 468, 469, 470, 471, 472, 473, 474, 475, 476, 477, 478, 479,
+  480, 481, 482, 483, 484, 485, 486, 487, 488, 489, 490, 491, 492, 493, 494,
+  495, 496, 497, 498, 499, 500, 501, 502, 503, 504, 505, 506, 507, 508, 509,
+  510, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_32x16[512]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134,
+  135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149,
+  150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164,
+  165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
+  180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194,
+  195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209,
+  210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224,
+  225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
+  240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254,
+  255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269,
+  270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284,
+  285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299,
+  300, 301, 302, 303, 304, 305, 306, 307, 308, 309, 310, 311, 312, 313, 314,
+  315, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329,
+  330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 343, 344,
+  345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359,
+  360, 361, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 373, 374,
+  375, 376, 377, 378, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389,
+  390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404,
+  405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419,
+  420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431, 432, 433, 434,
+  435, 436, 437, 438, 439, 440, 441, 442, 443, 444, 445, 446, 447, 448, 449,
+  450, 451, 452, 453, 454, 455, 456, 457, 458, 459, 460, 461, 462, 463, 464,
+  465, 466, 467, 468, 469, 470, 471, 472, 473, 474, 475, 476, 477, 478, 479,
+  480, 481, 482, 483, 484, 485, 486, 487, 488, 489, 490, 491, 492, 493, 494,
+  495, 496, 497, 498, 499, 500, 501, 502, 503, 504, 505, 506, 507, 508, 509,
+  510, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_16x16[256]) = {
+  0,   16,  1,   32,  17,  2,   48,  33,  18,  3,   64,  34,  49,  19,  65,
+  80,  50,  4,   35,  66,  20,  81,  96,  51,  5,   36,  82,  97,  67,  112,
+  21,  52,  98,  37,  83,  113, 6,   68,  128, 53,  22,  99,  114, 84,  7,
+  129, 38,  69,  100, 115, 144, 130, 85,  54,  23,  8,   145, 39,  70,  116,
+  101, 131, 160, 146, 55,  86,  24,  71,  132, 117, 161, 40,  9,   102, 147,
+  176, 162, 87,  56,  25,  133, 118, 177, 148, 72,  103, 41,  163, 10,  192,
+  178, 88,  57,  134, 149, 119, 26,  164, 73,  104, 193, 42,  179, 208, 11,
+  135, 89,  165, 120, 150, 58,  194, 180, 27,  74,  209, 105, 151, 136, 43,
+  90,  224, 166, 195, 181, 121, 210, 59,  12,  152, 106, 167, 196, 75,  137,
+  225, 211, 240, 182, 122, 91,  28,  197, 13,  226, 168, 183, 153, 44,  212,
+  138, 107, 241, 60,  29,  123, 198, 184, 227, 169, 242, 76,  213, 154, 45,
+  92,  14,  199, 139, 61,  228, 214, 170, 185, 243, 108, 77,  155, 30,  15,
+  200, 229, 124, 215, 244, 93,  46,  186, 171, 201, 109, 140, 230, 62,  216,
+  245, 31,  125, 78,  156, 231, 47,  187, 202, 217, 94,  246, 141, 63,  232,
+  172, 110, 247, 157, 79,  218, 203, 126, 233, 188, 248, 95,  173, 142, 219,
+  111, 249, 234, 158, 127, 189, 204, 250, 235, 143, 174, 220, 205, 159, 251,
+  190, 221, 175, 236, 237, 191, 206, 252, 222, 253, 207, 238, 223, 254, 239,
+  255,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_16x16[256]) = {
+  0,  16, 32, 48, 64, 80, 96,  112, 128, 144, 160, 176, 192, 208, 224, 240,
+  1,  17, 33, 49, 65, 81, 97,  113, 129, 145, 161, 177, 193, 209, 225, 241,
+  2,  18, 34, 50, 66, 82, 98,  114, 130, 146, 162, 178, 194, 210, 226, 242,
+  3,  19, 35, 51, 67, 83, 99,  115, 131, 147, 163, 179, 195, 211, 227, 243,
+  4,  20, 36, 52, 68, 84, 100, 116, 132, 148, 164, 180, 196, 212, 228, 244,
+  5,  21, 37, 53, 69, 85, 101, 117, 133, 149, 165, 181, 197, 213, 229, 245,
+  6,  22, 38, 54, 70, 86, 102, 118, 134, 150, 166, 182, 198, 214, 230, 246,
+  7,  23, 39, 55, 71, 87, 103, 119, 135, 151, 167, 183, 199, 215, 231, 247,
+  8,  24, 40, 56, 72, 88, 104, 120, 136, 152, 168, 184, 200, 216, 232, 248,
+  9,  25, 41, 57, 73, 89, 105, 121, 137, 153, 169, 185, 201, 217, 233, 249,
+  10, 26, 42, 58, 74, 90, 106, 122, 138, 154, 170, 186, 202, 218, 234, 250,
+  11, 27, 43, 59, 75, 91, 107, 123, 139, 155, 171, 187, 203, 219, 235, 251,
+  12, 28, 44, 60, 76, 92, 108, 124, 140, 156, 172, 188, 204, 220, 236, 252,
+  13, 29, 45, 61, 77, 93, 109, 125, 141, 157, 173, 189, 205, 221, 237, 253,
+  14, 30, 46, 62, 78, 94, 110, 126, 142, 158, 174, 190, 206, 222, 238, 254,
+  15, 31, 47, 63, 79, 95, 111, 127, 143, 159, 175, 191, 207, 223, 239, 255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_16x16[256]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134,
+  135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149,
+  150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164,
+  165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
+  180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194,
+  195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209,
+  210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224,
+  225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
+  240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254,
+  255,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, col_scan_16x16[256]) = {
+  0,   16,  32,  48,  1,   64,  17,  80,  33,  96,  49,  2,   65,  112, 18,
+  81,  34,  128, 50,  97,  3,   66,  144, 19,  113, 35,  82,  160, 98,  51,
+  129, 4,   67,  176, 20,  114, 145, 83,  36,  99,  130, 52,  192, 5,   161,
+  68,  115, 21,  146, 84,  208, 177, 37,  131, 100, 53,  162, 224, 69,  6,
+  116, 193, 147, 85,  22,  240, 132, 38,  178, 101, 163, 54,  209, 117, 70,
+  7,   148, 194, 86,  179, 225, 23,  133, 39,  164, 8,   102, 210, 241, 55,
+  195, 118, 149, 71,  180, 24,  87,  226, 134, 165, 211, 40,  103, 56,  72,
+  150, 196, 242, 119, 9,   181, 227, 88,  166, 25,  135, 41,  104, 212, 57,
+  151, 197, 120, 73,  243, 182, 136, 167, 213, 89,  10,  228, 105, 152, 198,
+  26,  42,  121, 183, 244, 168, 58,  137, 229, 74,  214, 90,  153, 199, 184,
+  11,  106, 245, 27,  122, 230, 169, 43,  215, 59,  200, 138, 185, 246, 75,
+  12,  91,  154, 216, 231, 107, 28,  44,  201, 123, 170, 60,  247, 232, 76,
+  139, 13,  92,  217, 186, 248, 155, 108, 29,  124, 45,  202, 233, 171, 61,
+  14,  77,  140, 15,  249, 93,  30,  187, 156, 218, 46,  109, 125, 62,  172,
+  78,  203, 31,  141, 234, 94,  47,  188, 63,  157, 110, 250, 219, 79,  126,
+  204, 173, 142, 95,  189, 111, 235, 158, 220, 251, 127, 174, 143, 205, 236,
+  159, 190, 221, 252, 175, 206, 237, 191, 253, 222, 238, 207, 254, 223, 239,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, row_scan_16x16[256]) = {
+  0,   1,   2,   16,  3,   17,  4,   18,  32,  5,   33,  19,  6,   34,  48,
+  20,  49,  7,   35,  21,  50,  64,  8,   36,  65,  22,  51,  37,  80,  9,
+  66,  52,  23,  38,  81,  67,  10,  53,  24,  82,  68,  96,  39,  11,  54,
+  83,  97,  69,  25,  98,  84,  40,  112, 55,  12,  70,  99,  113, 85,  26,
+  41,  56,  114, 100, 13,  71,  128, 86,  27,  115, 101, 129, 42,  57,  72,
+  116, 14,  87,  130, 102, 144, 73,  131, 117, 28,  58,  15,  88,  43,  145,
+  103, 132, 146, 118, 74,  160, 89,  133, 104, 29,  59,  147, 119, 44,  161,
+  148, 90,  105, 134, 162, 120, 176, 75,  135, 149, 30,  60,  163, 177, 45,
+  121, 91,  106, 164, 178, 150, 192, 136, 165, 179, 31,  151, 193, 76,  122,
+  61,  137, 194, 107, 152, 180, 208, 46,  166, 167, 195, 92,  181, 138, 209,
+  123, 153, 224, 196, 77,  168, 210, 182, 240, 108, 197, 62,  154, 225, 183,
+  169, 211, 47,  139, 93,  184, 226, 212, 241, 198, 170, 124, 155, 199, 78,
+  213, 185, 109, 227, 200, 63,  228, 242, 140, 214, 171, 186, 156, 229, 243,
+  125, 94,  201, 244, 215, 216, 230, 141, 187, 202, 79,  172, 110, 157, 245,
+  217, 231, 95,  246, 232, 126, 203, 247, 233, 173, 218, 142, 111, 158, 188,
+  248, 127, 234, 219, 249, 189, 204, 143, 174, 159, 250, 235, 205, 220, 175,
+  190, 251, 221, 191, 206, 236, 207, 237, 252, 222, 253, 223, 238, 239, 254,
+  255,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, mcol_scan_32x32[1024]) = {
+  0,   32,   64,  96,   128, 160,  192, 224,  256, 288,  320, 352,  384, 416,
+  448, 480,  512, 544,  576, 608,  640, 672,  704, 736,  768, 800,  832, 864,
+  896, 928,  960, 992,  1,   33,   65,  97,   129, 161,  193, 225,  257, 289,
+  321, 353,  385, 417,  449, 481,  513, 545,  577, 609,  641, 673,  705, 737,
+  769, 801,  833, 865,  897, 929,  961, 993,  2,   34,   66,  98,   130, 162,
+  194, 226,  258, 290,  322, 354,  386, 418,  450, 482,  514, 546,  578, 610,
+  642, 674,  706, 738,  770, 802,  834, 866,  898, 930,  962, 994,  3,   35,
+  67,  99,   131, 163,  195, 227,  259, 291,  323, 355,  387, 419,  451, 483,
+  515, 547,  579, 611,  643, 675,  707, 739,  771, 803,  835, 867,  899, 931,
+  963, 995,  4,   36,   68,  100,  132, 164,  196, 228,  260, 292,  324, 356,
+  388, 420,  452, 484,  516, 548,  580, 612,  644, 676,  708, 740,  772, 804,
+  836, 868,  900, 932,  964, 996,  5,   37,   69,  101,  133, 165,  197, 229,
+  261, 293,  325, 357,  389, 421,  453, 485,  517, 549,  581, 613,  645, 677,
+  709, 741,  773, 805,  837, 869,  901, 933,  965, 997,  6,   38,   70,  102,
+  134, 166,  198, 230,  262, 294,  326, 358,  390, 422,  454, 486,  518, 550,
+  582, 614,  646, 678,  710, 742,  774, 806,  838, 870,  902, 934,  966, 998,
+  7,   39,   71,  103,  135, 167,  199, 231,  263, 295,  327, 359,  391, 423,
+  455, 487,  519, 551,  583, 615,  647, 679,  711, 743,  775, 807,  839, 871,
+  903, 935,  967, 999,  8,   40,   72,  104,  136, 168,  200, 232,  264, 296,
+  328, 360,  392, 424,  456, 488,  520, 552,  584, 616,  648, 680,  712, 744,
+  776, 808,  840, 872,  904, 936,  968, 1000, 9,   41,   73,  105,  137, 169,
+  201, 233,  265, 297,  329, 361,  393, 425,  457, 489,  521, 553,  585, 617,
+  649, 681,  713, 745,  777, 809,  841, 873,  905, 937,  969, 1001, 10,  42,
+  74,  106,  138, 170,  202, 234,  266, 298,  330, 362,  394, 426,  458, 490,
+  522, 554,  586, 618,  650, 682,  714, 746,  778, 810,  842, 874,  906, 938,
+  970, 1002, 11,  43,   75,  107,  139, 171,  203, 235,  267, 299,  331, 363,
+  395, 427,  459, 491,  523, 555,  587, 619,  651, 683,  715, 747,  779, 811,
+  843, 875,  907, 939,  971, 1003, 12,  44,   76,  108,  140, 172,  204, 236,
+  268, 300,  332, 364,  396, 428,  460, 492,  524, 556,  588, 620,  652, 684,
+  716, 748,  780, 812,  844, 876,  908, 940,  972, 1004, 13,  45,   77,  109,
+  141, 173,  205, 237,  269, 301,  333, 365,  397, 429,  461, 493,  525, 557,
+  589, 621,  653, 685,  717, 749,  781, 813,  845, 877,  909, 941,  973, 1005,
+  14,  46,   78,  110,  142, 174,  206, 238,  270, 302,  334, 366,  398, 430,
+  462, 494,  526, 558,  590, 622,  654, 686,  718, 750,  782, 814,  846, 878,
+  910, 942,  974, 1006, 15,  47,   79,  111,  143, 175,  207, 239,  271, 303,
+  335, 367,  399, 431,  463, 495,  527, 559,  591, 623,  655, 687,  719, 751,
+  783, 815,  847, 879,  911, 943,  975, 1007, 16,  48,   80,  112,  144, 176,
+  208, 240,  272, 304,  336, 368,  400, 432,  464, 496,  528, 560,  592, 624,
+  656, 688,  720, 752,  784, 816,  848, 880,  912, 944,  976, 1008, 17,  49,
+  81,  113,  145, 177,  209, 241,  273, 305,  337, 369,  401, 433,  465, 497,
+  529, 561,  593, 625,  657, 689,  721, 753,  785, 817,  849, 881,  913, 945,
+  977, 1009, 18,  50,   82,  114,  146, 178,  210, 242,  274, 306,  338, 370,
+  402, 434,  466, 498,  530, 562,  594, 626,  658, 690,  722, 754,  786, 818,
+  850, 882,  914, 946,  978, 1010, 19,  51,   83,  115,  147, 179,  211, 243,
+  275, 307,  339, 371,  403, 435,  467, 499,  531, 563,  595, 627,  659, 691,
+  723, 755,  787, 819,  851, 883,  915, 947,  979, 1011, 20,  52,   84,  116,
+  148, 180,  212, 244,  276, 308,  340, 372,  404, 436,  468, 500,  532, 564,
+  596, 628,  660, 692,  724, 756,  788, 820,  852, 884,  916, 948,  980, 1012,
+  21,  53,   85,  117,  149, 181,  213, 245,  277, 309,  341, 373,  405, 437,
+  469, 501,  533, 565,  597, 629,  661, 693,  725, 757,  789, 821,  853, 885,
+  917, 949,  981, 1013, 22,  54,   86,  118,  150, 182,  214, 246,  278, 310,
+  342, 374,  406, 438,  470, 502,  534, 566,  598, 630,  662, 694,  726, 758,
+  790, 822,  854, 886,  918, 950,  982, 1014, 23,  55,   87,  119,  151, 183,
+  215, 247,  279, 311,  343, 375,  407, 439,  471, 503,  535, 567,  599, 631,
+  663, 695,  727, 759,  791, 823,  855, 887,  919, 951,  983, 1015, 24,  56,
+  88,  120,  152, 184,  216, 248,  280, 312,  344, 376,  408, 440,  472, 504,
+  536, 568,  600, 632,  664, 696,  728, 760,  792, 824,  856, 888,  920, 952,
+  984, 1016, 25,  57,   89,  121,  153, 185,  217, 249,  281, 313,  345, 377,
+  409, 441,  473, 505,  537, 569,  601, 633,  665, 697,  729, 761,  793, 825,
+  857, 889,  921, 953,  985, 1017, 26,  58,   90,  122,  154, 186,  218, 250,
+  282, 314,  346, 378,  410, 442,  474, 506,  538, 570,  602, 634,  666, 698,
+  730, 762,  794, 826,  858, 890,  922, 954,  986, 1018, 27,  59,   91,  123,
+  155, 187,  219, 251,  283, 315,  347, 379,  411, 443,  475, 507,  539, 571,
+  603, 635,  667, 699,  731, 763,  795, 827,  859, 891,  923, 955,  987, 1019,
+  28,  60,   92,  124,  156, 188,  220, 252,  284, 316,  348, 380,  412, 444,
+  476, 508,  540, 572,  604, 636,  668, 700,  732, 764,  796, 828,  860, 892,
+  924, 956,  988, 1020, 29,  61,   93,  125,  157, 189,  221, 253,  285, 317,
+  349, 381,  413, 445,  477, 509,  541, 573,  605, 637,  669, 701,  733, 765,
+  797, 829,  861, 893,  925, 957,  989, 1021, 30,  62,   94,  126,  158, 190,
+  222, 254,  286, 318,  350, 382,  414, 446,  478, 510,  542, 574,  606, 638,
+  670, 702,  734, 766,  798, 830,  862, 894,  926, 958,  990, 1022, 31,  63,
+  95,  127,  159, 191,  223, 255,  287, 319,  351, 383,  415, 447,  479, 511,
+  543, 575,  607, 639,  671, 703,  735, 767,  799, 831,  863, 895,  927, 959,
+  991, 1023,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, mrow_scan_32x32[1024]) = {
+  0,    1,    2,    3,    4,    5,    6,    7,    8,    9,    10,   11,   12,
+  13,   14,   15,   16,   17,   18,   19,   20,   21,   22,   23,   24,   25,
+  26,   27,   28,   29,   30,   31,   32,   33,   34,   35,   36,   37,   38,
+  39,   40,   41,   42,   43,   44,   45,   46,   47,   48,   49,   50,   51,
+  52,   53,   54,   55,   56,   57,   58,   59,   60,   61,   62,   63,   64,
+  65,   66,   67,   68,   69,   70,   71,   72,   73,   74,   75,   76,   77,
+  78,   79,   80,   81,   82,   83,   84,   85,   86,   87,   88,   89,   90,
+  91,   92,   93,   94,   95,   96,   97,   98,   99,   100,  101,  102,  103,
+  104,  105,  106,  107,  108,  109,  110,  111,  112,  113,  114,  115,  116,
+  117,  118,  119,  120,  121,  122,  123,  124,  125,  126,  127,  128,  129,
+  130,  131,  132,  133,  134,  135,  136,  137,  138,  139,  140,  141,  142,
+  143,  144,  145,  146,  147,  148,  149,  150,  151,  152,  153,  154,  155,
+  156,  157,  158,  159,  160,  161,  162,  163,  164,  165,  166,  167,  168,
+  169,  170,  171,  172,  173,  174,  175,  176,  177,  178,  179,  180,  181,
+  182,  183,  184,  185,  186,  187,  188,  189,  190,  191,  192,  193,  194,
+  195,  196,  197,  198,  199,  200,  201,  202,  203,  204,  205,  206,  207,
+  208,  209,  210,  211,  212,  213,  214,  215,  216,  217,  218,  219,  220,
+  221,  222,  223,  224,  225,  226,  227,  228,  229,  230,  231,  232,  233,
+  234,  235,  236,  237,  238,  239,  240,  241,  242,  243,  244,  245,  246,
+  247,  248,  249,  250,  251,  252,  253,  254,  255,  256,  257,  258,  259,
+  260,  261,  262,  263,  264,  265,  266,  267,  268,  269,  270,  271,  272,
+  273,  274,  275,  276,  277,  278,  279,  280,  281,  282,  283,  284,  285,
+  286,  287,  288,  289,  290,  291,  292,  293,  294,  295,  296,  297,  298,
+  299,  300,  301,  302,  303,  304,  305,  306,  307,  308,  309,  310,  311,
+  312,  313,  314,  315,  316,  317,  318,  319,  320,  321,  322,  323,  324,
+  325,  326,  327,  328,  329,  330,  331,  332,  333,  334,  335,  336,  337,
+  338,  339,  340,  341,  342,  343,  344,  345,  346,  347,  348,  349,  350,
+  351,  352,  353,  354,  355,  356,  357,  358,  359,  360,  361,  362,  363,
+  364,  365,  366,  367,  368,  369,  370,  371,  372,  373,  374,  375,  376,
+  377,  378,  379,  380,  381,  382,  383,  384,  385,  386,  387,  388,  389,
+  390,  391,  392,  393,  394,  395,  396,  397,  398,  399,  400,  401,  402,
+  403,  404,  405,  406,  407,  408,  409,  410,  411,  412,  413,  414,  415,
+  416,  417,  418,  419,  420,  421,  422,  423,  424,  425,  426,  427,  428,
+  429,  430,  431,  432,  433,  434,  435,  436,  437,  438,  439,  440,  441,
+  442,  443,  444,  445,  446,  447,  448,  449,  450,  451,  452,  453,  454,
+  455,  456,  457,  458,  459,  460,  461,  462,  463,  464,  465,  466,  467,
+  468,  469,  470,  471,  472,  473,  474,  475,  476,  477,  478,  479,  480,
+  481,  482,  483,  484,  485,  486,  487,  488,  489,  490,  491,  492,  493,
+  494,  495,  496,  497,  498,  499,  500,  501,  502,  503,  504,  505,  506,
+  507,  508,  509,  510,  511,  512,  513,  514,  515,  516,  517,  518,  519,
+  520,  521,  522,  523,  524,  525,  526,  527,  528,  529,  530,  531,  532,
+  533,  534,  535,  536,  537,  538,  539,  540,  541,  542,  543,  544,  545,
+  546,  547,  548,  549,  550,  551,  552,  553,  554,  555,  556,  557,  558,
+  559,  560,  561,  562,  563,  564,  565,  566,  567,  568,  569,  570,  571,
+  572,  573,  574,  575,  576,  577,  578,  579,  580,  581,  582,  583,  584,
+  585,  586,  587,  588,  589,  590,  591,  592,  593,  594,  595,  596,  597,
+  598,  599,  600,  601,  602,  603,  604,  605,  606,  607,  608,  609,  610,
+  611,  612,  613,  614,  615,  616,  617,  618,  619,  620,  621,  622,  623,
+  624,  625,  626,  627,  628,  629,  630,  631,  632,  633,  634,  635,  636,
+  637,  638,  639,  640,  641,  642,  643,  644,  645,  646,  647,  648,  649,
+  650,  651,  652,  653,  654,  655,  656,  657,  658,  659,  660,  661,  662,
+  663,  664,  665,  666,  667,  668,  669,  670,  671,  672,  673,  674,  675,
+  676,  677,  678,  679,  680,  681,  682,  683,  684,  685,  686,  687,  688,
+  689,  690,  691,  692,  693,  694,  695,  696,  697,  698,  699,  700,  701,
+  702,  703,  704,  705,  706,  707,  708,  709,  710,  711,  712,  713,  714,
+  715,  716,  717,  718,  719,  720,  721,  722,  723,  724,  725,  726,  727,
+  728,  729,  730,  731,  732,  733,  734,  735,  736,  737,  738,  739,  740,
+  741,  742,  743,  744,  745,  746,  747,  748,  749,  750,  751,  752,  753,
+  754,  755,  756,  757,  758,  759,  760,  761,  762,  763,  764,  765,  766,
+  767,  768,  769,  770,  771,  772,  773,  774,  775,  776,  777,  778,  779,
+  780,  781,  782,  783,  784,  785,  786,  787,  788,  789,  790,  791,  792,
+  793,  794,  795,  796,  797,  798,  799,  800,  801,  802,  803,  804,  805,
+  806,  807,  808,  809,  810,  811,  812,  813,  814,  815,  816,  817,  818,
+  819,  820,  821,  822,  823,  824,  825,  826,  827,  828,  829,  830,  831,
+  832,  833,  834,  835,  836,  837,  838,  839,  840,  841,  842,  843,  844,
+  845,  846,  847,  848,  849,  850,  851,  852,  853,  854,  855,  856,  857,
+  858,  859,  860,  861,  862,  863,  864,  865,  866,  867,  868,  869,  870,
+  871,  872,  873,  874,  875,  876,  877,  878,  879,  880,  881,  882,  883,
+  884,  885,  886,  887,  888,  889,  890,  891,  892,  893,  894,  895,  896,
+  897,  898,  899,  900,  901,  902,  903,  904,  905,  906,  907,  908,  909,
+  910,  911,  912,  913,  914,  915,  916,  917,  918,  919,  920,  921,  922,
+  923,  924,  925,  926,  927,  928,  929,  930,  931,  932,  933,  934,  935,
+  936,  937,  938,  939,  940,  941,  942,  943,  944,  945,  946,  947,  948,
+  949,  950,  951,  952,  953,  954,  955,  956,  957,  958,  959,  960,  961,
+  962,  963,  964,  965,  966,  967,  968,  969,  970,  971,  972,  973,  974,
+  975,  976,  977,  978,  979,  980,  981,  982,  983,  984,  985,  986,  987,
+  988,  989,  990,  991,  992,  993,  994,  995,  996,  997,  998,  999,  1000,
+  1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009, 1010, 1011, 1012, 1013,
+  1014, 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022, 1023,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, default_scan_32x32[1024]) = {
+  0,    32,   1,    64,  33,   2,    96,   65,   34,   128,  3,    97,   66,
+  160,  129,  35,   98,  4,    67,   130,  161,  192,  36,   99,   224,  5,
+  162,  193,  68,   131, 37,   100,  225,  194,  256,  163,  69,   132,  6,
+  226,  257,  288,  195, 101,  164,  38,   258,  7,    227,  289,  133,  320,
+  70,   196,  165,  290, 259,  228,  39,   321,  102,  352,  8,    197,  71,
+  134,  322,  291,  260, 353,  384,  229,  166,  103,  40,   354,  323,  292,
+  135,  385,  198,  261, 72,   9,    416,  167,  386,  355,  230,  324,  104,
+  293,  41,   417,  199, 136,  262,  387,  448,  325,  356,  10,   73,   418,
+  231,  168,  449,  294, 388,  105,  419,  263,  42,   200,  357,  450,  137,
+  480,  74,   326,  232, 11,   389,  169,  295,  420,  106,  451,  481,  358,
+  264,  327,  201,  43,  138,  512,  482,  390,  296,  233,  170,  421,  75,
+  452,  359,  12,   513, 265,  483,  328,  107,  202,  514,  544,  422,  391,
+  453,  139,  44,   234, 484,  297,  360,  171,  76,   515,  545,  266,  329,
+  454,  13,   423,  203, 108,  546,  485,  576,  298,  235,  140,  361,  330,
+  172,  547,  45,   455, 267,  577,  486,  77,   204,  362,  608,  14,   299,
+  578,  109,  236,  487, 609,  331,  141,  579,  46,   15,   173,  610,  363,
+  78,   205,  16,   110, 237,  611,  142,  47,   174,  79,   206,  17,   111,
+  238,  48,   143,  80,  175,  112,  207,  49,   18,   239,  81,   113,  19,
+  50,   82,   114,  51,  83,   115,  640,  516,  392,  268,  144,  20,   672,
+  641,  548,  517,  424, 393,  300,  269,  176,  145,  52,   21,   704,  673,
+  642,  580,  549,  518, 456,  425,  394,  332,  301,  270,  208,  177,  146,
+  84,   53,   22,   736, 705,  674,  643,  612,  581,  550,  519,  488,  457,
+  426,  395,  364,  333, 302,  271,  240,  209,  178,  147,  116,  85,   54,
+  23,   737,  706,  675, 613,  582,  551,  489,  458,  427,  365,  334,  303,
+  241,  210,  179,  117, 86,   55,   738,  707,  614,  583,  490,  459,  366,
+  335,  242,  211,  118, 87,   739,  615,  491,  367,  243,  119,  768,  644,
+  520,  396,  272,  148, 24,   800,  769,  676,  645,  552,  521,  428,  397,
+  304,  273,  180,  149, 56,   25,   832,  801,  770,  708,  677,  646,  584,
+  553,  522,  460,  429, 398,  336,  305,  274,  212,  181,  150,  88,   57,
+  26,   864,  833,  802, 771,  740,  709,  678,  647,  616,  585,  554,  523,
+  492,  461,  430,  399, 368,  337,  306,  275,  244,  213,  182,  151,  120,
+  89,   58,   27,   865, 834,  803,  741,  710,  679,  617,  586,  555,  493,
+  462,  431,  369,  338, 307,  245,  214,  183,  121,  90,   59,   866,  835,
+  742,  711,  618,  587, 494,  463,  370,  339,  246,  215,  122,  91,   867,
+  743,  619,  495,  371, 247,  123,  896,  772,  648,  524,  400,  276,  152,
+  28,   928,  897,  804, 773,  680,  649,  556,  525,  432,  401,  308,  277,
+  184,  153,  60,   29,  960,  929,  898,  836,  805,  774,  712,  681,  650,
+  588,  557,  526,  464, 433,  402,  340,  309,  278,  216,  185,  154,  92,
+  61,   30,   992,  961, 930,  899,  868,  837,  806,  775,  744,  713,  682,
+  651,  620,  589,  558, 527,  496,  465,  434,  403,  372,  341,  310,  279,
+  248,  217,  186,  155, 124,  93,   62,   31,   993,  962,  931,  869,  838,
+  807,  745,  714,  683, 621,  590,  559,  497,  466,  435,  373,  342,  311,
+  249,  218,  187,  125, 94,   63,   994,  963,  870,  839,  746,  715,  622,
+  591,  498,  467,  374, 343,  250,  219,  126,  95,   995,  871,  747,  623,
+  499,  375,  251,  127, 900,  776,  652,  528,  404,  280,  156,  932,  901,
+  808,  777,  684,  653, 560,  529,  436,  405,  312,  281,  188,  157,  964,
+  933,  902,  840,  809, 778,  716,  685,  654,  592,  561,  530,  468,  437,
+  406,  344,  313,  282, 220,  189,  158,  996,  965,  934,  903,  872,  841,
+  810,  779,  748,  717, 686,  655,  624,  593,  562,  531,  500,  469,  438,
+  407,  376,  345,  314, 283,  252,  221,  190,  159,  997,  966,  935,  873,
+  842,  811,  749,  718, 687,  625,  594,  563,  501,  470,  439,  377,  346,
+  315,  253,  222,  191, 998,  967,  874,  843,  750,  719,  626,  595,  502,
+  471,  378,  347,  254, 223,  999,  875,  751,  627,  503,  379,  255,  904,
+  780,  656,  532,  408, 284,  936,  905,  812,  781,  688,  657,  564,  533,
+  440,  409,  316,  285, 968,  937,  906,  844,  813,  782,  720,  689,  658,
+  596,  565,  534,  472, 441,  410,  348,  317,  286,  1000, 969,  938,  907,
+  876,  845,  814,  783, 752,  721,  690,  659,  628,  597,  566,  535,  504,
+  473,  442,  411,  380, 349,  318,  287,  1001, 970,  939,  877,  846,  815,
+  753,  722,  691,  629, 598,  567,  505,  474,  443,  381,  350,  319,  1002,
+  971,  878,  847,  754, 723,  630,  599,  506,  475,  382,  351,  1003, 879,
+  755,  631,  507,  383, 908,  784,  660,  536,  412,  940,  909,  816,  785,
+  692,  661,  568,  537, 444,  413,  972,  941,  910,  848,  817,  786,  724,
+  693,  662,  600,  569, 538,  476,  445,  414,  1004, 973,  942,  911,  880,
+  849,  818,  787,  756, 725,  694,  663,  632,  601,  570,  539,  508,  477,
+  446,  415,  1005, 974, 943,  881,  850,  819,  757,  726,  695,  633,  602,
+  571,  509,  478,  447, 1006, 975,  882,  851,  758,  727,  634,  603,  510,
+  479,  1007, 883,  759, 635,  511,  912,  788,  664,  540,  944,  913,  820,
+  789,  696,  665,  572, 541,  976,  945,  914,  852,  821,  790,  728,  697,
+  666,  604,  573,  542, 1008, 977,  946,  915,  884,  853,  822,  791,  760,
+  729,  698,  667,  636, 605,  574,  543,  1009, 978,  947,  885,  854,  823,
+  761,  730,  699,  637, 606,  575,  1010, 979,  886,  855,  762,  731,  638,
+  607,  1011, 887,  763, 639,  916,  792,  668,  948,  917,  824,  793,  700,
+  669,  980,  949,  918, 856,  825,  794,  732,  701,  670,  1012, 981,  950,
+  919,  888,  857,  826, 795,  764,  733,  702,  671,  1013, 982,  951,  889,
+  858,  827,  765,  734, 703,  1014, 983,  890,  859,  766,  735,  1015, 891,
+  767,  920,  796,  952, 921,  828,  797,  984,  953,  922,  860,  829,  798,
+  1016, 985,  954,  923, 892,  861,  830,  799,  1017, 986,  955,  893,  862,
+  831,  1018, 987,  894, 863,  1019, 895,  924,  956,  925,  988,  957,  926,
+  1020, 989,  958,  927, 1021, 990,  959,  1022, 991,  1023,
+};
+
+// Scan over two rectangular vertical partitions one after the other
+DECLARE_ALIGNED(16, static const int16_t, v2_scan_32x32[1024]) = {
+  0,    1,    32,   33,   2,    64,   34,   65,   66,   3,    96,   35,   97,
+  67,   98,   4,    128,  36,   129,  99,   68,   130,  5,    100,  131,  160,
+  37,   161,  69,   162,  132,  101,  163,  6,    192,  38,   193,  70,   194,
+  133,  164,  102,  195,  7,    224,  39,   165,  225,  134,  196,  71,   226,
+  103,  227,  166,  197,  8,    256,  40,   135,  228,  257,  72,   258,  198,
+  104,  259,  167,  229,  136,  260,  9,    288,  41,   289,  73,   199,  230,
+  290,  168,  261,  105,  291,  137,  292,  231,  10,   200,  262,  320,  42,
+  321,  74,   322,  169,  293,  106,  323,  232,  263,  138,  324,  201,  294,
+  11,   352,  43,   353,  75,   170,  325,  354,  264,  107,  233,  295,  355,
+  202,  326,  139,  356,  12,   384,  44,   265,  296,  385,  171,  357,  76,
+  386,  234,  327,  108,  387,  203,  358,  140,  388,  297,  266,  328,  13,
+  172,  389,  416,  45,   235,  359,  417,  77,   418,  109,  419,  204,  390,
+  298,  329,  141,  267,  360,  420,  236,  391,  173,  421,  14,   448,  46,
+  449,  78,   330,  450,  299,  361,  110,  205,  422,  451,  268,  392,  142,
+  452,  237,  423,  174,  331,  362,  453,  15,   300,  393,  480,  47,   481,
+  79,   482,  206,  454,  269,  424,  111,  483,  143,  484,  363,  332,  394,
+  238,  455,  175,  301,  425,  485,  512,  513,  270,  456,  514,  207,  486,
+  364,  395,  515,  333,  426,  516,  239,  487,  302,  457,  517,  396,  271,
+  488,  544,  365,  427,  545,  518,  546,  334,  458,  547,  519,  548,  303,
+  489,  397,  428,  549,  366,  459,  520,  576,  335,  490,  550,  577,  578,
+  579,  521,  429,  551,  398,  460,  580,  367,  491,  581,  552,  522,  582,
+  608,  609,  430,  461,  610,  399,  492,  553,  611,  583,  523,  612,  613,
+  584,  554,  462,  431,  493,  614,  524,  640,  641,  642,  585,  643,  555,
+  615,  644,  463,  494,  586,  525,  616,  645,  556,  646,  672,  617,  673,
+  587,  674,  647,  495,  675,  526,  676,  557,  618,  648,  677,  588,  678,
+  527,  649,  619,  704,  558,  705,  706,  679,  589,  707,  650,  708,  620,
+  680,  709,  559,  590,  710,  651,  681,  736,  621,  737,  711,  738,  739,
+  682,  652,  740,  712,  591,  741,  622,  683,  713,  742,  653,  768,  769,
+  743,  770,  714,  684,  771,  623,  772,  744,  654,  773,  715,  685,  745,
+  774,  655,  775,  800,  801,  716,  746,  802,  803,  686,  776,  804,  747,
+  805,  717,  777,  806,  687,  748,  807,  778,  832,  833,  718,  834,  835,
+  808,  836,  779,  749,  837,  809,  719,  838,  780,  750,  810,  839,  864,
+  865,  866,  867,  840,  781,  868,  811,  751,  869,  841,  870,  812,  782,
+  842,  871,  896,  897,  898,  872,  899,  813,  843,  900,  783,  901,  873,
+  844,  902,  814,  874,  903,  928,  929,  845,  930,  904,  815,  875,  931,
+  932,  905,  933,  846,  876,  934,  906,  935,  877,  960,  847,  961,  962,
+  907,  936,  963,  964,  937,  878,  965,  908,  966,  938,  967,  909,  879,
+  992,  939,  993,  968,  994,  995,  996,  910,  969,  940,  997,  998,  970,
+  911,  941,  999,  971,  1000, 942,  1001, 972,  1002, 943,  973,  1003, 974,
+  1004, 975,  1005, 1006, 1007, 16,   48,   80,   112,  144,  176,  17,   49,
+  208,  81,   113,  145,  240,  177,  272,  18,   50,   209,  82,   114,  304,
+  241,  146,  178,  273,  336,  210,  19,   51,   83,   115,  305,  242,  147,
+  368,  179,  274,  337,  211,  20,   400,  52,   84,   306,  116,  243,  369,
+  148,  338,  180,  275,  432,  401,  212,  21,   53,   307,  85,   370,  244,
+  117,  464,  149,  433,  339,  276,  181,  402,  213,  308,  496,  371,  22,
+  54,   465,  86,   245,  118,  434,  150,  340,  277,  403,  182,  528,  497,
+  214,  466,  372,  309,  23,   55,   435,  87,   246,  119,  341,  404,  151,
+  529,  560,  278,  498,  183,  467,  373,  215,  310,  436,  24,   56,   247,
+  561,  88,   530,  592,  342,  120,  405,  499,  152,  279,  468,  184,  374,
+  311,  437,  216,  562,  593,  531,  624,  25,   248,  500,  57,   406,  89,
+  343,  121,  469,  280,  153,  594,  185,  375,  563,  625,  438,  532,  656,
+  312,  217,  501,  407,  249,  26,   344,  58,   90,   470,  122,  595,  626,
+  281,  564,  657,  154,  376,  533,  688,  439,  186,  313,  502,  218,  408,
+  627,  596,  658,  250,  345,  471,  27,   59,   565,  689,  91,   123,  282,
+  534,  720,  155,  440,  377,  187,  503,  314,  628,  659,  219,  597,  690,
+  409,  472,  566,  721,  346,  251,  28,   60,   535,  752,  92,   124,  283,
+  441,  378,  156,  660,  504,  629,  691,  598,  722,  188,  315,  567,  753,
+  220,  410,  473,  347,  536,  784,  252,  29,   661,  692,  61,   93,   442,
+  630,  723,  284,  125,  379,  505,  599,  754,  157,  316,  568,  785,  189,
+  474,  411,  221,  537,  816,  693,  348,  662,  724,  253,  631,  755,  443,
+  30,   600,  786,  62,   506,  94,   285,  380,  126,  569,  817,  158,  317,
+  190,  475,  694,  725,  412,  663,  756,  538,  848,  222,  632,  787,  349,
+  254,  601,  818,  444,  507,  31,   63,   381,  286,  95,   570,  849,  726,
+  127,  695,  757,  664,  788,  159,  476,  318,  413,  539,  880,  191,  633,
+  819,  223,  350,  602,  850,  508,  255,  445,  727,  758,  696,  789,  571,
+  881,  382,  287,  665,  820,  477,  634,  851,  540,  912,  319,  414,  603,
+  882,  759,  728,  790,  351,  509,  697,  821,  446,  572,  913,  666,  852,
+  383,  635,  883,  478,  541,  944,  415,  760,  791,  604,  914,  729,  822,
+  698,  853,  510,  667,  884,  447,  573,  945,  636,  915,  792,  761,  823,
+  542,  976,  479,  730,  854,  605,  946,  699,  885,  668,  916,  511,  574,
+  977,  793,  824,  637,  947,  762,  855,  731,  886,  543,  1008, 606,  978,
+  700,  917,  669,  948,  575,  825,  1009, 794,  856,  763,  887,  638,  979,
+  732,  918,  701,  949,  607,  1010, 670,  980,  826,  857,  795,  888,  764,
+  919,  639,  1011, 733,  950,  702,  981,  858,  827,  889,  796,  920,  671,
+  1012, 765,  951,  734,  982,  703,  1013, 859,  890,  828,  921,  797,  952,
+  766,  983,  735,  1014, 891,  860,  922,  829,  953,  798,  984,  767,  1015,
+  892,  923,  861,  954,  830,  985,  799,  1016, 924,  893,  955,  862,  986,
+  831,  1017, 925,  956,  894,  987,  863,  1018, 957,  926,  988,  895,  1019,
+  958,  989,  927,  1020, 990,  959,  1021, 991,  1022, 1023,
+};
+
+// Scan over two rectangular horizontal partitions one after the other
+DECLARE_ALIGNED(16, static const int16_t, h2_scan_32x32[1024]) = {
+  0,    1,    32,   33,   2,    64,   34,   65,   66,   3,    96,   35,   97,
+  67,   98,   4,    128,  36,   129,  99,   68,   130,  5,    100,  131,  160,
+  37,   161,  69,   162,  132,  101,  163,  6,    192,  38,   193,  70,   194,
+  133,  164,  102,  195,  7,    224,  39,   165,  225,  134,  196,  71,   226,
+  103,  227,  166,  197,  8,    256,  40,   135,  228,  257,  72,   258,  198,
+  104,  259,  167,  229,  136,  260,  9,    288,  41,   289,  73,   199,  230,
+  290,  168,  261,  105,  291,  137,  292,  231,  10,   200,  262,  320,  42,
+  321,  74,   322,  169,  293,  106,  323,  232,  263,  138,  324,  201,  294,
+  11,   352,  43,   353,  75,   170,  325,  354,  264,  107,  233,  295,  355,
+  202,  326,  139,  356,  12,   384,  44,   265,  296,  385,  171,  357,  76,
+  386,  234,  327,  108,  387,  203,  358,  140,  388,  297,  266,  328,  13,
+  172,  389,  416,  45,   235,  359,  417,  77,   418,  109,  419,  204,  390,
+  298,  329,  141,  267,  360,  420,  236,  391,  173,  421,  14,   448,  46,
+  449,  78,   330,  450,  299,  361,  110,  205,  422,  451,  268,  392,  142,
+  452,  237,  423,  174,  331,  362,  453,  15,   300,  393,  480,  47,   481,
+  79,   482,  206,  454,  269,  424,  111,  483,  143,  484,  363,  332,  394,
+  238,  455,  175,  301,  425,  485,  16,   48,   80,   270,  456,  207,  486,
+  112,  364,  395,  333,  426,  144,  239,  487,  302,  457,  176,  396,  17,
+  271,  488,  49,   365,  427,  208,  81,   334,  458,  113,  145,  240,  303,
+  489,  397,  428,  177,  366,  459,  272,  18,   50,   209,  335,  490,  82,
+  114,  304,  241,  429,  146,  398,  460,  367,  491,  178,  273,  336,  210,
+  19,   51,   83,   430,  461,  399,  492,  115,  305,  242,  147,  368,  179,
+  274,  337,  462,  431,  493,  211,  20,   400,  52,   84,   306,  116,  243,
+  369,  148,  463,  494,  338,  180,  275,  432,  401,  212,  21,   53,   307,
+  85,   370,  244,  117,  495,  464,  149,  433,  339,  276,  181,  402,  213,
+  308,  496,  371,  22,   54,   465,  86,   245,  118,  434,  150,  340,  277,
+  403,  182,  497,  214,  466,  372,  309,  23,   55,   435,  87,   246,  119,
+  341,  404,  151,  278,  498,  183,  467,  373,  215,  310,  436,  24,   56,
+  247,  88,   342,  120,  405,  499,  152,  279,  468,  184,  374,  311,  437,
+  216,  25,   248,  500,  57,   406,  89,   343,  121,  469,  280,  153,  185,
+  375,  438,  312,  217,  501,  407,  249,  26,   344,  58,   90,   470,  122,
+  281,  154,  376,  439,  186,  313,  502,  218,  408,  250,  345,  471,  27,
+  59,   91,   123,  282,  155,  440,  377,  187,  503,  314,  219,  409,  472,
+  346,  251,  28,   60,   92,   124,  283,  441,  378,  156,  504,  188,  315,
+  220,  410,  473,  347,  252,  29,   61,   93,   442,  284,  125,  379,  505,
+  157,  316,  189,  474,  411,  221,  348,  253,  443,  30,   62,   506,  94,
+  285,  380,  126,  158,  317,  190,  475,  412,  222,  349,  254,  444,  507,
+  31,   63,   381,  286,  95,   127,  159,  476,  318,  413,  191,  223,  350,
+  508,  255,  445,  382,  287,  477,  319,  414,  351,  509,  446,  383,  478,
+  415,  510,  447,  479,  511,  512,  513,  514,  515,  516,  517,  544,  545,
+  518,  546,  547,  519,  548,  549,  520,  576,  550,  577,  578,  579,  521,
+  551,  580,  581,  552,  522,  582,  608,  609,  610,  553,  611,  583,  523,
+  612,  613,  584,  554,  614,  524,  640,  641,  642,  585,  643,  555,  615,
+  644,  586,  525,  616,  645,  556,  646,  672,  617,  673,  587,  674,  647,
+  675,  526,  676,  557,  618,  648,  677,  588,  678,  527,  649,  619,  704,
+  558,  705,  706,  679,  589,  707,  650,  708,  620,  680,  709,  528,  559,
+  590,  710,  651,  681,  736,  621,  737,  711,  738,  739,  682,  652,  529,
+  560,  740,  712,  591,  741,  622,  683,  713,  742,  653,  768,  769,  561,
+  743,  530,  592,  770,  714,  684,  771,  623,  772,  744,  654,  773,  715,
+  685,  745,  774,  562,  593,  531,  624,  655,  775,  800,  801,  716,  746,
+  802,  803,  686,  776,  804,  594,  563,  625,  747,  805,  717,  532,  656,
+  777,  806,  687,  748,  807,  778,  832,  833,  718,  834,  595,  626,  835,
+  564,  657,  808,  836,  533,  688,  779,  749,  837,  809,  719,  838,  780,
+  627,  596,  658,  750,  810,  839,  864,  565,  689,  865,  866,  867,  534,
+  720,  840,  781,  868,  811,  751,  869,  841,  628,  659,  597,  690,  870,
+  812,  782,  566,  721,  842,  871,  896,  535,  752,  897,  898,  872,  899,
+  813,  843,  660,  900,  783,  629,  691,  598,  722,  901,  873,  567,  753,
+  844,  902,  814,  874,  536,  784,  903,  661,  692,  928,  929,  630,  723,
+  845,  930,  904,  815,  875,  931,  599,  754,  932,  568,  785,  905,  933,
+  846,  876,  934,  537,  816,  693,  662,  724,  906,  631,  755,  935,  877,
+  600,  786,  960,  847,  961,  962,  907,  936,  963,  569,  817,  964,  937,
+  694,  725,  878,  965,  908,  663,  756,  538,  848,  966,  632,  787,  938,
+  601,  818,  967,  909,  879,  992,  939,  993,  968,  570,  849,  994,  726,
+  695,  757,  995,  664,  788,  996,  910,  969,  539,  880,  940,  633,  819,
+  997,  998,  602,  850,  970,  911,  941,  999,  727,  758,  696,  789,  571,
+  881,  971,  665,  820,  1000, 634,  851,  942,  540,  912,  1001, 972,  603,
+  882,  759,  728,  790,  1002, 697,  821,  943,  973,  572,  913,  666,  852,
+  1003, 635,  883,  974,  541,  944,  760,  791,  1004, 604,  914,  729,  822,
+  698,  853,  975,  667,  884,  573,  945,  1005, 636,  915,  792,  761,  823,
+  542,  976,  1006, 730,  854,  605,  946,  699,  885,  668,  916,  1007, 574,
+  977,  793,  824,  637,  947,  762,  855,  731,  886,  543,  1008, 606,  978,
+  700,  917,  669,  948,  575,  825,  1009, 794,  856,  763,  887,  638,  979,
+  732,  918,  701,  949,  607,  1010, 670,  980,  826,  857,  795,  888,  764,
+  919,  639,  1011, 733,  950,  702,  981,  858,  827,  889,  796,  920,  671,
+  1012, 765,  951,  734,  982,  703,  1013, 859,  890,  828,  921,  797,  952,
+  766,  983,  735,  1014, 891,  860,  922,  829,  953,  798,  984,  767,  1015,
+  892,  923,  861,  954,  830,  985,  799,  1016, 924,  893,  955,  862,  986,
+  831,  1017, 925,  956,  894,  987,  863,  1018, 957,  926,  988,  895,  1019,
+  958,  989,  927,  1020, 990,  959,  1021, 991,  1022, 1023,
+};
+
+// Scan where the top left quarter is scanned first
+DECLARE_ALIGNED(16, static const int16_t, qtr_scan_32x32[1024]) = {
+  0,    1,    32,   33,   2,    64,   34,   65,   66,   3,    96,   35,   97,
+  67,   98,   4,    128,  36,   129,  99,   68,   130,  5,    100,  131,  160,
+  37,   161,  69,   162,  132,  101,  163,  6,    192,  38,   193,  70,   194,
+  133,  164,  102,  195,  7,    224,  39,   165,  225,  134,  196,  71,   226,
+  103,  227,  166,  197,  8,    256,  40,   135,  228,  257,  72,   258,  198,
+  104,  259,  167,  229,  136,  260,  9,    288,  41,   289,  73,   199,  230,
+  290,  168,  261,  105,  291,  137,  292,  231,  10,   200,  262,  320,  42,
+  321,  74,   322,  169,  293,  106,  323,  232,  263,  138,  324,  201,  294,
+  11,   352,  43,   353,  75,   170,  325,  354,  264,  107,  233,  295,  355,
+  202,  326,  139,  356,  12,   384,  44,   265,  296,  385,  171,  357,  76,
+  386,  234,  327,  108,  387,  203,  358,  140,  388,  297,  266,  328,  13,
+  172,  389,  416,  45,   235,  359,  417,  77,   418,  109,  419,  204,  390,
+  298,  329,  141,  267,  360,  420,  236,  391,  173,  421,  14,   448,  46,
+  449,  78,   330,  450,  299,  361,  110,  205,  422,  451,  268,  392,  142,
+  452,  237,  423,  174,  331,  362,  453,  15,   300,  393,  480,  47,   481,
+  79,   482,  206,  454,  269,  424,  111,  483,  143,  484,  363,  332,  394,
+  238,  455,  175,  301,  425,  485,  270,  456,  207,  486,  364,  395,  333,
+  426,  239,  487,  302,  457,  396,  271,  488,  365,  427,  334,  458,  303,
+  489,  397,  428,  366,  459,  335,  490,  429,  398,  460,  367,  491,  430,
+  461,  399,  492,  462,  431,  493,  463,  494,  495,  16,   512,  48,   513,
+  80,   514,  112,  515,  144,  516,  176,  517,  17,   544,  49,   545,  208,
+  518,  81,   546,  113,  547,  145,  240,  519,  548,  177,  549,  272,  520,
+  18,   576,  50,   209,  550,  577,  82,   578,  114,  579,  304,  521,  241,
+  551,  146,  580,  178,  581,  273,  552,  336,  522,  210,  582,  19,   608,
+  51,   609,  83,   610,  115,  305,  553,  611,  242,  583,  147,  368,  523,
+  612,  179,  613,  274,  584,  337,  554,  211,  614,  20,   400,  524,  640,
+  52,   641,  84,   642,  306,  585,  116,  643,  243,  369,  555,  615,  148,
+  644,  338,  586,  180,  275,  432,  525,  616,  645,  401,  556,  212,  646,
+  21,   672,  53,   307,  617,  673,  85,   370,  587,  674,  244,  647,  117,
+  675,  464,  526,  149,  676,  433,  557,  339,  618,  276,  648,  181,  677,
+  402,  588,  213,  678,  308,  496,  527,  649,  371,  619,  22,   704,  54,
+  465,  558,  705,  86,   706,  245,  679,  118,  434,  589,  707,  150,  340,
+  650,  708,  277,  403,  620,  680,  182,  709,  528,  497,  559,  214,  466,
+  590,  710,  372,  651,  309,  681,  23,   736,  55,   435,  621,  737,  87,
+  246,  711,  738,  119,  739,  341,  682,  404,  652,  151,  529,  560,  740,
+  278,  712,  498,  591,  183,  741,  467,  622,  373,  683,  215,  310,  713,
+  742,  436,  653,  24,   768,  56,   769,  247,  561,  743,  88,   530,  592,
+  770,  342,  714,  120,  405,  684,  771,  499,  623,  152,  772,  279,  744,
+  468,  654,  184,  773,  374,  715,  311,  437,  685,  745,  216,  774,  562,
+  593,  531,  624,  25,   248,  500,  655,  775,  800,  57,   801,  406,  716,
+  89,   343,  746,  802,  121,  803,  469,  686,  280,  776,  153,  804,  594,
+  185,  375,  563,  625,  747,  805,  438,  717,  532,  656,  312,  777,  217,
+  806,  501,  687,  407,  748,  249,  807,  26,   344,  778,  832,  58,   833,
+  90,   470,  718,  834,  122,  595,  626,  835,  281,  564,  657,  808,  154,
+  836,  376,  533,  688,  779,  439,  749,  186,  837,  313,  809,  502,  719,
+  218,  838,  408,  780,  627,  596,  658,  250,  345,  471,  750,  810,  839,
+  27,   864,  59,   565,  689,  865,  91,   866,  123,  867,  282,  534,  720,
+  840,  155,  440,  781,  868,  377,  811,  187,  503,  751,  869,  314,  841,
+  628,  659,  219,  597,  690,  870,  409,  812,  472,  782,  566,  721,  346,
+  842,  251,  871,  28,   896,  60,   535,  752,  897,  92,   898,  124,  283,
+  872,  899,  441,  813,  378,  843,  156,  660,  900,  504,  783,  629,  691,
+  598,  722,  188,  901,  315,  873,  567,  753,  220,  410,  844,  902,  473,
+  814,  347,  874,  536,  784,  252,  903,  29,   661,  692,  928,  61,   929,
+  93,   442,  630,  723,  845,  930,  284,  904,  125,  379,  505,  815,  875,
+  931,  599,  754,  157,  932,  316,  568,  785,  905,  189,  933,  474,  846,
+  411,  876,  221,  934,  537,  816,  693,  348,  662,  724,  906,  253,  631,
+  755,  935,  443,  877,  30,   600,  786,  960,  62,   506,  847,  961,  94,
+  962,  285,  380,  907,  936,  126,  963,  569,  817,  158,  964,  317,  937,
+  190,  475,  694,  725,  878,  965,  412,  908,  663,  756,  538,  848,  222,
+  966,  632,  787,  349,  938,  254,  601,  818,  967,  444,  909,  507,  879,
+  31,   992,  63,   381,  939,  993,  286,  968,  95,   570,  849,  994,  726,
+  127,  695,  757,  995,  664,  788,  159,  996,  476,  910,  318,  969,  413,
+  539,  880,  940,  191,  633,  819,  997,  223,  998,  350,  602,  850,  970,
+  508,  911,  255,  445,  941,  999,  727,  758,  696,  789,  571,  881,  382,
+  971,  287,  665,  820,  1000, 477,  634,  851,  942,  540,  912,  319,  1001,
+  414,  972,  603,  882,  759,  728,  790,  351,  1002, 509,  697,  821,  943,
+  446,  973,  572,  913,  666,  852,  383,  1003, 635,  883,  478,  974,  541,
+  944,  415,  760,  791,  1004, 604,  914,  729,  822,  698,  853,  510,  975,
+  667,  884,  447,  573,  945,  1005, 636,  915,  792,  761,  823,  542,  976,
+  479,  1006, 730,  854,  605,  946,  699,  885,  668,  916,  511,  1007, 574,
+  977,  793,  824,  637,  947,  762,  855,  731,  886,  543,  1008, 606,  978,
+  700,  917,  669,  948,  575,  825,  1009, 794,  856,  763,  887,  638,  979,
+  732,  918,  701,  949,  607,  1010, 670,  980,  826,  857,  795,  888,  764,
+  919,  639,  1011, 733,  950,  702,  981,  858,  827,  889,  796,  920,  671,
+  1012, 765,  951,  734,  982,  703,  1013, 859,  890,  828,  921,  797,  952,
+  766,  983,  735,  1014, 891,  860,  922,  829,  953,  798,  984,  767,  1015,
+  892,  923,  861,  954,  830,  985,  799,  1016, 924,  893,  955,  862,  986,
+  831,  1017, 925,  956,  894,  987,  863,  1018, 957,  926,  988,  895,  1019,
+  958,  989,  927,  1020, 990,  959,  1021, 991,  1022, 1023,
+};
+
+#if CONFIG_TX64X64
+DECLARE_ALIGNED(16, static const int16_t, default_scan_64x64[4096]) = {
+  0,    1,    64,   65,   2,    128,  66,   129,  130,  3,    192,  67,   193,
+  131,  194,  4,    256,  68,   257,  195,  132,  258,  5,    196,  259,  320,
+  69,   321,  133,  322,  260,  197,  323,  6,    384,  70,   385,  134,  386,
+  261,  324,  198,  387,  7,    448,  71,   325,  449,  262,  388,  135,  450,
+  199,  451,  326,  389,  8,    512,  72,   263,  452,  513,  136,  514,  390,
+  200,  515,  327,  453,  264,  516,  9,    576,  73,   577,  137,  391,  454,
+  578,  328,  517,  201,  579,  265,  580,  455,  10,   392,  518,  640,  74,
+  641,  138,  642,  329,  581,  202,  643,  456,  519,  266,  644,  393,  582,
+  11,   704,  75,   705,  139,  330,  645,  706,  520,  203,  457,  583,  707,
+  394,  646,  267,  708,  12,   768,  76,   521,  584,  769,  331,  709,  140,
+  770,  458,  647,  204,  771,  395,  710,  268,  772,  585,  522,  648,  13,
+  332,  773,  832,  77,   459,  711,  833,  141,  834,  205,  835,  396,  774,
+  586,  649,  269,  523,  712,  836,  460,  775,  333,  837,  14,   896,  78,
+  897,  142,  650,  898,  587,  713,  206,  397,  838,  899,  524,  776,  270,
+  900,  461,  839,  334,  651,  714,  901,  15,   588,  777,  960,  79,   961,
+  143,  962,  398,  902,  525,  840,  207,  963,  271,  964,  715,  652,  778,
+  462,  903,  335,  589,  841,  965,  16,   1024, 80,   1025, 144,  526,  904,
+  1026, 399,  966,  208,  716,  779,  1027, 653,  842,  272,  1028, 463,  967,
+  590,  905,  336,  1029, 780,  17,   527,  968,  1088, 81,   717,  843,  1089,
+  400,  1030, 145,  1090, 654,  906,  209,  1091, 273,  464,  1031, 1092, 591,
+  969,  781,  844,  337,  1093, 718,  907,  528,  1032, 18,   1152, 82,   401,
+  655,  970,  1094, 1153, 146,  1154, 210,  1155, 592,  1033, 465,  845,  1095,
+  274,  782,  908,  1156, 719,  971,  338,  1157, 529,  1096, 656,  1034, 402,
+  1158, 19,   1216, 83,   1217, 147,  846,  909,  1218, 783,  972,  211,  593,
+  1097, 1219, 466,  1159, 275,  720,  1035, 1220, 339,  1221, 530,  1160, 657,
+  1098, 910,  847,  973,  403,  1222, 20,   784,  1036, 1280, 84,   1281, 148,
+  1282, 594,  1161, 212,  1283, 467,  721,  1099, 1223, 276,  1284, 911,  974,
+  658,  1162, 340,  531,  848,  1037, 1224, 1285, 785,  1100, 404,  1286, 21,
+  1344, 85,   595,  1225, 1345, 149,  722,  1163, 1346, 468,  1287, 213,  975,
+  1347, 912,  1038, 277,  1348, 849,  1101, 659,  1226, 532,  1288, 341,  1349,
+  786,  1164, 405,  1350, 596,  976,  1039, 1289, 723,  1227, 22,   1408, 86,
+  913,  1102, 1409, 150,  1410, 469,  1351, 214,  850,  1165, 1411, 278,  660,
+  1290, 1412, 533,  787,  1228, 1352, 342,  1413, 1040, 977,  1103, 406,  914,
+  1166, 1414, 724,  1291, 597,  1353, 23,   1472, 87,   851,  1229, 1473, 151,
+  470,  1415, 1474, 215,  1475, 661,  1354, 788,  1292, 279,  1041, 1104, 1476,
+  534,  1416, 978,  1167, 343,  1477, 915,  1230, 725,  1355, 407,  598,  1417,
+  1478, 852,  1293, 24,   1536, 88,   1537, 471,  1105, 1479, 152,  1042, 1168,
+  1538, 662,  1418, 216,  789,  1356, 1539, 979,  1231, 280,  1540, 535,  1480,
+  916,  1294, 344,  1541, 726,  1419, 599,  853,  1357, 1481, 408,  1542, 1106,
+  1169, 1043, 1232, 25,   472,  980,  1295, 1543, 1600, 89,   1601, 790,  1420,
+  153,  663,  1482, 1602, 217,  1603, 917,  1358, 536,  1544, 281,  1604, 1170,
+  345,  727,  1107, 1233, 1483, 1605, 854,  1421, 1044, 1296, 600,  1545, 409,
+  1606, 981,  1359, 791,  1484, 473,  1607, 26,   664,  1546, 1664, 90,   1665,
+  154,  918,  1422, 1666, 218,  1171, 1234, 1667, 537,  1108, 1297, 1608, 282,
+  1668, 728,  1045, 1360, 1547, 855,  1485, 346,  1669, 601,  1609, 982,  1423,
+  410,  1670, 792,  1548, 1235, 1172, 1298, 474,  665,  919,  1486, 1610, 1671,
+  27,   1728, 91,   1109, 1361, 1729, 155,  1730, 219,  1731, 538,  1046, 1424,
+  1672, 283,  856,  1549, 1732, 729,  1611, 347,  983,  1487, 1733, 602,  1673,
+  1236, 1299, 411,  1173, 1362, 1734, 793,  1612, 920,  1550, 1110, 1425, 666,
+  1674, 475,  1735, 28,   1792, 92,   1047, 1488, 1793, 156,  1794, 220,  539,
+  1736, 1795, 857,  1613, 730,  1675, 284,  1300, 1796, 984,  1551, 1237, 1363,
+  1174, 1426, 348,  1797, 603,  1737, 1111, 1489, 412,  794,  1676, 1798, 921,
+  1614, 667,  1738, 1048, 1552, 476,  1799, 29,   1301, 1364, 1856, 93,   1857,
+  157,  858,  1238, 1427, 1677, 1858, 540,  1800, 221,  731,  985,  1615, 1739,
+  1859, 1175, 1490, 285,  1860, 604,  1112, 1553, 1801, 349,  1861, 922,  1678,
+  795,  1740, 413,  1862, 1049, 1616, 1365, 668,  1302, 1428, 1802, 477,  1239,
+  1491, 1863, 859,  1741, 30,   1176, 1554, 1920, 94,   986,  1679, 1921, 158,
+  1922, 541,  732,  1803, 1864, 222,  1923, 1113, 1617, 286,  1924, 605,  1865,
+  350,  923,  1366, 1429, 1742, 1925, 796,  1804, 1303, 1492, 1050, 1680, 414,
+  1926, 1240, 1555, 669,  1866, 478,  1177, 1618, 1927, 860,  1805, 987,  1743,
+  31,   1984, 95,   733,  1867, 1985, 542,  1928, 159,  1114, 1681, 1986, 1430,
+  223,  1367, 1493, 1987, 1304, 1556, 287,  1988, 924,  1806, 606,  1929, 797,
+  1051, 1744, 1868, 351,  1241, 1619, 1989, 415,  1990, 670,  1178, 1682, 1930,
+  988,  1807, 479,  861,  1869, 1991, 1431, 1494, 1368, 1557, 1115, 1745, 734,
+  1931, 32,   2048, 96,   543,  1305, 1620, 1992, 2049, 160,  2050, 224,  2051,
+  925,  1242, 1683, 1870, 288,  1052, 1808, 2052, 607,  1993, 798,  1932, 352,
+  2053, 1179, 1746, 1495, 416,  1432, 1558, 2054, 671,  1994, 989,  1369, 1621,
+  1871, 862,  1933, 480,  1116, 1809, 2055, 1306, 1684, 735,  1995, 544,  2056,
+  33,   2112, 97,   1243, 1747, 2113, 161,  2114, 926,  1934, 1053, 1872, 225,
+  2115, 289,  608,  799,  1496, 1559, 1996, 2057, 2116, 1180, 1810, 1433, 1622,
+  353,  2117, 1370, 1685, 672,  2058, 417,  990,  1935, 2118, 1307, 1748, 863,
+  1117, 1873, 1997, 481,  2119, 736,  1244, 1811, 2059, 1560, 545,  2120, 1497,
+  1623, 34,   1054, 1936, 2176, 98,   927,  1998, 2177, 162,  1434, 1686, 2178,
+  226,  1181, 1874, 2179, 800,  2060, 609,  1371, 1749, 2121, 290,  2180, 354,
+  2181, 1308, 1812, 991,  1999, 673,  1118, 1937, 2122, 418,  2182, 864,  2061,
+  1561, 1624, 1245, 1875, 482,  1498, 1687, 2183, 737,  2123, 1435, 1750, 1055,
+  2000, 546,  928,  2062, 2184, 1182, 1938, 35,   1372, 1813, 2240, 99,   2241,
+  163,  2242, 801,  2124, 227,  2243, 610,  2185, 291,  1309, 1876, 2244, 992,
+  2063, 355,  1119, 1625, 2001, 2245, 1562, 1688, 674,  2186, 865,  1499, 1751,
+  2125, 419,  1246, 1939, 2246, 1436, 1814, 483,  2247, 738,  2187, 1056, 2064,
+  1373, 1877, 929,  1183, 2002, 2126, 547,  2248, 36,   2304, 100,  2305, 164,
+  802,  1310, 1940, 2188, 2306, 1626, 1689, 228,  1563, 1752, 2307, 611,  2249,
+  292,  2308, 1120, 1500, 1815, 2065, 993,  2127, 356,  2309, 1247, 2003, 675,
+  866,  1437, 1878, 2189, 2250, 420,  2310, 1374, 1941, 484,  1057, 2128, 2311,
+  739,  2251, 1184, 2066, 930,  1690, 2190, 1627, 1753, 548,  1564, 1816, 2312,
+  1311, 2004, 37,   803,  2252, 2368, 101,  1501, 1879, 2369, 165,  2370, 612,
+  2313, 229,  1121, 2129, 2371, 994,  2191, 1438, 1942, 293,  1248, 2067, 2372,
+  357,  867,  2253, 2373, 676,  2314, 1375, 2005, 421,  1691, 1754, 2374, 1628,
+  1817, 1058, 2192, 1185, 2130, 740,  1565, 1880, 2315, 485,  2375, 931,  2254,
+  1312, 2068, 1502, 1943, 549,  2376, 804,  2316, 38,   2432, 102,  1122, 1439,
+  2006, 2193, 2433, 166,  2434, 613,  995,  1249, 2131, 2255, 2377, 230,  2435,
+  1755, 294,  1692, 1818, 2436, 868,  1376, 2069, 2317, 1629, 1881, 358,  677,
+  2378, 2437, 1566, 1944, 422,  1186, 2194, 2438, 1059, 2256, 1313, 2132, 741,
+  1503, 2007, 2379, 932,  2318, 486,  2439, 550,  1440, 2070, 2440, 805,  1756,
+  1819, 2380, 1123, 2257, 1250, 1693, 1882, 2195, 39,   996,  2319, 2496, 103,
+  2497, 167,  614,  1630, 1945, 2441, 2498, 231,  1377, 2133, 2499, 295,  1567,
+  2008, 2500, 869,  2381, 678,  2442, 359,  2501, 1187, 2258, 1060, 2320, 1504,
+  2071, 1314, 2196, 423,  2502, 742,  933,  2382, 2443, 1820, 487,  1757, 1883,
+  2503, 1441, 2134, 1694, 1946, 551,  1124, 2321, 2504, 1251, 1631, 2009, 2259,
+  806,  2444, 997,  2383, 1378, 2197, 40,   1568, 2072, 2560, 104,  2561, 615,
+  2505, 168,  2562, 232,  2563, 870,  2445, 296,  2564, 1505, 2135, 1188, 2322,
+  679,  2506, 360,  1061, 1315, 1821, 1884, 2260, 2384, 2565, 1758, 1947, 424,
+  2566, 1695, 2010, 934,  1442, 2198, 2446, 743,  2507, 488,  1632, 2073, 2567,
+  1252, 2323, 1125, 2385, 552,  2568, 807,  1569, 2136, 2508, 1379, 2261, 998,
+  2447, 41,   616,  2569, 2624, 105,  1885, 2625, 1822, 1948, 169,  1506, 2199,
+  2626, 233,  871,  1759, 2011, 2509, 2627, 1189, 2386, 1316, 2324, 297,  2628,
+  680,  1062, 1696, 2074, 2448, 2570, 361,  2629, 1443, 2262, 1633, 2137, 425,
+  935,  2510, 2630, 744,  2571, 489,  1253, 2387, 2631, 1570, 2200, 1126, 2449,
+  1380, 2325, 1886, 1949, 808,  2572, 553,  1823, 2012, 2632, 999,  2511, 1760,
+  2075, 1507, 2263, 617,  2633, 42,   2688, 106,  1697, 2138, 2689, 170,  1190,
+  2450, 2690, 872,  1317, 2388, 2573, 234,  2691, 1063, 2512, 298,  1444, 2326,
+  2692, 681,  1634, 2201, 2634, 362,  2693, 936,  2574, 426,  1950, 2694, 1571,
+  2264, 745,  1887, 2013, 2635, 1254, 2451, 1824, 2076, 1127, 1381, 2389, 2513,
+  490,  2695, 1761, 2139, 809,  1000, 1508, 2327, 2575, 2636, 554,  2696, 1698,
+  2202, 1318, 2452, 618,  1191, 2514, 2697, 43,   2752, 107,  873,  1635, 2265,
+  2637, 2753, 171,  1445, 2390, 2754, 1064, 2576, 235,  2755, 1951, 2014, 682,
+  2698, 299,  1888, 2077, 2756, 1572, 2328, 1825, 2140, 363,  2757, 937,  2638,
+  1255, 2515, 427,  746,  1382, 1762, 2203, 2453, 2699, 2758, 1128, 2577, 491,
+  1509, 2391, 2759, 1699, 2266, 1001, 2639, 810,  2700, 555,  2760, 1319, 1636,
+  2329, 2516, 2015, 1192, 1952, 2078, 2578, 1446, 2454, 619,  1889, 2141, 2761,
+  874,  2701, 44,   2816, 108,  1065, 2640, 2817, 172,  1826, 2204, 2818, 236,
+  1573, 2392, 2819, 683,  2762, 300,  2820, 1763, 2267, 938,  2702, 364,  1256,
+  2579, 2821, 1383, 2517, 747,  1129, 2641, 2763, 428,  1700, 2330, 2822, 1510,
+  2455, 492,  2016, 2079, 2823, 1002, 1953, 2142, 2703, 811,  2764, 1637, 2393,
+  1890, 2205, 556,  1320, 2580, 2824, 1193, 1447, 2518, 2642, 1827, 2268, 620,
+  2825, 875,  2765, 1066, 1574, 2456, 2704, 45,   1764, 2331, 2880, 109,  2881,
+  173,  2882, 237,  2883, 684,  2826, 301,  1384, 2581, 2884, 1257, 2643, 939,
+  1701, 2394, 2766, 2080, 365,  1511, 2017, 2143, 2519, 2885, 1130, 2705, 1954,
+  2206, 748,  2827, 429,  2886, 1891, 2269, 1638, 2457, 493,  1003, 2767, 2887,
+  812,  1828, 2332, 2828, 1321, 2644, 1448, 2582, 1194, 2706, 557,  2888, 1575,
+  2520, 1765, 2395, 876,  1067, 2768, 2829, 621,  2889, 2081, 2144, 46,   2944,
+  110,  2018, 2207, 2945, 174,  1702, 2458, 2946, 1385, 2645, 238,  685,  1258,
+  1955, 2270, 2707, 2890, 2947, 1512, 2583, 302,  940,  2830, 2948, 1892, 2333,
+  1131, 2769, 366,  2949, 749,  1639, 2521, 2891, 430,  2950, 1829, 2396, 1004,
+  2831, 1322, 2708, 494,  1449, 2646, 2951, 813,  2892, 1195, 1766, 2459, 2770,
+  1576, 2584, 2145, 558,  2082, 2208, 2952, 2019, 2271, 1068, 2832, 877,  2893,
+  1956, 2334, 622,  1703, 2522, 2953, 1386, 2709, 47,   3008, 111,  1259, 1513,
+  1893, 2397, 2647, 2771, 3009, 175,  3010, 686,  2954, 239,  3011, 941,  2894,
+  303,  1132, 1640, 2585, 2833, 3012, 1830, 2460, 367,  3013, 750,  2955, 431,
+  2146, 2209, 3014, 1450, 2710, 1323, 2083, 2272, 2772, 1005, 1767, 2523, 2895,
+  1577, 2020, 2335, 2648, 495,  3015, 814,  1196, 2834, 2956, 1957, 2398, 559,
+  3016, 1704, 2586, 1069, 2896, 878,  1894, 2461, 2957, 623,  1387, 2773, 3017,
+  1514, 2711, 1260, 2835, 48,   3072, 112,  1831, 2524, 3073, 1641, 2649, 176,
+  3074, 687,  3018, 942,  2210, 2958, 240,  3075, 1133, 2147, 2273, 2897, 304,
+  2084, 2336, 3076, 368,  1768, 2587, 3077, 751,  2021, 2399, 3019, 1451, 2774,
+  1324, 2836, 432,  1578, 2712, 3078, 1006, 2959, 1958, 2462, 1197, 2898, 496,
+  815,  3020, 3079, 1705, 2650, 1895, 2525, 560,  3080, 1070, 2960, 1388, 2837,
+  879,  1515, 2775, 3021, 2211, 2274, 1832, 2588, 624,  2148, 2337, 3081, 1261,
+  2899, 1642, 2713, 2085, 2400, 49,   3136, 113,  3137, 688,  3082, 177,  943,
+  1134, 2022, 2463, 2961, 3022, 3138, 241,  1769, 2651, 3139, 305,  3140, 1452,
+  2838, 1959, 2526, 752,  1325, 1579, 2776, 2900, 3083, 369,  3141, 1007, 3023,
+  433,  3142, 1198, 1706, 2714, 2962, 1896, 2589, 816,  3084, 497,  2275, 3143,
+  2212, 2338, 2149, 2401, 561,  1071, 1516, 1833, 2652, 2839, 3024, 3144, 1389,
+  2901, 2086, 2464, 880,  3085, 1643, 2777, 1262, 2963, 625,  2023, 2527, 3145,
+  1770, 2715, 1135, 3025, 50,   944,  1960, 2590, 3086, 3200, 114,  689,  3146,
+  3201, 178,  3202, 242,  1453, 2902, 3203, 1580, 2840, 306,  1326, 2964, 3204,
+  2276, 2339, 753,  1897, 2653, 3147, 370,  1707, 2213, 2402, 2778, 3205, 1008,
+  3087, 1199, 2150, 2465, 3026, 434,  3206, 817,  2087, 2528, 3148, 1834, 2716,
+  498,  3207, 1517, 2903, 1390, 2965, 1072, 3088, 1644, 2024, 2591, 2841, 562,
+  3208, 881,  1263, 3027, 3149, 1771, 2779, 626,  1961, 2654, 3209, 2340, 1136,
+  3089, 2277, 2403, 945,  3150, 690,  1454, 2214, 2466, 2966, 3210, 51,   1581,
+  2904, 3264, 115,  3265, 179,  1898, 2717, 3266, 1327, 3028, 243,  2151, 2529,
+  3267, 1708, 2842, 307,  3268, 754,  3211, 2088, 2592, 371,  1009, 3151, 3269,
+  1200, 3090, 1835, 2780, 435,  3270, 2025, 2655, 818,  3212, 1518, 2967, 499,
+  1391, 1645, 2905, 3029, 3271, 1073, 3152, 1962, 2718, 563,  1264, 1772, 2341,
+  2404, 2843, 3091, 3272, 882,  2278, 2467, 3213, 2215, 2530, 627,  3273, 2152,
+  2593, 1137, 1899, 2781, 3153, 1582, 2968, 1455, 3030, 946,  3214, 691,  1709,
+  2906, 3274, 52,   1328, 3092, 3328, 116,  2089, 2656, 3329, 180,  3330, 244,
+  3331, 308,  1836, 2844, 3332, 755,  3275, 1010, 1201, 2026, 2719, 3154, 3215,
+  372,  3333, 1519, 2405, 3031, 436,  2342, 2468, 3334, 1646, 2969, 819,  1392,
+  3093, 3276, 2279, 2531, 1963, 2782, 500,  3335, 1773, 2907, 1074, 2216, 2594,
+  3216, 1265, 3155, 564,  3336, 883,  2153, 2657, 3277, 1900, 2845, 628,  1583,
+  3032, 3337, 1456, 2090, 2720, 3094, 1138, 3217, 1710, 2970, 947,  3278, 1329,
+  3156, 692,  3338, 53,   1837, 2908, 3392, 117,  2027, 2783, 3393, 181,  2406,
+  2469, 3394, 2343, 2532, 245,  3395, 1202, 3218, 309,  756,  2280, 2595, 3339,
+  3396, 1011, 3279, 1520, 3095, 373,  1647, 3033, 3397, 1964, 2846, 2217, 2658,
+  1393, 3157, 437,  1774, 2971, 3398, 820,  3340, 2154, 2721, 1075, 3280, 501,
+  3399, 1266, 3219, 1901, 2909, 565,  884,  2091, 2784, 3341, 3400, 1584, 3096,
+  1457, 1711, 3034, 3158, 2470, 629,  1139, 2407, 2533, 3281, 3401, 2344, 2596,
+  2028, 2847, 948,  1330, 1838, 2972, 3220, 3342, 2281, 2659, 693,  3402, 54,
+  3456, 118,  3457, 182,  2218, 2722, 3458, 246,  1203, 1965, 2910, 3282, 3459,
+  1012, 1648, 3097, 3343, 757,  1521, 3159, 3403, 310,  3460, 1775, 2155, 2785,
+  3035, 374,  1394, 3221, 3461, 438,  3462, 821,  3404, 1902, 2973, 1076, 2092,
+  2848, 3344, 1267, 3283, 502,  2471, 2534, 3463, 2408, 2597, 1585, 2345, 2660,
+  3160, 885,  3405, 566,  1712, 3098, 3464, 1458, 3222, 2029, 2911, 2282, 2723,
+  1140, 1839, 3036, 3345, 630,  3465, 1331, 3284, 949,  2219, 2786, 3406, 694,
+  1966, 2974, 3466, 55,   2156, 2849, 3520, 119,  1649, 3161, 3521, 1204, 3346,
+  183,  1522, 3223, 3522, 1776, 3099, 247,  1013, 3407, 3523, 758,  3467, 311,
+  3524, 1395, 2535, 3285, 2472, 2598, 2093, 2912, 375,  1903, 2409, 2661, 3037,
+  3525, 822,  2346, 2724, 3468, 439,  3526, 1077, 1268, 3347, 3408, 503,  2283,
+  2787, 3527, 1586, 3224, 1713, 2030, 2975, 3162, 886,  1459, 3286, 3469, 1840,
+  3100, 567,  3528, 2220, 2850, 1141, 3409, 1332, 3348, 631,  3529, 1967, 3038,
+  950,  3470, 2157, 2913, 2536, 2599, 695,  1650, 2473, 2662, 3225, 3530, 1523,
+  1777, 3163, 3287, 1205, 2410, 2725, 3410, 56,   3584, 120,  3585, 184,  2094,
+  2976, 3586, 1014, 3471, 248,  1396, 1904, 2347, 2788, 3101, 3349, 3587, 759,
+  3531, 312,  3588, 376,  2284, 2851, 3589, 823,  3532, 1269, 2031, 3039, 3411,
+  440,  1078, 3472, 3590, 1714, 3226, 1587, 3288, 2221, 2914, 504,  1841, 3164,
+  3591, 1460, 3350, 887,  3533, 568,  2600, 3592, 2537, 2663, 1968, 3102, 1142,
+  2158, 2977, 3473, 2474, 2726, 1333, 3412, 632,  3593, 2411, 2789, 951,  3534,
+  1651, 3289, 1778, 3227, 2348, 2852, 1524, 2095, 3040, 3351, 696,  3594, 1206,
+  3474, 1905, 3165, 57,   3648, 121,  1015, 1397, 2285, 2915, 3413, 3535, 3649,
+  185,  3650, 760,  3595, 249,  3651, 313,  2032, 3103, 3652, 2222, 2978, 377,
+  3653, 1270, 1715, 3290, 3475, 824,  1588, 3352, 3596, 1079, 2601, 2664, 3536,
+  1842, 3228, 441,  2538, 2727, 3654, 1461, 2475, 2790, 3414, 505,  2159, 3041,
+  3655, 1969, 3166, 888,  2412, 2853, 3597, 569,  3656, 1143, 3537, 1334, 3476,
+  2349, 2916, 2096, 3104, 1652, 3353, 633,  1779, 3291, 3657, 952,  3598, 1525,
+  3415, 1906, 2286, 2979, 3229, 697,  1207, 3538, 3658, 1398, 3477, 1016, 3599,
+  2033, 2665, 3167, 58,   2602, 2728, 3712, 122,  2223, 3042, 3713, 186,  3714,
+  761,  2539, 2791, 3659, 250,  3715, 314,  1716, 2476, 2854, 3354, 3716, 1589,
+  1843, 3292, 3416, 1271, 3539, 378,  3717, 1080, 3600, 825,  2160, 3105, 3660,
+  2413, 2917, 442,  1462, 1970, 3230, 3478, 3718, 2350, 2980, 506,  3719, 889,
+  3661, 1144, 1335, 2097, 3168, 3540, 3601, 570,  3720, 1780, 3355, 1653, 2287,
+  3043, 3417, 1907, 3293, 634,  953,  1526, 2666, 2729, 3479, 3662, 3721, 2603,
+  2792, 2540, 2855, 1208, 2224, 3106, 3602, 2034, 3231, 698,  3722, 1399, 3541,
+  2477, 2918, 1017, 3663, 59,   3776, 123,  3777, 187,  762,  1717, 2414, 2981,
+  3418, 3723, 3778, 1844, 3356, 251,  2161, 3169, 3779, 1590, 3480, 315,  1272,
+  3603, 3780, 1971, 3294, 1081, 2351, 3044, 3664, 379,  3781, 826,  3724, 1463,
+  3542, 443,  3782, 2098, 3232, 2730, 2288, 3107, 507,  2667, 2793, 3783, 890,
+  3725, 1336, 2604, 2856, 3604, 1145, 1781, 3419, 3665, 1654, 3481, 571,  1908,
+  3357, 3784, 2541, 2919, 1527, 3543, 2225, 3170, 954,  2478, 2982, 3726, 635,
+  2035, 3295, 3785, 1209, 3666, 1400, 3605, 2415, 3045, 699,  3786, 1018, 2162,
+  3233, 3727, 1718, 3482, 1845, 3420, 60,   2352, 3108, 3840, 124,  1591, 3544,
+  3841, 763,  3787, 188,  1972, 3358, 3842, 252,  3843, 1273, 3667, 2731, 2794,
+  316,  3844, 2668, 2857, 1082, 1464, 3606, 3728, 380,  827,  2099, 2605, 2920,
+  3296, 3788, 3845, 2289, 3171, 444,  3846, 2542, 2983, 1782, 3483, 508,  1337,
+  3668, 3847, 891,  1655, 1909, 3421, 3545, 3789, 1146, 2479, 3046, 3729, 2226,
+  3234, 572,  3848, 1528, 2036, 3359, 3607, 2416, 3109, 955,  3790, 636,  3849,
+  1210, 3730, 1401, 2163, 3297, 3669, 2353, 3172, 2795, 700,  1846, 2732, 2858,
+  3484, 3850, 1719, 3546, 1019, 2669, 2921, 3791, 1973, 3422, 1592, 3608, 2606,
+  2984, 61,   764,  3851, 3904, 125,  3905, 189,  1274, 2290, 3235, 3731, 3906,
+  2100, 3360, 253,  2543, 3047, 3907, 1465, 3670, 317,  1083, 3792, 3908, 828,
+  3852, 381,  3909, 2480, 3110, 1783, 3547, 445,  1910, 2227, 3298, 3485, 3910,
+  1656, 3609, 1338, 3732, 892,  3853, 509,  1147, 2037, 2417, 3173, 3423, 3793,
+  3911, 1529, 3671, 573,  2796, 2859, 3912, 2733, 2922, 2164, 3361, 956,  2354,
+  3236, 3854, 2670, 2985, 637,  3913, 1211, 1402, 3733, 3794, 1847, 2607, 3048,
+  3548, 1720, 3610, 1974, 3486, 701,  3914, 1020, 1593, 2544, 3111, 3672, 3855,
+  2291, 3299, 2101, 3424, 765,  1275, 3795, 3915, 62,   3968, 126,  2481, 3174,
+  3969, 190,  1466, 3734, 3970, 254,  3971, 1084, 3856, 318,  2228, 3362, 3972,
+  829,  1784, 3611, 3916, 1911, 3549, 382,  2418, 3237, 3973, 2860, 1657, 2797,
+  2923, 3673, 2038, 3487, 446,  2734, 2986, 3974, 1339, 3796, 1148, 3857, 893,
+  2671, 3049, 3917, 510,  1530, 3735, 3975, 2355, 3300, 2165, 3425, 2608, 3112,
+  574,  3976, 957,  3918, 1848, 3612, 1403, 2545, 3175, 3797, 1212, 3858, 638,
+  1721, 1975, 3550, 3674, 3977, 2292, 3363, 1594, 2102, 3488, 3736, 702,  2482,
+  3238, 3978, 1021, 3919, 1276, 2861, 2924, 3859, 766,  1467, 2229, 2798, 2987,
+  3426, 3798, 3979, 63,   4032, 127,  2419, 3301, 4033, 191,  2735, 3050, 4034,
+  1085, 1912, 3613, 3920, 255,  1785, 3675, 4035, 319,  2672, 3113, 4036, 2039,
+  3551, 830,  3980, 1658, 3737, 383,  4037, 1340, 2356, 3364, 3860, 2609, 3176,
+  447,  2166, 3489, 4038, 1149, 1531, 3799, 3921, 894,  3981, 511,  4039, 2546,
+  3239, 575,  1849, 3676, 4040, 2293, 3427, 1976, 3614, 958,  1722, 3738, 3982,
+  1404, 3861, 1213, 2483, 3302, 3922, 2103, 3552, 639,  2925, 4041, 2862, 2988,
+  1595, 3800, 2799, 3051, 2736, 3114, 703,  1022, 3983, 4042, 2230, 3490, 2420,
+  3365, 1277, 2673, 3177, 3923, 1468, 3862, 767,  1913, 3677, 4043, 1786, 3739,
+  2040, 3615, 1086, 2610, 3240, 3984, 2357, 3428, 1659, 3801, 831,  4044, 2167,
+  3553, 1341, 3924, 2547, 3303, 1532, 3863, 1150, 3985, 895,  4045, 2294, 2926,
+  2989, 3491, 2863, 3052, 1850, 2484, 3366, 3740, 1977, 3678, 2800, 3115, 1723,
+  3802, 2104, 3616, 1405, 3925, 959,  2737, 3178, 4046, 1214, 3986, 1596, 3864,
+  2421, 3429, 2231, 2674, 3241, 3554, 1023, 4047, 2611, 3304, 1278, 1469, 1914,
+  3741, 3926, 3987, 1787, 2041, 3679, 3803, 2358, 3492, 1087, 1660, 2168, 2548,
+  3367, 3617, 3865, 4048, 2990, 2927, 3053, 2864, 3116, 1342, 3988, 1533, 2295,
+  2801, 3179, 3555, 3927, 2485, 3430, 1151, 4049, 1978, 2738, 3242, 3742, 1851,
+  3804, 2105, 3680, 1724, 3866, 2675, 3305, 1406, 2422, 3493, 3989, 2232, 3618,
+  1215, 4050, 1597, 3928, 2612, 3368, 2359, 3556, 1915, 3805, 2042, 2991, 3054,
+  3743, 1470, 3990, 1788, 2928, 3117, 3867, 1279, 2549, 3431, 4051, 2865, 3180,
+  2169, 3681, 1661, 3929, 2802, 3243, 2486, 3494, 2296, 3619, 2739, 3306, 1343,
+  4052, 1534, 3991, 1979, 3806, 1852, 3868, 2676, 3369, 2106, 3744, 2423, 3557,
+  1725, 3930, 2233, 3682, 2613, 3432, 1407, 4053, 3055, 1598, 2992, 3118, 3992,
+  2929, 3181, 2360, 3620, 2866, 3244, 2550, 3495, 1916, 3869, 2043, 3807, 1789,
+  2803, 3307, 3931, 1471, 2170, 3745, 4054, 2740, 3370, 1662, 2487, 3558, 3993,
+  2297, 3683, 2677, 3433, 1535, 4055, 1980, 3870, 1853, 2107, 2424, 3621, 3808,
+  3932, 3056, 3119, 2614, 3496, 2993, 3182, 1726, 2234, 3746, 3994, 2930, 3245,
+  2867, 3308, 1599, 2361, 3684, 4056, 2551, 3559, 2804, 3371, 2044, 3871, 1917,
+  3933, 2171, 3809, 1790, 2741, 3434, 3995, 2488, 3622, 2298, 3747, 1663, 4057,
+  2678, 3497, 3120, 3057, 3183, 2994, 3246, 2425, 3685, 1981, 3934, 2108, 3872,
+  2615, 3560, 2931, 3309, 1854, 3996, 2235, 3810, 2868, 3372, 1727, 4058, 2552,
+  3623, 2805, 3435, 2362, 3748, 2742, 3498, 2045, 3935, 1918, 3997, 2172, 3873,
+  2489, 3686, 1791, 4059, 3121, 3184, 2299, 2679, 3561, 3811, 3058, 3247, 2995,
+  3310, 2932, 3373, 2426, 3749, 2616, 3624, 1982, 3998, 2109, 2869, 3436, 3936,
+  1855, 4060, 2236, 3874, 2806, 3499, 2553, 3687, 2363, 3812, 2743, 3562, 3185,
+  3122, 3248, 2046, 3999, 2490, 3750, 1919, 2173, 3059, 3311, 3937, 4061, 2680,
+  3625, 2996, 3374, 2300, 3875, 2933, 3437, 2617, 3688, 2427, 3813, 2870, 3500,
+  2110, 4000, 1983, 4062, 2807, 3563, 2237, 3938, 2554, 3751, 2364, 3876, 2744,
+  3626, 3186, 3249, 3123, 3312, 3060, 3375, 2491, 2997, 3438, 3814, 2047, 2681,
+  3689, 4063, 2174, 4001, 2934, 3501, 2301, 3939, 2871, 3564, 2618, 3752, 2428,
+  3877, 2808, 3627, 2111, 4064, 2238, 3250, 4002, 2555, 3187, 3313, 3815, 3124,
+  3376, 2745, 3690, 2365, 3940, 3061, 3439, 2998, 3502, 2492, 3878, 2682, 3753,
+  2935, 3565, 2175, 4065, 2302, 4003, 2872, 3628, 2619, 3816, 2429, 3941, 2809,
+  3691, 3251, 3314, 3188, 3377, 3125, 3440, 2556, 3879, 2239, 3062, 3503, 4066,
+  2746, 3754, 2366, 4004, 2999, 3566, 2936, 3629, 2683, 3817, 2493, 3942, 2873,
+  3692, 2303, 4067, 2620, 3880, 3315, 3252, 3378, 3189, 3441, 2430, 2810, 3755,
+  4005, 3126, 3504, 3063, 3567, 2557, 3943, 2747, 3818, 3000, 3630, 2367, 4068,
+  2937, 3693, 2684, 3881, 2494, 4006, 2874, 3756, 3316, 3379, 3253, 3442, 3190,
+  3505, 2621, 3944, 3127, 3568, 2811, 3819, 2431, 4069, 3064, 3631, 2748, 3882,
+  2558, 3001, 3694, 4007, 2938, 3757, 2685, 3945, 3380, 3317, 3443, 2495, 4070,
+  3254, 3506, 2875, 3820, 3191, 3569, 3128, 3632, 2622, 4008, 2812, 3883, 3065,
+  3695, 3002, 3758, 2749, 3946, 2559, 4071, 2939, 3821, 3381, 3444, 3318, 3507,
+  2686, 3255, 3570, 4009, 2876, 3884, 3192, 3633, 3129, 3696, 2623, 4072, 2813,
+  3947, 3066, 3759, 3003, 3822, 2750, 4010, 3445, 3382, 3508, 2940, 3885, 3319,
+  3571, 3256, 3634, 2687, 3193, 3697, 4073, 2877, 3948, 3130, 3760, 3067, 3823,
+  2814, 4011, 3004, 3886, 3446, 3509, 3383, 3572, 2751, 4074, 3320, 3635, 2941,
+  3949, 3257, 3698, 3194, 3761, 2878, 4012, 3131, 3824, 3068, 3887, 2815, 4075,
+  3510, 3447, 3573, 3005, 3950, 3384, 3636, 3321, 3699, 3258, 3762, 2942, 4013,
+  3195, 3825, 3132, 3888, 2879, 4076, 3069, 3951, 3511, 3574, 3448, 3637, 3385,
+  3700, 3006, 4014, 3322, 3763, 3259, 3826, 2943, 4077, 3196, 3889, 3133, 3952,
+  3575, 3512, 3638, 3070, 4015, 3449, 3701, 3386, 3764, 3323, 3827, 3007, 4078,
+  3260, 3890, 3197, 3953, 3134, 4016, 3576, 3639, 3513, 3702, 3450, 3765, 3071,
+  4079, 3387, 3828, 3324, 3891, 3261, 3954, 3198, 4017, 3640, 3135, 4080, 3577,
+  3703, 3514, 3766, 3451, 3829, 3388, 3892, 3325, 3955, 3262, 4018, 3199, 4081,
+  3641, 3704, 3578, 3767, 3515, 3830, 3452, 3893, 3389, 3956, 3326, 4019, 3263,
+  4082, 3705, 3642, 3768, 3579, 3831, 3516, 3894, 3453, 3957, 3390, 4020, 3327,
+  4083, 3706, 3769, 3643, 3832, 3580, 3895, 3517, 3958, 3454, 4021, 3391, 4084,
+  3770, 3707, 3833, 3644, 3896, 3581, 3959, 3518, 4022, 3455, 4085, 3771, 3834,
+  3708, 3897, 3645, 3960, 3582, 4023, 3519, 4086, 3835, 3772, 3898, 3709, 3961,
+  3646, 4024, 3583, 4087, 3836, 3899, 3773, 3962, 3710, 4025, 3647, 4088, 3900,
+  3837, 3963, 3774, 4026, 3711, 4089, 3901, 3964, 3838, 4027, 3775, 4090, 3965,
+  3902, 4028, 3839, 4091, 3966, 4029, 3903, 4092, 4030, 3967, 4093, 4031, 4094,
+  4095,
+};
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_CB4X4
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_2x2_neighbors[5 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 1, 1, 2, 0, 0,
+};
+#endif
+
+// Neighborhood 2-tuples for various scans and blocksizes,
+// in {top, left} order for each position in corresponding scan order.
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 4,  0, 1, 4, 4, 5,  5,  1, 8,  8,  5,  8, 2,
+  2, 2, 5, 9, 12, 6, 9, 3, 6, 10, 13, 7, 10, 11, 14, 0, 0,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 4, 4,  8,  8, 0, 0, 1, 4, 5,  8,  9,  12, 1,
+  1, 2, 5, 6, 9, 10, 13, 2, 2, 3, 6, 7, 10, 11, 14, 0,  0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 1, 1, 2,  2, 0, 0, 1,  4,  2,  5,  3,  6, 4,
+  4, 5, 8, 6, 9, 7, 10, 8, 8, 9, 12, 10, 13, 11, 14, 0, 0,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t,
+                col_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0,  0, 0, 0, 4, 4, 4, 0, 8, 8,  1,  4, 5,  8,  5,  1, 9,
+  12, 2, 5, 6, 9, 6, 2, 3, 6, 10, 13, 7, 10, 11, 14, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                row_scan_4x4_neighbors[17 * MAX_NEIGHBORS]) = {
+  0, 0, 0, 0, 0, 1, 1, 1, 1,  4, 2,  2,  2,  5,  4,  5, 5,
+  8, 3, 6, 8, 9, 6, 9, 9, 12, 7, 10, 10, 13, 11, 14, 0, 0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_4x8_neighbors[33 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  0,  0,  1,  4,  1,  1,  4,  4,  2,  5,  5,  8,  6,
+  9,  2,  2,  8,  8,  3,  6,  9,  12, 7,  10, 10, 13, 12, 12, 13, 16,
+  11, 14, 14, 17, 15, 18, 16, 16, 17, 20, 18, 21, 19, 22, 20, 20, 21,
+  24, 22, 25, 23, 26, 24, 24, 25, 28, 26, 29, 27, 30, 0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_4x8_neighbors[33 * MAX_NEIGHBORS]) = {
+  0, 0, 0,  0,  4,  4,  8,  8,  12, 12, 16, 16, 20, 20, 24, 24, 0,
+  0, 1, 4,  5,  8,  9,  12, 13, 16, 17, 20, 21, 24, 25, 28, 1,  1,
+  2, 5, 6,  9,  10, 13, 14, 17, 18, 21, 22, 25, 26, 29, 2,  2,  3,
+  6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_4x8_neighbors[33 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  1,  1,  2,  2,  0,  0,  1,  4,  2,  5,  3,  6,  4,
+  4,  5,  8,  6,  9,  7,  10, 8,  8,  9,  12, 10, 13, 11, 14, 12, 12,
+  13, 16, 14, 17, 15, 18, 16, 16, 17, 20, 18, 21, 19, 22, 20, 20, 21,
+  24, 22, 25, 23, 26, 24, 24, 25, 28, 26, 29, 27, 30, 0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_8x4_neighbors[33 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  0,  0, 1,  8,  1,  1,  8,  8,  2,  9,  9, 16, 10,
+  17, 2,  2,  16, 16, 3, 10, 17, 24, 11, 18, 18, 25, 3,  3, 4,  11,
+  19, 26, 12, 19, 4,  4, 20, 27, 5,  12, 13, 20, 21, 28, 5, 5,  6,
+  13, 14, 21, 22, 29, 6, 6,  7,  14, 15, 22, 23, 30, 0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_8x4_neighbors[33 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  8,  8,  16, 16, 0,  0,  1,  8,  9,  16, 17, 24, 1,
+  1,  2,  9,  10, 17, 18, 25, 2,  2,  3,  10, 11, 18, 19, 26, 3,  3,
+  4,  11, 12, 19, 20, 27, 4,  4,  5,  12, 13, 20, 21, 28, 5,  5,  6,
+  13, 14, 21, 22, 29, 6,  6,  7,  14, 15, 22, 23, 30, 0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_8x4_neighbors[33 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  1,  1,  2,  2,  3,  3,  4,  4,  5,  5,  6,  6,  0,
+  0,  1,  8,  2,  9,  3,  10, 4,  11, 5,  12, 6,  13, 7,  14, 8,  8,
+  9,  16, 10, 17, 11, 18, 12, 19, 13, 20, 14, 21, 15, 22, 16, 16, 17,
+  24, 18, 25, 19, 26, 20, 27, 21, 28, 22, 29, 23, 30, 0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_4x16_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  0,  0,  1,  1,  1,  4,  4,  4,  2,  2,  2,  5,  5,  8,  8,
+  8,  3,  6,  6,  9,  9,  12, 12, 12, 7,  10, 10, 13, 13, 16, 16, 16, 11, 14,
+  14, 17, 17, 20, 20, 20, 15, 18, 18, 21, 21, 24, 24, 24, 19, 22, 22, 25, 25,
+  28, 28, 28, 23, 26, 26, 29, 29, 32, 32, 32, 27, 30, 30, 33, 33, 36, 36, 36,
+  31, 34, 34, 37, 37, 40, 40, 40, 35, 38, 38, 41, 41, 44, 44, 44, 39, 42, 42,
+  45, 45, 48, 48, 48, 43, 46, 46, 49, 49, 52, 52, 52, 47, 50, 50, 53, 53, 56,
+  56, 56, 51, 54, 54, 57, 57, 60, 55, 58, 58, 61, 59, 62, 0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_16x4_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  0,  0,  1,  1,  1,  16, 16, 16, 2,  2,  2,  17, 17, 32, 32,
+  32, 3,  3,  3,  18, 18, 33, 33, 48, 4,  4,  4,  19, 19, 34, 34, 49, 5,  5,
+  5,  20, 20, 35, 35, 50, 6,  6,  6,  21, 21, 36, 36, 51, 7,  7,  7,  22, 22,
+  37, 37, 52, 8,  8,  8,  23, 23, 38, 38, 53, 9,  9,  9,  24, 24, 39, 39, 54,
+  10, 10, 10, 25, 25, 40, 40, 55, 11, 11, 11, 26, 26, 41, 41, 56, 12, 12, 12,
+  27, 27, 42, 42, 57, 13, 13, 13, 28, 28, 43, 43, 58, 14, 14, 14, 29, 29, 44,
+  44, 59, 15, 30, 30, 45, 45, 60, 31, 46, 46, 61, 47, 62, 0,  0
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_4x16_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  0,  0,  1,  1,  1,  16, 16, 16, 2,  2,  2,  17, 17, 32, 32,
+  32, 3,  3,  3,  18, 18, 33, 33, 48, 4,  4,  4,  19, 19, 34, 34, 49, 5,  5,
+  5,  20, 20, 35, 35, 50, 6,  6,  6,  21, 21, 36, 36, 51, 7,  7,  7,  22, 22,
+  37, 37, 52, 8,  8,  8,  23, 23, 38, 38, 53, 9,  9,  9,  24, 24, 39, 39, 54,
+  10, 10, 10, 25, 25, 40, 40, 55, 11, 11, 11, 26, 26, 41, 41, 56, 12, 12, 12,
+  27, 27, 42, 42, 57, 13, 13, 13, 28, 28, 43, 43, 58, 14, 14, 14, 29, 29, 44,
+  44, 59, 15, 30, 30, 45, 45, 60, 31, 46, 46, 61, 47, 62, 0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_16x4_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  1,  1,  2,  2,  3,  3,  4,  4,  5,  5,  6,  6,  7,  7,  8,
+  8,  9,  9,  10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 0,  0,  1,  16, 2,  17,
+  3,  18, 4,  19, 5,  20, 6,  21, 7,  22, 8,  23, 9,  24, 10, 25, 11, 26, 12,
+  27, 13, 28, 14, 29, 15, 30, 16, 16, 17, 32, 18, 33, 19, 34, 20, 35, 21, 36,
+  22, 37, 23, 38, 24, 39, 25, 40, 26, 41, 27, 42, 28, 43, 29, 44, 30, 45, 31,
+  46, 32, 32, 33, 48, 34, 49, 35, 50, 36, 51, 37, 52, 38, 53, 39, 54, 40, 55,
+  41, 56, 42, 57, 43, 58, 44, 59, 45, 60, 46, 61, 47, 62, 0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_4x16_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  4,  4,  8,  8,  12, 12, 16, 16, 20, 20, 24, 24, 28, 28, 32,
+  32, 36, 36, 40, 40, 44, 44, 48, 48, 52, 52, 56, 56, 0,  0,  1,  4,  5,  8,
+  9,  12, 13, 16, 17, 20, 21, 24, 25, 28, 29, 32, 33, 36, 37, 40, 41, 44, 45,
+  48, 49, 52, 53, 56, 57, 60, 1,  1,  2,  5,  6,  9,  10, 13, 14, 17, 18, 21,
+  22, 25, 26, 29, 30, 33, 34, 37, 38, 41, 42, 45, 46, 49, 50, 53, 54, 57, 58,
+  61, 2,  2,  3,  6,  7,  10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31, 34,
+  35, 38, 39, 42, 43, 46, 47, 50, 51, 54, 55, 58, 59, 62, 0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_16x4_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  16, 16, 32, 32, 0,  0,  1,  16, 17, 32, 33, 48, 1,  1,  2,
+  17, 18, 33, 34, 49, 2,  2,  3,  18, 19, 34, 35, 50, 3,  3,  4,  19, 20, 35,
+  36, 51, 4,  4,  5,  20, 21, 36, 37, 52, 5,  5,  6,  21, 22, 37, 38, 53, 6,
+  6,  7,  22, 23, 38, 39, 54, 7,  7,  8,  23, 24, 39, 40, 55, 8,  8,  9,  24,
+  25, 40, 41, 56, 9,  9,  10, 25, 26, 41, 42, 57, 10, 10, 11, 26, 27, 42, 43,
+  58, 11, 11, 12, 27, 28, 43, 44, 59, 12, 12, 13, 28, 29, 44, 45, 60, 13, 13,
+  14, 29, 30, 45, 46, 61, 14, 14, 15, 30, 31, 46, 47, 62, 0,  0
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_8x32_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   0,   0,   1,   1,   1,   8,   8,   8,   2,   2,   2,
+  9,   9,   16,  16,  16,  3,   3,   3,   10,  10,  17,  17,  24,  24,  24,
+  4,   4,   4,   11,  11,  18,  18,  25,  25,  32,  32,  32,  5,   5,   5,
+  12,  12,  19,  19,  26,  26,  33,  33,  40,  40,  40,  6,   6,   6,   13,
+  13,  20,  20,  27,  27,  34,  34,  41,  41,  48,  48,  48,  7,   14,  14,
+  21,  21,  28,  28,  35,  35,  42,  42,  49,  49,  56,  56,  56,  15,  22,
+  22,  29,  29,  36,  36,  43,  43,  50,  50,  57,  57,  64,  64,  64,  23,
+  30,  30,  37,  37,  44,  44,  51,  51,  58,  58,  65,  65,  72,  72,  72,
+  31,  38,  38,  45,  45,  52,  52,  59,  59,  66,  66,  73,  73,  80,  80,
+  80,  39,  46,  46,  53,  53,  60,  60,  67,  67,  74,  74,  81,  81,  88,
+  88,  88,  47,  54,  54,  61,  61,  68,  68,  75,  75,  82,  82,  89,  89,
+  96,  96,  96,  55,  62,  62,  69,  69,  76,  76,  83,  83,  90,  90,  97,
+  97,  104, 104, 104, 63,  70,  70,  77,  77,  84,  84,  91,  91,  98,  98,
+  105, 105, 112, 112, 112, 71,  78,  78,  85,  85,  92,  92,  99,  99,  106,
+  106, 113, 113, 120, 120, 120, 79,  86,  86,  93,  93,  100, 100, 107, 107,
+  114, 114, 121, 121, 128, 128, 128, 87,  94,  94,  101, 101, 108, 108, 115,
+  115, 122, 122, 129, 129, 136, 136, 136, 95,  102, 102, 109, 109, 116, 116,
+  123, 123, 130, 130, 137, 137, 144, 144, 144, 103, 110, 110, 117, 117, 124,
+  124, 131, 131, 138, 138, 145, 145, 152, 152, 152, 111, 118, 118, 125, 125,
+  132, 132, 139, 139, 146, 146, 153, 153, 160, 160, 160, 119, 126, 126, 133,
+  133, 140, 140, 147, 147, 154, 154, 161, 161, 168, 168, 168, 127, 134, 134,
+  141, 141, 148, 148, 155, 155, 162, 162, 169, 169, 176, 176, 176, 135, 142,
+  142, 149, 149, 156, 156, 163, 163, 170, 170, 177, 177, 184, 184, 184, 143,
+  150, 150, 157, 157, 164, 164, 171, 171, 178, 178, 185, 185, 192, 192, 192,
+  151, 158, 158, 165, 165, 172, 172, 179, 179, 186, 186, 193, 193, 200, 200,
+  200, 159, 166, 166, 173, 173, 180, 180, 187, 187, 194, 194, 201, 201, 208,
+  208, 208, 167, 174, 174, 181, 181, 188, 188, 195, 195, 202, 202, 209, 209,
+  216, 216, 216, 175, 182, 182, 189, 189, 196, 196, 203, 203, 210, 210, 217,
+  217, 224, 224, 224, 183, 190, 190, 197, 197, 204, 204, 211, 211, 218, 218,
+  225, 225, 232, 232, 232, 191, 198, 198, 205, 205, 212, 212, 219, 219, 226,
+  226, 233, 233, 240, 240, 240, 199, 206, 206, 213, 213, 220, 220, 227, 227,
+  234, 234, 241, 241, 248, 207, 214, 214, 221, 221, 228, 228, 235, 235, 242,
+  242, 249, 215, 222, 222, 229, 229, 236, 236, 243, 243, 250, 223, 230, 230,
+  237, 237, 244, 244, 251, 231, 238, 238, 245, 245, 252, 239, 246, 246, 253,
+  247, 254, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_32x8_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   0,   0,   1,   1,   1,   32,  32,  32,  2,   2,   2,
+  33,  33,  64,  64,  64,  3,   3,   3,   34,  34,  65,  65,  96,  96,  96,
+  4,   4,   4,   35,  35,  66,  66,  97,  97,  128, 128, 128, 5,   5,   5,
+  36,  36,  67,  67,  98,  98,  129, 129, 160, 160, 160, 6,   6,   6,   37,
+  37,  68,  68,  99,  99,  130, 130, 161, 161, 192, 192, 192, 7,   7,   7,
+  38,  38,  69,  69,  100, 100, 131, 131, 162, 162, 193, 193, 224, 8,   8,
+  8,   39,  39,  70,  70,  101, 101, 132, 132, 163, 163, 194, 194, 225, 9,
+  9,   9,   40,  40,  71,  71,  102, 102, 133, 133, 164, 164, 195, 195, 226,
+  10,  10,  10,  41,  41,  72,  72,  103, 103, 134, 134, 165, 165, 196, 196,
+  227, 11,  11,  11,  42,  42,  73,  73,  104, 104, 135, 135, 166, 166, 197,
+  197, 228, 12,  12,  12,  43,  43,  74,  74,  105, 105, 136, 136, 167, 167,
+  198, 198, 229, 13,  13,  13,  44,  44,  75,  75,  106, 106, 137, 137, 168,
+  168, 199, 199, 230, 14,  14,  14,  45,  45,  76,  76,  107, 107, 138, 138,
+  169, 169, 200, 200, 231, 15,  15,  15,  46,  46,  77,  77,  108, 108, 139,
+  139, 170, 170, 201, 201, 232, 16,  16,  16,  47,  47,  78,  78,  109, 109,
+  140, 140, 171, 171, 202, 202, 233, 17,  17,  17,  48,  48,  79,  79,  110,
+  110, 141, 141, 172, 172, 203, 203, 234, 18,  18,  18,  49,  49,  80,  80,
+  111, 111, 142, 142, 173, 173, 204, 204, 235, 19,  19,  19,  50,  50,  81,
+  81,  112, 112, 143, 143, 174, 174, 205, 205, 236, 20,  20,  20,  51,  51,
+  82,  82,  113, 113, 144, 144, 175, 175, 206, 206, 237, 21,  21,  21,  52,
+  52,  83,  83,  114, 114, 145, 145, 176, 176, 207, 207, 238, 22,  22,  22,
+  53,  53,  84,  84,  115, 115, 146, 146, 177, 177, 208, 208, 239, 23,  23,
+  23,  54,  54,  85,  85,  116, 116, 147, 147, 178, 178, 209, 209, 240, 24,
+  24,  24,  55,  55,  86,  86,  117, 117, 148, 148, 179, 179, 210, 210, 241,
+  25,  25,  25,  56,  56,  87,  87,  118, 118, 149, 149, 180, 180, 211, 211,
+  242, 26,  26,  26,  57,  57,  88,  88,  119, 119, 150, 150, 181, 181, 212,
+  212, 243, 27,  27,  27,  58,  58,  89,  89,  120, 120, 151, 151, 182, 182,
+  213, 213, 244, 28,  28,  28,  59,  59,  90,  90,  121, 121, 152, 152, 183,
+  183, 214, 214, 245, 29,  29,  29,  60,  60,  91,  91,  122, 122, 153, 153,
+  184, 184, 215, 215, 246, 30,  30,  30,  61,  61,  92,  92,  123, 123, 154,
+  154, 185, 185, 216, 216, 247, 31,  62,  62,  93,  93,  124, 124, 155, 155,
+  186, 186, 217, 217, 248, 63,  94,  94,  125, 125, 156, 156, 187, 187, 218,
+  218, 249, 95,  126, 126, 157, 157, 188, 188, 219, 219, 250, 127, 158, 158,
+  189, 189, 220, 220, 251, 159, 190, 190, 221, 221, 252, 191, 222, 222, 253,
+  223, 254, 0,   0
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_8x32_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   1,   1,   2,   2,   3,   3,   4,   4,   5,   5,   6,
+  6,   0,   0,   1,   8,   2,   9,   3,   10,  4,   11,  5,   12,  6,   13,
+  7,   14,  8,   8,   9,   16,  10,  17,  11,  18,  12,  19,  13,  20,  14,
+  21,  15,  22,  16,  16,  17,  24,  18,  25,  19,  26,  20,  27,  21,  28,
+  22,  29,  23,  30,  24,  24,  25,  32,  26,  33,  27,  34,  28,  35,  29,
+  36,  30,  37,  31,  38,  32,  32,  33,  40,  34,  41,  35,  42,  36,  43,
+  37,  44,  38,  45,  39,  46,  40,  40,  41,  48,  42,  49,  43,  50,  44,
+  51,  45,  52,  46,  53,  47,  54,  48,  48,  49,  56,  50,  57,  51,  58,
+  52,  59,  53,  60,  54,  61,  55,  62,  56,  56,  57,  64,  58,  65,  59,
+  66,  60,  67,  61,  68,  62,  69,  63,  70,  64,  64,  65,  72,  66,  73,
+  67,  74,  68,  75,  69,  76,  70,  77,  71,  78,  72,  72,  73,  80,  74,
+  81,  75,  82,  76,  83,  77,  84,  78,  85,  79,  86,  80,  80,  81,  88,
+  82,  89,  83,  90,  84,  91,  85,  92,  86,  93,  87,  94,  88,  88,  89,
+  96,  90,  97,  91,  98,  92,  99,  93,  100, 94,  101, 95,  102, 96,  96,
+  97,  104, 98,  105, 99,  106, 100, 107, 101, 108, 102, 109, 103, 110, 104,
+  104, 105, 112, 106, 113, 107, 114, 108, 115, 109, 116, 110, 117, 111, 118,
+  112, 112, 113, 120, 114, 121, 115, 122, 116, 123, 117, 124, 118, 125, 119,
+  126, 120, 120, 121, 128, 122, 129, 123, 130, 124, 131, 125, 132, 126, 133,
+  127, 134, 128, 128, 129, 136, 130, 137, 131, 138, 132, 139, 133, 140, 134,
+  141, 135, 142, 136, 136, 137, 144, 138, 145, 139, 146, 140, 147, 141, 148,
+  142, 149, 143, 150, 144, 144, 145, 152, 146, 153, 147, 154, 148, 155, 149,
+  156, 150, 157, 151, 158, 152, 152, 153, 160, 154, 161, 155, 162, 156, 163,
+  157, 164, 158, 165, 159, 166, 160, 160, 161, 168, 162, 169, 163, 170, 164,
+  171, 165, 172, 166, 173, 167, 174, 168, 168, 169, 176, 170, 177, 171, 178,
+  172, 179, 173, 180, 174, 181, 175, 182, 176, 176, 177, 184, 178, 185, 179,
+  186, 180, 187, 181, 188, 182, 189, 183, 190, 184, 184, 185, 192, 186, 193,
+  187, 194, 188, 195, 189, 196, 190, 197, 191, 198, 192, 192, 193, 200, 194,
+  201, 195, 202, 196, 203, 197, 204, 198, 205, 199, 206, 200, 200, 201, 208,
+  202, 209, 203, 210, 204, 211, 205, 212, 206, 213, 207, 214, 208, 208, 209,
+  216, 210, 217, 211, 218, 212, 219, 213, 220, 214, 221, 215, 222, 216, 216,
+  217, 224, 218, 225, 219, 226, 220, 227, 221, 228, 222, 229, 223, 230, 224,
+  224, 225, 232, 226, 233, 227, 234, 228, 235, 229, 236, 230, 237, 231, 238,
+  232, 232, 233, 240, 234, 241, 235, 242, 236, 243, 237, 244, 238, 245, 239,
+  246, 240, 240, 241, 248, 242, 249, 243, 250, 244, 251, 245, 252, 246, 253,
+  247, 254, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_32x8_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   1,   1,   2,   2,   3,   3,   4,   4,   5,   5,   6,
+  6,   7,   7,   8,   8,   9,   9,   10,  10,  11,  11,  12,  12,  13,  13,
+  14,  14,  15,  15,  16,  16,  17,  17,  18,  18,  19,  19,  20,  20,  21,
+  21,  22,  22,  23,  23,  24,  24,  25,  25,  26,  26,  27,  27,  28,  28,
+  29,  29,  30,  30,  0,   0,   1,   32,  2,   33,  3,   34,  4,   35,  5,
+  36,  6,   37,  7,   38,  8,   39,  9,   40,  10,  41,  11,  42,  12,  43,
+  13,  44,  14,  45,  15,  46,  16,  47,  17,  48,  18,  49,  19,  50,  20,
+  51,  21,  52,  22,  53,  23,  54,  24,  55,  25,  56,  26,  57,  27,  58,
+  28,  59,  29,  60,  30,  61,  31,  62,  32,  32,  33,  64,  34,  65,  35,
+  66,  36,  67,  37,  68,  38,  69,  39,  70,  40,  71,  41,  72,  42,  73,
+  43,  74,  44,  75,  45,  76,  46,  77,  47,  78,  48,  79,  49,  80,  50,
+  81,  51,  82,  52,  83,  53,  84,  54,  85,  55,  86,  56,  87,  57,  88,
+  58,  89,  59,  90,  60,  91,  61,  92,  62,  93,  63,  94,  64,  64,  65,
+  96,  66,  97,  67,  98,  68,  99,  69,  100, 70,  101, 71,  102, 72,  103,
+  73,  104, 74,  105, 75,  106, 76,  107, 77,  108, 78,  109, 79,  110, 80,
+  111, 81,  112, 82,  113, 83,  114, 84,  115, 85,  116, 86,  117, 87,  118,
+  88,  119, 89,  120, 90,  121, 91,  122, 92,  123, 93,  124, 94,  125, 95,
+  126, 96,  96,  97,  128, 98,  129, 99,  130, 100, 131, 101, 132, 102, 133,
+  103, 134, 104, 135, 105, 136, 106, 137, 107, 138, 108, 139, 109, 140, 110,
+  141, 111, 142, 112, 143, 113, 144, 114, 145, 115, 146, 116, 147, 117, 148,
+  118, 149, 119, 150, 120, 151, 121, 152, 122, 153, 123, 154, 124, 155, 125,
+  156, 126, 157, 127, 158, 128, 128, 129, 160, 130, 161, 131, 162, 132, 163,
+  133, 164, 134, 165, 135, 166, 136, 167, 137, 168, 138, 169, 139, 170, 140,
+  171, 141, 172, 142, 173, 143, 174, 144, 175, 145, 176, 146, 177, 147, 178,
+  148, 179, 149, 180, 150, 181, 151, 182, 152, 183, 153, 184, 154, 185, 155,
+  186, 156, 187, 157, 188, 158, 189, 159, 190, 160, 160, 161, 192, 162, 193,
+  163, 194, 164, 195, 165, 196, 166, 197, 167, 198, 168, 199, 169, 200, 170,
+  201, 171, 202, 172, 203, 173, 204, 174, 205, 175, 206, 176, 207, 177, 208,
+  178, 209, 179, 210, 180, 211, 181, 212, 182, 213, 183, 214, 184, 215, 185,
+  216, 186, 217, 187, 218, 188, 219, 189, 220, 190, 221, 191, 222, 192, 192,
+  193, 224, 194, 225, 195, 226, 196, 227, 197, 228, 198, 229, 199, 230, 200,
+  231, 201, 232, 202, 233, 203, 234, 204, 235, 205, 236, 206, 237, 207, 238,
+  208, 239, 209, 240, 210, 241, 211, 242, 212, 243, 213, 244, 214, 245, 215,
+  246, 216, 247, 217, 248, 218, 249, 219, 250, 220, 251, 221, 252, 222, 253,
+  223, 254, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_8x32_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   8,   8,   16,  16,  24,  24,  32,  32,  40,  40,  48,
+  48,  56,  56,  64,  64,  72,  72,  80,  80,  88,  88,  96,  96,  104, 104,
+  112, 112, 120, 120, 128, 128, 136, 136, 144, 144, 152, 152, 160, 160, 168,
+  168, 176, 176, 184, 184, 192, 192, 200, 200, 208, 208, 216, 216, 224, 224,
+  232, 232, 240, 240, 0,   0,   1,   8,   9,   16,  17,  24,  25,  32,  33,
+  40,  41,  48,  49,  56,  57,  64,  65,  72,  73,  80,  81,  88,  89,  96,
+  97,  104, 105, 112, 113, 120, 121, 128, 129, 136, 137, 144, 145, 152, 153,
+  160, 161, 168, 169, 176, 177, 184, 185, 192, 193, 200, 201, 208, 209, 216,
+  217, 224, 225, 232, 233, 240, 241, 248, 1,   1,   2,   9,   10,  17,  18,
+  25,  26,  33,  34,  41,  42,  49,  50,  57,  58,  65,  66,  73,  74,  81,
+  82,  89,  90,  97,  98,  105, 106, 113, 114, 121, 122, 129, 130, 137, 138,
+  145, 146, 153, 154, 161, 162, 169, 170, 177, 178, 185, 186, 193, 194, 201,
+  202, 209, 210, 217, 218, 225, 226, 233, 234, 241, 242, 249, 2,   2,   3,
+  10,  11,  18,  19,  26,  27,  34,  35,  42,  43,  50,  51,  58,  59,  66,
+  67,  74,  75,  82,  83,  90,  91,  98,  99,  106, 107, 114, 115, 122, 123,
+  130, 131, 138, 139, 146, 147, 154, 155, 162, 163, 170, 171, 178, 179, 186,
+  187, 194, 195, 202, 203, 210, 211, 218, 219, 226, 227, 234, 235, 242, 243,
+  250, 3,   3,   4,   11,  12,  19,  20,  27,  28,  35,  36,  43,  44,  51,
+  52,  59,  60,  67,  68,  75,  76,  83,  84,  91,  92,  99,  100, 107, 108,
+  115, 116, 123, 124, 131, 132, 139, 140, 147, 148, 155, 156, 163, 164, 171,
+  172, 179, 180, 187, 188, 195, 196, 203, 204, 211, 212, 219, 220, 227, 228,
+  235, 236, 243, 244, 251, 4,   4,   5,   12,  13,  20,  21,  28,  29,  36,
+  37,  44,  45,  52,  53,  60,  61,  68,  69,  76,  77,  84,  85,  92,  93,
+  100, 101, 108, 109, 116, 117, 124, 125, 132, 133, 140, 141, 148, 149, 156,
+  157, 164, 165, 172, 173, 180, 181, 188, 189, 196, 197, 204, 205, 212, 213,
+  220, 221, 228, 229, 236, 237, 244, 245, 252, 5,   5,   6,   13,  14,  21,
+  22,  29,  30,  37,  38,  45,  46,  53,  54,  61,  62,  69,  70,  77,  78,
+  85,  86,  93,  94,  101, 102, 109, 110, 117, 118, 125, 126, 133, 134, 141,
+  142, 149, 150, 157, 158, 165, 166, 173, 174, 181, 182, 189, 190, 197, 198,
+  205, 206, 213, 214, 221, 222, 229, 230, 237, 238, 245, 246, 253, 6,   6,
+  7,   14,  15,  22,  23,  30,  31,  38,  39,  46,  47,  54,  55,  62,  63,
+  70,  71,  78,  79,  86,  87,  94,  95,  102, 103, 110, 111, 118, 119, 126,
+  127, 134, 135, 142, 143, 150, 151, 158, 159, 166, 167, 174, 175, 182, 183,
+  190, 191, 198, 199, 206, 207, 214, 215, 222, 223, 230, 231, 238, 239, 246,
+  247, 254, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_32x8_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  32, 32, 64, 64,  96,  96,  128, 128, 160, 160, 192, 192,
+  0,  0,  1,  32, 33, 64, 65, 96,  97,  128, 129, 160, 161, 192, 193, 224,
+  1,  1,  2,  33, 34, 65, 66, 97,  98,  129, 130, 161, 162, 193, 194, 225,
+  2,  2,  3,  34, 35, 66, 67, 98,  99,  130, 131, 162, 163, 194, 195, 226,
+  3,  3,  4,  35, 36, 67, 68, 99,  100, 131, 132, 163, 164, 195, 196, 227,
+  4,  4,  5,  36, 37, 68, 69, 100, 101, 132, 133, 164, 165, 196, 197, 228,
+  5,  5,  6,  37, 38, 69, 70, 101, 102, 133, 134, 165, 166, 197, 198, 229,
+  6,  6,  7,  38, 39, 70, 71, 102, 103, 134, 135, 166, 167, 198, 199, 230,
+  7,  7,  8,  39, 40, 71, 72, 103, 104, 135, 136, 167, 168, 199, 200, 231,
+  8,  8,  9,  40, 41, 72, 73, 104, 105, 136, 137, 168, 169, 200, 201, 232,
+  9,  9,  10, 41, 42, 73, 74, 105, 106, 137, 138, 169, 170, 201, 202, 233,
+  10, 10, 11, 42, 43, 74, 75, 106, 107, 138, 139, 170, 171, 202, 203, 234,
+  11, 11, 12, 43, 44, 75, 76, 107, 108, 139, 140, 171, 172, 203, 204, 235,
+  12, 12, 13, 44, 45, 76, 77, 108, 109, 140, 141, 172, 173, 204, 205, 236,
+  13, 13, 14, 45, 46, 77, 78, 109, 110, 141, 142, 173, 174, 205, 206, 237,
+  14, 14, 15, 46, 47, 78, 79, 110, 111, 142, 143, 174, 175, 206, 207, 238,
+  15, 15, 16, 47, 48, 79, 80, 111, 112, 143, 144, 175, 176, 207, 208, 239,
+  16, 16, 17, 48, 49, 80, 81, 112, 113, 144, 145, 176, 177, 208, 209, 240,
+  17, 17, 18, 49, 50, 81, 82, 113, 114, 145, 146, 177, 178, 209, 210, 241,
+  18, 18, 19, 50, 51, 82, 83, 114, 115, 146, 147, 178, 179, 210, 211, 242,
+  19, 19, 20, 51, 52, 83, 84, 115, 116, 147, 148, 179, 180, 211, 212, 243,
+  20, 20, 21, 52, 53, 84, 85, 116, 117, 148, 149, 180, 181, 212, 213, 244,
+  21, 21, 22, 53, 54, 85, 86, 117, 118, 149, 150, 181, 182, 213, 214, 245,
+  22, 22, 23, 54, 55, 86, 87, 118, 119, 150, 151, 182, 183, 214, 215, 246,
+  23, 23, 24, 55, 56, 87, 88, 119, 120, 151, 152, 183, 184, 215, 216, 247,
+  24, 24, 25, 56, 57, 88, 89, 120, 121, 152, 153, 184, 185, 216, 217, 248,
+  25, 25, 26, 57, 58, 89, 90, 121, 122, 153, 154, 185, 186, 217, 218, 249,
+  26, 26, 27, 58, 59, 90, 91, 122, 123, 154, 155, 186, 187, 218, 219, 250,
+  27, 27, 28, 59, 60, 91, 92, 123, 124, 155, 156, 187, 188, 219, 220, 251,
+  28, 28, 29, 60, 61, 92, 93, 124, 125, 156, 157, 188, 189, 220, 221, 252,
+  29, 29, 30, 61, 62, 93, 94, 125, 126, 157, 158, 189, 190, 221, 222, 253,
+  30, 30, 31, 62, 63, 94, 95, 126, 127, 158, 159, 190, 191, 222, 223, 254,
+  0,  0
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t,
+                col_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  8,  8,  8,  0,  16, 16, 1,  8,  24, 24, 9,  16, 9,  1,  32,
+  32, 17, 24, 2,  9,  25, 32, 10, 17, 40, 40, 10, 2,  18, 25, 33, 40, 3,  10,
+  48, 48, 11, 18, 26, 33, 11, 3,  41, 48, 19, 26, 34, 41, 4,  11, 27, 34, 12,
+  19, 49, 56, 42, 49, 20, 27, 12, 4,  35, 42, 5,  12, 28, 35, 50, 57, 43, 50,
+  13, 20, 36, 43, 13, 5,  21, 28, 51, 58, 29, 36, 6,  13, 44, 51, 14, 21, 14,
+  6,  37, 44, 52, 59, 22, 29, 7,  14, 30, 37, 45, 52, 15, 22, 38, 45, 23, 30,
+  53, 60, 31, 38, 46, 53, 39, 46, 54, 61, 47, 54, 55, 62, 0,  0,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  8,  8,  16, 16, 24, 24, 32, 32, 40, 40, 48, 48, 0,  0,  1,
+  8,  9,  16, 17, 24, 25, 32, 33, 40, 41, 48, 49, 56, 1,  1,  2,  9,  10, 17,
+  18, 25, 26, 33, 34, 41, 42, 49, 50, 57, 2,  2,  3,  10, 11, 18, 19, 26, 27,
+  34, 35, 42, 43, 50, 51, 58, 3,  3,  4,  11, 12, 19, 20, 27, 28, 35, 36, 43,
+  44, 51, 52, 59, 4,  4,  5,  12, 13, 20, 21, 28, 29, 36, 37, 44, 45, 52, 53,
+  60, 5,  5,  6,  13, 14, 21, 22, 29, 30, 37, 38, 45, 46, 53, 54, 61, 6,  6,
+  7,  14, 15, 22, 23, 30, 31, 38, 39, 46, 47, 54, 55, 62, 0,  0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  1,  1,  2,  2,  3,  3,  4,  4,  5,  5,  6,  6,  0,  0,  1,
+  8,  2,  9,  3,  10, 4,  11, 5,  12, 6,  13, 7,  14, 8,  8,  9,  16, 10, 17,
+  11, 18, 12, 19, 13, 20, 14, 21, 15, 22, 16, 16, 17, 24, 18, 25, 19, 26, 20,
+  27, 21, 28, 22, 29, 23, 30, 24, 24, 25, 32, 26, 33, 27, 34, 28, 35, 29, 36,
+  30, 37, 31, 38, 32, 32, 33, 40, 34, 41, 35, 42, 36, 43, 37, 44, 38, 45, 39,
+  46, 40, 40, 41, 48, 42, 49, 43, 50, 44, 51, 45, 52, 46, 53, 47, 54, 48, 48,
+  49, 56, 50, 57, 51, 58, 52, 59, 53, 60, 54, 61, 55, 62, 0,  0,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t,
+                row_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  1,  1,  0,  1,  1,  8,  2,  2,  8,  9,  2,  9,  3,  3,  9,
+  16, 3,  10, 16, 17, 4,  4,  10, 17, 17, 24, 4,  11, 11, 18, 18, 25, 24, 25,
+  5,  5,  5,  12, 12, 19, 25, 32, 19, 26, 6,  6,  26, 33, 32, 33, 13, 20, 20,
+  27, 33, 40, 6,  13, 27, 34, 40, 41, 34, 41, 21, 28, 28, 35, 41, 48, 14, 21,
+  35, 42, 7,  14, 48, 49, 29, 36, 42, 49, 36, 43, 22, 29, 49, 56, 15, 22, 43,
+  50, 50, 57, 37, 44, 30, 37, 44, 51, 23, 30, 51, 58, 45, 52, 38, 45, 52, 59,
+  31, 38, 53, 60, 39, 46, 46, 53, 47, 54, 54, 61, 55, 62, 0,  0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_8x8_neighbors[65 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  8,  0,  8,  8,  1,  8,  9,  1,  9,  16, 16, 17, 2,  9,  10,
+  2,  10, 17, 17, 24, 24, 25, 3,  10, 11, 3,  18, 25, 25, 32, 11, 18, 32, 33,
+  4,  11, 26, 33, 19, 26, 12, 4,  33, 40, 12, 19, 40, 41, 5,  12, 27, 34, 34,
+  41, 20, 27, 13, 20, 13, 5,  41, 48, 48, 49, 28, 35, 35, 42, 21, 28, 6,  6,
+  6,  13, 42, 49, 49, 56, 36, 43, 14, 21, 29, 36, 7,  14, 43, 50, 50, 57, 22,
+  29, 37, 44, 15, 22, 44, 51, 51, 58, 30, 37, 23, 30, 52, 59, 45, 52, 38, 45,
+  31, 38, 53, 60, 46, 53, 39, 46, 54, 61, 47, 54, 55, 62, 0,  0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_8x16_neighbors[129 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   0,   0,   1,   1,   1,   8,   8,   8,   2,   2,   2,
+  9,   9,   16,  16,  16,  3,   3,   3,   10,  10,  17,  17,  24,  24,  24,
+  4,   4,   4,   11,  11,  18,  18,  25,  25,  32,  32,  32,  5,   5,   5,
+  12,  12,  19,  19,  26,  26,  33,  33,  40,  40,  40,  6,   6,   6,   13,
+  13,  20,  20,  27,  27,  34,  34,  41,  41,  48,  48,  48,  7,   14,  14,
+  21,  21,  28,  28,  35,  35,  42,  42,  49,  49,  56,  56,  56,  15,  22,
+  22,  29,  29,  36,  36,  43,  43,  50,  50,  57,  57,  64,  64,  64,  23,
+  30,  30,  37,  37,  44,  44,  51,  51,  58,  58,  65,  65,  72,  72,  72,
+  31,  38,  38,  45,  45,  52,  52,  59,  59,  66,  66,  73,  73,  80,  80,
+  80,  39,  46,  46,  53,  53,  60,  60,  67,  67,  74,  74,  81,  81,  88,
+  88,  88,  47,  54,  54,  61,  61,  68,  68,  75,  75,  82,  82,  89,  89,
+  96,  96,  96,  55,  62,  62,  69,  69,  76,  76,  83,  83,  90,  90,  97,
+  97,  104, 104, 104, 63,  70,  70,  77,  77,  84,  84,  91,  91,  98,  98,
+  105, 105, 112, 112, 112, 71,  78,  78,  85,  85,  92,  92,  99,  99,  106,
+  106, 113, 113, 120, 79,  86,  86,  93,  93,  100, 100, 107, 107, 114, 114,
+  121, 87,  94,  94,  101, 101, 108, 108, 115, 115, 122, 95,  102, 102, 109,
+  109, 116, 116, 123, 103, 110, 110, 117, 117, 124, 111, 118, 118, 125, 119,
+  126, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_16x8_neighbors[129 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   0,   0,  1,  1,   1,   16,  16,  16,  2,   2,   2,
+  17,  17,  32,  32,  32,  3,  3,  3,   18,  18,  33,  33,  48,  48,  48,
+  4,   4,   4,   19,  19,  34, 34, 49,  49,  64,  64,  64,  5,   5,   5,
+  20,  20,  35,  35,  50,  50, 65, 65,  80,  80,  80,  6,   6,   6,   21,
+  21,  36,  36,  51,  51,  66, 66, 81,  81,  96,  96,  96,  7,   7,   7,
+  22,  22,  37,  37,  52,  52, 67, 67,  82,  82,  97,  97,  112, 8,   8,
+  8,   23,  23,  38,  38,  53, 53, 68,  68,  83,  83,  98,  98,  113, 9,
+  9,   9,   24,  24,  39,  39, 54, 54,  69,  69,  84,  84,  99,  99,  114,
+  10,  10,  10,  25,  25,  40, 40, 55,  55,  70,  70,  85,  85,  100, 100,
+  115, 11,  11,  11,  26,  26, 41, 41,  56,  56,  71,  71,  86,  86,  101,
+  101, 116, 12,  12,  12,  27, 27, 42,  42,  57,  57,  72,  72,  87,  87,
+  102, 102, 117, 13,  13,  13, 28, 28,  43,  43,  58,  58,  73,  73,  88,
+  88,  103, 103, 118, 14,  14, 14, 29,  29,  44,  44,  59,  59,  74,  74,
+  89,  89,  104, 104, 119, 15, 30, 30,  45,  45,  60,  60,  75,  75,  90,
+  90,  105, 105, 120, 31,  46, 46, 61,  61,  76,  76,  91,  91,  106, 106,
+  121, 47,  62,  62,  77,  77, 92, 92,  107, 107, 122, 63,  78,  78,  93,
+  93,  108, 108, 123, 79,  94, 94, 109, 109, 124, 95,  110, 110, 125, 111,
+  126, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_8x16_neighbors[129 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  8,  8,  16, 16, 24, 24,  32,  32,  40,  40,  48,  48,
+  56, 56, 64, 64, 72, 72, 80, 80, 88, 88,  96,  96,  104, 104, 112, 112,
+  0,  0,  1,  8,  9,  16, 17, 24, 25, 32,  33,  40,  41,  48,  49,  56,
+  57, 64, 65, 72, 73, 80, 81, 88, 89, 96,  97,  104, 105, 112, 113, 120,
+  1,  1,  2,  9,  10, 17, 18, 25, 26, 33,  34,  41,  42,  49,  50,  57,
+  58, 65, 66, 73, 74, 81, 82, 89, 90, 97,  98,  105, 106, 113, 114, 121,
+  2,  2,  3,  10, 11, 18, 19, 26, 27, 34,  35,  42,  43,  50,  51,  58,
+  59, 66, 67, 74, 75, 82, 83, 90, 91, 98,  99,  106, 107, 114, 115, 122,
+  3,  3,  4,  11, 12, 19, 20, 27, 28, 35,  36,  43,  44,  51,  52,  59,
+  60, 67, 68, 75, 76, 83, 84, 91, 92, 99,  100, 107, 108, 115, 116, 123,
+  4,  4,  5,  12, 13, 20, 21, 28, 29, 36,  37,  44,  45,  52,  53,  60,
+  61, 68, 69, 76, 77, 84, 85, 92, 93, 100, 101, 108, 109, 116, 117, 124,
+  5,  5,  6,  13, 14, 21, 22, 29, 30, 37,  38,  45,  46,  53,  54,  61,
+  62, 69, 70, 77, 78, 85, 86, 93, 94, 101, 102, 109, 110, 117, 118, 125,
+  6,  6,  7,  14, 15, 22, 23, 30, 31, 38,  39,  46,  47,  54,  55,  62,
+  63, 70, 71, 78, 79, 86, 87, 94, 95, 102, 103, 110, 111, 118, 119, 126,
+  0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_16x8_neighbors[129 * MAX_NEIGHBORS]) = {
+  0,  0,  0,  0,  16, 16, 32, 32, 48, 48, 64, 64, 80, 80,  96,  96,
+  0,  0,  1,  16, 17, 32, 33, 48, 49, 64, 65, 80, 81, 96,  97,  112,
+  1,  1,  2,  17, 18, 33, 34, 49, 50, 65, 66, 81, 82, 97,  98,  113,
+  2,  2,  3,  18, 19, 34, 35, 50, 51, 66, 67, 82, 83, 98,  99,  114,
+  3,  3,  4,  19, 20, 35, 36, 51, 52, 67, 68, 83, 84, 99,  100, 115,
+  4,  4,  5,  20, 21, 36, 37, 52, 53, 68, 69, 84, 85, 100, 101, 116,
+  5,  5,  6,  21, 22, 37, 38, 53, 54, 69, 70, 85, 86, 101, 102, 117,
+  6,  6,  7,  22, 23, 38, 39, 54, 55, 70, 71, 86, 87, 102, 103, 118,
+  7,  7,  8,  23, 24, 39, 40, 55, 56, 71, 72, 87, 88, 103, 104, 119,
+  8,  8,  9,  24, 25, 40, 41, 56, 57, 72, 73, 88, 89, 104, 105, 120,
+  9,  9,  10, 25, 26, 41, 42, 57, 58, 73, 74, 89, 90, 105, 106, 121,
+  10, 10, 11, 26, 27, 42, 43, 58, 59, 74, 75, 90, 91, 106, 107, 122,
+  11, 11, 12, 27, 28, 43, 44, 59, 60, 75, 76, 91, 92, 107, 108, 123,
+  12, 12, 13, 28, 29, 44, 45, 60, 61, 76, 77, 92, 93, 108, 109, 124,
+  13, 13, 14, 29, 30, 45, 46, 61, 62, 77, 78, 93, 94, 109, 110, 125,
+  14, 14, 15, 30, 31, 46, 47, 62, 63, 78, 79, 94, 95, 110, 111, 126,
+  0,  0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_8x16_neighbors[129 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   1,   1,   2,   2,   3,   3,   4,   4,   5,   5,   6,
+  6,   0,   0,   1,   8,   2,   9,   3,   10,  4,   11,  5,   12,  6,   13,
+  7,   14,  8,   8,   9,   16,  10,  17,  11,  18,  12,  19,  13,  20,  14,
+  21,  15,  22,  16,  16,  17,  24,  18,  25,  19,  26,  20,  27,  21,  28,
+  22,  29,  23,  30,  24,  24,  25,  32,  26,  33,  27,  34,  28,  35,  29,
+  36,  30,  37,  31,  38,  32,  32,  33,  40,  34,  41,  35,  42,  36,  43,
+  37,  44,  38,  45,  39,  46,  40,  40,  41,  48,  42,  49,  43,  50,  44,
+  51,  45,  52,  46,  53,  47,  54,  48,  48,  49,  56,  50,  57,  51,  58,
+  52,  59,  53,  60,  54,  61,  55,  62,  56,  56,  57,  64,  58,  65,  59,
+  66,  60,  67,  61,  68,  62,  69,  63,  70,  64,  64,  65,  72,  66,  73,
+  67,  74,  68,  75,  69,  76,  70,  77,  71,  78,  72,  72,  73,  80,  74,
+  81,  75,  82,  76,  83,  77,  84,  78,  85,  79,  86,  80,  80,  81,  88,
+  82,  89,  83,  90,  84,  91,  85,  92,  86,  93,  87,  94,  88,  88,  89,
+  96,  90,  97,  91,  98,  92,  99,  93,  100, 94,  101, 95,  102, 96,  96,
+  97,  104, 98,  105, 99,  106, 100, 107, 101, 108, 102, 109, 103, 110, 104,
+  104, 105, 112, 106, 113, 107, 114, 108, 115, 109, 116, 110, 117, 111, 118,
+  112, 112, 113, 120, 114, 121, 115, 122, 116, 123, 117, 124, 118, 125, 119,
+  126, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_16x8_neighbors[129 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   1,   1,   2,   2,   3,   3,   4,   4,   5,   5,   6,
+  6,   7,   7,   8,   8,   9,   9,   10,  10,  11,  11,  12,  12,  13,  13,
+  14,  14,  0,   0,   1,   16,  2,   17,  3,   18,  4,   19,  5,   20,  6,
+  21,  7,   22,  8,   23,  9,   24,  10,  25,  11,  26,  12,  27,  13,  28,
+  14,  29,  15,  30,  16,  16,  17,  32,  18,  33,  19,  34,  20,  35,  21,
+  36,  22,  37,  23,  38,  24,  39,  25,  40,  26,  41,  27,  42,  28,  43,
+  29,  44,  30,  45,  31,  46,  32,  32,  33,  48,  34,  49,  35,  50,  36,
+  51,  37,  52,  38,  53,  39,  54,  40,  55,  41,  56,  42,  57,  43,  58,
+  44,  59,  45,  60,  46,  61,  47,  62,  48,  48,  49,  64,  50,  65,  51,
+  66,  52,  67,  53,  68,  54,  69,  55,  70,  56,  71,  57,  72,  58,  73,
+  59,  74,  60,  75,  61,  76,  62,  77,  63,  78,  64,  64,  65,  80,  66,
+  81,  67,  82,  68,  83,  69,  84,  70,  85,  71,  86,  72,  87,  73,  88,
+  74,  89,  75,  90,  76,  91,  77,  92,  78,  93,  79,  94,  80,  80,  81,
+  96,  82,  97,  83,  98,  84,  99,  85,  100, 86,  101, 87,  102, 88,  103,
+  89,  104, 90,  105, 91,  106, 92,  107, 93,  108, 94,  109, 95,  110, 96,
+  96,  97,  112, 98,  113, 99,  114, 100, 115, 101, 116, 102, 117, 103, 118,
+  104, 119, 105, 120, 106, 121, 107, 122, 108, 123, 109, 124, 110, 125, 111,
+  126, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_16x32_neighbors[513 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   0,   0,   1,   1,   1,   16,  16,  16,  2,   2,   2,
+  17,  17,  32,  32,  32,  3,   3,   3,   18,  18,  33,  33,  48,  48,  48,
+  4,   4,   4,   19,  19,  34,  34,  49,  49,  64,  64,  64,  5,   5,   5,
+  20,  20,  35,  35,  50,  50,  65,  65,  80,  80,  80,  6,   6,   6,   21,
+  21,  36,  36,  51,  51,  66,  66,  81,  81,  96,  96,  96,  7,   7,   7,
+  22,  22,  37,  37,  52,  52,  67,  67,  82,  82,  97,  97,  112, 112, 112,
+  8,   8,   8,   23,  23,  38,  38,  53,  53,  68,  68,  83,  83,  98,  98,
+  113, 113, 128, 128, 128, 9,   9,   9,   24,  24,  39,  39,  54,  54,  69,
+  69,  84,  84,  99,  99,  114, 114, 129, 129, 144, 144, 144, 10,  10,  10,
+  25,  25,  40,  40,  55,  55,  70,  70,  85,  85,  100, 100, 115, 115, 130,
+  130, 145, 145, 160, 160, 160, 11,  11,  11,  26,  26,  41,  41,  56,  56,
+  71,  71,  86,  86,  101, 101, 116, 116, 131, 131, 146, 146, 161, 161, 176,
+  176, 176, 12,  12,  12,  27,  27,  42,  42,  57,  57,  72,  72,  87,  87,
+  102, 102, 117, 117, 132, 132, 147, 147, 162, 162, 177, 177, 192, 192, 192,
+  13,  13,  13,  28,  28,  43,  43,  58,  58,  73,  73,  88,  88,  103, 103,
+  118, 118, 133, 133, 148, 148, 163, 163, 178, 178, 193, 193, 208, 208, 208,
+  14,  14,  14,  29,  29,  44,  44,  59,  59,  74,  74,  89,  89,  104, 104,
+  119, 119, 134, 134, 149, 149, 164, 164, 179, 179, 194, 194, 209, 209, 224,
+  224, 224, 15,  30,  30,  45,  45,  60,  60,  75,  75,  90,  90,  105, 105,
+  120, 120, 135, 135, 150, 150, 165, 165, 180, 180, 195, 195, 210, 210, 225,
+  225, 240, 240, 240, 31,  46,  46,  61,  61,  76,  76,  91,  91,  106, 106,
+  121, 121, 136, 136, 151, 151, 166, 166, 181, 181, 196, 196, 211, 211, 226,
+  226, 241, 241, 256, 256, 256, 47,  62,  62,  77,  77,  92,  92,  107, 107,
+  122, 122, 137, 137, 152, 152, 167, 167, 182, 182, 197, 197, 212, 212, 227,
+  227, 242, 242, 257, 257, 272, 272, 272, 63,  78,  78,  93,  93,  108, 108,
+  123, 123, 138, 138, 153, 153, 168, 168, 183, 183, 198, 198, 213, 213, 228,
+  228, 243, 243, 258, 258, 273, 273, 288, 288, 288, 79,  94,  94,  109, 109,
+  124, 124, 139, 139, 154, 154, 169, 169, 184, 184, 199, 199, 214, 214, 229,
+  229, 244, 244, 259, 259, 274, 274, 289, 289, 304, 304, 304, 95,  110, 110,
+  125, 125, 140, 140, 155, 155, 170, 170, 185, 185, 200, 200, 215, 215, 230,
+  230, 245, 245, 260, 260, 275, 275, 290, 290, 305, 305, 320, 320, 320, 111,
+  126, 126, 141, 141, 156, 156, 171, 171, 186, 186, 201, 201, 216, 216, 231,
+  231, 246, 246, 261, 261, 276, 276, 291, 291, 306, 306, 321, 321, 336, 336,
+  336, 127, 142, 142, 157, 157, 172, 172, 187, 187, 202, 202, 217, 217, 232,
+  232, 247, 247, 262, 262, 277, 277, 292, 292, 307, 307, 322, 322, 337, 337,
+  352, 352, 352, 143, 158, 158, 173, 173, 188, 188, 203, 203, 218, 218, 233,
+  233, 248, 248, 263, 263, 278, 278, 293, 293, 308, 308, 323, 323, 338, 338,
+  353, 353, 368, 368, 368, 159, 174, 174, 189, 189, 204, 204, 219, 219, 234,
+  234, 249, 249, 264, 264, 279, 279, 294, 294, 309, 309, 324, 324, 339, 339,
+  354, 354, 369, 369, 384, 384, 384, 175, 190, 190, 205, 205, 220, 220, 235,
+  235, 250, 250, 265, 265, 280, 280, 295, 295, 310, 310, 325, 325, 340, 340,
+  355, 355, 370, 370, 385, 385, 400, 400, 400, 191, 206, 206, 221, 221, 236,
+  236, 251, 251, 266, 266, 281, 281, 296, 296, 311, 311, 326, 326, 341, 341,
+  356, 356, 371, 371, 386, 386, 401, 401, 416, 416, 416, 207, 222, 222, 237,
+  237, 252, 252, 267, 267, 282, 282, 297, 297, 312, 312, 327, 327, 342, 342,
+  357, 357, 372, 372, 387, 387, 402, 402, 417, 417, 432, 432, 432, 223, 238,
+  238, 253, 253, 268, 268, 283, 283, 298, 298, 313, 313, 328, 328, 343, 343,
+  358, 358, 373, 373, 388, 388, 403, 403, 418, 418, 433, 433, 448, 448, 448,
+  239, 254, 254, 269, 269, 284, 284, 299, 299, 314, 314, 329, 329, 344, 344,
+  359, 359, 374, 374, 389, 389, 404, 404, 419, 419, 434, 434, 449, 449, 464,
+  464, 464, 255, 270, 270, 285, 285, 300, 300, 315, 315, 330, 330, 345, 345,
+  360, 360, 375, 375, 390, 390, 405, 405, 420, 420, 435, 435, 450, 450, 465,
+  465, 480, 480, 480, 271, 286, 286, 301, 301, 316, 316, 331, 331, 346, 346,
+  361, 361, 376, 376, 391, 391, 406, 406, 421, 421, 436, 436, 451, 451, 466,
+  466, 481, 481, 496, 287, 302, 302, 317, 317, 332, 332, 347, 347, 362, 362,
+  377, 377, 392, 392, 407, 407, 422, 422, 437, 437, 452, 452, 467, 467, 482,
+  482, 497, 303, 318, 318, 333, 333, 348, 348, 363, 363, 378, 378, 393, 393,
+  408, 408, 423, 423, 438, 438, 453, 453, 468, 468, 483, 483, 498, 319, 334,
+  334, 349, 349, 364, 364, 379, 379, 394, 394, 409, 409, 424, 424, 439, 439,
+  454, 454, 469, 469, 484, 484, 499, 335, 350, 350, 365, 365, 380, 380, 395,
+  395, 410, 410, 425, 425, 440, 440, 455, 455, 470, 470, 485, 485, 500, 351,
+  366, 366, 381, 381, 396, 396, 411, 411, 426, 426, 441, 441, 456, 456, 471,
+  471, 486, 486, 501, 367, 382, 382, 397, 397, 412, 412, 427, 427, 442, 442,
+  457, 457, 472, 472, 487, 487, 502, 383, 398, 398, 413, 413, 428, 428, 443,
+  443, 458, 458, 473, 473, 488, 488, 503, 399, 414, 414, 429, 429, 444, 444,
+  459, 459, 474, 474, 489, 489, 504, 415, 430, 430, 445, 445, 460, 460, 475,
+  475, 490, 490, 505, 431, 446, 446, 461, 461, 476, 476, 491, 491, 506, 447,
+  462, 462, 477, 477, 492, 492, 507, 463, 478, 478, 493, 493, 508, 479, 494,
+  494, 509, 495, 510, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_32x16_neighbors[513 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   0,   0,   1,   1,   1,   32,  32,  32,  2,   2,   2,
+  33,  33,  64,  64,  64,  3,   3,   3,   34,  34,  65,  65,  96,  96,  96,
+  4,   4,   4,   35,  35,  66,  66,  97,  97,  128, 128, 128, 5,   5,   5,
+  36,  36,  67,  67,  98,  98,  129, 129, 160, 160, 160, 6,   6,   6,   37,
+  37,  68,  68,  99,  99,  130, 130, 161, 161, 192, 192, 192, 7,   7,   7,
+  38,  38,  69,  69,  100, 100, 131, 131, 162, 162, 193, 193, 224, 224, 224,
+  8,   8,   8,   39,  39,  70,  70,  101, 101, 132, 132, 163, 163, 194, 194,
+  225, 225, 256, 256, 256, 9,   9,   9,   40,  40,  71,  71,  102, 102, 133,
+  133, 164, 164, 195, 195, 226, 226, 257, 257, 288, 288, 288, 10,  10,  10,
+  41,  41,  72,  72,  103, 103, 134, 134, 165, 165, 196, 196, 227, 227, 258,
+  258, 289, 289, 320, 320, 320, 11,  11,  11,  42,  42,  73,  73,  104, 104,
+  135, 135, 166, 166, 197, 197, 228, 228, 259, 259, 290, 290, 321, 321, 352,
+  352, 352, 12,  12,  12,  43,  43,  74,  74,  105, 105, 136, 136, 167, 167,
+  198, 198, 229, 229, 260, 260, 291, 291, 322, 322, 353, 353, 384, 384, 384,
+  13,  13,  13,  44,  44,  75,  75,  106, 106, 137, 137, 168, 168, 199, 199,
+  230, 230, 261, 261, 292, 292, 323, 323, 354, 354, 385, 385, 416, 416, 416,
+  14,  14,  14,  45,  45,  76,  76,  107, 107, 138, 138, 169, 169, 200, 200,
+  231, 231, 262, 262, 293, 293, 324, 324, 355, 355, 386, 386, 417, 417, 448,
+  448, 448, 15,  15,  15,  46,  46,  77,  77,  108, 108, 139, 139, 170, 170,
+  201, 201, 232, 232, 263, 263, 294, 294, 325, 325, 356, 356, 387, 387, 418,
+  418, 449, 449, 480, 16,  16,  16,  47,  47,  78,  78,  109, 109, 140, 140,
+  171, 171, 202, 202, 233, 233, 264, 264, 295, 295, 326, 326, 357, 357, 388,
+  388, 419, 419, 450, 450, 481, 17,  17,  17,  48,  48,  79,  79,  110, 110,
+  141, 141, 172, 172, 203, 203, 234, 234, 265, 265, 296, 296, 327, 327, 358,
+  358, 389, 389, 420, 420, 451, 451, 482, 18,  18,  18,  49,  49,  80,  80,
+  111, 111, 142, 142, 173, 173, 204, 204, 235, 235, 266, 266, 297, 297, 328,
+  328, 359, 359, 390, 390, 421, 421, 452, 452, 483, 19,  19,  19,  50,  50,
+  81,  81,  112, 112, 143, 143, 174, 174, 205, 205, 236, 236, 267, 267, 298,
+  298, 329, 329, 360, 360, 391, 391, 422, 422, 453, 453, 484, 20,  20,  20,
+  51,  51,  82,  82,  113, 113, 144, 144, 175, 175, 206, 206, 237, 237, 268,
+  268, 299, 299, 330, 330, 361, 361, 392, 392, 423, 423, 454, 454, 485, 21,
+  21,  21,  52,  52,  83,  83,  114, 114, 145, 145, 176, 176, 207, 207, 238,
+  238, 269, 269, 300, 300, 331, 331, 362, 362, 393, 393, 424, 424, 455, 455,
+  486, 22,  22,  22,  53,  53,  84,  84,  115, 115, 146, 146, 177, 177, 208,
+  208, 239, 239, 270, 270, 301, 301, 332, 332, 363, 363, 394, 394, 425, 425,
+  456, 456, 487, 23,  23,  23,  54,  54,  85,  85,  116, 116, 147, 147, 178,
+  178, 209, 209, 240, 240, 271, 271, 302, 302, 333, 333, 364, 364, 395, 395,
+  426, 426, 457, 457, 488, 24,  24,  24,  55,  55,  86,  86,  117, 117, 148,
+  148, 179, 179, 210, 210, 241, 241, 272, 272, 303, 303, 334, 334, 365, 365,
+  396, 396, 427, 427, 458, 458, 489, 25,  25,  25,  56,  56,  87,  87,  118,
+  118, 149, 149, 180, 180, 211, 211, 242, 242, 273, 273, 304, 304, 335, 335,
+  366, 366, 397, 397, 428, 428, 459, 459, 490, 26,  26,  26,  57,  57,  88,
+  88,  119, 119, 150, 150, 181, 181, 212, 212, 243, 243, 274, 274, 305, 305,
+  336, 336, 367, 367, 398, 398, 429, 429, 460, 460, 491, 27,  27,  27,  58,
+  58,  89,  89,  120, 120, 151, 151, 182, 182, 213, 213, 244, 244, 275, 275,
+  306, 306, 337, 337, 368, 368, 399, 399, 430, 430, 461, 461, 492, 28,  28,
+  28,  59,  59,  90,  90,  121, 121, 152, 152, 183, 183, 214, 214, 245, 245,
+  276, 276, 307, 307, 338, 338, 369, 369, 400, 400, 431, 431, 462, 462, 493,
+  29,  29,  29,  60,  60,  91,  91,  122, 122, 153, 153, 184, 184, 215, 215,
+  246, 246, 277, 277, 308, 308, 339, 339, 370, 370, 401, 401, 432, 432, 463,
+  463, 494, 30,  30,  30,  61,  61,  92,  92,  123, 123, 154, 154, 185, 185,
+  216, 216, 247, 247, 278, 278, 309, 309, 340, 340, 371, 371, 402, 402, 433,
+  433, 464, 464, 495, 31,  62,  62,  93,  93,  124, 124, 155, 155, 186, 186,
+  217, 217, 248, 248, 279, 279, 310, 310, 341, 341, 372, 372, 403, 403, 434,
+  434, 465, 465, 496, 63,  94,  94,  125, 125, 156, 156, 187, 187, 218, 218,
+  249, 249, 280, 280, 311, 311, 342, 342, 373, 373, 404, 404, 435, 435, 466,
+  466, 497, 95,  126, 126, 157, 157, 188, 188, 219, 219, 250, 250, 281, 281,
+  312, 312, 343, 343, 374, 374, 405, 405, 436, 436, 467, 467, 498, 127, 158,
+  158, 189, 189, 220, 220, 251, 251, 282, 282, 313, 313, 344, 344, 375, 375,
+  406, 406, 437, 437, 468, 468, 499, 159, 190, 190, 221, 221, 252, 252, 283,
+  283, 314, 314, 345, 345, 376, 376, 407, 407, 438, 438, 469, 469, 500, 191,
+  222, 222, 253, 253, 284, 284, 315, 315, 346, 346, 377, 377, 408, 408, 439,
+  439, 470, 470, 501, 223, 254, 254, 285, 285, 316, 316, 347, 347, 378, 378,
+  409, 409, 440, 440, 471, 471, 502, 255, 286, 286, 317, 317, 348, 348, 379,
+  379, 410, 410, 441, 441, 472, 472, 503, 287, 318, 318, 349, 349, 380, 380,
+  411, 411, 442, 442, 473, 473, 504, 319, 350, 350, 381, 381, 412, 412, 443,
+  443, 474, 474, 505, 351, 382, 382, 413, 413, 444, 444, 475, 475, 506, 383,
+  414, 414, 445, 445, 476, 476, 507, 415, 446, 446, 477, 477, 508, 447, 478,
+  478, 509, 479, 510, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_16x32_neighbors[513 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   16,  16,  32,  32,  48,  48,  64,  64,  80,  80,  96,
+  96,  112, 112, 128, 128, 144, 144, 160, 160, 176, 176, 192, 192, 208, 208,
+  224, 224, 240, 240, 256, 256, 272, 272, 288, 288, 304, 304, 320, 320, 336,
+  336, 352, 352, 368, 368, 384, 384, 400, 400, 416, 416, 432, 432, 448, 448,
+  464, 464, 480, 480, 0,   0,   1,   16,  17,  32,  33,  48,  49,  64,  65,
+  80,  81,  96,  97,  112, 113, 128, 129, 144, 145, 160, 161, 176, 177, 192,
+  193, 208, 209, 224, 225, 240, 241, 256, 257, 272, 273, 288, 289, 304, 305,
+  320, 321, 336, 337, 352, 353, 368, 369, 384, 385, 400, 401, 416, 417, 432,
+  433, 448, 449, 464, 465, 480, 481, 496, 1,   1,   2,   17,  18,  33,  34,
+  49,  50,  65,  66,  81,  82,  97,  98,  113, 114, 129, 130, 145, 146, 161,
+  162, 177, 178, 193, 194, 209, 210, 225, 226, 241, 242, 257, 258, 273, 274,
+  289, 290, 305, 306, 321, 322, 337, 338, 353, 354, 369, 370, 385, 386, 401,
+  402, 417, 418, 433, 434, 449, 450, 465, 466, 481, 482, 497, 2,   2,   3,
+  18,  19,  34,  35,  50,  51,  66,  67,  82,  83,  98,  99,  114, 115, 130,
+  131, 146, 147, 162, 163, 178, 179, 194, 195, 210, 211, 226, 227, 242, 243,
+  258, 259, 274, 275, 290, 291, 306, 307, 322, 323, 338, 339, 354, 355, 370,
+  371, 386, 387, 402, 403, 418, 419, 434, 435, 450, 451, 466, 467, 482, 483,
+  498, 3,   3,   4,   19,  20,  35,  36,  51,  52,  67,  68,  83,  84,  99,
+  100, 115, 116, 131, 132, 147, 148, 163, 164, 179, 180, 195, 196, 211, 212,
+  227, 228, 243, 244, 259, 260, 275, 276, 291, 292, 307, 308, 323, 324, 339,
+  340, 355, 356, 371, 372, 387, 388, 403, 404, 419, 420, 435, 436, 451, 452,
+  467, 468, 483, 484, 499, 4,   4,   5,   20,  21,  36,  37,  52,  53,  68,
+  69,  84,  85,  100, 101, 116, 117, 132, 133, 148, 149, 164, 165, 180, 181,
+  196, 197, 212, 213, 228, 229, 244, 245, 260, 261, 276, 277, 292, 293, 308,
+  309, 324, 325, 340, 341, 356, 357, 372, 373, 388, 389, 404, 405, 420, 421,
+  436, 437, 452, 453, 468, 469, 484, 485, 500, 5,   5,   6,   21,  22,  37,
+  38,  53,  54,  69,  70,  85,  86,  101, 102, 117, 118, 133, 134, 149, 150,
+  165, 166, 181, 182, 197, 198, 213, 214, 229, 230, 245, 246, 261, 262, 277,
+  278, 293, 294, 309, 310, 325, 326, 341, 342, 357, 358, 373, 374, 389, 390,
+  405, 406, 421, 422, 437, 438, 453, 454, 469, 470, 485, 486, 501, 6,   6,
+  7,   22,  23,  38,  39,  54,  55,  70,  71,  86,  87,  102, 103, 118, 119,
+  134, 135, 150, 151, 166, 167, 182, 183, 198, 199, 214, 215, 230, 231, 246,
+  247, 262, 263, 278, 279, 294, 295, 310, 311, 326, 327, 342, 343, 358, 359,
+  374, 375, 390, 391, 406, 407, 422, 423, 438, 439, 454, 455, 470, 471, 486,
+  487, 502, 7,   7,   8,   23,  24,  39,  40,  55,  56,  71,  72,  87,  88,
+  103, 104, 119, 120, 135, 136, 151, 152, 167, 168, 183, 184, 199, 200, 215,
+  216, 231, 232, 247, 248, 263, 264, 279, 280, 295, 296, 311, 312, 327, 328,
+  343, 344, 359, 360, 375, 376, 391, 392, 407, 408, 423, 424, 439, 440, 455,
+  456, 471, 472, 487, 488, 503, 8,   8,   9,   24,  25,  40,  41,  56,  57,
+  72,  73,  88,  89,  104, 105, 120, 121, 136, 137, 152, 153, 168, 169, 184,
+  185, 200, 201, 216, 217, 232, 233, 248, 249, 264, 265, 280, 281, 296, 297,
+  312, 313, 328, 329, 344, 345, 360, 361, 376, 377, 392, 393, 408, 409, 424,
+  425, 440, 441, 456, 457, 472, 473, 488, 489, 504, 9,   9,   10,  25,  26,
+  41,  42,  57,  58,  73,  74,  89,  90,  105, 106, 121, 122, 137, 138, 153,
+  154, 169, 170, 185, 186, 201, 202, 217, 218, 233, 234, 249, 250, 265, 266,
+  281, 282, 297, 298, 313, 314, 329, 330, 345, 346, 361, 362, 377, 378, 393,
+  394, 409, 410, 425, 426, 441, 442, 457, 458, 473, 474, 489, 490, 505, 10,
+  10,  11,  26,  27,  42,  43,  58,  59,  74,  75,  90,  91,  106, 107, 122,
+  123, 138, 139, 154, 155, 170, 171, 186, 187, 202, 203, 218, 219, 234, 235,
+  250, 251, 266, 267, 282, 283, 298, 299, 314, 315, 330, 331, 346, 347, 362,
+  363, 378, 379, 394, 395, 410, 411, 426, 427, 442, 443, 458, 459, 474, 475,
+  490, 491, 506, 11,  11,  12,  27,  28,  43,  44,  59,  60,  75,  76,  91,
+  92,  107, 108, 123, 124, 139, 140, 155, 156, 171, 172, 187, 188, 203, 204,
+  219, 220, 235, 236, 251, 252, 267, 268, 283, 284, 299, 300, 315, 316, 331,
+  332, 347, 348, 363, 364, 379, 380, 395, 396, 411, 412, 427, 428, 443, 444,
+  459, 460, 475, 476, 491, 492, 507, 12,  12,  13,  28,  29,  44,  45,  60,
+  61,  76,  77,  92,  93,  108, 109, 124, 125, 140, 141, 156, 157, 172, 173,
+  188, 189, 204, 205, 220, 221, 236, 237, 252, 253, 268, 269, 284, 285, 300,
+  301, 316, 317, 332, 333, 348, 349, 364, 365, 380, 381, 396, 397, 412, 413,
+  428, 429, 444, 445, 460, 461, 476, 477, 492, 493, 508, 13,  13,  14,  29,
+  30,  45,  46,  61,  62,  77,  78,  93,  94,  109, 110, 125, 126, 141, 142,
+  157, 158, 173, 174, 189, 190, 205, 206, 221, 222, 237, 238, 253, 254, 269,
+  270, 285, 286, 301, 302, 317, 318, 333, 334, 349, 350, 365, 366, 381, 382,
+  397, 398, 413, 414, 429, 430, 445, 446, 461, 462, 477, 478, 493, 494, 509,
+  14,  14,  15,  30,  31,  46,  47,  62,  63,  78,  79,  94,  95,  110, 111,
+  126, 127, 142, 143, 158, 159, 174, 175, 190, 191, 206, 207, 222, 223, 238,
+  239, 254, 255, 270, 271, 286, 287, 302, 303, 318, 319, 334, 335, 350, 351,
+  366, 367, 382, 383, 398, 399, 414, 415, 430, 431, 446, 447, 462, 463, 478,
+  479, 494, 495, 510, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_32x16_neighbors[513 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   32,  32,  64,  64,  96,  96,  128, 128, 160, 160, 192,
+  192, 224, 224, 256, 256, 288, 288, 320, 320, 352, 352, 384, 384, 416, 416,
+  448, 448, 0,   0,   1,   32,  33,  64,  65,  96,  97,  128, 129, 160, 161,
+  192, 193, 224, 225, 256, 257, 288, 289, 320, 321, 352, 353, 384, 385, 416,
+  417, 448, 449, 480, 1,   1,   2,   33,  34,  65,  66,  97,  98,  129, 130,
+  161, 162, 193, 194, 225, 226, 257, 258, 289, 290, 321, 322, 353, 354, 385,
+  386, 417, 418, 449, 450, 481, 2,   2,   3,   34,  35,  66,  67,  98,  99,
+  130, 131, 162, 163, 194, 195, 226, 227, 258, 259, 290, 291, 322, 323, 354,
+  355, 386, 387, 418, 419, 450, 451, 482, 3,   3,   4,   35,  36,  67,  68,
+  99,  100, 131, 132, 163, 164, 195, 196, 227, 228, 259, 260, 291, 292, 323,
+  324, 355, 356, 387, 388, 419, 420, 451, 452, 483, 4,   4,   5,   36,  37,
+  68,  69,  100, 101, 132, 133, 164, 165, 196, 197, 228, 229, 260, 261, 292,
+  293, 324, 325, 356, 357, 388, 389, 420, 421, 452, 453, 484, 5,   5,   6,
+  37,  38,  69,  70,  101, 102, 133, 134, 165, 166, 197, 198, 229, 230, 261,
+  262, 293, 294, 325, 326, 357, 358, 389, 390, 421, 422, 453, 454, 485, 6,
+  6,   7,   38,  39,  70,  71,  102, 103, 134, 135, 166, 167, 198, 199, 230,
+  231, 262, 263, 294, 295, 326, 327, 358, 359, 390, 391, 422, 423, 454, 455,
+  486, 7,   7,   8,   39,  40,  71,  72,  103, 104, 135, 136, 167, 168, 199,
+  200, 231, 232, 263, 264, 295, 296, 327, 328, 359, 360, 391, 392, 423, 424,
+  455, 456, 487, 8,   8,   9,   40,  41,  72,  73,  104, 105, 136, 137, 168,
+  169, 200, 201, 232, 233, 264, 265, 296, 297, 328, 329, 360, 361, 392, 393,
+  424, 425, 456, 457, 488, 9,   9,   10,  41,  42,  73,  74,  105, 106, 137,
+  138, 169, 170, 201, 202, 233, 234, 265, 266, 297, 298, 329, 330, 361, 362,
+  393, 394, 425, 426, 457, 458, 489, 10,  10,  11,  42,  43,  74,  75,  106,
+  107, 138, 139, 170, 171, 202, 203, 234, 235, 266, 267, 298, 299, 330, 331,
+  362, 363, 394, 395, 426, 427, 458, 459, 490, 11,  11,  12,  43,  44,  75,
+  76,  107, 108, 139, 140, 171, 172, 203, 204, 235, 236, 267, 268, 299, 300,
+  331, 332, 363, 364, 395, 396, 427, 428, 459, 460, 491, 12,  12,  13,  44,
+  45,  76,  77,  108, 109, 140, 141, 172, 173, 204, 205, 236, 237, 268, 269,
+  300, 301, 332, 333, 364, 365, 396, 397, 428, 429, 460, 461, 492, 13,  13,
+  14,  45,  46,  77,  78,  109, 110, 141, 142, 173, 174, 205, 206, 237, 238,
+  269, 270, 301, 302, 333, 334, 365, 366, 397, 398, 429, 430, 461, 462, 493,
+  14,  14,  15,  46,  47,  78,  79,  110, 111, 142, 143, 174, 175, 206, 207,
+  238, 239, 270, 271, 302, 303, 334, 335, 366, 367, 398, 399, 430, 431, 462,
+  463, 494, 15,  15,  16,  47,  48,  79,  80,  111, 112, 143, 144, 175, 176,
+  207, 208, 239, 240, 271, 272, 303, 304, 335, 336, 367, 368, 399, 400, 431,
+  432, 463, 464, 495, 16,  16,  17,  48,  49,  80,  81,  112, 113, 144, 145,
+  176, 177, 208, 209, 240, 241, 272, 273, 304, 305, 336, 337, 368, 369, 400,
+  401, 432, 433, 464, 465, 496, 17,  17,  18,  49,  50,  81,  82,  113, 114,
+  145, 146, 177, 178, 209, 210, 241, 242, 273, 274, 305, 306, 337, 338, 369,
+  370, 401, 402, 433, 434, 465, 466, 497, 18,  18,  19,  50,  51,  82,  83,
+  114, 115, 146, 147, 178, 179, 210, 211, 242, 243, 274, 275, 306, 307, 338,
+  339, 370, 371, 402, 403, 434, 435, 466, 467, 498, 19,  19,  20,  51,  52,
+  83,  84,  115, 116, 147, 148, 179, 180, 211, 212, 243, 244, 275, 276, 307,
+  308, 339, 340, 371, 372, 403, 404, 435, 436, 467, 468, 499, 20,  20,  21,
+  52,  53,  84,  85,  116, 117, 148, 149, 180, 181, 212, 213, 244, 245, 276,
+  277, 308, 309, 340, 341, 372, 373, 404, 405, 436, 437, 468, 469, 500, 21,
+  21,  22,  53,  54,  85,  86,  117, 118, 149, 150, 181, 182, 213, 214, 245,
+  246, 277, 278, 309, 310, 341, 342, 373, 374, 405, 406, 437, 438, 469, 470,
+  501, 22,  22,  23,  54,  55,  86,  87,  118, 119, 150, 151, 182, 183, 214,
+  215, 246, 247, 278, 279, 310, 311, 342, 343, 374, 375, 406, 407, 438, 439,
+  470, 471, 502, 23,  23,  24,  55,  56,  87,  88,  119, 120, 151, 152, 183,
+  184, 215, 216, 247, 248, 279, 280, 311, 312, 343, 344, 375, 376, 407, 408,
+  439, 440, 471, 472, 503, 24,  24,  25,  56,  57,  88,  89,  120, 121, 152,
+  153, 184, 185, 216, 217, 248, 249, 280, 281, 312, 313, 344, 345, 376, 377,
+  408, 409, 440, 441, 472, 473, 504, 25,  25,  26,  57,  58,  89,  90,  121,
+  122, 153, 154, 185, 186, 217, 218, 249, 250, 281, 282, 313, 314, 345, 346,
+  377, 378, 409, 410, 441, 442, 473, 474, 505, 26,  26,  27,  58,  59,  90,
+  91,  122, 123, 154, 155, 186, 187, 218, 219, 250, 251, 282, 283, 314, 315,
+  346, 347, 378, 379, 410, 411, 442, 443, 474, 475, 506, 27,  27,  28,  59,
+  60,  91,  92,  123, 124, 155, 156, 187, 188, 219, 220, 251, 252, 283, 284,
+  315, 316, 347, 348, 379, 380, 411, 412, 443, 444, 475, 476, 507, 28,  28,
+  29,  60,  61,  92,  93,  124, 125, 156, 157, 188, 189, 220, 221, 252, 253,
+  284, 285, 316, 317, 348, 349, 380, 381, 412, 413, 444, 445, 476, 477, 508,
+  29,  29,  30,  61,  62,  93,  94,  125, 126, 157, 158, 189, 190, 221, 222,
+  253, 254, 285, 286, 317, 318, 349, 350, 381, 382, 413, 414, 445, 446, 477,
+  478, 509, 30,  30,  31,  62,  63,  94,  95,  126, 127, 158, 159, 190, 191,
+  222, 223, 254, 255, 286, 287, 318, 319, 350, 351, 382, 383, 414, 415, 446,
+  447, 478, 479, 510, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_16x32_neighbors[513 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   1,   1,   2,   2,   3,   3,   4,   4,   5,   5,   6,
+  6,   7,   7,   8,   8,   9,   9,   10,  10,  11,  11,  12,  12,  13,  13,
+  14,  14,  0,   0,   1,   16,  2,   17,  3,   18,  4,   19,  5,   20,  6,
+  21,  7,   22,  8,   23,  9,   24,  10,  25,  11,  26,  12,  27,  13,  28,
+  14,  29,  15,  30,  16,  16,  17,  32,  18,  33,  19,  34,  20,  35,  21,
+  36,  22,  37,  23,  38,  24,  39,  25,  40,  26,  41,  27,  42,  28,  43,
+  29,  44,  30,  45,  31,  46,  32,  32,  33,  48,  34,  49,  35,  50,  36,
+  51,  37,  52,  38,  53,  39,  54,  40,  55,  41,  56,  42,  57,  43,  58,
+  44,  59,  45,  60,  46,  61,  47,  62,  48,  48,  49,  64,  50,  65,  51,
+  66,  52,  67,  53,  68,  54,  69,  55,  70,  56,  71,  57,  72,  58,  73,
+  59,  74,  60,  75,  61,  76,  62,  77,  63,  78,  64,  64,  65,  80,  66,
+  81,  67,  82,  68,  83,  69,  84,  70,  85,  71,  86,  72,  87,  73,  88,
+  74,  89,  75,  90,  76,  91,  77,  92,  78,  93,  79,  94,  80,  80,  81,
+  96,  82,  97,  83,  98,  84,  99,  85,  100, 86,  101, 87,  102, 88,  103,
+  89,  104, 90,  105, 91,  106, 92,  107, 93,  108, 94,  109, 95,  110, 96,
+  96,  97,  112, 98,  113, 99,  114, 100, 115, 101, 116, 102, 117, 103, 118,
+  104, 119, 105, 120, 106, 121, 107, 122, 108, 123, 109, 124, 110, 125, 111,
+  126, 112, 112, 113, 128, 114, 129, 115, 130, 116, 131, 117, 132, 118, 133,
+  119, 134, 120, 135, 121, 136, 122, 137, 123, 138, 124, 139, 125, 140, 126,
+  141, 127, 142, 128, 128, 129, 144, 130, 145, 131, 146, 132, 147, 133, 148,
+  134, 149, 135, 150, 136, 151, 137, 152, 138, 153, 139, 154, 140, 155, 141,
+  156, 142, 157, 143, 158, 144, 144, 145, 160, 146, 161, 147, 162, 148, 163,
+  149, 164, 150, 165, 151, 166, 152, 167, 153, 168, 154, 169, 155, 170, 156,
+  171, 157, 172, 158, 173, 159, 174, 160, 160, 161, 176, 162, 177, 163, 178,
+  164, 179, 165, 180, 166, 181, 167, 182, 168, 183, 169, 184, 170, 185, 171,
+  186, 172, 187, 173, 188, 174, 189, 175, 190, 176, 176, 177, 192, 178, 193,
+  179, 194, 180, 195, 181, 196, 182, 197, 183, 198, 184, 199, 185, 200, 186,
+  201, 187, 202, 188, 203, 189, 204, 190, 205, 191, 206, 192, 192, 193, 208,
+  194, 209, 195, 210, 196, 211, 197, 212, 198, 213, 199, 214, 200, 215, 201,
+  216, 202, 217, 203, 218, 204, 219, 205, 220, 206, 221, 207, 222, 208, 208,
+  209, 224, 210, 225, 211, 226, 212, 227, 213, 228, 214, 229, 215, 230, 216,
+  231, 217, 232, 218, 233, 219, 234, 220, 235, 221, 236, 222, 237, 223, 238,
+  224, 224, 225, 240, 226, 241, 227, 242, 228, 243, 229, 244, 230, 245, 231,
+  246, 232, 247, 233, 248, 234, 249, 235, 250, 236, 251, 237, 252, 238, 253,
+  239, 254, 240, 240, 241, 256, 242, 257, 243, 258, 244, 259, 245, 260, 246,
+  261, 247, 262, 248, 263, 249, 264, 250, 265, 251, 266, 252, 267, 253, 268,
+  254, 269, 255, 270, 256, 256, 257, 272, 258, 273, 259, 274, 260, 275, 261,
+  276, 262, 277, 263, 278, 264, 279, 265, 280, 266, 281, 267, 282, 268, 283,
+  269, 284, 270, 285, 271, 286, 272, 272, 273, 288, 274, 289, 275, 290, 276,
+  291, 277, 292, 278, 293, 279, 294, 280, 295, 281, 296, 282, 297, 283, 298,
+  284, 299, 285, 300, 286, 301, 287, 302, 288, 288, 289, 304, 290, 305, 291,
+  306, 292, 307, 293, 308, 294, 309, 295, 310, 296, 311, 297, 312, 298, 313,
+  299, 314, 300, 315, 301, 316, 302, 317, 303, 318, 304, 304, 305, 320, 306,
+  321, 307, 322, 308, 323, 309, 324, 310, 325, 311, 326, 312, 327, 313, 328,
+  314, 329, 315, 330, 316, 331, 317, 332, 318, 333, 319, 334, 320, 320, 321,
+  336, 322, 337, 323, 338, 324, 339, 325, 340, 326, 341, 327, 342, 328, 343,
+  329, 344, 330, 345, 331, 346, 332, 347, 333, 348, 334, 349, 335, 350, 336,
+  336, 337, 352, 338, 353, 339, 354, 340, 355, 341, 356, 342, 357, 343, 358,
+  344, 359, 345, 360, 346, 361, 347, 362, 348, 363, 349, 364, 350, 365, 351,
+  366, 352, 352, 353, 368, 354, 369, 355, 370, 356, 371, 357, 372, 358, 373,
+  359, 374, 360, 375, 361, 376, 362, 377, 363, 378, 364, 379, 365, 380, 366,
+  381, 367, 382, 368, 368, 369, 384, 370, 385, 371, 386, 372, 387, 373, 388,
+  374, 389, 375, 390, 376, 391, 377, 392, 378, 393, 379, 394, 380, 395, 381,
+  396, 382, 397, 383, 398, 384, 384, 385, 400, 386, 401, 387, 402, 388, 403,
+  389, 404, 390, 405, 391, 406, 392, 407, 393, 408, 394, 409, 395, 410, 396,
+  411, 397, 412, 398, 413, 399, 414, 400, 400, 401, 416, 402, 417, 403, 418,
+  404, 419, 405, 420, 406, 421, 407, 422, 408, 423, 409, 424, 410, 425, 411,
+  426, 412, 427, 413, 428, 414, 429, 415, 430, 416, 416, 417, 432, 418, 433,
+  419, 434, 420, 435, 421, 436, 422, 437, 423, 438, 424, 439, 425, 440, 426,
+  441, 427, 442, 428, 443, 429, 444, 430, 445, 431, 446, 432, 432, 433, 448,
+  434, 449, 435, 450, 436, 451, 437, 452, 438, 453, 439, 454, 440, 455, 441,
+  456, 442, 457, 443, 458, 444, 459, 445, 460, 446, 461, 447, 462, 448, 448,
+  449, 464, 450, 465, 451, 466, 452, 467, 453, 468, 454, 469, 455, 470, 456,
+  471, 457, 472, 458, 473, 459, 474, 460, 475, 461, 476, 462, 477, 463, 478,
+  464, 464, 465, 480, 466, 481, 467, 482, 468, 483, 469, 484, 470, 485, 471,
+  486, 472, 487, 473, 488, 474, 489, 475, 490, 476, 491, 477, 492, 478, 493,
+  479, 494, 480, 480, 481, 496, 482, 497, 483, 498, 484, 499, 485, 500, 486,
+  501, 487, 502, 488, 503, 489, 504, 490, 505, 491, 506, 492, 507, 493, 508,
+  494, 509, 495, 510, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_32x16_neighbors[513 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   1,   1,   2,   2,   3,   3,   4,   4,   5,   5,   6,
+  6,   7,   7,   8,   8,   9,   9,   10,  10,  11,  11,  12,  12,  13,  13,
+  14,  14,  15,  15,  16,  16,  17,  17,  18,  18,  19,  19,  20,  20,  21,
+  21,  22,  22,  23,  23,  24,  24,  25,  25,  26,  26,  27,  27,  28,  28,
+  29,  29,  30,  30,  0,   0,   1,   32,  2,   33,  3,   34,  4,   35,  5,
+  36,  6,   37,  7,   38,  8,   39,  9,   40,  10,  41,  11,  42,  12,  43,
+  13,  44,  14,  45,  15,  46,  16,  47,  17,  48,  18,  49,  19,  50,  20,
+  51,  21,  52,  22,  53,  23,  54,  24,  55,  25,  56,  26,  57,  27,  58,
+  28,  59,  29,  60,  30,  61,  31,  62,  32,  32,  33,  64,  34,  65,  35,
+  66,  36,  67,  37,  68,  38,  69,  39,  70,  40,  71,  41,  72,  42,  73,
+  43,  74,  44,  75,  45,  76,  46,  77,  47,  78,  48,  79,  49,  80,  50,
+  81,  51,  82,  52,  83,  53,  84,  54,  85,  55,  86,  56,  87,  57,  88,
+  58,  89,  59,  90,  60,  91,  61,  92,  62,  93,  63,  94,  64,  64,  65,
+  96,  66,  97,  67,  98,  68,  99,  69,  100, 70,  101, 71,  102, 72,  103,
+  73,  104, 74,  105, 75,  106, 76,  107, 77,  108, 78,  109, 79,  110, 80,
+  111, 81,  112, 82,  113, 83,  114, 84,  115, 85,  116, 86,  117, 87,  118,
+  88,  119, 89,  120, 90,  121, 91,  122, 92,  123, 93,  124, 94,  125, 95,
+  126, 96,  96,  97,  128, 98,  129, 99,  130, 100, 131, 101, 132, 102, 133,
+  103, 134, 104, 135, 105, 136, 106, 137, 107, 138, 108, 139, 109, 140, 110,
+  141, 111, 142, 112, 143, 113, 144, 114, 145, 115, 146, 116, 147, 117, 148,
+  118, 149, 119, 150, 120, 151, 121, 152, 122, 153, 123, 154, 124, 155, 125,
+  156, 126, 157, 127, 158, 128, 128, 129, 160, 130, 161, 131, 162, 132, 163,
+  133, 164, 134, 165, 135, 166, 136, 167, 137, 168, 138, 169, 139, 170, 140,
+  171, 141, 172, 142, 173, 143, 174, 144, 175, 145, 176, 146, 177, 147, 178,
+  148, 179, 149, 180, 150, 181, 151, 182, 152, 183, 153, 184, 154, 185, 155,
+  186, 156, 187, 157, 188, 158, 189, 159, 190, 160, 160, 161, 192, 162, 193,
+  163, 194, 164, 195, 165, 196, 166, 197, 167, 198, 168, 199, 169, 200, 170,
+  201, 171, 202, 172, 203, 173, 204, 174, 205, 175, 206, 176, 207, 177, 208,
+  178, 209, 179, 210, 180, 211, 181, 212, 182, 213, 183, 214, 184, 215, 185,
+  216, 186, 217, 187, 218, 188, 219, 189, 220, 190, 221, 191, 222, 192, 192,
+  193, 224, 194, 225, 195, 226, 196, 227, 197, 228, 198, 229, 199, 230, 200,
+  231, 201, 232, 202, 233, 203, 234, 204, 235, 205, 236, 206, 237, 207, 238,
+  208, 239, 209, 240, 210, 241, 211, 242, 212, 243, 213, 244, 214, 245, 215,
+  246, 216, 247, 217, 248, 218, 249, 219, 250, 220, 251, 221, 252, 222, 253,
+  223, 254, 224, 224, 225, 256, 226, 257, 227, 258, 228, 259, 229, 260, 230,
+  261, 231, 262, 232, 263, 233, 264, 234, 265, 235, 266, 236, 267, 237, 268,
+  238, 269, 239, 270, 240, 271, 241, 272, 242, 273, 243, 274, 244, 275, 245,
+  276, 246, 277, 247, 278, 248, 279, 249, 280, 250, 281, 251, 282, 252, 283,
+  253, 284, 254, 285, 255, 286, 256, 256, 257, 288, 258, 289, 259, 290, 260,
+  291, 261, 292, 262, 293, 263, 294, 264, 295, 265, 296, 266, 297, 267, 298,
+  268, 299, 269, 300, 270, 301, 271, 302, 272, 303, 273, 304, 274, 305, 275,
+  306, 276, 307, 277, 308, 278, 309, 279, 310, 280, 311, 281, 312, 282, 313,
+  283, 314, 284, 315, 285, 316, 286, 317, 287, 318, 288, 288, 289, 320, 290,
+  321, 291, 322, 292, 323, 293, 324, 294, 325, 295, 326, 296, 327, 297, 328,
+  298, 329, 299, 330, 300, 331, 301, 332, 302, 333, 303, 334, 304, 335, 305,
+  336, 306, 337, 307, 338, 308, 339, 309, 340, 310, 341, 311, 342, 312, 343,
+  313, 344, 314, 345, 315, 346, 316, 347, 317, 348, 318, 349, 319, 350, 320,
+  320, 321, 352, 322, 353, 323, 354, 324, 355, 325, 356, 326, 357, 327, 358,
+  328, 359, 329, 360, 330, 361, 331, 362, 332, 363, 333, 364, 334, 365, 335,
+  366, 336, 367, 337, 368, 338, 369, 339, 370, 340, 371, 341, 372, 342, 373,
+  343, 374, 344, 375, 345, 376, 346, 377, 347, 378, 348, 379, 349, 380, 350,
+  381, 351, 382, 352, 352, 353, 384, 354, 385, 355, 386, 356, 387, 357, 388,
+  358, 389, 359, 390, 360, 391, 361, 392, 362, 393, 363, 394, 364, 395, 365,
+  396, 366, 397, 367, 398, 368, 399, 369, 400, 370, 401, 371, 402, 372, 403,
+  373, 404, 374, 405, 375, 406, 376, 407, 377, 408, 378, 409, 379, 410, 380,
+  411, 381, 412, 382, 413, 383, 414, 384, 384, 385, 416, 386, 417, 387, 418,
+  388, 419, 389, 420, 390, 421, 391, 422, 392, 423, 393, 424, 394, 425, 395,
+  426, 396, 427, 397, 428, 398, 429, 399, 430, 400, 431, 401, 432, 402, 433,
+  403, 434, 404, 435, 405, 436, 406, 437, 407, 438, 408, 439, 409, 440, 410,
+  441, 411, 442, 412, 443, 413, 444, 414, 445, 415, 446, 416, 416, 417, 448,
+  418, 449, 419, 450, 420, 451, 421, 452, 422, 453, 423, 454, 424, 455, 425,
+  456, 426, 457, 427, 458, 428, 459, 429, 460, 430, 461, 431, 462, 432, 463,
+  433, 464, 434, 465, 435, 466, 436, 467, 437, 468, 438, 469, 439, 470, 440,
+  471, 441, 472, 442, 473, 443, 474, 444, 475, 445, 476, 446, 477, 447, 478,
+  448, 448, 449, 480, 450, 481, 451, 482, 452, 483, 453, 484, 454, 485, 455,
+  486, 456, 487, 457, 488, 458, 489, 459, 490, 460, 491, 461, 492, 462, 493,
+  463, 494, 464, 495, 465, 496, 466, 497, 467, 498, 468, 499, 469, 500, 470,
+  501, 471, 502, 472, 503, 473, 504, 474, 505, 475, 506, 476, 507, 477, 508,
+  478, 509, 479, 510, 0,   0
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   16,  16,  32,  32,  48,  48,  64,  64,  80,  80,  96,
+  96,  112, 112, 128, 128, 144, 144, 160, 160, 176, 176, 192, 192, 208, 208,
+  224, 224, 0,   0,   1,   16,  17,  32,  33,  48,  49,  64,  65,  80,  81,
+  96,  97,  112, 113, 128, 129, 144, 145, 160, 161, 176, 177, 192, 193, 208,
+  209, 224, 225, 240, 1,   1,   2,   17,  18,  33,  34,  49,  50,  65,  66,
+  81,  82,  97,  98,  113, 114, 129, 130, 145, 146, 161, 162, 177, 178, 193,
+  194, 209, 210, 225, 226, 241, 2,   2,   3,   18,  19,  34,  35,  50,  51,
+  66,  67,  82,  83,  98,  99,  114, 115, 130, 131, 146, 147, 162, 163, 178,
+  179, 194, 195, 210, 211, 226, 227, 242, 3,   3,   4,   19,  20,  35,  36,
+  51,  52,  67,  68,  83,  84,  99,  100, 115, 116, 131, 132, 147, 148, 163,
+  164, 179, 180, 195, 196, 211, 212, 227, 228, 243, 4,   4,   5,   20,  21,
+  36,  37,  52,  53,  68,  69,  84,  85,  100, 101, 116, 117, 132, 133, 148,
+  149, 164, 165, 180, 181, 196, 197, 212, 213, 228, 229, 244, 5,   5,   6,
+  21,  22,  37,  38,  53,  54,  69,  70,  85,  86,  101, 102, 117, 118, 133,
+  134, 149, 150, 165, 166, 181, 182, 197, 198, 213, 214, 229, 230, 245, 6,
+  6,   7,   22,  23,  38,  39,  54,  55,  70,  71,  86,  87,  102, 103, 118,
+  119, 134, 135, 150, 151, 166, 167, 182, 183, 198, 199, 214, 215, 230, 231,
+  246, 7,   7,   8,   23,  24,  39,  40,  55,  56,  71,  72,  87,  88,  103,
+  104, 119, 120, 135, 136, 151, 152, 167, 168, 183, 184, 199, 200, 215, 216,
+  231, 232, 247, 8,   8,   9,   24,  25,  40,  41,  56,  57,  72,  73,  88,
+  89,  104, 105, 120, 121, 136, 137, 152, 153, 168, 169, 184, 185, 200, 201,
+  216, 217, 232, 233, 248, 9,   9,   10,  25,  26,  41,  42,  57,  58,  73,
+  74,  89,  90,  105, 106, 121, 122, 137, 138, 153, 154, 169, 170, 185, 186,
+  201, 202, 217, 218, 233, 234, 249, 10,  10,  11,  26,  27,  42,  43,  58,
+  59,  74,  75,  90,  91,  106, 107, 122, 123, 138, 139, 154, 155, 170, 171,
+  186, 187, 202, 203, 218, 219, 234, 235, 250, 11,  11,  12,  27,  28,  43,
+  44,  59,  60,  75,  76,  91,  92,  107, 108, 123, 124, 139, 140, 155, 156,
+  171, 172, 187, 188, 203, 204, 219, 220, 235, 236, 251, 12,  12,  13,  28,
+  29,  44,  45,  60,  61,  76,  77,  92,  93,  108, 109, 124, 125, 140, 141,
+  156, 157, 172, 173, 188, 189, 204, 205, 220, 221, 236, 237, 252, 13,  13,
+  14,  29,  30,  45,  46,  61,  62,  77,  78,  93,  94,  109, 110, 125, 126,
+  141, 142, 157, 158, 173, 174, 189, 190, 205, 206, 221, 222, 237, 238, 253,
+  14,  14,  15,  30,  31,  46,  47,  62,  63,  78,  79,  94,  95,  110, 111,
+  126, 127, 142, 143, 158, 159, 174, 175, 190, 191, 206, 207, 222, 223, 238,
+  239, 254, 0,   0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   1,   1,   2,   2,   3,   3,   4,   4,   5,   5,   6,
+  6,   7,   7,   8,   8,   9,   9,   10,  10,  11,  11,  12,  12,  13,  13,
+  14,  14,  0,   0,   1,   16,  2,   17,  3,   18,  4,   19,  5,   20,  6,
+  21,  7,   22,  8,   23,  9,   24,  10,  25,  11,  26,  12,  27,  13,  28,
+  14,  29,  15,  30,  16,  16,  17,  32,  18,  33,  19,  34,  20,  35,  21,
+  36,  22,  37,  23,  38,  24,  39,  25,  40,  26,  41,  27,  42,  28,  43,
+  29,  44,  30,  45,  31,  46,  32,  32,  33,  48,  34,  49,  35,  50,  36,
+  51,  37,  52,  38,  53,  39,  54,  40,  55,  41,  56,  42,  57,  43,  58,
+  44,  59,  45,  60,  46,  61,  47,  62,  48,  48,  49,  64,  50,  65,  51,
+  66,  52,  67,  53,  68,  54,  69,  55,  70,  56,  71,  57,  72,  58,  73,
+  59,  74,  60,  75,  61,  76,  62,  77,  63,  78,  64,  64,  65,  80,  66,
+  81,  67,  82,  68,  83,  69,  84,  70,  85,  71,  86,  72,  87,  73,  88,
+  74,  89,  75,  90,  76,  91,  77,  92,  78,  93,  79,  94,  80,  80,  81,
+  96,  82,  97,  83,  98,  84,  99,  85,  100, 86,  101, 87,  102, 88,  103,
+  89,  104, 90,  105, 91,  106, 92,  107, 93,  108, 94,  109, 95,  110, 96,
+  96,  97,  112, 98,  113, 99,  114, 100, 115, 101, 116, 102, 117, 103, 118,
+  104, 119, 105, 120, 106, 121, 107, 122, 108, 123, 109, 124, 110, 125, 111,
+  126, 112, 112, 113, 128, 114, 129, 115, 130, 116, 131, 117, 132, 118, 133,
+  119, 134, 120, 135, 121, 136, 122, 137, 123, 138, 124, 139, 125, 140, 126,
+  141, 127, 142, 128, 128, 129, 144, 130, 145, 131, 146, 132, 147, 133, 148,
+  134, 149, 135, 150, 136, 151, 137, 152, 138, 153, 139, 154, 140, 155, 141,
+  156, 142, 157, 143, 158, 144, 144, 145, 160, 146, 161, 147, 162, 148, 163,
+  149, 164, 150, 165, 151, 166, 152, 167, 153, 168, 154, 169, 155, 170, 156,
+  171, 157, 172, 158, 173, 159, 174, 160, 160, 161, 176, 162, 177, 163, 178,
+  164, 179, 165, 180, 166, 181, 167, 182, 168, 183, 169, 184, 170, 185, 171,
+  186, 172, 187, 173, 188, 174, 189, 175, 190, 176, 176, 177, 192, 178, 193,
+  179, 194, 180, 195, 181, 196, 182, 197, 183, 198, 184, 199, 185, 200, 186,
+  201, 187, 202, 188, 203, 189, 204, 190, 205, 191, 206, 192, 192, 193, 208,
+  194, 209, 195, 210, 196, 211, 197, 212, 198, 213, 199, 214, 200, 215, 201,
+  216, 202, 217, 203, 218, 204, 219, 205, 220, 206, 221, 207, 222, 208, 208,
+  209, 224, 210, 225, 211, 226, 212, 227, 213, 228, 214, 229, 215, 230, 216,
+  231, 217, 232, 218, 233, 219, 234, 220, 235, 221, 236, 222, 237, 223, 238,
+  224, 224, 225, 240, 226, 241, 227, 242, 228, 243, 229, 244, 230, 245, 231,
+  246, 232, 247, 233, 248, 234, 249, 235, 250, 236, 251, 237, 252, 238, 253,
+  239, 254, 0,   0,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t,
+                col_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   16,  16,  32,  32,  16,  0,   48,  48,  1,   16,  64,
+  64,  17,  32,  80,  80,  33,  48,  17,  1,   49,  64,  96,  96,  2,   17,
+  65,  80,  18,  33,  112, 112, 34,  49,  81,  96,  18,  2,   50,  65,  128,
+  128, 3,   18,  97,  112, 19,  34,  66,  81,  144, 144, 82,  97,  35,  50,
+  113, 128, 19,  3,   51,  66,  160, 160, 4,   19,  98,  113, 129, 144, 67,
+  82,  20,  35,  83,  98,  114, 129, 36,  51,  176, 176, 20,  4,   145, 160,
+  52,  67,  99,  114, 5,   20,  130, 145, 68,  83,  192, 192, 161, 176, 21,
+  36,  115, 130, 84,  99,  37,  52,  146, 161, 208, 208, 53,  68,  21,  5,
+  100, 115, 177, 192, 131, 146, 69,  84,  6,   21,  224, 224, 116, 131, 22,
+  37,  162, 177, 85,  100, 147, 162, 38,  53,  193, 208, 101, 116, 54,  69,
+  22,  6,   132, 147, 178, 193, 70,  85,  163, 178, 209, 224, 7,   22,  117,
+  132, 23,  38,  148, 163, 23,  7,   86,  101, 194, 209, 225, 240, 39,  54,
+  179, 194, 102, 117, 133, 148, 55,  70,  164, 179, 8,   23,  71,  86,  210,
+  225, 118, 133, 149, 164, 195, 210, 24,  39,  87,  102, 40,  55,  56,  71,
+  134, 149, 180, 195, 226, 241, 103, 118, 24,  8,   165, 180, 211, 226, 72,
+  87,  150, 165, 9,   24,  119, 134, 25,  40,  88,  103, 196, 211, 41,  56,
+  135, 150, 181, 196, 104, 119, 57,  72,  227, 242, 166, 181, 120, 135, 151,
+  166, 197, 212, 73,  88,  25,  9,   212, 227, 89,  104, 136, 151, 182, 197,
+  10,  25,  26,  41,  105, 120, 167, 182, 228, 243, 152, 167, 42,  57,  121,
+  136, 213, 228, 58,  73,  198, 213, 74,  89,  137, 152, 183, 198, 168, 183,
+  26,  10,  90,  105, 229, 244, 11,  26,  106, 121, 214, 229, 153, 168, 27,
+  42,  199, 214, 43,  58,  184, 199, 122, 137, 169, 184, 230, 245, 59,  74,
+  27,  11,  75,  90,  138, 153, 200, 215, 215, 230, 91,  106, 12,  27,  28,
+  43,  185, 200, 107, 122, 154, 169, 44,  59,  231, 246, 216, 231, 60,  75,
+  123, 138, 28,  12,  76,  91,  201, 216, 170, 185, 232, 247, 139, 154, 92,
+  107, 13,  28,  108, 123, 29,  44,  186, 201, 217, 232, 155, 170, 45,  60,
+  29,  13,  61,  76,  124, 139, 14,  14,  233, 248, 77,  92,  14,  29,  171,
+  186, 140, 155, 202, 217, 30,  45,  93,  108, 109, 124, 46,  61,  156, 171,
+  62,  77,  187, 202, 15,  30,  125, 140, 218, 233, 78,  93,  31,  46,  172,
+  187, 47,  62,  141, 156, 94,  109, 234, 249, 203, 218, 63,  78,  110, 125,
+  188, 203, 157, 172, 126, 141, 79,  94,  173, 188, 95,  110, 219, 234, 142,
+  157, 204, 219, 235, 250, 111, 126, 158, 173, 127, 142, 189, 204, 220, 235,
+  143, 158, 174, 189, 205, 220, 236, 251, 159, 174, 190, 205, 221, 236, 175,
+  190, 237, 252, 206, 221, 222, 237, 191, 206, 238, 253, 207, 222, 223, 238,
+  239, 254, 0,   0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                row_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   1,   1,   0,   1,   2,   2,   1,   16,  3,   3,   2,
+  17,  16,  17,  4,   4,   17,  32,  3,   18,  5,   5,   18,  33,  32,  33,
+  4,   19,  33,  48,  6,   6,   19,  34,  5,   20,  34,  49,  48,  49,  7,
+  7,   20,  35,  49,  64,  6,   21,  35,  50,  21,  36,  64,  65,  8,   8,
+  50,  65,  36,  51,  7,   22,  22,  37,  65,  80,  51,  66,  9,   9,   37,
+  52,  8,   23,  66,  81,  52,  67,  80,  81,  23,  38,  10,  10,  38,  53,
+  67,  82,  81,  96,  53,  68,  9,   24,  82,  97,  68,  83,  24,  39,  96,
+  97,  39,  54,  11,  11,  54,  69,  83,  98,  97,  112, 69,  84,  10,  25,
+  25,  40,  40,  55,  98,  113, 84,  99,  12,  12,  55,  70,  112, 113, 70,
+  85,  11,  26,  99,  114, 85,  100, 113, 128, 26,  41,  41,  56,  56,  71,
+  100, 115, 13,  13,  71,  86,  114, 129, 86,  101, 128, 129, 57,  72,  115,
+  130, 101, 116, 12,  27,  42,  57,  14,  14,  72,  87,  27,  42,  129, 144,
+  87,  102, 116, 131, 130, 145, 102, 117, 58,  73,  144, 145, 73,  88,  117,
+  132, 88,  103, 13,  28,  43,  58,  131, 146, 103, 118, 28,  43,  145, 160,
+  132, 147, 74,  89,  89,  104, 118, 133, 146, 161, 104, 119, 160, 161, 59,
+  74,  119, 134, 133, 148, 14,  29,  44,  59,  147, 162, 161, 176, 29,  44,
+  105, 120, 75,  90,  90,  105, 148, 163, 162, 177, 134, 149, 176, 177, 120,
+  135, 149, 164, 163, 178, 15,  30,  135, 150, 177, 192, 60,  75,  106, 121,
+  45,  60,  121, 136, 178, 193, 91,  106, 136, 151, 164, 179, 192, 193, 30,
+  45,  150, 165, 151, 166, 179, 194, 76,  91,  165, 180, 122, 137, 193, 208,
+  107, 122, 137, 152, 208, 209, 180, 195, 61,  76,  152, 167, 194, 209, 166,
+  181, 224, 224, 92,  107, 181, 196, 46,  61,  138, 153, 209, 224, 167, 182,
+  153, 168, 195, 210, 31,  46,  123, 138, 77,  92,  168, 183, 210, 225, 196,
+  211, 225, 240, 182, 197, 154, 169, 108, 123, 139, 154, 183, 198, 62,  77,
+  197, 212, 169, 184, 93,  108, 211, 226, 184, 199, 47,  62,  212, 227, 226,
+  241, 124, 139, 198, 213, 155, 170, 170, 185, 140, 155, 213, 228, 227, 242,
+  109, 124, 78,  93,  185, 200, 228, 243, 199, 214, 200, 215, 214, 229, 125,
+  140, 171, 186, 186, 201, 63,  78,  156, 171, 94,  109, 141, 156, 229, 244,
+  201, 216, 215, 230, 79,  94,  230, 245, 216, 231, 110, 125, 187, 202, 231,
+  246, 217, 232, 157, 172, 202, 217, 126, 141, 95,  110, 142, 157, 172, 187,
+  232, 247, 111, 126, 218, 233, 203, 218, 233, 248, 173, 188, 188, 203, 127,
+  142, 158, 173, 143, 158, 234, 249, 219, 234, 189, 204, 204, 219, 159, 174,
+  174, 189, 235, 250, 205, 220, 175, 190, 190, 205, 220, 235, 191, 206, 221,
+  236, 236, 251, 206, 221, 237, 252, 207, 222, 222, 237, 223, 238, 238, 253,
+  239, 254, 0,   0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_16x16_neighbors[257 * MAX_NEIGHBORS]) = {
+  0,   0,   0,   0,   16,  0,   16,  16,  1,   16,  17,  1,   32,  32,  17,
+  32,  2,   17,  18,  2,   48,  48,  18,  33,  33,  48,  3,   18,  49,  64,
+  64,  65,  34,  49,  19,  3,   19,  34,  50,  65,  4,   19,  65,  80,  80,
+  81,  35,  50,  20,  4,   20,  35,  66,  81,  81,  96,  51,  66,  96,  97,
+  5,   20,  36,  51,  82,  97,  21,  36,  67,  82,  97,  112, 21,  5,   52,
+  67,  112, 113, 37,  52,  6,   21,  83,  98,  98,  113, 68,  83,  22,  6,
+  113, 128, 22,  37,  53,  68,  84,  99,  99,  114, 128, 129, 114, 129, 69,
+  84,  38,  53,  7,   22,  23,  7,   129, 144, 23,  38,  54,  69,  100, 115,
+  85,  100, 115, 130, 144, 145, 130, 145, 39,  54,  70,  85,  8,   23,  55,
+  70,  116, 131, 101, 116, 145, 160, 24,  39,  24,  8,   86,  101, 131, 146,
+  160, 161, 146, 161, 71,  86,  40,  55,  9,   24,  117, 132, 102, 117, 161,
+  176, 132, 147, 56,  71,  87,  102, 25,  40,  147, 162, 25,  9,   176, 177,
+  162, 177, 72,  87,  41,  56,  118, 133, 133, 148, 103, 118, 10,  25,  148,
+  163, 57,  72,  88,  103, 177, 192, 26,  41,  163, 178, 192, 193, 26,  10,
+  119, 134, 73,  88,  149, 164, 104, 119, 134, 149, 42,  57,  178, 193, 164,
+  179, 11,  26,  58,  73,  193, 208, 89,  104, 135, 150, 120, 135, 27,  42,
+  74,  89,  208, 209, 150, 165, 179, 194, 165, 180, 105, 120, 194, 209, 43,
+  58,  27,  11,  136, 151, 90,  105, 151, 166, 180, 195, 59,  74,  121, 136,
+  209, 224, 195, 210, 224, 225, 166, 181, 106, 121, 75,  90,  12,  27,  181,
+  196, 28,  12,  210, 225, 152, 167, 167, 182, 137, 152, 28,  43,  196, 211,
+  122, 137, 91,  106, 225, 240, 44,  59,  13,  28,  107, 122, 182, 197, 168,
+  183, 211, 226, 153, 168, 226, 241, 60,  75,  197, 212, 138, 153, 29,  44,
+  76,  91,  29,  13,  183, 198, 123, 138, 45,  60,  212, 227, 198, 213, 154,
+  169, 169, 184, 227, 242, 92,  107, 61,  76,  139, 154, 14,  29,  30,  14,
+  184, 199, 213, 228, 108, 123, 199, 214, 228, 243, 77,  92,  30,  45,  170,
+  185, 155, 170, 185, 200, 93,  108, 124, 139, 214, 229, 46,  61,  200, 215,
+  229, 244, 15,  30,  109, 124, 62,  77,  140, 155, 215, 230, 31,  46,  171,
+  186, 186, 201, 201, 216, 78,  93,  230, 245, 125, 140, 47,  62,  216, 231,
+  156, 171, 94,  109, 231, 246, 141, 156, 63,  78,  202, 217, 187, 202, 110,
+  125, 217, 232, 172, 187, 232, 247, 79,  94,  157, 172, 126, 141, 203, 218,
+  95,  110, 233, 248, 218, 233, 142, 157, 111, 126, 173, 188, 188, 203, 234,
+  249, 219, 234, 127, 142, 158, 173, 204, 219, 189, 204, 143, 158, 235, 250,
+  174, 189, 205, 220, 159, 174, 220, 235, 221, 236, 175, 190, 190, 205, 236,
+  251, 206, 221, 237, 252, 191, 206, 222, 237, 207, 222, 238, 253, 223, 238,
+  239, 254, 0,   0,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t,
+                mcol_scan_32x32_neighbors[1025 * MAX_NEIGHBORS]) = {
+  0,   0,    0,   0,    32,  32,   64,  64,   96,  96,   128, 128,  160, 160,
+  192, 192,  224, 224,  256, 256,  288, 288,  320, 320,  352, 352,  384, 384,
+  416, 416,  448, 448,  480, 480,  512, 512,  544, 544,  576, 576,  608, 608,
+  640, 640,  672, 672,  704, 704,  736, 736,  768, 768,  800, 800,  832, 832,
+  864, 864,  896, 896,  928, 928,  960, 960,  0,   0,    1,   32,   33,  64,
+  65,  96,   97,  128,  129, 160,  161, 192,  193, 224,  225, 256,  257, 288,
+  289, 320,  321, 352,  353, 384,  385, 416,  417, 448,  449, 480,  481, 512,
+  513, 544,  545, 576,  577, 608,  609, 640,  641, 672,  673, 704,  705, 736,
+  737, 768,  769, 800,  801, 832,  833, 864,  865, 896,  897, 928,  929, 960,
+  961, 992,  1,   1,    2,   33,   34,  65,   66,  97,   98,  129,  130, 161,
+  162, 193,  194, 225,  226, 257,  258, 289,  290, 321,  322, 353,  354, 385,
+  386, 417,  418, 449,  450, 481,  482, 513,  514, 545,  546, 577,  578, 609,
+  610, 641,  642, 673,  674, 705,  706, 737,  738, 769,  770, 801,  802, 833,
+  834, 865,  866, 897,  898, 929,  930, 961,  962, 993,  2,   2,    3,   34,
+  35,  66,   67,  98,   99,  130,  131, 162,  163, 194,  195, 226,  227, 258,
+  259, 290,  291, 322,  323, 354,  355, 386,  387, 418,  419, 450,  451, 482,
+  483, 514,  515, 546,  547, 578,  579, 610,  611, 642,  643, 674,  675, 706,
+  707, 738,  739, 770,  771, 802,  803, 834,  835, 866,  867, 898,  899, 930,
+  931, 962,  963, 994,  3,   3,    4,   35,   36,  67,   68,  99,   100, 131,
+  132, 163,  164, 195,  196, 227,  228, 259,  260, 291,  292, 323,  324, 355,
+  356, 387,  388, 419,  420, 451,  452, 483,  484, 515,  516, 547,  548, 579,
+  580, 611,  612, 643,  644, 675,  676, 707,  708, 739,  740, 771,  772, 803,
+  804, 835,  836, 867,  868, 899,  900, 931,  932, 963,  964, 995,  4,   4,
+  5,   36,   37,  68,   69,  100,  101, 132,  133, 164,  165, 196,  197, 228,
+  229, 260,  261, 292,  293, 324,  325, 356,  357, 388,  389, 420,  421, 452,
+  453, 484,  485, 516,  517, 548,  549, 580,  581, 612,  613, 644,  645, 676,
+  677, 708,  709, 740,  741, 772,  773, 804,  805, 836,  837, 868,  869, 900,
+  901, 932,  933, 964,  965, 996,  5,   5,    6,   37,   38,  69,   70,  101,
+  102, 133,  134, 165,  166, 197,  198, 229,  230, 261,  262, 293,  294, 325,
+  326, 357,  358, 389,  390, 421,  422, 453,  454, 485,  486, 517,  518, 549,
+  550, 581,  582, 613,  614, 645,  646, 677,  678, 709,  710, 741,  742, 773,
+  774, 805,  806, 837,  838, 869,  870, 901,  902, 933,  934, 965,  966, 997,
+  6,   6,    7,   38,   39,  70,   71,  102,  103, 134,  135, 166,  167, 198,
+  199, 230,  231, 262,  263, 294,  295, 326,  327, 358,  359, 390,  391, 422,
+  423, 454,  455, 486,  487, 518,  519, 550,  551, 582,  583, 614,  615, 646,
+  647, 678,  679, 710,  711, 742,  743, 774,  775, 806,  807, 838,  839, 870,
+  871, 902,  903, 934,  935, 966,  967, 998,  7,   7,    8,   39,   40,  71,
+  72,  103,  104, 135,  136, 167,  168, 199,  200, 231,  232, 263,  264, 295,
+  296, 327,  328, 359,  360, 391,  392, 423,  424, 455,  456, 487,  488, 519,
+  520, 551,  552, 583,  584, 615,  616, 647,  648, 679,  680, 711,  712, 743,
+  744, 775,  776, 807,  808, 839,  840, 871,  872, 903,  904, 935,  936, 967,
+  968, 999,  8,   8,    9,   40,   41,  72,   73,  104,  105, 136,  137, 168,
+  169, 200,  201, 232,  233, 264,  265, 296,  297, 328,  329, 360,  361, 392,
+  393, 424,  425, 456,  457, 488,  489, 520,  521, 552,  553, 584,  585, 616,
+  617, 648,  649, 680,  681, 712,  713, 744,  745, 776,  777, 808,  809, 840,
+  841, 872,  873, 904,  905, 936,  937, 968,  969, 1000, 9,   9,    10,  41,
+  42,  73,   74,  105,  106, 137,  138, 169,  170, 201,  202, 233,  234, 265,
+  266, 297,  298, 329,  330, 361,  362, 393,  394, 425,  426, 457,  458, 489,
+  490, 521,  522, 553,  554, 585,  586, 617,  618, 649,  650, 681,  682, 713,
+  714, 745,  746, 777,  778, 809,  810, 841,  842, 873,  874, 905,  906, 937,
+  938, 969,  970, 1001, 10,  10,   11,  42,   43,  74,   75,  106,  107, 138,
+  139, 170,  171, 202,  203, 234,  235, 266,  267, 298,  299, 330,  331, 362,
+  363, 394,  395, 426,  427, 458,  459, 490,  491, 522,  523, 554,  555, 586,
+  587, 618,  619, 650,  651, 682,  683, 714,  715, 746,  747, 778,  779, 810,
+  811, 842,  843, 874,  875, 906,  907, 938,  939, 970,  971, 1002, 11,  11,
+  12,  43,   44,  75,   76,  107,  108, 139,  140, 171,  172, 203,  204, 235,
+  236, 267,  268, 299,  300, 331,  332, 363,  364, 395,  396, 427,  428, 459,
+  460, 491,  492, 523,  524, 555,  556, 587,  588, 619,  620, 651,  652, 683,
+  684, 715,  716, 747,  748, 779,  780, 811,  812, 843,  844, 875,  876, 907,
+  908, 939,  940, 971,  972, 1003, 12,  12,   13,  44,   45,  76,   77,  108,
+  109, 140,  141, 172,  173, 204,  205, 236,  237, 268,  269, 300,  301, 332,
+  333, 364,  365, 396,  397, 428,  429, 460,  461, 492,  493, 524,  525, 556,
+  557, 588,  589, 620,  621, 652,  653, 684,  685, 716,  717, 748,  749, 780,
+  781, 812,  813, 844,  845, 876,  877, 908,  909, 940,  941, 972,  973, 1004,
+  13,  13,   14,  45,   46,  77,   78,  109,  110, 141,  142, 173,  174, 205,
+  206, 237,  238, 269,  270, 301,  302, 333,  334, 365,  366, 397,  398, 429,
+  430, 461,  462, 493,  494, 525,  526, 557,  558, 589,  590, 621,  622, 653,
+  654, 685,  686, 717,  718, 749,  750, 781,  782, 813,  814, 845,  846, 877,
+  878, 909,  910, 941,  942, 973,  974, 1005, 14,  14,   15,  46,   47,  78,
+  79,  110,  111, 142,  143, 174,  175, 206,  207, 238,  239, 270,  271, 302,
+  303, 334,  335, 366,  367, 398,  399, 430,  431, 462,  463, 494,  495, 526,
+  527, 558,  559, 590,  591, 622,  623, 654,  655, 686,  687, 718,  719, 750,
+  751, 782,  783, 814,  815, 846,  847, 878,  879, 910,  911, 942,  943, 974,
+  975, 1006, 15,  15,   16,  47,   48,  79,   80,  111,  112, 143,  144, 175,
+  176, 207,  208, 239,  240, 271,  272, 303,  304, 335,  336, 367,  368, 399,
+  400, 431,  432, 463,  464, 495,  496, 527,  528, 559,  560, 591,  592, 623,
+  624, 655,  656, 687,  688, 719,  720, 751,  752, 783,  784, 815,  816, 847,
+  848, 879,  880, 911,  912, 943,  944, 975,  976, 1007, 16,  16,   17,  48,
+  49,  80,   81,  112,  113, 144,  145, 176,  177, 208,  209, 240,  241, 272,
+  273, 304,  305, 336,  337, 368,  369, 400,  401, 432,  433, 464,  465, 496,
+  497, 528,  529, 560,  561, 592,  593, 624,  625, 656,  657, 688,  689, 720,
+  721, 752,  753, 784,  785, 816,  817, 848,  849, 880,  881, 912,  913, 944,
+  945, 976,  977, 1008, 17,  17,   18,  49,   50,  81,   82,  113,  114, 145,
+  146, 177,  178, 209,  210, 241,  242, 273,  274, 305,  306, 337,  338, 369,
+  370, 401,  402, 433,  434, 465,  466, 497,  498, 529,  530, 561,  562, 593,
+  594, 625,  626, 657,  658, 689,  690, 721,  722, 753,  754, 785,  786, 817,
+  818, 849,  850, 881,  882, 913,  914, 945,  946, 977,  978, 1009, 18,  18,
+  19,  50,   51,  82,   83,  114,  115, 146,  147, 178,  179, 210,  211, 242,
+  243, 274,  275, 306,  307, 338,  339, 370,  371, 402,  403, 434,  435, 466,
+  467, 498,  499, 530,  531, 562,  563, 594,  595, 626,  627, 658,  659, 690,
+  691, 722,  723, 754,  755, 786,  787, 818,  819, 850,  851, 882,  883, 914,
+  915, 946,  947, 978,  979, 1010, 19,  19,   20,  51,   52,  83,   84,  115,
+  116, 147,  148, 179,  180, 211,  212, 243,  244, 275,  276, 307,  308, 339,
+  340, 371,  372, 403,  404, 435,  436, 467,  468, 499,  500, 531,  532, 563,
+  564, 595,  596, 627,  628, 659,  660, 691,  692, 723,  724, 755,  756, 787,
+  788, 819,  820, 851,  852, 883,  884, 915,  916, 947,  948, 979,  980, 1011,
+  20,  20,   21,  52,   53,  84,   85,  116,  117, 148,  149, 180,  181, 212,
+  213, 244,  245, 276,  277, 308,  309, 340,  341, 372,  373, 404,  405, 436,
+  437, 468,  469, 500,  501, 532,  533, 564,  565, 596,  597, 628,  629, 660,
+  661, 692,  693, 724,  725, 756,  757, 788,  789, 820,  821, 852,  853, 884,
+  885, 916,  917, 948,  949, 980,  981, 1012, 21,  21,   22,  53,   54,  85,
+  86,  117,  118, 149,  150, 181,  182, 213,  214, 245,  246, 277,  278, 309,
+  310, 341,  342, 373,  374, 405,  406, 437,  438, 469,  470, 501,  502, 533,
+  534, 565,  566, 597,  598, 629,  630, 661,  662, 693,  694, 725,  726, 757,
+  758, 789,  790, 821,  822, 853,  854, 885,  886, 917,  918, 949,  950, 981,
+  982, 1013, 22,  22,   23,  54,   55,  86,   87,  118,  119, 150,  151, 182,
+  183, 214,  215, 246,  247, 278,  279, 310,  311, 342,  343, 374,  375, 406,
+  407, 438,  439, 470,  471, 502,  503, 534,  535, 566,  567, 598,  599, 630,
+  631, 662,  663, 694,  695, 726,  727, 758,  759, 790,  791, 822,  823, 854,
+  855, 886,  887, 918,  919, 950,  951, 982,  983, 1014, 23,  23,   24,  55,
+  56,  87,   88,  119,  120, 151,  152, 183,  184, 215,  216, 247,  248, 279,
+  280, 311,  312, 343,  344, 375,  376, 407,  408, 439,  440, 471,  472, 503,
+  504, 535,  536, 567,  568, 599,  600, 631,  632, 663,  664, 695,  696, 727,
+  728, 759,  760, 791,  792, 823,  824, 855,  856, 887,  888, 919,  920, 951,
+  952, 983,  984, 1015, 24,  24,   25,  56,   57,  88,   89,  120,  121, 152,
+  153, 184,  185, 216,  217, 248,  249, 280,  281, 312,  313, 344,  345, 376,
+  377, 408,  409, 440,  441, 472,  473, 504,  505, 536,  537, 568,  569, 600,
+  601, 632,  633, 664,  665, 696,  697, 728,  729, 760,  761, 792,  793, 824,
+  825, 856,  857, 888,  889, 920,  921, 952,  953, 984,  985, 1016, 25,  25,
+  26,  57,   58,  89,   90,  121,  122, 153,  154, 185,  186, 217,  218, 249,
+  250, 281,  282, 313,  314, 345,  346, 377,  378, 409,  410, 441,  442, 473,
+  474, 505,  506, 537,  538, 569,  570, 601,  602, 633,  634, 665,  666, 697,
+  698, 729,  730, 761,  762, 793,  794, 825,  826, 857,  858, 889,  890, 921,
+  922, 953,  954, 985,  986, 1017, 26,  26,   27,  58,   59,  90,   91,  122,
+  123, 154,  155, 186,  187, 218,  219, 250,  251, 282,  283, 314,  315, 346,
+  347, 378,  379, 410,  411, 442,  443, 474,  475, 506,  507, 538,  539, 570,
+  571, 602,  603, 634,  635, 666,  667, 698,  699, 730,  731, 762,  763, 794,
+  795, 826,  827, 858,  859, 890,  891, 922,  923, 954,  955, 986,  987, 1018,
+  27,  27,   28,  59,   60,  91,   92,  123,  124, 155,  156, 187,  188, 219,
+  220, 251,  252, 283,  284, 315,  316, 347,  348, 379,  380, 411,  412, 443,
+  444, 475,  476, 507,  508, 539,  540, 571,  572, 603,  604, 635,  636, 667,
+  668, 699,  700, 731,  732, 763,  764, 795,  796, 827,  828, 859,  860, 891,
+  892, 923,  924, 955,  956, 987,  988, 1019, 28,  28,   29,  60,   61,  92,
+  93,  124,  125, 156,  157, 188,  189, 220,  221, 252,  253, 284,  285, 316,
+  317, 348,  349, 380,  381, 412,  413, 444,  445, 476,  477, 508,  509, 540,
+  541, 572,  573, 604,  605, 636,  637, 668,  669, 700,  701, 732,  733, 764,
+  765, 796,  797, 828,  829, 860,  861, 892,  893, 924,  925, 956,  957, 988,
+  989, 1020, 29,  29,   30,  61,   62,  93,   94,  125,  126, 157,  158, 189,
+  190, 221,  222, 253,  254, 285,  286, 317,  318, 349,  350, 381,  382, 413,
+  414, 445,  446, 477,  478, 509,  510, 541,  542, 573,  574, 605,  606, 637,
+  638, 669,  670, 701,  702, 733,  734, 765,  766, 797,  798, 829,  830, 861,
+  862, 893,  894, 925,  926, 957,  958, 989,  990, 1021, 30,  30,   31,  62,
+  63,  94,   95,  126,  127, 158,  159, 190,  191, 222,  223, 254,  255, 286,
+  287, 318,  319, 350,  351, 382,  383, 414,  415, 446,  447, 478,  479, 510,
+  511, 542,  543, 574,  575, 606,  607, 638,  639, 670,  671, 702,  703, 734,
+  735, 766,  767, 798,  799, 830,  831, 862,  863, 894,  895, 926,  927, 958,
+  959, 990,  991, 1022, 0,   0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                mrow_scan_32x32_neighbors[1025 * MAX_NEIGHBORS]) = {
+  0,   0,    0,   0,    1,   1,    2,   2,    3,   3,    4,   4,    5,   5,
+  6,   6,    7,   7,    8,   8,    9,   9,    10,  10,   11,  11,   12,  12,
+  13,  13,   14,  14,   15,  15,   16,  16,   17,  17,   18,  18,   19,  19,
+  20,  20,   21,  21,   22,  22,   23,  23,   24,  24,   25,  25,   26,  26,
+  27,  27,   28,  28,   29,  29,   30,  30,   0,   0,    1,   32,   2,   33,
+  3,   34,   4,   35,   5,   36,   6,   37,   7,   38,   8,   39,   9,   40,
+  10,  41,   11,  42,   12,  43,   13,  44,   14,  45,   15,  46,   16,  47,
+  17,  48,   18,  49,   19,  50,   20,  51,   21,  52,   22,  53,   23,  54,
+  24,  55,   25,  56,   26,  57,   27,  58,   28,  59,   29,  60,   30,  61,
+  31,  62,   32,  32,   33,  64,   34,  65,   35,  66,   36,  67,   37,  68,
+  38,  69,   39,  70,   40,  71,   41,  72,   42,  73,   43,  74,   44,  75,
+  45,  76,   46,  77,   47,  78,   48,  79,   49,  80,   50,  81,   51,  82,
+  52,  83,   53,  84,   54,  85,   55,  86,   56,  87,   57,  88,   58,  89,
+  59,  90,   60,  91,   61,  92,   62,  93,   63,  94,   64,  64,   65,  96,
+  66,  97,   67,  98,   68,  99,   69,  100,  70,  101,  71,  102,  72,  103,
+  73,  104,  74,  105,  75,  106,  76,  107,  77,  108,  78,  109,  79,  110,
+  80,  111,  81,  112,  82,  113,  83,  114,  84,  115,  85,  116,  86,  117,
+  87,  118,  88,  119,  89,  120,  90,  121,  91,  122,  92,  123,  93,  124,
+  94,  125,  95,  126,  96,  96,   97,  128,  98,  129,  99,  130,  100, 131,
+  101, 132,  102, 133,  103, 134,  104, 135,  105, 136,  106, 137,  107, 138,
+  108, 139,  109, 140,  110, 141,  111, 142,  112, 143,  113, 144,  114, 145,
+  115, 146,  116, 147,  117, 148,  118, 149,  119, 150,  120, 151,  121, 152,
+  122, 153,  123, 154,  124, 155,  125, 156,  126, 157,  127, 158,  128, 128,
+  129, 160,  130, 161,  131, 162,  132, 163,  133, 164,  134, 165,  135, 166,
+  136, 167,  137, 168,  138, 169,  139, 170,  140, 171,  141, 172,  142, 173,
+  143, 174,  144, 175,  145, 176,  146, 177,  147, 178,  148, 179,  149, 180,
+  150, 181,  151, 182,  152, 183,  153, 184,  154, 185,  155, 186,  156, 187,
+  157, 188,  158, 189,  159, 190,  160, 160,  161, 192,  162, 193,  163, 194,
+  164, 195,  165, 196,  166, 197,  167, 198,  168, 199,  169, 200,  170, 201,
+  171, 202,  172, 203,  173, 204,  174, 205,  175, 206,  176, 207,  177, 208,
+  178, 209,  179, 210,  180, 211,  181, 212,  182, 213,  183, 214,  184, 215,
+  185, 216,  186, 217,  187, 218,  188, 219,  189, 220,  190, 221,  191, 222,
+  192, 192,  193, 224,  194, 225,  195, 226,  196, 227,  197, 228,  198, 229,
+  199, 230,  200, 231,  201, 232,  202, 233,  203, 234,  204, 235,  205, 236,
+  206, 237,  207, 238,  208, 239,  209, 240,  210, 241,  211, 242,  212, 243,
+  213, 244,  214, 245,  215, 246,  216, 247,  217, 248,  218, 249,  219, 250,
+  220, 251,  221, 252,  222, 253,  223, 254,  224, 224,  225, 256,  226, 257,
+  227, 258,  228, 259,  229, 260,  230, 261,  231, 262,  232, 263,  233, 264,
+  234, 265,  235, 266,  236, 267,  237, 268,  238, 269,  239, 270,  240, 271,
+  241, 272,  242, 273,  243, 274,  244, 275,  245, 276,  246, 277,  247, 278,
+  248, 279,  249, 280,  250, 281,  251, 282,  252, 283,  253, 284,  254, 285,
+  255, 286,  256, 256,  257, 288,  258, 289,  259, 290,  260, 291,  261, 292,
+  262, 293,  263, 294,  264, 295,  265, 296,  266, 297,  267, 298,  268, 299,
+  269, 300,  270, 301,  271, 302,  272, 303,  273, 304,  274, 305,  275, 306,
+  276, 307,  277, 308,  278, 309,  279, 310,  280, 311,  281, 312,  282, 313,
+  283, 314,  284, 315,  285, 316,  286, 317,  287, 318,  288, 288,  289, 320,
+  290, 321,  291, 322,  292, 323,  293, 324,  294, 325,  295, 326,  296, 327,
+  297, 328,  298, 329,  299, 330,  300, 331,  301, 332,  302, 333,  303, 334,
+  304, 335,  305, 336,  306, 337,  307, 338,  308, 339,  309, 340,  310, 341,
+  311, 342,  312, 343,  313, 344,  314, 345,  315, 346,  316, 347,  317, 348,
+  318, 349,  319, 350,  320, 320,  321, 352,  322, 353,  323, 354,  324, 355,
+  325, 356,  326, 357,  327, 358,  328, 359,  329, 360,  330, 361,  331, 362,
+  332, 363,  333, 364,  334, 365,  335, 366,  336, 367,  337, 368,  338, 369,
+  339, 370,  340, 371,  341, 372,  342, 373,  343, 374,  344, 375,  345, 376,
+  346, 377,  347, 378,  348, 379,  349, 380,  350, 381,  351, 382,  352, 352,
+  353, 384,  354, 385,  355, 386,  356, 387,  357, 388,  358, 389,  359, 390,
+  360, 391,  361, 392,  362, 393,  363, 394,  364, 395,  365, 396,  366, 397,
+  367, 398,  368, 399,  369, 400,  370, 401,  371, 402,  372, 403,  373, 404,
+  374, 405,  375, 406,  376, 407,  377, 408,  378, 409,  379, 410,  380, 411,
+  381, 412,  382, 413,  383, 414,  384, 384,  385, 416,  386, 417,  387, 418,
+  388, 419,  389, 420,  390, 421,  391, 422,  392, 423,  393, 424,  394, 425,
+  395, 426,  396, 427,  397, 428,  398, 429,  399, 430,  400, 431,  401, 432,
+  402, 433,  403, 434,  404, 435,  405, 436,  406, 437,  407, 438,  408, 439,
+  409, 440,  410, 441,  411, 442,  412, 443,  413, 444,  414, 445,  415, 446,
+  416, 416,  417, 448,  418, 449,  419, 450,  420, 451,  421, 452,  422, 453,
+  423, 454,  424, 455,  425, 456,  426, 457,  427, 458,  428, 459,  429, 460,
+  430, 461,  431, 462,  432, 463,  433, 464,  434, 465,  435, 466,  436, 467,
+  437, 468,  438, 469,  439, 470,  440, 471,  441, 472,  442, 473,  443, 474,
+  444, 475,  445, 476,  446, 477,  447, 478,  448, 448,  449, 480,  450, 481,
+  451, 482,  452, 483,  453, 484,  454, 485,  455, 486,  456, 487,  457, 488,
+  458, 489,  459, 490,  460, 491,  461, 492,  462, 493,  463, 494,  464, 495,
+  465, 496,  466, 497,  467, 498,  468, 499,  469, 500,  470, 501,  471, 502,
+  472, 503,  473, 504,  474, 505,  475, 506,  476, 507,  477, 508,  478, 509,
+  479, 510,  480, 480,  481, 512,  482, 513,  483, 514,  484, 515,  485, 516,
+  486, 517,  487, 518,  488, 519,  489, 520,  490, 521,  491, 522,  492, 523,
+  493, 524,  494, 525,  495, 526,  496, 527,  497, 528,  498, 529,  499, 530,
+  500, 531,  501, 532,  502, 533,  503, 534,  504, 535,  505, 536,  506, 537,
+  507, 538,  508, 539,  509, 540,  510, 541,  511, 542,  512, 512,  513, 544,
+  514, 545,  515, 546,  516, 547,  517, 548,  518, 549,  519, 550,  520, 551,
+  521, 552,  522, 553,  523, 554,  524, 555,  525, 556,  526, 557,  527, 558,
+  528, 559,  529, 560,  530, 561,  531, 562,  532, 563,  533, 564,  534, 565,
+  535, 566,  536, 567,  537, 568,  538, 569,  539, 570,  540, 571,  541, 572,
+  542, 573,  543, 574,  544, 544,  545, 576,  546, 577,  547, 578,  548, 579,
+  549, 580,  550, 581,  551, 582,  552, 583,  553, 584,  554, 585,  555, 586,
+  556, 587,  557, 588,  558, 589,  559, 590,  560, 591,  561, 592,  562, 593,
+  563, 594,  564, 595,  565, 596,  566, 597,  567, 598,  568, 599,  569, 600,
+  570, 601,  571, 602,  572, 603,  573, 604,  574, 605,  575, 606,  576, 576,
+  577, 608,  578, 609,  579, 610,  580, 611,  581, 612,  582, 613,  583, 614,
+  584, 615,  585, 616,  586, 617,  587, 618,  588, 619,  589, 620,  590, 621,
+  591, 622,  592, 623,  593, 624,  594, 625,  595, 626,  596, 627,  597, 628,
+  598, 629,  599, 630,  600, 631,  601, 632,  602, 633,  603, 634,  604, 635,
+  605, 636,  606, 637,  607, 638,  608, 608,  609, 640,  610, 641,  611, 642,
+  612, 643,  613, 644,  614, 645,  615, 646,  616, 647,  617, 648,  618, 649,
+  619, 650,  620, 651,  621, 652,  622, 653,  623, 654,  624, 655,  625, 656,
+  626, 657,  627, 658,  628, 659,  629, 660,  630, 661,  631, 662,  632, 663,
+  633, 664,  634, 665,  635, 666,  636, 667,  637, 668,  638, 669,  639, 670,
+  640, 640,  641, 672,  642, 673,  643, 674,  644, 675,  645, 676,  646, 677,
+  647, 678,  648, 679,  649, 680,  650, 681,  651, 682,  652, 683,  653, 684,
+  654, 685,  655, 686,  656, 687,  657, 688,  658, 689,  659, 690,  660, 691,
+  661, 692,  662, 693,  663, 694,  664, 695,  665, 696,  666, 697,  667, 698,
+  668, 699,  669, 700,  670, 701,  671, 702,  672, 672,  673, 704,  674, 705,
+  675, 706,  676, 707,  677, 708,  678, 709,  679, 710,  680, 711,  681, 712,
+  682, 713,  683, 714,  684, 715,  685, 716,  686, 717,  687, 718,  688, 719,
+  689, 720,  690, 721,  691, 722,  692, 723,  693, 724,  694, 725,  695, 726,
+  696, 727,  697, 728,  698, 729,  699, 730,  700, 731,  701, 732,  702, 733,
+  703, 734,  704, 704,  705, 736,  706, 737,  707, 738,  708, 739,  709, 740,
+  710, 741,  711, 742,  712, 743,  713, 744,  714, 745,  715, 746,  716, 747,
+  717, 748,  718, 749,  719, 750,  720, 751,  721, 752,  722, 753,  723, 754,
+  724, 755,  725, 756,  726, 757,  727, 758,  728, 759,  729, 760,  730, 761,
+  731, 762,  732, 763,  733, 764,  734, 765,  735, 766,  736, 736,  737, 768,
+  738, 769,  739, 770,  740, 771,  741, 772,  742, 773,  743, 774,  744, 775,
+  745, 776,  746, 777,  747, 778,  748, 779,  749, 780,  750, 781,  751, 782,
+  752, 783,  753, 784,  754, 785,  755, 786,  756, 787,  757, 788,  758, 789,
+  759, 790,  760, 791,  761, 792,  762, 793,  763, 794,  764, 795,  765, 796,
+  766, 797,  767, 798,  768, 768,  769, 800,  770, 801,  771, 802,  772, 803,
+  773, 804,  774, 805,  775, 806,  776, 807,  777, 808,  778, 809,  779, 810,
+  780, 811,  781, 812,  782, 813,  783, 814,  784, 815,  785, 816,  786, 817,
+  787, 818,  788, 819,  789, 820,  790, 821,  791, 822,  792, 823,  793, 824,
+  794, 825,  795, 826,  796, 827,  797, 828,  798, 829,  799, 830,  800, 800,
+  801, 832,  802, 833,  803, 834,  804, 835,  805, 836,  806, 837,  807, 838,
+  808, 839,  809, 840,  810, 841,  811, 842,  812, 843,  813, 844,  814, 845,
+  815, 846,  816, 847,  817, 848,  818, 849,  819, 850,  820, 851,  821, 852,
+  822, 853,  823, 854,  824, 855,  825, 856,  826, 857,  827, 858,  828, 859,
+  829, 860,  830, 861,  831, 862,  832, 832,  833, 864,  834, 865,  835, 866,
+  836, 867,  837, 868,  838, 869,  839, 870,  840, 871,  841, 872,  842, 873,
+  843, 874,  844, 875,  845, 876,  846, 877,  847, 878,  848, 879,  849, 880,
+  850, 881,  851, 882,  852, 883,  853, 884,  854, 885,  855, 886,  856, 887,
+  857, 888,  858, 889,  859, 890,  860, 891,  861, 892,  862, 893,  863, 894,
+  864, 864,  865, 896,  866, 897,  867, 898,  868, 899,  869, 900,  870, 901,
+  871, 902,  872, 903,  873, 904,  874, 905,  875, 906,  876, 907,  877, 908,
+  878, 909,  879, 910,  880, 911,  881, 912,  882, 913,  883, 914,  884, 915,
+  885, 916,  886, 917,  887, 918,  888, 919,  889, 920,  890, 921,  891, 922,
+  892, 923,  893, 924,  894, 925,  895, 926,  896, 896,  897, 928,  898, 929,
+  899, 930,  900, 931,  901, 932,  902, 933,  903, 934,  904, 935,  905, 936,
+  906, 937,  907, 938,  908, 939,  909, 940,  910, 941,  911, 942,  912, 943,
+  913, 944,  914, 945,  915, 946,  916, 947,  917, 948,  918, 949,  919, 950,
+  920, 951,  921, 952,  922, 953,  923, 954,  924, 955,  925, 956,  926, 957,
+  927, 958,  928, 928,  929, 960,  930, 961,  931, 962,  932, 963,  933, 964,
+  934, 965,  935, 966,  936, 967,  937, 968,  938, 969,  939, 970,  940, 971,
+  941, 972,  942, 973,  943, 974,  944, 975,  945, 976,  946, 977,  947, 978,
+  948, 979,  949, 980,  950, 981,  951, 982,  952, 983,  953, 984,  954, 985,
+  955, 986,  956, 987,  957, 988,  958, 989,  959, 990,  960, 960,  961, 992,
+  962, 993,  963, 994,  964, 995,  965, 996,  966, 997,  967, 998,  968, 999,
+  969, 1000, 970, 1001, 971, 1002, 972, 1003, 973, 1004, 974, 1005, 975, 1006,
+  976, 1007, 977, 1008, 978, 1009, 979, 1010, 980, 1011, 981, 1012, 982, 1013,
+  983, 1014, 984, 1015, 985, 1016, 986, 1017, 987, 1018, 988, 1019, 989, 1020,
+  990, 1021, 991, 1022, 0,   0,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_32x32_neighbors[1025 * MAX_NEIGHBORS]) = {
+  0,   0,    0,   0,    32,  0,    32,  32,   1,   32,  33,  1,    64,  64,
+  33,  64,   2,   33,   96,  96,   34,  2,    65,  96,  34,  65,   128, 128,
+  97,  128,  3,   34,   66,  97,   35,  3,    35,  66,  98,  129,  129, 160,
+  160, 161,  4,   35,   67,  98,   192, 192,  36,  4,   130, 161,  161, 192,
+  36,  67,   99,  130,  5,   36,   68,  99,   193, 224, 162, 193,  224, 225,
+  131, 162,  37,  68,   100, 131,  37,  5,    194, 225, 225, 256,  256, 257,
+  163, 194,  69,  100,  132, 163,  6,   37,   226, 257, 38,  6,    195, 226,
+  257, 288,  101, 132,  288, 289,  38,  69,   164, 195, 133, 164,  258, 289,
+  227, 258,  196, 227,  7,   38,   289, 320,  70,  101, 320, 321,  39,  7,
+  165, 196,  39,  70,   102, 133,  290, 321,  259, 290, 228, 259,  321, 352,
+  352, 353,  197, 228,  134, 165,  71,  102,  8,   39,  322, 353,  291, 322,
+  260, 291,  103, 134,  353, 384,  166, 197,  229, 260, 40,  71,   40,  8,
+  384, 385,  135, 166,  354, 385,  323, 354,  198, 229, 292, 323,  72,  103,
+  261, 292,  9,   40,   385, 416,  167, 198,  104, 135, 230, 261,  355, 386,
+  416, 417,  293, 324,  324, 355,  41,  9,    41,  72,  386, 417,  199, 230,
+  136, 167,  417, 448,  262, 293,  356, 387,  73,  104, 387, 418,  231, 262,
+  10,  41,   168, 199,  325, 356,  418, 449,  105, 136, 448, 449,  42,  73,
+  294, 325,  200, 231,  42,  10,   357, 388,  137, 168, 263, 294,  388, 419,
+  74,  105,  419, 450,  449, 480,  326, 357,  232, 263, 295, 326,  169, 200,
+  11,  42,   106, 137,  480, 481,  450, 481,  358, 389, 264, 295,  201, 232,
+  138, 169,  389, 420,  43,  74,   420, 451,  327, 358, 43,  11,   481, 512,
+  233, 264,  451, 482,  296, 327,  75,  106,  170, 201, 482, 513,  512, 513,
+  390, 421,  359, 390,  421, 452,  107, 138,  12,  43,  202, 233,  452, 483,
+  265, 296,  328, 359,  139, 170,  44,  75,   483, 514, 513, 544,  234, 265,
+  297, 328,  422, 453,  44,  12,   391, 422,  171, 202, 76,  107,  514, 545,
+  453, 484,  544, 545,  266, 297,  203, 234,  108, 139, 329, 360,  298, 329,
+  140, 171,  515, 546,  13,  44,   423, 454,  235, 266, 545, 576,  454, 485,
+  45,  76,   172, 203,  330, 361,  576, 577,  45,  13,  267, 298,  546, 577,
+  77,  108,  204, 235,  455, 486,  577, 608,  299, 330, 109, 140,  547, 578,
+  14,  45,   46,  14,   141, 172,  578, 609,  331, 362, 46,  77,   173, 204,
+  15,  15,   78,  109,  205, 236,  579, 610,  110, 141, 15,  46,   142, 173,
+  47,  78,   174, 205,  16,  16,   79,  110,  206, 237, 16,  47,   111, 142,
+  48,  79,   143, 174,  80,  111,  175, 206,  17,  48,  49,  17,   207, 238,
+  49,  80,   81,  112,  18,  18,   18,  49,   50,  81,  82,  113,  19,  50,
+  51,  82,   83,  114,  608, 609,  484, 515,  360, 391, 236, 267,  112, 143,
+  51,  19,   640, 640,  609, 640,  516, 547,  485, 516, 392, 423,  361, 392,
+  268, 299,  237, 268,  144, 175,  113, 144,  20,  51,  52,  20,   672, 672,
+  641, 672,  610, 641,  548, 579,  517, 548,  486, 517, 424, 455,  393, 424,
+  362, 393,  300, 331,  269, 300,  238, 269,  176, 207, 145, 176,  114, 145,
+  52,  83,   21,  52,   53,  21,   704, 704,  673, 704, 642, 673,  611, 642,
+  580, 611,  549, 580,  518, 549,  487, 518,  456, 487, 425, 456,  394, 425,
+  363, 394,  332, 363,  301, 332,  270, 301,  239, 270, 208, 239,  177, 208,
+  146, 177,  115, 146,  84,  115,  53,  84,   22,  53,  54,  22,   705, 736,
+  674, 705,  643, 674,  581, 612,  550, 581,  519, 550, 457, 488,  426, 457,
+  395, 426,  333, 364,  302, 333,  271, 302,  209, 240, 178, 209,  147, 178,
+  85,  116,  54,  85,   23,  54,   706, 737,  675, 706, 582, 613,  551, 582,
+  458, 489,  427, 458,  334, 365,  303, 334,  210, 241, 179, 210,  86,  117,
+  55,  86,   707, 738,  583, 614,  459, 490,  335, 366, 211, 242,  87,  118,
+  736, 737,  612, 643,  488, 519,  364, 395,  240, 271, 116, 147,  55,  23,
+  768, 768,  737, 768,  644, 675,  613, 644,  520, 551, 489, 520,  396, 427,
+  365, 396,  272, 303,  241, 272,  148, 179,  117, 148, 24,  55,   56,  24,
+  800, 800,  769, 800,  738, 769,  676, 707,  645, 676, 614, 645,  552, 583,
+  521, 552,  490, 521,  428, 459,  397, 428,  366, 397, 304, 335,  273, 304,
+  242, 273,  180, 211,  149, 180,  118, 149,  56,  87,  25,  56,   57,  25,
+  832, 832,  801, 832,  770, 801,  739, 770,  708, 739, 677, 708,  646, 677,
+  615, 646,  584, 615,  553, 584,  522, 553,  491, 522, 460, 491,  429, 460,
+  398, 429,  367, 398,  336, 367,  305, 336,  274, 305, 243, 274,  212, 243,
+  181, 212,  150, 181,  119, 150,  88,  119,  57,  88,  26,  57,   58,  26,
+  833, 864,  802, 833,  771, 802,  709, 740,  678, 709, 647, 678,  585, 616,
+  554, 585,  523, 554,  461, 492,  430, 461,  399, 430, 337, 368,  306, 337,
+  275, 306,  213, 244,  182, 213,  151, 182,  89,  120, 58,  89,   27,  58,
+  834, 865,  803, 834,  710, 741,  679, 710,  586, 617, 555, 586,  462, 493,
+  431, 462,  338, 369,  307, 338,  214, 245,  183, 214, 90,  121,  59,  90,
+  835, 866,  711, 742,  587, 618,  463, 494,  339, 370, 215, 246,  91,  122,
+  864, 865,  740, 771,  616, 647,  492, 523,  368, 399, 244, 275,  120, 151,
+  59,  27,   896, 896,  865, 896,  772, 803,  741, 772, 648, 679,  617, 648,
+  524, 555,  493, 524,  400, 431,  369, 400,  276, 307, 245, 276,  152, 183,
+  121, 152,  28,  59,   60,  28,   928, 928,  897, 928, 866, 897,  804, 835,
+  773, 804,  742, 773,  680, 711,  649, 680,  618, 649, 556, 587,  525, 556,
+  494, 525,  432, 463,  401, 432,  370, 401,  308, 339, 277, 308,  246, 277,
+  184, 215,  153, 184,  122, 153,  60,  91,   29,  60,  61,  29,   960, 960,
+  929, 960,  898, 929,  867, 898,  836, 867,  805, 836, 774, 805,  743, 774,
+  712, 743,  681, 712,  650, 681,  619, 650,  588, 619, 557, 588,  526, 557,
+  495, 526,  464, 495,  433, 464,  402, 433,  371, 402, 340, 371,  309, 340,
+  278, 309,  247, 278,  216, 247,  185, 216,  154, 185, 123, 154,  92,  123,
+  61,  92,   30,  61,   62,  30,   961, 992,  930, 961, 899, 930,  837, 868,
+  806, 837,  775, 806,  713, 744,  682, 713,  651, 682, 589, 620,  558, 589,
+  527, 558,  465, 496,  434, 465,  403, 434,  341, 372, 310, 341,  279, 310,
+  217, 248,  186, 217,  155, 186,  93,  124,  62,  93,  31,  62,   962, 993,
+  931, 962,  838, 869,  807, 838,  714, 745,  683, 714, 590, 621,  559, 590,
+  466, 497,  435, 466,  342, 373,  311, 342,  218, 249, 187, 218,  94,  125,
+  63,  94,   963, 994,  839, 870,  715, 746,  591, 622, 467, 498,  343, 374,
+  219, 250,  95,  126,  868, 899,  744, 775,  620, 651, 496, 527,  372, 403,
+  248, 279,  124, 155,  900, 931,  869, 900,  776, 807, 745, 776,  652, 683,
+  621, 652,  528, 559,  497, 528,  404, 435,  373, 404, 280, 311,  249, 280,
+  156, 187,  125, 156,  932, 963,  901, 932,  870, 901, 808, 839,  777, 808,
+  746, 777,  684, 715,  653, 684,  622, 653,  560, 591, 529, 560,  498, 529,
+  436, 467,  405, 436,  374, 405,  312, 343,  281, 312, 250, 281,  188, 219,
+  157, 188,  126, 157,  964, 995,  933, 964,  902, 933, 871, 902,  840, 871,
+  809, 840,  778, 809,  747, 778,  716, 747,  685, 716, 654, 685,  623, 654,
+  592, 623,  561, 592,  530, 561,  499, 530,  468, 499, 437, 468,  406, 437,
+  375, 406,  344, 375,  313, 344,  282, 313,  251, 282, 220, 251,  189, 220,
+  158, 189,  127, 158,  965, 996,  934, 965,  903, 934, 841, 872,  810, 841,
+  779, 810,  717, 748,  686, 717,  655, 686,  593, 624, 562, 593,  531, 562,
+  469, 500,  438, 469,  407, 438,  345, 376,  314, 345, 283, 314,  221, 252,
+  190, 221,  159, 190,  966, 997,  935, 966,  842, 873, 811, 842,  718, 749,
+  687, 718,  594, 625,  563, 594,  470, 501,  439, 470, 346, 377,  315, 346,
+  222, 253,  191, 222,  967, 998,  843, 874,  719, 750, 595, 626,  471, 502,
+  347, 378,  223, 254,  872, 903,  748, 779,  624, 655, 500, 531,  376, 407,
+  252, 283,  904, 935,  873, 904,  780, 811,  749, 780, 656, 687,  625, 656,
+  532, 563,  501, 532,  408, 439,  377, 408,  284, 315, 253, 284,  936, 967,
+  905, 936,  874, 905,  812, 843,  781, 812,  750, 781, 688, 719,  657, 688,
+  626, 657,  564, 595,  533, 564,  502, 533,  440, 471, 409, 440,  378, 409,
+  316, 347,  285, 316,  254, 285,  968, 999,  937, 968, 906, 937,  875, 906,
+  844, 875,  813, 844,  782, 813,  751, 782,  720, 751, 689, 720,  658, 689,
+  627, 658,  596, 627,  565, 596,  534, 565,  503, 534, 472, 503,  441, 472,
+  410, 441,  379, 410,  348, 379,  317, 348,  286, 317, 255, 286,  969, 1000,
+  938, 969,  907, 938,  845, 876,  814, 845,  783, 814, 721, 752,  690, 721,
+  659, 690,  597, 628,  566, 597,  535, 566,  473, 504, 442, 473,  411, 442,
+  349, 380,  318, 349,  287, 318,  970, 1001, 939, 970, 846, 877,  815, 846,
+  722, 753,  691, 722,  598, 629,  567, 598,  474, 505, 443, 474,  350, 381,
+  319, 350,  971, 1002, 847, 878,  723, 754,  599, 630, 475, 506,  351, 382,
+  876, 907,  752, 783,  628, 659,  504, 535,  380, 411, 908, 939,  877, 908,
+  784, 815,  753, 784,  660, 691,  629, 660,  536, 567, 505, 536,  412, 443,
+  381, 412,  940, 971,  909, 940,  878, 909,  816, 847, 785, 816,  754, 785,
+  692, 723,  661, 692,  630, 661,  568, 599,  537, 568, 506, 537,  444, 475,
+  413, 444,  382, 413,  972, 1003, 941, 972,  910, 941, 879, 910,  848, 879,
+  817, 848,  786, 817,  755, 786,  724, 755,  693, 724, 662, 693,  631, 662,
+  600, 631,  569, 600,  538, 569,  507, 538,  476, 507, 445, 476,  414, 445,
+  383, 414,  973, 1004, 942, 973,  911, 942,  849, 880, 818, 849,  787, 818,
+  725, 756,  694, 725,  663, 694,  601, 632,  570, 601, 539, 570,  477, 508,
+  446, 477,  415, 446,  974, 1005, 943, 974,  850, 881, 819, 850,  726, 757,
+  695, 726,  602, 633,  571, 602,  478, 509,  447, 478, 975, 1006, 851, 882,
+  727, 758,  603, 634,  479, 510,  880, 911,  756, 787, 632, 663,  508, 539,
+  912, 943,  881, 912,  788, 819,  757, 788,  664, 695, 633, 664,  540, 571,
+  509, 540,  944, 975,  913, 944,  882, 913,  820, 851, 789, 820,  758, 789,
+  696, 727,  665, 696,  634, 665,  572, 603,  541, 572, 510, 541,  976, 1007,
+  945, 976,  914, 945,  883, 914,  852, 883,  821, 852, 790, 821,  759, 790,
+  728, 759,  697, 728,  666, 697,  635, 666,  604, 635, 573, 604,  542, 573,
+  511, 542,  977, 1008, 946, 977,  915, 946,  853, 884, 822, 853,  791, 822,
+  729, 760,  698, 729,  667, 698,  605, 636,  574, 605, 543, 574,  978, 1009,
+  947, 978,  854, 885,  823, 854,  730, 761,  699, 730, 606, 637,  575, 606,
+  979, 1010, 855, 886,  731, 762,  607, 638,  884, 915, 760, 791,  636, 667,
+  916, 947,  885, 916,  792, 823,  761, 792,  668, 699, 637, 668,  948, 979,
+  917, 948,  886, 917,  824, 855,  793, 824,  762, 793, 700, 731,  669, 700,
+  638, 669,  980, 1011, 949, 980,  918, 949,  887, 918, 856, 887,  825, 856,
+  794, 825,  763, 794,  732, 763,  701, 732,  670, 701, 639, 670,  981, 1012,
+  950, 981,  919, 950,  857, 888,  826, 857,  795, 826, 733, 764,  702, 733,
+  671, 702,  982, 1013, 951, 982,  858, 889,  827, 858, 734, 765,  703, 734,
+  983, 1014, 859, 890,  735, 766,  888, 919,  764, 795, 920, 951,  889, 920,
+  796, 827,  765, 796,  952, 983,  921, 952,  890, 921, 828, 859,  797, 828,
+  766, 797,  984, 1015, 953, 984,  922, 953,  891, 922, 860, 891,  829, 860,
+  798, 829,  767, 798,  985, 1016, 954, 985,  923, 954, 861, 892,  830, 861,
+  799, 830,  986, 1017, 955, 986,  862, 893,  831, 862, 987, 1018, 863, 894,
+  892, 923,  924, 955,  893, 924,  956, 987,  925, 956, 894, 925,  988, 1019,
+  957, 988,  926, 957,  895, 926,  989, 1020, 958, 989, 927, 958,  990, 1021,
+  959, 990,  991, 1022, 0,   0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                v2_scan_32x32_neighbors[1025 * MAX_NEIGHBORS]) = {
+  0,   0,    0,   0,    0,   0,    1,   32,   1,   1,   32,  32,   2,   33,
+  33,  64,   34,  65,   2,   2,    64,  64,   3,   34,  65,  96,   35,  66,
+  66,  97,   3,   3,    96,  96,   4,   35,   97,  128, 67,  98,   36,  67,
+  98,  129,  4,   4,    68,  99,   99,  130,  128, 128, 5,   36,   129, 160,
+  37,  68,   130, 161,  100, 131,  69,  100,  131, 162, 5,   5,    160, 160,
+  6,   37,   161, 192,  38,  69,   162, 193,  101, 132, 132, 163,  70,  101,
+  163, 194,  6,   6,    192, 192,  7,   38,   133, 164, 193, 224,  102, 133,
+  164, 195,  39,  70,   194, 225,  71,  102,  195, 226, 134, 165,  165, 196,
+  7,   7,    224, 224,  8,   39,   103, 134,  196, 227, 225, 256,  40,  71,
+  226, 257,  166, 197,  72,  103,  227, 258,  135, 166, 197, 228,  104, 135,
+  228, 259,  8,   8,    256, 256,  9,   40,   257, 288, 41,  72,   167, 198,
+  198, 229,  258, 289,  136, 167,  229, 260,  73,  104, 259, 290,  105, 136,
+  260, 291,  199, 230,  9,   9,    168, 199,  230, 261, 288, 288,  10,  41,
+  289, 320,  42,  73,   290, 321,  137, 168,  261, 292, 74,  105,  291, 322,
+  200, 231,  231, 262,  106, 137,  292, 323,  169, 200, 262, 293,  10,  10,
+  320, 320,  11,  42,   321, 352,  43,  74,   138, 169, 293, 324,  322, 353,
+  232, 263,  75,  106,  201, 232,  263, 294,  323, 354, 170, 201,  294, 325,
+  107, 138,  324, 355,  11,  11,   352, 352,  12,  43,  233, 264,  264, 295,
+  353, 384,  139, 170,  325, 356,  44,  75,   354, 385, 202, 233,  295, 326,
+  76,  107,  355, 386,  171, 202,  326, 357,  108, 139, 356, 387,  265, 296,
+  234, 265,  296, 327,  12,  12,   140, 171,  357, 388, 384, 384,  13,  44,
+  203, 234,  327, 358,  385, 416,  45,  76,   386, 417, 77,  108,  387, 418,
+  172, 203,  358, 389,  266, 297,  297, 328,  109, 140, 235, 266,  328, 359,
+  388, 419,  204, 235,  359, 390,  141, 172,  389, 420, 13,  13,   416, 416,
+  14,  45,   417, 448,  46,  77,   298, 329,  418, 449, 267, 298,  329, 360,
+  78,  109,  173, 204,  390, 421,  419, 450,  236, 267, 360, 391,  110, 141,
+  420, 451,  205, 236,  391, 422,  142, 173,  299, 330, 330, 361,  421, 452,
+  14,  14,   268, 299,  361, 392,  448, 448,  15,  46,  449, 480,  47,  78,
+  450, 481,  174, 205,  422, 453,  237, 268,  392, 423, 79,  110,  451, 482,
+  111, 142,  452, 483,  331, 362,  300, 331,  362, 393, 206, 237,  423, 454,
+  143, 174,  269, 300,  393, 424,  453, 484,  480, 480, 481, 512,  238, 269,
+  424, 455,  482, 513,  175, 206,  454, 485,  332, 363, 363, 394,  483, 514,
+  301, 332,  394, 425,  484, 515,  207, 238,  455, 486, 270, 301,  425, 456,
+  485, 516,  364, 395,  239, 270,  456, 487,  512, 512, 333, 364,  395, 426,
+  513, 544,  486, 517,  514, 545,  302, 333,  426, 457, 515, 546,  487, 518,
+  516, 547,  271, 302,  457, 488,  365, 396,  396, 427, 517, 548,  334, 365,
+  427, 458,  488, 519,  544, 544,  303, 334,  458, 489, 518, 549,  545, 576,
+  546, 577,  547, 578,  489, 520,  397, 428,  519, 550, 366, 397,  428, 459,
+  548, 579,  335, 366,  459, 490,  549, 580,  520, 551, 490, 521,  550, 581,
+  576, 576,  577, 608,  398, 429,  429, 460,  578, 609, 367, 398,  460, 491,
+  521, 552,  579, 610,  551, 582,  491, 522,  580, 611, 581, 612,  552, 583,
+  522, 553,  430, 461,  399, 430,  461, 492,  582, 613, 492, 523,  608, 608,
+  609, 640,  610, 641,  553, 584,  611, 642,  523, 554, 583, 614,  612, 643,
+  431, 462,  462, 493,  554, 585,  493, 524,  584, 615, 613, 644,  524, 555,
+  614, 645,  640, 640,  585, 616,  641, 672,  555, 586, 642, 673,  615, 646,
+  463, 494,  643, 674,  494, 525,  644, 675,  525, 556, 586, 617,  616, 647,
+  645, 676,  556, 587,  646, 677,  495, 526,  617, 648, 587, 618,  672, 672,
+  526, 557,  673, 704,  674, 705,  647, 678,  557, 588, 675, 706,  618, 649,
+  676, 707,  588, 619,  648, 679,  677, 708,  527, 558, 558, 589,  678, 709,
+  619, 650,  649, 680,  704, 704,  589, 620,  705, 736, 679, 710,  706, 737,
+  707, 738,  650, 681,  620, 651,  708, 739,  680, 711, 559, 590,  709, 740,
+  590, 621,  651, 682,  681, 712,  710, 741,  621, 652, 736, 736,  737, 768,
+  711, 742,  738, 769,  682, 713,  652, 683,  739, 770, 591, 622,  740, 771,
+  712, 743,  622, 653,  741, 772,  683, 714,  653, 684, 713, 744,  742, 773,
+  623, 654,  743, 774,  768, 768,  769, 800,  684, 715, 714, 745,  770, 801,
+  771, 802,  654, 685,  744, 775,  772, 803,  715, 746, 773, 804,  685, 716,
+  745, 776,  774, 805,  655, 686,  716, 747,  775, 806, 746, 777,  800, 800,
+  801, 832,  686, 717,  802, 833,  803, 834,  776, 807, 804, 835,  747, 778,
+  717, 748,  805, 836,  777, 808,  687, 718,  806, 837, 748, 779,  718, 749,
+  778, 809,  807, 838,  832, 832,  833, 864,  834, 865, 835, 866,  808, 839,
+  749, 780,  836, 867,  779, 810,  719, 750,  837, 868, 809, 840,  838, 869,
+  780, 811,  750, 781,  810, 841,  839, 870,  864, 864, 865, 896,  866, 897,
+  840, 871,  867, 898,  781, 812,  811, 842,  868, 899, 751, 782,  869, 900,
+  841, 872,  812, 843,  870, 901,  782, 813,  842, 873, 871, 902,  896, 896,
+  897, 928,  813, 844,  898, 929,  872, 903,  783, 814, 843, 874,  899, 930,
+  900, 931,  873, 904,  901, 932,  814, 845,  844, 875, 902, 933,  874, 905,
+  903, 934,  845, 876,  928, 928,  815, 846,  929, 960, 930, 961,  875, 906,
+  904, 935,  931, 962,  932, 963,  905, 936,  846, 877, 933, 964,  876, 907,
+  934, 965,  906, 937,  935, 966,  877, 908,  847, 878, 960, 960,  907, 938,
+  961, 992,  936, 967,  962, 993,  963, 994,  964, 995, 878, 909,  937, 968,
+  908, 939,  965, 996,  966, 997,  938, 969,  879, 910, 909, 940,  967, 998,
+  939, 970,  968, 999,  910, 941,  969, 1000, 940, 971, 970, 1001, 911, 942,
+  941, 972,  971, 1002, 942, 973,  972, 1003, 943, 974, 973, 1004, 974, 1005,
+  975, 1006, 15,  15,   16,  47,   48,  79,   80,  111, 112, 143,  144, 175,
+  16,  16,   17,  48,   176, 207,  49,  80,   81,  112, 113, 144,  208, 239,
+  145, 176,  240, 271,  17,  17,   18,  49,   177, 208, 50,  81,   82,  113,
+  272, 303,  209, 240,  114, 145,  146, 177,  241, 272, 304, 335,  178, 209,
+  18,  18,   19,  50,   51,  82,   83,  114,  273, 304, 210, 241,  115, 146,
+  336, 367,  147, 178,  242, 273,  305, 336,  179, 210, 19,  19,   368, 399,
+  20,  51,   52,  83,   274, 305,  84,  115,  211, 242, 337, 368,  116, 147,
+  306, 337,  148, 179,  243, 274,  400, 431,  369, 400, 180, 211,  20,  20,
+  21,  52,   275, 306,  53,  84,   338, 369,  212, 243, 85,  116,  432, 463,
+  117, 148,  401, 432,  307, 338,  244, 275,  149, 180, 370, 401,  181, 212,
+  276, 307,  464, 495,  339, 370,  21,  21,   22,  53,  433, 464,  54,  85,
+  213, 244,  86,  117,  402, 433,  118, 149,  308, 339, 245, 276,  371, 402,
+  150, 181,  496, 527,  465, 496,  182, 213,  434, 465, 340, 371,  277, 308,
+  22,  22,   23,  54,   403, 434,  55,  86,   214, 245, 87,  118,  309, 340,
+  372, 403,  119, 150,  497, 528,  528, 559,  246, 277, 466, 497,  151, 182,
+  435, 466,  341, 372,  183, 214,  278, 309,  404, 435, 23,  23,   24,  55,
+  215, 246,  529, 560,  56,  87,   498, 529,  560, 591, 310, 341,  88,  119,
+  373, 404,  467, 498,  120, 151,  247, 278,  436, 467, 152, 183,  342, 373,
+  279, 310,  405, 436,  184, 215,  530, 561,  561, 592, 499, 530,  592, 623,
+  24,  24,   216, 247,  468, 499,  25,  56,   374, 405, 57,  88,   311, 342,
+  89,  120,  437, 468,  248, 279,  121, 152,  562, 593, 153, 184,  343, 374,
+  531, 562,  593, 624,  406, 437,  500, 531,  624, 655, 280, 311,  185, 216,
+  469, 500,  375, 406,  217, 248,  25,  25,   312, 343, 26,  57,   58,  89,
+  438, 469,  90,  121,  563, 594,  594, 625,  249, 280, 532, 563,  625, 656,
+  122, 153,  344, 375,  501, 532,  656, 687,  407, 438, 154, 185,  281, 312,
+  470, 501,  186, 217,  376, 407,  595, 626,  564, 595, 626, 657,  218, 249,
+  313, 344,  439, 470,  26,  26,   27,  58,   533, 564, 657, 688,  59,  90,
+  91,  122,  250, 281,  502, 533,  688, 719,  123, 154, 408, 439,  345, 376,
+  155, 186,  471, 502,  282, 313,  596, 627,  627, 658, 187, 218,  565, 596,
+  658, 689,  377, 408,  440, 471,  534, 565,  689, 720, 314, 345,  219, 250,
+  27,  27,   28,  59,   503, 534,  720, 751,  60,  91,  92,  123,  251, 282,
+  409, 440,  346, 377,  124, 155,  628, 659,  472, 503, 597, 628,  659, 690,
+  566, 597,  690, 721,  156, 187,  283, 314,  535, 566, 721, 752,  188, 219,
+  378, 409,  441, 472,  315, 346,  504, 535,  752, 783, 220, 251,  28,  28,
+  629, 660,  660, 691,  29,  60,   61,  92,   410, 441, 598, 629,  691, 722,
+  252, 283,  93,  124,  347, 378,  473, 504,  567, 598, 722, 753,  125, 156,
+  284, 315,  536, 567,  753, 784,  157, 188,  442, 473, 379, 410,  189, 220,
+  505, 536,  784, 815,  661, 692,  316, 347,  630, 661, 692, 723,  221, 252,
+  599, 630,  723, 754,  411, 442,  29,  29,   568, 599, 754, 785,  30,  61,
+  474, 505,  62,  93,   253, 284,  348, 379,  94,  125, 537, 568,  785, 816,
+  126, 157,  285, 316,  158, 189,  443, 474,  662, 693, 693, 724,  380, 411,
+  631, 662,  724, 755,  506, 537,  816, 847,  190, 221, 600, 631,  755, 786,
+  317, 348,  222, 253,  569, 600,  786, 817,  412, 443, 475, 506,  30,  30,
+  31,  62,   349, 380,  254, 285,  63,  94,   538, 569, 817, 848,  694, 725,
+  95,  126,  663, 694,  725, 756,  632, 663,  756, 787, 127, 158,  444, 475,
+  286, 317,  381, 412,  507, 538,  848, 879,  159, 190, 601, 632,  787, 818,
+  191, 222,  318, 349,  570, 601,  818, 849,  476, 507, 223, 254,  413, 444,
+  695, 726,  726, 757,  664, 695,  757, 788,  539, 570, 849, 880,  350, 381,
+  255, 286,  633, 664,  788, 819,  445, 476,  602, 633, 819, 850,  508, 539,
+  880, 911,  287, 318,  382, 413,  571, 602,  850, 881, 727, 758,  696, 727,
+  758, 789,  319, 350,  477, 508,  665, 696,  789, 820, 414, 445,  540, 571,
+  881, 912,  634, 665,  820, 851,  351, 382,  603, 634, 851, 882,  446, 477,
+  509, 540,  912, 943,  383, 414,  728, 759,  759, 790, 572, 603,  882, 913,
+  697, 728,  790, 821,  666, 697,  821, 852,  478, 509, 635, 666,  852, 883,
+  415, 446,  541, 572,  913, 944,  604, 635,  883, 914, 760, 791,  729, 760,
+  791, 822,  510, 541,  944, 975,  447, 478,  698, 729, 822, 853,  573, 604,
+  914, 945,  667, 698,  853, 884,  636, 667,  884, 915, 479, 510,  542, 573,
+  945, 976,  761, 792,  792, 823,  605, 636,  915, 946, 730, 761,  823, 854,
+  699, 730,  854, 885,  511, 542,  976, 1007, 574, 605, 946, 977,  668, 699,
+  885, 916,  637, 668,  916, 947,  543, 574,  793, 824, 977, 1008, 762, 793,
+  824, 855,  731, 762,  855, 886,  606, 637,  947, 978, 700, 731,  886, 917,
+  669, 700,  917, 948,  575, 606,  978, 1009, 638, 669, 948, 979,  794, 825,
+  825, 856,  763, 794,  856, 887,  732, 763,  887, 918, 607, 638,  979, 1010,
+  701, 732,  918, 949,  670, 701,  949, 980,  826, 857, 795, 826,  857, 888,
+  764, 795,  888, 919,  639, 670,  980, 1011, 733, 764, 919, 950,  702, 733,
+  950, 981,  671, 702,  981, 1012, 827, 858,  858, 889, 796, 827,  889, 920,
+  765, 796,  920, 951,  734, 765,  951, 982,  703, 734, 982, 1013, 859, 890,
+  828, 859,  890, 921,  797, 828,  921, 952,  766, 797, 952, 983,  735, 766,
+  983, 1014, 860, 891,  891, 922,  829, 860,  922, 953, 798, 829,  953, 984,
+  767, 798,  984, 1015, 892, 923,  861, 892,  923, 954, 830, 861,  954, 985,
+  799, 830,  985, 1016, 893, 924,  924, 955,  862, 893, 955, 986,  831, 862,
+  986, 1017, 925, 956,  894, 925,  956, 987,  863, 894, 987, 1018, 926, 957,
+  957, 988,  895, 926,  988, 1019, 958, 989,  927, 958, 989, 1020, 959, 990,
+  990, 1021, 991, 1022, 0,   0,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                h2_scan_32x32_neighbors[1025 * MAX_NEIGHBORS]) = {
+  0,   0,    0,   0,    0,   0,    1,   32,   1,   1,    32,  32,   2,   33,
+  33,  64,   34,  65,   2,   2,    64,  64,   3,   34,   65,  96,   35,  66,
+  66,  97,   3,   3,    96,  96,   4,   35,   97,  128,  67,  98,   36,  67,
+  98,  129,  4,   4,    68,  99,   99,  130,  128, 128,  5,   36,   129, 160,
+  37,  68,   130, 161,  100, 131,  69,  100,  131, 162,  5,   5,    160, 160,
+  6,   37,   161, 192,  38,  69,   162, 193,  101, 132,  132, 163,  70,  101,
+  163, 194,  6,   6,    192, 192,  7,   38,   133, 164,  193, 224,  102, 133,
+  164, 195,  39,  70,   194, 225,  71,  102,  195, 226,  134, 165,  165, 196,
+  7,   7,    224, 224,  8,   39,   103, 134,  196, 227,  225, 256,  40,  71,
+  226, 257,  166, 197,  72,  103,  227, 258,  135, 166,  197, 228,  104, 135,
+  228, 259,  8,   8,    256, 256,  9,   40,   257, 288,  41,  72,   167, 198,
+  198, 229,  258, 289,  136, 167,  229, 260,  73,  104,  259, 290,  105, 136,
+  260, 291,  199, 230,  9,   9,    168, 199,  230, 261,  288, 288,  10,  41,
+  289, 320,  42,  73,   290, 321,  137, 168,  261, 292,  74,  105,  291, 322,
+  200, 231,  231, 262,  106, 137,  292, 323,  169, 200,  262, 293,  10,  10,
+  320, 320,  11,  42,   321, 352,  43,  74,   138, 169,  293, 324,  322, 353,
+  232, 263,  75,  106,  201, 232,  263, 294,  323, 354,  170, 201,  294, 325,
+  107, 138,  324, 355,  11,  11,   352, 352,  12,  43,   233, 264,  264, 295,
+  353, 384,  139, 170,  325, 356,  44,  75,   354, 385,  202, 233,  295, 326,
+  76,  107,  355, 386,  171, 202,  326, 357,  108, 139,  356, 387,  265, 296,
+  234, 265,  296, 327,  12,  12,   140, 171,  357, 388,  384, 384,  13,  44,
+  203, 234,  327, 358,  385, 416,  45,  76,   386, 417,  77,  108,  387, 418,
+  172, 203,  358, 389,  266, 297,  297, 328,  109, 140,  235, 266,  328, 359,
+  388, 419,  204, 235,  359, 390,  141, 172,  389, 420,  13,  13,   416, 416,
+  14,  45,   417, 448,  46,  77,   298, 329,  418, 449,  267, 298,  329, 360,
+  78,  109,  173, 204,  390, 421,  419, 450,  236, 267,  360, 391,  110, 141,
+  420, 451,  205, 236,  391, 422,  142, 173,  299, 330,  330, 361,  421, 452,
+  14,  14,   268, 299,  361, 392,  448, 448,  15,  46,   449, 480,  47,  78,
+  450, 481,  174, 205,  422, 453,  237, 268,  392, 423,  79,  110,  451, 482,
+  111, 142,  452, 483,  331, 362,  300, 331,  362, 393,  206, 237,  423, 454,
+  143, 174,  269, 300,  393, 424,  453, 484,  15,  15,   16,  47,   48,  79,
+  238, 269,  424, 455,  175, 206,  454, 485,  80,  111,  332, 363,  363, 394,
+  301, 332,  394, 425,  112, 143,  207, 238,  455, 486,  270, 301,  425, 456,
+  144, 175,  364, 395,  16,  16,   239, 270,  456, 487,  17,  48,   333, 364,
+  395, 426,  176, 207,  49,  80,   302, 333,  426, 457,  81,  112,  113, 144,
+  208, 239,  271, 302,  457, 488,  365, 396,  396, 427,  145, 176,  334, 365,
+  427, 458,  240, 271,  17,  17,   18,  49,   177, 208,  303, 334,  458, 489,
+  50,  81,   82,  113,  272, 303,  209, 240,  397, 428,  114, 145,  366, 397,
+  428, 459,  335, 366,  459, 490,  146, 177,  241, 272,  304, 335,  178, 209,
+  18,  18,   19,  50,   51,  82,   398, 429,  429, 460,  367, 398,  460, 491,
+  83,  114,  273, 304,  210, 241,  115, 146,  336, 367,  147, 178,  242, 273,
+  305, 336,  430, 461,  399, 430,  461, 492,  179, 210,  19,  19,   368, 399,
+  20,  51,   52,  83,   274, 305,  84,  115,  211, 242,  337, 368,  116, 147,
+  431, 462,  462, 493,  306, 337,  148, 179,  243, 274,  400, 431,  369, 400,
+  180, 211,  20,  20,   21,  52,   275, 306,  53,  84,   338, 369,  212, 243,
+  85,  116,  463, 494,  432, 463,  117, 148,  401, 432,  307, 338,  244, 275,
+  149, 180,  370, 401,  181, 212,  276, 307,  464, 495,  339, 370,  21,  21,
+  22,  53,   433, 464,  54,  85,   213, 244,  86,  117,  402, 433,  118, 149,
+  308, 339,  245, 276,  371, 402,  150, 181,  465, 496,  182, 213,  434, 465,
+  340, 371,  277, 308,  22,  22,   23,  54,   403, 434,  55,  86,   214, 245,
+  87,  118,  309, 340,  372, 403,  119, 150,  246, 277,  466, 497,  151, 182,
+  435, 466,  341, 372,  183, 214,  278, 309,  404, 435,  23,  23,   24,  55,
+  215, 246,  56,  87,   310, 341,  88,  119,  373, 404,  467, 498,  120, 151,
+  247, 278,  436, 467,  152, 183,  342, 373,  279, 310,  405, 436,  184, 215,
+  24,  24,   216, 247,  468, 499,  25,  56,   374, 405,  57,  88,   311, 342,
+  89,  120,  437, 468,  248, 279,  121, 152,  153, 184,  343, 374,  406, 437,
+  280, 311,  185, 216,  469, 500,  375, 406,  217, 248,  25,  25,   312, 343,
+  26,  57,   58,  89,   438, 469,  90,  121,  249, 280,  122, 153,  344, 375,
+  407, 438,  154, 185,  281, 312,  470, 501,  186, 217,  376, 407,  218, 249,
+  313, 344,  439, 470,  26,  26,   27,  58,   59,  90,   91,  122,  250, 281,
+  123, 154,  408, 439,  345, 376,  155, 186,  471, 502,  282, 313,  187, 218,
+  377, 408,  440, 471,  314, 345,  219, 250,  27,  27,   28,  59,   60,  91,
+  92,  123,  251, 282,  409, 440,  346, 377,  124, 155,  472, 503,  156, 187,
+  283, 314,  188, 219,  378, 409,  441, 472,  315, 346,  220, 251,  28,  28,
+  29,  60,   61,  92,   410, 441,  252, 283,  93,  124,  347, 378,  473, 504,
+  125, 156,  284, 315,  157, 188,  442, 473,  379, 410,  189, 220,  316, 347,
+  221, 252,  411, 442,  29,  29,   30,  61,   474, 505,  62,  93,   253, 284,
+  348, 379,  94,  125,  126, 157,  285, 316,  158, 189,  443, 474,  380, 411,
+  190, 221,  317, 348,  222, 253,  412, 443,  475, 506,  30,  30,   31,  62,
+  349, 380,  254, 285,  63,  94,   95,  126,  127, 158,  444, 475,  286, 317,
+  381, 412,  159, 190,  191, 222,  318, 349,  476, 507,  223, 254,  413, 444,
+  350, 381,  255, 286,  445, 476,  287, 318,  382, 413,  319, 350,  477, 508,
+  414, 445,  351, 382,  446, 477,  383, 414,  478, 509,  415, 446,  447, 478,
+  479, 510,  480, 480,  481, 512,  482, 513,  483, 514,  484, 515,  485, 516,
+  512, 512,  513, 544,  486, 517,  514, 545,  515, 546,  487, 518,  516, 547,
+  517, 548,  488, 519,  544, 544,  518, 549,  545, 576,  546, 577,  547, 578,
+  489, 520,  519, 550,  548, 579,  549, 580,  520, 551,  490, 521,  550, 581,
+  576, 576,  577, 608,  578, 609,  521, 552,  579, 610,  551, 582,  491, 522,
+  580, 611,  581, 612,  552, 583,  522, 553,  582, 613,  492, 523,  608, 608,
+  609, 640,  610, 641,  553, 584,  611, 642,  523, 554,  583, 614,  612, 643,
+  554, 585,  493, 524,  584, 615,  613, 644,  524, 555,  614, 645,  640, 640,
+  585, 616,  641, 672,  555, 586,  642, 673,  615, 646,  643, 674,  494, 525,
+  644, 675,  525, 556,  586, 617,  616, 647,  645, 676,  556, 587,  646, 677,
+  495, 526,  617, 648,  587, 618,  672, 672,  526, 557,  673, 704,  674, 705,
+  647, 678,  557, 588,  675, 706,  618, 649,  676, 707,  588, 619,  648, 679,
+  677, 708,  496, 527,  527, 558,  558, 589,  678, 709,  619, 650,  649, 680,
+  704, 704,  589, 620,  705, 736,  679, 710,  706, 737,  707, 738,  650, 681,
+  620, 651,  497, 528,  528, 559,  708, 739,  680, 711,  559, 590,  709, 740,
+  590, 621,  651, 682,  681, 712,  710, 741,  621, 652,  736, 736,  737, 768,
+  529, 560,  711, 742,  498, 529,  560, 591,  738, 769,  682, 713,  652, 683,
+  739, 770,  591, 622,  740, 771,  712, 743,  622, 653,  741, 772,  683, 714,
+  653, 684,  713, 744,  742, 773,  530, 561,  561, 592,  499, 530,  592, 623,
+  623, 654,  743, 774,  768, 768,  769, 800,  684, 715,  714, 745,  770, 801,
+  771, 802,  654, 685,  744, 775,  772, 803,  562, 593,  531, 562,  593, 624,
+  715, 746,  773, 804,  685, 716,  500, 531,  624, 655,  745, 776,  774, 805,
+  655, 686,  716, 747,  775, 806,  746, 777,  800, 800,  801, 832,  686, 717,
+  802, 833,  563, 594,  594, 625,  803, 834,  532, 563,  625, 656,  776, 807,
+  804, 835,  501, 532,  656, 687,  747, 778,  717, 748,  805, 836,  777, 808,
+  687, 718,  806, 837,  748, 779,  595, 626,  564, 595,  626, 657,  718, 749,
+  778, 809,  807, 838,  832, 832,  533, 564,  657, 688,  833, 864,  834, 865,
+  835, 866,  502, 533,  688, 719,  808, 839,  749, 780,  836, 867,  779, 810,
+  719, 750,  837, 868,  809, 840,  596, 627,  627, 658,  565, 596,  658, 689,
+  838, 869,  780, 811,  750, 781,  534, 565,  689, 720,  810, 841,  839, 870,
+  864, 864,  503, 534,  720, 751,  865, 896,  866, 897,  840, 871,  867, 898,
+  781, 812,  811, 842,  628, 659,  868, 899,  751, 782,  597, 628,  659, 690,
+  566, 597,  690, 721,  869, 900,  841, 872,  535, 566,  721, 752,  812, 843,
+  870, 901,  782, 813,  842, 873,  504, 535,  752, 783,  871, 902,  629, 660,
+  660, 691,  896, 896,  897, 928,  598, 629,  691, 722,  813, 844,  898, 929,
+  872, 903,  783, 814,  843, 874,  899, 930,  567, 598,  722, 753,  900, 931,
+  536, 567,  753, 784,  873, 904,  901, 932,  814, 845,  844, 875,  902, 933,
+  505, 536,  784, 815,  661, 692,  630, 661,  692, 723,  874, 905,  599, 630,
+  723, 754,  903, 934,  845, 876,  568, 599,  754, 785,  928, 928,  815, 846,
+  929, 960,  930, 961,  875, 906,  904, 935,  931, 962,  537, 568,  785, 816,
+  932, 963,  905, 936,  662, 693,  693, 724,  846, 877,  933, 964,  876, 907,
+  631, 662,  724, 755,  506, 537,  816, 847,  934, 965,  600, 631,  755, 786,
+  906, 937,  569, 600,  786, 817,  935, 966,  877, 908,  847, 878,  960, 960,
+  907, 938,  961, 992,  936, 967,  538, 569,  817, 848,  962, 993,  694, 725,
+  663, 694,  725, 756,  963, 994,  632, 663,  756, 787,  964, 995,  878, 909,
+  937, 968,  507, 538,  848, 879,  908, 939,  601, 632,  787, 818,  965, 996,
+  966, 997,  570, 601,  818, 849,  938, 969,  879, 910,  909, 940,  967, 998,
+  695, 726,  726, 757,  664, 695,  757, 788,  539, 570,  849, 880,  939, 970,
+  633, 664,  788, 819,  968, 999,  602, 633,  819, 850,  910, 941,  508, 539,
+  880, 911,  969, 1000, 940, 971,  571, 602,  850, 881,  727, 758,  696, 727,
+  758, 789,  970, 1001, 665, 696,  789, 820,  911, 942,  941, 972,  540, 571,
+  881, 912,  634, 665,  820, 851,  971, 1002, 603, 634,  851, 882,  942, 973,
+  509, 540,  912, 943,  728, 759,  759, 790,  972, 1003, 572, 603,  882, 913,
+  697, 728,  790, 821,  666, 697,  821, 852,  943, 974,  635, 666,  852, 883,
+  541, 572,  913, 944,  973, 1004, 604, 635,  883, 914,  760, 791,  729, 760,
+  791, 822,  510, 541,  944, 975,  974, 1005, 698, 729,  822, 853,  573, 604,
+  914, 945,  667, 698,  853, 884,  636, 667,  884, 915,  975, 1006, 542, 573,
+  945, 976,  761, 792,  792, 823,  605, 636,  915, 946,  730, 761,  823, 854,
+  699, 730,  854, 885,  511, 542,  976, 1007, 574, 605,  946, 977,  668, 699,
+  885, 916,  637, 668,  916, 947,  543, 574,  793, 824,  977, 1008, 762, 793,
+  824, 855,  731, 762,  855, 886,  606, 637,  947, 978,  700, 731,  886, 917,
+  669, 700,  917, 948,  575, 606,  978, 1009, 638, 669,  948, 979,  794, 825,
+  825, 856,  763, 794,  856, 887,  732, 763,  887, 918,  607, 638,  979, 1010,
+  701, 732,  918, 949,  670, 701,  949, 980,  826, 857,  795, 826,  857, 888,
+  764, 795,  888, 919,  639, 670,  980, 1011, 733, 764,  919, 950,  702, 733,
+  950, 981,  671, 702,  981, 1012, 827, 858,  858, 889,  796, 827,  889, 920,
+  765, 796,  920, 951,  734, 765,  951, 982,  703, 734,  982, 1013, 859, 890,
+  828, 859,  890, 921,  797, 828,  921, 952,  766, 797,  952, 983,  735, 766,
+  983, 1014, 860, 891,  891, 922,  829, 860,  922, 953,  798, 829,  953, 984,
+  767, 798,  984, 1015, 892, 923,  861, 892,  923, 954,  830, 861,  954, 985,
+  799, 830,  985, 1016, 893, 924,  924, 955,  862, 893,  955, 986,  831, 862,
+  986, 1017, 925, 956,  894, 925,  956, 987,  863, 894,  987, 1018, 926, 957,
+  957, 988,  895, 926,  988, 1019, 958, 989,  927, 958,  989, 1020, 959, 990,
+  990, 1021, 991, 1022, 0,   0
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                qtr_scan_32x32_neighbors[1025 * MAX_NEIGHBORS]) = {
+  0,   0,    0,   0,    0,   0,    1,   32,   1,   1,    32,  32,   2,   33,
+  33,  64,   34,  65,   2,   2,    64,  64,   3,   34,   65,  96,   35,  66,
+  66,  97,   3,   3,    96,  96,   4,   35,   97,  128,  67,  98,   36,  67,
+  98,  129,  4,   4,    68,  99,   99,  130,  128, 128,  5,   36,   129, 160,
+  37,  68,   130, 161,  100, 131,  69,  100,  131, 162,  5,   5,    160, 160,
+  6,   37,   161, 192,  38,  69,   162, 193,  101, 132,  132, 163,  70,  101,
+  163, 194,  6,   6,    192, 192,  7,   38,   133, 164,  193, 224,  102, 133,
+  164, 195,  39,  70,   194, 225,  71,  102,  195, 226,  134, 165,  165, 196,
+  7,   7,    224, 224,  8,   39,   103, 134,  196, 227,  225, 256,  40,  71,
+  226, 257,  166, 197,  72,  103,  227, 258,  135, 166,  197, 228,  104, 135,
+  228, 259,  8,   8,    256, 256,  9,   40,   257, 288,  41,  72,   167, 198,
+  198, 229,  258, 289,  136, 167,  229, 260,  73,  104,  259, 290,  105, 136,
+  260, 291,  199, 230,  9,   9,    168, 199,  230, 261,  288, 288,  10,  41,
+  289, 320,  42,  73,   290, 321,  137, 168,  261, 292,  74,  105,  291, 322,
+  200, 231,  231, 262,  106, 137,  292, 323,  169, 200,  262, 293,  10,  10,
+  320, 320,  11,  42,   321, 352,  43,  74,   138, 169,  293, 324,  322, 353,
+  232, 263,  75,  106,  201, 232,  263, 294,  323, 354,  170, 201,  294, 325,
+  107, 138,  324, 355,  11,  11,   352, 352,  12,  43,   233, 264,  264, 295,
+  353, 384,  139, 170,  325, 356,  44,  75,   354, 385,  202, 233,  295, 326,
+  76,  107,  355, 386,  171, 202,  326, 357,  108, 139,  356, 387,  265, 296,
+  234, 265,  296, 327,  12,  12,   140, 171,  357, 388,  384, 384,  13,  44,
+  203, 234,  327, 358,  385, 416,  45,  76,   386, 417,  77,  108,  387, 418,
+  172, 203,  358, 389,  266, 297,  297, 328,  109, 140,  235, 266,  328, 359,
+  388, 419,  204, 235,  359, 390,  141, 172,  389, 420,  13,  13,   416, 416,
+  14,  45,   417, 448,  46,  77,   298, 329,  418, 449,  267, 298,  329, 360,
+  78,  109,  173, 204,  390, 421,  419, 450,  236, 267,  360, 391,  110, 141,
+  420, 451,  205, 236,  391, 422,  142, 173,  299, 330,  330, 361,  421, 452,
+  14,  14,   268, 299,  361, 392,  448, 448,  15,  46,   449, 480,  47,  78,
+  450, 481,  174, 205,  422, 453,  237, 268,  392, 423,  79,  110,  451, 482,
+  111, 142,  452, 483,  331, 362,  300, 331,  362, 393,  206, 237,  423, 454,
+  143, 174,  269, 300,  393, 424,  453, 484,  238, 269,  424, 455,  175, 206,
+  454, 485,  332, 363,  363, 394,  301, 332,  394, 425,  207, 238,  455, 486,
+  270, 301,  425, 456,  364, 395,  239, 270,  456, 487,  333, 364,  395, 426,
+  302, 333,  426, 457,  271, 302,  457, 488,  365, 396,  396, 427,  334, 365,
+  427, 458,  303, 334,  458, 489,  397, 428,  366, 397,  428, 459,  335, 366,
+  459, 490,  398, 429,  429, 460,  367, 398,  460, 491,  430, 461,  399, 430,
+  461, 492,  431, 462,  462, 493,  463, 494,  15,  15,   480, 480,  16,  47,
+  481, 512,  48,  79,   482, 513,  80,  111,  483, 514,  112, 143,  484, 515,
+  144, 175,  485, 516,  16,  16,   512, 512,  17,  48,   513, 544,  176, 207,
+  486, 517,  49,  80,   514, 545,  81,  112,  515, 546,  113, 144,  208, 239,
+  487, 518,  516, 547,  145, 176,  517, 548,  240, 271,  488, 519,  17,  17,
+  544, 544,  18,  49,   177, 208,  518, 549,  545, 576,  50,  81,   546, 577,
+  82,  113,  547, 578,  272, 303,  489, 520,  209, 240,  519, 550,  114, 145,
+  548, 579,  146, 177,  549, 580,  241, 272,  520, 551,  304, 335,  490, 521,
+  178, 209,  550, 581,  18,  18,   576, 576,  19,  50,   577, 608,  51,  82,
+  578, 609,  83,  114,  273, 304,  521, 552,  579, 610,  210, 241,  551, 582,
+  115, 146,  336, 367,  491, 522,  580, 611,  147, 178,  581, 612,  242, 273,
+  552, 583,  305, 336,  522, 553,  179, 210,  582, 613,  19,  19,   368, 399,
+  492, 523,  608, 608,  20,  51,   609, 640,  52,  83,   610, 641,  274, 305,
+  553, 584,  84,  115,  611, 642,  211, 242,  337, 368,  523, 554,  583, 614,
+  116, 147,  612, 643,  306, 337,  554, 585,  148, 179,  243, 274,  400, 431,
+  493, 524,  584, 615,  613, 644,  369, 400,  524, 555,  180, 211,  614, 645,
+  20,  20,   640, 640,  21,  52,   275, 306,  585, 616,  641, 672,  53,  84,
+  338, 369,  555, 586,  642, 673,  212, 243,  615, 646,  85,  116,  643, 674,
+  432, 463,  494, 525,  117, 148,  644, 675,  401, 432,  525, 556,  307, 338,
+  586, 617,  244, 275,  616, 647,  149, 180,  645, 676,  370, 401,  556, 587,
+  181, 212,  646, 677,  276, 307,  464, 495,  495, 526,  617, 648,  339, 370,
+  587, 618,  21,  21,   672, 672,  22,  53,   433, 464,  526, 557,  673, 704,
+  54,  85,   674, 705,  213, 244,  647, 678,  86,  117,  402, 433,  557, 588,
+  675, 706,  118, 149,  308, 339,  618, 649,  676, 707,  245, 276,  371, 402,
+  588, 619,  648, 679,  150, 181,  677, 708,  496, 527,  465, 496,  527, 558,
+  182, 213,  434, 465,  558, 589,  678, 709,  340, 371,  619, 650,  277, 308,
+  649, 680,  22,  22,   704, 704,  23,  54,   403, 434,  589, 620,  705, 736,
+  55,  86,   214, 245,  679, 710,  706, 737,  87,  118,  707, 738,  309, 340,
+  650, 681,  372, 403,  620, 651,  119, 150,  497, 528,  528, 559,  708, 739,
+  246, 277,  680, 711,  466, 497,  559, 590,  151, 182,  709, 740,  435, 466,
+  590, 621,  341, 372,  651, 682,  183, 214,  278, 309,  681, 712,  710, 741,
+  404, 435,  621, 652,  23,  23,   736, 736,  24,  55,   737, 768,  215, 246,
+  529, 560,  711, 742,  56,  87,   498, 529,  560, 591,  738, 769,  310, 341,
+  682, 713,  88,  119,  373, 404,  652, 683,  739, 770,  467, 498,  591, 622,
+  120, 151,  740, 771,  247, 278,  712, 743,  436, 467,  622, 653,  152, 183,
+  741, 772,  342, 373,  683, 714,  279, 310,  405, 436,  653, 684,  713, 744,
+  184, 215,  742, 773,  530, 561,  561, 592,  499, 530,  592, 623,  24,  24,
+  216, 247,  468, 499,  623, 654,  743, 774,  768, 768,  25,  56,   769, 800,
+  374, 405,  684, 715,  57,  88,   311, 342,  714, 745,  770, 801,  89,  120,
+  771, 802,  437, 468,  654, 685,  248, 279,  744, 775,  121, 152,  772, 803,
+  562, 593,  153, 184,  343, 374,  531, 562,  593, 624,  715, 746,  773, 804,
+  406, 437,  685, 716,  500, 531,  624, 655,  280, 311,  745, 776,  185, 216,
+  774, 805,  469, 500,  655, 686,  375, 406,  716, 747,  217, 248,  775, 806,
+  25,  25,   312, 343,  746, 777,  800, 800,  26,  57,   801, 832,  58,  89,
+  438, 469,  686, 717,  802, 833,  90,  121,  563, 594,  594, 625,  803, 834,
+  249, 280,  532, 563,  625, 656,  776, 807,  122, 153,  804, 835,  344, 375,
+  501, 532,  656, 687,  747, 778,  407, 438,  717, 748,  154, 185,  805, 836,
+  281, 312,  777, 808,  470, 501,  687, 718,  186, 217,  806, 837,  376, 407,
+  748, 779,  595, 626,  564, 595,  626, 657,  218, 249,  313, 344,  439, 470,
+  718, 749,  778, 809,  807, 838,  26,  26,   832, 832,  27,  58,   533, 564,
+  657, 688,  833, 864,  59,  90,   834, 865,  91,  122,  835, 866,  250, 281,
+  502, 533,  688, 719,  808, 839,  123, 154,  408, 439,  749, 780,  836, 867,
+  345, 376,  779, 810,  155, 186,  471, 502,  719, 750,  837, 868,  282, 313,
+  809, 840,  596, 627,  627, 658,  187, 218,  565, 596,  658, 689,  838, 869,
+  377, 408,  780, 811,  440, 471,  750, 781,  534, 565,  689, 720,  314, 345,
+  810, 841,  219, 250,  839, 870,  27,  27,   864, 864,  28,  59,   503, 534,
+  720, 751,  865, 896,  60,  91,   866, 897,  92,  123,  251, 282,  840, 871,
+  867, 898,  409, 440,  781, 812,  346, 377,  811, 842,  124, 155,  628, 659,
+  868, 899,  472, 503,  751, 782,  597, 628,  659, 690,  566, 597,  690, 721,
+  156, 187,  869, 900,  283, 314,  841, 872,  535, 566,  721, 752,  188, 219,
+  378, 409,  812, 843,  870, 901,  441, 472,  782, 813,  315, 346,  842, 873,
+  504, 535,  752, 783,  220, 251,  871, 902,  28,  28,   629, 660,  660, 691,
+  896, 896,  29,  60,   897, 928,  61,  92,   410, 441,  598, 629,  691, 722,
+  813, 844,  898, 929,  252, 283,  872, 903,  93,  124,  347, 378,  473, 504,
+  783, 814,  843, 874,  899, 930,  567, 598,  722, 753,  125, 156,  900, 931,
+  284, 315,  536, 567,  753, 784,  873, 904,  157, 188,  901, 932,  442, 473,
+  814, 845,  379, 410,  844, 875,  189, 220,  902, 933,  505, 536,  784, 815,
+  661, 692,  316, 347,  630, 661,  692, 723,  874, 905,  221, 252,  599, 630,
+  723, 754,  903, 934,  411, 442,  845, 876,  29,  29,   568, 599,  754, 785,
+  928, 928,  30,  61,   474, 505,  815, 846,  929, 960,  62,  93,   930, 961,
+  253, 284,  348, 379,  875, 906,  904, 935,  94,  125,  931, 962,  537, 568,
+  785, 816,  126, 157,  932, 963,  285, 316,  905, 936,  158, 189,  443, 474,
+  662, 693,  693, 724,  846, 877,  933, 964,  380, 411,  876, 907,  631, 662,
+  724, 755,  506, 537,  816, 847,  190, 221,  934, 965,  600, 631,  755, 786,
+  317, 348,  906, 937,  222, 253,  569, 600,  786, 817,  935, 966,  412, 443,
+  877, 908,  475, 506,  847, 878,  30,  30,   960, 960,  31,  62,   349, 380,
+  907, 938,  961, 992,  254, 285,  936, 967,  63,  94,   538, 569,  817, 848,
+  962, 993,  694, 725,  95,  126,  663, 694,  725, 756,  963, 994,  632, 663,
+  756, 787,  127, 158,  964, 995,  444, 475,  878, 909,  286, 317,  937, 968,
+  381, 412,  507, 538,  848, 879,  908, 939,  159, 190,  601, 632,  787, 818,
+  965, 996,  191, 222,  966, 997,  318, 349,  570, 601,  818, 849,  938, 969,
+  476, 507,  879, 910,  223, 254,  413, 444,  909, 940,  967, 998,  695, 726,
+  726, 757,  664, 695,  757, 788,  539, 570,  849, 880,  350, 381,  939, 970,
+  255, 286,  633, 664,  788, 819,  968, 999,  445, 476,  602, 633,  819, 850,
+  910, 941,  508, 539,  880, 911,  287, 318,  969, 1000, 382, 413,  940, 971,
+  571, 602,  850, 881,  727, 758,  696, 727,  758, 789,  319, 350,  970, 1001,
+  477, 508,  665, 696,  789, 820,  911, 942,  414, 445,  941, 972,  540, 571,
+  881, 912,  634, 665,  820, 851,  351, 382,  971, 1002, 603, 634,  851, 882,
+  446, 477,  942, 973,  509, 540,  912, 943,  383, 414,  728, 759,  759, 790,
+  972, 1003, 572, 603,  882, 913,  697, 728,  790, 821,  666, 697,  821, 852,
+  478, 509,  943, 974,  635, 666,  852, 883,  415, 446,  541, 572,  913, 944,
+  973, 1004, 604, 635,  883, 914,  760, 791,  729, 760,  791, 822,  510, 541,
+  944, 975,  447, 478,  974, 1005, 698, 729,  822, 853,  573, 604,  914, 945,
+  667, 698,  853, 884,  636, 667,  884, 915,  479, 510,  975, 1006, 542, 573,
+  945, 976,  761, 792,  792, 823,  605, 636,  915, 946,  730, 761,  823, 854,
+  699, 730,  854, 885,  511, 542,  976, 1007, 574, 605,  946, 977,  668, 699,
+  885, 916,  637, 668,  916, 947,  543, 574,  793, 824,  977, 1008, 762, 793,
+  824, 855,  731, 762,  855, 886,  606, 637,  947, 978,  700, 731,  886, 917,
+  669, 700,  917, 948,  575, 606,  978, 1009, 638, 669,  948, 979,  794, 825,
+  825, 856,  763, 794,  856, 887,  732, 763,  887, 918,  607, 638,  979, 1010,
+  701, 732,  918, 949,  670, 701,  949, 980,  826, 857,  795, 826,  857, 888,
+  764, 795,  888, 919,  639, 670,  980, 1011, 733, 764,  919, 950,  702, 733,
+  950, 981,  671, 702,  981, 1012, 827, 858,  858, 889,  796, 827,  889, 920,
+  765, 796,  920, 951,  734, 765,  951, 982,  703, 734,  982, 1013, 859, 890,
+  828, 859,  890, 921,  797, 828,  921, 952,  766, 797,  952, 983,  735, 766,
+  983, 1014, 860, 891,  891, 922,  829, 860,  922, 953,  798, 829,  953, 984,
+  767, 798,  984, 1015, 892, 923,  861, 892,  923, 954,  830, 861,  954, 985,
+  799, 830,  985, 1016, 893, 924,  924, 955,  862, 893,  955, 986,  831, 862,
+  986, 1017, 925, 956,  894, 925,  956, 987,  863, 894,  987, 1018, 926, 957,
+  957, 988,  895, 926,  988, 1019, 958, 989,  927, 958,  989, 1020, 959, 990,
+  990, 1021, 991, 1022, 0,   0
+};
+
+#if CONFIG_TX64X64
+DECLARE_ALIGNED(16, static const int16_t,
+                default_scan_64x64_neighbors[4097 * MAX_NEIGHBORS]) = {
+  0,    0,    0,    0,    0,    0,    1,    64,   1,    1,    64,   64,   2,
+  65,   65,   128,  66,   129,  2,    2,    128,  128,  3,    66,   129,  192,
+  67,   130,  130,  193,  3,    3,    192,  192,  4,    67,   193,  256,  131,
+  194,  68,   131,  194,  257,  4,    4,    132,  195,  195,  258,  256,  256,
+  5,    68,   257,  320,  69,   132,  258,  321,  196,  259,  133,  196,  259,
+  322,  5,    5,    320,  320,  6,    69,   321,  384,  70,   133,  322,  385,
+  197,  260,  260,  323,  134,  197,  323,  386,  6,    6,    384,  384,  7,
+  70,   261,  324,  385,  448,  198,  261,  324,  387,  71,   134,  386,  449,
+  135,  198,  387,  450,  262,  325,  325,  388,  7,    7,    448,  448,  8,
+  71,   199,  262,  388,  451,  449,  512,  72,   135,  450,  513,  326,  389,
+  136,  199,  451,  514,  263,  326,  389,  452,  200,  263,  452,  515,  8,
+  8,    512,  512,  9,    72,   513,  576,  73,   136,  327,  390,  390,  453,
+  514,  577,  264,  327,  453,  516,  137,  200,  515,  578,  201,  264,  516,
+  579,  391,  454,  9,    9,    328,  391,  454,  517,  576,  576,  10,   73,
+  577,  640,  74,   137,  578,  641,  265,  328,  517,  580,  138,  201,  579,
+  642,  392,  455,  455,  518,  202,  265,  580,  643,  329,  392,  518,  581,
+  10,   10,   640,  640,  11,   74,   641,  704,  75,   138,  266,  329,  581,
+  644,  642,  705,  456,  519,  139,  202,  393,  456,  519,  582,  643,  706,
+  330,  393,  582,  645,  203,  266,  644,  707,  11,   11,   704,  704,  12,
+  75,   457,  520,  520,  583,  705,  768,  267,  330,  645,  708,  76,   139,
+  706,  769,  394,  457,  583,  646,  140,  203,  707,  770,  331,  394,  646,
+  709,  204,  267,  708,  771,  521,  584,  458,  521,  584,  647,  12,   12,
+  268,  331,  709,  772,  768,  768,  13,   76,   395,  458,  647,  710,  769,
+  832,  77,   140,  770,  833,  141,  204,  771,  834,  332,  395,  710,  773,
+  522,  585,  585,  648,  205,  268,  459,  522,  648,  711,  772,  835,  396,
+  459,  711,  774,  269,  332,  773,  836,  13,   13,   832,  832,  14,   77,
+  833,  896,  78,   141,  586,  649,  834,  897,  523,  586,  649,  712,  142,
+  205,  333,  396,  774,  837,  835,  898,  460,  523,  712,  775,  206,  269,
+  836,  899,  397,  460,  775,  838,  270,  333,  587,  650,  650,  713,  837,
+  900,  14,   14,   524,  587,  713,  776,  896,  896,  15,   78,   897,  960,
+  79,   142,  898,  961,  334,  397,  838,  901,  461,  524,  776,  839,  143,
+  206,  899,  962,  207,  270,  900,  963,  651,  714,  588,  651,  714,  777,
+  398,  461,  839,  902,  271,  334,  525,  588,  777,  840,  901,  964,  15,
+  15,   960,  960,  16,   79,   961,  1024, 80,   143,  462,  525,  840,  903,
+  962,  1025, 335,  398,  902,  965,  144,  207,  652,  715,  715,  778,  963,
+  1026, 589,  652,  778,  841,  208,  271,  964,  1027, 399,  462,  903,  966,
+  526,  589,  841,  904,  272,  335,  965,  1028, 716,  779,  16,   16,   463,
+  526,  904,  967,  1024, 1024, 17,   80,   653,  716,  779,  842,  1025, 1088,
+  336,  399,  966,  1029, 81,   144,  1026, 1089, 590,  653,  842,  905,  145,
+  208,  1027, 1090, 209,  272,  400,  463,  967,  1030, 1028, 1091, 527,  590,
+  905,  968,  717,  780,  780,  843,  273,  336,  1029, 1092, 654,  717,  843,
+  906,  464,  527,  968,  1031, 17,   17,   1088, 1088, 18,   81,   337,  400,
+  591,  654,  906,  969,  1030, 1093, 1089, 1152, 82,   145,  1090, 1153, 146,
+  209,  1091, 1154, 528,  591,  969,  1032, 401,  464,  781,  844,  1031, 1094,
+  210,  273,  718,  781,  844,  907,  1092, 1155, 655,  718,  907,  970,  274,
+  337,  1093, 1156, 465,  528,  1032, 1095, 592,  655,  970,  1033, 338,  401,
+  1094, 1157, 18,   18,   1152, 1152, 19,   82,   1153, 1216, 83,   146,  782,
+  845,  845,  908,  1154, 1217, 719,  782,  908,  971,  147,  210,  529,  592,
+  1033, 1096, 1155, 1218, 402,  465,  1095, 1158, 211,  274,  656,  719,  971,
+  1034, 1156, 1219, 275,  338,  1157, 1220, 466,  529,  1096, 1159, 593,  656,
+  1034, 1097, 846,  909,  783,  846,  909,  972,  339,  402,  1158, 1221, 19,
+  19,   720,  783,  972,  1035, 1216, 1216, 20,   83,   1217, 1280, 84,   147,
+  1218, 1281, 530,  593,  1097, 1160, 148,  211,  1219, 1282, 403,  466,  657,
+  720,  1035, 1098, 1159, 1222, 212,  275,  1220, 1283, 847,  910,  910,  973,
+  594,  657,  1098, 1161, 276,  339,  467,  530,  784,  847,  973,  1036, 1160,
+  1223, 1221, 1284, 721,  784,  1036, 1099, 340,  403,  1222, 1285, 20,   20,
+  1280, 1280, 21,   84,   531,  594,  1161, 1224, 1281, 1344, 85,   148,  658,
+  721,  1099, 1162, 1282, 1345, 404,  467,  1223, 1286, 149,  212,  911,  974,
+  1283, 1346, 848,  911,  974,  1037, 213,  276,  1284, 1347, 785,  848,  1037,
+  1100, 595,  658,  1162, 1225, 468,  531,  1224, 1287, 277,  340,  1285, 1348,
+  722,  785,  1100, 1163, 341,  404,  1286, 1349, 532,  595,  912,  975,  975,
+  1038, 1225, 1288, 659,  722,  1163, 1226, 21,   21,   1344, 1344, 22,   85,
+  849,  912,  1038, 1101, 1345, 1408, 86,   149,  1346, 1409, 405,  468,  1287,
+  1350, 150,  213,  786,  849,  1101, 1164, 1347, 1410, 214,  277,  596,  659,
+  1226, 1289, 1348, 1411, 469,  532,  723,  786,  1164, 1227, 1288, 1351, 278,
+  341,  1349, 1412, 976,  1039, 913,  976,  1039, 1102, 342,  405,  850,  913,
+  1102, 1165, 1350, 1413, 660,  723,  1227, 1290, 533,  596,  1289, 1352, 22,
+  22,   1408, 1408, 23,   86,   787,  850,  1165, 1228, 1409, 1472, 87,   150,
+  406,  469,  1351, 1414, 1410, 1473, 151,  214,  1411, 1474, 597,  660,  1290,
+  1353, 724,  787,  1228, 1291, 215,  278,  977,  1040, 1040, 1103, 1412, 1475,
+  470,  533,  1352, 1415, 914,  977,  1103, 1166, 279,  342,  1413, 1476, 851,
+  914,  1166, 1229, 661,  724,  1291, 1354, 343,  406,  534,  597,  1353, 1416,
+  1414, 1477, 788,  851,  1229, 1292, 23,   23,   1472, 1472, 24,   87,   1473,
+  1536, 407,  470,  1041, 1104, 1415, 1478, 88,   151,  978,  1041, 1104, 1167,
+  1474, 1537, 598,  661,  1354, 1417, 152,  215,  725,  788,  1292, 1355, 1475,
+  1538, 915,  978,  1167, 1230, 216,  279,  1476, 1539, 471,  534,  1416, 1479,
+  852,  915,  1230, 1293, 280,  343,  1477, 1540, 662,  725,  1355, 1418, 535,
+  598,  789,  852,  1293, 1356, 1417, 1480, 344,  407,  1478, 1541, 1042, 1105,
+  1105, 1168, 979,  1042, 1168, 1231, 24,   24,   408,  471,  916,  979,  1231,
+  1294, 1479, 1542, 1536, 1536, 25,   88,   1537, 1600, 726,  789,  1356, 1419,
+  89,   152,  599,  662,  1418, 1481, 1538, 1601, 153,  216,  1539, 1602, 853,
+  916,  1294, 1357, 472,  535,  1480, 1543, 217,  280,  1540, 1603, 1106, 1169,
+  281,  344,  663,  726,  1043, 1106, 1169, 1232, 1419, 1482, 1541, 1604, 790,
+  853,  1357, 1420, 980,  1043, 1232, 1295, 536,  599,  1481, 1544, 345,  408,
+  1542, 1605, 917,  980,  1295, 1358, 727,  790,  1420, 1483, 409,  472,  1543,
+  1606, 25,   25,   600,  663,  1482, 1545, 1600, 1600, 26,   89,   1601, 1664,
+  90,   153,  854,  917,  1358, 1421, 1602, 1665, 154,  217,  1107, 1170, 1170,
+  1233, 1603, 1666, 473,  536,  1044, 1107, 1233, 1296, 1544, 1607, 218,  281,
+  1604, 1667, 664,  727,  981,  1044, 1296, 1359, 1483, 1546, 791,  854,  1421,
+  1484, 282,  345,  1605, 1668, 537,  600,  1545, 1608, 918,  981,  1359, 1422,
+  346,  409,  1606, 1669, 728,  791,  1484, 1547, 1171, 1234, 1108, 1171, 1234,
+  1297, 410,  473,  601,  664,  855,  918,  1422, 1485, 1546, 1609, 1607, 1670,
+  26,   26,   1664, 1664, 27,   90,   1045, 1108, 1297, 1360, 1665, 1728, 91,
+  154,  1666, 1729, 155,  218,  1667, 1730, 474,  537,  982,  1045, 1360, 1423,
+  1608, 1671, 219,  282,  792,  855,  1485, 1548, 1668, 1731, 665,  728,  1547,
+  1610, 283,  346,  919,  982,  1423, 1486, 1669, 1732, 538,  601,  1609, 1672,
+  1172, 1235, 1235, 1298, 347,  410,  1109, 1172, 1298, 1361, 1670, 1733, 729,
+  792,  1548, 1611, 856,  919,  1486, 1549, 1046, 1109, 1361, 1424, 602,  665,
+  1610, 1673, 411,  474,  1671, 1734, 27,   27,   1728, 1728, 28,   91,   983,
+  1046, 1424, 1487, 1729, 1792, 92,   155,  1730, 1793, 156,  219,  475,  538,
+  1672, 1735, 1731, 1794, 793,  856,  1549, 1612, 666,  729,  1611, 1674, 220,
+  283,  1236, 1299, 1732, 1795, 920,  983,  1487, 1550, 1173, 1236, 1299, 1362,
+  1110, 1173, 1362, 1425, 284,  347,  1733, 1796, 539,  602,  1673, 1736, 1047,
+  1110, 1425, 1488, 348,  411,  730,  793,  1612, 1675, 1734, 1797, 857,  920,
+  1550, 1613, 603,  666,  1674, 1737, 984,  1047, 1488, 1551, 412,  475,  1735,
+  1798, 28,   28,   1237, 1300, 1300, 1363, 1792, 1792, 29,   92,   1793, 1856,
+  93,   156,  794,  857,  1174, 1237, 1363, 1426, 1613, 1676, 1794, 1857, 476,
+  539,  1736, 1799, 157,  220,  667,  730,  921,  984,  1551, 1614, 1675, 1738,
+  1795, 1858, 1111, 1174, 1426, 1489, 221,  284,  1796, 1859, 540,  603,  1048,
+  1111, 1489, 1552, 1737, 1800, 285,  348,  1797, 1860, 858,  921,  1614, 1677,
+  731,  794,  1676, 1739, 349,  412,  1798, 1861, 985,  1048, 1552, 1615, 1301,
+  1364, 604,  667,  1238, 1301, 1364, 1427, 1738, 1801, 413,  476,  1175, 1238,
+  1427, 1490, 1799, 1862, 795,  858,  1677, 1740, 29,   29,   1112, 1175, 1490,
+  1553, 1856, 1856, 30,   93,   922,  985,  1615, 1678, 1857, 1920, 94,   157,
+  1858, 1921, 477,  540,  668,  731,  1739, 1802, 1800, 1863, 158,  221,  1859,
+  1922, 1049, 1112, 1553, 1616, 222,  285,  1860, 1923, 541,  604,  1801, 1864,
+  286,  349,  859,  922,  1302, 1365, 1365, 1428, 1678, 1741, 1861, 1924, 732,
+  795,  1740, 1803, 1239, 1302, 1428, 1491, 986,  1049, 1616, 1679, 350,  413,
+  1862, 1925, 1176, 1239, 1491, 1554, 605,  668,  1802, 1865, 414,  477,  1113,
+  1176, 1554, 1617, 1863, 1926, 796,  859,  1741, 1804, 923,  986,  1679, 1742,
+  30,   30,   1920, 1920, 31,   94,   669,  732,  1803, 1866, 1921, 1984, 478,
+  541,  1864, 1927, 95,   158,  1050, 1113, 1617, 1680, 1922, 1985, 1366, 1429,
+  159,  222,  1303, 1366, 1429, 1492, 1923, 1986, 1240, 1303, 1492, 1555, 223,
+  286,  1924, 1987, 860,  923,  1742, 1805, 542,  605,  1865, 1928, 733,  796,
+  987,  1050, 1680, 1743, 1804, 1867, 287,  350,  1177, 1240, 1555, 1618, 1925,
+  1988, 351,  414,  1926, 1989, 606,  669,  1114, 1177, 1618, 1681, 1866, 1929,
+  924,  987,  1743, 1806, 415,  478,  797,  860,  1805, 1868, 1927, 1990, 1367,
+  1430, 1430, 1493, 1304, 1367, 1493, 1556, 1051, 1114, 1681, 1744, 670,  733,
+  1867, 1930, 31,   31,   1984, 1984, 32,   95,   479,  542,  1241, 1304, 1556,
+  1619, 1928, 1991, 1985, 2048, 96,   159,  1986, 2049, 160,  223,  1987, 2050,
+  861,  924,  1178, 1241, 1619, 1682, 1806, 1869, 224,  287,  988,  1051, 1744,
+  1807, 1988, 2051, 543,  606,  1929, 1992, 734,  797,  1868, 1931, 288,  351,
+  1989, 2052, 1115, 1178, 1682, 1745, 1431, 1494, 352,  415,  1368, 1431, 1494,
+  1557, 1990, 2053, 607,  670,  1930, 1993, 925,  988,  1305, 1368, 1557, 1620,
+  1807, 1870, 798,  861,  1869, 1932, 416,  479,  1052, 1115, 1745, 1808, 1991,
+  2054, 1242, 1305, 1620, 1683, 671,  734,  1931, 1994, 480,  543,  1992, 2055,
+  32,   32,   2048, 2048, 33,   96,   1179, 1242, 1683, 1746, 2049, 2112, 97,
+  160,  2050, 2113, 862,  925,  1870, 1933, 989,  1052, 1808, 1871, 161,  224,
+  2051, 2114, 225,  288,  544,  607,  735,  798,  1432, 1495, 1495, 1558, 1932,
+  1995, 1993, 2056, 2052, 2115, 1116, 1179, 1746, 1809, 1369, 1432, 1558, 1621,
+  289,  352,  2053, 2116, 1306, 1369, 1621, 1684, 608,  671,  1994, 2057, 353,
+  416,  926,  989,  1871, 1934, 2054, 2117, 1243, 1306, 1684, 1747, 799,  862,
+  1053, 1116, 1809, 1872, 1933, 1996, 417,  480,  2055, 2118, 672,  735,  1180,
+  1243, 1747, 1810, 1995, 2058, 1496, 1559, 481,  544,  2056, 2119, 1433, 1496,
+  1559, 1622, 33,   33,   990,  1053, 1872, 1935, 2112, 2112, 34,   97,   863,
+  926,  1934, 1997, 2113, 2176, 98,   161,  1370, 1433, 1622, 1685, 2114, 2177,
+  162,  225,  1117, 1180, 1810, 1873, 2115, 2178, 736,  799,  1996, 2059, 545,
+  608,  1307, 1370, 1685, 1748, 2057, 2120, 226,  289,  2116, 2179, 290,  353,
+  2117, 2180, 1244, 1307, 1748, 1811, 927,  990,  1935, 1998, 609,  672,  1054,
+  1117, 1873, 1936, 2058, 2121, 354,  417,  2118, 2181, 800,  863,  1997, 2060,
+  1497, 1560, 1560, 1623, 1181, 1244, 1811, 1874, 418,  481,  1434, 1497, 1623,
+  1686, 2119, 2182, 673,  736,  2059, 2122, 1371, 1434, 1686, 1749, 991,  1054,
+  1936, 1999, 482,  545,  864,  927,  1998, 2061, 2120, 2183, 1118, 1181, 1874,
+  1937, 34,   34,   1308, 1371, 1749, 1812, 2176, 2176, 35,   98,   2177, 2240,
+  99,   162,  2178, 2241, 737,  800,  2060, 2123, 163,  226,  2179, 2242, 546,
+  609,  2121, 2184, 227,  290,  1245, 1308, 1812, 1875, 2180, 2243, 928,  991,
+  1999, 2062, 291,  354,  1055, 1118, 1561, 1624, 1937, 2000, 2181, 2244, 1498,
+  1561, 1624, 1687, 610,  673,  2122, 2185, 801,  864,  1435, 1498, 1687, 1750,
+  2061, 2124, 355,  418,  1182, 1245, 1875, 1938, 2182, 2245, 1372, 1435, 1750,
+  1813, 419,  482,  2183, 2246, 674,  737,  2123, 2186, 992,  1055, 2000, 2063,
+  1309, 1372, 1813, 1876, 865,  928,  1119, 1182, 1938, 2001, 2062, 2125, 483,
+  546,  2184, 2247, 35,   35,   2240, 2240, 36,   99,   2241, 2304, 100,  163,
+  738,  801,  1246, 1309, 1876, 1939, 2124, 2187, 2242, 2305, 1562, 1625, 1625,
+  1688, 164,  227,  1499, 1562, 1688, 1751, 2243, 2306, 547,  610,  2185, 2248,
+  228,  291,  2244, 2307, 1056, 1119, 1436, 1499, 1751, 1814, 2001, 2064, 929,
+  992,  2063, 2126, 292,  355,  2245, 2308, 1183, 1246, 1939, 2002, 611,  674,
+  802,  865,  1373, 1436, 1814, 1877, 2125, 2188, 2186, 2249, 356,  419,  2246,
+  2309, 1310, 1373, 1877, 1940, 420,  483,  993,  1056, 2064, 2127, 2247, 2310,
+  675,  738,  2187, 2250, 1120, 1183, 2002, 2065, 866,  929,  1626, 1689, 2126,
+  2189, 1563, 1626, 1689, 1752, 484,  547,  1500, 1563, 1752, 1815, 2248, 2311,
+  1247, 1310, 1940, 2003, 36,   36,   739,  802,  2188, 2251, 2304, 2304, 37,
+  100,  1437, 1500, 1815, 1878, 2305, 2368, 101,  164,  2306, 2369, 548,  611,
+  2249, 2312, 165,  228,  1057, 1120, 2065, 2128, 2307, 2370, 930,  993,  2127,
+  2190, 1374, 1437, 1878, 1941, 229,  292,  1184, 1247, 2003, 2066, 2308, 2371,
+  293,  356,  803,  866,  2189, 2252, 2309, 2372, 612,  675,  2250, 2313, 1311,
+  1374, 1941, 2004, 357,  420,  1627, 1690, 1690, 1753, 2310, 2373, 1564, 1627,
+  1753, 1816, 994,  1057, 2128, 2191, 1121, 1184, 2066, 2129, 676,  739,  1501,
+  1564, 1816, 1879, 2251, 2314, 421,  484,  2311, 2374, 867,  930,  2190, 2253,
+  1248, 1311, 2004, 2067, 1438, 1501, 1879, 1942, 485,  548,  2312, 2375, 740,
+  803,  2252, 2315, 37,   37,   2368, 2368, 38,   101,  1058, 1121, 1375, 1438,
+  1942, 2005, 2129, 2192, 2369, 2432, 102,  165,  2370, 2433, 549,  612,  931,
+  994,  1185, 1248, 2067, 2130, 2191, 2254, 2313, 2376, 166,  229,  2371, 2434,
+  1691, 1754, 230,  293,  1628, 1691, 1754, 1817, 2372, 2435, 804,  867,  1312,
+  1375, 2005, 2068, 2253, 2316, 1565, 1628, 1817, 1880, 294,  357,  613,  676,
+  2314, 2377, 2373, 2436, 1502, 1565, 1880, 1943, 358,  421,  1122, 1185, 2130,
+  2193, 2374, 2437, 995,  1058, 2192, 2255, 1249, 1312, 2068, 2131, 677,  740,
+  1439, 1502, 1943, 2006, 2315, 2378, 868,  931,  2254, 2317, 422,  485,  2375,
+  2438, 486,  549,  1376, 1439, 2006, 2069, 2376, 2439, 741,  804,  1692, 1755,
+  1755, 1818, 2316, 2379, 1059, 1122, 2193, 2256, 1186, 1249, 1629, 1692, 1818,
+  1881, 2131, 2194, 38,   38,   932,  995,  2255, 2318, 2432, 2432, 39,   102,
+  2433, 2496, 103,  166,  550,  613,  1566, 1629, 1881, 1944, 2377, 2440, 2434,
+  2497, 167,  230,  1313, 1376, 2069, 2132, 2435, 2498, 231,  294,  1503, 1566,
+  1944, 2007, 2436, 2499, 805,  868,  2317, 2380, 614,  677,  2378, 2441, 295,
+  358,  2437, 2500, 1123, 1186, 2194, 2257, 996,  1059, 2256, 2319, 1440, 1503,
+  2007, 2070, 1250, 1313, 2132, 2195, 359,  422,  2438, 2501, 678,  741,  869,
+  932,  2318, 2381, 2379, 2442, 1756, 1819, 423,  486,  1693, 1756, 1819, 1882,
+  2439, 2502, 1377, 1440, 2070, 2133, 1630, 1693, 1882, 1945, 487,  550,  1060,
+  1123, 2257, 2320, 2440, 2503, 1187, 1250, 1567, 1630, 1945, 2008, 2195, 2258,
+  742,  805,  2380, 2443, 933,  996,  2319, 2382, 1314, 1377, 2133, 2196, 39,
+  39,   1504, 1567, 2008, 2071, 2496, 2496, 40,   103,  2497, 2560, 551,  614,
+  2441, 2504, 104,  167,  2498, 2561, 168,  231,  2499, 2562, 806,  869,  2381,
+  2444, 232,  295,  2500, 2563, 1441, 1504, 2071, 2134, 1124, 1187, 2258, 2321,
+  615,  678,  2442, 2505, 296,  359,  997,  1060, 1251, 1314, 1757, 1820, 1820,
+  1883, 2196, 2259, 2320, 2383, 2501, 2564, 1694, 1757, 1883, 1946, 360,  423,
+  2502, 2565, 1631, 1694, 1946, 2009, 870,  933,  1378, 1441, 2134, 2197, 2382,
+  2445, 679,  742,  2443, 2506, 424,  487,  1568, 1631, 2009, 2072, 2503, 2566,
+  1188, 1251, 2259, 2322, 1061, 1124, 2321, 2384, 488,  551,  2504, 2567, 743,
+  806,  1505, 1568, 2072, 2135, 2444, 2507, 1315, 1378, 2197, 2260, 934,  997,
+  2383, 2446, 40,   40,   552,  615,  2505, 2568, 2560, 2560, 41,   104,  1821,
+  1884, 2561, 2624, 1758, 1821, 1884, 1947, 105,  168,  1442, 1505, 2135, 2198,
+  2562, 2625, 169,  232,  807,  870,  1695, 1758, 1947, 2010, 2445, 2508, 2563,
+  2626, 1125, 1188, 2322, 2385, 1252, 1315, 2260, 2323, 233,  296,  2564, 2627,
+  616,  679,  998,  1061, 1632, 1695, 2010, 2073, 2384, 2447, 2506, 2569, 297,
+  360,  2565, 2628, 1379, 1442, 2198, 2261, 1569, 1632, 2073, 2136, 361,  424,
+  871,  934,  2446, 2509, 2566, 2629, 680,  743,  2507, 2570, 425,  488,  1189,
+  1252, 2323, 2386, 2567, 2630, 1506, 1569, 2136, 2199, 1062, 1125, 2385, 2448,
+  1316, 1379, 2261, 2324, 1822, 1885, 1885, 1948, 744,  807,  2508, 2571, 489,
+  552,  1759, 1822, 1948, 2011, 2568, 2631, 935,  998,  2447, 2510, 1696, 1759,
+  2011, 2074, 1443, 1506, 2199, 2262, 553,  616,  2569, 2632, 41,   41,   2624,
+  2624, 42,   105,  1633, 1696, 2074, 2137, 2625, 2688, 106,  169,  1126, 1189,
+  2386, 2449, 2626, 2689, 808,  871,  1253, 1316, 2324, 2387, 2509, 2572, 170,
+  233,  2627, 2690, 999,  1062, 2448, 2511, 234,  297,  1380, 1443, 2262, 2325,
+  2628, 2691, 617,  680,  1570, 1633, 2137, 2200, 2570, 2633, 298,  361,  2629,
+  2692, 872,  935,  2510, 2573, 362,  425,  1886, 1949, 2630, 2693, 1507, 1570,
+  2200, 2263, 681,  744,  1823, 1886, 1949, 2012, 2571, 2634, 1190, 1253, 2387,
+  2450, 1760, 1823, 2012, 2075, 1063, 1126, 1317, 1380, 2325, 2388, 2449, 2512,
+  426,  489,  2631, 2694, 1697, 1760, 2075, 2138, 745,  808,  936,  999,  1444,
+  1507, 2263, 2326, 2511, 2574, 2572, 2635, 490,  553,  2632, 2695, 1634, 1697,
+  2138, 2201, 1254, 1317, 2388, 2451, 554,  617,  1127, 1190, 2450, 2513, 2633,
+  2696, 42,   42,   2688, 2688, 43,   106,  809,  872,  1571, 1634, 2201, 2264,
+  2573, 2636, 2689, 2752, 107,  170,  1381, 1444, 2326, 2389, 2690, 2753, 1000,
+  1063, 2512, 2575, 171,  234,  2691, 2754, 1887, 1950, 1950, 2013, 618,  681,
+  2634, 2697, 235,  298,  1824, 1887, 2013, 2076, 2692, 2755, 1508, 1571, 2264,
+  2327, 1761, 1824, 2076, 2139, 299,  362,  2693, 2756, 873,  936,  2574, 2637,
+  1191, 1254, 2451, 2514, 363,  426,  682,  745,  1318, 1381, 1698, 1761, 2139,
+  2202, 2389, 2452, 2635, 2698, 2694, 2757, 1064, 1127, 2513, 2576, 427,  490,
+  1445, 1508, 2327, 2390, 2695, 2758, 1635, 1698, 2202, 2265, 937,  1000, 2575,
+  2638, 746,  809,  2636, 2699, 491,  554,  2696, 2759, 1255, 1318, 1572, 1635,
+  2265, 2328, 2452, 2515, 1951, 2014, 1128, 1191, 1888, 1951, 2014, 2077, 2514,
+  2577, 1382, 1445, 2390, 2453, 555,  618,  1825, 1888, 2077, 2140, 2697, 2760,
+  810,  873,  2637, 2700, 43,   43,   2752, 2752, 44,   107,  1001, 1064, 2576,
+  2639, 2753, 2816, 108,  171,  1762, 1825, 2140, 2203, 2754, 2817, 172,  235,
+  1509, 1572, 2328, 2391, 2755, 2818, 619,  682,  2698, 2761, 236,  299,  2756,
+  2819, 1699, 1762, 2203, 2266, 874,  937,  2638, 2701, 300,  363,  1192, 1255,
+  2515, 2578, 2757, 2820, 1319, 1382, 2453, 2516, 683,  746,  1065, 1128, 2577,
+  2640, 2699, 2762, 364,  427,  1636, 1699, 2266, 2329, 2758, 2821, 1446, 1509,
+  2391, 2454, 428,  491,  1952, 2015, 2015, 2078, 2759, 2822, 938,  1001, 1889,
+  1952, 2078, 2141, 2639, 2702, 747,  810,  2700, 2763, 1573, 1636, 2329, 2392,
+  1826, 1889, 2141, 2204, 492,  555,  1256, 1319, 2516, 2579, 2760, 2823, 1129,
+  1192, 1383, 1446, 2454, 2517, 2578, 2641, 1763, 1826, 2204, 2267, 556,  619,
+  2761, 2824, 811,  874,  2701, 2764, 1002, 1065, 1510, 1573, 2392, 2455, 2640,
+  2703, 44,   44,   1700, 1763, 2267, 2330, 2816, 2816, 45,   108,  2817, 2880,
+  109,  172,  2818, 2881, 173,  236,  2819, 2882, 620,  683,  2762, 2825, 237,
+  300,  1320, 1383, 2517, 2580, 2820, 2883, 1193, 1256, 2579, 2642, 875,  938,
+  1637, 1700, 2330, 2393, 2702, 2765, 2016, 2079, 301,  364,  1447, 1510, 1953,
+  2016, 2079, 2142, 2455, 2518, 2821, 2884, 1066, 1129, 2641, 2704, 1890, 1953,
+  2142, 2205, 684,  747,  2763, 2826, 365,  428,  2822, 2885, 1827, 1890, 2205,
+  2268, 1574, 1637, 2393, 2456, 429,  492,  939,  1002, 2703, 2766, 2823, 2886,
+  748,  811,  1764, 1827, 2268, 2331, 2764, 2827, 1257, 1320, 2580, 2643, 1384,
+  1447, 2518, 2581, 1130, 1193, 2642, 2705, 493,  556,  2824, 2887, 1511, 1574,
+  2456, 2519, 1701, 1764, 2331, 2394, 812,  875,  1003, 1066, 2704, 2767, 2765,
+  2828, 557,  620,  2825, 2888, 2017, 2080, 2080, 2143, 45,   45,   2880, 2880,
+  46,   109,  1954, 2017, 2143, 2206, 2881, 2944, 110,  173,  1638, 1701, 2394,
+  2457, 2882, 2945, 1321, 1384, 2581, 2644, 174,  237,  621,  684,  1194, 1257,
+  1891, 1954, 2206, 2269, 2643, 2706, 2826, 2889, 2883, 2946, 1448, 1511, 2519,
+  2582, 238,  301,  876,  939,  2766, 2829, 2884, 2947, 1828, 1891, 2269, 2332,
+  1067, 1130, 2705, 2768, 302,  365,  2885, 2948, 685,  748,  1575, 1638, 2457,
+  2520, 2827, 2890, 366,  429,  2886, 2949, 1765, 1828, 2332, 2395, 940,  1003,
+  2767, 2830, 1258, 1321, 2644, 2707, 430,  493,  1385, 1448, 2582, 2645, 2887,
+  2950, 749,  812,  2828, 2891, 1131, 1194, 1702, 1765, 2395, 2458, 2706, 2769,
+  1512, 1575, 2520, 2583, 2081, 2144, 494,  557,  2018, 2081, 2144, 2207, 2888,
+  2951, 1955, 2018, 2207, 2270, 1004, 1067, 2768, 2831, 813,  876,  2829, 2892,
+  1892, 1955, 2270, 2333, 558,  621,  1639, 1702, 2458, 2521, 2889, 2952, 1322,
+  1385, 2645, 2708, 46,   46,   2944, 2944, 47,   110,  1195, 1258, 1449, 1512,
+  1829, 1892, 2333, 2396, 2583, 2646, 2707, 2770, 2945, 3008, 111,  174,  2946,
+  3009, 622,  685,  2890, 2953, 175,  238,  2947, 3010, 877,  940,  2830, 2893,
+  239,  302,  1068, 1131, 1576, 1639, 2521, 2584, 2769, 2832, 2948, 3011, 1766,
+  1829, 2396, 2459, 303,  366,  2949, 3012, 686,  749,  2891, 2954, 367,  430,
+  2082, 2145, 2145, 2208, 2950, 3013, 1386, 1449, 2646, 2709, 1259, 1322, 2019,
+  2082, 2208, 2271, 2708, 2771, 941,  1004, 1703, 1766, 2459, 2522, 2831, 2894,
+  1513, 1576, 1956, 2019, 2271, 2334, 2584, 2647, 431,  494,  2951, 3014, 750,
+  813,  1132, 1195, 2770, 2833, 2892, 2955, 1893, 1956, 2334, 2397, 495,  558,
+  2952, 3015, 1640, 1703, 2522, 2585, 1005, 1068, 2832, 2895, 814,  877,  1830,
+  1893, 2397, 2460, 2893, 2956, 559,  622,  1323, 1386, 2709, 2772, 2953, 3016,
+  1450, 1513, 2647, 2710, 1196, 1259, 2771, 2834, 47,   47,   3008, 3008, 48,
+  111,  1767, 1830, 2460, 2523, 3009, 3072, 1577, 1640, 2585, 2648, 112,  175,
+  3010, 3073, 623,  686,  2954, 3017, 878,  941,  2146, 2209, 2894, 2957, 176,
+  239,  3011, 3074, 1069, 1132, 2083, 2146, 2209, 2272, 2833, 2896, 240,  303,
+  2020, 2083, 2272, 2335, 3012, 3075, 304,  367,  1704, 1767, 2523, 2586, 3013,
+  3076, 687,  750,  1957, 2020, 2335, 2398, 2955, 3018, 1387, 1450, 2710, 2773,
+  1260, 1323, 2772, 2835, 368,  431,  1514, 1577, 2648, 2711, 3014, 3077, 942,
+  1005, 2895, 2958, 1894, 1957, 2398, 2461, 1133, 1196, 2834, 2897, 432,  495,
+  751,  814,  2956, 3019, 3015, 3078, 1641, 1704, 2586, 2649, 1831, 1894, 2461,
+  2524, 496,  559,  3016, 3079, 1006, 1069, 2896, 2959, 1324, 1387, 2773, 2836,
+  815,  878,  1451, 1514, 2711, 2774, 2957, 3020, 2147, 2210, 2210, 2273, 1768,
+  1831, 2524, 2587, 560,  623,  2084, 2147, 2273, 2336, 3017, 3080, 1197, 1260,
+  2835, 2898, 1578, 1641, 2649, 2712, 2021, 2084, 2336, 2399, 48,   48,   3072,
+  3072, 49,   112,  3073, 3136, 624,  687,  3018, 3081, 113,  176,  879,  942,
+  1070, 1133, 1958, 2021, 2399, 2462, 2897, 2960, 2958, 3021, 3074, 3137, 177,
+  240,  1705, 1768, 2587, 2650, 3075, 3138, 241,  304,  3076, 3139, 1388, 1451,
+  2774, 2837, 1895, 1958, 2462, 2525, 688,  751,  1261, 1324, 1515, 1578, 2712,
+  2775, 2836, 2899, 3019, 3082, 305,  368,  3077, 3140, 943,  1006, 2959, 3022,
+  369,  432,  3078, 3141, 1134, 1197, 1642, 1705, 2650, 2713, 2898, 2961, 1832,
+  1895, 2525, 2588, 752,  815,  3020, 3083, 433,  496,  2211, 2274, 3079, 3142,
+  2148, 2211, 2274, 2337, 2085, 2148, 2337, 2400, 497,  560,  1007, 1070, 1452,
+  1515, 1769, 1832, 2588, 2651, 2775, 2838, 2960, 3023, 3080, 3143, 1325, 1388,
+  2837, 2900, 2022, 2085, 2400, 2463, 816,  879,  3021, 3084, 1579, 1642, 2713,
+  2776, 1198, 1261, 2899, 2962, 561,  624,  1959, 2022, 2463, 2526, 3081, 3144,
+  1706, 1769, 2651, 2714, 1071, 1134, 2961, 3024, 49,   49,   880,  943,  1896,
+  1959, 2526, 2589, 3022, 3085, 3136, 3136, 50,   113,  625,  688,  3082, 3145,
+  3137, 3200, 114,  177,  3138, 3201, 178,  241,  1389, 1452, 2838, 2901, 3139,
+  3202, 1516, 1579, 2776, 2839, 242,  305,  1262, 1325, 2900, 2963, 3140, 3203,
+  2212, 2275, 2275, 2338, 689,  752,  1833, 1896, 2589, 2652, 3083, 3146, 306,
+  369,  1643, 1706, 2149, 2212, 2338, 2401, 2714, 2777, 3141, 3204, 944,  1007,
+  3023, 3086, 1135, 1198, 2086, 2149, 2401, 2464, 2962, 3025, 370,  433,  3142,
+  3205, 753,  816,  2023, 2086, 2464, 2527, 3084, 3147, 1770, 1833, 2652, 2715,
+  434,  497,  3143, 3206, 1453, 1516, 2839, 2902, 1326, 1389, 2901, 2964, 1008,
+  1071, 3024, 3087, 1580, 1643, 1960, 2023, 2527, 2590, 2777, 2840, 498,  561,
+  3144, 3207, 817,  880,  1199, 1262, 2963, 3026, 3085, 3148, 1707, 1770, 2715,
+  2778, 562,  625,  1897, 1960, 2590, 2653, 3145, 3208, 2276, 2339, 1072, 1135,
+  3025, 3088, 2213, 2276, 2339, 2402, 881,  944,  3086, 3149, 626,  689,  1390,
+  1453, 2150, 2213, 2402, 2465, 2902, 2965, 3146, 3209, 50,   50,   1517, 1580,
+  2840, 2903, 3200, 3200, 51,   114,  3201, 3264, 115,  178,  1834, 1897, 2653,
+  2716, 3202, 3265, 1263, 1326, 2964, 3027, 179,  242,  2087, 2150, 2465, 2528,
+  3203, 3266, 1644, 1707, 2778, 2841, 243,  306,  3204, 3267, 690,  753,  3147,
+  3210, 2024, 2087, 2528, 2591, 307,  370,  945,  1008, 3087, 3150, 3205, 3268,
+  1136, 1199, 3026, 3089, 1771, 1834, 2716, 2779, 371,  434,  3206, 3269, 1961,
+  2024, 2591, 2654, 754,  817,  3148, 3211, 1454, 1517, 2903, 2966, 435,  498,
+  1327, 1390, 1581, 1644, 2841, 2904, 2965, 3028, 3207, 3270, 1009, 1072, 3088,
+  3151, 1898, 1961, 2654, 2717, 499,  562,  1200, 1263, 1708, 1771, 2277, 2340,
+  2340, 2403, 2779, 2842, 3027, 3090, 3208, 3271, 818,  881,  2214, 2277, 2403,
+  2466, 3149, 3212, 2151, 2214, 2466, 2529, 563,  626,  3209, 3272, 2088, 2151,
+  2529, 2592, 1073, 1136, 1835, 1898, 2717, 2780, 3089, 3152, 1518, 1581, 2904,
+  2967, 1391, 1454, 2966, 3029, 882,  945,  3150, 3213, 627,  690,  1645, 1708,
+  2842, 2905, 3210, 3273, 51,   51,   1264, 1327, 3028, 3091, 3264, 3264, 52,
+  115,  2025, 2088, 2592, 2655, 3265, 3328, 116,  179,  3266, 3329, 180,  243,
+  3267, 3330, 244,  307,  1772, 1835, 2780, 2843, 3268, 3331, 691,  754,  3211,
+  3274, 946,  1009, 1137, 1200, 1962, 2025, 2655, 2718, 3090, 3153, 3151, 3214,
+  308,  371,  3269, 3332, 1455, 1518, 2341, 2404, 2967, 3030, 372,  435,  2278,
+  2341, 2404, 2467, 3270, 3333, 1582, 1645, 2905, 2968, 755,  818,  1328, 1391,
+  3029, 3092, 3212, 3275, 2215, 2278, 2467, 2530, 1899, 1962, 2718, 2781, 436,
+  499,  3271, 3334, 1709, 1772, 2843, 2906, 1010, 1073, 2152, 2215, 2530, 2593,
+  3152, 3215, 1201, 1264, 3091, 3154, 500,  563,  3272, 3335, 819,  882,  2089,
+  2152, 2593, 2656, 3213, 3276, 1836, 1899, 2781, 2844, 564,  627,  1519, 1582,
+  2968, 3031, 3273, 3336, 1392, 1455, 2026, 2089, 2656, 2719, 3030, 3093, 1074,
+  1137, 3153, 3216, 1646, 1709, 2906, 2969, 883,  946,  3214, 3277, 1265, 1328,
+  3092, 3155, 628,  691,  3274, 3337, 52,   52,   1773, 1836, 2844, 2907, 3328,
+  3328, 53,   116,  1963, 2026, 2719, 2782, 3329, 3392, 117,  180,  2342, 2405,
+  2405, 2468, 3330, 3393, 2279, 2342, 2468, 2531, 181,  244,  3331, 3394, 1138,
+  1201, 3154, 3217, 245,  308,  692,  755,  2216, 2279, 2531, 2594, 3275, 3338,
+  3332, 3395, 947,  1010, 3215, 3278, 1456, 1519, 3031, 3094, 309,  372,  1583,
+  1646, 2969, 3032, 3333, 3396, 1900, 1963, 2782, 2845, 2153, 2216, 2594, 2657,
+  1329, 1392, 3093, 3156, 373,  436,  1710, 1773, 2907, 2970, 3334, 3397, 756,
+  819,  3276, 3339, 2090, 2153, 2657, 2720, 1011, 1074, 3216, 3279, 437,  500,
+  3335, 3398, 1202, 1265, 3155, 3218, 1837, 1900, 2845, 2908, 501,  564,  820,
+  883,  2027, 2090, 2720, 2783, 3277, 3340, 3336, 3399, 1520, 1583, 3032, 3095,
+  1393, 1456, 1647, 1710, 2970, 3033, 3094, 3157, 2406, 2469, 565,  628,  1075,
+  1138, 2343, 2406, 2469, 2532, 3217, 3280, 3337, 3400, 2280, 2343, 2532, 2595,
+  1964, 2027, 2783, 2846, 884,  947,  1266, 1329, 1774, 1837, 2908, 2971, 3156,
+  3219, 3278, 3341, 2217, 2280, 2595, 2658, 629,  692,  3338, 3401, 53,   53,
+  3392, 3392, 54,   117,  3393, 3456, 118,  181,  2154, 2217, 2658, 2721, 3394,
+  3457, 182,  245,  1139, 1202, 1901, 1964, 2846, 2909, 3218, 3281, 3395, 3458,
+  948,  1011, 1584, 1647, 3033, 3096, 3279, 3342, 693,  756,  1457, 1520, 3095,
+  3158, 3339, 3402, 246,  309,  3396, 3459, 1711, 1774, 2091, 2154, 2721, 2784,
+  2971, 3034, 310,  373,  1330, 1393, 3157, 3220, 3397, 3460, 374,  437,  3398,
+  3461, 757,  820,  3340, 3403, 1838, 1901, 2909, 2972, 1012, 1075, 2028, 2091,
+  2784, 2847, 3280, 3343, 1203, 1266, 3219, 3282, 438,  501,  2407, 2470, 2470,
+  2533, 3399, 3462, 2344, 2407, 2533, 2596, 1521, 1584, 2281, 2344, 2596, 2659,
+  3096, 3159, 821,  884,  3341, 3404, 502,  565,  1648, 1711, 3034, 3097, 3400,
+  3463, 1394, 1457, 3158, 3221, 1965, 2028, 2847, 2910, 2218, 2281, 2659, 2722,
+  1076, 1139, 1775, 1838, 2972, 3035, 3281, 3344, 566,  629,  3401, 3464, 1267,
+  1330, 3220, 3283, 885,  948,  2155, 2218, 2722, 2785, 3342, 3405, 630,  693,
+  1902, 1965, 2910, 2973, 3402, 3465, 54,   54,   2092, 2155, 2785, 2848, 3456,
+  3456, 55,   118,  1585, 1648, 3097, 3160, 3457, 3520, 1140, 1203, 3282, 3345,
+  119,  182,  1458, 1521, 3159, 3222, 3458, 3521, 1712, 1775, 3035, 3098, 183,
+  246,  949,  1012, 3343, 3406, 3459, 3522, 694,  757,  3403, 3466, 247,  310,
+  3460, 3523, 1331, 1394, 2471, 2534, 3221, 3284, 2408, 2471, 2534, 2597, 2029,
+  2092, 2848, 2911, 311,  374,  1839, 1902, 2345, 2408, 2597, 2660, 2973, 3036,
+  3461, 3524, 758,  821,  2282, 2345, 2660, 2723, 3404, 3467, 375,  438,  3462,
+  3525, 1013, 1076, 1204, 1267, 3283, 3346, 3344, 3407, 439,  502,  2219, 2282,
+  2723, 2786, 3463, 3526, 1522, 1585, 3160, 3223, 1649, 1712, 1966, 2029, 2911,
+  2974, 3098, 3161, 822,  885,  1395, 1458, 3222, 3285, 3405, 3468, 1776, 1839,
+  3036, 3099, 503,  566,  3464, 3527, 2156, 2219, 2786, 2849, 1077, 1140, 3345,
+  3408, 1268, 1331, 3284, 3347, 567,  630,  3465, 3528, 1903, 1966, 2974, 3037,
+  886,  949,  3406, 3469, 2093, 2156, 2849, 2912, 2472, 2535, 2535, 2598, 631,
+  694,  1586, 1649, 2409, 2472, 2598, 2661, 3161, 3224, 3466, 3529, 1459, 1522,
+  1713, 1776, 3099, 3162, 3223, 3286, 1141, 1204, 2346, 2409, 2661, 2724, 3346,
+  3409, 55,   55,   3520, 3520, 56,   119,  3521, 3584, 120,  183,  2030, 2093,
+  2912, 2975, 3522, 3585, 950,  1013, 3407, 3470, 184,  247,  1332, 1395, 1840,
+  1903, 2283, 2346, 2724, 2787, 3037, 3100, 3285, 3348, 3523, 3586, 695,  758,
+  3467, 3530, 248,  311,  3524, 3587, 312,  375,  2220, 2283, 2787, 2850, 3525,
+  3588, 759,  822,  3468, 3531, 1205, 1268, 1967, 2030, 2975, 3038, 3347, 3410,
+  376,  439,  1014, 1077, 3408, 3471, 3526, 3589, 1650, 1713, 3162, 3225, 1523,
+  1586, 3224, 3287, 2157, 2220, 2850, 2913, 440,  503,  1777, 1840, 3100, 3163,
+  3527, 3590, 1396, 1459, 3286, 3349, 823,  886,  3469, 3532, 504,  567,  2536,
+  2599, 3528, 3591, 2473, 2536, 2599, 2662, 1904, 1967, 3038, 3101, 1078, 1141,
+  2094, 2157, 2913, 2976, 3409, 3472, 2410, 2473, 2662, 2725, 1269, 1332, 3348,
+  3411, 568,  631,  3529, 3592, 2347, 2410, 2725, 2788, 887,  950,  3470, 3533,
+  1587, 1650, 3225, 3288, 1714, 1777, 3163, 3226, 2284, 2347, 2788, 2851, 1460,
+  1523, 2031, 2094, 2976, 3039, 3287, 3350, 632,  695,  3530, 3593, 1142, 1205,
+  3410, 3473, 1841, 1904, 3101, 3164, 56,   56,   3584, 3584, 57,   120,  951,
+  1014, 1333, 1396, 2221, 2284, 2851, 2914, 3349, 3412, 3471, 3534, 3585, 3648,
+  121,  184,  3586, 3649, 696,  759,  3531, 3594, 185,  248,  3587, 3650, 249,
+  312,  1968, 2031, 3039, 3102, 3588, 3651, 2158, 2221, 2914, 2977, 313,  376,
+  3589, 3652, 1206, 1269, 1651, 1714, 3226, 3289, 3411, 3474, 760,  823,  1524,
+  1587, 3288, 3351, 3532, 3595, 1015, 1078, 2537, 2600, 2600, 2663, 3472, 3535,
+  1778, 1841, 3164, 3227, 377,  440,  2474, 2537, 2663, 2726, 3590, 3653, 1397,
+  1460, 2411, 2474, 2726, 2789, 3350, 3413, 441,  504,  2095, 2158, 2977, 3040,
+  3591, 3654, 1905, 1968, 3102, 3165, 824,  887,  2348, 2411, 2789, 2852, 3533,
+  3596, 505,  568,  3592, 3655, 1079, 1142, 3473, 3536, 1270, 1333, 3412, 3475,
+  2285, 2348, 2852, 2915, 2032, 2095, 3040, 3103, 1588, 1651, 3289, 3352, 569,
+  632,  1715, 1778, 3227, 3290, 3593, 3656, 888,  951,  3534, 3597, 1461, 1524,
+  3351, 3414, 1842, 1905, 2222, 2285, 2915, 2978, 3165, 3228, 633,  696,  1143,
+  1206, 3474, 3537, 3594, 3657, 1334, 1397, 3413, 3476, 952,  1015, 3535, 3598,
+  1969, 2032, 2601, 2664, 3103, 3166, 57,   57,   2538, 2601, 2664, 2727, 3648,
+  3648, 58,   121,  2159, 2222, 2978, 3041, 3649, 3712, 122,  185,  3650, 3713,
+  697,  760,  2475, 2538, 2727, 2790, 3595, 3658, 186,  249,  3651, 3714, 250,
+  313,  1652, 1715, 2412, 2475, 2790, 2853, 3290, 3353, 3652, 3715, 1525, 1588,
+  1779, 1842, 3228, 3291, 3352, 3415, 1207, 1270, 3475, 3538, 314,  377,  3653,
+  3716, 1016, 1079, 3536, 3599, 761,  824,  2096, 2159, 3041, 3104, 3596, 3659,
+  2349, 2412, 2853, 2916, 378,  441,  1398, 1461, 1906, 1969, 3166, 3229, 3414,
+  3477, 3654, 3717, 2286, 2349, 2916, 2979, 442,  505,  3655, 3718, 825,  888,
+  3597, 3660, 1080, 1143, 1271, 1334, 2033, 2096, 3104, 3167, 3476, 3539, 3537,
+  3600, 506,  569,  3656, 3719, 1716, 1779, 3291, 3354, 1589, 1652, 2223, 2286,
+  2979, 3042, 3353, 3416, 1843, 1906, 3229, 3292, 570,  633,  889,  952,  1462,
+  1525, 2602, 2665, 2665, 2728, 3415, 3478, 3598, 3661, 3657, 3720, 2539, 2602,
+  2728, 2791, 2476, 2539, 2791, 2854, 1144, 1207, 2160, 2223, 3042, 3105, 3538,
+  3601, 1970, 2033, 3167, 3230, 634,  697,  3658, 3721, 1335, 1398, 3477, 3540,
+  2413, 2476, 2854, 2917, 953,  1016, 3599, 3662, 58,   58,   3712, 3712, 59,
+  122,  3713, 3776, 123,  186,  698,  761,  1653, 1716, 2350, 2413, 2917, 2980,
+  3354, 3417, 3659, 3722, 3714, 3777, 1780, 1843, 3292, 3355, 187,  250,  2097,
+  2160, 3105, 3168, 3715, 3778, 1526, 1589, 3416, 3479, 251,  314,  1208, 1271,
+  3539, 3602, 3716, 3779, 1907, 1970, 3230, 3293, 1017, 1080, 2287, 2350, 2980,
+  3043, 3600, 3663, 315,  378,  3717, 3780, 762,  825,  3660, 3723, 1399, 1462,
+  3478, 3541, 379,  442,  3718, 3781, 2034, 2097, 3168, 3231, 2666, 2729, 2224,
+  2287, 3043, 3106, 443,  506,  2603, 2666, 2729, 2792, 3719, 3782, 826,  889,
+  3661, 3724, 1272, 1335, 2540, 2603, 2792, 2855, 3540, 3603, 1081, 1144, 1717,
+  1780, 3355, 3418, 3601, 3664, 1590, 1653, 3417, 3480, 507,  570,  1844, 1907,
+  3293, 3356, 3720, 3783, 2477, 2540, 2855, 2918, 1463, 1526, 3479, 3542, 2161,
+  2224, 3106, 3169, 890,  953,  2414, 2477, 2918, 2981, 3662, 3725, 571,  634,
+  1971, 2034, 3231, 3294, 3721, 3784, 1145, 1208, 3602, 3665, 1336, 1399, 3541,
+  3604, 2351, 2414, 2981, 3044, 635,  698,  3722, 3785, 954,  1017, 2098, 2161,
+  3169, 3232, 3663, 3726, 1654, 1717, 3418, 3481, 1781, 1844, 3356, 3419, 59,
+  59,   2288, 2351, 3044, 3107, 3776, 3776, 60,   123,  1527, 1590, 3480, 3543,
+  3777, 3840, 699,  762,  3723, 3786, 124,  187,  1908, 1971, 3294, 3357, 3778,
+  3841, 188,  251,  3779, 3842, 1209, 1272, 3603, 3666, 2667, 2730, 2730, 2793,
+  252,  315,  3780, 3843, 2604, 2667, 2793, 2856, 1018, 1081, 1400, 1463, 3542,
+  3605, 3664, 3727, 316,  379,  763,  826,  2035, 2098, 2541, 2604, 2856, 2919,
+  3232, 3295, 3724, 3787, 3781, 3844, 2225, 2288, 3107, 3170, 380,  443,  3782,
+  3845, 2478, 2541, 2919, 2982, 1718, 1781, 3419, 3482, 444,  507,  1273, 1336,
+  3604, 3667, 3783, 3846, 827,  890,  1591, 1654, 1845, 1908, 3357, 3420, 3481,
+  3544, 3725, 3788, 1082, 1145, 2415, 2478, 2982, 3045, 3665, 3728, 2162, 2225,
+  3170, 3233, 508,  571,  3784, 3847, 1464, 1527, 1972, 2035, 3295, 3358, 3543,
+  3606, 2352, 2415, 3045, 3108, 891,  954,  3726, 3789, 572,  635,  3785, 3848,
+  1146, 1209, 3666, 3729, 1337, 1400, 2099, 2162, 3233, 3296, 3605, 3668, 2289,
+  2352, 3108, 3171, 2731, 2794, 636,  699,  1782, 1845, 2668, 2731, 2794, 2857,
+  3420, 3483, 3786, 3849, 1655, 1718, 3482, 3545, 955,  1018, 2605, 2668, 2857,
+  2920, 3727, 3790, 1909, 1972, 3358, 3421, 1528, 1591, 3544, 3607, 2542, 2605,
+  2920, 2983, 60,   60,   700,  763,  3787, 3850, 3840, 3840, 61,   124,  3841,
+  3904, 125,  188,  1210, 1273, 2226, 2289, 3171, 3234, 3667, 3730, 3842, 3905,
+  2036, 2099, 3296, 3359, 189,  252,  2479, 2542, 2983, 3046, 3843, 3906, 1401,
+  1464, 3606, 3669, 253,  316,  1019, 1082, 3728, 3791, 3844, 3907, 764,  827,
+  3788, 3851, 317,  380,  3845, 3908, 2416, 2479, 3046, 3109, 1719, 1782, 3483,
+  3546, 381,  444,  1846, 1909, 2163, 2226, 3234, 3297, 3421, 3484, 3846, 3909,
+  1592, 1655, 3545, 3608, 1274, 1337, 3668, 3731, 828,  891,  3789, 3852, 445,
+  508,  1083, 1146, 1973, 2036, 2353, 2416, 3109, 3172, 3359, 3422, 3729, 3792,
+  3847, 3910, 1465, 1528, 3607, 3670, 509,  572,  2732, 2795, 2795, 2858, 3848,
+  3911, 2669, 2732, 2858, 2921, 2100, 2163, 3297, 3360, 892,  955,  2290, 2353,
+  3172, 3235, 3790, 3853, 2606, 2669, 2921, 2984, 573,  636,  3849, 3912, 1147,
+  1210, 1338, 1401, 3669, 3732, 3730, 3793, 1783, 1846, 2543, 2606, 2984, 3047,
+  3484, 3547, 1656, 1719, 3546, 3609, 1910, 1973, 3422, 3485, 637,  700,  3850,
+  3913, 956,  1019, 1529, 1592, 2480, 2543, 3047, 3110, 3608, 3671, 3791, 3854,
+  2227, 2290, 3235, 3298, 2037, 2100, 3360, 3423, 701,  764,  1211, 1274, 3731,
+  3794, 3851, 3914, 61,   61,   3904, 3904, 62,   125,  2417, 2480, 3110, 3173,
+  3905, 3968, 126,  189,  1402, 1465, 3670, 3733, 3906, 3969, 190,  253,  3907,
+  3970, 1020, 1083, 3792, 3855, 254,  317,  2164, 2227, 3298, 3361, 3908, 3971,
+  765,  828,  1720, 1783, 3547, 3610, 3852, 3915, 1847, 1910, 3485, 3548, 318,
+  381,  2354, 2417, 3173, 3236, 3909, 3972, 2796, 2859, 1593, 1656, 2733, 2796,
+  2859, 2922, 3609, 3672, 1974, 2037, 3423, 3486, 382,  445,  2670, 2733, 2922,
+  2985, 3910, 3973, 1275, 1338, 3732, 3795, 1084, 1147, 3793, 3856, 829,  892,
+  2607, 2670, 2985, 3048, 3853, 3916, 446,  509,  1466, 1529, 3671, 3734, 3911,
+  3974, 2291, 2354, 3236, 3299, 2101, 2164, 3361, 3424, 2544, 2607, 3048, 3111,
+  510,  573,  3912, 3975, 893,  956,  3854, 3917, 1784, 1847, 3548, 3611, 1339,
+  1402, 2481, 2544, 3111, 3174, 3733, 3796, 1148, 1211, 3794, 3857, 574,  637,
+  1657, 1720, 1911, 1974, 3486, 3549, 3610, 3673, 3913, 3976, 2228, 2291, 3299,
+  3362, 1530, 1593, 2038, 2101, 3424, 3487, 3672, 3735, 638,  701,  2418, 2481,
+  3174, 3237, 3914, 3977, 957,  1020, 3855, 3918, 1212, 1275, 2797, 2860, 2860,
+  2923, 3795, 3858, 702,  765,  1403, 1466, 2165, 2228, 2734, 2797, 2923, 2986,
+  3362, 3425, 3734, 3797, 3915, 3978, 62,   62,   3968, 3968, 63,   126,  2355,
+  2418, 3237, 3300, 3969, 4032, 127,  190,  2671, 2734, 2986, 3049, 3970, 4033,
+  1021, 1084, 1848, 1911, 3549, 3612, 3856, 3919, 191,  254,  1721, 1784, 3611,
+  3674, 3971, 4034, 255,  318,  2608, 2671, 3049, 3112, 3972, 4035, 1975, 2038,
+  3487, 3550, 766,  829,  3916, 3979, 1594, 1657, 3673, 3736, 319,  382,  3973,
+  4036, 1276, 1339, 2292, 2355, 3300, 3363, 3796, 3859, 2545, 2608, 3112, 3175,
+  383,  446,  2102, 2165, 3425, 3488, 3974, 4037, 1085, 1148, 1467, 1530, 3735,
+  3798, 3857, 3920, 830,  893,  3917, 3980, 447,  510,  3975, 4038, 2482, 2545,
+  3175, 3238, 511,  574,  1785, 1848, 3612, 3675, 3976, 4039, 2229, 2292, 3363,
+  3426, 1912, 1975, 3550, 3613, 894,  957,  1658, 1721, 3674, 3737, 3918, 3981,
+  1340, 1403, 3797, 3860, 1149, 1212, 2419, 2482, 3238, 3301, 3858, 3921, 2039,
+  2102, 3488, 3551, 575,  638,  2861, 2924, 3977, 4040, 2798, 2861, 2924, 2987,
+  1531, 1594, 3736, 3799, 2735, 2798, 2987, 3050, 2672, 2735, 3050, 3113, 639,
+  702,  958,  1021, 3919, 3982, 3978, 4041, 2166, 2229, 3426, 3489, 2356, 2419,
+  3301, 3364, 1213, 1276, 2609, 2672, 3113, 3176, 3859, 3922, 1404, 1467, 3798,
+  3861, 703,  766,  1849, 1912, 3613, 3676, 3979, 4042, 1722, 1785, 3675, 3738,
+  1976, 2039, 3551, 3614, 1022, 1085, 2546, 2609, 3176, 3239, 3920, 3983, 2293,
+  2356, 3364, 3427, 1595, 1658, 3737, 3800, 767,  830,  3980, 4043, 2103, 2166,
+  3489, 3552, 1277, 1340, 3860, 3923, 2483, 2546, 3239, 3302, 1468, 1531, 3799,
+  3862, 1086, 1149, 3921, 3984, 831,  894,  3981, 4044, 2230, 2293, 2862, 2925,
+  2925, 2988, 3427, 3490, 2799, 2862, 2988, 3051, 1786, 1849, 2420, 2483, 3302,
+  3365, 3676, 3739, 1913, 1976, 3614, 3677, 2736, 2799, 3051, 3114, 1659, 1722,
+  3738, 3801, 2040, 2103, 3552, 3615, 1341, 1404, 3861, 3924, 895,  958,  2673,
+  2736, 3114, 3177, 3982, 4045, 1150, 1213, 3922, 3985, 1532, 1595, 3800, 3863,
+  2357, 2420, 3365, 3428, 2167, 2230, 2610, 2673, 3177, 3240, 3490, 3553, 959,
+  1022, 3983, 4046, 2547, 2610, 3240, 3303, 1214, 1277, 1405, 1468, 1850, 1913,
+  3677, 3740, 3862, 3925, 3923, 3986, 1723, 1786, 1977, 2040, 3615, 3678, 3739,
+  3802, 2294, 2357, 3428, 3491, 1023, 1086, 1596, 1659, 2104, 2167, 2484, 2547,
+  3303, 3366, 3553, 3616, 3801, 3864, 3984, 4047, 2926, 2989, 2863, 2926, 2989,
+  3052, 2800, 2863, 3052, 3115, 1278, 1341, 3924, 3987, 1469, 1532, 2231, 2294,
+  2737, 2800, 3115, 3178, 3491, 3554, 3863, 3926, 2421, 2484, 3366, 3429, 1087,
+  1150, 3985, 4048, 1914, 1977, 2674, 2737, 3178, 3241, 3678, 3741, 1787, 1850,
+  3740, 3803, 2041, 2104, 3616, 3679, 1660, 1723, 3802, 3865, 2611, 2674, 3241,
+  3304, 1342, 1405, 2358, 2421, 3429, 3492, 3925, 3988, 2168, 2231, 3554, 3617,
+  1151, 1214, 3986, 4049, 1533, 1596, 3864, 3927, 2548, 2611, 3304, 3367, 2295,
+  2358, 3492, 3555, 1851, 1914, 3741, 3804, 1978, 2041, 2927, 2990, 2990, 3053,
+  3679, 3742, 1406, 1469, 3926, 3989, 1724, 1787, 2864, 2927, 3053, 3116, 3803,
+  3866, 1215, 1278, 2485, 2548, 3367, 3430, 3987, 4050, 2801, 2864, 3116, 3179,
+  2105, 2168, 3617, 3680, 1597, 1660, 3865, 3928, 2738, 2801, 3179, 3242, 2422,
+  2485, 3430, 3493, 2232, 2295, 3555, 3618, 2675, 2738, 3242, 3305, 1279, 1342,
+  3988, 4051, 1470, 1533, 3927, 3990, 1915, 1978, 3742, 3805, 1788, 1851, 3804,
+  3867, 2612, 2675, 3305, 3368, 2042, 2105, 3680, 3743, 2359, 2422, 3493, 3556,
+  1661, 1724, 3866, 3929, 2169, 2232, 3618, 3681, 2549, 2612, 3368, 3431, 1343,
+  1406, 3989, 4052, 2991, 3054, 1534, 1597, 2928, 2991, 3054, 3117, 3928, 3991,
+  2865, 2928, 3117, 3180, 2296, 2359, 3556, 3619, 2802, 2865, 3180, 3243, 2486,
+  2549, 3431, 3494, 1852, 1915, 3805, 3868, 1979, 2042, 3743, 3806, 1725, 1788,
+  2739, 2802, 3243, 3306, 3867, 3930, 1407, 1470, 2106, 2169, 3681, 3744, 3990,
+  4053, 2676, 2739, 3306, 3369, 1598, 1661, 2423, 2486, 3494, 3557, 3929, 3992,
+  2233, 2296, 3619, 3682, 2613, 2676, 3369, 3432, 1471, 1534, 3991, 4054, 1916,
+  1979, 3806, 3869, 1789, 1852, 2043, 2106, 2360, 2423, 3557, 3620, 3744, 3807,
+  3868, 3931, 2992, 3055, 3055, 3118, 2550, 2613, 3432, 3495, 2929, 2992, 3118,
+  3181, 1662, 1725, 2170, 2233, 3682, 3745, 3930, 3993, 2866, 2929, 3181, 3244,
+  2803, 2866, 3244, 3307, 1535, 1598, 2297, 2360, 3620, 3683, 3992, 4055, 2487,
+  2550, 3495, 3558, 2740, 2803, 3307, 3370, 1980, 2043, 3807, 3870, 1853, 1916,
+  3869, 3932, 2107, 2170, 3745, 3808, 1726, 1789, 2677, 2740, 3370, 3433, 3931,
+  3994, 2424, 2487, 3558, 3621, 2234, 2297, 3683, 3746, 1599, 1662, 3993, 4056,
+  2614, 2677, 3433, 3496, 3056, 3119, 2993, 3056, 3119, 3182, 2930, 2993, 3182,
+  3245, 2361, 2424, 3621, 3684, 1917, 1980, 3870, 3933, 2044, 2107, 3808, 3871,
+  2551, 2614, 3496, 3559, 2867, 2930, 3245, 3308, 1790, 1853, 3932, 3995, 2171,
+  2234, 3746, 3809, 2804, 2867, 3308, 3371, 1663, 1726, 3994, 4057, 2488, 2551,
+  3559, 3622, 2741, 2804, 3371, 3434, 2298, 2361, 3684, 3747, 2678, 2741, 3434,
+  3497, 1981, 2044, 3871, 3934, 1854, 1917, 3933, 3996, 2108, 2171, 3809, 3872,
+  2425, 2488, 3622, 3685, 1727, 1790, 3995, 4058, 3057, 3120, 3120, 3183, 2235,
+  2298, 2615, 2678, 3497, 3560, 3747, 3810, 2994, 3057, 3183, 3246, 2931, 2994,
+  3246, 3309, 2868, 2931, 3309, 3372, 2362, 2425, 3685, 3748, 2552, 2615, 3560,
+  3623, 1918, 1981, 3934, 3997, 2045, 2108, 2805, 2868, 3372, 3435, 3872, 3935,
+  1791, 1854, 3996, 4059, 2172, 2235, 3810, 3873, 2742, 2805, 3435, 3498, 2489,
+  2552, 3623, 3686, 2299, 2362, 3748, 3811, 2679, 2742, 3498, 3561, 3121, 3184,
+  3058, 3121, 3184, 3247, 1982, 2045, 3935, 3998, 2426, 2489, 3686, 3749, 1855,
+  1918, 2109, 2172, 2995, 3058, 3247, 3310, 3873, 3936, 3997, 4060, 2616, 2679,
+  3561, 3624, 2932, 2995, 3310, 3373, 2236, 2299, 3811, 3874, 2869, 2932, 3373,
+  3436, 2553, 2616, 3624, 3687, 2363, 2426, 3749, 3812, 2806, 2869, 3436, 3499,
+  2046, 2109, 3936, 3999, 1919, 1982, 3998, 4061, 2743, 2806, 3499, 3562, 2173,
+  2236, 3874, 3937, 2490, 2553, 3687, 3750, 2300, 2363, 3812, 3875, 2680, 2743,
+  3562, 3625, 3122, 3185, 3185, 3248, 3059, 3122, 3248, 3311, 2996, 3059, 3311,
+  3374, 2427, 2490, 2933, 2996, 3374, 3437, 3750, 3813, 1983, 2046, 2617, 2680,
+  3625, 3688, 3999, 4062, 2110, 2173, 3937, 4000, 2870, 2933, 3437, 3500, 2237,
+  2300, 3875, 3938, 2807, 2870, 3500, 3563, 2554, 2617, 3688, 3751, 2364, 2427,
+  3813, 3876, 2744, 2807, 3563, 3626, 2047, 2110, 4000, 4063, 2174, 2237, 3186,
+  3249, 3938, 4001, 2491, 2554, 3123, 3186, 3249, 3312, 3751, 3814, 3060, 3123,
+  3312, 3375, 2681, 2744, 3626, 3689, 2301, 2364, 3876, 3939, 2997, 3060, 3375,
+  3438, 2934, 2997, 3438, 3501, 2428, 2491, 3814, 3877, 2618, 2681, 3689, 3752,
+  2871, 2934, 3501, 3564, 2111, 2174, 4001, 4064, 2238, 2301, 3939, 4002, 2808,
+  2871, 3564, 3627, 2555, 2618, 3752, 3815, 2365, 2428, 3877, 3940, 2745, 2808,
+  3627, 3690, 3187, 3250, 3250, 3313, 3124, 3187, 3313, 3376, 3061, 3124, 3376,
+  3439, 2492, 2555, 3815, 3878, 2175, 2238, 2998, 3061, 3439, 3502, 4002, 4065,
+  2682, 2745, 3690, 3753, 2302, 2365, 3940, 4003, 2935, 2998, 3502, 3565, 2872,
+  2935, 3565, 3628, 2619, 2682, 3753, 3816, 2429, 2492, 3878, 3941, 2809, 2872,
+  3628, 3691, 2239, 2302, 4003, 4066, 2556, 2619, 3816, 3879, 3251, 3314, 3188,
+  3251, 3314, 3377, 3125, 3188, 3377, 3440, 2366, 2429, 2746, 2809, 3691, 3754,
+  3941, 4004, 3062, 3125, 3440, 3503, 2999, 3062, 3503, 3566, 2493, 2556, 3879,
+  3942, 2683, 2746, 3754, 3817, 2936, 2999, 3566, 3629, 2303, 2366, 4004, 4067,
+  2873, 2936, 3629, 3692, 2620, 2683, 3817, 3880, 2430, 2493, 3942, 4005, 2810,
+  2873, 3692, 3755, 3252, 3315, 3315, 3378, 3189, 3252, 3378, 3441, 3126, 3189,
+  3441, 3504, 2557, 2620, 3880, 3943, 3063, 3126, 3504, 3567, 2747, 2810, 3755,
+  3818, 2367, 2430, 4005, 4068, 3000, 3063, 3567, 3630, 2684, 2747, 3818, 3881,
+  2494, 2557, 2937, 3000, 3630, 3693, 3943, 4006, 2874, 2937, 3693, 3756, 2621,
+  2684, 3881, 3944, 3316, 3379, 3253, 3316, 3379, 3442, 2431, 2494, 4006, 4069,
+  3190, 3253, 3442, 3505, 2811, 2874, 3756, 3819, 3127, 3190, 3505, 3568, 3064,
+  3127, 3568, 3631, 2558, 2621, 3944, 4007, 2748, 2811, 3819, 3882, 3001, 3064,
+  3631, 3694, 2938, 3001, 3694, 3757, 2685, 2748, 3882, 3945, 2495, 2558, 4007,
+  4070, 2875, 2938, 3757, 3820, 3317, 3380, 3380, 3443, 3254, 3317, 3443, 3506,
+  2622, 2685, 3191, 3254, 3506, 3569, 3945, 4008, 2812, 2875, 3820, 3883, 3128,
+  3191, 3569, 3632, 3065, 3128, 3632, 3695, 2559, 2622, 4008, 4071, 2749, 2812,
+  3883, 3946, 3002, 3065, 3695, 3758, 2939, 3002, 3758, 3821, 2686, 2749, 3946,
+  4009, 3381, 3444, 3318, 3381, 3444, 3507, 2876, 2939, 3821, 3884, 3255, 3318,
+  3507, 3570, 3192, 3255, 3570, 3633, 2623, 2686, 3129, 3192, 3633, 3696, 4009,
+  4072, 2813, 2876, 3884, 3947, 3066, 3129, 3696, 3759, 3003, 3066, 3759, 3822,
+  2750, 2813, 3947, 4010, 2940, 3003, 3822, 3885, 3382, 3445, 3445, 3508, 3319,
+  3382, 3508, 3571, 2687, 2750, 4010, 4073, 3256, 3319, 3571, 3634, 2877, 2940,
+  3885, 3948, 3193, 3256, 3634, 3697, 3130, 3193, 3697, 3760, 2814, 2877, 3948,
+  4011, 3067, 3130, 3760, 3823, 3004, 3067, 3823, 3886, 2751, 2814, 4011, 4074,
+  3446, 3509, 3383, 3446, 3509, 3572, 2941, 3004, 3886, 3949, 3320, 3383, 3572,
+  3635, 3257, 3320, 3635, 3698, 3194, 3257, 3698, 3761, 2878, 2941, 3949, 4012,
+  3131, 3194, 3761, 3824, 3068, 3131, 3824, 3887, 2815, 2878, 4012, 4075, 3005,
+  3068, 3887, 3950, 3447, 3510, 3510, 3573, 3384, 3447, 3573, 3636, 3321, 3384,
+  3636, 3699, 2942, 3005, 3950, 4013, 3258, 3321, 3699, 3762, 3195, 3258, 3762,
+  3825, 2879, 2942, 4013, 4076, 3132, 3195, 3825, 3888, 3069, 3132, 3888, 3951,
+  3511, 3574, 3448, 3511, 3574, 3637, 3006, 3069, 3951, 4014, 3385, 3448, 3637,
+  3700, 3322, 3385, 3700, 3763, 3259, 3322, 3763, 3826, 2943, 3006, 4014, 4077,
+  3196, 3259, 3826, 3889, 3133, 3196, 3889, 3952, 3070, 3133, 3952, 4015, 3512,
+  3575, 3575, 3638, 3449, 3512, 3638, 3701, 3386, 3449, 3701, 3764, 3007, 3070,
+  4015, 4078, 3323, 3386, 3764, 3827, 3260, 3323, 3827, 3890, 3197, 3260, 3890,
+  3953, 3134, 3197, 3953, 4016, 3576, 3639, 3071, 3134, 4016, 4079, 3513, 3576,
+  3639, 3702, 3450, 3513, 3702, 3765, 3387, 3450, 3765, 3828, 3324, 3387, 3828,
+  3891, 3261, 3324, 3891, 3954, 3198, 3261, 3954, 4017, 3135, 3198, 4017, 4080,
+  3577, 3640, 3640, 3703, 3514, 3577, 3703, 3766, 3451, 3514, 3766, 3829, 3388,
+  3451, 3829, 3892, 3325, 3388, 3892, 3955, 3262, 3325, 3955, 4018, 3199, 3262,
+  4018, 4081, 3641, 3704, 3578, 3641, 3704, 3767, 3515, 3578, 3767, 3830, 3452,
+  3515, 3830, 3893, 3389, 3452, 3893, 3956, 3326, 3389, 3956, 4019, 3263, 3326,
+  4019, 4082, 3642, 3705, 3705, 3768, 3579, 3642, 3768, 3831, 3516, 3579, 3831,
+  3894, 3453, 3516, 3894, 3957, 3390, 3453, 3957, 4020, 3327, 3390, 4020, 4083,
+  3706, 3769, 3643, 3706, 3769, 3832, 3580, 3643, 3832, 3895, 3517, 3580, 3895,
+  3958, 3454, 3517, 3958, 4021, 3391, 3454, 4021, 4084, 3707, 3770, 3770, 3833,
+  3644, 3707, 3833, 3896, 3581, 3644, 3896, 3959, 3518, 3581, 3959, 4022, 3455,
+  3518, 4022, 4085, 3771, 3834, 3708, 3771, 3834, 3897, 3645, 3708, 3897, 3960,
+  3582, 3645, 3960, 4023, 3519, 3582, 4023, 4086, 3772, 3835, 3835, 3898, 3709,
+  3772, 3898, 3961, 3646, 3709, 3961, 4024, 3583, 3646, 4024, 4087, 3836, 3899,
+  3773, 3836, 3899, 3962, 3710, 3773, 3962, 4025, 3647, 3710, 4025, 4088, 3837,
+  3900, 3900, 3963, 3774, 3837, 3963, 4026, 3711, 3774, 4026, 4089, 3901, 3964,
+  3838, 3901, 3964, 4027, 3775, 3838, 4027, 4090, 3902, 3965, 3965, 4028, 3839,
+  3902, 4028, 4091, 3966, 4029, 3903, 3966, 4029, 4092, 3967, 4030, 4030, 4093,
+  4031, 4094, 0,    0,
+};
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_CB4X4
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_2x2[4]) = { 0, 1, 2,
+                                                                        3 };
+#endif
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_4x4[16]) = {
+  0, 2, 5, 8, 1, 3, 9, 12, 4, 7, 11, 14, 6, 10, 13, 15,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_4x4[16]) = {
+  0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_4x4[16]) = {
+  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, av1_col_iscan_4x4[16]) = {
+  0, 3, 7, 11, 1, 5, 9, 12, 2, 6, 10, 14, 4, 8, 13, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_row_iscan_4x4[16]) = {
+  0, 1, 3, 5, 2, 4, 6, 9, 7, 8, 11, 13, 10, 12, 14, 15,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_4x8[32]) = {
+  0,  1,  4,  9,  2,  3,  6,  11, 5,  7,  8,  13, 10, 12, 14, 17,
+  15, 16, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_4x8[32]) = {
+  0, 8,  16, 24, 1, 9,  17, 25, 2, 10, 18, 26, 3, 11, 19, 27,
+  4, 12, 20, 28, 5, 13, 21, 29, 6, 14, 22, 30, 7, 15, 23, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_4x8[32]) = {
+  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_8x4[32]) = {
+  0, 1, 4, 9,  15, 19, 24, 28, 2,  3,  6,  11, 16, 21, 25, 29,
+  5, 7, 8, 13, 18, 22, 26, 30, 10, 12, 14, 17, 20, 23, 27, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_8x4[32]) = {
+  0, 4, 8,  12, 16, 20, 24, 28, 1, 5, 9,  13, 17, 21, 25, 29,
+  2, 6, 10, 14, 18, 22, 26, 30, 3, 7, 11, 15, 19, 23, 27, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_8x4[32]) = {
+  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_4x16[64]) = {
+  0,  1,  3,  6,  2,  4,  7,  10, 5,  8,  11, 14, 9,  12, 15, 18,
+  13, 16, 19, 22, 17, 20, 23, 26, 21, 24, 27, 30, 25, 28, 31, 34,
+  29, 32, 35, 38, 33, 36, 39, 42, 37, 40, 43, 46, 41, 44, 47, 50,
+  45, 48, 51, 54, 49, 52, 55, 58, 53, 56, 59, 61, 57, 60, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_16x4[64]) = {
+  0, 1,  3,  6,  10, 14, 18, 22, 26, 30, 34, 38, 42, 46, 50, 54,
+  2, 4,  7,  11, 15, 19, 23, 27, 31, 35, 39, 43, 47, 51, 55, 58,
+  5, 8,  12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 59, 61,
+  9, 13, 17, 21, 25, 29, 33, 37, 41, 45, 49, 53, 57, 60, 62, 63,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_4x16[64]) = {
+  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+  32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+  48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_16x4[64]) = {
+  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+  32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+  48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_4x16[64]) = {
+  0,  16, 32, 48, 1,  17, 33, 49, 2,  18, 34, 50, 3,  19, 35, 51,
+  4,  20, 36, 52, 5,  21, 37, 53, 6,  22, 38, 54, 7,  23, 39, 55,
+  8,  24, 40, 56, 9,  25, 41, 57, 10, 26, 42, 58, 11, 27, 43, 59,
+  12, 28, 44, 60, 13, 29, 45, 61, 14, 30, 46, 62, 15, 31, 47, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_16x4[64]) = {
+  0,  16, 32, 48, 1,  17, 33, 49, 2,  18, 34, 50, 3,  19, 35, 51,
+  4,  20, 36, 52, 5,  21, 37, 53, 6,  22, 38, 54, 7,  23, 39, 55,
+  8,  24, 40, 56, 9,  25, 41, 57, 10, 26, 42, 58, 11, 27, 43, 59,
+  12, 28, 44, 60, 13, 29, 45, 61, 14, 30, 46, 62, 15, 31, 47, 63,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_8x32[256]) = {
+  0,   1,   3,   6,   10,  15,  21,  28,  2,   4,   7,   11,  16,  22,  29,
+  36,  5,   8,   12,  17,  23,  30,  37,  44,  9,   13,  18,  24,  31,  38,
+  45,  52,  14,  19,  25,  32,  39,  46,  53,  60,  20,  26,  33,  40,  47,
+  54,  61,  68,  27,  34,  41,  48,  55,  62,  69,  76,  35,  42,  49,  56,
+  63,  70,  77,  84,  43,  50,  57,  64,  71,  78,  85,  92,  51,  58,  65,
+  72,  79,  86,  93,  100, 59,  66,  73,  80,  87,  94,  101, 108, 67,  74,
+  81,  88,  95,  102, 109, 116, 75,  82,  89,  96,  103, 110, 117, 124, 83,
+  90,  97,  104, 111, 118, 125, 132, 91,  98,  105, 112, 119, 126, 133, 140,
+  99,  106, 113, 120, 127, 134, 141, 148, 107, 114, 121, 128, 135, 142, 149,
+  156, 115, 122, 129, 136, 143, 150, 157, 164, 123, 130, 137, 144, 151, 158,
+  165, 172, 131, 138, 145, 152, 159, 166, 173, 180, 139, 146, 153, 160, 167,
+  174, 181, 188, 147, 154, 161, 168, 175, 182, 189, 196, 155, 162, 169, 176,
+  183, 190, 197, 204, 163, 170, 177, 184, 191, 198, 205, 212, 171, 178, 185,
+  192, 199, 206, 213, 220, 179, 186, 193, 200, 207, 214, 221, 228, 187, 194,
+  201, 208, 215, 222, 229, 235, 195, 202, 209, 216, 223, 230, 236, 241, 203,
+  210, 217, 224, 231, 237, 242, 246, 211, 218, 225, 232, 238, 243, 247, 250,
+  219, 226, 233, 239, 244, 248, 251, 253, 227, 234, 240, 245, 249, 252, 254,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_32x8[256]) = {
+  0,   1,   3,   6,   10,  15,  21,  28,  36,  44,  52,  60,  68,  76,  84,
+  92,  100, 108, 116, 124, 132, 140, 148, 156, 164, 172, 180, 188, 196, 204,
+  212, 220, 2,   4,   7,   11,  16,  22,  29,  37,  45,  53,  61,  69,  77,
+  85,  93,  101, 109, 117, 125, 133, 141, 149, 157, 165, 173, 181, 189, 197,
+  205, 213, 221, 228, 5,   8,   12,  17,  23,  30,  38,  46,  54,  62,  70,
+  78,  86,  94,  102, 110, 118, 126, 134, 142, 150, 158, 166, 174, 182, 190,
+  198, 206, 214, 222, 229, 235, 9,   13,  18,  24,  31,  39,  47,  55,  63,
+  71,  79,  87,  95,  103, 111, 119, 127, 135, 143, 151, 159, 167, 175, 183,
+  191, 199, 207, 215, 223, 230, 236, 241, 14,  19,  25,  32,  40,  48,  56,
+  64,  72,  80,  88,  96,  104, 112, 120, 128, 136, 144, 152, 160, 168, 176,
+  184, 192, 200, 208, 216, 224, 231, 237, 242, 246, 20,  26,  33,  41,  49,
+  57,  65,  73,  81,  89,  97,  105, 113, 121, 129, 137, 145, 153, 161, 169,
+  177, 185, 193, 201, 209, 217, 225, 232, 238, 243, 247, 250, 27,  34,  42,
+  50,  58,  66,  74,  82,  90,  98,  106, 114, 122, 130, 138, 146, 154, 162,
+  170, 178, 186, 194, 202, 210, 218, 226, 233, 239, 244, 248, 251, 253, 35,
+  43,  51,  59,  67,  75,  83,  91,  99,  107, 115, 123, 131, 139, 147, 155,
+  163, 171, 179, 187, 195, 203, 211, 219, 227, 234, 240, 245, 249, 252, 254,
+  255,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_8x32[256]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134,
+  135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149,
+  150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164,
+  165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
+  180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194,
+  195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209,
+  210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224,
+  225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
+  240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_32x8[256]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134,
+  135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149,
+  150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164,
+  165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
+  180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194,
+  195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209,
+  210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224,
+  225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
+  240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_8x32[256]) = {
+  0,  32, 64, 96,  128, 160, 192, 224, 1,  33, 65, 97,  129, 161, 193, 225,
+  2,  34, 66, 98,  130, 162, 194, 226, 3,  35, 67, 99,  131, 163, 195, 227,
+  4,  36, 68, 100, 132, 164, 196, 228, 5,  37, 69, 101, 133, 165, 197, 229,
+  6,  38, 70, 102, 134, 166, 198, 230, 7,  39, 71, 103, 135, 167, 199, 231,
+  8,  40, 72, 104, 136, 168, 200, 232, 9,  41, 73, 105, 137, 169, 201, 233,
+  10, 42, 74, 106, 138, 170, 202, 234, 11, 43, 75, 107, 139, 171, 203, 235,
+  12, 44, 76, 108, 140, 172, 204, 236, 13, 45, 77, 109, 141, 173, 205, 237,
+  14, 46, 78, 110, 142, 174, 206, 238, 15, 47, 79, 111, 143, 175, 207, 239,
+  16, 48, 80, 112, 144, 176, 208, 240, 17, 49, 81, 113, 145, 177, 209, 241,
+  18, 50, 82, 114, 146, 178, 210, 242, 19, 51, 83, 115, 147, 179, 211, 243,
+  20, 52, 84, 116, 148, 180, 212, 244, 21, 53, 85, 117, 149, 181, 213, 245,
+  22, 54, 86, 118, 150, 182, 214, 246, 23, 55, 87, 119, 151, 183, 215, 247,
+  24, 56, 88, 120, 152, 184, 216, 248, 25, 57, 89, 121, 153, 185, 217, 249,
+  26, 58, 90, 122, 154, 186, 218, 250, 27, 59, 91, 123, 155, 187, 219, 251,
+  28, 60, 92, 124, 156, 188, 220, 252, 29, 61, 93, 125, 157, 189, 221, 253,
+  30, 62, 94, 126, 158, 190, 222, 254, 31, 63, 95, 127, 159, 191, 223, 255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_32x8[256]) = {
+  0,   8,   16,  24,  32,  40,  48,  56,  64,  72,  80,  88,  96,  104, 112,
+  120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232,
+  240, 248, 1,   9,   17,  25,  33,  41,  49,  57,  65,  73,  81,  89,  97,
+  105, 113, 121, 129, 137, 145, 153, 161, 169, 177, 185, 193, 201, 209, 217,
+  225, 233, 241, 249, 2,   10,  18,  26,  34,  42,  50,  58,  66,  74,  82,
+  90,  98,  106, 114, 122, 130, 138, 146, 154, 162, 170, 178, 186, 194, 202,
+  210, 218, 226, 234, 242, 250, 3,   11,  19,  27,  35,  43,  51,  59,  67,
+  75,  83,  91,  99,  107, 115, 123, 131, 139, 147, 155, 163, 171, 179, 187,
+  195, 203, 211, 219, 227, 235, 243, 251, 4,   12,  20,  28,  36,  44,  52,
+  60,  68,  76,  84,  92,  100, 108, 116, 124, 132, 140, 148, 156, 164, 172,
+  180, 188, 196, 204, 212, 220, 228, 236, 244, 252, 5,   13,  21,  29,  37,
+  45,  53,  61,  69,  77,  85,  93,  101, 109, 117, 125, 133, 141, 149, 157,
+  165, 173, 181, 189, 197, 205, 213, 221, 229, 237, 245, 253, 6,   14,  22,
+  30,  38,  46,  54,  62,  70,  78,  86,  94,  102, 110, 118, 126, 134, 142,
+  150, 158, 166, 174, 182, 190, 198, 206, 214, 222, 230, 238, 246, 254, 7,
+  15,  23,  31,  39,  47,  55,  63,  71,  79,  87,  95,  103, 111, 119, 127,
+  135, 143, 151, 159, 167, 175, 183, 191, 199, 207, 215, 223, 231, 239, 247,
+  255,
+};
+#endif  // CONFIG_EXT_TX
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_8x8[64]) = {
+  0, 8,  16, 24, 32, 40, 48, 56, 1, 9,  17, 25, 33, 41, 49, 57,
+  2, 10, 18, 26, 34, 42, 50, 58, 3, 11, 19, 27, 35, 43, 51, 59,
+  4, 12, 20, 28, 36, 44, 52, 60, 5, 13, 21, 29, 37, 45, 53, 61,
+  6, 14, 22, 30, 38, 46, 54, 62, 7, 15, 23, 31, 39, 47, 55, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_8x8[64]) = {
+  0,  1,  2,  3,  4,  5,  6,  7,  8,  9,  10, 11, 12, 13, 14, 15,
+  16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+  32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+  48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, av1_col_iscan_8x8[64]) = {
+  0,  3,  8,  15, 22, 32, 40, 47, 1,  5,  11, 18, 26, 34, 44, 51,
+  2,  7,  13, 20, 28, 38, 46, 54, 4,  10, 16, 24, 31, 41, 50, 56,
+  6,  12, 21, 27, 35, 43, 52, 58, 9,  17, 25, 33, 39, 48, 55, 60,
+  14, 23, 30, 37, 45, 53, 59, 62, 19, 29, 36, 42, 49, 57, 61, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_row_iscan_8x8[64]) = {
+  0,  1,  2,  5,  8,  12, 19, 24, 3,  4,  7,  10, 15, 20, 30, 39,
+  6,  9,  13, 16, 21, 27, 37, 46, 11, 14, 17, 23, 28, 34, 44, 52,
+  18, 22, 25, 31, 35, 41, 50, 57, 26, 29, 33, 38, 43, 49, 55, 59,
+  32, 36, 42, 47, 51, 54, 60, 61, 40, 45, 48, 53, 56, 58, 62, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_8x8[64]) = {
+  0,  2,  5,  9,  14, 22, 31, 37, 1,  4,  8,  13, 19, 26, 38, 44,
+  3,  6,  10, 17, 24, 30, 42, 49, 7,  11, 15, 21, 29, 36, 47, 53,
+  12, 16, 20, 27, 34, 43, 52, 57, 18, 23, 28, 35, 41, 48, 56, 60,
+  25, 32, 39, 45, 50, 55, 59, 62, 33, 40, 46, 51, 54, 58, 61, 63,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_8x16[128]) = {
+  0,  1,  3,   6,   10,  15,  21,  28,  2,  4,   7,   11,  16,  22,  29,  36,
+  5,  8,  12,  17,  23,  30,  37,  44,  9,  13,  18,  24,  31,  38,  45,  52,
+  14, 19, 25,  32,  39,  46,  53,  60,  20, 26,  33,  40,  47,  54,  61,  68,
+  27, 34, 41,  48,  55,  62,  69,  76,  35, 42,  49,  56,  63,  70,  77,  84,
+  43, 50, 57,  64,  71,  78,  85,  92,  51, 58,  65,  72,  79,  86,  93,  100,
+  59, 66, 73,  80,  87,  94,  101, 107, 67, 74,  81,  88,  95,  102, 108, 113,
+  75, 82, 89,  96,  103, 109, 114, 118, 83, 90,  97,  104, 110, 115, 119, 122,
+  91, 98, 105, 111, 116, 120, 123, 125, 99, 106, 112, 117, 121, 124, 126, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_16x8[128]) = {
+  0,  1,  3,  6,  10, 15, 21, 28, 36, 44,  52,  60,  68,  76,  84,  92,
+  2,  4,  7,  11, 16, 22, 29, 37, 45, 53,  61,  69,  77,  85,  93,  100,
+  5,  8,  12, 17, 23, 30, 38, 46, 54, 62,  70,  78,  86,  94,  101, 107,
+  9,  13, 18, 24, 31, 39, 47, 55, 63, 71,  79,  87,  95,  102, 108, 113,
+  14, 19, 25, 32, 40, 48, 56, 64, 72, 80,  88,  96,  103, 109, 114, 118,
+  20, 26, 33, 41, 49, 57, 65, 73, 81, 89,  97,  104, 110, 115, 119, 122,
+  27, 34, 42, 50, 58, 66, 74, 82, 90, 98,  105, 111, 116, 120, 123, 125,
+  35, 43, 51, 59, 67, 75, 83, 91, 99, 106, 112, 117, 121, 124, 126, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_8x16[128]) = {
+  0,  16, 32, 48, 64, 80, 96,  112, 1,  17, 33, 49, 65, 81, 97,  113,
+  2,  18, 34, 50, 66, 82, 98,  114, 3,  19, 35, 51, 67, 83, 99,  115,
+  4,  20, 36, 52, 68, 84, 100, 116, 5,  21, 37, 53, 69, 85, 101, 117,
+  6,  22, 38, 54, 70, 86, 102, 118, 7,  23, 39, 55, 71, 87, 103, 119,
+  8,  24, 40, 56, 72, 88, 104, 120, 9,  25, 41, 57, 73, 89, 105, 121,
+  10, 26, 42, 58, 74, 90, 106, 122, 11, 27, 43, 59, 75, 91, 107, 123,
+  12, 28, 44, 60, 76, 92, 108, 124, 13, 29, 45, 61, 77, 93, 109, 125,
+  14, 30, 46, 62, 78, 94, 110, 126, 15, 31, 47, 63, 79, 95, 111, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_16x8[128]) = {
+  0, 8,  16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96,  104, 112, 120,
+  1, 9,  17, 25, 33, 41, 49, 57, 65, 73, 81, 89, 97,  105, 113, 121,
+  2, 10, 18, 26, 34, 42, 50, 58, 66, 74, 82, 90, 98,  106, 114, 122,
+  3, 11, 19, 27, 35, 43, 51, 59, 67, 75, 83, 91, 99,  107, 115, 123,
+  4, 12, 20, 28, 36, 44, 52, 60, 68, 76, 84, 92, 100, 108, 116, 124,
+  5, 13, 21, 29, 37, 45, 53, 61, 69, 77, 85, 93, 101, 109, 117, 125,
+  6, 14, 22, 30, 38, 46, 54, 62, 70, 78, 86, 94, 102, 110, 118, 126,
+  7, 15, 23, 31, 39, 47, 55, 63, 71, 79, 87, 95, 103, 111, 119, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_8x16[128]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_16x8[128]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_16x32[512]) = {
+  0,   1,   3,   6,   10,  15,  21,  28,  36,  45,  55,  66,  78,  91,  105,
+  120, 2,   4,   7,   11,  16,  22,  29,  37,  46,  56,  67,  79,  92,  106,
+  121, 136, 5,   8,   12,  17,  23,  30,  38,  47,  57,  68,  80,  93,  107,
+  122, 137, 152, 9,   13,  18,  24,  31,  39,  48,  58,  69,  81,  94,  108,
+  123, 138, 153, 168, 14,  19,  25,  32,  40,  49,  59,  70,  82,  95,  109,
+  124, 139, 154, 169, 184, 20,  26,  33,  41,  50,  60,  71,  83,  96,  110,
+  125, 140, 155, 170, 185, 200, 27,  34,  42,  51,  61,  72,  84,  97,  111,
+  126, 141, 156, 171, 186, 201, 216, 35,  43,  52,  62,  73,  85,  98,  112,
+  127, 142, 157, 172, 187, 202, 217, 232, 44,  53,  63,  74,  86,  99,  113,
+  128, 143, 158, 173, 188, 203, 218, 233, 248, 54,  64,  75,  87,  100, 114,
+  129, 144, 159, 174, 189, 204, 219, 234, 249, 264, 65,  76,  88,  101, 115,
+  130, 145, 160, 175, 190, 205, 220, 235, 250, 265, 280, 77,  89,  102, 116,
+  131, 146, 161, 176, 191, 206, 221, 236, 251, 266, 281, 296, 90,  103, 117,
+  132, 147, 162, 177, 192, 207, 222, 237, 252, 267, 282, 297, 312, 104, 118,
+  133, 148, 163, 178, 193, 208, 223, 238, 253, 268, 283, 298, 313, 328, 119,
+  134, 149, 164, 179, 194, 209, 224, 239, 254, 269, 284, 299, 314, 329, 344,
+  135, 150, 165, 180, 195, 210, 225, 240, 255, 270, 285, 300, 315, 330, 345,
+  360, 151, 166, 181, 196, 211, 226, 241, 256, 271, 286, 301, 316, 331, 346,
+  361, 376, 167, 182, 197, 212, 227, 242, 257, 272, 287, 302, 317, 332, 347,
+  362, 377, 392, 183, 198, 213, 228, 243, 258, 273, 288, 303, 318, 333, 348,
+  363, 378, 393, 407, 199, 214, 229, 244, 259, 274, 289, 304, 319, 334, 349,
+  364, 379, 394, 408, 421, 215, 230, 245, 260, 275, 290, 305, 320, 335, 350,
+  365, 380, 395, 409, 422, 434, 231, 246, 261, 276, 291, 306, 321, 336, 351,
+  366, 381, 396, 410, 423, 435, 446, 247, 262, 277, 292, 307, 322, 337, 352,
+  367, 382, 397, 411, 424, 436, 447, 457, 263, 278, 293, 308, 323, 338, 353,
+  368, 383, 398, 412, 425, 437, 448, 458, 467, 279, 294, 309, 324, 339, 354,
+  369, 384, 399, 413, 426, 438, 449, 459, 468, 476, 295, 310, 325, 340, 355,
+  370, 385, 400, 414, 427, 439, 450, 460, 469, 477, 484, 311, 326, 341, 356,
+  371, 386, 401, 415, 428, 440, 451, 461, 470, 478, 485, 491, 327, 342, 357,
+  372, 387, 402, 416, 429, 441, 452, 462, 471, 479, 486, 492, 497, 343, 358,
+  373, 388, 403, 417, 430, 442, 453, 463, 472, 480, 487, 493, 498, 502, 359,
+  374, 389, 404, 418, 431, 443, 454, 464, 473, 481, 488, 494, 499, 503, 506,
+  375, 390, 405, 419, 432, 444, 455, 465, 474, 482, 489, 495, 500, 504, 507,
+  509, 391, 406, 420, 433, 445, 456, 466, 475, 483, 490, 496, 501, 505, 508,
+  510, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_32x16[512]) = {
+  0,   1,   3,   6,   10,  15,  21,  28,  36,  45,  55,  66,  78,  91,  105,
+  120, 136, 152, 168, 184, 200, 216, 232, 248, 264, 280, 296, 312, 328, 344,
+  360, 376, 2,   4,   7,   11,  16,  22,  29,  37,  46,  56,  67,  79,  92,
+  106, 121, 137, 153, 169, 185, 201, 217, 233, 249, 265, 281, 297, 313, 329,
+  345, 361, 377, 392, 5,   8,   12,  17,  23,  30,  38,  47,  57,  68,  80,
+  93,  107, 122, 138, 154, 170, 186, 202, 218, 234, 250, 266, 282, 298, 314,
+  330, 346, 362, 378, 393, 407, 9,   13,  18,  24,  31,  39,  48,  58,  69,
+  81,  94,  108, 123, 139, 155, 171, 187, 203, 219, 235, 251, 267, 283, 299,
+  315, 331, 347, 363, 379, 394, 408, 421, 14,  19,  25,  32,  40,  49,  59,
+  70,  82,  95,  109, 124, 140, 156, 172, 188, 204, 220, 236, 252, 268, 284,
+  300, 316, 332, 348, 364, 380, 395, 409, 422, 434, 20,  26,  33,  41,  50,
+  60,  71,  83,  96,  110, 125, 141, 157, 173, 189, 205, 221, 237, 253, 269,
+  285, 301, 317, 333, 349, 365, 381, 396, 410, 423, 435, 446, 27,  34,  42,
+  51,  61,  72,  84,  97,  111, 126, 142, 158, 174, 190, 206, 222, 238, 254,
+  270, 286, 302, 318, 334, 350, 366, 382, 397, 411, 424, 436, 447, 457, 35,
+  43,  52,  62,  73,  85,  98,  112, 127, 143, 159, 175, 191, 207, 223, 239,
+  255, 271, 287, 303, 319, 335, 351, 367, 383, 398, 412, 425, 437, 448, 458,
+  467, 44,  53,  63,  74,  86,  99,  113, 128, 144, 160, 176, 192, 208, 224,
+  240, 256, 272, 288, 304, 320, 336, 352, 368, 384, 399, 413, 426, 438, 449,
+  459, 468, 476, 54,  64,  75,  87,  100, 114, 129, 145, 161, 177, 193, 209,
+  225, 241, 257, 273, 289, 305, 321, 337, 353, 369, 385, 400, 414, 427, 439,
+  450, 460, 469, 477, 484, 65,  76,  88,  101, 115, 130, 146, 162, 178, 194,
+  210, 226, 242, 258, 274, 290, 306, 322, 338, 354, 370, 386, 401, 415, 428,
+  440, 451, 461, 470, 478, 485, 491, 77,  89,  102, 116, 131, 147, 163, 179,
+  195, 211, 227, 243, 259, 275, 291, 307, 323, 339, 355, 371, 387, 402, 416,
+  429, 441, 452, 462, 471, 479, 486, 492, 497, 90,  103, 117, 132, 148, 164,
+  180, 196, 212, 228, 244, 260, 276, 292, 308, 324, 340, 356, 372, 388, 403,
+  417, 430, 442, 453, 463, 472, 480, 487, 493, 498, 502, 104, 118, 133, 149,
+  165, 181, 197, 213, 229, 245, 261, 277, 293, 309, 325, 341, 357, 373, 389,
+  404, 418, 431, 443, 454, 464, 473, 481, 488, 494, 499, 503, 506, 119, 134,
+  150, 166, 182, 198, 214, 230, 246, 262, 278, 294, 310, 326, 342, 358, 374,
+  390, 405, 419, 432, 444, 455, 465, 474, 482, 489, 495, 500, 504, 507, 509,
+  135, 151, 167, 183, 199, 215, 231, 247, 263, 279, 295, 311, 327, 343, 359,
+  375, 391, 406, 420, 433, 445, 456, 466, 475, 483, 490, 496, 501, 505, 508,
+  510, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_16x32[512]) = {
+  0,  32, 64, 96,  128, 160, 192, 224, 256, 288, 320, 352, 384, 416, 448, 480,
+  1,  33, 65, 97,  129, 161, 193, 225, 257, 289, 321, 353, 385, 417, 449, 481,
+  2,  34, 66, 98,  130, 162, 194, 226, 258, 290, 322, 354, 386, 418, 450, 482,
+  3,  35, 67, 99,  131, 163, 195, 227, 259, 291, 323, 355, 387, 419, 451, 483,
+  4,  36, 68, 100, 132, 164, 196, 228, 260, 292, 324, 356, 388, 420, 452, 484,
+  5,  37, 69, 101, 133, 165, 197, 229, 261, 293, 325, 357, 389, 421, 453, 485,
+  6,  38, 70, 102, 134, 166, 198, 230, 262, 294, 326, 358, 390, 422, 454, 486,
+  7,  39, 71, 103, 135, 167, 199, 231, 263, 295, 327, 359, 391, 423, 455, 487,
+  8,  40, 72, 104, 136, 168, 200, 232, 264, 296, 328, 360, 392, 424, 456, 488,
+  9,  41, 73, 105, 137, 169, 201, 233, 265, 297, 329, 361, 393, 425, 457, 489,
+  10, 42, 74, 106, 138, 170, 202, 234, 266, 298, 330, 362, 394, 426, 458, 490,
+  11, 43, 75, 107, 139, 171, 203, 235, 267, 299, 331, 363, 395, 427, 459, 491,
+  12, 44, 76, 108, 140, 172, 204, 236, 268, 300, 332, 364, 396, 428, 460, 492,
+  13, 45, 77, 109, 141, 173, 205, 237, 269, 301, 333, 365, 397, 429, 461, 493,
+  14, 46, 78, 110, 142, 174, 206, 238, 270, 302, 334, 366, 398, 430, 462, 494,
+  15, 47, 79, 111, 143, 175, 207, 239, 271, 303, 335, 367, 399, 431, 463, 495,
+  16, 48, 80, 112, 144, 176, 208, 240, 272, 304, 336, 368, 400, 432, 464, 496,
+  17, 49, 81, 113, 145, 177, 209, 241, 273, 305, 337, 369, 401, 433, 465, 497,
+  18, 50, 82, 114, 146, 178, 210, 242, 274, 306, 338, 370, 402, 434, 466, 498,
+  19, 51, 83, 115, 147, 179, 211, 243, 275, 307, 339, 371, 403, 435, 467, 499,
+  20, 52, 84, 116, 148, 180, 212, 244, 276, 308, 340, 372, 404, 436, 468, 500,
+  21, 53, 85, 117, 149, 181, 213, 245, 277, 309, 341, 373, 405, 437, 469, 501,
+  22, 54, 86, 118, 150, 182, 214, 246, 278, 310, 342, 374, 406, 438, 470, 502,
+  23, 55, 87, 119, 151, 183, 215, 247, 279, 311, 343, 375, 407, 439, 471, 503,
+  24, 56, 88, 120, 152, 184, 216, 248, 280, 312, 344, 376, 408, 440, 472, 504,
+  25, 57, 89, 121, 153, 185, 217, 249, 281, 313, 345, 377, 409, 441, 473, 505,
+  26, 58, 90, 122, 154, 186, 218, 250, 282, 314, 346, 378, 410, 442, 474, 506,
+  27, 59, 91, 123, 155, 187, 219, 251, 283, 315, 347, 379, 411, 443, 475, 507,
+  28, 60, 92, 124, 156, 188, 220, 252, 284, 316, 348, 380, 412, 444, 476, 508,
+  29, 61, 93, 125, 157, 189, 221, 253, 285, 317, 349, 381, 413, 445, 477, 509,
+  30, 62, 94, 126, 158, 190, 222, 254, 286, 318, 350, 382, 414, 446, 478, 510,
+  31, 63, 95, 127, 159, 191, 223, 255, 287, 319, 351, 383, 415, 447, 479, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_32x16[512]) = {
+  0,   16,  32,  48,  64,  80,  96,  112, 128, 144, 160, 176, 192, 208, 224,
+  240, 256, 272, 288, 304, 320, 336, 352, 368, 384, 400, 416, 432, 448, 464,
+  480, 496, 1,   17,  33,  49,  65,  81,  97,  113, 129, 145, 161, 177, 193,
+  209, 225, 241, 257, 273, 289, 305, 321, 337, 353, 369, 385, 401, 417, 433,
+  449, 465, 481, 497, 2,   18,  34,  50,  66,  82,  98,  114, 130, 146, 162,
+  178, 194, 210, 226, 242, 258, 274, 290, 306, 322, 338, 354, 370, 386, 402,
+  418, 434, 450, 466, 482, 498, 3,   19,  35,  51,  67,  83,  99,  115, 131,
+  147, 163, 179, 195, 211, 227, 243, 259, 275, 291, 307, 323, 339, 355, 371,
+  387, 403, 419, 435, 451, 467, 483, 499, 4,   20,  36,  52,  68,  84,  100,
+  116, 132, 148, 164, 180, 196, 212, 228, 244, 260, 276, 292, 308, 324, 340,
+  356, 372, 388, 404, 420, 436, 452, 468, 484, 500, 5,   21,  37,  53,  69,
+  85,  101, 117, 133, 149, 165, 181, 197, 213, 229, 245, 261, 277, 293, 309,
+  325, 341, 357, 373, 389, 405, 421, 437, 453, 469, 485, 501, 6,   22,  38,
+  54,  70,  86,  102, 118, 134, 150, 166, 182, 198, 214, 230, 246, 262, 278,
+  294, 310, 326, 342, 358, 374, 390, 406, 422, 438, 454, 470, 486, 502, 7,
+  23,  39,  55,  71,  87,  103, 119, 135, 151, 167, 183, 199, 215, 231, 247,
+  263, 279, 295, 311, 327, 343, 359, 375, 391, 407, 423, 439, 455, 471, 487,
+  503, 8,   24,  40,  56,  72,  88,  104, 120, 136, 152, 168, 184, 200, 216,
+  232, 248, 264, 280, 296, 312, 328, 344, 360, 376, 392, 408, 424, 440, 456,
+  472, 488, 504, 9,   25,  41,  57,  73,  89,  105, 121, 137, 153, 169, 185,
+  201, 217, 233, 249, 265, 281, 297, 313, 329, 345, 361, 377, 393, 409, 425,
+  441, 457, 473, 489, 505, 10,  26,  42,  58,  74,  90,  106, 122, 138, 154,
+  170, 186, 202, 218, 234, 250, 266, 282, 298, 314, 330, 346, 362, 378, 394,
+  410, 426, 442, 458, 474, 490, 506, 11,  27,  43,  59,  75,  91,  107, 123,
+  139, 155, 171, 187, 203, 219, 235, 251, 267, 283, 299, 315, 331, 347, 363,
+  379, 395, 411, 427, 443, 459, 475, 491, 507, 12,  28,  44,  60,  76,  92,
+  108, 124, 140, 156, 172, 188, 204, 220, 236, 252, 268, 284, 300, 316, 332,
+  348, 364, 380, 396, 412, 428, 444, 460, 476, 492, 508, 13,  29,  45,  61,
+  77,  93,  109, 125, 141, 157, 173, 189, 205, 221, 237, 253, 269, 285, 301,
+  317, 333, 349, 365, 381, 397, 413, 429, 445, 461, 477, 493, 509, 14,  30,
+  46,  62,  78,  94,  110, 126, 142, 158, 174, 190, 206, 222, 238, 254, 270,
+  286, 302, 318, 334, 350, 366, 382, 398, 414, 430, 446, 462, 478, 494, 510,
+  15,  31,  47,  63,  79,  95,  111, 127, 143, 159, 175, 191, 207, 223, 239,
+  255, 271, 287, 303, 319, 335, 351, 367, 383, 399, 415, 431, 447, 463, 479,
+  495, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_16x32[512]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134,
+  135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149,
+  150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164,
+  165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
+  180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194,
+  195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209,
+  210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224,
+  225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
+  240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254,
+  255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269,
+  270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284,
+  285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299,
+  300, 301, 302, 303, 304, 305, 306, 307, 308, 309, 310, 311, 312, 313, 314,
+  315, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329,
+  330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 343, 344,
+  345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359,
+  360, 361, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 373, 374,
+  375, 376, 377, 378, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389,
+  390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404,
+  405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419,
+  420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431, 432, 433, 434,
+  435, 436, 437, 438, 439, 440, 441, 442, 443, 444, 445, 446, 447, 448, 449,
+  450, 451, 452, 453, 454, 455, 456, 457, 458, 459, 460, 461, 462, 463, 464,
+  465, 466, 467, 468, 469, 470, 471, 472, 473, 474, 475, 476, 477, 478, 479,
+  480, 481, 482, 483, 484, 485, 486, 487, 488, 489, 490, 491, 492, 493, 494,
+  495, 496, 497, 498, 499, 500, 501, 502, 503, 504, 505, 506, 507, 508, 509,
+  510, 511,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_32x16[512]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134,
+  135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149,
+  150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164,
+  165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
+  180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194,
+  195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209,
+  210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224,
+  225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
+  240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254,
+  255, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269,
+  270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284,
+  285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 297, 298, 299,
+  300, 301, 302, 303, 304, 305, 306, 307, 308, 309, 310, 311, 312, 313, 314,
+  315, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329,
+  330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 343, 344,
+  345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359,
+  360, 361, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 373, 374,
+  375, 376, 377, 378, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389,
+  390, 391, 392, 393, 394, 395, 396, 397, 398, 399, 400, 401, 402, 403, 404,
+  405, 406, 407, 408, 409, 410, 411, 412, 413, 414, 415, 416, 417, 418, 419,
+  420, 421, 422, 423, 424, 425, 426, 427, 428, 429, 430, 431, 432, 433, 434,
+  435, 436, 437, 438, 439, 440, 441, 442, 443, 444, 445, 446, 447, 448, 449,
+  450, 451, 452, 453, 454, 455, 456, 457, 458, 459, 460, 461, 462, 463, 464,
+  465, 466, 467, 468, 469, 470, 471, 472, 473, 474, 475, 476, 477, 478, 479,
+  480, 481, 482, 483, 484, 485, 486, 487, 488, 489, 490, 491, 492, 493, 494,
+  495, 496, 497, 498, 499, 500, 501, 502, 503, 504, 505, 506, 507, 508, 509,
+  510, 511,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_16x16[256]) = {
+  0,  16, 32, 48, 64, 80, 96,  112, 128, 144, 160, 176, 192, 208, 224, 240,
+  1,  17, 33, 49, 65, 81, 97,  113, 129, 145, 161, 177, 193, 209, 225, 241,
+  2,  18, 34, 50, 66, 82, 98,  114, 130, 146, 162, 178, 194, 210, 226, 242,
+  3,  19, 35, 51, 67, 83, 99,  115, 131, 147, 163, 179, 195, 211, 227, 243,
+  4,  20, 36, 52, 68, 84, 100, 116, 132, 148, 164, 180, 196, 212, 228, 244,
+  5,  21, 37, 53, 69, 85, 101, 117, 133, 149, 165, 181, 197, 213, 229, 245,
+  6,  22, 38, 54, 70, 86, 102, 118, 134, 150, 166, 182, 198, 214, 230, 246,
+  7,  23, 39, 55, 71, 87, 103, 119, 135, 151, 167, 183, 199, 215, 231, 247,
+  8,  24, 40, 56, 72, 88, 104, 120, 136, 152, 168, 184, 200, 216, 232, 248,
+  9,  25, 41, 57, 73, 89, 105, 121, 137, 153, 169, 185, 201, 217, 233, 249,
+  10, 26, 42, 58, 74, 90, 106, 122, 138, 154, 170, 186, 202, 218, 234, 250,
+  11, 27, 43, 59, 75, 91, 107, 123, 139, 155, 171, 187, 203, 219, 235, 251,
+  12, 28, 44, 60, 76, 92, 108, 124, 140, 156, 172, 188, 204, 220, 236, 252,
+  13, 29, 45, 61, 77, 93, 109, 125, 141, 157, 173, 189, 205, 221, 237, 253,
+  14, 30, 46, 62, 78, 94, 110, 126, 142, 158, 174, 190, 206, 222, 238, 254,
+  15, 31, 47, 63, 79, 95, 111, 127, 143, 159, 175, 191, 207, 223, 239, 255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_16x16[256]) = {
+  0,   1,   2,   3,   4,   5,   6,   7,   8,   9,   10,  11,  12,  13,  14,
+  15,  16,  17,  18,  19,  20,  21,  22,  23,  24,  25,  26,  27,  28,  29,
+  30,  31,  32,  33,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,  44,
+  45,  46,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  59,
+  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  74,
+  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  86,  87,  88,  89,
+  90,  91,  92,  93,  94,  95,  96,  97,  98,  99,  100, 101, 102, 103, 104,
+  105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119,
+  120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134,
+  135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149,
+  150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160, 161, 162, 163, 164,
+  165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175, 176, 177, 178, 179,
+  180, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 191, 192, 193, 194,
+  195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209,
+  210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224,
+  225, 226, 227, 228, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238, 239,
+  240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 254,
+  255,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, av1_col_iscan_16x16[256]) = {
+  0,  4,  11,  20,  31,  43,  59,  75,  85,  109, 130, 150, 165, 181, 195, 198,
+  1,  6,  14,  23,  34,  47,  64,  81,  95,  114, 135, 153, 171, 188, 201, 212,
+  2,  8,  16,  25,  38,  52,  67,  83,  101, 116, 136, 157, 172, 190, 205, 216,
+  3,  10, 18,  29,  41,  55,  71,  89,  103, 119, 141, 159, 176, 194, 208, 218,
+  5,  12, 21,  32,  45,  58,  74,  93,  104, 123, 144, 164, 179, 196, 210, 223,
+  7,  15, 26,  37,  49,  63,  78,  96,  112, 129, 146, 166, 182, 200, 215, 228,
+  9,  19, 28,  39,  54,  69,  86,  102, 117, 132, 151, 170, 187, 206, 220, 230,
+  13, 24, 35,  46,  60,  73,  91,  108, 122, 137, 154, 174, 189, 207, 224, 235,
+  17, 30, 40,  53,  66,  82,  98,  115, 126, 142, 161, 180, 197, 213, 227, 237,
+  22, 36, 48,  62,  76,  92,  105, 120, 133, 147, 167, 186, 203, 219, 232, 240,
+  27, 44, 56,  70,  84,  99,  113, 127, 140, 156, 175, 193, 209, 226, 236, 244,
+  33, 51, 68,  79,  94,  110, 125, 138, 149, 162, 184, 202, 217, 229, 241, 247,
+  42, 61, 77,  90,  106, 121, 134, 148, 160, 173, 191, 211, 225, 238, 245, 251,
+  50, 72, 87,  100, 118, 128, 145, 158, 168, 183, 204, 222, 233, 242, 249, 253,
+  57, 80, 97,  111, 131, 143, 155, 169, 178, 192, 214, 231, 239, 246, 250, 254,
+  65, 88, 107, 124, 139, 152, 163, 177, 185, 199, 221, 234, 243, 248, 252, 255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_row_iscan_16x16[256]) = {
+  0,   1,   2,   4,   6,   9,   12,  17,  22,  29,  36,  43,  54,  64,  76,
+  86,  3,   5,   7,   11,  15,  19,  25,  32,  38,  48,  59,  68,  84,  99,
+  115, 130, 8,   10,  13,  18,  23,  27,  33,  42,  51,  60,  72,  88,  103,
+  119, 142, 167, 14,  16,  20,  26,  31,  37,  44,  53,  61,  73,  85,  100,
+  116, 135, 161, 185, 21,  24,  30,  35,  40,  47,  55,  65,  74,  81,  94,
+  112, 133, 154, 179, 205, 28,  34,  39,  45,  50,  58,  67,  77,  87,  96,
+  106, 121, 146, 169, 196, 212, 41,  46,  49,  56,  63,  70,  79,  90,  98,
+  107, 122, 138, 159, 182, 207, 222, 52,  57,  62,  69,  75,  83,  93,  102,
+  110, 120, 134, 150, 176, 195, 215, 226, 66,  71,  78,  82,  91,  97,  108,
+  113, 127, 136, 148, 168, 188, 202, 221, 232, 80,  89,  92,  101, 105, 114,
+  125, 131, 139, 151, 162, 177, 192, 208, 223, 234, 95,  104, 109, 117, 123,
+  128, 143, 144, 155, 165, 175, 190, 206, 219, 233, 239, 111, 118, 124, 129,
+  140, 147, 157, 164, 170, 181, 191, 203, 224, 230, 240, 243, 126, 132, 137,
+  145, 153, 160, 174, 178, 184, 197, 204, 216, 231, 237, 244, 246, 141, 149,
+  156, 166, 172, 180, 189, 199, 200, 210, 220, 228, 238, 242, 249, 251, 152,
+  163, 171, 183, 186, 193, 201, 211, 214, 218, 227, 236, 245, 247, 252, 253,
+  158, 173, 187, 194, 198, 209, 213, 217, 225, 229, 235, 241, 248, 250, 254,
+  255,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_16x16[256]) = {
+  0,   2,   5,   9,   17,  24,  36,  44,  55,  72,  88,  104, 128, 143, 166,
+  179, 1,   4,   8,   13,  20,  30,  40,  54,  66,  79,  96,  113, 141, 154,
+  178, 196, 3,   7,   11,  18,  25,  33,  46,  57,  71,  86,  101, 119, 148,
+  164, 186, 201, 6,   12,  16,  23,  31,  39,  53,  64,  78,  92,  110, 127,
+  153, 169, 193, 208, 10,  14,  19,  28,  37,  47,  58,  67,  84,  98,  114,
+  133, 161, 176, 198, 214, 15,  21,  26,  34,  43,  52,  65,  77,  91,  106,
+  120, 140, 165, 185, 205, 221, 22,  27,  32,  41,  48,  60,  73,  85,  99,
+  116, 130, 151, 175, 190, 211, 225, 29,  35,  42,  49,  59,  69,  81,  95,
+  108, 125, 139, 155, 182, 197, 217, 229, 38,  45,  51,  61,  68,  80,  93,
+  105, 118, 134, 150, 168, 191, 207, 223, 234, 50,  56,  63,  74,  83,  94,
+  109, 117, 129, 147, 163, 177, 199, 213, 228, 238, 62,  70,  76,  87,  97,
+  107, 122, 131, 145, 159, 172, 188, 210, 222, 235, 242, 75,  82,  90,  102,
+  112, 124, 138, 146, 157, 173, 187, 202, 219, 230, 240, 245, 89,  100, 111,
+  123, 132, 142, 156, 167, 180, 189, 203, 216, 231, 237, 246, 250, 103, 115,
+  126, 136, 149, 162, 171, 183, 194, 204, 215, 224, 236, 241, 248, 252, 121,
+  135, 144, 158, 170, 181, 192, 200, 209, 218, 227, 233, 243, 244, 251, 254,
+  137, 152, 160, 174, 184, 195, 206, 212, 220, 226, 232, 239, 247, 249, 253,
+  255,
+};
+
+#if CONFIG_EXT_TX
+DECLARE_ALIGNED(16, static const int16_t, av1_mcol_iscan_32x32[1024]) = {
+  0,   32,   64,  96,   128, 160,  192, 224,  256, 288,  320, 352,  384, 416,
+  448, 480,  512, 544,  576, 608,  640, 672,  704, 736,  768, 800,  832, 864,
+  896, 928,  960, 992,  1,   33,   65,  97,   129, 161,  193, 225,  257, 289,
+  321, 353,  385, 417,  449, 481,  513, 545,  577, 609,  641, 673,  705, 737,
+  769, 801,  833, 865,  897, 929,  961, 993,  2,   34,   66,  98,   130, 162,
+  194, 226,  258, 290,  322, 354,  386, 418,  450, 482,  514, 546,  578, 610,
+  642, 674,  706, 738,  770, 802,  834, 866,  898, 930,  962, 994,  3,   35,
+  67,  99,   131, 163,  195, 227,  259, 291,  323, 355,  387, 419,  451, 483,
+  515, 547,  579, 611,  643, 675,  707, 739,  771, 803,  835, 867,  899, 931,
+  963, 995,  4,   36,   68,  100,  132, 164,  196, 228,  260, 292,  324, 356,
+  388, 420,  452, 484,  516, 548,  580, 612,  644, 676,  708, 740,  772, 804,
+  836, 868,  900, 932,  964, 996,  5,   37,   69,  101,  133, 165,  197, 229,
+  261, 293,  325, 357,  389, 421,  453, 485,  517, 549,  581, 613,  645, 677,
+  709, 741,  773, 805,  837, 869,  901, 933,  965, 997,  6,   38,   70,  102,
+  134, 166,  198, 230,  262, 294,  326, 358,  390, 422,  454, 486,  518, 550,
+  582, 614,  646, 678,  710, 742,  774, 806,  838, 870,  902, 934,  966, 998,
+  7,   39,   71,  103,  135, 167,  199, 231,  263, 295,  327, 359,  391, 423,
+  455, 487,  519, 551,  583, 615,  647, 679,  711, 743,  775, 807,  839, 871,
+  903, 935,  967, 999,  8,   40,   72,  104,  136, 168,  200, 232,  264, 296,
+  328, 360,  392, 424,  456, 488,  520, 552,  584, 616,  648, 680,  712, 744,
+  776, 808,  840, 872,  904, 936,  968, 1000, 9,   41,   73,  105,  137, 169,
+  201, 233,  265, 297,  329, 361,  393, 425,  457, 489,  521, 553,  585, 617,
+  649, 681,  713, 745,  777, 809,  841, 873,  905, 937,  969, 1001, 10,  42,
+  74,  106,  138, 170,  202, 234,  266, 298,  330, 362,  394, 426,  458, 490,
+  522, 554,  586, 618,  650, 682,  714, 746,  778, 810,  842, 874,  906, 938,
+  970, 1002, 11,  43,   75,  107,  139, 171,  203, 235,  267, 299,  331, 363,
+  395, 427,  459, 491,  523, 555,  587, 619,  651, 683,  715, 747,  779, 811,
+  843, 875,  907, 939,  971, 1003, 12,  44,   76,  108,  140, 172,  204, 236,
+  268, 300,  332, 364,  396, 428,  460, 492,  524, 556,  588, 620,  652, 684,
+  716, 748,  780, 812,  844, 876,  908, 940,  972, 1004, 13,  45,   77,  109,
+  141, 173,  205, 237,  269, 301,  333, 365,  397, 429,  461, 493,  525, 557,
+  589, 621,  653, 685,  717, 749,  781, 813,  845, 877,  909, 941,  973, 1005,
+  14,  46,   78,  110,  142, 174,  206, 238,  270, 302,  334, 366,  398, 430,
+  462, 494,  526, 558,  590, 622,  654, 686,  718, 750,  782, 814,  846, 878,
+  910, 942,  974, 1006, 15,  47,   79,  111,  143, 175,  207, 239,  271, 303,
+  335, 367,  399, 431,  463, 495,  527, 559,  591, 623,  655, 687,  719, 751,
+  783, 815,  847, 879,  911, 943,  975, 1007, 16,  48,   80,  112,  144, 176,
+  208, 240,  272, 304,  336, 368,  400, 432,  464, 496,  528, 560,  592, 624,
+  656, 688,  720, 752,  784, 816,  848, 880,  912, 944,  976, 1008, 17,  49,
+  81,  113,  145, 177,  209, 241,  273, 305,  337, 369,  401, 433,  465, 497,
+  529, 561,  593, 625,  657, 689,  721, 753,  785, 817,  849, 881,  913, 945,
+  977, 1009, 18,  50,   82,  114,  146, 178,  210, 242,  274, 306,  338, 370,
+  402, 434,  466, 498,  530, 562,  594, 626,  658, 690,  722, 754,  786, 818,
+  850, 882,  914, 946,  978, 1010, 19,  51,   83,  115,  147, 179,  211, 243,
+  275, 307,  339, 371,  403, 435,  467, 499,  531, 563,  595, 627,  659, 691,
+  723, 755,  787, 819,  851, 883,  915, 947,  979, 1011, 20,  52,   84,  116,
+  148, 180,  212, 244,  276, 308,  340, 372,  404, 436,  468, 500,  532, 564,
+  596, 628,  660, 692,  724, 756,  788, 820,  852, 884,  916, 948,  980, 1012,
+  21,  53,   85,  117,  149, 181,  213, 245,  277, 309,  341, 373,  405, 437,
+  469, 501,  533, 565,  597, 629,  661, 693,  725, 757,  789, 821,  853, 885,
+  917, 949,  981, 1013, 22,  54,   86,  118,  150, 182,  214, 246,  278, 310,
+  342, 374,  406, 438,  470, 502,  534, 566,  598, 630,  662, 694,  726, 758,
+  790, 822,  854, 886,  918, 950,  982, 1014, 23,  55,   87,  119,  151, 183,
+  215, 247,  279, 311,  343, 375,  407, 439,  471, 503,  535, 567,  599, 631,
+  663, 695,  727, 759,  791, 823,  855, 887,  919, 951,  983, 1015, 24,  56,
+  88,  120,  152, 184,  216, 248,  280, 312,  344, 376,  408, 440,  472, 504,
+  536, 568,  600, 632,  664, 696,  728, 760,  792, 824,  856, 888,  920, 952,
+  984, 1016, 25,  57,   89,  121,  153, 185,  217, 249,  281, 313,  345, 377,
+  409, 441,  473, 505,  537, 569,  601, 633,  665, 697,  729, 761,  793, 825,
+  857, 889,  921, 953,  985, 1017, 26,  58,   90,  122,  154, 186,  218, 250,
+  282, 314,  346, 378,  410, 442,  474, 506,  538, 570,  602, 634,  666, 698,
+  730, 762,  794, 826,  858, 890,  922, 954,  986, 1018, 27,  59,   91,  123,
+  155, 187,  219, 251,  283, 315,  347, 379,  411, 443,  475, 507,  539, 571,
+  603, 635,  667, 699,  731, 763,  795, 827,  859, 891,  923, 955,  987, 1019,
+  28,  60,   92,  124,  156, 188,  220, 252,  284, 316,  348, 380,  412, 444,
+  476, 508,  540, 572,  604, 636,  668, 700,  732, 764,  796, 828,  860, 892,
+  924, 956,  988, 1020, 29,  61,   93,  125,  157, 189,  221, 253,  285, 317,
+  349, 381,  413, 445,  477, 509,  541, 573,  605, 637,  669, 701,  733, 765,
+  797, 829,  861, 893,  925, 957,  989, 1021, 30,  62,   94,  126,  158, 190,
+  222, 254,  286, 318,  350, 382,  414, 446,  478, 510,  542, 574,  606, 638,
+  670, 702,  734, 766,  798, 830,  862, 894,  926, 958,  990, 1022, 31,  63,
+  95,  127,  159, 191,  223, 255,  287, 319,  351, 383,  415, 447,  479, 511,
+  543, 575,  607, 639,  671, 703,  735, 767,  799, 831,  863, 895,  927, 959,
+  991, 1023,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_mrow_iscan_32x32[1024]) = {
+  0,    1,    2,    3,    4,    5,    6,    7,    8,    9,    10,   11,   12,
+  13,   14,   15,   16,   17,   18,   19,   20,   21,   22,   23,   24,   25,
+  26,   27,   28,   29,   30,   31,   32,   33,   34,   35,   36,   37,   38,
+  39,   40,   41,   42,   43,   44,   45,   46,   47,   48,   49,   50,   51,
+  52,   53,   54,   55,   56,   57,   58,   59,   60,   61,   62,   63,   64,
+  65,   66,   67,   68,   69,   70,   71,   72,   73,   74,   75,   76,   77,
+  78,   79,   80,   81,   82,   83,   84,   85,   86,   87,   88,   89,   90,
+  91,   92,   93,   94,   95,   96,   97,   98,   99,   100,  101,  102,  103,
+  104,  105,  106,  107,  108,  109,  110,  111,  112,  113,  114,  115,  116,
+  117,  118,  119,  120,  121,  122,  123,  124,  125,  126,  127,  128,  129,
+  130,  131,  132,  133,  134,  135,  136,  137,  138,  139,  140,  141,  142,
+  143,  144,  145,  146,  147,  148,  149,  150,  151,  152,  153,  154,  155,
+  156,  157,  158,  159,  160,  161,  162,  163,  164,  165,  166,  167,  168,
+  169,  170,  171,  172,  173,  174,  175,  176,  177,  178,  179,  180,  181,
+  182,  183,  184,  185,  186,  187,  188,  189,  190,  191,  192,  193,  194,
+  195,  196,  197,  198,  199,  200,  201,  202,  203,  204,  205,  206,  207,
+  208,  209,  210,  211,  212,  213,  214,  215,  216,  217,  218,  219,  220,
+  221,  222,  223,  224,  225,  226,  227,  228,  229,  230,  231,  232,  233,
+  234,  235,  236,  237,  238,  239,  240,  241,  242,  243,  244,  245,  246,
+  247,  248,  249,  250,  251,  252,  253,  254,  255,  256,  257,  258,  259,
+  260,  261,  262,  263,  264,  265,  266,  267,  268,  269,  270,  271,  272,
+  273,  274,  275,  276,  277,  278,  279,  280,  281,  282,  283,  284,  285,
+  286,  287,  288,  289,  290,  291,  292,  293,  294,  295,  296,  297,  298,
+  299,  300,  301,  302,  303,  304,  305,  306,  307,  308,  309,  310,  311,
+  312,  313,  314,  315,  316,  317,  318,  319,  320,  321,  322,  323,  324,
+  325,  326,  327,  328,  329,  330,  331,  332,  333,  334,  335,  336,  337,
+  338,  339,  340,  341,  342,  343,  344,  345,  346,  347,  348,  349,  350,
+  351,  352,  353,  354,  355,  356,  357,  358,  359,  360,  361,  362,  363,
+  364,  365,  366,  367,  368,  369,  370,  371,  372,  373,  374,  375,  376,
+  377,  378,  379,  380,  381,  382,  383,  384,  385,  386,  387,  388,  389,
+  390,  391,  392,  393,  394,  395,  396,  397,  398,  399,  400,  401,  402,
+  403,  404,  405,  406,  407,  408,  409,  410,  411,  412,  413,  414,  415,
+  416,  417,  418,  419,  420,  421,  422,  423,  424,  425,  426,  427,  428,
+  429,  430,  431,  432,  433,  434,  435,  436,  437,  438,  439,  440,  441,
+  442,  443,  444,  445,  446,  447,  448,  449,  450,  451,  452,  453,  454,
+  455,  456,  457,  458,  459,  460,  461,  462,  463,  464,  465,  466,  467,
+  468,  469,  470,  471,  472,  473,  474,  475,  476,  477,  478,  479,  480,
+  481,  482,  483,  484,  485,  486,  487,  488,  489,  490,  491,  492,  493,
+  494,  495,  496,  497,  498,  499,  500,  501,  502,  503,  504,  505,  506,
+  507,  508,  509,  510,  511,  512,  513,  514,  515,  516,  517,  518,  519,
+  520,  521,  522,  523,  524,  525,  526,  527,  528,  529,  530,  531,  532,
+  533,  534,  535,  536,  537,  538,  539,  540,  541,  542,  543,  544,  545,
+  546,  547,  548,  549,  550,  551,  552,  553,  554,  555,  556,  557,  558,
+  559,  560,  561,  562,  563,  564,  565,  566,  567,  568,  569,  570,  571,
+  572,  573,  574,  575,  576,  577,  578,  579,  580,  581,  582,  583,  584,
+  585,  586,  587,  588,  589,  590,  591,  592,  593,  594,  595,  596,  597,
+  598,  599,  600,  601,  602,  603,  604,  605,  606,  607,  608,  609,  610,
+  611,  612,  613,  614,  615,  616,  617,  618,  619,  620,  621,  622,  623,
+  624,  625,  626,  627,  628,  629,  630,  631,  632,  633,  634,  635,  636,
+  637,  638,  639,  640,  641,  642,  643,  644,  645,  646,  647,  648,  649,
+  650,  651,  652,  653,  654,  655,  656,  657,  658,  659,  660,  661,  662,
+  663,  664,  665,  666,  667,  668,  669,  670,  671,  672,  673,  674,  675,
+  676,  677,  678,  679,  680,  681,  682,  683,  684,  685,  686,  687,  688,
+  689,  690,  691,  692,  693,  694,  695,  696,  697,  698,  699,  700,  701,
+  702,  703,  704,  705,  706,  707,  708,  709,  710,  711,  712,  713,  714,
+  715,  716,  717,  718,  719,  720,  721,  722,  723,  724,  725,  726,  727,
+  728,  729,  730,  731,  732,  733,  734,  735,  736,  737,  738,  739,  740,
+  741,  742,  743,  744,  745,  746,  747,  748,  749,  750,  751,  752,  753,
+  754,  755,  756,  757,  758,  759,  760,  761,  762,  763,  764,  765,  766,
+  767,  768,  769,  770,  771,  772,  773,  774,  775,  776,  777,  778,  779,
+  780,  781,  782,  783,  784,  785,  786,  787,  788,  789,  790,  791,  792,
+  793,  794,  795,  796,  797,  798,  799,  800,  801,  802,  803,  804,  805,
+  806,  807,  808,  809,  810,  811,  812,  813,  814,  815,  816,  817,  818,
+  819,  820,  821,  822,  823,  824,  825,  826,  827,  828,  829,  830,  831,
+  832,  833,  834,  835,  836,  837,  838,  839,  840,  841,  842,  843,  844,
+  845,  846,  847,  848,  849,  850,  851,  852,  853,  854,  855,  856,  857,
+  858,  859,  860,  861,  862,  863,  864,  865,  866,  867,  868,  869,  870,
+  871,  872,  873,  874,  875,  876,  877,  878,  879,  880,  881,  882,  883,
+  884,  885,  886,  887,  888,  889,  890,  891,  892,  893,  894,  895,  896,
+  897,  898,  899,  900,  901,  902,  903,  904,  905,  906,  907,  908,  909,
+  910,  911,  912,  913,  914,  915,  916,  917,  918,  919,  920,  921,  922,
+  923,  924,  925,  926,  927,  928,  929,  930,  931,  932,  933,  934,  935,
+  936,  937,  938,  939,  940,  941,  942,  943,  944,  945,  946,  947,  948,
+  949,  950,  951,  952,  953,  954,  955,  956,  957,  958,  959,  960,  961,
+  962,  963,  964,  965,  966,  967,  968,  969,  970,  971,  972,  973,  974,
+  975,  976,  977,  978,  979,  980,  981,  982,  983,  984,  985,  986,  987,
+  988,  989,  990,  991,  992,  993,  994,  995,  996,  997,  998,  999,  1000,
+  1001, 1002, 1003, 1004, 1005, 1006, 1007, 1008, 1009, 1010, 1011, 1012, 1013,
+  1014, 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022, 1023,
+};
+#endif  // CONFIG_EXT_TX
+
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_32x32[1024]) = {
+  0,    2,    5,    10,   17,   25,   38,   47,   62,   83,   101,  121,  145,
+  170,  193,  204,  210,  219,  229,  233,  245,  257,  275,  299,  342,  356,
+  377,  405,  455,  471,  495,  527,  1,    4,    8,    15,   22,   30,   45,
+  58,   74,   92,   112,  133,  158,  184,  203,  215,  222,  228,  234,  237,
+  256,  274,  298,  317,  355,  376,  404,  426,  470,  494,  526,  551,  3,
+  7,    12,   18,   28,   36,   52,   64,   82,   102,  118,  142,  164,  189,
+  208,  217,  224,  231,  235,  238,  273,  297,  316,  329,  375,  403,  425,
+  440,  493,  525,  550,  567,  6,    11,   16,   23,   31,   43,   60,   73,
+  90,   109,  126,  150,  173,  196,  211,  220,  226,  232,  236,  239,  296,
+  315,  328,  335,  402,  424,  439,  447,  524,  549,  566,  575,  9,    14,
+  19,   29,   37,   50,   65,   78,   95,   116,  134,  157,  179,  201,  214,
+  223,  244,  255,  272,  295,  341,  354,  374,  401,  454,  469,  492,  523,
+  582,  596,  617,  645,  13,   20,   26,   35,   44,   54,   72,   85,   105,
+  123,  140,  163,  182,  205,  216,  225,  254,  271,  294,  314,  353,  373,
+  400,  423,  468,  491,  522,  548,  595,  616,  644,  666,  21,   27,   33,
+  42,   53,   63,   80,   94,   113,  132,  151,  172,  190,  209,  218,  227,
+  270,  293,  313,  327,  372,  399,  422,  438,  490,  521,  547,  565,  615,
+  643,  665,  680,  24,   32,   39,   48,   57,   71,   88,   104,  120,  139,
+  159,  178,  197,  212,  221,  230,  292,  312,  326,  334,  398,  421,  437,
+  446,  520,  546,  564,  574,  642,  664,  679,  687,  34,   40,   46,   56,
+  68,   81,   96,   111,  130,  147,  167,  186,  243,  253,  269,  291,  340,
+  352,  371,  397,  453,  467,  489,  519,  581,  594,  614,  641,  693,  705,
+  723,  747,  41,   49,   55,   67,   77,   91,   107,  124,  138,  161,  177,
+  194,  252,  268,  290,  311,  351,  370,  396,  420,  466,  488,  518,  545,
+  593,  613,  640,  663,  704,  722,  746,  765,  51,   59,   66,   76,   89,
+  99,   119,  131,  149,  168,  181,  200,  267,  289,  310,  325,  369,  395,
+  419,  436,  487,  517,  544,  563,  612,  639,  662,  678,  721,  745,  764,
+  777,  61,   69,   75,   87,   100,  114,  129,  144,  162,  180,  191,  207,
+  288,  309,  324,  333,  394,  418,  435,  445,  516,  543,  562,  573,  638,
+  661,  677,  686,  744,  763,  776,  783,  70,   79,   86,   97,   108,  122,
+  137,  155,  242,  251,  266,  287,  339,  350,  368,  393,  452,  465,  486,
+  515,  580,  592,  611,  637,  692,  703,  720,  743,  788,  798,  813,  833,
+  84,   93,   103,  110,  125,  141,  154,  171,  250,  265,  286,  308,  349,
+  367,  392,  417,  464,  485,  514,  542,  591,  610,  636,  660,  702,  719,
+  742,  762,  797,  812,  832,  848,  98,   106,  115,  127,  143,  156,  169,
+  185,  264,  285,  307,  323,  366,  391,  416,  434,  484,  513,  541,  561,
+  609,  635,  659,  676,  718,  741,  761,  775,  811,  831,  847,  858,  117,
+  128,  136,  148,  160,  175,  188,  198,  284,  306,  322,  332,  390,  415,
+  433,  444,  512,  540,  560,  572,  634,  658,  675,  685,  740,  760,  774,
+  782,  830,  846,  857,  863,  135,  146,  152,  165,  241,  249,  263,  283,
+  338,  348,  365,  389,  451,  463,  483,  511,  579,  590,  608,  633,  691,
+  701,  717,  739,  787,  796,  810,  829,  867,  875,  887,  903,  153,  166,
+  174,  183,  248,  262,  282,  305,  347,  364,  388,  414,  462,  482,  510,
+  539,  589,  607,  632,  657,  700,  716,  738,  759,  795,  809,  828,  845,
+  874,  886,  902,  915,  176,  187,  195,  202,  261,  281,  304,  321,  363,
+  387,  413,  432,  481,  509,  538,  559,  606,  631,  656,  674,  715,  737,
+  758,  773,  808,  827,  844,  856,  885,  901,  914,  923,  192,  199,  206,
+  213,  280,  303,  320,  331,  386,  412,  431,  443,  508,  537,  558,  571,
+  630,  655,  673,  684,  736,  757,  772,  781,  826,  843,  855,  862,  900,
+  913,  922,  927,  240,  247,  260,  279,  337,  346,  362,  385,  450,  461,
+  480,  507,  578,  588,  605,  629,  690,  699,  714,  735,  786,  794,  807,
+  825,  866,  873,  884,  899,  930,  936,  945,  957,  246,  259,  278,  302,
+  345,  361,  384,  411,  460,  479,  506,  536,  587,  604,  628,  654,  698,
+  713,  734,  756,  793,  806,  824,  842,  872,  883,  898,  912,  935,  944,
+  956,  966,  258,  277,  301,  319,  360,  383,  410,  430,  478,  505,  535,
+  557,  603,  627,  653,  672,  712,  733,  755,  771,  805,  823,  841,  854,
+  882,  897,  911,  921,  943,  955,  965,  972,  276,  300,  318,  330,  382,
+  409,  429,  442,  504,  534,  556,  570,  626,  652,  671,  683,  732,  754,
+  770,  780,  822,  840,  853,  861,  896,  910,  920,  926,  954,  964,  971,
+  975,  336,  344,  359,  381,  449,  459,  477,  503,  577,  586,  602,  625,
+  689,  697,  711,  731,  785,  792,  804,  821,  865,  871,  881,  895,  929,
+  934,  942,  953,  977,  981,  987,  995,  343,  358,  380,  408,  458,  476,
+  502,  533,  585,  601,  624,  651,  696,  710,  730,  753,  791,  803,  820,
+  839,  870,  880,  894,  909,  933,  941,  952,  963,  980,  986,  994,  1001,
+  357,  379,  407,  428,  475,  501,  532,  555,  600,  623,  650,  670,  709,
+  729,  752,  769,  802,  819,  838,  852,  879,  893,  908,  919,  940,  951,
+  962,  970,  985,  993,  1000, 1005, 378,  406,  427,  441,  500,  531,  554,
+  569,  622,  649,  669,  682,  728,  751,  768,  779,  818,  837,  851,  860,
+  892,  907,  918,  925,  950,  961,  969,  974,  992,  999,  1004, 1007, 448,
+  457,  474,  499,  576,  584,  599,  621,  688,  695,  708,  727,  784,  790,
+  801,  817,  864,  869,  878,  891,  928,  932,  939,  949,  976,  979,  984,
+  991,  1008, 1010, 1013, 1017, 456,  473,  498,  530,  583,  598,  620,  648,
+  694,  707,  726,  750,  789,  800,  816,  836,  868,  877,  890,  906,  931,
+  938,  948,  960,  978,  983,  990,  998,  1009, 1012, 1016, 1020, 472,  497,
+  529,  553,  597,  619,  647,  668,  706,  725,  749,  767,  799,  815,  835,
+  850,  876,  889,  905,  917,  937,  947,  959,  968,  982,  989,  997,  1003,
+  1011, 1015, 1019, 1022, 496,  528,  552,  568,  618,  646,  667,  681,  724,
+  748,  766,  778,  814,  834,  849,  859,  888,  904,  916,  924,  946,  958,
+  967,  973,  988,  996,  1002, 1006, 1014, 1018, 1021, 1023,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_v2_iscan_32x32[1024]) = {
+  0,    1,    4,    9,    15,   22,   33,   43,   56,   71,   86,   104,  121,
+  142,  166,  189,  512,  518,  527,  539,  551,  566,  584,  602,  621,  644,
+  668,  695,  721,  748,  780,  811,  2,    3,    6,    11,   17,   26,   35,
+  45,   58,   73,   90,   106,  123,  146,  168,  193,  513,  519,  528,  540,
+  553,  567,  585,  603,  622,  647,  670,  696,  722,  751,  783,  812,  5,
+  7,    8,    13,   20,   28,   37,   50,   62,   75,   92,   108,  129,  150,
+  170,  195,  514,  521,  530,  541,  554,  569,  587,  605,  625,  649,  671,
+  699,  725,  752,  785,  815,  10,   12,   14,   19,   23,   31,   41,   52,
+  65,   81,   96,   113,  133,  152,  175,  201,  515,  522,  531,  542,  556,
+  572,  589,  607,  629,  651,  673,  700,  726,  757,  788,  819,  16,   18,
+  21,   24,   30,   39,   48,   59,   69,   83,   100,  119,  137,  158,  181,
+  203,  516,  523,  534,  545,  559,  574,  591,  610,  632,  654,  679,  704,
+  730,  762,  791,  824,  25,   27,   29,   32,   40,   46,   54,   67,   79,
+  94,   109,  127,  143,  164,  185,  210,  517,  525,  535,  547,  561,  578,
+  595,  615,  635,  656,  684,  707,  737,  766,  793,  830,  34,   36,   38,
+  42,   49,   55,   64,   76,   87,   102,  117,  135,  154,  176,  197,  219,
+  520,  529,  538,  550,  565,  580,  598,  618,  639,  664,  687,  712,  741,
+  769,  802,  833,  44,   47,   51,   53,   60,   68,   77,   85,   98,   114,
+  131,  147,  162,  183,  208,  227,  524,  533,  544,  557,  571,  588,  606,
+  623,  645,  667,  692,  720,  747,  776,  806,  838,  57,   61,   63,   66,
+  70,   80,   88,   99,   112,  124,  140,  159,  179,  199,  216,  233,  526,
+  536,  548,  562,  577,  593,  613,  633,  653,  676,  701,  727,  756,  786,
+  814,  847,  72,   74,   78,   82,   84,   95,   103,  115,  125,  139,  156,
+  173,  190,  211,  229,  246,  532,  543,  555,  568,  581,  601,  619,  637,
+  663,  685,  709,  738,  763,  792,  826,  855,  89,   91,   93,   97,   101,
+  110,  118,  132,  141,  157,  171,  186,  206,  224,  241,  255,  537,  549,
+  560,  576,  592,  608,  628,  650,  669,  693,  719,  744,  773,  805,  834,
+  862,  105,  107,  111,  116,  120,  128,  136,  148,  160,  174,  187,  205,
+  221,  236,  251,  267,  546,  558,  570,  583,  600,  617,  636,  657,  680,
+  706,  729,  758,  787,  813,  846,  871,  122,  126,  130,  134,  138,  144,
+  155,  163,  180,  191,  207,  222,  232,  248,  264,  278,  552,  564,  579,
+  594,  609,  630,  648,  666,  688,  715,  742,  768,  797,  827,  856,  877,
+  145,  149,  151,  153,  161,  165,  177,  184,  200,  212,  225,  237,  249,
+  262,  275,  289,  563,  575,  590,  604,  620,  638,  660,  683,  705,  728,
+  753,  779,  809,  839,  866,  889,  167,  169,  172,  178,  182,  188,  198,
+  209,  217,  230,  242,  252,  265,  276,  288,  301,  573,  586,  599,  616,
+  634,  652,  672,  694,  716,  743,  767,  794,  825,  850,  874,  899,  192,
+  194,  196,  202,  204,  213,  220,  228,  234,  247,  256,  268,  279,  290,
+  302,  315,  582,  597,  614,  631,  646,  665,  686,  708,  732,  759,  784,
+  810,  837,  863,  886,  908,  214,  215,  218,  223,  226,  231,  239,  244,
+  253,  261,  271,  283,  292,  304,  317,  325,  596,  611,  626,  642,  661,
+  681,  702,  723,  745,  770,  800,  828,  853,  875,  897,  919,  235,  238,
+  240,  243,  245,  250,  257,  263,  270,  280,  287,  298,  307,  319,  329,
+  340,  612,  624,  640,  658,  677,  697,  717,  739,  764,  789,  816,  844,
+  867,  890,  909,  927,  254,  258,  259,  260,  266,  269,  272,  282,  286,
+  296,  303,  312,  323,  333,  341,  355,  627,  641,  655,  674,  690,  713,
+  735,  760,  781,  807,  835,  857,  880,  902,  921,  940,  273,  274,  277,
+  281,  284,  285,  291,  299,  305,  310,  320,  327,  337,  346,  357,  369,
+  643,  659,  675,  689,  710,  733,  754,  777,  803,  831,  851,  872,  892,
+  913,  934,  950,  293,  294,  295,  297,  300,  306,  308,  314,  321,  326,
+  335,  343,  352,  361,  372,  378,  662,  678,  691,  711,  731,  749,  774,
+  798,  822,  848,  869,  887,  906,  925,  942,  961,  309,  311,  313,  316,
+  318,  322,  324,  332,  338,  344,  351,  358,  367,  375,  386,  394,  682,
+  698,  714,  734,  750,  772,  795,  820,  842,  864,  884,  904,  923,  938,
+  954,  967,  328,  330,  331,  334,  336,  339,  342,  348,  354,  359,  366,
+  374,  382,  391,  400,  409,  703,  718,  736,  755,  775,  796,  818,  840,
+  860,  882,  900,  917,  936,  952,  965,  977,  345,  347,  349,  350,  353,
+  356,  360,  364,  371,  376,  383,  389,  395,  406,  412,  423,  724,  740,
+  761,  778,  799,  821,  841,  859,  878,  895,  915,  932,  948,  963,  975,
+  986,  362,  363,  365,  368,  370,  373,  377,  379,  387,  392,  397,  405,
+  411,  420,  428,  439,  746,  765,  782,  804,  823,  843,  861,  879,  894,
+  911,  930,  946,  959,  973,  984,  994,  380,  381,  384,  385,  388,  390,
+  393,  396,  403,  408,  413,  422,  427,  436,  444,  452,  771,  790,  808,
+  832,  849,  865,  883,  896,  912,  928,  944,  957,  971,  982,  992,  1001,
+  398,  399,  401,  402,  404,  407,  410,  414,  419,  425,  429,  437,  442,
+  449,  458,  465,  801,  817,  836,  852,  870,  885,  901,  916,  931,  945,
+  956,  969,  980,  990,  999,  1007, 415,  416,  417,  418,  421,  424,  426,
+  430,  434,  441,  445,  453,  459,  463,  473,  480,  829,  845,  858,  873,
+  888,  905,  918,  933,  947,  958,  970,  979,  988,  997,  1005, 1012, 431,
+  432,  433,  435,  438,  440,  443,  446,  451,  456,  461,  468,  475,  479,
+  488,  494,  854,  868,  881,  893,  907,  924,  937,  949,  960,  972,  981,
+  989,  996,  1003, 1010, 1016, 447,  448,  450,  454,  455,  457,  460,  462,
+  469,  472,  477,  482,  490,  495,  499,  503,  876,  891,  903,  914,  926,
+  939,  953,  964,  974,  983,  991,  998,  1004, 1009, 1014, 1019, 464,  466,
+  467,  470,  471,  474,  476,  478,  484,  489,  493,  497,  501,  504,  506,
+  508,  898,  910,  922,  935,  943,  955,  966,  976,  985,  993,  1000, 1006,
+  1011, 1015, 1018, 1021, 481,  483,  485,  486,  487,  491,  492,  496,  498,
+  500,  502,  505,  507,  509,  510,  511,  920,  929,  941,  951,  962,  968,
+  978,  987,  995,  1002, 1008, 1013, 1017, 1020, 1022, 1023,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_h2_iscan_32x32[1024]) = {
+  0,    1,    4,    9,    15,   22,   33,   43,   56,   71,   86,   104,  121,
+  142,  166,  189,  214,  233,  254,  273,  292,  309,  328,  345,  362,  378,
+  397,  415,  431,  447,  464,  481,  2,    3,    6,    11,   17,   26,   35,
+  45,   58,   73,   90,   106,  123,  146,  168,  193,  215,  236,  255,  274,
+  294,  310,  329,  346,  363,  381,  399,  416,  432,  448,  465,  482,  5,
+  7,    8,    13,   20,   28,   37,   50,   62,   75,   92,   108,  129,  150,
+  170,  195,  216,  240,  259,  275,  295,  312,  331,  348,  365,  383,  400,
+  417,  433,  449,  467,  485,  10,   12,   14,   19,   23,   31,   41,   52,
+  65,   81,   96,   113,  133,  152,  175,  201,  221,  243,  260,  280,  297,
+  315,  333,  350,  367,  385,  402,  418,  434,  452,  470,  486,  16,   18,
+  21,   24,   30,   39,   48,   59,   69,   83,   100,  119,  137,  158,  181,
+  203,  226,  244,  264,  283,  300,  318,  335,  353,  370,  388,  404,  420,
+  438,  455,  471,  487,  25,   27,   29,   32,   40,   46,   54,   67,   79,
+  94,   109,  127,  143,  164,  185,  210,  231,  250,  269,  285,  304,  322,
+  339,  356,  373,  389,  407,  423,  440,  457,  473,  491,  34,   36,   38,
+  42,   49,   55,   64,   76,   87,   102,  117,  135,  154,  176,  197,  219,
+  239,  256,  272,  291,  308,  324,  341,  359,  377,  393,  410,  426,  442,
+  460,  476,  492,  44,   47,   51,   53,   60,   68,   77,   85,   98,   114,
+  131,  147,  162,  183,  208,  227,  245,  262,  282,  298,  314,  332,  349,
+  364,  379,  396,  412,  430,  446,  462,  478,  495,  57,   61,   63,   66,
+  70,   80,   88,   99,   112,  124,  140,  159,  179,  199,  217,  234,  253,
+  270,  286,  305,  321,  337,  354,  371,  387,  403,  419,  435,  451,  468,
+  484,  498,  72,   74,   78,   82,   84,   95,   103,  115,  125,  139,  156,
+  173,  190,  211,  229,  246,  261,  281,  296,  311,  325,  344,  360,  375,
+  392,  408,  425,  441,  456,  472,  489,  500,  89,   91,   93,   97,   101,
+  110,  118,  132,  141,  157,  171,  186,  206,  224,  241,  257,  271,  287,
+  303,  320,  336,  351,  366,  384,  398,  413,  429,  445,  461,  477,  493,
+  502,  105,  107,  111,  116,  120,  128,  136,  148,  160,  174,  187,  205,
+  222,  237,  251,  267,  284,  299,  313,  327,  343,  358,  374,  390,  405,
+  422,  437,  453,  469,  483,  497,  505,  122,  126,  130,  134,  138,  144,
+  155,  163,  180,  191,  207,  223,  232,  248,  265,  278,  293,  307,  323,
+  338,  352,  368,  382,  395,  411,  427,  443,  459,  475,  490,  501,  507,
+  145,  149,  151,  153,  161,  165,  177,  184,  200,  212,  225,  238,  249,
+  263,  276,  289,  306,  319,  334,  347,  361,  376,  391,  406,  421,  436,
+  450,  463,  479,  496,  504,  509,  167,  169,  172,  178,  182,  188,  198,
+  209,  218,  230,  242,  252,  266,  277,  288,  301,  317,  330,  342,  357,
+  372,  386,  401,  414,  428,  444,  458,  474,  488,  499,  506,  510,  192,
+  194,  196,  202,  204,  213,  220,  228,  235,  247,  258,  268,  279,  290,
+  302,  316,  326,  340,  355,  369,  380,  394,  409,  424,  439,  454,  466,
+  480,  494,  503,  508,  511,  512,  513,  514,  515,  516,  517,  520,  523,
+  526,  532,  537,  545,  551,  561,  573,  581,  596,  610,  625,  642,  661,
+  680,  701,  722,  745,  770,  800,  827,  853,  875,  897,  919,  518,  519,
+  521,  522,  524,  525,  528,  533,  536,  542,  549,  557,  564,  575,  585,
+  597,  611,  623,  640,  656,  676,  696,  717,  739,  763,  789,  815,  844,
+  867,  889,  909,  927,  527,  529,  530,  531,  534,  535,  538,  544,  548,
+  555,  560,  569,  579,  589,  598,  614,  626,  641,  655,  673,  690,  712,
+  735,  760,  780,  806,  834,  857,  880,  902,  921,  940,  539,  540,  541,
+  543,  546,  547,  550,  558,  562,  567,  576,  583,  593,  603,  616,  631,
+  643,  657,  674,  689,  710,  733,  752,  776,  803,  830,  850,  872,  892,
+  913,  934,  950,  552,  553,  554,  556,  559,  563,  565,  571,  577,  582,
+  591,  600,  609,  620,  634,  644,  662,  677,  691,  711,  730,  748,  773,
+  798,  822,  847,  869,  887,  906,  925,  942,  961,  566,  568,  570,  572,
+  574,  578,  580,  588,  594,  601,  608,  617,  629,  637,  652,  665,  681,
+  697,  713,  734,  749,  772,  793,  819,  842,  863,  884,  904,  923,  938,
+  954,  967,  584,  586,  587,  590,  592,  595,  599,  605,  613,  618,  628,
+  636,  648,  660,  671,  686,  702,  718,  736,  753,  774,  794,  818,  840,
+  860,  882,  900,  917,  936,  952,  965,  977,  602,  604,  606,  607,  612,
+  615,  619,  624,  633,  638,  649,  658,  666,  683,  692,  707,  723,  740,
+  761,  777,  799,  820,  841,  859,  877,  895,  915,  932,  948,  963,  975,
+  986,  621,  622,  627,  630,  632,  635,  639,  645,  653,  663,  668,  682,
+  688,  704,  716,  732,  746,  764,  781,  804,  823,  843,  861,  878,  894,
+  911,  930,  946,  959,  973,  984,  994,  646,  647,  650,  651,  654,  659,
+  664,  667,  678,  685,  693,  706,  715,  728,  743,  757,  771,  790,  807,
+  831,  848,  864,  883,  896,  912,  928,  944,  957,  971,  982,  992,  1001,
+  669,  670,  672,  675,  679,  684,  687,  694,  703,  709,  719,  729,  741,
+  754,  767,  783,  801,  816,  835,  851,  870,  885,  901,  916,  931,  945,
+  956,  969,  980,  990,  999,  1007, 695,  698,  699,  700,  705,  708,  714,
+  720,  726,  738,  744,  758,  768,  779,  795,  810,  828,  845,  858,  873,
+  888,  905,  918,  933,  947,  958,  970,  979,  988,  997,  1005, 1012, 721,
+  724,  725,  727,  731,  737,  742,  747,  756,  765,  775,  786,  797,  809,
+  825,  837,  854,  868,  881,  893,  907,  924,  937,  949,  960,  972,  981,
+  989,  996,  1003, 1010, 1016, 750,  751,  755,  759,  762,  766,  769,  778,
+  787,  792,  805,  812,  829,  838,  852,  865,  876,  890,  903,  914,  926,
+  939,  953,  964,  974,  983,  991,  998,  1004, 1009, 1014, 1019, 782,  784,
+  785,  788,  791,  796,  802,  808,  814,  826,  836,  846,  856,  866,  874,
+  886,  898,  910,  922,  935,  943,  955,  966,  976,  985,  993,  1000, 1006,
+  1011, 1015, 1018, 1021, 811,  813,  817,  821,  824,  832,  833,  839,  849,
+  855,  862,  871,  879,  891,  899,  908,  920,  929,  941,  951,  962,  968,
+  978,  987,  995,  1002, 1008, 1013, 1017, 1020, 1022, 1023,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_qtr_iscan_32x32[1024]) = {
+  0,    1,    4,    9,    15,   22,   33,   43,   56,   71,   86,   104,  121,
+  142,  166,  189,  256,  268,  286,  310,  334,  364,  400,  435,  471,  510,
+  553,  598,  640,  683,  732,  780,  2,    3,    6,    11,   17,   26,   35,
+  45,   58,   73,   90,   106,  123,  146,  168,  193,  258,  270,  288,  312,
+  338,  366,  402,  437,  473,  516,  557,  600,  642,  687,  736,  782,  5,
+  7,    8,    13,   20,   28,   37,   50,   62,   75,   92,   108,  129,  150,
+  170,  195,  260,  274,  292,  314,  340,  370,  406,  441,  478,  520,  559,
+  604,  646,  689,  740,  788,  10,   12,   14,   19,   23,   31,   41,   52,
+  65,   81,   96,   113,  133,  152,  175,  201,  262,  276,  294,  316,  344,
+  376,  410,  445,  484,  524,  563,  606,  648,  697,  746,  793,  16,   18,
+  21,   24,   30,   39,   48,   59,   69,   83,   100,  119,  137,  158,  181,
+  203,  264,  278,  300,  322,  350,  380,  414,  451,  490,  530,  571,  612,
+  656,  705,  750,  799,  25,   27,   29,   32,   40,   46,   54,   67,   79,
+  94,   109,  127,  143,  164,  185,  210,  266,  282,  302,  326,  354,  388,
+  422,  459,  496,  533,  579,  618,  665,  711,  754,  809,  34,   36,   38,
+  42,   49,   55,   64,   76,   87,   102,  117,  135,  154,  176,  197,  216,
+  272,  289,  308,  332,  362,  392,  427,  465,  504,  545,  585,  626,  671,
+  717,  766,  813,  44,   47,   51,   53,   60,   68,   77,   85,   98,   114,
+  131,  147,  162,  183,  208,  222,  279,  298,  320,  346,  374,  408,  442,
+  475,  511,  551,  592,  638,  681,  726,  772,  821,  57,   61,   63,   66,
+  70,   80,   88,   99,   112,  124,  140,  159,  179,  199,  214,  227,  284,
+  304,  328,  355,  386,  418,  455,  492,  528,  567,  608,  649,  695,  742,
+  786,  833,  72,   74,   78,   82,   84,   95,   103,  115,  125,  139,  156,
+  173,  190,  211,  224,  233,  296,  317,  342,  367,  394,  433,  466,  500,
+  543,  581,  622,  667,  707,  752,  803,  843,  89,   91,   93,   97,   101,
+  110,  118,  132,  141,  157,  171,  186,  206,  220,  231,  239,  306,  330,
+  352,  384,  415,  447,  482,  521,  554,  593,  636,  677,  722,  770,  815,
+  852,  105,  107,  111,  116,  120,  128,  136,  148,  160,  174,  187,  205,
+  218,  229,  237,  244,  323,  347,  371,  398,  431,  463,  498,  534,  573,
+  616,  654,  698,  743,  783,  831,  864,  122,  126,  130,  134,  138,  144,
+  155,  163,  180,  191,  207,  219,  226,  235,  242,  248,  335,  360,  390,
+  419,  449,  485,  518,  549,  587,  630,  672,  715,  760,  805,  845,  872,
+  145,  149,  151,  153,  161,  165,  177,  184,  200,  212,  221,  230,  236,
+  241,  246,  251,  356,  382,  411,  438,  469,  501,  539,  577,  613,  652,
+  690,  730,  776,  822,  858,  886,  167,  169,  172,  178,  182,  188,  198,
+  209,  215,  225,  232,  238,  243,  247,  250,  253,  378,  403,  428,  461,
+  494,  526,  560,  594,  632,  675,  713,  755,  801,  837,  868,  897,  192,
+  194,  196,  202,  204,  213,  217,  223,  228,  234,  240,  245,  249,  252,
+  254,  255,  395,  425,  457,  488,  512,  547,  583,  619,  659,  699,  737,
+  778,  819,  854,  882,  907,  257,  259,  261,  263,  265,  267,  273,  280,
+  285,  297,  307,  324,  336,  357,  379,  396,  424,  452,  479,  508,  541,
+  574,  609,  643,  679,  719,  764,  806,  841,  870,  895,  919,  269,  271,
+  275,  277,  281,  283,  290,  299,  305,  318,  331,  348,  361,  383,  404,
+  426,  453,  476,  506,  535,  568,  601,  634,  669,  708,  748,  789,  829,
+  860,  887,  909,  927,  287,  291,  293,  295,  301,  303,  309,  321,  329,
+  343,  353,  372,  391,  412,  429,  458,  480,  507,  532,  564,  590,  627,
+  663,  703,  733,  773,  816,  847,  876,  901,  921,  940,  311,  313,  315,
+  319,  325,  327,  333,  349,  358,  368,  385,  399,  420,  439,  462,  489,
+  509,  536,  565,  589,  624,  661,  691,  727,  768,  810,  838,  866,  890,
+  913,  934,  950,  337,  339,  341,  345,  351,  359,  363,  375,  387,  397,
+  416,  432,  450,  470,  495,  513,  542,  569,  591,  625,  657,  684,  723,
+  762,  797,  834,  862,  884,  905,  925,  942,  961,  365,  369,  373,  377,
+  381,  389,  393,  409,  421,  434,  448,  464,  486,  502,  527,  548,  575,
+  602,  628,  662,  685,  721,  756,  794,  827,  855,  880,  903,  923,  938,
+  954,  967,  401,  405,  407,  413,  417,  423,  430,  443,  456,  467,  483,
+  499,  519,  540,  561,  584,  610,  635,  664,  692,  724,  757,  792,  825,
+  850,  878,  899,  917,  936,  952,  965,  977,  436,  440,  444,  446,  454,
+  460,  468,  477,  493,  503,  522,  537,  550,  578,  595,  620,  644,  670,
+  704,  728,  763,  795,  826,  849,  873,  893,  915,  932,  948,  963,  975,
+  986,  472,  474,  481,  487,  491,  497,  505,  514,  529,  544,  555,  576,
+  588,  614,  633,  660,  680,  709,  734,  769,  798,  828,  851,  874,  892,
+  911,  930,  946,  959,  973,  984,  994,  515,  517,  523,  525,  531,  538,
+  546,  552,  570,  582,  596,  617,  631,  653,  676,  700,  720,  749,  774,
+  811,  835,  856,  879,  894,  912,  928,  944,  957,  971,  982,  992,  1001,
+  556,  558,  562,  566,  572,  580,  586,  597,  611,  623,  637,  655,  673,
+  693,  714,  738,  765,  790,  817,  839,  863,  881,  900,  916,  931,  945,
+  956,  969,  980,  990,  999,  1007, 599,  603,  605,  607,  615,  621,  629,
+  639,  650,  668,  678,  701,  716,  731,  758,  779,  807,  830,  848,  867,
+  885,  904,  918,  933,  947,  958,  970,  979,  988,  997,  1005, 1012, 641,
+  645,  647,  651,  658,  666,  674,  682,  696,  710,  725,  744,  761,  777,
+  802,  820,  842,  861,  877,  891,  906,  924,  937,  949,  960,  972,  981,
+  989,  996,  1003, 1010, 1016, 686,  688,  694,  702,  706,  712,  718,  729,
+  745,  753,  771,  784,  808,  823,  840,  857,  871,  888,  902,  914,  926,
+  939,  953,  964,  974,  983,  991,  998,  1004, 1009, 1014, 1019, 735,  739,
+  741,  747,  751,  759,  767,  775,  787,  804,  818,  832,  846,  859,  869,
+  883,  896,  910,  922,  935,  943,  955,  966,  976,  985,  993,  1000, 1006,
+  1011, 1015, 1018, 1021, 781,  785,  791,  796,  800,  812,  814,  824,  836,
+  844,  853,  865,  875,  889,  898,  908,  920,  929,  941,  951,  962,  968,
+  978,  987,  995,  1002, 1008, 1013, 1017, 1020, 1022, 1023,
+};
+
+#if CONFIG_TX64X64
+DECLARE_ALIGNED(16, static const int16_t, av1_default_iscan_64x64[4096]) = {
+  0,    1,    4,    9,    15,   22,   33,   43,   56,   71,   86,   104,  121,
+  142,  166,  189,  214,  239,  269,  300,  331,  363,  400,  435,  471,  510,
+  553,  598,  640,  683,  732,  780,  833,  884,  937,  995,  1048, 1107, 1165,
+  1230, 1293, 1353, 1422, 1489, 1562, 1632, 1701, 1776, 1850, 1929, 2006, 2091,
+  2173, 2252, 2339, 2421, 2516, 2603, 2694, 2786, 2879, 2978, 3076, 3175, 2,
+  3,    6,    11,   17,   26,   35,   45,   58,   73,   90,   106,  123,  146,
+  168,  193,  216,  243,  271,  302,  335,  365,  402,  437,  473,  516,  557,
+  600,  642,  687,  736,  782,  835,  886,  941,  999,  1050, 1111, 1167, 1234,
+  1297, 1357, 1424, 1491, 1564, 1636, 1703, 1778, 1852, 1931, 2012, 2095, 2177,
+  2256, 2341, 2425, 2518, 2605, 2698, 2788, 2883, 2982, 3078, 3177, 5,    7,
+  8,    13,   20,   28,   37,   50,   62,   75,   92,   108,  129,  150,  170,
+  195,  218,  249,  277,  304,  337,  369,  406,  441,  478,  520,  559,  604,
+  646,  689,  740,  788,  841,  890,  945,  1001, 1052, 1115, 1173, 1236, 1301,
+  1362, 1428, 1497, 1568, 1638, 1707, 1786, 1858, 1935, 2016, 2097, 2181, 2260,
+  2343, 2431, 2520, 2613, 2702, 2790, 2889, 2984, 3082, 3181, 10,   12,   14,
+  19,   23,   31,   41,   52,   65,   81,   96,   113,  133,  152,  175,  201,
+  224,  253,  279,  310,  341,  375,  410,  445,  484,  524,  563,  606,  648,
+  697,  746,  793,  843,  896,  949,  1005, 1060, 1119, 1181, 1242, 1303, 1366,
+  1436, 1503, 1572, 1640, 1713, 1790, 1865, 1943, 2018, 2103, 2183, 2266, 2347,
+  2437, 2526, 2617, 2708, 2800, 2893, 2992, 3086, 3189, 16,   18,   21,   24,
+  30,   39,   48,   59,   69,   83,   100,  119,  137,  158,  181,  203,  230,
+  255,  286,  316,  347,  380,  414,  451,  490,  530,  571,  612,  656,  705,
+  750,  799,  849,  898,  959,  1009, 1066, 1127, 1184, 1246, 1307, 1376, 1440,
+  1509, 1578, 1644, 1723, 1794, 1871, 1947, 2024, 2109, 2185, 2270, 2361, 2443,
+  2536, 2619, 2710, 2806, 2899, 2998, 3090, 3193, 25,   27,   29,   32,   40,
+  46,   54,   67,   79,   94,   109,  127,  143,  164,  185,  210,  236,  263,
+  292,  320,  353,  388,  422,  459,  496,  533,  579,  618,  665,  711,  754,
+  809,  857,  910,  961,  1015, 1074, 1131, 1194, 1254, 1315, 1384, 1448, 1517,
+  1584, 1655, 1731, 1802, 1875, 1959, 2034, 2115, 2197, 2280, 2367, 2452, 2538,
+  2625, 2722, 2816, 2907, 3004, 3100, 3203, 34,   36,   38,   42,   49,   55,
+  64,   76,   87,   102,  117,  135,  154,  176,  197,  222,  247,  272,  298,
+  329,  361,  392,  427,  465,  504,  545,  585,  626,  671,  717,  766,  813,
+  862,  916,  971,  1028, 1084, 1139, 1200, 1264, 1325, 1390, 1452, 1523, 1594,
+  1667, 1737, 1806, 1887, 1963, 2046, 2123, 2202, 2290, 2371, 2462, 2548, 2641,
+  2732, 2822, 2917, 3010, 3111, 3211, 44,   47,   51,   53,   60,   68,   77,
+  85,   98,   114,  131,  147,  162,  183,  208,  232,  256,  283,  314,  343,
+  373,  408,  442,  475,  511,  551,  592,  638,  681,  726,  772,  821,  874,
+  926,  979,  1034, 1088, 1153, 1214, 1271, 1335, 1396, 1469, 1533, 1600, 1673,
+  1745, 1824, 1897, 1973, 2054, 2131, 2216, 2300, 2383, 2468, 2558, 2649, 2740,
+  2829, 2923, 3022, 3123, 3221, 57,   61,   63,   66,   70,   80,   88,   99,
+  112,  124,  140,  159,  179,  199,  219,  240,  267,  294,  322,  354,  386,
+  418,  455,  492,  528,  567,  608,  649,  695,  742,  786,  836,  882,  933,
+  989,  1046, 1101, 1161, 1216, 1279, 1343, 1410, 1479, 1543, 1614, 1687, 1758,
+  1832, 1905, 1980, 2066, 2141, 2226, 2306, 2395, 2484, 2566, 2659, 2750, 2845,
+  2939, 3032, 3133, 3225, 72,   74,   78,   82,   84,   95,   103,  115,  125,
+  139,  156,  173,  190,  211,  234,  259,  281,  311,  339,  366,  394,  433,
+  466,  500,  543,  581,  622,  667,  707,  752,  803,  853,  899,  955,  1007,
+  1064, 1117, 1175, 1237, 1299, 1354, 1420, 1485, 1556, 1624, 1697, 1770, 1842,
+  1919, 1998, 2074, 2155, 2234, 2319, 2409, 2492, 2581, 2671, 2760, 2859, 2949,
+  3046, 3145, 3245, 89,   91,   93,   97,   101,  110,  118,  132,  141,  157,
+  171,  186,  206,  228,  251,  273,  296,  324,  351,  384,  415,  447,  482,
+  521,  554,  593,  636,  677,  722,  770,  815,  866,  914,  967,  1022, 1078,
+  1135, 1195, 1252, 1313, 1378, 1444, 1507, 1576, 1642, 1714, 1788, 1860, 1933,
+  2013, 2085, 2169, 2250, 2337, 2417, 2502, 2597, 2683, 2778, 2869, 2960, 3060,
+  3157, 3256, 105,  107,  111,  116,  120,  128,  136,  148,  160,  174,  187,
+  205,  225,  244,  265,  290,  317,  344,  370,  398,  431,  463,  498,  534,
+  573,  616,  654,  698,  743,  783,  831,  880,  928,  983,  1036, 1092, 1149,
+  1208, 1266, 1333, 1394, 1457, 1524, 1590, 1665, 1733, 1804, 1879, 1953, 2030,
+  2111, 2189, 2271, 2357, 2441, 2534, 2615, 2704, 2791, 2887, 2979, 3072, 3167,
+  3270, 122,  126,  130,  134,  138,  144,  155,  163,  180,  191,  207,  226,
+  238,  261,  287,  308,  332,  359,  390,  419,  449,  485,  518,  549,  587,
+  630,  672,  715,  760,  805,  855,  900,  953,  1003, 1053, 1108, 1163, 1220,
+  1287, 1345, 1408, 1473, 1541, 1608, 1677, 1749, 1826, 1898, 1971, 2048, 2127,
+  2208, 2294, 2373, 2458, 2542, 2631, 2726, 2818, 2908, 3002, 3094, 3199, 3286,
+  145,  149,  151,  153,  161,  165,  177,  184,  200,  212,  229,  245,  262,
+  284,  305,  327,  355,  382,  411,  438,  469,  501,  539,  577,  613,  652,
+  690,  730,  776,  822,  872,  922,  973,  1024, 1079, 1132, 1188, 1250, 1305,
+  1367, 1432, 1492, 1560, 1626, 1693, 1766, 1838, 1911, 1992, 2068, 2149, 2228,
+  2307, 2393, 2478, 2564, 2655, 2742, 2833, 2927, 3020, 3119, 3219, 3298, 167,
+  169,  172,  178,  182,  188,  198,  209,  220,  235,  252,  266,  288,  306,
+  326,  349,  378,  403,  428,  461,  494,  526,  560,  594,  632,  675,  713,
+  755,  801,  845,  892,  942,  990,  1042, 1096, 1155, 1212, 1267, 1329, 1391,
+  1450, 1519, 1582, 1650, 1724, 1792, 1862, 1936, 2007, 2083, 2167, 2246, 2329,
+  2413, 2496, 2585, 2675, 2761, 2855, 2947, 3040, 3135, 3233, 3320, 192,  194,
+  196,  202,  204,  213,  223,  233,  241,  260,  274,  291,  309,  328,  350,
+  376,  395,  425,  457,  488,  512,  547,  583,  619,  659,  699,  737,  778,
+  819,  868,  917,  965,  1013, 1072, 1123, 1176, 1231, 1289, 1351, 1414, 1474,
+  1539, 1604, 1674, 1741, 1816, 1891, 1961, 2040, 2116, 2191, 2276, 2353, 2438,
+  2524, 2606, 2689, 2784, 2871, 2968, 3062, 3161, 3257, 3334, 215,  217,  221,
+  227,  231,  237,  248,  257,  268,  282,  297,  318,  333,  356,  379,  396,
+  424,  452,  479,  508,  541,  574,  609,  643,  679,  719,  764,  806,  850,
+  894,  938,  987,  1038, 1089, 1145, 1204, 1258, 1316, 1379, 1438, 1501, 1565,
+  1628, 1694, 1764, 1836, 1907, 1981, 2060, 2137, 2220, 2298, 2377, 2464, 2549,
+  2635, 2724, 2812, 2903, 2999, 3088, 3185, 3278, 3350, 242,  246,  250,  254,
+  258,  264,  275,  285,  295,  312,  325,  345,  360,  383,  404,  426,  453,
+  476,  506,  535,  568,  601,  634,  669,  708,  748,  789,  829,  875,  923,
+  968,  1016, 1068, 1120, 1168, 1224, 1280, 1341, 1402, 1465, 1531, 1591, 1661,
+  1729, 1795, 1867, 1937, 2004, 2079, 2159, 2242, 2320, 2405, 2488, 2573, 2661,
+  2744, 2839, 2933, 3023, 3117, 3215, 3296, 3373, 270,  276,  278,  280,  289,
+  293,  299,  315,  323,  340,  352,  371,  391,  412,  429,  458,  480,  507,
+  532,  564,  590,  627,  663,  703,  733,  773,  816,  859,  906,  950,  993,
+  1043, 1094, 1147, 1201, 1256, 1311, 1372, 1429, 1486, 1550, 1618, 1685, 1751,
+  1827, 1895, 1965, 2042, 2119, 2192, 2268, 2348, 2429, 2512, 2599, 2684, 2772,
+  2863, 2951, 3048, 3143, 3239, 3324, 3393, 301,  303,  307,  313,  319,  321,
+  330,  346,  357,  367,  385,  399,  420,  439,  462,  489,  509,  536,  565,
+  589,  624,  661,  691,  727,  768,  810,  846,  887,  929,  977,  1029, 1076,
+  1128, 1177, 1226, 1283, 1339, 1397, 1461, 1521, 1585, 1648, 1715, 1779, 1848,
+  1923, 1996, 2069, 2142, 2224, 2302, 2381, 2465, 2544, 2627, 2720, 2807, 2895,
+  2985, 3073, 3163, 3264, 3338, 3413, 334,  336,  338,  342,  348,  358,  362,
+  374,  387,  397,  416,  432,  450,  470,  495,  513,  542,  569,  591,  625,
+  657,  684,  723,  762,  797,  837,  878,  920,  963,  1010, 1054, 1105, 1157,
+  1206, 1262, 1317, 1374, 1433, 1483, 1545, 1615, 1681, 1743, 1812, 1885, 1954,
+  2025, 2101, 2174, 2248, 2330, 2411, 2490, 2579, 2663, 2745, 2835, 2924, 3018,
+  3115, 3205, 3290, 3363, 3431, 364,  368,  372,  377,  381,  389,  393,  409,
+  421,  434,  448,  464,  486,  502,  527,  548,  575,  602,  628,  662,  685,
+  721,  756,  794,  827,  869,  912,  956,  996,  1040, 1086, 1137, 1189, 1243,
+  1291, 1349, 1404, 1466, 1525, 1588, 1645, 1711, 1774, 1843, 1909, 1988, 2058,
+  2132, 2209, 2288, 2368, 2445, 2527, 2607, 2687, 2780, 2865, 2953, 3049, 3139,
+  3237, 3318, 3387, 3451, 401,  405,  407,  413,  417,  423,  430,  443,  456,
+  467,  483,  499,  519,  540,  561,  584,  610,  635,  664,  692,  724,  757,
+  792,  825,  863,  908,  946,  985,  1032, 1080, 1125, 1169, 1217, 1275, 1330,
+  1386, 1441, 1498, 1554, 1619, 1683, 1746, 1810, 1883, 1949, 2019, 2086, 2165,
+  2238, 2314, 2399, 2479, 2562, 2645, 2733, 2820, 2904, 2996, 3083, 3168, 3268,
+  3339, 3407, 3474, 436,  440,  444,  446,  454,  460,  468,  477,  493,  503,
+  522,  537,  550,  578,  595,  620,  644,  670,  704,  728,  763,  795,  826,
+  861,  901,  935,  980,  1025, 1069, 1112, 1159, 1209, 1260, 1309, 1363, 1418,
+  1475, 1534, 1598, 1656, 1721, 1780, 1846, 1912, 1982, 2056, 2129, 2199, 2278,
+  2358, 2432, 2508, 2593, 2677, 2762, 2851, 2941, 3030, 3124, 3216, 3294, 3365,
+  3433, 3488, 472,  474,  481,  487,  491,  497,  505,  514,  529,  544,  555,
+  576,  588,  614,  633,  660,  680,  709,  734,  769,  798,  828,  864,  902,
+  932,  975,  1020, 1061, 1102, 1150, 1198, 1247, 1294, 1346, 1400, 1455, 1513,
+  1573, 1629, 1689, 1755, 1820, 1888, 1955, 2022, 2092, 2163, 2235, 2312, 2389,
+  2472, 2554, 2632, 2716, 2804, 2884, 2974, 3063, 3153, 3250, 3326, 3395, 3454,
+  3512, 515,  517,  523,  525,  531,  538,  546,  552,  570,  582,  596,  617,
+  631,  653,  676,  700,  720,  749,  774,  811,  838,  870,  909,  936,  976,
+  1017, 1058, 1099, 1143, 1192, 1238, 1284, 1336, 1388, 1445, 1493, 1546, 1610,
+  1671, 1734, 1796, 1856, 1925, 1994, 2062, 2133, 2206, 2281, 2354, 2426, 2503,
+  2587, 2669, 2754, 2843, 2928, 3016, 3105, 3201, 3284, 3351, 3421, 3480, 3534,
+  556,  558,  562,  566,  572,  580,  586,  597,  611,  623,  637,  655,  673,
+  693,  714,  738,  765,  790,  817,  847,  879,  913,  947,  981,  1021, 1059,
+  1097, 1140, 1185, 1227, 1277, 1327, 1380, 1425, 1481, 1537, 1595, 1651, 1708,
+  1771, 1834, 1901, 1966, 2035, 2107, 2170, 2244, 2315, 2396, 2474, 2552, 2628,
+  2711, 2792, 2875, 2966, 3056, 3146, 3234, 3314, 3383, 3445, 3504, 3559, 599,
+  603,  605,  607,  615,  621,  629,  639,  650,  668,  678,  701,  716,  731,
+  758,  779,  807,  830,  860,  888,  921,  957,  986,  1026, 1062, 1100, 1141,
+  1183, 1221, 1272, 1323, 1368, 1416, 1471, 1526, 1580, 1633, 1691, 1752, 1817,
+  1876, 1944, 2002, 2072, 2143, 2218, 2291, 2363, 2435, 2509, 2589, 2672, 2752,
+  2840, 2921, 3008, 3095, 3190, 3274, 3344, 3409, 3470, 3526, 3577, 641,  645,
+  647,  651,  658,  666,  674,  682,  696,  710,  725,  744,  761,  777,  802,
+  820,  851,  876,  907,  930,  964,  997,  1033, 1070, 1103, 1144, 1186, 1222,
+  1270, 1318, 1360, 1411, 1463, 1515, 1569, 1622, 1678, 1739, 1800, 1853, 1917,
+  1983, 2052, 2121, 2186, 2253, 2331, 2406, 2482, 2559, 2639, 2717, 2798, 2877,
+  2961, 3052, 3137, 3226, 3306, 3379, 3437, 3492, 3553, 3601, 686,  688,  694,
+  702,  706,  712,  718,  729,  745,  753,  771,  784,  808,  823,  848,  871,
+  895,  924,  951,  978,  1011, 1041, 1081, 1113, 1151, 1193, 1228, 1273, 1319,
+  1358, 1406, 1458, 1510, 1557, 1612, 1669, 1727, 1781, 1839, 1903, 1969, 2031,
+  2098, 2160, 2232, 2304, 2375, 2453, 2528, 2601, 2679, 2758, 2846, 2929, 3011,
+  3098, 3186, 3271, 3340, 3401, 3466, 3522, 3571, 3620, 735,  739,  741,  747,
+  751,  759,  767,  775,  787,  804,  818,  832,  856,  873,  893,  918,  939,
+  969,  994,  1030, 1055, 1087, 1126, 1160, 1199, 1239, 1278, 1324, 1361, 1407,
+  1453, 1505, 1551, 1605, 1663, 1716, 1768, 1830, 1893, 1951, 2008, 2075, 2139,
+  2214, 2284, 2349, 2418, 2494, 2571, 2653, 2734, 2810, 2890, 2972, 3058, 3147,
+  3231, 3310, 3375, 3435, 3490, 3545, 3595, 3642, 781,  785,  791,  796,  800,
+  812,  814,  824,  839,  854,  867,  881,  903,  925,  943,  966,  988,  1018,
+  1044, 1077, 1106, 1138, 1170, 1210, 1248, 1285, 1328, 1369, 1412, 1459, 1506,
+  1549, 1601, 1657, 1704, 1762, 1821, 1880, 1938, 1999, 2063, 2125, 2193, 2257,
+  2327, 2401, 2475, 2545, 2620, 2691, 2776, 2860, 2942, 3024, 3109, 3197, 3276,
+  3345, 3403, 3468, 3520, 3569, 3616, 3664, 834,  840,  842,  844,  852,  858,
+  865,  877,  883,  904,  915,  931,  954,  974,  991,  1014, 1039, 1071, 1095,
+  1129, 1158, 1190, 1218, 1261, 1295, 1337, 1381, 1417, 1464, 1511, 1552, 1602,
+  1654, 1699, 1759, 1813, 1872, 1927, 1990, 2049, 2113, 2178, 2239, 2308, 2378,
+  2450, 2521, 2594, 2667, 2746, 2824, 2909, 2990, 3070, 3154, 3243, 3316, 3381,
+  3441, 3493, 3547, 3597, 3640, 3682, 885,  889,  891,  897,  905,  911,  919,
+  927,  934,  958,  970,  984,  1004, 1027, 1045, 1073, 1090, 1121, 1148, 1178,
+  1207, 1244, 1276, 1310, 1347, 1389, 1426, 1472, 1516, 1558, 1606, 1658, 1700,
+  1757, 1807, 1868, 1920, 1978, 2043, 2104, 2157, 2229, 2296, 2364, 2422, 2498,
+  2574, 2650, 2727, 2801, 2872, 2954, 3038, 3129, 3212, 3288, 3352, 3419, 3475,
+  3524, 3573, 3621, 3668, 3707, 940,  944,  948,  952,  960,  962,  972,  982,
+  992,  1008, 1023, 1037, 1056, 1082, 1098, 1124, 1146, 1171, 1202, 1229, 1263,
+  1292, 1331, 1364, 1401, 1446, 1482, 1527, 1570, 1613, 1664, 1705, 1760, 1808,
+  1863, 1915, 1976, 2036, 2087, 2153, 2221, 2286, 2344, 2414, 2486, 2556, 2623,
+  2699, 2773, 2853, 2937, 3012, 3091, 3169, 3260, 3330, 3391, 3447, 3505, 3555,
+  3603, 3646, 3684, 3727, 998,  1000, 1002, 1006, 1012, 1019, 1031, 1035, 1047,
+  1065, 1083, 1093, 1109, 1133, 1156, 1179, 1205, 1225, 1257, 1286, 1320, 1350,
+  1387, 1419, 1456, 1494, 1538, 1581, 1623, 1670, 1717, 1763, 1814, 1869, 1916,
+  1974, 2028, 2081, 2150, 2212, 2272, 2335, 2403, 2469, 2539, 2608, 2680, 2755,
+  2827, 2915, 2986, 3068, 3151, 3229, 3300, 3366, 3427, 3484, 3532, 3581, 3630,
+  3672, 3709, 3745, 1049, 1051, 1057, 1063, 1067, 1075, 1085, 1091, 1104, 1118,
+  1136, 1152, 1164, 1191, 1213, 1232, 1259, 1281, 1312, 1340, 1375, 1405, 1442,
+  1476, 1514, 1547, 1596, 1634, 1679, 1728, 1769, 1822, 1873, 1921, 1977, 2029,
+  2078, 2144, 2203, 2264, 2325, 2390, 2459, 2529, 2591, 2665, 2738, 2813, 2880,
+  2957, 3041, 3127, 3206, 3282, 3348, 3399, 3460, 3513, 3565, 3609, 3650, 3695,
+  3733, 3768, 1110, 1114, 1116, 1122, 1130, 1134, 1142, 1154, 1162, 1180, 1196,
+  1211, 1223, 1251, 1268, 1290, 1321, 1342, 1373, 1398, 1434, 1467, 1499, 1535,
+  1574, 1611, 1652, 1692, 1740, 1782, 1831, 1881, 1928, 1979, 2037, 2082, 2145,
+  2200, 2261, 2321, 2387, 2454, 2513, 2583, 2656, 2730, 2793, 2867, 2945, 3025,
+  3101, 3178, 3262, 3328, 3388, 3443, 3494, 3543, 3591, 3636, 3678, 3715, 3754,
+  3790, 1166, 1172, 1174, 1182, 1187, 1197, 1203, 1215, 1219, 1240, 1253, 1269,
+  1288, 1306, 1332, 1352, 1382, 1403, 1430, 1462, 1484, 1528, 1555, 1599, 1630,
+  1672, 1709, 1753, 1801, 1840, 1894, 1939, 1991, 2044, 2088, 2151, 2204, 2262,
+  2318, 2384, 2448, 2504, 2577, 2646, 2712, 2782, 2856, 2934, 3006, 3079, 3158,
+  3240, 3307, 3371, 3425, 3481, 3530, 3575, 3618, 3660, 3701, 3741, 3774, 3807,
+  1233, 1235, 1241, 1245, 1249, 1255, 1265, 1274, 1282, 1300, 1314, 1334, 1348,
+  1370, 1392, 1415, 1439, 1468, 1487, 1522, 1548, 1589, 1620, 1659, 1690, 1735,
+  1772, 1818, 1854, 1904, 1952, 2000, 2050, 2105, 2154, 2213, 2265, 2322, 2385,
+  2446, 2500, 2569, 2642, 2705, 2770, 2849, 2919, 2993, 3064, 3140, 3223, 3292,
+  3353, 3414, 3464, 3516, 3561, 3607, 3648, 3687, 3725, 3762, 3796, 3827, 1296,
+  1298, 1302, 1304, 1308, 1322, 1326, 1338, 1344, 1355, 1383, 1395, 1409, 1435,
+  1451, 1477, 1502, 1532, 1553, 1586, 1616, 1646, 1684, 1722, 1756, 1797, 1835,
+  1877, 1918, 1970, 2009, 2064, 2114, 2158, 2222, 2273, 2326, 2388, 2449, 2501,
+  2567, 2636, 2695, 2768, 2836, 2910, 2976, 3053, 3131, 3209, 3279, 3336, 3397,
+  3449, 3500, 3549, 3593, 3634, 3676, 3713, 3747, 3784, 3817, 3845, 1356, 1359,
+  1365, 1371, 1377, 1385, 1393, 1399, 1413, 1421, 1447, 1460, 1478, 1495, 1520,
+  1540, 1566, 1592, 1621, 1649, 1682, 1712, 1747, 1783, 1823, 1857, 1902, 1945,
+  1984, 2032, 2076, 2126, 2179, 2230, 2287, 2336, 2391, 2455, 2505, 2570, 2637,
+  2692, 2763, 2830, 2901, 2969, 3044, 3120, 3194, 3265, 3331, 3385, 3439, 3486,
+  3536, 3582, 3626, 3665, 3703, 3739, 3772, 3802, 3835, 3864, 1423, 1427, 1431,
+  1437, 1443, 1449, 1454, 1470, 1480, 1488, 1508, 1529, 1542, 1561, 1583, 1607,
+  1631, 1662, 1686, 1718, 1744, 1775, 1811, 1847, 1889, 1926, 1967, 2003, 2053,
+  2099, 2140, 2194, 2240, 2297, 2345, 2404, 2460, 2514, 2578, 2643, 2696, 2764,
+  2826, 2897, 2962, 3036, 3112, 3182, 3254, 3321, 3376, 3429, 3478, 3527, 3567,
+  3611, 3652, 3693, 3731, 3764, 3794, 3825, 3853, 3882, 1490, 1496, 1500, 1504,
+  1512, 1518, 1530, 1536, 1544, 1559, 1577, 1593, 1609, 1627, 1653, 1675, 1695,
+  1730, 1754, 1784, 1815, 1844, 1884, 1913, 1956, 1995, 2038, 2073, 2122, 2161,
+  2215, 2258, 2309, 2365, 2415, 2470, 2530, 2584, 2647, 2706, 2769, 2831, 2898,
+  2959, 3033, 3106, 3170, 3252, 3312, 3367, 3423, 3471, 3518, 3563, 3605, 3644,
+  3680, 3717, 3755, 3788, 3819, 3847, 3874, 3898, 1563, 1567, 1571, 1575, 1579,
+  1587, 1597, 1603, 1617, 1625, 1643, 1666, 1680, 1696, 1725, 1742, 1765, 1798,
+  1828, 1849, 1886, 1910, 1950, 1985, 2023, 2065, 2108, 2146, 2187, 2233, 2285,
+  2328, 2379, 2423, 2487, 2540, 2592, 2657, 2713, 2771, 2837, 2902, 2963, 3034,
+  3104, 3164, 3248, 3304, 3361, 3417, 3462, 3510, 3557, 3598, 3638, 3674, 3711,
+  3743, 3776, 3811, 3839, 3868, 3892, 3917, 1635, 1637, 1639, 1641, 1647, 1660,
+  1668, 1676, 1688, 1698, 1719, 1736, 1750, 1767, 1793, 1819, 1837, 1870, 1896,
+  1924, 1957, 1989, 2020, 2057, 2093, 2134, 2171, 2219, 2254, 2305, 2350, 2402,
+  2451, 2499, 2557, 2609, 2666, 2731, 2783, 2850, 2911, 2970, 3037, 3107, 3165,
+  3246, 3301, 3359, 3410, 3458, 3508, 3551, 3589, 3632, 3670, 3705, 3737, 3770,
+  3800, 3829, 3858, 3886, 3911, 3933, 1702, 1706, 1710, 1720, 1726, 1732, 1738,
+  1748, 1761, 1773, 1789, 1805, 1829, 1841, 1864, 1892, 1908, 1940, 1968, 1997,
+  2026, 2059, 2089, 2130, 2164, 2207, 2245, 2292, 2332, 2376, 2419, 2476, 2522,
+  2575, 2624, 2681, 2739, 2794, 2857, 2920, 2977, 3045, 3113, 3171, 3249, 3302,
+  3358, 3404, 3455, 3502, 3541, 3587, 3628, 3661, 3699, 3735, 3766, 3797, 3823,
+  3851, 3876, 3903, 3927, 3950, 1777, 1785, 1787, 1791, 1799, 1803, 1809, 1825,
+  1833, 1845, 1861, 1882, 1899, 1914, 1941, 1962, 1986, 2005, 2045, 2070, 2102,
+  2135, 2166, 2201, 2236, 2282, 2316, 2366, 2407, 2456, 2495, 2546, 2595, 2651,
+  2700, 2756, 2814, 2868, 2935, 2994, 3054, 3121, 3183, 3253, 3305, 3360, 3405,
+  3453, 3498, 3539, 3585, 3622, 3658, 3697, 3728, 3760, 3792, 3821, 3849, 3872,
+  3896, 3919, 3942, 3964, 1851, 1855, 1859, 1866, 1874, 1878, 1890, 1900, 1906,
+  1922, 1934, 1958, 1972, 1993, 2010, 2041, 2061, 2080, 2120, 2147, 2175, 2210,
+  2241, 2279, 2313, 2355, 2397, 2436, 2483, 2531, 2572, 2621, 2668, 2728, 2774,
+  2828, 2881, 2946, 3007, 3065, 3132, 3195, 3255, 3313, 3362, 3411, 3456, 3499,
+  3538, 3579, 3614, 3656, 3691, 3723, 3758, 3786, 3815, 3843, 3870, 3894, 3915,
+  3937, 3956, 3975, 1930, 1932, 1942, 1946, 1948, 1960, 1964, 1975, 1987, 2001,
+  2014, 2033, 2051, 2071, 2084, 2117, 2138, 2162, 2195, 2225, 2249, 2289, 2317,
+  2359, 2392, 2427, 2477, 2510, 2560, 2602, 2654, 2693, 2747, 2802, 2854, 2916,
+  2958, 3026, 3080, 3141, 3210, 3266, 3322, 3368, 3418, 3459, 3503, 3540, 3580,
+  3613, 3654, 3688, 3721, 3752, 3782, 3813, 3841, 3865, 3890, 3913, 3935, 3954,
+  3972, 3989, 2011, 2015, 2017, 2021, 2027, 2039, 2047, 2055, 2067, 2077, 2090,
+  2112, 2128, 2152, 2168, 2196, 2223, 2243, 2269, 2303, 2333, 2369, 2400, 2433,
+  2473, 2506, 2553, 2590, 2640, 2682, 2735, 2777, 2825, 2873, 2938, 2987, 3042,
+  3102, 3159, 3224, 3280, 3332, 3377, 3424, 3463, 3509, 3542, 3586, 3615, 3655,
+  3685, 3719, 3750, 3780, 3809, 3836, 3862, 3888, 3909, 3931, 3952, 3970, 3987,
+  4003, 2094, 2096, 2100, 2106, 2110, 2118, 2124, 2136, 2148, 2156, 2172, 2190,
+  2211, 2231, 2247, 2277, 2299, 2323, 2351, 2382, 2412, 2447, 2480, 2511, 2555,
+  2588, 2629, 2673, 2718, 2759, 2811, 2861, 2912, 2955, 3013, 3069, 3128, 3179,
+  3241, 3293, 3337, 3386, 3430, 3472, 3511, 3552, 3588, 3623, 3657, 3689, 3720,
+  3749, 3778, 3805, 3833, 3860, 3884, 3907, 3929, 3948, 3968, 3985, 4001, 4016,
+  2176, 2180, 2182, 2184, 2188, 2198, 2205, 2217, 2227, 2237, 2251, 2274, 2295,
+  2310, 2334, 2356, 2380, 2408, 2430, 2466, 2491, 2532, 2563, 2596, 2633, 2670,
+  2714, 2753, 2799, 2847, 2891, 2943, 2991, 3039, 3092, 3152, 3207, 3263, 3308,
+  3354, 3398, 3440, 3479, 3519, 3558, 3590, 3629, 3659, 3692, 3722, 3751, 3779,
+  3804, 3831, 3856, 3880, 3905, 3925, 3946, 3966, 3983, 3999, 4014, 4028, 2255,
+  2259, 2263, 2267, 2275, 2283, 2293, 2301, 2311, 2324, 2338, 2360, 2374, 2394,
+  2416, 2439, 2467, 2489, 2515, 2547, 2580, 2610, 2648, 2678, 2719, 2757, 2795,
+  2841, 2878, 2930, 2973, 3027, 3071, 3130, 3172, 3230, 3283, 3329, 3372, 3415,
+  3450, 3487, 3528, 3564, 3599, 3633, 3662, 3698, 3724, 3753, 3781, 3806, 3832,
+  3855, 3878, 3901, 3923, 3944, 3962, 3981, 3997, 4012, 4026, 4039, 2340, 2342,
+  2346, 2352, 2362, 2370, 2372, 2386, 2398, 2410, 2420, 2442, 2461, 2481, 2497,
+  2525, 2550, 2576, 2600, 2630, 2664, 2688, 2736, 2765, 2805, 2844, 2876, 2922,
+  2964, 3014, 3059, 3110, 3155, 3213, 3261, 3303, 3349, 3389, 3426, 3465, 3501,
+  3537, 3568, 3606, 3639, 3671, 3700, 3729, 3759, 3783, 3810, 3834, 3857, 3879,
+  3900, 3921, 3940, 3960, 3979, 3995, 4010, 4024, 4037, 4049, 2424, 2428, 2434,
+  2440, 2444, 2457, 2463, 2471, 2485, 2493, 2507, 2535, 2543, 2565, 2586, 2611,
+  2638, 2662, 2685, 2721, 2748, 2781, 2821, 2852, 2885, 2931, 2967, 3009, 3055,
+  3099, 3148, 3198, 3244, 3289, 3333, 3369, 3400, 3444, 3482, 3517, 3550, 3583,
+  3612, 3645, 3675, 3706, 3736, 3761, 3787, 3814, 3837, 3861, 3881, 3902, 3922,
+  3939, 3958, 3977, 3993, 4008, 4022, 4035, 4047, 4058, 2517, 2519, 2523, 2533,
+  2537, 2541, 2551, 2561, 2568, 2582, 2598, 2616, 2634, 2658, 2676, 2690, 2725,
+  2749, 2775, 2808, 2838, 2866, 2905, 2944, 2975, 3017, 3057, 3096, 3138, 3187,
+  3232, 3277, 3317, 3355, 3392, 3428, 3461, 3495, 3531, 3562, 3594, 3627, 3653,
+  3681, 3712, 3738, 3767, 3793, 3816, 3842, 3863, 3885, 3906, 3924, 3941, 3959,
+  3974, 3991, 4006, 4020, 4033, 4045, 4056, 4066, 2604, 2612, 2614, 2618, 2622,
+  2626, 2644, 2652, 2660, 2674, 2686, 2707, 2729, 2743, 2766, 2785, 2815, 2842,
+  2864, 2896, 2925, 2956, 2997, 3031, 3066, 3108, 3149, 3191, 3227, 3272, 3311,
+  3346, 3382, 3420, 3448, 3485, 3514, 3544, 3576, 3608, 3635, 3666, 3694, 3718,
+  3744, 3771, 3798, 3822, 3844, 3866, 3889, 3908, 3926, 3945, 3961, 3978, 3992,
+  4005, 4018, 4031, 4043, 4054, 4064, 4073, 2697, 2701, 2703, 2709, 2715, 2723,
+  2737, 2741, 2751, 2767, 2779, 2796, 2819, 2834, 2858, 2874, 2906, 2936, 2952,
+  2988, 3019, 3050, 3084, 3125, 3156, 3202, 3235, 3275, 3309, 3341, 3378, 3406,
+  3442, 3476, 3506, 3533, 3566, 3592, 3619, 3649, 3677, 3704, 3732, 3756, 3777,
+  3801, 3824, 3850, 3871, 3891, 3910, 3930, 3947, 3963, 3980, 3994, 4007, 4019,
+  4030, 4041, 4052, 4062, 4071, 4079, 2787, 2789, 2797, 2803, 2809, 2817, 2823,
+  2832, 2848, 2862, 2870, 2888, 2913, 2932, 2948, 2971, 3000, 3028, 3051, 3074,
+  3116, 3142, 3173, 3217, 3251, 3285, 3315, 3347, 3380, 3402, 3436, 3469, 3496,
+  3525, 3556, 3584, 3610, 3637, 3663, 3690, 3714, 3740, 3765, 3789, 3812, 3830,
+  3852, 3873, 3895, 3914, 3932, 3949, 3967, 3982, 3996, 4009, 4021, 4032, 4042,
+  4051, 4060, 4069, 4077, 4084, 2882, 2886, 2892, 2894, 2900, 2914, 2918, 2926,
+  2940, 2950, 2965, 2980, 3003, 3021, 3043, 3067, 3089, 3118, 3144, 3166, 3208,
+  3238, 3269, 3295, 3327, 3356, 3384, 3412, 3438, 3467, 3491, 3521, 3548, 3574,
+  3604, 3631, 3651, 3679, 3702, 3726, 3748, 3773, 3795, 3820, 3840, 3859, 3877,
+  3897, 3916, 3936, 3953, 3969, 3984, 3998, 4011, 4023, 4034, 4044, 4053, 4061,
+  4068, 4075, 4082, 4088, 2981, 2983, 2989, 2995, 3001, 3005, 3015, 3029, 3035,
+  3047, 3061, 3075, 3097, 3122, 3136, 3162, 3188, 3218, 3242, 3267, 3291, 3319,
+  3342, 3370, 3396, 3422, 3446, 3473, 3497, 3523, 3546, 3570, 3600, 3624, 3647,
+  3673, 3696, 3716, 3742, 3763, 3785, 3803, 3826, 3848, 3869, 3887, 3904, 3920,
+  3938, 3955, 3971, 3986, 4000, 4013, 4025, 4036, 4046, 4055, 4063, 4070, 4076,
+  4081, 4086, 4091, 3077, 3081, 3085, 3087, 3093, 3103, 3114, 3126, 3134, 3150,
+  3160, 3174, 3200, 3220, 3236, 3258, 3281, 3297, 3325, 3343, 3364, 3390, 3408,
+  3434, 3457, 3483, 3507, 3529, 3554, 3572, 3596, 3617, 3641, 3669, 3686, 3710,
+  3734, 3757, 3775, 3799, 3818, 3838, 3854, 3875, 3893, 3912, 3928, 3943, 3957,
+  3973, 3988, 4002, 4015, 4027, 4038, 4048, 4057, 4065, 4072, 4078, 4083, 4087,
+  4090, 4093, 3176, 3180, 3184, 3192, 3196, 3204, 3214, 3222, 3228, 3247, 3259,
+  3273, 3287, 3299, 3323, 3335, 3357, 3374, 3394, 3416, 3432, 3452, 3477, 3489,
+  3515, 3535, 3560, 3578, 3602, 3625, 3643, 3667, 3683, 3708, 3730, 3746, 3769,
+  3791, 3808, 3828, 3846, 3867, 3883, 3899, 3918, 3934, 3951, 3965, 3976, 3990,
+  4004, 4017, 4029, 4040, 4050, 4059, 4067, 4074, 4080, 4085, 4089, 4092, 4094,
+  4095,
+};
+#endif  // CONFIG_TX64X64
+
+const SCAN_ORDER av1_default_scan_orders[TX_SIZES] = {
+#if CONFIG_CB4X4
+  { default_scan_2x2, av1_default_iscan_2x2, default_scan_2x2_neighbors },
+#endif
+  { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+  { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+  { default_scan_16x16, av1_default_iscan_16x16, default_scan_16x16_neighbors },
+  { default_scan_32x32, av1_default_iscan_32x32, default_scan_32x32_neighbors },
+#if CONFIG_TX64X64
+  { default_scan_64x64, av1_default_iscan_64x64, default_scan_64x64_neighbors },
+#endif  // CONFIG_TX64X64
+};
+
+const SCAN_ORDER av1_intra_scan_orders[TX_SIZES_ALL][TX_TYPES] = {
+#if CONFIG_CB4X4
+  {
+      // TX_2X2
+      { default_scan_2x2, av1_default_iscan_2x2, default_scan_2x2_neighbors },
+      { default_scan_2x2, av1_default_iscan_2x2, default_scan_2x2_neighbors },
+      { default_scan_2x2, av1_default_iscan_2x2, default_scan_2x2_neighbors },
+      { default_scan_2x2, av1_default_iscan_2x2, default_scan_2x2_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { mrow_scan_4x4, av1_mrow_iscan_4x4, mrow_scan_4x4_neighbors },
+      { row_scan_4x4, av1_row_iscan_4x4, row_scan_4x4_neighbors },
+      { col_scan_4x4, av1_col_iscan_4x4, col_scan_4x4_neighbors },
+      { row_scan_4x4, av1_row_iscan_4x4, row_scan_4x4_neighbors },
+      { col_scan_4x4, av1_col_iscan_4x4, col_scan_4x4_neighbors },
+      { row_scan_4x4, av1_row_iscan_4x4, row_scan_4x4_neighbors },
+      { col_scan_4x4, av1_col_iscan_4x4, col_scan_4x4_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+#endif
+  {
+      // TX_4X4
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { row_scan_4x4, av1_row_iscan_4x4, row_scan_4x4_neighbors },
+      { col_scan_4x4, av1_col_iscan_4x4, col_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { mrow_scan_4x4, av1_mrow_iscan_4x4, mrow_scan_4x4_neighbors },
+      { row_scan_4x4, av1_row_iscan_4x4, row_scan_4x4_neighbors },
+      { col_scan_4x4, av1_col_iscan_4x4, col_scan_4x4_neighbors },
+      { row_scan_4x4, av1_row_iscan_4x4, row_scan_4x4_neighbors },
+      { col_scan_4x4, av1_col_iscan_4x4, col_scan_4x4_neighbors },
+      { row_scan_4x4, av1_row_iscan_4x4, row_scan_4x4_neighbors },
+      { col_scan_4x4, av1_col_iscan_4x4, col_scan_4x4_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_8X8
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { row_scan_8x8, av1_row_iscan_8x8, row_scan_8x8_neighbors },
+      { col_scan_8x8, av1_col_iscan_8x8, col_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { mrow_scan_8x8, av1_mrow_iscan_8x8, mrow_scan_8x8_neighbors },
+      { row_scan_8x8, av1_row_iscan_8x8, row_scan_8x8_neighbors },
+      { col_scan_8x8, av1_col_iscan_8x8, col_scan_8x8_neighbors },
+      { row_scan_8x8, av1_row_iscan_8x8, row_scan_8x8_neighbors },
+      { col_scan_8x8, av1_col_iscan_8x8, col_scan_8x8_neighbors },
+      { row_scan_8x8, av1_row_iscan_8x8, row_scan_8x8_neighbors },
+      { col_scan_8x8, av1_col_iscan_8x8, col_scan_8x8_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_16X16
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { row_scan_16x16, av1_row_iscan_16x16, row_scan_16x16_neighbors },
+      { col_scan_16x16, av1_col_iscan_16x16, col_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { mrow_scan_16x16, av1_mrow_iscan_16x16, mrow_scan_16x16_neighbors },
+      { row_scan_16x16, av1_row_iscan_16x16, row_scan_16x16_neighbors },
+      { col_scan_16x16, av1_col_iscan_16x16, col_scan_16x16_neighbors },
+      { row_scan_16x16, av1_row_iscan_16x16, row_scan_16x16_neighbors },
+      { col_scan_16x16, av1_col_iscan_16x16, col_scan_16x16_neighbors },
+      { row_scan_16x16, av1_row_iscan_16x16, row_scan_16x16_neighbors },
+      { col_scan_16x16, av1_col_iscan_16x16, col_scan_16x16_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_32X32
+      { default_scan_32x32, av1_default_iscan_32x32,
+        default_scan_32x32_neighbors },
+      { h2_scan_32x32, av1_h2_iscan_32x32, h2_scan_32x32_neighbors },
+      { v2_scan_32x32, av1_v2_iscan_32x32, v2_scan_32x32_neighbors },
+      { qtr_scan_32x32, av1_qtr_iscan_32x32, qtr_scan_32x32_neighbors },
+#if CONFIG_EXT_TX
+      { h2_scan_32x32, av1_h2_iscan_32x32, h2_scan_32x32_neighbors },
+      { v2_scan_32x32, av1_v2_iscan_32x32, v2_scan_32x32_neighbors },
+      { qtr_scan_32x32, av1_qtr_iscan_32x32, qtr_scan_32x32_neighbors },
+      { qtr_scan_32x32, av1_qtr_iscan_32x32, qtr_scan_32x32_neighbors },
+      { qtr_scan_32x32, av1_qtr_iscan_32x32, qtr_scan_32x32_neighbors },
+      { mrow_scan_32x32, av1_mrow_iscan_32x32, mrow_scan_32x32_neighbors },
+      { mrow_scan_32x32, av1_mrow_iscan_32x32, mrow_scan_32x32_neighbors },
+      { mcol_scan_32x32, av1_mcol_iscan_32x32, mcol_scan_32x32_neighbors },
+      { mrow_scan_32x32, av1_mrow_iscan_32x32, mrow_scan_32x32_neighbors },
+      { mcol_scan_32x32, av1_mcol_iscan_32x32, mcol_scan_32x32_neighbors },
+      { mrow_scan_32x32, av1_mrow_iscan_32x32, mrow_scan_32x32_neighbors },
+      { mcol_scan_32x32, av1_mcol_iscan_32x32, mcol_scan_32x32_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+#if CONFIG_TX64X64
+  {
+      // TX_64X64
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+#endif  // CONFIG_TX64X64
+  {
+      // TX_4X8
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { mrow_scan_4x8, av1_mrow_iscan_4x8, mrow_scan_4x8_neighbors },
+      { mcol_scan_4x8, av1_mcol_iscan_4x8, mcol_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { mrow_scan_4x8, av1_mrow_iscan_4x8, mrow_scan_4x8_neighbors },
+      { mrow_scan_4x8, av1_mrow_iscan_4x8, mrow_scan_4x8_neighbors },
+      { mcol_scan_4x8, av1_mcol_iscan_4x8, mcol_scan_4x8_neighbors },
+      { mrow_scan_4x8, av1_mrow_iscan_4x8, mrow_scan_4x8_neighbors },
+      { mcol_scan_4x8, av1_mcol_iscan_4x8, mcol_scan_4x8_neighbors },
+      { mrow_scan_4x8, av1_mrow_iscan_4x8, mrow_scan_4x8_neighbors },
+      { mcol_scan_4x8, av1_mcol_iscan_4x8, mcol_scan_4x8_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_8X4
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { mrow_scan_8x4, av1_mrow_iscan_8x4, mrow_scan_8x4_neighbors },
+      { mcol_scan_8x4, av1_mcol_iscan_8x4, mcol_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { mrow_scan_8x4, av1_mrow_iscan_8x4, mrow_scan_8x4_neighbors },
+      { mrow_scan_8x4, av1_mrow_iscan_8x4, mrow_scan_8x4_neighbors },
+      { mcol_scan_8x4, av1_mcol_iscan_8x4, mcol_scan_8x4_neighbors },
+      { mrow_scan_8x4, av1_mrow_iscan_8x4, mrow_scan_8x4_neighbors },
+      { mcol_scan_8x4, av1_mcol_iscan_8x4, mcol_scan_8x4_neighbors },
+      { mrow_scan_8x4, av1_mrow_iscan_8x4, mrow_scan_8x4_neighbors },
+      { mcol_scan_8x4, av1_mcol_iscan_8x4, mcol_scan_8x4_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_8X16
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { mrow_scan_8x16, av1_mrow_iscan_8x16, mrow_scan_8x16_neighbors },
+      { mcol_scan_8x16, av1_mcol_iscan_8x16, mcol_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { mrow_scan_8x16, av1_mrow_iscan_8x16, mrow_scan_8x16_neighbors },
+      { mrow_scan_8x16, av1_mrow_iscan_8x16, mrow_scan_8x16_neighbors },
+      { mcol_scan_8x16, av1_mcol_iscan_8x16, mcol_scan_8x16_neighbors },
+      { mrow_scan_8x16, av1_mrow_iscan_8x16, mrow_scan_8x16_neighbors },
+      { mcol_scan_8x16, av1_mcol_iscan_8x16, mcol_scan_8x16_neighbors },
+      { mrow_scan_8x16, av1_mrow_iscan_8x16, mrow_scan_8x16_neighbors },
+      { mcol_scan_8x16, av1_mcol_iscan_8x16, mcol_scan_8x16_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_16X8
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { mrow_scan_16x8, av1_mrow_iscan_16x8, mrow_scan_16x8_neighbors },
+      { mcol_scan_16x8, av1_mcol_iscan_16x8, mcol_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { mrow_scan_16x8, av1_mrow_iscan_16x8, mrow_scan_16x8_neighbors },
+      { mrow_scan_16x8, av1_mrow_iscan_16x8, mrow_scan_16x8_neighbors },
+      { mcol_scan_16x8, av1_mcol_iscan_16x8, mcol_scan_16x8_neighbors },
+      { mrow_scan_16x8, av1_mrow_iscan_16x8, mrow_scan_16x8_neighbors },
+      { mcol_scan_16x8, av1_mcol_iscan_16x8, mcol_scan_16x8_neighbors },
+      { mrow_scan_16x8, av1_mrow_iscan_16x8, mrow_scan_16x8_neighbors },
+      { mcol_scan_16x8, av1_mcol_iscan_16x8, mcol_scan_16x8_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_16X32
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { mrow_scan_16x32, av1_mrow_iscan_16x32, mrow_scan_16x32_neighbors },
+      { mcol_scan_16x32, av1_mcol_iscan_16x32, mcol_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { mrow_scan_16x32, av1_mrow_iscan_16x32, mrow_scan_16x32_neighbors },
+      { mrow_scan_16x32, av1_mrow_iscan_16x32, mrow_scan_16x32_neighbors },
+      { mcol_scan_16x32, av1_mcol_iscan_16x32, mcol_scan_16x32_neighbors },
+      { mrow_scan_16x32, av1_mrow_iscan_16x32, mrow_scan_16x32_neighbors },
+      { mcol_scan_16x32, av1_mcol_iscan_16x32, mcol_scan_16x32_neighbors },
+      { mrow_scan_16x32, av1_mrow_iscan_16x32, mrow_scan_16x32_neighbors },
+      { mcol_scan_16x32, av1_mcol_iscan_16x32, mcol_scan_16x32_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_32X16
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { mrow_scan_32x16, av1_mrow_iscan_32x16, mrow_scan_32x16_neighbors },
+      { mcol_scan_32x16, av1_mcol_iscan_32x16, mcol_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { mrow_scan_32x16, av1_mrow_iscan_32x16, mrow_scan_32x16_neighbors },
+      { mrow_scan_32x16, av1_mrow_iscan_32x16, mrow_scan_32x16_neighbors },
+      { mcol_scan_32x16, av1_mcol_iscan_32x16, mcol_scan_32x16_neighbors },
+      { mrow_scan_32x16, av1_mrow_iscan_32x16, mrow_scan_32x16_neighbors },
+      { mcol_scan_32x16, av1_mcol_iscan_32x16, mcol_scan_32x16_neighbors },
+      { mrow_scan_32x16, av1_mrow_iscan_32x16, mrow_scan_32x16_neighbors },
+      { mcol_scan_32x16, av1_mcol_iscan_32x16, mcol_scan_32x16_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+};
+
+const SCAN_ORDER av1_inter_scan_orders[TX_SIZES_ALL][TX_TYPES] = {
+#if CONFIG_CB4X4
+  {
+      // TX_2X2
+      { default_scan_2x2, av1_default_iscan_2x2, default_scan_2x2_neighbors },
+      { default_scan_2x2, av1_default_iscan_2x2, default_scan_2x2_neighbors },
+      { default_scan_2x2, av1_default_iscan_2x2, default_scan_2x2_neighbors },
+      { default_scan_2x2, av1_default_iscan_2x2, default_scan_2x2_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { mrow_scan_4x4, av1_mrow_iscan_4x4, mrow_scan_4x4_neighbors },
+      { mrow_scan_4x4, av1_mrow_iscan_4x4, mrow_scan_4x4_neighbors },
+      { mcol_scan_4x4, av1_mcol_iscan_4x4, mcol_scan_4x4_neighbors },
+      { mrow_scan_4x4, av1_mrow_iscan_4x4, mrow_scan_4x4_neighbors },
+      { mcol_scan_4x4, av1_mcol_iscan_4x4, mcol_scan_4x4_neighbors },
+      { mrow_scan_4x4, av1_mrow_iscan_4x4, mrow_scan_4x4_neighbors },
+      { mcol_scan_4x4, av1_mcol_iscan_4x4, mcol_scan_4x4_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+#endif
+  {
+      // TX_4X4
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { default_scan_4x4, av1_default_iscan_4x4, default_scan_4x4_neighbors },
+      { mrow_scan_4x4, av1_mrow_iscan_4x4, mrow_scan_4x4_neighbors },
+      { mrow_scan_4x4, av1_mrow_iscan_4x4, mrow_scan_4x4_neighbors },
+      { mcol_scan_4x4, av1_mcol_iscan_4x4, mcol_scan_4x4_neighbors },
+      { mrow_scan_4x4, av1_mrow_iscan_4x4, mrow_scan_4x4_neighbors },
+      { mcol_scan_4x4, av1_mcol_iscan_4x4, mcol_scan_4x4_neighbors },
+      { mrow_scan_4x4, av1_mrow_iscan_4x4, mrow_scan_4x4_neighbors },
+      { mcol_scan_4x4, av1_mcol_iscan_4x4, mcol_scan_4x4_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_8X8
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { default_scan_8x8, av1_default_iscan_8x8, default_scan_8x8_neighbors },
+      { mrow_scan_8x8, av1_mrow_iscan_8x8, mrow_scan_8x8_neighbors },
+      { mrow_scan_8x8, av1_mrow_iscan_8x8, mrow_scan_8x8_neighbors },
+      { mcol_scan_8x8, av1_mcol_iscan_8x8, mcol_scan_8x8_neighbors },
+      { mrow_scan_8x8, av1_mrow_iscan_8x8, mrow_scan_8x8_neighbors },
+      { mcol_scan_8x8, av1_mcol_iscan_8x8, mcol_scan_8x8_neighbors },
+      { mrow_scan_8x8, av1_mrow_iscan_8x8, mrow_scan_8x8_neighbors },
+      { mcol_scan_8x8, av1_mcol_iscan_8x8, mcol_scan_8x8_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_16X16
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { default_scan_16x16, av1_default_iscan_16x16,
+        default_scan_16x16_neighbors },
+      { mrow_scan_16x16, av1_mrow_iscan_16x16, mrow_scan_16x16_neighbors },
+      { mrow_scan_16x16, av1_mrow_iscan_16x16, mrow_scan_16x16_neighbors },
+      { mcol_scan_16x16, av1_mcol_iscan_16x16, mcol_scan_16x16_neighbors },
+      { mrow_scan_16x16, av1_mrow_iscan_16x16, mrow_scan_16x16_neighbors },
+      { mcol_scan_16x16, av1_mcol_iscan_16x16, mcol_scan_16x16_neighbors },
+      { mrow_scan_16x16, av1_mrow_iscan_16x16, mrow_scan_16x16_neighbors },
+      { mcol_scan_16x16, av1_mcol_iscan_16x16, mcol_scan_16x16_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_32X32
+      { default_scan_32x32, av1_default_iscan_32x32,
+        default_scan_32x32_neighbors },
+      { h2_scan_32x32, av1_h2_iscan_32x32, h2_scan_32x32_neighbors },
+      { v2_scan_32x32, av1_v2_iscan_32x32, v2_scan_32x32_neighbors },
+      { qtr_scan_32x32, av1_qtr_iscan_32x32, qtr_scan_32x32_neighbors },
+#if CONFIG_EXT_TX
+      { h2_scan_32x32, av1_h2_iscan_32x32, h2_scan_32x32_neighbors },
+      { v2_scan_32x32, av1_v2_iscan_32x32, v2_scan_32x32_neighbors },
+      { qtr_scan_32x32, av1_qtr_iscan_32x32, qtr_scan_32x32_neighbors },
+      { qtr_scan_32x32, av1_qtr_iscan_32x32, qtr_scan_32x32_neighbors },
+      { qtr_scan_32x32, av1_qtr_iscan_32x32, qtr_scan_32x32_neighbors },
+      { mrow_scan_32x32, av1_mrow_iscan_32x32, mrow_scan_32x32_neighbors },
+      { mrow_scan_32x32, av1_mrow_iscan_32x32, mrow_scan_32x32_neighbors },
+      { mcol_scan_32x32, av1_mcol_iscan_32x32, mcol_scan_32x32_neighbors },
+      { mrow_scan_32x32, av1_mrow_iscan_32x32, mrow_scan_32x32_neighbors },
+      { mcol_scan_32x32, av1_mcol_iscan_32x32, mcol_scan_32x32_neighbors },
+      { mrow_scan_32x32, av1_mrow_iscan_32x32, mrow_scan_32x32_neighbors },
+      { mcol_scan_32x32, av1_mcol_iscan_32x32, mcol_scan_32x32_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+#if CONFIG_TX64X64
+  {
+      // TX_64X64
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+      { default_scan_64x64, av1_default_iscan_64x64,
+        default_scan_64x64_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+#endif  // CONFIG_TX64X64
+  {
+      // TX_4X8
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { default_scan_4x8, av1_default_iscan_4x8, default_scan_4x8_neighbors },
+      { mrow_scan_4x8, av1_mrow_iscan_4x8, mrow_scan_4x8_neighbors },
+      { mrow_scan_4x8, av1_mrow_iscan_4x8, mrow_scan_4x8_neighbors },
+      { mcol_scan_4x8, av1_mcol_iscan_4x8, mcol_scan_4x8_neighbors },
+      { mrow_scan_4x8, av1_mrow_iscan_4x8, mrow_scan_4x8_neighbors },
+      { mcol_scan_4x8, av1_mcol_iscan_4x8, mcol_scan_4x8_neighbors },
+      { mrow_scan_4x8, av1_mrow_iscan_4x8, mrow_scan_4x8_neighbors },
+      { mcol_scan_4x8, av1_mcol_iscan_4x8, mcol_scan_4x8_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_8X4
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { default_scan_8x4, av1_default_iscan_8x4, default_scan_8x4_neighbors },
+      { mrow_scan_8x4, av1_mrow_iscan_8x4, mrow_scan_8x4_neighbors },
+      { mrow_scan_8x4, av1_mrow_iscan_8x4, mrow_scan_8x4_neighbors },
+      { mcol_scan_8x4, av1_mcol_iscan_8x4, mcol_scan_8x4_neighbors },
+      { mrow_scan_8x4, av1_mrow_iscan_8x4, mrow_scan_8x4_neighbors },
+      { mcol_scan_8x4, av1_mcol_iscan_8x4, mcol_scan_8x4_neighbors },
+      { mrow_scan_8x4, av1_mrow_iscan_8x4, mrow_scan_8x4_neighbors },
+      { mcol_scan_8x4, av1_mcol_iscan_8x4, mcol_scan_8x4_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_8X16
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { default_scan_8x16, av1_default_iscan_8x16,
+        default_scan_8x16_neighbors },
+      { mrow_scan_8x16, av1_mrow_iscan_8x16, mrow_scan_8x16_neighbors },
+      { mrow_scan_8x16, av1_mrow_iscan_8x16, mrow_scan_8x16_neighbors },
+      { mcol_scan_8x16, av1_mcol_iscan_8x16, mcol_scan_8x16_neighbors },
+      { mrow_scan_8x16, av1_mrow_iscan_8x16, mrow_scan_8x16_neighbors },
+      { mcol_scan_8x16, av1_mcol_iscan_8x16, mcol_scan_8x16_neighbors },
+      { mrow_scan_8x16, av1_mrow_iscan_8x16, mrow_scan_8x16_neighbors },
+      { mcol_scan_8x16, av1_mcol_iscan_8x16, mcol_scan_8x16_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_16X8
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { default_scan_16x8, av1_default_iscan_16x8,
+        default_scan_16x8_neighbors },
+      { mrow_scan_16x8, av1_mrow_iscan_16x8, mrow_scan_16x8_neighbors },
+      { mrow_scan_16x8, av1_mrow_iscan_16x8, mrow_scan_16x8_neighbors },
+      { mcol_scan_16x8, av1_mcol_iscan_16x8, mcol_scan_16x8_neighbors },
+      { mrow_scan_16x8, av1_mrow_iscan_16x8, mrow_scan_16x8_neighbors },
+      { mcol_scan_16x8, av1_mcol_iscan_16x8, mcol_scan_16x8_neighbors },
+      { mrow_scan_16x8, av1_mrow_iscan_16x8, mrow_scan_16x8_neighbors },
+      { mcol_scan_16x8, av1_mcol_iscan_16x8, mcol_scan_16x8_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_16X32
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { default_scan_16x32, av1_default_iscan_16x32,
+        default_scan_16x32_neighbors },
+      { mrow_scan_16x32, av1_mrow_iscan_16x32, mrow_scan_16x32_neighbors },
+      { mrow_scan_16x32, av1_mrow_iscan_16x32, mrow_scan_16x32_neighbors },
+      { mcol_scan_16x32, av1_mcol_iscan_16x32, mcol_scan_16x32_neighbors },
+      { mrow_scan_16x32, av1_mrow_iscan_16x32, mrow_scan_16x32_neighbors },
+      { mcol_scan_16x32, av1_mcol_iscan_16x32, mcol_scan_16x32_neighbors },
+      { mrow_scan_16x32, av1_mrow_iscan_16x32, mrow_scan_16x32_neighbors },
+      { mcol_scan_16x32, av1_mcol_iscan_16x32, mcol_scan_16x32_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_32X16
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { default_scan_32x16, av1_default_iscan_32x16,
+        default_scan_32x16_neighbors },
+      { mrow_scan_32x16, av1_mrow_iscan_32x16, mrow_scan_32x16_neighbors },
+      { mrow_scan_32x16, av1_mrow_iscan_32x16, mrow_scan_32x16_neighbors },
+      { mcol_scan_32x16, av1_mcol_iscan_32x16, mcol_scan_32x16_neighbors },
+      { mrow_scan_32x16, av1_mrow_iscan_32x16, mrow_scan_32x16_neighbors },
+      { mcol_scan_32x16, av1_mcol_iscan_32x16, mcol_scan_32x16_neighbors },
+      { mrow_scan_32x16, av1_mrow_iscan_32x16, mrow_scan_32x16_neighbors },
+      { mcol_scan_32x16, av1_mcol_iscan_32x16, mcol_scan_32x16_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_4X16
+      { default_scan_4x16, av1_default_iscan_4x16,
+        default_scan_4x16_neighbors },
+      { default_scan_4x16, av1_default_iscan_4x16,
+        default_scan_4x16_neighbors },
+      { default_scan_4x16, av1_default_iscan_4x16,
+        default_scan_4x16_neighbors },
+      { default_scan_4x16, av1_default_iscan_4x16,
+        default_scan_4x16_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_4x16, av1_default_iscan_4x16,
+        default_scan_4x16_neighbors },
+      { default_scan_4x16, av1_default_iscan_4x16,
+        default_scan_4x16_neighbors },
+      { default_scan_4x16, av1_default_iscan_4x16,
+        default_scan_4x16_neighbors },
+      { default_scan_4x16, av1_default_iscan_4x16,
+        default_scan_4x16_neighbors },
+      { default_scan_4x16, av1_default_iscan_4x16,
+        default_scan_4x16_neighbors },
+      { mrow_scan_4x16, av1_mrow_iscan_4x16, mrow_scan_4x16_neighbors },
+      { mrow_scan_4x16, av1_mrow_iscan_4x16, mrow_scan_4x16_neighbors },
+      { mcol_scan_4x16, av1_mcol_iscan_4x16, mcol_scan_4x16_neighbors },
+      { mrow_scan_4x16, av1_mrow_iscan_4x16, mrow_scan_4x16_neighbors },
+      { mcol_scan_4x16, av1_mcol_iscan_4x16, mcol_scan_4x16_neighbors },
+      { mrow_scan_4x16, av1_mrow_iscan_4x16, mrow_scan_4x16_neighbors },
+      { mcol_scan_4x16, av1_mcol_iscan_4x16, mcol_scan_4x16_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_16X4
+      { default_scan_16x4, av1_default_iscan_16x4,
+        default_scan_16x4_neighbors },
+      { default_scan_16x4, av1_default_iscan_16x4,
+        default_scan_16x4_neighbors },
+      { default_scan_16x4, av1_default_iscan_16x4,
+        default_scan_16x4_neighbors },
+      { default_scan_16x4, av1_default_iscan_16x4,
+        default_scan_16x4_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_16x4, av1_default_iscan_16x4,
+        default_scan_16x4_neighbors },
+      { default_scan_16x4, av1_default_iscan_16x4,
+        default_scan_16x4_neighbors },
+      { default_scan_16x4, av1_default_iscan_16x4,
+        default_scan_16x4_neighbors },
+      { default_scan_16x4, av1_default_iscan_16x4,
+        default_scan_16x4_neighbors },
+      { default_scan_16x4, av1_default_iscan_16x4,
+        default_scan_16x4_neighbors },
+      { mrow_scan_16x4, av1_mrow_iscan_16x4, mrow_scan_16x4_neighbors },
+      { mrow_scan_16x4, av1_mrow_iscan_16x4, mrow_scan_16x4_neighbors },
+      { mcol_scan_16x4, av1_mcol_iscan_16x4, mcol_scan_16x4_neighbors },
+      { mrow_scan_16x4, av1_mrow_iscan_16x4, mrow_scan_16x4_neighbors },
+      { mcol_scan_16x4, av1_mcol_iscan_16x4, mcol_scan_16x4_neighbors },
+      { mrow_scan_16x4, av1_mrow_iscan_16x4, mrow_scan_16x4_neighbors },
+      { mcol_scan_16x4, av1_mcol_iscan_16x4, mcol_scan_16x4_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_8X32
+      { default_scan_8x32, av1_default_iscan_8x32,
+        default_scan_8x32_neighbors },
+      { default_scan_8x32, av1_default_iscan_8x32,
+        default_scan_8x32_neighbors },
+      { default_scan_8x32, av1_default_iscan_8x32,
+        default_scan_8x32_neighbors },
+      { default_scan_8x32, av1_default_iscan_8x32,
+        default_scan_8x32_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_8x32, av1_default_iscan_8x32,
+        default_scan_8x32_neighbors },
+      { default_scan_8x32, av1_default_iscan_8x32,
+        default_scan_8x32_neighbors },
+      { default_scan_8x32, av1_default_iscan_8x32,
+        default_scan_8x32_neighbors },
+      { default_scan_8x32, av1_default_iscan_8x32,
+        default_scan_8x32_neighbors },
+      { default_scan_8x32, av1_default_iscan_8x32,
+        default_scan_8x32_neighbors },
+      { mrow_scan_8x32, av1_mrow_iscan_8x32, mrow_scan_8x32_neighbors },
+      { mrow_scan_8x32, av1_mrow_iscan_8x32, mrow_scan_8x32_neighbors },
+      { mcol_scan_8x32, av1_mcol_iscan_8x32, mcol_scan_8x32_neighbors },
+      { mrow_scan_8x32, av1_mrow_iscan_8x32, mrow_scan_8x32_neighbors },
+      { mcol_scan_8x32, av1_mcol_iscan_8x32, mcol_scan_8x32_neighbors },
+      { mrow_scan_8x32, av1_mrow_iscan_8x32, mrow_scan_8x32_neighbors },
+      { mcol_scan_8x32, av1_mcol_iscan_8x32, mcol_scan_8x32_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // TX_32X8
+      { default_scan_32x8, av1_default_iscan_32x8,
+        default_scan_32x8_neighbors },
+      { default_scan_32x8, av1_default_iscan_32x8,
+        default_scan_32x8_neighbors },
+      { default_scan_32x8, av1_default_iscan_32x8,
+        default_scan_32x8_neighbors },
+      { default_scan_32x8, av1_default_iscan_32x8,
+        default_scan_32x8_neighbors },
+#if CONFIG_EXT_TX
+      { default_scan_32x8, av1_default_iscan_32x8,
+        default_scan_32x8_neighbors },
+      { default_scan_32x8, av1_default_iscan_32x8,
+        default_scan_32x8_neighbors },
+      { default_scan_32x8, av1_default_iscan_32x8,
+        default_scan_32x8_neighbors },
+      { default_scan_32x8, av1_default_iscan_32x8,
+        default_scan_32x8_neighbors },
+      { default_scan_32x8, av1_default_iscan_32x8,
+        default_scan_32x8_neighbors },
+      { mrow_scan_32x8, av1_mrow_iscan_32x8, mrow_scan_32x8_neighbors },
+      { mrow_scan_32x8, av1_mrow_iscan_32x8, mrow_scan_32x8_neighbors },
+      { mcol_scan_32x8, av1_mcol_iscan_32x8, mcol_scan_32x8_neighbors },
+      { mrow_scan_32x8, av1_mrow_iscan_32x8, mrow_scan_32x8_neighbors },
+      { mcol_scan_32x8, av1_mcol_iscan_32x8, mcol_scan_32x8_neighbors },
+      { mrow_scan_32x8, av1_mrow_iscan_32x8, mrow_scan_32x8_neighbors },
+      { mcol_scan_32x8, av1_mcol_iscan_32x8, mcol_scan_32x8_neighbors },
+#endif  // CONFIG_EXT_TX
+  },
+};
+
+#if CONFIG_ADAPT_SCAN
+// TX_32X32 will has 1024 coefficients whose indexes can be represented in 10
+// bits
+#define COEFF_IDX_BITS (10 + CONFIG_TX64X64)
+#define COEFF_IDX_SIZE (1 << COEFF_IDX_BITS)
+#define COEFF_IDX_MASK (COEFF_IDX_SIZE - 1)
+
+static uint32_t *get_non_zero_prob(FRAME_CONTEXT *fc, TX_SIZE tx_size,
+                                   TX_TYPE tx_type) {
+  switch (tx_size) {
+#if CONFIG_CB4X4
+    case TX_2X2: return fc->non_zero_prob_2x2[tx_type];
+#endif
+    case TX_4X4: return fc->non_zero_prob_4X4[tx_type];
+    case TX_8X8: return fc->non_zero_prob_8X8[tx_type];
+    case TX_16X16: return fc->non_zero_prob_16X16[tx_type];
+    case TX_32X32: return fc->non_zero_prob_32X32[tx_type];
+#if CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    case TX_4X8: return fc->non_zero_prob_4X8[tx_type];
+    case TX_8X4: return fc->non_zero_prob_8X4[tx_type];
+    case TX_8X16: return fc->non_zero_prob_8X16[tx_type];
+    case TX_16X8: return fc->non_zero_prob_16X8[tx_type];
+    case TX_16X32: return fc->non_zero_prob_16X32[tx_type];
+    case TX_32X16: return fc->non_zero_prob_32X16[tx_type];
+#endif  // CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    default: assert(0); return NULL;
+  }
+}
+
+static int16_t *get_adapt_scan(FRAME_CONTEXT *fc, TX_SIZE tx_size,
+                               TX_TYPE tx_type) {
+  switch (tx_size) {
+#if CONFIG_CB4X4
+    case TX_2X2: return fc->scan_2x2[tx_type];
+#endif
+    case TX_4X4: return fc->scan_4X4[tx_type];
+    case TX_8X8: return fc->scan_8X8[tx_type];
+    case TX_16X16: return fc->scan_16X16[tx_type];
+    case TX_32X32: return fc->scan_32X32[tx_type];
+#if CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    case TX_4X8: return fc->scan_4X8[tx_type];
+    case TX_8X4: return fc->scan_8X4[tx_type];
+    case TX_8X16: return fc->scan_8X16[tx_type];
+    case TX_16X8: return fc->scan_16X8[tx_type];
+    case TX_16X32: return fc->scan_16X32[tx_type];
+    case TX_32X16: return fc->scan_32X16[tx_type];
+#endif  // CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    default: assert(0); return NULL;
+  }
+}
+
+static int16_t *get_adapt_iscan(FRAME_CONTEXT *fc, TX_SIZE tx_size,
+                                TX_TYPE tx_type) {
+  switch (tx_size) {
+#if CONFIG_CB4X4
+    case TX_2X2: return fc->iscan_2x2[tx_type];
+#endif
+    case TX_4X4: return fc->iscan_4X4[tx_type];
+    case TX_8X8: return fc->iscan_8X8[tx_type];
+    case TX_16X16: return fc->iscan_16X16[tx_type];
+    case TX_32X32: return fc->iscan_32X32[tx_type];
+#if CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    case TX_4X8: return fc->iscan_4X8[tx_type];
+    case TX_8X4: return fc->iscan_8X4[tx_type];
+    case TX_8X16: return fc->iscan_8X16[tx_type];
+    case TX_16X8: return fc->iscan_16X8[tx_type];
+    case TX_16X32: return fc->iscan_16X32[tx_type];
+    case TX_32X16: return fc->iscan_32X16[tx_type];
+#endif  // CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    default: assert(0); return NULL;
+  }
+}
+
+static int16_t *get_adapt_nb(FRAME_CONTEXT *fc, TX_SIZE tx_size,
+                             TX_TYPE tx_type) {
+  switch (tx_size) {
+#if CONFIG_CB4X4
+    case TX_2X2: return fc->nb_2x2[tx_type];
+#endif
+    case TX_4X4: return fc->nb_4X4[tx_type];
+    case TX_8X8: return fc->nb_8X8[tx_type];
+    case TX_16X16: return fc->nb_16X16[tx_type];
+    case TX_32X32: return fc->nb_32X32[tx_type];
+#if CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    case TX_4X8: return fc->nb_4X8[tx_type];
+    case TX_8X4: return fc->nb_8X4[tx_type];
+    case TX_8X16: return fc->nb_8X16[tx_type];
+    case TX_16X8: return fc->nb_16X8[tx_type];
+    case TX_16X32: return fc->nb_16X32[tx_type];
+    case TX_32X16: return fc->nb_32X16[tx_type];
+#endif  // CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    default: assert(0); return NULL;
+  }
+}
+
+static uint32_t *get_non_zero_counts(FRAME_COUNTS *counts, TX_SIZE tx_size,
+                                     TX_TYPE tx_type) {
+  switch (tx_size) {
+#if CONFIG_CB4X4
+    case TX_2X2: return counts->non_zero_count_2x2[tx_type];
+#endif
+    case TX_4X4: return counts->non_zero_count_4X4[tx_type];
+    case TX_8X8: return counts->non_zero_count_8X8[tx_type];
+    case TX_16X16: return counts->non_zero_count_16X16[tx_type];
+    case TX_32X32: return counts->non_zero_count_32X32[tx_type];
+#if CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    case TX_4X8: return counts->non_zero_count_4x8[tx_type];
+    case TX_8X4: return counts->non_zero_count_8x4[tx_type];
+    case TX_8X16: return counts->non_zero_count_8x16[tx_type];
+    case TX_16X8: return counts->non_zero_count_16x8[tx_type];
+    case TX_16X32: return counts->non_zero_count_16x32[tx_type];
+    case TX_32X16: return counts->non_zero_count_32x16[tx_type];
+#endif  // CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    default: assert(0); return NULL;
+  }
+}
+
+static INLINE int clamp_64(int64_t value, int low, int high) {
+  return value < low ? low : (value > high ? high : (int)value);
+}
+
+static void update_scan_prob(AV1_COMMON *cm, TX_SIZE tx_size, TX_TYPE tx_type,
+                             int rate_16) {
+  FRAME_CONTEXT *pre_fc = &cm->frame_contexts[cm->frame_context_idx];
+  uint32_t *prev_non_zero_prob = get_non_zero_prob(pre_fc, tx_size, tx_type);
+  uint32_t *non_zero_prob = get_non_zero_prob(cm->fc, tx_size, tx_type);
+  uint32_t *non_zero_count = get_non_zero_counts(&cm->counts, tx_size, tx_type);
+  const int tx2d_size = tx_size_2d[tx_size];
+  unsigned int block_num = cm->counts.txb_count[tx_size][tx_type];
+  int i;
+  for (i = 0; i < tx2d_size; i++) {
+    int64_t curr_prob =
+        block_num == 0 ? 0 : (non_zero_count[i] << 16) / block_num;
+    int64_t prev_prob = prev_non_zero_prob[i];
+    int64_t pred_prob =
+        (curr_prob * rate_16 + prev_prob * ((1 << 16) - rate_16)) >> 16;
+    // TODO(angiebird): reduce the bit usage of probabilities and remove
+    // clamp_64()
+    non_zero_prob[i] = clamp_64(pred_prob, 0, UINT16_MAX);
+  }
+}
+
+static void update_scan_count(int16_t *scan, int max_scan,
+                              const tran_low_t *dqcoeffs,
+                              uint32_t *non_zero_count) {
+  int i;
+  for (i = 0; i < max_scan; ++i) {
+    int coeff_idx = scan[i];
+    non_zero_count[coeff_idx] += (dqcoeffs[coeff_idx] != 0);
+  }
+}
+
+void av1_update_scan_count_facade(AV1_COMMON *cm, FRAME_COUNTS *counts,
+                                  TX_SIZE tx_size, TX_TYPE tx_type,
+                                  const tran_low_t *dqcoeffs, int max_scan) {
+  int16_t *scan = get_adapt_scan(cm->fc, tx_size, tx_type);
+  uint32_t *non_zero_count = get_non_zero_counts(counts, tx_size, tx_type);
+  update_scan_count(scan, max_scan, dqcoeffs, non_zero_count);
+  ++counts->txb_count[tx_size][tx_type];
+}
+
+static int cmp_prob(const void *a, const void *b) {
+  return *(const uint32_t *)b > *(const uint32_t *)a ? 1 : -1;
+}
+
+void av1_augment_prob(TX_SIZE tx_size, TX_TYPE tx_type, uint32_t *prob) {
+  // TODO(angiebird): check if we need is_inter here
+  const SCAN_ORDER *sc = get_default_scan(tx_size, tx_type, 0);
+  const int tx1d_wide = tx_size_wide[tx_size];
+  const int tx1d_high = tx_size_high[tx_size];
+  int r, c;
+  for (r = 0; r < tx1d_high; r++) {
+    for (c = 0; c < tx1d_wide; c++) {
+      const int idx = r * tx1d_wide + c;
+      const uint32_t mask_16 = ((1 << 16) - 1);
+      const uint32_t tie_breaker = ~((uint32_t)sc->iscan[idx]);
+      // prob[idx]: 16 bits  dummy: 6 bits  scan_idx: 10 bits
+      prob[idx] = (prob[idx] << 16) | (mask_16 & tie_breaker);
+    }
+  }
+}
+
+// topological sort
+static void dfs_scan(int tx1d_size, int *scan_idx, int coeff_idx, int16_t *scan,
+                     int16_t *iscan) {
+  const int r = coeff_idx / tx1d_size;
+  const int c = coeff_idx % tx1d_size;
+
+  if (iscan[coeff_idx] != -1) return;
+
+  if (r > 0) dfs_scan(tx1d_size, scan_idx, coeff_idx - tx1d_size, scan, iscan);
+
+  if (c > 0) dfs_scan(tx1d_size, scan_idx, coeff_idx - 1, scan, iscan);
+
+  scan[*scan_idx] = coeff_idx;
+  iscan[coeff_idx] = *scan_idx;
+  ++(*scan_idx);
+}
+
+void av1_update_neighbors(int tx_size, const int16_t *scan,
+                          const int16_t *iscan, int16_t *neighbors) {
+  const int tx1d_wide = tx_size_wide[tx_size];
+  const int tx1d_high = tx_size_high[tx_size];
+  const int tx2d_size = tx_size_2d[tx_size];
+  int scan_idx;
+  for (scan_idx = 0; scan_idx < tx2d_size; ++scan_idx) {
+    const int coeff_idx = scan[scan_idx];
+    const int r = coeff_idx / tx1d_wide;
+    const int c = coeff_idx % tx1d_wide;
+    const int nb_offset_r[5] = { -1, 0, -1, -1, 1 };
+    const int nb_offset_c[5] = { 0, -1, -1, 1, -1 };
+    const int nb_num = 5;
+    int nb_count = 0;
+    int nb_idx;
+
+    for (nb_idx = 0; nb_idx < nb_num; ++nb_idx) {
+      if (nb_count < 2) {
+        int nb_r = r + nb_offset_r[nb_idx];
+        int nb_c = c + nb_offset_c[nb_idx];
+        int nb_coeff_idx = nb_r * tx1d_wide + nb_c;
+        int valid_pos =
+            nb_r >= 0 && nb_r < tx1d_high && nb_c >= 0 && nb_c < tx1d_wide;
+        if (valid_pos && iscan[nb_coeff_idx] < scan_idx) {
+          neighbors[scan_idx * MAX_NEIGHBORS + nb_count] = nb_coeff_idx;
+          ++nb_count;
+        }
+      } else {
+        break;
+      }
+    }
+
+    if (nb_count == 1) {
+      neighbors[scan_idx * MAX_NEIGHBORS + 1] =
+          neighbors[scan_idx * MAX_NEIGHBORS + 0];
+    } else if (nb_count == 0) {
+      neighbors[scan_idx * MAX_NEIGHBORS + 0] = scan[0];
+      neighbors[scan_idx * MAX_NEIGHBORS + 1] = scan[0];
+    }
+  }
+  neighbors[tx2d_size * MAX_NEIGHBORS + 0] = scan[0];
+  neighbors[tx2d_size * MAX_NEIGHBORS + 1] = scan[0];
+}
+
+void av1_update_sort_order(TX_SIZE tx_size, TX_TYPE tx_type,
+                           const uint32_t *non_zero_prob, int16_t *sort_order) {
+  const SCAN_ORDER *sc = get_default_scan(tx_size, tx_type, 0);
+  uint32_t temp[COEFF_IDX_SIZE];
+  const int tx2d_size = tx_size_2d[tx_size];
+  int sort_idx;
+  assert(tx2d_size <= COEFF_IDX_SIZE);
+  memcpy(temp, non_zero_prob, tx2d_size * sizeof(*non_zero_prob));
+  av1_augment_prob(tx_size, tx_type, temp);
+  qsort(temp, tx2d_size, sizeof(*temp), cmp_prob);
+  for (sort_idx = 0; sort_idx < tx2d_size; ++sort_idx) {
+    const int default_scan_idx =
+        (temp[sort_idx] & COEFF_IDX_MASK) ^ COEFF_IDX_MASK;
+    const int coeff_idx = sc->scan[default_scan_idx];
+    sort_order[sort_idx] = coeff_idx;
+  }
+}
+
+void av1_update_scan_order(TX_SIZE tx_size, int16_t *sort_order, int16_t *scan,
+                           int16_t *iscan) {
+  int coeff_idx;
+  int scan_idx;
+  int sort_idx;
+  const int tx1d_size = tx_size_wide[tx_size];
+  const int tx2d_size = tx_size_2d[tx_size];
+
+  for (coeff_idx = 0; coeff_idx < tx2d_size; ++coeff_idx) {
+    iscan[coeff_idx] = -1;
+  }
+
+  scan_idx = 0;
+  for (sort_idx = 0; sort_idx < tx2d_size; ++sort_idx) {
+    coeff_idx = sort_order[sort_idx];
+    dfs_scan(tx1d_size, &scan_idx, coeff_idx, scan, iscan);
+  }
+}
+
+static void update_scan_order_facade(AV1_COMMON *cm, TX_SIZE tx_size,
+                                     TX_TYPE tx_type) {
+  int16_t sort_order[COEFF_IDX_SIZE];
+  uint32_t *non_zero_prob = get_non_zero_prob(cm->fc, tx_size, tx_type);
+  int16_t *scan = get_adapt_scan(cm->fc, tx_size, tx_type);
+  int16_t *iscan = get_adapt_iscan(cm->fc, tx_size, tx_type);
+  int16_t *nb = get_adapt_nb(cm->fc, tx_size, tx_type);
+  assert(tx_size_2d[tx_size] <= COEFF_IDX_SIZE);
+  av1_update_sort_order(tx_size, tx_type, non_zero_prob, sort_order);
+  av1_update_scan_order(tx_size, sort_order, scan, iscan);
+  av1_update_neighbors(tx_size, scan, iscan, nb);
+}
+
+static void update_eob_threshold(AV1_COMMON *cm, TX_SIZE tx_size,
+                                 TX_TYPE tx_type) {
+  int i, row, col, row_limit, col_limit, cal_idx = 0;
+  const int tx_width = tx_size_wide[tx_size];
+  const int tx_height = tx_size_high[tx_size];
+
+  row_limit = tx_width >> 1;
+  col_limit = tx_height >> 1;
+
+  if (tx_width >= 8 && tx_height >= 8) {
+    SCAN_ORDER *sc = &cm->fc->sc[tx_size][tx_type];
+    int16_t *threshold = &cm->fc->eob_threshold[tx_size][tx_type][0];
+    const int tx2d_size = tx_size_2d[tx_size];
+
+    while (cal_idx < EOB_THRESHOLD_NUM) {
+      for (i = 0; i < tx2d_size; ++i) {
+        row = sc->scan[i] / tx_height;
+        col = sc->scan[i] % tx_width;
+        if (row >= row_limit || col >= col_limit) break;
+      }
+      row_limit >>= 1;
+      col_limit >>= 1;
+      threshold[cal_idx] = i;
+      cal_idx++;
+    }
+  }
+}
+
+void av1_init_scan_order(AV1_COMMON *cm) {
+  TX_SIZE tx_size;
+  TX_TYPE tx_type;
+  for (tx_size = 0; tx_size < TX_SIZES_ALL; ++tx_size) {
+#if CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    if (tx_size > TX_32X16) continue;
+#else
+    if (tx_size >= TX_SIZES) continue;
+#endif  // CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    for (tx_type = DCT_DCT; tx_type < TX_TYPES; ++tx_type) {
+      uint32_t *non_zero_prob = get_non_zero_prob(cm->fc, tx_size, tx_type);
+      const int tx2d_size = tx_size_2d[tx_size];
+      int i;
+      SCAN_ORDER *sc = &cm->fc->sc[tx_size][tx_type];
+      for (i = 0; i < tx2d_size; ++i) {
+        non_zero_prob[i] = (1 << 16) / 2;  // init non_zero_prob to 0.5
+      }
+      update_scan_order_facade(cm, tx_size, tx_type);
+      sc->scan = get_adapt_scan(cm->fc, tx_size, tx_type);
+      sc->iscan = get_adapt_iscan(cm->fc, tx_size, tx_type);
+      sc->neighbors = get_adapt_nb(cm->fc, tx_size, tx_type);
+      update_eob_threshold(cm, tx_size, tx_type);
+    }
+  }
+}
+
+void av1_adapt_scan_order(AV1_COMMON *cm) {
+  TX_SIZE tx_size;
+  for (tx_size = 0; tx_size < TX_SIZES_ALL; ++tx_size) {
+#if CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    if (tx_size > TX_32X16) continue;
+#else
+    if (tx_size >= TX_SIZES) continue;
+#endif  // CONFIG_RECT_TX && (CONFIG_EXT_TX || CONFIG_VAR_TX)
+    TX_TYPE tx_type;
+    for (tx_type = DCT_DCT; tx_type < TX_TYPES; ++tx_type) {
+      update_scan_prob(cm, tx_size, tx_type, ADAPT_SCAN_UPDATE_RATE_16);
+      update_scan_order_facade(cm, tx_size, tx_type);
+      update_eob_threshold(cm, tx_size, tx_type);
+    }
+  }
+}
+
+void av1_deliver_eob_threshold(const AV1_COMMON *cm, MACROBLOCKD *xd) {
+  xd->eob_threshold_md = (const EobThresholdMD *)cm->fc->eob_threshold;
+}
+#endif  // CONFIG_ADAPT_SCAN
diff --git a/third_party/aom/av1/common/scan.h b/third_party/aom/av1/common/scan.h
new file mode 100644
index 0000000000..ecef113686
--- /dev/null
+++ b/third_party/aom/av1/common/scan.h
@@ -0,0 +1,95 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_SCAN_H_
+#define AV1_COMMON_SCAN_H_
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+#include "av1/common/enums.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_NEIGHBORS 2
+
+extern const SCAN_ORDER av1_default_scan_orders[TX_SIZES];
+extern const SCAN_ORDER av1_intra_scan_orders[TX_SIZES_ALL][TX_TYPES];
+extern const SCAN_ORDER av1_inter_scan_orders[TX_SIZES_ALL][TX_TYPES];
+
+#if CONFIG_ADAPT_SCAN
+void av1_update_scan_count_facade(AV1_COMMON *cm, FRAME_COUNTS *counts,
+                                  TX_SIZE tx_size, TX_TYPE tx_type,
+                                  const tran_low_t *dqcoeffs, int max_scan);
+
+// embed r + c and coeff_idx info with nonzero probabilities. When sorting the
+// nonzero probabilities, if there is a tie, the coefficient with smaller r + c
+// will be scanned first
+void av1_augment_prob(TX_SIZE tx_size, TX_TYPE tx_type, uint32_t *prob);
+
+// apply quick sort on nonzero probabilities to obtain a sort order
+void av1_update_sort_order(TX_SIZE tx_size, TX_TYPE tx_type,
+                           const uint32_t *non_zero_prob, int16_t *sort_order);
+
+// apply topological sort on the nonzero probabilities sorting order to
+// guarantee each to-be-scanned coefficient's upper and left coefficient will be
+// scanned before the to-be-scanned coefficient.
+void av1_update_scan_order(TX_SIZE tx_size, int16_t *sort_order, int16_t *scan,
+                           int16_t *iscan);
+
+// For each coeff_idx in scan[], update its above and left neighbors in
+// neighbors[] accordingly.
+void av1_update_neighbors(int tx_size, const int16_t *scan,
+                          const int16_t *iscan, int16_t *neighbors);
+void av1_init_scan_order(AV1_COMMON *cm);
+void av1_adapt_scan_order(AV1_COMMON *cm);
+#endif
+void av1_deliver_eob_threshold(const AV1_COMMON *cm, MACROBLOCKD *xd);
+
+static INLINE int get_coef_context(const int16_t *neighbors,
+                                   const uint8_t *token_cache, int c) {
+  return (1 + token_cache[neighbors[MAX_NEIGHBORS * c + 0]] +
+          token_cache[neighbors[MAX_NEIGHBORS * c + 1]]) >>
+         1;
+}
+
+static INLINE const SCAN_ORDER *get_default_scan(TX_SIZE tx_size,
+                                                 TX_TYPE tx_type,
+                                                 int is_inter) {
+#if CONFIG_EXT_TX || CONFIG_VAR_TX
+  return is_inter ? &av1_inter_scan_orders[tx_size][tx_type]
+                  : &av1_intra_scan_orders[tx_size][tx_type];
+#else
+  (void)is_inter;
+  return &av1_intra_scan_orders[tx_size][tx_type];
+#endif  // CONFIG_EXT_TX
+}
+
+static INLINE const SCAN_ORDER *get_scan(const AV1_COMMON *cm, TX_SIZE tx_size,
+                                         TX_TYPE tx_type, int is_inter) {
+#if CONFIG_ADAPT_SCAN
+  (void)is_inter;
+  return &cm->fc->sc[tx_size][tx_type];
+#else   // CONFIG_ADAPT_SCAN
+  (void)cm;
+  return get_default_scan(tx_size, tx_type, is_inter);
+#endif  // CONFIG_ADAPT_SCAN
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_SCAN_H_
diff --git a/third_party/aom/av1/common/seg_common.c b/third_party/aom/av1/common/seg_common.c
new file mode 100644
index 0000000000..21a8536297
--- /dev/null
+++ b/third_party/aom/av1/common/seg_common.c
@@ -0,0 +1,62 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/av1_loopfilter.h"
+#include "av1/common/blockd.h"
+#include "av1/common/seg_common.h"
+#include "av1/common/quant_common.h"
+
+static const int seg_feature_data_signed[SEG_LVL_MAX] = { 1, 1, 0, 0 };
+
+static const int seg_feature_data_max[SEG_LVL_MAX] = { MAXQ, MAX_LOOP_FILTER, 3,
+                                                       0 };
+
+// These functions provide access to new segment level features.
+// Eventually these function may be "optimized out" but for the moment,
+// the coding mechanism is still subject to change so these provide a
+// convenient single point of change.
+
+void av1_clearall_segfeatures(struct segmentation *seg) {
+  av1_zero(seg->feature_data);
+  av1_zero(seg->feature_mask);
+}
+
+void av1_enable_segfeature(struct segmentation *seg, int segment_id,
+                           SEG_LVL_FEATURES feature_id) {
+  seg->feature_mask[segment_id] |= 1 << feature_id;
+}
+
+int av1_seg_feature_data_max(SEG_LVL_FEATURES feature_id) {
+  return seg_feature_data_max[feature_id];
+}
+
+int av1_is_segfeature_signed(SEG_LVL_FEATURES feature_id) {
+  return seg_feature_data_signed[feature_id];
+}
+
+void av1_set_segdata(struct segmentation *seg, int segment_id,
+                     SEG_LVL_FEATURES feature_id, int seg_data) {
+  assert(seg_data <= seg_feature_data_max[feature_id]);
+  if (seg_data < 0) {
+    assert(seg_feature_data_signed[feature_id]);
+    assert(-seg_data <= seg_feature_data_max[feature_id]);
+  }
+
+  seg->feature_data[segment_id][feature_id] = seg_data;
+}
+
+const aom_tree_index av1_segment_tree[TREE_SIZE(MAX_SEGMENTS)] = {
+  2, 4, 6, 8, 10, 12, 0, -1, -2, -3, -4, -5, -6, -7
+};
+
+// TBD? Functions to read and write segment data with range / validity checking
diff --git a/third_party/aom/av1/common/seg_common.h b/third_party/aom/av1/common/seg_common.h
new file mode 100644
index 0000000000..03ed38e790
--- /dev/null
+++ b/third_party/aom/av1/common/seg_common.h
@@ -0,0 +1,86 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_SEG_COMMON_H_
+#define AV1_COMMON_SEG_COMMON_H_
+
+#include "aom_dsp/prob.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define SEGMENT_DELTADATA 0
+#define SEGMENT_ABSDATA 1
+
+#define MAX_SEGMENTS 8
+#define SEG_TREE_PROBS (MAX_SEGMENTS - 1)
+
+#define PREDICTION_PROBS 3
+
+// Segment level features.
+typedef enum {
+  SEG_LVL_ALT_Q = 0,      // Use alternate Quantizer ....
+  SEG_LVL_ALT_LF = 1,     // Use alternate loop filter value...
+  SEG_LVL_REF_FRAME = 2,  // Optional Segment reference frame
+  SEG_LVL_SKIP = 3,       // Optional Segment (0,0) + skip mode
+  SEG_LVL_MAX = 4         // Number of features supported
+} SEG_LVL_FEATURES;
+
+struct segmentation {
+  uint8_t enabled;
+  uint8_t update_map;
+  uint8_t update_data;
+  uint8_t abs_delta;
+  uint8_t temporal_update;
+
+  int16_t feature_data[MAX_SEGMENTS][SEG_LVL_MAX];
+  unsigned int feature_mask[MAX_SEGMENTS];
+};
+
+struct segmentation_probs {
+  aom_prob tree_probs[SEG_TREE_PROBS];
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob tree_cdf[CDF_SIZE(MAX_SEGMENTS)];
+#endif
+  aom_prob pred_probs[PREDICTION_PROBS];
+};
+
+static INLINE int segfeature_active(const struct segmentation *seg,
+                                    int segment_id,
+                                    SEG_LVL_FEATURES feature_id) {
+  return seg->enabled && (seg->feature_mask[segment_id] & (1 << feature_id));
+}
+
+void av1_clearall_segfeatures(struct segmentation *seg);
+
+void av1_enable_segfeature(struct segmentation *seg, int segment_id,
+                           SEG_LVL_FEATURES feature_id);
+
+int av1_seg_feature_data_max(SEG_LVL_FEATURES feature_id);
+
+int av1_is_segfeature_signed(SEG_LVL_FEATURES feature_id);
+
+void av1_set_segdata(struct segmentation *seg, int segment_id,
+                     SEG_LVL_FEATURES feature_id, int seg_data);
+
+static INLINE int get_segdata(const struct segmentation *seg, int segment_id,
+                              SEG_LVL_FEATURES feature_id) {
+  return seg->feature_data[segment_id][feature_id];
+}
+
+extern const aom_tree_index av1_segment_tree[TREE_SIZE(MAX_SEGMENTS)];
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_SEG_COMMON_H_
diff --git a/third_party/aom/av1/common/thread_common.c b/third_party/aom/av1/common/thread_common.c
new file mode 100644
index 0000000000..ca8b1b3bd3
--- /dev/null
+++ b/third_party/aom/av1/common/thread_common.c
@@ -0,0 +1,529 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/thread_common.h"
+#include "av1/common/reconinter.h"
+
+#if CONFIG_MULTITHREAD
+static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
+  const int kMaxTryLocks = 4000;
+  int locked = 0;
+  int i;
+
+  for (i = 0; i < kMaxTryLocks; ++i) {
+    if (!pthread_mutex_trylock(mutex)) {
+      locked = 1;
+      break;
+    }
+  }
+
+  if (!locked) pthread_mutex_lock(mutex);
+}
+#endif  // CONFIG_MULTITHREAD
+
+static INLINE void sync_read(AV1LfSync *const lf_sync, int r, int c) {
+#if CONFIG_MULTITHREAD
+  const int nsync = lf_sync->sync_range;
+
+  if (r && !(c & (nsync - 1))) {
+    pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1];
+    mutex_lock(mutex);
+
+    while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
+      pthread_cond_wait(&lf_sync->cond_[r - 1], mutex);
+    }
+    pthread_mutex_unlock(mutex);
+  }
+#else
+  (void)lf_sync;
+  (void)r;
+  (void)c;
+#endif  // CONFIG_MULTITHREAD
+}
+
+static INLINE void sync_write(AV1LfSync *const lf_sync, int r, int c,
+                              const int sb_cols) {
+#if CONFIG_MULTITHREAD
+  const int nsync = lf_sync->sync_range;
+  int cur;
+  // Only signal when there are enough filtered SB for next row to run.
+  int sig = 1;
+
+  if (c < sb_cols - 1) {
+    cur = c;
+    if (c % nsync) sig = 0;
+  } else {
+    cur = sb_cols + nsync;
+  }
+
+  if (sig) {
+    mutex_lock(&lf_sync->mutex_[r]);
+
+    lf_sync->cur_sb_col[r] = cur;
+
+    pthread_cond_signal(&lf_sync->cond_[r]);
+    pthread_mutex_unlock(&lf_sync->mutex_[r]);
+  }
+#else
+  (void)lf_sync;
+  (void)r;
+  (void)c;
+  (void)sb_cols;
+#endif  // CONFIG_MULTITHREAD
+}
+
+#if !CONFIG_EXT_PARTITION_TYPES
+static INLINE enum lf_path get_loop_filter_path(
+    int y_only, struct macroblockd_plane planes[MAX_MB_PLANE]) {
+  if (y_only)
+    return LF_PATH_444;
+  else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
+    return LF_PATH_420;
+  else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
+    return LF_PATH_444;
+  else
+    return LF_PATH_SLOW;
+}
+
+static INLINE void loop_filter_block_plane_ver(
+    AV1_COMMON *cm, struct macroblockd_plane planes[MAX_MB_PLANE], int plane,
+    MODE_INFO **mi, int mi_row, int mi_col, enum lf_path path,
+    LOOP_FILTER_MASK *lfm) {
+  if (plane == 0) {
+    av1_filter_block_plane_ss00_ver(cm, &planes[0], mi_row, lfm);
+  } else {
+    switch (path) {
+      case LF_PATH_420:
+        av1_filter_block_plane_ss11_ver(cm, &planes[plane], mi_row, lfm);
+        break;
+      case LF_PATH_444:
+        av1_filter_block_plane_ss00_ver(cm, &planes[plane], mi_row, lfm);
+        break;
+      case LF_PATH_SLOW:
+        av1_filter_block_plane_non420_ver(cm, &planes[plane], mi, mi_row,
+                                          mi_col);
+        break;
+    }
+  }
+}
+
+static INLINE void loop_filter_block_plane_hor(
+    AV1_COMMON *cm, struct macroblockd_plane planes[MAX_MB_PLANE], int plane,
+    MODE_INFO **mi, int mi_row, int mi_col, enum lf_path path,
+    LOOP_FILTER_MASK *lfm) {
+  if (plane == 0) {
+    av1_filter_block_plane_ss00_hor(cm, &planes[0], mi_row, lfm);
+  } else {
+    switch (path) {
+      case LF_PATH_420:
+        av1_filter_block_plane_ss11_hor(cm, &planes[plane], mi_row, lfm);
+        break;
+      case LF_PATH_444:
+        av1_filter_block_plane_ss00_hor(cm, &planes[plane], mi_row, lfm);
+        break;
+      case LF_PATH_SLOW:
+        av1_filter_block_plane_non420_hor(cm, &planes[plane], mi, mi_row,
+                                          mi_col);
+        break;
+    }
+  }
+}
+#endif
+// Row-based multi-threaded loopfilter hook
+#if CONFIG_PARALLEL_DEBLOCKING
+static int loop_filter_ver_row_worker(AV1LfSync *const lf_sync,
+                                      LFWorkerData *const lf_data) {
+  const int num_planes = lf_data->y_only ? 1 : MAX_MB_PLANE;
+  int mi_row, mi_col;
+#if !CONFIG_EXT_PARTITION_TYPES
+  enum lf_path path = get_loop_filter_path(lf_data->y_only, lf_data->planes);
+#endif
+  for (mi_row = lf_data->start; mi_row < lf_data->stop;
+       mi_row += lf_sync->num_workers * lf_data->cm->mib_size) {
+    MODE_INFO **const mi =
+        lf_data->cm->mi_grid_visible + mi_row * lf_data->cm->mi_stride;
+
+    for (mi_col = 0; mi_col < lf_data->cm->mi_cols;
+         mi_col += lf_data->cm->mib_size) {
+      LOOP_FILTER_MASK lfm;
+      int plane;
+
+      av1_setup_dst_planes(lf_data->planes, lf_data->cm->sb_size,
+                           lf_data->frame_buffer, mi_row, mi_col);
+      av1_setup_mask(lf_data->cm, mi_row, mi_col, mi + mi_col,
+                     lf_data->cm->mi_stride, &lfm);
+
+#if CONFIG_EXT_PARTITION_TYPES
+      for (plane = 0; plane < num_planes; ++plane)
+        av1_filter_block_plane_non420_ver(lf_data->cm, &lf_data->planes[plane],
+                                          mi + mi_col, mi_row, mi_col);
+#else
+
+      for (plane = 0; plane < num_planes; ++plane)
+        loop_filter_block_plane_ver(lf_data->cm, lf_data->planes, plane,
+                                    mi + mi_col, mi_row, mi_col, path, &lfm);
+#endif
+    }
+  }
+  return 1;
+}
+
+static int loop_filter_hor_row_worker(AV1LfSync *const lf_sync,
+                                      LFWorkerData *const lf_data) {
+  const int num_planes = lf_data->y_only ? 1 : MAX_MB_PLANE;
+  const int sb_cols =
+      mi_cols_aligned_to_sb(lf_data->cm) >> lf_data->cm->mib_size_log2;
+  int mi_row, mi_col;
+#if !CONFIG_EXT_PARTITION_TYPES
+  enum lf_path path = get_loop_filter_path(lf_data->y_only, lf_data->planes);
+#endif
+
+  for (mi_row = lf_data->start; mi_row < lf_data->stop;
+       mi_row += lf_sync->num_workers * lf_data->cm->mib_size) {
+    MODE_INFO **const mi =
+        lf_data->cm->mi_grid_visible + mi_row * lf_data->cm->mi_stride;
+
+    for (mi_col = 0; mi_col < lf_data->cm->mi_cols;
+         mi_col += lf_data->cm->mib_size) {
+      const int r = mi_row >> lf_data->cm->mib_size_log2;
+      const int c = mi_col >> lf_data->cm->mib_size_log2;
+      LOOP_FILTER_MASK lfm;
+      int plane;
+
+      // TODO(wenhao.zhang@intel.com): For better parallelization, reorder
+      // the outer loop to column-based and remove the synchronizations here.
+      sync_read(lf_sync, r, c);
+
+      av1_setup_dst_planes(lf_data->planes, lf_data->cm->sb_size,
+                           lf_data->frame_buffer, mi_row, mi_col);
+      av1_setup_mask(lf_data->cm, mi_row, mi_col, mi + mi_col,
+                     lf_data->cm->mi_stride, &lfm);
+#if CONFIG_EXT_PARTITION_TYPES
+      for (plane = 0; plane < num_planes; ++plane)
+        av1_filter_block_plane_non420_hor(lf_data->cm, &lf_data->planes[plane],
+                                          mi + mi_col, mi_row, mi_col);
+#else
+      for (plane = 0; plane < num_planes; ++plane)
+        loop_filter_block_plane_hor(lf_data->cm, lf_data->planes, plane,
+                                    mi + mi_col, mi_row, mi_col, path, &lfm);
+#endif
+      sync_write(lf_sync, r, c, sb_cols);
+    }
+  }
+  return 1;
+}
+#else  //  CONFIG_PARALLEL_DEBLOCKING
+static int loop_filter_row_worker(AV1LfSync *const lf_sync,
+                                  LFWorkerData *const lf_data) {
+  const int num_planes = lf_data->y_only ? 1 : MAX_MB_PLANE;
+  const int sb_cols =
+      mi_cols_aligned_to_sb(lf_data->cm) >> lf_data->cm->mib_size_log2;
+  int mi_row, mi_col;
+#if !CONFIG_EXT_PARTITION_TYPES
+  enum lf_path path = get_loop_filter_path(lf_data->y_only, lf_data->planes);
+#endif  // !CONFIG_EXT_PARTITION_TYPES
+
+#if CONFIG_EXT_PARTITION
+  printf(
+      "STOPPING: This code has not been modified to work with the "
+      "extended coding unit size experiment");
+  exit(EXIT_FAILURE);
+#endif  // CONFIG_EXT_PARTITION
+
+  for (mi_row = lf_data->start; mi_row < lf_data->stop;
+       mi_row += lf_sync->num_workers * lf_data->cm->mib_size) {
+    MODE_INFO **const mi =
+        lf_data->cm->mi_grid_visible + mi_row * lf_data->cm->mi_stride;
+
+    for (mi_col = 0; mi_col < lf_data->cm->mi_cols;
+         mi_col += lf_data->cm->mib_size) {
+      const int r = mi_row >> lf_data->cm->mib_size_log2;
+      const int c = mi_col >> lf_data->cm->mib_size_log2;
+#if !CONFIG_EXT_PARTITION_TYPES
+      LOOP_FILTER_MASK lfm;
+#endif
+      int plane;
+
+      sync_read(lf_sync, r, c);
+
+      av1_setup_dst_planes(lf_data->planes, lf_data->cm->sb_size,
+                           lf_data->frame_buffer, mi_row, mi_col);
+#if CONFIG_EXT_PARTITION_TYPES
+      for (plane = 0; plane < num_planes; ++plane) {
+        av1_filter_block_plane_non420_ver(lf_data->cm, &lf_data->planes[plane],
+                                          mi + mi_col, mi_row, mi_col);
+        av1_filter_block_plane_non420_hor(lf_data->cm, &lf_data->planes[plane],
+                                          mi + mi_col, mi_row, mi_col);
+      }
+#else
+      av1_setup_mask(lf_data->cm, mi_row, mi_col, mi + mi_col,
+                     lf_data->cm->mi_stride, &lfm);
+
+      for (plane = 0; plane < num_planes; ++plane) {
+        loop_filter_block_plane_ver(lf_data->cm, lf_data->planes, plane,
+                                    mi + mi_col, mi_row, mi_col, path, &lfm);
+        loop_filter_block_plane_hor(lf_data->cm, lf_data->planes, plane,
+                                    mi + mi_col, mi_row, mi_col, path, &lfm);
+      }
+#endif  // CONFIG_EXT_PARTITION_TYPES
+      sync_write(lf_sync, r, c, sb_cols);
+    }
+  }
+  return 1;
+}
+#endif  //  CONFIG_PARALLEL_DEBLOCKING
+
+static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
+                                struct macroblockd_plane planes[MAX_MB_PLANE],
+                                int start, int stop, int y_only,
+                                AVxWorker *workers, int nworkers,
+                                AV1LfSync *lf_sync) {
+  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+  // Number of superblock rows and cols
+  const int sb_rows = mi_rows_aligned_to_sb(cm) >> cm->mib_size_log2;
+  // Decoder may allocate more threads than number of tiles based on user's
+  // input.
+  const int tile_cols = cm->tile_cols;
+  const int num_workers = AOMMIN(nworkers, tile_cols);
+  int i;
+
+#if CONFIG_EXT_PARTITION
+  printf(
+      "STOPPING: This code has not been modified to work with the "
+      "extended coding unit size experiment");
+  exit(EXIT_FAILURE);
+#endif  // CONFIG_EXT_PARTITION
+
+  if (!lf_sync->sync_range || sb_rows != lf_sync->rows ||
+      num_workers > lf_sync->num_workers) {
+    av1_loop_filter_dealloc(lf_sync);
+    av1_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers);
+  }
+
+// Set up loopfilter thread data.
+// The decoder is capping num_workers because it has been observed that using
+// more threads on the loopfilter than there are cores will hurt performance
+// on Android. This is because the system will only schedule the tile decode
+// workers on cores equal to the number of tile columns. Then if the decoder
+// tries to use more threads for the loopfilter, it will hurt performance
+// because of contention. If the multithreading code changes in the future
+// then the number of workers used by the loopfilter should be revisited.
+
+#if CONFIG_PARALLEL_DEBLOCKING
+  // Initialize cur_sb_col to -1 for all SB rows.
+  memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
+
+  // Filter all the vertical edges in the whole frame
+  for (i = 0; i < num_workers; ++i) {
+    AVxWorker *const worker = &workers[i];
+    LFWorkerData *const lf_data = &lf_sync->lfdata[i];
+
+    worker->hook = (AVxWorkerHook)loop_filter_ver_row_worker;
+    worker->data1 = lf_sync;
+    worker->data2 = lf_data;
+
+    // Loopfilter data
+    av1_loop_filter_data_reset(lf_data, frame, cm, planes);
+    lf_data->start = start + i * cm->mib_size;
+    lf_data->stop = stop;
+    lf_data->y_only = y_only;
+
+    // Start loopfiltering
+    if (i == num_workers - 1) {
+      winterface->execute(worker);
+    } else {
+      winterface->launch(worker);
+    }
+  }
+
+  // Wait till all rows are finished
+  for (i = 0; i < num_workers; ++i) {
+    winterface->sync(&workers[i]);
+  }
+
+  memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
+  // Filter all the horizontal edges in the whole frame
+  for (i = 0; i < num_workers; ++i) {
+    AVxWorker *const worker = &workers[i];
+    LFWorkerData *const lf_data = &lf_sync->lfdata[i];
+
+    worker->hook = (AVxWorkerHook)loop_filter_hor_row_worker;
+    worker->data1 = lf_sync;
+    worker->data2 = lf_data;
+
+    // Loopfilter data
+    av1_loop_filter_data_reset(lf_data, frame, cm, planes);
+    lf_data->start = start + i * cm->mib_size;
+    lf_data->stop = stop;
+    lf_data->y_only = y_only;
+
+    // Start loopfiltering
+    if (i == num_workers - 1) {
+      winterface->execute(worker);
+    } else {
+      winterface->launch(worker);
+    }
+  }
+
+  // Wait till all rows are finished
+  for (i = 0; i < num_workers; ++i) {
+    winterface->sync(&workers[i]);
+  }
+#else   // CONFIG_PARALLEL_DEBLOCKING
+  // Initialize cur_sb_col to -1 for all SB rows.
+  memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows);
+
+  for (i = 0; i < num_workers; ++i) {
+    AVxWorker *const worker = &workers[i];
+    LFWorkerData *const lf_data = &lf_sync->lfdata[i];
+
+    worker->hook = (AVxWorkerHook)loop_filter_row_worker;
+    worker->data1 = lf_sync;
+    worker->data2 = lf_data;
+
+    // Loopfilter data
+    av1_loop_filter_data_reset(lf_data, frame, cm, planes);
+    lf_data->start = start + i * cm->mib_size;
+    lf_data->stop = stop;
+    lf_data->y_only = y_only;
+
+    // Start loopfiltering
+    if (i == num_workers - 1) {
+      winterface->execute(worker);
+    } else {
+      winterface->launch(worker);
+    }
+  }
+
+  // Wait till all rows are finished
+  for (i = 0; i < num_workers; ++i) {
+    winterface->sync(&workers[i]);
+  }
+#endif  // CONFIG_PARALLEL_DEBLOCKING
+}
+
+void av1_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame, AV1_COMMON *cm,
+                              struct macroblockd_plane planes[MAX_MB_PLANE],
+                              int frame_filter_level, int y_only,
+                              int partial_frame, AVxWorker *workers,
+                              int num_workers, AV1LfSync *lf_sync) {
+  int start_mi_row, end_mi_row, mi_rows_to_filter;
+
+  if (!frame_filter_level) return;
+
+  start_mi_row = 0;
+  mi_rows_to_filter = cm->mi_rows;
+  if (partial_frame && cm->mi_rows > 8) {
+    start_mi_row = cm->mi_rows >> 1;
+    start_mi_row &= 0xfffffff8;
+    mi_rows_to_filter = AOMMAX(cm->mi_rows / 8, 8);
+  }
+  end_mi_row = start_mi_row + mi_rows_to_filter;
+  av1_loop_filter_frame_init(cm, frame_filter_level);
+
+  loop_filter_rows_mt(frame, cm, planes, start_mi_row, end_mi_row, y_only,
+                      workers, num_workers, lf_sync);
+}
+
+// Set up nsync by width.
+static INLINE int get_sync_range(int width) {
+  // nsync numbers are picked by testing. For example, for 4k
+  // video, using 4 gives best performance.
+  if (width < 640)
+    return 1;
+  else if (width <= 1280)
+    return 2;
+  else if (width <= 4096)
+    return 4;
+  else
+    return 8;
+}
+
+// Allocate memory for lf row synchronization
+void av1_loop_filter_alloc(AV1LfSync *lf_sync, AV1_COMMON *cm, int rows,
+                           int width, int num_workers) {
+  lf_sync->rows = rows;
+#if CONFIG_MULTITHREAD
+  {
+    int i;
+
+    CHECK_MEM_ERROR(cm, lf_sync->mutex_,
+                    aom_malloc(sizeof(*lf_sync->mutex_) * rows));
+    if (lf_sync->mutex_) {
+      for (i = 0; i < rows; ++i) {
+        pthread_mutex_init(&lf_sync->mutex_[i], NULL);
+      }
+    }
+
+    CHECK_MEM_ERROR(cm, lf_sync->cond_,
+                    aom_malloc(sizeof(*lf_sync->cond_) * rows));
+    if (lf_sync->cond_) {
+      for (i = 0; i < rows; ++i) {
+        pthread_cond_init(&lf_sync->cond_[i], NULL);
+      }
+    }
+  }
+#endif  // CONFIG_MULTITHREAD
+
+  CHECK_MEM_ERROR(cm, lf_sync->lfdata,
+                  aom_malloc(num_workers * sizeof(*lf_sync->lfdata)));
+  lf_sync->num_workers = num_workers;
+
+  CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
+                  aom_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
+
+  // Set up nsync.
+  lf_sync->sync_range = get_sync_range(width);
+}
+
+// Deallocate lf synchronization related mutex and data
+void av1_loop_filter_dealloc(AV1LfSync *lf_sync) {
+  if (lf_sync != NULL) {
+#if CONFIG_MULTITHREAD
+    int i;
+
+    if (lf_sync->mutex_ != NULL) {
+      for (i = 0; i < lf_sync->rows; ++i) {
+        pthread_mutex_destroy(&lf_sync->mutex_[i]);
+      }
+      aom_free(lf_sync->mutex_);
+    }
+    if (lf_sync->cond_ != NULL) {
+      for (i = 0; i < lf_sync->rows; ++i) {
+        pthread_cond_destroy(&lf_sync->cond_[i]);
+      }
+      aom_free(lf_sync->cond_);
+    }
+#endif  // CONFIG_MULTITHREAD
+    aom_free(lf_sync->lfdata);
+    aom_free(lf_sync->cur_sb_col);
+    // clear the structure as the source of this call may be a resize in which
+    // case this call will be followed by an _alloc() which may fail.
+    av1_zero(*lf_sync);
+  }
+}
+
+// Accumulate frame counts. FRAME_COUNTS consist solely of 'unsigned int'
+// members, so we treat it as an array, and sum over the whole length.
+void av1_accumulate_frame_counts(FRAME_COUNTS *acc_counts,
+                                 FRAME_COUNTS *counts) {
+  unsigned int *const acc = (unsigned int *)acc_counts;
+  const unsigned int *const cnt = (unsigned int *)counts;
+
+  const unsigned int n_counts = sizeof(FRAME_COUNTS) / sizeof(unsigned int);
+  unsigned int i;
+
+  for (i = 0; i < n_counts; i++) acc[i] += cnt[i];
+}
diff --git a/third_party/aom/av1/common/thread_common.h b/third_party/aom/av1/common/thread_common.h
new file mode 100644
index 0000000000..7b57ae8f34
--- /dev/null
+++ b/third_party/aom/av1/common/thread_common.h
@@ -0,0 +1,64 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_LOOPFILTER_THREAD_H_
+#define AV1_COMMON_LOOPFILTER_THREAD_H_
+#include "./aom_config.h"
+#include "av1/common/av1_loopfilter.h"
+#include "aom_util/aom_thread.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1Common;
+struct FRAME_COUNTS;
+
+// Loopfilter row synchronization
+typedef struct AV1LfSyncData {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t *mutex_;
+  pthread_cond_t *cond_;
+#endif
+  // Allocate memory to store the loop-filtered superblock index in each row.
+  int *cur_sb_col;
+  // The optimal sync_range for different resolution and platform should be
+  // determined by testing. Currently, it is chosen to be a power-of-2 number.
+  int sync_range;
+  int rows;
+
+  // Row-based parallel loopfilter data
+  LFWorkerData *lfdata;
+  int num_workers;
+} AV1LfSync;
+
+// Allocate memory for loopfilter row synchronization.
+void av1_loop_filter_alloc(AV1LfSync *lf_sync, struct AV1Common *cm, int rows,
+                           int width, int num_workers);
+
+// Deallocate loopfilter synchronization related mutex and data.
+void av1_loop_filter_dealloc(AV1LfSync *lf_sync);
+
+// Multi-threaded loopfilter that uses the tile threads.
+void av1_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame, struct AV1Common *cm,
+                              struct macroblockd_plane planes[MAX_MB_PLANE],
+                              int frame_filter_level, int y_only,
+                              int partial_frame, AVxWorker *workers,
+                              int num_workers, AV1LfSync *lf_sync);
+
+void av1_accumulate_frame_counts(struct FRAME_COUNTS *acc_counts,
+                                 struct FRAME_COUNTS *counts);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_LOOPFILTER_THREAD_H_
diff --git a/third_party/aom/av1/common/tile_common.c b/third_party/aom/av1/common/tile_common.c
new file mode 100644
index 0000000000..b8008ac2e0
--- /dev/null
+++ b/third_party/aom/av1/common/tile_common.c
@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/tile_common.h"
+#include "av1/common/onyxc_int.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+void av1_tile_set_row(TileInfo *tile, const AV1_COMMON *cm, int row) {
+  tile->mi_row_start = row * cm->tile_height;
+  tile->mi_row_end = AOMMIN(tile->mi_row_start + cm->tile_height, cm->mi_rows);
+}
+
+void av1_tile_set_col(TileInfo *tile, const AV1_COMMON *cm, int col) {
+  tile->mi_col_start = col * cm->tile_width;
+  tile->mi_col_end = AOMMIN(tile->mi_col_start + cm->tile_width, cm->mi_cols);
+}
+
+#if CONFIG_DEPENDENT_HORZTILES && CONFIG_TILE_GROUPS
+void av1_tile_set_tg_boundary(TileInfo *tile, const AV1_COMMON *const cm,
+                              int row, int col) {
+  if (row < cm->tile_rows - 1) {
+    tile->tg_horz_boundary =
+        col >= cm->tile_group_start_col[row][col]
+            ? (row == cm->tile_group_start_row[row][col] ? 1 : 0)
+            : (row == cm->tile_group_start_row[row + 1][col] ? 1 : 0);
+  } else {
+    assert(col >= cm->tile_group_start_col[row][col]);
+    tile->tg_horz_boundary =
+        (row == cm->tile_group_start_row[row][col] ? 1 : 0);
+  }
+}
+#endif
+void av1_tile_init(TileInfo *tile, const AV1_COMMON *cm, int row, int col) {
+  av1_tile_set_row(tile, cm, row);
+  av1_tile_set_col(tile, cm, col);
+#if CONFIG_DEPENDENT_HORZTILES && CONFIG_TILE_GROUPS
+  av1_tile_set_tg_boundary(tile, cm, row, col);
+#endif
+}
+
+#if !CONFIG_EXT_TILE
+
+#if CONFIG_EXT_PARTITION
+#define MIN_TILE_WIDTH_MAX_SB 2
+#define MAX_TILE_WIDTH_MAX_SB 32
+#else
+#define MIN_TILE_WIDTH_MAX_SB 4
+#define MAX_TILE_WIDTH_MAX_SB 64
+#endif  // CONFIG_EXT_PARTITION
+
+static int get_min_log2_tile_cols(int max_sb_cols) {
+  int min_log2 = 0;
+  while ((MAX_TILE_WIDTH_MAX_SB << min_log2) < max_sb_cols) ++min_log2;
+  return min_log2;
+}
+
+static int get_max_log2_tile_cols(int max_sb_cols) {
+  int max_log2 = 1;
+  while ((max_sb_cols >> max_log2) >= MIN_TILE_WIDTH_MAX_SB) ++max_log2;
+  return max_log2 - 1;
+}
+
+void av1_get_tile_n_bits(int mi_cols, int *min_log2_tile_cols,
+                         int *max_log2_tile_cols) {
+  const int max_sb_cols =
+      ALIGN_POWER_OF_TWO(mi_cols, MAX_MIB_SIZE_LOG2) >> MAX_MIB_SIZE_LOG2;
+  *min_log2_tile_cols = get_min_log2_tile_cols(max_sb_cols);
+  *max_log2_tile_cols = get_max_log2_tile_cols(max_sb_cols);
+  assert(*min_log2_tile_cols <= *max_log2_tile_cols);
+}
+#endif  // !CONFIG_EXT_TILE
+
+void av1_update_boundary_info(const struct AV1Common *cm,
+                              const TileInfo *const tile_info, int mi_row,
+                              int mi_col) {
+  int row, col;
+  for (row = mi_row; row < (mi_row + cm->mib_size); row++)
+    for (col = mi_col; col < (mi_col + cm->mib_size); col++) {
+      MODE_INFO *const mi = cm->mi + row * cm->mi_stride + col;
+      mi->mbmi.boundary_info = 0;
+      if (cm->tile_cols * cm->tile_rows > 1) {
+#if CONFIG_DEPENDENT_HORZTILES
+        if (row == tile_info->mi_row_start &&
+            (!cm->dependent_horz_tiles || tile_info->tg_horz_boundary))
+#if CONFIG_TILE_GROUPS
+#else
+          if (row == tile_info->mi_row_start && !cm->dependent_horz_tiles)
+#endif  // CONFIG_TILE_GROUPS
+#else
+        if (row == tile_info->mi_row_start)
+#endif  // CONFIG_DEPENDENT_HORZTILES
+          mi->mbmi.boundary_info |= TILE_ABOVE_BOUNDARY;
+        if (col == tile_info->mi_col_start)
+          mi->mbmi.boundary_info |= TILE_LEFT_BOUNDARY;
+        if ((row + 1) >= tile_info->mi_row_end)
+          mi->mbmi.boundary_info |= TILE_BOTTOM_BOUNDARY;
+        if ((col + 1) >= tile_info->mi_col_end)
+          mi->mbmi.boundary_info |= TILE_RIGHT_BOUNDARY;
+      }
+      // Frame boundary is treated as tile boundary
+      if (row == 0)
+        mi->mbmi.boundary_info |= FRAME_ABOVE_BOUNDARY | TILE_ABOVE_BOUNDARY;
+      if (col == 0)
+        mi->mbmi.boundary_info |= FRAME_LEFT_BOUNDARY | TILE_LEFT_BOUNDARY;
+      if ((row + 1) >= cm->mi_rows)
+        mi->mbmi.boundary_info |= FRAME_BOTTOM_BOUNDARY | TILE_BOTTOM_BOUNDARY;
+      if ((col + 1) >= cm->mi_cols)
+        mi->mbmi.boundary_info |= FRAME_RIGHT_BOUNDARY | TILE_RIGHT_BOUNDARY;
+    }
+}
+
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+int av1_disable_loopfilter_on_tile_boundary(const struct AV1Common *cm) {
+  return (!cm->loop_filter_across_tiles_enabled &&
+          (cm->tile_cols * cm->tile_rows > 1));
+}
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
diff --git a/third_party/aom/av1/common/tile_common.h b/third_party/aom/av1/common/tile_common.h
new file mode 100644
index 0000000000..617dda2028
--- /dev/null
+++ b/third_party/aom/av1/common/tile_common.h
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_TILE_COMMON_H_
+#define AV1_COMMON_TILE_COMMON_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "./aom_config.h"
+
+struct AV1Common;
+
+#if CONFIG_TILE_GROUPS
+#define DEFAULT_MAX_NUM_TG 1
+#endif
+
+typedef struct TileInfo {
+  int mi_row_start, mi_row_end;
+  int mi_col_start, mi_col_end;
+  int tg_horz_boundary;
+} TileInfo;
+
+// initializes 'tile->mi_(row|col)_(start|end)' for (row, col) based on
+// 'cm->log2_tile_(rows|cols)' & 'cm->mi_(rows|cols)'
+void av1_tile_init(TileInfo *tile, const struct AV1Common *cm, int row,
+                   int col);
+
+void av1_tile_set_row(TileInfo *tile, const struct AV1Common *cm, int row);
+void av1_tile_set_col(TileInfo *tile, const struct AV1Common *cm, int col);
+#if CONFIG_DEPENDENT_HORZTILES && CONFIG_TILE_GROUPS
+void av1_tile_set_tg_boundary(TileInfo *tile, const struct AV1Common *cm,
+                              int row, int col);
+#endif
+void av1_get_tile_n_bits(int mi_cols, int *min_log2_tile_cols,
+                         int *max_log2_tile_cols);
+
+void av1_update_boundary_info(const struct AV1Common *cm,
+                              const TileInfo *const tile_info, int mi_row,
+                              int mi_col);
+
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+int av1_disable_loopfilter_on_tile_boundary(const struct AV1Common *cm);
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_COMMON_TILE_COMMON_H_
diff --git a/third_party/aom/av1/common/txb_common.c b/third_party/aom/av1/common/txb_common.c
new file mode 100644
index 0000000000..08a685b597
--- /dev/null
+++ b/third_party/aom/av1/common/txb_common.c
@@ -0,0 +1,149 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "aom/aom_integer.h"
+#include "av1/common/onyxc_int.h"
+
+const int16_t av1_coeff_band_4x4[16] = { 0, 1, 2,  3,  4,  5,  6,  7,
+                                         8, 9, 10, 11, 12, 13, 14, 15 };
+
+const int16_t av1_coeff_band_8x8[64] = {
+  0,  1,  2,  2,  3,  3,  4,  4,  5,  6,  2,  2,  3,  3,  4,  4,
+  7,  7,  8,  8,  9,  9,  10, 10, 7,  7,  8,  8,  9,  9,  10, 10,
+  11, 11, 12, 12, 13, 13, 14, 14, 11, 11, 12, 12, 13, 13, 14, 14,
+  15, 15, 16, 16, 17, 17, 18, 18, 15, 15, 16, 16, 17, 17, 18, 18,
+};
+
+const int16_t av1_coeff_band_16x16[256] = {
+  0,  1,  4,  4,  7,  7,  7,  7,  8,  8,  8,  8,  9,  9,  9,  9,  2,  3,  4,
+  4,  7,  7,  7,  7,  8,  8,  8,  8,  9,  9,  9,  9,  5,  5,  6,  6,  7,  7,
+  7,  7,  8,  8,  8,  8,  9,  9,  9,  9,  5,  5,  6,  6,  7,  7,  7,  7,  8,
+  8,  8,  8,  9,  9,  9,  9,  10, 10, 10, 10, 11, 11, 11, 11, 12, 12, 12, 12,
+  13, 13, 13, 13, 10, 10, 10, 10, 11, 11, 11, 11, 12, 12, 12, 12, 13, 13, 13,
+  13, 10, 10, 10, 10, 11, 11, 11, 11, 12, 12, 12, 12, 13, 13, 13, 13, 10, 10,
+  10, 10, 11, 11, 11, 11, 12, 12, 12, 12, 13, 13, 13, 13, 14, 14, 14, 14, 15,
+  15, 15, 15, 16, 16, 16, 16, 17, 17, 17, 17, 14, 14, 14, 14, 15, 15, 15, 15,
+  16, 16, 16, 16, 17, 17, 17, 17, 14, 14, 14, 14, 15, 15, 15, 15, 16, 16, 16,
+  16, 17, 17, 17, 17, 14, 14, 14, 14, 15, 15, 15, 15, 16, 16, 16, 16, 17, 17,
+  17, 17, 18, 18, 18, 18, 19, 19, 19, 19, 20, 20, 20, 20, 21, 21, 21, 21, 18,
+  18, 18, 18, 19, 19, 19, 19, 20, 20, 20, 20, 21, 21, 21, 21, 18, 18, 18, 18,
+  19, 19, 19, 19, 20, 20, 20, 20, 21, 21, 21, 21, 18, 18, 18, 18, 19, 19, 19,
+  19, 20, 20, 20, 20, 21, 21, 21, 21,
+};
+
+const int16_t av1_coeff_band_32x32[1024] = {
+  0,  1,  4,  4,  7,  7,  7,  7,  10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11,
+  11, 11, 11, 11, 11, 12, 12, 12, 12, 12, 12, 12, 12, 2,  3,  4,  4,  7,  7,
+  7,  7,  10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11, 11, 11, 11, 11, 11, 12,
+  12, 12, 12, 12, 12, 12, 12, 5,  5,  6,  6,  7,  7,  7,  7,  10, 10, 10, 10,
+  10, 10, 10, 10, 11, 11, 11, 11, 11, 11, 11, 11, 12, 12, 12, 12, 12, 12, 12,
+  12, 5,  5,  6,  6,  7,  7,  7,  7,  10, 10, 10, 10, 10, 10, 10, 10, 11, 11,
+  11, 11, 11, 11, 11, 11, 12, 12, 12, 12, 12, 12, 12, 12, 8,  8,  8,  8,  9,
+  9,  9,  9,  10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11, 11, 11, 11, 11, 11,
+  12, 12, 12, 12, 12, 12, 12, 12, 8,  8,  8,  8,  9,  9,  9,  9,  10, 10, 10,
+  10, 10, 10, 10, 10, 11, 11, 11, 11, 11, 11, 11, 11, 12, 12, 12, 12, 12, 12,
+  12, 12, 8,  8,  8,  8,  9,  9,  9,  9,  10, 10, 10, 10, 10, 10, 10, 10, 11,
+  11, 11, 11, 11, 11, 11, 11, 12, 12, 12, 12, 12, 12, 12, 12, 8,  8,  8,  8,
+  9,  9,  9,  9,  10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11, 11, 11, 11, 11,
+  11, 12, 12, 12, 12, 12, 12, 12, 12, 13, 13, 13, 13, 13, 13, 13, 13, 14, 14,
+  14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 16, 16, 16, 16, 16,
+  16, 16, 16, 13, 13, 13, 13, 13, 13, 13, 13, 14, 14, 14, 14, 14, 14, 14, 14,
+  15, 15, 15, 15, 15, 15, 15, 15, 16, 16, 16, 16, 16, 16, 16, 16, 13, 13, 13,
+  13, 13, 13, 13, 13, 14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15,
+  15, 15, 16, 16, 16, 16, 16, 16, 16, 16, 13, 13, 13, 13, 13, 13, 13, 13, 14,
+  14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 16, 16, 16, 16,
+  16, 16, 16, 16, 13, 13, 13, 13, 13, 13, 13, 13, 14, 14, 14, 14, 14, 14, 14,
+  14, 15, 15, 15, 15, 15, 15, 15, 15, 16, 16, 16, 16, 16, 16, 16, 16, 13, 13,
+  13, 13, 13, 13, 13, 13, 14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15,
+  15, 15, 15, 16, 16, 16, 16, 16, 16, 16, 16, 13, 13, 13, 13, 13, 13, 13, 13,
+  14, 14, 14, 14, 14, 14, 14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 16, 16, 16,
+  16, 16, 16, 16, 16, 13, 13, 13, 13, 13, 13, 13, 13, 14, 14, 14, 14, 14, 14,
+  14, 14, 15, 15, 15, 15, 15, 15, 15, 15, 16, 16, 16, 16, 16, 16, 16, 16, 17,
+  17, 17, 17, 17, 17, 17, 17, 18, 18, 18, 18, 18, 18, 18, 18, 19, 19, 19, 19,
+  19, 19, 19, 19, 20, 20, 20, 20, 20, 20, 20, 20, 17, 17, 17, 17, 17, 17, 17,
+  17, 18, 18, 18, 18, 18, 18, 18, 18, 19, 19, 19, 19, 19, 19, 19, 19, 20, 20,
+  20, 20, 20, 20, 20, 20, 17, 17, 17, 17, 17, 17, 17, 17, 18, 18, 18, 18, 18,
+  18, 18, 18, 19, 19, 19, 19, 19, 19, 19, 19, 20, 20, 20, 20, 20, 20, 20, 20,
+  17, 17, 17, 17, 17, 17, 17, 17, 18, 18, 18, 18, 18, 18, 18, 18, 19, 19, 19,
+  19, 19, 19, 19, 19, 20, 20, 20, 20, 20, 20, 20, 20, 17, 17, 17, 17, 17, 17,
+  17, 17, 18, 18, 18, 18, 18, 18, 18, 18, 19, 19, 19, 19, 19, 19, 19, 19, 20,
+  20, 20, 20, 20, 20, 20, 20, 17, 17, 17, 17, 17, 17, 17, 17, 18, 18, 18, 18,
+  18, 18, 18, 18, 19, 19, 19, 19, 19, 19, 19, 19, 20, 20, 20, 20, 20, 20, 20,
+  20, 17, 17, 17, 17, 17, 17, 17, 17, 18, 18, 18, 18, 18, 18, 18, 18, 19, 19,
+  19, 19, 19, 19, 19, 19, 20, 20, 20, 20, 20, 20, 20, 20, 17, 17, 17, 17, 17,
+  17, 17, 17, 18, 18, 18, 18, 18, 18, 18, 18, 19, 19, 19, 19, 19, 19, 19, 19,
+  20, 20, 20, 20, 20, 20, 20, 20, 21, 21, 21, 21, 21, 21, 21, 21, 22, 22, 22,
+  22, 22, 22, 22, 22, 23, 23, 23, 23, 23, 23, 23, 23, 24, 24, 24, 24, 24, 24,
+  24, 24, 21, 21, 21, 21, 21, 21, 21, 21, 22, 22, 22, 22, 22, 22, 22, 22, 23,
+  23, 23, 23, 23, 23, 23, 23, 24, 24, 24, 24, 24, 24, 24, 24, 21, 21, 21, 21,
+  21, 21, 21, 21, 22, 22, 22, 22, 22, 22, 22, 22, 23, 23, 23, 23, 23, 23, 23,
+  23, 24, 24, 24, 24, 24, 24, 24, 24, 21, 21, 21, 21, 21, 21, 21, 21, 22, 22,
+  22, 22, 22, 22, 22, 22, 23, 23, 23, 23, 23, 23, 23, 23, 24, 24, 24, 24, 24,
+  24, 24, 24, 21, 21, 21, 21, 21, 21, 21, 21, 22, 22, 22, 22, 22, 22, 22, 22,
+  23, 23, 23, 23, 23, 23, 23, 23, 24, 24, 24, 24, 24, 24, 24, 24, 21, 21, 21,
+  21, 21, 21, 21, 21, 22, 22, 22, 22, 22, 22, 22, 22, 23, 23, 23, 23, 23, 23,
+  23, 23, 24, 24, 24, 24, 24, 24, 24, 24, 21, 21, 21, 21, 21, 21, 21, 21, 22,
+  22, 22, 22, 22, 22, 22, 22, 23, 23, 23, 23, 23, 23, 23, 23, 24, 24, 24, 24,
+  24, 24, 24, 24, 21, 21, 21, 21, 21, 21, 21, 21, 22, 22, 22, 22, 22, 22, 22,
+  22, 23, 23, 23, 23, 23, 23, 23, 23, 24, 24, 24, 24, 24, 24, 24, 24,
+};
+
+void av1_adapt_txb_probs(AV1_COMMON *cm, unsigned int count_sat,
+                         unsigned int update_factor) {
+  FRAME_CONTEXT *fc = cm->fc;
+  const FRAME_CONTEXT *pre_fc = &cm->frame_contexts[cm->frame_context_idx];
+  const FRAME_COUNTS *counts = &cm->counts;
+  TX_SIZE tx_size;
+  int plane, ctx, level;
+
+  // Update probability models for transform block skip flag
+  for (tx_size = 0; tx_size < TX_SIZES; ++tx_size)
+    for (ctx = 0; ctx < TXB_SKIP_CONTEXTS; ++ctx)
+      fc->txb_skip[tx_size][ctx] = mode_mv_merge_probs(
+          pre_fc->txb_skip[tx_size][ctx], counts->txb_skip[tx_size][ctx]);
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < DC_SIGN_CONTEXTS; ++ctx)
+      fc->dc_sign[plane][ctx] = mode_mv_merge_probs(
+          pre_fc->dc_sign[plane][ctx], counts->dc_sign[plane][ctx]);
+  // Update probability models for non-zero coefficient map and eob flag.
+  for (level = 0; level < NUM_BASE_LEVELS; ++level)
+    for (tx_size = 0; tx_size < TX_SIZES; ++tx_size)
+      for (plane = 0; plane < PLANE_TYPES; ++plane)
+        for (ctx = 0; ctx < COEFF_BASE_CONTEXTS; ++ctx)
+          fc->coeff_base[tx_size][plane][level][ctx] =
+              merge_probs(pre_fc->coeff_base[tx_size][plane][level][ctx],
+                          counts->coeff_base[tx_size][plane][level][ctx],
+                          count_sat, update_factor);
+
+  for (tx_size = 0; tx_size < TX_SIZES; ++tx_size) {
+    for (plane = 0; plane < PLANE_TYPES; ++plane) {
+      for (ctx = 0; ctx < SIG_COEF_CONTEXTS; ++ctx) {
+        fc->nz_map[tx_size][plane][ctx] = merge_probs(
+            pre_fc->nz_map[tx_size][plane][ctx],
+            counts->nz_map[tx_size][plane][ctx], count_sat, update_factor);
+      }
+
+      for (ctx = 0; ctx < EOB_COEF_CONTEXTS; ++ctx) {
+        fc->eob_flag[tx_size][plane][ctx] = merge_probs(
+            pre_fc->eob_flag[tx_size][plane][ctx],
+            counts->eob_flag[tx_size][plane][ctx], count_sat, update_factor);
+      }
+    }
+  }
+
+  for (tx_size = 0; tx_size < TX_SIZES; ++tx_size) {
+    for (plane = 0; plane < PLANE_TYPES; ++plane)
+      for (ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx)
+        fc->coeff_lps[tx_size][plane][ctx] = merge_probs(
+            pre_fc->coeff_lps[tx_size][plane][ctx],
+            counts->coeff_lps[tx_size][plane][ctx], count_sat, update_factor);
+  }
+}
diff --git a/third_party/aom/av1/common/txb_common.h b/third_party/aom/av1/common/txb_common.h
new file mode 100644
index 0000000000..cdd9ca26eb
--- /dev/null
+++ b/third_party/aom/av1/common/txb_common.h
@@ -0,0 +1,304 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_TXB_COMMON_H_
+#define AV1_COMMON_TXB_COMMON_H_
+extern const int16_t av1_coeff_band_4x4[16];
+
+extern const int16_t av1_coeff_band_8x8[64];
+
+extern const int16_t av1_coeff_band_16x16[256];
+
+extern const int16_t av1_coeff_band_32x32[1024];
+
+typedef struct txb_ctx {
+  int txb_skip_ctx;
+  int dc_sign_ctx;
+} TXB_CTX;
+
+#define BASE_CONTEXT_POSITION_NUM 12
+static int base_ref_offset[BASE_CONTEXT_POSITION_NUM][2] = {
+  /* clang-format off*/
+  { -2, 0 }, { -1, -1 }, { -1, 0 }, { -1, 1 }, { 0, -2 }, { 0, -1 }, { 0, 1 },
+  { 0, 2 },  { 1, -1 },  { 1, 0 },  { 1, 1 },  { 2, 0 }
+  /* clang-format on*/
+};
+
+static INLINE int get_base_ctx(const tran_low_t *tcoeffs,
+                               int c,  // raster order
+                               const int bwl, const int level) {
+  const int row = c >> bwl;
+  const int col = c - (row << bwl);
+  const int stride = 1 << bwl;
+  const int level_minus_1 = level - 1;
+  int ctx = 0;
+  int mag = 0;
+  int idx;
+  int ctx_idx = -1;
+  tran_low_t abs_coeff;
+
+  ctx = 0;
+  for (idx = 0; idx < BASE_CONTEXT_POSITION_NUM; ++idx) {
+    int ref_row = row + base_ref_offset[idx][0];
+    int ref_col = col + base_ref_offset[idx][1];
+    int pos = (ref_row << bwl) + ref_col;
+
+    if (ref_row < 0 || ref_col < 0 || ref_row >= stride || ref_col >= stride)
+      continue;
+
+    abs_coeff = abs(tcoeffs[pos]);
+    ctx += abs_coeff > level_minus_1;
+
+    if (base_ref_offset[idx][0] >= 0 && base_ref_offset[idx][1] >= 0)
+      mag |= abs_coeff > level;
+  }
+  ctx = (ctx + 1) >> 1;
+  if (row == 0 && col == 0) {
+    ctx_idx = (ctx << 1) + mag;
+    assert(ctx_idx < 8);
+  } else if (row == 0) {
+    ctx_idx = 8 + (ctx << 1) + mag;
+    assert(ctx_idx < 18);
+  } else if (col == 0) {
+    ctx_idx = 8 + 10 + (ctx << 1) + mag;
+    assert(ctx_idx < 28);
+  } else {
+    ctx_idx = 8 + 10 + 10 + (ctx << 1) + mag;
+    assert(ctx_idx < COEFF_BASE_CONTEXTS);
+  }
+  return ctx_idx;
+}
+
+#define BR_CONTEXT_POSITION_NUM 8  // Base range coefficient context
+static int br_ref_offset[BR_CONTEXT_POSITION_NUM][2] = {
+  /* clang-format off*/
+  { -1, -1 }, { -1, 0 }, { -1, 1 }, { 0, -1 },
+  { 0, 1 },   { 1, -1 }, { 1, 0 },  { 1, 1 },
+  /* clang-format on*/
+};
+
+static int br_level_map[9] = {
+  0, 0, 1, 1, 2, 2, 3, 3, 3,
+};
+
+static INLINE int get_level_ctx(const tran_low_t *tcoeffs,
+                                const int c,  // raster order
+                                const int bwl) {
+  const int row = c >> bwl;
+  const int col = c - (row << bwl);
+  const int stride = 1 << bwl;
+  const int level_minus_1 = NUM_BASE_LEVELS;
+  int ctx = 0;
+  int idx;
+  tran_low_t abs_coeff;
+  int mag = 0, offset = 0;
+
+  for (idx = 0; idx < BR_CONTEXT_POSITION_NUM; ++idx) {
+    int ref_row = row + br_ref_offset[idx][0];
+    int ref_col = col + br_ref_offset[idx][1];
+    int pos = (ref_row << bwl) + ref_col;
+
+    if (ref_row < 0 || ref_col < 0 || ref_row >= stride || ref_col >= stride)
+      continue;
+
+    abs_coeff = abs(tcoeffs[pos]);
+    ctx += abs_coeff > level_minus_1;
+
+    if (br_ref_offset[idx][0] >= 0 && br_ref_offset[idx][1] >= 0)
+      mag = AOMMAX(mag, abs_coeff);
+  }
+
+  if (mag <= 1)
+    offset = 0;
+  else if (mag <= 3)
+    offset = 1;
+  else if (mag <= 6)
+    offset = 2;
+  else
+    offset = 3;
+
+  ctx = br_level_map[ctx];
+
+  ctx += offset * BR_TMP_OFFSET;
+
+  // DC: 0 - 1
+  if (row == 0 && col == 0) return ctx;
+
+  // Top row: 2 - 4
+  if (row == 0) return 2 + ctx;
+
+  // Left column: 5 - 7
+  if (col == 0) return 5 + ctx;
+
+  // others: 8 - 11
+  return 8 + ctx;
+}
+
+static int sig_ref_offset[11][2] = {
+  { -2, -1 }, { -2, 0 }, { -2, 1 }, { -1, -2 }, { -1, -1 }, { -1, 0 },
+  { -1, 1 },  { 0, -2 }, { 0, -1 }, { 1, -2 },  { 1, -1 },
+};
+
+static INLINE int get_nz_map_ctx(const tran_low_t *tcoeffs,
+                                 const uint8_t *txb_mask,
+                                 const int c,  // raster order
+                                 const int bwl) {
+  const int row = c >> bwl;
+  const int col = c - (row << bwl);
+  int ctx = 0;
+  int idx;
+  int stride = 1 << bwl;
+
+  if (row == 0 && col == 0) return 0;
+
+  if (row == 0 && col == 1) return 1 + (tcoeffs[0] != 0);
+
+  if (row == 1 && col == 0) return 3 + (tcoeffs[0] != 0);
+
+  if (row == 1 && col == 1) {
+    int pos;
+    ctx = (tcoeffs[0] != 0);
+
+    if (txb_mask[1]) ctx += (tcoeffs[1] != 0);
+    pos = 1 << bwl;
+    if (txb_mask[pos]) ctx += (tcoeffs[pos] != 0);
+
+    ctx = (ctx + 1) >> 1;
+
+    assert(5 + ctx <= 7);
+
+    return 5 + ctx;
+  }
+
+  for (idx = 0; idx < 11; ++idx) {
+    int ref_row = row + sig_ref_offset[idx][0];
+    int ref_col = col + sig_ref_offset[idx][1];
+    int pos;
+
+    if (ref_row < 0 || ref_col < 0 || ref_row >= stride || ref_col >= stride)
+      continue;
+
+    pos = (ref_row << bwl) + ref_col;
+
+    if (txb_mask[pos]) ctx += (tcoeffs[pos] != 0);
+  }
+
+  if (row == 0) {
+    ctx = (ctx + 1) >> 1;
+
+    assert(ctx < 3);
+    return 8 + ctx;
+  }
+
+  if (col == 0) {
+    ctx = (ctx + 1) >> 1;
+
+    assert(ctx < 3);
+    return 11 + ctx;
+  }
+
+  ctx >>= 1;
+
+  assert(14 + ctx < 20);
+
+  return 14 + ctx;
+}
+
+static INLINE int get_eob_ctx(const tran_low_t *tcoeffs,
+                              const int c,  // raster order
+                              const int bwl) {
+  (void)tcoeffs;
+  if (bwl == 2) return av1_coeff_band_4x4[c];
+  if (bwl == 3) return av1_coeff_band_8x8[c];
+  if (bwl == 4) return av1_coeff_band_16x16[c];
+  if (bwl == 5) return av1_coeff_band_32x32[c];
+
+  assert(0);
+  return 0;
+}
+
+static INLINE void set_dc_sign(int *cul_level, tran_low_t v) {
+  if (v < 0)
+    *cul_level |= 1 << COEFF_CONTEXT_BITS;
+  else if (v > 0)
+    *cul_level += 2 << COEFF_CONTEXT_BITS;
+}
+
+static INLINE int get_dc_sign_ctx(int dc_sign) {
+  int dc_sign_ctx = 0;
+  if (dc_sign < 0)
+    dc_sign_ctx = 1;
+  else if (dc_sign > 0)
+    dc_sign_ctx = 2;
+
+  return dc_sign_ctx;
+}
+
+static INLINE void get_txb_ctx(BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                               int plane, const ENTROPY_CONTEXT *a,
+                               const ENTROPY_CONTEXT *l, TXB_CTX *txb_ctx) {
+  const int tx_size_in_blocks = 1 << tx_size;
+  int ctx_offset = (plane == 0) ? 0 : 7;
+  int k;
+
+  if (plane_bsize > txsize_to_bsize[tx_size]) ctx_offset += 3;
+
+  int dc_sign = 0;
+  for (k = 0; k < tx_size_in_blocks; ++k) {
+    int sign = ((uint8_t)a[k]) >> COEFF_CONTEXT_BITS;
+    if (sign == 1)
+      --dc_sign;
+    else if (sign == 2)
+      ++dc_sign;
+    else if (sign != 0)
+      assert(0);
+
+    sign = ((uint8_t)l[k]) >> COEFF_CONTEXT_BITS;
+    if (sign == 1)
+      --dc_sign;
+    else if (sign == 2)
+      ++dc_sign;
+    else if (sign != 0)
+      assert(0);
+  }
+  txb_ctx->dc_sign_ctx = get_dc_sign_ctx(dc_sign);
+
+  if (plane == 0) {
+    int top = 0;
+    int left = 0;
+    for (k = 0; k < tx_size_in_blocks; ++k) {
+      top = AOMMAX(top, ((uint8_t)a[k] & COEFF_CONTEXT_MASK));
+      left = AOMMAX(left, ((uint8_t)l[k] & COEFF_CONTEXT_MASK));
+    }
+    top = AOMMIN(top, 255);
+    left = AOMMIN(left, 255);
+
+    if (plane_bsize == txsize_to_bsize[tx_size])
+      txb_ctx->txb_skip_ctx = 0;
+    else if (top == 0 && left == 0)
+      txb_ctx->txb_skip_ctx = 1;
+    else if (top == 0 || left == 0)
+      txb_ctx->txb_skip_ctx = 2 + (AOMMAX(top, left) > 3);
+    else if (AOMMAX(top, left) <= 3)
+      txb_ctx->txb_skip_ctx = 4;
+    else if (AOMMIN(top, left) <= 3)
+      txb_ctx->txb_skip_ctx = 5;
+    else
+      txb_ctx->txb_skip_ctx = 6;
+  } else {
+    int ctx_base = get_entropy_context(tx_size, a, l);
+    txb_ctx->txb_skip_ctx = ctx_offset + ctx_base;
+  }
+}
+
+void av1_adapt_txb_probs(AV1_COMMON *cm, unsigned int count_sat,
+                         unsigned int update_factor);
+#endif  // AV1_COMMON_TXB_COMMON_H_
diff --git a/third_party/aom/av1/common/warped_motion.c b/third_party/aom/av1/common/warped_motion.c
new file mode 100644
index 0000000000..9d13dc705a
--- /dev/null
+++ b/third_party/aom/av1/common/warped_motion.c
@@ -0,0 +1,1773 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+#include <math.h>
+#include <assert.h>
+
+#include "./av1_rtcd.h"
+#include "av1/common/warped_motion.h"
+
+/* clang-format off */
+static const int error_measure_lut[512] = {
+  // pow 0.7
+  16384, 16339, 16294, 16249, 16204, 16158, 16113, 16068,
+  16022, 15977, 15932, 15886, 15840, 15795, 15749, 15703,
+  15657, 15612, 15566, 15520, 15474, 15427, 15381, 15335,
+  15289, 15242, 15196, 15149, 15103, 15056, 15010, 14963,
+  14916, 14869, 14822, 14775, 14728, 14681, 14634, 14587,
+  14539, 14492, 14445, 14397, 14350, 14302, 14254, 14206,
+  14159, 14111, 14063, 14015, 13967, 13918, 13870, 13822,
+  13773, 13725, 13676, 13628, 13579, 13530, 13481, 13432,
+  13383, 13334, 13285, 13236, 13187, 13137, 13088, 13038,
+  12988, 12939, 12889, 12839, 12789, 12739, 12689, 12639,
+  12588, 12538, 12487, 12437, 12386, 12335, 12285, 12234,
+  12183, 12132, 12080, 12029, 11978, 11926, 11875, 11823,
+  11771, 11719, 11667, 11615, 11563, 11511, 11458, 11406,
+  11353, 11301, 11248, 11195, 11142, 11089, 11036, 10982,
+  10929, 10875, 10822, 10768, 10714, 10660, 10606, 10552,
+  10497, 10443, 10388, 10333, 10279, 10224, 10168, 10113,
+  10058, 10002,  9947,  9891,  9835,  9779,  9723,  9666,
+  9610, 9553, 9497, 9440, 9383, 9326, 9268, 9211,
+  9153, 9095, 9037, 8979, 8921, 8862, 8804, 8745,
+  8686, 8627, 8568, 8508, 8449, 8389, 8329, 8269,
+  8208, 8148, 8087, 8026, 7965, 7903, 7842, 7780,
+  7718, 7656, 7593, 7531, 7468, 7405, 7341, 7278,
+  7214, 7150, 7086, 7021, 6956, 6891, 6826, 6760,
+  6695, 6628, 6562, 6495, 6428, 6361, 6293, 6225,
+  6157, 6089, 6020, 5950, 5881, 5811, 5741, 5670,
+  5599, 5527, 5456, 5383, 5311, 5237, 5164, 5090,
+  5015, 4941, 4865, 4789, 4713, 4636, 4558, 4480,
+  4401, 4322, 4242, 4162, 4080, 3998, 3916, 3832,
+  3748, 3663, 3577, 3490, 3402, 3314, 3224, 3133,
+  3041, 2948, 2854, 2758, 2661, 2562, 2461, 2359,
+  2255, 2148, 2040, 1929, 1815, 1698, 1577, 1452,
+  1323, 1187, 1045,  894,  731,  550,  339,    0,
+  339,  550,  731,  894, 1045, 1187, 1323, 1452,
+  1577, 1698, 1815, 1929, 2040, 2148, 2255, 2359,
+  2461, 2562, 2661, 2758, 2854, 2948, 3041, 3133,
+  3224, 3314, 3402, 3490, 3577, 3663, 3748, 3832,
+  3916, 3998, 4080, 4162, 4242, 4322, 4401, 4480,
+  4558, 4636, 4713, 4789, 4865, 4941, 5015, 5090,
+  5164, 5237, 5311, 5383, 5456, 5527, 5599, 5670,
+  5741, 5811, 5881, 5950, 6020, 6089, 6157, 6225,
+  6293, 6361, 6428, 6495, 6562, 6628, 6695, 6760,
+  6826, 6891, 6956, 7021, 7086, 7150, 7214, 7278,
+  7341, 7405, 7468, 7531, 7593, 7656, 7718, 7780,
+  7842, 7903, 7965, 8026, 8087, 8148, 8208, 8269,
+  8329, 8389, 8449, 8508, 8568, 8627, 8686, 8745,
+  8804, 8862, 8921, 8979, 9037, 9095, 9153, 9211,
+  9268, 9326, 9383, 9440, 9497, 9553, 9610, 9666,
+  9723,  9779,  9835,  9891,  9947, 10002, 10058, 10113,
+  10168, 10224, 10279, 10333, 10388, 10443, 10497, 10552,
+  10606, 10660, 10714, 10768, 10822, 10875, 10929, 10982,
+  11036, 11089, 11142, 11195, 11248, 11301, 11353, 11406,
+  11458, 11511, 11563, 11615, 11667, 11719, 11771, 11823,
+  11875, 11926, 11978, 12029, 12080, 12132, 12183, 12234,
+  12285, 12335, 12386, 12437, 12487, 12538, 12588, 12639,
+  12689, 12739, 12789, 12839, 12889, 12939, 12988, 13038,
+  13088, 13137, 13187, 13236, 13285, 13334, 13383, 13432,
+  13481, 13530, 13579, 13628, 13676, 13725, 13773, 13822,
+  13870, 13918, 13967, 14015, 14063, 14111, 14159, 14206,
+  14254, 14302, 14350, 14397, 14445, 14492, 14539, 14587,
+  14634, 14681, 14728, 14775, 14822, 14869, 14916, 14963,
+  15010, 15056, 15103, 15149, 15196, 15242, 15289, 15335,
+  15381, 15427, 15474, 15520, 15566, 15612, 15657, 15703,
+  15749, 15795, 15840, 15886, 15932, 15977, 16022, 16068,
+  16113, 16158, 16204, 16249, 16294, 16339, 16384, 16384,
+};
+/* clang-format on */
+
+static ProjectPointsFunc get_project_points_type(TransformationType type) {
+  switch (type) {
+    case HOMOGRAPHY: return project_points_homography;
+    case AFFINE: return project_points_affine;
+    case ROTZOOM: return project_points_rotzoom;
+    case TRANSLATION: return project_points_translation;
+    default: assert(0); return NULL;
+  }
+}
+
+void project_points_translation(int32_t *mat, int *points, int *proj,
+                                const int n, const int stride_points,
+                                const int stride_proj, const int subsampling_x,
+                                const int subsampling_y) {
+  int i;
+  for (i = 0; i < n; ++i) {
+    const int x = *(points++), y = *(points++);
+    if (subsampling_x)
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(
+          ((x * (1 << (WARPEDMODEL_PREC_BITS + 1))) + mat[0]),
+          WARPEDDIFF_PREC_BITS + 1);
+    else
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(
+          ((x * (1 << WARPEDMODEL_PREC_BITS)) + mat[0]), WARPEDDIFF_PREC_BITS);
+    if (subsampling_y)
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(
+          ((y * (1 << (WARPEDMODEL_PREC_BITS + 1))) + mat[1]),
+          WARPEDDIFF_PREC_BITS + 1);
+    else
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(
+          ((y * (1 << WARPEDMODEL_PREC_BITS))) + mat[1], WARPEDDIFF_PREC_BITS);
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+void project_points_rotzoom(int32_t *mat, int *points, int *proj, const int n,
+                            const int stride_points, const int stride_proj,
+                            const int subsampling_x, const int subsampling_y) {
+  int i;
+  for (i = 0; i < n; ++i) {
+    const int x = *(points++), y = *(points++);
+    if (subsampling_x)
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(
+          mat[2] * 2 * x + mat[3] * 2 * y + mat[0] +
+              (mat[2] + mat[3] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+          WARPEDDIFF_PREC_BITS + 1);
+    else
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(mat[2] * x + mat[3] * y + mat[0],
+                                            WARPEDDIFF_PREC_BITS);
+    if (subsampling_y)
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(
+          -mat[3] * 2 * x + mat[2] * 2 * y + mat[1] +
+              (-mat[3] + mat[2] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+          WARPEDDIFF_PREC_BITS + 1);
+    else
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(-mat[3] * x + mat[2] * y + mat[1],
+                                            WARPEDDIFF_PREC_BITS);
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+void project_points_affine(int32_t *mat, int *points, int *proj, const int n,
+                           const int stride_points, const int stride_proj,
+                           const int subsampling_x, const int subsampling_y) {
+  int i;
+  for (i = 0; i < n; ++i) {
+    const int x = *(points++), y = *(points++);
+    if (subsampling_x)
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(
+          mat[2] * 2 * x + mat[3] * 2 * y + mat[0] +
+              (mat[2] + mat[3] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+          WARPEDDIFF_PREC_BITS + 1);
+    else
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(mat[2] * x + mat[3] * y + mat[0],
+                                            WARPEDDIFF_PREC_BITS);
+    if (subsampling_y)
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(
+          mat[4] * 2 * x + mat[5] * 2 * y + mat[1] +
+              (mat[4] + mat[5] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+          WARPEDDIFF_PREC_BITS + 1);
+    else
+      *(proj++) = ROUND_POWER_OF_TWO_SIGNED(mat[4] * x + mat[5] * y + mat[1],
+                                            WARPEDDIFF_PREC_BITS);
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+void project_points_hortrapezoid(int32_t *mat, int *points, int *proj,
+                                 const int n, const int stride_points,
+                                 const int stride_proj, const int subsampling_x,
+                                 const int subsampling_y) {
+  int i;
+  int64_t x, y, Z;
+  int64_t xp, yp;
+  for (i = 0; i < n; ++i) {
+    x = *(points++), y = *(points++);
+    x = (subsampling_x ? 4 * x + 1 : 2 * x);
+    y = (subsampling_y ? 4 * y + 1 : 2 * y);
+
+    Z = (mat[7] * y + (1 << (WARPEDMODEL_ROW3HOMO_PREC_BITS + 1)));
+    xp = (mat[2] * x + mat[3] * y + 2 * mat[0]) *
+         (1 << (WARPEDPIXEL_PREC_BITS + WARPEDMODEL_ROW3HOMO_PREC_BITS -
+                WARPEDMODEL_PREC_BITS));
+    yp = (mat[5] * y + 2 * mat[1]) *
+         (1 << (WARPEDPIXEL_PREC_BITS + WARPEDMODEL_ROW3HOMO_PREC_BITS -
+                WARPEDMODEL_PREC_BITS));
+
+    xp = xp > 0 ? (xp + Z / 2) / Z : (xp - Z / 2) / Z;
+    yp = yp > 0 ? (yp + Z / 2) / Z : (yp - Z / 2) / Z;
+
+    if (subsampling_x) xp = (xp - (1 << (WARPEDPIXEL_PREC_BITS - 1))) / 2;
+    if (subsampling_y) yp = (yp - (1 << (WARPEDPIXEL_PREC_BITS - 1))) / 2;
+    *(proj++) = (int)xp;
+    *(proj++) = (int)yp;
+
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+void project_points_vertrapezoid(int32_t *mat, int *points, int *proj,
+                                 const int n, const int stride_points,
+                                 const int stride_proj, const int subsampling_x,
+                                 const int subsampling_y) {
+  int i;
+  int64_t x, y, Z;
+  int64_t xp, yp;
+  for (i = 0; i < n; ++i) {
+    x = *(points++), y = *(points++);
+    x = (subsampling_x ? 4 * x + 1 : 2 * x);
+    y = (subsampling_y ? 4 * y + 1 : 2 * y);
+
+    Z = (mat[6] * x + (1 << (WARPEDMODEL_ROW3HOMO_PREC_BITS + 1)));
+    xp = (mat[2] * x + 2 * mat[0]) *
+         (1 << (WARPEDPIXEL_PREC_BITS + WARPEDMODEL_ROW3HOMO_PREC_BITS -
+                WARPEDMODEL_PREC_BITS));
+    yp = (mat[4] * x + mat[5] * y + 2 * mat[1]) *
+         (1 << (WARPEDPIXEL_PREC_BITS + WARPEDMODEL_ROW3HOMO_PREC_BITS -
+                WARPEDMODEL_PREC_BITS));
+
+    xp = xp > 0 ? (xp + Z / 2) / Z : (xp - Z / 2) / Z;
+    yp = yp > 0 ? (yp + Z / 2) / Z : (yp - Z / 2) / Z;
+
+    if (subsampling_x) xp = (xp - (1 << (WARPEDPIXEL_PREC_BITS - 1))) / 2;
+    if (subsampling_y) yp = (yp - (1 << (WARPEDPIXEL_PREC_BITS - 1))) / 2;
+    *(proj++) = (int)xp;
+    *(proj++) = (int)yp;
+
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+void project_points_homography(int32_t *mat, int *points, int *proj,
+                               const int n, const int stride_points,
+                               const int stride_proj, const int subsampling_x,
+                               const int subsampling_y) {
+  int i;
+  int64_t x, y, Z;
+  int64_t xp, yp;
+  for (i = 0; i < n; ++i) {
+    x = *(points++), y = *(points++);
+    x = (subsampling_x ? 4 * x + 1 : 2 * x);
+    y = (subsampling_y ? 4 * y + 1 : 2 * y);
+
+    Z = (mat[6] * x + mat[7] * y + (1 << (WARPEDMODEL_ROW3HOMO_PREC_BITS + 1)));
+    xp = (mat[2] * x + mat[3] * y + 2 * mat[0]) *
+         (1 << (WARPEDPIXEL_PREC_BITS + WARPEDMODEL_ROW3HOMO_PREC_BITS -
+                WARPEDMODEL_PREC_BITS));
+    yp = (mat[4] * x + mat[5] * y + 2 * mat[1]) *
+         (1 << (WARPEDPIXEL_PREC_BITS + WARPEDMODEL_ROW3HOMO_PREC_BITS -
+                WARPEDMODEL_PREC_BITS));
+
+    xp = xp > 0 ? (xp + Z / 2) / Z : (xp - Z / 2) / Z;
+    yp = yp > 0 ? (yp + Z / 2) / Z : (yp - Z / 2) / Z;
+
+    if (subsampling_x) xp = (xp - (1 << (WARPEDPIXEL_PREC_BITS - 1))) / 2;
+    if (subsampling_y) yp = (yp - (1 << (WARPEDPIXEL_PREC_BITS - 1))) / 2;
+    *(proj++) = (int)xp;
+    *(proj++) = (int)yp;
+
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+// 'points' are at original scale, output 'proj's are scaled up by
+// 1 << WARPEDPIXEL_PREC_BITS
+void project_points(WarpedMotionParams *wm_params, int *points, int *proj,
+                    const int n, const int stride_points, const int stride_proj,
+                    const int subsampling_x, const int subsampling_y) {
+  switch (wm_params->wmtype) {
+    case AFFINE:
+      project_points_affine(wm_params->wmmat, points, proj, n, stride_points,
+                            stride_proj, subsampling_x, subsampling_y);
+      break;
+    case ROTZOOM:
+      project_points_rotzoom(wm_params->wmmat, points, proj, n, stride_points,
+                             stride_proj, subsampling_x, subsampling_y);
+      break;
+    case HOMOGRAPHY:
+      project_points_homography(wm_params->wmmat, points, proj, n,
+                                stride_points, stride_proj, subsampling_x,
+                                subsampling_y);
+      break;
+    default: assert(0 && "Invalid warped motion type!"); return;
+  }
+}
+
+static const int16_t
+    filter_ntap[WARPEDPIXEL_PREC_SHIFTS][WARPEDPIXEL_FILTER_TAPS] = {
+#if WARPEDPIXEL_PREC_BITS == 6
+      { 0, 0, 128, 0, 0, 0 },      { 0, -1, 128, 2, -1, 0 },
+      { 1, -3, 127, 4, -1, 0 },    { 1, -4, 126, 6, -2, 1 },
+      { 1, -5, 126, 8, -3, 1 },    { 1, -6, 125, 11, -4, 1 },
+      { 1, -7, 124, 13, -4, 1 },   { 2, -8, 123, 15, -5, 1 },
+      { 2, -9, 122, 18, -6, 1 },   { 2, -10, 121, 20, -6, 1 },
+      { 2, -11, 120, 22, -7, 2 },  { 2, -12, 119, 25, -8, 2 },
+      { 3, -13, 117, 27, -8, 2 },  { 3, -13, 116, 29, -9, 2 },
+      { 3, -14, 114, 32, -10, 3 }, { 3, -15, 113, 35, -10, 2 },
+      { 3, -15, 111, 37, -11, 3 }, { 3, -16, 109, 40, -11, 3 },
+      { 3, -16, 108, 42, -12, 3 }, { 4, -17, 106, 45, -13, 3 },
+      { 4, -17, 104, 47, -13, 3 }, { 4, -17, 102, 50, -14, 3 },
+      { 4, -17, 100, 52, -14, 3 }, { 4, -18, 98, 55, -15, 4 },
+      { 4, -18, 96, 58, -15, 3 },  { 4, -18, 94, 60, -16, 4 },
+      { 4, -18, 91, 63, -16, 4 },  { 4, -18, 89, 65, -16, 4 },
+      { 4, -18, 87, 68, -17, 4 },  { 4, -18, 85, 70, -17, 4 },
+      { 4, -18, 82, 73, -17, 4 },  { 4, -18, 80, 75, -17, 4 },
+      { 4, -18, 78, 78, -18, 4 },  { 4, -17, 75, 80, -18, 4 },
+      { 4, -17, 73, 82, -18, 4 },  { 4, -17, 70, 85, -18, 4 },
+      { 4, -17, 68, 87, -18, 4 },  { 4, -16, 65, 89, -18, 4 },
+      { 4, -16, 63, 91, -18, 4 },  { 4, -16, 60, 94, -18, 4 },
+      { 3, -15, 58, 96, -18, 4 },  { 4, -15, 55, 98, -18, 4 },
+      { 3, -14, 52, 100, -17, 4 }, { 3, -14, 50, 102, -17, 4 },
+      { 3, -13, 47, 104, -17, 4 }, { 3, -13, 45, 106, -17, 4 },
+      { 3, -12, 42, 108, -16, 3 }, { 3, -11, 40, 109, -16, 3 },
+      { 3, -11, 37, 111, -15, 3 }, { 2, -10, 35, 113, -15, 3 },
+      { 3, -10, 32, 114, -14, 3 }, { 2, -9, 29, 116, -13, 3 },
+      { 2, -8, 27, 117, -13, 3 },  { 2, -8, 25, 119, -12, 2 },
+      { 2, -7, 22, 120, -11, 2 },  { 1, -6, 20, 121, -10, 2 },
+      { 1, -6, 18, 122, -9, 2 },   { 1, -5, 15, 123, -8, 2 },
+      { 1, -4, 13, 124, -7, 1 },   { 1, -4, 11, 125, -6, 1 },
+      { 1, -3, 8, 126, -5, 1 },    { 1, -2, 6, 126, -4, 1 },
+      { 0, -1, 4, 127, -3, 1 },    { 0, -1, 2, 128, -1, 0 },
+#elif WARPEDPIXEL_PREC_BITS == 5
+      { 0, 0, 128, 0, 0, 0 },      { 1, -3, 127, 4, -1, 0 },
+      { 1, -5, 126, 8, -3, 1 },    { 1, -7, 124, 13, -4, 1 },
+      { 2, -9, 122, 18, -6, 1 },   { 2, -11, 120, 22, -7, 2 },
+      { 3, -13, 117, 27, -8, 2 },  { 3, -14, 114, 32, -10, 3 },
+      { 3, -15, 111, 37, -11, 3 }, { 3, -16, 108, 42, -12, 3 },
+      { 4, -17, 104, 47, -13, 3 }, { 4, -17, 100, 52, -14, 3 },
+      { 4, -18, 96, 58, -15, 3 },  { 4, -18, 91, 63, -16, 4 },
+      { 4, -18, 87, 68, -17, 4 },  { 4, -18, 82, 73, -17, 4 },
+      { 4, -18, 78, 78, -18, 4 },  { 4, -17, 73, 82, -18, 4 },
+      { 4, -17, 68, 87, -18, 4 },  { 4, -16, 63, 91, -18, 4 },
+      { 3, -15, 58, 96, -18, 4 },  { 3, -14, 52, 100, -17, 4 },
+      { 3, -13, 47, 104, -17, 4 }, { 3, -12, 42, 108, -16, 3 },
+      { 3, -11, 37, 111, -15, 3 }, { 3, -10, 32, 114, -14, 3 },
+      { 2, -8, 27, 117, -13, 3 },  { 2, -7, 22, 120, -11, 2 },
+      { 1, -6, 18, 122, -9, 2 },   { 1, -4, 13, 124, -7, 1 },
+      { 1, -3, 8, 126, -5, 1 },    { 0, -1, 4, 127, -3, 1 },
+#endif  // WARPEDPIXEL_PREC_BITS == 6
+    };
+
+static int32_t do_ntap_filter(int32_t *p, int x) {
+  int i;
+  int32_t sum = 0;
+  for (i = 0; i < WARPEDPIXEL_FILTER_TAPS; ++i) {
+    sum += p[i - WARPEDPIXEL_FILTER_TAPS / 2 + 1] * filter_ntap[x][i];
+  }
+  return sum;
+}
+
+static int32_t do_cubic_filter(int32_t *p, int x) {
+  if (x == 0) {
+    return p[0] * (1 << WARPEDPIXEL_FILTER_BITS);
+  } else if (x == (1 << WARPEDPIXEL_PREC_BITS)) {
+    return p[1] * (1 << WARPEDPIXEL_FILTER_BITS);
+  } else {
+    const int64_t v1 = (int64_t)x * x * x * (3 * (p[0] - p[1]) + p[2] - p[-1]);
+    const int64_t v2 =
+        (int64_t)x * x * (2 * p[-1] - 5 * p[0] + 4 * p[1] - p[2]);
+    const int64_t v3 = x * (p[1] - p[-1]);
+    const int64_t v4 = 2 * p[0];
+    return (int32_t)ROUND_POWER_OF_TWO_SIGNED(
+        (v4 * (1 << (3 * WARPEDPIXEL_PREC_BITS))) +
+            (v3 * (1 << (2 * WARPEDPIXEL_PREC_BITS))) +
+            (v2 * (1 << WARPEDPIXEL_PREC_BITS)) + v1,
+        3 * WARPEDPIXEL_PREC_BITS + 1 - WARPEDPIXEL_FILTER_BITS);
+  }
+}
+
+static INLINE void get_subcolumn(int taps, uint8_t *ref, int32_t *col,
+                                 int stride, int x, int y_start) {
+  int i;
+  for (i = 0; i < taps; ++i) {
+    col[i] = ref[(i + y_start) * stride + x];
+  }
+}
+
+static uint8_t bi_ntap_filter(uint8_t *ref, int x, int y, int stride) {
+  int32_t val, arr[WARPEDPIXEL_FILTER_TAPS];
+  int k;
+  int i = (int)x >> WARPEDPIXEL_PREC_BITS;
+  int j = (int)y >> WARPEDPIXEL_PREC_BITS;
+  for (k = 0; k < WARPEDPIXEL_FILTER_TAPS; ++k) {
+    int32_t arr_temp[WARPEDPIXEL_FILTER_TAPS];
+    get_subcolumn(WARPEDPIXEL_FILTER_TAPS, ref, arr_temp, stride,
+                  i + k + 1 - WARPEDPIXEL_FILTER_TAPS / 2,
+                  j + 1 - WARPEDPIXEL_FILTER_TAPS / 2);
+    arr[k] = do_ntap_filter(arr_temp + WARPEDPIXEL_FILTER_TAPS / 2 - 1,
+                            y - (j * (1 << WARPEDPIXEL_PREC_BITS)));
+  }
+  val = do_ntap_filter(arr + WARPEDPIXEL_FILTER_TAPS / 2 - 1,
+                       x - (i * (1 << WARPEDPIXEL_PREC_BITS)));
+  val = ROUND_POWER_OF_TWO_SIGNED(val, WARPEDPIXEL_FILTER_BITS * 2);
+  return (uint8_t)clip_pixel(val);
+}
+
+static uint8_t bi_cubic_filter(uint8_t *ref, int x, int y, int stride) {
+  int32_t val, arr[4];
+  int k;
+  int i = (int)x >> WARPEDPIXEL_PREC_BITS;
+  int j = (int)y >> WARPEDPIXEL_PREC_BITS;
+  for (k = 0; k < 4; ++k) {
+    int32_t arr_temp[4];
+    get_subcolumn(4, ref, arr_temp, stride, i + k - 1, j - 1);
+    arr[k] =
+        do_cubic_filter(arr_temp + 1, y - (j * (1 << WARPEDPIXEL_PREC_BITS)));
+  }
+  val = do_cubic_filter(arr + 1, x - (i * (1 << WARPEDPIXEL_PREC_BITS)));
+  val = ROUND_POWER_OF_TWO_SIGNED(val, WARPEDPIXEL_FILTER_BITS * 2);
+  return (uint8_t)clip_pixel(val);
+}
+
+static uint8_t bi_linear_filter(uint8_t *ref, int x, int y, int stride) {
+  const int ix = x >> WARPEDPIXEL_PREC_BITS;
+  const int iy = y >> WARPEDPIXEL_PREC_BITS;
+  const int sx = x - (ix * (1 << WARPEDPIXEL_PREC_BITS));
+  const int sy = y - (iy * (1 << WARPEDPIXEL_PREC_BITS));
+  int32_t val;
+  val = ROUND_POWER_OF_TWO_SIGNED(
+      ref[iy * stride + ix] * (WARPEDPIXEL_PREC_SHIFTS - sy) *
+              (WARPEDPIXEL_PREC_SHIFTS - sx) +
+          ref[iy * stride + ix + 1] * (WARPEDPIXEL_PREC_SHIFTS - sy) * sx +
+          ref[(iy + 1) * stride + ix] * sy * (WARPEDPIXEL_PREC_SHIFTS - sx) +
+          ref[(iy + 1) * stride + ix + 1] * sy * sx,
+      WARPEDPIXEL_PREC_BITS * 2);
+  return (uint8_t)clip_pixel(val);
+}
+
+static uint8_t warp_interpolate(uint8_t *ref, int x, int y, int width,
+                                int height, int stride) {
+  int ix = x >> WARPEDPIXEL_PREC_BITS;
+  int iy = y >> WARPEDPIXEL_PREC_BITS;
+  int sx = x - (ix * (1 << WARPEDPIXEL_PREC_BITS));
+  int sy = y - (iy * (1 << WARPEDPIXEL_PREC_BITS));
+  int32_t v;
+
+  if (ix < 0 && iy < 0)
+    return ref[0];
+  else if (ix < 0 && iy >= height - 1)
+    return ref[(height - 1) * stride];
+  else if (ix >= width - 1 && iy < 0)
+    return ref[width - 1];
+  else if (ix >= width - 1 && iy >= height - 1)
+    return ref[(height - 1) * stride + (width - 1)];
+  else if (ix < 0) {
+    v = ROUND_POWER_OF_TWO_SIGNED(
+        ref[iy * stride] * (WARPEDPIXEL_PREC_SHIFTS - sy) +
+            ref[(iy + 1) * stride] * sy,
+        WARPEDPIXEL_PREC_BITS);
+    return clip_pixel(v);
+  } else if (iy < 0) {
+    v = ROUND_POWER_OF_TWO_SIGNED(
+        ref[ix] * (WARPEDPIXEL_PREC_SHIFTS - sx) + ref[ix + 1] * sx,
+        WARPEDPIXEL_PREC_BITS);
+    return clip_pixel(v);
+  } else if (ix >= width - 1) {
+    v = ROUND_POWER_OF_TWO_SIGNED(
+        ref[iy * stride + width - 1] * (WARPEDPIXEL_PREC_SHIFTS - sy) +
+            ref[(iy + 1) * stride + width - 1] * sy,
+        WARPEDPIXEL_PREC_BITS);
+    return clip_pixel(v);
+  } else if (iy >= height - 1) {
+    v = ROUND_POWER_OF_TWO_SIGNED(
+        ref[(height - 1) * stride + ix] * (WARPEDPIXEL_PREC_SHIFTS - sx) +
+            ref[(height - 1) * stride + ix + 1] * sx,
+        WARPEDPIXEL_PREC_BITS);
+    return clip_pixel(v);
+  } else if (ix >= WARPEDPIXEL_FILTER_TAPS / 2 - 1 &&
+             iy >= WARPEDPIXEL_FILTER_TAPS / 2 - 1 &&
+             ix < width - WARPEDPIXEL_FILTER_TAPS / 2 &&
+             iy < height - WARPEDPIXEL_FILTER_TAPS / 2) {
+    return bi_ntap_filter(ref, x, y, stride);
+  } else if (ix >= 1 && iy >= 1 && ix < width - 2 && iy < height - 2) {
+    return bi_cubic_filter(ref, x, y, stride);
+  } else {
+    return bi_linear_filter(ref, x, y, stride);
+  }
+}
+
+// For warping, we really use a 6-tap filter, but we do blocks of 8 pixels
+// at a time. The zoom/rotation/shear in the model are applied to the
+// "fractional" position of each pixel, which therefore varies within
+// [-1, 2) * WARPEDPIXEL_PREC_SHIFTS.
+// We need an extra 2 taps to fit this in, for a total of 8 taps.
+/* clang-format off */
+const int16_t warped_filter[WARPEDPIXEL_PREC_SHIFTS * 3 + 1][8] = {
+#if WARPEDPIXEL_PREC_BITS == 6
+  // [-1, 0)
+  { 0,   0, 127,   1,   0, 0, 0, 0 }, { 0, - 1, 127,   2,   0, 0, 0, 0 },
+  { 1, - 3, 127,   4, - 1, 0, 0, 0 }, { 1, - 4, 126,   6, - 2, 1, 0, 0 },
+  { 1, - 5, 126,   8, - 3, 1, 0, 0 }, { 1, - 6, 125,  11, - 4, 1, 0, 0 },
+  { 1, - 7, 124,  13, - 4, 1, 0, 0 }, { 2, - 8, 123,  15, - 5, 1, 0, 0 },
+  { 2, - 9, 122,  18, - 6, 1, 0, 0 }, { 2, -10, 121,  20, - 6, 1, 0, 0 },
+  { 2, -11, 120,  22, - 7, 2, 0, 0 }, { 2, -12, 119,  25, - 8, 2, 0, 0 },
+  { 3, -13, 117,  27, - 8, 2, 0, 0 }, { 3, -13, 116,  29, - 9, 2, 0, 0 },
+  { 3, -14, 114,  32, -10, 3, 0, 0 }, { 3, -15, 113,  35, -10, 2, 0, 0 },
+  { 3, -15, 111,  37, -11, 3, 0, 0 }, { 3, -16, 109,  40, -11, 3, 0, 0 },
+  { 3, -16, 108,  42, -12, 3, 0, 0 }, { 4, -17, 106,  45, -13, 3, 0, 0 },
+  { 4, -17, 104,  47, -13, 3, 0, 0 }, { 4, -17, 102,  50, -14, 3, 0, 0 },
+  { 4, -17, 100,  52, -14, 3, 0, 0 }, { 4, -18,  98,  55, -15, 4, 0, 0 },
+  { 4, -18,  96,  58, -15, 3, 0, 0 }, { 4, -18,  94,  60, -16, 4, 0, 0 },
+  { 4, -18,  91,  63, -16, 4, 0, 0 }, { 4, -18,  89,  65, -16, 4, 0, 0 },
+  { 4, -18,  87,  68, -17, 4, 0, 0 }, { 4, -18,  85,  70, -17, 4, 0, 0 },
+  { 4, -18,  82,  73, -17, 4, 0, 0 }, { 4, -18,  80,  75, -17, 4, 0, 0 },
+  { 4, -18,  78,  78, -18, 4, 0, 0 }, { 4, -17,  75,  80, -18, 4, 0, 0 },
+  { 4, -17,  73,  82, -18, 4, 0, 0 }, { 4, -17,  70,  85, -18, 4, 0, 0 },
+  { 4, -17,  68,  87, -18, 4, 0, 0 }, { 4, -16,  65,  89, -18, 4, 0, 0 },
+  { 4, -16,  63,  91, -18, 4, 0, 0 }, { 4, -16,  60,  94, -18, 4, 0, 0 },
+  { 3, -15,  58,  96, -18, 4, 0, 0 }, { 4, -15,  55,  98, -18, 4, 0, 0 },
+  { 3, -14,  52, 100, -17, 4, 0, 0 }, { 3, -14,  50, 102, -17, 4, 0, 0 },
+  { 3, -13,  47, 104, -17, 4, 0, 0 }, { 3, -13,  45, 106, -17, 4, 0, 0 },
+  { 3, -12,  42, 108, -16, 3, 0, 0 }, { 3, -11,  40, 109, -16, 3, 0, 0 },
+  { 3, -11,  37, 111, -15, 3, 0, 0 }, { 2, -10,  35, 113, -15, 3, 0, 0 },
+  { 3, -10,  32, 114, -14, 3, 0, 0 }, { 2, - 9,  29, 116, -13, 3, 0, 0 },
+  { 2, - 8,  27, 117, -13, 3, 0, 0 }, { 2, - 8,  25, 119, -12, 2, 0, 0 },
+  { 2, - 7,  22, 120, -11, 2, 0, 0 }, { 1, - 6,  20, 121, -10, 2, 0, 0 },
+  { 1, - 6,  18, 122, - 9, 2, 0, 0 }, { 1, - 5,  15, 123, - 8, 2, 0, 0 },
+  { 1, - 4,  13, 124, - 7, 1, 0, 0 }, { 1, - 4,  11, 125, - 6, 1, 0, 0 },
+  { 1, - 3,   8, 126, - 5, 1, 0, 0 }, { 1, - 2,   6, 126, - 4, 1, 0, 0 },
+  { 0, - 1,   4, 127, - 3, 1, 0, 0 }, { 0,   0,   2, 127, - 1, 0, 0, 0 },
+
+  // [0, 1)
+  { 0,  0,   0, 127,   1,   0,  0,  0}, { 0,  0,  -1, 127,   2,   0,  0,  0},
+  { 0,  1,  -3, 127,   4,  -2,  1,  0}, { 0,  1,  -5, 127,   6,  -2,  1,  0},
+  { 0,  2,  -6, 126,   8,  -3,  1,  0}, {-1,  2,  -7, 126,  11,  -4,  2, -1},
+  {-1,  3,  -8, 125,  13,  -5,  2, -1}, {-1,  3, -10, 124,  16,  -6,  3, -1},
+  {-1,  4, -11, 123,  18,  -7,  3, -1}, {-1,  4, -12, 122,  20,  -7,  3, -1},
+  {-1,  4, -13, 121,  23,  -8,  3, -1}, {-2,  5, -14, 120,  25,  -9,  4, -1},
+  {-1,  5, -15, 119,  27, -10,  4, -1}, {-1,  5, -16, 118,  30, -11,  4, -1},
+  {-2,  6, -17, 116,  33, -12,  5, -1}, {-2,  6, -17, 114,  35, -12,  5, -1},
+  {-2,  6, -18, 113,  38, -13,  5, -1}, {-2,  7, -19, 111,  41, -14,  6, -2},
+  {-2,  7, -19, 110,  43, -15,  6, -2}, {-2,  7, -20, 108,  46, -15,  6, -2},
+  {-2,  7, -20, 106,  49, -16,  6, -2}, {-2,  7, -21, 104,  51, -16,  7, -2},
+  {-2,  7, -21, 102,  54, -17,  7, -2}, {-2,  8, -21, 100,  56, -18,  7, -2},
+  {-2,  8, -22,  98,  59, -18,  7, -2}, {-2,  8, -22,  96,  62, -19,  7, -2},
+  {-2,  8, -22,  94,  64, -19,  7, -2}, {-2,  8, -22,  91,  67, -20,  8, -2},
+  {-2,  8, -22,  89,  69, -20,  8, -2}, {-2,  8, -22,  87,  72, -21,  8, -2},
+  {-2,  8, -21,  84,  74, -21,  8, -2}, {-2,  8, -22,  82,  77, -21,  8, -2},
+  {-2,  8, -21,  79,  79, -21,  8, -2}, {-2,  8, -21,  77,  82, -22,  8, -2},
+  {-2,  8, -21,  74,  84, -21,  8, -2}, {-2,  8, -21,  72,  87, -22,  8, -2},
+  {-2,  8, -20,  69,  89, -22,  8, -2}, {-2,  8, -20,  67,  91, -22,  8, -2},
+  {-2,  7, -19,  64,  94, -22,  8, -2}, {-2,  7, -19,  62,  96, -22,  8, -2},
+  {-2,  7, -18,  59,  98, -22,  8, -2}, {-2,  7, -18,  56, 100, -21,  8, -2},
+  {-2,  7, -17,  54, 102, -21,  7, -2}, {-2,  7, -16,  51, 104, -21,  7, -2},
+  {-2,  6, -16,  49, 106, -20,  7, -2}, {-2,  6, -15,  46, 108, -20,  7, -2},
+  {-2,  6, -15,  43, 110, -19,  7, -2}, {-2,  6, -14,  41, 111, -19,  7, -2},
+  {-1,  5, -13,  38, 113, -18,  6, -2}, {-1,  5, -12,  35, 114, -17,  6, -2},
+  {-1,  5, -12,  33, 116, -17,  6, -2}, {-1,  4, -11,  30, 118, -16,  5, -1},
+  {-1,  4, -10,  27, 119, -15,  5, -1}, {-1,  4,  -9,  25, 120, -14,  5, -2},
+  {-1,  3,  -8,  23, 121, -13,  4, -1}, {-1,  3,  -7,  20, 122, -12,  4, -1},
+  {-1,  3,  -7,  18, 123, -11,  4, -1}, {-1,  3,  -6,  16, 124, -10,  3, -1},
+  {-1,  2,  -5,  13, 125,  -8,  3, -1}, {-1,  2,  -4,  11, 126,  -7,  2, -1},
+  { 0,  1,  -3,   8, 126,  -6,  2,  0}, { 0,  1,  -2,   6, 127,  -5,  1,  0},
+  { 0,  1,  -2,   4, 127,  -3,  1,  0}, { 0,  0,   0,   2, 127,  -1,  0,  0},
+
+  // [1, 2)
+  { 0, 0, 0,   1, 127,   0,   0, 0 }, { 0, 0, 0, - 1, 127,   2,   0, 0 },
+  { 0, 0, 1, - 3, 127,   4, - 1, 0 }, { 0, 0, 1, - 4, 126,   6, - 2, 1 },
+  { 0, 0, 1, - 5, 126,   8, - 3, 1 }, { 0, 0, 1, - 6, 125,  11, - 4, 1 },
+  { 0, 0, 1, - 7, 124,  13, - 4, 1 }, { 0, 0, 2, - 8, 123,  15, - 5, 1 },
+  { 0, 0, 2, - 9, 122,  18, - 6, 1 }, { 0, 0, 2, -10, 121,  20, - 6, 1 },
+  { 0, 0, 2, -11, 120,  22, - 7, 2 }, { 0, 0, 2, -12, 119,  25, - 8, 2 },
+  { 0, 0, 3, -13, 117,  27, - 8, 2 }, { 0, 0, 3, -13, 116,  29, - 9, 2 },
+  { 0, 0, 3, -14, 114,  32, -10, 3 }, { 0, 0, 3, -15, 113,  35, -10, 2 },
+  { 0, 0, 3, -15, 111,  37, -11, 3 }, { 0, 0, 3, -16, 109,  40, -11, 3 },
+  { 0, 0, 3, -16, 108,  42, -12, 3 }, { 0, 0, 4, -17, 106,  45, -13, 3 },
+  { 0, 0, 4, -17, 104,  47, -13, 3 }, { 0, 0, 4, -17, 102,  50, -14, 3 },
+  { 0, 0, 4, -17, 100,  52, -14, 3 }, { 0, 0, 4, -18,  98,  55, -15, 4 },
+  { 0, 0, 4, -18,  96,  58, -15, 3 }, { 0, 0, 4, -18,  94,  60, -16, 4 },
+  { 0, 0, 4, -18,  91,  63, -16, 4 }, { 0, 0, 4, -18,  89,  65, -16, 4 },
+  { 0, 0, 4, -18,  87,  68, -17, 4 }, { 0, 0, 4, -18,  85,  70, -17, 4 },
+  { 0, 0, 4, -18,  82,  73, -17, 4 }, { 0, 0, 4, -18,  80,  75, -17, 4 },
+  { 0, 0, 4, -18,  78,  78, -18, 4 }, { 0, 0, 4, -17,  75,  80, -18, 4 },
+  { 0, 0, 4, -17,  73,  82, -18, 4 }, { 0, 0, 4, -17,  70,  85, -18, 4 },
+  { 0, 0, 4, -17,  68,  87, -18, 4 }, { 0, 0, 4, -16,  65,  89, -18, 4 },
+  { 0, 0, 4, -16,  63,  91, -18, 4 }, { 0, 0, 4, -16,  60,  94, -18, 4 },
+  { 0, 0, 3, -15,  58,  96, -18, 4 }, { 0, 0, 4, -15,  55,  98, -18, 4 },
+  { 0, 0, 3, -14,  52, 100, -17, 4 }, { 0, 0, 3, -14,  50, 102, -17, 4 },
+  { 0, 0, 3, -13,  47, 104, -17, 4 }, { 0, 0, 3, -13,  45, 106, -17, 4 },
+  { 0, 0, 3, -12,  42, 108, -16, 3 }, { 0, 0, 3, -11,  40, 109, -16, 3 },
+  { 0, 0, 3, -11,  37, 111, -15, 3 }, { 0, 0, 2, -10,  35, 113, -15, 3 },
+  { 0, 0, 3, -10,  32, 114, -14, 3 }, { 0, 0, 2, - 9,  29, 116, -13, 3 },
+  { 0, 0, 2, - 8,  27, 117, -13, 3 }, { 0, 0, 2, - 8,  25, 119, -12, 2 },
+  { 0, 0, 2, - 7,  22, 120, -11, 2 }, { 0, 0, 1, - 6,  20, 121, -10, 2 },
+  { 0, 0, 1, - 6,  18, 122, - 9, 2 }, { 0, 0, 1, - 5,  15, 123, - 8, 2 },
+  { 0, 0, 1, - 4,  13, 124, - 7, 1 }, { 0, 0, 1, - 4,  11, 125, - 6, 1 },
+  { 0, 0, 1, - 3,   8, 126, - 5, 1 }, { 0, 0, 1, - 2,   6, 126, - 4, 1 },
+  { 0, 0, 0, - 1,   4, 127, - 3, 1 }, { 0, 0, 0,   0,   2, 127, - 1, 0 },
+
+#elif WARPEDPIXEL_PREC_BITS == 5
+  // [-1, 0)
+  {0,   0, 127,   1,   0, 0, 0, 0}, {1,  -3, 127,   4,  -1, 0, 0, 0},
+  {1,  -5, 126,   8,  -3, 1, 0, 0}, {1,  -7, 124,  13,  -4, 1, 0, 0},
+  {2,  -9, 122,  18,  -6, 1, 0, 0}, {2, -11, 120,  22,  -7, 2, 0, 0},
+  {3, -13, 117,  27,  -8, 2, 0, 0}, {3, -14, 114,  32, -10, 3, 0, 0},
+  {3, -15, 111,  37, -11, 3, 0, 0}, {3, -16, 108,  42, -12, 3, 0, 0},
+  {4, -17, 104,  47, -13, 3, 0, 0}, {4, -17, 100,  52, -14, 3, 0, 0},
+  {4, -18,  96,  58, -15, 3, 0, 0}, {4, -18,  91,  63, -16, 4, 0, 0},
+  {4, -18,  87,  68, -17, 4, 0, 0}, {4, -18,  82,  73, -17, 4, 0, 0},
+  {4, -18,  78,  78, -18, 4, 0, 0}, {4, -17,  73,  82, -18, 4, 0, 0},
+  {4, -17,  68,  87, -18, 4, 0, 0}, {4, -16,  63,  91, -18, 4, 0, 0},
+  {3, -15,  58,  96, -18, 4, 0, 0}, {3, -14,  52, 100, -17, 4, 0, 0},
+  {3, -13,  47, 104, -17, 4, 0, 0}, {3, -12,  42, 108, -16, 3, 0, 0},
+  {3, -11,  37, 111, -15, 3, 0, 0}, {3, -10,  32, 114, -14, 3, 0, 0},
+  {2,  -8,  27, 117, -13, 3, 0, 0}, {2,  -7,  22, 120, -11, 2, 0, 0},
+  {1,  -6,  18, 122,  -9, 2, 0, 0}, {1,  -4,  13, 124,  -7, 1, 0, 0},
+  {1,  -3,   8, 126,  -5, 1, 0, 0}, {0,  -1,   4, 127,  -3, 1, 0, 0},
+  // [0, 1)
+  { 0,  0,   0, 127,   1,   0,   0,  0}, { 0,  1,  -3, 127,   4,  -2,   1,  0},
+  { 0,  2,  -6, 126,   8,  -3,   1,  0}, {-1,  3,  -8, 125,  13,  -5,   2, -1},
+  {-1,  4, -11, 123,  18,  -7,   3, -1}, {-1,  4, -13, 121,  23,  -8,   3, -1},
+  {-1,  5, -15, 119,  27, -10,   4, -1}, {-2,  6, -17, 116,  33, -12,   5, -1},
+  {-2,  6, -18, 113,  38, -13,   5, -1}, {-2,  7, -19, 110,  43, -15,   6, -2},
+  {-2,  7, -20, 106,  49, -16,   6, -2}, {-2,  7, -21, 102,  54, -17,   7, -2},
+  {-2,  8, -22,  98,  59, -18,   7, -2}, {-2,  8, -22,  94,  64, -19,   7, -2},
+  {-2,  8, -22,  89,  69, -20,   8, -2}, {-2,  8, -21,  84,  74, -21,   8, -2},
+  {-2,  8, -21,  79,  79, -21,   8, -2}, {-2,  8, -21,  74,  84, -21,   8, -2},
+  {-2,  8, -20,  69,  89, -22,   8, -2}, {-2,  7, -19,  64,  94, -22,   8, -2},
+  {-2,  7, -18,  59,  98, -22,   8, -2}, {-2,  7, -17,  54, 102, -21,   7, -2},
+  {-2,  6, -16,  49, 106, -20,   7, -2}, {-2,  6, -15,  43, 110, -19,   7, -2},
+  {-1,  5, -13,  38, 113, -18,   6, -2}, {-1,  5, -12,  33, 116, -17,   6, -2},
+  {-1,  4, -10,  27, 119, -15,   5, -1}, {-1,  3,  -8,  23, 121, -13,   4, -1},
+  {-1,  3,  -7,  18, 123, -11,   4, -1}, {-1,  2,  -5,  13, 125,  -8,   3, -1},
+  { 0,  1,  -3,   8, 126,  -6,   2,  0}, { 0,  1,  -2,   4, 127,  -3,   1,  0},
+  // [1, 2)
+  {0, 0, 0,   1, 127,   0,   0, 0}, {0, 0, 1,  -3, 127,   4,  -1, 0},
+  {0, 0, 1,  -5, 126,   8,  -3, 1}, {0, 0, 1,  -7, 124,  13,  -4, 1},
+  {0, 0, 2,  -9, 122,  18,  -6, 1}, {0, 0, 2, -11, 120,  22,  -7, 2},
+  {0, 0, 3, -13, 117,  27,  -8, 2}, {0, 0, 3, -14, 114,  32, -10, 3},
+  {0, 0, 3, -15, 111,  37, -11, 3}, {0, 0, 3, -16, 108,  42, -12, 3},
+  {0, 0, 4, -17, 104,  47, -13, 3}, {0, 0, 4, -17, 100,  52, -14, 3},
+  {0, 0, 4, -18,  96,  58, -15, 3}, {0, 0, 4, -18,  91,  63, -16, 4},
+  {0, 0, 4, -18,  87,  68, -17, 4}, {0, 0, 4, -18,  82,  73, -17, 4},
+  {0, 0, 4, -18,  78,  78, -18, 4}, {0, 0, 4, -17,  73,  82, -18, 4},
+  {0, 0, 4, -17,  68,  87, -18, 4}, {0, 0, 4, -16,  63,  91, -18, 4},
+  {0, 0, 3, -15,  58,  96, -18, 4}, {0, 0, 3, -14,  52, 100, -17, 4},
+  {0, 0, 3, -13,  47, 104, -17, 4}, {0, 0, 3, -12,  42, 108, -16, 3},
+  {0, 0, 3, -11,  37, 111, -15, 3}, {0, 0, 3, -10,  32, 114, -14, 3},
+  {0, 0, 2,  -8,  27, 117, -13, 3}, {0, 0, 2,  -7,  22, 120, -11, 2},
+  {0, 0, 1,  -6,  18, 122,  -9, 2}, {0, 0, 1,  -4,  13, 124,  -7, 1},
+  {0, 0, 1,  -3,   8, 126,  -5, 1}, {0, 0, 0,  -1,   4, 127,  -3, 1},
+
+#endif  // WARPEDPIXEL_PREC_BITS == 6
+
+  // dummy
+  { 0, 0, 0, 0,   1, 127, 0, 0 },
+};
+
+/* clang-format on */
+
+#define DIV_LUT_PREC_BITS 14
+#define DIV_LUT_BITS 8
+#define DIV_LUT_NUM (1 << DIV_LUT_BITS)
+
+static const uint16_t div_lut[DIV_LUT_NUM + 1] = {
+  16384, 16320, 16257, 16194, 16132, 16070, 16009, 15948, 15888, 15828, 15768,
+  15709, 15650, 15592, 15534, 15477, 15420, 15364, 15308, 15252, 15197, 15142,
+  15087, 15033, 14980, 14926, 14873, 14821, 14769, 14717, 14665, 14614, 14564,
+  14513, 14463, 14413, 14364, 14315, 14266, 14218, 14170, 14122, 14075, 14028,
+  13981, 13935, 13888, 13843, 13797, 13752, 13707, 13662, 13618, 13574, 13530,
+  13487, 13443, 13400, 13358, 13315, 13273, 13231, 13190, 13148, 13107, 13066,
+  13026, 12985, 12945, 12906, 12866, 12827, 12788, 12749, 12710, 12672, 12633,
+  12596, 12558, 12520, 12483, 12446, 12409, 12373, 12336, 12300, 12264, 12228,
+  12193, 12157, 12122, 12087, 12053, 12018, 11984, 11950, 11916, 11882, 11848,
+  11815, 11782, 11749, 11716, 11683, 11651, 11619, 11586, 11555, 11523, 11491,
+  11460, 11429, 11398, 11367, 11336, 11305, 11275, 11245, 11215, 11185, 11155,
+  11125, 11096, 11067, 11038, 11009, 10980, 10951, 10923, 10894, 10866, 10838,
+  10810, 10782, 10755, 10727, 10700, 10673, 10645, 10618, 10592, 10565, 10538,
+  10512, 10486, 10460, 10434, 10408, 10382, 10356, 10331, 10305, 10280, 10255,
+  10230, 10205, 10180, 10156, 10131, 10107, 10082, 10058, 10034, 10010, 9986,
+  9963,  9939,  9916,  9892,  9869,  9846,  9823,  9800,  9777,  9754,  9732,
+  9709,  9687,  9664,  9642,  9620,  9598,  9576,  9554,  9533,  9511,  9489,
+  9468,  9447,  9425,  9404,  9383,  9362,  9341,  9321,  9300,  9279,  9259,
+  9239,  9218,  9198,  9178,  9158,  9138,  9118,  9098,  9079,  9059,  9039,
+  9020,  9001,  8981,  8962,  8943,  8924,  8905,  8886,  8867,  8849,  8830,
+  8812,  8793,  8775,  8756,  8738,  8720,  8702,  8684,  8666,  8648,  8630,
+  8613,  8595,  8577,  8560,  8542,  8525,  8508,  8490,  8473,  8456,  8439,
+  8422,  8405,  8389,  8372,  8355,  8339,  8322,  8306,  8289,  8273,  8257,
+  8240,  8224,  8208,  8192,
+};
+
+static INLINE int16_t saturate_int16(int32_t v) {
+  if (v > 32767)
+    return 32767;
+  else if (v < -32768)
+    return -32768;
+  return v;
+}
+
+#if CONFIG_WARPED_MOTION
+// Decomposes a divisor D such that 1/D = y/2^shift, where y is returned
+// at precision of DIV_LUT_PREC_BITS along with the shift.
+static int16_t resolve_divisor_64(uint64_t D, int16_t *shift) {
+  int64_t e, f;
+  *shift = (int16_t)((D >> 32) ? get_msb((unsigned int)(D >> 32)) + 32
+                               : get_msb((unsigned int)D));
+  // e is obtained from D after resetting the most significant 1 bit.
+  e = D - ((uint64_t)1 << *shift);
+  // Get the most significant DIV_LUT_BITS (8) bits of e into f
+  if (*shift > DIV_LUT_BITS)
+    f = ROUND_POWER_OF_TWO_64(e, *shift - DIV_LUT_BITS);
+  else
+    f = e << (DIV_LUT_BITS - *shift);
+  assert(f <= DIV_LUT_NUM);
+  *shift += DIV_LUT_PREC_BITS;
+  // Use f as lookup into the precomputed table of multipliers
+  return div_lut[f];
+}
+#endif  // CONFIG_WARPED_MOTION
+
+static int16_t resolve_divisor_32(uint32_t D, int16_t *shift) {
+  int32_t e, f;
+  *shift = get_msb(D);
+  // e is obtained from D after resetting the most significant 1 bit.
+  e = D - ((uint32_t)1 << *shift);
+  // Get the most significant DIV_LUT_BITS (8) bits of e into f
+  if (*shift > DIV_LUT_BITS)
+    f = ROUND_POWER_OF_TWO(e, *shift - DIV_LUT_BITS);
+  else
+    f = e << (DIV_LUT_BITS - *shift);
+  assert(f <= DIV_LUT_NUM);
+  *shift += DIV_LUT_PREC_BITS;
+  // Use f as lookup into the precomputed table of multipliers
+  return div_lut[f];
+}
+
+static int is_affine_valid(WarpedMotionParams *wm) {
+  const int32_t *mat = wm->wmmat;
+  return (mat[2] > 0);
+}
+
+static int is_affine_shear_allowed(int16_t alpha, int16_t beta, int16_t gamma,
+                                   int16_t delta) {
+  if ((4 * abs(alpha) + 7 * abs(beta) >= (1 << WARPEDMODEL_PREC_BITS)) ||
+      (4 * abs(gamma) + 4 * abs(delta) >= (1 << WARPEDMODEL_PREC_BITS)))
+    return 0;
+  else
+    return 1;
+}
+
+// Returns 1 on success or 0 on an invalid affine set
+int get_shear_params(WarpedMotionParams *wm) {
+  const int32_t *mat = wm->wmmat;
+  if (!is_affine_valid(wm)) return 0;
+  wm->alpha =
+      clamp(mat[2] - (1 << WARPEDMODEL_PREC_BITS), INT16_MIN, INT16_MAX);
+  wm->beta = clamp(mat[3], INT16_MIN, INT16_MAX);
+  int16_t shift;
+  int16_t y = resolve_divisor_32(abs(mat[2]), &shift) * (mat[2] < 0 ? -1 : 1);
+  int64_t v;
+  v = ((int64_t)mat[4] * (1 << WARPEDMODEL_PREC_BITS)) * y;
+  wm->gamma =
+      clamp((int)ROUND_POWER_OF_TWO_SIGNED_64(v, shift), INT16_MIN, INT16_MAX);
+  v = ((int64_t)mat[3] * mat[4]) * y;
+  wm->delta = clamp(mat[5] - (int)ROUND_POWER_OF_TWO_SIGNED_64(v, shift) -
+                        (1 << WARPEDMODEL_PREC_BITS),
+                    INT16_MIN, INT16_MAX);
+  if (!is_affine_shear_allowed(wm->alpha, wm->beta, wm->gamma, wm->delta))
+    return 0;
+  return 1;
+}
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE void highbd_get_subcolumn(int taps, uint16_t *ref, int32_t *col,
+                                        int stride, int x, int y_start) {
+  int i;
+  for (i = 0; i < taps; ++i) {
+    col[i] = ref[(i + y_start) * stride + x];
+  }
+}
+
+static uint16_t highbd_bi_ntap_filter(uint16_t *ref, int x, int y, int stride,
+                                      int bd) {
+  int32_t val, arr[WARPEDPIXEL_FILTER_TAPS];
+  int k;
+  int i = (int)x >> WARPEDPIXEL_PREC_BITS;
+  int j = (int)y >> WARPEDPIXEL_PREC_BITS;
+  for (k = 0; k < WARPEDPIXEL_FILTER_TAPS; ++k) {
+    int32_t arr_temp[WARPEDPIXEL_FILTER_TAPS];
+    highbd_get_subcolumn(WARPEDPIXEL_FILTER_TAPS, ref, arr_temp, stride,
+                         i + k + 1 - WARPEDPIXEL_FILTER_TAPS / 2,
+                         j + 1 - WARPEDPIXEL_FILTER_TAPS / 2);
+    arr[k] = do_ntap_filter(arr_temp + WARPEDPIXEL_FILTER_TAPS / 2 - 1,
+                            y - (j * (1 << WARPEDPIXEL_PREC_BITS)));
+  }
+  val = do_ntap_filter(arr + WARPEDPIXEL_FILTER_TAPS / 2 - 1,
+                       x - (i * (1 << WARPEDPIXEL_PREC_BITS)));
+  val = ROUND_POWER_OF_TWO_SIGNED(val, WARPEDPIXEL_FILTER_BITS * 2);
+  return (uint16_t)clip_pixel_highbd(val, bd);
+}
+
+static uint16_t highbd_bi_cubic_filter(uint16_t *ref, int x, int y, int stride,
+                                       int bd) {
+  int32_t val, arr[4];
+  int k;
+  int i = (int)x >> WARPEDPIXEL_PREC_BITS;
+  int j = (int)y >> WARPEDPIXEL_PREC_BITS;
+  for (k = 0; k < 4; ++k) {
+    int32_t arr_temp[4];
+    highbd_get_subcolumn(4, ref, arr_temp, stride, i + k - 1, j - 1);
+    arr[k] =
+        do_cubic_filter(arr_temp + 1, y - (j * (1 << WARPEDPIXEL_PREC_BITS)));
+  }
+  val = do_cubic_filter(arr + 1, x - (i * (1 << WARPEDPIXEL_PREC_BITS)));
+  val = ROUND_POWER_OF_TWO_SIGNED(val, WARPEDPIXEL_FILTER_BITS * 2);
+  return (uint16_t)clip_pixel_highbd(val, bd);
+}
+
+static uint16_t highbd_bi_linear_filter(uint16_t *ref, int x, int y, int stride,
+                                        int bd) {
+  const int ix = x >> WARPEDPIXEL_PREC_BITS;
+  const int iy = y >> WARPEDPIXEL_PREC_BITS;
+  const int sx = x - (ix * (1 << WARPEDPIXEL_PREC_BITS));
+  const int sy = y - (iy * (1 << WARPEDPIXEL_PREC_BITS));
+  int32_t val;
+  val = ROUND_POWER_OF_TWO_SIGNED(
+      ref[iy * stride + ix] * (WARPEDPIXEL_PREC_SHIFTS - sy) *
+              (WARPEDPIXEL_PREC_SHIFTS - sx) +
+          ref[iy * stride + ix + 1] * (WARPEDPIXEL_PREC_SHIFTS - sy) * sx +
+          ref[(iy + 1) * stride + ix] * sy * (WARPEDPIXEL_PREC_SHIFTS - sx) +
+          ref[(iy + 1) * stride + ix + 1] * sy * sx,
+      WARPEDPIXEL_PREC_BITS * 2);
+  return (uint16_t)clip_pixel_highbd(val, bd);
+}
+
+static uint16_t highbd_warp_interpolate(uint16_t *ref, int x, int y, int width,
+                                        int height, int stride, int bd) {
+  int ix = x >> WARPEDPIXEL_PREC_BITS;
+  int iy = y >> WARPEDPIXEL_PREC_BITS;
+  int sx = x - (ix * (1 << WARPEDPIXEL_PREC_BITS));
+  int sy = y - (iy * (1 << WARPEDPIXEL_PREC_BITS));
+  int32_t v;
+
+  if (ix < 0 && iy < 0)
+    return ref[0];
+  else if (ix < 0 && iy > height - 1)
+    return ref[(height - 1) * stride];
+  else if (ix > width - 1 && iy < 0)
+    return ref[width - 1];
+  else if (ix > width - 1 && iy > height - 1)
+    return ref[(height - 1) * stride + (width - 1)];
+  else if (ix < 0) {
+    v = ROUND_POWER_OF_TWO_SIGNED(
+        ref[iy * stride] * (WARPEDPIXEL_PREC_SHIFTS - sy) +
+            ref[(iy + 1) * stride] * sy,
+        WARPEDPIXEL_PREC_BITS);
+    return clip_pixel_highbd(v, bd);
+  } else if (iy < 0) {
+    v = ROUND_POWER_OF_TWO_SIGNED(
+        ref[ix] * (WARPEDPIXEL_PREC_SHIFTS - sx) + ref[ix + 1] * sx,
+        WARPEDPIXEL_PREC_BITS);
+    return clip_pixel_highbd(v, bd);
+  } else if (ix > width - 1) {
+    v = ROUND_POWER_OF_TWO_SIGNED(
+        ref[iy * stride + width - 1] * (WARPEDPIXEL_PREC_SHIFTS - sy) +
+            ref[(iy + 1) * stride + width - 1] * sy,
+        WARPEDPIXEL_PREC_BITS);
+    return clip_pixel_highbd(v, bd);
+  } else if (iy > height - 1) {
+    v = ROUND_POWER_OF_TWO_SIGNED(
+        ref[(height - 1) * stride + ix] * (WARPEDPIXEL_PREC_SHIFTS - sx) +
+            ref[(height - 1) * stride + ix + 1] * sx,
+        WARPEDPIXEL_PREC_BITS);
+    return clip_pixel_highbd(v, bd);
+  } else if (ix >= WARPEDPIXEL_FILTER_TAPS / 2 - 1 &&
+             iy >= WARPEDPIXEL_FILTER_TAPS / 2 - 1 &&
+             ix < width - WARPEDPIXEL_FILTER_TAPS / 2 &&
+             iy < height - WARPEDPIXEL_FILTER_TAPS / 2) {
+    return highbd_bi_ntap_filter(ref, x, y, stride, bd);
+  } else if (ix >= 1 && iy >= 1 && ix < width - 2 && iy < height - 2) {
+    return highbd_bi_cubic_filter(ref, x, y, stride, bd);
+  } else {
+    return highbd_bi_linear_filter(ref, x, y, stride, bd);
+  }
+}
+
+static INLINE int highbd_error_measure(int err, int bd) {
+  const int b = bd - 8;
+  const int bmask = (1 << b) - 1;
+  const int v = (1 << b);
+  int e1, e2;
+  err = abs(err);
+  e1 = err >> b;
+  e2 = err & bmask;
+  return error_measure_lut[255 + e1] * (v - e2) +
+         error_measure_lut[256 + e1] * e2;
+}
+
+static void highbd_warp_plane_old(WarpedMotionParams *wm, uint8_t *ref8,
+                                  int width, int height, int stride,
+                                  uint8_t *pred8, int p_col, int p_row,
+                                  int p_width, int p_height, int p_stride,
+                                  int subsampling_x, int subsampling_y,
+                                  int x_scale, int y_scale, int bd,
+                                  int ref_frm) {
+  int i, j;
+  ProjectPointsFunc projectpoints = get_project_points_type(wm->wmtype);
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  if (projectpoints == NULL) return;
+  for (i = p_row; i < p_row + p_height; ++i) {
+    for (j = p_col; j < p_col + p_width; ++j) {
+      int in[2], out[2];
+      in[0] = j;
+      in[1] = i;
+      projectpoints(wm->wmmat, in, out, 1, 2, 2, subsampling_x, subsampling_y);
+      out[0] = ROUND_POWER_OF_TWO_SIGNED(out[0] * x_scale, 4);
+      out[1] = ROUND_POWER_OF_TWO_SIGNED(out[1] * y_scale, 4);
+      if (ref_frm)
+        pred[(j - p_col) + (i - p_row) * p_stride] = ROUND_POWER_OF_TWO(
+            pred[(j - p_col) + (i - p_row) * p_stride] +
+                highbd_warp_interpolate(ref, out[0], out[1], width, height,
+                                        stride, bd),
+            1);
+      else
+        pred[(j - p_col) + (i - p_row) * p_stride] = highbd_warp_interpolate(
+            ref, out[0], out[1], width, height, stride, bd);
+    }
+  }
+}
+
+// Note: For an explanation of the warp algorithm, see the comment
+// above warp_plane()
+//
+// Note also: The "worst case" in terms of modulus of the data stored into 'tmp'
+// (ie, the result of 'sum' in the horizontal filter) occurs when:
+// coeffs = { -2,   8, -22,  87,  72, -21,   8, -2}, and
+// ref =    {  0, 255,   0, 255, 255,   0, 255,  0}
+// Before rounding, this gives sum = 716625. After rounding,
+// HORSHEAR_REDUCE_PREC_BITS = 4 => sum = 44789 > 2^15
+// HORSHEAR_REDUCE_PREC_BITS = 5 => sum = 22395 < 2^15
+//
+// So, as long as HORSHEAR_REDUCE_PREC_BITS >= 5, we can safely use a 16-bit
+// intermediate array.
+void av1_highbd_warp_affine_c(int32_t *mat, uint16_t *ref, int width,
+                              int height, int stride, uint16_t *pred, int p_col,
+                              int p_row, int p_width, int p_height,
+                              int p_stride, int subsampling_x,
+                              int subsampling_y, int bd, int ref_frm,
+                              int16_t alpha, int16_t beta, int16_t gamma,
+                              int16_t delta) {
+#if HORSHEAR_REDUCE_PREC_BITS >= 5
+  int16_t tmp[15 * 8];
+#else
+  int32_t tmp[15 * 8];
+#endif
+  int i, j, k, l, m;
+
+  /* Note: For this code to work, the left/right frame borders need to be
+     extended by at least 13 pixels each. By the time we get here, other
+     code will have set up this border, but we allow an explicit check
+     for debugging purposes.
+  */
+  /*for (i = 0; i < height; ++i) {
+    for (j = 0; j < 13; ++j) {
+      assert(ref[i * stride - 13 + j] == ref[i * stride]);
+      assert(ref[i * stride + width + j] == ref[i * stride + (width - 1)]);
+    }
+  }*/
+
+  for (i = p_row; i < p_row + p_height; i += 8) {
+    for (j = p_col; j < p_col + p_width; j += 8) {
+      int32_t x4, y4, ix4, sx4, iy4, sy4;
+      if (subsampling_x)
+        x4 = ROUND_POWER_OF_TWO_SIGNED(
+            mat[2] * 2 * (j + 4) + mat[3] * 2 * (i + 4) + mat[0] +
+                (mat[2] + mat[3] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+            1);
+      else
+        x4 = mat[2] * (j + 4) + mat[3] * (i + 4) + mat[0];
+
+      if (subsampling_y)
+        y4 = ROUND_POWER_OF_TWO_SIGNED(
+            mat[4] * 2 * (j + 4) + mat[5] * 2 * (i + 4) + mat[1] +
+                (mat[4] + mat[5] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+            1);
+      else
+        y4 = mat[4] * (j + 4) + mat[5] * (i + 4) + mat[1];
+
+      ix4 = x4 >> WARPEDMODEL_PREC_BITS;
+      sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+      iy4 = y4 >> WARPEDMODEL_PREC_BITS;
+      sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+
+      // Horizontal filter
+      for (k = -7; k < 8; ++k) {
+        int iy = iy4 + k;
+        if (iy < 0)
+          iy = 0;
+        else if (iy > height - 1)
+          iy = height - 1;
+
+        if (ix4 <= -7) {
+          for (l = 0; l < 8; ++l) {
+            tmp[(k + 7) * 8 + l] =
+                ref[iy * stride] *
+                (1 << (WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS));
+          }
+        } else if (ix4 >= width + 6) {
+          for (l = 0; l < 8; ++l) {
+            tmp[(k + 7) * 8 + l] =
+                ref[iy * stride + (width - 1)] *
+                (1 << (WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS));
+          }
+        } else {
+          int sx = sx4 + alpha * (-4) + beta * k;
+
+          for (l = -4; l < 4; ++l) {
+            int ix = ix4 + l - 3;
+            const int offs = ROUND_POWER_OF_TWO(sx, WARPEDDIFF_PREC_BITS) +
+                             WARPEDPIXEL_PREC_SHIFTS;
+            const int16_t *coeffs = warped_filter[offs];
+            int32_t sum = 0;
+            // assert(offs >= 0 && offs <= WARPEDPIXEL_PREC_SHIFTS * 3);
+            for (m = 0; m < 8; ++m) {
+              sum += ref[iy * stride + ix + m] * coeffs[m];
+            }
+            sum = ROUND_POWER_OF_TWO(sum, HORSHEAR_REDUCE_PREC_BITS);
+#if HORSHEAR_REDUCE_PREC_BITS >= 5
+            tmp[(k + 7) * 8 + (l + 4)] = saturate_int16(sum);
+#else
+            tmp[(k + 7) * 8 + (l + 4)] = sum;
+#endif
+            sx += alpha;
+          }
+        }
+      }
+
+      // Vertical filter
+      for (k = -4; k < AOMMIN(4, p_row + p_height - i - 4); ++k) {
+        int sy = sy4 + gamma * (-4) + delta * k;
+        for (l = -4; l < 4; ++l) {
+          uint16_t *p =
+              &pred[(i - p_row + k + 4) * p_stride + (j - p_col + l + 4)];
+          const int offs = ROUND_POWER_OF_TWO(sy, WARPEDDIFF_PREC_BITS) +
+                           WARPEDPIXEL_PREC_SHIFTS;
+          const int16_t *coeffs = warped_filter[offs];
+          int32_t sum = 0;
+          // assert(offs >= 0 && offs <= WARPEDPIXEL_PREC_SHIFTS * 3);
+          for (m = 0; m < 8; ++m) {
+            sum += tmp[(k + m + 4) * 8 + (l + 4)] * coeffs[m];
+          }
+          sum = clip_pixel_highbd(
+              ROUND_POWER_OF_TWO(sum, VERSHEAR_REDUCE_PREC_BITS), bd);
+          if (ref_frm)
+            *p = ROUND_POWER_OF_TWO(*p + sum, 1);
+          else
+            *p = sum;
+          sy += gamma;
+        }
+      }
+    }
+  }
+}
+
+static void highbd_warp_plane(WarpedMotionParams *wm, uint8_t *ref8, int width,
+                              int height, int stride, uint8_t *pred8, int p_col,
+                              int p_row, int p_width, int p_height,
+                              int p_stride, int subsampling_x,
+                              int subsampling_y, int x_scale, int y_scale,
+                              int bd, int ref_frm) {
+  if (wm->wmtype == ROTZOOM) {
+    wm->wmmat[5] = wm->wmmat[2];
+    wm->wmmat[4] = -wm->wmmat[3];
+  }
+  if ((wm->wmtype == ROTZOOM || wm->wmtype == AFFINE) && x_scale == 16 &&
+      y_scale == 16) {
+    int32_t *mat = wm->wmmat;
+    const int16_t alpha = wm->alpha;
+    const int16_t beta = wm->beta;
+    const int16_t gamma = wm->gamma;
+    const int16_t delta = wm->delta;
+
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+    uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+    av1_highbd_warp_affine(mat, ref, width, height, stride, pred, p_col, p_row,
+                           p_width, p_height, p_stride, subsampling_x,
+                           subsampling_y, bd, ref_frm, alpha, beta, gamma,
+                           delta);
+  } else {
+    highbd_warp_plane_old(wm, ref8, width, height, stride, pred8, p_col, p_row,
+                          p_width, p_height, p_stride, subsampling_x,
+                          subsampling_y, x_scale, y_scale, bd, ref_frm);
+  }
+}
+
+static double highbd_warp_erroradv(WarpedMotionParams *wm, uint8_t *ref8,
+                                   int width, int height, int stride,
+                                   uint8_t *dst8, int p_col, int p_row,
+                                   int p_width, int p_height, int p_stride,
+                                   int subsampling_x, int subsampling_y,
+                                   int x_scale, int y_scale, int bd) {
+  int gm_err = 0, no_gm_err = 0;
+  int64_t gm_sumerr = 0, no_gm_sumerr = 0;
+  int i, j;
+  uint16_t *tmp = aom_malloc(p_width * p_height * sizeof(*tmp));
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_warp_plane(wm, ref8, width, height, stride, CONVERT_TO_BYTEPTR(tmp),
+                    p_col, p_row, p_width, p_height, p_width, subsampling_x,
+                    subsampling_y, x_scale, y_scale, bd, 0);
+  for (i = 0; i < p_height; ++i) {
+    for (j = 0; j < p_width; ++j) {
+      gm_err = dst[j + i * p_stride] - tmp[j + i * p_width];
+      no_gm_err =
+          dst[j + i * p_stride] - ref[(j + p_col) + (i + p_row) * stride];
+      gm_sumerr += highbd_error_measure(gm_err, bd);
+      no_gm_sumerr += highbd_error_measure(no_gm_err, bd);
+    }
+  }
+  aom_free(tmp);
+  return (double)gm_sumerr / no_gm_sumerr;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static INLINE int error_measure(int err) {
+  return error_measure_lut[255 + err];
+}
+
+static void warp_plane_old(WarpedMotionParams *wm, uint8_t *ref, int width,
+                           int height, int stride, uint8_t *pred, int p_col,
+                           int p_row, int p_width, int p_height, int p_stride,
+                           int subsampling_x, int subsampling_y, int x_scale,
+                           int y_scale, int ref_frm) {
+  int i, j;
+  ProjectPointsFunc projectpoints = get_project_points_type(wm->wmtype);
+  if (projectpoints == NULL) return;
+  for (i = p_row; i < p_row + p_height; ++i) {
+    for (j = p_col; j < p_col + p_width; ++j) {
+      int in[2], out[2];
+      in[0] = j;
+      in[1] = i;
+      projectpoints(wm->wmmat, in, out, 1, 2, 2, subsampling_x, subsampling_y);
+      out[0] = ROUND_POWER_OF_TWO_SIGNED(out[0] * x_scale, 4);
+      out[1] = ROUND_POWER_OF_TWO_SIGNED(out[1] * y_scale, 4);
+      if (ref_frm)
+        pred[(j - p_col) + (i - p_row) * p_stride] = ROUND_POWER_OF_TWO(
+            pred[(j - p_col) + (i - p_row) * p_stride] +
+                warp_interpolate(ref, out[0], out[1], width, height, stride),
+            1);
+      else
+        pred[(j - p_col) + (i - p_row) * p_stride] =
+            warp_interpolate(ref, out[0], out[1], width, height, stride);
+    }
+  }
+}
+
+/* The warp filter for ROTZOOM and AFFINE models works as follows:
+   * Split the input into 8x8 blocks
+   * For each block, project the point (4, 4) within the block, to get the
+     overall block position. Split into integer and fractional coordinates,
+     maintaining full WARPEDMODEL precision
+   * Filter horizontally: Generate 15 rows of 8 pixels each. Each pixel gets a
+     variable horizontal offset. This means that, while the rows of the
+     intermediate buffer align with the rows of the *reference* image, the
+     columns align with the columns of the *destination* image.
+   * Filter vertically: Generate the output block (up to 8x8 pixels, but if the
+     destination is too small we crop the output at this stage). Each pixel has
+     a variable vertical offset, so that the resulting rows are aligned with
+     the rows of the destination image.
+
+   To accomplish these alignments, we factor the warp matrix as a
+   product of two shear / asymmetric zoom matrices:
+   / a b \  = /   1       0    \ * / 1+alpha  beta \
+   \ c d /    \ gamma  1+delta /   \    0      1   /
+   where a, b, c, d are wmmat[2], wmmat[3], wmmat[4], wmmat[5] respectively.
+   The second shear (with alpha and beta) is applied by the horizontal filter,
+   then the first shear (with gamma and delta) is applied by the vertical
+   filter.
+
+   The only limitation is that, to fit this in a fixed 8-tap filter size,
+   the fractional pixel offsets must be at most +-1. Since the horizontal filter
+   generates 15 rows of 8 columns, and the initial point we project is at (4, 4)
+   within the block, the parameters must satisfy
+   4 * |alpha| + 7 * |beta| <= 1   and   4 * |gamma| + 7 * |delta| <= 1
+   for this filter to be applicable.
+
+   Note: warp_affine() assumes that the caller has done all of the relevant
+   checks, ie. that we have a ROTZOOM or AFFINE model, that wm[4] and wm[5]
+   are set appropriately (if using a ROTZOOM model), and that alpha, beta,
+   gamma, delta are all in range.
+
+   TODO(david.barker): Maybe support scaled references?
+*/
+void av1_warp_affine_c(int32_t *mat, uint8_t *ref, int width, int height,
+                       int stride, uint8_t *pred, int p_col, int p_row,
+                       int p_width, int p_height, int p_stride,
+                       int subsampling_x, int subsampling_y, int ref_frm,
+                       int16_t alpha, int16_t beta, int16_t gamma,
+                       int16_t delta) {
+  int16_t tmp[15 * 8];
+  int i, j, k, l, m;
+
+  /* Note: For this code to work, the left/right frame borders need to be
+     extended by at least 13 pixels each. By the time we get here, other
+     code will have set up this border, but we allow an explicit check
+     for debugging purposes.
+  */
+  /*for (i = 0; i < height; ++i) {
+    for (j = 0; j < 13; ++j) {
+      assert(ref[i * stride - 13 + j] == ref[i * stride]);
+      assert(ref[i * stride + width + j] == ref[i * stride + (width - 1)]);
+    }
+  }*/
+
+  for (i = p_row; i < p_row + p_height; i += 8) {
+    for (j = p_col; j < p_col + p_width; j += 8) {
+      int32_t x4, y4, ix4, sx4, iy4, sy4;
+      if (subsampling_x)
+        x4 = ROUND_POWER_OF_TWO_SIGNED(
+            mat[2] * 2 * (j + 4) + mat[3] * 2 * (i + 4) + mat[0] +
+                (mat[2] + mat[3] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+            1);
+      else
+        x4 = mat[2] * (j + 4) + mat[3] * (i + 4) + mat[0];
+
+      if (subsampling_y)
+        y4 = ROUND_POWER_OF_TWO_SIGNED(
+            mat[4] * 2 * (j + 4) + mat[5] * 2 * (i + 4) + mat[1] +
+                (mat[4] + mat[5] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+            1);
+      else
+        y4 = mat[4] * (j + 4) + mat[5] * (i + 4) + mat[1];
+
+      ix4 = x4 >> WARPEDMODEL_PREC_BITS;
+      sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+      iy4 = y4 >> WARPEDMODEL_PREC_BITS;
+      sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+
+      // Horizontal filter
+      for (k = -7; k < 8; ++k) {
+        int iy = iy4 + k;
+        if (iy < 0)
+          iy = 0;
+        else if (iy > height - 1)
+          iy = height - 1;
+
+        if (ix4 <= -7) {
+          // In this case, the rightmost pixel sampled is in column
+          // ix4 + 3 + 7 - 3 = ix4 + 7 <= 0, ie. the entire block
+          // will sample only from the leftmost column
+          // (once border extension is taken into account)
+          for (l = 0; l < 8; ++l) {
+            tmp[(k + 7) * 8 + l] =
+                ref[iy * stride] *
+                (1 << (WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS));
+          }
+        } else if (ix4 >= width + 6) {
+          // In this case, the leftmost pixel sampled is in column
+          // ix4 - 4 + 0 - 3 = ix4 - 7 >= width - 1, ie. the entire block
+          // will sample only from the rightmost column
+          // (once border extension is taken into account)
+          for (l = 0; l < 8; ++l) {
+            tmp[(k + 7) * 8 + l] =
+                ref[iy * stride + (width - 1)] *
+                (1 << (WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS));
+          }
+        } else {
+          // If we get here, then
+          // the leftmost pixel sampled is
+          // ix4 - 4 + 0 - 3 = ix4 - 7 >= -13
+          // and the rightmost pixel sampled is at most
+          // ix4 + 3 + 7 - 3 = ix4 + 7 <= width + 12
+          // So, assuming that border extension has been done, we
+          // don't need to explicitly clamp values.
+          int sx = sx4 + alpha * (-4) + beta * k;
+
+          for (l = -4; l < 4; ++l) {
+            int ix = ix4 + l - 3;
+            // At this point, sx = sx4 + alpha * l + beta * k
+            const int offs = ROUND_POWER_OF_TWO(sx, WARPEDDIFF_PREC_BITS) +
+                             WARPEDPIXEL_PREC_SHIFTS;
+            const int16_t *coeffs = warped_filter[offs];
+            int32_t sum = 0;
+            // assert(offs >= 0 && offs <= WARPEDPIXEL_PREC_SHIFTS * 3);
+            for (m = 0; m < 8; ++m) {
+              sum += ref[iy * stride + ix + m] * coeffs[m];
+            }
+            sum = ROUND_POWER_OF_TWO(sum, HORSHEAR_REDUCE_PREC_BITS);
+            tmp[(k + 7) * 8 + (l + 4)] = saturate_int16(sum);
+            sx += alpha;
+          }
+        }
+      }
+
+      // Vertical filter
+      for (k = -4; k < AOMMIN(4, p_row + p_height - i - 4); ++k) {
+        int sy = sy4 + gamma * (-4) + delta * k;
+        for (l = -4; l < AOMMIN(4, p_col + p_width - j - 4); ++l) {
+          uint8_t *p =
+              &pred[(i - p_row + k + 4) * p_stride + (j - p_col + l + 4)];
+          // At this point, sy = sy4 + gamma * l + delta * k
+          const int offs = ROUND_POWER_OF_TWO(sy, WARPEDDIFF_PREC_BITS) +
+                           WARPEDPIXEL_PREC_SHIFTS;
+          const int16_t *coeffs = warped_filter[offs];
+          int32_t sum = 0;
+          // assert(offs >= 0 && offs <= WARPEDPIXEL_PREC_SHIFTS * 3);
+          for (m = 0; m < 8; ++m) {
+            sum += tmp[(k + m + 4) * 8 + (l + 4)] * coeffs[m];
+          }
+          sum = clip_pixel(ROUND_POWER_OF_TWO(sum, VERSHEAR_REDUCE_PREC_BITS));
+          if (ref_frm)
+            *p = ROUND_POWER_OF_TWO(*p + sum, 1);
+          else
+            *p = sum;
+          sy += gamma;
+        }
+      }
+    }
+  }
+}
+
+static void warp_plane(WarpedMotionParams *wm, uint8_t *ref, int width,
+                       int height, int stride, uint8_t *pred, int p_col,
+                       int p_row, int p_width, int p_height, int p_stride,
+                       int subsampling_x, int subsampling_y, int x_scale,
+                       int y_scale, int ref_frm) {
+  if (wm->wmtype == ROTZOOM) {
+    wm->wmmat[5] = wm->wmmat[2];
+    wm->wmmat[4] = -wm->wmmat[3];
+  }
+  if ((wm->wmtype == ROTZOOM || wm->wmtype == AFFINE) && x_scale == 16 &&
+      y_scale == 16) {
+    int32_t *mat = wm->wmmat;
+    const int16_t alpha = wm->alpha;
+    const int16_t beta = wm->beta;
+    const int16_t gamma = wm->gamma;
+    const int16_t delta = wm->delta;
+
+    av1_warp_affine(mat, ref, width, height, stride, pred, p_col, p_row,
+                    p_width, p_height, p_stride, subsampling_x, subsampling_y,
+                    ref_frm, alpha, beta, gamma, delta);
+  } else {
+    warp_plane_old(wm, ref, width, height, stride, pred, p_col, p_row, p_width,
+                   p_height, p_stride, subsampling_x, subsampling_y, x_scale,
+                   y_scale, ref_frm);
+  }
+}
+
+static double warp_erroradv(WarpedMotionParams *wm, uint8_t *ref, int width,
+                            int height, int stride, uint8_t *dst, int p_col,
+                            int p_row, int p_width, int p_height, int p_stride,
+                            int subsampling_x, int subsampling_y, int x_scale,
+                            int y_scale) {
+  int gm_err = 0, no_gm_err = 0;
+  int gm_sumerr = 0, no_gm_sumerr = 0;
+  int i, j;
+  uint8_t *tmp = aom_malloc(p_width * p_height);
+  warp_plane(wm, ref, width, height, stride, tmp, p_col, p_row, p_width,
+             p_height, p_width, subsampling_x, subsampling_y, x_scale, y_scale,
+             0);
+
+  for (i = 0; i < p_height; ++i) {
+    for (j = 0; j < p_width; ++j) {
+      gm_err = dst[j + i * p_stride] - tmp[j + i * p_width];
+      no_gm_err =
+          dst[j + i * p_stride] - ref[(j + p_col) + (i + p_row) * stride];
+      gm_sumerr += error_measure(gm_err);
+      no_gm_sumerr += error_measure(no_gm_err);
+    }
+  }
+
+  aom_free(tmp);
+  return (double)gm_sumerr / no_gm_sumerr;
+}
+
+double av1_warp_erroradv(WarpedMotionParams *wm,
+#if CONFIG_HIGHBITDEPTH
+                         int use_hbd, int bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                         uint8_t *ref, int width, int height, int stride,
+                         uint8_t *dst, int p_col, int p_row, int p_width,
+                         int p_height, int p_stride, int subsampling_x,
+                         int subsampling_y, int x_scale, int y_scale) {
+  if (wm->wmtype <= AFFINE)
+    if (!get_shear_params(wm)) return 1;
+#if CONFIG_HIGHBITDEPTH
+  if (use_hbd)
+    return highbd_warp_erroradv(
+        wm, ref, width, height, stride, dst, p_col, p_row, p_width, p_height,
+        p_stride, subsampling_x, subsampling_y, x_scale, y_scale, bd);
+#endif  // CONFIG_HIGHBITDEPTH
+  return warp_erroradv(wm, ref, width, height, stride, dst, p_col, p_row,
+                       p_width, p_height, p_stride, subsampling_x,
+                       subsampling_y, x_scale, y_scale);
+}
+
+void av1_warp_plane(WarpedMotionParams *wm,
+#if CONFIG_HIGHBITDEPTH
+                    int use_hbd, int bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                    uint8_t *ref, int width, int height, int stride,
+                    uint8_t *pred, int p_col, int p_row, int p_width,
+                    int p_height, int p_stride, int subsampling_x,
+                    int subsampling_y, int x_scale, int y_scale, int ref_frm) {
+#if CONFIG_HIGHBITDEPTH
+  if (use_hbd)
+    highbd_warp_plane(wm, ref, width, height, stride, pred, p_col, p_row,
+                      p_width, p_height, p_stride, subsampling_x, subsampling_y,
+                      x_scale, y_scale, bd, ref_frm);
+  else
+#endif  // CONFIG_HIGHBITDEPTH
+    warp_plane(wm, ref, width, height, stride, pred, p_col, p_row, p_width,
+               p_height, p_stride, subsampling_x, subsampling_y, x_scale,
+               y_scale, ref_frm);
+}
+
+#if CONFIG_WARPED_MOTION
+#define LEAST_SQUARES_ORDER 2
+
+#define LS_MV_MAX 256  // max mv in 1/8-pel
+#define LS_STEP 2
+
+// Assuming LS_MV_MAX is < MAX_SB_SIZE * 8,
+// the precision needed is:
+//   (MAX_SB_SIZE_LOG2 + 3) [for sx * sx magnitude] +
+//   (MAX_SB_SIZE_LOG2 + 4) [for sx * dx magnitude] +
+//   1 [for sign] +
+//   LEAST_SQUARES_SAMPLES_MAX_BITS
+//        [for adding up to LEAST_SQUARES_SAMPLES_MAX samples]
+// The value is 23
+#define LS_MAT_RANGE_BITS \
+  ((MAX_SB_SIZE_LOG2 + 4) * 2 + LEAST_SQUARES_SAMPLES_MAX_BITS)
+
+// Bit-depth reduction from the full-range
+#define LS_MAT_DOWN_BITS 2
+
+// bits range of A, Bx and By after downshifting
+#define LS_MAT_BITS (LS_MAT_RANGE_BITS - LS_MAT_DOWN_BITS)
+#define LS_MAT_MIN (-(1 << (LS_MAT_BITS - 1)))
+#define LS_MAT_MAX ((1 << (LS_MAT_BITS - 1)) - 1)
+
+#define LS_SUM(a) ((a)*4 + LS_STEP * 2)
+#define LS_SQUARE(a) \
+  (((a) * (a)*4 + (a)*4 * LS_STEP + LS_STEP * LS_STEP * 2) >> 2)
+#define LS_PRODUCT1(a, b) \
+  (((a) * (b)*4 + ((a) + (b)) * 2 * LS_STEP + LS_STEP * LS_STEP) >> 2)
+#define LS_PRODUCT2(a, b) \
+  (((a) * (b)*4 + ((a) + (b)) * 2 * LS_STEP + LS_STEP * LS_STEP * 2) >> 2)
+
+#if LEAST_SQUARES_ORDER == 2
+static int find_affine_int(int np, int *pts1, int *pts2, BLOCK_SIZE bsize,
+                           int mvy, int mvx, WarpedMotionParams *wm, int mi_row,
+                           int mi_col) {
+  int32_t A[2][2] = { { 0, 0 }, { 0, 0 } };
+  int32_t Bx[2] = { 0, 0 };
+  int32_t By[2] = { 0, 0 };
+  int i, n = 0;
+
+  const int bw = block_size_wide[bsize];
+  const int bh = block_size_high[bsize];
+  const int suy = (mi_row * MI_SIZE + AOMMAX(bh, MI_SIZE) / 2 - 1) * 8;
+  const int sux = (mi_col * MI_SIZE + AOMMAX(bw, MI_SIZE) / 2 - 1) * 8;
+  const int duy = suy + mvy;
+  const int dux = sux + mvx;
+
+  // Assume the center pixel of the block has exactly the same motion vector
+  // as transmitted for the block. First shift the origin of the source
+  // points to the block center, and the origin of the destination points to
+  // the block center added to the motion vector transmitted.
+  // Let (xi, yi) denote the source points and (xi', yi') denote destination
+  // points after origin shfifting, for i = 0, 1, 2, .... n-1.
+  // Then if  P = [x0, y0,
+  //               x1, y1
+  //               x2, y1,
+  //                ....
+  //              ]
+  //          q = [x0', x1', x2', ... ]'
+  //          r = [y0', y1', y2', ... ]'
+  // the least squares problems that need to be solved are:
+  //          [h1, h2]' = inv(P'P)P'q and
+  //          [h3, h4]' = inv(P'P)P'r
+  // where the affine transformation is given by:
+  //          x' = h1.x + h2.y
+  //          y' = h3.x + h4.y
+  //
+  // The loop below computes: A = P'P, Bx = P'q, By = P'r
+  // We need to just compute inv(A).Bx and inv(A).By for the solutions.
+  int sx, sy, dx, dy;
+  // Contribution from neighbor block
+  for (i = 0; i < np && n < LEAST_SQUARES_SAMPLES_MAX; i++) {
+    dx = pts2[i * 2] - dux;
+    dy = pts2[i * 2 + 1] - duy;
+    sx = pts1[i * 2] - sux;
+    sy = pts1[i * 2 + 1] - suy;
+    if (abs(sx - dx) < LS_MV_MAX && abs(sy - dy) < LS_MV_MAX) {
+      A[0][0] += LS_SQUARE(sx);
+      A[0][1] += LS_PRODUCT1(sx, sy);
+      A[1][1] += LS_SQUARE(sy);
+      Bx[0] += LS_PRODUCT2(sx, dx);
+      Bx[1] += LS_PRODUCT1(sy, dx);
+      By[0] += LS_PRODUCT1(sx, dy);
+      By[1] += LS_PRODUCT2(sy, dy);
+      n++;
+    }
+  }
+  int downshift;
+  if (n >= 4)
+    downshift = LS_MAT_DOWN_BITS;
+  else if (n >= 2)
+    downshift = LS_MAT_DOWN_BITS - 1;
+  else
+    downshift = LS_MAT_DOWN_BITS - 2;
+
+  // Reduce precision by downshift bits
+  A[0][0] = clamp(ROUND_POWER_OF_TWO_SIGNED(A[0][0], downshift), LS_MAT_MIN,
+                  LS_MAT_MAX);
+  A[0][1] = clamp(ROUND_POWER_OF_TWO_SIGNED(A[0][1], downshift), LS_MAT_MIN,
+                  LS_MAT_MAX);
+  A[1][1] = clamp(ROUND_POWER_OF_TWO_SIGNED(A[1][1], downshift), LS_MAT_MIN,
+                  LS_MAT_MAX);
+  Bx[0] = clamp(ROUND_POWER_OF_TWO_SIGNED(Bx[0], downshift), LS_MAT_MIN,
+                LS_MAT_MAX);
+  Bx[1] = clamp(ROUND_POWER_OF_TWO_SIGNED(Bx[1], downshift), LS_MAT_MIN,
+                LS_MAT_MAX);
+  By[0] = clamp(ROUND_POWER_OF_TWO_SIGNED(By[0], downshift), LS_MAT_MIN,
+                LS_MAT_MAX);
+  By[1] = clamp(ROUND_POWER_OF_TWO_SIGNED(By[1], downshift), LS_MAT_MIN,
+                LS_MAT_MAX);
+
+  int64_t Px[2], Py[2], Det;
+  int16_t iDet, shift;
+
+  // These divided by the Det, are the least squares solutions
+  Px[0] = (int64_t)A[1][1] * Bx[0] - (int64_t)A[0][1] * Bx[1];
+  Px[1] = -(int64_t)A[0][1] * Bx[0] + (int64_t)A[0][0] * Bx[1];
+  Py[0] = (int64_t)A[1][1] * By[0] - (int64_t)A[0][1] * By[1];
+  Py[1] = -(int64_t)A[0][1] * By[0] + (int64_t)A[0][0] * By[1];
+
+  // Compute Determinant of A
+  Det = (int64_t)A[0][0] * A[1][1] - (int64_t)A[0][1] * A[0][1];
+  if (Det == 0) return 1;
+  iDet = resolve_divisor_64(llabs(Det), &shift) * (Det < 0 ? -1 : 1);
+  shift -= WARPEDMODEL_PREC_BITS;
+  if (shift < 0) {
+    iDet <<= (-shift);
+    shift = 0;
+  }
+
+  int64_t v;
+  v = Px[0] * (int64_t)iDet;
+  wm->wmmat[2] = (int32_t)(ROUND_POWER_OF_TWO_SIGNED_64(v, shift));
+  v = Px[1] * (int64_t)iDet;
+  wm->wmmat[3] = (int32_t)(ROUND_POWER_OF_TWO_SIGNED_64(v, shift));
+  v = ((int64_t)dux * (1 << WARPEDMODEL_PREC_BITS)) -
+      (int64_t)sux * wm->wmmat[2] - (int64_t)suy * wm->wmmat[3];
+  wm->wmmat[0] = (int32_t)(ROUND_POWER_OF_TWO_SIGNED(v, 3));
+
+  v = Py[0] * (int64_t)iDet;
+  wm->wmmat[4] = (int32_t)(ROUND_POWER_OF_TWO_SIGNED_64(v, shift));
+  v = Py[1] * (int64_t)iDet;
+  wm->wmmat[5] = (int32_t)(ROUND_POWER_OF_TWO_SIGNED_64(v, shift));
+  v = ((int64_t)duy * (1 << WARPEDMODEL_PREC_BITS)) -
+      (int64_t)sux * wm->wmmat[4] - (int64_t)suy * wm->wmmat[5];
+  wm->wmmat[1] = (int32_t)(ROUND_POWER_OF_TWO_SIGNED(v, 3));
+
+  wm->wmmat[6] = wm->wmmat[7] = 0;
+
+  // Clamp values
+  wm->wmmat[0] = clamp(wm->wmmat[0], -WARPEDMODEL_TRANS_CLAMP,
+                       WARPEDMODEL_TRANS_CLAMP - 1);
+  wm->wmmat[1] = clamp(wm->wmmat[1], -WARPEDMODEL_TRANS_CLAMP,
+                       WARPEDMODEL_TRANS_CLAMP - 1);
+  wm->wmmat[2] = clamp(wm->wmmat[2], -WARPEDMODEL_DIAGAFFINE_CLAMP,
+                       WARPEDMODEL_DIAGAFFINE_CLAMP - 1);
+  wm->wmmat[5] = clamp(wm->wmmat[5], -WARPEDMODEL_DIAGAFFINE_CLAMP,
+                       WARPEDMODEL_DIAGAFFINE_CLAMP - 1);
+  wm->wmmat[3] = clamp(wm->wmmat[3], -WARPEDMODEL_NONDIAGAFFINE_CLAMP,
+                       WARPEDMODEL_NONDIAGAFFINE_CLAMP - 1);
+  wm->wmmat[4] = clamp(wm->wmmat[4], -WARPEDMODEL_NONDIAGAFFINE_CLAMP,
+                       WARPEDMODEL_NONDIAGAFFINE_CLAMP - 1);
+  return 0;
+}
+
+#else
+
+static int find_affine_int(int np, int *pts1, int *pts2, BLOCK_SIZE bsize,
+                           int mvy, int mvx, WarpedMotionParams *wm, int mi_row,
+                           int mi_col) {
+  int32_t A[3][3] = { { 0, 0, 0 }, { 0, 0, 0 }, { 0, 0, 0 } };
+  int32_t Bx[3] = { 0, 0, 0 };
+  int32_t By[3] = { 0, 0, 0 };
+  int i, n = 0, off;
+
+  int64_t C00, C01, C02, C11, C12, C22;
+  int64_t Px[3], Py[3];
+  int64_t Det, v;
+  const int bw = block_size_wide[bsize];
+  const int bh = block_size_high[bsize];
+  const int cy_offset = AOMMAX(bh, MI_SIZE) / 2 - 1;
+  const int cx_offset = AOMMAX(bw, MI_SIZE) / 2 - 1;
+
+  // Offsets to make the values in the arrays smaller
+  const int ux = mi_col * MI_SIZE * 8, uy = mi_row * MI_SIZE * 8;
+  // Let source points (xi, yi) map to destimation points (xi', yi'),
+  //     for i = 0, 1, 2, .... n-1
+  // Then if  P = [x0, y0, 1,
+  //               x1, y1, 1
+  //               x2, y2, 1,
+  //                ....
+  //              ]
+  //          q = [x0', x1', x2', ... ]'
+  //          r = [y0', y1', y2', ... ]'
+  // the least squares problems that need to be solved are:
+  //          [h1, h2, dx]' = inv(P'P)P'q and
+  //          [h3, h4, dy]' = inv(P'P)P'r
+  // where the affine transformation is given by:
+  //          x' = h1.x + h2.y + dx
+  //          y' = h3.x + h4.y + dy
+  //
+  // The loop below computes: A = P'P, Bx = P'q, By = P'r
+  // We need to just compute inv(A).Bx and inv(A).By for the solutions.
+  //
+  int sx, sy, dx, dy;
+  // Contribution from sample in current block
+  sx = cx_offset * 8;
+  sy = cy_offset * 8;
+  dx = sx + mvx;
+  dy = sy + mvy;
+  if (abs(sx - dx) < LS_MV_MAX && abs(sy - dy) < LS_MV_MAX) {
+    A[0][0] += LS_SQUARE(sx);
+    A[0][1] += LS_PRODUCT1(sx, sy);
+    A[0][2] += LS_SUM(sx);
+    A[1][1] += LS_SQUARE(sy);
+    A[1][2] += LS_SUM(sy);
+    A[2][2] += 4;
+    Bx[0] += LS_PRODUCT2(sx, dx);
+    Bx[1] += LS_PRODUCT1(sy, dx);
+    Bx[2] += LS_SUM(dx);
+    By[0] += LS_PRODUCT1(sx, dy);
+    By[1] += LS_PRODUCT2(sy, dy);
+    By[2] += LS_SUM(dy);
+    n++;
+  }
+  // Contribution from neighbor block
+  for (i = 0; i < np && n < LEAST_SQUARES_SAMPLES_MAX; i++) {
+    dx = pts2[i * 2] - ux;
+    dy = pts2[i * 2 + 1] - uy;
+    sx = pts1[i * 2] - ux;
+    sy = pts1[i * 2 + 1] - uy;
+    if (abs(sx - dx) < LS_MV_MAX && abs(sy - dy) < LS_MV_MAX) {
+      A[0][0] += LS_SQUARE(sx);
+      A[0][1] += LS_PRODUCT1(sx, sy);
+      A[0][2] += LS_SUM(sx);
+      A[1][1] += LS_SQUARE(sy);
+      A[1][2] += LS_SUM(sy);
+      A[2][2] += 4;
+      Bx[0] += LS_PRODUCT2(sx, dx);
+      Bx[1] += LS_PRODUCT1(sy, dx);
+      Bx[2] += LS_SUM(dx);
+      By[0] += LS_PRODUCT1(sx, dy);
+      By[1] += LS_PRODUCT2(sy, dy);
+      By[2] += LS_SUM(dy);
+      n++;
+    }
+  }
+  // Compute Cofactors of A
+  C00 = (int64_t)A[1][1] * A[2][2] - (int64_t)A[1][2] * A[1][2];
+  C01 = (int64_t)A[1][2] * A[0][2] - (int64_t)A[0][1] * A[2][2];
+  C02 = (int64_t)A[0][1] * A[1][2] - (int64_t)A[0][2] * A[1][1];
+  C11 = (int64_t)A[0][0] * A[2][2] - (int64_t)A[0][2] * A[0][2];
+  C12 = (int64_t)A[0][1] * A[0][2] - (int64_t)A[0][0] * A[1][2];
+  C22 = (int64_t)A[0][0] * A[1][1] - (int64_t)A[0][1] * A[0][1];
+
+  // Scale by 1/64
+  C00 = ROUND_POWER_OF_TWO_SIGNED(C00, 6);
+  C01 = ROUND_POWER_OF_TWO_SIGNED(C01, 6);
+  C02 = ROUND_POWER_OF_TWO_SIGNED(C02, 6);
+  C11 = ROUND_POWER_OF_TWO_SIGNED(C11, 6);
+  C12 = ROUND_POWER_OF_TWO_SIGNED(C12, 6);
+  C22 = ROUND_POWER_OF_TWO_SIGNED(C22, 6);
+
+  // Compute Determinant of A
+  Det = C00 * A[0][0] + C01 * A[0][1] + C02 * A[0][2];
+  if (Det == 0) return 1;
+
+  // These divided by the Det, are the least squares solutions
+  Px[0] = C00 * Bx[0] + C01 * Bx[1] + C02 * Bx[2];
+  Px[1] = C01 * Bx[0] + C11 * Bx[1] + C12 * Bx[2];
+  Px[2] = C02 * Bx[0] + C12 * Bx[1] + C22 * Bx[2];
+  Py[0] = C00 * By[0] + C01 * By[1] + C02 * By[2];
+  Py[1] = C01 * By[0] + C11 * By[1] + C12 * By[2];
+  Py[2] = C02 * By[0] + C12 * By[1] + C22 * By[2];
+
+  int16_t shift;
+  int64_t iDet;
+  iDet = resolve_divisor_64(llabs(Det), &shift) * (Det < 0 ? -1 : 1);
+  shift -= WARPEDMODEL_PREC_BITS;
+  if (shift < 0) {
+    iDet <<= (-shift);
+    shift = 0;
+  }
+
+  v = Px[0] * iDet;
+  wm->wmmat[2] = ROUND_POWER_OF_TWO_SIGNED_64(v, shift);
+  v = Px[1] * iDet;
+  wm->wmmat[3] = ROUND_POWER_OF_TWO_SIGNED_64(v, shift);
+  v = Px[2] * iDet;
+  wm->wmmat[0] = ROUND_POWER_OF_TWO_SIGNED_64(v, shift + 3);
+  // Adjust x displacement for the offset
+  off = (ux << WARPEDMODEL_PREC_BITS) - ux * wm->wmmat[2] - uy * wm->wmmat[3];
+  wm->wmmat[0] += ROUND_POWER_OF_TWO_SIGNED(off, 3);
+
+  v = Py[0] * iDet;
+  wm->wmmat[4] = ROUND_POWER_OF_TWO_SIGNED_64(v, shift);
+  v = Py[1] * iDet;
+  wm->wmmat[5] = ROUND_POWER_OF_TWO_SIGNED_64(v, shift);
+  v = Py[2] * iDet;
+  wm->wmmat[1] = ROUND_POWER_OF_TWO_SIGNED_64(v, shift + 3);
+  // Adjust y displacement for the offset
+  off = (uy << WARPEDMODEL_PREC_BITS) - ux * wm->wmmat[4] - uy * wm->wmmat[5];
+  wm->wmmat[1] += ROUND_POWER_OF_TWO_SIGNED(off, 3);
+  wm->wmmat[6] = wm->wmmat[7] = 0;
+
+  // Clamp values
+  wm->wmmat[0] = clamp(wm->wmmat[0], -WARPEDMODEL_TRANS_CLAMP,
+                       WARPEDMODEL_TRANS_CLAMP - 1);
+  wm->wmmat[1] = clamp(wm->wmmat[1], -WARPEDMODEL_TRANS_CLAMP,
+                       WARPEDMODEL_TRANS_CLAMP - 1);
+  wm->wmmat[2] = clamp(wm->wmmat[2], -WARPEDMODEL_DIAGAFFINE_CLAMP,
+                       WARPEDMODEL_DIAGAFFINE_CLAMP - 1);
+  wm->wmmat[5] = clamp(wm->wmmat[5], -WARPEDMODEL_DIAGAFFINE_CLAMP,
+                       WARPEDMODEL_DIAGAFFINE_CLAMP - 1);
+  wm->wmmat[3] = clamp(wm->wmmat[3], -WARPEDMODEL_NONDIAGAFFINE_CLAMP,
+                       WARPEDMODEL_NONDIAGAFFINE_CLAMP - 1);
+  wm->wmmat[4] = clamp(wm->wmmat[4], -WARPEDMODEL_NONDIAGAFFINE_CLAMP,
+                       WARPEDMODEL_NONDIAGAFFINE_CLAMP - 1);
+
+  return 0;
+}
+#endif  // LEAST_SQUARES_ORDER == 2
+
+int find_projection(int np, int *pts1, int *pts2, BLOCK_SIZE bsize, int mvy,
+                    int mvx, WarpedMotionParams *wm_params, int mi_row,
+                    int mi_col) {
+  int result = 1;
+  switch (wm_params->wmtype) {
+    case AFFINE:
+      result = find_affine_int(np, pts1, pts2, bsize, mvy, mvx, wm_params,
+                               mi_row, mi_col);
+      break;
+    default: assert(0 && "Invalid warped motion type!"); return 1;
+  }
+  if (result == 0) {
+    if (wm_params->wmtype == ROTZOOM) {
+      wm_params->wmmat[5] = wm_params->wmmat[2];
+      wm_params->wmmat[4] = -wm_params->wmmat[3];
+    }
+    if (wm_params->wmtype == AFFINE || wm_params->wmtype == ROTZOOM) {
+      // check compatibility with the fast warp filter
+      if (!get_shear_params(wm_params)) return 1;
+    }
+  }
+
+  return result;
+}
+#endif  // CONFIG_WARPED_MOTION
diff --git a/third_party/aom/av1/common/warped_motion.h b/third_party/aom/av1/common/warped_motion.h
new file mode 100644
index 0000000000..dfd8dae34b
--- /dev/null
+++ b/third_party/aom/av1/common/warped_motion.h
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_COMMON_WARPED_MOTION_H_
+#define AV1_COMMON_WARPED_MOTION_H_
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+#include <math.h>
+#include <assert.h>
+
+#include "./aom_config.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/common/mv.h"
+
+#define MAX_PARAMDIM 9
+#if CONFIG_WARPED_MOTION
+#define SAMPLES_ARRAY_SIZE ((2 * MAX_MIB_SIZE + 2) * 2)
+
+#define LEAST_SQUARES_SAMPLES_MAX_BITS 3
+#define LEAST_SQUARES_SAMPLES_MAX (1 << LEAST_SQUARES_SAMPLES_MAX_BITS)
+
+#define DEFAULT_WMTYPE AFFINE
+#endif  // CONFIG_WARPED_MOTION
+
+extern const int16_t warped_filter[WARPEDPIXEL_PREC_SHIFTS * 3 + 1][8];
+
+typedef void (*ProjectPointsFunc)(int32_t *mat, int *points, int *proj,
+                                  const int n, const int stride_points,
+                                  const int stride_proj,
+                                  const int subsampling_x,
+                                  const int subsampling_y);
+
+void project_points_translation(int32_t *mat, int *points, int *proj,
+                                const int n, const int stride_points,
+                                const int stride_proj, const int subsampling_x,
+                                const int subsampling_y);
+
+void project_points_rotzoom(int32_t *mat, int *points, int *proj, const int n,
+                            const int stride_points, const int stride_proj,
+                            const int subsampling_x, const int subsampling_y);
+
+void project_points_affine(int32_t *mat, int *points, int *proj, const int n,
+                           const int stride_points, const int stride_proj,
+                           const int subsampling_x, const int subsampling_y);
+
+void project_points_hortrapezoid(int32_t *mat, int *points, int *proj,
+                                 const int n, const int stride_points,
+                                 const int stride_proj, const int subsampling_x,
+                                 const int subsampling_y);
+void project_points_vertrapezoid(int32_t *mat, int *points, int *proj,
+                                 const int n, const int stride_points,
+                                 const int stride_proj, const int subsampling_x,
+                                 const int subsampling_y);
+void project_points_homography(int32_t *mat, int *points, int *proj,
+                               const int n, const int stride_points,
+                               const int stride_proj, const int subsampling_x,
+                               const int subsampling_y);
+
+void project_points(WarpedMotionParams *wm_params, int *points, int *proj,
+                    const int n, const int stride_points, const int stride_proj,
+                    const int subsampling_x, const int subsampling_y);
+
+double av1_warp_erroradv(WarpedMotionParams *wm,
+#if CONFIG_HIGHBITDEPTH
+                         int use_hbd, int bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                         uint8_t *ref, int width, int height, int stride,
+                         uint8_t *dst, int p_col, int p_row, int p_width,
+                         int p_height, int p_stride, int subsampling_x,
+                         int subsampling_y, int x_scale, int y_scale);
+
+void av1_warp_plane(WarpedMotionParams *wm,
+#if CONFIG_HIGHBITDEPTH
+                    int use_hbd, int bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                    uint8_t *ref, int width, int height, int stride,
+                    uint8_t *pred, int p_col, int p_row, int p_width,
+                    int p_height, int p_stride, int subsampling_x,
+                    int subsampling_y, int x_scale, int y_scale, int ref_frm);
+
+int find_projection(int np, int *pts1, int *pts2, BLOCK_SIZE bsize, int mvy,
+                    int mvx, WarpedMotionParams *wm_params, int mi_row,
+                    int mi_col);
+
+int get_shear_params(WarpedMotionParams *wm);
+#endif  // AV1_COMMON_WARPED_MOTION_H_
diff --git a/third_party/aom/av1/common/x86/av1_convolve_ssse3.c b/third_party/aom/av1/common/x86/av1_convolve_ssse3.c
new file mode 100644
index 0000000000..91102bbaf5
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_convolve_ssse3.c
@@ -0,0 +1,1029 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <tmmintrin.h>
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "av1/common/filter.h"
+
+#define WIDTH_BOUND (16)
+#define HEIGHT_BOUND (16)
+
+#if CONFIG_DUAL_FILTER
+DECLARE_ALIGNED(16, static int8_t,
+                sub_pel_filters_12sharp_signal_dir[15][2][16]);
+
+DECLARE_ALIGNED(16, static int8_t,
+                sub_pel_filters_12sharp_ver_signal_dir[15][6][16]);
+#endif  // CONFIG_DUAL_FILTER
+
+#if USE_TEMPORALFILTER_12TAP
+DECLARE_ALIGNED(16, static int8_t,
+                sub_pel_filters_temporalfilter_12_signal_dir[15][2][16]);
+
+DECLARE_ALIGNED(16, static int8_t,
+                sub_pel_filters_temporalfilter_12_ver_signal_dir[15][6][16]);
+#endif
+
+typedef int8_t (*SubpelFilterCoeffs)[16];
+
+static INLINE SubpelFilterCoeffs
+get_subpel_filter_signal_dir(const InterpFilterParams p, int index) {
+#if CONFIG_DUAL_FILTER
+  if (p.interp_filter == MULTITAP_SHARP) {
+    return &sub_pel_filters_12sharp_signal_dir[index][0];
+  }
+#endif
+#if USE_TEMPORALFILTER_12TAP
+  if (p.interp_filter == TEMPORALFILTER_12TAP) {
+    return &sub_pel_filters_temporalfilter_12_signal_dir[index][0];
+  }
+#endif
+  (void)p;
+  (void)index;
+  return NULL;
+}
+
+static INLINE SubpelFilterCoeffs
+get_subpel_filter_ver_signal_dir(const InterpFilterParams p, int index) {
+#if CONFIG_DUAL_FILTER
+  if (p.interp_filter == MULTITAP_SHARP) {
+    return &sub_pel_filters_12sharp_ver_signal_dir[index][0];
+  }
+#endif
+#if USE_TEMPORALFILTER_12TAP
+  if (p.interp_filter == TEMPORALFILTER_12TAP) {
+    return &sub_pel_filters_temporalfilter_12_ver_signal_dir[index][0];
+  }
+#endif
+  (void)p;
+  (void)index;
+  return NULL;
+}
+
+static INLINE void transpose_4x8(const __m128i *in, __m128i *out) {
+  __m128i t0, t1;
+
+  t0 = _mm_unpacklo_epi16(in[0], in[1]);
+  t1 = _mm_unpacklo_epi16(in[2], in[3]);
+
+  out[0] = _mm_unpacklo_epi32(t0, t1);
+  out[1] = _mm_srli_si128(out[0], 8);
+  out[2] = _mm_unpackhi_epi32(t0, t1);
+  out[3] = _mm_srli_si128(out[2], 8);
+
+  t0 = _mm_unpackhi_epi16(in[0], in[1]);
+  t1 = _mm_unpackhi_epi16(in[2], in[3]);
+
+  out[4] = _mm_unpacklo_epi32(t0, t1);
+  out[5] = _mm_srli_si128(out[4], 8);
+  // Note: We ignore out[6] and out[7] because
+  // they're zero vectors.
+}
+
+typedef void (*store_pixel_t)(const __m128i *x, uint8_t *dst);
+
+static INLINE __m128i accumulate_store(const __m128i *x, uint8_t *src) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i y = _mm_loadl_epi64((__m128i const *)src);
+  y = _mm_unpacklo_epi8(y, zero);
+  y = _mm_add_epi16(*x, y);
+  y = _mm_add_epi16(y, one);
+  y = _mm_srai_epi16(y, 1);
+  y = _mm_packus_epi16(y, y);
+  return y;
+}
+
+static INLINE void store_2_pixel_only(const __m128i *x, uint8_t *dst) {
+  uint32_t temp;
+  __m128i u = _mm_packus_epi16(*x, *x);
+  temp = _mm_cvtsi128_si32(u);
+  *(uint16_t *)dst = (uint16_t)temp;
+}
+
+static INLINE void accumulate_store_2_pixel(const __m128i *x, uint8_t *dst) {
+  uint32_t temp;
+  __m128i y = accumulate_store(x, dst);
+  temp = _mm_cvtsi128_si32(y);
+  *(uint16_t *)dst = (uint16_t)temp;
+}
+
+static store_pixel_t store2pixelTab[2] = { store_2_pixel_only,
+                                           accumulate_store_2_pixel };
+
+static INLINE void store_4_pixel_only(const __m128i *x, uint8_t *dst) {
+  __m128i u = _mm_packus_epi16(*x, *x);
+  *(int *)dst = _mm_cvtsi128_si32(u);
+}
+
+static INLINE void accumulate_store_4_pixel(const __m128i *x, uint8_t *dst) {
+  __m128i y = accumulate_store(x, dst);
+  *(int *)dst = _mm_cvtsi128_si32(y);
+}
+
+static store_pixel_t store4pixelTab[2] = { store_4_pixel_only,
+                                           accumulate_store_4_pixel };
+
+static void horiz_w4_ssse3(const uint8_t *src, const __m128i *f, int tapsNum,
+                           store_pixel_t store_func, uint8_t *dst) {
+  __m128i sumPairRow[4];
+  __m128i sumPairCol[8];
+  __m128i pixel;
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i zero = _mm_setzero_si128();
+
+  if (10 == tapsNum) {
+    src -= 1;
+  }
+
+  pixel = _mm_loadu_si128((__m128i const *)src);
+  sumPairRow[0] = _mm_maddubs_epi16(pixel, f[0]);
+  sumPairRow[2] = _mm_maddubs_epi16(pixel, f[1]);
+  sumPairRow[2] = _mm_srli_si128(sumPairRow[2], 2);
+
+  pixel = _mm_loadu_si128((__m128i const *)(src + 1));
+  sumPairRow[1] = _mm_maddubs_epi16(pixel, f[0]);
+  sumPairRow[3] = _mm_maddubs_epi16(pixel, f[1]);
+  sumPairRow[3] = _mm_srli_si128(sumPairRow[3], 2);
+
+  transpose_4x8(sumPairRow, sumPairCol);
+
+  sumPairRow[0] = _mm_adds_epi16(sumPairCol[0], sumPairCol[1]);
+  sumPairRow[1] = _mm_adds_epi16(sumPairCol[4], sumPairCol[5]);
+
+  sumPairRow[2] = _mm_min_epi16(sumPairCol[2], sumPairCol[3]);
+  sumPairRow[3] = _mm_max_epi16(sumPairCol[2], sumPairCol[3]);
+
+  sumPairRow[0] = _mm_adds_epi16(sumPairRow[0], sumPairRow[1]);
+  sumPairRow[0] = _mm_adds_epi16(sumPairRow[0], sumPairRow[2]);
+  sumPairRow[0] = _mm_adds_epi16(sumPairRow[0], sumPairRow[3]);
+
+  sumPairRow[1] = _mm_mulhrs_epi16(sumPairRow[0], k_256);
+  sumPairRow[1] = _mm_packus_epi16(sumPairRow[1], sumPairRow[1]);
+  sumPairRow[1] = _mm_unpacklo_epi8(sumPairRow[1], zero);
+
+  store_func(&sumPairRow[1], dst);
+}
+
+static void horiz_w8_ssse3(const uint8_t *src, const __m128i *f, int tapsNum,
+                           store_pixel_t store, uint8_t *buf) {
+  horiz_w4_ssse3(src, f, tapsNum, store, buf);
+  src += 4;
+  buf += 4;
+  horiz_w4_ssse3(src, f, tapsNum, store, buf);
+}
+
+static void horiz_w16_ssse3(const uint8_t *src, const __m128i *f, int tapsNum,
+                            store_pixel_t store, uint8_t *buf) {
+  horiz_w8_ssse3(src, f, tapsNum, store, buf);
+  src += 8;
+  buf += 8;
+  horiz_w8_ssse3(src, f, tapsNum, store, buf);
+}
+
+static void horiz_w32_ssse3(const uint8_t *src, const __m128i *f, int tapsNum,
+                            store_pixel_t store, uint8_t *buf) {
+  horiz_w16_ssse3(src, f, tapsNum, store, buf);
+  src += 16;
+  buf += 16;
+  horiz_w16_ssse3(src, f, tapsNum, store, buf);
+}
+
+static void horiz_w64_ssse3(const uint8_t *src, const __m128i *f, int tapsNum,
+                            store_pixel_t store, uint8_t *buf) {
+  horiz_w32_ssse3(src, f, tapsNum, store, buf);
+  src += 32;
+  buf += 32;
+  horiz_w32_ssse3(src, f, tapsNum, store, buf);
+}
+
+static void horiz_w128_ssse3(const uint8_t *src, const __m128i *f, int tapsNum,
+                             store_pixel_t store, uint8_t *buf) {
+  horiz_w64_ssse3(src, f, tapsNum, store, buf);
+  src += 64;
+  buf += 64;
+  horiz_w64_ssse3(src, f, tapsNum, store, buf);
+}
+
+static void (*horizTab[6])(const uint8_t *, const __m128i *, int, store_pixel_t,
+                           uint8_t *) = {
+  horiz_w4_ssse3,  horiz_w8_ssse3,  horiz_w16_ssse3,
+  horiz_w32_ssse3, horiz_w64_ssse3, horiz_w128_ssse3,
+};
+
+static void filter_horiz_ssse3(const uint8_t *src, __m128i *f, int tapsNum,
+                               int width, store_pixel_t store, uint8_t *dst) {
+  switch (width) {
+    // Note:
+    // For width=2 and 4, store function must be different
+    case 2:
+    case 4: horizTab[0](src, f, tapsNum, store, dst); break;
+    case 8: horizTab[1](src, f, tapsNum, store, dst); break;
+    case 16: horizTab[2](src, f, tapsNum, store, dst); break;
+    case 32: horizTab[3](src, f, tapsNum, store, dst); break;
+    case 64: horizTab[4](src, f, tapsNum, store, dst); break;
+    case 128: horizTab[5](src, f, tapsNum, store, dst); break;
+    default: assert(0);
+  }
+}
+
+// Vertical 8-pixel parallel
+typedef void (*transpose_to_dst_t)(const uint16_t *src, int src_stride,
+                                   uint8_t *dst, int dst_stride);
+
+static INLINE void transpose8x8_direct_to_dst(const uint16_t *src,
+                                              int src_stride, uint8_t *dst,
+                                              int dst_stride) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  __m128i v0, v1, v2, v3;
+
+  __m128i u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride));
+  __m128i u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride));
+  __m128i u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  __m128i u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+  __m128i u4 = _mm_loadu_si128((__m128i const *)(src + 4 * src_stride));
+  __m128i u5 = _mm_loadu_si128((__m128i const *)(src + 5 * src_stride));
+  __m128i u6 = _mm_loadu_si128((__m128i const *)(src + 6 * src_stride));
+  __m128i u7 = _mm_loadu_si128((__m128i const *)(src + 7 * src_stride));
+
+  u0 = _mm_mulhrs_epi16(u0, k_256);
+  u1 = _mm_mulhrs_epi16(u1, k_256);
+  u2 = _mm_mulhrs_epi16(u2, k_256);
+  u3 = _mm_mulhrs_epi16(u3, k_256);
+  u4 = _mm_mulhrs_epi16(u4, k_256);
+  u5 = _mm_mulhrs_epi16(u5, k_256);
+  u6 = _mm_mulhrs_epi16(u6, k_256);
+  u7 = _mm_mulhrs_epi16(u7, k_256);
+
+  v0 = _mm_packus_epi16(u0, u1);
+  v1 = _mm_packus_epi16(u2, u3);
+  v2 = _mm_packus_epi16(u4, u5);
+  v3 = _mm_packus_epi16(u6, u7);
+
+  u0 = _mm_unpacklo_epi8(v0, v1);
+  u1 = _mm_unpackhi_epi8(v0, v1);
+  u2 = _mm_unpacklo_epi8(v2, v3);
+  u3 = _mm_unpackhi_epi8(v2, v3);
+
+  u4 = _mm_unpacklo_epi8(u0, u1);
+  u5 = _mm_unpacklo_epi8(u2, u3);
+  u6 = _mm_unpackhi_epi8(u0, u1);
+  u7 = _mm_unpackhi_epi8(u2, u3);
+
+  u0 = _mm_unpacklo_epi32(u4, u5);
+  u1 = _mm_unpackhi_epi32(u4, u5);
+  u2 = _mm_unpacklo_epi32(u6, u7);
+  u3 = _mm_unpackhi_epi32(u6, u7);
+
+  u4 = _mm_srli_si128(u0, 8);
+  u5 = _mm_srli_si128(u1, 8);
+  u6 = _mm_srli_si128(u2, 8);
+  u7 = _mm_srli_si128(u3, 8);
+
+  _mm_storel_epi64((__m128i *)dst, u0);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 1), u4);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 2), u1);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 3), u5);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 4), u2);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 5), u6);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 6), u3);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 7), u7);
+}
+
+static INLINE void transpose8x8_accumu_to_dst(const uint16_t *src,
+                                              int src_stride, uint8_t *dst,
+                                              int dst_stride) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+
+  __m128i u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride));
+  __m128i u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride));
+  __m128i u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  __m128i u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+  __m128i u4 = _mm_loadu_si128((__m128i const *)(src + 4 * src_stride));
+  __m128i u5 = _mm_loadu_si128((__m128i const *)(src + 5 * src_stride));
+  __m128i u6 = _mm_loadu_si128((__m128i const *)(src + 6 * src_stride));
+  __m128i u7 = _mm_loadu_si128((__m128i const *)(src + 7 * src_stride));
+
+  u0 = _mm_mulhrs_epi16(u0, k_256);
+  u1 = _mm_mulhrs_epi16(u1, k_256);
+  u2 = _mm_mulhrs_epi16(u2, k_256);
+  u3 = _mm_mulhrs_epi16(u3, k_256);
+  u4 = _mm_mulhrs_epi16(u4, k_256);
+  u5 = _mm_mulhrs_epi16(u5, k_256);
+  u6 = _mm_mulhrs_epi16(u6, k_256);
+  u7 = _mm_mulhrs_epi16(u7, k_256);
+
+  v0 = _mm_packus_epi16(u0, u1);
+  v1 = _mm_packus_epi16(u2, u3);
+  v2 = _mm_packus_epi16(u4, u5);
+  v3 = _mm_packus_epi16(u6, u7);
+
+  u0 = _mm_unpacklo_epi8(v0, v1);
+  u1 = _mm_unpackhi_epi8(v0, v1);
+  u2 = _mm_unpacklo_epi8(v2, v3);
+  u3 = _mm_unpackhi_epi8(v2, v3);
+
+  u4 = _mm_unpacklo_epi8(u0, u1);
+  u5 = _mm_unpacklo_epi8(u2, u3);
+  u6 = _mm_unpackhi_epi8(u0, u1);
+  u7 = _mm_unpackhi_epi8(u2, u3);
+
+  u0 = _mm_unpacklo_epi32(u4, u5);
+  u1 = _mm_unpackhi_epi32(u4, u5);
+  u2 = _mm_unpacklo_epi32(u6, u7);
+  u3 = _mm_unpackhi_epi32(u6, u7);
+
+  u4 = _mm_srli_si128(u0, 8);
+  u5 = _mm_srli_si128(u1, 8);
+  u6 = _mm_srli_si128(u2, 8);
+  u7 = _mm_srli_si128(u3, 8);
+
+  v0 = _mm_loadl_epi64((__m128i const *)(dst + 0 * dst_stride));
+  v1 = _mm_loadl_epi64((__m128i const *)(dst + 1 * dst_stride));
+  v2 = _mm_loadl_epi64((__m128i const *)(dst + 2 * dst_stride));
+  v3 = _mm_loadl_epi64((__m128i const *)(dst + 3 * dst_stride));
+  v4 = _mm_loadl_epi64((__m128i const *)(dst + 4 * dst_stride));
+  v5 = _mm_loadl_epi64((__m128i const *)(dst + 5 * dst_stride));
+  v6 = _mm_loadl_epi64((__m128i const *)(dst + 6 * dst_stride));
+  v7 = _mm_loadl_epi64((__m128i const *)(dst + 7 * dst_stride));
+
+  u0 = _mm_unpacklo_epi8(u0, zero);
+  u1 = _mm_unpacklo_epi8(u1, zero);
+  u2 = _mm_unpacklo_epi8(u2, zero);
+  u3 = _mm_unpacklo_epi8(u3, zero);
+  u4 = _mm_unpacklo_epi8(u4, zero);
+  u5 = _mm_unpacklo_epi8(u5, zero);
+  u6 = _mm_unpacklo_epi8(u6, zero);
+  u7 = _mm_unpacklo_epi8(u7, zero);
+
+  v0 = _mm_unpacklo_epi8(v0, zero);
+  v1 = _mm_unpacklo_epi8(v1, zero);
+  v2 = _mm_unpacklo_epi8(v2, zero);
+  v3 = _mm_unpacklo_epi8(v3, zero);
+  v4 = _mm_unpacklo_epi8(v4, zero);
+  v5 = _mm_unpacklo_epi8(v5, zero);
+  v6 = _mm_unpacklo_epi8(v6, zero);
+  v7 = _mm_unpacklo_epi8(v7, zero);
+
+  v0 = _mm_adds_epi16(u0, v0);
+  v1 = _mm_adds_epi16(u4, v1);
+  v2 = _mm_adds_epi16(u1, v2);
+  v3 = _mm_adds_epi16(u5, v3);
+  v4 = _mm_adds_epi16(u2, v4);
+  v5 = _mm_adds_epi16(u6, v5);
+  v6 = _mm_adds_epi16(u3, v6);
+  v7 = _mm_adds_epi16(u7, v7);
+
+  v0 = _mm_adds_epi16(v0, one);
+  v1 = _mm_adds_epi16(v1, one);
+  v2 = _mm_adds_epi16(v2, one);
+  v3 = _mm_adds_epi16(v3, one);
+  v4 = _mm_adds_epi16(v4, one);
+  v5 = _mm_adds_epi16(v5, one);
+  v6 = _mm_adds_epi16(v6, one);
+  v7 = _mm_adds_epi16(v7, one);
+
+  v0 = _mm_srai_epi16(v0, 1);
+  v1 = _mm_srai_epi16(v1, 1);
+  v2 = _mm_srai_epi16(v2, 1);
+  v3 = _mm_srai_epi16(v3, 1);
+  v4 = _mm_srai_epi16(v4, 1);
+  v5 = _mm_srai_epi16(v5, 1);
+  v6 = _mm_srai_epi16(v6, 1);
+  v7 = _mm_srai_epi16(v7, 1);
+
+  u0 = _mm_packus_epi16(v0, v1);
+  u1 = _mm_packus_epi16(v2, v3);
+  u2 = _mm_packus_epi16(v4, v5);
+  u3 = _mm_packus_epi16(v6, v7);
+
+  u4 = _mm_srli_si128(u0, 8);
+  u5 = _mm_srli_si128(u1, 8);
+  u6 = _mm_srli_si128(u2, 8);
+  u7 = _mm_srli_si128(u3, 8);
+
+  _mm_storel_epi64((__m128i *)dst, u0);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 1), u4);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 2), u1);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 3), u5);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 4), u2);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 5), u6);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 6), u3);
+  _mm_storel_epi64((__m128i *)(dst + dst_stride * 7), u7);
+}
+
+static transpose_to_dst_t trans8x8Tab[2] = { transpose8x8_direct_to_dst,
+                                             transpose8x8_accumu_to_dst };
+
+static INLINE void transpose_8x16(const __m128i *in, __m128i *out) {
+  __m128i t0, t1, t2, t3, u0, u1;
+
+  t0 = _mm_unpacklo_epi16(in[0], in[1]);
+  t1 = _mm_unpacklo_epi16(in[2], in[3]);
+  t2 = _mm_unpacklo_epi16(in[4], in[5]);
+  t3 = _mm_unpacklo_epi16(in[6], in[7]);
+
+  u0 = _mm_unpacklo_epi32(t0, t1);
+  u1 = _mm_unpacklo_epi32(t2, t3);
+
+  out[0] = _mm_unpacklo_epi64(u0, u1);
+  out[1] = _mm_unpackhi_epi64(u0, u1);
+
+  u0 = _mm_unpackhi_epi32(t0, t1);
+  u1 = _mm_unpackhi_epi32(t2, t3);
+
+  out[2] = _mm_unpacklo_epi64(u0, u1);
+  out[3] = _mm_unpackhi_epi64(u0, u1);
+
+  t0 = _mm_unpackhi_epi16(in[0], in[1]);
+  t1 = _mm_unpackhi_epi16(in[2], in[3]);
+  t2 = _mm_unpackhi_epi16(in[4], in[5]);
+  t3 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  u0 = _mm_unpacklo_epi32(t0, t1);
+  u1 = _mm_unpacklo_epi32(t2, t3);
+
+  out[4] = _mm_unpacklo_epi64(u0, u1);
+  out[5] = _mm_unpackhi_epi64(u0, u1);
+
+  // Ignore out[6] and out[7]
+  // they're zero vectors.
+}
+
+static void filter_horiz_v8p_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                   __m128i *f, int tapsNum, uint16_t *buf) {
+  __m128i s[8], t[6];
+  __m128i min_x2x3, max_x2x3;
+  __m128i temp;
+
+  if (tapsNum == 10) {
+    src_ptr -= 1;
+  }
+  s[0] = _mm_loadu_si128((const __m128i *)src_ptr);
+  s[1] = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch));
+  s[2] = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
+  s[3] = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
+  s[4] = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4));
+  s[5] = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5));
+  s[6] = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6));
+  s[7] = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+
+  // TRANSPOSE...
+  // Vecotor represents column pixel pairs instead of a row
+  transpose_8x16(s, t);
+
+  // multiply 2 adjacent elements with the filter and add the result
+  s[0] = _mm_maddubs_epi16(t[0], f[0]);
+  s[1] = _mm_maddubs_epi16(t[1], f[1]);
+  s[2] = _mm_maddubs_epi16(t[2], f[2]);
+  s[3] = _mm_maddubs_epi16(t[3], f[3]);
+  s[4] = _mm_maddubs_epi16(t[4], f[4]);
+  s[5] = _mm_maddubs_epi16(t[5], f[5]);
+
+  // add and saturate the results together
+  min_x2x3 = _mm_min_epi16(s[2], s[3]);
+  max_x2x3 = _mm_max_epi16(s[2], s[3]);
+  temp = _mm_adds_epi16(s[0], s[1]);
+  temp = _mm_adds_epi16(temp, s[5]);
+  temp = _mm_adds_epi16(temp, s[4]);
+
+  temp = _mm_adds_epi16(temp, min_x2x3);
+  temp = _mm_adds_epi16(temp, max_x2x3);
+
+  _mm_storeu_si128((__m128i *)buf, temp);
+}
+
+// Vertical 4-pixel parallel
+static INLINE void transpose4x4_direct_to_dst(const uint16_t *src,
+                                              int src_stride, uint8_t *dst,
+                                              int dst_stride) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  __m128i v0, v1, v2, v3;
+
+  // TODO(luoyi): two loads, 8 elements per load (two bytes per element)
+  __m128i u0 = _mm_loadl_epi64((__m128i const *)(src + 0 * src_stride));
+  __m128i u1 = _mm_loadl_epi64((__m128i const *)(src + 1 * src_stride));
+  __m128i u2 = _mm_loadl_epi64((__m128i const *)(src + 2 * src_stride));
+  __m128i u3 = _mm_loadl_epi64((__m128i const *)(src + 3 * src_stride));
+
+  v0 = _mm_unpacklo_epi16(u0, u1);
+  v1 = _mm_unpacklo_epi16(u2, u3);
+
+  v2 = _mm_unpacklo_epi32(v0, v1);
+  v3 = _mm_unpackhi_epi32(v0, v1);
+
+  u0 = _mm_mulhrs_epi16(v2, k_256);
+  u1 = _mm_mulhrs_epi16(v3, k_256);
+
+  u0 = _mm_packus_epi16(u0, u1);
+  u1 = _mm_srli_si128(u0, 4);
+  u2 = _mm_srli_si128(u0, 8);
+  u3 = _mm_srli_si128(u0, 12);
+
+  *(int *)(dst) = _mm_cvtsi128_si32(u0);
+  *(int *)(dst + dst_stride) = _mm_cvtsi128_si32(u1);
+  *(int *)(dst + dst_stride * 2) = _mm_cvtsi128_si32(u2);
+  *(int *)(dst + dst_stride * 3) = _mm_cvtsi128_si32(u3);
+}
+
+static INLINE void transpose4x4_accumu_to_dst(const uint16_t *src,
+                                              int src_stride, uint8_t *dst,
+                                              int dst_stride) {
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+
+  __m128i v0, v1, v2, v3;
+
+  __m128i u0 = _mm_loadl_epi64((__m128i const *)(src));
+  __m128i u1 = _mm_loadl_epi64((__m128i const *)(src + src_stride));
+  __m128i u2 = _mm_loadl_epi64((__m128i const *)(src + 2 * src_stride));
+  __m128i u3 = _mm_loadl_epi64((__m128i const *)(src + 3 * src_stride));
+
+  v0 = _mm_unpacklo_epi16(u0, u1);
+  v1 = _mm_unpacklo_epi16(u2, u3);
+
+  v2 = _mm_unpacklo_epi32(v0, v1);
+  v3 = _mm_unpackhi_epi32(v0, v1);
+
+  u0 = _mm_mulhrs_epi16(v2, k_256);
+  u1 = _mm_mulhrs_epi16(v3, k_256);
+
+  u2 = _mm_packus_epi16(u0, u1);
+  u0 = _mm_unpacklo_epi8(u2, zero);
+  u1 = _mm_unpackhi_epi8(u2, zero);
+
+  // load pixel values
+  v0 = _mm_loadl_epi64((__m128i const *)(dst));
+  v1 = _mm_loadl_epi64((__m128i const *)(dst + dst_stride));
+  v2 = _mm_loadl_epi64((__m128i const *)(dst + 2 * dst_stride));
+  v3 = _mm_loadl_epi64((__m128i const *)(dst + 3 * dst_stride));
+
+  v0 = _mm_unpacklo_epi8(v0, zero);
+  v1 = _mm_unpacklo_epi8(v1, zero);
+  v2 = _mm_unpacklo_epi8(v2, zero);
+  v3 = _mm_unpacklo_epi8(v3, zero);
+
+  v0 = _mm_unpacklo_epi64(v0, v1);
+  v1 = _mm_unpacklo_epi64(v2, v3);
+
+  u0 = _mm_adds_epi16(u0, v0);
+  u1 = _mm_adds_epi16(u1, v1);
+
+  u0 = _mm_adds_epi16(u0, one);
+  u1 = _mm_adds_epi16(u1, one);
+
+  u0 = _mm_srai_epi16(u0, 1);
+  u1 = _mm_srai_epi16(u1, 1);
+
+  // saturation and pack to pixels
+  u0 = _mm_packus_epi16(u0, u1);
+  u1 = _mm_srli_si128(u0, 4);
+  u2 = _mm_srli_si128(u0, 8);
+  u3 = _mm_srli_si128(u0, 12);
+
+  *(int *)(dst) = _mm_cvtsi128_si32(u0);
+  *(int *)(dst + dst_stride) = _mm_cvtsi128_si32(u1);
+  *(int *)(dst + dst_stride * 2) = _mm_cvtsi128_si32(u2);
+  *(int *)(dst + dst_stride * 3) = _mm_cvtsi128_si32(u3);
+}
+
+static transpose_to_dst_t trans4x4Tab[2] = { transpose4x4_direct_to_dst,
+                                             transpose4x4_accumu_to_dst };
+
+static void filter_horiz_v4p_ssse3(const uint8_t *src_ptr, ptrdiff_t src_pitch,
+                                   __m128i *f, int tapsNum, uint16_t *buf) {
+  __m128i A, B, C, D;
+  __m128i tr0_0, tr0_1, s1s0, s3s2, s5s4, s7s6, s9s8, sbsa;
+  __m128i x0, x1, x2, x3, x4, x5;
+  __m128i min_x2x3, max_x2x3, temp;
+
+  if (tapsNum == 10) {
+    src_ptr -= 1;
+  }
+  A = _mm_loadu_si128((const __m128i *)src_ptr);
+  B = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch));
+  C = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
+  D = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
+
+  // TRANSPOSE...
+  // Vecotor represents column pixel pairs instead of a row
+  // 00 01 10 11 02 03 12 13 04 05 14 15 06 07 16 17
+  tr0_0 = _mm_unpacklo_epi16(A, B);
+  // 20 21 30 31 22 23 32 33 24 25 34 35 26 27 36 37
+  tr0_1 = _mm_unpacklo_epi16(C, D);
+  // 00 01 10 11 20 21 30 31 02 03 12 13 22 23 32 33
+  s1s0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  // 04 05 14 15 24 25 34 35 06 07 16 17 26 27 36 37
+  s5s4 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  // 02 03 12 13 22 23 32 33
+  s3s2 = _mm_srli_si128(s1s0, 8);
+  // 06 07 16 17 26 27 36 37
+  s7s6 = _mm_srli_si128(s5s4, 8);
+
+  tr0_0 = _mm_unpackhi_epi16(A, B);
+  tr0_1 = _mm_unpackhi_epi16(C, D);
+  s9s8 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  sbsa = _mm_srli_si128(s9s8, 8);
+
+  // multiply 2 adjacent elements with the filter and add the result
+  x0 = _mm_maddubs_epi16(s1s0, f[0]);
+  x1 = _mm_maddubs_epi16(s3s2, f[1]);
+  x2 = _mm_maddubs_epi16(s5s4, f[2]);
+  x3 = _mm_maddubs_epi16(s7s6, f[3]);
+  x4 = _mm_maddubs_epi16(s9s8, f[4]);
+  x5 = _mm_maddubs_epi16(sbsa, f[5]);
+  // add and saturate the results together
+  min_x2x3 = _mm_min_epi16(x2, x3);
+  max_x2x3 = _mm_max_epi16(x2, x3);
+  temp = _mm_adds_epi16(x0, x1);
+  temp = _mm_adds_epi16(temp, x5);
+  temp = _mm_adds_epi16(temp, x4);
+
+  temp = _mm_adds_epi16(temp, min_x2x3);
+  temp = _mm_adds_epi16(temp, max_x2x3);
+  _mm_storel_epi64((__m128i *)buf, temp);
+}
+
+// Note:
+//  This function assumes:
+// (1) 10/12-taps filters
+// (2) x_step_q4 = 16 then filter is fixed at the call
+
+void av1_convolve_horiz_ssse3(const uint8_t *src, int src_stride, uint8_t *dst,
+                              int dst_stride, int w, int h,
+                              const InterpFilterParams filter_params,
+                              const int subpel_x_q4, int x_step_q4,
+                              ConvolveParams *conv_params) {
+  DECLARE_ALIGNED(16, uint16_t, temp[8 * 8]);
+  __m128i verf[6];
+  __m128i horf[2];
+  SubpelFilterCoeffs hCoeffs, vCoeffs;
+  const uint8_t *src_ptr;
+  store_pixel_t store2p = store2pixelTab[conv_params->ref];
+  store_pixel_t store4p = store4pixelTab[conv_params->ref];
+  transpose_to_dst_t transpose_4x4 = trans4x4Tab[conv_params->ref];
+  transpose_to_dst_t transpose_8x8 = trans8x8Tab[conv_params->ref];
+
+  const int tapsNum = filter_params.taps;
+  int block_height, block_residu;
+  int i, col, count;
+  (void)x_step_q4;
+
+  if (0 == subpel_x_q4 || 16 != x_step_q4) {
+    av1_convolve_horiz_c(src, src_stride, dst, dst_stride, w, h, filter_params,
+                         subpel_x_q4, x_step_q4, conv_params);
+    return;
+  }
+
+  hCoeffs = get_subpel_filter_signal_dir(filter_params, subpel_x_q4 - 1);
+  vCoeffs = get_subpel_filter_ver_signal_dir(filter_params, subpel_x_q4 - 1);
+
+  if (!hCoeffs || !vCoeffs) {
+    av1_convolve_horiz_c(src, src_stride, dst, dst_stride, w, h, filter_params,
+                         subpel_x_q4, x_step_q4, conv_params);
+    return;
+  }
+
+  verf[0] = *((const __m128i *)(vCoeffs));
+  verf[1] = *((const __m128i *)(vCoeffs + 1));
+  verf[2] = *((const __m128i *)(vCoeffs + 2));
+  verf[3] = *((const __m128i *)(vCoeffs + 3));
+  verf[4] = *((const __m128i *)(vCoeffs + 4));
+  verf[5] = *((const __m128i *)(vCoeffs + 5));
+
+  horf[0] = *((const __m128i *)(hCoeffs));
+  horf[1] = *((const __m128i *)(hCoeffs + 1));
+
+  count = 0;
+
+  // here tapsNum is filter size
+  src -= (tapsNum >> 1) - 1;
+  src_ptr = src;
+  if (w > WIDTH_BOUND && h > HEIGHT_BOUND) {
+    // 8-pixels parallel
+    block_height = h >> 3;
+    block_residu = h & 7;
+
+    do {
+      for (col = 0; col < w; col += 8) {
+        for (i = 0; i < 8; ++i) {
+          filter_horiz_v8p_ssse3(src_ptr, src_stride, verf, tapsNum,
+                                 temp + (i * 8));
+          src_ptr += 1;
+        }
+        transpose_8x8(temp, 8, dst + col, dst_stride);
+      }
+      count++;
+      src_ptr = src + count * src_stride * 8;
+      dst += dst_stride * 8;
+    } while (count < block_height);
+
+    for (i = 0; i < block_residu; ++i) {
+      filter_horiz_ssse3(src_ptr, horf, tapsNum, w, store4p, dst);
+      src_ptr += src_stride;
+      dst += dst_stride;
+    }
+  } else {
+    if (w > 2) {
+      // 4-pixels parallel
+      block_height = h >> 2;
+      block_residu = h & 3;
+
+      do {
+        for (col = 0; col < w; col += 4) {
+          for (i = 0; i < 4; ++i) {
+            filter_horiz_v4p_ssse3(src_ptr, src_stride, verf, tapsNum,
+                                   temp + (i * 4));
+            src_ptr += 1;
+          }
+          transpose_4x4(temp, 4, dst + col, dst_stride);
+        }
+        count++;
+        src_ptr = src + count * src_stride * 4;
+        dst += dst_stride * 4;
+      } while (count < block_height);
+
+      for (i = 0; i < block_residu; ++i) {
+        filter_horiz_ssse3(src_ptr, horf, tapsNum, w, store4p, dst);
+        src_ptr += src_stride;
+        dst += dst_stride;
+      }
+    } else {
+      for (i = 0; i < h; i++) {
+        filter_horiz_ssse3(src_ptr, horf, tapsNum, w, store2p, dst);
+        src_ptr += src_stride;
+        dst += dst_stride;
+      }
+    }
+  }
+}
+
+// Vertical convolution filtering
+static INLINE void store_8_pixel_only(const __m128i *x, uint8_t *dst) {
+  __m128i u = _mm_packus_epi16(*x, *x);
+  _mm_storel_epi64((__m128i *)dst, u);
+}
+
+static INLINE void accumulate_store_8_pixel(const __m128i *x, uint8_t *dst) {
+  __m128i y = accumulate_store(x, dst);
+  _mm_storel_epi64((__m128i *)dst, y);
+}
+
+static store_pixel_t store8pixelTab[2] = { store_8_pixel_only,
+                                           accumulate_store_8_pixel };
+
+static __m128i filter_vert_ssse3(const uint8_t *src, int src_stride,
+                                 int tapsNum, __m128i *f) {
+  __m128i s[12];
+  const __m128i k_256 = _mm_set1_epi16(1 << 8);
+  const __m128i zero = _mm_setzero_si128();
+  __m128i min_x2x3, max_x2x3, sum;
+  int i = 0;
+  int r = 0;
+
+  if (10 == tapsNum) {
+    i += 1;
+    s[0] = zero;
+  }
+  while (i < 12) {
+    s[i] = _mm_loadu_si128((__m128i const *)(src + r * src_stride));
+    i += 1;
+    r += 1;
+  }
+
+  s[0] = _mm_unpacklo_epi8(s[0], s[1]);
+  s[2] = _mm_unpacklo_epi8(s[2], s[3]);
+  s[4] = _mm_unpacklo_epi8(s[4], s[5]);
+  s[6] = _mm_unpacklo_epi8(s[6], s[7]);
+  s[8] = _mm_unpacklo_epi8(s[8], s[9]);
+  s[10] = _mm_unpacklo_epi8(s[10], s[11]);
+
+  s[0] = _mm_maddubs_epi16(s[0], f[0]);
+  s[2] = _mm_maddubs_epi16(s[2], f[1]);
+  s[4] = _mm_maddubs_epi16(s[4], f[2]);
+  s[6] = _mm_maddubs_epi16(s[6], f[3]);
+  s[8] = _mm_maddubs_epi16(s[8], f[4]);
+  s[10] = _mm_maddubs_epi16(s[10], f[5]);
+
+  min_x2x3 = _mm_min_epi16(s[4], s[6]);
+  max_x2x3 = _mm_max_epi16(s[4], s[6]);
+  sum = _mm_adds_epi16(s[0], s[2]);
+  sum = _mm_adds_epi16(sum, s[10]);
+  sum = _mm_adds_epi16(sum, s[8]);
+
+  sum = _mm_adds_epi16(sum, min_x2x3);
+  sum = _mm_adds_epi16(sum, max_x2x3);
+
+  sum = _mm_mulhrs_epi16(sum, k_256);
+  sum = _mm_packus_epi16(sum, sum);
+  sum = _mm_unpacklo_epi8(sum, zero);
+  return sum;
+}
+
+static void filter_vert_horiz_parallel_ssse3(const uint8_t *src, int src_stride,
+                                             __m128i *f, int tapsNum,
+                                             store_pixel_t store_func,
+                                             uint8_t *dst) {
+  __m128i sum = filter_vert_ssse3(src, src_stride, tapsNum, f);
+  store_func(&sum, dst);
+}
+
+static void filter_vert_compute_small(const uint8_t *src, int src_stride,
+                                      __m128i *f, int tapsNum,
+                                      store_pixel_t store_func, int h,
+                                      uint8_t *dst, int dst_stride) {
+  int rowIndex = 0;
+  do {
+    filter_vert_horiz_parallel_ssse3(src, src_stride, f, tapsNum, store_func,
+                                     dst);
+    rowIndex++;
+    src += src_stride;
+    dst += dst_stride;
+  } while (rowIndex < h);
+}
+
+static void filter_vert_compute_large(const uint8_t *src, int src_stride,
+                                      __m128i *f, int tapsNum,
+                                      store_pixel_t store_func, int w, int h,
+                                      uint8_t *dst, int dst_stride) {
+  int col;
+  int rowIndex = 0;
+  const uint8_t *src_ptr = src;
+  uint8_t *dst_ptr = dst;
+
+  do {
+    for (col = 0; col < w; col += 8) {
+      filter_vert_horiz_parallel_ssse3(src_ptr, src_stride, f, tapsNum,
+                                       store_func, dst_ptr);
+      src_ptr += 8;
+      dst_ptr += 8;
+    }
+    rowIndex++;
+    src_ptr = src + rowIndex * src_stride;
+    dst_ptr = dst + rowIndex * dst_stride;
+  } while (rowIndex < h);
+}
+
+void av1_convolve_vert_ssse3(const uint8_t *src, int src_stride, uint8_t *dst,
+                             int dst_stride, int w, int h,
+                             const InterpFilterParams filter_params,
+                             const int subpel_y_q4, int y_step_q4,
+                             ConvolveParams *conv_params) {
+  __m128i verf[6];
+  SubpelFilterCoeffs vCoeffs;
+  const uint8_t *src_ptr;
+  uint8_t *dst_ptr = dst;
+  store_pixel_t store2p = store2pixelTab[conv_params->ref];
+  store_pixel_t store4p = store4pixelTab[conv_params->ref];
+  store_pixel_t store8p = store8pixelTab[conv_params->ref];
+  const int tapsNum = filter_params.taps;
+
+  if (0 == subpel_y_q4 || 16 != y_step_q4) {
+    av1_convolve_vert_c(src, src_stride, dst, dst_stride, w, h, filter_params,
+                        subpel_y_q4, y_step_q4, conv_params);
+    return;
+  }
+
+  vCoeffs = get_subpel_filter_ver_signal_dir(filter_params, subpel_y_q4 - 1);
+
+  if (!vCoeffs) {
+    av1_convolve_vert_c(src, src_stride, dst, dst_stride, w, h, filter_params,
+                        subpel_y_q4, y_step_q4, conv_params);
+    return;
+  }
+
+  verf[0] = *((const __m128i *)(vCoeffs));
+  verf[1] = *((const __m128i *)(vCoeffs + 1));
+  verf[2] = *((const __m128i *)(vCoeffs + 2));
+  verf[3] = *((const __m128i *)(vCoeffs + 3));
+  verf[4] = *((const __m128i *)(vCoeffs + 4));
+  verf[5] = *((const __m128i *)(vCoeffs + 5));
+
+  src -= src_stride * ((tapsNum >> 1) - 1);
+  src_ptr = src;
+
+  if (w > 4) {
+    filter_vert_compute_large(src_ptr, src_stride, verf, tapsNum, store8p, w, h,
+                              dst_ptr, dst_stride);
+  } else if (4 == w) {
+    filter_vert_compute_small(src_ptr, src_stride, verf, tapsNum, store4p, h,
+                              dst_ptr, dst_stride);
+  } else if (2 == w) {
+    filter_vert_compute_small(src_ptr, src_stride, verf, tapsNum, store2p, h,
+                              dst_ptr, dst_stride);
+  } else {
+    assert(0);
+  }
+}
+
+static void init_simd_horiz_filter(const int16_t *filter_ptr, int taps,
+                                   int8_t (*simd_horiz_filter)[2][16]) {
+  int shift;
+  int offset = (12 - taps) / 2;
+  const int16_t *filter_row;
+  for (shift = 1; shift < SUBPEL_SHIFTS; ++shift) {
+    int i;
+    filter_row = filter_ptr + shift * taps;
+    for (i = 0; i < offset; ++i) simd_horiz_filter[shift - 1][0][i] = 0;
+
+    for (i = 0; i < offset + 2; ++i) simd_horiz_filter[shift - 1][1][i] = 0;
+
+    for (i = 0; i < taps; ++i) {
+      simd_horiz_filter[shift - 1][0][i + offset] = (int8_t)filter_row[i];
+      simd_horiz_filter[shift - 1][1][i + offset + 2] = (int8_t)filter_row[i];
+    }
+
+    for (i = offset + taps; i < 16; ++i) simd_horiz_filter[shift - 1][0][i] = 0;
+
+    for (i = offset + 2 + taps; i < 16; ++i)
+      simd_horiz_filter[shift - 1][1][i] = 0;
+  }
+}
+
+static void init_simd_vert_filter(const int16_t *filter_ptr, int taps,
+                                  int8_t (*simd_vert_filter)[6][16]) {
+  int shift;
+  int offset = (12 - taps) / 2;
+  const int16_t *filter_row;
+  for (shift = 1; shift < SUBPEL_SHIFTS; ++shift) {
+    int i;
+    filter_row = filter_ptr + shift * taps;
+    for (i = 0; i < 6; ++i) {
+      int j;
+      for (j = 0; j < 16; ++j) {
+        int c = i * 2 + (j % 2) - offset;
+        if (c >= 0 && c < taps)
+          simd_vert_filter[shift - 1][i][j] = (int8_t)filter_row[c];
+        else
+          simd_vert_filter[shift - 1][i][j] = 0;
+      }
+    }
+  }
+}
+
+typedef struct SimdFilter {
+  InterpFilter interp_filter;
+  int8_t (*simd_horiz_filter)[2][16];
+  int8_t (*simd_vert_filter)[6][16];
+} SimdFilter;
+
+#if CONFIG_DUAL_FILTER
+#define MULTITAP_FILTER_NUM 1
+SimdFilter simd_filters[MULTITAP_FILTER_NUM] = {
+  { MULTITAP_SHARP, &sub_pel_filters_12sharp_signal_dir[0],
+    &sub_pel_filters_12sharp_ver_signal_dir[0] },
+};
+#endif
+
+#if USE_TEMPORALFILTER_12TAP
+SimdFilter temporal_simd_filter = {
+  TEMPORALFILTER_12TAP, &sub_pel_filters_temporalfilter_12_signal_dir[0],
+  &sub_pel_filters_temporalfilter_12_ver_signal_dir[0]
+};
+#endif
+
+void av1_lowbd_convolve_init_ssse3(void) {
+#if USE_TEMPORALFILTER_12TAP
+  {
+    InterpFilterParams filter_params =
+        av1_get_interp_filter_params(temporal_simd_filter.interp_filter);
+    int taps = filter_params.taps;
+    const int16_t *filter_ptr = filter_params.filter_ptr;
+    init_simd_horiz_filter(filter_ptr, taps,
+                           temporal_simd_filter.simd_horiz_filter);
+    init_simd_vert_filter(filter_ptr, taps,
+                          temporal_simd_filter.simd_vert_filter);
+  }
+#endif
+#if CONFIG_DUAL_FILTER
+  {
+    int i;
+    for (i = 0; i < MULTITAP_FILTER_NUM; ++i) {
+      InterpFilter interp_filter = simd_filters[i].interp_filter;
+      InterpFilterParams filter_params =
+          av1_get_interp_filter_params(interp_filter);
+      int taps = filter_params.taps;
+      const int16_t *filter_ptr = filter_params.filter_ptr;
+      init_simd_horiz_filter(filter_ptr, taps,
+                             simd_filters[i].simd_horiz_filter);
+      init_simd_vert_filter(filter_ptr, taps, simd_filters[i].simd_vert_filter);
+    }
+  }
+#endif
+  return;
+}
diff --git a/third_party/aom/av1/common/x86/av1_fwd_txfm1d_sse4.c b/third_party/aom/av1/common/x86/av1_fwd_txfm1d_sse4.c
new file mode 100644
index 0000000000..d04b667f19
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_fwd_txfm1d_sse4.c
@@ -0,0 +1,839 @@
+#include "av1/common/x86/av1_txfm1d_sse4.h"
+
+void av1_fdct32_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range) {
+  const int txfm_size = 32;
+  const int num_per_128 = 4;
+  const int32_t *cospi;
+  __m128i buf0[32];
+  __m128i buf1[32];
+  int col_num = txfm_size / num_per_128;
+  int bit;
+  int col;
+  (void)stage_range;
+  for (col = 0; col < col_num; col++) {
+    // stage 0;
+    int32_t stage_idx = 0;
+    int j;
+    for (j = 0; j < 32; ++j) {
+      buf0[j] = input[j * col_num + col];
+    }
+
+    // stage 1
+    stage_idx++;
+    buf1[0] = _mm_add_epi32(buf0[0], buf0[31]);
+    buf1[31] = _mm_sub_epi32(buf0[0], buf0[31]);
+    buf1[1] = _mm_add_epi32(buf0[1], buf0[30]);
+    buf1[30] = _mm_sub_epi32(buf0[1], buf0[30]);
+    buf1[2] = _mm_add_epi32(buf0[2], buf0[29]);
+    buf1[29] = _mm_sub_epi32(buf0[2], buf0[29]);
+    buf1[3] = _mm_add_epi32(buf0[3], buf0[28]);
+    buf1[28] = _mm_sub_epi32(buf0[3], buf0[28]);
+    buf1[4] = _mm_add_epi32(buf0[4], buf0[27]);
+    buf1[27] = _mm_sub_epi32(buf0[4], buf0[27]);
+    buf1[5] = _mm_add_epi32(buf0[5], buf0[26]);
+    buf1[26] = _mm_sub_epi32(buf0[5], buf0[26]);
+    buf1[6] = _mm_add_epi32(buf0[6], buf0[25]);
+    buf1[25] = _mm_sub_epi32(buf0[6], buf0[25]);
+    buf1[7] = _mm_add_epi32(buf0[7], buf0[24]);
+    buf1[24] = _mm_sub_epi32(buf0[7], buf0[24]);
+    buf1[8] = _mm_add_epi32(buf0[8], buf0[23]);
+    buf1[23] = _mm_sub_epi32(buf0[8], buf0[23]);
+    buf1[9] = _mm_add_epi32(buf0[9], buf0[22]);
+    buf1[22] = _mm_sub_epi32(buf0[9], buf0[22]);
+    buf1[10] = _mm_add_epi32(buf0[10], buf0[21]);
+    buf1[21] = _mm_sub_epi32(buf0[10], buf0[21]);
+    buf1[11] = _mm_add_epi32(buf0[11], buf0[20]);
+    buf1[20] = _mm_sub_epi32(buf0[11], buf0[20]);
+    buf1[12] = _mm_add_epi32(buf0[12], buf0[19]);
+    buf1[19] = _mm_sub_epi32(buf0[12], buf0[19]);
+    buf1[13] = _mm_add_epi32(buf0[13], buf0[18]);
+    buf1[18] = _mm_sub_epi32(buf0[13], buf0[18]);
+    buf1[14] = _mm_add_epi32(buf0[14], buf0[17]);
+    buf1[17] = _mm_sub_epi32(buf0[14], buf0[17]);
+    buf1[15] = _mm_add_epi32(buf0[15], buf0[16]);
+    buf1[16] = _mm_sub_epi32(buf0[15], buf0[16]);
+
+    // stage 2
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf0[0] = _mm_add_epi32(buf1[0], buf1[15]);
+    buf0[15] = _mm_sub_epi32(buf1[0], buf1[15]);
+    buf0[1] = _mm_add_epi32(buf1[1], buf1[14]);
+    buf0[14] = _mm_sub_epi32(buf1[1], buf1[14]);
+    buf0[2] = _mm_add_epi32(buf1[2], buf1[13]);
+    buf0[13] = _mm_sub_epi32(buf1[2], buf1[13]);
+    buf0[3] = _mm_add_epi32(buf1[3], buf1[12]);
+    buf0[12] = _mm_sub_epi32(buf1[3], buf1[12]);
+    buf0[4] = _mm_add_epi32(buf1[4], buf1[11]);
+    buf0[11] = _mm_sub_epi32(buf1[4], buf1[11]);
+    buf0[5] = _mm_add_epi32(buf1[5], buf1[10]);
+    buf0[10] = _mm_sub_epi32(buf1[5], buf1[10]);
+    buf0[6] = _mm_add_epi32(buf1[6], buf1[9]);
+    buf0[9] = _mm_sub_epi32(buf1[6], buf1[9]);
+    buf0[7] = _mm_add_epi32(buf1[7], buf1[8]);
+    buf0[8] = _mm_sub_epi32(buf1[7], buf1[8]);
+    buf0[16] = buf1[16];
+    buf0[17] = buf1[17];
+    buf0[18] = buf1[18];
+    buf0[19] = buf1[19];
+    btf_32_sse4_1_type0(-cospi[32], cospi[32], buf1[20], buf1[27], buf0[20],
+                        buf0[27], bit);
+    btf_32_sse4_1_type0(-cospi[32], cospi[32], buf1[21], buf1[26], buf0[21],
+                        buf0[26], bit);
+    btf_32_sse4_1_type0(-cospi[32], cospi[32], buf1[22], buf1[25], buf0[22],
+                        buf0[25], bit);
+    btf_32_sse4_1_type0(-cospi[32], cospi[32], buf1[23], buf1[24], buf0[23],
+                        buf0[24], bit);
+    buf0[28] = buf1[28];
+    buf0[29] = buf1[29];
+    buf0[30] = buf1[30];
+    buf0[31] = buf1[31];
+
+    // stage 3
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf1[0] = _mm_add_epi32(buf0[0], buf0[7]);
+    buf1[7] = _mm_sub_epi32(buf0[0], buf0[7]);
+    buf1[1] = _mm_add_epi32(buf0[1], buf0[6]);
+    buf1[6] = _mm_sub_epi32(buf0[1], buf0[6]);
+    buf1[2] = _mm_add_epi32(buf0[2], buf0[5]);
+    buf1[5] = _mm_sub_epi32(buf0[2], buf0[5]);
+    buf1[3] = _mm_add_epi32(buf0[3], buf0[4]);
+    buf1[4] = _mm_sub_epi32(buf0[3], buf0[4]);
+    buf1[8] = buf0[8];
+    buf1[9] = buf0[9];
+    btf_32_sse4_1_type0(-cospi[32], cospi[32], buf0[10], buf0[13], buf1[10],
+                        buf1[13], bit);
+    btf_32_sse4_1_type0(-cospi[32], cospi[32], buf0[11], buf0[12], buf1[11],
+                        buf1[12], bit);
+    buf1[14] = buf0[14];
+    buf1[15] = buf0[15];
+    buf1[16] = _mm_add_epi32(buf0[16], buf0[23]);
+    buf1[23] = _mm_sub_epi32(buf0[16], buf0[23]);
+    buf1[17] = _mm_add_epi32(buf0[17], buf0[22]);
+    buf1[22] = _mm_sub_epi32(buf0[17], buf0[22]);
+    buf1[18] = _mm_add_epi32(buf0[18], buf0[21]);
+    buf1[21] = _mm_sub_epi32(buf0[18], buf0[21]);
+    buf1[19] = _mm_add_epi32(buf0[19], buf0[20]);
+    buf1[20] = _mm_sub_epi32(buf0[19], buf0[20]);
+    buf1[24] = _mm_sub_epi32(buf0[31], buf0[24]);
+    buf1[31] = _mm_add_epi32(buf0[31], buf0[24]);
+    buf1[25] = _mm_sub_epi32(buf0[30], buf0[25]);
+    buf1[30] = _mm_add_epi32(buf0[30], buf0[25]);
+    buf1[26] = _mm_sub_epi32(buf0[29], buf0[26]);
+    buf1[29] = _mm_add_epi32(buf0[29], buf0[26]);
+    buf1[27] = _mm_sub_epi32(buf0[28], buf0[27]);
+    buf1[28] = _mm_add_epi32(buf0[28], buf0[27]);
+
+    // stage 4
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf0[0] = _mm_add_epi32(buf1[0], buf1[3]);
+    buf0[3] = _mm_sub_epi32(buf1[0], buf1[3]);
+    buf0[1] = _mm_add_epi32(buf1[1], buf1[2]);
+    buf0[2] = _mm_sub_epi32(buf1[1], buf1[2]);
+    buf0[4] = buf1[4];
+    btf_32_sse4_1_type0(-cospi[32], cospi[32], buf1[5], buf1[6], buf0[5],
+                        buf0[6], bit);
+    buf0[7] = buf1[7];
+    buf0[8] = _mm_add_epi32(buf1[8], buf1[11]);
+    buf0[11] = _mm_sub_epi32(buf1[8], buf1[11]);
+    buf0[9] = _mm_add_epi32(buf1[9], buf1[10]);
+    buf0[10] = _mm_sub_epi32(buf1[9], buf1[10]);
+    buf0[12] = _mm_sub_epi32(buf1[15], buf1[12]);
+    buf0[15] = _mm_add_epi32(buf1[15], buf1[12]);
+    buf0[13] = _mm_sub_epi32(buf1[14], buf1[13]);
+    buf0[14] = _mm_add_epi32(buf1[14], buf1[13]);
+    buf0[16] = buf1[16];
+    buf0[17] = buf1[17];
+    btf_32_sse4_1_type0(-cospi[16], cospi[48], buf1[18], buf1[29], buf0[18],
+                        buf0[29], bit);
+    btf_32_sse4_1_type0(-cospi[16], cospi[48], buf1[19], buf1[28], buf0[19],
+                        buf0[28], bit);
+    btf_32_sse4_1_type0(-cospi[48], -cospi[16], buf1[20], buf1[27], buf0[20],
+                        buf0[27], bit);
+    btf_32_sse4_1_type0(-cospi[48], -cospi[16], buf1[21], buf1[26], buf0[21],
+                        buf0[26], bit);
+    buf0[22] = buf1[22];
+    buf0[23] = buf1[23];
+    buf0[24] = buf1[24];
+    buf0[25] = buf1[25];
+    buf0[30] = buf1[30];
+    buf0[31] = buf1[31];
+
+    // stage 5
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    btf_32_sse4_1_type0(cospi[32], cospi[32], buf0[0], buf0[1], buf1[0],
+                        buf1[1], bit);
+    btf_32_sse4_1_type1(cospi[48], cospi[16], buf0[2], buf0[3], buf1[2],
+                        buf1[3], bit);
+    buf1[4] = _mm_add_epi32(buf0[4], buf0[5]);
+    buf1[5] = _mm_sub_epi32(buf0[4], buf0[5]);
+    buf1[6] = _mm_sub_epi32(buf0[7], buf0[6]);
+    buf1[7] = _mm_add_epi32(buf0[7], buf0[6]);
+    buf1[8] = buf0[8];
+    btf_32_sse4_1_type0(-cospi[16], cospi[48], buf0[9], buf0[14], buf1[9],
+                        buf1[14], bit);
+    btf_32_sse4_1_type0(-cospi[48], -cospi[16], buf0[10], buf0[13], buf1[10],
+                        buf1[13], bit);
+    buf1[11] = buf0[11];
+    buf1[12] = buf0[12];
+    buf1[15] = buf0[15];
+    buf1[16] = _mm_add_epi32(buf0[16], buf0[19]);
+    buf1[19] = _mm_sub_epi32(buf0[16], buf0[19]);
+    buf1[17] = _mm_add_epi32(buf0[17], buf0[18]);
+    buf1[18] = _mm_sub_epi32(buf0[17], buf0[18]);
+    buf1[20] = _mm_sub_epi32(buf0[23], buf0[20]);
+    buf1[23] = _mm_add_epi32(buf0[23], buf0[20]);
+    buf1[21] = _mm_sub_epi32(buf0[22], buf0[21]);
+    buf1[22] = _mm_add_epi32(buf0[22], buf0[21]);
+    buf1[24] = _mm_add_epi32(buf0[24], buf0[27]);
+    buf1[27] = _mm_sub_epi32(buf0[24], buf0[27]);
+    buf1[25] = _mm_add_epi32(buf0[25], buf0[26]);
+    buf1[26] = _mm_sub_epi32(buf0[25], buf0[26]);
+    buf1[28] = _mm_sub_epi32(buf0[31], buf0[28]);
+    buf1[31] = _mm_add_epi32(buf0[31], buf0[28]);
+    buf1[29] = _mm_sub_epi32(buf0[30], buf0[29]);
+    buf1[30] = _mm_add_epi32(buf0[30], buf0[29]);
+
+    // stage 6
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf0[0] = buf1[0];
+    buf0[1] = buf1[1];
+    buf0[2] = buf1[2];
+    buf0[3] = buf1[3];
+    btf_32_sse4_1_type1(cospi[56], cospi[8], buf1[4], buf1[7], buf0[4], buf0[7],
+                        bit);
+    btf_32_sse4_1_type1(cospi[24], cospi[40], buf1[5], buf1[6], buf0[5],
+                        buf0[6], bit);
+    buf0[8] = _mm_add_epi32(buf1[8], buf1[9]);
+    buf0[9] = _mm_sub_epi32(buf1[8], buf1[9]);
+    buf0[10] = _mm_sub_epi32(buf1[11], buf1[10]);
+    buf0[11] = _mm_add_epi32(buf1[11], buf1[10]);
+    buf0[12] = _mm_add_epi32(buf1[12], buf1[13]);
+    buf0[13] = _mm_sub_epi32(buf1[12], buf1[13]);
+    buf0[14] = _mm_sub_epi32(buf1[15], buf1[14]);
+    buf0[15] = _mm_add_epi32(buf1[15], buf1[14]);
+    buf0[16] = buf1[16];
+    btf_32_sse4_1_type0(-cospi[8], cospi[56], buf1[17], buf1[30], buf0[17],
+                        buf0[30], bit);
+    btf_32_sse4_1_type0(-cospi[56], -cospi[8], buf1[18], buf1[29], buf0[18],
+                        buf0[29], bit);
+    buf0[19] = buf1[19];
+    buf0[20] = buf1[20];
+    btf_32_sse4_1_type0(-cospi[40], cospi[24], buf1[21], buf1[26], buf0[21],
+                        buf0[26], bit);
+    btf_32_sse4_1_type0(-cospi[24], -cospi[40], buf1[22], buf1[25], buf0[22],
+                        buf0[25], bit);
+    buf0[23] = buf1[23];
+    buf0[24] = buf1[24];
+    buf0[27] = buf1[27];
+    buf0[28] = buf1[28];
+    buf0[31] = buf1[31];
+
+    // stage 7
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf1[0] = buf0[0];
+    buf1[1] = buf0[1];
+    buf1[2] = buf0[2];
+    buf1[3] = buf0[3];
+    buf1[4] = buf0[4];
+    buf1[5] = buf0[5];
+    buf1[6] = buf0[6];
+    buf1[7] = buf0[7];
+    btf_32_sse4_1_type1(cospi[60], cospi[4], buf0[8], buf0[15], buf1[8],
+                        buf1[15], bit);
+    btf_32_sse4_1_type1(cospi[28], cospi[36], buf0[9], buf0[14], buf1[9],
+                        buf1[14], bit);
+    btf_32_sse4_1_type1(cospi[44], cospi[20], buf0[10], buf0[13], buf1[10],
+                        buf1[13], bit);
+    btf_32_sse4_1_type1(cospi[12], cospi[52], buf0[11], buf0[12], buf1[11],
+                        buf1[12], bit);
+    buf1[16] = _mm_add_epi32(buf0[16], buf0[17]);
+    buf1[17] = _mm_sub_epi32(buf0[16], buf0[17]);
+    buf1[18] = _mm_sub_epi32(buf0[19], buf0[18]);
+    buf1[19] = _mm_add_epi32(buf0[19], buf0[18]);
+    buf1[20] = _mm_add_epi32(buf0[20], buf0[21]);
+    buf1[21] = _mm_sub_epi32(buf0[20], buf0[21]);
+    buf1[22] = _mm_sub_epi32(buf0[23], buf0[22]);
+    buf1[23] = _mm_add_epi32(buf0[23], buf0[22]);
+    buf1[24] = _mm_add_epi32(buf0[24], buf0[25]);
+    buf1[25] = _mm_sub_epi32(buf0[24], buf0[25]);
+    buf1[26] = _mm_sub_epi32(buf0[27], buf0[26]);
+    buf1[27] = _mm_add_epi32(buf0[27], buf0[26]);
+    buf1[28] = _mm_add_epi32(buf0[28], buf0[29]);
+    buf1[29] = _mm_sub_epi32(buf0[28], buf0[29]);
+    buf1[30] = _mm_sub_epi32(buf0[31], buf0[30]);
+    buf1[31] = _mm_add_epi32(buf0[31], buf0[30]);
+
+    // stage 8
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf0[0] = buf1[0];
+    buf0[1] = buf1[1];
+    buf0[2] = buf1[2];
+    buf0[3] = buf1[3];
+    buf0[4] = buf1[4];
+    buf0[5] = buf1[5];
+    buf0[6] = buf1[6];
+    buf0[7] = buf1[7];
+    buf0[8] = buf1[8];
+    buf0[9] = buf1[9];
+    buf0[10] = buf1[10];
+    buf0[11] = buf1[11];
+    buf0[12] = buf1[12];
+    buf0[13] = buf1[13];
+    buf0[14] = buf1[14];
+    buf0[15] = buf1[15];
+    btf_32_sse4_1_type1(cospi[62], cospi[2], buf1[16], buf1[31], buf0[16],
+                        buf0[31], bit);
+    btf_32_sse4_1_type1(cospi[30], cospi[34], buf1[17], buf1[30], buf0[17],
+                        buf0[30], bit);
+    btf_32_sse4_1_type1(cospi[46], cospi[18], buf1[18], buf1[29], buf0[18],
+                        buf0[29], bit);
+    btf_32_sse4_1_type1(cospi[14], cospi[50], buf1[19], buf1[28], buf0[19],
+                        buf0[28], bit);
+    btf_32_sse4_1_type1(cospi[54], cospi[10], buf1[20], buf1[27], buf0[20],
+                        buf0[27], bit);
+    btf_32_sse4_1_type1(cospi[22], cospi[42], buf1[21], buf1[26], buf0[21],
+                        buf0[26], bit);
+    btf_32_sse4_1_type1(cospi[38], cospi[26], buf1[22], buf1[25], buf0[22],
+                        buf0[25], bit);
+    btf_32_sse4_1_type1(cospi[6], cospi[58], buf1[23], buf1[24], buf0[23],
+                        buf0[24], bit);
+
+    // stage 9
+    stage_idx++;
+    buf1[0] = buf0[0];
+    buf1[1] = buf0[16];
+    buf1[2] = buf0[8];
+    buf1[3] = buf0[24];
+    buf1[4] = buf0[4];
+    buf1[5] = buf0[20];
+    buf1[6] = buf0[12];
+    buf1[7] = buf0[28];
+    buf1[8] = buf0[2];
+    buf1[9] = buf0[18];
+    buf1[10] = buf0[10];
+    buf1[11] = buf0[26];
+    buf1[12] = buf0[6];
+    buf1[13] = buf0[22];
+    buf1[14] = buf0[14];
+    buf1[15] = buf0[30];
+    buf1[16] = buf0[1];
+    buf1[17] = buf0[17];
+    buf1[18] = buf0[9];
+    buf1[19] = buf0[25];
+    buf1[20] = buf0[5];
+    buf1[21] = buf0[21];
+    buf1[22] = buf0[13];
+    buf1[23] = buf0[29];
+    buf1[24] = buf0[3];
+    buf1[25] = buf0[19];
+    buf1[26] = buf0[11];
+    buf1[27] = buf0[27];
+    buf1[28] = buf0[7];
+    buf1[29] = buf0[23];
+    buf1[30] = buf0[15];
+    buf1[31] = buf0[31];
+
+    for (j = 0; j < 32; ++j) {
+      output[j * col_num + col] = buf1[j];
+    }
+  }
+}
+
+void av1_fadst4_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range) {
+  const int txfm_size = 4;
+  const int num_per_128 = 4;
+  const int32_t *cospi;
+  __m128i buf0[4];
+  __m128i buf1[4];
+  int col_num = txfm_size / num_per_128;
+  int bit;
+  int col;
+  (void)stage_range;
+  for (col = 0; col < col_num; col++) {
+    // stage 0;
+    int32_t stage_idx = 0;
+    int j;
+    for (j = 0; j < 4; ++j) {
+      buf0[j] = input[j * col_num + col];
+    }
+
+    // stage 1
+    stage_idx++;
+    buf1[0] = buf0[3];
+    buf1[1] = buf0[0];
+    buf1[2] = buf0[1];
+    buf1[3] = buf0[2];
+
+    // stage 2
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    btf_32_sse4_1_type0(cospi[8], cospi[56], buf1[0], buf1[1], buf0[0], buf0[1],
+                        bit);
+    btf_32_sse4_1_type0(cospi[40], cospi[24], buf1[2], buf1[3], buf0[2],
+                        buf0[3], bit);
+
+    // stage 3
+    stage_idx++;
+    buf1[0] = _mm_add_epi32(buf0[0], buf0[2]);
+    buf1[2] = _mm_sub_epi32(buf0[0], buf0[2]);
+    buf1[1] = _mm_add_epi32(buf0[1], buf0[3]);
+    buf1[3] = _mm_sub_epi32(buf0[1], buf0[3]);
+
+    // stage 4
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf0[0] = buf1[0];
+    buf0[1] = buf1[1];
+    btf_32_sse4_1_type0(cospi[32], cospi[32], buf1[2], buf1[3], buf0[2],
+                        buf0[3], bit);
+
+    // stage 5
+    stage_idx++;
+    buf1[0] = buf0[0];
+    buf1[1] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[2]);
+    buf1[2] = buf0[3];
+    buf1[3] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[1]);
+
+    for (j = 0; j < 4; ++j) {
+      output[j * col_num + col] = buf1[j];
+    }
+  }
+}
+
+void av1_fadst32_new_sse4_1(const __m128i *input, __m128i *output,
+                            const int8_t *cos_bit, const int8_t *stage_range) {
+  const int txfm_size = 32;
+  const int num_per_128 = 4;
+  const int32_t *cospi;
+  __m128i buf0[32];
+  __m128i buf1[32];
+  int col_num = txfm_size / num_per_128;
+  int bit;
+  int col;
+  (void)stage_range;
+  for (col = 0; col < col_num; col++) {
+    // stage 0;
+    int32_t stage_idx = 0;
+    int j;
+    for (j = 0; j < 32; ++j) {
+      buf0[j] = input[j * col_num + col];
+    }
+
+    // stage 1
+    stage_idx++;
+    buf1[0] = buf0[31];
+    buf1[1] = buf0[0];
+    buf1[2] = buf0[29];
+    buf1[3] = buf0[2];
+    buf1[4] = buf0[27];
+    buf1[5] = buf0[4];
+    buf1[6] = buf0[25];
+    buf1[7] = buf0[6];
+    buf1[8] = buf0[23];
+    buf1[9] = buf0[8];
+    buf1[10] = buf0[21];
+    buf1[11] = buf0[10];
+    buf1[12] = buf0[19];
+    buf1[13] = buf0[12];
+    buf1[14] = buf0[17];
+    buf1[15] = buf0[14];
+    buf1[16] = buf0[15];
+    buf1[17] = buf0[16];
+    buf1[18] = buf0[13];
+    buf1[19] = buf0[18];
+    buf1[20] = buf0[11];
+    buf1[21] = buf0[20];
+    buf1[22] = buf0[9];
+    buf1[23] = buf0[22];
+    buf1[24] = buf0[7];
+    buf1[25] = buf0[24];
+    buf1[26] = buf0[5];
+    buf1[27] = buf0[26];
+    buf1[28] = buf0[3];
+    buf1[29] = buf0[28];
+    buf1[30] = buf0[1];
+    buf1[31] = buf0[30];
+
+    // stage 2
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    btf_32_sse4_1_type0(cospi[1], cospi[63], buf1[0], buf1[1], buf0[0], buf0[1],
+                        bit);
+    btf_32_sse4_1_type0(cospi[5], cospi[59], buf1[2], buf1[3], buf0[2], buf0[3],
+                        bit);
+    btf_32_sse4_1_type0(cospi[9], cospi[55], buf1[4], buf1[5], buf0[4], buf0[5],
+                        bit);
+    btf_32_sse4_1_type0(cospi[13], cospi[51], buf1[6], buf1[7], buf0[6],
+                        buf0[7], bit);
+    btf_32_sse4_1_type0(cospi[17], cospi[47], buf1[8], buf1[9], buf0[8],
+                        buf0[9], bit);
+    btf_32_sse4_1_type0(cospi[21], cospi[43], buf1[10], buf1[11], buf0[10],
+                        buf0[11], bit);
+    btf_32_sse4_1_type0(cospi[25], cospi[39], buf1[12], buf1[13], buf0[12],
+                        buf0[13], bit);
+    btf_32_sse4_1_type0(cospi[29], cospi[35], buf1[14], buf1[15], buf0[14],
+                        buf0[15], bit);
+    btf_32_sse4_1_type0(cospi[33], cospi[31], buf1[16], buf1[17], buf0[16],
+                        buf0[17], bit);
+    btf_32_sse4_1_type0(cospi[37], cospi[27], buf1[18], buf1[19], buf0[18],
+                        buf0[19], bit);
+    btf_32_sse4_1_type0(cospi[41], cospi[23], buf1[20], buf1[21], buf0[20],
+                        buf0[21], bit);
+    btf_32_sse4_1_type0(cospi[45], cospi[19], buf1[22], buf1[23], buf0[22],
+                        buf0[23], bit);
+    btf_32_sse4_1_type0(cospi[49], cospi[15], buf1[24], buf1[25], buf0[24],
+                        buf0[25], bit);
+    btf_32_sse4_1_type0(cospi[53], cospi[11], buf1[26], buf1[27], buf0[26],
+                        buf0[27], bit);
+    btf_32_sse4_1_type0(cospi[57], cospi[7], buf1[28], buf1[29], buf0[28],
+                        buf0[29], bit);
+    btf_32_sse4_1_type0(cospi[61], cospi[3], buf1[30], buf1[31], buf0[30],
+                        buf0[31], bit);
+
+    // stage 3
+    stage_idx++;
+    buf1[0] = _mm_add_epi32(buf0[0], buf0[16]);
+    buf1[16] = _mm_sub_epi32(buf0[0], buf0[16]);
+    buf1[1] = _mm_add_epi32(buf0[1], buf0[17]);
+    buf1[17] = _mm_sub_epi32(buf0[1], buf0[17]);
+    buf1[2] = _mm_add_epi32(buf0[2], buf0[18]);
+    buf1[18] = _mm_sub_epi32(buf0[2], buf0[18]);
+    buf1[3] = _mm_add_epi32(buf0[3], buf0[19]);
+    buf1[19] = _mm_sub_epi32(buf0[3], buf0[19]);
+    buf1[4] = _mm_add_epi32(buf0[4], buf0[20]);
+    buf1[20] = _mm_sub_epi32(buf0[4], buf0[20]);
+    buf1[5] = _mm_add_epi32(buf0[5], buf0[21]);
+    buf1[21] = _mm_sub_epi32(buf0[5], buf0[21]);
+    buf1[6] = _mm_add_epi32(buf0[6], buf0[22]);
+    buf1[22] = _mm_sub_epi32(buf0[6], buf0[22]);
+    buf1[7] = _mm_add_epi32(buf0[7], buf0[23]);
+    buf1[23] = _mm_sub_epi32(buf0[7], buf0[23]);
+    buf1[8] = _mm_add_epi32(buf0[8], buf0[24]);
+    buf1[24] = _mm_sub_epi32(buf0[8], buf0[24]);
+    buf1[9] = _mm_add_epi32(buf0[9], buf0[25]);
+    buf1[25] = _mm_sub_epi32(buf0[9], buf0[25]);
+    buf1[10] = _mm_add_epi32(buf0[10], buf0[26]);
+    buf1[26] = _mm_sub_epi32(buf0[10], buf0[26]);
+    buf1[11] = _mm_add_epi32(buf0[11], buf0[27]);
+    buf1[27] = _mm_sub_epi32(buf0[11], buf0[27]);
+    buf1[12] = _mm_add_epi32(buf0[12], buf0[28]);
+    buf1[28] = _mm_sub_epi32(buf0[12], buf0[28]);
+    buf1[13] = _mm_add_epi32(buf0[13], buf0[29]);
+    buf1[29] = _mm_sub_epi32(buf0[13], buf0[29]);
+    buf1[14] = _mm_add_epi32(buf0[14], buf0[30]);
+    buf1[30] = _mm_sub_epi32(buf0[14], buf0[30]);
+    buf1[15] = _mm_add_epi32(buf0[15], buf0[31]);
+    buf1[31] = _mm_sub_epi32(buf0[15], buf0[31]);
+
+    // stage 4
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf0[0] = buf1[0];
+    buf0[1] = buf1[1];
+    buf0[2] = buf1[2];
+    buf0[3] = buf1[3];
+    buf0[4] = buf1[4];
+    buf0[5] = buf1[5];
+    buf0[6] = buf1[6];
+    buf0[7] = buf1[7];
+    buf0[8] = buf1[8];
+    buf0[9] = buf1[9];
+    buf0[10] = buf1[10];
+    buf0[11] = buf1[11];
+    buf0[12] = buf1[12];
+    buf0[13] = buf1[13];
+    buf0[14] = buf1[14];
+    buf0[15] = buf1[15];
+    btf_32_sse4_1_type0(cospi[4], cospi[60], buf1[16], buf1[17], buf0[16],
+                        buf0[17], bit);
+    btf_32_sse4_1_type0(cospi[20], cospi[44], buf1[18], buf1[19], buf0[18],
+                        buf0[19], bit);
+    btf_32_sse4_1_type0(cospi[36], cospi[28], buf1[20], buf1[21], buf0[20],
+                        buf0[21], bit);
+    btf_32_sse4_1_type0(cospi[52], cospi[12], buf1[22], buf1[23], buf0[22],
+                        buf0[23], bit);
+    btf_32_sse4_1_type0(-cospi[60], cospi[4], buf1[24], buf1[25], buf0[24],
+                        buf0[25], bit);
+    btf_32_sse4_1_type0(-cospi[44], cospi[20], buf1[26], buf1[27], buf0[26],
+                        buf0[27], bit);
+    btf_32_sse4_1_type0(-cospi[28], cospi[36], buf1[28], buf1[29], buf0[28],
+                        buf0[29], bit);
+    btf_32_sse4_1_type0(-cospi[12], cospi[52], buf1[30], buf1[31], buf0[30],
+                        buf0[31], bit);
+
+    // stage 5
+    stage_idx++;
+    buf1[0] = _mm_add_epi32(buf0[0], buf0[8]);
+    buf1[8] = _mm_sub_epi32(buf0[0], buf0[8]);
+    buf1[1] = _mm_add_epi32(buf0[1], buf0[9]);
+    buf1[9] = _mm_sub_epi32(buf0[1], buf0[9]);
+    buf1[2] = _mm_add_epi32(buf0[2], buf0[10]);
+    buf1[10] = _mm_sub_epi32(buf0[2], buf0[10]);
+    buf1[3] = _mm_add_epi32(buf0[3], buf0[11]);
+    buf1[11] = _mm_sub_epi32(buf0[3], buf0[11]);
+    buf1[4] = _mm_add_epi32(buf0[4], buf0[12]);
+    buf1[12] = _mm_sub_epi32(buf0[4], buf0[12]);
+    buf1[5] = _mm_add_epi32(buf0[5], buf0[13]);
+    buf1[13] = _mm_sub_epi32(buf0[5], buf0[13]);
+    buf1[6] = _mm_add_epi32(buf0[6], buf0[14]);
+    buf1[14] = _mm_sub_epi32(buf0[6], buf0[14]);
+    buf1[7] = _mm_add_epi32(buf0[7], buf0[15]);
+    buf1[15] = _mm_sub_epi32(buf0[7], buf0[15]);
+    buf1[16] = _mm_add_epi32(buf0[16], buf0[24]);
+    buf1[24] = _mm_sub_epi32(buf0[16], buf0[24]);
+    buf1[17] = _mm_add_epi32(buf0[17], buf0[25]);
+    buf1[25] = _mm_sub_epi32(buf0[17], buf0[25]);
+    buf1[18] = _mm_add_epi32(buf0[18], buf0[26]);
+    buf1[26] = _mm_sub_epi32(buf0[18], buf0[26]);
+    buf1[19] = _mm_add_epi32(buf0[19], buf0[27]);
+    buf1[27] = _mm_sub_epi32(buf0[19], buf0[27]);
+    buf1[20] = _mm_add_epi32(buf0[20], buf0[28]);
+    buf1[28] = _mm_sub_epi32(buf0[20], buf0[28]);
+    buf1[21] = _mm_add_epi32(buf0[21], buf0[29]);
+    buf1[29] = _mm_sub_epi32(buf0[21], buf0[29]);
+    buf1[22] = _mm_add_epi32(buf0[22], buf0[30]);
+    buf1[30] = _mm_sub_epi32(buf0[22], buf0[30]);
+    buf1[23] = _mm_add_epi32(buf0[23], buf0[31]);
+    buf1[31] = _mm_sub_epi32(buf0[23], buf0[31]);
+
+    // stage 6
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf0[0] = buf1[0];
+    buf0[1] = buf1[1];
+    buf0[2] = buf1[2];
+    buf0[3] = buf1[3];
+    buf0[4] = buf1[4];
+    buf0[5] = buf1[5];
+    buf0[6] = buf1[6];
+    buf0[7] = buf1[7];
+    btf_32_sse4_1_type0(cospi[8], cospi[56], buf1[8], buf1[9], buf0[8], buf0[9],
+                        bit);
+    btf_32_sse4_1_type0(cospi[40], cospi[24], buf1[10], buf1[11], buf0[10],
+                        buf0[11], bit);
+    btf_32_sse4_1_type0(-cospi[56], cospi[8], buf1[12], buf1[13], buf0[12],
+                        buf0[13], bit);
+    btf_32_sse4_1_type0(-cospi[24], cospi[40], buf1[14], buf1[15], buf0[14],
+                        buf0[15], bit);
+    buf0[16] = buf1[16];
+    buf0[17] = buf1[17];
+    buf0[18] = buf1[18];
+    buf0[19] = buf1[19];
+    buf0[20] = buf1[20];
+    buf0[21] = buf1[21];
+    buf0[22] = buf1[22];
+    buf0[23] = buf1[23];
+    btf_32_sse4_1_type0(cospi[8], cospi[56], buf1[24], buf1[25], buf0[24],
+                        buf0[25], bit);
+    btf_32_sse4_1_type0(cospi[40], cospi[24], buf1[26], buf1[27], buf0[26],
+                        buf0[27], bit);
+    btf_32_sse4_1_type0(-cospi[56], cospi[8], buf1[28], buf1[29], buf0[28],
+                        buf0[29], bit);
+    btf_32_sse4_1_type0(-cospi[24], cospi[40], buf1[30], buf1[31], buf0[30],
+                        buf0[31], bit);
+
+    // stage 7
+    stage_idx++;
+    buf1[0] = _mm_add_epi32(buf0[0], buf0[4]);
+    buf1[4] = _mm_sub_epi32(buf0[0], buf0[4]);
+    buf1[1] = _mm_add_epi32(buf0[1], buf0[5]);
+    buf1[5] = _mm_sub_epi32(buf0[1], buf0[5]);
+    buf1[2] = _mm_add_epi32(buf0[2], buf0[6]);
+    buf1[6] = _mm_sub_epi32(buf0[2], buf0[6]);
+    buf1[3] = _mm_add_epi32(buf0[3], buf0[7]);
+    buf1[7] = _mm_sub_epi32(buf0[3], buf0[7]);
+    buf1[8] = _mm_add_epi32(buf0[8], buf0[12]);
+    buf1[12] = _mm_sub_epi32(buf0[8], buf0[12]);
+    buf1[9] = _mm_add_epi32(buf0[9], buf0[13]);
+    buf1[13] = _mm_sub_epi32(buf0[9], buf0[13]);
+    buf1[10] = _mm_add_epi32(buf0[10], buf0[14]);
+    buf1[14] = _mm_sub_epi32(buf0[10], buf0[14]);
+    buf1[11] = _mm_add_epi32(buf0[11], buf0[15]);
+    buf1[15] = _mm_sub_epi32(buf0[11], buf0[15]);
+    buf1[16] = _mm_add_epi32(buf0[16], buf0[20]);
+    buf1[20] = _mm_sub_epi32(buf0[16], buf0[20]);
+    buf1[17] = _mm_add_epi32(buf0[17], buf0[21]);
+    buf1[21] = _mm_sub_epi32(buf0[17], buf0[21]);
+    buf1[18] = _mm_add_epi32(buf0[18], buf0[22]);
+    buf1[22] = _mm_sub_epi32(buf0[18], buf0[22]);
+    buf1[19] = _mm_add_epi32(buf0[19], buf0[23]);
+    buf1[23] = _mm_sub_epi32(buf0[19], buf0[23]);
+    buf1[24] = _mm_add_epi32(buf0[24], buf0[28]);
+    buf1[28] = _mm_sub_epi32(buf0[24], buf0[28]);
+    buf1[25] = _mm_add_epi32(buf0[25], buf0[29]);
+    buf1[29] = _mm_sub_epi32(buf0[25], buf0[29]);
+    buf1[26] = _mm_add_epi32(buf0[26], buf0[30]);
+    buf1[30] = _mm_sub_epi32(buf0[26], buf0[30]);
+    buf1[27] = _mm_add_epi32(buf0[27], buf0[31]);
+    buf1[31] = _mm_sub_epi32(buf0[27], buf0[31]);
+
+    // stage 8
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf0[0] = buf1[0];
+    buf0[1] = buf1[1];
+    buf0[2] = buf1[2];
+    buf0[3] = buf1[3];
+    btf_32_sse4_1_type0(cospi[16], cospi[48], buf1[4], buf1[5], buf0[4],
+                        buf0[5], bit);
+    btf_32_sse4_1_type0(-cospi[48], cospi[16], buf1[6], buf1[7], buf0[6],
+                        buf0[7], bit);
+    buf0[8] = buf1[8];
+    buf0[9] = buf1[9];
+    buf0[10] = buf1[10];
+    buf0[11] = buf1[11];
+    btf_32_sse4_1_type0(cospi[16], cospi[48], buf1[12], buf1[13], buf0[12],
+                        buf0[13], bit);
+    btf_32_sse4_1_type0(-cospi[48], cospi[16], buf1[14], buf1[15], buf0[14],
+                        buf0[15], bit);
+    buf0[16] = buf1[16];
+    buf0[17] = buf1[17];
+    buf0[18] = buf1[18];
+    buf0[19] = buf1[19];
+    btf_32_sse4_1_type0(cospi[16], cospi[48], buf1[20], buf1[21], buf0[20],
+                        buf0[21], bit);
+    btf_32_sse4_1_type0(-cospi[48], cospi[16], buf1[22], buf1[23], buf0[22],
+                        buf0[23], bit);
+    buf0[24] = buf1[24];
+    buf0[25] = buf1[25];
+    buf0[26] = buf1[26];
+    buf0[27] = buf1[27];
+    btf_32_sse4_1_type0(cospi[16], cospi[48], buf1[28], buf1[29], buf0[28],
+                        buf0[29], bit);
+    btf_32_sse4_1_type0(-cospi[48], cospi[16], buf1[30], buf1[31], buf0[30],
+                        buf0[31], bit);
+
+    // stage 9
+    stage_idx++;
+    buf1[0] = _mm_add_epi32(buf0[0], buf0[2]);
+    buf1[2] = _mm_sub_epi32(buf0[0], buf0[2]);
+    buf1[1] = _mm_add_epi32(buf0[1], buf0[3]);
+    buf1[3] = _mm_sub_epi32(buf0[1], buf0[3]);
+    buf1[4] = _mm_add_epi32(buf0[4], buf0[6]);
+    buf1[6] = _mm_sub_epi32(buf0[4], buf0[6]);
+    buf1[5] = _mm_add_epi32(buf0[5], buf0[7]);
+    buf1[7] = _mm_sub_epi32(buf0[5], buf0[7]);
+    buf1[8] = _mm_add_epi32(buf0[8], buf0[10]);
+    buf1[10] = _mm_sub_epi32(buf0[8], buf0[10]);
+    buf1[9] = _mm_add_epi32(buf0[9], buf0[11]);
+    buf1[11] = _mm_sub_epi32(buf0[9], buf0[11]);
+    buf1[12] = _mm_add_epi32(buf0[12], buf0[14]);
+    buf1[14] = _mm_sub_epi32(buf0[12], buf0[14]);
+    buf1[13] = _mm_add_epi32(buf0[13], buf0[15]);
+    buf1[15] = _mm_sub_epi32(buf0[13], buf0[15]);
+    buf1[16] = _mm_add_epi32(buf0[16], buf0[18]);
+    buf1[18] = _mm_sub_epi32(buf0[16], buf0[18]);
+    buf1[17] = _mm_add_epi32(buf0[17], buf0[19]);
+    buf1[19] = _mm_sub_epi32(buf0[17], buf0[19]);
+    buf1[20] = _mm_add_epi32(buf0[20], buf0[22]);
+    buf1[22] = _mm_sub_epi32(buf0[20], buf0[22]);
+    buf1[21] = _mm_add_epi32(buf0[21], buf0[23]);
+    buf1[23] = _mm_sub_epi32(buf0[21], buf0[23]);
+    buf1[24] = _mm_add_epi32(buf0[24], buf0[26]);
+    buf1[26] = _mm_sub_epi32(buf0[24], buf0[26]);
+    buf1[25] = _mm_add_epi32(buf0[25], buf0[27]);
+    buf1[27] = _mm_sub_epi32(buf0[25], buf0[27]);
+    buf1[28] = _mm_add_epi32(buf0[28], buf0[30]);
+    buf1[30] = _mm_sub_epi32(buf0[28], buf0[30]);
+    buf1[29] = _mm_add_epi32(buf0[29], buf0[31]);
+    buf1[31] = _mm_sub_epi32(buf0[29], buf0[31]);
+
+    // stage 10
+    stage_idx++;
+    bit = cos_bit[stage_idx];
+    cospi = cospi_arr[bit - cos_bit_min];
+    buf0[0] = buf1[0];
+    buf0[1] = buf1[1];
+    btf_32_sse4_1_type0(cospi[32], cospi[32], buf1[2], buf1[3], buf0[2],
+                        buf0[3], bit);
+    buf0[4] = buf1[4];
+    buf0[5] = buf1[5];
+    btf_32_sse4_1_type0(cospi[32], cospi[32], buf1[6], buf1[7], buf0[6],
+                        buf0[7], bit);
+    buf0[8] = buf1[8];
+    buf0[9] = buf1[9];
+    btf_32_sse4_1_type0(cospi[32], cospi[32], buf1[10], buf1[11], buf0[10],
+                        buf0[11], bit);
+    buf0[12] = buf1[12];
+    buf0[13] = buf1[13];
+    btf_32_sse4_1_type0(cospi[32], cospi[32], buf1[14], buf1[15], buf0[14],
+                        buf0[15], bit);
+    buf0[16] = buf1[16];
+    buf0[17] = buf1[17];
+    btf_32_sse4_1_type0(cospi[32], cospi[32], buf1[18], buf1[19], buf0[18],
+                        buf0[19], bit);
+    buf0[20] = buf1[20];
+    buf0[21] = buf1[21];
+    btf_32_sse4_1_type0(cospi[32], cospi[32], buf1[22], buf1[23], buf0[22],
+                        buf0[23], bit);
+    buf0[24] = buf1[24];
+    buf0[25] = buf1[25];
+    btf_32_sse4_1_type0(cospi[32], cospi[32], buf1[26], buf1[27], buf0[26],
+                        buf0[27], bit);
+    buf0[28] = buf1[28];
+    buf0[29] = buf1[29];
+    btf_32_sse4_1_type0(cospi[32], cospi[32], buf1[30], buf1[31], buf0[30],
+                        buf0[31], bit);
+
+    // stage 11
+    stage_idx++;
+    buf1[0] = buf0[0];
+    buf1[1] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[16]);
+    buf1[2] = buf0[24];
+    buf1[3] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[8]);
+    buf1[4] = buf0[12];
+    buf1[5] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[28]);
+    buf1[6] = buf0[20];
+    buf1[7] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[4]);
+    buf1[8] = buf0[6];
+    buf1[9] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[22]);
+    buf1[10] = buf0[30];
+    buf1[11] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[14]);
+    buf1[12] = buf0[10];
+    buf1[13] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[26]);
+    buf1[14] = buf0[18];
+    buf1[15] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[2]);
+    buf1[16] = buf0[3];
+    buf1[17] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[19]);
+    buf1[18] = buf0[27];
+    buf1[19] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[11]);
+    buf1[20] = buf0[15];
+    buf1[21] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[31]);
+    buf1[22] = buf0[23];
+    buf1[23] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[7]);
+    buf1[24] = buf0[5];
+    buf1[25] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[21]);
+    buf1[26] = buf0[29];
+    buf1[27] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[13]);
+    buf1[28] = buf0[9];
+    buf1[29] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[25]);
+    buf1[30] = buf0[17];
+    buf1[31] = _mm_sub_epi32(_mm_set1_epi32(0), buf0[1]);
+
+    for (j = 0; j < 32; ++j) {
+      output[j * col_num + col] = buf1[j];
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/av1_fwd_txfm2d_sse4.c b/third_party/aom/av1/common/x86/av1_fwd_txfm2d_sse4.c
new file mode 100644
index 0000000000..78c261374a
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_fwd_txfm2d_sse4.c
@@ -0,0 +1,81 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "av1/common/enums.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/x86/av1_txfm1d_sse4.h"
+
+static INLINE void int16_array_with_stride_to_int32_array_without_stride(
+    const int16_t *input, int stride, int32_t *output, int txfm1d_size) {
+  int r, c;
+  for (r = 0; r < txfm1d_size; r++) {
+    for (c = 0; c < txfm1d_size; c++) {
+      output[r * txfm1d_size + c] = (int32_t)input[r * stride + c];
+    }
+  }
+}
+
+typedef void (*TxfmFuncSSE2)(const __m128i *input, __m128i *output,
+                             const int8_t *cos_bit, const int8_t *stage_range);
+
+static INLINE TxfmFuncSSE2 fwd_txfm_type_to_func(TXFM_TYPE txfm_type) {
+  switch (txfm_type) {
+    case TXFM_TYPE_DCT32: return av1_fdct32_new_sse4_1; break;
+    case TXFM_TYPE_ADST32: return av1_fadst32_new_sse4_1; break;
+    default: assert(0);
+  }
+  return NULL;
+}
+
+static INLINE void fwd_txfm2d_sse4_1(const int16_t *input, int32_t *output,
+                                     const int stride, const TXFM_2D_CFG *cfg,
+                                     int32_t *txfm_buf) {
+  const int txfm_size = cfg->txfm_size;
+  const int8_t *shift = cfg->shift;
+  const int8_t *stage_range_col = cfg->stage_range_col;
+  const int8_t *stage_range_row = cfg->stage_range_row;
+  const int8_t *cos_bit_col = cfg->cos_bit_col;
+  const int8_t *cos_bit_row = cfg->cos_bit_row;
+  const TxfmFuncSSE2 txfm_func_col = fwd_txfm_type_to_func(cfg->txfm_type_col);
+  const TxfmFuncSSE2 txfm_func_row = fwd_txfm_type_to_func(cfg->txfm_type_row);
+
+  __m128i *buf_128 = (__m128i *)txfm_buf;
+  __m128i *out_128 = (__m128i *)output;
+  int num_per_128 = 4;
+  int txfm2d_size_128 = txfm_size * txfm_size / num_per_128;
+
+  int16_array_with_stride_to_int32_array_without_stride(input, stride, txfm_buf,
+                                                        txfm_size);
+  round_shift_array_32_sse4_1(buf_128, out_128, txfm2d_size_128, -shift[0]);
+  txfm_func_col(out_128, buf_128, cos_bit_col, stage_range_col);
+  round_shift_array_32_sse4_1(buf_128, out_128, txfm2d_size_128, -shift[1]);
+  transpose_32(txfm_size, out_128, buf_128);
+  txfm_func_row(buf_128, out_128, cos_bit_row, stage_range_row);
+  round_shift_array_32_sse4_1(out_128, buf_128, txfm2d_size_128, -shift[2]);
+  transpose_32(txfm_size, buf_128, out_128);
+}
+
+void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output,
+                                 int stride, int tx_type, int bd) {
+  DECLARE_ALIGNED(16, int32_t, txfm_buf[1024]);
+  TXFM_2D_FLIP_CFG cfg = av1_get_fwd_txfm_cfg(tx_type, TX_32X32);
+  (void)bd;
+  fwd_txfm2d_sse4_1(input, output, stride, cfg.cfg, txfm_buf);
+}
+
+void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output,
+                                 int stride, int tx_type, int bd) {
+  DECLARE_ALIGNED(16, int32_t, txfm_buf[4096]);
+  TXFM_2D_FLIP_CFG cfg = av1_get_fwd_txfm_64x64_cfg(tx_type);
+  (void)bd;
+  fwd_txfm2d_sse4_1(input, output, stride, cfg.cfg, txfm_buf);
+}
diff --git a/third_party/aom/av1/common/x86/av1_highbd_convolve_sse4.c b/third_party/aom/av1/common/x86/av1_highbd_convolve_sse4.c
new file mode 100644
index 0000000000..cf6249bdc9
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_highbd_convolve_sse4.c
@@ -0,0 +1,533 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <smmintrin.h>
+
+#include "./av1_rtcd.h"
+#include "av1/common/filter.h"
+
+#if CONFIG_DUAL_FILTER
+DECLARE_ALIGNED(16, static int16_t, subpel_filters_sharp[15][6][8]);
+#endif
+
+#if USE_TEMPORALFILTER_12TAP
+DECLARE_ALIGNED(16, static int16_t, subpel_temporalfilter[15][6][8]);
+#endif
+
+typedef int16_t (*HbdSubpelFilterCoeffs)[8];
+
+typedef void (*TransposeSave)(int width, int pixelsNum, uint32_t *src,
+                              int src_stride, uint16_t *dst, int dst_stride,
+                              int bd);
+
+static INLINE HbdSubpelFilterCoeffs
+hbd_get_subpel_filter_ver_signal_dir(const InterpFilterParams p, int index) {
+#if CONFIG_DUAL_FILTER
+  if (p.interp_filter == MULTITAP_SHARP) {
+    return &subpel_filters_sharp[index][0];
+  }
+#endif
+#if USE_TEMPORALFILTER_12TAP
+  if (p.interp_filter == TEMPORALFILTER_12TAP) {
+    return &subpel_temporalfilter[index][0];
+  }
+#endif
+  (void)p;
+  (void)index;
+  return NULL;
+}
+
+static void init_simd_filter(const int16_t *filter_ptr, int taps,
+                             int16_t (*simd_filter)[6][8]) {
+  int shift;
+  int offset = (12 - taps) / 2;
+  for (shift = 1; shift < SUBPEL_SHIFTS; ++shift) {
+    const int16_t *filter_row = filter_ptr + shift * taps;
+    int i, j;
+    for (i = 0; i < 12; ++i) {
+      for (j = 0; j < 4; ++j) {
+        int r = i / 2;
+        int c = j * 2 + (i % 2);
+        if (i - offset >= 0 && i - offset < taps)
+          simd_filter[shift - 1][r][c] = filter_row[i - offset];
+        else
+          simd_filter[shift - 1][r][c] = 0;
+      }
+    }
+  }
+}
+
+void av1_highbd_convolve_init_sse4_1(void) {
+#if USE_TEMPORALFILTER_12TAP
+  {
+    InterpFilterParams filter_params =
+        av1_get_interp_filter_params(TEMPORALFILTER_12TAP);
+    int taps = filter_params.taps;
+    const int16_t *filter_ptr = filter_params.filter_ptr;
+    init_simd_filter(filter_ptr, taps, subpel_temporalfilter);
+  }
+#endif
+#if CONFIG_DUAL_FILTER
+  {
+    InterpFilterParams filter_params =
+        av1_get_interp_filter_params(MULTITAP_SHARP);
+    int taps = filter_params.taps;
+    const int16_t *filter_ptr = filter_params.filter_ptr;
+    init_simd_filter(filter_ptr, taps, subpel_filters_sharp);
+  }
+#endif
+}
+
+// pixelsNum 0: write all 4 pixels
+//           1/2/3: residual pixels 1/2/3
+static void writePixel(__m128i *u, int width, int pixelsNum, uint16_t *dst,
+                       int dst_stride) {
+  if (2 == width) {
+    if (0 == pixelsNum) {
+      *(int *)dst = _mm_cvtsi128_si32(u[0]);
+      *(int *)(dst + dst_stride) = _mm_cvtsi128_si32(u[1]);
+      *(int *)(dst + 2 * dst_stride) = _mm_cvtsi128_si32(u[2]);
+      *(int *)(dst + 3 * dst_stride) = _mm_cvtsi128_si32(u[3]);
+    } else if (1 == pixelsNum) {
+      *(int *)dst = _mm_cvtsi128_si32(u[0]);
+    } else if (2 == pixelsNum) {
+      *(int *)dst = _mm_cvtsi128_si32(u[0]);
+      *(int *)(dst + dst_stride) = _mm_cvtsi128_si32(u[1]);
+    } else if (3 == pixelsNum) {
+      *(int *)dst = _mm_cvtsi128_si32(u[0]);
+      *(int *)(dst + dst_stride) = _mm_cvtsi128_si32(u[1]);
+      *(int *)(dst + 2 * dst_stride) = _mm_cvtsi128_si32(u[2]);
+    }
+  } else {
+    if (0 == pixelsNum) {
+      _mm_storel_epi64((__m128i *)dst, u[0]);
+      _mm_storel_epi64((__m128i *)(dst + dst_stride), u[1]);
+      _mm_storel_epi64((__m128i *)(dst + 2 * dst_stride), u[2]);
+      _mm_storel_epi64((__m128i *)(dst + 3 * dst_stride), u[3]);
+    } else if (1 == pixelsNum) {
+      _mm_storel_epi64((__m128i *)dst, u[0]);
+    } else if (2 == pixelsNum) {
+      _mm_storel_epi64((__m128i *)dst, u[0]);
+      _mm_storel_epi64((__m128i *)(dst + dst_stride), u[1]);
+    } else if (3 == pixelsNum) {
+      _mm_storel_epi64((__m128i *)dst, u[0]);
+      _mm_storel_epi64((__m128i *)(dst + dst_stride), u[1]);
+      _mm_storel_epi64((__m128i *)(dst + 2 * dst_stride), u[2]);
+    }
+  }
+}
+
+// 16-bit pixels clip with bd (10/12)
+static void highbd_clip(__m128i *p, int numVecs, int bd) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
+  __m128i clamped, mask;
+  int i;
+
+  for (i = 0; i < numVecs; i++) {
+    mask = _mm_cmpgt_epi16(p[i], max);
+    clamped = _mm_andnot_si128(mask, p[i]);
+    mask = _mm_and_si128(mask, max);
+    clamped = _mm_or_si128(mask, clamped);
+    mask = _mm_cmpgt_epi16(clamped, zero);
+    p[i] = _mm_and_si128(clamped, mask);
+  }
+}
+
+static void transClipPixel(uint32_t *src, int src_stride, __m128i *u, int bd) {
+  __m128i v0, v1;
+  __m128i rnd = _mm_set1_epi32(1 << (FILTER_BITS - 1));
+
+  u[0] = _mm_loadu_si128((__m128i const *)src);
+  u[1] = _mm_loadu_si128((__m128i const *)(src + src_stride));
+  u[2] = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  u[3] = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+
+  u[0] = _mm_add_epi32(u[0], rnd);
+  u[1] = _mm_add_epi32(u[1], rnd);
+  u[2] = _mm_add_epi32(u[2], rnd);
+  u[3] = _mm_add_epi32(u[3], rnd);
+
+  u[0] = _mm_srai_epi32(u[0], FILTER_BITS);
+  u[1] = _mm_srai_epi32(u[1], FILTER_BITS);
+  u[2] = _mm_srai_epi32(u[2], FILTER_BITS);
+  u[3] = _mm_srai_epi32(u[3], FILTER_BITS);
+
+  u[0] = _mm_packus_epi32(u[0], u[1]);
+  u[1] = _mm_packus_epi32(u[2], u[3]);
+
+  highbd_clip(u, 2, bd);
+
+  v0 = _mm_unpacklo_epi16(u[0], u[1]);
+  v1 = _mm_unpackhi_epi16(u[0], u[1]);
+
+  u[0] = _mm_unpacklo_epi16(v0, v1);
+  u[2] = _mm_unpackhi_epi16(v0, v1);
+
+  u[1] = _mm_srli_si128(u[0], 8);
+  u[3] = _mm_srli_si128(u[2], 8);
+}
+
+// pixelsNum = 0     : all 4 rows of pixels will be saved.
+// pixelsNum = 1/2/3 : residual 1/2/4 rows of pixels will be saved.
+void trans_save_4x4(int width, int pixelsNum, uint32_t *src, int src_stride,
+                    uint16_t *dst, int dst_stride, int bd) {
+  __m128i u[4];
+  transClipPixel(src, src_stride, u, bd);
+  writePixel(u, width, pixelsNum, dst, dst_stride);
+}
+
+void trans_accum_save_4x4(int width, int pixelsNum, uint32_t *src,
+                          int src_stride, uint16_t *dst, int dst_stride,
+                          int bd) {
+  __m128i u[4], v[4];
+  const __m128i ones = _mm_set1_epi16(1);
+
+  transClipPixel(src, src_stride, u, bd);
+
+  v[0] = _mm_loadl_epi64((__m128i const *)dst);
+  v[1] = _mm_loadl_epi64((__m128i const *)(dst + dst_stride));
+  v[2] = _mm_loadl_epi64((__m128i const *)(dst + 2 * dst_stride));
+  v[3] = _mm_loadl_epi64((__m128i const *)(dst + 3 * dst_stride));
+
+  u[0] = _mm_add_epi16(u[0], v[0]);
+  u[1] = _mm_add_epi16(u[1], v[1]);
+  u[2] = _mm_add_epi16(u[2], v[2]);
+  u[3] = _mm_add_epi16(u[3], v[3]);
+
+  u[0] = _mm_add_epi16(u[0], ones);
+  u[1] = _mm_add_epi16(u[1], ones);
+  u[2] = _mm_add_epi16(u[2], ones);
+  u[3] = _mm_add_epi16(u[3], ones);
+
+  u[0] = _mm_srai_epi16(u[0], 1);
+  u[1] = _mm_srai_epi16(u[1], 1);
+  u[2] = _mm_srai_epi16(u[2], 1);
+  u[3] = _mm_srai_epi16(u[3], 1);
+
+  writePixel(u, width, pixelsNum, dst, dst_stride);
+}
+
+static TransposeSave transSaveTab[2] = { trans_save_4x4, trans_accum_save_4x4 };
+
+static INLINE void transpose_pair(__m128i *in, __m128i *out) {
+  __m128i x0, x1;
+
+  x0 = _mm_unpacklo_epi32(in[0], in[1]);
+  x1 = _mm_unpacklo_epi32(in[2], in[3]);
+
+  out[0] = _mm_unpacklo_epi64(x0, x1);
+  out[1] = _mm_unpackhi_epi64(x0, x1);
+
+  x0 = _mm_unpackhi_epi32(in[0], in[1]);
+  x1 = _mm_unpackhi_epi32(in[2], in[3]);
+
+  out[2] = _mm_unpacklo_epi64(x0, x1);
+  out[3] = _mm_unpackhi_epi64(x0, x1);
+
+  x0 = _mm_unpacklo_epi32(in[4], in[5]);
+  x1 = _mm_unpacklo_epi32(in[6], in[7]);
+
+  out[4] = _mm_unpacklo_epi64(x0, x1);
+  out[5] = _mm_unpackhi_epi64(x0, x1);
+}
+
+static void highbd_filter_horiz(const uint16_t *src, int src_stride, __m128i *f,
+                                int tapsNum, uint32_t *buf) {
+  __m128i u[8], v[6];
+
+  if (tapsNum == 10) {
+    src -= 1;
+  }
+
+  u[0] = _mm_loadu_si128((__m128i const *)src);
+  u[1] = _mm_loadu_si128((__m128i const *)(src + src_stride));
+  u[2] = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  u[3] = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+
+  u[4] = _mm_loadu_si128((__m128i const *)(src + 8));
+  u[5] = _mm_loadu_si128((__m128i const *)(src + src_stride + 8));
+  u[6] = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride + 8));
+  u[7] = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride + 8));
+
+  transpose_pair(u, v);
+
+  u[0] = _mm_madd_epi16(v[0], f[0]);
+  u[1] = _mm_madd_epi16(v[1], f[1]);
+  u[2] = _mm_madd_epi16(v[2], f[2]);
+  u[3] = _mm_madd_epi16(v[3], f[3]);
+  u[4] = _mm_madd_epi16(v[4], f[4]);
+  u[5] = _mm_madd_epi16(v[5], f[5]);
+
+  u[6] = _mm_min_epi32(u[2], u[3]);
+  u[7] = _mm_max_epi32(u[2], u[3]);
+
+  u[0] = _mm_add_epi32(u[0], u[1]);
+  u[0] = _mm_add_epi32(u[0], u[5]);
+  u[0] = _mm_add_epi32(u[0], u[4]);
+  u[0] = _mm_add_epi32(u[0], u[6]);
+  u[0] = _mm_add_epi32(u[0], u[7]);
+
+  _mm_storeu_si128((__m128i *)buf, u[0]);
+}
+
+void av1_highbd_convolve_horiz_sse4_1(const uint16_t *src, int src_stride,
+                                      uint16_t *dst, int dst_stride, int w,
+                                      int h,
+                                      const InterpFilterParams filter_params,
+                                      const int subpel_x_q4, int x_step_q4,
+                                      int avg, int bd) {
+  DECLARE_ALIGNED(16, uint32_t, temp[4 * 4]);
+  __m128i verf[6];
+  HbdSubpelFilterCoeffs vCoeffs;
+  const uint16_t *srcPtr;
+  const int tapsNum = filter_params.taps;
+  int i, col, count, blkResidu, blkHeight;
+  TransposeSave transSave = transSaveTab[avg];
+  (void)x_step_q4;
+
+  if (0 == subpel_x_q4 || 16 != x_step_q4) {
+    av1_highbd_convolve_horiz_c(src, src_stride, dst, dst_stride, w, h,
+                                filter_params, subpel_x_q4, x_step_q4, avg, bd);
+    return;
+  }
+
+  vCoeffs =
+      hbd_get_subpel_filter_ver_signal_dir(filter_params, subpel_x_q4 - 1);
+  if (!vCoeffs) {
+    av1_highbd_convolve_horiz_c(src, src_stride, dst, dst_stride, w, h,
+                                filter_params, subpel_x_q4, x_step_q4, avg, bd);
+    return;
+  }
+
+  verf[0] = *((const __m128i *)(vCoeffs));
+  verf[1] = *((const __m128i *)(vCoeffs + 1));
+  verf[2] = *((const __m128i *)(vCoeffs + 2));
+  verf[3] = *((const __m128i *)(vCoeffs + 3));
+  verf[4] = *((const __m128i *)(vCoeffs + 4));
+  verf[5] = *((const __m128i *)(vCoeffs + 5));
+
+  src -= (tapsNum >> 1) - 1;
+  srcPtr = src;
+
+  count = 0;
+  blkHeight = h >> 2;
+  blkResidu = h & 3;
+
+  while (blkHeight != 0) {
+    for (col = 0; col < w; col += 4) {
+      for (i = 0; i < 4; ++i) {
+        highbd_filter_horiz(srcPtr, src_stride, verf, tapsNum, temp + (i * 4));
+        srcPtr += 1;
+      }
+      transSave(w, 0, temp, 4, dst + col, dst_stride, bd);
+    }
+    count++;
+    srcPtr = src + count * src_stride * 4;
+    dst += dst_stride * 4;
+    blkHeight--;
+  }
+
+  if (blkResidu == 0) return;
+
+  for (col = 0; col < w; col += 4) {
+    for (i = 0; i < 4; ++i) {
+      highbd_filter_horiz(srcPtr, src_stride, verf, tapsNum, temp + (i * 4));
+      srcPtr += 1;
+    }
+    transSave(w, blkResidu, temp, 4, dst + col, dst_stride, bd);
+  }
+}
+
+// Vertical convolutional filter
+
+typedef void (*WritePixels)(__m128i *u, int bd, uint16_t *dst);
+
+static void highbdRndingPacks(__m128i *u) {
+  __m128i rnd = _mm_set1_epi32(1 << (FILTER_BITS - 1));
+  u[0] = _mm_add_epi32(u[0], rnd);
+  u[0] = _mm_srai_epi32(u[0], FILTER_BITS);
+  u[0] = _mm_packus_epi32(u[0], u[0]);
+}
+
+static void write2pixelsOnly(__m128i *u, int bd, uint16_t *dst) {
+  highbdRndingPacks(u);
+  highbd_clip(u, 1, bd);
+  *(uint32_t *)dst = _mm_cvtsi128_si32(u[0]);
+}
+
+static void write2pixelsAccum(__m128i *u, int bd, uint16_t *dst) {
+  __m128i v = _mm_loadl_epi64((__m128i const *)dst);
+  const __m128i ones = _mm_set1_epi16(1);
+
+  highbdRndingPacks(u);
+  highbd_clip(u, 1, bd);
+
+  v = _mm_add_epi16(v, u[0]);
+  v = _mm_add_epi16(v, ones);
+  v = _mm_srai_epi16(v, 1);
+  *(uint32_t *)dst = _mm_cvtsi128_si32(v);
+}
+
+WritePixels write2pixelsTab[2] = { write2pixelsOnly, write2pixelsAccum };
+
+static void write4pixelsOnly(__m128i *u, int bd, uint16_t *dst) {
+  highbdRndingPacks(u);
+  highbd_clip(u, 1, bd);
+  _mm_storel_epi64((__m128i *)dst, u[0]);
+}
+
+static void write4pixelsAccum(__m128i *u, int bd, uint16_t *dst) {
+  __m128i v = _mm_loadl_epi64((__m128i const *)dst);
+  const __m128i ones = _mm_set1_epi16(1);
+
+  highbdRndingPacks(u);
+  highbd_clip(u, 1, bd);
+
+  v = _mm_add_epi16(v, u[0]);
+  v = _mm_add_epi16(v, ones);
+  v = _mm_srai_epi16(v, 1);
+  _mm_storel_epi64((__m128i *)dst, v);
+}
+
+WritePixels write4pixelsTab[2] = { write4pixelsOnly, write4pixelsAccum };
+
+static void filter_vert_horiz_parallel(const uint16_t *src, int src_stride,
+                                       const __m128i *f, int taps,
+                                       uint16_t *dst, WritePixels saveFunc,
+                                       int bd) {
+  __m128i s[12];
+  __m128i zero = _mm_setzero_si128();
+  int i = 0;
+  int r = 0;
+
+  // TODO(luoyi) treat s[12] as a circular buffer in width = 2 case
+  if (10 == taps) {
+    i += 1;
+    s[0] = zero;
+  }
+  while (i < 12) {
+    s[i] = _mm_loadu_si128((__m128i const *)(src + r * src_stride));
+    i += 1;
+    r += 1;
+  }
+
+  s[0] = _mm_unpacklo_epi16(s[0], s[1]);
+  s[2] = _mm_unpacklo_epi16(s[2], s[3]);
+  s[4] = _mm_unpacklo_epi16(s[4], s[5]);
+  s[6] = _mm_unpacklo_epi16(s[6], s[7]);
+  s[8] = _mm_unpacklo_epi16(s[8], s[9]);
+  s[10] = _mm_unpacklo_epi16(s[10], s[11]);
+
+  s[0] = _mm_madd_epi16(s[0], f[0]);
+  s[2] = _mm_madd_epi16(s[2], f[1]);
+  s[4] = _mm_madd_epi16(s[4], f[2]);
+  s[6] = _mm_madd_epi16(s[6], f[3]);
+  s[8] = _mm_madd_epi16(s[8], f[4]);
+  s[10] = _mm_madd_epi16(s[10], f[5]);
+
+  s[1] = _mm_min_epi32(s[4], s[6]);
+  s[3] = _mm_max_epi32(s[4], s[6]);
+
+  s[0] = _mm_add_epi32(s[0], s[2]);
+  s[0] = _mm_add_epi32(s[0], s[10]);
+  s[0] = _mm_add_epi32(s[0], s[8]);
+  s[0] = _mm_add_epi32(s[0], s[1]);
+  s[0] = _mm_add_epi32(s[0], s[3]);
+
+  saveFunc(s, bd, dst);
+}
+
+static void highbd_filter_vert_compute_large(const uint16_t *src,
+                                             int src_stride, const __m128i *f,
+                                             int taps, int w, int h,
+                                             uint16_t *dst, int dst_stride,
+                                             int avg, int bd) {
+  int col;
+  int rowIndex = 0;
+  const uint16_t *src_ptr = src;
+  uint16_t *dst_ptr = dst;
+  const int step = 4;
+  WritePixels write4pixels = write4pixelsTab[avg];
+
+  do {
+    for (col = 0; col < w; col += step) {
+      filter_vert_horiz_parallel(src_ptr, src_stride, f, taps, dst_ptr,
+                                 write4pixels, bd);
+      src_ptr += step;
+      dst_ptr += step;
+    }
+    rowIndex++;
+    src_ptr = src + rowIndex * src_stride;
+    dst_ptr = dst + rowIndex * dst_stride;
+  } while (rowIndex < h);
+}
+
+static void highbd_filter_vert_compute_small(const uint16_t *src,
+                                             int src_stride, const __m128i *f,
+                                             int taps, int w, int h,
+                                             uint16_t *dst, int dst_stride,
+                                             int avg, int bd) {
+  int rowIndex = 0;
+  WritePixels write2pixels = write2pixelsTab[avg];
+  (void)w;
+
+  do {
+    filter_vert_horiz_parallel(src, src_stride, f, taps, dst, write2pixels, bd);
+    rowIndex++;
+    src += src_stride;
+    dst += dst_stride;
+  } while (rowIndex < h);
+}
+
+void av1_highbd_convolve_vert_sse4_1(const uint16_t *src, int src_stride,
+                                     uint16_t *dst, int dst_stride, int w,
+                                     int h,
+                                     const InterpFilterParams filter_params,
+                                     const int subpel_y_q4, int y_step_q4,
+                                     int avg, int bd) {
+  __m128i verf[6];
+  HbdSubpelFilterCoeffs vCoeffs;
+  const int tapsNum = filter_params.taps;
+
+  if (0 == subpel_y_q4 || 16 != y_step_q4) {
+    av1_highbd_convolve_vert_c(src, src_stride, dst, dst_stride, w, h,
+                               filter_params, subpel_y_q4, y_step_q4, avg, bd);
+    return;
+  }
+
+  vCoeffs =
+      hbd_get_subpel_filter_ver_signal_dir(filter_params, subpel_y_q4 - 1);
+  if (!vCoeffs) {
+    av1_highbd_convolve_vert_c(src, src_stride, dst, dst_stride, w, h,
+                               filter_params, subpel_y_q4, y_step_q4, avg, bd);
+    return;
+  }
+
+  verf[0] = *((const __m128i *)(vCoeffs));
+  verf[1] = *((const __m128i *)(vCoeffs + 1));
+  verf[2] = *((const __m128i *)(vCoeffs + 2));
+  verf[3] = *((const __m128i *)(vCoeffs + 3));
+  verf[4] = *((const __m128i *)(vCoeffs + 4));
+  verf[5] = *((const __m128i *)(vCoeffs + 5));
+
+  src -= src_stride * ((tapsNum >> 1) - 1);
+
+  if (w > 2) {
+    highbd_filter_vert_compute_large(src, src_stride, verf, tapsNum, w, h, dst,
+                                     dst_stride, avg, bd);
+  } else {
+    highbd_filter_vert_compute_small(src, src_stride, verf, tapsNum, w, h, dst,
+                                     dst_stride, avg, bd);
+  }
+}
diff --git a/third_party/aom/av1/common/x86/av1_txfm1d_sse4.h b/third_party/aom/av1/common/x86/av1_txfm1d_sse4.h
new file mode 100644
index 0000000000..af7afb7ee5
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_txfm1d_sse4.h
@@ -0,0 +1,144 @@
+#ifndef AV1_TXMF1D_SSE2_H_
+#define AV1_TXMF1D_SSE2_H_
+
+#include <smmintrin.h>
+#include "av1/common/av1_txfm.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_fdct4_new_sse4_1(const __m128i *input, __m128i *output,
+                          const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fdct8_new_sse4_1(const __m128i *input, __m128i *output,
+                          const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fdct16_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fdct32_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fdct64_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range);
+
+void av1_fadst4_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fadst8_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fadst16_new_sse4_1(const __m128i *input, __m128i *output,
+                            const int8_t *cos_bit, const int8_t *stage_range);
+void av1_fadst32_new_sse4_1(const __m128i *input, __m128i *output,
+                            const int8_t *cos_bit, const int8_t *stage_range);
+
+void av1_idct4_new_sse4_1(const __m128i *input, __m128i *output,
+                          const int8_t *cos_bit, const int8_t *stage_range);
+void av1_idct8_new_sse4_1(const __m128i *input, __m128i *output,
+                          const int8_t *cos_bit, const int8_t *stage_range);
+void av1_idct16_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range);
+void av1_idct32_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range);
+void av1_idct64_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range);
+
+void av1_iadst4_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range);
+void av1_iadst8_new_sse4_1(const __m128i *input, __m128i *output,
+                           const int8_t *cos_bit, const int8_t *stage_range);
+void av1_iadst16_new_sse4_1(const __m128i *input, __m128i *output,
+                            const int8_t *cos_bit, const int8_t *stage_range);
+void av1_iadst32_new_sse4_1(const __m128i *input, __m128i *output,
+                            const int8_t *cos_bit, const int8_t *stage_range);
+
+static INLINE void transpose_32_4x4(int stride, const __m128i *input,
+                                    __m128i *output) {
+  __m128i temp0 = _mm_unpacklo_epi32(input[0 * stride], input[2 * stride]);
+  __m128i temp1 = _mm_unpackhi_epi32(input[0 * stride], input[2 * stride]);
+  __m128i temp2 = _mm_unpacklo_epi32(input[1 * stride], input[3 * stride]);
+  __m128i temp3 = _mm_unpackhi_epi32(input[1 * stride], input[3 * stride]);
+
+  output[0 * stride] = _mm_unpacklo_epi32(temp0, temp2);
+  output[1 * stride] = _mm_unpackhi_epi32(temp0, temp2);
+  output[2 * stride] = _mm_unpacklo_epi32(temp1, temp3);
+  output[3 * stride] = _mm_unpackhi_epi32(temp1, temp3);
+}
+
+// the entire input block can be represent by a grid of 4x4 blocks
+// each 4x4 blocks can be represent by 4 vertical __m128i
+// we first transpose each 4x4 block internally
+// than transpose the grid
+static INLINE void transpose_32(int txfm_size, const __m128i *input,
+                                __m128i *output) {
+  const int num_per_128 = 4;
+  const int row_size = txfm_size;
+  const int col_size = txfm_size / num_per_128;
+  int r, c;
+
+  // transpose each 4x4 block internally
+  for (r = 0; r < row_size; r += 4) {
+    for (c = 0; c < col_size; c++) {
+      transpose_32_4x4(col_size, &input[r * col_size + c],
+                       &output[c * 4 * col_size + r / 4]);
+    }
+  }
+}
+
+static INLINE __m128i round_shift_32_sse4_1(__m128i vec, int bit) {
+  __m128i tmp, round;
+  round = _mm_set1_epi32(1 << (bit - 1));
+  tmp = _mm_add_epi32(vec, round);
+  return _mm_srai_epi32(tmp, bit);
+}
+
+static INLINE void round_shift_array_32_sse4_1(__m128i *input, __m128i *output,
+                                               const int size, const int bit) {
+  if (bit > 0) {
+    int i;
+    for (i = 0; i < size; i++) {
+      output[i] = round_shift_32_sse4_1(input[i], bit);
+    }
+  } else {
+    int i;
+    for (i = 0; i < size; i++) {
+      output[i] = _mm_slli_epi32(input[i], -bit);
+    }
+  }
+}
+
+// out0 = in0*w0 + in1*w1
+// out1 = -in1*w0 + in0*w1
+#define btf_32_sse4_1_type0(w0, w1, in0, in1, out0, out1, bit) \
+  do {                                                         \
+    __m128i ww0, ww1, in0_w0, in1_w1, in0_w1, in1_w0;          \
+    ww0 = _mm_set1_epi32(w0);                                  \
+    ww1 = _mm_set1_epi32(w1);                                  \
+    in0_w0 = _mm_mullo_epi32(in0, ww0);                        \
+    in1_w1 = _mm_mullo_epi32(in1, ww1);                        \
+    out0 = _mm_add_epi32(in0_w0, in1_w1);                      \
+    out0 = round_shift_32_sse4_1(out0, bit);                   \
+    in0_w1 = _mm_mullo_epi32(in0, ww1);                        \
+    in1_w0 = _mm_mullo_epi32(in1, ww0);                        \
+    out1 = _mm_sub_epi32(in0_w1, in1_w0);                      \
+    out1 = round_shift_32_sse4_1(out1, bit);                   \
+  } while (0)
+
+// out0 = in0*w0 + in1*w1
+// out1 = in1*w0 - in0*w1
+#define btf_32_sse4_1_type1(w0, w1, in0, in1, out0, out1, bit) \
+  do {                                                         \
+    __m128i ww0, ww1, in0_w0, in1_w1, in0_w1, in1_w0;          \
+    ww0 = _mm_set1_epi32(w0);                                  \
+    ww1 = _mm_set1_epi32(w1);                                  \
+    in0_w0 = _mm_mullo_epi32(in0, ww0);                        \
+    in1_w1 = _mm_mullo_epi32(in1, ww1);                        \
+    out0 = _mm_add_epi32(in0_w0, in1_w1);                      \
+    out0 = round_shift_32_sse4_1(out0, bit);                   \
+    in0_w1 = _mm_mullo_epi32(in0, ww1);                        \
+    in1_w0 = _mm_mullo_epi32(in1, ww0);                        \
+    out1 = _mm_sub_epi32(in1_w0, in0_w1);                      \
+    out1 = round_shift_32_sse4_1(out1, bit);                   \
+  } while (0)
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // AV1_TXMF1D_SSE2_H_
diff --git a/third_party/aom/av1/common/x86/filterintra_sse4.c b/third_party/aom/av1/common/x86/filterintra_sse4.c
new file mode 100644
index 0000000000..4f77da4463
--- /dev/null
+++ b/third_party/aom/av1/common/x86/filterintra_sse4.c
@@ -0,0 +1,898 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <smmintrin.h>
+
+#include "./av1_rtcd.h"
+#include "aom_ports/mem.h"
+#include "av1/common/enums.h"
+#include "av1/common/reconintra.h"
+
+#if USE_3TAP_INTRA_FILTER
+void filterintra_sse4_3tap_dummy_func(void);
+void filterintra_sse4_3tap_dummy_func(void) {}
+#else
+
+static INLINE void AddPixelsSmall(const uint8_t *above, const uint8_t *left,
+                                  __m128i *sum) {
+  const __m128i a = _mm_loadu_si128((const __m128i *)above);
+  const __m128i l = _mm_loadu_si128((const __m128i *)left);
+  const __m128i zero = _mm_setzero_si128();
+
+  __m128i u0 = _mm_unpacklo_epi8(a, zero);
+  __m128i u1 = _mm_unpacklo_epi8(l, zero);
+
+  sum[0] = _mm_add_epi16(u0, u1);
+}
+
+static INLINE int GetMeanValue4x4(const uint8_t *above, const uint8_t *left,
+                                  __m128i *params) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i sum_vector, u;
+  uint16_t sum_value;
+
+  AddPixelsSmall(above, left, &sum_vector);
+
+  sum_vector = _mm_hadd_epi16(sum_vector, zero);  // still has 2 values
+  u = _mm_srli_si128(sum_vector, 2);
+  sum_vector = _mm_add_epi16(sum_vector, u);
+
+  sum_value = _mm_extract_epi16(sum_vector, 0);
+  sum_value += 4;
+  sum_value >>= 3;
+  *params = _mm_set1_epi32(sum_value);
+  return sum_value;
+}
+
+static INLINE int GetMeanValue8x8(const uint8_t *above, const uint8_t *left,
+                                  __m128i *params) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i sum_vector, u;
+  uint16_t sum_value;
+
+  AddPixelsSmall(above, left, &sum_vector);
+
+  sum_vector = _mm_hadd_epi16(sum_vector, zero);  // still has 4 values
+  sum_vector = _mm_hadd_epi16(sum_vector, zero);  // still has 2 values
+
+  u = _mm_srli_si128(sum_vector, 2);
+  sum_vector = _mm_add_epi16(sum_vector, u);
+
+  sum_value = _mm_extract_epi16(sum_vector, 0);
+  sum_value += 8;
+  sum_value >>= 4;
+  *params = _mm_set1_epi32(sum_value);
+  return sum_value;
+}
+
+static INLINE void AddPixelsLarge(const uint8_t *above, const uint8_t *left,
+                                  __m128i *sum) {
+  const __m128i a = _mm_loadu_si128((const __m128i *)above);
+  const __m128i l = _mm_loadu_si128((const __m128i *)left);
+  const __m128i zero = _mm_setzero_si128();
+
+  __m128i u0 = _mm_unpacklo_epi8(a, zero);
+  __m128i u1 = _mm_unpacklo_epi8(l, zero);
+
+  sum[0] = _mm_add_epi16(u0, u1);
+
+  u0 = _mm_unpackhi_epi8(a, zero);
+  u1 = _mm_unpackhi_epi8(l, zero);
+
+  sum[0] = _mm_add_epi16(sum[0], u0);
+  sum[0] = _mm_add_epi16(sum[0], u1);
+}
+
+static INLINE int GetMeanValue16x16(const uint8_t *above, const uint8_t *left,
+                                    __m128i *params) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i sum_vector, u;
+  uint16_t sum_value;
+
+  AddPixelsLarge(above, left, &sum_vector);
+
+  sum_vector = _mm_hadd_epi16(sum_vector, zero);  // still has 4 values
+  sum_vector = _mm_hadd_epi16(sum_vector, zero);  // still has 2 values
+
+  u = _mm_srli_si128(sum_vector, 2);
+  sum_vector = _mm_add_epi16(sum_vector, u);
+
+  sum_value = _mm_extract_epi16(sum_vector, 0);
+  sum_value += 16;
+  sum_value >>= 5;
+  *params = _mm_set1_epi32(sum_value);
+  return sum_value;
+}
+
+static INLINE int GetMeanValue32x32(const uint8_t *above, const uint8_t *left,
+                                    __m128i *params) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i sum_vector[2], u;
+  uint16_t sum_value;
+
+  AddPixelsLarge(above, left, &sum_vector[0]);
+  AddPixelsLarge(above + 16, left + 16, &sum_vector[1]);
+
+  sum_vector[0] = _mm_add_epi16(sum_vector[0], sum_vector[1]);
+  sum_vector[0] = _mm_hadd_epi16(sum_vector[0], zero);  // still has 4 values
+  sum_vector[0] = _mm_hadd_epi16(sum_vector[0], zero);  // still has 2 values
+
+  u = _mm_srli_si128(sum_vector[0], 2);
+  sum_vector[0] = _mm_add_epi16(sum_vector[0], u);
+
+  sum_value = _mm_extract_epi16(sum_vector[0], 0);
+  sum_value += 32;
+  sum_value >>= 6;
+  *params = _mm_set1_epi32(sum_value);
+  return sum_value;
+}
+
+// Note:
+//  params[4] : mean value, 4 int32_t repetition
+//
+static INLINE int CalcRefPixelsMeanValue(const uint8_t *above,
+                                         const uint8_t *left, int bs,
+                                         __m128i *params) {
+  int meanValue = 0;
+  switch (bs) {
+    case 4: meanValue = GetMeanValue4x4(above, left, params); break;
+    case 8: meanValue = GetMeanValue8x8(above, left, params); break;
+    case 16: meanValue = GetMeanValue16x16(above, left, params); break;
+    case 32: meanValue = GetMeanValue32x32(above, left, params); break;
+    default: assert(0);
+  }
+  return meanValue;
+}
+
+// Note:
+//  params[0-3] : 4-tap filter coefficients (int32_t per coefficient)
+//
+static INLINE void GetIntraFilterParams(int bs, int mode, __m128i *params) {
+  const TX_SIZE tx_size =
+      (bs == 32) ? TX_32X32
+                 : ((bs == 16) ? TX_16X16 : ((bs == 8) ? TX_8X8 : (TX_4X4)));
+  // c0
+  params[0] = _mm_set_epi32(av1_filter_intra_taps_4[tx_size][mode][0],
+                            av1_filter_intra_taps_4[tx_size][mode][0],
+                            av1_filter_intra_taps_4[tx_size][mode][0],
+                            av1_filter_intra_taps_4[tx_size][mode][0]);
+  // c1
+  params[1] = _mm_set_epi32(av1_filter_intra_taps_4[tx_size][mode][1],
+                            av1_filter_intra_taps_4[tx_size][mode][1],
+                            av1_filter_intra_taps_4[tx_size][mode][1],
+                            av1_filter_intra_taps_4[tx_size][mode][1]);
+  // c2
+  params[2] = _mm_set_epi32(av1_filter_intra_taps_4[tx_size][mode][2],
+                            av1_filter_intra_taps_4[tx_size][mode][2],
+                            av1_filter_intra_taps_4[tx_size][mode][2],
+                            av1_filter_intra_taps_4[tx_size][mode][2]);
+  // c3
+  params[3] = _mm_set_epi32(av1_filter_intra_taps_4[tx_size][mode][3],
+                            av1_filter_intra_taps_4[tx_size][mode][3],
+                            av1_filter_intra_taps_4[tx_size][mode][3],
+                            av1_filter_intra_taps_4[tx_size][mode][3]);
+}
+
+static const int maxBlkSize = 32;
+
+static INLINE void SavePred4x4(int *pred, const __m128i *mean, uint8_t *dst,
+                               ptrdiff_t stride) {
+  const int predStride = (maxBlkSize << 1) + 1;
+  __m128i p0 = _mm_loadu_si128((const __m128i *)pred);
+  __m128i p1 = _mm_loadu_si128((const __m128i *)(pred + predStride));
+  __m128i p2 = _mm_loadu_si128((const __m128i *)(pred + 2 * predStride));
+  __m128i p3 = _mm_loadu_si128((const __m128i *)(pred + 3 * predStride));
+
+  p0 = _mm_add_epi32(p0, mean[0]);
+  p1 = _mm_add_epi32(p1, mean[0]);
+  p2 = _mm_add_epi32(p2, mean[0]);
+  p3 = _mm_add_epi32(p3, mean[0]);
+
+  p0 = _mm_packus_epi32(p0, p1);
+  p1 = _mm_packus_epi32(p2, p3);
+  p0 = _mm_packus_epi16(p0, p1);
+
+  *((int *)dst) = _mm_cvtsi128_si32(p0);
+  p0 = _mm_srli_si128(p0, 4);
+  *((int *)(dst + stride)) = _mm_cvtsi128_si32(p0);
+  p0 = _mm_srli_si128(p0, 4);
+  *((int *)(dst + 2 * stride)) = _mm_cvtsi128_si32(p0);
+  p0 = _mm_srli_si128(p0, 4);
+  *((int *)(dst + 3 * stride)) = _mm_cvtsi128_si32(p0);
+}
+
+static void SavePred8x8(int *pred, const __m128i *mean, uint8_t *dst,
+                        ptrdiff_t stride) {
+  const int predStride = (maxBlkSize << 1) + 1;
+  __m128i p0, p1, p2, p3;
+  int r = 0;
+
+  while (r < 8) {
+    p0 = _mm_loadu_si128((const __m128i *)(pred + r * predStride));
+    p1 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 4));
+    r += 1;
+    p2 = _mm_loadu_si128((const __m128i *)(pred + r * predStride));
+    p3 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 4));
+
+    p0 = _mm_add_epi32(p0, mean[0]);
+    p1 = _mm_add_epi32(p1, mean[0]);
+    p2 = _mm_add_epi32(p2, mean[0]);
+    p3 = _mm_add_epi32(p3, mean[0]);
+
+    p0 = _mm_packus_epi32(p0, p1);
+    p1 = _mm_packus_epi32(p2, p3);
+    p0 = _mm_packus_epi16(p0, p1);
+
+    _mm_storel_epi64((__m128i *)dst, p0);
+    dst += stride;
+    p0 = _mm_srli_si128(p0, 8);
+    _mm_storel_epi64((__m128i *)dst, p0);
+    dst += stride;
+    r += 1;
+  }
+}
+
+static void SavePred16x16(int *pred, const __m128i *mean, uint8_t *dst,
+                          ptrdiff_t stride) {
+  const int predStride = (maxBlkSize << 1) + 1;
+  __m128i p0, p1, p2, p3;
+  int r = 0;
+
+  while (r < 16) {
+    p0 = _mm_loadu_si128((const __m128i *)(pred + r * predStride));
+    p1 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 4));
+    p2 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 8));
+    p3 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 12));
+
+    p0 = _mm_add_epi32(p0, mean[0]);
+    p1 = _mm_add_epi32(p1, mean[0]);
+    p2 = _mm_add_epi32(p2, mean[0]);
+    p3 = _mm_add_epi32(p3, mean[0]);
+
+    p0 = _mm_packus_epi32(p0, p1);
+    p1 = _mm_packus_epi32(p2, p3);
+    p0 = _mm_packus_epi16(p0, p1);
+
+    _mm_storel_epi64((__m128i *)dst, p0);
+    p0 = _mm_srli_si128(p0, 8);
+    _mm_storel_epi64((__m128i *)(dst + 8), p0);
+    dst += stride;
+    r += 1;
+  }
+}
+
+static void SavePred32x32(int *pred, const __m128i *mean, uint8_t *dst,
+                          ptrdiff_t stride) {
+  const int predStride = (maxBlkSize << 1) + 1;
+  __m128i p0, p1, p2, p3, p4, p5, p6, p7;
+  int r = 0;
+
+  while (r < 32) {
+    p0 = _mm_loadu_si128((const __m128i *)(pred + r * predStride));
+    p1 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 4));
+    p2 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 8));
+    p3 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 12));
+
+    p4 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 16));
+    p5 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 20));
+    p6 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 24));
+    p7 = _mm_loadu_si128((const __m128i *)(pred + r * predStride + 28));
+
+    p0 = _mm_add_epi32(p0, mean[0]);
+    p1 = _mm_add_epi32(p1, mean[0]);
+    p2 = _mm_add_epi32(p2, mean[0]);
+    p3 = _mm_add_epi32(p3, mean[0]);
+
+    p4 = _mm_add_epi32(p4, mean[0]);
+    p5 = _mm_add_epi32(p5, mean[0]);
+    p6 = _mm_add_epi32(p6, mean[0]);
+    p7 = _mm_add_epi32(p7, mean[0]);
+
+    p0 = _mm_packus_epi32(p0, p1);
+    p1 = _mm_packus_epi32(p2, p3);
+    p0 = _mm_packus_epi16(p0, p1);
+
+    p4 = _mm_packus_epi32(p4, p5);
+    p5 = _mm_packus_epi32(p6, p7);
+    p4 = _mm_packus_epi16(p4, p5);
+
+    _mm_storel_epi64((__m128i *)dst, p0);
+    p0 = _mm_srli_si128(p0, 8);
+    _mm_storel_epi64((__m128i *)(dst + 8), p0);
+
+    _mm_storel_epi64((__m128i *)(dst + 16), p4);
+    p4 = _mm_srli_si128(p4, 8);
+    _mm_storel_epi64((__m128i *)(dst + 24), p4);
+
+    dst += stride;
+    r += 1;
+  }
+}
+
+static void SavePrediction(int *pred, const __m128i *mean, int bs, uint8_t *dst,
+                           ptrdiff_t stride) {
+  switch (bs) {
+    case 4: SavePred4x4(pred, mean, dst, stride); break;
+    case 8: SavePred8x8(pred, mean, dst, stride); break;
+    case 16: SavePred16x16(pred, mean, dst, stride); break;
+    case 32: SavePred32x32(pred, mean, dst, stride); break;
+    default: assert(0);
+  }
+}
+
+typedef void (*ProducePixelsFunc)(__m128i *p, const __m128i *prm, int *pred,
+                                  const int predStride);
+
+static void ProduceFourPixels(__m128i *p, const __m128i *prm, int *pred,
+                              const int predStride) {
+  __m128i u0, u1, u2;
+  int c0 = _mm_extract_epi32(prm[1], 0);
+  int x = *(pred + predStride);
+  int sum;
+
+  u0 = _mm_mullo_epi32(p[0], prm[2]);
+  u1 = _mm_mullo_epi32(p[1], prm[0]);
+  u2 = _mm_mullo_epi32(p[2], prm[3]);
+
+  u0 = _mm_add_epi32(u0, u1);
+  u0 = _mm_add_epi32(u0, u2);
+
+  sum = _mm_extract_epi32(u0, 0);
+  sum += c0 * x;
+  x = ROUND_POWER_OF_TWO_SIGNED(sum, FILTER_INTRA_PREC_BITS);
+  *(pred + predStride + 1) = x;
+
+  sum = _mm_extract_epi32(u0, 1);
+  sum += c0 * x;
+  x = ROUND_POWER_OF_TWO_SIGNED(sum, FILTER_INTRA_PREC_BITS);
+  *(pred + predStride + 2) = x;
+
+  sum = _mm_extract_epi32(u0, 2);
+  sum += c0 * x;
+  x = ROUND_POWER_OF_TWO_SIGNED(sum, FILTER_INTRA_PREC_BITS);
+  *(pred + predStride + 3) = x;
+
+  sum = _mm_extract_epi32(u0, 3);
+  sum += c0 * x;
+  x = ROUND_POWER_OF_TWO_SIGNED(sum, FILTER_INTRA_PREC_BITS);
+  *(pred + predStride + 4) = x;
+}
+
+static void ProduceThreePixels(__m128i *p, const __m128i *prm, int *pred,
+                               const int predStride) {
+  __m128i u0, u1, u2;
+  int c0 = _mm_extract_epi32(prm[1], 0);
+  int x = *(pred + predStride);
+  int sum;
+
+  u0 = _mm_mullo_epi32(p[0], prm[2]);
+  u1 = _mm_mullo_epi32(p[1], prm[0]);
+  u2 = _mm_mullo_epi32(p[2], prm[3]);
+
+  u0 = _mm_add_epi32(u0, u1);
+  u0 = _mm_add_epi32(u0, u2);
+
+  sum = _mm_extract_epi32(u0, 0);
+  sum += c0 * x;
+  x = ROUND_POWER_OF_TWO_SIGNED(sum, FILTER_INTRA_PREC_BITS);
+  *(pred + predStride + 1) = x;
+
+  sum = _mm_extract_epi32(u0, 1);
+  sum += c0 * x;
+  x = ROUND_POWER_OF_TWO_SIGNED(sum, FILTER_INTRA_PREC_BITS);
+  *(pred + predStride + 2) = x;
+
+  sum = _mm_extract_epi32(u0, 2);
+  sum += c0 * x;
+  x = ROUND_POWER_OF_TWO_SIGNED(sum, FILTER_INTRA_PREC_BITS);
+  *(pred + predStride + 3) = x;
+}
+
+static void ProduceTwoPixels(__m128i *p, const __m128i *prm, int *pred,
+                             const int predStride) {
+  __m128i u0, u1, u2;
+  int c0 = _mm_extract_epi32(prm[1], 0);
+  int x = *(pred + predStride);
+  int sum;
+
+  u0 = _mm_mullo_epi32(p[0], prm[2]);
+  u1 = _mm_mullo_epi32(p[1], prm[0]);
+  u2 = _mm_mullo_epi32(p[2], prm[3]);
+
+  u0 = _mm_add_epi32(u0, u1);
+  u0 = _mm_add_epi32(u0, u2);
+
+  sum = _mm_extract_epi32(u0, 0);
+  sum += c0 * x;
+  x = ROUND_POWER_OF_TWO_SIGNED(sum, FILTER_INTRA_PREC_BITS);
+  *(pred + predStride + 1) = x;
+
+  sum = _mm_extract_epi32(u0, 1);
+  sum += c0 * x;
+  x = ROUND_POWER_OF_TWO_SIGNED(sum, FILTER_INTRA_PREC_BITS);
+  *(pred + predStride + 2) = x;
+}
+
+static void ProduceOnePixels(__m128i *p, const __m128i *prm, int *pred,
+                             const int predStride) {
+  __m128i u0, u1, u2;
+  int c0 = _mm_extract_epi32(prm[1], 0);
+  int x = *(pred + predStride);
+  int sum;
+
+  u0 = _mm_mullo_epi32(p[0], prm[2]);
+  u1 = _mm_mullo_epi32(p[1], prm[0]);
+  u2 = _mm_mullo_epi32(p[2], prm[3]);
+
+  u0 = _mm_add_epi32(u0, u1);
+  u0 = _mm_add_epi32(u0, u2);
+
+  sum = _mm_extract_epi32(u0, 0);
+  sum += c0 * x;
+  x = ROUND_POWER_OF_TWO_SIGNED(sum, FILTER_INTRA_PREC_BITS);
+  *(pred + predStride + 1) = x;
+}
+
+static ProducePixelsFunc prodPixelsFuncTab[4] = {
+  ProduceOnePixels, ProduceTwoPixels, ProduceThreePixels, ProduceFourPixels
+};
+
+static void ProducePixels(int *pred, const __m128i *prm, int remain) {
+  __m128i p[3];
+  const int predStride = (maxBlkSize << 1) + 1;
+  int index;
+
+  p[0] = _mm_loadu_si128((const __m128i *)pred);
+  p[1] = _mm_loadu_si128((const __m128i *)(pred + 1));
+  p[2] = _mm_loadu_si128((const __m128i *)(pred + 2));
+
+  if (remain <= 2) {
+    return;
+  }
+  if (remain > 5) {
+    index = 3;
+  } else {
+    index = remain - 3;
+  }
+  prodPixelsFuncTab[index](p, prm, pred, predStride);
+}
+
+// Note:
+//  At column index c, the remaining pixels are R = 2 * bs + 1 - r - c
+//  the number of pixels to produce is R - 2 = 2 * bs - r - c - 1
+static void GeneratePrediction(const uint8_t *above, const uint8_t *left,
+                               const int bs, const __m128i *prm, int meanValue,
+                               uint8_t *dst, ptrdiff_t stride) {
+  int pred[33][65];
+  int r, c, colBound;
+  int remainings;
+
+  for (r = 0; r < bs; ++r) {
+    pred[r + 1][0] = (int)left[r] - meanValue;
+  }
+
+  above -= 1;
+  for (c = 0; c < 2 * bs + 1; ++c) {
+    pred[0][c] = (int)above[c] - meanValue;
+  }
+
+  r = 0;
+  c = 0;
+  while (r < bs) {
+    colBound = (bs << 1) - r;
+    for (c = 0; c < colBound; c += 4) {
+      remainings = colBound - c + 1;
+      ProducePixels(&pred[r][c], prm, remainings);
+    }
+    r += 1;
+  }
+
+  SavePrediction(&pred[1][1], &prm[4], bs, dst, stride);
+}
+
+static void FilterPrediction(const uint8_t *above, const uint8_t *left, int bs,
+                             __m128i *prm, uint8_t *dst, ptrdiff_t stride) {
+  int meanValue = 0;
+  meanValue = CalcRefPixelsMeanValue(above, left, bs, &prm[4]);
+  GeneratePrediction(above, left, bs, prm, meanValue, dst, stride);
+}
+
+void av1_dc_filter_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, int bs,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, DC_PRED, &prm[0]);
+  FilterPrediction(above, left, bs, prm, dst, stride);
+}
+
+void av1_v_filter_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, int bs,
+                                   const uint8_t *above, const uint8_t *left) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, V_PRED, &prm[0]);
+  FilterPrediction(above, left, bs, prm, dst, stride);
+}
+
+void av1_h_filter_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, int bs,
+                                   const uint8_t *above, const uint8_t *left) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, H_PRED, &prm[0]);
+  FilterPrediction(above, left, bs, prm, dst, stride);
+}
+
+void av1_d45_filter_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, int bs,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D45_PRED, &prm[0]);
+  FilterPrediction(above, left, bs, prm, dst, stride);
+}
+
+void av1_d135_filter_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, int bs,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D135_PRED, &prm[0]);
+  FilterPrediction(above, left, bs, prm, dst, stride);
+}
+
+void av1_d117_filter_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, int bs,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D117_PRED, &prm[0]);
+  FilterPrediction(above, left, bs, prm, dst, stride);
+}
+
+void av1_d153_filter_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, int bs,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D153_PRED, &prm[0]);
+  FilterPrediction(above, left, bs, prm, dst, stride);
+}
+
+void av1_d207_filter_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, int bs,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D207_PRED, &prm[0]);
+  FilterPrediction(above, left, bs, prm, dst, stride);
+}
+
+void av1_d63_filter_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, int bs,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D63_PRED, &prm[0]);
+  FilterPrediction(above, left, bs, prm, dst, stride);
+}
+
+void av1_tm_filter_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride, int bs,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, TM_PRED, &prm[0]);
+  FilterPrediction(above, left, bs, prm, dst, stride);
+}
+
+// ============== High Bit Depth ==============
+#if CONFIG_HIGHBITDEPTH
+static INLINE int HighbdGetMeanValue4x4(const uint16_t *above,
+                                        const uint16_t *left, const int bd,
+                                        __m128i *params) {
+  const __m128i a = _mm_loadu_si128((const __m128i *)above);
+  const __m128i l = _mm_loadu_si128((const __m128i *)left);
+  const __m128i zero = _mm_setzero_si128();
+  __m128i sum_vector, u;
+  uint16_t sum_value;
+  (void)bd;
+
+  sum_vector = _mm_add_epi16(a, l);
+
+  sum_vector = _mm_hadd_epi16(sum_vector, zero);  // still has 2 values
+  u = _mm_srli_si128(sum_vector, 2);
+  sum_vector = _mm_add_epi16(sum_vector, u);
+
+  sum_value = _mm_extract_epi16(sum_vector, 0);
+  sum_value += 4;
+  sum_value >>= 3;
+  *params = _mm_set1_epi32(sum_value);
+  return sum_value;
+}
+
+static INLINE int HighbdGetMeanValue8x8(const uint16_t *above,
+                                        const uint16_t *left, const int bd,
+                                        __m128i *params) {
+  const __m128i a = _mm_loadu_si128((const __m128i *)above);
+  const __m128i l = _mm_loadu_si128((const __m128i *)left);
+  const __m128i zero = _mm_setzero_si128();
+  __m128i sum_vector, u;
+  uint16_t sum_value;
+  (void)bd;
+
+  sum_vector = _mm_add_epi16(a, l);
+
+  sum_vector = _mm_hadd_epi16(sum_vector, zero);  // still has 4 values
+  sum_vector = _mm_hadd_epi16(sum_vector, zero);  // still has 2 values
+
+  u = _mm_srli_si128(sum_vector, 2);
+  sum_vector = _mm_add_epi16(sum_vector, u);
+
+  sum_value = _mm_extract_epi16(sum_vector, 0);
+  sum_value += 8;
+  sum_value >>= 4;
+  *params = _mm_set1_epi32(sum_value);
+  return sum_value;
+}
+
+// Note:
+//  Process 16 pixels above and left, 10-bit depth
+//  Add to the last 8 pixels sum
+static INLINE void AddPixels10bit(const uint16_t *above, const uint16_t *left,
+                                  __m128i *sum) {
+  __m128i a = _mm_loadu_si128((const __m128i *)above);
+  __m128i l = _mm_loadu_si128((const __m128i *)left);
+  sum[0] = _mm_add_epi16(a, l);
+  a = _mm_loadu_si128((const __m128i *)(above + 8));
+  l = _mm_loadu_si128((const __m128i *)(left + 8));
+  sum[0] = _mm_add_epi16(sum[0], a);
+  sum[0] = _mm_add_epi16(sum[0], l);
+}
+
+// Note:
+//  Process 16 pixels above and left, 12-bit depth
+//  Add to the last 8 pixels sum
+static INLINE void AddPixels12bit(const uint16_t *above, const uint16_t *left,
+                                  __m128i *sum) {
+  __m128i a = _mm_loadu_si128((const __m128i *)above);
+  __m128i l = _mm_loadu_si128((const __m128i *)left);
+  const __m128i zero = _mm_setzero_si128();
+  __m128i v0, v1;
+
+  v0 = _mm_unpacklo_epi16(a, zero);
+  v1 = _mm_unpacklo_epi16(l, zero);
+  sum[0] = _mm_add_epi32(v0, v1);
+
+  v0 = _mm_unpackhi_epi16(a, zero);
+  v1 = _mm_unpackhi_epi16(l, zero);
+  sum[0] = _mm_add_epi32(sum[0], v0);
+  sum[0] = _mm_add_epi32(sum[0], v1);
+
+  a = _mm_loadu_si128((const __m128i *)(above + 8));
+  l = _mm_loadu_si128((const __m128i *)(left + 8));
+
+  v0 = _mm_unpacklo_epi16(a, zero);
+  v1 = _mm_unpacklo_epi16(l, zero);
+  sum[0] = _mm_add_epi32(sum[0], v0);
+  sum[0] = _mm_add_epi32(sum[0], v1);
+
+  v0 = _mm_unpackhi_epi16(a, zero);
+  v1 = _mm_unpackhi_epi16(l, zero);
+  sum[0] = _mm_add_epi32(sum[0], v0);
+  sum[0] = _mm_add_epi32(sum[0], v1);
+}
+
+static INLINE int HighbdGetMeanValue16x16(const uint16_t *above,
+                                          const uint16_t *left, const int bd,
+                                          __m128i *params) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i sum_vector, u;
+  uint32_t sum_value = 0;
+
+  if (10 == bd) {
+    AddPixels10bit(above, left, &sum_vector);
+    sum_vector = _mm_hadd_epi16(sum_vector, zero);  // still has 4 values
+    sum_vector = _mm_hadd_epi16(sum_vector, zero);  // still has 2 values
+
+    u = _mm_srli_si128(sum_vector, 2);
+    sum_vector = _mm_add_epi16(sum_vector, u);
+    sum_value = _mm_extract_epi16(sum_vector, 0);
+  } else if (12 == bd) {
+    AddPixels12bit(above, left, &sum_vector);
+
+    sum_vector = _mm_hadd_epi32(sum_vector, zero);
+    u = _mm_srli_si128(sum_vector, 4);
+    sum_vector = _mm_add_epi32(u, sum_vector);
+    sum_value = _mm_extract_epi32(sum_vector, 0);
+  }
+
+  sum_value += 16;
+  sum_value >>= 5;
+  *params = _mm_set1_epi32(sum_value);
+  return sum_value;
+}
+
+static INLINE int HighbdGetMeanValue32x32(const uint16_t *above,
+                                          const uint16_t *left, const int bd,
+                                          __m128i *params) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i sum_vector[2], u;
+  uint32_t sum_value = 0;
+
+  if (10 == bd) {
+    AddPixels10bit(above, left, &sum_vector[0]);
+    AddPixels10bit(above + 16, left + 16, &sum_vector[1]);
+
+    sum_vector[0] = _mm_add_epi16(sum_vector[0], sum_vector[1]);
+    sum_vector[0] = _mm_hadd_epi16(sum_vector[0], zero);  // still has 4 values
+    sum_vector[0] = _mm_hadd_epi16(sum_vector[0], zero);  // still has 2 values
+
+    u = _mm_srli_si128(sum_vector[0], 2);
+    sum_vector[0] = _mm_add_epi16(sum_vector[0], u);
+    sum_value = _mm_extract_epi16(sum_vector[0], 0);
+  } else if (12 == bd) {
+    AddPixels12bit(above, left, &sum_vector[0]);
+    AddPixels12bit(above + 16, left + 16, &sum_vector[1]);
+
+    sum_vector[0] = _mm_add_epi32(sum_vector[0], sum_vector[1]);
+    sum_vector[0] = _mm_hadd_epi32(sum_vector[0], zero);
+    u = _mm_srli_si128(sum_vector[0], 4);
+    sum_vector[0] = _mm_add_epi32(u, sum_vector[0]);
+    sum_value = _mm_extract_epi32(sum_vector[0], 0);
+  }
+
+  sum_value += 32;
+  sum_value >>= 6;
+  *params = _mm_set1_epi32(sum_value);
+  return sum_value;
+}
+
+// Note:
+//  params[4] : mean value, 4 int32_t repetition
+//
+static INLINE int HighbdCalcRefPixelsMeanValue(const uint16_t *above,
+                                               const uint16_t *left, int bs,
+                                               const int bd, __m128i *params) {
+  int meanValue = 0;
+  switch (bs) {
+    case 4: meanValue = HighbdGetMeanValue4x4(above, left, bd, params); break;
+    case 8: meanValue = HighbdGetMeanValue8x8(above, left, bd, params); break;
+    case 16:
+      meanValue = HighbdGetMeanValue16x16(above, left, bd, params);
+      break;
+    case 32:
+      meanValue = HighbdGetMeanValue32x32(above, left, bd, params);
+      break;
+    default: assert(0);
+  }
+  return meanValue;
+}
+
+// Note:
+//  At column index c, the remaining pixels are R = 2 * bs + 1 - r - c
+//  the number of pixels to produce is R - 2 = 2 * bs - r - c - 1
+static void HighbdGeneratePrediction(const uint16_t *above,
+                                     const uint16_t *left, const int bs,
+                                     const int bd, const __m128i *prm,
+                                     int meanValue, uint16_t *dst,
+                                     ptrdiff_t stride) {
+  int pred[33][65];
+  int r, c, colBound;
+  int remainings;
+  int ipred;
+
+  for (r = 0; r < bs; ++r) {
+    pred[r + 1][0] = (int)left[r] - meanValue;
+  }
+
+  above -= 1;
+  for (c = 0; c < 2 * bs + 1; ++c) {
+    pred[0][c] = (int)above[c] - meanValue;
+  }
+
+  r = 0;
+  c = 0;
+  while (r < bs) {
+    colBound = (bs << 1) - r;
+    for (c = 0; c < colBound; c += 4) {
+      remainings = colBound - c + 1;
+      ProducePixels(&pred[r][c], prm, remainings);
+    }
+    r += 1;
+  }
+
+  for (r = 0; r < bs; ++r) {
+    for (c = 0; c < bs; ++c) {
+      ipred = pred[r + 1][c + 1] + meanValue;
+      dst[c] = clip_pixel_highbd(ipred, bd);
+    }
+    dst += stride;
+  }
+}
+
+static void HighbdFilterPrediction(const uint16_t *above, const uint16_t *left,
+                                   int bs, const int bd, __m128i *prm,
+                                   uint16_t *dst, ptrdiff_t stride) {
+  int meanValue = 0;
+  meanValue = HighbdCalcRefPixelsMeanValue(above, left, bs, bd, &prm[4]);
+  HighbdGeneratePrediction(above, left, bs, bd, prm, meanValue, dst, stride);
+}
+
+void av1_highbd_dc_filter_predictor_sse4_1(uint16_t *dst, ptrdiff_t stride,
+                                           int bs, const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, DC_PRED, &prm[0]);
+  HighbdFilterPrediction(above, left, bs, bd, prm, dst, stride);
+}
+
+void av1_highbd_v_filter_predictor_sse4_1(uint16_t *dst, ptrdiff_t stride,
+                                          int bs, const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, V_PRED, &prm[0]);
+  HighbdFilterPrediction(above, left, bs, bd, prm, dst, stride);
+}
+
+void av1_highbd_h_filter_predictor_sse4_1(uint16_t *dst, ptrdiff_t stride,
+                                          int bs, const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, H_PRED, &prm[0]);
+  HighbdFilterPrediction(above, left, bs, bd, prm, dst, stride);
+}
+
+void av1_highbd_d45_filter_predictor_sse4_1(uint16_t *dst, ptrdiff_t stride,
+                                            int bs, const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D45_PRED, &prm[0]);
+  HighbdFilterPrediction(above, left, bs, bd, prm, dst, stride);
+}
+
+void av1_highbd_d135_filter_predictor_sse4_1(uint16_t *dst, ptrdiff_t stride,
+                                             int bs, const uint16_t *above,
+                                             const uint16_t *left, int bd) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D135_PRED, &prm[0]);
+  HighbdFilterPrediction(above, left, bs, bd, prm, dst, stride);
+}
+
+void av1_highbd_d117_filter_predictor_sse4_1(uint16_t *dst, ptrdiff_t stride,
+                                             int bs, const uint16_t *above,
+                                             const uint16_t *left, int bd) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D117_PRED, &prm[0]);
+  HighbdFilterPrediction(above, left, bs, bd, prm, dst, stride);
+}
+
+void av1_highbd_d153_filter_predictor_sse4_1(uint16_t *dst, ptrdiff_t stride,
+                                             int bs, const uint16_t *above,
+                                             const uint16_t *left, int bd) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D153_PRED, &prm[0]);
+  HighbdFilterPrediction(above, left, bs, bd, prm, dst, stride);
+}
+
+void av1_highbd_d207_filter_predictor_sse4_1(uint16_t *dst, ptrdiff_t stride,
+                                             int bs, const uint16_t *above,
+                                             const uint16_t *left, int bd) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D207_PRED, &prm[0]);
+  HighbdFilterPrediction(above, left, bs, bd, prm, dst, stride);
+}
+
+void av1_highbd_d63_filter_predictor_sse4_1(uint16_t *dst, ptrdiff_t stride,
+                                            int bs, const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, D63_PRED, &prm[0]);
+  HighbdFilterPrediction(above, left, bs, bd, prm, dst, stride);
+}
+
+void av1_highbd_tm_filter_predictor_sse4_1(uint16_t *dst, ptrdiff_t stride,
+                                           int bs, const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  __m128i prm[5];
+  GetIntraFilterParams(bs, TM_PRED, &prm[0]);
+  HighbdFilterPrediction(above, left, bs, bd, prm, dst, stride);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#endif  // USE_3TAP_INTRA_FILTER
diff --git a/third_party/aom/av1/common/x86/highbd_inv_txfm_avx2.c b/third_party/aom/av1/common/x86/highbd_inv_txfm_avx2.c
new file mode 100644
index 0000000000..d10f1ccc21
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_inv_txfm_avx2.c
@@ -0,0 +1,557 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <assert.h>
+#include <immintrin.h>
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "av1/common/av1_inv_txfm2d_cfg.h"
+
+// Note:
+//  Total 32x4 registers to represent 32x32 block coefficients.
+//  For high bit depth, each coefficient is 4-byte.
+//  Each __m256i register holds 8 coefficients.
+//  So each "row" we needs 4 register. Totally 32 rows
+//  Register layout:
+//   v0,   v1,   v2,   v3,
+//   v4,   v5,   v6,   v7,
+//   ... ...
+//   v124, v125, v126, v127
+
+static void transpose_32x32_8x8(const __m256i *in, __m256i *out) {
+  __m256i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m256i x0, x1;
+
+  u0 = _mm256_unpacklo_epi32(in[0], in[4]);
+  u1 = _mm256_unpackhi_epi32(in[0], in[4]);
+
+  u2 = _mm256_unpacklo_epi32(in[8], in[12]);
+  u3 = _mm256_unpackhi_epi32(in[8], in[12]);
+
+  u4 = _mm256_unpacklo_epi32(in[16], in[20]);
+  u5 = _mm256_unpackhi_epi32(in[16], in[20]);
+
+  u6 = _mm256_unpacklo_epi32(in[24], in[28]);
+  u7 = _mm256_unpackhi_epi32(in[24], in[28]);
+
+  x0 = _mm256_unpacklo_epi64(u0, u2);
+  x1 = _mm256_unpacklo_epi64(u4, u6);
+  out[0] = _mm256_permute2f128_si256(x0, x1, 0x20);
+  out[16] = _mm256_permute2f128_si256(x0, x1, 0x31);
+
+  x0 = _mm256_unpackhi_epi64(u0, u2);
+  x1 = _mm256_unpackhi_epi64(u4, u6);
+  out[4] = _mm256_permute2f128_si256(x0, x1, 0x20);
+  out[20] = _mm256_permute2f128_si256(x0, x1, 0x31);
+
+  x0 = _mm256_unpacklo_epi64(u1, u3);
+  x1 = _mm256_unpacklo_epi64(u5, u7);
+  out[8] = _mm256_permute2f128_si256(x0, x1, 0x20);
+  out[24] = _mm256_permute2f128_si256(x0, x1, 0x31);
+
+  x0 = _mm256_unpackhi_epi64(u1, u3);
+  x1 = _mm256_unpackhi_epi64(u5, u7);
+  out[12] = _mm256_permute2f128_si256(x0, x1, 0x20);
+  out[28] = _mm256_permute2f128_si256(x0, x1, 0x31);
+}
+
+static void transpose_32x32_16x16(const __m256i *in, __m256i *out) {
+  transpose_32x32_8x8(&in[0], &out[0]);
+  transpose_32x32_8x8(&in[1], &out[32]);
+  transpose_32x32_8x8(&in[32], &out[1]);
+  transpose_32x32_8x8(&in[33], &out[33]);
+}
+
+static void transpose_32x32(const __m256i *in, __m256i *out) {
+  transpose_32x32_16x16(&in[0], &out[0]);
+  transpose_32x32_16x16(&in[2], &out[64]);
+  transpose_32x32_16x16(&in[64], &out[2]);
+  transpose_32x32_16x16(&in[66], &out[66]);
+}
+
+static void load_buffer_32x32(const int32_t *coeff, __m256i *in) {
+  int i;
+  for (i = 0; i < 128; ++i) {
+    in[i] = _mm256_loadu_si256((const __m256i *)coeff);
+    coeff += 8;
+  }
+}
+
+static void round_shift_32x32(__m256i *in, int shift) {
+  __m256i rnding = _mm256_set1_epi32(1 << (shift - 1));
+  int i = 0;
+
+  while (i < 128) {
+    in[i] = _mm256_add_epi32(in[i], rnding);
+    in[i] = _mm256_srai_epi32(in[i], shift);
+    i++;
+  }
+}
+
+static __m256i highbd_clamp_epi32(__m256i x, int bd) {
+  const __m256i zero = _mm256_setzero_si256();
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i max = _mm256_sub_epi16(_mm256_slli_epi16(one, bd), one);
+  __m256i clamped, mask;
+
+  mask = _mm256_cmpgt_epi16(x, max);
+  clamped = _mm256_andnot_si256(mask, x);
+  mask = _mm256_and_si256(mask, max);
+  clamped = _mm256_or_si256(mask, clamped);
+  mask = _mm256_cmpgt_epi16(clamped, zero);
+  clamped = _mm256_and_si256(clamped, mask);
+
+  return clamped;
+}
+
+static void write_buffer_32x32(__m256i *in, uint16_t *output, int stride,
+                               int fliplr, int flipud, int shift, int bd) {
+  __m256i u0, u1, x0, x1, x2, x3, v0, v1, v2, v3;
+  const __m256i zero = _mm256_setzero_si256();
+  int i = 0;
+  (void)fliplr;
+  (void)flipud;
+
+  round_shift_32x32(in, shift);
+
+  while (i < 128) {
+    u0 = _mm256_loadu_si256((const __m256i *)output);
+    u1 = _mm256_loadu_si256((const __m256i *)(output + 16));
+
+    x0 = _mm256_unpacklo_epi16(u0, zero);
+    x1 = _mm256_unpackhi_epi16(u0, zero);
+    x2 = _mm256_unpacklo_epi16(u1, zero);
+    x3 = _mm256_unpackhi_epi16(u1, zero);
+
+    v0 = _mm256_permute2f128_si256(in[i], in[i + 1], 0x20);
+    v1 = _mm256_permute2f128_si256(in[i], in[i + 1], 0x31);
+    v2 = _mm256_permute2f128_si256(in[i + 2], in[i + 3], 0x20);
+    v3 = _mm256_permute2f128_si256(in[i + 2], in[i + 3], 0x31);
+
+    v0 = _mm256_add_epi32(v0, x0);
+    v1 = _mm256_add_epi32(v1, x1);
+    v2 = _mm256_add_epi32(v2, x2);
+    v3 = _mm256_add_epi32(v3, x3);
+
+    v0 = _mm256_packus_epi32(v0, v1);
+    v2 = _mm256_packus_epi32(v2, v3);
+
+    v0 = highbd_clamp_epi32(v0, bd);
+    v2 = highbd_clamp_epi32(v2, bd);
+
+    _mm256_storeu_si256((__m256i *)output, v0);
+    _mm256_storeu_si256((__m256i *)(output + 16), v2);
+    output += stride;
+    i += 4;
+  }
+}
+
+static INLINE __m256i half_btf_avx2(__m256i w0, __m256i n0, __m256i w1,
+                                    __m256i n1, __m256i rounding, int bit) {
+  __m256i x, y;
+
+  x = _mm256_mullo_epi32(w0, n0);
+  y = _mm256_mullo_epi32(w1, n1);
+  x = _mm256_add_epi32(x, y);
+  x = _mm256_add_epi32(x, rounding);
+  x = _mm256_srai_epi32(x, bit);
+  return x;
+}
+
+static void idct32_avx2(__m256i *in, __m256i *out, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m256i cospi62 = _mm256_set1_epi32(cospi[62]);
+  const __m256i cospi30 = _mm256_set1_epi32(cospi[30]);
+  const __m256i cospi46 = _mm256_set1_epi32(cospi[46]);
+  const __m256i cospi14 = _mm256_set1_epi32(cospi[14]);
+  const __m256i cospi54 = _mm256_set1_epi32(cospi[54]);
+  const __m256i cospi22 = _mm256_set1_epi32(cospi[22]);
+  const __m256i cospi38 = _mm256_set1_epi32(cospi[38]);
+  const __m256i cospi6 = _mm256_set1_epi32(cospi[6]);
+  const __m256i cospi58 = _mm256_set1_epi32(cospi[58]);
+  const __m256i cospi26 = _mm256_set1_epi32(cospi[26]);
+  const __m256i cospi42 = _mm256_set1_epi32(cospi[42]);
+  const __m256i cospi10 = _mm256_set1_epi32(cospi[10]);
+  const __m256i cospi50 = _mm256_set1_epi32(cospi[50]);
+  const __m256i cospi18 = _mm256_set1_epi32(cospi[18]);
+  const __m256i cospi34 = _mm256_set1_epi32(cospi[34]);
+  const __m256i cospi2 = _mm256_set1_epi32(cospi[2]);
+  const __m256i cospim58 = _mm256_set1_epi32(-cospi[58]);
+  const __m256i cospim26 = _mm256_set1_epi32(-cospi[26]);
+  const __m256i cospim42 = _mm256_set1_epi32(-cospi[42]);
+  const __m256i cospim10 = _mm256_set1_epi32(-cospi[10]);
+  const __m256i cospim50 = _mm256_set1_epi32(-cospi[50]);
+  const __m256i cospim18 = _mm256_set1_epi32(-cospi[18]);
+  const __m256i cospim34 = _mm256_set1_epi32(-cospi[34]);
+  const __m256i cospim2 = _mm256_set1_epi32(-cospi[2]);
+  const __m256i cospi60 = _mm256_set1_epi32(cospi[60]);
+  const __m256i cospi28 = _mm256_set1_epi32(cospi[28]);
+  const __m256i cospi44 = _mm256_set1_epi32(cospi[44]);
+  const __m256i cospi12 = _mm256_set1_epi32(cospi[12]);
+  const __m256i cospi52 = _mm256_set1_epi32(cospi[52]);
+  const __m256i cospi20 = _mm256_set1_epi32(cospi[20]);
+  const __m256i cospi36 = _mm256_set1_epi32(cospi[36]);
+  const __m256i cospi4 = _mm256_set1_epi32(cospi[4]);
+  const __m256i cospim52 = _mm256_set1_epi32(-cospi[52]);
+  const __m256i cospim20 = _mm256_set1_epi32(-cospi[20]);
+  const __m256i cospim36 = _mm256_set1_epi32(-cospi[36]);
+  const __m256i cospim4 = _mm256_set1_epi32(-cospi[4]);
+  const __m256i cospi56 = _mm256_set1_epi32(cospi[56]);
+  const __m256i cospi24 = _mm256_set1_epi32(cospi[24]);
+  const __m256i cospi40 = _mm256_set1_epi32(cospi[40]);
+  const __m256i cospi8 = _mm256_set1_epi32(cospi[8]);
+  const __m256i cospim40 = _mm256_set1_epi32(-cospi[40]);
+  const __m256i cospim8 = _mm256_set1_epi32(-cospi[8]);
+  const __m256i cospim56 = _mm256_set1_epi32(-cospi[56]);
+  const __m256i cospim24 = _mm256_set1_epi32(-cospi[24]);
+  const __m256i cospi32 = _mm256_set1_epi32(cospi[32]);
+  const __m256i cospim32 = _mm256_set1_epi32(-cospi[32]);
+  const __m256i cospi48 = _mm256_set1_epi32(cospi[48]);
+  const __m256i cospim48 = _mm256_set1_epi32(-cospi[48]);
+  const __m256i cospi16 = _mm256_set1_epi32(cospi[16]);
+  const __m256i cospim16 = _mm256_set1_epi32(-cospi[16]);
+  const __m256i rounding = _mm256_set1_epi32(1 << (bit - 1));
+  __m256i bf1[32], bf0[32];
+  int col;
+
+  for (col = 0; col < 4; ++col) {
+    // stage 0
+    // stage 1
+    bf1[0] = in[0 * 4 + col];
+    bf1[1] = in[16 * 4 + col];
+    bf1[2] = in[8 * 4 + col];
+    bf1[3] = in[24 * 4 + col];
+    bf1[4] = in[4 * 4 + col];
+    bf1[5] = in[20 * 4 + col];
+    bf1[6] = in[12 * 4 + col];
+    bf1[7] = in[28 * 4 + col];
+    bf1[8] = in[2 * 4 + col];
+    bf1[9] = in[18 * 4 + col];
+    bf1[10] = in[10 * 4 + col];
+    bf1[11] = in[26 * 4 + col];
+    bf1[12] = in[6 * 4 + col];
+    bf1[13] = in[22 * 4 + col];
+    bf1[14] = in[14 * 4 + col];
+    bf1[15] = in[30 * 4 + col];
+    bf1[16] = in[1 * 4 + col];
+    bf1[17] = in[17 * 4 + col];
+    bf1[18] = in[9 * 4 + col];
+    bf1[19] = in[25 * 4 + col];
+    bf1[20] = in[5 * 4 + col];
+    bf1[21] = in[21 * 4 + col];
+    bf1[22] = in[13 * 4 + col];
+    bf1[23] = in[29 * 4 + col];
+    bf1[24] = in[3 * 4 + col];
+    bf1[25] = in[19 * 4 + col];
+    bf1[26] = in[11 * 4 + col];
+    bf1[27] = in[27 * 4 + col];
+    bf1[28] = in[7 * 4 + col];
+    bf1[29] = in[23 * 4 + col];
+    bf1[30] = in[15 * 4 + col];
+    bf1[31] = in[31 * 4 + col];
+
+    // stage 2
+    bf0[0] = bf1[0];
+    bf0[1] = bf1[1];
+    bf0[2] = bf1[2];
+    bf0[3] = bf1[3];
+    bf0[4] = bf1[4];
+    bf0[5] = bf1[5];
+    bf0[6] = bf1[6];
+    bf0[7] = bf1[7];
+    bf0[8] = bf1[8];
+    bf0[9] = bf1[9];
+    bf0[10] = bf1[10];
+    bf0[11] = bf1[11];
+    bf0[12] = bf1[12];
+    bf0[13] = bf1[13];
+    bf0[14] = bf1[14];
+    bf0[15] = bf1[15];
+    bf0[16] = half_btf_avx2(cospi62, bf1[16], cospim2, bf1[31], rounding, bit);
+    bf0[17] = half_btf_avx2(cospi30, bf1[17], cospim34, bf1[30], rounding, bit);
+    bf0[18] = half_btf_avx2(cospi46, bf1[18], cospim18, bf1[29], rounding, bit);
+    bf0[19] = half_btf_avx2(cospi14, bf1[19], cospim50, bf1[28], rounding, bit);
+    bf0[20] = half_btf_avx2(cospi54, bf1[20], cospim10, bf1[27], rounding, bit);
+    bf0[21] = half_btf_avx2(cospi22, bf1[21], cospim42, bf1[26], rounding, bit);
+    bf0[22] = half_btf_avx2(cospi38, bf1[22], cospim26, bf1[25], rounding, bit);
+    bf0[23] = half_btf_avx2(cospi6, bf1[23], cospim58, bf1[24], rounding, bit);
+    bf0[24] = half_btf_avx2(cospi58, bf1[23], cospi6, bf1[24], rounding, bit);
+    bf0[25] = half_btf_avx2(cospi26, bf1[22], cospi38, bf1[25], rounding, bit);
+    bf0[26] = half_btf_avx2(cospi42, bf1[21], cospi22, bf1[26], rounding, bit);
+    bf0[27] = half_btf_avx2(cospi10, bf1[20], cospi54, bf1[27], rounding, bit);
+    bf0[28] = half_btf_avx2(cospi50, bf1[19], cospi14, bf1[28], rounding, bit);
+    bf0[29] = half_btf_avx2(cospi18, bf1[18], cospi46, bf1[29], rounding, bit);
+    bf0[30] = half_btf_avx2(cospi34, bf1[17], cospi30, bf1[30], rounding, bit);
+    bf0[31] = half_btf_avx2(cospi2, bf1[16], cospi62, bf1[31], rounding, bit);
+
+    // stage 3
+    bf1[0] = bf0[0];
+    bf1[1] = bf0[1];
+    bf1[2] = bf0[2];
+    bf1[3] = bf0[3];
+    bf1[4] = bf0[4];
+    bf1[5] = bf0[5];
+    bf1[6] = bf0[6];
+    bf1[7] = bf0[7];
+    bf1[8] = half_btf_avx2(cospi60, bf0[8], cospim4, bf0[15], rounding, bit);
+    bf1[9] = half_btf_avx2(cospi28, bf0[9], cospim36, bf0[14], rounding, bit);
+    bf1[10] = half_btf_avx2(cospi44, bf0[10], cospim20, bf0[13], rounding, bit);
+    bf1[11] = half_btf_avx2(cospi12, bf0[11], cospim52, bf0[12], rounding, bit);
+    bf1[12] = half_btf_avx2(cospi52, bf0[11], cospi12, bf0[12], rounding, bit);
+    bf1[13] = half_btf_avx2(cospi20, bf0[10], cospi44, bf0[13], rounding, bit);
+    bf1[14] = half_btf_avx2(cospi36, bf0[9], cospi28, bf0[14], rounding, bit);
+    bf1[15] = half_btf_avx2(cospi4, bf0[8], cospi60, bf0[15], rounding, bit);
+    bf1[16] = _mm256_add_epi32(bf0[16], bf0[17]);
+    bf1[17] = _mm256_sub_epi32(bf0[16], bf0[17]);
+    bf1[18] = _mm256_sub_epi32(bf0[19], bf0[18]);
+    bf1[19] = _mm256_add_epi32(bf0[18], bf0[19]);
+    bf1[20] = _mm256_add_epi32(bf0[20], bf0[21]);
+    bf1[21] = _mm256_sub_epi32(bf0[20], bf0[21]);
+    bf1[22] = _mm256_sub_epi32(bf0[23], bf0[22]);
+    bf1[23] = _mm256_add_epi32(bf0[22], bf0[23]);
+    bf1[24] = _mm256_add_epi32(bf0[24], bf0[25]);
+    bf1[25] = _mm256_sub_epi32(bf0[24], bf0[25]);
+    bf1[26] = _mm256_sub_epi32(bf0[27], bf0[26]);
+    bf1[27] = _mm256_add_epi32(bf0[26], bf0[27]);
+    bf1[28] = _mm256_add_epi32(bf0[28], bf0[29]);
+    bf1[29] = _mm256_sub_epi32(bf0[28], bf0[29]);
+    bf1[30] = _mm256_sub_epi32(bf0[31], bf0[30]);
+    bf1[31] = _mm256_add_epi32(bf0[30], bf0[31]);
+
+    // stage 4
+    bf0[0] = bf1[0];
+    bf0[1] = bf1[1];
+    bf0[2] = bf1[2];
+    bf0[3] = bf1[3];
+    bf0[4] = half_btf_avx2(cospi56, bf1[4], cospim8, bf1[7], rounding, bit);
+    bf0[5] = half_btf_avx2(cospi24, bf1[5], cospim40, bf1[6], rounding, bit);
+    bf0[6] = half_btf_avx2(cospi40, bf1[5], cospi24, bf1[6], rounding, bit);
+    bf0[7] = half_btf_avx2(cospi8, bf1[4], cospi56, bf1[7], rounding, bit);
+    bf0[8] = _mm256_add_epi32(bf1[8], bf1[9]);
+    bf0[9] = _mm256_sub_epi32(bf1[8], bf1[9]);
+    bf0[10] = _mm256_sub_epi32(bf1[11], bf1[10]);
+    bf0[11] = _mm256_add_epi32(bf1[10], bf1[11]);
+    bf0[12] = _mm256_add_epi32(bf1[12], bf1[13]);
+    bf0[13] = _mm256_sub_epi32(bf1[12], bf1[13]);
+    bf0[14] = _mm256_sub_epi32(bf1[15], bf1[14]);
+    bf0[15] = _mm256_add_epi32(bf1[14], bf1[15]);
+    bf0[16] = bf1[16];
+    bf0[17] = half_btf_avx2(cospim8, bf1[17], cospi56, bf1[30], rounding, bit);
+    bf0[18] = half_btf_avx2(cospim56, bf1[18], cospim8, bf1[29], rounding, bit);
+    bf0[19] = bf1[19];
+    bf0[20] = bf1[20];
+    bf0[21] = half_btf_avx2(cospim40, bf1[21], cospi24, bf1[26], rounding, bit);
+    bf0[22] =
+        half_btf_avx2(cospim24, bf1[22], cospim40, bf1[25], rounding, bit);
+    bf0[23] = bf1[23];
+    bf0[24] = bf1[24];
+    bf0[25] = half_btf_avx2(cospim40, bf1[22], cospi24, bf1[25], rounding, bit);
+    bf0[26] = half_btf_avx2(cospi24, bf1[21], cospi40, bf1[26], rounding, bit);
+    bf0[27] = bf1[27];
+    bf0[28] = bf1[28];
+    bf0[29] = half_btf_avx2(cospim8, bf1[18], cospi56, bf1[29], rounding, bit);
+    bf0[30] = half_btf_avx2(cospi56, bf1[17], cospi8, bf1[30], rounding, bit);
+    bf0[31] = bf1[31];
+
+    // stage 5
+    bf1[0] = half_btf_avx2(cospi32, bf0[0], cospi32, bf0[1], rounding, bit);
+    bf1[1] = half_btf_avx2(cospi32, bf0[0], cospim32, bf0[1], rounding, bit);
+    bf1[2] = half_btf_avx2(cospi48, bf0[2], cospim16, bf0[3], rounding, bit);
+    bf1[3] = half_btf_avx2(cospi16, bf0[2], cospi48, bf0[3], rounding, bit);
+    bf1[4] = _mm256_add_epi32(bf0[4], bf0[5]);
+    bf1[5] = _mm256_sub_epi32(bf0[4], bf0[5]);
+    bf1[6] = _mm256_sub_epi32(bf0[7], bf0[6]);
+    bf1[7] = _mm256_add_epi32(bf0[6], bf0[7]);
+    bf1[8] = bf0[8];
+    bf1[9] = half_btf_avx2(cospim16, bf0[9], cospi48, bf0[14], rounding, bit);
+    bf1[10] =
+        half_btf_avx2(cospim48, bf0[10], cospim16, bf0[13], rounding, bit);
+    bf1[11] = bf0[11];
+    bf1[12] = bf0[12];
+    bf1[13] = half_btf_avx2(cospim16, bf0[10], cospi48, bf0[13], rounding, bit);
+    bf1[14] = half_btf_avx2(cospi48, bf0[9], cospi16, bf0[14], rounding, bit);
+    bf1[15] = bf0[15];
+    bf1[16] = _mm256_add_epi32(bf0[16], bf0[19]);
+    bf1[17] = _mm256_add_epi32(bf0[17], bf0[18]);
+    bf1[18] = _mm256_sub_epi32(bf0[17], bf0[18]);
+    bf1[19] = _mm256_sub_epi32(bf0[16], bf0[19]);
+    bf1[20] = _mm256_sub_epi32(bf0[23], bf0[20]);
+    bf1[21] = _mm256_sub_epi32(bf0[22], bf0[21]);
+    bf1[22] = _mm256_add_epi32(bf0[21], bf0[22]);
+    bf1[23] = _mm256_add_epi32(bf0[20], bf0[23]);
+    bf1[24] = _mm256_add_epi32(bf0[24], bf0[27]);
+    bf1[25] = _mm256_add_epi32(bf0[25], bf0[26]);
+    bf1[26] = _mm256_sub_epi32(bf0[25], bf0[26]);
+    bf1[27] = _mm256_sub_epi32(bf0[24], bf0[27]);
+    bf1[28] = _mm256_sub_epi32(bf0[31], bf0[28]);
+    bf1[29] = _mm256_sub_epi32(bf0[30], bf0[29]);
+    bf1[30] = _mm256_add_epi32(bf0[29], bf0[30]);
+    bf1[31] = _mm256_add_epi32(bf0[28], bf0[31]);
+
+    // stage 6
+    bf0[0] = _mm256_add_epi32(bf1[0], bf1[3]);
+    bf0[1] = _mm256_add_epi32(bf1[1], bf1[2]);
+    bf0[2] = _mm256_sub_epi32(bf1[1], bf1[2]);
+    bf0[3] = _mm256_sub_epi32(bf1[0], bf1[3]);
+    bf0[4] = bf1[4];
+    bf0[5] = half_btf_avx2(cospim32, bf1[5], cospi32, bf1[6], rounding, bit);
+    bf0[6] = half_btf_avx2(cospi32, bf1[5], cospi32, bf1[6], rounding, bit);
+    bf0[7] = bf1[7];
+    bf0[8] = _mm256_add_epi32(bf1[8], bf1[11]);
+    bf0[9] = _mm256_add_epi32(bf1[9], bf1[10]);
+    bf0[10] = _mm256_sub_epi32(bf1[9], bf1[10]);
+    bf0[11] = _mm256_sub_epi32(bf1[8], bf1[11]);
+    bf0[12] = _mm256_sub_epi32(bf1[15], bf1[12]);
+    bf0[13] = _mm256_sub_epi32(bf1[14], bf1[13]);
+    bf0[14] = _mm256_add_epi32(bf1[13], bf1[14]);
+    bf0[15] = _mm256_add_epi32(bf1[12], bf1[15]);
+    bf0[16] = bf1[16];
+    bf0[17] = bf1[17];
+    bf0[18] = half_btf_avx2(cospim16, bf1[18], cospi48, bf1[29], rounding, bit);
+    bf0[19] = half_btf_avx2(cospim16, bf1[19], cospi48, bf1[28], rounding, bit);
+    bf0[20] =
+        half_btf_avx2(cospim48, bf1[20], cospim16, bf1[27], rounding, bit);
+    bf0[21] =
+        half_btf_avx2(cospim48, bf1[21], cospim16, bf1[26], rounding, bit);
+    bf0[22] = bf1[22];
+    bf0[23] = bf1[23];
+    bf0[24] = bf1[24];
+    bf0[25] = bf1[25];
+    bf0[26] = half_btf_avx2(cospim16, bf1[21], cospi48, bf1[26], rounding, bit);
+    bf0[27] = half_btf_avx2(cospim16, bf1[20], cospi48, bf1[27], rounding, bit);
+    bf0[28] = half_btf_avx2(cospi48, bf1[19], cospi16, bf1[28], rounding, bit);
+    bf0[29] = half_btf_avx2(cospi48, bf1[18], cospi16, bf1[29], rounding, bit);
+    bf0[30] = bf1[30];
+    bf0[31] = bf1[31];
+
+    // stage 7
+    bf1[0] = _mm256_add_epi32(bf0[0], bf0[7]);
+    bf1[1] = _mm256_add_epi32(bf0[1], bf0[6]);
+    bf1[2] = _mm256_add_epi32(bf0[2], bf0[5]);
+    bf1[3] = _mm256_add_epi32(bf0[3], bf0[4]);
+    bf1[4] = _mm256_sub_epi32(bf0[3], bf0[4]);
+    bf1[5] = _mm256_sub_epi32(bf0[2], bf0[5]);
+    bf1[6] = _mm256_sub_epi32(bf0[1], bf0[6]);
+    bf1[7] = _mm256_sub_epi32(bf0[0], bf0[7]);
+    bf1[8] = bf0[8];
+    bf1[9] = bf0[9];
+    bf1[10] = half_btf_avx2(cospim32, bf0[10], cospi32, bf0[13], rounding, bit);
+    bf1[11] = half_btf_avx2(cospim32, bf0[11], cospi32, bf0[12], rounding, bit);
+    bf1[12] = half_btf_avx2(cospi32, bf0[11], cospi32, bf0[12], rounding, bit);
+    bf1[13] = half_btf_avx2(cospi32, bf0[10], cospi32, bf0[13], rounding, bit);
+    bf1[14] = bf0[14];
+    bf1[15] = bf0[15];
+    bf1[16] = _mm256_add_epi32(bf0[16], bf0[23]);
+    bf1[17] = _mm256_add_epi32(bf0[17], bf0[22]);
+    bf1[18] = _mm256_add_epi32(bf0[18], bf0[21]);
+    bf1[19] = _mm256_add_epi32(bf0[19], bf0[20]);
+    bf1[20] = _mm256_sub_epi32(bf0[19], bf0[20]);
+    bf1[21] = _mm256_sub_epi32(bf0[18], bf0[21]);
+    bf1[22] = _mm256_sub_epi32(bf0[17], bf0[22]);
+    bf1[23] = _mm256_sub_epi32(bf0[16], bf0[23]);
+    bf1[24] = _mm256_sub_epi32(bf0[31], bf0[24]);
+    bf1[25] = _mm256_sub_epi32(bf0[30], bf0[25]);
+    bf1[26] = _mm256_sub_epi32(bf0[29], bf0[26]);
+    bf1[27] = _mm256_sub_epi32(bf0[28], bf0[27]);
+    bf1[28] = _mm256_add_epi32(bf0[27], bf0[28]);
+    bf1[29] = _mm256_add_epi32(bf0[26], bf0[29]);
+    bf1[30] = _mm256_add_epi32(bf0[25], bf0[30]);
+    bf1[31] = _mm256_add_epi32(bf0[24], bf0[31]);
+
+    // stage 8
+    bf0[0] = _mm256_add_epi32(bf1[0], bf1[15]);
+    bf0[1] = _mm256_add_epi32(bf1[1], bf1[14]);
+    bf0[2] = _mm256_add_epi32(bf1[2], bf1[13]);
+    bf0[3] = _mm256_add_epi32(bf1[3], bf1[12]);
+    bf0[4] = _mm256_add_epi32(bf1[4], bf1[11]);
+    bf0[5] = _mm256_add_epi32(bf1[5], bf1[10]);
+    bf0[6] = _mm256_add_epi32(bf1[6], bf1[9]);
+    bf0[7] = _mm256_add_epi32(bf1[7], bf1[8]);
+    bf0[8] = _mm256_sub_epi32(bf1[7], bf1[8]);
+    bf0[9] = _mm256_sub_epi32(bf1[6], bf1[9]);
+    bf0[10] = _mm256_sub_epi32(bf1[5], bf1[10]);
+    bf0[11] = _mm256_sub_epi32(bf1[4], bf1[11]);
+    bf0[12] = _mm256_sub_epi32(bf1[3], bf1[12]);
+    bf0[13] = _mm256_sub_epi32(bf1[2], bf1[13]);
+    bf0[14] = _mm256_sub_epi32(bf1[1], bf1[14]);
+    bf0[15] = _mm256_sub_epi32(bf1[0], bf1[15]);
+    bf0[16] = bf1[16];
+    bf0[17] = bf1[17];
+    bf0[18] = bf1[18];
+    bf0[19] = bf1[19];
+    bf0[20] = half_btf_avx2(cospim32, bf1[20], cospi32, bf1[27], rounding, bit);
+    bf0[21] = half_btf_avx2(cospim32, bf1[21], cospi32, bf1[26], rounding, bit);
+    bf0[22] = half_btf_avx2(cospim32, bf1[22], cospi32, bf1[25], rounding, bit);
+    bf0[23] = half_btf_avx2(cospim32, bf1[23], cospi32, bf1[24], rounding, bit);
+    bf0[24] = half_btf_avx2(cospi32, bf1[23], cospi32, bf1[24], rounding, bit);
+    bf0[25] = half_btf_avx2(cospi32, bf1[22], cospi32, bf1[25], rounding, bit);
+    bf0[26] = half_btf_avx2(cospi32, bf1[21], cospi32, bf1[26], rounding, bit);
+    bf0[27] = half_btf_avx2(cospi32, bf1[20], cospi32, bf1[27], rounding, bit);
+    bf0[28] = bf1[28];
+    bf0[29] = bf1[29];
+    bf0[30] = bf1[30];
+    bf0[31] = bf1[31];
+
+    // stage 9
+    out[0 * 4 + col] = _mm256_add_epi32(bf0[0], bf0[31]);
+    out[1 * 4 + col] = _mm256_add_epi32(bf0[1], bf0[30]);
+    out[2 * 4 + col] = _mm256_add_epi32(bf0[2], bf0[29]);
+    out[3 * 4 + col] = _mm256_add_epi32(bf0[3], bf0[28]);
+    out[4 * 4 + col] = _mm256_add_epi32(bf0[4], bf0[27]);
+    out[5 * 4 + col] = _mm256_add_epi32(bf0[5], bf0[26]);
+    out[6 * 4 + col] = _mm256_add_epi32(bf0[6], bf0[25]);
+    out[7 * 4 + col] = _mm256_add_epi32(bf0[7], bf0[24]);
+    out[8 * 4 + col] = _mm256_add_epi32(bf0[8], bf0[23]);
+    out[9 * 4 + col] = _mm256_add_epi32(bf0[9], bf0[22]);
+    out[10 * 4 + col] = _mm256_add_epi32(bf0[10], bf0[21]);
+    out[11 * 4 + col] = _mm256_add_epi32(bf0[11], bf0[20]);
+    out[12 * 4 + col] = _mm256_add_epi32(bf0[12], bf0[19]);
+    out[13 * 4 + col] = _mm256_add_epi32(bf0[13], bf0[18]);
+    out[14 * 4 + col] = _mm256_add_epi32(bf0[14], bf0[17]);
+    out[15 * 4 + col] = _mm256_add_epi32(bf0[15], bf0[16]);
+    out[16 * 4 + col] = _mm256_sub_epi32(bf0[15], bf0[16]);
+    out[17 * 4 + col] = _mm256_sub_epi32(bf0[14], bf0[17]);
+    out[18 * 4 + col] = _mm256_sub_epi32(bf0[13], bf0[18]);
+    out[19 * 4 + col] = _mm256_sub_epi32(bf0[12], bf0[19]);
+    out[20 * 4 + col] = _mm256_sub_epi32(bf0[11], bf0[20]);
+    out[21 * 4 + col] = _mm256_sub_epi32(bf0[10], bf0[21]);
+    out[22 * 4 + col] = _mm256_sub_epi32(bf0[9], bf0[22]);
+    out[23 * 4 + col] = _mm256_sub_epi32(bf0[8], bf0[23]);
+    out[24 * 4 + col] = _mm256_sub_epi32(bf0[7], bf0[24]);
+    out[25 * 4 + col] = _mm256_sub_epi32(bf0[6], bf0[25]);
+    out[26 * 4 + col] = _mm256_sub_epi32(bf0[5], bf0[26]);
+    out[27 * 4 + col] = _mm256_sub_epi32(bf0[4], bf0[27]);
+    out[28 * 4 + col] = _mm256_sub_epi32(bf0[3], bf0[28]);
+    out[29 * 4 + col] = _mm256_sub_epi32(bf0[2], bf0[29]);
+    out[30 * 4 + col] = _mm256_sub_epi32(bf0[1], bf0[30]);
+    out[31 * 4 + col] = _mm256_sub_epi32(bf0[0], bf0[31]);
+  }
+}
+
+void av1_inv_txfm2d_add_32x32_avx2(const int32_t *coeff, uint16_t *output,
+                                   int stride, int tx_type, int bd) {
+  __m256i in[128], out[128];
+  const TXFM_2D_CFG *cfg = NULL;
+
+  switch (tx_type) {
+    case DCT_DCT:
+      cfg = &inv_txfm_2d_cfg_dct_dct_32;
+      load_buffer_32x32(coeff, in);
+      transpose_32x32(in, out);
+      idct32_avx2(out, in, cfg->cos_bit_row[2]);
+      round_shift_32x32(in, -cfg->shift[0]);
+      transpose_32x32(in, out);
+      idct32_avx2(out, in, cfg->cos_bit_col[2]);
+      write_buffer_32x32(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+    default: assert(0);
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_inv_txfm_sse4.c b/third_party/aom/av1/common/x86/highbd_inv_txfm_sse4.c
new file mode 100644
index 0000000000..24b2760b94
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_inv_txfm_sse4.c
@@ -0,0 +1,1398 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <assert.h>
+#include <smmintrin.h> /* SSE4.1 */
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "av1/common/av1_inv_txfm2d_cfg.h"
+#include "av1/common/x86/highbd_txfm_utility_sse4.h"
+
+static INLINE void load_buffer_4x4(const int32_t *coeff, __m128i *in) {
+  in[0] = _mm_load_si128((const __m128i *)(coeff + 0));
+  in[1] = _mm_load_si128((const __m128i *)(coeff + 4));
+  in[2] = _mm_load_si128((const __m128i *)(coeff + 8));
+  in[3] = _mm_load_si128((const __m128i *)(coeff + 12));
+}
+
+static void idct4x4_sse4_1(__m128i *in, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3, x, y;
+
+  v0 = _mm_unpacklo_epi32(in[0], in[1]);
+  v1 = _mm_unpackhi_epi32(in[0], in[1]);
+  v2 = _mm_unpacklo_epi32(in[2], in[3]);
+  v3 = _mm_unpackhi_epi32(in[2], in[3]);
+
+  u0 = _mm_unpacklo_epi64(v0, v2);
+  u1 = _mm_unpackhi_epi64(v0, v2);
+  u2 = _mm_unpacklo_epi64(v1, v3);
+  u3 = _mm_unpackhi_epi64(v1, v3);
+
+  x = _mm_mullo_epi32(u0, cospi32);
+  y = _mm_mullo_epi32(u2, cospi32);
+  v0 = _mm_add_epi32(x, y);
+  v0 = _mm_add_epi32(v0, rnding);
+  v0 = _mm_srai_epi32(v0, bit);
+
+  v1 = _mm_sub_epi32(x, y);
+  v1 = _mm_add_epi32(v1, rnding);
+  v1 = _mm_srai_epi32(v1, bit);
+
+  x = _mm_mullo_epi32(u1, cospi48);
+  y = _mm_mullo_epi32(u3, cospim16);
+  v2 = _mm_add_epi32(x, y);
+  v2 = _mm_add_epi32(v2, rnding);
+  v2 = _mm_srai_epi32(v2, bit);
+
+  x = _mm_mullo_epi32(u1, cospi16);
+  y = _mm_mullo_epi32(u3, cospi48);
+  v3 = _mm_add_epi32(x, y);
+  v3 = _mm_add_epi32(v3, rnding);
+  v3 = _mm_srai_epi32(v3, bit);
+
+  in[0] = _mm_add_epi32(v0, v3);
+  in[1] = _mm_add_epi32(v1, v2);
+  in[2] = _mm_sub_epi32(v1, v2);
+  in[3] = _mm_sub_epi32(v0, v3);
+}
+
+static void iadst4x4_sse4_1(__m128i *in, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const __m128i zero = _mm_setzero_si128();
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3, x, y;
+
+  v0 = _mm_unpacklo_epi32(in[0], in[1]);
+  v1 = _mm_unpackhi_epi32(in[0], in[1]);
+  v2 = _mm_unpacklo_epi32(in[2], in[3]);
+  v3 = _mm_unpackhi_epi32(in[2], in[3]);
+
+  u0 = _mm_unpacklo_epi64(v0, v2);
+  u1 = _mm_unpackhi_epi64(v0, v2);
+  u2 = _mm_unpacklo_epi64(v1, v3);
+  u3 = _mm_unpackhi_epi64(v1, v3);
+
+  // stage 0
+  // stage 1
+  u1 = _mm_sub_epi32(zero, u1);
+  u3 = _mm_sub_epi32(zero, u3);
+
+  // stage 2
+  v0 = u0;
+  v1 = u3;
+  x = _mm_mullo_epi32(u1, cospi32);
+  y = _mm_mullo_epi32(u2, cospi32);
+  v2 = _mm_add_epi32(x, y);
+  v2 = _mm_add_epi32(v2, rnding);
+  v2 = _mm_srai_epi32(v2, bit);
+
+  v3 = _mm_sub_epi32(x, y);
+  v3 = _mm_add_epi32(v3, rnding);
+  v3 = _mm_srai_epi32(v3, bit);
+
+  // stage 3
+  u0 = _mm_add_epi32(v0, v2);
+  u1 = _mm_add_epi32(v1, v3);
+  u2 = _mm_sub_epi32(v0, v2);
+  u3 = _mm_sub_epi32(v1, v3);
+
+  // stage 4
+  x = _mm_mullo_epi32(u0, cospi8);
+  y = _mm_mullo_epi32(u1, cospi56);
+  in[3] = _mm_add_epi32(x, y);
+  in[3] = _mm_add_epi32(in[3], rnding);
+  in[3] = _mm_srai_epi32(in[3], bit);
+
+  x = _mm_mullo_epi32(u0, cospi56);
+  y = _mm_mullo_epi32(u1, cospim8);
+  in[0] = _mm_add_epi32(x, y);
+  in[0] = _mm_add_epi32(in[0], rnding);
+  in[0] = _mm_srai_epi32(in[0], bit);
+
+  x = _mm_mullo_epi32(u2, cospi40);
+  y = _mm_mullo_epi32(u3, cospi24);
+  in[1] = _mm_add_epi32(x, y);
+  in[1] = _mm_add_epi32(in[1], rnding);
+  in[1] = _mm_srai_epi32(in[1], bit);
+
+  x = _mm_mullo_epi32(u2, cospi24);
+  y = _mm_mullo_epi32(u3, cospim40);
+  in[2] = _mm_add_epi32(x, y);
+  in[2] = _mm_add_epi32(in[2], rnding);
+  in[2] = _mm_srai_epi32(in[2], bit);
+}
+
+static INLINE void round_shift_4x4(__m128i *in, int shift) {
+  __m128i rnding = _mm_set1_epi32(1 << (shift - 1));
+
+  in[0] = _mm_add_epi32(in[0], rnding);
+  in[1] = _mm_add_epi32(in[1], rnding);
+  in[2] = _mm_add_epi32(in[2], rnding);
+  in[3] = _mm_add_epi32(in[3], rnding);
+
+  in[0] = _mm_srai_epi32(in[0], shift);
+  in[1] = _mm_srai_epi32(in[1], shift);
+  in[2] = _mm_srai_epi32(in[2], shift);
+  in[3] = _mm_srai_epi32(in[3], shift);
+}
+
+static INLINE __m128i highbd_clamp_epi16(__m128i u, int bd) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
+  __m128i clamped, mask;
+
+  mask = _mm_cmpgt_epi16(u, max);
+  clamped = _mm_andnot_si128(mask, u);
+  mask = _mm_and_si128(mask, max);
+  clamped = _mm_or_si128(mask, clamped);
+  mask = _mm_cmpgt_epi16(clamped, zero);
+  clamped = _mm_and_si128(clamped, mask);
+
+  return clamped;
+}
+
+static void write_buffer_4x4(__m128i *in, uint16_t *output, int stride,
+                             int fliplr, int flipud, int shift, int bd) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3;
+
+  round_shift_4x4(in, shift);
+
+  v0 = _mm_loadl_epi64((__m128i const *)(output + 0 * stride));
+  v1 = _mm_loadl_epi64((__m128i const *)(output + 1 * stride));
+  v2 = _mm_loadl_epi64((__m128i const *)(output + 2 * stride));
+  v3 = _mm_loadl_epi64((__m128i const *)(output + 3 * stride));
+
+  v0 = _mm_unpacklo_epi16(v0, zero);
+  v1 = _mm_unpacklo_epi16(v1, zero);
+  v2 = _mm_unpacklo_epi16(v2, zero);
+  v3 = _mm_unpacklo_epi16(v3, zero);
+
+  if (fliplr) {
+    in[0] = _mm_shuffle_epi32(in[0], 0x1B);
+    in[1] = _mm_shuffle_epi32(in[1], 0x1B);
+    in[2] = _mm_shuffle_epi32(in[2], 0x1B);
+    in[3] = _mm_shuffle_epi32(in[3], 0x1B);
+  }
+
+  if (flipud) {
+    u0 = _mm_add_epi32(in[3], v0);
+    u1 = _mm_add_epi32(in[2], v1);
+    u2 = _mm_add_epi32(in[1], v2);
+    u3 = _mm_add_epi32(in[0], v3);
+  } else {
+    u0 = _mm_add_epi32(in[0], v0);
+    u1 = _mm_add_epi32(in[1], v1);
+    u2 = _mm_add_epi32(in[2], v2);
+    u3 = _mm_add_epi32(in[3], v3);
+  }
+
+  v0 = _mm_packus_epi32(u0, u1);
+  v2 = _mm_packus_epi32(u2, u3);
+
+  u0 = highbd_clamp_epi16(v0, bd);
+  u2 = highbd_clamp_epi16(v2, bd);
+
+  v0 = _mm_unpacklo_epi64(u0, u0);
+  v1 = _mm_unpackhi_epi64(u0, u0);
+  v2 = _mm_unpacklo_epi64(u2, u2);
+  v3 = _mm_unpackhi_epi64(u2, u2);
+
+  _mm_storel_epi64((__m128i *)(output + 0 * stride), v0);
+  _mm_storel_epi64((__m128i *)(output + 1 * stride), v1);
+  _mm_storel_epi64((__m128i *)(output + 2 * stride), v2);
+  _mm_storel_epi64((__m128i *)(output + 3 * stride), v3);
+}
+
+void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *coeff, uint16_t *output,
+                                   int stride, int tx_type, int bd) {
+  __m128i in[4];
+  const TXFM_2D_CFG *cfg = NULL;
+
+  switch (tx_type) {
+    case DCT_DCT:
+      cfg = &inv_txfm_2d_cfg_dct_dct_4;
+      load_buffer_4x4(coeff, in);
+      idct4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      idct4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      write_buffer_4x4(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+    case ADST_DCT:
+      cfg = &inv_txfm_2d_cfg_adst_dct_4;
+      load_buffer_4x4(coeff, in);
+      idct4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      iadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      write_buffer_4x4(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+    case DCT_ADST:
+      cfg = &inv_txfm_2d_cfg_dct_adst_4;
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      idct4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      write_buffer_4x4(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+    case ADST_ADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_4;
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      iadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      write_buffer_4x4(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      cfg = &inv_txfm_2d_cfg_adst_dct_4;
+      load_buffer_4x4(coeff, in);
+      idct4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      iadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      write_buffer_4x4(in, output, stride, 0, 1, -cfg->shift[1], bd);
+      break;
+    case DCT_FLIPADST:
+      cfg = &inv_txfm_2d_cfg_dct_adst_4;
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      idct4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      write_buffer_4x4(in, output, stride, 1, 0, -cfg->shift[1], bd);
+      break;
+    case FLIPADST_FLIPADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_4;
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      iadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      write_buffer_4x4(in, output, stride, 1, 1, -cfg->shift[1], bd);
+      break;
+    case ADST_FLIPADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_4;
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      iadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      write_buffer_4x4(in, output, stride, 1, 0, -cfg->shift[1], bd);
+      break;
+    case FLIPADST_ADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_4;
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      iadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      write_buffer_4x4(in, output, stride, 0, 1, -cfg->shift[1], bd);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+}
+
+// 8x8
+static void load_buffer_8x8(const int32_t *coeff, __m128i *in) {
+  in[0] = _mm_load_si128((const __m128i *)(coeff + 0));
+  in[1] = _mm_load_si128((const __m128i *)(coeff + 4));
+  in[2] = _mm_load_si128((const __m128i *)(coeff + 8));
+  in[3] = _mm_load_si128((const __m128i *)(coeff + 12));
+  in[4] = _mm_load_si128((const __m128i *)(coeff + 16));
+  in[5] = _mm_load_si128((const __m128i *)(coeff + 20));
+  in[6] = _mm_load_si128((const __m128i *)(coeff + 24));
+  in[7] = _mm_load_si128((const __m128i *)(coeff + 28));
+  in[8] = _mm_load_si128((const __m128i *)(coeff + 32));
+  in[9] = _mm_load_si128((const __m128i *)(coeff + 36));
+  in[10] = _mm_load_si128((const __m128i *)(coeff + 40));
+  in[11] = _mm_load_si128((const __m128i *)(coeff + 44));
+  in[12] = _mm_load_si128((const __m128i *)(coeff + 48));
+  in[13] = _mm_load_si128((const __m128i *)(coeff + 52));
+  in[14] = _mm_load_si128((const __m128i *)(coeff + 56));
+  in[15] = _mm_load_si128((const __m128i *)(coeff + 60));
+}
+
+static void idct8x8_sse4_1(__m128i *in, __m128i *out, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i x, y;
+  int col;
+
+  // Note:
+  //  Even column: 0, 2, ..., 14
+  //  Odd column: 1, 3, ..., 15
+  //  one even column plus one odd column constructs one row (8 coeffs)
+  //  total we have 8 rows (8x8).
+  for (col = 0; col < 2; ++col) {
+    // stage 0
+    // stage 1
+    // stage 2
+    u0 = in[0 * 2 + col];
+    u1 = in[4 * 2 + col];
+    u2 = in[2 * 2 + col];
+    u3 = in[6 * 2 + col];
+
+    x = _mm_mullo_epi32(in[1 * 2 + col], cospi56);
+    y = _mm_mullo_epi32(in[7 * 2 + col], cospim8);
+    u4 = _mm_add_epi32(x, y);
+    u4 = _mm_add_epi32(u4, rnding);
+    u4 = _mm_srai_epi32(u4, bit);
+
+    x = _mm_mullo_epi32(in[1 * 2 + col], cospi8);
+    y = _mm_mullo_epi32(in[7 * 2 + col], cospi56);
+    u7 = _mm_add_epi32(x, y);
+    u7 = _mm_add_epi32(u7, rnding);
+    u7 = _mm_srai_epi32(u7, bit);
+
+    x = _mm_mullo_epi32(in[5 * 2 + col], cospi24);
+    y = _mm_mullo_epi32(in[3 * 2 + col], cospim40);
+    u5 = _mm_add_epi32(x, y);
+    u5 = _mm_add_epi32(u5, rnding);
+    u5 = _mm_srai_epi32(u5, bit);
+
+    x = _mm_mullo_epi32(in[5 * 2 + col], cospi40);
+    y = _mm_mullo_epi32(in[3 * 2 + col], cospi24);
+    u6 = _mm_add_epi32(x, y);
+    u6 = _mm_add_epi32(u6, rnding);
+    u6 = _mm_srai_epi32(u6, bit);
+
+    // stage 3
+    x = _mm_mullo_epi32(u0, cospi32);
+    y = _mm_mullo_epi32(u1, cospi32);
+    v0 = _mm_add_epi32(x, y);
+    v0 = _mm_add_epi32(v0, rnding);
+    v0 = _mm_srai_epi32(v0, bit);
+
+    v1 = _mm_sub_epi32(x, y);
+    v1 = _mm_add_epi32(v1, rnding);
+    v1 = _mm_srai_epi32(v1, bit);
+
+    x = _mm_mullo_epi32(u2, cospi48);
+    y = _mm_mullo_epi32(u3, cospim16);
+    v2 = _mm_add_epi32(x, y);
+    v2 = _mm_add_epi32(v2, rnding);
+    v2 = _mm_srai_epi32(v2, bit);
+
+    x = _mm_mullo_epi32(u2, cospi16);
+    y = _mm_mullo_epi32(u3, cospi48);
+    v3 = _mm_add_epi32(x, y);
+    v3 = _mm_add_epi32(v3, rnding);
+    v3 = _mm_srai_epi32(v3, bit);
+
+    v4 = _mm_add_epi32(u4, u5);
+    v5 = _mm_sub_epi32(u4, u5);
+    v6 = _mm_sub_epi32(u7, u6);
+    v7 = _mm_add_epi32(u6, u7);
+
+    // stage 4
+    u0 = _mm_add_epi32(v0, v3);
+    u1 = _mm_add_epi32(v1, v2);
+    u2 = _mm_sub_epi32(v1, v2);
+    u3 = _mm_sub_epi32(v0, v3);
+    u4 = v4;
+    u7 = v7;
+
+    x = _mm_mullo_epi32(v5, cospi32);
+    y = _mm_mullo_epi32(v6, cospi32);
+    u6 = _mm_add_epi32(y, x);
+    u6 = _mm_add_epi32(u6, rnding);
+    u6 = _mm_srai_epi32(u6, bit);
+
+    u5 = _mm_sub_epi32(y, x);
+    u5 = _mm_add_epi32(u5, rnding);
+    u5 = _mm_srai_epi32(u5, bit);
+
+    // stage 5
+    out[0 * 2 + col] = _mm_add_epi32(u0, u7);
+    out[1 * 2 + col] = _mm_add_epi32(u1, u6);
+    out[2 * 2 + col] = _mm_add_epi32(u2, u5);
+    out[3 * 2 + col] = _mm_add_epi32(u3, u4);
+    out[4 * 2 + col] = _mm_sub_epi32(u3, u4);
+    out[5 * 2 + col] = _mm_sub_epi32(u2, u5);
+    out[6 * 2 + col] = _mm_sub_epi32(u1, u6);
+    out[7 * 2 + col] = _mm_sub_epi32(u0, u7);
+  }
+}
+
+static void iadst8x8_sse4_1(__m128i *in, __m128i *out, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospim4 = _mm_set1_epi32(-cospi[4]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospim20 = _mm_set1_epi32(-cospi[20]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+  const __m128i cospim36 = _mm_set1_epi32(-cospi[36]);
+  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const __m128i zero = _mm_setzero_si128();
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i x, y;
+  int col;
+
+  // Note:
+  //  Even column: 0, 2, ..., 14
+  //  Odd column: 1, 3, ..., 15
+  //  one even column plus one odd column constructs one row (8 coeffs)
+  //  total we have 8 rows (8x8).
+  for (col = 0; col < 2; ++col) {
+    // stage 0
+    // stage 1
+    u0 = in[2 * 0 + col];
+    u1 = _mm_sub_epi32(zero, in[2 * 7 + col]);
+    u2 = _mm_sub_epi32(zero, in[2 * 3 + col]);
+    u3 = in[2 * 4 + col];
+    u4 = _mm_sub_epi32(zero, in[2 * 1 + col]);
+    u5 = in[2 * 6 + col];
+    u6 = in[2 * 2 + col];
+    u7 = _mm_sub_epi32(zero, in[2 * 5 + col]);
+
+    // stage 2
+    v0 = u0;
+    v1 = u1;
+
+    x = _mm_mullo_epi32(u2, cospi32);
+    y = _mm_mullo_epi32(u3, cospi32);
+    v2 = _mm_add_epi32(x, y);
+    v2 = _mm_add_epi32(v2, rnding);
+    v2 = _mm_srai_epi32(v2, bit);
+
+    v3 = _mm_sub_epi32(x, y);
+    v3 = _mm_add_epi32(v3, rnding);
+    v3 = _mm_srai_epi32(v3, bit);
+
+    v4 = u4;
+    v5 = u5;
+
+    x = _mm_mullo_epi32(u6, cospi32);
+    y = _mm_mullo_epi32(u7, cospi32);
+    v6 = _mm_add_epi32(x, y);
+    v6 = _mm_add_epi32(v6, rnding);
+    v6 = _mm_srai_epi32(v6, bit);
+
+    v7 = _mm_sub_epi32(x, y);
+    v7 = _mm_add_epi32(v7, rnding);
+    v7 = _mm_srai_epi32(v7, bit);
+
+    // stage 3
+    u0 = _mm_add_epi32(v0, v2);
+    u1 = _mm_add_epi32(v1, v3);
+    u2 = _mm_sub_epi32(v0, v2);
+    u3 = _mm_sub_epi32(v1, v3);
+    u4 = _mm_add_epi32(v4, v6);
+    u5 = _mm_add_epi32(v5, v7);
+    u6 = _mm_sub_epi32(v4, v6);
+    u7 = _mm_sub_epi32(v5, v7);
+
+    // stage 4
+    v0 = u0;
+    v1 = u1;
+    v2 = u2;
+    v3 = u3;
+
+    x = _mm_mullo_epi32(u4, cospi16);
+    y = _mm_mullo_epi32(u5, cospi48);
+    v4 = _mm_add_epi32(x, y);
+    v4 = _mm_add_epi32(v4, rnding);
+    v4 = _mm_srai_epi32(v4, bit);
+
+    x = _mm_mullo_epi32(u4, cospi48);
+    y = _mm_mullo_epi32(u5, cospim16);
+    v5 = _mm_add_epi32(x, y);
+    v5 = _mm_add_epi32(v5, rnding);
+    v5 = _mm_srai_epi32(v5, bit);
+
+    x = _mm_mullo_epi32(u6, cospim48);
+    y = _mm_mullo_epi32(u7, cospi16);
+    v6 = _mm_add_epi32(x, y);
+    v6 = _mm_add_epi32(v6, rnding);
+    v6 = _mm_srai_epi32(v6, bit);
+
+    x = _mm_mullo_epi32(u6, cospi16);
+    y = _mm_mullo_epi32(u7, cospi48);
+    v7 = _mm_add_epi32(x, y);
+    v7 = _mm_add_epi32(v7, rnding);
+    v7 = _mm_srai_epi32(v7, bit);
+
+    // stage 5
+    u0 = _mm_add_epi32(v0, v4);
+    u1 = _mm_add_epi32(v1, v5);
+    u2 = _mm_add_epi32(v2, v6);
+    u3 = _mm_add_epi32(v3, v7);
+    u4 = _mm_sub_epi32(v0, v4);
+    u5 = _mm_sub_epi32(v1, v5);
+    u6 = _mm_sub_epi32(v2, v6);
+    u7 = _mm_sub_epi32(v3, v7);
+
+    // stage 6
+    x = _mm_mullo_epi32(u0, cospi4);
+    y = _mm_mullo_epi32(u1, cospi60);
+    v0 = _mm_add_epi32(x, y);
+    v0 = _mm_add_epi32(v0, rnding);
+    v0 = _mm_srai_epi32(v0, bit);
+
+    x = _mm_mullo_epi32(u0, cospi60);
+    y = _mm_mullo_epi32(u1, cospim4);
+    v1 = _mm_add_epi32(x, y);
+    v1 = _mm_add_epi32(v1, rnding);
+    v1 = _mm_srai_epi32(v1, bit);
+
+    x = _mm_mullo_epi32(u2, cospi20);
+    y = _mm_mullo_epi32(u3, cospi44);
+    v2 = _mm_add_epi32(x, y);
+    v2 = _mm_add_epi32(v2, rnding);
+    v2 = _mm_srai_epi32(v2, bit);
+
+    x = _mm_mullo_epi32(u2, cospi44);
+    y = _mm_mullo_epi32(u3, cospim20);
+    v3 = _mm_add_epi32(x, y);
+    v3 = _mm_add_epi32(v3, rnding);
+    v3 = _mm_srai_epi32(v3, bit);
+
+    x = _mm_mullo_epi32(u4, cospi36);
+    y = _mm_mullo_epi32(u5, cospi28);
+    v4 = _mm_add_epi32(x, y);
+    v4 = _mm_add_epi32(v4, rnding);
+    v4 = _mm_srai_epi32(v4, bit);
+
+    x = _mm_mullo_epi32(u4, cospi28);
+    y = _mm_mullo_epi32(u5, cospim36);
+    v5 = _mm_add_epi32(x, y);
+    v5 = _mm_add_epi32(v5, rnding);
+    v5 = _mm_srai_epi32(v5, bit);
+
+    x = _mm_mullo_epi32(u6, cospi52);
+    y = _mm_mullo_epi32(u7, cospi12);
+    v6 = _mm_add_epi32(x, y);
+    v6 = _mm_add_epi32(v6, rnding);
+    v6 = _mm_srai_epi32(v6, bit);
+
+    x = _mm_mullo_epi32(u6, cospi12);
+    y = _mm_mullo_epi32(u7, cospim52);
+    v7 = _mm_add_epi32(x, y);
+    v7 = _mm_add_epi32(v7, rnding);
+    v7 = _mm_srai_epi32(v7, bit);
+
+    // stage 7
+    out[2 * 0 + col] = v1;
+    out[2 * 1 + col] = v6;
+    out[2 * 2 + col] = v3;
+    out[2 * 3 + col] = v4;
+    out[2 * 4 + col] = v5;
+    out[2 * 5 + col] = v2;
+    out[2 * 6 + col] = v7;
+    out[2 * 7 + col] = v0;
+  }
+}
+
+static void round_shift_8x8(__m128i *in, int shift) {
+  round_shift_4x4(&in[0], shift);
+  round_shift_4x4(&in[4], shift);
+  round_shift_4x4(&in[8], shift);
+  round_shift_4x4(&in[12], shift);
+}
+
+static __m128i get_recon_8x8(const __m128i pred, __m128i res_lo, __m128i res_hi,
+                             int fliplr, int bd) {
+  __m128i x0, x1;
+  const __m128i zero = _mm_setzero_si128();
+
+  x0 = _mm_unpacklo_epi16(pred, zero);
+  x1 = _mm_unpackhi_epi16(pred, zero);
+
+  if (fliplr) {
+    res_lo = _mm_shuffle_epi32(res_lo, 0x1B);
+    res_hi = _mm_shuffle_epi32(res_hi, 0x1B);
+    x0 = _mm_add_epi32(res_hi, x0);
+    x1 = _mm_add_epi32(res_lo, x1);
+
+  } else {
+    x0 = _mm_add_epi32(res_lo, x0);
+    x1 = _mm_add_epi32(res_hi, x1);
+  }
+
+  x0 = _mm_packus_epi32(x0, x1);
+  return highbd_clamp_epi16(x0, bd);
+}
+
+static void write_buffer_8x8(__m128i *in, uint16_t *output, int stride,
+                             int fliplr, int flipud, int shift, int bd) {
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+
+  round_shift_8x8(in, shift);
+
+  v0 = _mm_load_si128((__m128i const *)(output + 0 * stride));
+  v1 = _mm_load_si128((__m128i const *)(output + 1 * stride));
+  v2 = _mm_load_si128((__m128i const *)(output + 2 * stride));
+  v3 = _mm_load_si128((__m128i const *)(output + 3 * stride));
+  v4 = _mm_load_si128((__m128i const *)(output + 4 * stride));
+  v5 = _mm_load_si128((__m128i const *)(output + 5 * stride));
+  v6 = _mm_load_si128((__m128i const *)(output + 6 * stride));
+  v7 = _mm_load_si128((__m128i const *)(output + 7 * stride));
+
+  if (flipud) {
+    u0 = get_recon_8x8(v0, in[14], in[15], fliplr, bd);
+    u1 = get_recon_8x8(v1, in[12], in[13], fliplr, bd);
+    u2 = get_recon_8x8(v2, in[10], in[11], fliplr, bd);
+    u3 = get_recon_8x8(v3, in[8], in[9], fliplr, bd);
+    u4 = get_recon_8x8(v4, in[6], in[7], fliplr, bd);
+    u5 = get_recon_8x8(v5, in[4], in[5], fliplr, bd);
+    u6 = get_recon_8x8(v6, in[2], in[3], fliplr, bd);
+    u7 = get_recon_8x8(v7, in[0], in[1], fliplr, bd);
+  } else {
+    u0 = get_recon_8x8(v0, in[0], in[1], fliplr, bd);
+    u1 = get_recon_8x8(v1, in[2], in[3], fliplr, bd);
+    u2 = get_recon_8x8(v2, in[4], in[5], fliplr, bd);
+    u3 = get_recon_8x8(v3, in[6], in[7], fliplr, bd);
+    u4 = get_recon_8x8(v4, in[8], in[9], fliplr, bd);
+    u5 = get_recon_8x8(v5, in[10], in[11], fliplr, bd);
+    u6 = get_recon_8x8(v6, in[12], in[13], fliplr, bd);
+    u7 = get_recon_8x8(v7, in[14], in[15], fliplr, bd);
+  }
+
+  _mm_store_si128((__m128i *)(output + 0 * stride), u0);
+  _mm_store_si128((__m128i *)(output + 1 * stride), u1);
+  _mm_store_si128((__m128i *)(output + 2 * stride), u2);
+  _mm_store_si128((__m128i *)(output + 3 * stride), u3);
+  _mm_store_si128((__m128i *)(output + 4 * stride), u4);
+  _mm_store_si128((__m128i *)(output + 5 * stride), u5);
+  _mm_store_si128((__m128i *)(output + 6 * stride), u6);
+  _mm_store_si128((__m128i *)(output + 7 * stride), u7);
+}
+
+void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *coeff, uint16_t *output,
+                                   int stride, int tx_type, int bd) {
+  __m128i in[16], out[16];
+  const TXFM_2D_CFG *cfg = NULL;
+
+  switch (tx_type) {
+    case DCT_DCT:
+      cfg = &inv_txfm_2d_cfg_dct_dct_8;
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_8x8(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+    case DCT_ADST:
+      cfg = &inv_txfm_2d_cfg_dct_adst_8;
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_8x8(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+    case ADST_DCT:
+      cfg = &inv_txfm_2d_cfg_adst_dct_8;
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_8x8(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+    case ADST_ADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_8;
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_8x8(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      cfg = &inv_txfm_2d_cfg_adst_dct_8;
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_8x8(in, output, stride, 0, 1, -cfg->shift[1], bd);
+      break;
+    case DCT_FLIPADST:
+      cfg = &inv_txfm_2d_cfg_dct_adst_8;
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_8x8(in, output, stride, 1, 0, -cfg->shift[1], bd);
+      break;
+    case ADST_FLIPADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_8;
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_8x8(in, output, stride, 1, 0, -cfg->shift[1], bd);
+      break;
+    case FLIPADST_FLIPADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_8;
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_8x8(in, output, stride, 1, 1, -cfg->shift[1], bd);
+      break;
+    case FLIPADST_ADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_8;
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_row[2]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_8x8(in, output, stride, 0, 1, -cfg->shift[1], bd);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+}
+
+// 16x16
+static void load_buffer_16x16(const int32_t *coeff, __m128i *in) {
+  int i;
+  for (i = 0; i < 64; ++i) {
+    in[i] = _mm_load_si128((const __m128i *)(coeff + (i << 2)));
+  }
+}
+
+static void assign_8x8_input_from_16x16(const __m128i *in, __m128i *in8x8,
+                                        int col) {
+  int i;
+  for (i = 0; i < 16; i += 2) {
+    in8x8[i] = in[col];
+    in8x8[i + 1] = in[col + 1];
+    col += 4;
+  }
+}
+
+static void swap_addr(uint16_t **output1, uint16_t **output2) {
+  uint16_t *tmp;
+  tmp = *output1;
+  *output1 = *output2;
+  *output2 = tmp;
+}
+
+static void write_buffer_16x16(__m128i *in, uint16_t *output, int stride,
+                               int fliplr, int flipud, int shift, int bd) {
+  __m128i in8x8[16];
+  uint16_t *leftUp = &output[0];
+  uint16_t *rightUp = &output[8];
+  uint16_t *leftDown = &output[8 * stride];
+  uint16_t *rightDown = &output[8 * stride + 8];
+
+  if (fliplr) {
+    swap_addr(&leftUp, &rightUp);
+    swap_addr(&leftDown, &rightDown);
+  }
+
+  if (flipud) {
+    swap_addr(&leftUp, &leftDown);
+    swap_addr(&rightUp, &rightDown);
+  }
+
+  // Left-up quarter
+  assign_8x8_input_from_16x16(in, in8x8, 0);
+  write_buffer_8x8(in8x8, leftUp, stride, fliplr, flipud, shift, bd);
+
+  // Right-up quarter
+  assign_8x8_input_from_16x16(in, in8x8, 2);
+  write_buffer_8x8(in8x8, rightUp, stride, fliplr, flipud, shift, bd);
+
+  // Left-down quarter
+  assign_8x8_input_from_16x16(in, in8x8, 32);
+  write_buffer_8x8(in8x8, leftDown, stride, fliplr, flipud, shift, bd);
+
+  // Right-down quarter
+  assign_8x8_input_from_16x16(in, in8x8, 34);
+  write_buffer_8x8(in8x8, rightDown, stride, fliplr, flipud, shift, bd);
+}
+
+static void idct16x16_sse4_1(__m128i *in, __m128i *out, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospim4 = _mm_set1_epi32(-cospi[4]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospim36 = _mm_set1_epi32(-cospi[36]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospim20 = _mm_set1_epi32(-cospi[20]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
+  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  __m128i u[16], v[16], x, y;
+  int col;
+
+  for (col = 0; col < 4; ++col) {
+    // stage 0
+    // stage 1
+    u[0] = in[0 * 4 + col];
+    u[1] = in[8 * 4 + col];
+    u[2] = in[4 * 4 + col];
+    u[3] = in[12 * 4 + col];
+    u[4] = in[2 * 4 + col];
+    u[5] = in[10 * 4 + col];
+    u[6] = in[6 * 4 + col];
+    u[7] = in[14 * 4 + col];
+    u[8] = in[1 * 4 + col];
+    u[9] = in[9 * 4 + col];
+    u[10] = in[5 * 4 + col];
+    u[11] = in[13 * 4 + col];
+    u[12] = in[3 * 4 + col];
+    u[13] = in[11 * 4 + col];
+    u[14] = in[7 * 4 + col];
+    u[15] = in[15 * 4 + col];
+
+    // stage 2
+    v[0] = u[0];
+    v[1] = u[1];
+    v[2] = u[2];
+    v[3] = u[3];
+    v[4] = u[4];
+    v[5] = u[5];
+    v[6] = u[6];
+    v[7] = u[7];
+
+    v[8] = half_btf_sse4_1(cospi60, u[8], cospim4, u[15], rnding, bit);
+    v[9] = half_btf_sse4_1(cospi28, u[9], cospim36, u[14], rnding, bit);
+    v[10] = half_btf_sse4_1(cospi44, u[10], cospim20, u[13], rnding, bit);
+    v[11] = half_btf_sse4_1(cospi12, u[11], cospim52, u[12], rnding, bit);
+    v[12] = half_btf_sse4_1(cospi52, u[11], cospi12, u[12], rnding, bit);
+    v[13] = half_btf_sse4_1(cospi20, u[10], cospi44, u[13], rnding, bit);
+    v[14] = half_btf_sse4_1(cospi36, u[9], cospi28, u[14], rnding, bit);
+    v[15] = half_btf_sse4_1(cospi4, u[8], cospi60, u[15], rnding, bit);
+
+    // stage 3
+    u[0] = v[0];
+    u[1] = v[1];
+    u[2] = v[2];
+    u[3] = v[3];
+    u[4] = half_btf_sse4_1(cospi56, v[4], cospim8, v[7], rnding, bit);
+    u[5] = half_btf_sse4_1(cospi24, v[5], cospim40, v[6], rnding, bit);
+    u[6] = half_btf_sse4_1(cospi40, v[5], cospi24, v[6], rnding, bit);
+    u[7] = half_btf_sse4_1(cospi8, v[4], cospi56, v[7], rnding, bit);
+    u[8] = _mm_add_epi32(v[8], v[9]);
+    u[9] = _mm_sub_epi32(v[8], v[9]);
+    u[10] = _mm_sub_epi32(v[11], v[10]);
+    u[11] = _mm_add_epi32(v[10], v[11]);
+    u[12] = _mm_add_epi32(v[12], v[13]);
+    u[13] = _mm_sub_epi32(v[12], v[13]);
+    u[14] = _mm_sub_epi32(v[15], v[14]);
+    u[15] = _mm_add_epi32(v[14], v[15]);
+
+    // stage 4
+    x = _mm_mullo_epi32(u[0], cospi32);
+    y = _mm_mullo_epi32(u[1], cospi32);
+    v[0] = _mm_add_epi32(x, y);
+    v[0] = _mm_add_epi32(v[0], rnding);
+    v[0] = _mm_srai_epi32(v[0], bit);
+
+    v[1] = _mm_sub_epi32(x, y);
+    v[1] = _mm_add_epi32(v[1], rnding);
+    v[1] = _mm_srai_epi32(v[1], bit);
+
+    v[2] = half_btf_sse4_1(cospi48, u[2], cospim16, u[3], rnding, bit);
+    v[3] = half_btf_sse4_1(cospi16, u[2], cospi48, u[3], rnding, bit);
+    v[4] = _mm_add_epi32(u[4], u[5]);
+    v[5] = _mm_sub_epi32(u[4], u[5]);
+    v[6] = _mm_sub_epi32(u[7], u[6]);
+    v[7] = _mm_add_epi32(u[6], u[7]);
+    v[8] = u[8];
+    v[9] = half_btf_sse4_1(cospim16, u[9], cospi48, u[14], rnding, bit);
+    v[10] = half_btf_sse4_1(cospim48, u[10], cospim16, u[13], rnding, bit);
+    v[11] = u[11];
+    v[12] = u[12];
+    v[13] = half_btf_sse4_1(cospim16, u[10], cospi48, u[13], rnding, bit);
+    v[14] = half_btf_sse4_1(cospi48, u[9], cospi16, u[14], rnding, bit);
+    v[15] = u[15];
+
+    // stage 5
+    u[0] = _mm_add_epi32(v[0], v[3]);
+    u[1] = _mm_add_epi32(v[1], v[2]);
+    u[2] = _mm_sub_epi32(v[1], v[2]);
+    u[3] = _mm_sub_epi32(v[0], v[3]);
+    u[4] = v[4];
+
+    x = _mm_mullo_epi32(v[5], cospi32);
+    y = _mm_mullo_epi32(v[6], cospi32);
+    u[5] = _mm_sub_epi32(y, x);
+    u[5] = _mm_add_epi32(u[5], rnding);
+    u[5] = _mm_srai_epi32(u[5], bit);
+
+    u[6] = _mm_add_epi32(y, x);
+    u[6] = _mm_add_epi32(u[6], rnding);
+    u[6] = _mm_srai_epi32(u[6], bit);
+
+    u[7] = v[7];
+    u[8] = _mm_add_epi32(v[8], v[11]);
+    u[9] = _mm_add_epi32(v[9], v[10]);
+    u[10] = _mm_sub_epi32(v[9], v[10]);
+    u[11] = _mm_sub_epi32(v[8], v[11]);
+    u[12] = _mm_sub_epi32(v[15], v[12]);
+    u[13] = _mm_sub_epi32(v[14], v[13]);
+    u[14] = _mm_add_epi32(v[13], v[14]);
+    u[15] = _mm_add_epi32(v[12], v[15]);
+
+    // stage 6
+    v[0] = _mm_add_epi32(u[0], u[7]);
+    v[1] = _mm_add_epi32(u[1], u[6]);
+    v[2] = _mm_add_epi32(u[2], u[5]);
+    v[3] = _mm_add_epi32(u[3], u[4]);
+    v[4] = _mm_sub_epi32(u[3], u[4]);
+    v[5] = _mm_sub_epi32(u[2], u[5]);
+    v[6] = _mm_sub_epi32(u[1], u[6]);
+    v[7] = _mm_sub_epi32(u[0], u[7]);
+    v[8] = u[8];
+    v[9] = u[9];
+
+    x = _mm_mullo_epi32(u[10], cospi32);
+    y = _mm_mullo_epi32(u[13], cospi32);
+    v[10] = _mm_sub_epi32(y, x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[13] = _mm_add_epi32(x, y);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    x = _mm_mullo_epi32(u[11], cospi32);
+    y = _mm_mullo_epi32(u[12], cospi32);
+    v[11] = _mm_sub_epi32(y, x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = _mm_add_epi32(x, y);
+    v[12] = _mm_add_epi32(v[12], rnding);
+    v[12] = _mm_srai_epi32(v[12], bit);
+
+    v[14] = u[14];
+    v[15] = u[15];
+
+    // stage 7
+    out[0 * 4 + col] = _mm_add_epi32(v[0], v[15]);
+    out[1 * 4 + col] = _mm_add_epi32(v[1], v[14]);
+    out[2 * 4 + col] = _mm_add_epi32(v[2], v[13]);
+    out[3 * 4 + col] = _mm_add_epi32(v[3], v[12]);
+    out[4 * 4 + col] = _mm_add_epi32(v[4], v[11]);
+    out[5 * 4 + col] = _mm_add_epi32(v[5], v[10]);
+    out[6 * 4 + col] = _mm_add_epi32(v[6], v[9]);
+    out[7 * 4 + col] = _mm_add_epi32(v[7], v[8]);
+    out[8 * 4 + col] = _mm_sub_epi32(v[7], v[8]);
+    out[9 * 4 + col] = _mm_sub_epi32(v[6], v[9]);
+    out[10 * 4 + col] = _mm_sub_epi32(v[5], v[10]);
+    out[11 * 4 + col] = _mm_sub_epi32(v[4], v[11]);
+    out[12 * 4 + col] = _mm_sub_epi32(v[3], v[12]);
+    out[13 * 4 + col] = _mm_sub_epi32(v[2], v[13]);
+    out[14 * 4 + col] = _mm_sub_epi32(v[1], v[14]);
+    out[15 * 4 + col] = _mm_sub_epi32(v[0], v[15]);
+  }
+}
+
+static void iadst16x16_sse4_1(__m128i *in, __m128i *out, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospim56 = _mm_set1_epi32(-cospi[56]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospim24 = _mm_set1_epi32(-cospi[24]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+  const __m128i cospim2 = _mm_set1_epi32(-cospi[2]);
+  const __m128i cospi10 = _mm_set1_epi32(cospi[10]);
+  const __m128i cospi54 = _mm_set1_epi32(cospi[54]);
+  const __m128i cospim10 = _mm_set1_epi32(-cospi[10]);
+  const __m128i cospi18 = _mm_set1_epi32(cospi[18]);
+  const __m128i cospi46 = _mm_set1_epi32(cospi[46]);
+  const __m128i cospim18 = _mm_set1_epi32(-cospi[18]);
+  const __m128i cospi26 = _mm_set1_epi32(cospi[26]);
+  const __m128i cospi38 = _mm_set1_epi32(cospi[38]);
+  const __m128i cospim26 = _mm_set1_epi32(-cospi[26]);
+  const __m128i cospi34 = _mm_set1_epi32(cospi[34]);
+  const __m128i cospi30 = _mm_set1_epi32(cospi[30]);
+  const __m128i cospim34 = _mm_set1_epi32(-cospi[34]);
+  const __m128i cospi42 = _mm_set1_epi32(cospi[42]);
+  const __m128i cospi22 = _mm_set1_epi32(cospi[22]);
+  const __m128i cospim42 = _mm_set1_epi32(-cospi[42]);
+  const __m128i cospi50 = _mm_set1_epi32(cospi[50]);
+  const __m128i cospi14 = _mm_set1_epi32(cospi[14]);
+  const __m128i cospim50 = _mm_set1_epi32(-cospi[50]);
+  const __m128i cospi58 = _mm_set1_epi32(cospi[58]);
+  const __m128i cospi6 = _mm_set1_epi32(cospi[6]);
+  const __m128i cospim58 = _mm_set1_epi32(-cospi[58]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const __m128i zero = _mm_setzero_si128();
+
+  __m128i u[16], v[16], x, y;
+  int col;
+
+  for (col = 0; col < 4; ++col) {
+    // stage 0
+    // stage 1
+    u[0] = in[0 * 4 + col];
+    u[1] = _mm_sub_epi32(zero, in[15 * 4 + col]);
+    u[2] = _mm_sub_epi32(zero, in[7 * 4 + col]);
+    u[3] = in[8 * 4 + col];
+    u[4] = _mm_sub_epi32(zero, in[3 * 4 + col]);
+    u[5] = in[12 * 4 + col];
+    u[6] = in[4 * 4 + col];
+    u[7] = _mm_sub_epi32(zero, in[11 * 4 + col]);
+    u[8] = _mm_sub_epi32(zero, in[1 * 4 + col]);
+    u[9] = in[14 * 4 + col];
+    u[10] = in[6 * 4 + col];
+    u[11] = _mm_sub_epi32(zero, in[9 * 4 + col]);
+    u[12] = in[2 * 4 + col];
+    u[13] = _mm_sub_epi32(zero, in[13 * 4 + col]);
+    u[14] = _mm_sub_epi32(zero, in[5 * 4 + col]);
+    u[15] = in[10 * 4 + col];
+
+    // stage 2
+    v[0] = u[0];
+    v[1] = u[1];
+
+    x = _mm_mullo_epi32(u[2], cospi32);
+    y = _mm_mullo_epi32(u[3], cospi32);
+    v[2] = _mm_add_epi32(x, y);
+    v[2] = _mm_add_epi32(v[2], rnding);
+    v[2] = _mm_srai_epi32(v[2], bit);
+
+    v[3] = _mm_sub_epi32(x, y);
+    v[3] = _mm_add_epi32(v[3], rnding);
+    v[3] = _mm_srai_epi32(v[3], bit);
+
+    v[4] = u[4];
+    v[5] = u[5];
+
+    x = _mm_mullo_epi32(u[6], cospi32);
+    y = _mm_mullo_epi32(u[7], cospi32);
+    v[6] = _mm_add_epi32(x, y);
+    v[6] = _mm_add_epi32(v[6], rnding);
+    v[6] = _mm_srai_epi32(v[6], bit);
+
+    v[7] = _mm_sub_epi32(x, y);
+    v[7] = _mm_add_epi32(v[7], rnding);
+    v[7] = _mm_srai_epi32(v[7], bit);
+
+    v[8] = u[8];
+    v[9] = u[9];
+
+    x = _mm_mullo_epi32(u[10], cospi32);
+    y = _mm_mullo_epi32(u[11], cospi32);
+    v[10] = _mm_add_epi32(x, y);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[11] = _mm_sub_epi32(x, y);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = u[12];
+    v[13] = u[13];
+
+    x = _mm_mullo_epi32(u[14], cospi32);
+    y = _mm_mullo_epi32(u[15], cospi32);
+    v[14] = _mm_add_epi32(x, y);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[15] = _mm_sub_epi32(x, y);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    // stage 3
+    u[0] = _mm_add_epi32(v[0], v[2]);
+    u[1] = _mm_add_epi32(v[1], v[3]);
+    u[2] = _mm_sub_epi32(v[0], v[2]);
+    u[3] = _mm_sub_epi32(v[1], v[3]);
+    u[4] = _mm_add_epi32(v[4], v[6]);
+    u[5] = _mm_add_epi32(v[5], v[7]);
+    u[6] = _mm_sub_epi32(v[4], v[6]);
+    u[7] = _mm_sub_epi32(v[5], v[7]);
+    u[8] = _mm_add_epi32(v[8], v[10]);
+    u[9] = _mm_add_epi32(v[9], v[11]);
+    u[10] = _mm_sub_epi32(v[8], v[10]);
+    u[11] = _mm_sub_epi32(v[9], v[11]);
+    u[12] = _mm_add_epi32(v[12], v[14]);
+    u[13] = _mm_add_epi32(v[13], v[15]);
+    u[14] = _mm_sub_epi32(v[12], v[14]);
+    u[15] = _mm_sub_epi32(v[13], v[15]);
+
+    // stage 4
+    v[0] = u[0];
+    v[1] = u[1];
+    v[2] = u[2];
+    v[3] = u[3];
+    v[4] = half_btf_sse4_1(cospi16, u[4], cospi48, u[5], rnding, bit);
+    v[5] = half_btf_sse4_1(cospi48, u[4], cospim16, u[5], rnding, bit);
+    v[6] = half_btf_sse4_1(cospim48, u[6], cospi16, u[7], rnding, bit);
+    v[7] = half_btf_sse4_1(cospi16, u[6], cospi48, u[7], rnding, bit);
+    v[8] = u[8];
+    v[9] = u[9];
+    v[10] = u[10];
+    v[11] = u[11];
+    v[12] = half_btf_sse4_1(cospi16, u[12], cospi48, u[13], rnding, bit);
+    v[13] = half_btf_sse4_1(cospi48, u[12], cospim16, u[13], rnding, bit);
+    v[14] = half_btf_sse4_1(cospim48, u[14], cospi16, u[15], rnding, bit);
+    v[15] = half_btf_sse4_1(cospi16, u[14], cospi48, u[15], rnding, bit);
+
+    // stage 5
+    u[0] = _mm_add_epi32(v[0], v[4]);
+    u[1] = _mm_add_epi32(v[1], v[5]);
+    u[2] = _mm_add_epi32(v[2], v[6]);
+    u[3] = _mm_add_epi32(v[3], v[7]);
+    u[4] = _mm_sub_epi32(v[0], v[4]);
+    u[5] = _mm_sub_epi32(v[1], v[5]);
+    u[6] = _mm_sub_epi32(v[2], v[6]);
+    u[7] = _mm_sub_epi32(v[3], v[7]);
+    u[8] = _mm_add_epi32(v[8], v[12]);
+    u[9] = _mm_add_epi32(v[9], v[13]);
+    u[10] = _mm_add_epi32(v[10], v[14]);
+    u[11] = _mm_add_epi32(v[11], v[15]);
+    u[12] = _mm_sub_epi32(v[8], v[12]);
+    u[13] = _mm_sub_epi32(v[9], v[13]);
+    u[14] = _mm_sub_epi32(v[10], v[14]);
+    u[15] = _mm_sub_epi32(v[11], v[15]);
+
+    // stage 6
+    v[0] = u[0];
+    v[1] = u[1];
+    v[2] = u[2];
+    v[3] = u[3];
+    v[4] = u[4];
+    v[5] = u[5];
+    v[6] = u[6];
+    v[7] = u[7];
+    v[8] = half_btf_sse4_1(cospi8, u[8], cospi56, u[9], rnding, bit);
+    v[9] = half_btf_sse4_1(cospi56, u[8], cospim8, u[9], rnding, bit);
+    v[10] = half_btf_sse4_1(cospi40, u[10], cospi24, u[11], rnding, bit);
+    v[11] = half_btf_sse4_1(cospi24, u[10], cospim40, u[11], rnding, bit);
+    v[12] = half_btf_sse4_1(cospim56, u[12], cospi8, u[13], rnding, bit);
+    v[13] = half_btf_sse4_1(cospi8, u[12], cospi56, u[13], rnding, bit);
+    v[14] = half_btf_sse4_1(cospim24, u[14], cospi40, u[15], rnding, bit);
+    v[15] = half_btf_sse4_1(cospi40, u[14], cospi24, u[15], rnding, bit);
+
+    // stage 7
+    u[0] = _mm_add_epi32(v[0], v[8]);
+    u[1] = _mm_add_epi32(v[1], v[9]);
+    u[2] = _mm_add_epi32(v[2], v[10]);
+    u[3] = _mm_add_epi32(v[3], v[11]);
+    u[4] = _mm_add_epi32(v[4], v[12]);
+    u[5] = _mm_add_epi32(v[5], v[13]);
+    u[6] = _mm_add_epi32(v[6], v[14]);
+    u[7] = _mm_add_epi32(v[7], v[15]);
+    u[8] = _mm_sub_epi32(v[0], v[8]);
+    u[9] = _mm_sub_epi32(v[1], v[9]);
+    u[10] = _mm_sub_epi32(v[2], v[10]);
+    u[11] = _mm_sub_epi32(v[3], v[11]);
+    u[12] = _mm_sub_epi32(v[4], v[12]);
+    u[13] = _mm_sub_epi32(v[5], v[13]);
+    u[14] = _mm_sub_epi32(v[6], v[14]);
+    u[15] = _mm_sub_epi32(v[7], v[15]);
+
+    // stage 8
+    v[0] = half_btf_sse4_1(cospi2, u[0], cospi62, u[1], rnding, bit);
+    v[1] = half_btf_sse4_1(cospi62, u[0], cospim2, u[1], rnding, bit);
+    v[2] = half_btf_sse4_1(cospi10, u[2], cospi54, u[3], rnding, bit);
+    v[3] = half_btf_sse4_1(cospi54, u[2], cospim10, u[3], rnding, bit);
+    v[4] = half_btf_sse4_1(cospi18, u[4], cospi46, u[5], rnding, bit);
+    v[5] = half_btf_sse4_1(cospi46, u[4], cospim18, u[5], rnding, bit);
+    v[6] = half_btf_sse4_1(cospi26, u[6], cospi38, u[7], rnding, bit);
+    v[7] = half_btf_sse4_1(cospi38, u[6], cospim26, u[7], rnding, bit);
+    v[8] = half_btf_sse4_1(cospi34, u[8], cospi30, u[9], rnding, bit);
+    v[9] = half_btf_sse4_1(cospi30, u[8], cospim34, u[9], rnding, bit);
+    v[10] = half_btf_sse4_1(cospi42, u[10], cospi22, u[11], rnding, bit);
+    v[11] = half_btf_sse4_1(cospi22, u[10], cospim42, u[11], rnding, bit);
+    v[12] = half_btf_sse4_1(cospi50, u[12], cospi14, u[13], rnding, bit);
+    v[13] = half_btf_sse4_1(cospi14, u[12], cospim50, u[13], rnding, bit);
+    v[14] = half_btf_sse4_1(cospi58, u[14], cospi6, u[15], rnding, bit);
+    v[15] = half_btf_sse4_1(cospi6, u[14], cospim58, u[15], rnding, bit);
+
+    // stage 9
+    out[0 * 4 + col] = v[1];
+    out[1 * 4 + col] = v[14];
+    out[2 * 4 + col] = v[3];
+    out[3 * 4 + col] = v[12];
+    out[4 * 4 + col] = v[5];
+    out[5 * 4 + col] = v[10];
+    out[6 * 4 + col] = v[7];
+    out[7 * 4 + col] = v[8];
+    out[8 * 4 + col] = v[9];
+    out[9 * 4 + col] = v[6];
+    out[10 * 4 + col] = v[11];
+    out[11 * 4 + col] = v[4];
+    out[12 * 4 + col] = v[13];
+    out[13 * 4 + col] = v[2];
+    out[14 * 4 + col] = v[15];
+    out[15 * 4 + col] = v[0];
+  }
+}
+
+static void round_shift_16x16(__m128i *in, int shift) {
+  round_shift_8x8(&in[0], shift);
+  round_shift_8x8(&in[16], shift);
+  round_shift_8x8(&in[32], shift);
+  round_shift_8x8(&in[48], shift);
+}
+
+void av1_inv_txfm2d_add_16x16_sse4_1(const int32_t *coeff, uint16_t *output,
+                                     int stride, int tx_type, int bd) {
+  __m128i in[64], out[64];
+  const TXFM_2D_CFG *cfg = NULL;
+
+  switch (tx_type) {
+    case DCT_DCT:
+      cfg = &inv_txfm_2d_cfg_dct_dct_16;
+      load_buffer_16x16(coeff, in);
+      transpose_16x16(in, out);
+      idct16x16_sse4_1(out, in, cfg->cos_bit_row[2]);
+      round_shift_16x16(in, -cfg->shift[0]);
+      transpose_16x16(in, out);
+      idct16x16_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_16x16(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+    case DCT_ADST:
+      cfg = &inv_txfm_2d_cfg_dct_adst_16;
+      load_buffer_16x16(coeff, in);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_row[2]);
+      round_shift_16x16(in, -cfg->shift[0]);
+      transpose_16x16(in, out);
+      idct16x16_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_16x16(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+    case ADST_DCT:
+      cfg = &inv_txfm_2d_cfg_adst_dct_16;
+      load_buffer_16x16(coeff, in);
+      transpose_16x16(in, out);
+      idct16x16_sse4_1(out, in, cfg->cos_bit_row[2]);
+      round_shift_16x16(in, -cfg->shift[0]);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_16x16(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+    case ADST_ADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_16;
+      load_buffer_16x16(coeff, in);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_row[2]);
+      round_shift_16x16(in, -cfg->shift[0]);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_16x16(in, output, stride, 0, 0, -cfg->shift[1], bd);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      cfg = &inv_txfm_2d_cfg_adst_dct_16;
+      load_buffer_16x16(coeff, in);
+      transpose_16x16(in, out);
+      idct16x16_sse4_1(out, in, cfg->cos_bit_row[2]);
+      round_shift_16x16(in, -cfg->shift[0]);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_16x16(in, output, stride, 0, 1, -cfg->shift[1], bd);
+      break;
+    case DCT_FLIPADST:
+      cfg = &inv_txfm_2d_cfg_dct_adst_16;
+      load_buffer_16x16(coeff, in);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_row[2]);
+      round_shift_16x16(in, -cfg->shift[0]);
+      transpose_16x16(in, out);
+      idct16x16_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_16x16(in, output, stride, 1, 0, -cfg->shift[1], bd);
+      break;
+    case ADST_FLIPADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_16;
+      load_buffer_16x16(coeff, in);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_row[2]);
+      round_shift_16x16(in, -cfg->shift[0]);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_16x16(in, output, stride, 1, 0, -cfg->shift[1], bd);
+      break;
+    case FLIPADST_FLIPADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_16;
+      load_buffer_16x16(coeff, in);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_row[2]);
+      round_shift_16x16(in, -cfg->shift[0]);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_16x16(in, output, stride, 1, 1, -cfg->shift[1], bd);
+      break;
+    case FLIPADST_ADST:
+      cfg = &inv_txfm_2d_cfg_adst_adst_16;
+      load_buffer_16x16(coeff, in);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_row[2]);
+      round_shift_16x16(in, -cfg->shift[0]);
+      transpose_16x16(in, out);
+      iadst16x16_sse4_1(out, in, cfg->cos_bit_col[2]);
+      write_buffer_16x16(in, output, stride, 0, 1, -cfg->shift[1], bd);
+      break;
+#endif
+    default: assert(0);
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_txfm_utility_sse4.h b/third_party/aom/av1/common/x86/highbd_txfm_utility_sse4.h
new file mode 100644
index 0000000000..bc96defe3b
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_txfm_utility_sse4.h
@@ -0,0 +1,92 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef _HIGHBD_TXFM_UTILITY_SSE4_H
+#define _HIGHBD_TXFM_UTILITY_SSE4_H
+
+#include <smmintrin.h> /* SSE4.1 */
+
+#define TRANSPOSE_4X4(x0, x1, x2, x3, y0, y1, y2, y3) \
+  do {                                                \
+    __m128i u0, u1, u2, u3;                           \
+    u0 = _mm_unpacklo_epi32(x0, x1);                  \
+    u1 = _mm_unpackhi_epi32(x0, x1);                  \
+    u2 = _mm_unpacklo_epi32(x2, x3);                  \
+    u3 = _mm_unpackhi_epi32(x2, x3);                  \
+    y0 = _mm_unpacklo_epi64(u0, u2);                  \
+    y1 = _mm_unpackhi_epi64(u0, u2);                  \
+    y2 = _mm_unpacklo_epi64(u1, u3);                  \
+    y3 = _mm_unpackhi_epi64(u1, u3);                  \
+  } while (0)
+
+static INLINE void transpose_8x8(const __m128i *in, __m128i *out) {
+  TRANSPOSE_4X4(in[0], in[2], in[4], in[6], out[0], out[2], out[4], out[6]);
+  TRANSPOSE_4X4(in[1], in[3], in[5], in[7], out[8], out[10], out[12], out[14]);
+  TRANSPOSE_4X4(in[8], in[10], in[12], in[14], out[1], out[3], out[5], out[7]);
+  TRANSPOSE_4X4(in[9], in[11], in[13], in[15], out[9], out[11], out[13],
+                out[15]);
+}
+
+static INLINE void transpose_16x16(const __m128i *in, __m128i *out) {
+  // Upper left 8x8
+  TRANSPOSE_4X4(in[0], in[4], in[8], in[12], out[0], out[4], out[8], out[12]);
+  TRANSPOSE_4X4(in[1], in[5], in[9], in[13], out[16], out[20], out[24],
+                out[28]);
+  TRANSPOSE_4X4(in[16], in[20], in[24], in[28], out[1], out[5], out[9],
+                out[13]);
+  TRANSPOSE_4X4(in[17], in[21], in[25], in[29], out[17], out[21], out[25],
+                out[29]);
+
+  // Upper right 8x8
+  TRANSPOSE_4X4(in[2], in[6], in[10], in[14], out[32], out[36], out[40],
+                out[44]);
+  TRANSPOSE_4X4(in[3], in[7], in[11], in[15], out[48], out[52], out[56],
+                out[60]);
+  TRANSPOSE_4X4(in[18], in[22], in[26], in[30], out[33], out[37], out[41],
+                out[45]);
+  TRANSPOSE_4X4(in[19], in[23], in[27], in[31], out[49], out[53], out[57],
+                out[61]);
+
+  // Lower left 8x8
+  TRANSPOSE_4X4(in[32], in[36], in[40], in[44], out[2], out[6], out[10],
+                out[14]);
+  TRANSPOSE_4X4(in[33], in[37], in[41], in[45], out[18], out[22], out[26],
+                out[30]);
+  TRANSPOSE_4X4(in[48], in[52], in[56], in[60], out[3], out[7], out[11],
+                out[15]);
+  TRANSPOSE_4X4(in[49], in[53], in[57], in[61], out[19], out[23], out[27],
+                out[31]);
+  // Lower right 8x8
+  TRANSPOSE_4X4(in[34], in[38], in[42], in[46], out[34], out[38], out[42],
+                out[46]);
+  TRANSPOSE_4X4(in[35], in[39], in[43], in[47], out[50], out[54], out[58],
+                out[62]);
+  TRANSPOSE_4X4(in[50], in[54], in[58], in[62], out[35], out[39], out[43],
+                out[47]);
+  TRANSPOSE_4X4(in[51], in[55], in[59], in[63], out[51], out[55], out[59],
+                out[63]);
+}
+
+// Note:
+//  rounding = 1 << (bit - 1)
+static INLINE __m128i half_btf_sse4_1(__m128i w0, __m128i n0, __m128i w1,
+                                      __m128i n1, __m128i rounding, int bit) {
+  __m128i x, y;
+
+  x = _mm_mullo_epi32(w0, n0);
+  y = _mm_mullo_epi32(w1, n1);
+  x = _mm_add_epi32(x, y);
+  x = _mm_add_epi32(x, rounding);
+  x = _mm_srai_epi32(x, bit);
+  return x;
+}
+
+#endif  // _HIGHBD_TXFM_UTILITY_SSE4_H
diff --git a/third_party/aom/av1/common/x86/highbd_warp_plane_ssse3.c b/third_party/aom/av1/common/x86/highbd_warp_plane_ssse3.c
new file mode 100644
index 0000000000..c25db88b70
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_warp_plane_ssse3.c
@@ -0,0 +1,286 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <tmmintrin.h>
+
+#include "./av1_rtcd.h"
+#include "av1/common/warped_motion.h"
+
+static const __m128i *const filter = (const __m128i *const)warped_filter;
+
+/* SSE2 version of the rotzoom/affine warp filter */
+void av1_highbd_warp_affine_ssse3(int32_t *mat, uint16_t *ref, int width,
+                                  int height, int stride, uint16_t *pred,
+                                  int p_col, int p_row, int p_width,
+                                  int p_height, int p_stride, int subsampling_x,
+                                  int subsampling_y, int bd, int ref_frm,
+                                  int16_t alpha, int16_t beta, int16_t gamma,
+                                  int16_t delta) {
+#if HORSHEAR_REDUCE_PREC_BITS >= 5
+  __m128i tmp[15];
+#else
+#error "HORSHEAR_REDUCE_PREC_BITS < 5 not currently supported by SSSE3 filter"
+#endif
+  int i, j, k;
+
+  /* Note: For this code to work, the left/right frame borders need to be
+     extended by at least 13 pixels each. By the time we get here, other
+     code will have set up this border, but we allow an explicit check
+     for debugging purposes.
+  */
+  /*for (i = 0; i < height; ++i) {
+    for (j = 0; j < 13; ++j) {
+      assert(ref[i * stride - 13 + j] == ref[i * stride]);
+      assert(ref[i * stride + width + j] == ref[i * stride + (width - 1)]);
+    }
+  }*/
+
+  for (i = 0; i < p_height; i += 8) {
+    for (j = 0; j < p_width; j += 8) {
+      // (x, y) coordinates of the center of this block in the destination
+      // image
+      int32_t dst_x = p_col + j + 4;
+      int32_t dst_y = p_row + i + 4;
+
+      int32_t x4, y4, ix4, sx4, iy4, sy4;
+      if (subsampling_x)
+        x4 = ROUND_POWER_OF_TWO_SIGNED(
+            mat[2] * 2 * dst_x + mat[3] * 2 * dst_y + mat[0] +
+                (mat[2] + mat[3] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+            1);
+      else
+        x4 = mat[2] * dst_x + mat[3] * dst_y + mat[0];
+
+      if (subsampling_y)
+        y4 = ROUND_POWER_OF_TWO_SIGNED(
+            mat[4] * 2 * dst_x + mat[5] * 2 * dst_y + mat[1] +
+                (mat[4] + mat[5] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+            1);
+      else
+        y4 = mat[4] * dst_x + mat[5] * dst_y + mat[1];
+
+      ix4 = x4 >> WARPEDMODEL_PREC_BITS;
+      sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+      iy4 = y4 >> WARPEDMODEL_PREC_BITS;
+      sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+
+      // Horizontal filter
+      for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+        int iy = iy4 + k;
+        if (iy < 0)
+          iy = 0;
+        else if (iy > height - 1)
+          iy = height - 1;
+
+        // If the block is aligned such that, after clamping, every sample
+        // would be taken from the leftmost/rightmost column, then we can
+        // skip the expensive horizontal filter.
+        if (ix4 <= -7) {
+          tmp[k + 7] = _mm_set1_epi16(
+              ref[iy * stride] *
+              (1 << (WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS)));
+        } else if (ix4 >= width + 6) {
+          tmp[k + 7] = _mm_set1_epi16(
+              ref[iy * stride + (width - 1)] *
+              (1 << (WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS)));
+        } else {
+          int sx = sx4 + alpha * (-4) + beta * k +
+                   // Include rounding and offset here
+                   (1 << (WARPEDDIFF_PREC_BITS - 1)) +
+                   (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
+
+          // Load source pixels
+          __m128i src =
+              _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+          __m128i src2 =
+              _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 + 1));
+
+          // Filter even-index pixels
+          __m128i tmp_0 = filter[(sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS];
+          __m128i tmp_2 = filter[(sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS];
+          __m128i tmp_4 = filter[(sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS];
+          __m128i tmp_6 = filter[(sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS];
+
+          // coeffs 0 1 0 1 2 3 2 3 for pixels 0, 2
+          __m128i tmp_8 = _mm_unpacklo_epi32(tmp_0, tmp_2);
+          // coeffs 0 1 0 1 2 3 2 3 for pixels 4, 6
+          __m128i tmp_10 = _mm_unpacklo_epi32(tmp_4, tmp_6);
+          // coeffs 4 5 4 5 6 7 6 7 for pixels 0, 2
+          __m128i tmp_12 = _mm_unpackhi_epi32(tmp_0, tmp_2);
+          // coeffs 4 5 4 5 6 7 6 7 for pixels 4, 6
+          __m128i tmp_14 = _mm_unpackhi_epi32(tmp_4, tmp_6);
+
+          // coeffs 0 1 0 1 0 1 0 1 for pixels 0, 2, 4, 6
+          __m128i coeff_0 = _mm_unpacklo_epi64(tmp_8, tmp_10);
+          // coeffs 2 3 2 3 2 3 2 3 for pixels 0, 2, 4, 6
+          __m128i coeff_2 = _mm_unpackhi_epi64(tmp_8, tmp_10);
+          // coeffs 4 5 4 5 4 5 4 5 for pixels 0, 2, 4, 6
+          __m128i coeff_4 = _mm_unpacklo_epi64(tmp_12, tmp_14);
+          // coeffs 6 7 6 7 6 7 6 7 for pixels 0, 2, 4, 6
+          __m128i coeff_6 = _mm_unpackhi_epi64(tmp_12, tmp_14);
+
+          __m128i round_const =
+              _mm_set1_epi32((1 << HORSHEAR_REDUCE_PREC_BITS) >> 1);
+
+          // Calculate filtered results
+          __m128i res_0 = _mm_madd_epi16(src, coeff_0);
+          __m128i res_2 =
+              _mm_madd_epi16(_mm_alignr_epi8(src2, src, 4), coeff_2);
+          __m128i res_4 =
+              _mm_madd_epi16(_mm_alignr_epi8(src2, src, 8), coeff_4);
+          __m128i res_6 =
+              _mm_madd_epi16(_mm_alignr_epi8(src2, src, 12), coeff_6);
+
+          __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
+                                           _mm_add_epi32(res_2, res_6));
+          res_even = _mm_srai_epi32(_mm_add_epi32(res_even, round_const),
+                                    HORSHEAR_REDUCE_PREC_BITS);
+
+          // Filter odd-index pixels
+          __m128i tmp_1 = filter[(sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS];
+          __m128i tmp_3 = filter[(sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS];
+          __m128i tmp_5 = filter[(sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS];
+          __m128i tmp_7 = filter[(sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS];
+
+          __m128i tmp_9 = _mm_unpacklo_epi32(tmp_1, tmp_3);
+          __m128i tmp_11 = _mm_unpacklo_epi32(tmp_5, tmp_7);
+          __m128i tmp_13 = _mm_unpackhi_epi32(tmp_1, tmp_3);
+          __m128i tmp_15 = _mm_unpackhi_epi32(tmp_5, tmp_7);
+
+          __m128i coeff_1 = _mm_unpacklo_epi64(tmp_9, tmp_11);
+          __m128i coeff_3 = _mm_unpackhi_epi64(tmp_9, tmp_11);
+          __m128i coeff_5 = _mm_unpacklo_epi64(tmp_13, tmp_15);
+          __m128i coeff_7 = _mm_unpackhi_epi64(tmp_13, tmp_15);
+
+          __m128i res_1 =
+              _mm_madd_epi16(_mm_alignr_epi8(src2, src, 2), coeff_1);
+          __m128i res_3 =
+              _mm_madd_epi16(_mm_alignr_epi8(src2, src, 6), coeff_3);
+          __m128i res_5 =
+              _mm_madd_epi16(_mm_alignr_epi8(src2, src, 10), coeff_5);
+          __m128i res_7 =
+              _mm_madd_epi16(_mm_alignr_epi8(src2, src, 14), coeff_7);
+
+          __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
+                                          _mm_add_epi32(res_3, res_7));
+          res_odd = _mm_srai_epi32(_mm_add_epi32(res_odd, round_const),
+                                   HORSHEAR_REDUCE_PREC_BITS);
+
+          // Combine results into one register.
+          // We store the columns in the order 0, 2, 4, 6, 1, 3, 5, 7
+          // as this order helps with the vertical filter.
+          tmp[k + 7] = _mm_packs_epi32(res_even, res_odd);
+        }
+      }
+
+      // Vertical filter
+      for (k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
+        int sy = sy4 + gamma * (-4) + delta * k +
+                 (1 << (WARPEDDIFF_PREC_BITS - 1)) +
+                 (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
+
+        // Load from tmp and rearrange pairs of consecutive rows into the
+        // column order 0 0 2 2 4 4 6 6; 1 1 3 3 5 5 7 7
+        __m128i *src = tmp + (k + 4);
+        __m128i src_0 = _mm_unpacklo_epi16(src[0], src[1]);
+        __m128i src_2 = _mm_unpacklo_epi16(src[2], src[3]);
+        __m128i src_4 = _mm_unpacklo_epi16(src[4], src[5]);
+        __m128i src_6 = _mm_unpacklo_epi16(src[6], src[7]);
+
+        // Filter even-index pixels
+        __m128i tmp_0 = filter[(sy + 0 * gamma) >> WARPEDDIFF_PREC_BITS];
+        __m128i tmp_2 = filter[(sy + 2 * gamma) >> WARPEDDIFF_PREC_BITS];
+        __m128i tmp_4 = filter[(sy + 4 * gamma) >> WARPEDDIFF_PREC_BITS];
+        __m128i tmp_6 = filter[(sy + 6 * gamma) >> WARPEDDIFF_PREC_BITS];
+
+        __m128i tmp_8 = _mm_unpacklo_epi32(tmp_0, tmp_2);
+        __m128i tmp_10 = _mm_unpacklo_epi32(tmp_4, tmp_6);
+        __m128i tmp_12 = _mm_unpackhi_epi32(tmp_0, tmp_2);
+        __m128i tmp_14 = _mm_unpackhi_epi32(tmp_4, tmp_6);
+
+        __m128i coeff_0 = _mm_unpacklo_epi64(tmp_8, tmp_10);
+        __m128i coeff_2 = _mm_unpackhi_epi64(tmp_8, tmp_10);
+        __m128i coeff_4 = _mm_unpacklo_epi64(tmp_12, tmp_14);
+        __m128i coeff_6 = _mm_unpackhi_epi64(tmp_12, tmp_14);
+
+        __m128i res_0 = _mm_madd_epi16(src_0, coeff_0);
+        __m128i res_2 = _mm_madd_epi16(src_2, coeff_2);
+        __m128i res_4 = _mm_madd_epi16(src_4, coeff_4);
+        __m128i res_6 = _mm_madd_epi16(src_6, coeff_6);
+
+        __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
+                                         _mm_add_epi32(res_4, res_6));
+
+        // Filter odd-index pixels
+        __m128i src_1 = _mm_unpackhi_epi16(src[0], src[1]);
+        __m128i src_3 = _mm_unpackhi_epi16(src[2], src[3]);
+        __m128i src_5 = _mm_unpackhi_epi16(src[4], src[5]);
+        __m128i src_7 = _mm_unpackhi_epi16(src[6], src[7]);
+
+        __m128i tmp_1 = filter[(sy + 1 * gamma) >> WARPEDDIFF_PREC_BITS];
+        __m128i tmp_3 = filter[(sy + 3 * gamma) >> WARPEDDIFF_PREC_BITS];
+        __m128i tmp_5 = filter[(sy + 5 * gamma) >> WARPEDDIFF_PREC_BITS];
+        __m128i tmp_7 = filter[(sy + 7 * gamma) >> WARPEDDIFF_PREC_BITS];
+
+        __m128i tmp_9 = _mm_unpacklo_epi32(tmp_1, tmp_3);
+        __m128i tmp_11 = _mm_unpacklo_epi32(tmp_5, tmp_7);
+        __m128i tmp_13 = _mm_unpackhi_epi32(tmp_1, tmp_3);
+        __m128i tmp_15 = _mm_unpackhi_epi32(tmp_5, tmp_7);
+
+        __m128i coeff_1 = _mm_unpacklo_epi64(tmp_9, tmp_11);
+        __m128i coeff_3 = _mm_unpackhi_epi64(tmp_9, tmp_11);
+        __m128i coeff_5 = _mm_unpacklo_epi64(tmp_13, tmp_15);
+        __m128i coeff_7 = _mm_unpackhi_epi64(tmp_13, tmp_15);
+
+        __m128i res_1 = _mm_madd_epi16(src_1, coeff_1);
+        __m128i res_3 = _mm_madd_epi16(src_3, coeff_3);
+        __m128i res_5 = _mm_madd_epi16(src_5, coeff_5);
+        __m128i res_7 = _mm_madd_epi16(src_7, coeff_7);
+
+        __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
+                                        _mm_add_epi32(res_5, res_7));
+
+        // Rearrange pixels back into the order 0 ... 7
+        __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+        __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+
+        // Round and pack into 8 bits
+        __m128i round_const =
+            _mm_set1_epi32((1 << VERSHEAR_REDUCE_PREC_BITS) >> 1);
+
+        __m128i res_lo_round = _mm_srai_epi32(
+            _mm_add_epi32(res_lo, round_const), VERSHEAR_REDUCE_PREC_BITS);
+        __m128i res_hi_round = _mm_srai_epi32(
+            _mm_add_epi32(res_hi, round_const), VERSHEAR_REDUCE_PREC_BITS);
+
+        __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round);
+        // Clamp res_16bit to the range [0, 2^bd - 1]
+        __m128i max_val = _mm_set1_epi16((1 << bd) - 1);
+        __m128i zero = _mm_setzero_si128();
+        res_16bit = _mm_max_epi16(_mm_min_epi16(res_16bit, max_val), zero);
+
+        // Store, blending with 'pred' if needed
+        __m128i *p = (__m128i *)&pred[(i + k + 4) * p_stride + j];
+
+        // Note: If we're outputting a 4x4 block, we need to be very careful
+        // to only output 4 pixels at this point, to avoid encode/decode
+        // mismatches when encoding with multiple threads.
+        if (p_width == 4) {
+          if (ref_frm) res_16bit = _mm_avg_epu16(res_16bit, _mm_loadl_epi64(p));
+          _mm_storel_epi64(p, res_16bit);
+        } else {
+          if (ref_frm) res_16bit = _mm_avg_epu16(res_16bit, _mm_loadu_si128(p));
+          _mm_storeu_si128(p, res_16bit);
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/hybrid_inv_txfm_avx2.c b/third_party/aom/av1/common/x86/hybrid_inv_txfm_avx2.c
new file mode 100644
index 0000000000..efc8d1e240
--- /dev/null
+++ b/third_party/aom/av1/common/x86/hybrid_inv_txfm_avx2.c
@@ -0,0 +1,507 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>  // avx2
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+
+#include "aom_dsp/x86/txfm_common_avx2.h"
+
+static INLINE void load_coeff(const tran_low_t *coeff, __m256i *in) {
+#if CONFIG_HIGHBITDEPTH
+  *in = _mm256_setr_epi16(
+      (int16_t)coeff[0], (int16_t)coeff[1], (int16_t)coeff[2],
+      (int16_t)coeff[3], (int16_t)coeff[4], (int16_t)coeff[5],
+      (int16_t)coeff[6], (int16_t)coeff[7], (int16_t)coeff[8],
+      (int16_t)coeff[9], (int16_t)coeff[10], (int16_t)coeff[11],
+      (int16_t)coeff[12], (int16_t)coeff[13], (int16_t)coeff[14],
+      (int16_t)coeff[15]);
+#else
+  *in = _mm256_loadu_si256((const __m256i *)coeff);
+#endif
+}
+
+static void load_buffer_16x16(const tran_low_t *coeff, __m256i *in) {
+  int i = 0;
+  while (i < 16) {
+    load_coeff(coeff + (i << 4), &in[i]);
+    i += 1;
+  }
+}
+
+static void recon_and_store(const __m256i *res, uint8_t *output) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i x = _mm_loadu_si128((__m128i const *)output);
+  __m128i p0 = _mm_unpacklo_epi8(x, zero);
+  __m128i p1 = _mm_unpackhi_epi8(x, zero);
+
+  p0 = _mm_add_epi16(p0, _mm256_castsi256_si128(*res));
+  p1 = _mm_add_epi16(p1, _mm256_extractf128_si256(*res, 1));
+  x = _mm_packus_epi16(p0, p1);
+  _mm_storeu_si128((__m128i *)output, x);
+}
+
+#define IDCT_ROUNDING_POS (6)
+
+static void write_buffer_16x16(__m256i *in, const int stride, uint8_t *output) {
+  const __m256i rounding = _mm256_set1_epi16(1 << (IDCT_ROUNDING_POS - 1));
+  int i = 0;
+
+  while (i < 16) {
+    in[i] = _mm256_add_epi16(in[i], rounding);
+    in[i] = _mm256_srai_epi16(in[i], IDCT_ROUNDING_POS);
+    recon_and_store(&in[i], output + i * stride);
+    i += 1;
+  }
+}
+
+static INLINE void unpack_butter_fly(const __m256i *a0, const __m256i *a1,
+                                     const __m256i *c0, const __m256i *c1,
+                                     __m256i *b0, __m256i *b1) {
+  __m256i x0, x1;
+  x0 = _mm256_unpacklo_epi16(*a0, *a1);
+  x1 = _mm256_unpackhi_epi16(*a0, *a1);
+  *b0 = butter_fly(x0, x1, *c0);
+  *b1 = butter_fly(x0, x1, *c1);
+}
+
+static void idct16_avx2(__m256i *in) {
+  const __m256i cospi_p30_m02 = pair256_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m256i cospi_p02_p30 = pair256_set_epi16(cospi_2_64, cospi_30_64);
+  const __m256i cospi_p14_m18 = pair256_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m256i cospi_p18_p14 = pair256_set_epi16(cospi_18_64, cospi_14_64);
+  const __m256i cospi_p22_m10 = pair256_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m256i cospi_p10_p22 = pair256_set_epi16(cospi_10_64, cospi_22_64);
+  const __m256i cospi_p06_m26 = pair256_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m256i cospi_p26_p06 = pair256_set_epi16(cospi_26_64, cospi_6_64);
+  const __m256i cospi_p28_m04 = pair256_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m256i cospi_p04_p28 = pair256_set_epi16(cospi_4_64, cospi_28_64);
+  const __m256i cospi_p12_m20 = pair256_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m256i cospi_p20_p12 = pair256_set_epi16(cospi_20_64, cospi_12_64);
+  const __m256i cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64);
+  const __m256i cospi_p16_m16 = pair256_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m256i cospi_p24_m08 = pair256_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m256i cospi_p08_p24 = pair256_set_epi16(cospi_8_64, cospi_24_64);
+  const __m256i cospi_m08_p24 = pair256_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m256i cospi_p24_p08 = pair256_set_epi16(cospi_24_64, cospi_8_64);
+  const __m256i cospi_m24_m08 = pair256_set_epi16(-cospi_24_64, -cospi_8_64);
+  __m256i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m256i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m256i t0, t1, t2, t3, t4, t5, t6, t7;
+
+  // stage 1, (0-7)
+  u0 = in[0];
+  u1 = in[8];
+  u2 = in[4];
+  u3 = in[12];
+  u4 = in[2];
+  u5 = in[10];
+  u6 = in[6];
+  u7 = in[14];
+
+  // stage 2, (0-7)
+  // stage 3, (0-7)
+  t0 = u0;
+  t1 = u1;
+  t2 = u2;
+  t3 = u3;
+  unpack_butter_fly(&u4, &u7, &cospi_p28_m04, &cospi_p04_p28, &t4, &t7);
+  unpack_butter_fly(&u5, &u6, &cospi_p12_m20, &cospi_p20_p12, &t5, &t6);
+
+  // stage 4, (0-7)
+  unpack_butter_fly(&t0, &t1, &cospi_p16_p16, &cospi_p16_m16, &u0, &u1);
+  unpack_butter_fly(&t2, &t3, &cospi_p24_m08, &cospi_p08_p24, &u2, &u3);
+  u4 = _mm256_add_epi16(t4, t5);
+  u5 = _mm256_sub_epi16(t4, t5);
+  u6 = _mm256_sub_epi16(t7, t6);
+  u7 = _mm256_add_epi16(t7, t6);
+
+  // stage 5, (0-7)
+  t0 = _mm256_add_epi16(u0, u3);
+  t1 = _mm256_add_epi16(u1, u2);
+  t2 = _mm256_sub_epi16(u1, u2);
+  t3 = _mm256_sub_epi16(u0, u3);
+  t4 = u4;
+  t7 = u7;
+  unpack_butter_fly(&u6, &u5, &cospi_p16_m16, &cospi_p16_p16, &t5, &t6);
+
+  // stage 6, (0-7)
+  u0 = _mm256_add_epi16(t0, t7);
+  u1 = _mm256_add_epi16(t1, t6);
+  u2 = _mm256_add_epi16(t2, t5);
+  u3 = _mm256_add_epi16(t3, t4);
+  u4 = _mm256_sub_epi16(t3, t4);
+  u5 = _mm256_sub_epi16(t2, t5);
+  u6 = _mm256_sub_epi16(t1, t6);
+  u7 = _mm256_sub_epi16(t0, t7);
+
+  // stage 1, (8-15)
+  v0 = in[1];
+  v1 = in[9];
+  v2 = in[5];
+  v3 = in[13];
+  v4 = in[3];
+  v5 = in[11];
+  v6 = in[7];
+  v7 = in[15];
+
+  // stage 2, (8-15)
+  unpack_butter_fly(&v0, &v7, &cospi_p30_m02, &cospi_p02_p30, &t0, &t7);
+  unpack_butter_fly(&v1, &v6, &cospi_p14_m18, &cospi_p18_p14, &t1, &t6);
+  unpack_butter_fly(&v2, &v5, &cospi_p22_m10, &cospi_p10_p22, &t2, &t5);
+  unpack_butter_fly(&v3, &v4, &cospi_p06_m26, &cospi_p26_p06, &t3, &t4);
+
+  // stage 3, (8-15)
+  v0 = _mm256_add_epi16(t0, t1);
+  v1 = _mm256_sub_epi16(t0, t1);
+  v2 = _mm256_sub_epi16(t3, t2);
+  v3 = _mm256_add_epi16(t2, t3);
+  v4 = _mm256_add_epi16(t4, t5);
+  v5 = _mm256_sub_epi16(t4, t5);
+  v6 = _mm256_sub_epi16(t7, t6);
+  v7 = _mm256_add_epi16(t6, t7);
+
+  // stage 4, (8-15)
+  t0 = v0;
+  t7 = v7;
+  t3 = v3;
+  t4 = v4;
+  unpack_butter_fly(&v1, &v6, &cospi_m08_p24, &cospi_p24_p08, &t1, &t6);
+  unpack_butter_fly(&v2, &v5, &cospi_m24_m08, &cospi_m08_p24, &t2, &t5);
+
+  // stage 5, (8-15)
+  v0 = _mm256_add_epi16(t0, t3);
+  v1 = _mm256_add_epi16(t1, t2);
+  v2 = _mm256_sub_epi16(t1, t2);
+  v3 = _mm256_sub_epi16(t0, t3);
+  v4 = _mm256_sub_epi16(t7, t4);
+  v5 = _mm256_sub_epi16(t6, t5);
+  v6 = _mm256_add_epi16(t6, t5);
+  v7 = _mm256_add_epi16(t7, t4);
+
+  // stage 6, (8-15)
+  t0 = v0;
+  t1 = v1;
+  t6 = v6;
+  t7 = v7;
+  unpack_butter_fly(&v5, &v2, &cospi_p16_m16, &cospi_p16_p16, &t2, &t5);
+  unpack_butter_fly(&v4, &v3, &cospi_p16_m16, &cospi_p16_p16, &t3, &t4);
+
+  // stage 7
+  in[0] = _mm256_add_epi16(u0, t7);
+  in[1] = _mm256_add_epi16(u1, t6);
+  in[2] = _mm256_add_epi16(u2, t5);
+  in[3] = _mm256_add_epi16(u3, t4);
+  in[4] = _mm256_add_epi16(u4, t3);
+  in[5] = _mm256_add_epi16(u5, t2);
+  in[6] = _mm256_add_epi16(u6, t1);
+  in[7] = _mm256_add_epi16(u7, t0);
+  in[8] = _mm256_sub_epi16(u7, t0);
+  in[9] = _mm256_sub_epi16(u6, t1);
+  in[10] = _mm256_sub_epi16(u5, t2);
+  in[11] = _mm256_sub_epi16(u4, t3);
+  in[12] = _mm256_sub_epi16(u3, t4);
+  in[13] = _mm256_sub_epi16(u2, t5);
+  in[14] = _mm256_sub_epi16(u1, t6);
+  in[15] = _mm256_sub_epi16(u0, t7);
+}
+
+static void idct16(__m256i *in) {
+  mm256_transpose_16x16(in);
+  idct16_avx2(in);
+}
+
+static INLINE void butterfly_32b(const __m256i *a0, const __m256i *a1,
+                                 const __m256i *c0, const __m256i *c1,
+                                 __m256i *b) {
+  __m256i x0, x1;
+  x0 = _mm256_unpacklo_epi16(*a0, *a1);
+  x1 = _mm256_unpackhi_epi16(*a0, *a1);
+  b[0] = _mm256_madd_epi16(x0, *c0);
+  b[1] = _mm256_madd_epi16(x1, *c0);
+  b[2] = _mm256_madd_epi16(x0, *c1);
+  b[3] = _mm256_madd_epi16(x1, *c1);
+}
+
+static INLINE void group_rounding(__m256i *a, int num) {
+  const __m256i dct_rounding = _mm256_set1_epi32(DCT_CONST_ROUNDING);
+  int i;
+  for (i = 0; i < num; ++i) {
+    a[i] = _mm256_add_epi32(a[i], dct_rounding);
+    a[i] = _mm256_srai_epi32(a[i], DCT_CONST_BITS);
+  }
+}
+
+static INLINE void add_rnd(const __m256i *a, const __m256i *b, __m256i *out) {
+  __m256i x[4];
+  x[0] = _mm256_add_epi32(a[0], b[0]);
+  x[1] = _mm256_add_epi32(a[1], b[1]);
+  x[2] = _mm256_add_epi32(a[2], b[2]);
+  x[3] = _mm256_add_epi32(a[3], b[3]);
+
+  group_rounding(x, 4);
+
+  out[0] = _mm256_packs_epi32(x[0], x[1]);
+  out[1] = _mm256_packs_epi32(x[2], x[3]);
+}
+
+static INLINE void sub_rnd(const __m256i *a, const __m256i *b, __m256i *out) {
+  __m256i x[4];
+  x[0] = _mm256_sub_epi32(a[0], b[0]);
+  x[1] = _mm256_sub_epi32(a[1], b[1]);
+  x[2] = _mm256_sub_epi32(a[2], b[2]);
+  x[3] = _mm256_sub_epi32(a[3], b[3]);
+
+  group_rounding(x, 4);
+
+  out[0] = _mm256_packs_epi32(x[0], x[1]);
+  out[1] = _mm256_packs_epi32(x[2], x[3]);
+}
+
+static INLINE void butterfly_rnd(__m256i *a, __m256i *out) {
+  group_rounding(a, 4);
+  out[0] = _mm256_packs_epi32(a[0], a[1]);
+  out[1] = _mm256_packs_epi32(a[2], a[3]);
+}
+
+static void iadst16_avx2(__m256i *in) {
+  const __m256i cospi_p01_p31 = pair256_set_epi16(cospi_1_64, cospi_31_64);
+  const __m256i cospi_p31_m01 = pair256_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m256i cospi_p05_p27 = pair256_set_epi16(cospi_5_64, cospi_27_64);
+  const __m256i cospi_p27_m05 = pair256_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m256i cospi_p09_p23 = pair256_set_epi16(cospi_9_64, cospi_23_64);
+  const __m256i cospi_p23_m09 = pair256_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m256i cospi_p13_p19 = pair256_set_epi16(cospi_13_64, cospi_19_64);
+  const __m256i cospi_p19_m13 = pair256_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m256i cospi_p17_p15 = pair256_set_epi16(cospi_17_64, cospi_15_64);
+  const __m256i cospi_p15_m17 = pair256_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m256i cospi_p21_p11 = pair256_set_epi16(cospi_21_64, cospi_11_64);
+  const __m256i cospi_p11_m21 = pair256_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m256i cospi_p25_p07 = pair256_set_epi16(cospi_25_64, cospi_7_64);
+  const __m256i cospi_p07_m25 = pair256_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m256i cospi_p29_p03 = pair256_set_epi16(cospi_29_64, cospi_3_64);
+  const __m256i cospi_p03_m29 = pair256_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m256i cospi_p04_p28 = pair256_set_epi16(cospi_4_64, cospi_28_64);
+  const __m256i cospi_p28_m04 = pair256_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m256i cospi_p20_p12 = pair256_set_epi16(cospi_20_64, cospi_12_64);
+  const __m256i cospi_p12_m20 = pair256_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m256i cospi_m28_p04 = pair256_set_epi16(-cospi_28_64, cospi_4_64);
+  const __m256i cospi_m12_p20 = pair256_set_epi16(-cospi_12_64, cospi_20_64);
+  const __m256i cospi_p08_p24 = pair256_set_epi16(cospi_8_64, cospi_24_64);
+  const __m256i cospi_p24_m08 = pair256_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m256i cospi_m24_p08 = pair256_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m256i cospi_m16_m16 = _mm256_set1_epi16((int16_t)-cospi_16_64);
+  const __m256i cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64);
+  const __m256i cospi_p16_m16 = pair256_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m256i cospi_m16_p16 = pair256_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i x[16], s[16];
+  __m256i u[4], v[4];
+
+  // stage 1
+  butterfly_32b(&in[15], &in[0], &cospi_p01_p31, &cospi_p31_m01, u);
+  butterfly_32b(&in[7], &in[8], &cospi_p17_p15, &cospi_p15_m17, v);
+  add_rnd(u, v, &x[0]);
+  sub_rnd(u, v, &x[8]);
+
+  butterfly_32b(&in[13], &in[2], &cospi_p05_p27, &cospi_p27_m05, u);
+  butterfly_32b(&in[5], &in[10], &cospi_p21_p11, &cospi_p11_m21, v);
+  add_rnd(u, v, &x[2]);
+  sub_rnd(u, v, &x[10]);
+
+  butterfly_32b(&in[11], &in[4], &cospi_p09_p23, &cospi_p23_m09, u);
+  butterfly_32b(&in[3], &in[12], &cospi_p25_p07, &cospi_p07_m25, v);
+  add_rnd(u, v, &x[4]);
+  sub_rnd(u, v, &x[12]);
+
+  butterfly_32b(&in[9], &in[6], &cospi_p13_p19, &cospi_p19_m13, u);
+  butterfly_32b(&in[1], &in[14], &cospi_p29_p03, &cospi_p03_m29, v);
+  add_rnd(u, v, &x[6]);
+  sub_rnd(u, v, &x[14]);
+
+  // stage 2
+  s[0] = _mm256_add_epi16(x[0], x[4]);
+  s[1] = _mm256_add_epi16(x[1], x[5]);
+  s[2] = _mm256_add_epi16(x[2], x[6]);
+  s[3] = _mm256_add_epi16(x[3], x[7]);
+  s[4] = _mm256_sub_epi16(x[0], x[4]);
+  s[5] = _mm256_sub_epi16(x[1], x[5]);
+  s[6] = _mm256_sub_epi16(x[2], x[6]);
+  s[7] = _mm256_sub_epi16(x[3], x[7]);
+  butterfly_32b(&x[8], &x[9], &cospi_p04_p28, &cospi_p28_m04, u);
+  butterfly_32b(&x[12], &x[13], &cospi_m28_p04, &cospi_p04_p28, v);
+  add_rnd(u, v, &s[8]);
+  sub_rnd(u, v, &s[12]);
+
+  butterfly_32b(&x[10], &x[11], &cospi_p20_p12, &cospi_p12_m20, u);
+  butterfly_32b(&x[14], &x[15], &cospi_m12_p20, &cospi_p20_p12, v);
+  add_rnd(u, v, &s[10]);
+  sub_rnd(u, v, &s[14]);
+
+  // stage 3
+  x[0] = _mm256_add_epi16(s[0], s[2]);
+  x[1] = _mm256_add_epi16(s[1], s[3]);
+  x[2] = _mm256_sub_epi16(s[0], s[2]);
+  x[3] = _mm256_sub_epi16(s[1], s[3]);
+
+  x[8] = _mm256_add_epi16(s[8], s[10]);
+  x[9] = _mm256_add_epi16(s[9], s[11]);
+  x[10] = _mm256_sub_epi16(s[8], s[10]);
+  x[11] = _mm256_sub_epi16(s[9], s[11]);
+
+  butterfly_32b(&s[4], &s[5], &cospi_p08_p24, &cospi_p24_m08, u);
+  butterfly_32b(&s[6], &s[7], &cospi_m24_p08, &cospi_p08_p24, v);
+  add_rnd(u, v, &x[4]);
+  sub_rnd(u, v, &x[6]);
+
+  butterfly_32b(&s[12], &s[13], &cospi_p08_p24, &cospi_p24_m08, u);
+  butterfly_32b(&s[14], &s[15], &cospi_m24_p08, &cospi_p08_p24, v);
+  add_rnd(u, v, &x[12]);
+  sub_rnd(u, v, &x[14]);
+
+  // stage 4
+  butterfly_32b(&x[2], &x[3], &cospi_m16_m16, &cospi_p16_m16, u);
+  butterfly_32b(&x[6], &x[7], &cospi_p16_p16, &cospi_m16_p16, v);
+  butterfly_rnd(u, &x[2]);
+  butterfly_rnd(v, &x[6]);
+
+  butterfly_32b(&x[10], &x[11], &cospi_p16_p16, &cospi_m16_p16, u);
+  butterfly_32b(&x[14], &x[15], &cospi_m16_m16, &cospi_p16_m16, v);
+  butterfly_rnd(u, &x[10]);
+  butterfly_rnd(v, &x[14]);
+
+  in[0] = x[0];
+  in[1] = _mm256_sub_epi16(zero, x[8]);
+  in[2] = x[12];
+  in[3] = _mm256_sub_epi16(zero, x[4]);
+  in[4] = x[6];
+  in[5] = x[14];
+  in[6] = x[10];
+  in[7] = x[2];
+  in[8] = x[3];
+  in[9] = x[11];
+  in[10] = x[15];
+  in[11] = x[7];
+  in[12] = x[5];
+  in[13] = _mm256_sub_epi16(zero, x[13]);
+  in[14] = x[9];
+  in[15] = _mm256_sub_epi16(zero, x[1]);
+}
+
+static void iadst16(__m256i *in) {
+  mm256_transpose_16x16(in);
+  iadst16_avx2(in);
+}
+
+#if CONFIG_EXT_TX
+static void flip_row(__m256i *in, int rows) {
+  int i;
+  for (i = 0; i < rows; ++i) {
+    mm256_reverse_epi16(&in[i]);
+  }
+}
+
+static void flip_col(uint8_t **dest, int *stride, int rows) {
+  *dest = *dest + (rows - 1) * (*stride);
+  *stride = -*stride;
+}
+
+static void iidtx16(__m256i *in) {
+  mm256_transpose_16x16(in);
+  txfm_scaling16_avx2(Sqrt2, in);
+}
+#endif
+
+void av1_iht16x16_256_add_avx2(const tran_low_t *input, uint8_t *dest,
+                               int stride, int tx_type) {
+  __m256i in[16];
+
+  load_buffer_16x16(input, in);
+  switch (tx_type) {
+    case DCT_DCT:
+      idct16(in);
+      idct16(in);
+      break;
+    case ADST_DCT:
+      idct16(in);
+      iadst16(in);
+      break;
+    case DCT_ADST:
+      iadst16(in);
+      idct16(in);
+      break;
+    case ADST_ADST:
+      iadst16(in);
+      iadst16(in);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      idct16(in);
+      iadst16(in);
+      flip_col(&dest, &stride, 16);
+      break;
+    case DCT_FLIPADST:
+      iadst16(in);
+      idct16(in);
+      flip_row(in, 16);
+      break;
+    case FLIPADST_FLIPADST:
+      iadst16(in);
+      iadst16(in);
+      flip_row(in, 16);
+      flip_col(&dest, &stride, 16);
+      break;
+    case ADST_FLIPADST:
+      iadst16(in);
+      iadst16(in);
+      flip_row(in, 16);
+      break;
+    case FLIPADST_ADST:
+      iadst16(in);
+      iadst16(in);
+      flip_col(&dest, &stride, 16);
+      break;
+    case IDTX:
+      iidtx16(in);
+      iidtx16(in);
+      break;
+    case V_DCT:
+      iidtx16(in);
+      idct16(in);
+      break;
+    case H_DCT:
+      idct16(in);
+      iidtx16(in);
+      break;
+    case V_ADST:
+      iidtx16(in);
+      iadst16(in);
+      break;
+    case H_ADST:
+      iadst16(in);
+      iidtx16(in);
+      break;
+    case V_FLIPADST:
+      iidtx16(in);
+      iadst16(in);
+      flip_col(&dest, &stride, 16);
+      break;
+    case H_FLIPADST:
+      iadst16(in);
+      iidtx16(in);
+      flip_row(in, 16);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+  write_buffer_16x16(in, stride, dest);
+}
diff --git a/third_party/aom/av1/common/x86/idct_intrin_sse2.c b/third_party/aom/av1/common/x86/idct_intrin_sse2.c
new file mode 100644
index 0000000000..522e8988c3
--- /dev/null
+++ b/third_party/aom/av1/common/x86/idct_intrin_sse2.c
@@ -0,0 +1,1402 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "aom_dsp/x86/inv_txfm_sse2.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+#include "aom_ports/mem.h"
+#include "av1/common/enums.h"
+
+#if CONFIG_EXT_TX
+static INLINE void fliplr_4x4(__m128i in[2]) {
+  in[0] = _mm_shufflelo_epi16(in[0], 0x1b);
+  in[0] = _mm_shufflehi_epi16(in[0], 0x1b);
+  in[1] = _mm_shufflelo_epi16(in[1], 0x1b);
+  in[1] = _mm_shufflehi_epi16(in[1], 0x1b);
+}
+
+static INLINE void fliplr_8x8(__m128i in[8]) {
+  in[0] = mm_reverse_epi16(in[0]);
+  in[1] = mm_reverse_epi16(in[1]);
+  in[2] = mm_reverse_epi16(in[2]);
+  in[3] = mm_reverse_epi16(in[3]);
+
+  in[4] = mm_reverse_epi16(in[4]);
+  in[5] = mm_reverse_epi16(in[5]);
+  in[6] = mm_reverse_epi16(in[6]);
+  in[7] = mm_reverse_epi16(in[7]);
+}
+
+static INLINE void fliplr_16x8(__m128i in[16]) {
+  fliplr_8x8(&in[0]);
+  fliplr_8x8(&in[8]);
+}
+
+#define FLIPLR_16x16(in0, in1) \
+  do {                         \
+    __m128i *tmp;              \
+    fliplr_16x8(in0);          \
+    fliplr_16x8(in1);          \
+    tmp = (in0);               \
+    (in0) = (in1);             \
+    (in1) = tmp;               \
+  } while (0)
+
+#define FLIPUD_PTR(dest, stride, size)       \
+  do {                                       \
+    (dest) = (dest) + ((size)-1) * (stride); \
+    (stride) = -(stride);                    \
+  } while (0)
+#endif
+
+void av1_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  __m128i in[2];
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i eight = _mm_set1_epi16(8);
+
+  in[0] = load_input_data(input);
+  in[1] = load_input_data(input + 8);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      aom_idct4_sse2(in);
+      aom_idct4_sse2(in);
+      break;
+    case ADST_DCT:
+      aom_idct4_sse2(in);
+      aom_iadst4_sse2(in);
+      break;
+    case DCT_ADST:
+      aom_iadst4_sse2(in);
+      aom_idct4_sse2(in);
+      break;
+    case ADST_ADST:
+      aom_iadst4_sse2(in);
+      aom_iadst4_sse2(in);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      aom_idct4_sse2(in);
+      aom_iadst4_sse2(in);
+      FLIPUD_PTR(dest, stride, 4);
+      break;
+    case DCT_FLIPADST:
+      aom_iadst4_sse2(in);
+      aom_idct4_sse2(in);
+      fliplr_4x4(in);
+      break;
+    case FLIPADST_FLIPADST:
+      aom_iadst4_sse2(in);
+      aom_iadst4_sse2(in);
+      FLIPUD_PTR(dest, stride, 4);
+      fliplr_4x4(in);
+      break;
+    case ADST_FLIPADST:
+      aom_iadst4_sse2(in);
+      aom_iadst4_sse2(in);
+      fliplr_4x4(in);
+      break;
+    case FLIPADST_ADST:
+      aom_iadst4_sse2(in);
+      aom_iadst4_sse2(in);
+      FLIPUD_PTR(dest, stride, 4);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+
+  // Final round and shift
+  in[0] = _mm_add_epi16(in[0], eight);
+  in[1] = _mm_add_epi16(in[1], eight);
+
+  in[0] = _mm_srai_epi16(in[0], 4);
+  in[1] = _mm_srai_epi16(in[1], 4);
+
+  // Reconstruction and Store
+  {
+    __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 0));
+    __m128i d1 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 1));
+    __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2));
+    __m128i d3 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 3));
+    d0 = _mm_unpacklo_epi32(d0, d1);
+    d2 = _mm_unpacklo_epi32(d2, d3);
+    d0 = _mm_unpacklo_epi8(d0, zero);
+    d2 = _mm_unpacklo_epi8(d2, zero);
+    d0 = _mm_add_epi16(d0, in[0]);
+    d2 = _mm_add_epi16(d2, in[1]);
+    d0 = _mm_packus_epi16(d0, d2);
+    // store result[0]
+    *(int *)dest = _mm_cvtsi128_si32(d0);
+    // store result[1]
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride) = _mm_cvtsi128_si32(d0);
+    // store result[2]
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0);
+    // store result[3]
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0);
+  }
+}
+
+void av1_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  __m128i in[8];
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+
+  // load input data
+  in[0] = load_input_data(input);
+  in[1] = load_input_data(input + 8 * 1);
+  in[2] = load_input_data(input + 8 * 2);
+  in[3] = load_input_data(input + 8 * 3);
+  in[4] = load_input_data(input + 8 * 4);
+  in[5] = load_input_data(input + 8 * 5);
+  in[6] = load_input_data(input + 8 * 6);
+  in[7] = load_input_data(input + 8 * 7);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      aom_idct8_sse2(in);
+      aom_idct8_sse2(in);
+      break;
+    case ADST_DCT:
+      aom_idct8_sse2(in);
+      aom_iadst8_sse2(in);
+      break;
+    case DCT_ADST:
+      aom_iadst8_sse2(in);
+      aom_idct8_sse2(in);
+      break;
+    case ADST_ADST:
+      aom_iadst8_sse2(in);
+      aom_iadst8_sse2(in);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      aom_idct8_sse2(in);
+      aom_iadst8_sse2(in);
+      FLIPUD_PTR(dest, stride, 8);
+      break;
+    case DCT_FLIPADST:
+      aom_iadst8_sse2(in);
+      aom_idct8_sse2(in);
+      fliplr_8x8(in);
+      break;
+    case FLIPADST_FLIPADST:
+      aom_iadst8_sse2(in);
+      aom_iadst8_sse2(in);
+      FLIPUD_PTR(dest, stride, 8);
+      fliplr_8x8(in);
+      break;
+    case ADST_FLIPADST:
+      aom_iadst8_sse2(in);
+      aom_iadst8_sse2(in);
+      fliplr_8x8(in);
+      break;
+    case FLIPADST_ADST:
+      aom_iadst8_sse2(in);
+      aom_iadst8_sse2(in);
+      FLIPUD_PTR(dest, stride, 8);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+  in[4] = _mm_adds_epi16(in[4], final_rounding);
+  in[5] = _mm_adds_epi16(in[5], final_rounding);
+  in[6] = _mm_adds_epi16(in[6], final_rounding);
+  in[7] = _mm_adds_epi16(in[7], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 5);
+  in[1] = _mm_srai_epi16(in[1], 5);
+  in[2] = _mm_srai_epi16(in[2], 5);
+  in[3] = _mm_srai_epi16(in[3], 5);
+  in[4] = _mm_srai_epi16(in[4], 5);
+  in[5] = _mm_srai_epi16(in[5], 5);
+  in[6] = _mm_srai_epi16(in[6], 5);
+  in[7] = _mm_srai_epi16(in[7], 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in[0]);
+  RECON_AND_STORE(dest + 1 * stride, in[1]);
+  RECON_AND_STORE(dest + 2 * stride, in[2]);
+  RECON_AND_STORE(dest + 3 * stride, in[3]);
+  RECON_AND_STORE(dest + 4 * stride, in[4]);
+  RECON_AND_STORE(dest + 5 * stride, in[5]);
+  RECON_AND_STORE(dest + 6 * stride, in[6]);
+  RECON_AND_STORE(dest + 7 * stride, in[7]);
+}
+
+#if CONFIG_EXT_TX
+static void iidtx16_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  idtx16_8col(in0);
+  idtx16_8col(in1);
+}
+#endif  // CONFIG_EXT_TX
+
+void av1_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *dest,
+                               int stride, int tx_type) {
+  __m128i in[32];
+  __m128i *in0 = &in[0];
+  __m128i *in1 = &in[16];
+
+  load_buffer_8x16(input, in0);
+  input += 8;
+  load_buffer_8x16(input, in1);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      aom_idct16_sse2(in0, in1);
+      aom_idct16_sse2(in0, in1);
+      break;
+    case ADST_DCT:
+      aom_idct16_sse2(in0, in1);
+      aom_iadst16_sse2(in0, in1);
+      break;
+    case DCT_ADST:
+      aom_iadst16_sse2(in0, in1);
+      aom_idct16_sse2(in0, in1);
+      break;
+    case ADST_ADST:
+      aom_iadst16_sse2(in0, in1);
+      aom_iadst16_sse2(in0, in1);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      aom_idct16_sse2(in0, in1);
+      aom_iadst16_sse2(in0, in1);
+      FLIPUD_PTR(dest, stride, 16);
+      break;
+    case DCT_FLIPADST:
+      aom_iadst16_sse2(in0, in1);
+      aom_idct16_sse2(in0, in1);
+      FLIPLR_16x16(in0, in1);
+      break;
+    case FLIPADST_FLIPADST:
+      aom_iadst16_sse2(in0, in1);
+      aom_iadst16_sse2(in0, in1);
+      FLIPUD_PTR(dest, stride, 16);
+      FLIPLR_16x16(in0, in1);
+      break;
+    case ADST_FLIPADST:
+      aom_iadst16_sse2(in0, in1);
+      aom_iadst16_sse2(in0, in1);
+      FLIPLR_16x16(in0, in1);
+      break;
+    case FLIPADST_ADST:
+      aom_iadst16_sse2(in0, in1);
+      aom_iadst16_sse2(in0, in1);
+      FLIPUD_PTR(dest, stride, 16);
+      break;
+    case IDTX:
+      iidtx16_sse2(in0, in1);
+      iidtx16_sse2(in0, in1);
+      break;
+    case V_DCT:
+      iidtx16_sse2(in0, in1);
+      aom_idct16_sse2(in0, in1);
+      break;
+    case H_DCT:
+      aom_idct16_sse2(in0, in1);
+      iidtx16_sse2(in0, in1);
+      break;
+    case V_ADST:
+      iidtx16_sse2(in0, in1);
+      aom_iadst16_sse2(in0, in1);
+      break;
+    case H_ADST:
+      aom_iadst16_sse2(in0, in1);
+      iidtx16_sse2(in0, in1);
+      break;
+    case V_FLIPADST:
+      iidtx16_sse2(in0, in1);
+      aom_iadst16_sse2(in0, in1);
+      FLIPUD_PTR(dest, stride, 16);
+      break;
+    case H_FLIPADST:
+      aom_iadst16_sse2(in0, in1);
+      iidtx16_sse2(in0, in1);
+      FLIPLR_16x16(in0, in1);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+
+  write_buffer_8x16(dest, in0, stride);
+  dest += 8;
+  write_buffer_8x16(dest, in1, stride);
+}
+
+#if CONFIG_EXT_TX
+static void iidtx8_sse2(__m128i *in) {
+  in[0] = _mm_slli_epi16(in[0], 1);
+  in[1] = _mm_slli_epi16(in[1], 1);
+  in[2] = _mm_slli_epi16(in[2], 1);
+  in[3] = _mm_slli_epi16(in[3], 1);
+  in[4] = _mm_slli_epi16(in[4], 1);
+  in[5] = _mm_slli_epi16(in[5], 1);
+  in[6] = _mm_slli_epi16(in[6], 1);
+  in[7] = _mm_slli_epi16(in[7], 1);
+}
+
+static INLINE void iidtx4_sse2(__m128i *in) {
+  const __m128i v_scale_w = _mm_set1_epi16(Sqrt2);
+
+  const __m128i v_p0l_w = _mm_mullo_epi16(in[0], v_scale_w);
+  const __m128i v_p0h_w = _mm_mulhi_epi16(in[0], v_scale_w);
+  const __m128i v_p1l_w = _mm_mullo_epi16(in[1], v_scale_w);
+  const __m128i v_p1h_w = _mm_mulhi_epi16(in[1], v_scale_w);
+
+  const __m128i v_p0a_d = _mm_unpacklo_epi16(v_p0l_w, v_p0h_w);
+  const __m128i v_p0b_d = _mm_unpackhi_epi16(v_p0l_w, v_p0h_w);
+  const __m128i v_p1a_d = _mm_unpacklo_epi16(v_p1l_w, v_p1h_w);
+  const __m128i v_p1b_d = _mm_unpackhi_epi16(v_p1l_w, v_p1h_w);
+
+  in[0] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p0a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p0b_d, DCT_CONST_BITS));
+  in[1] = _mm_packs_epi32(xx_roundn_epi32_unsigned(v_p1a_d, DCT_CONST_BITS),
+                          xx_roundn_epi32_unsigned(v_p1b_d, DCT_CONST_BITS));
+}
+
+// load 8x8 array
+static INLINE void flip_buffer_lr_8x8(__m128i *in) {
+  in[0] = mm_reverse_epi16(in[0]);
+  in[1] = mm_reverse_epi16(in[1]);
+  in[2] = mm_reverse_epi16(in[2]);
+  in[3] = mm_reverse_epi16(in[3]);
+  in[4] = mm_reverse_epi16(in[4]);
+  in[5] = mm_reverse_epi16(in[5]);
+  in[6] = mm_reverse_epi16(in[6]);
+  in[7] = mm_reverse_epi16(in[7]);
+}
+#endif  // CONFIG_EXT_TX
+
+void av1_iht8x16_128_add_sse2(const tran_low_t *input, uint8_t *dest,
+                              int stride, int tx_type) {
+  __m128i in[16];
+
+  in[0] = load_input_data(input + 0 * 8);
+  in[1] = load_input_data(input + 1 * 8);
+  in[2] = load_input_data(input + 2 * 8);
+  in[3] = load_input_data(input + 3 * 8);
+  in[4] = load_input_data(input + 4 * 8);
+  in[5] = load_input_data(input + 5 * 8);
+  in[6] = load_input_data(input + 6 * 8);
+  in[7] = load_input_data(input + 7 * 8);
+
+  in[8] = load_input_data(input + 8 * 8);
+  in[9] = load_input_data(input + 9 * 8);
+  in[10] = load_input_data(input + 10 * 8);
+  in[11] = load_input_data(input + 11 * 8);
+  in[12] = load_input_data(input + 12 * 8);
+  in[13] = load_input_data(input + 13 * 8);
+  in[14] = load_input_data(input + 14 * 8);
+  in[15] = load_input_data(input + 15 * 8);
+
+  // Row transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case H_DCT:
+#endif
+      aom_idct8_sse2(in);
+      array_transpose_8x8(in, in);
+      aom_idct8_sse2(in + 8);
+      array_transpose_8x8(in + 8, in + 8);
+      break;
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case H_ADST:
+    case H_FLIPADST:
+#endif
+      aom_iadst8_sse2(in);
+      array_transpose_8x8(in, in);
+      aom_iadst8_sse2(in + 8);
+      array_transpose_8x8(in + 8, in + 8);
+      break;
+#if CONFIG_EXT_TX
+    case V_FLIPADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX:
+      iidtx8_sse2(in);
+      iidtx8_sse2(in + 8);
+      break;
+#endif
+    default: assert(0); break;
+  }
+  scale_sqrt2_8x8(in);
+  scale_sqrt2_8x8(in + 8);
+
+  // Column transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case DCT_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case V_DCT:
+#endif
+      idct16_8col(in);
+      break;
+    case ADST_DCT:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case FLIPADST_ADST:
+    case ADST_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case FLIPADST_DCT:
+    case V_ADST:
+    case V_FLIPADST:
+#endif
+      iadst16_8col(in);
+      break;
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case H_FLIPADST:
+    case IDTX: idtx16_8col(in); break;
+#endif
+    default: assert(0); break;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+#if CONFIG_EXT_TX
+    case H_DCT:
+#endif
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case H_ADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX:
+#endif
+      write_buffer_8x16(dest, in, stride);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case FLIPADST_ADST:
+    case V_FLIPADST: write_buffer_8x16(dest + stride * 15, in, -stride); break;
+    case DCT_FLIPADST:
+    case ADST_FLIPADST:
+    case H_FLIPADST:
+      flip_buffer_lr_8x8(in);
+      flip_buffer_lr_8x8(in + 8);
+      write_buffer_8x16(dest, in, stride);
+      break;
+    case FLIPADST_FLIPADST:
+      flip_buffer_lr_8x8(in);
+      flip_buffer_lr_8x8(in + 8);
+      write_buffer_8x16(dest + stride * 15, in, -stride);
+      break;
+#endif
+    default: assert(0); break;
+  }
+}
+
+static INLINE void write_buffer_8x8_round6(uint8_t *dest, __m128i *in,
+                                           int stride) {
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  const __m128i zero = _mm_setzero_si128();
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+  in[4] = _mm_adds_epi16(in[4], final_rounding);
+  in[5] = _mm_adds_epi16(in[5], final_rounding);
+  in[6] = _mm_adds_epi16(in[6], final_rounding);
+  in[7] = _mm_adds_epi16(in[7], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 6);
+  in[1] = _mm_srai_epi16(in[1], 6);
+  in[2] = _mm_srai_epi16(in[2], 6);
+  in[3] = _mm_srai_epi16(in[3], 6);
+  in[4] = _mm_srai_epi16(in[4], 6);
+  in[5] = _mm_srai_epi16(in[5], 6);
+  in[6] = _mm_srai_epi16(in[6], 6);
+  in[7] = _mm_srai_epi16(in[7], 6);
+
+  RECON_AND_STORE(dest + 0 * stride, in[0]);
+  RECON_AND_STORE(dest + 1 * stride, in[1]);
+  RECON_AND_STORE(dest + 2 * stride, in[2]);
+  RECON_AND_STORE(dest + 3 * stride, in[3]);
+  RECON_AND_STORE(dest + 4 * stride, in[4]);
+  RECON_AND_STORE(dest + 5 * stride, in[5]);
+  RECON_AND_STORE(dest + 6 * stride, in[6]);
+  RECON_AND_STORE(dest + 7 * stride, in[7]);
+}
+
+void av1_iht16x8_128_add_sse2(const tran_low_t *input, uint8_t *dest,
+                              int stride, int tx_type) {
+  __m128i in[16];
+
+  // Transpose 16x8 input into in[]
+  in[0] = load_input_data(input + 0 * 16);
+  in[1] = load_input_data(input + 1 * 16);
+  in[2] = load_input_data(input + 2 * 16);
+  in[3] = load_input_data(input + 3 * 16);
+  in[4] = load_input_data(input + 4 * 16);
+  in[5] = load_input_data(input + 5 * 16);
+  in[6] = load_input_data(input + 6 * 16);
+  in[7] = load_input_data(input + 7 * 16);
+  array_transpose_8x8(in, in);
+
+  in[8] = load_input_data(input + 8 + 0 * 16);
+  in[9] = load_input_data(input + 8 + 1 * 16);
+  in[10] = load_input_data(input + 8 + 2 * 16);
+  in[11] = load_input_data(input + 8 + 3 * 16);
+  in[12] = load_input_data(input + 8 + 4 * 16);
+  in[13] = load_input_data(input + 8 + 5 * 16);
+  in[14] = load_input_data(input + 8 + 6 * 16);
+  in[15] = load_input_data(input + 8 + 7 * 16);
+  array_transpose_8x8(in + 8, in + 8);
+
+  // Row transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case H_DCT:
+#endif
+      idct16_8col(in);
+      break;
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case H_ADST:
+    case H_FLIPADST:
+#endif
+      iadst16_8col(in);
+      break;
+#if CONFIG_EXT_TX
+    case V_FLIPADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX: idtx16_8col(in); break;
+#endif
+    default: assert(0); break;
+  }
+
+  // Scale
+  scale_sqrt2_8x8(in);
+  scale_sqrt2_8x8(in + 8);
+
+  // Column transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case DCT_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case V_DCT:
+#endif
+      aom_idct8_sse2(in);
+      aom_idct8_sse2(in + 8);
+      break;
+    case ADST_DCT:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case FLIPADST_ADST:
+    case ADST_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case FLIPADST_DCT:
+    case V_ADST:
+    case V_FLIPADST:
+#endif
+      aom_iadst8_sse2(in);
+      aom_iadst8_sse2(in + 8);
+      break;
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case H_FLIPADST:
+    case IDTX:
+      array_transpose_8x8(in, in);
+      array_transpose_8x8(in + 8, in + 8);
+      iidtx8_sse2(in);
+      iidtx8_sse2(in + 8);
+      break;
+#endif
+    default: assert(0); break;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX:
+#endif
+      write_buffer_8x8_round6(dest, in, stride);
+      write_buffer_8x8_round6(dest + 8, in + 8, stride);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case FLIPADST_ADST:
+    case V_FLIPADST:
+      write_buffer_8x8_round6(dest + stride * 7, in, -stride);
+      write_buffer_8x8_round6(dest + stride * 7 + 8, in + 8, -stride);
+      break;
+    case DCT_FLIPADST:
+    case ADST_FLIPADST:
+    case H_FLIPADST:
+      flip_buffer_lr_8x8(in);
+      flip_buffer_lr_8x8(in + 8);
+      write_buffer_8x8_round6(dest, in + 8, stride);
+      write_buffer_8x8_round6(dest + 8, in, stride);
+      break;
+    case FLIPADST_FLIPADST:
+      flip_buffer_lr_8x8(in);
+      flip_buffer_lr_8x8(in + 8);
+      write_buffer_8x8_round6(dest + stride * 7, in + 8, -stride);
+      write_buffer_8x8_round6(dest + stride * 7 + 8, in, -stride);
+      break;
+#endif
+    default: assert(0); break;
+  }
+}
+
+static INLINE void write_buffer_8x4_round5(uint8_t *dest, __m128i *in,
+                                           int stride) {
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i zero = _mm_setzero_si128();
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 5);
+  in[1] = _mm_srai_epi16(in[1], 5);
+  in[2] = _mm_srai_epi16(in[2], 5);
+  in[3] = _mm_srai_epi16(in[3], 5);
+
+  RECON_AND_STORE(dest + 0 * stride, in[0]);
+  RECON_AND_STORE(dest + 1 * stride, in[1]);
+  RECON_AND_STORE(dest + 2 * stride, in[2]);
+  RECON_AND_STORE(dest + 3 * stride, in[3]);
+}
+
+void av1_iht8x4_32_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  __m128i in[8];
+
+  in[0] = load_input_data(input + 0 * 8);
+  in[1] = load_input_data(input + 1 * 8);
+  in[2] = load_input_data(input + 2 * 8);
+  in[3] = load_input_data(input + 3 * 8);
+
+  // Row transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case H_DCT:
+#endif
+      aom_idct8_sse2(in);
+      break;
+    case DCT_ADST:
+    case ADST_ADST: aom_iadst8_sse2(in); break;
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case H_ADST:
+    case H_FLIPADST: aom_iadst8_sse2(in); break;
+    case V_FLIPADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX: iidtx8_sse2(in); array_transpose_8x8(in, in);
+#endif
+      break;
+    default: assert(0); break;
+  }
+
+  scale_sqrt2_8x8(in);
+
+  // Repack data. We pack into the bottom half of 'in'
+  // so that the next repacking stage can pack into the
+  // top half without overwriting anything
+  in[7] = _mm_unpacklo_epi64(in[6], in[7]);
+  in[6] = _mm_unpacklo_epi64(in[4], in[5]);
+  in[5] = _mm_unpacklo_epi64(in[2], in[3]);
+  in[4] = _mm_unpacklo_epi64(in[0], in[1]);
+
+  // Column transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case DCT_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case V_DCT:
+#endif
+      aom_idct4_sse2(in + 4);
+      aom_idct4_sse2(in + 6);
+      break;
+    case ADST_DCT:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case FLIPADST_ADST:
+    case ADST_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case FLIPADST_DCT:
+    case V_ADST:
+    case V_FLIPADST:
+#endif
+      aom_iadst4_sse2(in + 4);
+      aom_iadst4_sse2(in + 6);
+      break;
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case H_FLIPADST:
+    case IDTX:
+      iidtx4_sse2(in + 4);
+      array_transpose_4x4(in + 4);
+      iidtx4_sse2(in + 6);
+      array_transpose_4x4(in + 6);
+      break;
+#endif
+    default: assert(0); break;
+  }
+
+  // Repack data
+  in[0] = _mm_unpacklo_epi64(in[4], in[6]);
+  in[1] = _mm_unpackhi_epi64(in[4], in[6]);
+  in[2] = _mm_unpacklo_epi64(in[5], in[7]);
+  in[3] = _mm_unpackhi_epi64(in[5], in[7]);
+
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX: break;
+    case FLIPADST_DCT:
+    case FLIPADST_ADST:
+    case V_FLIPADST: FLIPUD_PTR(dest, stride, 4); break;
+    case DCT_FLIPADST:
+    case ADST_FLIPADST:
+    case H_FLIPADST:
+      in[0] = mm_reverse_epi16(in[0]);
+      in[1] = mm_reverse_epi16(in[1]);
+      in[2] = mm_reverse_epi16(in[2]);
+      in[3] = mm_reverse_epi16(in[3]);
+      break;
+    case FLIPADST_FLIPADST:
+      in[0] = mm_reverse_epi16(in[0]);
+      in[1] = mm_reverse_epi16(in[1]);
+      in[2] = mm_reverse_epi16(in[2]);
+      in[3] = mm_reverse_epi16(in[3]);
+      FLIPUD_PTR(dest, stride, 4);
+#endif
+      break;
+    default: assert(0); break;
+  }
+  write_buffer_8x4_round5(dest, in, stride);
+}
+
+static INLINE void write_buffer_4x8_round5(uint8_t *dest, __m128i *in,
+                                           int stride) {
+  const __m128i final_rounding = _mm_set1_epi16(1 << 4);
+  const __m128i zero = _mm_setzero_si128();
+  // Final rounding and shift
+  in[0] = _mm_adds_epi16(in[0], final_rounding);
+  in[1] = _mm_adds_epi16(in[1], final_rounding);
+  in[2] = _mm_adds_epi16(in[2], final_rounding);
+  in[3] = _mm_adds_epi16(in[3], final_rounding);
+
+  in[0] = _mm_srai_epi16(in[0], 5);
+  in[1] = _mm_srai_epi16(in[1], 5);
+  in[2] = _mm_srai_epi16(in[2], 5);
+  in[3] = _mm_srai_epi16(in[3], 5);
+
+  // Reconstruction and Store
+  {
+    __m128i d0 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 0));
+    __m128i d1 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 1));
+    __m128i d2 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 2));
+    __m128i d3 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 3));
+    __m128i d4 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 4));
+    __m128i d5 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 5));
+    __m128i d6 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 6));
+    __m128i d7 = _mm_cvtsi32_si128(*(const int *)(dest + stride * 7));
+
+    d0 = _mm_unpacklo_epi32(d0, d1);
+    d2 = _mm_unpacklo_epi32(d2, d3);
+    d4 = _mm_unpacklo_epi32(d4, d5);
+    d6 = _mm_unpacklo_epi32(d6, d7);
+    d0 = _mm_unpacklo_epi8(d0, zero);
+    d2 = _mm_unpacklo_epi8(d2, zero);
+    d4 = _mm_unpacklo_epi8(d4, zero);
+    d6 = _mm_unpacklo_epi8(d6, zero);
+    d0 = _mm_add_epi16(d0, in[0]);
+    d2 = _mm_add_epi16(d2, in[1]);
+    d4 = _mm_add_epi16(d4, in[2]);
+    d6 = _mm_add_epi16(d6, in[3]);
+
+    d0 = _mm_packus_epi16(d0, d2);
+    *(int *)dest = _mm_cvtsi128_si32(d0);
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride) = _mm_cvtsi128_si32(d0);
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 2) = _mm_cvtsi128_si32(d0);
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 3) = _mm_cvtsi128_si32(d0);
+    d0 = _mm_packus_epi16(d4, d6);
+    *(int *)(dest + stride * 4) = _mm_cvtsi128_si32(d0);
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 5) = _mm_cvtsi128_si32(d0);
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 6) = _mm_cvtsi128_si32(d0);
+    d0 = _mm_srli_si128(d0, 4);
+    *(int *)(dest + stride * 7) = _mm_cvtsi128_si32(d0);
+  }
+}
+
+void av1_iht4x8_32_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
+                            int tx_type) {
+  __m128i in[8];
+
+  // Load rows, packed two per element of 'in'.
+  // We pack into the bottom half of 'in' so that the
+  // later repacking stage can pack into the
+  // top half without overwriting anything
+  in[4] = load_input_data(input + 0 * 8);
+  in[5] = load_input_data(input + 1 * 8);
+  in[6] = load_input_data(input + 2 * 8);
+  in[7] = load_input_data(input + 3 * 8);
+
+  // Row transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case H_DCT:
+#endif
+      aom_idct4_sse2(in + 4);
+      aom_idct4_sse2(in + 6);
+      break;
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case H_ADST:
+    case H_FLIPADST:
+#endif
+      aom_iadst4_sse2(in + 4);
+      aom_iadst4_sse2(in + 6);
+      break;
+#if CONFIG_EXT_TX
+    case V_FLIPADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX:
+      iidtx4_sse2(in + 4);
+      array_transpose_4x4(in + 4);
+      iidtx4_sse2(in + 6);
+      array_transpose_4x4(in + 6);
+      break;
+#endif
+    default: assert(0); break;
+  }
+
+  scale_sqrt2_8x4(in + 4);
+
+  // Repack data
+  in[0] = _mm_unpacklo_epi64(in[4], in[6]);
+  in[1] = _mm_unpackhi_epi64(in[4], in[6]);
+  in[2] = _mm_unpacklo_epi64(in[5], in[7]);
+  in[3] = _mm_unpackhi_epi64(in[5], in[7]);
+
+  // Column transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case DCT_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case V_DCT:
+#endif
+      aom_idct8_sse2(in);
+      break;
+    case ADST_DCT:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case FLIPADST_ADST:
+    case ADST_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case FLIPADST_DCT:
+    case V_ADST:
+    case V_FLIPADST:
+#endif
+      aom_iadst8_sse2(in);
+      break;
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case H_FLIPADST:
+    case IDTX:
+      iidtx8_sse2(in);
+      array_transpose_8x8(in, in);
+      break;
+#endif
+    default: assert(0); break;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX:
+#endif
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case FLIPADST_ADST:
+    case V_FLIPADST: FLIPUD_PTR(dest, stride, 8); break;
+    case DCT_FLIPADST:
+    case ADST_FLIPADST:
+    case H_FLIPADST:
+      in[0] = _mm_shufflelo_epi16(in[0], 0x1b);
+      in[1] = _mm_shufflelo_epi16(in[1], 0x1b);
+      in[2] = _mm_shufflelo_epi16(in[2], 0x1b);
+      in[3] = _mm_shufflelo_epi16(in[3], 0x1b);
+      in[4] = _mm_shufflelo_epi16(in[4], 0x1b);
+      in[5] = _mm_shufflelo_epi16(in[5], 0x1b);
+      in[6] = _mm_shufflelo_epi16(in[6], 0x1b);
+      in[7] = _mm_shufflelo_epi16(in[7], 0x1b);
+      break;
+    case FLIPADST_FLIPADST:
+      in[0] = _mm_shufflelo_epi16(in[0], 0x1b);
+      in[1] = _mm_shufflelo_epi16(in[1], 0x1b);
+      in[2] = _mm_shufflelo_epi16(in[2], 0x1b);
+      in[3] = _mm_shufflelo_epi16(in[3], 0x1b);
+      in[4] = _mm_shufflelo_epi16(in[4], 0x1b);
+      in[5] = _mm_shufflelo_epi16(in[5], 0x1b);
+      in[6] = _mm_shufflelo_epi16(in[6], 0x1b);
+      in[7] = _mm_shufflelo_epi16(in[7], 0x1b);
+      FLIPUD_PTR(dest, stride, 8);
+      break;
+#endif
+    default: assert(0); break;
+  }
+  in[0] = _mm_unpacklo_epi64(in[0], in[1]);
+  in[1] = _mm_unpacklo_epi64(in[2], in[3]);
+  in[2] = _mm_unpacklo_epi64(in[4], in[5]);
+  in[3] = _mm_unpacklo_epi64(in[6], in[7]);
+  write_buffer_4x8_round5(dest, in, stride);
+}
+
+// Note: The 16-column 32-element transforms take input in the form of four
+// 8x16 blocks (each stored as a __m128i[16]), which are the four quadrants
+// of the overall 16x32 input buffer.
+static INLINE void idct32_16col(__m128i *tl, __m128i *tr, __m128i *bl,
+                                __m128i *br) {
+  array_transpose_16x16(tl, tr);
+  array_transpose_16x16(bl, br);
+  idct32_8col(tl, bl);
+  idct32_8col(tr, br);
+}
+
+static INLINE void ihalfright32_16col(__m128i *tl, __m128i *tr, __m128i *bl,
+                                      __m128i *br) {
+  __m128i tmpl[16], tmpr[16];
+  int i;
+
+  // Copy the top half of the input to temporary storage
+  for (i = 0; i < 16; ++i) {
+    tmpl[i] = tl[i];
+    tmpr[i] = tr[i];
+  }
+
+  // Generate the top half of the output
+  for (i = 0; i < 16; ++i) {
+    tl[i] = _mm_slli_epi16(bl[i], 2);
+    tr[i] = _mm_slli_epi16(br[i], 2);
+  }
+  array_transpose_16x16(tl, tr);
+
+  // Copy the temporary storage back to the bottom half of the input
+  for (i = 0; i < 16; ++i) {
+    bl[i] = tmpl[i];
+    br[i] = tmpr[i];
+  }
+
+  // Generate the bottom half of the output
+  scale_sqrt2_8x16(bl);
+  scale_sqrt2_8x16(br);
+  aom_idct16_sse2(bl, br);  // Includes a transposition
+}
+
+#if CONFIG_EXT_TX
+static INLINE void iidtx32_16col(__m128i *tl, __m128i *tr, __m128i *bl,
+                                 __m128i *br) {
+  int i;
+  array_transpose_16x16(tl, tr);
+  array_transpose_16x16(bl, br);
+  for (i = 0; i < 16; ++i) {
+    tl[i] = _mm_slli_epi16(tl[i], 2);
+    tr[i] = _mm_slli_epi16(tr[i], 2);
+    bl[i] = _mm_slli_epi16(bl[i], 2);
+    br[i] = _mm_slli_epi16(br[i], 2);
+  }
+}
+#endif  // CONFIG_EXT_TX
+
+static INLINE void write_buffer_16x32_round6(uint8_t *dest, __m128i *intl,
+                                             __m128i *intr, __m128i *inbl,
+                                             __m128i *inbr, int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  int i;
+
+  for (i = 0; i < 16; ++i) {
+    intl[i] = _mm_adds_epi16(intl[i], final_rounding);
+    intr[i] = _mm_adds_epi16(intr[i], final_rounding);
+    inbl[i] = _mm_adds_epi16(inbl[i], final_rounding);
+    inbr[i] = _mm_adds_epi16(inbr[i], final_rounding);
+    intl[i] = _mm_srai_epi16(intl[i], 6);
+    intr[i] = _mm_srai_epi16(intr[i], 6);
+    inbl[i] = _mm_srai_epi16(inbl[i], 6);
+    inbr[i] = _mm_srai_epi16(inbr[i], 6);
+    RECON_AND_STORE(dest + i * stride + 0, intl[i]);
+    RECON_AND_STORE(dest + i * stride + 8, intr[i]);
+    RECON_AND_STORE(dest + (i + 16) * stride + 0, inbl[i]);
+    RECON_AND_STORE(dest + (i + 16) * stride + 8, inbr[i]);
+  }
+}
+
+void av1_iht16x32_512_add_sse2(const tran_low_t *input, uint8_t *dest,
+                               int stride, int tx_type) {
+  __m128i intl[16], intr[16], inbl[16], inbr[16];
+
+  int i;
+  for (i = 0; i < 16; ++i) {
+    intl[i] = load_input_data(input + i * 16 + 0);
+    intr[i] = load_input_data(input + i * 16 + 8);
+    inbl[i] = load_input_data(input + (i + 16) * 16 + 0);
+    inbr[i] = load_input_data(input + (i + 16) * 16 + 8);
+  }
+
+  // Row transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case H_DCT:
+#endif
+      aom_idct16_sse2(intl, intr);
+      aom_idct16_sse2(inbl, inbr);
+      break;
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case H_ADST:
+    case H_FLIPADST:
+#endif
+      aom_iadst16_sse2(intl, intr);
+      aom_iadst16_sse2(inbl, inbr);
+      break;
+#if CONFIG_EXT_TX
+    case V_FLIPADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX:
+      iidtx16_sse2(intl, intr);
+      iidtx16_sse2(inbl, inbr);
+      break;
+#endif
+    default: assert(0); break;
+  }
+
+  scale_sqrt2_8x16(intl);
+  scale_sqrt2_8x16(intr);
+  scale_sqrt2_8x16(inbl);
+  scale_sqrt2_8x16(inbr);
+
+  // Column transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case DCT_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case V_DCT:
+#endif
+      idct32_16col(intl, intr, inbl, inbr);
+      break;
+    case ADST_DCT:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case FLIPADST_ADST:
+    case ADST_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case FLIPADST_DCT:
+    case V_ADST:
+    case V_FLIPADST:
+#endif
+      ihalfright32_16col(intl, intr, inbl, inbr);
+      break;
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case H_FLIPADST:
+    case IDTX: iidtx32_16col(intl, intr, inbl, inbr); break;
+#endif
+    default: assert(0); break;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX:
+#endif
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case FLIPADST_ADST:
+    case V_FLIPADST: FLIPUD_PTR(dest, stride, 32); break;
+    case DCT_FLIPADST:
+    case ADST_FLIPADST:
+    case H_FLIPADST:
+      for (i = 0; i < 16; ++i) {
+        __m128i tmp = intl[i];
+        intl[i] = mm_reverse_epi16(intr[i]);
+        intr[i] = mm_reverse_epi16(tmp);
+        tmp = inbl[i];
+        inbl[i] = mm_reverse_epi16(inbr[i]);
+        inbr[i] = mm_reverse_epi16(tmp);
+      }
+      break;
+    case FLIPADST_FLIPADST:
+      for (i = 0; i < 16; ++i) {
+        __m128i tmp = intl[i];
+        intl[i] = mm_reverse_epi16(intr[i]);
+        intr[i] = mm_reverse_epi16(tmp);
+        tmp = inbl[i];
+        inbl[i] = mm_reverse_epi16(inbr[i]);
+        inbr[i] = mm_reverse_epi16(tmp);
+      }
+      FLIPUD_PTR(dest, stride, 32);
+      break;
+#endif
+    default: assert(0); break;
+  }
+  write_buffer_16x32_round6(dest, intl, intr, inbl, inbr, stride);
+}
+
+static INLINE void write_buffer_32x16_round6(uint8_t *dest, __m128i *in0,
+                                             __m128i *in1, __m128i *in2,
+                                             __m128i *in3, int stride) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i final_rounding = _mm_set1_epi16(1 << 5);
+  int i;
+
+  for (i = 0; i < 16; ++i) {
+    in0[i] = _mm_adds_epi16(in0[i], final_rounding);
+    in1[i] = _mm_adds_epi16(in1[i], final_rounding);
+    in2[i] = _mm_adds_epi16(in2[i], final_rounding);
+    in3[i] = _mm_adds_epi16(in3[i], final_rounding);
+    in0[i] = _mm_srai_epi16(in0[i], 6);
+    in1[i] = _mm_srai_epi16(in1[i], 6);
+    in2[i] = _mm_srai_epi16(in2[i], 6);
+    in3[i] = _mm_srai_epi16(in3[i], 6);
+    RECON_AND_STORE(dest + i * stride + 0, in0[i]);
+    RECON_AND_STORE(dest + i * stride + 8, in1[i]);
+    RECON_AND_STORE(dest + i * stride + 16, in2[i]);
+    RECON_AND_STORE(dest + i * stride + 24, in3[i]);
+  }
+}
+
+void av1_iht32x16_512_add_sse2(const tran_low_t *input, uint8_t *dest,
+                               int stride, int tx_type) {
+  __m128i in0[16], in1[16], in2[16], in3[16];
+  int i;
+
+  for (i = 0; i < 16; ++i) {
+    in0[i] = load_input_data(input + i * 32 + 0);
+    in1[i] = load_input_data(input + i * 32 + 8);
+    in2[i] = load_input_data(input + i * 32 + 16);
+    in3[i] = load_input_data(input + i * 32 + 24);
+  }
+
+  // Row transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case H_DCT:
+#endif
+      idct32_16col(in0, in1, in2, in3);
+      break;
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case H_ADST:
+    case H_FLIPADST:
+#endif
+      ihalfright32_16col(in0, in1, in2, in3);
+      break;
+#if CONFIG_EXT_TX
+    case V_FLIPADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX: iidtx32_16col(in0, in1, in2, in3); break;
+#endif
+    default: assert(0); break;
+  }
+
+  scale_sqrt2_8x16(in0);
+  scale_sqrt2_8x16(in1);
+  scale_sqrt2_8x16(in2);
+  scale_sqrt2_8x16(in3);
+
+  // Column transform
+  switch (tx_type) {
+    case DCT_DCT:
+    case DCT_ADST:
+#if CONFIG_EXT_TX
+    case DCT_FLIPADST:
+    case V_DCT:
+#endif
+      aom_idct16_sse2(in0, in1);
+      aom_idct16_sse2(in2, in3);
+      break;
+    case ADST_DCT:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case FLIPADST_ADST:
+    case ADST_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case FLIPADST_DCT:
+    case V_ADST:
+    case V_FLIPADST:
+#endif
+      aom_iadst16_sse2(in0, in1);
+      aom_iadst16_sse2(in2, in3);
+      break;
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case H_FLIPADST:
+    case IDTX:
+      iidtx16_sse2(in0, in1);
+      iidtx16_sse2(in2, in3);
+      break;
+#endif
+    default: assert(0); break;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+#if CONFIG_EXT_TX
+    case H_DCT:
+    case H_ADST:
+    case V_ADST:
+    case V_DCT:
+    case IDTX:
+#endif
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case FLIPADST_ADST:
+    case V_FLIPADST: FLIPUD_PTR(dest, stride, 16); break;
+    case DCT_FLIPADST:
+    case ADST_FLIPADST:
+    case H_FLIPADST:
+      for (i = 0; i < 16; ++i) {
+        __m128i tmp1 = in0[i];
+        __m128i tmp2 = in1[i];
+        in0[i] = mm_reverse_epi16(in3[i]);
+        in1[i] = mm_reverse_epi16(in2[i]);
+        in2[i] = mm_reverse_epi16(tmp2);
+        in3[i] = mm_reverse_epi16(tmp1);
+      }
+      break;
+    case FLIPADST_FLIPADST:
+      for (i = 0; i < 16; ++i) {
+        __m128i tmp1 = in0[i];
+        __m128i tmp2 = in1[i];
+        in0[i] = mm_reverse_epi16(in3[i]);
+        in1[i] = mm_reverse_epi16(in2[i]);
+        in2[i] = mm_reverse_epi16(tmp2);
+        in3[i] = mm_reverse_epi16(tmp1);
+      }
+      FLIPUD_PTR(dest, stride, 16);
+      break;
+#endif
+    default: assert(0); break;
+  }
+  write_buffer_32x16_round6(dest, in0, in1, in2, in3, stride);
+}
diff --git a/third_party/aom/av1/common/x86/pvq_sse4.c b/third_party/aom/av1/common/x86/pvq_sse4.c
new file mode 100644
index 0000000000..b3ed9efdf4
--- /dev/null
+++ b/third_party/aom/av1/common/x86/pvq_sse4.c
@@ -0,0 +1,252 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <smmintrin.h>
+#include <emmintrin.h>
+#include <tmmintrin.h>
+#include <float.h>
+
+#include "./av1_rtcd.h"
+#include "av1/common/x86/pvq_sse4.h"
+#include "../odintrin.h"
+#include "av1/common/pvq.h"
+
+#define EPSILON 1e-15f
+
+static __m128 horizontal_sum_ps(__m128 x) {
+  x = _mm_add_ps(x, _mm_shuffle_ps(x, x, _MM_SHUFFLE(1, 0, 3, 2)));
+  x = _mm_add_ps(x, _mm_shuffle_ps(x, x, _MM_SHUFFLE(2, 3, 0, 1)));
+  return x;
+}
+
+static __m128i horizontal_sum_epi32(__m128i x) {
+  x = _mm_add_epi32(x, _mm_shuffle_epi32(x, _MM_SHUFFLE(1, 0, 3, 2)));
+  x = _mm_add_epi32(x, _mm_shuffle_epi32(x, _MM_SHUFFLE(2, 3, 0, 1)));
+  return x;
+}
+
+static INLINE float rsqrtf(float x) {
+  float y;
+  _mm_store_ss(&y, _mm_rsqrt_ss(_mm_load_ss(&x)));
+  return y;
+}
+
+/** Find the codepoint on the given PSphere closest to the desired
+ * vector. This is a float-precision PVQ search just to make sure
+ * our tests aren't limited by numerical accuracy. It's close to the
+ * pvq_search_rdo_double_c implementation, but is not bit accurate and
+ * it performs slightly worse on PSNR. One reason is that this code runs
+ * more RDO iterations than the C code. It also uses single precision
+ * floating point math, whereas the C version uses double precision.
+ *
+ * @param [in]      xcoeff  input vector to quantize (x in the math doc)
+ * @param [in]      n       number of dimensions
+ * @param [in]      k       number of pulses
+ * @param [out]     ypulse  optimal codevector found (y in the math doc)
+ * @param [in]      g2      multiplier for the distortion (typically squared
+ *                          gain units)
+ * @param [in] pvq_norm_lambda enc->pvq_norm_lambda for quantized RDO
+ * @param [in]      prev_k  number of pulses already in ypulse that we should
+ *                          reuse for the search (or 0 for a new search)
+ * @return                  cosine distance between x and y (between 0 and 1)
+ */
+double pvq_search_rdo_double_sse4_1(const od_val16 *xcoeff, int n, int k,
+                                    int *ypulse, double g2,
+                                    double pvq_norm_lambda, int prev_k) {
+  int i, j;
+  int reuse_pulses = prev_k > 0 && prev_k <= k;
+  /* TODO - This blows our 8kB stack space budget and should be fixed when
+   converting PVQ to fixed point. */
+  float xx = 0, xy = 0, yy = 0;
+  float x[MAXN + 3];
+  float y[MAXN + 3];
+  float sign_y[MAXN + 3];
+  for (i = 0; i < n; i++) {
+    float tmp = (float)xcoeff[i];
+    xx += tmp * tmp;
+    x[i] = xcoeff[i];
+  }
+
+  x[n] = x[n + 1] = x[n + 2] = 0;
+  ypulse[n] = ypulse[n + 1] = ypulse[n + 2] = 0;
+
+  __m128 sums = _mm_setzero_ps();
+  for (i = 0; i < n; i += 4) {
+    __m128 x4 = _mm_loadu_ps(&x[i]);
+    __m128 s4 = _mm_cmplt_ps(x4, _mm_setzero_ps());
+    /* Save the sign, we'll put it back later. */
+    _mm_storeu_ps(&sign_y[i], s4);
+    /* Get rid of the sign. */
+    x4 = _mm_andnot_ps(_mm_set_ps1(-0.f), x4);
+    sums = _mm_add_ps(sums, x4);
+    if (!reuse_pulses) {
+      /* Clear y and ypulse in case we don't do the projection. */
+      _mm_storeu_ps(&y[i], _mm_setzero_ps());
+      _mm_storeu_si128((__m128i *)&ypulse[i], _mm_setzero_si128());
+    }
+    _mm_storeu_ps(&x[i], x4);
+  }
+  sums = horizontal_sum_ps(sums);
+  int pulses_left = k;
+  {
+    __m128i pulses_sum;
+    __m128 yy4, xy4;
+    xy4 = yy4 = _mm_setzero_ps();
+    pulses_sum = _mm_setzero_si128();
+    if (reuse_pulses) {
+      /* We reuse pulses from a previous search so we don't have to search them
+          again. */
+      for (j = 0; j < n; j += 4) {
+        __m128 x4, y4;
+        __m128i iy4;
+        iy4 = _mm_abs_epi32(_mm_loadu_si128((__m128i *)&ypulse[j]));
+        pulses_sum = _mm_add_epi32(pulses_sum, iy4);
+        _mm_storeu_si128((__m128i *)&ypulse[j], iy4);
+        y4 = _mm_cvtepi32_ps(iy4);
+        x4 = _mm_loadu_ps(&x[j]);
+        xy4 = _mm_add_ps(xy4, _mm_mul_ps(x4, y4));
+        yy4 = _mm_add_ps(yy4, _mm_mul_ps(y4, y4));
+        /* Double the y[] vector so we don't have to do it in the search loop.
+         */
+        _mm_storeu_ps(&y[j], _mm_add_ps(y4, y4));
+      }
+      pulses_left -= _mm_cvtsi128_si32(horizontal_sum_epi32(pulses_sum));
+      xy4 = horizontal_sum_ps(xy4);
+      xy = _mm_cvtss_f32(xy4);
+      yy4 = horizontal_sum_ps(yy4);
+      yy = _mm_cvtss_f32(yy4);
+    } else if (k > (n >> 1)) {
+      /* Do a pre-search by projecting on the pyramid. */
+      __m128 rcp4;
+      float sum = _mm_cvtss_f32(sums);
+      /* If x is too small, just replace it with a pulse at 0. This prevents
+         infinities and NaNs from causing too many pulses to be allocated. Here,
+         64 is an
+         approximation of infinity. */
+      if (sum <= EPSILON) {
+        x[0] = 1.f;
+        for (i = 1; i < n; i++) {
+          x[i] = 0;
+        }
+        sums = _mm_set_ps1(1.f);
+      }
+      /* Using k + e with e < 1 guarantees we cannot get more than k pulses. */
+      rcp4 = _mm_mul_ps(_mm_set_ps1((float)k + .8f), _mm_rcp_ps(sums));
+      xy4 = yy4 = _mm_setzero_ps();
+      pulses_sum = _mm_setzero_si128();
+      for (j = 0; j < n; j += 4) {
+        __m128 rx4, x4, y4;
+        __m128i iy4;
+        x4 = _mm_loadu_ps(&x[j]);
+        rx4 = _mm_mul_ps(x4, rcp4);
+        iy4 = _mm_cvttps_epi32(rx4);
+        pulses_sum = _mm_add_epi32(pulses_sum, iy4);
+        _mm_storeu_si128((__m128i *)&ypulse[j], iy4);
+        y4 = _mm_cvtepi32_ps(iy4);
+        xy4 = _mm_add_ps(xy4, _mm_mul_ps(x4, y4));
+        yy4 = _mm_add_ps(yy4, _mm_mul_ps(y4, y4));
+        /* Double the y[] vector so we don't have to do it in the search loop.
+         */
+        _mm_storeu_ps(&y[j], _mm_add_ps(y4, y4));
+      }
+      pulses_left -= _mm_cvtsi128_si32(horizontal_sum_epi32(pulses_sum));
+      xy = _mm_cvtss_f32(horizontal_sum_ps(xy4));
+      yy = _mm_cvtss_f32(horizontal_sum_ps(yy4));
+    }
+    x[n] = x[n + 1] = x[n + 2] = -100;
+    y[n] = y[n + 1] = y[n + 2] = 100;
+  }
+
+  /* This should never happen. */
+  OD_ASSERT(pulses_left <= n + 3);
+
+  float lambda_delta_rate[MAXN + 3];
+  if (pulses_left) {
+    /* Hoist lambda to avoid the multiply in the loop. */
+    float lambda =
+        0.5f * sqrtf(xx) * (float)pvq_norm_lambda / (FLT_MIN + (float)g2);
+    float delta_rate = 3.f / n;
+    __m128 count = _mm_set_ps(3, 2, 1, 0);
+    for (i = 0; i < n; i += 4) {
+      _mm_storeu_ps(&lambda_delta_rate[i],
+                    _mm_mul_ps(count, _mm_set_ps1(lambda * delta_rate)));
+      count = _mm_add_ps(count, _mm_set_ps(4, 4, 4, 4));
+    }
+  }
+  lambda_delta_rate[n] = lambda_delta_rate[n + 1] = lambda_delta_rate[n + 2] =
+      1e30f;
+
+  for (i = 0; i < pulses_left; i++) {
+    int best_id = 0;
+    __m128 xy4, yy4;
+    __m128 max, max2;
+    __m128i count;
+    __m128i pos;
+
+    /* The squared magnitude term gets added anyway, so we might as well
+        add it outside the loop. */
+    yy = yy + 1;
+    xy4 = _mm_load1_ps(&xy);
+    yy4 = _mm_load1_ps(&yy);
+    max = _mm_setzero_ps();
+    pos = _mm_setzero_si128();
+    count = _mm_set_epi32(3, 2, 1, 0);
+    for (j = 0; j < n; j += 4) {
+      __m128 x4, y4, r4;
+      x4 = _mm_loadu_ps(&x[j]);
+      y4 = _mm_loadu_ps(&y[j]);
+      x4 = _mm_add_ps(x4, xy4);
+      y4 = _mm_add_ps(y4, yy4);
+      y4 = _mm_rsqrt_ps(y4);
+      r4 = _mm_mul_ps(x4, y4);
+      /* Subtract lambda. */
+      r4 = _mm_sub_ps(r4, _mm_loadu_ps(&lambda_delta_rate[j]));
+      /* Update the index of the max. */
+      pos = _mm_max_epi16(
+          pos, _mm_and_si128(count, _mm_castps_si128(_mm_cmpgt_ps(r4, max))));
+      /* Update the max. */
+      max = _mm_max_ps(max, r4);
+      /* Update the indices (+4) */
+      count = _mm_add_epi32(count, _mm_set_epi32(4, 4, 4, 4));
+    }
+    /* Horizontal max. */
+    max2 = _mm_max_ps(max, _mm_shuffle_ps(max, max, _MM_SHUFFLE(1, 0, 3, 2)));
+    max2 =
+        _mm_max_ps(max2, _mm_shuffle_ps(max2, max2, _MM_SHUFFLE(2, 3, 0, 1)));
+    /* Now that max2 contains the max at all positions, look at which value(s)
+       of the
+        partial max is equal to the global max. */
+    pos = _mm_and_si128(pos, _mm_castps_si128(_mm_cmpeq_ps(max, max2)));
+    pos = _mm_max_epi16(pos, _mm_unpackhi_epi64(pos, pos));
+    pos = _mm_max_epi16(pos, _mm_shufflelo_epi16(pos, _MM_SHUFFLE(1, 0, 3, 2)));
+    best_id = _mm_cvtsi128_si32(pos);
+    OD_ASSERT(best_id < n);
+    /* Updating the sums of the new pulse(s) */
+    xy = xy + x[best_id];
+    /* We're multiplying y[j] by two so we don't have to do it here. */
+    yy = yy + y[best_id];
+    /* Only now that we've made the final choice, update y/ypulse. */
+    /* Multiplying y[j] by 2 so we don't have to do it everywhere else. */
+    y[best_id] += 2;
+    ypulse[best_id]++;
+  }
+
+  /* Put the original sign back. */
+  for (i = 0; i < n; i += 4) {
+    __m128i y4;
+    __m128i s4;
+    y4 = _mm_loadu_si128((__m128i *)&ypulse[i]);
+    s4 = _mm_castps_si128(_mm_loadu_ps(&sign_y[i]));
+    y4 = _mm_xor_si128(_mm_add_epi32(y4, s4), s4);
+    _mm_storeu_si128((__m128i *)&ypulse[i], y4);
+  }
+  return xy * rsqrtf(xx * yy + FLT_MIN);
+}
diff --git a/third_party/aom/av1/common/x86/pvq_sse4.h b/third_party/aom/av1/common/x86/pvq_sse4.h
new file mode 100644
index 0000000000..3c4ce85435
--- /dev/null
+++ b/third_party/aom/av1/common/x86/pvq_sse4.h
@@ -0,0 +1,13 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_COMMON_PVQ_X86_SSE4_H_
+#define AOM_COMMON_PVQ_X86_SSE4_H_
+#endif  // AOM_COMMON_PVQ_X86_SSE4_H_
diff --git a/third_party/aom/av1/common/x86/selfguided_sse4.c b/third_party/aom/av1/common/x86/selfguided_sse4.c
new file mode 100644
index 0000000000..260faa8c97
--- /dev/null
+++ b/third_party/aom/av1/common/x86/selfguided_sse4.c
@@ -0,0 +1,1805 @@
+#include <smmintrin.h>
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "av1/common/restoration.h"
+
+/* Calculate four consecutive entries of the intermediate A and B arrays
+   (corresponding to the first loop in the C version of
+   av1_selfguided_restoration)
+*/
+static void calc_block(__m128i sum, __m128i sum_sq, __m128i n,
+                       __m128i one_over_n, __m128i s, int bit_depth, int idx,
+                       int32_t *A, int32_t *B) {
+  __m128i a, b, p;
+#if CONFIG_HIGHBITDEPTH
+  if (bit_depth > 8) {
+    __m128i rounding_a = _mm_set1_epi32((1 << (2 * (bit_depth - 8))) >> 1);
+    __m128i rounding_b = _mm_set1_epi32((1 << (bit_depth - 8)) >> 1);
+    __m128i shift_a = _mm_set_epi64x(0, 2 * (bit_depth - 8));
+    __m128i shift_b = _mm_set_epi64x(0, bit_depth - 8);
+    a = _mm_srl_epi32(_mm_add_epi32(sum_sq, rounding_a), shift_a);
+    b = _mm_srl_epi32(_mm_add_epi32(sum, rounding_b), shift_b);
+    a = _mm_mullo_epi32(a, n);
+    b = _mm_mullo_epi32(b, b);
+    p = _mm_sub_epi32(_mm_max_epi32(a, b), b);
+  } else {
+#endif
+    (void)bit_depth;
+    a = _mm_mullo_epi32(sum_sq, n);
+    b = _mm_mullo_epi32(sum, sum);
+    p = _mm_sub_epi32(a, b);
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif
+
+  __m128i rounding_z = _mm_set1_epi32((1 << SGRPROJ_MTABLE_BITS) >> 1);
+  __m128i z = _mm_srli_epi32(_mm_add_epi32(_mm_mullo_epi32(p, s), rounding_z),
+                             SGRPROJ_MTABLE_BITS);
+  z = _mm_min_epi32(z, _mm_set1_epi32(255));
+
+  // 'Gather' type instructions are not available pre-AVX2, so synthesize a
+  // gather using scalar loads.
+  __m128i a_res = _mm_set_epi32(x_by_xplus1[_mm_extract_epi32(z, 3)],
+                                x_by_xplus1[_mm_extract_epi32(z, 2)],
+                                x_by_xplus1[_mm_extract_epi32(z, 1)],
+                                x_by_xplus1[_mm_extract_epi32(z, 0)]);
+
+  _mm_storeu_si128((__m128i *)&A[idx], a_res);
+
+  __m128i rounding_res = _mm_set1_epi32((1 << SGRPROJ_RECIP_BITS) >> 1);
+  __m128i a_complement = _mm_sub_epi32(_mm_set1_epi32(SGRPROJ_SGR), a_res);
+  __m128i b_int =
+      _mm_mullo_epi32(a_complement, _mm_mullo_epi32(sum, one_over_n));
+  __m128i b_res =
+      _mm_srli_epi32(_mm_add_epi32(b_int, rounding_res), SGRPROJ_RECIP_BITS);
+
+  _mm_storeu_si128((__m128i *)&B[idx], b_res);
+}
+
+static void selfguided_restoration_1_v(uint8_t *src, int width, int height,
+                                       int src_stride, int32_t *A, int32_t *B,
+                                       int buf_stride) {
+  int i, j;
+
+  // Vertical sum
+  // When the width is not a multiple of 4, we know that 'stride' is rounded up
+  // to a multiple of 4. So it is safe for this loop to calculate extra columns
+  // at the right-hand edge of the frame.
+  int width_extend = (width + 3) & ~3;
+  for (j = 0; j < width_extend; j += 4) {
+    __m128i a, b, x, y, x2, y2;
+    __m128i sum, sum_sq, tmp;
+
+    a = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i *)&src[j]));
+    b = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i *)&src[src_stride + j]));
+
+    sum = _mm_cvtepi16_epi32(_mm_add_epi16(a, b));
+    tmp = _mm_unpacklo_epi16(a, b);
+    sum_sq = _mm_madd_epi16(tmp, tmp);
+
+    _mm_store_si128((__m128i *)&B[j], sum);
+    _mm_store_si128((__m128i *)&A[j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i *)&src[2 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    for (i = 1; i < height - 2; ++i) {
+      _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+      _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+      x = _mm_cvtepu8_epi32(
+          _mm_loadl_epi64((__m128i *)&src[(i - 1) * src_stride + j]));
+      y = _mm_cvtepu8_epi32(
+          _mm_loadl_epi64((__m128i *)&src[(i + 2) * src_stride + j]));
+
+      sum = _mm_add_epi32(sum, _mm_sub_epi32(y, x));
+
+      x2 = _mm_mullo_epi32(x, x);
+      y2 = _mm_mullo_epi32(y, y);
+
+      sum_sq = _mm_add_epi32(sum_sq, _mm_sub_epi32(y2, x2));
+    }
+    _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 1) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 1) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 1) * buf_stride + j], sum_sq);
+  }
+}
+
+static void selfguided_restoration_1_h(int32_t *A, int32_t *B, int width,
+                                       int height, int buf_stride, int eps,
+                                       int bit_depth) {
+  int i, j;
+
+  // Horizontal sum
+  int width_extend = (width + 3) & ~3;
+  for (i = 0; i < height; ++i) {
+    int h = AOMMIN(2, height - i) + AOMMIN(1, i);
+
+    __m128i a1 = _mm_loadu_si128((__m128i *)&A[i * buf_stride]);
+    __m128i b1 = _mm_loadu_si128((__m128i *)&B[i * buf_stride]);
+    __m128i a2 = _mm_loadu_si128((__m128i *)&A[i * buf_stride + 4]);
+    __m128i b2 = _mm_loadu_si128((__m128i *)&B[i * buf_stride + 4]);
+
+    // Note: The _mm_slli_si128 call sets up a register containing
+    // {0, A[i * buf_stride], ..., A[i * buf_stride + 2]},
+    // so that the first element of 'sum' (which should only add two values
+    // together) ends up calculated correctly.
+    __m128i sum_ = _mm_add_epi32(_mm_slli_si128(b1, 4),
+                                 _mm_add_epi32(b1, _mm_alignr_epi8(b2, b1, 4)));
+    __m128i sum_sq_ = _mm_add_epi32(
+        _mm_slli_si128(a1, 4), _mm_add_epi32(a1, _mm_alignr_epi8(a2, a1, 4)));
+    __m128i n = _mm_set_epi32(3 * h, 3 * h, 3 * h, 2 * h);
+    __m128i one_over_n =
+        _mm_set_epi32(one_by_x[3 * h - 1], one_by_x[3 * h - 1],
+                      one_by_x[3 * h - 1], one_by_x[2 * h - 1]);
+    __m128i s = _mm_set_epi32(
+        sgrproj_mtable[eps - 1][3 * h - 1], sgrproj_mtable[eps - 1][3 * h - 1],
+        sgrproj_mtable[eps - 1][3 * h - 1], sgrproj_mtable[eps - 1][2 * h - 1]);
+    calc_block(sum_, sum_sq_, n, one_over_n, s, bit_depth, i * buf_stride, A,
+               B);
+
+    n = _mm_set1_epi32(3 * h);
+    one_over_n = _mm_set1_epi32(one_by_x[3 * h - 1]);
+    s = _mm_set1_epi32(sgrproj_mtable[eps - 1][3 * h - 1]);
+
+    // Re-align a1 and b1 so that they start at index i * buf_stride + 3
+    a2 = _mm_alignr_epi8(a2, a1, 12);
+    b2 = _mm_alignr_epi8(b2, b1, 12);
+
+    // Note: When the width is not a multiple of 4, this loop may end up
+    // writing to the last 4 columns of the frame, potentially with incorrect
+    // values (especially for r=2 and r=3).
+    // This is fine, since we fix up those values in the block after this
+    // loop, and in exchange we never have more than four values to
+    // write / fix up after this loop finishes.
+    for (j = 4; j < width_extend - 4; j += 4) {
+      a1 = a2;
+      b1 = b2;
+      a2 = _mm_loadu_si128((__m128i *)&A[i * buf_stride + j + 3]);
+      b2 = _mm_loadu_si128((__m128i *)&B[i * buf_stride + j + 3]);
+      /* Loop invariant: At this point,
+         a1 = original A[i * buf_stride + j - 1 : i * buf_stride + j + 3]
+         a2 = original A[i * buf_stride + j + 3 : i * buf_stride + j + 7]
+         and similar for b1,b2 and B
+      */
+      sum_ = _mm_add_epi32(b1, _mm_add_epi32(_mm_alignr_epi8(b2, b1, 4),
+                                             _mm_alignr_epi8(b2, b1, 8)));
+      sum_sq_ = _mm_add_epi32(a1, _mm_add_epi32(_mm_alignr_epi8(a2, a1, 4),
+                                                _mm_alignr_epi8(a2, a1, 8)));
+      calc_block(sum_, sum_sq_, n, one_over_n, s, bit_depth, i * buf_stride + j,
+                 A, B);
+    }
+    __m128i a3 = _mm_loadu_si128((__m128i *)&A[i * buf_stride + j + 3]);
+    __m128i b3 = _mm_loadu_si128((__m128i *)&B[i * buf_stride + j + 3]);
+
+    j = width - 4;
+    switch (width % 4) {
+      case 0:
+        a1 = a2;
+        b1 = b2;
+        a2 = a3;
+        b2 = b3;
+        break;
+      case 1:
+        a1 = _mm_alignr_epi8(a2, a1, 4);
+        b1 = _mm_alignr_epi8(b2, b1, 4);
+        a2 = _mm_alignr_epi8(a3, a2, 4);
+        b2 = _mm_alignr_epi8(b3, b2, 4);
+        break;
+      case 2:
+        a1 = _mm_alignr_epi8(a2, a1, 8);
+        b1 = _mm_alignr_epi8(b2, b1, 8);
+        a2 = _mm_alignr_epi8(a3, a2, 8);
+        b2 = _mm_alignr_epi8(b3, b2, 8);
+        break;
+      case 3:
+        a1 = _mm_alignr_epi8(a2, a1, 12);
+        b1 = _mm_alignr_epi8(b2, b1, 12);
+        a2 = _mm_alignr_epi8(a3, a2, 12);
+        b2 = _mm_alignr_epi8(b3, b2, 12);
+        break;
+    }
+
+    // Zero out the data loaded from "off the edge" of the array
+    __m128i zero = _mm_setzero_si128();
+    a2 = _mm_blend_epi16(a2, zero, 0xfc);
+    b2 = _mm_blend_epi16(b2, zero, 0xfc);
+
+    sum_ = _mm_add_epi32(b1, _mm_add_epi32(_mm_alignr_epi8(b2, b1, 4),
+                                           _mm_alignr_epi8(b2, b1, 8)));
+    sum_sq_ = _mm_add_epi32(a1, _mm_add_epi32(_mm_alignr_epi8(a2, a1, 4),
+                                              _mm_alignr_epi8(a2, a1, 8)));
+    n = _mm_set_epi32(2 * h, 3 * h, 3 * h, 3 * h);
+    one_over_n = _mm_set_epi32(one_by_x[2 * h - 1], one_by_x[3 * h - 1],
+                               one_by_x[3 * h - 1], one_by_x[3 * h - 1]);
+    s = _mm_set_epi32(
+        sgrproj_mtable[eps - 1][2 * h - 1], sgrproj_mtable[eps - 1][3 * h - 1],
+        sgrproj_mtable[eps - 1][3 * h - 1], sgrproj_mtable[eps - 1][3 * h - 1]);
+    calc_block(sum_, sum_sq_, n, one_over_n, s, bit_depth, i * buf_stride + j,
+               A, B);
+  }
+}
+
+static void selfguided_restoration_2_v(uint8_t *src, int width, int height,
+                                       int src_stride, int32_t *A, int32_t *B,
+                                       int buf_stride) {
+  int i, j;
+
+  // Vertical sum
+  int width_extend = (width + 3) & ~3;
+  for (j = 0; j < width_extend; j += 4) {
+    __m128i a, b, c, c2, x, y, x2, y2;
+    __m128i sum, sum_sq, tmp;
+
+    a = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i *)&src[j]));
+    b = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i *)&src[src_stride + j]));
+    c = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i *)&src[2 * src_stride + j]));
+
+    sum = _mm_cvtepi16_epi32(_mm_add_epi16(_mm_add_epi16(a, b), c));
+    // Important: Since c may be up to 2^8, the result on squaring may
+    // be up to 2^16. So we need to zero-extend, not sign-extend.
+    c2 = _mm_cvtepu16_epi32(_mm_mullo_epi16(c, c));
+    tmp = _mm_unpacklo_epi16(a, b);
+    sum_sq = _mm_add_epi32(_mm_madd_epi16(tmp, tmp), c2);
+
+    _mm_store_si128((__m128i *)&B[j], sum);
+    _mm_store_si128((__m128i *)&A[j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i *)&src[3 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i *)&src[4 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    for (i = 2; i < height - 3; ++i) {
+      _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+      _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+      x = _mm_cvtepu8_epi32(
+          _mm_cvtsi32_si128(*((int *)&src[(i - 2) * src_stride + j])));
+      y = _mm_cvtepu8_epi32(
+          _mm_cvtsi32_si128(*((int *)&src[(i + 3) * src_stride + j])));
+
+      sum = _mm_add_epi32(sum, _mm_sub_epi32(y, x));
+
+      x2 = _mm_mullo_epi32(x, x);
+      y2 = _mm_mullo_epi32(y, y);
+
+      sum_sq = _mm_add_epi32(sum_sq, _mm_sub_epi32(y2, x2));
+    }
+    _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 2) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 1) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 1) * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 1) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 2) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 2) * buf_stride + j], sum_sq);
+  }
+}
+
+static void selfguided_restoration_2_h(int32_t *A, int32_t *B, int width,
+                                       int height, int buf_stride, int eps,
+                                       int bit_depth) {
+  int i, j;
+
+  // Horizontal sum
+  int width_extend = (width + 3) & ~3;
+  for (i = 0; i < height; ++i) {
+    int h = AOMMIN(3, height - i) + AOMMIN(2, i);
+
+    __m128i a1 = _mm_loadu_si128((__m128i *)&A[i * buf_stride]);
+    __m128i b1 = _mm_loadu_si128((__m128i *)&B[i * buf_stride]);
+    __m128i a2 = _mm_loadu_si128((__m128i *)&A[i * buf_stride + 4]);
+    __m128i b2 = _mm_loadu_si128((__m128i *)&B[i * buf_stride + 4]);
+
+    __m128i sum_ = _mm_add_epi32(
+        _mm_add_epi32(
+            _mm_add_epi32(_mm_slli_si128(b1, 8), _mm_slli_si128(b1, 4)),
+            _mm_add_epi32(b1, _mm_alignr_epi8(b2, b1, 4))),
+        _mm_alignr_epi8(b2, b1, 8));
+    __m128i sum_sq_ = _mm_add_epi32(
+        _mm_add_epi32(
+            _mm_add_epi32(_mm_slli_si128(a1, 8), _mm_slli_si128(a1, 4)),
+            _mm_add_epi32(a1, _mm_alignr_epi8(a2, a1, 4))),
+        _mm_alignr_epi8(a2, a1, 8));
+
+    __m128i n = _mm_set_epi32(5 * h, 5 * h, 4 * h, 3 * h);
+    __m128i one_over_n =
+        _mm_set_epi32(one_by_x[5 * h - 1], one_by_x[5 * h - 1],
+                      one_by_x[4 * h - 1], one_by_x[3 * h - 1]);
+    __m128i s = _mm_set_epi32(
+        sgrproj_mtable[eps - 1][5 * h - 1], sgrproj_mtable[eps - 1][5 * h - 1],
+        sgrproj_mtable[eps - 1][4 * h - 1], sgrproj_mtable[eps - 1][3 * h - 1]);
+    calc_block(sum_, sum_sq_, n, one_over_n, s, bit_depth, i * buf_stride, A,
+               B);
+
+    // Re-align a1 and b1 so that they start at index i * buf_stride + 2
+    a2 = _mm_alignr_epi8(a2, a1, 8);
+    b2 = _mm_alignr_epi8(b2, b1, 8);
+
+    n = _mm_set1_epi32(5 * h);
+    one_over_n = _mm_set1_epi32(one_by_x[5 * h - 1]);
+    s = _mm_set1_epi32(sgrproj_mtable[eps - 1][5 * h - 1]);
+
+    for (j = 4; j < width_extend - 4; j += 4) {
+      a1 = a2;
+      a2 = _mm_loadu_si128((__m128i *)&A[i * buf_stride + j + 2]);
+      b1 = b2;
+      b2 = _mm_loadu_si128((__m128i *)&B[i * buf_stride + j + 2]);
+      /* Loop invariant: At this point,
+         a1 = original A[i * buf_stride + j - 2 : i * buf_stride + j + 2]
+         a2 = original A[i * buf_stride + j + 2 : i * buf_stride + j + 6]
+         and similar for b1,b2 and B
+      */
+      sum_ = _mm_add_epi32(
+          _mm_add_epi32(b1, _mm_add_epi32(_mm_alignr_epi8(b2, b1, 4),
+                                          _mm_alignr_epi8(b2, b1, 8))),
+          _mm_add_epi32(_mm_alignr_epi8(b2, b1, 12), b2));
+      sum_sq_ = _mm_add_epi32(
+          _mm_add_epi32(a1, _mm_add_epi32(_mm_alignr_epi8(a2, a1, 4),
+                                          _mm_alignr_epi8(a2, a1, 8))),
+          _mm_add_epi32(_mm_alignr_epi8(a2, a1, 12), a2));
+
+      calc_block(sum_, sum_sq_, n, one_over_n, s, bit_depth, i * buf_stride + j,
+                 A, B);
+    }
+    // If the width is not a multiple of 4, we need to reset j to width - 4
+    // and adjust a1, a2, b1, b2 so that the loop invariant above is maintained
+    __m128i a3 = _mm_loadu_si128((__m128i *)&A[i * buf_stride + j + 2]);
+    __m128i b3 = _mm_loadu_si128((__m128i *)&B[i * buf_stride + j + 2]);
+
+    j = width - 4;
+    switch (width % 4) {
+      case 0:
+        a1 = a2;
+        b1 = b2;
+        a2 = a3;
+        b2 = b3;
+        break;
+      case 1:
+        a1 = _mm_alignr_epi8(a2, a1, 4);
+        b1 = _mm_alignr_epi8(b2, b1, 4);
+        a2 = _mm_alignr_epi8(a3, a2, 4);
+        b2 = _mm_alignr_epi8(b3, b2, 4);
+        break;
+      case 2:
+        a1 = _mm_alignr_epi8(a2, a1, 8);
+        b1 = _mm_alignr_epi8(b2, b1, 8);
+        a2 = _mm_alignr_epi8(a3, a2, 8);
+        b2 = _mm_alignr_epi8(b3, b2, 8);
+        break;
+      case 3:
+        a1 = _mm_alignr_epi8(a2, a1, 12);
+        b1 = _mm_alignr_epi8(b2, b1, 12);
+        a2 = _mm_alignr_epi8(a3, a2, 12);
+        b2 = _mm_alignr_epi8(b3, b2, 12);
+        break;
+    }
+
+    // Zero out the data loaded from "off the edge" of the array
+    __m128i zero = _mm_setzero_si128();
+    a2 = _mm_blend_epi16(a2, zero, 0xf0);
+    b2 = _mm_blend_epi16(b2, zero, 0xf0);
+
+    sum_ = _mm_add_epi32(
+        _mm_add_epi32(b1, _mm_add_epi32(_mm_alignr_epi8(b2, b1, 4),
+                                        _mm_alignr_epi8(b2, b1, 8))),
+        _mm_add_epi32(_mm_alignr_epi8(b2, b1, 12), b2));
+    sum_sq_ = _mm_add_epi32(
+        _mm_add_epi32(a1, _mm_add_epi32(_mm_alignr_epi8(a2, a1, 4),
+                                        _mm_alignr_epi8(a2, a1, 8))),
+        _mm_add_epi32(_mm_alignr_epi8(a2, a1, 12), a2));
+
+    n = _mm_set_epi32(3 * h, 4 * h, 5 * h, 5 * h);
+    one_over_n = _mm_set_epi32(one_by_x[3 * h - 1], one_by_x[4 * h - 1],
+                               one_by_x[5 * h - 1], one_by_x[5 * h - 1]);
+    s = _mm_set_epi32(
+        sgrproj_mtable[eps - 1][3 * h - 1], sgrproj_mtable[eps - 1][4 * h - 1],
+        sgrproj_mtable[eps - 1][5 * h - 1], sgrproj_mtable[eps - 1][5 * h - 1]);
+    calc_block(sum_, sum_sq_, n, one_over_n, s, bit_depth, i * buf_stride + j,
+               A, B);
+  }
+}
+
+static void selfguided_restoration_3_v(uint8_t *src, int width, int height,
+                                       int src_stride, int32_t *A, int32_t *B,
+                                       int buf_stride) {
+  int i, j;
+
+  // Vertical sum over 7-pixel regions, 4 columns at a time
+  int width_extend = (width + 3) & ~3;
+  for (j = 0; j < width_extend; j += 4) {
+    __m128i a, b, c, d, x, y, x2, y2;
+    __m128i sum, sum_sq, tmp, tmp2;
+
+    a = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i *)&src[j]));
+    b = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i *)&src[src_stride + j]));
+    c = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i *)&src[2 * src_stride + j]));
+    d = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i *)&src[3 * src_stride + j]));
+
+    sum = _mm_cvtepi16_epi32(
+        _mm_add_epi16(_mm_add_epi16(a, b), _mm_add_epi16(c, d)));
+    tmp = _mm_unpacklo_epi16(a, b);
+    tmp2 = _mm_unpacklo_epi16(c, d);
+    sum_sq =
+        _mm_add_epi32(_mm_madd_epi16(tmp, tmp), _mm_madd_epi16(tmp2, tmp2));
+
+    _mm_store_si128((__m128i *)&B[j], sum);
+    _mm_store_si128((__m128i *)&A[j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i *)&src[4 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i *)&src[5 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[2 * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[2 * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(_mm_loadl_epi64((__m128i *)&src[6 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    for (i = 3; i < height - 4; ++i) {
+      _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+      _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+      x = _mm_cvtepu8_epi32(
+          _mm_cvtsi32_si128(*((int *)&src[(i - 3) * src_stride + j])));
+      y = _mm_cvtepu8_epi32(
+          _mm_cvtsi32_si128(*((int *)&src[(i + 4) * src_stride + j])));
+
+      sum = _mm_add_epi32(sum, _mm_sub_epi32(y, x));
+
+      x2 = _mm_mullo_epi32(x, x);
+      y2 = _mm_mullo_epi32(y, y);
+
+      sum_sq = _mm_add_epi32(sum_sq, _mm_sub_epi32(y2, x2));
+    }
+    _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 3) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 1) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 1) * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 2) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 2) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 2) * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu8_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 1) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 3) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 3) * buf_stride + j], sum_sq);
+  }
+}
+
+static void selfguided_restoration_3_h(int32_t *A, int32_t *B, int width,
+                                       int height, int buf_stride, int eps,
+                                       int bit_depth) {
+  int i, j;
+  // Horizontal sum over 7-pixel regions of dst
+  int width_extend = (width + 3) & ~3;
+  for (i = 0; i < height; ++i) {
+    int h = AOMMIN(4, height - i) + AOMMIN(3, i);
+
+    __m128i a1 = _mm_loadu_si128((__m128i *)&A[i * buf_stride]);
+    __m128i b1 = _mm_loadu_si128((__m128i *)&B[i * buf_stride]);
+    __m128i a2 = _mm_loadu_si128((__m128i *)&A[i * buf_stride + 4]);
+    __m128i b2 = _mm_loadu_si128((__m128i *)&B[i * buf_stride + 4]);
+
+    __m128i sum_ = _mm_add_epi32(
+        _mm_add_epi32(
+            _mm_add_epi32(_mm_slli_si128(b1, 12), _mm_slli_si128(b1, 8)),
+            _mm_add_epi32(_mm_slli_si128(b1, 4), b1)),
+        _mm_add_epi32(_mm_add_epi32(_mm_alignr_epi8(b2, b1, 4),
+                                    _mm_alignr_epi8(b2, b1, 8)),
+                      _mm_alignr_epi8(b2, b1, 12)));
+    __m128i sum_sq_ = _mm_add_epi32(
+        _mm_add_epi32(
+            _mm_add_epi32(_mm_slli_si128(a1, 12), _mm_slli_si128(a1, 8)),
+            _mm_add_epi32(_mm_slli_si128(a1, 4), a1)),
+        _mm_add_epi32(_mm_add_epi32(_mm_alignr_epi8(a2, a1, 4),
+                                    _mm_alignr_epi8(a2, a1, 8)),
+                      _mm_alignr_epi8(a2, a1, 12)));
+
+    __m128i n = _mm_set_epi32(7 * h, 6 * h, 5 * h, 4 * h);
+    __m128i one_over_n =
+        _mm_set_epi32(one_by_x[7 * h - 1], one_by_x[6 * h - 1],
+                      one_by_x[5 * h - 1], one_by_x[4 * h - 1]);
+    __m128i s = _mm_set_epi32(
+        sgrproj_mtable[eps - 1][7 * h - 1], sgrproj_mtable[eps - 1][6 * h - 1],
+        sgrproj_mtable[eps - 1][5 * h - 1], sgrproj_mtable[eps - 1][4 * h - 1]);
+    calc_block(sum_, sum_sq_, n, one_over_n, s, bit_depth, i * buf_stride, A,
+               B);
+
+    // Re-align a1 and b1 so that they start at index i * buf_stride + 1
+    a2 = _mm_alignr_epi8(a2, a1, 4);
+    b2 = _mm_alignr_epi8(b2, b1, 4);
+
+    n = _mm_set1_epi32(7 * h);
+    one_over_n = _mm_set1_epi32(one_by_x[7 * h - 1]);
+    s = _mm_set1_epi32(sgrproj_mtable[eps - 1][7 * h - 1]);
+
+    for (j = 4; j < width_extend - 4; j += 4) {
+      a1 = a2;
+      a2 = _mm_loadu_si128((__m128i *)&A[i * buf_stride + j + 1]);
+      b1 = b2;
+      b2 = _mm_loadu_si128((__m128i *)&B[i * buf_stride + j + 1]);
+      __m128i a3 = _mm_loadu_si128((__m128i *)&A[i * buf_stride + j + 5]);
+      __m128i b3 = _mm_loadu_si128((__m128i *)&B[i * buf_stride + j + 5]);
+      /* Loop invariant: At this point,
+         a1 = original A[i * buf_stride + j - 3 : i * buf_stride + j + 1]
+         a2 = original A[i * buf_stride + j + 1 : i * buf_stride + j + 5]
+         a3 = original A[i * buf_stride + j + 5 : i * buf_stride + j + 9]
+         and similar for b1,b2,b3 and B
+      */
+      sum_ = _mm_add_epi32(
+          _mm_add_epi32(_mm_add_epi32(b1, _mm_alignr_epi8(b2, b1, 4)),
+                        _mm_add_epi32(_mm_alignr_epi8(b2, b1, 8),
+                                      _mm_alignr_epi8(b2, b1, 12))),
+          _mm_add_epi32(_mm_add_epi32(b2, _mm_alignr_epi8(b3, b2, 4)),
+                        _mm_alignr_epi8(b3, b2, 8)));
+      sum_sq_ = _mm_add_epi32(
+          _mm_add_epi32(_mm_add_epi32(a1, _mm_alignr_epi8(a2, a1, 4)),
+                        _mm_add_epi32(_mm_alignr_epi8(a2, a1, 8),
+                                      _mm_alignr_epi8(a2, a1, 12))),
+          _mm_add_epi32(_mm_add_epi32(a2, _mm_alignr_epi8(a3, a2, 4)),
+                        _mm_alignr_epi8(a3, a2, 8)));
+
+      calc_block(sum_, sum_sq_, n, one_over_n, s, bit_depth, i * buf_stride + j,
+                 A, B);
+    }
+    __m128i a3 = _mm_loadu_si128((__m128i *)&A[i * buf_stride + j + 1]);
+    __m128i b3 = _mm_loadu_si128((__m128i *)&B[i * buf_stride + j + 1]);
+
+    j = width - 4;
+    switch (width % 4) {
+      case 0:
+        a1 = a2;
+        b1 = b2;
+        a2 = a3;
+        b2 = b3;
+        break;
+      case 1:
+        a1 = _mm_alignr_epi8(a2, a1, 4);
+        b1 = _mm_alignr_epi8(b2, b1, 4);
+        a2 = _mm_alignr_epi8(a3, a2, 4);
+        b2 = _mm_alignr_epi8(b3, b2, 4);
+        break;
+      case 2:
+        a1 = _mm_alignr_epi8(a2, a1, 8);
+        b1 = _mm_alignr_epi8(b2, b1, 8);
+        a2 = _mm_alignr_epi8(a3, a2, 8);
+        b2 = _mm_alignr_epi8(b3, b2, 8);
+        break;
+      case 3:
+        a1 = _mm_alignr_epi8(a2, a1, 12);
+        b1 = _mm_alignr_epi8(b2, b1, 12);
+        a2 = _mm_alignr_epi8(a3, a2, 12);
+        b2 = _mm_alignr_epi8(b3, b2, 12);
+        break;
+    }
+
+    // Zero out the data loaded from "off the edge" of the array
+    __m128i zero = _mm_setzero_si128();
+    a2 = _mm_blend_epi16(a2, zero, 0xc0);
+    b2 = _mm_blend_epi16(b2, zero, 0xc0);
+
+    sum_ = _mm_add_epi32(
+        _mm_add_epi32(_mm_add_epi32(b1, _mm_alignr_epi8(b2, b1, 4)),
+                      _mm_add_epi32(_mm_alignr_epi8(b2, b1, 8),
+                                    _mm_alignr_epi8(b2, b1, 12))),
+        _mm_add_epi32(_mm_add_epi32(b2, _mm_alignr_epi8(zero, b2, 4)),
+                      _mm_alignr_epi8(zero, b2, 8)));
+    sum_sq_ = _mm_add_epi32(
+        _mm_add_epi32(_mm_add_epi32(a1, _mm_alignr_epi8(a2, a1, 4)),
+                      _mm_add_epi32(_mm_alignr_epi8(a2, a1, 8),
+                                    _mm_alignr_epi8(a2, a1, 12))),
+        _mm_add_epi32(_mm_add_epi32(a2, _mm_alignr_epi8(zero, a2, 4)),
+                      _mm_alignr_epi8(zero, a2, 8)));
+
+    n = _mm_set_epi32(4 * h, 5 * h, 6 * h, 7 * h);
+    one_over_n = _mm_set_epi32(one_by_x[4 * h - 1], one_by_x[5 * h - 1],
+                               one_by_x[6 * h - 1], one_by_x[7 * h - 1]);
+    s = _mm_set_epi32(
+        sgrproj_mtable[eps - 1][4 * h - 1], sgrproj_mtable[eps - 1][5 * h - 1],
+        sgrproj_mtable[eps - 1][6 * h - 1], sgrproj_mtable[eps - 1][7 * h - 1]);
+    calc_block(sum_, sum_sq_, n, one_over_n, s, bit_depth, i * buf_stride + j,
+               A, B);
+  }
+}
+
+void av1_selfguided_restoration_sse4_1(uint8_t *dgd, int width, int height,
+                                       int stride, int32_t *dst, int dst_stride,
+                                       int r, int eps, int32_t *tmpbuf) {
+  int32_t *A = tmpbuf;
+  int32_t *B = A + SGRPROJ_OUTBUF_SIZE;
+  int i, j;
+  // Adjusting the stride of A and B here appears to avoid bad cache effects,
+  // leading to a significant speed improvement.
+  // We also align the stride to a multiple of 16 bytes for efficiency.
+  int buf_stride = ((width + 3) & ~3) + 16;
+
+  // Don't filter tiles with dimensions < 5 on any axis
+  if ((width < 5) || (height < 5)) return;
+
+  if (r == 1) {
+    selfguided_restoration_1_v(dgd, width, height, stride, A, B, buf_stride);
+    selfguided_restoration_1_h(A, B, width, height, buf_stride, eps, 8);
+  } else if (r == 2) {
+    selfguided_restoration_2_v(dgd, width, height, stride, A, B, buf_stride);
+    selfguided_restoration_2_h(A, B, width, height, buf_stride, eps, 8);
+  } else if (r == 3) {
+    selfguided_restoration_3_v(dgd, width, height, stride, A, B, buf_stride);
+    selfguided_restoration_3_h(A, B, width, height, buf_stride, eps, 8);
+  } else {
+    assert(0);
+  }
+
+  {
+    i = 0;
+    j = 0;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = 3 * A[k] + 2 * A[k + 1] + 2 * A[k + buf_stride] +
+                        A[k + buf_stride + 1];
+      const int32_t b = 3 * B[k] + 2 * B[k + 1] + 2 * B[k + buf_stride] +
+                        B[k + buf_stride + 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = A[k] + 2 * (A[k - 1] + A[k + 1]) + A[k + buf_stride] +
+                        A[k + buf_stride - 1] + A[k + buf_stride + 1];
+      const int32_t b = B[k] + 2 * (B[k - 1] + B[k + 1]) + B[k + buf_stride] +
+                        B[k + buf_stride - 1] + B[k + buf_stride + 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+    j = width - 1;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = 3 * A[k] + 2 * A[k - 1] + 2 * A[k + buf_stride] +
+                        A[k + buf_stride - 1];
+      const int32_t b = 3 * B[k] + 2 * B[k - 1] + 2 * B[k + buf_stride] +
+                        B[k + buf_stride - 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+  }
+  for (i = 1; i < height - 1; ++i) {
+    j = 0;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = A[k] + 2 * (A[k - buf_stride] + A[k + buf_stride]) +
+                        A[k + 1] + A[k - buf_stride + 1] +
+                        A[k + buf_stride + 1];
+      const int32_t b = B[k] + 2 * (B[k - buf_stride] + B[k + buf_stride]) +
+                        B[k + 1] + B[k - buf_stride + 1] +
+                        B[k + buf_stride + 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+
+    // Vectorize the innermost loop
+    for (j = 1; j < width - 1; j += 4) {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 5;
+
+      __m128i tmp0 = _mm_loadu_si128((__m128i *)&A[k - 1 - buf_stride]);
+      __m128i tmp1 = _mm_loadu_si128((__m128i *)&A[k + 3 - buf_stride]);
+      __m128i tmp2 = _mm_loadu_si128((__m128i *)&A[k - 1]);
+      __m128i tmp3 = _mm_loadu_si128((__m128i *)&A[k + 3]);
+      __m128i tmp4 = _mm_loadu_si128((__m128i *)&A[k - 1 + buf_stride]);
+      __m128i tmp5 = _mm_loadu_si128((__m128i *)&A[k + 3 + buf_stride]);
+
+      __m128i a0 = _mm_add_epi32(
+          _mm_add_epi32(_mm_add_epi32(_mm_alignr_epi8(tmp3, tmp2, 4), tmp2),
+                        _mm_add_epi32(_mm_alignr_epi8(tmp3, tmp2, 8),
+                                      _mm_alignr_epi8(tmp5, tmp4, 4))),
+          _mm_alignr_epi8(tmp1, tmp0, 4));
+      __m128i a1 = _mm_add_epi32(_mm_add_epi32(tmp0, tmp4),
+                                 _mm_add_epi32(_mm_alignr_epi8(tmp1, tmp0, 8),
+                                               _mm_alignr_epi8(tmp5, tmp4, 8)));
+      __m128i a = _mm_sub_epi32(_mm_slli_epi32(_mm_add_epi32(a0, a1), 2), a1);
+
+      __m128i tmp6 = _mm_loadu_si128((__m128i *)&B[k - 1 - buf_stride]);
+      __m128i tmp7 = _mm_loadu_si128((__m128i *)&B[k + 3 - buf_stride]);
+      __m128i tmp8 = _mm_loadu_si128((__m128i *)&B[k - 1]);
+      __m128i tmp9 = _mm_loadu_si128((__m128i *)&B[k + 3]);
+      __m128i tmp10 = _mm_loadu_si128((__m128i *)&B[k - 1 + buf_stride]);
+      __m128i tmp11 = _mm_loadu_si128((__m128i *)&B[k + 3 + buf_stride]);
+
+      __m128i b0 = _mm_add_epi32(
+          _mm_add_epi32(_mm_add_epi32(_mm_alignr_epi8(tmp9, tmp8, 4), tmp8),
+                        _mm_add_epi32(_mm_alignr_epi8(tmp9, tmp8, 8),
+                                      _mm_alignr_epi8(tmp11, tmp10, 4))),
+          _mm_alignr_epi8(tmp7, tmp6, 4));
+      __m128i b1 =
+          _mm_add_epi32(_mm_add_epi32(tmp6, tmp10),
+                        _mm_add_epi32(_mm_alignr_epi8(tmp7, tmp6, 8),
+                                      _mm_alignr_epi8(tmp11, tmp10, 8)));
+      __m128i b = _mm_sub_epi32(_mm_slli_epi32(_mm_add_epi32(b0, b1), 2), b1);
+
+      __m128i src = _mm_cvtepu8_epi32(_mm_loadu_si128((__m128i *)&dgd[l]));
+
+      __m128i rounding = _mm_set1_epi32(
+          (1 << (SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS)) >> 1);
+      __m128i v = _mm_add_epi32(_mm_mullo_epi32(a, src), b);
+      __m128i w = _mm_srai_epi32(_mm_add_epi32(v, rounding),
+                                 SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+      _mm_storeu_si128((__m128i *)&dst[m], w);
+    }
+
+    // Deal with any extra pixels at the right-hand edge of the frame
+    // (typically have 2 such pixels, but may have anywhere between 0 and 3)
+    for (; j < width - 1; ++j) {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 5;
+      const int32_t a =
+          (A[k] + A[k - 1] + A[k + 1] + A[k - buf_stride] + A[k + buf_stride]) *
+              4 +
+          (A[k - 1 - buf_stride] + A[k - 1 + buf_stride] +
+           A[k + 1 - buf_stride] + A[k + 1 + buf_stride]) *
+              3;
+      const int32_t b =
+          (B[k] + B[k - 1] + B[k + 1] + B[k - buf_stride] + B[k + buf_stride]) *
+              4 +
+          (B[k - 1 - buf_stride] + B[k - 1 + buf_stride] +
+           B[k + 1 - buf_stride] + B[k + 1 + buf_stride]) *
+              3;
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+
+    j = width - 1;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = A[k] + 2 * (A[k - buf_stride] + A[k + buf_stride]) +
+                        A[k - 1] + A[k - buf_stride - 1] +
+                        A[k + buf_stride - 1];
+      const int32_t b = B[k] + 2 * (B[k - buf_stride] + B[k + buf_stride]) +
+                        B[k - 1] + B[k - buf_stride - 1] +
+                        B[k + buf_stride - 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+  }
+
+  {
+    i = height - 1;
+    j = 0;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = 3 * A[k] + 2 * A[k + 1] + 2 * A[k - buf_stride] +
+                        A[k - buf_stride + 1];
+      const int32_t b = 3 * B[k] + 2 * B[k + 1] + 2 * B[k - buf_stride] +
+                        B[k - buf_stride + 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = A[k] + 2 * (A[k - 1] + A[k + 1]) + A[k - buf_stride] +
+                        A[k - buf_stride - 1] + A[k - buf_stride + 1];
+      const int32_t b = B[k] + 2 * (B[k - 1] + B[k + 1]) + B[k - buf_stride] +
+                        B[k - buf_stride - 1] + B[k - buf_stride + 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+    j = width - 1;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = 3 * A[k] + 2 * A[k - 1] + 2 * A[k - buf_stride] +
+                        A[k - buf_stride - 1];
+      const int32_t b = 3 * B[k] + 2 * B[k - 1] + 2 * B[k - buf_stride] +
+                        B[k - buf_stride - 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+  }
+}
+
+void av1_highpass_filter_sse4_1(uint8_t *dgd, int width, int height, int stride,
+                                int32_t *dst, int dst_stride, int corner,
+                                int edge) {
+  int i, j;
+  const int center = (1 << SGRPROJ_RST_BITS) - 4 * (corner + edge);
+
+  {
+    i = 0;
+    j = 0;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] + edge * (dgd[k + 1] + dgd[k + stride] + dgd[k] * 2) +
+          corner *
+              (dgd[k + stride + 1] + dgd[k + 1] + dgd[k + stride] + dgd[k]);
+    }
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] = center * dgd[k] +
+               edge * (dgd[k - 1] + dgd[k + stride] + dgd[k + 1] + dgd[k]) +
+               corner * (dgd[k + stride - 1] + dgd[k + stride + 1] +
+                         dgd[k - 1] + dgd[k + 1]);
+    }
+    j = width - 1;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] + edge * (dgd[k - 1] + dgd[k + stride] + dgd[k] * 2) +
+          corner *
+              (dgd[k + stride - 1] + dgd[k - 1] + dgd[k + stride] + dgd[k]);
+    }
+  }
+  {
+    i = height - 1;
+    j = 0;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] + edge * (dgd[k + 1] + dgd[k - stride] + dgd[k] * 2) +
+          corner *
+              (dgd[k - stride + 1] + dgd[k + 1] + dgd[k - stride] + dgd[k]);
+    }
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] = center * dgd[k] +
+               edge * (dgd[k - 1] + dgd[k - stride] + dgd[k + 1] + dgd[k]) +
+               corner * (dgd[k - stride - 1] + dgd[k - stride + 1] +
+                         dgd[k - 1] + dgd[k + 1]);
+    }
+    j = width - 1;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] + edge * (dgd[k - 1] + dgd[k - stride] + dgd[k] * 2) +
+          corner *
+              (dgd[k - stride - 1] + dgd[k - 1] + dgd[k - stride] + dgd[k]);
+    }
+  }
+  __m128i center_ = _mm_set1_epi16(center);
+  __m128i edge_ = _mm_set1_epi16(edge);
+  __m128i corner_ = _mm_set1_epi16(corner);
+  for (i = 1; i < height - 1; ++i) {
+    j = 0;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] +
+          edge * (dgd[k - stride] + dgd[k + 1] + dgd[k + stride] + dgd[k]) +
+          corner * (dgd[k + stride + 1] + dgd[k - stride + 1] +
+                    dgd[k - stride] + dgd[k + stride]);
+    }
+    // Process in units of 8 pixels at a time.
+    for (j = 1; j < width - 8; j += 8) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+
+      __m128i a = _mm_loadu_si128((__m128i *)&dgd[k - stride - 1]);
+      __m128i b = _mm_loadu_si128((__m128i *)&dgd[k - 1]);
+      __m128i c = _mm_loadu_si128((__m128i *)&dgd[k + stride - 1]);
+
+      __m128i tl = _mm_cvtepu8_epi16(a);
+      __m128i tr = _mm_cvtepu8_epi16(_mm_srli_si128(a, 8));
+      __m128i cl = _mm_cvtepu8_epi16(b);
+      __m128i cr = _mm_cvtepu8_epi16(_mm_srli_si128(b, 8));
+      __m128i bl = _mm_cvtepu8_epi16(c);
+      __m128i br = _mm_cvtepu8_epi16(_mm_srli_si128(c, 8));
+
+      __m128i x = _mm_alignr_epi8(cr, cl, 2);
+      __m128i y = _mm_add_epi16(_mm_add_epi16(_mm_alignr_epi8(tr, tl, 2), cl),
+                                _mm_add_epi16(_mm_alignr_epi8(br, bl, 2),
+                                              _mm_alignr_epi8(cr, cl, 4)));
+      __m128i z = _mm_add_epi16(_mm_add_epi16(tl, bl),
+                                _mm_add_epi16(_mm_alignr_epi8(tr, tl, 4),
+                                              _mm_alignr_epi8(br, bl, 4)));
+
+      __m128i res = _mm_add_epi16(_mm_mullo_epi16(x, center_),
+                                  _mm_add_epi16(_mm_mullo_epi16(y, edge_),
+                                                _mm_mullo_epi16(z, corner_)));
+
+      _mm_storeu_si128((__m128i *)&dst[l], _mm_cvtepi16_epi32(res));
+      _mm_storeu_si128((__m128i *)&dst[l + 4],
+                       _mm_cvtepi16_epi32(_mm_srli_si128(res, 8)));
+    }
+    // If there are enough pixels left in this row, do another batch of 4
+    // pixels.
+    for (; j < width - 4; j += 4) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+
+      __m128i a = _mm_loadl_epi64((__m128i *)&dgd[k - stride - 1]);
+      __m128i b = _mm_loadl_epi64((__m128i *)&dgd[k - 1]);
+      __m128i c = _mm_loadl_epi64((__m128i *)&dgd[k + stride - 1]);
+
+      __m128i tl = _mm_cvtepu8_epi16(a);
+      __m128i cl = _mm_cvtepu8_epi16(b);
+      __m128i bl = _mm_cvtepu8_epi16(c);
+
+      __m128i x = _mm_srli_si128(cl, 2);
+      __m128i y = _mm_add_epi16(
+          _mm_add_epi16(_mm_srli_si128(tl, 2), cl),
+          _mm_add_epi16(_mm_srli_si128(bl, 2), _mm_srli_si128(cl, 4)));
+      __m128i z = _mm_add_epi16(
+          _mm_add_epi16(tl, bl),
+          _mm_add_epi16(_mm_srli_si128(tl, 4), _mm_srli_si128(bl, 4)));
+
+      __m128i res = _mm_add_epi16(_mm_mullo_epi16(x, center_),
+                                  _mm_add_epi16(_mm_mullo_epi16(y, edge_),
+                                                _mm_mullo_epi16(z, corner_)));
+
+      _mm_storeu_si128((__m128i *)&dst[l], _mm_cvtepi16_epi32(res));
+    }
+    // Handle any leftover pixels
+    for (; j < width - 1; ++j) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] +
+          edge * (dgd[k - stride] + dgd[k - 1] + dgd[k + stride] + dgd[k + 1]) +
+          corner * (dgd[k + stride - 1] + dgd[k - stride - 1] +
+                    dgd[k - stride + 1] + dgd[k + stride + 1]);
+    }
+    j = width - 1;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] +
+          edge * (dgd[k - stride] + dgd[k - 1] + dgd[k + stride] + dgd[k]) +
+          corner * (dgd[k + stride - 1] + dgd[k - stride - 1] +
+                    dgd[k - stride] + dgd[k + stride]);
+    }
+  }
+}
+
+void apply_selfguided_restoration_sse4_1(uint8_t *dat, int width, int height,
+                                         int stride, int eps, int *xqd,
+                                         uint8_t *dst, int dst_stride,
+                                         int32_t *tmpbuf) {
+  int xq[2];
+  int32_t *flt1 = tmpbuf;
+  int32_t *flt2 = flt1 + RESTORATION_TILEPELS_MAX;
+  int32_t *tmpbuf2 = flt2 + RESTORATION_TILEPELS_MAX;
+  int i, j;
+  assert(width * height <= RESTORATION_TILEPELS_MAX);
+#if USE_HIGHPASS_IN_SGRPROJ
+  av1_highpass_filter_sse4_1(dat, width, height, stride, flt1, width,
+                             sgr_params[eps].corner, sgr_params[eps].edge);
+#else
+    av1_selfguided_restoration_sse4_1(dat, width, height, stride, flt1, width,
+                                      sgr_params[eps].r1, sgr_params[eps].e1,
+                                      tmpbuf2);
+#endif  // USE_HIGHPASS_IN_SGRPROJ
+  av1_selfguided_restoration_sse4_1(dat, width, height, stride, flt2, width,
+                                    sgr_params[eps].r2, sgr_params[eps].e2,
+                                    tmpbuf2);
+  decode_xq(xqd, xq);
+
+  __m128i xq0 = _mm_set1_epi32(xq[0]);
+  __m128i xq1 = _mm_set1_epi32(xq[1]);
+  for (i = 0; i < height; ++i) {
+    // Calculate output in batches of 8 pixels
+    for (j = 0; j < width; j += 8) {
+      const int k = i * width + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      __m128i src =
+          _mm_slli_epi16(_mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i *)&dat[l])),
+                         SGRPROJ_RST_BITS);
+
+      const __m128i u_0 = _mm_cvtepu16_epi32(src);
+      const __m128i u_1 = _mm_cvtepu16_epi32(_mm_srli_si128(src, 8));
+
+      const __m128i f1_0 =
+          _mm_sub_epi32(_mm_loadu_si128((__m128i *)&flt1[k]), u_0);
+      const __m128i f2_0 =
+          _mm_sub_epi32(_mm_loadu_si128((__m128i *)&flt2[k]), u_0);
+      const __m128i f1_1 =
+          _mm_sub_epi32(_mm_loadu_si128((__m128i *)&flt1[k + 4]), u_1);
+      const __m128i f2_1 =
+          _mm_sub_epi32(_mm_loadu_si128((__m128i *)&flt2[k + 4]), u_1);
+
+      const __m128i v_0 = _mm_add_epi32(
+          _mm_add_epi32(_mm_mullo_epi32(xq0, f1_0), _mm_mullo_epi32(xq1, f2_0)),
+          _mm_slli_epi32(u_0, SGRPROJ_PRJ_BITS));
+      const __m128i v_1 = _mm_add_epi32(
+          _mm_add_epi32(_mm_mullo_epi32(xq0, f1_1), _mm_mullo_epi32(xq1, f2_1)),
+          _mm_slli_epi32(u_1, SGRPROJ_PRJ_BITS));
+
+      const __m128i rounding =
+          _mm_set1_epi32((1 << (SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS)) >> 1);
+      const __m128i w_0 = _mm_srai_epi32(_mm_add_epi32(v_0, rounding),
+                                         SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+      const __m128i w_1 = _mm_srai_epi32(_mm_add_epi32(v_1, rounding),
+                                         SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+
+      const __m128i tmp = _mm_packs_epi32(w_0, w_1);
+      const __m128i res = _mm_packus_epi16(tmp, tmp /* "don't care" value */);
+      _mm_storel_epi64((__m128i *)&dst[m], res);
+    }
+    // Process leftover pixels
+    for (; j < width; ++j) {
+      const int k = i * width + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int32_t u = ((int32_t)dat[l] << SGRPROJ_RST_BITS);
+      const int32_t f1 = (int32_t)flt1[k] - u;
+      const int32_t f2 = (int32_t)flt2[k] - u;
+      const int32_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
+      const int16_t w =
+          (int16_t)ROUND_POWER_OF_TWO(v, SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+      dst[m] = (uint16_t)clip_pixel(w);
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+// Only the vertical sums need to be adjusted for highbitdepth
+
+static void highbd_selfguided_restoration_1_v(uint16_t *src, int width,
+                                              int height, int src_stride,
+                                              int32_t *A, int32_t *B,
+                                              int buf_stride) {
+  int i, j;
+
+  int width_extend = (width + 3) & ~3;
+  for (j = 0; j < width_extend; j += 4) {
+    __m128i a, b, x, y, x2, y2;
+    __m128i sum, sum_sq, tmp;
+
+    a = _mm_loadl_epi64((__m128i *)&src[j]);
+    b = _mm_loadl_epi64((__m128i *)&src[src_stride + j]);
+
+    sum = _mm_cvtepi16_epi32(_mm_add_epi16(a, b));
+    tmp = _mm_unpacklo_epi16(a, b);
+    sum_sq = _mm_madd_epi16(tmp, tmp);
+
+    _mm_store_si128((__m128i *)&B[j], sum);
+    _mm_store_si128((__m128i *)&A[j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[2 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    for (i = 1; i < height - 2; ++i) {
+      _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+      _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+      x = _mm_cvtepu16_epi32(
+          _mm_loadl_epi64((__m128i *)&src[(i - 1) * src_stride + j]));
+      y = _mm_cvtepu16_epi32(
+          _mm_loadl_epi64((__m128i *)&src[(i + 2) * src_stride + j]));
+
+      sum = _mm_add_epi32(sum, _mm_sub_epi32(y, x));
+
+      x2 = _mm_mullo_epi32(x, x);
+      y2 = _mm_mullo_epi32(y, y);
+
+      sum_sq = _mm_add_epi32(sum_sq, _mm_sub_epi32(y2, x2));
+    }
+    _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 1) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 1) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 1) * buf_stride + j], sum_sq);
+  }
+}
+
+static void highbd_selfguided_restoration_2_v(uint16_t *src, int width,
+                                              int height, int src_stride,
+                                              int32_t *A, int32_t *B,
+                                              int buf_stride) {
+  int i, j;
+
+  int width_extend = (width + 3) & ~3;
+  for (j = 0; j < width_extend; j += 4) {
+    __m128i a, b, c, c2, x, y, x2, y2;
+    __m128i sum, sum_sq, tmp;
+
+    a = _mm_loadl_epi64((__m128i *)&src[j]);
+    b = _mm_loadl_epi64((__m128i *)&src[src_stride + j]);
+    c = _mm_loadl_epi64((__m128i *)&src[2 * src_stride + j]);
+
+    sum = _mm_cvtepi16_epi32(_mm_add_epi16(_mm_add_epi16(a, b), c));
+    // Important: We need to widen *before* squaring here, since
+    // c^2 may be up to 2^24.
+    c = _mm_cvtepu16_epi32(c);
+    c2 = _mm_mullo_epi32(c, c);
+    tmp = _mm_unpacklo_epi16(a, b);
+    sum_sq = _mm_add_epi32(_mm_madd_epi16(tmp, tmp), c2);
+
+    _mm_store_si128((__m128i *)&B[j], sum);
+    _mm_store_si128((__m128i *)&A[j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[3 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[4 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    for (i = 2; i < height - 3; ++i) {
+      _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+      _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+      x = _mm_cvtepu16_epi32(
+          _mm_loadl_epi64((__m128i *)&src[(i - 2) * src_stride + j]));
+      y = _mm_cvtepu16_epi32(
+          _mm_loadl_epi64((__m128i *)&src[(i + 3) * src_stride + j]));
+
+      sum = _mm_add_epi32(sum, _mm_sub_epi32(y, x));
+
+      x2 = _mm_mullo_epi32(x, x);
+      y2 = _mm_mullo_epi32(y, y);
+
+      sum_sq = _mm_add_epi32(sum_sq, _mm_sub_epi32(y2, x2));
+    }
+    _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 2) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 1) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 1) * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 1) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 2) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 2) * buf_stride + j], sum_sq);
+  }
+}
+
+static void highbd_selfguided_restoration_3_v(uint16_t *src, int width,
+                                              int height, int src_stride,
+                                              int32_t *A, int32_t *B,
+                                              int buf_stride) {
+  int i, j;
+
+  int width_extend = (width + 3) & ~3;
+  for (j = 0; j < width_extend; j += 4) {
+    __m128i a, b, c, d, x, y, x2, y2;
+    __m128i sum, sum_sq, tmp, tmp2;
+
+    a = _mm_loadl_epi64((__m128i *)&src[j]);
+    b = _mm_loadl_epi64((__m128i *)&src[src_stride + j]);
+    c = _mm_loadl_epi64((__m128i *)&src[2 * src_stride + j]);
+    d = _mm_loadl_epi64((__m128i *)&src[3 * src_stride + j]);
+
+    sum = _mm_cvtepi16_epi32(
+        _mm_add_epi16(_mm_add_epi16(a, b), _mm_add_epi16(c, d)));
+    tmp = _mm_unpacklo_epi16(a, b);
+    tmp2 = _mm_unpacklo_epi16(c, d);
+    sum_sq =
+        _mm_add_epi32(_mm_madd_epi16(tmp, tmp), _mm_madd_epi16(tmp2, tmp2));
+
+    _mm_store_si128((__m128i *)&B[j], sum);
+    _mm_store_si128((__m128i *)&A[j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[4 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[5 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[2 * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[2 * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[6 * src_stride + j]));
+    sum = _mm_add_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_add_epi32(sum_sq, x2);
+
+    for (i = 3; i < height - 4; ++i) {
+      _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+      _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+      x = _mm_cvtepu16_epi32(
+          _mm_loadl_epi64((__m128i *)&src[(i - 3) * src_stride + j]));
+      y = _mm_cvtepu16_epi32(
+          _mm_loadl_epi64((__m128i *)&src[(i + 4) * src_stride + j]));
+
+      sum = _mm_add_epi32(sum, _mm_sub_epi32(y, x));
+
+      x2 = _mm_mullo_epi32(x, x);
+      y2 = _mm_mullo_epi32(y, y);
+
+      sum_sq = _mm_add_epi32(sum_sq, _mm_sub_epi32(y2, x2));
+    }
+    _mm_store_si128((__m128i *)&B[i * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[i * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 3) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 1) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 1) * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 2) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 2) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 2) * buf_stride + j], sum_sq);
+
+    x = _mm_cvtepu16_epi32(
+        _mm_loadl_epi64((__m128i *)&src[(i - 1) * src_stride + j]));
+    sum = _mm_sub_epi32(sum, x);
+    x2 = _mm_mullo_epi32(x, x);
+    sum_sq = _mm_sub_epi32(sum_sq, x2);
+
+    _mm_store_si128((__m128i *)&B[(i + 3) * buf_stride + j], sum);
+    _mm_store_si128((__m128i *)&A[(i + 3) * buf_stride + j], sum_sq);
+  }
+}
+
+void av1_selfguided_restoration_highbd_sse4_1(uint16_t *dgd, int width,
+                                              int height, int stride,
+                                              int32_t *dst, int dst_stride,
+                                              int bit_depth, int r, int eps,
+                                              int32_t *tmpbuf) {
+  int32_t *A = tmpbuf;
+  int32_t *B = A + SGRPROJ_OUTBUF_SIZE;
+  int i, j;
+  // Adjusting the stride of A and B here appears to avoid bad cache effects,
+  // leading to a significant speed improvement.
+  // We also align the stride to a multiple of 16 bytes for efficiency.
+  int buf_stride = ((width + 3) & ~3) + 16;
+
+  // Don't filter tiles with dimensions < 5 on any axis
+  if ((width < 5) || (height < 5)) return;
+
+  if (r == 1) {
+    highbd_selfguided_restoration_1_v(dgd, width, height, stride, A, B,
+                                      buf_stride);
+    selfguided_restoration_1_h(A, B, width, height, buf_stride, eps, bit_depth);
+  } else if (r == 2) {
+    highbd_selfguided_restoration_2_v(dgd, width, height, stride, A, B,
+                                      buf_stride);
+    selfguided_restoration_2_h(A, B, width, height, buf_stride, eps, bit_depth);
+  } else if (r == 3) {
+    highbd_selfguided_restoration_3_v(dgd, width, height, stride, A, B,
+                                      buf_stride);
+    selfguided_restoration_3_h(A, B, width, height, buf_stride, eps, bit_depth);
+  } else {
+    assert(0);
+  }
+
+  {
+    i = 0;
+    j = 0;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = 3 * A[k] + 2 * A[k + 1] + 2 * A[k + buf_stride] +
+                        A[k + buf_stride + 1];
+      const int32_t b = 3 * B[k] + 2 * B[k + 1] + 2 * B[k + buf_stride] +
+                        B[k + buf_stride + 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = A[k] + 2 * (A[k - 1] + A[k + 1]) + A[k + buf_stride] +
+                        A[k + buf_stride - 1] + A[k + buf_stride + 1];
+      const int32_t b = B[k] + 2 * (B[k - 1] + B[k + 1]) + B[k + buf_stride] +
+                        B[k + buf_stride - 1] + B[k + buf_stride + 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+    j = width - 1;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = 3 * A[k] + 2 * A[k - 1] + 2 * A[k + buf_stride] +
+                        A[k + buf_stride - 1];
+      const int32_t b = 3 * B[k] + 2 * B[k - 1] + 2 * B[k + buf_stride] +
+                        B[k + buf_stride - 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+  }
+  for (i = 1; i < height - 1; ++i) {
+    j = 0;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = A[k] + 2 * (A[k - buf_stride] + A[k + buf_stride]) +
+                        A[k + 1] + A[k - buf_stride + 1] +
+                        A[k + buf_stride + 1];
+      const int32_t b = B[k] + 2 * (B[k - buf_stride] + B[k + buf_stride]) +
+                        B[k + 1] + B[k - buf_stride + 1] +
+                        B[k + buf_stride + 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+
+    // Vectorize the innermost loop
+    for (j = 1; j < width - 1; j += 4) {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 5;
+
+      __m128i tmp0 = _mm_loadu_si128((__m128i *)&A[k - 1 - buf_stride]);
+      __m128i tmp1 = _mm_loadu_si128((__m128i *)&A[k + 3 - buf_stride]);
+      __m128i tmp2 = _mm_loadu_si128((__m128i *)&A[k - 1]);
+      __m128i tmp3 = _mm_loadu_si128((__m128i *)&A[k + 3]);
+      __m128i tmp4 = _mm_loadu_si128((__m128i *)&A[k - 1 + buf_stride]);
+      __m128i tmp5 = _mm_loadu_si128((__m128i *)&A[k + 3 + buf_stride]);
+
+      __m128i a0 = _mm_add_epi32(
+          _mm_add_epi32(_mm_add_epi32(_mm_alignr_epi8(tmp3, tmp2, 4), tmp2),
+                        _mm_add_epi32(_mm_alignr_epi8(tmp3, tmp2, 8),
+                                      _mm_alignr_epi8(tmp5, tmp4, 4))),
+          _mm_alignr_epi8(tmp1, tmp0, 4));
+      __m128i a1 = _mm_add_epi32(_mm_add_epi32(tmp0, tmp4),
+                                 _mm_add_epi32(_mm_alignr_epi8(tmp1, tmp0, 8),
+                                               _mm_alignr_epi8(tmp5, tmp4, 8)));
+      __m128i a = _mm_sub_epi32(_mm_slli_epi32(_mm_add_epi32(a0, a1), 2), a1);
+
+      __m128i tmp6 = _mm_loadu_si128((__m128i *)&B[k - 1 - buf_stride]);
+      __m128i tmp7 = _mm_loadu_si128((__m128i *)&B[k + 3 - buf_stride]);
+      __m128i tmp8 = _mm_loadu_si128((__m128i *)&B[k - 1]);
+      __m128i tmp9 = _mm_loadu_si128((__m128i *)&B[k + 3]);
+      __m128i tmp10 = _mm_loadu_si128((__m128i *)&B[k - 1 + buf_stride]);
+      __m128i tmp11 = _mm_loadu_si128((__m128i *)&B[k + 3 + buf_stride]);
+
+      __m128i b0 = _mm_add_epi32(
+          _mm_add_epi32(_mm_add_epi32(_mm_alignr_epi8(tmp9, tmp8, 4), tmp8),
+                        _mm_add_epi32(_mm_alignr_epi8(tmp9, tmp8, 8),
+                                      _mm_alignr_epi8(tmp11, tmp10, 4))),
+          _mm_alignr_epi8(tmp7, tmp6, 4));
+      __m128i b1 =
+          _mm_add_epi32(_mm_add_epi32(tmp6, tmp10),
+                        _mm_add_epi32(_mm_alignr_epi8(tmp7, tmp6, 8),
+                                      _mm_alignr_epi8(tmp11, tmp10, 8)));
+      __m128i b = _mm_sub_epi32(_mm_slli_epi32(_mm_add_epi32(b0, b1), 2), b1);
+
+      __m128i src = _mm_cvtepu16_epi32(_mm_loadu_si128((__m128i *)&dgd[l]));
+
+      __m128i rounding = _mm_set1_epi32(
+          (1 << (SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS)) >> 1);
+      __m128i v = _mm_add_epi32(_mm_mullo_epi32(a, src), b);
+      __m128i w = _mm_srai_epi32(_mm_add_epi32(v, rounding),
+                                 SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+      _mm_storeu_si128((__m128i *)&dst[m], w);
+    }
+
+    // Deal with any extra pixels at the right-hand edge of the frame
+    // (typically have 2 such pixels, but may have anywhere between 0 and 3)
+    for (; j < width - 1; ++j) {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 5;
+      const int32_t a =
+          (A[k] + A[k - 1] + A[k + 1] + A[k - buf_stride] + A[k + buf_stride]) *
+              4 +
+          (A[k - 1 - buf_stride] + A[k - 1 + buf_stride] +
+           A[k + 1 - buf_stride] + A[k + 1 + buf_stride]) *
+              3;
+      const int32_t b =
+          (B[k] + B[k - 1] + B[k + 1] + B[k - buf_stride] + B[k + buf_stride]) *
+              4 +
+          (B[k - 1 - buf_stride] + B[k - 1 + buf_stride] +
+           B[k + 1 - buf_stride] + B[k + 1 + buf_stride]) *
+              3;
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+
+    j = width - 1;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = A[k] + 2 * (A[k - buf_stride] + A[k + buf_stride]) +
+                        A[k - 1] + A[k - buf_stride - 1] +
+                        A[k + buf_stride - 1];
+      const int32_t b = B[k] + 2 * (B[k - buf_stride] + B[k + buf_stride]) +
+                        B[k - 1] + B[k - buf_stride - 1] +
+                        B[k + buf_stride - 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+  }
+
+  {
+    i = height - 1;
+    j = 0;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = 3 * A[k] + 2 * A[k + 1] + 2 * A[k - buf_stride] +
+                        A[k - buf_stride + 1];
+      const int32_t b = 3 * B[k] + 2 * B[k + 1] + 2 * B[k - buf_stride] +
+                        B[k - buf_stride + 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = A[k] + 2 * (A[k - 1] + A[k + 1]) + A[k - buf_stride] +
+                        A[k - buf_stride - 1] + A[k - buf_stride + 1];
+      const int32_t b = B[k] + 2 * (B[k - 1] + B[k + 1]) + B[k - buf_stride] +
+                        B[k - buf_stride - 1] + B[k - buf_stride + 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+    j = width - 1;
+    {
+      const int k = i * buf_stride + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int nb = 3;
+      const int32_t a = 3 * A[k] + 2 * A[k - 1] + 2 * A[k - buf_stride] +
+                        A[k - buf_stride - 1];
+      const int32_t b = 3 * B[k] + 2 * B[k - 1] + 2 * B[k - buf_stride] +
+                        B[k - buf_stride - 1];
+      const int32_t v = a * dgd[l] + b;
+      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+    }
+  }
+}
+
+void av1_highpass_filter_highbd_sse4_1(uint16_t *dgd, int width, int height,
+                                       int stride, int32_t *dst, int dst_stride,
+                                       int corner, int edge) {
+  int i, j;
+  const int center = (1 << SGRPROJ_RST_BITS) - 4 * (corner + edge);
+
+  {
+    i = 0;
+    j = 0;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] + edge * (dgd[k + 1] + dgd[k + stride] + dgd[k] * 2) +
+          corner *
+              (dgd[k + stride + 1] + dgd[k + 1] + dgd[k + stride] + dgd[k]);
+    }
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] = center * dgd[k] +
+               edge * (dgd[k - 1] + dgd[k + stride] + dgd[k + 1] + dgd[k]) +
+               corner * (dgd[k + stride - 1] + dgd[k + stride + 1] +
+                         dgd[k - 1] + dgd[k + 1]);
+    }
+    j = width - 1;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] + edge * (dgd[k - 1] + dgd[k + stride] + dgd[k] * 2) +
+          corner *
+              (dgd[k + stride - 1] + dgd[k - 1] + dgd[k + stride] + dgd[k]);
+    }
+  }
+  __m128i center_ = _mm_set1_epi32(center);
+  __m128i edge_ = _mm_set1_epi32(edge);
+  __m128i corner_ = _mm_set1_epi32(corner);
+  for (i = 1; i < height - 1; ++i) {
+    j = 0;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] +
+          edge * (dgd[k - stride] + dgd[k + 1] + dgd[k + stride] + dgd[k]) +
+          corner * (dgd[k + stride + 1] + dgd[k - stride + 1] +
+                    dgd[k - stride] + dgd[k + stride]);
+    }
+    // Process 4 pixels at a time
+    for (j = 1; j < width - 4; j += 4) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+
+      __m128i a = _mm_loadu_si128((__m128i *)&dgd[k - stride - 1]);
+      __m128i b = _mm_loadu_si128((__m128i *)&dgd[k - 1]);
+      __m128i c = _mm_loadu_si128((__m128i *)&dgd[k + stride - 1]);
+
+      __m128i tl = _mm_cvtepu16_epi32(a);
+      __m128i tr = _mm_cvtepu16_epi32(_mm_srli_si128(a, 8));
+      __m128i cl = _mm_cvtepu16_epi32(b);
+      __m128i cr = _mm_cvtepu16_epi32(_mm_srli_si128(b, 8));
+      __m128i bl = _mm_cvtepu16_epi32(c);
+      __m128i br = _mm_cvtepu16_epi32(_mm_srli_si128(c, 8));
+
+      __m128i x = _mm_alignr_epi8(cr, cl, 4);
+      __m128i y = _mm_add_epi32(_mm_add_epi32(_mm_alignr_epi8(tr, tl, 4), cl),
+                                _mm_add_epi32(_mm_alignr_epi8(br, bl, 4),
+                                              _mm_alignr_epi8(cr, cl, 8)));
+      __m128i z = _mm_add_epi32(_mm_add_epi32(tl, bl),
+                                _mm_add_epi32(_mm_alignr_epi8(tr, tl, 8),
+                                              _mm_alignr_epi8(br, bl, 8)));
+
+      __m128i res = _mm_add_epi32(_mm_mullo_epi32(x, center_),
+                                  _mm_add_epi32(_mm_mullo_epi32(y, edge_),
+                                                _mm_mullo_epi32(z, corner_)));
+
+      _mm_storeu_si128((__m128i *)&dst[l], res);
+    }
+    // Handle any leftover pixels
+    for (; j < width - 1; ++j) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] +
+          edge * (dgd[k - stride] + dgd[k - 1] + dgd[k + stride] + dgd[k + 1]) +
+          corner * (dgd[k + stride - 1] + dgd[k - stride - 1] +
+                    dgd[k - stride + 1] + dgd[k + stride + 1]);
+    }
+    j = width - 1;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] +
+          edge * (dgd[k - stride] + dgd[k - 1] + dgd[k + stride] + dgd[k]) +
+          corner * (dgd[k + stride - 1] + dgd[k - stride - 1] +
+                    dgd[k - stride] + dgd[k + stride]);
+    }
+  }
+  {
+    i = height - 1;
+    j = 0;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] + edge * (dgd[k + 1] + dgd[k - stride] + dgd[k] * 2) +
+          corner *
+              (dgd[k - stride + 1] + dgd[k + 1] + dgd[k - stride] + dgd[k]);
+    }
+    for (j = 1; j < width - 1; ++j) {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] = center * dgd[k] +
+               edge * (dgd[k - 1] + dgd[k - stride] + dgd[k + 1] + dgd[k]) +
+               corner * (dgd[k - stride - 1] + dgd[k - stride + 1] +
+                         dgd[k - 1] + dgd[k + 1]);
+    }
+    j = width - 1;
+    {
+      const int k = i * stride + j;
+      const int l = i * dst_stride + j;
+      dst[l] =
+          center * dgd[k] + edge * (dgd[k - 1] + dgd[k - stride] + dgd[k] * 2) +
+          corner *
+              (dgd[k - stride - 1] + dgd[k - 1] + dgd[k - stride] + dgd[k]);
+    }
+  }
+}
+
+void apply_selfguided_restoration_highbd_sse4_1(
+    uint16_t *dat, int width, int height, int stride, int bit_depth, int eps,
+    int *xqd, uint16_t *dst, int dst_stride, int32_t *tmpbuf) {
+  int xq[2];
+  int32_t *flt1 = tmpbuf;
+  int32_t *flt2 = flt1 + RESTORATION_TILEPELS_MAX;
+  int32_t *tmpbuf2 = flt2 + RESTORATION_TILEPELS_MAX;
+  int i, j;
+  assert(width * height <= RESTORATION_TILEPELS_MAX);
+#if USE_HIGHPASS_IN_SGRPROJ
+  av1_highpass_filter_highbd_sse4_1(dat, width, height, stride, flt1, width,
+                                    sgr_params[eps].corner,
+                                    sgr_params[eps].edge);
+#else
+  av1_selfguided_restoration_highbd_sse4_1(dat, width, height, stride, flt1,
+                                           width, bit_depth, sgr_params[eps].r1,
+                                           sgr_params[eps].e1, tmpbuf2);
+#endif  // USE_HIGHPASS_IN_SGRPROJ
+  av1_selfguided_restoration_highbd_sse4_1(dat, width, height, stride, flt2,
+                                           width, bit_depth, sgr_params[eps].r2,
+                                           sgr_params[eps].e2, tmpbuf2);
+  decode_xq(xqd, xq);
+
+  __m128i xq0 = _mm_set1_epi32(xq[0]);
+  __m128i xq1 = _mm_set1_epi32(xq[1]);
+  for (i = 0; i < height; ++i) {
+    // Calculate output in batches of 8 pixels
+    for (j = 0; j < width; j += 8) {
+      const int k = i * width + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      __m128i src =
+          _mm_slli_epi16(_mm_load_si128((__m128i *)&dat[l]), SGRPROJ_RST_BITS);
+
+      const __m128i u_0 = _mm_cvtepu16_epi32(src);
+      const __m128i u_1 = _mm_cvtepu16_epi32(_mm_srli_si128(src, 8));
+
+      const __m128i f1_0 =
+          _mm_sub_epi32(_mm_loadu_si128((__m128i *)&flt1[k]), u_0);
+      const __m128i f2_0 =
+          _mm_sub_epi32(_mm_loadu_si128((__m128i *)&flt2[k]), u_0);
+      const __m128i f1_1 =
+          _mm_sub_epi32(_mm_loadu_si128((__m128i *)&flt1[k + 4]), u_1);
+      const __m128i f2_1 =
+          _mm_sub_epi32(_mm_loadu_si128((__m128i *)&flt2[k + 4]), u_1);
+
+      const __m128i v_0 = _mm_add_epi32(
+          _mm_add_epi32(_mm_mullo_epi32(xq0, f1_0), _mm_mullo_epi32(xq1, f2_0)),
+          _mm_slli_epi32(u_0, SGRPROJ_PRJ_BITS));
+      const __m128i v_1 = _mm_add_epi32(
+          _mm_add_epi32(_mm_mullo_epi32(xq0, f1_1), _mm_mullo_epi32(xq1, f2_1)),
+          _mm_slli_epi32(u_1, SGRPROJ_PRJ_BITS));
+
+      const __m128i rounding =
+          _mm_set1_epi32((1 << (SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS)) >> 1);
+      const __m128i w_0 = _mm_srai_epi32(_mm_add_epi32(v_0, rounding),
+                                         SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+      const __m128i w_1 = _mm_srai_epi32(_mm_add_epi32(v_1, rounding),
+                                         SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+
+      // Pack into 16 bits and clamp to [0, 2^bit_depth)
+      const __m128i tmp = _mm_packus_epi32(w_0, w_1);
+      const __m128i max = _mm_set1_epi16((1 << bit_depth) - 1);
+      const __m128i res = _mm_min_epi16(tmp, max);
+
+      _mm_store_si128((__m128i *)&dst[m], res);
+    }
+    // Process leftover pixels
+    for (; j < width; ++j) {
+      const int k = i * width + j;
+      const int l = i * stride + j;
+      const int m = i * dst_stride + j;
+      const int32_t u = ((int32_t)dat[l] << SGRPROJ_RST_BITS);
+      const int32_t f1 = (int32_t)flt1[k] - u;
+      const int32_t f2 = (int32_t)flt2[k] - u;
+      const int32_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
+      const int16_t w =
+          (int16_t)ROUND_POWER_OF_TWO(v, SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+      dst[m] = (uint16_t)clip_pixel_highbd(w, bit_depth);
+    }
+  }
+}
+
+#endif
diff --git a/third_party/aom/av1/common/x86/warp_plane_sse2.c b/third_party/aom/av1/common/x86/warp_plane_sse2.c
new file mode 100644
index 0000000000..925e4650d9
--- /dev/null
+++ b/third_party/aom/av1/common/x86/warp_plane_sse2.c
@@ -0,0 +1,297 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>
+
+#include "./av1_rtcd.h"
+#include "av1/common/warped_motion.h"
+
+static const __m128i *const filter = (const __m128i *const)warped_filter;
+
+/* SSE2 version of the rotzoom/affine warp filter */
+void av1_warp_affine_sse2(int32_t *mat, uint8_t *ref, int width, int height,
+                          int stride, uint8_t *pred, int p_col, int p_row,
+                          int p_width, int p_height, int p_stride,
+                          int subsampling_x, int subsampling_y, int ref_frm,
+                          int16_t alpha, int16_t beta, int16_t gamma,
+                          int16_t delta) {
+  __m128i tmp[15];
+  int i, j, k;
+
+  /* Note: For this code to work, the left/right frame borders need to be
+     extended by at least 13 pixels each. By the time we get here, other
+     code will have set up this border, but we allow an explicit check
+     for debugging purposes.
+  */
+  /*for (i = 0; i < height; ++i) {
+    for (j = 0; j < 13; ++j) {
+      assert(ref[i * stride - 13 + j] == ref[i * stride]);
+      assert(ref[i * stride + width + j] == ref[i * stride + (width - 1)]);
+    }
+  }*/
+
+  for (i = 0; i < p_height; i += 8) {
+    for (j = 0; j < p_width; j += 8) {
+      // (x, y) coordinates of the center of this block in the destination
+      // image
+      int32_t dst_x = p_col + j + 4;
+      int32_t dst_y = p_row + i + 4;
+
+      int32_t x4, y4, ix4, sx4, iy4, sy4;
+      if (subsampling_x)
+        x4 = ROUND_POWER_OF_TWO_SIGNED(
+            mat[2] * 2 * dst_x + mat[3] * 2 * dst_y + mat[0] +
+                (mat[2] + mat[3] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+            1);
+      else
+        x4 = mat[2] * dst_x + mat[3] * dst_y + mat[0];
+
+      if (subsampling_y)
+        y4 = ROUND_POWER_OF_TWO_SIGNED(
+            mat[4] * 2 * dst_x + mat[5] * 2 * dst_y + mat[1] +
+                (mat[4] + mat[5] - (1 << WARPEDMODEL_PREC_BITS)) / 2,
+            1);
+      else
+        y4 = mat[4] * dst_x + mat[5] * dst_y + mat[1];
+
+      ix4 = x4 >> WARPEDMODEL_PREC_BITS;
+      sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+      iy4 = y4 >> WARPEDMODEL_PREC_BITS;
+      sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+
+      // Horizontal filter
+      for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+        int iy = iy4 + k;
+        if (iy < 0)
+          iy = 0;
+        else if (iy > height - 1)
+          iy = height - 1;
+
+        // If the block is aligned such that, after clamping, every sample
+        // would be taken from the leftmost/rightmost column, then we can
+        // skip the expensive horizontal filter.
+        if (ix4 <= -7) {
+          tmp[k + 7] = _mm_set1_epi16(
+              ref[iy * stride] *
+              (1 << (WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS)));
+        } else if (ix4 >= width + 6) {
+          tmp[k + 7] = _mm_set1_epi16(
+              ref[iy * stride + (width - 1)] *
+              (1 << (WARPEDPIXEL_FILTER_BITS - HORSHEAR_REDUCE_PREC_BITS)));
+        } else {
+          int sx = sx4 + alpha * (-4) + beta * k +
+                   // Include rounding and offset here
+                   (1 << (WARPEDDIFF_PREC_BITS - 1)) +
+                   (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
+
+          // Load source pixels
+          __m128i zero = _mm_setzero_si128();
+          __m128i src =
+              _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+
+          // Filter even-index pixels
+          __m128i tmp_0 = _mm_loadu_si128(
+              (__m128i *)(filter + ((sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS)));
+          __m128i tmp_2 = _mm_loadu_si128(
+              (__m128i *)(filter + ((sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS)));
+          __m128i tmp_4 = _mm_loadu_si128(
+              (__m128i *)(filter + ((sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS)));
+          __m128i tmp_6 = _mm_loadu_si128(
+              (__m128i *)(filter + ((sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS)));
+
+          // coeffs 0 1 0 1 2 3 2 3 for pixels 0, 2
+          __m128i tmp_8 = _mm_unpacklo_epi32(tmp_0, tmp_2);
+          // coeffs 0 1 0 1 2 3 2 3 for pixels 4, 6
+          __m128i tmp_10 = _mm_unpacklo_epi32(tmp_4, tmp_6);
+          // coeffs 4 5 4 5 6 7 6 7 for pixels 0, 2
+          __m128i tmp_12 = _mm_unpackhi_epi32(tmp_0, tmp_2);
+          // coeffs 4 5 4 5 6 7 6 7 for pixels 4, 6
+          __m128i tmp_14 = _mm_unpackhi_epi32(tmp_4, tmp_6);
+
+          // coeffs 0 1 0 1 0 1 0 1 for pixels 0, 2, 4, 6
+          __m128i coeff_0 = _mm_unpacklo_epi64(tmp_8, tmp_10);
+          // coeffs 2 3 2 3 2 3 2 3 for pixels 0, 2, 4, 6
+          __m128i coeff_2 = _mm_unpackhi_epi64(tmp_8, tmp_10);
+          // coeffs 4 5 4 5 4 5 4 5 for pixels 0, 2, 4, 6
+          __m128i coeff_4 = _mm_unpacklo_epi64(tmp_12, tmp_14);
+          // coeffs 6 7 6 7 6 7 6 7 for pixels 0, 2, 4, 6
+          __m128i coeff_6 = _mm_unpackhi_epi64(tmp_12, tmp_14);
+
+          __m128i round_const =
+              _mm_set1_epi32((1 << HORSHEAR_REDUCE_PREC_BITS) >> 1);
+
+          // Calculate filtered results
+          __m128i src_0 = _mm_unpacklo_epi8(src, zero);
+          __m128i res_0 = _mm_madd_epi16(src_0, coeff_0);
+          __m128i src_2 = _mm_unpacklo_epi8(_mm_srli_si128(src, 2), zero);
+          __m128i res_2 = _mm_madd_epi16(src_2, coeff_2);
+          __m128i src_4 = _mm_unpacklo_epi8(_mm_srli_si128(src, 4), zero);
+          __m128i res_4 = _mm_madd_epi16(src_4, coeff_4);
+          __m128i src_6 = _mm_unpacklo_epi8(_mm_srli_si128(src, 6), zero);
+          __m128i res_6 = _mm_madd_epi16(src_6, coeff_6);
+
+          __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
+                                           _mm_add_epi32(res_2, res_6));
+          res_even = _mm_srai_epi32(_mm_add_epi32(res_even, round_const),
+                                    HORSHEAR_REDUCE_PREC_BITS);
+
+          // Filter odd-index pixels
+          __m128i tmp_1 = _mm_loadu_si128(
+              (__m128i *)(filter + ((sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS)));
+          __m128i tmp_3 = _mm_loadu_si128(
+              (__m128i *)(filter + ((sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS)));
+          __m128i tmp_5 = _mm_loadu_si128(
+              (__m128i *)(filter + ((sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS)));
+          __m128i tmp_7 = _mm_loadu_si128(
+              (__m128i *)(filter + ((sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS)));
+
+          __m128i tmp_9 = _mm_unpacklo_epi32(tmp_1, tmp_3);
+          __m128i tmp_11 = _mm_unpacklo_epi32(tmp_5, tmp_7);
+          __m128i tmp_13 = _mm_unpackhi_epi32(tmp_1, tmp_3);
+          __m128i tmp_15 = _mm_unpackhi_epi32(tmp_5, tmp_7);
+
+          __m128i coeff_1 = _mm_unpacklo_epi64(tmp_9, tmp_11);
+          __m128i coeff_3 = _mm_unpackhi_epi64(tmp_9, tmp_11);
+          __m128i coeff_5 = _mm_unpacklo_epi64(tmp_13, tmp_15);
+          __m128i coeff_7 = _mm_unpackhi_epi64(tmp_13, tmp_15);
+
+          __m128i src_1 = _mm_unpacklo_epi8(_mm_srli_si128(src, 1), zero);
+          __m128i res_1 = _mm_madd_epi16(src_1, coeff_1);
+          __m128i src_3 = _mm_unpacklo_epi8(_mm_srli_si128(src, 3), zero);
+          __m128i res_3 = _mm_madd_epi16(src_3, coeff_3);
+          __m128i src_5 = _mm_unpacklo_epi8(_mm_srli_si128(src, 5), zero);
+          __m128i res_5 = _mm_madd_epi16(src_5, coeff_5);
+          __m128i src_7 = _mm_unpacklo_epi8(_mm_srli_si128(src, 7), zero);
+          __m128i res_7 = _mm_madd_epi16(src_7, coeff_7);
+
+          __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
+                                          _mm_add_epi32(res_3, res_7));
+          res_odd = _mm_srai_epi32(_mm_add_epi32(res_odd, round_const),
+                                   HORSHEAR_REDUCE_PREC_BITS);
+
+          // Combine results into one register.
+          // We store the columns in the order 0, 2, 4, 6, 1, 3, 5, 7
+          // as this order helps with the vertical filter.
+          tmp[k + 7] = _mm_packs_epi32(res_even, res_odd);
+        }
+      }
+
+      // Vertical filter
+      for (k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
+        int sy = sy4 + gamma * (-4) + delta * k +
+                 (1 << (WARPEDDIFF_PREC_BITS - 1)) +
+                 (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
+
+        // Load from tmp and rearrange pairs of consecutive rows into the
+        // column order 0 0 2 2 4 4 6 6; 1 1 3 3 5 5 7 7
+        __m128i *src = tmp + (k + 4);
+        __m128i src_0 = _mm_unpacklo_epi16(src[0], src[1]);
+        __m128i src_2 = _mm_unpacklo_epi16(src[2], src[3]);
+        __m128i src_4 = _mm_unpacklo_epi16(src[4], src[5]);
+        __m128i src_6 = _mm_unpacklo_epi16(src[6], src[7]);
+
+        // Filter even-index pixels
+        __m128i tmp_0 = _mm_loadu_si128(
+            (__m128i *)(filter + ((sy + 0 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        __m128i tmp_2 = _mm_loadu_si128(
+            (__m128i *)(filter + ((sy + 2 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        __m128i tmp_4 = _mm_loadu_si128(
+            (__m128i *)(filter + ((sy + 4 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        __m128i tmp_6 = _mm_loadu_si128(
+            (__m128i *)(filter + ((sy + 6 * gamma) >> WARPEDDIFF_PREC_BITS)));
+
+        __m128i tmp_8 = _mm_unpacklo_epi32(tmp_0, tmp_2);
+        __m128i tmp_10 = _mm_unpacklo_epi32(tmp_4, tmp_6);
+        __m128i tmp_12 = _mm_unpackhi_epi32(tmp_0, tmp_2);
+        __m128i tmp_14 = _mm_unpackhi_epi32(tmp_4, tmp_6);
+
+        __m128i coeff_0 = _mm_unpacklo_epi64(tmp_8, tmp_10);
+        __m128i coeff_2 = _mm_unpackhi_epi64(tmp_8, tmp_10);
+        __m128i coeff_4 = _mm_unpacklo_epi64(tmp_12, tmp_14);
+        __m128i coeff_6 = _mm_unpackhi_epi64(tmp_12, tmp_14);
+
+        __m128i res_0 = _mm_madd_epi16(src_0, coeff_0);
+        __m128i res_2 = _mm_madd_epi16(src_2, coeff_2);
+        __m128i res_4 = _mm_madd_epi16(src_4, coeff_4);
+        __m128i res_6 = _mm_madd_epi16(src_6, coeff_6);
+
+        __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
+                                         _mm_add_epi32(res_4, res_6));
+
+        // Filter odd-index pixels
+        __m128i src_1 = _mm_unpackhi_epi16(src[0], src[1]);
+        __m128i src_3 = _mm_unpackhi_epi16(src[2], src[3]);
+        __m128i src_5 = _mm_unpackhi_epi16(src[4], src[5]);
+        __m128i src_7 = _mm_unpackhi_epi16(src[6], src[7]);
+
+        __m128i tmp_1 = _mm_loadu_si128(
+            (__m128i *)(filter + ((sy + 1 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        __m128i tmp_3 = _mm_loadu_si128(
+            (__m128i *)(filter + ((sy + 3 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        __m128i tmp_5 = _mm_loadu_si128(
+            (__m128i *)(filter + ((sy + 5 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        __m128i tmp_7 = _mm_loadu_si128(
+            (__m128i *)(filter + ((sy + 7 * gamma) >> WARPEDDIFF_PREC_BITS)));
+
+        __m128i tmp_9 = _mm_unpacklo_epi32(tmp_1, tmp_3);
+        __m128i tmp_11 = _mm_unpacklo_epi32(tmp_5, tmp_7);
+        __m128i tmp_13 = _mm_unpackhi_epi32(tmp_1, tmp_3);
+        __m128i tmp_15 = _mm_unpackhi_epi32(tmp_5, tmp_7);
+
+        __m128i coeff_1 = _mm_unpacklo_epi64(tmp_9, tmp_11);
+        __m128i coeff_3 = _mm_unpackhi_epi64(tmp_9, tmp_11);
+        __m128i coeff_5 = _mm_unpacklo_epi64(tmp_13, tmp_15);
+        __m128i coeff_7 = _mm_unpackhi_epi64(tmp_13, tmp_15);
+
+        __m128i res_1 = _mm_madd_epi16(src_1, coeff_1);
+        __m128i res_3 = _mm_madd_epi16(src_3, coeff_3);
+        __m128i res_5 = _mm_madd_epi16(src_5, coeff_5);
+        __m128i res_7 = _mm_madd_epi16(src_7, coeff_7);
+
+        __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
+                                        _mm_add_epi32(res_5, res_7));
+
+        // Rearrange pixels back into the order 0 ... 7
+        __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+        __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+
+        // Round and pack into 8 bits
+        __m128i round_const =
+            _mm_set1_epi32((1 << VERSHEAR_REDUCE_PREC_BITS) >> 1);
+
+        __m128i res_lo_round = _mm_srai_epi32(
+            _mm_add_epi32(res_lo, round_const), VERSHEAR_REDUCE_PREC_BITS);
+        __m128i res_hi_round = _mm_srai_epi32(
+            _mm_add_epi32(res_hi, round_const), VERSHEAR_REDUCE_PREC_BITS);
+
+        __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round);
+        __m128i res_8bit = _mm_packus_epi16(res_16bit, res_16bit);
+
+        // Store, blending with 'pred' if needed
+        __m128i *p = (__m128i *)&pred[(i + k + 4) * p_stride + j];
+
+        // Note: If we're outputting a 4x4 block, we need to be very careful
+        // to only output 4 pixels at this point, to avoid encode/decode
+        // mismatches when encoding with multiple threads.
+        if (p_width == 4) {
+          if (ref_frm) {
+            const __m128i orig = _mm_cvtsi32_si128(*(uint32_t *)p);
+            res_8bit = _mm_avg_epu8(res_8bit, orig);
+          }
+          *(uint32_t *)p = _mm_cvtsi128_si32(res_8bit);
+        } else {
+          if (ref_frm) res_8bit = _mm_avg_epu8(res_8bit, _mm_loadl_epi64(p));
+          _mm_storel_epi64(p, res_8bit);
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/zigzag.h b/third_party/aom/av1/common/zigzag.h
new file mode 100644
index 0000000000..c58b18b571
--- /dev/null
+++ b/third_party/aom/av1/common/zigzag.h
@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_zigzag_H)
+# define _zigzag_H (1)
+
+extern const unsigned char OD_ZIGZAG4_DCT_DCT[15][2];
+extern const unsigned char OD_ZIGZAG4_ADST_DCT[15][2];
+extern const unsigned char OD_ZIGZAG4_DCT_ADST[15][2];
+#define OD_ZIGZAG4_ADST_ADST OD_ZIGZAG4_DCT_DCT
+
+extern const unsigned char OD_ZIGZAG8_DCT_DCT[48][2];
+extern const unsigned char OD_ZIGZAG8_ADST_DCT[48][2];
+extern const unsigned char OD_ZIGZAG8_DCT_ADST[48][2];
+#define OD_ZIGZAG8_ADST_ADST OD_ZIGZAG8_DCT_DCT
+
+extern const unsigned char OD_ZIGZAG16_DCT_DCT[192][2];
+extern const unsigned char OD_ZIGZAG16_ADST_DCT[192][2];
+extern const unsigned char OD_ZIGZAG16_DCT_ADST[192][2];
+#define OD_ZIGZAG16_ADST_ADST OD_ZIGZAG16_DCT_DCT
+
+extern const unsigned char OD_ZIGZAG32_DCT_DCT[768][2];
+#endif
diff --git a/third_party/aom/av1/common/zigzag16.c b/third_party/aom/av1/common/zigzag16.c
new file mode 100644
index 0000000000..6df6e3855f
--- /dev/null
+++ b/third_party/aom/av1/common/zigzag16.c
@@ -0,0 +1,157 @@
+/* This file is generated by gen_zigzag16.m */
+
+/* clang-format off */
+
+#include "odintrin.h"
+OD_EXTERN const unsigned char OD_ZIGZAG16_DCT_DCT[192][2] = {
+  {8, 0}, {8, 1}, {8, 2}, {9, 0},
+  {8, 3}, {9, 1}, {9, 2}, {10, 0},
+  {9, 3}, {10, 1}, {10, 2}, {11, 0},
+  {10, 3}, {11, 1}, {11, 2}, {11, 3},
+  {12, 0}, {12, 1}, {13, 0}, {12, 2},
+  {12, 3}, {13, 1}, {13, 2}, {14, 0},
+  {13, 3}, {14, 1}, {15, 0}, {14, 2},
+  {14, 3}, {15, 1}, {15, 2}, {15, 3},
+  {0, 8}, {1, 8}, {0, 9}, {2, 8},
+  {1, 9}, {3, 8}, {0, 10}, {2, 9},
+  {1, 10}, {3, 9}, {0, 11}, {2, 10},
+  {1, 11}, {3, 10}, {0, 12}, {2, 11},
+  {1, 12}, {3, 11}, {0, 13}, {2, 12},
+  {1, 13}, {0, 14}, {3, 12}, {2, 13},
+  {1, 14}, {3, 13}, {0, 15}, {2, 14},
+  {1, 15}, {3, 14}, {2, 15}, {3, 15},
+  {4, 8}, {5, 8}, {4, 9}, {8, 4},
+  {8, 5}, {6, 8}, {5, 9}, {4, 10},
+  {9, 4}, {8, 6}, {7, 8}, {9, 5},
+  {5, 10}, {8, 7}, {6, 9}, {4, 11},
+  {10, 4}, {9, 6}, {7, 9}, {8, 8},
+  {10, 5}, {6, 10}, {5, 11}, {9, 7},
+  {8, 9}, {10, 6}, {7, 10}, {4, 12},
+  {11, 4}, {9, 8}, {6, 11}, {10, 7},
+  {11, 5}, {5, 12}, {8, 10}, {7, 11},
+  {9, 9}, {4, 13}, {10, 8}, {11, 6},
+  {11, 7}, {6, 12}, {8, 11}, {9, 10},
+  {12, 4}, {5, 13}, {10, 9}, {12, 5},
+  {7, 12}, {11, 8}, {4, 14}, {6, 13},
+  {10, 10}, {9, 11}, {12, 6}, {13, 4},
+  {11, 9}, {8, 12}, {5, 14}, {12, 7},
+  {7, 13}, {4, 15}, {13, 5}, {10, 11},
+  {11, 10}, {9, 12}, {13, 6}, {12, 8},
+  {6, 14}, {8, 13}, {5, 15}, {13, 7},
+  {14, 4}, {12, 9}, {7, 14}, {11, 11},
+  {10, 12}, {9, 13}, {14, 5}, {6, 15},
+  {13, 8}, {8, 14}, {12, 10}, {14, 6},
+  {7, 15}, {13, 9}, {15, 4}, {10, 13},
+  {11, 12}, {14, 7}, {9, 14}, {12, 11},
+  {8, 15}, {15, 5}, {13, 10}, {14, 8},
+  {11, 13}, {15, 6}, {9, 15}, {10, 14},
+  {14, 9}, {15, 7}, {13, 11}, {12, 12},
+  {10, 15}, {11, 14}, {15, 8}, {14, 10},
+  {12, 13}, {13, 12}, {15, 9}, {11, 15},
+  {14, 11}, {13, 13}, {15, 10}, {12, 14},
+  {13, 14}, {15, 11}, {14, 12}, {12, 15},
+  {14, 13}, {13, 15}, {15, 12}, {14, 14},
+  {15, 13}, {14, 15}, {15, 14}, {15, 15}
+  };
+
+OD_EXTERN const unsigned char OD_ZIGZAG16_ADST_DCT[192][2] = {
+  {8, 0}, {9, 0}, {10, 0}, {8, 1},
+  {11, 0}, {9, 1}, {8, 2}, {12, 0},
+  {10, 1}, {9, 2}, {8, 3}, {13, 0},
+  {11, 1}, {10, 2}, {9, 3}, {14, 0},
+  {12, 1}, {10, 3}, {15, 0}, {11, 2},
+  {13, 1}, {11, 3}, {12, 2}, {14, 1},
+  {12, 3}, {13, 2}, {15, 1}, {13, 3},
+  {14, 2}, {14, 3}, {15, 2}, {15, 3},
+  {0, 8}, {1, 8}, {2, 8}, {0, 9},
+  {3, 8}, {1, 9}, {2, 9}, {0, 10},
+  {3, 9}, {1, 10}, {2, 10}, {0, 11},
+  {3, 10}, {1, 11}, {2, 11}, {0, 12},
+  {3, 11}, {1, 12}, {2, 12}, {0, 13},
+  {3, 12}, {1, 13}, {0, 14}, {2, 13},
+  {0, 15}, {1, 14}, {3, 13}, {2, 14},
+  {1, 15}, {3, 14}, {2, 15}, {3, 15},
+  {8, 4}, {9, 4}, {8, 5}, {4, 8},
+  {10, 4}, {9, 5}, {5, 8}, {8, 6},
+  {4, 9}, {10, 5}, {9, 6}, {6, 8},
+  {8, 7}, {11, 4}, {7, 8}, {5, 9},
+  {9, 7}, {11, 5}, {10, 6}, {4, 10},
+  {6, 9}, {8, 8}, {5, 10}, {7, 9},
+  {12, 4}, {10, 7}, {9, 8}, {11, 6},
+  {8, 9}, {4, 11}, {6, 10}, {7, 10},
+  {12, 5}, {5, 11}, {10, 8}, {11, 7},
+  {9, 9}, {4, 12}, {13, 4}, {8, 10},
+  {6, 11}, {12, 6}, {5, 12}, {10, 9},
+  {7, 11}, {9, 10}, {11, 8}, {13, 5},
+  {8, 11}, {4, 13}, {6, 12}, {10, 10},
+  {12, 7}, {11, 9}, {7, 12}, {14, 4},
+  {5, 13}, {9, 11}, {13, 6}, {8, 12},
+  {4, 14}, {12, 8}, {6, 13}, {11, 10},
+  {10, 11}, {12, 9}, {5, 14}, {13, 7},
+  {14, 5}, {9, 12}, {4, 15}, {7, 13},
+  {8, 13}, {6, 14}, {13, 8}, {11, 11},
+  {10, 12}, {15, 4}, {12, 10}, {14, 6},
+  {13, 9}, {5, 15}, {9, 13}, {7, 14},
+  {15, 5}, {6, 15}, {8, 14}, {14, 7},
+  {11, 12}, {7, 15}, {9, 14}, {13, 10},
+  {10, 13}, {14, 8}, {15, 6}, {14, 9},
+  {12, 11}, {8, 15}, {15, 7}, {10, 14},
+  {11, 13}, {9, 15}, {13, 11}, {12, 12},
+  {15, 8}, {14, 10}, {15, 9}, {10, 15},
+  {11, 14}, {13, 12}, {12, 13}, {15, 10},
+  {14, 11}, {11, 15}, {13, 13}, {15, 11},
+  {14, 12}, {12, 14}, {15, 12}, {13, 14},
+  {12, 15}, {14, 13}, {13, 15}, {15, 13},
+  {14, 14}, {15, 14}, {14, 15}, {15, 15}
+  };
+
+OD_EXTERN const unsigned char OD_ZIGZAG16_DCT_ADST[192][2] = {
+  {8, 0}, {8, 1}, {8, 2}, {8, 3},
+  {9, 0}, {9, 1}, {9, 2}, {9, 3},
+  {10, 0}, {10, 1}, {10, 2}, {10, 3},
+  {11, 0}, {11, 1}, {11, 2}, {11, 3},
+  {12, 0}, {12, 1}, {12, 2}, {12, 3},
+  {13, 0}, {13, 1}, {13, 2}, {13, 3},
+  {14, 0}, {15, 0}, {14, 1}, {14, 2},
+  {14, 3}, {15, 1}, {15, 2}, {15, 3},
+  {0, 8}, {0, 9}, {0, 10}, {1, 8},
+  {0, 11}, {1, 9}, {2, 8}, {0, 12},
+  {1, 10}, {2, 9}, {0, 13}, {1, 11},
+  {3, 8}, {2, 10}, {0, 14}, {1, 12},
+  {3, 9}, {0, 15}, {2, 11}, {3, 10},
+  {1, 13}, {2, 12}, {3, 11}, {1, 14},
+  {2, 13}, {1, 15}, {3, 12}, {2, 14},
+  {3, 13}, {2, 15}, {3, 14}, {3, 15},
+  {4, 8}, {4, 9}, {5, 8}, {4, 10},
+  {5, 9}, {4, 11}, {6, 8}, {5, 10},
+  {8, 4}, {6, 9}, {4, 12}, {5, 11},
+  {8, 5}, {6, 10}, {7, 8}, {8, 6},
+  {4, 13}, {7, 9}, {5, 12}, {8, 7},
+  {9, 4}, {6, 11}, {8, 8}, {7, 10},
+  {5, 13}, {9, 5}, {4, 14}, {9, 6},
+  {8, 9}, {6, 12}, {9, 7}, {7, 11},
+  {4, 15}, {8, 10}, {9, 8}, {5, 14},
+  {10, 4}, {6, 13}, {10, 5}, {9, 9},
+  {7, 12}, {8, 11}, {10, 6}, {5, 15},
+  {10, 7}, {6, 14}, {9, 10}, {7, 13},
+  {8, 12}, {10, 8}, {9, 11}, {6, 15},
+  {11, 4}, {11, 5}, {10, 9}, {8, 13},
+  {7, 14}, {11, 6}, {9, 12}, {11, 7},
+  {10, 10}, {7, 15}, {8, 14}, {12, 4},
+  {11, 8}, {12, 5}, {9, 13}, {10, 11},
+  {8, 15}, {11, 9}, {12, 6}, {12, 7},
+  {10, 12}, {9, 14}, {11, 10}, {13, 4},
+  {12, 8}, {9, 15}, {13, 5}, {11, 11},
+  {12, 9}, {10, 13}, {13, 6}, {13, 7},
+  {12, 10}, {14, 4}, {11, 12}, {13, 8},
+  {10, 14}, {14, 5}, {12, 11}, {13, 9},
+  {14, 6}, {10, 15}, {11, 13}, {15, 4},
+  {14, 7}, {12, 12}, {13, 10}, {14, 8},
+  {15, 5}, {13, 11}, {15, 6}, {11, 14},
+  {14, 9}, {12, 13}, {11, 15}, {15, 7},
+  {14, 10}, {15, 8}, {13, 12}, {12, 14},
+  {15, 9}, {14, 11}, {13, 13}, {12, 15},
+  {15, 10}, {14, 12}, {13, 14}, {15, 11},
+  {13, 15}, {14, 13}, {14, 14}, {15, 12},
+  {14, 15}, {15, 13}, {15, 14}, {15, 15}
+  };
diff --git a/third_party/aom/av1/common/zigzag32.c b/third_party/aom/av1/common/zigzag32.c
new file mode 100644
index 0000000000..cb3b9bc632
--- /dev/null
+++ b/third_party/aom/av1/common/zigzag32.c
@@ -0,0 +1,199 @@
+/* This file is generated by gen_zigzag32.m */
+
+/* clang-format off */
+
+#include "odintrin.h"
+OD_EXTERN const unsigned char OD_ZIGZAG32_DCT_DCT[768][2] = {
+  { 16, 0 }, { 17, 0 }, { 18, 0 }, { 19, 0 },
+  { 16, 1 }, { 17, 1 }, { 20, 0 }, { 16, 2 },
+  { 18, 1 }, { 21, 0 }, { 17, 2 }, { 16, 3 },
+  { 19, 1 }, { 22, 0 }, { 18, 2 }, { 17, 3 },
+  { 20, 1 }, { 16, 4 }, { 23, 0 }, { 19, 2 },
+  { 24, 0 }, { 16, 5 }, { 21, 1 }, { 17, 4 },
+  { 18, 3 }, { 20, 2 }, { 17, 5 }, { 16, 6 },
+  { 19, 3 }, { 18, 4 }, { 25, 0 }, { 22, 1 },
+  { 16, 7 }, { 21, 2 }, { 17, 6 }, { 20, 3 },
+  { 26, 0 }, { 18, 5 }, { 19, 4 }, { 17, 7 },
+  { 23, 1 }, { 22, 2 }, { 18, 6 }, { 27, 0 },
+  { 19, 5 }, { 24, 1 }, { 21, 3 }, { 28, 0 },
+  { 20, 4 }, { 18, 7 }, { 19, 6 }, { 23, 2 },
+  { 29, 0 }, { 25, 1 }, { 21, 4 }, { 30, 0 },
+  { 20, 5 }, { 22, 3 }, { 31, 0 }, { 19, 7 },
+  { 24, 2 }, { 26, 1 }, { 20, 6 }, { 21, 5 },
+  { 22, 4 }, { 23, 3 }, { 27, 1 }, { 25, 2 },
+  { 20, 7 }, { 28, 1 }, { 24, 3 }, { 21, 6 },
+  { 22, 5 }, { 23, 4 }, { 26, 2 }, { 21, 7 },
+  { 29, 1 }, { 25, 3 }, { 30, 1 }, { 27, 2 },
+  { 22, 6 }, { 23, 5 }, { 31, 1 }, { 24, 4 },
+  { 26, 3 }, { 28, 2 }, { 22, 7 }, { 23, 6 },
+  { 25, 4 }, { 24, 5 }, { 29, 2 }, { 30, 2 },
+  { 27, 3 }, { 23, 7 }, { 31, 2 }, { 24, 6 },
+  { 26, 4 }, { 25, 5 }, { 28, 3 }, { 24, 7 },
+  { 27, 4 }, { 29, 3 }, { 25, 6 }, { 26, 5 },
+  { 30, 3 }, { 31, 3 }, { 28, 4 }, { 27, 5 },
+  { 25, 7 }, { 29, 4 }, { 26, 6 }, { 28, 5 },
+  { 30, 4 }, { 26, 7 }, { 27, 6 }, { 31, 4 },
+  { 29, 5 }, { 27, 7 }, { 30, 5 }, { 28, 6 },
+  { 31, 5 }, { 29, 6 }, { 28, 7 }, { 30, 6 },
+  { 31, 6 }, { 29, 7 }, { 30, 7 }, { 31, 7 },
+  { 0, 16 }, { 0, 17 }, { 1, 16 }, { 0, 18 },
+  { 1, 17 }, { 0, 19 }, { 2, 16 }, { 1, 18 },
+  { 0, 20 }, { 2, 17 }, { 3, 16 }, { 1, 19 },
+  { 2, 18 }, { 0, 21 }, { 3, 17 }, { 4, 16 },
+  { 1, 20 }, { 2, 19 }, { 0, 22 }, { 3, 18 },
+  { 4, 17 }, { 5, 16 }, { 0, 23 }, { 3, 19 },
+  { 2, 20 }, { 1, 21 }, { 4, 18 }, { 6, 16 },
+  { 5, 17 }, { 3, 20 }, { 2, 21 }, { 1, 22 },
+  { 0, 24 }, { 0, 25 }, { 4, 19 }, { 7, 16 },
+  { 6, 17 }, { 5, 18 }, { 0, 26 }, { 3, 21 },
+  { 2, 22 }, { 1, 23 }, { 4, 20 }, { 5, 19 },
+  { 6, 18 }, { 1, 24 }, { 7, 17 }, { 0, 27 },
+  { 2, 23 }, { 3, 22 }, { 4, 21 }, { 1, 25 },
+  { 5, 20 }, { 7, 18 }, { 0, 28 }, { 6, 19 },
+  { 2, 24 }, { 1, 26 }, { 0, 29 }, { 4, 22 },
+  { 3, 23 }, { 2, 25 }, { 5, 21 }, { 0, 31 },
+  { 7, 19 }, { 6, 20 }, { 0, 30 }, { 1, 27 },
+  { 3, 24 }, { 2, 26 }, { 4, 23 }, { 5, 22 },
+  { 7, 20 }, { 1, 28 }, { 6, 21 }, { 3, 25 },
+  { 2, 27 }, { 1, 29 }, { 4, 24 }, { 2, 28 },
+  { 1, 30 }, { 7, 21 }, { 5, 23 }, { 3, 26 },
+  { 6, 22 }, { 1, 31 }, { 4, 25 }, { 7, 22 },
+  { 3, 27 }, { 2, 29 }, { 2, 30 }, { 5, 24 },
+  { 2, 31 }, { 6, 23 }, { 4, 26 }, { 3, 28 },
+  { 5, 25 }, { 3, 29 }, { 6, 24 }, { 7, 23 },
+  { 3, 30 }, { 4, 27 }, { 3, 31 }, { 5, 26 },
+  { 6, 25 }, { 4, 28 }, { 7, 24 }, { 4, 29 },
+  { 5, 27 }, { 4, 30 }, { 4, 31 }, { 6, 26 },
+  { 5, 28 }, { 7, 25 }, { 6, 27 }, { 5, 29 },
+  { 7, 26 }, { 5, 30 }, { 5, 31 }, { 6, 28 },
+  { 7, 27 }, { 6, 29 }, { 6, 30 }, { 7, 28 },
+  { 6, 31 }, { 7, 29 }, { 7, 30 }, { 7, 31 },
+  { 8, 16 }, { 9, 16 }, { 8, 17 }, { 10, 16 },
+  { 9, 17 }, { 16, 8 }, { 8, 18 }, { 16, 9 },
+  { 10, 17 }, { 11, 16 }, { 17, 8 }, { 9, 18 },
+  { 8, 19 }, { 16, 10 }, { 11, 17 }, { 12, 16 },
+  { 10, 18 }, { 17, 9 }, { 9, 19 }, { 16, 11 },
+  { 8, 20 }, { 18, 8 }, { 17, 10 }, { 10, 19 },
+  { 12, 17 }, { 11, 18 }, { 9, 20 }, { 16, 12 },
+  { 18, 9 }, { 8, 21 }, { 13, 16 }, { 17, 11 },
+  { 19, 8 }, { 18, 10 }, { 13, 17 }, { 16, 13 },
+  { 11, 19 }, { 12, 18 }, { 10, 20 }, { 17, 12 },
+  { 9, 21 }, { 19, 9 }, { 8, 22 }, { 14, 16 },
+  { 18, 11 }, { 11, 20 }, { 10, 21 }, { 20, 8 },
+  { 13, 18 }, { 16, 14 }, { 12, 19 }, { 17, 13 },
+  { 19, 10 }, { 14, 17 }, { 9, 22 }, { 18, 12 },
+  { 8, 23 }, { 17, 14 }, { 20, 9 }, { 15, 16 },
+  { 16, 15 }, { 13, 19 }, { 10, 22 }, { 19, 11 },
+  { 11, 21 }, { 14, 18 }, { 12, 20 }, { 18, 13 },
+  { 20, 10 }, { 21, 8 }, { 15, 17 }, { 9, 23 },
+  { 19, 12 }, { 11, 22 }, { 8, 24 }, { 21, 9 },
+  { 17, 15 }, { 16, 16 }, { 14, 19 }, { 18, 14 },
+  { 12, 21 }, { 13, 20 }, { 20, 11 }, { 10, 23 },
+  { 19, 13 }, { 15, 18 }, { 16, 17 }, { 21, 10 },
+  { 22, 8 }, { 9, 24 }, { 8, 25 }, { 20, 12 },
+  { 15, 19 }, { 11, 23 }, { 17, 16 }, { 18, 15 },
+  { 14, 20 }, { 12, 22 }, { 10, 24 }, { 22, 9 },
+  { 21, 11 }, { 19, 14 }, { 13, 21 }, { 16, 18 },
+  { 9, 25 }, { 17, 17 }, { 8, 26 }, { 20, 13 },
+  { 23, 8 }, { 12, 23 }, { 13, 22 }, { 22, 10 },
+  { 19, 15 }, { 15, 20 }, { 16, 19 }, { 21, 12 },
+  { 11, 24 }, { 14, 21 }, { 8, 27 }, { 18, 16 },
+  { 10, 25 }, { 9, 26 }, { 22, 11 }, { 20, 14 },
+  { 23, 9 }, { 18, 17 }, { 17, 18 }, { 17, 19 },
+  { 19, 16 }, { 21, 13 }, { 10, 26 }, { 12, 24 },
+  { 23, 10 }, { 24, 8 }, { 8, 28 }, { 16, 20 },
+  { 9, 27 }, { 15, 21 }, { 22, 12 }, { 14, 22 },
+  { 13, 23 }, { 20, 15 }, { 11, 25 }, { 24, 9 },
+  { 18, 18 }, { 19, 17 }, { 23, 11 }, { 10, 27 },
+  { 8, 29 }, { 12, 25 }, { 9, 28 }, { 8, 30 },
+  { 21, 14 }, { 13, 24 }, { 11, 26 }, { 25, 8 },
+  { 24, 10 }, { 20, 16 }, { 19, 18 }, { 14, 23 },
+  { 22, 13 }, { 8, 31 }, { 17, 20 }, { 9, 29 },
+  { 23, 12 }, { 15, 22 }, { 25, 9 }, { 11, 27 },
+  { 10, 28 }, { 20, 17 }, { 21, 15 }, { 18, 19 },
+  { 16, 21 }, { 24, 11 }, { 9, 30 }, { 12, 26 },
+  { 10, 29 }, { 22, 14 }, { 14, 24 }, { 9, 31 },
+  { 26, 8 }, { 13, 25 }, { 25, 10 }, { 18, 20 },
+  { 19, 19 }, { 11, 28 }, { 15, 23 }, { 20, 18 },
+  { 10, 30 }, { 12, 27 }, { 17, 21 }, { 23, 13 },
+  { 24, 12 }, { 21, 16 }, { 16, 22 }, { 26, 9 },
+  { 27, 8 }, { 13, 26 }, { 22, 15 }, { 10, 31 },
+  { 14, 25 }, { 12, 28 }, { 25, 11 }, { 21, 17 },
+  { 26, 10 }, { 20, 19 }, { 11, 29 }, { 15, 24 },
+  { 23, 14 }, { 27, 9 }, { 11, 30 }, { 13, 27 },
+  { 19, 20 }, { 24, 13 }, { 28, 8 }, { 11, 31 },
+  { 22, 16 }, { 17, 22 }, { 16, 23 }, { 25, 12 },
+  { 18, 21 }, { 12, 29 }, { 21, 18 }, { 28, 9 },
+  { 27, 10 }, { 26, 11 }, { 29, 8 }, { 14, 26 },
+  { 15, 25 }, { 13, 28 }, { 12, 30 }, { 23, 15 },
+  { 30, 8 }, { 16, 24 }, { 13, 29 }, { 25, 13 },
+  { 24, 14 }, { 20, 20 }, { 31, 8 }, { 12, 31 },
+  { 14, 27 }, { 28, 10 }, { 26, 12 }, { 22, 17 },
+  { 21, 19 }, { 17, 23 }, { 18, 22 }, { 29, 9 },
+  { 27, 11 }, { 19, 21 }, { 27, 12 }, { 30, 9 },
+  { 31, 9 }, { 13, 30 }, { 24, 15 }, { 23, 16 },
+  { 15, 26 }, { 14, 28 }, { 29, 10 }, { 28, 11 },
+  { 26, 13 }, { 17, 24 }, { 13, 31 }, { 25, 14 },
+  { 22, 18 }, { 16, 25 }, { 30, 10 }, { 14, 29 },
+  { 15, 27 }, { 19, 22 }, { 21, 20 }, { 20, 21 },
+  { 27, 13 }, { 29, 11 }, { 18, 23 }, { 23, 17 },
+  { 16, 26 }, { 31, 10 }, { 24, 16 }, { 14, 30 },
+  { 22, 19 }, { 14, 31 }, { 28, 12 }, { 26, 14 },
+  { 30, 11 }, { 15, 28 }, { 25, 15 }, { 17, 25 },
+  { 23, 18 }, { 18, 24 }, { 15, 30 }, { 29, 12 },
+  { 31, 11 }, { 16, 27 }, { 24, 17 }, { 28, 13 },
+  { 19, 23 }, { 15, 29 }, { 25, 16 }, { 17, 26 },
+  { 27, 14 }, { 22, 20 }, { 15, 31 }, { 20, 22 },
+  { 21, 21 }, { 16, 28 }, { 17, 27 }, { 30, 12 },
+  { 26, 15 }, { 19, 24 }, { 18, 25 }, { 23, 19 },
+  { 29, 13 }, { 31, 12 }, { 24, 18 }, { 26, 16 },
+  { 25, 17 }, { 16, 29 }, { 28, 14 }, { 20, 23 },
+  { 18, 26 }, { 21, 22 }, { 19, 25 }, { 22, 21 },
+  { 27, 15 }, { 17, 28 }, { 16, 30 }, { 26, 17 },
+  { 23, 20 }, { 16, 31 }, { 25, 18 }, { 27, 16 },
+  { 20, 24 }, { 24, 19 }, { 31, 13 }, { 30, 13 },
+  { 29, 14 }, { 18, 27 }, { 28, 15 }, { 17, 29 },
+  { 19, 26 }, { 17, 30 }, { 21, 23 }, { 22, 22 },
+  { 30, 14 }, { 20, 25 }, { 23, 21 }, { 17, 31 },
+  { 18, 28 }, { 25, 19 }, { 24, 20 }, { 28, 16 },
+  { 31, 14 }, { 26, 18 }, { 19, 27 }, { 29, 15 },
+  { 27, 17 }, { 30, 15 }, { 21, 24 }, { 22, 23 },
+  { 26, 19 }, { 23, 22 }, { 28, 17 }, { 29, 16 },
+  { 18, 30 }, { 24, 21 }, { 25, 20 }, { 18, 31 },
+  { 18, 29 }, { 20, 26 }, { 19, 28 }, { 27, 18 },
+  { 31, 15 }, { 20, 27 }, { 30, 16 }, { 19, 29 },
+  { 29, 17 }, { 31, 16 }, { 27, 19 }, { 21, 25 },
+  { 28, 18 }, { 26, 20 }, { 22, 24 }, { 25, 21 },
+  { 19, 30 }, { 24, 22 }, { 30, 17 }, { 21, 26 },
+  { 23, 23 }, { 19, 31 }, { 20, 28 }, { 31, 17 },
+  { 28, 19 }, { 27, 20 }, { 21, 27 }, { 29, 18 },
+  { 30, 18 }, { 25, 22 }, { 26, 21 }, { 20, 29 },
+  { 22, 25 }, { 24, 23 }, { 29, 19 }, { 23, 24 },
+  { 20, 31 }, { 20, 30 }, { 28, 20 }, { 21, 28 },
+  { 22, 26 }, { 31, 18 }, { 27, 21 }, { 30, 19 },
+  { 22, 27 }, { 29, 20 }, { 23, 25 }, { 24, 24 },
+  { 26, 22 }, { 21, 29 }, { 25, 23 }, { 31, 19 },
+  { 21, 30 }, { 23, 26 }, { 28, 21 }, { 21, 31 },
+  { 22, 28 }, { 30, 20 }, { 25, 24 }, { 27, 22 },
+  { 29, 21 }, { 26, 23 }, { 24, 25 }, { 31, 20 },
+  { 23, 27 }, { 22, 29 }, { 30, 21 }, { 28, 22 },
+  { 24, 26 }, { 25, 25 }, { 27, 23 }, { 22, 30 },
+  { 23, 28 }, { 22, 31 }, { 26, 24 }, { 31, 21 },
+  { 24, 27 }, { 29, 22 }, { 27, 24 }, { 30, 22 },
+  { 25, 26 }, { 28, 23 }, { 23, 30 }, { 23, 29 },
+  { 24, 28 }, { 25, 27 }, { 31, 22 }, { 23, 31 },
+  { 26, 25 }, { 28, 24 }, { 29, 23 }, { 24, 29 },
+  { 24, 30 }, { 27, 25 }, { 25, 28 }, { 26, 26 },
+  { 30, 23 }, { 26, 27 }, { 31, 23 }, { 28, 25 },
+  { 27, 26 }, { 25, 29 }, { 24, 31 }, { 29, 24 },
+  { 30, 24 }, { 27, 27 }, { 29, 25 }, { 26, 28 },
+  { 31, 24 }, { 25, 30 }, { 25, 31 }, { 28, 26 },
+  { 27, 28 }, { 26, 29 }, { 30, 25 }, { 29, 26 },
+  { 28, 27 }, { 26, 30 }, { 31, 25 }, { 27, 29 },
+  { 26, 31 }, { 30, 26 }, { 28, 28 }, { 31, 26 },
+  { 29, 27 }, { 27, 30 }, { 28, 29 }, { 27, 31 },
+  { 30, 27 }, { 31, 27 }, { 28, 30 }, { 29, 28 },
+  { 30, 28 }, { 29, 29 }, { 30, 29 }, { 31, 28 },
+  { 28, 31 }, { 29, 30 }, { 29, 31 }, { 31, 29 },
+  { 30, 30 }, { 30, 31 }, { 31, 30 }, { 31, 31 }
+};
diff --git a/third_party/aom/av1/common/zigzag4.c b/third_party/aom/av1/common/zigzag4.c
new file mode 100644
index 0000000000..1fb5a320b7
--- /dev/null
+++ b/third_party/aom/av1/common/zigzag4.c
@@ -0,0 +1,22 @@
+/* This file is generated by gen_zigzag4.m */
+
+/* clang-format off */
+
+#include "odintrin.h"
+OD_EXTERN const unsigned char OD_ZIGZAG4_DCT_DCT[15][2] = {
+  {0, 1}, {1, 0}, {1, 1}, {0, 2},
+  {2, 0}, {0, 3}, {1, 2}, {3, 0},
+  {2, 1}, {1, 3}, {2, 2}, {3, 1},
+  {2, 3}, {3, 2}, {3, 3} };
+
+OD_EXTERN const unsigned char OD_ZIGZAG4_ADST_DCT[15][2] = {
+  {1, 0}, {0, 1}, {2, 0}, {1, 1},
+  {3, 0}, {2, 1}, {0, 2}, {1, 2},
+  {3, 1}, {0, 3}, {2, 2}, {1, 3},
+  {3, 2}, {2, 3}, {3, 3} };
+
+OD_EXTERN const unsigned char OD_ZIGZAG4_DCT_ADST[15][2] = {
+  {0, 1}, {0, 2}, {1, 0}, {0, 3},
+  {1, 1}, {1, 2}, {2, 0}, {1, 3},
+  {2, 1}, {2, 2}, {3, 0}, {3, 1},
+  {2, 3}, {3, 2}, {3, 3} };
diff --git a/third_party/aom/av1/common/zigzag8.c b/third_party/aom/av1/common/zigzag8.c
new file mode 100644
index 0000000000..3f11e0c035
--- /dev/null
+++ b/third_party/aom/av1/common/zigzag8.c
@@ -0,0 +1,50 @@
+/* This file is generated by gen_zigzag8.m */
+
+/* clang-format off */
+
+#include "odintrin.h"
+
+OD_EXTERN const unsigned char OD_ZIGZAG8_DCT_DCT[48][2] = {
+  {4, 0}, {4, 1}, {5, 0}, {5, 1},
+  {6, 0}, {7, 0}, {6, 1}, {7, 1},
+  {0, 4}, {1, 4}, {0, 5}, {1, 5},
+  {0, 6}, {1, 6}, {0, 7}, {1, 7},
+  {2, 4}, {4, 2}, {3, 4}, {2, 5},
+  {4, 3}, {5, 2}, {4, 4}, {3, 5},
+  {5, 3}, {2, 6}, {4, 5}, {6, 2},
+  {5, 4}, {3, 6}, {2, 7}, {6, 3},
+  {5, 5}, {7, 2}, {4, 6}, {3, 7},
+  {6, 4}, {7, 3}, {4, 7}, {5, 6},
+  {6, 5}, {7, 4}, {5, 7}, {6, 6},
+  {7, 5}, {6, 7}, {7, 6}, {7, 7}
+  };
+
+OD_EXTERN const unsigned char OD_ZIGZAG8_ADST_DCT[48][2] = {
+  {4, 0}, {5, 0}, {4, 1}, {6, 0},
+  {5, 1}, {7, 0}, {6, 1}, {7, 1},
+  {0, 4}, {1, 4}, {0, 5}, {1, 5},
+  {0, 6}, {1, 6}, {0, 7}, {1, 7},
+  {4, 2}, {2, 4}, {5, 2}, {4, 3},
+  {3, 4}, {2, 5}, {5, 3}, {4, 4},
+  {6, 2}, {3, 5}, {5, 4}, {2, 6},
+  {4, 5}, {6, 3}, {7, 2}, {3, 6},
+  {2, 7}, {5, 5}, {6, 4}, {4, 6},
+  {7, 3}, {3, 7}, {5, 6}, {6, 5},
+  {4, 7}, {7, 4}, {5, 7}, {7, 5},
+  {6, 6}, {7, 6}, {6, 7}, {7, 7}
+  };
+
+OD_EXTERN const unsigned char OD_ZIGZAG8_DCT_ADST[48][2] = {
+  {4, 0}, {4, 1}, {5, 0}, {5, 1},
+  {6, 0}, {6, 1}, {7, 0}, {7, 1},
+  {0, 4}, {0, 5}, {1, 4}, {0, 6},
+  {1, 5}, {0, 7}, {1, 6}, {1, 7},
+  {2, 4}, {2, 5}, {3, 4}, {4, 2},
+  {2, 6}, {4, 3}, {3, 5}, {4, 4},
+  {2, 7}, {3, 6}, {5, 2}, {4, 5},
+  {5, 3}, {3, 7}, {5, 4}, {4, 6},
+  {6, 2}, {5, 5}, {4, 7}, {6, 3},
+  {6, 4}, {5, 6}, {7, 2}, {6, 5},
+  {7, 3}, {5, 7}, {7, 4}, {6, 6},
+  {7, 5}, {6, 7}, {7, 6}, {7, 7}
+  };
diff --git a/third_party/aom/av1/decoder/accounting.c b/third_party/aom/av1/decoder/accounting.c
new file mode 100644
index 0000000000..ba243c9e1f
--- /dev/null
+++ b/third_party/aom/av1/decoder/accounting.c
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aom_integer.h"
+#include "./accounting.h"
+
+static int aom_accounting_hash(const char *str) {
+  uint32_t val;
+  const unsigned char *ustr;
+  val = 0;
+  ustr = (const unsigned char *)str;
+  /* This is about the worst hash one can design, but it should be good enough
+     here. */
+  while (*ustr) val += *ustr++;
+  return val % AOM_ACCOUNTING_HASH_SIZE;
+}
+
+/* Dictionary lookup based on an open-addressing hash table. */
+int aom_accounting_dictionary_lookup(Accounting *accounting, const char *str) {
+  int hash;
+  int len;
+  AccountingDictionary *dictionary;
+  dictionary = &accounting->syms.dictionary;
+  hash = aom_accounting_hash(str);
+  while (accounting->hash_dictionary[hash] != -1) {
+    if (strcmp(dictionary->strs[accounting->hash_dictionary[hash]], str) == 0) {
+      return accounting->hash_dictionary[hash];
+    }
+    hash++;
+    if (hash == AOM_ACCOUNTING_HASH_SIZE) hash = 0;
+  }
+  /* No match found. */
+  assert(dictionary->num_strs + 1 < MAX_SYMBOL_TYPES);
+  accounting->hash_dictionary[hash] = dictionary->num_strs;
+  len = strlen(str);
+  dictionary->strs[dictionary->num_strs] = malloc(len + 1);
+  snprintf(dictionary->strs[dictionary->num_strs], len + 1, "%s", str);
+  dictionary->num_strs++;
+  return dictionary->num_strs - 1;
+}
+
+void aom_accounting_init(Accounting *accounting) {
+  int i;
+  accounting->num_syms_allocated = 1000;
+  accounting->syms.syms =
+      malloc(sizeof(AccountingSymbol) * accounting->num_syms_allocated);
+  accounting->syms.dictionary.num_strs = 0;
+  assert(AOM_ACCOUNTING_HASH_SIZE > 2 * MAX_SYMBOL_TYPES);
+  for (i = 0; i < AOM_ACCOUNTING_HASH_SIZE; i++)
+    accounting->hash_dictionary[i] = -1;
+  aom_accounting_reset(accounting);
+}
+
+void aom_accounting_reset(Accounting *accounting) {
+  accounting->syms.num_syms = 0;
+  accounting->syms.num_binary_syms = 0;
+  accounting->syms.num_multi_syms = 0;
+  accounting->context.x = -1;
+  accounting->context.y = -1;
+  accounting->last_tell_frac = 0;
+}
+
+void aom_accounting_clear(Accounting *accounting) {
+  int i;
+  AccountingDictionary *dictionary;
+  free(accounting->syms.syms);
+  dictionary = &accounting->syms.dictionary;
+  for (i = 0; i < dictionary->num_strs; i++) {
+    free(dictionary->strs[i]);
+  }
+}
+
+void aom_accounting_set_context(Accounting *accounting, int16_t x, int16_t y) {
+  accounting->context.x = x;
+  accounting->context.y = y;
+}
+
+void aom_accounting_record(Accounting *accounting, const char *str,
+                           uint32_t bits) {
+  AccountingSymbol sym;
+  // Reuse previous symbol if it has the same context and symbol id.
+  if (accounting->syms.num_syms) {
+    AccountingSymbol *last_sym;
+    last_sym = &accounting->syms.syms[accounting->syms.num_syms - 1];
+    if (memcmp(&last_sym->context, &accounting->context,
+               sizeof(AccountingSymbolContext)) == 0) {
+      uint32_t id;
+      id = aom_accounting_dictionary_lookup(accounting, str);
+      if (id == last_sym->id) {
+        last_sym->bits += bits;
+        last_sym->samples++;
+        return;
+      }
+    }
+  }
+  sym.context = accounting->context;
+  sym.samples = 1;
+  sym.bits = bits;
+  sym.id = aom_accounting_dictionary_lookup(accounting, str);
+  assert(sym.id <= 255);
+  if (accounting->syms.num_syms == accounting->num_syms_allocated) {
+    accounting->num_syms_allocated *= 2;
+    accounting->syms.syms =
+        realloc(accounting->syms.syms,
+                sizeof(AccountingSymbol) * accounting->num_syms_allocated);
+    assert(accounting->syms.syms != NULL);
+  }
+  accounting->syms.syms[accounting->syms.num_syms++] = sym;
+}
+
+void aom_accounting_dump(Accounting *accounting) {
+  int i;
+  AccountingSymbol *sym;
+  printf("\n----- Number of recorded syntax elements = %d -----\n",
+         accounting->syms.num_syms);
+  printf("----- Total number of symbol calls = %d (%d binary) -----\n",
+         accounting->syms.num_multi_syms + accounting->syms.num_binary_syms,
+         accounting->syms.num_binary_syms);
+  for (i = 0; i < accounting->syms.num_syms; i++) {
+    sym = &accounting->syms.syms[i];
+    printf("%s x: %d, y: %d bits: %f samples: %d\n",
+           accounting->syms.dictionary.strs[sym->id], sym->context.x,
+           sym->context.y, (float)sym->bits / 8.0, sym->samples);
+  }
+}
diff --git a/third_party/aom/av1/decoder/accounting.h b/third_party/aom/av1/decoder/accounting.h
new file mode 100644
index 0000000000..889865b2e7
--- /dev/null
+++ b/third_party/aom/av1/decoder/accounting.h
@@ -0,0 +1,83 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_ACCOUNTING_H_
+#define AOM_ACCOUNTING_H_
+#include <stdlib.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+#define AOM_ACCOUNTING_HASH_SIZE (1021)
+
+/* Max number of entries for symbol types in the dictionary (increase as
+   necessary). */
+#define MAX_SYMBOL_TYPES (256)
+
+/*The resolution of fractional-precision bit usage measurements, i.e.,
+   3 => 1/8th bits.*/
+#define AOM_ACCT_BITRES (3)
+
+typedef struct {
+  int16_t x;
+  int16_t y;
+} AccountingSymbolContext;
+
+typedef struct {
+  AccountingSymbolContext context;
+  uint32_t id;
+  /** Number of bits in units of 1/8 bit. */
+  uint32_t bits;
+  uint32_t samples;
+} AccountingSymbol;
+
+/** Dictionary for translating strings into id. */
+typedef struct {
+  char *(strs[MAX_SYMBOL_TYPES]);
+  int num_strs;
+} AccountingDictionary;
+
+typedef struct {
+  /** All recorded symbols decoded. */
+  AccountingSymbol *syms;
+  /** Number of syntax actually recorded. */
+  int num_syms;
+  /** Raw symbol decoding calls for non-binary values. */
+  int num_multi_syms;
+  /** Raw binary symbol decoding calls. */
+  int num_binary_syms;
+  /** Dictionary for translating strings into id. */
+  AccountingDictionary dictionary;
+} AccountingSymbols;
+
+typedef struct Accounting Accounting;
+
+struct Accounting {
+  AccountingSymbols syms;
+  /** Size allocated for symbols (not all may be used). */
+  int num_syms_allocated;
+  int16_t hash_dictionary[AOM_ACCOUNTING_HASH_SIZE];
+  AccountingSymbolContext context;
+  uint32_t last_tell_frac;
+};
+
+void aom_accounting_init(Accounting *accounting);
+void aom_accounting_reset(Accounting *accounting);
+void aom_accounting_clear(Accounting *accounting);
+void aom_accounting_set_context(Accounting *accounting, int16_t x, int16_t y);
+int aom_accounting_dictionary_lookup(Accounting *accounting, const char *str);
+void aom_accounting_record(Accounting *accounting, const char *str,
+                           uint32_t bits);
+void aom_accounting_dump(Accounting *accounting);
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_ACCOUNTING_H_
diff --git a/third_party/aom/av1/decoder/decint.h b/third_party/aom/av1/decoder/decint.h
new file mode 100644
index 0000000000..e887ad5e0d
--- /dev/null
+++ b/third_party/aom/av1/decoder/decint.h
@@ -0,0 +1,35 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_decint_H)
+# define _decint_H (1)
+# include "av1/common/pvq_state.h"
+# include "aom_dsp/bitreader.h"
+# include "aom_dsp/entdec.h"
+
+typedef struct daala_dec_ctx daala_dec_ctx;
+
+typedef struct daala_dec_ctx od_dec_ctx;
+
+
+struct daala_dec_ctx {
+  /* Stores context-adaptive CDFs for PVQ. */
+  od_state state;
+  /* AOM entropy decoder. */
+  aom_reader *r;
+  int use_activity_masking;
+  /* Mode of quantization matrice : FLAT (0) or HVS (1) */
+  int qm;
+};
+
+#endif
diff --git a/third_party/aom/av1/decoder/decodeframe.c b/third_party/aom/av1/decoder/decodeframe.c
new file mode 100644
index 0000000000..289d386705
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodeframe.c
@@ -0,0 +1,5159 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdlib.h>  // qsort()
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "./aom_scale_rtcd.h"
+#include "./av1_rtcd.h"
+
+#include "aom/aom_codec.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/bitreader.h"
+#include "aom_dsp/bitreader_buffer.h"
+#include "aom_dsp/binary_codes_reader.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/mem_ops.h"
+#include "aom_scale/aom_scale.h"
+#include "aom_util/aom_thread.h"
+
+#if CONFIG_BITSTREAM_DEBUG
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#include "av1/common/alloccommon.h"
+#if CONFIG_CDEF
+#include "av1/common/cdef.h"
+#include "av1/common/clpf.h"
+#endif
+#if CONFIG_INSPECTION
+#include "av1/decoder/inspection.h"
+#endif
+#include "av1/common/common.h"
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/entropymv.h"
+#include "av1/common/idct.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+#include "av1/common/seg_common.h"
+#include "av1/common/thread_common.h"
+#include "av1/common/tile_common.h"
+
+#include "av1/decoder/decodeframe.h"
+#include "av1/decoder/decodemv.h"
+#include "av1/decoder/decoder.h"
+#if CONFIG_LV_MAP
+#include "av1/decoder/decodetxb.h"
+#endif
+#include "av1/decoder/detokenize.h"
+#include "av1/decoder/dsubexp.h"
+
+#if CONFIG_WARPED_MOTION || CONFIG_GLOBAL_MOTION
+#include "av1/common/warped_motion.h"
+#endif  // CONFIG_WARPED_MOTION || CONFIG_GLOBAL_MOTION
+
+#define MAX_AV1_HEADER_SIZE 80
+#define ACCT_STR __func__
+
+#if CONFIG_PVQ
+#include "av1/common/partition.h"
+#include "av1/common/pvq.h"
+#include "av1/common/scan.h"
+#include "av1/decoder/decint.h"
+#include "av1/decoder/pvq_decoder.h"
+#include "av1/encoder/encodemb.h"
+#include "av1/encoder/hybrid_fwd_txfm.h"
+#endif
+
+#if CONFIG_CFL
+#include "av1/common/cfl.h"
+#endif
+
+static struct aom_read_bit_buffer *init_read_bit_buffer(
+    AV1Decoder *pbi, struct aom_read_bit_buffer *rb, const uint8_t *data,
+    const uint8_t *data_end, uint8_t clear_data[MAX_AV1_HEADER_SIZE]);
+static int read_compressed_header(AV1Decoder *pbi, const uint8_t *data,
+                                  size_t partition_size);
+static size_t read_uncompressed_header(AV1Decoder *pbi,
+                                       struct aom_read_bit_buffer *rb);
+
+static int is_compound_reference_allowed(const AV1_COMMON *cm) {
+#if CONFIG_LOWDELAY_COMPOUND  // Normative in decoder
+  return !frame_is_intra_only(cm);
+#else
+  int i;
+  if (frame_is_intra_only(cm)) return 0;
+  for (i = 1; i < INTER_REFS_PER_FRAME; ++i)
+    if (cm->ref_frame_sign_bias[i + 1] != cm->ref_frame_sign_bias[1]) return 1;
+
+  return 0;
+#endif
+}
+
+static void setup_compound_reference_mode(AV1_COMMON *cm) {
+#if CONFIG_EXT_REFS
+  cm->comp_fwd_ref[0] = LAST_FRAME;
+  cm->comp_fwd_ref[1] = LAST2_FRAME;
+  cm->comp_fwd_ref[2] = LAST3_FRAME;
+  cm->comp_fwd_ref[3] = GOLDEN_FRAME;
+
+  cm->comp_bwd_ref[0] = BWDREF_FRAME;
+  cm->comp_bwd_ref[1] = ALTREF_FRAME;
+#else
+  if (cm->ref_frame_sign_bias[LAST_FRAME] ==
+      cm->ref_frame_sign_bias[GOLDEN_FRAME]) {
+    cm->comp_fixed_ref = ALTREF_FRAME;
+    cm->comp_var_ref[0] = LAST_FRAME;
+    cm->comp_var_ref[1] = GOLDEN_FRAME;
+  } else if (cm->ref_frame_sign_bias[LAST_FRAME] ==
+             cm->ref_frame_sign_bias[ALTREF_FRAME]) {
+    cm->comp_fixed_ref = GOLDEN_FRAME;
+    cm->comp_var_ref[0] = LAST_FRAME;
+    cm->comp_var_ref[1] = ALTREF_FRAME;
+  } else {
+    cm->comp_fixed_ref = LAST_FRAME;
+    cm->comp_var_ref[0] = GOLDEN_FRAME;
+    cm->comp_var_ref[1] = ALTREF_FRAME;
+  }
+#endif  // CONFIG_EXT_REFS
+}
+
+static int read_is_valid(const uint8_t *start, size_t len, const uint8_t *end) {
+  return len != 0 && len <= (size_t)(end - start);
+}
+
+static int decode_unsigned_max(struct aom_read_bit_buffer *rb, int max) {
+  const int data = aom_rb_read_literal(rb, get_unsigned_bits(max));
+  return data > max ? max : data;
+}
+
+static TX_MODE read_tx_mode(AV1_COMMON *cm, MACROBLOCKD *xd,
+                            struct aom_read_bit_buffer *rb) {
+  int i, all_lossless = 1;
+#if CONFIG_TX64X64
+  TX_MODE tx_mode;
+#endif
+
+  if (cm->seg.enabled) {
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      if (!xd->lossless[i]) {
+        all_lossless = 0;
+        break;
+      }
+    }
+  } else {
+    all_lossless = xd->lossless[0];
+  }
+
+  if (all_lossless) return ONLY_4X4;
+#if CONFIG_TX64X64
+  tx_mode = aom_rb_read_bit(rb) ? TX_MODE_SELECT : aom_rb_read_literal(rb, 2);
+  if (tx_mode == ALLOW_32X32) tx_mode += aom_rb_read_bit(rb);
+  return tx_mode;
+#else
+  return aom_rb_read_bit(rb) ? TX_MODE_SELECT : aom_rb_read_literal(rb, 2);
+#endif  // CONFIG_TX64X64
+}
+
+#if !CONFIG_EC_ADAPT
+static void read_tx_size_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j, k;
+  for (i = 0; i < MAX_TX_DEPTH; ++i)
+    for (j = 0; j < TX_SIZE_CONTEXTS; ++j)
+      for (k = 0; k < i + 1; ++k)
+        av1_diff_update_prob(r, &fc->tx_size_probs[i][j][k], ACCT_STR);
+}
+#endif
+
+#if !CONFIG_EC_ADAPT
+static void read_switchable_interp_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j;
+  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j) {
+    for (i = 0; i < SWITCHABLE_FILTERS - 1; ++i)
+      av1_diff_update_prob(r, &fc->switchable_interp_prob[j][i], ACCT_STR);
+  }
+}
+#endif
+
+static void read_inter_mode_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+#if CONFIG_REF_MV
+  int i;
+  for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i)
+    av1_diff_update_prob(r, &fc->newmv_prob[i], ACCT_STR);
+  for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i)
+    av1_diff_update_prob(r, &fc->zeromv_prob[i], ACCT_STR);
+  for (i = 0; i < REFMV_MODE_CONTEXTS; ++i)
+    av1_diff_update_prob(r, &fc->refmv_prob[i], ACCT_STR);
+  for (i = 0; i < DRL_MODE_CONTEXTS; ++i)
+    av1_diff_update_prob(r, &fc->drl_prob[i], ACCT_STR);
+#else
+#if !CONFIG_EC_ADAPT
+  int i, j;
+  for (i = 0; i < INTER_MODE_CONTEXTS; ++i) {
+    for (j = 0; j < INTER_MODES - 1; ++j)
+      av1_diff_update_prob(r, &fc->inter_mode_probs[i][j], ACCT_STR);
+  }
+#else
+  (void)fc;
+  (void)r;
+#endif
+#endif
+}
+
+#if CONFIG_EXT_INTER
+static void read_inter_compound_mode_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j;
+  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+    for (j = 0; j < INTER_MODE_CONTEXTS; ++j) {
+      for (i = 0; i < INTER_COMPOUND_MODES - 1; ++i) {
+        av1_diff_update_prob(r, &fc->inter_compound_mode_probs[j][i], ACCT_STR);
+      }
+    }
+  }
+}
+#endif  // CONFIG_EXT_INTER
+#if !CONFIG_EC_ADAPT
+#if !CONFIG_EXT_TX
+static void read_ext_tx_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j, k;
+  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+      for (j = 0; j < TX_TYPES; ++j) {
+        for (k = 0; k < TX_TYPES - 1; ++k)
+          av1_diff_update_prob(r, &fc->intra_ext_tx_prob[i][j][k], ACCT_STR);
+      }
+    }
+  }
+  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+      for (k = 0; k < TX_TYPES - 1; ++k)
+        av1_diff_update_prob(r, &fc->inter_ext_tx_prob[i][k], ACCT_STR);
+    }
+  }
+}
+#endif
+#endif
+
+static REFERENCE_MODE read_frame_reference_mode(
+    const AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
+  if (is_compound_reference_allowed(cm)) {
+#if CONFIG_REF_ADAPT
+    return aom_rb_read_bit(rb) ? REFERENCE_MODE_SELECT : SINGLE_REFERENCE;
+#else
+    return aom_rb_read_bit(rb)
+               ? REFERENCE_MODE_SELECT
+               : (aom_rb_read_bit(rb) ? COMPOUND_REFERENCE : SINGLE_REFERENCE);
+#endif  // CONFIG_REF_ADAPT
+  } else {
+    return SINGLE_REFERENCE;
+  }
+}
+
+static void read_frame_reference_mode_probs(AV1_COMMON *cm, aom_reader *r) {
+  FRAME_CONTEXT *const fc = cm->fc;
+  int i, j;
+
+  if (cm->reference_mode == REFERENCE_MODE_SELECT)
+    for (i = 0; i < COMP_INTER_CONTEXTS; ++i)
+      av1_diff_update_prob(r, &fc->comp_inter_prob[i], ACCT_STR);
+
+  if (cm->reference_mode != COMPOUND_REFERENCE) {
+    for (i = 0; i < REF_CONTEXTS; ++i) {
+      for (j = 0; j < (SINGLE_REFS - 1); ++j) {
+        av1_diff_update_prob(r, &fc->single_ref_prob[i][j], ACCT_STR);
+      }
+    }
+  }
+
+  if (cm->reference_mode != SINGLE_REFERENCE) {
+    for (i = 0; i < REF_CONTEXTS; ++i) {
+#if CONFIG_EXT_REFS
+      for (j = 0; j < (FWD_REFS - 1); ++j)
+        av1_diff_update_prob(r, &fc->comp_ref_prob[i][j], ACCT_STR);
+      for (j = 0; j < (BWD_REFS - 1); ++j)
+        av1_diff_update_prob(r, &fc->comp_bwdref_prob[i][j], ACCT_STR);
+#else
+      for (j = 0; j < (COMP_REFS - 1); ++j)
+        av1_diff_update_prob(r, &fc->comp_ref_prob[i][j], ACCT_STR);
+#endif  // CONFIG_EXT_REFS
+    }
+  }
+}
+
+static void update_mv_probs(aom_prob *p, int n, aom_reader *r) {
+  int i;
+  for (i = 0; i < n; ++i) av1_diff_update_prob(r, &p[i], ACCT_STR);
+}
+
+static void read_mv_probs(nmv_context *ctx, int allow_hp, aom_reader *r) {
+  int i;
+
+#if !CONFIG_EC_ADAPT
+  int j;
+  update_mv_probs(ctx->joints, MV_JOINTS - 1, r);
+
+  for (i = 0; i < 2; ++i) {
+    nmv_component *const comp_ctx = &ctx->comps[i];
+    update_mv_probs(&comp_ctx->sign, 1, r);
+    update_mv_probs(comp_ctx->classes, MV_CLASSES - 1, r);
+    update_mv_probs(comp_ctx->class0, CLASS0_SIZE - 1, r);
+    update_mv_probs(comp_ctx->bits, MV_OFFSET_BITS, r);
+  }
+  for (i = 0; i < 2; ++i) {
+    nmv_component *const comp_ctx = &ctx->comps[i];
+    for (j = 0; j < CLASS0_SIZE; ++j) {
+      update_mv_probs(comp_ctx->class0_fp[j], MV_FP_SIZE - 1, r);
+    }
+    update_mv_probs(comp_ctx->fp, MV_FP_SIZE - 1, r);
+  }
+#endif  // !CONFIG_EC_ADAPT
+
+  if (allow_hp) {
+    for (i = 0; i < 2; ++i) {
+      nmv_component *const comp_ctx = &ctx->comps[i];
+      update_mv_probs(&comp_ctx->class0_hp, 1, r);
+      update_mv_probs(&comp_ctx->hp, 1, r);
+    }
+  }
+}
+
+static void inverse_transform_block(MACROBLOCKD *xd, int plane,
+                                    const TX_TYPE tx_type,
+                                    const TX_SIZE tx_size, uint8_t *dst,
+                                    int stride, int16_t scan_line, int eob) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *const dqcoeff = pd->dqcoeff;
+  av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, dst, stride, eob);
+  memset(dqcoeff, 0, (scan_line + 1) * sizeof(dqcoeff[0]));
+}
+
+#if CONFIG_PVQ
+static int av1_pvq_decode_helper(MACROBLOCKD *xd, tran_low_t *ref_coeff,
+                                 tran_low_t *dqcoeff, int16_t *quant, int pli,
+                                 int bs, TX_TYPE tx_type, int xdec,
+                                 PVQ_SKIP_TYPE ac_dc_coded) {
+  unsigned int flags;  // used for daala's stream analyzer.
+  int off;
+  const int is_keyframe = 0;
+  const int has_dc_skip = 1;
+  int coeff_shift = 3 - av1_get_tx_scale(bs);
+  int hbd_downshift = 0;
+  int rounding_mask;
+  // DC quantizer for PVQ
+  int pvq_dc_quant;
+  int lossless = (quant[0] == 0);
+  const int blk_size = tx_size_wide[bs];
+  int eob = 0;
+  int i;
+  od_dec_ctx *dec = &xd->daala_dec;
+  int use_activity_masking = dec->use_activity_masking;
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_coeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+
+  od_coeff ref_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX];
+  od_coeff out_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX];
+
+#if CONFIG_HIGHBITDEPTH
+  hbd_downshift = xd->bd - 8;
+#endif  // CONFIG_HIGHBITDEPTH
+
+  od_raster_to_coding_order(ref_coeff_pvq, blk_size, tx_type, ref_coeff,
+                            blk_size);
+
+  assert(OD_COEFF_SHIFT >= 4);
+  if (lossless)
+    pvq_dc_quant = 1;
+  else {
+    if (use_activity_masking)
+      pvq_dc_quant = OD_MAXI(
+          1, (quant[0] << (OD_COEFF_SHIFT - 3) >> hbd_downshift) *
+                     dec->state.pvq_qm_q4[pli][od_qm_get_index(bs, 0)] >>
+                 4);
+    else
+      pvq_dc_quant =
+          OD_MAXI(1, quant[0] << (OD_COEFF_SHIFT - 3) >> hbd_downshift);
+  }
+
+  off = od_qm_offset(bs, xdec);
+
+  // copy int16 inputs to int32
+  for (i = 0; i < blk_size * blk_size; i++) {
+    ref_int32[i] =
+        AOM_SIGNED_SHL(ref_coeff_pvq[i], OD_COEFF_SHIFT - coeff_shift) >>
+        hbd_downshift;
+  }
+
+  od_pvq_decode(dec, ref_int32, out_int32,
+                OD_MAXI(1, quant[1] << (OD_COEFF_SHIFT - 3) >> hbd_downshift),
+                pli, bs, OD_PVQ_BETA[use_activity_masking][pli][bs],
+                is_keyframe, &flags, ac_dc_coded, dec->state.qm + off,
+                dec->state.qm_inv + off);
+
+  if (!has_dc_skip || out_int32[0]) {
+    out_int32[0] =
+        has_dc_skip + generic_decode(dec->r, &dec->state.adapt->model_dc[pli],
+                                     &dec->state.adapt->ex_dc[pli][bs][0], 2,
+                                     "dc:mag");
+    if (out_int32[0]) out_int32[0] *= aom_read_bit(dec->r, "dc:sign") ? -1 : 1;
+  }
+  out_int32[0] = out_int32[0] * pvq_dc_quant + ref_int32[0];
+
+  // copy int32 result back to int16
+  assert(OD_COEFF_SHIFT > coeff_shift);
+  rounding_mask = (1 << (OD_COEFF_SHIFT - coeff_shift - 1)) - 1;
+  for (i = 0; i < blk_size * blk_size; i++) {
+    out_int32[i] = AOM_SIGNED_SHL(out_int32[i], hbd_downshift);
+    dqcoeff_pvq[i] = (out_int32[i] + (out_int32[i] < 0) + rounding_mask) >>
+                     (OD_COEFF_SHIFT - coeff_shift);
+  }
+
+  od_coding_order_to_raster(dqcoeff, blk_size, tx_type, dqcoeff_pvq, blk_size);
+
+  eob = blk_size * blk_size;
+
+  return eob;
+}
+
+static PVQ_SKIP_TYPE read_pvq_skip(AV1_COMMON *cm, MACROBLOCKD *const xd,
+                                   int plane, TX_SIZE tx_size) {
+  // decode ac/dc coded flag. bit0: DC coded, bit1 : AC coded
+  // NOTE : we don't use 5 symbols for luma here in aom codebase,
+  // since block partition is taken care of by aom.
+  // So, only AC/DC skip info is coded
+  const int ac_dc_coded = aom_read_symbol(
+      xd->daala_dec.r,
+      xd->daala_dec.state.adapt->skip_cdf[2 * tx_size + (plane != 0)], 4,
+      "skip");
+  if (ac_dc_coded < 0 || ac_dc_coded > 3) {
+    aom_internal_error(&cm->error, AOM_CODEC_INVALID_PARAM,
+                       "Invalid PVQ Skip Type");
+  }
+  return ac_dc_coded;
+}
+
+static int av1_pvq_decode_helper2(AV1_COMMON *cm, MACROBLOCKD *const xd,
+                                  MB_MODE_INFO *const mbmi, int plane, int row,
+                                  int col, TX_SIZE tx_size, TX_TYPE tx_type) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  // transform block size in pixels
+  int tx_blk_size = tx_size_wide[tx_size];
+  int i, j;
+  tran_low_t *pvq_ref_coeff = pd->pvq_ref_coeff;
+  const int diff_stride = tx_blk_size;
+  int16_t *pred = pd->pred;
+  tran_low_t *const dqcoeff = pd->dqcoeff;
+  uint8_t *dst;
+  int eob;
+  const PVQ_SKIP_TYPE ac_dc_coded = read_pvq_skip(cm, xd, plane, tx_size);
+
+  eob = 0;
+  dst = &pd->dst.buf[4 * row * pd->dst.stride + 4 * col];
+
+  if (ac_dc_coded) {
+    int xdec = pd->subsampling_x;
+    int seg_id = mbmi->segment_id;
+    int16_t *quant;
+    FWD_TXFM_PARAM fwd_txfm_param;
+    // ToDo(yaowu): correct this with optimal number from decoding process.
+    const int max_scan_line = tx_size_2d[tx_size];
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      for (j = 0; j < tx_blk_size; j++)
+        for (i = 0; i < tx_blk_size; i++)
+          pred[diff_stride * j + i] =
+              CONVERT_TO_SHORTPTR(dst)[pd->dst.stride * j + i];
+    } else {
+#endif
+      for (j = 0; j < tx_blk_size; j++)
+        for (i = 0; i < tx_blk_size; i++)
+          pred[diff_stride * j + i] = dst[pd->dst.stride * j + i];
+#if CONFIG_HIGHBITDEPTH
+    }
+#endif
+
+    fwd_txfm_param.tx_type = tx_type;
+    fwd_txfm_param.tx_size = tx_size;
+    fwd_txfm_param.lossless = xd->lossless[seg_id];
+
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      fwd_txfm_param.bd = xd->bd;
+      av1_highbd_fwd_txfm(pred, pvq_ref_coeff, diff_stride, &fwd_txfm_param);
+    } else {
+#endif  // CONFIG_HIGHBITDEPTH
+      av1_fwd_txfm(pred, pvq_ref_coeff, diff_stride, &fwd_txfm_param);
+#if CONFIG_HIGHBITDEPTH
+    }
+#endif  // CONFIG_HIGHBITDEPTH
+
+    quant = &pd->seg_dequant[seg_id][0];  // aom's quantizer
+
+    eob = av1_pvq_decode_helper(xd, pvq_ref_coeff, dqcoeff, quant, plane,
+                                tx_size, tx_type, xdec, ac_dc_coded);
+
+    inverse_transform_block(xd, plane, tx_type, tx_size, dst, pd->dst.stride,
+                            max_scan_line, eob);
+  }
+
+  return eob;
+}
+#endif
+
+static int get_block_idx(const MACROBLOCKD *xd, int plane, int row, int col) {
+  const int bsize = xd->mi[0]->mbmi.sb_type;
+  const struct macroblockd_plane *pd = &xd->plane[plane];
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+  const BLOCK_SIZE plane_bsize =
+      AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#endif  // CONFIG_CHROMA_2X2
+#else
+  const BLOCK_SIZE plane_bsize =
+      get_plane_block_size(AOMMAX(BLOCK_8X8, bsize), pd);
+#endif
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+  const TX_SIZE tx_size = get_tx_size(plane, xd);
+  const uint8_t txh_unit = tx_size_high_unit[tx_size];
+  return row * max_blocks_wide + col * txh_unit;
+}
+
+static void predict_and_reconstruct_intra_block(
+    AV1_COMMON *cm, MACROBLOCKD *const xd, aom_reader *const r,
+    MB_MODE_INFO *const mbmi, int plane, int row, int col, TX_SIZE tx_size) {
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  const int block_idx = get_block_idx(xd, plane, row, col);
+#if CONFIG_PVQ
+  (void)r;
+#endif
+  av1_predict_intra_block_facade(xd, plane, block_idx, col, row, tx_size);
+
+  if (!mbmi->skip) {
+#if !CONFIG_PVQ
+    struct macroblockd_plane *const pd = &xd->plane[plane];
+#if CONFIG_LV_MAP
+    int16_t max_scan_line = 0;
+    int eob;
+    av1_read_coeffs_txb_facade(cm, xd, r, row, col, block_idx, plane,
+                               pd->dqcoeff, &max_scan_line, &eob);
+    // tx_type will be read out in av1_read_coeffs_txb_facade
+    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+#else   // CONFIG_LV_MAP
+    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+    const SCAN_ORDER *scan_order = get_scan(cm, tx_size, tx_type, 0);
+    int16_t max_scan_line = 0;
+    const int eob =
+        av1_decode_block_tokens(cm, xd, plane, scan_order, col, row, tx_size,
+                                tx_type, &max_scan_line, r, mbmi->segment_id);
+#endif  // CONFIG_LV_MAP
+    if (eob) {
+      uint8_t *dst =
+          &pd->dst.buf[(row * pd->dst.stride + col) << tx_size_wide_log2[0]];
+      inverse_transform_block(xd, plane, tx_type, tx_size, dst, pd->dst.stride,
+                              max_scan_line, eob);
+    }
+#else
+    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+    av1_pvq_decode_helper2(cm, xd, mbmi, plane, row, col, tx_size, tx_type);
+#endif
+  }
+#if CONFIG_CFL
+  if (plane == AOM_PLANE_Y) {
+    struct macroblockd_plane *const pd = &xd->plane[plane];
+    uint8_t *dst =
+        &pd->dst.buf[(row * pd->dst.stride + col) << tx_size_wide_log2[0]];
+    cfl_store(xd->cfl, dst, pd->dst.stride, row, col, tx_size);
+  }
+#endif
+}
+
+#if CONFIG_VAR_TX && !CONFIG_COEF_INTERLEAVE
+static void decode_reconstruct_tx(AV1_COMMON *cm, MACROBLOCKD *const xd,
+                                  aom_reader *r, MB_MODE_INFO *const mbmi,
+                                  int plane, BLOCK_SIZE plane_bsize,
+                                  int blk_row, int blk_col, TX_SIZE tx_size,
+                                  int *eob_total) {
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
+  const int tx_row = blk_row >> (1 - pd->subsampling_y);
+  const int tx_col = blk_col >> (1 - pd->subsampling_x);
+  const TX_SIZE plane_tx_size =
+      plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
+            : mbmi->inter_tx_size[tx_row][tx_col];
+  // Scale to match transform block unit.
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  if (tx_size == plane_tx_size) {
+    PLANE_TYPE plane_type = get_plane_type(plane);
+    int block_idx = get_block_idx(xd, plane, blk_row, blk_col);
+#if CONFIG_LV_MAP
+    (void)segment_id;
+    int16_t max_scan_line = 0;
+    int eob;
+    av1_read_coeffs_txb_facade(cm, xd, r, row, col, block_idx, plane,
+                               pd->dqcoeff, &max_scan_line, &eob);
+    // tx_type will be read out in av1_read_coeffs_txb_facade
+    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, plane_tx_size);
+#else   // CONFIG_LV_MAP
+    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, plane_tx_size);
+    const SCAN_ORDER *sc = get_scan(cm, plane_tx_size, tx_type, 1);
+    int16_t max_scan_line = 0;
+    const int eob = av1_decode_block_tokens(
+        cm, xd, plane, sc, blk_col, blk_row, plane_tx_size, tx_type,
+        &max_scan_line, r, mbmi->segment_id);
+#endif  // CONFIG_LV_MAP
+    inverse_transform_block(xd, plane, tx_type, plane_tx_size,
+                            &pd->dst.buf[(blk_row * pd->dst.stride + blk_col)
+                                         << tx_size_wide_log2[0]],
+                            pd->dst.stride, max_scan_line, eob);
+    *eob_total += eob;
+  } else {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int i;
+
+    assert(bsl > 0);
+
+    for (i = 0; i < 4; ++i) {
+      const int offsetr = blk_row + (i >> 1) * bsl;
+      const int offsetc = blk_col + (i & 0x01) * bsl;
+
+      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
+
+      decode_reconstruct_tx(cm, xd, r, mbmi, plane, plane_bsize, offsetr,
+                            offsetc, sub_txs, eob_total);
+    }
+  }
+}
+#endif  // CONFIG_VAR_TX
+
+#if !CONFIG_VAR_TX || CONFIG_SUPERTX || CONFIG_COEF_INTERLEAVE || \
+    (!CONFIG_VAR_TX && CONFIG_EXT_TX && CONFIG_RECT_TX)
+static int reconstruct_inter_block(AV1_COMMON *cm, MACROBLOCKD *const xd,
+                                   aom_reader *const r, int segment_id,
+                                   int plane, int row, int col,
+                                   TX_SIZE tx_size) {
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  int block_idx = get_block_idx(xd, plane, row, col);
+#if CONFIG_PVQ
+  int eob;
+  (void)r;
+  (void)segment_id;
+#else
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+#endif
+
+#if !CONFIG_PVQ
+#if CONFIG_LV_MAP
+  (void)segment_id;
+  int16_t max_scan_line = 0;
+  int eob;
+  av1_read_coeffs_txb_facade(cm, xd, r, row, col, block_idx, plane, pd->dqcoeff,
+                             &max_scan_line, &eob);
+  // tx_type will be read out in av1_read_coeffs_txb_facade
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+#else   // CONFIG_LV_MAP
+  int16_t max_scan_line = 0;
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+  const SCAN_ORDER *scan_order = get_scan(cm, tx_size, tx_type, 1);
+  const int eob =
+      av1_decode_block_tokens(cm, xd, plane, scan_order, col, row, tx_size,
+                              tx_type, &max_scan_line, r, segment_id);
+#endif  // CONFIG_LV_MAP
+  uint8_t *dst =
+      &pd->dst.buf[(row * pd->dst.stride + col) << tx_size_wide_log2[0]];
+  if (eob)
+    inverse_transform_block(xd, plane, tx_type, tx_size, dst, pd->dst.stride,
+                            max_scan_line, eob);
+#else
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
+  eob = av1_pvq_decode_helper2(cm, xd, &xd->mi[0]->mbmi, plane, row, col,
+                               tx_size, tx_type);
+#endif
+  return eob;
+}
+#endif  // !CONFIG_VAR_TX || CONFIG_SUPER_TX
+
+static void set_offsets(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                        BLOCK_SIZE bsize, int mi_row, int mi_col, int bw,
+                        int bh, int x_mis, int y_mis) {
+  const int offset = mi_row * cm->mi_stride + mi_col;
+  int x, y;
+  const TileInfo *const tile = &xd->tile;
+
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = &cm->mi[offset];
+  // TODO(slavarnway): Generate sb_type based on bwl and bhl, instead of
+  // passing bsize from decode_partition().
+  xd->mi[0]->mbmi.sb_type = bsize;
+#if CONFIG_RD_DEBUG
+  xd->mi[0]->mbmi.mi_row = mi_row;
+  xd->mi[0]->mbmi.mi_col = mi_col;
+#endif
+  for (y = 0; y < y_mis; ++y)
+    for (x = !y; x < x_mis; ++x) xd->mi[y * cm->mi_stride + x] = xd->mi[0];
+
+  set_plane_n4(xd, bw, bh);
+  set_skip_context(xd, mi_row, mi_col);
+
+#if CONFIG_VAR_TX
+  xd->max_tx_size = max_txsize_lookup[bsize];
+#endif
+
+  // Distance of Mb to the various image edges. These are specified to 8th pel
+  // as they are always compared to values that are in 1/8th pel units
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+  av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                       mi_col);
+}
+
+#if CONFIG_SUPERTX
+static MB_MODE_INFO *set_offsets_extend(AV1_COMMON *const cm,
+                                        MACROBLOCKD *const xd,
+                                        const TileInfo *const tile,
+                                        BLOCK_SIZE bsize_pred, int mi_row_pred,
+                                        int mi_col_pred, int mi_row_ori,
+                                        int mi_col_ori) {
+  // Used in supertx
+  // (mi_row_ori, mi_col_ori): location for mv
+  // (mi_row_pred, mi_col_pred, bsize_pred): region to predict
+  const int bw = mi_size_wide[bsize_pred];
+  const int bh = mi_size_high[bsize_pred];
+  const int offset = mi_row_ori * cm->mi_stride + mi_col_ori;
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = cm->mi + offset;
+  set_mi_row_col(xd, tile, mi_row_pred, bh, mi_col_pred, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+  xd->up_available = (mi_row_ori > tile->mi_row_start);
+  xd->left_available = (mi_col_ori > tile->mi_col_start);
+
+  set_plane_n4(xd, bw, bh);
+
+  return &xd->mi[0]->mbmi;
+}
+
+#if CONFIG_SUPERTX
+static MB_MODE_INFO *set_mb_offsets(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                    BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                    int bw, int bh, int x_mis, int y_mis) {
+  const int offset = mi_row * cm->mi_stride + mi_col;
+  const TileInfo *const tile = &xd->tile;
+  int x, y;
+
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = cm->mi + offset;
+  xd->mi[0]->mbmi.sb_type = bsize;
+  for (y = 0; y < y_mis; ++y)
+    for (x = !y; x < x_mis; ++x) xd->mi[y * cm->mi_stride + x] = xd->mi[0];
+
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+  return &xd->mi[0]->mbmi;
+}
+#endif
+
+static void set_offsets_topblock(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                 const TileInfo *const tile, BLOCK_SIZE bsize,
+                                 int mi_row, int mi_col) {
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int offset = mi_row * cm->mi_stride + mi_col;
+
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = cm->mi + offset;
+
+  set_plane_n4(xd, bw, bh);
+
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+  av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                       mi_col);
+}
+
+static void set_param_topblock(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                               BLOCK_SIZE bsize, int mi_row, int mi_col,
+                               int txfm, int skip) {
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int x_mis = AOMMIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = AOMMIN(bh, cm->mi_rows - mi_row);
+  const int offset = mi_row * cm->mi_stride + mi_col;
+  int x, y;
+
+  xd->mi = cm->mi_grid_visible + offset;
+  xd->mi[0] = cm->mi + offset;
+
+  for (y = 0; y < y_mis; ++y)
+    for (x = 0; x < x_mis; ++x) {
+      xd->mi[y * cm->mi_stride + x]->mbmi.skip = skip;
+      xd->mi[y * cm->mi_stride + x]->mbmi.tx_type = txfm;
+    }
+#if CONFIG_VAR_TX
+  xd->above_txfm_context = cm->above_txfm_context + mi_col;
+  xd->left_txfm_context =
+      xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+  set_txfm_ctxs(xd->mi[0]->mbmi.tx_size, bw, bh, skip, xd);
+#endif
+}
+
+static void set_ref(AV1_COMMON *const cm, MACROBLOCKD *const xd, int idx,
+                    int mi_row, int mi_col) {
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  RefBuffer *ref_buffer = &cm->frame_refs[mbmi->ref_frame[idx] - LAST_FRAME];
+  xd->block_refs[idx] = ref_buffer;
+  if (!av1_is_valid_scale(&ref_buffer->sf))
+    aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                       "Invalid scale factors");
+  av1_setup_pre_planes(xd, idx, ref_buffer->buf, mi_row, mi_col,
+                       &ref_buffer->sf);
+  aom_merge_corrupted_flag(&xd->corrupted, ref_buffer->buf->corrupted);
+}
+
+static void dec_predict_b_extend(
+    AV1Decoder *const pbi, MACROBLOCKD *const xd, const TileInfo *const tile,
+    int block, int mi_row_ori, int mi_col_ori, int mi_row_pred, int mi_col_pred,
+    int mi_row_top, int mi_col_top, uint8_t *dst_buf[3], int dst_stride[3],
+    BLOCK_SIZE bsize_top, BLOCK_SIZE bsize_pred, int b_sub8x8, int bextend) {
+  // Used in supertx
+  // (mi_row_ori, mi_col_ori): location for mv
+  // (mi_row_pred, mi_col_pred, bsize_pred): region to predict
+  // (mi_row_top, mi_col_top, bsize_top): region of the top partition size
+  // block: sub location of sub8x8 blocks
+  // b_sub8x8: 1: ori is sub8x8; 0: ori is not sub8x8
+  // bextend: 1: region to predict is an extension of ori; 0: not
+  int r = (mi_row_pred - mi_row_top) * MI_SIZE;
+  int c = (mi_col_pred - mi_col_top) * MI_SIZE;
+  const int mi_width_top = mi_size_wide[bsize_top];
+  const int mi_height_top = mi_size_high[bsize_top];
+  MB_MODE_INFO *mbmi;
+  AV1_COMMON *const cm = &pbi->common;
+
+  if (mi_row_pred < mi_row_top || mi_col_pred < mi_col_top ||
+      mi_row_pred >= mi_row_top + mi_height_top ||
+      mi_col_pred >= mi_col_top + mi_width_top || mi_row_pred >= cm->mi_rows ||
+      mi_col_pred >= cm->mi_cols)
+    return;
+
+  mbmi = set_offsets_extend(cm, xd, tile, bsize_pred, mi_row_pred, mi_col_pred,
+                            mi_row_ori, mi_col_ori);
+  set_ref(cm, xd, 0, mi_row_pred, mi_col_pred);
+  if (has_second_ref(&xd->mi[0]->mbmi))
+    set_ref(cm, xd, 1, mi_row_pred, mi_col_pred);
+
+  if (!bextend) mbmi->tx_size = max_txsize_lookup[bsize_top];
+
+  xd->plane[0].dst.stride = dst_stride[0];
+  xd->plane[1].dst.stride = dst_stride[1];
+  xd->plane[2].dst.stride = dst_stride[2];
+  xd->plane[0].dst.buf = dst_buf[0] +
+                         (r >> xd->plane[0].subsampling_y) * dst_stride[0] +
+                         (c >> xd->plane[0].subsampling_x);
+  xd->plane[1].dst.buf = dst_buf[1] +
+                         (r >> xd->plane[1].subsampling_y) * dst_stride[1] +
+                         (c >> xd->plane[1].subsampling_x);
+  xd->plane[2].dst.buf = dst_buf[2] +
+                         (r >> xd->plane[2].subsampling_y) * dst_stride[2] +
+                         (c >> xd->plane[2].subsampling_x);
+
+  if (!b_sub8x8)
+    av1_build_inter_predictors_sb_extend(xd,
+#if CONFIG_EXT_INTER
+                                         mi_row_ori, mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                                         mi_row_pred, mi_col_pred, bsize_pred);
+  else
+    av1_build_inter_predictors_sb_sub8x8_extend(xd,
+#if CONFIG_EXT_INTER
+                                                mi_row_ori, mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                                                mi_row_pred, mi_col_pred,
+                                                bsize_pred, block);
+}
+
+static void dec_extend_dir(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+                           const TileInfo *const tile, int block,
+                           BLOCK_SIZE bsize, BLOCK_SIZE top_bsize, int mi_row,
+                           int mi_col, int mi_row_top, int mi_col_top,
+                           uint8_t *dst_buf[3], int dst_stride[3], int dir) {
+  // dir: 0-lower, 1-upper, 2-left, 3-right
+  //      4-lowerleft, 5-upperleft, 6-lowerright, 7-upperright
+  const int mi_width = mi_size_wide[bsize];
+  const int mi_height = mi_size_high[bsize];
+  int xss = xd->plane[1].subsampling_x;
+  int yss = xd->plane[1].subsampling_y;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+  int b_sub8x8 = (bsize < BLOCK_8X8) && !unify_bsize ? 1 : 0;
+  BLOCK_SIZE extend_bsize;
+  int mi_row_pred, mi_col_pred;
+
+  int wide_unit, high_unit;
+  int i, j;
+  int ext_offset = 0;
+
+  if (dir == 0 || dir == 1) {
+    extend_bsize =
+        (mi_width == mi_size_wide[BLOCK_8X8] || bsize < BLOCK_8X8 || xss < yss)
+            ? BLOCK_8X8
+            : BLOCK_16X8;
+#if CONFIG_CB4X4
+    if (bsize < BLOCK_8X8) {
+      extend_bsize = BLOCK_4X4;
+      ext_offset = mi_size_wide[BLOCK_8X8];
+    }
+#endif
+
+    wide_unit = mi_size_wide[extend_bsize];
+    high_unit = mi_size_high[extend_bsize];
+
+    mi_row_pred = mi_row + ((dir == 0) ? mi_height : -(mi_height + ext_offset));
+    mi_col_pred = mi_col;
+
+    for (j = 0; j < mi_height + ext_offset; j += high_unit)
+      for (i = 0; i < mi_width + ext_offset; i += wide_unit)
+        dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col,
+                             mi_row_pred + j, mi_col_pred + i, mi_row_top,
+                             mi_col_top, dst_buf, dst_stride, top_bsize,
+                             extend_bsize, b_sub8x8, 1);
+  } else if (dir == 2 || dir == 3) {
+    extend_bsize =
+        (mi_height == mi_size_high[BLOCK_8X8] || bsize < BLOCK_8X8 || yss < xss)
+            ? BLOCK_8X8
+            : BLOCK_8X16;
+#if CONFIG_CB4X4
+    if (bsize < BLOCK_8X8) {
+      extend_bsize = BLOCK_4X4;
+      ext_offset = mi_size_wide[BLOCK_8X8];
+    }
+#endif
+
+    wide_unit = mi_size_wide[extend_bsize];
+    high_unit = mi_size_high[extend_bsize];
+
+    mi_row_pred = mi_row;
+    mi_col_pred = mi_col + ((dir == 3) ? mi_width : -(mi_width + ext_offset));
+
+    for (j = 0; j < mi_height + ext_offset; j += high_unit)
+      for (i = 0; i < mi_width + ext_offset; i += wide_unit)
+        dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col,
+                             mi_row_pred + j, mi_col_pred + i, mi_row_top,
+                             mi_col_top, dst_buf, dst_stride, top_bsize,
+                             extend_bsize, b_sub8x8, 1);
+  } else {
+    extend_bsize = BLOCK_8X8;
+#if CONFIG_CB4X4
+    if (bsize < BLOCK_8X8) {
+      extend_bsize = BLOCK_4X4;
+      ext_offset = mi_size_wide[BLOCK_8X8];
+    }
+#endif
+    wide_unit = mi_size_wide[extend_bsize];
+    high_unit = mi_size_high[extend_bsize];
+
+    mi_row_pred = mi_row + ((dir == 4 || dir == 6) ? mi_height
+                                                   : -(mi_height + ext_offset));
+    mi_col_pred =
+        mi_col + ((dir == 6 || dir == 7) ? mi_width : -(mi_width + ext_offset));
+
+    for (j = 0; j < mi_height + ext_offset; j += high_unit)
+      for (i = 0; i < mi_width + ext_offset; i += wide_unit)
+        dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col,
+                             mi_row_pred + j, mi_col_pred + i, mi_row_top,
+                             mi_col_top, dst_buf, dst_stride, top_bsize,
+                             extend_bsize, b_sub8x8, 1);
+  }
+}
+
+static void dec_extend_all(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+                           const TileInfo *const tile, int block,
+                           BLOCK_SIZE bsize, BLOCK_SIZE top_bsize, int mi_row,
+                           int mi_col, int mi_row_top, int mi_col_top,
+                           uint8_t *dst_buf[3], int dst_stride[3]) {
+  for (int i = 0; i < 8; ++i) {
+    dec_extend_dir(pbi, xd, tile, block, bsize, top_bsize, mi_row, mi_col,
+                   mi_row_top, mi_col_top, dst_buf, dst_stride, i);
+  }
+}
+
+static void dec_predict_sb_complex(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+                                   const TileInfo *const tile, int mi_row,
+                                   int mi_col, int mi_row_top, int mi_col_top,
+                                   BLOCK_SIZE bsize, BLOCK_SIZE top_bsize,
+                                   uint8_t *dst_buf[3], int dst_stride[3]) {
+  const AV1_COMMON *const cm = &pbi->common;
+  const int hbs = mi_size_wide[bsize] / 2;
+  const PARTITION_TYPE partition = get_partition(cm, mi_row, mi_col, bsize);
+  const BLOCK_SIZE subsize = get_subsize(bsize, partition);
+#if CONFIG_EXT_PARTITION_TYPES
+  const BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+  int i;
+  const int mi_offset = mi_row * cm->mi_stride + mi_col;
+  uint8_t *dst_buf1[3], *dst_buf2[3], *dst_buf3[3];
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf1[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf2[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf3[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
+  int dst_stride1[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
+  int dst_stride2[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
+  int dst_stride3[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    int len = sizeof(uint16_t);
+    dst_buf1[0] = CONVERT_TO_BYTEPTR(tmp_buf1);
+    dst_buf1[1] = CONVERT_TO_BYTEPTR(tmp_buf1 + MAX_TX_SQUARE * len);
+    dst_buf1[2] = CONVERT_TO_BYTEPTR(tmp_buf1 + 2 * MAX_TX_SQUARE * len);
+    dst_buf2[0] = CONVERT_TO_BYTEPTR(tmp_buf2);
+    dst_buf2[1] = CONVERT_TO_BYTEPTR(tmp_buf2 + MAX_TX_SQUARE * len);
+    dst_buf2[2] = CONVERT_TO_BYTEPTR(tmp_buf2 + 2 * MAX_TX_SQUARE * len);
+    dst_buf3[0] = CONVERT_TO_BYTEPTR(tmp_buf3);
+    dst_buf3[1] = CONVERT_TO_BYTEPTR(tmp_buf3 + MAX_TX_SQUARE * len);
+    dst_buf3[2] = CONVERT_TO_BYTEPTR(tmp_buf3 + 2 * MAX_TX_SQUARE * len);
+  } else {
+#endif
+    dst_buf1[0] = tmp_buf1;
+    dst_buf1[1] = tmp_buf1 + MAX_TX_SQUARE;
+    dst_buf1[2] = tmp_buf1 + 2 * MAX_TX_SQUARE;
+    dst_buf2[0] = tmp_buf2;
+    dst_buf2[1] = tmp_buf2 + MAX_TX_SQUARE;
+    dst_buf2[2] = tmp_buf2 + 2 * MAX_TX_SQUARE;
+    dst_buf3[0] = tmp_buf3;
+    dst_buf3[1] = tmp_buf3 + MAX_TX_SQUARE;
+    dst_buf3[2] = tmp_buf3 + 2 * MAX_TX_SQUARE;
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  xd->mi = cm->mi_grid_visible + mi_offset;
+  xd->mi[0] = cm->mi + mi_offset;
+
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    xd->plane[i].dst.buf = dst_buf[i];
+    xd->plane[i].dst.stride = dst_stride[i];
+  }
+
+  switch (partition) {
+    case PARTITION_NONE:
+      assert(bsize < top_bsize);
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                           mi_row_top, mi_col_top, dst_buf, dst_stride,
+                           top_bsize, bsize, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dst_buf, dst_stride);
+      break;
+    case PARTITION_HORZ:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        // For sub8x8, predict in 8x8 unit
+        // First half
+        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf, dst_stride,
+                             top_bsize, BLOCK_8X8, 1, 0);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride);
+
+        // Second half
+        dec_predict_b_extend(pbi, xd, tile, 2, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                             top_bsize, BLOCK_8X8, 1, 1);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 2, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+        // weighted average to smooth the boundary
+        xd->plane[0].dst.buf = dst_buf[0];
+        xd->plane[0].dst.stride = dst_stride[0];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[0], dst_stride[0], dst_buf1[0], dst_stride1[0], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            0);
+      } else {
+        // First half
+        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf, dst_stride,
+                             top_bsize, subsize, 0, 0);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride);
+        else
+          dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride, 0);
+
+        if (mi_row + hbs < cm->mi_rows) {
+          // Second half
+          dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col,
+                               mi_row + hbs, mi_col, mi_row_top, mi_col_top,
+                               dst_buf1, dst_stride1, top_bsize, subsize, 0, 0);
+          if (bsize < top_bsize)
+            dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1);
+          else
+            dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, 1);
+
+          // weighted average to smooth the boundary
+          for (i = 0; i < MAX_MB_PLANE; i++) {
+            xd->plane[i].dst.buf = dst_buf[i];
+            xd->plane[i].dst.stride = dst_stride[i];
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_HORZ, i);
+          }
+        }
+      }
+      break;
+    case PARTITION_VERT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        // First half
+        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf, dst_stride,
+                             top_bsize, BLOCK_8X8, 1, 0);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride);
+
+        // Second half
+        dec_predict_b_extend(pbi, xd, tile, 1, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                             top_bsize, BLOCK_8X8, 1, 1);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 1, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+        // Smooth
+        xd->plane[0].dst.buf = dst_buf[0];
+        xd->plane[0].dst.stride = dst_stride[0];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[0], dst_stride[0], dst_buf1[0], dst_stride1[0], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            0);
+      } else {
+        // First half
+        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf, dst_stride,
+                             top_bsize, subsize, 0, 0);
+        if (bsize < top_bsize)
+          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride);
+        else
+          dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride, 3);
+
+        // Second half
+        if (mi_col + hbs < cm->mi_cols) {
+          dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col + hbs, mi_row,
+                               mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                               dst_stride1, top_bsize, subsize, 0, 0);
+          if (bsize < top_bsize)
+            dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1);
+          else
+            dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, 2);
+
+          // Smooth
+          for (i = 0; i < MAX_MB_PLANE; i++) {
+            xd->plane[i].dst.buf = dst_buf[i];
+            xd->plane[i].dst.stride = dst_stride[i];
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_VERT, i);
+          }
+        }
+      }
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf, dst_stride,
+                             top_bsize, BLOCK_8X8, 1, 0);
+        dec_predict_b_extend(pbi, xd, tile, 1, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                             top_bsize, BLOCK_8X8, 1, 1);
+        dec_predict_b_extend(pbi, xd, tile, 2, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf2, dst_stride2,
+                             top_bsize, BLOCK_8X8, 1, 1);
+        dec_predict_b_extend(pbi, xd, tile, 3, mi_row, mi_col, mi_row, mi_col,
+                             mi_row_top, mi_col_top, dst_buf3, dst_stride3,
+                             top_bsize, BLOCK_8X8, 1, 1);
+        if (bsize < top_bsize) {
+          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride);
+          dec_extend_all(pbi, xd, tile, 1, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+          dec_extend_all(pbi, xd, tile, 2, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf2, dst_stride2);
+          dec_extend_all(pbi, xd, tile, 3, subsize, top_bsize, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf3, dst_stride3);
+        }
+      } else {
+        dec_predict_sb_complex(pbi, xd, tile, mi_row, mi_col, mi_row_top,
+                               mi_col_top, subsize, top_bsize, dst_buf,
+                               dst_stride);
+        if (mi_row < cm->mi_rows && mi_col + hbs < cm->mi_cols)
+          dec_predict_sb_complex(pbi, xd, tile, mi_row, mi_col + hbs,
+                                 mi_row_top, mi_col_top, subsize, top_bsize,
+                                 dst_buf1, dst_stride1);
+        if (mi_row + hbs < cm->mi_rows && mi_col < cm->mi_cols)
+          dec_predict_sb_complex(pbi, xd, tile, mi_row + hbs, mi_col,
+                                 mi_row_top, mi_col_top, subsize, top_bsize,
+                                 dst_buf2, dst_stride2);
+        if (mi_row + hbs < cm->mi_rows && mi_col + hbs < cm->mi_cols)
+          dec_predict_sb_complex(pbi, xd, tile, mi_row + hbs, mi_col + hbs,
+                                 mi_row_top, mi_col_top, subsize, top_bsize,
+                                 dst_buf3, dst_stride3);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+#if !CONFIG_CB4X4
+        if (bsize == BLOCK_8X8 && i != 0)
+          continue;  // Skip <4x4 chroma smoothing
+#endif
+        if (mi_row < cm->mi_rows && mi_col + hbs < cm->mi_cols) {
+          av1_build_masked_inter_predictor_complex(
+              xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i],
+              mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+              PARTITION_VERT, i);
+          if (mi_row + hbs < cm->mi_rows) {
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf2[i], dst_stride2[i], dst_buf3[i], dst_stride3[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_VERT, i);
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_HORZ, i);
+          }
+        } else if (mi_row + hbs < cm->mi_rows && mi_col < cm->mi_cols) {
+          av1_build_masked_inter_predictor_complex(
+              xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i],
+              mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+              PARTITION_HORZ, i);
+        }
+      }
+      break;
+#if CONFIG_EXT_PARTITION_TYPES
+    case PARTITION_HORZ_A:
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                           mi_row_top, mi_col_top, dst_buf, dst_stride,
+                           top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dst_buf, dst_stride);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col + hbs, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row, mi_col + hbs,
+                     mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf2,
+                           dst_stride2, top_bsize, subsize, 0, 0);
+      if (bsize < top_bsize)
+        dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row + hbs,
+                       mi_col, mi_row_top, mi_col_top, dst_buf2, dst_stride2);
+      else
+        dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row + hbs,
+                       mi_col, mi_row_top, mi_col_top, dst_buf2, dst_stride2,
+                       1);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            i);
+      }
+      break;
+    case PARTITION_VERT_A:
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                           mi_row_top, mi_col_top, dst_buf, dst_stride,
+                           top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dst_buf, dst_stride);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row + hbs, mi_col,
+                     mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col + hbs, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                           dst_stride2, top_bsize, subsize, 0, 0);
+      if (bsize < top_bsize)
+        dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row,
+                       mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                       dst_stride2);
+      else
+        dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row,
+                       mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                       dst_stride2, 2);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            i);
+      }
+      break;
+    case PARTITION_HORZ_B:
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                           mi_row_top, mi_col_top, dst_buf, dst_stride,
+                           top_bsize, subsize, 0, 0);
+      if (bsize < top_bsize)
+        dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride);
+      else
+        dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride, 0);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row + hbs, mi_col,
+                     mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col + hbs,
+                           mi_row + hbs, mi_col + hbs, mi_row_top, mi_col_top,
+                           dst_buf2, dst_stride2, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row + hbs,
+                     mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                     dst_stride2);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf1[i];
+        xd->plane[i].dst.stride = dst_stride1[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf1[i], dst_stride1[i], dst_buf2[i], dst_stride2[i],
+            mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+            PARTITION_VERT, i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            i);
+      }
+      break;
+    case PARTITION_VERT_B:
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                           mi_row_top, mi_col_top, dst_buf, dst_stride,
+                           top_bsize, subsize, 0, 0);
+      if (bsize < top_bsize)
+        dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride);
+      else
+        dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride, 3);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col + hbs, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row, mi_col + hbs,
+                     mi_row_top, mi_col_top, dst_buf1, dst_stride1);
+
+      dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col + hbs,
+                           mi_row + hbs, mi_col + hbs, mi_row_top, mi_col_top,
+                           dst_buf2, dst_stride2, top_bsize, bsize2, 0, 0);
+      dec_extend_all(pbi, xd, tile, 0, bsize2, top_bsize, mi_row + hbs,
+                     mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                     dst_stride2);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf1[i];
+        xd->plane[i].dst.stride = dst_stride1[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf1[i], dst_stride1[i], dst_buf2[i], dst_stride2[i],
+            mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+            PARTITION_HORZ, i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            i);
+      }
+      break;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+    default: assert(0);
+  }
+}
+
+static void set_segment_id_supertx(const AV1_COMMON *const cm, int mi_row,
+                                   int mi_col, BLOCK_SIZE bsize) {
+  const struct segmentation *seg = &cm->seg;
+  const int miw = AOMMIN(mi_size_wide[bsize], cm->mi_cols - mi_col);
+  const int mih = AOMMIN(mi_size_high[bsize], cm->mi_rows - mi_row);
+  const int mi_offset = mi_row * cm->mi_stride + mi_col;
+  MODE_INFO **const mip = cm->mi_grid_visible + mi_offset;
+  int r, c;
+  int seg_id_supertx = MAX_SEGMENTS;
+
+  if (!seg->enabled) {
+    seg_id_supertx = 0;
+  } else {
+    // Find the minimum segment_id
+    for (r = 0; r < mih; r++)
+      for (c = 0; c < miw; c++)
+        seg_id_supertx =
+            AOMMIN(mip[r * cm->mi_stride + c]->mbmi.segment_id, seg_id_supertx);
+    assert(0 <= seg_id_supertx && seg_id_supertx < MAX_SEGMENTS);
+  }
+
+  // Assign the the segment_id back to segment_id_supertx
+  for (r = 0; r < mih; r++)
+    for (c = 0; c < miw; c++)
+      mip[r * cm->mi_stride + c]->mbmi.segment_id_supertx = seg_id_supertx;
+}
+#endif  // CONFIG_SUPERTX
+
+static void decode_mbmi_block(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+#if CONFIG_SUPERTX
+                              int supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                              int mi_row, int mi_col, aom_reader *r,
+#if CONFIG_EXT_PARTITION_TYPES
+                              PARTITION_TYPE partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                              BLOCK_SIZE bsize) {
+  AV1_COMMON *const cm = &pbi->common;
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int x_mis = AOMMIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = AOMMIN(bh, cm->mi_rows - mi_row);
+
+#if CONFIG_ACCOUNTING
+  aom_accounting_set_context(&pbi->accounting, mi_col, mi_row);
+#endif
+#if CONFIG_SUPERTX
+  if (supertx_enabled) {
+    set_mb_offsets(cm, xd, bsize, mi_row, mi_col, bw, bh, x_mis, y_mis);
+  } else {
+    set_offsets(cm, xd, bsize, mi_row, mi_col, bw, bh, x_mis, y_mis);
+  }
+#if CONFIG_EXT_PARTITION_TYPES
+  xd->mi[0]->mbmi.partition = partition;
+#endif
+  av1_read_mode_info(pbi, xd, supertx_enabled, mi_row, mi_col, r, x_mis, y_mis);
+#else
+  set_offsets(cm, xd, bsize, mi_row, mi_col, bw, bh, x_mis, y_mis);
+#if CONFIG_EXT_PARTITION_TYPES
+  xd->mi[0]->mbmi.partition = partition;
+#endif
+  av1_read_mode_info(pbi, xd, mi_row, mi_col, r, x_mis, y_mis);
+#endif  // CONFIG_SUPERTX
+
+  if (bsize >= BLOCK_8X8 && (cm->subsampling_x || cm->subsampling_y)) {
+    const BLOCK_SIZE uv_subsize =
+        ss_size_lookup[bsize][cm->subsampling_x][cm->subsampling_y];
+    if (uv_subsize == BLOCK_INVALID)
+      aom_internal_error(xd->error_info, AOM_CODEC_CORRUPT_FRAME,
+                         "Invalid block size.");
+  }
+
+#if CONFIG_SUPERTX
+  xd->mi[0]->mbmi.segment_id_supertx = MAX_SEGMENTS;
+#endif  // CONFIG_SUPERTX
+
+  int reader_corrupted_flag = aom_reader_has_error(r);
+  aom_merge_corrupted_flag(&xd->corrupted, reader_corrupted_flag);
+}
+
+static void decode_token_and_recon_block(AV1Decoder *const pbi,
+                                         MACROBLOCKD *const xd, int mi_row,
+                                         int mi_col, aom_reader *r,
+                                         BLOCK_SIZE bsize) {
+  AV1_COMMON *const cm = &pbi->common;
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int x_mis = AOMMIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = AOMMIN(bh, cm->mi_rows - mi_row);
+
+  set_offsets(cm, xd, bsize, mi_row, mi_col, bw, bh, x_mis, y_mis);
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+
+#if CONFIG_DELTA_Q
+  if (cm->delta_q_present_flag) {
+    int i;
+    for (i = 0; i < MAX_SEGMENTS; i++) {
+#if CONFIG_EXT_DELTA_Q
+      xd->plane[0].seg_dequant[i][0] =
+          av1_dc_quant(av1_get_qindex(&cm->seg, i, xd->current_qindex),
+                       cm->y_dc_delta_q, cm->bit_depth);
+      xd->plane[0].seg_dequant[i][1] = av1_ac_quant(
+          av1_get_qindex(&cm->seg, i, xd->current_qindex), 0, cm->bit_depth);
+      xd->plane[1].seg_dequant[i][0] =
+          av1_dc_quant(av1_get_qindex(&cm->seg, i, xd->current_qindex),
+                       cm->uv_dc_delta_q, cm->bit_depth);
+      xd->plane[1].seg_dequant[i][1] =
+          av1_ac_quant(av1_get_qindex(&cm->seg, i, xd->current_qindex),
+                       cm->uv_ac_delta_q, cm->bit_depth);
+      xd->plane[2].seg_dequant[i][0] =
+          av1_dc_quant(av1_get_qindex(&cm->seg, i, xd->current_qindex),
+                       cm->uv_dc_delta_q, cm->bit_depth);
+      xd->plane[2].seg_dequant[i][1] =
+          av1_ac_quant(av1_get_qindex(&cm->seg, i, xd->current_qindex),
+                       cm->uv_ac_delta_q, cm->bit_depth);
+#else
+      xd->plane[0].seg_dequant[i][0] =
+          av1_dc_quant(xd->current_qindex, cm->y_dc_delta_q, cm->bit_depth);
+      xd->plane[0].seg_dequant[i][1] =
+          av1_ac_quant(xd->current_qindex, 0, cm->bit_depth);
+      xd->plane[1].seg_dequant[i][0] =
+          av1_dc_quant(xd->current_qindex, cm->uv_dc_delta_q, cm->bit_depth);
+      xd->plane[1].seg_dequant[i][1] =
+          av1_ac_quant(xd->current_qindex, cm->uv_ac_delta_q, cm->bit_depth);
+      xd->plane[2].seg_dequant[i][0] =
+          av1_dc_quant(xd->current_qindex, cm->uv_dc_delta_q, cm->bit_depth);
+      xd->plane[2].seg_dequant[i][1] =
+          av1_ac_quant(xd->current_qindex, cm->uv_ac_delta_q, cm->bit_depth);
+#endif
+    }
+  }
+#endif
+
+#if CONFIG_CB4X4
+  if (mbmi->skip) reset_skip_context(xd, bsize);
+#else
+  if (mbmi->skip) reset_skip_context(xd, AOMMAX(BLOCK_8X8, bsize));
+#endif
+
+#if CONFIG_COEF_INTERLEAVE
+  {
+    const struct macroblockd_plane *const pd_y = &xd->plane[0];
+    const struct macroblockd_plane *const pd_c = &xd->plane[1];
+    const TX_SIZE tx_log2_y = mbmi->tx_size;
+    const TX_SIZE tx_log2_c = get_uv_tx_size(mbmi, pd_c);
+    const int tx_sz_y = (1 << tx_log2_y);
+    const int tx_sz_c = (1 << tx_log2_c);
+    const int num_4x4_w_y = pd_y->n4_w;
+    const int num_4x4_h_y = pd_y->n4_h;
+    const int num_4x4_w_c = pd_c->n4_w;
+    const int num_4x4_h_c = pd_c->n4_h;
+    const int max_4x4_w_y = get_max_4x4_size(num_4x4_w_y, xd->mb_to_right_edge,
+                                             pd_y->subsampling_x);
+    const int max_4x4_h_y = get_max_4x4_size(num_4x4_h_y, xd->mb_to_bottom_edge,
+                                             pd_y->subsampling_y);
+    const int max_4x4_w_c = get_max_4x4_size(num_4x4_w_c, xd->mb_to_right_edge,
+                                             pd_c->subsampling_x);
+    const int max_4x4_h_c = get_max_4x4_size(num_4x4_h_c, xd->mb_to_bottom_edge,
+                                             pd_c->subsampling_y);
+
+    // The max_4x4_w/h may be smaller than tx_sz under some corner cases,
+    // i.e. when the SB is splitted by tile boundaries.
+    const int tu_num_w_y = (max_4x4_w_y + tx_sz_y - 1) / tx_sz_y;
+    const int tu_num_h_y = (max_4x4_h_y + tx_sz_y - 1) / tx_sz_y;
+    const int tu_num_w_c = (max_4x4_w_c + tx_sz_c - 1) / tx_sz_c;
+    const int tu_num_h_c = (max_4x4_h_c + tx_sz_c - 1) / tx_sz_c;
+    const int tu_num_c = tu_num_w_c * tu_num_h_c;
+
+    if (!is_inter_block(mbmi)) {
+      int tu_idx_c = 0;
+      int row_y, col_y, row_c, col_c;
+      int plane;
+
+#if CONFIG_PALETTE
+      for (plane = 0; plane <= 1; ++plane) {
+        if (mbmi->palette_mode_info.palette_size[plane])
+          av1_decode_palette_tokens(xd, plane, r);
+      }
+#endif
+
+      for (row_y = 0; row_y < tu_num_h_y; row_y++) {
+        for (col_y = 0; col_y < tu_num_w_y; col_y++) {
+          // luma
+          predict_and_reconstruct_intra_block(
+              cm, xd, r, mbmi, 0, row_y * tx_sz_y, col_y * tx_sz_y, tx_log2_y);
+          // chroma
+          if (tu_idx_c < tu_num_c) {
+            row_c = (tu_idx_c / tu_num_w_c) * tx_sz_c;
+            col_c = (tu_idx_c % tu_num_w_c) * tx_sz_c;
+            predict_and_reconstruct_intra_block(cm, xd, r, mbmi, 1, row_c,
+                                                col_c, tx_log2_c);
+            predict_and_reconstruct_intra_block(cm, xd, r, mbmi, 2, row_c,
+                                                col_c, tx_log2_c);
+            tu_idx_c++;
+          }
+        }
+      }
+
+      // In 422 case, it's possilbe that Chroma has more TUs than Luma
+      while (tu_idx_c < tu_num_c) {
+        row_c = (tu_idx_c / tu_num_w_c) * tx_sz_c;
+        col_c = (tu_idx_c % tu_num_w_c) * tx_sz_c;
+        predict_and_reconstruct_intra_block(cm, xd, r, mbmi, 1, row_c, col_c,
+                                            tx_log2_c);
+        predict_and_reconstruct_intra_block(cm, xd, r, mbmi, 2, row_c, col_c,
+                                            tx_log2_c);
+        tu_idx_c++;
+      }
+    } else {
+      // Prediction
+      av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL,
+                                    AOMMAX(bsize, BLOCK_8X8));
+
+      // Reconstruction
+      if (!mbmi->skip) {
+        int eobtotal = 0;
+        int tu_idx_c = 0;
+        int row_y, col_y, row_c, col_c;
+
+        for (row_y = 0; row_y < tu_num_h_y; row_y++) {
+          for (col_y = 0; col_y < tu_num_w_y; col_y++) {
+            // luma
+            eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id, 0,
+                                                row_y * tx_sz_y,
+                                                col_y * tx_sz_y, tx_log2_y);
+            // chroma
+            if (tu_idx_c < tu_num_c) {
+              row_c = (tu_idx_c / tu_num_w_c) * tx_sz_c;
+              col_c = (tu_idx_c % tu_num_w_c) * tx_sz_c;
+              eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id,
+                                                  1, row_c, col_c, tx_log2_c);
+              eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id,
+                                                  2, row_c, col_c, tx_log2_c);
+              tu_idx_c++;
+            }
+          }
+        }
+
+        // In 422 case, it's possilbe that Chroma has more TUs than Luma
+        while (tu_idx_c < tu_num_c) {
+          row_c = (tu_idx_c / tu_num_w_c) * tx_sz_c;
+          col_c = (tu_idx_c % tu_num_w_c) * tx_sz_c;
+          eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id, 1,
+                                              row_c, col_c, tx_log2_c);
+          eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id, 2,
+                                              row_c, col_c, tx_log2_c);
+          tu_idx_c++;
+        }
+
+        // TODO(CONFIG_COEF_INTERLEAVE owners): bring eob == 0 corner case
+        // into line with the defaut configuration
+        if (bsize >= BLOCK_8X8 && eobtotal == 0) mbmi->skip = 1;
+      }
+    }
+  }
+#else  // CONFIG_COEF_INTERLEAVE
+  if (!is_inter_block(mbmi)) {
+    int plane;
+#if CONFIG_PALETTE
+    for (plane = 0; plane <= 1; ++plane) {
+      if (mbmi->palette_mode_info.palette_size[plane])
+        av1_decode_palette_tokens(xd, plane, r);
+    }
+#endif  // CONFIG_PALETTE
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+      const struct macroblockd_plane *const pd = &xd->plane[plane];
+      const TX_SIZE tx_size = get_tx_size(plane, xd);
+      const int stepr = tx_size_high_unit[tx_size];
+      const int stepc = tx_size_wide_unit[tx_size];
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+      const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+      const BLOCK_SIZE plane_bsize =
+          AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#endif  // CONFIG_CHROMA_2X2
+#else
+      const BLOCK_SIZE plane_bsize =
+          get_plane_block_size(AOMMAX(BLOCK_8X8, bsize), pd);
+#endif
+      int row, col;
+      const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+      const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+#if CONFIG_CB4X4
+      if (!is_chroma_reference(mi_row, mi_col, bsize, pd->subsampling_x,
+                               pd->subsampling_y))
+        continue;
+#endif
+
+      for (row = 0; row < max_blocks_high; row += stepr)
+        for (col = 0; col < max_blocks_wide; col += stepc)
+          predict_and_reconstruct_intra_block(cm, xd, r, mbmi, plane, row, col,
+                                              tx_size);
+    }
+  } else {
+    int ref;
+
+    for (ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) {
+      const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
+      if (frame < LAST_FRAME) {
+#if CONFIG_INTRABC
+        assert(is_intrabc_block(mbmi));
+        assert(frame == INTRA_FRAME);
+        assert(ref == 0);
+#else
+        assert(0);
+#endif  // CONFIG_INTRABC
+      } else {
+        RefBuffer *ref_buf = &cm->frame_refs[frame - LAST_FRAME];
+
+        xd->block_refs[ref] = ref_buf;
+        if ((!av1_is_valid_scale(&ref_buf->sf)))
+          aom_internal_error(xd->error_info, AOM_CODEC_UNSUP_BITSTREAM,
+                             "Reference frame has invalid dimensions");
+        av1_setup_pre_planes(xd, ref, ref_buf->buf, mi_row, mi_col,
+                             &ref_buf->sf);
+      }
+    }
+
+#if CONFIG_CB4X4
+    av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL, bsize);
+#else
+    av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL,
+                                  AOMMAX(bsize, BLOCK_8X8));
+#endif
+
+#if CONFIG_MOTION_VAR
+    if (mbmi->motion_mode == OBMC_CAUSAL) {
+#if CONFIG_NCOBMC
+      av1_build_ncobmc_inter_predictors_sb(cm, xd, mi_row, mi_col);
+#else
+      av1_build_obmc_inter_predictors_sb(cm, xd, mi_row, mi_col);
+#endif
+    }
+#endif  // CONFIG_MOTION_VAR
+
+    // Reconstruction
+    if (!mbmi->skip) {
+      int eobtotal = 0;
+      int plane;
+
+      for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+        const struct macroblockd_plane *const pd = &xd->plane[plane];
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+        const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+        const BLOCK_SIZE plane_bsize =
+            AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#endif  // CONFIG_CHROMA_2X2
+#else
+        const BLOCK_SIZE plane_bsize =
+            get_plane_block_size(AOMMAX(BLOCK_8X8, bsize), pd);
+#endif
+        const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+        const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+        int row, col;
+
+#if CONFIG_CB4X4
+        if (!is_chroma_reference(mi_row, mi_col, bsize, pd->subsampling_x,
+                                 pd->subsampling_y))
+          continue;
+#endif
+
+#if CONFIG_VAR_TX
+        const TX_SIZE max_tx_size = get_vartx_max_txsize(mbmi, plane_bsize);
+        const int bh_var_tx = tx_size_high_unit[max_tx_size];
+        const int bw_var_tx = tx_size_wide_unit[max_tx_size];
+        for (row = 0; row < max_blocks_high; row += bh_var_tx)
+          for (col = 0; col < max_blocks_wide; col += bw_var_tx)
+            decode_reconstruct_tx(cm, xd, r, mbmi, plane, plane_bsize, row, col,
+                                  max_tx_size, &eobtotal);
+#else
+        const TX_SIZE tx_size = get_tx_size(plane, xd);
+        const int stepr = tx_size_high_unit[tx_size];
+        const int stepc = tx_size_wide_unit[tx_size];
+        for (row = 0; row < max_blocks_high; row += stepr)
+          for (col = 0; col < max_blocks_wide; col += stepc)
+            eobtotal += reconstruct_inter_block(cm, xd, r, mbmi->segment_id,
+                                                plane, row, col, tx_size);
+#endif
+      }
+    }
+  }
+#endif  // CONFIG_COEF_INTERLEAVE
+
+  int reader_corrupted_flag = aom_reader_has_error(r);
+  aom_merge_corrupted_flag(&xd->corrupted, reader_corrupted_flag);
+}
+
+#if CONFIG_NCOBMC && CONFIG_MOTION_VAR
+static void detoken_and_recon_sb(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+                                 int mi_row, int mi_col, aom_reader *r,
+                                 BLOCK_SIZE bsize) {
+  AV1_COMMON *const cm = &pbi->common;
+  const int hbs = mi_size_wide[bsize] >> 1;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+  BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+  const int has_rows = (mi_row + hbs) < cm->mi_rows;
+  const int has_cols = (mi_col + hbs) < cm->mi_cols;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  partition = get_partition(cm, mi_row, mi_col, bsize);
+  subsize = subsize_lookup[partition][bsize];
+
+  if (!hbs && !unify_bsize) {
+    xd->bmode_blocks_wl = 1 >> !!(partition & PARTITION_VERT);
+    xd->bmode_blocks_hl = 1 >> !!(partition & PARTITION_HORZ);
+    decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, subsize);
+  } else {
+    switch (partition) {
+      case PARTITION_NONE:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, bsize);
+        break;
+      case PARTITION_HORZ:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, subsize);
+        if (has_rows)
+          decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col, r,
+                                       subsize);
+        break;
+      case PARTITION_VERT:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, subsize);
+        if (has_cols)
+          decode_token_and_recon_block(pbi, xd, mi_row, mi_col + hbs, r,
+                                       subsize);
+        break;
+      case PARTITION_SPLIT:
+        detoken_and_recon_sb(pbi, xd, mi_row, mi_col, r, subsize);
+        detoken_and_recon_sb(pbi, xd, mi_row, mi_col + hbs, r, subsize);
+        detoken_and_recon_sb(pbi, xd, mi_row + hbs, mi_col, r, subsize);
+        detoken_and_recon_sb(pbi, xd, mi_row + hbs, mi_col + hbs, r, subsize);
+        break;
+#if CONFIG_EXT_PARTITION_TYPES
+      case PARTITION_HORZ_A:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col + hbs, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col, r, subsize);
+        break;
+      case PARTITION_HORZ_B:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, subsize);
+        decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col + hbs, r,
+                                     bsize2);
+        break;
+      case PARTITION_VERT_A:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col + hbs, r, subsize);
+        break;
+      case PARTITION_VERT_B:
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, subsize);
+        decode_token_and_recon_block(pbi, xd, mi_row, mi_col + hbs, r, bsize2);
+        decode_token_and_recon_block(pbi, xd, mi_row + hbs, mi_col + hbs, r,
+                                     bsize2);
+        break;
+#endif
+      default: assert(0 && "Invalid partition type");
+    }
+  }
+}
+#endif
+
+static void decode_block(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+#if CONFIG_SUPERTX
+                         int supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         int mi_row, int mi_col, aom_reader *r,
+#if CONFIG_EXT_PARTITION_TYPES
+                         PARTITION_TYPE partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                         BLOCK_SIZE bsize) {
+  decode_mbmi_block(pbi, xd,
+#if CONFIG_SUPERTX
+                    supertx_enabled,
+#endif
+                    mi_row, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                    partition,
+#endif
+                    bsize);
+#if !(CONFIG_MOTION_VAR && CONFIG_NCOBMC)
+#if CONFIG_SUPERTX
+  if (!supertx_enabled)
+#endif  // CONFIG_SUPERTX
+    decode_token_and_recon_block(pbi, xd, mi_row, mi_col, r, bsize);
+#endif
+}
+
+static PARTITION_TYPE read_partition(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                     int mi_row, int mi_col, aom_reader *r,
+                                     int has_rows, int has_cols,
+                                     BLOCK_SIZE bsize) {
+#if CONFIG_UNPOISON_PARTITION_CTX
+  const int ctx =
+      partition_plane_context(xd, mi_row, mi_col, has_rows, has_cols, bsize);
+  const aom_prob *const probs =
+      ctx < PARTITION_CONTEXTS ? cm->fc->partition_prob[ctx] : NULL;
+  FRAME_COUNTS *const counts = ctx < PARTITION_CONTEXTS ? xd->counts : NULL;
+#else
+  const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
+  const aom_prob *const probs = cm->fc->partition_prob[ctx];
+  FRAME_COUNTS *const counts = xd->counts;
+#endif
+  PARTITION_TYPE p;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob *partition_cdf = (ctx >= 0) ? ec_ctx->partition_cdf[ctx] : NULL;
+#endif
+
+  if (has_rows && has_cols)
+#if CONFIG_EXT_PARTITION_TYPES
+    if (bsize <= BLOCK_8X8)
+#if CONFIG_EC_MULTISYMBOL
+      p = (PARTITION_TYPE)aom_read_symbol(r, partition_cdf, PARTITION_TYPES,
+                                          ACCT_STR);
+#else
+      p = (PARTITION_TYPE)aom_read_tree(r, av1_partition_tree, probs, ACCT_STR);
+#endif
+    else
+#if CONFIG_EC_MULTISYMBOL
+      p = (PARTITION_TYPE)aom_read_symbol(r, partition_cdf, EXT_PARTITION_TYPES,
+                                          ACCT_STR);
+#else
+      p = (PARTITION_TYPE)aom_read_tree(r, av1_ext_partition_tree, probs,
+                                        ACCT_STR);
+#endif
+#else
+#if CONFIG_EC_MULTISYMBOL
+    p = (PARTITION_TYPE)aom_read_symbol(r, partition_cdf, PARTITION_TYPES,
+                                        ACCT_STR);
+#else
+    p = (PARTITION_TYPE)aom_read_tree(r, av1_partition_tree, probs, ACCT_STR);
+#endif
+#endif  // CONFIG_EXT_PARTITION_TYPES
+  else if (!has_rows && has_cols)
+    p = aom_read(r, probs[1], ACCT_STR) ? PARTITION_SPLIT : PARTITION_HORZ;
+  else if (has_rows && !has_cols)
+    p = aom_read(r, probs[2], ACCT_STR) ? PARTITION_SPLIT : PARTITION_VERT;
+  else
+    p = PARTITION_SPLIT;
+
+  if (counts) ++counts->partition[ctx][p];
+
+  return p;
+}
+
+#if CONFIG_SUPERTX
+static int read_skip(AV1_COMMON *cm, const MACROBLOCKD *xd, int segment_id,
+                     aom_reader *r) {
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
+    return 1;
+  } else {
+    const int ctx = av1_get_skip_context(xd);
+    const int skip = aom_read(r, cm->fc->skip_probs[ctx], ACCT_STR);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts) ++counts->skip[ctx][skip];
+    return skip;
+  }
+}
+#endif  // CONFIG_SUPERTX
+
+// TODO(slavarnway): eliminate bsize and subsize in future commits
+static void decode_partition(AV1Decoder *const pbi, MACROBLOCKD *const xd,
+#if CONFIG_SUPERTX
+                             int supertx_enabled,
+#endif
+                             int mi_row, int mi_col, aom_reader *r,
+                             BLOCK_SIZE bsize, int n4x4_l2) {
+  AV1_COMMON *const cm = &pbi->common;
+  const int n8x8_l2 = n4x4_l2 - 1;
+  const int num_8x8_wh = mi_size_wide[bsize];
+  const int hbs = num_8x8_wh >> 1;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+#if CONFIG_EXT_PARTITION_TYPES
+  BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+  const int has_rows = (mi_row + hbs) < cm->mi_rows;
+  const int has_cols = (mi_col + hbs) < cm->mi_cols;
+#if CONFIG_SUPERTX
+  const int read_token = !supertx_enabled;
+  int skip = 0;
+  TX_SIZE supertx_size = max_txsize_lookup[bsize];
+  const TileInfo *const tile = &xd->tile;
+  int txfm = DCT_DCT;
+#endif  // CONFIG_SUPERTX
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  partition = (bsize < BLOCK_8X8) ? PARTITION_NONE
+                                  : read_partition(cm, xd, mi_row, mi_col, r,
+                                                   has_rows, has_cols, bsize);
+  subsize = subsize_lookup[partition][bsize];  // get_subsize(bsize, partition);
+
+#if CONFIG_PVQ
+  assert(partition < PARTITION_TYPES);
+  assert(subsize < BLOCK_SIZES);
+#endif
+#if CONFIG_SUPERTX
+  if (!frame_is_intra_only(cm) && partition != PARTITION_NONE &&
+      bsize <= MAX_SUPERTX_BLOCK_SIZE && !supertx_enabled && !xd->lossless[0]) {
+    const int supertx_context = partition_supertx_context_lookup[partition];
+    supertx_enabled = aom_read(
+        r, cm->fc->supertx_prob[supertx_context][supertx_size], ACCT_STR);
+    if (xd->counts)
+      xd->counts->supertx[supertx_context][supertx_size][supertx_enabled]++;
+#if CONFIG_VAR_TX
+    if (supertx_enabled) xd->supertx_size = supertx_size;
+#endif
+  }
+#endif  // CONFIG_SUPERTX
+  if (!hbs && !unify_bsize) {
+    // calculate bmode block dimensions (log 2)
+    xd->bmode_blocks_wl = 1 >> !!(partition & PARTITION_VERT);
+    xd->bmode_blocks_hl = 1 >> !!(partition & PARTITION_HORZ);
+    decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                 supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                 mi_row, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                 partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                 subsize);
+  } else {
+    switch (partition) {
+      case PARTITION_NONE:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                     mi_row, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                     partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                     subsize);
+        break;
+      case PARTITION_HORZ:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                     mi_row, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                     partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                     subsize);
+        if (has_rows)
+          decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                       supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                       mi_row + hbs, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                       partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                       subsize);
+        break;
+      case PARTITION_VERT:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                     mi_row, mi_col, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                     partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                     subsize);
+        if (has_cols)
+          decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                       supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                       mi_row, mi_col + hbs, r,
+#if CONFIG_EXT_PARTITION_TYPES
+                       partition,
+#endif  // CONFIG_EXT_PARTITION_TYPES
+                       subsize);
+        break;
+      case PARTITION_SPLIT:
+        decode_partition(pbi, xd,
+#if CONFIG_SUPERTX
+                         supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         mi_row, mi_col, r, subsize, n8x8_l2);
+        decode_partition(pbi, xd,
+#if CONFIG_SUPERTX
+                         supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         mi_row, mi_col + hbs, r, subsize, n8x8_l2);
+        decode_partition(pbi, xd,
+#if CONFIG_SUPERTX
+                         supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         mi_row + hbs, mi_col, r, subsize, n8x8_l2);
+        decode_partition(pbi, xd,
+#if CONFIG_SUPERTX
+                         supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                         mi_row + hbs, mi_col + hbs, r, subsize, n8x8_l2);
+        break;
+#if CONFIG_EXT_PARTITION_TYPES
+      case PARTITION_HORZ_A:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col + hbs, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row + hbs, mi_col, r, partition, subsize);
+        break;
+      case PARTITION_HORZ_B:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col, r, partition, subsize);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row + hbs, mi_col, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row + hbs, mi_col + hbs, r, partition, bsize2);
+        break;
+      case PARTITION_VERT_A:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row + hbs, mi_col, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col + hbs, r, partition, subsize);
+        break;
+      case PARTITION_VERT_B:
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col, r, partition, subsize);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row, mi_col + hbs, r, partition, bsize2);
+        decode_block(pbi, xd,
+#if CONFIG_SUPERTX
+                     supertx_enabled,
+#endif
+                     mi_row + hbs, mi_col + hbs, r, partition, bsize2);
+        break;
+#endif
+      default: assert(0 && "Invalid partition type");
+    }
+  }
+
+#if CONFIG_SUPERTX
+  if (supertx_enabled && read_token) {
+    uint8_t *dst_buf[3];
+    int dst_stride[3], i;
+    int offset = mi_row * cm->mi_stride + mi_col;
+
+    set_segment_id_supertx(cm, mi_row, mi_col, bsize);
+
+#if CONFIG_DELTA_Q
+    if (cm->delta_q_present_flag) {
+      for (i = 0; i < MAX_SEGMENTS; i++) {
+        xd->plane[0].seg_dequant[i][0] =
+            av1_dc_quant(xd->current_qindex, cm->y_dc_delta_q, cm->bit_depth);
+        xd->plane[0].seg_dequant[i][1] =
+            av1_ac_quant(xd->current_qindex, 0, cm->bit_depth);
+        xd->plane[1].seg_dequant[i][0] =
+            av1_dc_quant(xd->current_qindex, cm->uv_dc_delta_q, cm->bit_depth);
+        xd->plane[1].seg_dequant[i][1] =
+            av1_ac_quant(xd->current_qindex, cm->uv_ac_delta_q, cm->bit_depth);
+        xd->plane[2].seg_dequant[i][0] =
+            av1_dc_quant(xd->current_qindex, cm->uv_dc_delta_q, cm->bit_depth);
+        xd->plane[2].seg_dequant[i][1] =
+            av1_ac_quant(xd->current_qindex, cm->uv_ac_delta_q, cm->bit_depth);
+      }
+    }
+#endif
+
+    xd->mi = cm->mi_grid_visible + offset;
+    xd->mi[0] = cm->mi + offset;
+    set_mi_row_col(xd, tile, mi_row, mi_size_high[bsize], mi_col,
+                   mi_size_wide[bsize],
+#if CONFIG_DEPENDENT_HORZTILES
+                   cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                   cm->mi_rows, cm->mi_cols);
+    set_skip_context(xd, mi_row, mi_col);
+    skip = read_skip(cm, xd, xd->mi[0]->mbmi.segment_id_supertx, r);
+    if (skip) {
+      reset_skip_context(xd, bsize);
+    } else {
+#if CONFIG_EXT_TX
+      if (get_ext_tx_types(supertx_size, bsize, 1, cm->reduced_tx_set_used) >
+          1) {
+        const int eset =
+            get_ext_tx_set(supertx_size, bsize, 1, cm->reduced_tx_set_used);
+        if (eset > 0) {
+          txfm = aom_read_tree(r, av1_ext_tx_inter_tree[eset],
+                               cm->fc->inter_ext_tx_prob[eset][supertx_size],
+                               ACCT_STR);
+          if (xd->counts) ++xd->counts->inter_ext_tx[eset][supertx_size][txfm];
+        }
+      }
+#else
+      if (supertx_size < TX_32X32) {
+        txfm = aom_read_tree(r, av1_ext_tx_tree,
+                             cm->fc->inter_ext_tx_prob[supertx_size], ACCT_STR);
+        if (xd->counts) ++xd->counts->inter_ext_tx[supertx_size][txfm];
+      }
+#endif  // CONFIG_EXT_TX
+    }
+
+    av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                         mi_col);
+    for (i = 0; i < MAX_MB_PLANE; i++) {
+      dst_buf[i] = xd->plane[i].dst.buf;
+      dst_stride[i] = xd->plane[i].dst.stride;
+    }
+    dec_predict_sb_complex(pbi, xd, tile, mi_row, mi_col, mi_row, mi_col, bsize,
+                           bsize, dst_buf, dst_stride);
+
+    if (!skip) {
+      int eobtotal = 0;
+      MB_MODE_INFO *mbmi;
+      set_offsets_topblock(cm, xd, tile, bsize, mi_row, mi_col);
+      mbmi = &xd->mi[0]->mbmi;
+      mbmi->tx_type = txfm;
+      assert(mbmi->segment_id_supertx != MAX_SEGMENTS);
+      for (i = 0; i < MAX_MB_PLANE; ++i) {
+        const struct macroblockd_plane *const pd = &xd->plane[i];
+        int row, col;
+        const TX_SIZE tx_size = get_tx_size(i, xd);
+        const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+        const int stepr = tx_size_high_unit[tx_size];
+        const int stepc = tx_size_wide_unit[tx_size];
+        const int max_blocks_wide = max_block_wide(xd, plane_bsize, i);
+        const int max_blocks_high = max_block_high(xd, plane_bsize, i);
+
+        for (row = 0; row < max_blocks_high; row += stepr)
+          for (col = 0; col < max_blocks_wide; col += stepc)
+            eobtotal += reconstruct_inter_block(
+                cm, xd, r, mbmi->segment_id_supertx, i, row, col, tx_size);
+      }
+      if ((unify_bsize || !(subsize < BLOCK_8X8)) && eobtotal == 0) skip = 1;
+    }
+    set_param_topblock(cm, xd, bsize, mi_row, mi_col, txfm, skip);
+  }
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_EXT_PARTITION_TYPES
+  update_ext_partition_context(xd, mi_row, mi_col, subsize, bsize, partition);
+#else
+  // update partition context
+  if (bsize >= BLOCK_8X8 &&
+      (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
+    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+#if CONFIG_CDEF
+#if CONFIG_EXT_PARTITION
+  if (cm->sb_size == BLOCK_128X128 && bsize == BLOCK_128X128) {
+    if (!sb_all_skip(cm, mi_row, mi_col)) {
+      cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col]->mbmi.cdef_strength =
+          aom_read_literal(r, cm->cdef_bits, ACCT_STR);
+    } else {
+      cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col]->mbmi.cdef_strength =
+          0;
+    }
+  } else if (cm->sb_size == BLOCK_64X64 && bsize == BLOCK_64X64) {
+#else
+  if (bsize == BLOCK_64X64) {
+#endif
+    if (!sb_all_skip(cm, mi_row, mi_col)) {
+      cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col]->mbmi.cdef_strength =
+          aom_read_literal(r, cm->cdef_bits, ACCT_STR);
+    } else {
+      cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col]->mbmi.cdef_strength =
+          -1;
+    }
+  }
+#endif  // CONFIG_CDEF
+}
+
+static void setup_bool_decoder(const uint8_t *data, const uint8_t *data_end,
+                               const size_t read_size,
+                               struct aom_internal_error_info *error_info,
+                               aom_reader *r,
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+                               int window_size,
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+                               aom_decrypt_cb decrypt_cb, void *decrypt_state) {
+  // Validate the calculated partition length. If the buffer
+  // described by the partition can't be fully read, then restrict
+  // it to the portion that can be (for EC mode) or throw an error.
+  if (!read_is_valid(data, read_size, data_end))
+    aom_internal_error(error_info, AOM_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt tile length");
+
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  r->window_size = window_size;
+#endif
+  if (aom_reader_init(r, data, read_size, decrypt_cb, decrypt_state))
+    aom_internal_error(error_info, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate bool decoder %d", 1);
+}
+
+#if !CONFIG_PVQ && !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET) && !CONFIG_LV_MAP
+static void read_coef_probs_common(av1_coeff_probs_model *coef_probs,
+                                   aom_reader *r) {
+  int i, j, k, l, m;
+#if CONFIG_EC_ADAPT
+  const int node_limit = UNCONSTRAINED_NODES - 1;
+#else
+  const int node_limit = UNCONSTRAINED_NODES;
+#endif
+
+  if (aom_read_bit(r, ACCT_STR))
+    for (i = 0; i < PLANE_TYPES; ++i)
+      for (j = 0; j < REF_TYPES; ++j)
+        for (k = 0; k < COEF_BANDS; ++k)
+          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
+            for (m = 0; m < node_limit; ++m)
+              av1_diff_update_prob(r, &coef_probs[i][j][k][l][m], ACCT_STR);
+}
+
+static void read_coef_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode, aom_reader *r) {
+  const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+  TX_SIZE tx_size;
+  for (tx_size = 0; tx_size <= max_tx_size; ++tx_size)
+    read_coef_probs_common(fc->coef_probs[tx_size], r);
+}
+#endif
+
+static void setup_segmentation(AV1_COMMON *const cm,
+                               struct aom_read_bit_buffer *rb) {
+  struct segmentation *const seg = &cm->seg;
+  int i, j;
+
+  seg->update_map = 0;
+  seg->update_data = 0;
+
+  seg->enabled = aom_rb_read_bit(rb);
+  if (!seg->enabled) return;
+
+  // Segmentation map update
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
+    seg->update_map = 1;
+  } else {
+    seg->update_map = aom_rb_read_bit(rb);
+  }
+  if (seg->update_map) {
+    if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
+      seg->temporal_update = 0;
+    } else {
+      seg->temporal_update = aom_rb_read_bit(rb);
+    }
+  }
+
+  // Segmentation data update
+  seg->update_data = aom_rb_read_bit(rb);
+  if (seg->update_data) {
+    seg->abs_delta = aom_rb_read_bit(rb);
+
+    av1_clearall_segfeatures(seg);
+
+    for (i = 0; i < MAX_SEGMENTS; i++) {
+      for (j = 0; j < SEG_LVL_MAX; j++) {
+        int data = 0;
+        const int feature_enabled = aom_rb_read_bit(rb);
+        if (feature_enabled) {
+          av1_enable_segfeature(seg, i, j);
+          data = decode_unsigned_max(rb, av1_seg_feature_data_max(j));
+          if (av1_is_segfeature_signed(j))
+            data = aom_rb_read_bit(rb) ? -data : data;
+        }
+        av1_set_segdata(seg, i, j, data);
+      }
+    }
+  }
+}
+
+#if CONFIG_LOOP_RESTORATION
+static void decode_restoration_mode(AV1_COMMON *cm,
+                                    struct aom_read_bit_buffer *rb) {
+  int p;
+  RestorationInfo *rsi = &cm->rst_info[0];
+  if (aom_rb_read_bit(rb)) {
+    rsi->frame_restoration_type =
+        aom_rb_read_bit(rb) ? RESTORE_SGRPROJ : RESTORE_WIENER;
+  } else {
+    rsi->frame_restoration_type =
+        aom_rb_read_bit(rb) ? RESTORE_SWITCHABLE : RESTORE_NONE;
+  }
+  for (p = 1; p < MAX_MB_PLANE; ++p) {
+    cm->rst_info[p].frame_restoration_type =
+        aom_rb_read_bit(rb) ? RESTORE_WIENER : RESTORE_NONE;
+  }
+
+  cm->rst_info[0].restoration_tilesize = RESTORATION_TILESIZE_MAX;
+  cm->rst_info[1].restoration_tilesize = RESTORATION_TILESIZE_MAX;
+  cm->rst_info[2].restoration_tilesize = RESTORATION_TILESIZE_MAX;
+  if (cm->rst_info[0].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[1].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[2].frame_restoration_type != RESTORE_NONE) {
+    rsi = &cm->rst_info[0];
+    rsi->restoration_tilesize >>= aom_rb_read_bit(rb);
+    if (rsi->restoration_tilesize != RESTORATION_TILESIZE_MAX) {
+      rsi->restoration_tilesize >>= aom_rb_read_bit(rb);
+    }
+    cm->rst_info[1].restoration_tilesize = cm->rst_info[0].restoration_tilesize;
+    cm->rst_info[2].restoration_tilesize = cm->rst_info[0].restoration_tilesize;
+  }
+}
+
+static void read_wiener_filter(WienerInfo *wiener_info,
+                               WienerInfo *ref_wiener_info, aom_reader *rb) {
+  wiener_info->vfilter[0] = wiener_info->vfilter[WIENER_WIN - 1] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
+          WIENER_FILT_TAP0_SUBEXP_K,
+          ref_wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV) +
+      WIENER_FILT_TAP0_MINV;
+  wiener_info->vfilter[1] = wiener_info->vfilter[WIENER_WIN - 2] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
+          WIENER_FILT_TAP1_SUBEXP_K,
+          ref_wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV) +
+      WIENER_FILT_TAP1_MINV;
+  wiener_info->vfilter[2] = wiener_info->vfilter[WIENER_WIN - 3] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
+          WIENER_FILT_TAP2_SUBEXP_K,
+          ref_wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV) +
+      WIENER_FILT_TAP2_MINV;
+  // The central element has an implicit +WIENER_FILT_STEP
+  wiener_info->vfilter[WIENER_HALFWIN] =
+      -2 * (wiener_info->vfilter[0] + wiener_info->vfilter[1] +
+            wiener_info->vfilter[2]);
+
+  wiener_info->hfilter[0] = wiener_info->hfilter[WIENER_WIN - 1] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
+          WIENER_FILT_TAP0_SUBEXP_K,
+          ref_wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV) +
+      WIENER_FILT_TAP0_MINV;
+  wiener_info->hfilter[1] = wiener_info->hfilter[WIENER_WIN - 2] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
+          WIENER_FILT_TAP1_SUBEXP_K,
+          ref_wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV) +
+      WIENER_FILT_TAP1_MINV;
+  wiener_info->hfilter[2] = wiener_info->hfilter[WIENER_WIN - 3] =
+      aom_read_primitive_refsubexpfin(
+          rb, WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
+          WIENER_FILT_TAP2_SUBEXP_K,
+          ref_wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV) +
+      WIENER_FILT_TAP2_MINV;
+  // The central element has an implicit +WIENER_FILT_STEP
+  wiener_info->hfilter[WIENER_HALFWIN] =
+      -2 * (wiener_info->hfilter[0] + wiener_info->hfilter[1] +
+            wiener_info->hfilter[2]);
+  memcpy(ref_wiener_info, wiener_info, sizeof(*wiener_info));
+}
+
+static void read_sgrproj_filter(SgrprojInfo *sgrproj_info,
+                                SgrprojInfo *ref_sgrproj_info, aom_reader *rb) {
+  sgrproj_info->ep = aom_read_literal(rb, SGRPROJ_PARAMS_BITS, ACCT_STR);
+  sgrproj_info->xqd[0] =
+      aom_read_primitive_refsubexpfin(
+          rb, SGRPROJ_PRJ_MAX0 - SGRPROJ_PRJ_MIN0 + 1, SGRPROJ_PRJ_SUBEXP_K,
+          ref_sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0) +
+      SGRPROJ_PRJ_MIN0;
+  sgrproj_info->xqd[1] =
+      aom_read_primitive_refsubexpfin(
+          rb, SGRPROJ_PRJ_MAX1 - SGRPROJ_PRJ_MIN1 + 1, SGRPROJ_PRJ_SUBEXP_K,
+          ref_sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1) +
+      SGRPROJ_PRJ_MIN1;
+  memcpy(ref_sgrproj_info, sgrproj_info, sizeof(*sgrproj_info));
+}
+
+static void decode_restoration(AV1_COMMON *cm, aom_reader *rb) {
+  int i, p;
+  SgrprojInfo ref_sgrproj_info;
+  WienerInfo ref_wiener_info;
+  set_default_wiener(&ref_wiener_info);
+  set_default_sgrproj(&ref_sgrproj_info);
+  const int ntiles = av1_get_rest_ntiles(cm->width, cm->height,
+                                         cm->rst_info[0].restoration_tilesize,
+                                         NULL, NULL, NULL, NULL);
+  const int ntiles_uv = av1_get_rest_ntiles(
+      ROUND_POWER_OF_TWO(cm->width, cm->subsampling_x),
+      ROUND_POWER_OF_TWO(cm->height, cm->subsampling_y),
+      cm->rst_info[1].restoration_tilesize, NULL, NULL, NULL, NULL);
+  RestorationInfo *rsi = &cm->rst_info[0];
+  if (rsi->frame_restoration_type != RESTORE_NONE) {
+    if (rsi->frame_restoration_type == RESTORE_SWITCHABLE) {
+      for (i = 0; i < ntiles; ++i) {
+        rsi->restoration_type[i] =
+            aom_read_tree(rb, av1_switchable_restore_tree,
+                          cm->fc->switchable_restore_prob, ACCT_STR);
+        if (rsi->restoration_type[i] == RESTORE_WIENER) {
+          read_wiener_filter(&rsi->wiener_info[i], &ref_wiener_info, rb);
+        } else if (rsi->restoration_type[i] == RESTORE_SGRPROJ) {
+          read_sgrproj_filter(&rsi->sgrproj_info[i], &ref_sgrproj_info, rb);
+        }
+      }
+    } else if (rsi->frame_restoration_type == RESTORE_WIENER) {
+      for (i = 0; i < ntiles; ++i) {
+        if (aom_read(rb, RESTORE_NONE_WIENER_PROB, ACCT_STR)) {
+          rsi->restoration_type[i] = RESTORE_WIENER;
+          read_wiener_filter(&rsi->wiener_info[i], &ref_wiener_info, rb);
+        } else {
+          rsi->restoration_type[i] = RESTORE_NONE;
+        }
+      }
+    } else if (rsi->frame_restoration_type == RESTORE_SGRPROJ) {
+      for (i = 0; i < ntiles; ++i) {
+        if (aom_read(rb, RESTORE_NONE_SGRPROJ_PROB, ACCT_STR)) {
+          rsi->restoration_type[i] = RESTORE_SGRPROJ;
+          read_sgrproj_filter(&rsi->sgrproj_info[i], &ref_sgrproj_info, rb);
+        } else {
+          rsi->restoration_type[i] = RESTORE_NONE;
+        }
+      }
+    }
+  }
+  for (p = 1; p < MAX_MB_PLANE; ++p) {
+    set_default_wiener(&ref_wiener_info);
+    rsi = &cm->rst_info[p];
+    if (rsi->frame_restoration_type == RESTORE_WIENER) {
+      for (i = 0; i < ntiles_uv; ++i) {
+        if (ntiles_uv > 1)
+          rsi->restoration_type[i] =
+              aom_read(rb, RESTORE_NONE_WIENER_PROB, ACCT_STR) ? RESTORE_WIENER
+                                                               : RESTORE_NONE;
+        else
+          rsi->restoration_type[i] = RESTORE_WIENER;
+        if (rsi->restoration_type[i] == RESTORE_WIENER) {
+          read_wiener_filter(&rsi->wiener_info[i], &ref_wiener_info, rb);
+        }
+      }
+    }
+  }
+}
+#endif  // CONFIG_LOOP_RESTORATION
+
+static void setup_loopfilter(AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
+  struct loopfilter *lf = &cm->lf;
+  lf->filter_level = aom_rb_read_literal(rb, 6);
+  lf->sharpness_level = aom_rb_read_literal(rb, 3);
+
+  // Read in loop filter deltas applied at the MB level based on mode or ref
+  // frame.
+  lf->mode_ref_delta_update = 0;
+
+  lf->mode_ref_delta_enabled = aom_rb_read_bit(rb);
+  if (lf->mode_ref_delta_enabled) {
+    lf->mode_ref_delta_update = aom_rb_read_bit(rb);
+    if (lf->mode_ref_delta_update) {
+      int i;
+
+      for (i = 0; i < TOTAL_REFS_PER_FRAME; i++)
+        if (aom_rb_read_bit(rb))
+          lf->ref_deltas[i] = aom_rb_read_inv_signed_literal(rb, 6);
+
+      for (i = 0; i < MAX_MODE_LF_DELTAS; i++)
+        if (aom_rb_read_bit(rb))
+          lf->mode_deltas[i] = aom_rb_read_inv_signed_literal(rb, 6);
+    }
+  }
+}
+
+#if CONFIG_CDEF
+static void setup_cdef(AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
+  int i;
+  cm->cdef_dering_damping = aom_rb_read_literal(rb, 1) + 5;
+  cm->cdef_clpf_damping = aom_rb_read_literal(rb, 2) + 3;
+  cm->cdef_bits = aom_rb_read_literal(rb, 2);
+  cm->nb_cdef_strengths = 1 << cm->cdef_bits;
+  for (i = 0; i < cm->nb_cdef_strengths; i++) {
+    cm->cdef_strengths[i] = aom_rb_read_literal(rb, CDEF_STRENGTH_BITS);
+    cm->cdef_uv_strengths[i] = aom_rb_read_literal(rb, CDEF_STRENGTH_BITS);
+  }
+}
+#endif  // CONFIG_CDEF
+
+static INLINE int read_delta_q(struct aom_read_bit_buffer *rb) {
+  return aom_rb_read_bit(rb) ? aom_rb_read_inv_signed_literal(rb, 6) : 0;
+}
+
+static void setup_quantization(AV1_COMMON *const cm,
+                               struct aom_read_bit_buffer *rb) {
+  cm->base_qindex = aom_rb_read_literal(rb, QINDEX_BITS);
+  cm->y_dc_delta_q = read_delta_q(rb);
+  cm->uv_dc_delta_q = read_delta_q(rb);
+  cm->uv_ac_delta_q = read_delta_q(rb);
+  cm->dequant_bit_depth = cm->bit_depth;
+#if CONFIG_AOM_QM
+  cm->using_qmatrix = aom_rb_read_bit(rb);
+  if (cm->using_qmatrix) {
+    cm->min_qmlevel = aom_rb_read_literal(rb, QM_LEVEL_BITS);
+    cm->max_qmlevel = aom_rb_read_literal(rb, QM_LEVEL_BITS);
+  } else {
+    cm->min_qmlevel = 0;
+    cm->max_qmlevel = 0;
+  }
+#endif
+}
+
+static void setup_segmentation_dequant(AV1_COMMON *const cm) {
+  // Build y/uv dequant values based on segmentation.
+  int i = 0;
+#if CONFIG_AOM_QM
+  int lossless;
+  int j = 0;
+  int qmlevel;
+  int using_qm = cm->using_qmatrix;
+  int minqm = cm->min_qmlevel;
+  int maxqm = cm->max_qmlevel;
+#endif
+#if CONFIG_NEW_QUANT
+  int b;
+  int dq;
+#endif  //  CONFIG_NEW_QUANT
+  if (cm->seg.enabled) {
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      const int qindex = av1_get_qindex(&cm->seg, i, cm->base_qindex);
+      cm->y_dequant[i][0] =
+          av1_dc_quant(qindex, cm->y_dc_delta_q, cm->bit_depth);
+      cm->y_dequant[i][1] = av1_ac_quant(qindex, 0, cm->bit_depth);
+      cm->uv_dequant[i][0] =
+          av1_dc_quant(qindex, cm->uv_dc_delta_q, cm->bit_depth);
+      cm->uv_dequant[i][1] =
+          av1_ac_quant(qindex, cm->uv_ac_delta_q, cm->bit_depth);
+#if CONFIG_AOM_QM
+      lossless = qindex == 0 && cm->y_dc_delta_q == 0 &&
+                 cm->uv_dc_delta_q == 0 && cm->uv_ac_delta_q == 0;
+      // NB: depends on base index so there is only 1 set per frame
+      // No quant weighting when lossless or signalled not using QM
+      qmlevel = (lossless || using_qm == 0)
+                    ? NUM_QM_LEVELS - 1
+                    : aom_get_qmlevel(cm->base_qindex, minqm, maxqm);
+      for (j = 0; j < TX_SIZES; ++j) {
+        cm->y_iqmatrix[i][1][j] = aom_iqmatrix(cm, qmlevel, 0, j, 1);
+        cm->y_iqmatrix[i][0][j] = aom_iqmatrix(cm, qmlevel, 0, j, 0);
+        cm->uv_iqmatrix[i][1][j] = aom_iqmatrix(cm, qmlevel, 1, j, 1);
+        cm->uv_iqmatrix[i][0][j] = aom_iqmatrix(cm, qmlevel, 1, j, 0);
+      }
+#endif  // CONFIG_AOM_QM
+#if CONFIG_NEW_QUANT
+      for (dq = 0; dq < QUANT_PROFILES; dq++) {
+        for (b = 0; b < COEF_BANDS; ++b) {
+          av1_get_dequant_val_nuq(cm->y_dequant[i][b != 0], b,
+                                  cm->y_dequant_nuq[i][dq][b], NULL, dq);
+          av1_get_dequant_val_nuq(cm->uv_dequant[i][b != 0], b,
+                                  cm->uv_dequant_nuq[i][dq][b], NULL, dq);
+        }
+      }
+#endif  //  CONFIG_NEW_QUANT
+    }
+  } else {
+    const int qindex = cm->base_qindex;
+    // When segmentation is disabled, only the first value is used.  The
+    // remaining are don't cares.
+    cm->y_dequant[0][0] = av1_dc_quant(qindex, cm->y_dc_delta_q, cm->bit_depth);
+    cm->y_dequant[0][1] = av1_ac_quant(qindex, 0, cm->bit_depth);
+    cm->uv_dequant[0][0] =
+        av1_dc_quant(qindex, cm->uv_dc_delta_q, cm->bit_depth);
+    cm->uv_dequant[0][1] =
+        av1_ac_quant(qindex, cm->uv_ac_delta_q, cm->bit_depth);
+#if CONFIG_AOM_QM
+    lossless = qindex == 0 && cm->y_dc_delta_q == 0 && cm->uv_dc_delta_q == 0 &&
+               cm->uv_ac_delta_q == 0;
+    // No quant weighting when lossless or signalled not using QM
+    qmlevel = (lossless || using_qm == 0)
+                  ? NUM_QM_LEVELS - 1
+                  : aom_get_qmlevel(cm->base_qindex, minqm, maxqm);
+    for (j = 0; j < TX_SIZES; ++j) {
+      cm->y_iqmatrix[i][1][j] = aom_iqmatrix(cm, qmlevel, 0, j, 1);
+      cm->y_iqmatrix[i][0][j] = aom_iqmatrix(cm, qmlevel, 0, j, 0);
+      cm->uv_iqmatrix[i][1][j] = aom_iqmatrix(cm, qmlevel, 1, j, 1);
+      cm->uv_iqmatrix[i][0][j] = aom_iqmatrix(cm, qmlevel, 1, j, 0);
+    }
+#endif
+#if CONFIG_NEW_QUANT
+    for (dq = 0; dq < QUANT_PROFILES; dq++) {
+      for (b = 0; b < COEF_BANDS; ++b) {
+        av1_get_dequant_val_nuq(cm->y_dequant[0][b != 0], b,
+                                cm->y_dequant_nuq[0][dq][b], NULL, dq);
+        av1_get_dequant_val_nuq(cm->uv_dequant[0][b != 0], b,
+                                cm->uv_dequant_nuq[0][dq][b], NULL, dq);
+      }
+    }
+#endif  //  CONFIG_NEW_QUANT
+  }
+}
+
+static InterpFilter read_frame_interp_filter(struct aom_read_bit_buffer *rb) {
+  return aom_rb_read_bit(rb) ? SWITCHABLE
+                             : aom_rb_read_literal(rb, LOG_SWITCHABLE_FILTERS);
+}
+
+static void setup_render_size(AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
+  cm->render_width = cm->width;
+  cm->render_height = cm->height;
+  if (aom_rb_read_bit(rb))
+    av1_read_frame_size(rb, &cm->render_width, &cm->render_height);
+}
+
+#if CONFIG_FRAME_SUPERRES
+// TODO(afergs): make "struct aom_read_bit_buffer *const rb"?
+static void setup_superres_size(AV1_COMMON *const cm,
+                                struct aom_read_bit_buffer *rb, int *width,
+                                int *height) {
+  // TODO(afergs): Test this behaviour
+  // Frame superres is probably in compatible with this render resolution
+  assert(cm->width == cm->render_width && cm->height == cm->render_height);
+
+  cm->superres_width = cm->width;
+  cm->superres_height = cm->height;
+  if (aom_rb_read_bit(rb)) {
+    cm->superres_scale_numerator =
+        (uint8_t)aom_rb_read_literal(rb, SUPERRES_SCALE_BITS);
+    cm->superres_scale_numerator += SUPERRES_SCALE_NUMERATOR_MIN;
+    // Don't edit cm->width or cm->height directly, or the buffers won't get
+    // resized correctly
+    // TODO(afergs): Should the render resolution not be modified? It's the same
+    // by default (ie. when it isn't sent)...
+    // resize_context_buffers() will change cm->width to equal cm->render_width,
+    // then they'll be the same again
+    *width = cm->render_width =
+        cm->width * cm->superres_scale_numerator / SUPERRES_SCALE_DENOMINATOR;
+    *height = cm->render_height =
+        cm->height * cm->superres_scale_numerator / SUPERRES_SCALE_DENOMINATOR;
+  } else {
+    // 1:1 scaling - ie. no scaling, scale not provided
+    cm->superres_scale_numerator = SUPERRES_SCALE_DENOMINATOR;
+  }
+}
+#endif  // CONFIG_FRAME_SUPERRES
+
+static void resize_mv_buffer(AV1_COMMON *cm) {
+  aom_free(cm->cur_frame->mvs);
+  cm->cur_frame->mi_rows = cm->mi_rows;
+  cm->cur_frame->mi_cols = cm->mi_cols;
+  CHECK_MEM_ERROR(cm, cm->cur_frame->mvs,
+                  (MV_REF *)aom_calloc(cm->mi_rows * cm->mi_cols,
+                                       sizeof(*cm->cur_frame->mvs)));
+}
+
+static void resize_context_buffers(AV1_COMMON *cm, int width, int height) {
+#if CONFIG_SIZE_LIMIT
+  if (width > DECODE_WIDTH_LIMIT || height > DECODE_HEIGHT_LIMIT)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Dimensions of %dx%d beyond allowed size of %dx%d.",
+                       width, height, DECODE_WIDTH_LIMIT, DECODE_HEIGHT_LIMIT);
+#endif
+  if (cm->width != width || cm->height != height) {
+    const int new_mi_rows =
+        ALIGN_POWER_OF_TWO(height, MI_SIZE_LOG2) >> MI_SIZE_LOG2;
+    const int new_mi_cols =
+        ALIGN_POWER_OF_TWO(width, MI_SIZE_LOG2) >> MI_SIZE_LOG2;
+
+    // Allocations in av1_alloc_context_buffers() depend on individual
+    // dimensions as well as the overall size.
+    if (new_mi_cols > cm->mi_cols || new_mi_rows > cm->mi_rows) {
+      if (av1_alloc_context_buffers(cm, width, height))
+        aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                           "Failed to allocate context buffers");
+    } else {
+      av1_set_mb_mi(cm, width, height);
+    }
+    av1_init_context_buffers(cm);
+    cm->width = width;
+    cm->height = height;
+  }
+  if (cm->cur_frame->mvs == NULL || cm->mi_rows > cm->cur_frame->mi_rows ||
+      cm->mi_cols > cm->cur_frame->mi_cols) {
+    resize_mv_buffer(cm);
+  }
+}
+
+static void setup_frame_size(AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
+  int width, height;
+  BufferPool *const pool = cm->buffer_pool;
+  av1_read_frame_size(rb, &width, &height);
+  setup_render_size(cm, rb);
+#if CONFIG_FRAME_SUPERRES
+  setup_superres_size(cm, rb, &width, &height);
+#endif  // CONFIG_FRAME_SUPERRES
+  resize_context_buffers(cm, width, height);
+
+  lock_buffer_pool(pool);
+  if (aom_realloc_frame_buffer(
+          get_frame_new_buffer(cm), cm->width, cm->height, cm->subsampling_x,
+          cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+          cm->use_highbitdepth,
+#endif
+          AOM_BORDER_IN_PIXELS, cm->byte_alignment,
+          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
+          pool->cb_priv)) {
+    unlock_buffer_pool(pool);
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate frame buffer");
+  }
+  unlock_buffer_pool(pool);
+
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
+  pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_space = cm->color_space;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_range = cm->color_range;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_width = cm->render_width;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_height = cm->render_height;
+}
+
+static INLINE int valid_ref_frame_img_fmt(aom_bit_depth_t ref_bit_depth,
+                                          int ref_xss, int ref_yss,
+                                          aom_bit_depth_t this_bit_depth,
+                                          int this_xss, int this_yss) {
+  return ref_bit_depth == this_bit_depth && ref_xss == this_xss &&
+         ref_yss == this_yss;
+}
+
+static void setup_frame_size_with_refs(AV1_COMMON *cm,
+                                       struct aom_read_bit_buffer *rb) {
+  int width, height;
+  int found = 0, i;
+  int has_valid_ref_frame = 0;
+  BufferPool *const pool = cm->buffer_pool;
+  for (i = 0; i < INTER_REFS_PER_FRAME; ++i) {
+    if (aom_rb_read_bit(rb)) {
+      YV12_BUFFER_CONFIG *const buf = cm->frame_refs[i].buf;
+      width = buf->y_crop_width;
+      height = buf->y_crop_height;
+      cm->render_width = buf->render_width;
+      cm->render_height = buf->render_height;
+      found = 1;
+      break;
+    }
+  }
+
+  if (!found) {
+    av1_read_frame_size(rb, &width, &height);
+    setup_render_size(cm, rb);
+  }
+
+  if (width <= 0 || height <= 0)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Invalid frame size");
+
+  // Check to make sure at least one of frames that this frame references
+  // has valid dimensions.
+  for (i = 0; i < INTER_REFS_PER_FRAME; ++i) {
+    RefBuffer *const ref_frame = &cm->frame_refs[i];
+    has_valid_ref_frame |=
+        valid_ref_frame_size(ref_frame->buf->y_crop_width,
+                             ref_frame->buf->y_crop_height, width, height);
+  }
+  if (!has_valid_ref_frame)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Referenced frame has invalid size");
+  for (i = 0; i < INTER_REFS_PER_FRAME; ++i) {
+    RefBuffer *const ref_frame = &cm->frame_refs[i];
+    if (!valid_ref_frame_img_fmt(ref_frame->buf->bit_depth,
+                                 ref_frame->buf->subsampling_x,
+                                 ref_frame->buf->subsampling_y, cm->bit_depth,
+                                 cm->subsampling_x, cm->subsampling_y))
+      aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                         "Referenced frame has incompatible color format");
+  }
+
+  resize_context_buffers(cm, width, height);
+
+  lock_buffer_pool(pool);
+  if (aom_realloc_frame_buffer(
+          get_frame_new_buffer(cm), cm->width, cm->height, cm->subsampling_x,
+          cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+          cm->use_highbitdepth,
+#endif
+          AOM_BORDER_IN_PIXELS, cm->byte_alignment,
+          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
+          pool->cb_priv)) {
+    unlock_buffer_pool(pool);
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate frame buffer");
+  }
+  unlock_buffer_pool(pool);
+
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_x = cm->subsampling_x;
+  pool->frame_bufs[cm->new_fb_idx].buf.subsampling_y = cm->subsampling_y;
+  pool->frame_bufs[cm->new_fb_idx].buf.bit_depth = (unsigned int)cm->bit_depth;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_space = cm->color_space;
+  pool->frame_bufs[cm->new_fb_idx].buf.color_range = cm->color_range;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_width = cm->render_width;
+  pool->frame_bufs[cm->new_fb_idx].buf.render_height = cm->render_height;
+}
+
+static void read_tile_info(AV1Decoder *const pbi,
+                           struct aom_read_bit_buffer *const rb) {
+  AV1_COMMON *const cm = &pbi->common;
+#if CONFIG_EXT_TILE
+  cm->tile_encoding_mode = aom_rb_read_literal(rb, 1);
+// Read the tile width/height
+#if CONFIG_EXT_PARTITION
+  if (cm->sb_size == BLOCK_128X128) {
+    cm->tile_width = aom_rb_read_literal(rb, 5) + 1;
+    cm->tile_height = aom_rb_read_literal(rb, 5) + 1;
+  } else
+#endif  // CONFIG_EXT_PARTITION
+  {
+    cm->tile_width = aom_rb_read_literal(rb, 6) + 1;
+    cm->tile_height = aom_rb_read_literal(rb, 6) + 1;
+  }
+
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  cm->loop_filter_across_tiles_enabled = aom_rb_read_bit(rb);
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+  cm->tile_width <<= cm->mib_size_log2;
+  cm->tile_height <<= cm->mib_size_log2;
+
+  cm->tile_width = AOMMIN(cm->tile_width, cm->mi_cols);
+  cm->tile_height = AOMMIN(cm->tile_height, cm->mi_rows);
+
+  // Get the number of tiles
+  cm->tile_cols = 1;
+  while (cm->tile_cols * cm->tile_width < cm->mi_cols) ++cm->tile_cols;
+
+  cm->tile_rows = 1;
+  while (cm->tile_rows * cm->tile_height < cm->mi_rows) ++cm->tile_rows;
+
+  if (cm->tile_cols * cm->tile_rows > 1) {
+    // Read the number of bytes used to store tile size
+    pbi->tile_col_size_bytes = aom_rb_read_literal(rb, 2) + 1;
+    pbi->tile_size_bytes = aom_rb_read_literal(rb, 2) + 1;
+  }
+
+#if CONFIG_DEPENDENT_HORZTILES
+  if (cm->tile_rows <= 1)
+    cm->dependent_horz_tiles = aom_rb_read_bit(rb);
+  else
+    cm->dependent_horz_tiles = 0;
+#endif
+#else
+  int min_log2_tile_cols, max_log2_tile_cols, max_ones;
+  av1_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
+
+  // columns
+  max_ones = max_log2_tile_cols - min_log2_tile_cols;
+  cm->log2_tile_cols = min_log2_tile_cols;
+  while (max_ones-- && aom_rb_read_bit(rb)) cm->log2_tile_cols++;
+
+  if (cm->log2_tile_cols > 6)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Invalid number of tile columns");
+
+  // rows
+  cm->log2_tile_rows = aom_rb_read_bit(rb);
+  if (cm->log2_tile_rows) cm->log2_tile_rows += aom_rb_read_bit(rb);
+#if CONFIG_DEPENDENT_HORZTILES
+  if (cm->log2_tile_rows != 0)
+    cm->dependent_horz_tiles = aom_rb_read_bit(rb);
+  else
+    cm->dependent_horz_tiles = 0;
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  cm->loop_filter_across_tiles_enabled = aom_rb_read_bit(rb);
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+  cm->tile_cols = 1 << cm->log2_tile_cols;
+  cm->tile_rows = 1 << cm->log2_tile_rows;
+
+  cm->tile_width = ALIGN_POWER_OF_TWO(cm->mi_cols, MAX_MIB_SIZE_LOG2);
+  cm->tile_width >>= cm->log2_tile_cols;
+  cm->tile_height = ALIGN_POWER_OF_TWO(cm->mi_rows, MAX_MIB_SIZE_LOG2);
+  cm->tile_height >>= cm->log2_tile_rows;
+
+  // round to integer multiples of superblock size
+  cm->tile_width = ALIGN_POWER_OF_TWO(cm->tile_width, MAX_MIB_SIZE_LOG2);
+  cm->tile_height = ALIGN_POWER_OF_TWO(cm->tile_height, MAX_MIB_SIZE_LOG2);
+
+// tile size magnitude
+#if !CONFIG_TILE_GROUPS
+  if (cm->tile_rows > 1 || cm->tile_cols > 1)
+#endif
+    pbi->tile_size_bytes = aom_rb_read_literal(rb, 2) + 1;
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_TILE_GROUPS
+  // Store an index to the location of the tile group information
+  pbi->tg_size_bit_offset = rb->bit_offset;
+  pbi->tg_size = 1 << (cm->log2_tile_rows + cm->log2_tile_cols);
+  if (cm->log2_tile_rows + cm->log2_tile_cols > 0) {
+    pbi->tg_start =
+        aom_rb_read_literal(rb, cm->log2_tile_rows + cm->log2_tile_cols);
+    pbi->tg_size =
+        1 + aom_rb_read_literal(rb, cm->log2_tile_rows + cm->log2_tile_cols);
+  }
+#endif
+}
+
+static int mem_get_varsize(const uint8_t *src, int sz) {
+  switch (sz) {
+    case 1: return src[0];
+    case 2: return mem_get_le16(src);
+    case 3: return mem_get_le24(src);
+    case 4: return mem_get_le32(src);
+    default: assert("Invalid size" && 0); return -1;
+  }
+}
+
+#if CONFIG_EXT_TILE
+// Reads the next tile returning its size and adjusting '*data' accordingly
+// based on 'is_last'.
+static void get_tile_buffer(const uint8_t *const data_end,
+                            struct aom_internal_error_info *error_info,
+                            const uint8_t **data, aom_decrypt_cb decrypt_cb,
+                            void *decrypt_state,
+                            TileBufferDec (*const tile_buffers)[MAX_TILE_COLS],
+                            int tile_size_bytes, int col, int row,
+                            unsigned int tile_encoding_mode) {
+  size_t size;
+
+  size_t copy_size = 0;
+  const uint8_t *copy_data = NULL;
+
+  if (!read_is_valid(*data, tile_size_bytes, data_end))
+    aom_internal_error(error_info, AOM_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt tile length");
+  if (decrypt_cb) {
+    uint8_t be_data[4];
+    decrypt_cb(decrypt_state, *data, be_data, tile_size_bytes);
+
+    // Only read number of bytes in cm->tile_size_bytes.
+    size = mem_get_varsize(be_data, tile_size_bytes);
+  } else {
+    size = mem_get_varsize(*data, tile_size_bytes);
+  }
+
+  // If cm->tile_encoding_mode = 1 (i.e. TILE_VR), then the top bit of the tile
+  // header indicates copy mode.
+  if (tile_encoding_mode && (size >> (tile_size_bytes * 8 - 1)) == 1) {
+    // The remaining bits in the top byte signal the row offset
+    int offset = (size >> (tile_size_bytes - 1) * 8) & 0x7f;
+
+    // Currently, only use tiles in same column as reference tiles.
+    copy_data = tile_buffers[row - offset][col].data;
+    copy_size = tile_buffers[row - offset][col].size;
+    size = 0;
+  }
+
+  *data += tile_size_bytes;
+
+  if (size > (size_t)(data_end - *data))
+    aom_internal_error(error_info, AOM_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt tile size");
+
+  if (size > 0) {
+    tile_buffers[row][col].data = *data;
+    tile_buffers[row][col].size = size;
+  } else {
+    tile_buffers[row][col].data = copy_data;
+    tile_buffers[row][col].size = copy_size;
+  }
+
+  *data += size;
+
+  tile_buffers[row][col].raw_data_end = *data;
+}
+
+static void get_tile_buffers(
+    AV1Decoder *pbi, const uint8_t *data, const uint8_t *data_end,
+    TileBufferDec (*const tile_buffers)[MAX_TILE_COLS]) {
+  AV1_COMMON *const cm = &pbi->common;
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  const int have_tiles = tile_cols * tile_rows > 1;
+
+  if (!have_tiles) {
+    const size_t tile_size = data_end - data;
+    tile_buffers[0][0].data = data;
+    tile_buffers[0][0].size = tile_size;
+    tile_buffers[0][0].raw_data_end = NULL;
+  } else {
+    // We locate only the tile buffers that are required, which are the ones
+    // specified by pbi->dec_tile_col and pbi->dec_tile_row. Also, we always
+    // need the last (bottom right) tile buffer, as we need to know where the
+    // end of the compressed frame buffer is for proper superframe decoding.
+
+    const uint8_t *tile_col_data_end[MAX_TILE_COLS];
+    const uint8_t *const data_start = data;
+
+    const int dec_tile_row = AOMMIN(pbi->dec_tile_row, tile_rows);
+    const int single_row = pbi->dec_tile_row >= 0;
+    const int tile_rows_start = single_row ? dec_tile_row : 0;
+    const int tile_rows_end = single_row ? tile_rows_start + 1 : tile_rows;
+    const int dec_tile_col = AOMMIN(pbi->dec_tile_col, tile_cols);
+    const int single_col = pbi->dec_tile_col >= 0;
+    const int tile_cols_start = single_col ? dec_tile_col : 0;
+    const int tile_cols_end = single_col ? tile_cols_start + 1 : tile_cols;
+
+    const int tile_col_size_bytes = pbi->tile_col_size_bytes;
+    const int tile_size_bytes = pbi->tile_size_bytes;
+
+    size_t tile_col_size;
+    int r, c;
+
+    // Read tile column sizes for all columns (we need the last tile buffer)
+    for (c = 0; c < tile_cols; ++c) {
+      const int is_last = c == tile_cols - 1;
+      if (!is_last) {
+        tile_col_size = mem_get_varsize(data, tile_col_size_bytes);
+        data += tile_col_size_bytes;
+        tile_col_data_end[c] = data + tile_col_size;
+      } else {
+        tile_col_size = data_end - data;
+        tile_col_data_end[c] = data_end;
+      }
+      data += tile_col_size;
+    }
+
+    data = data_start;
+
+    // Read the required tile sizes.
+    for (c = tile_cols_start; c < tile_cols_end; ++c) {
+      const int is_last = c == tile_cols - 1;
+
+      if (c > 0) data = tile_col_data_end[c - 1];
+
+      if (!is_last) data += tile_col_size_bytes;
+
+      // Get the whole of the last column, otherwise stop at the required tile.
+      for (r = 0; r < (is_last ? tile_rows : tile_rows_end); ++r) {
+        tile_buffers[r][c].col = c;
+
+        get_tile_buffer(tile_col_data_end[c], &pbi->common.error, &data,
+                        pbi->decrypt_cb, pbi->decrypt_state, tile_buffers,
+                        tile_size_bytes, c, r, cm->tile_encoding_mode);
+      }
+    }
+
+    // If we have not read the last column, then read it to get the last tile.
+    if (tile_cols_end != tile_cols) {
+      c = tile_cols - 1;
+
+      data = tile_col_data_end[c - 1];
+
+      for (r = 0; r < tile_rows; ++r) {
+        tile_buffers[r][c].col = c;
+
+        get_tile_buffer(tile_col_data_end[c], &pbi->common.error, &data,
+                        pbi->decrypt_cb, pbi->decrypt_state, tile_buffers,
+                        tile_size_bytes, c, r, cm->tile_encoding_mode);
+      }
+    }
+  }
+}
+#else
+// Reads the next tile returning its size and adjusting '*data' accordingly
+// based on 'is_last'.
+static void get_tile_buffer(const uint8_t *const data_end,
+                            const int tile_size_bytes, int is_last,
+                            struct aom_internal_error_info *error_info,
+                            const uint8_t **data, aom_decrypt_cb decrypt_cb,
+                            void *decrypt_state, TileBufferDec *const buf) {
+  size_t size;
+
+  if (!is_last) {
+    if (!read_is_valid(*data, tile_size_bytes, data_end))
+      aom_internal_error(error_info, AOM_CODEC_CORRUPT_FRAME,
+                         "Truncated packet or corrupt tile length");
+
+    if (decrypt_cb) {
+      uint8_t be_data[4];
+      decrypt_cb(decrypt_state, *data, be_data, tile_size_bytes);
+      size = mem_get_varsize(be_data, tile_size_bytes);
+    } else {
+      size = mem_get_varsize(*data, tile_size_bytes);
+    }
+    *data += tile_size_bytes;
+
+    if (size > (size_t)(data_end - *data))
+      aom_internal_error(error_info, AOM_CODEC_CORRUPT_FRAME,
+                         "Truncated packet or corrupt tile size");
+  } else {
+    size = data_end - *data;
+  }
+
+  buf->data = *data;
+  buf->size = size;
+
+  *data += size;
+}
+
+static void get_tile_buffers(
+    AV1Decoder *pbi, const uint8_t *data, const uint8_t *data_end,
+    TileBufferDec (*const tile_buffers)[MAX_TILE_COLS]) {
+  AV1_COMMON *const cm = &pbi->common;
+#if CONFIG_TILE_GROUPS
+  int r, c;
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  int tc = 0;
+  int first_tile_in_tg = 0;
+  struct aom_read_bit_buffer rb_tg_hdr;
+  uint8_t clear_data[MAX_AV1_HEADER_SIZE];
+  const int num_tiles = tile_rows * tile_cols;
+  const int num_bits = OD_ILOG(num_tiles) - 1;
+  const size_t hdr_size = pbi->uncomp_hdr_size + pbi->first_partition_size;
+  const int tg_size_bit_offset = pbi->tg_size_bit_offset;
+#if CONFIG_DEPENDENT_HORZTILES
+  int tile_group_start_col = 0;
+  int tile_group_start_row = 0;
+#endif
+
+  for (r = 0; r < tile_rows; ++r) {
+    for (c = 0; c < tile_cols; ++c, ++tc) {
+      TileBufferDec *const buf = &tile_buffers[r][c];
+      const int is_last = (r == tile_rows - 1) && (c == tile_cols - 1);
+      const size_t hdr_offset = (tc && tc == first_tile_in_tg) ? hdr_size : 0;
+
+      buf->col = c;
+      if (hdr_offset) {
+        init_read_bit_buffer(pbi, &rb_tg_hdr, data, data_end, clear_data);
+        rb_tg_hdr.bit_offset = tg_size_bit_offset;
+        if (num_tiles) {
+          pbi->tg_start = aom_rb_read_literal(&rb_tg_hdr, num_bits);
+          pbi->tg_size = 1 + aom_rb_read_literal(&rb_tg_hdr, num_bits);
+#if CONFIG_DEPENDENT_HORZTILES
+          tile_group_start_row = r;
+          tile_group_start_col = c;
+#endif
+        }
+      }
+      first_tile_in_tg += tc == first_tile_in_tg ? pbi->tg_size : 0;
+      data += hdr_offset;
+      get_tile_buffer(data_end, pbi->tile_size_bytes, is_last,
+                      &pbi->common.error, &data, pbi->decrypt_cb,
+                      pbi->decrypt_state, buf);
+#if CONFIG_DEPENDENT_HORZTILES
+      cm->tile_group_start_row[r][c] = tile_group_start_row;
+      cm->tile_group_start_col[r][c] = tile_group_start_col;
+#endif
+    }
+  }
+#else
+  int r, c;
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+
+  for (r = 0; r < tile_rows; ++r) {
+    for (c = 0; c < tile_cols; ++c) {
+      const int is_last = (r == tile_rows - 1) && (c == tile_cols - 1);
+      TileBufferDec *const buf = &tile_buffers[r][c];
+      buf->col = c;
+      get_tile_buffer(data_end, pbi->tile_size_bytes, is_last, &cm->error,
+                      &data, pbi->decrypt_cb, pbi->decrypt_state, buf);
+    }
+  }
+#endif
+}
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_PVQ
+static void daala_dec_init(AV1_COMMON *const cm, daala_dec_ctx *daala_dec,
+                           aom_reader *r) {
+  daala_dec->r = r;
+
+  // TODO(yushin) : activity masking info needs be signaled by a bitstream
+  daala_dec->use_activity_masking = AV1_PVQ_ENABLE_ACTIVITY_MASKING;
+
+#if !CONFIG_DAALA_DIST
+  daala_dec->use_activity_masking = 0;
+#endif
+
+  if (daala_dec->use_activity_masking)
+    daala_dec->qm = OD_HVS_QM;
+  else
+    daala_dec->qm = OD_FLAT_QM;
+
+  od_init_qm(daala_dec->state.qm, daala_dec->state.qm_inv,
+             daala_dec->qm == OD_HVS_QM ? OD_QM8_Q4_HVS : OD_QM8_Q4_FLAT);
+
+  if (daala_dec->use_activity_masking) {
+    int pli;
+    int use_masking = daala_dec->use_activity_masking;
+    int segment_id = 0;
+    int qindex = av1_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+
+    for (pli = 0; pli < MAX_MB_PLANE; pli++) {
+      int i;
+      int q;
+
+      q = qindex;
+      if (q <= OD_DEFAULT_QMS[use_masking][0][pli].interp_q << OD_COEFF_SHIFT) {
+        od_interp_qm(&daala_dec->state.pvq_qm_q4[pli][0], q,
+                     &OD_DEFAULT_QMS[use_masking][0][pli], NULL);
+      } else {
+        i = 0;
+        while (OD_DEFAULT_QMS[use_masking][i + 1][pli].qm_q4 != NULL &&
+               q > OD_DEFAULT_QMS[use_masking][i + 1][pli].interp_q
+                       << OD_COEFF_SHIFT) {
+          i++;
+        }
+        od_interp_qm(&daala_dec->state.pvq_qm_q4[pli][0], q,
+                     &OD_DEFAULT_QMS[use_masking][i][pli],
+                     &OD_DEFAULT_QMS[use_masking][i + 1][pli]);
+      }
+    }
+  }
+}
+#endif  // #if CONFIG_PVQ
+
+static const uint8_t *decode_tiles(AV1Decoder *pbi, const uint8_t *data,
+                                   const uint8_t *data_end) {
+  AV1_COMMON *const cm = &pbi->common;
+  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  const int n_tiles = tile_cols * tile_rows;
+  TileBufferDec(*const tile_buffers)[MAX_TILE_COLS] = pbi->tile_buffers;
+#if CONFIG_EXT_TILE
+  const int dec_tile_row = AOMMIN(pbi->dec_tile_row, tile_rows);
+  const int single_row = pbi->dec_tile_row >= 0;
+  const int tile_rows_start = single_row ? dec_tile_row : 0;
+  const int tile_rows_end = single_row ? dec_tile_row + 1 : tile_rows;
+  const int dec_tile_col = AOMMIN(pbi->dec_tile_col, tile_cols);
+  const int single_col = pbi->dec_tile_col >= 0;
+  const int tile_cols_start = single_col ? dec_tile_col : 0;
+  const int tile_cols_end = single_col ? tile_cols_start + 1 : tile_cols;
+  const int inv_col_order = pbi->inv_tile_order && !single_col;
+  const int inv_row_order = pbi->inv_tile_order && !single_row;
+#else
+  const int tile_rows_start = 0;
+  const int tile_rows_end = tile_rows;
+  const int tile_cols_start = 0;
+  const int tile_cols_end = tile_cols;
+  const int inv_col_order = pbi->inv_tile_order;
+  const int inv_row_order = pbi->inv_tile_order;
+#endif  // CONFIG_EXT_TILE
+  int tile_row, tile_col;
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  cm->do_subframe_update = n_tiles == 1;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+  if (cm->lf.filter_level && !cm->skip_loop_filter &&
+      pbi->lf_worker.data1 == NULL) {
+    CHECK_MEM_ERROR(cm, pbi->lf_worker.data1,
+                    aom_memalign(32, sizeof(LFWorkerData)));
+    pbi->lf_worker.hook = (AVxWorkerHook)av1_loop_filter_worker;
+    if (pbi->max_threads > 1 && !winterface->reset(&pbi->lf_worker)) {
+      aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                         "Loop filter thread creation failed");
+    }
+  }
+
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    LFWorkerData *const lf_data = (LFWorkerData *)pbi->lf_worker.data1;
+    // Be sure to sync as we might be resuming after a failed frame decode.
+    winterface->sync(&pbi->lf_worker);
+    av1_loop_filter_data_reset(lf_data, get_frame_new_buffer(cm), cm,
+                               pbi->mb.plane);
+  }
+
+  assert(tile_rows <= MAX_TILE_ROWS);
+  assert(tile_cols <= MAX_TILE_COLS);
+
+  get_tile_buffers(pbi, data, data_end, tile_buffers);
+
+  if (pbi->tile_data == NULL || n_tiles != pbi->allocated_tiles) {
+    aom_free(pbi->tile_data);
+    CHECK_MEM_ERROR(cm, pbi->tile_data,
+                    aom_memalign(32, n_tiles * (sizeof(*pbi->tile_data))));
+    pbi->allocated_tiles = n_tiles;
+  }
+#if CONFIG_ACCOUNTING
+  if (pbi->acct_enabled) {
+    aom_accounting_reset(&pbi->accounting);
+  }
+#endif
+  // Load all tile information into tile_data.
+  for (tile_row = tile_rows_start; tile_row < tile_rows_end; ++tile_row) {
+    for (tile_col = tile_cols_start; tile_col < tile_cols_end; ++tile_col) {
+      const TileBufferDec *const buf = &tile_buffers[tile_row][tile_col];
+      TileData *const td = pbi->tile_data + tile_cols * tile_row + tile_col;
+
+      td->cm = cm;
+      td->xd = pbi->mb;
+      td->xd.corrupted = 0;
+      td->xd.counts =
+          cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD
+              ? &cm->counts
+              : NULL;
+      av1_zero(td->dqcoeff);
+#if CONFIG_PVQ
+      av1_zero(td->pvq_ref_coeff);
+#endif
+      av1_tile_init(&td->xd.tile, td->cm, tile_row, tile_col);
+      setup_bool_decoder(buf->data, data_end, buf->size, &cm->error,
+                         &td->bit_reader,
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+                         1 << cm->ans_window_size_log2,
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+                         pbi->decrypt_cb, pbi->decrypt_state);
+#if CONFIG_ACCOUNTING
+      if (pbi->acct_enabled) {
+        td->bit_reader.accounting = &pbi->accounting;
+      } else {
+        td->bit_reader.accounting = NULL;
+      }
+#endif
+      av1_init_macroblockd(cm, &td->xd,
+#if CONFIG_PVQ
+                           td->pvq_ref_coeff,
+#endif
+#if CONFIG_CFL
+                           &td->cfl,
+#endif
+                           td->dqcoeff);
+
+#if CONFIG_EC_ADAPT
+      // Initialise the tile context from the frame context
+      td->tctx = *cm->fc;
+      td->xd.tile_ctx = &td->tctx;
+#endif
+
+#if CONFIG_PVQ
+      daala_dec_init(cm, &td->xd.daala_dec, &td->bit_reader);
+      td->xd.daala_dec.state.adapt = &td->tctx.pvq_context;
+#endif
+
+#if CONFIG_PALETTE
+      td->xd.plane[0].color_index_map = td->color_index_map[0];
+      td->xd.plane[1].color_index_map = td->color_index_map[1];
+#endif  // CONFIG_PALETTE
+    }
+  }
+
+  for (tile_row = tile_rows_start; tile_row < tile_rows_end; ++tile_row) {
+    const int row = inv_row_order ? tile_rows - 1 - tile_row : tile_row;
+    int mi_row = 0;
+    TileInfo tile_info;
+
+    av1_tile_set_row(&tile_info, cm, row);
+
+    for (tile_col = tile_cols_start; tile_col < tile_cols_end; ++tile_col) {
+      const int col = inv_col_order ? tile_cols - 1 - tile_col : tile_col;
+      TileData *const td = pbi->tile_data + tile_cols * row + col;
+#if CONFIG_ACCOUNTING
+      if (pbi->acct_enabled) {
+        td->bit_reader.accounting->last_tell_frac =
+            aom_reader_tell_frac(&td->bit_reader);
+      }
+#endif
+
+      av1_tile_set_col(&tile_info, cm, col);
+
+#if CONFIG_DEPENDENT_HORZTILES
+#if CONFIG_TILE_GROUPS
+      av1_tile_set_tg_boundary(&tile_info, cm, tile_row, tile_col);
+      if (!cm->dependent_horz_tiles || tile_row == 0 ||
+          tile_info.tg_horz_boundary) {
+#else
+      if (!cm->dependent_horz_tiles || tile_row == 0) {
+#endif
+        av1_zero_above_context(cm, tile_info.mi_col_start,
+                               tile_info.mi_col_end);
+      }
+#else
+      av1_zero_above_context(cm, tile_info.mi_col_start, tile_info.mi_col_end);
+#endif
+
+      for (mi_row = tile_info.mi_row_start; mi_row < tile_info.mi_row_end;
+           mi_row += cm->mib_size) {
+        int mi_col;
+
+        av1_zero_left_context(&td->xd);
+
+        for (mi_col = tile_info.mi_col_start; mi_col < tile_info.mi_col_end;
+             mi_col += cm->mib_size) {
+          av1_update_boundary_info(cm, &tile_info, mi_row, mi_col);
+          decode_partition(pbi, &td->xd,
+#if CONFIG_SUPERTX
+                           0,
+#endif  // CONFIG_SUPERTX
+                           mi_row, mi_col, &td->bit_reader, cm->sb_size,
+                           b_width_log2_lookup[cm->sb_size]);
+#if CONFIG_NCOBMC && CONFIG_MOTION_VAR
+          detoken_and_recon_sb(pbi, &td->xd, mi_row, mi_col, &td->bit_reader,
+                               cm->sb_size);
+#endif
+        }
+        aom_merge_corrupted_flag(&pbi->mb.corrupted, td->xd.corrupted);
+        if (pbi->mb.corrupted)
+          aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                             "Failed to decode tile data");
+#if CONFIG_SUBFRAME_PROB_UPDATE
+        if (cm->do_subframe_update &&
+            cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+          const int mi_rows_per_update =
+              MI_SIZE * AOMMAX(cm->mi_rows / MI_SIZE / COEF_PROBS_BUFS, 1);
+          if ((mi_row + MI_SIZE) % mi_rows_per_update == 0 &&
+              mi_row + MI_SIZE < cm->mi_rows &&
+              cm->coef_probs_update_idx < COEF_PROBS_BUFS - 1) {
+            av1_partial_adapt_probs(cm, mi_row, mi_col);
+            ++cm->coef_probs_update_idx;
+          }
+        }
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+      }
+    }
+
+    assert(mi_row > 0);
+
+// when Parallel deblocking is enabled, deblocking should not
+// be interleaved with decoding. Instead, deblocking should be done
+// after the entire frame is decoded.
+#if !CONFIG_VAR_TX && !CONFIG_PARALLEL_DEBLOCKING && !CONFIG_CB4X4
+    // Loopfilter one tile row.
+    // Note: If out-of-order tile decoding is used(for example, inv_row_order
+    // = 1), the loopfiltering has be done after all tile rows are decoded.
+    if (!inv_row_order && cm->lf.filter_level && !cm->skip_loop_filter) {
+      LFWorkerData *const lf_data = (LFWorkerData *)pbi->lf_worker.data1;
+      const int lf_start = AOMMAX(0, tile_info.mi_row_start - cm->mib_size);
+      const int lf_end = tile_info.mi_row_end - cm->mib_size;
+
+      // Delay the loopfilter if the first tile row is only
+      // a single superblock high.
+      if (lf_end <= 0) continue;
+
+      // Decoding has completed. Finish up the loop filter in this thread.
+      if (tile_info.mi_row_end >= cm->mi_rows) continue;
+
+      winterface->sync(&pbi->lf_worker);
+      lf_data->start = lf_start;
+      lf_data->stop = lf_end;
+      if (pbi->max_threads > 1) {
+        winterface->launch(&pbi->lf_worker);
+      } else {
+        winterface->execute(&pbi->lf_worker);
+      }
+    }
+#endif  // !CONFIG_VAR_TX && !CONFIG_PARALLEL_DEBLOCKING
+
+    // After loopfiltering, the last 7 row pixels in each superblock row may
+    // still be changed by the longest loopfilter of the next superblock row.
+    if (cm->frame_parallel_decode)
+      av1_frameworker_broadcast(pbi->cur_buf, mi_row << cm->mib_size_log2);
+  }
+
+#if CONFIG_VAR_TX || CONFIG_CB4X4
+  // Loopfilter the whole frame.
+  av1_loop_filter_frame(get_frame_new_buffer(cm), cm, &pbi->mb,
+                        cm->lf.filter_level, 0, 0);
+#else
+#if CONFIG_PARALLEL_DEBLOCKING
+  // Loopfilter all rows in the frame in the frame.
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    LFWorkerData *const lf_data = (LFWorkerData *)pbi->lf_worker.data1;
+    winterface->sync(&pbi->lf_worker);
+    lf_data->start = 0;
+    lf_data->stop = cm->mi_rows;
+    winterface->execute(&pbi->lf_worker);
+  }
+#else
+  // Loopfilter remaining rows in the frame.
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    LFWorkerData *const lf_data = (LFWorkerData *)pbi->lf_worker.data1;
+    winterface->sync(&pbi->lf_worker);
+    lf_data->start = lf_data->stop;
+    lf_data->stop = cm->mi_rows;
+    winterface->execute(&pbi->lf_worker);
+  }
+#endif  // CONFIG_PARALLEL_DEBLOCKING
+#endif  // CONFIG_VAR_TX
+  if (cm->frame_parallel_decode)
+    av1_frameworker_broadcast(pbi->cur_buf, INT_MAX);
+
+#if CONFIG_EXT_TILE
+  if (n_tiles == 1) {
+#if CONFIG_ANS
+    return data_end;
+#else
+    // Find the end of the single tile buffer
+    return aom_reader_find_end(&pbi->tile_data->bit_reader);
+#endif  // CONFIG_ANS
+  } else {
+    // Return the end of the last tile buffer
+    return tile_buffers[tile_rows - 1][tile_cols - 1].raw_data_end;
+  }
+#else
+#if CONFIG_ANS
+  return data_end;
+#else
+  {
+    // Get last tile data.
+    TileData *const td = pbi->tile_data + tile_cols * tile_rows - 1;
+    return aom_reader_find_end(&td->bit_reader);
+  }
+#endif  // CONFIG_ANS
+#endif  // CONFIG_EXT_TILE
+}
+
+static int tile_worker_hook(TileWorkerData *const tile_data,
+                            const TileInfo *const tile) {
+  AV1Decoder *const pbi = tile_data->pbi;
+  const AV1_COMMON *const cm = &pbi->common;
+  int mi_row, mi_col;
+
+  if (setjmp(tile_data->error_info.jmp)) {
+    tile_data->error_info.setjmp = 0;
+    aom_merge_corrupted_flag(&tile_data->xd.corrupted, 1);
+    return 0;
+  }
+
+  tile_data->error_info.setjmp = 1;
+  tile_data->xd.error_info = &tile_data->error_info;
+#if CONFIG_DEPENDENT_HORZTILES
+#if CONFIG_TILE_GROUPS
+  if (!cm->dependent_horz_tiles || tile->tg_horz_boundary) {
+#else
+  if (!cm->dependent_horz_tiles) {
+#endif
+    av1_zero_above_context(&pbi->common, tile->mi_col_start, tile->mi_col_end);
+  }
+#else
+  av1_zero_above_context(&pbi->common, tile->mi_col_start, tile->mi_col_end);
+#endif
+
+  for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
+       mi_row += cm->mib_size) {
+    av1_zero_left_context(&tile_data->xd);
+
+    for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
+         mi_col += cm->mib_size) {
+      decode_partition(pbi, &tile_data->xd,
+#if CONFIG_SUPERTX
+                       0,
+#endif
+                       mi_row, mi_col, &tile_data->bit_reader, cm->sb_size,
+                       b_width_log2_lookup[cm->sb_size]);
+#if CONFIG_NCOBMC && CONFIG_MOTION_VAR
+      detoken_and_recon_sb(pbi, &tile_data->xd, mi_row, mi_col,
+                           &tile_data->bit_reader, cm->sb_size);
+#endif
+    }
+  }
+  return !tile_data->xd.corrupted;
+}
+
+// sorts in descending order
+static int compare_tile_buffers(const void *a, const void *b) {
+  const TileBufferDec *const buf1 = (const TileBufferDec *)a;
+  const TileBufferDec *const buf2 = (const TileBufferDec *)b;
+  return (int)(buf2->size - buf1->size);
+}
+
+static const uint8_t *decode_tiles_mt(AV1Decoder *pbi, const uint8_t *data,
+                                      const uint8_t *data_end) {
+  AV1_COMMON *const cm = &pbi->common;
+  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  const int num_workers = AOMMIN(pbi->max_threads & ~1, tile_cols);
+  TileBufferDec(*const tile_buffers)[MAX_TILE_COLS] = pbi->tile_buffers;
+#if CONFIG_EXT_TILE
+  const int dec_tile_row = AOMMIN(pbi->dec_tile_row, tile_rows);
+  const int single_row = pbi->dec_tile_row >= 0;
+  const int tile_rows_start = single_row ? dec_tile_row : 0;
+  const int tile_rows_end = single_row ? dec_tile_row + 1 : tile_rows;
+  const int dec_tile_col = AOMMIN(pbi->dec_tile_col, tile_cols);
+  const int single_col = pbi->dec_tile_col >= 0;
+  const int tile_cols_start = single_col ? dec_tile_col : 0;
+  const int tile_cols_end = single_col ? tile_cols_start + 1 : tile_cols;
+#else
+  const int tile_rows_start = 0;
+  const int tile_rows_end = tile_rows;
+  const int tile_cols_start = 0;
+  const int tile_cols_end = tile_cols;
+#endif  // CONFIG_EXT_TILE
+  int tile_row, tile_col;
+  int i;
+
+#if !(CONFIG_ANS || CONFIG_EXT_TILE)
+  int final_worker = -1;
+#endif  // !(CONFIG_ANS || CONFIG_EXT_TILE)
+
+  assert(tile_rows <= MAX_TILE_ROWS);
+  assert(tile_cols <= MAX_TILE_COLS);
+
+  assert(tile_cols * tile_rows > 1);
+
+  // TODO(jzern): See if we can remove the restriction of passing in max
+  // threads to the decoder.
+  if (pbi->num_tile_workers == 0) {
+    const int num_threads = pbi->max_threads & ~1;
+    CHECK_MEM_ERROR(cm, pbi->tile_workers,
+                    aom_malloc(num_threads * sizeof(*pbi->tile_workers)));
+    // Ensure tile data offsets will be properly aligned. This may fail on
+    // platforms without DECLARE_ALIGNED().
+    assert((sizeof(*pbi->tile_worker_data) % 16) == 0);
+    CHECK_MEM_ERROR(
+        cm, pbi->tile_worker_data,
+        aom_memalign(32, num_threads * sizeof(*pbi->tile_worker_data)));
+    CHECK_MEM_ERROR(cm, pbi->tile_worker_info,
+                    aom_malloc(num_threads * sizeof(*pbi->tile_worker_info)));
+    for (i = 0; i < num_threads; ++i) {
+      AVxWorker *const worker = &pbi->tile_workers[i];
+      ++pbi->num_tile_workers;
+
+      winterface->init(worker);
+      if (i < num_threads - 1 && !winterface->reset(worker)) {
+        aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                           "Tile decoder thread creation failed");
+      }
+    }
+  }
+
+  // Reset tile decoding hook
+  for (i = 0; i < num_workers; ++i) {
+    AVxWorker *const worker = &pbi->tile_workers[i];
+    winterface->sync(worker);
+    worker->hook = (AVxWorkerHook)tile_worker_hook;
+    worker->data1 = &pbi->tile_worker_data[i];
+    worker->data2 = &pbi->tile_worker_info[i];
+  }
+
+  // Initialize thread frame counts.
+  if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+    for (i = 0; i < num_workers; ++i) {
+      TileWorkerData *const twd = (TileWorkerData *)pbi->tile_workers[i].data1;
+      av1_zero(twd->counts);
+    }
+  }
+
+  // Load tile data into tile_buffers
+  get_tile_buffers(pbi, data, data_end, tile_buffers);
+
+  for (tile_row = tile_rows_start; tile_row < tile_rows_end; ++tile_row) {
+    // Sort the buffers in this tile row based on size in descending order.
+    qsort(&tile_buffers[tile_row][tile_cols_start],
+          tile_cols_end - tile_cols_start, sizeof(tile_buffers[0][0]),
+          compare_tile_buffers);
+
+    // Rearrange the tile buffers in this tile row such that per-tile group
+    // the largest, and presumably the most difficult tile will be decoded in
+    // the main thread. This should help minimize the number of instances
+    // where the main thread is waiting for a worker to complete.
+    {
+      int group_start;
+      for (group_start = tile_cols_start; group_start < tile_cols_end;
+           group_start += num_workers) {
+        const int group_end = AOMMIN(group_start + num_workers, tile_cols);
+        const TileBufferDec largest = tile_buffers[tile_row][group_start];
+        memmove(&tile_buffers[tile_row][group_start],
+                &tile_buffers[tile_row][group_start + 1],
+                (group_end - group_start - 1) * sizeof(tile_buffers[0][0]));
+        tile_buffers[tile_row][group_end - 1] = largest;
+      }
+    }
+
+    for (tile_col = tile_cols_start; tile_col < tile_cols_end;) {
+      // Launch workers for individual columns
+      for (i = 0; i < num_workers && tile_col < tile_cols_end;
+           ++i, ++tile_col) {
+        TileBufferDec *const buf = &tile_buffers[tile_row][tile_col];
+        AVxWorker *const worker = &pbi->tile_workers[i];
+        TileWorkerData *const twd = (TileWorkerData *)worker->data1;
+        TileInfo *const tile_info = (TileInfo *)worker->data2;
+
+        twd->pbi = pbi;
+        twd->xd = pbi->mb;
+        twd->xd.corrupted = 0;
+        twd->xd.counts =
+            cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD
+                ? &twd->counts
+                : NULL;
+        av1_zero(twd->dqcoeff);
+        av1_tile_init(tile_info, cm, tile_row, buf->col);
+        av1_tile_init(&twd->xd.tile, cm, tile_row, buf->col);
+        setup_bool_decoder(buf->data, data_end, buf->size, &cm->error,
+                           &twd->bit_reader,
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+                           1 << cm->ans_window_size_log2,
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+                           pbi->decrypt_cb, pbi->decrypt_state);
+        av1_init_macroblockd(cm, &twd->xd,
+#if CONFIG_PVQ
+                             twd->pvq_ref_coeff,
+#endif
+#if CONFIG_CFL
+                             &twd->cfl,
+#endif
+                             twd->dqcoeff);
+#if CONFIG_PVQ
+        daala_dec_init(cm, &twd->xd.daala_dec, &twd->bit_reader);
+        twd->xd.daala_dec.state.adapt = &twd->tctx.pvq_context;
+#endif
+#if CONFIG_EC_ADAPT
+        // Initialise the tile context from the frame context
+        twd->tctx = *cm->fc;
+        twd->xd.tile_ctx = &twd->tctx;
+#endif
+#if CONFIG_PALETTE
+        twd->xd.plane[0].color_index_map = twd->color_index_map[0];
+        twd->xd.plane[1].color_index_map = twd->color_index_map[1];
+#endif  // CONFIG_PALETTE
+
+        worker->had_error = 0;
+        if (i == num_workers - 1 || tile_col == tile_cols_end - 1) {
+          winterface->execute(worker);
+        } else {
+          winterface->launch(worker);
+        }
+
+#if !(CONFIG_ANS || CONFIG_EXT_TILE)
+        if (tile_row == tile_rows - 1 && buf->col == tile_cols - 1) {
+          final_worker = i;
+        }
+#endif  // !(CONFIG_ANS || CONFIG_EXT_TILE)
+      }
+
+      // Sync all workers
+      for (; i > 0; --i) {
+        AVxWorker *const worker = &pbi->tile_workers[i - 1];
+        // TODO(jzern): The tile may have specific error data associated with
+        // its aom_internal_error_info which could be propagated to the main
+        // info in cm. Additionally once the threads have been synced and an
+        // error is detected, there's no point in continuing to decode tiles.
+        pbi->mb.corrupted |= !winterface->sync(worker);
+      }
+    }
+  }
+
+  // Accumulate thread frame counts.
+  if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+    for (i = 0; i < num_workers; ++i) {
+      TileWorkerData *const twd = (TileWorkerData *)pbi->tile_workers[i].data1;
+      av1_accumulate_frame_counts(&cm->counts, &twd->counts);
+    }
+  }
+
+#if CONFIG_EXT_TILE
+  // Return the end of the last tile buffer
+  return tile_buffers[tile_rows - 1][tile_cols - 1].raw_data_end;
+#else
+#if CONFIG_ANS
+  return data_end;
+#else
+  assert(final_worker != -1);
+  {
+    TileWorkerData *const twd =
+        (TileWorkerData *)pbi->tile_workers[final_worker].data1;
+    return aom_reader_find_end(&twd->bit_reader);
+  }
+#endif  // CONFIG_ANS
+#endif  // CONFIG_EXT_TILE
+}
+
+static void error_handler(void *data) {
+  AV1_COMMON *const cm = (AV1_COMMON *)data;
+  aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME, "Truncated packet");
+}
+
+static void read_bitdepth_colorspace_sampling(AV1_COMMON *cm,
+                                              struct aom_read_bit_buffer *rb) {
+  if (cm->profile >= PROFILE_2) {
+    cm->bit_depth = aom_rb_read_bit(rb) ? AOM_BITS_12 : AOM_BITS_10;
+  } else {
+    cm->bit_depth = AOM_BITS_8;
+  }
+
+#if CONFIG_HIGHBITDEPTH
+  if (cm->bit_depth > AOM_BITS_8) {
+    cm->use_highbitdepth = 1;
+  } else {
+#if CONFIG_LOWBITDEPTH
+    cm->use_highbitdepth = 0;
+#else
+    cm->use_highbitdepth = 1;
+#endif
+  }
+#endif
+
+  cm->color_space = aom_rb_read_literal(rb, 3);
+  if (cm->color_space != AOM_CS_SRGB) {
+    // [16,235] (including xvycc) vs [0,255] range
+    cm->color_range = aom_rb_read_bit(rb);
+    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
+      cm->subsampling_x = aom_rb_read_bit(rb);
+      cm->subsampling_y = aom_rb_read_bit(rb);
+      if (cm->subsampling_x == 1 && cm->subsampling_y == 1)
+        aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                           "4:2:0 color not supported in profile 1 or 3");
+      if (aom_rb_read_bit(rb))
+        aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                           "Reserved bit set");
+    } else {
+      cm->subsampling_y = cm->subsampling_x = 1;
+    }
+  } else {
+    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
+      // Note if colorspace is SRGB then 4:4:4 chroma sampling is assumed.
+      // 4:2:2 or 4:4:0 chroma sampling is not allowed.
+      cm->subsampling_y = cm->subsampling_x = 0;
+      if (aom_rb_read_bit(rb))
+        aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                           "Reserved bit set");
+    } else {
+      aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                         "4:4:4 color not supported in profile 0 or 2");
+    }
+  }
+}
+
+#if CONFIG_REFERENCE_BUFFER
+void read_sequence_header(SequenceHeader *seq_params) {
+  /* Placeholder for actually reading from the bitstream */
+  seq_params->frame_id_numbers_present_flag = FRAME_ID_NUMBERS_PRESENT_FLAG;
+  seq_params->frame_id_length_minus7 = FRAME_ID_LENGTH_MINUS7;
+  seq_params->delta_frame_id_length_minus2 = DELTA_FRAME_ID_LENGTH_MINUS2;
+}
+#endif
+
+static size_t read_uncompressed_header(AV1Decoder *pbi,
+                                       struct aom_read_bit_buffer *rb) {
+  AV1_COMMON *const cm = &pbi->common;
+  MACROBLOCKD *const xd = &pbi->mb;
+  BufferPool *const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = pool->frame_bufs;
+  int i, mask, ref_index = 0;
+  size_t sz;
+
+#if CONFIG_REFERENCE_BUFFER
+  /* TODO: Move outside frame loop or inside key-frame branch */
+  read_sequence_header(&pbi->seq_params);
+#endif
+
+  cm->last_frame_type = cm->frame_type;
+  cm->last_intra_only = cm->intra_only;
+
+#if CONFIG_EXT_REFS
+  // NOTE: By default all coded frames to be used as a reference
+  cm->is_reference_frame = 1;
+#endif  // CONFIG_EXT_REFS
+
+  if (aom_rb_read_literal(rb, 2) != AOM_FRAME_MARKER)
+    aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                       "Invalid frame marker");
+
+  cm->profile = av1_read_profile(rb);
+
+  const BITSTREAM_PROFILE MAX_SUPPORTED_PROFILE =
+      CONFIG_HIGHBITDEPTH ? MAX_PROFILES : PROFILE_2;
+
+  if (cm->profile >= MAX_SUPPORTED_PROFILE)
+    aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                       "Unsupported bitstream profile");
+
+  cm->show_existing_frame = aom_rb_read_bit(rb);
+
+  if (cm->show_existing_frame) {
+    // Show an existing frame directly.
+    const int existing_frame_idx = aom_rb_read_literal(rb, 3);
+    const int frame_to_show = cm->ref_frame_map[existing_frame_idx];
+#if CONFIG_REFERENCE_BUFFER
+    if (pbi->seq_params.frame_id_numbers_present_flag) {
+      int frame_id_length = pbi->seq_params.frame_id_length_minus7 + 7;
+      int display_frame_id = aom_rb_read_literal(rb, frame_id_length);
+      /* Compare display_frame_id with ref_frame_id and check valid for
+      * referencing */
+      if (display_frame_id != cm->ref_frame_id[existing_frame_idx] ||
+          cm->valid_for_referencing[existing_frame_idx] == 0)
+        aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                           "Reference buffer frame ID mismatch");
+    }
+#endif
+    lock_buffer_pool(pool);
+    if (frame_to_show < 0 || frame_bufs[frame_to_show].ref_count < 1) {
+      unlock_buffer_pool(pool);
+      aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                         "Buffer %d does not contain a decoded frame",
+                         frame_to_show);
+    }
+    ref_cnt_fb(frame_bufs, &cm->new_fb_idx, frame_to_show);
+    unlock_buffer_pool(pool);
+
+    cm->lf.filter_level = 0;
+    cm->show_frame = 1;
+    pbi->refresh_frame_flags = 0;
+
+    if (cm->frame_parallel_decode) {
+      for (i = 0; i < REF_FRAMES; ++i)
+        cm->next_ref_frame_map[i] = cm->ref_frame_map[i];
+    }
+
+    return 0;
+  }
+
+  cm->frame_type = (FRAME_TYPE)aom_rb_read_bit(rb);
+  cm->show_frame = aom_rb_read_bit(rb);
+  cm->error_resilient_mode = aom_rb_read_bit(rb);
+#if CONFIG_REFERENCE_BUFFER
+  if (pbi->seq_params.frame_id_numbers_present_flag) {
+    int frame_id_length = pbi->seq_params.frame_id_length_minus7 + 7;
+    int diff_len = pbi->seq_params.delta_frame_id_length_minus2 + 2;
+    int prev_frame_id = 0;
+    if (cm->frame_type != KEY_FRAME) {
+      prev_frame_id = cm->current_frame_id;
+    }
+    cm->current_frame_id = aom_rb_read_literal(rb, frame_id_length);
+
+    if (cm->frame_type != KEY_FRAME) {
+      int diff_frame_id;
+      if (cm->current_frame_id > prev_frame_id) {
+        diff_frame_id = cm->current_frame_id - prev_frame_id;
+      } else {
+        diff_frame_id =
+            (1 << frame_id_length) + cm->current_frame_id - prev_frame_id;
+      }
+      /* Check current_frame_id for conformance */
+      if (prev_frame_id == cm->current_frame_id ||
+          diff_frame_id >= (1 << (frame_id_length - 1))) {
+        aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                           "Invalid value of current_frame_id");
+      }
+    }
+    /* Check if some frames need to be marked as not valid for referencing */
+    for (i = 0; i < REF_FRAMES; i++) {
+      if (cm->frame_type == KEY_FRAME) {
+        cm->valid_for_referencing[i] = 0;
+      } else if (cm->current_frame_id - (1 << diff_len) > 0) {
+        if (cm->ref_frame_id[i] > cm->current_frame_id ||
+            cm->ref_frame_id[i] < cm->current_frame_id - (1 << diff_len))
+          cm->valid_for_referencing[i] = 0;
+      } else {
+        if (cm->ref_frame_id[i] > cm->current_frame_id &&
+            cm->ref_frame_id[i] <
+                (1 << frame_id_length) + cm->current_frame_id - (1 << diff_len))
+          cm->valid_for_referencing[i] = 0;
+      }
+    }
+  }
+#endif
+  if (cm->frame_type == KEY_FRAME) {
+    if (!av1_read_sync_code(rb))
+      aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                         "Invalid frame sync code");
+
+    read_bitdepth_colorspace_sampling(cm, rb);
+    pbi->refresh_frame_flags = (1 << REF_FRAMES) - 1;
+
+    for (i = 0; i < INTER_REFS_PER_FRAME; ++i) {
+      cm->frame_refs[i].idx = INVALID_IDX;
+      cm->frame_refs[i].buf = NULL;
+    }
+
+    setup_frame_size(cm, rb);
+    if (pbi->need_resync) {
+      memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+      pbi->need_resync = 0;
+    }
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+    cm->ans_window_size_log2 = aom_rb_read_literal(rb, 4) + 8;
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+#if CONFIG_PALETTE
+    cm->allow_screen_content_tools = aom_rb_read_bit(rb);
+#endif  // CONFIG_PALETTE
+  } else {
+    cm->intra_only = cm->show_frame ? 0 : aom_rb_read_bit(rb);
+#if CONFIG_PALETTE
+    if (cm->intra_only) cm->allow_screen_content_tools = aom_rb_read_bit(rb);
+#endif  // CONFIG_PALETTE
+    if (cm->error_resilient_mode) {
+      cm->reset_frame_context = RESET_FRAME_CONTEXT_ALL;
+    } else {
+      if (cm->intra_only) {
+        cm->reset_frame_context = aom_rb_read_bit(rb)
+                                      ? RESET_FRAME_CONTEXT_ALL
+                                      : RESET_FRAME_CONTEXT_CURRENT;
+      } else {
+        cm->reset_frame_context = aom_rb_read_bit(rb)
+                                      ? RESET_FRAME_CONTEXT_CURRENT
+                                      : RESET_FRAME_CONTEXT_NONE;
+        if (cm->reset_frame_context == RESET_FRAME_CONTEXT_CURRENT)
+          cm->reset_frame_context = aom_rb_read_bit(rb)
+                                        ? RESET_FRAME_CONTEXT_ALL
+                                        : RESET_FRAME_CONTEXT_CURRENT;
+      }
+    }
+
+    if (cm->intra_only) {
+      if (!av1_read_sync_code(rb))
+        aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                           "Invalid frame sync code");
+
+      read_bitdepth_colorspace_sampling(cm, rb);
+
+      pbi->refresh_frame_flags = aom_rb_read_literal(rb, REF_FRAMES);
+      setup_frame_size(cm, rb);
+      if (pbi->need_resync) {
+        memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+        pbi->need_resync = 0;
+      }
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+      cm->ans_window_size_log2 = aom_rb_read_literal(rb, 4) + 8;
+#endif
+    } else if (pbi->need_resync != 1) { /* Skip if need resync */
+      pbi->refresh_frame_flags = aom_rb_read_literal(rb, REF_FRAMES);
+
+#if CONFIG_EXT_REFS
+      if (!pbi->refresh_frame_flags) {
+        // NOTE: "pbi->refresh_frame_flags == 0" indicates that the coded frame
+        //       will not be used as a reference
+        cm->is_reference_frame = 0;
+      }
+#endif  // CONFIG_EXT_REFS
+
+      for (i = 0; i < INTER_REFS_PER_FRAME; ++i) {
+        const int ref = aom_rb_read_literal(rb, REF_FRAMES_LOG2);
+        const int idx = cm->ref_frame_map[ref];
+        RefBuffer *const ref_frame = &cm->frame_refs[i];
+        ref_frame->idx = idx;
+        ref_frame->buf = &frame_bufs[idx].buf;
+        cm->ref_frame_sign_bias[LAST_FRAME + i] = aom_rb_read_bit(rb);
+#if CONFIG_REFERENCE_BUFFER
+        if (pbi->seq_params.frame_id_numbers_present_flag) {
+          int frame_id_length = pbi->seq_params.frame_id_length_minus7 + 7;
+          int diff_len = pbi->seq_params.delta_frame_id_length_minus2 + 2;
+          int delta_frame_id_minus1 = aom_rb_read_literal(rb, diff_len);
+          int ref_frame_id =
+              ((cm->current_frame_id - (delta_frame_id_minus1 + 1) +
+                (1 << frame_id_length)) %
+               (1 << frame_id_length));
+          /* Compare values derived from delta_frame_id_minus1 and
+          * refresh_frame_flags. Also, check valid for referencing */
+          if (ref_frame_id != cm->ref_frame_id[ref] ||
+              cm->valid_for_referencing[ref] == 0)
+            aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                               "Reference buffer frame ID mismatch");
+        }
+#endif
+      }
+
+#if CONFIG_FRAME_SIZE
+      if (cm->error_resilient_mode == 0) {
+        setup_frame_size_with_refs(cm, rb);
+      } else {
+        setup_frame_size(cm, rb);
+      }
+#else
+      setup_frame_size_with_refs(cm, rb);
+#endif
+
+      cm->allow_high_precision_mv = aom_rb_read_bit(rb);
+      cm->interp_filter = read_frame_interp_filter(rb);
+#if CONFIG_TEMPMV_SIGNALING
+      if (!cm->error_resilient_mode) {
+        cm->use_prev_frame_mvs = aom_rb_read_bit(rb);
+      }
+#endif
+      for (i = 0; i < INTER_REFS_PER_FRAME; ++i) {
+        RefBuffer *const ref_buf = &cm->frame_refs[i];
+#if CONFIG_HIGHBITDEPTH
+        av1_setup_scale_factors_for_frame(
+            &ref_buf->sf, ref_buf->buf->y_crop_width,
+            ref_buf->buf->y_crop_height, cm->width, cm->height,
+            cm->use_highbitdepth);
+#else
+        av1_setup_scale_factors_for_frame(
+            &ref_buf->sf, ref_buf->buf->y_crop_width,
+            ref_buf->buf->y_crop_height, cm->width, cm->height);
+#endif
+      }
+    }
+  }
+#if CONFIG_TEMPMV_SIGNALING
+  cm->cur_frame->intra_only = cm->frame_type == KEY_FRAME || cm->intra_only;
+#endif
+
+#if CONFIG_REFERENCE_BUFFER
+  if (pbi->seq_params.frame_id_numbers_present_flag) {
+    /* If bitmask is set, update reference frame id values and
+    mark frames as valid for reference */
+    int refresh_frame_flags =
+        cm->frame_type == KEY_FRAME ? 0xFF : pbi->refresh_frame_flags;
+    for (i = 0; i < REF_FRAMES; i++) {
+      if ((refresh_frame_flags >> i) & 1) {
+        cm->ref_frame_id[i] = cm->current_frame_id;
+        cm->valid_for_referencing[i] = 1;
+      }
+    }
+  }
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+  get_frame_new_buffer(cm)->bit_depth = cm->bit_depth;
+#endif
+  get_frame_new_buffer(cm)->color_space = cm->color_space;
+  get_frame_new_buffer(cm)->color_range = cm->color_range;
+  get_frame_new_buffer(cm)->render_width = cm->render_width;
+  get_frame_new_buffer(cm)->render_height = cm->render_height;
+
+  if (pbi->need_resync) {
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Keyframe / intra-only frame required to reset decoder"
+                       " state");
+  }
+
+  if (!cm->error_resilient_mode) {
+    cm->refresh_frame_context = aom_rb_read_bit(rb)
+                                    ? REFRESH_FRAME_CONTEXT_FORWARD
+                                    : REFRESH_FRAME_CONTEXT_BACKWARD;
+  } else {
+    cm->refresh_frame_context = REFRESH_FRAME_CONTEXT_FORWARD;
+  }
+
+  // This flag will be overridden by the call to av1_setup_past_independence
+  // below, forcing the use of context 0 for those frame types.
+  cm->frame_context_idx = aom_rb_read_literal(rb, FRAME_CONTEXTS_LOG2);
+
+  // Generate next_ref_frame_map.
+  lock_buffer_pool(pool);
+  for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+    if (mask & 1) {
+      cm->next_ref_frame_map[ref_index] = cm->new_fb_idx;
+      ++frame_bufs[cm->new_fb_idx].ref_count;
+    } else {
+      cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+    }
+    // Current thread holds the reference frame.
+    if (cm->ref_frame_map[ref_index] >= 0)
+      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+    ++ref_index;
+  }
+
+  for (; ref_index < REF_FRAMES; ++ref_index) {
+    cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+
+    // Current thread holds the reference frame.
+    if (cm->ref_frame_map[ref_index] >= 0)
+      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+  }
+  unlock_buffer_pool(pool);
+  pbi->hold_ref_buf = 1;
+
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode)
+    av1_setup_past_independence(cm);
+
+#if CONFIG_EXT_PARTITION
+  set_sb_size(cm, aom_rb_read_bit(rb) ? BLOCK_128X128 : BLOCK_64X64);
+#else
+  set_sb_size(cm, BLOCK_64X64);
+#endif  // CONFIG_EXT_PARTITION
+
+  setup_loopfilter(cm, rb);
+#if CONFIG_CDEF
+  setup_cdef(cm, rb);
+#endif
+#if CONFIG_LOOP_RESTORATION
+  decode_restoration_mode(cm, rb);
+#endif  // CONFIG_LOOP_RESTORATION
+  setup_quantization(cm, rb);
+#if CONFIG_HIGHBITDEPTH
+  xd->bd = (int)cm->bit_depth;
+#endif
+
+#if CONFIG_Q_ADAPT_PROBS
+  av1_default_coef_probs(cm);
+  if (cm->frame_type == KEY_FRAME || cm->error_resilient_mode ||
+      cm->reset_frame_context == RESET_FRAME_CONTEXT_ALL) {
+    for (i = 0; i < FRAME_CONTEXTS; ++i) cm->frame_contexts[i] = *cm->fc;
+  } else if (cm->reset_frame_context == RESET_FRAME_CONTEXT_CURRENT) {
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+  }
+#endif  // CONFIG_Q_ADAPT_PROBS
+
+  setup_segmentation(cm, rb);
+
+#if CONFIG_DELTA_Q
+  {
+    struct segmentation *const seg = &cm->seg;
+    int segment_quantizer_active = 0;
+    for (i = 0; i < MAX_SEGMENTS; i++) {
+      if (segfeature_active(seg, i, SEG_LVL_ALT_Q)) {
+        segment_quantizer_active = 1;
+      }
+    }
+
+    cm->delta_q_res = 1;
+#if CONFIG_EXT_DELTA_Q
+    cm->delta_lf_res = 1;
+#endif
+    if (segment_quantizer_active == 0 && cm->base_qindex > 0) {
+      cm->delta_q_present_flag = aom_rb_read_bit(rb);
+    } else {
+      cm->delta_q_present_flag = 0;
+    }
+    if (cm->delta_q_present_flag) {
+      xd->prev_qindex = cm->base_qindex;
+      cm->delta_q_res = 1 << aom_rb_read_literal(rb, 2);
+#if CONFIG_EXT_DELTA_Q
+      if (segment_quantizer_active) {
+        assert(seg->abs_delta == SEGMENT_DELTADATA);
+      }
+      cm->delta_lf_present_flag = aom_rb_read_bit(rb);
+      if (cm->delta_lf_present_flag) {
+        xd->prev_delta_lf_from_base = 0;
+        cm->delta_lf_res = 1 << aom_rb_read_literal(rb, 2);
+      } else {
+        cm->delta_lf_present_flag = 0;
+      }
+#endif  // CONFIG_EXT_DELTA_Q
+    }
+  }
+#endif
+
+  for (i = 0; i < MAX_SEGMENTS; ++i) {
+    const int qindex = cm->seg.enabled
+                           ? av1_get_qindex(&cm->seg, i, cm->base_qindex)
+                           : cm->base_qindex;
+    xd->lossless[i] = qindex == 0 && cm->y_dc_delta_q == 0 &&
+                      cm->uv_dc_delta_q == 0 && cm->uv_ac_delta_q == 0;
+    xd->qindex[i] = qindex;
+  }
+
+  setup_segmentation_dequant(cm);
+  cm->tx_mode = read_tx_mode(cm, xd, rb);
+  cm->reference_mode = read_frame_reference_mode(cm, rb);
+
+#if CONFIG_EXT_TX
+  cm->reduced_tx_set_used = aom_rb_read_bit(rb);
+#endif  // CONFIG_EXT_TX
+
+  read_tile_info(pbi, rb);
+  sz = aom_rb_read_literal(rb, 16);
+
+  if (sz == 0)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Invalid header size");
+  return sz;
+}
+
+#if CONFIG_EXT_TX
+#if !CONFIG_EC_ADAPT
+static void read_ext_tx_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j, k;
+  int s;
+  for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
+    if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+      for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+        if (!use_inter_ext_tx_for_txsize[s][i]) continue;
+        for (j = 0; j < num_ext_tx_set[ext_tx_set_type_inter[s]] - 1; ++j)
+          av1_diff_update_prob(r, &fc->inter_ext_tx_prob[s][i][j], ACCT_STR);
+      }
+    }
+  }
+
+  for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
+    if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+      for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+        if (!use_intra_ext_tx_for_txsize[s][i]) continue;
+        for (j = 0; j < INTRA_MODES; ++j)
+          for (k = 0; k < num_ext_tx_set[ext_tx_set_type_intra[s]] - 1; ++k)
+            av1_diff_update_prob(r, &fc->intra_ext_tx_prob[s][i][j][k],
+                                 ACCT_STR);
+      }
+    }
+  }
+}
+#endif  // !CONFIG_EC_ADAPT
+#else
+
+#endif  // CONFIG_EXT_TX
+#if CONFIG_SUPERTX
+static void read_supertx_probs(FRAME_CONTEXT *fc, aom_reader *r) {
+  int i, j;
+  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
+    for (i = 0; i < PARTITION_SUPERTX_CONTEXTS; ++i) {
+      for (j = TX_8X8; j < TX_SIZES; ++j) {
+        av1_diff_update_prob(r, &fc->supertx_prob[i][j], ACCT_STR);
+      }
+    }
+  }
+}
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_GLOBAL_MOTION
+static void read_global_motion_params(WarpedMotionParams *params,
+                                      WarpedMotionParams *ref_params,
+                                      aom_prob *probs, aom_reader *r,
+                                      int allow_hp) {
+  TransformationType type =
+      aom_read_tree(r, av1_global_motion_types_tree, probs, ACCT_STR);
+  int trans_bits;
+  int trans_dec_factor;
+  int trans_prec_diff;
+  set_default_warp_params(params);
+  params->wmtype = type;
+  switch (type) {
+    case HOMOGRAPHY:
+    case HORTRAPEZOID:
+    case VERTRAPEZOID:
+      if (type != HORTRAPEZOID)
+        params->wmmat[6] =
+            aom_read_signed_primitive_refsubexpfin(
+                r, GM_ROW3HOMO_MAX + 1, SUBEXPFIN_K,
+                (ref_params->wmmat[6] >> GM_ROW3HOMO_PREC_DIFF)) *
+            GM_ROW3HOMO_DECODE_FACTOR;
+      if (type != VERTRAPEZOID)
+        params->wmmat[7] =
+            aom_read_signed_primitive_refsubexpfin(
+                r, GM_ROW3HOMO_MAX + 1, SUBEXPFIN_K,
+                (ref_params->wmmat[7] >> GM_ROW3HOMO_PREC_DIFF)) *
+            GM_ROW3HOMO_DECODE_FACTOR;
+    case AFFINE:
+    case ROTZOOM:
+      params->wmmat[2] = aom_read_signed_primitive_refsubexpfin(
+                             r, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+                             (ref_params->wmmat[2] >> GM_ALPHA_PREC_DIFF) -
+                                 (1 << GM_ALPHA_PREC_BITS)) *
+                             GM_ALPHA_DECODE_FACTOR +
+                         (1 << WARPEDMODEL_PREC_BITS);
+      if (type != VERTRAPEZOID)
+        params->wmmat[3] = aom_read_signed_primitive_refsubexpfin(
+                               r, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+                               (ref_params->wmmat[3] >> GM_ALPHA_PREC_DIFF)) *
+                           GM_ALPHA_DECODE_FACTOR;
+      if (type >= AFFINE) {
+        if (type != HORTRAPEZOID)
+          params->wmmat[4] = aom_read_signed_primitive_refsubexpfin(
+                                 r, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+                                 (ref_params->wmmat[4] >> GM_ALPHA_PREC_DIFF)) *
+                             GM_ALPHA_DECODE_FACTOR;
+        params->wmmat[5] = aom_read_signed_primitive_refsubexpfin(
+                               r, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+                               (ref_params->wmmat[5] >> GM_ALPHA_PREC_DIFF) -
+                                   (1 << GM_ALPHA_PREC_BITS)) *
+                               GM_ALPHA_DECODE_FACTOR +
+                           (1 << WARPEDMODEL_PREC_BITS);
+      } else {
+        params->wmmat[4] = -params->wmmat[3];
+        params->wmmat[5] = params->wmmat[2];
+      }
+    // fallthrough intended
+    case TRANSLATION:
+      trans_bits = (type == TRANSLATION) ? GM_ABS_TRANS_ONLY_BITS - !allow_hp
+                                         : GM_ABS_TRANS_BITS;
+      trans_dec_factor = (type == TRANSLATION)
+                             ? GM_TRANS_ONLY_DECODE_FACTOR * (1 << !allow_hp)
+                             : GM_TRANS_DECODE_FACTOR;
+      trans_prec_diff = (type == TRANSLATION)
+                            ? GM_TRANS_ONLY_PREC_DIFF + !allow_hp
+                            : GM_TRANS_PREC_DIFF;
+      params->wmmat[0] = aom_read_signed_primitive_refsubexpfin(
+                             r, (1 << trans_bits) + 1, SUBEXPFIN_K,
+                             (ref_params->wmmat[0] >> trans_prec_diff)) *
+                         trans_dec_factor;
+      params->wmmat[1] = aom_read_signed_primitive_refsubexpfin(
+                             r, (1 << trans_bits) + 1, SUBEXPFIN_K,
+                             (ref_params->wmmat[1] >> trans_prec_diff)) *
+                         trans_dec_factor;
+    case IDENTITY: break;
+    default: assert(0);
+  }
+  if (params->wmtype <= AFFINE)
+    if (!get_shear_params(params)) assert(0);
+}
+
+static void read_global_motion(AV1_COMMON *cm, aom_reader *r) {
+  int frame;
+  for (frame = LAST_FRAME; frame <= ALTREF_FRAME; ++frame) {
+    read_global_motion_params(
+        &cm->global_motion[frame], &cm->prev_frame->global_motion[frame],
+        cm->fc->global_motion_types_prob, r, cm->allow_high_precision_mv);
+    /*
+    printf("Dec Ref %d [%d/%d]: %d %d %d %d\n",
+           frame, cm->current_video_frame, cm->show_frame,
+           cm->global_motion[frame].wmmat[0],
+           cm->global_motion[frame].wmmat[1],
+           cm->global_motion[frame].wmmat[2],
+           cm->global_motion[frame].wmmat[3]);
+           */
+  }
+  memcpy(cm->cur_frame->global_motion, cm->global_motion,
+         TOTAL_REFS_PER_FRAME * sizeof(WarpedMotionParams));
+}
+#endif  // CONFIG_GLOBAL_MOTION
+
+static int read_compressed_header(AV1Decoder *pbi, const uint8_t *data,
+                                  size_t partition_size) {
+  AV1_COMMON *const cm = &pbi->common;
+#if CONFIG_SUPERTX
+  MACROBLOCKD *const xd = &pbi->mb;
+#endif
+  FRAME_CONTEXT *const fc = cm->fc;
+  aom_reader r;
+  int k, i;
+#if !CONFIG_EC_ADAPT || \
+    (CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION || CONFIG_EXT_INTER)
+  int j;
+#endif
+
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  r.window_size = 1 << cm->ans_window_size_log2;
+#endif
+  if (aom_reader_init(&r, data, partition_size, pbi->decrypt_cb,
+                      pbi->decrypt_state))
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate bool decoder 0");
+
+#if CONFIG_LOOP_RESTORATION
+  if (cm->rst_info[0].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[1].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[2].frame_restoration_type != RESTORE_NONE) {
+    av1_alloc_restoration_buffers(cm);
+    decode_restoration(cm, &r);
+  }
+#endif
+
+#if !CONFIG_EC_ADAPT
+  if (cm->tx_mode == TX_MODE_SELECT) read_tx_size_probs(fc, &r);
+#endif
+
+#if CONFIG_LV_MAP
+  av1_read_txb_probs(fc, cm->tx_mode, &r);
+#else  // CONFIG_LV_MAP
+#if !CONFIG_PVQ
+#if !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET)
+  read_coef_probs(fc, cm->tx_mode, &r);
+#endif  // !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET)
+#endif  // !CONFIG_PVQ
+#endif  // CONFIG_LV_MAP
+
+#if CONFIG_VAR_TX
+  for (k = 0; k < TXFM_PARTITION_CONTEXTS; ++k)
+    av1_diff_update_prob(&r, &fc->txfm_partition_prob[k], ACCT_STR);
+#endif  // CONFIG_VAR_TX
+  for (k = 0; k < SKIP_CONTEXTS; ++k)
+    av1_diff_update_prob(&r, &fc->skip_probs[k], ACCT_STR);
+
+#if CONFIG_DELTA_Q && !CONFIG_EC_ADAPT
+#if CONFIG_EXT_DELTA_Q
+  if (cm->delta_q_present_flag) {
+    for (k = 0; k < DELTA_Q_PROBS; ++k)
+      av1_diff_update_prob(&r, &fc->delta_q_prob[k], ACCT_STR);
+  }
+  if (cm->delta_lf_present_flag) {
+    for (k = 0; k < DELTA_LF_PROBS; ++k)
+      av1_diff_update_prob(&r, &fc->delta_lf_prob[k], ACCT_STR);
+  }
+#else
+  for (k = 0; k < DELTA_Q_PROBS; ++k)
+    av1_diff_update_prob(&r, &fc->delta_q_prob[k], ACCT_STR);
+#endif
+#endif
+
+#if !CONFIG_EC_ADAPT
+  if (cm->seg.enabled && cm->seg.update_map) {
+    if (cm->seg.temporal_update) {
+      for (k = 0; k < PREDICTION_PROBS; k++)
+        av1_diff_update_prob(&r, &cm->fc->seg.pred_probs[k], ACCT_STR);
+    }
+    for (k = 0; k < MAX_SEGMENTS - 1; k++)
+      av1_diff_update_prob(&r, &cm->fc->seg.tree_probs[k], ACCT_STR);
+  }
+
+  for (j = 0; j < INTRA_MODES; j++) {
+    for (i = 0; i < INTRA_MODES - 1; ++i)
+      av1_diff_update_prob(&r, &fc->uv_mode_prob[j][i], ACCT_STR);
+  }
+
+#if CONFIG_EXT_PARTITION_TYPES
+  for (j = 0; j < PARTITION_PLOFFSET; ++j)
+    for (i = 0; i < PARTITION_TYPES - 1; ++i)
+      av1_diff_update_prob(&r, &fc->partition_prob[j][i], ACCT_STR);
+  for (; j < PARTITION_CONTEXTS_PRIMARY; ++j)
+    for (i = 0; i < EXT_PARTITION_TYPES - 1; ++i)
+      av1_diff_update_prob(&r, &fc->partition_prob[j][i], ACCT_STR);
+#else
+  for (j = 0; j < PARTITION_CONTEXTS_PRIMARY; ++j)
+    for (i = 0; i < PARTITION_TYPES - 1; ++i)
+      av1_diff_update_prob(&r, &fc->partition_prob[j][i], ACCT_STR);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+#if CONFIG_UNPOISON_PARTITION_CTX
+  for (; j < PARTITION_CONTEXTS_PRIMARY + PARTITION_BLOCK_SIZES; ++j)
+    av1_diff_update_prob(&r, &fc->partition_prob[j][PARTITION_VERT], ACCT_STR);
+  for (; j < PARTITION_CONTEXTS_PRIMARY + 2 * PARTITION_BLOCK_SIZES; ++j)
+    av1_diff_update_prob(&r, &fc->partition_prob[j][PARTITION_HORZ], ACCT_STR);
+#endif  // CONFIG_UNPOISON_PARTITION_CTX
+
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+  for (i = 0; i < INTRA_FILTERS + 1; ++i)
+    for (j = 0; j < INTRA_FILTERS - 1; ++j)
+      av1_diff_update_prob(&r, &fc->intra_filter_probs[i][j], ACCT_STR);
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+#endif  // !CONFIG_EC_ADAPT
+
+  if (frame_is_intra_only(cm)) {
+    av1_copy(cm->kf_y_prob, av1_kf_y_mode_prob);
+#if CONFIG_EC_MULTISYMBOL
+    av1_copy(cm->fc->kf_y_cdf, av1_kf_y_mode_cdf);
+#endif
+#if !CONFIG_EC_ADAPT
+    for (k = 0; k < INTRA_MODES; k++)
+      for (j = 0; j < INTRA_MODES; j++)
+        for (i = 0; i < INTRA_MODES - 1; ++i)
+          av1_diff_update_prob(&r, &cm->kf_y_prob[k][j][i], ACCT_STR);
+#endif
+  } else {
+#if !CONFIG_REF_MV
+    nmv_context *const nmvc = &fc->nmvc;
+#endif
+    read_inter_mode_probs(fc, &r);
+
+#if CONFIG_EXT_INTER
+    read_inter_compound_mode_probs(fc, &r);
+    if (cm->reference_mode != COMPOUND_REFERENCE) {
+      for (i = 0; i < BLOCK_SIZE_GROUPS; i++) {
+        if (is_interintra_allowed_bsize_group(i)) {
+          av1_diff_update_prob(&r, &fc->interintra_prob[i], ACCT_STR);
+        }
+      }
+      for (i = 0; i < BLOCK_SIZE_GROUPS; i++) {
+        for (j = 0; j < INTERINTRA_MODES - 1; j++)
+          av1_diff_update_prob(&r, &fc->interintra_mode_prob[i][j], ACCT_STR);
+      }
+      for (i = 0; i < BLOCK_SIZES; i++) {
+        if (is_interintra_allowed_bsize(i) && is_interintra_wedge_used(i)) {
+          av1_diff_update_prob(&r, &fc->wedge_interintra_prob[i], ACCT_STR);
+        }
+      }
+    }
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+    if (cm->reference_mode != SINGLE_REFERENCE) {
+      for (i = 0; i < BLOCK_SIZES; i++) {
+        for (j = 0; j < COMPOUND_TYPES - 1; j++) {
+          av1_diff_update_prob(&r, &fc->compound_type_prob[i][j], ACCT_STR);
+        }
+      }
+    }
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    for (i = BLOCK_8X8; i < BLOCK_SIZES; ++i) {
+      for (j = 0; j < MOTION_MODES - 1; ++j)
+        av1_diff_update_prob(&r, &fc->motion_mode_prob[i][j], ACCT_STR);
+    }
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if !CONFIG_EC_ADAPT
+    if (cm->interp_filter == SWITCHABLE) read_switchable_interp_probs(fc, &r);
+#endif
+
+    for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+      av1_diff_update_prob(&r, &fc->intra_inter_prob[i], ACCT_STR);
+
+    if (cm->reference_mode != SINGLE_REFERENCE)
+      setup_compound_reference_mode(cm);
+    read_frame_reference_mode_probs(cm, &r);
+
+#if !CONFIG_EC_ADAPT
+    for (j = 0; j < BLOCK_SIZE_GROUPS; j++) {
+      for (i = 0; i < INTRA_MODES - 1; ++i)
+        av1_diff_update_prob(&r, &fc->y_mode_prob[j][i], ACCT_STR);
+    }
+#endif
+
+#if CONFIG_REF_MV
+    for (i = 0; i < NMV_CONTEXTS; ++i)
+      read_mv_probs(&fc->nmvc[i], cm->allow_high_precision_mv, &r);
+#else
+    read_mv_probs(nmvc, cm->allow_high_precision_mv, &r);
+#endif
+#if !CONFIG_EC_ADAPT
+    read_ext_tx_probs(fc, &r);
+#endif  // EC_ADAPT
+#if CONFIG_SUPERTX
+    if (!xd->lossless[0]) read_supertx_probs(fc, &r);
+#endif
+#if CONFIG_GLOBAL_MOTION
+    read_global_motion(cm, &r);
+#endif  // EC_ADAPT, DAALA_EC
+  }
+#if CONFIG_EC_MULTISYMBOL && !CONFIG_EC_ADAPT
+#if CONFIG_NEW_TOKENSET
+  av1_coef_head_cdfs(fc);
+#endif
+  /* Make tail distribution from head */
+  av1_coef_pareto_cdfs(fc);
+#if CONFIG_REF_MV
+  for (i = 0; i < NMV_CONTEXTS; ++i) av1_set_mv_cdfs(&fc->nmvc[i]);
+#else
+  av1_set_mv_cdfs(&fc->nmvc);
+#endif
+  av1_set_mode_cdfs(cm);
+#endif  // CONFIG_EC_MULTISYMBOL && !CONFIG_EC_ADAPT
+
+  return aom_reader_has_error(&r);
+}
+
+#ifdef NDEBUG
+#define debug_check_frame_counts(cm) (void)0
+#else  // !NDEBUG
+// Counts should only be incremented when frame_parallel_decoding_mode and
+// error_resilient_mode are disabled.
+static void debug_check_frame_counts(const AV1_COMMON *const cm) {
+  FRAME_COUNTS zero_counts;
+  av1_zero(zero_counts);
+  assert(cm->refresh_frame_context != REFRESH_FRAME_CONTEXT_BACKWARD ||
+         cm->error_resilient_mode);
+  assert(!memcmp(cm->counts.y_mode, zero_counts.y_mode,
+                 sizeof(cm->counts.y_mode)));
+  assert(!memcmp(cm->counts.uv_mode, zero_counts.uv_mode,
+                 sizeof(cm->counts.uv_mode)));
+  assert(!memcmp(cm->counts.partition, zero_counts.partition,
+                 sizeof(cm->counts.partition)));
+  assert(!memcmp(cm->counts.coef, zero_counts.coef, sizeof(cm->counts.coef)));
+  assert(!memcmp(cm->counts.eob_branch, zero_counts.eob_branch,
+                 sizeof(cm->counts.eob_branch)));
+#if CONFIG_EC_MULTISYMBOL
+  assert(!memcmp(cm->counts.blockz_count, zero_counts.blockz_count,
+                 sizeof(cm->counts.blockz_count)));
+#endif
+  assert(!memcmp(cm->counts.switchable_interp, zero_counts.switchable_interp,
+                 sizeof(cm->counts.switchable_interp)));
+  assert(!memcmp(cm->counts.inter_mode, zero_counts.inter_mode,
+                 sizeof(cm->counts.inter_mode)));
+#if CONFIG_EXT_INTER
+  assert(!memcmp(cm->counts.inter_compound_mode,
+                 zero_counts.inter_compound_mode,
+                 sizeof(cm->counts.inter_compound_mode)));
+  assert(!memcmp(cm->counts.interintra, zero_counts.interintra,
+                 sizeof(cm->counts.interintra)));
+  assert(!memcmp(cm->counts.wedge_interintra, zero_counts.wedge_interintra,
+                 sizeof(cm->counts.wedge_interintra)));
+  assert(!memcmp(cm->counts.compound_interinter,
+                 zero_counts.compound_interinter,
+                 sizeof(cm->counts.compound_interinter)));
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  assert(!memcmp(cm->counts.motion_mode, zero_counts.motion_mode,
+                 sizeof(cm->counts.motion_mode)));
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  assert(!memcmp(cm->counts.intra_inter, zero_counts.intra_inter,
+                 sizeof(cm->counts.intra_inter)));
+  assert(!memcmp(cm->counts.comp_inter, zero_counts.comp_inter,
+                 sizeof(cm->counts.comp_inter)));
+  assert(!memcmp(cm->counts.single_ref, zero_counts.single_ref,
+                 sizeof(cm->counts.single_ref)));
+  assert(!memcmp(cm->counts.comp_ref, zero_counts.comp_ref,
+                 sizeof(cm->counts.comp_ref)));
+#if CONFIG_EXT_REFS
+  assert(!memcmp(cm->counts.comp_bwdref, zero_counts.comp_bwdref,
+                 sizeof(cm->counts.comp_bwdref)));
+#endif  // CONFIG_EXT_REFS
+  assert(!memcmp(&cm->counts.tx_size, &zero_counts.tx_size,
+                 sizeof(cm->counts.tx_size)));
+  assert(!memcmp(cm->counts.skip, zero_counts.skip, sizeof(cm->counts.skip)));
+#if CONFIG_REF_MV
+  assert(
+      !memcmp(&cm->counts.mv[0], &zero_counts.mv[0], sizeof(cm->counts.mv[0])));
+  assert(
+      !memcmp(&cm->counts.mv[1], &zero_counts.mv[1], sizeof(cm->counts.mv[0])));
+#else
+  assert(!memcmp(&cm->counts.mv, &zero_counts.mv, sizeof(cm->counts.mv)));
+#endif
+  assert(!memcmp(cm->counts.inter_ext_tx, zero_counts.inter_ext_tx,
+                 sizeof(cm->counts.inter_ext_tx)));
+  assert(!memcmp(cm->counts.intra_ext_tx, zero_counts.intra_ext_tx,
+                 sizeof(cm->counts.intra_ext_tx)));
+}
+#endif  // NDEBUG
+
+static struct aom_read_bit_buffer *init_read_bit_buffer(
+    AV1Decoder *pbi, struct aom_read_bit_buffer *rb, const uint8_t *data,
+    const uint8_t *data_end, uint8_t clear_data[MAX_AV1_HEADER_SIZE]) {
+  rb->bit_offset = 0;
+  rb->error_handler = error_handler;
+  rb->error_handler_data = &pbi->common;
+  if (pbi->decrypt_cb) {
+    const int n = (int)AOMMIN(MAX_AV1_HEADER_SIZE, data_end - data);
+    pbi->decrypt_cb(pbi->decrypt_state, data, clear_data, n);
+    rb->bit_buffer = clear_data;
+    rb->bit_buffer_end = clear_data + n;
+  } else {
+    rb->bit_buffer = data;
+    rb->bit_buffer_end = data_end;
+  }
+  return rb;
+}
+
+//------------------------------------------------------------------------------
+
+int av1_read_sync_code(struct aom_read_bit_buffer *const rb) {
+  return aom_rb_read_literal(rb, 8) == AV1_SYNC_CODE_0 &&
+         aom_rb_read_literal(rb, 8) == AV1_SYNC_CODE_1 &&
+         aom_rb_read_literal(rb, 8) == AV1_SYNC_CODE_2;
+}
+
+void av1_read_frame_size(struct aom_read_bit_buffer *rb, int *width,
+                         int *height) {
+  *width = aom_rb_read_literal(rb, 16) + 1;
+  *height = aom_rb_read_literal(rb, 16) + 1;
+}
+
+BITSTREAM_PROFILE av1_read_profile(struct aom_read_bit_buffer *rb) {
+  int profile = aom_rb_read_bit(rb);
+  profile |= aom_rb_read_bit(rb) << 1;
+  if (profile > 2) profile += aom_rb_read_bit(rb);
+  return (BITSTREAM_PROFILE)profile;
+}
+
+#if CONFIG_EC_ADAPT
+static void make_update_tile_list_dec(AV1Decoder *pbi, int tile_rows,
+                                      int tile_cols, FRAME_CONTEXT *ec_ctxs[]) {
+  int i;
+  for (i = 0; i < tile_rows * tile_cols; ++i)
+    ec_ctxs[i] = &pbi->tile_data[i].tctx;
+}
+#endif
+
+void av1_decode_frame(AV1Decoder *pbi, const uint8_t *data,
+                      const uint8_t *data_end, const uint8_t **p_data_end) {
+  AV1_COMMON *const cm = &pbi->common;
+  MACROBLOCKD *const xd = &pbi->mb;
+  struct aom_read_bit_buffer rb;
+  int context_updated = 0;
+  uint8_t clear_data[MAX_AV1_HEADER_SIZE];
+  size_t first_partition_size;
+  YV12_BUFFER_CONFIG *new_fb;
+
+#if CONFIG_ADAPT_SCAN
+  av1_deliver_eob_threshold(cm, xd);
+#endif
+#if CONFIG_BITSTREAM_DEBUG
+  bitstream_queue_set_frame_read(cm->current_video_frame * 2 + cm->show_frame);
+#endif
+
+  first_partition_size = read_uncompressed_header(
+      pbi, init_read_bit_buffer(pbi, &rb, data, data_end, clear_data));
+
+#if CONFIG_EXT_TILE
+  // If cm->tile_encoding_mode == TILE_NORMAL, the independent decoding of a
+  // single tile or a section of a frame is not allowed.
+  if (!cm->tile_encoding_mode &&
+      (pbi->dec_tile_row >= 0 || pbi->dec_tile_col >= 0)) {
+    pbi->dec_tile_row = -1;
+    pbi->dec_tile_col = -1;
+  }
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_TILE_GROUPS
+  pbi->first_partition_size = first_partition_size;
+  pbi->uncomp_hdr_size = aom_rb_bytes_read(&rb);
+#endif
+  new_fb = get_frame_new_buffer(cm);
+  xd->cur_buf = new_fb;
+#if CONFIG_GLOBAL_MOTION
+  int i;
+  for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
+    set_default_warp_params(&cm->global_motion[i]);
+    set_default_warp_params(&cm->cur_frame->global_motion[i]);
+  }
+  xd->global_motion = cm->global_motion;
+#endif  // CONFIG_GLOBAL_MOTION
+
+  if (!first_partition_size) {
+    // showing a frame directly
+    *p_data_end = data + aom_rb_bytes_read(&rb);
+    return;
+  }
+
+  data += aom_rb_bytes_read(&rb);
+  if (!read_is_valid(data, first_partition_size, data_end))
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Truncated packet or corrupt header length");
+
+#if CONFIG_REF_MV
+  cm->setup_mi(cm);
+#endif
+
+#if CONFIG_TEMPMV_SIGNALING
+  if (cm->use_prev_frame_mvs) {
+    RefBuffer *last_fb_ref_buf = &cm->frame_refs[LAST_FRAME - LAST_FRAME];
+    cm->prev_frame = &cm->buffer_pool->frame_bufs[last_fb_ref_buf->idx];
+    assert(!cm->error_resilient_mode &&
+           cm->width == last_fb_ref_buf->buf->y_width &&
+           cm->height == last_fb_ref_buf->buf->y_height &&
+           !cm->prev_frame->intra_only);
+  }
+#else
+  cm->use_prev_frame_mvs =
+      !cm->error_resilient_mode && cm->width == cm->last_width &&
+      cm->height == cm->last_height && !cm->last_intra_only &&
+      cm->last_show_frame && (cm->last_frame_type != KEY_FRAME);
+#endif
+#if CONFIG_EXT_REFS
+  // NOTE(zoeliu): As cm->prev_frame can take neither a frame of
+  //               show_exisiting_frame=1, nor can it take a frame not used as
+  //               a reference, it is probable that by the time it is being
+  //               referred to, the frame buffer it originally points to may
+  //               already get expired and have been reassigned to the current
+  //               newly coded frame. Hence, we need to check whether this is
+  //               the case, and if yes, we have 2 choices:
+  //               (1) Simply disable the use of previous frame mvs; or
+  //               (2) Have cm->prev_frame point to one reference frame buffer,
+  //                   e.g. LAST_FRAME.
+  if (cm->use_prev_frame_mvs && !dec_is_ref_frame_buf(pbi, cm->prev_frame)) {
+    // Reassign the LAST_FRAME buffer to cm->prev_frame.
+    RefBuffer *last_fb_ref_buf = &cm->frame_refs[LAST_FRAME - LAST_FRAME];
+    cm->prev_frame = &cm->buffer_pool->frame_bufs[last_fb_ref_buf->idx];
+  }
+#endif  // CONFIG_EXT_REFS
+
+  av1_setup_block_planes(xd, cm->subsampling_x, cm->subsampling_y);
+
+  *cm->fc = cm->frame_contexts[cm->frame_context_idx];
+  if (!cm->fc->initialized)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Uninitialized entropy context.");
+
+  av1_zero(cm->counts);
+
+  xd->corrupted = 0;
+  new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);
+  if (new_fb->corrupted)
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Decode failed. Frame data header is corrupted.");
+
+  if (cm->lf.filter_level && !cm->skip_loop_filter) {
+    av1_loop_filter_frame_init(cm, cm->lf.filter_level);
+  }
+
+  // If encoded in frame parallel mode, frame context is ready after decoding
+  // the frame header.
+  if (cm->frame_parallel_decode &&
+      cm->refresh_frame_context != REFRESH_FRAME_CONTEXT_BACKWARD) {
+    AVxWorker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_FORWARD) {
+      context_updated = 1;
+      cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+    }
+    av1_frameworker_lock_stats(worker);
+    pbi->cur_buf->row = -1;
+    pbi->cur_buf->col = -1;
+    frame_worker_data->frame_context_ready = 1;
+    // Signal the main thread that context is ready.
+    av1_frameworker_signal_stats(worker);
+    av1_frameworker_unlock_stats(worker);
+  }
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  av1_copy(cm->starting_coef_probs, cm->fc->coef_probs);
+  cm->coef_probs_update_idx = 0;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+  if (pbi->max_threads > 1 && !CONFIG_CB4X4 &&
+#if CONFIG_EXT_TILE
+      pbi->dec_tile_col < 0 &&  // Decoding all columns
+#endif                          // CONFIG_EXT_TILE
+      cm->tile_cols > 1) {
+    // Multi-threaded tile decoder
+    *p_data_end = decode_tiles_mt(pbi, data + first_partition_size, data_end);
+    if (!xd->corrupted) {
+      if (!cm->skip_loop_filter) {
+        // If multiple threads are used to decode tiles, then we use those
+        // threads to do parallel loopfiltering.
+        av1_loop_filter_frame_mt(new_fb, cm, pbi->mb.plane, cm->lf.filter_level,
+                                 0, 0, pbi->tile_workers, pbi->num_tile_workers,
+                                 &pbi->lf_row_sync);
+      }
+    } else {
+      aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                         "Decode failed. Frame data is corrupted.");
+    }
+  } else {
+    *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
+  }
+
+#if CONFIG_CDEF
+  if (!cm->skip_loop_filter) {
+    av1_cdef_frame(&pbi->cur_buf->buf, cm, &pbi->mb);
+  }
+#endif  // CONFIG_CDEF
+
+#if CONFIG_LOOP_RESTORATION
+  if (cm->rst_info[0].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[1].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[2].frame_restoration_type != RESTORE_NONE) {
+    av1_loop_restoration_frame(new_fb, cm, cm->rst_info, 7, 0, NULL);
+  }
+#endif  // CONFIG_LOOP_RESTORATION
+
+  if (!xd->corrupted) {
+    if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+#if CONFIG_EC_ADAPT
+      FRAME_CONTEXT **tile_ctxs = aom_malloc(cm->tile_rows * cm->tile_cols *
+                                             sizeof(&pbi->tile_data[0].tctx));
+      aom_cdf_prob **cdf_ptrs =
+          aom_malloc(cm->tile_rows * cm->tile_cols *
+                     sizeof(&pbi->tile_data[0].tctx.partition_cdf[0][0]));
+      make_update_tile_list_dec(pbi, cm->tile_rows, cm->tile_cols, tile_ctxs);
+#endif
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+      cm->partial_prob_update = 0;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+      av1_adapt_coef_probs(cm);
+      av1_adapt_intra_frame_probs(cm);
+#if CONFIG_EC_ADAPT
+      av1_average_tile_coef_cdfs(pbi->common.fc, tile_ctxs, cdf_ptrs,
+                                 cm->tile_rows * cm->tile_cols);
+      av1_average_tile_intra_cdfs(pbi->common.fc, tile_ctxs, cdf_ptrs,
+                                  cm->tile_rows * cm->tile_cols);
+#if CONFIG_PVQ
+      av1_average_tile_pvq_cdfs(pbi->common.fc, tile_ctxs,
+                                cm->tile_rows * cm->tile_cols);
+#endif  // CONFIG_PVQ
+#endif  // CONFIG_EC_ADAPT
+#if CONFIG_ADAPT_SCAN
+      av1_adapt_scan_order(cm);
+#endif  // CONFIG_ADAPT_SCAN
+
+      if (!frame_is_intra_only(cm)) {
+        av1_adapt_inter_frame_probs(cm);
+        av1_adapt_mv_probs(cm, cm->allow_high_precision_mv);
+#if CONFIG_EC_ADAPT
+        av1_average_tile_inter_cdfs(&pbi->common, pbi->common.fc, tile_ctxs,
+                                    cdf_ptrs, cm->tile_rows * cm->tile_cols);
+        av1_average_tile_mv_cdfs(pbi->common.fc, tile_ctxs, cdf_ptrs,
+                                 cm->tile_rows * cm->tile_cols);
+#endif
+      }
+#if CONFIG_EC_ADAPT
+      aom_free(tile_ctxs);
+      aom_free(cdf_ptrs);
+#endif
+    } else {
+      debug_check_frame_counts(cm);
+    }
+  } else {
+    aom_internal_error(&cm->error, AOM_CODEC_CORRUPT_FRAME,
+                       "Decode failed. Frame data is corrupted.");
+  }
+
+#if CONFIG_INSPECTION
+  if (pbi->inspect_cb != NULL) {
+    (*pbi->inspect_cb)(pbi, pbi->inspect_ctx);
+  }
+#endif
+
+  // Non frame parallel update frame context here.
+  if (!cm->error_resilient_mode && !context_updated)
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+}
diff --git a/third_party/aom/av1/decoder/decodeframe.h b/third_party/aom/av1/decoder/decodeframe.h
new file mode 100644
index 0000000000..a904658b08
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodeframe.h
@@ -0,0 +1,39 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DECODEFRAME_H_
+#define AV1_DECODER_DECODEFRAME_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1Decoder;
+struct aom_read_bit_buffer;
+
+#if CONFIG_REFERENCE_BUFFER
+/* Placeholder for now */
+void read_sequence_header(SequenceHeader *seq_params);
+#endif
+
+int av1_read_sync_code(struct aom_read_bit_buffer *const rb);
+void av1_read_frame_size(struct aom_read_bit_buffer *rb, int *width,
+                         int *height);
+BITSTREAM_PROFILE av1_read_profile(struct aom_read_bit_buffer *rb);
+
+void av1_decode_frame(struct AV1Decoder *pbi, const uint8_t *data,
+                      const uint8_t *data_end, const uint8_t **p_data_end);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_DECODER_DECODEFRAME_H_
diff --git a/third_party/aom/av1/decoder/decodemv.c b/third_party/aom/av1/decoder/decodemv.c
new file mode 100644
index 0000000000..ec0f877512
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodemv.c
@@ -0,0 +1,2405 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/common.h"
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/entropymv.h"
+#include "av1/common/mvref_common.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/reconinter.h"
+#if CONFIG_EXT_INTRA
+#include "av1/common/reconintra.h"
+#endif  // CONFIG_EXT_INTRA
+#include "av1/common/seg_common.h"
+#if CONFIG_WARPED_MOTION
+#include "av1/common/warped_motion.h"
+#endif  // CONFIG_WARPED_MOTION
+
+#include "av1/decoder/decodeframe.h"
+#include "av1/decoder/decodemv.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+
+#define ACCT_STR __func__
+#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
+static INLINE int read_uniform(aom_reader *r, int n) {
+  const int l = get_unsigned_bits(n);
+  const int m = (1 << l) - n;
+  const int v = aom_read_literal(r, l - 1, ACCT_STR);
+  assert(l != 0);
+  if (v < m)
+    return v;
+  else
+    return (v << 1) - m + aom_read_literal(r, 1, ACCT_STR);
+}
+#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
+
+#if CONFIG_EC_MULTISYMBOL
+static PREDICTION_MODE read_intra_mode(aom_reader *r, aom_cdf_prob *cdf) {
+  return (PREDICTION_MODE)
+      av1_intra_mode_inv[aom_read_symbol(r, cdf, INTRA_MODES, ACCT_STR)];
+}
+#else
+static PREDICTION_MODE read_intra_mode(aom_reader *r, const aom_prob *p) {
+  return (PREDICTION_MODE)aom_read_tree(r, av1_intra_mode_tree, p, ACCT_STR);
+}
+#endif
+
+#if CONFIG_DELTA_Q
+static int read_delta_qindex(AV1_COMMON *cm, MACROBLOCKD *xd, aom_reader *r,
+                             MB_MODE_INFO *const mbmi, int mi_col, int mi_row) {
+  FRAME_COUNTS *counts = xd->counts;
+  int sign, abs, reduced_delta_qindex = 0;
+  BLOCK_SIZE bsize = mbmi->sb_type;
+  const int b_col = mi_col & MAX_MIB_MASK;
+  const int b_row = mi_row & MAX_MIB_MASK;
+  const int read_delta_q_flag = (b_col == 0 && b_row == 0);
+  int rem_bits, thr;
+  int i, smallval;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  if ((bsize != BLOCK_LARGEST || mbmi->skip == 0) && read_delta_q_flag) {
+#if !CONFIG_EC_MULTISYMBOL
+    int bit = 1;
+    abs = 0;
+    while (abs < DELTA_Q_SMALL && bit) {
+      bit = aom_read(r, ec_ctx->delta_q_prob[abs], ACCT_STR);
+      abs += bit;
+    }
+#else
+    abs = aom_read_symbol(r, ec_ctx->delta_q_cdf, DELTA_Q_PROBS + 1, ACCT_STR);
+#endif
+    smallval = (abs < DELTA_Q_SMALL);
+    if (counts) {
+      for (i = 0; i < abs; ++i) counts->delta_q[i][1]++;
+      if (smallval) counts->delta_q[abs][0]++;
+    }
+
+    if (!smallval) {
+      rem_bits = aom_read_literal(r, 3, ACCT_STR);
+      thr = (1 << rem_bits) + 1;
+      abs = aom_read_literal(r, rem_bits, ACCT_STR) + thr;
+    }
+
+    if (abs) {
+      sign = aom_read_bit(r, ACCT_STR);
+    } else {
+      sign = 1;
+    }
+
+    reduced_delta_qindex = sign ? -abs : abs;
+  }
+  return reduced_delta_qindex;
+}
+#if CONFIG_EXT_DELTA_Q
+static int read_delta_lflevel(AV1_COMMON *cm, MACROBLOCKD *xd, aom_reader *r,
+                              MB_MODE_INFO *const mbmi, int mi_col,
+                              int mi_row) {
+  FRAME_COUNTS *counts = xd->counts;
+  int sign, abs, reduced_delta_lflevel = 0;
+  BLOCK_SIZE bsize = mbmi->sb_type;
+  const int b_col = mi_col & MAX_MIB_MASK;
+  const int b_row = mi_row & MAX_MIB_MASK;
+  const int read_delta_lf_flag = (b_col == 0 && b_row == 0);
+  int rem_bits, thr;
+  int i, smallval;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  if ((bsize != BLOCK_64X64 || mbmi->skip == 0) && read_delta_lf_flag) {
+#if !CONFIG_EC_MULTISYMBOL
+    int bit = 1;
+    abs = 0;
+    while (abs < DELTA_LF_SMALL && bit) {
+      bit = aom_read(r, ec_ctx->delta_lf_prob[abs], ACCT_STR);
+      abs += bit;
+    }
+#else
+    abs =
+        aom_read_symbol(r, ec_ctx->delta_lf_cdf, DELTA_LF_PROBS + 1, ACCT_STR);
+#endif
+    smallval = (abs < DELTA_LF_SMALL);
+    if (counts) {
+      for (i = 0; i < abs; ++i) counts->delta_lf[i][1]++;
+      if (smallval) counts->delta_lf[abs][0]++;
+    }
+    if (!smallval) {
+      rem_bits = aom_read_literal(r, 3, ACCT_STR);
+      thr = (1 << rem_bits) + 1;
+      abs = aom_read_literal(r, rem_bits, ACCT_STR) + thr;
+    }
+
+    if (abs) {
+      sign = aom_read_bit(r, ACCT_STR);
+    } else {
+      sign = 1;
+    }
+
+    reduced_delta_lflevel = sign ? -abs : abs;
+  }
+  return reduced_delta_lflevel;
+}
+#endif
+#endif
+
+static PREDICTION_MODE read_intra_mode_y(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                         aom_reader *r, int size_group) {
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  const PREDICTION_MODE y_mode =
+#if CONFIG_EC_MULTISYMBOL
+      read_intra_mode(r, ec_ctx->y_mode_cdf[size_group]);
+#else
+      read_intra_mode(r, cm->fc->y_mode_prob[size_group]);
+#endif
+  FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_EC_ADAPT
+  (void)cm;
+#endif
+  if (counts) ++counts->y_mode[size_group][y_mode];
+  return y_mode;
+}
+
+static PREDICTION_MODE read_intra_mode_uv(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                          aom_reader *r,
+                                          PREDICTION_MODE y_mode) {
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  const PREDICTION_MODE uv_mode =
+#if CONFIG_EC_MULTISYMBOL
+      read_intra_mode(r, ec_ctx->uv_mode_cdf[y_mode]);
+#else
+      read_intra_mode(r, cm->fc->uv_mode_prob[y_mode]);
+#endif
+  FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_EC_ADAPT
+  (void)cm;
+#endif
+  if (counts) ++counts->uv_mode[y_mode][uv_mode];
+  return uv_mode;
+}
+
+#if CONFIG_EXT_INTER
+static INTERINTRA_MODE read_interintra_mode(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                            aom_reader *r, int size_group) {
+  const INTERINTRA_MODE ii_mode = (INTERINTRA_MODE)aom_read_tree(
+      r, av1_interintra_mode_tree, cm->fc->interintra_mode_prob[size_group],
+      ACCT_STR);
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts) ++counts->interintra_mode[size_group][ii_mode];
+  return ii_mode;
+}
+#endif  // CONFIG_EXT_INTER
+
+static PREDICTION_MODE read_inter_mode(FRAME_CONTEXT *ec_ctx, MACROBLOCKD *xd,
+                                       aom_reader *r, int16_t ctx) {
+#if CONFIG_REF_MV
+  FRAME_COUNTS *counts = xd->counts;
+  int16_t mode_ctx = ctx & NEWMV_CTX_MASK;
+  aom_prob mode_prob = ec_ctx->newmv_prob[mode_ctx];
+
+  if (aom_read(r, mode_prob, ACCT_STR) == 0) {
+    if (counts) ++counts->newmv_mode[mode_ctx][0];
+    return NEWMV;
+  }
+  if (counts) ++counts->newmv_mode[mode_ctx][1];
+
+  if (ctx & (1 << ALL_ZERO_FLAG_OFFSET)) return ZEROMV;
+
+  mode_ctx = (ctx >> ZEROMV_OFFSET) & ZEROMV_CTX_MASK;
+
+  mode_prob = ec_ctx->zeromv_prob[mode_ctx];
+  if (aom_read(r, mode_prob, ACCT_STR) == 0) {
+    if (counts) ++counts->zeromv_mode[mode_ctx][0];
+    return ZEROMV;
+  }
+  if (counts) ++counts->zeromv_mode[mode_ctx][1];
+
+  mode_ctx = (ctx >> REFMV_OFFSET) & REFMV_CTX_MASK;
+
+  if (ctx & (1 << SKIP_NEARESTMV_OFFSET)) mode_ctx = 6;
+  if (ctx & (1 << SKIP_NEARMV_OFFSET)) mode_ctx = 7;
+  if (ctx & (1 << SKIP_NEARESTMV_SUB8X8_OFFSET)) mode_ctx = 8;
+
+  mode_prob = ec_ctx->refmv_prob[mode_ctx];
+
+  if (aom_read(r, mode_prob, ACCT_STR) == 0) {
+    if (counts) ++counts->refmv_mode[mode_ctx][0];
+
+    return NEARESTMV;
+  } else {
+    if (counts) ++counts->refmv_mode[mode_ctx][1];
+    return NEARMV;
+  }
+
+  // Invalid prediction mode.
+  assert(0);
+#else
+#if CONFIG_EC_MULTISYMBOL
+  const int mode = av1_inter_mode_inv[aom_read_symbol(
+      r, ec_ctx->inter_mode_cdf[ctx], INTER_MODES, ACCT_STR)];
+#else
+  const int mode = aom_read_tree(r, av1_inter_mode_tree,
+                                 ec_ctx->inter_mode_probs[ctx], ACCT_STR);
+#endif
+  FRAME_COUNTS *counts = xd->counts;
+  if (counts) ++counts->inter_mode[ctx][mode];
+
+  return NEARESTMV + mode;
+#endif
+}
+
+#if CONFIG_REF_MV
+static void read_drl_idx(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                         MB_MODE_INFO *mbmi, aom_reader *r) {
+  uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+  mbmi->ref_mv_idx = 0;
+
+#if CONFIG_EXT_INTER
+  if (mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV) {
+#else
+  if (mbmi->mode == NEWMV) {
+#endif
+    int idx;
+    for (idx = 0; idx < 2; ++idx) {
+      if (xd->ref_mv_count[ref_frame_type] > idx + 1) {
+        uint8_t drl_ctx = av1_drl_ctx(xd->ref_mv_stack[ref_frame_type], idx);
+        aom_prob drl_prob = cm->fc->drl_prob[drl_ctx];
+        if (!aom_read(r, drl_prob, ACCT_STR)) {
+          mbmi->ref_mv_idx = idx;
+          if (xd->counts) ++xd->counts->drl_mode[drl_ctx][0];
+          return;
+        }
+        mbmi->ref_mv_idx = idx + 1;
+        if (xd->counts) ++xd->counts->drl_mode[drl_ctx][1];
+      }
+    }
+  }
+
+  if (have_nearmv_in_inter_mode(mbmi->mode)) {
+    int idx;
+    // Offset the NEARESTMV mode.
+    // TODO(jingning): Unify the two syntax decoding loops after the NEARESTMV
+    // mode is factored in.
+    for (idx = 1; idx < 3; ++idx) {
+      if (xd->ref_mv_count[ref_frame_type] > idx + 1) {
+        uint8_t drl_ctx = av1_drl_ctx(xd->ref_mv_stack[ref_frame_type], idx);
+        aom_prob drl_prob = cm->fc->drl_prob[drl_ctx];
+        if (!aom_read(r, drl_prob, ACCT_STR)) {
+          mbmi->ref_mv_idx = idx - 1;
+          if (xd->counts) ++xd->counts->drl_mode[drl_ctx][0];
+          return;
+        }
+        mbmi->ref_mv_idx = idx;
+        if (xd->counts) ++xd->counts->drl_mode[drl_ctx][1];
+      }
+    }
+  }
+}
+#endif
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+static MOTION_MODE read_motion_mode(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                    MODE_INFO *mi, aom_reader *r) {
+  MB_MODE_INFO *mbmi = &mi->mbmi;
+  const MOTION_MODE last_motion_mode_allowed = motion_mode_allowed(
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      0, xd->global_motion,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      mi);
+  int motion_mode;
+  FRAME_COUNTS *counts = xd->counts;
+
+  if (last_motion_mode_allowed == SIMPLE_TRANSLATION) return SIMPLE_TRANSLATION;
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  if (last_motion_mode_allowed == OBMC_CAUSAL) {
+    motion_mode = aom_read(r, cm->fc->obmc_prob[mbmi->sb_type], ACCT_STR);
+    if (counts) ++counts->obmc[mbmi->sb_type][motion_mode];
+    return (MOTION_MODE)(SIMPLE_TRANSLATION + motion_mode);
+  } else {
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+    motion_mode =
+        aom_read_tree(r, av1_motion_mode_tree,
+                      cm->fc->motion_mode_prob[mbmi->sb_type], ACCT_STR);
+    if (counts) ++counts->motion_mode[mbmi->sb_type][motion_mode];
+    return (MOTION_MODE)(SIMPLE_TRANSLATION + motion_mode);
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  }
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+}
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_EXT_INTER
+static PREDICTION_MODE read_inter_compound_mode(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                                aom_reader *r, int16_t ctx) {
+  const int mode =
+      aom_read_tree(r, av1_inter_compound_mode_tree,
+                    cm->fc->inter_compound_mode_probs[ctx], ACCT_STR);
+  FRAME_COUNTS *counts = xd->counts;
+
+  if (counts) ++counts->inter_compound_mode[ctx][mode];
+
+  assert(is_inter_compound_mode(NEAREST_NEARESTMV + mode));
+  return NEAREST_NEARESTMV + mode;
+}
+#endif  // CONFIG_EXT_INTER
+
+static int read_segment_id(aom_reader *r, struct segmentation_probs *segp) {
+#if CONFIG_EC_MULTISYMBOL
+  return aom_read_symbol(r, segp->tree_cdf, MAX_SEGMENTS, ACCT_STR);
+#else
+  return aom_read_tree(r, av1_segment_tree, segp->tree_probs, ACCT_STR);
+#endif
+}
+
+#if CONFIG_VAR_TX
+static void read_tx_size_vartx(AV1_COMMON *cm, MACROBLOCKD *xd,
+                               MB_MODE_INFO *mbmi, FRAME_COUNTS *counts,
+                               TX_SIZE tx_size, int depth, int blk_row,
+                               int blk_col, aom_reader *r) {
+  int is_split = 0;
+  const int tx_row = blk_row >> 1;
+  const int tx_col = blk_col >> 1;
+  const int max_blocks_high = max_block_high(xd, mbmi->sb_type, 0);
+  const int max_blocks_wide = max_block_wide(xd, mbmi->sb_type, 0);
+  int ctx = txfm_partition_context(xd->above_txfm_context + tx_col,
+                                   xd->left_txfm_context + tx_row,
+                                   mbmi->sb_type, tx_size);
+  TX_SIZE(*const inter_tx_size)
+  [MAX_MIB_SIZE] =
+      (TX_SIZE(*)[MAX_MIB_SIZE]) & mbmi->inter_tx_size[tx_row][tx_col];
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  if (depth == MAX_VARTX_DEPTH) {
+    int idx, idy;
+    inter_tx_size[0][0] = tx_size;
+    for (idy = 0; idy < tx_size_high_unit[tx_size] / 2; ++idy)
+      for (idx = 0; idx < tx_size_wide_unit[tx_size] / 2; ++idx)
+        inter_tx_size[idy][idx] = tx_size;
+    mbmi->tx_size = tx_size;
+    mbmi->min_tx_size = AOMMIN(mbmi->min_tx_size, get_min_tx_size(tx_size));
+    if (counts) ++counts->txfm_partition[ctx][0];
+    txfm_partition_update(xd->above_txfm_context + tx_col,
+                          xd->left_txfm_context + tx_row, tx_size, tx_size);
+    return;
+  }
+
+  is_split = aom_read(r, cm->fc->txfm_partition_prob[ctx], ACCT_STR);
+
+  if (is_split) {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int i;
+
+    if (counts) ++counts->txfm_partition[ctx][1];
+
+    if (tx_size == TX_8X8) {
+      int idx, idy;
+      inter_tx_size[0][0] = sub_txs;
+      for (idy = 0; idy < tx_size_high_unit[tx_size] / 2; ++idy)
+        for (idx = 0; idx < tx_size_wide_unit[tx_size] / 2; ++idx)
+          inter_tx_size[idy][idx] = inter_tx_size[0][0];
+      mbmi->tx_size = sub_txs;
+      mbmi->min_tx_size = get_min_tx_size(mbmi->tx_size);
+      txfm_partition_update(xd->above_txfm_context + tx_col,
+                            xd->left_txfm_context + tx_row, sub_txs, tx_size);
+      return;
+    }
+
+    assert(bsl > 0);
+    for (i = 0; i < 4; ++i) {
+      int offsetr = blk_row + (i >> 1) * bsl;
+      int offsetc = blk_col + (i & 0x01) * bsl;
+      read_tx_size_vartx(cm, xd, mbmi, counts, sub_txs, depth + 1, offsetr,
+                         offsetc, r);
+    }
+  } else {
+    int idx, idy;
+    inter_tx_size[0][0] = tx_size;
+    for (idy = 0; idy < tx_size_high_unit[tx_size] / 2; ++idy)
+      for (idx = 0; idx < tx_size_wide_unit[tx_size] / 2; ++idx)
+        inter_tx_size[idy][idx] = tx_size;
+    mbmi->tx_size = tx_size;
+    mbmi->min_tx_size = AOMMIN(mbmi->min_tx_size, get_min_tx_size(tx_size));
+    if (counts) ++counts->txfm_partition[ctx][0];
+    txfm_partition_update(xd->above_txfm_context + tx_col,
+                          xd->left_txfm_context + tx_row, tx_size, tx_size);
+  }
+}
+#endif
+
+static TX_SIZE read_selected_tx_size(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                     int tx_size_cat, aom_reader *r) {
+  FRAME_COUNTS *counts = xd->counts;
+  const int ctx = get_tx_size_context(xd);
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  const int depth =
+#if CONFIG_EC_MULTISYMBOL
+      aom_read_symbol(r, ec_ctx->tx_size_cdf[tx_size_cat][ctx], tx_size_cat + 2,
+                      ACCT_STR);
+#else
+      aom_read_tree(r, av1_tx_size_tree[tx_size_cat],
+                    ec_ctx->tx_size_probs[tx_size_cat][ctx], ACCT_STR);
+#endif
+  const TX_SIZE tx_size = depth_to_tx_size(depth);
+#if CONFIG_RECT_TX
+  assert(!is_rect_tx(tx_size));
+#endif  // CONFIG_RECT_TX
+  if (counts) ++counts->tx_size[tx_size_cat][ctx][depth];
+  return tx_size;
+}
+
+static TX_SIZE read_tx_size(AV1_COMMON *cm, MACROBLOCKD *xd, int is_inter,
+                            int allow_select_inter, aom_reader *r) {
+  const TX_MODE tx_mode = cm->tx_mode;
+  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  if (xd->lossless[xd->mi[0]->mbmi.segment_id]) return TX_4X4;
+#if CONFIG_CB4X4 && (CONFIG_VAR_TX || CONFIG_EXT_TX) && CONFIG_RECT_TX
+  if (bsize > BLOCK_4X4) {
+#else
+  if (bsize >= BLOCK_8X8) {
+#endif  // CONFIG_CB4X4 && CONFIG_VAR_TX
+    if ((!is_inter || allow_select_inter) && tx_mode == TX_MODE_SELECT) {
+      const int32_t tx_size_cat = is_inter ? inter_tx_size_cat_lookup[bsize]
+                                           : intra_tx_size_cat_lookup[bsize];
+      const TX_SIZE coded_tx_size =
+          read_selected_tx_size(cm, xd, tx_size_cat, r);
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+      if (coded_tx_size > max_txsize_lookup[bsize]) {
+        assert(coded_tx_size == max_txsize_lookup[bsize] + 1);
+        return max_txsize_rect_lookup[bsize];
+      }
+#else
+      assert(coded_tx_size <= max_txsize_lookup[bsize]);
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+      return coded_tx_size;
+    } else {
+      return tx_size_from_tx_mode(bsize, tx_mode, is_inter);
+    }
+  } else {
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+    assert(IMPLIES(tx_mode == ONLY_4X4, bsize == BLOCK_4X4));
+    return max_txsize_rect_lookup[bsize];
+#else
+    return TX_4X4;
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+  }
+}
+
+static int dec_get_segment_id(const AV1_COMMON *cm, const uint8_t *segment_ids,
+                              int mi_offset, int x_mis, int y_mis) {
+  int x, y, segment_id = INT_MAX;
+
+  for (y = 0; y < y_mis; y++)
+    for (x = 0; x < x_mis; x++)
+      segment_id =
+          AOMMIN(segment_id, segment_ids[mi_offset + y * cm->mi_cols + x]);
+
+  assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
+  return segment_id;
+}
+
+static void set_segment_id(AV1_COMMON *cm, int mi_offset, int x_mis, int y_mis,
+                           int segment_id) {
+  int x, y;
+
+  assert(segment_id >= 0 && segment_id < MAX_SEGMENTS);
+
+  for (y = 0; y < y_mis; y++)
+    for (x = 0; x < x_mis; x++)
+      cm->current_frame_seg_map[mi_offset + y * cm->mi_cols + x] = segment_id;
+}
+
+static int read_intra_segment_id(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                 int mi_offset, int x_mis, int y_mis,
+                                 aom_reader *r) {
+  struct segmentation *const seg = &cm->seg;
+  FRAME_COUNTS *counts = xd->counts;
+  struct segmentation_probs *const segp = &cm->fc->seg;
+  int segment_id;
+
+  if (!seg->enabled) return 0;  // Default for disabled segmentation
+
+  assert(seg->update_map && !seg->temporal_update);
+
+  segment_id = read_segment_id(r, segp);
+  if (counts) ++counts->seg.tree_total[segment_id];
+  set_segment_id(cm, mi_offset, x_mis, y_mis, segment_id);
+  return segment_id;
+}
+
+static void copy_segment_id(const AV1_COMMON *cm,
+                            const uint8_t *last_segment_ids,
+                            uint8_t *current_segment_ids, int mi_offset,
+                            int x_mis, int y_mis) {
+  int x, y;
+
+  for (y = 0; y < y_mis; y++)
+    for (x = 0; x < x_mis; x++)
+      current_segment_ids[mi_offset + y * cm->mi_cols + x] =
+          last_segment_ids ? last_segment_ids[mi_offset + y * cm->mi_cols + x]
+                           : 0;
+}
+
+static int read_inter_segment_id(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                 int mi_row, int mi_col, aom_reader *r) {
+  struct segmentation *const seg = &cm->seg;
+  FRAME_COUNTS *counts = xd->counts;
+  struct segmentation_probs *const segp = &cm->fc->seg;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int predicted_segment_id, segment_id;
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = mi_size_wide[mbmi->sb_type];
+  const int bh = mi_size_high[mbmi->sb_type];
+
+  // TODO(slavarnway): move x_mis, y_mis into xd ?????
+  const int x_mis = AOMMIN(cm->mi_cols - mi_col, bw);
+  const int y_mis = AOMMIN(cm->mi_rows - mi_row, bh);
+
+  if (!seg->enabled) return 0;  // Default for disabled segmentation
+
+  predicted_segment_id = cm->last_frame_seg_map
+                             ? dec_get_segment_id(cm, cm->last_frame_seg_map,
+                                                  mi_offset, x_mis, y_mis)
+                             : 0;
+
+  if (!seg->update_map) {
+    copy_segment_id(cm, cm->last_frame_seg_map, cm->current_frame_seg_map,
+                    mi_offset, x_mis, y_mis);
+    return predicted_segment_id;
+  }
+
+  if (seg->temporal_update) {
+    const int ctx = av1_get_pred_context_seg_id(xd);
+    const aom_prob pred_prob = segp->pred_probs[ctx];
+    mbmi->seg_id_predicted = aom_read(r, pred_prob, ACCT_STR);
+    if (counts) ++counts->seg.pred[ctx][mbmi->seg_id_predicted];
+    if (mbmi->seg_id_predicted) {
+      segment_id = predicted_segment_id;
+    } else {
+      segment_id = read_segment_id(r, segp);
+      if (counts) ++counts->seg.tree_mispred[segment_id];
+    }
+  } else {
+    segment_id = read_segment_id(r, segp);
+    if (counts) ++counts->seg.tree_total[segment_id];
+  }
+  set_segment_id(cm, mi_offset, x_mis, y_mis, segment_id);
+  return segment_id;
+}
+
+static int read_skip(AV1_COMMON *cm, const MACROBLOCKD *xd, int segment_id,
+                     aom_reader *r) {
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
+    return 1;
+  } else {
+    const int ctx = av1_get_skip_context(xd);
+    const int skip = aom_read(r, cm->fc->skip_probs[ctx], ACCT_STR);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts) ++counts->skip[ctx][skip];
+    return skip;
+  }
+}
+
+#if CONFIG_PALETTE
+static void read_palette_mode_info(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                   aom_reader *r) {
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  int i, n;
+  PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+
+  if (mbmi->mode == DC_PRED) {
+    int palette_y_mode_ctx = 0;
+    if (above_mi)
+      palette_y_mode_ctx +=
+          (above_mi->mbmi.palette_mode_info.palette_size[0] > 0);
+    if (left_mi)
+      palette_y_mode_ctx +=
+          (left_mi->mbmi.palette_mode_info.palette_size[0] > 0);
+    if (aom_read(r, av1_default_palette_y_mode_prob[bsize - BLOCK_8X8]
+                                                   [palette_y_mode_ctx],
+                 ACCT_STR)) {
+      pmi->palette_size[0] =
+          aom_read_tree(r, av1_palette_size_tree,
+                        av1_default_palette_y_size_prob[bsize - BLOCK_8X8],
+                        ACCT_STR) +
+          2;
+      n = pmi->palette_size[0];
+#if CONFIG_PALETTE_DELTA_ENCODING
+      const int min_bits = cm->bit_depth - 3;
+      int bits = min_bits + aom_read_literal(r, 2, ACCT_STR);
+      pmi->palette_colors[0] = aom_read_literal(r, cm->bit_depth, ACCT_STR);
+      for (i = 1; i < n; ++i) {
+        pmi->palette_colors[i] = pmi->palette_colors[i - 1] +
+                                 aom_read_literal(r, bits, ACCT_STR) + 1;
+        bits = AOMMIN(
+            bits, av1_ceil_log2((1 << cm->bit_depth) - pmi->palette_colors[i]));
+      }
+#else
+      for (i = 0; i < n; ++i)
+        pmi->palette_colors[i] = aom_read_literal(r, cm->bit_depth, ACCT_STR);
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+      xd->plane[0].color_index_map[0] = read_uniform(r, n);
+      assert(xd->plane[0].color_index_map[0] < n);
+    }
+  }
+
+  if (mbmi->uv_mode == DC_PRED) {
+    const int palette_uv_mode_ctx = (pmi->palette_size[0] > 0);
+    if (aom_read(r, av1_default_palette_uv_mode_prob[palette_uv_mode_ctx],
+                 ACCT_STR)) {
+      pmi->palette_size[1] =
+          aom_read_tree(r, av1_palette_size_tree,
+                        av1_default_palette_uv_size_prob[bsize - BLOCK_8X8],
+                        ACCT_STR) +
+          2;
+      n = pmi->palette_size[1];
+#if CONFIG_PALETTE_DELTA_ENCODING
+      // U channel colors.
+      const int min_bits_u = cm->bit_depth - 3;
+      int bits = min_bits_u + aom_read_literal(r, 2, ACCT_STR);
+      pmi->palette_colors[PALETTE_MAX_SIZE] =
+          aom_read_literal(r, cm->bit_depth, ACCT_STR);
+      for (i = 1; i < n; ++i) {
+        pmi->palette_colors[PALETTE_MAX_SIZE + i] =
+            pmi->palette_colors[PALETTE_MAX_SIZE + i - 1] +
+            aom_read_literal(r, bits, ACCT_STR);
+        bits = AOMMIN(bits,
+                      av1_ceil_log2(1 + (1 << cm->bit_depth) -
+                                    pmi->palette_colors[PALETTE_MAX_SIZE + i]));
+      }
+      // V channel colors.
+      if (aom_read_bit(r, ACCT_STR)) {  // Delta encoding.
+        const int min_bits_v = cm->bit_depth - 4;
+        const int max_val = 1 << cm->bit_depth;
+        bits = min_bits_v + aom_read_literal(r, 2, ACCT_STR);
+        pmi->palette_colors[2 * PALETTE_MAX_SIZE] =
+            aom_read_literal(r, cm->bit_depth, ACCT_STR);
+        for (i = 1; i < n; ++i) {
+          int delta = aom_read_literal(r, bits, ACCT_STR);
+          if (delta && aom_read_bit(r, ACCT_STR)) delta = -delta;
+          int val =
+              (int)pmi->palette_colors[2 * PALETTE_MAX_SIZE + i - 1] + delta;
+          if (val < 0) val += max_val;
+          if (val >= max_val) val -= max_val;
+          pmi->palette_colors[2 * PALETTE_MAX_SIZE + i] = val;
+        }
+      } else {
+        for (i = 0; i < n; ++i) {
+          pmi->palette_colors[2 * PALETTE_MAX_SIZE + i] =
+              aom_read_literal(r, cm->bit_depth, ACCT_STR);
+        }
+      }
+#else
+      for (i = 0; i < n; ++i) {
+        pmi->palette_colors[PALETTE_MAX_SIZE + i] =
+            aom_read_literal(r, cm->bit_depth, ACCT_STR);
+        pmi->palette_colors[2 * PALETTE_MAX_SIZE + i] =
+            aom_read_literal(r, cm->bit_depth, ACCT_STR);
+      }
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+      xd->plane[1].color_index_map[0] = read_uniform(r, n);
+      assert(xd->plane[1].color_index_map[0] < n);
+    }
+  }
+}
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_FILTER_INTRA
+static void read_filter_intra_mode_info(AV1_COMMON *const cm,
+                                        MACROBLOCKD *const xd, aom_reader *r) {
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  FRAME_COUNTS *counts = xd->counts;
+  FILTER_INTRA_MODE_INFO *filter_intra_mode_info =
+      &mbmi->filter_intra_mode_info;
+
+  if (mbmi->mode == DC_PRED
+#if CONFIG_PALETTE
+      && mbmi->palette_mode_info.palette_size[0] == 0
+#endif  // CONFIG_PALETTE
+      ) {
+    filter_intra_mode_info->use_filter_intra_mode[0] =
+        aom_read(r, cm->fc->filter_intra_probs[0], ACCT_STR);
+    if (filter_intra_mode_info->use_filter_intra_mode[0]) {
+      filter_intra_mode_info->filter_intra_mode[0] =
+          read_uniform(r, FILTER_INTRA_MODES);
+    }
+    if (counts) {
+      ++counts
+            ->filter_intra[0][filter_intra_mode_info->use_filter_intra_mode[0]];
+    }
+  }
+  if (mbmi->uv_mode == DC_PRED
+#if CONFIG_PALETTE
+      && mbmi->palette_mode_info.palette_size[1] == 0
+#endif  // CONFIG_PALETTE
+      ) {
+    filter_intra_mode_info->use_filter_intra_mode[1] =
+        aom_read(r, cm->fc->filter_intra_probs[1], ACCT_STR);
+    if (filter_intra_mode_info->use_filter_intra_mode[1]) {
+      filter_intra_mode_info->filter_intra_mode[1] =
+          read_uniform(r, FILTER_INTRA_MODES);
+    }
+    if (counts) {
+      ++counts
+            ->filter_intra[1][filter_intra_mode_info->use_filter_intra_mode[1]];
+    }
+  }
+}
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_EXT_INTRA
+static void read_intra_angle_info(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                                  aom_reader *r) {
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+#if CONFIG_INTRA_INTERP
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *const ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *const ec_ctx = cm->fc;
+#endif  // CONFIG_EC_ADAPT
+  const int ctx = av1_get_pred_context_intra_interp(xd);
+  int p_angle;
+#endif  // CONFIG_INTRA_INTERP
+
+  (void)cm;
+  if (bsize < BLOCK_8X8) return;
+
+  if (av1_is_directional_mode(mbmi->mode, bsize)) {
+    mbmi->angle_delta[0] =
+        read_uniform(r, 2 * MAX_ANGLE_DELTA + 1) - MAX_ANGLE_DELTA;
+#if CONFIG_INTRA_INTERP
+    p_angle = mode_to_angle_map[mbmi->mode] + mbmi->angle_delta[0] * ANGLE_STEP;
+    if (av1_is_intra_filter_switchable(p_angle)) {
+      FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_EC_MULTISYMBOL
+      mbmi->intra_filter = aom_read_symbol(r, ec_ctx->intra_filter_cdf[ctx],
+                                           INTRA_FILTERS, ACCT_STR);
+#else
+      mbmi->intra_filter = aom_read_tree(
+          r, av1_intra_filter_tree, ec_ctx->intra_filter_probs[ctx], ACCT_STR);
+#endif  // CONFIG_EC_MULTISYMBOL
+      if (counts) ++counts->intra_filter[ctx][mbmi->intra_filter];
+    } else {
+      mbmi->intra_filter = INTRA_FILTER_LINEAR;
+    }
+#endif  // CONFIG_INTRA_INTERP
+  }
+
+  if (av1_is_directional_mode(mbmi->uv_mode, bsize)) {
+    mbmi->angle_delta[1] =
+        read_uniform(r, 2 * MAX_ANGLE_DELTA + 1) - MAX_ANGLE_DELTA;
+  }
+}
+#endif  // CONFIG_EXT_INTRA
+
+void av1_read_tx_type(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                      int supertx_enabled,
+#endif
+#if CONFIG_TXK_SEL
+                      int block, int plane,
+#endif
+                      aom_reader *r) {
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const int inter_block = is_inter_block(mbmi);
+#if CONFIG_VAR_TX
+  const TX_SIZE tx_size = inter_block ? mbmi->min_tx_size : mbmi->tx_size;
+#else
+  const TX_SIZE tx_size = mbmi->tx_size;
+#endif
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+#if !CONFIG_TXK_SEL
+  TX_TYPE *tx_type = &mbmi->tx_type;
+#else
+  // only y plane's tx_type is transmitted
+  if (plane > 0) return;
+  TX_TYPE *tx_type = &mbmi->txk_type[block];
+#endif
+
+  if (!FIXED_TX_TYPE) {
+#if CONFIG_EXT_TX
+    const TX_SIZE square_tx_size = txsize_sqr_map[tx_size];
+    if (get_ext_tx_types(tx_size, mbmi->sb_type, inter_block,
+                         cm->reduced_tx_set_used) > 1 &&
+        ((!cm->seg.enabled && cm->base_qindex > 0) ||
+         (cm->seg.enabled && xd->qindex[mbmi->segment_id] > 0)) &&
+        !mbmi->skip &&
+#if CONFIG_SUPERTX
+        !supertx_enabled &&
+#endif  // CONFIG_SUPERTX
+        !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+      const int eset = get_ext_tx_set(tx_size, mbmi->sb_type, inter_block,
+                                      cm->reduced_tx_set_used);
+      FRAME_COUNTS *counts = xd->counts;
+
+      if (inter_block) {
+        if (eset > 0) {
+#if CONFIG_EC_MULTISYMBOL
+          *tx_type = av1_ext_tx_inter_inv[eset][aom_read_symbol(
+              r, ec_ctx->inter_ext_tx_cdf[eset][square_tx_size],
+              ext_tx_cnt_inter[eset], ACCT_STR)];
+#else
+          *tx_type = aom_read_tree(
+              r, av1_ext_tx_inter_tree[eset],
+              ec_ctx->inter_ext_tx_prob[eset][square_tx_size], ACCT_STR);
+#endif
+          if (counts) ++counts->inter_ext_tx[eset][square_tx_size][*tx_type];
+        }
+      } else if (ALLOW_INTRA_EXT_TX) {
+        if (eset > 0) {
+#if CONFIG_EC_MULTISYMBOL
+          *tx_type = av1_ext_tx_intra_inv[eset][aom_read_symbol(
+              r, ec_ctx->intra_ext_tx_cdf[eset][square_tx_size][mbmi->mode],
+              ext_tx_cnt_intra[eset], ACCT_STR)];
+#else
+          *tx_type = aom_read_tree(
+              r, av1_ext_tx_intra_tree[eset],
+              ec_ctx->intra_ext_tx_prob[eset][square_tx_size][mbmi->mode],
+              ACCT_STR);
+#endif
+          if (counts)
+            ++counts->intra_ext_tx[eset][square_tx_size][mbmi->mode][*tx_type];
+        }
+      }
+    } else {
+      *tx_type = DCT_DCT;
+    }
+#else
+
+    if (tx_size < TX_32X32 &&
+        ((!cm->seg.enabled && cm->base_qindex > 0) ||
+         (cm->seg.enabled && xd->qindex[mbmi->segment_id] > 0)) &&
+        !mbmi->skip &&
+#if CONFIG_SUPERTX
+        !supertx_enabled &&
+#endif  // CONFIG_SUPERTX
+        !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+      FRAME_COUNTS *counts = xd->counts;
+
+      if (inter_block) {
+#if CONFIG_EC_MULTISYMBOL
+        *tx_type = av1_ext_tx_inv[aom_read_symbol(
+            r, ec_ctx->inter_ext_tx_cdf[tx_size], TX_TYPES, ACCT_STR)];
+#else
+        *tx_type = aom_read_tree(r, av1_ext_tx_tree,
+                                 ec_ctx->inter_ext_tx_prob[tx_size], ACCT_STR);
+#endif
+        if (counts) ++counts->inter_ext_tx[tx_size][*tx_type];
+      } else {
+        const TX_TYPE tx_type_nom = intra_mode_to_tx_type_context[mbmi->mode];
+#if CONFIG_EC_MULTISYMBOL
+        *tx_type = av1_ext_tx_inv[aom_read_symbol(
+            r, ec_ctx->intra_ext_tx_cdf[tx_size][tx_type_nom], TX_TYPES,
+            ACCT_STR)];
+#else
+        *tx_type = aom_read_tree(
+            r, av1_ext_tx_tree, ec_ctx->intra_ext_tx_prob[tx_size][tx_type_nom],
+            ACCT_STR);
+#endif
+        if (counts) ++counts->intra_ext_tx[tx_size][tx_type_nom][*tx_type];
+      }
+    } else {
+      *tx_type = DCT_DCT;
+    }
+#endif  // CONFIG_EXT_TX
+  }
+}
+
+#if CONFIG_INTRABC
+static INLINE void read_mv(aom_reader *r, MV *mv, const MV *ref,
+                           nmv_context *ctx, nmv_context_counts *counts,
+                           int allow_hp);
+
+static INLINE int is_mv_valid(const MV *mv);
+
+static INLINE int assign_dv(AV1_COMMON *cm, MACROBLOCKD *xd, int_mv *mv,
+                            const int_mv *ref_mv, int mi_row, int mi_col,
+                            BLOCK_SIZE bsize, aom_reader *r) {
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+  FRAME_COUNTS *counts = xd->counts;
+  nmv_context_counts *const dv_counts = counts ? &counts->dv : NULL;
+  read_mv(r, &mv->as_mv, &ref_mv->as_mv, &ec_ctx->ndvc, dv_counts, 0);
+  int valid = is_mv_valid(&mv->as_mv) &&
+              is_dv_valid(mv->as_mv, &xd->tile, mi_row, mi_col, bsize);
+  // TODO(aconverse@google.com): additional validation
+  return valid;
+}
+#endif  // CONFIG_INTRABC
+
+static void read_intra_frame_mode_info(AV1_COMMON *const cm,
+                                       MACROBLOCKD *const xd, int mi_row,
+                                       int mi_col, aom_reader *r) {
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const MODE_INFO *above_mi = xd->above_mi;
+  const MODE_INFO *left_mi = xd->left_mi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  int i;
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+
+  // TODO(slavarnway): move x_mis, y_mis into xd ?????
+  const int x_mis = AOMMIN(cm->mi_cols - mi_col, bw);
+  const int y_mis = AOMMIN(cm->mi_rows - mi_row, bh);
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  mbmi->segment_id = read_intra_segment_id(cm, xd, mi_offset, x_mis, y_mis, r);
+  mbmi->skip = read_skip(cm, xd, mbmi->segment_id, r);
+
+#if CONFIG_DELTA_Q
+  if (cm->delta_q_present_flag) {
+    xd->current_qindex =
+        xd->prev_qindex +
+        read_delta_qindex(cm, xd, r, mbmi, mi_col, mi_row) * cm->delta_q_res;
+    /* Normative: Clamp to [1,MAXQ] to not interfere with lossless mode */
+    xd->current_qindex = clamp(xd->current_qindex, 1, MAXQ);
+    xd->prev_qindex = xd->current_qindex;
+#if CONFIG_EXT_DELTA_Q
+    if (cm->delta_lf_present_flag) {
+      mbmi->current_delta_lf_from_base = xd->current_delta_lf_from_base =
+          xd->prev_delta_lf_from_base +
+          read_delta_lflevel(cm, xd, r, mbmi, mi_col, mi_row) *
+              cm->delta_lf_res;
+      xd->prev_delta_lf_from_base = xd->current_delta_lf_from_base;
+    }
+#endif
+  }
+#endif
+
+  mbmi->tx_size = read_tx_size(cm, xd, 0, 1, r);
+  mbmi->ref_frame[0] = INTRA_FRAME;
+  mbmi->ref_frame[1] = NONE_FRAME;
+
+#if CONFIG_INTRABC
+  if (bsize >= BLOCK_8X8 && cm->allow_screen_content_tools) {
+    mbmi->use_intrabc = aom_read(r, INTRABC_PROB, ACCT_STR);
+    if (mbmi->use_intrabc) {
+      int_mv dv_ref;
+      mbmi->mode = mbmi->uv_mode = DC_PRED;
+#if CONFIG_DUAL_FILTER
+      for (int idx = 0; idx < 4; ++idx) mbmi->interp_filter[idx] = BILINEAR;
+#else
+      mbmi->interp_filter = BILINEAR;
+#endif
+      av1_find_ref_dv(&dv_ref, mi_row, mi_col);
+      xd->corrupted |=
+          !assign_dv(cm, xd, &mbmi->mv[0], &dv_ref, mi_row, mi_col, bsize, r);
+      return;
+    }
+  }
+#endif  // CONFIG_INTRABC
+
+#if CONFIG_CB4X4
+  (void)i;
+  mbmi->mode =
+#if CONFIG_EC_MULTISYMBOL
+      read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 0));
+#else
+      read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 0));
+#endif
+#else
+  switch (bsize) {
+    case BLOCK_4X4:
+      for (i = 0; i < 4; ++i)
+        mi->bmi[i].as_mode =
+#if CONFIG_EC_MULTISYMBOL
+            read_intra_mode(r,
+                            get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, i));
+#else
+            read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, i));
+#endif
+      mbmi->mode = mi->bmi[3].as_mode;
+      break;
+    case BLOCK_4X8:
+      mi->bmi[0].as_mode = mi->bmi[2].as_mode =
+#if CONFIG_EC_MULTISYMBOL
+          read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 0));
+#else
+          read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 0));
+#endif
+      mi->bmi[1].as_mode = mi->bmi[3].as_mode = mbmi->mode =
+#if CONFIG_EC_MULTISYMBOL
+          read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 1));
+#else
+          read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 1));
+#endif
+      break;
+    case BLOCK_8X4:
+      mi->bmi[0].as_mode = mi->bmi[1].as_mode =
+#if CONFIG_EC_MULTISYMBOL
+          read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 0));
+#else
+          read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 0));
+#endif
+      mi->bmi[2].as_mode = mi->bmi[3].as_mode = mbmi->mode =
+#if CONFIG_EC_MULTISYMBOL
+          read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 2));
+#else
+          read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 2));
+#endif
+      break;
+    default:
+      mbmi->mode =
+#if CONFIG_EC_MULTISYMBOL
+          read_intra_mode(r, get_y_mode_cdf(ec_ctx, mi, above_mi, left_mi, 0));
+#else
+          read_intra_mode(r, get_y_mode_probs(cm, mi, above_mi, left_mi, 0));
+#endif
+  }
+#endif
+
+#if CONFIG_CB4X4
+  if (is_chroma_reference(mi_row, mi_col, bsize, xd->plane[1].subsampling_x,
+                          xd->plane[1].subsampling_y))
+    mbmi->uv_mode = read_intra_mode_uv(cm, xd, r, mbmi->mode);
+#else
+  mbmi->uv_mode = read_intra_mode_uv(cm, xd, r, mbmi->mode);
+#endif
+
+#if CONFIG_EXT_INTRA
+  read_intra_angle_info(cm, xd, r);
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_PALETTE
+  mbmi->palette_mode_info.palette_size[0] = 0;
+  mbmi->palette_mode_info.palette_size[1] = 0;
+  if (bsize >= BLOCK_8X8 && cm->allow_screen_content_tools)
+    read_palette_mode_info(cm, xd, r);
+#endif  // CONFIG_PALETTE
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+  if (bsize >= BLOCK_8X8 || CONFIG_CB4X4)
+    read_filter_intra_mode_info(cm, xd, r);
+#endif  // CONFIG_FILTER_INTRA
+
+#if !CONFIG_TXK_SEL
+  av1_read_tx_type(cm, xd,
+#if CONFIG_SUPERTX
+                   0,
+#endif
+                   r);
+#endif  // !CONFIG_TXK_SEL
+}
+
+static int read_mv_component(aom_reader *r, nmv_component *mvcomp, int usehp) {
+  int mag, d, fr, hp;
+  const int sign = aom_read(r, mvcomp->sign, ACCT_STR);
+  const int mv_class =
+#if CONFIG_EC_MULTISYMBOL
+      aom_read_symbol(r, mvcomp->class_cdf, MV_CLASSES, ACCT_STR);
+#else
+      aom_read_tree(r, av1_mv_class_tree, mvcomp->classes, ACCT_STR);
+#endif
+  const int class0 = mv_class == MV_CLASS_0;
+
+  // Integer part
+  if (class0) {
+    d = aom_read(r, mvcomp->class0[0], ACCT_STR);
+    mag = 0;
+  } else {
+    int i;
+    const int n = mv_class + CLASS0_BITS - 1;  // number of bits
+
+    d = 0;
+    for (i = 0; i < n; ++i) d |= aom_read(r, mvcomp->bits[i], ACCT_STR) << i;
+    mag = CLASS0_SIZE << (mv_class + 2);
+  }
+
+// Fractional part
+#if CONFIG_EC_MULTISYMBOL
+  fr = aom_read_symbol(r, class0 ? mvcomp->class0_fp_cdf[d] : mvcomp->fp_cdf,
+                       MV_FP_SIZE, ACCT_STR);
+#else
+  fr = aom_read_tree(r, av1_mv_fp_tree,
+                     class0 ? mvcomp->class0_fp[d] : mvcomp->fp, ACCT_STR);
+#endif
+
+  // High precision part (if hp is not used, the default value of the hp is 1)
+  hp = usehp ? aom_read(r, class0 ? mvcomp->class0_hp : mvcomp->hp, ACCT_STR)
+             : 1;
+
+  // Result
+  mag += ((d << 3) | (fr << 1) | hp) + 1;
+  return sign ? -mag : mag;
+}
+
+static INLINE void read_mv(aom_reader *r, MV *mv, const MV *ref,
+                           nmv_context *ctx, nmv_context_counts *counts,
+                           int allow_hp) {
+  MV_JOINT_TYPE joint_type;
+  MV diff = { 0, 0 };
+  joint_type =
+#if CONFIG_EC_MULTISYMBOL
+      (MV_JOINT_TYPE)aom_read_symbol(r, ctx->joint_cdf, MV_JOINTS, ACCT_STR);
+#else
+      (MV_JOINT_TYPE)aom_read_tree(r, av1_mv_joint_tree, ctx->joints, ACCT_STR);
+#endif
+
+  if (mv_joint_vertical(joint_type))
+    diff.row = read_mv_component(r, &ctx->comps[0], allow_hp);
+
+  if (mv_joint_horizontal(joint_type))
+    diff.col = read_mv_component(r, &ctx->comps[1], allow_hp);
+
+  av1_inc_mv(&diff, counts, allow_hp);
+
+  mv->row = ref->row + diff.row;
+  mv->col = ref->col + diff.col;
+}
+
+static REFERENCE_MODE read_block_reference_mode(AV1_COMMON *cm,
+                                                const MACROBLOCKD *xd,
+                                                aom_reader *r) {
+#if !SUB8X8_COMP_REF
+  if (xd->mi[0]->mbmi.sb_type < BLOCK_8X8) return SINGLE_REFERENCE;
+#endif
+  if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+    const int ctx = av1_get_reference_mode_context(cm, xd);
+    const REFERENCE_MODE mode =
+        (REFERENCE_MODE)aom_read(r, cm->fc->comp_inter_prob[ctx], ACCT_STR);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts) ++counts->comp_inter[ctx][mode];
+    return mode;  // SINGLE_REFERENCE or COMPOUND_REFERENCE
+  } else {
+    return cm->reference_mode;
+  }
+}
+
+// Read the referncence frame
+static void read_ref_frames(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                            aom_reader *r, int segment_id,
+                            MV_REFERENCE_FRAME ref_frame[2]) {
+  FRAME_CONTEXT *const fc = cm->fc;
+  FRAME_COUNTS *counts = xd->counts;
+
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
+    ref_frame[0] = (MV_REFERENCE_FRAME)get_segdata(&cm->seg, segment_id,
+                                                   SEG_LVL_REF_FRAME);
+    ref_frame[1] = NONE_FRAME;
+  } else {
+    const REFERENCE_MODE mode = read_block_reference_mode(cm, xd, r);
+    // FIXME(rbultje) I'm pretty sure this breaks segmentation ref frame coding
+    if (mode == COMPOUND_REFERENCE) {
+#if CONFIG_LOWDELAY_COMPOUND  // Normative in decoder (for low delay)
+      const int idx = 1;
+#else
+#if CONFIG_EXT_REFS
+      const int idx = cm->ref_frame_sign_bias[cm->comp_bwd_ref[0]];
+#else
+      const int idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
+#endif  // CONFIG_EXT_REFS
+#endif
+      const int ctx = av1_get_pred_context_comp_ref_p(cm, xd);
+
+      const int bit = aom_read(r, fc->comp_ref_prob[ctx][0], ACCT_STR);
+      if (counts) ++counts->comp_ref[ctx][0][bit];
+
+#if CONFIG_EXT_REFS
+      // Decode forward references.
+      if (!bit) {
+        const int ctx1 = av1_get_pred_context_comp_ref_p1(cm, xd);
+        const int bit1 = aom_read(r, fc->comp_ref_prob[ctx1][1], ACCT_STR);
+        if (counts) ++counts->comp_ref[ctx1][1][bit1];
+        ref_frame[!idx] = cm->comp_fwd_ref[bit1 ? 0 : 1];
+      } else {
+        const int ctx2 = av1_get_pred_context_comp_ref_p2(cm, xd);
+        const int bit2 = aom_read(r, fc->comp_ref_prob[ctx2][2], ACCT_STR);
+        if (counts) ++counts->comp_ref[ctx2][2][bit2];
+        ref_frame[!idx] = cm->comp_fwd_ref[bit2 ? 3 : 2];
+      }
+
+      // Decode backward references.
+      {
+        const int ctx_bwd = av1_get_pred_context_comp_bwdref_p(cm, xd);
+        const int bit_bwd =
+            aom_read(r, fc->comp_bwdref_prob[ctx_bwd][0], ACCT_STR);
+        if (counts) ++counts->comp_bwdref[ctx_bwd][0][bit_bwd];
+        ref_frame[idx] = cm->comp_bwd_ref[bit_bwd];
+      }
+#else
+      ref_frame[!idx] = cm->comp_var_ref[bit];
+      ref_frame[idx] = cm->comp_fixed_ref;
+#endif  // CONFIG_EXT_REFS
+    } else if (mode == SINGLE_REFERENCE) {
+#if CONFIG_EXT_REFS
+      const int ctx0 = av1_get_pred_context_single_ref_p1(xd);
+      const int bit0 = aom_read(r, fc->single_ref_prob[ctx0][0], ACCT_STR);
+      if (counts) ++counts->single_ref[ctx0][0][bit0];
+
+      if (bit0) {
+        const int ctx1 = av1_get_pred_context_single_ref_p2(xd);
+        const int bit1 = aom_read(r, fc->single_ref_prob[ctx1][1], ACCT_STR);
+        if (counts) ++counts->single_ref[ctx1][1][bit1];
+        ref_frame[0] = bit1 ? ALTREF_FRAME : BWDREF_FRAME;
+      } else {
+        const int ctx2 = av1_get_pred_context_single_ref_p3(xd);
+        const int bit2 = aom_read(r, fc->single_ref_prob[ctx2][2], ACCT_STR);
+        if (counts) ++counts->single_ref[ctx2][2][bit2];
+        if (bit2) {
+          const int ctx4 = av1_get_pred_context_single_ref_p5(xd);
+          const int bit4 = aom_read(r, fc->single_ref_prob[ctx4][4], ACCT_STR);
+          if (counts) ++counts->single_ref[ctx4][4][bit4];
+          ref_frame[0] = bit4 ? GOLDEN_FRAME : LAST3_FRAME;
+        } else {
+          const int ctx3 = av1_get_pred_context_single_ref_p4(xd);
+          const int bit3 = aom_read(r, fc->single_ref_prob[ctx3][3], ACCT_STR);
+          if (counts) ++counts->single_ref[ctx3][3][bit3];
+          ref_frame[0] = bit3 ? LAST2_FRAME : LAST_FRAME;
+        }
+      }
+#else
+      const int ctx0 = av1_get_pred_context_single_ref_p1(xd);
+      const int bit0 = aom_read(r, fc->single_ref_prob[ctx0][0], ACCT_STR);
+      if (counts) ++counts->single_ref[ctx0][0][bit0];
+
+      if (bit0) {
+        const int ctx1 = av1_get_pred_context_single_ref_p2(xd);
+        const int bit1 = aom_read(r, fc->single_ref_prob[ctx1][1], ACCT_STR);
+        if (counts) ++counts->single_ref[ctx1][1][bit1];
+        ref_frame[0] = bit1 ? ALTREF_FRAME : GOLDEN_FRAME;
+      } else {
+        ref_frame[0] = LAST_FRAME;
+      }
+#endif  // CONFIG_EXT_REFS
+
+      ref_frame[1] = NONE_FRAME;
+    } else {
+      assert(0 && "Invalid prediction mode.");
+    }
+  }
+}
+
+static INLINE void read_mb_interp_filter(AV1_COMMON *const cm,
+                                         MACROBLOCKD *const xd,
+                                         MB_MODE_INFO *const mbmi,
+                                         aom_reader *r) {
+  FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  if (!av1_is_interp_needed(xd)) {
+    set_default_interp_filters(mbmi, cm->interp_filter);
+    return;
+  }
+
+#if CONFIG_DUAL_FILTER
+  if (cm->interp_filter != SWITCHABLE) {
+    int dir;
+
+    for (dir = 0; dir < 4; ++dir) mbmi->interp_filter[dir] = cm->interp_filter;
+  } else {
+    int dir;
+
+    for (dir = 0; dir < 2; ++dir) {
+      const int ctx = av1_get_pred_context_switchable_interp(xd, dir);
+      mbmi->interp_filter[dir] = EIGHTTAP_REGULAR;
+
+      if (has_subpel_mv_component(xd->mi[0], xd, dir) ||
+          (mbmi->ref_frame[1] > INTRA_FRAME &&
+           has_subpel_mv_component(xd->mi[0], xd, dir + 2))) {
+#if CONFIG_EC_MULTISYMBOL
+        mbmi->interp_filter[dir] =
+            (InterpFilter)av1_switchable_interp_inv[aom_read_symbol(
+                r, ec_ctx->switchable_interp_cdf[ctx], SWITCHABLE_FILTERS,
+                ACCT_STR)];
+#else
+        mbmi->interp_filter[dir] = (InterpFilter)aom_read_tree(
+            r, av1_switchable_interp_tree, ec_ctx->switchable_interp_prob[ctx],
+            ACCT_STR);
+#endif
+        if (counts) ++counts->switchable_interp[ctx][mbmi->interp_filter[dir]];
+      }
+    }
+    // The index system works as:
+    // (0, 1) -> (vertical, horizontal) filter types for the first ref frame.
+    // (2, 3) -> (vertical, horizontal) filter types for the second ref frame.
+    mbmi->interp_filter[2] = mbmi->interp_filter[0];
+    mbmi->interp_filter[3] = mbmi->interp_filter[1];
+  }
+#else  // CONFIG_DUAL_FILTER
+  if (cm->interp_filter != SWITCHABLE) {
+    mbmi->interp_filter = cm->interp_filter;
+  } else {
+    const int ctx = av1_get_pred_context_switchable_interp(xd);
+#if CONFIG_EC_MULTISYMBOL
+    mbmi->interp_filter =
+        (InterpFilter)av1_switchable_interp_inv[aom_read_symbol(
+            r, ec_ctx->switchable_interp_cdf[ctx], SWITCHABLE_FILTERS,
+            ACCT_STR)];
+#else
+    mbmi->interp_filter = (InterpFilter)aom_read_tree(
+        r, av1_switchable_interp_tree, ec_ctx->switchable_interp_prob[ctx],
+        ACCT_STR);
+#endif
+    if (counts) ++counts->switchable_interp[ctx][mbmi->interp_filter];
+  }
+#endif  // CONFIG_DUAL_FILTER
+}
+
+static void read_intra_block_mode_info(AV1_COMMON *const cm, const int mi_row,
+                                       const int mi_col, MACROBLOCKD *const xd,
+                                       MODE_INFO *mi, aom_reader *r) {
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const BLOCK_SIZE bsize = mi->mbmi.sb_type;
+  int i;
+
+  mbmi->ref_frame[0] = INTRA_FRAME;
+  mbmi->ref_frame[1] = NONE_FRAME;
+
+#if CONFIG_CB4X4
+  (void)i;
+  mbmi->mode = read_intra_mode_y(cm, xd, r, size_group_lookup[bsize]);
+#else
+  switch (bsize) {
+    case BLOCK_4X4:
+      for (i = 0; i < 4; ++i)
+        mi->bmi[i].as_mode = read_intra_mode_y(cm, xd, r, 0);
+      mbmi->mode = mi->bmi[3].as_mode;
+      break;
+    case BLOCK_4X8:
+      mi->bmi[0].as_mode = mi->bmi[2].as_mode = read_intra_mode_y(cm, xd, r, 0);
+      mi->bmi[1].as_mode = mi->bmi[3].as_mode = mbmi->mode =
+          read_intra_mode_y(cm, xd, r, 0);
+      break;
+    case BLOCK_8X4:
+      mi->bmi[0].as_mode = mi->bmi[1].as_mode = read_intra_mode_y(cm, xd, r, 0);
+      mi->bmi[2].as_mode = mi->bmi[3].as_mode = mbmi->mode =
+          read_intra_mode_y(cm, xd, r, 0);
+      break;
+    default:
+      mbmi->mode = read_intra_mode_y(cm, xd, r, size_group_lookup[bsize]);
+  }
+#endif
+
+#if CONFIG_CB4X4
+  if (is_chroma_reference(mi_row, mi_col, bsize, xd->plane[1].subsampling_x,
+                          xd->plane[1].subsampling_y))
+    mbmi->uv_mode = read_intra_mode_uv(cm, xd, r, mbmi->mode);
+#else
+  mbmi->uv_mode = read_intra_mode_uv(cm, xd, r, mbmi->mode);
+  (void)mi_row;
+  (void)mi_col;
+#endif
+
+#if CONFIG_EXT_INTRA
+  read_intra_angle_info(cm, xd, r);
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_PALETTE
+  mbmi->palette_mode_info.palette_size[0] = 0;
+  mbmi->palette_mode_info.palette_size[1] = 0;
+  if (bsize >= BLOCK_8X8 && cm->allow_screen_content_tools)
+    read_palette_mode_info(cm, xd, r);
+#endif  // CONFIG_PALETTE
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+  if (bsize >= BLOCK_8X8 || CONFIG_CB4X4)
+    read_filter_intra_mode_info(cm, xd, r);
+#endif  // CONFIG_FILTER_INTRA
+}
+
+static INLINE int is_mv_valid(const MV *mv) {
+  return mv->row > MV_LOW && mv->row < MV_UPP && mv->col > MV_LOW &&
+         mv->col < MV_UPP;
+}
+
+static INLINE int assign_mv(AV1_COMMON *cm, MACROBLOCKD *xd,
+                            PREDICTION_MODE mode,
+                            MV_REFERENCE_FRAME ref_frame[2], int block,
+                            int_mv mv[2], int_mv ref_mv[2],
+                            int_mv nearest_mv[2], int_mv near_mv[2], int mi_row,
+                            int mi_col, int is_compound, int allow_hp,
+                            aom_reader *r) {
+  int i;
+  int ret = 1;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+  BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+#if CONFIG_REF_MV
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+#if CONFIG_CB4X4
+  int_mv *pred_mv = mbmi->pred_mv;
+  (void)block;
+#else
+  int_mv *pred_mv =
+      (bsize >= BLOCK_8X8) ? mbmi->pred_mv : xd->mi[0]->bmi[block].pred_mv;
+#endif  // CONFIG_CB4X4
+#else
+  (void)block;
+#endif  // CONFIG_REF_MV
+  (void)ref_frame;
+  (void)cm;
+  (void)mi_row;
+  (void)mi_col;
+  (void)bsize;
+
+  switch (mode) {
+    case NEWMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if !CONFIG_REF_MV
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      for (i = 0; i < 1 + is_compound; ++i) {
+#if CONFIG_REF_MV
+        int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+        int nmv_ctx =
+            av1_nmv_ctx(xd->ref_mv_count[rf_type], xd->ref_mv_stack[rf_type], i,
+                        mbmi->ref_mv_idx);
+        nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+        nmv_context_counts *const mv_counts =
+            counts ? &counts->mv[nmv_ctx] : NULL;
+#endif
+        read_mv(r, &mv[i].as_mv, &ref_mv[i].as_mv, nmvc, mv_counts, allow_hp);
+        ret = ret && is_mv_valid(&mv[i].as_mv);
+
+#if CONFIG_REF_MV
+        pred_mv[i].as_int = ref_mv[i].as_int;
+#endif
+      }
+      break;
+    }
+    case NEARESTMV: {
+      mv[0].as_int = nearest_mv[0].as_int;
+      if (is_compound) mv[1].as_int = nearest_mv[1].as_int;
+
+#if CONFIG_REF_MV
+      pred_mv[0].as_int = nearest_mv[0].as_int;
+      if (is_compound) pred_mv[1].as_int = nearest_mv[1].as_int;
+#endif
+      break;
+    }
+    case NEARMV: {
+      mv[0].as_int = near_mv[0].as_int;
+      if (is_compound) mv[1].as_int = near_mv[1].as_int;
+
+#if CONFIG_REF_MV
+      pred_mv[0].as_int = near_mv[0].as_int;
+      if (is_compound) pred_mv[1].as_int = near_mv[1].as_int;
+#endif
+      break;
+    }
+    case ZEROMV: {
+#if CONFIG_GLOBAL_MOTION
+      mv[0].as_int = gm_get_motion_vector(&cm->global_motion[ref_frame[0]],
+                                          cm->allow_high_precision_mv, bsize,
+                                          mi_col, mi_row, block)
+                         .as_int;
+      if (is_compound)
+        mv[1].as_int = gm_get_motion_vector(&cm->global_motion[ref_frame[1]],
+                                            cm->allow_high_precision_mv, bsize,
+                                            mi_col, mi_row, block)
+                           .as_int;
+#else
+      mv[0].as_int = 0;
+      if (is_compound) mv[1].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+
+#if CONFIG_REF_MV
+      pred_mv[0].as_int = mv[0].as_int;
+      if (is_compound) pred_mv[1].as_int = mv[1].as_int;
+#endif
+      break;
+    }
+#if CONFIG_EXT_INTER
+    case NEW_NEWMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if !CONFIG_REF_MV
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      assert(is_compound);
+      for (i = 0; i < 2; ++i) {
+#if CONFIG_REF_MV
+        int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+        int nmv_ctx =
+            av1_nmv_ctx(xd->ref_mv_count[rf_type], xd->ref_mv_stack[rf_type], i,
+                        mbmi->ref_mv_idx);
+        nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+        nmv_context_counts *const mv_counts =
+            counts ? &counts->mv[nmv_ctx] : NULL;
+#endif
+        read_mv(r, &mv[i].as_mv, &ref_mv[i].as_mv, nmvc, mv_counts, allow_hp);
+        ret = ret && is_mv_valid(&mv[i].as_mv);
+      }
+      break;
+    }
+    case NEAREST_NEARESTMV: {
+      assert(is_compound);
+      mv[0].as_int = nearest_mv[0].as_int;
+      mv[1].as_int = nearest_mv[1].as_int;
+      break;
+    }
+    case NEAREST_NEARMV: {
+      assert(is_compound);
+      mv[0].as_int = nearest_mv[0].as_int;
+      mv[1].as_int = near_mv[1].as_int;
+      break;
+    }
+    case NEAR_NEARESTMV: {
+      assert(is_compound);
+      mv[0].as_int = near_mv[0].as_int;
+      mv[1].as_int = nearest_mv[1].as_int;
+      break;
+    }
+    case NEAR_NEARMV: {
+      assert(is_compound);
+      mv[0].as_int = near_mv[0].as_int;
+      mv[1].as_int = near_mv[1].as_int;
+      break;
+    }
+    case NEW_NEARESTMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx = av1_nmv_ctx(xd->ref_mv_count[rf_type],
+                                xd->ref_mv_stack[rf_type], 0, mbmi->ref_mv_idx);
+      nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+      nmv_context_counts *const mv_counts =
+          counts ? &counts->mv[nmv_ctx] : NULL;
+#else
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      read_mv(r, &mv[0].as_mv, &ref_mv[0].as_mv, nmvc, mv_counts, allow_hp);
+      assert(is_compound);
+      ret = ret && is_mv_valid(&mv[0].as_mv);
+      mv[1].as_int = nearest_mv[1].as_int;
+      break;
+    }
+    case NEAREST_NEWMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx = av1_nmv_ctx(xd->ref_mv_count[rf_type],
+                                xd->ref_mv_stack[rf_type], 1, mbmi->ref_mv_idx);
+      nmv_context_counts *const mv_counts =
+          counts ? &counts->mv[nmv_ctx] : NULL;
+      nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+#else
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      mv[0].as_int = nearest_mv[0].as_int;
+      read_mv(r, &mv[1].as_mv, &ref_mv[1].as_mv, nmvc, mv_counts, allow_hp);
+      assert(is_compound);
+      ret = ret && is_mv_valid(&mv[1].as_mv);
+      break;
+    }
+    case NEAR_NEWMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx = av1_nmv_ctx(xd->ref_mv_count[rf_type],
+                                xd->ref_mv_stack[rf_type], 1, mbmi->ref_mv_idx);
+      nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+      nmv_context_counts *const mv_counts =
+          counts ? &counts->mv[nmv_ctx] : NULL;
+#else
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      mv[0].as_int = near_mv[0].as_int;
+      read_mv(r, &mv[1].as_mv, &ref_mv[1].as_mv, nmvc, mv_counts, allow_hp);
+      assert(is_compound);
+
+      ret = ret && is_mv_valid(&mv[1].as_mv);
+      break;
+    }
+    case NEW_NEARMV: {
+      FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx = av1_nmv_ctx(xd->ref_mv_count[rf_type],
+                                xd->ref_mv_stack[rf_type], 0, mbmi->ref_mv_idx);
+      nmv_context *const nmvc = &ec_ctx->nmvc[nmv_ctx];
+      nmv_context_counts *const mv_counts =
+          counts ? &counts->mv[nmv_ctx] : NULL;
+#else
+      nmv_context *const nmvc = &ec_ctx->nmvc;
+      nmv_context_counts *const mv_counts = counts ? &counts->mv : NULL;
+#endif
+      read_mv(r, &mv[0].as_mv, &ref_mv[0].as_mv, nmvc, mv_counts, allow_hp);
+      assert(is_compound);
+      ret = ret && is_mv_valid(&mv[0].as_mv);
+      mv[1].as_int = near_mv[1].as_int;
+      break;
+    }
+    case ZERO_ZEROMV: {
+      assert(is_compound);
+#if CONFIG_GLOBAL_MOTION
+      mv[0].as_int = gm_get_motion_vector(&cm->global_motion[ref_frame[0]],
+                                          cm->allow_high_precision_mv, bsize,
+                                          mi_col, mi_row, block)
+                         .as_int;
+      mv[1].as_int = gm_get_motion_vector(&cm->global_motion[ref_frame[1]],
+                                          cm->allow_high_precision_mv, bsize,
+                                          mi_col, mi_row, block)
+                         .as_int;
+#else
+      mv[0].as_int = 0;
+      mv[1].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+      break;
+    }
+#endif  // CONFIG_EXT_INTER
+    default: { return 0; }
+  }
+  return ret;
+}
+
+static int read_is_inter_block(AV1_COMMON *const cm, MACROBLOCKD *const xd,
+                               int segment_id, aom_reader *r) {
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
+    return get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME) != INTRA_FRAME;
+  } else {
+    const int ctx = av1_get_intra_inter_context(xd);
+    const int is_inter = aom_read(r, cm->fc->intra_inter_prob[ctx], ACCT_STR);
+    FRAME_COUNTS *counts = xd->counts;
+    if (counts) ++counts->intra_inter[ctx][is_inter];
+    return is_inter;
+  }
+}
+
+static void fpm_sync(void *const data, int mi_row) {
+  AV1Decoder *const pbi = (AV1Decoder *)data;
+  av1_frameworker_wait(pbi->frame_worker_owner, pbi->common.prev_frame,
+                       mi_row << pbi->common.mib_size_log2);
+}
+
+static void read_inter_block_mode_info(AV1Decoder *const pbi,
+                                       MACROBLOCKD *const xd,
+                                       MODE_INFO *const mi,
+#if (CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION || CONFIG_EXT_INTER) && \
+    CONFIG_SUPERTX
+                                       int mi_row, int mi_col, aom_reader *r,
+                                       int supertx_enabled) {
+#else
+                                       int mi_row, int mi_col, aom_reader *r) {
+#endif  // CONFIG_MOTION_VAR && CONFIG_SUPERTX
+  AV1_COMMON *const cm = &pbi->common;
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  const int allow_hp = cm->allow_high_precision_mv;
+  const int unify_bsize = CONFIG_CB4X4;
+  int_mv nearestmv[2], nearmv[2];
+  int_mv ref_mvs[MODE_CTX_REF_FRAMES][MAX_MV_REF_CANDIDATES];
+  int ref, is_compound;
+  int16_t inter_mode_ctx[MODE_CTX_REF_FRAMES];
+#if CONFIG_REF_MV && CONFIG_EXT_INTER
+  int16_t compound_inter_mode_ctx[MODE_CTX_REF_FRAMES];
+#endif  // CONFIG_REF_MV && CONFIG_EXT_INTER
+  int16_t mode_ctx = 0;
+#if CONFIG_WARPED_MOTION
+  int pts[SAMPLES_ARRAY_SIZE], pts_inref[SAMPLES_ARRAY_SIZE];
+#endif  // CONFIG_WARPED_MOTION
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+#if CONFIG_PALETTE
+  mbmi->palette_mode_info.palette_size[0] = 0;
+  mbmi->palette_mode_info.palette_size[1] = 0;
+#endif  // CONFIG_PALETTE
+
+  memset(ref_mvs, 0, sizeof(ref_mvs));
+
+  read_ref_frames(cm, xd, r, mbmi->segment_id, mbmi->ref_frame);
+  is_compound = has_second_ref(mbmi);
+
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
+
+    av1_find_mv_refs(cm, xd, mi, frame,
+#if CONFIG_REF_MV
+                     &xd->ref_mv_count[frame], xd->ref_mv_stack[frame],
+#if CONFIG_EXT_INTER
+                     compound_inter_mode_ctx,
+#endif  // CONFIG_EXT_INTER
+#endif
+                     ref_mvs[frame], mi_row, mi_col, fpm_sync, (void *)pbi,
+                     inter_mode_ctx);
+  }
+
+#if CONFIG_REF_MV
+  if (is_compound) {
+    MV_REFERENCE_FRAME ref_frame = av1_ref_frame_type(mbmi->ref_frame);
+    av1_find_mv_refs(cm, xd, mi, ref_frame, &xd->ref_mv_count[ref_frame],
+                     xd->ref_mv_stack[ref_frame],
+#if CONFIG_EXT_INTER
+                     compound_inter_mode_ctx,
+#endif  // CONFIG_EXT_INTER
+                     ref_mvs[ref_frame], mi_row, mi_col, fpm_sync, (void *)pbi,
+                     inter_mode_ctx);
+
+    if (xd->ref_mv_count[ref_frame] < 2) {
+      MV_REFERENCE_FRAME rf[2];
+      int_mv zeromv[2];
+      av1_set_ref_frame(rf, ref_frame);
+#if CONFIG_GLOBAL_MOTION
+      zeromv[0].as_int = gm_get_motion_vector(&cm->global_motion[rf[0]],
+                                              cm->allow_high_precision_mv,
+                                              bsize, mi_col, mi_row, 0)
+                             .as_int;
+      zeromv[1].as_int = (rf[1] != NONE_FRAME)
+                             ? gm_get_motion_vector(&cm->global_motion[rf[1]],
+                                                    cm->allow_high_precision_mv,
+                                                    bsize, mi_col, mi_row, 0)
+                                   .as_int
+                             : 0;
+#else
+      zeromv[0].as_int = zeromv[1].as_int = 0;
+#endif
+      for (ref = 0; ref < 2; ++ref) {
+        if (rf[ref] == NONE_FRAME) continue;
+        lower_mv_precision(&ref_mvs[rf[ref]][0].as_mv, allow_hp);
+        lower_mv_precision(&ref_mvs[rf[ref]][1].as_mv, allow_hp);
+        if (ref_mvs[rf[ref]][0].as_int != zeromv[ref].as_int ||
+            ref_mvs[rf[ref]][1].as_int != zeromv[ref].as_int)
+          inter_mode_ctx[ref_frame] &= ~(1 << ALL_ZERO_FLAG_OFFSET);
+      }
+    }
+  }
+
+#if CONFIG_EXT_INTER
+  if (is_compound)
+    mode_ctx = compound_inter_mode_ctx[mbmi->ref_frame[0]];
+  else
+#endif  // CONFIG_EXT_INTER
+    mode_ctx =
+        av1_mode_context_analyzer(inter_mode_ctx, mbmi->ref_frame, bsize, -1);
+  mbmi->ref_mv_idx = 0;
+#else
+  mode_ctx = inter_mode_ctx[mbmi->ref_frame[0]];
+#endif
+
+  if (segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+    mbmi->mode = ZEROMV;
+    if (bsize < BLOCK_8X8 && !unify_bsize) {
+      aom_internal_error(xd->error_info, AOM_CODEC_UNSUP_BITSTREAM,
+                         "Invalid usage of segement feature on small blocks");
+      return;
+    }
+  } else {
+    if (bsize >= BLOCK_8X8 || unify_bsize) {
+#if CONFIG_EXT_INTER
+      if (is_compound)
+        mbmi->mode = read_inter_compound_mode(cm, xd, r, mode_ctx);
+      else
+#endif  // CONFIG_EXT_INTER
+        mbmi->mode = read_inter_mode(ec_ctx, xd, r, mode_ctx);
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+      if (mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV ||
+          have_nearmv_in_inter_mode(mbmi->mode))
+#else
+      if (mbmi->mode == NEARMV || mbmi->mode == NEWMV)
+#endif
+        read_drl_idx(cm, xd, mbmi, r);
+#endif
+    }
+  }
+
+#if CONFIG_EXT_INTER
+  if ((bsize < BLOCK_8X8 && unify_bsize) ||
+      (mbmi->mode != ZEROMV && mbmi->mode != ZERO_ZEROMV)) {
+#else
+  if ((bsize < BLOCK_8X8 && !unify_bsize) || mbmi->mode != ZEROMV) {
+#endif  // CONFIG_EXT_INTER
+    for (ref = 0; ref < 1 + is_compound; ++ref) {
+      av1_find_best_ref_mvs(allow_hp, ref_mvs[mbmi->ref_frame[ref]],
+                            &nearestmv[ref], &nearmv[ref]);
+    }
+  }
+
+#if CONFIG_REF_MV
+  if (mbmi->ref_mv_idx > 0) {
+    int_mv cur_mv =
+        xd->ref_mv_stack[mbmi->ref_frame[0]][1 + mbmi->ref_mv_idx].this_mv;
+    nearmv[0] = cur_mv;
+  }
+
+#if CONFIG_EXT_INTER
+  if (is_compound && (bsize >= BLOCK_8X8 || unify_bsize) &&
+      mbmi->mode != ZERO_ZEROMV) {
+#else
+  if (is_compound && (bsize >= BLOCK_8X8 || unify_bsize) &&
+      mbmi->mode != NEWMV && mbmi->mode != ZEROMV) {
+#endif  // CONFIG_EXT_INTER
+    uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+
+#if CONFIG_EXT_INTER
+    if (xd->ref_mv_count[ref_frame_type] > 0) {
+#else
+    if (xd->ref_mv_count[ref_frame_type] == 1 && mbmi->mode == NEARESTMV) {
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_EXT_INTER
+      if (mbmi->mode == NEAREST_NEARESTMV) {
+#endif  // CONFIG_EXT_INTER
+        nearestmv[0] = xd->ref_mv_stack[ref_frame_type][0].this_mv;
+        nearestmv[1] = xd->ref_mv_stack[ref_frame_type][0].comp_mv;
+        lower_mv_precision(&nearestmv[0].as_mv, allow_hp);
+        lower_mv_precision(&nearestmv[1].as_mv, allow_hp);
+#if CONFIG_EXT_INTER
+      } else if (mbmi->mode == NEAREST_NEWMV || mbmi->mode == NEAREST_NEARMV) {
+        nearestmv[0] = xd->ref_mv_stack[ref_frame_type][0].this_mv;
+        lower_mv_precision(&nearestmv[0].as_mv, allow_hp);
+      } else if (mbmi->mode == NEW_NEARESTMV || mbmi->mode == NEAR_NEARESTMV) {
+        nearestmv[1] = xd->ref_mv_stack[ref_frame_type][0].comp_mv;
+        lower_mv_precision(&nearestmv[1].as_mv, allow_hp);
+      }
+#endif  // CONFIG_EXT_INTER
+    }
+
+#if CONFIG_EXT_INTER
+    if (xd->ref_mv_count[ref_frame_type] > 1) {
+      int ref_mv_idx = 1 + mbmi->ref_mv_idx;
+      if (compound_ref0_mode(mbmi->mode) == NEARMV) {
+        nearmv[0] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].this_mv;
+        lower_mv_precision(&nearmv[0].as_mv, allow_hp);
+      }
+
+      if (compound_ref1_mode(mbmi->mode) == NEARMV) {
+        nearmv[1] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].comp_mv;
+        lower_mv_precision(&nearmv[1].as_mv, allow_hp);
+      }
+    }
+#else
+    if (xd->ref_mv_count[ref_frame_type] > 1) {
+      int ref_mv_idx = 1 + mbmi->ref_mv_idx;
+      nearestmv[0] = xd->ref_mv_stack[ref_frame_type][0].this_mv;
+      nearestmv[1] = xd->ref_mv_stack[ref_frame_type][0].comp_mv;
+      nearmv[0] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].this_mv;
+      nearmv[1] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].comp_mv;
+    }
+#endif  // CONFIG_EXT_INTER
+  }
+#endif
+
+#if !CONFIG_DUAL_FILTER && !CONFIG_WARPED_MOTION && !CONFIG_GLOBAL_MOTION
+  read_mb_interp_filter(cm, xd, mbmi, r);
+#endif  // !CONFIG_DUAL_FILTER && !CONFIG_WARPED_MOTION
+
+  if (bsize < BLOCK_8X8 && !unify_bsize) {
+    const int num_4x4_w = 1 << xd->bmode_blocks_wl;
+    const int num_4x4_h = 1 << xd->bmode_blocks_hl;
+    int idx, idy;
+    PREDICTION_MODE b_mode;
+    int_mv nearest_sub8x8[2], near_sub8x8[2];
+#if CONFIG_EXT_INTER
+    int_mv ref_mv[2][2];
+#endif  // CONFIG_EXT_INTER
+    for (idy = 0; idy < 2; idy += num_4x4_h) {
+      for (idx = 0; idx < 2; idx += num_4x4_w) {
+        int_mv block[2];
+        const int j = idy * 2 + idx;
+        int_mv ref_mv_s8[2];
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+        if (!is_compound)
+#endif  // CONFIG_EXT_INTER
+          mode_ctx = av1_mode_context_analyzer(inter_mode_ctx, mbmi->ref_frame,
+                                               bsize, j);
+#endif
+#if CONFIG_EXT_INTER
+        if (is_compound)
+          b_mode = read_inter_compound_mode(cm, xd, r, mode_ctx);
+        else
+#endif  // CONFIG_EXT_INTER
+          b_mode = read_inter_mode(ec_ctx, xd, r, mode_ctx);
+
+#if CONFIG_EXT_INTER
+        if (b_mode != ZEROMV && b_mode != ZERO_ZEROMV) {
+#else
+        if (b_mode != ZEROMV) {
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_REF_MV
+          CANDIDATE_MV ref_mv_stack[2][MAX_REF_MV_STACK_SIZE];
+          uint8_t ref_mv_count[2];
+#endif
+          for (ref = 0; ref < 1 + is_compound; ++ref)
+#if CONFIG_EXT_INTER
+          {
+            int_mv mv_ref_list[MAX_MV_REF_CANDIDATES];
+            av1_update_mv_context(cm, xd, mi, mbmi->ref_frame[ref], mv_ref_list,
+                                  j, mi_row, mi_col, NULL);
+#endif  // CONFIG_EXT_INTER
+            av1_append_sub8x8_mvs_for_idx(cm, xd, j, ref, mi_row, mi_col,
+#if CONFIG_REF_MV
+                                          ref_mv_stack[ref], &ref_mv_count[ref],
+#endif
+#if CONFIG_EXT_INTER
+                                          mv_ref_list,
+#endif  // CONFIG_EXT_INTER
+                                          &nearest_sub8x8[ref],
+                                          &near_sub8x8[ref]);
+#if CONFIG_EXT_INTER
+            if (have_newmv_in_inter_mode(b_mode)) {
+              mv_ref_list[0].as_int = nearest_sub8x8[ref].as_int;
+              mv_ref_list[1].as_int = near_sub8x8[ref].as_int;
+              av1_find_best_ref_mvs(allow_hp, mv_ref_list, &ref_mv[0][ref],
+                                    &ref_mv[1][ref]);
+            }
+          }
+#endif  // CONFIG_EXT_INTER
+        }
+
+        for (ref = 0; ref < 1 + is_compound && b_mode != ZEROMV; ++ref) {
+#if CONFIG_REF_MV
+          ref_mv_s8[ref] = nearest_sub8x8[ref];
+          lower_mv_precision(&ref_mv_s8[ref].as_mv, allow_hp);
+#else
+          ref_mv_s8[ref] = nearestmv[ref];
+#endif
+        }
+#if CONFIG_EXT_INTER
+        (void)ref_mv_s8;
+#endif
+
+        if (!assign_mv(cm, xd, b_mode, mbmi->ref_frame, j, block,
+#if CONFIG_EXT_INTER
+                       ref_mv[0],
+#else   // !CONFIG_EXT_INTER
+                       ref_mv_s8,
+#endif  // CONFIG_EXT_INTER
+                       nearest_sub8x8, near_sub8x8, mi_row, mi_col, is_compound,
+                       allow_hp, r)) {
+          aom_merge_corrupted_flag(&xd->corrupted, 1);
+          break;
+        };
+
+        mi->bmi[j].as_mv[0].as_int = block[0].as_int;
+        mi->bmi[j].as_mode = b_mode;
+        if (is_compound) mi->bmi[j].as_mv[1].as_int = block[1].as_int;
+
+        if (num_4x4_h == 2) mi->bmi[j + 2] = mi->bmi[j];
+        if (num_4x4_w == 2) mi->bmi[j + 1] = mi->bmi[j];
+      }
+    }
+
+#if CONFIG_REF_MV
+    mbmi->pred_mv[0].as_int = mi->bmi[3].pred_mv[0].as_int;
+    mbmi->pred_mv[1].as_int = mi->bmi[3].pred_mv[1].as_int;
+#endif
+    mi->mbmi.mode = b_mode;
+
+    mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
+    mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
+  } else {
+    int_mv ref_mv[2];
+    ref_mv[0] = nearestmv[0];
+    ref_mv[1] = nearestmv[1];
+
+#if CONFIG_EXT_INTER
+    if (is_compound) {
+#if CONFIG_REF_MV
+      int ref_mv_idx = mbmi->ref_mv_idx;
+      // Special case: NEAR_NEWMV and NEW_NEARMV modes use
+      // 1 + mbmi->ref_mv_idx (like NEARMV) instead of
+      // mbmi->ref_mv_idx (like NEWMV)
+      if (mbmi->mode == NEAR_NEWMV || mbmi->mode == NEW_NEARMV)
+        ref_mv_idx = 1 + mbmi->ref_mv_idx;
+#endif
+
+      if (compound_ref0_mode(mbmi->mode) == NEWMV) {
+#if CONFIG_REF_MV
+        uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+        if (xd->ref_mv_count[ref_frame_type] > 1) {
+          ref_mv[0] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].this_mv;
+          clamp_mv_ref(&ref_mv[0].as_mv, xd->n8_w << MI_SIZE_LOG2,
+                       xd->n8_h << MI_SIZE_LOG2, xd);
+        }
+#endif
+        nearestmv[0] = ref_mv[0];
+      }
+      if (compound_ref1_mode(mbmi->mode) == NEWMV) {
+#if CONFIG_REF_MV
+        uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+        if (xd->ref_mv_count[ref_frame_type] > 1) {
+          ref_mv[1] = xd->ref_mv_stack[ref_frame_type][ref_mv_idx].comp_mv;
+          clamp_mv_ref(&ref_mv[1].as_mv, xd->n8_w << MI_SIZE_LOG2,
+                       xd->n8_h << MI_SIZE_LOG2, xd);
+        }
+#endif
+        nearestmv[1] = ref_mv[1];
+      }
+    } else {
+#endif  // CONFIG_EXT_INTER
+      if (mbmi->mode == NEWMV) {
+        for (ref = 0; ref < 1 + is_compound; ++ref) {
+#if CONFIG_REF_MV
+          uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+          if (xd->ref_mv_count[ref_frame_type] > 1) {
+            ref_mv[ref] =
+                (ref == 0)
+                    ? xd->ref_mv_stack[ref_frame_type][mbmi->ref_mv_idx].this_mv
+                    : xd->ref_mv_stack[ref_frame_type][mbmi->ref_mv_idx]
+                          .comp_mv;
+            clamp_mv_ref(&ref_mv[ref].as_mv, xd->n8_w << MI_SIZE_LOG2,
+                         xd->n8_h << MI_SIZE_LOG2, xd);
+          }
+#endif
+          nearestmv[ref] = ref_mv[ref];
+        }
+      }
+#if CONFIG_EXT_INTER
+    }
+#endif  // CONFIG_EXT_INTER
+
+    int mv_corrupted_flag =
+        !assign_mv(cm, xd, mbmi->mode, mbmi->ref_frame, 0, mbmi->mv, ref_mv,
+                   nearestmv, nearmv, mi_row, mi_col, is_compound, allow_hp, r);
+    aom_merge_corrupted_flag(&xd->corrupted, mv_corrupted_flag);
+  }
+
+#if CONFIG_EXT_INTER
+  mbmi->use_wedge_interintra = 0;
+  if (cm->reference_mode != COMPOUND_REFERENCE &&
+#if CONFIG_SUPERTX
+      !supertx_enabled &&
+#endif
+      is_interintra_allowed(mbmi)) {
+    const int bsize_group = size_group_lookup[bsize];
+    const int interintra =
+        aom_read(r, cm->fc->interintra_prob[bsize_group], ACCT_STR);
+    if (xd->counts) xd->counts->interintra[bsize_group][interintra]++;
+    assert(mbmi->ref_frame[1] == NONE_FRAME);
+    if (interintra) {
+      const INTERINTRA_MODE interintra_mode =
+          read_interintra_mode(cm, xd, r, bsize_group);
+      mbmi->ref_frame[1] = INTRA_FRAME;
+      mbmi->interintra_mode = interintra_mode;
+#if CONFIG_EXT_INTRA
+      mbmi->angle_delta[0] = 0;
+      mbmi->angle_delta[1] = 0;
+#if CONFIG_INTRA_INTERP
+      mbmi->intra_filter = INTRA_FILTER_LINEAR;
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+      mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+      mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+#endif  // CONFIG_FILTER_INTRA
+      if (is_interintra_wedge_used(bsize)) {
+        mbmi->use_wedge_interintra =
+            aom_read(r, cm->fc->wedge_interintra_prob[bsize], ACCT_STR);
+        if (xd->counts)
+          xd->counts->wedge_interintra[bsize][mbmi->use_wedge_interintra]++;
+        if (mbmi->use_wedge_interintra) {
+          mbmi->interintra_wedge_index =
+              aom_read_literal(r, get_wedge_bits_lookup(bsize), ACCT_STR);
+          mbmi->interintra_wedge_sign = 0;
+        }
+      }
+    }
+  }
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  mbmi->motion_mode = SIMPLE_TRANSLATION;
+#if CONFIG_WARPED_MOTION
+  if (mbmi->sb_type >= BLOCK_8X8 && !has_second_ref(mbmi))
+    mbmi->num_proj_ref[0] = findSamples(cm, xd, mi_row, mi_col, pts, pts_inref);
+#endif  // CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR
+  av1_count_overlappable_neighbors(cm, xd, mi_row, mi_col);
+#endif
+
+#if CONFIG_SUPERTX
+  if (!supertx_enabled) {
+#endif  // CONFIG_SUPERTX
+#if CONFIG_EXT_INTER
+    if (mbmi->ref_frame[1] != INTRA_FRAME)
+#endif  // CONFIG_EXT_INTER
+      mbmi->motion_mode = read_motion_mode(cm, xd, mi, r);
+#if CONFIG_WARPED_MOTION
+    if (mbmi->motion_mode == WARPED_CAUSAL) {
+      mbmi->wm_params[0].wmtype = DEFAULT_WMTYPE;
+      if (find_projection(mbmi->num_proj_ref[0], pts, pts_inref, bsize,
+                          mbmi->mv[0].as_mv.row, mbmi->mv[0].as_mv.col,
+                          &mbmi->wm_params[0], mi_row, mi_col)) {
+        assert(0 && "Invalid Warped Model.");
+      }
+    }
+#endif  // CONFIG_WARPED_MOTION
+#if CONFIG_SUPERTX
+  }
+#endif  // CONFIG_SUPERTX
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_EXT_INTER
+  mbmi->interinter_compound_type = COMPOUND_AVERAGE;
+  if (cm->reference_mode != SINGLE_REFERENCE &&
+      is_inter_compound_mode(mbmi->mode)
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      && mbmi->motion_mode == SIMPLE_TRANSLATION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      ) {
+    if (is_any_masked_compound_used(bsize)) {
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+      mbmi->interinter_compound_type =
+          aom_read_tree(r, av1_compound_type_tree,
+                        cm->fc->compound_type_prob[bsize], ACCT_STR);
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#if CONFIG_WEDGE
+      if (mbmi->interinter_compound_type == COMPOUND_WEDGE) {
+        mbmi->wedge_index =
+            aom_read_literal(r, get_wedge_bits_lookup(bsize), ACCT_STR);
+        mbmi->wedge_sign = aom_read_bit(r, ACCT_STR);
+      }
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+      if (mbmi->interinter_compound_type == COMPOUND_SEG) {
+        mbmi->mask_type = aom_read_literal(r, MAX_SEG_MASK_BITS, ACCT_STR);
+      }
+#endif  // CONFIG_COMPOUND_SEGMENT
+    } else {
+      mbmi->interinter_compound_type = COMPOUND_AVERAGE;
+    }
+    if (xd->counts)
+      xd->counts->compound_interinter[bsize][mbmi->interinter_compound_type]++;
+  }
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_DUAL_FILTER || CONFIG_WARPED_MOTION || CONFIG_GLOBAL_MOTION
+  read_mb_interp_filter(cm, xd, mbmi, r);
+#endif  // CONFIG_DUAL_FILTER || CONFIG_WARPED_MOTION
+}
+
+static void read_inter_frame_mode_info(AV1Decoder *const pbi,
+                                       MACROBLOCKD *const xd,
+#if CONFIG_SUPERTX
+                                       int supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                                       int mi_row, int mi_col, aom_reader *r) {
+  AV1_COMMON *const cm = &pbi->common;
+  MODE_INFO *const mi = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mi->mbmi;
+  int inter_block = 1;
+#if CONFIG_VAR_TX
+  BLOCK_SIZE bsize = mbmi->sb_type;
+#endif  // CONFIG_VAR_TX
+
+  mbmi->mv[0].as_int = 0;
+  mbmi->mv[1].as_int = 0;
+  mbmi->segment_id = read_inter_segment_id(cm, xd, mi_row, mi_col, r);
+#if CONFIG_SUPERTX
+  if (!supertx_enabled)
+#endif  // CONFIG_SUPERTX
+    mbmi->skip = read_skip(cm, xd, mbmi->segment_id, r);
+
+#if CONFIG_DELTA_Q
+  if (cm->delta_q_present_flag) {
+    xd->current_qindex =
+        xd->prev_qindex +
+        read_delta_qindex(cm, xd, r, mbmi, mi_col, mi_row) * cm->delta_q_res;
+    /* Normative: Clamp to [1,MAXQ] to not interfere with lossless mode */
+    xd->current_qindex = clamp(xd->current_qindex, 1, MAXQ);
+    xd->prev_qindex = xd->current_qindex;
+#if CONFIG_EXT_DELTA_Q
+    if (cm->delta_lf_present_flag) {
+      mbmi->current_delta_lf_from_base = xd->current_delta_lf_from_base =
+          xd->prev_delta_lf_from_base +
+          read_delta_lflevel(cm, xd, r, mbmi, mi_col, mi_row) *
+              cm->delta_lf_res;
+      xd->prev_delta_lf_from_base = xd->current_delta_lf_from_base;
+    }
+#endif
+  }
+#endif
+
+#if CONFIG_SUPERTX
+  if (!supertx_enabled) {
+#endif  // CONFIG_SUPERTX
+    inter_block = read_is_inter_block(cm, xd, mbmi->segment_id, r);
+
+#if CONFIG_VAR_TX
+    xd->above_txfm_context = cm->above_txfm_context + mi_col;
+    xd->left_txfm_context =
+        xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+
+    if (cm->tx_mode == TX_MODE_SELECT &&
+#if CONFIG_CB4X4
+        bsize > BLOCK_4X4 &&
+#else
+        bsize >= BLOCK_8X8 &&
+#endif
+        !mbmi->skip && inter_block) {
+      const TX_SIZE max_tx_size = max_txsize_rect_lookup[bsize];
+      const int bh = tx_size_high_unit[max_tx_size];
+      const int bw = tx_size_wide_unit[max_tx_size];
+      const int width = block_size_wide[bsize] >> tx_size_wide_log2[0];
+      const int height = block_size_high[bsize] >> tx_size_wide_log2[0];
+      int idx, idy;
+
+      mbmi->min_tx_size = TX_SIZES_ALL;
+      for (idy = 0; idy < height; idy += bh)
+        for (idx = 0; idx < width; idx += bw)
+          read_tx_size_vartx(cm, xd, mbmi, xd->counts, max_tx_size,
+                             height != width, idy, idx, r);
+    } else {
+      mbmi->tx_size = read_tx_size(cm, xd, inter_block, !mbmi->skip, r);
+
+      if (inter_block) {
+        const int width = block_size_wide[bsize] >> tx_size_wide_log2[0];
+        const int height = block_size_high[bsize] >> tx_size_high_log2[0];
+        int idx, idy;
+        for (idy = 0; idy < height; ++idy)
+          for (idx = 0; idx < width; ++idx)
+            mbmi->inter_tx_size[idy >> 1][idx >> 1] = mbmi->tx_size;
+      }
+      mbmi->min_tx_size = get_min_tx_size(mbmi->tx_size);
+      set_txfm_ctxs(mbmi->tx_size, xd->n8_w, xd->n8_h, mbmi->skip, xd);
+    }
+#else
+  mbmi->tx_size = read_tx_size(cm, xd, inter_block, !mbmi->skip, r);
+#endif  // CONFIG_VAR_TX
+#if CONFIG_SUPERTX
+  }
+#if CONFIG_VAR_TX
+  else if (inter_block) {
+    const int width = num_4x4_blocks_wide_lookup[bsize];
+    const int height = num_4x4_blocks_high_lookup[bsize];
+    int idx, idy;
+    xd->mi[0]->mbmi.tx_size = xd->supertx_size;
+    for (idy = 0; idy < height; ++idy)
+      for (idx = 0; idx < width; ++idx)
+        xd->mi[0]->mbmi.inter_tx_size[idy >> 1][idx >> 1] = xd->supertx_size;
+  }
+#endif  // CONFIG_VAR_TX
+#endif  // CONFIG_SUPERTX
+
+  if (inter_block)
+    read_inter_block_mode_info(pbi, xd,
+#if (CONFIG_MOTION_VAR || CONFIG_EXT_INTER || CONFIG_WARPED_MOTION) && \
+    CONFIG_SUPERTX
+
+                               mi, mi_row, mi_col, r, supertx_enabled);
+#else
+                               mi, mi_row, mi_col, r);
+#endif  // CONFIG_MOTION_VAR && CONFIG_SUPERTX
+  else
+    read_intra_block_mode_info(cm, mi_row, mi_col, xd, mi, r);
+
+#if !CONFIG_TXK_SEL
+  av1_read_tx_type(cm, xd,
+#if CONFIG_SUPERTX
+                   supertx_enabled,
+#endif
+                   r);
+#endif  // !CONFIG_TXK_SEL
+}
+
+void av1_read_mode_info(AV1Decoder *const pbi, MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                        int supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                        int mi_row, int mi_col, aom_reader *r, int x_mis,
+                        int y_mis) {
+  AV1_COMMON *const cm = &pbi->common;
+  MODE_INFO *const mi = xd->mi[0];
+  MV_REF *frame_mvs = cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
+  int w, h;
+
+#if CONFIG_INTRABC
+  mi->mbmi.use_intrabc = 0;
+#endif  // CONFIG_INTRABC
+
+  if (frame_is_intra_only(cm)) {
+    read_intra_frame_mode_info(cm, xd, mi_row, mi_col, r);
+#if CONFIG_REF_MV
+    for (h = 0; h < y_mis; ++h) {
+      MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
+      for (w = 0; w < x_mis; ++w) {
+        MV_REF *const mv = frame_mv + w;
+        mv->ref_frame[0] = NONE_FRAME;
+        mv->ref_frame[1] = NONE_FRAME;
+      }
+    }
+#endif
+  } else {
+    read_inter_frame_mode_info(pbi, xd,
+#if CONFIG_SUPERTX
+                               supertx_enabled,
+#endif  // CONFIG_SUPERTX
+                               mi_row, mi_col, r);
+    for (h = 0; h < y_mis; ++h) {
+      MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
+      for (w = 0; w < x_mis; ++w) {
+        MV_REF *const mv = frame_mv + w;
+        mv->ref_frame[0] = mi->mbmi.ref_frame[0];
+        mv->ref_frame[1] = mi->mbmi.ref_frame[1];
+        mv->mv[0].as_int = mi->mbmi.mv[0].as_int;
+        mv->mv[1].as_int = mi->mbmi.mv[1].as_int;
+#if CONFIG_REF_MV
+        mv->pred_mv[0].as_int = mi->mbmi.pred_mv[0].as_int;
+        mv->pred_mv[1].as_int = mi->mbmi.pred_mv[1].as_int;
+#endif
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/decoder/decodemv.h b/third_party/aom/av1/decoder/decodemv.h
new file mode 100644
index 0000000000..ceaee1d6b0
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodemv.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DECODEMV_H_
+#define AV1_DECODER_DECODEMV_H_
+
+#include "aom_dsp/bitreader.h"
+
+#include "av1/decoder/decoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_read_mode_info(AV1Decoder *const pbi, MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                        int supertx_enabled,
+#endif
+
+                        int mi_row, int mi_col, aom_reader *r, int x_mis,
+                        int y_mis);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+void av1_read_tx_type(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                      int supertx_enabled,
+#endif
+#if CONFIG_TXK_SEL
+                      int block, int plane,
+#endif
+                      aom_reader *r);
+
+#endif  // AV1_DECODER_DECODEMV_H_
diff --git a/third_party/aom/av1/decoder/decoder.c b/third_party/aom/av1/decoder/decoder.c
new file mode 100644
index 0000000000..1bd91086e5
--- /dev/null
+++ b/third_party/aom/av1/decoder/decoder.c
@@ -0,0 +1,583 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <stdio.h>
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "./aom_scale_rtcd.h"
+
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/system_state.h"
+#include "aom_ports/aom_once.h"
+#include "aom_ports/aom_timer.h"
+#include "aom_scale/aom_scale.h"
+#include "aom_util/aom_thread.h"
+
+#include "av1/common/alloccommon.h"
+#include "av1/common/av1_loopfilter.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+
+#include "av1/decoder/decodeframe.h"
+#include "av1/decoder/decoder.h"
+
+#if !CONFIG_PVQ
+#include "av1/decoder/detokenize.h"
+#endif
+
+static void initialize_dec(void) {
+  static volatile int init_done = 0;
+
+  if (!init_done) {
+    av1_rtcd();
+    aom_dsp_rtcd();
+    aom_scale_rtcd();
+    av1_init_intra_predictors();
+#if CONFIG_EXT_INTER
+    av1_init_wedge_masks();
+#endif  // CONFIG_EXT_INTER
+    init_done = 1;
+#if CONFIG_EC_MULTISYMBOL
+    av1_indices_from_tree(av1_intra_mode_ind, av1_intra_mode_inv,
+                          av1_intra_mode_tree);
+    av1_indices_from_tree(av1_switchable_interp_ind, av1_switchable_interp_inv,
+                          av1_switchable_interp_tree);
+#if CONFIG_EXT_TX
+    int s;
+    for (s = 1; s < EXT_TX_SETS_INTRA; ++s)
+      av1_indices_from_tree(av1_ext_tx_intra_ind[s], av1_ext_tx_intra_inv[s],
+                            av1_ext_tx_intra_tree[s]);
+    for (s = 1; s < EXT_TX_SETS_INTER; ++s)
+      av1_indices_from_tree(av1_ext_tx_inter_ind[s], av1_ext_tx_inter_inv[s],
+                            av1_ext_tx_inter_tree[s]);
+#else
+    av1_indices_from_tree(av1_ext_tx_ind, av1_ext_tx_inv, av1_ext_tx_tree);
+#endif
+    av1_indices_from_tree(av1_inter_mode_ind, av1_inter_mode_inv,
+                          av1_inter_mode_tree);
+#endif
+  }
+}
+
+static void av1_dec_setup_mi(AV1_COMMON *cm) {
+  cm->mi = cm->mip + cm->mi_stride + 1;
+  cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
+  memset(cm->mi_grid_base, 0,
+         cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
+}
+
+static int av1_dec_alloc_mi(AV1_COMMON *cm, int mi_size) {
+  cm->mip = aom_calloc(mi_size, sizeof(*cm->mip));
+  if (!cm->mip) return 1;
+  cm->mi_alloc_size = mi_size;
+  cm->mi_grid_base = (MODE_INFO **)aom_calloc(mi_size, sizeof(MODE_INFO *));
+  if (!cm->mi_grid_base) return 1;
+  return 0;
+}
+
+static void av1_dec_free_mi(AV1_COMMON *cm) {
+  aom_free(cm->mip);
+  cm->mip = NULL;
+  aom_free(cm->mi_grid_base);
+  cm->mi_grid_base = NULL;
+}
+
+AV1Decoder *av1_decoder_create(BufferPool *const pool) {
+  AV1Decoder *volatile const pbi = aom_memalign(32, sizeof(*pbi));
+  AV1_COMMON *volatile const cm = pbi ? &pbi->common : NULL;
+
+  if (!cm) return NULL;
+
+  av1_zero(*pbi);
+
+  if (setjmp(cm->error.jmp)) {
+    cm->error.setjmp = 0;
+    av1_decoder_remove(pbi);
+    return NULL;
+  }
+
+  cm->error.setjmp = 1;
+
+  CHECK_MEM_ERROR(cm, cm->fc,
+                  (FRAME_CONTEXT *)aom_memalign(32, sizeof(*cm->fc)));
+  CHECK_MEM_ERROR(cm, cm->frame_contexts,
+                  (FRAME_CONTEXT *)aom_memalign(
+                      32, FRAME_CONTEXTS * sizeof(*cm->frame_contexts)));
+  memset(cm->fc, 0, sizeof(*cm->fc));
+  memset(cm->frame_contexts, 0, FRAME_CONTEXTS * sizeof(*cm->frame_contexts));
+
+  pbi->need_resync = 1;
+  once(initialize_dec);
+
+  // Initialize the references to not point to any frame buffers.
+  memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+  memset(&cm->next_ref_frame_map, -1, sizeof(cm->next_ref_frame_map));
+
+  cm->current_video_frame = 0;
+  pbi->ready_for_new_data = 1;
+  pbi->common.buffer_pool = pool;
+
+  cm->bit_depth = AOM_BITS_8;
+  cm->dequant_bit_depth = AOM_BITS_8;
+
+  cm->alloc_mi = av1_dec_alloc_mi;
+  cm->free_mi = av1_dec_free_mi;
+  cm->setup_mi = av1_dec_setup_mi;
+
+  av1_loop_filter_init(cm);
+
+#if CONFIG_AOM_QM
+  aom_qm_init(cm);
+#endif
+#if CONFIG_LOOP_RESTORATION
+  av1_loop_restoration_precal();
+#endif  // CONFIG_LOOP_RESTORATION
+#if CONFIG_ACCOUNTING
+  pbi->acct_enabled = 1;
+  aom_accounting_init(&pbi->accounting);
+#endif
+
+  cm->error.setjmp = 0;
+
+  aom_get_worker_interface()->init(&pbi->lf_worker);
+
+  return pbi;
+}
+
+void av1_decoder_remove(AV1Decoder *pbi) {
+  int i;
+
+  if (!pbi) return;
+
+  aom_get_worker_interface()->end(&pbi->lf_worker);
+  aom_free(pbi->lf_worker.data1);
+  aom_free(pbi->tile_data);
+  for (i = 0; i < pbi->num_tile_workers; ++i) {
+    AVxWorker *const worker = &pbi->tile_workers[i];
+    aom_get_worker_interface()->end(worker);
+  }
+  aom_free(pbi->tile_worker_data);
+  aom_free(pbi->tile_worker_info);
+  aom_free(pbi->tile_workers);
+
+  if (pbi->num_tile_workers > 0) {
+    av1_loop_filter_dealloc(&pbi->lf_row_sync);
+  }
+
+#if CONFIG_ACCOUNTING
+  aom_accounting_clear(&pbi->accounting);
+#endif
+
+  aom_free(pbi);
+}
+
+static int equal_dimensions(const YV12_BUFFER_CONFIG *a,
+                            const YV12_BUFFER_CONFIG *b) {
+  return a->y_height == b->y_height && a->y_width == b->y_width &&
+         a->uv_height == b->uv_height && a->uv_width == b->uv_width;
+}
+
+aom_codec_err_t av1_copy_reference_dec(AV1Decoder *pbi,
+                                       AOM_REFFRAME ref_frame_flag,
+                                       YV12_BUFFER_CONFIG *sd) {
+  AV1_COMMON *cm = &pbi->common;
+
+  /* TODO(jkoleszar): The decoder doesn't have any real knowledge of what the
+   * encoder is using the frame buffers for. This is just a stub to keep the
+   * aomenc --test-decode functionality working, and will be replaced in a
+   * later commit that adds AV1-specific controls for this functionality.
+   */
+  if (ref_frame_flag == AOM_LAST_FLAG) {
+    const YV12_BUFFER_CONFIG *const cfg = get_ref_frame(cm, 0);
+    if (cfg == NULL) {
+      aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                         "No 'last' reference frame");
+      return AOM_CODEC_ERROR;
+    }
+    if (!equal_dimensions(cfg, sd))
+      aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                         "Incorrect buffer dimensions");
+    else
+      aom_yv12_copy_frame(cfg, sd);
+  } else {
+    aom_internal_error(&cm->error, AOM_CODEC_ERROR, "Invalid reference frame");
+  }
+
+  return cm->error.error_code;
+}
+
+aom_codec_err_t av1_set_reference_dec(AV1_COMMON *cm,
+                                      AOM_REFFRAME ref_frame_flag,
+                                      YV12_BUFFER_CONFIG *sd) {
+  int idx;
+  YV12_BUFFER_CONFIG *ref_buf = NULL;
+
+  // TODO(jkoleszar): The decoder doesn't have any real knowledge of what the
+  // encoder is using the frame buffers for. This is just a stub to keep the
+  // aomenc --test-decode functionality working, and will be replaced in a
+  // later commit that adds AV1-specific controls for this functionality.
+
+  // (Yunqing) The set_reference control depends on the following setting in
+  // encoder.
+  //   cpi->lst_fb_idx = 0;
+  // #if CONFIG_EXT_REFS
+  //   cpi->lst2_fb_idx = 1;
+  //   cpi->lst3_fb_idx = 2;
+  //   cpi->gld_fb_idx = 3;
+  //   cpi->bwd_fb_idx = 4;
+  //   cpi->alt_fb_idx = 5;
+  // #else  // CONFIG_EXT_REFS
+  //   cpi->gld_fb_idx = 1;
+  //   cpi->alt_fb_idx = 2;
+  // #endif  // CONFIG_EXT_REFS
+
+  // TODO(zoeliu): To revisit following code and reconsider what assumption we
+  // may take on the reference frame buffer virtual indexes
+  if (ref_frame_flag == AOM_LAST_FLAG) {
+    idx = cm->ref_frame_map[0];
+#if CONFIG_EXT_REFS
+  } else if (ref_frame_flag == AOM_LAST2_FLAG) {
+    idx = cm->ref_frame_map[1];
+  } else if (ref_frame_flag == AOM_LAST3_FLAG) {
+    idx = cm->ref_frame_map[2];
+  } else if (ref_frame_flag == AOM_GOLD_FLAG) {
+    idx = cm->ref_frame_map[3];
+  } else if (ref_frame_flag == AOM_BWD_FLAG) {
+    idx = cm->ref_frame_map[4];
+  } else if (ref_frame_flag == AOM_ALT_FLAG) {
+    idx = cm->ref_frame_map[5];
+#else
+  } else if (ref_frame_flag == AOM_GOLD_FLAG) {
+    idx = cm->ref_frame_map[1];
+  } else if (ref_frame_flag == AOM_ALT_FLAG) {
+    idx = cm->ref_frame_map[2];
+#endif  // CONFIG_EXT_REFS
+  } else {
+    aom_internal_error(&cm->error, AOM_CODEC_ERROR, "Invalid reference frame");
+    return cm->error.error_code;
+  }
+
+  if (idx < 0 || idx >= FRAME_BUFFERS) {
+    aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                       "Invalid reference frame map");
+    return cm->error.error_code;
+  }
+
+  // Get the destination reference buffer.
+  ref_buf = &cm->buffer_pool->frame_bufs[idx].buf;
+
+  if (!equal_dimensions(ref_buf, sd)) {
+    aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                       "Incorrect buffer dimensions");
+  } else {
+    // Overwrite the reference frame buffer.
+    aom_yv12_copy_frame(sd, ref_buf);
+  }
+
+  return cm->error.error_code;
+}
+
+/* If any buffer updating is signaled it should be done here. */
+static void swap_frame_buffers(AV1Decoder *pbi) {
+  int ref_index = 0, mask;
+  AV1_COMMON *const cm = &pbi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+
+  lock_buffer_pool(pool);
+  for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+    const int old_idx = cm->ref_frame_map[ref_index];
+    // Current thread releases the holding of reference frame.
+    decrease_ref_count(old_idx, frame_bufs, pool);
+
+    // Release the reference frame holding in the reference map for the decoding
+    // of the next frame.
+    if (mask & 1) decrease_ref_count(old_idx, frame_bufs, pool);
+    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
+    ++ref_index;
+  }
+
+  // Current thread releases the holding of reference frame.
+  for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
+    const int old_idx = cm->ref_frame_map[ref_index];
+    decrease_ref_count(old_idx, frame_bufs, pool);
+    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
+  }
+
+  unlock_buffer_pool(pool);
+  pbi->hold_ref_buf = 0;
+  cm->frame_to_show = get_frame_new_buffer(cm);
+
+  // TODO(zoeliu): To fix the ref frame buffer update for the scenario of
+  //               cm->frame_parellel_decode == 1
+  if (!cm->frame_parallel_decode || !cm->show_frame) {
+    lock_buffer_pool(pool);
+    --frame_bufs[cm->new_fb_idx].ref_count;
+    unlock_buffer_pool(pool);
+  }
+
+  // Invalidate these references until the next frame starts.
+  for (ref_index = 0; ref_index < INTER_REFS_PER_FRAME; ref_index++) {
+    cm->frame_refs[ref_index].idx = INVALID_IDX;
+    cm->frame_refs[ref_index].buf = NULL;
+  }
+}
+
+int av1_receive_compressed_data(AV1Decoder *pbi, size_t size,
+                                const uint8_t **psource) {
+  AV1_COMMON *volatile const cm = &pbi->common;
+  BufferPool *volatile const pool = cm->buffer_pool;
+  RefCntBuffer *volatile const frame_bufs = cm->buffer_pool->frame_bufs;
+  const uint8_t *source = *psource;
+  int retcode = 0;
+  cm->error.error_code = AOM_CODEC_OK;
+
+  if (size == 0) {
+    // This is used to signal that we are missing frames.
+    // We do not know if the missing frame(s) was supposed to update
+    // any of the reference buffers, but we act conservative and
+    // mark only the last buffer as corrupted.
+    //
+    // TODO(jkoleszar): Error concealment is undefined and non-normative
+    // at this point, but if it becomes so, [0] may not always be the correct
+    // thing to do here.
+    if (cm->frame_refs[0].idx > 0) {
+      assert(cm->frame_refs[0].buf != NULL);
+      cm->frame_refs[0].buf->corrupted = 1;
+    }
+  }
+
+  pbi->ready_for_new_data = 0;
+
+  // Find a free buffer for the new frame, releasing the reference previously
+  // held.
+
+  // Check if the previous frame was a frame without any references to it.
+  // Release frame buffer if not decoding in frame parallel mode.
+  if (!cm->frame_parallel_decode && cm->new_fb_idx >= 0 &&
+      frame_bufs[cm->new_fb_idx].ref_count == 0)
+    pool->release_fb_cb(pool->cb_priv,
+                        &frame_bufs[cm->new_fb_idx].raw_frame_buffer);
+
+  // Find a free frame buffer. Return error if can not find any.
+  cm->new_fb_idx = get_free_fb(cm);
+  if (cm->new_fb_idx == INVALID_IDX) return AOM_CODEC_MEM_ERROR;
+
+  // Assign a MV array to the frame buffer.
+  cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
+
+  pbi->hold_ref_buf = 0;
+  if (cm->frame_parallel_decode) {
+    AVxWorker *const worker = pbi->frame_worker_owner;
+    av1_frameworker_lock_stats(worker);
+    frame_bufs[cm->new_fb_idx].frame_worker_owner = worker;
+    // Reset decoding progress.
+    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
+    pbi->cur_buf->row = -1;
+    pbi->cur_buf->col = -1;
+    av1_frameworker_unlock_stats(worker);
+  } else {
+    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
+  }
+
+  if (setjmp(cm->error.jmp)) {
+    const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+    int i;
+
+    cm->error.setjmp = 0;
+    pbi->ready_for_new_data = 1;
+
+    // Synchronize all threads immediately as a subsequent decode call may
+    // cause a resize invalidating some allocations.
+    winterface->sync(&pbi->lf_worker);
+    for (i = 0; i < pbi->num_tile_workers; ++i) {
+      winterface->sync(&pbi->tile_workers[i]);
+    }
+
+    lock_buffer_pool(pool);
+    // Release all the reference buffers if worker thread is holding them.
+    if (pbi->hold_ref_buf == 1) {
+      int ref_index = 0, mask;
+      for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+        const int old_idx = cm->ref_frame_map[ref_index];
+        // Current thread releases the holding of reference frame.
+        decrease_ref_count(old_idx, frame_bufs, pool);
+
+        // Release the reference frame holding in the reference map for the
+        // decoding of the next frame.
+        if (mask & 1) decrease_ref_count(old_idx, frame_bufs, pool);
+        ++ref_index;
+      }
+
+      // Current thread releases the holding of reference frame.
+      for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
+        const int old_idx = cm->ref_frame_map[ref_index];
+        decrease_ref_count(old_idx, frame_bufs, pool);
+      }
+      pbi->hold_ref_buf = 0;
+    }
+    // Release current frame.
+    decrease_ref_count(cm->new_fb_idx, frame_bufs, pool);
+    unlock_buffer_pool(pool);
+
+    aom_clear_system_state();
+    return -1;
+  }
+
+  cm->error.setjmp = 1;
+  av1_decode_frame(pbi, source, source + size, psource);
+
+  swap_frame_buffers(pbi);
+
+#if CONFIG_EXT_TILE
+  // For now, we only extend the frame borders when the whole frame is decoded.
+  // Later, if needed, extend the border for the decoded tile on the frame
+  // border.
+  if (pbi->dec_tile_row == -1 && pbi->dec_tile_col == -1)
+#endif  // CONFIG_EXT_TILE
+    aom_extend_frame_inner_borders(cm->frame_to_show);
+
+  aom_clear_system_state();
+
+  if (!cm->show_existing_frame) {
+    cm->last_show_frame = cm->show_frame;
+
+#if CONFIG_EXT_REFS
+    // NOTE: It is not supposed to ref to any frame not used as reference
+    if (cm->is_reference_frame)
+#endif  // CONFIG_EXT_REFS
+      cm->prev_frame = cm->cur_frame;
+
+    if (cm->seg.enabled && !cm->frame_parallel_decode)
+      av1_swap_current_and_last_seg_map(cm);
+  }
+
+  // Update progress in frame parallel decode.
+  if (cm->frame_parallel_decode) {
+    // Need to lock the mutex here as another thread may
+    // be accessing this buffer.
+    AVxWorker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    av1_frameworker_lock_stats(worker);
+
+    if (cm->show_frame) {
+      cm->current_video_frame++;
+    }
+    frame_worker_data->frame_decoded = 1;
+    frame_worker_data->frame_context_ready = 1;
+    av1_frameworker_signal_stats(worker);
+    av1_frameworker_unlock_stats(worker);
+  } else {
+    cm->last_width = cm->width;
+    cm->last_height = cm->height;
+    if (cm->show_frame) {
+      cm->current_video_frame++;
+    }
+  }
+
+  cm->error.setjmp = 0;
+  return retcode;
+}
+
+int av1_get_raw_frame(AV1Decoder *pbi, YV12_BUFFER_CONFIG *sd) {
+  AV1_COMMON *const cm = &pbi->common;
+  int ret = -1;
+  if (pbi->ready_for_new_data == 1) return ret;
+
+  pbi->ready_for_new_data = 1;
+
+  /* no raw frame to show!!! */
+  if (!cm->show_frame) return ret;
+
+  pbi->ready_for_new_data = 1;
+  *sd = *cm->frame_to_show;
+  ret = 0;
+  aom_clear_system_state();
+  return ret;
+}
+
+int av1_get_frame_to_show(AV1Decoder *pbi, YV12_BUFFER_CONFIG *frame) {
+  AV1_COMMON *const cm = &pbi->common;
+
+  if (!cm->show_frame || !cm->frame_to_show) return -1;
+
+  *frame = *cm->frame_to_show;
+  return 0;
+}
+
+aom_codec_err_t av1_parse_superframe_index(const uint8_t *data, size_t data_sz,
+                                           uint32_t sizes[8], int *count,
+                                           aom_decrypt_cb decrypt_cb,
+                                           void *decrypt_state) {
+  // A chunk ending with a byte matching 0xc0 is an invalid chunk unless
+  // it is a super frame index. If the last byte of real video compression
+  // data is 0xc0 the encoder must add a 0 byte. If we have the marker but
+  // not the associated matching marker byte at the front of the index we have
+  // an invalid bitstream and need to return an error.
+
+  uint8_t marker;
+  size_t frame_sz_sum = 0;
+
+  assert(data_sz);
+  marker = read_marker(decrypt_cb, decrypt_state, data + data_sz - 1);
+  *count = 0;
+
+  if ((marker & 0xe0) == 0xc0) {
+    const uint32_t frames = (marker & 0x7) + 1;
+    const uint32_t mag = ((marker >> 3) & 0x3) + 1;
+    const size_t index_sz = 2 + mag * (frames - 1);
+
+    // This chunk is marked as having a superframe index but doesn't have
+    // enough data for it, thus it's an invalid superframe index.
+    if (data_sz < index_sz) return AOM_CODEC_CORRUPT_FRAME;
+
+    {
+      const uint8_t marker2 =
+          read_marker(decrypt_cb, decrypt_state, data + data_sz - index_sz);
+
+      // This chunk is marked as having a superframe index but doesn't have
+      // the matching marker byte at the front of the index therefore it's an
+      // invalid chunk.
+      if (marker != marker2) return AOM_CODEC_CORRUPT_FRAME;
+    }
+
+    {
+      // Found a valid superframe index.
+      uint32_t i, j;
+      const uint8_t *x = &data[data_sz - index_sz + 1];
+
+      // Frames has a maximum of 8 and mag has a maximum of 4.
+      uint8_t clear_buffer[28];
+      assert(sizeof(clear_buffer) >= (frames - 1) * mag);
+      if (decrypt_cb) {
+        decrypt_cb(decrypt_state, x, clear_buffer, (frames - 1) * mag);
+        x = clear_buffer;
+      }
+
+      for (i = 0; i < frames - 1; ++i) {
+        uint32_t this_sz = 0;
+
+        for (j = 0; j < mag; ++j) this_sz |= (*x++) << (j * 8);
+        this_sz += 1;
+        sizes[i] = this_sz;
+        frame_sz_sum += this_sz;
+      }
+      sizes[i] = (uint32_t)(data_sz - index_sz - frame_sz_sum);
+      *count = frames;
+    }
+  }
+  return AOM_CODEC_OK;
+}
diff --git a/third_party/aom/av1/decoder/decoder.h b/third_party/aom/av1/decoder/decoder.h
new file mode 100644
index 0000000000..4a90b4ad55
--- /dev/null
+++ b/third_party/aom/av1/decoder/decoder.h
@@ -0,0 +1,224 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DECODER_H_
+#define AV1_DECODER_DECODER_H_
+
+#include "./aom_config.h"
+
+#include "aom/aom_codec.h"
+#include "aom_dsp/bitreader.h"
+#include "aom_scale/yv12config.h"
+#include "aom_util/aom_thread.h"
+
+#include "av1/common/thread_common.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/decoder/dthread.h"
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#endif
+#if CONFIG_INSPECTION
+#include "av1/decoder/inspection.h"
+#endif
+
+#if CONFIG_PVQ
+#include "aom_dsp/entdec.h"
+#include "av1/decoder/decint.h"
+#include "av1/encoder/encodemb.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// TODO(hkuang): combine this with TileWorkerData.
+typedef struct TileData {
+  AV1_COMMON *cm;
+  aom_reader bit_reader;
+  DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+  /* dqcoeff are shared by all the planes. So planes must be decoded serially */
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[MAX_TX_SQUARE]);
+#if CONFIG_PVQ
+  /* forward transformed predicted image, a reference for PVQ */
+  DECLARE_ALIGNED(16, tran_low_t, pvq_ref_coeff[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+#endif
+#if CONFIG_CFL
+  CFL_CTX cfl;
+#endif
+#if CONFIG_EC_ADAPT
+  DECLARE_ALIGNED(16, FRAME_CONTEXT, tctx);
+#endif
+#if CONFIG_PALETTE
+  DECLARE_ALIGNED(16, uint8_t, color_index_map[2][MAX_SB_SQUARE]);
+#endif  // CONFIG_PALETTE
+} TileData;
+
+typedef struct TileWorkerData {
+  struct AV1Decoder *pbi;
+  aom_reader bit_reader;
+  FRAME_COUNTS counts;
+  DECLARE_ALIGNED(16, MACROBLOCKD, xd);
+  /* dqcoeff are shared by all the planes. So planes must be decoded serially */
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[MAX_TX_SQUARE]);
+#if CONFIG_PVQ
+  /* forward transformed predicted image, a reference for PVQ */
+  DECLARE_ALIGNED(16, tran_low_t, pvq_ref_coeff[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+#endif
+#if CONFIG_CFL
+  CFL_CTX cfl;
+#endif
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT tctx;
+#endif
+#if CONFIG_PALETTE
+  DECLARE_ALIGNED(16, uint8_t, color_index_map[2][MAX_SB_SQUARE]);
+#endif  // CONFIG_PALETTE
+  struct aom_internal_error_info error_info;
+} TileWorkerData;
+
+typedef struct TileBufferDec {
+  const uint8_t *data;
+  size_t size;
+  const uint8_t *raw_data_end;  // The end of the raw tile buffer in the
+                                // bit stream.
+  int col;                      // only used with multi-threaded decoding
+} TileBufferDec;
+
+typedef struct AV1Decoder {
+  DECLARE_ALIGNED(16, MACROBLOCKD, mb);
+
+  DECLARE_ALIGNED(16, AV1_COMMON, common);
+
+  int ready_for_new_data;
+
+  int refresh_frame_flags;
+
+  // TODO(hkuang): Combine this with cur_buf in macroblockd as they are
+  // the same.
+  RefCntBuffer *cur_buf;  //  Current decoding frame buffer.
+
+  AVxWorker *frame_worker_owner;  // frame_worker that owns this pbi.
+  AVxWorker lf_worker;
+  AVxWorker *tile_workers;
+  TileWorkerData *tile_worker_data;
+  TileInfo *tile_worker_info;
+  int num_tile_workers;
+
+  TileData *tile_data;
+  int allocated_tiles;
+
+  TileBufferDec tile_buffers[MAX_TILE_ROWS][MAX_TILE_COLS];
+
+  AV1LfSync lf_row_sync;
+
+  aom_decrypt_cb decrypt_cb;
+  void *decrypt_state;
+
+  int max_threads;
+  int inv_tile_order;
+  int need_resync;   // wait for key/intra-only frame.
+  int hold_ref_buf;  // hold the reference buffer.
+
+  int tile_size_bytes;
+#if CONFIG_EXT_TILE
+  int tile_col_size_bytes;
+  int dec_tile_row, dec_tile_col;
+#endif  // CONFIG_EXT_TILE
+#if CONFIG_ACCOUNTING
+  int acct_enabled;
+  Accounting accounting;
+#endif
+  size_t uncomp_hdr_size;       // Size of the uncompressed header
+  size_t first_partition_size;  // Size of the compressed header
+#if CONFIG_TILE_GROUPS
+  int tg_size;   // Number of tiles in the current tilegroup
+  int tg_start;  // First tile in the current tilegroup
+  int tg_size_bit_offset;
+#endif
+#if CONFIG_REFERENCE_BUFFER
+  SequenceHeader seq_params;
+#endif
+#if CONFIG_INSPECTION
+  aom_inspect_cb inspect_cb;
+  void *inspect_ctx;
+#endif
+} AV1Decoder;
+
+int av1_receive_compressed_data(struct AV1Decoder *pbi, size_t size,
+                                const uint8_t **dest);
+
+int av1_get_raw_frame(struct AV1Decoder *pbi, YV12_BUFFER_CONFIG *sd);
+
+int av1_get_frame_to_show(struct AV1Decoder *pbi, YV12_BUFFER_CONFIG *frame);
+
+aom_codec_err_t av1_copy_reference_dec(struct AV1Decoder *pbi,
+                                       AOM_REFFRAME ref_frame_flag,
+                                       YV12_BUFFER_CONFIG *sd);
+
+aom_codec_err_t av1_set_reference_dec(AV1_COMMON *cm,
+                                      AOM_REFFRAME ref_frame_flag,
+                                      YV12_BUFFER_CONFIG *sd);
+
+static INLINE uint8_t read_marker(aom_decrypt_cb decrypt_cb,
+                                  void *decrypt_state, const uint8_t *data) {
+  if (decrypt_cb) {
+    uint8_t marker;
+    decrypt_cb(decrypt_state, data, &marker, 1);
+    return marker;
+  }
+  return *data;
+}
+
+// This function is exposed for use in tests, as well as the inlined function
+// "read_marker".
+aom_codec_err_t av1_parse_superframe_index(const uint8_t *data, size_t data_sz,
+                                           uint32_t sizes[8], int *count,
+                                           aom_decrypt_cb decrypt_cb,
+                                           void *decrypt_state);
+
+struct AV1Decoder *av1_decoder_create(BufferPool *const pool);
+
+void av1_decoder_remove(struct AV1Decoder *pbi);
+
+static INLINE void decrease_ref_count(int idx, RefCntBuffer *const frame_bufs,
+                                      BufferPool *const pool) {
+  if (idx >= 0) {
+    --frame_bufs[idx].ref_count;
+    // A worker may only get a free framebuffer index when calling get_free_fb.
+    // But the private buffer is not set up until finish decoding header.
+    // So any error happens during decoding header, the frame_bufs will not
+    // have valid priv buffer.
+    if (frame_bufs[idx].ref_count == 0 &&
+        frame_bufs[idx].raw_frame_buffer.priv) {
+      pool->release_fb_cb(pool->cb_priv, &frame_bufs[idx].raw_frame_buffer);
+    }
+  }
+}
+
+#if CONFIG_EXT_REFS
+static INLINE int dec_is_ref_frame_buf(AV1Decoder *const pbi,
+                                       RefCntBuffer *frame_buf) {
+  AV1_COMMON *const cm = &pbi->common;
+  int i;
+  for (i = 0; i < INTER_REFS_PER_FRAME; ++i) {
+    RefBuffer *const ref_frame = &cm->frame_refs[i];
+    if (ref_frame->idx == INVALID_IDX) continue;
+    if (frame_buf == &cm->buffer_pool->frame_bufs[ref_frame->idx]) break;
+  }
+  return (i < INTER_REFS_PER_FRAME);
+}
+#endif  // CONFIG_EXT_REFS
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_DECODER_DECODER_H_
diff --git a/third_party/aom/av1/decoder/decodetxb.c b/third_party/aom/av1/decoder/decodetxb.c
new file mode 100644
index 0000000000..e1db09775a
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodetxb.c
@@ -0,0 +1,286 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/scan.h"
+#include "av1/common/idct.h"
+#include "av1/common/txb_common.h"
+#include "av1/decoder/decodemv.h"
+#include "av1/decoder/decodetxb.h"
+#include "av1/decoder/dsubexp.h"
+
+#define ACCT_STR __func__
+
+static int read_golomb(MACROBLOCKD *xd, aom_reader *r) {
+  int x = 1;
+  int length = 0;
+  int i = 0;
+
+  while (!i) {
+    i = aom_read_bit(r, ACCT_STR);
+    ++length;
+    if (length >= 32) {
+      aom_internal_error(xd->error_info, AOM_CODEC_CORRUPT_FRAME,
+                         "Invalid length in read_golomb");
+      break;
+    }
+  }
+
+  for (i = 0; i < length - 1; ++i) {
+    x <<= 1;
+    x += aom_read_bit(r, ACCT_STR);
+  }
+
+  return x - 1;
+}
+
+uint8_t av1_read_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+                            aom_reader *r, int block, int plane,
+                            tran_low_t *tcoeffs, TXB_CTX *txb_ctx,
+                            int16_t *max_scan_line, int *eob) {
+  FRAME_COUNTS *counts = xd->counts;
+  TX_SIZE tx_size = get_tx_size(plane, xd);
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  aom_prob *nz_map = cm->fc->nz_map[tx_size][plane_type];
+  aom_prob *eob_flag = cm->fc->eob_flag[tx_size][plane_type];
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const int seg_eob = tx_size_2d[tx_size];
+  int c = 0;
+  int update_eob = -1;
+  const int16_t *const dequant = xd->plane[plane].seg_dequant[mbmi->segment_id];
+  const int shift = av1_get_tx_scale(tx_size);
+  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
+  int cul_level = 0;
+  unsigned int(*nz_map_count)[SIG_COEF_CONTEXTS][2];
+  uint8_t txb_mask[32 * 32] = { 0 };
+
+  nz_map_count = (counts) ? &counts->nz_map[tx_size][plane_type] : NULL;
+
+  memset(tcoeffs, 0, sizeof(*tcoeffs) * seg_eob);
+
+  int all_zero =
+      aom_read(r, cm->fc->txb_skip[tx_size][txb_ctx->txb_skip_ctx], ACCT_STR);
+  if (xd->counts)
+    ++xd->counts->txb_skip[tx_size][txb_ctx->txb_skip_ctx][all_zero];
+
+  *eob = 0;
+  if (all_zero) {
+    *max_scan_line = 0;
+    return 0;
+  }
+
+#if CONFIG_TXK_SEL
+  av1_read_tx_type(cm, xd, block, plane, r);
+#endif
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  const int16_t *scan = scan_order->scan;
+
+  for (c = 0; c < seg_eob; ++c) {
+    int is_nz;
+    int coeff_ctx = get_nz_map_ctx(tcoeffs, txb_mask, scan[c], bwl);
+    int eob_ctx = get_eob_ctx(tcoeffs, scan[c], bwl);
+
+    if (c < seg_eob - 1)
+      is_nz = aom_read(r, nz_map[coeff_ctx], tx_size);
+    else
+      is_nz = 1;
+
+    // set non-zero coefficient map.
+    tcoeffs[scan[c]] = is_nz;
+
+    if (c == seg_eob - 1) {
+      ++c;
+      break;
+    }
+
+    if (counts) ++(*nz_map_count)[coeff_ctx][is_nz];
+
+    if (is_nz) {
+      int is_eob = aom_read(r, eob_flag[eob_ctx], tx_size);
+      if (counts) ++counts->eob_flag[tx_size][plane_type][eob_ctx][is_eob];
+      if (is_eob) break;
+    }
+    txb_mask[scan[c]] = 1;
+  }
+
+  *eob = AOMMIN(seg_eob, c + 1);
+  *max_scan_line = *eob;
+
+  int i;
+  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+    aom_prob *coeff_base = cm->fc->coeff_base[tx_size][plane_type][i];
+
+    update_eob = 0;
+    for (c = *eob - 1; c >= 0; --c) {
+      tran_low_t *v = &tcoeffs[scan[c]];
+      int sign;
+      int ctx;
+
+      if (*v <= i) continue;
+
+      ctx = get_base_ctx(tcoeffs, scan[c], bwl, i + 1);
+
+      if (aom_read(r, coeff_base[ctx], tx_size)) {
+        *v = i + 1;
+        cul_level += i + 1;
+
+        if (counts) ++counts->coeff_base[tx_size][plane_type][i][ctx][1];
+
+        if (c == 0) {
+          int dc_sign_ctx = txb_ctx->dc_sign_ctx;
+          sign = aom_read(r, cm->fc->dc_sign[plane_type][dc_sign_ctx], tx_size);
+          if (counts) ++counts->dc_sign[plane_type][dc_sign_ctx][sign];
+        } else {
+          sign = aom_read_bit(r, ACCT_STR);
+        }
+        if (sign) *v = -(*v);
+        continue;
+      }
+      *v = i + 2;
+      if (counts) ++counts->coeff_base[tx_size][plane_type][i][ctx][0];
+
+      // update the eob flag for coefficients with magnitude above 1.
+      update_eob = AOMMAX(update_eob, c);
+    }
+  }
+
+  for (c = update_eob; c >= 0; --c) {
+    tran_low_t *v = &tcoeffs[scan[c]];
+    int sign;
+    int idx;
+    int ctx;
+
+    if (*v <= NUM_BASE_LEVELS) continue;
+
+    if (c == 0) {
+      int dc_sign_ctx = txb_ctx->dc_sign_ctx;
+      sign = aom_read(r, cm->fc->dc_sign[plane_type][dc_sign_ctx], tx_size);
+      if (counts) ++counts->dc_sign[plane_type][dc_sign_ctx][sign];
+    } else {
+      sign = aom_read_bit(r, ACCT_STR);
+    }
+
+    ctx = get_level_ctx(tcoeffs, scan[c], bwl);
+
+    if (cm->fc->coeff_lps[tx_size][plane_type][ctx] == 0) exit(0);
+
+    for (idx = 0; idx < COEFF_BASE_RANGE; ++idx) {
+      if (aom_read(r, cm->fc->coeff_lps[tx_size][plane_type][ctx], tx_size)) {
+        *v = (idx + 1 + NUM_BASE_LEVELS);
+        if (sign) *v = -(*v);
+        cul_level += abs(*v);
+
+        if (counts) ++counts->coeff_lps[tx_size][plane_type][ctx][1];
+        break;
+      }
+      if (counts) ++counts->coeff_lps[tx_size][plane_type][ctx][0];
+    }
+    if (idx < COEFF_BASE_RANGE) continue;
+
+    // decode 0-th order Golomb code
+    *v = read_golomb(xd, r) + COEFF_BASE_RANGE + 1 + NUM_BASE_LEVELS;
+    if (sign) *v = -(*v);
+    cul_level += abs(*v);
+  }
+
+  for (c = 0; c < *eob; ++c) {
+    int16_t dqv = (c == 0) ? dequant[0] : dequant[1];
+    tran_low_t *v = &tcoeffs[scan[c]];
+    int sign = (*v) < 0;
+    *v = (abs(*v) * dqv) >> shift;
+    if (sign) *v = -(*v);
+  }
+
+  cul_level = AOMMIN(63, cul_level);
+
+  // DC value
+  set_dc_sign(&cul_level, tcoeffs[0]);
+
+  return cul_level;
+}
+
+uint8_t av1_read_coeffs_txb_facade(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                   aom_reader *r, int row, int col, int block,
+                                   int plane, tran_low_t *tcoeffs,
+                                   int16_t *max_scan_line, int *eob) {
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct macroblockd_plane *pd = &xd->plane[plane];
+
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+  const BLOCK_SIZE plane_bsize =
+      AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#endif  // CONFIG_CHROMA_2X2
+#else   // CONFIG_CB4X4
+  const BLOCK_SIZE plane_bsize =
+      get_plane_block_size(AOMMAX(BLOCK_8X8, bsize), pd);
+#endif  // CONFIG_CB4X4
+
+  TX_SIZE tx_size = get_tx_size(plane, xd);
+  TXB_CTX txb_ctx;
+  get_txb_ctx(plane_bsize, tx_size, plane, pd->above_context + col,
+              pd->left_context + row, &txb_ctx);
+  uint8_t cul_level = av1_read_coeffs_txb(cm, xd, r, block, plane, tcoeffs,
+                                          &txb_ctx, max_scan_line, eob);
+#if CONFIG_ADAPT_SCAN
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  if (xd->counts && *eob > 0)
+    av1_update_scan_count_facade(cm, xd->counts, tx_size, tx_type, pd->dqcoeff,
+                                 *eob);
+#endif
+  av1_set_contexts(xd, pd, plane, tx_size, cul_level, col, row);
+  return cul_level;
+}
+
+static void read_txb_probs(FRAME_CONTEXT *fc, const TX_SIZE tx_size,
+                           aom_reader *r) {
+  int plane, ctx, level;
+
+  if (aom_read_bit(r, ACCT_STR) == 0) return;
+
+  for (ctx = 0; ctx < TXB_SKIP_CONTEXTS; ++ctx)
+    av1_diff_update_prob(r, &fc->txb_skip[tx_size][ctx], ACCT_STR);
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < SIG_COEF_CONTEXTS; ++ctx)
+      av1_diff_update_prob(r, &fc->nz_map[tx_size][plane][ctx], ACCT_STR);
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < EOB_COEF_CONTEXTS; ++ctx)
+      av1_diff_update_prob(r, &fc->eob_flag[tx_size][plane][ctx], ACCT_STR);
+
+  for (level = 0; level < NUM_BASE_LEVELS; ++level)
+    for (plane = 0; plane < PLANE_TYPES; ++plane)
+      for (ctx = 0; ctx < COEFF_BASE_CONTEXTS; ++ctx)
+        av1_diff_update_prob(r, &fc->coeff_base[tx_size][plane][level][ctx],
+                             ACCT_STR);
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx)
+      av1_diff_update_prob(r, &fc->coeff_lps[tx_size][plane][ctx], ACCT_STR);
+}
+
+void av1_read_txb_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode, aom_reader *r) {
+  const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+  TX_SIZE tx_size;
+  int ctx, plane;
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < DC_SIGN_CONTEXTS; ++ctx)
+      av1_diff_update_prob(r, &fc->dc_sign[plane][ctx], ACCT_STR);
+
+  for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
+    read_txb_probs(fc, tx_size, r);
+}
diff --git a/third_party/aom/av1/decoder/decodetxb.h b/third_party/aom/av1/decoder/decodetxb.h
new file mode 100644
index 0000000000..ee1bf6a3df
--- /dev/null
+++ b/third_party/aom/av1/decoder/decodetxb.h
@@ -0,0 +1,31 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef DECODETXB_H_
+#define DECODETXB_H_
+
+#include "./aom_config.h"
+#include "av1/common/blockd.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/txb_common.h"
+#include "aom_dsp/bitreader.h"
+
+uint8_t av1_read_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+                            aom_reader *r, int block, int plane,
+                            tran_low_t *tcoeffs, TXB_CTX *txb_ctx,
+                            int16_t *max_scan_line, int *eob);
+
+uint8_t av1_read_coeffs_txb_facade(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                   aom_reader *r, int row, int col, int block,
+                                   int plane, tran_low_t *tcoeffs,
+                                   int16_t *max_scan_line, int *eob);
+void av1_read_txb_probs(FRAME_CONTEXT *fc, TX_MODE tx_mode, aom_reader *r);
+#endif  //  DECODETXB_H_
diff --git a/third_party/aom/av1/decoder/detokenize.c b/third_party/aom/av1/decoder/detokenize.c
new file mode 100644
index 0000000000..494f1681fc
--- /dev/null
+++ b/third_party/aom/av1/decoder/detokenize.c
@@ -0,0 +1,467 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#if !CONFIG_PVQ
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#endif  // !CONFIG_PVQ
+
+#include "av1/common/blockd.h"
+
+#define ACCT_STR __func__
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+#include "av1/common/common.h"
+#include "av1/common/entropy.h"
+#include "av1/common/idct.h"
+#include "av1/decoder/detokenize.h"
+
+#define EOB_CONTEXT_NODE 0
+#define ZERO_CONTEXT_NODE 1
+#define ONE_CONTEXT_NODE 2
+#define LOW_VAL_CONTEXT_NODE 0
+#define TWO_CONTEXT_NODE 1
+#define THREE_CONTEXT_NODE 2
+#define HIGH_LOW_CONTEXT_NODE 3
+#define CAT_ONE_CONTEXT_NODE 4
+#define CAT_THREEFOUR_CONTEXT_NODE 5
+#define CAT_THREE_CONTEXT_NODE 6
+#define CAT_FIVE_CONTEXT_NODE 7
+
+#define INCREMENT_COUNT(token)                   \
+  do {                                           \
+    if (counts) ++coef_counts[band][ctx][token]; \
+  } while (0)
+
+#if CONFIG_NEW_MULTISYMBOL
+#define READ_COEFF(prob_name, cdf_name, num, r) read_coeff(cdf_name, num, r);
+static INLINE int read_coeff(const aom_cdf_prob *const *cdf, int n,
+                             aom_reader *r) {
+  int val = 0;
+  int i = 0;
+  int count = 0;
+  while (count < n) {
+    const int size = AOMMIN(n - count, 4);
+    val |= aom_read_cdf(r, cdf[i++], 1 << size, ACCT_STR) << count;
+    count += size;
+  }
+  return val;
+}
+#else
+#define READ_COEFF(prob_name, cdf_name, num, r) read_coeff(prob_name, num, r);
+static INLINE int read_coeff(const aom_prob *probs, int n, aom_reader *r) {
+  int i, val = 0;
+  for (i = 0; i < n; ++i) val = (val << 1) | aom_read(r, probs[i], ACCT_STR);
+  return val;
+}
+
+#endif
+
+static int token_to_value(aom_reader *const r, int token, TX_SIZE tx_size,
+                          int bit_depth) {
+#if !CONFIG_HIGHBITDEPTH
+  assert(bit_depth == 8);
+#endif  // !CONFIG_HIGHBITDEPTH
+
+  switch (token) {
+    case ZERO_TOKEN:
+    case ONE_TOKEN:
+    case TWO_TOKEN:
+    case THREE_TOKEN:
+    case FOUR_TOKEN: return token;
+    case CATEGORY1_TOKEN:
+      return CAT1_MIN_VAL + READ_COEFF(av1_cat1_prob, av1_cat1_cdf, 1, r);
+    case CATEGORY2_TOKEN:
+      return CAT2_MIN_VAL + READ_COEFF(av1_cat2_prob, av1_cat2_cdf, 2, r);
+    case CATEGORY3_TOKEN:
+      return CAT3_MIN_VAL + READ_COEFF(av1_cat3_prob, av1_cat3_cdf, 3, r);
+    case CATEGORY4_TOKEN:
+      return CAT4_MIN_VAL + READ_COEFF(av1_cat4_prob, av1_cat4_cdf, 4, r);
+    case CATEGORY5_TOKEN:
+      return CAT5_MIN_VAL + READ_COEFF(av1_cat5_prob, av1_cat5_cdf, 5, r);
+    case CATEGORY6_TOKEN: {
+      const int skip_bits = (int)sizeof(av1_cat6_prob) -
+                            av1_get_cat6_extrabits_size(tx_size, bit_depth);
+      return CAT6_MIN_VAL + READ_COEFF(av1_cat6_prob + skip_bits, av1_cat6_cdf,
+                                       18 - skip_bits, r);
+    }
+    default:
+      assert(0);  // Invalid token.
+      return -1;
+  }
+}
+
+static int decode_coefs(MACROBLOCKD *xd, PLANE_TYPE type, tran_low_t *dqcoeff,
+                        TX_SIZE tx_size, TX_TYPE tx_type, const int16_t *dq,
+#if CONFIG_NEW_QUANT
+                        dequant_val_type_nuq *dq_val,
+#endif  // CONFIG_NEW_QUANT
+#if CONFIG_AOM_QM
+                        const qm_val_t *iqm[2][TX_SIZES],
+#endif  // CONFIG_AOM_QM
+                        int ctx, const int16_t *scan, const int16_t *nb,
+                        int16_t *max_scan_line, aom_reader *r) {
+  FRAME_COUNTS *counts = xd->counts;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *const ec_ctx = xd->fc;
+#endif
+  const int max_eob = tx_size_2d[tx_size];
+  const int ref = is_inter_block(&xd->mi[0]->mbmi);
+#if CONFIG_AOM_QM
+  const qm_val_t *iqmatrix = iqm[!ref][tx_size];
+#endif  // CONFIG_AOM_QM
+  int band, c = 0;
+  const int tx_size_ctx = txsize_sqr_map[tx_size];
+#if CONFIG_NEW_TOKENSET
+  aom_cdf_prob(*coef_head_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_head_cdfs[tx_size_ctx][type][ref];
+  aom_cdf_prob(*coef_tail_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_tail_cdfs[tx_size_ctx][type][ref];
+  int val = 0;
+
+#if !CONFIG_EC_ADAPT
+  unsigned int *blockz_count;
+  unsigned int(*coef_counts)[COEFF_CONTEXTS][UNCONSTRAINED_NODES + 1] = NULL;
+  unsigned int(*eob_branch_count)[COEFF_CONTEXTS] = NULL;
+#endif
+#else
+  aom_prob(*coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
+      ec_ctx->coef_probs[tx_size_ctx][type][ref];
+  const aom_prob *prob;
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob(*coef_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_cdfs[tx_size_ctx][type][ref];
+  aom_cdf_prob(*cdf)[CDF_SIZE(ENTROPY_TOKENS)];
+#endif  // CONFIG_EC_MULTISYMBOL
+  unsigned int(*coef_counts)[COEFF_CONTEXTS][UNCONSTRAINED_NODES + 1] = NULL;
+  unsigned int(*eob_branch_count)[COEFF_CONTEXTS] = NULL;
+#endif  // CONFIG_NEW_TOKENSET
+  uint8_t token_cache[MAX_TX_SQUARE];
+  const uint8_t *band_translate = get_band_translate(tx_size);
+  int dq_shift;
+  int v, token;
+  int16_t dqv = dq[0];
+#if CONFIG_NEW_QUANT
+  const tran_low_t *dqv_val = &dq_val[0][0];
+#endif  // CONFIG_NEW_QUANT
+  (void)tx_type;
+#if CONFIG_AOM_QM
+  (void)iqmatrix;
+#endif  // CONFIG_AOM_QM
+
+  if (counts) {
+#if !CONFIG_NEW_TOKENSET || !CONFIG_EC_ADAPT
+    coef_counts = counts->coef[tx_size_ctx][type][ref];
+    eob_branch_count = counts->eob_branch[tx_size_ctx][type][ref];
+#endif
+#if CONFIG_NEW_TOKENSET && !CONFIG_EC_ADAPT
+    blockz_count = counts->blockz_count[tx_size_ctx][type][ref][ctx];
+#endif
+  }
+
+  dq_shift = av1_get_tx_scale(tx_size);
+
+#if CONFIG_NEW_TOKENSET
+  band = *band_translate++;
+
+  int more_data = 1;
+  while (more_data) {
+    int comb_token;
+    int last_pos = (c + 1 == max_eob);
+    int first_pos = (c == 0);
+
+#if CONFIG_NEW_QUANT
+    dqv_val = &dq_val[band][0];
+#endif  // CONFIG_NEW_QUANT
+
+    comb_token = last_pos ? 2 * aom_read_bit(r, ACCT_STR) + 2
+                          : aom_read_symbol(r, coef_head_cdfs[band][ctx],
+                                            HEAD_TOKENS + first_pos, ACCT_STR) +
+                                !first_pos;
+    if (first_pos) {
+#if !CONFIG_EC_ADAPT
+      if (counts) ++blockz_count[comb_token != 0];
+#endif
+      if (comb_token == 0) return 0;
+    }
+    token = comb_token >> 1;
+
+    while (!token) {
+      *max_scan_line = AOMMAX(*max_scan_line, scan[c]);
+      token_cache[scan[c]] = 0;
+#if !CONFIG_EC_ADAPT
+      if (counts && !last_pos) {
+        ++coef_counts[band][ctx][ZERO_TOKEN];
+      }
+#endif
+      ++c;
+      dqv = dq[1];
+      ctx = get_coef_context(nb, token_cache, c);
+      band = *band_translate++;
+
+      last_pos = (c + 1 == max_eob);
+
+      comb_token = last_pos ? 2 * aom_read_bit(r, ACCT_STR) + 2
+                            : aom_read_symbol(r, coef_head_cdfs[band][ctx],
+                                              HEAD_TOKENS, ACCT_STR) +
+                                  1;
+      token = comb_token >> 1;
+    }
+
+    more_data = comb_token & 1;
+#if !CONFIG_EC_ADAPT
+    if (counts && !last_pos) {
+      ++coef_counts[band][ctx][token];
+      ++eob_branch_count[band][ctx];
+      if (!more_data) ++coef_counts[band][ctx][EOB_MODEL_TOKEN];
+    }
+#endif
+
+    if (token > ONE_TOKEN)
+      token +=
+          aom_read_symbol(r, coef_tail_cdfs[band][ctx], TAIL_TOKENS, ACCT_STR);
+#if CONFIG_NEW_QUANT
+    dqv_val = &dq_val[band][0];
+#endif  // CONFIG_NEW_QUANT
+
+    *max_scan_line = AOMMAX(*max_scan_line, scan[c]);
+    token_cache[scan[c]] = av1_pt_energy_class[token];
+
+    val = token_to_value(r, token, tx_size,
+#if CONFIG_HIGHBITDEPTH
+                         xd->bd);
+#else
+                         8);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_NEW_QUANT
+    v = av1_dequant_abscoeff_nuq(val, dqv, dqv_val);
+    v = dq_shift ? ROUND_POWER_OF_TWO(v, dq_shift) : v;
+#else
+#if CONFIG_AOM_QM
+    dqv = ((iqmatrix[scan[c]] * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >>
+          AOM_QM_BITS;
+#endif
+    v = (val * dqv) >> dq_shift;
+#endif
+
+    v = aom_read_bit(r, ACCT_STR) ? -v : v;
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+#if CONFIG_HIGHBITDEPTH
+    check_range(v, xd->bd);
+#else
+    check_range(v, 8);
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+
+    dqcoeff[scan[c]] = v;
+
+    ++c;
+    more_data &= (c < max_eob);
+    if (!more_data) break;
+    dqv = dq[1];
+    ctx = get_coef_context(nb, token_cache, c);
+    band = *band_translate++;
+
+#else  // CONFIG_NEW_TOKENSET
+  while (c < max_eob) {
+    int val = -1;
+    band = *band_translate++;
+    prob = coef_probs[band][ctx];
+    if (counts) ++eob_branch_count[band][ctx];
+    if (!aom_read(r, prob[EOB_CONTEXT_NODE], ACCT_STR)) {
+      INCREMENT_COUNT(EOB_MODEL_TOKEN);
+      break;
+    }
+
+#if CONFIG_NEW_QUANT
+    dqv_val = &dq_val[band][0];
+#endif  // CONFIG_NEW_QUANT
+
+    while (!aom_read(r, prob[ZERO_CONTEXT_NODE], ACCT_STR)) {
+      INCREMENT_COUNT(ZERO_TOKEN);
+      dqv = dq[1];
+      token_cache[scan[c]] = 0;
+      ++c;
+      if (c >= max_eob) return c;  // zero tokens at the end (no eob token)
+      ctx = get_coef_context(nb, token_cache, c);
+      band = *band_translate++;
+      prob = coef_probs[band][ctx];
+#if CONFIG_NEW_QUANT
+      dqv_val = &dq_val[band][0];
+#endif  // CONFIG_NEW_QUANT
+    }
+
+    *max_scan_line = AOMMAX(*max_scan_line, scan[c]);
+
+#if CONFIG_EC_MULTISYMBOL
+    cdf = &coef_cdfs[band][ctx];
+    token = ONE_TOKEN +
+            aom_read_symbol(r, *cdf, CATEGORY6_TOKEN - ONE_TOKEN + 1, ACCT_STR);
+    INCREMENT_COUNT(ONE_TOKEN + (token > ONE_TOKEN));
+    assert(token != ZERO_TOKEN);
+    val = token_to_value(r, token, tx_size,
+#if CONFIG_HIGHBITDEPTH
+                         xd->bd);
+#else
+                         8);
+#endif  // CONFIG_HIGHBITDEPTH
+#else   // CONFIG_EC_MULTISYMBOL
+    if (!aom_read(r, prob[ONE_CONTEXT_NODE], ACCT_STR)) {
+      INCREMENT_COUNT(ONE_TOKEN);
+      token = ONE_TOKEN;
+      val = 1;
+    } else {
+      INCREMENT_COUNT(TWO_TOKEN);
+      token = aom_read_tree(r, av1_coef_con_tree,
+                            av1_pareto8_full[prob[PIVOT_NODE] - 1], ACCT_STR);
+      assert(token != ZERO_TOKEN && token != ONE_TOKEN);
+      val = token_to_value(r, token, tx_size,
+#if CONFIG_HIGHBITDEPTH
+                           xd->bd);
+#else
+                           8);
+#endif  // CONFIG_HIGHBITDEPTH
+    }
+#endif  // CONFIG_EC_MULTISYMBOL
+#if CONFIG_NEW_QUANT
+    v = av1_dequant_abscoeff_nuq(val, dqv, dqv_val);
+    v = dq_shift ? ROUND_POWER_OF_TWO(v, dq_shift) : v;
+#else
+#if CONFIG_AOM_QM
+    dqv = ((iqmatrix[scan[c]] * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >>
+          AOM_QM_BITS;
+#endif
+    v = (val * dqv) >> dq_shift;
+#endif  // CONFIG_NEW_QUANT
+
+#if CONFIG_COEFFICIENT_RANGE_CHECKING
+#if CONFIG_HIGHBITDEPTH
+    dqcoeff[scan[c]] =
+        highbd_check_range((aom_read_bit(r, ACCT_STR) ? -v : v), xd->bd);
+#else
+    dqcoeff[scan[c]] = check_range(aom_read_bit(r, ACCT_STR) ? -v : v, 8);
+#endif  // CONFIG_HIGHBITDEPTH
+#else
+    dqcoeff[scan[c]] = aom_read_bit(r, ACCT_STR) ? -v : v;
+#endif  // CONFIG_COEFFICIENT_RANGE_CHECKING
+    token_cache[scan[c]] = av1_pt_energy_class[token];
+    ++c;
+    ctx = get_coef_context(nb, token_cache, c);
+    dqv = dq[1];
+#endif  // CONFIG_NEW_TOKENSET
+  }
+
+  return c;
+}
+#endif  // !CONFIG_PVQ
+
+#if CONFIG_PALETTE
+void av1_decode_palette_tokens(MACROBLOCKD *const xd, int plane,
+                               aom_reader *r) {
+  const MODE_INFO *const mi = xd->mi[0];
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  uint8_t color_order[PALETTE_MAX_SIZE];
+  const int n = mbmi->palette_mode_info.palette_size[plane];
+  int i, j;
+  uint8_t *const color_map = xd->plane[plane].color_index_map;
+  const aom_prob(
+      *const prob)[PALETTE_COLOR_INDEX_CONTEXTS][PALETTE_COLORS - 1] =
+      plane ? av1_default_palette_uv_color_index_prob
+            : av1_default_palette_y_color_index_prob;
+  int plane_block_width, plane_block_height, rows, cols;
+  av1_get_block_dimensions(mbmi->sb_type, plane, xd, &plane_block_width,
+                           &plane_block_height, &rows, &cols);
+  assert(plane == 0 || plane == 1);
+
+#if CONFIG_PALETTE_THROUGHPUT
+  // Run wavefront on the palette map index decoding.
+  for (i = 1; i < rows + cols - 1; ++i) {
+    for (j = AOMMIN(i, cols - 1); j >= AOMMAX(0, i - rows + 1); --j) {
+      const int color_ctx = av1_get_palette_color_index_context(
+          color_map, plane_block_width, (i - j), j, n, color_order, NULL);
+      const int color_idx =
+          aom_read_tree(r, av1_palette_color_index_tree[n - 2],
+                        prob[n - 2][color_ctx], ACCT_STR);
+      assert(color_idx >= 0 && color_idx < n);
+      color_map[(i - j) * plane_block_width + j] = color_order[color_idx];
+    }
+  }
+  // Copy last column to extra columns.
+  if (cols < plane_block_width) {
+    for (i = 0; i < plane_block_height; ++i) {
+      memset(color_map + i * plane_block_width + cols,
+             color_map[i * plane_block_width + cols - 1],
+             (plane_block_width - cols));
+    }
+  }
+#else
+  for (i = 0; i < rows; ++i) {
+    for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
+      const int color_ctx = av1_get_palette_color_index_context(
+          color_map, plane_block_width, i, j, n, color_order, NULL);
+      const int color_idx =
+          aom_read_tree(r, av1_palette_color_index_tree[n - PALETTE_MIN_SIZE],
+                        prob[n - PALETTE_MIN_SIZE][color_ctx], ACCT_STR);
+      assert(color_idx >= 0 && color_idx < n);
+      color_map[i * plane_block_width + j] = color_order[color_idx];
+    }
+    memset(color_map + i * plane_block_width + cols,
+           color_map[i * plane_block_width + cols - 1],
+           (plane_block_width - cols));  // Copy last column to extra columns.
+  }
+#endif  // CONFIG_PALETTE_THROUGHPUT
+  // Copy last row to extra rows.
+  for (i = rows; i < plane_block_height; ++i) {
+    memcpy(color_map + i * plane_block_width,
+           color_map + (rows - 1) * plane_block_width, plane_block_width);
+  }
+}
+#endif  // CONFIG_PALETTE
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+int av1_decode_block_tokens(AV1_COMMON *cm, MACROBLOCKD *const xd, int plane,
+                            const SCAN_ORDER *sc, int x, int y, TX_SIZE tx_size,
+                            TX_TYPE tx_type, int16_t *max_scan_line,
+                            aom_reader *r, int seg_id) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int16_t *const dequant = pd->seg_dequant[seg_id];
+  const int ctx =
+      get_entropy_context(tx_size, pd->above_context + x, pd->left_context + y);
+#if CONFIG_NEW_QUANT
+  const int ref = is_inter_block(&xd->mi[0]->mbmi);
+  int dq =
+      get_dq_profile_from_ctx(xd->qindex[seg_id], ctx, ref, pd->plane_type);
+#endif  //  CONFIG_NEW_QUANT
+
+  const int eob =
+      decode_coefs(xd, pd->plane_type, pd->dqcoeff, tx_size, tx_type, dequant,
+#if CONFIG_NEW_QUANT
+                   pd->seg_dequant_nuq[seg_id][dq],
+#endif  // CONFIG_NEW_QUANT
+#if CONFIG_AOM_QM
+                   pd->seg_iqmatrix[seg_id],
+#endif  // CONFIG_AOM_QM
+                   ctx, sc->scan, sc->neighbors, max_scan_line, r);
+  av1_set_contexts(xd, pd, plane, tx_size, eob > 0, x, y);
+#if CONFIG_ADAPT_SCAN
+  if (xd->counts)
+    av1_update_scan_count_facade(cm, xd->counts, tx_size, tx_type, pd->dqcoeff,
+                                 eob);
+#else
+  (void)cm;
+#endif
+  return eob;
+}
+#endif  // !CONFIG_PVQ
diff --git a/third_party/aom/av1/decoder/detokenize.h b/third_party/aom/av1/decoder/detokenize.h
new file mode 100644
index 0000000000..ba40666037
--- /dev/null
+++ b/third_party/aom/av1/decoder/detokenize.h
@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DETOKENIZE_H_
+#define AV1_DECODER_DETOKENIZE_H_
+
+#include "./aom_config.h"
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+#include "av1/decoder/decoder.h"
+#include "av1/common/scan.h"
+#endif  // !CONFIG_PVQ
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_PALETTE
+void av1_decode_palette_tokens(MACROBLOCKD *const xd, int plane, aom_reader *r);
+#endif  // CONFIG_PALETTE
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+int av1_decode_block_tokens(AV1_COMMON *cm, MACROBLOCKD *const xd, int plane,
+                            const SCAN_ORDER *sc, int x, int y, TX_SIZE tx_size,
+                            TX_TYPE tx_type, int16_t *max_scan_line,
+                            aom_reader *r, int seg_id);
+#endif  // !CONFIG_PVQ
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+#endif  // AV1_DECODER_DETOKENIZE_H_
diff --git a/third_party/aom/av1/decoder/dsubexp.c b/third_party/aom/av1/decoder/dsubexp.c
new file mode 100644
index 0000000000..5171f1144e
--- /dev/null
+++ b/third_party/aom/av1/decoder/dsubexp.c
@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/entropy.h"
+
+#include "av1/decoder/dsubexp.h"
+
+static int inv_recenter_nonneg(int v, int m) {
+  if (v > 2 * m) return v;
+
+  return (v & 1) ? m - ((v + 1) >> 1) : m + (v >> 1);
+}
+
+#define decode_uniform(r, ACCT_STR_NAME) \
+  decode_uniform_(r ACCT_STR_ARG(ACCT_STR_NAME))
+#define decode_term_subexp(r, ACCT_STR_NAME) \
+  decode_term_subexp_(r ACCT_STR_ARG(ACCT_STR_NAME))
+
+static int decode_uniform_(aom_reader *r ACCT_STR_PARAM) {
+  const int l = 8;
+  const int m = (1 << l) - 190;
+  const int v = aom_read_literal(r, l - 1, ACCT_STR_NAME);
+  return v < m ? v : (v << 1) - m + aom_read_bit(r, ACCT_STR_NAME);
+}
+
+static int inv_remap_prob(int v, int m) {
+  /* clang-format off */
+  static uint8_t inv_map_table[MAX_PROB - 1] = {
+      7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176, 189,
+    202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,  10,  11,
+     12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,  25,  26,  27,
+     28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,  40,  41,  42,  43,
+     44,  45,  47,  48,  49,  50,  51,  52,  53,  54,  55,  56,  57,  58,  60,
+     61,  62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  73,  74,  75,  76,
+     77,  78,  79,  80,  81,  82,  83,  84,  86,  87,  88,  89,  90,  91,  92,
+     93,  94,  95,  96,  97,  99, 100, 101, 102, 103, 104, 105, 106, 107, 108,
+    109, 110, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 125,
+    126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 138, 139, 140, 141,
+    142, 143, 144, 145, 146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157,
+    158, 159, 160, 161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173,
+    174, 175, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190,
+    191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,
+    207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221, 222,
+    223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236, 237, 238,
+    239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253
+  }; /* clang-format on */
+  assert(v < (int)(sizeof(inv_map_table) / sizeof(inv_map_table[0])));
+  v = inv_map_table[v];
+  m--;
+  if ((m << 1) <= MAX_PROB) {
+    return 1 + inv_recenter_nonneg(v, m);
+  } else {
+    return MAX_PROB - inv_recenter_nonneg(v, MAX_PROB - 1 - m);
+  }
+}
+
+static int decode_term_subexp_(aom_reader *r ACCT_STR_PARAM) {
+  if (!aom_read_bit(r, ACCT_STR_NAME))
+    return aom_read_literal(r, 4, ACCT_STR_NAME);
+  if (!aom_read_bit(r, ACCT_STR_NAME))
+    return aom_read_literal(r, 4, ACCT_STR_NAME) + 16;
+  if (!aom_read_bit(r, ACCT_STR_NAME))
+    return aom_read_literal(r, 5, ACCT_STR_NAME) + 32;
+  return decode_uniform(r, ACCT_STR_NAME) + 64;
+}
+
+void av1_diff_update_prob_(aom_reader *r, aom_prob *p ACCT_STR_PARAM) {
+  if (aom_read(r, DIFF_UPDATE_PROB, ACCT_STR_NAME)) {
+    const int delp = decode_term_subexp(r, ACCT_STR_NAME);
+    *p = (aom_prob)inv_remap_prob(delp, *p);
+  }
+}
diff --git a/third_party/aom/av1/decoder/dsubexp.h b/third_party/aom/av1/decoder/dsubexp.h
new file mode 100644
index 0000000000..4bc38578c9
--- /dev/null
+++ b/third_party/aom/av1/decoder/dsubexp.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DSUBEXP_H_
+#define AV1_DECODER_DSUBEXP_H_
+
+#include "aom_dsp/bitreader.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_ACCOUNTING
+#define av1_diff_update_prob(r, p, str) av1_diff_update_prob_(r, p, str)
+#else
+#define av1_diff_update_prob(r, p, str) av1_diff_update_prob_(r, p)
+#endif
+
+void av1_diff_update_prob_(aom_reader *r, aom_prob *p ACCT_STR_PARAM);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+#endif  // AV1_DECODER_DSUBEXP_H_
diff --git a/third_party/aom/av1/decoder/dthread.c b/third_party/aom/av1/decoder/dthread.c
new file mode 100644
index 0000000000..50f8ed192c
--- /dev/null
+++ b/third_party/aom/av1/decoder/dthread.c
@@ -0,0 +1,194 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./aom_config.h"
+#include "aom_mem/aom_mem.h"
+#include "av1/common/reconinter.h"
+#include "av1/decoder/dthread.h"
+#include "av1/decoder/decoder.h"
+
+// #define DEBUG_THREAD
+
+// TODO(hkuang): Clean up all the #ifdef in this file.
+void av1_frameworker_lock_stats(AVxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  pthread_mutex_lock(&worker_data->stats_mutex);
+#else
+  (void)worker;
+#endif
+}
+
+void av1_frameworker_unlock_stats(AVxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  pthread_mutex_unlock(&worker_data->stats_mutex);
+#else
+  (void)worker;
+#endif
+}
+
+void av1_frameworker_signal_stats(AVxWorker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+
+// TODO(hkuang): Fix the pthread_cond_broadcast in windows wrapper.
+#if defined(_WIN32) && !HAVE_PTHREAD_H
+  pthread_cond_signal(&worker_data->stats_cond);
+#else
+  pthread_cond_broadcast(&worker_data->stats_cond);
+#endif
+
+#else
+  (void)worker;
+#endif
+}
+
+// This macro prevents thread_sanitizer from reporting known concurrent writes.
+#if defined(__has_feature)
+#if __has_feature(thread_sanitizer)
+#define BUILDING_WITH_TSAN
+#endif
+#endif
+
+// TODO(hkuang): Remove worker parameter as it is only used in debug code.
+void av1_frameworker_wait(AVxWorker *const worker, RefCntBuffer *const ref_buf,
+                          int row) {
+#if CONFIG_MULTITHREAD
+  if (!ref_buf) return;
+
+#ifndef BUILDING_WITH_TSAN
+  // The following line of code will get harmless tsan error but it is the key
+  // to get best performance.
+  if (ref_buf->row >= row && ref_buf->buf.corrupted != 1) return;
+#endif
+
+  {
+    // Find the worker thread that owns the reference frame. If the reference
+    // frame has been fully decoded, it may not have owner.
+    AVxWorker *const ref_worker = ref_buf->frame_worker_owner;
+    FrameWorkerData *const ref_worker_data =
+        (FrameWorkerData *)ref_worker->data1;
+    const AV1Decoder *const pbi = ref_worker_data->pbi;
+
+#ifdef DEBUG_THREAD
+    {
+      FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+      printf("%d %p worker is waiting for %d %p worker (%d)  ref %d \r\n",
+             worker_data->worker_id, worker, ref_worker_data->worker_id,
+             ref_buf->frame_worker_owner, row, ref_buf->row);
+    }
+#endif
+
+    av1_frameworker_lock_stats(ref_worker);
+    while (ref_buf->row < row && pbi->cur_buf == ref_buf &&
+           ref_buf->buf.corrupted != 1) {
+      pthread_cond_wait(&ref_worker_data->stats_cond,
+                        &ref_worker_data->stats_mutex);
+    }
+
+    if (ref_buf->buf.corrupted == 1) {
+      FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+      av1_frameworker_unlock_stats(ref_worker);
+      aom_internal_error(&worker_data->pbi->common.error,
+                         AOM_CODEC_CORRUPT_FRAME,
+                         "Worker %p failed to decode frame", worker);
+    }
+    av1_frameworker_unlock_stats(ref_worker);
+  }
+#else
+  (void)worker;
+  (void)ref_buf;
+  (void)row;
+  (void)ref_buf;
+#endif  // CONFIG_MULTITHREAD
+}
+
+void av1_frameworker_broadcast(RefCntBuffer *const buf, int row) {
+#if CONFIG_MULTITHREAD
+  AVxWorker *worker = buf->frame_worker_owner;
+
+#ifdef DEBUG_THREAD
+  {
+    FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+    printf("%d %p worker decode to (%d) \r\n", worker_data->worker_id,
+           buf->frame_worker_owner, row);
+  }
+#endif
+
+  av1_frameworker_lock_stats(worker);
+  buf->row = row;
+  av1_frameworker_signal_stats(worker);
+  av1_frameworker_unlock_stats(worker);
+#else
+  (void)buf;
+  (void)row;
+#endif  // CONFIG_MULTITHREAD
+}
+
+void av1_frameworker_copy_context(AVxWorker *const dst_worker,
+                                  AVxWorker *const src_worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const src_worker_data = (FrameWorkerData *)src_worker->data1;
+  FrameWorkerData *const dst_worker_data = (FrameWorkerData *)dst_worker->data1;
+  AV1_COMMON *const src_cm = &src_worker_data->pbi->common;
+  AV1_COMMON *const dst_cm = &dst_worker_data->pbi->common;
+  int i;
+
+  // Wait until source frame's context is ready.
+  av1_frameworker_lock_stats(src_worker);
+  while (!src_worker_data->frame_context_ready) {
+    pthread_cond_wait(&src_worker_data->stats_cond,
+                      &src_worker_data->stats_mutex);
+  }
+
+  dst_cm->last_frame_seg_map = src_cm->seg.enabled
+                                   ? src_cm->current_frame_seg_map
+                                   : src_cm->last_frame_seg_map;
+  dst_worker_data->pbi->need_resync = src_worker_data->pbi->need_resync;
+  av1_frameworker_unlock_stats(src_worker);
+
+  dst_cm->bit_depth = src_cm->bit_depth;
+#if CONFIG_HIGHBITDEPTH
+  dst_cm->use_highbitdepth = src_cm->use_highbitdepth;
+#endif
+#if CONFIG_EXT_REFS
+// TODO(zoeliu): To handle parallel decoding
+#endif  // CONFIG_EXT_REFS
+  dst_cm->prev_frame =
+      src_cm->show_existing_frame ? src_cm->prev_frame : src_cm->cur_frame;
+  dst_cm->last_width =
+      !src_cm->show_existing_frame ? src_cm->width : src_cm->last_width;
+  dst_cm->last_height =
+      !src_cm->show_existing_frame ? src_cm->height : src_cm->last_height;
+  dst_cm->subsampling_x = src_cm->subsampling_x;
+  dst_cm->subsampling_y = src_cm->subsampling_y;
+  dst_cm->frame_type = src_cm->frame_type;
+  dst_cm->last_show_frame = !src_cm->show_existing_frame
+                                ? src_cm->show_frame
+                                : src_cm->last_show_frame;
+  for (i = 0; i < REF_FRAMES; ++i)
+    dst_cm->ref_frame_map[i] = src_cm->next_ref_frame_map[i];
+
+  memcpy(dst_cm->lf_info.lfthr, src_cm->lf_info.lfthr,
+         (MAX_LOOP_FILTER + 1) * sizeof(loop_filter_thresh));
+  dst_cm->lf.last_sharpness_level = src_cm->lf.sharpness_level;
+  dst_cm->lf.filter_level = src_cm->lf.filter_level;
+  memcpy(dst_cm->lf.ref_deltas, src_cm->lf.ref_deltas, TOTAL_REFS_PER_FRAME);
+  memcpy(dst_cm->lf.mode_deltas, src_cm->lf.mode_deltas, MAX_MODE_LF_DELTAS);
+  dst_cm->seg = src_cm->seg;
+  memcpy(dst_cm->frame_contexts, src_cm->frame_contexts,
+         FRAME_CONTEXTS * sizeof(dst_cm->frame_contexts[0]));
+#else
+  (void)dst_worker;
+  (void)src_worker;
+#endif  // CONFIG_MULTITHREAD
+}
diff --git a/third_party/aom/av1/decoder/dthread.h b/third_party/aom/av1/decoder/dthread.h
new file mode 100644
index 0000000000..c17053d9c2
--- /dev/null
+++ b/third_party/aom/av1/decoder/dthread.h
@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_DECODER_DTHREAD_H_
+#define AV1_DECODER_DTHREAD_H_
+
+#include "./aom_config.h"
+#include "aom_util/aom_thread.h"
+#include "aom/internal/aom_codec_internal.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1Common;
+struct AV1Decoder;
+
+// WorkerData for the FrameWorker thread. It contains all the information of
+// the worker and decode structures for decoding a frame.
+typedef struct FrameWorkerData {
+  struct AV1Decoder *pbi;
+  const uint8_t *data;
+  const uint8_t *data_end;
+  size_t data_size;
+  void *user_priv;
+  int result;
+  int worker_id;
+  int received_frame;
+
+  // scratch_buffer is used in frame parallel mode only.
+  // It is used to make a copy of the compressed data.
+  uint8_t *scratch_buffer;
+  size_t scratch_buffer_size;
+
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t stats_mutex;
+  pthread_cond_t stats_cond;
+#endif
+
+  int frame_context_ready;  // Current frame's context is ready to read.
+  int frame_decoded;        // Finished decoding current frame.
+} FrameWorkerData;
+
+void av1_frameworker_lock_stats(AVxWorker *const worker);
+void av1_frameworker_unlock_stats(AVxWorker *const worker);
+void av1_frameworker_signal_stats(AVxWorker *const worker);
+
+// Wait until ref_buf has been decoded to row in real pixel unit.
+// Note: worker may already finish decoding ref_buf and release it in order to
+// start decoding next frame. So need to check whether worker is still decoding
+// ref_buf.
+void av1_frameworker_wait(AVxWorker *const worker, RefCntBuffer *const ref_buf,
+                          int row);
+
+// FrameWorker broadcasts its decoding progress so other workers that are
+// waiting on it can resume decoding.
+void av1_frameworker_broadcast(RefCntBuffer *const buf, int row);
+
+// Copy necessary decoding context from src worker to dst worker.
+void av1_frameworker_copy_context(AVxWorker *const dst_worker,
+                                  AVxWorker *const src_worker);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_DECODER_DTHREAD_H_
diff --git a/third_party/aom/av1/decoder/generic_decoder.c b/third_party/aom/av1/decoder/generic_decoder.c
new file mode 100644
index 0000000000..0c7d71b9fc
--- /dev/null
+++ b/third_party/aom/av1/decoder/generic_decoder.c
@@ -0,0 +1,110 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdio.h>
+
+#include "aom_dsp/bitreader.h"
+#include "av1/common/generic_code.h"
+#include "av1/common/odintrin.h"
+#include "pvq_decoder.h"
+
+/** Decodes a value from 0 to N-1 (with N up to 16) based on a cdf and adapts
+ * the cdf accordingly.
+ *
+ * @param [in,out] r     multi-symbol entropy decoder
+ * @param [in,out] cdf   CDF of the variable (Q15)
+ * @param [in]     n     number of values possible
+ * @param [in,out] count number of symbols encoded with that cdf so far
+ * @param [in]     rate  adaptation rate shift (smaller is faster)
+ * @return decoded variable
+ */
+int aom_decode_cdf_adapt_q15_(aom_reader *r, uint16_t *cdf, int n,
+ int *count, int rate ACCT_STR_PARAM) {
+  int val;
+  int i;
+  if (*count == 0) {
+    int ft;
+    ft = cdf[n - 1];
+    for (i = 0; i < n; i++) {
+      cdf[i] = AOM_ICDF(cdf[i]*32768/ft);
+    }
+  }
+  val = aom_read_cdf(r, cdf, n, ACCT_STR_NAME);
+  aom_cdf_adapt_q15(val, cdf, n, count, rate);
+  return val;
+}
+
+/** Encodes a random variable using a "generic" model, assuming that the
+ * distribution is one-sided (zero and up), has a single mode, and decays
+ * exponentially past the model.
+ *
+ * @param [in,out] r     multi-symbol entropy decoder
+ * @param [in,out] model generic probability model
+ * @param [in]     x     variable being encoded
+ * @param [in,out] ExQ16 expectation of x (adapted)
+ * @param [in]     integration integration period of ExQ16 (leaky average over
+ * 1<<integration samples)
+ *
+ * @retval decoded variable x
+ */
+int generic_decode_(aom_reader *r, generic_encoder *model,
+ int *ex_q16, int integration ACCT_STR_PARAM) {
+  int lg_q1;
+  int shift;
+  int id;
+  uint16_t *cdf;
+  int xs;
+  int lsb;
+  int x;
+  lsb = 0;
+  lg_q1 = log_ex(*ex_q16);
+  /* If expectation is too large, shift x to ensure that
+     all we have past xs=15 is the exponentially decaying tail
+     of the distribution. */
+  shift = OD_MAXI(0, (lg_q1 - 5) >> 1);
+  /* Choose the cdf to use: we have two per "octave" of ExQ16. */
+  id = OD_MINI(GENERIC_TABLES - 1, lg_q1);
+  cdf = model->cdf[id];
+  xs = aom_read_symbol_pvq(r, cdf, 16, ACCT_STR_NAME);
+  if (xs == 15) {
+    int e;
+    unsigned decay;
+    /* Estimate decay based on the assumption that the distribution is close
+       to Laplacian for large values. We should probably have an adaptive
+       estimate instead. Note: The 2* is a kludge that's not fully understood
+       yet. */
+    OD_ASSERT(*ex_q16 < INT_MAX >> 1);
+    e = ((2**ex_q16 >> 8) + (1 << shift >> 1)) >> shift;
+    decay = OD_MAXI(2, OD_MINI(254, 256*e/(e + 256)));
+    xs += aom_laplace_decode_special(r, decay, ACCT_STR_NAME);
+  }
+  if (shift != 0) {
+    int special;
+    /* Because of the rounding, there's only half the number of possibilities
+       for xs=0 */
+    special = xs == 0;
+    if (shift - special > 0) {
+      lsb = aom_read_literal(r, shift - special, ACCT_STR_NAME);
+    }
+    lsb -= !special << (shift - 1);
+  }
+  x = (xs << shift) + lsb;
+  generic_model_update(ex_q16, x, integration);
+  OD_LOG((OD_LOG_ENTROPY_CODER, OD_LOG_DEBUG,
+   "dec: %d %d %d %d %d %x", *ex_q16, x, shift, id, xs, dec->rng));
+  return x;
+}
diff --git a/third_party/aom/av1/decoder/inspection.c b/third_party/aom/av1/decoder/inspection.c
new file mode 100644
index 0000000000..2e8a610872
--- /dev/null
+++ b/third_party/aom/av1/decoder/inspection.c
@@ -0,0 +1,103 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "av1/decoder/decoder.h"
+#include "av1/decoder/inspection.h"
+#include "av1/common/enums.h"
+#if CONFIG_CDEF
+#include "av1/common/cdef.h"
+#endif
+
+void ifd_init(insp_frame_data *fd, int frame_width, int frame_height) {
+  fd->mi_cols = ALIGN_POWER_OF_TWO(frame_width, 3) >> MI_SIZE_LOG2;
+  fd->mi_rows = ALIGN_POWER_OF_TWO(frame_height, 3) >> MI_SIZE_LOG2;
+  fd->mi_grid = (insp_mi_data *)aom_malloc(sizeof(insp_mi_data) * fd->mi_rows *
+                                           fd->mi_cols);
+}
+
+void ifd_clear(insp_frame_data *fd) {
+  aom_free(fd->mi_grid);
+  fd->mi_grid = NULL;
+}
+
+/* TODO(negge) This function may be called by more than one thread when using
+               a multi-threaded decoder and this may cause a data race. */
+int ifd_inspect(insp_frame_data *fd, void *decoder) {
+  struct AV1Decoder *pbi = (struct AV1Decoder *)decoder;
+  AV1_COMMON *const cm = &pbi->common;
+  // TODO(negge): Should this function just call ifd_clear() and ifd_init()?
+  if (fd->mi_rows != cm->mi_rows || fd->mi_cols != cm->mi_cols) {
+    return 0;
+  }
+  fd->show_frame = cm->show_frame;
+  fd->frame_type = cm->frame_type;
+  fd->base_qindex = cm->base_qindex;
+  fd->tile_mi_cols = cm->tile_width;
+  fd->tile_mi_rows = cm->tile_height;
+#if CONFIG_ACCOUNTING
+  fd->accounting = &pbi->accounting;
+#endif
+#if CONFIG_CDEF
+// TODO(negge): copy per frame CDEF data
+#endif
+  int i, j;
+  for (i = 0; i < MAX_SEGMENTS; i++) {
+    for (j = 0; j < 2; j++) {
+      fd->y_dequant[i][j] = cm->y_dequant[i][j];
+      fd->uv_dequant[i][j] = cm->uv_dequant[i][j];
+    }
+  }
+  for (j = 0; j < cm->mi_rows; j++) {
+    for (i = 0; i < cm->mi_cols; i++) {
+      const MB_MODE_INFO *mbmi =
+          &cm->mi_grid_visible[j * cm->mi_stride + i]->mbmi;
+      insp_mi_data *mi = &fd->mi_grid[j * cm->mi_cols + i];
+      // Segment
+      mi->segment_id = mbmi->segment_id;
+      // Motion Vectors
+      mi->mv[0].row = mbmi->mv[0].as_mv.row;
+      mi->mv[0].col = mbmi->mv[0].as_mv.col;
+      mi->mv[1].row = mbmi->mv[1].as_mv.row;
+      mi->mv[1].col = mbmi->mv[1].as_mv.col;
+      // Reference Frames
+      mi->ref_frame[0] = mbmi->ref_frame[0];
+      mi->ref_frame[1] = mbmi->ref_frame[1];
+      // Prediction Mode
+      mi->mode = mbmi->mode;
+      // Prediction Mode for Chromatic planes
+      if (mi->mode < INTRA_MODES) {
+        mi->uv_mode = mbmi->uv_mode;
+      } else {
+        mi->uv_mode = INTRA_INVALID;
+      }
+      // Block Size
+      mi->sb_type = mbmi->sb_type;
+      // Skip Flag
+      mi->skip = mbmi->skip;
+#if CONFIG_DUAL_FILTER
+      mi->filter[0] = mbmi->interp_filter[0];
+      mi->filter[1] = mbmi->interp_filter[1];
+#else
+      mi->filter = mbmi->interp_filter;
+#endif
+      // Transform
+      mi->tx_type = mbmi->tx_type;
+      mi->tx_size = mbmi->tx_size;
+
+#if CONFIG_CDEF
+      mi->cdef_level = cm->cdef_strengths[mbmi->cdef_strength] / CLPF_STRENGTHS;
+      mi->cdef_strength =
+          cm->cdef_strengths[mbmi->cdef_strength] % CLPF_STRENGTHS;
+      mi->cdef_strength += mi->cdef_strength == 3;
+#endif
+    }
+  }
+  return 1;
+}
diff --git a/third_party/aom/av1/decoder/inspection.h b/third_party/aom/av1/decoder/inspection.h
new file mode 100644
index 0000000000..d6cf4319a6
--- /dev/null
+++ b/third_party/aom/av1/decoder/inspection.h
@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AOM_INSPECTION_H_
+#define AOM_INSPECTION_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+#if CONFIG_ACCOUNTING
+#include "av1/decoder/accounting.h"
+#endif
+
+typedef void (*aom_inspect_cb)(void *decoder, void *data);
+
+typedef struct insp_mv insp_mv;
+
+struct insp_mv {
+  int16_t row;
+  int16_t col;
+};
+
+typedef struct insp_mi_data insp_mi_data;
+
+struct insp_mi_data {
+  insp_mv mv[2];
+  int8_t ref_frame[2];
+  int8_t mode;
+  int8_t uv_mode;
+  int8_t sb_type;
+  int8_t skip;
+  int8_t segment_id;
+#if CONFIG_DUAL_FILTER
+  int8_t filter[2];
+#else
+  int8_t filter;
+#endif
+  int8_t tx_type;
+  int8_t tx_size;
+#if CONFIG_CDEF
+  int8_t cdef_level;
+  int8_t cdef_strength;
+#endif
+};
+
+typedef struct insp_frame_data insp_frame_data;
+
+struct insp_frame_data {
+#if CONFIG_ACCOUNTING
+  Accounting *accounting;
+#endif
+  insp_mi_data *mi_grid;
+  int show_frame;
+  int frame_type;
+  int base_qindex;
+  int mi_rows;
+  int mi_cols;
+  int tile_mi_rows;
+  int tile_mi_cols;
+  int16_t y_dequant[MAX_SEGMENTS][2];
+  int16_t uv_dequant[MAX_SEGMENTS][2];
+#if CONFIG_CDEF
+// TODO(negge): add per frame CDEF data
+#endif
+};
+
+void ifd_init(insp_frame_data *fd, int frame_width, int frame_height);
+void ifd_clear(insp_frame_data *fd);
+int ifd_inspect(insp_frame_data *fd, void *decoder);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // AOM_INSPECTION_H_
diff --git a/third_party/aom/av1/decoder/laplace_decoder.c b/third_party/aom/av1/decoder/laplace_decoder.c
new file mode 100644
index 0000000000..b6cf50bc7f
--- /dev/null
+++ b/third_party/aom/av1/decoder/laplace_decoder.c
@@ -0,0 +1,121 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdio.h>
+
+#include "aom_dsp/bitreader.h"
+#include "av1/common/pvq.h"
+#include "pvq_decoder.h"
+
+#define aom_decode_pvq_split(r, adapt, sum, ctx, ACCT_STR_NAME) \
+  aom_decode_pvq_split_(r, adapt, sum, ctx ACCT_STR_ARG(ACCT_STR_NAME))
+
+static int aom_decode_pvq_split_(aom_reader *r, od_pvq_codeword_ctx *adapt,
+ int sum, int ctx ACCT_STR_PARAM) {
+  int shift;
+  int count;
+  int msbs;
+  int fctx;
+  count = 0;
+  if (sum == 0) return 0;
+  shift = OD_MAXI(0, OD_ILOG(sum) - 3);
+  fctx = 7*ctx + (sum >> shift) - 1;
+  msbs = aom_read_symbol_pvq(r, adapt->pvq_split_cdf[fctx], (sum >> shift) + 1,
+      ACCT_STR_NAME);
+  if (shift) count = aom_read_literal(r, shift, ACCT_STR_NAME);
+  count += msbs << shift;
+  if (count > sum) {
+    count = sum;
+#if CONFIG_DAALA_EC
+    r->ec.error = 1;
+#else
+# error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  }
+  return count;
+}
+
+void aom_decode_band_pvq_splits(aom_reader *r, od_pvq_codeword_ctx *adapt,
+ od_coeff *y, int n, int k, int level) {
+  int mid;
+  int count_right;
+  if (n == 1) {
+    y[0] = k;
+  }
+  else if (k == 0) {
+    OD_CLEAR(y, n);
+  }
+  else if (k == 1 && n <= 16) {
+    int cdf_id;
+    int pos;
+    cdf_id = od_pvq_k1_ctx(n, level == 0);
+    OD_CLEAR(y, n);
+    pos = aom_read_symbol_pvq(r, adapt->pvq_k1_cdf[cdf_id], n, "pvq:k1");
+    y[pos] = 1;
+  }
+  else {
+    mid = n >> 1;
+    count_right = aom_decode_pvq_split(r, adapt, k, od_pvq_size_ctx(n),
+     "pvq:split");
+    aom_decode_band_pvq_splits(r, adapt, y, mid, k - count_right, level + 1);
+    aom_decode_band_pvq_splits(r, adapt, y + mid, n - mid, count_right,
+     level + 1);
+  }
+}
+
+/** Decodes the tail of a Laplace-distributed variable, i.e. it doesn't
+ * do anything special for the zero case.
+ *
+ * @param [dec] range decoder
+ * @param [decay] decay factor of the distribution, i.e. pdf ~= decay^x
+ *
+ * @retval decoded variable x
+ */
+int aom_laplace_decode_special_(aom_reader *r, unsigned decay ACCT_STR_PARAM) {
+  int pos;
+  int shift;
+  int xs;
+  int sym;
+  const uint16_t *cdf;
+  shift = 0;
+  /* We don't want a large decay value because that would require too many
+     symbols. */
+  while (decay > 235) {
+    decay = (decay*decay + 128) >> 8;
+    shift++;
+  }
+  decay = OD_MINI(decay, 254);
+  decay = OD_MAXI(decay, 2);
+  cdf = EXP_CDF_TABLE[(decay + 1) >> 1];
+  OD_LOG((OD_LOG_PVQ, OD_LOG_DEBUG, "decay = %d\n", decay));
+  xs = 0;
+  do {
+    sym = OD_MINI(xs, 15);
+    {
+      int i;
+      OD_LOG((OD_LOG_PVQ, OD_LOG_DEBUG, "%d %d %d", xs, shift, sym));
+      for (i = 0; i < 16; i++) {
+        OD_LOG_PARTIAL((OD_LOG_PVQ, OD_LOG_DEBUG, "%d ", cdf[i]));
+      }
+      OD_LOG_PARTIAL((OD_LOG_PVQ, OD_LOG_DEBUG, "\n"));
+    }
+    sym = aom_read_cdf(r, cdf, 16, ACCT_STR_NAME);
+    xs += sym;
+  } while (sym >= 15);
+  if (shift) pos = (xs << shift) + aom_read_literal(r, shift, ACCT_STR_NAME);
+  else pos = xs;
+  return pos;
+}
diff --git a/third_party/aom/av1/decoder/pvq_decoder.c b/third_party/aom/av1/decoder/pvq_decoder.c
new file mode 100644
index 0000000000..d9a8e80563
--- /dev/null
+++ b/third_party/aom/av1/decoder/pvq_decoder.c
@@ -0,0 +1,378 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdio.h>
+#include <stdlib.h>
+#include "./aom_config.h"
+#include "aom_dsp/bitreader.h"
+#include "aom_dsp/entcode.h"
+#include "aom_dsp/entdec.h"
+#include "av1/common/odintrin.h"
+#include "av1/common/partition.h"
+#include "av1/common/pvq_state.h"
+#include "av1/decoder/decint.h"
+#include "av1/decoder/pvq_decoder.h"
+#include "aom_ports/system_state.h"
+
+int aom_read_symbol_pvq_(aom_reader *r, aom_cdf_prob *cdf, int nsymbs
+ ACCT_STR_PARAM) {
+  if (cdf[0] == 0)
+    aom_cdf_init_q15_1D(cdf, nsymbs, CDF_SIZE(nsymbs));
+  return aom_read_symbol(r, cdf, nsymbs, ACCT_STR_NAME);
+}
+
+static void aom_decode_pvq_codeword(aom_reader *r, od_pvq_codeword_ctx *ctx,
+ od_coeff *y, int n, int k) {
+  int i;
+  aom_decode_band_pvq_splits(r, ctx, y, n, k, 0);
+  for (i = 0; i < n; i++) {
+    if (y[i] && aom_read_bit(r, "pvq:sign")) y[i] = -y[i];
+  }
+}
+
+/** Inverse of neg_interleave; decodes the interleaved gain.
+ *
+ * @param [in]      x      quantized/interleaved gain to decode
+ * @param [in]      ref    quantized gain of the reference
+ * @return                 original quantized gain value
+ */
+static int neg_deinterleave(int x, int ref) {
+  if (x < 2*ref-1) {
+    if (x & 1) return ref - 1 - (x >> 1);
+    else return ref + (x >> 1);
+  }
+  else return x+1;
+}
+
+/** Synthesizes one parition of coefficient values from a PVQ-encoded
+ * vector.
+ *
+ * @param [out]     xcoeff  output coefficient partition (x in math doc)
+ * @param [in]      ypulse  PVQ-encoded values (y in math doc); in the noref
+ *                          case, this vector has n entries, in the
+ *                          reference case it contains n-1 entries
+ *                          (the m-th entry is not included)
+ * @param [in]      ref     reference vector (prediction)
+ * @param [in]      n       number of elements in this partition
+ * @param [in]      gr      gain of the reference vector (prediction)
+ * @param [in]      noref   indicates presence or lack of prediction
+ * @param [in]      g       decoded quantized vector gain
+ * @param [in]      theta   decoded theta (prediction error)
+ * @param [in]      qm      QM with magnitude compensation
+ * @param [in]      qm_inv  Inverse of QM with magnitude compensation
+ */
+static void pvq_synthesis(od_coeff *xcoeff, od_coeff *ypulse, od_val16 *r16,
+ int n, od_val32 gr, int noref, od_val32 g, od_val32 theta, const int16_t *qm_inv,
+ int shift) {
+  int s;
+  int m;
+  /* Sign of the Householder reflection vector */
+  s = 0;
+  /* Direction of the Householder reflection vector */
+  m = noref ? 0 : od_compute_householder(r16, n, gr, &s, shift);
+  od_pvq_synthesis_partial(xcoeff, ypulse, r16, n, noref, g, theta, m, s,
+   qm_inv);
+}
+
+typedef struct {
+  od_coeff *ref;
+  int nb_coeffs;
+  int allow_flip;
+} cfl_ctx;
+
+/** Decodes a single vector of integers (eg, a partition within a
+ *  coefficient block) encoded using PVQ
+ *
+ * @param [in,out] ec          range encoder
+ * @param [in]     q0          scale/quantizer
+ * @param [in]     n           number of coefficients in partition
+ * @param [in,out] model       entropy decoder state
+ * @param [in,out] adapt       adaptation context
+ * @param [in,out] exg         ExQ16 expectation of decoded gain value
+ * @param [in,out] ext         ExQ16 expectation of decoded theta value
+ * @param [in]     ref         'reference' (prediction) vector
+ * @param [out]    out         decoded partition
+ * @param [out]    noref       boolean indicating absence of reference
+ * @param [in]     beta        per-band activity masking beta param
+ * @param [in]     is_keyframe whether we're encoding a keyframe
+ * @param [in]     pli         plane index
+ * @param [in]     cdf_ctx     selects which cdf context to use
+ * @param [in,out] skip_rest   whether to skip further bands in each direction
+ * @param [in]     band        index of the band being decoded
+ * @param [in]     band        index of the band being decoded
+ * @param [out]    skip        skip flag with range [0,1]
+ * @param [in]     qm          QM with magnitude compensation
+ * @param [in]     qm_inv      Inverse of QM with magnitude compensation
+ */
+static void pvq_decode_partition(aom_reader *r,
+                                 int q0,
+                                 int n,
+                                 generic_encoder model[3],
+                                 od_adapt_ctx *adapt,
+                                 int *exg,
+                                 int *ext,
+                                 od_coeff *ref,
+                                 od_coeff *out,
+                                 int *noref,
+                                 od_val16 beta,
+                                 int is_keyframe,
+                                 int pli,
+                                 int cdf_ctx,
+                                 cfl_ctx *cfl,
+                                 int has_skip,
+                                 int *skip_rest,
+                                 int band,
+                                 int *skip,
+                                 const int16_t *qm,
+                                 const int16_t *qm_inv) {
+  int k;
+  od_val32 qcg;
+  int itheta;
+  od_val32 theta;
+  od_val32 gr;
+  od_val32 gain_offset;
+  od_coeff y[MAXN];
+  int qg;
+  int id;
+  int i;
+  od_val16 ref16[MAXN];
+  int rshift;
+  theta = 0;
+  gr = 0;
+  gain_offset = 0;
+  /* Skip is per-direction. For band=0, we can use any of the flags. */
+  if (skip_rest[(band + 2) % 3]) {
+    qg = 0;
+    if (is_keyframe) {
+      itheta = -1;
+      *noref = 1;
+    }
+    else {
+      itheta = 0;
+      *noref = 0;
+    }
+  }
+  else {
+    /* Jointly decode gain, itheta and noref for small values. Then we handle
+       larger gain. */
+    id = aom_read_symbol_pvq(r, &adapt->pvq.pvq_gaintheta_cdf[cdf_ctx][0],
+     8 + 7*has_skip, "pvq:gaintheta");
+    if (!is_keyframe && id >= 10) id++;
+    if (is_keyframe && id >= 8) id++;
+    if (id >= 8) {
+      id -= 8;
+      skip_rest[0] = skip_rest[1] = skip_rest[2] = 1;
+    }
+    qg = id & 1;
+    itheta = (id >> 1) - 1;
+    *noref = (itheta == -1);
+  }
+  /* The CfL flip bit is only decoded on the first band that has noref=0. */
+  if (cfl->allow_flip && !*noref) {
+    int flip;
+    flip = aom_read_bit(r, "cfl:flip");
+    if (flip) {
+      for (i = 0; i < cfl->nb_coeffs; i++) cfl->ref[i] = -cfl->ref[i];
+    }
+    cfl->allow_flip = 0;
+  }
+  if (qg > 0) {
+    int tmp;
+    tmp = *exg;
+    qg = 1 + generic_decode(r, &model[!*noref], &tmp, 2, "pvq:gain");
+    OD_IIR_DIADIC(*exg, qg << 16, 2);
+  }
+  *skip = 0;
+#if defined(OD_FLOAT_PVQ)
+  rshift = 0;
+#else
+  /* Shift needed to make the reference fit in 15 bits, so that the Householder
+     vector can fit in 16 bits. */
+  rshift = OD_MAXI(0, od_vector_log_mag(ref, n) - 14);
+#endif
+  for (i = 0; i < n; i++) {
+#if defined(OD_FLOAT_PVQ)
+    ref16[i] = ref[i]*(double)qm[i]*OD_QM_SCALE_1;
+#else
+    ref16[i] = OD_SHR_ROUND(ref[i]*qm[i], OD_QM_SHIFT + rshift);
+#endif
+  }
+  if(!*noref){
+    /* we have a reference; compute its gain */
+    od_val32 cgr;
+    int icgr;
+    int cfl_enabled;
+    cfl_enabled = pli != 0 && is_keyframe && !OD_DISABLE_CFL;
+    cgr = od_pvq_compute_gain(ref16, n, q0, &gr, beta, rshift);
+    if (cfl_enabled) cgr = OD_CGAIN_SCALE;
+#if defined(OD_FLOAT_PVQ)
+    icgr = (int)floor(.5 + cgr);
+#else
+    icgr = OD_SHR_ROUND(cgr, OD_CGAIN_SHIFT);
+#endif
+    /* quantized gain is interleave encoded when there's a reference;
+       deinterleave it now */
+    if (is_keyframe) qg = neg_deinterleave(qg, icgr);
+    else {
+      qg = neg_deinterleave(qg, icgr + 1) - 1;
+      if (qg == 0) *skip = (icgr ? OD_PVQ_SKIP_ZERO : OD_PVQ_SKIP_COPY);
+    }
+    if (qg == icgr && itheta == 0 && !cfl_enabled) *skip = OD_PVQ_SKIP_COPY;
+    gain_offset = cgr - OD_SHL(icgr, OD_CGAIN_SHIFT);
+    qcg = OD_SHL(qg, OD_CGAIN_SHIFT) + gain_offset;
+    /* read and decode first-stage PVQ error theta */
+    if (itheta > 1) {
+      int tmp;
+      tmp = *ext;
+      itheta = 2 + generic_decode(r, &model[2], &tmp, 2, "pvq:theta");
+      OD_IIR_DIADIC(*ext, itheta << 16, 2);
+    }
+    theta = od_pvq_compute_theta(itheta, od_pvq_compute_max_theta(qcg, beta));
+  }
+  else{
+    itheta = 0;
+    if (!is_keyframe) qg++;
+    qcg = OD_SHL(qg, OD_CGAIN_SHIFT);
+    if (qg == 0) *skip = OD_PVQ_SKIP_ZERO;
+  }
+
+  k = od_pvq_compute_k(qcg, itheta, *noref, n, beta);
+  if (k != 0) {
+    /* when noref==0, y is actually size n-1 */
+    aom_decode_pvq_codeword(r, &adapt->pvq.pvq_codeword_ctx, y,
+     n - !*noref, k);
+  }
+  else {
+    OD_CLEAR(y, n);
+  }
+  if (*skip) {
+    if (*skip == OD_PVQ_SKIP_COPY) OD_COPY(out, ref, n);
+    else OD_CLEAR(out, n);
+  }
+  else {
+    od_val32 g;
+    g = od_gain_expand(qcg, q0, beta);
+    pvq_synthesis(out, y, ref16, n, gr, *noref, g, theta, qm_inv, rshift);
+  }
+  /* If OD_PVQ_SKIP_ZERO or OD_PVQ_SKIP_COPY, set skip to 1 for visualization */
+  if (*skip) *skip = 1;
+}
+
+/** Decodes a coefficient block (except for DC) encoded using PVQ
+ *
+ * @param [in,out] dec         daala decoder context
+ * @param [in]     ref         'reference' (prediction) vector
+ * @param [out]    out         decoded partition
+ * @param [in]     q0          quantizer
+ * @param [in]     pli         plane index
+ * @param [in]     bs          log of the block size minus two
+ * @param [in]     beta        per-band activity masking beta param
+ * @param [in]     is_keyframe whether we're encoding a keyframe
+ * @param [out]    flags       bitmask of the per band skip and noref flags
+ * @param [in]     ac_dc_coded skip flag for the block (range 0-3)
+ * @param [in]     qm          QM with magnitude compensation
+ * @param [in]     qm_inv      Inverse of QM with magnitude compensation
+ */
+void od_pvq_decode(daala_dec_ctx *dec,
+                   od_coeff *ref,
+                   od_coeff *out,
+                   int q0,
+                   int pli,
+                   int bs,
+                   const od_val16 *beta,
+                   int is_keyframe,
+                   unsigned int *flags,
+                   PVQ_SKIP_TYPE ac_dc_coded,
+                   const int16_t *qm,
+                   const int16_t *qm_inv){
+
+  int noref[PVQ_MAX_PARTITIONS];
+  int skip[PVQ_MAX_PARTITIONS];
+  int *exg;
+  int *ext;
+  int nb_bands;
+  int i;
+  const int *off;
+  int size[PVQ_MAX_PARTITIONS];
+  generic_encoder *model;
+  int skip_rest[3] = {0};
+  cfl_ctx cfl;
+  const unsigned char *pvq_qm;
+  int use_masking;
+
+  aom_clear_system_state();
+
+  /*Default to skip=1 and noref=0 for all bands.*/
+  for (i = 0; i < PVQ_MAX_PARTITIONS; i++) {
+    noref[i] = 0;
+    skip[i] = 1;
+  }
+
+  use_masking = dec->use_activity_masking;
+
+  if (use_masking)
+    pvq_qm = &dec->state.pvq_qm_q4[pli][0];
+  else
+    pvq_qm = 0;
+
+  exg = &dec->state.adapt->pvq.pvq_exg[pli][bs][0];
+  ext = dec->state.adapt->pvq.pvq_ext + bs*PVQ_MAX_PARTITIONS;
+  model = dec->state.adapt->pvq.pvq_param_model;
+  nb_bands = OD_BAND_OFFSETS[bs][0];
+  off = &OD_BAND_OFFSETS[bs][1];
+  out[0] = ac_dc_coded & DC_CODED;
+  if (ac_dc_coded < AC_CODED) {
+    if (is_keyframe) for (i = 1; i < 1 << (2*bs + 4); i++) out[i] = 0;
+    else for (i = 1; i < 1 << (2*bs + 4); i++) out[i] = ref[i];
+  }
+  else {
+    for (i = 0; i < nb_bands; i++) size[i] = off[i+1] - off[i];
+    cfl.ref = ref;
+    cfl.nb_coeffs = off[nb_bands];
+    cfl.allow_flip = pli != 0 && is_keyframe;
+    for (i = 0; i < nb_bands; i++) {
+      int q;
+
+      if (use_masking)
+        q = OD_MAXI(1, q0 * pvq_qm[od_qm_get_index(bs, i + 1)] >> 4);
+      else
+        q = OD_MAXI(1, q0);
+
+      pvq_decode_partition(dec->r, q, size[i],
+       model, dec->state.adapt, exg + i, ext + i, ref + off[i], out + off[i],
+       &noref[i], beta[i], is_keyframe, pli,
+       (pli != 0)*OD_TXSIZES*PVQ_MAX_PARTITIONS + bs*PVQ_MAX_PARTITIONS + i,
+       &cfl, i == 0 && (i < nb_bands - 1), skip_rest, i, &skip[i],
+       qm + off[i], qm_inv + off[i]);
+      if (i == 0 && !skip_rest[0] && bs > 0) {
+        int skip_dir;
+        int j;
+        skip_dir = aom_read_symbol(dec->r,
+         &dec->state.adapt->pvq.pvq_skip_dir_cdf[(pli != 0) + 2*(bs - 1)][0], 7,
+         "pvq:skiprest");
+        for (j = 0; j < 3; j++) skip_rest[j] = !!(skip_dir & (1 << j));
+      }
+    }
+  }
+  *flags = 0;
+  for (i = nb_bands - 1; i >= 0; i--) {
+    *flags <<= 1;
+    *flags |= noref[i]&1;
+    *flags <<= 1;
+    *flags |= skip[i]&1;
+  }
+}
diff --git a/third_party/aom/av1/decoder/pvq_decoder.h b/third_party/aom/av1/decoder/pvq_decoder.h
new file mode 100644
index 0000000000..98970663be
--- /dev/null
+++ b/third_party/aom/av1/decoder/pvq_decoder.h
@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_pvq_decoder_H)
+# define _pvq_decoder_H (1)
+# include "aom_dsp/bitreader.h"
+# include "aom_dsp/entdec.h"
+# include "av1/common/pvq.h"
+# include "av1/decoder/decint.h"
+
+#define aom_read_symbol_pvq(r, cdf, nsymbs, ACCT_STR_NAME) \
+  aom_read_symbol_pvq_(r, cdf, nsymbs ACCT_STR_ARG(ACCT_STR_NAME))
+
+int aom_read_symbol_pvq_(aom_reader *r, aom_cdf_prob *cdf, int nsymbs
+  ACCT_STR_PARAM);
+
+void aom_decode_band_pvq_splits(aom_reader *r, od_pvq_codeword_ctx *adapt,
+ od_coeff *y, int n, int k, int level);
+
+#define aom_laplace_decode_special(r, decay, ACCT_STR_NAME) \
+  aom_laplace_decode_special_(r, decay ACCT_STR_ARG(ACCT_STR_NAME))
+
+int aom_laplace_decode_special_(aom_reader *r, unsigned decay ACCT_STR_PARAM);
+
+void od_pvq_decode(daala_dec_ctx *dec, od_coeff *ref, od_coeff *out, int q0,
+    int pli, int bs, const od_val16 *beta, int is_keyframe,
+    unsigned int *flags, PVQ_SKIP_TYPE ac_dc_coded, const int16_t *qm,
+    const int16_t *qm_inv);
+
+#endif
diff --git a/third_party/aom/av1/encoder/aq_complexity.c b/third_party/aom/av1/encoder/aq_complexity.c
new file mode 100644
index 0000000000..054b0e062b
--- /dev/null
+++ b/third_party/aom/av1/encoder/aq_complexity.c
@@ -0,0 +1,163 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+#include <math.h>
+
+#include "av1/encoder/aq_complexity.h"
+#include "av1/encoder/aq_variance.h"
+#include "av1/encoder/encodeframe.h"
+#include "av1/common/seg_common.h"
+#include "av1/encoder/segmentation.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/system_state.h"
+
+#define AQ_C_SEGMENTS 5
+#define DEFAULT_AQ2_SEG 3  // Neutral Q segment
+#define AQ_C_STRENGTHS 3
+static const double aq_c_q_adj_factor[AQ_C_STRENGTHS][AQ_C_SEGMENTS] = {
+  { 1.75, 1.25, 1.05, 1.00, 0.90 },
+  { 2.00, 1.50, 1.15, 1.00, 0.85 },
+  { 2.50, 1.75, 1.25, 1.00, 0.80 }
+};
+static const double aq_c_transitions[AQ_C_STRENGTHS][AQ_C_SEGMENTS] = {
+  { 0.15, 0.30, 0.55, 2.00, 100.0 },
+  { 0.20, 0.40, 0.65, 2.00, 100.0 },
+  { 0.25, 0.50, 0.75, 2.00, 100.0 }
+};
+static const double aq_c_var_thresholds[AQ_C_STRENGTHS][AQ_C_SEGMENTS] = {
+  { -4.0, -3.0, -2.0, 100.00, 100.0 },
+  { -3.5, -2.5, -1.5, 100.00, 100.0 },
+  { -3.0, -2.0, -1.0, 100.00, 100.0 }
+};
+
+#define DEFAULT_COMPLEXITY 64
+
+static int get_aq_c_strength(int q_index, aom_bit_depth_t bit_depth) {
+  // Approximate base quatizer (truncated to int)
+  const int base_quant = av1_ac_quant(q_index, 0, bit_depth) / 4;
+  return (base_quant > 10) + (base_quant > 25);
+}
+
+void av1_setup_in_frame_q_adj(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  struct segmentation *const seg = &cm->seg;
+
+  // Make SURE use of floating point in this function is safe.
+  aom_clear_system_state();
+
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode ||
+      cpi->refresh_alt_ref_frame ||
+      (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+    int segment;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex, cm->bit_depth);
+
+    // Clear down the segment map.
+    memset(cpi->segmentation_map, DEFAULT_AQ2_SEG, cm->mi_rows * cm->mi_cols);
+
+    av1_clearall_segfeatures(seg);
+
+    // Segmentation only makes sense if the target bits per SB is above a
+    // threshold. Below this the overheads will usually outweigh any benefit.
+    if (cpi->rc.sb64_target_rate < 256) {
+      av1_disable_segmentation(seg);
+      return;
+    }
+
+    av1_enable_segmentation(seg);
+
+    // Select delta coding method.
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    // Default segment "Q" feature is disabled so it defaults to the baseline Q.
+    av1_disable_segfeature(seg, DEFAULT_AQ2_SEG, SEG_LVL_ALT_Q);
+
+    // Use some of the segments for in frame Q adjustment.
+    for (segment = 0; segment < AQ_C_SEGMENTS; ++segment) {
+      int qindex_delta;
+
+      if (segment == DEFAULT_AQ2_SEG) continue;
+
+      qindex_delta = av1_compute_qdelta_by_rate(
+          &cpi->rc, cm->frame_type, cm->base_qindex,
+          aq_c_q_adj_factor[aq_strength][segment], cm->bit_depth);
+
+      // For AQ complexity mode, we dont allow Q0 in a segment if the base
+      // Q is not 0. Q0 (lossless) implies 4x4 only and in AQ mode 2 a segment
+      // Q delta is sometimes applied without going back around the rd loop.
+      // This could lead to an illegal combination of partition size and q.
+      if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
+        qindex_delta = -cm->base_qindex + 1;
+      }
+      if ((cm->base_qindex + qindex_delta) > 0) {
+        av1_enable_segfeature(seg, segment, SEG_LVL_ALT_Q);
+        av1_set_segdata(seg, segment, SEG_LVL_ALT_Q, qindex_delta);
+      }
+    }
+  }
+}
+
+#define DEFAULT_LV_THRESH 10.0
+#define MIN_DEFAULT_LV_THRESH 8.0
+#define VAR_STRENGTH_STEP 0.25
+// Select a segment for the current block.
+// The choice of segment for a block depends on the ratio of the projected
+// bits for the block vs a target average and its spatial complexity.
+void av1_caq_select_segment(const AV1_COMP *cpi, MACROBLOCK *mb, BLOCK_SIZE bs,
+                            int mi_row, int mi_col, int projected_rate) {
+  const AV1_COMMON *const cm = &cpi->common;
+
+  const int mi_offset = mi_row * cm->mi_cols + mi_col;
+  const int xmis = AOMMIN(cm->mi_cols - mi_col, mi_size_wide[bs]);
+  const int ymis = AOMMIN(cm->mi_rows - mi_row, mi_size_high[bs]);
+  int x, y;
+  int i;
+  unsigned char segment;
+
+  if (0) {
+    segment = DEFAULT_AQ2_SEG;
+  } else {
+    // Rate depends on fraction of a SB64 in frame (xmis * ymis / bw * bh).
+    // It is converted to bits * 256 units.
+    const int64_t num = (int64_t)cpi->rc.sb64_target_rate * xmis * ymis * 256;
+    const int denom = cm->mib_size * cm->mib_size;
+    const int target_rate = (int)(num / denom);
+    double logvar;
+    double low_var_thresh;
+    const int aq_strength = get_aq_c_strength(cm->base_qindex, cm->bit_depth);
+
+    aom_clear_system_state();
+    low_var_thresh = (cpi->oxcf.pass == 2) ? AOMMAX(cpi->twopass.mb_av_energy,
+                                                    MIN_DEFAULT_LV_THRESH)
+                                           : DEFAULT_LV_THRESH;
+
+    av1_setup_src_planes(mb, cpi->source, mi_row, mi_col);
+    logvar = av1_log_block_var(cpi, mb, bs);
+
+    segment = AQ_C_SEGMENTS - 1;  // Just in case no break out below.
+    for (i = 0; i < AQ_C_SEGMENTS; ++i) {
+      // Test rate against a threshold value and variance against a threshold.
+      // Increasing segment number (higher variance and complexity) = higher Q.
+      if ((projected_rate < target_rate * aq_c_transitions[aq_strength][i]) &&
+          (logvar < (low_var_thresh + aq_c_var_thresholds[aq_strength][i]))) {
+        segment = i;
+        break;
+      }
+    }
+  }
+
+  // Fill in the entires in the segment map corresponding to this SB64.
+  for (y = 0; y < ymis; y++) {
+    for (x = 0; x < xmis; x++) {
+      cpi->segmentation_map[mi_offset + y * cm->mi_cols + x] = segment;
+    }
+  }
+}
diff --git a/third_party/aom/av1/encoder/aq_complexity.h b/third_party/aom/av1/encoder/aq_complexity.h
new file mode 100644
index 0000000000..af525b36de
--- /dev/null
+++ b/third_party/aom/av1/encoder/aq_complexity.h
@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_AQ_COMPLEXITY_H_
+#define AV1_ENCODER_AQ_COMPLEXITY_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "av1/common/enums.h"
+
+struct AV1_COMP;
+struct macroblock;
+
+// Select a segment for the current Block.
+void av1_caq_select_segment(const struct AV1_COMP *cpi, struct macroblock *,
+                            BLOCK_SIZE bs, int mi_row, int mi_col,
+                            int projected_rate);
+
+// This function sets up a set of segments with delta Q values around
+// the baseline frame quantizer.
+void av1_setup_in_frame_q_adj(struct AV1_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_AQ_COMPLEXITY_H_
diff --git a/third_party/aom/av1/encoder/aq_cyclicrefresh.c b/third_party/aom/av1/encoder/aq_cyclicrefresh.c
new file mode 100644
index 0000000000..e41c608b64
--- /dev/null
+++ b/third_party/aom/av1/encoder/aq_cyclicrefresh.c
@@ -0,0 +1,566 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+#include <math.h>
+
+#include "av1/common/seg_common.h"
+#include "av1/encoder/aq_cyclicrefresh.h"
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/segmentation.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/system_state.h"
+
+struct CYCLIC_REFRESH {
+  // Percentage of blocks per frame that are targeted as candidates
+  // for cyclic refresh.
+  int percent_refresh;
+  // Maximum q-delta as percentage of base q.
+  int max_qdelta_perc;
+  // Superblock starting index for cycling through the frame.
+  int sb_index;
+  // Controls how long block will need to wait to be refreshed again, in
+  // excess of the cycle time, i.e., in the case of all zero motion, block
+  // will be refreshed every (100/percent_refresh + time_for_refresh) frames.
+  int time_for_refresh;
+  // Target number of (8x8) blocks that are set for delta-q.
+  int target_num_seg_blocks;
+  // Actual number of (8x8) blocks that were applied delta-q.
+  int actual_num_seg1_blocks;
+  int actual_num_seg2_blocks;
+  // RD mult. parameters for segment 1.
+  int rdmult;
+  // Cyclic refresh map.
+  signed char *map;
+  // Map of the last q a block was coded at.
+  uint8_t *last_coded_q_map;
+  // Thresholds applied to the projected rate/distortion of the coding block,
+  // when deciding whether block should be refreshed.
+  int64_t thresh_rate_sb;
+  int64_t thresh_dist_sb;
+  // Threshold applied to the motion vector (in units of 1/8 pel) of the
+  // coding block, when deciding whether block should be refreshed.
+  int16_t motion_thresh;
+  // Rate target ratio to set q delta.
+  double rate_ratio_qdelta;
+  // Boost factor for rate target ratio, for segment CR_SEGMENT_ID_BOOST2.
+  int rate_boost_fac;
+  double low_content_avg;
+  int qindex_delta[3];
+};
+
+CYCLIC_REFRESH *av1_cyclic_refresh_alloc(int mi_rows, int mi_cols) {
+  size_t last_coded_q_map_size;
+  CYCLIC_REFRESH *const cr = aom_calloc(1, sizeof(*cr));
+  if (cr == NULL) return NULL;
+
+  cr->map = aom_calloc(mi_rows * mi_cols, sizeof(*cr->map));
+  if (cr->map == NULL) {
+    av1_cyclic_refresh_free(cr);
+    return NULL;
+  }
+  last_coded_q_map_size = mi_rows * mi_cols * sizeof(*cr->last_coded_q_map);
+  cr->last_coded_q_map = aom_malloc(last_coded_q_map_size);
+  if (cr->last_coded_q_map == NULL) {
+    av1_cyclic_refresh_free(cr);
+    return NULL;
+  }
+  assert(MAXQ <= 255);
+  memset(cr->last_coded_q_map, MAXQ, last_coded_q_map_size);
+
+  return cr;
+}
+
+void av1_cyclic_refresh_free(CYCLIC_REFRESH *cr) {
+  aom_free(cr->map);
+  aom_free(cr->last_coded_q_map);
+  aom_free(cr);
+}
+
+// Check if we should turn off cyclic refresh based on bitrate condition.
+static int apply_cyclic_refresh_bitrate(const AV1_COMMON *cm,
+                                        const RATE_CONTROL *rc) {
+  // Turn off cyclic refresh if bits available per frame is not sufficiently
+  // larger than bit cost of segmentation. Segment map bit cost should scale
+  // with number of seg blocks, so compare available bits to number of blocks.
+  // Average bits available per frame = avg_frame_bandwidth
+  // Number of (8x8) blocks in frame = mi_rows * mi_cols;
+  const float factor = 0.25;
+  const int number_blocks = cm->mi_rows * cm->mi_cols;
+  // The condition below corresponds to turning off at target bitrates:
+  // (at 30fps), ~12kbps for CIF, 36kbps for VGA, 100kps for HD/720p.
+  // Also turn off at very small frame sizes, to avoid too large fraction of
+  // superblocks to be refreshed per frame. Threshold below is less than QCIF.
+  if (rc->avg_frame_bandwidth < factor * number_blocks ||
+      number_blocks / 64 < 5)
+    return 0;
+  else
+    return 1;
+}
+
+// Check if this coding block, of size bsize, should be considered for refresh
+// (lower-qp coding). Decision can be based on various factors, such as
+// size of the coding block (i.e., below min_block size rejected), coding
+// mode, and rate/distortion.
+static int candidate_refresh_aq(const CYCLIC_REFRESH *cr,
+                                const MB_MODE_INFO *mbmi, int64_t rate,
+                                int64_t dist, int bsize) {
+  MV mv = mbmi->mv[0].as_mv;
+  // Reject the block for lower-qp coding if projected distortion
+  // is above the threshold, and any of the following is true:
+  // 1) mode uses large mv
+  // 2) mode is an intra-mode
+  // Otherwise accept for refresh.
+  if (dist > cr->thresh_dist_sb &&
+      (mv.row > cr->motion_thresh || mv.row < -cr->motion_thresh ||
+       mv.col > cr->motion_thresh || mv.col < -cr->motion_thresh ||
+       !is_inter_block(mbmi)))
+    return CR_SEGMENT_ID_BASE;
+  else if (bsize >= BLOCK_16X16 && rate < cr->thresh_rate_sb &&
+           is_inter_block(mbmi) && mbmi->mv[0].as_int == 0 &&
+           cr->rate_boost_fac > 10)
+    // More aggressive delta-q for bigger blocks with zero motion.
+    return CR_SEGMENT_ID_BOOST2;
+  else
+    return CR_SEGMENT_ID_BOOST1;
+}
+
+// Compute delta-q for the segment.
+static int compute_deltaq(const AV1_COMP *cpi, int q, double rate_factor) {
+  const CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  int deltaq = av1_compute_qdelta_by_rate(rc, cpi->common.frame_type, q,
+                                          rate_factor, cpi->common.bit_depth);
+  if ((-deltaq) > cr->max_qdelta_perc * q / 100) {
+    deltaq = -cr->max_qdelta_perc * q / 100;
+  }
+  return deltaq;
+}
+
+// For the just encoded frame, estimate the bits, incorporating the delta-q
+// from non-base segment. For now ignore effect of multiple segments
+// (with different delta-q). Note this function is called in the postencode
+// (called from rc_update_rate_correction_factors()).
+int av1_cyclic_refresh_estimate_bits_at_q(const AV1_COMP *cpi,
+                                          double correction_factor) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int estimated_bits;
+  int mbs = cm->MBs;
+  int num8x8bl = mbs << 2;
+  // Weight for non-base segments: use actual number of blocks refreshed in
+  // previous/just encoded frame. Note number of blocks here is in 8x8 units.
+  double weight_segment1 = (double)cr->actual_num_seg1_blocks / num8x8bl;
+  double weight_segment2 = (double)cr->actual_num_seg2_blocks / num8x8bl;
+  // Take segment weighted average for estimated bits.
+  estimated_bits =
+      (int)((1.0 - weight_segment1 - weight_segment2) *
+                av1_estimate_bits_at_q(cm->frame_type, cm->base_qindex, mbs,
+                                       correction_factor, cm->bit_depth) +
+            weight_segment1 *
+                av1_estimate_bits_at_q(cm->frame_type,
+                                       cm->base_qindex + cr->qindex_delta[1],
+                                       mbs, correction_factor, cm->bit_depth) +
+            weight_segment2 *
+                av1_estimate_bits_at_q(cm->frame_type,
+                                       cm->base_qindex + cr->qindex_delta[2],
+                                       mbs, correction_factor, cm->bit_depth));
+  return estimated_bits;
+}
+
+// Prior to encoding the frame, estimate the bits per mb, for a given q = i and
+// a corresponding delta-q (for segment 1). This function is called in the
+// rc_regulate_q() to set the base qp index.
+// Note: the segment map is set to either 0/CR_SEGMENT_ID_BASE (no refresh) or
+// to 1/CR_SEGMENT_ID_BOOST1 (refresh) for each superblock, prior to encoding.
+int av1_cyclic_refresh_rc_bits_per_mb(const AV1_COMP *cpi, int i,
+                                      double correction_factor) {
+  const AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int bits_per_mb;
+  int num8x8bl = cm->MBs << 2;
+  // Weight for segment prior to encoding: take the average of the target
+  // number for the frame to be encoded and the actual from the previous frame.
+  double weight_segment =
+      (double)((cr->target_num_seg_blocks + cr->actual_num_seg1_blocks +
+                cr->actual_num_seg2_blocks) >>
+               1) /
+      num8x8bl;
+  // Compute delta-q corresponding to qindex i.
+  int deltaq = compute_deltaq(cpi, i, cr->rate_ratio_qdelta);
+  // Take segment weighted average for bits per mb.
+  bits_per_mb = (int)((1.0 - weight_segment) *
+                          av1_rc_bits_per_mb(cm->frame_type, i,
+                                             correction_factor, cm->bit_depth) +
+                      weight_segment *
+                          av1_rc_bits_per_mb(cm->frame_type, i + deltaq,
+                                             correction_factor, cm->bit_depth));
+  return bits_per_mb;
+}
+
+// Prior to coding a given prediction block, of size bsize at (mi_row, mi_col),
+// check if we should reset the segment_id, and update the cyclic_refresh map
+// and segmentation map.
+void av1_cyclic_refresh_update_segment(const AV1_COMP *cpi,
+                                       MB_MODE_INFO *const mbmi, int mi_row,
+                                       int mi_col, BLOCK_SIZE bsize,
+                                       int64_t rate, int64_t dist, int skip) {
+  const AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  const int bw = mi_size_wide[bsize];
+  const int bh = mi_size_high[bsize];
+  const int xmis = AOMMIN(cm->mi_cols - mi_col, bw);
+  const int ymis = AOMMIN(cm->mi_rows - mi_row, bh);
+  const int block_index = mi_row * cm->mi_cols + mi_col;
+  const int refresh_this_block =
+      candidate_refresh_aq(cr, mbmi, rate, dist, bsize);
+  // Default is to not update the refresh map.
+  int new_map_value = cr->map[block_index];
+  int x = 0;
+  int y = 0;
+
+  // If this block is labeled for refresh, check if we should reset the
+  // segment_id.
+  if (cyclic_refresh_segment_id_boosted(mbmi->segment_id)) {
+    mbmi->segment_id = refresh_this_block;
+    // Reset segment_id if will be skipped.
+    if (skip) mbmi->segment_id = CR_SEGMENT_ID_BASE;
+  }
+
+  // Update the cyclic refresh map, to be used for setting segmentation map
+  // for the next frame. If the block  will be refreshed this frame, mark it
+  // as clean. The magnitude of the -ve influences how long before we consider
+  // it for refresh again.
+  if (cyclic_refresh_segment_id_boosted(mbmi->segment_id)) {
+    new_map_value = -cr->time_for_refresh;
+  } else if (refresh_this_block) {
+    // Else if it is accepted as candidate for refresh, and has not already
+    // been refreshed (marked as 1) then mark it as a candidate for cleanup
+    // for future time (marked as 0), otherwise don't update it.
+    if (cr->map[block_index] == 1) new_map_value = 0;
+  } else {
+    // Leave it marked as block that is not candidate for refresh.
+    new_map_value = 1;
+  }
+
+  // Update entries in the cyclic refresh map with new_map_value, and
+  // copy mbmi->segment_id into global segmentation map.
+  for (y = 0; y < ymis; y++)
+    for (x = 0; x < xmis; x++) {
+      int map_offset = block_index + y * cm->mi_cols + x;
+      cr->map[map_offset] = new_map_value;
+      cpi->segmentation_map[map_offset] = mbmi->segment_id;
+      // Inter skip blocks were clearly not coded at the current qindex, so
+      // don't update the map for them. For cases where motion is non-zero or
+      // the reference frame isn't the previous frame, the previous value in
+      // the map for this spatial location is not entirely correct.
+      if ((!is_inter_block(mbmi) || !skip) &&
+          mbmi->segment_id <= CR_SEGMENT_ID_BOOST2) {
+        cr->last_coded_q_map[map_offset] = clamp(
+            cm->base_qindex + cr->qindex_delta[mbmi->segment_id], 0, MAXQ);
+      } else if (is_inter_block(mbmi) && skip &&
+                 mbmi->segment_id <= CR_SEGMENT_ID_BOOST2) {
+        cr->last_coded_q_map[map_offset] =
+            AOMMIN(clamp(cm->base_qindex + cr->qindex_delta[mbmi->segment_id],
+                         0, MAXQ),
+                   cr->last_coded_q_map[map_offset]);
+      }
+    }
+}
+
+// Update the actual number of blocks that were applied the segment delta q.
+void av1_cyclic_refresh_postencode(AV1_COMP *const cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  unsigned char *const seg_map = cpi->segmentation_map;
+  int mi_row, mi_col;
+  cr->actual_num_seg1_blocks = 0;
+  cr->actual_num_seg2_blocks = 0;
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++)
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
+      if (cyclic_refresh_segment_id(seg_map[mi_row * cm->mi_cols + mi_col]) ==
+          CR_SEGMENT_ID_BOOST1)
+        cr->actual_num_seg1_blocks++;
+      else if (cyclic_refresh_segment_id(
+                   seg_map[mi_row * cm->mi_cols + mi_col]) ==
+               CR_SEGMENT_ID_BOOST2)
+        cr->actual_num_seg2_blocks++;
+    }
+}
+
+// Set golden frame update interval, for 1 pass CBR mode.
+void av1_cyclic_refresh_set_golden_update(AV1_COMP *const cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  // Set minimum gf_interval for GF update to a multiple (== 2) of refresh
+  // period. Depending on past encoding stats, GF flag may be reset and update
+  // may not occur until next baseline_gf_interval.
+  if (cr->percent_refresh > 0)
+    rc->baseline_gf_interval = 4 * (100 / cr->percent_refresh);
+  else
+    rc->baseline_gf_interval = 40;
+}
+
+// Update some encoding stats (from the just encoded frame). If this frame's
+// background has high motion, refresh the golden frame. Otherwise, if the
+// golden reference is to be updated check if we should NOT update the golden
+// ref.
+void av1_cyclic_refresh_check_golden_update(AV1_COMP *const cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  int mi_row, mi_col;
+  double fraction_low = 0.0;
+  int low_content_frame = 0;
+
+  MODE_INFO **mi;
+  RATE_CONTROL *const rc = &cpi->rc;
+  const int rows = cm->mi_rows, cols = cm->mi_cols;
+  int cnt1 = 0, cnt2 = 0;
+  int force_gf_refresh = 0;
+
+  for (mi_row = 0; mi_row < rows; mi_row++) {
+    mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
+
+    for (mi_col = 0; mi_col < cols; mi_col++) {
+      int16_t abs_mvr = mi[0]->mbmi.mv[0].as_mv.row >= 0
+                            ? mi[0]->mbmi.mv[0].as_mv.row
+                            : -1 * mi[0]->mbmi.mv[0].as_mv.row;
+      int16_t abs_mvc = mi[0]->mbmi.mv[0].as_mv.col >= 0
+                            ? mi[0]->mbmi.mv[0].as_mv.col
+                            : -1 * mi[0]->mbmi.mv[0].as_mv.col;
+
+      // Calculate the motion of the background.
+      if (abs_mvr <= 16 && abs_mvc <= 16) {
+        cnt1++;
+        if (abs_mvr == 0 && abs_mvc == 0) cnt2++;
+      }
+      mi++;
+
+      // Accumulate low_content_frame.
+      if (cr->map[mi_row * cols + mi_col] < 1) low_content_frame++;
+    }
+  }
+
+  // For video conference clips, if the background has high motion in current
+  // frame because of the camera movement, set this frame as the golden frame.
+  // Use 70% and 5% as the thresholds for golden frame refreshing.
+  // Also, force this frame as a golden update frame if this frame will change
+  // the resolution (resize_pending != 0).
+  if (cpi->resize_pending != 0 ||
+      (cnt1 * 10 > (70 * rows * cols) && cnt2 * 20 < cnt1)) {
+    av1_cyclic_refresh_set_golden_update(cpi);
+    rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+
+    if (rc->frames_till_gf_update_due > rc->frames_to_key)
+      rc->frames_till_gf_update_due = rc->frames_to_key;
+    cpi->refresh_golden_frame = 1;
+    force_gf_refresh = 1;
+  }
+
+  fraction_low = (double)low_content_frame / (rows * cols);
+  // Update average.
+  cr->low_content_avg = (fraction_low + 3 * cr->low_content_avg) / 4;
+  if (!force_gf_refresh && cpi->refresh_golden_frame == 1) {
+    // Don't update golden reference if the amount of low_content for the
+    // current encoded frame is small, or if the recursive average of the
+    // low_content over the update interval window falls below threshold.
+    if (fraction_low < 0.8 || cr->low_content_avg < 0.7)
+      cpi->refresh_golden_frame = 0;
+    // Reset for next internal.
+    cr->low_content_avg = fraction_low;
+  }
+}
+
+// Update the segmentation map, and related quantities: cyclic refresh map,
+// refresh sb_index, and target number of blocks to be refreshed.
+// The map is set to either 0/CR_SEGMENT_ID_BASE (no refresh) or to
+// 1/CR_SEGMENT_ID_BOOST1 (refresh) for each superblock.
+// Blocks labeled as BOOST1 may later get set to BOOST2 (during the
+// encoding of the superblock).
+static void cyclic_refresh_update_map(AV1_COMP *const cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  unsigned char *const seg_map = cpi->segmentation_map;
+  int i, block_count, bl_index, sb_rows, sb_cols, sbs_in_frame;
+  int xmis, ymis, x, y;
+  memset(seg_map, CR_SEGMENT_ID_BASE, cm->mi_rows * cm->mi_cols);
+  sb_cols = (cm->mi_cols + cm->mib_size - 1) / cm->mib_size;
+  sb_rows = (cm->mi_rows + cm->mib_size - 1) / cm->mib_size;
+  sbs_in_frame = sb_cols * sb_rows;
+  // Number of target blocks to get the q delta (segment 1).
+  block_count = cr->percent_refresh * cm->mi_rows * cm->mi_cols / 100;
+  // Set the segmentation map: cycle through the superblocks, starting at
+  // cr->mb_index, and stopping when either block_count blocks have been found
+  // to be refreshed, or we have passed through whole frame.
+  assert(cr->sb_index < sbs_in_frame);
+  i = cr->sb_index;
+  cr->target_num_seg_blocks = 0;
+  do {
+    int sum_map = 0;
+    // Get the mi_row/mi_col corresponding to superblock index i.
+    int sb_row_index = (i / sb_cols);
+    int sb_col_index = i - sb_row_index * sb_cols;
+    int mi_row = sb_row_index * cm->mib_size;
+    int mi_col = sb_col_index * cm->mib_size;
+    int qindex_thresh =
+        cpi->oxcf.content == AOM_CONTENT_SCREEN
+            ? av1_get_qindex(&cm->seg, CR_SEGMENT_ID_BOOST2, cm->base_qindex)
+            : 0;
+    assert(mi_row >= 0 && mi_row < cm->mi_rows);
+    assert(mi_col >= 0 && mi_col < cm->mi_cols);
+    bl_index = mi_row * cm->mi_cols + mi_col;
+    // Loop through all MI blocks in superblock and update map.
+    xmis = AOMMIN(cm->mi_cols - mi_col, cm->mib_size);
+    ymis = AOMMIN(cm->mi_rows - mi_row, cm->mib_size);
+    for (y = 0; y < ymis; y++) {
+      for (x = 0; x < xmis; x++) {
+        const int bl_index2 = bl_index + y * cm->mi_cols + x;
+        // If the block is as a candidate for clean up then mark it
+        // for possible boost/refresh (segment 1). The segment id may get
+        // reset to 0 later if block gets coded anything other than ZEROMV.
+        if (cr->map[bl_index2] == 0) {
+          if (cr->last_coded_q_map[bl_index2] > qindex_thresh) sum_map++;
+        } else if (cr->map[bl_index2] < 0) {
+          cr->map[bl_index2]++;
+        }
+      }
+    }
+    // Enforce constant segment over superblock.
+    // If segment is at least half of superblock, set to 1.
+    if (sum_map >= xmis * ymis / 2) {
+      for (y = 0; y < ymis; y++)
+        for (x = 0; x < xmis; x++) {
+          seg_map[bl_index + y * cm->mi_cols + x] = CR_SEGMENT_ID_BOOST1;
+        }
+      cr->target_num_seg_blocks += xmis * ymis;
+    }
+    i++;
+    if (i == sbs_in_frame) {
+      i = 0;
+    }
+  } while (cr->target_num_seg_blocks < block_count && i != cr->sb_index);
+  cr->sb_index = i;
+}
+
+// Set cyclic refresh parameters.
+void av1_cyclic_refresh_update_parameters(AV1_COMP *const cpi) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  cr->percent_refresh = 10;
+  cr->max_qdelta_perc = 50;
+  cr->time_for_refresh = 0;
+  // Use larger delta-qp (increase rate_ratio_qdelta) for first few (~4)
+  // periods of the refresh cycle, after a key frame.
+  if (rc->frames_since_key < 4 * cr->percent_refresh)
+    cr->rate_ratio_qdelta = 3.0;
+  else
+    cr->rate_ratio_qdelta = 2.0;
+  // Adjust some parameters for low resolutions at low bitrates.
+  if (cm->width <= 352 && cm->height <= 288 && rc->avg_frame_bandwidth < 3400) {
+    cr->motion_thresh = 4;
+    cr->rate_boost_fac = 10;
+  } else {
+    cr->motion_thresh = 32;
+    cr->rate_boost_fac = 17;
+  }
+}
+
+// Setup cyclic background refresh: set delta q and segmentation map.
+void av1_cyclic_refresh_setup(AV1_COMP *const cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  struct segmentation *const seg = &cm->seg;
+  const int apply_cyclic_refresh = apply_cyclic_refresh_bitrate(cm, rc);
+  if (cm->current_video_frame == 0) cr->low_content_avg = 0.0;
+  // Don't apply refresh on key frame or enhancement layer frames.
+  if (!apply_cyclic_refresh || cm->frame_type == KEY_FRAME) {
+    // Set segmentation map to 0 and disable.
+    unsigned char *const seg_map = cpi->segmentation_map;
+    memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
+    av1_disable_segmentation(&cm->seg);
+    if (cm->frame_type == KEY_FRAME) {
+      memset(cr->last_coded_q_map, MAXQ,
+             cm->mi_rows * cm->mi_cols * sizeof(*cr->last_coded_q_map));
+      cr->sb_index = 0;
+    }
+    return;
+  } else {
+    int qindex_delta = 0;
+    int qindex2;
+    const double q = av1_convert_qindex_to_q(cm->base_qindex, cm->bit_depth);
+    aom_clear_system_state();
+    // Set rate threshold to some multiple (set to 2 for now) of the target
+    // rate (target is given by sb64_target_rate and scaled by 256).
+    cr->thresh_rate_sb = ((int64_t)(rc->sb64_target_rate) << 8) << 2;
+    // Distortion threshold, quadratic in Q, scale factor to be adjusted.
+    // q will not exceed 457, so (q * q) is within 32bit; see:
+    // av1_convert_qindex_to_q(), av1_ac_quant(), ac_qlookup*[].
+    cr->thresh_dist_sb = ((int64_t)(q * q)) << 2;
+
+    // Set up segmentation.
+    // Clear down the segment map.
+    av1_enable_segmentation(&cm->seg);
+    av1_clearall_segfeatures(seg);
+    // Select delta coding method.
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    // Note: setting temporal_update has no effect, as the seg-map coding method
+    // (temporal or spatial) is determined in
+    // av1_choose_segmap_coding_method(),
+    // based on the coding cost of each method. For error_resilient mode on the
+    // last_frame_seg_map is set to 0, so if temporal coding is used, it is
+    // relative to 0 previous map.
+    // seg->temporal_update = 0;
+
+    // Segment BASE "Q" feature is disabled so it defaults to the baseline Q.
+    av1_disable_segfeature(seg, CR_SEGMENT_ID_BASE, SEG_LVL_ALT_Q);
+    // Use segment BOOST1 for in-frame Q adjustment.
+    av1_enable_segfeature(seg, CR_SEGMENT_ID_BOOST1, SEG_LVL_ALT_Q);
+    // Use segment BOOST2 for more aggressive in-frame Q adjustment.
+    av1_enable_segfeature(seg, CR_SEGMENT_ID_BOOST2, SEG_LVL_ALT_Q);
+
+    // Set the q delta for segment BOOST1.
+    qindex_delta = compute_deltaq(cpi, cm->base_qindex, cr->rate_ratio_qdelta);
+    cr->qindex_delta[1] = qindex_delta;
+
+    // Compute rd-mult for segment BOOST1.
+    qindex2 = clamp(cm->base_qindex + cm->y_dc_delta_q + qindex_delta, 0, MAXQ);
+
+    cr->rdmult = av1_compute_rd_mult(cpi, qindex2);
+
+    av1_set_segdata(seg, CR_SEGMENT_ID_BOOST1, SEG_LVL_ALT_Q, qindex_delta);
+
+    // Set a more aggressive (higher) q delta for segment BOOST2.
+    qindex_delta = compute_deltaq(
+        cpi, cm->base_qindex,
+        AOMMIN(CR_MAX_RATE_TARGET_RATIO,
+               0.1 * cr->rate_boost_fac * cr->rate_ratio_qdelta));
+    cr->qindex_delta[2] = qindex_delta;
+    av1_set_segdata(seg, CR_SEGMENT_ID_BOOST2, SEG_LVL_ALT_Q, qindex_delta);
+
+    // Update the segmentation and refresh map.
+    cyclic_refresh_update_map(cpi);
+  }
+}
+
+int av1_cyclic_refresh_get_rdmult(const CYCLIC_REFRESH *cr) {
+  return cr->rdmult;
+}
+
+void av1_cyclic_refresh_reset_resize(AV1_COMP *const cpi) {
+  const AV1_COMMON *const cm = &cpi->common;
+  CYCLIC_REFRESH *const cr = cpi->cyclic_refresh;
+  memset(cr->map, 0, cm->mi_rows * cm->mi_cols);
+  cr->sb_index = 0;
+  cpi->refresh_golden_frame = 1;
+}
diff --git a/third_party/aom/av1/encoder/aq_cyclicrefresh.h b/third_party/aom/av1/encoder/aq_cyclicrefresh.h
new file mode 100644
index 0000000000..459ab80b8e
--- /dev/null
+++ b/third_party/aom/av1/encoder/aq_cyclicrefresh.h
@@ -0,0 +1,98 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_AQ_CYCLICREFRESH_H_
+#define AV1_ENCODER_AQ_CYCLICREFRESH_H_
+
+#include "av1/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The segment ids used in cyclic refresh: from base (no boost) to increasing
+// boost (higher delta-qp).
+#define CR_SEGMENT_ID_BASE 0
+#define CR_SEGMENT_ID_BOOST1 1
+#define CR_SEGMENT_ID_BOOST2 2
+
+// Maximum rate target ratio for setting segment delta-qp.
+#define CR_MAX_RATE_TARGET_RATIO 4.0
+
+struct AV1_COMP;
+
+struct CYCLIC_REFRESH;
+typedef struct CYCLIC_REFRESH CYCLIC_REFRESH;
+
+CYCLIC_REFRESH *av1_cyclic_refresh_alloc(int mi_rows, int mi_cols);
+
+void av1_cyclic_refresh_free(CYCLIC_REFRESH *cr);
+
+// Estimate the bits, incorporating the delta-q from segment 1, after encoding
+// the frame.
+int av1_cyclic_refresh_estimate_bits_at_q(const struct AV1_COMP *cpi,
+                                          double correction_factor);
+
+// Estimate the bits per mb, for a given q = i and a corresponding delta-q
+// (for segment 1), prior to encoding the frame.
+int av1_cyclic_refresh_rc_bits_per_mb(const struct AV1_COMP *cpi, int i,
+                                      double correction_factor);
+
+// Prior to coding a given prediction block, of size bsize at (mi_row, mi_col),
+// check if we should reset the segment_id, and update the cyclic_refresh map
+// and segmentation map.
+void av1_cyclic_refresh_update_segment(const struct AV1_COMP *cpi,
+                                       MB_MODE_INFO *const mbmi, int mi_row,
+                                       int mi_col, BLOCK_SIZE bsize,
+                                       int64_t rate, int64_t dist, int skip);
+
+// Update the segmentation map, and related quantities: cyclic refresh map,
+// refresh sb_index, and target number of blocks to be refreshed.
+void av1_cyclic_refresh_update__map(struct AV1_COMP *const cpi);
+
+// Update the actual number of blocks that were applied the segment delta q.
+void av1_cyclic_refresh_postencode(struct AV1_COMP *const cpi);
+
+// Set golden frame update interval, for 1 pass CBR mode.
+void av1_cyclic_refresh_set_golden_update(struct AV1_COMP *const cpi);
+
+// Check if we should not update golden reference, based on past refresh stats.
+void av1_cyclic_refresh_check_golden_update(struct AV1_COMP *const cpi);
+
+// Set/update global/frame level refresh parameters.
+void av1_cyclic_refresh_update_parameters(struct AV1_COMP *const cpi);
+
+// Setup cyclic background refresh: set delta q and segmentation map.
+void av1_cyclic_refresh_setup(struct AV1_COMP *const cpi);
+
+int av1_cyclic_refresh_get_rdmult(const CYCLIC_REFRESH *cr);
+
+void av1_cyclic_refresh_reset_resize(struct AV1_COMP *const cpi);
+
+static INLINE int cyclic_refresh_segment_id_boosted(int segment_id) {
+  return segment_id == CR_SEGMENT_ID_BOOST1 ||
+         segment_id == CR_SEGMENT_ID_BOOST2;
+}
+
+static INLINE int cyclic_refresh_segment_id(int segment_id) {
+  if (segment_id == CR_SEGMENT_ID_BOOST1)
+    return CR_SEGMENT_ID_BOOST1;
+  else if (segment_id == CR_SEGMENT_ID_BOOST2)
+    return CR_SEGMENT_ID_BOOST2;
+  else
+    return CR_SEGMENT_ID_BASE;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_AQ_CYCLICREFRESH_H_
diff --git a/third_party/aom/av1/encoder/aq_variance.c b/third_party/aom/av1/encoder/aq_variance.c
new file mode 100644
index 0000000000..ab9b3790bf
--- /dev/null
+++ b/third_party/aom/av1/encoder/aq_variance.c
@@ -0,0 +1,207 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+
+#include "aom_ports/mem.h"
+
+#include "av1/encoder/aq_variance.h"
+
+#include "av1/common/seg_common.h"
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/rd.h"
+#include "av1/encoder/segmentation.h"
+#include "aom_ports/system_state.h"
+
+#define ENERGY_MIN (-4)
+#define ENERGY_MAX (1)
+#define ENERGY_SPAN (ENERGY_MAX - ENERGY_MIN + 1)
+#define ENERGY_IN_BOUNDS(energy) \
+  assert((energy) >= ENERGY_MIN && (energy) <= ENERGY_MAX)
+
+static const double rate_ratio[MAX_SEGMENTS] = { 2.5,  2.0, 1.5, 1.0,
+                                                 0.75, 1.0, 1.0, 1.0 };
+static const int segment_id[ENERGY_SPAN] = { 0, 1, 1, 2, 3, 4 };
+
+#define SEGMENT_ID(i) segment_id[(i)-ENERGY_MIN]
+
+DECLARE_ALIGNED(16, static const uint8_t, av1_all_zeros[MAX_SB_SIZE]) = { 0 };
+#if CONFIG_HIGHBITDEPTH
+DECLARE_ALIGNED(16, static const uint16_t,
+                av1_highbd_all_zeros[MAX_SB_SIZE]) = { 0 };
+#endif
+
+unsigned int av1_vaq_segment_id(int energy) {
+  ENERGY_IN_BOUNDS(energy);
+  return SEGMENT_ID(energy);
+}
+
+void av1_vaq_frame_setup(AV1_COMP *cpi) {
+  AV1_COMMON *cm = &cpi->common;
+  struct segmentation *seg = &cm->seg;
+  int i;
+
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode ||
+      cpi->refresh_alt_ref_frame ||
+      (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
+    cpi->vaq_refresh = 1;
+
+    av1_enable_segmentation(seg);
+    av1_clearall_segfeatures(seg);
+
+    seg->abs_delta = SEGMENT_DELTADATA;
+
+    aom_clear_system_state();
+
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      int qindex_delta =
+          av1_compute_qdelta_by_rate(&cpi->rc, cm->frame_type, cm->base_qindex,
+                                     rate_ratio[i], cm->bit_depth);
+
+      // We don't allow qindex 0 in a segment if the base value is not 0.
+      // Q index 0 (lossless) implies 4x4 encoding only and in AQ mode a segment
+      // Q delta is sometimes applied without going back around the rd loop.
+      // This could lead to an illegal combination of partition size and q.
+      if ((cm->base_qindex != 0) && ((cm->base_qindex + qindex_delta) == 0)) {
+        qindex_delta = -cm->base_qindex + 1;
+      }
+
+      // No need to enable SEG_LVL_ALT_Q for this segment.
+      if (rate_ratio[i] == 1.0) {
+        continue;
+      }
+
+      av1_set_segdata(seg, i, SEG_LVL_ALT_Q, qindex_delta);
+      av1_enable_segfeature(seg, i, SEG_LVL_ALT_Q);
+    }
+  }
+}
+
+/* TODO(agrange, paulwilkins): The block_variance calls the unoptimized versions
+ * of variance() and highbd_8_variance(). It should not.
+ */
+static void aq_variance(const uint8_t *a, int a_stride, const uint8_t *b,
+                        int b_stride, int w, int h, unsigned int *sse,
+                        int *sum) {
+  int i, j;
+
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void aq_highbd_variance64(const uint8_t *a8, int a_stride,
+                                 const uint8_t *b8, int b_stride, int w, int h,
+                                 uint64_t *sse, uint64_t *sum) {
+  int i, j;
+
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  *sum = 0;
+  *sse = 0;
+
+  for (i = 0; i < h; i++) {
+    for (j = 0; j < w; j++) {
+      const int diff = a[j] - b[j];
+      *sum += diff;
+      *sse += diff * diff;
+    }
+    a += a_stride;
+    b += b_stride;
+  }
+}
+
+static void aq_highbd_8_variance(const uint8_t *a8, int a_stride,
+                                 const uint8_t *b8, int b_stride, int w, int h,
+                                 unsigned int *sse, int *sum) {
+  uint64_t sse_long = 0;
+  uint64_t sum_long = 0;
+  aq_highbd_variance64(a8, a_stride, b8, b_stride, w, h, &sse_long, &sum_long);
+  *sse = (unsigned int)sse_long;
+  *sum = (int)sum_long;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static unsigned int block_variance(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                   BLOCK_SIZE bs) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  unsigned int var, sse;
+  int right_overflow =
+      (xd->mb_to_right_edge < 0) ? ((-xd->mb_to_right_edge) >> 3) : 0;
+  int bottom_overflow =
+      (xd->mb_to_bottom_edge < 0) ? ((-xd->mb_to_bottom_edge) >> 3) : 0;
+
+  if (right_overflow || bottom_overflow) {
+    const int bw = 8 * mi_size_wide[bs] - right_overflow;
+    const int bh = 8 * mi_size_high[bs] - bottom_overflow;
+    int avg;
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      aq_highbd_8_variance(x->plane[0].src.buf, x->plane[0].src.stride,
+                           CONVERT_TO_BYTEPTR(av1_highbd_all_zeros), 0, bw, bh,
+                           &sse, &avg);
+      sse >>= 2 * (xd->bd - 8);
+      avg >>= (xd->bd - 8);
+    } else {
+      aq_variance(x->plane[0].src.buf, x->plane[0].src.stride, av1_all_zeros, 0,
+                  bw, bh, &sse, &avg);
+    }
+#else
+    aq_variance(x->plane[0].src.buf, x->plane[0].src.stride, av1_all_zeros, 0,
+                bw, bh, &sse, &avg);
+#endif  // CONFIG_HIGHBITDEPTH
+    var = sse - (unsigned int)(((int64_t)avg * avg) / (bw * bh));
+    return (unsigned int)((uint64_t)var * 256) / (bw * bh);
+  } else {
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      var =
+          cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
+                             CONVERT_TO_BYTEPTR(av1_highbd_all_zeros), 0, &sse);
+    } else {
+      var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
+                               av1_all_zeros, 0, &sse);
+    }
+#else
+    var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
+                             av1_all_zeros, 0, &sse);
+#endif  // CONFIG_HIGHBITDEPTH
+    return (unsigned int)((uint64_t)var * 256) >> num_pels_log2_lookup[bs];
+  }
+}
+
+double av1_log_block_var(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs) {
+  unsigned int var = block_variance(cpi, x, bs);
+  aom_clear_system_state();
+  return log(var + 1.0);
+}
+
+#define DEFAULT_E_MIDPOINT 10.0
+int av1_block_energy(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs) {
+  double energy;
+  double energy_midpoint;
+  aom_clear_system_state();
+  energy_midpoint =
+      (cpi->oxcf.pass == 2) ? cpi->twopass.mb_av_energy : DEFAULT_E_MIDPOINT;
+  energy = av1_log_block_var(cpi, x, bs) - energy_midpoint;
+  return clamp((int)round(energy), ENERGY_MIN, ENERGY_MAX);
+}
diff --git a/third_party/aom/av1/encoder/aq_variance.h b/third_party/aom/av1/encoder/aq_variance.h
new file mode 100644
index 0000000000..05725c5def
--- /dev/null
+++ b/third_party/aom/av1/encoder/aq_variance.h
@@ -0,0 +1,31 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_AQ_VARIANCE_H_
+#define AV1_ENCODER_AQ_VARIANCE_H_
+
+#include "av1/encoder/encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+unsigned int av1_vaq_segment_id(int energy);
+void av1_vaq_frame_setup(AV1_COMP *cpi);
+
+int av1_block_energy(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs);
+double av1_log_block_var(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_AQ_VARIANCE_H_
diff --git a/third_party/aom/av1/encoder/arm/neon/dct_neon.c b/third_party/aom/av1/encoder/arm/neon/dct_neon.c
new file mode 100644
index 0000000000..f6ce24a3dd
--- /dev/null
+++ b/third_party/aom/av1/encoder/arm/neon/dct_neon.c
@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "av1/common/blockd.h"
+#include "aom_dsp/txfm_common.h"
+
+void av1_fdct8x8_quant_neon(const int16_t *input, int stride,
+                            int16_t *coeff_ptr, intptr_t n_coeffs,
+                            int skip_block, const int16_t *zbin_ptr,
+                            const int16_t *round_ptr, const int16_t *quant_ptr,
+                            const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                            int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                            uint16_t *eob_ptr, const int16_t *scan_ptr,
+                            const int16_t *iscan_ptr) {
+  int16_t temp_buffer[64];
+  (void)coeff_ptr;
+
+  aom_fdct8x8_neon(input, temp_buffer, stride);
+  av1_quantize_fp_neon(temp_buffer, n_coeffs, skip_block, zbin_ptr, round_ptr,
+                       quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
+                       dequant_ptr, eob_ptr, scan_ptr, iscan_ptr);
+}
diff --git a/third_party/aom/av1/encoder/arm/neon/error_neon.c b/third_party/aom/av1/encoder/arm/neon/error_neon.c
new file mode 100644
index 0000000000..fe5233f89f
--- /dev/null
+++ b/third_party/aom/av1/encoder/arm/neon/error_neon.c
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./av1_rtcd.h"
+
+int64_t av1_block_error_fp_neon(const int16_t *coeff, const int16_t *dqcoeff,
+                                int block_size) {
+  int64x2_t error = vdupq_n_s64(0);
+
+  assert(block_size >= 8);
+  assert((block_size % 8) == 0);
+
+  do {
+    const int16x8_t c = vld1q_s16(coeff);
+    const int16x8_t d = vld1q_s16(dqcoeff);
+    const int16x8_t diff = vsubq_s16(c, d);
+    const int16x4_t diff_lo = vget_low_s16(diff);
+    const int16x4_t diff_hi = vget_high_s16(diff);
+    // diff is 15-bits, the squares 30, so we can store 2 in 31-bits before
+    // accumulating them in 64-bits.
+    const int32x4_t err0 = vmull_s16(diff_lo, diff_lo);
+    const int32x4_t err1 = vmlal_s16(err0, diff_hi, diff_hi);
+    const int64x2_t err2 = vaddl_s32(vget_low_s32(err1), vget_high_s32(err1));
+    error = vaddq_s64(error, err2);
+    coeff += 8;
+    dqcoeff += 8;
+    block_size -= 8;
+  } while (block_size != 0);
+
+  return vgetq_lane_s64(error, 0) + vgetq_lane_s64(error, 1);
+}
diff --git a/third_party/aom/av1/encoder/arm/neon/quantize_neon.c b/third_party/aom/av1/encoder/arm/neon/quantize_neon.c
new file mode 100644
index 0000000000..36e7d33702
--- /dev/null
+++ b/third_party/aom/av1/encoder/arm/neon/quantize_neon.c
@@ -0,0 +1,118 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+
+#include <math.h>
+
+#include "aom_mem/aom_mem.h"
+
+#include "av1/common/quant_common.h"
+#include "av1/common/seg_common.h"
+
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/rd.h"
+
+void av1_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
+                          int skip_block, const int16_t *zbin_ptr,
+                          const int16_t *round_ptr, const int16_t *quant_ptr,
+                          const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                          int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                          uint16_t *eob_ptr, const int16_t *scan,
+                          const int16_t *iscan) {
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)scan;
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    int i;
+    const int16x8_t v_zero = vdupq_n_s16(0);
+    const int16x8_t v_one = vdupq_n_s16(1);
+    int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
+    int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
+    int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
+    int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    // adjust for dc
+    v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
+    v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
+    v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
+    // process dc and the first seven ac coeffs
+    {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[0]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+      const int32x4_t v_tmp_lo =
+          vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi =
+          vmull_s16(vget_high_s16(v_tmp), vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 =
+          vcombine_s16(vshrn_n_s32(v_tmp_lo, 16), vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+      vst1q_s16(&qcoeff_ptr[0], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[0], v_dqcoeff);
+      v_round = vmovq_n_s16(round_ptr[1]);
+      v_quant = vmovq_n_s16(quant_ptr[1]);
+      v_dequant = vmovq_n_s16(dequant_ptr[1]);
+    }
+    // now process the rest of the ac coeffs
+    for (i = 8; i < count; i += 8) {
+      const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
+      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[i]);
+      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
+      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
+      const int32x4_t v_tmp_lo =
+          vmull_s16(vget_low_s16(v_tmp), vget_low_s16(v_quant));
+      const int32x4_t v_tmp_hi =
+          vmull_s16(vget_high_s16(v_tmp), vget_high_s16(v_quant));
+      const int16x8_t v_tmp2 =
+          vcombine_s16(vshrn_n_s32(v_tmp_lo, 16), vshrn_n_s32(v_tmp_hi, 16));
+      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
+      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
+      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
+      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
+      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
+      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
+      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
+      vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
+      vst1q_s16(&dqcoeff_ptr[i], v_dqcoeff);
+    }
+    {
+      const int16x4_t v_eobmax_3210 = vmax_s16(
+          vget_low_s16(v_eobmax_76543210), vget_high_s16(v_eobmax_76543210));
+      const int64x1_t v_eobmax_xx32 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
+      const int16x4_t v_eobmax_tmp =
+          vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
+      const int64x1_t v_eobmax_xxx3 =
+          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
+      const int16x4_t v_eobmax_final =
+          vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
+
+      *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
+    }
+  } else {
+    memset(qcoeff_ptr, 0, count * sizeof(int16_t));
+    memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
+    *eob_ptr = 0;
+  }
+}
diff --git a/third_party/aom/av1/encoder/av1_quantize.c b/third_party/aom/av1/encoder/av1_quantize.c
new file mode 100644
index 0000000000..6cffac264b
--- /dev/null
+++ b/third_party/aom/av1/encoder/av1_quantize.c
@@ -0,0 +1,1790 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/quantize.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#include "av1/common/idct.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/scan.h"
+#include "av1/common/seg_common.h"
+
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/rd.h"
+
+#if CONFIG_NEW_QUANT
+static INLINE int quantize_coeff_nuq(
+    const tran_low_t coeffv, const int16_t quant, const int16_t quant_shift,
+    const int16_t dequant, const tran_low_t *cuml_bins_ptr,
+    const tran_low_t *dequant_val, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr) {
+  const int coeff = coeffv;
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int i, q;
+  int tmp = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+  for (i = 0; i < NUQ_KNOTS; i++) {
+    if (tmp < cuml_bins_ptr[i]) {
+      q = i;
+      break;
+    }
+  }
+  if (i == NUQ_KNOTS) {
+    tmp -= cuml_bins_ptr[NUQ_KNOTS - 1];
+    q = NUQ_KNOTS + (((((tmp * quant) >> 16) + tmp) * quant_shift) >> 16);
+  }
+  if (q) {
+    *dqcoeff_ptr = av1_dequant_abscoeff_nuq(q, dequant, dequant_val);
+    *qcoeff_ptr = (q ^ coeff_sign) - coeff_sign;
+    *dqcoeff_ptr = *qcoeff_ptr < 0 ? -*dqcoeff_ptr : *dqcoeff_ptr;
+  } else {
+    *qcoeff_ptr = 0;
+    *dqcoeff_ptr = 0;
+  }
+  return (q != 0);
+}
+
+static INLINE int quantize_coeff_bigtx_nuq(
+    const tran_low_t coeffv, const int16_t quant, const int16_t quant_shift,
+    const int16_t dequant, const tran_low_t *cuml_bins_ptr,
+    const tran_low_t *dequant_val, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, int logsizeby16) {
+  const int coeff = coeffv;
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int i, q;
+  int tmp = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+  for (i = 0; i < NUQ_KNOTS; i++) {
+    if (tmp < ROUND_POWER_OF_TWO(cuml_bins_ptr[i], logsizeby16)) {
+      q = i;
+      break;
+    }
+  }
+  if (i == NUQ_KNOTS) {
+    tmp -= ROUND_POWER_OF_TWO(cuml_bins_ptr[NUQ_KNOTS - 1], logsizeby16);
+    q = NUQ_KNOTS +
+        (((((tmp * quant) >> 16) + tmp) * quant_shift) >> (16 - logsizeby16));
+  }
+  if (q) {
+    *dqcoeff_ptr = ROUND_POWER_OF_TWO(
+        av1_dequant_abscoeff_nuq(q, dequant, dequant_val), logsizeby16);
+    // *dqcoeff_ptr = av1_dequant_abscoeff_nuq(q, dequant, dequant_val) >>
+    // (logsizeby16);
+    *qcoeff_ptr = (q ^ coeff_sign) - coeff_sign;
+    *dqcoeff_ptr = *qcoeff_ptr < 0 ? -*dqcoeff_ptr : *dqcoeff_ptr;
+  } else {
+    *qcoeff_ptr = 0;
+    *dqcoeff_ptr = 0;
+  }
+  return (q != 0);
+}
+
+static INLINE int quantize_coeff_fp_nuq(
+    const tran_low_t coeffv, const int16_t quant, const int16_t dequant,
+    const tran_low_t *cuml_bins_ptr, const tran_low_t *dequant_val,
+    tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr) {
+  const int coeff = coeffv;
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int i, q;
+  int tmp = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+  for (i = 0; i < NUQ_KNOTS; i++) {
+    if (tmp < cuml_bins_ptr[i]) {
+      q = i;
+      break;
+    }
+  }
+  if (i == NUQ_KNOTS) {
+    q = NUQ_KNOTS +
+        ((((int64_t)tmp - cuml_bins_ptr[NUQ_KNOTS - 1]) * quant) >> 16);
+  }
+  if (q) {
+    *dqcoeff_ptr = av1_dequant_abscoeff_nuq(q, dequant, dequant_val);
+    *qcoeff_ptr = (q ^ coeff_sign) - coeff_sign;
+    *dqcoeff_ptr = *qcoeff_ptr < 0 ? -*dqcoeff_ptr : *dqcoeff_ptr;
+  } else {
+    *qcoeff_ptr = 0;
+    *dqcoeff_ptr = 0;
+  }
+  return (q != 0);
+}
+
+static INLINE int quantize_coeff_bigtx_fp_nuq(
+    const tran_low_t coeffv, const int16_t quant, const int16_t dequant,
+    const tran_low_t *cuml_bins_ptr, const tran_low_t *dequant_val,
+    tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, int logsizeby16) {
+  const int coeff = coeffv;
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int i, q;
+  int tmp = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+  for (i = 0; i < NUQ_KNOTS; i++) {
+    if (tmp < ROUND_POWER_OF_TWO(cuml_bins_ptr[i], logsizeby16)) {
+      q = i;
+      break;
+    }
+  }
+  if (i == NUQ_KNOTS) {
+    q = NUQ_KNOTS +
+        ((((int64_t)tmp -
+           ROUND_POWER_OF_TWO(cuml_bins_ptr[NUQ_KNOTS - 1], logsizeby16)) *
+          quant) >>
+         (16 - logsizeby16));
+  }
+  if (q) {
+    *dqcoeff_ptr = ROUND_POWER_OF_TWO(
+        av1_dequant_abscoeff_nuq(q, dequant, dequant_val), logsizeby16);
+    // *dqcoeff_ptr = av1_dequant_abscoeff_nuq(q, dequant, dequant_val) >>
+    // (logsizeby16);
+    *qcoeff_ptr = (q ^ coeff_sign) - coeff_sign;
+    *dqcoeff_ptr = *qcoeff_ptr < 0 ? -*dqcoeff_ptr : *dqcoeff_ptr;
+  } else {
+    *qcoeff_ptr = 0;
+    *dqcoeff_ptr = 0;
+  }
+  return (q != 0);
+}
+
+void quantize_dc_nuq(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                     int skip_block, const int16_t quant,
+                     const int16_t quant_shift, const int16_t dequant,
+                     const tran_low_t *cuml_bins_ptr,
+                     const tran_low_t *dequant_val, tran_low_t *qcoeff_ptr,
+                     tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (quantize_coeff_nuq(coeff_ptr[rc], quant, quant_shift, dequant,
+                           cuml_bins_ptr, dequant_val, qcoeff_ptr, dqcoeff_ptr))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void quantize_dc_fp_nuq(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                        int skip_block, const int16_t quant,
+                        const int16_t dequant, const tran_low_t *cuml_bins_ptr,
+                        const tran_low_t *dequant_val, tran_low_t *qcoeff_ptr,
+                        tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (quantize_coeff_fp_nuq(coeff_ptr[rc], quant, dequant, cuml_bins_ptr,
+                              dequant_val, qcoeff_ptr, dqcoeff_ptr))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void quantize_dc_32x32_nuq(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                           int skip_block, const int16_t quant,
+                           const int16_t quant_shift, const int16_t dequant,
+                           const tran_low_t *cuml_bins_ptr,
+                           const tran_low_t *dequant_val,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (quantize_coeff_bigtx_nuq(coeff_ptr[rc], quant, quant_shift, dequant,
+                                 cuml_bins_ptr, dequant_val, qcoeff_ptr,
+                                 dqcoeff_ptr, av1_get_tx_scale(TX_32X32)))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void quantize_dc_32x32_fp_nuq(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                              int skip_block, const int16_t quant,
+                              const int16_t dequant,
+                              const tran_low_t *cuml_bins_ptr,
+                              const tran_low_t *dequant_val,
+                              tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                              uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (quantize_coeff_bigtx_fp_nuq(coeff_ptr[rc], quant, dequant,
+                                    cuml_bins_ptr, dequant_val, qcoeff_ptr,
+                                    dqcoeff_ptr, av1_get_tx_scale(TX_32X32)))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_TX64X64
+void quantize_dc_64x64_nuq(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                           int skip_block, const int16_t quant,
+                           const int16_t quant_shift, const int16_t dequant,
+                           const tran_low_t *cuml_bins_ptr,
+                           const tran_low_t *dequant_val,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (quantize_coeff_bigtx_nuq(coeff_ptr[rc], quant, quant_shift, dequant,
+                                 cuml_bins_ptr, dequant_val, qcoeff_ptr,
+                                 dqcoeff_ptr, av1_get_tx_scale(TX_64X64)))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void quantize_dc_64x64_fp_nuq(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                              int skip_block, const int16_t quant,
+                              const int16_t dequant,
+                              const tran_low_t *cuml_bins_ptr,
+                              const tran_low_t *dequant_val,
+                              tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                              uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (quantize_coeff_bigtx_fp_nuq(coeff_ptr[rc], quant, dequant,
+                                    cuml_bins_ptr, dequant_val, qcoeff_ptr,
+                                    dqcoeff_ptr, av1_get_tx_scale(TX_64X64)))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_TX64X64
+
+void quantize_nuq_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                    int skip_block, const int16_t *quant_ptr,
+                    const int16_t *quant_shift_ptr, const int16_t *dequant_ptr,
+                    const cuml_bins_type_nuq *cuml_bins_ptr,
+                    const dequant_val_type_nuq *dequant_val,
+                    tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                    uint16_t *eob_ptr, const int16_t *scan,
+                    const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (quantize_coeff_nuq(coeff_ptr[rc], quant_ptr[rc != 0],
+                             quant_shift_ptr[rc != 0], dequant_ptr[rc != 0],
+                             cuml_bins_ptr[band[i]], dequant_val[band[i]],
+                             &qcoeff_ptr[rc], &dqcoeff_ptr[rc]))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void quantize_fp_nuq_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                       int skip_block, const int16_t *quant_ptr,
+                       const int16_t *dequant_ptr,
+                       const cuml_bins_type_nuq *cuml_bins_ptr,
+                       const dequant_val_type_nuq *dequant_val,
+                       tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                       uint16_t *eob_ptr, const int16_t *scan,
+                       const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (quantize_coeff_fp_nuq(coeff_ptr[rc], quant_ptr[rc != 0],
+                                dequant_ptr[rc != 0], cuml_bins_ptr[band[i]],
+                                dequant_val[band[i]], &qcoeff_ptr[rc],
+                                &dqcoeff_ptr[rc]))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void quantize_32x32_nuq_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                          int skip_block, const int16_t *quant_ptr,
+                          const int16_t *quant_shift_ptr,
+                          const int16_t *dequant_ptr,
+                          const cuml_bins_type_nuq *cuml_bins_ptr,
+                          const dequant_val_type_nuq *dequant_val,
+                          tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                          uint16_t *eob_ptr, const int16_t *scan,
+                          const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (quantize_coeff_bigtx_nuq(
+              coeff_ptr[rc], quant_ptr[rc != 0], quant_shift_ptr[rc != 0],
+              dequant_ptr[rc != 0], cuml_bins_ptr[band[i]],
+              dequant_val[band[i]], &qcoeff_ptr[rc], &dqcoeff_ptr[rc],
+              av1_get_tx_scale(TX_32X32)))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void quantize_32x32_fp_nuq_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t *quant_ptr,
+                             const int16_t *dequant_ptr,
+                             const cuml_bins_type_nuq *cuml_bins_ptr,
+                             const dequant_val_type_nuq *dequant_val,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             uint16_t *eob_ptr, const int16_t *scan,
+                             const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (quantize_coeff_bigtx_fp_nuq(
+              coeff_ptr[rc], quant_ptr[rc != 0], dequant_ptr[rc != 0],
+              cuml_bins_ptr[band[i]], dequant_val[band[i]], &qcoeff_ptr[rc],
+              &dqcoeff_ptr[rc], av1_get_tx_scale(TX_32X32)))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_TX64X64
+void quantize_64x64_nuq_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                          int skip_block, const int16_t *quant_ptr,
+                          const int16_t *quant_shift_ptr,
+                          const int16_t *dequant_ptr,
+                          const cuml_bins_type_nuq *cuml_bins_ptr,
+                          const dequant_val_type_nuq *dequant_val,
+                          tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                          uint16_t *eob_ptr, const int16_t *scan,
+                          const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (quantize_coeff_bigtx_nuq(
+              coeff_ptr[rc], quant_ptr[rc != 0], quant_shift_ptr[rc != 0],
+              dequant_ptr[rc != 0], cuml_bins_ptr[band[i]],
+              dequant_val[band[i]], &qcoeff_ptr[rc], &dqcoeff_ptr[rc],
+              av1_get_tx_scale(TX_64X64)))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void quantize_64x64_fp_nuq_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t *quant_ptr,
+                             const int16_t *dequant_ptr,
+                             const cuml_bins_type_nuq *cuml_bins_ptr,
+                             const dequant_val_type_nuq *dequant_val,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             uint16_t *eob_ptr, const int16_t *scan,
+                             const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (quantize_coeff_bigtx_fp_nuq(
+              coeff_ptr[rc], quant_ptr[rc != 0], dequant_ptr[rc != 0],
+              cuml_bins_ptr[band[i]], dequant_val[band[i]], &qcoeff_ptr[rc],
+              &dqcoeff_ptr[rc], av1_get_tx_scale(TX_64X64)))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_NEW_QUANT
+
+void av1_quantize_skip(intptr_t n_coeffs, tran_low_t *qcoeff_ptr,
+                       tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr) {
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  *eob_ptr = 0;
+}
+
+static void quantize_fp_helper_c(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan,
+#if CONFIG_AOM_QM
+    const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr,
+#endif
+    int log_scale) {
+  int i, eob = -1;
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+#if CONFIG_AOM_QM
+      const qm_val_t wt = qm_ptr[rc];
+      const qm_val_t iwt = iqm_ptr[rc];
+      const int dequant =
+          (dequant_ptr[rc != 0] * iwt + (1 << (AOM_QM_BITS - 1))) >>
+          AOM_QM_BITS;
+#endif
+      const int coeff_sign = (coeff >> 31);
+      int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      int tmp32 = 0;
+#if CONFIG_AOM_QM
+      if (abs_coeff * wt >=
+          (dequant_ptr[rc != 0] << (AOM_QM_BITS - (1 + log_scale)))) {
+#else
+      if (abs_coeff >= (dequant_ptr[rc != 0] >> (1 + log_scale))) {
+#endif
+        abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], log_scale);
+        abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
+#if CONFIG_AOM_QM
+        tmp32 = (int)((abs_coeff * wt * quant_ptr[rc != 0]) >>
+                      ((16 - log_scale) + AOM_QM_BITS));
+        qcoeff_ptr[rc] = (tmp32 ^ coeff_sign) - coeff_sign;
+        dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant / (1 << log_scale);
+#else
+        tmp32 = (int)((abs_coeff * quant_ptr[rc != 0]) >> (16 - log_scale));
+        qcoeff_ptr[rc] = (tmp32 ^ coeff_sign) - coeff_sign;
+        dqcoeff_ptr[rc] =
+            qcoeff_ptr[rc] * dequant_ptr[rc != 0] / (1 << log_scale);
+#endif
+      }
+
+      if (tmp32) eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                       int skip_block, const int16_t *zbin_ptr,
+                       const int16_t *round_ptr, const int16_t *quant_ptr,
+                       const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+                       tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                       uint16_t *eob_ptr, const int16_t *scan,
+                       const int16_t *iscan
+#if CONFIG_AOM_QM
+                       ,
+                       const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr
+#endif
+                       ) {
+  quantize_fp_helper_c(coeff_ptr, n_coeffs, skip_block, zbin_ptr, round_ptr,
+                       quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
+                       dequant_ptr, eob_ptr, scan, iscan,
+#if CONFIG_AOM_QM
+                       qm_ptr, iqm_ptr,
+#endif
+                       0);
+}
+
+void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t *zbin_ptr,
+                             const int16_t *round_ptr, const int16_t *quant_ptr,
+                             const int16_t *quant_shift_ptr,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                             const int16_t *scan, const int16_t *iscan
+#if CONFIG_AOM_QM
+                             ,
+                             const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr
+#endif
+                             ) {
+  quantize_fp_helper_c(coeff_ptr, n_coeffs, skip_block, zbin_ptr, round_ptr,
+                       quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
+                       dequant_ptr, eob_ptr, scan, iscan,
+#if CONFIG_AOM_QM
+                       qm_ptr, iqm_ptr,
+#endif
+                       1);
+}
+
+#if CONFIG_TX64X64
+void av1_quantize_fp_64x64_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t *zbin_ptr,
+                             const int16_t *round_ptr, const int16_t *quant_ptr,
+                             const int16_t *quant_shift_ptr,
+                             tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                             const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                             const int16_t *scan, const int16_t *iscan
+#if CONFIG_AOM_QM
+                             ,
+                             const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr
+#endif
+                             ) {
+  quantize_fp_helper_c(coeff_ptr, n_coeffs, skip_block, zbin_ptr, round_ptr,
+                       quant_ptr, quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr,
+                       dequant_ptr, eob_ptr, scan, iscan,
+#if CONFIG_AOM_QM
+                       qm_ptr, iqm_ptr,
+#endif
+                       2);
+}
+#endif  // CONFIG_TX64X64
+
+void av1_quantize_fp_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            const MACROBLOCK_PLANE *p, tran_low_t *qcoeff_ptr,
+                            const MACROBLOCKD_PLANE *pd,
+                            tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                            const SCAN_ORDER *sc, const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+#if CONFIG_AOM_QM
+  const qm_val_t *qm_ptr = qparam->qmatrix;
+  const qm_val_t *iqm_ptr = qparam->iqmatrix;
+#endif  // CONFIG_AOM_QM
+
+  switch (qparam->log_scale) {
+    case 0:
+      if (n_coeffs < 16) {
+        // TODO(jingning): Need SIMD implementation for smaller block size
+        // quantization.
+        quantize_fp_helper_c(coeff_ptr, n_coeffs, skip_block, p->zbin,
+                             p->round_fp, p->quant_fp, p->quant_shift,
+                             qcoeff_ptr, dqcoeff_ptr, pd->dequant, eob_ptr,
+                             sc->scan, sc->iscan,
+#if CONFIG_AOM_QM
+                             qm_ptr, iqm_ptr,
+#endif
+                             qparam->log_scale);
+      } else {
+        av1_quantize_fp(coeff_ptr, n_coeffs, skip_block, p->zbin, p->round_fp,
+                        p->quant_fp, p->quant_shift, qcoeff_ptr, dqcoeff_ptr,
+                        pd->dequant, eob_ptr, sc->scan, sc->iscan
+#if CONFIG_AOM_QM
+                        ,
+                        qm_ptr, iqm_ptr
+#endif
+                        );
+      }
+      break;
+    case 1:
+      av1_quantize_fp_32x32(coeff_ptr, n_coeffs, skip_block, p->zbin,
+                            p->round_fp, p->quant_fp, p->quant_shift,
+                            qcoeff_ptr, dqcoeff_ptr, pd->dequant, eob_ptr,
+                            sc->scan, sc->iscan
+#if CONFIG_AOM_QM
+                            ,
+                            qm_ptr, iqm_ptr
+#endif
+                            );
+      break;
+#if CONFIG_TX64X64
+    case 2:
+      av1_quantize_fp_64x64(coeff_ptr, n_coeffs, skip_block, p->zbin,
+                            p->round_fp, p->quant_fp, p->quant_shift,
+                            qcoeff_ptr, dqcoeff_ptr, pd->dequant, eob_ptr,
+                            sc->scan, sc->iscan
+#if CONFIG_AOM_QM
+                            ,
+                            qm_ptr, iqm_ptr
+#endif
+                            );
+      break;
+#endif  // CONFIG_TX64X64
+    default: assert(0);
+  }
+}
+
+void av1_quantize_b_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                           const MACROBLOCK_PLANE *p, tran_low_t *qcoeff_ptr,
+                           const MACROBLOCKD_PLANE *pd, tran_low_t *dqcoeff_ptr,
+                           uint16_t *eob_ptr, const SCAN_ORDER *sc,
+                           const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+#if CONFIG_AOM_QM
+  const qm_val_t *qm_ptr = qparam->qmatrix;
+  const qm_val_t *iqm_ptr = qparam->iqmatrix;
+#endif  // CONFIG_AOM_QM
+
+  switch (qparam->log_scale) {
+    case 0:
+      aom_quantize_b(coeff_ptr, n_coeffs, skip_block, p->zbin, p->round,
+                     p->quant, p->quant_shift, qcoeff_ptr, dqcoeff_ptr,
+                     pd->dequant, eob_ptr, sc->scan, sc->iscan
+#if CONFIG_AOM_QM
+                     ,
+                     qm_ptr, iqm_ptr
+#endif
+                     );
+      break;
+    case 1:
+      aom_quantize_b_32x32(coeff_ptr, n_coeffs, skip_block, p->zbin, p->round,
+                           p->quant, p->quant_shift, qcoeff_ptr, dqcoeff_ptr,
+                           pd->dequant, eob_ptr, sc->scan, sc->iscan
+#if CONFIG_AOM_QM
+                           ,
+                           qm_ptr, iqm_ptr
+#endif
+                           );
+      break;
+#if CONFIG_TX64X64
+    case 2:
+      aom_quantize_b_64x64(coeff_ptr, n_coeffs, skip_block, p->zbin, p->round,
+                           p->quant, p->quant_shift, qcoeff_ptr, dqcoeff_ptr,
+                           pd->dequant, eob_ptr, sc->scan, sc->iscan
+#if CONFIG_AOM_QM
+                           ,
+                           qm_ptr, iqm_ptr
+#endif
+                           );
+      break;
+#endif  // CONFIG_TX64X64
+    default: assert(0);
+  }
+}
+
+void av1_quantize_dc_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            const MACROBLOCK_PLANE *p, tran_low_t *qcoeff_ptr,
+                            const MACROBLOCKD_PLANE *pd,
+                            tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                            const SCAN_ORDER *sc, const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+#if CONFIG_AOM_QM
+  const qm_val_t *qm_ptr = qparam->qmatrix;
+  const qm_val_t *iqm_ptr = qparam->iqmatrix;
+#endif  // CONFIG_AOM_QM
+
+  (void)sc;
+
+  switch (qparam->log_scale) {
+    case 0:
+      aom_quantize_dc(coeff_ptr, (int)n_coeffs, skip_block, p->round,
+                      p->quant_fp[0], qcoeff_ptr, dqcoeff_ptr, pd->dequant[0],
+                      eob_ptr
+#if CONFIG_AOM_QM
+                      ,
+                      qm_ptr, iqm_ptr
+#endif
+                      );
+      break;
+    case 1:
+      aom_quantize_dc_32x32(coeff_ptr, skip_block, p->round, p->quant_fp[0],
+                            qcoeff_ptr, dqcoeff_ptr, pd->dequant[0], eob_ptr
+#if CONFIG_AOM_QM
+                            ,
+                            qm_ptr, iqm_ptr
+#endif
+                            );
+      break;
+#if CONFIG_TX64X64
+      aom_quantize_dc_64x64(coeff_ptr, skip_block, p->round, p->quant_fp[0],
+                            qcoeff_ptr, dqcoeff_ptr, pd->dequant[0], eob_ptr
+#if CONFIG_AOM_QM
+                            ,
+                            qm_ptr, iqm_ptr
+#endif
+                            );
+    case 2: break;
+#endif  // CONFIG_TX64X64
+    default: assert(0);
+  }
+}
+
+#if CONFIG_NEW_QUANT
+void av1_quantize_b_nuq_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                               const MACROBLOCK_PLANE *p,
+                               tran_low_t *qcoeff_ptr,
+                               const MACROBLOCKD_PLANE *pd,
+                               tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                               const SCAN_ORDER *sc,
+                               const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+  const uint8_t *band = get_band_translate(qparam->tx_size);
+  int dq = qparam->dq;
+
+  switch (qparam->log_scale) {
+    case 0:
+      quantize_nuq(coeff_ptr, n_coeffs, skip_block, p->quant, p->quant_shift,
+                   pd->dequant,
+                   (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+                   (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq],
+                   qcoeff_ptr, dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+    case 1:
+      quantize_32x32_nuq(coeff_ptr, n_coeffs, skip_block, p->quant,
+                         p->quant_shift, pd->dequant,
+                         (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+                         (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq],
+                         qcoeff_ptr, dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+#if CONFIG_TX64X64
+    case 2:
+      quantize_64x64_nuq(coeff_ptr, n_coeffs, skip_block, p->quant,
+                         p->quant_shift, pd->dequant,
+                         (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+                         (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq],
+                         qcoeff_ptr, dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+#endif  // CONFIG_TX64X64
+    default: assert(0);
+  }
+}
+
+void av1_quantize_fp_nuq_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                const MACROBLOCK_PLANE *p,
+                                tran_low_t *qcoeff_ptr,
+                                const MACROBLOCKD_PLANE *pd,
+                                tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                const SCAN_ORDER *sc,
+                                const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+  const uint8_t *band = get_band_translate(qparam->tx_size);
+  int dq = qparam->dq;
+
+  switch (qparam->log_scale) {
+    case 0:
+      quantize_fp_nuq(coeff_ptr, n_coeffs, skip_block, p->quant_fp, pd->dequant,
+                      (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+                      (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq],
+                      qcoeff_ptr, dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+    case 1:
+      quantize_32x32_fp_nuq(
+          coeff_ptr, n_coeffs, skip_block, p->quant_fp, pd->dequant,
+          (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+          (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq], qcoeff_ptr,
+          dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+#if CONFIG_TX64X64
+    case 2:
+      quantize_64x64_fp_nuq(
+          coeff_ptr, n_coeffs, skip_block, p->quant_fp, pd->dequant,
+          (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+          (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq], qcoeff_ptr,
+          dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+#endif  // CONFIG_TX64X64
+    default: assert(0);
+  }
+}
+
+void av1_quantize_dc_nuq_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                const MACROBLOCK_PLANE *p,
+                                tran_low_t *qcoeff_ptr,
+                                const MACROBLOCKD_PLANE *pd,
+                                tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                const SCAN_ORDER *sc,
+                                const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+  int dq = qparam->dq;
+  (void)sc;
+
+  switch (qparam->log_scale) {
+    case 0:
+      quantize_dc_fp_nuq(coeff_ptr, n_coeffs, skip_block, p->quant_fp[0],
+                         pd->dequant[0], p->cuml_bins_nuq[dq][0],
+                         pd->dequant_val_nuq[dq][0], qcoeff_ptr, dqcoeff_ptr,
+                         eob_ptr);
+      break;
+    case 1:
+      quantize_dc_32x32_fp_nuq(coeff_ptr, n_coeffs, skip_block, p->quant_fp[0],
+                               pd->dequant[0], p->cuml_bins_nuq[dq][0],
+                               pd->dequant_val_nuq[dq][0], qcoeff_ptr,
+                               dqcoeff_ptr, eob_ptr);
+      break;
+#if CONFIG_TX64X64
+    case 2:
+      quantize_dc_64x64_fp_nuq(coeff_ptr, n_coeffs, skip_block, p->quant_fp[0],
+                               pd->dequant[0], p->cuml_bins_nuq[dq][0],
+                               pd->dequant_val_nuq[dq][0], qcoeff_ptr,
+                               dqcoeff_ptr, eob_ptr);
+      break;
+#endif  // CONFIG_TX64X64
+    default: assert(0);
+  }
+}
+#endif  // CONFIG_NEW_QUANT
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_quantize_fp_facade(const tran_low_t *coeff_ptr,
+                                   intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+                                   tran_low_t *qcoeff_ptr,
+                                   const MACROBLOCKD_PLANE *pd,
+                                   tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                   const SCAN_ORDER *sc,
+                                   const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+#if CONFIG_AOM_QM
+  const qm_val_t *qm_ptr = qparam->qmatrix;
+  const qm_val_t *iqm_ptr = qparam->iqmatrix;
+#endif  // CONFIG_AOM_QM
+
+  if (n_coeffs < 16) {
+    // TODO(jingning): Need SIMD implementation for smaller block size
+    // quantization.
+    av1_highbd_quantize_fp_c(coeff_ptr, n_coeffs, skip_block, p->zbin,
+                             p->round_fp, p->quant_fp, p->quant_shift,
+                             qcoeff_ptr, dqcoeff_ptr, pd->dequant, eob_ptr,
+                             sc->scan, sc->iscan,
+#if CONFIG_AOM_QM
+                             qm_ptr, iqm_ptr,
+#endif
+                             qparam->log_scale);
+    return;
+  }
+
+  av1_highbd_quantize_fp(coeff_ptr, n_coeffs, skip_block, p->zbin, p->round_fp,
+                         p->quant_fp, p->quant_shift, qcoeff_ptr, dqcoeff_ptr,
+                         pd->dequant, eob_ptr, sc->scan, sc->iscan,
+#if CONFIG_AOM_QM
+                         qm_ptr, iqm_ptr,
+#endif
+                         qparam->log_scale);
+}
+
+void av1_highbd_quantize_b_facade(const tran_low_t *coeff_ptr,
+                                  intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+                                  tran_low_t *qcoeff_ptr,
+                                  const MACROBLOCKD_PLANE *pd,
+                                  tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                  const SCAN_ORDER *sc,
+                                  const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+#if CONFIG_AOM_QM
+  const qm_val_t *qm_ptr = qparam->qmatrix;
+  const qm_val_t *iqm_ptr = qparam->iqmatrix;
+#endif  // CONFIG_AOM_QM
+
+  switch (qparam->log_scale) {
+    case 0:
+      aom_highbd_quantize_b(coeff_ptr, n_coeffs, skip_block, p->zbin, p->round,
+                            p->quant, p->quant_shift, qcoeff_ptr, dqcoeff_ptr,
+                            pd->dequant, eob_ptr, sc->scan, sc->iscan
+#if CONFIG_AOM_QM
+                            ,
+                            qm_ptr, iqm_ptr
+#endif
+                            );
+      break;
+    case 1:
+      aom_highbd_quantize_b_32x32(coeff_ptr, n_coeffs, skip_block, p->zbin,
+                                  p->round, p->quant, p->quant_shift,
+                                  qcoeff_ptr, dqcoeff_ptr, pd->dequant, eob_ptr,
+                                  sc->scan, sc->iscan
+#if CONFIG_AOM_QM
+                                  ,
+                                  qm_ptr, iqm_ptr
+#endif
+                                  );
+      break;
+#if CONFIG_TX64X64
+    case 2:
+      aom_highbd_quantize_b_64x64(coeff_ptr, n_coeffs, skip_block, p->zbin,
+                                  p->round, p->quant, p->quant_shift,
+                                  qcoeff_ptr, dqcoeff_ptr, pd->dequant, eob_ptr,
+                                  sc->scan, sc->iscan
+#if CONFIG_AOM_QM
+                                  ,
+                                  qm_ptr, iqm_ptr
+#endif
+                                  );
+      break;
+#endif  // CONFIG_TX64X64
+    default: assert(0);
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE void highbd_quantize_dc(
+    const tran_low_t *coeff_ptr, int n_coeffs, int skip_block,
+    const int16_t *round_ptr, const int16_t quant, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t dequant_ptr, uint16_t *eob_ptr,
+#if CONFIG_AOM_QM
+    const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr,
+#endif
+    const int log_scale) {
+  int eob = -1;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+#if CONFIG_AOM_QM
+  (void)qm_ptr;
+  (void)iqm_ptr;
+#endif
+  if (!skip_block) {
+    const int coeff = coeff_ptr[0];
+    const int coeff_sign = (coeff >> 31);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp = abs_coeff + round_ptr[0];
+    const uint32_t abs_qcoeff = (uint32_t)((tmp * quant) >> (16 - log_scale));
+    qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+    dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr / (1 << log_scale);
+    if (abs_qcoeff) eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_highbd_quantize_dc_facade(const tran_low_t *coeff_ptr,
+                                   intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+                                   tran_low_t *qcoeff_ptr,
+                                   const MACROBLOCKD_PLANE *pd,
+                                   tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                   const SCAN_ORDER *sc,
+                                   const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+#if CONFIG_AOM_QM
+  const qm_val_t *qm_ptr = qparam->qmatrix;
+  const qm_val_t *iqm_ptr = qparam->iqmatrix;
+#endif  // CONFIG_AOM_QM
+
+  (void)sc;
+
+  highbd_quantize_dc(coeff_ptr, (int)n_coeffs, skip_block, p->round,
+                     p->quant_fp[0], qcoeff_ptr, dqcoeff_ptr, pd->dequant[0],
+                     eob_ptr,
+#if CONFIG_AOM_QM
+                     qm_ptr, iqm_ptr,
+#endif
+                     qparam->log_scale);
+}
+
+#if CONFIG_NEW_QUANT
+static INLINE int highbd_quantize_coeff_nuq(
+    const tran_low_t coeffv, const int16_t quant, const int16_t quant_shift,
+    const int16_t dequant, const tran_low_t *cuml_bins_ptr,
+    const tran_low_t *dequant_val, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr) {
+  const int coeff = coeffv;
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int i, q;
+  int64_t tmp = clamp(abs_coeff, INT32_MIN, INT32_MAX);
+  for (i = 0; i < NUQ_KNOTS; i++) {
+    if (tmp < cuml_bins_ptr[i]) {
+      q = i;
+      break;
+    }
+  }
+  if (i == NUQ_KNOTS) {
+    tmp -= cuml_bins_ptr[NUQ_KNOTS - 1];
+    q = NUQ_KNOTS + (int)(((((tmp * quant) >> 16) + tmp) * quant_shift) >> 16);
+  }
+  if (q) {
+    *dqcoeff_ptr = av1_dequant_abscoeff_nuq(q, dequant, dequant_val);
+    *qcoeff_ptr = (q ^ coeff_sign) - coeff_sign;
+    *dqcoeff_ptr = *qcoeff_ptr < 0 ? -*dqcoeff_ptr : *dqcoeff_ptr;
+  } else {
+    *qcoeff_ptr = 0;
+    *dqcoeff_ptr = 0;
+  }
+  return (q != 0);
+}
+
+static INLINE int highbd_quantize_coeff_fp_nuq(
+    const tran_low_t coeffv, const int16_t quant, const int16_t dequant,
+    const tran_low_t *cuml_bins_ptr, const tran_low_t *dequant_val,
+    tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr) {
+  const int coeff = coeffv;
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int i, q;
+  int64_t tmp = clamp(abs_coeff, INT32_MIN, INT32_MAX);
+  for (i = 0; i < NUQ_KNOTS; i++) {
+    if (tmp < cuml_bins_ptr[i]) {
+      q = i;
+      break;
+    }
+  }
+  if (i == NUQ_KNOTS) {
+    q = NUQ_KNOTS + (int)(((tmp - cuml_bins_ptr[NUQ_KNOTS - 1]) * quant) >> 16);
+  }
+  if (q) {
+    *dqcoeff_ptr = av1_dequant_abscoeff_nuq(q, dequant, dequant_val);
+    *qcoeff_ptr = (q ^ coeff_sign) - coeff_sign;
+    *dqcoeff_ptr = *qcoeff_ptr < 0 ? -*dqcoeff_ptr : *dqcoeff_ptr;
+  } else {
+    *qcoeff_ptr = 0;
+    *dqcoeff_ptr = 0;
+  }
+  return (q != 0);
+}
+
+static INLINE int highbd_quantize_coeff_bigtx_fp_nuq(
+    const tran_low_t coeffv, const int16_t quant, const int16_t dequant,
+    const tran_low_t *cuml_bins_ptr, const tran_low_t *dequant_val,
+    tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, int logsizeby16) {
+  const int coeff = coeffv;
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int i, q;
+  int64_t tmp = clamp(abs_coeff, INT32_MIN, INT32_MAX);
+  for (i = 0; i < NUQ_KNOTS; i++) {
+    if (tmp < ROUND_POWER_OF_TWO(cuml_bins_ptr[i], logsizeby16)) {
+      q = i;
+      break;
+    }
+  }
+  if (i == NUQ_KNOTS) {
+    q = NUQ_KNOTS +
+        (int)(((tmp -
+                ROUND_POWER_OF_TWO(cuml_bins_ptr[NUQ_KNOTS - 1], logsizeby16)) *
+               quant) >>
+              (16 - logsizeby16));
+  }
+  if (q) {
+    *dqcoeff_ptr = ROUND_POWER_OF_TWO(
+        av1_dequant_abscoeff_nuq(q, dequant, dequant_val), logsizeby16);
+    *qcoeff_ptr = (q ^ coeff_sign) - coeff_sign;
+    *dqcoeff_ptr = *qcoeff_ptr < 0 ? -*dqcoeff_ptr : *dqcoeff_ptr;
+  } else {
+    *qcoeff_ptr = 0;
+    *dqcoeff_ptr = 0;
+  }
+  return (q != 0);
+}
+
+static INLINE int highbd_quantize_coeff_bigtx_nuq(
+    const tran_low_t coeffv, const int16_t quant, const int16_t quant_shift,
+    const int16_t dequant, const tran_low_t *cuml_bins_ptr,
+    const tran_low_t *dequant_val, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, int logsizeby16) {
+  const int coeff = coeffv;
+  const int coeff_sign = (coeff >> 31);
+  const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+  int i, q;
+  int64_t tmp = clamp(abs_coeff, INT32_MIN, INT32_MAX);
+  for (i = 0; i < NUQ_KNOTS; i++) {
+    if (tmp < ROUND_POWER_OF_TWO(cuml_bins_ptr[i], logsizeby16)) {
+      q = i;
+      break;
+    }
+  }
+  if (i == NUQ_KNOTS) {
+    tmp -= ROUND_POWER_OF_TWO(cuml_bins_ptr[NUQ_KNOTS - 1], logsizeby16);
+    q = NUQ_KNOTS + (int)(((((tmp * quant) >> 16) + tmp) * quant_shift) >>
+                          (16 - logsizeby16));
+  }
+  if (q) {
+    *dqcoeff_ptr = ROUND_POWER_OF_TWO(
+        av1_dequant_abscoeff_nuq(q, dequant, dequant_val), logsizeby16);
+    *qcoeff_ptr = (q ^ coeff_sign) - coeff_sign;
+    *dqcoeff_ptr = *qcoeff_ptr < 0 ? -*dqcoeff_ptr : *dqcoeff_ptr;
+  } else {
+    *qcoeff_ptr = 0;
+    *dqcoeff_ptr = 0;
+  }
+  return (q != 0);
+}
+
+void highbd_quantize_dc_nuq(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            int skip_block, const int16_t quant,
+                            const int16_t quant_shift, const int16_t dequant,
+                            const tran_low_t *cuml_bins_ptr,
+                            const tran_low_t *dequant_val,
+                            tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                            uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (highbd_quantize_coeff_nuq(coeff_ptr[rc], quant, quant_shift, dequant,
+                                  cuml_bins_ptr, dequant_val, qcoeff_ptr,
+                                  dqcoeff_ptr))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void highbd_quantize_dc_fp_nuq(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                               int skip_block, const int16_t quant,
+                               const int16_t dequant,
+                               const tran_low_t *cuml_bins_ptr,
+                               const tran_low_t *dequant_val,
+                               tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                               uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (highbd_quantize_coeff_fp_nuq(coeff_ptr[rc], quant, dequant,
+                                     cuml_bins_ptr, dequant_val, qcoeff_ptr,
+                                     dqcoeff_ptr))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void highbd_quantize_nuq_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                           int skip_block, const int16_t *quant_ptr,
+                           const int16_t *quant_shift_ptr,
+                           const int16_t *dequant_ptr,
+                           const cuml_bins_type_nuq *cuml_bins_ptr,
+                           const dequant_val_type_nuq *dequant_val,
+                           tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                           uint16_t *eob_ptr, const int16_t *scan,
+                           const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (highbd_quantize_coeff_nuq(
+              coeff_ptr[rc], quant_ptr[rc != 0], quant_shift_ptr[rc != 0],
+              dequant_ptr[rc != 0], cuml_bins_ptr[band[i]],
+              dequant_val[band[i]], &qcoeff_ptr[rc], &dqcoeff_ptr[rc]))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void highbd_quantize_32x32_nuq_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                 int skip_block, const int16_t *quant_ptr,
+                                 const int16_t *quant_shift_ptr,
+                                 const int16_t *dequant_ptr,
+                                 const cuml_bins_type_nuq *cuml_bins_ptr,
+                                 const dequant_val_type_nuq *dequant_val,
+                                 tran_low_t *qcoeff_ptr,
+                                 tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                 const int16_t *scan, const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (highbd_quantize_coeff_bigtx_nuq(
+              coeff_ptr[rc], quant_ptr[rc != 0], quant_shift_ptr[rc != 0],
+              dequant_ptr[rc != 0], cuml_bins_ptr[band[i]],
+              dequant_val[band[i]], &qcoeff_ptr[rc], &dqcoeff_ptr[rc],
+              av1_get_tx_scale(TX_32X32)))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void highbd_quantize_32x32_fp_nuq_c(const tran_low_t *coeff_ptr,
+                                    intptr_t n_coeffs, int skip_block,
+                                    const int16_t *quant_ptr,
+                                    const int16_t *dequant_ptr,
+                                    const cuml_bins_type_nuq *cuml_bins_ptr,
+                                    const dequant_val_type_nuq *dequant_val,
+                                    tran_low_t *qcoeff_ptr,
+                                    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                    const int16_t *scan, const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (highbd_quantize_coeff_bigtx_fp_nuq(
+              coeff_ptr[rc], quant_ptr[rc != 0], dequant_ptr[rc != 0],
+              cuml_bins_ptr[band[i]], dequant_val[band[i]], &qcoeff_ptr[rc],
+              &dqcoeff_ptr[rc], av1_get_tx_scale(TX_32X32)))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_TX64X64
+void highbd_quantize_64x64_nuq_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                 int skip_block, const int16_t *quant_ptr,
+                                 const int16_t *quant_shift_ptr,
+                                 const int16_t *dequant_ptr,
+                                 const cuml_bins_type_nuq *cuml_bins_ptr,
+                                 const dequant_val_type_nuq *dequant_val,
+                                 tran_low_t *qcoeff_ptr,
+                                 tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                 const int16_t *scan, const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (highbd_quantize_coeff_bigtx_nuq(
+              coeff_ptr[rc], quant_ptr[rc != 0], quant_shift_ptr[rc != 0],
+              dequant_ptr[rc != 0], cuml_bins_ptr[band[i]],
+              dequant_val[band[i]], &qcoeff_ptr[rc], &dqcoeff_ptr[rc],
+              av1_get_tx_scale(TX_64X64)))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void highbd_quantize_64x64_fp_nuq_c(const tran_low_t *coeff_ptr,
+                                    intptr_t n_coeffs, int skip_block,
+                                    const int16_t *quant_ptr,
+                                    const int16_t *dequant_ptr,
+                                    const cuml_bins_type_nuq *cuml_bins_ptr,
+                                    const dequant_val_type_nuq *dequant_val,
+                                    tran_low_t *qcoeff_ptr,
+                                    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                    const int16_t *scan, const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (highbd_quantize_coeff_bigtx_fp_nuq(
+              coeff_ptr[rc], quant_ptr[rc != 0], dequant_ptr[rc != 0],
+              cuml_bins_ptr[band[i]], dequant_val[band[i]], &qcoeff_ptr[rc],
+              &dqcoeff_ptr[rc], av1_get_tx_scale(TX_64X64)))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_TX64X64
+
+void highbd_quantize_fp_nuq_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                              int skip_block, const int16_t *quant_ptr,
+                              const int16_t *dequant_ptr,
+                              const cuml_bins_type_nuq *cuml_bins_ptr,
+                              const dequant_val_type_nuq *dequant_val,
+                              tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                              uint16_t *eob_ptr, const int16_t *scan,
+                              const uint8_t *band) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    int i;
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      if (highbd_quantize_coeff_fp_nuq(
+              coeff_ptr[rc], quant_ptr[rc != 0], dequant_ptr[rc != 0],
+              cuml_bins_ptr[band[i]], dequant_val[band[i]], &qcoeff_ptr[rc],
+              &dqcoeff_ptr[rc]))
+        eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void highbd_quantize_dc_32x32_nuq(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    const int16_t quant, const int16_t quant_shift, const int16_t dequant,
+    const tran_low_t *cuml_bins_ptr, const tran_low_t *dequant_val,
+    tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (highbd_quantize_coeff_bigtx_nuq(
+            coeff_ptr[rc], quant, quant_shift, dequant, cuml_bins_ptr,
+            dequant_val, qcoeff_ptr, dqcoeff_ptr, av1_get_tx_scale(TX_32X32)))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void highbd_quantize_dc_32x32_fp_nuq(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    const int16_t quant, const int16_t dequant, const tran_low_t *cuml_bins_ptr,
+    const tran_low_t *dequant_val, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (highbd_quantize_coeff_bigtx_fp_nuq(
+            coeff_ptr[rc], quant, dequant, cuml_bins_ptr, dequant_val,
+            qcoeff_ptr, dqcoeff_ptr, av1_get_tx_scale(TX_32X32)))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+#if CONFIG_TX64X64
+void highbd_quantize_dc_64x64_nuq(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    const int16_t quant, const int16_t quant_shift, const int16_t dequant,
+    const tran_low_t *cuml_bins_ptr, const tran_low_t *dequant_val,
+    tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (highbd_quantize_coeff_bigtx_nuq(
+            coeff_ptr[rc], quant, quant_shift, dequant, cuml_bins_ptr,
+            dequant_val, qcoeff_ptr, dqcoeff_ptr, av1_get_tx_scale(TX_64X64)))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void highbd_quantize_dc_64x64_fp_nuq(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
+    const int16_t quant, const int16_t dequant, const tran_low_t *cuml_bins_ptr,
+    const tran_low_t *dequant_val, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr) {
+  int eob = -1;
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+  if (!skip_block) {
+    const int rc = 0;
+    if (highbd_quantize_coeff_bigtx_fp_nuq(
+            coeff_ptr[rc], quant, dequant, cuml_bins_ptr, dequant_val,
+            qcoeff_ptr, dqcoeff_ptr, av1_get_tx_scale(TX_64X64)))
+      eob = 0;
+  }
+  *eob_ptr = eob + 1;
+}
+#endif  // CONFIG_TX64X64
+
+void av1_highbd_quantize_b_nuq_facade(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+    tran_low_t *qcoeff_ptr, const MACROBLOCKD_PLANE *pd,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const SCAN_ORDER *sc,
+    const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+  const uint8_t *band = get_band_translate(qparam->tx_size);
+  const int dq = qparam->dq;
+
+  switch (qparam->log_scale) {
+    case 0:
+      highbd_quantize_nuq(coeff_ptr, n_coeffs, skip_block, p->quant,
+                          p->quant_shift, pd->dequant,
+                          (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+                          (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq],
+                          qcoeff_ptr, dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+    case 1:
+      highbd_quantize_32x32_nuq(
+          coeff_ptr, n_coeffs, skip_block, p->quant, p->quant_shift,
+          pd->dequant, (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+          (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq], qcoeff_ptr,
+          dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+#if CONFIG_TX64X64
+    case 2:
+      highbd_quantize_64x64_nuq(
+          coeff_ptr, n_coeffs, skip_block, p->quant, p->quant_shift,
+          pd->dequant, (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+          (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq], qcoeff_ptr,
+          dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+#endif  // CONFIG_TX64X64
+    default: assert(0);
+  }
+}
+
+void av1_highbd_quantize_fp_nuq_facade(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+    tran_low_t *qcoeff_ptr, const MACROBLOCKD_PLANE *pd,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const SCAN_ORDER *sc,
+    const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+  const uint8_t *band = get_band_translate(qparam->tx_size);
+  const int dq = qparam->dq;
+
+  switch (qparam->log_scale) {
+    case 0:
+      highbd_quantize_fp_nuq(
+          coeff_ptr, n_coeffs, skip_block, p->quant_fp, pd->dequant,
+          (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+          (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq], qcoeff_ptr,
+          dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+    case 1:
+      highbd_quantize_32x32_fp_nuq(
+          coeff_ptr, n_coeffs, skip_block, p->quant_fp, pd->dequant,
+          (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+          (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq], qcoeff_ptr,
+          dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+#if CONFIG_TX64X64
+    case 2:
+      highbd_quantize_64x64_fp_nuq(
+          coeff_ptr, n_coeffs, skip_block, p->quant_fp, pd->dequant,
+          (const cuml_bins_type_nuq *)p->cuml_bins_nuq[dq],
+          (const dequant_val_type_nuq *)pd->dequant_val_nuq[dq], qcoeff_ptr,
+          dqcoeff_ptr, eob_ptr, sc->scan, band);
+      break;
+#endif  // CONFIG_TX64X64
+    default: assert(0);
+  }
+}
+
+void av1_highbd_quantize_dc_nuq_facade(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+    tran_low_t *qcoeff_ptr, const MACROBLOCKD_PLANE *pd,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const SCAN_ORDER *sc,
+    const QUANT_PARAM *qparam) {
+  // obsolete skip_block
+  const int skip_block = 0;
+  const int dq = qparam->dq;
+  (void)sc;
+
+  switch (qparam->log_scale) {
+    case 0:
+      highbd_quantize_dc_fp_nuq(coeff_ptr, n_coeffs, skip_block, p->quant_fp[0],
+                                pd->dequant[0], p->cuml_bins_nuq[dq][0],
+                                pd->dequant_val_nuq[dq][0], qcoeff_ptr,
+                                dqcoeff_ptr, eob_ptr);
+      break;
+    case 1:
+      highbd_quantize_dc_32x32_fp_nuq(
+          coeff_ptr, n_coeffs, skip_block, p->quant_fp[0], pd->dequant[0],
+          p->cuml_bins_nuq[dq][0], pd->dequant_val_nuq[dq][0], qcoeff_ptr,
+          dqcoeff_ptr, eob_ptr);
+      break;
+#if CONFIG_TX64X64
+    case 2:
+      highbd_quantize_dc_64x64_fp_nuq(
+          coeff_ptr, n_coeffs, skip_block, p->quant_fp[0], pd->dequant[0],
+          p->cuml_bins_nuq[dq][0], pd->dequant_val_nuq[dq][0], qcoeff_ptr,
+          dqcoeff_ptr, eob_ptr);
+      break;
+#endif  // CONFIG_TX64X64
+    default: assert(0);
+  }
+}
+#endif  // CONFIG_NEW_QUANT
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t count,
+                              int skip_block, const int16_t *zbin_ptr,
+                              const int16_t *round_ptr,
+                              const int16_t *quant_ptr,
+                              const int16_t *quant_shift_ptr,
+                              tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                              const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                              const int16_t *scan, const int16_t *iscan,
+#if CONFIG_AOM_QM
+                              const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr,
+#endif
+                              int log_scale) {
+  int i;
+  int eob = -1;
+  const int scale = 1 << log_scale;
+  const int shift = 16 - log_scale;
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < count; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+#if CONFIG_AOM_QM
+      const qm_val_t wt = qm_ptr[rc];
+      const qm_val_t iwt = iqm_ptr[rc];
+      const int dequant =
+          (dequant_ptr[rc != 0] * iwt + (1 << (AOM_QM_BITS - 1))) >>
+          AOM_QM_BITS;
+#endif
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int64_t tmp = abs_coeff + round_ptr[rc != 0];
+#if CONFIG_AOM_QM
+      const uint32_t abs_qcoeff =
+          (uint32_t)((tmp * quant_ptr[rc != 0] * wt) >> (shift + AOM_QM_BITS));
+      qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant / scale;
+#else
+      const uint32_t abs_qcoeff =
+          (uint32_t)((tmp * quant_ptr[rc != 0]) >> shift);
+      qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / scale;
+#endif
+      if (abs_qcoeff) eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+#endif  // CONFIG_HIGHBITDEPTH
+
+static void invert_quant(int16_t *quant, int16_t *shift, int d) {
+  uint32_t t;
+  int l, m;
+  t = d;
+  for (l = 0; t > 1; l++) t >>= 1;
+  m = 1 + (1 << (16 + l)) / d;
+  *quant = (int16_t)(m - (1 << 16));
+  *shift = 1 << (16 - l);
+}
+
+static int get_qzbin_factor(int q, aom_bit_depth_t bit_depth) {
+  const int quant = av1_dc_quant(q, 0, bit_depth);
+#if CONFIG_HIGHBITDEPTH
+  switch (bit_depth) {
+    case AOM_BITS_8: return q == 0 ? 64 : (quant < 148 ? 84 : 80);
+    case AOM_BITS_10: return q == 0 ? 64 : (quant < 592 ? 84 : 80);
+    case AOM_BITS_12: return q == 0 ? 64 : (quant < 2368 ? 84 : 80);
+    default:
+      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
+      return -1;
+  }
+#else
+  (void)bit_depth;
+  return q == 0 ? 64 : (quant < 148 ? 84 : 80);
+#endif
+}
+
+void av1_init_quantizer(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  QUANTS *const quants = &cpi->quants;
+  int i, q, quant;
+#if CONFIG_NEW_QUANT
+  int dq;
+#endif
+
+  for (q = 0; q < QINDEX_RANGE; q++) {
+    const int qzbin_factor = get_qzbin_factor(q, cm->bit_depth);
+    const int qrounding_factor = q == 0 ? 64 : 48;
+
+    for (i = 0; i < 2; ++i) {
+      int qrounding_factor_fp = 64;
+      // y
+      quant = i == 0 ? av1_dc_quant(q, cm->y_dc_delta_q, cm->bit_depth)
+                     : av1_ac_quant(q, 0, cm->bit_depth);
+      invert_quant(&quants->y_quant[q][i], &quants->y_quant_shift[q][i], quant);
+      quants->y_quant_fp[q][i] = (1 << 16) / quant;
+      quants->y_round_fp[q][i] = (qrounding_factor_fp * quant) >> 7;
+      quants->y_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
+      quants->y_round[q][i] = (qrounding_factor * quant) >> 7;
+      cpi->y_dequant[q][i] = quant;
+
+      // uv
+      quant = i == 0 ? av1_dc_quant(q, cm->uv_dc_delta_q, cm->bit_depth)
+                     : av1_ac_quant(q, cm->uv_ac_delta_q, cm->bit_depth);
+      invert_quant(&quants->uv_quant[q][i], &quants->uv_quant_shift[q][i],
+                   quant);
+      quants->uv_quant_fp[q][i] = (1 << 16) / quant;
+      quants->uv_round_fp[q][i] = (qrounding_factor_fp * quant) >> 7;
+      quants->uv_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
+      quants->uv_round[q][i] = (qrounding_factor * quant) >> 7;
+      cpi->uv_dequant[q][i] = quant;
+    }
+
+#if CONFIG_NEW_QUANT
+    for (dq = 0; dq < QUANT_PROFILES; dq++) {
+      for (i = 0; i < COEF_BANDS; i++) {
+        const int y_quant = cpi->y_dequant[q][i != 0];
+        const int uvquant = cpi->uv_dequant[q][i != 0];
+        av1_get_dequant_val_nuq(y_quant, i, cpi->y_dequant_val_nuq[dq][q][i],
+                                quants->y_cuml_bins_nuq[dq][q][i], dq);
+        av1_get_dequant_val_nuq(uvquant, i, cpi->uv_dequant_val_nuq[dq][q][i],
+                                quants->uv_cuml_bins_nuq[dq][q][i], dq);
+      }
+    }
+#endif  // CONFIG_NEW_QUANT
+
+    for (i = 2; i < 8; i++) {  // 8: SIMD width
+      quants->y_quant[q][i] = quants->y_quant[q][1];
+      quants->y_quant_fp[q][i] = quants->y_quant_fp[q][1];
+      quants->y_round_fp[q][i] = quants->y_round_fp[q][1];
+      quants->y_quant_shift[q][i] = quants->y_quant_shift[q][1];
+      quants->y_zbin[q][i] = quants->y_zbin[q][1];
+      quants->y_round[q][i] = quants->y_round[q][1];
+      cpi->y_dequant[q][i] = cpi->y_dequant[q][1];
+
+      quants->uv_quant[q][i] = quants->uv_quant[q][1];
+      quants->uv_quant_fp[q][i] = quants->uv_quant_fp[q][1];
+      quants->uv_round_fp[q][i] = quants->uv_round_fp[q][1];
+      quants->uv_quant_shift[q][i] = quants->uv_quant_shift[q][1];
+      quants->uv_zbin[q][i] = quants->uv_zbin[q][1];
+      quants->uv_round[q][i] = quants->uv_round[q][1];
+      cpi->uv_dequant[q][i] = cpi->uv_dequant[q][1];
+    }
+  }
+}
+
+void av1_init_plane_quantizers(const AV1_COMP *cpi, MACROBLOCK *x,
+                               int segment_id) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const QUANTS *const quants = &cpi->quants;
+
+#if CONFIG_DELTA_Q
+#if CONFIG_EXT_DELTA_Q
+  int current_q_index = AOMMAX(
+      0, AOMMIN(QINDEX_RANGE - 1, cpi->oxcf.deltaq_mode != NO_DELTA_Q
+                                      ? cm->base_qindex + xd->delta_qindex
+                                      : cm->base_qindex));
+#else
+  int current_q_index = AOMMAX(
+      0, AOMMIN(QINDEX_RANGE - 1, cm->delta_q_present_flag
+                                      ? cm->base_qindex + xd->delta_qindex
+                                      : cm->base_qindex));
+#endif
+  const int qindex = av1_get_qindex(&cm->seg, segment_id, current_q_index);
+#else
+  const int qindex = av1_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+#endif
+  const int rdmult = av1_compute_rd_mult(cpi, qindex + cm->y_dc_delta_q);
+  int i;
+#if CONFIG_AOM_QM
+  int minqm = cm->min_qmlevel;
+  int maxqm = cm->max_qmlevel;
+  // Quant matrix only depends on the base QP so there is only one set per frame
+  int qmlevel = (xd->lossless[segment_id] || cm->using_qmatrix == 0)
+                    ? NUM_QM_LEVELS - 1
+                    : aom_get_qmlevel(cm->base_qindex, minqm, maxqm);
+#endif
+#if CONFIG_NEW_QUANT
+  int dq;
+#endif
+
+  // Y
+  x->plane[0].quant = quants->y_quant[qindex];
+  x->plane[0].quant_fp = quants->y_quant_fp[qindex];
+  x->plane[0].round_fp = quants->y_round_fp[qindex];
+  x->plane[0].quant_shift = quants->y_quant_shift[qindex];
+  x->plane[0].zbin = quants->y_zbin[qindex];
+  x->plane[0].round = quants->y_round[qindex];
+#if CONFIG_AOM_QM
+  memcpy(&xd->plane[0].seg_qmatrix[segment_id], cm->gqmatrix[qmlevel][0],
+         sizeof(cm->gqmatrix[qmlevel][0]));
+  memcpy(&xd->plane[0].seg_iqmatrix[segment_id], cm->giqmatrix[qmlevel][0],
+         sizeof(cm->giqmatrix[qmlevel][0]));
+#endif
+  xd->plane[0].dequant = cpi->y_dequant[qindex];
+#if CONFIG_NEW_QUANT
+  for (dq = 0; dq < QUANT_PROFILES; dq++) {
+    x->plane[0].cuml_bins_nuq[dq] = quants->y_cuml_bins_nuq[dq][qindex];
+    xd->plane[0].dequant_val_nuq[dq] = cpi->y_dequant_val_nuq[dq][qindex];
+  }
+#endif  // CONFIG_NEW_QUANT
+
+  // UV
+  for (i = 1; i < 3; i++) {
+    x->plane[i].quant = quants->uv_quant[qindex];
+    x->plane[i].quant_fp = quants->uv_quant_fp[qindex];
+    x->plane[i].round_fp = quants->uv_round_fp[qindex];
+    x->plane[i].quant_shift = quants->uv_quant_shift[qindex];
+    x->plane[i].zbin = quants->uv_zbin[qindex];
+    x->plane[i].round = quants->uv_round[qindex];
+#if CONFIG_AOM_QM
+    memcpy(&xd->plane[i].seg_qmatrix[segment_id], cm->gqmatrix[qmlevel][1],
+           sizeof(cm->gqmatrix[qmlevel][1]));
+    memcpy(&xd->plane[i].seg_iqmatrix[segment_id], cm->giqmatrix[qmlevel][1],
+           sizeof(cm->giqmatrix[qmlevel][1]));
+#endif
+    xd->plane[i].dequant = cpi->uv_dequant[qindex];
+#if CONFIG_NEW_QUANT
+    for (dq = 0; dq < QUANT_PROFILES; dq++) {
+      x->plane[i].cuml_bins_nuq[dq] = quants->uv_cuml_bins_nuq[dq][qindex];
+      xd->plane[i].dequant_val_nuq[dq] = cpi->uv_dequant_val_nuq[dq][qindex];
+    }
+#endif  // CONFIG_NEW_QUANT
+  }
+
+  x->skip_block = segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP);
+  x->qindex = qindex;
+
+  set_error_per_bit(x, rdmult);
+
+  av1_initialize_me_consts(cpi, x, qindex);
+}
+
+void av1_frame_init_quantizer(AV1_COMP *cpi) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  av1_init_plane_quantizers(cpi, x, xd->mi[0]->mbmi.segment_id);
+}
+
+void av1_set_quantizer(AV1_COMMON *cm, int q) {
+  // quantizer has to be reinitialized with av1_init_quantizer() if any
+  // delta_q changes.
+  cm->base_qindex = q;
+  cm->y_dc_delta_q = 0;
+  cm->uv_dc_delta_q = 0;
+  cm->uv_ac_delta_q = 0;
+}
+
+// Table that converts 0-63 Q-range values passed in outside to the Qindex
+// range used internally.
+static const int quantizer_to_qindex[] = {
+  0,   4,   8,   12,  16,  20,  24,  28,  32,  36,  40,  44,  48,
+  52,  56,  60,  64,  68,  72,  76,  80,  84,  88,  92,  96,  100,
+  104, 108, 112, 116, 120, 124, 128, 132, 136, 140, 144, 148, 152,
+  156, 160, 164, 168, 172, 176, 180, 184, 188, 192, 196, 200, 204,
+  208, 212, 216, 220, 224, 228, 232, 236, 240, 244, 249, 255,
+};
+
+int av1_quantizer_to_qindex(int quantizer) {
+  return quantizer_to_qindex[quantizer];
+}
+
+int av1_qindex_to_quantizer(int qindex) {
+  int quantizer;
+
+  for (quantizer = 0; quantizer < 64; ++quantizer)
+    if (quantizer_to_qindex[quantizer] >= qindex) return quantizer;
+
+  return 63;
+}
diff --git a/third_party/aom/av1/encoder/av1_quantize.h b/third_party/aom/av1/encoder/av1_quantize.h
new file mode 100644
index 0000000000..c87b6b7dc0
--- /dev/null
+++ b/third_party/aom/av1/encoder/av1_quantize.h
@@ -0,0 +1,184 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_QUANTIZE_H_
+#define AV1_ENCODER_QUANTIZE_H_
+
+#include "./aom_config.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/scan.h"
+#include "av1/encoder/block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct QUANT_PARAM {
+  int log_scale;
+#if CONFIG_NEW_QUANT
+  TX_SIZE tx_size;
+  int dq;
+#endif  // CONFIG_NEW_QUANT
+#if CONFIG_AOM_QM
+  const qm_val_t *qmatrix;
+  const qm_val_t *iqmatrix;
+#endif  // CONFIG_AOM_QM
+} QUANT_PARAM;
+
+typedef void (*AV1_QUANT_FACADE)(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                 const MACROBLOCK_PLANE *p,
+                                 tran_low_t *qcoeff_ptr,
+                                 const MACROBLOCKD_PLANE *pd,
+                                 tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                 const SCAN_ORDER *sc,
+                                 const QUANT_PARAM *qparam);
+
+typedef struct {
+#if CONFIG_NEW_QUANT
+  DECLARE_ALIGNED(
+      16, tran_low_t,
+      y_cuml_bins_nuq[QUANT_PROFILES][QINDEX_RANGE][COEF_BANDS][NUQ_KNOTS]);
+  DECLARE_ALIGNED(
+      16, tran_low_t,
+      uv_cuml_bins_nuq[QUANT_PROFILES][QINDEX_RANGE][COEF_BANDS][NUQ_KNOTS]);
+#endif  // CONFIG_NEW_QUANT
+  // 0: dc 1: ac 2-8: ac repeated to SIMD width
+  DECLARE_ALIGNED(16, int16_t, y_quant[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_quant_shift[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_zbin[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_round[QINDEX_RANGE][8]);
+
+  // TODO(jingning): in progress of re-working the quantization. will decide
+  // if we want to deprecate the current use of y_quant.
+  DECLARE_ALIGNED(16, int16_t, y_quant_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_quant_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, y_round_fp[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_round_fp[QINDEX_RANGE][8]);
+
+  DECLARE_ALIGNED(16, int16_t, uv_quant[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_quant_shift[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_zbin[QINDEX_RANGE][8]);
+  DECLARE_ALIGNED(16, int16_t, uv_round[QINDEX_RANGE][8]);
+} QUANTS;
+
+struct AV1_COMP;
+struct AV1Common;
+
+void av1_frame_init_quantizer(struct AV1_COMP *cpi);
+
+void av1_init_plane_quantizers(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                               int segment_id);
+
+void av1_init_quantizer(struct AV1_COMP *cpi);
+
+void av1_set_quantizer(struct AV1Common *cm, int q);
+
+int av1_quantizer_to_qindex(int quantizer);
+
+int av1_qindex_to_quantizer(int qindex);
+
+void av1_quantize_skip(intptr_t n_coeffs, tran_low_t *qcoeff_ptr,
+                       tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr);
+
+void av1_quantize_fp_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            const MACROBLOCK_PLANE *p, tran_low_t *qcoeff_ptr,
+                            const MACROBLOCKD_PLANE *pd,
+                            tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                            const SCAN_ORDER *sc, const QUANT_PARAM *qparam);
+
+void av1_quantize_b_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                           const MACROBLOCK_PLANE *p, tran_low_t *qcoeff_ptr,
+                           const MACROBLOCKD_PLANE *pd, tran_low_t *dqcoeff_ptr,
+                           uint16_t *eob_ptr, const SCAN_ORDER *sc,
+                           const QUANT_PARAM *qparam);
+
+void av1_quantize_dc_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                            const MACROBLOCK_PLANE *p, tran_low_t *qcoeff_ptr,
+                            const MACROBLOCKD_PLANE *pd,
+                            tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                            const SCAN_ORDER *sc, const QUANT_PARAM *qparam);
+
+#if CONFIG_NEW_QUANT
+void av1_quantize_fp_nuq_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                const MACROBLOCK_PLANE *p,
+                                tran_low_t *qcoeff_ptr,
+                                const MACROBLOCKD_PLANE *pd,
+                                tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                const SCAN_ORDER *sc,
+                                const QUANT_PARAM *qparam);
+
+void av1_quantize_b_nuq_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                               const MACROBLOCK_PLANE *p,
+                               tran_low_t *qcoeff_ptr,
+                               const MACROBLOCKD_PLANE *pd,
+                               tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                               const SCAN_ORDER *sc, const QUANT_PARAM *qparam);
+
+void av1_quantize_dc_nuq_facade(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                const MACROBLOCK_PLANE *p,
+                                tran_low_t *qcoeff_ptr,
+                                const MACROBLOCKD_PLANE *pd,
+                                tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                const SCAN_ORDER *sc,
+                                const QUANT_PARAM *qparam);
+#endif  // CONFIG_NEW_QUANT
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_quantize_fp_facade(const tran_low_t *coeff_ptr,
+                                   intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+                                   tran_low_t *qcoeff_ptr,
+                                   const MACROBLOCKD_PLANE *pd,
+                                   tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                   const SCAN_ORDER *sc,
+                                   const QUANT_PARAM *qparam);
+
+void av1_highbd_quantize_b_facade(const tran_low_t *coeff_ptr,
+                                  intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+                                  tran_low_t *qcoeff_ptr,
+                                  const MACROBLOCKD_PLANE *pd,
+                                  tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                  const SCAN_ORDER *sc,
+                                  const QUANT_PARAM *qparam);
+
+void av1_highbd_quantize_dc_facade(const tran_low_t *coeff_ptr,
+                                   intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+                                   tran_low_t *qcoeff_ptr,
+                                   const MACROBLOCKD_PLANE *pd,
+                                   tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr,
+                                   const SCAN_ORDER *sc,
+                                   const QUANT_PARAM *qparam);
+
+#if CONFIG_NEW_QUANT
+void av1_highbd_quantize_fp_nuq_facade(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+    tran_low_t *qcoeff_ptr, const MACROBLOCKD_PLANE *pd,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const SCAN_ORDER *sc,
+    const QUANT_PARAM *qparam);
+
+void av1_highbd_quantize_b_nuq_facade(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+    tran_low_t *qcoeff_ptr, const MACROBLOCKD_PLANE *pd,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const SCAN_ORDER *sc,
+    const QUANT_PARAM *qparam);
+
+void av1_highbd_quantize_dc_nuq_facade(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const MACROBLOCK_PLANE *p,
+    tran_low_t *qcoeff_ptr, const MACROBLOCKD_PLANE *pd,
+    tran_low_t *dqcoeff_ptr, uint16_t *eob_ptr, const SCAN_ORDER *sc,
+    const QUANT_PARAM *qparam);
+#endif  // CONFIG_NEW_QUANT
+#endif  // CONFIG_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_QUANTIZE_H_
diff --git a/third_party/aom/av1/encoder/bitstream.c b/third_party/aom/av1/encoder/bitstream.c
new file mode 100644
index 0000000000..7cc6179ead
--- /dev/null
+++ b/third_party/aom/av1/encoder/bitstream.c
@@ -0,0 +1,5399 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <stdio.h>
+
+#include "aom/aom_encoder.h"
+#include "aom_dsp/bitwriter_buffer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/binary_codes_writer.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem_ops.h"
+#include "aom_ports/system_state.h"
+#if CONFIG_BITSTREAM_DEBUG
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#if CONFIG_CDEF
+#include "av1/common/cdef.h"
+#include "av1/common/clpf.h"
+#endif  // CONFIG_CDEF
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/entropymv.h"
+#include "av1/common/mvref_common.h"
+#include "av1/common/odintrin.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/reconinter.h"
+#if CONFIG_EXT_INTRA
+#include "av1/common/reconintra.h"
+#endif  // CONFIG_EXT_INTRA
+#include "av1/common/seg_common.h"
+#include "av1/common/tile_common.h"
+
+#if CONFIG_ANS
+#include "aom_dsp/buf_ans.h"
+#endif  // CONFIG_ANS
+#if CONFIG_LV_MAP
+#include "av1/encoder/encodetxb.h"
+#endif  // CONFIG_LV_MAP
+#include "av1/encoder/bitstream.h"
+#include "av1/encoder/cost.h"
+#include "av1/encoder/encodemv.h"
+#include "av1/encoder/mcomp.h"
+#if CONFIG_PALETTE && CONFIG_PALETTE_DELTA_ENCODING
+#include "av1/encoder/palette.h"
+#endif  // CONFIG_PALETTE && CONFIG_PALETTE_DELTA_ENCODING
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/subexp.h"
+#include "av1/encoder/tokenize.h"
+#if CONFIG_PVQ
+#include "av1/encoder/pvq_encoder.h"
+#endif
+
+static struct av1_token intra_mode_encodings[INTRA_MODES];
+static struct av1_token switchable_interp_encodings[SWITCHABLE_FILTERS];
+#if CONFIG_EXT_PARTITION_TYPES && !CONFIG_EC_MULTISYMBOL
+static const struct av1_token ext_partition_encodings[EXT_PARTITION_TYPES] = {
+  { 0, 1 },  { 4, 3 },  { 12, 4 }, { 7, 3 },
+  { 10, 4 }, { 11, 4 }, { 26, 5 }, { 27, 5 }
+};
+#endif
+static struct av1_token partition_encodings[PARTITION_TYPES];
+#if !CONFIG_REF_MV
+static struct av1_token inter_mode_encodings[INTER_MODES];
+#endif
+#if CONFIG_EXT_INTER
+static const struct av1_token
+    inter_compound_mode_encodings[INTER_COMPOUND_MODES] = {
+      { 2, 2 },  { 50, 6 }, { 51, 6 }, { 24, 5 }, { 52, 6 },
+      { 53, 6 }, { 54, 6 }, { 55, 6 }, { 0, 1 },  { 7, 3 }
+    };
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_PALETTE
+static struct av1_token palette_size_encodings[PALETTE_SIZES];
+static struct av1_token palette_color_index_encodings[PALETTE_SIZES]
+                                                     [PALETTE_COLORS];
+#endif  // CONFIG_PALETTE
+#if !CONFIG_EC_MULTISYMBOL
+static const struct av1_token tx_size_encodings[MAX_TX_DEPTH][TX_SIZES] = {
+  { { 0, 1 }, { 1, 1 } },                      // Max tx_size is 8X8
+  { { 0, 1 }, { 2, 2 }, { 3, 2 } },            // Max tx_size is 16X16
+  { { 0, 1 }, { 2, 2 }, { 6, 3 }, { 7, 3 } },  // Max tx_size is 32X32
+#if CONFIG_TX64X64
+  { { 0, 1 }, { 2, 2 }, { 6, 3 }, { 14, 4 }, { 15, 4 } },  // Max tx_size 64X64
+#endif                                                     // CONFIG_TX64X64
+};
+#endif
+
+#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
+static INLINE void write_uniform(aom_writer *w, int n, int v) {
+  const int l = get_unsigned_bits(n);
+  const int m = (1 << l) - n;
+  if (l == 0) return;
+  if (v < m) {
+    aom_write_literal(w, v, l - 1);
+  } else {
+    aom_write_literal(w, m + ((v - m) >> 1), l - 1);
+    aom_write_literal(w, (v - m) & 1, 1);
+  }
+}
+#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
+
+#if CONFIG_EXT_TX
+static struct av1_token ext_tx_inter_encodings[EXT_TX_SETS_INTER][TX_TYPES];
+static struct av1_token ext_tx_intra_encodings[EXT_TX_SETS_INTRA][TX_TYPES];
+#else
+static struct av1_token ext_tx_encodings[TX_TYPES];
+#endif  // CONFIG_EXT_TX
+#if CONFIG_GLOBAL_MOTION
+static struct av1_token global_motion_types_encodings[GLOBAL_TRANS_TYPES];
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+static struct av1_token intra_filter_encodings[INTRA_FILTERS];
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_EXT_INTER
+static struct av1_token interintra_mode_encodings[INTERINTRA_MODES];
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+static struct av1_token compound_type_encodings[COMPOUND_TYPES];
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+static struct av1_token motion_mode_encodings[MOTION_MODES];
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+#if CONFIG_LOOP_RESTORATION
+static struct av1_token switchable_restore_encodings[RESTORE_SWITCHABLE_TYPES];
+#endif  // CONFIG_LOOP_RESTORATION
+static void write_uncompressed_header(AV1_COMP *cpi,
+                                      struct aom_write_bit_buffer *wb);
+static uint32_t write_compressed_header(AV1_COMP *cpi, uint8_t *data);
+static int remux_tiles(const AV1_COMMON *const cm, uint8_t *dst,
+                       const uint32_t data_size, const uint32_t max_tile_size,
+                       const uint32_t max_tile_col_size,
+                       int *const tile_size_bytes,
+                       int *const tile_col_size_bytes);
+
+void av1_encode_token_init(void) {
+#if CONFIG_EXT_TX || CONFIG_PALETTE
+  int s;
+#endif  // CONFIG_EXT_TX || CONFIG_PALETTE
+#if CONFIG_EXT_TX
+  for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
+    av1_tokens_from_tree(ext_tx_inter_encodings[s], av1_ext_tx_inter_tree[s]);
+  }
+  for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
+    av1_tokens_from_tree(ext_tx_intra_encodings[s], av1_ext_tx_intra_tree[s]);
+  }
+#else
+  av1_tokens_from_tree(ext_tx_encodings, av1_ext_tx_tree);
+#endif  // CONFIG_EXT_TX
+  av1_tokens_from_tree(intra_mode_encodings, av1_intra_mode_tree);
+  av1_tokens_from_tree(switchable_interp_encodings, av1_switchable_interp_tree);
+  av1_tokens_from_tree(partition_encodings, av1_partition_tree);
+#if !CONFIG_REF_MV
+  av1_tokens_from_tree(inter_mode_encodings, av1_inter_mode_tree);
+#endif
+
+#if CONFIG_PALETTE
+  av1_tokens_from_tree(palette_size_encodings, av1_palette_size_tree);
+  for (s = 0; s < PALETTE_SIZES; ++s) {
+    av1_tokens_from_tree(palette_color_index_encodings[s],
+                         av1_palette_color_index_tree[s]);
+  }
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+  av1_tokens_from_tree(intra_filter_encodings, av1_intra_filter_tree);
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+#if CONFIG_EXT_INTER
+  av1_tokens_from_tree(interintra_mode_encodings, av1_interintra_mode_tree);
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+  av1_tokens_from_tree(compound_type_encodings, av1_compound_type_tree);
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  av1_tokens_from_tree(motion_mode_encodings, av1_motion_mode_tree);
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+#if CONFIG_GLOBAL_MOTION
+  av1_tokens_from_tree(global_motion_types_encodings,
+                       av1_global_motion_types_tree);
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_LOOP_RESTORATION
+  av1_tokens_from_tree(switchable_restore_encodings,
+                       av1_switchable_restore_tree);
+#endif  // CONFIG_LOOP_RESTORATION
+
+#if CONFIG_EC_MULTISYMBOL
+  /* This hack is necessary when CONFIG_DUAL_FILTER is enabled because the five
+      SWITCHABLE_FILTERS are not consecutive, e.g., 0, 1, 2, 3, 4, when doing
+      an in-order traversal of the av1_switchable_interp_tree structure. */
+  av1_indices_from_tree(av1_switchable_interp_ind, av1_switchable_interp_inv,
+                        av1_switchable_interp_tree);
+/* This hack is necessary because the four TX_TYPES are not consecutive,
+    e.g., 0, 1, 2, 3, when doing an in-order traversal of the av1_ext_tx_tree
+    structure. */
+#if CONFIG_EXT_TX
+  for (s = 1; s < EXT_TX_SETS_INTRA; ++s)
+    av1_indices_from_tree(av1_ext_tx_intra_ind[s], av1_ext_tx_intra_inv[s],
+                          av1_ext_tx_intra_tree[s]);
+  for (s = 1; s < EXT_TX_SETS_INTER; ++s)
+    av1_indices_from_tree(av1_ext_tx_inter_ind[s], av1_ext_tx_inter_inv[s],
+                          av1_ext_tx_inter_tree[s]);
+#else
+  av1_indices_from_tree(av1_ext_tx_ind, av1_ext_tx_inv, av1_ext_tx_tree);
+#endif
+  av1_indices_from_tree(av1_intra_mode_ind, av1_intra_mode_inv,
+                        av1_intra_mode_tree);
+  av1_indices_from_tree(av1_inter_mode_ind, av1_inter_mode_inv,
+                        av1_inter_mode_tree);
+#endif
+}
+
+static void write_intra_mode_kf(const AV1_COMMON *cm, FRAME_CONTEXT *frame_ctx,
+                                const MODE_INFO *mi, const MODE_INFO *above_mi,
+                                const MODE_INFO *left_mi, int block,
+                                PREDICTION_MODE mode, aom_writer *w) {
+#if CONFIG_INTRABC
+  assert(!is_intrabc_block(&mi->mbmi));
+#endif  // CONFIG_INTRABC
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, av1_intra_mode_ind[mode],
+                   get_y_mode_cdf(frame_ctx, mi, above_mi, left_mi, block),
+                   INTRA_MODES);
+  (void)cm;
+#else
+  av1_write_token(w, av1_intra_mode_tree,
+                  get_y_mode_probs(cm, mi, above_mi, left_mi, block),
+                  &intra_mode_encodings[mode]);
+  (void)frame_ctx;
+#endif
+}
+
+#if CONFIG_EXT_INTER
+static void write_interintra_mode(aom_writer *w, INTERINTRA_MODE mode,
+                                  const aom_prob *probs) {
+  av1_write_token(w, av1_interintra_mode_tree, probs,
+                  &interintra_mode_encodings[mode]);
+}
+#endif  // CONFIG_EXT_INTER
+
+static void write_inter_mode(aom_writer *w, PREDICTION_MODE mode,
+                             FRAME_CONTEXT *ec_ctx, const int16_t mode_ctx) {
+#if CONFIG_REF_MV
+  const int16_t newmv_ctx = mode_ctx & NEWMV_CTX_MASK;
+  const aom_prob newmv_prob = ec_ctx->newmv_prob[newmv_ctx];
+
+#define IS_NEWMV_MODE(mode) ((mode) == NEWMV)
+  aom_write(w, !IS_NEWMV_MODE(mode), newmv_prob);
+
+  if (!IS_NEWMV_MODE(mode)) {
+    const int16_t zeromv_ctx = (mode_ctx >> ZEROMV_OFFSET) & ZEROMV_CTX_MASK;
+    const aom_prob zeromv_prob = ec_ctx->zeromv_prob[zeromv_ctx];
+
+    if (mode_ctx & (1 << ALL_ZERO_FLAG_OFFSET)) {
+      assert(mode == ZEROMV);
+      return;
+    }
+
+    aom_write(w, mode != ZEROMV, zeromv_prob);
+
+    if (mode != ZEROMV) {
+      int16_t refmv_ctx = (mode_ctx >> REFMV_OFFSET) & REFMV_CTX_MASK;
+      aom_prob refmv_prob;
+
+      if (mode_ctx & (1 << SKIP_NEARESTMV_OFFSET)) refmv_ctx = 6;
+      if (mode_ctx & (1 << SKIP_NEARMV_OFFSET)) refmv_ctx = 7;
+      if (mode_ctx & (1 << SKIP_NEARESTMV_SUB8X8_OFFSET)) refmv_ctx = 8;
+
+      refmv_prob = ec_ctx->refmv_prob[refmv_ctx];
+      aom_write(w, mode != NEARESTMV, refmv_prob);
+    }
+  }
+
+#undef IS_NEWMV_MODE
+
+#else  // !CONFIG_REF_MV
+  assert(is_inter_mode(mode));
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, av1_inter_mode_ind[INTER_OFFSET(mode)],
+                   ec_ctx->inter_mode_cdf[mode_ctx], INTER_MODES);
+#else
+  {
+    const aom_prob *const inter_probs = ec_ctx->inter_mode_probs[mode_ctx];
+    av1_write_token(w, av1_inter_mode_tree, inter_probs,
+                    &inter_mode_encodings[INTER_OFFSET(mode)]);
+  }
+#endif
+#endif
+}
+
+#if CONFIG_REF_MV
+static void write_drl_idx(const AV1_COMMON *cm, const MB_MODE_INFO *mbmi,
+                          const MB_MODE_INFO_EXT *mbmi_ext, aom_writer *w) {
+  uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+
+  assert(mbmi->ref_mv_idx < 3);
+
+#if CONFIG_EXT_INTER
+  if (mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV) {
+#else
+  if (mbmi->mode == NEWMV) {
+#endif
+    int idx;
+    for (idx = 0; idx < 2; ++idx) {
+      if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
+        uint8_t drl_ctx =
+            av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx);
+        aom_prob drl_prob = cm->fc->drl_prob[drl_ctx];
+
+        aom_write(w, mbmi->ref_mv_idx != idx, drl_prob);
+        if (mbmi->ref_mv_idx == idx) return;
+      }
+    }
+    return;
+  }
+
+  if (have_nearmv_in_inter_mode(mbmi->mode)) {
+    int idx;
+    // TODO(jingning): Temporary solution to compensate the NEARESTMV offset.
+    for (idx = 1; idx < 3; ++idx) {
+      if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
+        uint8_t drl_ctx =
+            av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx);
+        aom_prob drl_prob = cm->fc->drl_prob[drl_ctx];
+
+        aom_write(w, mbmi->ref_mv_idx != (idx - 1), drl_prob);
+        if (mbmi->ref_mv_idx == (idx - 1)) return;
+      }
+    }
+    return;
+  }
+}
+#endif
+
+#if CONFIG_EXT_INTER
+static void write_inter_compound_mode(AV1_COMMON *cm, aom_writer *w,
+                                      PREDICTION_MODE mode,
+                                      const int16_t mode_ctx) {
+  const aom_prob *const inter_compound_probs =
+      cm->fc->inter_compound_mode_probs[mode_ctx];
+
+  assert(is_inter_compound_mode(mode));
+  av1_write_token(w, av1_inter_compound_mode_tree, inter_compound_probs,
+                  &inter_compound_mode_encodings[INTER_COMPOUND_OFFSET(mode)]);
+}
+#endif  // CONFIG_EXT_INTER
+
+static void encode_unsigned_max(struct aom_write_bit_buffer *wb, int data,
+                                int max) {
+  aom_wb_write_literal(wb, data, get_unsigned_bits(max));
+}
+
+#if !CONFIG_EC_ADAPT || \
+    (CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION || CONFIG_EXT_INTER)
+static void prob_diff_update(const aom_tree_index *tree,
+                             aom_prob probs[/*n - 1*/],
+                             const unsigned int counts[/*n - 1*/], int n,
+                             int probwt, aom_writer *w) {
+  int i;
+  unsigned int branch_ct[32][2];
+
+  // Assuming max number of probabilities <= 32
+  assert(n <= 32);
+
+  av1_tree_probs_from_distribution(tree, branch_ct, counts);
+  for (i = 0; i < n - 1; ++i)
+    av1_cond_prob_diff_update(w, &probs[i], branch_ct[i], probwt);
+}
+#endif
+
+#if CONFIG_EXT_INTER || !CONFIG_EC_ADAPT
+static int prob_diff_update_savings(const aom_tree_index *tree,
+                                    aom_prob probs[/*n - 1*/],
+                                    const unsigned int counts[/*n - 1*/], int n,
+                                    int probwt) {
+  int i;
+  unsigned int branch_ct[32][2];
+  int savings = 0;
+
+  // Assuming max number of probabilities <= 32
+  assert(n <= 32);
+  av1_tree_probs_from_distribution(tree, branch_ct, counts);
+  for (i = 0; i < n - 1; ++i) {
+    savings +=
+        av1_cond_prob_diff_update_savings(&probs[i], branch_ct[i], probwt);
+  }
+  return savings;
+}
+#endif  // CONFIG_EXT_INTER || !CONFIG_EC_ADAPT
+
+#if CONFIG_VAR_TX
+static void write_tx_size_vartx(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                                const MB_MODE_INFO *mbmi, TX_SIZE tx_size,
+                                int depth, int blk_row, int blk_col,
+                                aom_writer *w) {
+  const int tx_row = blk_row >> 1;
+  const int tx_col = blk_col >> 1;
+  const int max_blocks_high = max_block_high(xd, mbmi->sb_type, 0);
+  const int max_blocks_wide = max_block_wide(xd, mbmi->sb_type, 0);
+
+  int ctx = txfm_partition_context(xd->above_txfm_context + tx_col,
+                                   xd->left_txfm_context + tx_row,
+                                   mbmi->sb_type, tx_size);
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  if (depth == MAX_VARTX_DEPTH) {
+    txfm_partition_update(xd->above_txfm_context + tx_col,
+                          xd->left_txfm_context + tx_row, tx_size, tx_size);
+    return;
+  }
+
+  if (tx_size == mbmi->inter_tx_size[tx_row][tx_col]) {
+    aom_write(w, 0, cm->fc->txfm_partition_prob[ctx]);
+    txfm_partition_update(xd->above_txfm_context + tx_col,
+                          xd->left_txfm_context + tx_row, tx_size, tx_size);
+  } else {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int i;
+
+    aom_write(w, 1, cm->fc->txfm_partition_prob[ctx]);
+
+    if (tx_size == TX_8X8) {
+      txfm_partition_update(xd->above_txfm_context + tx_col,
+                            xd->left_txfm_context + tx_row, sub_txs, tx_size);
+      return;
+    }
+
+    assert(bsl > 0);
+    for (i = 0; i < 4; ++i) {
+      int offsetr = blk_row + (i >> 1) * bsl;
+      int offsetc = blk_col + (i & 0x01) * bsl;
+      write_tx_size_vartx(cm, xd, mbmi, sub_txs, depth + 1, offsetr, offsetc,
+                          w);
+    }
+  }
+}
+
+static void update_txfm_partition_probs(AV1_COMMON *cm, aom_writer *w,
+                                        FRAME_COUNTS *counts, int probwt) {
+  int k;
+  for (k = 0; k < TXFM_PARTITION_CONTEXTS; ++k)
+    av1_cond_prob_diff_update(w, &cm->fc->txfm_partition_prob[k],
+                              counts->txfm_partition[k], probwt);
+}
+#endif
+
+static void write_selected_tx_size(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                                   aom_writer *w) {
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+// For sub8x8 blocks the tx_size symbol does not need to be sent
+#if CONFIG_CB4X4 && (CONFIG_VAR_TX || CONFIG_EXT_TX) && CONFIG_RECT_TX
+  if (bsize > BLOCK_4X4) {
+#else
+  if (bsize >= BLOCK_8X8) {
+#endif
+    const TX_SIZE tx_size = mbmi->tx_size;
+    const int is_inter = is_inter_block(mbmi);
+    const int tx_size_ctx = get_tx_size_context(xd);
+    const int tx_size_cat = is_inter ? inter_tx_size_cat_lookup[bsize]
+                                     : intra_tx_size_cat_lookup[bsize];
+    const TX_SIZE coded_tx_size = txsize_sqr_up_map[tx_size];
+    const int depth = tx_size_to_depth(coded_tx_size);
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+    assert(IMPLIES(is_rect_tx(tx_size), is_rect_tx_allowed(xd, mbmi)));
+    assert(
+        IMPLIES(is_rect_tx(tx_size), tx_size == max_txsize_rect_lookup[bsize]));
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+
+#if CONFIG_EC_MULTISYMBOL
+    aom_write_symbol(w, depth, ec_ctx->tx_size_cdf[tx_size_cat][tx_size_ctx],
+                     tx_size_cat + 2);
+#else
+    av1_write_token(w, av1_tx_size_tree[tx_size_cat],
+                    ec_ctx->tx_size_probs[tx_size_cat][tx_size_ctx],
+                    &tx_size_encodings[tx_size_cat][depth]);
+#endif
+  }
+}
+
+#if CONFIG_REF_MV
+static void update_inter_mode_probs(AV1_COMMON *cm, aom_writer *w,
+                                    FRAME_COUNTS *counts) {
+  int i;
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+  for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i)
+    av1_cond_prob_diff_update(w, &cm->fc->newmv_prob[i], counts->newmv_mode[i],
+                              probwt);
+  for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i)
+    av1_cond_prob_diff_update(w, &cm->fc->zeromv_prob[i],
+                              counts->zeromv_mode[i], probwt);
+  for (i = 0; i < REFMV_MODE_CONTEXTS; ++i)
+    av1_cond_prob_diff_update(w, &cm->fc->refmv_prob[i], counts->refmv_mode[i],
+                              probwt);
+  for (i = 0; i < DRL_MODE_CONTEXTS; ++i)
+    av1_cond_prob_diff_update(w, &cm->fc->drl_prob[i], counts->drl_mode[i],
+                              probwt);
+}
+#endif
+
+#if CONFIG_EXT_INTER
+static void update_inter_compound_mode_probs(AV1_COMMON *cm, int probwt,
+                                             aom_writer *w) {
+  const int savings_thresh = av1_cost_one(GROUP_DIFF_UPDATE_PROB) -
+                             av1_cost_zero(GROUP_DIFF_UPDATE_PROB);
+  int i;
+  int savings = 0;
+  int do_update = 0;
+  for (i = 0; i < INTER_MODE_CONTEXTS; ++i) {
+    savings += prob_diff_update_savings(
+        av1_inter_compound_mode_tree, cm->fc->inter_compound_mode_probs[i],
+        cm->counts.inter_compound_mode[i], INTER_COMPOUND_MODES, probwt);
+  }
+  do_update = savings > savings_thresh;
+  aom_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
+  if (do_update) {
+    for (i = 0; i < INTER_MODE_CONTEXTS; ++i) {
+      prob_diff_update(
+          av1_inter_compound_mode_tree, cm->fc->inter_compound_mode_probs[i],
+          cm->counts.inter_compound_mode[i], INTER_COMPOUND_MODES, probwt, w);
+    }
+  }
+}
+#endif  // CONFIG_EXT_INTER
+
+static int write_skip(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                      int segment_id, const MODE_INFO *mi, aom_writer *w) {
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
+    return 1;
+  } else {
+    const int skip = mi->mbmi.skip;
+    aom_write(w, skip, av1_get_skip_prob(cm, xd));
+    return skip;
+  }
+}
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+static void write_motion_mode(const AV1_COMMON *cm, const MODE_INFO *mi,
+                              aom_writer *w) {
+  const MB_MODE_INFO *mbmi = &mi->mbmi;
+  MOTION_MODE last_motion_mode_allowed = motion_mode_allowed(
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      0, cm->global_motion,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      mi);
+
+  if (last_motion_mode_allowed == SIMPLE_TRANSLATION) return;
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  if (last_motion_mode_allowed == OBMC_CAUSAL) {
+    aom_write(w, mbmi->motion_mode == OBMC_CAUSAL,
+              cm->fc->obmc_prob[mbmi->sb_type]);
+  } else {
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+    av1_write_token(w, av1_motion_mode_tree,
+                    cm->fc->motion_mode_prob[mbmi->sb_type],
+                    &motion_mode_encodings[mbmi->motion_mode]);
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  }
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+}
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_DELTA_Q
+static void write_delta_qindex(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                               int delta_qindex, aom_writer *w) {
+  int sign = delta_qindex < 0;
+  int abs = sign ? -delta_qindex : delta_qindex;
+  int rem_bits, thr;
+  int smallval = abs < DELTA_Q_SMALL ? 1 : 0;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+  (void)xd;
+#endif
+
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, AOMMIN(abs, DELTA_Q_SMALL), ec_ctx->delta_q_cdf,
+                   DELTA_Q_PROBS + 1);
+#else
+  int i = 0;
+  while (i < DELTA_Q_SMALL && i <= abs) {
+    int bit = (i < abs);
+    aom_write(w, bit, ec_ctx->delta_q_prob[i]);
+    i++;
+  }
+#endif
+
+  if (!smallval) {
+    rem_bits = OD_ILOG_NZ(abs - 1) - 1;
+    thr = (1 << rem_bits) + 1;
+    aom_write_literal(w, rem_bits, 3);
+    aom_write_literal(w, abs - thr, rem_bits);
+  }
+  if (abs > 0) {
+    aom_write_bit(w, sign);
+  }
+}
+
+#if !CONFIG_EC_ADAPT
+static void update_delta_q_probs(AV1_COMMON *cm, aom_writer *w,
+                                 FRAME_COUNTS *counts) {
+  int k;
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+#if CONFIG_EXT_DELTA_Q
+  if (!cm->delta_q_present_flag) return;
+#endif  // CONFIG_EXT_DELTA_Q
+  for (k = 0; k < DELTA_Q_PROBS; ++k) {
+    av1_cond_prob_diff_update(w, &cm->fc->delta_q_prob[k], counts->delta_q[k],
+                              probwt);
+  }
+}
+#endif  // CONFIG_EC_ADAPT
+
+#if CONFIG_EXT_DELTA_Q
+static void write_delta_lflevel(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                                int delta_lflevel, aom_writer *w) {
+  int sign = delta_lflevel < 0;
+  int abs = sign ? -delta_lflevel : delta_lflevel;
+  int rem_bits, thr;
+  int smallval = abs < DELTA_LF_SMALL ? 1 : 0;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+  (void)xd;
+#endif
+
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, AOMMIN(abs, DELTA_LF_SMALL), ec_ctx->delta_lf_cdf,
+                   DELTA_LF_PROBS + 1);
+#else
+  int i = 0;
+  while (i < DELTA_LF_SMALL && i <= abs) {
+    int bit = (i < abs);
+    aom_write(w, bit, ec_ctx->delta_lf_prob[i]);
+    i++;
+  }
+#endif  // CONFIG_EC_MULTISYMBOL
+
+  if (!smallval) {
+    rem_bits = OD_ILOG_NZ(abs - 1) - 1;
+    thr = (1 << rem_bits) + 1;
+    aom_write_literal(w, rem_bits, 3);
+    aom_write_literal(w, abs - thr, rem_bits);
+  }
+  if (abs > 0) {
+    aom_write_bit(w, sign);
+  }
+}
+
+#if !CONFIG_EC_ADAPT
+static void update_delta_lf_probs(AV1_COMMON *cm, aom_writer *w,
+                                  FRAME_COUNTS *counts) {
+  int k;
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+  if (!cm->delta_lf_present_flag) return;
+  for (k = 0; k < DELTA_LF_PROBS; ++k) {
+    av1_cond_prob_diff_update(w, &cm->fc->delta_lf_prob[k], counts->delta_lf[k],
+                              probwt);
+  }
+}
+#endif  // CONFIG_EC_ADAPT
+#endif  // CONFIG_EXT_DELTA_Q
+#endif  // CONFIG_DELTA_Q
+
+static void update_skip_probs(AV1_COMMON *cm, aom_writer *w,
+                              FRAME_COUNTS *counts) {
+  int k;
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+  for (k = 0; k < SKIP_CONTEXTS; ++k) {
+    av1_cond_prob_diff_update(w, &cm->fc->skip_probs[k], counts->skip[k],
+                              probwt);
+  }
+}
+
+#if !CONFIG_EC_ADAPT
+static void update_switchable_interp_probs(AV1_COMMON *cm, aom_writer *w,
+                                           FRAME_COUNTS *counts) {
+  int j;
+  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j) {
+#if CONFIG_TILE_GROUPS
+    const int probwt = cm->num_tg;
+#else
+    const int probwt = 1;
+#endif
+    prob_diff_update(
+        av1_switchable_interp_tree, cm->fc->switchable_interp_prob[j],
+        counts->switchable_interp[j], SWITCHABLE_FILTERS, probwt, w);
+  }
+}
+#endif
+
+#if !CONFIG_EC_ADAPT
+#if CONFIG_EXT_TX
+static void update_ext_tx_probs(AV1_COMMON *cm, aom_writer *w) {
+  const int savings_thresh = av1_cost_one(GROUP_DIFF_UPDATE_PROB) -
+                             av1_cost_zero(GROUP_DIFF_UPDATE_PROB);
+  int i, j;
+  int s;
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+  for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
+    int savings = 0;
+    int do_update = 0;
+    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+      if (!use_inter_ext_tx_for_txsize[s][i]) continue;
+      savings += prob_diff_update_savings(
+          av1_ext_tx_inter_tree[s], cm->fc->inter_ext_tx_prob[s][i],
+          cm->counts.inter_ext_tx[s][i],
+          num_ext_tx_set[ext_tx_set_type_inter[s]], probwt);
+    }
+    do_update = savings > savings_thresh;
+    aom_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
+    if (do_update) {
+      for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+        if (!use_inter_ext_tx_for_txsize[s][i]) continue;
+        prob_diff_update(av1_ext_tx_inter_tree[s],
+                         cm->fc->inter_ext_tx_prob[s][i],
+                         cm->counts.inter_ext_tx[s][i],
+                         num_ext_tx_set[ext_tx_set_type_inter[s]], probwt, w);
+      }
+    }
+  }
+
+  for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
+    int savings = 0;
+    int do_update = 0;
+    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+      if (!use_intra_ext_tx_for_txsize[s][i]) continue;
+      for (j = 0; j < INTRA_MODES; ++j)
+        savings += prob_diff_update_savings(
+            av1_ext_tx_intra_tree[s], cm->fc->intra_ext_tx_prob[s][i][j],
+            cm->counts.intra_ext_tx[s][i][j],
+            num_ext_tx_set[ext_tx_set_type_intra[s]], probwt);
+    }
+    do_update = savings > savings_thresh;
+    aom_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
+    if (do_update) {
+      for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+        if (!use_intra_ext_tx_for_txsize[s][i]) continue;
+        for (j = 0; j < INTRA_MODES; ++j)
+          prob_diff_update(av1_ext_tx_intra_tree[s],
+                           cm->fc->intra_ext_tx_prob[s][i][j],
+                           cm->counts.intra_ext_tx[s][i][j],
+                           num_ext_tx_set[ext_tx_set_type_intra[s]], probwt, w);
+      }
+    }
+  }
+}
+
+#else
+static void update_ext_tx_probs(AV1_COMMON *cm, aom_writer *w) {
+  const int savings_thresh = av1_cost_one(GROUP_DIFF_UPDATE_PROB) -
+                             av1_cost_zero(GROUP_DIFF_UPDATE_PROB);
+  int i, j;
+
+  int savings = 0;
+  int do_update = 0;
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+    for (j = 0; j < TX_TYPES; ++j)
+      savings += prob_diff_update_savings(
+          av1_ext_tx_tree, cm->fc->intra_ext_tx_prob[i][j],
+          cm->counts.intra_ext_tx[i][j], TX_TYPES, probwt);
+  }
+  do_update = savings > savings_thresh;
+  aom_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
+  if (do_update) {
+    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+      for (j = 0; j < TX_TYPES; ++j) {
+        prob_diff_update(av1_ext_tx_tree, cm->fc->intra_ext_tx_prob[i][j],
+                         cm->counts.intra_ext_tx[i][j], TX_TYPES, probwt, w);
+      }
+    }
+  }
+
+  savings = 0;
+  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+    savings +=
+        prob_diff_update_savings(av1_ext_tx_tree, cm->fc->inter_ext_tx_prob[i],
+                                 cm->counts.inter_ext_tx[i], TX_TYPES, probwt);
+  }
+  do_update = savings > savings_thresh;
+  aom_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
+  if (do_update) {
+    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+      prob_diff_update(av1_ext_tx_tree, cm->fc->inter_ext_tx_prob[i],
+                       cm->counts.inter_ext_tx[i], TX_TYPES, probwt, w);
+    }
+  }
+}
+#endif  // CONFIG_EXT_TX
+#endif  // !CONFIG_EC_ADAPT
+#if CONFIG_PALETTE
+static void pack_palette_tokens(aom_writer *w, const TOKENEXTRA **tp, int n,
+                                int num) {
+  int i;
+  const TOKENEXTRA *p = *tp;
+
+  for (i = 0; i < num; ++i) {
+    av1_write_token(
+        w, av1_palette_color_index_tree[n - PALETTE_MIN_SIZE], p->context_tree,
+        &palette_color_index_encodings[n - PALETTE_MIN_SIZE][p->token]);
+    ++p;
+  }
+
+  *tp = p;
+}
+#endif  // CONFIG_PALETTE
+
+#if !CONFIG_PVQ
+#if CONFIG_SUPERTX
+static void update_supertx_probs(AV1_COMMON *cm, int probwt, aom_writer *w) {
+  const int savings_thresh = av1_cost_one(GROUP_DIFF_UPDATE_PROB) -
+                             av1_cost_zero(GROUP_DIFF_UPDATE_PROB);
+  int i, j;
+  int savings = 0;
+  int do_update = 0;
+  for (i = 0; i < PARTITION_SUPERTX_CONTEXTS; ++i) {
+    for (j = TX_8X8; j < TX_SIZES; ++j) {
+      savings += av1_cond_prob_diff_update_savings(
+          &cm->fc->supertx_prob[i][j], cm->counts.supertx[i][j], probwt);
+    }
+  }
+  do_update = savings > savings_thresh;
+  aom_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
+  if (do_update) {
+    for (i = 0; i < PARTITION_SUPERTX_CONTEXTS; ++i) {
+      for (j = TX_8X8; j < TX_SIZES; ++j) {
+        av1_cond_prob_diff_update(w, &cm->fc->supertx_prob[i][j],
+                                  cm->counts.supertx[i][j], probwt);
+      }
+    }
+  }
+}
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_NEW_MULTISYMBOL
+static INLINE void write_coeff_extra(const aom_cdf_prob *const *cdf, int val,
+                                     int n, aom_writer *w) {
+  // Code the extra bits from LSB to MSB in groups of 4
+  int i = 0;
+  int count = 0;
+  while (count < n) {
+    const int size = AOMMIN(n - count, 4);
+    const int mask = (1 << size) - 1;
+    aom_write_cdf(w, val & mask, cdf[i++], 1 << size);
+    val >>= size;
+    count += size;
+  }
+}
+#else
+static INLINE void write_coeff_extra(const aom_prob *pb, int value,
+                                     int num_bits, int skip_bits, aom_writer *w,
+                                     TOKEN_STATS *token_stats) {
+  // Code the extra bits from MSB to LSB 1 bit at a time
+  int index;
+  for (index = skip_bits; index < num_bits; ++index) {
+    const int shift = num_bits - index - 1;
+    const int bb = (value >> shift) & 1;
+    aom_write_record(w, bb, pb[index], token_stats);
+  }
+}
+#endif
+
+#if CONFIG_NEW_TOKENSET && !CONFIG_LV_MAP
+static void pack_mb_tokens(aom_writer *w, const TOKENEXTRA **tp,
+                           const TOKENEXTRA *const stop,
+                           aom_bit_depth_t bit_depth, const TX_SIZE tx_size,
+                           TOKEN_STATS *token_stats) {
+  const TOKENEXTRA *p = *tp;
+#if CONFIG_VAR_TX
+  int count = 0;
+  const int seg_eob = tx_size_2d[tx_size];
+#endif
+
+  while (p < stop && p->token != EOSB_TOKEN) {
+    const int token = p->token;
+    if (token == BLOCK_Z_TOKEN) {
+      aom_write_symbol(w, 0, *p->head_cdf, HEAD_TOKENS + 1);
+      p++;
+      continue;
+    }
+
+    const av1_extra_bit *const extra_bits = &av1_extra_bits[token];
+    if (p->eob_val == LAST_EOB) {
+      // Just code a flag indicating whether the value is >1 or 1.
+      aom_write_bit(w, token != ONE_TOKEN);
+    } else {
+      int comb_symb = 2 * AOMMIN(token, TWO_TOKEN) - p->eob_val + p->first_val;
+      aom_write_symbol(w, comb_symb, *p->head_cdf, HEAD_TOKENS + p->first_val);
+    }
+    if (token > ONE_TOKEN) {
+      aom_write_symbol(w, token - TWO_TOKEN, *p->tail_cdf, TAIL_TOKENS);
+    }
+
+    if (extra_bits->base_val) {
+      const int bit_string = p->extra;
+      const int bit_string_length = extra_bits->len;  // Length of extra bits to
+      const int is_cat6 = (extra_bits->base_val == CAT6_MIN_VAL);
+      // be written excluding
+      // the sign bit.
+      int skip_bits = is_cat6
+                          ? (int)sizeof(av1_cat6_prob) -
+                                av1_get_cat6_extrabits_size(tx_size, bit_depth)
+                          : 0;
+
+      assert(!(bit_string >> (bit_string_length - skip_bits + 1)));
+      if (bit_string_length > 0)
+#if CONFIG_NEW_MULTISYMBOL
+        write_coeff_extra(extra_bits->cdf, bit_string >> 1,
+                          bit_string_length - skip_bits, w);
+#else
+        write_coeff_extra(extra_bits->prob, bit_string >> 1, bit_string_length,
+                          skip_bits, w, token_stats);
+#endif
+
+      aom_write_bit_record(w, bit_string & 1, token_stats);
+    }
+    ++p;
+
+#if CONFIG_VAR_TX
+    ++count;
+    if (token == EOB_TOKEN || count == seg_eob) break;
+#endif
+  }
+
+  *tp = p;
+}
+#else  //  CONFIG_NEW_TOKENSET
+#if !CONFIG_LV_MAP
+static void pack_mb_tokens(aom_writer *w, const TOKENEXTRA **tp,
+                           const TOKENEXTRA *const stop,
+                           aom_bit_depth_t bit_depth, const TX_SIZE tx_size,
+                           TOKEN_STATS *token_stats) {
+  const TOKENEXTRA *p = *tp;
+#if CONFIG_VAR_TX
+  int count = 0;
+  const int seg_eob = tx_size_2d[tx_size];
+#endif
+
+  while (p < stop && p->token != EOSB_TOKEN) {
+    const int token = p->token;
+#if !CONFIG_EC_MULTISYMBOL
+    const struct av1_token *const coef_encoding = &av1_coef_encodings[token];
+    int coef_value = coef_encoding->value;
+    int coef_length = coef_encoding->len;
+#endif  // !CONFIG_EC_MULTISYMBOL
+    const av1_extra_bit *const extra_bits = &av1_extra_bits[token];
+
+#if CONFIG_EC_MULTISYMBOL
+    /* skip one or two nodes */
+    if (!p->skip_eob_node)
+      aom_write_record(w, token != EOB_TOKEN, p->context_tree[0], token_stats);
+    if (token != EOB_TOKEN) {
+      aom_write_record(w, token != ZERO_TOKEN, p->context_tree[1], token_stats);
+      if (token != ZERO_TOKEN) {
+        aom_write_symbol(w, token - ONE_TOKEN, *p->token_cdf,
+                         CATEGORY6_TOKEN - ONE_TOKEN + 1);
+      }
+    }
+#else
+    /* skip one or two nodes */
+    if (p->skip_eob_node)
+      coef_length -= p->skip_eob_node;
+    else
+      aom_write_record(w, token != EOB_TOKEN, p->context_tree[0], token_stats);
+
+    if (token != EOB_TOKEN) {
+      aom_write_record(w, token != ZERO_TOKEN, p->context_tree[1], token_stats);
+
+      if (token != ZERO_TOKEN) {
+        aom_write_record(w, token != ONE_TOKEN, p->context_tree[2],
+                         token_stats);
+
+        if (token != ONE_TOKEN) {
+          const int unconstrained_len = UNCONSTRAINED_NODES - p->skip_eob_node;
+          aom_write_tree_record(
+              w, av1_coef_con_tree,
+              av1_pareto8_full[p->context_tree[PIVOT_NODE] - 1], coef_value,
+              coef_length - unconstrained_len, 0, token_stats);
+        }
+      }
+    }
+#endif  // CONFIG_EC_MULTISYMBOL
+
+    if (extra_bits->base_val) {
+      const int bit_string = p->extra;
+      const int bit_string_length = extra_bits->len;  // Length of extra bits to
+      // be written excluding
+      // the sign bit.
+      int skip_bits = (extra_bits->base_val == CAT6_MIN_VAL)
+                          ? (int)sizeof(av1_cat6_prob) -
+                                av1_get_cat6_extrabits_size(tx_size, bit_depth)
+                          : 0;
+
+      assert(!(bit_string >> (bit_string_length - skip_bits + 1)));
+      if (bit_string_length > 0) {
+#if CONFIG_NEW_MULTISYMBOL
+        skip_bits &= ~3;
+        write_coeff_extra(extra_bits->cdf, bit_string >> 1,
+                          bit_string_length - skip_bits, w);
+#else
+        write_coeff_extra(extra_bits->prob, bit_string >> 1, bit_string_length,
+                          skip_bits, w, token_stats);
+#endif
+      }
+      aom_write_bit_record(w, bit_string & 1, token_stats);
+    }
+    ++p;
+
+#if CONFIG_VAR_TX
+    ++count;
+    if (token == EOB_TOKEN || count == seg_eob) break;
+#endif
+  }
+
+  *tp = p;
+}
+#endif  // !CONFIG_LV_MAP
+#endif  // CONFIG_NEW_TOKENSET
+#else   // !CONFIG_PVQ
+static PVQ_INFO *get_pvq_block(PVQ_QUEUE *pvq_q) {
+  PVQ_INFO *pvq;
+
+  assert(pvq_q->curr_pos <= pvq_q->last_pos);
+  assert(pvq_q->curr_pos < pvq_q->buf_len);
+
+  pvq = pvq_q->buf + pvq_q->curr_pos;
+  ++pvq_q->curr_pos;
+
+  return pvq;
+}
+
+static void pack_pvq_tokens(aom_writer *w, MACROBLOCK *const x,
+                            MACROBLOCKD *const xd, int plane, BLOCK_SIZE bsize,
+                            const TX_SIZE tx_size) {
+  PVQ_INFO *pvq;
+  int idx, idy;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  od_adapt_ctx *adapt;
+  int max_blocks_wide;
+  int max_blocks_high;
+  int step = (1 << tx_size);
+  const BLOCK_SIZE plane_bsize =
+      get_plane_block_size(AOMMAX(bsize, BLOCK_8X8), pd);
+
+  adapt = x->daala_enc.state.adapt;
+
+  max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+  max_blocks_high = max_block_high(xd, plane_bsize, plane);
+
+  for (idy = 0; idy < max_blocks_high; idy += step) {
+    for (idx = 0; idx < max_blocks_wide; idx += step) {
+      const int is_keyframe = 0;
+      const int encode_flip = 0;
+      const int flip = 0;
+      int i;
+      const int has_dc_skip = 1;
+      int *exg = &adapt->pvq.pvq_exg[plane][tx_size][0];
+      int *ext = adapt->pvq.pvq_ext + tx_size * PVQ_MAX_PARTITIONS;
+      generic_encoder *model = adapt->pvq.pvq_param_model;
+
+      pvq = get_pvq_block(x->pvq_q);
+
+      // encode block skip info
+      aom_write_symbol(w, pvq->ac_dc_coded,
+                       adapt->skip_cdf[2 * tx_size + (plane != 0)], 4);
+
+      // AC coeffs coded?
+      if (pvq->ac_dc_coded & AC_CODED) {
+        assert(pvq->bs == tx_size);
+        for (i = 0; i < pvq->nb_bands; i++) {
+          if (i == 0 ||
+              (!pvq->skip_rest && !(pvq->skip_dir & (1 << ((i - 1) % 3))))) {
+            pvq_encode_partition(
+                w, pvq->qg[i], pvq->theta[i], pvq->y + pvq->off[i],
+                pvq->size[i], pvq->k[i], model, adapt, exg + i, ext + i,
+                (plane != 0) * OD_TXSIZES * PVQ_MAX_PARTITIONS +
+                    pvq->bs * PVQ_MAX_PARTITIONS + i,
+                is_keyframe, i == 0 && (i < pvq->nb_bands - 1), pvq->skip_rest,
+                encode_flip, flip);
+          }
+          if (i == 0 && !pvq->skip_rest && pvq->bs > 0) {
+            aom_write_symbol(
+                w, pvq->skip_dir,
+                &adapt->pvq
+                     .pvq_skip_dir_cdf[(plane != 0) + 2 * (pvq->bs - 1)][0],
+                7);
+          }
+        }
+      }
+      // Encode residue of DC coeff, if exist.
+      if (!has_dc_skip || (pvq->ac_dc_coded & DC_CODED)) {
+        generic_encode(w, &adapt->model_dc[plane],
+                       abs(pvq->dq_dc_residue) - has_dc_skip,
+                       &adapt->ex_dc[plane][pvq->bs][0], 2);
+      }
+      if ((pvq->ac_dc_coded & DC_CODED)) {
+        aom_write_bit(w, pvq->dq_dc_residue < 0);
+      }
+    }
+  }  // for (idy = 0;
+}
+#endif  // !CONFIG_PVG
+
+#if CONFIG_VAR_TX && !CONFIG_COEF_INTERLEAVE
+static void pack_txb_tokens(aom_writer *w, const TOKENEXTRA **tp,
+                            const TOKENEXTRA *const tok_end,
+#if CONFIG_PVQ
+                            MACROBLOCK *const x,
+#endif
+                            MACROBLOCKD *xd, MB_MODE_INFO *mbmi, int plane,
+                            BLOCK_SIZE plane_bsize, aom_bit_depth_t bit_depth,
+                            int block, int blk_row, int blk_col,
+                            TX_SIZE tx_size, TOKEN_STATS *token_stats) {
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
+  const int tx_row = blk_row >> (1 - pd->subsampling_y);
+  const int tx_col = blk_col >> (1 - pd->subsampling_x);
+  TX_SIZE plane_tx_size;
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  plane_tx_size =
+      plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
+            : mbmi->inter_tx_size[tx_row][tx_col];
+
+  if (tx_size == plane_tx_size) {
+    TOKEN_STATS tmp_token_stats;
+    init_token_stats(&tmp_token_stats);
+#if !CONFIG_PVQ
+    pack_mb_tokens(w, tp, tok_end, bit_depth, tx_size, &tmp_token_stats);
+#else
+    pack_pvq_tokens(w, x, xd, plane, bsize, tx_size);
+#endif
+#if CONFIG_RD_DEBUG
+    token_stats->txb_coeff_cost_map[blk_row][blk_col] = tmp_token_stats.cost;
+    token_stats->cost += tmp_token_stats.cost;
+#endif
+  } else {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int i;
+
+    assert(bsl > 0);
+
+    for (i = 0; i < 4; ++i) {
+      const int offsetr = blk_row + (i >> 1) * bsl;
+      const int offsetc = blk_col + (i & 0x01) * bsl;
+      const int step = tx_size_wide_unit[sub_txs] * tx_size_high_unit[sub_txs];
+
+      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
+
+      pack_txb_tokens(w, tp, tok_end,
+#if CONFIG_PVQ
+                      x,
+#endif
+                      xd, mbmi, plane, plane_bsize, bit_depth, block, offsetr,
+                      offsetc, sub_txs, token_stats);
+      block += step;
+    }
+  }
+}
+#endif
+
+static void write_segment_id(aom_writer *w, const struct segmentation *seg,
+                             struct segmentation_probs *segp, int segment_id) {
+  if (seg->enabled && seg->update_map) {
+#if CONFIG_EC_MULTISYMBOL
+    aom_write_symbol(w, segment_id, segp->tree_cdf, MAX_SEGMENTS);
+#else
+    aom_write_tree(w, av1_segment_tree, segp->tree_probs, segment_id, 3, 0);
+#endif
+  }
+}
+
+// This function encodes the reference frame
+static void write_ref_frames(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                             aom_writer *w) {
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int is_compound = has_second_ref(mbmi);
+  const int segment_id = mbmi->segment_id;
+
+  // If segment level coding of this signal is disabled...
+  // or the segment allows multiple reference frame options
+  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
+    assert(!is_compound);
+    assert(mbmi->ref_frame[0] ==
+           get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME));
+  } else {
+    // does the feature use compound prediction or not
+    // (if not specified at the frame/segment level)
+    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+#if SUB8X8_COMP_REF
+      aom_write(w, is_compound, av1_get_reference_mode_prob(cm, xd));
+#else
+      if (mbmi->sb_type >= BLOCK_8X8)
+        aom_write(w, is_compound, av1_get_reference_mode_prob(cm, xd));
+#endif
+    } else {
+      assert((!is_compound) == (cm->reference_mode == SINGLE_REFERENCE));
+    }
+
+    if (is_compound) {
+#if CONFIG_EXT_REFS
+      const int bit = (mbmi->ref_frame[0] == GOLDEN_FRAME ||
+                       mbmi->ref_frame[0] == LAST3_FRAME);
+      const int bit_bwd = mbmi->ref_frame[1] == ALTREF_FRAME;
+#else  // CONFIG_EXT_REFS
+      const int bit = mbmi->ref_frame[0] == GOLDEN_FRAME;
+#endif  // CONFIG_EXT_REFS
+
+      aom_write(w, bit, av1_get_pred_prob_comp_ref_p(cm, xd));
+
+#if CONFIG_EXT_REFS
+      if (!bit) {
+        const int bit1 = mbmi->ref_frame[0] == LAST_FRAME;
+        aom_write(w, bit1, av1_get_pred_prob_comp_ref_p1(cm, xd));
+      } else {
+        const int bit2 = mbmi->ref_frame[0] == GOLDEN_FRAME;
+        aom_write(w, bit2, av1_get_pred_prob_comp_ref_p2(cm, xd));
+      }
+      aom_write(w, bit_bwd, av1_get_pred_prob_comp_bwdref_p(cm, xd));
+#endif  // CONFIG_EXT_REFS
+    } else {
+#if CONFIG_EXT_REFS
+      const int bit0 = (mbmi->ref_frame[0] == ALTREF_FRAME ||
+                        mbmi->ref_frame[0] == BWDREF_FRAME);
+      aom_write(w, bit0, av1_get_pred_prob_single_ref_p1(cm, xd));
+
+      if (bit0) {
+        const int bit1 = mbmi->ref_frame[0] == ALTREF_FRAME;
+        aom_write(w, bit1, av1_get_pred_prob_single_ref_p2(cm, xd));
+      } else {
+        const int bit2 = (mbmi->ref_frame[0] == LAST3_FRAME ||
+                          mbmi->ref_frame[0] == GOLDEN_FRAME);
+        aom_write(w, bit2, av1_get_pred_prob_single_ref_p3(cm, xd));
+
+        if (!bit2) {
+          const int bit3 = mbmi->ref_frame[0] != LAST_FRAME;
+          aom_write(w, bit3, av1_get_pred_prob_single_ref_p4(cm, xd));
+        } else {
+          const int bit4 = mbmi->ref_frame[0] != LAST3_FRAME;
+          aom_write(w, bit4, av1_get_pred_prob_single_ref_p5(cm, xd));
+        }
+      }
+#else   // CONFIG_EXT_REFS
+      const int bit0 = mbmi->ref_frame[0] != LAST_FRAME;
+      aom_write(w, bit0, av1_get_pred_prob_single_ref_p1(cm, xd));
+
+      if (bit0) {
+        const int bit1 = mbmi->ref_frame[0] != GOLDEN_FRAME;
+        aom_write(w, bit1, av1_get_pred_prob_single_ref_p2(cm, xd));
+      }
+#endif  // CONFIG_EXT_REFS
+    }
+  }
+}
+
+#if CONFIG_FILTER_INTRA
+static void write_filter_intra_mode_info(const AV1_COMMON *const cm,
+                                         const MB_MODE_INFO *const mbmi,
+                                         aom_writer *w) {
+  if (mbmi->mode == DC_PRED
+#if CONFIG_PALETTE
+      && mbmi->palette_mode_info.palette_size[0] == 0
+#endif  // CONFIG_PALETTE
+      ) {
+    aom_write(w, mbmi->filter_intra_mode_info.use_filter_intra_mode[0],
+              cm->fc->filter_intra_probs[0]);
+    if (mbmi->filter_intra_mode_info.use_filter_intra_mode[0]) {
+      const FILTER_INTRA_MODE mode =
+          mbmi->filter_intra_mode_info.filter_intra_mode[0];
+      write_uniform(w, FILTER_INTRA_MODES, mode);
+    }
+  }
+
+  if (mbmi->uv_mode == DC_PRED
+#if CONFIG_PALETTE
+      && mbmi->palette_mode_info.palette_size[1] == 0
+#endif  // CONFIG_PALETTE
+      ) {
+    aom_write(w, mbmi->filter_intra_mode_info.use_filter_intra_mode[1],
+              cm->fc->filter_intra_probs[1]);
+    if (mbmi->filter_intra_mode_info.use_filter_intra_mode[1]) {
+      const FILTER_INTRA_MODE mode =
+          mbmi->filter_intra_mode_info.filter_intra_mode[1];
+      write_uniform(w, FILTER_INTRA_MODES, mode);
+    }
+  }
+}
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_EXT_INTRA
+static void write_intra_angle_info(const MACROBLOCKD *xd,
+                                   FRAME_CONTEXT *const ec_ctx, aom_writer *w) {
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+#if CONFIG_INTRA_INTERP
+  const int intra_filter_ctx = av1_get_pred_context_intra_interp(xd);
+  int p_angle;
+#endif  // CONFIG_INTRA_INTERP
+
+  (void)ec_ctx;
+  if (bsize < BLOCK_8X8) return;
+
+  if (av1_is_directional_mode(mbmi->mode, bsize)) {
+    write_uniform(w, 2 * MAX_ANGLE_DELTA + 1,
+                  MAX_ANGLE_DELTA + mbmi->angle_delta[0]);
+#if CONFIG_INTRA_INTERP
+    p_angle = mode_to_angle_map[mbmi->mode] + mbmi->angle_delta[0] * ANGLE_STEP;
+    if (av1_is_intra_filter_switchable(p_angle)) {
+#if CONFIG_EC_MULTISYMBOL
+      aom_write_symbol(w, mbmi->intra_filter,
+                       ec_ctx->intra_filter_cdf[intra_filter_ctx],
+                       INTRA_FILTERS);
+#else
+      av1_write_token(w, av1_intra_filter_tree,
+                      ec_ctx->intra_filter_probs[intra_filter_ctx],
+                      &intra_filter_encodings[mbmi->intra_filter]);
+#endif  // CONFIG_EC_MULTISYMBOL
+    }
+#endif  // CONFIG_INTRA_INTERP
+  }
+
+  if (av1_is_directional_mode(mbmi->uv_mode, bsize)) {
+    write_uniform(w, 2 * MAX_ANGLE_DELTA + 1,
+                  MAX_ANGLE_DELTA + mbmi->angle_delta[1]);
+  }
+}
+#endif  // CONFIG_EXT_INTRA
+
+static void write_mb_interp_filter(AV1_COMP *cpi, const MACROBLOCKD *xd,
+                                   aom_writer *w) {
+  AV1_COMMON *const cm = &cpi->common;
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  if (!av1_is_interp_needed(xd)) {
+#if CONFIG_DUAL_FILTER
+    for (int i = 0; i < 4; ++i)
+      assert(mbmi->interp_filter[i] == (cm->interp_filter == SWITCHABLE
+                                            ? EIGHTTAP_REGULAR
+                                            : cm->interp_filter));
+#else
+    assert(mbmi->interp_filter == (cm->interp_filter == SWITCHABLE
+                                       ? EIGHTTAP_REGULAR
+                                       : cm->interp_filter));
+#endif  // CONFIG_DUAL_FILTER
+    return;
+  }
+  if (cm->interp_filter == SWITCHABLE) {
+#if CONFIG_DUAL_FILTER
+    int dir;
+    for (dir = 0; dir < 2; ++dir) {
+      if (has_subpel_mv_component(xd->mi[0], xd, dir) ||
+          (mbmi->ref_frame[1] > INTRA_FRAME &&
+           has_subpel_mv_component(xd->mi[0], xd, dir + 2))) {
+        const int ctx = av1_get_pred_context_switchable_interp(xd, dir);
+#if CONFIG_EC_MULTISYMBOL
+        aom_write_symbol(w, av1_switchable_interp_ind[mbmi->interp_filter[dir]],
+                         ec_ctx->switchable_interp_cdf[ctx],
+                         SWITCHABLE_FILTERS);
+#else
+        av1_write_token(w, av1_switchable_interp_tree,
+                        ec_ctx->switchable_interp_prob[ctx],
+                        &switchable_interp_encodings[mbmi->interp_filter[dir]]);
+#endif
+        ++cpi->interp_filter_selected[0][mbmi->interp_filter[dir]];
+      } else {
+        assert(mbmi->interp_filter[dir] == EIGHTTAP_REGULAR);
+      }
+    }
+#else
+    {
+      const int ctx = av1_get_pred_context_switchable_interp(xd);
+#if CONFIG_EC_MULTISYMBOL
+      aom_write_symbol(w, av1_switchable_interp_ind[mbmi->interp_filter],
+                       ec_ctx->switchable_interp_cdf[ctx], SWITCHABLE_FILTERS);
+#else
+      av1_write_token(w, av1_switchable_interp_tree,
+                      ec_ctx->switchable_interp_prob[ctx],
+                      &switchable_interp_encodings[mbmi->interp_filter]);
+#endif
+      ++cpi->interp_filter_selected[0][mbmi->interp_filter];
+    }
+#endif  // CONFIG_DUAL_FILTER
+  }
+}
+
+#if CONFIG_PALETTE
+#if CONFIG_PALETTE_DELTA_ENCODING
+// Write luma palette color values with delta encoding. Write the first value as
+// literal, and the deltas between each value and the previous one. The luma
+// palette is sorted so each delta is larger than 0.
+static void write_palette_colors_y(const PALETTE_MODE_INFO *const pmi,
+                                   int bit_depth, aom_writer *w) {
+  const int n = pmi->palette_size[0];
+  int min_bits, i;
+  int bits = av1_get_palette_delta_bits_y(pmi, bit_depth, &min_bits);
+  aom_write_literal(w, bits - min_bits, 2);
+  aom_write_literal(w, pmi->palette_colors[0], bit_depth);
+  for (i = 1; i < n; ++i) {
+    aom_write_literal(
+        w, pmi->palette_colors[i] - pmi->palette_colors[i - 1] - 1, bits);
+    bits =
+        AOMMIN(bits, av1_ceil_log2((1 << bit_depth) - pmi->palette_colors[i]));
+  }
+}
+
+// Write chroma palette color values. Use delta encoding for u channel as its
+// palette is sorted. For v channel, either use delta encoding or transmit
+// raw values directly, whichever costs less.
+static void write_palette_colors_uv(const PALETTE_MODE_INFO *const pmi,
+                                    int bit_depth, aom_writer *w) {
+  int i;
+  const int n = pmi->palette_size[1];
+#if CONFIG_HIGHBITDEPTH
+  const uint16_t *colors_u = pmi->palette_colors + PALETTE_MAX_SIZE;
+  const uint16_t *colors_v = pmi->palette_colors + 2 * PALETTE_MAX_SIZE;
+#else
+  const uint8_t *colors_u = pmi->palette_colors + PALETTE_MAX_SIZE;
+  const uint8_t *colors_v = pmi->palette_colors + 2 * PALETTE_MAX_SIZE;
+#endif  // CONFIG_HIGHBITDEPTH
+  // U channel colors.
+  int min_bits_u = 0;
+  int bits_u = av1_get_palette_delta_bits_u(pmi, bit_depth, &min_bits_u);
+  aom_write_literal(w, bits_u - min_bits_u, 2);
+  aom_write_literal(w, colors_u[0], bit_depth);
+  for (i = 1; i < n; ++i) {
+    aom_write_literal(w, colors_u[i] - colors_u[i - 1], bits_u);
+    bits_u = AOMMIN(bits_u, av1_ceil_log2(1 + (1 << bit_depth) - colors_u[i]));
+  }
+  // V channel colors.
+  const int max_val = 1 << bit_depth;
+  int zero_count = 0, min_bits_v = 0;
+  int bits_v =
+      av1_get_palette_delta_bits_v(pmi, bit_depth, &zero_count, &min_bits_v);
+  const int rate_using_delta =
+      2 + bit_depth + (bits_v + 1) * (n - 1) - zero_count;
+  const int rate_using_raw = bit_depth * n;
+  if (rate_using_delta < rate_using_raw) {  // delta encoding
+    aom_write_bit(w, 1);
+    aom_write_literal(w, bits_v - min_bits_v, 2);
+    aom_write_literal(w, colors_v[0], bit_depth);
+    for (i = 1; i < n; ++i) {
+      if (colors_v[i] == colors_v[i - 1]) {  // No need to signal sign bit.
+        aom_write_literal(w, 0, bits_v);
+        continue;
+      }
+      const int delta = abs((int)colors_v[i] - colors_v[i - 1]);
+      const int sign_bit = colors_v[i] < colors_v[i - 1];
+      if (delta <= max_val - delta) {
+        aom_write_literal(w, delta, bits_v);
+        aom_write_bit(w, sign_bit);
+      } else {
+        aom_write_literal(w, max_val - delta, bits_v);
+        aom_write_bit(w, !sign_bit);
+      }
+    }
+  } else {  // Transmit raw values.
+    aom_write_bit(w, 0);
+    for (i = 0; i < n; ++i) aom_write_literal(w, colors_v[i], bit_depth);
+  }
+}
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+
+static void write_palette_mode_info(const AV1_COMMON *cm, const MACROBLOCKD *xd,
+                                    const MODE_INFO *const mi, aom_writer *w) {
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+
+  if (mbmi->mode == DC_PRED) {
+    const int n = pmi->palette_size[0];
+    int palette_y_mode_ctx = 0;
+    if (above_mi)
+      palette_y_mode_ctx +=
+          (above_mi->mbmi.palette_mode_info.palette_size[0] > 0);
+    if (left_mi)
+      palette_y_mode_ctx +=
+          (left_mi->mbmi.palette_mode_info.palette_size[0] > 0);
+    aom_write(
+        w, n > 0,
+        av1_default_palette_y_mode_prob[bsize - BLOCK_8X8][palette_y_mode_ctx]);
+    if (n > 0) {
+      av1_write_token(w, av1_palette_size_tree,
+                      av1_default_palette_y_size_prob[bsize - BLOCK_8X8],
+                      &palette_size_encodings[n - PALETTE_MIN_SIZE]);
+#if CONFIG_PALETTE_DELTA_ENCODING
+      write_palette_colors_y(pmi, cm->bit_depth, w);
+#else
+      int i;
+      for (i = 0; i < n; ++i)
+        aom_write_literal(w, pmi->palette_colors[i], cm->bit_depth);
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+      write_uniform(w, n, pmi->palette_first_color_idx[0]);
+    }
+  }
+
+  if (mbmi->uv_mode == DC_PRED) {
+    const int n = pmi->palette_size[1];
+    const int palette_uv_mode_ctx = (pmi->palette_size[0] > 0);
+    aom_write(w, n > 0, av1_default_palette_uv_mode_prob[palette_uv_mode_ctx]);
+    if (n > 0) {
+      av1_write_token(w, av1_palette_size_tree,
+                      av1_default_palette_uv_size_prob[bsize - BLOCK_8X8],
+                      &palette_size_encodings[n - PALETTE_MIN_SIZE]);
+#if CONFIG_PALETTE_DELTA_ENCODING
+      write_palette_colors_uv(pmi, cm->bit_depth, w);
+#else
+      int i;
+      for (i = 0; i < n; ++i) {
+        aom_write_literal(w, pmi->palette_colors[PALETTE_MAX_SIZE + i],
+                          cm->bit_depth);
+        aom_write_literal(w, pmi->palette_colors[2 * PALETTE_MAX_SIZE + i],
+                          cm->bit_depth);
+      }
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+      write_uniform(w, n, pmi->palette_first_color_idx[1]);
+    }
+  }
+}
+#endif  // CONFIG_PALETTE
+
+void av1_write_tx_type(const AV1_COMMON *const cm, const MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                       const int supertx_enabled,
+#endif
+#if CONFIG_TXK_SEL
+                       int block, int plane,
+#endif
+                       aom_writer *w) {
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const int is_inter = is_inter_block(mbmi);
+#if CONFIG_VAR_TX
+  const TX_SIZE tx_size = is_inter ? mbmi->min_tx_size : mbmi->tx_size;
+#else
+  const TX_SIZE tx_size = mbmi->tx_size;
+#endif  // CONFIG_VAR_TX
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+#if !CONFIG_TXK_SEL
+  TX_TYPE tx_type = mbmi->tx_type;
+#else
+  // Only y plane's tx_type is transmitted
+  if (plane > 0) return;
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+#endif
+
+  if (!FIXED_TX_TYPE) {
+#if CONFIG_EXT_TX
+    const TX_SIZE square_tx_size = txsize_sqr_map[tx_size];
+    const BLOCK_SIZE bsize = mbmi->sb_type;
+    if (get_ext_tx_types(tx_size, bsize, is_inter, cm->reduced_tx_set_used) >
+            1 &&
+        ((!cm->seg.enabled && cm->base_qindex > 0) ||
+         (cm->seg.enabled && xd->qindex[mbmi->segment_id] > 0)) &&
+        !mbmi->skip &&
+#if CONFIG_SUPERTX
+        !supertx_enabled &&
+#endif  // CONFIG_SUPERTX
+        !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+      const int eset =
+          get_ext_tx_set(tx_size, bsize, is_inter, cm->reduced_tx_set_used);
+      if (is_inter) {
+        assert(ext_tx_used_inter[eset][tx_type]);
+        if (eset > 0) {
+#if CONFIG_EC_MULTISYMBOL
+          aom_write_symbol(w, av1_ext_tx_inter_ind[eset][tx_type],
+                           ec_ctx->inter_ext_tx_cdf[eset][square_tx_size],
+                           ext_tx_cnt_inter[eset]);
+#else
+          av1_write_token(w, av1_ext_tx_inter_tree[eset],
+                          ec_ctx->inter_ext_tx_prob[eset][square_tx_size],
+                          &ext_tx_inter_encodings[eset][tx_type]);
+#endif
+        }
+      } else if (ALLOW_INTRA_EXT_TX) {
+        assert(ext_tx_used_intra[eset][tx_type]);
+        if (eset > 0) {
+#if CONFIG_EC_MULTISYMBOL
+          aom_write_symbol(
+              w, av1_ext_tx_intra_ind[eset][tx_type],
+              ec_ctx->intra_ext_tx_cdf[eset][square_tx_size][mbmi->mode],
+              ext_tx_cnt_intra[eset]);
+#else
+          av1_write_token(
+              w, av1_ext_tx_intra_tree[eset],
+              ec_ctx->intra_ext_tx_prob[eset][square_tx_size][mbmi->mode],
+              &ext_tx_intra_encodings[eset][tx_type]);
+#endif
+        }
+      }
+    }
+#else
+    if (tx_size < TX_32X32 &&
+        ((!cm->seg.enabled && cm->base_qindex > 0) ||
+         (cm->seg.enabled && xd->qindex[mbmi->segment_id] > 0)) &&
+        !mbmi->skip &&
+#if CONFIG_SUPERTX
+        !supertx_enabled &&
+#endif  // CONFIG_SUPERTX
+        !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+      if (is_inter) {
+#if CONFIG_EC_MULTISYMBOL
+        aom_write_symbol(w, av1_ext_tx_ind[tx_type],
+                         ec_ctx->inter_ext_tx_cdf[tx_size], TX_TYPES);
+#else
+        av1_write_token(w, av1_ext_tx_tree, ec_ctx->inter_ext_tx_prob[tx_size],
+                        &ext_tx_encodings[tx_type]);
+#endif
+      } else {
+#if CONFIG_EC_MULTISYMBOL
+        aom_write_symbol(
+            w, av1_ext_tx_ind[tx_type],
+            ec_ctx->intra_ext_tx_cdf[tx_size]
+                                    [intra_mode_to_tx_type_context[mbmi->mode]],
+            TX_TYPES);
+#else
+        av1_write_token(
+            w, av1_ext_tx_tree,
+            ec_ctx
+                ->intra_ext_tx_prob[tx_size]
+                                   [intra_mode_to_tx_type_context[mbmi->mode]],
+            &ext_tx_encodings[tx_type]);
+#endif
+      }
+    }
+#endif  // CONFIG_EXT_TX
+  }
+}
+
+static void write_intra_mode(FRAME_CONTEXT *frame_ctx, BLOCK_SIZE bsize,
+                             PREDICTION_MODE mode, aom_writer *w) {
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, av1_intra_mode_ind[mode],
+                   frame_ctx->y_mode_cdf[size_group_lookup[bsize]],
+                   INTRA_MODES);
+#else
+  av1_write_token(w, av1_intra_mode_tree,
+                  frame_ctx->y_mode_prob[size_group_lookup[bsize]],
+                  &intra_mode_encodings[mode]);
+#endif
+}
+
+static void write_intra_uv_mode(FRAME_CONTEXT *frame_ctx,
+                                PREDICTION_MODE uv_mode, PREDICTION_MODE y_mode,
+                                aom_writer *w) {
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, av1_intra_mode_ind[uv_mode],
+                   frame_ctx->uv_mode_cdf[y_mode], INTRA_MODES);
+#else
+  av1_write_token(w, av1_intra_mode_tree, frame_ctx->uv_mode_prob[y_mode],
+                  &intra_mode_encodings[uv_mode]);
+#endif
+}
+
+static void pack_inter_mode_mvs(AV1_COMP *cpi, const int mi_row,
+                                const int mi_col,
+#if CONFIG_SUPERTX
+                                int supertx_enabled,
+#endif
+                                aom_writer *w) {
+  AV1_COMMON *const cm = &cpi->common;
+#if CONFIG_DELTA_Q || CONFIG_EC_ADAPT
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+#else
+  const MACROBLOCK *x = &cpi->td.mb;
+  const MACROBLOCKD *xd = &x->e_mbd;
+#endif
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+#if !CONFIG_REF_MV
+  nmv_context *nmvc = &ec_ctx->nmvc;
+#endif
+  const MODE_INFO *mi = xd->mi[0];
+
+  const struct segmentation *const seg = &cm->seg;
+  struct segmentation_probs *const segp = &cm->fc->seg;
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  const PREDICTION_MODE mode = mbmi->mode;
+  const int segment_id = mbmi->segment_id;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  const int allow_hp = cm->allow_high_precision_mv;
+  const int is_inter = is_inter_block(mbmi);
+  const int is_compound = has_second_ref(mbmi);
+  int skip, ref;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+  (void)mi_row;
+  (void)mi_col;
+
+  if (seg->update_map) {
+    if (seg->temporal_update) {
+      const int pred_flag = mbmi->seg_id_predicted;
+      aom_prob pred_prob = av1_get_pred_prob_seg_id(segp, xd);
+      aom_write(w, pred_flag, pred_prob);
+      if (!pred_flag) write_segment_id(w, seg, segp, segment_id);
+    } else {
+      write_segment_id(w, seg, segp, segment_id);
+    }
+  }
+
+#if CONFIG_SUPERTX
+  if (supertx_enabled)
+    skip = mbmi->skip;
+  else
+    skip = write_skip(cm, xd, segment_id, mi, w);
+#else
+  skip = write_skip(cm, xd, segment_id, mi, w);
+#endif  // CONFIG_SUPERTX
+#if CONFIG_DELTA_Q
+  if (cm->delta_q_present_flag) {
+    int super_block_upper_left =
+        ((mi_row & MAX_MIB_MASK) == 0) && ((mi_col & MAX_MIB_MASK) == 0);
+    if ((bsize != BLOCK_LARGEST || skip == 0) && super_block_upper_left) {
+      assert(mbmi->current_q_index > 0);
+      int reduced_delta_qindex =
+          (mbmi->current_q_index - xd->prev_qindex) / cm->delta_q_res;
+      write_delta_qindex(cm, xd, reduced_delta_qindex, w);
+      xd->prev_qindex = mbmi->current_q_index;
+#if CONFIG_EXT_DELTA_Q
+      if (cm->delta_lf_present_flag) {
+        int reduced_delta_lflevel =
+            (mbmi->current_delta_lf_from_base - xd->prev_delta_lf_from_base) /
+            cm->delta_lf_res;
+        write_delta_lflevel(cm, xd, reduced_delta_lflevel, w);
+        xd->prev_delta_lf_from_base = mbmi->current_delta_lf_from_base;
+      }
+#endif  // CONFIG_EXT_DELTA_Q
+    }
+  }
+#endif
+
+#if CONFIG_SUPERTX
+  if (!supertx_enabled)
+#endif  // CONFIG_SUPERTX
+    if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
+      aom_write(w, is_inter, av1_get_intra_inter_prob(cm, xd));
+
+  if (cm->tx_mode == TX_MODE_SELECT &&
+#if CONFIG_CB4X4 && (CONFIG_VAR_TX || CONFIG_RECT_TX)
+#if CONFIG_RECT_TX
+      bsize > BLOCK_4X4 &&
+#else
+      (bsize >= BLOCK_8X8 || (bsize > BLOCK_4X4 && is_inter)) &&
+#endif  // CONFIG_RECT_TX
+#else
+      bsize >= BLOCK_8X8 &&
+#endif
+#if CONFIG_SUPERTX
+      !supertx_enabled &&
+#endif  // CONFIG_SUPERTX
+      !(is_inter && skip) && !xd->lossless[segment_id]) {
+#if CONFIG_VAR_TX
+    if (is_inter) {  // This implies skip flag is 0.
+      const TX_SIZE max_tx_size = get_vartx_max_txsize(mbmi, bsize);
+      const int bh = tx_size_high_unit[max_tx_size];
+      const int bw = tx_size_wide_unit[max_tx_size];
+      const int width = block_size_wide[bsize] >> tx_size_wide_log2[0];
+      const int height = block_size_high[bsize] >> tx_size_wide_log2[0];
+      int idx, idy;
+      for (idy = 0; idy < height; idy += bh)
+        for (idx = 0; idx < width; idx += bw)
+          write_tx_size_vartx(cm, xd, mbmi, max_tx_size, height != width, idy,
+                              idx, w);
+    } else {
+      set_txfm_ctxs(mbmi->tx_size, xd->n8_w, xd->n8_h, skip, xd);
+      write_selected_tx_size(cm, xd, w);
+    }
+  } else {
+    set_txfm_ctxs(mbmi->tx_size, xd->n8_w, xd->n8_h, skip, xd);
+#else
+    write_selected_tx_size(cm, xd, w);
+#endif
+  }
+
+  if (!is_inter) {
+    if (bsize >= BLOCK_8X8 || unify_bsize) {
+      write_intra_mode(ec_ctx, bsize, mode, w);
+    } else {
+      int idx, idy;
+      const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+      const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+      for (idy = 0; idy < 2; idy += num_4x4_h) {
+        for (idx = 0; idx < 2; idx += num_4x4_w) {
+          const PREDICTION_MODE b_mode = mi->bmi[idy * 2 + idx].as_mode;
+          write_intra_mode(ec_ctx, bsize, b_mode, w);
+        }
+      }
+    }
+#if CONFIG_CB4X4
+    if (is_chroma_reference(mi_row, mi_col, bsize, xd->plane[1].subsampling_x,
+                            xd->plane[1].subsampling_y))
+      write_intra_uv_mode(ec_ctx, mbmi->uv_mode, mode, w);
+#else  // !CONFIG_CB4X4
+    write_intra_uv_mode(ec_ctx, mbmi->uv_mode, mode, w);
+#endif  // CONFIG_CB4X4
+
+#if CONFIG_EXT_INTRA
+    write_intra_angle_info(xd, ec_ctx, w);
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_PALETTE
+    if (bsize >= BLOCK_8X8 && cm->allow_screen_content_tools)
+      write_palette_mode_info(cm, xd, mi, w);
+#endif  // CONFIG_PALETTE
+#if CONFIG_FILTER_INTRA
+    if (bsize >= BLOCK_8X8 || unify_bsize)
+      write_filter_intra_mode_info(cm, mbmi, w);
+#endif  // CONFIG_FILTER_INTRA
+  } else {
+    int16_t mode_ctx;
+    write_ref_frames(cm, xd, w);
+
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+    if (is_compound)
+      mode_ctx = mbmi_ext->compound_mode_context[mbmi->ref_frame[0]];
+    else
+#endif  // CONFIG_EXT_INTER
+      mode_ctx = av1_mode_context_analyzer(mbmi_ext->mode_context,
+                                           mbmi->ref_frame, bsize, -1);
+#else  // CONFIG_REF_MV
+    mode_ctx = mbmi_ext->mode_context[mbmi->ref_frame[0]];
+#endif  // CONFIG_REF_MV
+
+    // If segment skip is not enabled code the mode.
+    if (!segfeature_active(seg, segment_id, SEG_LVL_SKIP)) {
+      if (bsize >= BLOCK_8X8 || unify_bsize) {
+#if CONFIG_EXT_INTER
+        if (is_inter_compound_mode(mode))
+          write_inter_compound_mode(cm, w, mode, mode_ctx);
+        else if (is_inter_singleref_mode(mode))
+#endif  // CONFIG_EXT_INTER
+          write_inter_mode(w, mode, ec_ctx, mode_ctx);
+
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+        if (mode == NEWMV || mode == NEW_NEWMV ||
+            have_nearmv_in_inter_mode(mode))
+#else
+        if (mode == NEARMV || mode == NEWMV)
+#endif
+          write_drl_idx(cm, mbmi, mbmi_ext, w);
+        else
+          assert(mbmi->ref_mv_idx == 0);
+#endif
+      }
+    }
+
+#if !CONFIG_DUAL_FILTER && !CONFIG_WARPED_MOTION && !CONFIG_GLOBAL_MOTION
+    write_mb_interp_filter(cpi, xd, w);
+#endif  // !CONFIG_DUAL_FILTER && !CONFIG_WARPED_MOTION
+
+    if (bsize < BLOCK_8X8 && !unify_bsize) {
+      const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+      const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+      int idx, idy;
+      for (idy = 0; idy < 2; idy += num_4x4_h) {
+        for (idx = 0; idx < 2; idx += num_4x4_w) {
+          const int j = idy * 2 + idx;
+          const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+          if (!is_compound)
+#endif  // CONFIG_EXT_INTER
+            mode_ctx = av1_mode_context_analyzer(mbmi_ext->mode_context,
+                                                 mbmi->ref_frame, bsize, j);
+#endif
+#if CONFIG_EXT_INTER
+          if (is_inter_compound_mode(b_mode))
+            write_inter_compound_mode(cm, w, b_mode, mode_ctx);
+          else if (is_inter_singleref_mode(b_mode))
+#endif  // CONFIG_EXT_INTER
+            write_inter_mode(w, b_mode, ec_ctx, mode_ctx);
+
+#if CONFIG_EXT_INTER
+          if (b_mode == NEWMV || b_mode == NEW_NEWMV) {
+#else
+          if (b_mode == NEWMV) {
+#endif  // CONFIG_EXT_INTER
+            for (ref = 0; ref < 1 + is_compound; ++ref) {
+#if CONFIG_REF_MV
+              int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+              int nmv_ctx = av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                                        mbmi_ext->ref_mv_stack[rf_type], ref,
+                                        mbmi->ref_mv_idx);
+              nmv_context *nmvc = &ec_ctx->nmvc[nmv_ctx];
+#endif
+              av1_encode_mv(cpi, w, &mi->bmi[j].as_mv[ref].as_mv,
+#if CONFIG_EXT_INTER
+                            &mi->bmi[j].ref_mv[ref].as_mv,
+#else
+#if CONFIG_REF_MV
+                            &mi->bmi[j].pred_mv[ref].as_mv,
+#else
+                            &mbmi_ext->ref_mvs[mbmi->ref_frame[ref]][0].as_mv,
+#endif  // CONFIG_REF_MV
+#endif  // CONFIG_EXT_INTER
+                            nmvc, allow_hp);
+            }
+          }
+#if CONFIG_EXT_INTER
+          else if (b_mode == NEAREST_NEWMV || b_mode == NEAR_NEWMV) {
+#if CONFIG_REF_MV
+            int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+            int nmv_ctx = av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                                      mbmi_ext->ref_mv_stack[rf_type], 1,
+                                      mbmi->ref_mv_idx);
+            nmv_context *nmvc = &ec_ctx->nmvc[nmv_ctx];
+#endif
+            av1_encode_mv(cpi, w, &mi->bmi[j].as_mv[1].as_mv,
+                          &mi->bmi[j].ref_mv[1].as_mv, nmvc, allow_hp);
+          } else if (b_mode == NEW_NEARESTMV || b_mode == NEW_NEARMV) {
+#if CONFIG_REF_MV
+            int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+            int nmv_ctx = av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                                      mbmi_ext->ref_mv_stack[rf_type], 0,
+                                      mbmi->ref_mv_idx);
+            nmv_context *nmvc = &ec_ctx->nmvc[nmv_ctx];
+#endif
+            av1_encode_mv(cpi, w, &mi->bmi[j].as_mv[0].as_mv,
+                          &mi->bmi[j].ref_mv[0].as_mv, nmvc, allow_hp);
+          }
+#endif  // CONFIG_EXT_INTER
+        }
+      }
+    } else {
+#if CONFIG_EXT_INTER
+      if (mode == NEWMV || mode == NEW_NEWMV) {
+#else
+      if (mode == NEWMV) {
+#endif  // CONFIG_EXT_INTER
+        int_mv ref_mv;
+        for (ref = 0; ref < 1 + is_compound; ++ref) {
+#if CONFIG_REF_MV
+          int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+          int nmv_ctx = av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                                    mbmi_ext->ref_mv_stack[rf_type], ref,
+                                    mbmi->ref_mv_idx);
+          nmv_context *nmvc = &ec_ctx->nmvc[nmv_ctx];
+#endif
+          ref_mv = mbmi_ext->ref_mvs[mbmi->ref_frame[ref]][0];
+          av1_encode_mv(cpi, w, &mbmi->mv[ref].as_mv, &ref_mv.as_mv, nmvc,
+                        allow_hp);
+        }
+#if CONFIG_EXT_INTER
+      } else if (mode == NEAREST_NEWMV || mode == NEAR_NEWMV) {
+#if CONFIG_REF_MV
+        int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+        int nmv_ctx =
+            av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                        mbmi_ext->ref_mv_stack[rf_type], 1, mbmi->ref_mv_idx);
+        nmv_context *nmvc = &ec_ctx->nmvc[nmv_ctx];
+#endif
+        av1_encode_mv(cpi, w, &mbmi->mv[1].as_mv,
+                      &mbmi_ext->ref_mvs[mbmi->ref_frame[1]][0].as_mv, nmvc,
+                      allow_hp);
+      } else if (mode == NEW_NEARESTMV || mode == NEW_NEARMV) {
+#if CONFIG_REF_MV
+        int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+        int nmv_ctx =
+            av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                        mbmi_ext->ref_mv_stack[rf_type], 0, mbmi->ref_mv_idx);
+        nmv_context *nmvc = &ec_ctx->nmvc[nmv_ctx];
+#endif
+        av1_encode_mv(cpi, w, &mbmi->mv[0].as_mv,
+                      &mbmi_ext->ref_mvs[mbmi->ref_frame[0]][0].as_mv, nmvc,
+                      allow_hp);
+#endif  // CONFIG_EXT_INTER
+      }
+    }
+
+#if CONFIG_EXT_INTER
+    if (cpi->common.reference_mode != COMPOUND_REFERENCE &&
+#if CONFIG_SUPERTX
+        !supertx_enabled &&
+#endif  // CONFIG_SUPERTX
+        is_interintra_allowed(mbmi)) {
+      const int interintra = mbmi->ref_frame[1] == INTRA_FRAME;
+      const int bsize_group = size_group_lookup[bsize];
+      aom_write(w, interintra, cm->fc->interintra_prob[bsize_group]);
+      if (interintra) {
+        write_interintra_mode(w, mbmi->interintra_mode,
+                              cm->fc->interintra_mode_prob[bsize_group]);
+        if (is_interintra_wedge_used(bsize)) {
+          aom_write(w, mbmi->use_wedge_interintra,
+                    cm->fc->wedge_interintra_prob[bsize]);
+          if (mbmi->use_wedge_interintra) {
+            aom_write_literal(w, mbmi->interintra_wedge_index,
+                              get_wedge_bits_lookup(bsize));
+            assert(mbmi->interintra_wedge_sign == 0);
+          }
+        }
+      }
+    }
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+#if CONFIG_SUPERTX
+    if (!supertx_enabled)
+#endif  // CONFIG_SUPERTX
+#if CONFIG_EXT_INTER
+      if (mbmi->ref_frame[1] != INTRA_FRAME)
+#endif  // CONFIG_EXT_INTER
+        write_motion_mode(cm, mi, w);
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_EXT_INTER
+    if (cpi->common.reference_mode != SINGLE_REFERENCE &&
+        is_inter_compound_mode(mbmi->mode)
+#if CONFIG_MOTION_VAR
+        && mbmi->motion_mode == SIMPLE_TRANSLATION
+#endif  // CONFIG_MOTION_VAR
+        && is_any_masked_compound_used(bsize)) {
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+      av1_write_token(w, av1_compound_type_tree,
+                      cm->fc->compound_type_prob[bsize],
+                      &compound_type_encodings[mbmi->interinter_compound_type]);
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#if CONFIG_WEDGE
+      if (mbmi->interinter_compound_type == COMPOUND_WEDGE) {
+        aom_write_literal(w, mbmi->wedge_index, get_wedge_bits_lookup(bsize));
+        aom_write_bit(w, mbmi->wedge_sign);
+      }
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+      if (mbmi->interinter_compound_type == COMPOUND_SEG) {
+        aom_write_literal(w, mbmi->mask_type, MAX_SEG_MASK_BITS);
+      }
+#endif  // CONFIG_COMPOUND_SEGMENT
+    }
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_DUAL_FILTER || CONFIG_WARPED_MOTION || CONFIG_GLOBAL_MOTION
+    write_mb_interp_filter(cpi, xd, w);
+#endif  // CONFIG_DUAL_FILTE || CONFIG_WARPED_MOTION
+  }
+
+#if !CONFIG_TXK_SEL
+  av1_write_tx_type(cm, xd,
+#if CONFIG_SUPERTX
+                    supertx_enabled,
+#endif
+                    w);
+#endif  // !CONFIG_TXK_SEL
+}
+
+#if CONFIG_DELTA_Q
+static void write_mb_modes_kf(AV1_COMMON *cm, MACROBLOCKD *xd, const int mi_row,
+                              const int mi_col, aom_writer *w) {
+  int skip;
+#else
+static void write_mb_modes_kf(AV1_COMMON *cm, const MACROBLOCKD *xd,
+                              const int mi_row, const int mi_col,
+                              aom_writer *w) {
+#endif
+  const struct segmentation *const seg = &cm->seg;
+  struct segmentation_probs *const segp = &cm->fc->seg;
+  const MODE_INFO *const mi = xd->mi[0];
+  const MODE_INFO *const above_mi = xd->above_mi;
+  const MODE_INFO *const left_mi = xd->left_mi;
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+  (void)mi_row;
+  (void)mi_col;
+
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  if (seg->update_map) write_segment_id(w, seg, segp, mbmi->segment_id);
+
+#if CONFIG_DELTA_Q
+  skip = write_skip(cm, xd, mbmi->segment_id, mi, w);
+  if (cm->delta_q_present_flag) {
+    int super_block_upper_left =
+        ((mi_row & MAX_MIB_MASK) == 0) && ((mi_col & MAX_MIB_MASK) == 0);
+    if ((bsize != BLOCK_LARGEST || skip == 0) && super_block_upper_left) {
+      assert(mbmi->current_q_index > 0);
+      int reduced_delta_qindex =
+          (mbmi->current_q_index - xd->prev_qindex) / cm->delta_q_res;
+      write_delta_qindex(cm, xd, reduced_delta_qindex, w);
+      xd->prev_qindex = mbmi->current_q_index;
+#if CONFIG_EXT_DELTA_Q
+      if (cm->delta_lf_present_flag) {
+        int reduced_delta_lflevel =
+            (mbmi->current_delta_lf_from_base - xd->prev_delta_lf_from_base) /
+            cm->delta_lf_res;
+        write_delta_lflevel(cm, xd, reduced_delta_lflevel, w);
+        xd->prev_delta_lf_from_base = mbmi->current_delta_lf_from_base;
+      }
+#endif  // CONFIG_EXT_DELTA_Q
+    }
+  }
+#else
+  write_skip(cm, xd, mbmi->segment_id, mi, w);
+#endif
+
+  if (cm->tx_mode == TX_MODE_SELECT &&
+#if CONFIG_CB4X4 && (CONFIG_VAR_TX || CONFIG_RECT_TX)
+#if CONFIG_RECT_TX
+      bsize > BLOCK_4X4 &&
+#else
+      bsize >= BLOCK_8X8 &&
+#endif  // CONFIG_RECT_TX
+#else
+      bsize >= BLOCK_8X8 &&
+#endif
+      !xd->lossless[mbmi->segment_id])
+    write_selected_tx_size(cm, xd, w);
+
+#if CONFIG_INTRABC
+  if (bsize >= BLOCK_8X8 && cm->allow_screen_content_tools) {
+    int use_intrabc = is_intrabc_block(mbmi);
+    aom_write(w, use_intrabc, INTRABC_PROB);
+    if (use_intrabc) {
+      assert(mbmi->mode == DC_PRED);
+      assert(mbmi->uv_mode == DC_PRED);
+      int_mv dv_ref;
+      av1_find_ref_dv(&dv_ref, mi_row, mi_col);
+      av1_encode_dv(w, &mbmi->mv[0].as_mv, &dv_ref.as_mv, &ec_ctx->ndvc);
+      return;
+    }
+  }
+#endif  // CONFIG_INTRABC
+
+  if (bsize >= BLOCK_8X8 || unify_bsize) {
+    write_intra_mode_kf(cm, ec_ctx, mi, above_mi, left_mi, 0, mbmi->mode, w);
+  } else {
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+    int idx, idy;
+
+    for (idy = 0; idy < 2; idy += num_4x4_h) {
+      for (idx = 0; idx < 2; idx += num_4x4_w) {
+        const int block = idy * 2 + idx;
+        write_intra_mode_kf(cm, ec_ctx, mi, above_mi, left_mi, block,
+                            mi->bmi[block].as_mode, w);
+      }
+    }
+  }
+
+#if CONFIG_CB4X4
+  if (is_chroma_reference(mi_row, mi_col, bsize, xd->plane[1].subsampling_x,
+                          xd->plane[1].subsampling_y))
+    write_intra_uv_mode(ec_ctx, mbmi->uv_mode, mbmi->mode, w);
+#else  // !CONFIG_CB4X4
+  write_intra_uv_mode(ec_ctx, mbmi->uv_mode, mbmi->mode, w);
+#endif  // CONFIG_CB4X4
+
+#if CONFIG_EXT_INTRA
+  write_intra_angle_info(xd, ec_ctx, w);
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_PALETTE
+  if (bsize >= BLOCK_8X8 && cm->allow_screen_content_tools)
+    write_palette_mode_info(cm, xd, mi, w);
+#endif  // CONFIG_PALETTE
+#if CONFIG_FILTER_INTRA
+  if (bsize >= BLOCK_8X8 || unify_bsize)
+    write_filter_intra_mode_info(cm, mbmi, w);
+#endif  // CONFIG_FILTER_INTRA
+
+#if !CONFIG_TXK_SEL
+  av1_write_tx_type(cm, xd,
+#if CONFIG_SUPERTX
+                    0,
+#endif
+                    w);
+#endif  // !CONFIG_TXK_SEL
+}
+
+#if CONFIG_SUPERTX
+#define write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled, \
+                              mi_row, mi_col)                              \
+  write_modes_b(cpi, tile, w, tok, tok_end, supertx_enabled, mi_row, mi_col)
+#else
+#define write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled, \
+                              mi_row, mi_col)                              \
+  write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col)
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_RD_DEBUG
+static void dump_mode_info(MODE_INFO *mi) {
+  printf("\nmi->mbmi.mi_row == %d\n", mi->mbmi.mi_row);
+  printf("&& mi->mbmi.mi_col == %d\n", mi->mbmi.mi_col);
+  printf("&& mi->mbmi.sb_type == %d\n", mi->mbmi.sb_type);
+  printf("&& mi->mbmi.tx_size == %d\n", mi->mbmi.tx_size);
+  if (mi->mbmi.sb_type >= BLOCK_8X8) {
+    printf("&& mi->mbmi.mode == %d\n", mi->mbmi.mode);
+  } else {
+    printf("&& mi->bmi[0].as_mode == %d\n", mi->bmi[0].as_mode);
+  }
+}
+static int rd_token_stats_mismatch(RD_STATS *rd_stats, TOKEN_STATS *token_stats,
+                                   int plane) {
+  if (rd_stats->txb_coeff_cost[plane] != token_stats->cost) {
+#if CONFIG_VAR_TX
+    int r, c;
+#endif
+    printf("\nplane %d rd_stats->txb_coeff_cost %d token_stats->cost %d\n",
+           plane, rd_stats->txb_coeff_cost[plane], token_stats->cost);
+#if CONFIG_VAR_TX
+    printf("rd txb_coeff_cost_map\n");
+    for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r) {
+      for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c) {
+        printf("%d ", rd_stats->txb_coeff_cost_map[plane][r][c]);
+      }
+      printf("\n");
+    }
+
+    printf("pack txb_coeff_cost_map\n");
+    for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r) {
+      for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c) {
+        printf("%d ", token_stats->txb_coeff_cost_map[r][c]);
+      }
+      printf("\n");
+    }
+#endif
+    return 1;
+  }
+  return 0;
+}
+#endif
+
+static void write_mbmi_b(AV1_COMP *cpi, const TileInfo *const tile,
+                         aom_writer *w,
+#if CONFIG_SUPERTX
+                         int supertx_enabled,
+#endif
+                         int mi_row, int mi_col) {
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+  MODE_INFO *m;
+  int bh, bw;
+  xd->mi = cm->mi_grid_visible + (mi_row * cm->mi_stride + mi_col);
+  m = xd->mi[0];
+
+  assert(m->mbmi.sb_type <= cm->sb_size);
+
+  bh = mi_size_high[m->mbmi.sb_type];
+  bw = mi_size_wide[m->mbmi.sb_type];
+
+  cpi->td.mb.mbmi_ext = cpi->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
+
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+  if (frame_is_intra_only(cm)) {
+    write_mb_modes_kf(cm, xd, mi_row, mi_col, w);
+  } else {
+#if CONFIG_VAR_TX
+    xd->above_txfm_context = cm->above_txfm_context + mi_col;
+    xd->left_txfm_context =
+        xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+#endif
+#if CONFIG_DUAL_FILTER
+    // has_subpel_mv_component needs the ref frame buffers set up to look
+    // up if they are scaled. has_subpel_mv_component is in turn needed by
+    // write_switchable_interp_filter, which is called by pack_inter_mode_mvs.
+    set_ref_ptrs(cm, xd, m->mbmi.ref_frame[0], m->mbmi.ref_frame[1]);
+#endif  // CONFIG_DUAL_FILTER
+#if 0
+    // NOTE(zoeliu): For debug
+    if (cm->current_video_frame == FRAME_TO_CHECK && cm->show_frame == 1) {
+      const PREDICTION_MODE mode = m->mbmi.mode;
+      const int segment_id = m->mbmi.segment_id;
+      const BLOCK_SIZE bsize = m->mbmi.sb_type;
+
+      // For sub8x8, simply dump out the first sub8x8 block info
+      const PREDICTION_MODE b_mode =
+          (bsize < BLOCK_8X8) ? m->bmi[0].as_mode : -1;
+      const int mv_x = (bsize < BLOCK_8X8) ?
+          m->bmi[0].as_mv[0].as_mv.row : m->mbmi.mv[0].as_mv.row;
+      const int mv_y = (bsize < BLOCK_8X8) ?
+          m->bmi[0].as_mv[0].as_mv.col : m->mbmi.mv[0].as_mv.col;
+
+      printf("Before pack_inter_mode_mvs(): "
+             "Frame=%d, (mi_row,mi_col)=(%d,%d), "
+             "mode=%d, segment_id=%d, bsize=%d, b_mode=%d, "
+             "mv[0]=(%d, %d), ref[0]=%d, ref[1]=%d\n",
+             cm->current_video_frame, mi_row, mi_col,
+             mode, segment_id, bsize, b_mode, mv_x, mv_y,
+             m->mbmi.ref_frame[0], m->mbmi.ref_frame[1]);
+    }
+#endif  // 0
+    pack_inter_mode_mvs(cpi, mi_row, mi_col,
+#if CONFIG_SUPERTX
+                        supertx_enabled,
+#endif
+                        w);
+  }
+}
+
+static void write_tokens_b(AV1_COMP *cpi, const TileInfo *const tile,
+                           aom_writer *w, const TOKENEXTRA **tok,
+                           const TOKENEXTRA *const tok_end, int mi_row,
+                           int mi_col) {
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+  MODE_INFO *const m = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &m->mbmi;
+  int plane;
+  int bh, bw;
+#if CONFIG_PVQ || CONFIG_LV_MAP
+  MACROBLOCK *const x = &cpi->td.mb;
+  (void)tok;
+  (void)tok_end;
+#endif
+  xd->mi = cm->mi_grid_visible + (mi_row * cm->mi_stride + mi_col);
+
+  assert(mbmi->sb_type <= cm->sb_size);
+
+  bh = mi_size_high[mbmi->sb_type];
+  bw = mi_size_wide[mbmi->sb_type];
+  cpi->td.mb.mbmi_ext = cpi->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
+
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+#if CONFIG_PALETTE
+  for (plane = 0; plane <= 1; ++plane) {
+    const uint8_t palette_size_plane =
+        mbmi->palette_mode_info.palette_size[plane];
+    if (palette_size_plane > 0) {
+#if CONFIG_INTRABC
+      assert(mbmi->use_intrabc == 0);
+#endif
+      int rows, cols;
+      assert(mbmi->sb_type >= BLOCK_8X8);
+      av1_get_block_dimensions(mbmi->sb_type, plane, xd, NULL, NULL, &rows,
+                               &cols);
+      assert(*tok < tok_end);
+      pack_palette_tokens(w, tok, palette_size_plane, rows * cols - 1);
+      assert(*tok < tok_end + mbmi->skip);
+    }
+  }
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_COEF_INTERLEAVE
+  if (!mbmi->skip) {
+    const struct macroblockd_plane *const pd_y = &xd->plane[0];
+    const struct macroblockd_plane *const pd_c = &xd->plane[1];
+    const TX_SIZE tx_log2_y = mbmi->tx_size;
+    const TX_SIZE tx_log2_c = get_uv_tx_size(mbmi, pd_c);
+    const int tx_sz_y = (1 << tx_log2_y);
+    const int tx_sz_c = (1 << tx_log2_c);
+
+    const BLOCK_SIZE plane_bsize_y =
+        get_plane_block_size(AOMMAX(mbmi->sb_type, 3), pd_y);
+    const BLOCK_SIZE plane_bsize_c =
+        get_plane_block_size(AOMMAX(mbmi->sb_type, 3), pd_c);
+
+    const int num_4x4_w_y = num_4x4_blocks_wide_lookup[plane_bsize_y];
+    const int num_4x4_w_c = num_4x4_blocks_wide_lookup[plane_bsize_c];
+    const int num_4x4_h_y = num_4x4_blocks_high_lookup[plane_bsize_y];
+    const int num_4x4_h_c = num_4x4_blocks_high_lookup[plane_bsize_c];
+
+    const int max_4x4_w_y = get_max_4x4_size(num_4x4_w_y, xd->mb_to_right_edge,
+                                             pd_y->subsampling_x);
+    const int max_4x4_h_y = get_max_4x4_size(num_4x4_h_y, xd->mb_to_bottom_edge,
+                                             pd_y->subsampling_y);
+    const int max_4x4_w_c = get_max_4x4_size(num_4x4_w_c, xd->mb_to_right_edge,
+                                             pd_c->subsampling_x);
+    const int max_4x4_h_c = get_max_4x4_size(num_4x4_h_c, xd->mb_to_bottom_edge,
+                                             pd_c->subsampling_y);
+
+    // The max_4x4_w/h may be smaller than tx_sz under some corner cases,
+    // i.e. when the SB is splitted by tile boundaries.
+    const int tu_num_w_y = (max_4x4_w_y + tx_sz_y - 1) / tx_sz_y;
+    const int tu_num_h_y = (max_4x4_h_y + tx_sz_y - 1) / tx_sz_y;
+    const int tu_num_w_c = (max_4x4_w_c + tx_sz_c - 1) / tx_sz_c;
+    const int tu_num_h_c = (max_4x4_h_c + tx_sz_c - 1) / tx_sz_c;
+    const int tu_num_y = tu_num_w_y * tu_num_h_y;
+    const int tu_num_c = tu_num_w_c * tu_num_h_c;
+
+    int tu_idx_y = 0, tu_idx_c = 0;
+    TOKEN_STATS token_stats;
+    init_token_stats(&token_stats);
+
+    assert(*tok < tok_end);
+
+    while (tu_idx_y < tu_num_y) {
+      pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx_log2_y, &token_stats);
+      assert(*tok < tok_end && (*tok)->token == EOSB_TOKEN);
+      (*tok)++;
+      tu_idx_y++;
+
+      if (tu_idx_c < tu_num_c) {
+        pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx_log2_c, &token_stats);
+        assert(*tok < tok_end && (*tok)->token == EOSB_TOKEN);
+        (*tok)++;
+
+        pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx_log2_c, &token_stats);
+        assert(*tok < tok_end && (*tok)->token == EOSB_TOKEN);
+        (*tok)++;
+
+        tu_idx_c++;
+      }
+    }
+
+    // In 422 case, it's possilbe that Chroma has more TUs than Luma
+    while (tu_idx_c < tu_num_c) {
+      pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx_log2_c, &token_stats);
+      assert(*tok < tok_end && (*tok)->token == EOSB_TOKEN);
+      (*tok)++;
+
+      pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx_log2_c, &token_stats);
+      assert(*tok < tok_end && (*tok)->token == EOSB_TOKEN);
+      (*tok)++;
+
+      tu_idx_c++;
+    }
+  }
+#else  // CONFIG_COEF_INTERLEAVE
+  if (!mbmi->skip) {
+#if !CONFIG_PVQ && !CONFIG_LV_MAP
+    assert(*tok < tok_end);
+#endif
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4
+      if (!is_chroma_reference(mi_row, mi_col, mbmi->sb_type,
+                               xd->plane[plane].subsampling_x,
+                               xd->plane[plane].subsampling_y)) {
+        (*tok)++;
+        continue;
+      }
+#endif
+#if CONFIG_VAR_TX
+      const struct macroblockd_plane *const pd = &xd->plane[plane];
+      BLOCK_SIZE bsize = mbmi->sb_type;
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+      const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+      const BLOCK_SIZE plane_bsize =
+          AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#endif
+#else
+      const BLOCK_SIZE plane_bsize =
+          get_plane_block_size(AOMMAX(bsize, BLOCK_8X8), pd);
+#endif
+
+      const int num_4x4_w =
+          block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+      const int num_4x4_h =
+          block_size_high[plane_bsize] >> tx_size_wide_log2[0];
+      int row, col;
+      TOKEN_STATS token_stats;
+      init_token_stats(&token_stats);
+
+      if (is_inter_block(mbmi)) {
+        const TX_SIZE max_tx_size = get_vartx_max_txsize(mbmi, plane_bsize);
+        int block = 0;
+        const int step =
+            tx_size_wide_unit[max_tx_size] * tx_size_high_unit[max_tx_size];
+        const int bkw = tx_size_wide_unit[max_tx_size];
+        const int bkh = tx_size_high_unit[max_tx_size];
+        for (row = 0; row < num_4x4_h; row += bkh) {
+          for (col = 0; col < num_4x4_w; col += bkw) {
+            pack_txb_tokens(w, tok, tok_end,
+#if CONFIG_PVQ
+                            x,
+#endif
+                            xd, mbmi, plane, plane_bsize, cm->bit_depth, block,
+                            row, col, max_tx_size, &token_stats);
+            block += step;
+          }
+        }
+#if CONFIG_RD_DEBUG
+        if (mbmi->sb_type >= BLOCK_8X8 &&
+            rd_token_stats_mismatch(&mbmi->rd_stats, &token_stats, plane)) {
+          dump_mode_info(m);
+          assert(0);
+        }
+#endif  // CONFIG_RD_DEBUG
+      } else {
+        TX_SIZE tx = get_tx_size(plane, xd);
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+        tx = AOMMAX(TX_4X4, tx);
+#endif
+        const int bkw = tx_size_wide_unit[tx];
+        const int bkh = tx_size_high_unit[tx];
+        for (row = 0; row < num_4x4_h; row += bkh) {
+          for (col = 0; col < num_4x4_w; col += bkw) {
+#if !CONFIG_PVQ
+            pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx, &token_stats);
+#else
+            pack_pvq_tokens(w, x, xd, plane, bsize, tx);
+#endif
+          }
+        }
+      }
+#else
+      TX_SIZE tx = get_tx_size(plane, xd);
+      TOKEN_STATS token_stats;
+#if !CONFIG_PVQ
+      init_token_stats(&token_stats);
+#if CONFIG_LV_MAP
+      (void)tx;
+      av1_write_coeffs_mb(cm, x, w, plane);
+#else   // CONFIG_LV_MAP
+      pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx, &token_stats);
+#endif  // CONFIG_LV_MAP
+
+#else
+      (void)token_stats;
+      pack_pvq_tokens(w, x, xd, plane, mbmi->sb_type, tx);
+#endif
+#if CONFIG_RD_DEBUG
+      if (is_inter_block(mbmi) && mbmi->sb_type >= BLOCK_8X8 &&
+          rd_token_stats_mismatch(&mbmi->rd_stats, &token_stats, plane)) {
+        dump_mode_info(m);
+        assert(0);
+      }
+#endif  // CONFIG_RD_DEBUG
+#endif  // CONFIG_VAR_TX
+
+#if !CONFIG_PVQ && !CONFIG_LV_MAP
+      assert(*tok < tok_end && (*tok)->token == EOSB_TOKEN);
+      (*tok)++;
+#endif
+    }
+  }
+#endif  // CONFIG_COEF_INTERLEAVE
+}
+
+#if CONFIG_MOTION_VAR && CONFIG_NCOBMC
+static void write_tokens_sb(AV1_COMP *cpi, const TileInfo *const tile,
+                            aom_writer *w, const TOKENEXTRA **tok,
+                            const TOKENEXTRA *const tok_end, int mi_row,
+                            int mi_col, BLOCK_SIZE bsize) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int hbs = mi_size_wide[bsize] / 2;
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  partition = get_partition(cm, mi_row, mi_col, bsize);
+  subsize = get_subsize(bsize, partition);
+
+  if (subsize < BLOCK_8X8 && !unify_bsize) {
+    write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+  } else {
+    switch (partition) {
+      case PARTITION_NONE:
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+        break;
+      case PARTITION_HORZ:
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+        if (mi_row + hbs < cm->mi_rows)
+          write_tokens_b(cpi, tile, w, tok, tok_end, mi_row + hbs, mi_col);
+        break;
+      case PARTITION_VERT:
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+        if (mi_col + hbs < cm->mi_cols)
+          write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col + hbs);
+        break;
+      case PARTITION_SPLIT:
+        write_tokens_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col, subsize);
+        write_tokens_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col + hbs,
+                        subsize);
+        write_tokens_sb(cpi, tile, w, tok, tok_end, mi_row + hbs, mi_col,
+                        subsize);
+        write_tokens_sb(cpi, tile, w, tok, tok_end, mi_row + hbs, mi_col + hbs,
+                        subsize);
+        break;
+#if CONFIG_EXT_PARTITION_TYPES
+      case PARTITION_HORZ_A:
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col + hbs);
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row + hbs, mi_col);
+        break;
+      case PARTITION_HORZ_B:
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row + hbs, mi_col);
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row + hbs, mi_col + hbs);
+        break;
+      case PARTITION_VERT_A:
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row + hbs, mi_col);
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col + hbs);
+        break;
+      case PARTITION_VERT_B:
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col + hbs);
+        write_tokens_b(cpi, tile, w, tok, tok_end, mi_row + hbs, mi_col + hbs);
+        break;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+      default: assert(0);
+    }
+  }
+}
+#endif
+
+static void write_modes_b(AV1_COMP *cpi, const TileInfo *const tile,
+                          aom_writer *w, const TOKENEXTRA **tok,
+                          const TOKENEXTRA *const tok_end,
+#if CONFIG_SUPERTX
+                          int supertx_enabled,
+#endif
+                          int mi_row, int mi_col) {
+  write_mbmi_b(cpi, tile, w,
+#if CONFIG_SUPERTX
+               supertx_enabled,
+#endif
+               mi_row, mi_col);
+#if CONFIG_MOTION_VAR && CONFIG_NCOBMC
+  (void)tok;
+  (void)tok_end;
+#else
+#if !CONFIG_PVQ && CONFIG_SUPERTX
+  if (!supertx_enabled)
+#endif
+    write_tokens_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
+#endif
+}
+
+static void write_partition(const AV1_COMMON *const cm,
+                            const MACROBLOCKD *const xd, int hbs, int mi_row,
+                            int mi_col, PARTITION_TYPE p, BLOCK_SIZE bsize,
+                            aom_writer *w) {
+  const int has_rows = (mi_row + hbs) < cm->mi_rows;
+  const int has_cols = (mi_col + hbs) < cm->mi_cols;
+  const int is_partition_point = bsize >= BLOCK_8X8;
+  const int ctx = is_partition_point
+                      ? partition_plane_context(xd, mi_row, mi_col,
+#if CONFIG_UNPOISON_PARTITION_CTX
+                                                has_rows, has_cols,
+#endif
+                                                bsize)
+                      : 0;
+#if CONFIG_UNPOISON_PARTITION_CTX
+  const aom_prob *const probs =
+      ctx < PARTITION_CONTEXTS ? cm->fc->partition_prob[ctx] : NULL;
+#else
+  const aom_prob *const probs = cm->fc->partition_prob[ctx];
+#endif
+
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+  (void)cm;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+
+  if (!is_partition_point) return;
+
+  if (has_rows && has_cols) {
+#if CONFIG_EXT_PARTITION_TYPES
+    if (bsize <= BLOCK_8X8)
+#if CONFIG_EC_MULTISYMBOL
+      aom_write_symbol(w, p, ec_ctx->partition_cdf[ctx], PARTITION_TYPES);
+#else
+      av1_write_token(w, av1_partition_tree, probs, &partition_encodings[p]);
+#endif
+    else
+#if CONFIG_EC_MULTISYMBOL
+      aom_write_symbol(w, p, ec_ctx->partition_cdf[ctx], EXT_PARTITION_TYPES);
+#else
+      av1_write_token(w, av1_ext_partition_tree, probs,
+                      &ext_partition_encodings[p]);
+#endif  // CONFIG_EC_MULTISYMBOL
+#else
+#if CONFIG_EC_MULTISYMBOL
+    aom_write_symbol(w, p, ec_ctx->partition_cdf[ctx], PARTITION_TYPES);
+#else
+    av1_write_token(w, av1_partition_tree, probs, &partition_encodings[p]);
+#endif
+#endif  // CONFIG_EXT_PARTITION_TYPES
+  } else if (!has_rows && has_cols) {
+    assert(p == PARTITION_SPLIT || p == PARTITION_HORZ);
+    aom_write(w, p == PARTITION_SPLIT, probs[1]);
+  } else if (has_rows && !has_cols) {
+    assert(p == PARTITION_SPLIT || p == PARTITION_VERT);
+    aom_write(w, p == PARTITION_SPLIT, probs[2]);
+  } else {
+    assert(p == PARTITION_SPLIT);
+  }
+}
+
+#if CONFIG_SUPERTX
+#define write_modes_sb_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,   \
+                               mi_row, mi_col, bsize)                         \
+  write_modes_sb(cpi, tile, w, tok, tok_end, supertx_enabled, mi_row, mi_col, \
+                 bsize)
+#else
+#define write_modes_sb_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled, \
+                               mi_row, mi_col, bsize)                       \
+  write_modes_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col, bsize)
+#endif  // CONFIG_SUPERTX
+
+static void write_modes_sb(AV1_COMP *const cpi, const TileInfo *const tile,
+                           aom_writer *const w, const TOKENEXTRA **tok,
+                           const TOKENEXTRA *const tok_end,
+#if CONFIG_SUPERTX
+                           int supertx_enabled,
+#endif
+                           int mi_row, int mi_col, BLOCK_SIZE bsize) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+  const int hbs = mi_size_wide[bsize] / 2;
+  const PARTITION_TYPE partition = get_partition(cm, mi_row, mi_col, bsize);
+  const BLOCK_SIZE subsize = get_subsize(bsize, partition);
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+#if CONFIG_SUPERTX
+  const int mi_offset = mi_row * cm->mi_stride + mi_col;
+  MB_MODE_INFO *mbmi;
+  const int pack_token = !supertx_enabled;
+  TX_SIZE supertx_size;
+  int plane;
+#endif
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  write_partition(cm, xd, hbs, mi_row, mi_col, partition, bsize, w);
+#if CONFIG_SUPERTX
+  mbmi = &cm->mi_grid_visible[mi_offset]->mbmi;
+  xd->mi = cm->mi_grid_visible + mi_offset;
+  set_mi_row_col(xd, tile, mi_row, mi_size_high[bsize], mi_col,
+                 mi_size_wide[bsize],
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+  if (!supertx_enabled && !frame_is_intra_only(cm) &&
+      partition != PARTITION_NONE && bsize <= MAX_SUPERTX_BLOCK_SIZE &&
+      !xd->lossless[0]) {
+    aom_prob prob;
+    supertx_size = max_txsize_lookup[bsize];
+    prob = cm->fc->supertx_prob[partition_supertx_context_lookup[partition]]
+                               [supertx_size];
+    supertx_enabled = (xd->mi[0]->mbmi.tx_size == supertx_size);
+    aom_write(w, supertx_enabled, prob);
+  }
+#endif  // CONFIG_SUPERTX
+  if (subsize < BLOCK_8X8 && !unify_bsize) {
+    write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled, mi_row,
+                          mi_col);
+  } else {
+    switch (partition) {
+      case PARTITION_NONE:
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row, mi_col);
+        break;
+      case PARTITION_HORZ:
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row, mi_col);
+        if (mi_row + hbs < cm->mi_rows)
+          write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                                mi_row + hbs, mi_col);
+        break;
+      case PARTITION_VERT:
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row, mi_col);
+        if (mi_col + hbs < cm->mi_cols)
+          write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                                mi_row, mi_col + hbs);
+        break;
+      case PARTITION_SPLIT:
+        write_modes_sb_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                               mi_row, mi_col, subsize);
+        write_modes_sb_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                               mi_row, mi_col + hbs, subsize);
+        write_modes_sb_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                               mi_row + hbs, mi_col, subsize);
+        write_modes_sb_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                               mi_row + hbs, mi_col + hbs, subsize);
+        break;
+#if CONFIG_EXT_PARTITION_TYPES
+      case PARTITION_HORZ_A:
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row, mi_col);
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row, mi_col + hbs);
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row + hbs, mi_col);
+        break;
+      case PARTITION_HORZ_B:
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row, mi_col);
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row + hbs, mi_col);
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row + hbs, mi_col + hbs);
+        break;
+      case PARTITION_VERT_A:
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row, mi_col);
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row + hbs, mi_col);
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row, mi_col + hbs);
+        break;
+      case PARTITION_VERT_B:
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row, mi_col);
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row, mi_col + hbs);
+        write_modes_b_wrapper(cpi, tile, w, tok, tok_end, supertx_enabled,
+                              mi_row + hbs, mi_col + hbs);
+        break;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+      default: assert(0);
+    }
+  }
+#if CONFIG_SUPERTX
+  if (partition != PARTITION_NONE && supertx_enabled && pack_token) {
+    int skip;
+    const int bsw = mi_size_wide[bsize];
+    const int bsh = mi_size_high[bsize];
+
+    xd->mi = cm->mi_grid_visible + mi_offset;
+    supertx_size = mbmi->tx_size;
+    set_mi_row_col(xd, tile, mi_row, bsh, mi_col, bsw,
+#if CONFIG_DEPENDENT_HORZTILES
+                   cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                   cm->mi_rows, cm->mi_cols);
+
+    assert(IMPLIES(!cm->seg.enabled, mbmi->segment_id_supertx == 0));
+    assert(mbmi->segment_id_supertx < MAX_SEGMENTS);
+
+    skip = write_skip(cm, xd, mbmi->segment_id_supertx, xd->mi[0], w);
+#if CONFIG_EXT_TX
+    if (get_ext_tx_types(supertx_size, bsize, 1, cm->reduced_tx_set_used) > 1 &&
+        !skip) {
+      const int eset =
+          get_ext_tx_set(supertx_size, bsize, 1, cm->reduced_tx_set_used);
+      if (eset > 0) {
+#if CONFIG_EC_MULTISYMBOL
+#if CONFIG_EC_ADAPT
+        FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#else
+        FRAME_CONTEXT *ec_ctx = cm->fc;
+#endif
+        aom_write_symbol(w, av1_ext_tx_inter_ind[eset][mbmi->tx_type],
+                         ec_ctx->inter_ext_tx_cdf[eset][supertx_size],
+                         ext_tx_cnt_inter[eset]);
+#else
+        av1_write_token(w, av1_ext_tx_inter_tree[eset],
+                        cm->fc->inter_ext_tx_prob[eset][supertx_size],
+                        &ext_tx_inter_encodings[eset][mbmi->tx_type]);
+#endif
+      }
+    }
+#else
+    if (supertx_size < TX_32X32 && !skip) {
+      av1_write_token(w, av1_ext_tx_tree,
+                      cm->fc->inter_ext_tx_prob[supertx_size],
+                      &ext_tx_encodings[mbmi->tx_type]);
+    }
+#endif  // CONFIG_EXT_TX
+
+    if (!skip) {
+      assert(*tok < tok_end);
+      for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+        const struct macroblockd_plane *const pd = &xd->plane[plane];
+        const int mbmi_txb_size = txsize_to_bsize[mbmi->tx_size];
+        const BLOCK_SIZE plane_bsize = get_plane_block_size(mbmi_txb_size, pd);
+
+        const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+        const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+
+        int row, col;
+        TX_SIZE tx = get_tx_size(plane, xd);
+        BLOCK_SIZE txb_size = txsize_to_bsize[tx];
+
+        const int stepr = tx_size_high_unit[txb_size];
+        const int stepc = tx_size_wide_unit[txb_size];
+
+        TOKEN_STATS token_stats;
+        token_stats.cost = 0;
+        for (row = 0; row < max_blocks_high; row += stepr)
+          for (col = 0; col < max_blocks_wide; col += stepc)
+            pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx, &token_stats);
+        assert(*tok < tok_end && (*tok)->token == EOSB_TOKEN);
+        (*tok)++;
+      }
+    }
+#if CONFIG_VAR_TX
+    xd->above_txfm_context = cm->above_txfm_context + mi_col;
+    xd->left_txfm_context =
+        xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+    set_txfm_ctxs(xd->mi[0]->mbmi.tx_size, bsw, bsh, skip, xd);
+#endif
+  }
+#endif  // CONFIG_SUPERTX
+
+// update partition context
+#if CONFIG_EXT_PARTITION_TYPES
+  update_ext_partition_context(xd, mi_row, mi_col, subsize, bsize, partition);
+#else
+  if (bsize >= BLOCK_8X8 &&
+      (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
+    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+#if CONFIG_CDEF
+#if CONFIG_EXT_PARTITION
+  if (cm->sb_size == BLOCK_128X128 && bsize == BLOCK_128X128 &&
+      !sb_all_skip(cm, mi_row, mi_col)) {
+    aom_write_literal(w, cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col]
+                             ->mbmi.cdef_strength,
+                      cm->cdef_bits);
+  } else if (cm->sb_size == BLOCK_64X64 && bsize == BLOCK_64X64 &&
+#else
+  if (bsize == BLOCK_64X64 &&
+#endif  // CONFIG_EXT_PARTITION
+             !sb_all_skip(cm, mi_row, mi_col)) {
+    if (cm->cdef_bits != 0)
+      aom_write_literal(w, cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col]
+                               ->mbmi.cdef_strength,
+                        cm->cdef_bits);
+  }
+#endif
+}
+
+static void write_modes(AV1_COMP *const cpi, const TileInfo *const tile,
+                        aom_writer *const w, const TOKENEXTRA **tok,
+                        const TOKENEXTRA *const tok_end) {
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+  const int mi_row_start = tile->mi_row_start;
+  const int mi_row_end = tile->mi_row_end;
+  const int mi_col_start = tile->mi_col_start;
+  const int mi_col_end = tile->mi_col_end;
+  int mi_row, mi_col;
+
+#if CONFIG_DEPENDENT_HORZTILES
+#if CONFIG_TILE_GROUPS
+  if (!cm->dependent_horz_tiles || mi_row_start == 0 ||
+      tile->tg_horz_boundary) {
+#else
+  if (!cm->dependent_horz_tiles || mi_row_start == 0) {
+#endif
+    av1_zero_above_context(cm, mi_col_start, mi_col_end);
+  }
+#else
+  av1_zero_above_context(cm, mi_col_start, mi_col_end);
+#endif
+#if CONFIG_PVQ
+  assert(cpi->td.mb.pvq_q->curr_pos == 0);
+#endif
+#if CONFIG_DELTA_Q
+  if (cpi->common.delta_q_present_flag) {
+    xd->prev_qindex = cpi->common.base_qindex;
+#if CONFIG_EXT_DELTA_Q
+    if (cpi->common.delta_lf_present_flag) {
+      xd->prev_delta_lf_from_base = 0;
+    }
+#endif  // CONFIG_EXT_DELTA_Q
+  }
+#endif
+
+  for (mi_row = mi_row_start; mi_row < mi_row_end; mi_row += cm->mib_size) {
+    av1_zero_left_context(xd);
+
+    for (mi_col = mi_col_start; mi_col < mi_col_end; mi_col += cm->mib_size) {
+      write_modes_sb_wrapper(cpi, tile, w, tok, tok_end, 0, mi_row, mi_col,
+                             cm->sb_size);
+#if CONFIG_MOTION_VAR && CONFIG_NCOBMC
+      write_tokens_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col, cm->sb_size);
+#endif
+    }
+  }
+#if CONFIG_PVQ
+  // Check that the number of PVQ blocks encoded and written to the bitstream
+  // are the same
+  assert(cpi->td.mb.pvq_q->curr_pos == cpi->td.mb.pvq_q->last_pos);
+  // Reset curr_pos in case we repack the bitstream
+  cpi->td.mb.pvq_q->curr_pos = 0;
+#endif
+}
+
+#if !CONFIG_LV_MAP
+#if !CONFIG_PVQ && !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET)
+static void build_tree_distribution(AV1_COMP *cpi, TX_SIZE tx_size,
+                                    av1_coeff_stats *coef_branch_ct,
+                                    av1_coeff_probs_model *coef_probs) {
+  av1_coeff_count *coef_counts = cpi->td.rd_counts.coef_counts[tx_size];
+  unsigned int(*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
+      cpi->common.counts.eob_branch[tx_size];
+  int i, j, k, l, m;
+#if CONFIG_RECT_TX
+  assert(!is_rect_tx(tx_size));
+#endif  // CONFIG_RECT_TX
+
+  for (i = 0; i < PLANE_TYPES; ++i) {
+    for (j = 0; j < REF_TYPES; ++j) {
+      for (k = 0; k < COEF_BANDS; ++k) {
+        for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+          av1_tree_probs_from_distribution(av1_coef_tree,
+                                           coef_branch_ct[i][j][k][l],
+                                           coef_counts[i][j][k][l]);
+          coef_branch_ct[i][j][k][l][0][1] =
+              eob_branch_ct[i][j][k][l] - coef_branch_ct[i][j][k][l][0][0];
+          for (m = 0; m < UNCONSTRAINED_NODES; ++m)
+            coef_probs[i][j][k][l][m] =
+                get_binary_prob(coef_branch_ct[i][j][k][l][m][0],
+                                coef_branch_ct[i][j][k][l][m][1]);
+        }
+      }
+    }
+  }
+}
+
+#if !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET)
+static void update_coef_probs_common(aom_writer *const bc, AV1_COMP *cpi,
+                                     TX_SIZE tx_size,
+                                     av1_coeff_stats *frame_branch_ct,
+                                     av1_coeff_probs_model *new_coef_probs) {
+  av1_coeff_probs_model *old_coef_probs = cpi->common.fc->coef_probs[tx_size];
+  const aom_prob upd = DIFF_UPDATE_PROB;
+#if CONFIG_EC_ADAPT
+  const int entropy_nodes_update = UNCONSTRAINED_NODES - 1;
+#else
+  const int entropy_nodes_update = UNCONSTRAINED_NODES;
+#endif
+  int i, j, k, l, t;
+  int stepsize = cpi->sf.coeff_prob_appx_step;
+#if CONFIG_TILE_GROUPS
+  const int probwt = cpi->common.num_tg;
+#else
+  const int probwt = 1;
+#endif
+#if CONFIG_RECT_TX
+  assert(!is_rect_tx(tx_size));
+#endif  // CONFIG_RECT_TX
+
+  switch (cpi->sf.use_fast_coef_updates) {
+    case TWO_LOOP: {
+      /* dry run to see if there is any update at all needed */
+      int savings = 0;
+      int update[2] = { 0, 0 };
+      for (i = 0; i < PLANE_TYPES; ++i) {
+        for (j = 0; j < REF_TYPES; ++j) {
+          for (k = 0; k < COEF_BANDS; ++k) {
+            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+              for (t = 0; t < entropy_nodes_update; ++t) {
+                aom_prob newp = new_coef_probs[i][j][k][l][t];
+                const aom_prob oldp = old_coef_probs[i][j][k][l][t];
+                int s;
+                int u = 0;
+                if (t == PIVOT_NODE)
+                  s = av1_prob_diff_update_savings_search_model(
+                      frame_branch_ct[i][j][k][l][0], oldp, &newp, upd,
+                      stepsize, probwt);
+                else
+                  s = av1_prob_diff_update_savings_search(
+                      frame_branch_ct[i][j][k][l][t], oldp, &newp, upd, probwt);
+
+                if (s > 0 && newp != oldp) u = 1;
+                if (u)
+                  savings += s - (int)(av1_cost_zero(upd));
+                else
+                  savings -= (int)(av1_cost_zero(upd));
+                update[u]++;
+              }
+            }
+          }
+        }
+      }
+
+      /* Is coef updated at all */
+      if (update[1] == 0 || savings < 0) {
+        aom_write_bit(bc, 0);
+        return;
+      }
+      aom_write_bit(bc, 1);
+      for (i = 0; i < PLANE_TYPES; ++i) {
+        for (j = 0; j < REF_TYPES; ++j) {
+          for (k = 0; k < COEF_BANDS; ++k) {
+            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+              // calc probs and branch cts for this frame only
+              for (t = 0; t < entropy_nodes_update; ++t) {
+                aom_prob newp = new_coef_probs[i][j][k][l][t];
+                aom_prob *oldp = old_coef_probs[i][j][k][l] + t;
+                int s;
+                int u = 0;
+                if (t == PIVOT_NODE)
+                  s = av1_prob_diff_update_savings_search_model(
+                      frame_branch_ct[i][j][k][l][0], *oldp, &newp, upd,
+                      stepsize, probwt);
+                else
+                  s = av1_prob_diff_update_savings_search(
+                      frame_branch_ct[i][j][k][l][t], *oldp, &newp, upd,
+                      probwt);
+                if (s > 0 && newp != *oldp) u = 1;
+                aom_write(bc, u, upd);
+                if (u) {
+                  /* send/use new probability */
+                  av1_write_prob_diff_update(bc, newp, *oldp);
+                  *oldp = newp;
+                }
+              }
+            }
+          }
+        }
+      }
+      return;
+    }
+
+    case ONE_LOOP_REDUCED: {
+      int updates = 0;
+      int noupdates_before_first = 0;
+      for (i = 0; i < PLANE_TYPES; ++i) {
+        for (j = 0; j < REF_TYPES; ++j) {
+          for (k = 0; k < COEF_BANDS; ++k) {
+            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+              // calc probs and branch cts for this frame only
+              for (t = 0; t < entropy_nodes_update; ++t) {
+                aom_prob newp = new_coef_probs[i][j][k][l][t];
+                aom_prob *oldp = old_coef_probs[i][j][k][l] + t;
+                int s;
+                int u = 0;
+                if (t == PIVOT_NODE) {
+                  s = av1_prob_diff_update_savings_search_model(
+                      frame_branch_ct[i][j][k][l][0], *oldp, &newp, upd,
+                      stepsize, probwt);
+                } else {
+                  s = av1_prob_diff_update_savings_search(
+                      frame_branch_ct[i][j][k][l][t], *oldp, &newp, upd,
+                      probwt);
+                }
+
+                if (s > 0 && newp != *oldp) u = 1;
+                updates += u;
+                if (u == 0 && updates == 0) {
+                  noupdates_before_first++;
+                  continue;
+                }
+                if (u == 1 && updates == 1) {
+                  int v;
+                  // first update
+                  aom_write_bit(bc, 1);
+                  for (v = 0; v < noupdates_before_first; ++v)
+                    aom_write(bc, 0, upd);
+                }
+                aom_write(bc, u, upd);
+                if (u) {
+                  /* send/use new probability */
+                  av1_write_prob_diff_update(bc, newp, *oldp);
+                  *oldp = newp;
+                }
+              }
+            }
+          }
+        }
+      }
+      if (updates == 0) {
+        aom_write_bit(bc, 0);  // no updates
+      }
+      return;
+    }
+    default: assert(0);
+  }
+}
+#endif
+#if CONFIG_SUBFRAME_PROB_UPDATE
+// Calculate the token counts between subsequent subframe updates.
+static void get_coef_counts_diff(
+    AV1_COMP *cpi, int index,
+    av1_coeff_count coef_counts[TX_SIZES][PLANE_TYPES],
+    unsigned int eob_counts[TX_SIZES][PLANE_TYPES][REF_TYPES][COEF_BANDS]
+                           [COEFF_CONTEXTS]) {
+  int i, j, k, l, m, tx_size, val;
+  const int max_idx = cpi->common.coef_probs_update_idx;
+  const TX_MODE tx_mode = cpi->common.tx_mode;
+  const int max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+  const SUBFRAME_STATS *subframe_stats = &cpi->subframe_stats;
+
+  assert(max_idx < COEF_PROBS_BUFS);
+
+  for (tx_size = 0; tx_size <= max_tx_size; ++tx_size)
+    for (i = 0; i < PLANE_TYPES; ++i)
+      for (j = 0; j < REF_TYPES; ++j)
+        for (k = 0; k < COEF_BANDS; ++k)
+          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+            if (index == max_idx) {
+              val =
+                  cpi->common.counts.eob_branch[tx_size][i][j][k][l] -
+                  subframe_stats->eob_counts_buf[max_idx][tx_size][i][j][k][l];
+            } else {
+              val = subframe_stats
+                        ->eob_counts_buf[index + 1][tx_size][i][j][k][l] -
+                    subframe_stats->eob_counts_buf[index][tx_size][i][j][k][l];
+            }
+            assert(val >= 0);
+            eob_counts[tx_size][i][j][k][l] = val;
+
+            for (m = 0; m < ENTROPY_TOKENS; ++m) {
+              if (index == max_idx) {
+                val = cpi->td.rd_counts.coef_counts[tx_size][i][j][k][l][m] -
+                      subframe_stats
+                          ->coef_counts_buf[max_idx][tx_size][i][j][k][l][m];
+              } else {
+                val = subframe_stats
+                          ->coef_counts_buf[index + 1][tx_size][i][j][k][l][m] -
+                      subframe_stats
+                          ->coef_counts_buf[index][tx_size][i][j][k][l][m];
+              }
+              assert(val >= 0);
+              coef_counts[tx_size][i][j][k][l][m] = val;
+            }
+          }
+}
+
+static void update_coef_probs_subframe(
+    aom_writer *const bc, AV1_COMP *cpi, TX_SIZE tx_size,
+    av1_coeff_stats branch_ct[COEF_PROBS_BUFS][TX_SIZES][PLANE_TYPES],
+    av1_coeff_probs_model *new_coef_probs) {
+  av1_coeff_probs_model *old_coef_probs = cpi->common.fc->coef_probs[tx_size];
+  const aom_prob upd = DIFF_UPDATE_PROB;
+  const int entropy_nodes_update = UNCONSTRAINED_NODES;
+  int i, j, k, l, t;
+  int stepsize = cpi->sf.coeff_prob_appx_step;
+  const int max_idx = cpi->common.coef_probs_update_idx;
+  int idx;
+  unsigned int this_branch_ct[ENTROPY_NODES][COEF_PROBS_BUFS][2];
+
+  switch (cpi->sf.use_fast_coef_updates) {
+    case TWO_LOOP: {
+      /* dry run to see if there is any update at all needed */
+      int savings = 0;
+      int update[2] = { 0, 0 };
+      for (i = 0; i < PLANE_TYPES; ++i) {
+        for (j = 0; j < REF_TYPES; ++j) {
+          for (k = 0; k < COEF_BANDS; ++k) {
+            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+              for (t = 0; t < ENTROPY_NODES; ++t) {
+                for (idx = 0; idx <= max_idx; ++idx) {
+                  memcpy(this_branch_ct[t][idx],
+                         branch_ct[idx][tx_size][i][j][k][l][t],
+                         2 * sizeof(this_branch_ct[t][idx][0]));
+                }
+              }
+              for (t = 0; t < entropy_nodes_update; ++t) {
+                aom_prob newp = new_coef_probs[i][j][k][l][t];
+                const aom_prob oldp = old_coef_probs[i][j][k][l][t];
+                int s, u = 0;
+
+                if (t == PIVOT_NODE)
+                  s = av1_prob_update_search_model_subframe(
+                      this_branch_ct, old_coef_probs[i][j][k][l], &newp, upd,
+                      stepsize, max_idx);
+                else
+                  s = av1_prob_update_search_subframe(this_branch_ct[t], oldp,
+                                                      &newp, upd, max_idx);
+                if (s > 0 && newp != oldp) u = 1;
+                if (u)
+                  savings += s - (int)(av1_cost_zero(upd));
+                else
+                  savings -= (int)(av1_cost_zero(upd));
+                update[u]++;
+              }
+            }
+          }
+        }
+      }
+
+      /* Is coef updated at all */
+      if (update[1] == 0 || savings < 0) {
+        aom_write_bit(bc, 0);
+        return;
+      }
+      aom_write_bit(bc, 1);
+      for (i = 0; i < PLANE_TYPES; ++i) {
+        for (j = 0; j < REF_TYPES; ++j) {
+          for (k = 0; k < COEF_BANDS; ++k) {
+            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+              for (t = 0; t < ENTROPY_NODES; ++t) {
+                for (idx = 0; idx <= max_idx; ++idx) {
+                  memcpy(this_branch_ct[t][idx],
+                         branch_ct[idx][tx_size][i][j][k][l][t],
+                         2 * sizeof(this_branch_ct[t][idx][0]));
+                }
+              }
+              for (t = 0; t < entropy_nodes_update; ++t) {
+                aom_prob newp = new_coef_probs[i][j][k][l][t];
+                aom_prob *oldp = old_coef_probs[i][j][k][l] + t;
+                int s;
+                int u = 0;
+
+                if (t == PIVOT_NODE)
+                  s = av1_prob_update_search_model_subframe(
+                      this_branch_ct, old_coef_probs[i][j][k][l], &newp, upd,
+                      stepsize, max_idx);
+                else
+                  s = av1_prob_update_search_subframe(this_branch_ct[t], *oldp,
+                                                      &newp, upd, max_idx);
+                if (s > 0 && newp != *oldp) u = 1;
+                aom_write(bc, u, upd);
+                if (u) {
+                  /* send/use new probability */
+                  av1_write_prob_diff_update(bc, newp, *oldp);
+                  *oldp = newp;
+                }
+              }
+            }
+          }
+        }
+      }
+      return;
+    }
+
+    case ONE_LOOP_REDUCED: {
+      int updates = 0;
+      int noupdates_before_first = 0;
+      for (i = 0; i < PLANE_TYPES; ++i) {
+        for (j = 0; j < REF_TYPES; ++j) {
+          for (k = 0; k < COEF_BANDS; ++k) {
+            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+              for (t = 0; t < ENTROPY_NODES; ++t) {
+                for (idx = 0; idx <= max_idx; ++idx) {
+                  memcpy(this_branch_ct[t][idx],
+                         branch_ct[idx][tx_size][i][j][k][l][t],
+                         2 * sizeof(this_branch_ct[t][idx][0]));
+                }
+              }
+              for (t = 0; t < entropy_nodes_update; ++t) {
+                aom_prob newp = new_coef_probs[i][j][k][l][t];
+                aom_prob *oldp = old_coef_probs[i][j][k][l] + t;
+                int s;
+                int u = 0;
+
+                if (t == PIVOT_NODE)
+                  s = av1_prob_update_search_model_subframe(
+                      this_branch_ct, old_coef_probs[i][j][k][l], &newp, upd,
+                      stepsize, max_idx);
+                else
+                  s = av1_prob_update_search_subframe(this_branch_ct[t], *oldp,
+                                                      &newp, upd, max_idx);
+                if (s > 0 && newp != *oldp) u = 1;
+                updates += u;
+                if (u == 0 && updates == 0) {
+                  noupdates_before_first++;
+                  continue;
+                }
+                if (u == 1 && updates == 1) {
+                  int v;
+                  // first update
+                  aom_write_bit(bc, 1);
+                  for (v = 0; v < noupdates_before_first; ++v)
+                    aom_write(bc, 0, upd);
+                }
+                aom_write(bc, u, upd);
+                if (u) {
+                  /* send/use new probability */
+                  av1_write_prob_diff_update(bc, newp, *oldp);
+                  *oldp = newp;
+                }
+              }
+            }
+          }
+        }
+      }
+      if (updates == 0) {
+        aom_write_bit(bc, 0);  // no updates
+      }
+      return;
+    }
+    default: assert(0);
+  }
+}
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+#if !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET)
+static void update_coef_probs(AV1_COMP *cpi, aom_writer *w) {
+  const TX_MODE tx_mode = cpi->common.tx_mode;
+  const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+  TX_SIZE tx_size;
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  AV1_COMMON *cm = &cpi->common;
+  SUBFRAME_STATS *subframe_stats = &cpi->subframe_stats;
+  int i;
+  av1_coeff_probs_model dummy_frame_coef_probs[PLANE_TYPES];
+
+  if (cm->do_subframe_update &&
+      cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+    av1_copy(cpi->common.fc->coef_probs,
+             subframe_stats->enc_starting_coef_probs);
+    for (i = 0; i <= cpi->common.coef_probs_update_idx; ++i) {
+      get_coef_counts_diff(cpi, i, cpi->wholeframe_stats.coef_counts_buf[i],
+                           cpi->wholeframe_stats.eob_counts_buf[i]);
+    }
+  }
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+  for (tx_size = 0; tx_size <= max_tx_size; ++tx_size) {
+    av1_coeff_stats frame_branch_ct[PLANE_TYPES];
+    av1_coeff_probs_model frame_coef_probs[PLANE_TYPES];
+    if (cpi->td.counts->tx_size_totals[tx_size] <= 20 || CONFIG_RD_DEBUG ||
+        (tx_size >= TX_16X16 && cpi->sf.tx_size_search_method == USE_TX_8X8)) {
+      aom_write_bit(w, 0);
+    } else {
+#if CONFIG_SUBFRAME_PROB_UPDATE
+      if (cm->do_subframe_update &&
+          cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+        unsigned int this_eob_counts_copy[PLANE_TYPES][REF_TYPES][COEF_BANDS]
+                                         [COEFF_CONTEXTS];
+        av1_coeff_count coef_counts_copy[PLANE_TYPES];
+        av1_copy(this_eob_counts_copy, cpi->common.counts.eob_branch[tx_size]);
+        av1_copy(coef_counts_copy, cpi->td.rd_counts.coef_counts[tx_size]);
+        build_tree_distribution(cpi, tx_size, frame_branch_ct,
+                                frame_coef_probs);
+        for (i = 0; i <= cpi->common.coef_probs_update_idx; ++i) {
+          av1_copy(cpi->common.counts.eob_branch[tx_size],
+                   cpi->wholeframe_stats.eob_counts_buf[i][tx_size]);
+          av1_copy(cpi->td.rd_counts.coef_counts[tx_size],
+                   cpi->wholeframe_stats.coef_counts_buf[i][tx_size]);
+          build_tree_distribution(cpi, tx_size, cpi->branch_ct_buf[i][tx_size],
+                                  dummy_frame_coef_probs);
+        }
+        av1_copy(cpi->common.counts.eob_branch[tx_size], this_eob_counts_copy);
+        av1_copy(cpi->td.rd_counts.coef_counts[tx_size], coef_counts_copy);
+
+        update_coef_probs_subframe(w, cpi, tx_size, cpi->branch_ct_buf,
+                                   frame_coef_probs);
+      } else {
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+        build_tree_distribution(cpi, tx_size, frame_branch_ct,
+                                frame_coef_probs);
+        update_coef_probs_common(w, cpi, tx_size, frame_branch_ct,
+                                 frame_coef_probs);
+#if CONFIG_SUBFRAME_PROB_UPDATE
+      }
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+    }
+  }
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  av1_copy(cm->starting_coef_probs, cm->fc->coef_probs);
+  av1_copy(subframe_stats->coef_probs_buf[0], cm->fc->coef_probs);
+  if (cm->do_subframe_update &&
+      cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+    unsigned int eob_counts_copy[TX_SIZES][PLANE_TYPES][REF_TYPES][COEF_BANDS]
+                                [COEFF_CONTEXTS];
+    av1_copy(eob_counts_copy, cm->counts.eob_branch);
+    for (i = 1; i <= cpi->common.coef_probs_update_idx; ++i) {
+      for (tx_size = 0; tx_size <= max_tx_size; ++tx_size)
+        av1_full_to_model_counts(cm->counts.coef[tx_size],
+                                 subframe_stats->coef_counts_buf[i][tx_size]);
+      av1_copy(cm->counts.eob_branch, subframe_stats->eob_counts_buf[i]);
+      av1_partial_adapt_probs(cm, 0, 0);
+      av1_copy(subframe_stats->coef_probs_buf[i], cm->fc->coef_probs);
+    }
+    av1_copy(cm->fc->coef_probs, subframe_stats->coef_probs_buf[0]);
+    av1_copy(cm->counts.eob_branch, eob_counts_copy);
+  }
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+}
+#endif  // !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET)
+#endif  // !CONFIG_EC_ADAPT
+#endif  // !CONFIG_LV_MAP
+
+#if CONFIG_LOOP_RESTORATION
+static void encode_restoration_mode(AV1_COMMON *cm,
+                                    struct aom_write_bit_buffer *wb) {
+  int p;
+  RestorationInfo *rsi = &cm->rst_info[0];
+  switch (rsi->frame_restoration_type) {
+    case RESTORE_NONE:
+      aom_wb_write_bit(wb, 0);
+      aom_wb_write_bit(wb, 0);
+      break;
+    case RESTORE_WIENER:
+      aom_wb_write_bit(wb, 1);
+      aom_wb_write_bit(wb, 0);
+      break;
+    case RESTORE_SGRPROJ:
+      aom_wb_write_bit(wb, 1);
+      aom_wb_write_bit(wb, 1);
+      break;
+    case RESTORE_SWITCHABLE:
+      aom_wb_write_bit(wb, 0);
+      aom_wb_write_bit(wb, 1);
+      break;
+    default: assert(0);
+  }
+  for (p = 1; p < MAX_MB_PLANE; ++p) {
+    rsi = &cm->rst_info[p];
+    switch (rsi->frame_restoration_type) {
+      case RESTORE_NONE: aom_wb_write_bit(wb, 0); break;
+      case RESTORE_WIENER: aom_wb_write_bit(wb, 1); break;
+      default: assert(0);
+    }
+  }
+  if (cm->rst_info[0].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[1].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[2].frame_restoration_type != RESTORE_NONE) {
+    rsi = &cm->rst_info[0];
+    aom_wb_write_bit(wb, rsi->restoration_tilesize != RESTORATION_TILESIZE_MAX);
+    if (rsi->restoration_tilesize != RESTORATION_TILESIZE_MAX) {
+      aom_wb_write_bit(
+          wb, rsi->restoration_tilesize != (RESTORATION_TILESIZE_MAX >> 1));
+    }
+  }
+}
+
+static void write_wiener_filter(WienerInfo *wiener_info,
+                                WienerInfo *ref_wiener_info, aom_writer *wb) {
+  aom_write_primitive_refsubexpfin(
+      wb, WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
+      WIENER_FILT_TAP0_SUBEXP_K,
+      ref_wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV,
+      wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV);
+  aom_write_primitive_refsubexpfin(
+      wb, WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
+      WIENER_FILT_TAP1_SUBEXP_K,
+      ref_wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV,
+      wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV);
+  aom_write_primitive_refsubexpfin(
+      wb, WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
+      WIENER_FILT_TAP2_SUBEXP_K,
+      ref_wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV,
+      wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV);
+  aom_write_primitive_refsubexpfin(
+      wb, WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
+      WIENER_FILT_TAP0_SUBEXP_K,
+      ref_wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV,
+      wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV);
+  aom_write_primitive_refsubexpfin(
+      wb, WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
+      WIENER_FILT_TAP1_SUBEXP_K,
+      ref_wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV,
+      wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV);
+  aom_write_primitive_refsubexpfin(
+      wb, WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
+      WIENER_FILT_TAP2_SUBEXP_K,
+      ref_wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV,
+      wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV);
+  memcpy(ref_wiener_info, wiener_info, sizeof(*wiener_info));
+}
+
+static void write_sgrproj_filter(SgrprojInfo *sgrproj_info,
+                                 SgrprojInfo *ref_sgrproj_info,
+                                 aom_writer *wb) {
+  aom_write_literal(wb, sgrproj_info->ep, SGRPROJ_PARAMS_BITS);
+  aom_write_primitive_refsubexpfin(wb, SGRPROJ_PRJ_MAX0 - SGRPROJ_PRJ_MIN0 + 1,
+                                   SGRPROJ_PRJ_SUBEXP_K,
+                                   ref_sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0,
+                                   sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0);
+  aom_write_primitive_refsubexpfin(wb, SGRPROJ_PRJ_MAX1 - SGRPROJ_PRJ_MIN1 + 1,
+                                   SGRPROJ_PRJ_SUBEXP_K,
+                                   ref_sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1,
+                                   sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1);
+  memcpy(ref_sgrproj_info, sgrproj_info, sizeof(*sgrproj_info));
+}
+
+static void encode_restoration(AV1_COMMON *cm, aom_writer *wb) {
+  int i, p;
+  const int ntiles = av1_get_rest_ntiles(cm->width, cm->height,
+                                         cm->rst_info[0].restoration_tilesize,
+                                         NULL, NULL, NULL, NULL);
+  WienerInfo ref_wiener_info;
+  SgrprojInfo ref_sgrproj_info;
+  set_default_wiener(&ref_wiener_info);
+  set_default_sgrproj(&ref_sgrproj_info);
+  const int ntiles_uv = av1_get_rest_ntiles(
+      ROUND_POWER_OF_TWO(cm->width, cm->subsampling_x),
+      ROUND_POWER_OF_TWO(cm->height, cm->subsampling_y),
+      cm->rst_info[1].restoration_tilesize, NULL, NULL, NULL, NULL);
+  RestorationInfo *rsi = &cm->rst_info[0];
+  if (rsi->frame_restoration_type != RESTORE_NONE) {
+    if (rsi->frame_restoration_type == RESTORE_SWITCHABLE) {
+      // RESTORE_SWITCHABLE
+      for (i = 0; i < ntiles; ++i) {
+        av1_write_token(
+            wb, av1_switchable_restore_tree, cm->fc->switchable_restore_prob,
+            &switchable_restore_encodings[rsi->restoration_type[i]]);
+        if (rsi->restoration_type[i] == RESTORE_WIENER) {
+          write_wiener_filter(&rsi->wiener_info[i], &ref_wiener_info, wb);
+        } else if (rsi->restoration_type[i] == RESTORE_SGRPROJ) {
+          write_sgrproj_filter(&rsi->sgrproj_info[i], &ref_sgrproj_info, wb);
+        }
+      }
+    } else if (rsi->frame_restoration_type == RESTORE_WIENER) {
+      for (i = 0; i < ntiles; ++i) {
+        aom_write(wb, rsi->restoration_type[i] != RESTORE_NONE,
+                  RESTORE_NONE_WIENER_PROB);
+        if (rsi->restoration_type[i] != RESTORE_NONE) {
+          write_wiener_filter(&rsi->wiener_info[i], &ref_wiener_info, wb);
+        }
+      }
+    } else if (rsi->frame_restoration_type == RESTORE_SGRPROJ) {
+      for (i = 0; i < ntiles; ++i) {
+        aom_write(wb, rsi->restoration_type[i] != RESTORE_NONE,
+                  RESTORE_NONE_SGRPROJ_PROB);
+        if (rsi->restoration_type[i] != RESTORE_NONE) {
+          write_sgrproj_filter(&rsi->sgrproj_info[i], &ref_sgrproj_info, wb);
+        }
+      }
+    }
+  }
+  for (p = 1; p < MAX_MB_PLANE; ++p) {
+    set_default_wiener(&ref_wiener_info);
+    rsi = &cm->rst_info[p];
+    if (rsi->frame_restoration_type == RESTORE_WIENER) {
+      for (i = 0; i < ntiles_uv; ++i) {
+        if (ntiles_uv > 1)
+          aom_write(wb, rsi->restoration_type[i] != RESTORE_NONE,
+                    RESTORE_NONE_WIENER_PROB);
+        if (rsi->restoration_type[i] != RESTORE_NONE) {
+          write_wiener_filter(&rsi->wiener_info[i], &ref_wiener_info, wb);
+        }
+      }
+    } else if (rsi->frame_restoration_type != RESTORE_NONE) {
+      assert(0);
+    }
+  }
+}
+#endif  // CONFIG_LOOP_RESTORATION
+
+static void encode_loopfilter(AV1_COMMON *cm, struct aom_write_bit_buffer *wb) {
+  int i;
+  struct loopfilter *lf = &cm->lf;
+
+  // Encode the loop filter level and type
+  aom_wb_write_literal(wb, lf->filter_level, 6);
+  aom_wb_write_literal(wb, lf->sharpness_level, 3);
+
+  // Write out loop filter deltas applied at the MB level based on mode or
+  // ref frame (if they are enabled).
+  aom_wb_write_bit(wb, lf->mode_ref_delta_enabled);
+
+  if (lf->mode_ref_delta_enabled) {
+    aom_wb_write_bit(wb, lf->mode_ref_delta_update);
+    if (lf->mode_ref_delta_update) {
+      for (i = 0; i < TOTAL_REFS_PER_FRAME; i++) {
+        const int delta = lf->ref_deltas[i];
+        const int changed = delta != lf->last_ref_deltas[i];
+        aom_wb_write_bit(wb, changed);
+        if (changed) {
+          lf->last_ref_deltas[i] = delta;
+          aom_wb_write_inv_signed_literal(wb, delta, 6);
+        }
+      }
+
+      for (i = 0; i < MAX_MODE_LF_DELTAS; i++) {
+        const int delta = lf->mode_deltas[i];
+        const int changed = delta != lf->last_mode_deltas[i];
+        aom_wb_write_bit(wb, changed);
+        if (changed) {
+          lf->last_mode_deltas[i] = delta;
+          aom_wb_write_inv_signed_literal(wb, delta, 6);
+        }
+      }
+    }
+  }
+}
+
+#if CONFIG_CDEF
+static void encode_cdef(const AV1_COMMON *cm, struct aom_write_bit_buffer *wb) {
+  int i;
+  aom_wb_write_literal(wb, cm->cdef_dering_damping - 5, 1);
+  aom_wb_write_literal(wb, cm->cdef_clpf_damping - 3, 2);
+  aom_wb_write_literal(wb, cm->cdef_bits, 2);
+  for (i = 0; i < cm->nb_cdef_strengths; i++) {
+    aom_wb_write_literal(wb, cm->cdef_strengths[i], CDEF_STRENGTH_BITS);
+    aom_wb_write_literal(wb, cm->cdef_uv_strengths[i], CDEF_STRENGTH_BITS);
+  }
+}
+#endif
+
+static void write_delta_q(struct aom_write_bit_buffer *wb, int delta_q) {
+  if (delta_q != 0) {
+    aom_wb_write_bit(wb, 1);
+    aom_wb_write_inv_signed_literal(wb, delta_q, 6);
+  } else {
+    aom_wb_write_bit(wb, 0);
+  }
+}
+
+static void encode_quantization(const AV1_COMMON *const cm,
+                                struct aom_write_bit_buffer *wb) {
+  aom_wb_write_literal(wb, cm->base_qindex, QINDEX_BITS);
+  write_delta_q(wb, cm->y_dc_delta_q);
+  write_delta_q(wb, cm->uv_dc_delta_q);
+  write_delta_q(wb, cm->uv_ac_delta_q);
+#if CONFIG_AOM_QM
+  aom_wb_write_bit(wb, cm->using_qmatrix);
+  if (cm->using_qmatrix) {
+    aom_wb_write_literal(wb, cm->min_qmlevel, QM_LEVEL_BITS);
+    aom_wb_write_literal(wb, cm->max_qmlevel, QM_LEVEL_BITS);
+  }
+#endif
+}
+
+static void encode_segmentation(AV1_COMMON *cm, MACROBLOCKD *xd,
+                                struct aom_write_bit_buffer *wb) {
+  int i, j;
+  const struct segmentation *seg = &cm->seg;
+
+  aom_wb_write_bit(wb, seg->enabled);
+  if (!seg->enabled) return;
+
+  // Segmentation map
+  if (!frame_is_intra_only(cm) && !cm->error_resilient_mode) {
+    aom_wb_write_bit(wb, seg->update_map);
+  } else {
+    assert(seg->update_map == 1);
+  }
+  if (seg->update_map) {
+    // Select the coding strategy (temporal or spatial)
+    av1_choose_segmap_coding_method(cm, xd);
+
+    // Write out the chosen coding method.
+    if (!frame_is_intra_only(cm) && !cm->error_resilient_mode) {
+      aom_wb_write_bit(wb, seg->temporal_update);
+    } else {
+      assert(seg->temporal_update == 0);
+    }
+  }
+
+  // Segmentation data
+  aom_wb_write_bit(wb, seg->update_data);
+  if (seg->update_data) {
+    aom_wb_write_bit(wb, seg->abs_delta);
+
+    for (i = 0; i < MAX_SEGMENTS; i++) {
+      for (j = 0; j < SEG_LVL_MAX; j++) {
+        const int active = segfeature_active(seg, i, j);
+        aom_wb_write_bit(wb, active);
+        if (active) {
+          const int data = get_segdata(seg, i, j);
+          const int data_max = av1_seg_feature_data_max(j);
+
+          if (av1_is_segfeature_signed(j)) {
+            encode_unsigned_max(wb, abs(data), data_max);
+            aom_wb_write_bit(wb, data < 0);
+          } else {
+            encode_unsigned_max(wb, data, data_max);
+          }
+        }
+      }
+    }
+  }
+}
+
+#if !CONFIG_EC_ADAPT
+static void update_seg_probs(AV1_COMP *cpi, aom_writer *w) {
+  AV1_COMMON *cm = &cpi->common;
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+
+  if (!cm->seg.enabled || !cm->seg.update_map) return;
+
+  if (cm->seg.temporal_update) {
+    int i;
+
+    for (i = 0; i < PREDICTION_PROBS; i++)
+      av1_cond_prob_diff_update(w, &cm->fc->seg.pred_probs[i],
+                                cm->counts.seg.pred[i], probwt);
+
+    prob_diff_update(av1_segment_tree, cm->fc->seg.tree_probs,
+                     cm->counts.seg.tree_mispred, MAX_SEGMENTS, probwt, w);
+  } else {
+    prob_diff_update(av1_segment_tree, cm->fc->seg.tree_probs,
+                     cm->counts.seg.tree_total, MAX_SEGMENTS, probwt, w);
+  }
+}
+#endif
+
+static void write_tx_mode(AV1_COMMON *cm, MACROBLOCKD *xd, TX_MODE *mode,
+                          struct aom_write_bit_buffer *wb) {
+  int i, all_lossless = 1;
+
+  if (cm->seg.enabled) {
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      if (!xd->lossless[i]) {
+        all_lossless = 0;
+        break;
+      }
+    }
+  } else {
+    all_lossless = xd->lossless[0];
+  }
+  if (all_lossless) {
+    *mode = ONLY_4X4;
+    return;
+  }
+#if CONFIG_TX64X64
+  aom_wb_write_bit(wb, *mode == TX_MODE_SELECT);
+  if (*mode != TX_MODE_SELECT) {
+    aom_wb_write_literal(wb, AOMMIN(*mode, ALLOW_32X32), 2);
+    if (*mode >= ALLOW_32X32) aom_wb_write_bit(wb, *mode == ALLOW_64X64);
+  }
+#else
+  aom_wb_write_bit(wb, *mode == TX_MODE_SELECT);
+  if (*mode != TX_MODE_SELECT) aom_wb_write_literal(wb, *mode, 2);
+#endif  // CONFIG_TX64X64
+}
+
+#if !CONFIG_EC_ADAPT
+static void update_txfm_probs(AV1_COMMON *cm, aom_writer *w,
+                              FRAME_COUNTS *counts) {
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+  if (cm->tx_mode == TX_MODE_SELECT) {
+    int i, j;
+    for (i = 0; i < MAX_TX_DEPTH; ++i)
+      for (j = 0; j < TX_SIZE_CONTEXTS; ++j)
+        prob_diff_update(av1_tx_size_tree[i], cm->fc->tx_size_probs[i][j],
+                         counts->tx_size[i][j], i + 2, probwt, w);
+  }
+}
+#endif
+
+static void write_frame_interp_filter(InterpFilter filter,
+                                      struct aom_write_bit_buffer *wb) {
+  aom_wb_write_bit(wb, filter == SWITCHABLE);
+  if (filter != SWITCHABLE)
+    aom_wb_write_literal(wb, filter, LOG_SWITCHABLE_FILTERS);
+}
+
+static void fix_interp_filter(AV1_COMMON *cm, FRAME_COUNTS *counts) {
+  if (cm->interp_filter == SWITCHABLE) {
+    // Check to see if only one of the filters is actually used
+    int count[SWITCHABLE_FILTERS];
+    int i, j, c = 0;
+    for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
+      count[i] = 0;
+      for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
+        count[i] += counts->switchable_interp[j][i];
+      c += (count[i] > 0);
+    }
+    if (c == 1) {
+      // Only one filter is used. So set the filter at frame level
+      for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
+        if (count[i]) {
+#if CONFIG_MOTION_VAR && (CONFIG_WARPED_MOTION || CONFIG_GLOBAL_MOTION)
+#if CONFIG_WARPED_MOTION
+          if (i == EIGHTTAP_REGULAR || WARP_WM_NEIGHBORS_WITH_OBMC)
+#else
+          if (i == EIGHTTAP_REGULAR || WARP_GM_NEIGHBORS_WITH_OBMC)
+#endif  // CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR && (CONFIG_WARPED_MOTION || CONFIG_GLOBAL_MOTION)
+            cm->interp_filter = i;
+          break;
+        }
+      }
+    }
+  }
+}
+
+static void write_tile_info(const AV1_COMMON *const cm,
+                            struct aom_write_bit_buffer *wb) {
+#if CONFIG_EXT_TILE
+  const int tile_width =
+      ALIGN_POWER_OF_TWO(cm->tile_width, cm->mib_size_log2) >>
+      cm->mib_size_log2;
+  const int tile_height =
+      ALIGN_POWER_OF_TWO(cm->tile_height, cm->mib_size_log2) >>
+      cm->mib_size_log2;
+
+  assert(tile_width > 0);
+  assert(tile_height > 0);
+
+  aom_wb_write_literal(wb, cm->tile_encoding_mode, 1);
+
+// Write the tile sizes
+#if CONFIG_EXT_PARTITION
+  if (cm->sb_size == BLOCK_128X128) {
+    assert(tile_width <= 32);
+    assert(tile_height <= 32);
+    aom_wb_write_literal(wb, tile_width - 1, 5);
+    aom_wb_write_literal(wb, tile_height - 1, 5);
+  } else
+#endif  // CONFIG_EXT_PARTITION
+  {
+    assert(tile_width <= 64);
+    assert(tile_height <= 64);
+    aom_wb_write_literal(wb, tile_width - 1, 6);
+    aom_wb_write_literal(wb, tile_height - 1, 6);
+  }
+#else
+  int min_log2_tile_cols, max_log2_tile_cols, ones;
+  av1_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
+
+  // columns
+  ones = cm->log2_tile_cols - min_log2_tile_cols;
+  while (ones--) aom_wb_write_bit(wb, 1);
+
+  if (cm->log2_tile_cols < max_log2_tile_cols) aom_wb_write_bit(wb, 0);
+
+  // rows
+  aom_wb_write_bit(wb, cm->log2_tile_rows != 0);
+  if (cm->log2_tile_rows != 0) aom_wb_write_bit(wb, cm->log2_tile_rows != 1);
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_DEPENDENT_HORZTILES
+  if (cm->log2_tile_rows != 0) aom_wb_write_bit(wb, cm->dependent_horz_tiles);
+#endif
+
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  aom_wb_write_bit(wb, cm->loop_filter_across_tiles_enabled);
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+}
+
+static int get_refresh_mask(AV1_COMP *cpi) {
+  int refresh_mask = 0;
+
+#if CONFIG_EXT_REFS
+  // NOTE(zoeliu): When LAST_FRAME is to get refreshed, the decoder will be
+  // notified to get LAST3_FRAME refreshed and then the virtual indexes for all
+  // the 3 LAST reference frames will be updated accordingly, i.e.:
+  // (1) The original virtual index for LAST3_FRAME will become the new virtual
+  //     index for LAST_FRAME; and
+  // (2) The original virtual indexes for LAST_FRAME and LAST2_FRAME will be
+  //     shifted and become the new virtual indexes for LAST2_FRAME and
+  //     LAST3_FRAME.
+  refresh_mask |=
+      (cpi->refresh_last_frame << cpi->lst_fb_idxes[LAST_REF_FRAMES - 1]);
+  if (cpi->rc.is_bwd_ref_frame && cpi->num_extra_arfs) {
+    // We have swapped the virtual indices
+    refresh_mask |= (cpi->refresh_bwd_ref_frame << cpi->arf_map[0]);
+  } else {
+    refresh_mask |= (cpi->refresh_bwd_ref_frame << cpi->bwd_fb_idx);
+  }
+#else
+  refresh_mask |= (cpi->refresh_last_frame << cpi->lst_fb_idx);
+#endif  // CONFIG_EXT_REFS
+
+  if (av1_preserve_existing_gf(cpi)) {
+    // We have decided to preserve the previously existing golden frame as our
+    // new ARF frame. However, in the short term we leave it in the GF slot and,
+    // if we're updating the GF with the current decoded frame, we save it
+    // instead to the ARF slot.
+    // Later, in the function av1_encoder.c:av1_update_reference_frames() we
+    // will swap gld_fb_idx and alt_fb_idx to achieve our objective. We do it
+    // there so that it can be done outside of the recode loop.
+    // Note: This is highly specific to the use of ARF as a forward reference,
+    // and this needs to be generalized as other uses are implemented
+    // (like RTC/temporal scalability).
+    return refresh_mask | (cpi->refresh_golden_frame << cpi->alt_fb_idx);
+  } else {
+#if CONFIG_EXT_REFS
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    int arf_idx = cpi->arf_map[gf_group->arf_update_idx[gf_group->index]];
+#else
+    int arf_idx = cpi->alt_fb_idx;
+    if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
+      const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+      arf_idx = gf_group->arf_update_idx[gf_group->index];
+    }
+#endif  // CONFIG_EXT_REFS
+    return refresh_mask | (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
+           (cpi->refresh_alt_ref_frame << arf_idx);
+  }
+}
+
+#if CONFIG_EXT_TILE
+static INLINE int find_identical_tile(
+    const int tile_row, const int tile_col,
+    TileBufferEnc (*const tile_buffers)[1024]) {
+  const MV32 candidate_offset[1] = { { 1, 0 } };
+  const uint8_t *const cur_tile_data =
+      tile_buffers[tile_row][tile_col].data + 4;
+  const size_t cur_tile_size = tile_buffers[tile_row][tile_col].size;
+
+  int i;
+
+  if (tile_row == 0) return 0;
+
+  // (TODO: yunqingwang) For now, only above tile is checked and used.
+  // More candidates such as left tile can be added later.
+  for (i = 0; i < 1; i++) {
+    int row_offset = candidate_offset[0].row;
+    int col_offset = candidate_offset[0].col;
+    int row = tile_row - row_offset;
+    int col = tile_col - col_offset;
+    uint8_t tile_hdr;
+    const uint8_t *tile_data;
+    TileBufferEnc *candidate;
+
+    if (row < 0 || col < 0) continue;
+
+    tile_hdr = *(tile_buffers[row][col].data);
+
+    // Read out tcm bit
+    if ((tile_hdr >> 7) == 1) {
+      // The candidate is a copy tile itself
+      row_offset += tile_hdr & 0x7f;
+      row = tile_row - row_offset;
+    }
+
+    candidate = &tile_buffers[row][col];
+
+    if (row_offset >= 128 || candidate->size != cur_tile_size) continue;
+
+    tile_data = candidate->data + 4;
+
+    if (memcmp(tile_data, cur_tile_data, cur_tile_size) != 0) continue;
+
+    // Identical tile found
+    assert(row_offset > 0);
+    return row_offset;
+  }
+
+  // No identical tile found
+  return 0;
+}
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_TILE_GROUPS
+static uint32_t write_tiles(AV1_COMP *const cpi,
+                            struct aom_write_bit_buffer *wb,
+                            unsigned int *max_tile_size,
+                            unsigned int *max_tile_col_size) {
+#else
+static uint32_t write_tiles(AV1_COMP *const cpi, uint8_t *const dst,
+                            unsigned int *max_tile_size,
+                            unsigned int *max_tile_col_size) {
+#endif
+  const AV1_COMMON *const cm = &cpi->common;
+#if CONFIG_ANS
+  struct BufAnsCoder *buf_ans = &cpi->buf_ans;
+#else
+  aom_writer mode_bc;
+#endif  // CONFIG_ANS
+  int tile_row, tile_col;
+  TOKENEXTRA *(*const tok_buffers)[MAX_TILE_COLS] = cpi->tile_tok;
+  TileBufferEnc(*const tile_buffers)[MAX_TILE_COLS] = cpi->tile_buffers;
+  uint32_t total_size = 0;
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  unsigned int tile_size = 0;
+#if CONFIG_TILE_GROUPS
+  const int n_log2_tiles = cm->log2_tile_rows + cm->log2_tile_cols;
+  const int have_tiles = n_log2_tiles > 0;
+  uint32_t comp_hdr_size;
+  // Fixed size tile groups for the moment
+  const int num_tg_hdrs = cm->num_tg;
+  const int tg_size = (tile_rows * tile_cols + num_tg_hdrs - 1) / num_tg_hdrs;
+  int tile_count = 0;
+  int tg_count = 1;
+  int tile_size_bytes = 4;
+  int tile_col_size_bytes;
+  uint32_t uncompressed_hdr_size = 0;
+  uint8_t *dst = NULL;
+  struct aom_write_bit_buffer comp_hdr_len_wb;
+  struct aom_write_bit_buffer tg_params_wb;
+  struct aom_write_bit_buffer tile_size_bytes_wb;
+  uint32_t saved_offset;
+  int mtu_size = cpi->oxcf.mtu;
+  int curr_tg_data_size = 0;
+  int hdr_size;
+#endif
+#if CONFIG_EXT_TILE
+  const int have_tiles = tile_cols * tile_rows > 1;
+#endif  // CONFIG_EXT_TILE
+
+  *max_tile_size = 0;
+  *max_tile_col_size = 0;
+
+// All tile size fields are output on 4 bytes. A call to remux_tiles will
+// later compact the data if smaller headers are adequate.
+
+#if CONFIG_EXT_TILE
+  for (tile_col = 0; tile_col < tile_cols; tile_col++) {
+    TileInfo tile_info;
+    const int is_last_col = (tile_col == tile_cols - 1);
+    const uint32_t col_offset = total_size;
+
+    av1_tile_set_col(&tile_info, cm, tile_col);
+
+    // The last column does not have a column header
+    if (!is_last_col) total_size += 4;
+
+    for (tile_row = 0; tile_row < tile_rows; tile_row++) {
+      TileBufferEnc *const buf = &tile_buffers[tile_row][tile_col];
+      const TOKENEXTRA *tok = tok_buffers[tile_row][tile_col];
+      const TOKENEXTRA *tok_end = tok + cpi->tok_count[tile_row][tile_col];
+      const int data_offset = have_tiles ? 4 : 0;
+#if CONFIG_EC_ADAPT
+      const int tile_idx = tile_row * tile_cols + tile_col;
+      TileDataEnc *this_tile = &cpi->tile_data[tile_idx];
+#endif
+      av1_tile_set_row(&tile_info, cm, tile_row);
+
+      buf->data = dst + total_size;
+
+      // Is CONFIG_EXT_TILE = 1, every tile in the row has a header,
+      // even for the last one, unless no tiling is used at all.
+      total_size += data_offset;
+#if CONFIG_EC_ADAPT
+      // Initialise tile context from the frame context
+      this_tile->tctx = *cm->fc;
+      cpi->td.mb.e_mbd.tile_ctx = &this_tile->tctx;
+#endif
+#if CONFIG_PVQ
+      cpi->td.mb.pvq_q = &this_tile->pvq_q;
+      cpi->td.mb.daala_enc.state.adapt = &this_tile->tctx.pvq_context;
+#endif  // CONFIG_PVQ
+#if !CONFIG_ANS
+      aom_start_encode(&mode_bc, buf->data + data_offset);
+      write_modes(cpi, &tile_info, &mode_bc, &tok, tok_end);
+      assert(tok == tok_end);
+      aom_stop_encode(&mode_bc);
+      tile_size = mode_bc.pos;
+#else
+      buf_ans_write_init(buf_ans, buf->data + data_offset);
+      write_modes(cpi, &tile_info, buf_ans, &tok, tok_end);
+      assert(tok == tok_end);
+      aom_buf_ans_flush(buf_ans);
+      tile_size = buf_ans_write_end(buf_ans);
+#endif  // !CONFIG_ANS
+#if CONFIG_PVQ
+      cpi->td.mb.pvq_q = NULL;
+#endif
+      buf->size = tile_size;
+
+      // Record the maximum tile size we see, so we can compact headers later.
+      *max_tile_size = AOMMAX(*max_tile_size, tile_size);
+
+      if (have_tiles) {
+        // tile header: size of this tile, or copy offset
+        uint32_t tile_header = tile_size;
+
+        // If the tile_encoding_mode is 1 (i.e. TILE_VR), check if this tile is
+        // a copy tile.
+        // Very low chances to have copy tiles on the key frames, so don't
+        // search on key frames to reduce unnecessary search.
+        if (cm->frame_type != KEY_FRAME && cm->tile_encoding_mode) {
+          const int idendical_tile_offset =
+              find_identical_tile(tile_row, tile_col, tile_buffers);
+
+          if (idendical_tile_offset > 0) {
+            tile_size = 0;
+            tile_header = idendical_tile_offset | 0x80;
+            tile_header <<= 24;
+          }
+        }
+
+        mem_put_le32(buf->data, tile_header);
+      }
+
+      total_size += tile_size;
+    }
+
+    if (!is_last_col) {
+      uint32_t col_size = total_size - col_offset - 4;
+      mem_put_le32(dst + col_offset, col_size);
+
+      // If it is not final packing, record the maximum tile column size we see,
+      // otherwise, check if the tile size is out of the range.
+      *max_tile_col_size = AOMMAX(*max_tile_col_size, col_size);
+    }
+  }
+#else
+#if CONFIG_TILE_GROUPS
+  write_uncompressed_header(cpi, wb);
+
+#if CONFIG_EXT_REFS
+  if (cm->show_existing_frame) {
+    total_size = aom_wb_bytes_written(wb);
+    return (uint32_t)total_size;
+  }
+#endif  // CONFIG_EXT_REFS
+
+  // Write the tile length code
+  tile_size_bytes_wb = *wb;
+  aom_wb_write_literal(wb, 3, 2);
+
+  /* Write a placeholder for the number of tiles in each tile group */
+  tg_params_wb = *wb;
+  saved_offset = wb->bit_offset;
+  if (have_tiles) {
+    aom_wb_overwrite_literal(wb, 3, n_log2_tiles);
+    aom_wb_overwrite_literal(wb, (1 << n_log2_tiles) - 1, n_log2_tiles);
+  }
+
+  /* Write a placeholder for the compressed header length */
+  comp_hdr_len_wb = *wb;
+  aom_wb_write_literal(wb, 0, 16);
+
+  uncompressed_hdr_size = aom_wb_bytes_written(wb);
+  dst = wb->bit_buffer;
+  comp_hdr_size = write_compressed_header(cpi, dst + uncompressed_hdr_size);
+  aom_wb_overwrite_literal(&comp_hdr_len_wb, (int)(comp_hdr_size), 16);
+  hdr_size = uncompressed_hdr_size + comp_hdr_size;
+  total_size += hdr_size;
+#endif
+
+  for (tile_row = 0; tile_row < tile_rows; tile_row++) {
+    TileInfo tile_info;
+    const int is_last_row = (tile_row == tile_rows - 1);
+    av1_tile_set_row(&tile_info, cm, tile_row);
+
+    for (tile_col = 0; tile_col < tile_cols; tile_col++) {
+      const int tile_idx = tile_row * tile_cols + tile_col;
+      TileBufferEnc *const buf = &tile_buffers[tile_row][tile_col];
+#if CONFIG_PVQ || CONFIG_EC_ADAPT
+      TileDataEnc *this_tile = &cpi->tile_data[tile_idx];
+#endif
+      const TOKENEXTRA *tok = tok_buffers[tile_row][tile_col];
+      const TOKENEXTRA *tok_end = tok + cpi->tok_count[tile_row][tile_col];
+      const int is_last_col = (tile_col == tile_cols - 1);
+      const int is_last_tile = is_last_col && is_last_row;
+#if !CONFIG_TILE_GROUPS
+      (void)tile_idx;
+#else
+
+      if ((!mtu_size && tile_count > tg_size) ||
+          (mtu_size && tile_count && curr_tg_data_size >= mtu_size)) {
+        // New tile group
+        tg_count++;
+        // We've exceeded the packet size
+        if (tile_count > 1) {
+          /* The last tile exceeded the packet size. The tile group size
+             should therefore be tile_count-1.
+             Move the last tile and insert headers before it
+           */
+          uint32_t old_total_size = total_size - tile_size - 4;
+          memmove(dst + old_total_size + hdr_size, dst + old_total_size,
+                  (tile_size + 4) * sizeof(uint8_t));
+          // Copy uncompressed header
+          memmove(dst + old_total_size, dst,
+                  uncompressed_hdr_size * sizeof(uint8_t));
+          // Write the number of tiles in the group into the last uncompressed
+          // header before the one we've just inserted
+          aom_wb_overwrite_literal(&tg_params_wb, tile_idx - tile_count,
+                                   n_log2_tiles);
+          aom_wb_overwrite_literal(&tg_params_wb, tile_count - 2, n_log2_tiles);
+          // Update the pointer to the last TG params
+          tg_params_wb.bit_offset = saved_offset + 8 * old_total_size;
+          // Copy compressed header
+          memmove(dst + old_total_size + uncompressed_hdr_size,
+                  dst + uncompressed_hdr_size, comp_hdr_size * sizeof(uint8_t));
+          total_size += hdr_size;
+          tile_count = 1;
+          curr_tg_data_size = hdr_size + tile_size + 4;
+
+        } else {
+          // We exceeded the packet size in just one tile
+          // Copy uncompressed header
+          memmove(dst + total_size, dst,
+                  uncompressed_hdr_size * sizeof(uint8_t));
+          // Write the number of tiles in the group into the last uncompressed
+          // header
+          aom_wb_overwrite_literal(&tg_params_wb, tile_idx - tile_count,
+                                   n_log2_tiles);
+          aom_wb_overwrite_literal(&tg_params_wb, tile_count - 1, n_log2_tiles);
+          tg_params_wb.bit_offset = saved_offset + 8 * total_size;
+          // Copy compressed header
+          memmove(dst + total_size + uncompressed_hdr_size,
+                  dst + uncompressed_hdr_size, comp_hdr_size * sizeof(uint8_t));
+          total_size += hdr_size;
+          tile_count = 0;
+          curr_tg_data_size = hdr_size;
+        }
+      }
+      tile_count++;
+#endif
+      av1_tile_set_col(&tile_info, cm, tile_col);
+
+#if CONFIG_DEPENDENT_HORZTILES && CONFIG_TILE_GROUPS
+      av1_tile_set_tg_boundary(&tile_info, cm, tile_row, tile_col);
+#endif
+      buf->data = dst + total_size;
+
+      // The last tile does not have a header.
+      if (!is_last_tile) total_size += 4;
+
+#if CONFIG_EC_ADAPT
+      // Initialise tile context from the frame context
+      this_tile->tctx = *cm->fc;
+      cpi->td.mb.e_mbd.tile_ctx = &this_tile->tctx;
+#endif
+#if CONFIG_PVQ
+      cpi->td.mb.pvq_q = &this_tile->pvq_q;
+      cpi->td.mb.daala_enc.state.adapt = &this_tile->tctx.pvq_context;
+#endif  // CONFIG_PVQ
+#if CONFIG_ANS
+      buf_ans_write_init(buf_ans, dst + total_size);
+      write_modes(cpi, &tile_info, buf_ans, &tok, tok_end);
+      assert(tok == tok_end);
+      aom_buf_ans_flush(buf_ans);
+      tile_size = buf_ans_write_end(buf_ans);
+#else
+      aom_start_encode(&mode_bc, dst + total_size);
+      write_modes(cpi, &tile_info, &mode_bc, &tok, tok_end);
+#if !CONFIG_LV_MAP
+      assert(tok == tok_end);
+#endif  // !CONFIG_LV_MAP
+      aom_stop_encode(&mode_bc);
+      tile_size = mode_bc.pos;
+#endif  // CONFIG_ANS
+#if CONFIG_PVQ
+      cpi->td.mb.pvq_q = NULL;
+#endif
+
+      assert(tile_size > 0);
+
+#if CONFIG_TILE_GROUPS
+      curr_tg_data_size += tile_size + 4;
+#endif
+      buf->size = tile_size;
+
+      if (!is_last_tile) {
+        *max_tile_size = AOMMAX(*max_tile_size, tile_size);
+        // size of this tile
+        mem_put_le32(buf->data, tile_size);
+      }
+
+      total_size += tile_size;
+    }
+  }
+#if CONFIG_TILE_GROUPS
+  // Write the final tile group size
+  if (n_log2_tiles) {
+    aom_wb_overwrite_literal(&tg_params_wb, (1 << n_log2_tiles) - tile_count,
+                             n_log2_tiles);
+    aom_wb_overwrite_literal(&tg_params_wb, tile_count - 1, n_log2_tiles);
+  }
+  // Remux if possible. TODO (Thomas Davies): do this for more than one tile
+  // group
+  if (have_tiles && tg_count == 1) {
+    int data_size = total_size - (uncompressed_hdr_size + comp_hdr_size);
+    data_size = remux_tiles(cm, dst + uncompressed_hdr_size + comp_hdr_size,
+                            data_size, *max_tile_size, *max_tile_col_size,
+                            &tile_size_bytes, &tile_col_size_bytes);
+    total_size = data_size + uncompressed_hdr_size + comp_hdr_size;
+    aom_wb_overwrite_literal(&tile_size_bytes_wb, tile_size_bytes - 1, 2);
+  }
+
+#endif
+#endif  // CONFIG_EXT_TILE
+  return (uint32_t)total_size;
+}
+
+static void write_render_size(const AV1_COMMON *cm,
+                              struct aom_write_bit_buffer *wb) {
+  const int scaling_active =
+      cm->width != cm->render_width || cm->height != cm->render_height;
+  aom_wb_write_bit(wb, scaling_active);
+  if (scaling_active) {
+    aom_wb_write_literal(wb, cm->render_width - 1, 16);
+    aom_wb_write_literal(wb, cm->render_height - 1, 16);
+  }
+}
+
+#if CONFIG_FRAME_SUPERRES
+static void write_superres_scale(const AV1_COMMON *const cm,
+                                 struct aom_write_bit_buffer *wb) {
+  // This scaling and frame superres are probably incompatible
+  assert(cm->width == cm->render_width && cm->height == cm->render_height);
+
+  // First bit is whether to to scale or not
+  if (cm->superres_scale_numerator == SUPERRES_SCALE_DENOMINATOR) {
+    aom_wb_write_bit(wb, 0);  // no scaling
+  } else {
+    aom_wb_write_bit(wb, 1);  // scaling, write scale factor
+    // TODO(afergs): write factor to the compressed header instead
+    aom_wb_write_literal(
+        wb, cm->superres_scale_numerator - SUPERRES_SCALE_NUMERATOR_MIN,
+        SUPERRES_SCALE_BITS);
+  }
+}
+#endif  // CONFIG_FRAME_SUPERRES
+
+static void write_frame_size(const AV1_COMMON *cm,
+                             struct aom_write_bit_buffer *wb) {
+#if CONFIG_FRAME_SUPERRES
+  // If SUPERRES scaling is happening, write the full resolution instead of the
+  // downscaled resolution. The decoder will reduce this resolution itself.
+  if (cm->superres_scale_numerator != SUPERRES_SCALE_DENOMINATOR) {
+    aom_wb_write_literal(wb, cm->superres_width - 1, 16);
+    aom_wb_write_literal(wb, cm->superres_height - 1, 16);
+  } else {
+#endif  // CONFIG_FRAME_SUPERRES
+    aom_wb_write_literal(wb, cm->width - 1, 16);
+    aom_wb_write_literal(wb, cm->height - 1, 16);
+#if CONFIG_FRAME_SUPERRES
+  }
+#endif  // CONFIG_FRAME_SUPERRES
+
+  // TODO(afergs): Also write something different to render_size?
+  //               When superres scales, they'll be almost guaranteed to be
+  //               different on the other side.
+  write_render_size(cm, wb);
+#if CONFIG_FRAME_SUPERRES
+  write_superres_scale(cm, wb);
+#endif  // CONFIG_FRAME_SUPERRES
+}
+
+static void write_frame_size_with_refs(AV1_COMP *cpi,
+                                       struct aom_write_bit_buffer *wb) {
+  AV1_COMMON *const cm = &cpi->common;
+  int found = 0;
+
+  MV_REFERENCE_FRAME ref_frame;
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, ref_frame);
+
+    if (cfg != NULL) {
+      found =
+          cm->width == cfg->y_crop_width && cm->height == cfg->y_crop_height;
+      found &= cm->render_width == cfg->render_width &&
+               cm->render_height == cfg->render_height;
+    }
+    aom_wb_write_bit(wb, found);
+    if (found) {
+      break;
+    }
+  }
+
+  if (!found) {
+    write_frame_size(cm, wb);
+  }
+}
+
+static void write_sync_code(struct aom_write_bit_buffer *wb) {
+  aom_wb_write_literal(wb, AV1_SYNC_CODE_0, 8);
+  aom_wb_write_literal(wb, AV1_SYNC_CODE_1, 8);
+  aom_wb_write_literal(wb, AV1_SYNC_CODE_2, 8);
+}
+
+static void write_profile(BITSTREAM_PROFILE profile,
+                          struct aom_write_bit_buffer *wb) {
+  switch (profile) {
+    case PROFILE_0: aom_wb_write_literal(wb, 0, 2); break;
+    case PROFILE_1: aom_wb_write_literal(wb, 2, 2); break;
+    case PROFILE_2: aom_wb_write_literal(wb, 1, 2); break;
+    case PROFILE_3: aom_wb_write_literal(wb, 6, 3); break;
+    default: assert(0);
+  }
+}
+
+static void write_bitdepth_colorspace_sampling(
+    AV1_COMMON *const cm, struct aom_write_bit_buffer *wb) {
+  if (cm->profile >= PROFILE_2) {
+    assert(cm->bit_depth > AOM_BITS_8);
+    aom_wb_write_bit(wb, cm->bit_depth == AOM_BITS_10 ? 0 : 1);
+  }
+  aom_wb_write_literal(wb, cm->color_space, 3);
+  if (cm->color_space != AOM_CS_SRGB) {
+    // 0: [16, 235] (i.e. xvYCC), 1: [0, 255]
+    aom_wb_write_bit(wb, cm->color_range);
+    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
+      assert(cm->subsampling_x != 1 || cm->subsampling_y != 1);
+      aom_wb_write_bit(wb, cm->subsampling_x);
+      aom_wb_write_bit(wb, cm->subsampling_y);
+      aom_wb_write_bit(wb, 0);  // unused
+    } else {
+      assert(cm->subsampling_x == 1 && cm->subsampling_y == 1);
+    }
+  } else {
+    assert(cm->profile == PROFILE_1 || cm->profile == PROFILE_3);
+    aom_wb_write_bit(wb, 0);  // unused
+  }
+}
+
+#if CONFIG_REFERENCE_BUFFER
+void write_sequence_header(SequenceHeader *seq_params) {
+  /* Placeholder for actually writing to the bitstream */
+  seq_params->frame_id_numbers_present_flag = FRAME_ID_NUMBERS_PRESENT_FLAG;
+  seq_params->frame_id_length_minus7 = FRAME_ID_LENGTH_MINUS7;
+  seq_params->delta_frame_id_length_minus2 = DELTA_FRAME_ID_LENGTH_MINUS2;
+}
+#endif
+
+static void write_uncompressed_header(AV1_COMP *cpi,
+                                      struct aom_write_bit_buffer *wb) {
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+
+#if CONFIG_REFERENCE_BUFFER
+  /* TODO: Move outside frame loop or inside key-frame branch */
+  write_sequence_header(&cpi->seq_params);
+#endif
+
+  aom_wb_write_literal(wb, AOM_FRAME_MARKER, 2);
+
+  write_profile(cm->profile, wb);
+
+#if CONFIG_EXT_REFS
+  // NOTE: By default all coded frames to be used as a reference
+  cm->is_reference_frame = 1;
+
+  if (cm->show_existing_frame) {
+    RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+    const int frame_to_show = cm->ref_frame_map[cpi->existing_fb_idx_to_show];
+
+    if (frame_to_show < 0 || frame_bufs[frame_to_show].ref_count < 1) {
+      aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
+                         "Buffer %d does not contain a reconstructed frame",
+                         frame_to_show);
+    }
+    ref_cnt_fb(frame_bufs, &cm->new_fb_idx, frame_to_show);
+
+    aom_wb_write_bit(wb, 1);  // show_existing_frame
+    aom_wb_write_literal(wb, cpi->existing_fb_idx_to_show, 3);
+
+#if CONFIG_REFERENCE_BUFFER
+    if (cpi->seq_params.frame_id_numbers_present_flag) {
+      int frame_id_len = cpi->seq_params.frame_id_length_minus7 + 7;
+      int display_frame_id = cm->ref_frame_id[cpi->existing_fb_idx_to_show];
+      aom_wb_write_literal(wb, display_frame_id, frame_id_len);
+      /* Add a zero byte to prevent emulation of superframe marker */
+      /* Same logic as when when terminating the entropy coder */
+      /* Consider to have this logic only one place */
+      aom_wb_write_literal(wb, 0, 8);
+    }
+#endif
+
+    return;
+  } else {
+#endif                        // CONFIG_EXT_REFS
+    aom_wb_write_bit(wb, 0);  // show_existing_frame
+#if CONFIG_EXT_REFS
+  }
+#endif  // CONFIG_EXT_REFS
+
+  aom_wb_write_bit(wb, cm->frame_type);
+  aom_wb_write_bit(wb, cm->show_frame);
+  aom_wb_write_bit(wb, cm->error_resilient_mode);
+
+#if CONFIG_REFERENCE_BUFFER
+  cm->invalid_delta_frame_id_minus1 = 0;
+  if (cpi->seq_params.frame_id_numbers_present_flag) {
+    int frame_id_len = cpi->seq_params.frame_id_length_minus7 + 7;
+    aom_wb_write_literal(wb, cm->current_frame_id, frame_id_len);
+  }
+#endif
+
+#if CONFIG_FRAME_SUPERRES
+  // TODO(afergs): Remove - this is just to stop superres from breaking
+  cm->superres_scale_numerator = SUPERRES_SCALE_DENOMINATOR;
+#endif  // CONFIG_FRAME_SUPERRES
+
+  if (cm->frame_type == KEY_FRAME) {
+    write_sync_code(wb);
+    write_bitdepth_colorspace_sampling(cm, wb);
+    write_frame_size(cm, wb);
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+    assert(cpi->common.ans_window_size_log2 >= 8);
+    assert(cpi->common.ans_window_size_log2 < 24);
+    aom_wb_write_literal(wb, cpi->common.ans_window_size_log2 - 8, 4);
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+#if CONFIG_PALETTE
+    aom_wb_write_bit(wb, cm->allow_screen_content_tools);
+#endif  // CONFIG_PALETTE
+  } else {
+    if (!cm->show_frame) aom_wb_write_bit(wb, cm->intra_only);
+#if CONFIG_PALETTE
+    if (cm->intra_only) aom_wb_write_bit(wb, cm->allow_screen_content_tools);
+#endif  // CONFIG_PALETTE
+    if (!cm->error_resilient_mode) {
+      if (cm->intra_only) {
+        aom_wb_write_bit(wb,
+                         cm->reset_frame_context == RESET_FRAME_CONTEXT_ALL);
+      } else {
+        aom_wb_write_bit(wb,
+                         cm->reset_frame_context != RESET_FRAME_CONTEXT_NONE);
+        if (cm->reset_frame_context != RESET_FRAME_CONTEXT_NONE)
+          aom_wb_write_bit(wb,
+                           cm->reset_frame_context == RESET_FRAME_CONTEXT_ALL);
+      }
+    }
+
+#if CONFIG_EXT_REFS
+    cpi->refresh_frame_mask = get_refresh_mask(cpi);
+#endif  // CONFIG_EXT_REFS
+
+    if (cm->intra_only) {
+      write_sync_code(wb);
+      write_bitdepth_colorspace_sampling(cm, wb);
+
+#if CONFIG_EXT_REFS
+      aom_wb_write_literal(wb, cpi->refresh_frame_mask, REF_FRAMES);
+#else
+      aom_wb_write_literal(wb, get_refresh_mask(cpi), REF_FRAMES);
+#endif  // CONFIG_EXT_REFS
+      write_frame_size(cm, wb);
+
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+      assert(cpi->common.ans_window_size_log2 >= 8);
+      assert(cpi->common.ans_window_size_log2 < 24);
+      aom_wb_write_literal(wb, cpi->common.ans_window_size_log2 - 8, 4);
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+    } else {
+      MV_REFERENCE_FRAME ref_frame;
+
+#if CONFIG_EXT_REFS
+      aom_wb_write_literal(wb, cpi->refresh_frame_mask, REF_FRAMES);
+#else
+      aom_wb_write_literal(wb, get_refresh_mask(cpi), REF_FRAMES);
+#endif  // CONFIG_EXT_REFS
+
+#if CONFIG_EXT_REFS
+      if (!cpi->refresh_frame_mask) {
+        // NOTE: "cpi->refresh_frame_mask == 0" indicates that the coded frame
+        //       will not be used as a reference
+        cm->is_reference_frame = 0;
+      }
+#endif  // CONFIG_EXT_REFS
+
+      for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+        assert(get_ref_frame_map_idx(cpi, ref_frame) != INVALID_IDX);
+        aom_wb_write_literal(wb, get_ref_frame_map_idx(cpi, ref_frame),
+                             REF_FRAMES_LOG2);
+        aom_wb_write_bit(wb, cm->ref_frame_sign_bias[ref_frame]);
+#if CONFIG_REFERENCE_BUFFER
+        if (cpi->seq_params.frame_id_numbers_present_flag) {
+          int i = get_ref_frame_map_idx(cpi, ref_frame);
+          int frame_id_len = cpi->seq_params.frame_id_length_minus7 + 7;
+          int diff_len = cpi->seq_params.delta_frame_id_length_minus2 + 2;
+          int delta_frame_id_minus1 =
+              ((cm->current_frame_id - cm->ref_frame_id[i] +
+                (1 << frame_id_len)) %
+               (1 << frame_id_len)) -
+              1;
+          if (delta_frame_id_minus1 < 0 ||
+              delta_frame_id_minus1 >= (1 << diff_len))
+            cm->invalid_delta_frame_id_minus1 = 1;
+          aom_wb_write_literal(wb, delta_frame_id_minus1, diff_len);
+        }
+#endif
+      }
+
+#if CONFIG_FRAME_SIZE
+      if (cm->error_resilient_mode == 0) {
+        write_frame_size_with_refs(cpi, wb);
+      } else {
+        write_frame_size(cm, wb);
+      }
+#else
+      write_frame_size_with_refs(cpi, wb);
+#endif
+
+      aom_wb_write_bit(wb, cm->allow_high_precision_mv);
+
+      fix_interp_filter(cm, cpi->td.counts);
+      write_frame_interp_filter(cm->interp_filter, wb);
+#if CONFIG_TEMPMV_SIGNALING
+      if (!cm->error_resilient_mode) {
+        aom_wb_write_bit(wb, cm->use_prev_frame_mvs);
+      }
+#endif
+    }
+  }
+
+#if CONFIG_REFERENCE_BUFFER
+  cm->refresh_mask = cm->frame_type == KEY_FRAME ? 0xFF : get_refresh_mask(cpi);
+#endif
+
+  if (!cm->error_resilient_mode) {
+    aom_wb_write_bit(
+        wb, cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_FORWARD);
+  }
+
+  aom_wb_write_literal(wb, cm->frame_context_idx, FRAME_CONTEXTS_LOG2);
+
+  assert(cm->mib_size == mi_size_wide[cm->sb_size]);
+  assert(cm->mib_size == 1 << cm->mib_size_log2);
+#if CONFIG_EXT_PARTITION
+  assert(cm->sb_size == BLOCK_128X128 || cm->sb_size == BLOCK_64X64);
+  aom_wb_write_bit(wb, cm->sb_size == BLOCK_128X128 ? 1 : 0);
+#else
+  assert(cm->sb_size == BLOCK_64X64);
+#endif  // CONFIG_EXT_PARTITION
+
+  encode_loopfilter(cm, wb);
+#if CONFIG_CDEF
+  encode_cdef(cm, wb);
+#endif
+#if CONFIG_LOOP_RESTORATION
+  encode_restoration_mode(cm, wb);
+#endif  // CONFIG_LOOP_RESTORATION
+  encode_quantization(cm, wb);
+  encode_segmentation(cm, xd, wb);
+#if CONFIG_DELTA_Q
+  {
+    int i;
+    struct segmentation *const seg = &cm->seg;
+    int segment_quantizer_active = 0;
+    for (i = 0; i < MAX_SEGMENTS; i++) {
+      if (segfeature_active(seg, i, SEG_LVL_ALT_Q)) {
+        segment_quantizer_active = 1;
+      }
+    }
+
+    if (cm->delta_q_present_flag)
+      assert(segment_quantizer_active == 0 && cm->base_qindex > 0);
+    if (segment_quantizer_active == 0 && cm->base_qindex > 0) {
+      aom_wb_write_bit(wb, cm->delta_q_present_flag);
+      if (cm->delta_q_present_flag) {
+        aom_wb_write_literal(wb, OD_ILOG_NZ(cm->delta_q_res) - 1, 2);
+        xd->prev_qindex = cm->base_qindex;
+#if CONFIG_EXT_DELTA_Q
+        assert(seg->abs_delta == SEGMENT_DELTADATA);
+        aom_wb_write_bit(wb, cm->delta_lf_present_flag);
+        if (cm->delta_lf_present_flag) {
+          aom_wb_write_literal(wb, OD_ILOG_NZ(cm->delta_lf_res) - 1, 2);
+          xd->prev_delta_lf_from_base = 0;
+        }
+#endif  // CONFIG_EXT_DELTA_Q
+      }
+    }
+  }
+#endif
+
+  write_tx_mode(cm, xd, &cm->tx_mode, wb);
+
+  if (cpi->allow_comp_inter_inter) {
+    const int use_hybrid_pred = cm->reference_mode == REFERENCE_MODE_SELECT;
+#if !CONFIG_REF_ADAPT
+    const int use_compound_pred = cm->reference_mode != SINGLE_REFERENCE;
+#endif  // !CONFIG_REF_ADAPT
+
+    aom_wb_write_bit(wb, use_hybrid_pred);
+#if !CONFIG_REF_ADAPT
+    if (!use_hybrid_pred) aom_wb_write_bit(wb, use_compound_pred);
+#endif  // !CONFIG_REF_ADAPT
+  }
+
+#if CONFIG_EXT_TX
+  aom_wb_write_bit(wb, cm->reduced_tx_set_used);
+#endif  // CONFIG_EXT_TX
+
+  write_tile_info(cm, wb);
+}
+
+#if CONFIG_GLOBAL_MOTION
+static void write_global_motion_params(WarpedMotionParams *params,
+                                       WarpedMotionParams *ref_params,
+                                       aom_prob *probs, aom_writer *w,
+                                       int allow_hp) {
+  TransformationType type = params->wmtype;
+  int trans_bits;
+  int trans_prec_diff;
+  av1_write_token(w, av1_global_motion_types_tree, probs,
+                  &global_motion_types_encodings[type]);
+  switch (type) {
+    case HOMOGRAPHY:
+    case HORTRAPEZOID:
+    case VERTRAPEZOID:
+      if (type != HORTRAPEZOID)
+        aom_write_signed_primitive_refsubexpfin(
+            w, GM_ROW3HOMO_MAX + 1, SUBEXPFIN_K,
+            (ref_params->wmmat[6] >> GM_ROW3HOMO_PREC_DIFF),
+            (params->wmmat[6] >> GM_ROW3HOMO_PREC_DIFF));
+      if (type != VERTRAPEZOID)
+        aom_write_signed_primitive_refsubexpfin(
+            w, GM_ROW3HOMO_MAX + 1, SUBEXPFIN_K,
+            (ref_params->wmmat[7] >> GM_ROW3HOMO_PREC_DIFF),
+            (params->wmmat[7] >> GM_ROW3HOMO_PREC_DIFF));
+    // fallthrough intended
+    case AFFINE:
+    case ROTZOOM:
+      aom_write_signed_primitive_refsubexpfin(
+          w, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+          (ref_params->wmmat[2] >> GM_ALPHA_PREC_DIFF) -
+              (1 << GM_ALPHA_PREC_BITS),
+          (params->wmmat[2] >> GM_ALPHA_PREC_DIFF) - (1 << GM_ALPHA_PREC_BITS));
+      if (type != VERTRAPEZOID)
+        aom_write_signed_primitive_refsubexpfin(
+            w, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+            (ref_params->wmmat[3] >> GM_ALPHA_PREC_DIFF),
+            (params->wmmat[3] >> GM_ALPHA_PREC_DIFF));
+      if (type >= AFFINE) {
+        if (type != HORTRAPEZOID)
+          aom_write_signed_primitive_refsubexpfin(
+              w, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+              (ref_params->wmmat[4] >> GM_ALPHA_PREC_DIFF),
+              (params->wmmat[4] >> GM_ALPHA_PREC_DIFF));
+        aom_write_signed_primitive_refsubexpfin(
+            w, GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+            (ref_params->wmmat[5] >> GM_ALPHA_PREC_DIFF) -
+                (1 << GM_ALPHA_PREC_BITS),
+            (params->wmmat[5] >> GM_ALPHA_PREC_DIFF) -
+                (1 << GM_ALPHA_PREC_BITS));
+      }
+    // fallthrough intended
+    case TRANSLATION:
+      trans_bits = (type == TRANSLATION) ? GM_ABS_TRANS_ONLY_BITS - !allow_hp
+                                         : GM_ABS_TRANS_BITS;
+      trans_prec_diff = (type == TRANSLATION)
+                            ? GM_TRANS_ONLY_PREC_DIFF + !allow_hp
+                            : GM_TRANS_PREC_DIFF;
+      aom_write_signed_primitive_refsubexpfin(
+          w, (1 << trans_bits) + 1, SUBEXPFIN_K,
+          (ref_params->wmmat[0] >> trans_prec_diff),
+          (params->wmmat[0] >> trans_prec_diff));
+      aom_write_signed_primitive_refsubexpfin(
+          w, (1 << trans_bits) + 1, SUBEXPFIN_K,
+          (ref_params->wmmat[1] >> trans_prec_diff),
+          (params->wmmat[1] >> trans_prec_diff));
+      break;
+    case IDENTITY: break;
+    default: assert(0);
+  }
+}
+
+static void write_global_motion(AV1_COMP *cpi, aom_writer *w) {
+  AV1_COMMON *const cm = &cpi->common;
+  int frame;
+  for (frame = LAST_FRAME; frame <= ALTREF_FRAME; ++frame) {
+#if !CONFIG_REF_MV
+    // With ref-mv, clearing unused global motion models here is
+    // unsafe, and we need to rely on the recode loop to do it
+    // instead. See av1_find_mv_refs for details.
+    if (!cpi->td.rd_counts.global_motion_used[frame]) {
+      set_default_warp_params(&cm->global_motion[frame]);
+    }
+#endif
+    write_global_motion_params(
+        &cm->global_motion[frame], &cm->prev_frame->global_motion[frame],
+        cm->fc->global_motion_types_prob, w, cm->allow_high_precision_mv);
+    /*
+    printf("Frame %d/%d: Enc Ref %d (used %d): %d %d %d %d\n",
+           cm->current_video_frame, cm->show_frame, frame,
+           cpi->global_motion_used[frame], cm->global_motion[frame].wmmat[0],
+           cm->global_motion[frame].wmmat[1], cm->global_motion[frame].wmmat[2],
+           cm->global_motion[frame].wmmat[3]);
+           */
+  }
+}
+#endif
+
+static uint32_t write_compressed_header(AV1_COMP *cpi, uint8_t *data) {
+  AV1_COMMON *const cm = &cpi->common;
+#if CONFIG_SUPERTX
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+#endif  // CONFIG_SUPERTX
+  FRAME_CONTEXT *const fc = cm->fc;
+  FRAME_COUNTS *counts = cpi->td.counts;
+  aom_writer *header_bc;
+  int i, j;
+
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+
+#if CONFIG_ANS
+  int header_size;
+  header_bc = &cpi->buf_ans;
+  buf_ans_write_init(header_bc, data);
+#else
+  aom_writer real_header_bc;
+  header_bc = &real_header_bc;
+  aom_start_encode(header_bc, data);
+#endif
+
+#if CONFIG_LOOP_RESTORATION
+  encode_restoration(cm, header_bc);
+#endif  // CONFIG_LOOP_RESTORATION
+#if !CONFIG_EC_ADAPT
+  update_txfm_probs(cm, header_bc, counts);
+#endif
+#if CONFIG_LV_MAP
+  av1_write_txb_probs(cpi, header_bc);
+#else
+#if !CONFIG_PVQ
+#if !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET)
+  update_coef_probs(cpi, header_bc);
+#endif  // !(CONFIG_EC_ADAPT && CONFIG_NEW_TOKENSET)
+#endif  // CONFIG_PVQ
+#endif  // CONFIG_LV_MAP
+
+#if CONFIG_VAR_TX
+  update_txfm_partition_probs(cm, header_bc, counts, probwt);
+#endif
+
+  update_skip_probs(cm, header_bc, counts);
+#if !CONFIG_EC_ADAPT && CONFIG_DELTA_Q
+  update_delta_q_probs(cm, header_bc, counts);
+#if CONFIG_EXT_DELTA_Q
+  update_delta_lf_probs(cm, header_bc, counts);
+#endif
+#endif
+#if !CONFIG_EC_ADAPT
+  update_seg_probs(cpi, header_bc);
+
+  for (i = 0; i < INTRA_MODES; ++i) {
+    prob_diff_update(av1_intra_mode_tree, fc->uv_mode_prob[i],
+                     counts->uv_mode[i], INTRA_MODES, probwt, header_bc);
+  }
+
+#if CONFIG_EXT_PARTITION_TYPES
+  for (i = 0; i < PARTITION_PLOFFSET; ++i)
+    prob_diff_update(av1_partition_tree, fc->partition_prob[i],
+                     counts->partition[i], PARTITION_TYPES, probwt, header_bc);
+  for (; i < PARTITION_CONTEXTS_PRIMARY; ++i)
+    prob_diff_update(av1_ext_partition_tree, fc->partition_prob[i],
+                     counts->partition[i], EXT_PARTITION_TYPES, probwt,
+                     header_bc);
+#else
+  for (i = 0; i < PARTITION_CONTEXTS_PRIMARY; ++i)
+    prob_diff_update(av1_partition_tree, fc->partition_prob[i],
+                     counts->partition[i], PARTITION_TYPES, probwt, header_bc);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+#if CONFIG_UNPOISON_PARTITION_CTX
+  for (; i < PARTITION_CONTEXTS_PRIMARY + PARTITION_BLOCK_SIZES; ++i) {
+    unsigned int ct[2] = { counts->partition[i][PARTITION_VERT],
+                           counts->partition[i][PARTITION_SPLIT] };
+    assert(counts->partition[i][PARTITION_NONE] == 0);
+    assert(counts->partition[i][PARTITION_HORZ] == 0);
+    assert(fc->partition_prob[i][PARTITION_NONE] == 0);
+    assert(fc->partition_prob[i][PARTITION_HORZ] == 0);
+    av1_cond_prob_diff_update(header_bc, &fc->partition_prob[i][PARTITION_VERT],
+                              ct, probwt);
+  }
+  for (; i < PARTITION_CONTEXTS_PRIMARY + 2 * PARTITION_BLOCK_SIZES; ++i) {
+    unsigned int ct[2] = { counts->partition[i][PARTITION_HORZ],
+                           counts->partition[i][PARTITION_SPLIT] };
+    assert(counts->partition[i][PARTITION_NONE] == 0);
+    assert(counts->partition[i][PARTITION_VERT] == 0);
+    assert(fc->partition_prob[i][PARTITION_NONE] == 0);
+    assert(fc->partition_prob[i][PARTITION_VERT] == 0);
+    av1_cond_prob_diff_update(header_bc, &fc->partition_prob[i][PARTITION_HORZ],
+                              ct, probwt);
+  }
+#endif
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+  for (i = 0; i < INTRA_FILTERS + 1; ++i)
+    prob_diff_update(av1_intra_filter_tree, fc->intra_filter_probs[i],
+                     counts->intra_filter[i], INTRA_FILTERS, probwt, header_bc);
+#endif  // CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+#endif  // !CONFIG_EC_ADAPT
+
+  if (frame_is_intra_only(cm)) {
+    av1_copy(cm->kf_y_prob, av1_kf_y_mode_prob);
+#if CONFIG_EC_MULTISYMBOL
+    av1_copy(cm->fc->kf_y_cdf, av1_kf_y_mode_cdf);
+#endif
+
+#if !CONFIG_EC_ADAPT
+    for (i = 0; i < INTRA_MODES; ++i)
+      for (j = 0; j < INTRA_MODES; ++j)
+        prob_diff_update(av1_intra_mode_tree, cm->kf_y_prob[i][j],
+                         counts->kf_y_mode[i][j], INTRA_MODES, probwt,
+                         header_bc);
+#endif  // CONFIG_EC_ADAPT
+  } else {
+#if CONFIG_REF_MV
+    update_inter_mode_probs(cm, header_bc, counts);
+#else
+#if !CONFIG_EC_ADAPT
+    for (i = 0; i < INTER_MODE_CONTEXTS; ++i) {
+      prob_diff_update(av1_inter_mode_tree, cm->fc->inter_mode_probs[i],
+                       counts->inter_mode[i], INTER_MODES, probwt, header_bc);
+    }
+#endif
+#endif
+#if CONFIG_EXT_INTER
+    update_inter_compound_mode_probs(cm, probwt, header_bc);
+
+    if (cm->reference_mode != COMPOUND_REFERENCE) {
+      for (i = 0; i < BLOCK_SIZE_GROUPS; i++) {
+        if (is_interintra_allowed_bsize_group(i)) {
+          av1_cond_prob_diff_update(header_bc, &fc->interintra_prob[i],
+                                    cm->counts.interintra[i], probwt);
+        }
+      }
+      for (i = 0; i < BLOCK_SIZE_GROUPS; i++) {
+        prob_diff_update(
+            av1_interintra_mode_tree, cm->fc->interintra_mode_prob[i],
+            counts->interintra_mode[i], INTERINTRA_MODES, probwt, header_bc);
+      }
+      for (i = 0; i < BLOCK_SIZES; i++) {
+        if (is_interintra_allowed_bsize(i) && is_interintra_wedge_used(i))
+          av1_cond_prob_diff_update(header_bc, &fc->wedge_interintra_prob[i],
+                                    cm->counts.wedge_interintra[i], probwt);
+      }
+    }
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+    if (cm->reference_mode != SINGLE_REFERENCE) {
+      for (i = 0; i < BLOCK_SIZES; i++)
+        prob_diff_update(av1_compound_type_tree, fc->compound_type_prob[i],
+                         cm->counts.compound_interinter[i], COMPOUND_TYPES,
+                         probwt, header_bc);
+    }
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    for (i = BLOCK_8X8; i < BLOCK_SIZES; ++i)
+      prob_diff_update(av1_motion_mode_tree, fc->motion_mode_prob[i],
+                       counts->motion_mode[i], MOTION_MODES, probwt, header_bc);
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+#if !CONFIG_EC_ADAPT
+    if (cm->interp_filter == SWITCHABLE)
+      update_switchable_interp_probs(cm, header_bc, counts);
+#endif
+
+    for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
+      av1_cond_prob_diff_update(header_bc, &fc->intra_inter_prob[i],
+                                counts->intra_inter[i], probwt);
+
+    if (cpi->allow_comp_inter_inter) {
+      const int use_hybrid_pred = cm->reference_mode == REFERENCE_MODE_SELECT;
+      if (use_hybrid_pred)
+        for (i = 0; i < COMP_INTER_CONTEXTS; i++)
+          av1_cond_prob_diff_update(header_bc, &fc->comp_inter_prob[i],
+                                    counts->comp_inter[i], probwt);
+    }
+
+    if (cm->reference_mode != COMPOUND_REFERENCE) {
+      for (i = 0; i < REF_CONTEXTS; i++) {
+        for (j = 0; j < (SINGLE_REFS - 1); j++) {
+          av1_cond_prob_diff_update(header_bc, &fc->single_ref_prob[i][j],
+                                    counts->single_ref[i][j], probwt);
+        }
+      }
+    }
+    if (cm->reference_mode != SINGLE_REFERENCE) {
+      for (i = 0; i < REF_CONTEXTS; i++) {
+#if CONFIG_EXT_REFS
+        for (j = 0; j < (FWD_REFS - 1); j++) {
+          av1_cond_prob_diff_update(header_bc, &fc->comp_ref_prob[i][j],
+                                    counts->comp_ref[i][j], probwt);
+        }
+        for (j = 0; j < (BWD_REFS - 1); j++) {
+          av1_cond_prob_diff_update(header_bc, &fc->comp_bwdref_prob[i][j],
+                                    counts->comp_bwdref[i][j], probwt);
+        }
+#else
+        for (j = 0; j < (COMP_REFS - 1); j++) {
+          av1_cond_prob_diff_update(header_bc, &fc->comp_ref_prob[i][j],
+                                    counts->comp_ref[i][j], probwt);
+        }
+#endif  // CONFIG_EXT_REFS
+      }
+    }
+
+#if !CONFIG_EC_ADAPT
+    for (i = 0; i < BLOCK_SIZE_GROUPS; ++i) {
+      prob_diff_update(av1_intra_mode_tree, cm->fc->y_mode_prob[i],
+                       counts->y_mode[i], INTRA_MODES, probwt, header_bc);
+    }
+#endif
+
+    av1_write_nmv_probs(cm, cm->allow_high_precision_mv, header_bc,
+#if CONFIG_REF_MV
+                        counts->mv);
+#else
+                        &counts->mv);
+#endif
+#if !CONFIG_EC_ADAPT
+    update_ext_tx_probs(cm, header_bc);
+#endif
+#if CONFIG_SUPERTX
+    if (!xd->lossless[0]) update_supertx_probs(cm, probwt, header_bc);
+#endif  // CONFIG_SUPERTX
+#if CONFIG_GLOBAL_MOTION
+    write_global_motion(cpi, header_bc);
+#endif  // CONFIG_GLOBAL_MOTION
+  }
+#if CONFIG_EC_MULTISYMBOL
+#if !CONFIG_EC_ADAPT
+#if CONFIG_NEW_TOKENSET
+  av1_coef_head_cdfs(fc);
+#endif
+  av1_coef_pareto_cdfs(fc);
+#if CONFIG_REF_MV
+  for (i = 0; i < NMV_CONTEXTS; ++i) av1_set_mv_cdfs(&fc->nmvc[i]);
+#else
+  av1_set_mv_cdfs(&fc->nmvc);
+#endif
+#if CONFIG_EC_MULTISYMBOL
+  av1_set_mode_cdfs(cm);
+#endif
+#endif  // !CONFIG_EC_ADAPT
+#endif
+#if CONFIG_ANS
+  aom_buf_ans_flush(header_bc);
+  header_size = buf_ans_write_end(header_bc);
+  assert(header_size <= 0xffff);
+  return header_size;
+#else
+  aom_stop_encode(header_bc);
+  assert(header_bc->pos <= 0xffff);
+  return header_bc->pos;
+#endif  // CONFIG_ANS
+}
+
+static int choose_size_bytes(uint32_t size, int spare_msbs) {
+  // Choose the number of bytes required to represent size, without
+  // using the 'spare_msbs' number of most significant bits.
+
+  // Make sure we will fit in 4 bytes to start with..
+  if (spare_msbs > 0 && size >> (32 - spare_msbs) != 0) return -1;
+
+  // Normalise to 32 bits
+  size <<= spare_msbs;
+
+  if (size >> 24 != 0)
+    return 4;
+  else if (size >> 16 != 0)
+    return 3;
+  else if (size >> 8 != 0)
+    return 2;
+  else
+    return 1;
+}
+
+static void mem_put_varsize(uint8_t *const dst, const int sz, const int val) {
+  switch (sz) {
+    case 1: dst[0] = (uint8_t)(val & 0xff); break;
+    case 2: mem_put_le16(dst, val); break;
+    case 3: mem_put_le24(dst, val); break;
+    case 4: mem_put_le32(dst, val); break;
+    default: assert(0 && "Invalid size"); break;
+  }
+}
+static int remux_tiles(const AV1_COMMON *const cm, uint8_t *dst,
+                       const uint32_t data_size, const uint32_t max_tile_size,
+                       const uint32_t max_tile_col_size,
+                       int *const tile_size_bytes,
+                       int *const tile_col_size_bytes) {
+// Choose the tile size bytes (tsb) and tile column size bytes (tcsb)
+#if CONFIG_EXT_TILE
+  // The top bit in the tile size field indicates tile copy mode, so we
+  // have 1 less bit to code the tile size
+  const int tsb = choose_size_bytes(max_tile_size, 1);
+  const int tcsb = choose_size_bytes(max_tile_col_size, 0);
+#else
+  const int tsb = choose_size_bytes(max_tile_size, 0);
+  const int tcsb = 4;  // This is ignored
+  (void)max_tile_col_size;
+#endif  // CONFIG_EXT_TILE
+
+  assert(tsb > 0);
+  assert(tcsb > 0);
+
+  *tile_size_bytes = tsb;
+  *tile_col_size_bytes = tcsb;
+
+  if (tsb == 4 && tcsb == 4) {
+    return data_size;
+  } else {
+    uint32_t wpos = 0;
+    uint32_t rpos = 0;
+
+#if CONFIG_EXT_TILE
+    int tile_row;
+    int tile_col;
+
+    for (tile_col = 0; tile_col < cm->tile_cols; tile_col++) {
+      // All but the last column has a column header
+      if (tile_col < cm->tile_cols - 1) {
+        uint32_t tile_col_size = mem_get_le32(dst + rpos);
+        rpos += 4;
+
+        // Adjust the tile column size by the number of bytes removed
+        // from the tile size fields.
+        tile_col_size -= (4 - tsb) * cm->tile_rows;
+
+        mem_put_varsize(dst + wpos, tcsb, tile_col_size);
+        wpos += tcsb;
+      }
+
+      for (tile_row = 0; tile_row < cm->tile_rows; tile_row++) {
+        // All, including the last row has a header
+        uint32_t tile_header = mem_get_le32(dst + rpos);
+        rpos += 4;
+
+        // If this is a copy tile, we need to shift the MSB to the
+        // top bit of the new width, and there is no data to copy.
+        if (tile_header >> 31 != 0) {
+          if (tsb < 4) tile_header >>= 32 - 8 * tsb;
+          mem_put_varsize(dst + wpos, tsb, tile_header);
+          wpos += tsb;
+        } else {
+          mem_put_varsize(dst + wpos, tsb, tile_header);
+          wpos += tsb;
+
+          memmove(dst + wpos, dst + rpos, tile_header);
+          rpos += tile_header;
+          wpos += tile_header;
+        }
+      }
+    }
+#else
+    const int n_tiles = cm->tile_cols * cm->tile_rows;
+    int n;
+
+    for (n = 0; n < n_tiles; n++) {
+      int tile_size;
+
+      if (n == n_tiles - 1) {
+        tile_size = data_size - rpos;
+      } else {
+        tile_size = mem_get_le32(dst + rpos);
+        rpos += 4;
+        mem_put_varsize(dst + wpos, tsb, tile_size);
+        wpos += tsb;
+      }
+
+      memmove(dst + wpos, dst + rpos, tile_size);
+
+      rpos += tile_size;
+      wpos += tile_size;
+    }
+#endif  // CONFIG_EXT_TILE
+
+    assert(rpos > wpos);
+    assert(rpos == data_size);
+
+    return wpos;
+  }
+}
+
+void av1_pack_bitstream(AV1_COMP *const cpi, uint8_t *dst, size_t *size) {
+  uint8_t *data = dst;
+#if !CONFIG_TILE_GROUPS
+  uint32_t compressed_header_size;
+  uint32_t uncompressed_header_size;
+  struct aom_write_bit_buffer saved_wb;
+#endif
+  uint32_t data_size;
+  struct aom_write_bit_buffer wb = { data, 0 };
+
+  unsigned int max_tile_size;
+  unsigned int max_tile_col_size;
+
+#if CONFIG_BITSTREAM_DEBUG
+  bitstream_queue_reset_write();
+#endif
+
+#if !CONFIG_TILE_GROUPS
+  int tile_size_bytes;
+  int tile_col_size_bytes;
+  AV1_COMMON *const cm = &cpi->common;
+  const int have_tiles = cm->tile_cols * cm->tile_rows > 1;
+
+  // Write the uncompressed header
+  write_uncompressed_header(cpi, &wb);
+
+#if CONFIG_EXT_REFS
+  if (cm->show_existing_frame) {
+    *size = aom_wb_bytes_written(&wb);
+    return;
+  }
+#endif  // CONFIG_EXT_REFS
+
+  // We do not know these in advance. Output placeholder bit.
+  saved_wb = wb;
+  // Write tile size magnitudes
+  if (have_tiles) {
+// Note that the last item in the uncompressed header is the data
+// describing tile configuration.
+#if CONFIG_EXT_TILE
+    // Number of bytes in tile column size - 1
+    aom_wb_write_literal(&wb, 0, 2);
+#endif  // CONFIG_EXT_TILE
+    // Number of bytes in tile size - 1
+    aom_wb_write_literal(&wb, 0, 2);
+  }
+  // Size of compressed header
+  aom_wb_write_literal(&wb, 0, 16);
+
+  uncompressed_header_size = (uint32_t)aom_wb_bytes_written(&wb);
+  data += uncompressed_header_size;
+
+  aom_clear_system_state();
+
+  // Write the compressed header
+  compressed_header_size = write_compressed_header(cpi, data);
+  data += compressed_header_size;
+
+  // Write the encoded tile data
+  data_size = write_tiles(cpi, data, &max_tile_size, &max_tile_col_size);
+#else
+  data_size = write_tiles(cpi, &wb, &max_tile_size, &max_tile_col_size);
+#endif
+#if !CONFIG_TILE_GROUPS
+  if (have_tiles) {
+    data_size =
+        remux_tiles(cm, data, data_size, max_tile_size, max_tile_col_size,
+                    &tile_size_bytes, &tile_col_size_bytes);
+  }
+
+  data += data_size;
+
+  // Now fill in the gaps in the uncompressed header.
+  if (have_tiles) {
+#if CONFIG_EXT_TILE
+    assert(tile_col_size_bytes >= 1 && tile_col_size_bytes <= 4);
+    aom_wb_write_literal(&saved_wb, tile_col_size_bytes - 1, 2);
+#endif  // CONFIG_EXT_TILE
+    assert(tile_size_bytes >= 1 && tile_size_bytes <= 4);
+    aom_wb_write_literal(&saved_wb, tile_size_bytes - 1, 2);
+  }
+  // TODO(jbb): Figure out what to do if compressed_header_size > 16 bits.
+  assert(compressed_header_size <= 0xffff);
+  aom_wb_write_literal(&saved_wb, compressed_header_size, 16);
+#else
+  data += data_size;
+#endif
+#if CONFIG_ANS && ANS_REVERSE
+  // Avoid aliasing the superframe index
+  *data++ = 0;
+#endif
+  *size = data - dst;
+}
diff --git a/third_party/aom/av1/encoder/bitstream.h b/third_party/aom/av1/encoder/bitstream.h
new file mode 100644
index 0000000000..c75d80891b
--- /dev/null
+++ b/third_party/aom/av1/encoder/bitstream.h
@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_BITSTREAM_H_
+#define AV1_ENCODER_BITSTREAM_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "av1/encoder/encoder.h"
+
+#if CONFIG_REFERENCE_BUFFER
+void write_sequence_header(SequenceHeader *seq_params);
+#endif
+
+void av1_pack_bitstream(AV1_COMP *const cpi, uint8_t *dest, size_t *size);
+
+void av1_encode_token_init(void);
+
+static INLINE int av1_preserve_existing_gf(AV1_COMP *cpi) {
+#if CONFIG_EXT_REFS
+  // Do not swap gf and arf indices for internal overlay frames
+  return !cpi->multi_arf_allowed && cpi->rc.is_src_frame_alt_ref &&
+         !cpi->rc.is_src_frame_ext_arf;
+#else
+  return !cpi->multi_arf_allowed && cpi->refresh_golden_frame &&
+         cpi->rc.is_src_frame_alt_ref;
+#endif  // CONFIG_EXT_REFS
+}
+
+void av1_write_tx_type(const AV1_COMMON *const cm, const MACROBLOCKD *xd,
+#if CONFIG_SUPERTX
+                       const int supertx_enabled,
+#endif
+#if CONFIG_TXK_SEL
+                       int block, int plane,
+#endif
+                       aom_writer *w);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_BITSTREAM_H_
diff --git a/third_party/aom/av1/encoder/block.h b/third_party/aom/av1/encoder/block.h
new file mode 100644
index 0000000000..39e08d5b4c
--- /dev/null
+++ b/third_party/aom/av1/encoder/block.h
@@ -0,0 +1,241 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_BLOCK_H_
+#define AV1_ENCODER_BLOCK_H_
+
+#include "av1/common/entropymv.h"
+#include "av1/common/entropy.h"
+#if CONFIG_PVQ
+#include "av1/encoder/encint.h"
+#endif
+#if CONFIG_REF_MV
+#include "av1/common/mvref_common.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_PVQ
+// Maximum possible # of tx blocks in luma plane, which is currently 256,
+// since there can be 16x16 of 4x4 tx.
+#define MAX_PVQ_BLOCKS_IN_SB (MAX_SB_SQUARE >> 2 * OD_LOG_BSIZE0)
+#endif
+
+typedef struct {
+  unsigned int sse;
+  int sum;
+  unsigned int var;
+} DIFF;
+
+typedef struct macroblock_plane {
+  DECLARE_ALIGNED(16, int16_t, src_diff[MAX_SB_SQUARE]);
+#if CONFIG_PVQ
+  DECLARE_ALIGNED(16, int16_t, src_int16[MAX_SB_SQUARE]);
+#endif
+  tran_low_t *qcoeff;
+  tran_low_t *coeff;
+  uint16_t *eobs;
+#if CONFIG_LV_MAP
+  uint8_t *txb_entropy_ctx;
+#endif
+  struct buf_2d src;
+
+  // Quantizer setings
+  const int16_t *quant_fp;
+  const int16_t *round_fp;
+  const int16_t *quant;
+  const int16_t *quant_shift;
+  const int16_t *zbin;
+  const int16_t *round;
+#if CONFIG_NEW_QUANT
+  const cuml_bins_type_nuq *cuml_bins_nuq[QUANT_PROFILES];
+#endif  // CONFIG_NEW_QUANT
+} MACROBLOCK_PLANE;
+
+/* The [2] dimension is for whether we skip the EOB node (i.e. if previous
+ * coefficient in this block was zero) or not. */
+typedef unsigned int av1_coeff_cost[PLANE_TYPES][REF_TYPES][COEF_BANDS][2]
+                                   [COEFF_CONTEXTS][ENTROPY_TOKENS];
+
+typedef struct {
+  int_mv ref_mvs[MODE_CTX_REF_FRAMES][MAX_MV_REF_CANDIDATES];
+  int16_t mode_context[MODE_CTX_REF_FRAMES];
+#if CONFIG_LV_MAP
+  // TODO(angiebird): Reduce the buffer size according to sb_type
+  tran_low_t tcoeff[MAX_MB_PLANE][MAX_SB_SQUARE];
+  uint16_t eobs[MAX_MB_PLANE][MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
+  uint8_t txb_skip_ctx[MAX_MB_PLANE]
+                      [MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
+  int dc_sign_ctx[MAX_MB_PLANE]
+                 [MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
+#endif
+#if CONFIG_REF_MV
+  uint8_t ref_mv_count[MODE_CTX_REF_FRAMES];
+  CANDIDATE_MV ref_mv_stack[MODE_CTX_REF_FRAMES][MAX_REF_MV_STACK_SIZE];
+#if CONFIG_EXT_INTER
+  int16_t compound_mode_context[MODE_CTX_REF_FRAMES];
+#endif  // CONFIG_EXT_INTER
+#endif
+} MB_MODE_INFO_EXT;
+
+typedef struct {
+  int col_min;
+  int col_max;
+  int row_min;
+  int row_max;
+} MvLimits;
+
+#if CONFIG_PALETTE
+typedef struct {
+  uint8_t best_palette_color_map[MAX_SB_SQUARE];
+  float kmeans_data_buf[2 * MAX_SB_SQUARE];
+} PALETTE_BUFFER;
+#endif  // CONFIG_PALETTE
+
+typedef struct macroblock MACROBLOCK;
+struct macroblock {
+  struct macroblock_plane plane[MAX_MB_PLANE];
+
+  MACROBLOCKD e_mbd;
+  MB_MODE_INFO_EXT *mbmi_ext;
+  int skip_block;
+  int qindex;
+
+  // The equivalent error at the current rdmult of one whole bit (not one
+  // bitcost unit).
+  int errorperbit;
+  // The equivalend SAD error of one (whole) bit at the current quantizer
+  // for large blocks.
+  int sadperbit16;
+  // The equivalend SAD error of one (whole) bit at the current quantizer
+  // for sub-8x8 blocks.
+  int sadperbit4;
+  int rddiv;
+  int rdmult;
+  int mb_energy;
+  int *m_search_count_ptr;
+  int *ex_search_count_ptr;
+
+#if CONFIG_VAR_TX
+  unsigned int txb_split_count;
+#endif
+
+  // These are set to their default values at the beginning, and then adjusted
+  // further in the encoding process.
+  BLOCK_SIZE min_partition_size;
+  BLOCK_SIZE max_partition_size;
+
+  int mv_best_ref_index[TOTAL_REFS_PER_FRAME];
+  unsigned int max_mv_context[TOTAL_REFS_PER_FRAME];
+  unsigned int source_variance;
+  unsigned int pred_sse[TOTAL_REFS_PER_FRAME];
+  int pred_mv_sad[TOTAL_REFS_PER_FRAME];
+
+#if CONFIG_REF_MV
+  int *nmvjointcost;
+  int nmv_vec_cost[NMV_CONTEXTS][MV_JOINTS];
+  int *nmvcost[NMV_CONTEXTS][2];
+  int *nmvcost_hp[NMV_CONTEXTS][2];
+  int **mv_cost_stack[NMV_CONTEXTS];
+  int *nmvjointsadcost;
+#else
+  int nmvjointcost[MV_JOINTS];
+  int *nmvcost[2];
+  int *nmvcost_hp[2];
+  int nmvjointsadcost[MV_JOINTS];
+#endif
+
+  int **mvcost;
+  int *nmvsadcost[2];
+  int *nmvsadcost_hp[2];
+  int **mvsadcost;
+#if CONFIG_MOTION_VAR
+  int32_t *wsrc_buf;
+  int32_t *mask_buf;
+#endif  // CONFIG_MOTION_VAR
+
+#if CONFIG_PALETTE
+  PALETTE_BUFFER *palette_buffer;
+#endif  // CONFIG_PALETTE
+
+  // These define limits to motion vector components to prevent them
+  // from extending outside the UMV borders
+  MvLimits mv_limits;
+
+#if CONFIG_VAR_TX
+  uint8_t blk_skip[MAX_MB_PLANE][MAX_MIB_SIZE * MAX_MIB_SIZE * 8];
+#if CONFIG_REF_MV
+  uint8_t blk_skip_drl[MAX_MB_PLANE][MAX_MIB_SIZE * MAX_MIB_SIZE * 8];
+#endif
+#endif
+
+  int skip;
+
+#if CONFIG_CB4X4
+  int skip_chroma_rd;
+#endif
+
+  // note that token_costs is the cost when eob node is skipped
+  av1_coeff_cost token_costs[TX_SIZES];
+
+  int optimize;
+
+  // Used to store sub partition's choices.
+  MV pred_mv[TOTAL_REFS_PER_FRAME];
+
+  // Store the best motion vector during motion search
+  int_mv best_mv;
+  // Store the second best motion vector during full-pixel motion search
+  int_mv second_best_mv;
+
+  // use default transform and skip transform type search for intra modes
+  int use_default_intra_tx_type;
+  // use default transform and skip transform type search for inter modes
+  int use_default_inter_tx_type;
+#if CONFIG_PVQ
+  int rate;
+  // 1 if neither AC nor DC is coded. Only used during RDO.
+  int pvq_skip[MAX_MB_PLANE];
+  PVQ_QUEUE *pvq_q;
+
+  // Storage for PVQ tx block encodings in a superblock.
+  // There can be max 16x16 of 4x4 blocks (and YUV) encode by PVQ
+  // 256 is the max # of 4x4 blocks in a SB (64x64), which comes from:
+  // 1) Since PVQ is applied to each trasnform-ed block
+  // 2) 4x4 is the smallest tx size in AV1
+  // 3) AV1 allows using smaller tx size than block (i.e. partition) size
+  // TODO(yushin) : The memory usage could be improved a lot, since this has
+  // storage for 10 bands and 128 coefficients for every 4x4 block,
+  PVQ_INFO pvq[MAX_PVQ_BLOCKS_IN_SB][MAX_MB_PLANE];
+  daala_enc_ctx daala_enc;
+  int pvq_speed;
+  int pvq_coded;  // Indicates whether pvq_info needs be stored to tokenize
+#endif
+#if CONFIG_DAALA_DIST
+  // Keep rate of each 4x4 block in the current macroblock during RDO
+  // This is needed when using the 8x8 Daala distortion metric during RDO,
+  // because it evaluates distortion in a different order than the underlying
+  // 4x4 blocks are coded.
+  int rate_4x4[256];
+#endif
+#if CONFIG_CFL
+  // Whether luma needs to be stored during RDO.
+  int cfl_store_y;
+#endif
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_BLOCK_H_
diff --git a/third_party/aom/av1/encoder/blockiness.c b/third_party/aom/av1/encoder/blockiness.c
new file mode 100644
index 0000000000..113ceb29d2
--- /dev/null
+++ b/third_party/aom/av1/encoder/blockiness.c
@@ -0,0 +1,142 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "av1/common/common.h"
+#include "av1/common/filter.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_convolve.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/system_state.h"
+
+static int horizontal_filter(const uint8_t *s) {
+  return (s[1] - s[-2]) * 2 + (s[-1] - s[0]) * 6;
+}
+
+static int vertical_filter(const uint8_t *s, int p) {
+  return (s[p] - s[-2 * p]) * 2 + (s[-p] - s[0]) * 6;
+}
+
+static int variance(int sum, int sum_squared, int size) {
+  return sum_squared / size - (sum / size) * (sum / size);
+}
+// Calculate a blockiness level for a vertical block edge.
+// This function returns a new blockiness metric that's defined as
+
+//              p0 p1 p2 p3
+//              q0 q1 q2 q3
+// block edge ->
+//              r0 r1 r2 r3
+//              s0 s1 s2 s3
+
+// blockiness =  p0*-2+q0*6+r0*-6+s0*2 +
+//               p1*-2+q1*6+r1*-6+s1*2 +
+//               p2*-2+q2*6+r2*-6+s2*2 +
+//               p3*-2+q3*6+r3*-6+s3*2 ;
+
+// reconstructed_blockiness = abs(blockiness from reconstructed buffer -
+//                                blockiness from source buffer,0)
+//
+// I make the assumption that flat blocks are much more visible than high
+// contrast blocks. As such, I scale the result of the blockiness calc
+// by dividing the blockiness by the variance of the pixels on either side
+// of the edge as follows:
+// var_0 = (q0^2+q1^2+q2^2+q3^2) - ((q0 + q1 + q2 + q3) / 4 )^2
+// var_1 = (r0^2+r1^2+r2^2+r3^2) - ((r0 + r1 + r2 + r3) / 4 )^2
+// The returned blockiness is the scaled value
+// Reconstructed blockiness / ( 1 + var_0 + var_1 ) ;
+static int blockiness_vertical(const uint8_t *s, int sp, const uint8_t *r,
+                               int rp, int size) {
+  int s_blockiness = 0;
+  int r_blockiness = 0;
+  int sum_0 = 0;
+  int sum_sq_0 = 0;
+  int sum_1 = 0;
+  int sum_sq_1 = 0;
+  int i;
+  int var_0;
+  int var_1;
+  for (i = 0; i < size; ++i, s += sp, r += rp) {
+    s_blockiness += horizontal_filter(s);
+    r_blockiness += horizontal_filter(r);
+    sum_0 += s[0];
+    sum_sq_0 += s[0] * s[0];
+    sum_1 += s[-1];
+    sum_sq_1 += s[-1] * s[-1];
+  }
+  var_0 = variance(sum_0, sum_sq_0, size);
+  var_1 = variance(sum_1, sum_sq_1, size);
+  r_blockiness = abs(r_blockiness);
+  s_blockiness = abs(s_blockiness);
+
+  if (r_blockiness > s_blockiness)
+    return (r_blockiness - s_blockiness) / (1 + var_0 + var_1);
+  else
+    return 0;
+}
+
+// Calculate a blockiness level for a horizontal block edge
+// same as above.
+static int blockiness_horizontal(const uint8_t *s, int sp, const uint8_t *r,
+                                 int rp, int size) {
+  int s_blockiness = 0;
+  int r_blockiness = 0;
+  int sum_0 = 0;
+  int sum_sq_0 = 0;
+  int sum_1 = 0;
+  int sum_sq_1 = 0;
+  int i;
+  int var_0;
+  int var_1;
+  for (i = 0; i < size; ++i, ++s, ++r) {
+    s_blockiness += vertical_filter(s, sp);
+    r_blockiness += vertical_filter(r, rp);
+    sum_0 += s[0];
+    sum_sq_0 += s[0] * s[0];
+    sum_1 += s[-sp];
+    sum_sq_1 += s[-sp] * s[-sp];
+  }
+  var_0 = variance(sum_0, sum_sq_0, size);
+  var_1 = variance(sum_1, sum_sq_1, size);
+  r_blockiness = abs(r_blockiness);
+  s_blockiness = abs(s_blockiness);
+
+  if (r_blockiness > s_blockiness)
+    return (r_blockiness - s_blockiness) / (1 + var_0 + var_1);
+  else
+    return 0;
+}
+
+// This function returns the blockiness for the entire frame currently by
+// looking at all borders in steps of 4.
+double av1_get_blockiness(const unsigned char *img1, int img1_pitch,
+                          const unsigned char *img2, int img2_pitch, int width,
+                          int height) {
+  double blockiness = 0;
+  int i, j;
+  aom_clear_system_state();
+  for (i = 0; i < height;
+       i += 4, img1 += img1_pitch * 4, img2 += img2_pitch * 4) {
+    for (j = 0; j < width; j += 4) {
+      if (i > 0 && i < height && j > 0 && j < width) {
+        blockiness +=
+            blockiness_vertical(img1 + j, img1_pitch, img2 + j, img2_pitch, 4);
+        blockiness += blockiness_horizontal(img1 + j, img1_pitch, img2 + j,
+                                            img2_pitch, 4);
+      }
+    }
+  }
+  blockiness /= width * height / 16;
+  return blockiness;
+}
diff --git a/third_party/aom/av1/encoder/context_tree.c b/third_party/aom/av1/encoder/context_tree.c
new file mode 100644
index 0000000000..4c7d6ff00a
--- /dev/null
+++ b/third_party/aom/av1/encoder/context_tree.c
@@ -0,0 +1,331 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/encoder/context_tree.h"
+#include "av1/encoder/encoder.h"
+
+static const BLOCK_SIZE square[MAX_SB_SIZE_LOG2 - 1] = {
+#if CONFIG_CB4X4
+  BLOCK_4X4,
+#endif
+  BLOCK_8X8,     BLOCK_16X16, BLOCK_32X32, BLOCK_64X64,
+#if CONFIG_EXT_PARTITION
+  BLOCK_128X128,
+#endif  // CONFIG_EXT_PARTITION
+};
+
+static void alloc_mode_context(AV1_COMMON *cm, int num_4x4_blk,
+#if CONFIG_EXT_PARTITION_TYPES
+                               PARTITION_TYPE partition,
+#endif
+                               PICK_MODE_CONTEXT *ctx) {
+  const int num_blk = (num_4x4_blk < 4 ? 4 : num_4x4_blk);
+  const int num_pix = num_blk * tx_size_2d[0];
+  int i;
+#if CONFIG_CB4X4 && CONFIG_VAR_TX
+  ctx->num_4x4_blk = num_blk / 4;
+#else
+  ctx->num_4x4_blk = num_blk;
+#endif
+
+#if CONFIG_EXT_PARTITION_TYPES
+  ctx->partition = partition;
+#endif
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+#if CONFIG_VAR_TX
+    CHECK_MEM_ERROR(cm, ctx->blk_skip[i], aom_calloc(num_blk, sizeof(uint8_t)));
+#endif
+    CHECK_MEM_ERROR(cm, ctx->coeff[i],
+                    aom_memalign(32, num_pix * sizeof(*ctx->coeff[i])));
+    CHECK_MEM_ERROR(cm, ctx->qcoeff[i],
+                    aom_memalign(32, num_pix * sizeof(*ctx->qcoeff[i])));
+    CHECK_MEM_ERROR(cm, ctx->dqcoeff[i],
+                    aom_memalign(32, num_pix * sizeof(*ctx->dqcoeff[i])));
+    CHECK_MEM_ERROR(cm, ctx->eobs[i],
+                    aom_memalign(32, num_blk * sizeof(*ctx->eobs[i])));
+#if CONFIG_LV_MAP
+    CHECK_MEM_ERROR(
+        cm, ctx->txb_entropy_ctx[i],
+        aom_memalign(32, num_blk * sizeof(*ctx->txb_entropy_ctx[i])));
+#endif
+
+#if CONFIG_PVQ
+    CHECK_MEM_ERROR(cm, ctx->pvq_ref_coeff[i],
+                    aom_memalign(32, num_pix * sizeof(*ctx->pvq_ref_coeff[i])));
+#endif
+  }
+
+#if CONFIG_PALETTE
+  if (cm->allow_screen_content_tools) {
+    for (i = 0; i < 2; ++i) {
+      CHECK_MEM_ERROR(
+          cm, ctx->color_index_map[i],
+          aom_memalign(32, num_pix * sizeof(*ctx->color_index_map[i])));
+    }
+  }
+#endif  // CONFIG_PALETTE
+}
+
+static void free_mode_context(PICK_MODE_CONTEXT *ctx) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+#if CONFIG_VAR_TX
+    aom_free(ctx->blk_skip[i]);
+    ctx->blk_skip[i] = 0;
+#endif
+    aom_free(ctx->coeff[i]);
+    ctx->coeff[i] = 0;
+    aom_free(ctx->qcoeff[i]);
+    ctx->qcoeff[i] = 0;
+    aom_free(ctx->dqcoeff[i]);
+    ctx->dqcoeff[i] = 0;
+#if CONFIG_PVQ
+    aom_free(ctx->pvq_ref_coeff[i]);
+    ctx->pvq_ref_coeff[i] = 0;
+#endif
+    aom_free(ctx->eobs[i]);
+    ctx->eobs[i] = 0;
+#if CONFIG_LV_MAP
+    aom_free(ctx->txb_entropy_ctx[i]);
+    ctx->txb_entropy_ctx[i] = 0;
+#endif
+  }
+
+#if CONFIG_PALETTE
+  for (i = 0; i < 2; ++i) {
+    aom_free(ctx->color_index_map[i]);
+    ctx->color_index_map[i] = 0;
+  }
+#endif  // CONFIG_PALETTE
+}
+
+static void alloc_tree_contexts(AV1_COMMON *cm, PC_TREE *tree,
+                                int num_4x4_blk) {
+#if CONFIG_EXT_PARTITION_TYPES
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_NONE, &tree->none);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_HORZ, &tree->horizontal[0]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_VERT, &tree->vertical[0]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_VERT, &tree->horizontal[1]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_VERT, &tree->vertical[1]);
+
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_HORZ_A,
+                     &tree->horizontala[0]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_HORZ_A,
+                     &tree->horizontala[1]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_HORZ_A,
+                     &tree->horizontala[2]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_HORZ_B,
+                     &tree->horizontalb[0]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_HORZ_B,
+                     &tree->horizontalb[1]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_HORZ_B,
+                     &tree->horizontalb[2]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_VERT_A,
+                     &tree->verticala[0]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_VERT_A,
+                     &tree->verticala[1]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_VERT_A,
+                     &tree->verticala[2]);
+  alloc_mode_context(cm, num_4x4_blk / 2, PARTITION_VERT_B,
+                     &tree->verticalb[0]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_VERT_B,
+                     &tree->verticalb[1]);
+  alloc_mode_context(cm, num_4x4_blk / 4, PARTITION_VERT_B,
+                     &tree->verticalb[2]);
+#ifdef CONFIG_SUPERTX
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_HORZ,
+                     &tree->horizontal_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_VERT, &tree->vertical_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_SPLIT, &tree->split_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_HORZ_A,
+                     &tree->horizontala_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_HORZ_B,
+                     &tree->horizontalb_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_VERT_A,
+                     &tree->verticala_supertx);
+  alloc_mode_context(cm, num_4x4_blk, PARTITION_VERT_B,
+                     &tree->verticalb_supertx);
+#endif  // CONFIG_SUPERTX
+#else
+  alloc_mode_context(cm, num_4x4_blk, &tree->none);
+  alloc_mode_context(cm, num_4x4_blk / 2, &tree->horizontal[0]);
+  alloc_mode_context(cm, num_4x4_blk / 2, &tree->vertical[0]);
+#ifdef CONFIG_SUPERTX
+  alloc_mode_context(cm, num_4x4_blk, &tree->horizontal_supertx);
+  alloc_mode_context(cm, num_4x4_blk, &tree->vertical_supertx);
+  alloc_mode_context(cm, num_4x4_blk, &tree->split_supertx);
+#endif
+
+  if (num_4x4_blk > 4) {
+    alloc_mode_context(cm, num_4x4_blk / 2, &tree->horizontal[1]);
+    alloc_mode_context(cm, num_4x4_blk / 2, &tree->vertical[1]);
+  } else {
+    memset(&tree->horizontal[1], 0, sizeof(tree->horizontal[1]));
+    memset(&tree->vertical[1], 0, sizeof(tree->vertical[1]));
+  }
+#endif  // CONFIG_EXT_PARTITION_TYPES
+}
+
+static void free_tree_contexts(PC_TREE *tree) {
+#if CONFIG_EXT_PARTITION_TYPES
+  int i;
+  for (i = 0; i < 3; i++) {
+    free_mode_context(&tree->horizontala[i]);
+    free_mode_context(&tree->horizontalb[i]);
+    free_mode_context(&tree->verticala[i]);
+    free_mode_context(&tree->verticalb[i]);
+  }
+#endif  // CONFIG_EXT_PARTITION_TYPES
+  free_mode_context(&tree->none);
+  free_mode_context(&tree->horizontal[0]);
+  free_mode_context(&tree->horizontal[1]);
+  free_mode_context(&tree->vertical[0]);
+  free_mode_context(&tree->vertical[1]);
+#ifdef CONFIG_SUPERTX
+  free_mode_context(&tree->horizontal_supertx);
+  free_mode_context(&tree->vertical_supertx);
+  free_mode_context(&tree->split_supertx);
+#if CONFIG_EXT_PARTITION_TYPES
+  free_mode_context(&tree->horizontala_supertx);
+  free_mode_context(&tree->horizontalb_supertx);
+  free_mode_context(&tree->verticala_supertx);
+  free_mode_context(&tree->verticalb_supertx);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+#endif  // CONFIG_SUPERTX
+}
+
+// This function sets up a tree of contexts such that at each square
+// partition level. There are contexts for none, horizontal, vertical, and
+// split.  Along with a block_size value and a selected block_size which
+// represents the state of our search.
+void av1_setup_pc_tree(AV1_COMMON *cm, ThreadData *td) {
+  int i, j;
+// TODO(jingning): The pc_tree allocation is redundant. We can take out all
+// the leaf nodes after cb4x4 mode is enabled.
+#if CONFIG_CB4X4
+#if CONFIG_EXT_PARTITION
+  const int tree_nodes_inc = 1024;
+#else
+  const int tree_nodes_inc = 256;
+#endif  // CONFIG_EXT_PARTITION
+  const int leaf_factor = 4;
+#else
+  const int tree_nodes_inc = 0;
+  const int leaf_factor = 1;
+#endif
+#if CONFIG_EXT_PARTITION
+  const int leaf_nodes = 256 * leaf_factor;
+  const int tree_nodes = tree_nodes_inc + 256 + 64 + 16 + 4 + 1;
+#else
+  const int leaf_nodes = 64 * leaf_factor;
+  const int tree_nodes = tree_nodes_inc + 64 + 16 + 4 + 1;
+#endif  // CONFIG_EXT_PARTITION
+  int pc_tree_index = 0;
+  PC_TREE *this_pc;
+  PICK_MODE_CONTEXT *this_leaf;
+  int square_index = 1;
+  int nodes;
+
+  aom_free(td->leaf_tree);
+  CHECK_MEM_ERROR(cm, td->leaf_tree,
+                  aom_calloc(leaf_nodes, sizeof(*td->leaf_tree)));
+  aom_free(td->pc_tree);
+  CHECK_MEM_ERROR(cm, td->pc_tree,
+                  aom_calloc(tree_nodes, sizeof(*td->pc_tree)));
+
+  this_pc = &td->pc_tree[0];
+  this_leaf = &td->leaf_tree[0];
+
+  // 4x4 blocks smaller than 8x8 but in the same 8x8 block share the same
+  // context so we only need to allocate 1 for each 8x8 block.
+  for (i = 0; i < leaf_nodes; ++i) {
+#if CONFIG_EXT_PARTITION_TYPES
+    alloc_mode_context(cm, 4, PARTITION_NONE, &td->leaf_tree[i]);
+#else
+    alloc_mode_context(cm, 16, &td->leaf_tree[i]);
+#endif
+  }
+
+  // Sets up all the leaf nodes in the tree.
+  for (pc_tree_index = 0; pc_tree_index < leaf_nodes; ++pc_tree_index) {
+    PC_TREE *const tree = &td->pc_tree[pc_tree_index];
+    tree->block_size = square[0];
+#if CONFIG_CB4X4
+    alloc_tree_contexts(cm, tree, 16);
+#else
+    alloc_tree_contexts(cm, tree, 4);
+#endif
+    tree->leaf_split[0] = this_leaf++;
+    for (j = 1; j < 4; j++) tree->leaf_split[j] = tree->leaf_split[0];
+  }
+
+  // Each node has 4 leaf nodes, fill each block_size level of the tree
+  // from leafs to the root.
+  for (nodes = leaf_nodes >> 2; nodes > 0; nodes >>= 2) {
+    for (i = 0; i < nodes; ++i) {
+      PC_TREE *const tree = &td->pc_tree[pc_tree_index];
+#if CONFIG_CB4X4
+      alloc_tree_contexts(cm, tree, 16 << (2 * square_index));
+#else
+      alloc_tree_contexts(cm, tree, 4 << (2 * square_index));
+#endif
+      tree->block_size = square[square_index];
+      for (j = 0; j < 4; j++) tree->split[j] = this_pc++;
+      ++pc_tree_index;
+    }
+    ++square_index;
+  }
+
+  // Set up the root node for the largest superblock size
+  i = MAX_MIB_SIZE_LOG2 - MIN_MIB_SIZE_LOG2;
+  td->pc_root[i] = &td->pc_tree[tree_nodes - 1];
+  td->pc_root[i]->none.best_mode_index = 2;
+  // Set up the root nodes for the rest of the possible superblock sizes
+  while (--i >= 0) {
+    td->pc_root[i] = td->pc_root[i + 1]->split[0];
+    td->pc_root[i]->none.best_mode_index = 2;
+  }
+}
+
+void av1_free_pc_tree(ThreadData *td) {
+#if CONFIG_CB4X4
+#if CONFIG_EXT_PARTITION
+  const int tree_nodes_inc = 1024;
+#else
+  const int tree_nodes_inc = 256;
+#endif  // CONFIG_EXT_PARTITION
+  const int leaf_factor = 4;
+#else
+  const int tree_nodes_inc = 0;
+  const int leaf_factor = 1;
+#endif
+
+#if CONFIG_EXT_PARTITION
+  const int leaf_nodes = 256 * leaf_factor;
+  const int tree_nodes = tree_nodes_inc + 256 + 64 + 16 + 4 + 1;
+#else
+  const int leaf_nodes = 64 * leaf_factor;
+  const int tree_nodes = tree_nodes_inc + 64 + 16 + 4 + 1;
+#endif  // CONFIG_EXT_PARTITION
+  int i;
+
+  // Set up all 4x4 mode contexts
+  for (i = 0; i < leaf_nodes; ++i) free_mode_context(&td->leaf_tree[i]);
+
+  // Sets up all the leaf nodes in the tree.
+  for (i = 0; i < tree_nodes; ++i) free_tree_contexts(&td->pc_tree[i]);
+
+  aom_free(td->pc_tree);
+  td->pc_tree = NULL;
+  aom_free(td->leaf_tree);
+  td->leaf_tree = NULL;
+}
diff --git a/third_party/aom/av1/encoder/context_tree.h b/third_party/aom/av1/encoder/context_tree.h
new file mode 100644
index 0000000000..67954126c6
--- /dev/null
+++ b/third_party/aom/av1/encoder/context_tree.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_CONTEXT_TREE_H_
+#define AV1_ENCODER_CONTEXT_TREE_H_
+
+#include "av1/common/blockd.h"
+#include "av1/encoder/block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1_COMP;
+struct AV1Common;
+struct ThreadData;
+
+// Structure to hold snapshot of coding context during the mode picking process
+typedef struct {
+  MODE_INFO mic;
+  MB_MODE_INFO_EXT mbmi_ext;
+#if CONFIG_PALETTE
+  uint8_t *color_index_map[2];
+#endif  // CONFIG_PALETTE
+#if CONFIG_VAR_TX
+  uint8_t *blk_skip[MAX_MB_PLANE];
+#endif
+
+  // dual buffer pointers, 0: in use, 1: best in store
+  tran_low_t *coeff[MAX_MB_PLANE];
+  tran_low_t *qcoeff[MAX_MB_PLANE];
+  tran_low_t *dqcoeff[MAX_MB_PLANE];
+#if CONFIG_PVQ
+  tran_low_t *pvq_ref_coeff[MAX_MB_PLANE];
+#endif
+  uint16_t *eobs[MAX_MB_PLANE];
+#if CONFIG_LV_MAP
+  uint8_t *txb_entropy_ctx[MAX_MB_PLANE];
+#endif
+
+  int num_4x4_blk;
+  int skip;
+  int pred_pixel_ready;
+  // For current partition, only if all Y, U, and V transform blocks'
+  // coefficients are quantized to 0, skippable is set to 0.
+  int skippable;
+  int best_mode_index;
+  int hybrid_pred_diff;
+  int comp_pred_diff;
+  int single_pred_diff;
+
+  // TODO(jingning) Use RD_COST struct here instead. This involves a boarder
+  // scope of refactoring.
+  int rate;
+  int64_t dist;
+
+  // motion vector cache for adaptive motion search control in partition
+  // search loop
+  MV pred_mv[TOTAL_REFS_PER_FRAME];
+  InterpFilter pred_interp_filter;
+#if CONFIG_EXT_PARTITION_TYPES
+  PARTITION_TYPE partition;
+#endif
+} PICK_MODE_CONTEXT;
+
+typedef struct PC_TREE {
+  int index;
+  PARTITION_TYPE partitioning;
+  BLOCK_SIZE block_size;
+  PICK_MODE_CONTEXT none;
+  PICK_MODE_CONTEXT horizontal[2];
+  PICK_MODE_CONTEXT vertical[2];
+#if CONFIG_EXT_PARTITION_TYPES
+  PICK_MODE_CONTEXT horizontala[3];
+  PICK_MODE_CONTEXT horizontalb[3];
+  PICK_MODE_CONTEXT verticala[3];
+  PICK_MODE_CONTEXT verticalb[3];
+#endif
+  union {
+    struct PC_TREE *split[4];
+    PICK_MODE_CONTEXT *leaf_split[4];
+  };
+#ifdef CONFIG_SUPERTX
+  PICK_MODE_CONTEXT horizontal_supertx;
+  PICK_MODE_CONTEXT vertical_supertx;
+  PICK_MODE_CONTEXT split_supertx;
+#if CONFIG_EXT_PARTITION_TYPES
+  PICK_MODE_CONTEXT horizontala_supertx;
+  PICK_MODE_CONTEXT horizontalb_supertx;
+  PICK_MODE_CONTEXT verticala_supertx;
+  PICK_MODE_CONTEXT verticalb_supertx;
+#endif
+#endif
+} PC_TREE;
+
+void av1_setup_pc_tree(struct AV1Common *cm, struct ThreadData *td);
+void av1_free_pc_tree(struct ThreadData *td);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif /* AV1_ENCODER_CONTEXT_TREE_H_ */
diff --git a/third_party/aom/av1/encoder/corner_detect.c b/third_party/aom/av1/encoder/corner_detect.c
new file mode 100644
index 0000000000..e4c59dd9c6
--- /dev/null
+++ b/third_party/aom/av1/encoder/corner_detect.c
@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <memory.h>
+#include <math.h>
+#include <assert.h>
+
+#include "third_party/fastfeat/fast.h"
+
+#include "av1/encoder/corner_detect.h"
+
+// Fast_9 wrapper
+#define FAST_BARRIER 18
+int fast_corner_detect(unsigned char *buf, int width, int height, int stride,
+                       int *points, int max_points) {
+  int num_points;
+  xy *const frm_corners_xy = fast9_detect_nonmax(buf, width, height, stride,
+                                                 FAST_BARRIER, &num_points);
+  num_points = (num_points <= max_points ? num_points : max_points);
+  if (num_points > 0 && frm_corners_xy) {
+    memcpy(points, frm_corners_xy, sizeof(*frm_corners_xy) * num_points);
+    free(frm_corners_xy);
+    return num_points;
+  }
+  free(frm_corners_xy);
+  return 0;
+}
diff --git a/third_party/aom/av1/encoder/corner_detect.h b/third_party/aom/av1/encoder/corner_detect.h
new file mode 100644
index 0000000000..0317db5b31
--- /dev/null
+++ b/third_party/aom/av1/encoder/corner_detect.h
@@ -0,0 +1,22 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_CORNER_DETECT_H_
+#define AV1_ENCODER_CORNER_DETECT_H_
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+
+int fast_corner_detect(unsigned char *buf, int width, int height, int stride,
+                       int *points, int max_points);
+
+#endif  // AV1_ENCODER_CORNER_DETECT_H_
diff --git a/third_party/aom/av1/encoder/corner_match.c b/third_party/aom/av1/encoder/corner_match.c
new file mode 100644
index 0000000000..64ee0c5ae1
--- /dev/null
+++ b/third_party/aom/av1/encoder/corner_match.c
@@ -0,0 +1,193 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+#include <math.h>
+
+#include "av1/encoder/corner_match.h"
+
+#define MATCH_SZ 13
+#define MATCH_SZ_BY2 ((MATCH_SZ - 1) / 2)
+#define MATCH_SZ_SQ (MATCH_SZ * MATCH_SZ)
+#define SEARCH_SZ 9
+#define SEARCH_SZ_BY2 ((SEARCH_SZ - 1) / 2)
+
+#define THRESHOLD_NCC 0.75
+
+/* Compute var(im) * MATCH_SZ_SQ over a MATCH_SZ by MATCH_SZ window of im,
+   centered at (x, y).
+*/
+static double compute_variance(unsigned char *im, int stride, int x, int y) {
+  int sum = 0.0;
+  int sumsq = 0.0;
+  int var;
+  int i, j;
+  for (i = 0; i < MATCH_SZ; ++i)
+    for (j = 0; j < MATCH_SZ; ++j) {
+      sum += im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)];
+      sumsq += im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)] *
+               im[(i + y - MATCH_SZ_BY2) * stride + (j + x - MATCH_SZ_BY2)];
+    }
+  var = sumsq * MATCH_SZ_SQ - sum * sum;
+  return (double)var;
+}
+
+/* Compute corr(im1, im2) * MATCH_SZ * stddev(im1), where the
+   correlation/standard deviation are taken over MATCH_SZ by MATCH_SZ windows
+   of each image, centered at (x1, y1) and (x2, y2) respectively.
+*/
+static double compute_cross_correlation(unsigned char *im1, int stride1, int x1,
+                                        int y1, unsigned char *im2, int stride2,
+                                        int x2, int y2) {
+  int v1, v2;
+  int sum1 = 0;
+  int sum2 = 0;
+  int sumsq2 = 0;
+  int cross = 0;
+  int var2, cov;
+  int i, j;
+  for (i = 0; i < MATCH_SZ; ++i)
+    for (j = 0; j < MATCH_SZ; ++j) {
+      v1 = im1[(i + y1 - MATCH_SZ_BY2) * stride1 + (j + x1 - MATCH_SZ_BY2)];
+      v2 = im2[(i + y2 - MATCH_SZ_BY2) * stride2 + (j + x2 - MATCH_SZ_BY2)];
+      sum1 += v1;
+      sum2 += v2;
+      sumsq2 += v2 * v2;
+      cross += v1 * v2;
+    }
+  var2 = sumsq2 * MATCH_SZ_SQ - sum2 * sum2;
+  cov = cross * MATCH_SZ_SQ - sum1 * sum2;
+  return cov / sqrt((double)var2);
+}
+
+static int is_eligible_point(int pointx, int pointy, int width, int height) {
+  return (pointx >= MATCH_SZ_BY2 && pointy >= MATCH_SZ_BY2 &&
+          pointx + MATCH_SZ_BY2 < width && pointy + MATCH_SZ_BY2 < height);
+}
+
+static int is_eligible_distance(int point1x, int point1y, int point2x,
+                                int point2y, int width, int height) {
+  const int thresh = (width < height ? height : width) >> 4;
+  return ((point1x - point2x) * (point1x - point2x) +
+          (point1y - point2y) * (point1y - point2y)) <= thresh * thresh;
+}
+
+static void improve_correspondence(unsigned char *frm, unsigned char *ref,
+                                   int width, int height, int frm_stride,
+                                   int ref_stride,
+                                   Correspondence *correspondences,
+                                   int num_correspondences) {
+  int i;
+  for (i = 0; i < num_correspondences; ++i) {
+    int x, y, best_x = 0, best_y = 0;
+    double best_match_ncc = 0.0;
+    for (y = -SEARCH_SZ_BY2; y <= SEARCH_SZ_BY2; ++y) {
+      for (x = -SEARCH_SZ_BY2; x <= SEARCH_SZ_BY2; ++x) {
+        double match_ncc;
+        if (!is_eligible_point(correspondences[i].rx + x,
+                               correspondences[i].ry + y, width, height))
+          continue;
+        if (!is_eligible_distance(correspondences[i].x, correspondences[i].y,
+                                  correspondences[i].rx + x,
+                                  correspondences[i].ry + y, width, height))
+          continue;
+        match_ncc = compute_cross_correlation(
+            frm, frm_stride, correspondences[i].x, correspondences[i].y, ref,
+            ref_stride, correspondences[i].rx + x, correspondences[i].ry + y);
+        if (match_ncc > best_match_ncc) {
+          best_match_ncc = match_ncc;
+          best_y = y;
+          best_x = x;
+        }
+      }
+    }
+    correspondences[i].rx += best_x;
+    correspondences[i].ry += best_y;
+  }
+  for (i = 0; i < num_correspondences; ++i) {
+    int x, y, best_x = 0, best_y = 0;
+    double best_match_ncc = 0.0;
+    for (y = -SEARCH_SZ_BY2; y <= SEARCH_SZ_BY2; ++y)
+      for (x = -SEARCH_SZ_BY2; x <= SEARCH_SZ_BY2; ++x) {
+        double match_ncc;
+        if (!is_eligible_point(correspondences[i].x + x,
+                               correspondences[i].y + y, width, height))
+          continue;
+        if (!is_eligible_distance(
+                correspondences[i].x + x, correspondences[i].y + y,
+                correspondences[i].rx, correspondences[i].ry, width, height))
+          continue;
+        match_ncc = compute_cross_correlation(
+            ref, ref_stride, correspondences[i].rx, correspondences[i].ry, frm,
+            frm_stride, correspondences[i].x + x, correspondences[i].y + y);
+        if (match_ncc > best_match_ncc) {
+          best_match_ncc = match_ncc;
+          best_y = y;
+          best_x = x;
+        }
+      }
+    correspondences[i].x += best_x;
+    correspondences[i].y += best_y;
+  }
+}
+
+int determine_correspondence(unsigned char *frm, int *frm_corners,
+                             int num_frm_corners, unsigned char *ref,
+                             int *ref_corners, int num_ref_corners, int width,
+                             int height, int frm_stride, int ref_stride,
+                             int *correspondence_pts) {
+  // TODO(sarahparker) Improve this to include 2-way match
+  int i, j;
+  Correspondence *correspondences = (Correspondence *)correspondence_pts;
+  int num_correspondences = 0;
+  for (i = 0; i < num_frm_corners; ++i) {
+    double best_match_ncc = 0.0;
+    double template_norm;
+    int best_match_j = -1;
+    if (!is_eligible_point(frm_corners[2 * i], frm_corners[2 * i + 1], width,
+                           height))
+      continue;
+    for (j = 0; j < num_ref_corners; ++j) {
+      double match_ncc;
+      if (!is_eligible_point(ref_corners[2 * j], ref_corners[2 * j + 1], width,
+                             height))
+        continue;
+      if (!is_eligible_distance(frm_corners[2 * i], frm_corners[2 * i + 1],
+                                ref_corners[2 * j], ref_corners[2 * j + 1],
+                                width, height))
+        continue;
+      match_ncc = compute_cross_correlation(
+          frm, frm_stride, frm_corners[2 * i], frm_corners[2 * i + 1], ref,
+          ref_stride, ref_corners[2 * j], ref_corners[2 * j + 1]);
+      if (match_ncc > best_match_ncc) {
+        best_match_ncc = match_ncc;
+        best_match_j = j;
+      }
+    }
+    // Note: We want to test if the best correlation is >= THRESHOLD_NCC,
+    // but need to account for the normalization in compute_cross_correlation.
+    template_norm = compute_variance(frm, frm_stride, frm_corners[2 * i],
+                                     frm_corners[2 * i + 1]);
+    if (best_match_ncc > THRESHOLD_NCC * sqrt(template_norm)) {
+      correspondences[num_correspondences].x = frm_corners[2 * i];
+      correspondences[num_correspondences].y = frm_corners[2 * i + 1];
+      correspondences[num_correspondences].rx = ref_corners[2 * best_match_j];
+      correspondences[num_correspondences].ry =
+          ref_corners[2 * best_match_j + 1];
+      num_correspondences++;
+    }
+  }
+  improve_correspondence(frm, ref, width, height, frm_stride, ref_stride,
+                         correspondences, num_correspondences);
+  return num_correspondences;
+}
diff --git a/third_party/aom/av1/encoder/corner_match.h b/third_party/aom/av1/encoder/corner_match.h
new file mode 100644
index 0000000000..c0458642c1
--- /dev/null
+++ b/third_party/aom/av1/encoder/corner_match.h
@@ -0,0 +1,29 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AV1_ENCODER_CORNER_MATCH_H_
+#define AV1_ENCODER_CORNER_MATCH_H_
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+
+typedef struct {
+  int x, y;
+  int rx, ry;
+} Correspondence;
+
+int determine_correspondence(unsigned char *frm, int *frm_corners,
+                             int num_frm_corners, unsigned char *ref,
+                             int *ref_corners, int num_ref_corners, int width,
+                             int height, int frm_stride, int ref_stride,
+                             int *correspondence_pts);
+
+#endif  // AV1_ENCODER_CORNER_MATCH_H_
diff --git a/third_party/aom/av1/encoder/cost.c b/third_party/aom/av1/encoder/cost.c
new file mode 100644
index 0000000000..e3151a5973
--- /dev/null
+++ b/third_party/aom/av1/encoder/cost.c
@@ -0,0 +1,67 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <assert.h>
+
+#include "av1/encoder/cost.h"
+#include "av1/common/entropy.h"
+
+/* round(-log2(i/256.) * (1 << AV1_PROB_COST_SHIFT))
+   Begins with a bogus entry for simpler addressing. */
+const uint16_t av1_prob_cost[256] = {
+  4096, 4096, 3584, 3284, 3072, 2907, 2772, 2659, 2560, 2473, 2395, 2325, 2260,
+  2201, 2147, 2096, 2048, 2003, 1961, 1921, 1883, 1847, 1813, 1780, 1748, 1718,
+  1689, 1661, 1635, 1609, 1584, 1559, 1536, 1513, 1491, 1470, 1449, 1429, 1409,
+  1390, 1371, 1353, 1335, 1318, 1301, 1284, 1268, 1252, 1236, 1221, 1206, 1192,
+  1177, 1163, 1149, 1136, 1123, 1110, 1097, 1084, 1072, 1059, 1047, 1036, 1024,
+  1013, 1001, 990,  979,  968,  958,  947,  937,  927,  917,  907,  897,  887,
+  878,  868,  859,  850,  841,  832,  823,  814,  806,  797,  789,  780,  772,
+  764,  756,  748,  740,  732,  724,  717,  709,  702,  694,  687,  680,  673,
+  665,  658,  651,  644,  637,  631,  624,  617,  611,  604,  598,  591,  585,
+  578,  572,  566,  560,  554,  547,  541,  535,  530,  524,  518,  512,  506,
+  501,  495,  489,  484,  478,  473,  467,  462,  456,  451,  446,  441,  435,
+  430,  425,  420,  415,  410,  405,  400,  395,  390,  385,  380,  375,  371,
+  366,  361,  356,  352,  347,  343,  338,  333,  329,  324,  320,  316,  311,
+  307,  302,  298,  294,  289,  285,  281,  277,  273,  268,  264,  260,  256,
+  252,  248,  244,  240,  236,  232,  228,  224,  220,  216,  212,  209,  205,
+  201,  197,  194,  190,  186,  182,  179,  175,  171,  168,  164,  161,  157,
+  153,  150,  146,  143,  139,  136,  132,  129,  125,  122,  119,  115,  112,
+  109,  105,  102,  99,   95,   92,   89,   86,   82,   79,   76,   73,   70,
+  66,   63,   60,   57,   54,   51,   48,   45,   42,   38,   35,   32,   29,
+  26,   23,   20,   18,   15,   12,   9,    6,    3
+};
+
+static void cost(int *costs, aom_tree tree, const aom_prob *probs, int i,
+                 int c) {
+  const aom_prob prob = probs[i / 2];
+  int b;
+
+  assert(prob != 0);
+  for (b = 0; b <= 1; ++b) {
+    const int cc = c + av1_cost_bit(prob, b);
+    const aom_tree_index ii = tree[i + b];
+
+    if (ii <= 0)
+      costs[-ii] = cc;
+    else
+      cost(costs, tree, probs, ii, cc);
+  }
+}
+
+void av1_cost_tokens(int *costs, const aom_prob *probs, aom_tree tree) {
+  cost(costs, tree, probs, 0, 0);
+}
+
+void av1_cost_tokens_skip(int *costs, const aom_prob *probs, aom_tree tree) {
+  assert(tree[0] <= 0 && tree[1] > 0);
+
+  costs[-tree[0]] = av1_cost_bit(probs[0], 0);
+  cost(costs, tree, probs, 2, 0);
+}
diff --git a/third_party/aom/av1/encoder/cost.h b/third_party/aom/av1/encoder/cost.h
new file mode 100644
index 0000000000..d8fb357e6d
--- /dev/null
+++ b/third_party/aom/av1/encoder/cost.h
@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_COST_H_
+#define AV1_ENCODER_COST_H_
+
+#include "aom_dsp/prob.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern const uint16_t av1_prob_cost[256];
+
+// The factor to scale from cost in bits to cost in av1_prob_cost units.
+#define AV1_PROB_COST_SHIFT 9
+
+#define av1_cost_zero(prob) (av1_prob_cost[prob])
+
+#define av1_cost_one(prob) av1_cost_zero(256 - (prob))
+
+#define av1_cost_bit(prob, bit) av1_cost_zero((bit) ? 256 - (prob) : (prob))
+
+// Cost of coding an n bit literal, using 128 (i.e. 50%) probability
+// for each bit.
+#define av1_cost_literal(n) ((n) * (1 << AV1_PROB_COST_SHIFT))
+
+static INLINE unsigned int cost_branch256(const unsigned int ct[2],
+                                          aom_prob p) {
+  return ct[0] * av1_cost_zero(p) + ct[1] * av1_cost_one(p);
+}
+
+static INLINE int treed_cost(aom_tree tree, const aom_prob *probs, int bits,
+                             int len) {
+  int cost = 0;
+  aom_tree_index i = 0;
+
+  do {
+    const int bit = (bits >> --len) & 1;
+    cost += av1_cost_bit(probs[i >> 1], bit);
+    i = tree[i + bit];
+  } while (len);
+
+  return cost;
+}
+
+void av1_cost_tokens(int *costs, const aom_prob *probs, aom_tree tree);
+void av1_cost_tokens_skip(int *costs, const aom_prob *probs, aom_tree tree);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_COST_H_
diff --git a/third_party/aom/av1/encoder/daala_compat_enc.c b/third_party/aom/av1/encoder/daala_compat_enc.c
new file mode 100644
index 0000000000..3df424cac2
--- /dev/null
+++ b/third_party/aom/av1/encoder/daala_compat_enc.c
@@ -0,0 +1,30 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "encint.h"
+
+void od_encode_checkpoint(const daala_enc_ctx *enc, od_rollback_buffer *rbuf) {
+#if CONFIG_DAALA_EC
+  od_ec_enc_checkpoint(&rbuf->ec, &enc->w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  OD_COPY(&rbuf->adapt, enc->state.adapt, 1);
+}
+
+void od_encode_rollback(daala_enc_ctx *enc, const od_rollback_buffer *rbuf) {
+#if CONFIG_DAALA_EC
+  od_ec_enc_rollback(&enc->w.ec, &rbuf->ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  OD_COPY(enc->state.adapt, &rbuf->adapt, 1);
+}
diff --git a/third_party/aom/av1/encoder/dct.c b/third_party/aom/av1/encoder/dct.c
new file mode 100644
index 0000000000..09e1b05637
--- /dev/null
+++ b/third_party/aom/av1/encoder/dct.c
@@ -0,0 +1,2228 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+#include "aom_dsp/fwd_txfm.h"
+#include "aom_ports/mem.h"
+#include "av1/common/blockd.h"
+#include "av1/common/av1_fwd_txfm1d.h"
+#include "av1/common/av1_fwd_txfm2d_cfg.h"
+#include "av1/common/idct.h"
+
+static INLINE void range_check(const tran_low_t *input, const int size,
+                               const int bit) {
+#if 0  // CONFIG_COEFFICIENT_RANGE_CHECKING
+// TODO(angiebird): the range_check is not used because the bit range
+// in fdct# is not correct. Since we are going to merge in a new version
+// of fdct# from nextgenv2, we won't fix the incorrect bit range now.
+  int i;
+  for (i = 0; i < size; ++i) {
+    assert(abs(input[i]) < (1 << bit));
+  }
+#else
+  (void)input;
+  (void)size;
+  (void)bit;
+#endif
+}
+
+static void fdct4(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t temp;
+  tran_low_t step[4];
+
+  // stage 0
+  range_check(input, 4, 14);
+
+  // stage 1
+  output[0] = input[0] + input[3];
+  output[1] = input[1] + input[2];
+  output[2] = input[1] - input[2];
+  output[3] = input[0] - input[3];
+
+  range_check(output, 4, 15);
+
+  // stage 2
+  temp = output[0] * cospi_16_64 + output[1] * cospi_16_64;
+  step[0] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[1] * -cospi_16_64 + output[0] * cospi_16_64;
+  step[1] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[2] * cospi_24_64 + output[3] * cospi_8_64;
+  step[2] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[3] * cospi_24_64 + output[2] * -cospi_8_64;
+  step[3] = (tran_low_t)fdct_round_shift(temp);
+
+  range_check(step, 4, 16);
+
+  // stage 3
+  output[0] = step[0];
+  output[1] = step[2];
+  output[2] = step[1];
+  output[3] = step[3];
+
+  range_check(output, 4, 16);
+}
+
+static void fdct8(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t temp;
+  tran_low_t step[8];
+
+  // stage 0
+  range_check(input, 8, 13);
+
+  // stage 1
+  output[0] = input[0] + input[7];
+  output[1] = input[1] + input[6];
+  output[2] = input[2] + input[5];
+  output[3] = input[3] + input[4];
+  output[4] = input[3] - input[4];
+  output[5] = input[2] - input[5];
+  output[6] = input[1] - input[6];
+  output[7] = input[0] - input[7];
+
+  range_check(output, 8, 14);
+
+  // stage 2
+  step[0] = output[0] + output[3];
+  step[1] = output[1] + output[2];
+  step[2] = output[1] - output[2];
+  step[3] = output[0] - output[3];
+  step[4] = output[4];
+  temp = output[5] * -cospi_16_64 + output[6] * cospi_16_64;
+  step[5] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[6] * cospi_16_64 + output[5] * cospi_16_64;
+  step[6] = (tran_low_t)fdct_round_shift(temp);
+  step[7] = output[7];
+
+  range_check(step, 8, 15);
+
+  // stage 3
+  temp = step[0] * cospi_16_64 + step[1] * cospi_16_64;
+  output[0] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[1] * -cospi_16_64 + step[0] * cospi_16_64;
+  output[1] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[2] * cospi_24_64 + step[3] * cospi_8_64;
+  output[2] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[3] * cospi_24_64 + step[2] * -cospi_8_64;
+  output[3] = (tran_low_t)fdct_round_shift(temp);
+  output[4] = step[4] + step[5];
+  output[5] = step[4] - step[5];
+  output[6] = step[7] - step[6];
+  output[7] = step[7] + step[6];
+
+  range_check(output, 8, 16);
+
+  // stage 4
+  step[0] = output[0];
+  step[1] = output[1];
+  step[2] = output[2];
+  step[3] = output[3];
+  temp = output[4] * cospi_28_64 + output[7] * cospi_4_64;
+  step[4] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[5] * cospi_12_64 + output[6] * cospi_20_64;
+  step[5] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[6] * cospi_12_64 + output[5] * -cospi_20_64;
+  step[6] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[7] * cospi_28_64 + output[4] * -cospi_4_64;
+  step[7] = (tran_low_t)fdct_round_shift(temp);
+
+  range_check(step, 8, 16);
+
+  // stage 5
+  output[0] = step[0];
+  output[1] = step[4];
+  output[2] = step[2];
+  output[3] = step[6];
+  output[4] = step[1];
+  output[5] = step[5];
+  output[6] = step[3];
+  output[7] = step[7];
+
+  range_check(output, 8, 16);
+}
+
+static void fdct16(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t temp;
+  tran_low_t step[16];
+
+  // stage 0
+  range_check(input, 16, 13);
+
+  // stage 1
+  output[0] = input[0] + input[15];
+  output[1] = input[1] + input[14];
+  output[2] = input[2] + input[13];
+  output[3] = input[3] + input[12];
+  output[4] = input[4] + input[11];
+  output[5] = input[5] + input[10];
+  output[6] = input[6] + input[9];
+  output[7] = input[7] + input[8];
+  output[8] = input[7] - input[8];
+  output[9] = input[6] - input[9];
+  output[10] = input[5] - input[10];
+  output[11] = input[4] - input[11];
+  output[12] = input[3] - input[12];
+  output[13] = input[2] - input[13];
+  output[14] = input[1] - input[14];
+  output[15] = input[0] - input[15];
+
+  range_check(output, 16, 14);
+
+  // stage 2
+  step[0] = output[0] + output[7];
+  step[1] = output[1] + output[6];
+  step[2] = output[2] + output[5];
+  step[3] = output[3] + output[4];
+  step[4] = output[3] - output[4];
+  step[5] = output[2] - output[5];
+  step[6] = output[1] - output[6];
+  step[7] = output[0] - output[7];
+  step[8] = output[8];
+  step[9] = output[9];
+  temp = output[10] * -cospi_16_64 + output[13] * cospi_16_64;
+  step[10] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[11] * -cospi_16_64 + output[12] * cospi_16_64;
+  step[11] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[12] * cospi_16_64 + output[11] * cospi_16_64;
+  step[12] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[13] * cospi_16_64 + output[10] * cospi_16_64;
+  step[13] = (tran_low_t)fdct_round_shift(temp);
+  step[14] = output[14];
+  step[15] = output[15];
+
+  range_check(step, 16, 15);
+
+  // stage 3
+  output[0] = step[0] + step[3];
+  output[1] = step[1] + step[2];
+  output[2] = step[1] - step[2];
+  output[3] = step[0] - step[3];
+  output[4] = step[4];
+  temp = step[5] * -cospi_16_64 + step[6] * cospi_16_64;
+  output[5] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[6] * cospi_16_64 + step[5] * cospi_16_64;
+  output[6] = (tran_low_t)fdct_round_shift(temp);
+  output[7] = step[7];
+  output[8] = step[8] + step[11];
+  output[9] = step[9] + step[10];
+  output[10] = step[9] - step[10];
+  output[11] = step[8] - step[11];
+  output[12] = step[15] - step[12];
+  output[13] = step[14] - step[13];
+  output[14] = step[14] + step[13];
+  output[15] = step[15] + step[12];
+
+  range_check(output, 16, 16);
+
+  // stage 4
+  temp = output[0] * cospi_16_64 + output[1] * cospi_16_64;
+  step[0] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[1] * -cospi_16_64 + output[0] * cospi_16_64;
+  step[1] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[2] * cospi_24_64 + output[3] * cospi_8_64;
+  step[2] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[3] * cospi_24_64 + output[2] * -cospi_8_64;
+  step[3] = (tran_low_t)fdct_round_shift(temp);
+  step[4] = output[4] + output[5];
+  step[5] = output[4] - output[5];
+  step[6] = output[7] - output[6];
+  step[7] = output[7] + output[6];
+  step[8] = output[8];
+  temp = output[9] * -cospi_8_64 + output[14] * cospi_24_64;
+  step[9] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[10] * -cospi_24_64 + output[13] * -cospi_8_64;
+  step[10] = (tran_low_t)fdct_round_shift(temp);
+  step[11] = output[11];
+  step[12] = output[12];
+  temp = output[13] * cospi_24_64 + output[10] * -cospi_8_64;
+  step[13] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[14] * cospi_8_64 + output[9] * cospi_24_64;
+  step[14] = (tran_low_t)fdct_round_shift(temp);
+  step[15] = output[15];
+
+  range_check(step, 16, 16);
+
+  // stage 5
+  output[0] = step[0];
+  output[1] = step[1];
+  output[2] = step[2];
+  output[3] = step[3];
+  temp = step[4] * cospi_28_64 + step[7] * cospi_4_64;
+  output[4] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[5] * cospi_12_64 + step[6] * cospi_20_64;
+  output[5] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[6] * cospi_12_64 + step[5] * -cospi_20_64;
+  output[6] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[7] * cospi_28_64 + step[4] * -cospi_4_64;
+  output[7] = (tran_low_t)fdct_round_shift(temp);
+  output[8] = step[8] + step[9];
+  output[9] = step[8] - step[9];
+  output[10] = step[11] - step[10];
+  output[11] = step[11] + step[10];
+  output[12] = step[12] + step[13];
+  output[13] = step[12] - step[13];
+  output[14] = step[15] - step[14];
+  output[15] = step[15] + step[14];
+
+  range_check(output, 16, 16);
+
+  // stage 6
+  step[0] = output[0];
+  step[1] = output[1];
+  step[2] = output[2];
+  step[3] = output[3];
+  step[4] = output[4];
+  step[5] = output[5];
+  step[6] = output[6];
+  step[7] = output[7];
+  temp = output[8] * cospi_30_64 + output[15] * cospi_2_64;
+  step[8] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[9] * cospi_14_64 + output[14] * cospi_18_64;
+  step[9] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[10] * cospi_22_64 + output[13] * cospi_10_64;
+  step[10] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[11] * cospi_6_64 + output[12] * cospi_26_64;
+  step[11] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[12] * cospi_6_64 + output[11] * -cospi_26_64;
+  step[12] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[13] * cospi_22_64 + output[10] * -cospi_10_64;
+  step[13] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[14] * cospi_14_64 + output[9] * -cospi_18_64;
+  step[14] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[15] * cospi_30_64 + output[8] * -cospi_2_64;
+  step[15] = (tran_low_t)fdct_round_shift(temp);
+
+  range_check(step, 16, 16);
+
+  // stage 7
+  output[0] = step[0];
+  output[1] = step[8];
+  output[2] = step[4];
+  output[3] = step[12];
+  output[4] = step[2];
+  output[5] = step[10];
+  output[6] = step[6];
+  output[7] = step[14];
+  output[8] = step[1];
+  output[9] = step[9];
+  output[10] = step[5];
+  output[11] = step[13];
+  output[12] = step[3];
+  output[13] = step[11];
+  output[14] = step[7];
+  output[15] = step[15];
+
+  range_check(output, 16, 16);
+}
+
+static void fdct32(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t temp;
+  tran_low_t step[32];
+
+  // stage 0
+  range_check(input, 32, 14);
+
+  // stage 1
+  output[0] = input[0] + input[31];
+  output[1] = input[1] + input[30];
+  output[2] = input[2] + input[29];
+  output[3] = input[3] + input[28];
+  output[4] = input[4] + input[27];
+  output[5] = input[5] + input[26];
+  output[6] = input[6] + input[25];
+  output[7] = input[7] + input[24];
+  output[8] = input[8] + input[23];
+  output[9] = input[9] + input[22];
+  output[10] = input[10] + input[21];
+  output[11] = input[11] + input[20];
+  output[12] = input[12] + input[19];
+  output[13] = input[13] + input[18];
+  output[14] = input[14] + input[17];
+  output[15] = input[15] + input[16];
+  output[16] = input[15] - input[16];
+  output[17] = input[14] - input[17];
+  output[18] = input[13] - input[18];
+  output[19] = input[12] - input[19];
+  output[20] = input[11] - input[20];
+  output[21] = input[10] - input[21];
+  output[22] = input[9] - input[22];
+  output[23] = input[8] - input[23];
+  output[24] = input[7] - input[24];
+  output[25] = input[6] - input[25];
+  output[26] = input[5] - input[26];
+  output[27] = input[4] - input[27];
+  output[28] = input[3] - input[28];
+  output[29] = input[2] - input[29];
+  output[30] = input[1] - input[30];
+  output[31] = input[0] - input[31];
+
+  range_check(output, 32, 15);
+
+  // stage 2
+  step[0] = output[0] + output[15];
+  step[1] = output[1] + output[14];
+  step[2] = output[2] + output[13];
+  step[3] = output[3] + output[12];
+  step[4] = output[4] + output[11];
+  step[5] = output[5] + output[10];
+  step[6] = output[6] + output[9];
+  step[7] = output[7] + output[8];
+  step[8] = output[7] - output[8];
+  step[9] = output[6] - output[9];
+  step[10] = output[5] - output[10];
+  step[11] = output[4] - output[11];
+  step[12] = output[3] - output[12];
+  step[13] = output[2] - output[13];
+  step[14] = output[1] - output[14];
+  step[15] = output[0] - output[15];
+  step[16] = output[16];
+  step[17] = output[17];
+  step[18] = output[18];
+  step[19] = output[19];
+  temp = output[20] * -cospi_16_64 + output[27] * cospi_16_64;
+  step[20] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[21] * -cospi_16_64 + output[26] * cospi_16_64;
+  step[21] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[22] * -cospi_16_64 + output[25] * cospi_16_64;
+  step[22] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[23] * -cospi_16_64 + output[24] * cospi_16_64;
+  step[23] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[24] * cospi_16_64 + output[23] * cospi_16_64;
+  step[24] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[25] * cospi_16_64 + output[22] * cospi_16_64;
+  step[25] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[26] * cospi_16_64 + output[21] * cospi_16_64;
+  step[26] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[27] * cospi_16_64 + output[20] * cospi_16_64;
+  step[27] = (tran_low_t)fdct_round_shift(temp);
+  step[28] = output[28];
+  step[29] = output[29];
+  step[30] = output[30];
+  step[31] = output[31];
+
+  range_check(step, 32, 16);
+
+  // stage 3
+  output[0] = step[0] + step[7];
+  output[1] = step[1] + step[6];
+  output[2] = step[2] + step[5];
+  output[3] = step[3] + step[4];
+  output[4] = step[3] - step[4];
+  output[5] = step[2] - step[5];
+  output[6] = step[1] - step[6];
+  output[7] = step[0] - step[7];
+  output[8] = step[8];
+  output[9] = step[9];
+  temp = step[10] * -cospi_16_64 + step[13] * cospi_16_64;
+  output[10] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[11] * -cospi_16_64 + step[12] * cospi_16_64;
+  output[11] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[12] * cospi_16_64 + step[11] * cospi_16_64;
+  output[12] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[13] * cospi_16_64 + step[10] * cospi_16_64;
+  output[13] = (tran_low_t)fdct_round_shift(temp);
+  output[14] = step[14];
+  output[15] = step[15];
+  output[16] = step[16] + step[23];
+  output[17] = step[17] + step[22];
+  output[18] = step[18] + step[21];
+  output[19] = step[19] + step[20];
+  output[20] = step[19] - step[20];
+  output[21] = step[18] - step[21];
+  output[22] = step[17] - step[22];
+  output[23] = step[16] - step[23];
+  output[24] = step[31] - step[24];
+  output[25] = step[30] - step[25];
+  output[26] = step[29] - step[26];
+  output[27] = step[28] - step[27];
+  output[28] = step[28] + step[27];
+  output[29] = step[29] + step[26];
+  output[30] = step[30] + step[25];
+  output[31] = step[31] + step[24];
+
+  range_check(output, 32, 17);
+
+  // stage 4
+  step[0] = output[0] + output[3];
+  step[1] = output[1] + output[2];
+  step[2] = output[1] - output[2];
+  step[3] = output[0] - output[3];
+  step[4] = output[4];
+  temp = output[5] * -cospi_16_64 + output[6] * cospi_16_64;
+  step[5] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[6] * cospi_16_64 + output[5] * cospi_16_64;
+  step[6] = (tran_low_t)fdct_round_shift(temp);
+  step[7] = output[7];
+  step[8] = output[8] + output[11];
+  step[9] = output[9] + output[10];
+  step[10] = output[9] - output[10];
+  step[11] = output[8] - output[11];
+  step[12] = output[15] - output[12];
+  step[13] = output[14] - output[13];
+  step[14] = output[14] + output[13];
+  step[15] = output[15] + output[12];
+  step[16] = output[16];
+  step[17] = output[17];
+  temp = output[18] * -cospi_8_64 + output[29] * cospi_24_64;
+  step[18] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[19] * -cospi_8_64 + output[28] * cospi_24_64;
+  step[19] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[20] * -cospi_24_64 + output[27] * -cospi_8_64;
+  step[20] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[21] * -cospi_24_64 + output[26] * -cospi_8_64;
+  step[21] = (tran_low_t)fdct_round_shift(temp);
+  step[22] = output[22];
+  step[23] = output[23];
+  step[24] = output[24];
+  step[25] = output[25];
+  temp = output[26] * cospi_24_64 + output[21] * -cospi_8_64;
+  step[26] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[27] * cospi_24_64 + output[20] * -cospi_8_64;
+  step[27] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[28] * cospi_8_64 + output[19] * cospi_24_64;
+  step[28] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[29] * cospi_8_64 + output[18] * cospi_24_64;
+  step[29] = (tran_low_t)fdct_round_shift(temp);
+  step[30] = output[30];
+  step[31] = output[31];
+
+  range_check(step, 32, 18);
+
+  // stage 5
+  temp = step[0] * cospi_16_64 + step[1] * cospi_16_64;
+  output[0] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[1] * -cospi_16_64 + step[0] * cospi_16_64;
+  output[1] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[2] * cospi_24_64 + step[3] * cospi_8_64;
+  output[2] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[3] * cospi_24_64 + step[2] * -cospi_8_64;
+  output[3] = (tran_low_t)fdct_round_shift(temp);
+  output[4] = step[4] + step[5];
+  output[5] = step[4] - step[5];
+  output[6] = step[7] - step[6];
+  output[7] = step[7] + step[6];
+  output[8] = step[8];
+  temp = step[9] * -cospi_8_64 + step[14] * cospi_24_64;
+  output[9] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[10] * -cospi_24_64 + step[13] * -cospi_8_64;
+  output[10] = (tran_low_t)fdct_round_shift(temp);
+  output[11] = step[11];
+  output[12] = step[12];
+  temp = step[13] * cospi_24_64 + step[10] * -cospi_8_64;
+  output[13] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[14] * cospi_8_64 + step[9] * cospi_24_64;
+  output[14] = (tran_low_t)fdct_round_shift(temp);
+  output[15] = step[15];
+  output[16] = step[16] + step[19];
+  output[17] = step[17] + step[18];
+  output[18] = step[17] - step[18];
+  output[19] = step[16] - step[19];
+  output[20] = step[23] - step[20];
+  output[21] = step[22] - step[21];
+  output[22] = step[22] + step[21];
+  output[23] = step[23] + step[20];
+  output[24] = step[24] + step[27];
+  output[25] = step[25] + step[26];
+  output[26] = step[25] - step[26];
+  output[27] = step[24] - step[27];
+  output[28] = step[31] - step[28];
+  output[29] = step[30] - step[29];
+  output[30] = step[30] + step[29];
+  output[31] = step[31] + step[28];
+
+  range_check(output, 32, 18);
+
+  // stage 6
+  step[0] = output[0];
+  step[1] = output[1];
+  step[2] = output[2];
+  step[3] = output[3];
+  temp = output[4] * cospi_28_64 + output[7] * cospi_4_64;
+  step[4] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[5] * cospi_12_64 + output[6] * cospi_20_64;
+  step[5] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[6] * cospi_12_64 + output[5] * -cospi_20_64;
+  step[6] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[7] * cospi_28_64 + output[4] * -cospi_4_64;
+  step[7] = (tran_low_t)fdct_round_shift(temp);
+  step[8] = output[8] + output[9];
+  step[9] = output[8] - output[9];
+  step[10] = output[11] - output[10];
+  step[11] = output[11] + output[10];
+  step[12] = output[12] + output[13];
+  step[13] = output[12] - output[13];
+  step[14] = output[15] - output[14];
+  step[15] = output[15] + output[14];
+  step[16] = output[16];
+  temp = output[17] * -cospi_4_64 + output[30] * cospi_28_64;
+  step[17] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[18] * -cospi_28_64 + output[29] * -cospi_4_64;
+  step[18] = (tran_low_t)fdct_round_shift(temp);
+  step[19] = output[19];
+  step[20] = output[20];
+  temp = output[21] * -cospi_20_64 + output[26] * cospi_12_64;
+  step[21] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[22] * -cospi_12_64 + output[25] * -cospi_20_64;
+  step[22] = (tran_low_t)fdct_round_shift(temp);
+  step[23] = output[23];
+  step[24] = output[24];
+  temp = output[25] * cospi_12_64 + output[22] * -cospi_20_64;
+  step[25] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[26] * cospi_20_64 + output[21] * cospi_12_64;
+  step[26] = (tran_low_t)fdct_round_shift(temp);
+  step[27] = output[27];
+  step[28] = output[28];
+  temp = output[29] * cospi_28_64 + output[18] * -cospi_4_64;
+  step[29] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[30] * cospi_4_64 + output[17] * cospi_28_64;
+  step[30] = (tran_low_t)fdct_round_shift(temp);
+  step[31] = output[31];
+
+  range_check(step, 32, 18);
+
+  // stage 7
+  output[0] = step[0];
+  output[1] = step[1];
+  output[2] = step[2];
+  output[3] = step[3];
+  output[4] = step[4];
+  output[5] = step[5];
+  output[6] = step[6];
+  output[7] = step[7];
+  temp = step[8] * cospi_30_64 + step[15] * cospi_2_64;
+  output[8] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[9] * cospi_14_64 + step[14] * cospi_18_64;
+  output[9] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[10] * cospi_22_64 + step[13] * cospi_10_64;
+  output[10] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[11] * cospi_6_64 + step[12] * cospi_26_64;
+  output[11] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[12] * cospi_6_64 + step[11] * -cospi_26_64;
+  output[12] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[13] * cospi_22_64 + step[10] * -cospi_10_64;
+  output[13] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[14] * cospi_14_64 + step[9] * -cospi_18_64;
+  output[14] = (tran_low_t)fdct_round_shift(temp);
+  temp = step[15] * cospi_30_64 + step[8] * -cospi_2_64;
+  output[15] = (tran_low_t)fdct_round_shift(temp);
+  output[16] = step[16] + step[17];
+  output[17] = step[16] - step[17];
+  output[18] = step[19] - step[18];
+  output[19] = step[19] + step[18];
+  output[20] = step[20] + step[21];
+  output[21] = step[20] - step[21];
+  output[22] = step[23] - step[22];
+  output[23] = step[23] + step[22];
+  output[24] = step[24] + step[25];
+  output[25] = step[24] - step[25];
+  output[26] = step[27] - step[26];
+  output[27] = step[27] + step[26];
+  output[28] = step[28] + step[29];
+  output[29] = step[28] - step[29];
+  output[30] = step[31] - step[30];
+  output[31] = step[31] + step[30];
+
+  range_check(output, 32, 18);
+
+  // stage 8
+  step[0] = output[0];
+  step[1] = output[1];
+  step[2] = output[2];
+  step[3] = output[3];
+  step[4] = output[4];
+  step[5] = output[5];
+  step[6] = output[6];
+  step[7] = output[7];
+  step[8] = output[8];
+  step[9] = output[9];
+  step[10] = output[10];
+  step[11] = output[11];
+  step[12] = output[12];
+  step[13] = output[13];
+  step[14] = output[14];
+  step[15] = output[15];
+  temp = output[16] * cospi_31_64 + output[31] * cospi_1_64;
+  step[16] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[17] * cospi_15_64 + output[30] * cospi_17_64;
+  step[17] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[18] * cospi_23_64 + output[29] * cospi_9_64;
+  step[18] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[19] * cospi_7_64 + output[28] * cospi_25_64;
+  step[19] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[20] * cospi_27_64 + output[27] * cospi_5_64;
+  step[20] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[21] * cospi_11_64 + output[26] * cospi_21_64;
+  step[21] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[22] * cospi_19_64 + output[25] * cospi_13_64;
+  step[22] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[23] * cospi_3_64 + output[24] * cospi_29_64;
+  step[23] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[24] * cospi_3_64 + output[23] * -cospi_29_64;
+  step[24] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[25] * cospi_19_64 + output[22] * -cospi_13_64;
+  step[25] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[26] * cospi_11_64 + output[21] * -cospi_21_64;
+  step[26] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[27] * cospi_27_64 + output[20] * -cospi_5_64;
+  step[27] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[28] * cospi_7_64 + output[19] * -cospi_25_64;
+  step[28] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[29] * cospi_23_64 + output[18] * -cospi_9_64;
+  step[29] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[30] * cospi_15_64 + output[17] * -cospi_17_64;
+  step[30] = (tran_low_t)fdct_round_shift(temp);
+  temp = output[31] * cospi_31_64 + output[16] * -cospi_1_64;
+  step[31] = (tran_low_t)fdct_round_shift(temp);
+
+  range_check(step, 32, 18);
+
+  // stage 9
+  output[0] = step[0];
+  output[1] = step[16];
+  output[2] = step[8];
+  output[3] = step[24];
+  output[4] = step[4];
+  output[5] = step[20];
+  output[6] = step[12];
+  output[7] = step[28];
+  output[8] = step[2];
+  output[9] = step[18];
+  output[10] = step[10];
+  output[11] = step[26];
+  output[12] = step[6];
+  output[13] = step[22];
+  output[14] = step[14];
+  output[15] = step[30];
+  output[16] = step[1];
+  output[17] = step[17];
+  output[18] = step[9];
+  output[19] = step[25];
+  output[20] = step[5];
+  output[21] = step[21];
+  output[22] = step[13];
+  output[23] = step[29];
+  output[24] = step[3];
+  output[25] = step[19];
+  output[26] = step[11];
+  output[27] = step[27];
+  output[28] = step[7];
+  output[29] = step[23];
+  output[30] = step[15];
+  output[31] = step[31];
+
+  range_check(output, 32, 18);
+}
+
+#ifndef AV1_DCT_GTEST
+
+static void fadst4(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t x0, x1, x2, x3;
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  x0 = input[0];
+  x1 = input[1];
+  x2 = input[2];
+  x3 = input[3];
+
+  if (!(x0 | x1 | x2 | x3)) {
+    output[0] = output[1] = output[2] = output[3] = 0;
+    return;
+  }
+
+  s0 = sinpi_1_9 * x0;
+  s1 = sinpi_4_9 * x0;
+  s2 = sinpi_2_9 * x1;
+  s3 = sinpi_1_9 * x1;
+  s4 = sinpi_3_9 * x2;
+  s5 = sinpi_4_9 * x3;
+  s6 = sinpi_2_9 * x3;
+  s7 = x0 + x1 - x3;
+
+  x0 = s0 + s2 + s5;
+  x1 = sinpi_3_9 * s7;
+  x2 = s1 - s3 + s6;
+  x3 = s4;
+
+  s0 = x0 + x3;
+  s1 = x1;
+  s2 = x2 - x3;
+  s3 = x2 - x0 + x3;
+
+  // 1-D transform scaling factor is sqrt(2).
+  output[0] = (tran_low_t)fdct_round_shift(s0);
+  output[1] = (tran_low_t)fdct_round_shift(s1);
+  output[2] = (tran_low_t)fdct_round_shift(s2);
+  output[3] = (tran_low_t)fdct_round_shift(s3);
+}
+
+static void fadst8(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
+
+  tran_high_t x0 = input[7];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[5];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[3];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[1];
+  tran_high_t x7 = input[6];
+
+  // stage 1
+  s0 = cospi_2_64 * x0 + cospi_30_64 * x1;
+  s1 = cospi_30_64 * x0 - cospi_2_64 * x1;
+  s2 = cospi_10_64 * x2 + cospi_22_64 * x3;
+  s3 = cospi_22_64 * x2 - cospi_10_64 * x3;
+  s4 = cospi_18_64 * x4 + cospi_14_64 * x5;
+  s5 = cospi_14_64 * x4 - cospi_18_64 * x5;
+  s6 = cospi_26_64 * x6 + cospi_6_64 * x7;
+  s7 = cospi_6_64 * x6 - cospi_26_64 * x7;
+
+  x0 = s0 + s4;
+  x1 = s1 + s5;
+  x2 = s2 + s6;
+  x3 = s3 + s7;
+  x4 = fdct_round_shift(s0 - s4);
+  x5 = fdct_round_shift(s1 - s5);
+  x6 = fdct_round_shift(s2 - s6);
+  x7 = fdct_round_shift(s3 - s7);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = cospi_8_64 * x4 + cospi_24_64 * x5;
+  s5 = cospi_24_64 * x4 - cospi_8_64 * x5;
+  s6 = -cospi_24_64 * x6 + cospi_8_64 * x7;
+  s7 = cospi_8_64 * x6 + cospi_24_64 * x7;
+
+  x0 = fdct_round_shift(s0 + s2);
+  x1 = fdct_round_shift(s1 + s3);
+  x2 = fdct_round_shift(s0 - s2);
+  x3 = fdct_round_shift(s1 - s3);
+  x4 = fdct_round_shift(s4 + s6);
+  x5 = fdct_round_shift(s5 + s7);
+  x6 = fdct_round_shift(s4 - s6);
+  x7 = fdct_round_shift(s5 - s7);
+
+  // stage 3
+  s2 = cospi_16_64 * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (x6 - x7);
+
+  x2 = fdct_round_shift(s2);
+  x3 = fdct_round_shift(s3);
+  x6 = fdct_round_shift(s6);
+  x7 = fdct_round_shift(s7);
+
+  output[0] = (tran_low_t)x0;
+  output[1] = (tran_low_t)-x4;
+  output[2] = (tran_low_t)x6;
+  output[3] = (tran_low_t)-x2;
+  output[4] = (tran_low_t)x3;
+  output[5] = (tran_low_t)-x7;
+  output[6] = (tran_low_t)x5;
+  output[7] = (tran_low_t)-x1;
+}
+
+static void fadst16(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1, s2, s3, s4, s5, s6, s7, s8;
+  tran_high_t s9, s10, s11, s12, s13, s14, s15;
+
+  tran_high_t x0 = input[15];
+  tran_high_t x1 = input[0];
+  tran_high_t x2 = input[13];
+  tran_high_t x3 = input[2];
+  tran_high_t x4 = input[11];
+  tran_high_t x5 = input[4];
+  tran_high_t x6 = input[9];
+  tran_high_t x7 = input[6];
+  tran_high_t x8 = input[7];
+  tran_high_t x9 = input[8];
+  tran_high_t x10 = input[5];
+  tran_high_t x11 = input[10];
+  tran_high_t x12 = input[3];
+  tran_high_t x13 = input[12];
+  tran_high_t x14 = input[1];
+  tran_high_t x15 = input[14];
+
+  // stage 1
+  s0 = x0 * cospi_1_64 + x1 * cospi_31_64;
+  s1 = x0 * cospi_31_64 - x1 * cospi_1_64;
+  s2 = x2 * cospi_5_64 + x3 * cospi_27_64;
+  s3 = x2 * cospi_27_64 - x3 * cospi_5_64;
+  s4 = x4 * cospi_9_64 + x5 * cospi_23_64;
+  s5 = x4 * cospi_23_64 - x5 * cospi_9_64;
+  s6 = x6 * cospi_13_64 + x7 * cospi_19_64;
+  s7 = x6 * cospi_19_64 - x7 * cospi_13_64;
+  s8 = x8 * cospi_17_64 + x9 * cospi_15_64;
+  s9 = x8 * cospi_15_64 - x9 * cospi_17_64;
+  s10 = x10 * cospi_21_64 + x11 * cospi_11_64;
+  s11 = x10 * cospi_11_64 - x11 * cospi_21_64;
+  s12 = x12 * cospi_25_64 + x13 * cospi_7_64;
+  s13 = x12 * cospi_7_64 - x13 * cospi_25_64;
+  s14 = x14 * cospi_29_64 + x15 * cospi_3_64;
+  s15 = x14 * cospi_3_64 - x15 * cospi_29_64;
+
+  x0 = s0 + s8;
+  x1 = s1 + s9;
+  x2 = s2 + s10;
+  x3 = s3 + s11;
+  x4 = s4 + s12;
+  x5 = s5 + s13;
+  x6 = s6 + s14;
+  x7 = s7 + s15;
+
+  x8 = fdct_round_shift(s0 - s8);
+  x9 = fdct_round_shift(s1 - s9);
+  x10 = fdct_round_shift(s2 - s10);
+  x11 = fdct_round_shift(s3 - s11);
+  x12 = fdct_round_shift(s4 - s12);
+  x13 = fdct_round_shift(s5 - s13);
+  x14 = fdct_round_shift(s6 - s14);
+  x15 = fdct_round_shift(s7 - s15);
+
+  // stage 2
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4;
+  s5 = x5;
+  s6 = x6;
+  s7 = x7;
+  s8 = x8 * cospi_4_64 + x9 * cospi_28_64;
+  s9 = x8 * cospi_28_64 - x9 * cospi_4_64;
+  s10 = x10 * cospi_20_64 + x11 * cospi_12_64;
+  s11 = x10 * cospi_12_64 - x11 * cospi_20_64;
+  s12 = -x12 * cospi_28_64 + x13 * cospi_4_64;
+  s13 = x12 * cospi_4_64 + x13 * cospi_28_64;
+  s14 = -x14 * cospi_12_64 + x15 * cospi_20_64;
+  s15 = x14 * cospi_20_64 + x15 * cospi_12_64;
+
+  x0 = s0 + s4;
+  x1 = s1 + s5;
+  x2 = s2 + s6;
+  x3 = s3 + s7;
+  x4 = fdct_round_shift(s0 - s4);
+  x5 = fdct_round_shift(s1 - s5);
+  x6 = fdct_round_shift(s2 - s6);
+  x7 = fdct_round_shift(s3 - s7);
+
+  x8 = s8 + s12;
+  x9 = s9 + s13;
+  x10 = s10 + s14;
+  x11 = s11 + s15;
+  x12 = fdct_round_shift(s8 - s12);
+  x13 = fdct_round_shift(s9 - s13);
+  x14 = fdct_round_shift(s10 - s14);
+  x15 = fdct_round_shift(s11 - s15);
+
+  // stage 3
+  s0 = x0;
+  s1 = x1;
+  s2 = x2;
+  s3 = x3;
+  s4 = x4 * cospi_8_64 + x5 * cospi_24_64;
+  s5 = x4 * cospi_24_64 - x5 * cospi_8_64;
+  s6 = -x6 * cospi_24_64 + x7 * cospi_8_64;
+  s7 = x6 * cospi_8_64 + x7 * cospi_24_64;
+  s8 = x8;
+  s9 = x9;
+  s10 = x10;
+  s11 = x11;
+  s12 = x12 * cospi_8_64 + x13 * cospi_24_64;
+  s13 = x12 * cospi_24_64 - x13 * cospi_8_64;
+  s14 = -x14 * cospi_24_64 + x15 * cospi_8_64;
+  s15 = x14 * cospi_8_64 + x15 * cospi_24_64;
+
+  x0 = fdct_round_shift(s0 + s2);
+  x1 = fdct_round_shift(s1 + s3);
+  x2 = fdct_round_shift(s0 - s2);
+  x3 = fdct_round_shift(s1 - s3);
+
+  x4 = fdct_round_shift(s4 + s6);
+  x5 = fdct_round_shift(s5 + s7);
+  x6 = fdct_round_shift(s4 - s6);
+  x7 = fdct_round_shift(s5 - s7);
+
+  x8 = fdct_round_shift(s8 + s10);
+  x9 = fdct_round_shift(s9 + s11);
+  x10 = fdct_round_shift(s8 - s10);
+  x11 = fdct_round_shift(s9 - s11);
+
+  x12 = fdct_round_shift(s12 + s14);
+  x13 = fdct_round_shift(s13 + s15);
+  x14 = fdct_round_shift(s12 - s14);
+  x15 = fdct_round_shift(s13 - s15);
+
+  // stage 4
+  s2 = (-cospi_16_64) * (x2 + x3);
+  s3 = cospi_16_64 * (x2 - x3);
+  s6 = cospi_16_64 * (x6 + x7);
+  s7 = cospi_16_64 * (-x6 + x7);
+  s10 = cospi_16_64 * (x10 + x11);
+  s11 = cospi_16_64 * (-x10 + x11);
+  s14 = (-cospi_16_64) * (x14 + x15);
+  s15 = cospi_16_64 * (x14 - x15);
+
+  x2 = fdct_round_shift(s2);
+  x3 = fdct_round_shift(s3);
+  x6 = fdct_round_shift(s6);
+  x7 = fdct_round_shift(s7);
+  x10 = fdct_round_shift(s10);
+  x11 = fdct_round_shift(s11);
+  x14 = fdct_round_shift(s14);
+  x15 = fdct_round_shift(s15);
+
+  output[0] = (tran_low_t)x0;
+  output[1] = (tran_low_t)-x8;
+  output[2] = (tran_low_t)x12;
+  output[3] = (tran_low_t)-x4;
+  output[4] = (tran_low_t)x6;
+  output[5] = (tran_low_t)x14;
+  output[6] = (tran_low_t)x10;
+  output[7] = (tran_low_t)x2;
+  output[8] = (tran_low_t)x3;
+  output[9] = (tran_low_t)x11;
+  output[10] = (tran_low_t)x15;
+  output[11] = (tran_low_t)x7;
+  output[12] = (tran_low_t)x5;
+  output[13] = (tran_low_t)-x13;
+  output[14] = (tran_low_t)x9;
+  output[15] = (tran_low_t)-x1;
+}
+
+// For use in lieu of ADST
+static void fhalfright32(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  tran_low_t inputhalf[16];
+  for (i = 0; i < 16; ++i) {
+    output[16 + i] = input[i] * 4;
+  }
+  // Multiply input by sqrt(2)
+  for (i = 0; i < 16; ++i) {
+    inputhalf[i] = (tran_low_t)fdct_round_shift(input[i + 16] * Sqrt2);
+  }
+  fdct16(inputhalf, output);
+  // Note overall scaling factor is 4 times orthogonal
+}
+
+#if CONFIG_EXT_TX
+static void fidtx4(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 4; ++i)
+    output[i] = (tran_low_t)fdct_round_shift(input[i] * Sqrt2);
+}
+
+static void fidtx8(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 8; ++i) output[i] = input[i] * 2;
+}
+
+static void fidtx16(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 16; ++i)
+    output[i] = (tran_low_t)fdct_round_shift(input[i] * 2 * Sqrt2);
+}
+
+static void fidtx32(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 32; ++i) output[i] = input[i] * 4;
+}
+
+static void copy_block(const int16_t *src, int src_stride, int l, int w,
+                       int16_t *dest, int dest_stride) {
+  int i;
+  for (i = 0; i < l; ++i) {
+    memcpy(dest + dest_stride * i, src + src_stride * i, w * sizeof(int16_t));
+  }
+}
+
+static void fliplr(int16_t *dest, int stride, int l, int w) {
+  int i, j;
+  for (i = 0; i < l; ++i) {
+    for (j = 0; j < w / 2; ++j) {
+      const int16_t tmp = dest[i * stride + j];
+      dest[i * stride + j] = dest[i * stride + w - 1 - j];
+      dest[i * stride + w - 1 - j] = tmp;
+    }
+  }
+}
+
+static void flipud(int16_t *dest, int stride, int l, int w) {
+  int i, j;
+  for (j = 0; j < w; ++j) {
+    for (i = 0; i < l / 2; ++i) {
+      const int16_t tmp = dest[i * stride + j];
+      dest[i * stride + j] = dest[(l - 1 - i) * stride + j];
+      dest[(l - 1 - i) * stride + j] = tmp;
+    }
+  }
+}
+
+static void fliplrud(int16_t *dest, int stride, int l, int w) {
+  int i, j;
+  for (i = 0; i < l / 2; ++i) {
+    for (j = 0; j < w; ++j) {
+      const int16_t tmp = dest[i * stride + j];
+      dest[i * stride + j] = dest[(l - 1 - i) * stride + w - 1 - j];
+      dest[(l - 1 - i) * stride + w - 1 - j] = tmp;
+    }
+  }
+}
+
+static void copy_fliplr(const int16_t *src, int src_stride, int l, int w,
+                        int16_t *dest, int dest_stride) {
+  copy_block(src, src_stride, l, w, dest, dest_stride);
+  fliplr(dest, dest_stride, l, w);
+}
+
+static void copy_flipud(const int16_t *src, int src_stride, int l, int w,
+                        int16_t *dest, int dest_stride) {
+  copy_block(src, src_stride, l, w, dest, dest_stride);
+  flipud(dest, dest_stride, l, w);
+}
+
+static void copy_fliplrud(const int16_t *src, int src_stride, int l, int w,
+                          int16_t *dest, int dest_stride) {
+  copy_block(src, src_stride, l, w, dest, dest_stride);
+  fliplrud(dest, dest_stride, l, w);
+}
+
+static void maybe_flip_input(const int16_t **src, int *src_stride, int l, int w,
+                             int16_t *buff, int tx_type) {
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case IDTX:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST: break;
+    case FLIPADST_DCT:
+    case FLIPADST_ADST:
+    case V_FLIPADST:
+      copy_flipud(*src, *src_stride, l, w, buff, w);
+      *src = buff;
+      *src_stride = w;
+      break;
+    case DCT_FLIPADST:
+    case ADST_FLIPADST:
+    case H_FLIPADST:
+      copy_fliplr(*src, *src_stride, l, w, buff, w);
+      *src = buff;
+      *src_stride = w;
+      break;
+    case FLIPADST_FLIPADST:
+      copy_fliplrud(*src, *src_stride, l, w, buff, w);
+      *src = buff;
+      *src_stride = w;
+      break;
+    default: assert(0); break;
+  }
+}
+#endif  // CONFIG_EXT_TX
+
+void av1_fht4x4_c(const int16_t *input, tran_low_t *output, int stride,
+                  int tx_type) {
+  if (tx_type == DCT_DCT) {
+    aom_fdct4x4_c(input, output, stride);
+  } else {
+    static const transform_2d FHT[] = {
+      { fdct4, fdct4 },    // DCT_DCT
+      { fadst4, fdct4 },   // ADST_DCT
+      { fdct4, fadst4 },   // DCT_ADST
+      { fadst4, fadst4 },  // ADST_ADST
+#if CONFIG_EXT_TX
+      { fadst4, fdct4 },   // FLIPADST_DCT
+      { fdct4, fadst4 },   // DCT_FLIPADST
+      { fadst4, fadst4 },  // FLIPADST_FLIPADST
+      { fadst4, fadst4 },  // ADST_FLIPADST
+      { fadst4, fadst4 },  // FLIPADST_ADST
+      { fidtx4, fidtx4 },  // IDTX
+      { fdct4, fidtx4 },   // V_DCT
+      { fidtx4, fdct4 },   // H_DCT
+      { fadst4, fidtx4 },  // V_ADST
+      { fidtx4, fadst4 },  // H_ADST
+      { fadst4, fidtx4 },  // V_FLIPADST
+      { fidtx4, fadst4 },  // H_FLIPADST
+#endif                     // CONFIG_EXT_TX
+    };
+    const transform_2d ht = FHT[tx_type];
+    tran_low_t out[4 * 4];
+    int i, j;
+    tran_low_t temp_in[4], temp_out[4];
+
+#if CONFIG_EXT_TX
+    int16_t flipped_input[4 * 4];
+    maybe_flip_input(&input, &stride, 4, 4, flipped_input, tx_type);
+#endif
+
+    // Columns
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j) temp_in[j] = input[j * stride + i] * 16;
+      if (i == 0 && temp_in[0]) temp_in[0] += 1;
+      ht.cols(temp_in, temp_out);
+      for (j = 0; j < 4; ++j) out[j * 4 + i] = temp_out[j];
+    }
+
+    // Rows
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j) temp_in[j] = out[j + i * 4];
+      ht.rows(temp_in, temp_out);
+      for (j = 0; j < 4; ++j) output[j + i * 4] = (temp_out[j] + 1) >> 2;
+    }
+  }
+}
+
+void av1_fht4x8_c(const int16_t *input, tran_low_t *output, int stride,
+                  int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct8, fdct4 },    // DCT_DCT
+    { fadst8, fdct4 },   // ADST_DCT
+    { fdct8, fadst4 },   // DCT_ADST
+    { fadst8, fadst4 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fadst8, fdct4 },   // FLIPADST_DCT
+    { fdct8, fadst4 },   // DCT_FLIPADST
+    { fadst8, fadst4 },  // FLIPADST_FLIPADST
+    { fadst8, fadst4 },  // ADST_FLIPADST
+    { fadst8, fadst4 },  // FLIPADST_ADST
+    { fidtx8, fidtx4 },  // IDTX
+    { fdct8, fidtx4 },   // V_DCT
+    { fidtx8, fdct4 },   // H_DCT
+    { fadst8, fidtx4 },  // V_ADST
+    { fidtx8, fadst4 },  // H_ADST
+    { fadst8, fidtx4 },  // V_FLIPADST
+    { fidtx8, fadst4 },  // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  const int n = 4;
+  const int n2 = 8;
+  tran_low_t out[8 * 4];
+  tran_low_t temp_in[8], temp_out[8];
+  int i, j;
+#if CONFIG_EXT_TX
+  int16_t flipped_input[8 * 4];
+  maybe_flip_input(&input, &stride, n2, n, flipped_input, tx_type);
+#endif
+
+  // Rows
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j)
+      temp_in[j] =
+          (tran_low_t)fdct_round_shift(input[i * stride + j] * 4 * Sqrt2);
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < n; ++j) out[j * n2 + i] = temp_out[j];
+  }
+
+  // Columns
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) temp_in[j] = out[j + i * n2];
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < n2; ++j)
+      output[i + j * n] = (temp_out[j] + (temp_out[j] < 0)) >> 1;
+  }
+  // Note: overall scale factor of transform is 8 times unitary
+}
+
+void av1_fht8x4_c(const int16_t *input, tran_low_t *output, int stride,
+                  int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct4, fdct8 },    // DCT_DCT
+    { fadst4, fdct8 },   // ADST_DCT
+    { fdct4, fadst8 },   // DCT_ADST
+    { fadst4, fadst8 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fadst4, fdct8 },   // FLIPADST_DCT
+    { fdct4, fadst8 },   // DCT_FLIPADST
+    { fadst4, fadst8 },  // FLIPADST_FLIPADST
+    { fadst4, fadst8 },  // ADST_FLIPADST
+    { fadst4, fadst8 },  // FLIPADST_ADST
+    { fidtx4, fidtx8 },  // IDTX
+    { fdct4, fidtx8 },   // V_DCT
+    { fidtx4, fdct8 },   // H_DCT
+    { fadst4, fidtx8 },  // V_ADST
+    { fidtx4, fadst8 },  // H_ADST
+    { fadst4, fidtx8 },  // V_FLIPADST
+    { fidtx4, fadst8 },  // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  const int n = 4;
+  const int n2 = 8;
+  tran_low_t out[8 * 4];
+  tran_low_t temp_in[8], temp_out[8];
+  int i, j;
+#if CONFIG_EXT_TX
+  int16_t flipped_input[8 * 4];
+  maybe_flip_input(&input, &stride, n, n2, flipped_input, tx_type);
+#endif
+
+  // Columns
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j)
+      temp_in[j] =
+          (tran_low_t)fdct_round_shift(input[j * stride + i] * 4 * Sqrt2);
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < n; ++j) out[j * n2 + i] = temp_out[j];
+  }
+
+  // Rows
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) temp_in[j] = out[j + i * n2];
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < n2; ++j)
+      output[j + i * n2] = (temp_out[j] + (temp_out[j] < 0)) >> 1;
+  }
+  // Note: overall scale factor of transform is 8 times unitary
+}
+
+void av1_fht4x16_c(const int16_t *input, tran_low_t *output, int stride,
+                   int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct16, fdct4 },    // DCT_DCT
+    { fadst16, fdct4 },   // ADST_DCT
+    { fdct16, fadst4 },   // DCT_ADST
+    { fadst16, fadst4 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fadst16, fdct4 },   // FLIPADST_DCT
+    { fdct16, fadst4 },   // DCT_FLIPADST
+    { fadst16, fadst4 },  // FLIPADST_FLIPADST
+    { fadst16, fadst4 },  // ADST_FLIPADST
+    { fadst16, fadst4 },  // FLIPADST_ADST
+    { fidtx16, fidtx4 },  // IDTX
+    { fdct16, fidtx4 },   // V_DCT
+    { fidtx16, fdct4 },   // H_DCT
+    { fadst16, fidtx4 },  // V_ADST
+    { fidtx16, fadst4 },  // H_ADST
+    { fadst16, fidtx4 },  // V_FLIPADST
+    { fidtx16, fadst4 },  // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  const int n = 4;
+  const int n4 = 16;
+  tran_low_t out[16 * 4];
+  tran_low_t temp_in[16], temp_out[16];
+  int i, j;
+#if CONFIG_EXT_TX
+  int16_t flipped_input[16 * 4];
+  maybe_flip_input(&input, &stride, n4, n, flipped_input, tx_type);
+#endif
+
+  // Rows
+  for (i = 0; i < n4; ++i) {
+    for (j = 0; j < n; ++j) temp_in[j] = input[i * stride + j] * 4;
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < n; ++j) out[j * n4 + i] = temp_out[j];
+  }
+
+  // Columns
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n4; ++j) temp_in[j] = out[j + i * n4];
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < n4; ++j)
+      output[i + j * n] = (temp_out[j] + (temp_out[j] < 0)) >> 1;
+  }
+  // Note: overall scale factor of transform is 8 times unitary
+}
+
+void av1_fht16x4_c(const int16_t *input, tran_low_t *output, int stride,
+                   int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct4, fdct16 },    // DCT_DCT
+    { fadst4, fdct16 },   // ADST_DCT
+    { fdct4, fadst16 },   // DCT_ADST
+    { fadst4, fadst16 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fadst4, fdct16 },   // FLIPADST_DCT
+    { fdct4, fadst16 },   // DCT_FLIPADST
+    { fadst4, fadst16 },  // FLIPADST_FLIPADST
+    { fadst4, fadst16 },  // ADST_FLIPADST
+    { fadst4, fadst16 },  // FLIPADST_ADST
+    { fidtx4, fidtx16 },  // IDTX
+    { fdct4, fidtx16 },   // V_DCT
+    { fidtx4, fdct16 },   // H_DCT
+    { fadst4, fidtx16 },  // V_ADST
+    { fidtx4, fadst16 },  // H_ADST
+    { fadst4, fidtx16 },  // V_FLIPADST
+    { fidtx4, fadst16 },  // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  const int n = 4;
+  const int n4 = 16;
+  tran_low_t out[16 * 4];
+  tran_low_t temp_in[16], temp_out[16];
+  int i, j;
+#if CONFIG_EXT_TX
+  int16_t flipped_input[16 * 4];
+  maybe_flip_input(&input, &stride, n, n4, flipped_input, tx_type);
+#endif
+
+  // Columns
+  for (i = 0; i < n4; ++i) {
+    for (j = 0; j < n; ++j) temp_in[j] = input[j * stride + i] * 4;
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < n; ++j) out[j * n4 + i] = temp_out[j];
+  }
+
+  // Rows
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n4; ++j) temp_in[j] = out[j + i * n4];
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < n4; ++j)
+      output[j + i * n4] = (temp_out[j] + (temp_out[j] < 0)) >> 1;
+  }
+  // Note: overall scale factor of transform is 8 times unitary
+}
+
+void av1_fht8x16_c(const int16_t *input, tran_low_t *output, int stride,
+                   int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct16, fdct8 },    // DCT_DCT
+    { fadst16, fdct8 },   // ADST_DCT
+    { fdct16, fadst8 },   // DCT_ADST
+    { fadst16, fadst8 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fadst16, fdct8 },   // FLIPADST_DCT
+    { fdct16, fadst8 },   // DCT_FLIPADST
+    { fadst16, fadst8 },  // FLIPADST_FLIPADST
+    { fadst16, fadst8 },  // ADST_FLIPADST
+    { fadst16, fadst8 },  // FLIPADST_ADST
+    { fidtx16, fidtx8 },  // IDTX
+    { fdct16, fidtx8 },   // V_DCT
+    { fidtx16, fdct8 },   // H_DCT
+    { fadst16, fidtx8 },  // V_ADST
+    { fidtx16, fadst8 },  // H_ADST
+    { fadst16, fidtx8 },  // V_FLIPADST
+    { fidtx16, fadst8 },  // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  const int n = 8;
+  const int n2 = 16;
+  tran_low_t out[16 * 8];
+  tran_low_t temp_in[16], temp_out[16];
+  int i, j;
+#if CONFIG_EXT_TX
+  int16_t flipped_input[16 * 8];
+  maybe_flip_input(&input, &stride, n2, n, flipped_input, tx_type);
+#endif
+
+  // Rows
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j)
+      temp_in[j] =
+          (tran_low_t)fdct_round_shift(input[i * stride + j] * 4 * Sqrt2);
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < n; ++j)
+      out[j * n2 + i] = ROUND_POWER_OF_TWO_SIGNED(temp_out[j], 2);
+  }
+
+  // Columns
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) temp_in[j] = out[j + i * n2];
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < n2; ++j) output[i + j * n] = temp_out[j];
+  }
+  // Note: overall scale factor of transform is 8 times unitary
+}
+
+void av1_fht16x8_c(const int16_t *input, tran_low_t *output, int stride,
+                   int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct8, fdct16 },    // DCT_DCT
+    { fadst8, fdct16 },   // ADST_DCT
+    { fdct8, fadst16 },   // DCT_ADST
+    { fadst8, fadst16 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fadst8, fdct16 },   // FLIPADST_DCT
+    { fdct8, fadst16 },   // DCT_FLIPADST
+    { fadst8, fadst16 },  // FLIPADST_FLIPADST
+    { fadst8, fadst16 },  // ADST_FLIPADST
+    { fadst8, fadst16 },  // FLIPADST_ADST
+    { fidtx8, fidtx16 },  // IDTX
+    { fdct8, fidtx16 },   // V_DCT
+    { fidtx8, fdct16 },   // H_DCT
+    { fadst8, fidtx16 },  // V_ADST
+    { fidtx8, fadst16 },  // H_ADST
+    { fadst8, fidtx16 },  // V_FLIPADST
+    { fidtx8, fadst16 },  // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  const int n = 8;
+  const int n2 = 16;
+  tran_low_t out[16 * 8];
+  tran_low_t temp_in[16], temp_out[16];
+  int i, j;
+#if CONFIG_EXT_TX
+  int16_t flipped_input[16 * 8];
+  maybe_flip_input(&input, &stride, n, n2, flipped_input, tx_type);
+#endif
+
+  // Columns
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j)
+      temp_in[j] =
+          (tran_low_t)fdct_round_shift(input[j * stride + i] * 4 * Sqrt2);
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < n; ++j)
+      out[j * n2 + i] = ROUND_POWER_OF_TWO_SIGNED(temp_out[j], 2);
+  }
+
+  // Rows
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) temp_in[j] = out[j + i * n2];
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < n2; ++j) output[j + i * n2] = temp_out[j];
+  }
+  // Note: overall scale factor of transform is 8 times unitary
+}
+
+void av1_fht8x32_c(const int16_t *input, tran_low_t *output, int stride,
+                   int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct32, fdct8 },         // DCT_DCT
+    { fhalfright32, fdct8 },   // ADST_DCT
+    { fdct32, fadst8 },        // DCT_ADST
+    { fhalfright32, fadst8 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fhalfright32, fdct8 },   // FLIPADST_DCT
+    { fdct32, fadst8 },        // DCT_FLIPADST
+    { fhalfright32, fadst8 },  // FLIPADST_FLIPADST
+    { fhalfright32, fadst8 },  // ADST_FLIPADST
+    { fhalfright32, fadst8 },  // FLIPADST_ADST
+    { fidtx32, fidtx8 },       // IDTX
+    { fdct32, fidtx8 },        // V_DCT
+    { fidtx32, fdct8 },        // H_DCT
+    { fhalfright32, fidtx8 },  // V_ADST
+    { fidtx32, fadst8 },       // H_ADST
+    { fhalfright32, fidtx8 },  // V_FLIPADST
+    { fidtx32, fadst8 },       // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  const int n = 8;
+  const int n4 = 32;
+  tran_low_t out[32 * 8];
+  tran_low_t temp_in[32], temp_out[32];
+  int i, j;
+#if CONFIG_EXT_TX
+  int16_t flipped_input[32 * 8];
+  maybe_flip_input(&input, &stride, n4, n, flipped_input, tx_type);
+#endif
+
+  // Rows
+  for (i = 0; i < n4; ++i) {
+    for (j = 0; j < n; ++j) temp_in[j] = input[i * stride + j] * 4;
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < n; ++j) out[j * n4 + i] = temp_out[j];
+  }
+
+  // Columns
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n4; ++j) temp_in[j] = out[j + i * n4];
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < n4; ++j)
+      output[i + j * n] = ROUND_POWER_OF_TWO_SIGNED(temp_out[j], 2);
+  }
+  // Note: overall scale factor of transform is 4 times unitary
+}
+
+void av1_fht32x8_c(const int16_t *input, tran_low_t *output, int stride,
+                   int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct8, fdct32 },         // DCT_DCT
+    { fadst8, fdct32 },        // ADST_DCT
+    { fdct8, fhalfright32 },   // DCT_ADST
+    { fadst8, fhalfright32 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fadst8, fdct32 },        // FLIPADST_DCT
+    { fdct8, fhalfright32 },   // DCT_FLIPADST
+    { fadst8, fhalfright32 },  // FLIPADST_FLIPADST
+    { fadst8, fhalfright32 },  // ADST_FLIPADST
+    { fadst8, fhalfright32 },  // FLIPADST_ADST
+    { fidtx8, fidtx32 },       // IDTX
+    { fdct8, fidtx32 },        // V_DCT
+    { fidtx8, fdct32 },        // H_DCT
+    { fadst8, fidtx32 },       // V_ADST
+    { fidtx8, fhalfright32 },  // H_ADST
+    { fadst8, fidtx32 },       // V_FLIPADST
+    { fidtx8, fhalfright32 },  // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  const int n = 8;
+  const int n4 = 32;
+  tran_low_t out[32 * 8];
+  tran_low_t temp_in[32], temp_out[32];
+  int i, j;
+#if CONFIG_EXT_TX
+  int16_t flipped_input[32 * 8];
+  maybe_flip_input(&input, &stride, n, n4, flipped_input, tx_type);
+#endif
+
+  // Columns
+  for (i = 0; i < n4; ++i) {
+    for (j = 0; j < n; ++j) temp_in[j] = input[j * stride + i] * 4;
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < n; ++j) out[j * n4 + i] = temp_out[j];
+  }
+
+  // Rows
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n4; ++j) temp_in[j] = out[j + i * n4];
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < n4; ++j)
+      output[j + i * n4] = ROUND_POWER_OF_TWO_SIGNED(temp_out[j], 2);
+  }
+  // Note: overall scale factor of transform is 4 times unitary
+}
+
+void av1_fht16x32_c(const int16_t *input, tran_low_t *output, int stride,
+                    int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct32, fdct16 },         // DCT_DCT
+    { fhalfright32, fdct16 },   // ADST_DCT
+    { fdct32, fadst16 },        // DCT_ADST
+    { fhalfright32, fadst16 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fhalfright32, fdct16 },   // FLIPADST_DCT
+    { fdct32, fadst16 },        // DCT_FLIPADST
+    { fhalfright32, fadst16 },  // FLIPADST_FLIPADST
+    { fhalfright32, fadst16 },  // ADST_FLIPADST
+    { fhalfright32, fadst16 },  // FLIPADST_ADST
+    { fidtx32, fidtx16 },       // IDTX
+    { fdct32, fidtx16 },        // V_DCT
+    { fidtx32, fdct16 },        // H_DCT
+    { fhalfright32, fidtx16 },  // V_ADST
+    { fidtx32, fadst16 },       // H_ADST
+    { fhalfright32, fidtx16 },  // V_FLIPADST
+    { fidtx32, fadst16 },       // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  const int n = 16;
+  const int n2 = 32;
+  tran_low_t out[32 * 16];
+  tran_low_t temp_in[32], temp_out[32];
+  int i, j;
+#if CONFIG_EXT_TX
+  int16_t flipped_input[32 * 16];
+  maybe_flip_input(&input, &stride, n2, n, flipped_input, tx_type);
+#endif
+
+  // Rows
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j)
+      temp_in[j] =
+          (tran_low_t)fdct_round_shift(input[i * stride + j] * 4 * Sqrt2);
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < n; ++j)
+      out[j * n2 + i] = ROUND_POWER_OF_TWO_SIGNED(temp_out[j], 4);
+  }
+
+  // Columns
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) temp_in[j] = out[j + i * n2];
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < n2; ++j) output[i + j * n] = temp_out[j];
+  }
+  // Note: overall scale factor of transform is 4 times unitary
+}
+
+void av1_fht32x16_c(const int16_t *input, tran_low_t *output, int stride,
+                    int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct16, fdct32 },         // DCT_DCT
+    { fadst16, fdct32 },        // ADST_DCT
+    { fdct16, fhalfright32 },   // DCT_ADST
+    { fadst16, fhalfright32 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fadst16, fdct32 },        // FLIPADST_DCT
+    { fdct16, fhalfright32 },   // DCT_FLIPADST
+    { fadst16, fhalfright32 },  // FLIPADST_FLIPADST
+    { fadst16, fhalfright32 },  // ADST_FLIPADST
+    { fadst16, fhalfright32 },  // FLIPADST_ADST
+    { fidtx16, fidtx32 },       // IDTX
+    { fdct16, fidtx32 },        // V_DCT
+    { fidtx16, fdct32 },        // H_DCT
+    { fadst16, fidtx32 },       // V_ADST
+    { fidtx16, fhalfright32 },  // H_ADST
+    { fadst16, fidtx32 },       // V_FLIPADST
+    { fidtx16, fhalfright32 },  // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  const int n = 16;
+  const int n2 = 32;
+  tran_low_t out[32 * 16];
+  tran_low_t temp_in[32], temp_out[32];
+  int i, j;
+#if CONFIG_EXT_TX
+  int16_t flipped_input[32 * 16];
+  maybe_flip_input(&input, &stride, n, n2, flipped_input, tx_type);
+#endif
+
+  // Columns
+  for (i = 0; i < n2; ++i) {
+    for (j = 0; j < n; ++j)
+      temp_in[j] =
+          (tran_low_t)fdct_round_shift(input[j * stride + i] * 4 * Sqrt2);
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < n; ++j)
+      out[j * n2 + i] = ROUND_POWER_OF_TWO_SIGNED(temp_out[j], 4);
+  }
+
+  // Rows
+  for (i = 0; i < n; ++i) {
+    for (j = 0; j < n2; ++j) temp_in[j] = out[j + i * n2];
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < n2; ++j) output[j + i * n2] = temp_out[j];
+  }
+  // Note: overall scale factor of transform is 4 times unitary
+}
+
+void av1_fdct8x8_quant_c(const int16_t *input, int stride,
+                         tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                         int skip_block, const int16_t *zbin_ptr,
+                         const int16_t *round_ptr, const int16_t *quant_ptr,
+                         const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+                         tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                         uint16_t *eob_ptr, const int16_t *scan,
+                         const int16_t *iscan
+#if CONFIG_AOM_QM
+                         ,
+                         const qm_val_t *qm_ptr, const qm_val_t *iqm_ptr
+#endif
+                         ) {
+  int eob = -1;
+
+  int i, j;
+  tran_low_t intermediate[64];
+
+  // Transform columns
+  {
+    tran_low_t *output = intermediate;
+    tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;  // canbe16
+    tran_high_t t0, t1, t2, t3;                  // needs32
+    tran_high_t x0, x1, x2, x3;                  // canbe16
+
+    for (i = 0; i < 8; i++) {
+      // stage 1
+      s0 = (input[0 * stride] + input[7 * stride]) * 4;
+      s1 = (input[1 * stride] + input[6 * stride]) * 4;
+      s2 = (input[2 * stride] + input[5 * stride]) * 4;
+      s3 = (input[3 * stride] + input[4 * stride]) * 4;
+      s4 = (input[3 * stride] - input[4 * stride]) * 4;
+      s5 = (input[2 * stride] - input[5 * stride]) * 4;
+      s6 = (input[1 * stride] - input[6 * stride]) * 4;
+      s7 = (input[0 * stride] - input[7 * stride]) * 4;
+
+      // fdct4(step, step);
+      x0 = s0 + s3;
+      x1 = s1 + s2;
+      x2 = s1 - s2;
+      x3 = s0 - s3;
+      t0 = (x0 + x1) * cospi_16_64;
+      t1 = (x0 - x1) * cospi_16_64;
+      t2 = x2 * cospi_24_64 + x3 * cospi_8_64;
+      t3 = -x2 * cospi_8_64 + x3 * cospi_24_64;
+      output[0 * 8] = (tran_low_t)fdct_round_shift(t0);
+      output[2 * 8] = (tran_low_t)fdct_round_shift(t2);
+      output[4 * 8] = (tran_low_t)fdct_round_shift(t1);
+      output[6 * 8] = (tran_low_t)fdct_round_shift(t3);
+
+      // stage 2
+      t0 = (s6 - s5) * cospi_16_64;
+      t1 = (s6 + s5) * cospi_16_64;
+      t2 = fdct_round_shift(t0);
+      t3 = fdct_round_shift(t1);
+
+      // stage 3
+      x0 = s4 + t2;
+      x1 = s4 - t2;
+      x2 = s7 - t3;
+      x3 = s7 + t3;
+
+      // stage 4
+      t0 = x0 * cospi_28_64 + x3 * cospi_4_64;
+      t1 = x1 * cospi_12_64 + x2 * cospi_20_64;
+      t2 = x2 * cospi_12_64 + x1 * -cospi_20_64;
+      t3 = x3 * cospi_28_64 + x0 * -cospi_4_64;
+      output[1 * 8] = (tran_low_t)fdct_round_shift(t0);
+      output[3 * 8] = (tran_low_t)fdct_round_shift(t2);
+      output[5 * 8] = (tran_low_t)fdct_round_shift(t1);
+      output[7 * 8] = (tran_low_t)fdct_round_shift(t3);
+      input++;
+      output++;
+    }
+  }
+
+  // Rows
+  for (i = 0; i < 8; ++i) {
+    fdct8(&intermediate[i * 8], &coeff_ptr[i * 8]);
+    for (j = 0; j < 8; ++j) coeff_ptr[j + i * 8] /= 2;
+  }
+
+  // TODO(jingning) Decide the need of these arguments after the
+  // quantization process is completed.
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)iscan;
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Quantization pass: All coefficients with index >= zero_flag are
+    // skippable. Note: zero_flag can be zero.
+    for (i = 0; i < n_coeffs; i++) {
+      const int rc = scan[i];
+      const int coeff = coeff_ptr[rc];
+#if CONFIG_AOM_QM
+      const qm_val_t wt = qm_ptr[rc];
+      const qm_val_t iwt = iqm_ptr[rc];
+      const int dequant =
+          (dequant_ptr[rc != 0] * iwt + (1 << (AOM_QM_BITS - 1))) >>
+          AOM_QM_BITS;
+#endif
+      const int coeff_sign = (coeff >> 31);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+
+      int64_t tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
+      int tmp32;
+#if CONFIG_AOM_QM
+      tmp32 = (int)((tmp * quant_ptr[rc != 0] * wt) >> (16 + AOM_QM_BITS));
+      qcoeff_ptr[rc] = (tmp32 ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant;
+#else
+      tmp32 = (int)((tmp * quant_ptr[rc != 0]) >> 16);
+      qcoeff_ptr[rc] = (tmp32 ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
+#endif
+
+      if (tmp32) eob = i;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+
+void av1_fht8x8_c(const int16_t *input, tran_low_t *output, int stride,
+                  int tx_type) {
+  if (tx_type == DCT_DCT) {
+    aom_fdct8x8_c(input, output, stride);
+  } else {
+    static const transform_2d FHT[] = {
+      { fdct8, fdct8 },    // DCT_DCT
+      { fadst8, fdct8 },   // ADST_DCT
+      { fdct8, fadst8 },   // DCT_ADST
+      { fadst8, fadst8 },  // ADST_ADST
+#if CONFIG_EXT_TX
+      { fadst8, fdct8 },   // FLIPADST_DCT
+      { fdct8, fadst8 },   // DCT_FLIPADST
+      { fadst8, fadst8 },  // FLIPADST_FLIPADST
+      { fadst8, fadst8 },  // ADST_FLIPADST
+      { fadst8, fadst8 },  // FLIPADST_ADST
+      { fidtx8, fidtx8 },  // IDTX
+      { fdct8, fidtx8 },   // V_DCT
+      { fidtx8, fdct8 },   // H_DCT
+      { fadst8, fidtx8 },  // V_ADST
+      { fidtx8, fadst8 },  // H_ADST
+      { fadst8, fidtx8 },  // V_FLIPADST
+      { fidtx8, fadst8 },  // H_FLIPADST
+#endif                     // CONFIG_EXT_TX
+    };
+    const transform_2d ht = FHT[tx_type];
+    tran_low_t out[64];
+    int i, j;
+    tran_low_t temp_in[8], temp_out[8];
+
+#if CONFIG_EXT_TX
+    int16_t flipped_input[8 * 8];
+    maybe_flip_input(&input, &stride, 8, 8, flipped_input, tx_type);
+#endif
+
+    // Columns
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j) temp_in[j] = input[j * stride + i] * 4;
+      ht.cols(temp_in, temp_out);
+      for (j = 0; j < 8; ++j) out[j * 8 + i] = temp_out[j];
+    }
+
+    // Rows
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j) temp_in[j] = out[j + i * 8];
+      ht.rows(temp_in, temp_out);
+      for (j = 0; j < 8; ++j)
+        output[j + i * 8] = (temp_out[j] + (temp_out[j] < 0)) >> 1;
+    }
+  }
+}
+
+/* 4-point reversible, orthonormal Walsh-Hadamard in 3.5 adds, 0.5 shifts per
+   pixel. */
+void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride) {
+  int i;
+  tran_high_t a1, b1, c1, d1, e1;
+  const int16_t *ip_pass0 = input;
+  const tran_low_t *ip = NULL;
+  tran_low_t *op = output;
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip_pass0[0 * stride];
+    b1 = ip_pass0[1 * stride];
+    c1 = ip_pass0[2 * stride];
+    d1 = ip_pass0[3 * stride];
+
+    a1 += b1;
+    d1 = d1 - c1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= c1;
+    d1 += b1;
+    op[0] = (tran_low_t)a1;
+    op[4] = (tran_low_t)c1;
+    op[8] = (tran_low_t)d1;
+    op[12] = (tran_low_t)b1;
+
+    ip_pass0++;
+    op++;
+  }
+  ip = output;
+  op = output;
+
+  for (i = 0; i < 4; i++) {
+    a1 = ip[0];
+    b1 = ip[1];
+    c1 = ip[2];
+    d1 = ip[3];
+
+    a1 += b1;
+    d1 -= c1;
+    e1 = (a1 - d1) >> 1;
+    b1 = e1 - b1;
+    c1 = e1 - c1;
+    a1 -= c1;
+    d1 += b1;
+    op[0] = (tran_low_t)(a1 * UNIT_QUANT_FACTOR);
+    op[1] = (tran_low_t)(c1 * UNIT_QUANT_FACTOR);
+    op[2] = (tran_low_t)(d1 * UNIT_QUANT_FACTOR);
+    op[3] = (tran_low_t)(b1 * UNIT_QUANT_FACTOR);
+
+    ip += 4;
+    op += 4;
+  }
+}
+
+void av1_fht16x16_c(const int16_t *input, tran_low_t *output, int stride,
+                    int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct16, fdct16 },    // DCT_DCT
+    { fadst16, fdct16 },   // ADST_DCT
+    { fdct16, fadst16 },   // DCT_ADST
+    { fadst16, fadst16 },  // ADST_ADST
+#if CONFIG_EXT_TX
+    { fadst16, fdct16 },   // FLIPADST_DCT
+    { fdct16, fadst16 },   // DCT_FLIPADST
+    { fadst16, fadst16 },  // FLIPADST_FLIPADST
+    { fadst16, fadst16 },  // ADST_FLIPADST
+    { fadst16, fadst16 },  // FLIPADST_ADST
+    { fidtx16, fidtx16 },  // IDTX
+    { fdct16, fidtx16 },   // V_DCT
+    { fidtx16, fdct16 },   // H_DCT
+    { fadst16, fidtx16 },  // V_ADST
+    { fidtx16, fadst16 },  // H_ADST
+    { fadst16, fidtx16 },  // V_FLIPADST
+    { fidtx16, fadst16 },  // H_FLIPADST
+#endif                     // CONFIG_EXT_TX
+  };
+
+  const transform_2d ht = FHT[tx_type];
+  tran_low_t out[256];
+  int i, j;
+  tran_low_t temp_in[16], temp_out[16];
+
+#if CONFIG_EXT_TX
+  int16_t flipped_input[16 * 16];
+  maybe_flip_input(&input, &stride, 16, 16, flipped_input, tx_type);
+#endif
+
+  // Columns
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j) temp_in[j] = input[j * stride + i] * 4;
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < 16; ++j)
+      out[j * 16 + i] = (temp_out[j] + 1 + (temp_out[j] < 0)) >> 2;
+  }
+
+  // Rows
+  for (i = 0; i < 16; ++i) {
+    for (j = 0; j < 16; ++j) temp_in[j] = out[j + i * 16];
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < 16; ++j) output[j + i * 16] = temp_out[j];
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_fht4x4_c(const int16_t *input, tran_low_t *output, int stride,
+                         int tx_type) {
+  av1_fht4x4_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht4x8_c(const int16_t *input, tran_low_t *output, int stride,
+                         int tx_type) {
+  av1_fht4x8_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht8x4_c(const int16_t *input, tran_low_t *output, int stride,
+                         int tx_type) {
+  av1_fht8x4_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht8x16_c(const int16_t *input, tran_low_t *output, int stride,
+                          int tx_type) {
+  av1_fht8x16_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht16x8_c(const int16_t *input, tran_low_t *output, int stride,
+                          int tx_type) {
+  av1_fht16x8_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht16x32_c(const int16_t *input, tran_low_t *output, int stride,
+                           int tx_type) {
+  av1_fht16x32_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht32x16_c(const int16_t *input, tran_low_t *output, int stride,
+                           int tx_type) {
+  av1_fht32x16_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht4x16_c(const int16_t *input, tran_low_t *output, int stride,
+                          int tx_type) {
+  av1_fht4x16_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht16x4_c(const int16_t *input, tran_low_t *output, int stride,
+                          int tx_type) {
+  av1_fht16x4_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht8x32_c(const int16_t *input, tran_low_t *output, int stride,
+                          int tx_type) {
+  av1_fht8x32_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht32x8_c(const int16_t *input, tran_low_t *output, int stride,
+                          int tx_type) {
+  av1_fht32x8_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fht8x8_c(const int16_t *input, tran_low_t *output, int stride,
+                         int tx_type) {
+  av1_fht8x8_c(input, output, stride, tx_type);
+}
+
+void av1_highbd_fwht4x4_c(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  av1_fwht4x4_c(input, output, stride);
+}
+
+void av1_highbd_fht16x16_c(const int16_t *input, tran_low_t *output, int stride,
+                           int tx_type) {
+  av1_fht16x16_c(input, output, stride, tx_type);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_fht32x32_c(const int16_t *input, tran_low_t *output, int stride,
+                    int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct32, fdct32 },  // DCT_DCT
+#if CONFIG_EXT_TX
+    { fhalfright32, fdct32 },        // ADST_DCT
+    { fdct32, fhalfright32 },        // DCT_ADST
+    { fhalfright32, fhalfright32 },  // ADST_ADST
+    { fhalfright32, fdct32 },        // FLIPADST_DCT
+    { fdct32, fhalfright32 },        // DCT_FLIPADST
+    { fhalfright32, fhalfright32 },  // FLIPADST_FLIPADST
+    { fhalfright32, fhalfright32 },  // ADST_FLIPADST
+    { fhalfright32, fhalfright32 },  // FLIPADST_ADST
+    { fidtx32, fidtx32 },            // IDTX
+    { fdct32, fidtx32 },             // V_DCT
+    { fidtx32, fdct32 },             // H_DCT
+    { fhalfright32, fidtx32 },       // V_ADST
+    { fidtx32, fhalfright32 },       // H_ADST
+    { fhalfright32, fidtx32 },       // V_FLIPADST
+    { fidtx32, fhalfright32 },       // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  tran_low_t out[1024];
+  int i, j;
+  tran_low_t temp_in[32], temp_out[32];
+
+#if CONFIG_EXT_TX
+  int16_t flipped_input[32 * 32];
+  maybe_flip_input(&input, &stride, 32, 32, flipped_input, tx_type);
+#endif
+
+  // Columns
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j) temp_in[j] = input[j * stride + i] * 4;
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < 32; ++j)
+      out[j * 32 + i] = ROUND_POWER_OF_TWO_SIGNED(temp_out[j], 4);
+  }
+
+  // Rows
+  for (i = 0; i < 32; ++i) {
+    for (j = 0; j < 32; ++j) temp_in[j] = out[j + i * 32];
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < 32; ++j) output[j + i * 32] = temp_out[j];
+  }
+}
+
+#if CONFIG_TX64X64
+#if CONFIG_EXT_TX
+static void fidtx64(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 64; ++i)
+    output[i] = (tran_low_t)fdct_round_shift(input[i] * 4 * Sqrt2);
+}
+
+// For use in lieu of ADST
+static void fhalfright64(const tran_low_t *input, tran_low_t *output) {
+  int i;
+  tran_low_t inputhalf[32];
+  for (i = 0; i < 32; ++i) {
+    output[32 + i] = (tran_low_t)fdct_round_shift(input[i] * 4 * Sqrt2);
+  }
+  // Multiply input by sqrt(2)
+  for (i = 0; i < 32; ++i) {
+    inputhalf[i] = (tran_low_t)fdct_round_shift(input[i + 32] * Sqrt2);
+  }
+  fdct32(inputhalf, output);
+  // Note overall scaling factor is 2 times unitary
+}
+#endif  // CONFIG_EXT_TX
+
+static void fdct64_col(const tran_low_t *input, tran_low_t *output) {
+  int32_t in[64], out[64];
+  int i;
+  for (i = 0; i < 64; ++i) in[i] = (int32_t)input[i];
+  av1_fdct64_new(in, out, fwd_cos_bit_col_dct_dct_64,
+                 fwd_stage_range_col_dct_dct_64);
+  for (i = 0; i < 64; ++i) output[i] = (tran_low_t)out[i];
+}
+
+static void fdct64_row(const tran_low_t *input, tran_low_t *output) {
+  int32_t in[64], out[64];
+  int i;
+  for (i = 0; i < 64; ++i) in[i] = (int32_t)input[i];
+  av1_fdct64_new(in, out, fwd_cos_bit_row_dct_dct_64,
+                 fwd_stage_range_row_dct_dct_64);
+  for (i = 0; i < 64; ++i) output[i] = (tran_low_t)out[i];
+}
+
+void av1_fht64x64_c(const int16_t *input, tran_low_t *output, int stride,
+                    int tx_type) {
+  static const transform_2d FHT[] = {
+    { fdct64_col, fdct64_row },  // DCT_DCT
+#if CONFIG_EXT_TX
+    { fhalfright64, fdct64_row },    // ADST_DCT
+    { fdct64_col, fhalfright64 },    // DCT_ADST
+    { fhalfright64, fhalfright64 },  // ADST_ADST
+    { fhalfright64, fdct64_row },    // FLIPADST_DCT
+    { fdct64_col, fhalfright64 },    // DCT_FLIPADST
+    { fhalfright64, fhalfright64 },  // FLIPADST_FLIPADST
+    { fhalfright64, fhalfright64 },  // ADST_FLIPADST
+    { fhalfright64, fhalfright64 },  // FLIPADST_ADST
+    { fidtx64, fidtx64 },            // IDTX
+    { fdct64_col, fidtx64 },         // V_DCT
+    { fidtx64, fdct64_row },         // H_DCT
+    { fhalfright64, fidtx64 },       // V_ADST
+    { fidtx64, fhalfright64 },       // H_ADST
+    { fhalfright64, fidtx64 },       // V_FLIPADST
+    { fidtx64, fhalfright64 },       // H_FLIPADST
+#endif
+  };
+  const transform_2d ht = FHT[tx_type];
+  tran_low_t out[4096];
+  int i, j;
+  tran_low_t temp_in[64], temp_out[64];
+#if CONFIG_EXT_TX
+  int16_t flipped_input[64 * 64];
+  maybe_flip_input(&input, &stride, 64, 64, flipped_input, tx_type);
+#endif
+  // Columns
+  for (i = 0; i < 64; ++i) {
+    for (j = 0; j < 64; ++j) temp_in[j] = input[j * stride + i];
+    ht.cols(temp_in, temp_out);
+    for (j = 0; j < 64; ++j)
+      out[j * 64 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
+  }
+
+  // Rows
+  for (i = 0; i < 64; ++i) {
+    for (j = 0; j < 64; ++j) temp_in[j] = out[j + i * 64];
+    ht.rows(temp_in, temp_out);
+    for (j = 0; j < 64; ++j)
+      output[j + i * 64] =
+          (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2);
+  }
+}
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_EXT_TX
+// Forward identity transform.
+void av1_fwd_idtx_c(const int16_t *src_diff, tran_low_t *coeff, int stride,
+                    int bs, int tx_type) {
+  int r, c;
+  const int shift = bs < 32 ? 3 : (bs < 64 ? 2 : 1);
+  if (tx_type == IDTX) {
+    for (r = 0; r < bs; ++r) {
+      for (c = 0; c < bs; ++c) coeff[c] = src_diff[c] * (1 << shift);
+      src_diff += stride;
+      coeff += bs;
+    }
+  }
+}
+#endif  // CONFIG_EXT_TX
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_fht32x32_c(const int16_t *input, tran_low_t *output, int stride,
+                           int tx_type) {
+  av1_fht32x32_c(input, output, stride, tx_type);
+}
+
+#if CONFIG_TX64X64
+void av1_highbd_fht64x64_c(const int16_t *input, tran_low_t *output, int stride,
+                           int tx_type) {
+  av1_fht64x64_c(input, output, stride, tx_type);
+}
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // !AV1_DCT_GTEST
diff --git a/third_party/aom/av1/encoder/encint.h b/third_party/aom/av1/encoder/encint.h
new file mode 100644
index 0000000000..30ea8521fb
--- /dev/null
+++ b/third_party/aom/av1/encoder/encint.h
@@ -0,0 +1,51 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+/* clang-format off */
+
+#if !defined(_encint_H)
+# define _encint_H (1)
+
+typedef struct daala_enc_ctx od_enc_ctx;
+typedef struct od_params_ctx od_params_ctx;
+typedef struct od_rollback_buffer od_rollback_buffer;
+
+# include "aom_dsp/entenc.h"
+# include "av1/common/odintrin.h"
+# include "av1/common/pvq_state.h"
+
+struct daala_enc_ctx{
+  /* Stores context-adaptive CDFs for PVQ. */
+  od_state state;
+  /* AOM entropy encoder. */
+  aom_writer w;
+  int use_activity_masking;
+  /* Mode of quantization matrice : FLAT (0) or HVS (1) */
+  int qm;
+  /*Normalized PVQ lambda for use where we've already performed
+     quantization.*/
+  double pvq_norm_lambda;
+  double pvq_norm_lambda_dc;
+};
+
+// from daalaenc.h
+/**The encoder context.*/
+typedef struct daala_enc_ctx daala_enc_ctx;
+
+/** Holds important encoder information so we can roll back decisions */
+struct od_rollback_buffer {
+  od_ec_enc ec;
+  od_adapt_ctx adapt;
+};
+
+void od_encode_checkpoint(const daala_enc_ctx *enc, od_rollback_buffer *rbuf);
+void od_encode_rollback(daala_enc_ctx *enc, const od_rollback_buffer *rbuf);
+
+#endif
diff --git a/third_party/aom/av1/encoder/encodeframe.c b/third_party/aom/av1/encoder/encodeframe.c
new file mode 100644
index 0000000000..d254157e72
--- /dev/null
+++ b/third_party/aom/av1/encoder/encodeframe.c
@@ -0,0 +1,7160 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "./aom_config.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/binary_codes_writer.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/aom_timer.h"
+#include "aom_ports/system_state.h"
+
+#include "av1/common/common.h"
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/idct.h"
+#include "av1/common/mv.h"
+#include "av1/common/mvref_common.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/reconintra.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/seg_common.h"
+#include "av1/common/tile_common.h"
+
+#include "av1/encoder/aq_complexity.h"
+#include "av1/encoder/aq_cyclicrefresh.h"
+#include "av1/encoder/aq_variance.h"
+#if CONFIG_SUPERTX
+#include "av1/encoder/cost.h"
+#endif
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#include "av1/common/warped_motion.h"
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+#if CONFIG_GLOBAL_MOTION
+#include "av1/encoder/global_motion.h"
+#endif  // CONFIG_GLOBAL_MOTION
+#include "av1/encoder/encodeframe.h"
+#include "av1/encoder/encodemb.h"
+#include "av1/encoder/encodemv.h"
+#if CONFIG_LV_MAP
+#include "av1/encoder/encodetxb.h"
+#endif
+#include "av1/encoder/ethread.h"
+#include "av1/encoder/extend.h"
+#include "av1/encoder/rd.h"
+#include "av1/encoder/rdopt.h"
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/tokenize.h"
+#if CONFIG_PVQ
+#include "av1/common/pvq.h"
+#include "av1/encoder/pvq_encoder.h"
+#endif
+#if CONFIG_HIGHBITDEPTH
+#define IF_HBD(...) __VA_ARGS__
+#else
+#define IF_HBD(...)
+#endif  // CONFIG_HIGHBITDEPTH
+
+static void encode_superblock(const AV1_COMP *const cpi, ThreadData *td,
+                              TOKENEXTRA **t, RUN_TYPE dry_run, int mi_row,
+                              int mi_col, BLOCK_SIZE bsize,
+                              PICK_MODE_CONTEXT *ctx, int *rate);
+
+#if CONFIG_SUPERTX
+static int check_intra_b(PICK_MODE_CONTEXT *ctx);
+
+static int check_intra_sb(const AV1_COMP *cpi, const TileInfo *const tile,
+                          int mi_row, int mi_col, BLOCK_SIZE bsize,
+                          PC_TREE *pc_tree);
+static void predict_superblock(const AV1_COMP *const cpi, ThreadData *td,
+#if CONFIG_EXT_INTER
+                               int mi_row_ori, int mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                               int mi_row_pred, int mi_col_pred,
+                               BLOCK_SIZE bsize_pred, int b_sub8x8, int block);
+static int check_supertx_sb(BLOCK_SIZE bsize, TX_SIZE supertx_size,
+                            PC_TREE *pc_tree);
+static void predict_sb_complex(const AV1_COMP *const cpi, ThreadData *td,
+                               const TileInfo *const tile, int mi_row,
+                               int mi_col, int mi_row_ori, int mi_col_ori,
+                               RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                               BLOCK_SIZE top_bsize, uint8_t *dst_buf[3],
+                               int dst_stride[3], PC_TREE *pc_tree);
+static void update_state_sb_supertx(const AV1_COMP *const cpi, ThreadData *td,
+                                    const TileInfo *const tile, int mi_row,
+                                    int mi_col, BLOCK_SIZE bsize,
+                                    RUN_TYPE dry_run, PC_TREE *pc_tree);
+static void rd_supertx_sb(const AV1_COMP *const cpi, ThreadData *td,
+                          const TileInfo *const tile, int mi_row, int mi_col,
+                          BLOCK_SIZE bsize, int *tmp_rate, int64_t *tmp_dist,
+                          TX_TYPE *best_tx, PC_TREE *pc_tree);
+#endif  // CONFIG_SUPERTX
+
+// This is used as a reference when computing the source variance for the
+//  purposes of activity masking.
+// Eventually this should be replaced by custom no-reference routines,
+//  which will be faster.
+static const uint8_t AV1_VAR_OFFS[MAX_SB_SIZE] = {
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+#if CONFIG_EXT_PARTITION
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
+#endif  // CONFIG_EXT_PARTITION
+};
+
+#if CONFIG_HIGHBITDEPTH
+static const uint16_t AV1_HIGH_VAR_OFFS_8[MAX_SB_SIZE] = {
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+#if CONFIG_EXT_PARTITION
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
+  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
+#endif  // CONFIG_EXT_PARTITION
+};
+
+static const uint16_t AV1_HIGH_VAR_OFFS_10[MAX_SB_SIZE] = {
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+#if CONFIG_EXT_PARTITION
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
+  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4
+#endif  // CONFIG_EXT_PARTITION
+};
+
+static const uint16_t AV1_HIGH_VAR_OFFS_12[MAX_SB_SIZE] = {
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16,
+#if CONFIG_EXT_PARTITION
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
+  128 * 16
+#endif  // CONFIG_EXT_PARTITION
+};
+#endif  // CONFIG_HIGHBITDEPTH
+
+unsigned int av1_get_sby_perpixel_variance(const AV1_COMP *cpi,
+                                           const struct buf_2d *ref,
+                                           BLOCK_SIZE bs) {
+  unsigned int sse;
+  const unsigned int var =
+      cpi->fn_ptr[bs].vf(ref->buf, ref->stride, AV1_VAR_OFFS, 0, &sse);
+  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
+}
+
+#if CONFIG_HIGHBITDEPTH
+unsigned int av1_high_get_sby_perpixel_variance(const AV1_COMP *cpi,
+                                                const struct buf_2d *ref,
+                                                BLOCK_SIZE bs, int bd) {
+  unsigned int var, sse;
+  switch (bd) {
+    case 10:
+      var =
+          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
+                             CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_10), 0, &sse);
+      break;
+    case 12:
+      var =
+          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
+                             CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_12), 0, &sse);
+      break;
+    case 8:
+    default:
+      var =
+          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
+                             CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_8), 0, &sse);
+      break;
+  }
+  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static unsigned int get_sby_perpixel_diff_variance(const AV1_COMP *const cpi,
+                                                   const struct buf_2d *ref,
+                                                   int mi_row, int mi_col,
+                                                   BLOCK_SIZE bs) {
+  unsigned int sse, var;
+  uint8_t *last_y;
+  const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
+
+  assert(last != NULL);
+  last_y =
+      &last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
+  var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
+  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
+}
+
+static BLOCK_SIZE get_rd_var_based_fixed_partition(AV1_COMP *cpi, MACROBLOCK *x,
+                                                   int mi_row, int mi_col) {
+  unsigned int var = get_sby_perpixel_diff_variance(
+      cpi, &x->plane[0].src, mi_row, mi_col, BLOCK_64X64);
+  if (var < 8)
+    return BLOCK_64X64;
+  else if (var < 128)
+    return BLOCK_32X32;
+  else if (var < 2048)
+    return BLOCK_16X16;
+  else
+    return BLOCK_8X8;
+}
+
+// Lighter version of set_offsets that only sets the mode info
+// pointers.
+static void set_mode_info_offsets(const AV1_COMP *const cpi,
+                                  MACROBLOCK *const x, MACROBLOCKD *const xd,
+                                  int mi_row, int mi_col) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int idx_str = xd->mi_stride * mi_row + mi_col;
+  xd->mi = cm->mi_grid_visible + idx_str;
+  xd->mi[0] = cm->mi + idx_str;
+  x->mbmi_ext = cpi->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
+}
+
+static void set_offsets_without_segment_id(const AV1_COMP *const cpi,
+                                           const TileInfo *const tile,
+                                           MACROBLOCK *const x, int mi_row,
+                                           int mi_col, BLOCK_SIZE bsize) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int mi_width = mi_size_wide[bsize];
+  const int mi_height = mi_size_high[bsize];
+
+  set_skip_context(xd, mi_row, mi_col);
+
+  set_mode_info_offsets(cpi, x, xd, mi_row, mi_col);
+
+#if CONFIG_VAR_TX
+  xd->above_txfm_context = cm->above_txfm_context + mi_col;
+  xd->left_txfm_context =
+      xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+  xd->max_tx_size = max_txsize_lookup[bsize];
+#endif
+
+  // Set up destination pointers.
+  av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                       mi_col);
+
+  // Set up limit values for MV components.
+  // Mv beyond the range do not produce new/different prediction block.
+  x->mv_limits.row_min =
+      -(((mi_row + mi_height) * MI_SIZE) + AOM_INTERP_EXTEND);
+  x->mv_limits.col_min = -(((mi_col + mi_width) * MI_SIZE) + AOM_INTERP_EXTEND);
+  x->mv_limits.row_max = (cm->mi_rows - mi_row) * MI_SIZE + AOM_INTERP_EXTEND;
+  x->mv_limits.col_max = (cm->mi_cols - mi_col) * MI_SIZE + AOM_INTERP_EXTEND;
+
+  set_plane_n4(xd, mi_width, mi_height);
+
+  // Set up distance of MB to edge of frame in 1/8th pel units.
+  assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
+  set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+  // Set up source buffers.
+  av1_setup_src_planes(x, cpi->source, mi_row, mi_col);
+
+  // R/D setup.
+  x->rddiv = cpi->rd.RDDIV;
+  x->rdmult = cpi->rd.RDMULT;
+
+  // required by av1_append_sub8x8_mvs_for_idx() and av1_find_best_ref_mvs()
+  xd->tile = *tile;
+}
+
+static void set_offsets(const AV1_COMP *const cpi, const TileInfo *const tile,
+                        MACROBLOCK *const x, int mi_row, int mi_col,
+                        BLOCK_SIZE bsize) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi;
+  const struct segmentation *const seg = &cm->seg;
+
+  set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize);
+
+  mbmi = &xd->mi[0]->mbmi;
+
+  // Setup segment ID.
+  if (seg->enabled) {
+    if (!cpi->vaq_refresh) {
+      const uint8_t *const map =
+          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
+      mbmi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
+    }
+    av1_init_plane_quantizers(cpi, x, mbmi->segment_id);
+  } else {
+    mbmi->segment_id = 0;
+  }
+
+#if CONFIG_SUPERTX
+  mbmi->segment_id_supertx = MAX_SEGMENTS;
+#endif  // CONFIG_SUPERTX
+}
+
+#if CONFIG_SUPERTX
+static void set_offsets_supertx(const AV1_COMP *const cpi, ThreadData *td,
+                                const TileInfo *const tile, int mi_row,
+                                int mi_col, BLOCK_SIZE bsize) {
+  MACROBLOCK *const x = &td->mb;
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int mi_width = mi_size_wide[bsize];
+  const int mi_height = mi_size_high[bsize];
+#if CONFIG_DEPENDENT_HORZTILES
+  set_mode_info_offsets(cpi, x, xd, mi_row, mi_col, cm->dependent_horz_tiles);
+#else
+  set_mode_info_offsets(cpi, x, xd, mi_row, mi_col);
+#endif
+
+  // Set up distance of MB to edge of frame in 1/8th pel units.
+  assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
+  set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+}
+
+static void set_offsets_extend(const AV1_COMP *const cpi, ThreadData *td,
+                               const TileInfo *const tile, int mi_row_pred,
+                               int mi_col_pred, int mi_row_ori, int mi_col_ori,
+                               BLOCK_SIZE bsize_pred) {
+  // Used in supertx
+  // (mi_row_ori, mi_col_ori, bsize_ori): region for mv
+  // (mi_row_pred, mi_col_pred, bsize_pred): region to predict
+  MACROBLOCK *const x = &td->mb;
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int mi_width = mi_size_wide[bsize_pred];
+  const int mi_height = mi_size_high[bsize_pred];
+
+#if CONFIG_DEPENDENT_HORZTILES
+  set_mode_info_offsets(cpi, x, xd, mi_row_ori, mi_col_ori,
+                        cm->dependent_horz_tiles);
+#else
+  set_mode_info_offsets(cpi, x, xd, mi_row_ori, mi_col_ori);
+#endif
+
+  // Set up limit values for MV components.
+  // Mv beyond the range do not produce new/different prediction block.
+  x->mv_limits.row_min =
+      -(((mi_row_pred + mi_height) * MI_SIZE) + AOM_INTERP_EXTEND);
+  x->mv_limits.col_min =
+      -(((mi_col_pred + mi_width) * MI_SIZE) + AOM_INTERP_EXTEND);
+  x->mv_limits.row_max =
+      (cm->mi_rows - mi_row_pred) * MI_SIZE + AOM_INTERP_EXTEND;
+  x->mv_limits.col_max =
+      (cm->mi_cols - mi_col_pred) * MI_SIZE + AOM_INTERP_EXTEND;
+
+// Set up distance of MB to edge of frame in 1/8th pel units.
+#if !CONFIG_CB4X4
+  assert(!(mi_col_pred & (mi_width - mi_size_wide[BLOCK_8X8])) &&
+         !(mi_row_pred & (mi_height - mi_size_high[BLOCK_8X8])));
+#endif
+  set_mi_row_col(xd, tile, mi_row_pred, mi_height, mi_col_pred, mi_width,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+  xd->up_available = (mi_row_ori > tile->mi_row_start);
+  xd->left_available = (mi_col_ori > tile->mi_col_start);
+
+  // R/D setup.
+  x->rddiv = cpi->rd.RDDIV;
+  x->rdmult = cpi->rd.RDMULT;
+}
+
+static void set_segment_id_supertx(const AV1_COMP *const cpi,
+                                   MACROBLOCK *const x, const int mi_row,
+                                   const int mi_col, const BLOCK_SIZE bsize) {
+  const AV1_COMMON *cm = &cpi->common;
+  const struct segmentation *seg = &cm->seg;
+  const int miw = AOMMIN(mi_size_wide[bsize], cm->mi_cols - mi_col);
+  const int mih = AOMMIN(mi_size_high[bsize], cm->mi_rows - mi_row);
+  const int mi_offset = mi_row * cm->mi_stride + mi_col;
+  MODE_INFO **const mip = cm->mi_grid_visible + mi_offset;
+  int r, c;
+  int seg_id_supertx = MAX_SEGMENTS;
+
+  if (!seg->enabled) {
+    seg_id_supertx = 0;
+  } else {
+    // Find the minimum segment_id
+    for (r = 0; r < mih; r++)
+      for (c = 0; c < miw; c++)
+        seg_id_supertx =
+            AOMMIN(mip[r * cm->mi_stride + c]->mbmi.segment_id, seg_id_supertx);
+    assert(0 <= seg_id_supertx && seg_id_supertx < MAX_SEGMENTS);
+
+    // Initialize plane quantisers
+    av1_init_plane_quantizers(cpi, x, seg_id_supertx);
+  }
+
+  // Assign the the segment_id back to segment_id_supertx
+  for (r = 0; r < mih; r++)
+    for (c = 0; c < miw; c++)
+      mip[r * cm->mi_stride + c]->mbmi.segment_id_supertx = seg_id_supertx;
+}
+#endif  // CONFIG_SUPERTX
+
+static void set_block_size(AV1_COMP *const cpi, MACROBLOCK *const x,
+                           MACROBLOCKD *const xd, int mi_row, int mi_col,
+                           BLOCK_SIZE bsize) {
+  if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
+    const int mi_width = AOMMAX(mi_size_wide[bsize], mi_size_wide[BLOCK_8X8]);
+    const int mi_height = AOMMAX(mi_size_high[bsize], mi_size_high[BLOCK_8X8]);
+    for (int r = 0; r < mi_height; ++r) {
+      for (int c = 0; c < mi_width; ++c) {
+        set_mode_info_offsets(cpi, x, xd, mi_row + r, mi_col + c);
+        xd->mi[0]->mbmi.sb_type = bsize;
+      }
+    }
+  }
+}
+
+static void set_vt_partitioning(AV1_COMP *cpi, MACROBLOCK *const x,
+                                MACROBLOCKD *const xd, VAR_TREE *vt, int mi_row,
+                                int mi_col, const int64_t *const threshold,
+                                const BLOCK_SIZE *const bsize_min) {
+  AV1_COMMON *const cm = &cpi->common;
+  const int hbw = mi_size_wide[vt->bsize] / 2;
+  const int hbh = mi_size_high[vt->bsize] / 2;
+  const int has_cols = mi_col + hbw < cm->mi_cols;
+  const int has_rows = mi_row + hbh < cm->mi_rows;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  assert(vt->bsize >= BLOCK_8X8);
+
+  assert(hbh == hbw);
+
+  if (vt->bsize == BLOCK_8X8 && cm->frame_type != KEY_FRAME) {
+    set_block_size(cpi, x, xd, mi_row, mi_col, BLOCK_8X8);
+    return;
+  }
+
+  if (vt->force_split || (!has_cols && !has_rows)) goto split;
+
+  // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
+  // variance is below threshold, otherwise split will be selected.
+  // No check for vert/horiz split as too few samples for variance.
+  if (vt->bsize == bsize_min[0]) {
+    if (has_cols && has_rows && vt->variances.none.variance < threshold[0]) {
+      set_block_size(cpi, x, xd, mi_row, mi_col, vt->bsize);
+      return;
+    } else {
+      BLOCK_SIZE subsize = get_subsize(vt->bsize, PARTITION_SPLIT);
+      set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
+      if (vt->bsize > BLOCK_8X8) {
+        set_block_size(cpi, x, xd, mi_row, mi_col + hbw, subsize);
+        set_block_size(cpi, x, xd, mi_row + hbh, mi_col, subsize);
+        set_block_size(cpi, x, xd, mi_row + hbh, mi_col + hbw, subsize);
+      }
+      return;
+    }
+  } else if (vt->bsize > bsize_min[0]) {
+    // For key frame: take split for bsize above 32X32 or very high variance.
+    if (cm->frame_type == KEY_FRAME &&
+        (vt->bsize > BLOCK_32X32 ||
+         vt->variances.none.variance > (threshold[0] << 4))) {
+      goto split;
+    }
+    // If variance is low, take the bsize (no split).
+    if (has_cols && has_rows && vt->variances.none.variance < threshold[0]) {
+      set_block_size(cpi, x, xd, mi_row, mi_col, vt->bsize);
+      return;
+    }
+
+    // Check vertical split.
+    if (has_rows) {
+      BLOCK_SIZE subsize = get_subsize(vt->bsize, PARTITION_VERT);
+      if (vt->variances.vert[0].variance < threshold[0] &&
+          vt->variances.vert[1].variance < threshold[0] &&
+          get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
+        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
+        set_block_size(cpi, x, xd, mi_row, mi_col + hbw, subsize);
+        return;
+      }
+    }
+    // Check horizontal split.
+    if (has_cols) {
+      BLOCK_SIZE subsize = get_subsize(vt->bsize, PARTITION_HORZ);
+      if (vt->variances.horz[0].variance < threshold[0] &&
+          vt->variances.horz[1].variance < threshold[0] &&
+          get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
+        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
+        set_block_size(cpi, x, xd, mi_row + hbh, mi_col, subsize);
+        return;
+      }
+    }
+  }
+
+split : {
+  set_vt_partitioning(cpi, x, xd, vt->split[0], mi_row, mi_col, threshold + 1,
+                      bsize_min + 1);
+  set_vt_partitioning(cpi, x, xd, vt->split[1], mi_row, mi_col + hbw,
+                      threshold + 1, bsize_min + 1);
+  set_vt_partitioning(cpi, x, xd, vt->split[2], mi_row + hbh, mi_col,
+                      threshold + 1, bsize_min + 1);
+  set_vt_partitioning(cpi, x, xd, vt->split[3], mi_row + hbh, mi_col + hbw,
+                      threshold + 1, bsize_min + 1);
+  return;
+}
+}
+
+// Set the variance split thresholds for following the block sizes:
+// 0 - threshold_64x64, 1 - threshold_32x32, 2 - threshold_16x16,
+// 3 - vbp_threshold_8x8. vbp_threshold_8x8 (to split to 4x4 partition) is
+// currently only used on key frame.
+static void set_vbp_thresholds(AV1_COMP *cpi, int64_t thresholds[], int q) {
+  AV1_COMMON *const cm = &cpi->common;
+  const int is_key_frame = (cm->frame_type == KEY_FRAME);
+  const int threshold_multiplier = is_key_frame ? 20 : 1;
+  const int64_t threshold_base =
+      (int64_t)(threshold_multiplier * cpi->y_dequant[q][1]);
+  if (is_key_frame) {
+    thresholds[1] = threshold_base;
+    thresholds[2] = threshold_base >> 2;
+    thresholds[3] = threshold_base >> 2;
+    thresholds[4] = threshold_base << 2;
+  } else {
+    thresholds[2] = threshold_base;
+    if (cm->width <= 352 && cm->height <= 288) {
+      thresholds[1] = threshold_base >> 2;
+      thresholds[3] = threshold_base << 3;
+    } else {
+      thresholds[1] = threshold_base;
+      thresholds[2] = (5 * threshold_base) >> 2;
+      if (cm->width >= 1920 && cm->height >= 1080)
+        thresholds[2] = (7 * threshold_base) >> 2;
+      thresholds[3] = threshold_base << cpi->oxcf.speed;
+    }
+  }
+  thresholds[0] = INT64_MIN;
+}
+
+void av1_set_variance_partition_thresholds(AV1_COMP *cpi, int q) {
+  AV1_COMMON *const cm = &cpi->common;
+  SPEED_FEATURES *const sf = &cpi->sf;
+  const int is_key_frame = (cm->frame_type == KEY_FRAME);
+  if (sf->partition_search_type != VAR_BASED_PARTITION &&
+      sf->partition_search_type != REFERENCE_PARTITION) {
+    return;
+  } else {
+    set_vbp_thresholds(cpi, cpi->vbp_thresholds, q);
+    // The thresholds below are not changed locally.
+    if (is_key_frame) {
+      cpi->vbp_threshold_sad = 0;
+      cpi->vbp_bsize_min = BLOCK_8X8;
+    } else {
+      if (cm->width <= 352 && cm->height <= 288)
+        cpi->vbp_threshold_sad = 100;
+      else
+        cpi->vbp_threshold_sad = (cpi->y_dequant[q][1] << 1) > 1000
+                                     ? (cpi->y_dequant[q][1] << 1)
+                                     : 1000;
+      cpi->vbp_bsize_min = BLOCK_16X16;
+    }
+    cpi->vbp_threshold_minmax = 15 + (q >> 3);
+  }
+}
+
+// Compute the minmax over the 8x8 subblocks.
+static int compute_minmax_8x8(const uint8_t *src, int src_stride,
+                              const uint8_t *ref, int ref_stride,
+#if CONFIG_HIGHBITDEPTH
+                              int highbd,
+#endif
+                              int pixels_wide, int pixels_high) {
+  int k;
+  int minmax_max = 0;
+  int minmax_min = 255;
+  // Loop over the 4 8x8 subblocks.
+  for (k = 0; k < 4; k++) {
+    const int x8_idx = ((k & 1) << 3);
+    const int y8_idx = ((k >> 1) << 3);
+    int min = 0;
+    int max = 0;
+    if (x8_idx < pixels_wide && y8_idx < pixels_high) {
+      const int src_offset = y8_idx * src_stride + x8_idx;
+      const int ref_offset = y8_idx * ref_stride + x8_idx;
+#if CONFIG_HIGHBITDEPTH
+      if (highbd) {
+        aom_highbd_minmax_8x8(src + src_offset, src_stride, ref + ref_offset,
+                              ref_stride, &min, &max);
+      } else {
+        aom_minmax_8x8(src + src_offset, src_stride, ref + ref_offset,
+                       ref_stride, &min, &max);
+      }
+#else
+      aom_minmax_8x8(src + src_offset, src_stride, ref + ref_offset, ref_stride,
+                     &min, &max);
+#endif
+      if ((max - min) > minmax_max) minmax_max = (max - min);
+      if ((max - min) < minmax_min) minmax_min = (max - min);
+    }
+  }
+  return (minmax_max - minmax_min);
+}
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE int avg_4x4(const uint8_t *const src, const int stride,
+                          const int highbd) {
+  if (highbd) {
+    return aom_highbd_avg_4x4(src, stride);
+  } else {
+    return aom_avg_4x4(src, stride);
+  }
+}
+#else
+static INLINE int avg_4x4(const uint8_t *const src, const int stride) {
+  return aom_avg_4x4(src, stride);
+}
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+static INLINE int avg_8x8(const uint8_t *const src, const int stride,
+                          const int highbd) {
+  if (highbd) {
+    return aom_highbd_avg_8x8(src, stride);
+  } else {
+    return aom_avg_8x8(src, stride);
+  }
+}
+#else
+static INLINE int avg_8x8(const uint8_t *const src, const int stride) {
+  return aom_avg_8x8(src, stride);
+}
+#endif
+
+static void init_variance_tree(VAR_TREE *const vt,
+#if CONFIG_HIGHBITDEPTH
+                               const int highbd,
+#endif
+                               BLOCK_SIZE bsize, BLOCK_SIZE leaf_size,
+                               const int width, const int height,
+                               const uint8_t *const src, const int src_stride,
+                               const uint8_t *const ref, const int ref_stride) {
+  assert(bsize >= leaf_size);
+
+  vt->bsize = bsize;
+
+  vt->force_split = 0;
+
+  vt->src = src;
+  vt->src_stride = src_stride;
+  vt->ref = ref;
+  vt->ref_stride = ref_stride;
+
+  vt->width = width;
+  vt->height = height;
+
+#if CONFIG_HIGHBITDEPTH
+  vt->highbd = highbd;
+#endif  // CONFIG_HIGHBITDEPTH
+
+  if (bsize > leaf_size) {
+    const BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
+    const int px = block_size_wide[subsize];
+
+    init_variance_tree(vt->split[0],
+#if CONFIG_HIGHBITDEPTH
+                       highbd,
+#endif  // CONFIG_HIGHBITDEPTH
+                       subsize, leaf_size, AOMMIN(px, width),
+                       AOMMIN(px, height), src, src_stride, ref, ref_stride);
+    init_variance_tree(vt->split[1],
+#if CONFIG_HIGHBITDEPTH
+                       highbd,
+#endif  // CONFIG_HIGHBITDEPTH
+                       subsize, leaf_size, width - px, AOMMIN(px, height),
+                       src + px, src_stride, ref + px, ref_stride);
+    init_variance_tree(vt->split[2],
+#if CONFIG_HIGHBITDEPTH
+                       highbd,
+#endif  // CONFIG_HIGHBITDEPTH
+                       subsize, leaf_size, AOMMIN(px, width), height - px,
+                       src + px * src_stride, src_stride, ref + px * ref_stride,
+                       ref_stride);
+    init_variance_tree(vt->split[3],
+#if CONFIG_HIGHBITDEPTH
+                       highbd,
+#endif  // CONFIG_HIGHBITDEPTH
+                       subsize, leaf_size, width - px, height - px,
+                       src + px * src_stride + px, src_stride,
+                       ref + px * ref_stride + px, ref_stride);
+  }
+}
+
+// Fill the variance tree based on averaging pixel values (sub-sampling), at
+// the leaf node size.
+static void fill_variance_tree(VAR_TREE *const vt, const BLOCK_SIZE leaf_size) {
+  if (vt->bsize > leaf_size) {
+    fill_variance_tree(vt->split[0], leaf_size);
+    fill_variance_tree(vt->split[1], leaf_size);
+    fill_variance_tree(vt->split[2], leaf_size);
+    fill_variance_tree(vt->split[3], leaf_size);
+    fill_variance_node(vt);
+  } else if (vt->width <= 0 || vt->height <= 0) {
+    fill_variance(0, 0, 0, &vt->variances.none);
+  } else {
+    unsigned int sse = 0;
+    int sum = 0;
+    int src_avg;
+    int ref_avg;
+    assert(leaf_size == BLOCK_4X4 || leaf_size == BLOCK_8X8);
+    if (leaf_size == BLOCK_4X4) {
+      src_avg = avg_4x4(vt->src, vt->src_stride IF_HBD(, vt->highbd));
+      ref_avg = avg_4x4(vt->ref, vt->ref_stride IF_HBD(, vt->highbd));
+    } else {
+      src_avg = avg_8x8(vt->src, vt->src_stride IF_HBD(, vt->highbd));
+      ref_avg = avg_8x8(vt->ref, vt->ref_stride IF_HBD(, vt->highbd));
+    }
+    sum = src_avg - ref_avg;
+    sse = sum * sum;
+    fill_variance(sse, sum, 0, &vt->variances.none);
+  }
+}
+
+static void refine_variance_tree(VAR_TREE *const vt, const int64_t threshold) {
+  if (vt->bsize >= BLOCK_8X8) {
+    if (vt->bsize == BLOCK_16X16) {
+      if (vt->variances.none.variance <= threshold)
+        return;
+      else
+        vt->force_split = 0;
+    }
+
+    refine_variance_tree(vt->split[0], threshold);
+    refine_variance_tree(vt->split[1], threshold);
+    refine_variance_tree(vt->split[2], threshold);
+    refine_variance_tree(vt->split[3], threshold);
+
+    if (vt->bsize <= BLOCK_16X16) fill_variance_node(vt);
+  } else if (vt->width <= 0 || vt->height <= 0) {
+    fill_variance(0, 0, 0, &vt->variances.none);
+  } else {
+    const int src_avg = avg_4x4(vt->src, vt->src_stride IF_HBD(, vt->highbd));
+    const int ref_avg = avg_4x4(vt->ref, vt->ref_stride IF_HBD(, vt->highbd));
+    const int sum = src_avg - ref_avg;
+    const unsigned int sse = sum * sum;
+    assert(vt->bsize == BLOCK_4X4);
+    fill_variance(sse, sum, 0, &vt->variances.none);
+  }
+}
+
+static int check_split_key_frame(VAR_TREE *const vt, const int64_t threshold) {
+  if (vt->bsize == BLOCK_32X32) {
+    vt->force_split = vt->variances.none.variance > threshold;
+  } else {
+    vt->force_split |= check_split_key_frame(vt->split[0], threshold);
+    vt->force_split |= check_split_key_frame(vt->split[1], threshold);
+    vt->force_split |= check_split_key_frame(vt->split[2], threshold);
+    vt->force_split |= check_split_key_frame(vt->split[3], threshold);
+  }
+  return vt->force_split;
+}
+
+static int check_split(AV1_COMP *const cpi, VAR_TREE *const vt,
+                       const int segment_id, const int64_t *const thresholds) {
+  if (vt->bsize == BLOCK_16X16) {
+    vt->force_split = vt->variances.none.variance > thresholds[0];
+    if (!vt->force_split && vt->variances.none.variance > thresholds[-1] &&
+        !cyclic_refresh_segment_id_boosted(segment_id)) {
+      // We have some nominal amount of 16x16 variance (based on average),
+      // compute the minmax over the 8x8 sub-blocks, and if above threshold,
+      // force split to 8x8 block for this 16x16 block.
+      int minmax =
+          compute_minmax_8x8(vt->src, vt->src_stride, vt->ref, vt->ref_stride,
+#if CONFIG_HIGHBITDEPTH
+                             vt->highbd,
+#endif
+                             vt->width, vt->height);
+      vt->force_split = minmax > cpi->vbp_threshold_minmax;
+    }
+  } else {
+    vt->force_split |=
+        check_split(cpi, vt->split[0], segment_id, thresholds + 1);
+    vt->force_split |=
+        check_split(cpi, vt->split[1], segment_id, thresholds + 1);
+    vt->force_split |=
+        check_split(cpi, vt->split[2], segment_id, thresholds + 1);
+    vt->force_split |=
+        check_split(cpi, vt->split[3], segment_id, thresholds + 1);
+
+    if (vt->bsize == BLOCK_32X32 && !vt->force_split) {
+      vt->force_split = vt->variances.none.variance > thresholds[0];
+    }
+  }
+
+  return vt->force_split;
+}
+
+// This function chooses partitioning based on the variance between source and
+// reconstructed last (or golden), where variance is computed for down-sampled
+// inputs.
+static void choose_partitioning(AV1_COMP *const cpi, ThreadData *const td,
+                                const TileInfo *const tile, MACROBLOCK *const x,
+                                const int mi_row, const int mi_col) {
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  VAR_TREE *const vt = td->var_root[cm->mib_size_log2 - MIN_MIB_SIZE_LOG2];
+#if CONFIG_DUAL_FILTER
+  int i;
+#endif
+  const uint8_t *src;
+  const uint8_t *ref;
+  int src_stride;
+  int ref_stride;
+  int pixels_wide = MI_SIZE * mi_size_wide[cm->sb_size];
+  int pixels_high = MI_SIZE * mi_size_high[cm->sb_size];
+  int64_t thresholds[5] = {
+    cpi->vbp_thresholds[0], cpi->vbp_thresholds[1], cpi->vbp_thresholds[2],
+    cpi->vbp_thresholds[3], cpi->vbp_thresholds[4],
+  };
+  BLOCK_SIZE bsize_min[5] = { BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
+                              cpi->vbp_bsize_min, BLOCK_8X8 };
+  const int start_level = cm->sb_size == BLOCK_64X64 ? 1 : 0;
+  const int64_t *const thre = thresholds + start_level;
+  const BLOCK_SIZE *const bmin = bsize_min + start_level;
+
+  const int is_key_frame = (cm->frame_type == KEY_FRAME);
+  const int low_res = (cm->width <= 352 && cm->height <= 288);
+
+  int segment_id = CR_SEGMENT_ID_BASE;
+
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
+    const uint8_t *const map =
+        cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
+    segment_id = get_segment_id(cm, map, cm->sb_size, mi_row, mi_col);
+
+    if (cyclic_refresh_segment_id_boosted(segment_id)) {
+      int q = av1_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+      set_vbp_thresholds(cpi, thresholds, q);
+    }
+  }
+
+  set_offsets(cpi, tile, x, mi_row, mi_col, cm->sb_size);
+
+  if (xd->mb_to_right_edge < 0) pixels_wide += (xd->mb_to_right_edge >> 3);
+  if (xd->mb_to_bottom_edge < 0) pixels_high += (xd->mb_to_bottom_edge >> 3);
+
+  src = x->plane[0].src.buf;
+  src_stride = x->plane[0].src.stride;
+
+  if (!is_key_frame) {
+    MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+    const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
+    const YV12_BUFFER_CONFIG *yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
+    unsigned int y_sad, y_sad_g;
+
+    const int hbs = cm->mib_size / 2;
+    const int split_vert = mi_col + hbs >= cm->mi_cols;
+    const int split_horz = mi_row + hbs >= cm->mi_rows;
+    BLOCK_SIZE bsize;
+
+    if (split_vert && split_horz)
+      bsize = get_subsize(cm->sb_size, PARTITION_SPLIT);
+    else if (split_vert)
+      bsize = get_subsize(cm->sb_size, PARTITION_VERT);
+    else if (split_horz)
+      bsize = get_subsize(cm->sb_size, PARTITION_HORZ);
+    else
+      bsize = cm->sb_size;
+
+    assert(yv12 != NULL);
+
+    if (yv12_g && yv12_g != yv12) {
+      av1_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
+                           &cm->frame_refs[GOLDEN_FRAME - 1].sf);
+      y_sad_g = cpi->fn_ptr[bsize].sdf(
+          x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
+          xd->plane[0].pre[0].stride);
+    } else {
+      y_sad_g = UINT_MAX;
+    }
+
+    av1_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
+                         &cm->frame_refs[LAST_FRAME - 1].sf);
+    mbmi->ref_frame[0] = LAST_FRAME;
+    mbmi->ref_frame[1] = NONE_FRAME;
+    mbmi->sb_type = cm->sb_size;
+    mbmi->mv[0].as_int = 0;
+#if CONFIG_DUAL_FILTER
+    for (i = 0; i < 4; ++i) mbmi->interp_filter[i] = BILINEAR;
+#else
+    mbmi->interp_filter = BILINEAR;
+#endif
+
+    y_sad = av1_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
+
+    if (y_sad_g < y_sad) {
+      av1_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
+                           &cm->frame_refs[GOLDEN_FRAME - 1].sf);
+      mbmi->ref_frame[0] = GOLDEN_FRAME;
+      mbmi->mv[0].as_int = 0;
+      y_sad = y_sad_g;
+    } else {
+      x->pred_mv[LAST_FRAME] = mbmi->mv[0].as_mv;
+    }
+
+    av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL, cm->sb_size);
+
+    ref = xd->plane[0].dst.buf;
+    ref_stride = xd->plane[0].dst.stride;
+
+    // If the y_sad is very small, take the largest partition and exit.
+    // Don't check on boosted segment for now, as largest is suppressed there.
+    if (segment_id == CR_SEGMENT_ID_BASE && y_sad < cpi->vbp_threshold_sad) {
+      if (!split_vert && !split_horz) {
+        set_block_size(cpi, x, xd, mi_row, mi_col, cm->sb_size);
+        return;
+      }
+    }
+  } else {
+    ref = AV1_VAR_OFFS;
+    ref_stride = 0;
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      switch (xd->bd) {
+        case 10: ref = CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_10); break;
+        case 12: ref = CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_12); break;
+        case 8:
+        default: ref = CONVERT_TO_BYTEPTR(AV1_HIGH_VAR_OFFS_8); break;
+      }
+    }
+#endif  // CONFIG_HIGHBITDEPTH
+  }
+
+  init_variance_tree(
+      vt,
+#if CONFIG_HIGHBITDEPTH
+      xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH,
+#endif  // CONFIG_HIGHBITDEPTH
+      cm->sb_size, (is_key_frame || low_res) ? BLOCK_4X4 : BLOCK_8X8,
+      pixels_wide, pixels_high, src, src_stride, ref, ref_stride);
+
+  // Fill in the entire tree of variances and compute splits.
+  if (is_key_frame) {
+    fill_variance_tree(vt, BLOCK_4X4);
+    check_split_key_frame(vt, thre[1]);
+  } else {
+    fill_variance_tree(vt, BLOCK_8X8);
+    check_split(cpi, vt, segment_id, thre);
+    if (low_res) {
+      refine_variance_tree(vt, thre[1] << 1);
+    }
+  }
+
+  vt->force_split |= mi_col + cm->mib_size > cm->mi_cols ||
+                     mi_row + cm->mib_size > cm->mi_rows;
+
+  // Now go through the entire structure, splitting every block size until
+  // we get to one that's got a variance lower than our threshold.
+  set_vt_partitioning(cpi, x, xd, vt, mi_row, mi_col, thre, bmin);
+}
+
+#if CONFIG_DUAL_FILTER
+static void reset_intmv_filter_type(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+                                    MB_MODE_INFO *mbmi) {
+  int dir;
+  for (dir = 0; dir < 2; ++dir) {
+    if (!has_subpel_mv_component(xd->mi[0], xd, dir) &&
+        (mbmi->ref_frame[1] == NONE_FRAME ||
+         !has_subpel_mv_component(xd->mi[0], xd, dir + 2)))
+      mbmi->interp_filter[dir] = (cm->interp_filter == SWITCHABLE)
+                                     ? EIGHTTAP_REGULAR
+                                     : cm->interp_filter;
+    mbmi->interp_filter[dir + 2] = mbmi->interp_filter[dir];
+  }
+}
+
+static void update_filter_type_count(FRAME_COUNTS *counts,
+                                     const MACROBLOCKD *xd,
+                                     const MB_MODE_INFO *mbmi) {
+  int dir;
+  for (dir = 0; dir < 2; ++dir) {
+    if (has_subpel_mv_component(xd->mi[0], xd, dir) ||
+        (mbmi->ref_frame[1] > INTRA_FRAME &&
+         has_subpel_mv_component(xd->mi[0], xd, dir + 2))) {
+      const int ctx = av1_get_pred_context_switchable_interp(xd, dir);
+      ++counts->switchable_interp[ctx][mbmi->interp_filter[dir]];
+    }
+  }
+}
+#endif
+#if CONFIG_GLOBAL_MOTION
+static void update_global_motion_used(PREDICTION_MODE mode, BLOCK_SIZE bsize,
+                                      const MB_MODE_INFO *mbmi,
+                                      RD_COUNTS *rdc) {
+  if (mode == ZEROMV
+#if CONFIG_EXT_INTER
+      || mode == ZERO_ZEROMV
+#endif
+      ) {
+    const int num_4x4s =
+        num_4x4_blocks_wide_lookup[bsize] * num_4x4_blocks_high_lookup[bsize];
+    int ref;
+    for (ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) {
+      rdc->global_motion_used[mbmi->ref_frame[ref]] += num_4x4s;
+    }
+  }
+}
+#endif  // CONFIG_GLOBAL_MOTION
+
+static void reset_tx_size(MACROBLOCKD *xd, MB_MODE_INFO *mbmi,
+                          const TX_MODE tx_mode) {
+  if (xd->lossless[mbmi->segment_id]) {
+    mbmi->tx_size = TX_4X4;
+  } else if (tx_mode != TX_MODE_SELECT) {
+    mbmi->tx_size =
+        tx_size_from_tx_mode(mbmi->sb_type, tx_mode, is_inter_block(mbmi));
+  }
+}
+
+#if CONFIG_REF_MV
+static void set_ref_and_pred_mvs(MACROBLOCK *const x, int_mv *const mi_pred_mv,
+                                 int8_t rf_type) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+
+  const int bw = xd->n8_w << MI_SIZE_LOG2;
+  const int bh = xd->n8_h << MI_SIZE_LOG2;
+  int ref_mv_idx = mbmi->ref_mv_idx;
+  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  CANDIDATE_MV *const curr_ref_mv_stack = mbmi_ext->ref_mv_stack[rf_type];
+
+#if CONFIG_EXT_INTER
+  if (has_second_ref(mbmi)) {
+    // Special case: NEAR_NEWMV and NEW_NEARMV modes use 1 + mbmi->ref_mv_idx
+    // (like NEARMV) instead
+    if (mbmi->mode == NEAR_NEWMV || mbmi->mode == NEW_NEARMV) ref_mv_idx += 1;
+
+    if (compound_ref0_mode(mbmi->mode) == NEWMV) {
+      int_mv this_mv = curr_ref_mv_stack[ref_mv_idx].this_mv;
+      clamp_mv_ref(&this_mv.as_mv, bw, bh, xd);
+      mbmi_ext->ref_mvs[mbmi->ref_frame[0]][0] = this_mv;
+      mbmi->pred_mv[0] = this_mv;
+      mi_pred_mv[0] = this_mv;
+    }
+    if (compound_ref1_mode(mbmi->mode) == NEWMV) {
+      int_mv this_mv = curr_ref_mv_stack[ref_mv_idx].comp_mv;
+      clamp_mv_ref(&this_mv.as_mv, bw, bh, xd);
+      mbmi_ext->ref_mvs[mbmi->ref_frame[1]][0] = this_mv;
+      mbmi->pred_mv[1] = this_mv;
+      mi_pred_mv[1] = this_mv;
+    }
+  } else {
+#endif  // CONFIG_EXT_INTER
+    if (mbmi->mode == NEWMV) {
+      int i;
+      for (i = 0; i < 1 + has_second_ref(mbmi); ++i) {
+        int_mv this_mv = (i == 0) ? curr_ref_mv_stack[ref_mv_idx].this_mv
+                                  : curr_ref_mv_stack[ref_mv_idx].comp_mv;
+        clamp_mv_ref(&this_mv.as_mv, bw, bh, xd);
+        mbmi_ext->ref_mvs[mbmi->ref_frame[i]][0] = this_mv;
+        mbmi->pred_mv[i] = this_mv;
+        mi_pred_mv[i] = this_mv;
+      }
+    }
+#if CONFIG_EXT_INTER
+  }
+#endif  // CONFIG_EXT_INTER
+}
+#endif  // CONFIG_REF_MV
+
+static void update_state(const AV1_COMP *const cpi, ThreadData *td,
+                         PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col,
+                         BLOCK_SIZE bsize, RUN_TYPE dry_run) {
+  int i, x_idx, y;
+  const AV1_COMMON *const cm = &cpi->common;
+  RD_COUNTS *const rdc = &td->rd_counts;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *const p = x->plane;
+  struct macroblockd_plane *const pd = xd->plane;
+  MODE_INFO *mi = &ctx->mic;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  MODE_INFO *mi_addr = xd->mi[0];
+  const struct segmentation *const seg = &cm->seg;
+  const int bw = mi_size_wide[mi->mbmi.sb_type];
+  const int bh = mi_size_high[mi->mbmi.sb_type];
+  const int x_mis = AOMMIN(bw, cm->mi_cols - mi_col);
+  const int y_mis = AOMMIN(bh, cm->mi_rows - mi_row);
+  MV_REF *const frame_mvs = cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
+  int w, h;
+
+  const int mis = cm->mi_stride;
+  const int mi_width = mi_size_wide[bsize];
+  const int mi_height = mi_size_high[bsize];
+  const int unify_bsize = CONFIG_CB4X4;
+
+#if CONFIG_REF_MV
+  int8_t rf_type;
+#endif
+
+#if !CONFIG_SUPERTX
+  assert(mi->mbmi.sb_type == bsize);
+#endif
+
+  *mi_addr = *mi;
+  *x->mbmi_ext = ctx->mbmi_ext;
+
+#if CONFIG_DUAL_FILTER
+  reset_intmv_filter_type(cm, xd, mbmi);
+#endif
+
+#if CONFIG_REF_MV
+  rf_type = av1_ref_frame_type(mbmi->ref_frame);
+  if (x->mbmi_ext->ref_mv_count[rf_type] > 1 &&
+      (mbmi->sb_type >= BLOCK_8X8 || unify_bsize)) {
+    set_ref_and_pred_mvs(x, mi->mbmi.pred_mv, rf_type);
+  }
+#endif  // CONFIG_REF_MV
+
+  // If segmentation in use
+  if (seg->enabled) {
+    // For in frame complexity AQ copy the segment id from the segment map.
+    if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
+      const uint8_t *const map =
+          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
+      mi_addr->mbmi.segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
+      reset_tx_size(xd, &mi_addr->mbmi, cm->tx_mode);
+    }
+    // Else for cyclic refresh mode update the segment map, set the segment id
+    // and then update the quantizer.
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
+      av1_cyclic_refresh_update_segment(cpi, &xd->mi[0]->mbmi, mi_row, mi_col,
+                                        bsize, ctx->rate, ctx->dist, x->skip);
+      reset_tx_size(xd, &mi_addr->mbmi, cm->tx_mode);
+    }
+  }
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    p[i].coeff = ctx->coeff[i];
+    p[i].qcoeff = ctx->qcoeff[i];
+    pd[i].dqcoeff = ctx->dqcoeff[i];
+#if CONFIG_PVQ
+    pd[i].pvq_ref_coeff = ctx->pvq_ref_coeff[i];
+#endif
+    p[i].eobs = ctx->eobs[i];
+#if CONFIG_LV_MAP
+    p[i].txb_entropy_ctx = ctx->txb_entropy_ctx[i];
+#endif  // CONFIG_LV_MAP
+  }
+#if CONFIG_PALETTE
+  for (i = 0; i < 2; ++i) pd[i].color_index_map = ctx->color_index_map[i];
+#endif  // CONFIG_PALETTE
+
+  // Restore the coding context of the MB to that that was in place
+  // when the mode was picked for it
+  for (y = 0; y < mi_height; y++)
+    for (x_idx = 0; x_idx < mi_width; x_idx++)
+      if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx &&
+          (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
+        xd->mi[x_idx + y * mis] = mi_addr;
+      }
+
+#if CONFIG_DELTA_Q && !CONFIG_EXT_DELTA_Q
+  if (cpi->oxcf.aq_mode > NO_AQ && cpi->oxcf.aq_mode < DELTA_AQ)
+    av1_init_plane_quantizers(cpi, x, xd->mi[0]->mbmi.segment_id);
+#else
+  if (cpi->oxcf.aq_mode)
+    av1_init_plane_quantizers(cpi, x, xd->mi[0]->mbmi.segment_id);
+#endif
+
+  if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8 && !unify_bsize) {
+    mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
+    mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
+  }
+
+  x->skip = ctx->skip;
+
+#if CONFIG_VAR_TX
+  for (i = 0; i < 1; ++i)
+    memcpy(x->blk_skip[i], ctx->blk_skip[i],
+           sizeof(uint8_t) * ctx->num_4x4_blk);
+#endif
+
+  if (dry_run) return;
+
+#if CONFIG_INTERNAL_STATS
+  {
+    unsigned int *const mode_chosen_counts =
+        (unsigned int *)cpi->mode_chosen_counts;  // Cast const away.
+    if (frame_is_intra_only(cm)) {
+      static const int kf_mode_index[] = {
+        THR_DC /*DC_PRED*/,
+        THR_V_PRED /*V_PRED*/,
+        THR_H_PRED /*H_PRED*/,
+        THR_D45_PRED /*D45_PRED*/,
+        THR_D135_PRED /*D135_PRED*/,
+        THR_D117_PRED /*D117_PRED*/,
+        THR_D153_PRED /*D153_PRED*/,
+        THR_D207_PRED /*D207_PRED*/,
+        THR_D63_PRED /*D63_PRED*/,
+#if CONFIG_ALT_INTRA
+        THR_SMOOTH, /*SMOOTH_PRED*/
+#endif              // CONFIG_ALT_INTRA
+        THR_TM /*TM_PRED*/,
+      };
+      ++mode_chosen_counts[kf_mode_index[mbmi->mode]];
+    } else {
+      // Note how often each mode chosen as best
+      ++mode_chosen_counts[ctx->best_mode_index];
+    }
+  }
+#endif
+  if (!frame_is_intra_only(cm)) {
+    if (is_inter_block(mbmi)) {
+      av1_update_mv_count(td);
+#if CONFIG_GLOBAL_MOTION
+      if (bsize >= BLOCK_8X8) {
+        // TODO(sarahparker): global motion stats need to be handled per-tile
+        // to be compatible with tile-based threading.
+        update_global_motion_used(mbmi->mode, bsize, mbmi, rdc);
+      } else {
+        const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+        const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+        int idx, idy;
+        for (idy = 0; idy < 2; idy += num_4x4_h) {
+          for (idx = 0; idx < 2; idx += num_4x4_w) {
+            const int j = idy * 2 + idx;
+            update_global_motion_used(mi->bmi[j].as_mode, bsize, mbmi, rdc);
+          }
+        }
+      }
+#endif  // CONFIG_GLOBAL_MOTION
+      if (cm->interp_filter == SWITCHABLE
+#if CONFIG_WARPED_MOTION
+          && mbmi->motion_mode != WARPED_CAUSAL
+#endif  // CONFIG_WARPED_MOTION
+#if CONFIG_GLOBAL_MOTION
+          && !is_nontrans_global_motion(xd)
+#endif  // CONFIG_GLOBAL_MOTION
+              ) {
+#if CONFIG_DUAL_FILTER
+        update_filter_type_count(td->counts, xd, mbmi);
+#else
+        const int switchable_ctx = av1_get_pred_context_switchable_interp(xd);
+        ++td->counts->switchable_interp[switchable_ctx][mbmi->interp_filter];
+#endif
+      }
+    }
+
+    rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
+    rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
+    rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
+  }
+
+  for (h = 0; h < y_mis; ++h) {
+    MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
+    for (w = 0; w < x_mis; ++w) {
+      MV_REF *const mv = frame_mv + w;
+      mv->ref_frame[0] = mi->mbmi.ref_frame[0];
+      mv->ref_frame[1] = mi->mbmi.ref_frame[1];
+      mv->mv[0].as_int = mi->mbmi.mv[0].as_int;
+      mv->mv[1].as_int = mi->mbmi.mv[1].as_int;
+    }
+  }
+}
+
+#if CONFIG_SUPERTX
+static void update_state_supertx(const AV1_COMP *const cpi, ThreadData *td,
+                                 PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col,
+                                 BLOCK_SIZE bsize, RUN_TYPE dry_run) {
+  int y, x_idx;
+#if CONFIG_VAR_TX
+  int i;
+#endif
+  const AV1_COMMON *const cm = &cpi->common;
+  RD_COUNTS *const rdc = &td->rd_counts;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *mi = &ctx->mic;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  MODE_INFO *mi_addr = xd->mi[0];
+  const struct segmentation *const seg = &cm->seg;
+  const int mis = cm->mi_stride;
+  const int mi_width = mi_size_wide[bsize];
+  const int mi_height = mi_size_high[bsize];
+  const int x_mis = AOMMIN(mi_width, cm->mi_cols - mi_col);
+  const int y_mis = AOMMIN(mi_height, cm->mi_rows - mi_row);
+  const int unify_bsize = CONFIG_CB4X4;
+  MV_REF *const frame_mvs = cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
+  int w, h;
+
+#if CONFIG_REF_MV
+  int8_t rf_type;
+#endif
+
+  *mi_addr = *mi;
+  *x->mbmi_ext = ctx->mbmi_ext;
+  assert(is_inter_block(mbmi));
+  assert(mbmi->tx_size == ctx->mic.mbmi.tx_size);
+
+#if CONFIG_DUAL_FILTER
+  reset_intmv_filter_type(cm, xd, mbmi);
+#endif
+
+#if CONFIG_REF_MV
+  rf_type = av1_ref_frame_type(mbmi->ref_frame);
+  if (x->mbmi_ext->ref_mv_count[rf_type] > 1 &&
+      (mbmi->sb_type >= BLOCK_8X8 || unify_bsize)) {
+    set_ref_and_pred_mvs(x, mi->mbmi.pred_mv, rf_type);
+  }
+#endif  // CONFIG_REF_MV
+
+  // If segmentation in use
+  if (seg->enabled) {
+    if (cpi->vaq_refresh) {
+      const int energy =
+          bsize <= BLOCK_16X16 ? x->mb_energy : av1_block_energy(cpi, x, bsize);
+      mi_addr->mbmi.segment_id = av1_vaq_segment_id(energy);
+    } else if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
+      // For cyclic refresh mode, now update the segment map
+      // and set the segment id.
+      av1_cyclic_refresh_update_segment(cpi, &xd->mi[0]->mbmi, mi_row, mi_col,
+                                        bsize, ctx->rate, ctx->dist, 1);
+    } else {
+      // Otherwise just set the segment id based on the current segment map
+      const uint8_t *const map =
+          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
+      mi_addr->mbmi.segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
+    }
+    mi_addr->mbmi.segment_id_supertx = MAX_SEGMENTS;
+  }
+
+  // Restore the coding context of the MB to that that was in place
+  // when the mode was picked for it
+  for (y = 0; y < mi_height; y++)
+    for (x_idx = 0; x_idx < mi_width; x_idx++)
+      if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx &&
+          (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
+        xd->mi[x_idx + y * mis] = mi_addr;
+      }
+
+#if !CONFIG_CB4X4
+  if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8) {
+    mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
+    mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
+  }
+#endif
+
+  x->skip = ctx->skip;
+
+#if CONFIG_VAR_TX
+  for (i = 0; i < 1; ++i)
+    memcpy(x->blk_skip[i], ctx->blk_skip[i],
+           sizeof(uint8_t) * ctx->num_4x4_blk);
+
+  if (!is_inter_block(mbmi) || mbmi->skip)
+    mbmi->min_tx_size = get_min_tx_size(mbmi->tx_size);
+#endif  // CONFIG_VAR_TX
+
+#if CONFIG_VAR_TX
+  {
+    const TX_SIZE mtx = mbmi->tx_size;
+    const int num_4x4_blocks_wide = tx_size_wide_unit[mtx] >> 1;
+    const int num_4x4_blocks_high = tx_size_high_unit[mtx] >> 1;
+    int idy, idx;
+    mbmi->inter_tx_size[0][0] = mtx;
+    for (idy = 0; idy < num_4x4_blocks_high; ++idy)
+      for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
+        mbmi->inter_tx_size[idy][idx] = mtx;
+  }
+#endif  // CONFIG_VAR_TX
+  // Turn motion variation off for supertx
+  mbmi->motion_mode = SIMPLE_TRANSLATION;
+
+  if (dry_run) return;
+
+  if (!frame_is_intra_only(cm)) {
+    av1_update_mv_count(td);
+
+#if CONFIG_GLOBAL_MOTION
+    if (is_inter_block(mbmi)) {
+      if (bsize >= BLOCK_8X8) {
+        // TODO(sarahparker): global motion stats need to be handled per-tile
+        // to be compatible with tile-based threading.
+        update_global_motion_used(mbmi->mode, bsize, mbmi, rdc);
+      } else {
+        const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+        const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+        int idx, idy;
+        for (idy = 0; idy < 2; idy += num_4x4_h) {
+          for (idx = 0; idx < 2; idx += num_4x4_w) {
+            const int j = idy * 2 + idx;
+            update_global_motion_used(mi->bmi[j].as_mode, bsize, mbmi, rdc);
+          }
+        }
+      }
+    }
+#endif  // CONFIG_GLOBAL_MOTION
+
+    if (cm->interp_filter == SWITCHABLE
+#if CONFIG_GLOBAL_MOTION
+        && !is_nontrans_global_motion(xd)
+#endif  // CONFIG_GLOBAL_MOTION
+            ) {
+#if CONFIG_DUAL_FILTER
+      update_filter_type_count(td->counts, xd, mbmi);
+#else
+      const int pred_ctx = av1_get_pred_context_switchable_interp(xd);
+      ++td->counts->switchable_interp[pred_ctx][mbmi->interp_filter];
+#endif
+    }
+
+    rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
+    rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
+    rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
+  }
+
+  for (h = 0; h < y_mis; ++h) {
+    MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
+    for (w = 0; w < x_mis; ++w) {
+      MV_REF *const mv = frame_mv + w;
+      mv->ref_frame[0] = mi->mbmi.ref_frame[0];
+      mv->ref_frame[1] = mi->mbmi.ref_frame[1];
+      mv->mv[0].as_int = mi->mbmi.mv[0].as_int;
+      mv->mv[1].as_int = mi->mbmi.mv[1].as_int;
+    }
+  }
+}
+
+static void update_state_sb_supertx(const AV1_COMP *const cpi, ThreadData *td,
+                                    const TileInfo *const tile, int mi_row,
+                                    int mi_col, BLOCK_SIZE bsize,
+                                    RUN_TYPE dry_run, PC_TREE *pc_tree) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *const p = x->plane;
+  struct macroblockd_plane *const pd = xd->plane;
+  int hbs = mi_size_wide[bsize] / 2;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+  PARTITION_TYPE partition = pc_tree->partitioning;
+  BLOCK_SIZE subsize = get_subsize(bsize, partition);
+  int i;
+#if CONFIG_EXT_PARTITION_TYPES
+  BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+  PICK_MODE_CONTEXT *pmc = NULL;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  if (bsize == BLOCK_16X16 && cpi->vaq_refresh)
+    x->mb_energy = av1_block_energy(cpi, x, bsize);
+
+  switch (partition) {
+    case PARTITION_NONE:
+      set_offsets_supertx(cpi, td, tile, mi_row, mi_col, subsize);
+      update_state_supertx(cpi, td, &pc_tree->none, mi_row, mi_col, subsize,
+                           dry_run);
+      break;
+    case PARTITION_VERT:
+      set_offsets_supertx(cpi, td, tile, mi_row, mi_col, subsize);
+      update_state_supertx(cpi, td, &pc_tree->vertical[0], mi_row, mi_col,
+                           subsize, dry_run);
+      if (mi_col + hbs < cm->mi_cols && (bsize > BLOCK_8X8 || unify_bsize)) {
+        set_offsets_supertx(cpi, td, tile, mi_row, mi_col + hbs, subsize);
+        update_state_supertx(cpi, td, &pc_tree->vertical[1], mi_row,
+                             mi_col + hbs, subsize, dry_run);
+      }
+      pmc = &pc_tree->vertical_supertx;
+      break;
+    case PARTITION_HORZ:
+      set_offsets_supertx(cpi, td, tile, mi_row, mi_col, subsize);
+      update_state_supertx(cpi, td, &pc_tree->horizontal[0], mi_row, mi_col,
+                           subsize, dry_run);
+      if (mi_row + hbs < cm->mi_rows && (bsize > BLOCK_8X8 || unify_bsize)) {
+        set_offsets_supertx(cpi, td, tile, mi_row + hbs, mi_col, subsize);
+        update_state_supertx(cpi, td, &pc_tree->horizontal[1], mi_row + hbs,
+                             mi_col, subsize, dry_run);
+      }
+      pmc = &pc_tree->horizontal_supertx;
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        set_offsets_supertx(cpi, td, tile, mi_row, mi_col, subsize);
+        update_state_supertx(cpi, td, pc_tree->leaf_split[0], mi_row, mi_col,
+                             subsize, dry_run);
+      } else {
+        set_offsets_supertx(cpi, td, tile, mi_row, mi_col, subsize);
+        update_state_sb_supertx(cpi, td, tile, mi_row, mi_col, subsize, dry_run,
+                                pc_tree->split[0]);
+        set_offsets_supertx(cpi, td, tile, mi_row, mi_col + hbs, subsize);
+        update_state_sb_supertx(cpi, td, tile, mi_row, mi_col + hbs, subsize,
+                                dry_run, pc_tree->split[1]);
+        set_offsets_supertx(cpi, td, tile, mi_row + hbs, mi_col, subsize);
+        update_state_sb_supertx(cpi, td, tile, mi_row + hbs, mi_col, subsize,
+                                dry_run, pc_tree->split[2]);
+        set_offsets_supertx(cpi, td, tile, mi_row + hbs, mi_col + hbs, subsize);
+        update_state_sb_supertx(cpi, td, tile, mi_row + hbs, mi_col + hbs,
+                                subsize, dry_run, pc_tree->split[3]);
+      }
+      pmc = &pc_tree->split_supertx;
+      break;
+#if CONFIG_EXT_PARTITION_TYPES
+    case PARTITION_HORZ_A:
+      set_offsets_supertx(cpi, td, tile, mi_row, mi_col, bsize2);
+      update_state_supertx(cpi, td, &pc_tree->horizontala[0], mi_row, mi_col,
+                           bsize2, dry_run);
+      set_offsets_supertx(cpi, td, tile, mi_row, mi_col + hbs, bsize2);
+      update_state_supertx(cpi, td, &pc_tree->horizontala[1], mi_row,
+                           mi_col + hbs, bsize2, dry_run);
+      set_offsets_supertx(cpi, td, tile, mi_row + hbs, mi_col, subsize);
+      update_state_supertx(cpi, td, &pc_tree->horizontala[2], mi_row + hbs,
+                           mi_col, subsize, dry_run);
+      pmc = &pc_tree->horizontala_supertx;
+      break;
+    case PARTITION_HORZ_B:
+      set_offsets_supertx(cpi, td, tile, mi_row, mi_col, subsize);
+      update_state_supertx(cpi, td, &pc_tree->horizontalb[0], mi_row, mi_col,
+                           subsize, dry_run);
+      set_offsets_supertx(cpi, td, tile, mi_row + hbs, mi_col, bsize2);
+      update_state_supertx(cpi, td, &pc_tree->horizontalb[1], mi_row + hbs,
+                           mi_col, bsize2, dry_run);
+      set_offsets_supertx(cpi, td, tile, mi_row + hbs, mi_col + hbs, bsize2);
+      update_state_supertx(cpi, td, &pc_tree->horizontalb[2], mi_row + hbs,
+                           mi_col + hbs, bsize2, dry_run);
+      pmc = &pc_tree->horizontalb_supertx;
+      break;
+    case PARTITION_VERT_A:
+      set_offsets_supertx(cpi, td, tile, mi_row, mi_col, bsize2);
+      update_state_supertx(cpi, td, &pc_tree->verticala[0], mi_row, mi_col,
+                           bsize2, dry_run);
+      set_offsets_supertx(cpi, td, tile, mi_row + hbs, mi_col, bsize2);
+      update_state_supertx(cpi, td, &pc_tree->verticala[1], mi_row + hbs,
+                           mi_col, bsize2, dry_run);
+      set_offsets_supertx(cpi, td, tile, mi_row, mi_col + hbs, subsize);
+      update_state_supertx(cpi, td, &pc_tree->verticala[2], mi_row,
+                           mi_col + hbs, subsize, dry_run);
+      pmc = &pc_tree->verticala_supertx;
+      break;
+    case PARTITION_VERT_B:
+      set_offsets_supertx(cpi, td, tile, mi_row, mi_col, subsize);
+      update_state_supertx(cpi, td, &pc_tree->verticalb[0], mi_row, mi_col,
+                           subsize, dry_run);
+      set_offsets_supertx(cpi, td, tile, mi_row, mi_col + hbs, bsize2);
+      update_state_supertx(cpi, td, &pc_tree->verticalb[1], mi_row,
+                           mi_col + hbs, bsize2, dry_run);
+      set_offsets_supertx(cpi, td, tile, mi_row + hbs, mi_col + hbs, bsize2);
+      update_state_supertx(cpi, td, &pc_tree->verticalb[2], mi_row + hbs,
+                           mi_col + hbs, bsize2, dry_run);
+      pmc = &pc_tree->verticalb_supertx;
+      break;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+    default: assert(0);
+  }
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    if (pmc != NULL) {
+      p[i].coeff = pmc->coeff[i];
+      p[i].qcoeff = pmc->qcoeff[i];
+      pd[i].dqcoeff = pmc->dqcoeff[i];
+      p[i].eobs = pmc->eobs[i];
+    } else {
+      // These should never be used
+      p[i].coeff = NULL;
+      p[i].qcoeff = NULL;
+      pd[i].dqcoeff = NULL;
+      p[i].eobs = NULL;
+    }
+  }
+}
+
+static void update_supertx_param(ThreadData *td, PICK_MODE_CONTEXT *ctx,
+                                 int best_tx, TX_SIZE supertx_size) {
+  MACROBLOCK *const x = &td->mb;
+#if CONFIG_VAR_TX
+  int i;
+
+  for (i = 0; i < 1; ++i)
+    memcpy(ctx->blk_skip[i], x->blk_skip[i],
+           sizeof(uint8_t) * ctx->num_4x4_blk);
+  ctx->mic.mbmi.min_tx_size = get_min_tx_size(supertx_size);
+#endif  // CONFIG_VAR_TX
+  ctx->mic.mbmi.tx_size = supertx_size;
+  ctx->skip = x->skip;
+  ctx->mic.mbmi.tx_type = best_tx;
+}
+
+static void update_supertx_param_sb(const AV1_COMP *const cpi, ThreadData *td,
+                                    int mi_row, int mi_col, BLOCK_SIZE bsize,
+                                    int best_tx, TX_SIZE supertx_size,
+                                    PC_TREE *pc_tree) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int hbs = mi_size_wide[bsize] / 2;
+  PARTITION_TYPE partition = pc_tree->partitioning;
+  BLOCK_SIZE subsize = get_subsize(bsize, partition);
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+  int i;
+#endif
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  switch (partition) {
+    case PARTITION_NONE:
+      update_supertx_param(td, &pc_tree->none, best_tx, supertx_size);
+      break;
+    case PARTITION_VERT:
+      update_supertx_param(td, &pc_tree->vertical[0], best_tx, supertx_size);
+      if (mi_col + hbs < cm->mi_cols && (bsize > BLOCK_8X8 || unify_bsize))
+        update_supertx_param(td, &pc_tree->vertical[1], best_tx, supertx_size);
+      break;
+    case PARTITION_HORZ:
+      update_supertx_param(td, &pc_tree->horizontal[0], best_tx, supertx_size);
+      if (mi_row + hbs < cm->mi_rows && (bsize > BLOCK_8X8 || unify_bsize))
+        update_supertx_param(td, &pc_tree->horizontal[1], best_tx,
+                             supertx_size);
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        update_supertx_param(td, pc_tree->leaf_split[0], best_tx, supertx_size);
+      } else {
+        update_supertx_param_sb(cpi, td, mi_row, mi_col, subsize, best_tx,
+                                supertx_size, pc_tree->split[0]);
+        update_supertx_param_sb(cpi, td, mi_row, mi_col + hbs, subsize, best_tx,
+                                supertx_size, pc_tree->split[1]);
+        update_supertx_param_sb(cpi, td, mi_row + hbs, mi_col, subsize, best_tx,
+                                supertx_size, pc_tree->split[2]);
+        update_supertx_param_sb(cpi, td, mi_row + hbs, mi_col + hbs, subsize,
+                                best_tx, supertx_size, pc_tree->split[3]);
+      }
+      break;
+#if CONFIG_EXT_PARTITION_TYPES
+    case PARTITION_HORZ_A:
+      for (i = 0; i < 3; i++)
+        update_supertx_param(td, &pc_tree->horizontala[i], best_tx,
+                             supertx_size);
+      break;
+    case PARTITION_HORZ_B:
+      for (i = 0; i < 3; i++)
+        update_supertx_param(td, &pc_tree->horizontalb[i], best_tx,
+                             supertx_size);
+      break;
+    case PARTITION_VERT_A:
+      for (i = 0; i < 3; i++)
+        update_supertx_param(td, &pc_tree->verticala[i], best_tx, supertx_size);
+      break;
+    case PARTITION_VERT_B:
+      for (i = 0; i < 3; i++)
+        update_supertx_param(td, &pc_tree->verticalb[i], best_tx, supertx_size);
+      break;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+    default: assert(0);
+  }
+}
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_MOTION_VAR && CONFIG_NCOBMC
+static void set_mode_info_b(const AV1_COMP *const cpi,
+                            const TileInfo *const tile, ThreadData *td,
+                            int mi_row, int mi_col, BLOCK_SIZE bsize,
+                            PICK_MODE_CONTEXT *ctx) {
+  MACROBLOCK *const x = &td->mb;
+  set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
+  update_state(cpi, td, ctx, mi_row, mi_col, bsize, 1);
+}
+
+static void set_mode_info_sb(const AV1_COMP *const cpi, ThreadData *td,
+                             const TileInfo *const tile, TOKENEXTRA **tp,
+                             int mi_row, int mi_col, BLOCK_SIZE bsize,
+                             PC_TREE *pc_tree) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int hbs = mi_size_wide[bsize] / 2;
+  const PARTITION_TYPE partition = pc_tree->partitioning;
+  BLOCK_SIZE subsize = get_subsize(bsize, partition);
+#if CONFIG_EXT_PARTITION_TYPES
+  const BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+  assert(bsize >= BLOCK_8X8);
+#endif
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  switch (partition) {
+    case PARTITION_NONE:
+      set_mode_info_b(cpi, tile, td, mi_row, mi_col, subsize, &pc_tree->none);
+      break;
+    case PARTITION_VERT:
+      set_mode_info_b(cpi, tile, td, mi_row, mi_col, subsize,
+                      &pc_tree->vertical[0]);
+      if (mi_col + hbs < cm->mi_cols && (bsize > BLOCK_8X8 || unify_bsize)) {
+        set_mode_info_b(cpi, tile, td, mi_row, mi_col + hbs, subsize,
+                        &pc_tree->vertical[1]);
+      }
+      break;
+    case PARTITION_HORZ:
+      set_mode_info_b(cpi, tile, td, mi_row, mi_col, subsize,
+                      &pc_tree->horizontal[0]);
+      if (mi_row + hbs < cm->mi_rows && (bsize > BLOCK_8X8 || unify_bsize)) {
+        set_mode_info_b(cpi, tile, td, mi_row + hbs, mi_col, subsize,
+                        &pc_tree->horizontal[1]);
+      }
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        set_mode_info_b(cpi, tile, td, mi_row, mi_col, subsize,
+                        pc_tree->leaf_split[0]);
+      } else {
+        set_mode_info_sb(cpi, td, tile, tp, mi_row, mi_col, subsize,
+                         pc_tree->split[0]);
+        set_mode_info_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, subsize,
+                         pc_tree->split[1]);
+        set_mode_info_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, subsize,
+                         pc_tree->split[2]);
+        set_mode_info_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, subsize,
+                         pc_tree->split[3]);
+      }
+      break;
+#if CONFIG_EXT_PARTITION_TYPES
+    case PARTITION_HORZ_A:
+      set_mode_info_b(cpi, tile, td, mi_row, mi_col, bsize2,
+                      &pc_tree->horizontala[0]);
+      set_mode_info_b(cpi, tile, td, mi_row, mi_col + hbs, bsize2,
+                      &pc_tree->horizontala[1]);
+      set_mode_info_b(cpi, tile, td, mi_row + hbs, mi_col, subsize,
+                      &pc_tree->horizontala[2]);
+      break;
+    case PARTITION_HORZ_B:
+      set_mode_info_b(cpi, tile, td, mi_row, mi_col, subsize,
+                      &pc_tree->horizontalb[0]);
+      set_mode_info_b(cpi, tile, td, mi_row + hbs, mi_col, bsize2,
+                      &pc_tree->horizontalb[1]);
+      set_mode_info_b(cpi, tile, td, mi_row + hbs, mi_col + hbs, bsize2,
+                      &pc_tree->horizontalb[2]);
+      break;
+    case PARTITION_VERT_A:
+      set_mode_info_b(cpi, tile, td, mi_row, mi_col, bsize2,
+                      &pc_tree->verticala[0]);
+      set_mode_info_b(cpi, tile, td, mi_row + hbs, mi_col, bsize2,
+                      &pc_tree->verticala[1]);
+      set_mode_info_b(cpi, tile, td, mi_row, mi_col + hbs, subsize,
+                      &pc_tree->verticala[2]);
+      break;
+    case PARTITION_VERT_B:
+      set_mode_info_b(cpi, tile, td, mi_row, mi_col, subsize,
+                      &pc_tree->verticalb[0]);
+      set_mode_info_b(cpi, tile, td, mi_row, mi_col + hbs, bsize2,
+                      &pc_tree->verticalb[1]);
+      set_mode_info_b(cpi, tile, td, mi_row + hbs, mi_col + hbs, bsize2,
+                      &pc_tree->verticalb[2]);
+      break;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+    default: assert(0 && "Invalid partition type."); break;
+  }
+}
+#endif
+
+void av1_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
+                          int mi_row, int mi_col) {
+  uint8_t *const buffers[3] = { src->y_buffer, src->u_buffer, src->v_buffer };
+  const int widths[3] = { src->y_crop_width, src->uv_crop_width,
+                          src->uv_crop_width };
+  const int heights[3] = { src->y_crop_height, src->uv_crop_height,
+                           src->uv_crop_height };
+  const int strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
+  int i;
+
+  // Set current frame pointer.
+  x->e_mbd.cur_buf = src;
+
+  for (i = 0; i < MAX_MB_PLANE; i++)
+    setup_pred_plane(&x->plane[i].src, x->e_mbd.mi[0]->mbmi.sb_type, buffers[i],
+                     widths[i], heights[i], strides[i], mi_row, mi_col, NULL,
+                     x->e_mbd.plane[i].subsampling_x,
+                     x->e_mbd.plane[i].subsampling_y);
+}
+
+static int set_segment_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x,
+                              int8_t segment_id) {
+  int segment_qindex;
+  const AV1_COMMON *const cm = &cpi->common;
+  av1_init_plane_quantizers(cpi, x, segment_id);
+  aom_clear_system_state();
+  segment_qindex = av1_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+  return av1_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
+}
+
+static void rd_pick_sb_modes(const AV1_COMP *const cpi, TileDataEnc *tile_data,
+                             MACROBLOCK *const x, int mi_row, int mi_col,
+                             RD_STATS *rd_cost,
+#if CONFIG_SUPERTX
+                             int *totalrate_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                             PARTITION_TYPE partition,
+#endif
+                             BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                             int64_t best_rd) {
+  const AV1_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi;
+  struct macroblock_plane *const p = x->plane;
+  struct macroblockd_plane *const pd = xd->plane;
+  const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
+  int i, orig_rdmult;
+  const int unify_bsize = CONFIG_CB4X4;
+
+  aom_clear_system_state();
+
+#if CONFIG_PVQ
+  x->pvq_speed = 1;
+  x->pvq_coded = 0;
+#endif
+#if CONFIG_CFL
+  // Don't store luma during RDO (we will store the best mode later).
+  x->cfl_store_y = 0;
+#endif
+
+  set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+  mbmi = &xd->mi[0]->mbmi;
+  mbmi->sb_type = bsize;
+#if CONFIG_RD_DEBUG
+  mbmi->mi_row = mi_row;
+  mbmi->mi_col = mi_col;
+#endif
+#if CONFIG_SUPERTX
+  // We set tx_size here as skip blocks would otherwise not set it.
+  // tx_size needs to be set at this point as supertx_enable in
+  // write_modes_sb is computed based on this, and if the garbage in memory
+  // just happens to be the supertx_size, then the packer will code this
+  // block as a supertx block, even if rdopt did not pick it as such.
+  mbmi->tx_size = max_txsize_lookup[bsize];
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+  mbmi->partition = partition;
+#endif
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    p[i].coeff = ctx->coeff[i];
+    p[i].qcoeff = ctx->qcoeff[i];
+    pd[i].dqcoeff = ctx->dqcoeff[i];
+#if CONFIG_PVQ
+    pd[i].pvq_ref_coeff = ctx->pvq_ref_coeff[i];
+#endif
+    p[i].eobs = ctx->eobs[i];
+#if CONFIG_LV_MAP
+    p[i].txb_entropy_ctx = ctx->txb_entropy_ctx[i];
+#endif
+  }
+
+#if CONFIG_PALETTE
+  for (i = 0; i < 2; ++i) pd[i].color_index_map = ctx->color_index_map[i];
+#endif  // CONFIG_PALETTE
+
+  ctx->skippable = 0;
+  ctx->pred_pixel_ready = 0;
+
+  // Set to zero to make sure we do not use the previous encoded frame stats
+  mbmi->skip = 0;
+
+#if CONFIG_CB4X4
+  x->skip_chroma_rd =
+      !is_chroma_reference(mi_row, mi_col, bsize, xd->plane[1].subsampling_x,
+                           xd->plane[1].subsampling_y);
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    x->source_variance = av1_high_get_sby_perpixel_variance(
+        cpi, &x->plane[0].src, bsize, xd->bd);
+  } else {
+    x->source_variance =
+        av1_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
+  }
+#else
+  x->source_variance =
+      av1_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
+#endif  // CONFIG_HIGHBITDEPTH
+
+  // Save rdmult before it might be changed, so it can be restored later.
+  orig_rdmult = x->rdmult;
+
+  if (aq_mode == VARIANCE_AQ) {
+    if (cpi->vaq_refresh) {
+      const int energy =
+          bsize <= BLOCK_16X16 ? x->mb_energy : av1_block_energy(cpi, x, bsize);
+      mbmi->segment_id = av1_vaq_segment_id(energy);
+      // Re-initialise quantiser
+      av1_init_plane_quantizers(cpi, x, mbmi->segment_id);
+    }
+    x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
+  } else if (aq_mode == COMPLEXITY_AQ) {
+    x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
+  } else if (aq_mode == CYCLIC_REFRESH_AQ) {
+    // If segment is boosted, use rdmult for that segment.
+    if (cyclic_refresh_segment_id_boosted(mbmi->segment_id))
+      x->rdmult = av1_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
+  }
+
+  // Find best coding mode & reconstruct the MB so it is available
+  // as a predictor for MBs that follow in the SB
+  if (frame_is_intra_only(cm)) {
+    av1_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
+#if CONFIG_SUPERTX
+    *totalrate_nocoef = 0;
+#endif  // CONFIG_SUPERTX
+  } else {
+    if (bsize >= BLOCK_8X8 || unify_bsize) {
+      if (segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+        av1_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, mi_row, mi_col,
+                                           rd_cost, bsize, ctx, best_rd);
+#if CONFIG_SUPERTX
+        *totalrate_nocoef = rd_cost->rate;
+#endif  // CONFIG_SUPERTX
+      } else {
+        av1_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col, rd_cost,
+#if CONFIG_SUPERTX
+                                  totalrate_nocoef,
+#endif  // CONFIG_SUPERTX
+                                  bsize, ctx, best_rd);
+#if CONFIG_SUPERTX
+        assert(*totalrate_nocoef >= 0);
+#endif  // CONFIG_SUPERTX
+      }
+    } else {
+      if (segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+        // The decoder rejects sub8x8 partitions when SEG_LVL_SKIP is set.
+        rd_cost->rate = INT_MAX;
+      } else {
+        av1_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col,
+                                      rd_cost,
+#if CONFIG_SUPERTX
+                                      totalrate_nocoef,
+#endif  // CONFIG_SUPERTX
+                                      bsize, ctx, best_rd);
+#if CONFIG_SUPERTX
+        assert(*totalrate_nocoef >= 0);
+#endif  // CONFIG_SUPERTX
+      }
+    }
+  }
+
+  // Examine the resulting rate and for AQ mode 2 make a segment choice.
+  if ((rd_cost->rate != INT_MAX) && (aq_mode == COMPLEXITY_AQ) &&
+      (bsize >= BLOCK_16X16) &&
+      (cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame ||
+       (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
+    av1_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
+  }
+
+  x->rdmult = orig_rdmult;
+
+  // TODO(jingning) The rate-distortion optimization flow needs to be
+  // refactored to provide proper exit/return handle.
+  if (rd_cost->rate == INT_MAX) rd_cost->rdcost = INT64_MAX;
+
+  ctx->rate = rd_cost->rate;
+  ctx->dist = rd_cost->dist;
+}
+
+#if CONFIG_REF_MV
+static void update_inter_mode_stats(FRAME_COUNTS *counts, PREDICTION_MODE mode,
+                                    int16_t mode_context) {
+  int16_t mode_ctx = mode_context & NEWMV_CTX_MASK;
+  if (mode == NEWMV) {
+    ++counts->newmv_mode[mode_ctx][0];
+    return;
+  } else {
+    ++counts->newmv_mode[mode_ctx][1];
+
+    if (mode_context & (1 << ALL_ZERO_FLAG_OFFSET)) {
+      return;
+    }
+
+    mode_ctx = (mode_context >> ZEROMV_OFFSET) & ZEROMV_CTX_MASK;
+    if (mode == ZEROMV) {
+      ++counts->zeromv_mode[mode_ctx][0];
+      return;
+    } else {
+      ++counts->zeromv_mode[mode_ctx][1];
+      mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK;
+
+      if (mode_context & (1 << SKIP_NEARESTMV_OFFSET)) mode_ctx = 6;
+      if (mode_context & (1 << SKIP_NEARMV_OFFSET)) mode_ctx = 7;
+      if (mode_context & (1 << SKIP_NEARESTMV_SUB8X8_OFFSET)) mode_ctx = 8;
+
+      ++counts->refmv_mode[mode_ctx][mode != NEARESTMV];
+    }
+  }
+}
+#endif
+
+static void update_stats(const AV1_COMMON *const cm, ThreadData *td, int mi_row,
+                         int mi_col
+#if CONFIG_SUPERTX
+                         ,
+                         int supertx_enabled
+#endif
+                         ) {
+#if CONFIG_DELTA_Q
+  MACROBLOCK *x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+#else
+  const MACROBLOCK *x = &td->mb;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+#endif
+  const MODE_INFO *const mi = xd->mi[0];
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  const int unify_bsize = CONFIG_CB4X4;
+
+#if CONFIG_DELTA_Q
+  // delta quant applies to both intra and inter
+  const int super_block_upper_left = ((mi_row & 7) == 0) && ((mi_col & 7) == 0);
+
+  if (cm->delta_q_present_flag && (bsize != BLOCK_64X64 || !mbmi->skip) &&
+      super_block_upper_left) {
+    const int dq = (mbmi->current_q_index - xd->prev_qindex) / cm->delta_q_res;
+    const int absdq = abs(dq);
+    int i;
+    for (i = 0; i < AOMMIN(absdq, DELTA_Q_SMALL); ++i) {
+      td->counts->delta_q[i][1]++;
+    }
+    if (absdq < DELTA_Q_SMALL) td->counts->delta_q[absdq][0]++;
+    xd->prev_qindex = mbmi->current_q_index;
+#if CONFIG_EXT_DELTA_Q
+    if (cm->delta_lf_present_flag) {
+      const int dlf =
+          (mbmi->current_delta_lf_from_base - xd->prev_delta_lf_from_base) /
+          cm->delta_lf_res;
+      const int absdlf = abs(dlf);
+      for (i = 0; i < AOMMIN(absdlf, DELTA_LF_SMALL); ++i) {
+        td->counts->delta_lf[i][1]++;
+      }
+      if (absdlf < DELTA_LF_SMALL) td->counts->delta_lf[absdlf][0]++;
+      xd->prev_delta_lf_from_base = mbmi->current_delta_lf_from_base;
+    }
+#endif
+  }
+#else
+  (void)mi_row;
+  (void)mi_col;
+#endif
+  if (!frame_is_intra_only(cm)) {
+    FRAME_COUNTS *const counts = td->counts;
+    const int inter_block = is_inter_block(mbmi);
+    const int seg_ref_active =
+        segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
+    if (!seg_ref_active) {
+#if CONFIG_SUPERTX
+      if (!supertx_enabled)
+#endif
+        counts->intra_inter[av1_get_intra_inter_context(xd)][inter_block]++;
+      // If the segment reference feature is enabled we have only a single
+      // reference frame allowed for the segment so exclude it from
+      // the reference frame counts used to work out probabilities.
+      if (inter_block) {
+        const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0];
+#if CONFIG_EXT_REFS
+        const MV_REFERENCE_FRAME ref1 = mbmi->ref_frame[1];
+#endif  // CONFIG_EXT_REFS
+
+        if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+#if !SUB8X8_COMP_REF
+          if (mbmi->sb_type >= BLOCK_8X8)
+            counts->comp_inter[av1_get_reference_mode_context(cm, xd)]
+                              [has_second_ref(mbmi)]++;
+#else
+          counts->comp_inter[av1_get_reference_mode_context(cm, xd)]
+                            [has_second_ref(mbmi)]++;
+#endif
+        }
+
+        if (has_second_ref(mbmi)) {
+#if CONFIG_EXT_REFS
+          const int bit = (ref0 == GOLDEN_FRAME || ref0 == LAST3_FRAME);
+
+          counts->comp_ref[av1_get_pred_context_comp_ref_p(cm, xd)][0][bit]++;
+          if (!bit) {
+            counts->comp_ref[av1_get_pred_context_comp_ref_p1(cm, xd)][1]
+                            [ref0 == LAST_FRAME]++;
+          } else {
+            counts->comp_ref[av1_get_pred_context_comp_ref_p2(cm, xd)][2]
+                            [ref0 == GOLDEN_FRAME]++;
+          }
+
+          counts->comp_bwdref[av1_get_pred_context_comp_bwdref_p(cm, xd)][0]
+                             [ref1 == ALTREF_FRAME]++;
+#else
+          counts->comp_ref[av1_get_pred_context_comp_ref_p(cm, xd)][0]
+                          [ref0 == GOLDEN_FRAME]++;
+#endif  // CONFIG_EXT_REFS
+        } else {
+#if CONFIG_EXT_REFS
+          const int bit = (ref0 == ALTREF_FRAME || ref0 == BWDREF_FRAME);
+
+          counts->single_ref[av1_get_pred_context_single_ref_p1(xd)][0][bit]++;
+          if (bit) {
+            counts->single_ref[av1_get_pred_context_single_ref_p2(xd)][1]
+                              [ref0 != BWDREF_FRAME]++;
+          } else {
+            const int bit1 = !(ref0 == LAST2_FRAME || ref0 == LAST_FRAME);
+            counts
+                ->single_ref[av1_get_pred_context_single_ref_p3(xd)][2][bit1]++;
+            if (!bit1) {
+              counts->single_ref[av1_get_pred_context_single_ref_p4(xd)][3]
+                                [ref0 != LAST_FRAME]++;
+            } else {
+              counts->single_ref[av1_get_pred_context_single_ref_p5(xd)][4]
+                                [ref0 != LAST3_FRAME]++;
+            }
+          }
+#else
+          counts->single_ref[av1_get_pred_context_single_ref_p1(xd)][0]
+                            [ref0 != LAST_FRAME]++;
+          if (ref0 != LAST_FRAME) {
+            counts->single_ref[av1_get_pred_context_single_ref_p2(xd)][1]
+                              [ref0 != GOLDEN_FRAME]++;
+          }
+#endif  // CONFIG_EXT_REFS
+        }
+
+#if CONFIG_EXT_INTER
+        if (cm->reference_mode != COMPOUND_REFERENCE &&
+#if CONFIG_SUPERTX
+            !supertx_enabled &&
+#endif
+            is_interintra_allowed(mbmi)) {
+          const int bsize_group = size_group_lookup[bsize];
+          if (mbmi->ref_frame[1] == INTRA_FRAME) {
+            counts->interintra[bsize_group][1]++;
+            counts->interintra_mode[bsize_group][mbmi->interintra_mode]++;
+            if (is_interintra_wedge_used(bsize))
+              counts->wedge_interintra[bsize][mbmi->use_wedge_interintra]++;
+          } else {
+            counts->interintra[bsize_group][0]++;
+          }
+        }
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+        const MOTION_MODE motion_allowed = motion_mode_allowed(
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+            0, xd->global_motion,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+            mi);
+#if CONFIG_SUPERTX
+        if (!supertx_enabled)
+#endif  // CONFIG_SUPERTX
+#if CONFIG_EXT_INTER
+          if (mbmi->ref_frame[1] != INTRA_FRAME)
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+          {
+            if (motion_allowed == WARPED_CAUSAL)
+              counts->motion_mode[mbmi->sb_type][mbmi->motion_mode]++;
+            else if (motion_allowed == OBMC_CAUSAL)
+              counts->obmc[mbmi->sb_type][mbmi->motion_mode == OBMC_CAUSAL]++;
+          }
+#else
+        if (motion_allowed > SIMPLE_TRANSLATION)
+          counts->motion_mode[mbmi->sb_type][mbmi->motion_mode]++;
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_EXT_INTER
+        if (cm->reference_mode != SINGLE_REFERENCE &&
+            is_inter_compound_mode(mbmi->mode)
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+            && mbmi->motion_mode == SIMPLE_TRANSLATION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+            ) {
+          counts->compound_interinter[bsize][mbmi->interinter_compound_type]++;
+        }
+#endif  // CONFIG_EXT_INTER
+      }
+    }
+
+    if (inter_block &&
+        !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+      int16_t mode_ctx;
+#if !CONFIG_REF_MV
+      mode_ctx = mbmi_ext->mode_context[mbmi->ref_frame[0]];
+#endif
+      if (bsize >= BLOCK_8X8 || unify_bsize) {
+        const PREDICTION_MODE mode = mbmi->mode;
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+        if (has_second_ref(mbmi)) {
+          mode_ctx = mbmi_ext->compound_mode_context[mbmi->ref_frame[0]];
+          ++counts->inter_compound_mode[mode_ctx][INTER_COMPOUND_OFFSET(mode)];
+        } else {
+#endif  // CONFIG_EXT_INTER
+          mode_ctx = av1_mode_context_analyzer(mbmi_ext->mode_context,
+                                               mbmi->ref_frame, bsize, -1);
+          update_inter_mode_stats(counts, mode, mode_ctx);
+#if CONFIG_EXT_INTER
+        }
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_EXT_INTER
+        if (mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV) {
+#else
+        if (mbmi->mode == NEWMV) {
+#endif
+          uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+          int idx;
+
+          for (idx = 0; idx < 2; ++idx) {
+            if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
+              uint8_t drl_ctx =
+                  av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx);
+              ++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx];
+
+              if (mbmi->ref_mv_idx == idx) break;
+            }
+          }
+        }
+
+#if CONFIG_EXT_INTER
+        if (have_nearmv_in_inter_mode(mbmi->mode)) {
+#else
+        if (mbmi->mode == NEARMV) {
+#endif
+          uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+          int idx;
+
+          for (idx = 1; idx < 3; ++idx) {
+            if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
+              uint8_t drl_ctx =
+                  av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx);
+              ++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx - 1];
+
+              if (mbmi->ref_mv_idx == idx - 1) break;
+            }
+          }
+        }
+#else
+#if CONFIG_EXT_INTER
+        if (is_inter_compound_mode(mode))
+          ++counts->inter_compound_mode[mode_ctx][INTER_COMPOUND_OFFSET(mode)];
+        else
+#endif  // CONFIG_EXT_INTER
+          ++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
+#endif
+      } else {
+        const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+        const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+        int idx, idy;
+        for (idy = 0; idy < 2; idy += num_4x4_h) {
+          for (idx = 0; idx < 2; idx += num_4x4_w) {
+            const int j = idy * 2 + idx;
+            const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+            if (has_second_ref(mbmi)) {
+              mode_ctx = mbmi_ext->compound_mode_context[mbmi->ref_frame[0]];
+              ++counts->inter_compound_mode[mode_ctx]
+                                           [INTER_COMPOUND_OFFSET(b_mode)];
+            } else {
+#endif  // CONFIG_EXT_INTER
+              mode_ctx = av1_mode_context_analyzer(mbmi_ext->mode_context,
+                                                   mbmi->ref_frame, bsize, j);
+              update_inter_mode_stats(counts, b_mode, mode_ctx);
+#if CONFIG_EXT_INTER
+            }
+#endif  // CONFIG_EXT_INTER
+#else
+#if CONFIG_EXT_INTER
+            if (is_inter_compound_mode(b_mode))
+              ++counts->inter_compound_mode[mode_ctx]
+                                           [INTER_COMPOUND_OFFSET(b_mode)];
+            else
+#endif  // CONFIG_EXT_INTER
+              ++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
+#endif
+          }
+        }
+      }
+    }
+  }
+}
+
+typedef struct {
+  ENTROPY_CONTEXT a[2 * MAX_MIB_SIZE * MAX_MB_PLANE];
+  ENTROPY_CONTEXT l[2 * MAX_MIB_SIZE * MAX_MB_PLANE];
+  PARTITION_CONTEXT sa[MAX_MIB_SIZE];
+  PARTITION_CONTEXT sl[MAX_MIB_SIZE];
+#if CONFIG_VAR_TX
+  TXFM_CONTEXT *p_ta;
+  TXFM_CONTEXT *p_tl;
+  TXFM_CONTEXT ta[MAX_MIB_SIZE];
+  TXFM_CONTEXT tl[MAX_MIB_SIZE];
+#endif
+} RD_SEARCH_MACROBLOCK_CONTEXT;
+
+static void restore_context(MACROBLOCK *x,
+                            const RD_SEARCH_MACROBLOCK_CONTEXT *ctx, int mi_row,
+                            int mi_col,
+#if CONFIG_PVQ
+                            od_rollback_buffer *rdo_buf,
+#endif
+                            BLOCK_SIZE bsize) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  int p;
+  const int num_4x4_blocks_wide =
+      block_size_wide[bsize] >> tx_size_wide_log2[0];
+  const int num_4x4_blocks_high =
+      block_size_high[bsize] >> tx_size_high_log2[0];
+  int mi_width = mi_size_wide[bsize];
+  int mi_height = mi_size_high[bsize];
+  for (p = 0; p < MAX_MB_PLANE; p++) {
+    memcpy(xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
+           ctx->a + num_4x4_blocks_wide * p,
+           (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
+               xd->plane[p].subsampling_x);
+    memcpy(xd->left_context[p] +
+               ((mi_row & MAX_MIB_MASK) * 2 >> xd->plane[p].subsampling_y),
+           ctx->l + num_4x4_blocks_high * p,
+           (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
+               xd->plane[p].subsampling_y);
+  }
+  memcpy(xd->above_seg_context + mi_col, ctx->sa,
+         sizeof(*xd->above_seg_context) * mi_width);
+  memcpy(xd->left_seg_context + (mi_row & MAX_MIB_MASK), ctx->sl,
+         sizeof(xd->left_seg_context[0]) * mi_height);
+#if CONFIG_VAR_TX
+  xd->above_txfm_context = ctx->p_ta;
+  xd->left_txfm_context = ctx->p_tl;
+  memcpy(xd->above_txfm_context, ctx->ta,
+         sizeof(*xd->above_txfm_context) * mi_width);
+  memcpy(xd->left_txfm_context, ctx->tl,
+         sizeof(*xd->left_txfm_context) * mi_height);
+#endif
+#if CONFIG_PVQ
+  od_encode_rollback(&x->daala_enc, rdo_buf);
+#endif
+}
+
+static void save_context(const MACROBLOCK *x, RD_SEARCH_MACROBLOCK_CONTEXT *ctx,
+                         int mi_row, int mi_col,
+#if CONFIG_PVQ
+                         od_rollback_buffer *rdo_buf,
+#endif
+                         BLOCK_SIZE bsize) {
+  const MACROBLOCKD *xd = &x->e_mbd;
+  int p;
+  const int num_4x4_blocks_wide =
+      block_size_wide[bsize] >> tx_size_wide_log2[0];
+  const int num_4x4_blocks_high =
+      block_size_high[bsize] >> tx_size_high_log2[0];
+  int mi_width = mi_size_wide[bsize];
+  int mi_height = mi_size_high[bsize];
+
+  // buffer the above/left context information of the block in search.
+  for (p = 0; p < MAX_MB_PLANE; ++p) {
+    memcpy(ctx->a + num_4x4_blocks_wide * p,
+           xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
+           (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
+               xd->plane[p].subsampling_x);
+    memcpy(ctx->l + num_4x4_blocks_high * p,
+           xd->left_context[p] +
+               ((mi_row & MAX_MIB_MASK) * 2 >> xd->plane[p].subsampling_y),
+           (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
+               xd->plane[p].subsampling_y);
+  }
+  memcpy(ctx->sa, xd->above_seg_context + mi_col,
+         sizeof(*xd->above_seg_context) * mi_width);
+  memcpy(ctx->sl, xd->left_seg_context + (mi_row & MAX_MIB_MASK),
+         sizeof(xd->left_seg_context[0]) * mi_height);
+#if CONFIG_VAR_TX
+  memcpy(ctx->ta, xd->above_txfm_context,
+         sizeof(*xd->above_txfm_context) * mi_width);
+  memcpy(ctx->tl, xd->left_txfm_context,
+         sizeof(*xd->left_txfm_context) * mi_height);
+  ctx->p_ta = xd->above_txfm_context;
+  ctx->p_tl = xd->left_txfm_context;
+#endif
+#if CONFIG_PVQ
+  od_encode_checkpoint(&x->daala_enc, rdo_buf);
+#endif
+}
+
+static void encode_b(const AV1_COMP *const cpi, const TileInfo *const tile,
+                     ThreadData *td, TOKENEXTRA **tp, int mi_row, int mi_col,
+                     RUN_TYPE dry_run, BLOCK_SIZE bsize,
+#if CONFIG_EXT_PARTITION_TYPES
+                     PARTITION_TYPE partition,
+#endif
+                     PICK_MODE_CONTEXT *ctx, int *rate) {
+  MACROBLOCK *const x = &td->mb;
+#if (CONFIG_MOTION_VAR && CONFIG_NCOBMC) | CONFIG_EXT_DELTA_Q
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi;
+#if CONFIG_MOTION_VAR && CONFIG_NCOBMC
+  int check_ncobmc;
+#endif
+#endif
+
+  set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
+#if CONFIG_EXT_PARTITION_TYPES
+  x->e_mbd.mi[0]->mbmi.partition = partition;
+#endif
+  update_state(cpi, td, ctx, mi_row, mi_col, bsize, dry_run);
+#if CONFIG_MOTION_VAR && CONFIG_NCOBMC
+  mbmi = &xd->mi[0]->mbmi;
+  const MOTION_MODE motion_allowed = motion_mode_allowed(
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      0, xd->global_motion,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      xd->mi[0]);
+  check_ncobmc = is_inter_block(mbmi) && motion_allowed >= OBMC_CAUSAL;
+  if (!dry_run && check_ncobmc) {
+    av1_check_ncobmc_rd(cpi, x, mi_row, mi_col);
+    av1_setup_dst_planes(x->e_mbd.plane, bsize,
+                         get_frame_new_buffer(&cpi->common), mi_row, mi_col);
+  }
+#endif
+  encode_superblock(cpi, td, tp, dry_run, mi_row, mi_col, bsize, ctx, rate);
+
+  if (!dry_run) {
+#if CONFIG_EXT_DELTA_Q
+    mbmi = &xd->mi[0]->mbmi;
+    if (bsize == BLOCK_64X64 && mbmi->skip == 1 && is_inter_block(mbmi) &&
+        cpi->common.delta_lf_present_flag) {
+      mbmi->current_delta_lf_from_base = xd->prev_delta_lf_from_base;
+    }
+#endif
+#if CONFIG_SUPERTX
+    update_stats(&cpi->common, td, mi_row, mi_col, 0);
+#else
+    update_stats(&cpi->common, td, mi_row, mi_col);
+#endif
+  }
+}
+
+static void encode_sb(const AV1_COMP *const cpi, ThreadData *td,
+                      const TileInfo *const tile, TOKENEXTRA **tp, int mi_row,
+                      int mi_col, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                      PC_TREE *pc_tree, int *rate) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int hbs = mi_size_wide[bsize] / 2;
+  const int is_partition_root = bsize >= BLOCK_8X8;
+  const int ctx = is_partition_root
+                      ? partition_plane_context(xd, mi_row, mi_col,
+#if CONFIG_UNPOISON_PARTITION_CTX
+                                                mi_row + hbs < cm->mi_rows,
+                                                mi_col + hbs < cm->mi_cols,
+#endif
+                                                bsize)
+                      : -1;
+  const PARTITION_TYPE partition = pc_tree->partitioning;
+  const BLOCK_SIZE subsize = get_subsize(bsize, partition);
+#if CONFIG_EXT_PARTITION_TYPES
+  const BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+  assert(bsize >= BLOCK_8X8);
+#endif
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  if (!dry_run && ctx >= 0) td->counts->partition[ctx][partition]++;
+
+#if CONFIG_SUPERTX
+  if (!frame_is_intra_only(cm) && bsize <= MAX_SUPERTX_BLOCK_SIZE &&
+      partition != PARTITION_NONE && !xd->lossless[0]) {
+    int supertx_enabled;
+    TX_SIZE supertx_size = max_txsize_lookup[bsize];
+    supertx_enabled = check_supertx_sb(bsize, supertx_size, pc_tree);
+    if (supertx_enabled) {
+      const int mi_width = mi_size_wide[bsize];
+      const int mi_height = mi_size_high[bsize];
+      int x_idx, y_idx, i;
+      uint8_t *dst_buf[3];
+      int dst_stride[3];
+      set_skip_context(xd, mi_row, mi_col);
+      set_mode_info_offsets(cpi, x, xd, mi_row, mi_col);
+      update_state_sb_supertx(cpi, td, tile, mi_row, mi_col, bsize, dry_run,
+                              pc_tree);
+
+      av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                           mi_col);
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        dst_buf[i] = xd->plane[i].dst.buf;
+        dst_stride[i] = xd->plane[i].dst.stride;
+      }
+      predict_sb_complex(cpi, td, tile, mi_row, mi_col, mi_row, mi_col, dry_run,
+                         bsize, bsize, dst_buf, dst_stride, pc_tree);
+
+      set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize);
+      set_segment_id_supertx(cpi, x, mi_row, mi_col, bsize);
+
+      if (!x->skip) {
+        int this_rate = 0;
+        av1_encode_sb_supertx((AV1_COMMON *)cm, x, bsize);
+        av1_tokenize_sb_supertx(cpi, td, tp, dry_run, bsize, rate);
+        if (rate) *rate += this_rate;
+      } else {
+        xd->mi[0]->mbmi.skip = 1;
+        if (!dry_run) td->counts->skip[av1_get_skip_context(xd)][1]++;
+        reset_skip_context(xd, bsize);
+      }
+      if (!dry_run) {
+        for (y_idx = 0; y_idx < mi_height; y_idx++)
+          for (x_idx = 0; x_idx < mi_width; x_idx++) {
+            if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width >
+                    x_idx &&
+                (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height >
+                    y_idx) {
+              xd->mi[x_idx + y_idx * cm->mi_stride]->mbmi.skip =
+                  xd->mi[0]->mbmi.skip;
+            }
+          }
+        td->counts->supertx[partition_supertx_context_lookup[partition]]
+                           [supertx_size][1]++;
+        td->counts->supertx_size[supertx_size]++;
+#if CONFIG_EXT_TX
+        if (get_ext_tx_types(supertx_size, bsize, 1, cm->reduced_tx_set_used) >
+                1 &&
+            !xd->mi[0]->mbmi.skip) {
+          const int eset =
+              get_ext_tx_set(supertx_size, bsize, 1, cm->reduced_tx_set_used);
+          if (eset > 0) {
+            ++td->counts
+                  ->inter_ext_tx[eset][supertx_size][xd->mi[0]->mbmi.tx_type];
+          }
+        }
+#else
+        if (supertx_size < TX_32X32 && !xd->mi[0]->mbmi.skip) {
+          ++td->counts->inter_ext_tx[supertx_size][xd->mi[0]->mbmi.tx_type];
+        }
+#endif  // CONFIG_EXT_TX
+      }
+#if CONFIG_EXT_PARTITION_TYPES
+      update_ext_partition_context(xd, mi_row, mi_col, subsize, bsize,
+                                   partition);
+#else
+      if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
+        update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+#endif
+#if CONFIG_VAR_TX
+      set_txfm_ctxs(supertx_size, mi_width, mi_height, xd->mi[0]->mbmi.skip,
+                    xd);
+#endif  // CONFIG_VAR_TX
+      return;
+    } else {
+      if (!dry_run) {
+        td->counts->supertx[partition_supertx_context_lookup[partition]]
+                           [supertx_size][0]++;
+      }
+    }
+  }
+#endif  // CONFIG_SUPERTX
+
+  switch (partition) {
+    case PARTITION_NONE:
+      encode_b(cpi, tile, td, tp, mi_row, mi_col, dry_run, subsize,
+#if CONFIG_EXT_PARTITION_TYPES
+               partition,
+#endif
+               &pc_tree->none, rate);
+      break;
+    case PARTITION_VERT:
+      encode_b(cpi, tile, td, tp, mi_row, mi_col, dry_run, subsize,
+#if CONFIG_EXT_PARTITION_TYPES
+               partition,
+#endif
+               &pc_tree->vertical[0], rate);
+      if (mi_col + hbs < cm->mi_cols && (bsize > BLOCK_8X8 || unify_bsize)) {
+        encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, dry_run, subsize,
+#if CONFIG_EXT_PARTITION_TYPES
+                 partition,
+#endif
+                 &pc_tree->vertical[1], rate);
+      }
+      break;
+    case PARTITION_HORZ:
+      encode_b(cpi, tile, td, tp, mi_row, mi_col, dry_run, subsize,
+#if CONFIG_EXT_PARTITION_TYPES
+               partition,
+#endif
+               &pc_tree->horizontal[0], rate);
+      if (mi_row + hbs < cm->mi_rows && (bsize > BLOCK_8X8 || unify_bsize)) {
+        encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, dry_run, subsize,
+#if CONFIG_EXT_PARTITION_TYPES
+                 partition,
+#endif
+                 &pc_tree->horizontal[1], rate);
+      }
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        encode_b(cpi, tile, td, tp, mi_row, mi_col, dry_run, subsize,
+#if CONFIG_EXT_PARTITION_TYPES
+                 partition,
+#endif
+                 pc_tree->leaf_split[0], rate);
+      } else {
+        encode_sb(cpi, td, tile, tp, mi_row, mi_col, dry_run, subsize,
+                  pc_tree->split[0], rate);
+        encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, dry_run, subsize,
+                  pc_tree->split[1], rate);
+        encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, dry_run, subsize,
+                  pc_tree->split[2], rate);
+        encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, dry_run,
+                  subsize, pc_tree->split[3], rate);
+      }
+      break;
+#if CONFIG_EXT_PARTITION_TYPES
+    case PARTITION_HORZ_A:
+      encode_b(cpi, tile, td, tp, mi_row, mi_col, dry_run, bsize2, partition,
+               &pc_tree->horizontala[0], rate);
+      encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, dry_run, bsize2,
+               partition, &pc_tree->horizontala[1], rate);
+      encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, dry_run, subsize,
+               partition, &pc_tree->horizontala[2], rate);
+      break;
+    case PARTITION_HORZ_B:
+      encode_b(cpi, tile, td, tp, mi_row, mi_col, dry_run, subsize, partition,
+               &pc_tree->horizontalb[0], rate);
+      encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, dry_run, bsize2,
+               partition, &pc_tree->horizontalb[1], rate);
+      encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col + hbs, dry_run, bsize2,
+               partition, &pc_tree->horizontalb[2], rate);
+      break;
+    case PARTITION_VERT_A:
+      encode_b(cpi, tile, td, tp, mi_row, mi_col, dry_run, bsize2, partition,
+               &pc_tree->verticala[0], rate);
+      encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, dry_run, bsize2,
+               partition, &pc_tree->verticala[1], rate);
+      encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, dry_run, subsize,
+               partition, &pc_tree->verticala[2], rate);
+
+      break;
+    case PARTITION_VERT_B:
+      encode_b(cpi, tile, td, tp, mi_row, mi_col, dry_run, subsize, partition,
+               &pc_tree->verticalb[0], rate);
+      encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, dry_run, bsize2,
+               partition, &pc_tree->verticalb[1], rate);
+      encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col + hbs, dry_run, bsize2,
+               partition, &pc_tree->verticalb[2], rate);
+      break;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+    default: assert(0 && "Invalid partition type."); break;
+  }
+
+#if CONFIG_EXT_PARTITION_TYPES
+  update_ext_partition_context(xd, mi_row, mi_col, subsize, bsize, partition);
+#else
+  if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
+    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+}
+
+// Check to see if the given partition size is allowed for a specified number
+// of mi block rows and columns remaining in the image.
+// If not then return the largest allowed partition size
+static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize, int rows_left,
+                                      int cols_left, int *bh, int *bw) {
+  if (rows_left <= 0 || cols_left <= 0) {
+    return AOMMIN(bsize, BLOCK_8X8);
+  } else {
+    for (; bsize > 0; bsize -= 3) {
+      *bh = mi_size_high[bsize];
+      *bw = mi_size_wide[bsize];
+      if ((*bh <= rows_left) && (*bw <= cols_left)) {
+        break;
+      }
+    }
+  }
+  return bsize;
+}
+
+static void set_partial_sb_partition(const AV1_COMMON *const cm, MODE_INFO *mi,
+                                     int bh_in, int bw_in,
+                                     int mi_rows_remaining,
+                                     int mi_cols_remaining, BLOCK_SIZE bsize,
+                                     MODE_INFO **mib) {
+  int bh = bh_in;
+  int r, c;
+  for (r = 0; r < cm->mib_size; r += bh) {
+    int bw = bw_in;
+    for (c = 0; c < cm->mib_size; c += bw) {
+      const int index = r * cm->mi_stride + c;
+      mib[index] = mi + index;
+      mib[index]->mbmi.sb_type = find_partition_size(
+          bsize, mi_rows_remaining - r, mi_cols_remaining - c, &bh, &bw);
+    }
+  }
+}
+
+// This function attempts to set all mode info entries in a given superblock
+// to the same block partition size.
+// However, at the bottom and right borders of the image the requested size
+// may not be allowed in which case this code attempts to choose the largest
+// allowable partition.
+static void set_fixed_partitioning(AV1_COMP *cpi, const TileInfo *const tile,
+                                   MODE_INFO **mib, int mi_row, int mi_col,
+                                   BLOCK_SIZE bsize) {
+  AV1_COMMON *const cm = &cpi->common;
+  const int mi_rows_remaining = tile->mi_row_end - mi_row;
+  const int mi_cols_remaining = tile->mi_col_end - mi_col;
+  int block_row, block_col;
+  MODE_INFO *const mi_upper_left = cm->mi + mi_row * cm->mi_stride + mi_col;
+  int bh = mi_size_high[bsize];
+  int bw = mi_size_wide[bsize];
+
+  assert((mi_rows_remaining > 0) && (mi_cols_remaining > 0));
+
+  // Apply the requested partition size to the SB if it is all "in image"
+  if ((mi_cols_remaining >= cm->mib_size) &&
+      (mi_rows_remaining >= cm->mib_size)) {
+    for (block_row = 0; block_row < cm->mib_size; block_row += bh) {
+      for (block_col = 0; block_col < cm->mib_size; block_col += bw) {
+        int index = block_row * cm->mi_stride + block_col;
+        mib[index] = mi_upper_left + index;
+        mib[index]->mbmi.sb_type = bsize;
+      }
+    }
+  } else {
+    // Else this is a partial SB.
+    set_partial_sb_partition(cm, mi_upper_left, bh, bw, mi_rows_remaining,
+                             mi_cols_remaining, bsize, mib);
+  }
+}
+
+static void rd_use_partition(AV1_COMP *cpi, ThreadData *td,
+                             TileDataEnc *tile_data, MODE_INFO **mib,
+                             TOKENEXTRA **tp, int mi_row, int mi_col,
+                             BLOCK_SIZE bsize, int *rate, int64_t *dist,
+#if CONFIG_SUPERTX
+                             int *rate_nocoef,
+#endif
+                             int do_recon, PC_TREE *pc_tree) {
+  AV1_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int bs = mi_size_wide[bsize];
+  const int hbs = bs / 2;
+  int i;
+  const int pl = (bsize >= BLOCK_8X8)
+                     ? partition_plane_context(xd, mi_row, mi_col,
+#if CONFIG_UNPOISON_PARTITION_CTX
+                                               mi_row + hbs < cm->mi_rows,
+                                               mi_col + hbs < cm->mi_cols,
+#endif
+                                               bsize)
+                     : 0;
+  const PARTITION_TYPE partition =
+      (bsize >= BLOCK_8X8) ? get_partition(cm, mi_row, mi_col, bsize)
+                           : PARTITION_NONE;
+  const BLOCK_SIZE subsize = get_subsize(bsize, partition);
+  RD_SEARCH_MACROBLOCK_CONTEXT x_ctx;
+  RD_STATS last_part_rdc, none_rdc, chosen_rdc;
+  BLOCK_SIZE sub_subsize = BLOCK_4X4;
+  int splits_below = 0;
+  BLOCK_SIZE bs_type = mib[0]->mbmi.sb_type;
+  int do_partition_search = 1;
+  PICK_MODE_CONTEXT *ctx_none = &pc_tree->none;
+  const int unify_bsize = CONFIG_CB4X4;
+#if CONFIG_SUPERTX
+  int last_part_rate_nocoef = INT_MAX;
+  int none_rate_nocoef = INT_MAX;
+  int chosen_rate_nocoef = INT_MAX;
+#endif
+#if CONFIG_PVQ
+  od_rollback_buffer pre_rdo_buf;
+#endif
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  assert(num_4x4_blocks_wide_lookup[bsize] ==
+         num_4x4_blocks_high_lookup[bsize]);
+
+  av1_invalid_rd_stats(&last_part_rdc);
+  av1_invalid_rd_stats(&none_rdc);
+  av1_invalid_rd_stats(&chosen_rdc);
+
+  pc_tree->partitioning = partition;
+
+#if CONFIG_VAR_TX
+  xd->above_txfm_context = cm->above_txfm_context + mi_col;
+  xd->left_txfm_context =
+      xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+#endif
+#if !CONFIG_PVQ
+  save_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+  save_context(x, &x_ctx, mi_row, mi_col, &pre_rdo_buf, bsize);
+#endif
+
+  if (bsize == BLOCK_16X16 && cpi->vaq_refresh) {
+    set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+    x->mb_energy = av1_block_energy(cpi, x, bsize);
+  }
+
+  if (do_partition_search &&
+      cpi->sf.partition_search_type == SEARCH_PARTITION &&
+      cpi->sf.adjust_partitioning_from_last_frame) {
+    // Check if any of the sub blocks are further split.
+    if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
+      sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
+      splits_below = 1;
+      for (i = 0; i < 4; i++) {
+        int jj = i >> 1, ii = i & 0x01;
+        MODE_INFO *this_mi = mib[jj * hbs * cm->mi_stride + ii * hbs];
+        if (this_mi && this_mi->mbmi.sb_type >= sub_subsize) {
+          splits_below = 0;
+        }
+      }
+    }
+
+    // If partition is not none try none unless each of the 4 splits are split
+    // even further..
+    if (partition != PARTITION_NONE && !splits_below &&
+        mi_row + hbs < cm->mi_rows && mi_col + hbs < cm->mi_cols) {
+      pc_tree->partitioning = PARTITION_NONE;
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc,
+#if CONFIG_SUPERTX
+                       &none_rate_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_NONE,
+#endif
+                       bsize, ctx_none, INT64_MAX);
+
+      if (none_rdc.rate < INT_MAX) {
+        none_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
+        none_rdc.rdcost =
+            RDCOST(x->rdmult, x->rddiv, none_rdc.rate, none_rdc.dist);
+#if CONFIG_SUPERTX
+        none_rate_nocoef += cpi->partition_cost[pl][PARTITION_NONE];
+#endif
+      }
+
+#if !CONFIG_PVQ
+      restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+      restore_context(x, &x_ctx, mi_row, mi_col, &pre_rdo_buf, bsize);
+#endif
+      mib[0]->mbmi.sb_type = bs_type;
+      pc_tree->partitioning = partition;
+    }
+  }
+
+  switch (partition) {
+    case PARTITION_NONE:
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
+#if CONFIG_SUPERTX
+                       &last_part_rate_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_NONE,
+#endif
+                       bsize, ctx_none, INT64_MAX);
+      break;
+    case PARTITION_HORZ:
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
+#if CONFIG_SUPERTX
+                       &last_part_rate_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_HORZ,
+#endif
+                       subsize, &pc_tree->horizontal[0], INT64_MAX);
+      if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
+          mi_row + hbs < cm->mi_rows) {
+        RD_STATS tmp_rdc;
+#if CONFIG_SUPERTX
+        int rt_nocoef = 0;
+#endif
+        PICK_MODE_CONTEXT *ctx_h = &pc_tree->horizontal[0];
+        av1_init_rd_stats(&tmp_rdc);
+        update_state(cpi, td, ctx_h, mi_row, mi_col, subsize, 1);
+        encode_superblock(cpi, td, tp, DRY_RUN_NORMAL, mi_row, mi_col, subsize,
+                          ctx_h, NULL);
+        rd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, &tmp_rdc,
+#if CONFIG_SUPERTX
+                         &rt_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                         PARTITION_HORZ,
+#endif
+                         subsize, &pc_tree->horizontal[1], INT64_MAX);
+        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
+          av1_invalid_rd_stats(&last_part_rdc);
+#if CONFIG_SUPERTX
+          last_part_rate_nocoef = INT_MAX;
+#endif
+          break;
+        }
+        last_part_rdc.rate += tmp_rdc.rate;
+        last_part_rdc.dist += tmp_rdc.dist;
+        last_part_rdc.rdcost += tmp_rdc.rdcost;
+#if CONFIG_SUPERTX
+        last_part_rate_nocoef += rt_nocoef;
+#endif
+      }
+      break;
+    case PARTITION_VERT:
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
+#if CONFIG_SUPERTX
+                       &last_part_rate_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_VERT,
+#endif
+                       subsize, &pc_tree->vertical[0], INT64_MAX);
+      if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
+          mi_col + hbs < cm->mi_cols) {
+        RD_STATS tmp_rdc;
+#if CONFIG_SUPERTX
+        int rt_nocoef = 0;
+#endif
+        PICK_MODE_CONTEXT *ctx_v = &pc_tree->vertical[0];
+        av1_init_rd_stats(&tmp_rdc);
+        update_state(cpi, td, ctx_v, mi_row, mi_col, subsize, 1);
+        encode_superblock(cpi, td, tp, DRY_RUN_NORMAL, mi_row, mi_col, subsize,
+                          ctx_v, NULL);
+        rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, &tmp_rdc,
+#if CONFIG_SUPERTX
+                         &rt_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                         PARTITION_VERT,
+#endif
+                         subsize, &pc_tree->vertical[bsize > BLOCK_8X8],
+                         INT64_MAX);
+        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
+          av1_invalid_rd_stats(&last_part_rdc);
+#if CONFIG_SUPERTX
+          last_part_rate_nocoef = INT_MAX;
+#endif
+          break;
+        }
+        last_part_rdc.rate += tmp_rdc.rate;
+        last_part_rdc.dist += tmp_rdc.dist;
+        last_part_rdc.rdcost += tmp_rdc.rdcost;
+#if CONFIG_SUPERTX
+        last_part_rate_nocoef += rt_nocoef;
+#endif
+      }
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
+#if CONFIG_SUPERTX
+                         &last_part_rate_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                         PARTITION_SPLIT,
+#endif
+                         subsize, pc_tree->leaf_split[0], INT64_MAX);
+        break;
+      }
+      last_part_rdc.rate = 0;
+      last_part_rdc.dist = 0;
+      last_part_rdc.rdcost = 0;
+#if CONFIG_SUPERTX
+      last_part_rate_nocoef = 0;
+#endif
+      for (i = 0; i < 4; i++) {
+        int x_idx = (i & 1) * hbs;
+        int y_idx = (i >> 1) * hbs;
+        int jj = i >> 1, ii = i & 0x01;
+        RD_STATS tmp_rdc;
+#if CONFIG_SUPERTX
+        int rt_nocoef;
+#endif
+        if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
+          continue;
+
+        av1_init_rd_stats(&tmp_rdc);
+        rd_use_partition(cpi, td, tile_data,
+                         mib + jj * hbs * cm->mi_stride + ii * hbs, tp,
+                         mi_row + y_idx, mi_col + x_idx, subsize, &tmp_rdc.rate,
+                         &tmp_rdc.dist,
+#if CONFIG_SUPERTX
+                         &rt_nocoef,
+#endif
+                         i != 3, pc_tree->split[i]);
+        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
+          av1_invalid_rd_stats(&last_part_rdc);
+#if CONFIG_SUPERTX
+          last_part_rate_nocoef = INT_MAX;
+#endif
+          break;
+        }
+        last_part_rdc.rate += tmp_rdc.rate;
+        last_part_rdc.dist += tmp_rdc.dist;
+#if CONFIG_SUPERTX
+        last_part_rate_nocoef += rt_nocoef;
+#endif
+      }
+      break;
+#if CONFIG_EXT_PARTITION_TYPES
+    case PARTITION_VERT_A:
+    case PARTITION_VERT_B:
+    case PARTITION_HORZ_A:
+    case PARTITION_HORZ_B: assert(0 && "Cannot handle extended partiton types");
+#endif  //  CONFIG_EXT_PARTITION_TYPES
+    default: assert(0); break;
+  }
+
+  if (last_part_rdc.rate < INT_MAX) {
+    last_part_rdc.rate += cpi->partition_cost[pl][partition];
+    last_part_rdc.rdcost =
+        RDCOST(x->rdmult, x->rddiv, last_part_rdc.rate, last_part_rdc.dist);
+#if CONFIG_SUPERTX
+    last_part_rate_nocoef += cpi->partition_cost[pl][partition];
+#endif
+  }
+
+  if (do_partition_search && cpi->sf.adjust_partitioning_from_last_frame &&
+      cpi->sf.partition_search_type == SEARCH_PARTITION &&
+      partition != PARTITION_SPLIT && bsize > BLOCK_8X8 &&
+      (mi_row + bs < cm->mi_rows || mi_row + hbs == cm->mi_rows) &&
+      (mi_col + bs < cm->mi_cols || mi_col + hbs == cm->mi_cols)) {
+    BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
+    chosen_rdc.rate = 0;
+    chosen_rdc.dist = 0;
+#if CONFIG_SUPERTX
+    chosen_rate_nocoef = 0;
+#endif
+#if !CONFIG_PVQ
+    restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+    restore_context(x, &x_ctx, mi_row, mi_col, &pre_rdo_buf, bsize);
+#endif
+    pc_tree->partitioning = PARTITION_SPLIT;
+
+    // Split partition.
+    for (i = 0; i < 4; i++) {
+      int x_idx = (i & 1) * hbs;
+      int y_idx = (i >> 1) * hbs;
+      RD_STATS tmp_rdc;
+#if CONFIG_SUPERTX
+      int rt_nocoef = 0;
+#endif
+#if CONFIG_PVQ
+      od_rollback_buffer buf;
+#endif
+      if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
+        continue;
+
+#if !CONFIG_PVQ
+      save_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+      save_context(x, &x_ctx, mi_row, mi_col, &buf, bsize);
+#endif
+      pc_tree->split[i]->partitioning = PARTITION_NONE;
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx,
+                       &tmp_rdc,
+#if CONFIG_SUPERTX
+                       &rt_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_SPLIT,
+#endif
+                       split_subsize, &pc_tree->split[i]->none, INT64_MAX);
+
+#if !CONFIG_PVQ
+      restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+      restore_context(x, &x_ctx, mi_row, mi_col, &buf, bsize);
+#endif
+      if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
+        av1_invalid_rd_stats(&chosen_rdc);
+#if CONFIG_SUPERTX
+        chosen_rate_nocoef = INT_MAX;
+#endif
+        break;
+      }
+
+      chosen_rdc.rate += tmp_rdc.rate;
+      chosen_rdc.dist += tmp_rdc.dist;
+#if CONFIG_SUPERTX
+      chosen_rate_nocoef += rt_nocoef;
+#endif
+
+      if (i != 3)
+        encode_sb(cpi, td, tile_info, tp, mi_row + y_idx, mi_col + x_idx,
+                  OUTPUT_ENABLED, split_subsize, pc_tree->split[i], NULL);
+
+      chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
+#if CONFIG_SUPERTX
+      chosen_rate_nocoef += cpi->partition_cost[pl][PARTITION_SPLIT];
+#endif
+    }
+    if (chosen_rdc.rate < INT_MAX) {
+      chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
+      chosen_rdc.rdcost =
+          RDCOST(x->rdmult, x->rddiv, chosen_rdc.rate, chosen_rdc.dist);
+#if CONFIG_SUPERTX
+      chosen_rate_nocoef += cpi->partition_cost[pl][PARTITION_NONE];
+#endif
+    }
+  }
+
+  // If last_part is better set the partitioning to that.
+  if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
+    mib[0]->mbmi.sb_type = bsize;
+    if (bsize >= BLOCK_8X8) pc_tree->partitioning = partition;
+    chosen_rdc = last_part_rdc;
+#if CONFIG_SUPERTX
+    chosen_rate_nocoef = last_part_rate_nocoef;
+#endif
+  }
+  // If none was better set the partitioning to that.
+  if (none_rdc.rdcost < chosen_rdc.rdcost) {
+    if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
+    chosen_rdc = none_rdc;
+#if CONFIG_SUPERTX
+    chosen_rate_nocoef = none_rate_nocoef;
+#endif
+  }
+
+#if !CONFIG_PVQ
+  restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+  restore_context(x, &x_ctx, mi_row, mi_col, &pre_rdo_buf, bsize);
+#endif
+
+  // We must have chosen a partitioning and encoding or we'll fail later on.
+  // No other opportunities for success.
+  if (bsize == cm->sb_size)
+    assert(chosen_rdc.rate < INT_MAX && chosen_rdc.dist < INT64_MAX);
+
+  if (do_recon) {
+    if (bsize == cm->sb_size) {
+      // NOTE: To get estimate for rate due to the tokens, use:
+      // int rate_coeffs = 0;
+      // encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, DRY_RUN_COSTCOEFFS,
+      //           bsize, pc_tree, &rate_coeffs);
+      encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, OUTPUT_ENABLED, bsize,
+                pc_tree, NULL);
+    } else {
+      encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, DRY_RUN_NORMAL, bsize,
+                pc_tree, NULL);
+    }
+  }
+
+  *rate = chosen_rdc.rate;
+  *dist = chosen_rdc.dist;
+#if CONFIG_SUPERTX
+  *rate_nocoef = chosen_rate_nocoef;
+#endif
+}
+
+/* clang-format off */
+static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  BLOCK_2X2,   BLOCK_2X2,   BLOCK_2X2,    //    2x2,    2x4,     4x2
+#endif
+                            BLOCK_4X4,    //                     4x4
+  BLOCK_4X4,   BLOCK_4X4,   BLOCK_4X4,    //    4x8,    8x4,     8x8
+  BLOCK_4X4,   BLOCK_4X4,   BLOCK_8X8,    //   8x16,   16x8,   16x16
+  BLOCK_8X8,   BLOCK_8X8,   BLOCK_16X16,  //  16x32,  32x16,   32x32
+  BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,  //  32x64,  64x32,   64x64
+#if CONFIG_EXT_PARTITION
+  BLOCK_16X16, BLOCK_16X16, BLOCK_16X16   // 64x128, 128x64, 128x128
+#endif  // CONFIG_EXT_PARTITION
+};
+
+static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  BLOCK_4X4,     BLOCK_4X4,       BLOCK_4X4,    //    2x2,    2x4,     4x2
+#endif
+                                  BLOCK_8X8,    //                     4x4
+  BLOCK_16X16,   BLOCK_16X16,   BLOCK_16X16,    //    4x8,    8x4,     8x8
+  BLOCK_32X32,   BLOCK_32X32,   BLOCK_32X32,    //   8x16,   16x8,   16x16
+  BLOCK_64X64,   BLOCK_64X64,   BLOCK_64X64,    //  16x32,  32x16,   32x32
+  BLOCK_LARGEST, BLOCK_LARGEST, BLOCK_LARGEST,  //  32x64,  64x32,   64x64
+#if CONFIG_EXT_PARTITION
+  BLOCK_LARGEST, BLOCK_LARGEST, BLOCK_LARGEST   // 64x128, 128x64, 128x128
+#endif  // CONFIG_EXT_PARTITION
+};
+
+// Next square block size less or equal than current block size.
+static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  BLOCK_2X2,   BLOCK_2X2,     BLOCK_2X2,    //    2x2,    2x4,     4x2
+#endif
+                              BLOCK_4X4,    //                     4x4
+  BLOCK_4X4,   BLOCK_4X4,     BLOCK_8X8,    //    4x8,    8x4,     8x8
+  BLOCK_8X8,   BLOCK_8X8,     BLOCK_16X16,  //   8x16,   16x8,   16x16
+  BLOCK_16X16, BLOCK_16X16,   BLOCK_32X32,  //  16x32,  32x16,   32x32
+  BLOCK_32X32, BLOCK_32X32,   BLOCK_64X64,  //  32x64,  64x32,   64x64
+#if CONFIG_EXT_PARTITION
+  BLOCK_64X64, BLOCK_64X64, BLOCK_128X128   // 64x128, 128x64, 128x128
+#endif  // CONFIG_EXT_PARTITION
+};
+/* clang-format on */
+
+// Look at all the mode_info entries for blocks that are part of this
+// partition and find the min and max values for sb_type.
+// At the moment this is designed to work on a superblock but could be
+// adjusted to use a size parameter.
+//
+// The min and max are assumed to have been initialized prior to calling this
+// function so repeat calls can accumulate a min and max of more than one
+// superblock.
+static void get_sb_partition_size_range(const AV1_COMMON *const cm,
+                                        MACROBLOCKD *xd, MODE_INFO **mib,
+                                        BLOCK_SIZE *min_block_size,
+                                        BLOCK_SIZE *max_block_size) {
+  int i, j;
+  int index = 0;
+
+  // Check the sb_type for each block that belongs to this region.
+  for (i = 0; i < cm->mib_size; ++i) {
+    for (j = 0; j < cm->mib_size; ++j) {
+      MODE_INFO *mi = mib[index + j];
+      BLOCK_SIZE sb_type = mi ? mi->mbmi.sb_type : BLOCK_4X4;
+      *min_block_size = AOMMIN(*min_block_size, sb_type);
+      *max_block_size = AOMMAX(*max_block_size, sb_type);
+    }
+    index += xd->mi_stride;
+  }
+}
+
+// Look at neighboring blocks and set a min and max partition size based on
+// what they chose.
+static void rd_auto_partition_range(AV1_COMP *cpi, const TileInfo *const tile,
+                                    MACROBLOCKD *const xd, int mi_row,
+                                    int mi_col, BLOCK_SIZE *min_block_size,
+                                    BLOCK_SIZE *max_block_size) {
+  AV1_COMMON *const cm = &cpi->common;
+  MODE_INFO **mi = xd->mi;
+  const int left_in_image = xd->left_available && mi[-1];
+  const int above_in_image = xd->up_available && mi[-xd->mi_stride];
+  const int mi_rows_remaining = tile->mi_row_end - mi_row;
+  const int mi_cols_remaining = tile->mi_col_end - mi_col;
+  int bh, bw;
+  BLOCK_SIZE min_size = BLOCK_4X4;
+  BLOCK_SIZE max_size = BLOCK_LARGEST;
+
+  // Trap case where we do not have a prediction.
+  if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
+    // Default "min to max" and "max to min"
+    min_size = BLOCK_LARGEST;
+    max_size = BLOCK_4X4;
+
+    // NOTE: each call to get_sb_partition_size_range() uses the previous
+    // passed in values for min and max as a starting point.
+    // Find the min and max partition used in previous frame at this location
+    if (cm->frame_type != KEY_FRAME) {
+      MODE_INFO **prev_mi =
+          &cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col];
+      get_sb_partition_size_range(cm, xd, prev_mi, &min_size, &max_size);
+    }
+    // Find the min and max partition sizes used in the left superblock
+    if (left_in_image) {
+      MODE_INFO **left_sb_mi = &mi[-cm->mib_size];
+      get_sb_partition_size_range(cm, xd, left_sb_mi, &min_size, &max_size);
+    }
+    // Find the min and max partition sizes used in the above suprblock.
+    if (above_in_image) {
+      MODE_INFO **above_sb_mi = &mi[-xd->mi_stride * cm->mib_size];
+      get_sb_partition_size_range(cm, xd, above_sb_mi, &min_size, &max_size);
+    }
+
+    // Adjust observed min and max for "relaxed" auto partition case.
+    if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
+      min_size = min_partition_size[min_size];
+      max_size = max_partition_size[max_size];
+    }
+  }
+
+  // Check border cases where max and min from neighbors may not be legal.
+  max_size = find_partition_size(max_size, mi_rows_remaining, mi_cols_remaining,
+                                 &bh, &bw);
+  min_size = AOMMIN(min_size, max_size);
+
+  // Test for blocks at the edge of the active image.
+  // This may be the actual edge of the image or where there are formatting
+  // bars.
+  if (av1_active_edge_sb(cpi, mi_row, mi_col)) {
+    min_size = BLOCK_4X4;
+  } else {
+    min_size = AOMMIN(cpi->sf.rd_auto_partition_min_limit, min_size);
+  }
+
+  // When use_square_partition_only is true, make sure at least one square
+  // partition is allowed by selecting the next smaller square size as
+  // *min_block_size.
+  if (cpi->sf.use_square_partition_only) {
+    min_size = AOMMIN(min_size, next_square_size[max_size]);
+  }
+
+  *min_block_size = AOMMIN(min_size, cm->sb_size);
+  *max_block_size = AOMMIN(max_size, cm->sb_size);
+}
+
+// TODO(jingning) refactor functions setting partition search range
+static void set_partition_range(const AV1_COMMON *const cm,
+                                const MACROBLOCKD *const xd, int mi_row,
+                                int mi_col, BLOCK_SIZE bsize,
+                                BLOCK_SIZE *const min_bs,
+                                BLOCK_SIZE *const max_bs) {
+  const int mi_width = mi_size_wide[bsize];
+  const int mi_height = mi_size_high[bsize];
+  int idx, idy;
+
+  const int idx_str = cm->mi_stride * mi_row + mi_col;
+  MODE_INFO **const prev_mi = &cm->prev_mi_grid_visible[idx_str];
+  BLOCK_SIZE min_size = BLOCK_64X64;  // default values
+  BLOCK_SIZE max_size = BLOCK_4X4;
+
+  if (prev_mi) {
+    for (idy = 0; idy < mi_height; ++idy) {
+      for (idx = 0; idx < mi_width; ++idx) {
+        const MODE_INFO *const mi = prev_mi[idy * cm->mi_stride + idx];
+        const BLOCK_SIZE bs = mi ? mi->mbmi.sb_type : bsize;
+        min_size = AOMMIN(min_size, bs);
+        max_size = AOMMAX(max_size, bs);
+      }
+    }
+  }
+
+  if (xd->left_available) {
+    for (idy = 0; idy < mi_height; ++idy) {
+      const MODE_INFO *const mi = xd->mi[idy * cm->mi_stride - 1];
+      const BLOCK_SIZE bs = mi ? mi->mbmi.sb_type : bsize;
+      min_size = AOMMIN(min_size, bs);
+      max_size = AOMMAX(max_size, bs);
+    }
+  }
+
+  if (xd->up_available) {
+    for (idx = 0; idx < mi_width; ++idx) {
+      const MODE_INFO *const mi = xd->mi[idx - cm->mi_stride];
+      const BLOCK_SIZE bs = mi ? mi->mbmi.sb_type : bsize;
+      min_size = AOMMIN(min_size, bs);
+      max_size = AOMMAX(max_size, bs);
+    }
+  }
+
+  if (min_size == max_size) {
+    min_size = min_partition_size[min_size];
+    max_size = max_partition_size[max_size];
+  }
+
+  *min_bs = AOMMIN(min_size, cm->sb_size);
+  *max_bs = AOMMIN(max_size, cm->sb_size);
+}
+
+static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
+  memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
+}
+
+static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
+  memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
+}
+
+#if CONFIG_FP_MB_STATS
+const int qindex_skip_threshold_lookup[BLOCK_SIZES] = {
+  0,
+  10,
+  10,
+  30,
+  40,
+  40,
+  60,
+  80,
+  80,
+  90,
+  100,
+  100,
+  120,
+#if CONFIG_EXT_PARTITION
+  // TODO(debargha): What are the correct numbers here?
+  130,
+  130,
+  150
+#endif  // CONFIG_EXT_PARTITION
+};
+const int qindex_split_threshold_lookup[BLOCK_SIZES] = {
+  0,
+  3,
+  3,
+  7,
+  15,
+  15,
+  30,
+  40,
+  40,
+  60,
+  80,
+  80,
+  120,
+#if CONFIG_EXT_PARTITION
+  // TODO(debargha): What are the correct numbers here?
+  160,
+  160,
+  240
+#endif  // CONFIG_EXT_PARTITION
+};
+const int complexity_16x16_blocks_threshold[BLOCK_SIZES] = {
+  1,
+  1,
+  1,
+  1,
+  1,
+  1,
+  1,
+  1,
+  1,
+  1,
+  4,
+  4,
+  6,
+#if CONFIG_EXT_PARTITION
+  // TODO(debargha): What are the correct numbers here?
+  8,
+  8,
+  10
+#endif  // CONFIG_EXT_PARTITION
+};
+
+typedef enum {
+  MV_ZERO = 0,
+  MV_LEFT = 1,
+  MV_UP = 2,
+  MV_RIGHT = 3,
+  MV_DOWN = 4,
+  MV_INVALID
+} MOTION_DIRECTION;
+
+static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
+  if (fp_byte & FPMB_MOTION_ZERO_MASK) {
+    return MV_ZERO;
+  } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
+    return MV_LEFT;
+  } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
+    return MV_RIGHT;
+  } else if (fp_byte & FPMB_MOTION_UP_MASK) {
+    return MV_UP;
+  } else {
+    return MV_DOWN;
+  }
+}
+
+static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
+                                           MOTION_DIRECTION that_mv) {
+  if (this_mv == that_mv) {
+    return 0;
+  } else {
+    return abs(this_mv - that_mv) == 2 ? 2 : 1;
+  }
+}
+#endif
+
+#if CONFIG_EXT_PARTITION_TYPES
+static void rd_test_partition3(
+    const AV1_COMP *const cpi, ThreadData *td, TileDataEnc *tile_data,
+    TOKENEXTRA **tp, PC_TREE *pc_tree, RD_STATS *best_rdc,
+    PICK_MODE_CONTEXT ctxs[3], PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col,
+    BLOCK_SIZE bsize, PARTITION_TYPE partition,
+#if CONFIG_SUPERTX
+    int64_t best_rd, int *best_rate_nocoef, RD_SEARCH_MACROBLOCK_CONTEXT *x_ctx,
+#endif
+    int mi_row0, int mi_col0, BLOCK_SIZE subsize0, int mi_row1, int mi_col1,
+    BLOCK_SIZE subsize1, int mi_row2, int mi_col2, BLOCK_SIZE subsize2) {
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  RD_STATS this_rdc, sum_rdc;
+#if CONFIG_SUPERTX
+  const AV1_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  int this_rate_nocoef, sum_rate_nocoef;
+  int abort_flag;
+  const int supertx_allowed = !frame_is_intra_only(cm) &&
+                              bsize <= MAX_SUPERTX_BLOCK_SIZE &&
+                              !xd->lossless[0];
+#endif
+  if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
+
+  rd_pick_sb_modes(cpi, tile_data, x, mi_row0, mi_col0, &sum_rdc,
+#if CONFIG_SUPERTX
+                   &sum_rate_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                   partition,
+#endif
+                   subsize0, &ctxs[0], best_rdc->rdcost);
+#if CONFIG_SUPERTX
+  abort_flag = sum_rdc.rdcost >= best_rd;
+#endif
+
+#if CONFIG_SUPERTX
+  if (sum_rdc.rdcost < INT64_MAX) {
+#else
+  if (sum_rdc.rdcost < best_rdc->rdcost) {
+#endif
+    PICK_MODE_CONTEXT *ctx_0 = &ctxs[0];
+    update_state(cpi, td, ctx_0, mi_row0, mi_col0, subsize0, 1);
+    encode_superblock(cpi, td, tp, DRY_RUN_NORMAL, mi_row0, mi_col0, subsize0,
+                      ctx_0, NULL);
+
+    if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_0);
+
+#if CONFIG_SUPERTX
+    rd_pick_sb_modes(cpi, tile_data, x, mi_row1, mi_col1, &this_rdc,
+                     &this_rate_nocoef,
+#if CONFIG_EXT_PARTITION_TYPES
+                     partition,
+#endif
+                     subsize1, &ctxs[1], INT64_MAX - sum_rdc.rdcost);
+#else
+    rd_pick_sb_modes(cpi, tile_data, x, mi_row1, mi_col1, &this_rdc,
+#if CONFIG_EXT_PARTITION_TYPES
+                     partition,
+#endif
+                     subsize1, &ctxs[1], best_rdc->rdcost - sum_rdc.rdcost);
+#endif  // CONFIG_SUPERTX
+
+    if (this_rdc.rate == INT_MAX) {
+      sum_rdc.rdcost = INT64_MAX;
+#if CONFIG_SUPERTX
+      sum_rate_nocoef = INT_MAX;
+#endif
+    } else {
+      sum_rdc.rate += this_rdc.rate;
+      sum_rdc.dist += this_rdc.dist;
+      sum_rdc.rdcost += this_rdc.rdcost;
+#if CONFIG_SUPERTX
+      sum_rate_nocoef += this_rate_nocoef;
+#endif
+    }
+
+#if CONFIG_SUPERTX
+    if (sum_rdc.rdcost < INT64_MAX) {
+#else
+    if (sum_rdc.rdcost < best_rdc->rdcost) {
+#endif
+      PICK_MODE_CONTEXT *ctx_1 = &ctxs[1];
+      update_state(cpi, td, ctx_1, mi_row1, mi_col1, subsize1, 1);
+      encode_superblock(cpi, td, tp, DRY_RUN_NORMAL, mi_row1, mi_col1, subsize1,
+                        ctx_1, NULL);
+
+      if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_1);
+
+#if CONFIG_SUPERTX
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row2, mi_col2, &this_rdc,
+                       &this_rate_nocoef,
+#if CONFIG_EXT_PARTITION_TYPES
+                       partition,
+#endif
+                       subsize2, &ctxs[2], INT64_MAX - sum_rdc.rdcost);
+#else
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row2, mi_col2, &this_rdc,
+#if CONFIG_EXT_PARTITION_TYPES
+                       partition,
+#endif
+                       subsize2, &ctxs[2], best_rdc->rdcost - sum_rdc.rdcost);
+#endif  // CONFIG_SUPERTX
+
+      if (this_rdc.rate == INT_MAX) {
+        sum_rdc.rdcost = INT64_MAX;
+#if CONFIG_SUPERTX
+        sum_rate_nocoef = INT_MAX;
+#endif
+      } else {
+        sum_rdc.rate += this_rdc.rate;
+        sum_rdc.dist += this_rdc.dist;
+        sum_rdc.rdcost += this_rdc.rdcost;
+#if CONFIG_SUPERTX
+        sum_rate_nocoef += this_rate_nocoef;
+#endif
+      }
+
+#if CONFIG_SUPERTX
+      if (supertx_allowed && !abort_flag && sum_rdc.rdcost < INT64_MAX) {
+        TX_SIZE supertx_size = max_txsize_lookup[bsize];
+        const PARTITION_TYPE best_partition = pc_tree->partitioning;
+        pc_tree->partitioning = partition;
+        sum_rdc.rate += av1_cost_bit(
+            cm->fc->supertx_prob[partition_supertx_context_lookup[partition]]
+                                [supertx_size],
+            0);
+        sum_rdc.rdcost =
+            RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+
+        if (!check_intra_sb(cpi, tile_info, mi_row, mi_col, bsize, pc_tree)) {
+          TX_TYPE best_tx = DCT_DCT;
+          RD_STATS tmp_rdc = { sum_rate_nocoef, 0, 0 };
+
+          restore_context(x, x_ctx, mi_row, mi_col, bsize);
+
+          rd_supertx_sb(cpi, td, tile_info, mi_row, mi_col, bsize,
+                        &tmp_rdc.rate, &tmp_rdc.dist, &best_tx, pc_tree);
+
+          tmp_rdc.rate += av1_cost_bit(
+              cm->fc->supertx_prob[partition_supertx_context_lookup[partition]]
+                                  [supertx_size],
+              1);
+          tmp_rdc.rdcost =
+              RDCOST(x->rdmult, x->rddiv, tmp_rdc.rate, tmp_rdc.dist);
+          if (tmp_rdc.rdcost < sum_rdc.rdcost) {
+            sum_rdc = tmp_rdc;
+            update_supertx_param_sb(cpi, td, mi_row, mi_col, bsize, best_tx,
+                                    supertx_size, pc_tree);
+          }
+        }
+
+        pc_tree->partitioning = best_partition;
+      }
+#endif  // CONFIG_SUPERTX
+
+      if (sum_rdc.rdcost < best_rdc->rdcost) {
+        int pl = partition_plane_context(xd, mi_row, mi_col,
+#if CONFIG_UNPOISON_PARTITION_CTX
+                                         has_rows, has_cols,
+#endif
+                                         bsize);
+        sum_rdc.rate += cpi->partition_cost[pl][partition];
+        sum_rdc.rdcost =
+            RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+#if CONFIG_SUPERTX
+        sum_rate_nocoef += cpi->partition_cost[pl][partition];
+#endif
+        if (sum_rdc.rdcost < best_rdc->rdcost) {
+#if CONFIG_SUPERTX
+          *best_rate_nocoef = sum_rate_nocoef;
+          assert(*best_rate_nocoef >= 0);
+#endif
+          *best_rdc = sum_rdc;
+          pc_tree->partitioning = partition;
+        }
+      }
+    }
+  }
+}
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+// TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
+// unlikely to be selected depending on previous rate-distortion optimization
+// results, for encoding speed-up.
+static void rd_pick_partition(const AV1_COMP *const cpi, ThreadData *td,
+                              TileDataEnc *tile_data, TOKENEXTRA **tp,
+                              int mi_row, int mi_col, BLOCK_SIZE bsize,
+                              RD_STATS *rd_cost,
+#if CONFIG_SUPERTX
+                              int *rate_nocoef,
+#endif
+                              int64_t best_rd, PC_TREE *pc_tree) {
+  const AV1_COMMON *const cm = &cpi->common;
+  TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int mi_step = mi_size_wide[bsize] / 2;
+  RD_SEARCH_MACROBLOCK_CONTEXT x_ctx;
+  const TOKENEXTRA *const tp_orig = *tp;
+  PICK_MODE_CONTEXT *ctx_none = &pc_tree->none;
+#if CONFIG_UNPOISON_PARTITION_CTX
+  const int hbs = mi_size_wide[bsize] / 2;
+  const int has_rows = mi_row + hbs < cm->mi_rows;
+  const int has_cols = mi_col + hbs < cm->mi_cols;
+#else
+  int tmp_partition_cost[PARTITION_TYPES];
+#endif
+  BLOCK_SIZE subsize;
+  RD_STATS this_rdc, sum_rdc, best_rdc;
+  const int bsize_at_least_8x8 = (bsize >= BLOCK_8X8);
+  int do_square_split = bsize_at_least_8x8;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+  const int pl = bsize_at_least_8x8
+                     ? partition_plane_context(xd, mi_row, mi_col,
+#if CONFIG_UNPOISON_PARTITION_CTX
+                                               has_rows, has_cols,
+#endif
+                                               bsize)
+                     : 0;
+#else
+  const int unify_bsize = 0;
+  const int pl = partition_plane_context(xd, mi_row, mi_col,
+#if CONFIG_UNPOISON_PARTITION_CTX
+                                         has_rows, has_cols,
+#endif
+                                         bsize);
+#endif  // CONFIG_CB4X4
+  const int *partition_cost = cpi->partition_cost[pl];
+#if CONFIG_SUPERTX
+  int this_rate_nocoef, sum_rate_nocoef = 0, best_rate_nocoef = INT_MAX;
+  int abort_flag;
+  const int supertx_allowed = !frame_is_intra_only(cm) &&
+                              bsize <= MAX_SUPERTX_BLOCK_SIZE &&
+                              !xd->lossless[0];
+#endif  // CONFIG_SUPERTX
+
+  int do_rectangular_split = 1;
+#if CONFIG_EXT_PARTITION_TYPES
+  BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+
+  // Override skipping rectangular partition operations for edge blocks
+  const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
+  const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
+  const int xss = x->e_mbd.plane[1].subsampling_x;
+  const int yss = x->e_mbd.plane[1].subsampling_y;
+
+  BLOCK_SIZE min_size = x->min_partition_size;
+  BLOCK_SIZE max_size = x->max_partition_size;
+
+#if CONFIG_FP_MB_STATS
+  unsigned int src_diff_var = UINT_MAX;
+  int none_complexity = 0;
+#endif
+
+  int partition_none_allowed = !force_horz_split && !force_vert_split;
+  int partition_horz_allowed =
+      !force_vert_split && yss <= xss && bsize_at_least_8x8;
+  int partition_vert_allowed =
+      !force_horz_split && xss <= yss && bsize_at_least_8x8;
+
+#if CONFIG_PVQ
+  od_rollback_buffer pre_rdo_buf;
+#endif
+
+  (void)*tp_orig;
+
+#if !CONFIG_UNPOISON_PARTITION_CTX
+  if (force_horz_split || force_vert_split) {
+    tmp_partition_cost[PARTITION_NONE] = INT_MAX;
+
+    if (!force_vert_split) {  // force_horz_split only
+      tmp_partition_cost[PARTITION_VERT] = INT_MAX;
+      tmp_partition_cost[PARTITION_HORZ] =
+          av1_cost_bit(cm->fc->partition_prob[pl][PARTITION_HORZ], 0);
+      tmp_partition_cost[PARTITION_SPLIT] =
+          av1_cost_bit(cm->fc->partition_prob[pl][PARTITION_HORZ], 1);
+    } else if (!force_horz_split) {  // force_vert_split only
+      tmp_partition_cost[PARTITION_HORZ] = INT_MAX;
+      tmp_partition_cost[PARTITION_VERT] =
+          av1_cost_bit(cm->fc->partition_prob[pl][PARTITION_VERT], 0);
+      tmp_partition_cost[PARTITION_SPLIT] =
+          av1_cost_bit(cm->fc->partition_prob[pl][PARTITION_VERT], 1);
+    } else {  // force_ horz_split && force_vert_split horz_split
+      tmp_partition_cost[PARTITION_HORZ] = INT_MAX;
+      tmp_partition_cost[PARTITION_VERT] = INT_MAX;
+      tmp_partition_cost[PARTITION_SPLIT] = 0;
+    }
+
+    partition_cost = tmp_partition_cost;
+  }
+#endif
+
+#if CONFIG_VAR_TX
+#ifndef NDEBUG
+  // Nothing should rely on the default value of this array (which is just
+  // leftover from encoding the previous block. Setting it to magic number
+  // when debugging.
+  memset(x->blk_skip[0], 234, sizeof(x->blk_skip[0]));
+#endif  // NDEBUG
+#endif  // CONFIG_VAR_TX
+
+  assert(mi_size_wide[bsize] == mi_size_high[bsize]);
+
+  av1_init_rd_stats(&this_rdc);
+  av1_init_rd_stats(&sum_rdc);
+  av1_invalid_rd_stats(&best_rdc);
+  best_rdc.rdcost = best_rd;
+
+  set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+
+  if (bsize == BLOCK_16X16 && cpi->vaq_refresh)
+    x->mb_energy = av1_block_energy(cpi, x, bsize);
+
+  if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
+    const int cb_partition_search_ctrl =
+        ((pc_tree->index == 0 || pc_tree->index == 3) +
+         get_chessboard_index(cm->current_video_frame)) &
+        0x1;
+
+    if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
+      set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
+  }
+
+  // Determine partition types in search according to the speed features.
+  // The threshold set here has to be of square block size.
+  if (cpi->sf.auto_min_max_partition_size) {
+    const int no_partition_allowed = (bsize <= max_size && bsize >= min_size);
+    // Note: Further partitioning is NOT allowed when bsize == min_size already.
+    const int partition_allowed = (bsize <= max_size && bsize > min_size);
+    partition_none_allowed &= no_partition_allowed;
+    partition_horz_allowed &= partition_allowed || force_horz_split;
+    partition_vert_allowed &= partition_allowed || force_vert_split;
+    do_square_split &= bsize > min_size;
+  }
+  if (cpi->sf.use_square_partition_only) {
+    partition_horz_allowed &= force_horz_split;
+    partition_vert_allowed &= force_vert_split;
+  }
+
+#if CONFIG_VAR_TX
+  xd->above_txfm_context = cm->above_txfm_context + mi_col;
+  xd->left_txfm_context =
+      xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+#endif
+#if !CONFIG_PVQ
+  save_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+  save_context(x, &x_ctx, mi_row, mi_col, &pre_rdo_buf, bsize);
+#endif
+
+#if CONFIG_FP_MB_STATS
+  if (cpi->use_fp_mb_stats) {
+    set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+    src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src, mi_row,
+                                                  mi_col, bsize);
+  }
+#endif
+
+#if CONFIG_FP_MB_STATS
+  // Decide whether we shall split directly and skip searching NONE by using
+  // the first pass block statistics
+  if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_square_split &&
+      partition_none_allowed && src_diff_var > 4 &&
+      cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
+    int mb_row = mi_row >> 1;
+    int mb_col = mi_col >> 1;
+    int mb_row_end =
+        AOMMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
+    int mb_col_end =
+        AOMMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
+    int r, c;
+
+    // compute a complexity measure, basically measure inconsistency of motion
+    // vectors obtained from the first pass in the current block
+    for (r = mb_row; r < mb_row_end; r++) {
+      for (c = mb_col; c < mb_col_end; c++) {
+        const int mb_index = r * cm->mb_cols + c;
+
+        MOTION_DIRECTION this_mv;
+        MOTION_DIRECTION right_mv;
+        MOTION_DIRECTION bottom_mv;
+
+        this_mv =
+            get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
+
+        // to its right
+        if (c != mb_col_end - 1) {
+          right_mv = get_motion_direction_fp(
+              cpi->twopass.this_frame_mb_stats[mb_index + 1]);
+          none_complexity += get_motion_inconsistency(this_mv, right_mv);
+        }
+
+        // to its bottom
+        if (r != mb_row_end - 1) {
+          bottom_mv = get_motion_direction_fp(
+              cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
+          none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
+        }
+
+        // do not count its left and top neighbors to avoid double counting
+      }
+    }
+
+    if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
+      partition_none_allowed = 0;
+    }
+  }
+#endif
+
+  // PARTITION_NONE
+  if (partition_none_allowed) {
+    rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc,
+#if CONFIG_SUPERTX
+                     &this_rate_nocoef,
+#endif
+#if CONFIG_EXT_PARTITION_TYPES
+                     PARTITION_NONE,
+#endif
+                     bsize, ctx_none, best_rdc.rdcost);
+    if (this_rdc.rate != INT_MAX) {
+      if (bsize_at_least_8x8) {
+        this_rdc.rate += partition_cost[PARTITION_NONE];
+        this_rdc.rdcost =
+            RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
+#if CONFIG_SUPERTX
+        this_rate_nocoef += partition_cost[PARTITION_NONE];
+#endif
+      }
+
+      if (this_rdc.rdcost < best_rdc.rdcost) {
+        // Adjust dist breakout threshold according to the partition size.
+        const int64_t dist_breakout_thr =
+            cpi->sf.partition_search_breakout_dist_thr >>
+            ((2 * (MAX_SB_SIZE_LOG2 - 2)) -
+             (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]));
+        const int rate_breakout_thr =
+            cpi->sf.partition_search_breakout_rate_thr *
+            num_pels_log2_lookup[bsize];
+
+        best_rdc = this_rdc;
+#if CONFIG_SUPERTX
+        best_rate_nocoef = this_rate_nocoef;
+        assert(best_rate_nocoef >= 0);
+#endif
+        if (bsize_at_least_8x8) pc_tree->partitioning = PARTITION_NONE;
+
+        // If all y, u, v transform blocks in this partition are skippable, and
+        // the dist & rate are within the thresholds, the partition search is
+        // terminated for current branch of the partition search tree.
+        // The dist & rate thresholds are set to 0 at speed 0 to disable the
+        // early termination at that speed.
+        if (!x->e_mbd.lossless[xd->mi[0]->mbmi.segment_id] &&
+            (ctx_none->skippable && best_rdc.dist < dist_breakout_thr &&
+             best_rdc.rate < rate_breakout_thr)) {
+          do_square_split = 0;
+          do_rectangular_split = 0;
+        }
+
+#if CONFIG_FP_MB_STATS
+        // Check if every 16x16 first pass block statistics has zero
+        // motion and the corresponding first pass residue is small enough.
+        // If that is the case, check the difference variance between the
+        // current frame and the last frame. If the variance is small enough,
+        // stop further splitting in RD optimization
+        if (cpi->use_fp_mb_stats && do_square_split &&
+            cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
+          int mb_row = mi_row >> 1;
+          int mb_col = mi_col >> 1;
+          int mb_row_end =
+              AOMMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
+          int mb_col_end =
+              AOMMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
+          int r, c;
+
+          int skip = 1;
+          for (r = mb_row; r < mb_row_end; r++) {
+            for (c = mb_col; c < mb_col_end; c++) {
+              const int mb_index = r * cm->mb_cols + c;
+              if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
+                    FPMB_MOTION_ZERO_MASK) ||
+                  !(cpi->twopass.this_frame_mb_stats[mb_index] &
+                    FPMB_ERROR_SMALL_MASK)) {
+                skip = 0;
+                break;
+              }
+            }
+            if (skip == 0) {
+              break;
+            }
+          }
+          if (skip) {
+            if (src_diff_var == UINT_MAX) {
+              set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
+              src_diff_var = get_sby_perpixel_diff_variance(
+                  cpi, &x->plane[0].src, mi_row, mi_col, bsize);
+            }
+            if (src_diff_var < 8) {
+              do_square_split = 0;
+              do_rectangular_split = 0;
+            }
+          }
+        }
+#endif
+      }
+    }
+#if !CONFIG_PVQ
+    restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+    restore_context(x, &x_ctx, mi_row, mi_col, &pre_rdo_buf, bsize);
+#endif
+  }
+
+  // store estimated motion vector
+  if (cpi->sf.adaptive_motion_search) store_pred_mv(x, ctx_none);
+
+  // PARTITION_SPLIT
+  // TODO(jingning): use the motion vectors given by the above search as
+  // the starting point of motion search in the following partition type check.
+  if (do_square_split) {
+    int reached_last_index = 0;
+    subsize = get_subsize(bsize, PARTITION_SPLIT);
+    if (bsize == BLOCK_8X8 && !unify_bsize) {
+#if CONFIG_DUAL_FILTER
+      if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
+        pc_tree->leaf_split[0]->pred_interp_filter =
+            ctx_none->mic.mbmi.interp_filter[0];
+#else
+      if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
+        pc_tree->leaf_split[0]->pred_interp_filter =
+            ctx_none->mic.mbmi.interp_filter;
+#endif
+#if CONFIG_SUPERTX
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc,
+                       &sum_rate_nocoef,
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_SPLIT,
+#endif
+                       subsize, pc_tree->leaf_split[0], INT64_MAX);
+#else
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc,
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_SPLIT,
+#endif
+                       subsize, pc_tree->leaf_split[0], best_rdc.rdcost);
+#endif  // CONFIG_SUPERTX
+      if (sum_rdc.rate == INT_MAX) {
+        sum_rdc.rdcost = INT64_MAX;
+#if CONFIG_SUPERTX
+        sum_rate_nocoef = INT_MAX;
+#endif
+      }
+#if CONFIG_SUPERTX
+      if (supertx_allowed && sum_rdc.rdcost < INT64_MAX) {
+        TX_SIZE supertx_size = max_txsize_lookup[bsize];
+        const PARTITION_TYPE best_partition = pc_tree->partitioning;
+
+        pc_tree->partitioning = PARTITION_SPLIT;
+
+        sum_rdc.rate += av1_cost_bit(
+            cm->fc->supertx_prob[partition_supertx_context_lookup
+                                     [PARTITION_SPLIT]][supertx_size],
+            0);
+        sum_rdc.rdcost =
+            RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+
+        if (is_inter_mode(pc_tree->leaf_split[0]->mic.mbmi.mode)) {
+          TX_TYPE best_tx = DCT_DCT;
+          RD_STATS tmp_rdc;
+          av1_init_rd_stats(&tmp_rdc);
+          tmp_rdc.rate = sum_rate_nocoef;
+
+          restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+
+          rd_supertx_sb(cpi, td, tile_info, mi_row, mi_col, bsize,
+                        &tmp_rdc.rate, &tmp_rdc.dist, &best_tx, pc_tree);
+
+          tmp_rdc.rate += av1_cost_bit(
+              cm->fc->supertx_prob[partition_supertx_context_lookup
+                                       [PARTITION_SPLIT]][supertx_size],
+              1);
+          tmp_rdc.rdcost =
+              RDCOST(x->rdmult, x->rddiv, tmp_rdc.rate, tmp_rdc.dist);
+          if (tmp_rdc.rdcost < sum_rdc.rdcost) {
+            sum_rdc = tmp_rdc;
+            update_supertx_param_sb(cpi, td, mi_row, mi_col, bsize, best_tx,
+                                    supertx_size, pc_tree);
+          }
+        }
+
+        pc_tree->partitioning = best_partition;
+      }
+#endif  // CONFIG_SUPERTX
+      reached_last_index = 1;
+    } else {
+      int idx;
+#if CONFIG_SUPERTX
+      for (idx = 0; idx < 4 && sum_rdc.rdcost < INT64_MAX; ++idx) {
+#else
+      for (idx = 0; idx < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++idx) {
+#endif  // CONFIG_SUPERTX
+        const int x_idx = (idx & 1) * mi_step;
+        const int y_idx = (idx >> 1) * mi_step;
+
+        if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
+          continue;
+
+        if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_none);
+
+        pc_tree->split[idx]->index = idx;
+#if CONFIG_SUPERTX
+        rd_pick_partition(cpi, td, tile_data, tp, mi_row + y_idx,
+                          mi_col + x_idx, subsize, &this_rdc, &this_rate_nocoef,
+                          INT64_MAX - sum_rdc.rdcost, pc_tree->split[idx]);
+#else
+        rd_pick_partition(
+            cpi, td, tile_data, tp, mi_row + y_idx, mi_col + x_idx, subsize,
+            &this_rdc, best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[idx]);
+#endif  // CONFIG_SUPERTX
+
+        if (this_rdc.rate == INT_MAX) {
+          sum_rdc.rdcost = INT64_MAX;
+#if CONFIG_SUPERTX
+          sum_rate_nocoef = INT_MAX;
+#endif  // CONFIG_SUPERTX
+          break;
+        } else {
+          sum_rdc.rate += this_rdc.rate;
+          sum_rdc.dist += this_rdc.dist;
+          sum_rdc.rdcost += this_rdc.rdcost;
+#if CONFIG_SUPERTX
+          sum_rate_nocoef += this_rate_nocoef;
+#endif  // CONFIG_SUPERTX
+        }
+      }
+      reached_last_index = (idx == 4);
+#if CONFIG_SUPERTX
+      if (supertx_allowed && sum_rdc.rdcost < INT64_MAX && reached_last_index) {
+        TX_SIZE supertx_size = max_txsize_lookup[bsize];
+        const PARTITION_TYPE best_partition = pc_tree->partitioning;
+
+        pc_tree->partitioning = PARTITION_SPLIT;
+
+        sum_rdc.rate += av1_cost_bit(
+            cm->fc->supertx_prob[partition_supertx_context_lookup
+                                     [PARTITION_SPLIT]][supertx_size],
+            0);
+        sum_rdc.rdcost =
+            RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+
+        if (!check_intra_sb(cpi, tile_info, mi_row, mi_col, bsize, pc_tree)) {
+          TX_TYPE best_tx = DCT_DCT;
+          RD_STATS tmp_rdc;
+          av1_init_rd_stats(&tmp_rdc);
+          tmp_rdc.rate = sum_rate_nocoef;
+
+          restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+
+          rd_supertx_sb(cpi, td, tile_info, mi_row, mi_col, bsize,
+                        &tmp_rdc.rate, &tmp_rdc.dist, &best_tx, pc_tree);
+
+          tmp_rdc.rate += av1_cost_bit(
+              cm->fc->supertx_prob[partition_supertx_context_lookup
+                                       [PARTITION_SPLIT]][supertx_size],
+              1);
+          tmp_rdc.rdcost =
+              RDCOST(x->rdmult, x->rddiv, tmp_rdc.rate, tmp_rdc.dist);
+          if (tmp_rdc.rdcost < sum_rdc.rdcost) {
+            sum_rdc = tmp_rdc;
+            update_supertx_param_sb(cpi, td, mi_row, mi_col, bsize, best_tx,
+                                    supertx_size, pc_tree);
+          }
+        }
+
+        pc_tree->partitioning = best_partition;
+      }
+#endif  // CONFIG_SUPERTX
+    }
+
+    if (reached_last_index && sum_rdc.rdcost < best_rdc.rdcost) {
+      sum_rdc.rate += partition_cost[PARTITION_SPLIT];
+      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+#if CONFIG_SUPERTX
+      sum_rate_nocoef += partition_cost[PARTITION_SPLIT];
+#endif  // CONFIG_SUPERTX
+
+      if (sum_rdc.rdcost < best_rdc.rdcost) {
+        best_rdc = sum_rdc;
+#if CONFIG_SUPERTX
+        best_rate_nocoef = sum_rate_nocoef;
+        assert(best_rate_nocoef >= 0);
+#endif  // CONFIG_SUPERTX
+        pc_tree->partitioning = PARTITION_SPLIT;
+      }
+    } else if (cpi->sf.less_rectangular_check) {
+      // skip rectangular partition test when larger block size
+      // gives better rd cost
+      do_rectangular_split &= !partition_none_allowed;
+    }
+#if !CONFIG_PVQ
+    restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+    restore_context(x, &x_ctx, mi_row, mi_col, &pre_rdo_buf, bsize);
+#endif
+  }  // if (do_split)
+
+  // PARTITION_HORZ
+  if (partition_horz_allowed &&
+      (do_rectangular_split || av1_active_h_edge(cpi, mi_row, mi_step))) {
+    subsize = get_subsize(bsize, PARTITION_HORZ);
+    if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_none);
+#if CONFIG_DUAL_FILTER
+    if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+        partition_none_allowed)
+      pc_tree->horizontal[0].pred_interp_filter =
+          ctx_none->mic.mbmi.interp_filter[0];
+#else
+    if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+        partition_none_allowed)
+      pc_tree->horizontal[0].pred_interp_filter =
+          ctx_none->mic.mbmi.interp_filter;
+#endif
+    rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc,
+#if CONFIG_SUPERTX
+                     &sum_rate_nocoef,
+#endif  // CONFIG_SUPERTX
+#if CONFIG_EXT_PARTITION_TYPES
+                     PARTITION_HORZ,
+#endif
+                     subsize, &pc_tree->horizontal[0], best_rdc.rdcost);
+
+#if CONFIG_SUPERTX
+    abort_flag =
+        (sum_rdc.rdcost >= best_rd && (bsize > BLOCK_8X8 || unify_bsize)) ||
+        (sum_rdc.rate == INT_MAX && bsize == BLOCK_8X8);
+    if (sum_rdc.rdcost < INT64_MAX &&
+#else
+    if (sum_rdc.rdcost < best_rdc.rdcost &&
+#endif  // CONFIG_SUPERTX
+        !force_horz_split && (bsize > BLOCK_8X8 || unify_bsize)) {
+      PICK_MODE_CONTEXT *ctx_h = &pc_tree->horizontal[0];
+      update_state(cpi, td, ctx_h, mi_row, mi_col, subsize, 1);
+      encode_superblock(cpi, td, tp, DRY_RUN_NORMAL, mi_row, mi_col, subsize,
+                        ctx_h, NULL);
+
+      if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_h);
+
+#if CONFIG_DUAL_FILTER
+      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+          partition_none_allowed)
+        pc_tree->horizontal[1].pred_interp_filter =
+            ctx_h->mic.mbmi.interp_filter[0];
+#else
+      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+          partition_none_allowed)
+        pc_tree->horizontal[1].pred_interp_filter =
+            ctx_none->mic.mbmi.interp_filter;
+#endif
+#if CONFIG_SUPERTX
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col, &this_rdc,
+                       &this_rate_nocoef,
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_HORZ,
+#endif
+                       subsize, &pc_tree->horizontal[1], INT64_MAX);
+#else
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col, &this_rdc,
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_HORZ,
+#endif
+                       subsize, &pc_tree->horizontal[1],
+                       best_rdc.rdcost - sum_rdc.rdcost);
+#endif  // CONFIG_SUPERTX
+      if (this_rdc.rate == INT_MAX) {
+        sum_rdc.rdcost = INT64_MAX;
+#if CONFIG_SUPERTX
+        sum_rate_nocoef = INT_MAX;
+#endif  // CONFIG_SUPERTX
+      } else {
+        sum_rdc.rate += this_rdc.rate;
+        sum_rdc.dist += this_rdc.dist;
+        sum_rdc.rdcost += this_rdc.rdcost;
+#if CONFIG_SUPERTX
+        sum_rate_nocoef += this_rate_nocoef;
+#endif  // CONFIG_SUPERTX
+      }
+    }
+
+#if CONFIG_SUPERTX
+    if (supertx_allowed && sum_rdc.rdcost < INT64_MAX && !abort_flag) {
+      TX_SIZE supertx_size = max_txsize_lookup[bsize];
+      const PARTITION_TYPE best_partition = pc_tree->partitioning;
+
+      pc_tree->partitioning = PARTITION_HORZ;
+
+      sum_rdc.rate += av1_cost_bit(
+          cm->fc->supertx_prob[partition_supertx_context_lookup[PARTITION_HORZ]]
+                              [supertx_size],
+          0);
+      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+
+      if (!check_intra_sb(cpi, tile_info, mi_row, mi_col, bsize, pc_tree)) {
+        TX_TYPE best_tx = DCT_DCT;
+        RD_STATS tmp_rdc;
+        av1_init_rd_stats(&tmp_rdc);
+        tmp_rdc.rate = sum_rate_nocoef;
+
+        restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+
+        rd_supertx_sb(cpi, td, tile_info, mi_row, mi_col, bsize, &tmp_rdc.rate,
+                      &tmp_rdc.dist, &best_tx, pc_tree);
+
+        tmp_rdc.rate += av1_cost_bit(
+            cm->fc
+                ->supertx_prob[partition_supertx_context_lookup[PARTITION_HORZ]]
+                              [supertx_size],
+            1);
+        tmp_rdc.rdcost =
+            RDCOST(x->rdmult, x->rddiv, tmp_rdc.rate, tmp_rdc.dist);
+        if (tmp_rdc.rdcost < sum_rdc.rdcost) {
+          sum_rdc = tmp_rdc;
+          update_supertx_param_sb(cpi, td, mi_row, mi_col, bsize, best_tx,
+                                  supertx_size, pc_tree);
+        }
+      }
+
+      pc_tree->partitioning = best_partition;
+    }
+#endif  // CONFIG_SUPERTX
+
+    if (sum_rdc.rdcost < best_rdc.rdcost) {
+      sum_rdc.rate += partition_cost[PARTITION_HORZ];
+      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+#if CONFIG_SUPERTX
+      sum_rate_nocoef += partition_cost[PARTITION_HORZ];
+#endif  // CONFIG_SUPERTX
+      if (sum_rdc.rdcost < best_rdc.rdcost) {
+        best_rdc = sum_rdc;
+#if CONFIG_SUPERTX
+        best_rate_nocoef = sum_rate_nocoef;
+        assert(best_rate_nocoef >= 0);
+#endif  // CONFIG_SUPERTX
+        pc_tree->partitioning = PARTITION_HORZ;
+      }
+    }
+#if !CONFIG_PVQ
+    restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+    restore_context(x, &x_ctx, mi_row, mi_col, &pre_rdo_buf, bsize);
+#endif
+  }
+
+  // PARTITION_VERT
+  if (partition_vert_allowed &&
+      (do_rectangular_split || av1_active_v_edge(cpi, mi_col, mi_step))) {
+    subsize = get_subsize(bsize, PARTITION_VERT);
+
+    if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_none);
+
+#if CONFIG_DUAL_FILTER
+    if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+        partition_none_allowed)
+      pc_tree->vertical[0].pred_interp_filter =
+          ctx_none->mic.mbmi.interp_filter[0];
+#else
+    if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+        partition_none_allowed)
+      pc_tree->vertical[0].pred_interp_filter =
+          ctx_none->mic.mbmi.interp_filter;
+#endif
+    rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc,
+#if CONFIG_SUPERTX
+                     &sum_rate_nocoef,
+#endif  // CONFIG_SUPERTX
+#if CONFIG_EXT_PARTITION_TYPES
+                     PARTITION_VERT,
+#endif
+                     subsize, &pc_tree->vertical[0], best_rdc.rdcost);
+#if CONFIG_SUPERTX
+    abort_flag =
+        (sum_rdc.rdcost >= best_rd && (bsize > BLOCK_8X8 || unify_bsize)) ||
+        (sum_rdc.rate == INT_MAX && bsize == BLOCK_8X8);
+    if (sum_rdc.rdcost < INT64_MAX &&
+#else
+    if (sum_rdc.rdcost < best_rdc.rdcost &&
+#endif  // CONFIG_SUPERTX
+        !force_vert_split && (bsize > BLOCK_8X8 || unify_bsize)) {
+      update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 1);
+      encode_superblock(cpi, td, tp, DRY_RUN_NORMAL, mi_row, mi_col, subsize,
+                        &pc_tree->vertical[0], NULL);
+
+      if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx_none);
+
+#if CONFIG_DUAL_FILTER
+      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+          partition_none_allowed)
+        pc_tree->vertical[1].pred_interp_filter =
+            ctx_none->mic.mbmi.interp_filter[0];
+#else
+      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
+          partition_none_allowed)
+        pc_tree->vertical[1].pred_interp_filter =
+            ctx_none->mic.mbmi.interp_filter;
+#endif
+#if CONFIG_SUPERTX
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step, &this_rdc,
+                       &this_rate_nocoef,
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_VERT,
+#endif
+                       subsize, &pc_tree->vertical[1],
+                       INT64_MAX - sum_rdc.rdcost);
+#else
+      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step, &this_rdc,
+#if CONFIG_EXT_PARTITION_TYPES
+                       PARTITION_VERT,
+#endif
+                       subsize, &pc_tree->vertical[1],
+                       best_rdc.rdcost - sum_rdc.rdcost);
+#endif  // CONFIG_SUPERTX
+      if (this_rdc.rate == INT_MAX) {
+        sum_rdc.rdcost = INT64_MAX;
+#if CONFIG_SUPERTX
+        sum_rate_nocoef = INT_MAX;
+#endif  // CONFIG_SUPERTX
+      } else {
+        sum_rdc.rate += this_rdc.rate;
+        sum_rdc.dist += this_rdc.dist;
+        sum_rdc.rdcost += this_rdc.rdcost;
+#if CONFIG_SUPERTX
+        sum_rate_nocoef += this_rate_nocoef;
+#endif  // CONFIG_SUPERTX
+      }
+    }
+#if CONFIG_SUPERTX
+    if (supertx_allowed && sum_rdc.rdcost < INT64_MAX && !abort_flag) {
+      TX_SIZE supertx_size = max_txsize_lookup[bsize];
+      const PARTITION_TYPE best_partition = pc_tree->partitioning;
+
+      pc_tree->partitioning = PARTITION_VERT;
+
+      sum_rdc.rate += av1_cost_bit(
+          cm->fc->supertx_prob[partition_supertx_context_lookup[PARTITION_VERT]]
+                              [supertx_size],
+          0);
+      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+
+      if (!check_intra_sb(cpi, tile_info, mi_row, mi_col, bsize, pc_tree)) {
+        TX_TYPE best_tx = DCT_DCT;
+        RD_STATS tmp_rdc;
+        av1_init_rd_stats(&tmp_rdc);
+        tmp_rdc.rate = sum_rate_nocoef;
+
+        restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+
+        rd_supertx_sb(cpi, td, tile_info, mi_row, mi_col, bsize, &tmp_rdc.rate,
+                      &tmp_rdc.dist, &best_tx, pc_tree);
+
+        tmp_rdc.rate += av1_cost_bit(
+            cm->fc
+                ->supertx_prob[partition_supertx_context_lookup[PARTITION_VERT]]
+                              [supertx_size],
+            1);
+        tmp_rdc.rdcost =
+            RDCOST(x->rdmult, x->rddiv, tmp_rdc.rate, tmp_rdc.dist);
+        if (tmp_rdc.rdcost < sum_rdc.rdcost) {
+          sum_rdc = tmp_rdc;
+          update_supertx_param_sb(cpi, td, mi_row, mi_col, bsize, best_tx,
+                                  supertx_size, pc_tree);
+        }
+      }
+
+      pc_tree->partitioning = best_partition;
+    }
+#endif  // CONFIG_SUPERTX
+
+    if (sum_rdc.rdcost < best_rdc.rdcost) {
+      sum_rdc.rate += partition_cost[PARTITION_VERT];
+      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
+#if CONFIG_SUPERTX
+      sum_rate_nocoef += partition_cost[PARTITION_VERT];
+#endif  // CONFIG_SUPERTX
+      if (sum_rdc.rdcost < best_rdc.rdcost) {
+        best_rdc = sum_rdc;
+#if CONFIG_SUPERTX
+        best_rate_nocoef = sum_rate_nocoef;
+        assert(best_rate_nocoef >= 0);
+#endif  // CONFIG_SUPERTX
+        pc_tree->partitioning = PARTITION_VERT;
+      }
+    }
+#if !CONFIG_PVQ
+    restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+#else
+    restore_context(x, &x_ctx, mi_row, mi_col, &pre_rdo_buf, bsize);
+#endif
+  }
+
+#if CONFIG_EXT_PARTITION_TYPES
+  // PARTITION_HORZ_A
+  if (partition_horz_allowed && do_rectangular_split && bsize > BLOCK_8X8 &&
+      partition_none_allowed) {
+    subsize = get_subsize(bsize, PARTITION_HORZ_A);
+    rd_test_partition3(cpi, td, tile_data, tp, pc_tree, &best_rdc,
+                       pc_tree->horizontala, ctx_none, mi_row, mi_col, bsize,
+                       PARTITION_HORZ_A,
+#if CONFIG_SUPERTX
+                       best_rd, &best_rate_nocoef, &x_ctx,
+#endif
+                       mi_row, mi_col, bsize2, mi_row, mi_col + mi_step, bsize2,
+                       mi_row + mi_step, mi_col, subsize);
+    restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+  }
+  // PARTITION_HORZ_B
+  if (partition_horz_allowed && do_rectangular_split && bsize > BLOCK_8X8 &&
+      partition_none_allowed) {
+    subsize = get_subsize(bsize, PARTITION_HORZ_B);
+    rd_test_partition3(cpi, td, tile_data, tp, pc_tree, &best_rdc,
+                       pc_tree->horizontalb, ctx_none, mi_row, mi_col, bsize,
+                       PARTITION_HORZ_B,
+#if CONFIG_SUPERTX
+                       best_rd, &best_rate_nocoef, &x_ctx,
+#endif
+                       mi_row, mi_col, subsize, mi_row + mi_step, mi_col,
+                       bsize2, mi_row + mi_step, mi_col + mi_step, bsize2);
+    restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+  }
+  // PARTITION_VERT_A
+  if (partition_vert_allowed && do_rectangular_split && bsize > BLOCK_8X8 &&
+      partition_none_allowed) {
+    subsize = get_subsize(bsize, PARTITION_VERT_A);
+    rd_test_partition3(cpi, td, tile_data, tp, pc_tree, &best_rdc,
+                       pc_tree->verticala, ctx_none, mi_row, mi_col, bsize,
+                       PARTITION_VERT_A,
+#if CONFIG_SUPERTX
+                       best_rd, &best_rate_nocoef, &x_ctx,
+#endif
+                       mi_row, mi_col, bsize2, mi_row + mi_step, mi_col, bsize2,
+                       mi_row, mi_col + mi_step, subsize);
+    restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+  }
+  // PARTITION_VERT_B
+  if (partition_vert_allowed && do_rectangular_split && bsize > BLOCK_8X8 &&
+      partition_none_allowed) {
+    subsize = get_subsize(bsize, PARTITION_VERT_B);
+    rd_test_partition3(cpi, td, tile_data, tp, pc_tree, &best_rdc,
+                       pc_tree->verticalb, ctx_none, mi_row, mi_col, bsize,
+                       PARTITION_VERT_B,
+#if CONFIG_SUPERTX
+                       best_rd, &best_rate_nocoef, &x_ctx,
+#endif
+                       mi_row, mi_col, subsize, mi_row, mi_col + mi_step,
+                       bsize2, mi_row + mi_step, mi_col + mi_step, bsize2);
+    restore_context(x, &x_ctx, mi_row, mi_col, bsize);
+  }
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+  // TODO(jbb): This code added so that we avoid static analysis
+  // warning related to the fact that best_rd isn't used after this
+  // point.  This code should be refactored so that the duplicate
+  // checks occur in some sub function and thus are used...
+  (void)best_rd;
+  *rd_cost = best_rdc;
+#if CONFIG_SUPERTX
+  *rate_nocoef = best_rate_nocoef;
+#endif  // CONFIG_SUPERTX
+
+#if CONFIG_CFL
+  // Store the luma for the best mode
+  x->cfl_store_y = 1;
+#endif
+  if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
+      pc_tree->index != 3) {
+    if (bsize == cm->sb_size) {
+#if CONFIG_MOTION_VAR && CONFIG_NCOBMC
+      set_mode_info_sb(cpi, td, tile_info, tp, mi_row, mi_col, bsize, pc_tree);
+#endif
+      encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, OUTPUT_ENABLED, bsize,
+                pc_tree, NULL);
+    } else {
+      encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, DRY_RUN_NORMAL, bsize,
+                pc_tree, NULL);
+    }
+  }
+#if CONFIG_CFL
+  x->cfl_store_y = 0;
+#endif
+
+  if (bsize == cm->sb_size) {
+#if !CONFIG_PVQ && !CONFIG_LV_MAP
+    assert(tp_orig < *tp || (tp_orig == *tp && xd->mi[0]->mbmi.skip));
+#endif
+    assert(best_rdc.rate < INT_MAX);
+    assert(best_rdc.dist < INT64_MAX);
+  } else {
+    assert(tp_orig == *tp);
+  }
+}
+
+static void encode_rd_sb_row(AV1_COMP *cpi, ThreadData *td,
+                             TileDataEnc *tile_data, int mi_row,
+                             TOKENEXTRA **tp) {
+  AV1_COMMON *const cm = &cpi->common;
+  const TileInfo *const tile_info = &tile_data->tile_info;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  SPEED_FEATURES *const sf = &cpi->sf;
+  int mi_col;
+#if CONFIG_EXT_PARTITION
+  const int leaf_nodes = 256;
+#else
+  const int leaf_nodes = 64;
+#endif  // CONFIG_EXT_PARTITION
+
+  // Initialize the left context for the new SB row
+  av1_zero_left_context(xd);
+
+#if CONFIG_DELTA_Q
+  // Reset delta for every tile
+  if (cm->delta_q_present_flag)
+    if (mi_row == tile_info->mi_row_start) xd->prev_qindex = cm->base_qindex;
+#if CONFIG_EXT_DELTA_Q
+  if (cm->delta_lf_present_flag)
+    if (mi_row == tile_info->mi_row_start) xd->prev_delta_lf_from_base = 0;
+#endif
+#endif
+
+  // Code each SB in the row
+  for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
+       mi_col += cm->mib_size) {
+    const struct segmentation *const seg = &cm->seg;
+    int dummy_rate;
+    int64_t dummy_dist;
+    RD_STATS dummy_rdc;
+#if CONFIG_SUPERTX
+    int dummy_rate_nocoef;
+#endif  // CONFIG_SUPERTX
+    int i;
+    int seg_skip = 0;
+
+    const int idx_str = cm->mi_stride * mi_row + mi_col;
+    MODE_INFO **mi = cm->mi_grid_visible + idx_str;
+    PC_TREE *const pc_root = td->pc_root[cm->mib_size_log2 - MIN_MIB_SIZE_LOG2];
+
+    av1_update_boundary_info(cm, tile_info, mi_row, mi_col);
+
+    if (sf->adaptive_pred_interp_filter) {
+      for (i = 0; i < leaf_nodes; ++i)
+        td->leaf_tree[i].pred_interp_filter = SWITCHABLE;
+
+      for (i = 0; i < leaf_nodes; ++i) {
+        td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
+        td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
+        td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
+        td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
+      }
+    }
+
+    av1_zero(x->pred_mv);
+    pc_root->index = 0;
+
+    if (seg->enabled) {
+      const uint8_t *const map =
+          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
+      int segment_id = get_segment_id(cm, map, cm->sb_size, mi_row, mi_col);
+      seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
+    }
+
+#if CONFIG_DELTA_Q
+    if (cm->delta_q_present_flag) {
+      // Test mode for delta quantization
+      int sb_row = mi_row >> 3;
+      int sb_col = mi_col >> 3;
+      int sb_stride = (cm->width + MAX_SB_SIZE - 1) >> MAX_SB_SIZE_LOG2;
+      int index = ((sb_row * sb_stride + sb_col + 8) & 31) - 16;
+
+      // Ensure divisibility of delta_qindex by delta_q_res
+      int offset_qindex = (index < 0 ? -index - 8 : index - 8);
+      int qmask = ~(cm->delta_q_res - 1);
+      int current_qindex = clamp(cm->base_qindex + offset_qindex,
+                                 cm->delta_q_res, 256 - cm->delta_q_res);
+
+      current_qindex =
+          ((current_qindex - cm->base_qindex + cm->delta_q_res / 2) & qmask) +
+          cm->base_qindex;
+      assert(current_qindex > 0);
+
+      xd->delta_qindex = current_qindex - cm->base_qindex;
+      set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
+      xd->mi[0]->mbmi.current_q_index = current_qindex;
+#if !CONFIG_EXT_DELTA_Q
+      xd->mi[0]->mbmi.segment_id = 0;
+#endif  // CONFIG_EXT_DELTA_Q
+      av1_init_plane_quantizers(cpi, x, xd->mi[0]->mbmi.segment_id);
+#if CONFIG_EXT_DELTA_Q
+      if (cpi->oxcf.deltaq_mode == DELTA_Q_LF) {
+        int j, k;
+        int lfmask = ~(cm->delta_lf_res - 1);
+        int current_delta_lf_from_base = offset_qindex / 2;
+        current_delta_lf_from_base =
+            ((current_delta_lf_from_base + cm->delta_lf_res / 2) & lfmask);
+
+        // pre-set the delta lf for loop filter. Note that this value is set
+        // before mi is assigned for each block in current superblock
+        for (j = 0; j < AOMMIN(cm->mib_size, cm->mi_rows - mi_row); j++) {
+          for (k = 0; k < AOMMIN(cm->mib_size, cm->mi_cols - mi_col); k++) {
+            cm->mi[(mi_row + j) * cm->mi_stride + (mi_col + k)]
+                .mbmi.current_delta_lf_from_base = current_delta_lf_from_base;
+          }
+        }
+      }
+#endif  // CONFIG_EXT_DELTA_Q
+    }
+#endif  // CONFIG_DELTA_Q
+
+    x->source_variance = UINT_MAX;
+    if (sf->partition_search_type == FIXED_PARTITION || seg_skip) {
+      BLOCK_SIZE bsize;
+      set_offsets(cpi, tile_info, x, mi_row, mi_col, cm->sb_size);
+      bsize = seg_skip ? cm->sb_size : sf->always_this_block_size;
+      set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
+      rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, cm->sb_size,
+                       &dummy_rate, &dummy_dist,
+#if CONFIG_SUPERTX
+                       &dummy_rate_nocoef,
+#endif  // CONFIG_SUPERTX
+                       1, pc_root);
+    } else if (cpi->partition_search_skippable_frame) {
+      BLOCK_SIZE bsize;
+      set_offsets(cpi, tile_info, x, mi_row, mi_col, cm->sb_size);
+      bsize = get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col);
+      set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
+      rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, cm->sb_size,
+                       &dummy_rate, &dummy_dist,
+#if CONFIG_SUPERTX
+                       &dummy_rate_nocoef,
+#endif  // CONFIG_SUPERTX
+                       1, pc_root);
+    } else if (sf->partition_search_type == VAR_BASED_PARTITION) {
+      choose_partitioning(cpi, td, tile_info, x, mi_row, mi_col);
+      rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, cm->sb_size,
+                       &dummy_rate, &dummy_dist,
+#if CONFIG_SUPERTX
+                       &dummy_rate_nocoef,
+#endif  // CONFIG_SUPERTX
+                       1, pc_root);
+    } else {
+      // If required set upper and lower partition size limits
+      if (sf->auto_min_max_partition_size) {
+        set_offsets(cpi, tile_info, x, mi_row, mi_col, cm->sb_size);
+        rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col,
+                                &x->min_partition_size, &x->max_partition_size);
+      }
+      rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, cm->sb_size,
+                        &dummy_rdc,
+#if CONFIG_SUPERTX
+                        &dummy_rate_nocoef,
+#endif  // CONFIG_SUPERTX
+                        INT64_MAX, pc_root);
+    }
+  }
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  if (cm->do_subframe_update &&
+      cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+    const int mi_rows_per_update =
+        MI_SIZE * AOMMAX(cm->mi_rows / MI_SIZE / COEF_PROBS_BUFS, 1);
+    if ((mi_row + MI_SIZE) % mi_rows_per_update == 0 &&
+        mi_row + MI_SIZE < cm->mi_rows &&
+        cm->coef_probs_update_idx < COEF_PROBS_BUFS - 1) {
+      TX_SIZE t;
+      SUBFRAME_STATS *subframe_stats = &cpi->subframe_stats;
+
+      for (t = 0; t < TX_SIZES; ++t)
+        av1_full_to_model_counts(cpi->td.counts->coef[t],
+                                 cpi->td.rd_counts.coef_counts[t]);
+      av1_partial_adapt_probs(cm, mi_row, mi_col);
+      ++cm->coef_probs_update_idx;
+      av1_copy(subframe_stats->coef_probs_buf[cm->coef_probs_update_idx],
+               cm->fc->coef_probs);
+      av1_copy(subframe_stats->coef_counts_buf[cm->coef_probs_update_idx],
+               cpi->td.rd_counts.coef_counts);
+      av1_copy(subframe_stats->eob_counts_buf[cm->coef_probs_update_idx],
+               cm->counts.eob_branch);
+      av1_fill_token_costs(x->token_costs, cm->fc->coef_probs);
+    }
+  }
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+}
+
+static void init_encode_frame_mb_context(AV1_COMP *cpi) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+
+  // Copy data over into macro block data structures.
+  av1_setup_src_planes(x, cpi->source, 0, 0);
+
+  av1_setup_block_planes(xd, cm->subsampling_x, cm->subsampling_y);
+}
+
+#if !CONFIG_REF_ADAPT
+static int check_dual_ref_flags(AV1_COMP *cpi) {
+  const int ref_flags = cpi->ref_frame_flags;
+
+  if (segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
+    return 0;
+  } else {
+    return (!!(ref_flags & AOM_GOLD_FLAG) + !!(ref_flags & AOM_LAST_FLAG) +
+#if CONFIG_EXT_REFS
+            !!(ref_flags & AOM_LAST2_FLAG) + !!(ref_flags & AOM_LAST3_FLAG) +
+            !!(ref_flags & AOM_BWD_FLAG) +
+#endif  // CONFIG_EXT_REFS
+            !!(ref_flags & AOM_ALT_FLAG)) >= 2;
+  }
+}
+#endif  // !CONFIG_REF_ADAPT
+
+#if !CONFIG_VAR_TX
+static void reset_skip_tx_size(AV1_COMMON *cm, TX_SIZE max_tx_size) {
+  int mi_row, mi_col;
+  const int mis = cm->mi_stride;
+  MODE_INFO **mi_ptr = cm->mi_grid_visible;
+
+  for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
+    for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
+      if (txsize_sqr_up_map[mi_ptr[mi_col]->mbmi.tx_size] > max_tx_size)
+        mi_ptr[mi_col]->mbmi.tx_size = max_tx_size;
+    }
+  }
+}
+#endif
+
+static MV_REFERENCE_FRAME get_frame_type(const AV1_COMP *cpi) {
+  if (frame_is_intra_only(&cpi->common)) return INTRA_FRAME;
+#if CONFIG_EXT_REFS
+  // We will not update the golden frame with an internal overlay frame
+  else if ((cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame) ||
+           cpi->rc.is_src_frame_ext_arf)
+#else
+  else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
+#endif
+    return ALTREF_FRAME;
+  else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
+    return GOLDEN_FRAME;
+  else
+    // TODO(zoeliu): To investigate whether a frame_type other than
+    // INTRA/ALTREF/GOLDEN/LAST needs to be specified seperately.
+    return LAST_FRAME;
+}
+
+static TX_MODE select_tx_mode(const AV1_COMP *cpi, MACROBLOCKD *const xd) {
+  int i, all_lossless = 1;
+
+  if (cpi->common.seg.enabled) {
+    for (i = 0; i < MAX_SEGMENTS; ++i) {
+      if (!xd->lossless[i]) {
+        all_lossless = 0;
+        break;
+      }
+    }
+  } else {
+    all_lossless = xd->lossless[0];
+  }
+  if (all_lossless) return ONLY_4X4;
+  if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
+    return ALLOW_32X32 + CONFIG_TX64X64;
+  else if (cpi->sf.tx_size_search_method == USE_FULL_RD ||
+           cpi->sf.tx_size_search_method == USE_TX_8X8)
+    return TX_MODE_SELECT;
+  else
+    return cpi->common.tx_mode;
+}
+
+void av1_init_tile_data(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  int tile_col, tile_row;
+  TOKENEXTRA *pre_tok = cpi->tile_tok[0][0];
+  unsigned int tile_tok = 0;
+
+  if (cpi->tile_data == NULL || cpi->allocated_tiles < tile_cols * tile_rows) {
+    if (cpi->tile_data != NULL) aom_free(cpi->tile_data);
+    CHECK_MEM_ERROR(
+        cm, cpi->tile_data,
+        aom_memalign(32, tile_cols * tile_rows * sizeof(*cpi->tile_data)));
+    cpi->allocated_tiles = tile_cols * tile_rows;
+
+    for (tile_row = 0; tile_row < tile_rows; ++tile_row)
+      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+        TileDataEnc *const tile_data =
+            &cpi->tile_data[tile_row * tile_cols + tile_col];
+        int i, j;
+        for (i = 0; i < BLOCK_SIZES; ++i) {
+          for (j = 0; j < MAX_MODES; ++j) {
+            tile_data->thresh_freq_fact[i][j] = 32;
+            tile_data->mode_map[i][j] = j;
+          }
+        }
+#if CONFIG_PVQ
+        // This will be dynamically increased as more pvq block is encoded.
+        tile_data->pvq_q.buf_len = 1000;
+        CHECK_MEM_ERROR(
+            cm, tile_data->pvq_q.buf,
+            aom_malloc(tile_data->pvq_q.buf_len * sizeof(PVQ_INFO)));
+        tile_data->pvq_q.curr_pos = 0;
+#endif
+      }
+  }
+
+  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
+    for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+      TileInfo *const tile_info =
+          &cpi->tile_data[tile_row * tile_cols + tile_col].tile_info;
+      av1_tile_init(tile_info, cm, tile_row, tile_col);
+
+      cpi->tile_tok[tile_row][tile_col] = pre_tok + tile_tok;
+      pre_tok = cpi->tile_tok[tile_row][tile_col];
+      tile_tok = allocated_tokens(*tile_info);
+#if CONFIG_PVQ
+      cpi->tile_data[tile_row * tile_cols + tile_col].pvq_q.curr_pos = 0;
+#endif
+    }
+  }
+}
+
+void av1_encode_tile(AV1_COMP *cpi, ThreadData *td, int tile_row,
+                     int tile_col) {
+  AV1_COMMON *const cm = &cpi->common;
+  TileDataEnc *const this_tile =
+      &cpi->tile_data[tile_row * cm->tile_cols + tile_col];
+  const TileInfo *const tile_info = &this_tile->tile_info;
+  TOKENEXTRA *tok = cpi->tile_tok[tile_row][tile_col];
+  int mi_row;
+
+#if CONFIG_DEPENDENT_HORZTILES
+#if CONFIG_TILE_GROUPS
+  if ((!cm->dependent_horz_tiles) || (tile_row == 0) ||
+      tile_info->tg_horz_boundary) {
+#else
+  if ((!cm->dependent_horz_tiles) || (tile_row == 0)) {
+#endif
+    av1_zero_above_context(cm, tile_info->mi_col_start, tile_info->mi_col_end);
+  }
+#else
+  av1_zero_above_context(cm, tile_info->mi_col_start, tile_info->mi_col_end);
+#endif
+
+  // Set up pointers to per thread motion search counters.
+  this_tile->m_search_count = 0;   // Count of motion search hits.
+  this_tile->ex_search_count = 0;  // Exhaustive mesh search hits.
+  td->mb.m_search_count_ptr = &this_tile->m_search_count;
+  td->mb.ex_search_count_ptr = &this_tile->ex_search_count;
+
+#if CONFIG_PVQ
+  td->mb.pvq_q = &this_tile->pvq_q;
+
+  // TODO(yushin) : activity masking info needs be signaled by a bitstream
+  td->mb.daala_enc.use_activity_masking = AV1_PVQ_ENABLE_ACTIVITY_MASKING;
+
+  if (td->mb.daala_enc.use_activity_masking)
+    td->mb.daala_enc.qm = OD_HVS_QM;  // Hard coded. Enc/dec required to sync.
+  else
+    td->mb.daala_enc.qm = OD_FLAT_QM;  // Hard coded. Enc/dec required to sync.
+
+  {
+    // FIXME: Multiple segments support
+    int segment_id = 0;
+    int rdmult = set_segment_rdmult(cpi, &td->mb, segment_id);
+    int qindex = av1_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+#if CONFIG_HIGHBITDEPTH
+    const int quantizer_shift = td->mb.e_mbd.bd - 8;
+#else
+    const int quantizer_shift = 0;
+#endif  // CONFIG_HIGHBITDEPTH
+    int64_t q_ac = OD_MAXI(
+        1, av1_ac_quant(qindex, 0, cpi->common.bit_depth) >> quantizer_shift);
+    int64_t q_dc = OD_MAXI(
+        1, av1_dc_quant(qindex, 0, cpi->common.bit_depth) >> quantizer_shift);
+    /* td->mb.daala_enc.pvq_norm_lambda = OD_PVQ_LAMBDA; */
+    td->mb.daala_enc.pvq_norm_lambda =
+        (double)rdmult * (64 / 16) / (q_ac * q_ac * (1 << RDDIV_BITS));
+    td->mb.daala_enc.pvq_norm_lambda_dc =
+        (double)rdmult * (64 / 16) / (q_dc * q_dc * (1 << RDDIV_BITS));
+    // printf("%f\n", td->mb.daala_enc.pvq_norm_lambda);
+  }
+  od_init_qm(td->mb.daala_enc.state.qm, td->mb.daala_enc.state.qm_inv,
+             td->mb.daala_enc.qm == OD_HVS_QM ? OD_QM8_Q4_HVS : OD_QM8_Q4_FLAT);
+
+  if (td->mb.daala_enc.use_activity_masking) {
+    int pli;
+    int use_masking = td->mb.daala_enc.use_activity_masking;
+    int segment_id = 0;
+    int qindex = av1_get_qindex(&cm->seg, segment_id, cm->base_qindex);
+
+    for (pli = 0; pli < MAX_MB_PLANE; pli++) {
+      int i;
+      int q;
+
+      q = qindex;
+      if (q <= OD_DEFAULT_QMS[use_masking][0][pli].interp_q << OD_COEFF_SHIFT) {
+        od_interp_qm(&td->mb.daala_enc.state.pvq_qm_q4[pli][0], q,
+                     &OD_DEFAULT_QMS[use_masking][0][pli], NULL);
+      } else {
+        i = 0;
+        while (OD_DEFAULT_QMS[use_masking][i + 1][pli].qm_q4 != NULL &&
+               q > OD_DEFAULT_QMS[use_masking][i + 1][pli].interp_q
+                       << OD_COEFF_SHIFT) {
+          i++;
+        }
+        od_interp_qm(&td->mb.daala_enc.state.pvq_qm_q4[pli][0], q,
+                     &OD_DEFAULT_QMS[use_masking][i][pli],
+                     &OD_DEFAULT_QMS[use_masking][i + 1][pli]);
+      }
+    }
+  }
+
+#if CONFIG_DAALA_EC
+  od_ec_enc_init(&td->mb.daala_enc.w.ec, 65025);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+
+#if CONFIG_DAALA_EC
+  od_ec_enc_reset(&td->mb.daala_enc.w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+#endif  // #if CONFIG_PVQ
+
+#if CONFIG_EC_ADAPT
+  this_tile->tctx = *cm->fc;
+  td->mb.e_mbd.tile_ctx = &this_tile->tctx;
+#endif  // #if CONFIG_EC_ADAPT
+
+#if CONFIG_CFL
+  MACROBLOCKD *const xd = &td->mb.e_mbd;
+  xd->cfl = &this_tile->cfl;
+  cfl_init(xd->cfl, cm, xd->plane[AOM_PLANE_U].subsampling_x,
+           xd->plane[AOM_PLANE_U].subsampling_y);
+#endif
+
+#if CONFIG_PVQ
+  td->mb.daala_enc.state.adapt = &this_tile->tctx.pvq_context;
+#endif  // CONFIG_PVQ
+
+  for (mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end;
+       mi_row += cm->mib_size) {
+    encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok);
+  }
+
+  cpi->tok_count[tile_row][tile_col] =
+      (unsigned int)(tok - cpi->tile_tok[tile_row][tile_col]);
+  assert(cpi->tok_count[tile_row][tile_col] <= allocated_tokens(*tile_info));
+#if CONFIG_PVQ
+#if CONFIG_DAALA_EC
+  od_ec_enc_clear(&td->mb.daala_enc.w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+
+  td->mb.pvq_q->last_pos = td->mb.pvq_q->curr_pos;
+  // rewind current position so that bitstream can be written
+  // from the 1st pvq block
+  td->mb.pvq_q->curr_pos = 0;
+
+  td->mb.pvq_q = NULL;
+#endif
+}
+
+static void encode_tiles(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  int tile_col, tile_row;
+
+  av1_init_tile_data(cpi);
+
+  for (tile_row = 0; tile_row < cm->tile_rows; ++tile_row)
+    for (tile_col = 0; tile_col < cm->tile_cols; ++tile_col)
+      av1_encode_tile(cpi, &cpi->td, tile_row, tile_col);
+}
+
+#if CONFIG_FP_MB_STATS
+static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
+                            AV1_COMMON *cm, uint8_t **this_frame_mb_stats) {
+  uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
+                         cm->current_video_frame * cm->MBs * sizeof(uint8_t);
+
+  if (mb_stats_in > firstpass_mb_stats->mb_stats_end) return EOF;
+
+  *this_frame_mb_stats = mb_stats_in;
+
+  return 1;
+}
+#endif
+
+#if CONFIG_GLOBAL_MOTION
+#define GLOBAL_TRANS_TYPES_ENC 3  // highest motion model to search
+static int gm_get_params_cost(WarpedMotionParams *gm,
+                              WarpedMotionParams *ref_gm, int allow_hp) {
+  assert(gm->wmtype < GLOBAL_TRANS_TYPES);
+  int params_cost = 0;
+  int trans_bits, trans_prec_diff;
+  switch (gm->wmtype) {
+    case HOMOGRAPHY:
+    case HORTRAPEZOID:
+    case VERTRAPEZOID:
+      if (gm->wmtype != HORTRAPEZOID)
+        params_cost += aom_count_signed_primitive_refsubexpfin(
+            GM_ROW3HOMO_MAX + 1, SUBEXPFIN_K,
+            (ref_gm->wmmat[6] >> GM_ROW3HOMO_PREC_DIFF),
+            (gm->wmmat[6] >> GM_ROW3HOMO_PREC_DIFF));
+      if (gm->wmtype != VERTRAPEZOID)
+        params_cost += aom_count_signed_primitive_refsubexpfin(
+            GM_ROW3HOMO_MAX + 1, SUBEXPFIN_K,
+            (ref_gm->wmmat[7] >> GM_ROW3HOMO_PREC_DIFF),
+            (gm->wmmat[7] >> GM_ROW3HOMO_PREC_DIFF));
+    // Fallthrough intended
+    case AFFINE:
+    case ROTZOOM:
+      params_cost += aom_count_signed_primitive_refsubexpfin(
+          GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+          (ref_gm->wmmat[2] >> GM_ALPHA_PREC_DIFF) - (1 << GM_ALPHA_PREC_BITS),
+          (gm->wmmat[2] >> GM_ALPHA_PREC_DIFF) - (1 << GM_ALPHA_PREC_BITS));
+      if (gm->wmtype != VERTRAPEZOID)
+        params_cost += aom_count_signed_primitive_refsubexpfin(
+            GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+            (ref_gm->wmmat[3] >> GM_ALPHA_PREC_DIFF),
+            (gm->wmmat[3] >> GM_ALPHA_PREC_DIFF));
+      if (gm->wmtype >= AFFINE) {
+        if (gm->wmtype != HORTRAPEZOID)
+          params_cost += aom_count_signed_primitive_refsubexpfin(
+              GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+              (ref_gm->wmmat[4] >> GM_ALPHA_PREC_DIFF),
+              (gm->wmmat[4] >> GM_ALPHA_PREC_DIFF));
+        params_cost += aom_count_signed_primitive_refsubexpfin(
+            GM_ALPHA_MAX + 1, SUBEXPFIN_K,
+            (ref_gm->wmmat[5] >> GM_ALPHA_PREC_DIFF) -
+                (1 << GM_ALPHA_PREC_BITS),
+            (gm->wmmat[5] >> GM_ALPHA_PREC_DIFF) - (1 << GM_ALPHA_PREC_BITS));
+      }
+    // Fallthrough intended
+    case TRANSLATION:
+      trans_bits = (gm->wmtype == TRANSLATION)
+                       ? GM_ABS_TRANS_ONLY_BITS - !allow_hp
+                       : GM_ABS_TRANS_BITS;
+      trans_prec_diff = (gm->wmtype == TRANSLATION)
+                            ? GM_TRANS_ONLY_PREC_DIFF + !allow_hp
+                            : GM_TRANS_PREC_DIFF;
+      params_cost += aom_count_signed_primitive_refsubexpfin(
+          (1 << trans_bits) + 1, SUBEXPFIN_K,
+          (ref_gm->wmmat[0] >> trans_prec_diff),
+          (gm->wmmat[0] >> trans_prec_diff));
+      params_cost += aom_count_signed_primitive_refsubexpfin(
+          (1 << trans_bits) + 1, SUBEXPFIN_K,
+          (ref_gm->wmmat[1] >> trans_prec_diff),
+          (gm->wmmat[1] >> trans_prec_diff));
+    // Fallthrough intended
+    case IDENTITY: break;
+    default: assert(0);
+  }
+  return (params_cost << AV1_PROB_COST_SHIFT);
+}
+#endif  // CONFIG_GLOBAL_MOTION
+
+static void encode_frame_internal(AV1_COMP *cpi) {
+  ThreadData *const td = &cpi->td;
+  MACROBLOCK *const x = &td->mb;
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  RD_COUNTS *const rdc = &cpi->td.rd_counts;
+  int i;
+#if CONFIG_TEMPMV_SIGNALING || CONFIG_EXT_REFS
+  const int last_fb_buf_idx = get_ref_frame_buf_idx(cpi, LAST_FRAME);
+#endif  // CONFIG_TEMPMV_SIGNALING || CONFIG_EXT_REFS
+
+#if CONFIG_ADAPT_SCAN
+  av1_deliver_eob_threshold(cm, xd);
+#endif
+
+  x->min_partition_size = AOMMIN(x->min_partition_size, cm->sb_size);
+  x->max_partition_size = AOMMIN(x->max_partition_size, cm->sb_size);
+#if CONFIG_REF_MV
+  cm->setup_mi(cm);
+#endif
+
+  xd->mi = cm->mi_grid_visible;
+  xd->mi[0] = cm->mi;
+
+  av1_zero(*td->counts);
+  av1_zero(rdc->coef_counts);
+  av1_zero(rdc->comp_pred_diff);
+
+#if CONFIG_GLOBAL_MOTION
+  av1_zero(rdc->global_motion_used);
+  if (cpi->common.frame_type == INTER_FRAME && cpi->source &&
+      !cpi->global_motion_search_done) {
+    YV12_BUFFER_CONFIG *ref_buf;
+    int frame;
+    double params_by_motion[RANSAC_NUM_MOTIONS * (MAX_PARAMDIM - 1)];
+    const double *params_this_motion;
+    int inliers_by_motion[RANSAC_NUM_MOTIONS];
+    WarpedMotionParams tmp_wm_params;
+    static const double kInfiniteErrAdv = 1e12;
+    static const double kIdentityParams[MAX_PARAMDIM - 1] = {
+      0.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 0.0
+    };
+
+    for (frame = LAST_FRAME; frame <= ALTREF_FRAME; ++frame) {
+      ref_buf = get_ref_frame_buffer(cpi, frame);
+      if (ref_buf) {
+        TransformationType model;
+        aom_clear_system_state();
+        for (model = ROTZOOM; model < GLOBAL_TRANS_TYPES_ENC; ++model) {
+          double best_erroradvantage = kInfiniteErrAdv;
+
+          // Initially set all params to identity.
+          for (i = 0; i < RANSAC_NUM_MOTIONS; ++i) {
+            memcpy(params_by_motion + (MAX_PARAMDIM - 1) * i, kIdentityParams,
+                   (MAX_PARAMDIM - 1) * sizeof(*params_by_motion));
+          }
+
+          compute_global_motion_feature_based(
+              model, cpi->source, ref_buf,
+#if CONFIG_HIGHBITDEPTH
+              cpi->common.bit_depth,
+#endif  // CONFIG_HIGHBITDEPTH
+              inliers_by_motion, params_by_motion, RANSAC_NUM_MOTIONS);
+
+          for (i = 0; i < RANSAC_NUM_MOTIONS; ++i) {
+            if (inliers_by_motion[i] == 0) continue;
+
+            params_this_motion = params_by_motion + (MAX_PARAMDIM - 1) * i;
+            convert_model_to_params(params_this_motion, &tmp_wm_params);
+
+            if (tmp_wm_params.wmtype != IDENTITY) {
+              const double erroradv_this_motion = refine_integerized_param(
+                  &tmp_wm_params, tmp_wm_params.wmtype,
+#if CONFIG_HIGHBITDEPTH
+                  xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH, xd->bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                  ref_buf->y_buffer, ref_buf->y_width, ref_buf->y_height,
+                  ref_buf->y_stride, cpi->source->y_buffer,
+                  cpi->source->y_width, cpi->source->y_height,
+                  cpi->source->y_stride, 3);
+              if (erroradv_this_motion < best_erroradvantage) {
+                best_erroradvantage = erroradv_this_motion;
+                // Save the wm_params modified by refine_integerized_param()
+                // rather than motion index to avoid rerunning refine() below.
+                memcpy(&(cm->global_motion[frame]), &tmp_wm_params,
+                       sizeof(WarpedMotionParams));
+              }
+            }
+          }
+          if (cm->global_motion[frame].wmtype <= AFFINE)
+            if (!get_shear_params(&cm->global_motion[frame]))
+              set_default_warp_params(&cm->global_motion[frame]);
+
+          if (cm->global_motion[frame].wmtype == TRANSLATION) {
+            cm->global_motion[frame].wmmat[0] =
+                convert_to_trans_prec(cm->allow_high_precision_mv,
+                                      cm->global_motion[frame].wmmat[0]) *
+                GM_TRANS_ONLY_DECODE_FACTOR;
+            cm->global_motion[frame].wmmat[1] =
+                convert_to_trans_prec(cm->allow_high_precision_mv,
+                                      cm->global_motion[frame].wmmat[1]) *
+                GM_TRANS_ONLY_DECODE_FACTOR;
+          }
+
+          // If the best error advantage found doesn't meet the threshold for
+          // this motion type, revert to IDENTITY.
+          if (!is_enough_erroradvantage(
+                  best_erroradvantage,
+                  gm_get_params_cost(&cm->global_motion[frame],
+                                     &cm->prev_frame->global_motion[frame],
+                                     cm->allow_high_precision_mv))) {
+            set_default_warp_params(&cm->global_motion[frame]);
+          }
+
+          if (cm->global_motion[frame].wmtype != IDENTITY) break;
+        }
+        aom_clear_system_state();
+      }
+      cpi->gmparams_cost[frame] =
+          gm_get_params_cost(&cm->global_motion[frame],
+                             &cm->prev_frame->global_motion[frame],
+                             cm->allow_high_precision_mv) +
+          cpi->gmtype_cost[cm->global_motion[frame].wmtype] -
+          cpi->gmtype_cost[IDENTITY];
+    }
+    cpi->global_motion_search_done = 1;
+  }
+  memcpy(cm->cur_frame->global_motion, cm->global_motion,
+         TOTAL_REFS_PER_FRAME * sizeof(WarpedMotionParams));
+#endif  // CONFIG_GLOBAL_MOTION
+
+  for (i = 0; i < MAX_SEGMENTS; ++i) {
+    const int qindex = cm->seg.enabled
+                           ? av1_get_qindex(&cm->seg, i, cm->base_qindex)
+                           : cm->base_qindex;
+    xd->lossless[i] = qindex == 0 && cm->y_dc_delta_q == 0 &&
+                      cm->uv_dc_delta_q == 0 && cm->uv_ac_delta_q == 0;
+    xd->qindex[i] = qindex;
+  }
+
+  if (!cm->seg.enabled && xd->lossless[0]) x->optimize = 0;
+
+  cm->tx_mode = select_tx_mode(cpi, xd);
+
+#if CONFIG_DELTA_Q
+  // Fix delta q resolution for the moment
+  cm->delta_q_res = DEFAULT_DELTA_Q_RES;
+// Set delta_q_present_flag before it is used for the first time
+#if CONFIG_EXT_DELTA_Q
+  cm->delta_lf_res = DEFAULT_DELTA_LF_RES;
+  // update delta_q_present_flag and delta_lf_present_flag based on base_qindex
+  cm->delta_q_present_flag &= cm->base_qindex > 0;
+  cm->delta_lf_present_flag &= cm->base_qindex > 0;
+#else
+  cm->delta_q_present_flag =
+      cpi->oxcf.aq_mode == DELTA_AQ && cm->base_qindex > 0;
+#endif  // CONFIG_EXT_DELTA_Q
+#endif
+
+  av1_frame_init_quantizer(cpi);
+
+  av1_initialize_rd_consts(cpi);
+  av1_initialize_me_consts(cpi, x, cm->base_qindex);
+  init_encode_frame_mb_context(cpi);
+#if CONFIG_TEMPMV_SIGNALING
+  if (last_fb_buf_idx != INVALID_IDX) {
+    cm->prev_frame = &cm->buffer_pool->frame_bufs[last_fb_buf_idx];
+    cm->use_prev_frame_mvs &= !cm->error_resilient_mode &&
+                              cm->width == cm->prev_frame->buf.y_width &&
+                              cm->height == cm->prev_frame->buf.y_height &&
+                              !cm->intra_only && !cm->prev_frame->intra_only;
+  }
+#else
+  cm->use_prev_frame_mvs =
+      !cm->error_resilient_mode && cm->width == cm->last_width &&
+      cm->height == cm->last_height && !cm->intra_only && cm->last_show_frame;
+#endif
+
+#if CONFIG_EXT_REFS
+  // NOTE(zoeliu): As cm->prev_frame can take neither a frame of
+  //               show_exisiting_frame=1, nor can it take a frame not used as
+  //               a reference, it is probable that by the time it is being
+  //               referred to, the frame buffer it originally points to may
+  //               already get expired and have been reassigned to the current
+  //               newly coded frame. Hence, we need to check whether this is
+  //               the case, and if yes, we have 2 choices:
+  //               (1) Simply disable the use of previous frame mvs; or
+  //               (2) Have cm->prev_frame point to one reference frame buffer,
+  //                   e.g. LAST_FRAME.
+  if (cm->use_prev_frame_mvs && !enc_is_ref_frame_buf(cpi, cm->prev_frame)) {
+    // Reassign the LAST_FRAME buffer to cm->prev_frame.
+    cm->prev_frame = &cm->buffer_pool->frame_bufs[last_fb_buf_idx];
+  }
+#endif  // CONFIG_EXT_REFS
+
+  // Special case: set prev_mi to NULL when the previous mode info
+  // context cannot be used.
+  cm->prev_mi =
+      cm->use_prev_frame_mvs ? cm->prev_mip + cm->mi_stride + 1 : NULL;
+
+#if CONFIG_VAR_TX
+  x->txb_split_count = 0;
+#if CONFIG_REF_MV
+  av1_zero(x->blk_skip_drl);
+#endif
+#endif
+
+  if (cpi->sf.partition_search_type == VAR_BASED_PARTITION &&
+      cpi->td.var_root[0] == NULL)
+    av1_setup_var_tree(&cpi->common, &cpi->td);
+
+  {
+    struct aom_usec_timer emr_timer;
+    aom_usec_timer_start(&emr_timer);
+
+#if CONFIG_FP_MB_STATS
+    if (cpi->use_fp_mb_stats) {
+      input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
+                       &cpi->twopass.this_frame_mb_stats);
+    }
+#endif
+
+    // If allowed, encoding tiles in parallel with one thread handling one tile.
+    // TODO(geza.lore): The multi-threaded encoder is not safe with more than
+    // 1 tile rows, as it uses the single above_context et al arrays from
+    // cpi->common
+    if (AOMMIN(cpi->oxcf.max_threads, cm->tile_cols) > 1 && cm->tile_rows == 1)
+      av1_encode_tiles_mt(cpi);
+    else
+      encode_tiles(cpi);
+
+    aom_usec_timer_mark(&emr_timer);
+    cpi->time_encode_sb_row += aom_usec_timer_elapsed(&emr_timer);
+  }
+
+#if 0
+  // Keep record of the total distortion this time around for future use
+  cpi->last_frame_distortion = cpi->frame_distortion;
+#endif
+}
+
+void av1_encode_frame(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+#if CONFIG_EXT_TX
+  // Indicates whether or not to use a default reduced set for ext-tx
+  // rather than the potential full set of 16 transforms
+  cm->reduced_tx_set_used = 0;
+#endif  // CONFIG_EXT_TX
+
+  // In the longer term the encoder should be generalized to match the
+  // decoder such that we allow compound where one of the 3 buffers has a
+  // different sign bias and that buffer is then the fixed ref. However, this
+  // requires further work in the rd loop. For now the only supported encoder
+  // side behavior is where the ALT ref buffer has opposite sign bias to
+  // the other two.
+  if (!frame_is_intra_only(cm)) {
+#if CONFIG_LOWDELAY_COMPOUND  // Normative in encoder
+    cpi->allow_comp_inter_inter = 1;
+#if CONFIG_EXT_REFS
+    cm->comp_fwd_ref[0] = LAST_FRAME;
+    cm->comp_fwd_ref[1] = LAST2_FRAME;
+    cm->comp_fwd_ref[2] = LAST3_FRAME;
+    cm->comp_fwd_ref[3] = GOLDEN_FRAME;
+    cm->comp_bwd_ref[0] = BWDREF_FRAME;
+    cm->comp_bwd_ref[1] = ALTREF_FRAME;
+#else
+    cm->comp_fixed_ref = ALTREF_FRAME;
+    cm->comp_var_ref[0] = LAST_FRAME;
+    cm->comp_var_ref[1] = GOLDEN_FRAME;
+#endif  // CONFIG_EXT_REFS
+#else
+    if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
+         cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
+        (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
+         cm->ref_frame_sign_bias[LAST_FRAME])) {
+      cpi->allow_comp_inter_inter = 0;
+    } else {
+      cpi->allow_comp_inter_inter = 1;
+
+#if CONFIG_EXT_REFS
+      cm->comp_fwd_ref[0] = LAST_FRAME;
+      cm->comp_fwd_ref[1] = LAST2_FRAME;
+      cm->comp_fwd_ref[2] = LAST3_FRAME;
+      cm->comp_fwd_ref[3] = GOLDEN_FRAME;
+      cm->comp_bwd_ref[0] = BWDREF_FRAME;
+      cm->comp_bwd_ref[1] = ALTREF_FRAME;
+#else
+      cm->comp_fixed_ref = ALTREF_FRAME;
+      cm->comp_var_ref[0] = LAST_FRAME;
+      cm->comp_var_ref[1] = GOLDEN_FRAME;
+#endif  // CONFIG_EXT_REFS
+    }
+#endif
+  } else {
+    cpi->allow_comp_inter_inter = 0;
+  }
+
+  if (cpi->sf.frame_parameter_update) {
+    int i;
+    RD_OPT *const rd_opt = &cpi->rd;
+    FRAME_COUNTS *counts = cpi->td.counts;
+    RD_COUNTS *const rdc = &cpi->td.rd_counts;
+
+    // This code does a single RD pass over the whole frame assuming
+    // either compound, single or hybrid prediction as per whatever has
+    // worked best for that type of frame in the past.
+    // It also predicts whether another coding mode would have worked
+    // better than this coding mode. If that is the case, it remembers
+    // that for subsequent frames.
+    // It does the same analysis for transform size selection also.
+    //
+    // TODO(zoeliu): To investigate whether a frame_type other than
+    // INTRA/ALTREF/GOLDEN/LAST needs to be specified seperately.
+    const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
+    int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
+    const int is_alt_ref = frame_type == ALTREF_FRAME;
+
+/* prediction (compound, single or hybrid) mode selection */
+#if CONFIG_REF_ADAPT
+    // NOTE(zoeliu): "is_alt_ref" is true only for OVERLAY/INTNL_OVERLAY frames
+    if (is_alt_ref || !cpi->allow_comp_inter_inter)
+      cm->reference_mode = SINGLE_REFERENCE;
+    else
+      cm->reference_mode = REFERENCE_MODE_SELECT;
+#else
+    if (is_alt_ref || !cpi->allow_comp_inter_inter)
+      cm->reference_mode = SINGLE_REFERENCE;
+    else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
+             mode_thrs[COMPOUND_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT] &&
+             check_dual_ref_flags(cpi) && cpi->static_mb_pct == 100)
+      cm->reference_mode = COMPOUND_REFERENCE;
+    else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
+      cm->reference_mode = SINGLE_REFERENCE;
+    else
+      cm->reference_mode = REFERENCE_MODE_SELECT;
+#endif  // CONFIG_REF_ADAPT
+
+#if CONFIG_DUAL_FILTER
+    cm->interp_filter = SWITCHABLE;
+#endif
+
+    encode_frame_internal(cpi);
+
+    for (i = 0; i < REFERENCE_MODES; ++i)
+      mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2;
+
+    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+      int single_count_zero = 0;
+      int comp_count_zero = 0;
+
+      for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
+        single_count_zero += counts->comp_inter[i][0];
+        comp_count_zero += counts->comp_inter[i][1];
+      }
+
+      if (comp_count_zero == 0) {
+        cm->reference_mode = SINGLE_REFERENCE;
+        av1_zero(counts->comp_inter);
+#if !CONFIG_REF_ADAPT
+      } else if (single_count_zero == 0) {
+        cm->reference_mode = COMPOUND_REFERENCE;
+        av1_zero(counts->comp_inter);
+#endif  // !CONFIG_REF_ADAPT
+      }
+    }
+
+#if CONFIG_VAR_TX
+    if (cm->tx_mode == TX_MODE_SELECT && cpi->td.mb.txb_split_count == 0)
+      cm->tx_mode = ALLOW_32X32 + CONFIG_TX64X64;
+#else
+    if (cm->tx_mode == TX_MODE_SELECT) {
+#if CONFIG_TX64X64
+      int count4x4 = 0;
+      int count8x8_8x8p = 0, count8x8_lp = 0;
+      int count16x16_16x16p = 0, count16x16_lp = 0;
+      int count32x32_32x32p = 0, count32x32_lp = 0;
+      int count64x64_64x64p = 0;
+      for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
+        // counts->tx_size[max_depth][context_idx][this_depth_level]
+        count4x4 += counts->tx_size[0][i][0];
+        count4x4 += counts->tx_size[1][i][0];
+        count4x4 += counts->tx_size[2][i][0];
+        count4x4 += counts->tx_size[3][i][0];
+
+        count8x8_8x8p += counts->tx_size[0][i][1];
+        count8x8_lp += counts->tx_size[1][i][1];
+        count8x8_lp += counts->tx_size[2][i][1];
+        count8x8_lp += counts->tx_size[3][i][1];
+
+        count16x16_16x16p += counts->tx_size[1][i][2];
+        count16x16_lp += counts->tx_size[2][i][2];
+        count16x16_lp += counts->tx_size[3][i][2];
+
+        count32x32_32x32p += counts->tx_size[2][i][3];
+        count32x32_lp += counts->tx_size[3][i][3];
+
+        count64x64_64x64p += counts->tx_size[3][i][4];
+      }
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+      count4x4 += counts->tx_size_implied[0][TX_4X4];
+      count4x4 += counts->tx_size_implied[1][TX_4X4];
+      count4x4 += counts->tx_size_implied[2][TX_4X4];
+      count4x4 += counts->tx_size_implied[3][TX_4X4];
+      count8x8_8x8p += counts->tx_size_implied[1][TX_8X8];
+      count8x8_lp += counts->tx_size_implied[2][TX_8X8];
+      count8x8_lp += counts->tx_size_implied[3][TX_8X8];
+      count8x8_lp += counts->tx_size_implied[4][TX_8X8];
+      count16x16_16x16p += counts->tx_size_implied[2][TX_16X16];
+      count16x16_lp += counts->tx_size_implied[3][TX_16X16];
+      count16x16_lp += counts->tx_size_implied[4][TX_16X16];
+      count32x32_32x32p += counts->tx_size_implied[3][TX_32X32];
+      count32x32_lp += counts->tx_size_implied[4][TX_32X32];
+      count64x64_64x64p += counts->tx_size_implied[4][TX_64X64];
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+      if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
+          count32x32_lp == 0 && count32x32_32x32p == 0 &&
+#if CONFIG_SUPERTX
+          cm->counts.supertx_size[TX_16X16] == 0 &&
+          cm->counts.supertx_size[TX_32X32] == 0 &&
+          cm->counts.supertx_size[TX_64X64] == 0 &&
+#endif
+          count64x64_64x64p == 0) {
+        cm->tx_mode = ALLOW_8X8;
+        reset_skip_tx_size(cm, TX_8X8);
+      } else if (count8x8_8x8p == 0 && count8x8_lp == 0 &&
+                 count16x16_16x16p == 0 && count16x16_lp == 0 &&
+                 count32x32_32x32p == 0 && count32x32_lp == 0 &&
+#if CONFIG_SUPERTX
+                 cm->counts.supertx_size[TX_8X8] == 0 &&
+                 cm->counts.supertx_size[TX_16X16] == 0 &&
+                 cm->counts.supertx_size[TX_32X32] == 0 &&
+                 cm->counts.supertx_size[TX_64X64] == 0 &&
+#endif
+                 count64x64_64x64p == 0) {
+        cm->tx_mode = ONLY_4X4;
+        reset_skip_tx_size(cm, TX_4X4);
+      } else if (count4x4 == 0 && count8x8_lp == 0 && count16x16_lp == 0 &&
+                 count32x32_lp == 0) {
+        cm->tx_mode = ALLOW_64X64;
+      } else if (count4x4 == 0 && count8x8_lp == 0 && count16x16_lp == 0 &&
+#if CONFIG_SUPERTX
+                 cm->counts.supertx_size[TX_64X64] == 0 &&
+#endif
+                 count64x64_64x64p == 0) {
+        cm->tx_mode = ALLOW_32X32;
+        reset_skip_tx_size(cm, TX_32X32);
+      } else if (count4x4 == 0 && count8x8_lp == 0 && count32x32_lp == 0 &&
+                 count32x32_32x32p == 0 &&
+#if CONFIG_SUPERTX
+                 cm->counts.supertx_size[TX_32X32] == 0 &&
+                 cm->counts.supertx_size[TX_64X64] == 0 &&
+#endif
+                 count64x64_64x64p == 0) {
+        cm->tx_mode = ALLOW_16X16;
+        reset_skip_tx_size(cm, TX_16X16);
+      }
+
+#else  // CONFIG_TX64X64
+
+      int count4x4 = 0;
+      int count8x8_lp = 0, count8x8_8x8p = 0;
+      int count16x16_16x16p = 0, count16x16_lp = 0;
+      int count32x32 = 0;
+      for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
+        // counts->tx_size[max_depth][context_idx][this_depth_level]
+        count4x4 += counts->tx_size[0][i][0];
+        count4x4 += counts->tx_size[1][i][0];
+        count4x4 += counts->tx_size[2][i][0];
+
+        count8x8_8x8p += counts->tx_size[0][i][1];
+        count8x8_lp += counts->tx_size[1][i][1];
+        count8x8_lp += counts->tx_size[2][i][1];
+
+        count16x16_16x16p += counts->tx_size[1][i][2];
+        count16x16_lp += counts->tx_size[2][i][2];
+        count32x32 += counts->tx_size[2][i][3];
+      }
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+      count4x4 += counts->tx_size_implied[0][TX_4X4];
+      count4x4 += counts->tx_size_implied[1][TX_4X4];
+      count4x4 += counts->tx_size_implied[2][TX_4X4];
+      count4x4 += counts->tx_size_implied[3][TX_4X4];
+      count8x8_8x8p += counts->tx_size_implied[1][TX_8X8];
+      count8x8_lp += counts->tx_size_implied[2][TX_8X8];
+      count8x8_lp += counts->tx_size_implied[3][TX_8X8];
+      count16x16_lp += counts->tx_size_implied[3][TX_16X16];
+      count16x16_16x16p += counts->tx_size_implied[2][TX_16X16];
+      count32x32 += counts->tx_size_implied[3][TX_32X32];
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+      if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
+#if CONFIG_SUPERTX
+          cm->counts.supertx_size[TX_16X16] == 0 &&
+          cm->counts.supertx_size[TX_32X32] == 0 &&
+#endif  // CONFIG_SUPERTX
+          count32x32 == 0) {
+        cm->tx_mode = ALLOW_8X8;
+        reset_skip_tx_size(cm, TX_8X8);
+      } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
+                 count8x8_lp == 0 && count16x16_lp == 0 &&
+#if CONFIG_SUPERTX
+                 cm->counts.supertx_size[TX_8X8] == 0 &&
+                 cm->counts.supertx_size[TX_16X16] == 0 &&
+                 cm->counts.supertx_size[TX_32X32] == 0 &&
+#endif  // CONFIG_SUPERTX
+                 count32x32 == 0) {
+        cm->tx_mode = ONLY_4X4;
+        reset_skip_tx_size(cm, TX_4X4);
+      } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
+        cm->tx_mode = ALLOW_32X32;
+      } else if (count32x32 == 0 && count8x8_lp == 0 &&
+#if CONFIG_SUPERTX
+                 cm->counts.supertx_size[TX_32X32] == 0 &&
+#endif  // CONFIG_SUPERTX
+                 count4x4 == 0) {
+        cm->tx_mode = ALLOW_16X16;
+        reset_skip_tx_size(cm, TX_16X16);
+      }
+#endif  // CONFIG_TX64X64
+    }
+#endif
+  } else {
+    encode_frame_internal(cpi);
+  }
+}
+
+static void sum_intra_stats(FRAME_COUNTS *counts, MACROBLOCKD *xd,
+                            const MODE_INFO *mi, const MODE_INFO *above_mi,
+                            const MODE_INFO *left_mi, const int intraonly,
+                            const int mi_row, const int mi_col) {
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const PREDICTION_MODE y_mode = mbmi->mode;
+  const PREDICTION_MODE uv_mode = mbmi->uv_mode;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  const int unify_bsize = CONFIG_CB4X4;
+
+  if (bsize < BLOCK_8X8 && !unify_bsize) {
+    int idx, idy;
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
+    for (idy = 0; idy < 2; idy += num_4x4_h)
+      for (idx = 0; idx < 2; idx += num_4x4_w) {
+        const int bidx = idy * 2 + idx;
+        const PREDICTION_MODE bmode = mi->bmi[bidx].as_mode;
+        if (intraonly) {
+          const PREDICTION_MODE a = av1_above_block_mode(mi, above_mi, bidx);
+          const PREDICTION_MODE l = av1_left_block_mode(mi, left_mi, bidx);
+          ++counts->kf_y_mode[a][l][bmode];
+        } else {
+          ++counts->y_mode[0][bmode];
+        }
+      }
+  } else {
+    if (intraonly) {
+      const PREDICTION_MODE above = av1_above_block_mode(mi, above_mi, 0);
+      const PREDICTION_MODE left = av1_left_block_mode(mi, left_mi, 0);
+      ++counts->kf_y_mode[above][left][y_mode];
+    } else {
+      ++counts->y_mode[size_group_lookup[bsize]][y_mode];
+    }
+#if CONFIG_FILTER_INTRA
+    if (mbmi->mode == DC_PRED
+#if CONFIG_PALETTE
+        && mbmi->palette_mode_info.palette_size[0] == 0
+#endif  // CONFIG_PALETTE
+        ) {
+      const int use_filter_intra_mode =
+          mbmi->filter_intra_mode_info.use_filter_intra_mode[0];
+      ++counts->filter_intra[0][use_filter_intra_mode];
+    }
+    if (mbmi->uv_mode == DC_PRED
+#if CONFIG_PALETTE
+        && mbmi->palette_mode_info.palette_size[1] == 0
+#endif  // CONFIG_PALETTE
+        ) {
+      const int use_filter_intra_mode =
+          mbmi->filter_intra_mode_info.use_filter_intra_mode[1];
+      ++counts->filter_intra[1][use_filter_intra_mode];
+    }
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_EXT_INTRA && CONFIG_INTRA_INTERP
+    if (av1_is_directional_mode(mbmi->mode, bsize)) {
+      const int intra_filter_ctx = av1_get_pred_context_intra_interp(xd);
+      const int p_angle =
+          mode_to_angle_map[mbmi->mode] + mbmi->angle_delta[0] * ANGLE_STEP;
+      if (av1_is_intra_filter_switchable(p_angle))
+        ++counts->intra_filter[intra_filter_ctx][mbmi->intra_filter];
+    }
+#endif  // CONFIG_INTRA_INTERP && CONFIG_INTRA_INTERP
+  }
+
+#if CONFIG_CB4X4
+  if (!is_chroma_reference(mi_row, mi_col, bsize, xd->plane[1].subsampling_x,
+                           xd->plane[1].subsampling_y))
+    return;
+#else
+  (void)mi_row;
+  (void)mi_col;
+  (void)xd;
+#endif
+  ++counts->uv_mode[y_mode][uv_mode];
+}
+
+#if CONFIG_VAR_TX
+static void update_txfm_count(MACROBLOCK *x, MACROBLOCKD *xd,
+                              FRAME_COUNTS *counts, TX_SIZE tx_size, int depth,
+                              int blk_row, int blk_col) {
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const int tx_row = blk_row >> 1;
+  const int tx_col = blk_col >> 1;
+  const int max_blocks_high = max_block_high(xd, mbmi->sb_type, 0);
+  const int max_blocks_wide = max_block_wide(xd, mbmi->sb_type, 0);
+  int ctx = txfm_partition_context(xd->above_txfm_context + tx_col,
+                                   xd->left_txfm_context + tx_row,
+                                   mbmi->sb_type, tx_size);
+  const TX_SIZE plane_tx_size = mbmi->inter_tx_size[tx_row][tx_col];
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  if (tx_size == plane_tx_size) {
+    ++counts->txfm_partition[ctx][0];
+    mbmi->tx_size = tx_size;
+    txfm_partition_update(xd->above_txfm_context + tx_col,
+                          xd->left_txfm_context + tx_row, tx_size, tx_size);
+  } else {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bs = tx_size_wide_unit[sub_txs];
+    int i;
+
+    ++counts->txfm_partition[ctx][1];
+    ++x->txb_split_count;
+
+    if (tx_size == TX_8X8) {
+      mbmi->inter_tx_size[tx_row][tx_col] = TX_4X4;
+      mbmi->tx_size = TX_4X4;
+      txfm_partition_update(xd->above_txfm_context + tx_col,
+                            xd->left_txfm_context + tx_row, TX_4X4, tx_size);
+      return;
+    }
+
+    for (i = 0; i < 4; ++i) {
+      int offsetr = (i >> 1) * bs;
+      int offsetc = (i & 0x01) * bs;
+      update_txfm_count(x, xd, counts, sub_txs, depth + 1, blk_row + offsetr,
+                        blk_col + offsetc);
+    }
+  }
+}
+
+static void tx_partition_count_update(const AV1_COMMON *const cm, MACROBLOCK *x,
+                                      BLOCK_SIZE plane_bsize, int mi_row,
+                                      int mi_col, FRAME_COUNTS *td_counts) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+  const int mi_height = block_size_high[plane_bsize] >> tx_size_wide_log2[0];
+  TX_SIZE max_tx_size = get_vartx_max_txsize(&xd->mi[0]->mbmi, plane_bsize);
+  const int bh = tx_size_high_unit[max_tx_size];
+  const int bw = tx_size_wide_unit[max_tx_size];
+  int idx, idy;
+
+  xd->above_txfm_context = cm->above_txfm_context + mi_col;
+  xd->left_txfm_context =
+      xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+
+  for (idy = 0; idy < mi_height; idy += bh)
+    for (idx = 0; idx < mi_width; idx += bw)
+      update_txfm_count(x, xd, td_counts, max_tx_size, mi_width != mi_height,
+                        idy, idx);
+}
+
+static void set_txfm_context(MACROBLOCKD *xd, TX_SIZE tx_size, int blk_row,
+                             int blk_col) {
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const int tx_row = blk_row >> 1;
+  const int tx_col = blk_col >> 1;
+  const int max_blocks_high = max_block_high(xd, mbmi->sb_type, 0);
+  const int max_blocks_wide = max_block_wide(xd, mbmi->sb_type, 0);
+  const TX_SIZE plane_tx_size = mbmi->inter_tx_size[tx_row][tx_col];
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  if (tx_size == plane_tx_size) {
+    mbmi->tx_size = tx_size;
+    txfm_partition_update(xd->above_txfm_context + tx_col,
+                          xd->left_txfm_context + tx_row, tx_size, tx_size);
+
+  } else {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int i;
+
+    if (tx_size == TX_8X8) {
+      mbmi->inter_tx_size[tx_row][tx_col] = TX_4X4;
+      mbmi->tx_size = TX_4X4;
+      txfm_partition_update(xd->above_txfm_context + tx_col,
+                            xd->left_txfm_context + tx_row, TX_4X4, tx_size);
+      return;
+    }
+
+    assert(bsl > 0);
+    for (i = 0; i < 4; ++i) {
+      int offsetr = (i >> 1) * bsl;
+      int offsetc = (i & 0x01) * bsl;
+      set_txfm_context(xd, sub_txs, blk_row + offsetr, blk_col + offsetc);
+    }
+  }
+}
+
+static void tx_partition_set_contexts(const AV1_COMMON *const cm,
+                                      MACROBLOCKD *xd, BLOCK_SIZE plane_bsize,
+                                      int mi_row, int mi_col) {
+  const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+  const int mi_height = block_size_high[plane_bsize] >> tx_size_high_log2[0];
+  TX_SIZE max_tx_size = get_vartx_max_txsize(&xd->mi[0]->mbmi, plane_bsize);
+  const int bh = tx_size_high_unit[max_tx_size];
+  const int bw = tx_size_wide_unit[max_tx_size];
+  int idx, idy;
+
+  xd->above_txfm_context = cm->above_txfm_context + mi_col;
+  xd->left_txfm_context =
+      xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
+
+  for (idy = 0; idy < mi_height; idy += bh)
+    for (idx = 0; idx < mi_width; idx += bw)
+      set_txfm_context(xd, max_tx_size, idy, idx);
+}
+#endif
+
+void av1_update_tx_type_count(const AV1_COMMON *cm, MACROBLOCKD *xd,
+#if CONFIG_TXK_SEL
+                              int block, int plane,
+#endif
+                              BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              FRAME_COUNTS *counts) {
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  int is_inter = is_inter_block(mbmi);
+#if !CONFIG_TXK_SEL
+  TX_TYPE tx_type = mbmi->tx_type;
+#else
+  // Only y plane's tx_type is updated
+  if (plane > 0) return;
+  TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block, tx_size);
+#endif
+#if CONFIG_EXT_TX
+  if (get_ext_tx_types(tx_size, bsize, is_inter, cm->reduced_tx_set_used) > 1 &&
+      cm->base_qindex > 0 && !mbmi->skip &&
+      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+    const int eset =
+        get_ext_tx_set(tx_size, bsize, is_inter, cm->reduced_tx_set_used);
+    if (eset > 0) {
+      if (is_inter) {
+        ++counts->inter_ext_tx[eset][txsize_sqr_map[tx_size]][tx_type];
+      } else {
+        ++counts->intra_ext_tx[eset][txsize_sqr_map[tx_size]][mbmi->mode]
+                              [tx_type];
+      }
+    }
+  }
+#else
+  (void)bsize;
+  if (tx_size < TX_32X32 &&
+      ((!cm->seg.enabled && cm->base_qindex > 0) ||
+       (cm->seg.enabled && xd->qindex[mbmi->segment_id] > 0)) &&
+      !mbmi->skip &&
+      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
+    if (is_inter) {
+      ++counts->inter_ext_tx[tx_size][tx_type];
+    } else {
+      ++counts->intra_ext_tx[tx_size][intra_mode_to_tx_type_context[mbmi->mode]]
+                            [tx_type];
+    }
+  }
+#endif  // CONFIG_EXT_TX
+}
+
+static void encode_superblock(const AV1_COMP *const cpi, ThreadData *td,
+                              TOKENEXTRA **t, RUN_TYPE dry_run, int mi_row,
+                              int mi_col, BLOCK_SIZE bsize,
+                              PICK_MODE_CONTEXT *ctx, int *rate) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO **mi_8x8 = xd->mi;
+  MODE_INFO *mi = mi_8x8[0];
+  MB_MODE_INFO *mbmi = &mi->mbmi;
+  const int seg_skip =
+      segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP);
+  const int mis = cm->mi_stride;
+  const int mi_width = mi_size_wide[bsize];
+  const int mi_height = mi_size_high[bsize];
+  const int is_inter = is_inter_block(mbmi);
+#if CONFIG_CB4X4
+  const BLOCK_SIZE block_size = bsize;
+#else
+  const BLOCK_SIZE block_size = AOMMAX(bsize, BLOCK_8X8);
+#endif
+
+#if CONFIG_PVQ
+  x->pvq_speed = 0;
+  x->pvq_coded = (dry_run == OUTPUT_ENABLED) ? 1 : 0;
+#endif
+#if CONFIG_CFL
+  x->cfl_store_y = (dry_run == OUTPUT_ENABLED) ? 1 : 0;
+#endif
+
+  if (!is_inter) {
+    int plane;
+    mbmi->skip = 1;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+      av1_encode_intra_block_plane((AV1_COMMON *)cm, x, block_size, plane, 1,
+                                   mi_row, mi_col);
+    }
+    if (!dry_run) {
+      sum_intra_stats(td->counts, xd, mi, xd->above_mi, xd->left_mi,
+                      frame_is_intra_only(cm), mi_row, mi_col);
+    }
+#if CONFIG_PALETTE
+    if (bsize >= BLOCK_8X8 && !dry_run) {
+      for (plane = 0; plane <= 1; ++plane) {
+        if (mbmi->palette_mode_info.palette_size[plane] > 0) {
+          mbmi->palette_mode_info.palette_first_color_idx[plane] =
+              xd->plane[plane].color_index_map[0];
+          // TODO(huisu): this increases the use of token buffer. Needs stretch
+          // test to verify.
+          av1_tokenize_palette_sb(cpi, td, plane, t, dry_run, bsize, rate);
+        }
+      }
+    }
+#endif  // CONFIG_PALETTE
+#if CONFIG_VAR_TX
+    mbmi->min_tx_size = get_min_tx_size(mbmi->tx_size);
+#endif
+#if CONFIG_LV_MAP
+    av1_update_txb_context(cpi, td, dry_run, block_size, rate, mi_row, mi_col);
+#else   // CONFIG_LV_MAP
+    av1_tokenize_sb(cpi, td, t, dry_run, block_size, rate, mi_row, mi_col);
+#endif  // CONFIG_LV_MAP
+  } else {
+    int ref;
+    const int is_compound = has_second_ref(mbmi);
+
+    set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
+    for (ref = 0; ref < 1 + is_compound; ++ref) {
+      YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, mbmi->ref_frame[ref]);
+#if CONFIG_INTRABC
+      assert(IMPLIES(!is_intrabc_block(mbmi), cfg));
+#else
+      assert(cfg != NULL);
+#endif  // !CONFIG_INTRABC
+      av1_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
+                           &xd->block_refs[ref]->sf);
+    }
+    if (!(cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready) || seg_skip)
+      av1_build_inter_predictors_sby(xd, mi_row, mi_col, NULL, block_size);
+
+    av1_build_inter_predictors_sbuv(xd, mi_row, mi_col, NULL, block_size);
+#if CONFIG_MOTION_VAR
+    if (mbmi->motion_mode == OBMC_CAUSAL) {
+#if CONFIG_NCOBMC
+      if (dry_run == OUTPUT_ENABLED)
+        av1_build_ncobmc_inter_predictors_sb(cm, xd, mi_row, mi_col);
+      else
+#endif
+        av1_build_obmc_inter_predictors_sb(cm, xd, mi_row, mi_col);
+    }
+#endif  // CONFIG_MOTION_VAR
+
+    av1_encode_sb((AV1_COMMON *)cm, x, block_size, mi_row, mi_col);
+#if CONFIG_VAR_TX
+    if (mbmi->skip) mbmi->min_tx_size = get_min_tx_size(mbmi->tx_size);
+    av1_tokenize_sb_vartx(cpi, td, t, dry_run, mi_row, mi_col, block_size,
+                          rate);
+#else
+#if CONFIG_LV_MAP
+    av1_update_txb_context(cpi, td, dry_run, block_size, rate, mi_row, mi_col);
+#else   // CONFIG_LV_MAP
+    av1_tokenize_sb(cpi, td, t, dry_run, block_size, rate, mi_row, mi_col);
+#endif  // CONFIG_LV_MAP
+#endif
+  }
+
+  if (!dry_run) {
+#if CONFIG_VAR_TX
+    TX_SIZE tx_size =
+        is_inter && !mbmi->skip ? mbmi->min_tx_size : mbmi->tx_size;
+#else
+    TX_SIZE tx_size = mbmi->tx_size;
+#endif
+    if (cm->tx_mode == TX_MODE_SELECT && !xd->lossless[mbmi->segment_id] &&
+#if CONFIG_CB4X4 && (CONFIG_VAR_TX || CONFIG_EXT_TX) && CONFIG_RECT_TX
+        mbmi->sb_type > BLOCK_4X4 &&
+#else
+        mbmi->sb_type >= BLOCK_8X8 &&
+#endif
+        !(is_inter && (mbmi->skip || seg_skip))) {
+#if CONFIG_VAR_TX
+      if (is_inter) {
+        tx_partition_count_update(cm, x, bsize, mi_row, mi_col, td->counts);
+      } else {
+        const int tx_size_ctx = get_tx_size_context(xd);
+        const int tx_size_cat = is_inter ? inter_tx_size_cat_lookup[bsize]
+                                         : intra_tx_size_cat_lookup[bsize];
+        const TX_SIZE coded_tx_size = txsize_sqr_up_map[tx_size];
+        const int depth = tx_size_to_depth(coded_tx_size);
+        ++td->counts->tx_size[tx_size_cat][tx_size_ctx][depth];
+        if (tx_size != max_txsize_lookup[bsize]) ++x->txb_split_count;
+      }
+#else
+      const int tx_size_ctx = get_tx_size_context(xd);
+      const int tx_size_cat = is_inter ? inter_tx_size_cat_lookup[bsize]
+                                       : intra_tx_size_cat_lookup[bsize];
+      const TX_SIZE coded_tx_size = txsize_sqr_up_map[tx_size];
+      const int depth = tx_size_to_depth(coded_tx_size);
+
+      ++td->counts->tx_size[tx_size_cat][tx_size_ctx][depth];
+#endif
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+      assert(IMPLIES(is_rect_tx(tx_size), is_rect_tx_allowed(xd, mbmi)));
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+    } else {
+      int i, j;
+      TX_SIZE intra_tx_size;
+      // The new intra coding scheme requires no change of transform size
+      if (is_inter) {
+        if (xd->lossless[mbmi->segment_id]) {
+          intra_tx_size = TX_4X4;
+        } else {
+          intra_tx_size = tx_size_from_tx_mode(bsize, cm->tx_mode, 1);
+        }
+      } else {
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+        intra_tx_size = tx_size;
+#else
+        intra_tx_size = (bsize >= BLOCK_8X8) ? tx_size : TX_4X4;
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+      }
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+      ++td->counts->tx_size_implied[max_txsize_lookup[bsize]]
+                                   [txsize_sqr_up_map[tx_size]];
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+
+      for (j = 0; j < mi_height; j++)
+        for (i = 0; i < mi_width; i++)
+          if (mi_col + i < cm->mi_cols && mi_row + j < cm->mi_rows)
+            mi_8x8[mis * j + i]->mbmi.tx_size = intra_tx_size;
+
+#if CONFIG_VAR_TX
+      mbmi->min_tx_size = get_min_tx_size(intra_tx_size);
+      if (intra_tx_size != max_txsize_lookup[bsize]) ++x->txb_split_count;
+#endif
+    }
+
+    ++td->counts->tx_size_totals[txsize_sqr_map[tx_size]];
+    ++td->counts
+          ->tx_size_totals[txsize_sqr_map[get_uv_tx_size(mbmi, &xd->plane[1])]];
+#if !CONFIG_TXK_SEL
+    av1_update_tx_type_count(cm, xd, bsize, tx_size, td->counts);
+#endif
+  }
+
+#if CONFIG_VAR_TX
+  if (cm->tx_mode == TX_MODE_SELECT &&
+#if CONFIG_CB4X4
+      mbmi->sb_type > BLOCK_4X4 &&
+#else
+      mbmi->sb_type >= BLOCK_8X8 &&
+#endif
+      is_inter && !(mbmi->skip || seg_skip)) {
+    if (dry_run) tx_partition_set_contexts(cm, xd, bsize, mi_row, mi_col);
+  } else {
+    TX_SIZE tx_size = mbmi->tx_size;
+    // The new intra coding scheme requires no change of transform size
+    if (is_inter)
+      tx_size = tx_size_from_tx_mode(bsize, cm->tx_mode, is_inter);
+    else
+      tx_size = (bsize > BLOCK_4X4) ? tx_size : TX_4X4;
+    mbmi->tx_size = tx_size;
+    set_txfm_ctxs(tx_size, xd->n8_w, xd->n8_h, (mbmi->skip || seg_skip), xd);
+  }
+#endif  // CONFIG_VAR_TX
+}
+
+#if CONFIG_SUPERTX
+static int check_intra_b(PICK_MODE_CONTEXT *ctx) {
+  if (!is_inter_mode((&ctx->mic)->mbmi.mode)) return 1;
+#if CONFIG_EXT_INTER
+  if (ctx->mic.mbmi.ref_frame[1] == INTRA_FRAME) return 1;
+#endif  // CONFIG_EXT_INTER
+  return 0;
+}
+
+static int check_intra_sb(const AV1_COMP *const cpi, const TileInfo *const tile,
+                          int mi_row, int mi_col, BLOCK_SIZE bsize,
+                          PC_TREE *pc_tree) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int hbs = mi_size_wide[bsize] / 2;
+  const PARTITION_TYPE partition = pc_tree->partitioning;
+  const BLOCK_SIZE subsize = get_subsize(bsize, partition);
+#if CONFIG_EXT_PARTITION_TYPES
+  int i;
+#endif
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+#if !CONFIG_CB4X4
+  assert(bsize >= BLOCK_8X8);
+#endif
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return 1;
+
+  switch (partition) {
+    case PARTITION_NONE: return check_intra_b(&pc_tree->none); break;
+    case PARTITION_VERT:
+      if (check_intra_b(&pc_tree->vertical[0])) return 1;
+      if (mi_col + hbs < cm->mi_cols && (bsize > BLOCK_8X8 || unify_bsize)) {
+        if (check_intra_b(&pc_tree->vertical[1])) return 1;
+      }
+      break;
+    case PARTITION_HORZ:
+      if (check_intra_b(&pc_tree->horizontal[0])) return 1;
+      if (mi_row + hbs < cm->mi_rows && (bsize > BLOCK_8X8 || unify_bsize)) {
+        if (check_intra_b(&pc_tree->horizontal[1])) return 1;
+      }
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        if (check_intra_b(pc_tree->leaf_split[0])) return 1;
+      } else {
+        if (check_intra_sb(cpi, tile, mi_row, mi_col, subsize,
+                           pc_tree->split[0]))
+          return 1;
+        if (check_intra_sb(cpi, tile, mi_row, mi_col + hbs, subsize,
+                           pc_tree->split[1]))
+          return 1;
+        if (check_intra_sb(cpi, tile, mi_row + hbs, mi_col, subsize,
+                           pc_tree->split[2]))
+          return 1;
+        if (check_intra_sb(cpi, tile, mi_row + hbs, mi_col + hbs, subsize,
+                           pc_tree->split[3]))
+          return 1;
+      }
+      break;
+#if CONFIG_EXT_PARTITION_TYPES
+    case PARTITION_HORZ_A:
+      for (i = 0; i < 3; i++) {
+        if (check_intra_b(&pc_tree->horizontala[i])) return 1;
+      }
+      break;
+    case PARTITION_HORZ_B:
+      for (i = 0; i < 3; i++) {
+        if (check_intra_b(&pc_tree->horizontalb[i])) return 1;
+      }
+      break;
+    case PARTITION_VERT_A:
+      for (i = 0; i < 3; i++) {
+        if (check_intra_b(&pc_tree->verticala[i])) return 1;
+      }
+      break;
+    case PARTITION_VERT_B:
+      for (i = 0; i < 3; i++) {
+        if (check_intra_b(&pc_tree->verticalb[i])) return 1;
+      }
+      break;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+    default: assert(0);
+  }
+  return 0;
+}
+
+static int check_supertx_b(TX_SIZE supertx_size, PICK_MODE_CONTEXT *ctx) {
+  return ctx->mic.mbmi.tx_size == supertx_size;
+}
+
+static int check_supertx_sb(BLOCK_SIZE bsize, TX_SIZE supertx_size,
+                            PC_TREE *pc_tree) {
+  PARTITION_TYPE partition;
+  BLOCK_SIZE subsize;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+  partition = pc_tree->partitioning;
+  subsize = get_subsize(bsize, partition);
+  switch (partition) {
+    case PARTITION_NONE: return check_supertx_b(supertx_size, &pc_tree->none);
+    case PARTITION_VERT:
+      return check_supertx_b(supertx_size, &pc_tree->vertical[0]);
+    case PARTITION_HORZ:
+      return check_supertx_b(supertx_size, &pc_tree->horizontal[0]);
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8 && !unify_bsize)
+        return check_supertx_b(supertx_size, pc_tree->leaf_split[0]);
+      else
+        return check_supertx_sb(subsize, supertx_size, pc_tree->split[0]);
+#if CONFIG_EXT_PARTITION_TYPES
+    case PARTITION_HORZ_A:
+      return check_supertx_b(supertx_size, &pc_tree->horizontala[0]);
+    case PARTITION_HORZ_B:
+      return check_supertx_b(supertx_size, &pc_tree->horizontalb[0]);
+    case PARTITION_VERT_A:
+      return check_supertx_b(supertx_size, &pc_tree->verticala[0]);
+    case PARTITION_VERT_B:
+      return check_supertx_b(supertx_size, &pc_tree->verticalb[0]);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+    default: assert(0); return 0;
+  }
+}
+
+static void predict_superblock(const AV1_COMP *const cpi, ThreadData *td,
+#if CONFIG_EXT_INTER
+                               int mi_row_ori, int mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                               int mi_row_pred, int mi_col_pred,
+                               BLOCK_SIZE bsize_pred, int b_sub8x8, int block) {
+  // Used in supertx
+  // (mi_row_ori, mi_col_ori): location for mv
+  // (mi_row_pred, mi_col_pred, bsize_pred): region to predict
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *mi_8x8 = xd->mi[0];
+  MODE_INFO *mi = mi_8x8;
+  MB_MODE_INFO *mbmi = &mi->mbmi;
+  int ref;
+  const int is_compound = has_second_ref(mbmi);
+
+  set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
+
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, mbmi->ref_frame[ref]);
+    av1_setup_pre_planes(xd, ref, cfg, mi_row_pred, mi_col_pred,
+                         &xd->block_refs[ref]->sf);
+  }
+
+  if (!b_sub8x8)
+    av1_build_inter_predictors_sb_extend(xd,
+#if CONFIG_EXT_INTER
+                                         mi_row_ori, mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                                         mi_row_pred, mi_col_pred, bsize_pred);
+  else
+    av1_build_inter_predictors_sb_sub8x8_extend(xd,
+#if CONFIG_EXT_INTER
+                                                mi_row_ori, mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                                                mi_row_pred, mi_col_pred,
+                                                bsize_pred, block);
+}
+
+static void predict_b_extend(const AV1_COMP *const cpi, ThreadData *td,
+                             const TileInfo *const tile, int block,
+                             int mi_row_ori, int mi_col_ori, int mi_row_pred,
+                             int mi_col_pred, int mi_row_top, int mi_col_top,
+                             uint8_t *dst_buf[3], int dst_stride[3],
+                             BLOCK_SIZE bsize_top, BLOCK_SIZE bsize_pred,
+                             RUN_TYPE dry_run, int b_sub8x8, int bextend) {
+  // Used in supertx
+  // (mi_row_ori, mi_col_ori): location for mv
+  // (mi_row_pred, mi_col_pred, bsize_pred): region to predict
+  // (mi_row_top, mi_col_top, bsize_top): region of the top partition size
+  // block: sub location of sub8x8 blocks
+  // b_sub8x8: 1: ori is sub8x8; 0: ori is not sub8x8
+  // bextend: 1: region to predict is an extension of ori; 0: not
+
+  MACROBLOCK *const x = &td->mb;
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int r = (mi_row_pred - mi_row_top) * MI_SIZE;
+  int c = (mi_col_pred - mi_col_top) * MI_SIZE;
+  const int mi_width_top = mi_size_wide[bsize_top];
+  const int mi_height_top = mi_size_high[bsize_top];
+
+  if (mi_row_pred < mi_row_top || mi_col_pred < mi_col_top ||
+      mi_row_pred >= mi_row_top + mi_height_top ||
+      mi_col_pred >= mi_col_top + mi_width_top || mi_row_pred >= cm->mi_rows ||
+      mi_col_pred >= cm->mi_cols)
+    return;
+
+  set_offsets_extend(cpi, td, tile, mi_row_pred, mi_col_pred, mi_row_ori,
+                     mi_col_ori, bsize_pred);
+  xd->plane[0].dst.stride = dst_stride[0];
+  xd->plane[1].dst.stride = dst_stride[1];
+  xd->plane[2].dst.stride = dst_stride[2];
+  xd->plane[0].dst.buf = dst_buf[0] +
+                         (r >> xd->plane[0].subsampling_y) * dst_stride[0] +
+                         (c >> xd->plane[0].subsampling_x);
+  xd->plane[1].dst.buf = dst_buf[1] +
+                         (r >> xd->plane[1].subsampling_y) * dst_stride[1] +
+                         (c >> xd->plane[1].subsampling_x);
+  xd->plane[2].dst.buf = dst_buf[2] +
+                         (r >> xd->plane[2].subsampling_y) * dst_stride[2] +
+                         (c >> xd->plane[2].subsampling_x);
+
+  predict_superblock(cpi, td,
+#if CONFIG_EXT_INTER
+                     mi_row_ori, mi_col_ori,
+#endif  // CONFIG_EXT_INTER
+                     mi_row_pred, mi_col_pred, bsize_pred, b_sub8x8, block);
+
+  if (!dry_run && !bextend)
+    update_stats(&cpi->common, td, mi_row_pred, mi_col_pred, 1);
+}
+
+static void extend_dir(const AV1_COMP *const cpi, ThreadData *td,
+                       const TileInfo *const tile, int block, BLOCK_SIZE bsize,
+                       BLOCK_SIZE top_bsize, int mi_row, int mi_col,
+                       int mi_row_top, int mi_col_top, RUN_TYPE dry_run,
+                       uint8_t *dst_buf[3], int dst_stride[3], int dir) {
+  // dir: 0-lower, 1-upper, 2-left, 3-right
+  //      4-lowerleft, 5-upperleft, 6-lowerright, 7-upperright
+  MACROBLOCKD *xd = &td->mb.e_mbd;
+  const int mi_width = mi_size_wide[bsize];
+  const int mi_height = mi_size_high[bsize];
+  int xss = xd->plane[1].subsampling_x;
+  int yss = xd->plane[1].subsampling_y;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+  int b_sub8x8 = (bsize < BLOCK_8X8) && !unify_bsize ? 1 : 0;
+  int wide_unit, high_unit;
+  int i, j;
+  int ext_offset = 0;
+
+  BLOCK_SIZE extend_bsize;
+  int mi_row_pred, mi_col_pred;
+
+  if (dir == 0 || dir == 1) {  // lower and upper
+    extend_bsize =
+        (mi_width == mi_size_wide[BLOCK_8X8] || bsize < BLOCK_8X8 || xss < yss)
+            ? BLOCK_8X8
+            : BLOCK_16X8;
+
+#if CONFIG_CB4X4
+    if (bsize < BLOCK_8X8) {
+      extend_bsize = BLOCK_4X4;
+      ext_offset = mi_size_wide[BLOCK_8X8];
+    }
+#endif
+    wide_unit = mi_size_wide[extend_bsize];
+    high_unit = mi_size_high[extend_bsize];
+
+    mi_row_pred = mi_row + ((dir == 0) ? mi_height : -(mi_height + ext_offset));
+    mi_col_pred = mi_col;
+
+    for (j = 0; j < mi_height + ext_offset; j += high_unit)
+      for (i = 0; i < mi_width + ext_offset; i += wide_unit)
+        predict_b_extend(cpi, td, tile, block, mi_row, mi_col, mi_row_pred + j,
+                         mi_col_pred + i, mi_row_top, mi_col_top, dst_buf,
+                         dst_stride, top_bsize, extend_bsize, dry_run, b_sub8x8,
+                         1);
+  } else if (dir == 2 || dir == 3) {  // left and right
+    extend_bsize =
+        (mi_height == mi_size_high[BLOCK_8X8] || bsize < BLOCK_8X8 || yss < xss)
+            ? BLOCK_8X8
+            : BLOCK_8X16;
+#if CONFIG_CB4X4
+    if (bsize < BLOCK_8X8) {
+      extend_bsize = BLOCK_4X4;
+      ext_offset = mi_size_wide[BLOCK_8X8];
+    }
+#endif
+    wide_unit = mi_size_wide[extend_bsize];
+    high_unit = mi_size_high[extend_bsize];
+
+    mi_row_pred = mi_row;
+    mi_col_pred = mi_col + ((dir == 3) ? mi_width : -(mi_width + ext_offset));
+
+    for (j = 0; j < mi_height + ext_offset; j += high_unit)
+      for (i = 0; i < mi_width + ext_offset; i += wide_unit)
+        predict_b_extend(cpi, td, tile, block, mi_row, mi_col, mi_row_pred + j,
+                         mi_col_pred + i, mi_row_top, mi_col_top, dst_buf,
+                         dst_stride, top_bsize, extend_bsize, dry_run, b_sub8x8,
+                         1);
+  } else {
+    extend_bsize = BLOCK_8X8;
+#if CONFIG_CB4X4
+    if (bsize < BLOCK_8X8) {
+      extend_bsize = BLOCK_4X4;
+      ext_offset = mi_size_wide[BLOCK_8X8];
+    }
+#endif
+    wide_unit = mi_size_wide[extend_bsize];
+    high_unit = mi_size_high[extend_bsize];
+
+    mi_row_pred = mi_row + ((dir == 4 || dir == 6) ? mi_height
+                                                   : -(mi_height + ext_offset));
+    mi_col_pred =
+        mi_col + ((dir == 6 || dir == 7) ? mi_width : -(mi_width + ext_offset));
+
+    for (j = 0; j < mi_height + ext_offset; j += high_unit)
+      for (i = 0; i < mi_width + ext_offset; i += wide_unit)
+        predict_b_extend(cpi, td, tile, block, mi_row, mi_col, mi_row_pred + j,
+                         mi_col_pred + i, mi_row_top, mi_col_top, dst_buf,
+                         dst_stride, top_bsize, extend_bsize, dry_run, b_sub8x8,
+                         1);
+  }
+}
+
+static void extend_all(const AV1_COMP *const cpi, ThreadData *td,
+                       const TileInfo *const tile, int block, BLOCK_SIZE bsize,
+                       BLOCK_SIZE top_bsize, int mi_row, int mi_col,
+                       int mi_row_top, int mi_col_top, RUN_TYPE dry_run,
+                       uint8_t *dst_buf[3], int dst_stride[3]) {
+  assert(block >= 0 && block < 4);
+  extend_dir(cpi, td, tile, block, bsize, top_bsize, mi_row, mi_col, mi_row_top,
+             mi_col_top, dry_run, dst_buf, dst_stride, 0);
+  extend_dir(cpi, td, tile, block, bsize, top_bsize, mi_row, mi_col, mi_row_top,
+             mi_col_top, dry_run, dst_buf, dst_stride, 1);
+  extend_dir(cpi, td, tile, block, bsize, top_bsize, mi_row, mi_col, mi_row_top,
+             mi_col_top, dry_run, dst_buf, dst_stride, 2);
+  extend_dir(cpi, td, tile, block, bsize, top_bsize, mi_row, mi_col, mi_row_top,
+             mi_col_top, dry_run, dst_buf, dst_stride, 3);
+  extend_dir(cpi, td, tile, block, bsize, top_bsize, mi_row, mi_col, mi_row_top,
+             mi_col_top, dry_run, dst_buf, dst_stride, 4);
+  extend_dir(cpi, td, tile, block, bsize, top_bsize, mi_row, mi_col, mi_row_top,
+             mi_col_top, dry_run, dst_buf, dst_stride, 5);
+  extend_dir(cpi, td, tile, block, bsize, top_bsize, mi_row, mi_col, mi_row_top,
+             mi_col_top, dry_run, dst_buf, dst_stride, 6);
+  extend_dir(cpi, td, tile, block, bsize, top_bsize, mi_row, mi_col, mi_row_top,
+             mi_col_top, dry_run, dst_buf, dst_stride, 7);
+}
+
+// This function generates prediction for multiple blocks, between which
+// discontinuity around boundary is reduced by smoothing masks. The basic
+// smoothing mask is a soft step function along horz/vert direction. In more
+// complicated case when a block is split into 4 subblocks, the basic mask is
+// first applied to neighboring subblocks (2 pairs) in horizontal direction and
+// then applied to the 2 masked prediction mentioned above in vertical direction
+// If the block is split into more than one level, at every stage, masked
+// prediction is stored in dst_buf[] passed from higher level.
+static void predict_sb_complex(const AV1_COMP *const cpi, ThreadData *td,
+                               const TileInfo *const tile, int mi_row,
+                               int mi_col, int mi_row_top, int mi_col_top,
+                               RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                               BLOCK_SIZE top_bsize, uint8_t *dst_buf[3],
+                               int dst_stride[3], PC_TREE *pc_tree) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const int hbs = mi_size_wide[bsize] / 2;
+  const int is_partition_root = bsize >= BLOCK_8X8;
+  const int ctx = is_partition_root
+                      ? partition_plane_context(xd, mi_row, mi_col,
+#if CONFIG_UNPOISON_PARTITION_CTX
+                                                mi_row + hbs < cm->mi_rows,
+                                                mi_col + hbs < cm->mi_cols,
+#endif
+                                                bsize)
+                      : -1;
+  const PARTITION_TYPE partition = pc_tree->partitioning;
+  const BLOCK_SIZE subsize = get_subsize(bsize, partition);
+#if CONFIG_EXT_PARTITION_TYPES
+  const BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
+#endif
+
+  int i;
+  uint8_t *dst_buf1[3], *dst_buf2[3], *dst_buf3[3];
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf1[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf2[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf3[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
+  int dst_stride1[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
+  int dst_stride2[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
+  int dst_stride3[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+  assert(bsize >= BLOCK_8X8);
+#endif
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    int len = sizeof(uint16_t);
+    dst_buf1[0] = CONVERT_TO_BYTEPTR(tmp_buf1);
+    dst_buf1[1] = CONVERT_TO_BYTEPTR(tmp_buf1 + MAX_TX_SQUARE * len);
+    dst_buf1[2] = CONVERT_TO_BYTEPTR(tmp_buf1 + 2 * MAX_TX_SQUARE * len);
+    dst_buf2[0] = CONVERT_TO_BYTEPTR(tmp_buf2);
+    dst_buf2[1] = CONVERT_TO_BYTEPTR(tmp_buf2 + MAX_TX_SQUARE * len);
+    dst_buf2[2] = CONVERT_TO_BYTEPTR(tmp_buf2 + 2 * MAX_TX_SQUARE * len);
+    dst_buf3[0] = CONVERT_TO_BYTEPTR(tmp_buf3);
+    dst_buf3[1] = CONVERT_TO_BYTEPTR(tmp_buf3 + MAX_TX_SQUARE * len);
+    dst_buf3[2] = CONVERT_TO_BYTEPTR(tmp_buf3 + 2 * MAX_TX_SQUARE * len);
+  } else {
+#endif  // CONFIG_HIGHBITDEPTH
+    dst_buf1[0] = tmp_buf1;
+    dst_buf1[1] = tmp_buf1 + MAX_TX_SQUARE;
+    dst_buf1[2] = tmp_buf1 + 2 * MAX_TX_SQUARE;
+    dst_buf2[0] = tmp_buf2;
+    dst_buf2[1] = tmp_buf2 + MAX_TX_SQUARE;
+    dst_buf2[2] = tmp_buf2 + 2 * MAX_TX_SQUARE;
+    dst_buf3[0] = tmp_buf3;
+    dst_buf3[1] = tmp_buf3 + MAX_TX_SQUARE;
+    dst_buf3[2] = tmp_buf3 + 2 * MAX_TX_SQUARE;
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  if (!dry_run && ctx >= 0 && bsize < top_bsize) {
+    // Explicitly cast away const.
+    FRAME_COUNTS *const frame_counts = (FRAME_COUNTS *)&cm->counts;
+    frame_counts->partition[ctx][partition]++;
+  }
+
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    xd->plane[i].dst.buf = dst_buf[i];
+    xd->plane[i].dst.stride = dst_stride[i];
+  }
+
+  switch (partition) {
+    case PARTITION_NONE:
+      assert(bsize < top_bsize);
+      predict_b_extend(cpi, td, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride, top_bsize,
+                       bsize, dry_run, 0, 0);
+      extend_all(cpi, td, tile, 0, bsize, top_bsize, mi_row, mi_col, mi_row_top,
+                 mi_col_top, dry_run, dst_buf, dst_stride);
+      break;
+    case PARTITION_HORZ:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        // Fisrt half
+        predict_b_extend(cpi, td, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride, top_bsize,
+                         BLOCK_8X8, dry_run, 1, 0);
+        if (bsize < top_bsize)
+          extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride);
+
+        // Second half
+        predict_b_extend(cpi, td, tile, 2, mi_row, mi_col, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                         top_bsize, BLOCK_8X8, dry_run, 1, 1);
+        if (bsize < top_bsize)
+          extend_all(cpi, td, tile, 2, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf1, dst_stride1);
+
+        // Smooth
+        xd->plane[0].dst.buf = dst_buf[0];
+        xd->plane[0].dst.stride = dst_stride[0];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[0], dst_stride[0], dst_buf1[0], dst_stride1[0], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            0);
+      } else {
+        // First half
+        predict_b_extend(cpi, td, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride, top_bsize,
+                         subsize, dry_run, 0, 0);
+        if (bsize < top_bsize)
+          extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride);
+        else
+          extend_dir(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride, 0);
+
+        if (mi_row + hbs < cm->mi_rows) {
+          // Second half
+          predict_b_extend(cpi, td, tile, 0, mi_row + hbs, mi_col, mi_row + hbs,
+                           mi_col, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, top_bsize, subsize, dry_run, 0, 0);
+          if (bsize < top_bsize)
+            extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row + hbs,
+                       mi_col, mi_row_top, mi_col_top, dry_run, dst_buf1,
+                       dst_stride1);
+          else
+            extend_dir(cpi, td, tile, 0, subsize, top_bsize, mi_row + hbs,
+                       mi_col, mi_row_top, mi_col_top, dry_run, dst_buf1,
+                       dst_stride1, 1);
+
+          // Smooth
+          for (i = 0; i < MAX_MB_PLANE; i++) {
+            xd->plane[i].dst.buf = dst_buf[i];
+            xd->plane[i].dst.stride = dst_stride[i];
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_HORZ, i);
+          }
+        }
+      }
+      break;
+    case PARTITION_VERT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        // First half
+        predict_b_extend(cpi, td, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride, top_bsize,
+                         BLOCK_8X8, dry_run, 1, 0);
+        if (bsize < top_bsize)
+          extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride);
+
+        // Second half
+        predict_b_extend(cpi, td, tile, 1, mi_row, mi_col, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                         top_bsize, BLOCK_8X8, dry_run, 1, 1);
+        if (bsize < top_bsize)
+          extend_all(cpi, td, tile, 1, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf1, dst_stride1);
+
+        // Smooth
+        xd->plane[0].dst.buf = dst_buf[0];
+        xd->plane[0].dst.stride = dst_stride[0];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[0], dst_stride[0], dst_buf1[0], dst_stride1[0], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            0);
+      } else {
+        // bsize: not important, not useful
+        predict_b_extend(cpi, td, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride, top_bsize,
+                         subsize, dry_run, 0, 0);
+        if (bsize < top_bsize)
+          extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride);
+        else
+          extend_dir(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride, 3);
+
+        if (mi_col + hbs < cm->mi_cols) {
+          predict_b_extend(cpi, td, tile, 0, mi_row, mi_col + hbs, mi_row,
+                           mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                           dst_stride1, top_bsize, subsize, dry_run, 0, 0);
+          if (bsize < top_bsize)
+            extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row,
+                       mi_col + hbs, mi_row_top, mi_col_top, dry_run, dst_buf1,
+                       dst_stride1);
+          else
+            extend_dir(cpi, td, tile, 0, subsize, top_bsize, mi_row,
+                       mi_col + hbs, mi_row_top, mi_col_top, dry_run, dst_buf1,
+                       dst_stride1, 2);
+
+          for (i = 0; i < MAX_MB_PLANE; i++) {
+            xd->plane[i].dst.buf = dst_buf[i];
+            xd->plane[i].dst.stride = dst_stride[i];
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_VERT, i);
+          }
+        }
+      }
+      break;
+    case PARTITION_SPLIT:
+      if (bsize == BLOCK_8X8 && !unify_bsize) {
+        predict_b_extend(cpi, td, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf, dst_stride, top_bsize,
+                         BLOCK_8X8, dry_run, 1, 0);
+        predict_b_extend(cpi, td, tile, 1, mi_row, mi_col, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                         top_bsize, BLOCK_8X8, dry_run, 1, 1);
+        predict_b_extend(cpi, td, tile, 2, mi_row, mi_col, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf2, dst_stride2,
+                         top_bsize, BLOCK_8X8, dry_run, 1, 1);
+        predict_b_extend(cpi, td, tile, 3, mi_row, mi_col, mi_row, mi_col,
+                         mi_row_top, mi_col_top, dst_buf3, dst_stride3,
+                         top_bsize, BLOCK_8X8, dry_run, 1, 1);
+
+        if (bsize < top_bsize) {
+          extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride);
+          extend_all(cpi, td, tile, 1, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf1, dst_stride1);
+          extend_all(cpi, td, tile, 2, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf2, dst_stride2);
+          extend_all(cpi, td, tile, 3, subsize, top_bsize, mi_row, mi_col,
+                     mi_row_top, mi_col_top, dry_run, dst_buf3, dst_stride3);
+        }
+      } else {
+        predict_sb_complex(cpi, td, tile, mi_row, mi_col, mi_row_top,
+                           mi_col_top, dry_run, subsize, top_bsize, dst_buf,
+                           dst_stride, pc_tree->split[0]);
+        if (mi_row < cm->mi_rows && mi_col + hbs < cm->mi_cols)
+          predict_sb_complex(cpi, td, tile, mi_row, mi_col + hbs, mi_row_top,
+                             mi_col_top, dry_run, subsize, top_bsize, dst_buf1,
+                             dst_stride1, pc_tree->split[1]);
+        if (mi_row + hbs < cm->mi_rows && mi_col < cm->mi_cols)
+          predict_sb_complex(cpi, td, tile, mi_row + hbs, mi_col, mi_row_top,
+                             mi_col_top, dry_run, subsize, top_bsize, dst_buf2,
+                             dst_stride2, pc_tree->split[2]);
+        if (mi_row + hbs < cm->mi_rows && mi_col + hbs < cm->mi_cols)
+          predict_sb_complex(cpi, td, tile, mi_row + hbs, mi_col + hbs,
+                             mi_row_top, mi_col_top, dry_run, subsize,
+                             top_bsize, dst_buf3, dst_stride3,
+                             pc_tree->split[3]);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+#if !CONFIG_CB4X4
+        if (bsize == BLOCK_8X8 && i != 0)
+          continue;  // Skip <4x4 chroma smoothing
+#endif
+        if (mi_row < cm->mi_rows && mi_col + hbs < cm->mi_cols) {
+          av1_build_masked_inter_predictor_complex(
+              xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i],
+              mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+              PARTITION_VERT, i);
+          if (mi_row + hbs < cm->mi_rows) {
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf2[i], dst_stride2[i], dst_buf3[i], dst_stride3[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_VERT, i);
+            av1_build_masked_inter_predictor_complex(
+                xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i],
+                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+                PARTITION_HORZ, i);
+          }
+        } else if (mi_row + hbs < cm->mi_rows && mi_col < cm->mi_cols) {
+          if (bsize == BLOCK_8X8 && i != 0)
+            continue;  // Skip <4x4 chroma smoothing
+
+          av1_build_masked_inter_predictor_complex(
+              xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i],
+              mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+              PARTITION_HORZ, i);
+        }
+      }
+      break;
+#if CONFIG_EXT_PARTITION_TYPES
+    case PARTITION_HORZ_A:
+      predict_b_extend(cpi, td, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride, top_bsize,
+                       bsize2, dry_run, 0, 0);
+      extend_all(cpi, td, tile, 0, bsize2, top_bsize, mi_row, mi_col,
+                 mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride);
+
+      predict_b_extend(cpi, td, tile, 0, mi_row, mi_col + hbs, mi_row,
+                       mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                       dst_stride1, top_bsize, bsize2, dry_run, 0, 0);
+      extend_all(cpi, td, tile, 0, bsize2, top_bsize, mi_row, mi_col + hbs,
+                 mi_row_top, mi_col_top, dry_run, dst_buf1, dst_stride1);
+
+      predict_b_extend(cpi, td, tile, 0, mi_row + hbs, mi_col, mi_row + hbs,
+                       mi_col, mi_row_top, mi_col_top, dst_buf2, dst_stride2,
+                       top_bsize, subsize, dry_run, 0, 0);
+      if (bsize < top_bsize)
+        extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row + hbs, mi_col,
+                   mi_row_top, mi_col_top, dry_run, dst_buf2, dst_stride2);
+      else
+        extend_dir(cpi, td, tile, 0, subsize, top_bsize, mi_row + hbs, mi_col,
+                   mi_row_top, mi_col_top, dry_run, dst_buf2, dst_stride2, 1);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            i);
+      }
+
+      break;
+    case PARTITION_VERT_A:
+
+      predict_b_extend(cpi, td, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride, top_bsize,
+                       bsize2, dry_run, 0, 0);
+      extend_all(cpi, td, tile, 0, bsize2, top_bsize, mi_row, mi_col,
+                 mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride);
+
+      predict_b_extend(cpi, td, tile, 0, mi_row + hbs, mi_col, mi_row + hbs,
+                       mi_col, mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                       top_bsize, bsize2, dry_run, 0, 0);
+      extend_all(cpi, td, tile, 0, bsize2, top_bsize, mi_row + hbs, mi_col,
+                 mi_row_top, mi_col_top, dry_run, dst_buf1, dst_stride1);
+
+      predict_b_extend(cpi, td, tile, 0, mi_row, mi_col + hbs, mi_row,
+                       mi_col + hbs, mi_row_top, mi_col_top, dst_buf2,
+                       dst_stride2, top_bsize, subsize, dry_run, 0, 0);
+      if (bsize < top_bsize)
+        extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col + hbs,
+                   mi_row_top, mi_col_top, dry_run, dst_buf2, dst_stride2);
+      else
+        extend_dir(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col + hbs,
+                   mi_row_top, mi_col_top, dry_run, dst_buf2, dst_stride2, 2);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf2[i], dst_stride2[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            i);
+      }
+      break;
+    case PARTITION_HORZ_B:
+
+      predict_b_extend(cpi, td, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride, top_bsize,
+                       subsize, dry_run, 0, 0);
+      if (bsize < top_bsize)
+        extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                   mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride);
+      else
+        extend_dir(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                   mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride, 0);
+
+      predict_b_extend(cpi, td, tile, 0, mi_row + hbs, mi_col, mi_row + hbs,
+                       mi_col, mi_row_top, mi_col_top, dst_buf1, dst_stride1,
+                       top_bsize, bsize2, dry_run, 0, 0);
+      extend_all(cpi, td, tile, 0, bsize2, top_bsize, mi_row + hbs, mi_col,
+                 mi_row_top, mi_col_top, dry_run, dst_buf1, dst_stride1);
+
+      predict_b_extend(cpi, td, tile, 0, mi_row + hbs, mi_col + hbs,
+                       mi_row + hbs, mi_col + hbs, mi_row_top, mi_col_top,
+                       dst_buf2, dst_stride2, top_bsize, bsize2, dry_run, 0, 0);
+      extend_all(cpi, td, tile, 0, bsize2, top_bsize, mi_row + hbs,
+                 mi_col + hbs, mi_row_top, mi_col_top, dry_run, dst_buf2,
+                 dst_stride2);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf1[i];
+        xd->plane[i].dst.stride = dst_stride1[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf1[i], dst_stride1[i], dst_buf2[i], dst_stride2[i],
+            mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+            PARTITION_VERT, i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
+            i);
+      }
+      break;
+    case PARTITION_VERT_B:
+
+      predict_b_extend(cpi, td, tile, 0, mi_row, mi_col, mi_row, mi_col,
+                       mi_row_top, mi_col_top, dst_buf, dst_stride, top_bsize,
+                       subsize, dry_run, 0, 0);
+      if (bsize < top_bsize)
+        extend_all(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                   mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride);
+      else
+        extend_dir(cpi, td, tile, 0, subsize, top_bsize, mi_row, mi_col,
+                   mi_row_top, mi_col_top, dry_run, dst_buf, dst_stride, 3);
+
+      predict_b_extend(cpi, td, tile, 0, mi_row, mi_col + hbs, mi_row,
+                       mi_col + hbs, mi_row_top, mi_col_top, dst_buf1,
+                       dst_stride1, top_bsize, bsize2, dry_run, 0, 0);
+      extend_all(cpi, td, tile, 0, bsize2, top_bsize, mi_row, mi_col + hbs,
+                 mi_row_top, mi_col_top, dry_run, dst_buf1, dst_stride1);
+
+      predict_b_extend(cpi, td, tile, 0, mi_row + hbs, mi_col + hbs,
+                       mi_row + hbs, mi_col + hbs, mi_row_top, mi_col_top,
+                       dst_buf2, dst_stride2, top_bsize, bsize2, dry_run, 0, 0);
+      extend_all(cpi, td, tile, 0, bsize2, top_bsize, mi_row + hbs,
+                 mi_col + hbs, mi_row_top, mi_col_top, dry_run, dst_buf2,
+                 dst_stride2);
+
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf1[i];
+        xd->plane[i].dst.stride = dst_stride1[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf1[i], dst_stride1[i], dst_buf2[i], dst_stride2[i],
+            mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
+            PARTITION_HORZ, i);
+      }
+      for (i = 0; i < MAX_MB_PLANE; i++) {
+        xd->plane[i].dst.buf = dst_buf[i];
+        xd->plane[i].dst.stride = dst_stride[i];
+        av1_build_masked_inter_predictor_complex(
+            xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i], mi_row,
+            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_VERT,
+            i);
+      }
+      break;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+    default: assert(0);
+  }
+
+#if CONFIG_EXT_PARTITION_TYPES
+  if (bsize < top_bsize)
+    update_ext_partition_context(xd, mi_row, mi_col, subsize, bsize, partition);
+#else
+  if (bsize < top_bsize && (partition != PARTITION_SPLIT || bsize == BLOCK_8X8))
+    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+}
+
+static void rd_supertx_sb(const AV1_COMP *const cpi, ThreadData *td,
+                          const TileInfo *const tile, int mi_row, int mi_col,
+                          BLOCK_SIZE bsize, int *tmp_rate, int64_t *tmp_dist,
+                          TX_TYPE *best_tx, PC_TREE *pc_tree) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int plane, pnskip, skippable, skippable_uv, rate_uv, this_rate,
+      base_rate = *tmp_rate;
+  int64_t sse, pnsse, sse_uv, this_dist, dist_uv;
+  uint8_t *dst_buf[3];
+  int dst_stride[3];
+  TX_SIZE tx_size;
+  MB_MODE_INFO *mbmi;
+  TX_TYPE tx_type, best_tx_nostx;
+#if CONFIG_EXT_TX
+  int ext_tx_set;
+#endif  // CONFIG_EXT_TX
+  int tmp_rate_tx = 0, skip_tx = 0;
+  int64_t tmp_dist_tx = 0, rd_tx, bestrd_tx = INT64_MAX;
+
+  set_skip_context(xd, mi_row, mi_col);
+  set_mode_info_offsets(cpi, x, xd, mi_row, mi_col);
+  update_state_sb_supertx(cpi, td, tile, mi_row, mi_col, bsize, 1, pc_tree);
+  av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                       mi_col);
+  for (plane = 0; plane < MAX_MB_PLANE; plane++) {
+    dst_buf[plane] = xd->plane[plane].dst.buf;
+    dst_stride[plane] = xd->plane[plane].dst.stride;
+  }
+  predict_sb_complex(cpi, td, tile, mi_row, mi_col, mi_row, mi_col, 1, bsize,
+                     bsize, dst_buf, dst_stride, pc_tree);
+
+  set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize);
+  set_segment_id_supertx(cpi, x, mi_row, mi_col, bsize);
+
+  mbmi = &xd->mi[0]->mbmi;
+  best_tx_nostx = mbmi->tx_type;
+
+  *best_tx = DCT_DCT;
+
+  // chroma
+  skippable_uv = 1;
+  rate_uv = 0;
+  dist_uv = 0;
+  sse_uv = 0;
+  for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_VAR_TX
+    ENTROPY_CONTEXT ctxa[2 * MAX_MIB_SIZE];
+    ENTROPY_CONTEXT ctxl[2 * MAX_MIB_SIZE];
+    const struct macroblockd_plane *const pd = &xd->plane[plane];
+    RD_STATS this_rd_stats;
+    av1_init_rd_stats(&this_rd_stats);
+
+    tx_size = max_txsize_lookup[bsize];
+    tx_size =
+        uv_txsize_lookup[bsize][tx_size][cm->subsampling_x][cm->subsampling_y];
+    av1_get_entropy_contexts(bsize, tx_size, pd, ctxa, ctxl);
+
+    av1_subtract_plane(x, bsize, plane);
+    av1_tx_block_rd_b(cpi, x, tx_size, 0, 0, plane, 0,
+                      get_plane_block_size(bsize, pd), &ctxa[0], &ctxl[0],
+                      &this_rd_stats);
+
+    this_rate = this_rd_stats.rate;
+    this_dist = this_rd_stats.dist;
+    pnsse = this_rd_stats.sse;
+    pnskip = this_rd_stats.skip;
+#else
+    tx_size = max_txsize_lookup[bsize];
+    tx_size =
+        uv_txsize_lookup[bsize][tx_size][cm->subsampling_x][cm->subsampling_y];
+    av1_subtract_plane(x, bsize, plane);
+    av1_txfm_rd_in_plane_supertx(x, cpi, &this_rate, &this_dist, &pnskip,
+                                 &pnsse, INT64_MAX, plane, bsize, tx_size, 0);
+#endif  // CONFIG_VAR_TX
+
+    rate_uv += this_rate;
+    dist_uv += this_dist;
+    sse_uv += pnsse;
+    skippable_uv &= pnskip;
+  }
+
+  // luma
+  tx_size = max_txsize_lookup[bsize];
+  av1_subtract_plane(x, bsize, 0);
+#if CONFIG_EXT_TX
+  ext_tx_set = get_ext_tx_set(tx_size, bsize, 1, cm->reduced_tx_set_used);
+#endif  // CONFIG_EXT_TX
+  for (tx_type = DCT_DCT; tx_type < TX_TYPES; ++tx_type) {
+#if CONFIG_VAR_TX
+    ENTROPY_CONTEXT ctxa[2 * MAX_MIB_SIZE];
+    ENTROPY_CONTEXT ctxl[2 * MAX_MIB_SIZE];
+    const struct macroblockd_plane *const pd = &xd->plane[0];
+    RD_STATS this_rd_stats;
+#endif  // CONFIG_VAR_TX
+
+#if CONFIG_EXT_TX
+    if (!ext_tx_used_inter[ext_tx_set][tx_type]) continue;
+#else
+    if (tx_size >= TX_32X32 && tx_type != DCT_DCT) continue;
+#endif  // CONFIG_EXT_TX
+    mbmi->tx_type = tx_type;
+
+#if CONFIG_VAR_TX
+    av1_init_rd_stats(&this_rd_stats);
+    av1_get_entropy_contexts(bsize, tx_size, pd, ctxa, ctxl);
+    av1_tx_block_rd_b(cpi, x, tx_size, 0, 0, 0, 0, bsize, &ctxa[0], &ctxl[0],
+                      &this_rd_stats);
+
+    this_rate = this_rd_stats.rate;
+    this_dist = this_rd_stats.dist;
+    pnsse = this_rd_stats.sse;
+    pnskip = this_rd_stats.skip;
+#else
+    av1_txfm_rd_in_plane_supertx(x, cpi, &this_rate, &this_dist, &pnskip,
+                                 &pnsse, INT64_MAX, 0, bsize, tx_size, 0);
+#endif  // CONFIG_VAR_TX
+
+#if CONFIG_EXT_TX
+    if (get_ext_tx_types(tx_size, bsize, 1, cm->reduced_tx_set_used) > 1 &&
+        !xd->lossless[xd->mi[0]->mbmi.segment_id] && this_rate != INT_MAX) {
+      if (ext_tx_set > 0)
+        this_rate +=
+            cpi->inter_tx_type_costs[ext_tx_set][mbmi->tx_size][mbmi->tx_type];
+    }
+#else
+    if (tx_size < TX_32X32 && !xd->lossless[xd->mi[0]->mbmi.segment_id] &&
+        this_rate != INT_MAX) {
+      this_rate += cpi->inter_tx_type_costs[tx_size][mbmi->tx_type];
+    }
+#endif  // CONFIG_EXT_TX
+    *tmp_rate = rate_uv + this_rate;
+    *tmp_dist = dist_uv + this_dist;
+    sse = sse_uv + pnsse;
+    skippable = skippable_uv && pnskip;
+    if (skippable) {
+      *tmp_rate = av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+      x->skip = 1;
+    } else {
+      if (RDCOST(x->rdmult, x->rddiv, *tmp_rate, *tmp_dist) <
+          RDCOST(x->rdmult, x->rddiv, 0, sse)) {
+        *tmp_rate += av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+        x->skip = 0;
+      } else {
+        *tmp_dist = sse;
+        *tmp_rate = av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+        x->skip = 1;
+      }
+    }
+    *tmp_rate += base_rate;
+    rd_tx = RDCOST(x->rdmult, x->rddiv, *tmp_rate, *tmp_dist);
+    if (rd_tx < bestrd_tx * 0.99 || tx_type == DCT_DCT) {
+      *best_tx = tx_type;
+      bestrd_tx = rd_tx;
+      tmp_rate_tx = *tmp_rate;
+      tmp_dist_tx = *tmp_dist;
+      skip_tx = x->skip;
+    }
+  }
+  *tmp_rate = tmp_rate_tx;
+  *tmp_dist = tmp_dist_tx;
+  x->skip = skip_tx;
+#if CONFIG_VAR_TX
+  for (plane = 0; plane < 1; ++plane)
+    memset(x->blk_skip[plane], x->skip,
+           sizeof(uint8_t) * pc_tree->none.num_4x4_blk);
+#endif  // CONFIG_VAR_TX
+  xd->mi[0]->mbmi.tx_type = best_tx_nostx;
+}
+#endif  // CONFIG_SUPERTX
diff --git a/third_party/aom/av1/encoder/encodeframe.h b/third_party/aom/av1/encoder/encodeframe.h
new file mode 100644
index 0000000000..08d6d20dee
--- /dev/null
+++ b/third_party/aom/av1/encoder/encodeframe.h
@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_ENCODEFRAME_H_
+#define AV1_ENCODER_ENCODEFRAME_H_
+
+#include "aom/aom_integer.h"
+#include "av1/common/blockd.h"
+#include "av1/common/enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct macroblock;
+struct yv12_buffer_config;
+struct AV1_COMP;
+struct ThreadData;
+
+// Constants used in SOURCE_VAR_BASED_PARTITION
+#define VAR_HIST_MAX_BG_VAR 1000
+#define VAR_HIST_FACTOR 10
+#define VAR_HIST_BINS (VAR_HIST_MAX_BG_VAR / VAR_HIST_FACTOR + 1)
+#define VAR_HIST_LARGE_CUT_OFF 75
+#define VAR_HIST_SMALL_CUT_OFF 45
+
+void av1_setup_src_planes(struct macroblock *x,
+                          const struct yv12_buffer_config *src, int mi_row,
+                          int mi_col);
+
+void av1_encode_frame(struct AV1_COMP *cpi);
+
+void av1_init_tile_data(struct AV1_COMP *cpi);
+void av1_encode_tile(struct AV1_COMP *cpi, struct ThreadData *td, int tile_row,
+                     int tile_col);
+
+void av1_set_variance_partition_thresholds(struct AV1_COMP *cpi, int q);
+
+void av1_update_tx_type_count(const struct AV1Common *cm, MACROBLOCKD *xd,
+#if CONFIG_TXK_SEL
+                              int block, int plane,
+#endif
+                              BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              FRAME_COUNTS *counts);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_ENCODEFRAME_H_
diff --git a/third_party/aom/av1/encoder/encodemb.c b/third_party/aom/av1/encoder/encodemb.c
new file mode 100644
index 0000000000..c450244b1c
--- /dev/null
+++ b/third_party/aom/av1/encoder/encodemb.c
@@ -0,0 +1,1671 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom_dsp/bitwriter.h"
+#include "aom_dsp/quantize.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#include "av1/common/idct.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+#include "av1/common/scan.h"
+
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/encodemb.h"
+#if CONFIG_LV_MAP
+#include "av1/encoder/encodetxb.h"
+#endif
+#include "av1/encoder/hybrid_fwd_txfm.h"
+#include "av1/encoder/rd.h"
+#include "av1/encoder/tokenize.h"
+
+#if CONFIG_PVQ
+#include "av1/encoder/encint.h"
+#include "av1/common/partition.h"
+#include "av1/encoder/pvq_encoder.h"
+#endif
+
+#if CONFIG_CFL
+#include "av1/common/cfl.h"
+#endif
+
+// Check if one needs to use c version subtraction.
+static int check_subtract_block_size(int w, int h) { return w < 4 || h < 4; }
+
+static void subtract_block(const MACROBLOCKD *xd, int rows, int cols,
+                           int16_t *diff, ptrdiff_t diff_stride,
+                           const uint8_t *src8, ptrdiff_t src_stride,
+                           const uint8_t *pred8, ptrdiff_t pred_stride) {
+#if !CONFIG_HIGHBITDEPTH
+  (void)xd;
+#endif
+
+  if (check_subtract_block_size(rows, cols)) {
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      aom_highbd_subtract_block_c(rows, cols, diff, diff_stride, src8,
+                                  src_stride, pred8, pred_stride, xd->bd);
+      return;
+    }
+#endif  // CONFIG_HIGHBITDEPTH
+    aom_subtract_block_c(rows, cols, diff, diff_stride, src8, src_stride, pred8,
+                         pred_stride);
+
+    return;
+  }
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    aom_highbd_subtract_block(rows, cols, diff, diff_stride, src8, src_stride,
+                              pred8, pred_stride, xd->bd);
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  aom_subtract_block(rows, cols, diff, diff_stride, src8, src_stride, pred8,
+                     pred_stride);
+}
+
+void av1_subtract_txb(MACROBLOCK *x, int plane, BLOCK_SIZE plane_bsize,
+                      int blk_col, int blk_row, TX_SIZE tx_size) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *const p = &x->plane[plane];
+  const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
+  const int diff_stride = block_size_wide[plane_bsize];
+  const int src_stride = p->src.stride;
+  const int dst_stride = pd->dst.stride;
+  const int tx1d_width = tx_size_wide[tx_size];
+  const int tx1d_height = tx_size_high[tx_size];
+  uint8_t *dst =
+      &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
+  uint8_t *src =
+      &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
+  int16_t *src_diff =
+      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
+  subtract_block(xd, tx1d_height, tx1d_width, src_diff, diff_stride, src,
+                 src_stride, dst, dst_stride);
+}
+
+void av1_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
+  struct macroblock_plane *const p = &x->plane[plane];
+  const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+  const int bw = block_size_wide[plane_bsize];
+  const int bh = block_size_high[plane_bsize];
+  const MACROBLOCKD *xd = &x->e_mbd;
+
+  subtract_block(xd, bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
+                 pd->dst.buf, pd->dst.stride);
+}
+
+// These numbers are empirically obtained.
+static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
+#if CONFIG_EC_ADAPT
+  { 10, 7 }, { 8, 5 },
+#else
+  { 10, 6 }, { 8, 5 },
+#endif
+};
+
+#define UPDATE_RD_COST()                             \
+  {                                                  \
+    rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0); \
+    rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1); \
+  }
+
+static INLINE int64_t
+get_token_bit_costs(unsigned int token_costs[2][COEFF_CONTEXTS][ENTROPY_TOKENS],
+                    int skip_eob, int ctx, int token) {
+#if CONFIG_NEW_TOKENSET
+  (void)skip_eob;
+  return token_costs[token == ZERO_TOKEN || token == EOB_TOKEN][ctx][token];
+#else
+  return token_costs[skip_eob][ctx][token];
+#endif
+}
+
+#define USE_GREEDY_OPTIMIZE_B 0
+
+#if USE_GREEDY_OPTIMIZE_B
+
+typedef struct av1_token_state {
+  int16_t token;
+  tran_low_t qc;
+  tran_low_t dqc;
+} av1_token_state;
+
+int av1_optimize_b(const AV1_COMMON *cm, MACROBLOCK *mb, int plane, int block,
+                   TX_SIZE tx_size, int ctx) {
+#if !CONFIG_PVQ
+  MACROBLOCKD *const xd = &mb->e_mbd;
+  struct macroblock_plane *const p = &mb->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int ref = is_inter_block(&xd->mi[0]->mbmi);
+  av1_token_state tokens[MAX_TX_SQUARE + 1][2];
+  uint8_t token_cache[MAX_TX_SQUARE];
+  const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  const int eob = p->eobs[block];
+  const PLANE_TYPE plane_type = pd->plane_type;
+  const int16_t *const dequant_ptr = pd->dequant;
+  const uint8_t *const band_translate = get_band_translate(tx_size);
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(&xd->mi[0]->mbmi));
+  const int16_t *const scan = scan_order->scan;
+  const int16_t *const nb = scan_order->neighbors;
+  int dqv;
+  const int shift = av1_get_tx_scale(tx_size);
+#if CONFIG_AOM_QM
+  int seg_id = xd->mi[0]->mbmi.segment_id;
+  const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!ref][tx_size];
+#endif
+#if CONFIG_NEW_QUANT
+  int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
+  const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq[dq];
+#elif !CONFIG_AOM_QM
+  const int dq_step[2] = { dequant_ptr[0] >> shift, dequant_ptr[1] >> shift };
+#endif  // CONFIG_NEW_QUANT
+  int sz = 0;
+  const int64_t rddiv = mb->rddiv;
+  int64_t rd_cost0, rd_cost1;
+  int16_t t0, t1;
+  int i, final_eob;
+#if CONFIG_HIGHBITDEPTH
+  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
+#else
+  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, 8);
+#endif
+  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
+      mb->token_costs[txsize_sqr_map[tx_size]][plane_type][ref];
+  const int default_eob = tx_size_2d[tx_size];
+
+  assert((mb->qindex == 0) ^ (xd->lossless[xd->mi[0]->mbmi.segment_id] == 0));
+
+  assert((!plane_type && !plane) || (plane_type && plane));
+  assert(eob <= default_eob);
+
+  int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
+/* CpuSpeedTest uses "--min-q=0 --max-q=0" and expects 100dB psnr
+* This creates conflict with search for a better EOB position
+* The line below is to make sure EOB search is disabled at this corner case.
+*/
+#if !CONFIG_NEW_QUANT && !CONFIG_AOM_QM
+  if (dq_step[1] <= 4) {
+    rdmult = 1;
+  }
+#endif
+
+  int64_t rate0, rate1;
+  for (i = 0; i < eob; i++) {
+    const int rc = scan[i];
+    int x = qcoeff[rc];
+    t0 = av1_get_token(x);
+
+    tokens[i][0].qc = x;
+    tokens[i][0].token = t0;
+    tokens[i][0].dqc = dqcoeff[rc];
+
+    token_cache[rc] = av1_pt_energy_class[t0];
+  }
+  tokens[eob][0].token = EOB_TOKEN;
+  tokens[eob][0].qc = 0;
+  tokens[eob][0].dqc = 0;
+  tokens[eob][1] = tokens[eob][0];
+
+  unsigned int(*token_costs_ptr)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
+      token_costs;
+
+  final_eob = 0;
+
+  int64_t eob_cost0, eob_cost1;
+
+  const int ctx0 = ctx;
+  /* Record the r-d cost */
+  int64_t accu_rate = 0;
+  int64_t accu_error = 0;
+
+  rate0 = get_token_bit_costs(*(token_costs_ptr + band_translate[0]), 0, ctx0,
+                              EOB_TOKEN);
+  int64_t best_block_rd_cost = RDCOST(rdmult, rddiv, rate0, accu_error);
+
+  // int64_t best_block_rd_cost_all0 = best_block_rd_cost;
+
+  int x_prev = 1;
+
+  for (i = 0; i < eob; i++) {
+    const int rc = scan[i];
+    int x = qcoeff[rc];
+    sz = -(x < 0);
+
+    int band_cur = band_translate[i];
+    int ctx_cur = (i == 0) ? ctx : get_coef_context(nb, token_cache, i);
+    int token_tree_sel_cur = (x_prev == 0);
+
+    if (x == 0) {
+      // no need to search when x == 0
+      rate0 =
+          get_token_bit_costs(*(token_costs_ptr + band_cur), token_tree_sel_cur,
+                              ctx_cur, tokens[i][0].token);
+      accu_rate += rate0;
+      x_prev = 0;
+      // accu_error does not change when x==0
+    } else {
+      /*  Computing distortion
+       */
+      // compute the distortion for the first candidate
+      // and the distortion for quantizing to 0.
+      int dx0 = (-coeff[rc]) * (1 << shift);
+#if CONFIG_HIGHBITDEPTH
+      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        dx0 >>= xd->bd - 8;
+      }
+#endif
+      int64_t d0 = (int64_t)dx0 * dx0;
+
+      int x_a = x - 2 * sz - 1;
+      int64_t d2, d2_a;
+
+      int dx;
+
+#if CONFIG_AOM_QM
+      int iwt = iqmatrix[rc];
+      dqv = dequant_ptr[rc != 0];
+      dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
+#else
+      dqv = dequant_ptr[rc != 0];
+#endif
+
+      dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
+#if CONFIG_HIGHBITDEPTH
+      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        dx >>= xd->bd - 8;
+      }
+#endif  // CONFIG_HIGHBITDEPTH
+      d2 = (int64_t)dx * dx;
+
+      /* compute the distortion for the second candidate
+       * x_a = x - 2 * sz + 1;
+       */
+      if (x_a != 0) {
+#if CONFIG_NEW_QUANT
+        dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
+             (coeff[rc] << shift);
+#if CONFIG_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          dx >>= xd->bd - 8;
+        }
+#endif  // CONFIG_HIGHBITDEPTH
+#else   // CONFIG_NEW_QUANT
+#if CONFIG_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
+        } else {
+          dx -= (dqv + sz) ^ sz;
+        }
+#else
+        dx -= (dqv + sz) ^ sz;
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_NEW_QUANT
+        d2_a = (int64_t)dx * dx;
+      } else {
+        d2_a = d0;
+      }
+      /*  Computing rates and r-d cost
+       */
+
+      int best_x, best_eob_x;
+      int64_t base_bits, next_bits0, next_bits1;
+      int64_t next_eob_bits0, next_eob_bits1;
+
+      // rate cost of x
+      base_bits = av1_get_token_cost(x, &t0, cat6_bits);
+      rate0 = base_bits + get_token_bit_costs(*(token_costs_ptr + band_cur),
+                                              token_tree_sel_cur, ctx_cur, t0);
+
+      base_bits = av1_get_token_cost(x_a, &t1, cat6_bits);
+      rate1 = base_bits + get_token_bit_costs(*(token_costs_ptr + band_cur),
+                                              token_tree_sel_cur, ctx_cur, t1);
+
+      next_bits0 = 0;
+      next_bits1 = 0;
+      next_eob_bits0 = 0;
+      next_eob_bits1 = 0;
+
+      if (i < default_eob - 1) {
+        int ctx_next, token_tree_sel_next;
+        int band_next = band_translate[i + 1];
+
+        token_cache[rc] = av1_pt_energy_class[t0];
+        ctx_next = get_coef_context(nb, token_cache, i + 1);
+        token_tree_sel_next = (x == 0);
+
+        next_bits0 = get_token_bit_costs(*(token_costs_ptr + band_next),
+                                         token_tree_sel_next, ctx_next,
+                                         tokens[i + 1][0].token);
+        next_eob_bits0 =
+            get_token_bit_costs(*(token_costs_ptr + band_next),
+                                token_tree_sel_next, ctx_next, EOB_TOKEN);
+
+        token_cache[rc] = av1_pt_energy_class[t1];
+        ctx_next = get_coef_context(nb, token_cache, i + 1);
+        token_tree_sel_next = (x_a == 0);
+
+        next_bits1 = get_token_bit_costs(*(token_costs_ptr + band_next),
+                                         token_tree_sel_next, ctx_next,
+                                         tokens[i + 1][0].token);
+
+        if (x_a != 0) {
+          next_eob_bits1 =
+              get_token_bit_costs(*(token_costs_ptr + band_next),
+                                  token_tree_sel_next, ctx_next, EOB_TOKEN);
+        }
+      }
+
+      rd_cost0 = RDCOST(rdmult, rddiv, (rate0 + next_bits0), d2);
+      rd_cost1 = RDCOST(rdmult, rddiv, (rate1 + next_bits1), d2_a);
+
+      best_x = (rd_cost1 < rd_cost0);
+
+      eob_cost0 = RDCOST(rdmult, rddiv, (accu_rate + rate0 + next_eob_bits0),
+                         (accu_error + d2 - d0));
+      eob_cost1 = eob_cost0;
+      if (x_a != 0) {
+        eob_cost1 = RDCOST(rdmult, rddiv, (accu_rate + rate1 + next_eob_bits1),
+                           (accu_error + d2_a - d0));
+        best_eob_x = (eob_cost1 < eob_cost0);
+      } else {
+        best_eob_x = 0;
+      }
+
+      int dqc, dqc_a = 0;
+
+      dqc = dqcoeff[rc];
+      if (best_x + best_eob_x) {
+        if (x_a != 0) {
+#if CONFIG_NEW_QUANT
+          dqc_a = av1_dequant_abscoeff_nuq(abs(x_a), dqv,
+                                           dequant_val[band_translate[i]]);
+          dqc_a = shift ? ROUND_POWER_OF_TWO(dqc_a, shift) : dqc_a;
+          if (sz) dqc_a = -dqc_a;
+#else
+// The 32x32 transform coefficient uses half quantization step size.
+// Account for the rounding difference in the dequantized coefficeint
+// value when the quantization index is dropped from an even number
+// to an odd number.
+
+#if CONFIG_AOM_QM
+          tran_low_t offset = dqv >> shift;
+#else
+          tran_low_t offset = dq_step[rc != 0];
+#endif
+          if (shift & x_a) offset += (dqv & 0x01);
+
+          if (sz == 0)
+            dqc_a = dqcoeff[rc] - offset;
+          else
+            dqc_a = dqcoeff[rc] + offset;
+#endif  // CONFIG_NEW_QUANT
+        } else {
+          dqc_a = 0;
+        }  // if (x_a != 0)
+      }
+
+      // record the better quantized value
+      if (best_x) {
+        qcoeff[rc] = x_a;
+        dqcoeff[rc] = dqc_a;
+
+        accu_rate += rate1;
+        accu_error += d2_a - d0;
+        assert(d2_a <= d0);
+
+        token_cache[rc] = av1_pt_energy_class[t1];
+      } else {
+        accu_rate += rate0;
+        accu_error += d2 - d0;
+        assert(d2 <= d0);
+
+        token_cache[rc] = av1_pt_energy_class[t0];
+      }
+
+      x_prev = qcoeff[rc];
+
+      // determine whether to move the eob position to i+1
+      int64_t best_eob_cost_i = eob_cost0;
+
+      tokens[i][1].token = t0;
+      tokens[i][1].qc = x;
+      tokens[i][1].dqc = dqc;
+
+      if ((x_a != 0) && (best_eob_x)) {
+        best_eob_cost_i = eob_cost1;
+
+        tokens[i][1].token = t1;
+        tokens[i][1].qc = x_a;
+        tokens[i][1].dqc = dqc_a;
+      }
+
+      if (best_eob_cost_i < best_block_rd_cost) {
+        best_block_rd_cost = best_eob_cost_i;
+        final_eob = i + 1;
+      }
+    }  // if (x==0)
+  }    // for (i)
+
+  assert(final_eob <= eob);
+  if (final_eob > 0) {
+    assert(tokens[final_eob - 1][1].qc != 0);
+    i = final_eob - 1;
+    int rc = scan[i];
+    qcoeff[rc] = tokens[i][1].qc;
+    dqcoeff[rc] = tokens[i][1].dqc;
+  }
+
+  for (i = final_eob; i < eob; i++) {
+    int rc = scan[i];
+    qcoeff[rc] = 0;
+    dqcoeff[rc] = 0;
+  }
+
+  mb->plane[plane].eobs[block] = final_eob;
+  return final_eob;
+
+#else   // !CONFIG_PVQ
+  (void)cm;
+  (void)tx_size;
+  (void)ctx;
+  struct macroblock_plane *const p = &mb->plane[plane];
+  return p->eobs[block];
+#endif  // !CONFIG_PVQ
+}
+
+#else  // USE_GREEDY_OPTIMIZE_B
+
+typedef struct av1_token_state {
+  int64_t error;
+  int rate;
+  int16_t next;
+  int16_t token;
+  tran_low_t qc;
+  tran_low_t dqc;
+  uint8_t best_index;
+} av1_token_state;
+
+int av1_optimize_b(const AV1_COMMON *cm, MACROBLOCK *mb, int plane, int block,
+                   TX_SIZE tx_size, int ctx) {
+#if !CONFIG_PVQ
+  MACROBLOCKD *const xd = &mb->e_mbd;
+  struct macroblock_plane *const p = &mb->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int ref = is_inter_block(&xd->mi[0]->mbmi);
+  av1_token_state tokens[MAX_TX_SQUARE + 1][2];
+  uint8_t token_cache[MAX_TX_SQUARE];
+  const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  const int eob = p->eobs[block];
+  const PLANE_TYPE plane_type = pd->plane_type;
+  const int default_eob = tx_size_2d[tx_size];
+  const int16_t *const dequant_ptr = pd->dequant;
+  const uint8_t *const band_translate = get_band_translate(tx_size);
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(&xd->mi[0]->mbmi));
+  const int16_t *const scan = scan_order->scan;
+  const int16_t *const nb = scan_order->neighbors;
+  int dqv;
+  const int shift = av1_get_tx_scale(tx_size);
+#if CONFIG_AOM_QM
+  int seg_id = xd->mi[0]->mbmi.segment_id;
+  const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!ref][tx_size];
+#endif
+#if CONFIG_NEW_QUANT
+  int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
+  const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq[dq];
+#elif !CONFIG_AOM_QM
+  const int dq_step[2] = { dequant_ptr[0] >> shift, dequant_ptr[1] >> shift };
+#endif  // CONFIG_NEW_QUANT
+  int next = eob, sz = 0;
+  const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
+  const int64_t rddiv = mb->rddiv;
+  int64_t rd_cost0, rd_cost1;
+  int rate0, rate1;
+  int64_t error0, error1;
+  int16_t t0, t1;
+  int best, band = (eob < default_eob) ? band_translate[eob]
+                                       : band_translate[eob - 1];
+  int pt, i, final_eob;
+#if CONFIG_HIGHBITDEPTH
+  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
+#else
+  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, 8);
+#endif
+  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
+      mb->token_costs[txsize_sqr_map[tx_size]][plane_type][ref];
+  const uint16_t *band_counts = &band_count_table[tx_size][band];
+  uint16_t band_left = eob - band_cum_count_table[tx_size][band] + 1;
+  int shortcut = 0;
+  int next_shortcut = 0;
+
+#if CONFIG_EXT_DELTA_Q
+  const int qindex = cm->seg.enabled
+                         ? av1_get_qindex(&cm->seg, xd->mi[0]->mbmi.segment_id,
+                                          cm->base_qindex)
+                         : cm->base_qindex;
+  if (qindex == 0) {
+    assert((qindex == 0) ^ (xd->lossless[xd->mi[0]->mbmi.segment_id] == 0));
+  }
+#else
+  assert((mb->qindex == 0) ^ (xd->lossless[xd->mi[0]->mbmi.segment_id] == 0));
+#endif
+
+  token_costs += band;
+
+  assert((!plane_type && !plane) || (plane_type && plane));
+  assert(eob <= default_eob);
+
+  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
+  /* Initialize the sentinel node of the trellis. */
+  tokens[eob][0].rate = 0;
+  tokens[eob][0].error = 0;
+  tokens[eob][0].next = default_eob;
+  tokens[eob][0].token = EOB_TOKEN;
+  tokens[eob][0].qc = 0;
+  tokens[eob][1] = tokens[eob][0];
+
+  for (i = 0; i < eob; i++) {
+    const int rc = scan[i];
+    tokens[i][0].rate = av1_get_token_cost(qcoeff[rc], &t0, cat6_bits);
+    tokens[i][0].token = t0;
+    token_cache[rc] = av1_pt_energy_class[t0];
+  }
+
+  for (i = eob; i-- > 0;) {
+    int base_bits, dx;
+    int64_t d2;
+    const int rc = scan[i];
+    int x = qcoeff[rc];
+#if CONFIG_AOM_QM
+    int iwt = iqmatrix[rc];
+    dqv = dequant_ptr[rc != 0];
+    dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
+#else
+    dqv = dequant_ptr[rc != 0];
+#endif
+    next_shortcut = shortcut;
+
+    /* Only add a trellis state for non-zero coefficients. */
+    if (UNLIKELY(x)) {
+      error0 = tokens[next][0].error;
+      error1 = tokens[next][1].error;
+      /* Evaluate the first possibility for this state. */
+      rate0 = tokens[next][0].rate;
+      rate1 = tokens[next][1].rate;
+
+      if (next_shortcut) {
+        /* Consider both possible successor states. */
+        if (next < default_eob) {
+          pt = get_coef_context(nb, token_cache, i + 1);
+          rate0 +=
+              get_token_bit_costs(*token_costs, 0, pt, tokens[next][0].token);
+          rate1 +=
+              get_token_bit_costs(*token_costs, 0, pt, tokens[next][1].token);
+        }
+        UPDATE_RD_COST();
+        /* And pick the best. */
+        best = rd_cost1 < rd_cost0;
+      } else {
+        if (next < default_eob) {
+          pt = get_coef_context(nb, token_cache, i + 1);
+          rate0 +=
+              get_token_bit_costs(*token_costs, 0, pt, tokens[next][0].token);
+        }
+        best = 0;
+      }
+
+      dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
+#if CONFIG_HIGHBITDEPTH
+      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        dx >>= xd->bd - 8;
+      }
+#endif  // CONFIG_HIGHBITDEPTH
+      d2 = (int64_t)dx * dx;
+      tokens[i][0].rate += (best ? rate1 : rate0);
+      tokens[i][0].error = d2 + (best ? error1 : error0);
+      tokens[i][0].next = next;
+      tokens[i][0].qc = x;
+      tokens[i][0].dqc = dqcoeff[rc];
+      tokens[i][0].best_index = best;
+
+      /* Evaluate the second possibility for this state. */
+      rate0 = tokens[next][0].rate;
+      rate1 = tokens[next][1].rate;
+
+      // The threshold of 3 is empirically obtained.
+      if (UNLIKELY(abs(x) > 3)) {
+        shortcut = 0;
+      } else {
+#if CONFIG_NEW_QUANT
+        shortcut = ((av1_dequant_abscoeff_nuq(abs(x), dqv,
+                                              dequant_val[band_translate[i]]) >
+                     (abs(coeff[rc]) << shift)) &&
+                    (av1_dequant_abscoeff_nuq(abs(x) - 1, dqv,
+                                              dequant_val[band_translate[i]]) <
+                     (abs(coeff[rc]) << shift)));
+#else  // CONFIG_NEW_QUANT
+#if CONFIG_AOM_QM
+        if ((abs(x) * dequant_ptr[rc != 0] * iwt >
+             ((abs(coeff[rc]) << shift) << AOM_QM_BITS)) &&
+            (abs(x) * dequant_ptr[rc != 0] * iwt <
+             (((abs(coeff[rc]) << shift) + dequant_ptr[rc != 0])
+              << AOM_QM_BITS)))
+#else
+        if ((abs(x) * dequant_ptr[rc != 0] > (abs(coeff[rc]) << shift)) &&
+            (abs(x) * dequant_ptr[rc != 0] <
+             (abs(coeff[rc]) << shift) + dequant_ptr[rc != 0]))
+#endif  // CONFIG_AOM_QM
+          shortcut = 1;
+        else
+          shortcut = 0;
+#endif  // CONFIG_NEW_QUANT
+      }
+
+      if (shortcut) {
+        sz = -(x < 0);
+        x -= 2 * sz + 1;
+      } else {
+        tokens[i][1] = tokens[i][0];
+        next = i;
+
+        if (UNLIKELY(!(--band_left))) {
+          --band_counts;
+          band_left = *band_counts;
+          --token_costs;
+        }
+        continue;
+      }
+
+      /* Consider both possible successor states. */
+      if (!x) {
+        /* If we reduced this coefficient to zero, check to see if
+         *  we need to move the EOB back here.
+         */
+        t0 = tokens[next][0].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
+        t1 = tokens[next][1].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
+        base_bits = 0;
+      } else {
+        base_bits = av1_get_token_cost(x, &t0, cat6_bits);
+        t1 = t0;
+      }
+
+      if (next_shortcut) {
+        if (LIKELY(next < default_eob)) {
+          if (t0 != EOB_TOKEN) {
+            token_cache[rc] = av1_pt_energy_class[t0];
+            pt = get_coef_context(nb, token_cache, i + 1);
+            rate0 += get_token_bit_costs(*token_costs, !x, pt,
+                                         tokens[next][0].token);
+          }
+          if (t1 != EOB_TOKEN) {
+            token_cache[rc] = av1_pt_energy_class[t1];
+            pt = get_coef_context(nb, token_cache, i + 1);
+            rate1 += get_token_bit_costs(*token_costs, !x, pt,
+                                         tokens[next][1].token);
+          }
+        }
+
+        UPDATE_RD_COST();
+        /* And pick the best. */
+        best = rd_cost1 < rd_cost0;
+      } else {
+        // The two states in next stage are identical.
+        if (next < default_eob && t0 != EOB_TOKEN) {
+          token_cache[rc] = av1_pt_energy_class[t0];
+          pt = get_coef_context(nb, token_cache, i + 1);
+          rate0 +=
+              get_token_bit_costs(*token_costs, !x, pt, tokens[next][0].token);
+        }
+        best = 0;
+      }
+
+#if CONFIG_NEW_QUANT
+      dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
+           (coeff[rc] << shift);
+#if CONFIG_HIGHBITDEPTH
+      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        dx >>= xd->bd - 8;
+      }
+#endif  // CONFIG_HIGHBITDEPTH
+#else   // CONFIG_NEW_QUANT
+#if CONFIG_HIGHBITDEPTH
+      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
+      } else {
+        dx -= (dqv + sz) ^ sz;
+      }
+#else
+      dx -= (dqv + sz) ^ sz;
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_NEW_QUANT
+      d2 = (int64_t)dx * dx;
+
+      tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
+      tokens[i][1].error = d2 + (best ? error1 : error0);
+      tokens[i][1].next = next;
+      tokens[i][1].token = best ? t1 : t0;
+      tokens[i][1].qc = x;
+
+      if (x) {
+#if CONFIG_NEW_QUANT
+        tokens[i][1].dqc = av1_dequant_abscoeff_nuq(
+            abs(x), dqv, dequant_val[band_translate[i]]);
+        tokens[i][1].dqc = shift ? ROUND_POWER_OF_TWO(tokens[i][1].dqc, shift)
+                                 : tokens[i][1].dqc;
+        if (sz) tokens[i][1].dqc = -tokens[i][1].dqc;
+#else
+// The 32x32 transform coefficient uses half quantization step size.
+// Account for the rounding difference in the dequantized coefficeint
+// value when the quantization index is dropped from an even number
+// to an odd number.
+
+#if CONFIG_AOM_QM
+        tran_low_t offset = dqv >> shift;
+#else
+        tran_low_t offset = dq_step[rc != 0];
+#endif
+        if (shift & x) offset += (dqv & 0x01);
+
+        if (sz == 0)
+          tokens[i][1].dqc = dqcoeff[rc] - offset;
+        else
+          tokens[i][1].dqc = dqcoeff[rc] + offset;
+#endif  // CONFIG_NEW_QUANT
+      } else {
+        tokens[i][1].dqc = 0;
+      }
+
+      tokens[i][1].best_index = best;
+      /* Finally, make this the new head of the trellis. */
+      next = i;
+    } else {
+      /* There's no choice to make for a zero coefficient, so we don't
+       *  add a new trellis node, but we do need to update the costs.
+       */
+      t0 = tokens[next][0].token;
+      t1 = tokens[next][1].token;
+      pt = get_coef_context(nb, token_cache, i + 1);
+      /* Update the cost of each path if we're past the EOB token. */
+      if (t0 != EOB_TOKEN) {
+        tokens[next][0].rate += get_token_bit_costs(*token_costs, 1, pt, t0);
+        tokens[next][0].token = ZERO_TOKEN;
+      }
+      if (t1 != EOB_TOKEN) {
+        tokens[next][1].rate += get_token_bit_costs(*token_costs, 1, pt, t1);
+        tokens[next][1].token = ZERO_TOKEN;
+      }
+      tokens[i][0].best_index = tokens[i][1].best_index = 0;
+      shortcut = (tokens[next][0].rate != tokens[next][1].rate);
+      /* Don't update next, because we didn't add a new node. */
+    }
+
+    if (UNLIKELY(!(--band_left))) {
+      --band_counts;
+      band_left = *band_counts;
+      --token_costs;
+    }
+  }
+
+  /* Now pick the best path through the whole trellis. */
+  rate0 = tokens[next][0].rate;
+  rate1 = tokens[next][1].rate;
+  error0 = tokens[next][0].error;
+  error1 = tokens[next][1].error;
+  t0 = tokens[next][0].token;
+  t1 = tokens[next][1].token;
+  rate0 += get_token_bit_costs(*token_costs, 0, ctx, t0);
+  rate1 += get_token_bit_costs(*token_costs, 0, ctx, t1);
+  UPDATE_RD_COST();
+  best = rd_cost1 < rd_cost0;
+
+  final_eob = -1;
+
+  for (i = next; i < eob; i = next) {
+    const int x = tokens[i][best].qc;
+    const int rc = scan[i];
+    if (x) final_eob = i;
+    qcoeff[rc] = x;
+    dqcoeff[rc] = tokens[i][best].dqc;
+
+    next = tokens[i][best].next;
+    best = tokens[i][best].best_index;
+  }
+  final_eob++;
+
+  mb->plane[plane].eobs[block] = final_eob;
+  assert(final_eob <= default_eob);
+  return final_eob;
+#else   // !CONFIG_PVQ
+  (void)cm;
+  (void)tx_size;
+  (void)ctx;
+  struct macroblock_plane *const p = &mb->plane[plane];
+  return p->eobs[block];
+#endif  // !CONFIG_PVQ
+}
+
+#endif  // USE_GREEDY_OPTIMIZE_B
+
+#if !CONFIG_PVQ
+#if CONFIG_HIGHBITDEPTH
+typedef enum QUANT_FUNC {
+  QUANT_FUNC_LOWBD = 0,
+  QUANT_FUNC_HIGHBD = 1,
+  QUANT_FUNC_TYPES = 2
+} QUANT_FUNC;
+
+static AV1_QUANT_FACADE
+    quant_func_list[AV1_XFORM_QUANT_TYPES][QUANT_FUNC_TYPES] = {
+#if !CONFIG_NEW_QUANT
+      { av1_quantize_fp_facade, av1_highbd_quantize_fp_facade },
+      { av1_quantize_b_facade, av1_highbd_quantize_b_facade },
+      { av1_quantize_dc_facade, av1_highbd_quantize_dc_facade },
+#else   // !CONFIG_NEW_QUANT
+      { av1_quantize_fp_nuq_facade, av1_highbd_quantize_fp_nuq_facade },
+      { av1_quantize_b_nuq_facade, av1_highbd_quantize_b_nuq_facade },
+      { av1_quantize_dc_nuq_facade, av1_highbd_quantize_dc_nuq_facade },
+#endif  // !CONFIG_NEW_QUANT
+      { NULL, NULL }
+    };
+
+#else
+
+typedef enum QUANT_FUNC {
+  QUANT_FUNC_LOWBD = 0,
+  QUANT_FUNC_TYPES = 1
+} QUANT_FUNC;
+
+static AV1_QUANT_FACADE quant_func_list[AV1_XFORM_QUANT_TYPES]
+                                       [QUANT_FUNC_TYPES] = {
+#if !CONFIG_NEW_QUANT
+                                         { av1_quantize_fp_facade },
+                                         { av1_quantize_b_facade },
+                                         { av1_quantize_dc_facade },
+#else   // !CONFIG_NEW_QUANT
+                                         { av1_quantize_fp_nuq_facade },
+                                         { av1_quantize_b_nuq_facade },
+                                         { av1_quantize_dc_nuq_facade },
+#endif  // !CONFIG_NEW_QUANT
+                                         { NULL }
+                                       };
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_PVQ
+
+void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
+                     int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
+                     TX_SIZE tx_size, int ctx,
+                     AV1_XFORM_QUANT xform_quant_idx) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+#if !(CONFIG_PVQ || CONFIG_DAALA_DIST)
+  const struct macroblock_plane *const p = &x->plane[plane];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+#else
+  struct macroblock_plane *const p = &x->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+#endif
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const int is_inter = is_inter_block(mbmi);
+  const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, is_inter);
+  tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  uint16_t *const eob = &p->eobs[block];
+  const int diff_stride = block_size_wide[plane_bsize];
+#if CONFIG_AOM_QM
+  int seg_id = mbmi->segment_id;
+  const qm_val_t *qmatrix = pd->seg_qmatrix[seg_id][!is_inter][tx_size];
+  const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!is_inter][tx_size];
+#endif
+
+  FWD_TXFM_PARAM fwd_txfm_param;
+
+#if CONFIG_PVQ || CONFIG_DAALA_DIST
+  uint8_t *dst;
+  int16_t *pred;
+  const int dst_stride = pd->dst.stride;
+  int tx_blk_size;
+  int i, j;
+#endif
+
+#if !CONFIG_PVQ
+  const int tx2d_size = tx_size_2d[tx_size];
+  QUANT_PARAM qparam;
+  const int16_t *src_diff;
+
+  src_diff =
+      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
+  qparam.log_scale = av1_get_tx_scale(tx_size);
+#if CONFIG_NEW_QUANT
+  qparam.tx_size = tx_size;
+  qparam.dq = get_dq_profile_from_ctx(x->qindex, ctx, is_inter, plane_type);
+#endif  // CONFIG_NEW_QUANT
+#if CONFIG_AOM_QM
+  qparam.qmatrix = qmatrix;
+  qparam.iqmatrix = iqmatrix;
+#endif  // CONFIG_AOM_QM
+#else
+  tran_low_t *ref_coeff = BLOCK_OFFSET(pd->pvq_ref_coeff, block);
+  int skip = 1;
+  PVQ_INFO *pvq_info = NULL;
+  uint8_t *src;
+  int16_t *src_int16;
+  const int src_stride = p->src.stride;
+
+  (void)ctx;
+  (void)scan_order;
+  (void)qcoeff;
+
+  if (x->pvq_coded) {
+    assert(block < MAX_PVQ_BLOCKS_IN_SB);
+    pvq_info = &x->pvq[block][plane];
+  }
+  src = &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
+  src_int16 =
+      &p->src_int16[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
+
+  // transform block size in pixels
+  tx_blk_size = tx_size_wide[tx_size];
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    for (j = 0; j < tx_blk_size; j++)
+      for (i = 0; i < tx_blk_size; i++)
+        src_int16[diff_stride * j + i] =
+            CONVERT_TO_SHORTPTR(src)[src_stride * j + i];
+  } else {
+#endif  // CONFIG_HIGHBITDEPTH
+    for (j = 0; j < tx_blk_size; j++)
+      for (i = 0; i < tx_blk_size; i++)
+        src_int16[diff_stride * j + i] = src[src_stride * j + i];
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+#endif
+
+#if CONFIG_PVQ || CONFIG_DAALA_DIST
+  dst = &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
+  pred = &pd->pred[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
+
+  // transform block size in pixels
+  tx_blk_size = tx_size_wide[tx_size];
+
+// copy uint8 orig and predicted block to int16 buffer
+// in order to use existing VP10 transform functions
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    for (j = 0; j < tx_blk_size; j++)
+      for (i = 0; i < tx_blk_size; i++)
+        pred[diff_stride * j + i] =
+            CONVERT_TO_SHORTPTR(dst)[dst_stride * j + i];
+  } else {
+#endif  // CONFIG_HIGHBITDEPTH
+    for (j = 0; j < tx_blk_size; j++)
+      for (i = 0; i < tx_blk_size; i++)
+        pred[diff_stride * j + i] = dst[dst_stride * j + i];
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+#endif
+
+  (void)ctx;
+
+  fwd_txfm_param.tx_type = tx_type;
+  fwd_txfm_param.tx_size = tx_size;
+  fwd_txfm_param.lossless = xd->lossless[mbmi->segment_id];
+
+#if !CONFIG_PVQ
+#if CONFIG_HIGHBITDEPTH
+  fwd_txfm_param.bd = xd->bd;
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    av1_highbd_fwd_txfm(src_diff, coeff, diff_stride, &fwd_txfm_param);
+    if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
+      if (LIKELY(!x->skip_block)) {
+        quant_func_list[xform_quant_idx][QUANT_FUNC_HIGHBD](
+            coeff, tx2d_size, p, qcoeff, pd, dqcoeff, eob, scan_order, &qparam);
+      } else {
+        av1_quantize_skip(tx2d_size, qcoeff, dqcoeff, eob);
+      }
+    }
+#if CONFIG_LV_MAP
+    p->txb_entropy_ctx[block] =
+        (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob);
+#endif  // CONFIG_LV_MAP
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  av1_fwd_txfm(src_diff, coeff, diff_stride, &fwd_txfm_param);
+  if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
+    if (LIKELY(!x->skip_block)) {
+      quant_func_list[xform_quant_idx][QUANT_FUNC_LOWBD](
+          coeff, tx2d_size, p, qcoeff, pd, dqcoeff, eob, scan_order, &qparam);
+    } else {
+      av1_quantize_skip(tx2d_size, qcoeff, dqcoeff, eob);
+    }
+  }
+#if CONFIG_LV_MAP
+  p->txb_entropy_ctx[block] =
+      (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob);
+#endif  // CONFIG_LV_MAP
+#else   // #if !CONFIG_PVQ
+  (void)xform_quant_idx;
+#if CONFIG_HIGHBITDEPTH
+  fwd_txfm_param.bd = xd->bd;
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    av1_highbd_fwd_txfm(src_int16, coeff, diff_stride, &fwd_txfm_param);
+    av1_highbd_fwd_txfm(pred, ref_coeff, diff_stride, &fwd_txfm_param);
+  } else {
+#endif
+    av1_fwd_txfm(src_int16, coeff, diff_stride, &fwd_txfm_param);
+    av1_fwd_txfm(pred, ref_coeff, diff_stride, &fwd_txfm_param);
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif
+
+  // PVQ for inter mode block
+  if (!x->skip_block) {
+    PVQ_SKIP_TYPE ac_dc_coded =
+        av1_pvq_encode_helper(x,
+                              coeff,        // target original vector
+                              ref_coeff,    // reference vector
+                              dqcoeff,      // de-quantized vector
+                              eob,          // End of Block marker
+                              pd->dequant,  // aom's quantizers
+                              plane,        // image plane
+                              tx_size,      // block size in log_2 - 2
+                              tx_type,
+                              &x->rate,  // rate measured
+                              x->pvq_speed,
+                              pvq_info);  // PVQ info for a block
+    skip = ac_dc_coded == PVQ_SKIP;
+  }
+  x->pvq_skip[plane] = skip;
+
+  if (!skip) mbmi->skip = 0;
+#endif  // #if !CONFIG_PVQ
+}
+
+static void encode_block(int plane, int block, int blk_row, int blk_col,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
+  struct encode_b_args *const args = arg;
+  AV1_COMMON *cm = args->cm;
+  MACROBLOCK *const x = args->x;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int ctx;
+  struct macroblock_plane *const p = &x->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  uint8_t *dst;
+#if !CONFIG_PVQ
+  ENTROPY_CONTEXT *a, *l;
+#endif
+#if CONFIG_VAR_TX
+  int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+#endif
+  dst = &pd->dst
+             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
+
+#if !CONFIG_PVQ
+  a = &args->ta[blk_col];
+  l = &args->tl[blk_row];
+#if CONFIG_VAR_TX
+  ctx = get_entropy_context(tx_size, a, l);
+#else
+  ctx = combine_entropy_contexts(*a, *l);
+#endif
+#else
+  ctx = 0;
+#endif  // CONFIG_PVQ
+
+#if CONFIG_VAR_TX
+  // Assert not magic number (uninitialized).
+  assert(x->blk_skip[plane][blk_row * bw + blk_col] != 234);
+
+  if (x->blk_skip[plane][blk_row * bw + blk_col] == 0) {
+#else
+  {
+#endif
+    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                    ctx, AV1_XFORM_QUANT_FP);
+  }
+#if CONFIG_VAR_TX
+  else {
+    p->eobs[block] = 0;
+  }
+#endif
+
+#if !CONFIG_PVQ
+  if (p->eobs[block] && !xd->lossless[xd->mi[0]->mbmi.segment_id])
+    av1_optimize_b(cm, x, plane, block, tx_size, ctx);
+
+  av1_set_txb_context(x, plane, block, tx_size, a, l);
+
+  if (p->eobs[block]) *(args->skip) = 0;
+
+  if (p->eobs[block] == 0) return;
+#else
+  (void)ctx;
+  if (!x->pvq_skip[plane]) *(args->skip) = 0;
+
+  if (x->pvq_skip[plane]) return;
+#endif
+  TX_TYPE tx_type = get_tx_type(pd->plane_type, xd, block, tx_size);
+  av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, dst,
+                              pd->dst.stride, p->eobs[block]);
+}
+
+#if CONFIG_VAR_TX
+static void encode_block_inter(int plane, int block, int blk_row, int blk_col,
+                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                               void *arg) {
+  struct encode_b_args *const args = arg;
+  MACROBLOCK *const x = args->x;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int tx_row = blk_row >> (1 - pd->subsampling_y);
+  const int tx_col = blk_col >> (1 - pd->subsampling_x);
+  TX_SIZE plane_tx_size;
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  plane_tx_size =
+      plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
+            : mbmi->inter_tx_size[tx_row][tx_col];
+
+  if (tx_size == plane_tx_size) {
+    encode_block(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
+  } else {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    // This is the square transform block partition entry point.
+    int bsl = tx_size_wide_unit[sub_txs];
+    int i;
+    assert(bsl > 0);
+    assert(tx_size < TX_SIZES_ALL);
+
+    for (i = 0; i < 4; ++i) {
+      const int offsetr = blk_row + ((i >> 1) * bsl);
+      const int offsetc = blk_col + ((i & 0x01) * bsl);
+      int step = tx_size_wide_unit[sub_txs] * tx_size_high_unit[sub_txs];
+
+      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
+
+      encode_block_inter(plane, block, offsetr, offsetc, plane_bsize, sub_txs,
+                         arg);
+      block += step;
+    }
+  }
+}
+#endif
+
+typedef struct encode_block_pass1_args {
+  AV1_COMMON *cm;
+  MACROBLOCK *x;
+} encode_block_pass1_args;
+
+static void encode_block_pass1(int plane, int block, int blk_row, int blk_col,
+                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                               void *arg) {
+  encode_block_pass1_args *args = (encode_block_pass1_args *)arg;
+  AV1_COMMON *cm = args->cm;
+  MACROBLOCK *const x = args->x;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *const p = &x->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  uint8_t *dst;
+  int ctx = 0;
+  dst = &pd->dst
+             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
+
+  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                  ctx, AV1_XFORM_QUANT_B);
+#if !CONFIG_PVQ
+  if (p->eobs[block] > 0) {
+#else
+  if (!x->pvq_skip[plane]) {
+    {
+      int tx_blk_size;
+      int i, j;
+      // transform block size in pixels
+      tx_blk_size = tx_size_wide[tx_size];
+
+// Since av1 does not have separate function which does inverse transform
+// but av1_inv_txfm_add_*x*() also does addition of predicted image to
+// inverse transformed image,
+// pass blank dummy image to av1_inv_txfm_add_*x*(), i.e. set dst as zeros
+#if CONFIG_HIGHBITDEPTH
+      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        for (j = 0; j < tx_blk_size; j++)
+          for (i = 0; i < tx_blk_size; i++)
+            CONVERT_TO_SHORTPTR(dst)[j * pd->dst.stride + i] = 0;
+      } else {
+#endif  // CONFIG_HIGHBITDEPTH
+        for (j = 0; j < tx_blk_size; j++)
+          for (i = 0; i < tx_blk_size; i++) dst[j * pd->dst.stride + i] = 0;
+#if CONFIG_HIGHBITDEPTH
+      }
+#endif  // CONFIG_HIGHBITDEPTH
+    }
+#endif  // !CONFIG_PVQ
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
+        av1_highbd_iwht4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block],
+                               xd->bd);
+      } else {
+        av1_highbd_idct4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block],
+                               xd->bd);
+      }
+      return;
+    }
+#endif  //  CONFIG_HIGHBITDEPTH
+    if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
+      av1_iwht4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
+    } else {
+      av1_idct4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
+    }
+  }
+}
+
+void av1_encode_sby_pass1(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize) {
+  encode_block_pass1_args args = { cm, x };
+  av1_subtract_plane(x, bsize, 0);
+  av1_foreach_transformed_block_in_plane(&x->e_mbd, bsize, 0,
+                                         encode_block_pass1, &args);
+}
+
+void av1_encode_sb(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize,
+                   const int mi_row, const int mi_col) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct optimize_ctx ctx;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct encode_b_args arg = { cm, x, &ctx, &mbmi->skip, NULL, NULL, 1 };
+  int plane;
+
+  mbmi->skip = 1;
+
+  if (x->skip) return;
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+    const int subsampling_x = xd->plane[plane].subsampling_x;
+    const int subsampling_y = xd->plane[plane].subsampling_y;
+
+    if (!is_chroma_reference(mi_row, mi_col, bsize, subsampling_x,
+                             subsampling_y))
+      continue;
+
+    bsize = scale_chroma_bsize(bsize, subsampling_x, subsampling_y);
+#else
+    (void)mi_row;
+    (void)mi_col;
+#endif
+
+#if CONFIG_VAR_TX
+    // TODO(jingning): Clean this up.
+    const struct macroblockd_plane *const pd = &xd->plane[plane];
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+    const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+    const int mi_height = block_size_high[plane_bsize] >> tx_size_wide_log2[0];
+    const TX_SIZE max_tx_size = get_vartx_max_txsize(mbmi, plane_bsize);
+    const BLOCK_SIZE txb_size = txsize_to_bsize[max_tx_size];
+    const int bw = block_size_wide[txb_size] >> tx_size_wide_log2[0];
+    const int bh = block_size_high[txb_size] >> tx_size_wide_log2[0];
+    int idx, idy;
+    int block = 0;
+    int step = tx_size_wide_unit[max_tx_size] * tx_size_high_unit[max_tx_size];
+    av1_get_entropy_contexts(bsize, 0, pd, ctx.ta[plane], ctx.tl[plane]);
+#else
+    const struct macroblockd_plane *const pd = &xd->plane[plane];
+    const TX_SIZE tx_size = get_tx_size(plane, xd);
+    av1_get_entropy_contexts(bsize, tx_size, pd, ctx.ta[plane], ctx.tl[plane]);
+#endif
+
+#if !CONFIG_PVQ
+    av1_subtract_plane(x, bsize, plane);
+#endif
+    arg.ta = ctx.ta[plane];
+    arg.tl = ctx.tl[plane];
+
+#if CONFIG_VAR_TX
+    for (idy = 0; idy < mi_height; idy += bh) {
+      for (idx = 0; idx < mi_width; idx += bw) {
+        encode_block_inter(plane, block, idy, idx, plane_bsize, max_tx_size,
+                           &arg);
+        block += step;
+      }
+    }
+#else
+    av1_foreach_transformed_block_in_plane(xd, bsize, plane, encode_block,
+                                           &arg);
+#endif
+  }
+}
+
+#if CONFIG_SUPERTX
+void av1_encode_sb_supertx(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct optimize_ctx ctx;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct encode_b_args arg = { cm, x, &ctx, &mbmi->skip, NULL, NULL, 1 };
+  int plane;
+
+  mbmi->skip = 1;
+  if (x->skip) return;
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    const struct macroblockd_plane *const pd = &xd->plane[plane];
+#if CONFIG_VAR_TX
+    const TX_SIZE tx_size = TX_4X4;
+#else
+    const TX_SIZE tx_size = get_tx_size(plane, xd);
+#endif
+    av1_subtract_plane(x, bsize, plane);
+    av1_get_entropy_contexts(bsize, tx_size, pd, ctx.ta[plane], ctx.tl[plane]);
+    arg.ta = ctx.ta[plane];
+    arg.tl = ctx.tl[plane];
+    av1_foreach_transformed_block_in_plane(xd, bsize, plane, encode_block,
+                                           &arg);
+  }
+}
+#endif  // CONFIG_SUPERTX
+
+#if !CONFIG_PVQ
+void av1_set_txb_context(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
+                         ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l) {
+  (void)tx_size;
+  struct macroblock_plane *p = &x->plane[plane];
+
+#if !CONFIG_LV_MAP
+  *a = *l = p->eobs[block] > 0;
+#else   // !CONFIG_LV_MAP
+  *a = *l = p->txb_entropy_ctx[block];
+#endif  // !CONFIG_LV_MAP
+
+#if CONFIG_VAR_TX || CONFIG_LV_MAP
+  int i;
+  for (i = 0; i < tx_size_wide_unit[tx_size]; ++i) a[i] = a[0];
+
+  for (i = 0; i < tx_size_high_unit[tx_size]; ++i) l[i] = l[0];
+#endif
+}
+#endif
+
+static void encode_block_intra_and_set_context(int plane, int block,
+                                               int blk_row, int blk_col,
+                                               BLOCK_SIZE plane_bsize,
+                                               TX_SIZE tx_size, void *arg) {
+  av1_encode_block_intra(plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                         arg);
+#if !CONFIG_PVQ
+  struct encode_b_args *const args = arg;
+  MACROBLOCK *x = args->x;
+  ENTROPY_CONTEXT *a = &args->ta[blk_col];
+  ENTROPY_CONTEXT *l = &args->tl[blk_row];
+  av1_set_txb_context(x, plane, block, tx_size, a, l);
+#endif
+}
+
+void av1_encode_block_intra(int plane, int block, int blk_row, int blk_col,
+                            BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                            void *arg) {
+  struct encode_b_args *const args = arg;
+  AV1_COMMON *cm = args->cm;
+  MACROBLOCK *const x = args->x;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *const p = &x->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  uint16_t *eob = &p->eobs[block];
+  const int dst_stride = pd->dst.stride;
+  uint8_t *dst =
+      &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
+  av1_predict_intra_block_facade(xd, plane, block, blk_col, blk_row, tx_size);
+  av1_subtract_txb(x, plane, plane_bsize, blk_col, blk_row, tx_size);
+
+  const ENTROPY_CONTEXT *a = &args->ta[blk_col];
+  const ENTROPY_CONTEXT *l = &args->tl[blk_row];
+  int ctx = combine_entropy_contexts(*a, *l);
+  if (args->enable_optimize_b) {
+    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                    ctx, AV1_XFORM_QUANT_FP);
+    if (p->eobs[block]) {
+      av1_optimize_b(cm, x, plane, block, tx_size, ctx);
+    }
+  } else {
+    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                    ctx, AV1_XFORM_QUANT_B);
+  }
+
+#if CONFIG_PVQ
+  // *(args->skip) == mbmi->skip
+  if (!x->pvq_skip[plane]) *(args->skip) = 0;
+
+  if (x->pvq_skip[plane]) return;
+#endif  // CONFIG_PVQ
+  av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, dst, dst_stride,
+                              *eob);
+#if !CONFIG_PVQ
+  if (*eob) *(args->skip) = 0;
+#else
+// Note : *(args->skip) == mbmi->skip
+#endif
+#if CONFIG_CFL
+  if (plane == AOM_PLANE_Y && x->cfl_store_y) {
+    cfl_store(xd->cfl, dst, dst_stride, blk_row, blk_col, tx_size);
+  }
+#endif
+}
+
+void av1_encode_intra_block_plane(AV1_COMMON *cm, MACROBLOCK *x,
+                                  BLOCK_SIZE bsize, int plane,
+                                  int enable_optimize_b, const int mi_row,
+                                  const int mi_col) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  ENTROPY_CONTEXT ta[2 * MAX_MIB_SIZE] = { 0 };
+  ENTROPY_CONTEXT tl[2 * MAX_MIB_SIZE] = { 0 };
+
+  struct encode_b_args arg = {
+    cm, x, NULL, &xd->mi[0]->mbmi.skip, ta, tl, enable_optimize_b
+  };
+
+#if CONFIG_CB4X4
+  if (!is_chroma_reference(mi_row, mi_col, bsize,
+                           xd->plane[plane].subsampling_x,
+                           xd->plane[plane].subsampling_y))
+    return;
+#else
+  (void)mi_row;
+  (void)mi_col;
+#endif
+
+  if (enable_optimize_b) {
+    const struct macroblockd_plane *const pd = &xd->plane[plane];
+    const TX_SIZE tx_size = get_tx_size(plane, xd);
+    av1_get_entropy_contexts(bsize, tx_size, pd, ta, tl);
+  }
+  av1_foreach_transformed_block_in_plane(
+      xd, bsize, plane, encode_block_intra_and_set_context, &arg);
+}
+
+#if CONFIG_PVQ
+PVQ_SKIP_TYPE av1_pvq_encode_helper(MACROBLOCK *x, tran_low_t *const coeff,
+                                    tran_low_t *ref_coeff,
+                                    tran_low_t *const dqcoeff, uint16_t *eob,
+                                    const int16_t *quant, int plane,
+                                    int tx_size, TX_TYPE tx_type, int *rate,
+                                    int speed, PVQ_INFO *pvq_info) {
+  const int tx_blk_size = tx_size_wide[tx_size];
+  daala_enc_ctx *daala_enc = &x->daala_enc;
+  PVQ_SKIP_TYPE ac_dc_coded;
+  int coeff_shift = 3 - av1_get_tx_scale(tx_size);
+  int hbd_downshift = 0;
+  int rounding_mask;
+  int pvq_dc_quant;
+  int use_activity_masking = daala_enc->use_activity_masking;
+  int tell;
+  int has_dc_skip = 1;
+  int i;
+  int off = od_qm_offset(tx_size, plane ? 1 : 0);
+
+  DECLARE_ALIGNED(16, tran_low_t, coeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+  DECLARE_ALIGNED(16, tran_low_t, ref_coeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+
+  DECLARE_ALIGNED(16, int32_t, in_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+  DECLARE_ALIGNED(16, int32_t, ref_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+  DECLARE_ALIGNED(16, int32_t, out_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
+
+#if CONFIG_HIGHBITDEPTH
+  hbd_downshift = x->e_mbd.bd - 8;
+#endif
+
+  assert(OD_COEFF_SHIFT >= 4);
+  // DC quantizer for PVQ
+  if (use_activity_masking)
+    pvq_dc_quant =
+        OD_MAXI(1, (quant[0] << (OD_COEFF_SHIFT - 3) >> hbd_downshift) *
+                           daala_enc->state
+                               .pvq_qm_q4[plane][od_qm_get_index(tx_size, 0)] >>
+                       4);
+  else
+    pvq_dc_quant =
+        OD_MAXI(1, quant[0] << (OD_COEFF_SHIFT - 3) >> hbd_downshift);
+
+  *eob = 0;
+
+#if CONFIG_DAALA_EC
+  tell = od_ec_enc_tell_frac(&daala_enc->w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+
+  // Change coefficient ordering for pvq encoding.
+  od_raster_to_coding_order(coeff_pvq, tx_blk_size, tx_type, coeff,
+                            tx_blk_size);
+  od_raster_to_coding_order(ref_coeff_pvq, tx_blk_size, tx_type, ref_coeff,
+                            tx_blk_size);
+
+  // copy int16 inputs to int32
+  for (i = 0; i < tx_blk_size * tx_blk_size; i++) {
+    ref_int32[i] =
+        AOM_SIGNED_SHL(ref_coeff_pvq[i], OD_COEFF_SHIFT - coeff_shift) >>
+        hbd_downshift;
+    in_int32[i] = AOM_SIGNED_SHL(coeff_pvq[i], OD_COEFF_SHIFT - coeff_shift) >>
+                  hbd_downshift;
+  }
+
+  if (abs(in_int32[0] - ref_int32[0]) < pvq_dc_quant * 141 / 256) { /* 0.55 */
+    out_int32[0] = 0;
+  } else {
+    out_int32[0] = OD_DIV_R0(in_int32[0] - ref_int32[0], pvq_dc_quant);
+  }
+
+  ac_dc_coded =
+      od_pvq_encode(daala_enc, ref_int32, in_int32, out_int32,
+                    OD_MAXI(1, quant[0] << (OD_COEFF_SHIFT - 3) >>
+                                   hbd_downshift),  // scale/quantizer
+                    OD_MAXI(1, quant[1] << (OD_COEFF_SHIFT - 3) >>
+                                   hbd_downshift),  // scale/quantizer
+                    plane,
+                    tx_size, OD_PVQ_BETA[use_activity_masking][plane][tx_size],
+                    0,  // is_keyframe,
+                    daala_enc->state.qm + off, daala_enc->state.qm_inv + off,
+                    speed,  // speed
+                    pvq_info);
+
+  // Encode residue of DC coeff, if required.
+  if (!has_dc_skip || out_int32[0]) {
+    generic_encode(&daala_enc->w, &daala_enc->state.adapt->model_dc[plane],
+                   abs(out_int32[0]) - has_dc_skip,
+                   &daala_enc->state.adapt->ex_dc[plane][tx_size][0], 2);
+  }
+  if (out_int32[0]) {
+    aom_write_bit(&daala_enc->w, out_int32[0] < 0);
+  }
+
+  // need to save quantized residue of DC coeff
+  // so that final pvq bitstream writing can know whether DC is coded.
+  if (pvq_info) pvq_info->dq_dc_residue = out_int32[0];
+
+  out_int32[0] = out_int32[0] * pvq_dc_quant;
+  out_int32[0] += ref_int32[0];
+
+  // copy int32 result back to int16
+  assert(OD_COEFF_SHIFT > coeff_shift);
+  rounding_mask = (1 << (OD_COEFF_SHIFT - coeff_shift - 1)) - 1;
+  for (i = 0; i < tx_blk_size * tx_blk_size; i++) {
+    out_int32[i] = AOM_SIGNED_SHL(out_int32[i], hbd_downshift);
+    dqcoeff_pvq[i] = (out_int32[i] + (out_int32[i] < 0) + rounding_mask) >>
+                     (OD_COEFF_SHIFT - coeff_shift);
+  }
+
+  // Back to original coefficient order
+  od_coding_order_to_raster(dqcoeff, tx_blk_size, tx_type, dqcoeff_pvq,
+                            tx_blk_size);
+
+  *eob = tx_blk_size * tx_blk_size;
+
+#if CONFIG_DAALA_EC
+  *rate = (od_ec_enc_tell_frac(&daala_enc->w.ec) - tell)
+          << (AV1_PROB_COST_SHIFT - OD_BITRES);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  assert(*rate >= 0);
+
+  return ac_dc_coded;
+}
+
+void av1_store_pvq_enc_info(PVQ_INFO *pvq_info, int *qg, int *theta, int *k,
+                            od_coeff *y, int nb_bands, const int *off,
+                            int *size, int skip_rest, int skip_dir,
+                            int bs) {  // block size in log_2 -2
+  int i;
+  const int tx_blk_size = tx_size_wide[bs];
+
+  for (i = 0; i < nb_bands; i++) {
+    pvq_info->qg[i] = qg[i];
+    pvq_info->theta[i] = theta[i];
+    pvq_info->k[i] = k[i];
+    pvq_info->off[i] = off[i];
+    pvq_info->size[i] = size[i];
+  }
+
+  memcpy(pvq_info->y, y, tx_blk_size * tx_blk_size * sizeof(od_coeff));
+
+  pvq_info->nb_bands = nb_bands;
+  pvq_info->skip_rest = skip_rest;
+  pvq_info->skip_dir = skip_dir;
+  pvq_info->bs = bs;
+}
+#endif
diff --git a/third_party/aom/av1/encoder/encodemb.h b/third_party/aom/av1/encoder/encodemb.h
new file mode 100644
index 0000000000..73fde1d884
--- /dev/null
+++ b/third_party/aom/av1/encoder/encodemb.h
@@ -0,0 +1,92 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_ENCODEMB_H_
+#define AV1_ENCODER_ENCODEMB_H_
+
+#include "./aom_config.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/encoder/block.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct optimize_ctx {
+  ENTROPY_CONTEXT ta[MAX_MB_PLANE][2 * MAX_MIB_SIZE];
+  ENTROPY_CONTEXT tl[MAX_MB_PLANE][2 * MAX_MIB_SIZE];
+};
+
+struct encode_b_args {
+  AV1_COMMON *cm;
+  MACROBLOCK *x;
+  struct optimize_ctx *ctx;
+  int8_t *skip;
+  ENTROPY_CONTEXT *ta;
+  ENTROPY_CONTEXT *tl;
+  int8_t enable_optimize_b;
+};
+
+typedef enum AV1_XFORM_QUANT {
+  AV1_XFORM_QUANT_FP = 0,
+  AV1_XFORM_QUANT_B = 1,
+  AV1_XFORM_QUANT_DC = 2,
+  AV1_XFORM_QUANT_SKIP_QUANT,
+  AV1_XFORM_QUANT_TYPES,
+} AV1_XFORM_QUANT;
+
+void av1_encode_sb(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row,
+                   int mi_col);
+#if CONFIG_SUPERTX
+void av1_encode_sb_supertx(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize);
+#endif  // CONFIG_SUPERTX
+void av1_encode_sby_pass1(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize);
+void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
+                     int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
+                     TX_SIZE tx_size, int ctx, AV1_XFORM_QUANT xform_quant_idx);
+
+int av1_optimize_b(const AV1_COMMON *cm, MACROBLOCK *mb, int plane, int block,
+                   TX_SIZE tx_size, int ctx);
+
+void av1_subtract_txb(MACROBLOCK *x, int plane, BLOCK_SIZE plane_bsize,
+                      int blk_col, int blk_row, TX_SIZE tx_size);
+
+void av1_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+
+void av1_set_txb_context(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
+                         ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l);
+
+void av1_encode_block_intra(int plane, int block, int blk_row, int blk_col,
+                            BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg);
+
+void av1_encode_intra_block_plane(AV1_COMMON *cm, MACROBLOCK *x,
+                                  BLOCK_SIZE bsize, int plane,
+                                  int enable_optimize_b, int mi_row,
+                                  int mi_col);
+
+#if CONFIG_PVQ
+PVQ_SKIP_TYPE av1_pvq_encode_helper(MACROBLOCK *x, tran_low_t *const coeff,
+                                    tran_low_t *ref_coeff,
+                                    tran_low_t *const dqcoeff, uint16_t *eob,
+                                    const int16_t *quant, int plane,
+                                    int tx_size, TX_TYPE tx_type, int *rate,
+                                    int speed, PVQ_INFO *pvq_info);
+
+void av1_store_pvq_enc_info(PVQ_INFO *pvq_info, int *qg, int *theta, int *k,
+                            od_coeff *y, int nb_bands, const int *off,
+                            int *size, int skip_rest, int skip_dir, int bs);
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_ENCODEMB_H_
diff --git a/third_party/aom/av1/encoder/encodemv.c b/third_party/aom/av1/encoder/encodemv.c
new file mode 100644
index 0000000000..a2a53f8408
--- /dev/null
+++ b/third_party/aom/av1/encoder/encodemv.c
@@ -0,0 +1,497 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+
+#include "av1/common/common.h"
+#include "av1/common/entropymode.h"
+
+#include "av1/encoder/cost.h"
+#include "av1/encoder/encodemv.h"
+#include "av1/encoder/subexp.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+
+static struct av1_token mv_joint_encodings[MV_JOINTS];
+static struct av1_token mv_class_encodings[MV_CLASSES];
+static struct av1_token mv_fp_encodings[MV_FP_SIZE];
+
+void av1_entropy_mv_init(void) {
+  av1_tokens_from_tree(mv_joint_encodings, av1_mv_joint_tree);
+  av1_tokens_from_tree(mv_class_encodings, av1_mv_class_tree);
+  av1_tokens_from_tree(mv_fp_encodings, av1_mv_fp_tree);
+}
+
+static void encode_mv_component(aom_writer *w, int comp, nmv_component *mvcomp,
+                                int usehp) {
+  int offset;
+  const int sign = comp < 0;
+  const int mag = sign ? -comp : comp;
+  const int mv_class = av1_get_mv_class(mag - 1, &offset);
+  const int d = offset >> 3;         // int mv data
+  const int fr = (offset >> 1) & 3;  // fractional mv data
+  const int hp = offset & 1;         // high precision mv data
+
+  assert(comp != 0);
+
+  // Sign
+  aom_write(w, sign, mvcomp->sign);
+
+// Class
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, mv_class, mvcomp->class_cdf, MV_CLASSES);
+#else
+  av1_write_token(w, av1_mv_class_tree, mvcomp->classes,
+                  &mv_class_encodings[mv_class]);
+#endif
+
+  // Integer bits
+  if (mv_class == MV_CLASS_0) {
+    aom_write(w, d, mvcomp->class0[0]);
+  } else {
+    int i;
+    const int n = mv_class + CLASS0_BITS - 1;  // number of bits
+    for (i = 0; i < n; ++i) aom_write(w, (d >> i) & 1, mvcomp->bits[i]);
+  }
+
+// Fractional bits
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(
+      w, fr, mv_class == MV_CLASS_0 ? mvcomp->class0_fp_cdf[d] : mvcomp->fp_cdf,
+      MV_FP_SIZE);
+#else
+  av1_write_token(w, av1_mv_fp_tree,
+                  mv_class == MV_CLASS_0 ? mvcomp->class0_fp[d] : mvcomp->fp,
+                  &mv_fp_encodings[fr]);
+#endif
+
+  // High precision bit
+  if (usehp)
+    aom_write(w, hp, mv_class == MV_CLASS_0 ? mvcomp->class0_hp : mvcomp->hp);
+}
+
+static void build_nmv_component_cost_table(int *mvcost,
+                                           const nmv_component *const mvcomp,
+                                           int usehp) {
+  int i, v;
+  int sign_cost[2], class_cost[MV_CLASSES], class0_cost[CLASS0_SIZE];
+  int bits_cost[MV_OFFSET_BITS][2];
+  int class0_fp_cost[CLASS0_SIZE][MV_FP_SIZE], fp_cost[MV_FP_SIZE];
+  int class0_hp_cost[2], hp_cost[2];
+
+  sign_cost[0] = av1_cost_zero(mvcomp->sign);
+  sign_cost[1] = av1_cost_one(mvcomp->sign);
+  av1_cost_tokens(class_cost, mvcomp->classes, av1_mv_class_tree);
+  av1_cost_tokens(class0_cost, mvcomp->class0, av1_mv_class0_tree);
+  for (i = 0; i < MV_OFFSET_BITS; ++i) {
+    bits_cost[i][0] = av1_cost_zero(mvcomp->bits[i]);
+    bits_cost[i][1] = av1_cost_one(mvcomp->bits[i]);
+  }
+
+  for (i = 0; i < CLASS0_SIZE; ++i)
+    av1_cost_tokens(class0_fp_cost[i], mvcomp->class0_fp[i], av1_mv_fp_tree);
+  av1_cost_tokens(fp_cost, mvcomp->fp, av1_mv_fp_tree);
+
+  if (usehp) {
+    class0_hp_cost[0] = av1_cost_zero(mvcomp->class0_hp);
+    class0_hp_cost[1] = av1_cost_one(mvcomp->class0_hp);
+    hp_cost[0] = av1_cost_zero(mvcomp->hp);
+    hp_cost[1] = av1_cost_one(mvcomp->hp);
+  }
+  mvcost[0] = 0;
+  for (v = 1; v <= MV_MAX; ++v) {
+    int z, c, o, d, e, f, cost = 0;
+    z = v - 1;
+    c = av1_get_mv_class(z, &o);
+    cost += class_cost[c];
+    d = (o >> 3);     /* int mv data */
+    f = (o >> 1) & 3; /* fractional pel mv data */
+    e = (o & 1);      /* high precision mv data */
+    if (c == MV_CLASS_0) {
+      cost += class0_cost[d];
+    } else {
+      const int b = c + CLASS0_BITS - 1; /* number of bits */
+      for (i = 0; i < b; ++i) cost += bits_cost[i][((d >> i) & 1)];
+    }
+    if (c == MV_CLASS_0) {
+      cost += class0_fp_cost[d][f];
+    } else {
+      cost += fp_cost[f];
+    }
+    if (usehp) {
+      if (c == MV_CLASS_0) {
+        cost += class0_hp_cost[e];
+      } else {
+        cost += hp_cost[e];
+      }
+    }
+    mvcost[v] = cost + sign_cost[0];
+    mvcost[-v] = cost + sign_cost[1];
+  }
+}
+
+static void update_mv(aom_writer *w, const unsigned int ct[2], aom_prob *cur_p,
+                      aom_prob upd_p) {
+  (void)upd_p;
+#if CONFIG_TILE_GROUPS
+  // Just use the default maximum number of tile groups to avoid passing in the
+  // actual
+  // number
+  av1_cond_prob_diff_update(w, cur_p, ct, DEFAULT_MAX_NUM_TG);
+#else
+  av1_cond_prob_diff_update(w, cur_p, ct, 1);
+#endif
+}
+
+#if !CONFIG_EC_ADAPT
+static void write_mv_update(const aom_tree_index *tree,
+                            aom_prob probs[/*n - 1*/],
+                            const unsigned int counts[/*n - 1*/], int n,
+                            aom_writer *w) {
+  int i;
+  unsigned int branch_ct[32][2];
+
+  // Assuming max number of probabilities <= 32
+  assert(n <= 32);
+
+  av1_tree_probs_from_distribution(tree, branch_ct, counts);
+  for (i = 0; i < n - 1; ++i)
+    update_mv(w, branch_ct[i], &probs[i], MV_UPDATE_PROB);
+}
+#endif
+
+void av1_write_nmv_probs(AV1_COMMON *cm, int usehp, aom_writer *w,
+                         nmv_context_counts *const nmv_counts) {
+  int i;
+#if CONFIG_REF_MV
+  int nmv_ctx = 0;
+  for (nmv_ctx = 0; nmv_ctx < NMV_CONTEXTS; ++nmv_ctx) {
+    nmv_context *const mvc = &cm->fc->nmvc[nmv_ctx];
+    nmv_context_counts *const counts = &nmv_counts[nmv_ctx];
+#if !CONFIG_EC_ADAPT
+    write_mv_update(av1_mv_joint_tree, mvc->joints, counts->joints, MV_JOINTS,
+                    w);
+
+    for (i = 0; i < 2; ++i) {
+      int j;
+      nmv_component *comp = &mvc->comps[i];
+      nmv_component_counts *comp_counts = &counts->comps[i];
+
+      update_mv(w, comp_counts->sign, &comp->sign, MV_UPDATE_PROB);
+      write_mv_update(av1_mv_class_tree, comp->classes, comp_counts->classes,
+                      MV_CLASSES, w);
+      write_mv_update(av1_mv_class0_tree, comp->class0, comp_counts->class0,
+                      CLASS0_SIZE, w);
+      for (j = 0; j < MV_OFFSET_BITS; ++j)
+        update_mv(w, comp_counts->bits[j], &comp->bits[j], MV_UPDATE_PROB);
+    }
+
+    for (i = 0; i < 2; ++i) {
+      int j;
+      for (j = 0; j < CLASS0_SIZE; ++j)
+        write_mv_update(av1_mv_fp_tree, mvc->comps[i].class0_fp[j],
+                        counts->comps[i].class0_fp[j], MV_FP_SIZE, w);
+
+      write_mv_update(av1_mv_fp_tree, mvc->comps[i].fp, counts->comps[i].fp,
+                      MV_FP_SIZE, w);
+    }
+#endif
+
+    if (usehp) {
+      for (i = 0; i < 2; ++i) {
+        update_mv(w, counts->comps[i].class0_hp, &mvc->comps[i].class0_hp,
+                  MV_UPDATE_PROB);
+        update_mv(w, counts->comps[i].hp, &mvc->comps[i].hp, MV_UPDATE_PROB);
+      }
+    }
+  }
+#else
+  nmv_context *const mvc = &cm->fc->nmvc;
+  nmv_context_counts *const counts = nmv_counts;
+
+#if !CONFIG_EC_ADAPT
+  write_mv_update(av1_mv_joint_tree, mvc->joints, counts->joints, MV_JOINTS, w);
+
+  for (i = 0; i < 2; ++i) {
+    int j;
+    nmv_component *comp = &mvc->comps[i];
+    nmv_component_counts *comp_counts = &counts->comps[i];
+
+    update_mv(w, comp_counts->sign, &comp->sign, MV_UPDATE_PROB);
+    write_mv_update(av1_mv_class_tree, comp->classes, comp_counts->classes,
+                    MV_CLASSES, w);
+    write_mv_update(av1_mv_class0_tree, comp->class0, comp_counts->class0,
+                    CLASS0_SIZE, w);
+    for (j = 0; j < MV_OFFSET_BITS; ++j)
+      update_mv(w, comp_counts->bits[j], &comp->bits[j], MV_UPDATE_PROB);
+  }
+
+  for (i = 0; i < 2; ++i) {
+    int j;
+    for (j = 0; j < CLASS0_SIZE; ++j) {
+      write_mv_update(av1_mv_fp_tree, mvc->comps[i].class0_fp[j],
+                      counts->comps[i].class0_fp[j], MV_FP_SIZE, w);
+    }
+    write_mv_update(av1_mv_fp_tree, mvc->comps[i].fp, counts->comps[i].fp,
+                    MV_FP_SIZE, w);
+  }
+#endif  // !CONFIG_EC_ADAPT
+
+  if (usehp) {
+    for (i = 0; i < 2; ++i) {
+      update_mv(w, counts->comps[i].class0_hp, &mvc->comps[i].class0_hp,
+                MV_UPDATE_PROB);
+      update_mv(w, counts->comps[i].hp, &mvc->comps[i].hp, MV_UPDATE_PROB);
+    }
+  }
+#endif
+}
+
+void av1_encode_mv(AV1_COMP *cpi, aom_writer *w, const MV *mv, const MV *ref,
+                   nmv_context *mvctx, int usehp) {
+  const MV diff = { mv->row - ref->row, mv->col - ref->col };
+  const MV_JOINT_TYPE j = av1_get_mv_joint(&diff);
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, j, mvctx->joint_cdf, MV_JOINTS);
+#else
+  av1_write_token(w, av1_mv_joint_tree, mvctx->joints, &mv_joint_encodings[j]);
+#endif
+  if (mv_joint_vertical(j))
+    encode_mv_component(w, diff.row, &mvctx->comps[0], usehp);
+
+  if (mv_joint_horizontal(j))
+    encode_mv_component(w, diff.col, &mvctx->comps[1], usehp);
+
+  // If auto_mv_step_size is enabled then keep track of the largest
+  // motion vector component used.
+  if (cpi->sf.mv.auto_mv_step_size) {
+    unsigned int maxv = AOMMAX(abs(mv->row), abs(mv->col)) >> 3;
+    cpi->max_mv_magnitude = AOMMAX(maxv, cpi->max_mv_magnitude);
+  }
+}
+
+#if CONFIG_INTRABC
+void av1_encode_dv(aom_writer *w, const MV *mv, const MV *ref,
+                   nmv_context *mvctx) {
+  const MV diff = { mv->row - ref->row, mv->col - ref->col };
+  const MV_JOINT_TYPE j = av1_get_mv_joint(&diff);
+
+#if CONFIG_EC_MULTISYMBOL
+  aom_write_symbol(w, j, mvctx->joint_cdf, MV_JOINTS);
+#else
+  av1_write_token(w, av1_mv_joint_tree, mvctx->joints, &mv_joint_encodings[j]);
+#endif
+  if (mv_joint_vertical(j))
+    encode_mv_component(w, diff.row, &mvctx->comps[0], 0);
+
+  if (mv_joint_horizontal(j))
+    encode_mv_component(w, diff.col, &mvctx->comps[1], 0);
+}
+#endif  // CONFIG_INTRABC
+
+void av1_build_nmv_cost_table(int *mvjoint, int *mvcost[2],
+                              const nmv_context *ctx, int usehp) {
+  av1_cost_tokens(mvjoint, ctx->joints, av1_mv_joint_tree);
+  build_nmv_component_cost_table(mvcost[0], &ctx->comps[0], usehp);
+  build_nmv_component_cost_table(mvcost[1], &ctx->comps[1], usehp);
+}
+
+#if CONFIG_EXT_INTER
+static void inc_mvs(const MB_MODE_INFO *mbmi, const MB_MODE_INFO_EXT *mbmi_ext,
+                    const int_mv mvs[2],
+#if CONFIG_REF_MV
+                    const int_mv pred_mvs[2],
+#endif
+                    nmv_context_counts *nmv_counts) {
+  int i;
+  PREDICTION_MODE mode = mbmi->mode;
+#if !CONFIG_REF_MV
+  nmv_context_counts *counts = nmv_counts;
+#endif
+
+  if (mode == NEWMV || mode == NEW_NEWMV) {
+    for (i = 0; i < 1 + has_second_ref(mbmi); ++i) {
+      const MV *ref = &mbmi_ext->ref_mvs[mbmi->ref_frame[i]][0].as_mv;
+      const MV diff = { mvs[i].as_mv.row - ref->row,
+                        mvs[i].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx =
+          av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                      mbmi_ext->ref_mv_stack[rf_type], i, mbmi->ref_mv_idx);
+      nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+      (void)pred_mvs;
+#endif
+      av1_inc_mv(&diff, counts, 1);
+    }
+  } else if (mode == NEAREST_NEWMV || mode == NEAR_NEWMV) {
+    const MV *ref = &mbmi_ext->ref_mvs[mbmi->ref_frame[1]][0].as_mv;
+    const MV diff = { mvs[1].as_mv.row - ref->row,
+                      mvs[1].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+    int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+    int nmv_ctx =
+        av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                    mbmi_ext->ref_mv_stack[rf_type], 1, mbmi->ref_mv_idx);
+    nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+#endif
+    av1_inc_mv(&diff, counts, 1);
+  } else if (mode == NEW_NEARESTMV || mode == NEW_NEARMV) {
+    const MV *ref = &mbmi_ext->ref_mvs[mbmi->ref_frame[0]][0].as_mv;
+    const MV diff = { mvs[0].as_mv.row - ref->row,
+                      mvs[0].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+    int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+    int nmv_ctx =
+        av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                    mbmi_ext->ref_mv_stack[rf_type], 0, mbmi->ref_mv_idx);
+    nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+#endif
+    av1_inc_mv(&diff, counts, 1);
+  }
+}
+
+static void inc_mvs_sub8x8(const MODE_INFO *mi, int block, const int_mv mvs[2],
+#if CONFIG_REF_MV
+                           const MB_MODE_INFO_EXT *mbmi_ext,
+#endif
+                           nmv_context_counts *nmv_counts) {
+  int i;
+  PREDICTION_MODE mode = mi->bmi[block].as_mode;
+#if CONFIG_REF_MV
+  const MB_MODE_INFO *mbmi = &mi->mbmi;
+#else
+  nmv_context_counts *counts = nmv_counts;
+#endif
+
+  if (mode == NEWMV || mode == NEW_NEWMV) {
+    for (i = 0; i < 1 + has_second_ref(&mi->mbmi); ++i) {
+      const MV *ref = &mi->bmi[block].ref_mv[i].as_mv;
+      const MV diff = { mvs[i].as_mv.row - ref->row,
+                        mvs[i].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+      int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+      int nmv_ctx =
+          av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                      mbmi_ext->ref_mv_stack[rf_type], i, mbmi->ref_mv_idx);
+      nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+#endif
+      av1_inc_mv(&diff, counts, 1);
+    }
+  } else if (mode == NEAREST_NEWMV || mode == NEAR_NEWMV) {
+    const MV *ref = &mi->bmi[block].ref_mv[1].as_mv;
+    const MV diff = { mvs[1].as_mv.row - ref->row,
+                      mvs[1].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+    int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+    int nmv_ctx =
+        av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                    mbmi_ext->ref_mv_stack[rf_type], 1, mbmi->ref_mv_idx);
+    nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+#endif
+    av1_inc_mv(&diff, counts, 1);
+  } else if (mode == NEW_NEARESTMV || mode == NEW_NEARMV) {
+    const MV *ref = &mi->bmi[block].ref_mv[0].as_mv;
+    const MV diff = { mvs[0].as_mv.row - ref->row,
+                      mvs[0].as_mv.col - ref->col };
+#if CONFIG_REF_MV
+    int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+    int nmv_ctx =
+        av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                    mbmi_ext->ref_mv_stack[rf_type], 0, mbmi->ref_mv_idx);
+    nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+#endif
+    av1_inc_mv(&diff, counts, 1);
+  }
+}
+#else
+static void inc_mvs(const MB_MODE_INFO *mbmi, const MB_MODE_INFO_EXT *mbmi_ext,
+                    const int_mv mvs[2],
+#if CONFIG_REF_MV
+                    const int_mv pred_mvs[2],
+#endif
+                    nmv_context_counts *nmv_counts) {
+  int i;
+#if !CONFIG_REF_MV
+  nmv_context_counts *counts = nmv_counts;
+#endif
+
+  for (i = 0; i < 1 + has_second_ref(mbmi); ++i) {
+#if CONFIG_REF_MV
+    int8_t rf_type = av1_ref_frame_type(mbmi->ref_frame);
+    int nmv_ctx =
+        av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                    mbmi_ext->ref_mv_stack[rf_type], i, mbmi->ref_mv_idx);
+    nmv_context_counts *counts = &nmv_counts[nmv_ctx];
+    const MV *ref = &pred_mvs[i].as_mv;
+#else
+    const MV *ref = &mbmi_ext->ref_mvs[mbmi->ref_frame[i]][0].as_mv;
+#endif
+    const MV diff = { mvs[i].as_mv.row - ref->row,
+                      mvs[i].as_mv.col - ref->col };
+    av1_inc_mv(&diff, counts, 1);
+  }
+}
+#endif  // CONFIG_EXT_INTER
+
+void av1_update_mv_count(ThreadData *td) {
+  const MACROBLOCKD *xd = &td->mb.e_mbd;
+  const MODE_INFO *mi = xd->mi[0];
+  const MB_MODE_INFO *const mbmi = &mi->mbmi;
+  const MB_MODE_INFO_EXT *mbmi_ext = td->mb.mbmi_ext;
+#if CONFIG_CB4X4
+  const int unify_bsize = 1;
+#else
+  const int unify_bsize = 0;
+#endif
+
+  if (mbmi->sb_type < BLOCK_8X8 && !unify_bsize) {
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[mbmi->sb_type];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[mbmi->sb_type];
+    int idx, idy;
+
+    for (idy = 0; idy < 2; idy += num_4x4_h) {
+      for (idx = 0; idx < 2; idx += num_4x4_w) {
+        const int i = idy * 2 + idx;
+
+#if CONFIG_EXT_INTER
+        if (have_newmv_in_inter_mode(mi->bmi[i].as_mode))
+          inc_mvs_sub8x8(mi, i, mi->bmi[i].as_mv,
+#if CONFIG_REF_MV
+                         mbmi_ext, td->counts->mv);
+#else
+                         &td->counts->mv);
+#endif
+#else
+        if (mi->bmi[i].as_mode == NEWMV)
+          inc_mvs(mbmi, mbmi_ext, mi->bmi[i].as_mv,
+#if CONFIG_REF_MV
+                  mi->bmi[i].pred_mv, td->counts->mv);
+#else
+                  &td->counts->mv);
+#endif
+#endif  // CONFIG_EXT_INTER
+      }
+    }
+  } else {
+#if CONFIG_EXT_INTER
+    if (have_newmv_in_inter_mode(mbmi->mode))
+#else
+    if (mbmi->mode == NEWMV)
+#endif  // CONFIG_EXT_INTER
+      inc_mvs(mbmi, mbmi_ext, mbmi->mv,
+#if CONFIG_REF_MV
+              mbmi->pred_mv, td->counts->mv);
+#else
+              &td->counts->mv);
+#endif
+  }
+}
diff --git a/third_party/aom/av1/encoder/encodemv.h b/third_party/aom/av1/encoder/encodemv.h
new file mode 100644
index 0000000000..6d442147fe
--- /dev/null
+++ b/third_party/aom/av1/encoder/encodemv.h
@@ -0,0 +1,43 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_ENCODEMV_H_
+#define AV1_ENCODER_ENCODEMV_H_
+
+#include "av1/encoder/encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_entropy_mv_init(void);
+
+void av1_write_nmv_probs(AV1_COMMON *cm, int usehp, aom_writer *w,
+                         nmv_context_counts *const counts);
+
+void av1_encode_mv(AV1_COMP *cpi, aom_writer *w, const MV *mv, const MV *ref,
+                   nmv_context *mvctx, int usehp);
+
+void av1_build_nmv_cost_table(int *mvjoint, int *mvcost[2],
+                              const nmv_context *mvctx, int usehp);
+
+void av1_update_mv_count(ThreadData *td);
+
+#if CONFIG_INTRABC
+void av1_encode_dv(aom_writer *w, const MV *mv, const MV *ref,
+                   nmv_context *mvctx);
+#endif  // CONFIG_INTRABC
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_ENCODEMV_H_
diff --git a/third_party/aom/av1/encoder/encoder.c b/third_party/aom/av1/encoder/encoder.c
new file mode 100644
index 0000000000..0271091519
--- /dev/null
+++ b/third_party/aom/av1/encoder/encoder.c
@@ -0,0 +1,5980 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./aom_config.h"
+
+#include "av1/common/alloccommon.h"
+#if CONFIG_CDEF
+#include "av1/common/cdef.h"
+#include "av1/common/clpf.h"
+#endif  // CONFIG_CDEF
+#include "av1/common/filter.h"
+#include "av1/common/idct.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+#include "av1/common/resize.h"
+#include "av1/common/tile_common.h"
+
+#include "av1/encoder/aq_complexity.h"
+#include "av1/encoder/aq_cyclicrefresh.h"
+#include "av1/encoder/aq_variance.h"
+#include "av1/encoder/bitstream.h"
+#if CONFIG_ANS
+#include "aom_dsp/buf_ans.h"
+#endif
+#include "av1/encoder/context_tree.h"
+#include "av1/encoder/encodeframe.h"
+#include "av1/encoder/encodemv.h"
+#include "av1/encoder/encoder.h"
+#if CONFIG_LV_MAP
+#include "av1/encoder/encodetxb.h"
+#endif
+#include "av1/encoder/ethread.h"
+#include "av1/encoder/firstpass.h"
+#include "av1/encoder/mbgraph.h"
+#include "av1/encoder/picklpf.h"
+#if CONFIG_LOOP_RESTORATION
+#include "av1/encoder/pickrst.h"
+#endif  // CONFIG_LOOP_RESTORATION
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/rd.h"
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/speed_features.h"
+#include "av1/encoder/temporal_filter.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "./aom_scale_rtcd.h"
+#include "aom_dsp/psnr.h"
+#if CONFIG_INTERNAL_STATS
+#include "aom_dsp/ssim.h"
+#endif
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_ports/aom_timer.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/system_state.h"
+#include "aom_scale/aom_scale.h"
+#if CONFIG_BITSTREAM_DEBUG
+#include "aom_util/debug_util.h"
+#endif  // CONFIG_BITSTREAM_DEBUG
+
+#if CONFIG_ENTROPY_STATS
+FRAME_COUNTS aggregate_fc;
+#endif  // CONFIG_ENTROPY_STATS
+
+#define AM_SEGMENT_ID_INACTIVE 7
+#define AM_SEGMENT_ID_ACTIVE 0
+
+#define SHARP_FILTER_QTHRESH 0 /* Q threshold for 8-tap sharp filter */
+
+#define ALTREF_HIGH_PRECISION_MV 1     // Whether to use high precision mv
+                                       //  for altref computation.
+#define HIGH_PRECISION_MV_QTHRESH 200  // Q threshold for high precision
+                                       // mv. Choose a very high value for
+                                       // now so that HIGH_PRECISION is always
+                                       // chosen.
+// #define OUTPUT_YUV_REC
+#ifdef OUTPUT_YUV_DENOISED
+FILE *yuv_denoised_file = NULL;
+#endif
+#ifdef OUTPUT_YUV_SKINMAP
+FILE *yuv_skinmap_file = NULL;
+#endif
+#ifdef OUTPUT_YUV_REC
+FILE *yuv_rec_file;
+#define FILE_NAME_LEN 100
+#endif
+
+#if 0
+FILE *framepsnr;
+FILE *kf_list;
+FILE *keyfile;
+#endif
+
+#if CONFIG_CFL
+CFL_CTX NULL_CFL;
+#endif
+
+#if CONFIG_INTERNAL_STATS
+typedef enum { Y, U, V, ALL } STAT_TYPE;
+#endif  // CONFIG_INTERNAL_STATS
+
+static INLINE void Scale2Ratio(AOM_SCALING mode, int *hr, int *hs) {
+  switch (mode) {
+    case NORMAL:
+      *hr = 1;
+      *hs = 1;
+      break;
+    case FOURFIVE:
+      *hr = 4;
+      *hs = 5;
+      break;
+    case THREEFIVE:
+      *hr = 3;
+      *hs = 5;
+      break;
+    case ONETWO:
+      *hr = 1;
+      *hs = 2;
+      break;
+    default:
+      *hr = 1;
+      *hs = 1;
+      assert(0);
+      break;
+  }
+}
+
+// Mark all inactive blocks as active. Other segmentation features may be set
+// so memset cannot be used, instead only inactive blocks should be reset.
+static void suppress_active_map(AV1_COMP *cpi) {
+  unsigned char *const seg_map = cpi->segmentation_map;
+  int i;
+  if (cpi->active_map.enabled || cpi->active_map.update)
+    for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i)
+      if (seg_map[i] == AM_SEGMENT_ID_INACTIVE)
+        seg_map[i] = AM_SEGMENT_ID_ACTIVE;
+}
+
+static void apply_active_map(AV1_COMP *cpi) {
+  struct segmentation *const seg = &cpi->common.seg;
+  unsigned char *const seg_map = cpi->segmentation_map;
+  const unsigned char *const active_map = cpi->active_map.map;
+  int i;
+
+  assert(AM_SEGMENT_ID_ACTIVE == CR_SEGMENT_ID_BASE);
+
+  if (frame_is_intra_only(&cpi->common)) {
+    cpi->active_map.enabled = 0;
+    cpi->active_map.update = 1;
+  }
+
+  if (cpi->active_map.update) {
+    if (cpi->active_map.enabled) {
+      for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i)
+        if (seg_map[i] == AM_SEGMENT_ID_ACTIVE) seg_map[i] = active_map[i];
+      av1_enable_segmentation(seg);
+      av1_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
+      av1_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
+      // Setting the data to -MAX_LOOP_FILTER will result in the computed loop
+      // filter level being zero regardless of the value of seg->abs_delta.
+      av1_set_segdata(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF,
+                      -MAX_LOOP_FILTER);
+    } else {
+      av1_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
+      av1_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
+      if (seg->enabled) {
+        seg->update_data = 1;
+        seg->update_map = 1;
+      }
+    }
+    cpi->active_map.update = 0;
+  }
+}
+
+int av1_set_active_map(AV1_COMP *cpi, unsigned char *new_map_16x16, int rows,
+                       int cols) {
+  if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols) {
+    unsigned char *const active_map_8x8 = cpi->active_map.map;
+    const int mi_rows = cpi->common.mi_rows;
+    const int mi_cols = cpi->common.mi_cols;
+    const int row_scale = mi_size_high[BLOCK_16X16] == 2 ? 1 : 2;
+    const int col_scale = mi_size_wide[BLOCK_16X16] == 2 ? 1 : 2;
+    cpi->active_map.update = 1;
+    if (new_map_16x16) {
+      int r, c;
+      for (r = 0; r < mi_rows; ++r) {
+        for (c = 0; c < mi_cols; ++c) {
+          active_map_8x8[r * mi_cols + c] =
+              new_map_16x16[(r >> row_scale) * cols + (c >> col_scale)]
+                  ? AM_SEGMENT_ID_ACTIVE
+                  : AM_SEGMENT_ID_INACTIVE;
+        }
+      }
+      cpi->active_map.enabled = 1;
+    } else {
+      cpi->active_map.enabled = 0;
+    }
+    return 0;
+  } else {
+    return -1;
+  }
+}
+
+int av1_get_active_map(AV1_COMP *cpi, unsigned char *new_map_16x16, int rows,
+                       int cols) {
+  if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols &&
+      new_map_16x16) {
+    unsigned char *const seg_map_8x8 = cpi->segmentation_map;
+    const int mi_rows = cpi->common.mi_rows;
+    const int mi_cols = cpi->common.mi_cols;
+    const int row_scale = mi_size_high[BLOCK_16X16] == 2 ? 1 : 2;
+    const int col_scale = mi_size_wide[BLOCK_16X16] == 2 ? 1 : 2;
+
+    memset(new_map_16x16, !cpi->active_map.enabled, rows * cols);
+    if (cpi->active_map.enabled) {
+      int r, c;
+      for (r = 0; r < mi_rows; ++r) {
+        for (c = 0; c < mi_cols; ++c) {
+          // Cyclic refresh segments are considered active despite not having
+          // AM_SEGMENT_ID_ACTIVE
+          new_map_16x16[(r >> row_scale) * cols + (c >> col_scale)] |=
+              seg_map_8x8[r * mi_cols + c] != AM_SEGMENT_ID_INACTIVE;
+        }
+      }
+    }
+    return 0;
+  } else {
+    return -1;
+  }
+}
+
+void av1_set_high_precision_mv(AV1_COMP *cpi, int allow_high_precision_mv) {
+  MACROBLOCK *const mb = &cpi->td.mb;
+  cpi->common.allow_high_precision_mv = allow_high_precision_mv;
+
+#if CONFIG_REF_MV
+  if (cpi->common.allow_high_precision_mv) {
+    int i;
+    for (i = 0; i < NMV_CONTEXTS; ++i) {
+      mb->mv_cost_stack[i] = mb->nmvcost_hp[i];
+      mb->mvsadcost = mb->nmvsadcost_hp;
+    }
+  } else {
+    int i;
+    for (i = 0; i < NMV_CONTEXTS; ++i) {
+      mb->mv_cost_stack[i] = mb->nmvcost[i];
+      mb->mvsadcost = mb->nmvsadcost;
+    }
+  }
+#else
+  if (cpi->common.allow_high_precision_mv) {
+    mb->mvcost = mb->nmvcost_hp;
+    mb->mvsadcost = mb->nmvcost_hp;
+  } else {
+    mb->mvcost = mb->nmvcost;
+    mb->mvsadcost = mb->nmvcost;
+  }
+#endif
+}
+
+static BLOCK_SIZE select_sb_size(const AV1_COMP *const cpi) {
+#if CONFIG_EXT_PARTITION
+  if (cpi->oxcf.superblock_size == AOM_SUPERBLOCK_SIZE_64X64)
+    return BLOCK_64X64;
+
+  if (cpi->oxcf.superblock_size == AOM_SUPERBLOCK_SIZE_128X128)
+    return BLOCK_128X128;
+
+  assert(cpi->oxcf.superblock_size == AOM_SUPERBLOCK_SIZE_DYNAMIC);
+
+  assert(IMPLIES(cpi->common.tile_cols > 1,
+                 cpi->common.tile_width % MAX_MIB_SIZE == 0));
+  assert(IMPLIES(cpi->common.tile_rows > 1,
+                 cpi->common.tile_height % MAX_MIB_SIZE == 0));
+
+  // TODO(any): Possibly could improve this with a heuristic.
+  return BLOCK_128X128;
+#else
+  (void)cpi;
+  return BLOCK_64X64;
+#endif  //  CONFIG_EXT_PARTITION
+}
+
+static void setup_frame(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  // Set up entropy context depending on frame type. The decoder mandates
+  // the use of the default context, index 0, for keyframes and inter
+  // frames where the error_resilient_mode or intra_only flag is set. For
+  // other inter-frames the encoder currently uses only two contexts;
+  // context 1 for ALTREF frames and context 0 for the others.
+  if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
+    av1_setup_past_independence(cm);
+  } else {
+#if CONFIG_EXT_REFS
+    const GF_GROUP *gf_group = &cpi->twopass.gf_group;
+    if (gf_group->rf_level[gf_group->index] == GF_ARF_LOW)
+      cm->frame_context_idx = EXT_ARF_FRAME;
+    else if (cpi->refresh_alt_ref_frame)
+      cm->frame_context_idx = ARF_FRAME;
+#else
+    if (cpi->refresh_alt_ref_frame) cm->frame_context_idx = ARF_FRAME;
+#endif  // CONFIG_EXT_REFS
+    else if (cpi->rc.is_src_frame_alt_ref)
+      cm->frame_context_idx = OVERLAY_FRAME;
+    else if (cpi->refresh_golden_frame)
+      cm->frame_context_idx = GLD_FRAME;
+#if CONFIG_EXT_REFS
+    else if (cpi->refresh_bwd_ref_frame)
+      cm->frame_context_idx = BRF_FRAME;
+#endif  // CONFIG_EXT_REFS
+    else
+      cm->frame_context_idx = REGULAR_FRAME;
+  }
+
+  if (cm->frame_type == KEY_FRAME) {
+    cpi->refresh_golden_frame = 1;
+    cpi->refresh_alt_ref_frame = 1;
+    av1_zero(cpi->interp_filter_selected);
+  } else {
+    *cm->fc = cm->frame_contexts[cm->frame_context_idx];
+    av1_zero(cpi->interp_filter_selected[0]);
+  }
+#if CONFIG_EXT_REFS
+#if CONFIG_LOWDELAY_COMPOUND  // No change to bitstream
+  if (cpi->sf.recode_loop == DISALLOW_RECODE) {
+    cpi->refresh_bwd_ref_frame = cpi->refresh_last_frame;
+    cpi->rc.is_bipred_frame = 1;
+  }
+#endif
+#endif
+
+  cpi->vaq_refresh = 0;
+
+  set_sb_size(cm, select_sb_size(cpi));
+}
+
+static void av1_enc_setup_mi(AV1_COMMON *cm) {
+  int i;
+  cm->mi = cm->mip + cm->mi_stride + 1;
+  memset(cm->mip, 0, cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mip));
+  cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
+  // Clear top border row
+  memset(cm->prev_mip, 0, sizeof(*cm->prev_mip) * cm->mi_stride);
+  // Clear left border column
+  for (i = 1; i < cm->mi_rows + 1; ++i)
+    memset(&cm->prev_mip[i * cm->mi_stride], 0, sizeof(*cm->prev_mip));
+
+  cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
+  cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
+
+  memset(cm->mi_grid_base, 0,
+         cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
+}
+
+static int av1_enc_alloc_mi(AV1_COMMON *cm, int mi_size) {
+  cm->mip = aom_calloc(mi_size, sizeof(*cm->mip));
+  if (!cm->mip) return 1;
+  cm->prev_mip = aom_calloc(mi_size, sizeof(*cm->prev_mip));
+  if (!cm->prev_mip) return 1;
+  cm->mi_alloc_size = mi_size;
+
+  cm->mi_grid_base = (MODE_INFO **)aom_calloc(mi_size, sizeof(MODE_INFO *));
+  if (!cm->mi_grid_base) return 1;
+  cm->prev_mi_grid_base =
+      (MODE_INFO **)aom_calloc(mi_size, sizeof(MODE_INFO *));
+  if (!cm->prev_mi_grid_base) return 1;
+
+  return 0;
+}
+
+static void av1_enc_free_mi(AV1_COMMON *cm) {
+  aom_free(cm->mip);
+  cm->mip = NULL;
+  aom_free(cm->prev_mip);
+  cm->prev_mip = NULL;
+  aom_free(cm->mi_grid_base);
+  cm->mi_grid_base = NULL;
+  aom_free(cm->prev_mi_grid_base);
+  cm->prev_mi_grid_base = NULL;
+}
+
+static void av1_swap_mi_and_prev_mi(AV1_COMMON *cm) {
+  // Current mip will be the prev_mip for the next frame.
+  MODE_INFO **temp_base = cm->prev_mi_grid_base;
+  MODE_INFO *temp = cm->prev_mip;
+  cm->prev_mip = cm->mip;
+  cm->mip = temp;
+
+  // Update the upper left visible macroblock ptrs.
+  cm->mi = cm->mip + cm->mi_stride + 1;
+  cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
+
+  cm->prev_mi_grid_base = cm->mi_grid_base;
+  cm->mi_grid_base = temp_base;
+  cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
+  cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
+}
+
+void av1_initialize_enc(void) {
+  static volatile int init_done = 0;
+
+  if (!init_done) {
+    av1_rtcd();
+    aom_dsp_rtcd();
+    aom_scale_rtcd();
+    av1_init_intra_predictors();
+    av1_init_me_luts();
+#if !CONFIG_XIPHRC
+    av1_rc_init_minq_luts();
+#endif
+    av1_entropy_mv_init();
+    av1_encode_token_init();
+#if CONFIG_EXT_INTER
+    av1_init_wedge_masks();
+#endif
+    init_done = 1;
+  }
+}
+
+static void dealloc_compressor_data(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  int i;
+
+  aom_free(cpi->mbmi_ext_base);
+  cpi->mbmi_ext_base = NULL;
+
+#if CONFIG_PVQ
+  if (cpi->oxcf.pass != 1) {
+    const int tile_cols = cm->tile_cols;
+    const int tile_rows = cm->tile_rows;
+    int tile_col, tile_row;
+
+    for (tile_row = 0; tile_row < tile_rows; ++tile_row)
+      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
+        TileDataEnc *tile_data =
+            &cpi->tile_data[tile_row * tile_cols + tile_col];
+        aom_free(tile_data->pvq_q.buf);
+      }
+  }
+#endif
+  aom_free(cpi->tile_data);
+  cpi->tile_data = NULL;
+
+  // Delete sementation map
+  aom_free(cpi->segmentation_map);
+  cpi->segmentation_map = NULL;
+
+  av1_cyclic_refresh_free(cpi->cyclic_refresh);
+  cpi->cyclic_refresh = NULL;
+
+  aom_free(cpi->active_map.map);
+  cpi->active_map.map = NULL;
+
+  // Free up-sampled reference buffers.
+  for (i = 0; i < (REF_FRAMES + 1); i++)
+    aom_free_frame_buffer(&cpi->upsampled_ref_bufs[i].buf);
+
+  av1_free_ref_frame_buffers(cm->buffer_pool);
+#if CONFIG_LV_MAP
+  av1_free_txb_buf(cpi);
+#endif
+  av1_free_context_buffers(cm);
+
+  aom_free_frame_buffer(&cpi->last_frame_uf);
+#if CONFIG_LOOP_RESTORATION
+  av1_free_restoration_buffers(cm);
+  aom_free_frame_buffer(&cpi->last_frame_db);
+  aom_free_frame_buffer(&cpi->trial_frame_rst);
+  aom_free(cpi->extra_rstbuf);
+  for (i = 0; i < MAX_MB_PLANE; ++i)
+    av1_free_restoration_struct(&cpi->rst_search[i]);
+#endif  // CONFIG_LOOP_RESTORATION
+  aom_free_frame_buffer(&cpi->scaled_source);
+  aom_free_frame_buffer(&cpi->scaled_last_source);
+  aom_free_frame_buffer(&cpi->alt_ref_buffer);
+  av1_lookahead_destroy(cpi->lookahead);
+
+  aom_free(cpi->tile_tok[0][0]);
+  cpi->tile_tok[0][0] = 0;
+
+  av1_free_pc_tree(&cpi->td);
+  av1_free_var_tree(&cpi->td);
+
+#if CONFIG_PALETTE
+  if (cpi->common.allow_screen_content_tools)
+    aom_free(cpi->td.mb.palette_buffer);
+#endif  // CONFIG_PALETTE
+
+  if (cpi->source_diff_var != NULL) {
+    aom_free(cpi->source_diff_var);
+    cpi->source_diff_var = NULL;
+  }
+#if CONFIG_ANS
+  aom_buf_ans_free(&cpi->buf_ans);
+#endif  // CONFIG_ANS
+}
+
+static void save_coding_context(AV1_COMP *cpi) {
+  CODING_CONTEXT *const cc = &cpi->coding_context;
+  AV1_COMMON *cm = &cpi->common;
+#if CONFIG_REF_MV
+  int i;
+#endif
+
+// Stores a snapshot of key state variables which can subsequently be
+// restored with a call to av1_restore_coding_context. These functions are
+// intended for use in a re-code loop in av1_compress_frame where the
+// quantizer value is adjusted between loop iterations.
+#if CONFIG_REF_MV
+  for (i = 0; i < NMV_CONTEXTS; ++i) {
+    av1_copy(cc->nmv_vec_cost[i], cpi->td.mb.nmv_vec_cost[i]);
+    av1_copy(cc->nmv_costs, cpi->nmv_costs);
+    av1_copy(cc->nmv_costs_hp, cpi->nmv_costs_hp);
+  }
+#else
+  av1_copy(cc->nmvjointcost, cpi->td.mb.nmvjointcost);
+#endif
+
+  av1_copy(cc->nmvcosts, cpi->nmvcosts);
+  av1_copy(cc->nmvcosts_hp, cpi->nmvcosts_hp);
+
+  av1_copy(cc->last_ref_lf_deltas, cm->lf.last_ref_deltas);
+  av1_copy(cc->last_mode_lf_deltas, cm->lf.last_mode_deltas);
+
+  cc->fc = *cm->fc;
+}
+
+static void restore_coding_context(AV1_COMP *cpi) {
+  CODING_CONTEXT *const cc = &cpi->coding_context;
+  AV1_COMMON *cm = &cpi->common;
+#if CONFIG_REF_MV
+  int i;
+#endif
+
+// Restore key state variables to the snapshot state stored in the
+// previous call to av1_save_coding_context.
+#if CONFIG_REF_MV
+  for (i = 0; i < NMV_CONTEXTS; ++i) {
+    av1_copy(cpi->td.mb.nmv_vec_cost[i], cc->nmv_vec_cost[i]);
+    av1_copy(cpi->nmv_costs, cc->nmv_costs);
+    av1_copy(cpi->nmv_costs_hp, cc->nmv_costs_hp);
+  }
+#else
+  av1_copy(cpi->td.mb.nmvjointcost, cc->nmvjointcost);
+#endif
+
+  av1_copy(cpi->nmvcosts, cc->nmvcosts);
+  av1_copy(cpi->nmvcosts_hp, cc->nmvcosts_hp);
+
+  av1_copy(cm->lf.last_ref_deltas, cc->last_ref_lf_deltas);
+  av1_copy(cm->lf.last_mode_deltas, cc->last_mode_lf_deltas);
+
+  *cm->fc = cc->fc;
+}
+
+static void configure_static_seg_features(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  struct segmentation *const seg = &cm->seg;
+
+  int high_q = (int)(rc->avg_q > 48.0);
+  int qi_delta;
+
+  // Disable and clear down for KF
+  if (cm->frame_type == KEY_FRAME) {
+    // Clear down the global segmentation map
+    memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
+    seg->update_map = 0;
+    seg->update_data = 0;
+    cpi->static_mb_pct = 0;
+
+    // Disable segmentation
+    av1_disable_segmentation(seg);
+
+    // Clear down the segment features.
+    av1_clearall_segfeatures(seg);
+  } else if (cpi->refresh_alt_ref_frame) {
+    // If this is an alt ref frame
+    // Clear down the global segmentation map
+    memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
+    seg->update_map = 0;
+    seg->update_data = 0;
+    cpi->static_mb_pct = 0;
+
+    // Disable segmentation and individual segment features by default
+    av1_disable_segmentation(seg);
+    av1_clearall_segfeatures(seg);
+
+    // Scan frames from current to arf frame.
+    // This function re-enables segmentation if appropriate.
+    av1_update_mbgraph_stats(cpi);
+
+    // If segmentation was enabled set those features needed for the
+    // arf itself.
+    if (seg->enabled) {
+      seg->update_map = 1;
+      seg->update_data = 1;
+
+      qi_delta =
+          av1_compute_qdelta(rc, rc->avg_q, rc->avg_q * 0.875, cm->bit_depth);
+      av1_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta - 2);
+      av1_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
+
+      av1_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
+      av1_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
+
+      // Where relevant assume segment data is delta data
+      seg->abs_delta = SEGMENT_DELTADATA;
+    }
+  } else if (seg->enabled) {
+    // All other frames if segmentation has been enabled
+
+    // First normal frame in a valid gf or alt ref group
+    if (rc->frames_since_golden == 0) {
+      // Set up segment features for normal frames in an arf group
+      if (rc->source_alt_ref_active) {
+        seg->update_map = 0;
+        seg->update_data = 1;
+        seg->abs_delta = SEGMENT_DELTADATA;
+
+        qi_delta =
+            av1_compute_qdelta(rc, rc->avg_q, rc->avg_q * 1.125, cm->bit_depth);
+        av1_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta + 2);
+        av1_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
+
+        av1_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
+        av1_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
+
+        // Segment coding disabled for compred testing
+        if (high_q || (cpi->static_mb_pct == 100)) {
+          av1_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
+          av1_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
+          av1_enable_segfeature(seg, 1, SEG_LVL_SKIP);
+        }
+      } else {
+        // Disable segmentation and clear down features if alt ref
+        // is not active for this group
+
+        av1_disable_segmentation(seg);
+
+        memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
+
+        seg->update_map = 0;
+        seg->update_data = 0;
+
+        av1_clearall_segfeatures(seg);
+      }
+    } else if (rc->is_src_frame_alt_ref) {
+      // Special case where we are coding over the top of a previous
+      // alt ref frame.
+      // Segment coding disabled for compred testing
+
+      // Enable ref frame features for segment 0 as well
+      av1_enable_segfeature(seg, 0, SEG_LVL_REF_FRAME);
+      av1_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
+
+      // All mbs should use ALTREF_FRAME
+      av1_clear_segdata(seg, 0, SEG_LVL_REF_FRAME);
+      av1_set_segdata(seg, 0, SEG_LVL_REF_FRAME, ALTREF_FRAME);
+      av1_clear_segdata(seg, 1, SEG_LVL_REF_FRAME);
+      av1_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
+
+      // Skip all MBs if high Q (0,0 mv and skip coeffs)
+      if (high_q) {
+        av1_enable_segfeature(seg, 0, SEG_LVL_SKIP);
+        av1_enable_segfeature(seg, 1, SEG_LVL_SKIP);
+      }
+      // Enable data update
+      seg->update_data = 1;
+    } else {
+      // All other frames.
+
+      // No updates.. leave things as they are.
+      seg->update_map = 0;
+      seg->update_data = 0;
+    }
+  }
+}
+
+static void update_reference_segmentation_map(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
+  uint8_t *cache_ptr = cm->last_frame_seg_map;
+  int row, col;
+
+  for (row = 0; row < cm->mi_rows; row++) {
+    MODE_INFO **mi_8x8 = mi_8x8_ptr;
+    uint8_t *cache = cache_ptr;
+    for (col = 0; col < cm->mi_cols; col++, mi_8x8++, cache++)
+      cache[0] = mi_8x8[0]->mbmi.segment_id;
+    mi_8x8_ptr += cm->mi_stride;
+    cache_ptr += cm->mi_cols;
+  }
+}
+
+static void alloc_raw_frame_buffers(AV1_COMP *cpi) {
+  AV1_COMMON *cm = &cpi->common;
+  const AV1EncoderConfig *oxcf = &cpi->oxcf;
+
+  if (!cpi->lookahead)
+    cpi->lookahead = av1_lookahead_init(oxcf->width, oxcf->height,
+                                        cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                        cm->use_highbitdepth,
+#endif
+                                        oxcf->lag_in_frames);
+  if (!cpi->lookahead)
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate lag buffers");
+
+  // TODO(agrange) Check if ARF is enabled and skip allocation if not.
+  if (aom_realloc_frame_buffer(&cpi->alt_ref_buffer, oxcf->width, oxcf->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               AOM_BORDER_IN_PIXELS, cm->byte_alignment, NULL,
+                               NULL, NULL))
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate altref buffer");
+}
+
+static void alloc_util_frame_buffers(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  if (aom_realloc_frame_buffer(&cpi->last_frame_uf, cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               AOM_BORDER_IN_PIXELS, cm->byte_alignment, NULL,
+                               NULL, NULL))
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate last frame buffer");
+
+#if CONFIG_LOOP_RESTORATION
+  if (aom_realloc_frame_buffer(&cpi->last_frame_db, cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               AOM_BORDER_IN_PIXELS, cm->byte_alignment, NULL,
+                               NULL, NULL))
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate last frame deblocked buffer");
+  if (aom_realloc_frame_buffer(&cpi->trial_frame_rst, cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               AOM_BORDER_IN_PIXELS, cm->byte_alignment, NULL,
+                               NULL, NULL))
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate trial restored frame buffer");
+  int extra_rstbuf_sz = RESTORATION_EXTBUF_SIZE;
+  if (extra_rstbuf_sz > 0) {
+    aom_free(cpi->extra_rstbuf);
+    CHECK_MEM_ERROR(cm, cpi->extra_rstbuf,
+                    (uint8_t *)aom_malloc(extra_rstbuf_sz));
+  } else {
+    cpi->extra_rstbuf = NULL;
+  }
+#endif  // CONFIG_LOOP_RESTORATION
+
+  if (aom_realloc_frame_buffer(&cpi->scaled_source, cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               AOM_BORDER_IN_PIXELS, cm->byte_alignment, NULL,
+                               NULL, NULL))
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate scaled source buffer");
+
+  if (aom_realloc_frame_buffer(&cpi->scaled_last_source, cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               AOM_BORDER_IN_PIXELS, cm->byte_alignment, NULL,
+                               NULL, NULL))
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate scaled last source buffer");
+}
+
+static int alloc_context_buffers_ext(AV1_COMP *cpi) {
+  AV1_COMMON *cm = &cpi->common;
+  int mi_size = cm->mi_cols * cm->mi_rows;
+
+  cpi->mbmi_ext_base = aom_calloc(mi_size, sizeof(*cpi->mbmi_ext_base));
+  if (!cpi->mbmi_ext_base) return 1;
+
+  return 0;
+}
+
+void av1_alloc_compressor_data(AV1_COMP *cpi) {
+  AV1_COMMON *cm = &cpi->common;
+
+  av1_alloc_context_buffers(cm, cm->width, cm->height);
+
+#if CONFIG_LV_MAP
+  av1_alloc_txb_buf(cpi);
+#endif
+
+  alloc_context_buffers_ext(cpi);
+
+  aom_free(cpi->tile_tok[0][0]);
+
+  {
+    unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols);
+    CHECK_MEM_ERROR(cm, cpi->tile_tok[0][0],
+                    aom_calloc(tokens, sizeof(*cpi->tile_tok[0][0])));
+#if CONFIG_ANS && !ANS_MAX_SYMBOLS
+    aom_buf_ans_alloc(&cpi->buf_ans, &cm->error, (int)tokens);
+#endif  // CONFIG_ANS
+  }
+
+  av1_setup_pc_tree(&cpi->common, &cpi->td);
+}
+
+void av1_new_framerate(AV1_COMP *cpi, double framerate) {
+  cpi->framerate = framerate < 0.1 ? 30 : framerate;
+#if CONFIG_XIPHRC
+  if (!cpi->od_rc.cur_frame) return;
+  cpi->od_rc.framerate = cpi->framerate;
+  od_enc_rc_resize(&cpi->od_rc);
+#else
+  av1_rc_update_framerate(cpi);
+#endif
+}
+
+static void set_tile_info(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+#if CONFIG_TILE_GROUPS && CONFIG_DEPENDENT_HORZTILES
+  int tile_row, tile_col, num_tiles_in_tg;
+  int tg_row_start, tg_col_start;
+#endif
+#if CONFIG_EXT_TILE
+#if CONFIG_EXT_PARTITION
+  if (cpi->oxcf.superblock_size != AOM_SUPERBLOCK_SIZE_64X64) {
+    cm->tile_width = clamp(cpi->oxcf.tile_columns, 1, 32);
+    cm->tile_height = clamp(cpi->oxcf.tile_rows, 1, 32);
+    cm->tile_width <<= MAX_MIB_SIZE_LOG2;
+    cm->tile_height <<= MAX_MIB_SIZE_LOG2;
+  } else {
+    cm->tile_width = clamp(cpi->oxcf.tile_columns, 1, 64);
+    cm->tile_height = clamp(cpi->oxcf.tile_rows, 1, 64);
+    cm->tile_width <<= MAX_MIB_SIZE_LOG2 - 1;
+    cm->tile_height <<= MAX_MIB_SIZE_LOG2 - 1;
+  }
+#else
+  cm->tile_width = clamp(cpi->oxcf.tile_columns, 1, 64);
+  cm->tile_height = clamp(cpi->oxcf.tile_rows, 1, 64);
+  cm->tile_width <<= MAX_MIB_SIZE_LOG2;
+  cm->tile_height <<= MAX_MIB_SIZE_LOG2;
+#endif  // CONFIG_EXT_PARTITION
+
+  cm->tile_width = AOMMIN(cm->tile_width, cm->mi_cols);
+  cm->tile_height = AOMMIN(cm->tile_height, cm->mi_rows);
+
+  assert(cm->tile_width >> MAX_MIB_SIZE <= 32);
+  assert(cm->tile_height >> MAX_MIB_SIZE <= 32);
+
+  // Get the number of tiles
+  cm->tile_cols = 1;
+  while (cm->tile_cols * cm->tile_width < cm->mi_cols) ++cm->tile_cols;
+
+  cm->tile_rows = 1;
+  while (cm->tile_rows * cm->tile_height < cm->mi_rows) ++cm->tile_rows;
+#else
+  int min_log2_tile_cols, max_log2_tile_cols;
+  av1_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
+
+  cm->log2_tile_cols =
+      clamp(cpi->oxcf.tile_columns, min_log2_tile_cols, max_log2_tile_cols);
+  cm->log2_tile_rows = cpi->oxcf.tile_rows;
+
+  cm->tile_cols = 1 << cm->log2_tile_cols;
+  cm->tile_rows = 1 << cm->log2_tile_rows;
+
+  cm->tile_width = ALIGN_POWER_OF_TWO(cm->mi_cols, MAX_MIB_SIZE_LOG2);
+  cm->tile_width >>= cm->log2_tile_cols;
+  cm->tile_height = ALIGN_POWER_OF_TWO(cm->mi_rows, MAX_MIB_SIZE_LOG2);
+  cm->tile_height >>= cm->log2_tile_rows;
+
+  // round to integer multiples of max superblock size
+  cm->tile_width = ALIGN_POWER_OF_TWO(cm->tile_width, MAX_MIB_SIZE_LOG2);
+  cm->tile_height = ALIGN_POWER_OF_TWO(cm->tile_height, MAX_MIB_SIZE_LOG2);
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_DEPENDENT_HORZTILES
+  cm->dependent_horz_tiles = cpi->oxcf.dependent_horz_tiles;
+  if (cm->log2_tile_rows == 0) cm->dependent_horz_tiles = 0;
+#if CONFIG_TILE_GROUPS
+  if (cpi->oxcf.mtu == 0) {
+    cm->num_tg = cpi->oxcf.num_tile_groups;
+  } else {
+    // Use a default value for the purposes of weighting costs in probability
+    // updates
+    cm->num_tg = DEFAULT_MAX_NUM_TG;
+  }
+  num_tiles_in_tg =
+      (cm->tile_cols * cm->tile_rows + cm->num_tg - 1) / cm->num_tg;
+  tg_row_start = 0;
+  tg_col_start = 0;
+  for (tile_row = 0; tile_row < cm->tile_rows; ++tile_row) {
+    for (tile_col = 0; tile_col < cm->tile_cols; ++tile_col) {
+      if ((tile_row * cm->tile_cols + tile_col) % num_tiles_in_tg == 0) {
+        tg_row_start = tile_row;
+        tg_col_start = tile_col;
+      }
+      cm->tile_group_start_row[tile_row][tile_col] = tg_row_start;
+      cm->tile_group_start_col[tile_row][tile_col] = tg_col_start;
+    }
+  }
+#endif
+#endif
+
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  cm->loop_filter_across_tiles_enabled =
+      cpi->oxcf.loop_filter_across_tiles_enabled;
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+}
+
+static void update_frame_size(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+
+  av1_set_mb_mi(cm, cm->width, cm->height);
+  av1_init_context_buffers(cm);
+  av1_init_macroblockd(cm, xd,
+#if CONFIG_PVQ
+                       NULL,
+#endif
+#if CONFIG_CFL
+                       &NULL_CFL,
+#endif
+                       NULL);
+  memset(cpi->mbmi_ext_base, 0,
+         cm->mi_rows * cm->mi_cols * sizeof(*cpi->mbmi_ext_base));
+
+  set_tile_info(cpi);
+}
+
+static void init_buffer_indices(AV1_COMP *cpi) {
+#if CONFIG_EXT_REFS
+  int fb_idx;
+  for (fb_idx = 0; fb_idx < LAST_REF_FRAMES; ++fb_idx)
+    cpi->lst_fb_idxes[fb_idx] = fb_idx;
+  cpi->gld_fb_idx = LAST_REF_FRAMES;
+  cpi->bwd_fb_idx = LAST_REF_FRAMES + 1;
+  cpi->alt_fb_idx = LAST_REF_FRAMES + 2;
+  for (fb_idx = 0; fb_idx < MAX_EXT_ARFS + 1; ++fb_idx)
+    cpi->arf_map[fb_idx] = LAST_REF_FRAMES + 2 + fb_idx;
+#else
+  cpi->lst_fb_idx = 0;
+  cpi->gld_fb_idx = 1;
+  cpi->alt_fb_idx = 2;
+#endif  // CONFIG_EXT_REFS
+}
+
+static void init_config(struct AV1_COMP *cpi, AV1EncoderConfig *oxcf) {
+  AV1_COMMON *const cm = &cpi->common;
+
+  cpi->oxcf = *oxcf;
+  cpi->framerate = oxcf->init_framerate;
+
+  cm->profile = oxcf->profile;
+  cm->bit_depth = oxcf->bit_depth;
+#if CONFIG_HIGHBITDEPTH
+  cm->use_highbitdepth = oxcf->use_highbitdepth;
+#endif
+  cm->color_space = oxcf->color_space;
+  cm->color_range = oxcf->color_range;
+
+  cm->width = oxcf->width;
+  cm->height = oxcf->height;
+  av1_alloc_compressor_data(cpi);
+
+  // Single thread case: use counts in common.
+  cpi->td.counts = &cm->counts;
+
+  // change includes all joint functionality
+  av1_change_config(cpi, oxcf);
+
+  cpi->static_mb_pct = 0;
+  cpi->ref_frame_flags = 0;
+
+  init_buffer_indices(cpi);
+}
+
+static void set_rc_buffer_sizes(RATE_CONTROL *rc,
+                                const AV1EncoderConfig *oxcf) {
+  const int64_t bandwidth = oxcf->target_bandwidth;
+  const int64_t starting = oxcf->starting_buffer_level_ms;
+  const int64_t optimal = oxcf->optimal_buffer_level_ms;
+  const int64_t maximum = oxcf->maximum_buffer_size_ms;
+
+  rc->starting_buffer_level = starting * bandwidth / 1000;
+  rc->optimal_buffer_level =
+      (optimal == 0) ? bandwidth / 8 : optimal * bandwidth / 1000;
+  rc->maximum_buffer_size =
+      (maximum == 0) ? bandwidth / 8 : maximum * bandwidth / 1000;
+}
+
+#if CONFIG_HIGHBITDEPTH
+#define HIGHBD_BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF) \
+  cpi->fn_ptr[BT].sdf = SDF;                                           \
+  cpi->fn_ptr[BT].sdaf = SDAF;                                         \
+  cpi->fn_ptr[BT].vf = VF;                                             \
+  cpi->fn_ptr[BT].svf = SVF;                                           \
+  cpi->fn_ptr[BT].svaf = SVAF;                                         \
+  cpi->fn_ptr[BT].sdx3f = SDX3F;                                       \
+  cpi->fn_ptr[BT].sdx8f = SDX8F;                                       \
+  cpi->fn_ptr[BT].sdx4df = SDX4DF;
+
+#define MAKE_BFP_SAD_WRAPPER(fnname)                                           \
+  static unsigned int fnname##_bits8(const uint8_t *src_ptr,                   \
+                                     int source_stride,                        \
+                                     const uint8_t *ref_ptr, int ref_stride) { \
+    return fnname(src_ptr, source_stride, ref_ptr, ref_stride);                \
+  }                                                                            \
+  static unsigned int fnname##_bits10(                                         \
+      const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
+      int ref_stride) {                                                        \
+    return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 2;           \
+  }                                                                            \
+  static unsigned int fnname##_bits12(                                         \
+      const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
+      int ref_stride) {                                                        \
+    return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 4;           \
+  }
+
+#define MAKE_BFP_SADAVG_WRAPPER(fnname)                                        \
+  static unsigned int fnname##_bits8(                                          \
+      const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
+      int ref_stride, const uint8_t *second_pred) {                            \
+    return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred);   \
+  }                                                                            \
+  static unsigned int fnname##_bits10(                                         \
+      const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
+      int ref_stride, const uint8_t *second_pred) {                            \
+    return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred) >> \
+           2;                                                                  \
+  }                                                                            \
+  static unsigned int fnname##_bits12(                                         \
+      const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
+      int ref_stride, const uint8_t *second_pred) {                            \
+    return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred) >> \
+           4;                                                                  \
+  }
+
+#define MAKE_BFP_SAD3_WRAPPER(fnname)                                    \
+  static void fnname##_bits8(const uint8_t *src_ptr, int source_stride,  \
+                             const uint8_t *ref_ptr, int ref_stride,     \
+                             unsigned int *sad_array) {                  \
+    fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
+  }                                                                      \
+  static void fnname##_bits10(const uint8_t *src_ptr, int source_stride, \
+                              const uint8_t *ref_ptr, int ref_stride,    \
+                              unsigned int *sad_array) {                 \
+    int i;                                                               \
+    fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
+    for (i = 0; i < 3; i++) sad_array[i] >>= 2;                          \
+  }                                                                      \
+  static void fnname##_bits12(const uint8_t *src_ptr, int source_stride, \
+                              const uint8_t *ref_ptr, int ref_stride,    \
+                              unsigned int *sad_array) {                 \
+    int i;                                                               \
+    fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
+    for (i = 0; i < 3; i++) sad_array[i] >>= 4;                          \
+  }
+
+#define MAKE_BFP_SAD8_WRAPPER(fnname)                                    \
+  static void fnname##_bits8(const uint8_t *src_ptr, int source_stride,  \
+                             const uint8_t *ref_ptr, int ref_stride,     \
+                             unsigned int *sad_array) {                  \
+    fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
+  }                                                                      \
+  static void fnname##_bits10(const uint8_t *src_ptr, int source_stride, \
+                              const uint8_t *ref_ptr, int ref_stride,    \
+                              unsigned int *sad_array) {                 \
+    int i;                                                               \
+    fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
+    for (i = 0; i < 8; i++) sad_array[i] >>= 2;                          \
+  }                                                                      \
+  static void fnname##_bits12(const uint8_t *src_ptr, int source_stride, \
+                              const uint8_t *ref_ptr, int ref_stride,    \
+                              unsigned int *sad_array) {                 \
+    int i;                                                               \
+    fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
+    for (i = 0; i < 8; i++) sad_array[i] >>= 4;                          \
+  }
+#define MAKE_BFP_SAD4D_WRAPPER(fnname)                                        \
+  static void fnname##_bits8(const uint8_t *src_ptr, int source_stride,       \
+                             const uint8_t *const ref_ptr[], int ref_stride,  \
+                             unsigned int *sad_array) {                       \
+    fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);           \
+  }                                                                           \
+  static void fnname##_bits10(const uint8_t *src_ptr, int source_stride,      \
+                              const uint8_t *const ref_ptr[], int ref_stride, \
+                              unsigned int *sad_array) {                      \
+    int i;                                                                    \
+    fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);           \
+    for (i = 0; i < 4; i++) sad_array[i] >>= 2;                               \
+  }                                                                           \
+  static void fnname##_bits12(const uint8_t *src_ptr, int source_stride,      \
+                              const uint8_t *const ref_ptr[], int ref_stride, \
+                              unsigned int *sad_array) {                      \
+    int i;                                                                    \
+    fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);           \
+    for (i = 0; i < 4; i++) sad_array[i] >>= 4;                               \
+  }
+
+#if CONFIG_EXT_PARTITION
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad128x128)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad128x128_avg)
+MAKE_BFP_SAD3_WRAPPER(aom_highbd_sad128x128x3)
+MAKE_BFP_SAD8_WRAPPER(aom_highbd_sad128x128x8)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad128x128x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad128x64)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad128x64_avg)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad128x64x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad64x128)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad64x128_avg)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad64x128x4d)
+#endif  // CONFIG_EXT_PARTITION
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad32x16)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad32x16_avg)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad32x16x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad16x32)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad16x32_avg)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad16x32x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad64x32)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad64x32_avg)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad64x32x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad32x64)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad32x64_avg)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad32x64x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad32x32)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad32x32_avg)
+MAKE_BFP_SAD3_WRAPPER(aom_highbd_sad32x32x3)
+MAKE_BFP_SAD8_WRAPPER(aom_highbd_sad32x32x8)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad32x32x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad64x64)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad64x64_avg)
+MAKE_BFP_SAD3_WRAPPER(aom_highbd_sad64x64x3)
+MAKE_BFP_SAD8_WRAPPER(aom_highbd_sad64x64x8)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad64x64x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad16x16)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad16x16_avg)
+MAKE_BFP_SAD3_WRAPPER(aom_highbd_sad16x16x3)
+MAKE_BFP_SAD8_WRAPPER(aom_highbd_sad16x16x8)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad16x16x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad16x8)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad16x8_avg)
+MAKE_BFP_SAD3_WRAPPER(aom_highbd_sad16x8x3)
+MAKE_BFP_SAD8_WRAPPER(aom_highbd_sad16x8x8)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad16x8x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad8x16)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad8x16_avg)
+MAKE_BFP_SAD3_WRAPPER(aom_highbd_sad8x16x3)
+MAKE_BFP_SAD8_WRAPPER(aom_highbd_sad8x16x8)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad8x16x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad8x8)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad8x8_avg)
+MAKE_BFP_SAD3_WRAPPER(aom_highbd_sad8x8x3)
+MAKE_BFP_SAD8_WRAPPER(aom_highbd_sad8x8x8)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad8x8x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad8x4)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad8x4_avg)
+MAKE_BFP_SAD8_WRAPPER(aom_highbd_sad8x4x8)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad8x4x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad4x8)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad4x8_avg)
+MAKE_BFP_SAD8_WRAPPER(aom_highbd_sad4x8x8)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad4x8x4d)
+MAKE_BFP_SAD_WRAPPER(aom_highbd_sad4x4)
+MAKE_BFP_SADAVG_WRAPPER(aom_highbd_sad4x4_avg)
+MAKE_BFP_SAD3_WRAPPER(aom_highbd_sad4x4x3)
+MAKE_BFP_SAD8_WRAPPER(aom_highbd_sad4x4x8)
+MAKE_BFP_SAD4D_WRAPPER(aom_highbd_sad4x4x4d)
+
+#if CONFIG_EXT_INTER
+#define HIGHBD_MBFP(BT, MSDF, MVF, MSVF) \
+  cpi->fn_ptr[BT].msdf = MSDF;           \
+  cpi->fn_ptr[BT].mvf = MVF;             \
+  cpi->fn_ptr[BT].msvf = MSVF;
+
+#define MAKE_MBFP_SAD_WRAPPER(fnname)                                          \
+  static unsigned int fnname##_bits8(                                          \
+      const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
+      int ref_stride, const uint8_t *m, int m_stride) {                        \
+    return fnname(src_ptr, source_stride, ref_ptr, ref_stride, m, m_stride);   \
+  }                                                                            \
+  static unsigned int fnname##_bits10(                                         \
+      const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
+      int ref_stride, const uint8_t *m, int m_stride) {                        \
+    return fnname(src_ptr, source_stride, ref_ptr, ref_stride, m, m_stride) >> \
+           2;                                                                  \
+  }                                                                            \
+  static unsigned int fnname##_bits12(                                         \
+      const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
+      int ref_stride, const uint8_t *m, int m_stride) {                        \
+    return fnname(src_ptr, source_stride, ref_ptr, ref_stride, m, m_stride) >> \
+           4;                                                                  \
+  }
+
+#if CONFIG_EXT_PARTITION
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad128x128)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad128x64)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad64x128)
+#endif  // CONFIG_EXT_PARTITION
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad64x64)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad64x32)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad32x64)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad32x32)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad32x16)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad16x32)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad16x16)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad16x8)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad8x16)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad8x8)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad8x4)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad4x8)
+MAKE_MBFP_SAD_WRAPPER(aom_highbd_masked_sad4x4)
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR
+#define HIGHBD_OBFP(BT, OSDF, OVF, OSVF) \
+  cpi->fn_ptr[BT].osdf = OSDF;           \
+  cpi->fn_ptr[BT].ovf = OVF;             \
+  cpi->fn_ptr[BT].osvf = OSVF;
+
+#define MAKE_OBFP_SAD_WRAPPER(fnname)                                     \
+  static unsigned int fnname##_bits8(const uint8_t *ref, int ref_stride,  \
+                                     const int32_t *wsrc,                 \
+                                     const int32_t *msk) {                \
+    return fnname(ref, ref_stride, wsrc, msk);                            \
+  }                                                                       \
+  static unsigned int fnname##_bits10(const uint8_t *ref, int ref_stride, \
+                                      const int32_t *wsrc,                \
+                                      const int32_t *msk) {               \
+    return fnname(ref, ref_stride, wsrc, msk) >> 2;                       \
+  }                                                                       \
+  static unsigned int fnname##_bits12(const uint8_t *ref, int ref_stride, \
+                                      const int32_t *wsrc,                \
+                                      const int32_t *msk) {               \
+    return fnname(ref, ref_stride, wsrc, msk) >> 4;                       \
+  }
+
+#if CONFIG_EXT_PARTITION
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad128x128)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad128x64)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad64x128)
+#endif  // CONFIG_EXT_PARTITION
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad64x64)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad64x32)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad32x64)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad32x32)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad32x16)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad16x32)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad16x16)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad16x8)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad8x16)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad8x8)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad8x4)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad4x8)
+MAKE_OBFP_SAD_WRAPPER(aom_highbd_obmc_sad4x4)
+#endif  // CONFIG_MOTION_VAR
+
+static void highbd_set_var_fns(AV1_COMP *const cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  if (cm->use_highbitdepth) {
+    switch (cm->bit_depth) {
+      case AOM_BITS_8:
+        HIGHBD_BFP(BLOCK_32X16, aom_highbd_sad32x16_bits8,
+                   aom_highbd_sad32x16_avg_bits8, aom_highbd_8_variance32x16,
+                   aom_highbd_8_sub_pixel_variance32x16,
+                   aom_highbd_8_sub_pixel_avg_variance32x16, NULL, NULL,
+                   aom_highbd_sad32x16x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_16X32, aom_highbd_sad16x32_bits8,
+                   aom_highbd_sad16x32_avg_bits8, aom_highbd_8_variance16x32,
+                   aom_highbd_8_sub_pixel_variance16x32,
+                   aom_highbd_8_sub_pixel_avg_variance16x32, NULL, NULL,
+                   aom_highbd_sad16x32x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_64X32, aom_highbd_sad64x32_bits8,
+                   aom_highbd_sad64x32_avg_bits8, aom_highbd_8_variance64x32,
+                   aom_highbd_8_sub_pixel_variance64x32,
+                   aom_highbd_8_sub_pixel_avg_variance64x32, NULL, NULL,
+                   aom_highbd_sad64x32x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_32X64, aom_highbd_sad32x64_bits8,
+                   aom_highbd_sad32x64_avg_bits8, aom_highbd_8_variance32x64,
+                   aom_highbd_8_sub_pixel_variance32x64,
+                   aom_highbd_8_sub_pixel_avg_variance32x64, NULL, NULL,
+                   aom_highbd_sad32x64x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_32X32, aom_highbd_sad32x32_bits8,
+                   aom_highbd_sad32x32_avg_bits8, aom_highbd_8_variance32x32,
+                   aom_highbd_8_sub_pixel_variance32x32,
+                   aom_highbd_8_sub_pixel_avg_variance32x32,
+                   aom_highbd_sad32x32x3_bits8, aom_highbd_sad32x32x8_bits8,
+                   aom_highbd_sad32x32x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_64X64, aom_highbd_sad64x64_bits8,
+                   aom_highbd_sad64x64_avg_bits8, aom_highbd_8_variance64x64,
+                   aom_highbd_8_sub_pixel_variance64x64,
+                   aom_highbd_8_sub_pixel_avg_variance64x64,
+                   aom_highbd_sad64x64x3_bits8, aom_highbd_sad64x64x8_bits8,
+                   aom_highbd_sad64x64x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_16X16, aom_highbd_sad16x16_bits8,
+                   aom_highbd_sad16x16_avg_bits8, aom_highbd_8_variance16x16,
+                   aom_highbd_8_sub_pixel_variance16x16,
+                   aom_highbd_8_sub_pixel_avg_variance16x16,
+                   aom_highbd_sad16x16x3_bits8, aom_highbd_sad16x16x8_bits8,
+                   aom_highbd_sad16x16x4d_bits8)
+
+        HIGHBD_BFP(
+            BLOCK_16X8, aom_highbd_sad16x8_bits8, aom_highbd_sad16x8_avg_bits8,
+            aom_highbd_8_variance16x8, aom_highbd_8_sub_pixel_variance16x8,
+            aom_highbd_8_sub_pixel_avg_variance16x8, aom_highbd_sad16x8x3_bits8,
+            aom_highbd_sad16x8x8_bits8, aom_highbd_sad16x8x4d_bits8)
+
+        HIGHBD_BFP(
+            BLOCK_8X16, aom_highbd_sad8x16_bits8, aom_highbd_sad8x16_avg_bits8,
+            aom_highbd_8_variance8x16, aom_highbd_8_sub_pixel_variance8x16,
+            aom_highbd_8_sub_pixel_avg_variance8x16, aom_highbd_sad8x16x3_bits8,
+            aom_highbd_sad8x16x8_bits8, aom_highbd_sad8x16x4d_bits8)
+
+        HIGHBD_BFP(
+            BLOCK_8X8, aom_highbd_sad8x8_bits8, aom_highbd_sad8x8_avg_bits8,
+            aom_highbd_8_variance8x8, aom_highbd_8_sub_pixel_variance8x8,
+            aom_highbd_8_sub_pixel_avg_variance8x8, aom_highbd_sad8x8x3_bits8,
+            aom_highbd_sad8x8x8_bits8, aom_highbd_sad8x8x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_8X4, aom_highbd_sad8x4_bits8,
+                   aom_highbd_sad8x4_avg_bits8, aom_highbd_8_variance8x4,
+                   aom_highbd_8_sub_pixel_variance8x4,
+                   aom_highbd_8_sub_pixel_avg_variance8x4, NULL,
+                   aom_highbd_sad8x4x8_bits8, aom_highbd_sad8x4x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_4X8, aom_highbd_sad4x8_bits8,
+                   aom_highbd_sad4x8_avg_bits8, aom_highbd_8_variance4x8,
+                   aom_highbd_8_sub_pixel_variance4x8,
+                   aom_highbd_8_sub_pixel_avg_variance4x8, NULL,
+                   aom_highbd_sad4x8x8_bits8, aom_highbd_sad4x8x4d_bits8)
+
+        HIGHBD_BFP(
+            BLOCK_4X4, aom_highbd_sad4x4_bits8, aom_highbd_sad4x4_avg_bits8,
+            aom_highbd_8_variance4x4, aom_highbd_8_sub_pixel_variance4x4,
+            aom_highbd_8_sub_pixel_avg_variance4x4, aom_highbd_sad4x4x3_bits8,
+            aom_highbd_sad4x4x8_bits8, aom_highbd_sad4x4x4d_bits8)
+
+#if CONFIG_CB4X4
+        HIGHBD_BFP(BLOCK_2X2, NULL, NULL, aom_highbd_8_variance2x2, NULL, NULL,
+                   NULL, NULL, NULL)
+        HIGHBD_BFP(BLOCK_4X2, NULL, NULL, aom_highbd_8_variance4x2, NULL, NULL,
+                   NULL, NULL, NULL)
+        HIGHBD_BFP(BLOCK_2X4, NULL, NULL, aom_highbd_8_variance2x4, NULL, NULL,
+                   NULL, NULL, NULL)
+#endif
+
+#if CONFIG_EXT_PARTITION
+        HIGHBD_BFP(BLOCK_128X128, aom_highbd_sad128x128_bits8,
+                   aom_highbd_sad128x128_avg_bits8,
+                   aom_highbd_8_variance128x128,
+                   aom_highbd_8_sub_pixel_variance128x128,
+                   aom_highbd_8_sub_pixel_avg_variance128x128,
+                   aom_highbd_sad128x128x3_bits8, aom_highbd_sad128x128x8_bits8,
+                   aom_highbd_sad128x128x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_128X64, aom_highbd_sad128x64_bits8,
+                   aom_highbd_sad128x64_avg_bits8, aom_highbd_8_variance128x64,
+                   aom_highbd_8_sub_pixel_variance128x64,
+                   aom_highbd_8_sub_pixel_avg_variance128x64, NULL, NULL,
+                   aom_highbd_sad128x64x4d_bits8)
+
+        HIGHBD_BFP(BLOCK_64X128, aom_highbd_sad64x128_bits8,
+                   aom_highbd_sad64x128_avg_bits8, aom_highbd_8_variance64x128,
+                   aom_highbd_8_sub_pixel_variance64x128,
+                   aom_highbd_8_sub_pixel_avg_variance64x128, NULL, NULL,
+                   aom_highbd_sad64x128x4d_bits8)
+#endif  // CONFIG_EXT_PARTITION
+
+#if CONFIG_EXT_INTER
+#if CONFIG_EXT_PARTITION
+        HIGHBD_MBFP(BLOCK_128X128, aom_highbd_masked_sad128x128_bits8,
+                    aom_highbd_masked_variance128x128,
+                    aom_highbd_masked_sub_pixel_variance128x128)
+        HIGHBD_MBFP(BLOCK_128X64, aom_highbd_masked_sad128x64_bits8,
+                    aom_highbd_masked_variance128x64,
+                    aom_highbd_masked_sub_pixel_variance128x64)
+        HIGHBD_MBFP(BLOCK_64X128, aom_highbd_masked_sad64x128_bits8,
+                    aom_highbd_masked_variance64x128,
+                    aom_highbd_masked_sub_pixel_variance64x128)
+#endif  // CONFIG_EXT_PARTITION
+        HIGHBD_MBFP(BLOCK_64X64, aom_highbd_masked_sad64x64_bits8,
+                    aom_highbd_masked_variance64x64,
+                    aom_highbd_masked_sub_pixel_variance64x64)
+        HIGHBD_MBFP(BLOCK_64X32, aom_highbd_masked_sad64x32_bits8,
+                    aom_highbd_masked_variance64x32,
+                    aom_highbd_masked_sub_pixel_variance64x32)
+        HIGHBD_MBFP(BLOCK_32X64, aom_highbd_masked_sad32x64_bits8,
+                    aom_highbd_masked_variance32x64,
+                    aom_highbd_masked_sub_pixel_variance32x64)
+        HIGHBD_MBFP(BLOCK_32X32, aom_highbd_masked_sad32x32_bits8,
+                    aom_highbd_masked_variance32x32,
+                    aom_highbd_masked_sub_pixel_variance32x32)
+        HIGHBD_MBFP(BLOCK_32X16, aom_highbd_masked_sad32x16_bits8,
+                    aom_highbd_masked_variance32x16,
+                    aom_highbd_masked_sub_pixel_variance32x16)
+        HIGHBD_MBFP(BLOCK_16X32, aom_highbd_masked_sad16x32_bits8,
+                    aom_highbd_masked_variance16x32,
+                    aom_highbd_masked_sub_pixel_variance16x32)
+        HIGHBD_MBFP(BLOCK_16X16, aom_highbd_masked_sad16x16_bits8,
+                    aom_highbd_masked_variance16x16,
+                    aom_highbd_masked_sub_pixel_variance16x16)
+        HIGHBD_MBFP(BLOCK_8X16, aom_highbd_masked_sad8x16_bits8,
+                    aom_highbd_masked_variance8x16,
+                    aom_highbd_masked_sub_pixel_variance8x16)
+        HIGHBD_MBFP(BLOCK_16X8, aom_highbd_masked_sad16x8_bits8,
+                    aom_highbd_masked_variance16x8,
+                    aom_highbd_masked_sub_pixel_variance16x8)
+        HIGHBD_MBFP(BLOCK_8X8, aom_highbd_masked_sad8x8_bits8,
+                    aom_highbd_masked_variance8x8,
+                    aom_highbd_masked_sub_pixel_variance8x8)
+        HIGHBD_MBFP(BLOCK_4X8, aom_highbd_masked_sad4x8_bits8,
+                    aom_highbd_masked_variance4x8,
+                    aom_highbd_masked_sub_pixel_variance4x8)
+        HIGHBD_MBFP(BLOCK_8X4, aom_highbd_masked_sad8x4_bits8,
+                    aom_highbd_masked_variance8x4,
+                    aom_highbd_masked_sub_pixel_variance8x4)
+        HIGHBD_MBFP(BLOCK_4X4, aom_highbd_masked_sad4x4_bits8,
+                    aom_highbd_masked_variance4x4,
+                    aom_highbd_masked_sub_pixel_variance4x4)
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR
+#if CONFIG_EXT_PARTITION
+        HIGHBD_OBFP(BLOCK_128X128, aom_highbd_obmc_sad128x128_bits8,
+                    aom_highbd_obmc_variance128x128,
+                    aom_highbd_obmc_sub_pixel_variance128x128)
+        HIGHBD_OBFP(BLOCK_128X64, aom_highbd_obmc_sad128x64_bits8,
+                    aom_highbd_obmc_variance128x64,
+                    aom_highbd_obmc_sub_pixel_variance128x64)
+        HIGHBD_OBFP(BLOCK_64X128, aom_highbd_obmc_sad64x128_bits8,
+                    aom_highbd_obmc_variance64x128,
+                    aom_highbd_obmc_sub_pixel_variance64x128)
+#endif  // CONFIG_EXT_PARTITION
+        HIGHBD_OBFP(BLOCK_64X64, aom_highbd_obmc_sad64x64_bits8,
+                    aom_highbd_obmc_variance64x64,
+                    aom_highbd_obmc_sub_pixel_variance64x64)
+        HIGHBD_OBFP(BLOCK_64X32, aom_highbd_obmc_sad64x32_bits8,
+                    aom_highbd_obmc_variance64x32,
+                    aom_highbd_obmc_sub_pixel_variance64x32)
+        HIGHBD_OBFP(BLOCK_32X64, aom_highbd_obmc_sad32x64_bits8,
+                    aom_highbd_obmc_variance32x64,
+                    aom_highbd_obmc_sub_pixel_variance32x64)
+        HIGHBD_OBFP(BLOCK_32X32, aom_highbd_obmc_sad32x32_bits8,
+                    aom_highbd_obmc_variance32x32,
+                    aom_highbd_obmc_sub_pixel_variance32x32)
+        HIGHBD_OBFP(BLOCK_32X16, aom_highbd_obmc_sad32x16_bits8,
+                    aom_highbd_obmc_variance32x16,
+                    aom_highbd_obmc_sub_pixel_variance32x16)
+        HIGHBD_OBFP(BLOCK_16X32, aom_highbd_obmc_sad16x32_bits8,
+                    aom_highbd_obmc_variance16x32,
+                    aom_highbd_obmc_sub_pixel_variance16x32)
+        HIGHBD_OBFP(BLOCK_16X16, aom_highbd_obmc_sad16x16_bits8,
+                    aom_highbd_obmc_variance16x16,
+                    aom_highbd_obmc_sub_pixel_variance16x16)
+        HIGHBD_OBFP(BLOCK_8X16, aom_highbd_obmc_sad8x16_bits8,
+                    aom_highbd_obmc_variance8x16,
+                    aom_highbd_obmc_sub_pixel_variance8x16)
+        HIGHBD_OBFP(BLOCK_16X8, aom_highbd_obmc_sad16x8_bits8,
+                    aom_highbd_obmc_variance16x8,
+                    aom_highbd_obmc_sub_pixel_variance16x8)
+        HIGHBD_OBFP(BLOCK_8X8, aom_highbd_obmc_sad8x8_bits8,
+                    aom_highbd_obmc_variance8x8,
+                    aom_highbd_obmc_sub_pixel_variance8x8)
+        HIGHBD_OBFP(BLOCK_4X8, aom_highbd_obmc_sad4x8_bits8,
+                    aom_highbd_obmc_variance4x8,
+                    aom_highbd_obmc_sub_pixel_variance4x8)
+        HIGHBD_OBFP(BLOCK_8X4, aom_highbd_obmc_sad8x4_bits8,
+                    aom_highbd_obmc_variance8x4,
+                    aom_highbd_obmc_sub_pixel_variance8x4)
+        HIGHBD_OBFP(BLOCK_4X4, aom_highbd_obmc_sad4x4_bits8,
+                    aom_highbd_obmc_variance4x4,
+                    aom_highbd_obmc_sub_pixel_variance4x4)
+#endif  // CONFIG_MOTION_VAR
+        break;
+
+      case AOM_BITS_10:
+        HIGHBD_BFP(BLOCK_32X16, aom_highbd_sad32x16_bits10,
+                   aom_highbd_sad32x16_avg_bits10, aom_highbd_10_variance32x16,
+                   aom_highbd_10_sub_pixel_variance32x16,
+                   aom_highbd_10_sub_pixel_avg_variance32x16, NULL, NULL,
+                   aom_highbd_sad32x16x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_16X32, aom_highbd_sad16x32_bits10,
+                   aom_highbd_sad16x32_avg_bits10, aom_highbd_10_variance16x32,
+                   aom_highbd_10_sub_pixel_variance16x32,
+                   aom_highbd_10_sub_pixel_avg_variance16x32, NULL, NULL,
+                   aom_highbd_sad16x32x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_64X32, aom_highbd_sad64x32_bits10,
+                   aom_highbd_sad64x32_avg_bits10, aom_highbd_10_variance64x32,
+                   aom_highbd_10_sub_pixel_variance64x32,
+                   aom_highbd_10_sub_pixel_avg_variance64x32, NULL, NULL,
+                   aom_highbd_sad64x32x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_32X64, aom_highbd_sad32x64_bits10,
+                   aom_highbd_sad32x64_avg_bits10, aom_highbd_10_variance32x64,
+                   aom_highbd_10_sub_pixel_variance32x64,
+                   aom_highbd_10_sub_pixel_avg_variance32x64, NULL, NULL,
+                   aom_highbd_sad32x64x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_32X32, aom_highbd_sad32x32_bits10,
+                   aom_highbd_sad32x32_avg_bits10, aom_highbd_10_variance32x32,
+                   aom_highbd_10_sub_pixel_variance32x32,
+                   aom_highbd_10_sub_pixel_avg_variance32x32,
+                   aom_highbd_sad32x32x3_bits10, aom_highbd_sad32x32x8_bits10,
+                   aom_highbd_sad32x32x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_64X64, aom_highbd_sad64x64_bits10,
+                   aom_highbd_sad64x64_avg_bits10, aom_highbd_10_variance64x64,
+                   aom_highbd_10_sub_pixel_variance64x64,
+                   aom_highbd_10_sub_pixel_avg_variance64x64,
+                   aom_highbd_sad64x64x3_bits10, aom_highbd_sad64x64x8_bits10,
+                   aom_highbd_sad64x64x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_16X16, aom_highbd_sad16x16_bits10,
+                   aom_highbd_sad16x16_avg_bits10, aom_highbd_10_variance16x16,
+                   aom_highbd_10_sub_pixel_variance16x16,
+                   aom_highbd_10_sub_pixel_avg_variance16x16,
+                   aom_highbd_sad16x16x3_bits10, aom_highbd_sad16x16x8_bits10,
+                   aom_highbd_sad16x16x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_16X8, aom_highbd_sad16x8_bits10,
+                   aom_highbd_sad16x8_avg_bits10, aom_highbd_10_variance16x8,
+                   aom_highbd_10_sub_pixel_variance16x8,
+                   aom_highbd_10_sub_pixel_avg_variance16x8,
+                   aom_highbd_sad16x8x3_bits10, aom_highbd_sad16x8x8_bits10,
+                   aom_highbd_sad16x8x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_8X16, aom_highbd_sad8x16_bits10,
+                   aom_highbd_sad8x16_avg_bits10, aom_highbd_10_variance8x16,
+                   aom_highbd_10_sub_pixel_variance8x16,
+                   aom_highbd_10_sub_pixel_avg_variance8x16,
+                   aom_highbd_sad8x16x3_bits10, aom_highbd_sad8x16x8_bits10,
+                   aom_highbd_sad8x16x4d_bits10)
+
+        HIGHBD_BFP(
+            BLOCK_8X8, aom_highbd_sad8x8_bits10, aom_highbd_sad8x8_avg_bits10,
+            aom_highbd_10_variance8x8, aom_highbd_10_sub_pixel_variance8x8,
+            aom_highbd_10_sub_pixel_avg_variance8x8, aom_highbd_sad8x8x3_bits10,
+            aom_highbd_sad8x8x8_bits10, aom_highbd_sad8x8x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_8X4, aom_highbd_sad8x4_bits10,
+                   aom_highbd_sad8x4_avg_bits10, aom_highbd_10_variance8x4,
+                   aom_highbd_10_sub_pixel_variance8x4,
+                   aom_highbd_10_sub_pixel_avg_variance8x4, NULL,
+                   aom_highbd_sad8x4x8_bits10, aom_highbd_sad8x4x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_4X8, aom_highbd_sad4x8_bits10,
+                   aom_highbd_sad4x8_avg_bits10, aom_highbd_10_variance4x8,
+                   aom_highbd_10_sub_pixel_variance4x8,
+                   aom_highbd_10_sub_pixel_avg_variance4x8, NULL,
+                   aom_highbd_sad4x8x8_bits10, aom_highbd_sad4x8x4d_bits10)
+
+        HIGHBD_BFP(
+            BLOCK_4X4, aom_highbd_sad4x4_bits10, aom_highbd_sad4x4_avg_bits10,
+            aom_highbd_10_variance4x4, aom_highbd_10_sub_pixel_variance4x4,
+            aom_highbd_10_sub_pixel_avg_variance4x4, aom_highbd_sad4x4x3_bits10,
+            aom_highbd_sad4x4x8_bits10, aom_highbd_sad4x4x4d_bits10)
+
+#if CONFIG_CB4X4
+        HIGHBD_BFP(BLOCK_2X2, NULL, NULL, aom_highbd_10_variance2x2, NULL, NULL,
+                   NULL, NULL, NULL)
+        HIGHBD_BFP(BLOCK_4X2, NULL, NULL, aom_highbd_10_variance4x2, NULL, NULL,
+                   NULL, NULL, NULL)
+        HIGHBD_BFP(BLOCK_2X4, NULL, NULL, aom_highbd_10_variance2x4, NULL, NULL,
+                   NULL, NULL, NULL)
+#endif
+
+#if CONFIG_EXT_PARTITION
+        HIGHBD_BFP(
+            BLOCK_128X128, aom_highbd_sad128x128_bits10,
+            aom_highbd_sad128x128_avg_bits10, aom_highbd_10_variance128x128,
+            aom_highbd_10_sub_pixel_variance128x128,
+            aom_highbd_10_sub_pixel_avg_variance128x128,
+            aom_highbd_sad128x128x3_bits10, aom_highbd_sad128x128x8_bits10,
+            aom_highbd_sad128x128x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_128X64, aom_highbd_sad128x64_bits10,
+                   aom_highbd_sad128x64_avg_bits10,
+                   aom_highbd_10_variance128x64,
+                   aom_highbd_10_sub_pixel_variance128x64,
+                   aom_highbd_10_sub_pixel_avg_variance128x64, NULL, NULL,
+                   aom_highbd_sad128x64x4d_bits10)
+
+        HIGHBD_BFP(BLOCK_64X128, aom_highbd_sad64x128_bits10,
+                   aom_highbd_sad64x128_avg_bits10,
+                   aom_highbd_10_variance64x128,
+                   aom_highbd_10_sub_pixel_variance64x128,
+                   aom_highbd_10_sub_pixel_avg_variance64x128, NULL, NULL,
+                   aom_highbd_sad64x128x4d_bits10)
+#endif  // CONFIG_EXT_PARTITION
+
+#if CONFIG_EXT_INTER
+#if CONFIG_EXT_PARTITION
+        HIGHBD_MBFP(BLOCK_128X128, aom_highbd_masked_sad128x128_bits10,
+                    aom_highbd_10_masked_variance128x128,
+                    aom_highbd_10_masked_sub_pixel_variance128x128)
+        HIGHBD_MBFP(BLOCK_128X64, aom_highbd_masked_sad128x64_bits10,
+                    aom_highbd_10_masked_variance128x64,
+                    aom_highbd_10_masked_sub_pixel_variance128x64)
+        HIGHBD_MBFP(BLOCK_64X128, aom_highbd_masked_sad64x128_bits10,
+                    aom_highbd_10_masked_variance64x128,
+                    aom_highbd_10_masked_sub_pixel_variance64x128)
+#endif  // CONFIG_EXT_PARTITION
+        HIGHBD_MBFP(BLOCK_64X64, aom_highbd_masked_sad64x64_bits10,
+                    aom_highbd_10_masked_variance64x64,
+                    aom_highbd_10_masked_sub_pixel_variance64x64)
+        HIGHBD_MBFP(BLOCK_64X32, aom_highbd_masked_sad64x32_bits10,
+                    aom_highbd_10_masked_variance64x32,
+                    aom_highbd_10_masked_sub_pixel_variance64x32)
+        HIGHBD_MBFP(BLOCK_32X64, aom_highbd_masked_sad32x64_bits10,
+                    aom_highbd_10_masked_variance32x64,
+                    aom_highbd_10_masked_sub_pixel_variance32x64)
+        HIGHBD_MBFP(BLOCK_32X32, aom_highbd_masked_sad32x32_bits10,
+                    aom_highbd_10_masked_variance32x32,
+                    aom_highbd_10_masked_sub_pixel_variance32x32)
+        HIGHBD_MBFP(BLOCK_32X16, aom_highbd_masked_sad32x16_bits10,
+                    aom_highbd_10_masked_variance32x16,
+                    aom_highbd_10_masked_sub_pixel_variance32x16)
+        HIGHBD_MBFP(BLOCK_16X32, aom_highbd_masked_sad16x32_bits10,
+                    aom_highbd_10_masked_variance16x32,
+                    aom_highbd_10_masked_sub_pixel_variance16x32)
+        HIGHBD_MBFP(BLOCK_16X16, aom_highbd_masked_sad16x16_bits10,
+                    aom_highbd_10_masked_variance16x16,
+                    aom_highbd_10_masked_sub_pixel_variance16x16)
+        HIGHBD_MBFP(BLOCK_8X16, aom_highbd_masked_sad8x16_bits10,
+                    aom_highbd_10_masked_variance8x16,
+                    aom_highbd_10_masked_sub_pixel_variance8x16)
+        HIGHBD_MBFP(BLOCK_16X8, aom_highbd_masked_sad16x8_bits10,
+                    aom_highbd_10_masked_variance16x8,
+                    aom_highbd_10_masked_sub_pixel_variance16x8)
+        HIGHBD_MBFP(BLOCK_8X8, aom_highbd_masked_sad8x8_bits10,
+                    aom_highbd_10_masked_variance8x8,
+                    aom_highbd_10_masked_sub_pixel_variance8x8)
+        HIGHBD_MBFP(BLOCK_4X8, aom_highbd_masked_sad4x8_bits10,
+                    aom_highbd_10_masked_variance4x8,
+                    aom_highbd_10_masked_sub_pixel_variance4x8)
+        HIGHBD_MBFP(BLOCK_8X4, aom_highbd_masked_sad8x4_bits10,
+                    aom_highbd_10_masked_variance8x4,
+                    aom_highbd_10_masked_sub_pixel_variance8x4)
+        HIGHBD_MBFP(BLOCK_4X4, aom_highbd_masked_sad4x4_bits10,
+                    aom_highbd_10_masked_variance4x4,
+                    aom_highbd_10_masked_sub_pixel_variance4x4)
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR
+#if CONFIG_EXT_PARTITION
+        HIGHBD_OBFP(BLOCK_128X128, aom_highbd_obmc_sad128x128_bits10,
+                    aom_highbd_10_obmc_variance128x128,
+                    aom_highbd_10_obmc_sub_pixel_variance128x128)
+        HIGHBD_OBFP(BLOCK_128X64, aom_highbd_obmc_sad128x64_bits10,
+                    aom_highbd_10_obmc_variance128x64,
+                    aom_highbd_10_obmc_sub_pixel_variance128x64)
+        HIGHBD_OBFP(BLOCK_64X128, aom_highbd_obmc_sad64x128_bits10,
+                    aom_highbd_10_obmc_variance64x128,
+                    aom_highbd_10_obmc_sub_pixel_variance64x128)
+#endif  // CONFIG_EXT_PARTITION
+        HIGHBD_OBFP(BLOCK_64X64, aom_highbd_obmc_sad64x64_bits10,
+                    aom_highbd_10_obmc_variance64x64,
+                    aom_highbd_10_obmc_sub_pixel_variance64x64)
+        HIGHBD_OBFP(BLOCK_64X32, aom_highbd_obmc_sad64x32_bits10,
+                    aom_highbd_10_obmc_variance64x32,
+                    aom_highbd_10_obmc_sub_pixel_variance64x32)
+        HIGHBD_OBFP(BLOCK_32X64, aom_highbd_obmc_sad32x64_bits10,
+                    aom_highbd_10_obmc_variance32x64,
+                    aom_highbd_10_obmc_sub_pixel_variance32x64)
+        HIGHBD_OBFP(BLOCK_32X32, aom_highbd_obmc_sad32x32_bits10,
+                    aom_highbd_10_obmc_variance32x32,
+                    aom_highbd_10_obmc_sub_pixel_variance32x32)
+        HIGHBD_OBFP(BLOCK_32X16, aom_highbd_obmc_sad32x16_bits10,
+                    aom_highbd_10_obmc_variance32x16,
+                    aom_highbd_10_obmc_sub_pixel_variance32x16)
+        HIGHBD_OBFP(BLOCK_16X32, aom_highbd_obmc_sad16x32_bits10,
+                    aom_highbd_10_obmc_variance16x32,
+                    aom_highbd_10_obmc_sub_pixel_variance16x32)
+        HIGHBD_OBFP(BLOCK_16X16, aom_highbd_obmc_sad16x16_bits10,
+                    aom_highbd_10_obmc_variance16x16,
+                    aom_highbd_10_obmc_sub_pixel_variance16x16)
+        HIGHBD_OBFP(BLOCK_8X16, aom_highbd_obmc_sad8x16_bits10,
+                    aom_highbd_10_obmc_variance8x16,
+                    aom_highbd_10_obmc_sub_pixel_variance8x16)
+        HIGHBD_OBFP(BLOCK_16X8, aom_highbd_obmc_sad16x8_bits10,
+                    aom_highbd_10_obmc_variance16x8,
+                    aom_highbd_10_obmc_sub_pixel_variance16x8)
+        HIGHBD_OBFP(BLOCK_8X8, aom_highbd_obmc_sad8x8_bits10,
+                    aom_highbd_10_obmc_variance8x8,
+                    aom_highbd_10_obmc_sub_pixel_variance8x8)
+        HIGHBD_OBFP(BLOCK_4X8, aom_highbd_obmc_sad4x8_bits10,
+                    aom_highbd_10_obmc_variance4x8,
+                    aom_highbd_10_obmc_sub_pixel_variance4x8)
+        HIGHBD_OBFP(BLOCK_8X4, aom_highbd_obmc_sad8x4_bits10,
+                    aom_highbd_10_obmc_variance8x4,
+                    aom_highbd_10_obmc_sub_pixel_variance8x4)
+        HIGHBD_OBFP(BLOCK_4X4, aom_highbd_obmc_sad4x4_bits10,
+                    aom_highbd_10_obmc_variance4x4,
+                    aom_highbd_10_obmc_sub_pixel_variance4x4)
+#endif  // CONFIG_MOTION_VAR
+        break;
+
+      case AOM_BITS_12:
+        HIGHBD_BFP(BLOCK_32X16, aom_highbd_sad32x16_bits12,
+                   aom_highbd_sad32x16_avg_bits12, aom_highbd_12_variance32x16,
+                   aom_highbd_12_sub_pixel_variance32x16,
+                   aom_highbd_12_sub_pixel_avg_variance32x16, NULL, NULL,
+                   aom_highbd_sad32x16x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_16X32, aom_highbd_sad16x32_bits12,
+                   aom_highbd_sad16x32_avg_bits12, aom_highbd_12_variance16x32,
+                   aom_highbd_12_sub_pixel_variance16x32,
+                   aom_highbd_12_sub_pixel_avg_variance16x32, NULL, NULL,
+                   aom_highbd_sad16x32x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_64X32, aom_highbd_sad64x32_bits12,
+                   aom_highbd_sad64x32_avg_bits12, aom_highbd_12_variance64x32,
+                   aom_highbd_12_sub_pixel_variance64x32,
+                   aom_highbd_12_sub_pixel_avg_variance64x32, NULL, NULL,
+                   aom_highbd_sad64x32x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_32X64, aom_highbd_sad32x64_bits12,
+                   aom_highbd_sad32x64_avg_bits12, aom_highbd_12_variance32x64,
+                   aom_highbd_12_sub_pixel_variance32x64,
+                   aom_highbd_12_sub_pixel_avg_variance32x64, NULL, NULL,
+                   aom_highbd_sad32x64x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_32X32, aom_highbd_sad32x32_bits12,
+                   aom_highbd_sad32x32_avg_bits12, aom_highbd_12_variance32x32,
+                   aom_highbd_12_sub_pixel_variance32x32,
+                   aom_highbd_12_sub_pixel_avg_variance32x32,
+                   aom_highbd_sad32x32x3_bits12, aom_highbd_sad32x32x8_bits12,
+                   aom_highbd_sad32x32x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_64X64, aom_highbd_sad64x64_bits12,
+                   aom_highbd_sad64x64_avg_bits12, aom_highbd_12_variance64x64,
+                   aom_highbd_12_sub_pixel_variance64x64,
+                   aom_highbd_12_sub_pixel_avg_variance64x64,
+                   aom_highbd_sad64x64x3_bits12, aom_highbd_sad64x64x8_bits12,
+                   aom_highbd_sad64x64x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_16X16, aom_highbd_sad16x16_bits12,
+                   aom_highbd_sad16x16_avg_bits12, aom_highbd_12_variance16x16,
+                   aom_highbd_12_sub_pixel_variance16x16,
+                   aom_highbd_12_sub_pixel_avg_variance16x16,
+                   aom_highbd_sad16x16x3_bits12, aom_highbd_sad16x16x8_bits12,
+                   aom_highbd_sad16x16x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_16X8, aom_highbd_sad16x8_bits12,
+                   aom_highbd_sad16x8_avg_bits12, aom_highbd_12_variance16x8,
+                   aom_highbd_12_sub_pixel_variance16x8,
+                   aom_highbd_12_sub_pixel_avg_variance16x8,
+                   aom_highbd_sad16x8x3_bits12, aom_highbd_sad16x8x8_bits12,
+                   aom_highbd_sad16x8x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_8X16, aom_highbd_sad8x16_bits12,
+                   aom_highbd_sad8x16_avg_bits12, aom_highbd_12_variance8x16,
+                   aom_highbd_12_sub_pixel_variance8x16,
+                   aom_highbd_12_sub_pixel_avg_variance8x16,
+                   aom_highbd_sad8x16x3_bits12, aom_highbd_sad8x16x8_bits12,
+                   aom_highbd_sad8x16x4d_bits12)
+
+        HIGHBD_BFP(
+            BLOCK_8X8, aom_highbd_sad8x8_bits12, aom_highbd_sad8x8_avg_bits12,
+            aom_highbd_12_variance8x8, aom_highbd_12_sub_pixel_variance8x8,
+            aom_highbd_12_sub_pixel_avg_variance8x8, aom_highbd_sad8x8x3_bits12,
+            aom_highbd_sad8x8x8_bits12, aom_highbd_sad8x8x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_8X4, aom_highbd_sad8x4_bits12,
+                   aom_highbd_sad8x4_avg_bits12, aom_highbd_12_variance8x4,
+                   aom_highbd_12_sub_pixel_variance8x4,
+                   aom_highbd_12_sub_pixel_avg_variance8x4, NULL,
+                   aom_highbd_sad8x4x8_bits12, aom_highbd_sad8x4x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_4X8, aom_highbd_sad4x8_bits12,
+                   aom_highbd_sad4x8_avg_bits12, aom_highbd_12_variance4x8,
+                   aom_highbd_12_sub_pixel_variance4x8,
+                   aom_highbd_12_sub_pixel_avg_variance4x8, NULL,
+                   aom_highbd_sad4x8x8_bits12, aom_highbd_sad4x8x4d_bits12)
+
+        HIGHBD_BFP(
+            BLOCK_4X4, aom_highbd_sad4x4_bits12, aom_highbd_sad4x4_avg_bits12,
+            aom_highbd_12_variance4x4, aom_highbd_12_sub_pixel_variance4x4,
+            aom_highbd_12_sub_pixel_avg_variance4x4, aom_highbd_sad4x4x3_bits12,
+            aom_highbd_sad4x4x8_bits12, aom_highbd_sad4x4x4d_bits12)
+
+#if CONFIG_CB4X4
+        HIGHBD_BFP(BLOCK_2X2, NULL, NULL, aom_highbd_12_variance2x2, NULL, NULL,
+                   NULL, NULL, NULL)
+        HIGHBD_BFP(BLOCK_4X2, NULL, NULL, aom_highbd_12_variance4x2, NULL, NULL,
+                   NULL, NULL, NULL)
+        HIGHBD_BFP(BLOCK_2X4, NULL, NULL, aom_highbd_12_variance2x4, NULL, NULL,
+                   NULL, NULL, NULL)
+#endif
+
+#if CONFIG_EXT_PARTITION
+        HIGHBD_BFP(
+            BLOCK_128X128, aom_highbd_sad128x128_bits12,
+            aom_highbd_sad128x128_avg_bits12, aom_highbd_12_variance128x128,
+            aom_highbd_12_sub_pixel_variance128x128,
+            aom_highbd_12_sub_pixel_avg_variance128x128,
+            aom_highbd_sad128x128x3_bits12, aom_highbd_sad128x128x8_bits12,
+            aom_highbd_sad128x128x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_128X64, aom_highbd_sad128x64_bits12,
+                   aom_highbd_sad128x64_avg_bits12,
+                   aom_highbd_12_variance128x64,
+                   aom_highbd_12_sub_pixel_variance128x64,
+                   aom_highbd_12_sub_pixel_avg_variance128x64, NULL, NULL,
+                   aom_highbd_sad128x64x4d_bits12)
+
+        HIGHBD_BFP(BLOCK_64X128, aom_highbd_sad64x128_bits12,
+                   aom_highbd_sad64x128_avg_bits12,
+                   aom_highbd_12_variance64x128,
+                   aom_highbd_12_sub_pixel_variance64x128,
+                   aom_highbd_12_sub_pixel_avg_variance64x128, NULL, NULL,
+                   aom_highbd_sad64x128x4d_bits12)
+#endif  // CONFIG_EXT_PARTITION
+
+#if CONFIG_EXT_INTER
+#if CONFIG_EXT_PARTITION
+        HIGHBD_MBFP(BLOCK_128X128, aom_highbd_masked_sad128x128_bits12,
+                    aom_highbd_12_masked_variance128x128,
+                    aom_highbd_12_masked_sub_pixel_variance128x128)
+        HIGHBD_MBFP(BLOCK_128X64, aom_highbd_masked_sad128x64_bits12,
+                    aom_highbd_12_masked_variance128x64,
+                    aom_highbd_12_masked_sub_pixel_variance128x64)
+        HIGHBD_MBFP(BLOCK_64X128, aom_highbd_masked_sad64x128_bits12,
+                    aom_highbd_12_masked_variance64x128,
+                    aom_highbd_12_masked_sub_pixel_variance64x128)
+#endif  // CONFIG_EXT_PARTITION
+        HIGHBD_MBFP(BLOCK_64X64, aom_highbd_masked_sad64x64_bits12,
+                    aom_highbd_12_masked_variance64x64,
+                    aom_highbd_12_masked_sub_pixel_variance64x64)
+        HIGHBD_MBFP(BLOCK_64X32, aom_highbd_masked_sad64x32_bits12,
+                    aom_highbd_12_masked_variance64x32,
+                    aom_highbd_12_masked_sub_pixel_variance64x32)
+        HIGHBD_MBFP(BLOCK_32X64, aom_highbd_masked_sad32x64_bits12,
+                    aom_highbd_12_masked_variance32x64,
+                    aom_highbd_12_masked_sub_pixel_variance32x64)
+        HIGHBD_MBFP(BLOCK_32X32, aom_highbd_masked_sad32x32_bits12,
+                    aom_highbd_12_masked_variance32x32,
+                    aom_highbd_12_masked_sub_pixel_variance32x32)
+        HIGHBD_MBFP(BLOCK_32X16, aom_highbd_masked_sad32x16_bits12,
+                    aom_highbd_12_masked_variance32x16,
+                    aom_highbd_12_masked_sub_pixel_variance32x16)
+        HIGHBD_MBFP(BLOCK_16X32, aom_highbd_masked_sad16x32_bits12,
+                    aom_highbd_12_masked_variance16x32,
+                    aom_highbd_12_masked_sub_pixel_variance16x32)
+        HIGHBD_MBFP(BLOCK_16X16, aom_highbd_masked_sad16x16_bits12,
+                    aom_highbd_12_masked_variance16x16,
+                    aom_highbd_12_masked_sub_pixel_variance16x16)
+        HIGHBD_MBFP(BLOCK_8X16, aom_highbd_masked_sad8x16_bits12,
+                    aom_highbd_12_masked_variance8x16,
+                    aom_highbd_12_masked_sub_pixel_variance8x16)
+        HIGHBD_MBFP(BLOCK_16X8, aom_highbd_masked_sad16x8_bits12,
+                    aom_highbd_12_masked_variance16x8,
+                    aom_highbd_12_masked_sub_pixel_variance16x8)
+        HIGHBD_MBFP(BLOCK_8X8, aom_highbd_masked_sad8x8_bits12,
+                    aom_highbd_12_masked_variance8x8,
+                    aom_highbd_12_masked_sub_pixel_variance8x8)
+        HIGHBD_MBFP(BLOCK_4X8, aom_highbd_masked_sad4x8_bits12,
+                    aom_highbd_12_masked_variance4x8,
+                    aom_highbd_12_masked_sub_pixel_variance4x8)
+        HIGHBD_MBFP(BLOCK_8X4, aom_highbd_masked_sad8x4_bits12,
+                    aom_highbd_12_masked_variance8x4,
+                    aom_highbd_12_masked_sub_pixel_variance8x4)
+        HIGHBD_MBFP(BLOCK_4X4, aom_highbd_masked_sad4x4_bits12,
+                    aom_highbd_12_masked_variance4x4,
+                    aom_highbd_12_masked_sub_pixel_variance4x4)
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR
+#if CONFIG_EXT_PARTITION
+        HIGHBD_OBFP(BLOCK_128X128, aom_highbd_obmc_sad128x128_bits12,
+                    aom_highbd_12_obmc_variance128x128,
+                    aom_highbd_12_obmc_sub_pixel_variance128x128)
+        HIGHBD_OBFP(BLOCK_128X64, aom_highbd_obmc_sad128x64_bits12,
+                    aom_highbd_12_obmc_variance128x64,
+                    aom_highbd_12_obmc_sub_pixel_variance128x64)
+        HIGHBD_OBFP(BLOCK_64X128, aom_highbd_obmc_sad64x128_bits12,
+                    aom_highbd_12_obmc_variance64x128,
+                    aom_highbd_12_obmc_sub_pixel_variance64x128)
+#endif  // CONFIG_EXT_PARTITION
+        HIGHBD_OBFP(BLOCK_64X64, aom_highbd_obmc_sad64x64_bits12,
+                    aom_highbd_12_obmc_variance64x64,
+                    aom_highbd_12_obmc_sub_pixel_variance64x64)
+        HIGHBD_OBFP(BLOCK_64X32, aom_highbd_obmc_sad64x32_bits12,
+                    aom_highbd_12_obmc_variance64x32,
+                    aom_highbd_12_obmc_sub_pixel_variance64x32)
+        HIGHBD_OBFP(BLOCK_32X64, aom_highbd_obmc_sad32x64_bits12,
+                    aom_highbd_12_obmc_variance32x64,
+                    aom_highbd_12_obmc_sub_pixel_variance32x64)
+        HIGHBD_OBFP(BLOCK_32X32, aom_highbd_obmc_sad32x32_bits12,
+                    aom_highbd_12_obmc_variance32x32,
+                    aom_highbd_12_obmc_sub_pixel_variance32x32)
+        HIGHBD_OBFP(BLOCK_32X16, aom_highbd_obmc_sad32x16_bits12,
+                    aom_highbd_12_obmc_variance32x16,
+                    aom_highbd_12_obmc_sub_pixel_variance32x16)
+        HIGHBD_OBFP(BLOCK_16X32, aom_highbd_obmc_sad16x32_bits12,
+                    aom_highbd_12_obmc_variance16x32,
+                    aom_highbd_12_obmc_sub_pixel_variance16x32)
+        HIGHBD_OBFP(BLOCK_16X16, aom_highbd_obmc_sad16x16_bits12,
+                    aom_highbd_12_obmc_variance16x16,
+                    aom_highbd_12_obmc_sub_pixel_variance16x16)
+        HIGHBD_OBFP(BLOCK_8X16, aom_highbd_obmc_sad8x16_bits12,
+                    aom_highbd_12_obmc_variance8x16,
+                    aom_highbd_12_obmc_sub_pixel_variance8x16)
+        HIGHBD_OBFP(BLOCK_16X8, aom_highbd_obmc_sad16x8_bits12,
+                    aom_highbd_12_obmc_variance16x8,
+                    aom_highbd_12_obmc_sub_pixel_variance16x8)
+        HIGHBD_OBFP(BLOCK_8X8, aom_highbd_obmc_sad8x8_bits12,
+                    aom_highbd_12_obmc_variance8x8,
+                    aom_highbd_12_obmc_sub_pixel_variance8x8)
+        HIGHBD_OBFP(BLOCK_4X8, aom_highbd_obmc_sad4x8_bits12,
+                    aom_highbd_12_obmc_variance4x8,
+                    aom_highbd_12_obmc_sub_pixel_variance4x8)
+        HIGHBD_OBFP(BLOCK_8X4, aom_highbd_obmc_sad8x4_bits12,
+                    aom_highbd_12_obmc_variance8x4,
+                    aom_highbd_12_obmc_sub_pixel_variance8x4)
+        HIGHBD_OBFP(BLOCK_4X4, aom_highbd_obmc_sad4x4_bits12,
+                    aom_highbd_12_obmc_variance4x4,
+                    aom_highbd_12_obmc_sub_pixel_variance4x4)
+#endif  // CONFIG_MOTION_VAR
+        break;
+
+      default:
+        assert(0 &&
+               "cm->bit_depth should be AOM_BITS_8, "
+               "AOM_BITS_10 or AOM_BITS_12");
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static void realloc_segmentation_maps(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+
+  // Create the encoder segmentation map and set all entries to 0
+  aom_free(cpi->segmentation_map);
+  CHECK_MEM_ERROR(cm, cpi->segmentation_map,
+                  aom_calloc(cm->mi_rows * cm->mi_cols, 1));
+
+  // Create a map used for cyclic background refresh.
+  if (cpi->cyclic_refresh) av1_cyclic_refresh_free(cpi->cyclic_refresh);
+  CHECK_MEM_ERROR(cm, cpi->cyclic_refresh,
+                  av1_cyclic_refresh_alloc(cm->mi_rows, cm->mi_cols));
+
+  // Create a map used to mark inactive areas.
+  aom_free(cpi->active_map.map);
+  CHECK_MEM_ERROR(cm, cpi->active_map.map,
+                  aom_calloc(cm->mi_rows * cm->mi_cols, 1));
+}
+
+void av1_change_config(struct AV1_COMP *cpi, const AV1EncoderConfig *oxcf) {
+  AV1_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  if (cm->profile != oxcf->profile) cm->profile = oxcf->profile;
+  cm->bit_depth = oxcf->bit_depth;
+  cm->color_space = oxcf->color_space;
+  cm->color_range = oxcf->color_range;
+
+  if (cm->profile <= PROFILE_1)
+    assert(cm->bit_depth == AOM_BITS_8);
+  else
+    assert(cm->bit_depth > AOM_BITS_8);
+
+  cpi->oxcf = *oxcf;
+#if CONFIG_HIGHBITDEPTH
+  cpi->td.mb.e_mbd.bd = (int)cm->bit_depth;
+#endif  // CONFIG_HIGHBITDEPTH
+#if CONFIG_GLOBAL_MOTION
+  cpi->td.mb.e_mbd.global_motion = cm->global_motion;
+#endif  // CONFIG_GLOBAL_MOTION
+
+  if ((oxcf->pass == 0) && (oxcf->rc_mode == AOM_Q)) {
+    rc->baseline_gf_interval = FIXED_GF_INTERVAL;
+  } else {
+    rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2;
+  }
+
+  cpi->refresh_last_frame = 1;
+  cpi->refresh_golden_frame = 0;
+#if CONFIG_EXT_REFS
+  cpi->refresh_bwd_ref_frame = 0;
+#endif  // CONFIG_EXT_REFS
+
+  cm->refresh_frame_context =
+      (oxcf->error_resilient_mode || oxcf->frame_parallel_decoding_mode)
+          ? REFRESH_FRAME_CONTEXT_FORWARD
+          : REFRESH_FRAME_CONTEXT_BACKWARD;
+  cm->reset_frame_context = RESET_FRAME_CONTEXT_NONE;
+
+#if CONFIG_PALETTE
+  cm->allow_screen_content_tools = (cpi->oxcf.content == AOM_CONTENT_SCREEN);
+  if (cm->allow_screen_content_tools) {
+    MACROBLOCK *x = &cpi->td.mb;
+    if (x->palette_buffer == 0) {
+      CHECK_MEM_ERROR(cm, x->palette_buffer,
+                      aom_memalign(16, sizeof(*x->palette_buffer)));
+    }
+    // Reallocate the pc_tree, as it's contents depends on
+    // the state of cm->allow_screen_content_tools
+    av1_free_pc_tree(&cpi->td);
+    av1_setup_pc_tree(&cpi->common, &cpi->td);
+  }
+#endif  // CONFIG_PALETTE
+
+  av1_reset_segment_features(cm);
+  av1_set_high_precision_mv(cpi, 0);
+
+  set_rc_buffer_sizes(rc, &cpi->oxcf);
+
+  // Under a configuration change, where maximum_buffer_size may change,
+  // keep buffer level clipped to the maximum allowed buffer size.
+  rc->bits_off_target = AOMMIN(rc->bits_off_target, rc->maximum_buffer_size);
+  rc->buffer_level = AOMMIN(rc->buffer_level, rc->maximum_buffer_size);
+
+  // Set up frame rate and related parameters rate control values.
+  av1_new_framerate(cpi, cpi->framerate);
+
+  // Set absolute upper and lower quality limits
+  rc->worst_quality = cpi->oxcf.worst_allowed_q;
+  rc->best_quality = cpi->oxcf.best_allowed_q;
+
+  cm->interp_filter = cpi->sf.default_interp_filter;
+
+  if (cpi->oxcf.render_width > 0 && cpi->oxcf.render_height > 0) {
+    cm->render_width = cpi->oxcf.render_width;
+    cm->render_height = cpi->oxcf.render_height;
+  } else {
+    cm->render_width = cpi->oxcf.width;
+    cm->render_height = cpi->oxcf.height;
+  }
+  cm->width = cpi->oxcf.width;
+  cm->height = cpi->oxcf.height;
+
+  if (cpi->initial_width) {
+    if (cm->width > cpi->initial_width || cm->height > cpi->initial_height) {
+      av1_free_context_buffers(cm);
+      av1_alloc_compressor_data(cpi);
+      realloc_segmentation_maps(cpi);
+      cpi->initial_width = cpi->initial_height = 0;
+    }
+  }
+  update_frame_size(cpi);
+
+  cpi->alt_ref_source = NULL;
+  rc->is_src_frame_alt_ref = 0;
+
+#if CONFIG_EXT_REFS
+  rc->is_bwd_ref_frame = 0;
+  rc->is_last_bipred_frame = 0;
+  rc->is_bipred_frame = 0;
+#endif  // CONFIG_EXT_REFS
+
+#if 0
+  // Experimental RD Code
+  cpi->frame_distortion = 0;
+  cpi->last_frame_distortion = 0;
+#endif
+
+  set_tile_info(cpi);
+
+  cpi->ext_refresh_frame_flags_pending = 0;
+  cpi->ext_refresh_frame_context_pending = 0;
+
+#if CONFIG_HIGHBITDEPTH
+  highbd_set_var_fns(cpi);
+#endif
+
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  cpi->common.ans_window_size_log2 = cpi->oxcf.ans_window_size_log2;
+  if (cpi->buf_ans.size != (1 << cpi->common.ans_window_size_log2)) {
+    aom_buf_ans_free(&cpi->buf_ans);
+    aom_buf_ans_alloc(&cpi->buf_ans, &cpi->common.error,
+                      1 << cpi->common.ans_window_size_log2);
+  }
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+}
+
+#ifndef M_LOG2_E
+#define M_LOG2_E 0.693147180559945309417
+#endif
+#define log2f(x) (log(x) / (float)M_LOG2_E)
+
+#if !CONFIG_REF_MV
+static void cal_nmvjointsadcost(int *mvjointsadcost) {
+  mvjointsadcost[0] = 600;
+  mvjointsadcost[1] = 300;
+  mvjointsadcost[2] = 300;
+  mvjointsadcost[3] = 300;
+}
+#endif
+
+static void cal_nmvsadcosts(int *mvsadcost[2]) {
+  int i = 1;
+
+  mvsadcost[0][0] = 0;
+  mvsadcost[1][0] = 0;
+
+  do {
+    double z = 256 * (2 * (log2f(8 * i) + .6));
+    mvsadcost[0][i] = (int)z;
+    mvsadcost[1][i] = (int)z;
+    mvsadcost[0][-i] = (int)z;
+    mvsadcost[1][-i] = (int)z;
+  } while (++i <= MV_MAX);
+}
+
+static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
+  int i = 1;
+
+  mvsadcost[0][0] = 0;
+  mvsadcost[1][0] = 0;
+
+  do {
+    double z = 256 * (2 * (log2f(8 * i) + .6));
+    mvsadcost[0][i] = (int)z;
+    mvsadcost[1][i] = (int)z;
+    mvsadcost[0][-i] = (int)z;
+    mvsadcost[1][-i] = (int)z;
+  } while (++i <= MV_MAX);
+}
+
+static INLINE void init_upsampled_ref_frame_bufs(AV1_COMP *cpi) {
+  int i;
+
+  for (i = 0; i < (REF_FRAMES + 1); ++i) {
+    cpi->upsampled_ref_bufs[i].ref_count = 0;
+    cpi->upsampled_ref_idx[i] = INVALID_IDX;
+  }
+}
+
+AV1_COMP *av1_create_compressor(AV1EncoderConfig *oxcf,
+                                BufferPool *const pool) {
+  unsigned int i;
+  AV1_COMP *volatile const cpi = aom_memalign(32, sizeof(AV1_COMP));
+  AV1_COMMON *volatile const cm = cpi != NULL ? &cpi->common : NULL;
+
+  if (!cm) return NULL;
+
+  av1_zero(*cpi);
+
+  if (setjmp(cm->error.jmp)) {
+    cm->error.setjmp = 0;
+    av1_remove_compressor(cpi);
+    return 0;
+  }
+
+  cm->error.setjmp = 1;
+  cm->alloc_mi = av1_enc_alloc_mi;
+  cm->free_mi = av1_enc_free_mi;
+  cm->setup_mi = av1_enc_setup_mi;
+
+  CHECK_MEM_ERROR(cm, cm->fc,
+                  (FRAME_CONTEXT *)aom_memalign(32, sizeof(*cm->fc)));
+  CHECK_MEM_ERROR(cm, cm->frame_contexts,
+                  (FRAME_CONTEXT *)aom_memalign(
+                      32, FRAME_CONTEXTS * sizeof(*cm->frame_contexts)));
+  memset(cm->fc, 0, sizeof(*cm->fc));
+  memset(cm->frame_contexts, 0, FRAME_CONTEXTS * sizeof(*cm->frame_contexts));
+
+  cpi->resize_state = 0;
+  cpi->resize_avg_qp = 0;
+  cpi->resize_buffer_underflow = 0;
+  cpi->common.buffer_pool = pool;
+
+  init_config(cpi, oxcf);
+#if CONFIG_XIPHRC
+  cpi->od_rc.framerate = cpi->framerate;
+  cpi->od_rc.frame_width = cm->render_width;
+  cpi->od_rc.frame_height = cm->render_height;
+  cpi->od_rc.keyframe_rate = oxcf->key_freq;
+  cpi->od_rc.goldenframe_rate = FIXED_GF_INTERVAL;
+  cpi->od_rc.altref_rate = 25;
+  cpi->od_rc.firstpass_quant = 1;
+  cpi->od_rc.bit_depth = cm->bit_depth;
+  cpi->od_rc.minq = oxcf->best_allowed_q;
+  cpi->od_rc.maxq = oxcf->worst_allowed_q;
+  if (cpi->oxcf.rc_mode == AOM_CQ) cpi->od_rc.minq = cpi->od_rc.quality;
+  cpi->od_rc.quality = cpi->oxcf.rc_mode == AOM_Q ? oxcf->cq_level : -1;
+  cpi->od_rc.periodic_boosts = oxcf->frame_periodic_boost;
+  od_enc_rc_init(&cpi->od_rc,
+                 cpi->oxcf.rc_mode == AOM_Q ? -1 : oxcf->target_bandwidth,
+                 oxcf->maximum_buffer_size_ms);
+#else
+  av1_rc_init(&cpi->oxcf, oxcf->pass, &cpi->rc);
+#endif
+
+  cm->current_video_frame = 0;
+  cpi->partition_search_skippable_frame = 0;
+  cpi->tile_data = NULL;
+  cpi->last_show_frame_buf_idx = INVALID_IDX;
+
+  realloc_segmentation_maps(cpi);
+
+#if CONFIG_REF_MV
+  for (i = 0; i < NMV_CONTEXTS; ++i) {
+    memset(cpi->nmv_costs, 0, sizeof(cpi->nmv_costs));
+    memset(cpi->nmv_costs_hp, 0, sizeof(cpi->nmv_costs_hp));
+  }
+#endif
+
+  memset(cpi->nmvcosts, 0, sizeof(cpi->nmvcosts));
+  memset(cpi->nmvcosts_hp, 0, sizeof(cpi->nmvcosts_hp));
+  memset(cpi->nmvsadcosts, 0, sizeof(cpi->nmvsadcosts));
+  memset(cpi->nmvsadcosts_hp, 0, sizeof(cpi->nmvsadcosts_hp));
+
+  for (i = 0; i < (sizeof(cpi->mbgraph_stats) / sizeof(cpi->mbgraph_stats[0]));
+       i++) {
+    CHECK_MEM_ERROR(
+        cm, cpi->mbgraph_stats[i].mb_stats,
+        aom_calloc(cm->MBs * sizeof(*cpi->mbgraph_stats[i].mb_stats), 1));
+  }
+
+#if CONFIG_FP_MB_STATS
+  cpi->use_fp_mb_stats = 0;
+  if (cpi->use_fp_mb_stats) {
+    // a place holder used to store the first pass mb stats in the first pass
+    CHECK_MEM_ERROR(cm, cpi->twopass.frame_mb_stats_buf,
+                    aom_calloc(cm->MBs * sizeof(uint8_t), 1));
+  } else {
+    cpi->twopass.frame_mb_stats_buf = NULL;
+  }
+#endif
+
+  cpi->refresh_alt_ref_frame = 0;
+  cpi->multi_arf_last_grp_enabled = 0;
+
+  cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
+#if CONFIG_INTERNAL_STATS
+  cpi->b_calculate_blockiness = 1;
+  cpi->b_calculate_consistency = 1;
+  cpi->total_inconsistency = 0;
+  cpi->psnr.worst = 100.0;
+  cpi->worst_ssim = 100.0;
+
+  cpi->count = 0;
+  cpi->bytes = 0;
+
+  if (cpi->b_calculate_psnr) {
+    cpi->total_sq_error = 0;
+    cpi->total_samples = 0;
+    cpi->tot_recode_hits = 0;
+    cpi->summed_quality = 0;
+    cpi->summed_weights = 0;
+  }
+
+  cpi->fastssim.worst = 100.0;
+  cpi->psnrhvs.worst = 100.0;
+
+  if (cpi->b_calculate_blockiness) {
+    cpi->total_blockiness = 0;
+    cpi->worst_blockiness = 0.0;
+  }
+
+  if (cpi->b_calculate_consistency) {
+    CHECK_MEM_ERROR(cm, cpi->ssim_vars,
+                    aom_malloc(sizeof(*cpi->ssim_vars) * 4 *
+                               cpi->common.mi_rows * cpi->common.mi_cols));
+    cpi->worst_consistency = 100.0;
+  }
+#endif
+#if CONFIG_ENTROPY_STATS
+  av1_zero(aggregate_fc);
+#endif  // CONFIG_ENTROPY_STATS
+
+  cpi->first_time_stamp_ever = INT64_MAX;
+
+#if CONFIG_REF_MV
+  for (i = 0; i < NMV_CONTEXTS; ++i) {
+    cpi->td.mb.nmvcost[i][0] = &cpi->nmv_costs[i][0][MV_MAX];
+    cpi->td.mb.nmvcost[i][1] = &cpi->nmv_costs[i][1][MV_MAX];
+    cpi->td.mb.nmvcost_hp[i][0] = &cpi->nmv_costs_hp[i][0][MV_MAX];
+    cpi->td.mb.nmvcost_hp[i][1] = &cpi->nmv_costs_hp[i][1][MV_MAX];
+  }
+#else
+  cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
+  cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX];
+  cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX];
+  cpi->td.mb.nmvcost_hp[0] = &cpi->nmvcosts_hp[0][MV_MAX];
+  cpi->td.mb.nmvcost_hp[1] = &cpi->nmvcosts_hp[1][MV_MAX];
+#endif
+  cpi->td.mb.nmvsadcost[0] = &cpi->nmvsadcosts[0][MV_MAX];
+  cpi->td.mb.nmvsadcost[1] = &cpi->nmvsadcosts[1][MV_MAX];
+  cal_nmvsadcosts(cpi->td.mb.nmvsadcost);
+
+  cpi->td.mb.nmvsadcost_hp[0] = &cpi->nmvsadcosts_hp[0][MV_MAX];
+  cpi->td.mb.nmvsadcost_hp[1] = &cpi->nmvsadcosts_hp[1][MV_MAX];
+  cal_nmvsadcosts_hp(cpi->td.mb.nmvsadcost_hp);
+
+#ifdef OUTPUT_YUV_SKINMAP
+  yuv_skinmap_file = fopen("skinmap.yuv", "ab");
+#endif
+#ifdef OUTPUT_YUV_REC
+  yuv_rec_file = fopen("rec.yuv", "wb");
+#endif
+
+#if 0
+  framepsnr = fopen("framepsnr.stt", "a");
+  kf_list = fopen("kf_list.stt", "w");
+#endif
+
+#if CONFIG_XIPHRC
+  if (oxcf->pass == 2) {
+    cpi->od_rc.twopass_allframes_buf = oxcf->two_pass_stats_in.buf;
+    cpi->od_rc.twopass_allframes_buf_size = oxcf->two_pass_stats_in.sz;
+  }
+#else
+  if (oxcf->pass == 1) {
+    av1_init_first_pass(cpi);
+  } else if (oxcf->pass == 2) {
+    const size_t packet_sz = sizeof(FIRSTPASS_STATS);
+    const int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz);
+
+#if CONFIG_FP_MB_STATS
+    if (cpi->use_fp_mb_stats) {
+      const size_t psz = cpi->common.MBs * sizeof(uint8_t);
+      const int ps = (int)(oxcf->firstpass_mb_stats_in.sz / psz);
+
+      cpi->twopass.firstpass_mb_stats.mb_stats_start =
+          oxcf->firstpass_mb_stats_in.buf;
+      cpi->twopass.firstpass_mb_stats.mb_stats_end =
+          cpi->twopass.firstpass_mb_stats.mb_stats_start +
+          (ps - 1) * cpi->common.MBs * sizeof(uint8_t);
+    }
+#endif
+
+    cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf;
+    cpi->twopass.stats_in = cpi->twopass.stats_in_start;
+    cpi->twopass.stats_in_end = &cpi->twopass.stats_in[packets - 1];
+
+    av1_init_second_pass(cpi);
+  }
+#endif
+
+  init_upsampled_ref_frame_bufs(cpi);
+
+  av1_set_speed_features_framesize_independent(cpi);
+  av1_set_speed_features_framesize_dependent(cpi);
+
+  // Allocate memory to store variances for a frame.
+  CHECK_MEM_ERROR(cm, cpi->source_diff_var,
+                  aom_calloc(cm->MBs, sizeof(*cpi->source_diff_var)));
+  cpi->source_var_thresh = 0;
+  cpi->frames_till_next_var_check = 0;
+
+#define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF) \
+  cpi->fn_ptr[BT].sdf = SDF;                                    \
+  cpi->fn_ptr[BT].sdaf = SDAF;                                  \
+  cpi->fn_ptr[BT].vf = VF;                                      \
+  cpi->fn_ptr[BT].svf = SVF;                                    \
+  cpi->fn_ptr[BT].svaf = SVAF;                                  \
+  cpi->fn_ptr[BT].sdx3f = SDX3F;                                \
+  cpi->fn_ptr[BT].sdx8f = SDX8F;                                \
+  cpi->fn_ptr[BT].sdx4df = SDX4DF;
+
+#if CONFIG_EXT_PARTITION
+  BFP(BLOCK_128X128, aom_sad128x128, aom_sad128x128_avg, aom_variance128x128,
+      aom_sub_pixel_variance128x128, aom_sub_pixel_avg_variance128x128,
+      aom_sad128x128x3, aom_sad128x128x8, aom_sad128x128x4d)
+
+  BFP(BLOCK_128X64, aom_sad128x64, aom_sad128x64_avg, aom_variance128x64,
+      aom_sub_pixel_variance128x64, aom_sub_pixel_avg_variance128x64, NULL,
+      NULL, aom_sad128x64x4d)
+
+  BFP(BLOCK_64X128, aom_sad64x128, aom_sad64x128_avg, aom_variance64x128,
+      aom_sub_pixel_variance64x128, aom_sub_pixel_avg_variance64x128, NULL,
+      NULL, aom_sad64x128x4d)
+#endif  // CONFIG_EXT_PARTITION
+
+  BFP(BLOCK_32X16, aom_sad32x16, aom_sad32x16_avg, aom_variance32x16,
+      aom_sub_pixel_variance32x16, aom_sub_pixel_avg_variance32x16, NULL, NULL,
+      aom_sad32x16x4d)
+
+  BFP(BLOCK_16X32, aom_sad16x32, aom_sad16x32_avg, aom_variance16x32,
+      aom_sub_pixel_variance16x32, aom_sub_pixel_avg_variance16x32, NULL, NULL,
+      aom_sad16x32x4d)
+
+  BFP(BLOCK_64X32, aom_sad64x32, aom_sad64x32_avg, aom_variance64x32,
+      aom_sub_pixel_variance64x32, aom_sub_pixel_avg_variance64x32, NULL, NULL,
+      aom_sad64x32x4d)
+
+  BFP(BLOCK_32X64, aom_sad32x64, aom_sad32x64_avg, aom_variance32x64,
+      aom_sub_pixel_variance32x64, aom_sub_pixel_avg_variance32x64, NULL, NULL,
+      aom_sad32x64x4d)
+
+  BFP(BLOCK_32X32, aom_sad32x32, aom_sad32x32_avg, aom_variance32x32,
+      aom_sub_pixel_variance32x32, aom_sub_pixel_avg_variance32x32,
+      aom_sad32x32x3, aom_sad32x32x8, aom_sad32x32x4d)
+
+  BFP(BLOCK_64X64, aom_sad64x64, aom_sad64x64_avg, aom_variance64x64,
+      aom_sub_pixel_variance64x64, aom_sub_pixel_avg_variance64x64,
+      aom_sad64x64x3, aom_sad64x64x8, aom_sad64x64x4d)
+
+  BFP(BLOCK_16X16, aom_sad16x16, aom_sad16x16_avg, aom_variance16x16,
+      aom_sub_pixel_variance16x16, aom_sub_pixel_avg_variance16x16,
+      aom_sad16x16x3, aom_sad16x16x8, aom_sad16x16x4d)
+
+  BFP(BLOCK_16X8, aom_sad16x8, aom_sad16x8_avg, aom_variance16x8,
+      aom_sub_pixel_variance16x8, aom_sub_pixel_avg_variance16x8, aom_sad16x8x3,
+      aom_sad16x8x8, aom_sad16x8x4d)
+
+  BFP(BLOCK_8X16, aom_sad8x16, aom_sad8x16_avg, aom_variance8x16,
+      aom_sub_pixel_variance8x16, aom_sub_pixel_avg_variance8x16, aom_sad8x16x3,
+      aom_sad8x16x8, aom_sad8x16x4d)
+
+  BFP(BLOCK_8X8, aom_sad8x8, aom_sad8x8_avg, aom_variance8x8,
+      aom_sub_pixel_variance8x8, aom_sub_pixel_avg_variance8x8, aom_sad8x8x3,
+      aom_sad8x8x8, aom_sad8x8x4d)
+
+  BFP(BLOCK_8X4, aom_sad8x4, aom_sad8x4_avg, aom_variance8x4,
+      aom_sub_pixel_variance8x4, aom_sub_pixel_avg_variance8x4, NULL,
+      aom_sad8x4x8, aom_sad8x4x4d)
+
+  BFP(BLOCK_4X8, aom_sad4x8, aom_sad4x8_avg, aom_variance4x8,
+      aom_sub_pixel_variance4x8, aom_sub_pixel_avg_variance4x8, NULL,
+      aom_sad4x8x8, aom_sad4x8x4d)
+
+  BFP(BLOCK_4X4, aom_sad4x4, aom_sad4x4_avg, aom_variance4x4,
+      aom_sub_pixel_variance4x4, aom_sub_pixel_avg_variance4x4, aom_sad4x4x3,
+      aom_sad4x4x8, aom_sad4x4x4d)
+
+#if CONFIG_CB4X4
+  BFP(BLOCK_2X2, NULL, NULL, aom_variance2x2, NULL, NULL, NULL, NULL, NULL)
+  BFP(BLOCK_2X4, NULL, NULL, aom_variance2x4, NULL, NULL, NULL, NULL, NULL)
+  BFP(BLOCK_4X2, NULL, NULL, aom_variance4x2, NULL, NULL, NULL, NULL, NULL)
+#endif
+
+#if CONFIG_MOTION_VAR
+#define OBFP(BT, OSDF, OVF, OSVF) \
+  cpi->fn_ptr[BT].osdf = OSDF;    \
+  cpi->fn_ptr[BT].ovf = OVF;      \
+  cpi->fn_ptr[BT].osvf = OSVF;
+
+#if CONFIG_EXT_PARTITION
+  OBFP(BLOCK_128X128, aom_obmc_sad128x128, aom_obmc_variance128x128,
+       aom_obmc_sub_pixel_variance128x128)
+  OBFP(BLOCK_128X64, aom_obmc_sad128x64, aom_obmc_variance128x64,
+       aom_obmc_sub_pixel_variance128x64)
+  OBFP(BLOCK_64X128, aom_obmc_sad64x128, aom_obmc_variance64x128,
+       aom_obmc_sub_pixel_variance64x128)
+#endif  // CONFIG_EXT_PARTITION
+  OBFP(BLOCK_64X64, aom_obmc_sad64x64, aom_obmc_variance64x64,
+       aom_obmc_sub_pixel_variance64x64)
+  OBFP(BLOCK_64X32, aom_obmc_sad64x32, aom_obmc_variance64x32,
+       aom_obmc_sub_pixel_variance64x32)
+  OBFP(BLOCK_32X64, aom_obmc_sad32x64, aom_obmc_variance32x64,
+       aom_obmc_sub_pixel_variance32x64)
+  OBFP(BLOCK_32X32, aom_obmc_sad32x32, aom_obmc_variance32x32,
+       aom_obmc_sub_pixel_variance32x32)
+  OBFP(BLOCK_32X16, aom_obmc_sad32x16, aom_obmc_variance32x16,
+       aom_obmc_sub_pixel_variance32x16)
+  OBFP(BLOCK_16X32, aom_obmc_sad16x32, aom_obmc_variance16x32,
+       aom_obmc_sub_pixel_variance16x32)
+  OBFP(BLOCK_16X16, aom_obmc_sad16x16, aom_obmc_variance16x16,
+       aom_obmc_sub_pixel_variance16x16)
+  OBFP(BLOCK_16X8, aom_obmc_sad16x8, aom_obmc_variance16x8,
+       aom_obmc_sub_pixel_variance16x8)
+  OBFP(BLOCK_8X16, aom_obmc_sad8x16, aom_obmc_variance8x16,
+       aom_obmc_sub_pixel_variance8x16)
+  OBFP(BLOCK_8X8, aom_obmc_sad8x8, aom_obmc_variance8x8,
+       aom_obmc_sub_pixel_variance8x8)
+  OBFP(BLOCK_4X8, aom_obmc_sad4x8, aom_obmc_variance4x8,
+       aom_obmc_sub_pixel_variance4x8)
+  OBFP(BLOCK_8X4, aom_obmc_sad8x4, aom_obmc_variance8x4,
+       aom_obmc_sub_pixel_variance8x4)
+  OBFP(BLOCK_4X4, aom_obmc_sad4x4, aom_obmc_variance4x4,
+       aom_obmc_sub_pixel_variance4x4)
+#endif  // CONFIG_MOTION_VAR
+
+#if CONFIG_EXT_INTER
+#define MBFP(BT, MSDF, MVF, MSVF) \
+  cpi->fn_ptr[BT].msdf = MSDF;    \
+  cpi->fn_ptr[BT].mvf = MVF;      \
+  cpi->fn_ptr[BT].msvf = MSVF;
+
+#if CONFIG_EXT_PARTITION
+  MBFP(BLOCK_128X128, aom_masked_sad128x128, aom_masked_variance128x128,
+       aom_masked_sub_pixel_variance128x128)
+  MBFP(BLOCK_128X64, aom_masked_sad128x64, aom_masked_variance128x64,
+       aom_masked_sub_pixel_variance128x64)
+  MBFP(BLOCK_64X128, aom_masked_sad64x128, aom_masked_variance64x128,
+       aom_masked_sub_pixel_variance64x128)
+#endif  // CONFIG_EXT_PARTITION
+  MBFP(BLOCK_64X64, aom_masked_sad64x64, aom_masked_variance64x64,
+       aom_masked_sub_pixel_variance64x64)
+  MBFP(BLOCK_64X32, aom_masked_sad64x32, aom_masked_variance64x32,
+       aom_masked_sub_pixel_variance64x32)
+  MBFP(BLOCK_32X64, aom_masked_sad32x64, aom_masked_variance32x64,
+       aom_masked_sub_pixel_variance32x64)
+  MBFP(BLOCK_32X32, aom_masked_sad32x32, aom_masked_variance32x32,
+       aom_masked_sub_pixel_variance32x32)
+  MBFP(BLOCK_32X16, aom_masked_sad32x16, aom_masked_variance32x16,
+       aom_masked_sub_pixel_variance32x16)
+  MBFP(BLOCK_16X32, aom_masked_sad16x32, aom_masked_variance16x32,
+       aom_masked_sub_pixel_variance16x32)
+  MBFP(BLOCK_16X16, aom_masked_sad16x16, aom_masked_variance16x16,
+       aom_masked_sub_pixel_variance16x16)
+  MBFP(BLOCK_16X8, aom_masked_sad16x8, aom_masked_variance16x8,
+       aom_masked_sub_pixel_variance16x8)
+  MBFP(BLOCK_8X16, aom_masked_sad8x16, aom_masked_variance8x16,
+       aom_masked_sub_pixel_variance8x16)
+  MBFP(BLOCK_8X8, aom_masked_sad8x8, aom_masked_variance8x8,
+       aom_masked_sub_pixel_variance8x8)
+  MBFP(BLOCK_4X8, aom_masked_sad4x8, aom_masked_variance4x8,
+       aom_masked_sub_pixel_variance4x8)
+  MBFP(BLOCK_8X4, aom_masked_sad8x4, aom_masked_variance8x4,
+       aom_masked_sub_pixel_variance8x4)
+  MBFP(BLOCK_4X4, aom_masked_sad4x4, aom_masked_variance4x4,
+       aom_masked_sub_pixel_variance4x4)
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_HIGHBITDEPTH
+  highbd_set_var_fns(cpi);
+#endif
+
+  /* av1_init_quantizer() is first called here. Add check in
+   * av1_frame_init_quantizer() so that av1_init_quantizer is only
+   * called later when needed. This will avoid unnecessary calls of
+   * av1_init_quantizer() for every frame.
+   */
+  av1_init_quantizer(cpi);
+#if CONFIG_AOM_QM
+  aom_qm_init(cm);
+#endif
+
+  av1_loop_filter_init(cm);
+#if CONFIG_LOOP_RESTORATION
+  av1_loop_restoration_precal();
+#endif  // CONFIG_LOOP_RESTORATION
+
+  cm->error.setjmp = 0;
+
+  return cpi;
+}
+
+#define SNPRINT(H, T) snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T))
+
+#define SNPRINT2(H, T, V) \
+  snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T), (V))
+
+void av1_remove_compressor(AV1_COMP *cpi) {
+  AV1_COMMON *cm;
+  unsigned int i;
+  int t;
+
+  if (!cpi) return;
+
+  cm = &cpi->common;
+  if (cm->current_video_frame > 0) {
+#if CONFIG_ENTROPY_STATS
+    if (cpi->oxcf.pass != 1) {
+      fprintf(stderr, "Writing counts.stt\n");
+      FILE *f = fopen("counts.stt", "wb");
+      fwrite(&aggregate_fc, sizeof(aggregate_fc), 1, f);
+      fclose(f);
+    }
+#endif  // CONFIG_ENTROPY_STATS
+#if CONFIG_INTERNAL_STATS
+    aom_clear_system_state();
+
+    if (cpi->oxcf.pass != 1) {
+      char headings[512] = { 0 };
+      char results[512] = { 0 };
+      FILE *f = fopen("opsnr.stt", "a");
+      double time_encoded =
+          (cpi->last_end_time_stamp_seen - cpi->first_time_stamp_ever) /
+          10000000.000;
+      double total_encode_time =
+          (cpi->time_receive_data + cpi->time_compress_data) / 1000.000;
+      const double dr =
+          (double)cpi->bytes * (double)8 / (double)1000 / time_encoded;
+      const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
+      const double target_rate = (double)cpi->oxcf.target_bandwidth / 1000;
+      const double rate_err = ((100.0 * (dr - target_rate)) / target_rate);
+
+      if (cpi->b_calculate_psnr) {
+        const double total_psnr = aom_sse_to_psnr(
+            (double)cpi->total_samples, peak, (double)cpi->total_sq_error);
+        const double total_ssim =
+            100 * pow(cpi->summed_quality / cpi->summed_weights, 8.0);
+        snprintf(headings, sizeof(headings),
+                 "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
+                 "AOMSSIM\tVPSSIMP\tFASTSIM\tPSNRHVS\t"
+                 "WstPsnr\tWstSsim\tWstFast\tWstHVS");
+        snprintf(results, sizeof(results),
+                 "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
+                 "%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
+                 "%7.3f\t%7.3f\t%7.3f\t%7.3f",
+                 dr, cpi->psnr.stat[ALL] / cpi->count, total_psnr,
+                 cpi->psnr.stat[ALL] / cpi->count, total_psnr, total_ssim,
+                 total_ssim, cpi->fastssim.stat[ALL] / cpi->count,
+                 cpi->psnrhvs.stat[ALL] / cpi->count, cpi->psnr.worst,
+                 cpi->worst_ssim, cpi->fastssim.worst, cpi->psnrhvs.worst);
+
+        if (cpi->b_calculate_blockiness) {
+          SNPRINT(headings, "\t  Block\tWstBlck");
+          SNPRINT2(results, "\t%7.3f", cpi->total_blockiness / cpi->count);
+          SNPRINT2(results, "\t%7.3f", cpi->worst_blockiness);
+        }
+
+        if (cpi->b_calculate_consistency) {
+          double consistency =
+              aom_sse_to_psnr((double)cpi->total_samples, peak,
+                              (double)cpi->total_inconsistency);
+
+          SNPRINT(headings, "\tConsist\tWstCons");
+          SNPRINT2(results, "\t%7.3f", consistency);
+          SNPRINT2(results, "\t%7.3f", cpi->worst_consistency);
+        }
+        fprintf(f, "%s\t    Time\tRcErr\tAbsErr\n", headings);
+        fprintf(f, "%s\t%8.0f\t%7.2f\t%7.2f\n", results, total_encode_time,
+                rate_err, fabs(rate_err));
+      }
+
+      fclose(f);
+    }
+
+#endif
+
+#if 0
+    {
+      printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000);
+      printf("\n_frames recive_data encod_mb_row compress_frame  Total\n");
+      printf("%6d %10ld %10ld %10ld %10ld\n", cpi->common.current_video_frame,
+             cpi->time_receive_data / 1000, cpi->time_encode_sb_row / 1000,
+             cpi->time_compress_data / 1000,
+             (cpi->time_receive_data + cpi->time_compress_data) / 1000);
+    }
+#endif
+  }
+
+  for (t = 0; t < cpi->num_workers; ++t) {
+    AVxWorker *const worker = &cpi->workers[t];
+    EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
+
+    // Deallocate allocated threads.
+    aom_get_worker_interface()->end(worker);
+
+    // Deallocate allocated thread data.
+    if (t < cpi->num_workers - 1) {
+#if CONFIG_PALETTE
+      if (cpi->common.allow_screen_content_tools)
+        aom_free(thread_data->td->mb.palette_buffer);
+#endif  // CONFIG_PALETTE
+      aom_free(thread_data->td->counts);
+      av1_free_pc_tree(thread_data->td);
+      av1_free_var_tree(thread_data->td);
+      aom_free(thread_data->td);
+    }
+  }
+  aom_free(cpi->tile_thr_data);
+  aom_free(cpi->workers);
+
+  if (cpi->num_workers > 1) av1_loop_filter_dealloc(&cpi->lf_row_sync);
+
+  dealloc_compressor_data(cpi);
+
+  for (i = 0; i < sizeof(cpi->mbgraph_stats) / sizeof(cpi->mbgraph_stats[0]);
+       ++i) {
+    aom_free(cpi->mbgraph_stats[i].mb_stats);
+  }
+
+#if CONFIG_FP_MB_STATS
+  if (cpi->use_fp_mb_stats) {
+    aom_free(cpi->twopass.frame_mb_stats_buf);
+    cpi->twopass.frame_mb_stats_buf = NULL;
+  }
+#endif
+#if CONFIG_INTERNAL_STATS
+  aom_free(cpi->ssim_vars);
+  cpi->ssim_vars = NULL;
+#endif  // CONFIG_INTERNAL_STATS
+
+  av1_remove_common(cm);
+  av1_free_ref_frame_buffers(cm->buffer_pool);
+  aom_free(cpi);
+
+#ifdef OUTPUT_YUV_SKINMAP
+  fclose(yuv_skinmap_file);
+#endif
+#ifdef OUTPUT_YUV_REC
+  fclose(yuv_rec_file);
+#endif
+
+#if 0
+
+  if (keyfile)
+    fclose(keyfile);
+
+  if (framepsnr)
+    fclose(framepsnr);
+
+  if (kf_list)
+    fclose(kf_list);
+
+#endif
+}
+
+static void generate_psnr_packet(AV1_COMP *cpi) {
+  struct aom_codec_cx_pkt pkt;
+  int i;
+  PSNR_STATS psnr;
+#if CONFIG_HIGHBITDEPTH
+  aom_calc_highbd_psnr(cpi->source, cpi->common.frame_to_show, &psnr,
+                       cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth);
+#else
+  aom_calc_psnr(cpi->source, cpi->common.frame_to_show, &psnr);
+#endif
+
+  for (i = 0; i < 4; ++i) {
+    pkt.data.psnr.samples[i] = psnr.samples[i];
+    pkt.data.psnr.sse[i] = psnr.sse[i];
+    pkt.data.psnr.psnr[i] = psnr.psnr[i];
+  }
+  pkt.kind = AOM_CODEC_PSNR_PKT;
+  aom_codec_pkt_list_add(cpi->output_pkt_list, &pkt);
+}
+
+int av1_use_as_reference(AV1_COMP *cpi, int ref_frame_flags) {
+  if (ref_frame_flags > ((1 << INTER_REFS_PER_FRAME) - 1)) return -1;
+
+  cpi->ref_frame_flags = ref_frame_flags;
+  return 0;
+}
+
+void av1_update_reference(AV1_COMP *cpi, int ref_frame_flags) {
+  cpi->ext_refresh_golden_frame = (ref_frame_flags & AOM_GOLD_FLAG) != 0;
+  cpi->ext_refresh_alt_ref_frame = (ref_frame_flags & AOM_ALT_FLAG) != 0;
+  cpi->ext_refresh_last_frame = (ref_frame_flags & AOM_LAST_FLAG) != 0;
+  cpi->ext_refresh_frame_flags_pending = 1;
+}
+
+static YV12_BUFFER_CONFIG *get_av1_ref_frame_buffer(
+    AV1_COMP *cpi, AOM_REFFRAME ref_frame_flag) {
+  MV_REFERENCE_FRAME ref_frame = NONE_FRAME;
+  if (ref_frame_flag == AOM_LAST_FLAG) ref_frame = LAST_FRAME;
+#if CONFIG_EXT_REFS
+  else if (ref_frame_flag == AOM_LAST2_FLAG)
+    ref_frame = LAST2_FRAME;
+  else if (ref_frame_flag == AOM_LAST3_FLAG)
+    ref_frame = LAST3_FRAME;
+#endif  // CONFIG_EXT_REFS
+  else if (ref_frame_flag == AOM_GOLD_FLAG)
+    ref_frame = GOLDEN_FRAME;
+#if CONFIG_EXT_REFS
+  else if (ref_frame_flag == AOM_BWD_FLAG)
+    ref_frame = BWDREF_FRAME;
+#endif  // CONFIG_EXT_REFS
+  else if (ref_frame_flag == AOM_ALT_FLAG)
+    ref_frame = ALTREF_FRAME;
+
+  return ref_frame == NONE_FRAME ? NULL : get_ref_frame_buffer(cpi, ref_frame);
+}
+
+int av1_copy_reference_enc(AV1_COMP *cpi, AOM_REFFRAME ref_frame_flag,
+                           YV12_BUFFER_CONFIG *sd) {
+  YV12_BUFFER_CONFIG *cfg = get_av1_ref_frame_buffer(cpi, ref_frame_flag);
+  if (cfg) {
+    aom_yv12_copy_frame(cfg, sd);
+    return 0;
+  } else {
+    return -1;
+  }
+}
+
+int av1_set_reference_enc(AV1_COMP *cpi, AOM_REFFRAME ref_frame_flag,
+                          YV12_BUFFER_CONFIG *sd) {
+  YV12_BUFFER_CONFIG *cfg = get_av1_ref_frame_buffer(cpi, ref_frame_flag);
+  if (cfg) {
+    aom_yv12_copy_frame(sd, cfg);
+    return 0;
+  } else {
+    return -1;
+  }
+}
+
+int av1_update_entropy(AV1_COMP *cpi, int update) {
+  cpi->ext_refresh_frame_context = update;
+  cpi->ext_refresh_frame_context_pending = 1;
+  return 0;
+}
+
+#if defined(OUTPUT_YUV_DENOISED) || defined(OUTPUT_YUV_SKINMAP)
+// The denoiser buffer is allocated as a YUV 440 buffer. This function writes it
+// as YUV 420. We simply use the top-left pixels of the UV buffers, since we do
+// not denoise the UV channels at this time. If ever we implement UV channel
+// denoising we will have to modify this.
+void aom_write_yuv_frame_420(YV12_BUFFER_CONFIG *s, FILE *f) {
+  uint8_t *src = s->y_buffer;
+  int h = s->y_height;
+
+  do {
+    fwrite(src, s->y_width, 1, f);
+    src += s->y_stride;
+  } while (--h);
+
+  src = s->u_buffer;
+  h = s->uv_height;
+
+  do {
+    fwrite(src, s->uv_width, 1, f);
+    src += s->uv_stride;
+  } while (--h);
+
+  src = s->v_buffer;
+  h = s->uv_height;
+
+  do {
+    fwrite(src, s->uv_width, 1, f);
+    src += s->uv_stride;
+  } while (--h);
+}
+#endif
+
+#if CONFIG_EXT_REFS && !CONFIG_XIPHRC
+static void check_show_existing_frame(AV1_COMP *cpi) {
+  const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+  AV1_COMMON *const cm = &cpi->common;
+  const FRAME_UPDATE_TYPE next_frame_update_type =
+      gf_group->update_type[gf_group->index];
+  const int which_arf = gf_group->arf_update_idx[gf_group->index];
+
+  if (cm->show_existing_frame == 1) {
+    cm->show_existing_frame = 0;
+  } else if (cpi->rc.is_last_bipred_frame) {
+    // NOTE(zoeliu): If the current frame is a last bi-predictive frame, it is
+    //               needed next to show the BWDREF_FRAME, which is pointed by
+    //               the last_fb_idxes[0] after reference frame buffer update
+    cpi->rc.is_last_bipred_frame = 0;
+    cm->show_existing_frame = 1;
+    cpi->existing_fb_idx_to_show = cpi->lst_fb_idxes[0];
+  } else if (cpi->is_arf_filter_off[which_arf] &&
+             (next_frame_update_type == OVERLAY_UPDATE ||
+              next_frame_update_type == INTNL_OVERLAY_UPDATE)) {
+    // Other parameters related to OVERLAY_UPDATE will be taken care of
+    // in av1_rc_get_second_pass_params(cpi)
+    cm->show_existing_frame = 1;
+    cpi->rc.is_src_frame_alt_ref = 1;
+    cpi->existing_fb_idx_to_show = cpi->alt_fb_idx;
+    cpi->is_arf_filter_off[which_arf] = 0;
+  }
+  cpi->rc.is_src_frame_ext_arf = 0;
+}
+#endif  // CONFIG_EXT_REFS && !CONFIG_XIPHRC
+
+#ifdef OUTPUT_YUV_REC
+void aom_write_one_yuv_frame(AV1_COMMON *cm, YV12_BUFFER_CONFIG *s) {
+  uint8_t *src = s->y_buffer;
+  int h = cm->height;
+
+#if CONFIG_HIGHBITDEPTH
+  if (s->flags & YV12_FLAG_HIGHBITDEPTH) {
+    uint16_t *src16 = CONVERT_TO_SHORTPTR(s->y_buffer);
+
+    do {
+      fwrite(src16, s->y_width, 2, yuv_rec_file);
+      src16 += s->y_stride;
+    } while (--h);
+
+    src16 = CONVERT_TO_SHORTPTR(s->u_buffer);
+    h = s->uv_height;
+
+    do {
+      fwrite(src16, s->uv_width, 2, yuv_rec_file);
+      src16 += s->uv_stride;
+    } while (--h);
+
+    src16 = CONVERT_TO_SHORTPTR(s->v_buffer);
+    h = s->uv_height;
+
+    do {
+      fwrite(src16, s->uv_width, 2, yuv_rec_file);
+      src16 += s->uv_stride;
+    } while (--h);
+
+    fflush(yuv_rec_file);
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  do {
+    fwrite(src, s->y_width, 1, yuv_rec_file);
+    src += s->y_stride;
+  } while (--h);
+
+  src = s->u_buffer;
+  h = s->uv_height;
+
+  do {
+    fwrite(src, s->uv_width, 1, yuv_rec_file);
+    src += s->uv_stride;
+  } while (--h);
+
+  src = s->v_buffer;
+  h = s->uv_height;
+
+  do {
+    fwrite(src, s->uv_width, 1, yuv_rec_file);
+    src += s->uv_stride;
+  } while (--h);
+
+  fflush(yuv_rec_file);
+}
+#endif  // OUTPUT_YUV_REC
+
+#if CONFIG_HIGHBITDEPTH
+static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
+                                                YV12_BUFFER_CONFIG *dst,
+                                                int bd) {
+#else
+static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
+                                                YV12_BUFFER_CONFIG *dst) {
+#endif  // CONFIG_HIGHBITDEPTH
+  // TODO(dkovalev): replace YV12_BUFFER_CONFIG with aom_image_t
+  int i;
+  const uint8_t *const srcs[3] = { src->y_buffer, src->u_buffer,
+                                   src->v_buffer };
+  const int src_strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
+  const int src_widths[3] = { src->y_crop_width, src->uv_crop_width,
+                              src->uv_crop_width };
+  const int src_heights[3] = { src->y_crop_height, src->uv_crop_height,
+                               src->uv_crop_height };
+  uint8_t *const dsts[3] = { dst->y_buffer, dst->u_buffer, dst->v_buffer };
+  const int dst_strides[3] = { dst->y_stride, dst->uv_stride, dst->uv_stride };
+  const int dst_widths[3] = { dst->y_crop_width, dst->uv_crop_width,
+                              dst->uv_crop_width };
+  const int dst_heights[3] = { dst->y_crop_height, dst->uv_crop_height,
+                               dst->uv_crop_height };
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+#if CONFIG_HIGHBITDEPTH
+    if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+      av1_highbd_resize_plane(srcs[i], src_heights[i], src_widths[i],
+                              src_strides[i], dsts[i], dst_heights[i],
+                              dst_widths[i], dst_strides[i], bd);
+    } else {
+      av1_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
+                       dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
+    }
+#else
+    av1_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
+                     dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
+#endif  // CONFIG_HIGHBITDEPTH
+  }
+  aom_extend_frame_borders(dst);
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                                   YV12_BUFFER_CONFIG *dst, int planes,
+                                   int bd) {
+#else
+static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                                   YV12_BUFFER_CONFIG *dst, int planes) {
+#endif  // CONFIG_HIGHBITDEPTH
+  const int src_w = src->y_crop_width;
+  const int src_h = src->y_crop_height;
+  const int dst_w = dst->y_crop_width;
+  const int dst_h = dst->y_crop_height;
+  const uint8_t *const srcs[3] = { src->y_buffer, src->u_buffer,
+                                   src->v_buffer };
+  const int src_strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
+  uint8_t *const dsts[3] = { dst->y_buffer, dst->u_buffer, dst->v_buffer };
+  const int dst_strides[3] = { dst->y_stride, dst->uv_stride, dst->uv_stride };
+  const InterpFilterParams interp_filter_params =
+      av1_get_interp_filter_params(EIGHTTAP_REGULAR);
+  const int16_t *kernel = interp_filter_params.filter_ptr;
+  const int taps = interp_filter_params.taps;
+  int x, y, i;
+
+  assert(planes <= 3);
+  for (y = 0; y < dst_h; y += 16) {
+    for (x = 0; x < dst_w; x += 16) {
+      for (i = 0; i < planes; ++i) {
+        const int factor = (i == 0 || i == 3 ? 1 : 2);
+        const int x_q4 = x * (16 / factor) * src_w / dst_w;
+        const int y_q4 = y * (16 / factor) * src_h / dst_h;
+        const int src_stride = src_strides[i];
+        const int dst_stride = dst_strides[i];
+        const uint8_t *src_ptr = srcs[i] +
+                                 (y / factor) * src_h / dst_h * src_stride +
+                                 (x / factor) * src_w / dst_w;
+        uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
+
+#if CONFIG_HIGHBITDEPTH
+        if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+          aom_highbd_convolve8(src_ptr, src_stride, dst_ptr, dst_stride,
+                               &kernel[(x_q4 & 0xf) * taps], 16 * src_w / dst_w,
+                               &kernel[(y_q4 & 0xf) * taps], 16 * src_h / dst_h,
+                               16 / factor, 16 / factor, bd);
+        } else {
+          aom_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
+                        &kernel[(x_q4 & 0xf) * taps], 16 * src_w / dst_w,
+                        &kernel[(y_q4 & 0xf) * taps], 16 * src_h / dst_h,
+                        16 / factor, 16 / factor);
+        }
+#else
+        aom_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
+                      &kernel[(x_q4 & 0xf) * taps], 16 * src_w / dst_w,
+                      &kernel[(y_q4 & 0xf) * taps], 16 * src_h / dst_h,
+                      16 / factor, 16 / factor);
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+    }
+  }
+
+  if (planes == 1)
+    aom_extend_frame_borders_y(dst);
+  else
+    aom_extend_frame_borders(dst);
+}
+
+static int scale_down(AV1_COMP *cpi, int q) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+  int scale = 0;
+  assert(frame_is_kf_gf_arf(cpi));
+
+  if (rc->frame_size_selector == UNSCALED &&
+      q >= rc->rf_level_maxq[gf_group->rf_level[gf_group->index]]) {
+    const int max_size_thresh =
+        (int)(rate_thresh_mult[SCALE_STEP1] *
+              AOMMAX(rc->this_frame_target, rc->avg_frame_bandwidth));
+    scale = rc->projected_frame_size > max_size_thresh ? 1 : 0;
+  }
+  return scale;
+}
+
+#if CONFIG_GLOBAL_MOTION
+#define GM_RECODE_LOOP_NUM4X4_FACTOR 192
+static int recode_loop_test_global_motion(AV1_COMP *cpi) {
+  int i;
+  int recode = 0;
+  RD_COUNTS *const rdc = &cpi->td.rd_counts;
+  AV1_COMMON *const cm = &cpi->common;
+  for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
+    if (cm->global_motion[i].wmtype != IDENTITY &&
+        rdc->global_motion_used[i] * GM_RECODE_LOOP_NUM4X4_FACTOR <
+            cpi->gmparams_cost[i]) {
+      set_default_warp_params(&cm->global_motion[i]);
+      cpi->gmparams_cost[i] = 0;
+#if CONFIG_REF_MV
+      recode = 1;
+#else
+      recode |= (rdc->global_motion_used[i] > 0);
+#endif
+    }
+  }
+  return recode;
+}
+#endif  // CONFIG_GLOBAL_MOTION
+
+// Function to test for conditions that indicate we should loop
+// back and recode a frame.
+static int recode_loop_test(AV1_COMP *cpi, int high_limit, int low_limit, int q,
+                            int maxq, int minq) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  const int frame_is_kfgfarf = frame_is_kf_gf_arf(cpi);
+  int force_recode = 0;
+
+  if ((rc->projected_frame_size >= rc->max_frame_bandwidth) ||
+      (cpi->sf.recode_loop == ALLOW_RECODE) ||
+      (frame_is_kfgfarf && (cpi->sf.recode_loop == ALLOW_RECODE_KFARFGF))) {
+    if (frame_is_kfgfarf && (oxcf->resize_mode == RESIZE_DYNAMIC) &&
+        scale_down(cpi, q)) {
+      // Code this group at a lower resolution.
+      cpi->resize_pending = 1;
+      return 1;
+    }
+
+    // TODO(agrange) high_limit could be greater than the scale-down threshold.
+    if ((rc->projected_frame_size > high_limit && q < maxq) ||
+        (rc->projected_frame_size < low_limit && q > minq)) {
+      force_recode = 1;
+    } else if (cpi->oxcf.rc_mode == AOM_CQ) {
+      // Deal with frame undershoot and whether or not we are
+      // below the automatically set cq level.
+      if (q > oxcf->cq_level &&
+          rc->projected_frame_size < ((rc->this_frame_target * 7) >> 3)) {
+        force_recode = 1;
+      }
+    }
+  }
+  return force_recode;
+}
+
+static INLINE int get_free_upsampled_ref_buf(EncRefCntBuffer *ubufs) {
+  int i;
+
+  for (i = 0; i < (REF_FRAMES + 1); i++) {
+    if (!ubufs[i].ref_count) {
+      return i;
+    }
+  }
+  return INVALID_IDX;
+}
+
+// Up-sample 1 reference frame.
+static INLINE int upsample_ref_frame(AV1_COMP *cpi,
+                                     const YV12_BUFFER_CONFIG *const ref) {
+  AV1_COMMON *const cm = &cpi->common;
+  EncRefCntBuffer *ubufs = cpi->upsampled_ref_bufs;
+  int new_uidx = get_free_upsampled_ref_buf(ubufs);
+
+  if (new_uidx == INVALID_IDX) {
+    return INVALID_IDX;
+  } else {
+    YV12_BUFFER_CONFIG *upsampled_ref = &ubufs[new_uidx].buf;
+
+    // Can allocate buffer for Y plane only.
+    if (upsampled_ref->buffer_alloc_sz < (ref->buffer_alloc_sz << 6))
+      if (aom_realloc_frame_buffer(upsampled_ref, (cm->width << 3),
+                                   (cm->height << 3), cm->subsampling_x,
+                                   cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                   cm->use_highbitdepth,
+#endif
+                                   (AOM_BORDER_IN_PIXELS << 3),
+                                   cm->byte_alignment, NULL, NULL, NULL))
+        aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                           "Failed to allocate up-sampled frame buffer");
+
+// Currently, only Y plane is up-sampled, U, V are not used.
+#if CONFIG_HIGHBITDEPTH
+    scale_and_extend_frame(ref, upsampled_ref, 1, (int)cm->bit_depth);
+#else
+    scale_and_extend_frame(ref, upsampled_ref, 1);
+#endif
+    return new_uidx;
+  }
+}
+
+#define DUMP_REF_FRAME_IMAGES 0
+
+#if DUMP_REF_FRAME_IMAGES == 1
+static int dump_one_image(AV1_COMMON *cm,
+                          const YV12_BUFFER_CONFIG *const ref_buf,
+                          char *file_name) {
+  int h;
+  FILE *f_ref = NULL;
+
+  if (ref_buf == NULL) {
+    printf("Frame data buffer is NULL.\n");
+    return AOM_CODEC_MEM_ERROR;
+  }
+
+  if ((f_ref = fopen(file_name, "wb")) == NULL) {
+    printf("Unable to open file %s to write.\n", file_name);
+    return AOM_CODEC_MEM_ERROR;
+  }
+
+  // --- Y ---
+  for (h = 0; h < cm->height; ++h) {
+    fwrite(&ref_buf->y_buffer[h * ref_buf->y_stride], 1, cm->width, f_ref);
+  }
+  // --- U ---
+  for (h = 0; h < (cm->height >> 1); ++h) {
+    fwrite(&ref_buf->u_buffer[h * ref_buf->uv_stride], 1, (cm->width >> 1),
+           f_ref);
+  }
+  // --- V ---
+  for (h = 0; h < (cm->height >> 1); ++h) {
+    fwrite(&ref_buf->v_buffer[h * ref_buf->uv_stride], 1, (cm->width >> 1),
+           f_ref);
+  }
+
+  fclose(f_ref);
+
+  return AOM_CODEC_OK;
+}
+
+static void dump_ref_frame_images(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  MV_REFERENCE_FRAME ref_frame;
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    char file_name[256] = "";
+    snprintf(file_name, sizeof(file_name), "/tmp/enc_F%d_ref_%d.yuv",
+             cm->current_video_frame, ref_frame);
+    dump_one_image(cm, get_ref_frame_buffer(cpi, ref_frame), file_name);
+  }
+}
+#endif  // DUMP_REF_FRAME_IMAGES == 1
+
+#if CONFIG_EXT_REFS
+// This function is used to shift the virtual indices of last reference frames
+// as follows:
+// LAST_FRAME -> LAST2_FRAME -> LAST3_FRAME
+// when the LAST_FRAME is updated.
+static INLINE void shift_last_ref_frames(AV1_COMP *cpi) {
+  int ref_frame;
+  for (ref_frame = LAST_REF_FRAMES - 1; ref_frame > 0; --ref_frame) {
+    cpi->lst_fb_idxes[ref_frame] = cpi->lst_fb_idxes[ref_frame - 1];
+
+    // [0] is allocated to the current coded frame. The statistics for the
+    // reference frames start at [LAST_FRAME], i.e. [1].
+    if (!cpi->rc.is_src_frame_alt_ref) {
+      memcpy(cpi->interp_filter_selected[ref_frame + LAST_FRAME],
+             cpi->interp_filter_selected[ref_frame - 1 + LAST_FRAME],
+             sizeof(cpi->interp_filter_selected[ref_frame - 1 + LAST_FRAME]));
+    }
+  }
+}
+#endif  // CONFIG_EXT_REFS
+
+void av1_update_reference_frames(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  const int use_upsampled_ref = cpi->sf.use_upsampled_references;
+  int new_uidx = 0;
+
+  // NOTE: Save the new show frame buffer index for --test-code=warn, i.e.,
+  //       for the purpose to verify no mismatch between encoder and decoder.
+  if (cm->show_frame) cpi->last_show_frame_buf_idx = cm->new_fb_idx;
+
+  if (use_upsampled_ref) {
+#if CONFIG_EXT_REFS
+    if (cm->show_existing_frame) {
+      new_uidx = cpi->upsampled_ref_idx[cpi->existing_fb_idx_to_show];
+      // TODO(zoeliu): Once following is confirmed, remove it.
+      assert(cpi->upsampled_ref_bufs[new_uidx].ref_count > 0);
+    } else {
+#endif  // CONFIG_EXT_REFS
+      // Up-sample the current encoded frame.
+      RefCntBuffer *bufs = pool->frame_bufs;
+      const YV12_BUFFER_CONFIG *const ref = &bufs[cm->new_fb_idx].buf;
+
+      new_uidx = upsample_ref_frame(cpi, ref);
+#if CONFIG_EXT_REFS
+      assert(new_uidx != INVALID_IDX);
+    }
+#endif  // CONFIG_EXT_REFS
+  }
+  // At this point the new frame has been encoded.
+  // If any buffer copy / swapping is signaled it should be done here.
+  if (cm->frame_type == KEY_FRAME) {
+    ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
+               cm->new_fb_idx);
+#if CONFIG_EXT_REFS
+    ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->bwd_fb_idx],
+               cm->new_fb_idx);
+#endif  // CONFIG_EXT_REFS
+    ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->alt_fb_idx],
+               cm->new_fb_idx);
+
+    if (use_upsampled_ref) {
+      uref_cnt_fb(cpi->upsampled_ref_bufs,
+                  &cpi->upsampled_ref_idx[cpi->gld_fb_idx], new_uidx);
+#if CONFIG_EXT_REFS
+      uref_cnt_fb(cpi->upsampled_ref_bufs,
+                  &cpi->upsampled_ref_idx[cpi->bwd_fb_idx], new_uidx);
+#endif  // CONFIG_EXT_REFS
+      uref_cnt_fb(cpi->upsampled_ref_bufs,
+                  &cpi->upsampled_ref_idx[cpi->alt_fb_idx], new_uidx);
+    }
+  } else if (av1_preserve_existing_gf(cpi)) {
+    // We have decided to preserve the previously existing golden frame as our
+    // new ARF frame. However, in the short term in function
+    // av1_bitstream.c::get_refresh_mask() we left it in the GF slot and, if
+    // we're updating the GF with the current decoded frame, we save it to the
+    // ARF slot instead.
+    // We now have to update the ARF with the current frame and swap gld_fb_idx
+    // and alt_fb_idx so that, overall, we've stored the old GF in the new ARF
+    // slot and, if we're updating the GF, the current frame becomes the new GF.
+    int tmp;
+
+    ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->alt_fb_idx],
+               cm->new_fb_idx);
+    if (use_upsampled_ref)
+      uref_cnt_fb(cpi->upsampled_ref_bufs,
+                  &cpi->upsampled_ref_idx[cpi->alt_fb_idx], new_uidx);
+
+    tmp = cpi->alt_fb_idx;
+    cpi->alt_fb_idx = cpi->gld_fb_idx;
+    cpi->gld_fb_idx = tmp;
+
+#if CONFIG_EXT_REFS
+    // We need to modify the mapping accordingly
+    cpi->arf_map[0] = cpi->alt_fb_idx;
+#endif
+// TODO(zoeliu): Do we need to copy cpi->interp_filter_selected[0] over to
+// cpi->interp_filter_selected[GOLDEN_FRAME]?
+#if CONFIG_EXT_REFS
+  } else if (cpi->rc.is_last_bipred_frame) {
+    // Refresh the LAST_FRAME with the BWDREF_FRAME and retire the LAST3_FRAME
+    // by updating the virtual indices. Note that the frame BWDREF_FRAME points
+    // to now should be retired, and it should not be used before refreshed.
+    int tmp = cpi->lst_fb_idxes[LAST_REF_FRAMES - 1];
+
+    shift_last_ref_frames(cpi);
+    cpi->lst_fb_idxes[0] = cpi->bwd_fb_idx;
+    cpi->bwd_fb_idx = tmp;
+
+    memcpy(cpi->interp_filter_selected[LAST_FRAME],
+           cpi->interp_filter_selected[BWDREF_FRAME],
+           sizeof(cpi->interp_filter_selected[BWDREF_FRAME]));
+  } else if (cpi->rc.is_src_frame_ext_arf && cm->show_existing_frame) {
+    // Deal with the special case for showing existing internal ALTREF_FRAME
+    // Refresh the LAST_FRAME with the ALTREF_FRAME and retire the LAST3_FRAME
+    // by updating the virtual indices.
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    int which_arf = gf_group->arf_ref_idx[gf_group->index];
+    int tmp = cpi->lst_fb_idxes[LAST_REF_FRAMES - 1];
+
+    shift_last_ref_frames(cpi);
+    cpi->lst_fb_idxes[0] = cpi->alt_fb_idx;
+    cpi->alt_fb_idx = tmp;
+
+    // We need to modify the mapping accordingly
+    cpi->arf_map[which_arf] = cpi->alt_fb_idx;
+
+    memcpy(cpi->interp_filter_selected[LAST_FRAME],
+           cpi->interp_filter_selected[ALTREF_FRAME + which_arf],
+           sizeof(cpi->interp_filter_selected[ALTREF_FRAME + which_arf]));
+#endif     // CONFIG_EXT_REFS
+  } else { /* For non key/golden frames */
+    if (cpi->refresh_alt_ref_frame) {
+      int arf_idx = cpi->alt_fb_idx;
+      int which_arf = 0;
+#if CONFIG_EXT_REFS
+      if (cpi->oxcf.pass == 2) {
+        const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+        which_arf = gf_group->arf_update_idx[gf_group->index];
+        arf_idx = cpi->arf_map[which_arf];
+      }
+#else
+      if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
+        const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+        arf_idx = gf_group->arf_update_idx[gf_group->index];
+      }
+#endif  // CONFIG_EXT_REFS
+      ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
+      if (use_upsampled_ref)
+        uref_cnt_fb(cpi->upsampled_ref_bufs, &cpi->upsampled_ref_idx[arf_idx],
+                    new_uidx);
+
+      memcpy(cpi->interp_filter_selected[ALTREF_FRAME + which_arf],
+             cpi->interp_filter_selected[0],
+             sizeof(cpi->interp_filter_selected[0]));
+    }
+
+    if (cpi->refresh_golden_frame) {
+      ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
+                 cm->new_fb_idx);
+      if (use_upsampled_ref)
+        uref_cnt_fb(cpi->upsampled_ref_bufs,
+                    &cpi->upsampled_ref_idx[cpi->gld_fb_idx], new_uidx);
+
+#if !CONFIG_EXT_REFS
+      if (!cpi->rc.is_src_frame_alt_ref)
+#endif  // !CONFIG_EXT_REFS
+        memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
+               cpi->interp_filter_selected[0],
+               sizeof(cpi->interp_filter_selected[0]));
+    }
+
+#if CONFIG_EXT_REFS
+    if (cpi->refresh_bwd_ref_frame) {
+      if (cpi->rc.is_bwd_ref_frame && cpi->num_extra_arfs) {
+        // We have swapped the virtual indices to allow bwd_ref_frame to use
+        // ALT0 as reference frame. We need to swap them back.
+        // NOTE: The ALT_REFs' are indexed reversely, and ALT0 refers to the
+        //       farthest ALT_REF from the first frame in the gf group.
+        int tmp = cpi->arf_map[0];
+        cpi->arf_map[0] = cpi->alt_fb_idx;
+        cpi->alt_fb_idx = cpi->bwd_fb_idx;
+        cpi->bwd_fb_idx = tmp;
+      }
+
+      ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->bwd_fb_idx],
+                 cm->new_fb_idx);
+      if (use_upsampled_ref)
+        uref_cnt_fb(cpi->upsampled_ref_bufs,
+                    &cpi->upsampled_ref_idx[cpi->bwd_fb_idx], new_uidx);
+
+      memcpy(cpi->interp_filter_selected[BWDREF_FRAME],
+             cpi->interp_filter_selected[0],
+             sizeof(cpi->interp_filter_selected[0]));
+    }
+#endif  // CONFIG_EXT_REFS
+  }
+
+  if (cpi->refresh_last_frame) {
+#if CONFIG_EXT_REFS
+    // NOTE(zoeliu): We have two layers of mapping (1) from the per-frame
+    // reference to the reference frame buffer virtual index; and then (2) from
+    // the virtual index to the reference frame buffer physical index:
+    //
+    // LAST_FRAME,      ..., LAST3_FRAME,     ..., ALTREF_FRAME
+    //      |                     |                     |
+    //      v                     v                     v
+    // lst_fb_idxes[0], ..., lst_fb_idxes[2], ..., alt_fb_idx
+    //      |                     |                     |
+    //      v                     v                     v
+    // ref_frame_map[], ..., ref_frame_map[], ..., ref_frame_map[]
+    //
+    // When refresh_last_frame is set, it is intended to retire LAST3_FRAME,
+    // have the other 2 LAST reference frames shifted as follows:
+    // LAST_FRAME -> LAST2_FRAME -> LAST3_FRAME
+    // , and then have LAST_FRAME refreshed by the newly coded frame.
+    //
+    // To fulfill it, the decoder will be notified to execute following 2 steps:
+    //
+    // (a) To change ref_frame_map[] and have the virtual index of LAST3_FRAME
+    //     to point to the newly coded frame, i.e.
+    //     ref_frame_map[lst_fb_idexes[2]] => new_fb_idx;
+    //
+    // (b) To change the 1st layer mapping to have LAST_FRAME mapped to the
+    //     original virtual index of LAST3_FRAME and have the other mappings
+    //     shifted as follows:
+    // LAST_FRAME,      LAST2_FRAME,     LAST3_FRAME
+    //      |                |                |
+    //      v                v                v
+    // lst_fb_idxes[2], lst_fb_idxes[0], lst_fb_idxes[1]
+    int ref_frame;
+
+    if (cpi->rc.is_bwd_ref_frame && cpi->num_extra_arfs) {
+      // We have swapped the virtual indices to use ALT0 as BWD_REF
+      // and we need to swap them back.
+      int tmp = cpi->arf_map[0];
+      cpi->arf_map[0] = cpi->alt_fb_idx;
+      cpi->alt_fb_idx = cpi->bwd_fb_idx;
+      cpi->bwd_fb_idx = tmp;
+    }
+
+    if (cm->frame_type == KEY_FRAME) {
+      for (ref_frame = 0; ref_frame < LAST_REF_FRAMES; ++ref_frame) {
+        ref_cnt_fb(pool->frame_bufs,
+                   &cm->ref_frame_map[cpi->lst_fb_idxes[ref_frame]],
+                   cm->new_fb_idx);
+
+        if (use_upsampled_ref)
+          uref_cnt_fb(cpi->upsampled_ref_bufs,
+                      &cpi->upsampled_ref_idx[cpi->lst_fb_idxes[ref_frame]],
+                      new_uidx);
+      }
+    } else {
+      int tmp;
+
+      ref_cnt_fb(pool->frame_bufs,
+                 &cm->ref_frame_map[cpi->lst_fb_idxes[LAST_REF_FRAMES - 1]],
+                 cm->new_fb_idx);
+
+      if (use_upsampled_ref)
+        uref_cnt_fb(
+            cpi->upsampled_ref_bufs,
+            &cpi->upsampled_ref_idx[cpi->lst_fb_idxes[LAST_REF_FRAMES - 1]],
+            new_uidx);
+
+      tmp = cpi->lst_fb_idxes[LAST_REF_FRAMES - 1];
+
+      shift_last_ref_frames(cpi);
+      cpi->lst_fb_idxes[0] = tmp;
+
+      assert(cm->show_existing_frame == 0);
+      // NOTE: Currently only LF_UPDATE and INTNL_OVERLAY_UPDATE frames are to
+      //       refresh the LAST_FRAME.
+      memcpy(cpi->interp_filter_selected[LAST_FRAME],
+             cpi->interp_filter_selected[0],
+             sizeof(cpi->interp_filter_selected[0]));
+    }
+#else
+    ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->lst_fb_idx],
+               cm->new_fb_idx);
+    if (use_upsampled_ref)
+      uref_cnt_fb(cpi->upsampled_ref_bufs,
+                  &cpi->upsampled_ref_idx[cpi->lst_fb_idx], new_uidx);
+    if (!cpi->rc.is_src_frame_alt_ref) {
+      memcpy(cpi->interp_filter_selected[LAST_FRAME],
+             cpi->interp_filter_selected[0],
+             sizeof(cpi->interp_filter_selected[0]));
+    }
+#endif  // CONFIG_EXT_REFS
+  }
+
+#if DUMP_REF_FRAME_IMAGES == 1
+  // Dump out all reference frame images.
+  dump_ref_frame_images(cpi);
+#endif  // DUMP_REF_FRAME_IMAGES
+}
+
+static void loopfilter_frame(AV1_COMP *cpi, AV1_COMMON *cm) {
+  MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
+  struct loopfilter *lf = &cm->lf;
+  if (is_lossless_requested(&cpi->oxcf)) {
+    lf->filter_level = 0;
+  } else {
+    struct aom_usec_timer timer;
+
+    aom_clear_system_state();
+
+    aom_usec_timer_start(&timer);
+
+    av1_pick_filter_level(cpi->source, cpi, cpi->sf.lpf_pick);
+
+    aom_usec_timer_mark(&timer);
+    cpi->time_pick_lpf += aom_usec_timer_elapsed(&timer);
+  }
+
+  if (lf->filter_level > 0) {
+#if CONFIG_VAR_TX || CONFIG_EXT_PARTITION || CONFIG_CB4X4
+    av1_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
+#else
+    if (cpi->num_workers > 1)
+      av1_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane,
+                               lf->filter_level, 0, 0, cpi->workers,
+                               cpi->num_workers, &cpi->lf_row_sync);
+    else
+      av1_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
+#endif
+  }
+#if CONFIG_CDEF
+  if (is_lossless_requested(&cpi->oxcf)) {
+    cm->cdef_bits = 0;
+    cm->cdef_strengths[0] = 0;
+    cm->nb_cdef_strengths = 1;
+  } else {
+    // Find cm->dering_level, cm->clpf_strength_u and cm->clpf_strength_v
+    av1_cdef_search(cm->frame_to_show, cpi->source, cm, xd);
+
+    // Apply the filter
+    av1_cdef_frame(cm->frame_to_show, cm, xd);
+  }
+#endif
+#if CONFIG_LOOP_RESTORATION
+  av1_pick_filter_restoration(cpi->source, cpi, cpi->sf.lpf_pick);
+  if (cm->rst_info[0].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[1].frame_restoration_type != RESTORE_NONE ||
+      cm->rst_info[2].frame_restoration_type != RESTORE_NONE) {
+    av1_loop_restoration_frame(cm->frame_to_show, cm, cm->rst_info, 7, 0, NULL);
+  }
+#endif  // CONFIG_LOOP_RESTORATION
+  aom_extend_frame_inner_borders(cm->frame_to_show);
+}
+
+static INLINE void alloc_frame_mvs(AV1_COMMON *const cm, int buffer_idx) {
+  RefCntBuffer *const new_fb_ptr = &cm->buffer_pool->frame_bufs[buffer_idx];
+  if (new_fb_ptr->mvs == NULL || new_fb_ptr->mi_rows < cm->mi_rows ||
+      new_fb_ptr->mi_cols < cm->mi_cols) {
+    aom_free(new_fb_ptr->mvs);
+    CHECK_MEM_ERROR(cm, new_fb_ptr->mvs,
+                    (MV_REF *)aom_calloc(cm->mi_rows * cm->mi_cols,
+                                         sizeof(*new_fb_ptr->mvs)));
+    new_fb_ptr->mi_rows = cm->mi_rows;
+    new_fb_ptr->mi_cols = cm->mi_cols;
+  }
+}
+
+void av1_scale_references(AV1_COMP *cpi) {
+  AV1_COMMON *cm = &cpi->common;
+  MV_REFERENCE_FRAME ref_frame;
+  const AOM_REFFRAME ref_mask[INTER_REFS_PER_FRAME] = {
+    AOM_LAST_FLAG,
+#if CONFIG_EXT_REFS
+    AOM_LAST2_FLAG,
+    AOM_LAST3_FLAG,
+#endif  // CONFIG_EXT_REFS
+    AOM_GOLD_FLAG,
+#if CONFIG_EXT_REFS
+    AOM_BWD_FLAG,
+#endif  // CONFIG_EXT_REFS
+    AOM_ALT_FLAG
+  };
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    // Need to convert from AOM_REFFRAME to index into ref_mask (subtract 1).
+    if (cpi->ref_frame_flags & ref_mask[ref_frame - 1]) {
+      BufferPool *const pool = cm->buffer_pool;
+      const YV12_BUFFER_CONFIG *const ref =
+          get_ref_frame_buffer(cpi, ref_frame);
+
+      if (ref == NULL) {
+        cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
+        continue;
+      }
+
+#if CONFIG_HIGHBITDEPTH
+      if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
+        RefCntBuffer *new_fb_ptr = NULL;
+        int force_scaling = 0;
+        int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
+        if (new_fb == INVALID_IDX) {
+          new_fb = get_free_fb(cm);
+          force_scaling = 1;
+        }
+        if (new_fb == INVALID_IDX) return;
+        new_fb_ptr = &pool->frame_bufs[new_fb];
+        if (force_scaling || new_fb_ptr->buf.y_crop_width != cm->width ||
+            new_fb_ptr->buf.y_crop_height != cm->height) {
+          if (aom_realloc_frame_buffer(
+                  &new_fb_ptr->buf, cm->width, cm->height, cm->subsampling_x,
+                  cm->subsampling_y, cm->use_highbitdepth, AOM_BORDER_IN_PIXELS,
+                  cm->byte_alignment, NULL, NULL, NULL))
+            aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                               "Failed to allocate frame buffer");
+          scale_and_extend_frame(ref, &new_fb_ptr->buf, MAX_MB_PLANE,
+                                 (int)cm->bit_depth);
+          cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
+          alloc_frame_mvs(cm, new_fb);
+        }
+#else
+      if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
+        RefCntBuffer *new_fb_ptr = NULL;
+        int force_scaling = 0;
+        int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
+        if (new_fb == INVALID_IDX) {
+          new_fb = get_free_fb(cm);
+          force_scaling = 1;
+        }
+        if (new_fb == INVALID_IDX) return;
+        new_fb_ptr = &pool->frame_bufs[new_fb];
+        if (force_scaling || new_fb_ptr->buf.y_crop_width != cm->width ||
+            new_fb_ptr->buf.y_crop_height != cm->height) {
+          if (aom_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
+                                       cm->subsampling_x, cm->subsampling_y,
+                                       AOM_BORDER_IN_PIXELS, cm->byte_alignment,
+                                       NULL, NULL, NULL))
+            aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                               "Failed to allocate frame buffer");
+          scale_and_extend_frame(ref, &new_fb_ptr->buf, MAX_MB_PLANE);
+          cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
+          alloc_frame_mvs(cm, new_fb);
+        }
+#endif  // CONFIG_HIGHBITDEPTH
+
+        if (cpi->sf.use_upsampled_references &&
+            (force_scaling || new_fb_ptr->buf.y_crop_width != cm->width ||
+             new_fb_ptr->buf.y_crop_height != cm->height)) {
+          const int map_idx = get_ref_frame_map_idx(cpi, ref_frame);
+          EncRefCntBuffer *ubuf =
+              &cpi->upsampled_ref_bufs[cpi->upsampled_ref_idx[map_idx]];
+
+          if (aom_realloc_frame_buffer(&ubuf->buf, (cm->width << 3),
+                                       (cm->height << 3), cm->subsampling_x,
+                                       cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                       cm->use_highbitdepth,
+#endif
+                                       (AOM_BORDER_IN_PIXELS << 3),
+                                       cm->byte_alignment, NULL, NULL, NULL))
+            aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                               "Failed to allocate up-sampled frame buffer");
+#if CONFIG_HIGHBITDEPTH
+          scale_and_extend_frame(&new_fb_ptr->buf, &ubuf->buf, 1,
+                                 (int)cm->bit_depth);
+#else
+          scale_and_extend_frame(&new_fb_ptr->buf, &ubuf->buf, 1);
+#endif
+        }
+      } else {
+        const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+        RefCntBuffer *const buf = &pool->frame_bufs[buf_idx];
+        buf->buf.y_crop_width = ref->y_crop_width;
+        buf->buf.y_crop_height = ref->y_crop_height;
+        cpi->scaled_ref_idx[ref_frame - 1] = buf_idx;
+        ++buf->ref_count;
+      }
+    } else {
+      if (cpi->oxcf.pass != 0) cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
+    }
+  }
+}
+
+static void release_scaled_references(AV1_COMP *cpi) {
+  AV1_COMMON *cm = &cpi->common;
+  int i;
+  if (cpi->oxcf.pass == 0) {
+    // Only release scaled references under certain conditions:
+    // if reference will be updated, or if scaled reference has same resolution.
+    int refresh[INTER_REFS_PER_FRAME];
+    refresh[0] = (cpi->refresh_last_frame) ? 1 : 0;
+#if CONFIG_EXT_REFS
+    refresh[1] = refresh[2] = 0;
+    refresh[3] = (cpi->refresh_golden_frame) ? 1 : 0;
+    refresh[4] = (cpi->refresh_bwd_ref_frame) ? 1 : 0;
+    refresh[5] = (cpi->refresh_alt_ref_frame) ? 1 : 0;
+#else
+    refresh[1] = (cpi->refresh_golden_frame) ? 1 : 0;
+    refresh[2] = (cpi->refresh_alt_ref_frame) ? 1 : 0;
+#endif  // CONFIG_EXT_REFS
+    for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
+      const int idx = cpi->scaled_ref_idx[i - 1];
+      RefCntBuffer *const buf =
+          idx != INVALID_IDX ? &cm->buffer_pool->frame_bufs[idx] : NULL;
+      const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, i);
+      if (buf != NULL &&
+          (refresh[i - 1] || (buf->buf.y_crop_width == ref->y_crop_width &&
+                              buf->buf.y_crop_height == ref->y_crop_height))) {
+        --buf->ref_count;
+        cpi->scaled_ref_idx[i - 1] = INVALID_IDX;
+      }
+    }
+  } else {
+    for (i = 0; i < TOTAL_REFS_PER_FRAME; ++i) {
+      const int idx = cpi->scaled_ref_idx[i];
+      RefCntBuffer *const buf =
+          idx != INVALID_IDX ? &cm->buffer_pool->frame_bufs[idx] : NULL;
+      if (buf != NULL) {
+        --buf->ref_count;
+        cpi->scaled_ref_idx[i] = INVALID_IDX;
+      }
+    }
+  }
+}
+
+static void full_to_model_count(unsigned int *model_count,
+                                unsigned int *full_count) {
+  int n;
+  model_count[ZERO_TOKEN] = full_count[ZERO_TOKEN];
+  model_count[ONE_TOKEN] = full_count[ONE_TOKEN];
+  model_count[TWO_TOKEN] = full_count[TWO_TOKEN];
+  for (n = THREE_TOKEN; n < EOB_TOKEN; ++n)
+    model_count[TWO_TOKEN] += full_count[n];
+  model_count[EOB_MODEL_TOKEN] = full_count[EOB_TOKEN];
+}
+
+void av1_full_to_model_counts(av1_coeff_count_model *model_count,
+                              av1_coeff_count *full_count) {
+  int i, j, k, l;
+
+  for (i = 0; i < PLANE_TYPES; ++i)
+    for (j = 0; j < REF_TYPES; ++j)
+      for (k = 0; k < COEF_BANDS; ++k)
+        for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
+          full_to_model_count(model_count[i][j][k][l], full_count[i][j][k][l]);
+}
+
+#if 0 && CONFIG_INTERNAL_STATS
+static void output_frame_level_debug_stats(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  FILE *const f = fopen("tmp.stt", cm->current_video_frame ? "a" : "w");
+  int64_t recon_err;
+
+  aom_clear_system_state();
+
+  recon_err = aom_get_y_sse(cpi->source, get_frame_new_buffer(cm));
+
+  if (cpi->twopass.total_left_stats.coded_error != 0.0)
+    fprintf(f, "%10u %dx%d %d %d %10d %10d %10d %10d"
+       "%10"PRId64" %10"PRId64" %5d %5d %10"PRId64" "
+       "%10"PRId64" %10"PRId64" %10d "
+       "%7.2lf %7.2lf %7.2lf %7.2lf %7.2lf"
+        "%6d %6d %5d %5d %5d "
+        "%10"PRId64" %10.3lf"
+        "%10lf %8u %10"PRId64" %10d %10d %10d\n",
+        cpi->common.current_video_frame,
+        cm->width, cm->height,
+        cpi->rc.source_alt_ref_pending,
+        cpi->rc.source_alt_ref_active,
+        cpi->rc.this_frame_target,
+        cpi->rc.projected_frame_size,
+        cpi->rc.projected_frame_size / cpi->common.MBs,
+        (cpi->rc.projected_frame_size - cpi->rc.this_frame_target),
+        cpi->rc.vbr_bits_off_target,
+        cpi->rc.vbr_bits_off_target_fast,
+        cpi->twopass.extend_minq,
+        cpi->twopass.extend_minq_fast,
+        cpi->rc.total_target_vs_actual,
+        (cpi->rc.starting_buffer_level - cpi->rc.bits_off_target),
+        cpi->rc.total_actual_bits, cm->base_qindex,
+        av1_convert_qindex_to_q(cm->base_qindex, cm->bit_depth),
+        (double)av1_dc_quant(cm->base_qindex, 0, cm->bit_depth) / 4.0,
+        av1_convert_qindex_to_q(cpi->twopass.active_worst_quality,
+                                cm->bit_depth),
+        cpi->rc.avg_q,
+        av1_convert_qindex_to_q(cpi->oxcf.cq_level, cm->bit_depth),
+        cpi->refresh_last_frame, cpi->refresh_golden_frame,
+        cpi->refresh_alt_ref_frame, cm->frame_type, cpi->rc.gfu_boost,
+        cpi->twopass.bits_left,
+        cpi->twopass.total_left_stats.coded_error,
+        cpi->twopass.bits_left /
+            (1 + cpi->twopass.total_left_stats.coded_error),
+        cpi->tot_recode_hits, recon_err, cpi->rc.kf_boost,
+        cpi->twopass.kf_zeromotion_pct,
+        cpi->twopass.fr_content_type);
+
+  fclose(f);
+
+  if (0) {
+    FILE *const fmodes = fopen("Modes.stt", "a");
+    int i;
+
+    fprintf(fmodes, "%6d:%1d:%1d:%1d ", cpi->common.current_video_frame,
+            cm->frame_type, cpi->refresh_golden_frame,
+            cpi->refresh_alt_ref_frame);
+
+    for (i = 0; i < MAX_MODES; ++i)
+      fprintf(fmodes, "%5d ", cpi->mode_chosen_counts[i]);
+
+    fprintf(fmodes, "\n");
+
+    fclose(fmodes);
+  }
+}
+#endif
+
+static void set_mv_search_params(AV1_COMP *cpi) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const unsigned int max_mv_def = AOMMIN(cm->width, cm->height);
+
+  // Default based on max resolution.
+  cpi->mv_step_param = av1_init_search_range(max_mv_def);
+
+  if (cpi->sf.mv.auto_mv_step_size) {
+    if (frame_is_intra_only(cm)) {
+      // Initialize max_mv_magnitude for use in the first INTER frame
+      // after a key/intra-only frame.
+      cpi->max_mv_magnitude = max_mv_def;
+    } else {
+      if (cm->show_frame) {
+        // Allow mv_steps to correspond to twice the max mv magnitude found
+        // in the previous frame, capped by the default max_mv_magnitude based
+        // on resolution.
+        cpi->mv_step_param = av1_init_search_range(
+            AOMMIN(max_mv_def, 2 * cpi->max_mv_magnitude));
+      }
+      cpi->max_mv_magnitude = 0;
+    }
+  }
+}
+
+static void set_size_independent_vars(AV1_COMP *cpi) {
+#if CONFIG_GLOBAL_MOTION
+  int i;
+  for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
+    set_default_warp_params(&cpi->common.global_motion[i]);
+  }
+  cpi->global_motion_search_done = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+  av1_set_speed_features_framesize_independent(cpi);
+  av1_set_rd_speed_thresholds(cpi);
+  av1_set_rd_speed_thresholds_sub8x8(cpi);
+  cpi->common.interp_filter = cpi->sf.default_interp_filter;
+}
+
+static void set_size_dependent_vars(AV1_COMP *cpi, int *q, int *bottom_index,
+                                    int *top_index) {
+  AV1_COMMON *const cm = &cpi->common;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+
+  // Setup variables that depend on the dimensions of the frame.
+  av1_set_speed_features_framesize_dependent(cpi);
+
+// Decide q and q bounds.
+#if CONFIG_XIPHRC
+  int frame_type = cm->frame_type == KEY_FRAME ? OD_I_FRAME : OD_P_FRAME;
+  *q = od_enc_rc_select_quantizers_and_lambdas(
+      &cpi->od_rc, cpi->refresh_golden_frame, cpi->refresh_alt_ref_frame,
+      frame_type, bottom_index, top_index);
+#else
+  *q = av1_rc_pick_q_and_bounds(cpi, bottom_index, top_index);
+#endif
+
+  if (!frame_is_intra_only(cm)) {
+    av1_set_high_precision_mv(cpi, (*q) < HIGH_PRECISION_MV_QTHRESH);
+  }
+
+  // Configure experimental use of segmentation for enhanced coding of
+  // static regions if indicated.
+  // Only allowed in the second pass of a two pass encode, as it requires
+  // lagged coding, and if the relevant speed feature flag is set.
+  if (oxcf->pass == 2 && cpi->sf.static_segmentation)
+    configure_static_seg_features(cpi);
+}
+
+static void init_motion_estimation(AV1_COMP *cpi) {
+  int y_stride = cpi->scaled_source.y_stride;
+
+  if (cpi->sf.mv.search_method == NSTEP) {
+    av1_init3smotion_compensation(&cpi->ss_cfg, y_stride);
+  } else if (cpi->sf.mv.search_method == DIAMOND) {
+    av1_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
+  }
+}
+
+#if CONFIG_LOOP_RESTORATION
+static void set_restoration_tilesize(int width, int height,
+                                     RestorationInfo *rst) {
+  (void)width;
+  (void)height;
+  rst[0].restoration_tilesize = (RESTORATION_TILESIZE_MAX >> 1);
+  rst[1].restoration_tilesize = rst[0].restoration_tilesize;
+  rst[2].restoration_tilesize = rst[0].restoration_tilesize;
+}
+#endif  // CONFIG_LOOP_RESTORATION
+
+static void set_frame_size(AV1_COMP *cpi) {
+  int ref_frame;
+  AV1_COMMON *const cm = &cpi->common;
+  AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
+
+  if (oxcf->pass == 2 && oxcf->rc_mode == AOM_VBR &&
+      ((oxcf->resize_mode == RESIZE_FIXED && cm->current_video_frame == 0) ||
+       (oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending))) {
+    av1_calculate_coded_size(cpi, &oxcf->scaled_frame_width,
+                             &oxcf->scaled_frame_height);
+
+    // There has been a change in frame size.
+    av1_set_size_literal(cpi, oxcf->scaled_frame_width,
+                         oxcf->scaled_frame_height);
+  }
+
+  if (oxcf->pass == 0 && oxcf->rc_mode == AOM_CBR &&
+      oxcf->resize_mode == RESIZE_DYNAMIC) {
+    if (cpi->resize_pending == 1) {
+      oxcf->scaled_frame_width =
+          (cm->width * cpi->resize_scale_num) / cpi->resize_scale_den;
+      oxcf->scaled_frame_height =
+          (cm->height * cpi->resize_scale_num) / cpi->resize_scale_den;
+    } else if (cpi->resize_pending == -1) {
+      // Go back up to original size.
+      oxcf->scaled_frame_width = oxcf->width;
+      oxcf->scaled_frame_height = oxcf->height;
+    }
+    if (cpi->resize_pending != 0) {
+      // There has been a change in frame size.
+      av1_set_size_literal(cpi, oxcf->scaled_frame_width,
+                           oxcf->scaled_frame_height);
+
+      // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
+      set_mv_search_params(cpi);
+    }
+  }
+
+#if !CONFIG_XIPHRC
+  if (oxcf->pass == 2) {
+    av1_set_target_rate(cpi);
+  }
+#endif
+
+  alloc_frame_mvs(cm, cm->new_fb_idx);
+
+  // Reset the frame pointers to the current frame size.
+  if (aom_realloc_frame_buffer(get_frame_new_buffer(cm), cm->width, cm->height,
+                               cm->subsampling_x, cm->subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                               cm->use_highbitdepth,
+#endif
+                               AOM_BORDER_IN_PIXELS, cm->byte_alignment, NULL,
+                               NULL, NULL))
+    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
+                       "Failed to allocate frame buffer");
+
+#if CONFIG_LOOP_RESTORATION
+  set_restoration_tilesize(cm->width, cm->height, cm->rst_info);
+  for (int i = 0; i < MAX_MB_PLANE; ++i)
+    cm->rst_info[i].frame_restoration_type = RESTORE_NONE;
+  av1_alloc_restoration_buffers(cm);
+  for (int i = 0; i < MAX_MB_PLANE; ++i) {
+    cpi->rst_search[i].restoration_tilesize =
+        cm->rst_info[i].restoration_tilesize;
+    av1_alloc_restoration_struct(cm, &cpi->rst_search[i], cm->width,
+                                 cm->height);
+  }
+#endif  // CONFIG_LOOP_RESTORATION
+  alloc_util_frame_buffers(cpi);
+  init_motion_estimation(cpi);
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - LAST_FRAME];
+    const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+
+    ref_buf->idx = buf_idx;
+
+    if (buf_idx != INVALID_IDX) {
+      YV12_BUFFER_CONFIG *const buf = &cm->buffer_pool->frame_bufs[buf_idx].buf;
+      ref_buf->buf = buf;
+#if CONFIG_HIGHBITDEPTH
+      av1_setup_scale_factors_for_frame(
+          &ref_buf->sf, buf->y_crop_width, buf->y_crop_height, cm->width,
+          cm->height, (buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0);
+#else
+      av1_setup_scale_factors_for_frame(&ref_buf->sf, buf->y_crop_width,
+                                        buf->y_crop_height, cm->width,
+                                        cm->height);
+#endif  // CONFIG_HIGHBITDEPTH
+      if (av1_is_scaled(&ref_buf->sf)) aom_extend_frame_borders(buf);
+    } else {
+      ref_buf->buf = NULL;
+    }
+  }
+
+  set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
+}
+
+static void reset_use_upsampled_references(AV1_COMP *cpi) {
+  MV_REFERENCE_FRAME ref_frame;
+
+  // reset up-sampled reference buffer structure.
+  init_upsampled_ref_frame_bufs(cpi);
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, ref_frame);
+    int new_uidx = upsample_ref_frame(cpi, ref);
+
+    // Update the up-sampled reference index.
+    cpi->upsampled_ref_idx[get_ref_frame_map_idx(cpi, ref_frame)] = new_uidx;
+    cpi->upsampled_ref_bufs[new_uidx].ref_count++;
+  }
+}
+
+static void encode_without_recode_loop(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  int q = 0, bottom_index = 0, top_index = 0;  // Dummy variables.
+  const int use_upsampled_ref = cpi->sf.use_upsampled_references;
+
+  aom_clear_system_state();
+
+  set_frame_size(cpi);
+
+  // For 1 pass CBR under dynamic resize mode: use faster scaling for source.
+  // Only for 2x2 scaling for now.
+  if (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == AOM_CBR &&
+      cpi->oxcf.resize_mode == RESIZE_DYNAMIC &&
+      cpi->un_scaled_source->y_width == (cm->width << 1) &&
+      cpi->un_scaled_source->y_height == (cm->height << 1)) {
+    cpi->source = av1_scale_if_required_fast(cm, cpi->un_scaled_source,
+                                             &cpi->scaled_source);
+    if (cpi->unscaled_last_source != NULL)
+      cpi->last_source = av1_scale_if_required_fast(
+          cm, cpi->unscaled_last_source, &cpi->scaled_last_source);
+  } else {
+    cpi->source =
+        av1_scale_if_required(cm, cpi->un_scaled_source, &cpi->scaled_source);
+    if (cpi->unscaled_last_source != NULL)
+      cpi->last_source = av1_scale_if_required(cm, cpi->unscaled_last_source,
+                                               &cpi->scaled_last_source);
+  }
+
+  if (frame_is_intra_only(cm) == 0) {
+    av1_scale_references(cpi);
+  }
+
+  set_size_independent_vars(cpi);
+  set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
+
+  // cpi->sf.use_upsampled_references can be different from frame to frame.
+  // Every time when cpi->sf.use_upsampled_references is changed from 0 to 1.
+  // The reference frames for this frame have to be up-sampled before encoding.
+  if (!use_upsampled_ref && cpi->sf.use_upsampled_references &&
+      cm->frame_type != KEY_FRAME)
+    reset_use_upsampled_references(cpi);
+
+  av1_set_quantizer(cm, q);
+  av1_set_variance_partition_thresholds(cpi, q);
+
+  setup_frame(cpi);
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  cm->do_subframe_update = cm->tile_cols == 1 && cm->tile_rows == 1;
+  av1_copy(cm->starting_coef_probs, cm->fc->coef_probs);
+  av1_copy(cpi->subframe_stats.enc_starting_coef_probs, cm->fc->coef_probs);
+  cm->coef_probs_update_idx = 0;
+  av1_copy(cpi->subframe_stats.coef_probs_buf[0], cm->fc->coef_probs);
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+  suppress_active_map(cpi);
+  // Variance adaptive and in frame q adjustment experiments are mutually
+  // exclusive.
+  if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
+    av1_vaq_frame_setup(cpi);
+  } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
+    av1_setup_in_frame_q_adj(cpi);
+  } else if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
+    av1_cyclic_refresh_setup(cpi);
+  }
+  apply_active_map(cpi);
+
+  // transform / motion compensation build reconstruction frame
+  av1_encode_frame(cpi);
+
+  // Update some stats from cyclic refresh, and check if we should not update
+  // golden reference, for 1 pass CBR.
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->frame_type != KEY_FRAME &&
+      (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == AOM_CBR))
+    av1_cyclic_refresh_check_golden_update(cpi);
+
+  // Update the skip mb flag probabilities based on the distribution
+  // seen in the last encoder iteration.
+  // update_base_skip_probs(cpi);
+  aom_clear_system_state();
+}
+
+static void encode_with_recode_loop(AV1_COMP *cpi, size_t *size,
+                                    uint8_t *dest) {
+  AV1_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int bottom_index, top_index;
+  int loop_count = 0;
+  int loop_at_this_size = 0;
+  int loop = 0;
+#if !CONFIG_XIPHRC
+  int overshoot_seen = 0;
+  int undershoot_seen = 0;
+#endif
+  int frame_over_shoot_limit;
+  int frame_under_shoot_limit;
+  int q = 0, q_low = 0, q_high = 0;
+  const int use_upsampled_ref = cpi->sf.use_upsampled_references;
+
+  set_size_independent_vars(cpi);
+
+  do {
+    aom_clear_system_state();
+
+    set_frame_size(cpi);
+
+    if (loop_count == 0 || cpi->resize_pending != 0) {
+      set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
+
+      // cpi->sf.use_upsampled_references can be different from frame to frame.
+      // Every time when cpi->sf.use_upsampled_references is changed from 0 to
+      // 1.
+      // The reference frames for this frame have to be up-sampled before
+      // encoding.
+      if (!use_upsampled_ref && cpi->sf.use_upsampled_references &&
+          cm->frame_type != KEY_FRAME)
+        reset_use_upsampled_references(cpi);
+
+      // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
+      set_mv_search_params(cpi);
+
+#if !CONFIG_XIPHRC
+      // Reset the loop state for new frame size.
+      overshoot_seen = 0;
+      undershoot_seen = 0;
+#endif
+
+      // Reconfiguration for change in frame size has concluded.
+      cpi->resize_pending = 0;
+
+      q_low = bottom_index;
+      q_high = top_index;
+
+      loop_at_this_size = 0;
+    }
+
+    // Decide frame size bounds first time through.
+    if (loop_count == 0) {
+      av1_rc_compute_frame_size_bounds(cpi, rc->this_frame_target,
+                                       &frame_under_shoot_limit,
+                                       &frame_over_shoot_limit);
+    }
+
+    cpi->source =
+        av1_scale_if_required(cm, cpi->un_scaled_source, &cpi->scaled_source);
+
+    if (cpi->unscaled_last_source != NULL)
+      cpi->last_source = av1_scale_if_required(cm, cpi->unscaled_last_source,
+                                               &cpi->scaled_last_source);
+
+    if (frame_is_intra_only(cm) == 0) {
+      if (loop_count > 0) {
+        release_scaled_references(cpi);
+      }
+      av1_scale_references(cpi);
+    }
+
+    av1_set_quantizer(cm, q);
+
+    if (loop_count == 0) setup_frame(cpi);
+
+#if CONFIG_Q_ADAPT_PROBS
+    // Base q-index may have changed, so we need to assign proper default coef
+    // probs before every iteration.
+    if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
+      int i;
+      av1_default_coef_probs(cm);
+      if (cm->frame_type == KEY_FRAME || cm->error_resilient_mode ||
+          cm->reset_frame_context == RESET_FRAME_CONTEXT_ALL) {
+        for (i = 0; i < FRAME_CONTEXTS; ++i) cm->frame_contexts[i] = *cm->fc;
+      } else if (cm->reset_frame_context == RESET_FRAME_CONTEXT_CURRENT) {
+        cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+      }
+    }
+#endif  // CONFIG_Q_ADAPT_PROBS
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+    cm->do_subframe_update = cm->tile_cols == 1 && cm->tile_rows == 1;
+    if (loop_count == 0 || frame_is_intra_only(cm) ||
+        cm->error_resilient_mode) {
+      av1_copy(cm->starting_coef_probs, cm->fc->coef_probs);
+      av1_copy(cpi->subframe_stats.enc_starting_coef_probs, cm->fc->coef_probs);
+    } else {
+      if (cm->do_subframe_update) {
+        av1_copy(cm->fc->coef_probs,
+                 cpi->subframe_stats.enc_starting_coef_probs);
+        av1_copy(cm->starting_coef_probs,
+                 cpi->subframe_stats.enc_starting_coef_probs);
+        av1_zero(cpi->subframe_stats.coef_counts_buf);
+        av1_zero(cpi->subframe_stats.eob_counts_buf);
+      }
+    }
+    cm->coef_probs_update_idx = 0;
+    av1_copy(cpi->subframe_stats.coef_probs_buf[0], cm->fc->coef_probs);
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+    // Variance adaptive and in frame q adjustment experiments are mutually
+    // exclusive.
+    if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
+      av1_vaq_frame_setup(cpi);
+    } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
+      av1_setup_in_frame_q_adj(cpi);
+    }
+
+    // transform / motion compensation build reconstruction frame
+    av1_encode_frame(cpi);
+
+    // Update the skip mb flag probabilities based on the distribution
+    // seen in the last encoder iteration.
+    // update_base_skip_probs(cpi);
+
+    aom_clear_system_state();
+
+    // Dummy pack of the bitstream using up to date stats to get an
+    // accurate estimate of output frame size to determine if we need
+    // to recode.
+    if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF) {
+      save_coding_context(cpi);
+
+      av1_pack_bitstream(cpi, dest, size);
+
+      rc->projected_frame_size = (int)(*size) << 3;
+      restore_coding_context(cpi);
+
+      if (frame_over_shoot_limit == 0) frame_over_shoot_limit = 1;
+    }
+
+    if (cpi->oxcf.rc_mode == AOM_Q) {
+      loop = 0;
+    } else {
+      if ((cm->frame_type == KEY_FRAME) && rc->this_key_frame_forced &&
+          (rc->projected_frame_size < rc->max_frame_bandwidth)) {
+        int last_q = q;
+        int64_t kf_err;
+
+        int64_t high_err_target = cpi->ambient_err;
+        int64_t low_err_target = cpi->ambient_err >> 1;
+
+#if CONFIG_HIGHBITDEPTH
+        if (cm->use_highbitdepth) {
+          kf_err = aom_highbd_get_y_sse(cpi->source, get_frame_new_buffer(cm));
+        } else {
+          kf_err = aom_get_y_sse(cpi->source, get_frame_new_buffer(cm));
+        }
+#else
+        kf_err = aom_get_y_sse(cpi->source, get_frame_new_buffer(cm));
+#endif  // CONFIG_HIGHBITDEPTH
+
+        // Prevent possible divide by zero error below for perfect KF
+        kf_err += !kf_err;
+
+        // The key frame is not good enough or we can afford
+        // to make it better without undue risk of popping.
+        if ((kf_err > high_err_target &&
+             rc->projected_frame_size <= frame_over_shoot_limit) ||
+            (kf_err > low_err_target &&
+             rc->projected_frame_size <= frame_under_shoot_limit)) {
+          // Lower q_high
+          q_high = q > q_low ? q - 1 : q_low;
+
+          // Adjust Q
+          q = (int)((q * high_err_target) / kf_err);
+          q = AOMMIN(q, (q_high + q_low) >> 1);
+        } else if (kf_err < low_err_target &&
+                   rc->projected_frame_size >= frame_under_shoot_limit) {
+          // The key frame is much better than the previous frame
+          // Raise q_low
+          q_low = q < q_high ? q + 1 : q_high;
+
+          // Adjust Q
+          q = (int)((q * low_err_target) / kf_err);
+          q = AOMMIN(q, (q_high + q_low + 1) >> 1);
+        }
+
+        // Clamp Q to upper and lower limits:
+        q = clamp(q, q_low, q_high);
+
+        loop = q != last_q;
+      } else if (recode_loop_test(cpi, frame_over_shoot_limit,
+                                  frame_under_shoot_limit, q,
+                                  AOMMAX(q_high, top_index), bottom_index)) {
+        // Is the projected frame size out of range and are we allowed
+        // to attempt to recode.
+        int last_q = q;
+#if !CONFIG_XIPHRC
+        int retries = 0;
+#endif
+
+        if (cpi->resize_pending == 1) {
+          // Change in frame size so go back around the recode loop.
+          cpi->rc.frame_size_selector =
+              SCALE_STEP1 - cpi->rc.frame_size_selector;
+          cpi->rc.next_frame_size_selector = cpi->rc.frame_size_selector;
+
+#if CONFIG_INTERNAL_STATS
+          ++cpi->tot_recode_hits;
+#endif
+          ++loop_count;
+          loop = 1;
+          continue;
+        }
+
+#if !CONFIG_XIPHRC
+        // Frame size out of permitted range:
+        // Update correction factor & compute new Q to try...
+        // Frame is too large
+        if (rc->projected_frame_size > rc->this_frame_target) {
+          // Special case if the projected size is > the max allowed.
+          if (rc->projected_frame_size >= rc->max_frame_bandwidth)
+            q_high = rc->worst_quality;
+
+          // Raise Qlow as to at least the current value
+          q_low = q < q_high ? q + 1 : q_high;
+
+          if (undershoot_seen || loop_at_this_size > 1) {
+            // Update rate_correction_factor unless
+            av1_rc_update_rate_correction_factors(cpi);
+
+            q = (q_high + q_low + 1) / 2;
+          } else {
+            // Update rate_correction_factor unless
+            av1_rc_update_rate_correction_factors(cpi);
+
+            q = av1_rc_regulate_q(cpi, rc->this_frame_target, bottom_index,
+                                  AOMMAX(q_high, top_index));
+
+            while (q < q_low && retries < 10) {
+              av1_rc_update_rate_correction_factors(cpi);
+              q = av1_rc_regulate_q(cpi, rc->this_frame_target, bottom_index,
+                                    AOMMAX(q_high, top_index));
+              retries++;
+            }
+          }
+
+          overshoot_seen = 1;
+        } else {
+          // Frame is too small
+          q_high = q > q_low ? q - 1 : q_low;
+
+          if (overshoot_seen || loop_at_this_size > 1) {
+            av1_rc_update_rate_correction_factors(cpi);
+            q = (q_high + q_low) / 2;
+          } else {
+            av1_rc_update_rate_correction_factors(cpi);
+            q = av1_rc_regulate_q(cpi, rc->this_frame_target, bottom_index,
+                                  top_index);
+            // Special case reset for qlow for constrained quality.
+            // This should only trigger where there is very substantial
+            // undershoot on a frame and the auto cq level is above
+            // the user passsed in value.
+            if (cpi->oxcf.rc_mode == AOM_CQ && q < q_low) {
+              q_low = q;
+            }
+
+            while (q > q_high && retries < 10) {
+              av1_rc_update_rate_correction_factors(cpi);
+              q = av1_rc_regulate_q(cpi, rc->this_frame_target, bottom_index,
+                                    top_index);
+              retries++;
+            }
+          }
+
+          undershoot_seen = 1;
+        }
+#endif
+
+        // Clamp Q to upper and lower limits:
+        q = clamp(q, q_low, q_high);
+
+        loop = (q != last_q);
+      } else {
+        loop = 0;
+      }
+    }
+
+    // Special case for overlay frame.
+    if (rc->is_src_frame_alt_ref &&
+        rc->projected_frame_size < rc->max_frame_bandwidth)
+      loop = 0;
+
+#if CONFIG_GLOBAL_MOTION
+    if (recode_loop_test_global_motion(cpi)) {
+      loop = 1;
+    }
+#endif  // CONFIG_GLOBAL_MOTION
+
+    if (loop) {
+      ++loop_count;
+      ++loop_at_this_size;
+
+#if CONFIG_INTERNAL_STATS
+      ++cpi->tot_recode_hits;
+#endif
+    }
+  } while (loop);
+}
+
+static int get_ref_frame_flags(const AV1_COMP *cpi) {
+  const int *const map = cpi->common.ref_frame_map;
+
+#if CONFIG_EXT_REFS
+  const int last2_is_last =
+      map[cpi->lst_fb_idxes[1]] == map[cpi->lst_fb_idxes[0]];
+  const int last3_is_last =
+      map[cpi->lst_fb_idxes[2]] == map[cpi->lst_fb_idxes[0]];
+  const int gld_is_last = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idxes[0]];
+#if CONFIG_LOWDELAY_COMPOUND
+  const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idxes[0]];
+  const int last3_is_last2 =
+      map[cpi->lst_fb_idxes[2]] == map[cpi->lst_fb_idxes[1]];
+  const int gld_is_last2 = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idxes[1]];
+  const int gld_is_last3 = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idxes[2]];
+#else
+  const int bwd_is_last = map[cpi->bwd_fb_idx] == map[cpi->lst_fb_idxes[0]];
+  const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idxes[0]];
+
+  const int last3_is_last2 =
+      map[cpi->lst_fb_idxes[2]] == map[cpi->lst_fb_idxes[1]];
+  const int gld_is_last2 = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idxes[1]];
+  const int bwd_is_last2 = map[cpi->bwd_fb_idx] == map[cpi->lst_fb_idxes[1]];
+
+  const int gld_is_last3 = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idxes[2]];
+  const int bwd_is_last3 = map[cpi->bwd_fb_idx] == map[cpi->lst_fb_idxes[2]];
+
+  const int bwd_is_gld = map[cpi->bwd_fb_idx] == map[cpi->gld_fb_idx];
+
+#endif
+  const int last2_is_alt = map[cpi->lst_fb_idxes[1]] == map[cpi->alt_fb_idx];
+  const int last3_is_alt = map[cpi->lst_fb_idxes[2]] == map[cpi->alt_fb_idx];
+  const int gld_is_alt = map[cpi->gld_fb_idx] == map[cpi->alt_fb_idx];
+  const int bwd_is_alt = map[cpi->bwd_fb_idx] == map[cpi->alt_fb_idx];
+#else
+  const int gld_is_last = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idx];
+  const int gld_is_alt = map[cpi->gld_fb_idx] == map[cpi->alt_fb_idx];
+  const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idx];
+#endif  // CONFIG_EXT_REFS
+
+  int flags = AOM_REFFRAME_ALL;
+
+#if CONFIG_EXT_REFS
+  // Disable the use of BWDREF_FRAME for non-bipredictive frames.
+  if (!(cpi->rc.is_bipred_frame || cpi->rc.is_last_bipred_frame ||
+        (cpi->rc.is_bwd_ref_frame && cpi->num_extra_arfs)))
+    flags &= ~AOM_BWD_FLAG;
+#endif  // CONFIG_EXT_REFS
+
+  if (gld_is_last || gld_is_alt) flags &= ~AOM_GOLD_FLAG;
+
+  if (cpi->rc.frames_till_gf_update_due == INT_MAX) flags &= ~AOM_GOLD_FLAG;
+
+  if (alt_is_last) flags &= ~AOM_ALT_FLAG;
+
+#if CONFIG_EXT_REFS
+  if (last2_is_last || last2_is_alt) flags &= ~AOM_LAST2_FLAG;
+
+  if (last3_is_last || last3_is_last2 || last3_is_alt) flags &= ~AOM_LAST3_FLAG;
+
+  if (gld_is_last2 || gld_is_last3) flags &= ~AOM_GOLD_FLAG;
+
+#if CONFIG_LOWDELAY_COMPOUND  // Changes LL & HL bitstream
+  /* Allow biprediction between two identical frames (e.g. bwd_is_last = 1) */
+  if (bwd_is_alt && (flags & AOM_BWD_FLAG)) flags &= ~AOM_BWD_FLAG;
+#else
+  if ((bwd_is_last || bwd_is_last2 || bwd_is_last3 || bwd_is_gld ||
+       bwd_is_alt) &&
+      (flags & AOM_BWD_FLAG))
+    flags &= ~AOM_BWD_FLAG;
+#endif
+#endif  // CONFIG_EXT_REFS
+
+  return flags;
+}
+
+static void set_ext_overrides(AV1_COMP *cpi) {
+  // Overrides the defaults with the externally supplied values with
+  // av1_update_reference() and av1_update_entropy() calls
+  // Note: The overrides are valid only for the next frame passed
+  // to encode_frame_to_data_rate() function
+  if (cpi->ext_refresh_frame_context_pending) {
+    cpi->common.refresh_frame_context = cpi->ext_refresh_frame_context;
+    cpi->ext_refresh_frame_context_pending = 0;
+  }
+  if (cpi->ext_refresh_frame_flags_pending) {
+    cpi->refresh_last_frame = cpi->ext_refresh_last_frame;
+    cpi->refresh_golden_frame = cpi->ext_refresh_golden_frame;
+    cpi->refresh_alt_ref_frame = cpi->ext_refresh_alt_ref_frame;
+    cpi->ext_refresh_frame_flags_pending = 0;
+  }
+}
+
+YV12_BUFFER_CONFIG *av1_scale_if_required_fast(AV1_COMMON *cm,
+                                               YV12_BUFFER_CONFIG *unscaled,
+                                               YV12_BUFFER_CONFIG *scaled) {
+  if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
+      cm->mi_rows * MI_SIZE != unscaled->y_height) {
+    // For 2x2 scaling down.
+    aom_scale_frame(unscaled, scaled, unscaled->y_buffer, 9, 2, 1, 2, 1, 0);
+    aom_extend_frame_borders(scaled);
+    return scaled;
+  } else {
+    return unscaled;
+  }
+}
+
+YV12_BUFFER_CONFIG *av1_scale_if_required(AV1_COMMON *cm,
+                                          YV12_BUFFER_CONFIG *unscaled,
+                                          YV12_BUFFER_CONFIG *scaled) {
+  if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
+      cm->mi_rows * MI_SIZE != unscaled->y_height) {
+#if CONFIG_HIGHBITDEPTH
+    scale_and_extend_frame_nonnormative(unscaled, scaled, (int)cm->bit_depth);
+#else
+    scale_and_extend_frame_nonnormative(unscaled, scaled);
+#endif  // CONFIG_HIGHBITDEPTH
+    return scaled;
+  } else {
+    return unscaled;
+  }
+}
+
+static void set_arf_sign_bias(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  int arf_sign_bias;
+#if CONFIG_EXT_REFS
+  const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+  // The arf_sign_bias will be one for internal ARFs'
+  arf_sign_bias = cpi->rc.source_alt_ref_active &&
+                  (!cpi->refresh_alt_ref_frame ||
+                   (gf_group->rf_level[gf_group->index] == GF_ARF_LOW));
+#else
+  if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    arf_sign_bias = cpi->rc.source_alt_ref_active &&
+                    (!cpi->refresh_alt_ref_frame ||
+                     (gf_group->rf_level[gf_group->index] == GF_ARF_LOW));
+  } else {
+    arf_sign_bias =
+        (cpi->rc.source_alt_ref_active && !cpi->refresh_alt_ref_frame);
+  }
+#endif  // CONFIG_EXT_REFS
+
+  cm->ref_frame_sign_bias[ALTREF_FRAME] = arf_sign_bias;
+#if CONFIG_EXT_REFS
+  cm->ref_frame_sign_bias[BWDREF_FRAME] = cm->ref_frame_sign_bias[ALTREF_FRAME];
+#endif  // CONFIG_EXT_REFS
+}
+
+static int setup_interp_filter_search_mask(AV1_COMP *cpi) {
+  InterpFilter ifilter;
+  int ref_total[TOTAL_REFS_PER_FRAME] = { 0 };
+  MV_REFERENCE_FRAME ref;
+  int mask = 0;
+  int arf_idx = ALTREF_FRAME;
+
+#if CONFIG_EXT_REFS
+  // Get which arf used as ALTREF_FRAME
+  if (cpi->oxcf.pass == 2)
+    arf_idx += cpi->twopass.gf_group.arf_ref_idx[cpi->twopass.gf_group.index];
+#endif  // CONFIG_EXT_REFS
+
+  if (cpi->common.last_frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame)
+    return mask;
+
+#if CONFIG_EXT_REFS
+  for (ref = LAST_FRAME; ref < ALTREF_FRAME; ++ref)
+    for (ifilter = EIGHTTAP_REGULAR; ifilter < SWITCHABLE_FILTERS; ++ifilter)
+      ref_total[ref] += cpi->interp_filter_selected[ref][ifilter];
+
+  for (ifilter = EIGHTTAP_REGULAR; ifilter < SWITCHABLE_FILTERS; ++ifilter)
+    ref_total[ref] += cpi->interp_filter_selected[arf_idx][ifilter];
+#else
+  for (ref = LAST_FRAME; ref <= ALTREF_FRAME; ++ref)
+    for (ifilter = EIGHTTAP_REGULAR; ifilter < SWITCHABLE_FILTERS; ++ifilter)
+      ref_total[ref] += cpi->interp_filter_selected[ref][ifilter];
+#endif  // CONFIG_EXT_REFS
+
+  for (ifilter = EIGHTTAP_REGULAR; ifilter < SWITCHABLE_FILTERS; ++ifilter) {
+    if ((ref_total[LAST_FRAME] &&
+         cpi->interp_filter_selected[LAST_FRAME][ifilter] == 0) &&
+#if CONFIG_EXT_REFS
+        (ref_total[LAST2_FRAME] == 0 ||
+         cpi->interp_filter_selected[LAST2_FRAME][ifilter] * 50 <
+             ref_total[LAST2_FRAME]) &&
+        (ref_total[LAST3_FRAME] == 0 ||
+         cpi->interp_filter_selected[LAST3_FRAME][ifilter] * 50 <
+             ref_total[LAST3_FRAME]) &&
+#endif  // CONFIG_EXT_REFS
+        (ref_total[GOLDEN_FRAME] == 0 ||
+         cpi->interp_filter_selected[GOLDEN_FRAME][ifilter] * 50 <
+             ref_total[GOLDEN_FRAME]) &&
+#if CONFIG_EXT_REFS
+        (ref_total[BWDREF_FRAME] == 0 ||
+         cpi->interp_filter_selected[BWDREF_FRAME][ifilter] * 50 <
+             ref_total[BWDREF_FRAME]) &&
+#endif  // CONFIG_EXT_REFS
+        (ref_total[ALTREF_FRAME] == 0 ||
+         cpi->interp_filter_selected[arf_idx][ifilter] * 50 <
+             ref_total[ALTREF_FRAME]))
+      mask |= 1 << ifilter;
+  }
+  return mask;
+}
+
+#define DUMP_RECON_FRAMES 0
+
+#if DUMP_RECON_FRAMES == 1
+// NOTE(zoeliu): For debug - Output the filtered reconstructed video.
+static void dump_filtered_recon_frames(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  const YV12_BUFFER_CONFIG *recon_buf = cm->frame_to_show;
+  int h;
+  char file_name[256] = "/tmp/enc_filtered_recon.yuv";
+  FILE *f_recon = NULL;
+
+  if (recon_buf == NULL || !cm->show_frame) {
+    printf("Frame %d is not ready or no show to dump.\n",
+           cm->current_video_frame);
+    return;
+  }
+
+  if (cm->current_video_frame == 0) {
+    if ((f_recon = fopen(file_name, "wb")) == NULL) {
+      printf("Unable to open file %s to write.\n", file_name);
+      return;
+    }
+  } else {
+    if ((f_recon = fopen(file_name, "ab")) == NULL) {
+      printf("Unable to open file %s to append.\n", file_name);
+      return;
+    }
+  }
+  printf(
+      "\nFrame=%5d, encode_update_type[%5d]=%1d, show_existing_frame=%d, "
+      "y_stride=%4d, uv_stride=%4d, width=%4d, height=%4d\n",
+      cm->current_video_frame, cpi->twopass.gf_group.index,
+      cpi->twopass.gf_group.update_type[cpi->twopass.gf_group.index],
+      cm->show_existing_frame, recon_buf->y_stride, recon_buf->uv_stride,
+      cm->width, cm->height);
+
+  // --- Y ---
+  for (h = 0; h < cm->height; ++h) {
+    fwrite(&recon_buf->y_buffer[h * recon_buf->y_stride], 1, cm->width,
+           f_recon);
+  }
+  // --- U ---
+  for (h = 0; h < (cm->height >> 1); ++h) {
+    fwrite(&recon_buf->u_buffer[h * recon_buf->uv_stride], 1, (cm->width >> 1),
+           f_recon);
+  }
+  // --- V ---
+  for (h = 0; h < (cm->height >> 1); ++h) {
+    fwrite(&recon_buf->v_buffer[h * recon_buf->uv_stride], 1, (cm->width >> 1),
+           f_recon);
+  }
+
+  fclose(f_recon);
+}
+#endif  // DUMP_RECON_FRAMES
+
+#if CONFIG_EC_ADAPT
+
+static void make_update_tile_list_enc(AV1_COMP *cpi, const int tile_rows,
+                                      const int tile_cols,
+                                      FRAME_CONTEXT *ec_ctxs[]) {
+  int i;
+  for (i = 0; i < tile_rows * tile_cols; ++i)
+    ec_ctxs[i] = &cpi->tile_data[i].tctx;
+}
+
+#endif
+static void encode_frame_to_data_rate(AV1_COMP *cpi, size_t *size,
+                                      uint8_t *dest, int skip_adapt,
+                                      unsigned int *frame_flags) {
+  AV1_COMMON *const cm = &cpi->common;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  struct segmentation *const seg = &cm->seg;
+  TX_SIZE t;
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT **tile_ctxs = aom_malloc(cm->tile_rows * cm->tile_cols *
+                                         sizeof(&cpi->tile_data[0].tctx));
+  aom_cdf_prob **cdf_ptrs =
+      aom_malloc(cm->tile_rows * cm->tile_cols *
+                 sizeof(&cpi->tile_data[0].tctx.partition_cdf[0][0]));
+#endif
+#if CONFIG_XIPHRC
+  int frame_type;
+  int drop_this_frame = 0;
+#endif  // CONFIG_XIPHRC
+  set_ext_overrides(cpi);
+  aom_clear_system_state();
+
+  // Set the arf sign bias for this frame.
+  set_arf_sign_bias(cpi);
+#if CONFIG_TEMPMV_SIGNALING
+  // frame type has been decided outside of this function call
+  cm->cur_frame->intra_only = cm->frame_type == KEY_FRAME || cm->intra_only;
+  cm->use_prev_frame_mvs =
+      !cpi->oxcf.disable_tempmv && !cm->cur_frame->intra_only;
+#endif
+
+#if CONFIG_EXT_REFS
+  // NOTE:
+  // (1) Move the setup of the ref_frame_flags upfront as it would be
+  //     determined by the current frame properties;
+  // (2) The setup of the ref_frame_flags applies to both show_existing_frame's
+  //     and the other cases.
+  if (cm->current_video_frame > 0)
+    cpi->ref_frame_flags = get_ref_frame_flags(cpi);
+
+  if (cm->show_existing_frame) {
+    // NOTE(zoeliu): In BIDIR_PRED, the existing frame to show is the current
+    //               BWDREF_FRAME in the reference frame buffer.
+    cm->frame_type = INTER_FRAME;
+    cm->show_frame = 1;
+    cpi->frame_flags = *frame_flags;
+
+    // In the case of show_existing frame, we will not send fresh flag
+    // to decoder. Any change in the reference frame buffer can be done by
+    // switching the virtual indices.
+
+    cpi->refresh_last_frame = 0;
+    cpi->refresh_golden_frame = 0;
+    cpi->refresh_bwd_ref_frame = 0;
+    cpi->refresh_alt_ref_frame = 0;
+
+    cpi->rc.is_bwd_ref_frame = 0;
+    cpi->rc.is_last_bipred_frame = 0;
+    cpi->rc.is_bipred_frame = 0;
+
+    // Build the bitstream
+    av1_pack_bitstream(cpi, dest, size);
+
+    // Set up frame to show to get ready for stats collection.
+    cm->frame_to_show = get_frame_new_buffer(cm);
+
+#if DUMP_RECON_FRAMES == 1
+    // NOTE(zoeliu): For debug - Output the filtered reconstructed video.
+    dump_filtered_recon_frames(cpi);
+#endif  // DUMP_RECON_FRAMES
+
+    // Update the LAST_FRAME in the reference frame buffer.
+    av1_update_reference_frames(cpi);
+
+    // Update frame flags
+    cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN;
+    cpi->frame_flags &= ~FRAMEFLAGS_BWDREF;
+    cpi->frame_flags &= ~FRAMEFLAGS_ALTREF;
+
+    *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY;
+
+    // Update the frame type
+    cm->last_frame_type = cm->frame_type;
+
+    // Since we allocate a spot for the OVERLAY frame in the gf group, we need
+    // to do post-encoding update accordingly.
+    if (cpi->rc.is_src_frame_alt_ref) {
+      av1_set_target_rate(cpi);
+#if CONFIG_XIPHRC
+      frame_type = cm->frame_type == INTER_FRAME ? OD_P_FRAME : OD_I_FRAME;
+      drop_this_frame = od_enc_rc_update_state(
+          &cpi->od_rc, *size << 3, cpi->refresh_golden_frame,
+          cpi->refresh_alt_ref_frame, frame_type, cpi->droppable);
+#else
+      av1_rc_postencode_update(cpi, *size);
+#endif
+    }
+
+    cm->last_width = cm->width;
+    cm->last_height = cm->height;
+
+    ++cm->current_video_frame;
+
+#if CONFIG_EC_ADAPT
+    aom_free(tile_ctxs);
+    aom_free(cdf_ptrs);
+#endif
+    return;
+  }
+#endif  // CONFIG_EXT_REFS
+
+  // Set default state for segment based loop filter update flags.
+  cm->lf.mode_ref_delta_update = 0;
+
+  if (cpi->oxcf.pass == 2 && cpi->sf.adaptive_interp_filter_search)
+    cpi->sf.interp_filter_search_mask = setup_interp_filter_search_mask(cpi);
+
+  // Set various flags etc to special state if it is a key frame.
+  if (frame_is_intra_only(cm)) {
+    // Reset the loop filter deltas and segmentation map.
+    av1_reset_segment_features(cm);
+
+    // If segmentation is enabled force a map update for key frames.
+    if (seg->enabled) {
+      seg->update_map = 1;
+      seg->update_data = 1;
+    }
+
+    // The alternate reference frame cannot be active for a key frame.
+    cpi->rc.source_alt_ref_active = 0;
+
+    cm->error_resilient_mode = oxcf->error_resilient_mode;
+
+    // By default, encoder assumes decoder can use prev_mi.
+    if (cm->error_resilient_mode) {
+      cm->reset_frame_context = RESET_FRAME_CONTEXT_NONE;
+      cm->refresh_frame_context = REFRESH_FRAME_CONTEXT_FORWARD;
+    } else if (cm->intra_only) {
+      // Only reset the current context.
+      cm->reset_frame_context = RESET_FRAME_CONTEXT_CURRENT;
+    }
+  }
+#if CONFIG_TILE_GROUPS
+  if (cpi->oxcf.mtu == 0) {
+    cm->num_tg = cpi->oxcf.num_tile_groups;
+  } else {
+    // Use a default value for the purposes of weighting costs in probability
+    // updates
+    cm->num_tg = DEFAULT_MAX_NUM_TG;
+  }
+#endif
+
+#if CONFIG_EXT_TILE
+  cm->tile_encoding_mode = cpi->oxcf.tile_encoding_mode;
+#endif  // CONFIG_EXT_TILE
+
+#if CONFIG_XIPHRC
+  if (drop_this_frame) {
+    av1_rc_postencode_update_drop_frame(cpi);
+    ++cm->current_video_frame;
+#if CONFIG_EC_ADAPT
+    aom_free(tile_ctxs);
+    aom_free(cdf_ptrs);
+#endif
+    return;
+  }
+#else
+  // For 1 pass CBR, check if we are dropping this frame.
+  // Never drop on key frame.
+  if (oxcf->pass == 0 && oxcf->rc_mode == AOM_CBR &&
+      cm->frame_type != KEY_FRAME) {
+    if (av1_rc_drop_frame(cpi)) {
+      av1_rc_postencode_update_drop_frame(cpi);
+      ++cm->current_video_frame;
+#if CONFIG_EC_ADAPT
+      aom_free(tile_ctxs);
+      aom_free(cdf_ptrs);
+#endif
+      return;
+    }
+  }
+#endif
+
+  aom_clear_system_state();
+
+#if CONFIG_INTERNAL_STATS
+  memset(cpi->mode_chosen_counts, 0,
+         MAX_MODES * sizeof(*cpi->mode_chosen_counts));
+#endif
+
+#if CONFIG_REFERENCE_BUFFER
+  {
+    /* Non-normative definition of current_frame_id ("frame counter" with
+    * wraparound) */
+    const int frame_id_length = FRAME_ID_LENGTH_MINUS7 + 7;
+    if (cm->current_frame_id == -1) {
+      int lsb, msb;
+/* quasi-random initialization of current_frame_id for a key frame */
+#if CONFIG_HIGHBITDEPTH
+      if (cpi->source->flags & YV12_FLAG_HIGHBITDEPTH) {
+        lsb = CONVERT_TO_SHORTPTR(cpi->source->y_buffer)[0] & 0xff;
+        msb = CONVERT_TO_SHORTPTR(cpi->source->y_buffer)[1] & 0xff;
+      } else {
+#endif
+        lsb = cpi->source->y_buffer[0] & 0xff;
+        msb = cpi->source->y_buffer[1] & 0xff;
+#if CONFIG_HIGHBITDEPTH
+      }
+#endif
+      cm->current_frame_id = ((msb << 8) + lsb) % (1 << frame_id_length);
+    } else {
+      cm->current_frame_id =
+          (cm->current_frame_id + 1 + (1 << frame_id_length)) %
+          (1 << frame_id_length);
+    }
+  }
+#endif
+
+#if CONFIG_EXT_DELTA_Q
+  cm->delta_q_present_flag = cpi->oxcf.deltaq_mode != NO_DELTA_Q;
+  cm->delta_lf_present_flag = cpi->oxcf.deltaq_mode == DELTA_Q_LF;
+#endif
+
+  if (cpi->sf.recode_loop == DISALLOW_RECODE) {
+    encode_without_recode_loop(cpi);
+  } else {
+    encode_with_recode_loop(cpi, size, dest);
+  }
+
+#ifdef OUTPUT_YUV_SKINMAP
+  if (cpi->common.current_video_frame > 1) {
+    av1_compute_skin_map(cpi, yuv_skinmap_file);
+  }
+#endif  // OUTPUT_YUV_SKINMAP
+
+  // Special case code to reduce pulsing when key frames are forced at a
+  // fixed interval. Note the reconstruction error if it is the frame before
+  // the force key frame
+  if (cpi->rc.next_key_frame_forced && cpi->rc.frames_to_key == 1) {
+#if CONFIG_HIGHBITDEPTH
+    if (cm->use_highbitdepth) {
+      cpi->ambient_err =
+          aom_highbd_get_y_sse(cpi->source, get_frame_new_buffer(cm));
+    } else {
+      cpi->ambient_err = aom_get_y_sse(cpi->source, get_frame_new_buffer(cm));
+    }
+#else
+    cpi->ambient_err = aom_get_y_sse(cpi->source, get_frame_new_buffer(cm));
+#endif  // CONFIG_HIGHBITDEPTH
+  }
+
+  // If the encoder forced a KEY_FRAME decision
+  if (cm->frame_type == KEY_FRAME) {
+    cpi->refresh_last_frame = 1;
+  }
+
+  cm->frame_to_show = get_frame_new_buffer(cm);
+  cm->frame_to_show->color_space = cm->color_space;
+  cm->frame_to_show->color_range = cm->color_range;
+  cm->frame_to_show->render_width = cm->render_width;
+  cm->frame_to_show->render_height = cm->render_height;
+
+#if CONFIG_EXT_REFS
+// TODO(zoeliu): For non-ref frames, loop filtering may need to be turned
+// off.
+#endif  // CONFIG_EXT_REFS
+
+  // Pick the loop filter level for the frame.
+  loopfilter_frame(cpi, cm);
+
+  // Build the bitstream
+  av1_pack_bitstream(cpi, dest, size);
+
+  if (skip_adapt) {
+#if CONFIG_EC_ADAPT
+    aom_free(tile_ctxs);
+    aom_free(cdf_ptrs);
+#endif
+    return;
+  }
+
+#if CONFIG_REFERENCE_BUFFER
+  {
+    int i;
+    /* Update reference frame id values based on the value of refresh_mask */
+    for (i = 0; i < REF_FRAMES; i++) {
+      if ((cm->refresh_mask >> i) & 1) {
+        cm->ref_frame_id[i] = cm->current_frame_id;
+      }
+    }
+  }
+#endif
+
+#if DUMP_RECON_FRAMES == 1
+  // NOTE(zoeliu): For debug - Output the filtered reconstructed video.
+  if (cm->show_frame) dump_filtered_recon_frames(cpi);
+#endif  // DUMP_RECON_FRAMES
+
+  if (cm->seg.update_map) update_reference_segmentation_map(cpi);
+
+  if (frame_is_intra_only(cm) == 0) {
+    release_scaled_references(cpi);
+  }
+
+  av1_update_reference_frames(cpi);
+
+  for (t = 0; t < TX_SIZES; t++)
+    av1_full_to_model_counts(cpi->td.counts->coef[t],
+                             cpi->td.rd_counts.coef_counts[t]);
+#if CONFIG_ENTROPY_STATS
+  av1_accumulate_frame_counts(&aggregate_fc, &cm->counts);
+#endif  // CONFIG_ENTROPY_STATS
+  if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+#if CONFIG_SUBFRAME_PROB_UPDATE
+    cm->partial_prob_update = 0;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+    av1_adapt_coef_probs(cm);
+    av1_adapt_intra_frame_probs(cm);
+#if CONFIG_EC_ADAPT
+    make_update_tile_list_enc(cpi, cm->tile_rows, cm->tile_cols, tile_ctxs);
+    av1_average_tile_coef_cdfs(cpi->common.fc, tile_ctxs, cdf_ptrs,
+                               cm->tile_rows * cm->tile_cols);
+    av1_average_tile_intra_cdfs(cpi->common.fc, tile_ctxs, cdf_ptrs,
+                                cm->tile_rows * cm->tile_cols);
+#if CONFIG_PVQ
+    av1_average_tile_pvq_cdfs(cpi->common.fc, tile_ctxs,
+                              cm->tile_rows * cm->tile_cols);
+#endif  // CONFIG_PVQ
+#endif  // CONFIG_EC_ADAPT
+#if CONFIG_ADAPT_SCAN
+    av1_adapt_scan_order(cm);
+#endif  // CONFIG_ADAPT_SCAN
+  }
+
+  if (!frame_is_intra_only(cm)) {
+    if (cm->refresh_frame_context == REFRESH_FRAME_CONTEXT_BACKWARD) {
+      av1_adapt_inter_frame_probs(cm);
+      av1_adapt_mv_probs(cm, cm->allow_high_precision_mv);
+#if CONFIG_EC_ADAPT
+      av1_average_tile_inter_cdfs(&cpi->common, cpi->common.fc, tile_ctxs,
+                                  cdf_ptrs, cm->tile_rows * cm->tile_cols);
+      av1_average_tile_mv_cdfs(cpi->common.fc, tile_ctxs, cdf_ptrs,
+                               cm->tile_rows * cm->tile_cols);
+#endif
+    }
+  }
+
+  if (cpi->refresh_golden_frame == 1)
+    cpi->frame_flags |= FRAMEFLAGS_GOLDEN;
+  else
+    cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN;
+
+  if (cpi->refresh_alt_ref_frame == 1)
+    cpi->frame_flags |= FRAMEFLAGS_ALTREF;
+  else
+    cpi->frame_flags &= ~FRAMEFLAGS_ALTREF;
+
+#if CONFIG_EXT_REFS
+  if (cpi->refresh_bwd_ref_frame == 1)
+    cpi->frame_flags |= FRAMEFLAGS_BWDREF;
+  else
+    cpi->frame_flags &= ~FRAMEFLAGS_BWDREF;
+#endif  // CONFIG_EXT_REFS
+
+#if !CONFIG_EXT_REFS
+  cpi->ref_frame_flags = get_ref_frame_flags(cpi);
+#endif  // !CONFIG_EXT_REFS
+
+  cm->last_frame_type = cm->frame_type;
+
+#if CONFIG_XIPHRC
+  frame_type = cm->frame_type == KEY_FRAME ? OD_I_FRAME : OD_P_FRAME;
+
+  drop_this_frame =
+      od_enc_rc_update_state(&cpi->od_rc, *size << 3, cpi->refresh_golden_frame,
+                             cpi->refresh_alt_ref_frame, frame_type, 0);
+  if (drop_this_frame) {
+    av1_rc_postencode_update_drop_frame(cpi);
+    ++cm->current_video_frame;
+#if CONFIG_EC_ADAPT
+    aom_free(tile_ctxs);
+    aom_free(cdf_ptrs);
+#endif
+    return;
+  }
+#else   // !CONFIG_XIPHRC
+  av1_rc_postencode_update(cpi, *size);
+#endif  // CONFIG_XIPHRC
+
+#if 0
+  output_frame_level_debug_stats(cpi);
+#endif
+
+  if (cm->frame_type == KEY_FRAME) {
+    // Tell the caller that the frame was coded as a key frame
+    *frame_flags = cpi->frame_flags | FRAMEFLAGS_KEY;
+  } else {
+    *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY;
+  }
+
+  // Clear the one shot update flags for segmentation map and mode/ref loop
+  // filter deltas.
+  cm->seg.update_map = 0;
+  cm->seg.update_data = 0;
+  cm->lf.mode_ref_delta_update = 0;
+
+  // keep track of the last coded dimensions
+  cm->last_width = cm->width;
+  cm->last_height = cm->height;
+
+  // reset to normal state now that we are done.
+  if (!cm->show_existing_frame) cm->last_show_frame = cm->show_frame;
+
+  if (cm->show_frame) {
+#if CONFIG_EXT_REFS
+// TODO(zoeliu): We may only swamp mi and prev_mi for those frames that are
+// being used as reference.
+#endif  // CONFIG_EXT_REFS
+    av1_swap_mi_and_prev_mi(cm);
+    // Don't increment frame counters if this was an altref buffer
+    // update not a real frame
+    ++cm->current_video_frame;
+  }
+
+#if CONFIG_EXT_REFS
+  // NOTE: Shall not refer to any frame not used as reference.
+  if (cm->is_reference_frame)
+#endif  // CONFIG_EXT_REFS
+    cm->prev_frame = cm->cur_frame;
+#if CONFIG_EC_ADAPT
+  aom_free(tile_ctxs);
+  aom_free(cdf_ptrs);
+#endif
+}
+
+static void Pass0Encode(AV1_COMP *cpi, size_t *size, uint8_t *dest,
+                        int skip_adapt, unsigned int *frame_flags) {
+#if CONFIG_XIPHRC
+  int64_t ip_count;
+  int frame_type, is_golden, is_altref;
+
+  /* Not updated during init so update it here */
+  if (cpi->oxcf.rc_mode == AOM_Q) cpi->od_rc.quality = cpi->oxcf.cq_level;
+
+  frame_type = od_frame_type(&cpi->od_rc, cpi->od_rc.cur_frame, &is_golden,
+                             &is_altref, &ip_count);
+
+  if (frame_type == OD_I_FRAME) {
+    frame_type = KEY_FRAME;
+    cpi->frame_flags &= FRAMEFLAGS_KEY;
+  } else if (frame_type == OD_P_FRAME) {
+    frame_type = INTER_FRAME;
+  }
+
+  if (is_altref) {
+    cpi->refresh_alt_ref_frame = 1;
+    cpi->rc.source_alt_ref_active = 1;
+  }
+
+  cpi->refresh_golden_frame = is_golden;
+  cpi->common.frame_type = frame_type;
+  if (is_golden) cpi->frame_flags &= FRAMEFLAGS_GOLDEN;
+#else
+  if (cpi->oxcf.rc_mode == AOM_CBR) {
+    av1_rc_get_one_pass_cbr_params(cpi);
+  } else {
+    av1_rc_get_one_pass_vbr_params(cpi);
+  }
+#endif
+  encode_frame_to_data_rate(cpi, size, dest, skip_adapt, frame_flags);
+}
+
+#if !CONFIG_XIPHRC
+static void Pass2Encode(AV1_COMP *cpi, size_t *size, uint8_t *dest,
+                        unsigned int *frame_flags) {
+  encode_frame_to_data_rate(cpi, size, dest, 0, frame_flags);
+
+#if CONFIG_EXT_REFS
+  // Do not do post-encoding update for those frames that do not have a spot in
+  // a gf group, but note that an OVERLAY frame always has a spot in a gf group,
+  // even when show_existing_frame is used.
+  if (!cpi->common.show_existing_frame || cpi->rc.is_src_frame_alt_ref) {
+    av1_twopass_postencode_update(cpi);
+  }
+  check_show_existing_frame(cpi);
+#else
+  av1_twopass_postencode_update(cpi);
+#endif  // CONFIG_EXT_REFS
+}
+#endif
+
+static void init_ref_frame_bufs(AV1_COMMON *cm) {
+  int i;
+  BufferPool *const pool = cm->buffer_pool;
+  cm->new_fb_idx = INVALID_IDX;
+  for (i = 0; i < REF_FRAMES; ++i) {
+    cm->ref_frame_map[i] = INVALID_IDX;
+    pool->frame_bufs[i].ref_count = 0;
+  }
+}
+
+static void check_initial_width(AV1_COMP *cpi,
+#if CONFIG_HIGHBITDEPTH
+                                int use_highbitdepth,
+#endif
+                                int subsampling_x, int subsampling_y) {
+  AV1_COMMON *const cm = &cpi->common;
+
+  if (!cpi->initial_width ||
+#if CONFIG_HIGHBITDEPTH
+      cm->use_highbitdepth != use_highbitdepth ||
+#endif
+      cm->subsampling_x != subsampling_x ||
+      cm->subsampling_y != subsampling_y) {
+    cm->subsampling_x = subsampling_x;
+    cm->subsampling_y = subsampling_y;
+#if CONFIG_HIGHBITDEPTH
+    cm->use_highbitdepth = use_highbitdepth;
+#endif
+
+    alloc_raw_frame_buffers(cpi);
+    init_ref_frame_bufs(cm);
+    alloc_util_frame_buffers(cpi);
+
+    init_motion_estimation(cpi);  // TODO(agrange) This can be removed.
+
+    cpi->initial_width = cm->width;
+    cpi->initial_height = cm->height;
+    cpi->initial_mbs = cm->MBs;
+  }
+}
+
+int av1_receive_raw_frame(AV1_COMP *cpi, aom_enc_frame_flags_t frame_flags,
+                          YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
+                          int64_t end_time) {
+  AV1_COMMON *const cm = &cpi->common;
+  struct aom_usec_timer timer;
+  int res = 0;
+  const int subsampling_x = sd->subsampling_x;
+  const int subsampling_y = sd->subsampling_y;
+#if CONFIG_HIGHBITDEPTH
+  const int use_highbitdepth = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0;
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+  check_initial_width(cpi, use_highbitdepth, subsampling_x, subsampling_y);
+#else
+  check_initial_width(cpi, subsampling_x, subsampling_y);
+#endif  // CONFIG_HIGHBITDEPTH
+
+  aom_usec_timer_start(&timer);
+
+  if (av1_lookahead_push(cpi->lookahead, sd, time_stamp, end_time,
+#if CONFIG_HIGHBITDEPTH
+                         use_highbitdepth,
+#endif  // CONFIG_HIGHBITDEPTH
+                         frame_flags))
+    res = -1;
+  aom_usec_timer_mark(&timer);
+  cpi->time_receive_data += aom_usec_timer_elapsed(&timer);
+
+  if ((cm->profile == PROFILE_0 || cm->profile == PROFILE_2) &&
+      (subsampling_x != 1 || subsampling_y != 1)) {
+    aom_internal_error(&cm->error, AOM_CODEC_INVALID_PARAM,
+                       "Non-4:2:0 color format requires profile 1 or 3");
+    res = -1;
+  }
+  if ((cm->profile == PROFILE_1 || cm->profile == PROFILE_3) &&
+      (subsampling_x == 1 && subsampling_y == 1)) {
+    aom_internal_error(&cm->error, AOM_CODEC_INVALID_PARAM,
+                       "4:2:0 color format requires profile 0 or 2");
+    res = -1;
+  }
+
+  return res;
+}
+
+static int frame_is_reference(const AV1_COMP *cpi) {
+  const AV1_COMMON *cm = &cpi->common;
+
+  return cm->frame_type == KEY_FRAME || cpi->refresh_last_frame ||
+         cpi->refresh_golden_frame ||
+#if CONFIG_EXT_REFS
+         cpi->refresh_bwd_ref_frame ||
+#endif  // CONFIG_EXT_REFS
+         cpi->refresh_alt_ref_frame || !cm->error_resilient_mode ||
+         cm->lf.mode_ref_delta_update || cm->seg.update_map ||
+         cm->seg.update_data;
+}
+
+static void adjust_frame_rate(AV1_COMP *cpi,
+                              const struct lookahead_entry *source) {
+  int64_t this_duration;
+  int step = 0;
+
+  if (source->ts_start == cpi->first_time_stamp_ever) {
+    this_duration = source->ts_end - source->ts_start;
+    step = 1;
+  } else {
+    int64_t last_duration =
+        cpi->last_end_time_stamp_seen - cpi->last_time_stamp_seen;
+
+    this_duration = source->ts_end - cpi->last_end_time_stamp_seen;
+
+    // do a step update if the duration changes by 10%
+    if (last_duration)
+      step = (int)((this_duration - last_duration) * 10 / last_duration);
+  }
+
+  if (this_duration) {
+    if (step) {
+      av1_new_framerate(cpi, 10000000.0 / this_duration);
+    } else {
+      // Average this frame's rate into the last second's average
+      // frame rate. If we haven't seen 1 second yet, then average
+      // over the whole interval seen.
+      const double interval = AOMMIN(
+          (double)(source->ts_end - cpi->first_time_stamp_ever), 10000000.0);
+      double avg_duration = 10000000.0 / cpi->framerate;
+      avg_duration *= (interval - avg_duration + this_duration);
+      avg_duration /= interval;
+
+      av1_new_framerate(cpi, 10000000.0 / avg_duration);
+    }
+  }
+  cpi->last_time_stamp_seen = source->ts_start;
+  cpi->last_end_time_stamp_seen = source->ts_end;
+}
+
+// Returns 0 if this is not an alt ref else the offset of the source frame
+// used as the arf midpoint.
+static int get_arf_src_index(AV1_COMP *cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  int arf_src_index = 0;
+  if (is_altref_enabled(cpi)) {
+    if (cpi->oxcf.pass == 2) {
+      const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+      if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
+        arf_src_index = gf_group->arf_src_offset[gf_group->index];
+      }
+    } else if (rc->source_alt_ref_pending) {
+      arf_src_index = rc->frames_till_gf_update_due;
+    }
+  }
+  return arf_src_index;
+}
+
+#if CONFIG_EXT_REFS
+static int get_brf_src_index(AV1_COMP *cpi) {
+  int brf_src_index = 0;
+  const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+
+  // TODO(zoeliu): We need to add the check on the -bwd_ref command line setup
+  //               flag.
+  if (gf_group->bidir_pred_enabled[gf_group->index]) {
+    if (cpi->oxcf.pass == 2) {
+      if (gf_group->update_type[gf_group->index] == BRF_UPDATE)
+        brf_src_index = gf_group->brf_src_offset[gf_group->index];
+    } else {
+      // TODO(zoeliu): To re-visit the setup for this scenario
+      brf_src_index = cpi->rc.bipred_group_interval - 1;
+    }
+  }
+
+  return brf_src_index;
+}
+#endif  // CONFIG_EXT_REFS
+
+static void check_src_altref(AV1_COMP *cpi,
+                             const struct lookahead_entry *source) {
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  // If pass == 2, the parameters set here will be reset in
+  // av1_rc_get_second_pass_params()
+
+  if (cpi->oxcf.pass == 2) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    rc->is_src_frame_alt_ref =
+#if CONFIG_EXT_REFS
+        (gf_group->update_type[gf_group->index] == INTNL_OVERLAY_UPDATE) ||
+#endif  // CONFIG_EXT_REFS
+        (gf_group->update_type[gf_group->index] == OVERLAY_UPDATE);
+  } else {
+    rc->is_src_frame_alt_ref =
+        cpi->alt_ref_source && (source == cpi->alt_ref_source);
+  }
+
+  if (rc->is_src_frame_alt_ref) {
+    // Current frame is an ARF overlay frame.
+    cpi->alt_ref_source = NULL;
+
+    // Don't refresh the last buffer for an ARF overlay frame. It will
+    // become the GF so preserve last as an alternative prediction option.
+    cpi->refresh_last_frame = 0;
+  }
+}
+
+#if CONFIG_INTERNAL_STATS
+extern double av1_get_blockiness(const unsigned char *img1, int img1_pitch,
+                                 const unsigned char *img2, int img2_pitch,
+                                 int width, int height);
+
+static void adjust_image_stat(double y, double u, double v, double all,
+                              ImageStat *s) {
+  s->stat[Y] += y;
+  s->stat[U] += u;
+  s->stat[V] += v;
+  s->stat[ALL] += all;
+  s->worst = AOMMIN(s->worst, all);
+}
+
+static void compute_internal_stats(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  double samples = 0.0;
+  uint32_t in_bit_depth = 8;
+  uint32_t bit_depth = 8;
+
+#if CONFIG_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    in_bit_depth = cpi->oxcf.input_bit_depth;
+    bit_depth = cm->bit_depth;
+  }
+#endif
+  if (cm->show_frame) {
+    const YV12_BUFFER_CONFIG *orig = cpi->source;
+    const YV12_BUFFER_CONFIG *recon = cpi->common.frame_to_show;
+    double y, u, v, frame_all;
+
+    cpi->count++;
+    if (cpi->b_calculate_psnr) {
+      PSNR_STATS psnr;
+      double frame_ssim2 = 0.0, weight = 0.0;
+      aom_clear_system_state();
+// TODO(yaowu): unify these two versions into one.
+#if CONFIG_HIGHBITDEPTH
+      aom_calc_highbd_psnr(orig, recon, &psnr, bit_depth, in_bit_depth);
+#else
+      aom_calc_psnr(orig, recon, &psnr);
+#endif  // CONFIG_HIGHBITDEPTH
+
+      adjust_image_stat(psnr.psnr[1], psnr.psnr[2], psnr.psnr[3], psnr.psnr[0],
+                        &cpi->psnr);
+      cpi->total_sq_error += psnr.sse[0];
+      cpi->total_samples += psnr.samples[0];
+      samples = psnr.samples[0];
+// TODO(yaowu): unify these two versions into one.
+#if CONFIG_HIGHBITDEPTH
+      if (cm->use_highbitdepth)
+        frame_ssim2 =
+            aom_highbd_calc_ssim(orig, recon, &weight, bit_depth, in_bit_depth);
+      else
+        frame_ssim2 = aom_calc_ssim(orig, recon, &weight);
+#else
+      frame_ssim2 = aom_calc_ssim(orig, recon, &weight);
+#endif  // CONFIG_HIGHBITDEPTH
+
+      cpi->worst_ssim = AOMMIN(cpi->worst_ssim, frame_ssim2);
+      cpi->summed_quality += frame_ssim2 * weight;
+      cpi->summed_weights += weight;
+
+#if 0
+      {
+        FILE *f = fopen("q_used.stt", "a");
+        fprintf(f, "%5d : Y%f7.3:U%f7.3:V%f7.3:F%f7.3:S%7.3f\n",
+                cpi->common.current_video_frame, y2, u2, v2,
+                frame_psnr2, frame_ssim2);
+        fclose(f);
+      }
+#endif
+    }
+    if (cpi->b_calculate_blockiness) {
+#if CONFIG_HIGHBITDEPTH
+      if (!cm->use_highbitdepth)
+#endif
+      {
+        const double frame_blockiness =
+            av1_get_blockiness(orig->y_buffer, orig->y_stride, recon->y_buffer,
+                               recon->y_stride, orig->y_width, orig->y_height);
+        cpi->worst_blockiness = AOMMAX(cpi->worst_blockiness, frame_blockiness);
+        cpi->total_blockiness += frame_blockiness;
+      }
+
+      if (cpi->b_calculate_consistency) {
+#if CONFIG_HIGHBITDEPTH
+        if (!cm->use_highbitdepth)
+#endif
+        {
+          const double this_inconsistency = aom_get_ssim_metrics(
+              orig->y_buffer, orig->y_stride, recon->y_buffer, recon->y_stride,
+              orig->y_width, orig->y_height, cpi->ssim_vars, &cpi->metrics, 1);
+
+          const double peak = (double)((1 << in_bit_depth) - 1);
+          const double consistency =
+              aom_sse_to_psnr(samples, peak, cpi->total_inconsistency);
+          if (consistency > 0.0)
+            cpi->worst_consistency =
+                AOMMIN(cpi->worst_consistency, consistency);
+          cpi->total_inconsistency += this_inconsistency;
+        }
+      }
+    }
+
+    frame_all =
+        aom_calc_fastssim(orig, recon, &y, &u, &v, bit_depth, in_bit_depth);
+    adjust_image_stat(y, u, v, frame_all, &cpi->fastssim);
+    frame_all = aom_psnrhvs(orig, recon, &y, &u, &v, bit_depth, in_bit_depth);
+    adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs);
+  }
+}
+#endif  // CONFIG_INTERNAL_STATS
+
+int av1_get_compressed_data(AV1_COMP *cpi, unsigned int *frame_flags,
+                            size_t *size, uint8_t *dest, int64_t *time_stamp,
+                            int64_t *time_end, int flush) {
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  AV1_COMMON *const cm = &cpi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  RATE_CONTROL *const rc = &cpi->rc;
+  struct aom_usec_timer cmptimer;
+  YV12_BUFFER_CONFIG *force_src_buffer = NULL;
+  struct lookahead_entry *last_source = NULL;
+  struct lookahead_entry *source = NULL;
+  int arf_src_index;
+#if CONFIG_EXT_REFS
+  int brf_src_index;
+#endif  // CONFIG_EXT_REFS
+  int i;
+
+#if CONFIG_XIPHRC
+  cpi->od_rc.end_of_input = flush;
+#endif
+
+#if CONFIG_BITSTREAM_DEBUG
+  assert(cpi->oxcf.max_threads == 0 &&
+         "bitstream debug tool does not support multithreading");
+  bitstream_queue_record_write();
+  bitstream_queue_set_frame_write(cm->current_video_frame * 2 + cm->show_frame);
+#endif
+
+  aom_usec_timer_start(&cmptimer);
+
+  av1_set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV);
+
+  // Is multi-arf enabled.
+  // Note that at the moment multi_arf is only configured for 2 pass VBR
+  if ((oxcf->pass == 2) && (cpi->oxcf.enable_auto_arf > 1))
+    cpi->multi_arf_allowed = 1;
+  else
+    cpi->multi_arf_allowed = 0;
+
+  // Normal defaults
+  cm->reset_frame_context = RESET_FRAME_CONTEXT_NONE;
+  cm->refresh_frame_context =
+      (oxcf->error_resilient_mode || oxcf->frame_parallel_decoding_mode)
+          ? REFRESH_FRAME_CONTEXT_FORWARD
+          : REFRESH_FRAME_CONTEXT_BACKWARD;
+
+  cpi->refresh_last_frame = 1;
+  cpi->refresh_golden_frame = 0;
+#if CONFIG_EXT_REFS
+  cpi->refresh_bwd_ref_frame = 0;
+#endif  // CONFIG_EXT_REFS
+  cpi->refresh_alt_ref_frame = 0;
+
+#if CONFIG_EXT_REFS && !CONFIG_XIPHRC
+  if (oxcf->pass == 2 && cm->show_existing_frame) {
+    // Manage the source buffer and flush out the source frame that has been
+    // coded already; Also get prepared for PSNR calculation if needed.
+    if ((source = av1_lookahead_pop(cpi->lookahead, flush)) == NULL) {
+      *size = 0;
+      return -1;
+    }
+    cpi->source = &source->img;
+    // TODO(zoeliu): To track down to determine whether it's needed to adjust
+    // the frame rate.
+    *time_stamp = source->ts_start;
+    *time_end = source->ts_end;
+
+    // We need to adjust frame rate for an overlay frame
+    if (cpi->rc.is_src_frame_alt_ref) adjust_frame_rate(cpi, source);
+
+    // Find a free buffer for the new frame, releasing the reference previously
+    // held.
+    if (cm->new_fb_idx != INVALID_IDX) {
+      --pool->frame_bufs[cm->new_fb_idx].ref_count;
+    }
+    cm->new_fb_idx = get_free_fb(cm);
+
+    if (cm->new_fb_idx == INVALID_IDX) return -1;
+
+    // Clear down mmx registers
+    aom_clear_system_state();
+
+    // Start with a 0 size frame.
+    *size = 0;
+
+    // We need to update the gf_group for show_existing overlay frame
+    if (cpi->rc.is_src_frame_alt_ref) av1_rc_get_second_pass_params(cpi);
+
+    Pass2Encode(cpi, size, dest, frame_flags);
+
+    if (cpi->b_calculate_psnr) generate_psnr_packet(cpi);
+
+#if CONFIG_INTERNAL_STATS
+    compute_internal_stats(cpi);
+    cpi->bytes += (int)(*size);
+#endif  // CONFIG_INTERNAL_STATS
+
+    // Clear down mmx registers
+    aom_clear_system_state();
+
+    cm->show_existing_frame = 0;
+    return 0;
+  }
+#endif  // CONFIG_EXT_REFS && !CONFIG_XIPHRC
+
+  // Should we encode an arf frame.
+  arf_src_index = get_arf_src_index(cpi);
+  if (arf_src_index) {
+    for (i = 0; i <= arf_src_index; ++i) {
+      struct lookahead_entry *e = av1_lookahead_peek(cpi->lookahead, i);
+      // Avoid creating an alt-ref if there's a forced keyframe pending.
+      if (e == NULL) {
+        break;
+      } else if (e->flags == AOM_EFLAG_FORCE_KF) {
+        arf_src_index = 0;
+        flush = 1;
+        break;
+      }
+    }
+  }
+
+  if (arf_src_index) {
+    assert(arf_src_index <= rc->frames_to_key);
+
+    if ((source = av1_lookahead_peek(cpi->lookahead, arf_src_index)) != NULL) {
+      cpi->alt_ref_source = source;
+
+      if (oxcf->arnr_max_frames > 0) {
+        // Produce the filtered ARF frame.
+        av1_temporal_filter(cpi, arf_src_index);
+        aom_extend_frame_borders(&cpi->alt_ref_buffer);
+        force_src_buffer = &cpi->alt_ref_buffer;
+      }
+
+      cm->show_frame = 0;
+      cm->intra_only = 0;
+      cpi->refresh_alt_ref_frame = 1;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_last_frame = 0;
+      rc->is_src_frame_alt_ref = 0;
+    }
+    rc->source_alt_ref_pending = 0;
+  }
+
+#if CONFIG_EXT_REFS
+  rc->is_bwd_ref_frame = 0;
+  brf_src_index = get_brf_src_index(cpi);
+  if (brf_src_index) {
+    assert(brf_src_index <= rc->frames_to_key);
+    if ((source = av1_lookahead_peek(cpi->lookahead, brf_src_index)) != NULL) {
+      cm->show_frame = 0;
+      cm->intra_only = 0;
+
+      cpi->refresh_bwd_ref_frame = 1;
+      cpi->refresh_last_frame = 0;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_alt_ref_frame = 0;
+
+      rc->is_bwd_ref_frame = 1;
+    }
+  }
+#endif  // CONFIG_EXT_REFS
+
+  if (!source) {
+    // Get last frame source.
+    if (cm->current_video_frame > 0) {
+      if ((last_source = av1_lookahead_peek(cpi->lookahead, -1)) == NULL)
+        return -1;
+    }
+
+    // Read in the source frame.
+    source = av1_lookahead_pop(cpi->lookahead, flush);
+
+    if (source != NULL) {
+      cm->show_frame = 1;
+      cm->intra_only = 0;
+
+      // Check to see if the frame should be encoded as an arf overlay.
+      check_src_altref(cpi, source);
+    }
+  }
+
+  if (source) {
+    cpi->un_scaled_source = cpi->source =
+        force_src_buffer ? force_src_buffer : &source->img;
+
+    cpi->unscaled_last_source = last_source != NULL ? &last_source->img : NULL;
+
+    *time_stamp = source->ts_start;
+    *time_end = source->ts_end;
+    *frame_flags = (source->flags & AOM_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
+
+  } else {
+    *size = 0;
+    if (flush && oxcf->pass == 1 && !cpi->twopass.first_pass_done) {
+#if CONFIG_XIPHRC
+      od_enc_rc_2pass_out(&cpi->od_rc, cpi->output_pkt_list, 1);
+#else
+      av1_end_first_pass(cpi); /* get last stats packet */
+#endif
+      cpi->twopass.first_pass_done = 1;
+    }
+    return -1;
+  }
+
+  if (source->ts_start < cpi->first_time_stamp_ever) {
+    cpi->first_time_stamp_ever = source->ts_start;
+    cpi->last_end_time_stamp_seen = source->ts_start;
+  }
+
+  // Clear down mmx registers
+  aom_clear_system_state();
+
+  // adjust frame rates based on timestamps given
+  if (cm->show_frame) adjust_frame_rate(cpi, source);
+
+  // Find a free buffer for the new frame, releasing the reference previously
+  // held.
+  if (cm->new_fb_idx != INVALID_IDX) {
+    --pool->frame_bufs[cm->new_fb_idx].ref_count;
+  }
+  cm->new_fb_idx = get_free_fb(cm);
+
+  if (cm->new_fb_idx == INVALID_IDX) return -1;
+
+  cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
+
+#if CONFIG_EXT_REFS
+  if (oxcf->pass == 2) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    cpi->alt_fb_idx = cpi->arf_map[gf_group->arf_ref_idx[gf_group->index]];
+  }
+#else
+  if (cpi->multi_arf_allowed) {
+    if (cm->frame_type == KEY_FRAME) {
+      init_buffer_indices(cpi);
+    } else if (oxcf->pass == 2) {
+      const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+      cpi->alt_fb_idx = gf_group->arf_ref_idx[gf_group->index];
+    }
+  }
+#endif  // CONFIG_EXT_REFS
+
+  // Start with a 0 size frame.
+  *size = 0;
+
+  cpi->frame_flags = *frame_flags;
+
+  if (oxcf->pass == 2) {
+#if CONFIG_XIPHRC
+    if (od_enc_rc_2pass_in(&cpi->od_rc) < 0) return -1;
+  }
+#else
+    av1_rc_get_second_pass_params(cpi);
+  } else if (oxcf->pass == 1) {
+    set_frame_size(cpi);
+  }
+#endif
+
+  if (cpi->oxcf.pass != 0 || frame_is_intra_only(cm) == 1) {
+    for (i = 0; i < TOTAL_REFS_PER_FRAME; ++i)
+      cpi->scaled_ref_idx[i] = INVALID_IDX;
+  }
+
+#if CONFIG_AOM_QM
+  cm->using_qmatrix = cpi->oxcf.using_qm;
+  cm->min_qmlevel = cpi->oxcf.qm_minlevel;
+  cm->max_qmlevel = cpi->oxcf.qm_maxlevel;
+#endif
+
+#if CONFIG_REFERENCE_BUFFER
+  if (*time_stamp == 0) {
+    cpi->common.current_frame_id = -1;
+  }
+#endif
+
+#if CONFIG_XIPHRC
+  if (oxcf->pass == 1) {
+    size_t tmp;
+    if (cpi->od_rc.cur_frame == 0) Pass0Encode(cpi, &tmp, dest, 1, frame_flags);
+    cpi->od_rc.firstpass_quant = cpi->od_rc.target_quantizer;
+    Pass0Encode(cpi, &tmp, dest, 0, frame_flags);
+    od_enc_rc_2pass_out(&cpi->od_rc, cpi->output_pkt_list, 0);
+  } else if (oxcf->pass == 2) {
+    Pass0Encode(cpi, size, dest, 0, frame_flags);
+  } else {
+    if (cpi->od_rc.cur_frame == 0) {
+      size_t tmp;
+      Pass0Encode(cpi, &tmp, dest, 1, frame_flags);
+    }
+    Pass0Encode(cpi, size, dest, 0, frame_flags);
+  }
+#else
+  if (oxcf->pass == 1) {
+    cpi->td.mb.e_mbd.lossless[0] = is_lossless_requested(oxcf);
+    av1_first_pass(cpi, source);
+  } else if (oxcf->pass == 2) {
+    Pass2Encode(cpi, size, dest, frame_flags);
+  } else {
+    // One pass encode
+    Pass0Encode(cpi, size, dest, 0, frame_flags);
+  }
+#endif
+
+  if (!cm->error_resilient_mode)
+    cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+
+  // No frame encoded, or frame was dropped, release scaled references.
+  if ((*size == 0) && (frame_is_intra_only(cm) == 0)) {
+    release_scaled_references(cpi);
+  }
+
+  if (*size > 0) {
+    cpi->droppable = !frame_is_reference(cpi);
+  }
+
+  aom_usec_timer_mark(&cmptimer);
+  cpi->time_compress_data += aom_usec_timer_elapsed(&cmptimer);
+
+  if (cpi->b_calculate_psnr && oxcf->pass != 1 && cm->show_frame)
+    generate_psnr_packet(cpi);
+
+#if CONFIG_INTERNAL_STATS
+  if (oxcf->pass != 1) {
+    compute_internal_stats(cpi);
+    cpi->bytes += (int)(*size);
+  }
+#endif  // CONFIG_INTERNAL_STATS
+
+#if CONFIG_XIPHRC
+  cpi->od_rc.cur_frame++;
+#endif
+
+  aom_clear_system_state();
+
+  return 0;
+}
+
+int av1_get_preview_raw_frame(AV1_COMP *cpi, YV12_BUFFER_CONFIG *dest) {
+  AV1_COMMON *cm = &cpi->common;
+  if (!cm->show_frame) {
+    return -1;
+  } else {
+    int ret;
+    if (cm->frame_to_show) {
+      *dest = *cm->frame_to_show;
+      dest->y_width = cm->width;
+      dest->y_height = cm->height;
+      dest->uv_width = cm->width >> cm->subsampling_x;
+      dest->uv_height = cm->height >> cm->subsampling_y;
+      ret = 0;
+    } else {
+      ret = -1;
+    }
+    aom_clear_system_state();
+    return ret;
+  }
+}
+
+int av1_get_last_show_frame(AV1_COMP *cpi, YV12_BUFFER_CONFIG *frame) {
+  if (cpi->last_show_frame_buf_idx == INVALID_IDX) return -1;
+
+  *frame =
+      cpi->common.buffer_pool->frame_bufs[cpi->last_show_frame_buf_idx].buf;
+  return 0;
+}
+
+int av1_set_internal_size(AV1_COMP *cpi, AOM_SCALING horiz_mode,
+                          AOM_SCALING vert_mode) {
+  AV1_COMMON *cm = &cpi->common;
+  int hr = 0, hs = 0, vr = 0, vs = 0;
+
+  if (horiz_mode > ONETWO || vert_mode > ONETWO) return -1;
+
+  Scale2Ratio(horiz_mode, &hr, &hs);
+  Scale2Ratio(vert_mode, &vr, &vs);
+
+  // always go to the next whole number
+  cm->width = (hs - 1 + cpi->oxcf.width * hr) / hs;
+  cm->height = (vs - 1 + cpi->oxcf.height * vr) / vs;
+  assert(cm->width <= cpi->initial_width);
+  assert(cm->height <= cpi->initial_height);
+
+  update_frame_size(cpi);
+
+  return 0;
+}
+
+int av1_set_size_literal(AV1_COMP *cpi, unsigned int width,
+                         unsigned int height) {
+  AV1_COMMON *cm = &cpi->common;
+#if CONFIG_HIGHBITDEPTH
+  check_initial_width(cpi, cm->use_highbitdepth, 1, 1);
+#else
+  check_initial_width(cpi, 1, 1);
+#endif  // CONFIG_HIGHBITDEPTH
+
+  if (width) {
+    cm->width = width;
+    if (cm->width > cpi->initial_width) {
+      cm->width = cpi->initial_width;
+      printf("Warning: Desired width too large, changed to %d\n", cm->width);
+    }
+  }
+
+  if (height) {
+    cm->height = height;
+    if (cm->height > cpi->initial_height) {
+      cm->height = cpi->initial_height;
+      printf("Warning: Desired height too large, changed to %d\n", cm->height);
+    }
+  }
+  assert(cm->width <= cpi->initial_width);
+  assert(cm->height <= cpi->initial_height);
+
+  update_frame_size(cpi);
+
+  return 0;
+}
+
+int av1_get_quantizer(AV1_COMP *cpi) { return cpi->common.base_qindex; }
+
+void av1_apply_encoding_flags(AV1_COMP *cpi, aom_enc_frame_flags_t flags) {
+  if (flags &
+      (AOM_EFLAG_NO_REF_LAST | AOM_EFLAG_NO_REF_GF | AOM_EFLAG_NO_REF_ARF)) {
+    int ref = AOM_REFFRAME_ALL;
+
+    if (flags & AOM_EFLAG_NO_REF_LAST) {
+      ref ^= AOM_LAST_FLAG;
+#if CONFIG_EXT_REFS
+      ref ^= AOM_LAST2_FLAG;
+      ref ^= AOM_LAST3_FLAG;
+#endif  // CONFIG_EXT_REFS
+    }
+
+    if (flags & AOM_EFLAG_NO_REF_GF) ref ^= AOM_GOLD_FLAG;
+
+    if (flags & AOM_EFLAG_NO_REF_ARF) ref ^= AOM_ALT_FLAG;
+
+    av1_use_as_reference(cpi, ref);
+  }
+
+  if (flags &
+      (AOM_EFLAG_NO_UPD_LAST | AOM_EFLAG_NO_UPD_GF | AOM_EFLAG_NO_UPD_ARF |
+       AOM_EFLAG_FORCE_GF | AOM_EFLAG_FORCE_ARF)) {
+    int upd = AOM_REFFRAME_ALL;
+
+    if (flags & AOM_EFLAG_NO_UPD_LAST) {
+      upd ^= AOM_LAST_FLAG;
+#if CONFIG_EXT_REFS
+      upd ^= AOM_LAST2_FLAG;
+      upd ^= AOM_LAST3_FLAG;
+#endif  // CONFIG_EXT_REFS
+    }
+
+    if (flags & AOM_EFLAG_NO_UPD_GF) upd ^= AOM_GOLD_FLAG;
+
+    if (flags & AOM_EFLAG_NO_UPD_ARF) upd ^= AOM_ALT_FLAG;
+
+    av1_update_reference(cpi, upd);
+  }
+
+  if (flags & AOM_EFLAG_NO_UPD_ENTROPY) {
+    av1_update_entropy(cpi, 0);
+  }
+}
diff --git a/third_party/aom/av1/encoder/encoder.h b/third_party/aom/av1/encoder/encoder.h
new file mode 100644
index 0000000000..4e7aef8fcd
--- /dev/null
+++ b/third_party/aom/av1/encoder/encoder.h
@@ -0,0 +1,883 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_ENCODER_H_
+#define AV1_ENCODER_ENCODER_H_
+
+#include <stdio.h>
+
+#include "./aom_config.h"
+#include "aom/aomcx.h"
+
+#include "av1/common/alloccommon.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/thread_common.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/encoder/aq_cyclicrefresh.h"
+#if CONFIG_ANS
+#include "aom_dsp/ans.h"
+#include "aom_dsp/buf_ans.h"
+#endif
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/context_tree.h"
+#include "av1/encoder/encodemb.h"
+#include "av1/encoder/firstpass.h"
+#include "av1/encoder/lookahead.h"
+#include "av1/encoder/mbgraph.h"
+#include "av1/encoder/mcomp.h"
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/rd.h"
+#include "av1/encoder/speed_features.h"
+#include "av1/encoder/tokenize.h"
+#include "av1/encoder/variance_tree.h"
+#if CONFIG_XIPHRC
+#include "av1/encoder/ratectrl_xiph.h"
+#endif
+
+#if CONFIG_INTERNAL_STATS
+#include "aom_dsp/ssim.h"
+#endif
+#include "aom_dsp/variance.h"
+#include "aom/internal/aom_codec_internal.h"
+#include "aom_util/aom_thread.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  int nmvjointcost[MV_JOINTS];
+  int nmvcosts[2][MV_VALS];
+  int nmvcosts_hp[2][MV_VALS];
+
+#if CONFIG_REF_MV
+  int nmv_vec_cost[NMV_CONTEXTS][MV_JOINTS];
+  int nmv_costs[NMV_CONTEXTS][2][MV_VALS];
+  int nmv_costs_hp[NMV_CONTEXTS][2][MV_VALS];
+#endif
+
+  // 0 = Intra, Last, GF, ARF
+  signed char last_ref_lf_deltas[TOTAL_REFS_PER_FRAME];
+  // 0 = ZERO_MV, MV
+  signed char last_mode_lf_deltas[MAX_MODE_LF_DELTAS];
+
+  FRAME_CONTEXT fc;
+} CODING_CONTEXT;
+
+typedef enum {
+  // regular inter frame
+  REGULAR_FRAME = 0,
+  // alternate reference frame
+  ARF_FRAME = 1,
+  // overlay frame
+  OVERLAY_FRAME = 2,
+  // golden frame
+  GLD_FRAME = 3,
+#if CONFIG_EXT_REFS
+  // backward reference frame
+  BRF_FRAME = 4,
+  // extra alternate reference frame
+  EXT_ARF_FRAME = 5
+#endif
+} FRAME_CONTEXT_INDEX;
+
+typedef enum {
+  NORMAL = 0,
+  FOURFIVE = 1,
+  THREEFIVE = 2,
+  ONETWO = 3
+} AOM_SCALING;
+
+typedef enum {
+  // Good Quality Fast Encoding. The encoder balances quality with the amount of
+  // time it takes to encode the output. Speed setting controls how fast.
+  GOOD
+} MODE;
+
+typedef enum {
+  FRAMEFLAGS_KEY = 1 << 0,
+  FRAMEFLAGS_GOLDEN = 1 << 1,
+#if CONFIG_EXT_REFS
+  FRAMEFLAGS_BWDREF = 1 << 2,
+  FRAMEFLAGS_ALTREF = 1 << 3,
+#else
+  FRAMEFLAGS_ALTREF = 1 << 2,
+#endif  // CONFIG_EXT_REFS
+} FRAMETYPE_FLAGS;
+
+typedef enum {
+  NO_AQ = 0,
+  VARIANCE_AQ = 1,
+  COMPLEXITY_AQ = 2,
+  CYCLIC_REFRESH_AQ = 3,
+#if CONFIG_DELTA_Q && !CONFIG_EXT_DELTA_Q
+  DELTA_AQ = 4,
+#endif
+  AQ_MODE_COUNT  // This should always be the last member of the enum
+} AQ_MODE;
+#if CONFIG_EXT_DELTA_Q
+typedef enum {
+  NO_DELTA_Q = 0,
+  DELTA_Q_ONLY = 1,
+  DELTA_Q_LF = 2,
+  DELTAQ_MODE_COUNT  // This should always be the last member of the enum
+} DELTAQ_MODE;
+#endif
+typedef enum {
+  RESIZE_NONE = 0,    // No frame resizing allowed.
+  RESIZE_FIXED = 1,   // All frames are coded at the specified dimension.
+  RESIZE_DYNAMIC = 2  // Coded size of each frame is determined by the codec.
+} RESIZE_TYPE;
+
+typedef struct AV1EncoderConfig {
+  BITSTREAM_PROFILE profile;
+  aom_bit_depth_t bit_depth;     // Codec bit-depth.
+  int width;                     // width of data passed to the compressor
+  int height;                    // height of data passed to the compressor
+  unsigned int input_bit_depth;  // Input bit depth.
+  double init_framerate;         // set to passed in framerate
+  int64_t target_bandwidth;      // bandwidth to be used in bits per second
+
+  int noise_sensitivity;  // pre processing blur: recommendation 0
+  int sharpness;          // sharpening output: recommendation 0:
+  int speed;
+  // maximum allowed bitrate for any intra frame in % of bitrate target.
+  unsigned int rc_max_intra_bitrate_pct;
+  // maximum allowed bitrate for any inter frame in % of bitrate target.
+  unsigned int rc_max_inter_bitrate_pct;
+  // percent of rate boost for golden frame in CBR mode.
+  unsigned int gf_cbr_boost_pct;
+
+  MODE mode;
+  int pass;
+
+  // Key Framing Operations
+  int auto_key;  // autodetect cut scenes and set the keyframes
+  int key_freq;  // maximum distance to key frame.
+
+  int lag_in_frames;  // how many frames lag before we start encoding
+
+  // ----------------------------------------------------------------
+  // DATARATE CONTROL OPTIONS
+
+  // vbr, cbr, constrained quality or constant quality
+  enum aom_rc_mode rc_mode;
+
+  // buffer targeting aggressiveness
+  int under_shoot_pct;
+  int over_shoot_pct;
+
+  // buffering parameters
+  int64_t starting_buffer_level_ms;
+  int64_t optimal_buffer_level_ms;
+  int64_t maximum_buffer_size_ms;
+
+  // Frame drop threshold.
+  int drop_frames_water_mark;
+
+  // controlling quality
+  int fixed_q;
+  int worst_allowed_q;
+  int best_allowed_q;
+  int cq_level;
+  AQ_MODE aq_mode;  // Adaptive Quantization mode
+#if CONFIG_EXT_DELTA_Q
+  DELTAQ_MODE deltaq_mode;
+#endif
+#if CONFIG_AOM_QM
+  int using_qm;
+  int qm_minlevel;
+  int qm_maxlevel;
+#endif
+#if CONFIG_TILE_GROUPS
+  unsigned int num_tile_groups;
+  unsigned int mtu;
+#endif
+
+#if CONFIG_TEMPMV_SIGNALING
+  unsigned int disable_tempmv;
+#endif
+  // Internal frame size scaling.
+  RESIZE_TYPE resize_mode;
+  int scaled_frame_width;
+  int scaled_frame_height;
+
+  // Enable feature to reduce the frame quantization every x frames.
+  int frame_periodic_boost;
+
+  // two pass datarate control
+  int two_pass_vbrbias;  // two pass datarate control tweaks
+  int two_pass_vbrmin_section;
+  int two_pass_vbrmax_section;
+  // END DATARATE CONTROL OPTIONS
+  // ----------------------------------------------------------------
+
+  int enable_auto_arf;
+#if CONFIG_EXT_REFS
+  int enable_auto_brf;  // (b)ackward (r)ef (f)rame
+#endif                  // CONFIG_EXT_REFS
+
+  /* Bitfield defining the error resiliency features to enable.
+   * Can provide decodable frames after losses in previous
+   * frames and decodable partitions after losses in the same frame.
+   */
+  unsigned int error_resilient_mode;
+
+  /* Bitfield defining the parallel decoding mode where the
+   * decoding in successive frames may be conducted in parallel
+   * just by decoding the frame headers.
+   */
+  unsigned int frame_parallel_decoding_mode;
+
+  int arnr_max_frames;
+  int arnr_strength;
+
+  int min_gf_interval;
+  int max_gf_interval;
+
+  int tile_columns;
+  int tile_rows;
+#if CONFIG_DEPENDENT_HORZTILES
+  int dependent_horz_tiles;
+#endif
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+  int loop_filter_across_tiles_enabled;
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+
+  int max_threads;
+
+  aom_fixed_buf_t two_pass_stats_in;
+  struct aom_codec_pkt_list *output_pkt_list;
+
+#if CONFIG_FP_MB_STATS
+  aom_fixed_buf_t firstpass_mb_stats_in;
+#endif
+
+  aom_tune_metric tuning;
+  aom_tune_content content;
+#if CONFIG_HIGHBITDEPTH
+  int use_highbitdepth;
+#endif
+  aom_color_space_t color_space;
+  int color_range;
+  int render_width;
+  int render_height;
+
+#if CONFIG_EXT_PARTITION
+  aom_superblock_size_t superblock_size;
+#endif  // CONFIG_EXT_PARTITION
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  int ans_window_size_log2;
+#endif  // CONFIG_ANS && ANS_MAX_SYMBOLS
+#if CONFIG_EXT_TILE
+  unsigned int tile_encoding_mode;
+#endif  // CONFIG_EXT_TILE
+
+  unsigned int motion_vector_unit_test;
+} AV1EncoderConfig;
+
+static INLINE int is_lossless_requested(const AV1EncoderConfig *cfg) {
+  return cfg->best_allowed_q == 0 && cfg->worst_allowed_q == 0;
+}
+
+// TODO(jingning) All spatially adaptive variables should go to TileDataEnc.
+typedef struct TileDataEnc {
+  TileInfo tile_info;
+  int thresh_freq_fact[BLOCK_SIZES][MAX_MODES];
+  int mode_map[BLOCK_SIZES][MAX_MODES];
+  int m_search_count;
+  int ex_search_count;
+#if CONFIG_PVQ
+  PVQ_QUEUE pvq_q;
+#endif
+#if CONFIG_CFL
+  CFL_CTX cfl;
+#endif
+#if CONFIG_EC_ADAPT
+  DECLARE_ALIGNED(16, FRAME_CONTEXT, tctx);
+#endif
+} TileDataEnc;
+
+typedef struct RD_COUNTS {
+  av1_coeff_count coef_counts[TX_SIZES][PLANE_TYPES];
+  int64_t comp_pred_diff[REFERENCE_MODES];
+#if CONFIG_GLOBAL_MOTION
+  // Stores number of 4x4 blocks using global motion per reference frame.
+  int global_motion_used[TOTAL_REFS_PER_FRAME];
+#endif  // CONFIG_GLOBAL_MOTION
+} RD_COUNTS;
+
+typedef struct ThreadData {
+  MACROBLOCK mb;
+  RD_COUNTS rd_counts;
+  FRAME_COUNTS *counts;
+
+  PICK_MODE_CONTEXT *leaf_tree;
+  PC_TREE *pc_tree;
+  PC_TREE *pc_root[MAX_MIB_SIZE_LOG2 - MIN_MIB_SIZE_LOG2 + 1];
+
+  VAR_TREE *var_tree;
+  VAR_TREE *var_root[MAX_MIB_SIZE_LOG2 - MIN_MIB_SIZE_LOG2 + 1];
+} ThreadData;
+
+struct EncWorkerData;
+
+typedef struct ActiveMap {
+  int enabled;
+  int update;
+  unsigned char *map;
+} ActiveMap;
+
+#define NUM_STAT_TYPES 4  // types of stats: Y, U, V and ALL
+
+typedef struct IMAGE_STAT {
+  double stat[NUM_STAT_TYPES];
+  double worst;
+} ImageStat;
+
+#undef NUM_STAT_TYPES
+
+typedef struct {
+  int ref_count;
+  YV12_BUFFER_CONFIG buf;
+} EncRefCntBuffer;
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+typedef struct SUBFRAME_STATS {
+  av1_coeff_probs_model coef_probs_buf[COEF_PROBS_BUFS][TX_SIZES][PLANE_TYPES];
+  av1_coeff_count coef_counts_buf[COEF_PROBS_BUFS][TX_SIZES][PLANE_TYPES];
+  unsigned int eob_counts_buf[COEF_PROBS_BUFS][TX_SIZES][PLANE_TYPES][REF_TYPES]
+                             [COEF_BANDS][COEFF_CONTEXTS];
+  av1_coeff_probs_model enc_starting_coef_probs[TX_SIZES][PLANE_TYPES];
+} SUBFRAME_STATS;
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+typedef struct TileBufferEnc {
+  uint8_t *data;
+  size_t size;
+} TileBufferEnc;
+
+typedef struct AV1_COMP {
+  QUANTS quants;
+  ThreadData td;
+  MB_MODE_INFO_EXT *mbmi_ext_base;
+  DECLARE_ALIGNED(16, int16_t, y_dequant[QINDEX_RANGE][8]);   // 8: SIMD width
+  DECLARE_ALIGNED(16, int16_t, uv_dequant[QINDEX_RANGE][8]);  // 8: SIMD width
+#if CONFIG_NEW_QUANT
+  DECLARE_ALIGNED(16, dequant_val_type_nuq,
+                  y_dequant_val_nuq[QUANT_PROFILES][QINDEX_RANGE][COEF_BANDS]);
+  DECLARE_ALIGNED(16, dequant_val_type_nuq,
+                  uv_dequant_val_nuq[QUANT_PROFILES][QINDEX_RANGE][COEF_BANDS]);
+#endif  // CONFIG_NEW_QUANT
+  AV1_COMMON common;
+  AV1EncoderConfig oxcf;
+  struct lookahead_ctx *lookahead;
+  struct lookahead_entry *alt_ref_source;
+
+  YV12_BUFFER_CONFIG *source;
+  YV12_BUFFER_CONFIG *last_source;  // NULL for first frame and alt_ref frames
+  YV12_BUFFER_CONFIG *un_scaled_source;
+  YV12_BUFFER_CONFIG scaled_source;
+  YV12_BUFFER_CONFIG *unscaled_last_source;
+  YV12_BUFFER_CONFIG scaled_last_source;
+
+  // Up-sampled reference buffers
+  // NOTE(zoeliu): It is needed to allocate sufficient space to the up-sampled
+  // reference buffers, which should include the up-sampled version of all the
+  // possibly stored references plus the currently coded frame itself.
+  EncRefCntBuffer upsampled_ref_bufs[REF_FRAMES + 1];
+  int upsampled_ref_idx[REF_FRAMES + 1];
+
+  // For a still frame, this flag is set to 1 to skip partition search.
+  int partition_search_skippable_frame;
+
+  int scaled_ref_idx[TOTAL_REFS_PER_FRAME];
+#if CONFIG_EXT_REFS
+  int lst_fb_idxes[LAST_REF_FRAMES];
+#else
+  int lst_fb_idx;
+#endif  // CONFIG_EXT_REFS
+  int gld_fb_idx;
+#if CONFIG_EXT_REFS
+  int bwd_fb_idx;  // BWD_REF_FRAME
+#endif             // CONFIG_EXT_REFS
+  int alt_fb_idx;
+
+  int last_show_frame_buf_idx;  // last show frame buffer index
+
+  int refresh_last_frame;
+  int refresh_golden_frame;
+#if CONFIG_EXT_REFS
+  int refresh_bwd_ref_frame;
+#endif  // CONFIG_EXT_REFS
+  int refresh_alt_ref_frame;
+
+  int ext_refresh_frame_flags_pending;
+  int ext_refresh_last_frame;
+  int ext_refresh_golden_frame;
+  int ext_refresh_alt_ref_frame;
+
+  int ext_refresh_frame_context_pending;
+  int ext_refresh_frame_context;
+
+  YV12_BUFFER_CONFIG last_frame_uf;
+#if CONFIG_LOOP_RESTORATION
+  YV12_BUFFER_CONFIG last_frame_db;
+  YV12_BUFFER_CONFIG trial_frame_rst;
+  uint8_t *extra_rstbuf;  // Extra buffers used in restoration search
+  RestorationInfo rst_search[MAX_MB_PLANE];  // Used for encoder side search
+#endif                                       // CONFIG_LOOP_RESTORATION
+
+  // Ambient reconstruction err target for force key frames
+  int64_t ambient_err;
+
+  RD_OPT rd;
+
+  CODING_CONTEXT coding_context;
+
+#if CONFIG_REF_MV
+  int nmv_costs[NMV_CONTEXTS][2][MV_VALS];
+  int nmv_costs_hp[NMV_CONTEXTS][2][MV_VALS];
+#endif
+
+  int nmvcosts[2][MV_VALS];
+  int nmvcosts_hp[2][MV_VALS];
+  int nmvsadcosts[2][MV_VALS];
+  int nmvsadcosts_hp[2][MV_VALS];
+
+  int64_t last_time_stamp_seen;
+  int64_t last_end_time_stamp_seen;
+  int64_t first_time_stamp_ever;
+
+  RATE_CONTROL rc;
+#if CONFIG_XIPHRC
+  od_rc_state od_rc;
+#endif
+  double framerate;
+
+  // NOTE(zoeliu): Any inter frame allows maximum of REF_FRAMES inter
+  // references; Plus the currently coded frame itself, it is needed to allocate
+  // sufficient space to the size of the maximum possible number of frames.
+  int interp_filter_selected[REF_FRAMES + 1][SWITCHABLE];
+
+  struct aom_codec_pkt_list *output_pkt_list;
+
+  MBGRAPH_FRAME_STATS mbgraph_stats[MAX_LAG_BUFFERS];
+  int mbgraph_n_frames;  // number of frames filled in the above
+  int static_mb_pct;     // % forced skip mbs by segmentation
+  int ref_frame_flags;
+
+  SPEED_FEATURES sf;
+
+  unsigned int max_mv_magnitude;
+  int mv_step_param;
+
+  int allow_comp_inter_inter;
+
+  uint8_t *segmentation_map;
+
+  CYCLIC_REFRESH *cyclic_refresh;
+  ActiveMap active_map;
+
+  fractional_mv_step_fp *find_fractional_mv_step;
+  av1_full_search_fn_t full_search_sad;  // It is currently unused.
+  av1_diamond_search_fn_t diamond_search_sad;
+  aom_variance_fn_ptr_t fn_ptr[BLOCK_SIZES];
+  uint64_t time_receive_data;
+  uint64_t time_compress_data;
+  uint64_t time_pick_lpf;
+  uint64_t time_encode_sb_row;
+
+#if CONFIG_FP_MB_STATS
+  int use_fp_mb_stats;
+#endif
+
+  TWO_PASS twopass;
+
+  YV12_BUFFER_CONFIG alt_ref_buffer;
+
+#if CONFIG_INTERNAL_STATS
+  unsigned int mode_chosen_counts[MAX_MODES];
+
+  int count;
+  uint64_t total_sq_error;
+  uint64_t total_samples;
+  ImageStat psnr;
+
+  double total_blockiness;
+  double worst_blockiness;
+
+  int bytes;
+  double summed_quality;
+  double summed_weights;
+  unsigned int tot_recode_hits;
+  double worst_ssim;
+
+  ImageStat fastssim;
+  ImageStat psnrhvs;
+
+  int b_calculate_blockiness;
+  int b_calculate_consistency;
+
+  double total_inconsistency;
+  double worst_consistency;
+  Ssimv *ssim_vars;
+  Metrics metrics;
+#endif
+  int b_calculate_psnr;
+
+  int droppable;
+
+  int initial_width;
+  int initial_height;
+  int initial_mbs;  // Number of MBs in the full-size frame; to be used to
+                    // normalize the firstpass stats. This will differ from the
+                    // number of MBs in the current frame when the frame is
+                    // scaled.
+
+  // Store frame variance info in SOURCE_VAR_BASED_PARTITION search type.
+  DIFF *source_diff_var;
+  // The threshold used in SOURCE_VAR_BASED_PARTITION search type.
+  unsigned int source_var_thresh;
+  int frames_till_next_var_check;
+
+  int frame_flags;
+
+  search_site_config ss_cfg;
+
+  int mbmode_cost[BLOCK_SIZE_GROUPS][INTRA_MODES];
+#if CONFIG_REF_MV
+  int newmv_mode_cost[NEWMV_MODE_CONTEXTS][2];
+  int zeromv_mode_cost[ZEROMV_MODE_CONTEXTS][2];
+  int refmv_mode_cost[REFMV_MODE_CONTEXTS][2];
+  int drl_mode_cost0[DRL_MODE_CONTEXTS][2];
+#endif
+
+  unsigned int inter_mode_cost[INTER_MODE_CONTEXTS][INTER_MODES];
+#if CONFIG_EXT_INTER
+  unsigned int inter_compound_mode_cost[INTER_MODE_CONTEXTS]
+                                       [INTER_COMPOUND_MODES];
+  unsigned int interintra_mode_cost[BLOCK_SIZE_GROUPS][INTERINTRA_MODES];
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  int motion_mode_cost[BLOCK_SIZES][MOTION_MODES];
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+  int motion_mode_cost1[BLOCK_SIZES][2];
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  int intra_uv_mode_cost[INTRA_MODES][INTRA_MODES];
+  int y_mode_costs[INTRA_MODES][INTRA_MODES][INTRA_MODES];
+  int switchable_interp_costs[SWITCHABLE_FILTER_CONTEXTS][SWITCHABLE_FILTERS];
+#if CONFIG_EXT_PARTITION_TYPES
+  int partition_cost[PARTITION_CONTEXTS + CONFIG_UNPOISON_PARTITION_CTX]
+                    [EXT_PARTITION_TYPES];
+#else
+  int partition_cost[PARTITION_CONTEXTS + CONFIG_UNPOISON_PARTITION_CTX]
+                    [PARTITION_TYPES];
+#endif
+#if CONFIG_PALETTE
+  int palette_y_size_cost[PALETTE_BLOCK_SIZES][PALETTE_SIZES];
+  int palette_uv_size_cost[PALETTE_BLOCK_SIZES][PALETTE_SIZES];
+  int palette_y_color_cost[PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS]
+                          [PALETTE_COLORS];
+  int palette_uv_color_cost[PALETTE_SIZES][PALETTE_COLOR_INDEX_CONTEXTS]
+                           [PALETTE_COLORS];
+#endif  // CONFIG_PALETTE
+  int tx_size_cost[TX_SIZES - 1][TX_SIZE_CONTEXTS][TX_SIZES];
+#if CONFIG_EXT_TX
+  int inter_tx_type_costs[EXT_TX_SETS_INTER][EXT_TX_SIZES][TX_TYPES];
+  int intra_tx_type_costs[EXT_TX_SETS_INTRA][EXT_TX_SIZES][INTRA_MODES]
+                         [TX_TYPES];
+#else
+  int intra_tx_type_costs[EXT_TX_SIZES][TX_TYPES][TX_TYPES];
+  int inter_tx_type_costs[EXT_TX_SIZES][TX_TYPES];
+#endif  // CONFIG_EXT_TX
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+  int intra_filter_cost[INTRA_FILTERS + 1][INTRA_FILTERS];
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_LOOP_RESTORATION
+  int switchable_restore_cost[RESTORE_SWITCHABLE_TYPES];
+#endif  // CONFIG_LOOP_RESTORATION
+#if CONFIG_GLOBAL_MOTION
+  int gmtype_cost[TRANS_TYPES];
+  int gmparams_cost[TOTAL_REFS_PER_FRAME];
+#endif  // CONFIG_GLOBAL_MOTION
+
+  int multi_arf_allowed;
+  int multi_arf_enabled;
+  int multi_arf_last_grp_enabled;
+
+  TileDataEnc *tile_data;
+  int allocated_tiles;  // Keep track of memory allocated for tiles.
+
+  TOKENEXTRA *tile_tok[MAX_TILE_ROWS][MAX_TILE_COLS];
+  unsigned int tok_count[MAX_TILE_ROWS][MAX_TILE_COLS];
+
+  TileBufferEnc tile_buffers[MAX_TILE_ROWS][MAX_TILE_COLS];
+
+  int resize_pending;
+  int resize_state;
+  int resize_scale_num;
+  int resize_scale_den;
+  int resize_avg_qp;
+  int resize_buffer_underflow;
+  int resize_count;
+
+  // VAR_BASED_PARTITION thresholds
+  // 0 - threshold_128x128;
+  // 1 - threshold_64x64;
+  // 2 - threshold_32x32;
+  // 3 - threshold_16x16;
+  // 4 - threshold_8x8;
+  int64_t vbp_thresholds[5];
+  int64_t vbp_threshold_minmax;
+  int64_t vbp_threshold_sad;
+  BLOCK_SIZE vbp_bsize_min;
+
+  // VARIANCE_AQ segment map refresh
+  int vaq_refresh;
+
+  // Multi-threading
+  int num_workers;
+  AVxWorker *workers;
+  struct EncWorkerData *tile_thr_data;
+  AV1LfSync lf_row_sync;
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  SUBFRAME_STATS subframe_stats;
+  // TODO(yaowu): minimize the size of count buffers
+  SUBFRAME_STATS wholeframe_stats;
+  av1_coeff_stats branch_ct_buf[COEF_PROBS_BUFS][TX_SIZES][PLANE_TYPES];
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+#if CONFIG_ANS
+  struct BufAnsCoder buf_ans;
+#endif
+#if CONFIG_EXT_REFS
+  int refresh_frame_mask;
+  int existing_fb_idx_to_show;
+  int is_arf_filter_off[MAX_EXT_ARFS + 1];
+  int num_extra_arfs;
+  int arf_map[MAX_EXT_ARFS + 1];
+#endif  // CONFIG_EXT_REFS
+#if CONFIG_GLOBAL_MOTION
+  int global_motion_search_done;
+#endif
+#if CONFIG_REFERENCE_BUFFER
+  SequenceHeader seq_params;
+#endif
+#if CONFIG_LV_MAP
+  tran_low_t *tcoeff_buf[MAX_MB_PLANE];
+#endif
+} AV1_COMP;
+
+void av1_initialize_enc(void);
+
+struct AV1_COMP *av1_create_compressor(AV1EncoderConfig *oxcf,
+                                       BufferPool *const pool);
+void av1_remove_compressor(AV1_COMP *cpi);
+
+void av1_change_config(AV1_COMP *cpi, const AV1EncoderConfig *oxcf);
+
+// receive a frames worth of data. caller can assume that a copy of this
+// frame is made and not just a copy of the pointer..
+int av1_receive_raw_frame(AV1_COMP *cpi, aom_enc_frame_flags_t frame_flags,
+                          YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
+                          int64_t end_time_stamp);
+
+int av1_get_compressed_data(AV1_COMP *cpi, unsigned int *frame_flags,
+                            size_t *size, uint8_t *dest, int64_t *time_stamp,
+                            int64_t *time_end, int flush);
+
+int av1_get_preview_raw_frame(AV1_COMP *cpi, YV12_BUFFER_CONFIG *dest);
+
+int av1_get_last_show_frame(AV1_COMP *cpi, YV12_BUFFER_CONFIG *frame);
+
+int av1_use_as_reference(AV1_COMP *cpi, int ref_frame_flags);
+
+void av1_update_reference(AV1_COMP *cpi, int ref_frame_flags);
+
+int av1_copy_reference_enc(AV1_COMP *cpi, AOM_REFFRAME ref_frame_flag,
+                           YV12_BUFFER_CONFIG *sd);
+
+int av1_set_reference_enc(AV1_COMP *cpi, AOM_REFFRAME ref_frame_flag,
+                          YV12_BUFFER_CONFIG *sd);
+
+int av1_update_entropy(AV1_COMP *cpi, int update);
+
+int av1_set_active_map(AV1_COMP *cpi, unsigned char *map, int rows, int cols);
+
+int av1_get_active_map(AV1_COMP *cpi, unsigned char *map, int rows, int cols);
+
+int av1_set_internal_size(AV1_COMP *cpi, AOM_SCALING horiz_mode,
+                          AOM_SCALING vert_mode);
+
+int av1_set_size_literal(AV1_COMP *cpi, unsigned int width,
+                         unsigned int height);
+
+int av1_get_quantizer(struct AV1_COMP *cpi);
+
+void av1_full_to_model_counts(av1_coeff_count_model *model_count,
+                              av1_coeff_count *full_count);
+
+static INLINE int frame_is_kf_gf_arf(const AV1_COMP *cpi) {
+  return frame_is_intra_only(&cpi->common) || cpi->refresh_alt_ref_frame ||
+         (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref);
+}
+
+static INLINE int get_ref_frame_map_idx(const AV1_COMP *cpi,
+                                        MV_REFERENCE_FRAME ref_frame) {
+#if CONFIG_EXT_REFS
+  if (ref_frame >= LAST_FRAME && ref_frame <= LAST3_FRAME)
+    return cpi->lst_fb_idxes[ref_frame - 1];
+#else
+  if (ref_frame == LAST_FRAME) return cpi->lst_fb_idx;
+#endif  // CONFIG_EXT_REFS
+  else if (ref_frame == GOLDEN_FRAME)
+    return cpi->gld_fb_idx;
+#if CONFIG_EXT_REFS
+  else if (ref_frame == BWDREF_FRAME)
+    return cpi->bwd_fb_idx;
+#endif  // CONFIG_EXT_REFS
+  else
+    return cpi->alt_fb_idx;
+}
+
+static INLINE int get_ref_frame_buf_idx(const AV1_COMP *cpi,
+                                        MV_REFERENCE_FRAME ref_frame) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int map_idx = get_ref_frame_map_idx(cpi, ref_frame);
+  return (map_idx != INVALID_IDX) ? cm->ref_frame_map[map_idx] : INVALID_IDX;
+}
+
+static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer(
+    const AV1_COMP *cpi, MV_REFERENCE_FRAME ref_frame) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+  return buf_idx != INVALID_IDX ? &cm->buffer_pool->frame_bufs[buf_idx].buf
+                                : NULL;
+}
+
+static INLINE const YV12_BUFFER_CONFIG *get_upsampled_ref(
+    const AV1_COMP *cpi, const MV_REFERENCE_FRAME ref_frame) {
+  // Use up-sampled reference frames.
+  const int buf_idx =
+      cpi->upsampled_ref_idx[get_ref_frame_map_idx(cpi, ref_frame)];
+  return &cpi->upsampled_ref_bufs[buf_idx].buf;
+}
+
+#if CONFIG_EXT_REFS
+static INLINE int enc_is_ref_frame_buf(AV1_COMP *cpi, RefCntBuffer *frame_buf) {
+  MV_REFERENCE_FRAME ref_frame;
+  AV1_COMMON *const cm = &cpi->common;
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+    if (buf_idx == INVALID_IDX) continue;
+    if (frame_buf == &cm->buffer_pool->frame_bufs[buf_idx]) break;
+  }
+  return (ref_frame <= ALTREF_FRAME);
+}
+#endif  // CONFIG_EXT_REFS
+
+static INLINE unsigned int get_token_alloc(int mb_rows, int mb_cols) {
+  // We assume 3 planes all at full resolution. We assume up to 1 token per
+  // pixel, and then allow a head room of 1 EOSB token per 4x4 block per plane,
+  // plus EOSB_TOKEN per plane.
+  return mb_rows * mb_cols * (16 * 16 + 17) * 3;
+}
+
+// Get the allocated token size for a tile. It does the same calculation as in
+// the frame token allocation.
+static INLINE unsigned int allocated_tokens(TileInfo tile) {
+#if CONFIG_CB4X4
+  int tile_mb_rows = (tile.mi_row_end - tile.mi_row_start + 2) >> 2;
+  int tile_mb_cols = (tile.mi_col_end - tile.mi_col_start + 2) >> 2;
+#else
+  int tile_mb_rows = (tile.mi_row_end - tile.mi_row_start + 1) >> 1;
+  int tile_mb_cols = (tile.mi_col_end - tile.mi_col_start + 1) >> 1;
+#endif
+
+  return get_token_alloc(tile_mb_rows, tile_mb_cols);
+}
+
+void av1_alloc_compressor_data(AV1_COMP *cpi);
+
+void av1_scale_references(AV1_COMP *cpi);
+
+void av1_update_reference_frames(AV1_COMP *cpi);
+
+void av1_set_high_precision_mv(AV1_COMP *cpi, int allow_high_precision_mv);
+#if CONFIG_TEMPMV_SIGNALING
+void av1_set_temporal_mv_prediction(AV1_COMP *cpi, int allow_tempmv_prediction);
+#endif
+
+YV12_BUFFER_CONFIG *av1_scale_if_required_fast(AV1_COMMON *cm,
+                                               YV12_BUFFER_CONFIG *unscaled,
+                                               YV12_BUFFER_CONFIG *scaled);
+
+YV12_BUFFER_CONFIG *av1_scale_if_required(AV1_COMMON *cm,
+                                          YV12_BUFFER_CONFIG *unscaled,
+                                          YV12_BUFFER_CONFIG *scaled);
+
+void av1_apply_encoding_flags(AV1_COMP *cpi, aom_enc_frame_flags_t flags);
+
+static INLINE int is_altref_enabled(const AV1_COMP *const cpi) {
+  return cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.enable_auto_arf;
+}
+
+// TODO(zoeliu): To set up cpi->oxcf.enable_auto_brf
+#if 0 && CONFIG_EXT_REFS
+static INLINE int is_bwdref_enabled(const AV1_COMP *const cpi) {
+  // NOTE(zoeliu): The enabling of bi-predictive frames depends on the use of
+  //               alt_ref, and now will be off when the alt_ref interval is
+  //               not sufficiently large.
+  return is_altref_enabled(cpi) && cpi->oxcf.enable_auto_brf;
+}
+#endif  // CONFIG_EXT_REFS
+
+static INLINE void set_ref_ptrs(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                                MV_REFERENCE_FRAME ref0,
+                                MV_REFERENCE_FRAME ref1) {
+  xd->block_refs[0] =
+      &cm->frame_refs[ref0 >= LAST_FRAME ? ref0 - LAST_FRAME : 0];
+  xd->block_refs[1] =
+      &cm->frame_refs[ref1 >= LAST_FRAME ? ref1 - LAST_FRAME : 0];
+}
+
+static INLINE int get_chessboard_index(int frame_index) {
+  return frame_index & 0x1;
+}
+
+static INLINE int *cond_cost_list(const struct AV1_COMP *cpi, int *cost_list) {
+  return cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL;
+}
+
+void av1_new_framerate(AV1_COMP *cpi, double framerate);
+
+#define LAYER_IDS_TO_IDX(sl, tl, num_tl) ((sl) * (num_tl) + (tl))
+
+// Update up-sampled reference frame index.
+static INLINE void uref_cnt_fb(EncRefCntBuffer *ubufs, int *uidx,
+                               int new_uidx) {
+  const int ref_index = *uidx;
+
+  if (ref_index >= 0 && ubufs[ref_index].ref_count > 0)
+    ubufs[ref_index].ref_count--;
+
+  *uidx = new_uidx;
+  ubufs[new_uidx].ref_count++;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_ENCODER_H_
diff --git a/third_party/aom/av1/encoder/encodetxb.c b/third_party/aom/av1/encoder/encodetxb.c
new file mode 100644
index 0000000000..3f71a4472e
--- /dev/null
+++ b/third_party/aom/av1/encoder/encodetxb.c
@@ -0,0 +1,784 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/scan.h"
+#include "av1/common/blockd.h"
+#include "av1/common/idct.h"
+#include "av1/common/pred_common.h"
+#include "av1/encoder/bitstream.h"
+#include "av1/encoder/encodeframe.h"
+#include "av1/encoder/cost.h"
+#include "av1/encoder/encodetxb.h"
+#include "av1/encoder/rdopt.h"
+#include "av1/encoder/subexp.h"
+#include "av1/encoder/tokenize.h"
+
+void av1_alloc_txb_buf(AV1_COMP *cpi) {
+#if 0
+  AV1_COMMON *cm = &cpi->common;
+  int mi_block_size = 1 << MI_SIZE_LOG2;
+  // TODO(angiebird): Make sure cm->subsampling_x/y is set correctly, and then
+  // use precise buffer size according to cm->subsampling_x/y
+  int pixel_stride = mi_block_size * cm->mi_cols;
+  int pixel_height = mi_block_size * cm->mi_rows;
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    CHECK_MEM_ERROR(
+        cm, cpi->tcoeff_buf[i],
+        aom_malloc(sizeof(*cpi->tcoeff_buf[i]) * pixel_stride * pixel_height));
+  }
+#else
+  (void)cpi;
+#endif
+}
+
+void av1_free_txb_buf(AV1_COMP *cpi) {
+#if 0
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    aom_free(cpi->tcoeff_buf[i]);
+  }
+#else
+  (void)cpi;
+#endif
+}
+
+static void write_golomb(aom_writer *w, int level) {
+  int x = level + 1;
+  int i = x;
+  int length = 0;
+
+  while (i) {
+    i >>= 1;
+    ++length;
+  }
+  assert(length > 0);
+
+  for (i = 0; i < length - 1; ++i) aom_write_bit(w, 0);
+
+  for (i = length - 1; i >= 0; --i) aom_write_bit(w, (x >> i) & 0x01);
+}
+
+void av1_write_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+                          aom_writer *w, int block, int plane,
+                          const tran_low_t *tcoeff, uint16_t eob,
+                          TXB_CTX *txb_ctx) {
+  aom_prob *nz_map;
+  aom_prob *eob_flag;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const PLANE_TYPE plane_type = get_plane_type(plane);
+  const TX_SIZE tx_size = get_tx_size(plane, xd);
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  const int16_t *scan = scan_order->scan;
+  int c;
+  int is_nz;
+  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
+  const int seg_eob = tx_size_2d[tx_size];
+  uint8_t txb_mask[32 * 32] = { 0 };
+  uint16_t update_eob = 0;
+
+  aom_write(w, eob == 0, cm->fc->txb_skip[tx_size][txb_ctx->txb_skip_ctx]);
+
+  if (eob == 0) return;
+#if CONFIG_TXK_SEL
+  av1_write_tx_type(cm, xd, block, plane, w);
+#endif
+
+  nz_map = cm->fc->nz_map[tx_size][plane_type];
+  eob_flag = cm->fc->eob_flag[tx_size][plane_type];
+
+  for (c = 0; c < eob; ++c) {
+    int coeff_ctx = get_nz_map_ctx(tcoeff, txb_mask, scan[c], bwl);
+    int eob_ctx = get_eob_ctx(tcoeff, scan[c], bwl);
+
+    tran_low_t v = tcoeff[scan[c]];
+    is_nz = (v != 0);
+
+    if (c == seg_eob - 1) break;
+
+    aom_write(w, is_nz, nz_map[coeff_ctx]);
+
+    if (is_nz) {
+      aom_write(w, c == (eob - 1), eob_flag[eob_ctx]);
+    }
+    txb_mask[scan[c]] = 1;
+  }
+
+  int i;
+  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+    aom_prob *coeff_base = cm->fc->coeff_base[tx_size][plane_type][i];
+
+    update_eob = 0;
+    for (c = eob - 1; c >= 0; --c) {
+      tran_low_t v = tcoeff[scan[c]];
+      tran_low_t level = abs(v);
+      int sign = (v < 0) ? 1 : 0;
+      int ctx;
+
+      if (level <= i) continue;
+
+      ctx = get_base_ctx(tcoeff, scan[c], bwl, i + 1);
+
+      if (level == i + 1) {
+        aom_write(w, 1, coeff_base[ctx]);
+        if (c == 0) {
+          aom_write(w, sign, cm->fc->dc_sign[plane_type][txb_ctx->dc_sign_ctx]);
+        } else {
+          aom_write_bit(w, sign);
+        }
+        continue;
+      }
+      aom_write(w, 0, coeff_base[ctx]);
+      update_eob = AOMMAX(update_eob, c);
+    }
+  }
+
+  for (c = update_eob; c >= 0; --c) {
+    tran_low_t v = tcoeff[scan[c]];
+    tran_low_t level = abs(v);
+    int sign = (v < 0) ? 1 : 0;
+    int idx;
+    int ctx;
+
+    if (level <= NUM_BASE_LEVELS) continue;
+
+    if (c == 0) {
+      aom_write(w, sign, cm->fc->dc_sign[plane_type][txb_ctx->dc_sign_ctx]);
+    } else {
+      aom_write_bit(w, sign);
+    }
+
+    // level is above 1.
+    ctx = get_level_ctx(tcoeff, scan[c], bwl);
+    for (idx = 0; idx < COEFF_BASE_RANGE; ++idx) {
+      if (level == (idx + 1 + NUM_BASE_LEVELS)) {
+        aom_write(w, 1, cm->fc->coeff_lps[tx_size][plane_type][ctx]);
+        break;
+      }
+      aom_write(w, 0, cm->fc->coeff_lps[tx_size][plane_type][ctx]);
+    }
+    if (idx < COEFF_BASE_RANGE) continue;
+
+    // use 0-th order Golomb code to handle the residual level.
+    write_golomb(w, level - COEFF_BASE_RANGE - 1 - NUM_BASE_LEVELS);
+  }
+}
+
+void av1_write_coeffs_mb(const AV1_COMMON *const cm, MACROBLOCK *x,
+                         aom_writer *w, int plane) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  BLOCK_SIZE bsize = mbmi->sb_type;
+  struct macroblockd_plane *pd = &xd->plane[plane];
+
+#if CONFIG_CB4X4
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+  const BLOCK_SIZE plane_bsize =
+      get_plane_block_size(AOMMAX(bsize, BLOCK_8X8), pd);
+#endif
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+  TX_SIZE tx_size = get_tx_size(plane, xd);
+  const int bkw = tx_size_wide_unit[tx_size];
+  const int bkh = tx_size_high_unit[tx_size];
+  const int step = tx_size_wide_unit[tx_size] * tx_size_high_unit[tx_size];
+  int row, col;
+  int block = 0;
+  for (row = 0; row < max_blocks_high; row += bkh) {
+    for (col = 0; col < max_blocks_wide; col += bkw) {
+      tran_low_t *tcoeff = BLOCK_OFFSET(x->mbmi_ext->tcoeff[plane], block);
+      uint16_t eob = x->mbmi_ext->eobs[plane][block];
+      TXB_CTX txb_ctx = { x->mbmi_ext->txb_skip_ctx[plane][block],
+                          x->mbmi_ext->dc_sign_ctx[plane][block] };
+      av1_write_coeffs_txb(cm, xd, w, block, plane, tcoeff, eob, &txb_ctx);
+      block += step;
+    }
+  }
+}
+
+static INLINE void get_base_ctx_set(const tran_low_t *tcoeffs,
+                                    int c,  // raster order
+                                    const int bwl,
+                                    int ctx_set[NUM_BASE_LEVELS]) {
+  const int row = c >> bwl;
+  const int col = c - (row << bwl);
+  const int stride = 1 << bwl;
+  int mag[NUM_BASE_LEVELS] = { 0 };
+  int idx;
+  tran_low_t abs_coeff;
+  int i;
+
+  for (idx = 0; idx < BASE_CONTEXT_POSITION_NUM; ++idx) {
+    int ref_row = row + base_ref_offset[idx][0];
+    int ref_col = col + base_ref_offset[idx][1];
+    int pos = (ref_row << bwl) + ref_col;
+
+    if (ref_row < 0 || ref_col < 0 || ref_row >= stride || ref_col >= stride)
+      continue;
+
+    abs_coeff = abs(tcoeffs[pos]);
+
+    for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+      ctx_set[i] += abs_coeff > i;
+      if (base_ref_offset[idx][0] >= 0 && base_ref_offset[idx][1] >= 0)
+        mag[i] |= abs_coeff > (i + 1);
+    }
+  }
+
+  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+    ctx_set[i] = (ctx_set[i] + 1) >> 1;
+
+    if (row == 0 && col == 0)
+      ctx_set[i] = (ctx_set[i] << 1) + mag[i];
+    else if (row == 0)
+      ctx_set[i] = 8 + (ctx_set[i] << 1) + mag[i];
+    else if (col == 0)
+      ctx_set[i] = 18 + (ctx_set[i] << 1) + mag[i];
+    else
+      ctx_set[i] = 28 + (ctx_set[i] << 1) + mag[i];
+  }
+  return;
+}
+
+int av1_cost_coeffs_txb(const AV1_COMP *const cpi, MACROBLOCK *x, int plane,
+                        int block, TXB_CTX *txb_ctx) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const TX_SIZE tx_size = get_tx_size(plane, xd);
+  const PLANE_TYPE plane_type = get_plane_type(plane);
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const struct macroblock_plane *p = &x->plane[plane];
+  const int eob = p->eobs[block];
+  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  int c, cost;
+  const int seg_eob = AOMMIN(eob, tx_size_2d[tx_size] - 1);
+  int txb_skip_ctx = txb_ctx->txb_skip_ctx;
+  aom_prob *nz_map = xd->fc->nz_map[tx_size][plane_type];
+
+  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
+  // txb_mask is only initialized for once here. After that, it will be set when
+  // coding zero map and then reset when coding level 1 info.
+  uint8_t txb_mask[32 * 32] = { 0 };
+  aom_prob(*coeff_base)[COEFF_BASE_CONTEXTS] =
+      xd->fc->coeff_base[tx_size][plane_type];
+
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  const int16_t *scan = scan_order->scan;
+
+  cost = 0;
+
+  if (eob == 0) {
+    cost = av1_cost_bit(xd->fc->txb_skip[tx_size][txb_skip_ctx], 1);
+    return cost;
+  }
+
+  cost = av1_cost_bit(xd->fc->txb_skip[tx_size][txb_skip_ctx], 0);
+
+#if CONFIG_TXK_SEL
+  cost += av1_tx_type_cost(cpi, xd, mbmi->sb_type, plane, tx_size, tx_type);
+#endif
+
+  for (c = 0; c < eob; ++c) {
+    tran_low_t v = qcoeff[scan[c]];
+    int is_nz = (v != 0);
+    int level = abs(v);
+
+    if (c < seg_eob) {
+      int coeff_ctx = get_nz_map_ctx(qcoeff, txb_mask, scan[c], bwl);
+      cost += av1_cost_bit(nz_map[coeff_ctx], is_nz);
+    }
+
+    if (is_nz) {
+      int ctx_ls[NUM_BASE_LEVELS] = { 0 };
+      int sign = (v < 0) ? 1 : 0;
+
+      // sign bit cost
+      if (c == 0) {
+        int dc_sign_ctx = txb_ctx->dc_sign_ctx;
+
+        cost += av1_cost_bit(xd->fc->dc_sign[plane_type][dc_sign_ctx], sign);
+      } else {
+        cost += av1_cost_bit(128, sign);
+      }
+
+      get_base_ctx_set(qcoeff, scan[c], bwl, ctx_ls);
+
+      int i;
+      for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+        if (level <= i) continue;
+
+        if (level == i + 1) {
+          cost += av1_cost_bit(coeff_base[i][ctx_ls[i]], 1);
+          continue;
+        }
+        cost += av1_cost_bit(coeff_base[i][ctx_ls[i]], 0);
+      }
+
+      if (level > NUM_BASE_LEVELS) {
+        int idx;
+        int ctx;
+
+        ctx = get_level_ctx(qcoeff, scan[c], bwl);
+
+        for (idx = 0; idx < COEFF_BASE_RANGE; ++idx) {
+          if (level == (idx + 1 + NUM_BASE_LEVELS)) {
+            cost +=
+                av1_cost_bit(xd->fc->coeff_lps[tx_size][plane_type][ctx], 1);
+            break;
+          }
+          cost += av1_cost_bit(xd->fc->coeff_lps[tx_size][plane_type][ctx], 0);
+        }
+
+        if (idx >= COEFF_BASE_RANGE) {
+          // residual cost
+          int r = level - COEFF_BASE_RANGE - NUM_BASE_LEVELS;
+          int ri = r;
+          int length = 0;
+
+          while (ri) {
+            ri >>= 1;
+            ++length;
+          }
+
+          for (ri = 0; ri < length - 1; ++ri) cost += av1_cost_bit(128, 0);
+
+          for (ri = length - 1; ri >= 0; --ri)
+            cost += av1_cost_bit(128, (r >> ri) & 0x01);
+        }
+      }
+
+      if (c < seg_eob) {
+        int eob_ctx = get_eob_ctx(qcoeff, scan[c], bwl);
+        cost += av1_cost_bit(xd->fc->eob_flag[tx_size][plane_type][eob_ctx],
+                             c == (eob - 1));
+      }
+    }
+
+    txb_mask[scan[c]] = 1;
+  }
+
+  return cost;
+}
+
+typedef struct TxbParams {
+  const AV1_COMP *cpi;
+  ThreadData *td;
+  int rate;
+} TxbParams;
+
+int av1_get_txb_entropy_context(const tran_low_t *qcoeff,
+                                const SCAN_ORDER *scan_order, int eob) {
+  const int16_t *scan = scan_order->scan;
+  int cul_level = 0;
+  int c;
+  for (c = 0; c < eob; ++c) {
+    cul_level += abs(qcoeff[scan[c]]);
+  }
+
+  cul_level = AOMMIN(COEFF_CONTEXT_MASK, cul_level);
+  set_dc_sign(&cul_level, qcoeff[0]);
+
+  return cul_level;
+}
+
+static void update_txb_context(int plane, int block, int blk_row, int blk_col,
+                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                               void *arg) {
+  TxbParams *const args = arg;
+  const AV1_COMP *cpi = args->cpi;
+  const AV1_COMMON *cm = &cpi->common;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  const uint16_t eob = p->eobs[block];
+  const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  const PLANE_TYPE plane_type = pd->plane_type;
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  (void)plane_bsize;
+
+  int cul_level = av1_get_txb_entropy_context(qcoeff, scan_order, eob);
+  av1_set_contexts(xd, pd, plane, tx_size, cul_level, blk_col, blk_row);
+}
+
+static void update_and_record_txb_context(int plane, int block, int blk_row,
+                                          int blk_col, BLOCK_SIZE plane_bsize,
+                                          TX_SIZE tx_size, void *arg) {
+  TxbParams *const args = arg;
+  const AV1_COMP *cpi = args->cpi;
+  const AV1_COMMON *cm = &cpi->common;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  int eob = p->eobs[block], update_eob = 0;
+  const PLANE_TYPE plane_type = pd->plane_type;
+  const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  tran_low_t *tcoeff = BLOCK_OFFSET(x->mbmi_ext->tcoeff[plane], block);
+  const int segment_id = mbmi->segment_id;
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  const int16_t *scan = scan_order->scan;
+  const int seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
+  int c, i;
+  TXB_CTX txb_ctx;
+  get_txb_ctx(plane_bsize, tx_size, plane, pd->above_context + blk_col,
+              pd->left_context + blk_row, &txb_ctx);
+  const int bwl = b_width_log2_lookup[txsize_to_bsize[tx_size]] + 2;
+  int cul_level = 0;
+  unsigned int(*nz_map_count)[SIG_COEF_CONTEXTS][2];
+  uint8_t txb_mask[32 * 32] = { 0 };
+
+  nz_map_count = &td->counts->nz_map[tx_size][plane_type];
+
+  memcpy(tcoeff, qcoeff, sizeof(*tcoeff) * seg_eob);
+
+  ++td->counts->txb_skip[tx_size][txb_ctx.txb_skip_ctx][eob == 0];
+  x->mbmi_ext->txb_skip_ctx[plane][block] = txb_ctx.txb_skip_ctx;
+
+  x->mbmi_ext->eobs[plane][block] = eob;
+
+  if (eob == 0) {
+    av1_set_contexts(xd, pd, plane, tx_size, 0, blk_col, blk_row);
+    return;
+  }
+
+#if CONFIG_TXK_SEL
+  av1_update_tx_type_count(cm, xd, block, plane, mbmi->sb_type, tx_size,
+                           td->counts);
+#endif
+
+  for (c = 0; c < eob; ++c) {
+    tran_low_t v = qcoeff[scan[c]];
+    int is_nz = (v != 0);
+    int coeff_ctx = get_nz_map_ctx(tcoeff, txb_mask, scan[c], bwl);
+    int eob_ctx = get_eob_ctx(tcoeff, scan[c], bwl);
+
+    if (c == seg_eob - 1) break;
+
+    ++(*nz_map_count)[coeff_ctx][is_nz];
+
+    if (is_nz) {
+      ++td->counts->eob_flag[tx_size][plane_type][eob_ctx][c == (eob - 1)];
+    }
+    txb_mask[scan[c]] = 1;
+  }
+
+  // Reverse process order to handle coefficient level and sign.
+  for (i = 0; i < NUM_BASE_LEVELS; ++i) {
+    update_eob = 0;
+    for (c = eob - 1; c >= 0; --c) {
+      tran_low_t v = qcoeff[scan[c]];
+      tran_low_t level = abs(v);
+      int ctx;
+
+      if (level <= i) continue;
+
+      ctx = get_base_ctx(tcoeff, scan[c], bwl, i + 1);
+
+      if (level == i + 1) {
+        ++td->counts->coeff_base[tx_size][plane_type][i][ctx][1];
+        if (c == 0) {
+          int dc_sign_ctx = txb_ctx.dc_sign_ctx;
+
+          ++td->counts->dc_sign[plane_type][dc_sign_ctx][v < 0];
+          x->mbmi_ext->dc_sign_ctx[plane][block] = dc_sign_ctx;
+        }
+        cul_level += level;
+        continue;
+      }
+      ++td->counts->coeff_base[tx_size][plane_type][i][ctx][0];
+      update_eob = AOMMAX(update_eob, c);
+    }
+  }
+
+  for (c = update_eob; c >= 0; --c) {
+    tran_low_t v = qcoeff[scan[c]];
+    tran_low_t level = abs(v);
+    int idx;
+    int ctx;
+
+    if (level <= NUM_BASE_LEVELS) continue;
+
+    cul_level += level;
+    if (c == 0) {
+      int dc_sign_ctx = txb_ctx.dc_sign_ctx;
+
+      ++td->counts->dc_sign[plane_type][dc_sign_ctx][v < 0];
+      x->mbmi_ext->dc_sign_ctx[plane][block] = dc_sign_ctx;
+    }
+
+    // level is above 1.
+    ctx = get_level_ctx(tcoeff, scan[c], bwl);
+    for (idx = 0; idx < COEFF_BASE_RANGE; ++idx) {
+      if (level == (idx + 1 + NUM_BASE_LEVELS)) {
+        ++td->counts->coeff_lps[tx_size][plane_type][ctx][1];
+        break;
+      }
+      ++td->counts->coeff_lps[tx_size][plane_type][ctx][0];
+    }
+    if (idx < COEFF_BASE_RANGE) continue;
+
+    // use 0-th order Golomb code to handle the residual level.
+  }
+
+  cul_level = AOMMIN(COEFF_CONTEXT_MASK, cul_level);
+
+  // DC value
+  set_dc_sign(&cul_level, tcoeff[0]);
+  av1_set_contexts(xd, pd, plane, tx_size, cul_level, blk_col, blk_row);
+
+#if CONFIG_ADAPT_SCAN
+  // Since dqcoeff is not available here, we pass qcoeff into
+  // av1_update_scan_count_facade(). The update behavior should be the same
+  // because av1_update_scan_count_facade() only cares if coefficients are zero
+  // or not.
+  av1_update_scan_count_facade((AV1_COMMON *)cm, td->counts, tx_size, tx_type,
+                               qcoeff, eob);
+#endif
+}
+
+void av1_update_txb_context(const AV1_COMP *cpi, ThreadData *td,
+                            RUN_TYPE dry_run, BLOCK_SIZE bsize, int *rate,
+                            int mi_row, int mi_col) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int ctx = av1_get_skip_context(xd);
+  const int skip_inc =
+      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP);
+  struct TxbParams arg = { cpi, td, 0 };
+  (void)rate;
+  (void)mi_row;
+  (void)mi_col;
+  if (mbmi->skip) {
+    if (!dry_run) td->counts->skip[ctx][1] += skip_inc;
+    reset_skip_context(xd, bsize);
+    return;
+  }
+
+  if (!dry_run) {
+    td->counts->skip[ctx][0] += skip_inc;
+    av1_foreach_transformed_block(xd, bsize, mi_row, mi_col,
+                                  update_and_record_txb_context, &arg);
+  } else if (dry_run == DRY_RUN_NORMAL) {
+    av1_foreach_transformed_block(xd, bsize, mi_row, mi_col, update_txb_context,
+                                  &arg);
+  } else {
+    printf("DRY_RUN_COSTCOEFFS is not supported yet\n");
+    assert(0);
+  }
+}
+
+static void find_new_prob(unsigned int *branch_cnt, aom_prob *oldp,
+                          int *savings, int *update, aom_writer *const bc) {
+  const aom_prob upd = DIFF_UPDATE_PROB;
+  int u = 0;
+  aom_prob newp = get_binary_prob(branch_cnt[0], branch_cnt[1]);
+  int s = av1_prob_diff_update_savings_search(branch_cnt, *oldp, &newp, upd, 1);
+
+  if (s > 0 && newp != *oldp) u = 1;
+
+  if (u)
+    *savings += s - (int)(av1_cost_zero(upd));  // TODO(jingning): 1?
+  else
+    *savings -= (int)(av1_cost_zero(upd));
+
+  if (update) {
+    ++update[u];
+    return;
+  }
+
+  aom_write(bc, u, upd);
+  if (u) {
+    /* send/use new probability */
+    av1_write_prob_diff_update(bc, newp, *oldp);
+    *oldp = newp;
+  }
+}
+
+static void write_txb_probs(aom_writer *const bc, AV1_COMP *cpi,
+                            TX_SIZE tx_size) {
+  FRAME_CONTEXT *fc = cpi->common.fc;
+  FRAME_COUNTS *counts = cpi->td.counts;
+  int savings = 0;
+  int update[2] = { 0, 0 };
+  int plane, ctx, level;
+
+  for (ctx = 0; ctx < TXB_SKIP_CONTEXTS; ++ctx) {
+    find_new_prob(counts->txb_skip[tx_size][ctx], &fc->txb_skip[tx_size][ctx],
+                  &savings, update, bc);
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < SIG_COEF_CONTEXTS; ++ctx) {
+      find_new_prob(counts->nz_map[tx_size][plane][ctx],
+                    &fc->nz_map[tx_size][plane][ctx], &savings, update, bc);
+    }
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < EOB_COEF_CONTEXTS; ++ctx) {
+      find_new_prob(counts->eob_flag[tx_size][plane][ctx],
+                    &fc->eob_flag[tx_size][plane][ctx], &savings, update, bc);
+    }
+  }
+
+  for (level = 0; level < NUM_BASE_LEVELS; ++level) {
+    for (plane = 0; plane < PLANE_TYPES; ++plane) {
+      for (ctx = 0; ctx < COEFF_BASE_CONTEXTS; ++ctx) {
+        find_new_prob(counts->coeff_base[tx_size][plane][level][ctx],
+                      &fc->coeff_base[tx_size][plane][level][ctx], &savings,
+                      update, bc);
+      }
+    }
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx) {
+      find_new_prob(counts->coeff_lps[tx_size][plane][ctx],
+                    &fc->coeff_lps[tx_size][plane][ctx], &savings, update, bc);
+    }
+  }
+
+  // Decide if to update the model for this tx_size
+  if (update[1] == 0 || savings < 0) {
+    aom_write_bit(bc, 0);
+    return;
+  }
+  aom_write_bit(bc, 1);
+
+  for (ctx = 0; ctx < TXB_SKIP_CONTEXTS; ++ctx) {
+    find_new_prob(counts->txb_skip[tx_size][ctx], &fc->txb_skip[tx_size][ctx],
+                  &savings, NULL, bc);
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < SIG_COEF_CONTEXTS; ++ctx) {
+      find_new_prob(counts->nz_map[tx_size][plane][ctx],
+                    &fc->nz_map[tx_size][plane][ctx], &savings, NULL, bc);
+    }
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < EOB_COEF_CONTEXTS; ++ctx) {
+      find_new_prob(counts->eob_flag[tx_size][plane][ctx],
+                    &fc->eob_flag[tx_size][plane][ctx], &savings, NULL, bc);
+    }
+  }
+
+  for (level = 0; level < NUM_BASE_LEVELS; ++level) {
+    for (plane = 0; plane < PLANE_TYPES; ++plane) {
+      for (ctx = 0; ctx < COEFF_BASE_CONTEXTS; ++ctx) {
+        find_new_prob(counts->coeff_base[tx_size][plane][level][ctx],
+                      &fc->coeff_base[tx_size][plane][level][ctx], &savings,
+                      NULL, bc);
+      }
+    }
+  }
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane) {
+    for (ctx = 0; ctx < LEVEL_CONTEXTS; ++ctx) {
+      find_new_prob(counts->coeff_lps[tx_size][plane][ctx],
+                    &fc->coeff_lps[tx_size][plane][ctx], &savings, NULL, bc);
+    }
+  }
+}
+
+void av1_write_txb_probs(AV1_COMP *cpi, aom_writer *w) {
+  const TX_MODE tx_mode = cpi->common.tx_mode;
+  const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
+  TX_SIZE tx_size;
+  int ctx, plane;
+
+  for (plane = 0; plane < PLANE_TYPES; ++plane)
+    for (ctx = 0; ctx < DC_SIGN_CONTEXTS; ++ctx)
+      av1_cond_prob_diff_update(w, &cpi->common.fc->dc_sign[plane][ctx],
+                                cpi->td.counts->dc_sign[plane][ctx], 1);
+
+  for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
+    write_txb_probs(w, cpi, tx_size);
+}
+
+#if CONFIG_TXK_SEL
+int64_t av1_search_txk_type(const AV1_COMP *cpi, MACROBLOCK *x, int plane,
+                            int block, int blk_row, int blk_col,
+                            BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                            const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
+                            int use_fast_coef_costing, RD_STATS *rd_stats) {
+  const AV1_COMMON *cm = &cpi->common;
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  TX_TYPE txk_start = DCT_DCT;
+  TX_TYPE txk_end = TX_TYPES - 1;
+  TX_TYPE best_tx_type = txk_start;
+  int64_t best_rd = INT64_MAX;
+  const int coeff_ctx = combine_entropy_contexts(*a, *l);
+  TX_TYPE tx_type;
+  for (tx_type = txk_start; tx_type <= txk_end; ++tx_type) {
+    if (plane == 0) mbmi->txk_type[block] = tx_type;
+    TX_TYPE ref_tx_type =
+        get_tx_type(get_plane_type(plane), xd, block, tx_size);
+    if (tx_type != ref_tx_type) {
+      // use get_tx_type() to check if the tx_type is valid for the current mode
+      // if it's not, we skip it here.
+      continue;
+    }
+    RD_STATS this_rd_stats;
+    av1_invalid_rd_stats(&this_rd_stats);
+    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                    coeff_ctx, AV1_XFORM_QUANT_FP);
+    if (x->plane[plane].eobs[block] && !xd->lossless[mbmi->segment_id])
+      av1_optimize_b(cm, x, plane, block, tx_size, coeff_ctx);
+    av1_dist_block(cpi, x, plane, plane_bsize, block, blk_row, blk_col, tx_size,
+                   &this_rd_stats.dist, &this_rd_stats.sse,
+                   OUTPUT_HAS_PREDICTED_PIXELS);
+    const SCAN_ORDER *scan_order =
+        get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+    this_rd_stats.rate = av1_cost_coeffs(
+        cpi, x, plane, block, tx_size, scan_order, a, l, use_fast_coef_costing);
+    int rd =
+        RDCOST(x->rdmult, x->rddiv, this_rd_stats.rate, this_rd_stats.dist);
+    if (rd < best_rd) {
+      best_rd = rd;
+      *rd_stats = this_rd_stats;
+      best_tx_type = tx_type;
+    }
+  }
+  if (plane == 0) mbmi->txk_type[block] = best_tx_type;
+  // TODO(angiebird): Instead of re-call av1_xform_quant and av1_optimize_b,
+  // copy the best result in the above tx_type search for loop
+  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                  coeff_ctx, AV1_XFORM_QUANT_FP);
+  if (x->plane[plane].eobs[block] && !xd->lossless[mbmi->segment_id])
+    av1_optimize_b(cm, x, plane, block, tx_size, coeff_ctx);
+  if (!is_inter_block(mbmi)) {
+    // intra mode needs decoded result such that the next transform block
+    // can use it for prediction.
+    av1_inverse_transform_block_facade(xd, plane, block, blk_row, blk_col,
+                                       x->plane[plane].eobs[block]);
+  }
+  return best_rd;
+}
+#endif  // CONFIG_TXK_SEL
diff --git a/third_party/aom/av1/encoder/encodetxb.h b/third_party/aom/av1/encoder/encodetxb.h
new file mode 100644
index 0000000000..552d47b542
--- /dev/null
+++ b/third_party/aom/av1/encoder/encodetxb.h
@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef ENCODETXB_H_
+#define ENCODETXB_H_
+
+#include "./aom_config.h"
+#include "av1/common/blockd.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/txb_common.h"
+#include "av1/encoder/block.h"
+#include "av1/encoder/encoder.h"
+#include "aom_dsp/bitwriter.h"
+#ifdef __cplusplus
+extern "C" {
+#endif
+void av1_alloc_txb_buf(AV1_COMP *cpi);
+void av1_free_txb_buf(AV1_COMP *cpi);
+int av1_cost_coeffs_txb(const AV1_COMP *const cpi, MACROBLOCK *x, int plane,
+                        int block, TXB_CTX *txb_ctx);
+void av1_write_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCKD *xd,
+                          aom_writer *w, int block, int plane,
+                          const tran_low_t *tcoeff, uint16_t eob,
+                          TXB_CTX *txb_ctx);
+void av1_write_coeffs_mb(const AV1_COMMON *const cm, MACROBLOCK *x,
+                         aom_writer *w, int plane);
+int av1_get_txb_entropy_context(const tran_low_t *qcoeff,
+                                const SCAN_ORDER *scan_order, int eob);
+void av1_update_txb_context(const AV1_COMP *cpi, ThreadData *td,
+                            RUN_TYPE dry_run, BLOCK_SIZE bsize, int *rate,
+                            const int mi_row, const int mi_col);
+void av1_write_txb_probs(AV1_COMP *cpi, aom_writer *w);
+
+#if CONFIG_TXK_SEL
+int64_t av1_search_txk_type(const AV1_COMP *cpi, MACROBLOCK *x, int plane,
+                            int block, int blk_row, int blk_col,
+                            BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
+                            const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
+                            int use_fast_coef_costing, RD_STATS *rd_stats);
+#endif
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // COEFFS_CODING_H_
diff --git a/third_party/aom/av1/encoder/ethread.c b/third_party/aom/av1/encoder/ethread.c
new file mode 100644
index 0000000000..34f0b95665
--- /dev/null
+++ b/third_party/aom/av1/encoder/ethread.c
@@ -0,0 +1,176 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/encoder/encodeframe.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/ethread.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+static void accumulate_rd_opt(ThreadData *td, ThreadData *td_t) {
+  int i, j, k, l, m, n;
+
+  for (i = 0; i < REFERENCE_MODES; i++)
+    td->rd_counts.comp_pred_diff[i] += td_t->rd_counts.comp_pred_diff[i];
+
+#if CONFIG_GLOBAL_MOTION
+  for (i = 0; i < TOTAL_REFS_PER_FRAME; i++)
+    td->rd_counts.global_motion_used[i] +=
+        td_t->rd_counts.global_motion_used[i];
+#endif  // CONFIG_GLOBAL_MOTION
+
+  for (i = 0; i < TX_SIZES; i++)
+    for (j = 0; j < PLANE_TYPES; j++)
+      for (k = 0; k < REF_TYPES; k++)
+        for (l = 0; l < COEF_BANDS; l++)
+          for (m = 0; m < COEFF_CONTEXTS; m++)
+            for (n = 0; n < ENTROPY_TOKENS; n++)
+              td->rd_counts.coef_counts[i][j][k][l][m][n] +=
+                  td_t->rd_counts.coef_counts[i][j][k][l][m][n];
+}
+
+static int enc_worker_hook(EncWorkerData *const thread_data, void *unused) {
+  AV1_COMP *const cpi = thread_data->cpi;
+  const AV1_COMMON *const cm = &cpi->common;
+  const int tile_cols = cm->tile_cols;
+  const int tile_rows = cm->tile_rows;
+  int t;
+
+  (void)unused;
+
+  for (t = thread_data->start; t < tile_rows * tile_cols;
+       t += cpi->num_workers) {
+    int tile_row = t / tile_cols;
+    int tile_col = t % tile_cols;
+
+    av1_encode_tile(cpi, thread_data->td, tile_row, tile_col);
+  }
+
+  return 0;
+}
+
+void av1_encode_tiles_mt(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  const int tile_cols = cm->tile_cols;
+  const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+  const int num_workers = AOMMIN(cpi->oxcf.max_threads, tile_cols);
+  int i;
+
+  av1_init_tile_data(cpi);
+
+  // Only run once to create threads and allocate thread data.
+  if (cpi->num_workers == 0) {
+    CHECK_MEM_ERROR(cm, cpi->workers,
+                    aom_malloc(num_workers * sizeof(*cpi->workers)));
+
+    CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
+                    aom_calloc(num_workers, sizeof(*cpi->tile_thr_data)));
+
+    for (i = 0; i < num_workers; i++) {
+      AVxWorker *const worker = &cpi->workers[i];
+      EncWorkerData *const thread_data = &cpi->tile_thr_data[i];
+
+      ++cpi->num_workers;
+      winterface->init(worker);
+
+      thread_data->cpi = cpi;
+
+      if (i < num_workers - 1) {
+        // Allocate thread data.
+        CHECK_MEM_ERROR(cm, thread_data->td,
+                        aom_memalign(32, sizeof(*thread_data->td)));
+        av1_zero(*thread_data->td);
+
+        // Set up pc_tree.
+        thread_data->td->leaf_tree = NULL;
+        thread_data->td->pc_tree = NULL;
+        av1_setup_pc_tree(cm, thread_data->td);
+
+        // Set up variance tree if needed.
+        if (cpi->sf.partition_search_type == VAR_BASED_PARTITION)
+          av1_setup_var_tree(cm, thread_data->td);
+
+        // Allocate frame counters in thread data.
+        CHECK_MEM_ERROR(cm, thread_data->td->counts,
+                        aom_calloc(1, sizeof(*thread_data->td->counts)));
+
+        // Create threads
+        if (!winterface->reset(worker))
+          aom_internal_error(&cm->error, AOM_CODEC_ERROR,
+                             "Tile encoder thread creation failed");
+      } else {
+        // Main thread acts as a worker and uses the thread data in cpi.
+        thread_data->td = &cpi->td;
+      }
+
+      winterface->sync(worker);
+    }
+  }
+
+  for (i = 0; i < num_workers; i++) {
+    AVxWorker *const worker = &cpi->workers[i];
+    EncWorkerData *thread_data;
+
+    worker->hook = (AVxWorkerHook)enc_worker_hook;
+    worker->data1 = &cpi->tile_thr_data[i];
+    worker->data2 = NULL;
+    thread_data = (EncWorkerData *)worker->data1;
+
+    // Before encoding a frame, copy the thread data from cpi.
+    if (thread_data->td != &cpi->td) {
+      thread_data->td->mb = cpi->td.mb;
+      thread_data->td->rd_counts = cpi->td.rd_counts;
+    }
+    if (thread_data->td->counts != &cpi->common.counts) {
+      memcpy(thread_data->td->counts, &cpi->common.counts,
+             sizeof(cpi->common.counts));
+    }
+
+#if CONFIG_PALETTE
+    // Allocate buffers used by palette coding mode.
+    if (cpi->common.allow_screen_content_tools && i < num_workers - 1) {
+      MACROBLOCK *x = &thread_data->td->mb;
+      CHECK_MEM_ERROR(cm, x->palette_buffer,
+                      aom_memalign(16, sizeof(*x->palette_buffer)));
+    }
+#endif  // CONFIG_PALETTE
+  }
+
+  // Encode a frame
+  for (i = 0; i < num_workers; i++) {
+    AVxWorker *const worker = &cpi->workers[i];
+    EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
+
+    // Set the starting tile for each thread.
+    thread_data->start = i;
+
+    if (i == cpi->num_workers - 1)
+      winterface->execute(worker);
+    else
+      winterface->launch(worker);
+  }
+
+  // Encoding ends.
+  for (i = 0; i < num_workers; i++) {
+    AVxWorker *const worker = &cpi->workers[i];
+    winterface->sync(worker);
+  }
+
+  for (i = 0; i < num_workers; i++) {
+    AVxWorker *const worker = &cpi->workers[i];
+    EncWorkerData *const thread_data = (EncWorkerData *)worker->data1;
+
+    // Accumulate counters.
+    if (i < cpi->num_workers - 1) {
+      av1_accumulate_frame_counts(&cm->counts, thread_data->td->counts);
+      accumulate_rd_opt(&cpi->td, thread_data->td);
+    }
+  }
+}
diff --git a/third_party/aom/av1/encoder/ethread.h b/third_party/aom/av1/encoder/ethread.h
new file mode 100644
index 0000000000..6c30a3e5cf
--- /dev/null
+++ b/third_party/aom/av1/encoder/ethread.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_ETHREAD_H_
+#define AV1_ENCODER_ETHREAD_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1_COMP;
+struct ThreadData;
+
+typedef struct EncWorkerData {
+  struct AV1_COMP *cpi;
+  struct ThreadData *td;
+  int start;
+} EncWorkerData;
+
+void av1_encode_tiles_mt(struct AV1_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_ETHREAD_H_
diff --git a/third_party/aom/av1/encoder/extend.c b/third_party/aom/av1/encoder/extend.c
new file mode 100644
index 0000000000..007694a38c
--- /dev/null
+++ b/third_party/aom/av1/encoder/extend.c
@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#include "av1/common/common.h"
+#include "av1/encoder/extend.h"
+
+static void copy_and_extend_plane(const uint8_t *src, int src_pitch,
+                                  uint8_t *dst, int dst_pitch, int w, int h,
+                                  int extend_top, int extend_left,
+                                  int extend_bottom, int extend_right) {
+  int i, linesize;
+
+  // copy the left and right most columns out
+  const uint8_t *src_ptr1 = src;
+  const uint8_t *src_ptr2 = src + w - 1;
+  uint8_t *dst_ptr1 = dst - extend_left;
+  uint8_t *dst_ptr2 = dst + w;
+
+  for (i = 0; i < h; i++) {
+    memset(dst_ptr1, src_ptr1[0], extend_left);
+    memcpy(dst_ptr1 + extend_left, src_ptr1, w);
+    memset(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_pitch;
+    src_ptr2 += src_pitch;
+    dst_ptr1 += dst_pitch;
+    dst_ptr2 += dst_pitch;
+  }
+
+  // Now copy the top and bottom lines into each line of the respective
+  // borders
+  src_ptr1 = dst - extend_left;
+  src_ptr2 = dst + dst_pitch * (h - 1) - extend_left;
+  dst_ptr1 = dst + dst_pitch * (-extend_top) - extend_left;
+  dst_ptr2 = dst + dst_pitch * (h)-extend_left;
+  linesize = extend_left + extend_right + w;
+
+  for (i = 0; i < extend_top; i++) {
+    memcpy(dst_ptr1, src_ptr1, linesize);
+    dst_ptr1 += dst_pitch;
+  }
+
+  for (i = 0; i < extend_bottom; i++) {
+    memcpy(dst_ptr2, src_ptr2, linesize);
+    dst_ptr2 += dst_pitch;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_copy_and_extend_plane(const uint8_t *src8, int src_pitch,
+                                         uint8_t *dst8, int dst_pitch, int w,
+                                         int h, int extend_top, int extend_left,
+                                         int extend_bottom, int extend_right) {
+  int i, linesize;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+
+  // copy the left and right most columns out
+  const uint16_t *src_ptr1 = src;
+  const uint16_t *src_ptr2 = src + w - 1;
+  uint16_t *dst_ptr1 = dst - extend_left;
+  uint16_t *dst_ptr2 = dst + w;
+
+  for (i = 0; i < h; i++) {
+    aom_memset16(dst_ptr1, src_ptr1[0], extend_left);
+    memcpy(dst_ptr1 + extend_left, src_ptr1, w * sizeof(src_ptr1[0]));
+    aom_memset16(dst_ptr2, src_ptr2[0], extend_right);
+    src_ptr1 += src_pitch;
+    src_ptr2 += src_pitch;
+    dst_ptr1 += dst_pitch;
+    dst_ptr2 += dst_pitch;
+  }
+
+  // Now copy the top and bottom lines into each line of the respective
+  // borders
+  src_ptr1 = dst - extend_left;
+  src_ptr2 = dst + dst_pitch * (h - 1) - extend_left;
+  dst_ptr1 = dst + dst_pitch * (-extend_top) - extend_left;
+  dst_ptr2 = dst + dst_pitch * (h)-extend_left;
+  linesize = extend_left + extend_right + w;
+
+  for (i = 0; i < extend_top; i++) {
+    memcpy(dst_ptr1, src_ptr1, linesize * sizeof(src_ptr1[0]));
+    dst_ptr1 += dst_pitch;
+  }
+
+  for (i = 0; i < extend_bottom; i++) {
+    memcpy(dst_ptr2, src_ptr2, linesize * sizeof(src_ptr2[0]));
+    dst_ptr2 += dst_pitch;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_copy_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                               YV12_BUFFER_CONFIG *dst) {
+  // Extend src frame in buffer
+  // Altref filtering assumes 16 pixel extension
+  const int et_y = 16;
+  const int el_y = 16;
+  // Motion estimation may use src block variance with the block size up
+  // to 64x64, so the right and bottom need to be extended to 64 multiple
+  // or up to 16, whichever is greater.
+  const int er_y =
+      AOMMAX(src->y_width + 16, ALIGN_POWER_OF_TWO(src->y_width, 6)) -
+      src->y_crop_width;
+  const int eb_y =
+      AOMMAX(src->y_height + 16, ALIGN_POWER_OF_TWO(src->y_height, 6)) -
+      src->y_crop_height;
+  const int uv_width_subsampling = (src->uv_width != src->y_width);
+  const int uv_height_subsampling = (src->uv_height != src->y_height);
+  const int et_uv = et_y >> uv_height_subsampling;
+  const int el_uv = el_y >> uv_width_subsampling;
+  const int eb_uv = eb_y >> uv_height_subsampling;
+  const int er_uv = er_y >> uv_width_subsampling;
+
+#if CONFIG_HIGHBITDEPTH
+  if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
+    highbd_copy_and_extend_plane(src->y_buffer, src->y_stride, dst->y_buffer,
+                                 dst->y_stride, src->y_crop_width,
+                                 src->y_crop_height, et_y, el_y, eb_y, er_y);
+
+    highbd_copy_and_extend_plane(
+        src->u_buffer, src->uv_stride, dst->u_buffer, dst->uv_stride,
+        src->uv_crop_width, src->uv_crop_height, et_uv, el_uv, eb_uv, er_uv);
+
+    highbd_copy_and_extend_plane(
+        src->v_buffer, src->uv_stride, dst->v_buffer, dst->uv_stride,
+        src->uv_crop_width, src->uv_crop_height, et_uv, el_uv, eb_uv, er_uv);
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  copy_and_extend_plane(src->y_buffer, src->y_stride, dst->y_buffer,
+                        dst->y_stride, src->y_crop_width, src->y_crop_height,
+                        et_y, el_y, eb_y, er_y);
+
+  copy_and_extend_plane(src->u_buffer, src->uv_stride, dst->u_buffer,
+                        dst->uv_stride, src->uv_crop_width, src->uv_crop_height,
+                        et_uv, el_uv, eb_uv, er_uv);
+
+  copy_and_extend_plane(src->v_buffer, src->uv_stride, dst->v_buffer,
+                        dst->uv_stride, src->uv_crop_width, src->uv_crop_height,
+                        et_uv, el_uv, eb_uv, er_uv);
+}
+
+void av1_copy_and_extend_frame_with_rect(const YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst, int srcy,
+                                         int srcx, int srch, int srcw) {
+  // If the side is not touching the bounder then don't extend.
+  const int et_y = srcy ? 0 : dst->border;
+  const int el_y = srcx ? 0 : dst->border;
+  const int eb_y = srcy + srch != src->y_height
+                       ? 0
+                       : dst->border + dst->y_height - src->y_height;
+  const int er_y = srcx + srcw != src->y_width
+                       ? 0
+                       : dst->border + dst->y_width - src->y_width;
+  const int src_y_offset = srcy * src->y_stride + srcx;
+  const int dst_y_offset = srcy * dst->y_stride + srcx;
+
+  const int et_uv = ROUND_POWER_OF_TWO(et_y, 1);
+  const int el_uv = ROUND_POWER_OF_TWO(el_y, 1);
+  const int eb_uv = ROUND_POWER_OF_TWO(eb_y, 1);
+  const int er_uv = ROUND_POWER_OF_TWO(er_y, 1);
+  const int src_uv_offset = ((srcy * src->uv_stride) >> 1) + (srcx >> 1);
+  const int dst_uv_offset = ((srcy * dst->uv_stride) >> 1) + (srcx >> 1);
+  const int srch_uv = ROUND_POWER_OF_TWO(srch, 1);
+  const int srcw_uv = ROUND_POWER_OF_TWO(srcw, 1);
+
+  copy_and_extend_plane(src->y_buffer + src_y_offset, src->y_stride,
+                        dst->y_buffer + dst_y_offset, dst->y_stride, srcw, srch,
+                        et_y, el_y, eb_y, er_y);
+
+  copy_and_extend_plane(src->u_buffer + src_uv_offset, src->uv_stride,
+                        dst->u_buffer + dst_uv_offset, dst->uv_stride, srcw_uv,
+                        srch_uv, et_uv, el_uv, eb_uv, er_uv);
+
+  copy_and_extend_plane(src->v_buffer + src_uv_offset, src->uv_stride,
+                        dst->v_buffer + dst_uv_offset, dst->uv_stride, srcw_uv,
+                        srch_uv, et_uv, el_uv, eb_uv, er_uv);
+}
diff --git a/third_party/aom/av1/encoder/extend.h b/third_party/aom/av1/encoder/extend.h
new file mode 100644
index 0000000000..48178b9647
--- /dev/null
+++ b/third_party/aom/av1/encoder/extend.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_EXTEND_H_
+#define AV1_ENCODER_EXTEND_H_
+
+#include "aom_scale/yv12config.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_copy_and_extend_frame(const YV12_BUFFER_CONFIG *src,
+                               YV12_BUFFER_CONFIG *dst);
+
+void av1_copy_and_extend_frame_with_rect(const YV12_BUFFER_CONFIG *src,
+                                         YV12_BUFFER_CONFIG *dst, int srcy,
+                                         int srcx, int srch, int srcw);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_EXTEND_H_
diff --git a/third_party/aom/av1/encoder/firstpass.c b/third_party/aom/av1/encoder/firstpass.c
new file mode 100644
index 0000000000..e35a54ef2f
--- /dev/null
+++ b/third_party/aom/av1/encoder/firstpass.c
@@ -0,0 +1,3026 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./aom_scale_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/system_state.h"
+#include "aom_scale/aom_scale.h"
+#include "aom_scale/yv12config.h"
+
+#include "aom_dsp/variance.h"
+#include "av1/common/entropymv.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/reconinter.h"  // av1_setup_dst_planes()
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/aq_variance.h"
+#include "av1/encoder/block.h"
+#include "av1/encoder/encodeframe.h"
+#include "av1/encoder/encodemb.h"
+#include "av1/encoder/encodemv.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/extend.h"
+#include "av1/encoder/firstpass.h"
+#include "av1/encoder/mcomp.h"
+#include "av1/encoder/rd.h"
+
+#define OUTPUT_FPF 0
+#define ARF_STATS_OUTPUT 0
+
+#define GROUP_ADAPTIVE_MAXQ 1
+
+#define BOOST_BREAKOUT 12.5
+#define BOOST_FACTOR 12.5
+#define FACTOR_PT_LOW 0.70
+#define FACTOR_PT_HIGH 0.90
+#define FIRST_PASS_Q 10.0
+#define GF_MAX_BOOST 96.0
+#define INTRA_MODE_PENALTY 1024
+#define KF_MAX_BOOST 128.0
+#define MIN_ARF_GF_BOOST 240
+#define MIN_DECAY_FACTOR 0.01
+#define MIN_KF_BOOST 300
+#define NEW_MV_MODE_PENALTY 32
+#define DARK_THRESH 64
+#define DEFAULT_GRP_WEIGHT 1.0
+#define RC_FACTOR_MIN 0.75
+#define RC_FACTOR_MAX 1.75
+
+#define NCOUNT_INTRA_THRESH 8192
+#define NCOUNT_INTRA_FACTOR 3
+#define NCOUNT_FRAME_II_THRESH 5.0
+
+#define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x)-0.000001 : (x) + 0.000001)
+
+#if ARF_STATS_OUTPUT
+unsigned int arf_count = 0;
+#endif
+
+// Resets the first pass file to the given position using a relative seek from
+// the current position.
+static void reset_fpf_position(TWO_PASS *p, const FIRSTPASS_STATS *position) {
+  p->stats_in = position;
+}
+
+// Read frame stats at an offset from the current position.
+static const FIRSTPASS_STATS *read_frame_stats(const TWO_PASS *p, int offset) {
+  if ((offset >= 0 && p->stats_in + offset >= p->stats_in_end) ||
+      (offset < 0 && p->stats_in + offset < p->stats_in_start)) {
+    return NULL;
+  }
+
+  return &p->stats_in[offset];
+}
+
+static int input_stats(TWO_PASS *p, FIRSTPASS_STATS *fps) {
+  if (p->stats_in >= p->stats_in_end) return EOF;
+
+  *fps = *p->stats_in;
+  ++p->stats_in;
+  return 1;
+}
+
+static void output_stats(FIRSTPASS_STATS *stats,
+                         struct aom_codec_pkt_list *pktlist) {
+  struct aom_codec_cx_pkt pkt;
+  pkt.kind = AOM_CODEC_STATS_PKT;
+  pkt.data.twopass_stats.buf = stats;
+  pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
+  aom_codec_pkt_list_add(pktlist, &pkt);
+
+// TEMP debug code
+#if OUTPUT_FPF
+  {
+    FILE *fpfile;
+    fpfile = fopen("firstpass.stt", "a");
+
+    fprintf(fpfile,
+            "%12.0lf %12.4lf %12.0lf %12.0lf %12.0lf %12.4lf %12.4lf"
+            "%12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf"
+            "%12.4lf %12.4lf %12.0lf %12.0lf %12.0lf %12.4lf\n",
+            stats->frame, stats->weight, stats->intra_error, stats->coded_error,
+            stats->sr_coded_error, stats->pcnt_inter, stats->pcnt_motion,
+            stats->pcnt_second_ref, stats->pcnt_neutral, stats->intra_skip_pct,
+            stats->inactive_zone_rows, stats->inactive_zone_cols, stats->MVr,
+            stats->mvr_abs, stats->MVc, stats->mvc_abs, stats->MVrv,
+            stats->MVcv, stats->mv_in_out_count, stats->new_mv_count,
+            stats->count, stats->duration);
+    fclose(fpfile);
+  }
+#endif
+}
+
+#if CONFIG_FP_MB_STATS
+static void output_fpmb_stats(uint8_t *this_frame_mb_stats, int stats_size,
+                              struct aom_codec_pkt_list *pktlist) {
+  struct aom_codec_cx_pkt pkt;
+  pkt.kind = AOM_CODEC_FPMB_STATS_PKT;
+  pkt.data.firstpass_mb_stats.buf = this_frame_mb_stats;
+  pkt.data.firstpass_mb_stats.sz = stats_size * sizeof(*this_frame_mb_stats);
+  aom_codec_pkt_list_add(pktlist, &pkt);
+}
+#endif
+
+static void zero_stats(FIRSTPASS_STATS *section) {
+  section->frame = 0.0;
+  section->weight = 0.0;
+  section->intra_error = 0.0;
+  section->coded_error = 0.0;
+  section->sr_coded_error = 0.0;
+  section->pcnt_inter = 0.0;
+  section->pcnt_motion = 0.0;
+  section->pcnt_second_ref = 0.0;
+  section->pcnt_neutral = 0.0;
+  section->intra_skip_pct = 0.0;
+  section->inactive_zone_rows = 0.0;
+  section->inactive_zone_cols = 0.0;
+  section->MVr = 0.0;
+  section->mvr_abs = 0.0;
+  section->MVc = 0.0;
+  section->mvc_abs = 0.0;
+  section->MVrv = 0.0;
+  section->MVcv = 0.0;
+  section->mv_in_out_count = 0.0;
+  section->new_mv_count = 0.0;
+  section->count = 0.0;
+  section->duration = 1.0;
+}
+
+static void accumulate_stats(FIRSTPASS_STATS *section,
+                             const FIRSTPASS_STATS *frame) {
+  section->frame += frame->frame;
+  section->weight += frame->weight;
+  section->intra_error += frame->intra_error;
+  section->coded_error += frame->coded_error;
+  section->sr_coded_error += frame->sr_coded_error;
+  section->pcnt_inter += frame->pcnt_inter;
+  section->pcnt_motion += frame->pcnt_motion;
+  section->pcnt_second_ref += frame->pcnt_second_ref;
+  section->pcnt_neutral += frame->pcnt_neutral;
+  section->intra_skip_pct += frame->intra_skip_pct;
+  section->inactive_zone_rows += frame->inactive_zone_rows;
+  section->inactive_zone_cols += frame->inactive_zone_cols;
+  section->MVr += frame->MVr;
+  section->mvr_abs += frame->mvr_abs;
+  section->MVc += frame->MVc;
+  section->mvc_abs += frame->mvc_abs;
+  section->MVrv += frame->MVrv;
+  section->MVcv += frame->MVcv;
+  section->mv_in_out_count += frame->mv_in_out_count;
+  section->new_mv_count += frame->new_mv_count;
+  section->count += frame->count;
+  section->duration += frame->duration;
+}
+
+static void subtract_stats(FIRSTPASS_STATS *section,
+                           const FIRSTPASS_STATS *frame) {
+  section->frame -= frame->frame;
+  section->weight -= frame->weight;
+  section->intra_error -= frame->intra_error;
+  section->coded_error -= frame->coded_error;
+  section->sr_coded_error -= frame->sr_coded_error;
+  section->pcnt_inter -= frame->pcnt_inter;
+  section->pcnt_motion -= frame->pcnt_motion;
+  section->pcnt_second_ref -= frame->pcnt_second_ref;
+  section->pcnt_neutral -= frame->pcnt_neutral;
+  section->intra_skip_pct -= frame->intra_skip_pct;
+  section->inactive_zone_rows -= frame->inactive_zone_rows;
+  section->inactive_zone_cols -= frame->inactive_zone_cols;
+  section->MVr -= frame->MVr;
+  section->mvr_abs -= frame->mvr_abs;
+  section->MVc -= frame->MVc;
+  section->mvc_abs -= frame->mvc_abs;
+  section->MVrv -= frame->MVrv;
+  section->MVcv -= frame->MVcv;
+  section->mv_in_out_count -= frame->mv_in_out_count;
+  section->new_mv_count -= frame->new_mv_count;
+  section->count -= frame->count;
+  section->duration -= frame->duration;
+}
+
+// Calculate the linear size relative to a baseline of 1080P
+#define BASE_SIZE 2073600.0  // 1920x1080
+static double get_linear_size_factor(const AV1_COMP *cpi) {
+  const double this_area = cpi->initial_width * cpi->initial_height;
+  return pow(this_area / BASE_SIZE, 0.5);
+}
+
+// Calculate an active area of the image that discounts formatting
+// bars and partially discounts other 0 energy areas.
+#define MIN_ACTIVE_AREA 0.5
+#define MAX_ACTIVE_AREA 1.0
+static double calculate_active_area(const AV1_COMP *cpi,
+                                    const FIRSTPASS_STATS *this_frame) {
+  double active_pct;
+
+  active_pct =
+      1.0 -
+      ((this_frame->intra_skip_pct / 2) +
+       ((this_frame->inactive_zone_rows * 2) / (double)cpi->common.mb_rows));
+  return fclamp(active_pct, MIN_ACTIVE_AREA, MAX_ACTIVE_AREA);
+}
+
+// Calculate a modified Error used in distributing bits between easier and
+// harder frames.
+#define ACT_AREA_CORRECTION 0.5
+static double calculate_modified_err(const AV1_COMP *cpi,
+                                     const TWO_PASS *twopass,
+                                     const AV1EncoderConfig *oxcf,
+                                     const FIRSTPASS_STATS *this_frame) {
+  const FIRSTPASS_STATS *const stats = &twopass->total_stats;
+  const double av_weight = stats->weight / stats->count;
+  const double av_err = (stats->coded_error * av_weight) / stats->count;
+  double modified_error =
+      av_err * pow(this_frame->coded_error * this_frame->weight /
+                       DOUBLE_DIVIDE_CHECK(av_err),
+                   oxcf->two_pass_vbrbias / 100.0);
+
+  // Correction for active area. Frames with a reduced active area
+  // (eg due to formatting bars) have a higher error per mb for the
+  // remaining active MBs. The correction here assumes that coding
+  // 0.5N blocks of complexity 2X is a little easier than coding N
+  // blocks of complexity X.
+  modified_error *=
+      pow(calculate_active_area(cpi, this_frame), ACT_AREA_CORRECTION);
+
+  return fclamp(modified_error, twopass->modified_error_min,
+                twopass->modified_error_max);
+}
+
+// This function returns the maximum target rate per frame.
+static int frame_max_bits(const RATE_CONTROL *rc,
+                          const AV1EncoderConfig *oxcf) {
+  int64_t max_bits = ((int64_t)rc->avg_frame_bandwidth *
+                      (int64_t)oxcf->two_pass_vbrmax_section) /
+                     100;
+  if (max_bits < 0)
+    max_bits = 0;
+  else if (max_bits > rc->max_frame_bandwidth)
+    max_bits = rc->max_frame_bandwidth;
+
+  return (int)max_bits;
+}
+
+void av1_init_first_pass(AV1_COMP *cpi) {
+  zero_stats(&cpi->twopass.total_stats);
+}
+
+void av1_end_first_pass(AV1_COMP *cpi) {
+  output_stats(&cpi->twopass.total_stats, cpi->output_pkt_list);
+}
+
+static aom_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
+  switch (bsize) {
+    case BLOCK_8X8: return aom_mse8x8;
+    case BLOCK_16X8: return aom_mse16x8;
+    case BLOCK_8X16: return aom_mse8x16;
+    default: return aom_mse16x16;
+  }
+}
+
+static unsigned int get_prediction_error(BLOCK_SIZE bsize,
+                                         const struct buf_2d *src,
+                                         const struct buf_2d *ref) {
+  unsigned int sse;
+  const aom_variance_fn_t fn = get_block_variance_fn(bsize);
+  fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
+  return sse;
+}
+
+#if CONFIG_HIGHBITDEPTH
+static aom_variance_fn_t highbd_get_block_variance_fn(BLOCK_SIZE bsize,
+                                                      int bd) {
+  switch (bd) {
+    default:
+      switch (bsize) {
+        case BLOCK_8X8: return aom_highbd_8_mse8x8;
+        case BLOCK_16X8: return aom_highbd_8_mse16x8;
+        case BLOCK_8X16: return aom_highbd_8_mse8x16;
+        default: return aom_highbd_8_mse16x16;
+      }
+      break;
+    case 10:
+      switch (bsize) {
+        case BLOCK_8X8: return aom_highbd_10_mse8x8;
+        case BLOCK_16X8: return aom_highbd_10_mse16x8;
+        case BLOCK_8X16: return aom_highbd_10_mse8x16;
+        default: return aom_highbd_10_mse16x16;
+      }
+      break;
+    case 12:
+      switch (bsize) {
+        case BLOCK_8X8: return aom_highbd_12_mse8x8;
+        case BLOCK_16X8: return aom_highbd_12_mse16x8;
+        case BLOCK_8X16: return aom_highbd_12_mse8x16;
+        default: return aom_highbd_12_mse16x16;
+      }
+      break;
+  }
+}
+
+static unsigned int highbd_get_prediction_error(BLOCK_SIZE bsize,
+                                                const struct buf_2d *src,
+                                                const struct buf_2d *ref,
+                                                int bd) {
+  unsigned int sse;
+  const aom_variance_fn_t fn = highbd_get_block_variance_fn(bsize, bd);
+  fn(src->buf, src->stride, ref->buf, ref->stride, &sse);
+  return sse;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+// Refine the motion search range according to the frame dimension
+// for first pass test.
+static int get_search_range(const AV1_COMP *cpi) {
+  int sr = 0;
+  const int dim = AOMMIN(cpi->initial_width, cpi->initial_height);
+
+  while ((dim << sr) < MAX_FULL_PEL_VAL) ++sr;
+  return sr;
+}
+
+static void first_pass_motion_search(AV1_COMP *cpi, MACROBLOCK *x,
+                                     const MV *ref_mv, MV *best_mv,
+                                     int *best_motion_err) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MV tmp_mv = { 0, 0 };
+  MV ref_mv_full = { ref_mv->row >> 3, ref_mv->col >> 3 };
+  int num00, tmp_err, n;
+  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  aom_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
+  const int new_mv_mode_penalty = NEW_MV_MODE_PENALTY;
+
+  int step_param = 3;
+  int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
+  const int sr = get_search_range(cpi);
+  step_param += sr;
+  further_steps -= sr;
+
+  // Override the default variance function to use MSE.
+  v_fn_ptr.vf = get_block_variance_fn(bsize);
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    v_fn_ptr.vf = highbd_get_block_variance_fn(bsize, xd->bd);
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  // Center the initial step/diamond search on best mv.
+  tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
+                                    step_param, x->sadperbit16, &num00,
+                                    &v_fn_ptr, ref_mv);
+  if (tmp_err < INT_MAX)
+    tmp_err = av1_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
+  if (tmp_err < INT_MAX - new_mv_mode_penalty) tmp_err += new_mv_mode_penalty;
+
+  if (tmp_err < *best_motion_err) {
+    *best_motion_err = tmp_err;
+    *best_mv = tmp_mv;
+  }
+
+  // Carry out further step/diamond searches as necessary.
+  n = num00;
+  num00 = 0;
+
+  while (n < further_steps) {
+    ++n;
+
+    if (num00) {
+      --num00;
+    } else {
+      tmp_err = cpi->diamond_search_sad(x, &cpi->ss_cfg, &ref_mv_full, &tmp_mv,
+                                        step_param + n, x->sadperbit16, &num00,
+                                        &v_fn_ptr, ref_mv);
+      if (tmp_err < INT_MAX)
+        tmp_err = av1_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
+      if (tmp_err < INT_MAX - new_mv_mode_penalty)
+        tmp_err += new_mv_mode_penalty;
+
+      if (tmp_err < *best_motion_err) {
+        *best_motion_err = tmp_err;
+        *best_mv = tmp_mv;
+      }
+    }
+  }
+}
+
+static BLOCK_SIZE get_bsize(const AV1_COMMON *cm, int mb_row, int mb_col) {
+  if (mi_size_wide[BLOCK_16X16] * mb_col + mi_size_wide[BLOCK_8X8] <
+      cm->mi_cols) {
+    return mi_size_wide[BLOCK_16X16] * mb_row + mi_size_wide[BLOCK_8X8] <
+                   cm->mi_rows
+               ? BLOCK_16X16
+               : BLOCK_16X8;
+  } else {
+    return mi_size_wide[BLOCK_16X16] * mb_row + mi_size_wide[BLOCK_8X8] <
+                   cm->mi_rows
+               ? BLOCK_8X16
+               : BLOCK_8X8;
+  }
+}
+
+static int find_fp_qindex(aom_bit_depth_t bit_depth) {
+  int i;
+
+  for (i = 0; i < QINDEX_RANGE; ++i)
+    if (av1_convert_qindex_to_q(i, bit_depth) >= FIRST_PASS_Q) break;
+
+  if (i == QINDEX_RANGE) i--;
+
+  return i;
+}
+
+static void set_first_pass_params(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  if (!cpi->refresh_alt_ref_frame &&
+      (cm->current_video_frame == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY))) {
+    cm->frame_type = KEY_FRAME;
+  } else {
+    cm->frame_type = INTER_FRAME;
+  }
+  // Do not use periodic key frames.
+  cpi->rc.frames_to_key = INT_MAX;
+}
+
+#define UL_INTRA_THRESH 50
+#define INVALID_ROW -1
+void av1_first_pass(AV1_COMP *cpi, const struct lookahead_entry *source) {
+  int mb_row, mb_col;
+  MACROBLOCK *const x = &cpi->td.mb;
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  TileInfo tile;
+  struct macroblock_plane *const p = x->plane;
+  struct macroblockd_plane *const pd = xd->plane;
+  const PICK_MODE_CONTEXT *ctx =
+      &cpi->td.pc_root[MAX_MIB_SIZE_LOG2 - MIN_MIB_SIZE_LOG2]->none;
+  int i;
+
+  int recon_yoffset, recon_uvoffset;
+  int64_t intra_error = 0;
+  int64_t coded_error = 0;
+  int64_t sr_coded_error = 0;
+
+  int sum_mvr = 0, sum_mvc = 0;
+  int sum_mvr_abs = 0, sum_mvc_abs = 0;
+  int64_t sum_mvrs = 0, sum_mvcs = 0;
+  int mvcount = 0;
+  int intercount = 0;
+  int second_ref_count = 0;
+  const int intrapenalty = INTRA_MODE_PENALTY;
+  double neutral_count;
+  int intra_skip_count = 0;
+  int image_data_start_row = INVALID_ROW;
+  int new_mv_count = 0;
+  int sum_in_vectors = 0;
+  MV lastmv = { 0, 0 };
+  TWO_PASS *twopass = &cpi->twopass;
+  const MV zero_mv = { 0, 0 };
+  int recon_y_stride, recon_uv_stride, uv_mb_height;
+
+  YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
+  YV12_BUFFER_CONFIG *gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
+  YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
+  const YV12_BUFFER_CONFIG *first_ref_buf = lst_yv12;
+  double intra_factor;
+  double brightness_factor;
+  BufferPool *const pool = cm->buffer_pool;
+  const int qindex = find_fp_qindex(cm->bit_depth);
+  const int mb_scale = mi_size_wide[BLOCK_16X16];
+#if CONFIG_PVQ
+  PVQ_QUEUE pvq_q;
+  od_adapt_ctx pvq_context;
+#endif
+
+  // First pass code requires valid last and new frame buffers.
+  assert(new_yv12 != NULL);
+  assert(frame_is_intra_only(cm) || (lst_yv12 != NULL));
+
+#if CONFIG_FP_MB_STATS
+  if (cpi->use_fp_mb_stats) {
+    av1_zero_array(cpi->twopass.frame_mb_stats_buf, cpi->initial_mbs);
+  }
+#endif
+
+  aom_clear_system_state();
+
+  xd->mi = cm->mi_grid_visible;
+  xd->mi[0] = cm->mi;
+  x->e_mbd.mi[0]->mbmi.sb_type = BLOCK_16X16;
+
+  intra_factor = 0.0;
+  brightness_factor = 0.0;
+  neutral_count = 0.0;
+
+  set_first_pass_params(cpi);
+  av1_set_quantizer(cm, qindex);
+
+  av1_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
+
+  av1_setup_src_planes(x, cpi->source, 0, 0);
+  av1_setup_dst_planes(xd->plane, cm->sb_size, new_yv12, 0, 0);
+
+  if (!frame_is_intra_only(cm)) {
+    av1_setup_pre_planes(xd, 0, first_ref_buf, 0, 0, NULL);
+  }
+
+  xd->mi = cm->mi_grid_visible;
+  xd->mi[0] = cm->mi;
+
+#if CONFIG_CFL
+  // Don't store luma on the fist pass since chroma is not computed
+  x->cfl_store_y = 0;
+#endif
+  av1_frame_init_quantizer(cpi);
+
+#if CONFIG_PVQ
+  // For pass 1 of 2-pass encoding, init here for PVQ for now.
+  {
+    pvq_q.buf_len = 5000;
+    CHECK_MEM_ERROR(cm, pvq_q.buf,
+                    aom_malloc(pvq_q.buf_len * sizeof(PVQ_INFO)));
+    pvq_q.curr_pos = 0;
+    x->pvq_coded = 0;
+
+    x->pvq_q = &pvq_q;
+
+    // TODO(yushin): Since this init step is also called in 2nd pass,
+    // or 1-pass encoding, consider factoring out it as a function.
+    // TODO(yushin)
+    // If activity masking is enabled, change below to OD_HVS_QM
+    x->daala_enc.qm = OD_FLAT_QM;  // Hard coded. Enc/dec required to sync.
+    x->daala_enc.pvq_norm_lambda = OD_PVQ_LAMBDA;
+    x->daala_enc.pvq_norm_lambda_dc = OD_PVQ_LAMBDA;
+
+    od_init_qm(x->daala_enc.state.qm, x->daala_enc.state.qm_inv,
+               x->daala_enc.qm == OD_HVS_QM ? OD_QM8_Q4_HVS : OD_QM8_Q4_FLAT);
+#if CONFIG_DAALA_EC
+    od_ec_enc_init(&x->daala_enc.w.ec, 65025);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+
+#if CONFIG_DAALA_EC
+    od_ec_enc_reset(&x->daala_enc.w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  }
+#endif
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    p[i].coeff = ctx->coeff[i];
+    p[i].qcoeff = ctx->qcoeff[i];
+    pd[i].dqcoeff = ctx->dqcoeff[i];
+#if CONFIG_PVQ
+    pd[i].pvq_ref_coeff = ctx->pvq_ref_coeff[i];
+#endif
+    p[i].eobs = ctx->eobs[i];
+#if CONFIG_LV_MAP
+    p[i].txb_entropy_ctx = ctx->txb_entropy_ctx[i];
+#endif
+  }
+
+  av1_init_mv_probs(cm);
+#if CONFIG_ADAPT_SCAN
+  av1_init_scan_order(cm);
+#endif
+  av1_convolve_init(cm);
+#if CONFIG_PVQ
+  od_adapt_ctx_reset(&pvq_context, 0);
+  x->daala_enc.state.adapt = &pvq_context;
+#endif  // CONFIG_PVQ
+  av1_initialize_rd_consts(cpi);
+
+  // Tiling is ignored in the first pass.
+  av1_tile_init(&tile, cm, 0, 0);
+
+  recon_y_stride = new_yv12->y_stride;
+  recon_uv_stride = new_yv12->uv_stride;
+  uv_mb_height = 16 >> (new_yv12->y_height > new_yv12->uv_height);
+
+  for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
+    MV best_ref_mv = { 0, 0 };
+
+    // Reset above block coeffs.
+    xd->up_available = (mb_row != 0);
+    recon_yoffset = (mb_row * recon_y_stride * 16);
+    recon_uvoffset = (mb_row * recon_uv_stride * uv_mb_height);
+
+    // Set up limit values for motion vectors to prevent them extending
+    // outside the UMV borders.
+    x->mv_limits.row_min = -((mb_row * 16) + BORDER_MV_PIXELS_B16);
+    x->mv_limits.row_max =
+        ((cm->mb_rows - 1 - mb_row) * 16) + BORDER_MV_PIXELS_B16;
+
+    for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
+      int this_error;
+      const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
+      const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);
+      double log_intra;
+      int level_sample;
+
+#if CONFIG_FP_MB_STATS
+      const int mb_index = mb_row * cm->mb_cols + mb_col;
+#endif
+
+      aom_clear_system_state();
+
+      xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
+      xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
+      xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
+      xd->left_available = (mb_col != 0);
+      xd->mi[0]->mbmi.sb_type = bsize;
+      xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
+      set_mi_row_col(xd, &tile, mb_row * mb_scale, mi_size_high[bsize],
+                     mb_col * mb_scale, mi_size_wide[bsize],
+#if CONFIG_DEPENDENT_HORZTILES
+                     cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                     cm->mi_rows, cm->mi_cols);
+
+      set_plane_n4(xd, mi_size_wide[bsize], mi_size_high[bsize]);
+
+      // Do intra 16x16 prediction.
+      xd->mi[0]->mbmi.segment_id = 0;
+#if CONFIG_SUPERTX
+      xd->mi[0]->mbmi.segment_id_supertx = 0;
+#endif  // CONFIG_SUPERTX
+      xd->lossless[xd->mi[0]->mbmi.segment_id] = (qindex == 0);
+      xd->mi[0]->mbmi.mode = DC_PRED;
+      xd->mi[0]->mbmi.tx_size =
+          use_dc_pred ? (bsize >= BLOCK_16X16 ? TX_16X16 : TX_8X8) : TX_4X4;
+      av1_encode_intra_block_plane(cm, x, bsize, 0, 0, mb_row * 2, mb_col * 2);
+      this_error = aom_get_mb_ss(x->plane[0].src_diff);
+
+      // Keep a record of blocks that have almost no intra error residual
+      // (i.e. are in effect completely flat and untextured in the intra
+      // domain). In natural videos this is uncommon, but it is much more
+      // common in animations, graphics and screen content, so may be used
+      // as a signal to detect these types of content.
+      if (this_error < UL_INTRA_THRESH) {
+        ++intra_skip_count;
+      } else if ((mb_col > 0) && (image_data_start_row == INVALID_ROW)) {
+        image_data_start_row = mb_row;
+      }
+
+#if CONFIG_HIGHBITDEPTH
+      if (cm->use_highbitdepth) {
+        switch (cm->bit_depth) {
+          case AOM_BITS_8: break;
+          case AOM_BITS_10: this_error >>= 4; break;
+          case AOM_BITS_12: this_error >>= 8; break;
+          default:
+            assert(0 &&
+                   "cm->bit_depth should be AOM_BITS_8, "
+                   "AOM_BITS_10 or AOM_BITS_12");
+            return;
+        }
+      }
+#endif  // CONFIG_HIGHBITDEPTH
+
+      aom_clear_system_state();
+      log_intra = log(this_error + 1.0);
+      if (log_intra < 10.0)
+        intra_factor += 1.0 + ((10.0 - log_intra) * 0.05);
+      else
+        intra_factor += 1.0;
+
+#if CONFIG_HIGHBITDEPTH
+      if (cm->use_highbitdepth)
+        level_sample = CONVERT_TO_SHORTPTR(x->plane[0].src.buf)[0];
+      else
+        level_sample = x->plane[0].src.buf[0];
+#else
+      level_sample = x->plane[0].src.buf[0];
+#endif
+      if ((level_sample < DARK_THRESH) && (log_intra < 9.0))
+        brightness_factor += 1.0 + (0.01 * (DARK_THRESH - level_sample));
+      else
+        brightness_factor += 1.0;
+
+      // Intrapenalty below deals with situations where the intra and inter
+      // error scores are very low (e.g. a plain black frame).
+      // We do not have special cases in first pass for 0,0 and nearest etc so
+      // all inter modes carry an overhead cost estimate for the mv.
+      // When the error score is very low this causes us to pick all or lots of
+      // INTRA modes and throw lots of key frames.
+      // This penalty adds a cost matching that of a 0,0 mv to the intra case.
+      this_error += intrapenalty;
+
+      // Accumulate the intra error.
+      intra_error += (int64_t)this_error;
+
+#if CONFIG_FP_MB_STATS
+      if (cpi->use_fp_mb_stats) {
+        // initialization
+        cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
+      }
+#endif
+
+      // Set up limit values for motion vectors to prevent them extending
+      // outside the UMV borders.
+      x->mv_limits.col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
+      x->mv_limits.col_max =
+          ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;
+
+      if (!frame_is_intra_only(cm)) {  // Do a motion search
+        int tmp_err, motion_error, raw_motion_error;
+        // Assume 0,0 motion with no mv overhead.
+        MV mv = { 0, 0 }, tmp_mv = { 0, 0 };
+        struct buf_2d unscaled_last_source_buf_2d;
+
+        xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
+#if CONFIG_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          motion_error = highbd_get_prediction_error(
+              bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
+        } else {
+          motion_error = get_prediction_error(bsize, &x->plane[0].src,
+                                              &xd->plane[0].pre[0]);
+        }
+#else
+        motion_error =
+            get_prediction_error(bsize, &x->plane[0].src, &xd->plane[0].pre[0]);
+#endif  // CONFIG_HIGHBITDEPTH
+
+        // Compute the motion error of the 0,0 motion using the last source
+        // frame as the reference. Skip the further motion search on
+        // reconstructed frame if this error is small.
+        unscaled_last_source_buf_2d.buf =
+            cpi->unscaled_last_source->y_buffer + recon_yoffset;
+        unscaled_last_source_buf_2d.stride =
+            cpi->unscaled_last_source->y_stride;
+#if CONFIG_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          raw_motion_error = highbd_get_prediction_error(
+              bsize, &x->plane[0].src, &unscaled_last_source_buf_2d, xd->bd);
+        } else {
+          raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
+                                                  &unscaled_last_source_buf_2d);
+        }
+#else
+        raw_motion_error = get_prediction_error(bsize, &x->plane[0].src,
+                                                &unscaled_last_source_buf_2d);
+#endif  // CONFIG_HIGHBITDEPTH
+
+        // TODO(pengchong): Replace the hard-coded threshold
+        if (raw_motion_error > 25) {
+          // Test last reference frame using the previous best mv as the
+          // starting point (best reference) for the search.
+          first_pass_motion_search(cpi, x, &best_ref_mv, &mv, &motion_error);
+
+          // If the current best reference mv is not centered on 0,0 then do a
+          // 0,0 based search as well.
+          if (!is_zero_mv(&best_ref_mv)) {
+            tmp_err = INT_MAX;
+            first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv, &tmp_err);
+
+            if (tmp_err < motion_error) {
+              motion_error = tmp_err;
+              mv = tmp_mv;
+            }
+          }
+
+          // Search in an older reference frame.
+          if ((cm->current_video_frame > 1) && gld_yv12 != NULL) {
+            // Assume 0,0 motion with no mv overhead.
+            int gf_motion_error;
+
+            xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
+#if CONFIG_HIGHBITDEPTH
+            if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+              gf_motion_error = highbd_get_prediction_error(
+                  bsize, &x->plane[0].src, &xd->plane[0].pre[0], xd->bd);
+            } else {
+              gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
+                                                     &xd->plane[0].pre[0]);
+            }
+#else
+            gf_motion_error = get_prediction_error(bsize, &x->plane[0].src,
+                                                   &xd->plane[0].pre[0]);
+#endif  // CONFIG_HIGHBITDEPTH
+
+            first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv,
+                                     &gf_motion_error);
+
+            if (gf_motion_error < motion_error && gf_motion_error < this_error)
+              ++second_ref_count;
+
+            // Reset to last frame as reference buffer.
+            xd->plane[0].pre[0].buf = first_ref_buf->y_buffer + recon_yoffset;
+            xd->plane[1].pre[0].buf = first_ref_buf->u_buffer + recon_uvoffset;
+            xd->plane[2].pre[0].buf = first_ref_buf->v_buffer + recon_uvoffset;
+
+            // In accumulating a score for the older reference frame take the
+            // best of the motion predicted score and the intra coded error
+            // (just as will be done for) accumulation of "coded_error" for
+            // the last frame.
+            if (gf_motion_error < this_error)
+              sr_coded_error += gf_motion_error;
+            else
+              sr_coded_error += this_error;
+          } else {
+            sr_coded_error += motion_error;
+          }
+        } else {
+          sr_coded_error += motion_error;
+        }
+
+        // Start by assuming that intra mode is best.
+        best_ref_mv.row = 0;
+        best_ref_mv.col = 0;
+
+#if CONFIG_FP_MB_STATS
+        if (cpi->use_fp_mb_stats) {
+          // intra predication statistics
+          cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
+          cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_DCINTRA_MASK;
+          cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
+          if (this_error > FPMB_ERROR_LARGE_TH) {
+            cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_LARGE_MASK;
+          } else if (this_error < FPMB_ERROR_SMALL_TH) {
+            cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_ERROR_SMALL_MASK;
+          }
+        }
+#endif
+
+        if (motion_error <= this_error) {
+          aom_clear_system_state();
+
+          // Keep a count of cases where the inter and intra were very close
+          // and very low. This helps with scene cut detection for example in
+          // cropped clips with black bars at the sides or top and bottom.
+          if (((this_error - intrapenalty) * 9 <= motion_error * 10) &&
+              (this_error < (2 * intrapenalty))) {
+            neutral_count += 1.0;
+            // Also track cases where the intra is not much worse than the inter
+            // and use this in limiting the GF/arf group length.
+          } else if ((this_error > NCOUNT_INTRA_THRESH) &&
+                     (this_error < (NCOUNT_INTRA_FACTOR * motion_error))) {
+            neutral_count +=
+                (double)motion_error / DOUBLE_DIVIDE_CHECK((double)this_error);
+          }
+
+          mv.row *= 8;
+          mv.col *= 8;
+          this_error = motion_error;
+          xd->mi[0]->mbmi.mode = NEWMV;
+          xd->mi[0]->mbmi.mv[0].as_mv = mv;
+          xd->mi[0]->mbmi.tx_size = TX_4X4;
+          xd->mi[0]->mbmi.ref_frame[0] = LAST_FRAME;
+          xd->mi[0]->mbmi.ref_frame[1] = NONE_FRAME;
+          av1_build_inter_predictors_sby(xd, mb_row * mb_scale,
+                                         mb_col * mb_scale, NULL, bsize);
+          av1_encode_sby_pass1(cm, x, bsize);
+          sum_mvr += mv.row;
+          sum_mvr_abs += abs(mv.row);
+          sum_mvc += mv.col;
+          sum_mvc_abs += abs(mv.col);
+          sum_mvrs += mv.row * mv.row;
+          sum_mvcs += mv.col * mv.col;
+          ++intercount;
+
+          best_ref_mv = mv;
+
+#if CONFIG_FP_MB_STATS
+          if (cpi->use_fp_mb_stats) {
+            // inter predication statistics
+            cpi->twopass.frame_mb_stats_buf[mb_index] = 0;
+            cpi->twopass.frame_mb_stats_buf[mb_index] &= ~FPMB_DCINTRA_MASK;
+            cpi->twopass.frame_mb_stats_buf[mb_index] |= FPMB_MOTION_ZERO_MASK;
+            if (this_error > FPMB_ERROR_LARGE_TH) {
+              cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                  FPMB_ERROR_LARGE_MASK;
+            } else if (this_error < FPMB_ERROR_SMALL_TH) {
+              cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                  FPMB_ERROR_SMALL_MASK;
+            }
+          }
+#endif
+
+          if (!is_zero_mv(&mv)) {
+            ++mvcount;
+
+#if CONFIG_FP_MB_STATS
+            if (cpi->use_fp_mb_stats) {
+              cpi->twopass.frame_mb_stats_buf[mb_index] &=
+                  ~FPMB_MOTION_ZERO_MASK;
+              // check estimated motion direction
+              if (mv.col > 0 && mv.col >= abs(mv.row)) {
+                // right direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_RIGHT_MASK;
+              } else if (mv.row < 0 && abs(mv.row) >= abs(mv.col)) {
+                // up direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_UP_MASK;
+              } else if (mv.col < 0 && abs(mv.col) >= abs(mv.row)) {
+                // left direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_LEFT_MASK;
+              } else {
+                // down direction
+                cpi->twopass.frame_mb_stats_buf[mb_index] |=
+                    FPMB_MOTION_DOWN_MASK;
+              }
+            }
+#endif
+
+            // Non-zero vector, was it different from the last non zero vector?
+            if (!is_equal_mv(&mv, &lastmv)) ++new_mv_count;
+            lastmv = mv;
+
+            // Does the row vector point inwards or outwards?
+            if (mb_row < cm->mb_rows / 2) {
+              if (mv.row > 0)
+                --sum_in_vectors;
+              else if (mv.row < 0)
+                ++sum_in_vectors;
+            } else if (mb_row > cm->mb_rows / 2) {
+              if (mv.row > 0)
+                ++sum_in_vectors;
+              else if (mv.row < 0)
+                --sum_in_vectors;
+            }
+
+            // Does the col vector point inwards or outwards?
+            if (mb_col < cm->mb_cols / 2) {
+              if (mv.col > 0)
+                --sum_in_vectors;
+              else if (mv.col < 0)
+                ++sum_in_vectors;
+            } else if (mb_col > cm->mb_cols / 2) {
+              if (mv.col > 0)
+                ++sum_in_vectors;
+              else if (mv.col < 0)
+                --sum_in_vectors;
+            }
+          }
+        }
+      } else {
+        sr_coded_error += (int64_t)this_error;
+      }
+      coded_error += (int64_t)this_error;
+
+      // Adjust to the next column of MBs.
+      x->plane[0].src.buf += 16;
+      x->plane[1].src.buf += uv_mb_height;
+      x->plane[2].src.buf += uv_mb_height;
+
+      recon_yoffset += 16;
+      recon_uvoffset += uv_mb_height;
+    }
+
+    // Adjust to the next row of MBs.
+    x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
+    x->plane[1].src.buf +=
+        uv_mb_height * x->plane[1].src.stride - uv_mb_height * cm->mb_cols;
+    x->plane[2].src.buf +=
+        uv_mb_height * x->plane[1].src.stride - uv_mb_height * cm->mb_cols;
+
+    aom_clear_system_state();
+  }
+
+#if CONFIG_PVQ
+#if CONFIG_DAALA_EC
+  od_ec_enc_clear(&x->daala_enc.w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+
+  x->pvq_q->last_pos = x->pvq_q->curr_pos;
+  x->pvq_q->curr_pos = 0;
+  x->pvq_q = NULL;
+
+  aom_free(pvq_q.buf);
+#endif
+
+  // Clamp the image start to rows/2. This number of rows is discarded top
+  // and bottom as dead data so rows / 2 means the frame is blank.
+  if ((image_data_start_row > cm->mb_rows / 2) ||
+      (image_data_start_row == INVALID_ROW)) {
+    image_data_start_row = cm->mb_rows / 2;
+  }
+  // Exclude any image dead zone
+  if (image_data_start_row > 0) {
+    intra_skip_count =
+        AOMMAX(0, intra_skip_count - (image_data_start_row * cm->mb_cols * 2));
+  }
+
+  {
+    FIRSTPASS_STATS fps;
+    // The minimum error here insures some bit allocation to frames even
+    // in static regions. The allocation per MB declines for larger formats
+    // where the typical "real" energy per MB also falls.
+    // Initial estimate here uses sqrt(mbs) to define the min_err, where the
+    // number of mbs is proportional to the image area.
+    const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
+                            ? cpi->initial_mbs
+                            : cpi->common.MBs;
+    const double min_err = 200 * sqrt(num_mbs);
+
+    intra_factor = intra_factor / (double)num_mbs;
+    brightness_factor = brightness_factor / (double)num_mbs;
+    fps.weight = intra_factor * brightness_factor;
+
+    fps.frame = cm->current_video_frame;
+    fps.coded_error = (double)(coded_error >> 8) + min_err;
+    fps.sr_coded_error = (double)(sr_coded_error >> 8) + min_err;
+    fps.intra_error = (double)(intra_error >> 8) + min_err;
+    fps.count = 1.0;
+    fps.pcnt_inter = (double)intercount / num_mbs;
+    fps.pcnt_second_ref = (double)second_ref_count / num_mbs;
+    fps.pcnt_neutral = (double)neutral_count / num_mbs;
+    fps.intra_skip_pct = (double)intra_skip_count / num_mbs;
+    fps.inactive_zone_rows = (double)image_data_start_row;
+    fps.inactive_zone_cols = (double)0;  // TODO(paulwilkins): fix
+
+    if (mvcount > 0) {
+      fps.MVr = (double)sum_mvr / mvcount;
+      fps.mvr_abs = (double)sum_mvr_abs / mvcount;
+      fps.MVc = (double)sum_mvc / mvcount;
+      fps.mvc_abs = (double)sum_mvc_abs / mvcount;
+      fps.MVrv =
+          ((double)sum_mvrs - ((double)sum_mvr * sum_mvr / mvcount)) / mvcount;
+      fps.MVcv =
+          ((double)sum_mvcs - ((double)sum_mvc * sum_mvc / mvcount)) / mvcount;
+      fps.mv_in_out_count = (double)sum_in_vectors / (mvcount * 2);
+      fps.new_mv_count = new_mv_count;
+      fps.pcnt_motion = (double)mvcount / num_mbs;
+    } else {
+      fps.MVr = 0.0;
+      fps.mvr_abs = 0.0;
+      fps.MVc = 0.0;
+      fps.mvc_abs = 0.0;
+      fps.MVrv = 0.0;
+      fps.MVcv = 0.0;
+      fps.mv_in_out_count = 0.0;
+      fps.new_mv_count = 0.0;
+      fps.pcnt_motion = 0.0;
+    }
+
+    // TODO(paulwilkins):  Handle the case when duration is set to 0, or
+    // something less than the full time between subsequent values of
+    // cpi->source_time_stamp.
+    fps.duration = (double)(source->ts_end - source->ts_start);
+
+    // Don't want to do output stats with a stack variable!
+    twopass->this_frame_stats = fps;
+    output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
+    accumulate_stats(&twopass->total_stats, &fps);
+
+#if CONFIG_FP_MB_STATS
+    if (cpi->use_fp_mb_stats) {
+      output_fpmb_stats(twopass->frame_mb_stats_buf, cpi->initial_mbs,
+                        cpi->output_pkt_list);
+    }
+#endif
+  }
+
+  // Copy the previous Last Frame back into gf and and arf buffers if
+  // the prediction is good enough... but also don't allow it to lag too far.
+  if ((twopass->sr_update_lag > 3) ||
+      ((cm->current_video_frame > 0) &&
+       (twopass->this_frame_stats.pcnt_inter > 0.20) &&
+       ((twopass->this_frame_stats.intra_error /
+         DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
+    if (gld_yv12 != NULL) {
+#if CONFIG_EXT_REFS
+      ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
+                 cm->ref_frame_map[cpi->lst_fb_idxes[LAST_FRAME - LAST_FRAME]]);
+#else
+      ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
+                 cm->ref_frame_map[cpi->lst_fb_idx]);
+#endif  // CONFIG_EXT_REFS
+    }
+    twopass->sr_update_lag = 1;
+  } else {
+    ++twopass->sr_update_lag;
+  }
+
+  aom_extend_frame_borders(new_yv12);
+
+// The frame we just compressed now becomes the last frame.
+#if CONFIG_EXT_REFS
+  ref_cnt_fb(pool->frame_bufs,
+             &cm->ref_frame_map[cpi->lst_fb_idxes[LAST_FRAME - LAST_FRAME]],
+             cm->new_fb_idx);
+#else
+  ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->lst_fb_idx],
+             cm->new_fb_idx);
+#endif  // CONFIG_EXT_REFS
+
+  // Special case for the first frame. Copy into the GF buffer as a second
+  // reference.
+  if (cm->current_video_frame == 0 && cpi->gld_fb_idx != INVALID_IDX) {
+#if CONFIG_EXT_REFS
+    ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
+               cm->ref_frame_map[cpi->lst_fb_idxes[LAST_FRAME - LAST_FRAME]]);
+#else
+    ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
+               cm->ref_frame_map[cpi->lst_fb_idx]);
+#endif  // CONFIG_EXT_REFS
+  }
+
+  // Use this to see what the first pass reconstruction looks like.
+  if (0) {
+    char filename[512];
+    FILE *recon_file;
+    snprintf(filename, sizeof(filename), "enc%04d.yuv",
+             (int)cm->current_video_frame);
+
+    if (cm->current_video_frame == 0)
+      recon_file = fopen(filename, "wb");
+    else
+      recon_file = fopen(filename, "ab");
+
+    (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
+    fclose(recon_file);
+  }
+
+  ++cm->current_video_frame;
+}
+
+static double calc_correction_factor(double err_per_mb, double err_divisor,
+                                     double pt_low, double pt_high, int q,
+                                     aom_bit_depth_t bit_depth) {
+  const double error_term = err_per_mb / err_divisor;
+
+  // Adjustment based on actual quantizer to power term.
+  const double power_term =
+      AOMMIN(av1_convert_qindex_to_q(q, bit_depth) * 0.01 + pt_low, pt_high);
+
+  // Calculate correction factor.
+  if (power_term < 1.0) assert(error_term >= 0.0);
+
+  return fclamp(pow(error_term, power_term), 0.05, 5.0);
+}
+
+#define ERR_DIVISOR 100.0
+static int get_twopass_worst_quality(const AV1_COMP *cpi,
+                                     const double section_err,
+                                     double inactive_zone,
+                                     int section_target_bandwidth,
+                                     double group_weight_factor) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+
+  inactive_zone = fclamp(inactive_zone, 0.0, 1.0);
+
+  if (section_target_bandwidth <= 0) {
+    return rc->worst_quality;  // Highest value allowed
+  } else {
+    const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
+                            ? cpi->initial_mbs
+                            : cpi->common.MBs;
+    const int active_mbs = AOMMAX(1, num_mbs - (int)(num_mbs * inactive_zone));
+    const double av_err_per_mb = section_err / active_mbs;
+    const double speed_term = 1.0 + 0.04 * oxcf->speed;
+    double ediv_size_correction;
+    const int target_norm_bits_per_mb =
+        (int)((uint64_t)section_target_bandwidth << BPER_MB_NORMBITS) /
+        active_mbs;
+    int q;
+
+    // Larger image formats are expected to be a little harder to code
+    // relatively given the same prediction error score. This in part at
+    // least relates to the increased size and hence coding overheads of
+    // motion vectors. Some account of this is made through adjustment of
+    // the error divisor.
+    ediv_size_correction =
+        AOMMAX(0.2, AOMMIN(5.0, get_linear_size_factor(cpi)));
+    if (ediv_size_correction < 1.0)
+      ediv_size_correction = -(1.0 / ediv_size_correction);
+    ediv_size_correction *= 4.0;
+
+    // Try and pick a max Q that will be high enough to encode the
+    // content at the given rate.
+    for (q = rc->best_quality; q < rc->worst_quality; ++q) {
+      const double factor = calc_correction_factor(
+          av_err_per_mb, ERR_DIVISOR - ediv_size_correction, FACTOR_PT_LOW,
+          FACTOR_PT_HIGH, q, cpi->common.bit_depth);
+      const int bits_per_mb = av1_rc_bits_per_mb(
+          INTER_FRAME, q, factor * speed_term * group_weight_factor,
+          cpi->common.bit_depth);
+      if (bits_per_mb <= target_norm_bits_per_mb) break;
+    }
+
+    // Restriction on active max q for constrained quality mode.
+    if (cpi->oxcf.rc_mode == AOM_CQ) q = AOMMAX(q, oxcf->cq_level);
+    return q;
+  }
+}
+
+static void setup_rf_level_maxq(AV1_COMP *cpi) {
+  int i;
+  RATE_CONTROL *const rc = &cpi->rc;
+  for (i = INTER_NORMAL; i < RATE_FACTOR_LEVELS; ++i) {
+    int qdelta = av1_frame_type_qdelta(cpi, i, rc->worst_quality);
+    rc->rf_level_maxq[i] = AOMMAX(rc->worst_quality + qdelta, rc->best_quality);
+  }
+}
+
+void av1_init_subsampling(AV1_COMP *cpi) {
+  const AV1_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  const int w = cm->width;
+  const int h = cm->height;
+  int i;
+
+  for (i = 0; i < FRAME_SCALE_STEPS; ++i) {
+    // Note: Frames with odd-sized dimensions may result from this scaling.
+    rc->frame_width[i] = (w * 16) / frame_scale_factor[i];
+    rc->frame_height[i] = (h * 16) / frame_scale_factor[i];
+  }
+
+  setup_rf_level_maxq(cpi);
+}
+
+void av1_calculate_coded_size(AV1_COMP *cpi, int *scaled_frame_width,
+                              int *scaled_frame_height) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  *scaled_frame_width = rc->frame_width[rc->frame_size_selector];
+  *scaled_frame_height = rc->frame_height[rc->frame_size_selector];
+}
+
+void av1_init_second_pass(AV1_COMP *cpi) {
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  TWO_PASS *const twopass = &cpi->twopass;
+  double frame_rate;
+  FIRSTPASS_STATS *stats;
+
+  zero_stats(&twopass->total_stats);
+  zero_stats(&twopass->total_left_stats);
+
+  if (!twopass->stats_in_end) return;
+
+  stats = &twopass->total_stats;
+
+  *stats = *twopass->stats_in_end;
+  twopass->total_left_stats = *stats;
+
+  frame_rate = 10000000.0 * stats->count / stats->duration;
+  // Each frame can have a different duration, as the frame rate in the source
+  // isn't guaranteed to be constant. The frame rate prior to the first frame
+  // encoded in the second pass is a guess. However, the sum duration is not.
+  // It is calculated based on the actual durations of all frames from the
+  // first pass.
+  av1_new_framerate(cpi, frame_rate);
+  twopass->bits_left =
+      (int64_t)(stats->duration * oxcf->target_bandwidth / 10000000.0);
+
+  // This variable monitors how far behind the second ref update is lagging.
+  twopass->sr_update_lag = 1;
+
+  // Scan the first pass file and calculate a modified total error based upon
+  // the bias/power function used to allocate bits.
+  {
+    const double avg_error =
+        stats->coded_error / DOUBLE_DIVIDE_CHECK(stats->count);
+    const FIRSTPASS_STATS *s = twopass->stats_in;
+    double modified_error_total = 0.0;
+    twopass->modified_error_min =
+        (avg_error * oxcf->two_pass_vbrmin_section) / 100;
+    twopass->modified_error_max =
+        (avg_error * oxcf->two_pass_vbrmax_section) / 100;
+    while (s < twopass->stats_in_end) {
+      modified_error_total += calculate_modified_err(cpi, twopass, oxcf, s);
+      ++s;
+    }
+    twopass->modified_error_left = modified_error_total;
+  }
+
+  // Reset the vbr bits off target counters
+  cpi->rc.vbr_bits_off_target = 0;
+  cpi->rc.vbr_bits_off_target_fast = 0;
+
+  cpi->rc.rate_error_estimate = 0;
+
+  // Static sequence monitor variables.
+  twopass->kf_zeromotion_pct = 100;
+  twopass->last_kfgroup_zeromotion_pct = 100;
+
+  if (oxcf->resize_mode != RESIZE_NONE) {
+    av1_init_subsampling(cpi);
+  }
+}
+
+#define SR_DIFF_PART 0.0015
+#define MOTION_AMP_PART 0.003
+#define INTRA_PART 0.005
+#define DEFAULT_DECAY_LIMIT 0.75
+#define LOW_SR_DIFF_TRHESH 0.1
+#define SR_DIFF_MAX 128.0
+
+static double get_sr_decay_rate(const AV1_COMP *cpi,
+                                const FIRSTPASS_STATS *frame) {
+  const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE) ? cpi->initial_mbs
+                                                             : cpi->common.MBs;
+  double sr_diff = (frame->sr_coded_error - frame->coded_error) / num_mbs;
+  double sr_decay = 1.0;
+  double modified_pct_inter;
+  double modified_pcnt_intra;
+  const double motion_amplitude_factor =
+      frame->pcnt_motion * ((frame->mvc_abs + frame->mvr_abs) / 2);
+
+  modified_pct_inter = frame->pcnt_inter;
+  if ((frame->intra_error / DOUBLE_DIVIDE_CHECK(frame->coded_error)) <
+      (double)NCOUNT_FRAME_II_THRESH) {
+    modified_pct_inter = frame->pcnt_inter - frame->pcnt_neutral;
+  }
+  modified_pcnt_intra = 100 * (1.0 - modified_pct_inter);
+
+  if ((sr_diff > LOW_SR_DIFF_TRHESH)) {
+    sr_diff = AOMMIN(sr_diff, SR_DIFF_MAX);
+    sr_decay = 1.0 - (SR_DIFF_PART * sr_diff) -
+               (MOTION_AMP_PART * motion_amplitude_factor) -
+               (INTRA_PART * modified_pcnt_intra);
+  }
+  return AOMMAX(sr_decay, AOMMIN(DEFAULT_DECAY_LIMIT, modified_pct_inter));
+}
+
+// This function gives an estimate of how badly we believe the prediction
+// quality is decaying from frame to frame.
+static double get_zero_motion_factor(const AV1_COMP *cpi,
+                                     const FIRSTPASS_STATS *frame) {
+  const double zero_motion_pct = frame->pcnt_inter - frame->pcnt_motion;
+  double sr_decay = get_sr_decay_rate(cpi, frame);
+  return AOMMIN(sr_decay, zero_motion_pct);
+}
+
+#define ZM_POWER_FACTOR 0.75
+
+static double get_prediction_decay_rate(const AV1_COMP *cpi,
+                                        const FIRSTPASS_STATS *next_frame) {
+  const double sr_decay_rate = get_sr_decay_rate(cpi, next_frame);
+  const double zero_motion_factor =
+      (0.95 * pow((next_frame->pcnt_inter - next_frame->pcnt_motion),
+                  ZM_POWER_FACTOR));
+
+  return AOMMAX(zero_motion_factor,
+                (sr_decay_rate + ((1.0 - sr_decay_rate) * zero_motion_factor)));
+}
+
+// Function to test for a condition where a complex transition is followed
+// by a static section. For example in slide shows where there is a fade
+// between slides. This is to help with more optimal kf and gf positioning.
+static int detect_transition_to_still(AV1_COMP *cpi, int frame_interval,
+                                      int still_interval,
+                                      double loop_decay_rate,
+                                      double last_decay_rate) {
+  TWO_PASS *const twopass = &cpi->twopass;
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  // Break clause to detect very still sections after motion
+  // For example a static image after a fade or other transition
+  // instead of a clean scene cut.
+  if (frame_interval > rc->min_gf_interval && loop_decay_rate >= 0.999 &&
+      last_decay_rate < 0.9) {
+    int j;
+
+    // Look ahead a few frames to see if static condition persists...
+    for (j = 0; j < still_interval; ++j) {
+      const FIRSTPASS_STATS *stats = &twopass->stats_in[j];
+      if (stats >= twopass->stats_in_end) break;
+
+      if (stats->pcnt_inter - stats->pcnt_motion < 0.999) break;
+    }
+
+    // Only if it does do we signal a transition to still.
+    return j == still_interval;
+  }
+
+  return 0;
+}
+
+// This function detects a flash through the high relative pcnt_second_ref
+// score in the frame following a flash frame. The offset passed in should
+// reflect this.
+static int detect_flash(const TWO_PASS *twopass, int offset) {
+  const FIRSTPASS_STATS *const next_frame = read_frame_stats(twopass, offset);
+
+  // What we are looking for here is a situation where there is a
+  // brief break in prediction (such as a flash) but subsequent frames
+  // are reasonably well predicted by an earlier (pre flash) frame.
+  // The recovery after a flash is indicated by a high pcnt_second_ref
+  // compared to pcnt_inter.
+  return next_frame != NULL &&
+         next_frame->pcnt_second_ref > next_frame->pcnt_inter &&
+         next_frame->pcnt_second_ref >= 0.5;
+}
+
+// Update the motion related elements to the GF arf boost calculation.
+static void accumulate_frame_motion_stats(const FIRSTPASS_STATS *stats,
+                                          double *mv_in_out,
+                                          double *mv_in_out_accumulator,
+                                          double *abs_mv_in_out_accumulator,
+                                          double *mv_ratio_accumulator) {
+  const double pct = stats->pcnt_motion;
+
+  // Accumulate Motion In/Out of frame stats.
+  *mv_in_out = stats->mv_in_out_count * pct;
+  *mv_in_out_accumulator += *mv_in_out;
+  *abs_mv_in_out_accumulator += fabs(*mv_in_out);
+
+  // Accumulate a measure of how uniform (or conversely how random) the motion
+  // field is (a ratio of abs(mv) / mv).
+  if (pct > 0.05) {
+    const double mvr_ratio =
+        fabs(stats->mvr_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVr));
+    const double mvc_ratio =
+        fabs(stats->mvc_abs) / DOUBLE_DIVIDE_CHECK(fabs(stats->MVc));
+
+    *mv_ratio_accumulator +=
+        pct * (mvr_ratio < stats->mvr_abs ? mvr_ratio : stats->mvr_abs);
+    *mv_ratio_accumulator +=
+        pct * (mvc_ratio < stats->mvc_abs ? mvc_ratio : stats->mvc_abs);
+  }
+}
+
+#define BASELINE_ERR_PER_MB 1000.0
+static double calc_frame_boost(AV1_COMP *cpi, const FIRSTPASS_STATS *this_frame,
+                               double this_frame_mv_in_out, double max_boost) {
+  double frame_boost;
+  const double lq = av1_convert_qindex_to_q(
+      cpi->rc.avg_frame_qindex[INTER_FRAME], cpi->common.bit_depth);
+  const double boost_q_correction = AOMMIN((0.5 + (lq * 0.015)), 1.5);
+  int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE) ? cpi->initial_mbs
+                                                       : cpi->common.MBs;
+
+  // Correct for any inactive region in the image
+  num_mbs = (int)AOMMAX(1, num_mbs * calculate_active_area(cpi, this_frame));
+
+  // Underlying boost factor is based on inter error ratio.
+  frame_boost = (BASELINE_ERR_PER_MB * num_mbs) /
+                DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
+  frame_boost = frame_boost * BOOST_FACTOR * boost_q_correction;
+
+  // Increase boost for frames where new data coming into frame (e.g. zoom out).
+  // Slightly reduce boost if there is a net balance of motion out of the frame
+  // (zoom in). The range for this_frame_mv_in_out is -1.0 to +1.0.
+  if (this_frame_mv_in_out > 0.0)
+    frame_boost += frame_boost * (this_frame_mv_in_out * 2.0);
+  // In the extreme case the boost is halved.
+  else
+    frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
+
+  return AOMMIN(frame_boost, max_boost * boost_q_correction);
+}
+
+static int calc_arf_boost(AV1_COMP *cpi, int offset, int f_frames, int b_frames,
+                          int *f_boost, int *b_boost) {
+  TWO_PASS *const twopass = &cpi->twopass;
+  int i;
+  double boost_score = 0.0;
+  double mv_ratio_accumulator = 0.0;
+  double decay_accumulator = 1.0;
+  double this_frame_mv_in_out = 0.0;
+  double mv_in_out_accumulator = 0.0;
+  double abs_mv_in_out_accumulator = 0.0;
+  int arf_boost;
+  int flash_detected = 0;
+
+  // Search forward from the proposed arf/next gf position.
+  for (i = 0; i < f_frames; ++i) {
+    const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
+    if (this_frame == NULL) break;
+
+    // Update the motion related elements to the boost calculation.
+    accumulate_frame_motion_stats(
+        this_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
+        &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
+
+    // We want to discount the flash frame itself and the recovery
+    // frame that follows as both will have poor scores.
+    flash_detected = detect_flash(twopass, i + offset) ||
+                     detect_flash(twopass, i + offset + 1);
+
+    // Accumulate the effect of prediction quality decay.
+    if (!flash_detected) {
+      decay_accumulator *= get_prediction_decay_rate(cpi, this_frame);
+      decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
+                              ? MIN_DECAY_FACTOR
+                              : decay_accumulator;
+    }
+
+    boost_score +=
+        decay_accumulator *
+        calc_frame_boost(cpi, this_frame, this_frame_mv_in_out, GF_MAX_BOOST);
+  }
+
+  *f_boost = (int)boost_score;
+
+  // Reset for backward looking loop.
+  boost_score = 0.0;
+  mv_ratio_accumulator = 0.0;
+  decay_accumulator = 1.0;
+  this_frame_mv_in_out = 0.0;
+  mv_in_out_accumulator = 0.0;
+  abs_mv_in_out_accumulator = 0.0;
+
+  // Search backward towards last gf position.
+  for (i = -1; i >= -b_frames; --i) {
+    const FIRSTPASS_STATS *this_frame = read_frame_stats(twopass, i + offset);
+    if (this_frame == NULL) break;
+
+    // Update the motion related elements to the boost calculation.
+    accumulate_frame_motion_stats(
+        this_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
+        &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
+
+    // We want to discount the the flash frame itself and the recovery
+    // frame that follows as both will have poor scores.
+    flash_detected = detect_flash(twopass, i + offset) ||
+                     detect_flash(twopass, i + offset + 1);
+
+    // Cumulative effect of prediction quality decay.
+    if (!flash_detected) {
+      decay_accumulator *= get_prediction_decay_rate(cpi, this_frame);
+      decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
+                              ? MIN_DECAY_FACTOR
+                              : decay_accumulator;
+    }
+
+    boost_score +=
+        decay_accumulator *
+        calc_frame_boost(cpi, this_frame, this_frame_mv_in_out, GF_MAX_BOOST);
+  }
+  *b_boost = (int)boost_score;
+
+  arf_boost = (*f_boost + *b_boost);
+  if (arf_boost < ((b_frames + f_frames) * 20))
+    arf_boost = ((b_frames + f_frames) * 20);
+  arf_boost = AOMMAX(arf_boost, MIN_ARF_GF_BOOST);
+
+  return arf_boost;
+}
+
+// Calculate a section intra ratio used in setting max loop filter.
+static int calculate_section_intra_ratio(const FIRSTPASS_STATS *begin,
+                                         const FIRSTPASS_STATS *end,
+                                         int section_length) {
+  const FIRSTPASS_STATS *s = begin;
+  double intra_error = 0.0;
+  double coded_error = 0.0;
+  int i = 0;
+
+  while (s < end && i < section_length) {
+    intra_error += s->intra_error;
+    coded_error += s->coded_error;
+    ++s;
+    ++i;
+  }
+
+  return (int)(intra_error / DOUBLE_DIVIDE_CHECK(coded_error));
+}
+
+// Calculate the total bits to allocate in this GF/ARF group.
+static int64_t calculate_total_gf_group_bits(AV1_COMP *cpi,
+                                             double gf_group_err) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const TWO_PASS *const twopass = &cpi->twopass;
+  const int max_bits = frame_max_bits(rc, &cpi->oxcf);
+  int64_t total_group_bits;
+
+  // Calculate the bits to be allocated to the group as a whole.
+  if ((twopass->kf_group_bits > 0) && (twopass->kf_group_error_left > 0)) {
+    total_group_bits = (int64_t)(twopass->kf_group_bits *
+                                 (gf_group_err / twopass->kf_group_error_left));
+  } else {
+    total_group_bits = 0;
+  }
+
+  // Clamp odd edge cases.
+  total_group_bits = (total_group_bits < 0)
+                         ? 0
+                         : (total_group_bits > twopass->kf_group_bits)
+                               ? twopass->kf_group_bits
+                               : total_group_bits;
+
+  // Clip based on user supplied data rate variability limit.
+  if (total_group_bits > (int64_t)max_bits * rc->baseline_gf_interval)
+    total_group_bits = (int64_t)max_bits * rc->baseline_gf_interval;
+
+  return total_group_bits;
+}
+
+// Calculate the number bits extra to assign to boosted frames in a group.
+static int calculate_boost_bits(int frame_count, int boost,
+                                int64_t total_group_bits) {
+  int allocation_chunks;
+
+  // return 0 for invalid inputs (could arise e.g. through rounding errors)
+  if (!boost || (total_group_bits <= 0) || (frame_count <= 0)) return 0;
+
+  allocation_chunks = (frame_count * 100) + boost;
+
+  // Prevent overflow.
+  if (boost > 1023) {
+    int divisor = boost >> 10;
+    boost /= divisor;
+    allocation_chunks /= divisor;
+  }
+
+  // Calculate the number of extra bits for use in the boosted frame or frames.
+  return AOMMAX((int)(((int64_t)boost * total_group_bits) / allocation_chunks),
+                0);
+}
+
+#if !CONFIG_EXT_REFS
+// Current limit on maximum number of active arfs in a GF/ARF group.
+#define MAX_ACTIVE_ARFS 2
+#define ARF_SLOT1 2
+#define ARF_SLOT2 3
+// This function indirects the choice of buffers for arfs.
+// At the moment the values are fixed but this may change as part of
+// the integration process with other codec features that swap buffers around.
+static void get_arf_buffer_indices(unsigned char *arf_buffer_indices) {
+  arf_buffer_indices[0] = ARF_SLOT1;
+  arf_buffer_indices[1] = ARF_SLOT2;
+}
+#endif
+
+static void allocate_gf_group_bits(AV1_COMP *cpi, int64_t gf_group_bits,
+                                   double group_error, int gf_arf_bits) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  TWO_PASS *const twopass = &cpi->twopass;
+  GF_GROUP *const gf_group = &twopass->gf_group;
+  FIRSTPASS_STATS frame_stats;
+  int i;
+  int frame_index = 0;
+  int target_frame_size;
+  int key_frame;
+  const int max_bits = frame_max_bits(&cpi->rc, &cpi->oxcf);
+  int64_t total_group_bits = gf_group_bits;
+  double modified_err = 0.0;
+  double err_fraction;
+  int mid_boost_bits = 0;
+#if CONFIG_EXT_REFS
+  // The use of bi-predictive frames are only enabled when following 3
+  // conditions are met:
+  // (1) Alt-ref is enabled;
+  // (2) The bi-predictive group interval is at least 2; and
+  // (3) The bi-predictive group interval is strictly smaller than the
+  //     golden group interval.
+  const int is_bipred_enabled =
+      rc->source_alt_ref_pending && rc->bipred_group_interval &&
+      rc->bipred_group_interval <=
+          (rc->baseline_gf_interval - rc->source_alt_ref_pending);
+  int bipred_group_end = 0;
+  int bipred_frame_index = 0;
+
+  int arf_pos[MAX_EXT_ARFS + 1];
+  const unsigned char ext_arf_interval =
+      (unsigned char)(rc->baseline_gf_interval / (cpi->num_extra_arfs + 1) - 1);
+  int which_arf = cpi->num_extra_arfs;
+  int subgroup_interval[MAX_EXT_ARFS + 1];
+  int ext_arf_boost[MAX_EXT_ARFS];
+  int is_sg_bipred_enabled = is_bipred_enabled;
+  int accumulative_subgroup_interval = 0;
+#else
+  int mid_frame_idx;
+  unsigned char arf_buffer_indices[MAX_ACTIVE_ARFS];
+#endif  // CONFIG_EXT_REFS
+
+#if CONFIG_EXT_REFS
+  av1_zero_array(ext_arf_boost, MAX_EXT_ARFS);
+#endif  // CONFIG_EXT_REFS
+
+  key_frame = cpi->common.frame_type == KEY_FRAME;
+
+#if !CONFIG_EXT_REFS
+  get_arf_buffer_indices(arf_buffer_indices);
+#endif  // !CONFIG_EXT_REFS
+
+  // For key frames the frame target rate is already set and it
+  // is also the golden frame.
+  if (!key_frame) {
+    if (rc->source_alt_ref_active) {
+      gf_group->update_type[frame_index] = OVERLAY_UPDATE;
+      gf_group->rf_level[frame_index] = INTER_NORMAL;
+      gf_group->bit_allocation[frame_index] = 0;
+    } else {
+      gf_group->update_type[frame_index] = GF_UPDATE;
+      gf_group->rf_level[frame_index] = GF_ARF_STD;
+      gf_group->bit_allocation[frame_index] = gf_arf_bits;
+    }
+#if CONFIG_EXT_REFS
+    gf_group->arf_update_idx[frame_index] = 0;
+    gf_group->arf_ref_idx[frame_index] = 0;
+#else
+    gf_group->arf_update_idx[frame_index] = arf_buffer_indices[0];
+    gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[0];
+#endif  // CONFIG_EXT_REFS
+    // Step over the golden frame / overlay frame
+    if (EOF == input_stats(twopass, &frame_stats)) return;
+  }
+
+#if CONFIG_EXT_REFS
+  gf_group->bidir_pred_enabled[frame_index] = 0;
+  gf_group->brf_src_offset[frame_index] = 0;
+#endif  // CONFIG_EXT_REFS
+
+  // Deduct the boost bits for arf (or gf if it is not a key frame)
+  // from the group total.
+  if (rc->source_alt_ref_pending || !key_frame) total_group_bits -= gf_arf_bits;
+
+  frame_index++;
+
+#if CONFIG_EXT_REFS
+  bipred_frame_index++;
+#endif  // CONFIG_EXT_REFS
+
+  // Store the bits to spend on the ARF if there is one.
+  if (rc->source_alt_ref_pending) {
+    gf_group->update_type[frame_index] = ARF_UPDATE;
+    gf_group->rf_level[frame_index] = GF_ARF_STD;
+    gf_group->bit_allocation[frame_index] = gf_arf_bits;
+
+    gf_group->arf_src_offset[frame_index] =
+        (unsigned char)(rc->baseline_gf_interval - 1);
+
+#if CONFIG_EXT_REFS
+    gf_group->arf_update_idx[frame_index] = 0;
+    gf_group->arf_ref_idx[frame_index] = 0;
+
+    gf_group->bidir_pred_enabled[frame_index] = 0;
+    gf_group->brf_src_offset[frame_index] = 0;
+// NOTE: "bidir_pred_frame_index" stays unchanged for ARF_UPDATE frames.
+#else
+    gf_group->arf_update_idx[frame_index] = arf_buffer_indices[0];
+    gf_group->arf_ref_idx[frame_index] =
+        arf_buffer_indices[cpi->multi_arf_last_grp_enabled &&
+                           rc->source_alt_ref_active];
+#endif  // CONFIG_EXT_REFS
+
+#if CONFIG_EXT_REFS
+    // Work out the ARFs' positions in this gf group
+    // NOTE(weitinglin): ALT_REFs' are indexed inversely, but coded in display
+    // order (except for the original ARF). In the example of three ALT_REF's,
+    // We index ALTREF's as: KEY ----- ALT2 ----- ALT1 ----- ALT0
+    // but code them in the following order:
+    // KEY-ALT0-ALT2 ----- OVERLAY2-ALT1 ----- OVERLAY1 ----- OVERLAY0
+    arf_pos[0] =
+        frame_index + cpi->num_extra_arfs + gf_group->arf_src_offset[1] + 1;
+    for (i = 0; i < cpi->num_extra_arfs; ++i) {
+      arf_pos[i + 1] =
+          frame_index + (cpi->num_extra_arfs - i) * (ext_arf_interval + 2);
+      subgroup_interval[i] = arf_pos[i] - arf_pos[i + 1] - (i == 0 ? 1 : 2);
+    }
+    subgroup_interval[cpi->num_extra_arfs] = arf_pos[cpi->num_extra_arfs] -
+                                             frame_index -
+                                             (cpi->num_extra_arfs == 0 ? 1 : 2);
+#endif  // CONFIG_EXT_REFS
+
+    ++frame_index;
+
+#if CONFIG_EXT_REFS
+    // Insert an extra ARF
+    if (cpi->num_extra_arfs) {
+      gf_group->update_type[frame_index] = ARF_UPDATE;
+      // Note (weitinglin): GF_ARF_LOW is also used as an identifier
+      //                    for internal ALT_REF's:
+      gf_group->rf_level[frame_index] = GF_ARF_LOW;
+      gf_group->arf_src_offset[frame_index] = ext_arf_interval;
+      gf_group->arf_update_idx[frame_index] = which_arf;
+      gf_group->arf_ref_idx[frame_index] = 0;
+      ++frame_index;
+    }
+    accumulative_subgroup_interval += subgroup_interval[cpi->num_extra_arfs];
+#else
+    if (cpi->multi_arf_enabled) {
+      // Set aside a slot for a level 1 arf.
+      gf_group->update_type[frame_index] = ARF_UPDATE;
+      gf_group->rf_level[frame_index] = GF_ARF_LOW;
+      gf_group->arf_src_offset[frame_index] =
+          (unsigned char)((rc->baseline_gf_interval >> 1) - 1);
+      gf_group->arf_update_idx[frame_index] = arf_buffer_indices[1];
+      gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[0];
+      ++frame_index;
+    }
+#endif  // CONFIG_EXT_ARFS
+  }
+
+#if !CONFIG_EXT_REFS
+  // Define middle frame
+  mid_frame_idx = frame_index + (rc->baseline_gf_interval >> 1) - 1;
+#endif  // !CONFIG_EXT_REFS
+
+  // Allocate bits to the other frames in the group.
+  for (i = 0; i < rc->baseline_gf_interval - rc->source_alt_ref_pending; ++i) {
+#if !CONFIG_EXT_REFS
+    int arf_idx = 0;
+#endif  // !CONFIG_EXT_REFS
+
+    if (EOF == input_stats(twopass, &frame_stats)) break;
+
+    modified_err = calculate_modified_err(cpi, twopass, oxcf, &frame_stats);
+
+    if (group_error > 0)
+      err_fraction = modified_err / DOUBLE_DIVIDE_CHECK(group_error);
+    else
+      err_fraction = 0.0;
+
+    target_frame_size = (int)((double)total_group_bits * err_fraction);
+
+    if (rc->source_alt_ref_pending && cpi->multi_arf_enabled) {
+      mid_boost_bits += (target_frame_size >> 4);
+      target_frame_size -= (target_frame_size >> 4);
+#if !CONFIG_EXT_REFS
+      if (frame_index <= mid_frame_idx) arf_idx = 1;
+#endif  // !CONFIG_EXT_REFS
+    }
+
+#if CONFIG_EXT_REFS
+    gf_group->arf_update_idx[frame_index] = which_arf;
+    gf_group->arf_ref_idx[frame_index] = which_arf;
+#else
+    gf_group->arf_update_idx[frame_index] = arf_buffer_indices[arf_idx];
+    gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[arf_idx];
+#endif  // CONFIG_EXT_REFS
+
+    target_frame_size =
+        clamp(target_frame_size, 0, AOMMIN(max_bits, (int)total_group_bits));
+
+#if CONFIG_EXT_REFS
+    // If we are going to have ARFs, check if we can have BWDREF in this
+    // subgroup.
+    if (rc->source_alt_ref_pending) {
+      is_sg_bipred_enabled =
+          is_bipred_enabled &&
+          (subgroup_interval[which_arf] > rc->bipred_group_interval);
+    }
+
+    // NOTE: BIDIR_PRED is only enabled when the length of the bi-predictive
+    //       frame group interval is strictly smaller than that of the GOLDEN
+    //       FRAME group interval.
+    // TODO(zoeliu): Currently BIDIR_PRED is only enabled when alt-ref is on.
+    if (is_sg_bipred_enabled && !bipred_group_end) {
+      const int cur_brf_src_offset = rc->bipred_group_interval - 1;
+
+      // --- BRF_UPDATE ---
+      if (bipred_frame_index == 1) {
+        gf_group->update_type[frame_index] = BRF_UPDATE;
+        gf_group->bidir_pred_enabled[frame_index] = 1;
+        gf_group->brf_src_offset[frame_index] = cur_brf_src_offset;
+        // --- LAST_BIPRED_UPDATE ---
+      } else if (bipred_frame_index == rc->bipred_group_interval) {
+        gf_group->update_type[frame_index] = LAST_BIPRED_UPDATE;
+        gf_group->bidir_pred_enabled[frame_index] = 1;
+        gf_group->brf_src_offset[frame_index] = 0;
+        // Reset the bi-predictive frame index.
+        bipred_frame_index = 0;
+        // --- BIPRED_UPDATE ---
+      } else {
+        gf_group->update_type[frame_index] = BIPRED_UPDATE;
+        gf_group->bidir_pred_enabled[frame_index] = 1;
+        gf_group->brf_src_offset[frame_index] = 0;
+      }
+
+      bipred_frame_index++;
+      // Check whether the next bi-predictive frame group would entirely be
+      // included within the current golden frame group.
+      // In addition, we need to avoid coding a BRF right before an ARF.
+      if (bipred_frame_index == 1 &&
+          (i + 2 + cur_brf_src_offset) >= accumulative_subgroup_interval) {
+        bipred_group_end = 1;
+      }
+    } else {
+#endif  // CONFIG_EXT_REFS
+      gf_group->update_type[frame_index] = LF_UPDATE;
+#if CONFIG_EXT_REFS
+      gf_group->bidir_pred_enabled[frame_index] = 0;
+      gf_group->brf_src_offset[frame_index] = 0;
+    }
+#endif  // CONFIG_EXT_REFS
+
+#if CONFIG_EXT_REFS
+    if (gf_group->update_type[frame_index] == BRF_UPDATE) {
+      // Boost up the allocated bits on BWDREF_FRAME
+      gf_group->rf_level[frame_index] = GF_ARF_LOW;
+      gf_group->bit_allocation[frame_index] =
+          target_frame_size + (target_frame_size >> 2);
+    } else if (gf_group->update_type[frame_index] == LAST_BIPRED_UPDATE) {
+      // Press down the allocated bits on LAST_BIPRED_UPDATE frames
+      gf_group->rf_level[frame_index] = INTER_NORMAL;
+      gf_group->bit_allocation[frame_index] =
+          target_frame_size - (target_frame_size >> 1);
+    } else if (gf_group->update_type[frame_index] == BIPRED_UPDATE) {
+      // TODO(zoeliu): To investigate whether the allocated bits on
+      // BIPRED_UPDATE frames need to be further adjusted.
+      gf_group->rf_level[frame_index] = INTER_NORMAL;
+      gf_group->bit_allocation[frame_index] = target_frame_size;
+    } else {
+#endif  // CONFIG_EXT_REFS
+      gf_group->rf_level[frame_index] = INTER_NORMAL;
+      gf_group->bit_allocation[frame_index] = target_frame_size;
+#if CONFIG_EXT_REFS
+    }
+#endif  // CONFIG_EXT_REFS
+
+    ++frame_index;
+
+#if CONFIG_EXT_REFS
+    // Check if we need to update the ARF
+    if (is_sg_bipred_enabled && cpi->num_extra_arfs && which_arf > 0 &&
+        frame_index > arf_pos[which_arf]) {
+      --which_arf;
+      accumulative_subgroup_interval += subgroup_interval[which_arf] + 1;
+      // Meet the new subgroup. Reset the bipred_group_end flag;
+      bipred_group_end = 0;
+      // Insert another extra ARF after the overlay frame
+      if (which_arf) {
+        gf_group->update_type[frame_index] = ARF_UPDATE;
+        gf_group->rf_level[frame_index] = GF_ARF_LOW;
+        gf_group->arf_src_offset[frame_index] = ext_arf_interval;
+        gf_group->arf_update_idx[frame_index] = which_arf;
+        gf_group->arf_ref_idx[frame_index] = 0;
+        ++frame_index;
+      }
+    }
+#endif  // CONFIG_EXT_REFS
+  }
+
+// Note:
+// We need to configure the frame at the end of the sequence + 1 that will be
+// the start frame for the next group. Otherwise prior to the call to
+// av1_rc_get_second_pass_params() the data will be undefined.
+#if CONFIG_EXT_REFS
+  gf_group->arf_update_idx[frame_index] = 0;
+  gf_group->arf_ref_idx[frame_index] = 0;
+#else
+  gf_group->arf_update_idx[frame_index] = arf_buffer_indices[0];
+  gf_group->arf_ref_idx[frame_index] = arf_buffer_indices[0];
+#endif  // CONFIG_EXT_REFS
+
+  if (rc->source_alt_ref_pending) {
+    gf_group->update_type[frame_index] = OVERLAY_UPDATE;
+    gf_group->rf_level[frame_index] = INTER_NORMAL;
+
+#if CONFIG_EXT_REFS
+    if (cpi->num_extra_arfs) {
+      for (i = cpi->num_extra_arfs; i > 0; --i) {
+        int arf_pos_in_gf = (i == cpi->num_extra_arfs ? 2 : arf_pos[i + 1] + 1);
+        gf_group->bit_allocation[arf_pos_in_gf] =
+            gf_group->bit_allocation[arf_pos[i]];
+        gf_group->update_type[arf_pos[i]] = INTNL_OVERLAY_UPDATE;
+        gf_group->bit_allocation[arf_pos[i]] = 0;
+        gf_group->rf_level[arf_pos[i]] = INTER_NORMAL;
+      }
+    }
+#else
+    // Final setup for second arf and its overlay.
+    if (cpi->multi_arf_enabled) {
+      gf_group->bit_allocation[2] =
+          gf_group->bit_allocation[mid_frame_idx] + mid_boost_bits;
+      gf_group->update_type[mid_frame_idx] = OVERLAY_UPDATE;
+      gf_group->bit_allocation[mid_frame_idx] = 0;
+    }
+#endif  // CONFIG_EXT_REFS
+  } else {
+    gf_group->update_type[frame_index] = GF_UPDATE;
+    gf_group->rf_level[frame_index] = GF_ARF_STD;
+  }
+
+#if CONFIG_EXT_REFS
+  gf_group->bidir_pred_enabled[frame_index] = 0;
+  gf_group->brf_src_offset[frame_index] = 0;
+#endif  // CONFIG_EXT_REFS
+
+  // Note whether multi-arf was enabled this group for next time.
+  cpi->multi_arf_last_grp_enabled = cpi->multi_arf_enabled;
+}
+
+// Analyse and define a gf/arf group.
+static void define_gf_group(AV1_COMP *cpi, FIRSTPASS_STATS *this_frame) {
+  AV1_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  TWO_PASS *const twopass = &cpi->twopass;
+  FIRSTPASS_STATS next_frame;
+  const FIRSTPASS_STATS *const start_pos = twopass->stats_in;
+  int i;
+
+  double boost_score = 0.0;
+  double old_boost_score = 0.0;
+  double gf_group_err = 0.0;
+#if GROUP_ADAPTIVE_MAXQ
+  double gf_group_raw_error = 0.0;
+#endif
+  double gf_group_skip_pct = 0.0;
+  double gf_group_inactive_zone_rows = 0.0;
+  double gf_first_frame_err = 0.0;
+  double mod_frame_err = 0.0;
+
+  double mv_ratio_accumulator = 0.0;
+  double decay_accumulator = 1.0;
+  double zero_motion_accumulator = 1.0;
+
+  double loop_decay_rate = 1.00;
+  double last_loop_decay_rate = 1.00;
+
+  double this_frame_mv_in_out = 0.0;
+  double mv_in_out_accumulator = 0.0;
+  double abs_mv_in_out_accumulator = 0.0;
+  double mv_ratio_accumulator_thresh;
+  unsigned int allow_alt_ref = is_altref_enabled(cpi);
+
+  int f_boost = 0;
+  int b_boost = 0;
+  int flash_detected;
+  int active_max_gf_interval;
+  int active_min_gf_interval;
+  int64_t gf_group_bits;
+  double gf_group_error_left;
+  int gf_arf_bits;
+  const int is_key_frame = frame_is_intra_only(cm);
+  const int arf_active_or_kf = is_key_frame || rc->source_alt_ref_active;
+
+  // Reset the GF group data structures unless this is a key
+  // frame in which case it will already have been done.
+  if (is_key_frame == 0) {
+    av1_zero(twopass->gf_group);
+  }
+
+  aom_clear_system_state();
+  av1_zero(next_frame);
+
+  // Load stats for the current frame.
+  mod_frame_err = calculate_modified_err(cpi, twopass, oxcf, this_frame);
+
+  // Note the error of the frame at the start of the group. This will be
+  // the GF frame error if we code a normal gf.
+  gf_first_frame_err = mod_frame_err;
+
+  // If this is a key frame or the overlay from a previous arf then
+  // the error score / cost of this frame has already been accounted for.
+  if (arf_active_or_kf) {
+    gf_group_err -= gf_first_frame_err;
+#if GROUP_ADAPTIVE_MAXQ
+    gf_group_raw_error -= this_frame->coded_error;
+#endif
+    gf_group_skip_pct -= this_frame->intra_skip_pct;
+    gf_group_inactive_zone_rows -= this_frame->inactive_zone_rows;
+  }
+
+  // Motion breakout threshold for loop below depends on image size.
+  mv_ratio_accumulator_thresh =
+      (cpi->initial_height + cpi->initial_width) / 4.0;
+
+  // Set a maximum and minimum interval for the GF group.
+  // If the image appears almost completely static we can extend beyond this.
+  {
+    int int_max_q = (int)(av1_convert_qindex_to_q(twopass->active_worst_quality,
+                                                  cpi->common.bit_depth));
+    int int_lbq = (int)(av1_convert_qindex_to_q(rc->last_boosted_qindex,
+                                                cpi->common.bit_depth));
+
+    active_min_gf_interval = rc->min_gf_interval + AOMMIN(2, int_max_q / 200);
+    if (active_min_gf_interval > rc->max_gf_interval)
+      active_min_gf_interval = rc->max_gf_interval;
+
+    if (cpi->multi_arf_allowed) {
+      active_max_gf_interval = rc->max_gf_interval;
+    } else {
+      // The value chosen depends on the active Q range. At low Q we have
+      // bits to spare and are better with a smaller interval and smaller boost.
+      // At high Q when there are few bits to spare we are better with a longer
+      // interval to spread the cost of the GF.
+      active_max_gf_interval = 12 + AOMMIN(4, (int_lbq / 6));
+
+      // We have: active_min_gf_interval <= rc->max_gf_interval
+      if (active_max_gf_interval < active_min_gf_interval)
+        active_max_gf_interval = active_min_gf_interval;
+      else if (active_max_gf_interval > rc->max_gf_interval)
+        active_max_gf_interval = rc->max_gf_interval;
+    }
+  }
+
+  i = 0;
+  while (i < rc->static_scene_max_gf_interval && i < rc->frames_to_key) {
+    ++i;
+
+    // Accumulate error score of frames in this gf group.
+    mod_frame_err = calculate_modified_err(cpi, twopass, oxcf, this_frame);
+    gf_group_err += mod_frame_err;
+#if GROUP_ADAPTIVE_MAXQ
+    gf_group_raw_error += this_frame->coded_error;
+#endif
+    gf_group_skip_pct += this_frame->intra_skip_pct;
+    gf_group_inactive_zone_rows += this_frame->inactive_zone_rows;
+
+    if (EOF == input_stats(twopass, &next_frame)) break;
+
+    // Test for the case where there is a brief flash but the prediction
+    // quality back to an earlier frame is then restored.
+    flash_detected = detect_flash(twopass, 0);
+
+    // Update the motion related elements to the boost calculation.
+    accumulate_frame_motion_stats(
+        &next_frame, &this_frame_mv_in_out, &mv_in_out_accumulator,
+        &abs_mv_in_out_accumulator, &mv_ratio_accumulator);
+
+    // Accumulate the effect of prediction quality decay.
+    if (!flash_detected) {
+      last_loop_decay_rate = loop_decay_rate;
+      loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
+
+      decay_accumulator = decay_accumulator * loop_decay_rate;
+
+      // Monitor for static sections.
+      zero_motion_accumulator = AOMMIN(
+          zero_motion_accumulator, get_zero_motion_factor(cpi, &next_frame));
+
+      // Break clause to detect very still sections after motion. For example,
+      // a static image after a fade or other transition.
+      if (detect_transition_to_still(cpi, i, 5, loop_decay_rate,
+                                     last_loop_decay_rate)) {
+        allow_alt_ref = 0;
+        break;
+      }
+    }
+
+    // Calculate a boost number for this frame.
+    boost_score +=
+        decay_accumulator *
+        calc_frame_boost(cpi, &next_frame, this_frame_mv_in_out, GF_MAX_BOOST);
+
+    // Break out conditions.
+    if (
+        // Break at active_max_gf_interval unless almost totally static.
+        (i >= (active_max_gf_interval + arf_active_or_kf) &&
+         zero_motion_accumulator < 0.995) ||
+        (
+            // Don't break out with a very short interval.
+            (i >= active_min_gf_interval + arf_active_or_kf) &&
+            (!flash_detected) &&
+            ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
+             (abs_mv_in_out_accumulator > 3.0) ||
+             (mv_in_out_accumulator < -2.0) ||
+             ((boost_score - old_boost_score) < BOOST_BREAKOUT)))) {
+      boost_score = old_boost_score;
+      break;
+    }
+
+    *this_frame = next_frame;
+    old_boost_score = boost_score;
+  }
+
+  twopass->gf_zeromotion_pct = (int)(zero_motion_accumulator * 1000.0);
+
+  // Was the group length constrained by the requirement for a new KF?
+  rc->constrained_gf_group = (i >= rc->frames_to_key) ? 1 : 0;
+
+  // Should we use the alternate reference frame.
+  if (allow_alt_ref && (i < cpi->oxcf.lag_in_frames) &&
+      (i >= rc->min_gf_interval)) {
+    // Calculate the boost for alt ref.
+    rc->gfu_boost =
+        calc_arf_boost(cpi, 0, (i - 1), (i - 1), &f_boost, &b_boost);
+    rc->source_alt_ref_pending = 1;
+
+    // Test to see if multi arf is appropriate.
+    cpi->multi_arf_enabled =
+        (cpi->multi_arf_allowed && (rc->baseline_gf_interval >= 6) &&
+         (zero_motion_accumulator < 0.995))
+            ? 1
+            : 0;
+  } else {
+    rc->gfu_boost = AOMMAX((int)boost_score, MIN_ARF_GF_BOOST);
+    rc->source_alt_ref_pending = 0;
+  }
+
+  // Set the interval until the next gf.
+  rc->baseline_gf_interval = i - (is_key_frame || rc->source_alt_ref_pending);
+
+#if CONFIG_EXT_REFS
+  // Compute how many extra alt_refs we can have
+  cpi->num_extra_arfs = get_number_of_extra_arfs(rc->baseline_gf_interval,
+                                                 rc->source_alt_ref_pending);
+  // Currently at maximum two extra ARFs' are allowed
+  assert(cpi->num_extra_arfs <= MAX_EXT_ARFS);
+#endif  // CONFIG_EXT_REFS
+
+  rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+
+#if CONFIG_EXT_REFS
+  rc->bipred_group_interval = BFG_INTERVAL;
+  // The minimum bi-predictive frame group interval is 2.
+  if (rc->bipred_group_interval < 2) rc->bipred_group_interval = 0;
+#endif  // CONFIG_EXT_REFS
+
+  // Reset the file position.
+  reset_fpf_position(twopass, start_pos);
+
+  // Calculate the bits to be allocated to the gf/arf group as a whole
+  gf_group_bits = calculate_total_gf_group_bits(cpi, gf_group_err);
+
+#if GROUP_ADAPTIVE_MAXQ
+  // Calculate an estimate of the maxq needed for the group.
+  // We are more agressive about correcting for sections
+  // where there could be significant overshoot than for easier
+  // sections where we do not wish to risk creating an overshoot
+  // of the allocated bit budget.
+  if ((cpi->oxcf.rc_mode != AOM_Q) && (rc->baseline_gf_interval > 1)) {
+    const int vbr_group_bits_per_frame =
+        (int)(gf_group_bits / rc->baseline_gf_interval);
+    const double group_av_err = gf_group_raw_error / rc->baseline_gf_interval;
+    const double group_av_skip_pct =
+        gf_group_skip_pct / rc->baseline_gf_interval;
+    const double group_av_inactive_zone =
+        ((gf_group_inactive_zone_rows * 2) /
+         (rc->baseline_gf_interval * (double)cm->mb_rows));
+
+    int tmp_q;
+    // rc factor is a weight factor that corrects for local rate control drift.
+    double rc_factor = 1.0;
+    if (rc->rate_error_estimate > 0) {
+      rc_factor = AOMMAX(RC_FACTOR_MIN,
+                         (double)(100 - rc->rate_error_estimate) / 100.0);
+    } else {
+      rc_factor = AOMMIN(RC_FACTOR_MAX,
+                         (double)(100 - rc->rate_error_estimate) / 100.0);
+    }
+    tmp_q = get_twopass_worst_quality(
+        cpi, group_av_err, (group_av_skip_pct + group_av_inactive_zone),
+        vbr_group_bits_per_frame, twopass->kfgroup_inter_fraction * rc_factor);
+    twopass->active_worst_quality =
+        AOMMAX(tmp_q, twopass->active_worst_quality >> 1);
+  }
+#endif
+
+  // Calculate the extra bits to be used for boosted frame(s)
+  gf_arf_bits = calculate_boost_bits(rc->baseline_gf_interval, rc->gfu_boost,
+                                     gf_group_bits);
+
+  // Adjust KF group bits and error remaining.
+  twopass->kf_group_error_left -= (int64_t)gf_group_err;
+
+  // If this is an arf update we want to remove the score for the overlay
+  // frame at the end which will usually be very cheap to code.
+  // The overlay frame has already, in effect, been coded so we want to spread
+  // the remaining bits among the other frames.
+  // For normal GFs remove the score for the GF itself unless this is
+  // also a key frame in which case it has already been accounted for.
+  if (rc->source_alt_ref_pending) {
+    gf_group_error_left = gf_group_err - mod_frame_err;
+  } else if (is_key_frame == 0) {
+    gf_group_error_left = gf_group_err - gf_first_frame_err;
+  } else {
+    gf_group_error_left = gf_group_err;
+  }
+
+  // Allocate bits to each of the frames in the GF group.
+  allocate_gf_group_bits(cpi, gf_group_bits, gf_group_error_left, gf_arf_bits);
+
+  // Reset the file position.
+  reset_fpf_position(twopass, start_pos);
+
+  // Calculate a section intra ratio used in setting max loop filter.
+  if (cpi->common.frame_type != KEY_FRAME) {
+    twopass->section_intra_rating = calculate_section_intra_ratio(
+        start_pos, twopass->stats_in_end, rc->baseline_gf_interval);
+  }
+
+  if (oxcf->resize_mode == RESIZE_DYNAMIC) {
+    // Default to starting GF groups at normal frame size.
+    cpi->rc.next_frame_size_selector = UNSCALED;
+  }
+}
+
+// Threshold for use of the lagging second reference frame. High second ref
+// usage may point to a transient event like a flash or occlusion rather than
+// a real scene cut.
+#define SECOND_REF_USEAGE_THRESH 0.1
+// Minimum % intra coding observed in first pass (1.0 = 100%)
+#define MIN_INTRA_LEVEL 0.25
+// Minimum ratio between the % of intra coding and inter coding in the first
+// pass after discounting neutral blocks (discounting neutral blocks in this
+// way helps catch scene cuts in clips with very flat areas or letter box
+// format clips with image padding.
+#define INTRA_VS_INTER_THRESH 2.0
+// Hard threshold where the first pass chooses intra for almost all blocks.
+// In such a case even if the frame is not a scene cut coding a key frame
+// may be a good option.
+#define VERY_LOW_INTER_THRESH 0.05
+// Maximum threshold for the relative ratio of intra error score vs best
+// inter error score.
+#define KF_II_ERR_THRESHOLD 2.5
+// In real scene cuts there is almost always a sharp change in the intra
+// or inter error score.
+#define ERR_CHANGE_THRESHOLD 0.4
+// For real scene cuts we expect an improvment in the intra inter error
+// ratio in the next frame.
+#define II_IMPROVEMENT_THRESHOLD 3.5
+#define KF_II_MAX 128.0
+
+static int test_candidate_kf(TWO_PASS *twopass,
+                             const FIRSTPASS_STATS *last_frame,
+                             const FIRSTPASS_STATS *this_frame,
+                             const FIRSTPASS_STATS *next_frame) {
+  int is_viable_kf = 0;
+  double pcnt_intra = 1.0 - this_frame->pcnt_inter;
+  double modified_pcnt_inter =
+      this_frame->pcnt_inter - this_frame->pcnt_neutral;
+
+  // Does the frame satisfy the primary criteria of a key frame?
+  // See above for an explanation of the test criteria.
+  // If so, then examine how well it predicts subsequent frames.
+  if ((this_frame->pcnt_second_ref < SECOND_REF_USEAGE_THRESH) &&
+      (next_frame->pcnt_second_ref < SECOND_REF_USEAGE_THRESH) &&
+      ((this_frame->pcnt_inter < VERY_LOW_INTER_THRESH) ||
+       ((pcnt_intra > MIN_INTRA_LEVEL) &&
+        (pcnt_intra > (INTRA_VS_INTER_THRESH * modified_pcnt_inter)) &&
+        ((this_frame->intra_error /
+          DOUBLE_DIVIDE_CHECK(this_frame->coded_error)) <
+         KF_II_ERR_THRESHOLD) &&
+        ((fabs(last_frame->coded_error - this_frame->coded_error) /
+              DOUBLE_DIVIDE_CHECK(this_frame->coded_error) >
+          ERR_CHANGE_THRESHOLD) ||
+         (fabs(last_frame->intra_error - this_frame->intra_error) /
+              DOUBLE_DIVIDE_CHECK(this_frame->intra_error) >
+          ERR_CHANGE_THRESHOLD) ||
+         ((next_frame->intra_error /
+           DOUBLE_DIVIDE_CHECK(next_frame->coded_error)) >
+          II_IMPROVEMENT_THRESHOLD))))) {
+    int i;
+    const FIRSTPASS_STATS *start_pos = twopass->stats_in;
+    FIRSTPASS_STATS local_next_frame = *next_frame;
+    double boost_score = 0.0;
+    double old_boost_score = 0.0;
+    double decay_accumulator = 1.0;
+
+    // Examine how well the key frame predicts subsequent frames.
+    for (i = 0; i < 16; ++i) {
+      double next_iiratio = (BOOST_FACTOR * local_next_frame.intra_error /
+                             DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
+
+      if (next_iiratio > KF_II_MAX) next_iiratio = KF_II_MAX;
+
+      // Cumulative effect of decay in prediction quality.
+      if (local_next_frame.pcnt_inter > 0.85)
+        decay_accumulator *= local_next_frame.pcnt_inter;
+      else
+        decay_accumulator *= (0.85 + local_next_frame.pcnt_inter) / 2.0;
+
+      // Keep a running total.
+      boost_score += (decay_accumulator * next_iiratio);
+
+      // Test various breakout clauses.
+      if ((local_next_frame.pcnt_inter < 0.05) || (next_iiratio < 1.5) ||
+          (((local_next_frame.pcnt_inter - local_next_frame.pcnt_neutral) <
+            0.20) &&
+           (next_iiratio < 3.0)) ||
+          ((boost_score - old_boost_score) < 3.0) ||
+          (local_next_frame.intra_error < 200)) {
+        break;
+      }
+
+      old_boost_score = boost_score;
+
+      // Get the next frame details
+      if (EOF == input_stats(twopass, &local_next_frame)) break;
+    }
+
+    // If there is tolerable prediction for at least the next 3 frames then
+    // break out else discard this potential key frame and move on
+    if (boost_score > 30.0 && (i > 3)) {
+      is_viable_kf = 1;
+    } else {
+      // Reset the file position
+      reset_fpf_position(twopass, start_pos);
+
+      is_viable_kf = 0;
+    }
+  }
+
+  return is_viable_kf;
+}
+
+#define FRAMES_TO_CHECK_DECAY 8
+
+static void find_next_key_frame(AV1_COMP *cpi, FIRSTPASS_STATS *this_frame) {
+  int i, j;
+  RATE_CONTROL *const rc = &cpi->rc;
+  TWO_PASS *const twopass = &cpi->twopass;
+  GF_GROUP *const gf_group = &twopass->gf_group;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  const FIRSTPASS_STATS first_frame = *this_frame;
+  const FIRSTPASS_STATS *const start_position = twopass->stats_in;
+  FIRSTPASS_STATS next_frame;
+  FIRSTPASS_STATS last_frame;
+  int kf_bits = 0;
+  int loop_decay_counter = 0;
+  double decay_accumulator = 1.0;
+  double av_decay_accumulator = 0.0;
+  double zero_motion_accumulator = 1.0;
+  double boost_score = 0.0;
+  double kf_mod_err = 0.0;
+  double kf_group_err = 0.0;
+  double recent_loop_decay[FRAMES_TO_CHECK_DECAY];
+
+  av1_zero(next_frame);
+
+  cpi->common.frame_type = KEY_FRAME;
+
+  // Reset the GF group data structures.
+  av1_zero(*gf_group);
+
+  // Is this a forced key frame by interval.
+  rc->this_key_frame_forced = rc->next_key_frame_forced;
+
+  // Clear the alt ref active flag and last group multi arf flags as they
+  // can never be set for a key frame.
+  rc->source_alt_ref_active = 0;
+  cpi->multi_arf_last_grp_enabled = 0;
+
+  // KF is always a GF so clear frames till next gf counter.
+  rc->frames_till_gf_update_due = 0;
+
+  rc->frames_to_key = 1;
+
+  twopass->kf_group_bits = 0;        // Total bits available to kf group
+  twopass->kf_group_error_left = 0;  // Group modified error score.
+
+  kf_mod_err = calculate_modified_err(cpi, twopass, oxcf, this_frame);
+
+  // Initialize the decay rates for the recent frames to check
+  for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j) recent_loop_decay[j] = 1.0;
+
+  // Find the next keyframe.
+  i = 0;
+  while (twopass->stats_in < twopass->stats_in_end &&
+         rc->frames_to_key < cpi->oxcf.key_freq) {
+    // Accumulate kf group error.
+    kf_group_err += calculate_modified_err(cpi, twopass, oxcf, this_frame);
+
+    // Load the next frame's stats.
+    last_frame = *this_frame;
+    input_stats(twopass, this_frame);
+
+    // Provided that we are not at the end of the file...
+    if (cpi->oxcf.auto_key && twopass->stats_in < twopass->stats_in_end) {
+      double loop_decay_rate;
+
+      // Check for a scene cut.
+      if (test_candidate_kf(twopass, &last_frame, this_frame,
+                            twopass->stats_in))
+        break;
+
+      // How fast is the prediction quality decaying?
+      loop_decay_rate = get_prediction_decay_rate(cpi, twopass->stats_in);
+
+      // We want to know something about the recent past... rather than
+      // as used elsewhere where we are concerned with decay in prediction
+      // quality since the last GF or KF.
+      recent_loop_decay[i % FRAMES_TO_CHECK_DECAY] = loop_decay_rate;
+      decay_accumulator = 1.0;
+      for (j = 0; j < FRAMES_TO_CHECK_DECAY; ++j)
+        decay_accumulator *= recent_loop_decay[j];
+
+      // Special check for transition or high motion followed by a
+      // static scene.
+      if (detect_transition_to_still(cpi, i, cpi->oxcf.key_freq - i,
+                                     loop_decay_rate, decay_accumulator))
+        break;
+
+      // Step on to the next frame.
+      ++rc->frames_to_key;
+
+      // If we don't have a real key frame within the next two
+      // key_freq intervals then break out of the loop.
+      if (rc->frames_to_key >= 2 * cpi->oxcf.key_freq) break;
+    } else {
+      ++rc->frames_to_key;
+    }
+    ++i;
+  }
+
+  // If there is a max kf interval set by the user we must obey it.
+  // We already breakout of the loop above at 2x max.
+  // This code centers the extra kf if the actual natural interval
+  // is between 1x and 2x.
+  if (cpi->oxcf.auto_key && rc->frames_to_key > cpi->oxcf.key_freq) {
+    FIRSTPASS_STATS tmp_frame = first_frame;
+
+    rc->frames_to_key /= 2;
+
+    // Reset to the start of the group.
+    reset_fpf_position(twopass, start_position);
+
+    kf_group_err = 0.0;
+
+    // Rescan to get the correct error data for the forced kf group.
+    for (i = 0; i < rc->frames_to_key; ++i) {
+      kf_group_err += calculate_modified_err(cpi, twopass, oxcf, &tmp_frame);
+      input_stats(twopass, &tmp_frame);
+    }
+    rc->next_key_frame_forced = 1;
+  } else if (twopass->stats_in == twopass->stats_in_end ||
+             rc->frames_to_key >= cpi->oxcf.key_freq) {
+    rc->next_key_frame_forced = 1;
+  } else {
+    rc->next_key_frame_forced = 0;
+  }
+
+  // Special case for the last key frame of the file.
+  if (twopass->stats_in >= twopass->stats_in_end) {
+    // Accumulate kf group error.
+    kf_group_err += calculate_modified_err(cpi, twopass, oxcf, this_frame);
+  }
+
+  // Calculate the number of bits that should be assigned to the kf group.
+  if (twopass->bits_left > 0 && twopass->modified_error_left > 0.0) {
+    // Maximum number of bits for a single normal frame (not key frame).
+    const int max_bits = frame_max_bits(rc, &cpi->oxcf);
+
+    // Maximum number of bits allocated to the key frame group.
+    int64_t max_grp_bits;
+
+    // Default allocation based on bits left and relative
+    // complexity of the section.
+    twopass->kf_group_bits = (int64_t)(
+        twopass->bits_left * (kf_group_err / twopass->modified_error_left));
+
+    // Clip based on maximum per frame rate defined by the user.
+    max_grp_bits = (int64_t)max_bits * (int64_t)rc->frames_to_key;
+    if (twopass->kf_group_bits > max_grp_bits)
+      twopass->kf_group_bits = max_grp_bits;
+  } else {
+    twopass->kf_group_bits = 0;
+  }
+  twopass->kf_group_bits = AOMMAX(0, twopass->kf_group_bits);
+
+  // Reset the first pass file position.
+  reset_fpf_position(twopass, start_position);
+
+  // Scan through the kf group collating various stats used to determine
+  // how many bits to spend on it.
+  decay_accumulator = 1.0;
+  boost_score = 0.0;
+  for (i = 0; i < (rc->frames_to_key - 1); ++i) {
+    if (EOF == input_stats(twopass, &next_frame)) break;
+
+    // Monitor for static sections.
+    zero_motion_accumulator = AOMMIN(zero_motion_accumulator,
+                                     get_zero_motion_factor(cpi, &next_frame));
+
+    // Not all frames in the group are necessarily used in calculating boost.
+    if ((i <= rc->max_gf_interval) ||
+        ((i <= (rc->max_gf_interval * 4)) && (decay_accumulator > 0.5))) {
+      const double frame_boost =
+          calc_frame_boost(cpi, this_frame, 0, KF_MAX_BOOST);
+
+      // How fast is prediction quality decaying.
+      if (!detect_flash(twopass, 0)) {
+        const double loop_decay_rate =
+            get_prediction_decay_rate(cpi, &next_frame);
+        decay_accumulator *= loop_decay_rate;
+        decay_accumulator = AOMMAX(decay_accumulator, MIN_DECAY_FACTOR);
+        av_decay_accumulator += decay_accumulator;
+        ++loop_decay_counter;
+      }
+      boost_score += (decay_accumulator * frame_boost);
+    }
+  }
+  av_decay_accumulator /= (double)loop_decay_counter;
+
+  reset_fpf_position(twopass, start_position);
+
+  // Store the zero motion percentage
+  twopass->kf_zeromotion_pct = (int)(zero_motion_accumulator * 100.0);
+
+  // Calculate a section intra ratio used in setting max loop filter.
+  twopass->section_intra_rating = calculate_section_intra_ratio(
+      start_position, twopass->stats_in_end, rc->frames_to_key);
+
+  // Apply various clamps for min and max boost
+  rc->kf_boost = (int)(av_decay_accumulator * boost_score);
+  rc->kf_boost = AOMMAX(rc->kf_boost, (rc->frames_to_key * 3));
+  rc->kf_boost = AOMMAX(rc->kf_boost, MIN_KF_BOOST);
+
+  // Work out how many bits to allocate for the key frame itself.
+  kf_bits = calculate_boost_bits((rc->frames_to_key - 1), rc->kf_boost,
+                                 twopass->kf_group_bits);
+
+  // Work out the fraction of the kf group bits reserved for the inter frames
+  // within the group after discounting the bits for the kf itself.
+  if (twopass->kf_group_bits) {
+    twopass->kfgroup_inter_fraction =
+        (double)(twopass->kf_group_bits - kf_bits) /
+        (double)twopass->kf_group_bits;
+  } else {
+    twopass->kfgroup_inter_fraction = 1.0;
+  }
+
+  twopass->kf_group_bits -= kf_bits;
+
+  // Save the bits to spend on the key frame.
+  gf_group->bit_allocation[0] = kf_bits;
+  gf_group->update_type[0] = KF_UPDATE;
+  gf_group->rf_level[0] = KF_STD;
+
+  // Note the total error score of the kf group minus the key frame itself.
+  twopass->kf_group_error_left = (int)(kf_group_err - kf_mod_err);
+
+  // Adjust the count of total modified error left.
+  // The count of bits left is adjusted elsewhere based on real coded frame
+  // sizes.
+  twopass->modified_error_left -= kf_group_err;
+
+  if (oxcf->resize_mode == RESIZE_DYNAMIC) {
+    // Default to normal-sized frame on keyframes.
+    cpi->rc.next_frame_size_selector = UNSCALED;
+  }
+}
+
+// Define the reference buffers that will be updated post encode.
+static void configure_buffer_updates(AV1_COMP *cpi) {
+  TWO_PASS *const twopass = &cpi->twopass;
+
+  // Wei-Ting: Should we define another function to take care of
+  // cpi->rc.is_$Source_Type to make this function as it is in the comment?
+
+  cpi->rc.is_src_frame_alt_ref = 0;
+#if CONFIG_EXT_REFS
+  cpi->rc.is_bwd_ref_frame = 0;
+  cpi->rc.is_last_bipred_frame = 0;
+  cpi->rc.is_bipred_frame = 0;
+  cpi->rc.is_src_frame_ext_arf = 0;
+#endif  // CONFIG_EXT_REFS
+
+  switch (twopass->gf_group.update_type[twopass->gf_group.index]) {
+    case KF_UPDATE:
+#if CONFIG_EXT_REFS
+      cpi->refresh_bwd_ref_frame = 1;
+#endif  // CONFIG_EXT_REFS
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 1;
+      cpi->refresh_alt_ref_frame = 1;
+      break;
+
+    case LF_UPDATE:
+#if CONFIG_EXT_REFS
+      // If we have extra ALT_REFs, we can use the farthest ALT (ALT0) as
+      // the BWD_REF.
+      if (cpi->num_extra_arfs) {
+        int tmp = cpi->bwd_fb_idx;
+
+        cpi->bwd_fb_idx = cpi->alt_fb_idx;
+        cpi->alt_fb_idx = cpi->arf_map[0];
+        cpi->arf_map[0] = tmp;
+
+        cpi->rc.is_bwd_ref_frame = 1;
+      } else {
+        cpi->rc.is_bwd_ref_frame = 0;
+      }
+#endif  // CONFIG_EXT_REFS
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_alt_ref_frame = 0;
+      break;
+
+    case GF_UPDATE:
+#if CONFIG_EXT_REFS
+      cpi->refresh_bwd_ref_frame = 0;
+#endif  // CONFIG_EXT_REFS
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 1;
+      cpi->refresh_alt_ref_frame = 0;
+      break;
+
+    case OVERLAY_UPDATE:
+      cpi->refresh_last_frame = 0;
+      cpi->refresh_golden_frame = 1;
+#if CONFIG_EXT_REFS
+      cpi->refresh_bwd_ref_frame = 0;
+#endif  // CONFIG_EXT_REFS
+      cpi->refresh_alt_ref_frame = 0;
+      cpi->rc.is_src_frame_alt_ref = 1;
+      break;
+
+    case ARF_UPDATE:
+#if CONFIG_EXT_REFS
+      cpi->refresh_bwd_ref_frame = 1;
+#endif  // CONFIG_EXT_REFS
+      cpi->refresh_last_frame = 0;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_alt_ref_frame = 1;
+      break;
+
+#if CONFIG_EXT_REFS
+    case BRF_UPDATE:
+      cpi->refresh_last_frame = 0;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_bwd_ref_frame = 1;
+      cpi->refresh_alt_ref_frame = 0;
+      cpi->rc.is_bwd_ref_frame = 1;
+      if (cpi->num_extra_arfs) {
+        // Allow BRF use the farthest ALT_REF (ALT0) as BWD_REF by swapping
+        // the virtual indices.
+        // NOTE: The indices will be swapped back after this frame is encoded
+        //       (in av1_update_reference_frames()).
+        int tmp = cpi->bwd_fb_idx;
+
+        cpi->bwd_fb_idx = cpi->alt_fb_idx;
+        cpi->alt_fb_idx = cpi->arf_map[0];
+        cpi->arf_map[0] = tmp;
+      }
+      break;
+
+    case LAST_BIPRED_UPDATE:
+      cpi->refresh_last_frame = 0;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_bwd_ref_frame = 0;
+      cpi->refresh_alt_ref_frame = 0;
+      cpi->rc.is_last_bipred_frame = 1;
+      break;
+
+    case BIPRED_UPDATE:
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_bwd_ref_frame = 0;
+      cpi->refresh_alt_ref_frame = 0;
+      cpi->rc.is_bipred_frame = 1;
+      break;
+
+    case INTNL_OVERLAY_UPDATE:
+      cpi->refresh_last_frame = 1;
+      cpi->refresh_golden_frame = 0;
+      cpi->refresh_bwd_ref_frame = 0;
+      cpi->refresh_alt_ref_frame = 0;
+      cpi->rc.is_src_frame_alt_ref = 1;
+      cpi->rc.is_src_frame_ext_arf = 1;
+      break;
+#endif  // CONFIG_EXT_REFS
+
+    default: assert(0); break;
+  }
+}
+
+static int is_skippable_frame(const AV1_COMP *cpi) {
+  // If the current frame does not have non-zero motion vector detected in the
+  // first  pass, and so do its previous and forward frames, then this frame
+  // can be skipped for partition check, and the partition size is assigned
+  // according to the variance
+  const TWO_PASS *const twopass = &cpi->twopass;
+
+  return (!frame_is_intra_only(&cpi->common) &&
+          twopass->stats_in - 2 > twopass->stats_in_start &&
+          twopass->stats_in < twopass->stats_in_end &&
+          (twopass->stats_in - 1)->pcnt_inter -
+                  (twopass->stats_in - 1)->pcnt_motion ==
+              1 &&
+          (twopass->stats_in - 2)->pcnt_inter -
+                  (twopass->stats_in - 2)->pcnt_motion ==
+              1 &&
+          twopass->stats_in->pcnt_inter - twopass->stats_in->pcnt_motion == 1);
+}
+
+void av1_rc_get_second_pass_params(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  TWO_PASS *const twopass = &cpi->twopass;
+  GF_GROUP *const gf_group = &twopass->gf_group;
+  int frames_left;
+  FIRSTPASS_STATS this_frame;
+
+  int target_rate;
+
+  frames_left = (int)(twopass->total_stats.count - cm->current_video_frame);
+
+  if (!twopass->stats_in) return;
+
+  // If this is an arf frame then we dont want to read the stats file or
+  // advance the input pointer as we already have what we need.
+  if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
+    configure_buffer_updates(cpi);
+    target_rate = gf_group->bit_allocation[gf_group->index];
+    target_rate = av1_rc_clamp_pframe_target_size(cpi, target_rate);
+    rc->base_frame_target = target_rate;
+
+    cm->frame_type = INTER_FRAME;
+
+    // Do the firstpass stats indicate that this frame is skippable for the
+    // partition search?
+    if (cpi->sf.allow_partition_search_skip && cpi->oxcf.pass == 2) {
+      cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
+    }
+
+    return;
+  }
+
+  aom_clear_system_state();
+
+  if (cpi->oxcf.rc_mode == AOM_Q) {
+    twopass->active_worst_quality = cpi->oxcf.cq_level;
+  } else if (cm->current_video_frame == 0) {
+    // Special case code for first frame.
+    const int section_target_bandwidth =
+        (int)(twopass->bits_left / frames_left);
+    const double section_length = twopass->total_left_stats.count;
+    const double section_error =
+        twopass->total_left_stats.coded_error / section_length;
+    const double section_intra_skip =
+        twopass->total_left_stats.intra_skip_pct / section_length;
+    const double section_inactive_zone =
+        (twopass->total_left_stats.inactive_zone_rows * 2) /
+        ((double)cm->mb_rows * section_length);
+    const int tmp_q = get_twopass_worst_quality(
+        cpi, section_error, section_intra_skip + section_inactive_zone,
+        section_target_bandwidth, DEFAULT_GRP_WEIGHT);
+
+    twopass->active_worst_quality = tmp_q;
+    twopass->baseline_active_worst_quality = tmp_q;
+    rc->ni_av_qi = tmp_q;
+    rc->last_q[INTER_FRAME] = tmp_q;
+    rc->avg_q = av1_convert_qindex_to_q(tmp_q, cm->bit_depth);
+    rc->avg_frame_qindex[INTER_FRAME] = tmp_q;
+    rc->last_q[KEY_FRAME] = (tmp_q + cpi->oxcf.best_allowed_q) / 2;
+    rc->avg_frame_qindex[KEY_FRAME] = rc->last_q[KEY_FRAME];
+  }
+
+  av1_zero(this_frame);
+  if (EOF == input_stats(twopass, &this_frame)) return;
+
+  // Set the frame content type flag.
+  if (this_frame.intra_skip_pct >= FC_ANIMATION_THRESH)
+    twopass->fr_content_type = FC_GRAPHICS_ANIMATION;
+  else
+    twopass->fr_content_type = FC_NORMAL;
+
+  // Keyframe and section processing.
+  if (rc->frames_to_key == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY)) {
+    FIRSTPASS_STATS this_frame_copy;
+    this_frame_copy = this_frame;
+    // Define next KF group and assign bits to it.
+    find_next_key_frame(cpi, &this_frame);
+    this_frame = this_frame_copy;
+  } else {
+    cm->frame_type = INTER_FRAME;
+  }
+
+  // Define a new GF/ARF group. (Should always enter here for key frames).
+  if (rc->frames_till_gf_update_due == 0) {
+    define_gf_group(cpi, &this_frame);
+
+    rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+
+#if ARF_STATS_OUTPUT
+    {
+      FILE *fpfile;
+      fpfile = fopen("arf.stt", "a");
+      ++arf_count;
+      fprintf(fpfile, "%10d %10ld %10d %10d %10ld\n", cm->current_video_frame,
+              rc->frames_till_gf_update_due, rc->kf_boost, arf_count,
+              rc->gfu_boost);
+
+      fclose(fpfile);
+    }
+#endif
+  }
+
+  configure_buffer_updates(cpi);
+
+  // Do the firstpass stats indicate that this frame is skippable for the
+  // partition search?
+  if (cpi->sf.allow_partition_search_skip && cpi->oxcf.pass == 2) {
+    cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
+  }
+
+  target_rate = gf_group->bit_allocation[gf_group->index];
+
+  if (cpi->common.frame_type == KEY_FRAME)
+    target_rate = av1_rc_clamp_iframe_target_size(cpi, target_rate);
+  else
+    target_rate = av1_rc_clamp_pframe_target_size(cpi, target_rate);
+
+  rc->base_frame_target = target_rate;
+
+  {
+    const int num_mbs = (cpi->oxcf.resize_mode != RESIZE_NONE)
+                            ? cpi->initial_mbs
+                            : cpi->common.MBs;
+    // The multiplication by 256 reverses a scaling factor of (>> 8)
+    // applied when combining MB error values for the frame.
+    twopass->mb_av_energy =
+        log(((this_frame.intra_error * 256.0) / num_mbs) + 1.0);
+  }
+
+  // Update the total stats remaining structure.
+  subtract_stats(&twopass->total_left_stats, &this_frame);
+}
+
+#define MINQ_ADJ_LIMIT 48
+#define MINQ_ADJ_LIMIT_CQ 20
+#define HIGH_UNDERSHOOT_RATIO 2
+void av1_twopass_postencode_update(AV1_COMP *cpi) {
+  TWO_PASS *const twopass = &cpi->twopass;
+  RATE_CONTROL *const rc = &cpi->rc;
+  const int bits_used = rc->base_frame_target;
+
+  // VBR correction is done through rc->vbr_bits_off_target. Based on the
+  // sign of this value, a limited % adjustment is made to the target rate
+  // of subsequent frames, to try and push it back towards 0. This method
+  // is designed to prevent extreme behaviour at the end of a clip
+  // or group of frames.
+  rc->vbr_bits_off_target += rc->base_frame_target - rc->projected_frame_size;
+  twopass->bits_left = AOMMAX(twopass->bits_left - bits_used, 0);
+
+  // Calculate the pct rc error.
+  if (rc->total_actual_bits) {
+    rc->rate_error_estimate =
+        (int)((rc->vbr_bits_off_target * 100) / rc->total_actual_bits);
+    rc->rate_error_estimate = clamp(rc->rate_error_estimate, -100, 100);
+  } else {
+    rc->rate_error_estimate = 0;
+  }
+
+  if (cpi->common.frame_type != KEY_FRAME) {
+    twopass->kf_group_bits -= bits_used;
+    twopass->last_kfgroup_zeromotion_pct = twopass->kf_zeromotion_pct;
+  }
+  twopass->kf_group_bits = AOMMAX(twopass->kf_group_bits, 0);
+
+  // Increment the gf group index ready for the next frame.
+  ++twopass->gf_group.index;
+
+  // If the rate control is drifting consider adjustment to min or maxq.
+  if ((cpi->oxcf.rc_mode != AOM_Q) &&
+      (cpi->twopass.gf_zeromotion_pct < VLOW_MOTION_THRESHOLD) &&
+      !cpi->rc.is_src_frame_alt_ref) {
+    const int maxq_adj_limit =
+        rc->worst_quality - twopass->active_worst_quality;
+    const int minq_adj_limit =
+        (cpi->oxcf.rc_mode == AOM_CQ ? MINQ_ADJ_LIMIT_CQ : MINQ_ADJ_LIMIT);
+
+    // Undershoot.
+    if (rc->rate_error_estimate > cpi->oxcf.under_shoot_pct) {
+      --twopass->extend_maxq;
+      if (rc->rolling_target_bits >= rc->rolling_actual_bits)
+        ++twopass->extend_minq;
+      // Overshoot.
+    } else if (rc->rate_error_estimate < -cpi->oxcf.over_shoot_pct) {
+      --twopass->extend_minq;
+      if (rc->rolling_target_bits < rc->rolling_actual_bits)
+        ++twopass->extend_maxq;
+    } else {
+      // Adjustment for extreme local overshoot.
+      if (rc->projected_frame_size > (2 * rc->base_frame_target) &&
+          rc->projected_frame_size > (2 * rc->avg_frame_bandwidth))
+        ++twopass->extend_maxq;
+
+      // Unwind undershoot or overshoot adjustment.
+      if (rc->rolling_target_bits < rc->rolling_actual_bits)
+        --twopass->extend_minq;
+      else if (rc->rolling_target_bits > rc->rolling_actual_bits)
+        --twopass->extend_maxq;
+    }
+
+    twopass->extend_minq = clamp(twopass->extend_minq, 0, minq_adj_limit);
+    twopass->extend_maxq = clamp(twopass->extend_maxq, 0, maxq_adj_limit);
+
+    // If there is a big and undexpected undershoot then feed the extra
+    // bits back in quickly. One situation where this may happen is if a
+    // frame is unexpectedly almost perfectly predicted by the ARF or GF
+    // but not very well predcited by the previous frame.
+    if (!frame_is_kf_gf_arf(cpi) && !cpi->rc.is_src_frame_alt_ref) {
+      int fast_extra_thresh = rc->base_frame_target / HIGH_UNDERSHOOT_RATIO;
+      if (rc->projected_frame_size < fast_extra_thresh) {
+        rc->vbr_bits_off_target_fast +=
+            fast_extra_thresh - rc->projected_frame_size;
+        rc->vbr_bits_off_target_fast =
+            AOMMIN(rc->vbr_bits_off_target_fast, (4 * rc->avg_frame_bandwidth));
+
+        // Fast adaptation of minQ if necessary to use up the extra bits.
+        if (rc->avg_frame_bandwidth) {
+          twopass->extend_minq_fast =
+              (int)(rc->vbr_bits_off_target_fast * 8 / rc->avg_frame_bandwidth);
+        }
+        twopass->extend_minq_fast = AOMMIN(
+            twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
+      } else if (rc->vbr_bits_off_target_fast) {
+        twopass->extend_minq_fast = AOMMIN(
+            twopass->extend_minq_fast, minq_adj_limit - twopass->extend_minq);
+      } else {
+        twopass->extend_minq_fast = 0;
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/encoder/firstpass.h b/third_party/aom/av1/encoder/firstpass.h
new file mode 100644
index 0000000000..db459cc229
--- /dev/null
+++ b/third_party/aom/av1/encoder/firstpass.h
@@ -0,0 +1,202 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_FIRSTPASS_H_
+#define AV1_ENCODER_FIRSTPASS_H_
+
+#include "av1/encoder/lookahead.h"
+#include "av1/encoder/ratectrl.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if CONFIG_FP_MB_STATS
+
+#define FPMB_DCINTRA_MASK 0x01
+
+#define FPMB_MOTION_ZERO_MASK 0x02
+#define FPMB_MOTION_LEFT_MASK 0x04
+#define FPMB_MOTION_RIGHT_MASK 0x08
+#define FPMB_MOTION_UP_MASK 0x10
+#define FPMB_MOTION_DOWN_MASK 0x20
+
+#define FPMB_ERROR_SMALL_MASK 0x40
+#define FPMB_ERROR_LARGE_MASK 0x80
+#define FPMB_ERROR_SMALL_TH 2000
+#define FPMB_ERROR_LARGE_TH 48000
+
+typedef struct {
+  uint8_t *mb_stats_start;
+  uint8_t *mb_stats_end;
+} FIRSTPASS_MB_STATS;
+#endif
+
+#if CONFIG_EXT_REFS
+// Length of the bi-predictive frame group (BFG)
+// NOTE: Currently each BFG contains one backward ref (BWF) frame plus a certain
+//       number of bi-predictive frames.
+#define BFG_INTERVAL 2
+// The maximum number of extra ALT_REF's
+// NOTE: This number cannot be greater than 2 or the reference frame buffer will
+//       overflow.
+#define MAX_EXT_ARFS 2
+#define MIN_EXT_ARF_INTERVAL 4
+#endif  // CONFIG_EXT_REFS
+
+#define VLOW_MOTION_THRESHOLD 950
+
+typedef struct {
+  double frame;
+  double weight;
+  double intra_error;
+  double coded_error;
+  double sr_coded_error;
+  double pcnt_inter;
+  double pcnt_motion;
+  double pcnt_second_ref;
+  double pcnt_neutral;
+  double intra_skip_pct;
+  double inactive_zone_rows;  // Image mask rows top and bottom.
+  double inactive_zone_cols;  // Image mask columns at left and right edges.
+  double MVr;
+  double mvr_abs;
+  double MVc;
+  double mvc_abs;
+  double MVrv;
+  double MVcv;
+  double mv_in_out_count;
+  double new_mv_count;
+  double duration;
+  double count;
+} FIRSTPASS_STATS;
+
+typedef enum {
+  KF_UPDATE = 0,
+  LF_UPDATE = 1,
+  GF_UPDATE = 2,
+  ARF_UPDATE = 3,
+  OVERLAY_UPDATE = 4,
+#if CONFIG_EXT_REFS
+  BRF_UPDATE = 5,            // Backward Reference Frame
+  LAST_BIPRED_UPDATE = 6,    // Last Bi-predictive Frame
+  BIPRED_UPDATE = 7,         // Bi-predictive Frame, but not the last one
+  INTNL_OVERLAY_UPDATE = 8,  // Internal Overlay Frame
+  FRAME_UPDATE_TYPES = 9
+#else
+  FRAME_UPDATE_TYPES = 5
+#endif  // CONFIG_EXT_REFS
+} FRAME_UPDATE_TYPE;
+
+#define FC_ANIMATION_THRESH 0.15
+typedef enum {
+  FC_NORMAL = 0,
+  FC_GRAPHICS_ANIMATION = 1,
+  FRAME_CONTENT_TYPES = 2
+} FRAME_CONTENT_TYPE;
+
+typedef struct {
+  unsigned char index;
+  RATE_FACTOR_LEVEL rf_level[(MAX_LAG_BUFFERS * 2) + 1];
+  FRAME_UPDATE_TYPE update_type[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_src_offset[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_update_idx[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char arf_ref_idx[(MAX_LAG_BUFFERS * 2) + 1];
+#if CONFIG_EXT_REFS
+  unsigned char brf_src_offset[(MAX_LAG_BUFFERS * 2) + 1];
+  unsigned char bidir_pred_enabled[(MAX_LAG_BUFFERS * 2) + 1];
+#endif  // CONFIG_EXT_REFS
+  int bit_allocation[(MAX_LAG_BUFFERS * 2) + 1];
+} GF_GROUP;
+
+typedef struct {
+  unsigned int section_intra_rating;
+  FIRSTPASS_STATS total_stats;
+  FIRSTPASS_STATS this_frame_stats;
+  const FIRSTPASS_STATS *stats_in;
+  const FIRSTPASS_STATS *stats_in_start;
+  const FIRSTPASS_STATS *stats_in_end;
+  FIRSTPASS_STATS total_left_stats;
+  int first_pass_done;
+  int64_t bits_left;
+  double modified_error_min;
+  double modified_error_max;
+  double modified_error_left;
+  double mb_av_energy;
+
+#if CONFIG_FP_MB_STATS
+  uint8_t *frame_mb_stats_buf;
+  uint8_t *this_frame_mb_stats;
+  FIRSTPASS_MB_STATS firstpass_mb_stats;
+#endif
+  // An indication of the content type of the current frame
+  FRAME_CONTENT_TYPE fr_content_type;
+
+  // Projected total bits available for a key frame group of frames
+  int64_t kf_group_bits;
+
+  // Error score of frames still to be coded in kf group
+  int64_t kf_group_error_left;
+
+  // The fraction for a kf groups total bits allocated to the inter frames
+  double kfgroup_inter_fraction;
+
+  int sr_update_lag;
+
+  int kf_zeromotion_pct;
+  int last_kfgroup_zeromotion_pct;
+  int gf_zeromotion_pct;
+  int active_worst_quality;
+  int baseline_active_worst_quality;
+  int extend_minq;
+  int extend_maxq;
+  int extend_minq_fast;
+
+  GF_GROUP gf_group;
+} TWO_PASS;
+
+struct AV1_COMP;
+
+void av1_init_first_pass(struct AV1_COMP *cpi);
+void av1_rc_get_first_pass_params(struct AV1_COMP *cpi);
+void av1_first_pass(struct AV1_COMP *cpi, const struct lookahead_entry *source);
+void av1_end_first_pass(struct AV1_COMP *cpi);
+
+void av1_init_second_pass(struct AV1_COMP *cpi);
+void av1_rc_get_second_pass_params(struct AV1_COMP *cpi);
+void av1_twopass_postencode_update(struct AV1_COMP *cpi);
+
+// Post encode update of the rate control parameters for 2-pass
+void av1_twopass_postencode_update(struct AV1_COMP *cpi);
+
+void av1_init_subsampling(struct AV1_COMP *cpi);
+
+void av1_calculate_coded_size(struct AV1_COMP *cpi, int *scaled_frame_width,
+                              int *scaled_frame_height);
+
+#if CONFIG_EXT_REFS
+static INLINE int get_number_of_extra_arfs(int interval, int arf_pending) {
+  if (arf_pending && MAX_EXT_ARFS > 0)
+    return interval >= MIN_EXT_ARF_INTERVAL * (MAX_EXT_ARFS + 1)
+               ? MAX_EXT_ARFS
+               : interval >= MIN_EXT_ARF_INTERVAL * MAX_EXT_ARFS
+                     ? MAX_EXT_ARFS - 1
+                     : 0;
+  else
+    return 0;
+}
+#endif  // CONFIG_EXT_REFS
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_FIRSTPASS_H_
diff --git a/third_party/aom/av1/encoder/generic_encoder.c b/third_party/aom/av1/encoder/generic_encoder.c
new file mode 100644
index 0000000000..a31bb9ef69
--- /dev/null
+++ b/third_party/aom/av1/encoder/generic_encoder.c
@@ -0,0 +1,157 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdio.h>
+
+#include "aom_dsp/bitwriter.h"
+#include "av1/common/generic_code.h"
+#include "av1/common/odintrin.h"
+#include "pvq_encoder.h"
+
+/** Encodes a value from 0 to N-1 (with N up to 16) based on a cdf and adapts
+ * the cdf accordingly.
+ *
+ * @param [in,out] w     multi-symbol entropy encoder
+ * @param [in]     val   variable being encoded
+ * @param [in,out] cdf   CDF of the variable (Q15)
+ * @param [in]     n     number of values possible
+ * @param [in,out] count number of symbols encoded with that cdf so far
+ * @param [in]     rate  adaptation rate shift (smaller is faster)
+ */
+void aom_encode_cdf_adapt_q15(aom_writer *w, int val, uint16_t *cdf, int n,
+ int *count, int rate) {
+  int i;
+  if (*count == 0) {
+    /* On the first call, we normalize the cdf to (32768 - n). This should
+       eventually be moved to the state init, but for now it makes it much
+       easier to experiment and convert symbols to the Q15 adaptation.*/
+    int ft;
+    ft = cdf[n - 1];
+    for (i = 0; i < n; i++) {
+      cdf[i] = AOM_ICDF(cdf[i]*32768/ft);
+    }
+  }
+  aom_write_cdf(w, val, cdf, n);
+  aom_cdf_adapt_q15(val, cdf, n, count, rate);
+}
+
+/** Encodes a random variable using a "generic" model, assuming that the
+ * distribution is one-sided (zero and up), has a single mode, and decays
+ * exponentially past the model.
+ *
+ * @param [in,out] w     multi-symbol entropy encoder
+ * @param [in,out] model generic probability model
+ * @param [in]     x     variable being encoded
+ * @param [in,out] ExQ16 expectation of x (adapted)
+ * @param [in]     integration integration period of ExQ16 (leaky average over
+ * 1<<integration samples)
+ */
+void generic_encode(aom_writer *w, generic_encoder *model, int x,
+ int *ex_q16, int integration) {
+  int lg_q1;
+  int shift;
+  int id;
+  uint16_t *cdf;
+  int xs;
+  lg_q1 = log_ex(*ex_q16);
+  OD_LOG((OD_LOG_ENTROPY_CODER, OD_LOG_DEBUG,
+   "%d %d", *ex_q16, lg_q1));
+  /* If expectation is too large, shift x to ensure that
+     all we have past xs=15 is the exponentially decaying tail
+     of the distribution */
+  shift = OD_MAXI(0, (lg_q1 - 5) >> 1);
+  /* Choose the cdf to use: we have two per "octave" of ExQ16 */
+  id = OD_MINI(GENERIC_TABLES - 1, lg_q1);
+  cdf = model->cdf[id];
+  xs = (x + (1 << shift >> 1)) >> shift;
+  aom_write_symbol_pvq(w, OD_MINI(15, xs), cdf, 16);
+  if (xs >= 15) {
+    int e;
+    unsigned decay;
+    /* Estimate decay based on the assumption that the distribution is close
+       to Laplacian for large values. We should probably have an adaptive
+       estimate instead. Note: The 2* is a kludge that's not fully understood
+       yet. */
+    OD_ASSERT(*ex_q16 < INT_MAX >> 1);
+    e = ((2**ex_q16 >> 8) + (1 << shift >> 1)) >> shift;
+    decay = OD_MAXI(2, OD_MINI(254, 256*e/(e + 256)));
+    /* Encode the tail of the distribution assuming exponential decay. */
+    aom_laplace_encode_special(w, xs - 15, decay);
+  }
+  if (shift != 0) {
+    int special;
+    /* Because of the rounding, there's only half the number of possibilities
+       for xs=0. */
+    special = xs == 0;
+    if (shift - special > 0) {
+      aom_write_literal(w, x - (xs << shift) + (!special << (shift - 1)),
+       shift - special);
+    }
+  }
+  generic_model_update(ex_q16, x, integration);
+  OD_LOG((OD_LOG_ENTROPY_CODER, OD_LOG_DEBUG,
+   "enc: %d %d %d %d %d %x", *ex_q16, x, shift, id, xs, enc->rng));
+}
+
+/** Estimates the cost of encoding a value with generic_encode().
+ *
+ * @param [in,out] model generic probability model
+ * @param [in]     x     variable being encoded
+ * @param [in,out] ExQ16 expectation of x (adapted)
+ * @return number of bits (approximation)
+ */
+double generic_encode_cost(generic_encoder *model, int x, int *ex_q16) {
+  int lg_q1;
+  int shift;
+  int id;
+  uint16_t *cdf;
+  int xs;
+  int extra;
+  lg_q1 = log_ex(*ex_q16);
+  /* If expectation is too large, shift x to ensure that
+       all we have past xs=15 is the exponentially decaying tail
+       of the distribution */
+  shift = OD_MAXI(0, (lg_q1 - 5) >> 1);
+  /* Choose the cdf to use: we have two per "octave" of ExQ16 */
+  id = OD_MINI(GENERIC_TABLES - 1, lg_q1);
+  cdf = model->cdf[id];
+  xs = (x + (1 << shift >> 1)) >> shift;
+  extra = 0;
+  if (shift) extra = shift - (xs == 0);
+  xs = OD_MINI(15, xs);
+  /* Shortcut: assume it's going to cost 2 bits for the Laplace coder. */
+  if (xs == 15) extra += 2;
+  return
+      extra - OD_LOG2((double)(cdf[xs] - (xs == 0 ? 0 : cdf[xs - 1]))/cdf[15]);
+}
+
+/*Estimates the cost of encoding a value with a given CDF.*/
+double od_encode_cdf_cost(int val, uint16_t *cdf, int n) {
+  int total_prob;
+  int prev_prob;
+  double val_prob;
+  OD_ASSERT(n > 0);
+  total_prob = cdf[n - 1];
+  if (val == 0) {
+    prev_prob = 0;
+  }
+  else {
+    prev_prob = cdf[val - 1];
+  }
+  val_prob = (cdf[val] - prev_prob) / (double)total_prob;
+  return -OD_LOG2(val_prob);
+}
diff --git a/third_party/aom/av1/encoder/global_motion.c b/third_party/aom/av1/encoder/global_motion.c
new file mode 100644
index 0000000000..2a62049391
--- /dev/null
+++ b/third_party/aom/av1/encoder/global_motion.c
@@ -0,0 +1,319 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <memory.h>
+#include <math.h>
+#include <assert.h>
+
+#include "av1/encoder/global_motion.h"
+
+#include "av1/common/warped_motion.h"
+
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/corner_detect.h"
+#include "av1/encoder/corner_match.h"
+#include "av1/encoder/ransac.h"
+
+#define MAX_CORNERS 4096
+#define MIN_INLIER_PROB 0.1
+
+#define MIN_TRANS_THRESH (1 * GM_TRANS_DECODE_FACTOR)
+
+// Border over which to compute the global motion
+#define ERRORADV_BORDER 0
+
+#define ERRORADV_MAX_THRESH 0.995
+#define ERRORADV_COST_PRODUCT_THRESH 26000
+
+int is_enough_erroradvantage(double best_erroradvantage, int params_cost) {
+  return best_erroradvantage < ERRORADV_MAX_THRESH &&
+         best_erroradvantage * params_cost < ERRORADV_COST_PRODUCT_THRESH;
+}
+
+static void convert_to_params(const double *params, int32_t *model) {
+  int i;
+  int alpha_present = 0;
+  model[0] = (int32_t)floor(params[0] * (1 << GM_TRANS_PREC_BITS) + 0.5);
+  model[1] = (int32_t)floor(params[1] * (1 << GM_TRANS_PREC_BITS) + 0.5);
+  model[0] = (int32_t)clamp(model[0], GM_TRANS_MIN, GM_TRANS_MAX) *
+             GM_TRANS_DECODE_FACTOR;
+  model[1] = (int32_t)clamp(model[1], GM_TRANS_MIN, GM_TRANS_MAX) *
+             GM_TRANS_DECODE_FACTOR;
+
+  for (i = 2; i < 6; ++i) {
+    const int diag_value = ((i == 2 || i == 5) ? (1 << GM_ALPHA_PREC_BITS) : 0);
+    model[i] = (int32_t)floor(params[i] * (1 << GM_ALPHA_PREC_BITS) + 0.5);
+    model[i] =
+        (int32_t)clamp(model[i] - diag_value, GM_ALPHA_MIN, GM_ALPHA_MAX);
+    alpha_present |= (model[i] != 0);
+    model[i] = (model[i] + diag_value) * GM_ALPHA_DECODE_FACTOR;
+  }
+  for (; i < 8; ++i) {
+    model[i] = (int32_t)floor(params[i] * (1 << GM_ROW3HOMO_PREC_BITS) + 0.5);
+    model[i] = (int32_t)clamp(model[i], GM_ROW3HOMO_MIN, GM_ROW3HOMO_MAX) *
+               GM_ROW3HOMO_DECODE_FACTOR;
+    alpha_present |= (model[i] != 0);
+  }
+
+  if (!alpha_present) {
+    if (abs(model[0]) < MIN_TRANS_THRESH && abs(model[1]) < MIN_TRANS_THRESH) {
+      model[0] = 0;
+      model[1] = 0;
+    }
+  }
+}
+
+void convert_model_to_params(const double *params, WarpedMotionParams *model) {
+  convert_to_params(params, model->wmmat);
+  model->wmtype = get_gmtype(model);
+}
+
+// Adds some offset to a global motion parameter and handles
+// all of the necessary precision shifts, clamping, and
+// zero-centering.
+static int32_t add_param_offset(int param_index, int32_t param_value,
+                                int32_t offset) {
+  const int scale_vals[3] = { GM_TRANS_PREC_DIFF, GM_ALPHA_PREC_DIFF,
+                              GM_ROW3HOMO_PREC_DIFF };
+  const int clamp_vals[3] = { GM_TRANS_MAX, GM_ALPHA_MAX, GM_ROW3HOMO_MAX };
+  // type of param: 0 - translation, 1 - affine, 2 - homography
+  const int param_type = (param_index < 2 ? 0 : (param_index < 6 ? 1 : 2));
+  const int is_one_centered = (param_index == 2 || param_index == 5);
+
+  // Make parameter zero-centered and offset the shift that was done to make
+  // it compatible with the warped model
+  param_value = (param_value - (is_one_centered << WARPEDMODEL_PREC_BITS)) >>
+                scale_vals[param_type];
+  // Add desired offset to the rescaled/zero-centered parameter
+  param_value += offset;
+  // Clamp the parameter so it does not overflow the number of bits allotted
+  // to it in the bitstream
+  param_value = (int32_t)clamp(param_value, -clamp_vals[param_type],
+                               clamp_vals[param_type]);
+  // Rescale the parameter to WARPEDMODEL_PRECISION_BITS so it is compatible
+  // with the warped motion library
+  param_value *= (1 << scale_vals[param_type]);
+
+  // Undo the zero-centering step if necessary
+  return param_value + (is_one_centered << WARPEDMODEL_PREC_BITS);
+}
+
+static void force_wmtype(WarpedMotionParams *wm, TransformationType wmtype) {
+  switch (wmtype) {
+    case IDENTITY: wm->wmmat[0] = 0; wm->wmmat[1] = 0;
+    case TRANSLATION:
+      wm->wmmat[2] = 1 << WARPEDMODEL_PREC_BITS;
+      wm->wmmat[3] = 0;
+    case ROTZOOM: wm->wmmat[4] = -wm->wmmat[3]; wm->wmmat[5] = wm->wmmat[2];
+    case AFFINE: wm->wmmat[6] = wm->wmmat[7] = 0; break;
+    case HORTRAPEZOID: wm->wmmat[6] = wm->wmmat[4] = 0; break;
+    case VERTRAPEZOID: wm->wmmat[7] = wm->wmmat[3] = 0; break;
+    case HOMOGRAPHY: break;
+    default: assert(0);
+  }
+  wm->wmtype = wmtype;
+}
+
+double refine_integerized_param(WarpedMotionParams *wm,
+                                TransformationType wmtype,
+#if CONFIG_HIGHBITDEPTH
+                                int use_hbd, int bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                                uint8_t *ref, int r_width, int r_height,
+                                int r_stride, uint8_t *dst, int d_width,
+                                int d_height, int d_stride, int n_refinements) {
+  static const int max_trans_model_params[TRANS_TYPES] = {
+    0, 2, 4, 6, 8, 8, 8
+  };
+  const int border = ERRORADV_BORDER;
+  int i = 0, p;
+  int n_params = max_trans_model_params[wmtype];
+  int32_t *param_mat = wm->wmmat;
+  double step_error;
+  int32_t step;
+  int32_t *param;
+  int32_t curr_param;
+  int32_t best_param;
+  double best_error;
+
+  force_wmtype(wm, wmtype);
+  best_error = av1_warp_erroradv(wm,
+#if CONFIG_HIGHBITDEPTH
+                                 use_hbd, bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                                 ref, r_width, r_height, r_stride,
+                                 dst + border * d_stride + border, border,
+                                 border, d_width - 2 * border,
+                                 d_height - 2 * border, d_stride, 0, 0, 16, 16);
+  step = 1 << (n_refinements + 1);
+  for (i = 0; i < n_refinements; i++, step >>= 1) {
+    for (p = 0; p < n_params; ++p) {
+      int step_dir = 0;
+      // Skip searches for parameters that are forced to be 0
+      if (wmtype == HORTRAPEZOID && (p == 4 || p == 6)) continue;
+      if (wmtype == VERTRAPEZOID && (p == 3 || p == 7)) continue;
+      param = param_mat + p;
+      curr_param = *param;
+      best_param = curr_param;
+      // look to the left
+      *param = add_param_offset(p, curr_param, -step);
+      step_error = av1_warp_erroradv(
+          wm,
+#if CONFIG_HIGHBITDEPTH
+          use_hbd, bd,
+#endif  // CONFIG_HIGHBITDEPTH
+          ref, r_width, r_height, r_stride, dst + border * d_stride + border,
+          border, border, d_width - 2 * border, d_height - 2 * border, d_stride,
+          0, 0, 16, 16);
+      if (step_error < best_error) {
+        best_error = step_error;
+        best_param = *param;
+        step_dir = -1;
+      }
+
+      // look to the right
+      *param = add_param_offset(p, curr_param, step);
+      step_error = av1_warp_erroradv(
+          wm,
+#if CONFIG_HIGHBITDEPTH
+          use_hbd, bd,
+#endif  // CONFIG_HIGHBITDEPTH
+          ref, r_width, r_height, r_stride, dst + border * d_stride + border,
+          border, border, d_width - 2 * border, d_height - 2 * border, d_stride,
+          0, 0, 16, 16);
+      if (step_error < best_error) {
+        best_error = step_error;
+        best_param = *param;
+        step_dir = 1;
+      }
+      *param = best_param;
+
+      // look to the direction chosen above repeatedly until error increases
+      // for the biggest step size
+      while (step_dir) {
+        *param = add_param_offset(p, best_param, step * step_dir);
+        step_error = av1_warp_erroradv(
+            wm,
+#if CONFIG_HIGHBITDEPTH
+            use_hbd, bd,
+#endif  // CONFIG_HIGHBITDEPTH
+            ref, r_width, r_height, r_stride, dst + border * d_stride + border,
+            border, border, d_width - 2 * border, d_height - 2 * border,
+            d_stride, 0, 0, 16, 16);
+        if (step_error < best_error) {
+          best_error = step_error;
+          best_param = *param;
+        } else {
+          *param = best_param;
+          step_dir = 0;
+        }
+      }
+    }
+  }
+  force_wmtype(wm, wmtype);
+  wm->wmtype = get_gmtype(wm);
+  return best_error;
+}
+
+static INLINE RansacFunc get_ransac_type(TransformationType type) {
+  switch (type) {
+    case HOMOGRAPHY: return ransac_homography;
+    case HORTRAPEZOID: return ransac_hortrapezoid;
+    case VERTRAPEZOID: return ransac_vertrapezoid;
+    case AFFINE: return ransac_affine;
+    case ROTZOOM: return ransac_rotzoom;
+    case TRANSLATION: return ransac_translation;
+    default: assert(0); return NULL;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static unsigned char *downconvert_frame(YV12_BUFFER_CONFIG *frm,
+                                        int bit_depth) {
+  int i, j;
+  uint16_t *orig_buf = CONVERT_TO_SHORTPTR(frm->y_buffer);
+  uint8_t *buf = malloc(frm->y_height * frm->y_stride * sizeof(*buf));
+
+  for (i = 0; i < frm->y_height; ++i)
+    for (j = 0; j < frm->y_width; ++j)
+      buf[i * frm->y_stride + j] =
+          orig_buf[i * frm->y_stride + j] >> (bit_depth - 8);
+
+  return buf;
+}
+#endif
+
+int compute_global_motion_feature_based(
+    TransformationType type, YV12_BUFFER_CONFIG *frm, YV12_BUFFER_CONFIG *ref,
+#if CONFIG_HIGHBITDEPTH
+    int bit_depth,
+#endif
+    int *num_inliers_by_motion, double *params_by_motion, int num_motions) {
+  int i;
+  int num_frm_corners, num_ref_corners;
+  int num_correspondences;
+  int *correspondences;
+  int frm_corners[2 * MAX_CORNERS], ref_corners[2 * MAX_CORNERS];
+  unsigned char *frm_buffer = frm->y_buffer;
+  unsigned char *ref_buffer = ref->y_buffer;
+  RansacFunc ransac = get_ransac_type(type);
+
+#if CONFIG_HIGHBITDEPTH
+  if (frm->flags & YV12_FLAG_HIGHBITDEPTH) {
+    // The frame buffer is 16-bit, so we need to convert to 8 bits for the
+    // following code. We cache the result until the frame is released.
+    if (frm->y_buffer_8bit)
+      frm_buffer = frm->y_buffer_8bit;
+    else
+      frm_buffer = frm->y_buffer_8bit = downconvert_frame(frm, bit_depth);
+  }
+  if (ref->flags & YV12_FLAG_HIGHBITDEPTH) {
+    if (ref->y_buffer_8bit)
+      ref_buffer = ref->y_buffer_8bit;
+    else
+      ref_buffer = ref->y_buffer_8bit = downconvert_frame(ref, bit_depth);
+  }
+#endif
+
+  // compute interest points in images using FAST features
+  num_frm_corners = fast_corner_detect(frm_buffer, frm->y_width, frm->y_height,
+                                       frm->y_stride, frm_corners, MAX_CORNERS);
+  num_ref_corners = fast_corner_detect(ref_buffer, ref->y_width, ref->y_height,
+                                       ref->y_stride, ref_corners, MAX_CORNERS);
+
+  // find correspondences between the two images
+  correspondences =
+      (int *)malloc(num_frm_corners * 4 * sizeof(*correspondences));
+  num_correspondences = determine_correspondence(
+      frm_buffer, (int *)frm_corners, num_frm_corners, ref_buffer,
+      (int *)ref_corners, num_ref_corners, frm->y_width, frm->y_height,
+      frm->y_stride, ref->y_stride, correspondences);
+
+  ransac(correspondences, num_correspondences, num_inliers_by_motion,
+         params_by_motion, num_motions);
+
+  free(correspondences);
+
+  // Set num_inliers = 0 for motions with too few inliers so they are ignored.
+  for (i = 0; i < num_motions; ++i) {
+    if (num_inliers_by_motion[i] < MIN_INLIER_PROB * num_correspondences) {
+      num_inliers_by_motion[i] = 0;
+    }
+  }
+
+  // Return true if any one of the motions has inliers.
+  for (i = 0; i < num_motions; ++i) {
+    if (num_inliers_by_motion[i] > 0) return 1;
+  }
+  return 0;
+}
diff --git a/third_party/aom/av1/encoder/global_motion.h b/third_party/aom/av1/encoder/global_motion.h
new file mode 100644
index 0000000000..8fc757f387
--- /dev/null
+++ b/third_party/aom/av1/encoder/global_motion.h
@@ -0,0 +1,62 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_GLOBAL_MOTION_H_
+#define AV1_ENCODER_GLOBAL_MOTION_H_
+
+#include "aom/aom_integer.h"
+#include "aom_scale/yv12config.h"
+#include "av1/common/mv.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define RANSAC_NUM_MOTIONS 1
+
+void convert_model_to_params(const double *params, WarpedMotionParams *model);
+
+int is_enough_erroradvantage(double erroradv, int params_cost);
+
+double refine_integerized_param(WarpedMotionParams *wm,
+                                TransformationType wmtype,
+#if CONFIG_HIGHBITDEPTH
+                                int use_hbd, int bd,
+#endif  // CONFIG_HIGHBITDEPTH
+                                uint8_t *ref, int r_width, int r_height,
+                                int r_stride, uint8_t *dst, int d_width,
+                                int d_height, int d_stride, int n_refinements);
+
+/*
+  Computes "num_motions" candidate global motion parameters between two frames.
+  The array "params_by_motion" should be length 8 * "num_motions". The ordering
+  of each set of parameters is best described  by the homography:
+
+        [x'     (m2 m3 m0   [x
+    z .  y'  =   m4 m5 m1 *  y
+         1]      m6 m7 1)    1]
+
+  where m{i} represents the ith value in any given set of parameters.
+
+  "num_inliers" should be length "num_motions", and will be populated with the
+  number of inlier feature points for each motion. Params for which the
+  num_inliers entry is 0 should be ignored by the caller.
+*/
+int compute_global_motion_feature_based(
+    TransformationType type, YV12_BUFFER_CONFIG *frm, YV12_BUFFER_CONFIG *ref,
+#if CONFIG_HIGHBITDEPTH
+    int bit_depth,
+#endif
+    int *num_inliers_by_motion, double *params_by_motion, int num_motions);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+#endif  // AV1_ENCODER_GLOBAL_MOTION_H_
diff --git a/third_party/aom/av1/encoder/hybrid_fwd_txfm.c b/third_party/aom/av1/encoder/hybrid_fwd_txfm.c
new file mode 100644
index 0000000000..4fd5631632
--- /dev/null
+++ b/third_party/aom/av1/encoder/hybrid_fwd_txfm.c
@@ -0,0 +1,499 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "av1/common/idct.h"
+#include "av1/encoder/hybrid_fwd_txfm.h"
+
+#if CONFIG_CB4X4
+static void fwd_txfm_2x2(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, TX_TYPE tx_type, int lossless) {
+  tran_high_t a1 = src_diff[0];
+  tran_high_t b1 = src_diff[1];
+  tran_high_t c1 = src_diff[diff_stride];
+  tran_high_t d1 = src_diff[1 + diff_stride];
+
+  tran_high_t a2 = a1 + c1;
+  tran_high_t b2 = b1 + d1;
+  tran_high_t c2 = a1 - c1;
+  tran_high_t d2 = b1 - d1;
+
+  a1 = a2 + b2;
+  b1 = a2 - b2;
+  c1 = c2 + d2;
+  d1 = c2 - d2;
+
+  coeff[0] = (tran_low_t)(4 * a1);
+  coeff[1] = (tran_low_t)(4 * b1);
+  coeff[2] = (tran_low_t)(4 * c1);
+  coeff[3] = (tran_low_t)(4 * d1);
+
+  (void)tx_type;
+  (void)lossless;
+}
+#endif
+
+static void fwd_txfm_4x4(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, TX_TYPE tx_type, int lossless) {
+  if (lossless) {
+    assert(tx_type == DCT_DCT);
+    av1_fwht4x4(src_diff, coeff, diff_stride);
+    return;
+  }
+
+  av1_fht4x4(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_4x8(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, TX_TYPE tx_type,
+                         FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht4x8(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_8x4(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, TX_TYPE tx_type,
+                         FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht8x4(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_8x16(const int16_t *src_diff, tran_low_t *coeff,
+                          int diff_stride, TX_TYPE tx_type,
+                          FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht8x16(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_16x8(const int16_t *src_diff, tran_low_t *coeff,
+                          int diff_stride, TX_TYPE tx_type,
+                          FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht16x8(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_16x32(const int16_t *src_diff, tran_low_t *coeff,
+                           int diff_stride, TX_TYPE tx_type,
+                           FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht16x32(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_32x16(const int16_t *src_diff, tran_low_t *coeff,
+                           int diff_stride, TX_TYPE tx_type,
+                           FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht32x16(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_8x8(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, TX_TYPE tx_type,
+                         FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht8x8(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_16x16(const int16_t *src_diff, tran_low_t *coeff,
+                           int diff_stride, TX_TYPE tx_type,
+                           FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht16x16(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void fwd_txfm_32x32(const int16_t *src_diff, tran_low_t *coeff,
+                           int diff_stride, TX_TYPE tx_type,
+                           FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+  av1_fht32x32(src_diff, coeff, diff_stride, tx_type);
+}
+
+#if CONFIG_TX64X64
+static void fwd_txfm_64x64(const int16_t *src_diff, tran_low_t *coeff,
+                           int diff_stride, TX_TYPE tx_type,
+                           FWD_TXFM_OPT fwd_txfm_opt) {
+  (void)fwd_txfm_opt;
+#if CONFIG_EXT_TX
+  if (tx_type == IDTX)
+    av1_fwd_idtx_c(src_diff, coeff, diff_stride, 64, tx_type);
+  else
+#endif
+    av1_fht64x64(src_diff, coeff, diff_stride, tx_type);
+}
+#endif  // CONFIG_TX64X64
+
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_CB4X4
+static void highbd_fwd_txfm_2x2(const int16_t *src_diff, tran_low_t *coeff,
+                                int diff_stride, TX_TYPE tx_type, int lossless,
+                                const int bd) {
+  tran_high_t a1 = src_diff[0];
+  tran_high_t b1 = src_diff[1];
+  tran_high_t c1 = src_diff[diff_stride];
+  tran_high_t d1 = src_diff[1 + diff_stride];
+
+  tran_high_t a2 = a1 + c1;
+  tran_high_t b2 = b1 + d1;
+  tran_high_t c2 = a1 - c1;
+  tran_high_t d2 = b1 - d1;
+
+  a1 = a2 + b2;
+  b1 = a2 - b2;
+  c1 = c2 + d2;
+  d1 = c2 - d2;
+
+  coeff[0] = (tran_low_t)(4 * a1);
+  coeff[1] = (tran_low_t)(4 * b1);
+  coeff[2] = (tran_low_t)(4 * c1);
+  coeff[3] = (tran_low_t)(4 * d1);
+
+  (void)tx_type;
+  (void)lossless;
+  (void)bd;
+}
+#endif
+
+static void highbd_fwd_txfm_4x4(const int16_t *src_diff, tran_low_t *coeff,
+                                int diff_stride, TX_TYPE tx_type, int lossless,
+                                const int bd) {
+  if (lossless) {
+    assert(tx_type == DCT_DCT);
+    av1_highbd_fwht4x4(src_diff, coeff, diff_stride);
+    return;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+      av1_fwd_txfm2d_4x4(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+      av1_fwd_txfm2d_4x4(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      av1_highbd_fht4x4_c(src_diff, coeff, diff_stride, tx_type);
+      break;
+    case IDTX: av1_fwd_idtx_c(src_diff, coeff, diff_stride, 4, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+}
+
+static void highbd_fwd_txfm_4x8(const int16_t *src_diff, tran_low_t *coeff,
+                                int diff_stride, TX_TYPE tx_type,
+                                FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht4x8(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_8x4(const int16_t *src_diff, tran_low_t *coeff,
+                                int diff_stride, TX_TYPE tx_type,
+                                FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht8x4(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_8x16(const int16_t *src_diff, tran_low_t *coeff,
+                                 int diff_stride, TX_TYPE tx_type,
+                                 FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht8x16(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_16x8(const int16_t *src_diff, tran_low_t *coeff,
+                                 int diff_stride, TX_TYPE tx_type,
+                                 FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht16x8(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_16x32(const int16_t *src_diff, tran_low_t *coeff,
+                                  int diff_stride, TX_TYPE tx_type,
+                                  FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht16x32(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_32x16(const int16_t *src_diff, tran_low_t *coeff,
+                                  int diff_stride, TX_TYPE tx_type,
+                                  FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  av1_highbd_fht32x16(src_diff, coeff, diff_stride, tx_type);
+}
+
+static void highbd_fwd_txfm_8x8(const int16_t *src_diff, tran_low_t *coeff,
+                                int diff_stride, TX_TYPE tx_type,
+                                FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+      av1_fwd_txfm2d_8x8(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+      av1_fwd_txfm2d_8x8(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      // Use C version since DST exists only in C
+      av1_highbd_fht8x8_c(src_diff, coeff, diff_stride, tx_type);
+      break;
+    case IDTX: av1_fwd_idtx_c(src_diff, coeff, diff_stride, 8, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+}
+
+static void highbd_fwd_txfm_16x16(const int16_t *src_diff, tran_low_t *coeff,
+                                  int diff_stride, TX_TYPE tx_type,
+                                  FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+      av1_fwd_txfm2d_16x16(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+      av1_fwd_txfm2d_16x16(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      // Use C version since DST exists only in C
+      av1_highbd_fht16x16_c(src_diff, coeff, diff_stride, tx_type);
+      break;
+    case IDTX: av1_fwd_idtx_c(src_diff, coeff, diff_stride, 16, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+}
+
+static void highbd_fwd_txfm_32x32(const int16_t *src_diff, tran_low_t *coeff,
+                                  int diff_stride, TX_TYPE tx_type,
+                                  FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  switch (tx_type) {
+    case DCT_DCT:
+      av1_fwd_txfm2d_32x32(src_diff, coeff, diff_stride, tx_type, bd);
+      break;
+#if CONFIG_EXT_TX
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      av1_highbd_fht32x32_c(src_diff, coeff, diff_stride, tx_type);
+      break;
+    case IDTX: av1_fwd_idtx_c(src_diff, coeff, diff_stride, 32, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+#if CONFIG_TX64X64
+static void highbd_fwd_txfm_64x64(const int16_t *src_diff, tran_low_t *coeff,
+                                  int diff_stride, TX_TYPE tx_type,
+                                  FWD_TXFM_OPT fwd_txfm_opt, const int bd) {
+  (void)fwd_txfm_opt;
+  (void)bd;
+  switch (tx_type) {
+    case DCT_DCT:
+      av1_highbd_fht64x64(src_diff, coeff, diff_stride, tx_type);
+      break;
+#if CONFIG_EXT_TX
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+      av1_highbd_fht64x64(src_diff, coeff, diff_stride, tx_type);
+      break;
+    case IDTX: av1_fwd_idtx_c(src_diff, coeff, diff_stride, 64, tx_type); break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+#endif  // CONFIG_TX64X64
+#endif  // CONFIG_HIGHBITDEPTH
+
+void av1_fwd_txfm(const int16_t *src_diff, tran_low_t *coeff, int diff_stride,
+                  FWD_TXFM_PARAM *fwd_txfm_param) {
+  const int fwd_txfm_opt = FWD_TXFM_OPT_NORMAL;
+  const TX_TYPE tx_type = fwd_txfm_param->tx_type;
+  const TX_SIZE tx_size = fwd_txfm_param->tx_size;
+  const int lossless = fwd_txfm_param->lossless;
+  switch (tx_size) {
+#if CONFIG_TX64X64
+    case TX_64X64:
+      fwd_txfm_64x64(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+#endif  // CONFIG_TX64X64
+    case TX_32X32:
+      fwd_txfm_32x32(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_16X16:
+      fwd_txfm_16x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_8X8:
+      fwd_txfm_8x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_4X8:
+      fwd_txfm_4x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_8X4:
+      fwd_txfm_8x4(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_8X16:
+      fwd_txfm_8x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_16X8:
+      fwd_txfm_16x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_16X32:
+      fwd_txfm_16x32(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_32X16:
+      fwd_txfm_32x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt);
+      break;
+    case TX_4X4:
+      fwd_txfm_4x4(src_diff, coeff, diff_stride, tx_type, lossless);
+      break;
+#if CONFIG_CB4X4
+    case TX_2X2:
+      fwd_txfm_2x2(src_diff, coeff, diff_stride, tx_type, lossless);
+      break;
+#endif
+    default: assert(0); break;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_fwd_txfm(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, FWD_TXFM_PARAM *fwd_txfm_param) {
+  const int fwd_txfm_opt = FWD_TXFM_OPT_NORMAL;
+  const TX_TYPE tx_type = fwd_txfm_param->tx_type;
+  const TX_SIZE tx_size = fwd_txfm_param->tx_size;
+  const int lossless = fwd_txfm_param->lossless;
+  const int bd = fwd_txfm_param->bd;
+  switch (tx_size) {
+#if CONFIG_TX64X64
+    case TX_64X64:
+      highbd_fwd_txfm_64x64(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                            bd);
+      break;
+#endif  // CONFIG_TX64X64
+    case TX_32X32:
+      highbd_fwd_txfm_32x32(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                            bd);
+      break;
+    case TX_16X16:
+      highbd_fwd_txfm_16x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                            bd);
+      break;
+    case TX_8X8:
+      highbd_fwd_txfm_8x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                          bd);
+      break;
+    case TX_4X8:
+      highbd_fwd_txfm_4x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                          bd);
+      break;
+    case TX_8X4:
+      highbd_fwd_txfm_8x4(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                          bd);
+      break;
+    case TX_8X16:
+      highbd_fwd_txfm_8x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                           bd);
+      break;
+    case TX_16X8:
+      highbd_fwd_txfm_16x8(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                           bd);
+      break;
+    case TX_16X32:
+      highbd_fwd_txfm_16x32(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                            bd);
+      break;
+    case TX_32X16:
+      highbd_fwd_txfm_32x16(src_diff, coeff, diff_stride, tx_type, fwd_txfm_opt,
+                            bd);
+      break;
+    case TX_4X4:
+      highbd_fwd_txfm_4x4(src_diff, coeff, diff_stride, tx_type, lossless, bd);
+      break;
+#if CONFIG_CB4X4
+    case TX_2X2:
+      highbd_fwd_txfm_2x2(src_diff, coeff, diff_stride, tx_type, lossless, bd);
+      break;
+#endif
+    default: assert(0); break;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/av1/encoder/hybrid_fwd_txfm.h b/third_party/aom/av1/encoder/hybrid_fwd_txfm.h
new file mode 100644
index 0000000000..e6fd17275e
--- /dev/null
+++ b/third_party/aom/av1/encoder/hybrid_fwd_txfm.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_HYBRID_FWD_TXFM_H_
+#define AV1_ENCODER_HYBRID_FWD_TXFM_H_
+
+#include "./aom_config.h"
+
+typedef enum FWD_TXFM_OPT { FWD_TXFM_OPT_NORMAL } FWD_TXFM_OPT;
+
+typedef struct FWD_TXFM_PARAM {
+  TX_TYPE tx_type;
+  TX_SIZE tx_size;
+  int lossless;
+#if CONFIG_HIGHBITDEPTH
+  int bd;
+#endif  // CONFIG_HIGHBITDEPTH
+} FWD_TXFM_PARAM;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_fwd_txfm(const int16_t *src_diff, tran_low_t *coeff, int diff_stride,
+                  FWD_TXFM_PARAM *fwd_txfm_param);
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_fwd_txfm(const int16_t *src_diff, tran_low_t *coeff,
+                         int diff_stride, FWD_TXFM_PARAM *fwd_txfm_param);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_HYBRID_FWD_TXFM_H_
diff --git a/third_party/aom/av1/encoder/laplace_encoder.c b/third_party/aom/av1/encoder/laplace_encoder.c
new file mode 100644
index 0000000000..54ffc88fb5
--- /dev/null
+++ b/third_party/aom/av1/encoder/laplace_encoder.c
@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <stdio.h>
+
+#include "aom_dsp/bitwriter.h"
+#include "av1/common/odintrin.h"
+#include "av1/common/pvq.h"
+#include "pvq_encoder.h"
+
+static void aom_encode_pvq_split(aom_writer *w, od_pvq_codeword_ctx *adapt,
+ int count, int sum, int ctx) {
+  int shift;
+  int rest;
+  int fctx;
+  if (sum == 0) return;
+  shift = OD_MAXI(0, OD_ILOG(sum) - 3);
+  if (shift) {
+    rest = count & ((1 << shift) - 1);
+    count >>= shift;
+    sum >>= shift;
+  }
+  fctx = 7*ctx + sum - 1;
+  aom_write_symbol_pvq(w, count, adapt->pvq_split_cdf[fctx], sum + 1);
+  if (shift) aom_write_literal(w, rest, shift);
+}
+
+void aom_encode_band_pvq_splits(aom_writer *w, od_pvq_codeword_ctx *adapt,
+ const int *y, int n, int k, int level) {
+  int mid;
+  int i;
+  int count_right;
+  if (n <= 1 || k == 0) return;
+  if (k == 1 && n <= 16) {
+    int cdf_id;
+    int pos;
+    cdf_id = od_pvq_k1_ctx(n, level == 0);
+    for (pos = 0; !y[pos]; pos++);
+    OD_ASSERT(pos < n);
+    aom_write_symbol_pvq(w, pos, adapt->pvq_k1_cdf[cdf_id], n);
+  }
+  else {
+    mid = n >> 1;
+    count_right = k;
+    for (i = 0; i < mid; i++) count_right -= abs(y[i]);
+    aom_encode_pvq_split(w, adapt, count_right, k, od_pvq_size_ctx(n));
+    aom_encode_band_pvq_splits(w, adapt, y, mid, k - count_right, level + 1);
+    aom_encode_band_pvq_splits(w, adapt, y + mid, n - mid, count_right,
+     level + 1);
+  }
+}
+
+/** Encodes the tail of a Laplace-distributed variable, i.e. it doesn't
+ * do anything special for the zero case.
+ *
+ * @param [in,out] enc     range encoder
+ * @param [in]     x       variable to encode (has to be positive)
+ * @param [in]     decay   decay factor of the distribution in Q8 format,
+ * i.e. pdf ~= decay^x
+ */
+void aom_laplace_encode_special(aom_writer *w, int x, unsigned decay) {
+  int shift;
+  int xs;
+  int sym;
+  const uint16_t *cdf;
+  shift = 0;
+  /* We don't want a large decay value because that would require too many
+     symbols. */
+  while (decay > 235) {
+    decay = (decay*decay + 128) >> 8;
+    shift++;
+  }
+  decay = OD_MINI(decay, 254);
+  decay = OD_MAXI(decay, 2);
+  xs = x >> shift;
+  cdf = EXP_CDF_TABLE[(decay + 1) >> 1];
+  OD_LOG((OD_LOG_PVQ, OD_LOG_DEBUG, "decay = %d", decay));
+  do {
+    sym = OD_MINI(xs, 15);
+    {
+      int i;
+      OD_LOG((OD_LOG_PVQ, OD_LOG_DEBUG, "%d %d %d %d %d\n", x, xs, shift,
+       sym, max));
+      for (i = 0; i < 16; i++) {
+        OD_LOG_PARTIAL((OD_LOG_PVQ, OD_LOG_DEBUG, "%d ", cdf[i]));
+      }
+      OD_LOG_PARTIAL((OD_LOG_PVQ, OD_LOG_DEBUG, "\n"));
+    }
+    aom_write_cdf(w, sym, cdf, 16);
+    xs -= 15;
+  } while (sym >= 15);
+  if (shift) aom_write_literal(w, x & ((1 << shift) - 1), shift);
+}
diff --git a/third_party/aom/av1/encoder/lookahead.c b/third_party/aom/av1/encoder/lookahead.c
new file mode 100644
index 0000000000..591ca61521
--- /dev/null
+++ b/third_party/aom/av1/encoder/lookahead.c
@@ -0,0 +1,225 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <assert.h>
+#include <stdlib.h>
+
+#include "./aom_config.h"
+
+#include "av1/common/common.h"
+
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/extend.h"
+#include "av1/encoder/lookahead.h"
+
+/* Return the buffer at the given absolute index and increment the index */
+static struct lookahead_entry *pop(struct lookahead_ctx *ctx, int *idx) {
+  int index = *idx;
+  struct lookahead_entry *buf = ctx->buf + index;
+
+  assert(index < ctx->max_sz);
+  if (++index >= ctx->max_sz) index -= ctx->max_sz;
+  *idx = index;
+  return buf;
+}
+
+void av1_lookahead_destroy(struct lookahead_ctx *ctx) {
+  if (ctx) {
+    if (ctx->buf) {
+      int i;
+
+      for (i = 0; i < ctx->max_sz; i++) aom_free_frame_buffer(&ctx->buf[i].img);
+      free(ctx->buf);
+    }
+    free(ctx);
+  }
+}
+
+struct lookahead_ctx *av1_lookahead_init(unsigned int width,
+                                         unsigned int height,
+                                         unsigned int subsampling_x,
+                                         unsigned int subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                         int use_highbitdepth,
+#endif
+                                         unsigned int depth) {
+  struct lookahead_ctx *ctx = NULL;
+
+  // Clamp the lookahead queue depth
+  depth = clamp(depth, 1, MAX_LAG_BUFFERS);
+
+  // Allocate memory to keep previous source frames available.
+  depth += MAX_PRE_FRAMES;
+
+  // Allocate the lookahead structures
+  ctx = calloc(1, sizeof(*ctx));
+  if (ctx) {
+    const int legacy_byte_alignment = 0;
+    unsigned int i;
+    ctx->max_sz = depth;
+    ctx->buf = calloc(depth, sizeof(*ctx->buf));
+    if (!ctx->buf) goto bail;
+    for (i = 0; i < depth; i++)
+      if (aom_alloc_frame_buffer(&ctx->buf[i].img, width, height, subsampling_x,
+                                 subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                 use_highbitdepth,
+#endif
+                                 AOM_BORDER_IN_PIXELS, legacy_byte_alignment))
+        goto bail;
+  }
+  return ctx;
+bail:
+  av1_lookahead_destroy(ctx);
+  return NULL;
+}
+
+#define USE_PARTIAL_COPY 0
+
+int av1_lookahead_push(struct lookahead_ctx *ctx, YV12_BUFFER_CONFIG *src,
+                       int64_t ts_start, int64_t ts_end,
+#if CONFIG_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       aom_enc_frame_flags_t flags) {
+  struct lookahead_entry *buf;
+#if USE_PARTIAL_COPY
+  int row, col, active_end;
+  int mb_rows = (src->y_height + 15) >> 4;
+  int mb_cols = (src->y_width + 15) >> 4;
+#endif
+  int width = src->y_crop_width;
+  int height = src->y_crop_height;
+  int uv_width = src->uv_crop_width;
+  int uv_height = src->uv_crop_height;
+  int subsampling_x = src->subsampling_x;
+  int subsampling_y = src->subsampling_y;
+  int larger_dimensions, new_dimensions;
+
+  if (ctx->sz + 1 + MAX_PRE_FRAMES > ctx->max_sz) return 1;
+  ctx->sz++;
+  buf = pop(ctx, &ctx->write_idx);
+
+  new_dimensions = width != buf->img.y_crop_width ||
+                   height != buf->img.y_crop_height ||
+                   uv_width != buf->img.uv_crop_width ||
+                   uv_height != buf->img.uv_crop_height;
+  larger_dimensions = width > buf->img.y_width || height > buf->img.y_height ||
+                      uv_width > buf->img.uv_width ||
+                      uv_height > buf->img.uv_height;
+  assert(!larger_dimensions || new_dimensions);
+
+#if USE_PARTIAL_COPY
+  // TODO(jkoleszar): This is disabled for now, as
+  // av1_copy_and_extend_frame_with_rect is not subsampling/alpha aware.
+
+  // Only do this partial copy if the following conditions are all met:
+  // 1. Lookahead queue has has size of 1.
+  // 2. Active map is provided.
+  // 3. This is not a key frame, golden nor altref frame.
+  if (!new_dimensions && ctx->max_sz == 1 && active_map && !flags) {
+    for (row = 0; row < mb_rows; ++row) {
+      col = 0;
+
+      while (1) {
+        // Find the first active macroblock in this row.
+        for (; col < mb_cols; ++col) {
+          if (active_map[col]) break;
+        }
+
+        // No more active macroblock in this row.
+        if (col == mb_cols) break;
+
+        // Find the end of active region in this row.
+        active_end = col;
+
+        for (; active_end < mb_cols; ++active_end) {
+          if (!active_map[active_end]) break;
+        }
+
+        // Only copy this active region.
+        av1_copy_and_extend_frame_with_rect(src, &buf->img, row << 4, col << 4,
+                                            16, (active_end - col) << 4);
+
+        // Start again from the end of this active region.
+        col = active_end;
+      }
+
+      active_map += mb_cols;
+    }
+  } else {
+#endif
+    if (larger_dimensions) {
+      YV12_BUFFER_CONFIG new_img;
+      memset(&new_img, 0, sizeof(new_img));
+      if (aom_alloc_frame_buffer(&new_img, width, height, subsampling_x,
+                                 subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                 use_highbitdepth,
+#endif
+                                 AOM_BORDER_IN_PIXELS, 0))
+        return 1;
+      aom_free_frame_buffer(&buf->img);
+      buf->img = new_img;
+    } else if (new_dimensions) {
+      buf->img.y_crop_width = src->y_crop_width;
+      buf->img.y_crop_height = src->y_crop_height;
+      buf->img.uv_crop_width = src->uv_crop_width;
+      buf->img.uv_crop_height = src->uv_crop_height;
+      buf->img.subsampling_x = src->subsampling_x;
+      buf->img.subsampling_y = src->subsampling_y;
+    }
+    // Partial copy not implemented yet
+    av1_copy_and_extend_frame(src, &buf->img);
+#if USE_PARTIAL_COPY
+  }
+#endif
+
+  buf->ts_start = ts_start;
+  buf->ts_end = ts_end;
+  buf->flags = flags;
+  return 0;
+}
+
+struct lookahead_entry *av1_lookahead_pop(struct lookahead_ctx *ctx,
+                                          int drain) {
+  struct lookahead_entry *buf = NULL;
+
+  if (ctx && ctx->sz && (drain || ctx->sz == ctx->max_sz - MAX_PRE_FRAMES)) {
+    buf = pop(ctx, &ctx->read_idx);
+    ctx->sz--;
+  }
+  return buf;
+}
+
+struct lookahead_entry *av1_lookahead_peek(struct lookahead_ctx *ctx,
+                                           int index) {
+  struct lookahead_entry *buf = NULL;
+
+  if (index >= 0) {
+    // Forward peek
+    if (index < ctx->sz) {
+      index += ctx->read_idx;
+      if (index >= ctx->max_sz) index -= ctx->max_sz;
+      buf = ctx->buf + index;
+    }
+  } else if (index < 0) {
+    // Backward peek
+    if (-index <= MAX_PRE_FRAMES) {
+      index += (int)(ctx->read_idx);
+      if (index < 0) index += (int)(ctx->max_sz);
+      buf = ctx->buf + index;
+    }
+  }
+
+  return buf;
+}
+
+unsigned int av1_lookahead_depth(struct lookahead_ctx *ctx) { return ctx->sz; }
diff --git a/third_party/aom/av1/encoder/lookahead.h b/third_party/aom/av1/encoder/lookahead.h
new file mode 100644
index 0000000000..19f75d7e45
--- /dev/null
+++ b/third_party/aom/av1/encoder/lookahead.h
@@ -0,0 +1,114 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_LOOKAHEAD_H_
+#define AV1_ENCODER_LOOKAHEAD_H_
+
+#include "aom_scale/yv12config.h"
+#include "aom/aom_integer.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define MAX_LAG_BUFFERS 25
+
+struct lookahead_entry {
+  YV12_BUFFER_CONFIG img;
+  int64_t ts_start;
+  int64_t ts_end;
+  aom_enc_frame_flags_t flags;
+};
+
+// The max of past frames we want to keep in the queue.
+#define MAX_PRE_FRAMES 1
+
+struct lookahead_ctx {
+  int max_sz;                  /* Absolute size of the queue */
+  int sz;                      /* Number of buffers currently in the queue */
+  int read_idx;                /* Read index */
+  int write_idx;               /* Write index */
+  struct lookahead_entry *buf; /* Buffer list */
+};
+
+/**\brief Initializes the lookahead stage
+ *
+ * The lookahead stage is a queue of frame buffers on which some analysis
+ * may be done when buffers are enqueued.
+ */
+struct lookahead_ctx *av1_lookahead_init(unsigned int width,
+                                         unsigned int height,
+                                         unsigned int subsampling_x,
+                                         unsigned int subsampling_y,
+#if CONFIG_HIGHBITDEPTH
+                                         int use_highbitdepth,
+#endif
+                                         unsigned int depth);
+
+/**\brief Destroys the lookahead stage
+ */
+void av1_lookahead_destroy(struct lookahead_ctx *ctx);
+
+/**\brief Enqueue a source buffer
+ *
+ * This function will copy the source image into a new framebuffer with
+ * the expected stride/border.
+ *
+ * If active_map is non-NULL and there is only one frame in the queue, then copy
+ * only active macroblocks.
+ *
+ * \param[in] ctx         Pointer to the lookahead context
+ * \param[in] src         Pointer to the image to enqueue
+ * \param[in] ts_start    Timestamp for the start of this frame
+ * \param[in] ts_end      Timestamp for the end of this frame
+ * \param[in] flags       Flags set on this frame
+ * \param[in] active_map  Map that specifies which macroblock is active
+ */
+int av1_lookahead_push(struct lookahead_ctx *ctx, YV12_BUFFER_CONFIG *src,
+                       int64_t ts_start, int64_t ts_end,
+#if CONFIG_HIGHBITDEPTH
+                       int use_highbitdepth,
+#endif
+                       aom_enc_frame_flags_t flags);
+
+/**\brief Get the next source buffer to encode
+ *
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ * \param[in] drain     Flag indicating the buffer should be drained
+ *                      (return a buffer regardless of the current queue depth)
+ *
+ * \retval NULL, if drain set and queue is empty
+ * \retval NULL, if drain not set and queue not of the configured depth
+ */
+struct lookahead_entry *av1_lookahead_pop(struct lookahead_ctx *ctx, int drain);
+
+/**\brief Get a future source buffer to encode
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ * \param[in] index     Index of the frame to be returned, 0 == next frame
+ *
+ * \retval NULL, if no buffer exists at the specified index
+ */
+struct lookahead_entry *av1_lookahead_peek(struct lookahead_ctx *ctx,
+                                           int index);
+
+/**\brief Get the number of frames currently in the lookahead queue
+ *
+ * \param[in] ctx       Pointer to the lookahead context
+ */
+unsigned int av1_lookahead_depth(struct lookahead_ctx *ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_LOOKAHEAD_H_
diff --git a/third_party/aom/av1/encoder/mbgraph.c b/third_party/aom/av1/encoder/mbgraph.c
new file mode 100644
index 0000000000..1296027dcb
--- /dev/null
+++ b/third_party/aom/av1/encoder/mbgraph.c
@@ -0,0 +1,398 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/system_state.h"
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/mcomp.h"
+#include "av1/common/blockd.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+
+static unsigned int do_16x16_motion_iteration(AV1_COMP *cpi, const MV *ref_mv,
+                                              int mb_row, int mb_col) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
+  const aom_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[BLOCK_16X16];
+
+  const MvLimits tmp_mv_limits = x->mv_limits;
+  MV ref_full;
+  int cost_list[5];
+
+  // Further step/diamond searches as necessary
+  int step_param = mv_sf->reduce_first_step_size;
+  step_param = AOMMIN(step_param, MAX_MVSEARCH_STEPS - 2);
+
+  av1_set_mv_search_range(&x->mv_limits, ref_mv);
+
+  ref_full.col = ref_mv->col >> 3;
+  ref_full.row = ref_mv->row >> 3;
+
+  /*cpi->sf.search_method == HEX*/
+  av1_hex_search(x, &ref_full, step_param, x->errorperbit, 0,
+                 cond_cost_list(cpi, cost_list), &v_fn_ptr, 0, ref_mv);
+
+  // Try sub-pixel MC
+  // if (bestsme > error_thresh && bestsme < INT_MAX)
+  {
+    int distortion;
+    unsigned int sse;
+    cpi->find_fractional_mv_step(x, ref_mv, cpi->common.allow_high_precision_mv,
+                                 x->errorperbit, &v_fn_ptr, 0,
+                                 mv_sf->subpel_iters_per_step,
+                                 cond_cost_list(cpi, cost_list), NULL, NULL,
+                                 &distortion, &sse, NULL, 0, 0, 0);
+  }
+
+#if CONFIG_EXT_INTER
+  if (has_second_ref(&xd->mi[0]->mbmi))
+    xd->mi[0]->mbmi.mode = NEW_NEWMV;
+  else
+#endif  // CONFIG_EXT_INTER
+    xd->mi[0]->mbmi.mode = NEWMV;
+
+  xd->mi[0]->mbmi.mv[0] = x->best_mv;
+#if CONFIG_EXT_INTER
+  xd->mi[0]->mbmi.ref_frame[1] = NONE_FRAME;
+#endif  // CONFIG_EXT_INTER
+
+  av1_build_inter_predictors_sby(xd, mb_row, mb_col, NULL, BLOCK_16X16);
+
+  /* restore UMV window */
+  x->mv_limits = tmp_mv_limits;
+
+  return aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                      xd->plane[0].dst.buf, xd->plane[0].dst.stride);
+}
+
+static int do_16x16_motion_search(AV1_COMP *cpi, const MV *ref_mv, int mb_row,
+                                  int mb_col) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  unsigned int err, tmp_err;
+  MV best_mv;
+
+  // Try zero MV first
+  // FIXME should really use something like near/nearest MV and/or MV prediction
+  err = aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
+  best_mv.col = best_mv.row = 0;
+
+  // Test last reference frame using the previous best mv as the
+  // starting point (best reference) for the search
+  tmp_err = do_16x16_motion_iteration(cpi, ref_mv, mb_row, mb_col);
+  if (tmp_err < err) {
+    err = tmp_err;
+    best_mv = x->best_mv.as_mv;
+  }
+
+  // If the current best reference mv is not centered on 0,0 then do a 0,0
+  // based search as well.
+  if (ref_mv->row != 0 || ref_mv->col != 0) {
+    MV zero_ref_mv = { 0, 0 };
+
+    tmp_err = do_16x16_motion_iteration(cpi, &zero_ref_mv, mb_row, mb_col);
+    if (tmp_err < err) {
+      err = tmp_err;
+      best_mv = x->best_mv.as_mv;
+    }
+  }
+
+  x->best_mv.as_mv = best_mv;
+  return err;
+}
+
+static int do_16x16_zerozero_search(AV1_COMP *cpi, int_mv *dst_mv) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  unsigned int err;
+
+  // Try zero MV first
+  // FIXME should really use something like near/nearest MV and/or MV prediction
+  err = aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                     xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride);
+
+  dst_mv->as_int = 0;
+
+  return err;
+}
+static int find_best_16x16_intra(AV1_COMP *cpi, PREDICTION_MODE *pbest_mode) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  PREDICTION_MODE best_mode = -1, mode;
+  unsigned int best_err = INT_MAX;
+
+  // calculate SATD for each intra prediction mode;
+  // we're intentionally not doing 4x4, we just want a rough estimate
+  for (mode = DC_PRED; mode <= TM_PRED; mode++) {
+    unsigned int err;
+
+    xd->mi[0]->mbmi.mode = mode;
+    av1_predict_intra_block(xd, 16, 16, BLOCK_16X16, mode, x->plane[0].src.buf,
+                            x->plane[0].src.stride, xd->plane[0].dst.buf,
+                            xd->plane[0].dst.stride, 0, 0, 0);
+    err = aom_sad16x16(x->plane[0].src.buf, x->plane[0].src.stride,
+                       xd->plane[0].dst.buf, xd->plane[0].dst.stride);
+
+    // find best
+    if (err < best_err) {
+      best_err = err;
+      best_mode = mode;
+    }
+  }
+
+  if (pbest_mode) *pbest_mode = best_mode;
+
+  return best_err;
+}
+
+static void update_mbgraph_mb_stats(AV1_COMP *cpi, MBGRAPH_MB_STATS *stats,
+                                    YV12_BUFFER_CONFIG *buf, int mb_y_offset,
+                                    YV12_BUFFER_CONFIG *golden_ref,
+                                    const MV *prev_golden_ref_mv,
+                                    YV12_BUFFER_CONFIG *alt_ref, int mb_row,
+                                    int mb_col) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int intra_error;
+  AV1_COMMON *cm = &cpi->common;
+
+  // FIXME in practice we're completely ignoring chroma here
+  x->plane[0].src.buf = buf->y_buffer + mb_y_offset;
+  x->plane[0].src.stride = buf->y_stride;
+
+  xd->plane[0].dst.buf = get_frame_new_buffer(cm)->y_buffer + mb_y_offset;
+  xd->plane[0].dst.stride = get_frame_new_buffer(cm)->y_stride;
+
+  // do intra 16x16 prediction
+  intra_error = find_best_16x16_intra(cpi, &stats->ref[INTRA_FRAME].m.mode);
+  if (intra_error <= 0) intra_error = 1;
+  stats->ref[INTRA_FRAME].err = intra_error;
+
+  // Golden frame MV search, if it exists and is different than last frame
+  if (golden_ref) {
+    int g_motion_error;
+    xd->plane[0].pre[0].buf = golden_ref->y_buffer + mb_y_offset;
+    xd->plane[0].pre[0].stride = golden_ref->y_stride;
+    g_motion_error =
+        do_16x16_motion_search(cpi, prev_golden_ref_mv, mb_row, mb_col);
+    stats->ref[GOLDEN_FRAME].m.mv = x->best_mv;
+    stats->ref[GOLDEN_FRAME].err = g_motion_error;
+  } else {
+    stats->ref[GOLDEN_FRAME].err = INT_MAX;
+    stats->ref[GOLDEN_FRAME].m.mv.as_int = 0;
+  }
+
+  // Do an Alt-ref frame MV search, if it exists and is different than
+  // last/golden frame.
+  if (alt_ref) {
+    int a_motion_error;
+    xd->plane[0].pre[0].buf = alt_ref->y_buffer + mb_y_offset;
+    xd->plane[0].pre[0].stride = alt_ref->y_stride;
+    a_motion_error =
+        do_16x16_zerozero_search(cpi, &stats->ref[ALTREF_FRAME].m.mv);
+
+    stats->ref[ALTREF_FRAME].err = a_motion_error;
+  } else {
+    stats->ref[ALTREF_FRAME].err = INT_MAX;
+    stats->ref[ALTREF_FRAME].m.mv.as_int = 0;
+  }
+}
+
+static void update_mbgraph_frame_stats(AV1_COMP *cpi,
+                                       MBGRAPH_FRAME_STATS *stats,
+                                       YV12_BUFFER_CONFIG *buf,
+                                       YV12_BUFFER_CONFIG *golden_ref,
+                                       YV12_BUFFER_CONFIG *alt_ref) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  AV1_COMMON *const cm = &cpi->common;
+
+  int mb_col, mb_row, offset = 0;
+  int mb_y_offset = 0, arf_y_offset = 0, gld_y_offset = 0;
+  MV gld_top_mv = { 0, 0 };
+  MODE_INFO mi_local;
+
+  av1_zero(mi_local);
+  // Set up limit values for motion vectors to prevent them extending outside
+  // the UMV borders.
+  x->mv_limits.row_min = -BORDER_MV_PIXELS_B16;
+  x->mv_limits.row_max = (cm->mb_rows - 1) * 8 + BORDER_MV_PIXELS_B16;
+  xd->up_available = 0;
+  xd->plane[0].dst.stride = buf->y_stride;
+  xd->plane[0].pre[0].stride = buf->y_stride;
+  xd->plane[1].dst.stride = buf->uv_stride;
+  xd->mi[0] = &mi_local;
+  mi_local.mbmi.sb_type = BLOCK_16X16;
+  mi_local.mbmi.ref_frame[0] = LAST_FRAME;
+  mi_local.mbmi.ref_frame[1] = NONE_FRAME;
+
+  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
+    MV gld_left_mv = gld_top_mv;
+    int mb_y_in_offset = mb_y_offset;
+    int arf_y_in_offset = arf_y_offset;
+    int gld_y_in_offset = gld_y_offset;
+
+    // Set up limit values for motion vectors to prevent them extending outside
+    // the UMV borders.
+    x->mv_limits.col_min = -BORDER_MV_PIXELS_B16;
+    x->mv_limits.col_max = (cm->mb_cols - 1) * 8 + BORDER_MV_PIXELS_B16;
+    xd->left_available = 0;
+
+    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
+      MBGRAPH_MB_STATS *mb_stats = &stats->mb_stats[offset + mb_col];
+
+      update_mbgraph_mb_stats(cpi, mb_stats, buf, mb_y_in_offset, golden_ref,
+                              &gld_left_mv, alt_ref, mb_row, mb_col);
+      gld_left_mv = mb_stats->ref[GOLDEN_FRAME].m.mv.as_mv;
+      if (mb_col == 0) {
+        gld_top_mv = gld_left_mv;
+      }
+      xd->left_available = 1;
+      mb_y_in_offset += 16;
+      gld_y_in_offset += 16;
+      arf_y_in_offset += 16;
+      x->mv_limits.col_min -= 16;
+      x->mv_limits.col_max -= 16;
+    }
+    xd->up_available = 1;
+    mb_y_offset += buf->y_stride * 16;
+    gld_y_offset += golden_ref->y_stride * 16;
+    if (alt_ref) arf_y_offset += alt_ref->y_stride * 16;
+    x->mv_limits.row_min -= 16;
+    x->mv_limits.row_max -= 16;
+    offset += cm->mb_cols;
+  }
+}
+
+// void separate_arf_mbs_byzz
+static void separate_arf_mbs(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  int mb_col, mb_row, offset, i;
+  int mi_row, mi_col;
+  int ncnt[4] = { 0 };
+  int n_frames = cpi->mbgraph_n_frames;
+
+  int *arf_not_zz;
+
+  CHECK_MEM_ERROR(
+      cm, arf_not_zz,
+      aom_calloc(cm->mb_rows * cm->mb_cols * sizeof(*arf_not_zz), 1));
+
+  // We are not interested in results beyond the alt ref itself.
+  if (n_frames > cpi->rc.frames_till_gf_update_due)
+    n_frames = cpi->rc.frames_till_gf_update_due;
+
+  // defer cost to reference frames
+  for (i = n_frames - 1; i >= 0; i--) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+
+    for (offset = 0, mb_row = 0; mb_row < cm->mb_rows;
+         offset += cm->mb_cols, mb_row++) {
+      for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
+        MBGRAPH_MB_STATS *mb_stats = &frame_stats->mb_stats[offset + mb_col];
+
+        int altref_err = mb_stats->ref[ALTREF_FRAME].err;
+        int intra_err = mb_stats->ref[INTRA_FRAME].err;
+        int golden_err = mb_stats->ref[GOLDEN_FRAME].err;
+
+        // Test for altref vs intra and gf and that its mv was 0,0.
+        if (altref_err > 1000 || altref_err > intra_err ||
+            altref_err > golden_err) {
+          arf_not_zz[offset + mb_col]++;
+        }
+      }
+    }
+  }
+
+  // arf_not_zz is indexed by MB, but this loop is indexed by MI to avoid out
+  // of bound access in segmentation_map
+  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
+    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++) {
+      // If any of the blocks in the sequence failed then the MB
+      // goes in segment 0
+      if (arf_not_zz[mi_row / 2 * cm->mb_cols + mi_col / 2]) {
+        ncnt[0]++;
+        cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 0;
+      } else {
+        cpi->segmentation_map[mi_row * cm->mi_cols + mi_col] = 1;
+        ncnt[1]++;
+      }
+    }
+  }
+
+  // Only bother with segmentation if over 10% of the MBs in static segment
+  // if ( ncnt[1] && (ncnt[0] / ncnt[1] < 10) )
+  if (1) {
+    // Note % of blocks that are marked as static
+    if (cm->MBs)
+      cpi->static_mb_pct = (ncnt[1] * 100) / (cm->mi_rows * cm->mi_cols);
+
+    // This error case should not be reachable as this function should
+    // never be called with the common data structure uninitialized.
+    else
+      cpi->static_mb_pct = 0;
+
+    av1_enable_segmentation(&cm->seg);
+  } else {
+    cpi->static_mb_pct = 0;
+    av1_disable_segmentation(&cm->seg);
+  }
+
+  // Free localy allocated storage
+  aom_free(arf_not_zz);
+}
+
+void av1_update_mbgraph_stats(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  int i, n_frames = av1_lookahead_depth(cpi->lookahead);
+  YV12_BUFFER_CONFIG *golden_ref = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
+
+  assert(golden_ref != NULL);
+
+  // we need to look ahead beyond where the ARF transitions into
+  // being a GF - so exit if we don't look ahead beyond that
+  if (n_frames <= cpi->rc.frames_till_gf_update_due) return;
+
+  if (n_frames > MAX_LAG_BUFFERS) n_frames = MAX_LAG_BUFFERS;
+
+  cpi->mbgraph_n_frames = n_frames;
+  for (i = 0; i < n_frames; i++) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+    memset(frame_stats->mb_stats, 0,
+           cm->mb_rows * cm->mb_cols * sizeof(*cpi->mbgraph_stats[i].mb_stats));
+  }
+
+  // do motion search to find contribution of each reference to data
+  // later on in this GF group
+  // FIXME really, the GF/last MC search should be done forward, and
+  // the ARF MC search backwards, to get optimal results for MV caching
+  for (i = 0; i < n_frames; i++) {
+    MBGRAPH_FRAME_STATS *frame_stats = &cpi->mbgraph_stats[i];
+    struct lookahead_entry *q_cur = av1_lookahead_peek(cpi->lookahead, i);
+
+    assert(q_cur != NULL);
+
+    update_mbgraph_frame_stats(cpi, frame_stats, &q_cur->img, golden_ref,
+                               cpi->source);
+  }
+
+  aom_clear_system_state();
+
+  separate_arf_mbs(cpi);
+}
diff --git a/third_party/aom/av1/encoder/mbgraph.h b/third_party/aom/av1/encoder/mbgraph.h
new file mode 100644
index 0000000000..758e2ad152
--- /dev/null
+++ b/third_party/aom/av1/encoder/mbgraph.h
@@ -0,0 +1,39 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_MBGRAPH_H_
+#define AV1_ENCODER_MBGRAPH_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  struct {
+    int err;
+    union {
+      int_mv mv;
+      PREDICTION_MODE mode;
+    } m;
+  } ref[TOTAL_REFS_PER_FRAME];
+} MBGRAPH_MB_STATS;
+
+typedef struct { MBGRAPH_MB_STATS *mb_stats; } MBGRAPH_FRAME_STATS;
+
+struct AV1_COMP;
+
+void av1_update_mbgraph_stats(struct AV1_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_MBGRAPH_H_
diff --git a/third_party/aom/av1/encoder/mcomp.c b/third_party/aom/av1/encoder/mcomp.c
new file mode 100644
index 0000000000..d069eefb0d
--- /dev/null
+++ b/third_party/aom/av1/encoder/mcomp.c
@@ -0,0 +1,3493 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#include "av1/common/common.h"
+#include "av1/common/mvref_common.h"
+#include "av1/common/reconinter.h"
+
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/mcomp.h"
+#include "av1/encoder/rdopt.h"
+
+// #define NEW_DIAMOND_SEARCH
+
+static INLINE const uint8_t *get_buf_from_mv(const struct buf_2d *buf,
+                                             const MV *mv) {
+  return &buf->buf[mv->row * buf->stride + mv->col];
+}
+
+void av1_set_mv_search_range(MvLimits *mv_limits, const MV *mv) {
+  int col_min = (mv->col >> 3) - MAX_FULL_PEL_VAL + (mv->col & 7 ? 1 : 0);
+  int row_min = (mv->row >> 3) - MAX_FULL_PEL_VAL + (mv->row & 7 ? 1 : 0);
+  int col_max = (mv->col >> 3) + MAX_FULL_PEL_VAL;
+  int row_max = (mv->row >> 3) + MAX_FULL_PEL_VAL;
+
+  col_min = AOMMAX(col_min, (MV_LOW >> 3) + 1);
+  row_min = AOMMAX(row_min, (MV_LOW >> 3) + 1);
+  col_max = AOMMIN(col_max, (MV_UPP >> 3) - 1);
+  row_max = AOMMIN(row_max, (MV_UPP >> 3) - 1);
+
+  // Get intersection of UMV window and valid MV window to reduce # of checks
+  // in diamond search.
+  if (mv_limits->col_min < col_min) mv_limits->col_min = col_min;
+  if (mv_limits->col_max > col_max) mv_limits->col_max = col_max;
+  if (mv_limits->row_min < row_min) mv_limits->row_min = row_min;
+  if (mv_limits->row_max > row_max) mv_limits->row_max = row_max;
+}
+
+static void av1_set_subpel_mv_search_range(const MvLimits *mv_limits,
+                                           int *col_min, int *col_max,
+                                           int *row_min, int *row_max,
+                                           const MV *ref_mv) {
+  const int max_mv = MAX_FULL_PEL_VAL * 8;
+  const int minc = AOMMAX(mv_limits->col_min * 8, ref_mv->col - max_mv);
+  const int maxc = AOMMIN(mv_limits->col_max * 8, ref_mv->col + max_mv);
+  const int minr = AOMMAX(mv_limits->row_min * 8, ref_mv->row - max_mv);
+  const int maxr = AOMMIN(mv_limits->row_max * 8, ref_mv->row + max_mv);
+
+  *col_min = AOMMAX(MV_LOW + 1, minc);
+  *col_max = AOMMIN(MV_UPP - 1, maxc);
+  *row_min = AOMMAX(MV_LOW + 1, minr);
+  *row_max = AOMMIN(MV_UPP - 1, maxr);
+}
+
+int av1_init_search_range(int size) {
+  int sr = 0;
+  // Minimum search size no matter what the passed in value.
+  size = AOMMAX(16, size);
+
+  while ((size << sr) < MAX_FULL_PEL_VAL) sr++;
+
+  sr = AOMMIN(sr, MAX_MVSEARCH_STEPS - 2);
+  return sr;
+}
+
+static INLINE int mv_cost(const MV *mv, const int *joint_cost,
+                          int *const comp_cost[2]) {
+  return joint_cost[av1_get_mv_joint(mv)] + comp_cost[0][mv->row] +
+         comp_cost[1][mv->col];
+}
+
+int av1_mv_bit_cost(const MV *mv, const MV *ref, const int *mvjcost,
+                    int *mvcost[2], int weight) {
+  const MV diff = { mv->row - ref->row, mv->col - ref->col };
+  return ROUND_POWER_OF_TWO(mv_cost(&diff, mvjcost, mvcost) * weight, 7);
+}
+
+#define PIXEL_TRANSFORM_ERROR_SCALE 4
+static int mv_err_cost(const MV *mv, const MV *ref, const int *mvjcost,
+                       int *mvcost[2], int error_per_bit) {
+  if (mvcost) {
+    const MV diff = { mv->row - ref->row, mv->col - ref->col };
+    return (int)ROUND_POWER_OF_TWO_64(
+        (int64_t)mv_cost(&diff, mvjcost, mvcost) * error_per_bit,
+        RDDIV_BITS + AV1_PROB_COST_SHIFT - RD_EPB_SHIFT +
+            PIXEL_TRANSFORM_ERROR_SCALE);
+  }
+  return 0;
+}
+
+static int mvsad_err_cost(const MACROBLOCK *x, const MV *mv, const MV *ref,
+                          int sad_per_bit) {
+  const MV diff = { (mv->row - ref->row) * 8, (mv->col - ref->col) * 8 };
+  return ROUND_POWER_OF_TWO(
+      (unsigned)mv_cost(&diff, x->nmvjointsadcost, x->mvsadcost) * sad_per_bit,
+      AV1_PROB_COST_SHIFT);
+}
+
+void av1_init_dsmotion_compensation(search_site_config *cfg, int stride) {
+  int len, ss_count = 1;
+
+  cfg->ss[0].mv.col = cfg->ss[0].mv.row = 0;
+  cfg->ss[0].offset = 0;
+
+  for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
+    // Generate offsets for 4 search sites per step.
+    const MV ss_mvs[] = { { -len, 0 }, { len, 0 }, { 0, -len }, { 0, len } };
+    int i;
+    for (i = 0; i < 4; ++i) {
+      search_site *const ss = &cfg->ss[ss_count++];
+      ss->mv = ss_mvs[i];
+      ss->offset = ss->mv.row * stride + ss->mv.col;
+    }
+  }
+
+  cfg->ss_count = ss_count;
+  cfg->searches_per_step = 4;
+}
+
+void av1_init3smotion_compensation(search_site_config *cfg, int stride) {
+  int len, ss_count = 1;
+
+  cfg->ss[0].mv.col = cfg->ss[0].mv.row = 0;
+  cfg->ss[0].offset = 0;
+
+  for (len = MAX_FIRST_STEP; len > 0; len /= 2) {
+    // Generate offsets for 8 search sites per step.
+    const MV ss_mvs[8] = { { -len, 0 },   { len, 0 },     { 0, -len },
+                           { 0, len },    { -len, -len }, { -len, len },
+                           { len, -len }, { len, len } };
+    int i;
+    for (i = 0; i < 8; ++i) {
+      search_site *const ss = &cfg->ss[ss_count++];
+      ss->mv = ss_mvs[i];
+      ss->offset = ss->mv.row * stride + ss->mv.col;
+    }
+  }
+
+  cfg->ss_count = ss_count;
+  cfg->searches_per_step = 8;
+}
+
+/*
+ * To avoid the penalty for crossing cache-line read, preload the reference
+ * area in a small buffer, which is aligned to make sure there won't be crossing
+ * cache-line read while reading from this buffer. This reduced the cpu
+ * cycles spent on reading ref data in sub-pixel filter functions.
+ * TODO: Currently, since sub-pixel search range here is -3 ~ 3, copy 22 rows x
+ * 32 cols area that is enough for 16x16 macroblock. Later, for SPLITMV, we
+ * could reduce the area.
+ */
+
+// convert motion vector component to offset for sv[a]f calc
+static INLINE int sp(int x) { return x & 7; }
+
+static INLINE const uint8_t *pre(const uint8_t *buf, int stride, int r, int c) {
+  return &buf[(r >> 3) * stride + (c >> 3)];
+}
+
+/* checks if (r, c) has better score than previous best */
+#define CHECK_BETTER(v, r, c)                                             \
+  if (c >= minc && c <= maxc && r >= minr && r <= maxr) {                 \
+    MV this_mv = { r, c };                                                \
+    v = mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit);    \
+    if (second_pred == NULL)                                              \
+      thismse = vfp->svf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r),  \
+                         src_address, src_stride, &sse);                  \
+    else                                                                  \
+      thismse = vfp->svaf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), \
+                          src_address, src_stride, &sse, second_pred);    \
+    v += thismse;                                                         \
+    if (v < besterr) {                                                    \
+      besterr = v;                                                        \
+      br = r;                                                             \
+      bc = c;                                                             \
+      *distortion = thismse;                                              \
+      *sse1 = sse;                                                        \
+    }                                                                     \
+  } else {                                                                \
+    v = INT_MAX;                                                          \
+  }
+
+#define CHECK_BETTER0(v, r, c) CHECK_BETTER(v, r, c)
+
+static INLINE const uint8_t *upre(const uint8_t *buf, int stride, int r,
+                                  int c) {
+  return &buf[(r)*stride + (c)];
+}
+
+/* checks if (r, c) has better score than previous best */
+#define CHECK_BETTER1(v, r, c)                                         \
+  if (c >= minc && c <= maxc && r >= minr && r <= maxr) {              \
+    MV this_mv = { r, c };                                             \
+    thismse = upsampled_pref_error(xd, vfp, src_address, src_stride,   \
+                                   upre(y, y_stride, r, c), y_stride,  \
+                                   second_pred, w, h, &sse);           \
+    v = mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit); \
+    v += thismse;                                                      \
+    if (v < besterr) {                                                 \
+      besterr = v;                                                     \
+      br = r;                                                          \
+      bc = c;                                                          \
+      *distortion = thismse;                                           \
+      *sse1 = sse;                                                     \
+    }                                                                  \
+  } else {                                                             \
+    v = INT_MAX;                                                       \
+  }
+
+#define FIRST_LEVEL_CHECKS                                       \
+  {                                                              \
+    unsigned int left, right, up, down, diag;                    \
+    CHECK_BETTER(left, tr, tc - hstep);                          \
+    CHECK_BETTER(right, tr, tc + hstep);                         \
+    CHECK_BETTER(up, tr - hstep, tc);                            \
+    CHECK_BETTER(down, tr + hstep, tc);                          \
+    whichdir = (left < right ? 0 : 1) + (up < down ? 0 : 2);     \
+    switch (whichdir) {                                          \
+      case 0: CHECK_BETTER(diag, tr - hstep, tc - hstep); break; \
+      case 1: CHECK_BETTER(diag, tr - hstep, tc + hstep); break; \
+      case 2: CHECK_BETTER(diag, tr + hstep, tc - hstep); break; \
+      case 3: CHECK_BETTER(diag, tr + hstep, tc + hstep); break; \
+    }                                                            \
+  }
+
+#define SECOND_LEVEL_CHECKS                                       \
+  {                                                               \
+    int kr, kc;                                                   \
+    unsigned int second;                                          \
+    if (tr != br && tc != bc) {                                   \
+      kr = br - tr;                                               \
+      kc = bc - tc;                                               \
+      CHECK_BETTER(second, tr + kr, tc + 2 * kc);                 \
+      CHECK_BETTER(second, tr + 2 * kr, tc + kc);                 \
+    } else if (tr == br && tc != bc) {                            \
+      kc = bc - tc;                                               \
+      CHECK_BETTER(second, tr + hstep, tc + 2 * kc);              \
+      CHECK_BETTER(second, tr - hstep, tc + 2 * kc);              \
+      switch (whichdir) {                                         \
+        case 0:                                                   \
+        case 1: CHECK_BETTER(second, tr + hstep, tc + kc); break; \
+        case 2:                                                   \
+        case 3: CHECK_BETTER(second, tr - hstep, tc + kc); break; \
+      }                                                           \
+    } else if (tr != br && tc == bc) {                            \
+      kr = br - tr;                                               \
+      CHECK_BETTER(second, tr + 2 * kr, tc + hstep);              \
+      CHECK_BETTER(second, tr + 2 * kr, tc - hstep);              \
+      switch (whichdir) {                                         \
+        case 0:                                                   \
+        case 2: CHECK_BETTER(second, tr + kr, tc + hstep); break; \
+        case 1:                                                   \
+        case 3: CHECK_BETTER(second, tr + kr, tc - hstep); break; \
+      }                                                           \
+    }                                                             \
+  }
+
+// TODO(yunqingwang): SECOND_LEVEL_CHECKS_BEST was a rewrote of
+// SECOND_LEVEL_CHECKS, and SECOND_LEVEL_CHECKS should be rewritten
+// later in the same way.
+#define SECOND_LEVEL_CHECKS_BEST(k)                \
+  {                                                \
+    unsigned int second;                           \
+    int br0 = br;                                  \
+    int bc0 = bc;                                  \
+    assert(tr == br || tc == bc);                  \
+    if (tr == br && tc != bc) {                    \
+      kc = bc - tc;                                \
+    } else if (tr != br && tc == bc) {             \
+      kr = br - tr;                                \
+    }                                              \
+    CHECK_BETTER##k(second, br0 + kr, bc0);        \
+    CHECK_BETTER##k(second, br0, bc0 + kc);        \
+    if (br0 != br || bc0 != bc) {                  \
+      CHECK_BETTER##k(second, br0 + kr, bc0 + kc); \
+    }                                              \
+  }
+
+#define SETUP_SUBPEL_SEARCH                                                 \
+  const uint8_t *const src_address = x->plane[0].src.buf;                   \
+  const int src_stride = x->plane[0].src.stride;                            \
+  const MACROBLOCKD *xd = &x->e_mbd;                                        \
+  unsigned int besterr = INT_MAX;                                           \
+  unsigned int sse;                                                         \
+  unsigned int whichdir;                                                    \
+  int thismse;                                                              \
+  MV *bestmv = &x->best_mv.as_mv;                                           \
+  const unsigned int halfiters = iters_per_step;                            \
+  const unsigned int quarteriters = iters_per_step;                         \
+  const unsigned int eighthiters = iters_per_step;                          \
+  const int y_stride = xd->plane[0].pre[0].stride;                          \
+  const int offset = bestmv->row * y_stride + bestmv->col;                  \
+  const uint8_t *const y = xd->plane[0].pre[0].buf;                         \
+                                                                            \
+  int br = bestmv->row * 8;                                                 \
+  int bc = bestmv->col * 8;                                                 \
+  int hstep = 4;                                                            \
+  int minc, maxc, minr, maxr;                                               \
+  int tr = br;                                                              \
+  int tc = bc;                                                              \
+                                                                            \
+  av1_set_subpel_mv_search_range(&x->mv_limits, &minc, &maxc, &minr, &maxr, \
+                                 ref_mv);                                   \
+                                                                            \
+  bestmv->row *= 8;                                                         \
+  bestmv->col *= 8;
+
+static unsigned int setup_center_error(
+    const MACROBLOCKD *xd, const MV *bestmv, const MV *ref_mv,
+    int error_per_bit, const aom_variance_fn_ptr_t *vfp,
+    const uint8_t *const src, const int src_stride, const uint8_t *const y,
+    int y_stride, const uint8_t *second_pred, int w, int h, int offset,
+    int *mvjcost, int *mvcost[2], unsigned int *sse1, int *distortion) {
+  unsigned int besterr;
+#if CONFIG_HIGHBITDEPTH
+  if (second_pred != NULL) {
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      DECLARE_ALIGNED(16, uint16_t, comp_pred16[MAX_SB_SQUARE]);
+      aom_highbd_comp_avg_pred(comp_pred16, second_pred, w, h, y + offset,
+                               y_stride);
+      besterr =
+          vfp->vf(CONVERT_TO_BYTEPTR(comp_pred16), w, src, src_stride, sse1);
+    } else {
+      DECLARE_ALIGNED(16, uint8_t, comp_pred[MAX_SB_SQUARE]);
+      aom_comp_avg_pred(comp_pred, second_pred, w, h, y + offset, y_stride);
+      besterr = vfp->vf(comp_pred, w, src, src_stride, sse1);
+    }
+  } else {
+    besterr = vfp->vf(y + offset, y_stride, src, src_stride, sse1);
+  }
+  *distortion = besterr;
+  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
+#else
+  (void)xd;
+  if (second_pred != NULL) {
+    DECLARE_ALIGNED(16, uint8_t, comp_pred[MAX_SB_SQUARE]);
+    aom_comp_avg_pred(comp_pred, second_pred, w, h, y + offset, y_stride);
+    besterr = vfp->vf(comp_pred, w, src, src_stride, sse1);
+  } else {
+    besterr = vfp->vf(y + offset, y_stride, src, src_stride, sse1);
+  }
+  *distortion = besterr;
+  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
+#endif  // CONFIG_HIGHBITDEPTH
+  return besterr;
+}
+
+static INLINE int divide_and_round(int n, int d) {
+  return ((n < 0) ^ (d < 0)) ? ((n - d / 2) / d) : ((n + d / 2) / d);
+}
+
+static INLINE int is_cost_list_wellbehaved(int *cost_list) {
+  return cost_list[0] < cost_list[1] && cost_list[0] < cost_list[2] &&
+         cost_list[0] < cost_list[3] && cost_list[0] < cost_list[4];
+}
+
+// Returns surface minima estimate at given precision in 1/2^n bits.
+// Assume a model for the cost surface: S = A(x - x0)^2 + B(y - y0)^2 + C
+// For a given set of costs S0, S1, S2, S3, S4 at points
+// (y, x) = (0, 0), (0, -1), (1, 0), (0, 1) and (-1, 0) respectively,
+// the solution for the location of the minima (x0, y0) is given by:
+// x0 = 1/2 (S1 - S3)/(S1 + S3 - 2*S0),
+// y0 = 1/2 (S4 - S2)/(S4 + S2 - 2*S0).
+// The code below is an integerized version of that.
+static void get_cost_surf_min(int *cost_list, int *ir, int *ic, int bits) {
+  *ic = divide_and_round((cost_list[1] - cost_list[3]) * (1 << (bits - 1)),
+                         (cost_list[1] - 2 * cost_list[0] + cost_list[3]));
+  *ir = divide_and_round((cost_list[4] - cost_list[2]) * (1 << (bits - 1)),
+                         (cost_list[4] - 2 * cost_list[0] + cost_list[2]));
+}
+
+int av1_find_best_sub_pixel_tree_pruned_evenmore(
+    MACROBLOCK *x, const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, int forced_stop, int iters_per_step,
+    int *cost_list, int *mvjcost, int *mvcost[2], int *distortion,
+    unsigned int *sse1, const uint8_t *second_pred, int w, int h,
+    int use_upsampled_ref) {
+  SETUP_SUBPEL_SEARCH;
+  besterr = setup_center_error(
+      xd, bestmv, ref_mv, error_per_bit, vfp, src_address, src_stride, y,
+      y_stride, second_pred, w, h, offset, mvjcost, mvcost, sse1, distortion);
+  (void)halfiters;
+  (void)quarteriters;
+  (void)eighthiters;
+  (void)whichdir;
+  (void)allow_hp;
+  (void)forced_stop;
+  (void)hstep;
+  (void)use_upsampled_ref;
+
+  if (cost_list && cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
+      cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
+      cost_list[4] != INT_MAX && is_cost_list_wellbehaved(cost_list)) {
+    int ir, ic;
+    unsigned int minpt;
+    get_cost_surf_min(cost_list, &ir, &ic, 2);
+    if (ir != 0 || ic != 0) {
+      CHECK_BETTER(minpt, tr + 2 * ir, tc + 2 * ic);
+    }
+  } else {
+    FIRST_LEVEL_CHECKS;
+    if (halfiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+
+    tr = br;
+    tc = bc;
+
+    // Each subsequent iteration checks at least one point in common with
+    // the last iteration could be 2 ( if diag selected) 1/4 pel
+    // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
+    if (forced_stop != 2) {
+      hstep >>= 1;
+      FIRST_LEVEL_CHECKS;
+      if (quarteriters > 1) {
+        SECOND_LEVEL_CHECKS;
+      }
+    }
+  }
+
+  tr = br;
+  tc = bc;
+
+  if (allow_hp && forced_stop == 0) {
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (eighthiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+  }
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  return besterr;
+}
+
+int av1_find_best_sub_pixel_tree_pruned_more(
+    MACROBLOCK *x, const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, int forced_stop, int iters_per_step,
+    int *cost_list, int *mvjcost, int *mvcost[2], int *distortion,
+    unsigned int *sse1, const uint8_t *second_pred, int w, int h,
+    int use_upsampled_ref) {
+  SETUP_SUBPEL_SEARCH;
+  (void)use_upsampled_ref;
+
+  besterr = setup_center_error(
+      xd, bestmv, ref_mv, error_per_bit, vfp, src_address, src_stride, y,
+      y_stride, second_pred, w, h, offset, mvjcost, mvcost, sse1, distortion);
+  if (cost_list && cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
+      cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
+      cost_list[4] != INT_MAX && is_cost_list_wellbehaved(cost_list)) {
+    unsigned int minpt;
+    int ir, ic;
+    get_cost_surf_min(cost_list, &ir, &ic, 1);
+    if (ir != 0 || ic != 0) {
+      CHECK_BETTER(minpt, tr + ir * hstep, tc + ic * hstep);
+    }
+  } else {
+    FIRST_LEVEL_CHECKS;
+    if (halfiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+  }
+
+  // Each subsequent iteration checks at least one point in common with
+  // the last iteration could be 2 ( if diag selected) 1/4 pel
+
+  // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
+  if (forced_stop != 2) {
+    tr = br;
+    tc = bc;
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (quarteriters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+  }
+
+  if (allow_hp && forced_stop == 0) {
+    tr = br;
+    tc = bc;
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (eighthiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+  }
+  // These lines insure static analysis doesn't warn that
+  // tr and tc aren't used after the above point.
+  (void)tr;
+  (void)tc;
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  return besterr;
+}
+
+int av1_find_best_sub_pixel_tree_pruned(
+    MACROBLOCK *x, const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, int forced_stop, int iters_per_step,
+    int *cost_list, int *mvjcost, int *mvcost[2], int *distortion,
+    unsigned int *sse1, const uint8_t *second_pred, int w, int h,
+    int use_upsampled_ref) {
+  SETUP_SUBPEL_SEARCH;
+  (void)use_upsampled_ref;
+
+  besterr = setup_center_error(
+      xd, bestmv, ref_mv, error_per_bit, vfp, src_address, src_stride, y,
+      y_stride, second_pred, w, h, offset, mvjcost, mvcost, sse1, distortion);
+  if (cost_list && cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
+      cost_list[2] != INT_MAX && cost_list[3] != INT_MAX &&
+      cost_list[4] != INT_MAX) {
+    unsigned int left, right, up, down, diag;
+    whichdir = (cost_list[1] < cost_list[3] ? 0 : 1) +
+               (cost_list[2] < cost_list[4] ? 0 : 2);
+    switch (whichdir) {
+      case 0:
+        CHECK_BETTER(left, tr, tc - hstep);
+        CHECK_BETTER(down, tr + hstep, tc);
+        CHECK_BETTER(diag, tr + hstep, tc - hstep);
+        break;
+      case 1:
+        CHECK_BETTER(right, tr, tc + hstep);
+        CHECK_BETTER(down, tr + hstep, tc);
+        CHECK_BETTER(diag, tr + hstep, tc + hstep);
+        break;
+      case 2:
+        CHECK_BETTER(left, tr, tc - hstep);
+        CHECK_BETTER(up, tr - hstep, tc);
+        CHECK_BETTER(diag, tr - hstep, tc - hstep);
+        break;
+      case 3:
+        CHECK_BETTER(right, tr, tc + hstep);
+        CHECK_BETTER(up, tr - hstep, tc);
+        CHECK_BETTER(diag, tr - hstep, tc + hstep);
+        break;
+    }
+  } else {
+    FIRST_LEVEL_CHECKS;
+    if (halfiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+  }
+
+  tr = br;
+  tc = bc;
+
+  // Each subsequent iteration checks at least one point in common with
+  // the last iteration could be 2 ( if diag selected) 1/4 pel
+
+  // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
+  if (forced_stop != 2) {
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (quarteriters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+    tr = br;
+    tc = bc;
+  }
+
+  if (allow_hp && forced_stop == 0) {
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (eighthiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+    tr = br;
+    tc = bc;
+  }
+  // These lines insure static analysis doesn't warn that
+  // tr and tc aren't used after the above point.
+  (void)tr;
+  (void)tc;
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  return besterr;
+}
+
+/* clang-format off */
+static const MV search_step_table[12] = {
+  // left, right, up, down
+  { 0, -4 }, { 0, 4 }, { -4, 0 }, { 4, 0 },
+  { 0, -2 }, { 0, 2 }, { -2, 0 }, { 2, 0 },
+  { 0, -1 }, { 0, 1 }, { -1, 0 }, { 1, 0 }
+};
+/* clang-format on */
+
+static int upsampled_pref_error(const MACROBLOCKD *xd,
+                                const aom_variance_fn_ptr_t *vfp,
+                                const uint8_t *const src, const int src_stride,
+                                const uint8_t *const y, int y_stride,
+                                const uint8_t *second_pred, int w, int h,
+                                unsigned int *sse) {
+  unsigned int besterr;
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    DECLARE_ALIGNED(16, uint16_t, pred16[MAX_SB_SQUARE]);
+    if (second_pred != NULL)
+      aom_highbd_comp_avg_upsampled_pred(pred16, second_pred, w, h, y,
+                                         y_stride);
+    else
+      aom_highbd_upsampled_pred(pred16, w, h, y, y_stride);
+
+    besterr = vfp->vf(CONVERT_TO_BYTEPTR(pred16), w, src, src_stride, sse);
+  } else {
+    DECLARE_ALIGNED(16, uint8_t, pred[MAX_SB_SQUARE]);
+#else
+  DECLARE_ALIGNED(16, uint8_t, pred[MAX_SB_SQUARE]);
+  (void)xd;
+#endif  // CONFIG_HIGHBITDEPTH
+    if (second_pred != NULL)
+      aom_comp_avg_upsampled_pred(pred, second_pred, w, h, y, y_stride);
+    else
+      aom_upsampled_pred(pred, w, h, y, y_stride);
+
+    besterr = vfp->vf(pred, w, src, src_stride, sse);
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif
+  return besterr;
+}
+
+static unsigned int upsampled_setup_center_error(
+    const MACROBLOCKD *xd, const MV *bestmv, const MV *ref_mv,
+    int error_per_bit, const aom_variance_fn_ptr_t *vfp,
+    const uint8_t *const src, const int src_stride, const uint8_t *const y,
+    int y_stride, const uint8_t *second_pred, int w, int h, int offset,
+    int *mvjcost, int *mvcost[2], unsigned int *sse1, int *distortion) {
+  unsigned int besterr = upsampled_pref_error(
+      xd, vfp, src, src_stride, y + offset, y_stride, second_pred, w, h, sse1);
+  *distortion = besterr;
+  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
+  return besterr;
+}
+
+int av1_find_best_sub_pixel_tree(MACROBLOCK *x, const MV *ref_mv, int allow_hp,
+                                 int error_per_bit,
+                                 const aom_variance_fn_ptr_t *vfp,
+                                 int forced_stop, int iters_per_step,
+                                 int *cost_list, int *mvjcost, int *mvcost[2],
+                                 int *distortion, unsigned int *sse1,
+                                 const uint8_t *second_pred, int w, int h,
+                                 int use_upsampled_ref) {
+  const uint8_t *const src_address = x->plane[0].src.buf;
+  const int src_stride = x->plane[0].src.stride;
+  const MACROBLOCKD *xd = &x->e_mbd;
+  unsigned int besterr = INT_MAX;
+  unsigned int sse;
+  unsigned int thismse;
+  const int y_stride = xd->plane[0].pre[0].stride;
+  MV *bestmv = &x->best_mv.as_mv;
+  const int offset = bestmv->row * y_stride + bestmv->col;
+  const uint8_t *const y = xd->plane[0].pre[0].buf;
+
+  int br = bestmv->row * 8;
+  int bc = bestmv->col * 8;
+  int hstep = 4;
+  int iter, round = 3 - forced_stop;
+  int tr = br;
+  int tc = bc;
+  const MV *search_step = search_step_table;
+  int idx, best_idx = -1;
+  unsigned int cost_array[5];
+  int kr, kc;
+  int minc, maxc, minr, maxr;
+
+  av1_set_subpel_mv_search_range(&x->mv_limits, &minc, &maxc, &minr, &maxr,
+                                 ref_mv);
+
+  if (!allow_hp)
+    if (round == 3) round = 2;
+
+  bestmv->row *= 8;
+  bestmv->col *= 8;
+
+  // use_upsampled_ref can be 0 or 1
+  if (use_upsampled_ref)
+    besterr = upsampled_setup_center_error(
+        xd, bestmv, ref_mv, error_per_bit, vfp, src_address, src_stride, y,
+        y_stride, second_pred, w, h, (offset * 8), mvjcost, mvcost, sse1,
+        distortion);
+  else
+    besterr = setup_center_error(
+        xd, bestmv, ref_mv, error_per_bit, vfp, src_address, src_stride, y,
+        y_stride, second_pred, w, h, offset, mvjcost, mvcost, sse1, distortion);
+
+  (void)cost_list;  // to silence compiler warning
+
+  for (iter = 0; iter < round; ++iter) {
+    // Check vertical and horizontal sub-pixel positions.
+    for (idx = 0; idx < 4; ++idx) {
+      tr = br + search_step[idx].row;
+      tc = bc + search_step[idx].col;
+      if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
+        MV this_mv = { tr, tc };
+
+        if (use_upsampled_ref) {
+          const uint8_t *const pre_address = y + tr * y_stride + tc;
+
+          thismse = upsampled_pref_error(xd, vfp, src_address, src_stride,
+                                         pre_address, y_stride, second_pred, w,
+                                         h, &sse);
+        } else {
+          const uint8_t *const pre_address =
+              y + (tr >> 3) * y_stride + (tc >> 3);
+          if (second_pred == NULL)
+            thismse = vfp->svf(pre_address, y_stride, sp(tc), sp(tr),
+                               src_address, src_stride, &sse);
+          else
+            thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
+                                src_address, src_stride, &sse, second_pred);
+        }
+
+        cost_array[idx] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost,
+                                                mvcost, error_per_bit);
+
+        if (cost_array[idx] < besterr) {
+          best_idx = idx;
+          besterr = cost_array[idx];
+          *distortion = thismse;
+          *sse1 = sse;
+        }
+      } else {
+        cost_array[idx] = INT_MAX;
+      }
+    }
+
+    // Check diagonal sub-pixel position
+    kc = (cost_array[0] <= cost_array[1] ? -hstep : hstep);
+    kr = (cost_array[2] <= cost_array[3] ? -hstep : hstep);
+
+    tc = bc + kc;
+    tr = br + kr;
+    if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
+      MV this_mv = { tr, tc };
+
+      if (use_upsampled_ref) {
+        const uint8_t *const pre_address = y + tr * y_stride + tc;
+
+        thismse =
+            upsampled_pref_error(xd, vfp, src_address, src_stride, pre_address,
+                                 y_stride, second_pred, w, h, &sse);
+      } else {
+        const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
+
+        if (second_pred == NULL)
+          thismse = vfp->svf(pre_address, y_stride, sp(tc), sp(tr), src_address,
+                             src_stride, &sse);
+        else
+          thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
+                              src_address, src_stride, &sse, second_pred);
+      }
+
+      cost_array[4] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost,
+                                            error_per_bit);
+
+      if (cost_array[4] < besterr) {
+        best_idx = 4;
+        besterr = cost_array[4];
+        *distortion = thismse;
+        *sse1 = sse;
+      }
+    } else {
+      cost_array[idx] = INT_MAX;
+    }
+
+    if (best_idx < 4 && best_idx >= 0) {
+      br += search_step[best_idx].row;
+      bc += search_step[best_idx].col;
+    } else if (best_idx == 4) {
+      br = tr;
+      bc = tc;
+    }
+
+    if (iters_per_step > 1 && best_idx != -1) {
+      if (use_upsampled_ref) {
+        SECOND_LEVEL_CHECKS_BEST(1);
+      } else {
+        SECOND_LEVEL_CHECKS_BEST(0);
+      }
+    }
+
+    search_step += 4;
+    hstep >>= 1;
+    best_idx = -1;
+  }
+
+  // These lines insure static analysis doesn't warn that
+  // tr and tc aren't used after the above point.
+  (void)tr;
+  (void)tc;
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  return besterr;
+}
+
+#undef PRE
+#undef CHECK_BETTER
+
+static INLINE int check_bounds(const MvLimits *mv_limits, int row, int col,
+                               int range) {
+  return ((row - range) >= mv_limits->row_min) &
+         ((row + range) <= mv_limits->row_max) &
+         ((col - range) >= mv_limits->col_min) &
+         ((col + range) <= mv_limits->col_max);
+}
+
+static INLINE int is_mv_in(const MvLimits *mv_limits, const MV *mv) {
+  return (mv->col >= mv_limits->col_min) && (mv->col <= mv_limits->col_max) &&
+         (mv->row >= mv_limits->row_min) && (mv->row <= mv_limits->row_max);
+}
+
+#define CHECK_BETTER                                                      \
+  {                                                                       \
+    if (thissad < bestsad) {                                              \
+      if (use_mvcost)                                                     \
+        thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit); \
+      if (thissad < bestsad) {                                            \
+        bestsad = thissad;                                                \
+        best_site = i;                                                    \
+      }                                                                   \
+    }                                                                     \
+  }
+
+#define MAX_PATTERN_SCALES 11
+#define MAX_PATTERN_CANDIDATES 8  // max number of canddiates per scale
+#define PATTERN_CANDIDATES_REF 3  // number of refinement candidates
+
+// Calculate and return a sad+mvcost list around an integer best pel.
+static INLINE void calc_int_cost_list(const MACROBLOCK *x,
+                                      const MV *const ref_mv, int sadpb,
+                                      const aom_variance_fn_ptr_t *fn_ptr,
+                                      const MV *best_mv, int *cost_list) {
+  static const MV neighbors[4] = { { 0, -1 }, { 1, 0 }, { 0, 1 }, { -1, 0 } };
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &x->e_mbd.plane[0].pre[0];
+  const MV fcenter_mv = { ref_mv->row >> 3, ref_mv->col >> 3 };
+  const int br = best_mv->row;
+  const int bc = best_mv->col;
+  int i;
+  unsigned int sse;
+  const MV this_mv = { br, bc };
+
+  cost_list[0] =
+      fn_ptr->vf(what->buf, what->stride, get_buf_from_mv(in_what, &this_mv),
+                 in_what->stride, &sse) +
+      mvsad_err_cost(x, &this_mv, &fcenter_mv, sadpb);
+  if (check_bounds(&x->mv_limits, br, bc, 1)) {
+    for (i = 0; i < 4; i++) {
+      const MV neighbor_mv = { br + neighbors[i].row, bc + neighbors[i].col };
+      cost_list[i + 1] = fn_ptr->vf(what->buf, what->stride,
+                                    get_buf_from_mv(in_what, &neighbor_mv),
+                                    in_what->stride, &sse) +
+                         mv_err_cost(&neighbor_mv, &fcenter_mv, x->nmvjointcost,
+                                     x->mvcost, x->errorperbit);
+    }
+  } else {
+    for (i = 0; i < 4; i++) {
+      const MV neighbor_mv = { br + neighbors[i].row, bc + neighbors[i].col };
+      if (!is_mv_in(&x->mv_limits, &neighbor_mv))
+        cost_list[i + 1] = INT_MAX;
+      else
+        cost_list[i + 1] =
+            fn_ptr->vf(what->buf, what->stride,
+                       get_buf_from_mv(in_what, &neighbor_mv), in_what->stride,
+                       &sse) +
+            mv_err_cost(&neighbor_mv, &fcenter_mv, x->nmvjointcost, x->mvcost,
+                        x->errorperbit);
+    }
+  }
+}
+
+static INLINE void calc_int_sad_list(const MACROBLOCK *x,
+                                     const MV *const ref_mv, int sadpb,
+                                     const aom_variance_fn_ptr_t *fn_ptr,
+                                     const MV *best_mv, int *cost_list,
+                                     const int use_mvcost, const int bestsad) {
+  static const MV neighbors[4] = { { 0, -1 }, { 1, 0 }, { 0, 1 }, { -1, 0 } };
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &x->e_mbd.plane[0].pre[0];
+  const MV fcenter_mv = { ref_mv->row >> 3, ref_mv->col >> 3 };
+  int i;
+  const int br = best_mv->row;
+  const int bc = best_mv->col;
+
+  if (cost_list[0] == INT_MAX) {
+    cost_list[0] = bestsad;
+    if (check_bounds(&x->mv_limits, br, bc, 1)) {
+      for (i = 0; i < 4; i++) {
+        const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
+        cost_list[i + 1] =
+            fn_ptr->sdf(what->buf, what->stride,
+                        get_buf_from_mv(in_what, &this_mv), in_what->stride);
+      }
+    } else {
+      for (i = 0; i < 4; i++) {
+        const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
+        if (!is_mv_in(&x->mv_limits, &this_mv))
+          cost_list[i + 1] = INT_MAX;
+        else
+          cost_list[i + 1] =
+              fn_ptr->sdf(what->buf, what->stride,
+                          get_buf_from_mv(in_what, &this_mv), in_what->stride);
+      }
+    }
+  } else {
+    if (use_mvcost) {
+      for (i = 0; i < 4; i++) {
+        const MV this_mv = { br + neighbors[i].row, bc + neighbors[i].col };
+        if (cost_list[i + 1] != INT_MAX) {
+          cost_list[i + 1] += mvsad_err_cost(x, &this_mv, &fcenter_mv, sadpb);
+        }
+      }
+    }
+  }
+}
+
+// Generic pattern search function that searches over multiple scales.
+// Each scale can have a different number of candidates and shape of
+// candidates as indicated in the num_candidates and candidates arrays
+// passed into this function
+//
+static int pattern_search(
+    MACROBLOCK *x, MV *start_mv, int search_param, int sad_per_bit,
+    int do_init_search, int *cost_list, const aom_variance_fn_ptr_t *vfp,
+    int use_mvcost, const MV *center_mv,
+    const int num_candidates[MAX_PATTERN_SCALES],
+    const MV candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES]) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  static const int search_param_to_steps[MAX_MVSEARCH_STEPS] = {
+    10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
+  };
+  int i, s, t;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const int last_is_4 = num_candidates[0] == 4;
+  int br, bc;
+  int bestsad = INT_MAX;
+  int thissad;
+  int k = -1;
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  int best_init_s = search_param_to_steps[search_param];
+  // adjust ref_mv to make sure it is within MV range
+  clamp_mv(start_mv, x->mv_limits.col_min, x->mv_limits.col_max,
+           x->mv_limits.row_min, x->mv_limits.row_max);
+  br = start_mv->row;
+  bc = start_mv->col;
+  if (cost_list != NULL) {
+    cost_list[0] = cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] =
+        INT_MAX;
+  }
+
+  // Work out the start point for the search
+  bestsad = vfp->sdf(what->buf, what->stride,
+                     get_buf_from_mv(in_what, start_mv), in_what->stride) +
+            mvsad_err_cost(x, start_mv, &fcenter_mv, sad_per_bit);
+
+  // Search all possible scales upto the search param around the center point
+  // pick the scale of the point that is best as the starting scale of
+  // further steps around it.
+  if (do_init_search) {
+    s = best_init_s;
+    best_init_s = -1;
+    for (t = 0; t <= s; ++t) {
+      int best_site = -1;
+      if (check_bounds(&x->mv_limits, br, bc, 1 << t)) {
+        for (i = 0; i < num_candidates[t]; i++) {
+          const MV this_mv = { br + candidates[t][i].row,
+                               bc + candidates[t][i].col };
+          thissad =
+              vfp->sdf(what->buf, what->stride,
+                       get_buf_from_mv(in_what, &this_mv), in_what->stride);
+          CHECK_BETTER
+        }
+      } else {
+        for (i = 0; i < num_candidates[t]; i++) {
+          const MV this_mv = { br + candidates[t][i].row,
+                               bc + candidates[t][i].col };
+          if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
+          thissad =
+              vfp->sdf(what->buf, what->stride,
+                       get_buf_from_mv(in_what, &this_mv), in_what->stride);
+          CHECK_BETTER
+        }
+      }
+      if (best_site == -1) {
+        continue;
+      } else {
+        best_init_s = t;
+        k = best_site;
+      }
+    }
+    if (best_init_s != -1) {
+      br += candidates[best_init_s][k].row;
+      bc += candidates[best_init_s][k].col;
+    }
+  }
+
+  // If the center point is still the best, just skip this and move to
+  // the refinement step.
+  if (best_init_s != -1) {
+    const int last_s = (last_is_4 && cost_list != NULL);
+    int best_site = -1;
+    s = best_init_s;
+
+    for (; s >= last_s; s--) {
+      // No need to search all points the 1st time if initial search was used
+      if (!do_init_search || s != best_init_s) {
+        if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
+          for (i = 0; i < num_candidates[s]; i++) {
+            const MV this_mv = { br + candidates[s][i].row,
+                                 bc + candidates[s][i].col };
+            thissad =
+                vfp->sdf(what->buf, what->stride,
+                         get_buf_from_mv(in_what, &this_mv), in_what->stride);
+            CHECK_BETTER
+          }
+        } else {
+          for (i = 0; i < num_candidates[s]; i++) {
+            const MV this_mv = { br + candidates[s][i].row,
+                                 bc + candidates[s][i].col };
+            if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
+            thissad =
+                vfp->sdf(what->buf, what->stride,
+                         get_buf_from_mv(in_what, &this_mv), in_what->stride);
+            CHECK_BETTER
+          }
+        }
+
+        if (best_site == -1) {
+          continue;
+        } else {
+          br += candidates[s][best_site].row;
+          bc += candidates[s][best_site].col;
+          k = best_site;
+        }
+      }
+
+      do {
+        int next_chkpts_indices[PATTERN_CANDIDATES_REF];
+        best_site = -1;
+        next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
+        next_chkpts_indices[1] = k;
+        next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
+
+        if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
+          for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
+            const MV this_mv = {
+              br + candidates[s][next_chkpts_indices[i]].row,
+              bc + candidates[s][next_chkpts_indices[i]].col
+            };
+            thissad =
+                vfp->sdf(what->buf, what->stride,
+                         get_buf_from_mv(in_what, &this_mv), in_what->stride);
+            CHECK_BETTER
+          }
+        } else {
+          for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
+            const MV this_mv = {
+              br + candidates[s][next_chkpts_indices[i]].row,
+              bc + candidates[s][next_chkpts_indices[i]].col
+            };
+            if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
+            thissad =
+                vfp->sdf(what->buf, what->stride,
+                         get_buf_from_mv(in_what, &this_mv), in_what->stride);
+            CHECK_BETTER
+          }
+        }
+
+        if (best_site != -1) {
+          k = next_chkpts_indices[best_site];
+          br += candidates[s][k].row;
+          bc += candidates[s][k].col;
+        }
+      } while (best_site != -1);
+    }
+
+    // Note: If we enter the if below, then cost_list must be non-NULL.
+    if (s == 0) {
+      cost_list[0] = bestsad;
+      if (!do_init_search || s != best_init_s) {
+        if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
+          for (i = 0; i < num_candidates[s]; i++) {
+            const MV this_mv = { br + candidates[s][i].row,
+                                 bc + candidates[s][i].col };
+            cost_list[i + 1] = thissad =
+                vfp->sdf(what->buf, what->stride,
+                         get_buf_from_mv(in_what, &this_mv), in_what->stride);
+            CHECK_BETTER
+          }
+        } else {
+          for (i = 0; i < num_candidates[s]; i++) {
+            const MV this_mv = { br + candidates[s][i].row,
+                                 bc + candidates[s][i].col };
+            if (!is_mv_in(&x->mv_limits, &this_mv)) continue;
+            cost_list[i + 1] = thissad =
+                vfp->sdf(what->buf, what->stride,
+                         get_buf_from_mv(in_what, &this_mv), in_what->stride);
+            CHECK_BETTER
+          }
+        }
+
+        if (best_site != -1) {
+          br += candidates[s][best_site].row;
+          bc += candidates[s][best_site].col;
+          k = best_site;
+        }
+      }
+      while (best_site != -1) {
+        int next_chkpts_indices[PATTERN_CANDIDATES_REF];
+        best_site = -1;
+        next_chkpts_indices[0] = (k == 0) ? num_candidates[s] - 1 : k - 1;
+        next_chkpts_indices[1] = k;
+        next_chkpts_indices[2] = (k == num_candidates[s] - 1) ? 0 : k + 1;
+        cost_list[1] = cost_list[2] = cost_list[3] = cost_list[4] = INT_MAX;
+        cost_list[((k + 2) % 4) + 1] = cost_list[0];
+        cost_list[0] = bestsad;
+
+        if (check_bounds(&x->mv_limits, br, bc, 1 << s)) {
+          for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
+            const MV this_mv = {
+              br + candidates[s][next_chkpts_indices[i]].row,
+              bc + candidates[s][next_chkpts_indices[i]].col
+            };
+            cost_list[next_chkpts_indices[i] + 1] = thissad =
+                vfp->sdf(what->buf, what->stride,
+                         get_buf_from_mv(in_what, &this_mv), in_what->stride);
+            CHECK_BETTER
+          }
+        } else {
+          for (i = 0; i < PATTERN_CANDIDATES_REF; i++) {
+            const MV this_mv = {
+              br + candidates[s][next_chkpts_indices[i]].row,
+              bc + candidates[s][next_chkpts_indices[i]].col
+            };
+            if (!is_mv_in(&x->mv_limits, &this_mv)) {
+              cost_list[next_chkpts_indices[i] + 1] = INT_MAX;
+              continue;
+            }
+            cost_list[next_chkpts_indices[i] + 1] = thissad =
+                vfp->sdf(what->buf, what->stride,
+                         get_buf_from_mv(in_what, &this_mv), in_what->stride);
+            CHECK_BETTER
+          }
+        }
+
+        if (best_site != -1) {
+          k = next_chkpts_indices[best_site];
+          br += candidates[s][k].row;
+          bc += candidates[s][k].col;
+        }
+      }
+    }
+  }
+
+  // Returns the one-away integer pel cost/sad around the best as follows:
+  // cost_list[0]: cost/sad at the best integer pel
+  // cost_list[1]: cost/sad at delta {0, -1} (left)   from the best integer pel
+  // cost_list[2]: cost/sad at delta { 1, 0} (bottom) from the best integer pel
+  // cost_list[3]: cost/sad at delta { 0, 1} (right)  from the best integer pel
+  // cost_list[4]: cost/sad at delta {-1, 0} (top)    from the best integer pel
+  if (cost_list) {
+    const MV best_int_mv = { br, bc };
+    if (last_is_4) {
+      calc_int_sad_list(x, center_mv, sad_per_bit, vfp, &best_int_mv, cost_list,
+                        use_mvcost, bestsad);
+    } else {
+      calc_int_cost_list(x, center_mv, sad_per_bit, vfp, &best_int_mv,
+                         cost_list);
+    }
+  }
+  x->best_mv.as_mv.row = br;
+  x->best_mv.as_mv.col = bc;
+  return bestsad;
+}
+
+int av1_get_mvpred_var(const MACROBLOCK *x, const MV *best_mv,
+                       const MV *center_mv, const aom_variance_fn_ptr_t *vfp,
+                       int use_mvcost) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const MV mv = { best_mv->row * 8, best_mv->col * 8 };
+  unsigned int unused;
+
+  return vfp->vf(what->buf, what->stride, get_buf_from_mv(in_what, best_mv),
+                 in_what->stride, &unused) +
+         (use_mvcost ? mv_err_cost(&mv, center_mv, x->nmvjointcost, x->mvcost,
+                                   x->errorperbit)
+                     : 0);
+}
+
+int av1_get_mvpred_av_var(const MACROBLOCK *x, const MV *best_mv,
+                          const MV *center_mv, const uint8_t *second_pred,
+                          const aom_variance_fn_ptr_t *vfp, int use_mvcost) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const MV mv = { best_mv->row * 8, best_mv->col * 8 };
+  unsigned int unused;
+
+  return vfp->svaf(get_buf_from_mv(in_what, best_mv), in_what->stride, 0, 0,
+                   what->buf, what->stride, &unused, second_pred) +
+         (use_mvcost ? mv_err_cost(&mv, center_mv, x->nmvjointcost, x->mvcost,
+                                   x->errorperbit)
+                     : 0);
+}
+
+int av1_hex_search(MACROBLOCK *x, MV *start_mv, int search_param,
+                   int sad_per_bit, int do_init_search, int *cost_list,
+                   const aom_variance_fn_ptr_t *vfp, int use_mvcost,
+                   const MV *center_mv) {
+  // First scale has 8-closest points, the rest have 6 points in hex shape
+  // at increasing scales
+  static const int hex_num_candidates[MAX_PATTERN_SCALES] = { 8, 6, 6, 6, 6, 6,
+                                                              6, 6, 6, 6, 6 };
+  // Note that the largest candidate step at each scale is 2^scale
+  /* clang-format off */
+  static const MV hex_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = {
+    { { -1, -1 }, { 0, -1 }, { 1, -1 }, { 1, 0 }, { 1, 1 }, { 0, 1 }, { -1, 1 },
+      { -1, 0 } },
+    { { -1, -2 }, { 1, -2 }, { 2, 0 }, { 1, 2 }, { -1, 2 }, { -2, 0 } },
+    { { -2, -4 }, { 2, -4 }, { 4, 0 }, { 2, 4 }, { -2, 4 }, { -4, 0 } },
+    { { -4, -8 }, { 4, -8 }, { 8, 0 }, { 4, 8 }, { -4, 8 }, { -8, 0 } },
+    { { -8, -16 }, { 8, -16 }, { 16, 0 }, { 8, 16 }, { -8, 16 }, { -16, 0 } },
+    { { -16, -32 }, { 16, -32 }, { 32, 0 }, { 16, 32 }, { -16, 32 },
+      { -32, 0 } },
+    { { -32, -64 }, { 32, -64 }, { 64, 0 }, { 32, 64 }, { -32, 64 },
+      { -64, 0 } },
+    { { -64, -128 }, { 64, -128 }, { 128, 0 }, { 64, 128 }, { -64, 128 },
+      { -128, 0 } },
+    { { -128, -256 }, { 128, -256 }, { 256, 0 }, { 128, 256 }, { -128, 256 },
+      { -256, 0 } },
+    { { -256, -512 }, { 256, -512 }, { 512, 0 }, { 256, 512 }, { -256, 512 },
+      { -512, 0 } },
+    { { -512, -1024 }, { 512, -1024 }, { 1024, 0 }, { 512, 1024 },
+      { -512, 1024 }, { -1024, 0 } },
+  };
+  /* clang-format on */
+  return pattern_search(x, start_mv, search_param, sad_per_bit, do_init_search,
+                        cost_list, vfp, use_mvcost, center_mv,
+                        hex_num_candidates, hex_candidates);
+}
+
+static int bigdia_search(MACROBLOCK *x, MV *start_mv, int search_param,
+                         int sad_per_bit, int do_init_search, int *cost_list,
+                         const aom_variance_fn_ptr_t *vfp, int use_mvcost,
+                         const MV *center_mv) {
+  // First scale has 4-closest points, the rest have 8 points in diamond
+  // shape at increasing scales
+  static const int bigdia_num_candidates[MAX_PATTERN_SCALES] = {
+    4, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  };
+  // Note that the largest candidate step at each scale is 2^scale
+  /* clang-format off */
+  static const MV
+      bigdia_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = {
+        { { 0, -1 }, { 1, 0 }, { 0, 1 }, { -1, 0 } },
+        { { -1, -1 }, { 0, -2 }, { 1, -1 }, { 2, 0 }, { 1, 1 }, { 0, 2 },
+          { -1, 1 }, { -2, 0 } },
+        { { -2, -2 }, { 0, -4 }, { 2, -2 }, { 4, 0 }, { 2, 2 }, { 0, 4 },
+          { -2, 2 }, { -4, 0 } },
+        { { -4, -4 }, { 0, -8 }, { 4, -4 }, { 8, 0 }, { 4, 4 }, { 0, 8 },
+          { -4, 4 }, { -8, 0 } },
+        { { -8, -8 }, { 0, -16 }, { 8, -8 }, { 16, 0 }, { 8, 8 }, { 0, 16 },
+          { -8, 8 }, { -16, 0 } },
+        { { -16, -16 }, { 0, -32 }, { 16, -16 }, { 32, 0 }, { 16, 16 },
+          { 0, 32 }, { -16, 16 }, { -32, 0 } },
+        { { -32, -32 }, { 0, -64 }, { 32, -32 }, { 64, 0 }, { 32, 32 },
+          { 0, 64 }, { -32, 32 }, { -64, 0 } },
+        { { -64, -64 }, { 0, -128 }, { 64, -64 }, { 128, 0 }, { 64, 64 },
+          { 0, 128 }, { -64, 64 }, { -128, 0 } },
+        { { -128, -128 }, { 0, -256 }, { 128, -128 }, { 256, 0 }, { 128, 128 },
+          { 0, 256 }, { -128, 128 }, { -256, 0 } },
+        { { -256, -256 }, { 0, -512 }, { 256, -256 }, { 512, 0 }, { 256, 256 },
+          { 0, 512 }, { -256, 256 }, { -512, 0 } },
+        { { -512, -512 }, { 0, -1024 }, { 512, -512 }, { 1024, 0 },
+          { 512, 512 }, { 0, 1024 }, { -512, 512 }, { -1024, 0 } },
+      };
+  /* clang-format on */
+  return pattern_search(x, start_mv, search_param, sad_per_bit, do_init_search,
+                        cost_list, vfp, use_mvcost, center_mv,
+                        bigdia_num_candidates, bigdia_candidates);
+}
+
+static int square_search(MACROBLOCK *x, MV *start_mv, int search_param,
+                         int sad_per_bit, int do_init_search, int *cost_list,
+                         const aom_variance_fn_ptr_t *vfp, int use_mvcost,
+                         const MV *center_mv) {
+  // All scales have 8 closest points in square shape
+  static const int square_num_candidates[MAX_PATTERN_SCALES] = {
+    8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  };
+  // Note that the largest candidate step at each scale is 2^scale
+  /* clang-format off */
+  static const MV
+      square_candidates[MAX_PATTERN_SCALES][MAX_PATTERN_CANDIDATES] = {
+        { { -1, -1 }, { 0, -1 }, { 1, -1 }, { 1, 0 }, { 1, 1 }, { 0, 1 },
+          { -1, 1 }, { -1, 0 } },
+        { { -2, -2 }, { 0, -2 }, { 2, -2 }, { 2, 0 }, { 2, 2 }, { 0, 2 },
+          { -2, 2 }, { -2, 0 } },
+        { { -4, -4 }, { 0, -4 }, { 4, -4 }, { 4, 0 }, { 4, 4 }, { 0, 4 },
+          { -4, 4 }, { -4, 0 } },
+        { { -8, -8 }, { 0, -8 }, { 8, -8 }, { 8, 0 }, { 8, 8 }, { 0, 8 },
+          { -8, 8 }, { -8, 0 } },
+        { { -16, -16 }, { 0, -16 }, { 16, -16 }, { 16, 0 }, { 16, 16 },
+          { 0, 16 }, { -16, 16 }, { -16, 0 } },
+        { { -32, -32 }, { 0, -32 }, { 32, -32 }, { 32, 0 }, { 32, 32 },
+          { 0, 32 }, { -32, 32 }, { -32, 0 } },
+        { { -64, -64 }, { 0, -64 }, { 64, -64 }, { 64, 0 }, { 64, 64 },
+          { 0, 64 }, { -64, 64 }, { -64, 0 } },
+        { { -128, -128 }, { 0, -128 }, { 128, -128 }, { 128, 0 }, { 128, 128 },
+          { 0, 128 }, { -128, 128 }, { -128, 0 } },
+        { { -256, -256 }, { 0, -256 }, { 256, -256 }, { 256, 0 }, { 256, 256 },
+          { 0, 256 }, { -256, 256 }, { -256, 0 } },
+        { { -512, -512 }, { 0, -512 }, { 512, -512 }, { 512, 0 }, { 512, 512 },
+          { 0, 512 }, { -512, 512 }, { -512, 0 } },
+        { { -1024, -1024 }, { 0, -1024 }, { 1024, -1024 }, { 1024, 0 },
+          { 1024, 1024 }, { 0, 1024 }, { -1024, 1024 }, { -1024, 0 } },
+      };
+  /* clang-format on */
+  return pattern_search(x, start_mv, search_param, sad_per_bit, do_init_search,
+                        cost_list, vfp, use_mvcost, center_mv,
+                        square_num_candidates, square_candidates);
+}
+
+static int fast_hex_search(MACROBLOCK *x, MV *ref_mv, int search_param,
+                           int sad_per_bit,
+                           int do_init_search,  // must be zero for fast_hex
+                           int *cost_list, const aom_variance_fn_ptr_t *vfp,
+                           int use_mvcost, const MV *center_mv) {
+  return av1_hex_search(x, ref_mv, AOMMAX(MAX_MVSEARCH_STEPS - 2, search_param),
+                        sad_per_bit, do_init_search, cost_list, vfp, use_mvcost,
+                        center_mv);
+}
+
+static int fast_dia_search(MACROBLOCK *x, MV *ref_mv, int search_param,
+                           int sad_per_bit, int do_init_search, int *cost_list,
+                           const aom_variance_fn_ptr_t *vfp, int use_mvcost,
+                           const MV *center_mv) {
+  return bigdia_search(x, ref_mv, AOMMAX(MAX_MVSEARCH_STEPS - 2, search_param),
+                       sad_per_bit, do_init_search, cost_list, vfp, use_mvcost,
+                       center_mv);
+}
+
+#undef CHECK_BETTER
+
+// Exhuastive motion search around a given centre position with a given
+// step size.
+static int exhuastive_mesh_search(MACROBLOCK *x, MV *ref_mv, MV *best_mv,
+                                  int range, int step, int sad_per_bit,
+                                  const aom_variance_fn_ptr_t *fn_ptr,
+                                  const MV *center_mv) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  MV fcenter_mv = { center_mv->row, center_mv->col };
+  unsigned int best_sad = INT_MAX;
+  int r, c, i;
+  int start_col, end_col, start_row, end_row;
+  int col_step = (step > 1) ? step : 4;
+
+  assert(step >= 1);
+
+  clamp_mv(&fcenter_mv, x->mv_limits.col_min, x->mv_limits.col_max,
+           x->mv_limits.row_min, x->mv_limits.row_max);
+  *best_mv = fcenter_mv;
+  best_sad =
+      fn_ptr->sdf(what->buf, what->stride,
+                  get_buf_from_mv(in_what, &fcenter_mv), in_what->stride) +
+      mvsad_err_cost(x, &fcenter_mv, ref_mv, sad_per_bit);
+  start_row = AOMMAX(-range, x->mv_limits.row_min - fcenter_mv.row);
+  start_col = AOMMAX(-range, x->mv_limits.col_min - fcenter_mv.col);
+  end_row = AOMMIN(range, x->mv_limits.row_max - fcenter_mv.row);
+  end_col = AOMMIN(range, x->mv_limits.col_max - fcenter_mv.col);
+
+  for (r = start_row; r <= end_row; r += step) {
+    for (c = start_col; c <= end_col; c += col_step) {
+      // Step > 1 means we are not checking every location in this pass.
+      if (step > 1) {
+        const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c };
+        unsigned int sad =
+            fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, &mv),
+                        in_what->stride);
+        if (sad < best_sad) {
+          sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
+          if (sad < best_sad) {
+            best_sad = sad;
+            x->second_best_mv.as_mv = *best_mv;
+            *best_mv = mv;
+          }
+        }
+      } else {
+        // 4 sads in a single call if we are checking every location
+        if (c + 3 <= end_col) {
+          unsigned int sads[4];
+          const uint8_t *addrs[4];
+          for (i = 0; i < 4; ++i) {
+            const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c + i };
+            addrs[i] = get_buf_from_mv(in_what, &mv);
+          }
+          fn_ptr->sdx4df(what->buf, what->stride, addrs, in_what->stride, sads);
+
+          for (i = 0; i < 4; ++i) {
+            if (sads[i] < best_sad) {
+              const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c + i };
+              const unsigned int sad =
+                  sads[i] + mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
+              if (sad < best_sad) {
+                best_sad = sad;
+                x->second_best_mv.as_mv = *best_mv;
+                *best_mv = mv;
+              }
+            }
+          }
+        } else {
+          for (i = 0; i < end_col - c; ++i) {
+            const MV mv = { fcenter_mv.row + r, fcenter_mv.col + c + i };
+            unsigned int sad =
+                fn_ptr->sdf(what->buf, what->stride,
+                            get_buf_from_mv(in_what, &mv), in_what->stride);
+            if (sad < best_sad) {
+              sad += mvsad_err_cost(x, &mv, ref_mv, sad_per_bit);
+              if (sad < best_sad) {
+                best_sad = sad;
+                x->second_best_mv.as_mv = *best_mv;
+                *best_mv = mv;
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  return best_sad;
+}
+
+int av1_diamond_search_sad_c(MACROBLOCK *x, const search_site_config *cfg,
+                             MV *ref_mv, MV *best_mv, int search_param,
+                             int sad_per_bit, int *num00,
+                             const aom_variance_fn_ptr_t *fn_ptr,
+                             const MV *center_mv) {
+  int i, j, step;
+
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  uint8_t *what = x->plane[0].src.buf;
+  const int what_stride = x->plane[0].src.stride;
+  const uint8_t *in_what;
+  const int in_what_stride = xd->plane[0].pre[0].stride;
+  const uint8_t *best_address;
+
+  unsigned int bestsad = INT_MAX;
+  int best_site = 0;
+  int last_site = 0;
+
+  int ref_row;
+  int ref_col;
+
+  // search_param determines the length of the initial step and hence the number
+  // of iterations.
+  // 0 = initial step (MAX_FIRST_STEP) pel
+  // 1 = (MAX_FIRST_STEP/2) pel,
+  // 2 = (MAX_FIRST_STEP/4) pel...
+  const search_site *ss = &cfg->ss[search_param * cfg->searches_per_step];
+  const int tot_steps = (cfg->ss_count / cfg->searches_per_step) - search_param;
+
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
+           x->mv_limits.row_min, x->mv_limits.row_max);
+  ref_row = ref_mv->row;
+  ref_col = ref_mv->col;
+  *num00 = 0;
+  best_mv->row = ref_row;
+  best_mv->col = ref_col;
+
+  // Work out the start point for the search
+  in_what = xd->plane[0].pre[0].buf + ref_row * in_what_stride + ref_col;
+  best_address = in_what;
+
+  // Check the starting position
+  bestsad = fn_ptr->sdf(what, what_stride, in_what, in_what_stride) +
+            mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit);
+
+  i = 1;
+
+  for (step = 0; step < tot_steps; step++) {
+    int all_in = 1, t;
+
+    // All_in is true if every one of the points we are checking are within
+    // the bounds of the image.
+    all_in &= ((best_mv->row + ss[i].mv.row) > x->mv_limits.row_min);
+    all_in &= ((best_mv->row + ss[i + 1].mv.row) < x->mv_limits.row_max);
+    all_in &= ((best_mv->col + ss[i + 2].mv.col) > x->mv_limits.col_min);
+    all_in &= ((best_mv->col + ss[i + 3].mv.col) < x->mv_limits.col_max);
+
+    // If all the pixels are within the bounds we don't check whether the
+    // search point is valid in this loop,  otherwise we check each point
+    // for validity..
+    if (all_in) {
+      unsigned int sad_array[4];
+
+      for (j = 0; j < cfg->searches_per_step; j += 4) {
+        unsigned char const *block_offset[4];
+
+        for (t = 0; t < 4; t++)
+          block_offset[t] = ss[i + t].offset + best_address;
+
+        fn_ptr->sdx4df(what, what_stride, block_offset, in_what_stride,
+                       sad_array);
+
+        for (t = 0; t < 4; t++, i++) {
+          if (sad_array[t] < bestsad) {
+            const MV this_mv = { best_mv->row + ss[i].mv.row,
+                                 best_mv->col + ss[i].mv.col };
+            sad_array[t] +=
+                mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
+            if (sad_array[t] < bestsad) {
+              bestsad = sad_array[t];
+              best_site = i;
+            }
+          }
+        }
+      }
+    } else {
+      for (j = 0; j < cfg->searches_per_step; j++) {
+        // Trap illegal vectors
+        const MV this_mv = { best_mv->row + ss[i].mv.row,
+                             best_mv->col + ss[i].mv.col };
+
+        if (is_mv_in(&x->mv_limits, &this_mv)) {
+          const uint8_t *const check_here = ss[i].offset + best_address;
+          unsigned int thissad =
+              fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
+
+          if (thissad < bestsad) {
+            thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
+            if (thissad < bestsad) {
+              bestsad = thissad;
+              best_site = i;
+            }
+          }
+        }
+        i++;
+      }
+    }
+    if (best_site != last_site) {
+      x->second_best_mv.as_mv = *best_mv;
+      best_mv->row += ss[best_site].mv.row;
+      best_mv->col += ss[best_site].mv.col;
+      best_address += ss[best_site].offset;
+      last_site = best_site;
+#if defined(NEW_DIAMOND_SEARCH)
+      while (1) {
+        const MV this_mv = { best_mv->row + ss[best_site].mv.row,
+                             best_mv->col + ss[best_site].mv.col };
+        if (is_mv_in(&x->mv_limits, &this_mv)) {
+          const uint8_t *const check_here = ss[best_site].offset + best_address;
+          unsigned int thissad =
+              fn_ptr->sdf(what, what_stride, check_here, in_what_stride);
+          if (thissad < bestsad) {
+            thissad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
+            if (thissad < bestsad) {
+              bestsad = thissad;
+              best_mv->row += ss[best_site].mv.row;
+              best_mv->col += ss[best_site].mv.col;
+              best_address += ss[best_site].offset;
+              continue;
+            }
+          }
+        }
+        break;
+      }
+#endif
+    } else if (best_address == in_what) {
+      (*num00)++;
+    }
+  }
+  return bestsad;
+}
+
+static int vector_match(int16_t *ref, int16_t *src, int bwl) {
+  int best_sad = INT_MAX;
+  int this_sad;
+  int d;
+  int center, offset = 0;
+  int bw = 4 << bwl;  // redundant variable, to be changed in the experiments.
+  for (d = 0; d <= bw; d += 16) {
+    this_sad = aom_vector_var(&ref[d], src, bwl);
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      offset = d;
+    }
+  }
+  center = offset;
+
+  for (d = -8; d <= 8; d += 16) {
+    int this_pos = offset + d;
+    // check limit
+    if (this_pos < 0 || this_pos > bw) continue;
+    this_sad = aom_vector_var(&ref[this_pos], src, bwl);
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      center = this_pos;
+    }
+  }
+  offset = center;
+
+  for (d = -4; d <= 4; d += 8) {
+    int this_pos = offset + d;
+    // check limit
+    if (this_pos < 0 || this_pos > bw) continue;
+    this_sad = aom_vector_var(&ref[this_pos], src, bwl);
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      center = this_pos;
+    }
+  }
+  offset = center;
+
+  for (d = -2; d <= 2; d += 4) {
+    int this_pos = offset + d;
+    // check limit
+    if (this_pos < 0 || this_pos > bw) continue;
+    this_sad = aom_vector_var(&ref[this_pos], src, bwl);
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      center = this_pos;
+    }
+  }
+  offset = center;
+
+  for (d = -1; d <= 1; d += 2) {
+    int this_pos = offset + d;
+    // check limit
+    if (this_pos < 0 || this_pos > bw) continue;
+    this_sad = aom_vector_var(&ref[this_pos], src, bwl);
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      center = this_pos;
+    }
+  }
+
+  return (center - (bw >> 1));
+}
+
+static const MV search_pos[4] = {
+  { -1, 0 }, { 0, -1 }, { 0, 1 }, { 1, 0 },
+};
+
+unsigned int av1_int_pro_motion_estimation(const AV1_COMP *cpi, MACROBLOCK *x,
+                                           BLOCK_SIZE bsize, int mi_row,
+                                           int mi_col) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0, 0, 0, 0 } };
+  DECLARE_ALIGNED(16, int16_t, hbuf[2 * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, int16_t, vbuf[2 * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, int16_t, src_hbuf[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(16, int16_t, src_vbuf[MAX_SB_SQUARE]);
+  int idx;
+  const int bw = 4 << b_width_log2_lookup[bsize];
+  const int bh = 4 << b_height_log2_lookup[bsize];
+  const int search_width = bw << 1;
+  const int search_height = bh << 1;
+  const int src_stride = x->plane[0].src.stride;
+  const int ref_stride = xd->plane[0].pre[0].stride;
+  uint8_t const *ref_buf, *src_buf;
+  MV *tmp_mv = &xd->mi[0]->mbmi.mv[0].as_mv;
+  unsigned int best_sad, tmp_sad, sad_arr[4];
+  MV this_mv;
+  const int norm_factor = 3 + (bw >> 5);
+  const YV12_BUFFER_CONFIG *scaled_ref_frame =
+      av1_get_scaled_ref_frame(cpi, mbmi->ref_frame[0]);
+
+  if (scaled_ref_frame) {
+    int i;
+    // Swap out the reference frame for a version that's been scaled to
+    // match the resolution of the current frame, allowing the existing
+    // motion search code to be used without additional modifications.
+    for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
+    av1_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
+  }
+
+#if CONFIG_HIGHBITDEPTH
+  {
+    unsigned int this_sad;
+    tmp_mv->row = 0;
+    tmp_mv->col = 0;
+    this_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf, src_stride,
+                                      xd->plane[0].pre[0].buf, ref_stride);
+
+    if (scaled_ref_frame) {
+      int i;
+      for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
+    }
+    return this_sad;
+  }
+#endif
+
+  // Set up prediction 1-D reference set
+  ref_buf = xd->plane[0].pre[0].buf - (bw >> 1);
+  for (idx = 0; idx < search_width; idx += 16) {
+    aom_int_pro_row(&hbuf[idx], ref_buf, ref_stride, bh);
+    ref_buf += 16;
+  }
+
+  ref_buf = xd->plane[0].pre[0].buf - (bh >> 1) * ref_stride;
+  for (idx = 0; idx < search_height; ++idx) {
+    vbuf[idx] = aom_int_pro_col(ref_buf, bw) >> norm_factor;
+    ref_buf += ref_stride;
+  }
+
+  // Set up src 1-D reference set
+  for (idx = 0; idx < bw; idx += 16) {
+    src_buf = x->plane[0].src.buf + idx;
+    aom_int_pro_row(&src_hbuf[idx], src_buf, src_stride, bh);
+  }
+
+  src_buf = x->plane[0].src.buf;
+  for (idx = 0; idx < bh; ++idx) {
+    src_vbuf[idx] = aom_int_pro_col(src_buf, bw) >> norm_factor;
+    src_buf += src_stride;
+  }
+
+  // Find the best match per 1-D search
+  tmp_mv->col = vector_match(hbuf, src_hbuf, b_width_log2_lookup[bsize]);
+  tmp_mv->row = vector_match(vbuf, src_vbuf, b_height_log2_lookup[bsize]);
+
+  this_mv = *tmp_mv;
+  src_buf = x->plane[0].src.buf;
+  ref_buf = xd->plane[0].pre[0].buf + this_mv.row * ref_stride + this_mv.col;
+  best_sad = cpi->fn_ptr[bsize].sdf(src_buf, src_stride, ref_buf, ref_stride);
+
+  {
+    const uint8_t *const pos[4] = {
+      ref_buf - ref_stride, ref_buf - 1, ref_buf + 1, ref_buf + ref_stride,
+    };
+
+    cpi->fn_ptr[bsize].sdx4df(src_buf, src_stride, pos, ref_stride, sad_arr);
+  }
+
+  for (idx = 0; idx < 4; ++idx) {
+    if (sad_arr[idx] < best_sad) {
+      best_sad = sad_arr[idx];
+      tmp_mv->row = search_pos[idx].row + this_mv.row;
+      tmp_mv->col = search_pos[idx].col + this_mv.col;
+    }
+  }
+
+  if (sad_arr[0] < sad_arr[3])
+    this_mv.row -= 1;
+  else
+    this_mv.row += 1;
+
+  if (sad_arr[1] < sad_arr[2])
+    this_mv.col -= 1;
+  else
+    this_mv.col += 1;
+
+  ref_buf = xd->plane[0].pre[0].buf + this_mv.row * ref_stride + this_mv.col;
+
+  tmp_sad = cpi->fn_ptr[bsize].sdf(src_buf, src_stride, ref_buf, ref_stride);
+  if (best_sad > tmp_sad) {
+    *tmp_mv = this_mv;
+    best_sad = tmp_sad;
+  }
+
+  tmp_mv->row *= 8;
+  tmp_mv->col *= 8;
+
+  if (scaled_ref_frame) {
+    int i;
+    for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
+  }
+
+  return best_sad;
+}
+
+/* do_refine: If last step (1-away) of n-step search doesn't pick the center
+              point as the best match, we will do a final 1-away diamond
+              refining search  */
+static int full_pixel_diamond(const AV1_COMP *const cpi, MACROBLOCK *x,
+                              MV *mvp_full, int step_param, int sadpb,
+                              int further_steps, int do_refine, int *cost_list,
+                              const aom_variance_fn_ptr_t *fn_ptr,
+                              const MV *ref_mv) {
+  MV temp_mv;
+  int thissme, n, num00 = 0;
+  int bestsme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
+                                        step_param, sadpb, &n, fn_ptr, ref_mv);
+  if (bestsme < INT_MAX)
+    bestsme = av1_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
+  x->best_mv.as_mv = temp_mv;
+
+  // If there won't be more n-step search, check to see if refining search is
+  // needed.
+  if (n > further_steps) do_refine = 0;
+
+  while (n < further_steps) {
+    ++n;
+
+    if (num00) {
+      num00--;
+    } else {
+      thissme = cpi->diamond_search_sad(x, &cpi->ss_cfg, mvp_full, &temp_mv,
+                                        step_param + n, sadpb, &num00, fn_ptr,
+                                        ref_mv);
+      if (thissme < INT_MAX)
+        thissme = av1_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
+
+      // check to see if refining search is needed.
+      if (num00 > further_steps - n) do_refine = 0;
+
+      if (thissme < bestsme) {
+        bestsme = thissme;
+        x->best_mv.as_mv = temp_mv;
+      }
+    }
+  }
+
+  // final 1-away diamond refining search
+  if (do_refine) {
+    const int search_range = 8;
+    MV best_mv = x->best_mv.as_mv;
+    thissme = av1_refining_search_sad(x, &best_mv, sadpb, search_range, fn_ptr,
+                                      ref_mv);
+    if (thissme < INT_MAX)
+      thissme = av1_get_mvpred_var(x, &best_mv, ref_mv, fn_ptr, 1);
+    if (thissme < bestsme) {
+      bestsme = thissme;
+      x->best_mv.as_mv = best_mv;
+    }
+  }
+
+  // Return cost list.
+  if (cost_list) {
+    calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, &x->best_mv.as_mv, cost_list);
+  }
+  return bestsme;
+}
+
+#define MIN_RANGE 7
+#define MAX_RANGE 256
+#define MIN_INTERVAL 1
+// Runs an limited range exhaustive mesh search using a pattern set
+// according to the encode speed profile.
+static int full_pixel_exhaustive(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                 const MV *centre_mv_full, int sadpb,
+                                 int *cost_list,
+                                 const aom_variance_fn_ptr_t *fn_ptr,
+                                 const MV *ref_mv, MV *dst_mv) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  MV temp_mv = { centre_mv_full->row, centre_mv_full->col };
+  MV f_ref_mv = { ref_mv->row >> 3, ref_mv->col >> 3 };
+  int bestsme;
+  int i;
+  int interval = sf->mesh_patterns[0].interval;
+  int range = sf->mesh_patterns[0].range;
+  int baseline_interval_divisor;
+
+  // Keep track of number of exhaustive calls (this frame in this thread).
+  ++(*x->ex_search_count_ptr);
+
+  // Trap illegal values for interval and range for this function.
+  if ((range < MIN_RANGE) || (range > MAX_RANGE) || (interval < MIN_INTERVAL) ||
+      (interval > range))
+    return INT_MAX;
+
+  baseline_interval_divisor = range / interval;
+
+  // Check size of proposed first range against magnitude of the centre
+  // value used as a starting point.
+  range = AOMMAX(range, (5 * AOMMAX(abs(temp_mv.row), abs(temp_mv.col))) / 4);
+  range = AOMMIN(range, MAX_RANGE);
+  interval = AOMMAX(interval, range / baseline_interval_divisor);
+
+  // initial search
+  bestsme = exhuastive_mesh_search(x, &f_ref_mv, &temp_mv, range, interval,
+                                   sadpb, fn_ptr, &temp_mv);
+
+  if ((interval > MIN_INTERVAL) && (range > MIN_RANGE)) {
+    // Progressive searches with range and step size decreasing each time
+    // till we reach a step size of 1. Then break out.
+    for (i = 1; i < MAX_MESH_STEP; ++i) {
+      // First pass with coarser step and longer range
+      bestsme = exhuastive_mesh_search(
+          x, &f_ref_mv, &temp_mv, sf->mesh_patterns[i].range,
+          sf->mesh_patterns[i].interval, sadpb, fn_ptr, &temp_mv);
+
+      if (sf->mesh_patterns[i].interval == 1) break;
+    }
+  }
+
+  if (bestsme < INT_MAX)
+    bestsme = av1_get_mvpred_var(x, &temp_mv, ref_mv, fn_ptr, 1);
+  *dst_mv = temp_mv;
+
+  // Return cost list.
+  if (cost_list) {
+    calc_int_cost_list(x, ref_mv, sadpb, fn_ptr, dst_mv, cost_list);
+  }
+  return bestsme;
+}
+
+int av1_full_search_sad_c(const MACROBLOCK *x, const MV *ref_mv,
+                          int sad_per_bit, int distance,
+                          const aom_variance_fn_ptr_t *fn_ptr,
+                          const MV *center_mv, MV *best_mv) {
+  int r, c;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const int row_min = AOMMAX(ref_mv->row - distance, x->mv_limits.row_min);
+  const int row_max = AOMMIN(ref_mv->row + distance, x->mv_limits.row_max);
+  const int col_min = AOMMAX(ref_mv->col - distance, x->mv_limits.col_min);
+  const int col_max = AOMMIN(ref_mv->col + distance, x->mv_limits.col_max);
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  int best_sad =
+      fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv),
+                  in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+  *best_mv = *ref_mv;
+
+  for (r = row_min; r < row_max; ++r) {
+    for (c = col_min; c < col_max; ++c) {
+      const MV mv = { r, c };
+      const int sad =
+          fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, &mv),
+                      in_what->stride) +
+          mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+      if (sad < best_sad) {
+        best_sad = sad;
+        *best_mv = mv;
+      }
+    }
+  }
+  return best_sad;
+}
+
+int av1_full_search_sadx3(const MACROBLOCK *x, const MV *ref_mv,
+                          int sad_per_bit, int distance,
+                          const aom_variance_fn_ptr_t *fn_ptr,
+                          const MV *center_mv, MV *best_mv) {
+  int r;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const int row_min = AOMMAX(ref_mv->row - distance, x->mv_limits.row_min);
+  const int row_max = AOMMIN(ref_mv->row + distance, x->mv_limits.row_max);
+  const int col_min = AOMMAX(ref_mv->col - distance, x->mv_limits.col_min);
+  const int col_max = AOMMIN(ref_mv->col + distance, x->mv_limits.col_max);
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  unsigned int best_sad =
+      fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv),
+                  in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+  *best_mv = *ref_mv;
+
+  for (r = row_min; r < row_max; ++r) {
+    int c = col_min;
+    const uint8_t *check_here = &in_what->buf[r * in_what->stride + c];
+
+    if (fn_ptr->sdx3f != NULL) {
+      while ((c + 2) < col_max) {
+        int i;
+        DECLARE_ALIGNED(16, uint32_t, sads[3]);
+
+        fn_ptr->sdx3f(what->buf, what->stride, check_here, in_what->stride,
+                      sads);
+
+        for (i = 0; i < 3; ++i) {
+          unsigned int sad = sads[i];
+          if (sad < best_sad) {
+            const MV mv = { r, c };
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              *best_mv = mv;
+            }
+          }
+          ++check_here;
+          ++c;
+        }
+      }
+    }
+
+    while (c < col_max) {
+      unsigned int sad =
+          fn_ptr->sdf(what->buf, what->stride, check_here, in_what->stride);
+      if (sad < best_sad) {
+        const MV mv = { r, c };
+        sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+        if (sad < best_sad) {
+          best_sad = sad;
+          *best_mv = mv;
+        }
+      }
+      ++check_here;
+      ++c;
+    }
+  }
+
+  return best_sad;
+}
+
+int av1_full_search_sadx8(const MACROBLOCK *x, const MV *ref_mv,
+                          int sad_per_bit, int distance,
+                          const aom_variance_fn_ptr_t *fn_ptr,
+                          const MV *center_mv, MV *best_mv) {
+  int r;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const int row_min = AOMMAX(ref_mv->row - distance, x->mv_limits.row_min);
+  const int row_max = AOMMIN(ref_mv->row + distance, x->mv_limits.row_max);
+  const int col_min = AOMMAX(ref_mv->col - distance, x->mv_limits.col_min);
+  const int col_max = AOMMIN(ref_mv->col + distance, x->mv_limits.col_max);
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  unsigned int best_sad =
+      fn_ptr->sdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv),
+                  in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, sad_per_bit);
+  *best_mv = *ref_mv;
+
+  for (r = row_min; r < row_max; ++r) {
+    int c = col_min;
+    const uint8_t *check_here = &in_what->buf[r * in_what->stride + c];
+
+    if (fn_ptr->sdx8f != NULL) {
+      while ((c + 7) < col_max) {
+        int i;
+        DECLARE_ALIGNED(16, uint32_t, sads[8]);
+
+        fn_ptr->sdx8f(what->buf, what->stride, check_here, in_what->stride,
+                      sads);
+
+        for (i = 0; i < 8; ++i) {
+          unsigned int sad = sads[i];
+          if (sad < best_sad) {
+            const MV mv = { r, c };
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              *best_mv = mv;
+            }
+          }
+          ++check_here;
+          ++c;
+        }
+      }
+    }
+
+    if (fn_ptr->sdx3f != NULL) {
+      while ((c + 2) < col_max) {
+        int i;
+        DECLARE_ALIGNED(16, uint32_t, sads[3]);
+
+        fn_ptr->sdx3f(what->buf, what->stride, check_here, in_what->stride,
+                      sads);
+
+        for (i = 0; i < 3; ++i) {
+          unsigned int sad = sads[i];
+          if (sad < best_sad) {
+            const MV mv = { r, c };
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              *best_mv = mv;
+            }
+          }
+          ++check_here;
+          ++c;
+        }
+      }
+    }
+
+    while (c < col_max) {
+      unsigned int sad =
+          fn_ptr->sdf(what->buf, what->stride, check_here, in_what->stride);
+      if (sad < best_sad) {
+        const MV mv = { r, c };
+        sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+        if (sad < best_sad) {
+          best_sad = sad;
+          *best_mv = mv;
+        }
+      }
+      ++check_here;
+      ++c;
+    }
+  }
+
+  return best_sad;
+}
+
+int av1_refining_search_sad(MACROBLOCK *x, MV *ref_mv, int error_per_bit,
+                            int search_range,
+                            const aom_variance_fn_ptr_t *fn_ptr,
+                            const MV *center_mv) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const MV neighbors[4] = { { -1, 0 }, { 0, -1 }, { 0, 1 }, { 1, 0 } };
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  const uint8_t *best_address = get_buf_from_mv(in_what, ref_mv);
+  unsigned int best_sad =
+      fn_ptr->sdf(what->buf, what->stride, best_address, in_what->stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
+  int i, j;
+
+  for (i = 0; i < search_range; i++) {
+    int best_site = -1;
+    const int all_in = ((ref_mv->row - 1) > x->mv_limits.row_min) &
+                       ((ref_mv->row + 1) < x->mv_limits.row_max) &
+                       ((ref_mv->col - 1) > x->mv_limits.col_min) &
+                       ((ref_mv->col + 1) < x->mv_limits.col_max);
+
+    if (all_in) {
+      unsigned int sads[4];
+      const uint8_t *const positions[4] = { best_address - in_what->stride,
+                                            best_address - 1, best_address + 1,
+                                            best_address + in_what->stride };
+
+      fn_ptr->sdx4df(what->buf, what->stride, positions, in_what->stride, sads);
+
+      for (j = 0; j < 4; ++j) {
+        if (sads[j] < best_sad) {
+          const MV mv = { ref_mv->row + neighbors[j].row,
+                          ref_mv->col + neighbors[j].col };
+          sads[j] += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
+          if (sads[j] < best_sad) {
+            best_sad = sads[j];
+            best_site = j;
+          }
+        }
+      }
+    } else {
+      for (j = 0; j < 4; ++j) {
+        const MV mv = { ref_mv->row + neighbors[j].row,
+                        ref_mv->col + neighbors[j].col };
+
+        if (is_mv_in(&x->mv_limits, &mv)) {
+          unsigned int sad =
+              fn_ptr->sdf(what->buf, what->stride,
+                          get_buf_from_mv(in_what, &mv), in_what->stride);
+          if (sad < best_sad) {
+            sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              best_site = j;
+            }
+          }
+        }
+      }
+    }
+
+    if (best_site == -1) {
+      break;
+    } else {
+      x->second_best_mv.as_mv = *ref_mv;
+      ref_mv->row += neighbors[best_site].row;
+      ref_mv->col += neighbors[best_site].col;
+      best_address = get_buf_from_mv(in_what, ref_mv);
+    }
+  }
+
+  return best_sad;
+}
+
+// This function is called when we do joint motion search in comp_inter_inter
+// mode.
+int av1_refining_search_8p_c(MACROBLOCK *x, int error_per_bit, int search_range,
+                             const aom_variance_fn_ptr_t *fn_ptr,
+                             const MV *center_mv, const uint8_t *second_pred) {
+  const MV neighbors[8] = { { -1, 0 },  { 0, -1 }, { 0, 1 },  { 1, 0 },
+                            { -1, -1 }, { 1, -1 }, { -1, 1 }, { 1, 1 } };
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[0];
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  MV *best_mv = &x->best_mv.as_mv;
+  unsigned int best_sad = INT_MAX;
+  int i, j;
+
+  clamp_mv(best_mv, x->mv_limits.col_min, x->mv_limits.col_max,
+           x->mv_limits.row_min, x->mv_limits.row_max);
+  best_sad =
+      fn_ptr->sdaf(what->buf, what->stride, get_buf_from_mv(in_what, best_mv),
+                   in_what->stride, second_pred) +
+      mvsad_err_cost(x, best_mv, &fcenter_mv, error_per_bit);
+
+  for (i = 0; i < search_range; ++i) {
+    int best_site = -1;
+
+    for (j = 0; j < 8; ++j) {
+      const MV mv = { best_mv->row + neighbors[j].row,
+                      best_mv->col + neighbors[j].col };
+
+      if (is_mv_in(&x->mv_limits, &mv)) {
+        unsigned int sad =
+            fn_ptr->sdaf(what->buf, what->stride, get_buf_from_mv(in_what, &mv),
+                         in_what->stride, second_pred);
+        if (sad < best_sad) {
+          sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
+          if (sad < best_sad) {
+            best_sad = sad;
+            best_site = j;
+          }
+        }
+      }
+    }
+
+    if (best_site == -1) {
+      break;
+    } else {
+      best_mv->row += neighbors[best_site].row;
+      best_mv->col += neighbors[best_site].col;
+    }
+  }
+  return best_sad;
+}
+
+#define MIN_EX_SEARCH_LIMIT 128
+static int is_exhaustive_allowed(const AV1_COMP *const cpi, MACROBLOCK *x) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  const int max_ex =
+      AOMMAX(MIN_EX_SEARCH_LIMIT,
+             (*x->m_search_count_ptr * sf->max_exaustive_pct) / 100);
+
+  return sf->allow_exhaustive_searches &&
+         (sf->exhaustive_searches_thresh < INT_MAX) &&
+         (*x->ex_search_count_ptr <= max_ex) && !cpi->rc.is_src_frame_alt_ref;
+}
+
+int av1_full_pixel_search(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
+                          MV *mvp_full, int step_param, int error_per_bit,
+                          int *cost_list, const MV *ref_mv, int var_max,
+                          int rd) {
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  const SEARCH_METHODS method = sf->mv.search_method;
+  const aom_variance_fn_ptr_t *fn_ptr = &cpi->fn_ptr[bsize];
+  int var = 0;
+
+  if (cost_list) {
+    cost_list[0] = INT_MAX;
+    cost_list[1] = INT_MAX;
+    cost_list[2] = INT_MAX;
+    cost_list[3] = INT_MAX;
+    cost_list[4] = INT_MAX;
+  }
+
+  // Keep track of number of searches (this frame in this thread).
+  ++(*x->m_search_count_ptr);
+
+  switch (method) {
+    case FAST_DIAMOND:
+      var = fast_dia_search(x, mvp_full, step_param, error_per_bit, 0,
+                            cost_list, fn_ptr, 1, ref_mv);
+      break;
+    case FAST_HEX:
+      var = fast_hex_search(x, mvp_full, step_param, error_per_bit, 0,
+                            cost_list, fn_ptr, 1, ref_mv);
+      break;
+    case HEX:
+      var = av1_hex_search(x, mvp_full, step_param, error_per_bit, 1, cost_list,
+                           fn_ptr, 1, ref_mv);
+      break;
+    case SQUARE:
+      var = square_search(x, mvp_full, step_param, error_per_bit, 1, cost_list,
+                          fn_ptr, 1, ref_mv);
+      break;
+    case BIGDIA:
+      var = bigdia_search(x, mvp_full, step_param, error_per_bit, 1, cost_list,
+                          fn_ptr, 1, ref_mv);
+      break;
+    case NSTEP:
+      var = full_pixel_diamond(cpi, x, mvp_full, step_param, error_per_bit,
+                               MAX_MVSEARCH_STEPS - 1 - step_param, 1,
+                               cost_list, fn_ptr, ref_mv);
+
+      // Should we allow a follow on exhaustive search?
+      if (is_exhaustive_allowed(cpi, x)) {
+        int exhuastive_thr = sf->exhaustive_searches_thresh;
+        exhuastive_thr >>=
+            10 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
+
+        // Threshold variance for an exhaustive full search.
+        if (var > exhuastive_thr) {
+          int var_ex;
+          MV tmp_mv_ex;
+          var_ex =
+              full_pixel_exhaustive(cpi, x, &x->best_mv.as_mv, error_per_bit,
+                                    cost_list, fn_ptr, ref_mv, &tmp_mv_ex);
+
+          if (var_ex < var) {
+            var = var_ex;
+            x->best_mv.as_mv = tmp_mv_ex;
+          }
+        }
+      }
+      break;
+
+      break;
+    default: assert(0 && "Invalid search method.");
+  }
+
+  if (method != NSTEP && rd && var < var_max)
+    var = av1_get_mvpred_var(x, &x->best_mv.as_mv, ref_mv, fn_ptr, 1);
+
+  return var;
+}
+
+#if CONFIG_EXT_INTER
+/* returns subpixel variance error function */
+#define DIST(r, c)                                                         \
+  vfp->msvf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z, src_stride, \
+            mask, mask_stride, &sse)
+
+/* checks if (r, c) has better score than previous best */
+
+#define MVC(r, c)                                                         \
+  (mvcost                                                                 \
+       ? ((mvjcost[((r) != rr) * 2 + ((c) != rc)] + mvcost[0][((r)-rr)] + \
+           mvcost[1][((c)-rc)]) *                                         \
+              error_per_bit +                                             \
+          4096) >>                                                        \
+             13                                                           \
+       : 0)
+
+#define CHECK_BETTER(v, r, c)                             \
+  if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
+    thismse = (DIST(r, c));                               \
+    if ((v = MVC(r, c) + thismse) < besterr) {            \
+      besterr = v;                                        \
+      br = r;                                             \
+      bc = c;                                             \
+      *distortion = thismse;                              \
+      *sse1 = sse;                                        \
+    }                                                     \
+  } else {                                                \
+    v = INT_MAX;                                          \
+  }
+
+#undef CHECK_BETTER0
+#define CHECK_BETTER0(v, r, c) CHECK_BETTER(v, r, c)
+
+#undef CHECK_BETTER1
+#define CHECK_BETTER1(v, r, c)                                                 \
+  if (c >= minc && c <= maxc && r >= minr && r <= maxr) {                      \
+    thismse = upsampled_masked_pref_error(xd, mask, mask_stride, vfp, z,       \
+                                          src_stride, upre(y, y_stride, r, c), \
+                                          y_stride, w, h, &sse);               \
+    if ((v = MVC(r, c) + thismse) < besterr) {                                 \
+      besterr = v;                                                             \
+      br = r;                                                                  \
+      bc = c;                                                                  \
+      *distortion = thismse;                                                   \
+      *sse1 = sse;                                                             \
+    }                                                                          \
+  } else {                                                                     \
+    v = INT_MAX;                                                               \
+  }
+
+int av1_find_best_masked_sub_pixel_tree(
+    const MACROBLOCK *x, const uint8_t *mask, int mask_stride, MV *bestmv,
+    const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, int forced_stop, int iters_per_step,
+    int *mvjcost, int *mvcost[2], int *distortion, unsigned int *sse1,
+    int is_second) {
+  const uint8_t *const z = x->plane[0].src.buf;
+  const int src_stride = x->plane[0].src.stride;
+  const MACROBLOCKD *xd = &x->e_mbd;
+  unsigned int besterr = INT_MAX;
+  unsigned int sse;
+  int thismse;
+  unsigned int whichdir;
+  unsigned int halfiters = iters_per_step;
+  unsigned int quarteriters = iters_per_step;
+  unsigned int eighthiters = iters_per_step;
+
+  const int y_stride = xd->plane[0].pre[is_second].stride;
+  const int offset = bestmv->row * y_stride + bestmv->col;
+  const uint8_t *const y = xd->plane[0].pre[is_second].buf;
+
+  int rr = ref_mv->row;
+  int rc = ref_mv->col;
+  int br = bestmv->row * 8;
+  int bc = bestmv->col * 8;
+  int hstep = 4;
+  int tr = br;
+  int tc = bc;
+  int minc, maxc, minr, maxr;
+
+  av1_set_subpel_mv_search_range(&x->mv_limits, &minc, &maxc, &minr, &maxr,
+                                 ref_mv);
+
+  // central mv
+  bestmv->row *= 8;
+  bestmv->col *= 8;
+
+  // calculate central point error
+  besterr =
+      vfp->mvf(y + offset, y_stride, z, src_stride, mask, mask_stride, sse1);
+  *distortion = besterr;
+  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
+
+  // 1/2 pel
+  FIRST_LEVEL_CHECKS;
+  if (halfiters > 1) {
+    SECOND_LEVEL_CHECKS;
+  }
+  tr = br;
+  tc = bc;
+
+  // Note forced_stop: 0 - full, 1 - qtr only, 2 - half only
+  if (forced_stop != 2) {
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (quarteriters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+    tr = br;
+    tc = bc;
+  }
+
+  if (allow_hp && forced_stop == 0) {
+    hstep >>= 1;
+    FIRST_LEVEL_CHECKS;
+    if (eighthiters > 1) {
+      SECOND_LEVEL_CHECKS;
+    }
+    tr = br;
+    tc = bc;
+  }
+  // These lines insure static analysis doesn't warn that
+  // tr and tc aren't used after the above point.
+  (void)tr;
+  (void)tc;
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  return besterr;
+}
+
+static unsigned int setup_masked_center_error(
+    const uint8_t *mask, int mask_stride, const MV *bestmv, const MV *ref_mv,
+    int error_per_bit, const aom_variance_fn_ptr_t *vfp,
+    const uint8_t *const src, const int src_stride, const uint8_t *const y,
+    int y_stride, int offset, int *mvjcost, int *mvcost[2], unsigned int *sse1,
+    int *distortion) {
+  unsigned int besterr;
+  besterr =
+      vfp->mvf(y + offset, y_stride, src, src_stride, mask, mask_stride, sse1);
+  *distortion = besterr;
+  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
+  return besterr;
+}
+
+static int upsampled_masked_pref_error(const MACROBLOCKD *xd,
+                                       const uint8_t *mask, int mask_stride,
+                                       const aom_variance_fn_ptr_t *vfp,
+                                       const uint8_t *const src,
+                                       const int src_stride,
+                                       const uint8_t *const y, int y_stride,
+                                       int w, int h, unsigned int *sse) {
+  unsigned int besterr;
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    DECLARE_ALIGNED(16, uint16_t, pred16[MAX_SB_SQUARE]);
+    aom_highbd_upsampled_pred(pred16, w, h, y, y_stride);
+
+    besterr = vfp->mvf(CONVERT_TO_BYTEPTR(pred16), w, src, src_stride, mask,
+                       mask_stride, sse);
+  } else {
+    DECLARE_ALIGNED(16, uint8_t, pred[MAX_SB_SQUARE]);
+#else
+  DECLARE_ALIGNED(16, uint8_t, pred[MAX_SB_SQUARE]);
+  (void)xd;
+#endif  // CONFIG_HIGHBITDEPTH
+    aom_upsampled_pred(pred, w, h, y, y_stride);
+
+    besterr = vfp->mvf(pred, w, src, src_stride, mask, mask_stride, sse);
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif
+  return besterr;
+}
+
+static unsigned int upsampled_setup_masked_center_error(
+    const MACROBLOCKD *xd, const uint8_t *mask, int mask_stride,
+    const MV *bestmv, const MV *ref_mv, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, const uint8_t *const src,
+    const int src_stride, const uint8_t *const y, int y_stride, int w, int h,
+    int offset, int *mvjcost, int *mvcost[2], unsigned int *sse1,
+    int *distortion) {
+  unsigned int besterr =
+      upsampled_masked_pref_error(xd, mask, mask_stride, vfp, src, src_stride,
+                                  y + offset, y_stride, w, h, sse1);
+  *distortion = besterr;
+  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
+  return besterr;
+}
+
+int av1_find_best_masked_sub_pixel_tree_up(
+    const AV1_COMP *cpi, MACROBLOCK *x, const uint8_t *mask, int mask_stride,
+    int mi_row, int mi_col, MV *bestmv, const MV *ref_mv, int allow_hp,
+    int error_per_bit, const aom_variance_fn_ptr_t *vfp, int forced_stop,
+    int iters_per_step, int *mvjcost, int *mvcost[2], int *distortion,
+    unsigned int *sse1, int is_second, int use_upsampled_ref) {
+  const uint8_t *const z = x->plane[0].src.buf;
+  const uint8_t *const src_address = z;
+  const int src_stride = x->plane[0].src.stride;
+  MACROBLOCKD *xd = &x->e_mbd;
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  unsigned int besterr = INT_MAX;
+  unsigned int sse;
+  unsigned int thismse;
+
+  int rr = ref_mv->row;
+  int rc = ref_mv->col;
+  int br = bestmv->row * 8;
+  int bc = bestmv->col * 8;
+  int hstep = 4;
+  int iter;
+  int round = 3 - forced_stop;
+  int tr = br;
+  int tc = bc;
+  const MV *search_step = search_step_table;
+  int idx, best_idx = -1;
+  unsigned int cost_array[5];
+  int kr, kc;
+  const int w = block_size_wide[mbmi->sb_type];
+  const int h = block_size_high[mbmi->sb_type];
+  int offset;
+  int y_stride;
+  const uint8_t *y;
+
+  const struct buf_2d backup_pred = pd->pre[is_second];
+  int minc, maxc, minr, maxr;
+
+  av1_set_subpel_mv_search_range(&x->mv_limits, &minc, &maxc, &minr, &maxr,
+                                 ref_mv);
+
+  if (use_upsampled_ref) {
+    int ref = xd->mi[0]->mbmi.ref_frame[is_second];
+    const YV12_BUFFER_CONFIG *upsampled_ref = get_upsampled_ref(cpi, ref);
+    setup_pred_plane(&pd->pre[is_second], mbmi->sb_type,
+                     upsampled_ref->y_buffer, upsampled_ref->y_crop_width,
+                     upsampled_ref->y_crop_height, upsampled_ref->y_stride,
+                     (mi_row << 3), (mi_col << 3), NULL, pd->subsampling_x,
+                     pd->subsampling_y);
+  }
+  y = pd->pre[is_second].buf;
+  y_stride = pd->pre[is_second].stride;
+  offset = bestmv->row * y_stride + bestmv->col;
+
+  if (!allow_hp)
+    if (round == 3) round = 2;
+
+  bestmv->row *= 8;
+  bestmv->col *= 8;
+
+  // use_upsampled_ref can be 0 or 1
+  if (use_upsampled_ref)
+    besterr = upsampled_setup_masked_center_error(
+        xd, mask, mask_stride, bestmv, ref_mv, error_per_bit, vfp, z,
+        src_stride, y, y_stride, w, h, (offset * 8), mvjcost, mvcost, sse1,
+        distortion);
+  else
+    besterr = setup_masked_center_error(
+        mask, mask_stride, bestmv, ref_mv, error_per_bit, vfp, z, src_stride, y,
+        y_stride, offset, mvjcost, mvcost, sse1, distortion);
+
+  for (iter = 0; iter < round; ++iter) {
+    // Check vertical and horizontal sub-pixel positions.
+    for (idx = 0; idx < 4; ++idx) {
+      tr = br + search_step[idx].row;
+      tc = bc + search_step[idx].col;
+      if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
+        MV this_mv = { tr, tc };
+
+        if (use_upsampled_ref) {
+          const uint8_t *const pre_address = y + tr * y_stride + tc;
+
+          thismse = upsampled_masked_pref_error(
+              xd, mask, mask_stride, vfp, src_address, src_stride, pre_address,
+              y_stride, w, h, &sse);
+        } else {
+          const uint8_t *const pre_address =
+              y + (tr >> 3) * y_stride + (tc >> 3);
+          thismse = vfp->msvf(pre_address, y_stride, sp(tc), sp(tr),
+                              src_address, src_stride, mask, mask_stride, &sse);
+        }
+
+        cost_array[idx] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost,
+                                                mvcost, error_per_bit);
+
+        if (cost_array[idx] < besterr) {
+          best_idx = idx;
+          besterr = cost_array[idx];
+          *distortion = thismse;
+          *sse1 = sse;
+        }
+      } else {
+        cost_array[idx] = INT_MAX;
+      }
+    }
+
+    // Check diagonal sub-pixel position
+    kc = (cost_array[0] <= cost_array[1] ? -hstep : hstep);
+    kr = (cost_array[2] <= cost_array[3] ? -hstep : hstep);
+
+    tc = bc + kc;
+    tr = br + kr;
+    if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
+      MV this_mv = { tr, tc };
+
+      if (use_upsampled_ref) {
+        const uint8_t *const pre_address = y + tr * y_stride + tc;
+
+        thismse = upsampled_masked_pref_error(
+            xd, mask, mask_stride, vfp, src_address, src_stride, pre_address,
+            y_stride, w, h, &sse);
+      } else {
+        const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
+
+        thismse = vfp->msvf(pre_address, y_stride, sp(tc), sp(tr), src_address,
+                            src_stride, mask, mask_stride, &sse);
+      }
+
+      cost_array[4] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost,
+                                            error_per_bit);
+
+      if (cost_array[4] < besterr) {
+        best_idx = 4;
+        besterr = cost_array[4];
+        *distortion = thismse;
+        *sse1 = sse;
+      }
+    } else {
+      cost_array[idx] = INT_MAX;
+    }
+
+    if (best_idx < 4 && best_idx >= 0) {
+      br += search_step[best_idx].row;
+      bc += search_step[best_idx].col;
+    } else if (best_idx == 4) {
+      br = tr;
+      bc = tc;
+    }
+
+    if (iters_per_step > 1 && best_idx != -1) {
+      if (use_upsampled_ref) {
+        SECOND_LEVEL_CHECKS_BEST(1);
+      } else {
+        SECOND_LEVEL_CHECKS_BEST(0);
+      }
+    }
+
+    tr = br;
+    tc = bc;
+
+    search_step += 4;
+    hstep >>= 1;
+    best_idx = -1;
+  }
+
+  // These lines insure static analysis doesn't warn that
+  // tr and tc aren't used after the above point.
+  (void)tr;
+  (void)tc;
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  if (use_upsampled_ref) {
+    pd->pre[is_second] = backup_pred;
+  }
+
+  return besterr;
+}
+
+#undef DIST
+#undef MVC
+#undef CHECK_BETTER
+
+static int get_masked_mvpred_var(const MACROBLOCK *x, const uint8_t *mask,
+                                 int mask_stride, const MV *best_mv,
+                                 const MV *center_mv,
+                                 const aom_variance_fn_ptr_t *vfp,
+                                 int use_mvcost, int is_second) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[is_second];
+  const MV mv = { best_mv->row * 8, best_mv->col * 8 };
+  unsigned int unused;
+
+  return vfp->mvf(what->buf, what->stride, get_buf_from_mv(in_what, best_mv),
+                  in_what->stride, mask, mask_stride, &unused) +
+         (use_mvcost ? mv_err_cost(&mv, center_mv, x->nmvjointcost, x->mvcost,
+                                   x->errorperbit)
+                     : 0);
+}
+
+int masked_refining_search_sad(const MACROBLOCK *x, const uint8_t *mask,
+                               int mask_stride, MV *ref_mv, int error_per_bit,
+                               int search_range,
+                               const aom_variance_fn_ptr_t *fn_ptr,
+                               const MV *center_mv, int is_second) {
+  const MV neighbors[4] = { { -1, 0 }, { 0, -1 }, { 0, 1 }, { 1, 0 } };
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[is_second];
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  unsigned int best_sad =
+      fn_ptr->msdf(what->buf, what->stride, get_buf_from_mv(in_what, ref_mv),
+                   in_what->stride, mask, mask_stride) +
+      mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
+  int i, j;
+
+  for (i = 0; i < search_range; i++) {
+    int best_site = -1;
+
+    for (j = 0; j < 4; j++) {
+      const MV mv = { ref_mv->row + neighbors[j].row,
+                      ref_mv->col + neighbors[j].col };
+      if (is_mv_in(&x->mv_limits, &mv)) {
+        unsigned int sad =
+            fn_ptr->msdf(what->buf, what->stride, get_buf_from_mv(in_what, &mv),
+                         in_what->stride, mask, mask_stride);
+        if (sad < best_sad) {
+          sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
+          if (sad < best_sad) {
+            best_sad = sad;
+            best_site = j;
+          }
+        }
+      }
+    }
+
+    if (best_site == -1) {
+      break;
+    } else {
+      ref_mv->row += neighbors[best_site].row;
+      ref_mv->col += neighbors[best_site].col;
+    }
+  }
+  return best_sad;
+}
+
+int masked_diamond_search_sad(const MACROBLOCK *x,
+                              const search_site_config *cfg,
+                              const uint8_t *mask, int mask_stride, MV *ref_mv,
+                              MV *best_mv, int search_param, int sad_per_bit,
+                              int *num00, const aom_variance_fn_ptr_t *fn_ptr,
+                              const MV *center_mv, int is_second) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const what = &x->plane[0].src;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[is_second];
+  // search_param determines the length of the initial step and hence the number
+  // of iterations
+  // 0 = initial step (MAX_FIRST_STEP) pel : 1 = (MAX_FIRST_STEP/2) pel, 2 =
+  // (MAX_FIRST_STEP/4) pel... etc.
+  const search_site *const ss = &cfg->ss[search_param * cfg->searches_per_step];
+  const int tot_steps = (cfg->ss_count / cfg->searches_per_step) - search_param;
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  const uint8_t *best_address, *in_what_ref;
+  int best_sad = INT_MAX;
+  int best_site = 0;
+  int last_site = 0;
+  int i, j, step;
+
+  clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
+           x->mv_limits.row_min, x->mv_limits.row_max);
+  in_what_ref = get_buf_from_mv(in_what, ref_mv);
+  best_address = in_what_ref;
+  *num00 = 0;
+  *best_mv = *ref_mv;
+
+  // Check the starting position
+  best_sad = fn_ptr->msdf(what->buf, what->stride, best_address,
+                          in_what->stride, mask, mask_stride) +
+             mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit);
+
+  i = 1;
+
+  for (step = 0; step < tot_steps; step++) {
+    for (j = 0; j < cfg->searches_per_step; j++) {
+      const MV mv = { best_mv->row + ss[i].mv.row,
+                      best_mv->col + ss[i].mv.col };
+      if (is_mv_in(&x->mv_limits, &mv)) {
+        int sad =
+            fn_ptr->msdf(what->buf, what->stride, best_address + ss[i].offset,
+                         in_what->stride, mask, mask_stride);
+        if (sad < best_sad) {
+          sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+          if (sad < best_sad) {
+            best_sad = sad;
+            best_site = i;
+          }
+        }
+      }
+
+      i++;
+    }
+
+    if (best_site != last_site) {
+      best_mv->row += ss[best_site].mv.row;
+      best_mv->col += ss[best_site].mv.col;
+      best_address += ss[best_site].offset;
+      last_site = best_site;
+#if defined(NEW_DIAMOND_SEARCH)
+      while (1) {
+        const MV this_mv = { best_mv->row + ss[best_site].mv.row,
+                             best_mv->col + ss[best_site].mv.col };
+        if (is_mv_in(&x->mv_limits, &this_mv)) {
+          int sad = fn_ptr->msdf(what->buf, what->stride,
+                                 best_address + ss[best_site].offset,
+                                 in_what->stride, mask, mask_stride);
+          if (sad < best_sad) {
+            sad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              best_mv->row += ss[best_site].mv.row;
+              best_mv->col += ss[best_site].mv.col;
+              best_address += ss[best_site].offset;
+              continue;
+            }
+          }
+        }
+        break;
+      }
+#endif
+    } else if (best_address == in_what_ref) {
+      (*num00)++;
+    }
+  }
+  return best_sad;
+}
+
+int av1_masked_full_pixel_diamond(const AV1_COMP *cpi, MACROBLOCK *x,
+                                  const uint8_t *mask, int mask_stride,
+                                  MV *mvp_full, int step_param, int sadpb,
+                                  int further_steps, int do_refine,
+                                  const aom_variance_fn_ptr_t *fn_ptr,
+                                  const MV *ref_mv, MV *dst_mv, int is_second) {
+  MV temp_mv;
+  int thissme, n, num00 = 0;
+  int bestsme = masked_diamond_search_sad(x, &cpi->ss_cfg, mask, mask_stride,
+                                          mvp_full, &temp_mv, step_param, sadpb,
+                                          &n, fn_ptr, ref_mv, is_second);
+  if (bestsme < INT_MAX)
+    bestsme = get_masked_mvpred_var(x, mask, mask_stride, &temp_mv, ref_mv,
+                                    fn_ptr, 1, is_second);
+  *dst_mv = temp_mv;
+
+  // If there won't be more n-step search, check to see if refining search is
+  // needed.
+  if (n > further_steps) do_refine = 0;
+
+  while (n < further_steps) {
+    ++n;
+
+    if (num00) {
+      num00--;
+    } else {
+      thissme = masked_diamond_search_sad(
+          x, &cpi->ss_cfg, mask, mask_stride, mvp_full, &temp_mv,
+          step_param + n, sadpb, &num00, fn_ptr, ref_mv, is_second);
+      if (thissme < INT_MAX)
+        thissme = get_masked_mvpred_var(x, mask, mask_stride, &temp_mv, ref_mv,
+                                        fn_ptr, 1, is_second);
+
+      // check to see if refining search is needed.
+      if (num00 > further_steps - n) do_refine = 0;
+
+      if (thissme < bestsme) {
+        bestsme = thissme;
+        *dst_mv = temp_mv;
+      }
+    }
+  }
+
+  // final 1-away diamond refining search
+  if (do_refine) {
+    const int search_range = 8;
+    MV best_mv = *dst_mv;
+    thissme =
+        masked_refining_search_sad(x, mask, mask_stride, &best_mv, sadpb,
+                                   search_range, fn_ptr, ref_mv, is_second);
+    if (thissme < INT_MAX)
+      thissme = get_masked_mvpred_var(x, mask, mask_stride, &best_mv, ref_mv,
+                                      fn_ptr, 1, is_second);
+    if (thissme < bestsme) {
+      bestsme = thissme;
+      *dst_mv = best_mv;
+    }
+  }
+  return bestsme;
+}
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR
+/* returns subpixel variance error function */
+#define DIST(r, c) \
+  vfp->osvf(pre(y, y_stride, r, c), y_stride, sp(c), sp(r), z, mask, &sse)
+
+/* checks if (r, c) has better score than previous best */
+#define MVC(r, c)                                                         \
+  (mvcost                                                                 \
+       ? ((mvjcost[((r) != rr) * 2 + ((c) != rc)] + mvcost[0][((r)-rr)] + \
+           mvcost[1][((c)-rc)]) *                                         \
+              error_per_bit +                                             \
+          4096) >>                                                        \
+             13                                                           \
+       : 0)
+
+#define CHECK_BETTER(v, r, c)                             \
+  if (c >= minc && c <= maxc && r >= minr && r <= maxr) { \
+    thismse = (DIST(r, c));                               \
+    if ((v = MVC(r, c) + thismse) < besterr) {            \
+      besterr = v;                                        \
+      br = r;                                             \
+      bc = c;                                             \
+      *distortion = thismse;                              \
+      *sse1 = sse;                                        \
+    }                                                     \
+  } else {                                                \
+    v = INT_MAX;                                          \
+  }
+
+#undef CHECK_BETTER0
+#define CHECK_BETTER0(v, r, c) CHECK_BETTER(v, r, c)
+
+#undef CHECK_BETTER1
+#define CHECK_BETTER1(v, r, c)                                            \
+  if (c >= minc && c <= maxc && r >= minr && r <= maxr) {                 \
+    thismse = upsampled_obmc_pref_error(                                  \
+        xd, mask, vfp, z, upre(y, y_stride, r, c), y_stride, w, h, &sse); \
+    if ((v = MVC(r, c) + thismse) < besterr) {                            \
+      besterr = v;                                                        \
+      br = r;                                                             \
+      bc = c;                                                             \
+      *distortion = thismse;                                              \
+      *sse1 = sse;                                                        \
+    }                                                                     \
+  } else {                                                                \
+    v = INT_MAX;                                                          \
+  }
+
+static unsigned int setup_obmc_center_error(
+    const int32_t *mask, const MV *bestmv, const MV *ref_mv, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, const int32_t *const wsrc,
+    const uint8_t *const y, int y_stride, int offset, int *mvjcost,
+    int *mvcost[2], unsigned int *sse1, int *distortion) {
+  unsigned int besterr;
+  besterr = vfp->ovf(y + offset, y_stride, wsrc, mask, sse1);
+  *distortion = besterr;
+  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
+  return besterr;
+}
+
+static int upsampled_obmc_pref_error(const MACROBLOCKD *xd, const int32_t *mask,
+                                     const aom_variance_fn_ptr_t *vfp,
+                                     const int32_t *const wsrc,
+                                     const uint8_t *const y, int y_stride,
+                                     int w, int h, unsigned int *sse) {
+  unsigned int besterr;
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    DECLARE_ALIGNED(16, uint16_t, pred16[MAX_SB_SQUARE]);
+    aom_highbd_upsampled_pred(pred16, w, h, y, y_stride);
+
+    besterr = vfp->ovf(CONVERT_TO_BYTEPTR(pred16), w, wsrc, mask, sse);
+  } else {
+    DECLARE_ALIGNED(16, uint8_t, pred[MAX_SB_SQUARE]);
+#else
+  DECLARE_ALIGNED(16, uint8_t, pred[MAX_SB_SQUARE]);
+  (void)xd;
+#endif  // CONFIG_HIGHBITDEPTH
+    aom_upsampled_pred(pred, w, h, y, y_stride);
+
+    besterr = vfp->ovf(pred, w, wsrc, mask, sse);
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif
+  return besterr;
+}
+
+static unsigned int upsampled_setup_obmc_center_error(
+    const MACROBLOCKD *xd, const int32_t *mask, const MV *bestmv,
+    const MV *ref_mv, int error_per_bit, const aom_variance_fn_ptr_t *vfp,
+    const int32_t *const wsrc, const uint8_t *const y, int y_stride, int w,
+    int h, int offset, int *mvjcost, int *mvcost[2], unsigned int *sse1,
+    int *distortion) {
+  unsigned int besterr = upsampled_obmc_pref_error(
+      xd, mask, vfp, wsrc, y + offset, y_stride, w, h, sse1);
+  *distortion = besterr;
+  besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
+  return besterr;
+}
+
+int av1_find_best_obmc_sub_pixel_tree_up(
+    const AV1_COMP *cpi, MACROBLOCK *x, int mi_row, int mi_col, MV *bestmv,
+    const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, int forced_stop, int iters_per_step,
+    int *mvjcost, int *mvcost[2], int *distortion, unsigned int *sse1,
+    int is_second, int use_upsampled_ref) {
+  const int32_t *wsrc = x->wsrc_buf;
+  const int32_t *mask = x->mask_buf;
+  const int *const z = wsrc;
+  const int *const src_address = z;
+  MACROBLOCKD *xd = &x->e_mbd;
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  unsigned int besterr = INT_MAX;
+  unsigned int sse;
+  unsigned int thismse;
+
+  int rr = ref_mv->row;
+  int rc = ref_mv->col;
+  int br = bestmv->row * 8;
+  int bc = bestmv->col * 8;
+  int hstep = 4;
+  int iter;
+  int round = 3 - forced_stop;
+  int tr = br;
+  int tc = bc;
+  const MV *search_step = search_step_table;
+  int idx, best_idx = -1;
+  unsigned int cost_array[5];
+  int kr, kc;
+  const int w = block_size_wide[mbmi->sb_type];
+  const int h = block_size_high[mbmi->sb_type];
+  int offset;
+  int y_stride;
+  const uint8_t *y;
+
+  const struct buf_2d backup_pred = pd->pre[is_second];
+  int minc, maxc, minr, maxr;
+
+  av1_set_subpel_mv_search_range(&x->mv_limits, &minc, &maxc, &minr, &maxr,
+                                 ref_mv);
+
+  if (use_upsampled_ref) {
+    int ref = xd->mi[0]->mbmi.ref_frame[is_second];
+    const YV12_BUFFER_CONFIG *upsampled_ref = get_upsampled_ref(cpi, ref);
+    setup_pred_plane(&pd->pre[is_second], mbmi->sb_type,
+                     upsampled_ref->y_buffer, upsampled_ref->y_crop_width,
+                     upsampled_ref->y_crop_height, upsampled_ref->y_stride,
+                     (mi_row << 3), (mi_col << 3), NULL, pd->subsampling_x,
+                     pd->subsampling_y);
+  }
+  y = pd->pre[is_second].buf;
+  y_stride = pd->pre[is_second].stride;
+  offset = bestmv->row * y_stride + bestmv->col;
+
+  if (!allow_hp)
+    if (round == 3) round = 2;
+
+  bestmv->row *= 8;
+  bestmv->col *= 8;
+  // use_upsampled_ref can be 0 or 1
+  if (use_upsampled_ref)
+    besterr = upsampled_setup_obmc_center_error(
+        xd, mask, bestmv, ref_mv, error_per_bit, vfp, z, y, y_stride, w, h,
+        (offset * 8), mvjcost, mvcost, sse1, distortion);
+  else
+    besterr = setup_obmc_center_error(mask, bestmv, ref_mv, error_per_bit, vfp,
+                                      z, y, y_stride, offset, mvjcost, mvcost,
+                                      sse1, distortion);
+
+  for (iter = 0; iter < round; ++iter) {
+    // Check vertical and horizontal sub-pixel positions.
+    for (idx = 0; idx < 4; ++idx) {
+      tr = br + search_step[idx].row;
+      tc = bc + search_step[idx].col;
+      if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
+        MV this_mv = { tr, tc };
+
+        if (use_upsampled_ref) {
+          const uint8_t *const pre_address = y + tr * y_stride + tc;
+
+          thismse = upsampled_obmc_pref_error(
+              xd, mask, vfp, src_address, pre_address, y_stride, w, h, &sse);
+        } else {
+          const uint8_t *const pre_address =
+              y + (tr >> 3) * y_stride + (tc >> 3);
+          thismse = vfp->osvf(pre_address, y_stride, sp(tc), sp(tr),
+                              src_address, mask, &sse);
+        }
+
+        cost_array[idx] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost,
+                                                mvcost, error_per_bit);
+        if (cost_array[idx] < besterr) {
+          best_idx = idx;
+          besterr = cost_array[idx];
+          *distortion = thismse;
+          *sse1 = sse;
+        }
+      } else {
+        cost_array[idx] = INT_MAX;
+      }
+    }
+
+    // Check diagonal sub-pixel position
+    kc = (cost_array[0] <= cost_array[1] ? -hstep : hstep);
+    kr = (cost_array[2] <= cost_array[3] ? -hstep : hstep);
+
+    tc = bc + kc;
+    tr = br + kr;
+    if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
+      MV this_mv = { tr, tc };
+
+      if (use_upsampled_ref) {
+        const uint8_t *const pre_address = y + tr * y_stride + tc;
+
+        thismse = upsampled_obmc_pref_error(xd, mask, vfp, src_address,
+                                            pre_address, y_stride, w, h, &sse);
+      } else {
+        const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
+
+        thismse = vfp->osvf(pre_address, y_stride, sp(tc), sp(tr), src_address,
+                            mask, &sse);
+      }
+
+      cost_array[4] = thismse + mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost,
+                                            error_per_bit);
+
+      if (cost_array[4] < besterr) {
+        best_idx = 4;
+        besterr = cost_array[4];
+        *distortion = thismse;
+        *sse1 = sse;
+      }
+    } else {
+      cost_array[idx] = INT_MAX;
+    }
+
+    if (best_idx < 4 && best_idx >= 0) {
+      br += search_step[best_idx].row;
+      bc += search_step[best_idx].col;
+    } else if (best_idx == 4) {
+      br = tr;
+      bc = tc;
+    }
+
+    if (iters_per_step > 1 && best_idx != -1) {
+      if (use_upsampled_ref) {
+        SECOND_LEVEL_CHECKS_BEST(1);
+      } else {
+        SECOND_LEVEL_CHECKS_BEST(0);
+      }
+    }
+
+    tr = br;
+    tc = bc;
+
+    search_step += 4;
+    hstep >>= 1;
+    best_idx = -1;
+  }
+
+  // These lines insure static analysis doesn't warn that
+  // tr and tc aren't used after the above point.
+  (void)tr;
+  (void)tc;
+
+  bestmv->row = br;
+  bestmv->col = bc;
+
+  if (use_upsampled_ref) {
+    pd->pre[is_second] = backup_pred;
+  }
+
+  return besterr;
+}
+
+#undef DIST
+#undef MVC
+#undef CHECK_BETTER
+
+static int get_obmc_mvpred_var(const MACROBLOCK *x, const int32_t *wsrc,
+                               const int32_t *mask, const MV *best_mv,
+                               const MV *center_mv,
+                               const aom_variance_fn_ptr_t *vfp, int use_mvcost,
+                               int is_second) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[is_second];
+  const MV mv = { best_mv->row * 8, best_mv->col * 8 };
+  unsigned int unused;
+
+  return vfp->ovf(get_buf_from_mv(in_what, best_mv), in_what->stride, wsrc,
+                  mask, &unused) +
+         (use_mvcost ? mv_err_cost(&mv, center_mv, x->nmvjointcost, x->mvcost,
+                                   x->errorperbit)
+                     : 0);
+}
+
+int obmc_refining_search_sad(const MACROBLOCK *x, const int32_t *wsrc,
+                             const int32_t *mask, MV *ref_mv, int error_per_bit,
+                             int search_range,
+                             const aom_variance_fn_ptr_t *fn_ptr,
+                             const MV *center_mv, int is_second) {
+  const MV neighbors[4] = { { -1, 0 }, { 0, -1 }, { 0, 1 }, { 1, 0 } };
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[is_second];
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  unsigned int best_sad = fn_ptr->osdf(get_buf_from_mv(in_what, ref_mv),
+                                       in_what->stride, wsrc, mask) +
+                          mvsad_err_cost(x, ref_mv, &fcenter_mv, error_per_bit);
+  int i, j;
+
+  for (i = 0; i < search_range; i++) {
+    int best_site = -1;
+
+    for (j = 0; j < 4; j++) {
+      const MV mv = { ref_mv->row + neighbors[j].row,
+                      ref_mv->col + neighbors[j].col };
+      if (is_mv_in(&x->mv_limits, &mv)) {
+        unsigned int sad = fn_ptr->osdf(get_buf_from_mv(in_what, &mv),
+                                        in_what->stride, wsrc, mask);
+        if (sad < best_sad) {
+          sad += mvsad_err_cost(x, &mv, &fcenter_mv, error_per_bit);
+          if (sad < best_sad) {
+            best_sad = sad;
+            best_site = j;
+          }
+        }
+      }
+    }
+
+    if (best_site == -1) {
+      break;
+    } else {
+      ref_mv->row += neighbors[best_site].row;
+      ref_mv->col += neighbors[best_site].col;
+    }
+  }
+  return best_sad;
+}
+
+int obmc_diamond_search_sad(const MACROBLOCK *x, const search_site_config *cfg,
+                            const int32_t *wsrc, const int32_t *mask,
+                            MV *ref_mv, MV *best_mv, int search_param,
+                            int sad_per_bit, int *num00,
+                            const aom_variance_fn_ptr_t *fn_ptr,
+                            const MV *center_mv, int is_second) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const in_what = &xd->plane[0].pre[is_second];
+  // search_param determines the length of the initial step and hence the number
+  // of iterations
+  // 0 = initial step (MAX_FIRST_STEP) pel : 1 = (MAX_FIRST_STEP/2) pel, 2 =
+  // (MAX_FIRST_STEP/4) pel... etc.
+  const search_site *const ss = &cfg->ss[search_param * cfg->searches_per_step];
+  const int tot_steps = (cfg->ss_count / cfg->searches_per_step) - search_param;
+  const MV fcenter_mv = { center_mv->row >> 3, center_mv->col >> 3 };
+  const uint8_t *best_address, *in_what_ref;
+  int best_sad = INT_MAX;
+  int best_site = 0;
+  int last_site = 0;
+  int i, j, step;
+
+  clamp_mv(ref_mv, x->mv_limits.col_min, x->mv_limits.col_max,
+           x->mv_limits.row_min, x->mv_limits.row_max);
+  in_what_ref = in_what->buf + ref_mv->row * in_what->stride + ref_mv->col;
+  best_address = in_what_ref;
+  *num00 = 0;
+  *best_mv = *ref_mv;
+
+  // Check the starting position
+  best_sad = fn_ptr->osdf(best_address, in_what->stride, wsrc, mask) +
+             mvsad_err_cost(x, best_mv, &fcenter_mv, sad_per_bit);
+
+  i = 1;
+
+  for (step = 0; step < tot_steps; step++) {
+    for (j = 0; j < cfg->searches_per_step; j++) {
+      const MV mv = { best_mv->row + ss[i].mv.row,
+                      best_mv->col + ss[i].mv.col };
+      if (is_mv_in(&x->mv_limits, &mv)) {
+        int sad = fn_ptr->osdf(best_address + ss[i].offset, in_what->stride,
+                               wsrc, mask);
+        if (sad < best_sad) {
+          sad += mvsad_err_cost(x, &mv, &fcenter_mv, sad_per_bit);
+          if (sad < best_sad) {
+            best_sad = sad;
+            best_site = i;
+          }
+        }
+      }
+
+      i++;
+    }
+
+    if (best_site != last_site) {
+      best_mv->row += ss[best_site].mv.row;
+      best_mv->col += ss[best_site].mv.col;
+      best_address += ss[best_site].offset;
+      last_site = best_site;
+#if defined(NEW_DIAMOND_SEARCH)
+      while (1) {
+        const MV this_mv = { best_mv->row + ss[best_site].mv.row,
+                             best_mv->col + ss[best_site].mv.col };
+        if (is_mv_in(&x->mv_limits, &this_mv)) {
+          int sad = fn_ptr->osdf(best_address + ss[best_site].offset,
+                                 in_what->stride, wsrc, mask);
+          if (sad < best_sad) {
+            sad += mvsad_err_cost(x, &this_mv, &fcenter_mv, sad_per_bit);
+            if (sad < best_sad) {
+              best_sad = sad;
+              best_mv->row += ss[best_site].mv.row;
+              best_mv->col += ss[best_site].mv.col;
+              best_address += ss[best_site].offset;
+              continue;
+            }
+          }
+        }
+        break;
+      }
+#endif
+    } else if (best_address == in_what_ref) {
+      (*num00)++;
+    }
+  }
+  return best_sad;
+}
+
+int av1_obmc_full_pixel_diamond(const AV1_COMP *cpi, MACROBLOCK *x,
+                                MV *mvp_full, int step_param, int sadpb,
+                                int further_steps, int do_refine,
+                                const aom_variance_fn_ptr_t *fn_ptr,
+                                const MV *ref_mv, MV *dst_mv, int is_second) {
+  const int32_t *wsrc = x->wsrc_buf;
+  const int32_t *mask = x->mask_buf;
+  MV temp_mv;
+  int thissme, n, num00 = 0;
+  int bestsme =
+      obmc_diamond_search_sad(x, &cpi->ss_cfg, wsrc, mask, mvp_full, &temp_mv,
+                              step_param, sadpb, &n, fn_ptr, ref_mv, is_second);
+  if (bestsme < INT_MAX)
+    bestsme = get_obmc_mvpred_var(x, wsrc, mask, &temp_mv, ref_mv, fn_ptr, 1,
+                                  is_second);
+  *dst_mv = temp_mv;
+
+  // If there won't be more n-step search, check to see if refining search is
+  // needed.
+  if (n > further_steps) do_refine = 0;
+
+  while (n < further_steps) {
+    ++n;
+
+    if (num00) {
+      num00--;
+    } else {
+      thissme = obmc_diamond_search_sad(x, &cpi->ss_cfg, wsrc, mask, mvp_full,
+                                        &temp_mv, step_param + n, sadpb, &num00,
+                                        fn_ptr, ref_mv, is_second);
+      if (thissme < INT_MAX)
+        thissme = get_obmc_mvpred_var(x, wsrc, mask, &temp_mv, ref_mv, fn_ptr,
+                                      1, is_second);
+
+      // check to see if refining search is needed.
+      if (num00 > further_steps - n) do_refine = 0;
+
+      if (thissme < bestsme) {
+        bestsme = thissme;
+        *dst_mv = temp_mv;
+      }
+    }
+  }
+
+  // final 1-away diamond refining search
+  if (do_refine) {
+    const int search_range = 8;
+    MV best_mv = *dst_mv;
+    thissme = obmc_refining_search_sad(x, wsrc, mask, &best_mv, sadpb,
+                                       search_range, fn_ptr, ref_mv, is_second);
+    if (thissme < INT_MAX)
+      thissme = get_obmc_mvpred_var(x, wsrc, mask, &best_mv, ref_mv, fn_ptr, 1,
+                                    is_second);
+    if (thissme < bestsme) {
+      bestsme = thissme;
+      *dst_mv = best_mv;
+    }
+  }
+  return bestsme;
+}
+#endif  // CONFIG_MOTION_VAR
+
+// Note(yunqingwang): The following 2 functions are only used in the motion
+// vector unit test, which return extreme motion vectors allowed by the MV
+// limits.
+#define COMMON_MV_TEST     \
+  SETUP_SUBPEL_SEARCH;     \
+                           \
+  (void)error_per_bit;     \
+  (void)vfp;               \
+  (void)src_address;       \
+  (void)src_stride;        \
+  (void)y;                 \
+  (void)y_stride;          \
+  (void)second_pred;       \
+  (void)w;                 \
+  (void)h;                 \
+  (void)use_upsampled_ref; \
+  (void)offset;            \
+  (void)mvjcost;           \
+  (void)mvcost;            \
+  (void)sse1;              \
+  (void)distortion;        \
+                           \
+  (void)halfiters;         \
+  (void)quarteriters;      \
+  (void)eighthiters;       \
+  (void)whichdir;          \
+  (void)forced_stop;       \
+  (void)hstep;             \
+                           \
+  (void)tr;                \
+  (void)tc;                \
+  (void)sse;               \
+  (void)thismse;           \
+  (void)cost_list;
+// Return the maximum MV.
+int av1_return_max_sub_pixel_mv(MACROBLOCK *x, const MV *ref_mv, int allow_hp,
+                                int error_per_bit,
+                                const aom_variance_fn_ptr_t *vfp,
+                                int forced_stop, int iters_per_step,
+                                int *cost_list, int *mvjcost, int *mvcost[2],
+                                int *distortion, unsigned int *sse1,
+                                const uint8_t *second_pred, int w, int h,
+                                int use_upsampled_ref) {
+  COMMON_MV_TEST;
+  (void)minr;
+  (void)minc;
+  bestmv->row = maxr;
+  bestmv->col = maxc;
+  besterr = 0;
+  // In the sub-pel motion search, if hp is not used, then the last bit of mv
+  // has to be 0.
+  lower_mv_precision(bestmv, allow_hp);
+  return besterr;
+}
+// Return the minimum MV.
+int av1_return_min_sub_pixel_mv(MACROBLOCK *x, const MV *ref_mv, int allow_hp,
+                                int error_per_bit,
+                                const aom_variance_fn_ptr_t *vfp,
+                                int forced_stop, int iters_per_step,
+                                int *cost_list, int *mvjcost, int *mvcost[2],
+                                int *distortion, unsigned int *sse1,
+                                const uint8_t *second_pred, int w, int h,
+                                int use_upsampled_ref) {
+  COMMON_MV_TEST;
+  (void)maxr;
+  (void)maxc;
+  bestmv->row = minr;
+  bestmv->col = minc;
+  besterr = 0;
+  // In the sub-pel motion search, if hp is not used, then the last bit of mv
+  // has to be 0.
+  lower_mv_precision(bestmv, allow_hp);
+  return besterr;
+}
diff --git a/third_party/aom/av1/encoder/mcomp.h b/third_party/aom/av1/encoder/mcomp.h
new file mode 100644
index 0000000000..8465860adf
--- /dev/null
+++ b/third_party/aom/av1/encoder/mcomp.h
@@ -0,0 +1,163 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_MCOMP_H_
+#define AV1_ENCODER_MCOMP_H_
+
+#include "av1/encoder/block.h"
+#include "aom_dsp/variance.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// The maximum number of steps in a step search given the largest
+// allowed initial step
+#define MAX_MVSEARCH_STEPS 11
+// Max full pel mv specified in the unit of full pixel
+// Enable the use of motion vector in range [-1023, 1023].
+#define MAX_FULL_PEL_VAL ((1 << (MAX_MVSEARCH_STEPS - 1)) - 1)
+// Maximum size of the first step in full pel units
+#define MAX_FIRST_STEP (1 << (MAX_MVSEARCH_STEPS - 1))
+// Allowed motion vector pixel distance outside image border
+// for Block_16x16
+#define BORDER_MV_PIXELS_B16 (16 + AOM_INTERP_EXTEND)
+
+// motion search site
+typedef struct search_site {
+  MV mv;
+  int offset;
+} search_site;
+
+typedef struct search_site_config {
+  search_site ss[8 * MAX_MVSEARCH_STEPS + 1];
+  int ss_count;
+  int searches_per_step;
+} search_site_config;
+
+void av1_init_dsmotion_compensation(search_site_config *cfg, int stride);
+void av1_init3smotion_compensation(search_site_config *cfg, int stride);
+
+void av1_set_mv_search_range(MvLimits *mv_limits, const MV *mv);
+
+int av1_mv_bit_cost(const MV *mv, const MV *ref, const int *mvjcost,
+                    int *mvcost[2], int weight);
+
+// Utility to compute variance + MV rate cost for a given MV
+int av1_get_mvpred_var(const MACROBLOCK *x, const MV *best_mv,
+                       const MV *center_mv, const aom_variance_fn_ptr_t *vfp,
+                       int use_mvcost);
+int av1_get_mvpred_av_var(const MACROBLOCK *x, const MV *best_mv,
+                          const MV *center_mv, const uint8_t *second_pred,
+                          const aom_variance_fn_ptr_t *vfp, int use_mvcost);
+
+struct AV1_COMP;
+struct SPEED_FEATURES;
+
+int av1_init_search_range(int size);
+
+int av1_refining_search_sad(struct macroblock *x, struct mv *ref_mv,
+                            int sad_per_bit, int distance,
+                            const aom_variance_fn_ptr_t *fn_ptr,
+                            const struct mv *center_mv);
+
+// Runs sequence of diamond searches in smaller steps for RD.
+int av1_full_pixel_diamond(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                           MV *mvp_full, int step_param, int sadpb,
+                           int further_steps, int do_refine, int *cost_list,
+                           const aom_variance_fn_ptr_t *fn_ptr,
+                           const MV *ref_mv, MV *dst_mv);
+
+// Perform integral projection based motion estimation.
+unsigned int av1_int_pro_motion_estimation(const struct AV1_COMP *cpi,
+                                           MACROBLOCK *x, BLOCK_SIZE bsize,
+                                           int mi_row, int mi_col);
+
+int av1_hex_search(MACROBLOCK *x, MV *start_mv, int search_param,
+                   int sad_per_bit, int do_init_search, int *cost_list,
+                   const aom_variance_fn_ptr_t *vfp, int use_mvcost,
+                   const MV *center_mv);
+
+typedef int(fractional_mv_step_fp)(
+    MACROBLOCK *x, const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp,
+    int forced_stop,  // 0 - full, 1 - qtr only, 2 - half only
+    int iters_per_step, int *cost_list, int *mvjcost, int *mvcost[2],
+    int *distortion, unsigned int *sse1, const uint8_t *second_pred, int w,
+    int h, int use_upsampled_ref);
+
+extern fractional_mv_step_fp av1_find_best_sub_pixel_tree;
+extern fractional_mv_step_fp av1_find_best_sub_pixel_tree_pruned;
+extern fractional_mv_step_fp av1_find_best_sub_pixel_tree_pruned_more;
+extern fractional_mv_step_fp av1_find_best_sub_pixel_tree_pruned_evenmore;
+extern fractional_mv_step_fp av1_return_max_sub_pixel_mv;
+extern fractional_mv_step_fp av1_return_min_sub_pixel_mv;
+
+typedef int (*av1_full_search_fn_t)(const MACROBLOCK *x, const MV *ref_mv,
+                                    int sad_per_bit, int distance,
+                                    const aom_variance_fn_ptr_t *fn_ptr,
+                                    const MV *center_mv, MV *best_mv);
+
+typedef int (*av1_diamond_search_fn_t)(
+    MACROBLOCK *x, const search_site_config *cfg, MV *ref_mv, MV *best_mv,
+    int search_param, int sad_per_bit, int *num00,
+    const aom_variance_fn_ptr_t *fn_ptr, const MV *center_mv);
+
+int av1_refining_search_8p_c(MACROBLOCK *x, int error_per_bit, int search_range,
+                             const aom_variance_fn_ptr_t *fn_ptr,
+                             const MV *center_mv, const uint8_t *second_pred);
+
+struct AV1_COMP;
+
+int av1_full_pixel_search(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                          BLOCK_SIZE bsize, MV *mvp_full, int step_param,
+                          int error_per_bit, int *cost_list, const MV *ref_mv,
+                          int var_max, int rd);
+
+#if CONFIG_EXT_INTER
+int av1_find_best_masked_sub_pixel_tree(
+    const MACROBLOCK *x, const uint8_t *mask, int mask_stride, MV *bestmv,
+    const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, int forced_stop, int iters_per_step,
+    int *mvjcost, int *mvcost[2], int *distortion, unsigned int *sse1,
+    int is_second);
+int av1_find_best_masked_sub_pixel_tree_up(
+    const struct AV1_COMP *cpi, MACROBLOCK *x, const uint8_t *mask,
+    int mask_stride, int mi_row, int mi_col, MV *bestmv, const MV *ref_mv,
+    int allow_hp, int error_per_bit, const aom_variance_fn_ptr_t *vfp,
+    int forced_stop, int iters_per_step, int *mvjcost, int *mvcost[2],
+    int *distortion, unsigned int *sse1, int is_second, int use_upsampled_ref);
+int av1_masked_full_pixel_diamond(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                                  const uint8_t *mask, int mask_stride,
+                                  MV *mvp_full, int step_param, int sadpb,
+                                  int further_steps, int do_refine,
+                                  const aom_variance_fn_ptr_t *fn_ptr,
+                                  const MV *ref_mv, MV *dst_mv, int is_second);
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_MOTION_VAR
+int av1_obmc_full_pixel_diamond(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                                MV *mvp_full, int step_param, int sadpb,
+                                int further_steps, int do_refine,
+                                const aom_variance_fn_ptr_t *fn_ptr,
+                                const MV *ref_mv, MV *dst_mv, int is_second);
+int av1_find_best_obmc_sub_pixel_tree_up(
+    const struct AV1_COMP *cpi, MACROBLOCK *x, int mi_row, int mi_col,
+    MV *bestmv, const MV *ref_mv, int allow_hp, int error_per_bit,
+    const aom_variance_fn_ptr_t *vfp, int forced_stop, int iters_per_step,
+    int *mvjcost, int *mvcost[2], int *distortion, unsigned int *sse1,
+    int is_second, int use_upsampled_ref);
+#endif  // CONFIG_MOTION_VAR
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_MCOMP_H_
diff --git a/third_party/aom/av1/encoder/mips/msa/error_msa.c b/third_party/aom/av1/encoder/mips/msa/error_msa.c
new file mode 100644
index 0000000000..8d13af7ad9
--- /dev/null
+++ b/third_party/aom/av1/encoder/mips/msa/error_msa.c
@@ -0,0 +1,108 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "aom_dsp/mips/macros_msa.h"
+
+#define BLOCK_ERROR_BLOCKSIZE_MSA(BSize)                                     \
+  static int64_t block_error_##BSize##size_msa(                              \
+      const int16_t *coeff_ptr, const int16_t *dq_coeff_ptr, int64_t *ssz) { \
+    int64_t err = 0;                                                         \
+    uint32_t loop_cnt;                                                       \
+    v8i16 coeff, dq_coeff, coeff_r_h, coeff_l_h;                             \
+    v4i32 diff_r, diff_l, coeff_r_w, coeff_l_w;                              \
+    v2i64 sq_coeff_r, sq_coeff_l;                                            \
+    v2i64 err0, err_dup0, err1, err_dup1;                                    \
+                                                                             \
+    coeff = LD_SH(coeff_ptr);                                                \
+    dq_coeff = LD_SH(dq_coeff_ptr);                                          \
+    UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                                \
+    ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                      \
+    HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                       \
+    DOTP_SW2_SD(coeff_r_w, coeff_l_w, coeff_r_w, coeff_l_w, sq_coeff_r,      \
+                sq_coeff_l);                                                 \
+    DOTP_SW2_SD(diff_r, diff_l, diff_r, diff_l, err0, err1);                 \
+                                                                             \
+    coeff = LD_SH(coeff_ptr + 8);                                            \
+    dq_coeff = LD_SH(dq_coeff_ptr + 8);                                      \
+    UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                                \
+    ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                      \
+    HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                       \
+    DPADD_SD2_SD(coeff_r_w, coeff_l_w, sq_coeff_r, sq_coeff_l);              \
+    DPADD_SD2_SD(diff_r, diff_l, err0, err1);                                \
+                                                                             \
+    coeff_ptr += 16;                                                         \
+    dq_coeff_ptr += 16;                                                      \
+                                                                             \
+    for (loop_cnt = ((BSize >> 4) - 1); loop_cnt--;) {                       \
+      coeff = LD_SH(coeff_ptr);                                              \
+      dq_coeff = LD_SH(dq_coeff_ptr);                                        \
+      UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                              \
+      ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                    \
+      HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                     \
+      DPADD_SD2_SD(coeff_r_w, coeff_l_w, sq_coeff_r, sq_coeff_l);            \
+      DPADD_SD2_SD(diff_r, diff_l, err0, err1);                              \
+                                                                             \
+      coeff = LD_SH(coeff_ptr + 8);                                          \
+      dq_coeff = LD_SH(dq_coeff_ptr + 8);                                    \
+      UNPCK_SH_SW(coeff, coeff_r_w, coeff_l_w);                              \
+      ILVRL_H2_SH(coeff, dq_coeff, coeff_r_h, coeff_l_h);                    \
+      HSUB_UH2_SW(coeff_r_h, coeff_l_h, diff_r, diff_l);                     \
+      DPADD_SD2_SD(coeff_r_w, coeff_l_w, sq_coeff_r, sq_coeff_l);            \
+      DPADD_SD2_SD(diff_r, diff_l, err0, err1);                              \
+                                                                             \
+      coeff_ptr += 16;                                                       \
+      dq_coeff_ptr += 16;                                                    \
+    }                                                                        \
+                                                                             \
+    err_dup0 = __msa_splati_d(sq_coeff_r, 1);                                \
+    err_dup1 = __msa_splati_d(sq_coeff_l, 1);                                \
+    sq_coeff_r += err_dup0;                                                  \
+    sq_coeff_l += err_dup1;                                                  \
+    *ssz = __msa_copy_s_d(sq_coeff_r, 0);                                    \
+    *ssz += __msa_copy_s_d(sq_coeff_l, 0);                                   \
+                                                                             \
+    err_dup0 = __msa_splati_d(err0, 1);                                      \
+    err_dup1 = __msa_splati_d(err1, 1);                                      \
+    err0 += err_dup0;                                                        \
+    err1 += err_dup1;                                                        \
+    err = __msa_copy_s_d(err0, 0);                                           \
+    err += __msa_copy_s_d(err1, 0);                                          \
+                                                                             \
+    return err;                                                              \
+  }
+
+/* clang-format off */
+BLOCK_ERROR_BLOCKSIZE_MSA(16)
+BLOCK_ERROR_BLOCKSIZE_MSA(64)
+BLOCK_ERROR_BLOCKSIZE_MSA(256)
+BLOCK_ERROR_BLOCKSIZE_MSA(1024)
+/* clang-format on */
+
+int64_t av1_block_error_msa(const tran_low_t *coeff_ptr,
+                            const tran_low_t *dq_coeff_ptr, intptr_t blk_size,
+                            int64_t *ssz) {
+  int64_t err;
+  const int16_t *coeff = (const int16_t *)coeff_ptr;
+  const int16_t *dq_coeff = (const int16_t *)dq_coeff_ptr;
+
+  switch (blk_size) {
+    case 16: err = block_error_16size_msa(coeff, dq_coeff, ssz); break;
+    case 64: err = block_error_64size_msa(coeff, dq_coeff, ssz); break;
+    case 256: err = block_error_256size_msa(coeff, dq_coeff, ssz); break;
+    case 1024: err = block_error_1024size_msa(coeff, dq_coeff, ssz); break;
+    default:
+      err = av1_block_error_c(coeff_ptr, dq_coeff_ptr, blk_size, ssz);
+      break;
+  }
+
+  return err;
+}
diff --git a/third_party/aom/av1/encoder/mips/msa/fdct16x16_msa.c b/third_party/aom/av1/encoder/mips/msa/fdct16x16_msa.c
new file mode 100644
index 0000000000..4b0364d6c8
--- /dev/null
+++ b/third_party/aom/av1/encoder/mips/msa/fdct16x16_msa.c
@@ -0,0 +1,436 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/enums.h"
+#include "av1/encoder/mips/msa/fdct_msa.h"
+#include "aom_dsp/mips/fwd_txfm_msa.h"
+
+static void fadst16_cols_step1_msa(const int16_t *input, int32_t stride,
+                                   const int32_t *const0, int16_t *int_buf) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 tp0, tp1, tp2, tp3, g0, g1, g2, g3, g8, g9, g10, g11, h0, h1, h2, h3;
+  v4i32 k0, k1, k2, k3;
+
+  /* load input data */
+  r0 = LD_SH(input);
+  r15 = LD_SH(input + 15 * stride);
+  r7 = LD_SH(input + 7 * stride);
+  r8 = LD_SH(input + 8 * stride);
+  SLLI_4V(r0, r15, r7, r8, 2);
+
+  /* stage 1 */
+  LD_SW2(const0, 4, k0, k1);
+  LD_SW2(const0 + 8, 4, k2, k3);
+  MADD_BF(r15, r0, r7, r8, k0, k1, k2, k3, g0, g1, g2, g3);
+
+  r3 = LD_SH(input + 3 * stride);
+  r4 = LD_SH(input + 4 * stride);
+  r11 = LD_SH(input + 11 * stride);
+  r12 = LD_SH(input + 12 * stride);
+  SLLI_4V(r3, r4, r11, r12, 2);
+
+  LD_SW2(const0 + 4 * 4, 4, k0, k1);
+  LD_SW2(const0 + 4 * 6, 4, k2, k3);
+  MADD_BF(r11, r4, r3, r12, k0, k1, k2, k3, g8, g9, g10, g11);
+
+  /* stage 2 */
+  BUTTERFLY_4(g0, g2, g10, g8, tp0, tp2, tp3, tp1);
+  ST_SH2(tp0, tp2, int_buf, 8);
+  ST_SH2(tp1, tp3, int_buf + 4 * 8, 8);
+
+  LD_SW2(const0 + 4 * 8, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 10);
+  MADD_BF(g1, g3, g9, g11, k0, k1, k2, k0, h0, h1, h2, h3);
+
+  ST_SH2(h0, h1, int_buf + 8 * 8, 8);
+  ST_SH2(h3, h2, int_buf + 12 * 8, 8);
+
+  r9 = LD_SH(input + 9 * stride);
+  r6 = LD_SH(input + 6 * stride);
+  r1 = LD_SH(input + stride);
+  r14 = LD_SH(input + 14 * stride);
+  SLLI_4V(r9, r6, r1, r14, 2);
+
+  LD_SW2(const0 + 4 * 11, 4, k0, k1);
+  LD_SW2(const0 + 4 * 13, 4, k2, k3);
+  MADD_BF(r9, r6, r1, r14, k0, k1, k2, k3, g0, g1, g2, g3);
+
+  ST_SH2(g1, g3, int_buf + 3 * 8, 4 * 8);
+
+  r13 = LD_SH(input + 13 * stride);
+  r2 = LD_SH(input + 2 * stride);
+  r5 = LD_SH(input + 5 * stride);
+  r10 = LD_SH(input + 10 * stride);
+  SLLI_4V(r13, r2, r5, r10, 2);
+
+  LD_SW2(const0 + 4 * 15, 4, k0, k1);
+  LD_SW2(const0 + 4 * 17, 4, k2, k3);
+  MADD_BF(r13, r2, r5, r10, k0, k1, k2, k3, h0, h1, h2, h3);
+
+  ST_SH2(h1, h3, int_buf + 11 * 8, 4 * 8);
+
+  BUTTERFLY_4(h0, h2, g2, g0, tp0, tp1, tp2, tp3);
+  ST_SH4(tp0, tp1, tp2, tp3, int_buf + 2 * 8, 4 * 8);
+}
+
+static void fadst16_step2_msa_helper(int16_t *int_buf, const int32_t *const0,
+                                     int16_t *out, int16_t *out_ptr) {
+  v8i16 tp0, tp1, tp2, tp3, g5, g7, g13, g15;
+  v8i16 h0, h1, h2, h3, h4, h5, h6, h7, h10, h11;
+  v8i16 out0, out1, out2, out3, out4, out5, out6, out7;
+  v8i16 out8, out9, out10, out11, out12, out13, out14, out15;
+  v4i32 k0, k1, k2, k3;
+
+  LD_SH2(int_buf + 3 * 8, 4 * 8, g13, g15);
+  LD_SH2(int_buf + 11 * 8, 4 * 8, g5, g7);
+  LD_SW2(const0 + 4 * 19, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 21);
+  MADD_BF(g7, g5, g15, g13, k0, k1, k2, k0, h4, h5, h6, h7);
+
+  tp0 = LD_SH(int_buf + 4 * 8);
+  tp1 = LD_SH(int_buf + 5 * 8);
+  tp3 = LD_SH(int_buf + 10 * 8);
+  tp2 = LD_SH(int_buf + 14 * 8);
+  LD_SW2(const0 + 4 * 22, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 24);
+  MADD_BF(tp0, tp1, tp2, tp3, k0, k1, k2, k0, out4, out6, out5, out7);
+  out4 = -out4;
+  ST_SH(out4, (out + 3 * 16));
+  ST_SH(out5, (out_ptr + 4 * 16));
+
+  h1 = LD_SH(int_buf + 9 * 8);
+  h3 = LD_SH(int_buf + 12 * 8);
+  MADD_BF(h1, h3, h5, h7, k0, k1, k2, k0, out12, out14, out13, out15);
+  out13 = -out13;
+  ST_SH(out12, (out + 2 * 16));
+  ST_SH(out13, (out_ptr + 5 * 16));
+
+  tp0 = LD_SH(int_buf);
+  tp1 = LD_SH(int_buf + 8);
+  tp2 = LD_SH(int_buf + 2 * 8);
+  tp3 = LD_SH(int_buf + 6 * 8);
+
+  BUTTERFLY_4(tp0, tp1, tp3, tp2, out0, out1, h11, h10);
+  out1 = -out1;
+  ST_SH(out0, (out));
+  ST_SH(out1, (out_ptr + 7 * 16));
+
+  h0 = LD_SH(int_buf + 8 * 8);
+  h2 = LD_SH(int_buf + 13 * 8);
+
+  BUTTERFLY_4(h0, h2, h6, h4, out8, out9, out11, out10);
+  out8 = -out8;
+  ST_SH(out8, (out + 16));
+  ST_SH(out9, (out_ptr + 6 * 16));
+
+  /* stage 4 */
+  LD_SW2(const0 + 4 * 25, 4, k0, k1);
+  LD_SW2(const0 + 4 * 27, 4, k2, k3);
+  MADD_SHORT(h10, h11, k1, k2, out2, out3);
+  ST_SH(out2, (out + 7 * 16));
+  ST_SH(out3, (out_ptr));
+
+  MADD_SHORT(out6, out7, k0, k3, out6, out7);
+  ST_SH(out6, (out + 4 * 16));
+  ST_SH(out7, (out_ptr + 3 * 16));
+
+  MADD_SHORT(out10, out11, k0, k3, out10, out11);
+  ST_SH(out10, (out + 6 * 16));
+  ST_SH(out11, (out_ptr + 16));
+
+  MADD_SHORT(out14, out15, k1, k2, out14, out15);
+  ST_SH(out14, (out + 5 * 16));
+  ST_SH(out15, (out_ptr + 2 * 16));
+}
+
+static void fadst16_cols_step2_msa(int16_t *int_buf, const int32_t *const0,
+                                   int16_t *out) {
+  fadst16_step2_msa_helper(int_buf, const0, out, out + 128);
+}
+
+static void fadst16_transpose_postproc_msa(int16_t *input, int16_t *out) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 l0, l1, l2, l3, l4, l5, l6, l7, l8, l9, l10, l11, l12, l13, l14, l15;
+
+  /* load input data */
+  LD_SH8(input, 16, l0, l1, l2, l3, l4, l5, l6, l7);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7, r0, r1, r2, r3, r4, r5, r6,
+                     r7);
+  FDCT_POSTPROC_2V_NEG_H(r0, r1);
+  FDCT_POSTPROC_2V_NEG_H(r2, r3);
+  FDCT_POSTPROC_2V_NEG_H(r4, r5);
+  FDCT_POSTPROC_2V_NEG_H(r6, r7);
+  ST_SH8(r0, r1, r2, r3, r4, r5, r6, r7, out, 8);
+  out += 64;
+
+  LD_SH8(input + 8, 16, l8, l9, l10, l11, l12, l13, l14, l15);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15, r8, r9, r10, r11,
+                     r12, r13, r14, r15);
+  FDCT_POSTPROC_2V_NEG_H(r8, r9);
+  FDCT_POSTPROC_2V_NEG_H(r10, r11);
+  FDCT_POSTPROC_2V_NEG_H(r12, r13);
+  FDCT_POSTPROC_2V_NEG_H(r14, r15);
+  ST_SH8(r8, r9, r10, r11, r12, r13, r14, r15, out, 8);
+  out += 64;
+
+  /* load input data */
+  input += 128;
+  LD_SH8(input, 16, l0, l1, l2, l3, l4, l5, l6, l7);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7, r0, r1, r2, r3, r4, r5, r6,
+                     r7);
+  FDCT_POSTPROC_2V_NEG_H(r0, r1);
+  FDCT_POSTPROC_2V_NEG_H(r2, r3);
+  FDCT_POSTPROC_2V_NEG_H(r4, r5);
+  FDCT_POSTPROC_2V_NEG_H(r6, r7);
+  ST_SH8(r0, r1, r2, r3, r4, r5, r6, r7, out, 8);
+  out += 64;
+
+  LD_SH8(input + 8, 16, l8, l9, l10, l11, l12, l13, l14, l15);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15, r8, r9, r10, r11,
+                     r12, r13, r14, r15);
+  FDCT_POSTPROC_2V_NEG_H(r8, r9);
+  FDCT_POSTPROC_2V_NEG_H(r10, r11);
+  FDCT_POSTPROC_2V_NEG_H(r12, r13);
+  FDCT_POSTPROC_2V_NEG_H(r14, r15);
+  ST_SH8(r8, r9, r10, r11, r12, r13, r14, r15, out, 8);
+}
+
+static void fadst16_rows_step1_msa(int16_t *input, const int32_t *const0,
+                                   int16_t *int_buf) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 tp0, tp1, tp2, tp3, g0, g1, g2, g3, g8, g9, g10, g11, h0, h1, h2, h3;
+  v4i32 k0, k1, k2, k3;
+
+  /* load input data */
+  r0 = LD_SH(input);
+  r7 = LD_SH(input + 7 * 8);
+  r8 = LD_SH(input + 8 * 8);
+  r15 = LD_SH(input + 15 * 8);
+
+  /* stage 1 */
+  LD_SW2(const0, 4, k0, k1);
+  LD_SW2(const0 + 4 * 2, 4, k2, k3);
+  MADD_BF(r15, r0, r7, r8, k0, k1, k2, k3, g0, g1, g2, g3);
+
+  r3 = LD_SH(input + 3 * 8);
+  r4 = LD_SH(input + 4 * 8);
+  r11 = LD_SH(input + 11 * 8);
+  r12 = LD_SH(input + 12 * 8);
+
+  LD_SW2(const0 + 4 * 4, 4, k0, k1);
+  LD_SW2(const0 + 4 * 6, 4, k2, k3);
+  MADD_BF(r11, r4, r3, r12, k0, k1, k2, k3, g8, g9, g10, g11);
+
+  /* stage 2 */
+  BUTTERFLY_4(g0, g2, g10, g8, tp0, tp2, tp3, tp1);
+  ST_SH2(tp0, tp1, int_buf, 4 * 8);
+  ST_SH2(tp2, tp3, int_buf + 8, 4 * 8);
+
+  LD_SW2(const0 + 4 * 8, 4, k0, k1);
+  k2 = LD_SW(const0 + 4 * 10);
+  MADD_BF(g1, g3, g9, g11, k0, k1, k2, k0, h0, h1, h2, h3);
+  ST_SH2(h0, h3, int_buf + 8 * 8, 4 * 8);
+  ST_SH2(h1, h2, int_buf + 9 * 8, 4 * 8);
+
+  r1 = LD_SH(input + 8);
+  r6 = LD_SH(input + 6 * 8);
+  r9 = LD_SH(input + 9 * 8);
+  r14 = LD_SH(input + 14 * 8);
+
+  LD_SW2(const0 + 4 * 11, 4, k0, k1);
+  LD_SW2(const0 + 4 * 13, 4, k2, k3);
+  MADD_BF(r9, r6, r1, r14, k0, k1, k2, k3, g0, g1, g2, g3);
+  ST_SH2(g1, g3, int_buf + 3 * 8, 4 * 8);
+
+  r2 = LD_SH(input + 2 * 8);
+  r5 = LD_SH(input + 5 * 8);
+  r10 = LD_SH(input + 10 * 8);
+  r13 = LD_SH(input + 13 * 8);
+
+  LD_SW2(const0 + 4 * 15, 4, k0, k1);
+  LD_SW2(const0 + 4 * 17, 4, k2, k3);
+  MADD_BF(r13, r2, r5, r10, k0, k1, k2, k3, h0, h1, h2, h3);
+  ST_SH2(h1, h3, int_buf + 11 * 8, 4 * 8);
+  BUTTERFLY_4(h0, h2, g2, g0, tp0, tp1, tp2, tp3);
+  ST_SH4(tp0, tp1, tp2, tp3, int_buf + 2 * 8, 4 * 8);
+}
+
+static void fadst16_rows_step2_msa(int16_t *int_buf, const int32_t *const0,
+                                   int16_t *out) {
+  fadst16_step2_msa_helper(int_buf, const0, out, out + 8);
+}
+
+static void fadst16_transpose_msa(int16_t *input, int16_t *out) {
+  v8i16 r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+  v8i16 l0, l1, l2, l3, l4, l5, l6, l7, l8, l9, l10, l11, l12, l13, l14, l15;
+
+  /* load input data */
+  LD_SH16(input, 8, l0, l8, l1, l9, l2, l10, l3, l11, l4, l12, l5, l13, l6, l14,
+          l7, l15);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7, r0, r1, r2, r3, r4, r5, r6,
+                     r7);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15, r8, r9, r10, r11,
+                     r12, r13, r14, r15);
+  ST_SH8(r0, r8, r1, r9, r2, r10, r3, r11, out, 8);
+  ST_SH8(r4, r12, r5, r13, r6, r14, r7, r15, (out + 64), 8);
+  out += 16 * 8;
+
+  /* load input data */
+  input += 128;
+  LD_SH16(input, 8, l0, l8, l1, l9, l2, l10, l3, l11, l4, l12, l5, l13, l6, l14,
+          l7, l15);
+  TRANSPOSE8x8_SH_SH(l0, l1, l2, l3, l4, l5, l6, l7, r0, r1, r2, r3, r4, r5, r6,
+                     r7);
+  TRANSPOSE8x8_SH_SH(l8, l9, l10, l11, l12, l13, l14, l15, r8, r9, r10, r11,
+                     r12, r13, r14, r15);
+  ST_SH8(r0, r8, r1, r9, r2, r10, r3, r11, out, 8);
+  ST_SH8(r4, r12, r5, r13, r6, r14, r7, r15, (out + 64), 8);
+}
+
+static void postproc_fdct16x8_1d_row(int16_t *intermediate, int16_t *output) {
+  int16_t *temp = intermediate;
+  int16_t *out = output;
+  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11;
+  v8i16 in12, in13, in14, in15;
+
+  LD_SH8(temp, 16, in0, in1, in2, in3, in4, in5, in6, in7);
+  temp = intermediate + 8;
+  LD_SH8(temp, 16, in8, in9, in10, in11, in12, in13, in14, in15);
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(in8, in9, in10, in11, in12, in13, in14, in15, in8, in9,
+                     in10, in11, in12, in13, in14, in15);
+  FDCT_POSTPROC_2V_NEG_H(in0, in1);
+  FDCT_POSTPROC_2V_NEG_H(in2, in3);
+  FDCT_POSTPROC_2V_NEG_H(in4, in5);
+  FDCT_POSTPROC_2V_NEG_H(in6, in7);
+  FDCT_POSTPROC_2V_NEG_H(in8, in9);
+  FDCT_POSTPROC_2V_NEG_H(in10, in11);
+  FDCT_POSTPROC_2V_NEG_H(in12, in13);
+  FDCT_POSTPROC_2V_NEG_H(in14, in15);
+  BUTTERFLY_16(in0, in1, in2, in3, in4, in5, in6, in7, in8, in9, in10, in11,
+               in12, in13, in14, in15, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6,
+               tmp7, in8, in9, in10, in11, in12, in13, in14, in15);
+  temp = intermediate;
+  ST_SH8(in8, in9, in10, in11, in12, in13, in14, in15, temp, 16);
+  FDCT8x16_EVEN(tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp0, tmp1,
+                tmp2, tmp3, tmp4, tmp5, tmp6, tmp7);
+  temp = intermediate;
+  LD_SH8(temp, 16, in8, in9, in10, in11, in12, in13, in14, in15);
+  FDCT8x16_ODD(in8, in9, in10, in11, in12, in13, in14, in15, in0, in1, in2, in3,
+               in4, in5, in6, in7);
+  TRANSPOSE8x8_SH_SH(tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3, tmp0, in0,
+                     tmp1, in1, tmp2, in2, tmp3, in3);
+  ST_SH8(tmp0, in0, tmp1, in1, tmp2, in2, tmp3, in3, out, 16);
+  TRANSPOSE8x8_SH_SH(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7, tmp4, in4,
+                     tmp5, in5, tmp6, in6, tmp7, in7);
+  out = output + 8;
+  ST_SH8(tmp4, in4, tmp5, in5, tmp6, in6, tmp7, in7, out, 16);
+}
+
+void av1_fht16x16_msa(const int16_t *input, int16_t *output, int32_t stride,
+                      int32_t tx_type) {
+  DECLARE_ALIGNED(32, int16_t, tmp[256]);
+  DECLARE_ALIGNED(32, int16_t, trans_buf[256]);
+  DECLARE_ALIGNED(32, int16_t, tmp_buf[128]);
+  int32_t i;
+  int16_t *ptmpbuf = &tmp_buf[0];
+  int16_t *trans = &trans_buf[0];
+  const int32_t const_arr[29 * 4] = {
+    52707308,    52707308,    52707308,    52707308,    -1072430300,
+    -1072430300, -1072430300, -1072430300, 795618043,   795618043,
+    795618043,   795618043,   -721080468,  -721080468,  -721080468,
+    -721080468,  459094491,   459094491,   459094491,   459094491,
+    -970646691,  -970646691,  -970646691,  -970646691,  1010963856,
+    1010963856,  1010963856,  1010963856,  -361743294,  -361743294,
+    -361743294,  -361743294,  209469125,   209469125,   209469125,
+    209469125,   -1053094788, -1053094788, -1053094788, -1053094788,
+    1053160324,  1053160324,  1053160324,  1053160324,  639644520,
+    639644520,   639644520,   639644520,   -862444000,  -862444000,
+    -862444000,  -862444000,  1062144356,  1062144356,  1062144356,
+    1062144356,  -157532337,  -157532337,  -157532337,  -157532337,
+    260914709,   260914709,   260914709,   260914709,   -1041559667,
+    -1041559667, -1041559667, -1041559667, 920985831,   920985831,
+    920985831,   920985831,   -551995675,  -551995675,  -551995675,
+    -551995675,  596522295,   596522295,   596522295,   596522295,
+    892853362,   892853362,   892853362,   892853362,   -892787826,
+    -892787826,  -892787826,  -892787826,  410925857,   410925857,
+    410925857,   410925857,   -992012162,  -992012162,  -992012162,
+    -992012162,  992077698,   992077698,   992077698,   992077698,
+    759246145,   759246145,   759246145,   759246145,   -759180609,
+    -759180609,  -759180609,  -759180609,  -759222975,  -759222975,
+    -759222975,  -759222975,  759288511,   759288511,   759288511,
+    759288511
+  };
+
+  switch (tx_type) {
+    case DCT_DCT:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fdct8x16_1d_column(input + 8 * i, tmp + 8 * i, stride);
+      }
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        fdct16x8_1d_row(tmp + (128 * i), output + (128 * i));
+      }
+      break;
+    case ADST_DCT:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_cols_step1_msa(input + (i << 3), stride, const_arr, ptmpbuf);
+        fadst16_cols_step2_msa(ptmpbuf, const_arr, tmp + (i << 3));
+      }
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        postproc_fdct16x8_1d_row(tmp + (128 * i), output + (128 * i));
+      }
+      break;
+    case DCT_ADST:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fdct8x16_1d_column(input + 8 * i, tmp + 8 * i, stride);
+      }
+
+      fadst16_transpose_postproc_msa(tmp, trans);
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_rows_step1_msa(trans + (i << 7), const_arr, ptmpbuf);
+        fadst16_rows_step2_msa(ptmpbuf, const_arr, tmp + (i << 7));
+      }
+
+      fadst16_transpose_msa(tmp, output);
+      break;
+    case ADST_ADST:
+      /* column transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_cols_step1_msa(input + (i << 3), stride, const_arr, ptmpbuf);
+        fadst16_cols_step2_msa(ptmpbuf, const_arr, tmp + (i << 3));
+      }
+
+      fadst16_transpose_postproc_msa(tmp, trans);
+
+      /* row transform */
+      for (i = 0; i < 2; ++i) {
+        fadst16_rows_step1_msa(trans + (i << 7), const_arr, ptmpbuf);
+        fadst16_rows_step2_msa(ptmpbuf, const_arr, tmp + (i << 7));
+      }
+
+      fadst16_transpose_msa(tmp, output);
+      break;
+    default: assert(0); break;
+  }
+}
diff --git a/third_party/aom/av1/encoder/mips/msa/fdct4x4_msa.c b/third_party/aom/av1/encoder/mips/msa/fdct4x4_msa.c
new file mode 100644
index 0000000000..da1ac74f06
--- /dev/null
+++ b/third_party/aom/av1/encoder/mips/msa/fdct4x4_msa.c
@@ -0,0 +1,98 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/enums.h"
+#include "av1/encoder/mips/msa/fdct_msa.h"
+
+void av1_fwht4x4_msa(const int16_t *input, int16_t *output,
+                     int32_t src_stride) {
+  v8i16 in0, in1, in2, in3, in4;
+
+  LD_SH4(input, src_stride, in0, in1, in2, in3);
+
+  in0 += in1;
+  in3 -= in2;
+  in4 = (in0 - in3) >> 1;
+  SUB2(in4, in1, in4, in2, in1, in2);
+  in0 -= in2;
+  in3 += in1;
+
+  TRANSPOSE4x4_SH_SH(in0, in2, in3, in1, in0, in2, in3, in1);
+
+  in0 += in2;
+  in1 -= in3;
+  in4 = (in0 - in1) >> 1;
+  SUB2(in4, in2, in4, in3, in2, in3);
+  in0 -= in3;
+  in1 += in2;
+
+  SLLI_4V(in0, in1, in2, in3, 2);
+
+  TRANSPOSE4x4_SH_SH(in0, in3, in1, in2, in0, in3, in1, in2);
+
+  ST4x2_UB(in0, output, 4);
+  ST4x2_UB(in3, output + 4, 4);
+  ST4x2_UB(in1, output + 8, 4);
+  ST4x2_UB(in2, output + 12, 4);
+}
+
+void av1_fht4x4_msa(const int16_t *input, int16_t *output, int32_t stride,
+                    int32_t tx_type) {
+  v8i16 in0, in1, in2, in3;
+
+  LD_SH4(input, stride, in0, in1, in2, in3);
+
+  /* fdct4 pre-process */
+  {
+    v8i16 temp, mask;
+    v16i8 zero = { 0 };
+    v16i8 one = __msa_ldi_b(1);
+
+    mask = (v8i16)__msa_sldi_b(zero, one, 15);
+    SLLI_4V(in0, in1, in2, in3, 4);
+    temp = __msa_ceqi_h(in0, 0);
+    temp = (v8i16)__msa_xori_b((v16u8)temp, 255);
+    temp = mask & temp;
+    in0 += temp;
+  }
+
+  switch (tx_type) {
+    case DCT_DCT:
+      AOM_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case ADST_DCT:
+      AOM_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case DCT_ADST:
+      AOM_FDCT4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    case ADST_ADST:
+      AOM_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+      AOM_FADST4(in0, in1, in2, in3, in0, in1, in2, in3);
+      break;
+    default: assert(0); break;
+  }
+
+  TRANSPOSE4x4_SH_SH(in0, in1, in2, in3, in0, in1, in2, in3);
+  ADD4(in0, 1, in1, 1, in2, 1, in3, 1, in0, in1, in2, in3);
+  SRA_4V(in0, in1, in2, in3, 2);
+  PCKEV_D2_SH(in1, in0, in3, in2, in0, in2);
+  ST_SH2(in0, in2, output, 8);
+}
diff --git a/third_party/aom/av1/encoder/mips/msa/fdct8x8_msa.c b/third_party/aom/av1/encoder/mips/msa/fdct8x8_msa.c
new file mode 100644
index 0000000000..4cbf60a11d
--- /dev/null
+++ b/third_party/aom/av1/encoder/mips/msa/fdct8x8_msa.c
@@ -0,0 +1,65 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "av1/common/enums.h"
+#include "av1/encoder/mips/msa/fdct_msa.h"
+
+void av1_fht8x8_msa(const int16_t *input, int16_t *output, int32_t stride,
+                    int32_t tx_type) {
+  v8i16 in0, in1, in2, in3, in4, in5, in6, in7;
+
+  LD_SH8(input, stride, in0, in1, in2, in3, in4, in5, in6, in7);
+  SLLI_4V(in0, in1, in2, in3, 2);
+  SLLI_4V(in4, in5, in6, in7, 2);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      break;
+    case ADST_DCT:
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      break;
+    case DCT_ADST:
+      AOM_FDCT8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      break;
+    case ADST_ADST:
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2,
+                         in3, in4, in5, in6, in7);
+      AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3, in4,
+                in5, in6, in7);
+      break;
+    default: assert(0); break;
+  }
+
+  TRANSPOSE8x8_SH_SH(in0, in1, in2, in3, in4, in5, in6, in7, in0, in1, in2, in3,
+                     in4, in5, in6, in7);
+  SRLI_AVE_S_4V_H(in0, in1, in2, in3, in4, in5, in6, in7);
+  ST_SH8(in0, in1, in2, in3, in4, in5, in6, in7, output, 8);
+}
diff --git a/third_party/aom/av1/encoder/mips/msa/fdct_msa.h b/third_party/aom/av1/encoder/mips/msa/fdct_msa.h
new file mode 100644
index 0000000000..52bcf790c9
--- /dev/null
+++ b/third_party/aom/av1/encoder/mips/msa/fdct_msa.h
@@ -0,0 +1,117 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_MIPS_MSA_AV1_FDCT_MSA_H_
+#define AV1_ENCODER_MIPS_MSA_AV1_FDCT_MSA_H_
+
+#include "aom_dsp/mips/fwd_txfm_msa.h"
+#include "aom_dsp/mips/txfm_macros_msa.h"
+#include "aom_ports/mem.h"
+
+#define AOM_ADST8(in0, in1, in2, in3, in4, in5, in6, in7, out0, out1, out2,  \
+                  out3, out4, out5, out6, out7)                              \
+  {                                                                          \
+    v8i16 cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst4_m;                       \
+    v8i16 vec0_m, vec1_m, vec2_m, vec3_m, s0_m, s1_m;                        \
+    v8i16 coeff0_m = { cospi_2_64,  cospi_6_64,  cospi_10_64, cospi_14_64,   \
+                       cospi_18_64, cospi_22_64, cospi_26_64, cospi_30_64 }; \
+    v8i16 coeff1_m = { cospi_8_64,  -cospi_8_64,  cospi_16_64, -cospi_16_64, \
+                       cospi_24_64, -cospi_24_64, 0,           0 };          \
+                                                                             \
+    SPLATI_H2_SH(coeff0_m, 0, 7, cnst0_m, cnst1_m);                          \
+    cnst2_m = -cnst0_m;                                                      \
+    ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);       \
+    SPLATI_H2_SH(coeff0_m, 4, 3, cnst2_m, cnst3_m);                          \
+    cnst4_m = -cnst2_m;                                                      \
+    ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);       \
+                                                                             \
+    ILVRL_H2_SH(in0, in7, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(in4, in3, vec3_m, vec2_m);                                   \
+    DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m, cnst1_m,  \
+                          cnst2_m, cnst3_m, in7, in0, in4, in3);             \
+                                                                             \
+    SPLATI_H2_SH(coeff0_m, 2, 5, cnst0_m, cnst1_m);                          \
+    cnst2_m = -cnst0_m;                                                      \
+    ILVEV_H2_SH(cnst0_m, cnst1_m, cnst1_m, cnst2_m, cnst0_m, cnst1_m);       \
+    SPLATI_H2_SH(coeff0_m, 6, 1, cnst2_m, cnst3_m);                          \
+    cnst4_m = -cnst2_m;                                                      \
+    ILVEV_H2_SH(cnst2_m, cnst3_m, cnst3_m, cnst4_m, cnst2_m, cnst3_m);       \
+                                                                             \
+    ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                   \
+                                                                             \
+    DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m, cnst1_m,  \
+                          cnst2_m, cnst3_m, in5, in2, in6, in1);             \
+    BUTTERFLY_4(in7, in0, in2, in5, s1_m, s0_m, in2, in5);                   \
+    out7 = -s0_m;                                                            \
+    out0 = s1_m;                                                             \
+                                                                             \
+    SPLATI_H4_SH(coeff1_m, 0, 4, 1, 5, cnst0_m, cnst1_m, cnst2_m, cnst3_m);  \
+                                                                             \
+    ILVEV_H2_SH(cnst3_m, cnst0_m, cnst1_m, cnst2_m, cnst3_m, cnst2_m);       \
+    cnst0_m = __msa_ilvev_h(cnst1_m, cnst0_m);                               \
+    cnst1_m = cnst0_m;                                                       \
+                                                                             \
+    ILVRL_H2_SH(in4, in3, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(in6, in1, vec3_m, vec2_m);                                   \
+    DOT_ADD_SUB_SRARI_PCK(vec0_m, vec1_m, vec2_m, vec3_m, cnst0_m, cnst2_m,  \
+                          cnst3_m, cnst1_m, out1, out6, s0_m, s1_m);         \
+                                                                             \
+    SPLATI_H2_SH(coeff1_m, 2, 3, cnst0_m, cnst1_m);                          \
+    cnst1_m = __msa_ilvev_h(cnst1_m, cnst0_m);                               \
+                                                                             \
+    ILVRL_H2_SH(in2, in5, vec1_m, vec0_m);                                   \
+    ILVRL_H2_SH(s0_m, s1_m, vec3_m, vec2_m);                                 \
+    out3 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst0_m);                   \
+    out4 = DOT_SHIFT_RIGHT_PCK_H(vec0_m, vec1_m, cnst1_m);                   \
+    out2 = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst0_m);                   \
+    out5 = DOT_SHIFT_RIGHT_PCK_H(vec2_m, vec3_m, cnst1_m);                   \
+                                                                             \
+    out1 = -out1;                                                            \
+    out3 = -out3;                                                            \
+    out5 = -out5;                                                            \
+  }
+
+#define AOM_FADST4(in0, in1, in2, in3, out0, out1, out2, out3)              \
+  {                                                                         \
+    v4i32 s0_m, s1_m, s2_m, s3_m, constant_m;                               \
+    v4i32 in0_r_m, in1_r_m, in2_r_m, in3_r_m;                               \
+                                                                            \
+    UNPCK_R_SH_SW(in0, in0_r_m);                                            \
+    UNPCK_R_SH_SW(in1, in1_r_m);                                            \
+    UNPCK_R_SH_SW(in2, in2_r_m);                                            \
+    UNPCK_R_SH_SW(in3, in3_r_m);                                            \
+                                                                            \
+    constant_m = __msa_fill_w(sinpi_4_9);                                   \
+    MUL2(in0_r_m, constant_m, in3_r_m, constant_m, s1_m, s0_m);             \
+                                                                            \
+    constant_m = __msa_fill_w(sinpi_1_9);                                   \
+    s0_m += in0_r_m * constant_m;                                           \
+    s1_m -= in1_r_m * constant_m;                                           \
+                                                                            \
+    constant_m = __msa_fill_w(sinpi_2_9);                                   \
+    s0_m += in1_r_m * constant_m;                                           \
+    s1_m += in3_r_m * constant_m;                                           \
+                                                                            \
+    s2_m = in0_r_m + in1_r_m - in3_r_m;                                     \
+                                                                            \
+    constant_m = __msa_fill_w(sinpi_3_9);                                   \
+    MUL2(in2_r_m, constant_m, s2_m, constant_m, s3_m, in1_r_m);             \
+                                                                            \
+    in0_r_m = s0_m + s3_m;                                                  \
+    s2_m = s1_m - s3_m;                                                     \
+    s3_m = s1_m - s0_m + s3_m;                                              \
+                                                                            \
+    SRARI_W4_SW(in0_r_m, in1_r_m, s2_m, s3_m, DCT_CONST_BITS);              \
+    PCKEV_H4_SH(in0_r_m, in0_r_m, in1_r_m, in1_r_m, s2_m, s2_m, s3_m, s3_m, \
+                out0, out1, out2, out3);                                    \
+  }
+#endif  // AV1_ENCODER_MIPS_MSA_AV1_FDCT_MSA_H_
diff --git a/third_party/aom/av1/encoder/mips/msa/temporal_filter_msa.c b/third_party/aom/av1/encoder/mips/msa/temporal_filter_msa.c
new file mode 100644
index 0000000000..4ec679642a
--- /dev/null
+++ b/third_party/aom/av1/encoder/mips/msa/temporal_filter_msa.c
@@ -0,0 +1,284 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./av1_rtcd.h"
+#include "aom_dsp/mips/macros_msa.h"
+
+static void temporal_filter_apply_8size_msa(uint8_t *frm1_ptr, uint32_t stride,
+                                            uint8_t *frm2_ptr, int32_t filt_sth,
+                                            int32_t filt_wgt, uint32_t *acc,
+                                            uint16_t *cnt) {
+  uint32_t row;
+  uint64_t f0, f1, f2, f3;
+  v16i8 frm2, frm1 = { 0 };
+  v16i8 frm4, frm3 = { 0 };
+  v16u8 frm_r, frm_l;
+  v8i16 frm2_r, frm2_l;
+  v8i16 diff0, diff1, mod0_h, mod1_h;
+  v4i32 cnst3, cnst16, filt_wt, strength;
+  v4i32 mod0_w, mod1_w, mod2_w, mod3_w;
+  v4i32 diff0_r, diff0_l, diff1_r, diff1_l;
+  v4i32 frm2_rr, frm2_rl, frm2_lr, frm2_ll;
+  v4i32 acc0, acc1, acc2, acc3;
+  v8i16 cnt0, cnt1;
+
+  filt_wt = __msa_fill_w(filt_wgt);
+  strength = __msa_fill_w(filt_sth);
+  cnst3 = __msa_ldi_w(3);
+  cnst16 = __msa_ldi_w(16);
+
+  for (row = 2; row--;) {
+    LD4(frm1_ptr, stride, f0, f1, f2, f3);
+    frm1_ptr += (4 * stride);
+
+    LD_SB2(frm2_ptr, 16, frm2, frm4);
+    frm2_ptr += 32;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc + 8, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    INSERT_D2_SB(f0, f1, frm1);
+    INSERT_D2_SB(f2, f3, frm3);
+    ILVRL_B2_UB(frm1, frm2, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l, diff1_l,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+
+    UNPCK_UB_SH(frm2, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3, mod0_w, mod1_w,
+         mod2_w, mod3_w);
+
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc + 8, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    ILVRL_B2_UB(frm3, frm4, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l, diff1_l,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+    UNPCK_UB_SH(frm4, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3, mod0_w, mod1_w,
+         mod2_w, mod3_w);
+
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+  }
+}
+
+static void temporal_filter_apply_16size_msa(uint8_t *frm1_ptr, uint32_t stride,
+                                             uint8_t *frm2_ptr,
+                                             int32_t filt_sth, int32_t filt_wgt,
+                                             uint32_t *acc, uint16_t *cnt) {
+  uint32_t row;
+  v16i8 frm1, frm2, frm3, frm4;
+  v16u8 frm_r, frm_l;
+  v16i8 zero = { 0 };
+  v8u16 frm2_r, frm2_l;
+  v8i16 diff0, diff1, mod0_h, mod1_h;
+  v4i32 cnst3, cnst16, filt_wt, strength;
+  v4i32 mod0_w, mod1_w, mod2_w, mod3_w;
+  v4i32 diff0_r, diff0_l, diff1_r, diff1_l;
+  v4i32 frm2_rr, frm2_rl, frm2_lr, frm2_ll;
+  v4i32 acc0, acc1, acc2, acc3;
+  v8i16 cnt0, cnt1;
+
+  filt_wt = __msa_fill_w(filt_wgt);
+  strength = __msa_fill_w(filt_sth);
+  cnst3 = __msa_ldi_w(3);
+  cnst16 = __msa_ldi_w(16);
+
+  for (row = 8; row--;) {
+    LD_SB2(frm1_ptr, stride, frm1, frm3);
+    frm1_ptr += stride;
+
+    LD_SB2(frm2_ptr, 16, frm2, frm4);
+    frm2_ptr += 16;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    ILVRL_B2_UB(frm1, frm2, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l, diff1_l,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+
+    ILVRL_B2_UH(zero, frm2, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3, mod0_w, mod1_w,
+         mod2_w, mod3_w);
+
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+
+    LD_SW2(acc, 4, acc0, acc1);
+    LD_SW2(acc + 8, 4, acc2, acc3);
+    LD_SH2(cnt, 8, cnt0, cnt1);
+
+    ILVRL_B2_UB(frm3, frm4, frm_r, frm_l);
+    HSUB_UB2_SH(frm_r, frm_l, diff0, diff1);
+    UNPCK_SH_SW(diff0, diff0_r, diff0_l);
+    UNPCK_SH_SW(diff1, diff1_r, diff1_l);
+    MUL4(diff0_r, diff0_r, diff0_l, diff0_l, diff1_r, diff1_r, diff1_l, diff1_l,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    MUL4(mod0_w, cnst3, mod1_w, cnst3, mod2_w, cnst3, mod3_w, cnst3, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+    SRAR_W4_SW(mod0_w, mod1_w, mod2_w, mod3_w, strength);
+
+    diff0_r = (mod0_w < cnst16);
+    diff0_l = (mod1_w < cnst16);
+    diff1_r = (mod2_w < cnst16);
+    diff1_l = (mod3_w < cnst16);
+
+    SUB4(cnst16, mod0_w, cnst16, mod1_w, cnst16, mod2_w, cnst16, mod3_w, mod0_w,
+         mod1_w, mod2_w, mod3_w);
+
+    mod0_w = diff0_r & mod0_w;
+    mod1_w = diff0_l & mod1_w;
+    mod2_w = diff1_r & mod2_w;
+    mod3_w = diff1_l & mod3_w;
+
+    MUL4(mod0_w, filt_wt, mod1_w, filt_wt, mod2_w, filt_wt, mod3_w, filt_wt,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    PCKEV_H2_SH(mod1_w, mod0_w, mod3_w, mod2_w, mod0_h, mod1_h);
+    ADD2(mod0_h, cnt0, mod1_h, cnt1, mod0_h, mod1_h);
+    ST_SH2(mod0_h, mod1_h, cnt, 8);
+    cnt += 16;
+
+    ILVRL_B2_UH(zero, frm4, frm2_r, frm2_l);
+    UNPCK_SH_SW(frm2_r, frm2_rr, frm2_rl);
+    UNPCK_SH_SW(frm2_l, frm2_lr, frm2_ll);
+    MUL4(mod0_w, frm2_rr, mod1_w, frm2_rl, mod2_w, frm2_lr, mod3_w, frm2_ll,
+         mod0_w, mod1_w, mod2_w, mod3_w);
+    ADD4(mod0_w, acc0, mod1_w, acc1, mod2_w, acc2, mod3_w, acc3, mod0_w, mod1_w,
+         mod2_w, mod3_w);
+    ST_SW2(mod0_w, mod1_w, acc, 4);
+    acc += 8;
+    ST_SW2(mod2_w, mod3_w, acc, 4);
+    acc += 8;
+
+    frm1_ptr += stride;
+    frm2_ptr += 16;
+  }
+}
+
+void av1_temporal_filter_apply_msa(uint8_t *frame1_ptr, uint32_t stride,
+                                   uint8_t *frame2_ptr, uint32_t blk_w,
+                                   uint32_t blk_h, int32_t strength,
+                                   int32_t filt_wgt, uint32_t *accu,
+                                   uint16_t *cnt) {
+  if (8 == (blk_w * blk_h)) {
+    temporal_filter_apply_8size_msa(frame1_ptr, stride, frame2_ptr, strength,
+                                    filt_wgt, accu, cnt);
+  } else if (16 == (blk_w * blk_h)) {
+    temporal_filter_apply_16size_msa(frame1_ptr, stride, frame2_ptr, strength,
+                                     filt_wgt, accu, cnt);
+  } else {
+    av1_temporal_filter_apply_c(frame1_ptr, stride, frame2_ptr, blk_w, blk_h,
+                                strength, filt_wgt, accu, cnt);
+  }
+}
diff --git a/third_party/aom/av1/encoder/palette.c b/third_party/aom/av1/encoder/palette.c
new file mode 100644
index 0000000000..355141de55
--- /dev/null
+++ b/third_party/aom/av1/encoder/palette.c
@@ -0,0 +1,277 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+
+#include "av1/encoder/cost.h"
+#include "av1/encoder/palette.h"
+
+static float calc_dist(const float *p1, const float *p2, int dim) {
+  float dist = 0;
+  int i;
+  for (i = 0; i < dim; ++i) {
+    const float diff = p1[i] - p2[i];
+    dist += diff * diff;
+  }
+  return dist;
+}
+
+void av1_calc_indices(const float *data, const float *centroids,
+                      uint8_t *indices, int n, int k, int dim) {
+  int i, j;
+  for (i = 0; i < n; ++i) {
+    float min_dist = calc_dist(data + i * dim, centroids, dim);
+    indices[i] = 0;
+    for (j = 1; j < k; ++j) {
+      const float this_dist =
+          calc_dist(data + i * dim, centroids + j * dim, dim);
+      if (this_dist < min_dist) {
+        min_dist = this_dist;
+        indices[i] = j;
+      }
+    }
+  }
+}
+
+// Generate a random number in the range [0, 32768).
+static unsigned int lcg_rand16(unsigned int *state) {
+  *state = (unsigned int)(*state * 1103515245ULL + 12345);
+  return *state / 65536 % 32768;
+}
+
+static void calc_centroids(const float *data, float *centroids,
+                           const uint8_t *indices, int n, int k, int dim) {
+  int i, j, index;
+  int count[PALETTE_MAX_SIZE];
+  unsigned int rand_state = (unsigned int)data[0];
+
+  assert(n <= 32768);
+
+  memset(count, 0, sizeof(count[0]) * k);
+  memset(centroids, 0, sizeof(centroids[0]) * k * dim);
+
+  for (i = 0; i < n; ++i) {
+    index = indices[i];
+    assert(index < k);
+    ++count[index];
+    for (j = 0; j < dim; ++j) {
+      centroids[index * dim + j] += data[i * dim + j];
+    }
+  }
+
+  for (i = 0; i < k; ++i) {
+    if (count[i] == 0) {
+      memcpy(centroids + i * dim, data + (lcg_rand16(&rand_state) % n) * dim,
+             sizeof(centroids[0]) * dim);
+    } else {
+      const float norm = 1.0f / count[i];
+      for (j = 0; j < dim; ++j) centroids[i * dim + j] *= norm;
+    }
+  }
+
+  // Round to nearest integers.
+  for (i = 0; i < k * dim; ++i) {
+    centroids[i] = roundf(centroids[i]);
+  }
+}
+
+static float calc_total_dist(const float *data, const float *centroids,
+                             const uint8_t *indices, int n, int k, int dim) {
+  float dist = 0;
+  int i;
+  (void)k;
+
+  for (i = 0; i < n; ++i)
+    dist += calc_dist(data + i * dim, centroids + indices[i] * dim, dim);
+
+  return dist;
+}
+
+void av1_k_means(const float *data, float *centroids, uint8_t *indices, int n,
+                 int k, int dim, int max_itr) {
+  int i;
+  float this_dist;
+  float pre_centroids[2 * PALETTE_MAX_SIZE];
+  uint8_t pre_indices[MAX_SB_SQUARE];
+
+  av1_calc_indices(data, centroids, indices, n, k, dim);
+  this_dist = calc_total_dist(data, centroids, indices, n, k, dim);
+
+  for (i = 0; i < max_itr; ++i) {
+    const float pre_dist = this_dist;
+    memcpy(pre_centroids, centroids, sizeof(pre_centroids[0]) * k * dim);
+    memcpy(pre_indices, indices, sizeof(pre_indices[0]) * n);
+
+    calc_centroids(data, centroids, indices, n, k, dim);
+    av1_calc_indices(data, centroids, indices, n, k, dim);
+    this_dist = calc_total_dist(data, centroids, indices, n, k, dim);
+
+    if (this_dist > pre_dist) {
+      memcpy(centroids, pre_centroids, sizeof(pre_centroids[0]) * k * dim);
+      memcpy(indices, pre_indices, sizeof(pre_indices[0]) * n);
+      break;
+    }
+    if (!memcmp(centroids, pre_centroids, sizeof(pre_centroids[0]) * k * dim))
+      break;
+  }
+}
+
+static int float_comparer(const void *a, const void *b) {
+  const float fa = *(const float *)a;
+  const float fb = *(const float *)b;
+  return (fa > fb) - (fa < fb);
+}
+
+int av1_remove_duplicates(float *centroids, int num_centroids) {
+  int num_unique;  // number of unique centroids
+  int i;
+  qsort(centroids, num_centroids, sizeof(*centroids), float_comparer);
+  // Remove duplicates.
+  num_unique = 1;
+  for (i = 1; i < num_centroids; ++i) {
+    if (centroids[i] != centroids[i - 1]) {  // found a new unique centroid
+      centroids[num_unique++] = centroids[i];
+    }
+  }
+  return num_unique;
+}
+
+int av1_count_colors(const uint8_t *src, int stride, int rows, int cols) {
+  int n = 0, r, c, i, val_count[256];
+  uint8_t val;
+  memset(val_count, 0, sizeof(val_count));
+
+  for (r = 0; r < rows; ++r) {
+    for (c = 0; c < cols; ++c) {
+      val = src[r * stride + c];
+      ++val_count[val];
+    }
+  }
+
+  for (i = 0; i < 256; ++i) {
+    if (val_count[i]) {
+      ++n;
+    }
+  }
+
+  return n;
+}
+
+#if CONFIG_PALETTE_DELTA_ENCODING
+int av1_get_palette_delta_bits_y(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *min_bits) {
+  const int n = pmi->palette_size[0];
+  int max_d = 0, i;
+  *min_bits = bit_depth - 3;
+  for (i = 1; i < n; ++i) {
+    const int delta = pmi->palette_colors[i] - pmi->palette_colors[i - 1];
+    assert(delta > 0);
+    if (delta > max_d) max_d = delta;
+  }
+  return AOMMAX(av1_ceil_log2(max_d), *min_bits);
+}
+
+int av1_get_palette_delta_bits_u(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *min_bits) {
+  const int n = pmi->palette_size[1];
+  int max_d = 0, i;
+  *min_bits = bit_depth - 3;
+  for (i = 1; i < n; ++i) {
+    const int delta = pmi->palette_colors[PALETTE_MAX_SIZE + i] -
+                      pmi->palette_colors[PALETTE_MAX_SIZE + i - 1];
+    assert(delta >= 0);
+    if (delta > max_d) max_d = delta;
+  }
+  return AOMMAX(av1_ceil_log2(max_d + 1), *min_bits);
+}
+
+int av1_get_palette_delta_bits_v(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *zero_count,
+                                 int *min_bits) {
+  const int n = pmi->palette_size[1];
+  const int max_val = 1 << bit_depth;
+  int max_d = 0, i;
+  *min_bits = bit_depth - 4;
+  *zero_count = 0;
+  for (i = 1; i < n; ++i) {
+    const int delta = pmi->palette_colors[2 * PALETTE_MAX_SIZE + i] -
+                      pmi->palette_colors[2 * PALETTE_MAX_SIZE + i - 1];
+    const int v = abs(delta);
+    const int d = AOMMIN(v, max_val - v);
+    if (d > max_d) max_d = d;
+    if (d == 0) ++(*zero_count);
+  }
+  return AOMMAX(av1_ceil_log2(max_d + 1), *min_bits);
+}
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+
+int av1_palette_color_cost_y(const PALETTE_MODE_INFO *const pmi,
+                             int bit_depth) {
+  const int n = pmi->palette_size[0];
+#if CONFIG_PALETTE_DELTA_ENCODING
+  int min_bits = 0;
+  const int bits = av1_get_palette_delta_bits_y(pmi, bit_depth, &min_bits);
+  return av1_cost_bit(128, 0) * (2 + bit_depth + bits * (n - 1));
+#else
+  return bit_depth * n * av1_cost_bit(128, 0);
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+}
+
+int av1_palette_color_cost_uv(const PALETTE_MODE_INFO *const pmi,
+                              int bit_depth) {
+  const int n = pmi->palette_size[1];
+#if CONFIG_PALETTE_DELTA_ENCODING
+  int cost = 0;
+  // U channel palette color cost.
+  int min_bits_u = 0;
+  const int bits_u = av1_get_palette_delta_bits_u(pmi, bit_depth, &min_bits_u);
+  cost += av1_cost_bit(128, 0) * (2 + bit_depth + bits_u * (n - 1));
+  // V channel palette color cost.
+  int zero_count = 0, min_bits_v = 0;
+  const int bits_v =
+      av1_get_palette_delta_bits_v(pmi, bit_depth, &zero_count, &min_bits_v);
+  const int bits_using_delta =
+      2 + bit_depth + (bits_v + 1) * (n - 1) - zero_count;
+  const int bits_using_raw = bit_depth * n;
+  cost += av1_cost_bit(128, 0) * (1 + AOMMIN(bits_using_delta, bits_using_raw));
+  return cost;
+#else
+  return 2 * bit_depth * n * av1_cost_bit(128, 0);
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+}
+
+#if CONFIG_HIGHBITDEPTH
+int av1_count_colors_highbd(const uint8_t *src8, int stride, int rows, int cols,
+                            int bit_depth) {
+  int n = 0, r, c, i;
+  uint16_t val;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  int val_count[1 << 12];
+
+  assert(bit_depth <= 12);
+  memset(val_count, 0, (1 << 12) * sizeof(val_count[0]));
+  for (r = 0; r < rows; ++r) {
+    for (c = 0; c < cols; ++c) {
+      val = src[r * stride + c];
+      ++val_count[val];
+    }
+  }
+
+  for (i = 0; i < (1 << bit_depth); ++i) {
+    if (val_count[i]) {
+      ++n;
+    }
+  }
+
+  return n;
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/av1/encoder/palette.h b/third_party/aom/av1/encoder/palette.h
new file mode 100644
index 0000000000..5403ac5e60
--- /dev/null
+++ b/third_party/aom/av1/encoder/palette.h
@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_PALETTE_H_
+#define AV1_ENCODER_PALETTE_H_
+
+#include "av1/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Given 'n' 'data' points and 'k' 'centroids' each of dimension 'dim',
+// calculate the centroid 'indices' for the data points.
+void av1_calc_indices(const float *data, const float *centroids,
+                      uint8_t *indices, int n, int k, int dim);
+
+// Given 'n' 'data' points and an initial guess of 'k' 'centroids' each of
+// dimension 'dim', runs up to 'max_itr' iterations of k-means algorithm to get
+// updated 'centroids' and the centroid 'indices' for elements in 'data'.
+// Note: the output centroids are rounded off to nearest integers.
+void av1_k_means(const float *data, float *centroids, uint8_t *indices, int n,
+                 int k, int dim, int max_itr);
+
+// Given a list of centroids, returns the unique number of centroids 'k', and
+// puts these unique centroids in first 'k' indices of 'centroids' array.
+// Ideally, the centroids should be rounded to integers before calling this
+// method.
+int av1_remove_duplicates(float *centroids, int num_centroids);
+
+// Returns the number of colors in 'src'.
+int av1_count_colors(const uint8_t *src, int stride, int rows, int cols);
+#if CONFIG_HIGHBITDEPTH
+// Same as av1_count_colors(), but for high-bitdepth mode.
+int av1_count_colors_highbd(const uint8_t *src8, int stride, int rows, int cols,
+                            int bit_depth);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_PALETTE_DELTA_ENCODING
+// Return the number of bits used to transmit each luma palette color delta.
+int av1_get_palette_delta_bits_y(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *min_bits);
+
+// Return the number of bits used to transmit each U palette color delta.
+int av1_get_palette_delta_bits_u(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *min_bits);
+
+// Return the number of bits used to transmit each v palette color delta;
+// assign zero_count with the number of deltas being 0.
+int av1_get_palette_delta_bits_v(const PALETTE_MODE_INFO *const pmi,
+                                 int bit_depth, int *zero_count, int *min_bits);
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+
+// Return the rate cost for transmitting luma palette color values.
+int av1_palette_color_cost_y(const PALETTE_MODE_INFO *const pmi, int bit_depth);
+
+// Return the rate cost for transmitting chroma palette color values.
+int av1_palette_color_cost_uv(const PALETTE_MODE_INFO *const pmi,
+                              int bit_depth);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif /* AV1_ENCODER_PALETTE_H_ */
diff --git a/third_party/aom/av1/encoder/pickcdef.c b/third_party/aom/av1/encoder/pickcdef.c
new file mode 100644
index 0000000000..da64fb48d6
--- /dev/null
+++ b/third_party/aom/av1/encoder/pickcdef.c
@@ -0,0 +1,490 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <string.h>
+
+#include "./aom_scale_rtcd.h"
+#include "aom/aom_integer.h"
+#include "av1/common/cdef.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/reconinter.h"
+#include "av1/encoder/encoder.h"
+
+#define TOTAL_STRENGTHS (DERING_STRENGTHS * CLPF_STRENGTHS)
+
+/* Search for the best strength to add as an option, knowing we
+   already selected nb_strengths options. */
+static uint64_t search_one(int *lev, int nb_strengths,
+                           uint64_t mse[][TOTAL_STRENGTHS], int sb_count) {
+  uint64_t tot_mse[TOTAL_STRENGTHS];
+  int i, j;
+  uint64_t best_tot_mse = (uint64_t)1 << 63;
+  int best_id = 0;
+  memset(tot_mse, 0, sizeof(tot_mse));
+  for (i = 0; i < sb_count; i++) {
+    int gi;
+    uint64_t best_mse = (uint64_t)1 << 63;
+    /* Find best mse among already selected options. */
+    for (gi = 0; gi < nb_strengths; gi++) {
+      if (mse[i][lev[gi]] < best_mse) {
+        best_mse = mse[i][lev[gi]];
+      }
+    }
+    /* Find best mse when adding each possible new option. */
+    for (j = 0; j < TOTAL_STRENGTHS; j++) {
+      uint64_t best = best_mse;
+      if (mse[i][j] < best) best = mse[i][j];
+      tot_mse[j] += best;
+    }
+  }
+  for (j = 0; j < TOTAL_STRENGTHS; j++) {
+    if (tot_mse[j] < best_tot_mse) {
+      best_tot_mse = tot_mse[j];
+      best_id = j;
+    }
+  }
+  lev[nb_strengths] = best_id;
+  return best_tot_mse;
+}
+
+/* Search for the best luma+chroma strength to add as an option, knowing we
+   already selected nb_strengths options. */
+static uint64_t search_one_dual(int *lev0, int *lev1, int nb_strengths,
+                                uint64_t (**mse)[TOTAL_STRENGTHS],
+                                int sb_count) {
+  uint64_t tot_mse[TOTAL_STRENGTHS][TOTAL_STRENGTHS];
+  int i, j;
+  uint64_t best_tot_mse = (uint64_t)1 << 63;
+  int best_id0 = 0;
+  int best_id1 = 0;
+  memset(tot_mse, 0, sizeof(tot_mse));
+  for (i = 0; i < sb_count; i++) {
+    int gi;
+    uint64_t best_mse = (uint64_t)1 << 63;
+    /* Find best mse among already selected options. */
+    for (gi = 0; gi < nb_strengths; gi++) {
+      uint64_t curr = mse[0][i][lev0[gi]];
+      curr += mse[1][i][lev1[gi]];
+      if (curr < best_mse) {
+        best_mse = curr;
+      }
+    }
+    /* Find best mse when adding each possible new option. */
+    for (j = 0; j < TOTAL_STRENGTHS; j++) {
+      int k;
+      for (k = 0; k < TOTAL_STRENGTHS; k++) {
+        uint64_t best = best_mse;
+        uint64_t curr = mse[0][i][j];
+        curr += mse[1][i][k];
+        if (curr < best) best = curr;
+        tot_mse[j][k] += best;
+      }
+    }
+  }
+  for (j = 0; j < TOTAL_STRENGTHS; j++) {
+    int k;
+    for (k = 0; k < TOTAL_STRENGTHS; k++) {
+      if (tot_mse[j][k] < best_tot_mse) {
+        best_tot_mse = tot_mse[j][k];
+        best_id0 = j;
+        best_id1 = k;
+      }
+    }
+  }
+  lev0[nb_strengths] = best_id0;
+  lev1[nb_strengths] = best_id1;
+  return best_tot_mse;
+}
+
+/* Search for the set of strengths that minimizes mse. */
+static uint64_t joint_strength_search(int *best_lev, int nb_strengths,
+                                      uint64_t mse[][TOTAL_STRENGTHS],
+                                      int sb_count) {
+  uint64_t best_tot_mse;
+  int i;
+  best_tot_mse = (uint64_t)1 << 63;
+  /* Greedy search: add one strength options at a time. */
+  for (i = 0; i < nb_strengths; i++) {
+    best_tot_mse = search_one(best_lev, i, mse, sb_count);
+  }
+  /* Trying to refine the greedy search by reconsidering each
+     already-selected option. */
+  for (i = 0; i < 4 * nb_strengths; i++) {
+    int j;
+    for (j = 0; j < nb_strengths - 1; j++) best_lev[j] = best_lev[j + 1];
+    best_tot_mse = search_one(best_lev, nb_strengths - 1, mse, sb_count);
+  }
+  return best_tot_mse;
+}
+
+/* Search for the set of luma+chroma strengths that minimizes mse. */
+static uint64_t joint_strength_search_dual(int *best_lev0, int *best_lev1,
+                                           int nb_strengths,
+                                           uint64_t (**mse)[TOTAL_STRENGTHS],
+                                           int sb_count) {
+  uint64_t best_tot_mse;
+  int i;
+  best_tot_mse = (uint64_t)1 << 63;
+  /* Greedy search: add one strength options at a time. */
+  for (i = 0; i < nb_strengths; i++) {
+    best_tot_mse = search_one_dual(best_lev0, best_lev1, i, mse, sb_count);
+  }
+  /* Trying to refine the greedy search by reconsidering each
+     already-selected option. */
+  for (i = 0; i < 4 * nb_strengths; i++) {
+    int j;
+    for (j = 0; j < nb_strengths - 1; j++) {
+      best_lev0[j] = best_lev0[j + 1];
+      best_lev1[j] = best_lev1[j + 1];
+    }
+    best_tot_mse =
+        search_one_dual(best_lev0, best_lev1, nb_strengths - 1, mse, sb_count);
+  }
+  return best_tot_mse;
+}
+
+/* FIXME: SSE-optimize this. */
+static void copy_sb16_16(uint16_t *dst, int dstride, const uint16_t *src,
+                         int src_voffset, int src_hoffset, int sstride,
+                         int vsize, int hsize) {
+  int r, c;
+  const uint16_t *base = &src[src_voffset * sstride + src_hoffset];
+  for (r = 0; r < vsize; r++) {
+    for (c = 0; c < hsize; c++) {
+      dst[r * dstride + c] = base[r * sstride + c];
+    }
+  }
+}
+
+static INLINE uint64_t dist_8x8_16bit(uint16_t *dst, int dstride, uint16_t *src,
+                                      int sstride, int coeff_shift) {
+  uint64_t svar = 0;
+  uint64_t dvar = 0;
+  uint64_t sum_s = 0;
+  uint64_t sum_d = 0;
+  uint64_t sum_s2 = 0;
+  uint64_t sum_d2 = 0;
+  uint64_t sum_sd = 0;
+  int i, j;
+  for (i = 0; i < 8; i++) {
+    for (j = 0; j < 8; j++) {
+      sum_s += src[i * sstride + j];
+      sum_d += dst[i * dstride + j];
+      sum_s2 += src[i * sstride + j] * src[i * sstride + j];
+      sum_d2 += dst[i * dstride + j] * dst[i * dstride + j];
+      sum_sd += src[i * sstride + j] * dst[i * dstride + j];
+    }
+  }
+  /* Compute the variance -- the calculation cannot go negative. */
+  svar = sum_s2 - ((sum_s * sum_s + 32) >> 6);
+  dvar = sum_d2 - ((sum_d * sum_d + 32) >> 6);
+  return (uint64_t)floor(
+      .5 +
+      (sum_d2 + sum_s2 - 2 * sum_sd) * .5 *
+          (svar + dvar + (400 << 2 * coeff_shift)) /
+          (sqrt((20000 << 4 * coeff_shift) + svar * (double)dvar)));
+}
+
+static INLINE uint64_t mse_8x8_16bit(uint16_t *dst, int dstride, uint16_t *src,
+                                     int sstride) {
+  uint64_t sum = 0;
+  int i, j;
+  for (i = 0; i < 8; i++) {
+    for (j = 0; j < 8; j++) {
+      int e = dst[i * dstride + j] - src[i * sstride + j];
+      sum += e * e;
+    }
+  }
+  return sum;
+}
+
+static INLINE uint64_t mse_4x4_16bit(uint16_t *dst, int dstride, uint16_t *src,
+                                     int sstride) {
+  uint64_t sum = 0;
+  int i, j;
+  for (i = 0; i < 4; i++) {
+    for (j = 0; j < 4; j++) {
+      int e = dst[i * dstride + j] - src[i * sstride + j];
+      sum += e * e;
+    }
+  }
+  return sum;
+}
+
+/* Compute MSE only on the blocks we filtered. */
+uint64_t compute_dering_dist(uint16_t *dst, int dstride, uint16_t *src,
+                             dering_list *dlist, int dering_count,
+                             BLOCK_SIZE bsize, int coeff_shift, int pli) {
+  uint64_t sum = 0;
+  int bi, bx, by;
+  if (bsize == BLOCK_8X8) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      if (pli == 0) {
+        sum += dist_8x8_16bit(&dst[(by << 3) * dstride + (bx << 3)], dstride,
+                              &src[bi << (3 + 3)], 8, coeff_shift);
+      } else {
+        sum += mse_8x8_16bit(&dst[(by << 3) * dstride + (bx << 3)], dstride,
+                             &src[bi << (3 + 3)], 8);
+      }
+    }
+  } else if (bsize == BLOCK_4X8) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      sum += mse_4x4_16bit(&dst[(by << 3) * dstride + (bx << 2)], dstride,
+                           &src[bi << (3 + 2)], 4);
+      sum += mse_4x4_16bit(&dst[((by << 3) + 4) * dstride + (bx << 2)], dstride,
+                           &src[(bi << (3 + 2)) + 4 * 4], 4);
+    }
+  } else if (bsize == BLOCK_8X4) {
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      sum += mse_4x4_16bit(&dst[(by << 2) * dstride + (bx << 3)], dstride,
+                           &src[bi << (2 + 3)], 8);
+      sum += mse_4x4_16bit(&dst[(by << 2) * dstride + (bx << 3) + 4], dstride,
+                           &src[(bi << (2 + 3)) + 4], 8);
+    }
+  } else {
+    assert(bsize == BLOCK_4X4);
+    for (bi = 0; bi < dering_count; bi++) {
+      by = dlist[bi].by;
+      bx = dlist[bi].bx;
+      sum += mse_4x4_16bit(&dst[(by << 2) * dstride + (bx << 2)], dstride,
+                           &src[bi << (2 + 2)], 4);
+    }
+  }
+  return sum >> 2 * coeff_shift;
+}
+
+void av1_cdef_search(YV12_BUFFER_CONFIG *frame, const YV12_BUFFER_CONFIG *ref,
+                     AV1_COMMON *cm, MACROBLOCKD *xd) {
+  int r, c;
+  int sbr, sbc;
+  uint16_t *src[3];
+  uint16_t *ref_coeff[3];
+  dering_list dlist[MAX_MIB_SIZE * MAX_MIB_SIZE];
+  int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS] = { { 0 } };
+  int var[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS] = { { 0 } };
+  int stride[3];
+  int bsize[3];
+  int mi_wide_l2[3];
+  int mi_high_l2[3];
+  int xdec[3];
+  int ydec[3];
+  int pli;
+  int dering_count;
+  int coeff_shift = AOMMAX(cm->bit_depth - 8, 0);
+  uint64_t best_tot_mse = (uint64_t)1 << 63;
+  uint64_t tot_mse;
+  int sb_count;
+  int nvsb = (cm->mi_rows + MAX_MIB_SIZE - 1) / MAX_MIB_SIZE;
+  int nhsb = (cm->mi_cols + MAX_MIB_SIZE - 1) / MAX_MIB_SIZE;
+  int *sb_index = aom_malloc(nvsb * nhsb * sizeof(*sb_index));
+  int *selected_strength = aom_malloc(nvsb * nhsb * sizeof(*sb_index));
+  uint64_t(*mse[2])[TOTAL_STRENGTHS];
+  int clpf_damping = 3 + (cm->base_qindex >> 6);
+  int dering_damping = 6;
+  int i;
+  int nb_strengths;
+  int nb_strength_bits;
+  int quantizer;
+  double lambda;
+  int nplanes = 3;
+  DECLARE_ALIGNED(32, uint16_t, inbuf[OD_DERING_INBUF_SIZE]);
+  uint16_t *in;
+  DECLARE_ALIGNED(32, uint16_t, tmp_dst[MAX_SB_SQUARE]);
+  int chroma_dering =
+      xd->plane[1].subsampling_x == xd->plane[1].subsampling_y &&
+      xd->plane[2].subsampling_x == xd->plane[2].subsampling_y;
+  quantizer =
+      av1_ac_quant(cm->base_qindex, 0, cm->bit_depth) >> (cm->bit_depth - 8);
+  lambda = .12 * quantizer * quantizer / 256.;
+
+  av1_setup_dst_planes(xd->plane, cm->sb_size, frame, 0, 0);
+  mse[0] = aom_malloc(sizeof(**mse) * nvsb * nhsb);
+  mse[1] = aom_malloc(sizeof(**mse) * nvsb * nhsb);
+  for (pli = 0; pli < nplanes; pli++) {
+    uint8_t *ref_buffer;
+    int ref_stride;
+    switch (pli) {
+      case 0:
+        ref_buffer = ref->y_buffer;
+        ref_stride = ref->y_stride;
+        break;
+      case 1:
+        ref_buffer = ref->u_buffer;
+        ref_stride = ref->uv_stride;
+        break;
+      case 2:
+        ref_buffer = ref->v_buffer;
+        ref_stride = ref->uv_stride;
+        break;
+    }
+    src[pli] = aom_memalign(
+        32, sizeof(*src) * cm->mi_rows * cm->mi_cols * MI_SIZE * MI_SIZE);
+    ref_coeff[pli] = aom_memalign(
+        32, sizeof(*ref_coeff) * cm->mi_rows * cm->mi_cols * MI_SIZE * MI_SIZE);
+    xdec[pli] = xd->plane[pli].subsampling_x;
+    ydec[pli] = xd->plane[pli].subsampling_y;
+    bsize[pli] = ydec[pli] ? (xdec[pli] ? BLOCK_4X4 : BLOCK_8X4)
+                           : (xdec[pli] ? BLOCK_4X8 : BLOCK_8X8);
+    stride[pli] = cm->mi_cols << MI_SIZE_LOG2;
+    mi_wide_l2[pli] = MI_SIZE_LOG2 - xd->plane[pli].subsampling_x;
+    mi_high_l2[pli] = MI_SIZE_LOG2 - xd->plane[pli].subsampling_y;
+
+    const int frame_height =
+        (cm->mi_rows * MI_SIZE) >> xd->plane[pli].subsampling_y;
+    const int frame_width =
+        (cm->mi_cols * MI_SIZE) >> xd->plane[pli].subsampling_x;
+
+    for (r = 0; r < frame_height; ++r) {
+      for (c = 0; c < frame_width; ++c) {
+#if CONFIG_HIGHBITDEPTH
+        if (cm->use_highbitdepth) {
+          src[pli][r * stride[pli] + c] = CONVERT_TO_SHORTPTR(
+              xd->plane[pli].dst.buf)[r * xd->plane[pli].dst.stride + c];
+          ref_coeff[pli][r * stride[pli] + c] =
+              CONVERT_TO_SHORTPTR(ref_buffer)[r * ref_stride + c];
+        } else {
+#endif
+          src[pli][r * stride[pli] + c] =
+              xd->plane[pli].dst.buf[r * xd->plane[pli].dst.stride + c];
+          ref_coeff[pli][r * stride[pli] + c] = ref_buffer[r * ref_stride + c];
+#if CONFIG_HIGHBITDEPTH
+        }
+#endif
+      }
+    }
+  }
+  in = inbuf + OD_FILT_VBORDER * OD_FILT_BSTRIDE + OD_FILT_HBORDER;
+  sb_count = 0;
+  for (sbr = 0; sbr < nvsb; ++sbr) {
+    for (sbc = 0; sbc < nhsb; ++sbc) {
+      int nvb, nhb;
+      int gi;
+      int dirinit = 0;
+      nhb = AOMMIN(MAX_MIB_SIZE, cm->mi_cols - MAX_MIB_SIZE * sbc);
+      nvb = AOMMIN(MAX_MIB_SIZE, cm->mi_rows - MAX_MIB_SIZE * sbr);
+      cm->mi_grid_visible[MAX_MIB_SIZE * sbr * cm->mi_stride +
+                          MAX_MIB_SIZE * sbc]
+          ->mbmi.cdef_strength = -1;
+      if (sb_all_skip(cm, sbr * MAX_MIB_SIZE, sbc * MAX_MIB_SIZE)) continue;
+      dering_count = sb_compute_dering_list(cm, sbr * MAX_MIB_SIZE,
+                                            sbc * MAX_MIB_SIZE, dlist, 1);
+      for (pli = 0; pli < nplanes; pli++) {
+        for (i = 0; i < OD_DERING_INBUF_SIZE; i++)
+          inbuf[i] = OD_DERING_VERY_LARGE;
+        for (gi = 0; gi < TOTAL_STRENGTHS; gi++) {
+          int threshold;
+          uint64_t curr_mse;
+          int clpf_strength;
+          threshold = gi / CLPF_STRENGTHS;
+          if (pli > 0 && !chroma_dering) threshold = 0;
+          /* We avoid filtering the pixels for which some of the pixels to
+             average
+             are outside the frame. We could change the filter instead, but it
+             would add special cases for any future vectorization. */
+          int yoff = OD_FILT_VBORDER * (sbr != 0);
+          int xoff = OD_FILT_HBORDER * (sbc != 0);
+          int ysize = (nvb << mi_high_l2[pli]) +
+                      OD_FILT_VBORDER * (sbr != nvsb - 1) + yoff;
+          int xsize = (nhb << mi_wide_l2[pli]) +
+                      OD_FILT_HBORDER * (sbc != nhsb - 1) + xoff;
+          clpf_strength = gi % CLPF_STRENGTHS;
+          if (clpf_strength == 0)
+            copy_sb16_16(&in[(-yoff * OD_FILT_BSTRIDE - xoff)], OD_FILT_BSTRIDE,
+                         src[pli],
+                         (sbr * MAX_MIB_SIZE << mi_high_l2[pli]) - yoff,
+                         (sbc * MAX_MIB_SIZE << mi_wide_l2[pli]) - xoff,
+                         stride[pli], ysize, xsize);
+          od_dering(clpf_strength ? NULL : (uint8_t *)in, OD_FILT_BSTRIDE,
+                    tmp_dst, in, xdec[pli], ydec[pli], dir, &dirinit, var, pli,
+                    dlist, dering_count, threshold,
+                    clpf_strength + (clpf_strength == 3), clpf_damping,
+                    dering_damping, coeff_shift, clpf_strength != 0, 1);
+          curr_mse = compute_dering_dist(
+              ref_coeff[pli] +
+                  (sbr * MAX_MIB_SIZE << mi_high_l2[pli]) * stride[pli] +
+                  (sbc * MAX_MIB_SIZE << mi_wide_l2[pli]),
+              stride[pli], tmp_dst, dlist, dering_count, bsize[pli],
+              coeff_shift, pli);
+          if (pli < 2)
+            mse[pli][sb_count][gi] = curr_mse;
+          else
+            mse[1][sb_count][gi] += curr_mse;
+          sb_index[sb_count] =
+              MAX_MIB_SIZE * sbr * cm->mi_stride + MAX_MIB_SIZE * sbc;
+        }
+      }
+      sb_count++;
+    }
+  }
+  nb_strength_bits = 0;
+  /* Search for different number of signalling bits. */
+  for (i = 0; i <= 3; i++) {
+    int j;
+    int best_lev0[CDEF_MAX_STRENGTHS];
+    int best_lev1[CDEF_MAX_STRENGTHS] = { 0 };
+    nb_strengths = 1 << i;
+    if (nplanes >= 3)
+      tot_mse = joint_strength_search_dual(best_lev0, best_lev1, nb_strengths,
+                                           mse, sb_count);
+    else
+      tot_mse =
+          joint_strength_search(best_lev0, nb_strengths, mse[0], sb_count);
+    /* Count superblock signalling cost. */
+    tot_mse += (uint64_t)(sb_count * lambda * i);
+    /* Count header signalling cost. */
+    tot_mse += (uint64_t)(nb_strengths * lambda * CDEF_STRENGTH_BITS);
+    if (tot_mse < best_tot_mse) {
+      best_tot_mse = tot_mse;
+      nb_strength_bits = i;
+      for (j = 0; j < 1 << nb_strength_bits; j++) {
+        cm->cdef_strengths[j] = best_lev0[j];
+        cm->cdef_uv_strengths[j] = best_lev1[j];
+      }
+    }
+  }
+  nb_strengths = 1 << nb_strength_bits;
+
+  cm->cdef_bits = nb_strength_bits;
+  cm->nb_cdef_strengths = nb_strengths;
+  for (i = 0; i < sb_count; i++) {
+    int gi;
+    int best_gi;
+    uint64_t best_mse = (uint64_t)1 << 63;
+    best_gi = 0;
+    for (gi = 0; gi < cm->nb_cdef_strengths; gi++) {
+      uint64_t curr = mse[0][i][cm->cdef_strengths[gi]];
+      if (nplanes >= 3) curr += mse[1][i][cm->cdef_uv_strengths[gi]];
+      if (curr < best_mse) {
+        best_gi = gi;
+        best_mse = curr;
+      }
+    }
+    selected_strength[i] = best_gi;
+    cm->mi_grid_visible[sb_index[i]]->mbmi.cdef_strength = best_gi;
+  }
+  cm->cdef_dering_damping = dering_damping;
+  cm->cdef_clpf_damping = clpf_damping;
+  aom_free(mse[0]);
+  aom_free(mse[1]);
+  for (pli = 0; pli < nplanes; pli++) {
+    aom_free(src[pli]);
+    aom_free(ref_coeff[pli]);
+  }
+  aom_free(sb_index);
+  aom_free(selected_strength);
+}
diff --git a/third_party/aom/av1/encoder/picklpf.c b/third_party/aom/av1/encoder/picklpf.c
new file mode 100644
index 0000000000..fc0ea485d8
--- /dev/null
+++ b/third_party/aom/av1/encoder/picklpf.c
@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <limits.h>
+
+#include "./aom_scale_rtcd.h"
+
+#include "aom_dsp/psnr.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#include "av1/common/av1_loopfilter.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/quant_common.h"
+
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/picklpf.h"
+
+int av1_get_max_filter_level(const AV1_COMP *cpi) {
+  if (cpi->oxcf.pass == 2) {
+    return cpi->twopass.section_intra_rating > 8 ? MAX_LOOP_FILTER * 3 / 4
+                                                 : MAX_LOOP_FILTER;
+  } else {
+    return MAX_LOOP_FILTER;
+  }
+}
+
+static int64_t try_filter_frame(const YV12_BUFFER_CONFIG *sd,
+                                AV1_COMP *const cpi, int filt_level,
+                                int partial_frame) {
+  AV1_COMMON *const cm = &cpi->common;
+  int64_t filt_err;
+
+#if CONFIG_VAR_TX || CONFIG_EXT_PARTITION || CONFIG_CB4X4
+  av1_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filt_level, 1,
+                        partial_frame);
+#else
+  if (cpi->num_workers > 1)
+    av1_loop_filter_frame_mt(cm->frame_to_show, cm, cpi->td.mb.e_mbd.plane,
+                             filt_level, 1, partial_frame, cpi->workers,
+                             cpi->num_workers, &cpi->lf_row_sync);
+  else
+    av1_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filt_level,
+                          1, partial_frame);
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    filt_err = aom_highbd_get_y_sse(sd, cm->frame_to_show);
+  } else {
+    filt_err = aom_get_y_sse(sd, cm->frame_to_show);
+  }
+#else
+  filt_err = aom_get_y_sse(sd, cm->frame_to_show);
+#endif  // CONFIG_HIGHBITDEPTH
+
+  // Re-instate the unfiltered frame
+  aom_yv12_copy_y(&cpi->last_frame_uf, cm->frame_to_show);
+
+  return filt_err;
+}
+
+int av1_search_filter_level(const YV12_BUFFER_CONFIG *sd, AV1_COMP *cpi,
+                            int partial_frame, double *best_cost_ret) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const struct loopfilter *const lf = &cm->lf;
+  const int min_filter_level = 0;
+  const int max_filter_level = av1_get_max_filter_level(cpi);
+  int filt_direction = 0;
+  int64_t best_err;
+  int filt_best;
+  MACROBLOCK *x = &cpi->td.mb;
+
+  // Start the search at the previous frame filter level unless it is now out of
+  // range.
+  int filt_mid = clamp(lf->filter_level, min_filter_level, max_filter_level);
+  int filter_step = filt_mid < 16 ? 4 : filt_mid / 4;
+  // Sum squared error at each filter level
+  int64_t ss_err[MAX_LOOP_FILTER + 1];
+
+  // Set each entry to -1
+  memset(ss_err, 0xFF, sizeof(ss_err));
+
+  //  Make a copy of the unfiltered / processed recon buffer
+  aom_yv12_copy_y(cm->frame_to_show, &cpi->last_frame_uf);
+
+  best_err = try_filter_frame(sd, cpi, filt_mid, partial_frame);
+  filt_best = filt_mid;
+  ss_err[filt_mid] = best_err;
+
+  while (filter_step > 0) {
+    const int filt_high = AOMMIN(filt_mid + filter_step, max_filter_level);
+    const int filt_low = AOMMAX(filt_mid - filter_step, min_filter_level);
+
+    // Bias against raising loop filter in favor of lowering it.
+    int64_t bias = (best_err >> (15 - (filt_mid / 8))) * filter_step;
+
+    if ((cpi->oxcf.pass == 2) && (cpi->twopass.section_intra_rating < 20))
+      bias = (bias * cpi->twopass.section_intra_rating) / 20;
+
+    // yx, bias less for large block size
+    if (cm->tx_mode != ONLY_4X4) bias >>= 1;
+
+    if (filt_direction <= 0 && filt_low != filt_mid) {
+      // Get Low filter error score
+      if (ss_err[filt_low] < 0) {
+        ss_err[filt_low] = try_filter_frame(sd, cpi, filt_low, partial_frame);
+      }
+      // If value is close to the best so far then bias towards a lower loop
+      // filter value.
+      if (ss_err[filt_low] < (best_err + bias)) {
+        // Was it actually better than the previous best?
+        if (ss_err[filt_low] < best_err) {
+          best_err = ss_err[filt_low];
+        }
+        filt_best = filt_low;
+      }
+    }
+
+    // Now look at filt_high
+    if (filt_direction >= 0 && filt_high != filt_mid) {
+      if (ss_err[filt_high] < 0) {
+        ss_err[filt_high] = try_filter_frame(sd, cpi, filt_high, partial_frame);
+      }
+      // If value is significantly better than previous best, bias added against
+      // raising filter value
+      if (ss_err[filt_high] < (best_err - bias)) {
+        best_err = ss_err[filt_high];
+        filt_best = filt_high;
+      }
+    }
+
+    // Half the step distance if the best filter value was the same as last time
+    if (filt_best == filt_mid) {
+      filter_step /= 2;
+      filt_direction = 0;
+    } else {
+      filt_direction = (filt_best < filt_mid) ? -1 : 1;
+      filt_mid = filt_best;
+    }
+  }
+
+  // Update best error
+  best_err = ss_err[filt_best];
+
+  if (best_cost_ret)
+    *best_cost_ret = RDCOST_DBL(x->rdmult, x->rddiv, 0, best_err);
+  return filt_best;
+}
+
+void av1_pick_filter_level(const YV12_BUFFER_CONFIG *sd, AV1_COMP *cpi,
+                           LPF_PICK_METHOD method) {
+  AV1_COMMON *const cm = &cpi->common;
+  struct loopfilter *const lf = &cm->lf;
+
+  lf->sharpness_level = cm->frame_type == KEY_FRAME ? 0 : cpi->oxcf.sharpness;
+
+  if (method == LPF_PICK_MINIMAL_LPF && lf->filter_level) {
+    lf->filter_level = 0;
+  } else if (method >= LPF_PICK_FROM_Q) {
+    const int min_filter_level = 0;
+    const int max_filter_level = av1_get_max_filter_level(cpi);
+    const int q = av1_ac_quant(cm->base_qindex, 0, cm->bit_depth);
+// These values were determined by linear fitting the result of the
+// searched level, filt_guess = q * 0.316206 + 3.87252
+#if CONFIG_HIGHBITDEPTH
+    int filt_guess;
+    switch (cm->bit_depth) {
+      case AOM_BITS_8:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 1015158, 18);
+        break;
+      case AOM_BITS_10:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 4060632, 20);
+        break;
+      case AOM_BITS_12:
+        filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 16242526, 22);
+        break;
+      default:
+        assert(0 &&
+               "bit_depth should be AOM_BITS_8, AOM_BITS_10 "
+               "or AOM_BITS_12");
+        return;
+    }
+#else
+    int filt_guess = ROUND_POWER_OF_TWO(q * 20723 + 1015158, 18);
+#endif  // CONFIG_HIGHBITDEPTH
+    if (cm->frame_type == KEY_FRAME) filt_guess -= 4;
+    lf->filter_level = clamp(filt_guess, min_filter_level, max_filter_level);
+  } else {
+    lf->filter_level = av1_search_filter_level(
+        sd, cpi, method == LPF_PICK_FROM_SUBIMAGE, NULL);
+  }
+
+#if CONFIG_EXT_TILE
+  // TODO(any): 0 loopfilter level is only necessary if individual tile
+  // decoding is required. We need to communicate this requirement to this
+  // code and force loop filter level 0 only if required.
+  if (cm->tile_encoding_mode) lf->filter_level = 0;
+#endif  // CONFIG_EXT_TILE
+}
diff --git a/third_party/aom/av1/encoder/picklpf.h b/third_party/aom/av1/encoder/picklpf.h
new file mode 100644
index 0000000000..3c0a83462b
--- /dev/null
+++ b/third_party/aom/av1/encoder/picklpf.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_PICKLPF_H_
+#define AV1_ENCODER_PICKLPF_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "av1/encoder/encoder.h"
+
+struct yv12_buffer_config;
+struct AV1_COMP;
+int av1_get_max_filter_level(const AV1_COMP *cpi);
+int av1_search_filter_level(const YV12_BUFFER_CONFIG *sd, AV1_COMP *cpi,
+                            int partial_frame, double *err);
+void av1_pick_filter_level(const struct yv12_buffer_config *sd,
+                           struct AV1_COMP *cpi, LPF_PICK_METHOD method);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_PICKLPF_H_
diff --git a/third_party/aom/av1/encoder/pickrst.c b/third_party/aom/av1/encoder/pickrst.c
new file mode 100644
index 0000000000..21410e0afa
--- /dev/null
+++ b/third_party/aom/av1/encoder/pickrst.c
@@ -0,0 +1,1269 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <float.h>
+#include <limits.h>
+#include <math.h>
+
+#include "./aom_scale_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/binary_codes_writer.h"
+#include "aom_dsp/psnr.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/system_state.h"
+
+#include "av1/common/onyxc_int.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/restoration.h"
+
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/picklpf.h"
+#include "av1/encoder/pickrst.h"
+
+// When set to RESTORE_WIENER or RESTORE_SGRPROJ only those are allowed.
+// When set to RESTORE_NONE (0) we allow switchable.
+const RestorationType force_restore_type = RESTORE_NONE;
+
+// Number of Wiener iterations
+#define NUM_WIENER_ITERS 10
+
+typedef double (*search_restore_type)(const YV12_BUFFER_CONFIG *src,
+                                      AV1_COMP *cpi, int partial_frame,
+                                      RestorationInfo *info,
+                                      RestorationType *rest_level,
+                                      double *best_tile_cost,
+                                      YV12_BUFFER_CONFIG *dst_frame);
+
+const int frame_level_restore_bits[RESTORE_TYPES] = { 2, 2, 2, 2 };
+
+static int64_t sse_restoration_tile(const YV12_BUFFER_CONFIG *src,
+                                    const YV12_BUFFER_CONFIG *dst,
+                                    const AV1_COMMON *cm, int h_start,
+                                    int width, int v_start, int height,
+                                    int components_pattern) {
+  int64_t filt_err = 0;
+  (void)cm;
+  // Y and UV components cannot be mixed
+  assert(components_pattern == 1 || components_pattern == 2 ||
+         components_pattern == 4 || components_pattern == 6);
+#if CONFIG_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    if ((components_pattern >> AOM_PLANE_Y) & 1) {
+      filt_err +=
+          aom_highbd_get_y_sse_part(src, dst, h_start, width, v_start, height);
+    }
+    if ((components_pattern >> AOM_PLANE_U) & 1) {
+      filt_err +=
+          aom_highbd_get_u_sse_part(src, dst, h_start, width, v_start, height);
+    }
+    if ((components_pattern >> AOM_PLANE_V) & 1) {
+      filt_err +=
+          aom_highbd_get_v_sse_part(src, dst, h_start, width, v_start, height);
+    }
+    return filt_err;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  if ((components_pattern >> AOM_PLANE_Y) & 1) {
+    filt_err += aom_get_y_sse_part(src, dst, h_start, width, v_start, height);
+  }
+  if ((components_pattern >> AOM_PLANE_U) & 1) {
+    filt_err += aom_get_u_sse_part(src, dst, h_start, width, v_start, height);
+  }
+  if ((components_pattern >> AOM_PLANE_V) & 1) {
+    filt_err += aom_get_v_sse_part(src, dst, h_start, width, v_start, height);
+  }
+  return filt_err;
+}
+
+static int64_t sse_restoration_frame(AV1_COMMON *const cm,
+                                     const YV12_BUFFER_CONFIG *src,
+                                     const YV12_BUFFER_CONFIG *dst,
+                                     int components_pattern) {
+  int64_t filt_err = 0;
+#if CONFIG_HIGHBITDEPTH
+  if (cm->use_highbitdepth) {
+    if ((components_pattern >> AOM_PLANE_Y) & 1) {
+      filt_err += aom_highbd_get_y_sse(src, dst);
+    }
+    if ((components_pattern >> AOM_PLANE_U) & 1) {
+      filt_err += aom_highbd_get_u_sse(src, dst);
+    }
+    if ((components_pattern >> AOM_PLANE_V) & 1) {
+      filt_err += aom_highbd_get_v_sse(src, dst);
+    }
+    return filt_err;
+  }
+#else
+  (void)cm;
+#endif  // CONFIG_HIGHBITDEPTH
+  if ((components_pattern >> AOM_PLANE_Y) & 1) {
+    filt_err = aom_get_y_sse(src, dst);
+  }
+  if ((components_pattern >> AOM_PLANE_U) & 1) {
+    filt_err += aom_get_u_sse(src, dst);
+  }
+  if ((components_pattern >> AOM_PLANE_V) & 1) {
+    filt_err += aom_get_v_sse(src, dst);
+  }
+  return filt_err;
+}
+
+static int64_t try_restoration_tile(const YV12_BUFFER_CONFIG *src,
+                                    AV1_COMP *const cpi, RestorationInfo *rsi,
+                                    int components_pattern, int partial_frame,
+                                    int tile_idx, int subtile_idx,
+                                    int subtile_bits,
+                                    YV12_BUFFER_CONFIG *dst_frame) {
+  AV1_COMMON *const cm = &cpi->common;
+  int64_t filt_err;
+  int tile_width, tile_height, nhtiles, nvtiles;
+  int h_start, h_end, v_start, v_end;
+  int ntiles, width, height;
+
+  // Y and UV components cannot be mixed
+  assert(components_pattern == 1 || components_pattern == 2 ||
+         components_pattern == 4 || components_pattern == 6);
+
+  if (components_pattern == 1) {  // Y only
+    width = src->y_crop_width;
+    height = src->y_crop_height;
+  } else {  // Color
+    width = src->uv_crop_width;
+    height = src->uv_crop_height;
+  }
+  ntiles = av1_get_rest_ntiles(
+      width, height, cm->rst_info[components_pattern > 1].restoration_tilesize,
+      &tile_width, &tile_height, &nhtiles, &nvtiles);
+  (void)ntiles;
+
+  av1_loop_restoration_frame(cm->frame_to_show, cm, rsi, components_pattern,
+                             partial_frame, dst_frame);
+  av1_get_rest_tile_limits(tile_idx, subtile_idx, subtile_bits, nhtiles,
+                           nvtiles, tile_width, tile_height, width, height, 0,
+                           0, &h_start, &h_end, &v_start, &v_end);
+  filt_err = sse_restoration_tile(src, dst_frame, cm, h_start, h_end - h_start,
+                                  v_start, v_end - v_start, components_pattern);
+
+  return filt_err;
+}
+
+static int64_t try_restoration_frame(const YV12_BUFFER_CONFIG *src,
+                                     AV1_COMP *const cpi, RestorationInfo *rsi,
+                                     int components_pattern, int partial_frame,
+                                     YV12_BUFFER_CONFIG *dst_frame) {
+  AV1_COMMON *const cm = &cpi->common;
+  int64_t filt_err;
+  av1_loop_restoration_frame(cm->frame_to_show, cm, rsi, components_pattern,
+                             partial_frame, dst_frame);
+  filt_err = sse_restoration_frame(cm, src, dst_frame, components_pattern);
+  return filt_err;
+}
+
+static int64_t get_pixel_proj_error(uint8_t *src8, int width, int height,
+                                    int src_stride, uint8_t *dat8,
+                                    int dat_stride, int bit_depth,
+                                    int32_t *flt1, int flt1_stride,
+                                    int32_t *flt2, int flt2_stride, int *xqd) {
+  int i, j;
+  int64_t err = 0;
+  int xq[2];
+  decode_xq(xqd, xq);
+  if (bit_depth == 8) {
+    const uint8_t *src = src8;
+    const uint8_t *dat = dat8;
+    for (i = 0; i < height; ++i) {
+      for (j = 0; j < width; ++j) {
+        const int32_t u =
+            (int32_t)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS);
+        const int32_t f1 = (int32_t)flt1[i * flt1_stride + j] - u;
+        const int32_t f2 = (int32_t)flt2[i * flt2_stride + j] - u;
+        const int32_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
+        const int32_t e =
+            ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) -
+            src[i * src_stride + j];
+        err += e * e;
+      }
+    }
+  } else {
+    const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+    const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
+    for (i = 0; i < height; ++i) {
+      for (j = 0; j < width; ++j) {
+        const int32_t u =
+            (int32_t)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS);
+        const int32_t f1 = (int32_t)flt1[i * flt1_stride + j] - u;
+        const int32_t f2 = (int32_t)flt2[i * flt2_stride + j] - u;
+        const int32_t v = xq[0] * f1 + xq[1] * f2 + (u << SGRPROJ_PRJ_BITS);
+        const int32_t e =
+            ROUND_POWER_OF_TWO(v, SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) -
+            src[i * src_stride + j];
+        err += e * e;
+      }
+    }
+  }
+  return err;
+}
+
+static void get_proj_subspace(uint8_t *src8, int width, int height,
+                              int src_stride, uint8_t *dat8, int dat_stride,
+                              int bit_depth, int32_t *flt1, int flt1_stride,
+                              int32_t *flt2, int flt2_stride, int *xq) {
+  int i, j;
+  double H[2][2] = { { 0, 0 }, { 0, 0 } };
+  double C[2] = { 0, 0 };
+  double Det;
+  double x[2];
+  const int size = width * height;
+
+  aom_clear_system_state();
+
+  // Default
+  xq[0] = 0;
+  xq[1] = 0;
+  if (bit_depth == 8) {
+    const uint8_t *src = src8;
+    const uint8_t *dat = dat8;
+    for (i = 0; i < height; ++i) {
+      for (j = 0; j < width; ++j) {
+        const double u = (double)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS);
+        const double s =
+            (double)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u;
+        const double f1 = (double)flt1[i * flt1_stride + j] - u;
+        const double f2 = (double)flt2[i * flt2_stride + j] - u;
+        H[0][0] += f1 * f1;
+        H[1][1] += f2 * f2;
+        H[0][1] += f1 * f2;
+        C[0] += f1 * s;
+        C[1] += f2 * s;
+      }
+    }
+  } else {
+    const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+    const uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
+    for (i = 0; i < height; ++i) {
+      for (j = 0; j < width; ++j) {
+        const double u = (double)(dat[i * dat_stride + j] << SGRPROJ_RST_BITS);
+        const double s =
+            (double)(src[i * src_stride + j] << SGRPROJ_RST_BITS) - u;
+        const double f1 = (double)flt1[i * flt1_stride + j] - u;
+        const double f2 = (double)flt2[i * flt2_stride + j] - u;
+        H[0][0] += f1 * f1;
+        H[1][1] += f2 * f2;
+        H[0][1] += f1 * f2;
+        C[0] += f1 * s;
+        C[1] += f2 * s;
+      }
+    }
+  }
+  H[0][0] /= size;
+  H[0][1] /= size;
+  H[1][1] /= size;
+  H[1][0] = H[0][1];
+  C[0] /= size;
+  C[1] /= size;
+  Det = (H[0][0] * H[1][1] - H[0][1] * H[1][0]);
+  if (Det < 1e-8) return;  // ill-posed, return default values
+  x[0] = (H[1][1] * C[0] - H[0][1] * C[1]) / Det;
+  x[1] = (H[0][0] * C[1] - H[1][0] * C[0]) / Det;
+  xq[0] = (int)rint(x[0] * (1 << SGRPROJ_PRJ_BITS));
+  xq[1] = (int)rint(x[1] * (1 << SGRPROJ_PRJ_BITS));
+}
+
+void encode_xq(int *xq, int *xqd) {
+  xqd[0] = xq[0];
+  xqd[0] = clamp(xqd[0], SGRPROJ_PRJ_MIN0, SGRPROJ_PRJ_MAX0);
+  xqd[1] = (1 << SGRPROJ_PRJ_BITS) - xqd[0] - xq[1];
+  xqd[1] = clamp(xqd[1], SGRPROJ_PRJ_MIN1, SGRPROJ_PRJ_MAX1);
+}
+
+static void search_selfguided_restoration(uint8_t *dat8, int width, int height,
+                                          int dat_stride, uint8_t *src8,
+                                          int src_stride, int bit_depth,
+                                          int *eps, int *xqd, int32_t *rstbuf) {
+  int32_t *flt1 = rstbuf;
+  int32_t *flt2 = flt1 + RESTORATION_TILEPELS_MAX;
+  int32_t *tmpbuf2 = flt2 + RESTORATION_TILEPELS_MAX;
+  int ep, bestep = 0;
+  int64_t err, besterr = -1;
+  int exqd[2], bestxqd[2] = { 0, 0 };
+
+  for (ep = 0; ep < SGRPROJ_PARAMS; ep++) {
+    int exq[2];
+#if CONFIG_HIGHBITDEPTH
+    if (bit_depth > 8) {
+      uint16_t *dat = CONVERT_TO_SHORTPTR(dat8);
+#if USE_HIGHPASS_IN_SGRPROJ
+      av1_highpass_filter_highbd(dat, width, height, dat_stride, flt1, width,
+                                 sgr_params[ep].corner, sgr_params[ep].edge);
+#else
+      av1_selfguided_restoration_highbd(dat, width, height, dat_stride, flt1,
+                                        width, bit_depth, sgr_params[ep].r1,
+                                        sgr_params[ep].e1, tmpbuf2);
+#endif  // USE_HIGHPASS_IN_SGRPROJ
+      av1_selfguided_restoration_highbd(dat, width, height, dat_stride, flt2,
+                                        width, bit_depth, sgr_params[ep].r2,
+                                        sgr_params[ep].e2, tmpbuf2);
+    } else {
+#endif
+#if USE_HIGHPASS_IN_SGRPROJ
+      av1_highpass_filter(dat8, width, height, dat_stride, flt1, width,
+                          sgr_params[ep].corner, sgr_params[ep].edge);
+#else
+    av1_selfguided_restoration(dat8, width, height, dat_stride, flt1, width,
+                               sgr_params[ep].r1, sgr_params[ep].e1, tmpbuf2);
+#endif  // USE_HIGHPASS_IN_SGRPROJ
+      av1_selfguided_restoration(dat8, width, height, dat_stride, flt2, width,
+                                 sgr_params[ep].r2, sgr_params[ep].e2, tmpbuf2);
+#if CONFIG_HIGHBITDEPTH
+    }
+#endif
+    get_proj_subspace(src8, width, height, src_stride, dat8, dat_stride,
+                      bit_depth, flt1, width, flt2, width, exq);
+    encode_xq(exq, exqd);
+    err =
+        get_pixel_proj_error(src8, width, height, src_stride, dat8, dat_stride,
+                             bit_depth, flt1, width, flt2, width, exqd);
+    if (besterr == -1 || err < besterr) {
+      bestep = ep;
+      besterr = err;
+      bestxqd[0] = exqd[0];
+      bestxqd[1] = exqd[1];
+    }
+  }
+  *eps = bestep;
+  xqd[0] = bestxqd[0];
+  xqd[1] = bestxqd[1];
+}
+
+static int count_sgrproj_bits(SgrprojInfo *sgrproj_info,
+                              SgrprojInfo *ref_sgrproj_info) {
+  int bits = SGRPROJ_PARAMS_BITS;
+  bits += aom_count_primitive_refsubexpfin(
+      SGRPROJ_PRJ_MAX0 - SGRPROJ_PRJ_MIN0 + 1, SGRPROJ_PRJ_SUBEXP_K,
+      ref_sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0,
+      sgrproj_info->xqd[0] - SGRPROJ_PRJ_MIN0);
+  bits += aom_count_primitive_refsubexpfin(
+      SGRPROJ_PRJ_MAX1 - SGRPROJ_PRJ_MIN1 + 1, SGRPROJ_PRJ_SUBEXP_K,
+      ref_sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1,
+      sgrproj_info->xqd[1] - SGRPROJ_PRJ_MIN1);
+  return bits;
+}
+
+static double search_sgrproj(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
+                             int partial_frame, RestorationInfo *info,
+                             RestorationType *type, double *best_tile_cost,
+                             YV12_BUFFER_CONFIG *dst_frame) {
+  SgrprojInfo *sgrproj_info = info->sgrproj_info;
+  double err, cost_norestore, cost_sgrproj;
+  int bits;
+  MACROBLOCK *x = &cpi->td.mb;
+  AV1_COMMON *const cm = &cpi->common;
+  const YV12_BUFFER_CONFIG *dgd = cm->frame_to_show;
+  RestorationInfo *rsi = &cpi->rst_search[0];
+  int tile_idx, tile_width, tile_height, nhtiles, nvtiles;
+  int h_start, h_end, v_start, v_end;
+  // Allocate for the src buffer at high precision
+  const int ntiles = av1_get_rest_ntiles(
+      cm->width, cm->height, cm->rst_info[0].restoration_tilesize, &tile_width,
+      &tile_height, &nhtiles, &nvtiles);
+  SgrprojInfo ref_sgrproj_info;
+  set_default_sgrproj(&ref_sgrproj_info);
+
+  rsi->frame_restoration_type = RESTORE_SGRPROJ;
+
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    rsi->restoration_type[tile_idx] = RESTORE_NONE;
+  }
+  // Compute best Sgrproj filters for each tile
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
+                             tile_height, cm->width, cm->height, 0, 0, &h_start,
+                             &h_end, &v_start, &v_end);
+    err = sse_restoration_tile(src, cm->frame_to_show, cm, h_start,
+                               h_end - h_start, v_start, v_end - v_start, 1);
+    // #bits when a tile is not restored
+    bits = av1_cost_bit(RESTORE_NONE_SGRPROJ_PROB, 0);
+    cost_norestore = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+    best_tile_cost[tile_idx] = DBL_MAX;
+    search_selfguided_restoration(
+        dgd->y_buffer + v_start * dgd->y_stride + h_start, h_end - h_start,
+        v_end - v_start, dgd->y_stride,
+        src->y_buffer + v_start * src->y_stride + h_start, src->y_stride,
+#if CONFIG_HIGHBITDEPTH
+        cm->bit_depth,
+#else
+        8,
+#endif  // CONFIG_HIGHBITDEPTH
+        &rsi->sgrproj_info[tile_idx].ep, rsi->sgrproj_info[tile_idx].xqd,
+        cm->rst_internal.tmpbuf);
+    rsi->restoration_type[tile_idx] = RESTORE_SGRPROJ;
+    err = try_restoration_tile(src, cpi, rsi, 1, partial_frame, tile_idx, 0, 0,
+                               dst_frame);
+    bits = count_sgrproj_bits(&rsi->sgrproj_info[tile_idx], &ref_sgrproj_info)
+           << AV1_PROB_COST_SHIFT;
+    bits += av1_cost_bit(RESTORE_NONE_SGRPROJ_PROB, 1);
+    cost_sgrproj = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+    if (cost_sgrproj >= cost_norestore) {
+      type[tile_idx] = RESTORE_NONE;
+    } else {
+      type[tile_idx] = RESTORE_SGRPROJ;
+      memcpy(&sgrproj_info[tile_idx], &rsi->sgrproj_info[tile_idx],
+             sizeof(sgrproj_info[tile_idx]));
+      bits = count_sgrproj_bits(&rsi->sgrproj_info[tile_idx], &ref_sgrproj_info)
+             << AV1_PROB_COST_SHIFT;
+      memcpy(&ref_sgrproj_info, &sgrproj_info[tile_idx],
+             sizeof(ref_sgrproj_info));
+      best_tile_cost[tile_idx] = err;
+    }
+    rsi->restoration_type[tile_idx] = RESTORE_NONE;
+  }
+  // Cost for Sgrproj filtering
+  set_default_sgrproj(&ref_sgrproj_info);
+  bits = frame_level_restore_bits[rsi->frame_restoration_type]
+         << AV1_PROB_COST_SHIFT;
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    bits +=
+        av1_cost_bit(RESTORE_NONE_SGRPROJ_PROB, type[tile_idx] != RESTORE_NONE);
+    memcpy(&rsi->sgrproj_info[tile_idx], &sgrproj_info[tile_idx],
+           sizeof(sgrproj_info[tile_idx]));
+    if (type[tile_idx] == RESTORE_SGRPROJ) {
+      bits +=
+          count_sgrproj_bits(&rsi->sgrproj_info[tile_idx], &ref_sgrproj_info)
+          << AV1_PROB_COST_SHIFT;
+      memcpy(&ref_sgrproj_info, &rsi->sgrproj_info[tile_idx],
+             sizeof(ref_sgrproj_info));
+    }
+    rsi->restoration_type[tile_idx] = type[tile_idx];
+  }
+  err = try_restoration_frame(src, cpi, rsi, 1, partial_frame, dst_frame);
+  cost_sgrproj = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+
+  return cost_sgrproj;
+}
+
+static double find_average(uint8_t *src, int h_start, int h_end, int v_start,
+                           int v_end, int stride) {
+  uint64_t sum = 0;
+  double avg = 0;
+  int i, j;
+  aom_clear_system_state();
+  for (i = v_start; i < v_end; i++)
+    for (j = h_start; j < h_end; j++) sum += src[i * stride + j];
+  avg = (double)sum / ((v_end - v_start) * (h_end - h_start));
+  return avg;
+}
+
+static void compute_stats(uint8_t *dgd, uint8_t *src, int h_start, int h_end,
+                          int v_start, int v_end, int dgd_stride,
+                          int src_stride, double *M, double *H) {
+  int i, j, k, l;
+  double Y[WIENER_WIN2];
+  const double avg =
+      find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
+
+  memset(M, 0, sizeof(*M) * WIENER_WIN2);
+  memset(H, 0, sizeof(*H) * WIENER_WIN2 * WIENER_WIN2);
+  for (i = v_start; i < v_end; i++) {
+    for (j = h_start; j < h_end; j++) {
+      const double X = (double)src[i * src_stride + j] - avg;
+      int idx = 0;
+      for (k = -WIENER_HALFWIN; k <= WIENER_HALFWIN; k++) {
+        for (l = -WIENER_HALFWIN; l <= WIENER_HALFWIN; l++) {
+          Y[idx] = (double)dgd[(i + l) * dgd_stride + (j + k)] - avg;
+          idx++;
+        }
+      }
+      for (k = 0; k < WIENER_WIN2; ++k) {
+        M[k] += Y[k] * X;
+        H[k * WIENER_WIN2 + k] += Y[k] * Y[k];
+        for (l = k + 1; l < WIENER_WIN2; ++l) {
+          // H is a symmetric matrix, so we only need to fill out the upper
+          // triangle here. We can copy it down to the lower triangle outside
+          // the (i, j) loops.
+          H[k * WIENER_WIN2 + l] += Y[k] * Y[l];
+        }
+      }
+    }
+  }
+  for (k = 0; k < WIENER_WIN2; ++k) {
+    for (l = k + 1; l < WIENER_WIN2; ++l) {
+      H[l * WIENER_WIN2 + k] = H[k * WIENER_WIN2 + l];
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static double find_average_highbd(uint16_t *src, int h_start, int h_end,
+                                  int v_start, int v_end, int stride) {
+  uint64_t sum = 0;
+  double avg = 0;
+  int i, j;
+  aom_clear_system_state();
+  for (i = v_start; i < v_end; i++)
+    for (j = h_start; j < h_end; j++) sum += src[i * stride + j];
+  avg = (double)sum / ((v_end - v_start) * (h_end - h_start));
+  return avg;
+}
+
+static void compute_stats_highbd(uint8_t *dgd8, uint8_t *src8, int h_start,
+                                 int h_end, int v_start, int v_end,
+                                 int dgd_stride, int src_stride, double *M,
+                                 double *H) {
+  int i, j, k, l;
+  double Y[WIENER_WIN2];
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dgd = CONVERT_TO_SHORTPTR(dgd8);
+  const double avg =
+      find_average_highbd(dgd, h_start, h_end, v_start, v_end, dgd_stride);
+
+  memset(M, 0, sizeof(*M) * WIENER_WIN2);
+  memset(H, 0, sizeof(*H) * WIENER_WIN2 * WIENER_WIN2);
+  for (i = v_start; i < v_end; i++) {
+    for (j = h_start; j < h_end; j++) {
+      const double X = (double)src[i * src_stride + j] - avg;
+      int idx = 0;
+      for (k = -WIENER_HALFWIN; k <= WIENER_HALFWIN; k++) {
+        for (l = -WIENER_HALFWIN; l <= WIENER_HALFWIN; l++) {
+          Y[idx] = (double)dgd[(i + l) * dgd_stride + (j + k)] - avg;
+          idx++;
+        }
+      }
+      for (k = 0; k < WIENER_WIN2; ++k) {
+        M[k] += Y[k] * X;
+        H[k * WIENER_WIN2 + k] += Y[k] * Y[k];
+        for (l = k + 1; l < WIENER_WIN2; ++l) {
+          // H is a symmetric matrix, so we only need to fill out the upper
+          // triangle here. We can copy it down to the lower triangle outside
+          // the (i, j) loops.
+          H[k * WIENER_WIN2 + l] += Y[k] * Y[l];
+        }
+      }
+    }
+  }
+  for (k = 0; k < WIENER_WIN2; ++k) {
+    for (l = k + 1; l < WIENER_WIN2; ++l) {
+      H[l * WIENER_WIN2 + k] = H[k * WIENER_WIN2 + l];
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+// Solves Ax = b, where x and b are column vectors
+static int linsolve(int n, double *A, int stride, double *b, double *x) {
+  int i, j, k;
+  double c;
+
+  aom_clear_system_state();
+
+  // Forward elimination
+  for (k = 0; k < n - 1; k++) {
+    // Bring the largest magitude to the diagonal position
+    for (i = n - 1; i > k; i--) {
+      if (fabs(A[(i - 1) * stride + k]) < fabs(A[i * stride + k])) {
+        for (j = 0; j < n; j++) {
+          c = A[i * stride + j];
+          A[i * stride + j] = A[(i - 1) * stride + j];
+          A[(i - 1) * stride + j] = c;
+        }
+        c = b[i];
+        b[i] = b[i - 1];
+        b[i - 1] = c;
+      }
+    }
+    for (i = k; i < n - 1; i++) {
+      if (fabs(A[k * stride + k]) < 1e-10) return 0;
+      c = A[(i + 1) * stride + k] / A[k * stride + k];
+      for (j = 0; j < n; j++) A[(i + 1) * stride + j] -= c * A[k * stride + j];
+      b[i + 1] -= c * b[k];
+    }
+  }
+  // Backward substitution
+  for (i = n - 1; i >= 0; i--) {
+    if (fabs(A[i * stride + i]) < 1e-10) return 0;
+    c = 0;
+    for (j = i + 1; j <= n - 1; j++) c += A[i * stride + j] * x[j];
+    x[i] = (b[i] - c) / A[i * stride + i];
+  }
+
+  return 1;
+}
+
+static INLINE int wrap_index(int i) {
+  return (i >= WIENER_HALFWIN1 ? WIENER_WIN - 1 - i : i);
+}
+
+// Fix vector b, update vector a
+static void update_a_sep_sym(double **Mc, double **Hc, double *a, double *b) {
+  int i, j;
+  double S[WIENER_WIN];
+  double A[WIENER_HALFWIN1], B[WIENER_HALFWIN1 * WIENER_HALFWIN1];
+  int w, w2;
+  memset(A, 0, sizeof(A));
+  memset(B, 0, sizeof(B));
+  for (i = 0; i < WIENER_WIN; i++) {
+    for (j = 0; j < WIENER_WIN; ++j) {
+      const int jj = wrap_index(j);
+      A[jj] += Mc[i][j] * b[i];
+    }
+  }
+  for (i = 0; i < WIENER_WIN; i++) {
+    for (j = 0; j < WIENER_WIN; j++) {
+      int k, l;
+      for (k = 0; k < WIENER_WIN; ++k)
+        for (l = 0; l < WIENER_WIN; ++l) {
+          const int kk = wrap_index(k);
+          const int ll = wrap_index(l);
+          B[ll * WIENER_HALFWIN1 + kk] +=
+              Hc[j * WIENER_WIN + i][k * WIENER_WIN2 + l] * b[i] * b[j];
+        }
+    }
+  }
+  // Normalization enforcement in the system of equations itself
+  w = WIENER_WIN;
+  w2 = (w >> 1) + 1;
+  for (i = 0; i < w2 - 1; ++i)
+    A[i] -=
+        A[w2 - 1] * 2 + B[i * w2 + w2 - 1] - 2 * B[(w2 - 1) * w2 + (w2 - 1)];
+  for (i = 0; i < w2 - 1; ++i)
+    for (j = 0; j < w2 - 1; ++j)
+      B[i * w2 + j] -= 2 * (B[i * w2 + (w2 - 1)] + B[(w2 - 1) * w2 + j] -
+                            2 * B[(w2 - 1) * w2 + (w2 - 1)]);
+  if (linsolve(w2 - 1, B, w2, A, S)) {
+    S[w2 - 1] = 1.0;
+    for (i = w2; i < w; ++i) {
+      S[i] = S[w - 1 - i];
+      S[w2 - 1] -= 2 * S[i];
+    }
+    memcpy(a, S, w * sizeof(*a));
+  }
+}
+
+// Fix vector a, update vector b
+static void update_b_sep_sym(double **Mc, double **Hc, double *a, double *b) {
+  int i, j;
+  double S[WIENER_WIN];
+  double A[WIENER_HALFWIN1], B[WIENER_HALFWIN1 * WIENER_HALFWIN1];
+  int w, w2;
+  memset(A, 0, sizeof(A));
+  memset(B, 0, sizeof(B));
+  for (i = 0; i < WIENER_WIN; i++) {
+    const int ii = wrap_index(i);
+    for (j = 0; j < WIENER_WIN; j++) A[ii] += Mc[i][j] * a[j];
+  }
+
+  for (i = 0; i < WIENER_WIN; i++) {
+    for (j = 0; j < WIENER_WIN; j++) {
+      const int ii = wrap_index(i);
+      const int jj = wrap_index(j);
+      int k, l;
+      for (k = 0; k < WIENER_WIN; ++k)
+        for (l = 0; l < WIENER_WIN; ++l)
+          B[jj * WIENER_HALFWIN1 + ii] +=
+              Hc[i * WIENER_WIN + j][k * WIENER_WIN2 + l] * a[k] * a[l];
+    }
+  }
+  // Normalization enforcement in the system of equations itself
+  w = WIENER_WIN;
+  w2 = WIENER_HALFWIN1;
+  for (i = 0; i < w2 - 1; ++i)
+    A[i] -=
+        A[w2 - 1] * 2 + B[i * w2 + w2 - 1] - 2 * B[(w2 - 1) * w2 + (w2 - 1)];
+  for (i = 0; i < w2 - 1; ++i)
+    for (j = 0; j < w2 - 1; ++j)
+      B[i * w2 + j] -= 2 * (B[i * w2 + (w2 - 1)] + B[(w2 - 1) * w2 + j] -
+                            2 * B[(w2 - 1) * w2 + (w2 - 1)]);
+  if (linsolve(w2 - 1, B, w2, A, S)) {
+    S[w2 - 1] = 1.0;
+    for (i = w2; i < w; ++i) {
+      S[i] = S[w - 1 - i];
+      S[w2 - 1] -= 2 * S[i];
+    }
+    memcpy(b, S, w * sizeof(*b));
+  }
+}
+
+static int wiener_decompose_sep_sym(double *M, double *H, double *a,
+                                    double *b) {
+  static const double init_filt[WIENER_WIN] = {
+    0.035623, -0.127154, 0.211436, 0.760190, 0.211436, -0.127154, 0.035623,
+  };
+  int i, j, iter;
+  double *Hc[WIENER_WIN2];
+  double *Mc[WIENER_WIN];
+  for (i = 0; i < WIENER_WIN; i++) {
+    Mc[i] = M + i * WIENER_WIN;
+    for (j = 0; j < WIENER_WIN; j++) {
+      Hc[i * WIENER_WIN + j] =
+          H + i * WIENER_WIN * WIENER_WIN2 + j * WIENER_WIN;
+    }
+  }
+  memcpy(a, init_filt, sizeof(*a) * WIENER_WIN);
+  memcpy(b, init_filt, sizeof(*b) * WIENER_WIN);
+
+  iter = 1;
+  while (iter < NUM_WIENER_ITERS) {
+    update_a_sep_sym(Mc, Hc, a, b);
+    update_b_sep_sym(Mc, Hc, a, b);
+    iter++;
+  }
+  return 1;
+}
+
+// Computes the function x'*H*x - x'*M for the learned 2D filter x, and compares
+// against identity filters; Final score is defined as the difference between
+// the function values
+static double compute_score(double *M, double *H, InterpKernel vfilt,
+                            InterpKernel hfilt) {
+  double ab[WIENER_WIN * WIENER_WIN];
+  int i, k, l;
+  double P = 0, Q = 0;
+  double iP = 0, iQ = 0;
+  double Score, iScore;
+  double a[WIENER_WIN], b[WIENER_WIN];
+
+  aom_clear_system_state();
+
+  a[WIENER_HALFWIN] = b[WIENER_HALFWIN] = 1.0;
+  for (i = 0; i < WIENER_HALFWIN; ++i) {
+    a[i] = a[WIENER_WIN - i - 1] = (double)vfilt[i] / WIENER_FILT_STEP;
+    b[i] = b[WIENER_WIN - i - 1] = (double)hfilt[i] / WIENER_FILT_STEP;
+    a[WIENER_HALFWIN] -= 2 * a[i];
+    b[WIENER_HALFWIN] -= 2 * b[i];
+  }
+  for (k = 0; k < WIENER_WIN; ++k) {
+    for (l = 0; l < WIENER_WIN; ++l) ab[k * WIENER_WIN + l] = a[l] * b[k];
+  }
+  for (k = 0; k < WIENER_WIN2; ++k) {
+    P += ab[k] * M[k];
+    for (l = 0; l < WIENER_WIN2; ++l)
+      Q += ab[k] * H[k * WIENER_WIN2 + l] * ab[l];
+  }
+  Score = Q - 2 * P;
+
+  iP = M[WIENER_WIN2 >> 1];
+  iQ = H[(WIENER_WIN2 >> 1) * WIENER_WIN2 + (WIENER_WIN2 >> 1)];
+  iScore = iQ - 2 * iP;
+
+  return Score - iScore;
+}
+
+static void quantize_sym_filter(double *f, InterpKernel fi) {
+  int i;
+  for (i = 0; i < WIENER_HALFWIN; ++i) {
+    fi[i] = RINT(f[i] * WIENER_FILT_STEP);
+  }
+  // Specialize for 7-tap filter
+  fi[0] = CLIP(fi[0], WIENER_FILT_TAP0_MINV, WIENER_FILT_TAP0_MAXV);
+  fi[1] = CLIP(fi[1], WIENER_FILT_TAP1_MINV, WIENER_FILT_TAP1_MAXV);
+  fi[2] = CLIP(fi[2], WIENER_FILT_TAP2_MINV, WIENER_FILT_TAP2_MAXV);
+  // Satisfy filter constraints
+  fi[WIENER_WIN - 1] = fi[0];
+  fi[WIENER_WIN - 2] = fi[1];
+  fi[WIENER_WIN - 3] = fi[2];
+  // The central element has an implicit +WIENER_FILT_STEP
+  fi[3] = -2 * (fi[0] + fi[1] + fi[2]);
+}
+
+static int count_wiener_bits(WienerInfo *wiener_info,
+                             WienerInfo *ref_wiener_info) {
+  int bits = 0;
+  bits += aom_count_primitive_refsubexpfin(
+      WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
+      WIENER_FILT_TAP0_SUBEXP_K,
+      ref_wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV,
+      wiener_info->vfilter[0] - WIENER_FILT_TAP0_MINV);
+  bits += aom_count_primitive_refsubexpfin(
+      WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
+      WIENER_FILT_TAP1_SUBEXP_K,
+      ref_wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV,
+      wiener_info->vfilter[1] - WIENER_FILT_TAP1_MINV);
+  bits += aom_count_primitive_refsubexpfin(
+      WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
+      WIENER_FILT_TAP2_SUBEXP_K,
+      ref_wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV,
+      wiener_info->vfilter[2] - WIENER_FILT_TAP2_MINV);
+  bits += aom_count_primitive_refsubexpfin(
+      WIENER_FILT_TAP0_MAXV - WIENER_FILT_TAP0_MINV + 1,
+      WIENER_FILT_TAP0_SUBEXP_K,
+      ref_wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV,
+      wiener_info->hfilter[0] - WIENER_FILT_TAP0_MINV);
+  bits += aom_count_primitive_refsubexpfin(
+      WIENER_FILT_TAP1_MAXV - WIENER_FILT_TAP1_MINV + 1,
+      WIENER_FILT_TAP1_SUBEXP_K,
+      ref_wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV,
+      wiener_info->hfilter[1] - WIENER_FILT_TAP1_MINV);
+  bits += aom_count_primitive_refsubexpfin(
+      WIENER_FILT_TAP2_MAXV - WIENER_FILT_TAP2_MINV + 1,
+      WIENER_FILT_TAP2_SUBEXP_K,
+      ref_wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV,
+      wiener_info->hfilter[2] - WIENER_FILT_TAP2_MINV);
+  return bits;
+}
+
+static double search_wiener_uv(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
+                               int partial_frame, int plane,
+                               RestorationInfo *info, RestorationType *type,
+                               YV12_BUFFER_CONFIG *dst_frame) {
+  WienerInfo *wiener_info = info->wiener_info;
+  AV1_COMMON *const cm = &cpi->common;
+  RestorationInfo *rsi = cpi->rst_search;
+  int64_t err;
+  int bits;
+  double cost_wiener, cost_norestore, cost_wiener_frame, cost_norestore_frame;
+  MACROBLOCK *x = &cpi->td.mb;
+  double M[WIENER_WIN2];
+  double H[WIENER_WIN2 * WIENER_WIN2];
+  double vfilterd[WIENER_WIN], hfilterd[WIENER_WIN];
+  const YV12_BUFFER_CONFIG *dgd = cm->frame_to_show;
+  const int width = src->uv_crop_width;
+  const int height = src->uv_crop_height;
+  const int src_stride = src->uv_stride;
+  const int dgd_stride = dgd->uv_stride;
+  double score;
+  int tile_idx, tile_width, tile_height, nhtiles, nvtiles;
+  int h_start, h_end, v_start, v_end;
+  const int ntiles =
+      av1_get_rest_ntiles(width, height, cm->rst_info[1].restoration_tilesize,
+                          &tile_width, &tile_height, &nhtiles, &nvtiles);
+  WienerInfo ref_wiener_info;
+  set_default_wiener(&ref_wiener_info);
+  assert(width == dgd->uv_crop_width);
+  assert(height == dgd->uv_crop_height);
+
+  rsi[plane].frame_restoration_type = RESTORE_NONE;
+  err = sse_restoration_frame(cm, src, cm->frame_to_show, (1 << plane));
+  bits = 0;
+  cost_norestore_frame = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+
+  rsi[plane].frame_restoration_type = RESTORE_WIENER;
+
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    rsi[plane].restoration_type[tile_idx] = RESTORE_NONE;
+  }
+
+  // Compute best Wiener filters for each tile
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
+                             tile_height, width, height, 0, 0, &h_start, &h_end,
+                             &v_start, &v_end);
+    err = sse_restoration_tile(src, cm->frame_to_show, cm, h_start,
+                               h_end - h_start, v_start, v_end - v_start,
+                               1 << plane);
+    // #bits when a tile is not restored
+    bits = av1_cost_bit(RESTORE_NONE_WIENER_PROB, 0);
+    cost_norestore = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+    // best_tile_cost[tile_idx] = DBL_MAX;
+
+    av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
+                             tile_height, width, height, WIENER_HALFWIN,
+                             WIENER_HALFWIN, &h_start, &h_end, &v_start,
+                             &v_end);
+    if (plane == AOM_PLANE_U) {
+#if CONFIG_HIGHBITDEPTH
+      if (cm->use_highbitdepth)
+        compute_stats_highbd(dgd->u_buffer, src->u_buffer, h_start, h_end,
+                             v_start, v_end, dgd_stride, src_stride, M, H);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        compute_stats(dgd->u_buffer, src->u_buffer, h_start, h_end, v_start,
+                      v_end, dgd_stride, src_stride, M, H);
+    } else if (plane == AOM_PLANE_V) {
+#if CONFIG_HIGHBITDEPTH
+      if (cm->use_highbitdepth)
+        compute_stats_highbd(dgd->v_buffer, src->v_buffer, h_start, h_end,
+                             v_start, v_end, dgd_stride, src_stride, M, H);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        compute_stats(dgd->v_buffer, src->v_buffer, h_start, h_end, v_start,
+                      v_end, dgd_stride, src_stride, M, H);
+    } else {
+      assert(0);
+    }
+
+    type[tile_idx] = RESTORE_WIENER;
+
+    if (!wiener_decompose_sep_sym(M, H, vfilterd, hfilterd)) {
+      type[tile_idx] = RESTORE_NONE;
+      continue;
+    }
+    quantize_sym_filter(vfilterd, rsi[plane].wiener_info[tile_idx].vfilter);
+    quantize_sym_filter(hfilterd, rsi[plane].wiener_info[tile_idx].hfilter);
+
+    // Filter score computes the value of the function x'*A*x - x'*b for the
+    // learned filter and compares it against identity filer. If there is no
+    // reduction in the function, the filter is reverted back to identity
+    score = compute_score(M, H, rsi[plane].wiener_info[tile_idx].vfilter,
+                          rsi[plane].wiener_info[tile_idx].hfilter);
+    if (score > 0.0) {
+      type[tile_idx] = RESTORE_NONE;
+      continue;
+    }
+
+    rsi[plane].restoration_type[tile_idx] = RESTORE_WIENER;
+    err = try_restoration_tile(src, cpi, rsi, 1 << plane, partial_frame,
+                               tile_idx, 0, 0, dst_frame);
+    bits =
+        count_wiener_bits(&rsi[plane].wiener_info[tile_idx], &ref_wiener_info)
+        << AV1_PROB_COST_SHIFT;
+    // bits = WIENER_FILT_BITS << AV1_PROB_COST_SHIFT;
+    bits += av1_cost_bit(RESTORE_NONE_WIENER_PROB, 1);
+    cost_wiener = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+    if (cost_wiener >= cost_norestore) {
+      type[tile_idx] = RESTORE_NONE;
+    } else {
+      type[tile_idx] = RESTORE_WIENER;
+      memcpy(&wiener_info[tile_idx], &rsi[plane].wiener_info[tile_idx],
+             sizeof(wiener_info[tile_idx]));
+      memcpy(&ref_wiener_info, &rsi[plane].wiener_info[tile_idx],
+             sizeof(ref_wiener_info));
+    }
+    rsi[plane].restoration_type[tile_idx] = RESTORE_NONE;
+  }
+  // Cost for Wiener filtering
+  set_default_wiener(&ref_wiener_info);
+  bits = 0;
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    bits +=
+        av1_cost_bit(RESTORE_NONE_WIENER_PROB, type[tile_idx] != RESTORE_NONE);
+    memcpy(&rsi[plane].wiener_info[tile_idx], &wiener_info[tile_idx],
+           sizeof(wiener_info[tile_idx]));
+    if (type[tile_idx] == RESTORE_WIENER) {
+      bits +=
+          count_wiener_bits(&rsi[plane].wiener_info[tile_idx], &ref_wiener_info)
+          << AV1_PROB_COST_SHIFT;
+      memcpy(&ref_wiener_info, &rsi[plane].wiener_info[tile_idx],
+             sizeof(ref_wiener_info));
+    }
+    rsi[plane].restoration_type[tile_idx] = type[tile_idx];
+  }
+  err = try_restoration_frame(src, cpi, rsi, 1 << plane, partial_frame,
+                              dst_frame);
+  cost_wiener_frame = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+
+  if (cost_wiener_frame < cost_norestore_frame) {
+    info->frame_restoration_type = RESTORE_WIENER;
+  } else {
+    info->frame_restoration_type = RESTORE_NONE;
+  }
+
+  return info->frame_restoration_type == RESTORE_WIENER ? cost_wiener_frame
+                                                        : cost_norestore_frame;
+}
+
+static double search_wiener(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
+                            int partial_frame, RestorationInfo *info,
+                            RestorationType *type, double *best_tile_cost,
+                            YV12_BUFFER_CONFIG *dst_frame) {
+  WienerInfo *wiener_info = info->wiener_info;
+  AV1_COMMON *const cm = &cpi->common;
+  RestorationInfo *rsi = cpi->rst_search;
+  int64_t err;
+  int bits;
+  double cost_wiener, cost_norestore;
+  MACROBLOCK *x = &cpi->td.mb;
+  double M[WIENER_WIN2];
+  double H[WIENER_WIN2 * WIENER_WIN2];
+  double vfilterd[WIENER_WIN], hfilterd[WIENER_WIN];
+  const YV12_BUFFER_CONFIG *dgd = cm->frame_to_show;
+  const int width = cm->width;
+  const int height = cm->height;
+  const int src_stride = src->y_stride;
+  const int dgd_stride = dgd->y_stride;
+  double score;
+  int tile_idx, tile_width, tile_height, nhtiles, nvtiles;
+  int h_start, h_end, v_start, v_end;
+  const int ntiles =
+      av1_get_rest_ntiles(width, height, cm->rst_info[0].restoration_tilesize,
+                          &tile_width, &tile_height, &nhtiles, &nvtiles);
+  WienerInfo ref_wiener_info;
+  set_default_wiener(&ref_wiener_info);
+
+  assert(width == dgd->y_crop_width);
+  assert(height == dgd->y_crop_height);
+  assert(width == src->y_crop_width);
+  assert(height == src->y_crop_height);
+
+  rsi->frame_restoration_type = RESTORE_WIENER;
+
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    rsi->restoration_type[tile_idx] = RESTORE_NONE;
+  }
+
+// Construct a (WIENER_HALFWIN)-pixel border around the frame
+#if CONFIG_HIGHBITDEPTH
+  if (cm->use_highbitdepth)
+    extend_frame_highbd(CONVERT_TO_SHORTPTR(dgd->y_buffer), width, height,
+                        dgd_stride);
+  else
+#endif
+    extend_frame(dgd->y_buffer, width, height, dgd_stride);
+
+  // Compute best Wiener filters for each tile
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
+                             tile_height, width, height, 0, 0, &h_start, &h_end,
+                             &v_start, &v_end);
+    err = sse_restoration_tile(src, cm->frame_to_show, cm, h_start,
+                               h_end - h_start, v_start, v_end - v_start, 1);
+    // #bits when a tile is not restored
+    bits = av1_cost_bit(RESTORE_NONE_WIENER_PROB, 0);
+    cost_norestore = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+    best_tile_cost[tile_idx] = DBL_MAX;
+
+    av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
+                             tile_height, width, height, 0, 0, &h_start, &h_end,
+                             &v_start, &v_end);
+#if CONFIG_HIGHBITDEPTH
+    if (cm->use_highbitdepth)
+      compute_stats_highbd(dgd->y_buffer, src->y_buffer, h_start, h_end,
+                           v_start, v_end, dgd_stride, src_stride, M, H);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      compute_stats(dgd->y_buffer, src->y_buffer, h_start, h_end, v_start,
+                    v_end, dgd_stride, src_stride, M, H);
+
+    type[tile_idx] = RESTORE_WIENER;
+
+    if (!wiener_decompose_sep_sym(M, H, vfilterd, hfilterd)) {
+      type[tile_idx] = RESTORE_NONE;
+      continue;
+    }
+    quantize_sym_filter(vfilterd, rsi->wiener_info[tile_idx].vfilter);
+    quantize_sym_filter(hfilterd, rsi->wiener_info[tile_idx].hfilter);
+
+    // Filter score computes the value of the function x'*A*x - x'*b for the
+    // learned filter and compares it against identity filer. If there is no
+    // reduction in the function, the filter is reverted back to identity
+    score = compute_score(M, H, rsi->wiener_info[tile_idx].vfilter,
+                          rsi->wiener_info[tile_idx].hfilter);
+    if (score > 0.0) {
+      type[tile_idx] = RESTORE_NONE;
+      continue;
+    }
+
+    rsi->restoration_type[tile_idx] = RESTORE_WIENER;
+    err = try_restoration_tile(src, cpi, rsi, 1, partial_frame, tile_idx, 0, 0,
+                               dst_frame);
+    bits = count_wiener_bits(&rsi->wiener_info[tile_idx], &ref_wiener_info)
+           << AV1_PROB_COST_SHIFT;
+    bits += av1_cost_bit(RESTORE_NONE_WIENER_PROB, 1);
+    cost_wiener = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+    if (cost_wiener >= cost_norestore) {
+      type[tile_idx] = RESTORE_NONE;
+    } else {
+      type[tile_idx] = RESTORE_WIENER;
+      memcpy(&wiener_info[tile_idx], &rsi->wiener_info[tile_idx],
+             sizeof(wiener_info[tile_idx]));
+      memcpy(&ref_wiener_info, &rsi->wiener_info[tile_idx],
+             sizeof(ref_wiener_info));
+      bits = count_wiener_bits(&wiener_info[tile_idx], &ref_wiener_info)
+             << AV1_PROB_COST_SHIFT;
+      best_tile_cost[tile_idx] = err;
+    }
+    rsi->restoration_type[tile_idx] = RESTORE_NONE;
+  }
+  // Cost for Wiener filtering
+  set_default_wiener(&ref_wiener_info);
+  bits = frame_level_restore_bits[rsi->frame_restoration_type]
+         << AV1_PROB_COST_SHIFT;
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    bits +=
+        av1_cost_bit(RESTORE_NONE_WIENER_PROB, type[tile_idx] != RESTORE_NONE);
+    memcpy(&rsi->wiener_info[tile_idx], &wiener_info[tile_idx],
+           sizeof(wiener_info[tile_idx]));
+    if (type[tile_idx] == RESTORE_WIENER) {
+      bits += count_wiener_bits(&rsi->wiener_info[tile_idx], &ref_wiener_info)
+              << AV1_PROB_COST_SHIFT;
+      memcpy(&ref_wiener_info, &rsi->wiener_info[tile_idx],
+             sizeof(ref_wiener_info));
+    }
+    rsi->restoration_type[tile_idx] = type[tile_idx];
+  }
+  err = try_restoration_frame(src, cpi, rsi, 1, partial_frame, dst_frame);
+  cost_wiener = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+
+  return cost_wiener;
+}
+
+static double search_norestore(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
+                               int partial_frame, RestorationInfo *info,
+                               RestorationType *type, double *best_tile_cost,
+                               YV12_BUFFER_CONFIG *dst_frame) {
+  double err, cost_norestore;
+  int bits;
+  MACROBLOCK *x = &cpi->td.mb;
+  AV1_COMMON *const cm = &cpi->common;
+  int tile_idx, tile_width, tile_height, nhtiles, nvtiles;
+  int h_start, h_end, v_start, v_end;
+  const int ntiles = av1_get_rest_ntiles(
+      cm->width, cm->height, cm->rst_info[0].restoration_tilesize, &tile_width,
+      &tile_height, &nhtiles, &nvtiles);
+  (void)info;
+  (void)dst_frame;
+  (void)partial_frame;
+
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    av1_get_rest_tile_limits(tile_idx, 0, 0, nhtiles, nvtiles, tile_width,
+                             tile_height, cm->width, cm->height, 0, 0, &h_start,
+                             &h_end, &v_start, &v_end);
+    err = sse_restoration_tile(src, cm->frame_to_show, cm, h_start,
+                               h_end - h_start, v_start, v_end - v_start, 1);
+    type[tile_idx] = RESTORE_NONE;
+    best_tile_cost[tile_idx] = err;
+  }
+  // RD cost associated with no restoration
+  err = sse_restoration_tile(src, cm->frame_to_show, cm, 0, cm->width, 0,
+                             cm->height, 1);
+  bits = frame_level_restore_bits[RESTORE_NONE] << AV1_PROB_COST_SHIFT;
+  cost_norestore = RDCOST_DBL(x->rdmult, x->rddiv, (bits >> 4), err);
+  return cost_norestore;
+}
+
+static double search_switchable_restoration(
+    AV1_COMP *cpi, int partial_frame, RestorationInfo *rsi,
+    double *tile_cost[RESTORE_SWITCHABLE_TYPES]) {
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *x = &cpi->td.mb;
+  double cost_switchable = 0;
+  int bits, tile_idx;
+  RestorationType r;
+  const int ntiles = av1_get_rest_ntiles(cm->width, cm->height,
+                                         cm->rst_info[0].restoration_tilesize,
+                                         NULL, NULL, NULL, NULL);
+  SgrprojInfo ref_sgrproj_info;
+  set_default_sgrproj(&ref_sgrproj_info);
+  WienerInfo ref_wiener_info;
+  set_default_wiener(&ref_wiener_info);
+  (void)partial_frame;
+
+  rsi->frame_restoration_type = RESTORE_SWITCHABLE;
+  bits = frame_level_restore_bits[rsi->frame_restoration_type]
+         << AV1_PROB_COST_SHIFT;
+  cost_switchable = RDCOST_DBL(x->rdmult, x->rddiv, bits >> 4, 0);
+  for (tile_idx = 0; tile_idx < ntiles; ++tile_idx) {
+    double best_cost = RDCOST_DBL(
+        x->rdmult, x->rddiv, (cpi->switchable_restore_cost[RESTORE_NONE] >> 4),
+        tile_cost[RESTORE_NONE][tile_idx]);
+    rsi->restoration_type[tile_idx] = RESTORE_NONE;
+    for (r = 1; r < RESTORE_SWITCHABLE_TYPES; r++) {
+      if (force_restore_type != 0)
+        if (r != force_restore_type) continue;
+      int tilebits = 0;
+      if (r == RESTORE_WIENER)
+        tilebits +=
+            count_wiener_bits(&rsi->wiener_info[tile_idx], &ref_wiener_info);
+      else if (r == RESTORE_SGRPROJ)
+        tilebits +=
+            count_sgrproj_bits(&rsi->sgrproj_info[tile_idx], &ref_sgrproj_info);
+      tilebits <<= AV1_PROB_COST_SHIFT;
+      tilebits += cpi->switchable_restore_cost[r];
+      double cost = RDCOST_DBL(x->rdmult, x->rddiv, tilebits >> 4,
+                               tile_cost[r][tile_idx]);
+
+      if (cost < best_cost) {
+        rsi->restoration_type[tile_idx] = r;
+        best_cost = cost;
+      }
+    }
+    if (rsi->restoration_type[tile_idx] == RESTORE_WIENER)
+      memcpy(&ref_wiener_info, &rsi->wiener_info[tile_idx],
+             sizeof(ref_wiener_info));
+    else if (rsi->restoration_type[tile_idx] == RESTORE_SGRPROJ)
+      memcpy(&ref_sgrproj_info, &rsi->sgrproj_info[tile_idx],
+             sizeof(ref_sgrproj_info));
+    if (force_restore_type != 0)
+      assert(rsi->restoration_type[tile_idx] == force_restore_type ||
+             rsi->restoration_type[tile_idx] == RESTORE_NONE);
+    cost_switchable += best_cost;
+  }
+  return cost_switchable;
+}
+
+void av1_pick_filter_restoration(const YV12_BUFFER_CONFIG *src, AV1_COMP *cpi,
+                                 LPF_PICK_METHOD method) {
+  static search_restore_type search_restore_fun[RESTORE_SWITCHABLE_TYPES] = {
+    search_norestore, search_wiener, search_sgrproj,
+  };
+  AV1_COMMON *const cm = &cpi->common;
+  double cost_restore[RESTORE_TYPES];
+  double *tile_cost[RESTORE_SWITCHABLE_TYPES];
+  RestorationType *restore_types[RESTORE_SWITCHABLE_TYPES];
+  double best_cost_restore;
+  RestorationType r, best_restore;
+
+  const int ntiles = av1_get_rest_ntiles(cm->width, cm->height,
+                                         cm->rst_info[0].restoration_tilesize,
+                                         NULL, NULL, NULL, NULL);
+
+  for (r = 0; r < RESTORE_SWITCHABLE_TYPES; r++) {
+    tile_cost[r] = (double *)aom_malloc(sizeof(*tile_cost[0]) * ntiles);
+    restore_types[r] =
+        (RestorationType *)aom_malloc(sizeof(*restore_types[0]) * ntiles);
+  }
+
+  for (r = 0; r < RESTORE_SWITCHABLE_TYPES; ++r) {
+    if (force_restore_type != 0)
+      if (r != RESTORE_NONE && r != force_restore_type) continue;
+    cost_restore[r] = search_restore_fun[r](
+        src, cpi, method == LPF_PICK_FROM_SUBIMAGE, &cm->rst_info[0],
+        restore_types[r], tile_cost[r], &cpi->trial_frame_rst);
+  }
+  cost_restore[RESTORE_SWITCHABLE] = search_switchable_restoration(
+      cpi, method == LPF_PICK_FROM_SUBIMAGE, &cm->rst_info[0], tile_cost);
+
+  best_cost_restore = DBL_MAX;
+  best_restore = 0;
+  for (r = 0; r < RESTORE_TYPES; ++r) {
+    if (force_restore_type != 0)
+      if (r != RESTORE_NONE && r != force_restore_type) continue;
+    if (cost_restore[r] < best_cost_restore) {
+      best_restore = r;
+      best_cost_restore = cost_restore[r];
+    }
+  }
+  cm->rst_info[0].frame_restoration_type = best_restore;
+  if (force_restore_type != 0)
+    assert(best_restore == force_restore_type || best_restore == RESTORE_NONE);
+  if (best_restore != RESTORE_SWITCHABLE) {
+    memcpy(cm->rst_info[0].restoration_type, restore_types[best_restore],
+           ntiles * sizeof(restore_types[best_restore][0]));
+  }
+
+  // Color components
+  search_wiener_uv(src, cpi, method == LPF_PICK_FROM_SUBIMAGE, AOM_PLANE_U,
+                   &cm->rst_info[AOM_PLANE_U],
+                   cm->rst_info[AOM_PLANE_U].restoration_type,
+                   &cpi->trial_frame_rst);
+  search_wiener_uv(src, cpi, method == LPF_PICK_FROM_SUBIMAGE, AOM_PLANE_V,
+                   &cm->rst_info[AOM_PLANE_V],
+                   cm->rst_info[AOM_PLANE_V].restoration_type,
+                   &cpi->trial_frame_rst);
+  /*
+  printf("Frame %d/%d restore types: %d %d %d\n",
+         cm->current_video_frame, cm->show_frame,
+         cm->rst_info[0].frame_restoration_type,
+         cm->rst_info[1].frame_restoration_type,
+         cm->rst_info[2].frame_restoration_type);
+  printf("Frame %d/%d frame_restore_type %d : %f %f %f %f\n",
+         cm->current_video_frame, cm->show_frame,
+         cm->rst_info[0].frame_restoration_type, cost_restore[0],
+         cost_restore[1], cost_restore[2], cost_restore[3]);
+         */
+
+  for (r = 0; r < RESTORE_SWITCHABLE_TYPES; r++) {
+    aom_free(tile_cost[r]);
+    aom_free(restore_types[r]);
+  }
+}
diff --git a/third_party/aom/av1/encoder/pickrst.h b/third_party/aom/av1/encoder/pickrst.h
new file mode 100644
index 0000000000..f6096ed1d1
--- /dev/null
+++ b/third_party/aom/av1/encoder/pickrst.h
@@ -0,0 +1,30 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef AV1_ENCODER_PICKRST_H_
+#define AV1_ENCODER_PICKRST_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "av1/encoder/encoder.h"
+
+struct yv12_buffer_config;
+struct AV1_COMP;
+
+void av1_pick_filter_restoration(const YV12_BUFFER_CONFIG *sd, AV1_COMP *cpi,
+                                 LPF_PICK_METHOD method);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_PICKRST_H_
diff --git a/third_party/aom/av1/encoder/pvq_encoder.c b/third_party/aom/av1/encoder/pvq_encoder.c
new file mode 100644
index 0000000000..ab63f1b7dd
--- /dev/null
+++ b/third_party/aom/av1/encoder/pvq_encoder.c
@@ -0,0 +1,988 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#ifdef HAVE_CONFIG_H
+# include "config.h"
+#endif
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include "aom_dsp/entcode.h"
+#include "aom_dsp/entenc.h"
+#include "av1/common/blockd.h"
+#include "av1/common/odintrin.h"
+#include "av1/common/partition.h"
+#include "av1/common/pvq_state.h"
+#include "av1/encoder/encodemb.h"
+#include "av1/encoder/pvq_encoder.h"
+#include "aom_ports/system_state.h"
+
+/*Shift to ensure that the upper bound (i.e. for the max blocksize) of the
+   dot-product of the 1st band of chroma with the luma ref doesn't overflow.*/
+#define OD_CFL_FLIP_SHIFT (OD_LIMIT_BSIZE_MAX + 0)
+
+void aom_write_symbol_pvq(aom_writer *w, int symb, aom_cdf_prob *cdf,
+    int nsymbs) {
+  if (cdf[0] == 0)
+    aom_cdf_init_q15_1D(cdf, nsymbs, CDF_SIZE(nsymbs));
+  aom_write_symbol(w, symb, cdf, nsymbs);
+}
+
+static void aom_encode_pvq_codeword(aom_writer *w, od_pvq_codeword_ctx *adapt,
+ const od_coeff *in, int n, int k) {
+  int i;
+  aom_encode_band_pvq_splits(w, adapt, in, n, k, 0);
+  for (i = 0; i < n; i++) if (in[i]) aom_write_bit(w, in[i] < 0);
+}
+
+/* Computes 1/sqrt(i) using a table for small values. */
+static double od_rsqrt_table(int i) {
+  static double table[16] = {
+    1.000000, 0.707107, 0.577350, 0.500000,
+    0.447214, 0.408248, 0.377964, 0.353553,
+    0.333333, 0.316228, 0.301511, 0.288675,
+    0.277350, 0.267261, 0.258199, 0.250000};
+  if (i <= 16) return table[i-1];
+  else return 1./sqrt(i);
+}
+
+/*Computes 1/sqrt(start+2*i+1) using a lookup table containing the results
+   where 0 <= i < table_size.*/
+static double od_custom_rsqrt_dynamic_table(const double* table,
+ const int table_size, const double start, const int i) {
+  if (i < table_size) return table[i];
+  else return od_rsqrt_table((int)(start + 2*i + 1));
+}
+
+/*Fills tables used in od_custom_rsqrt_dynamic_table for a given start.*/
+static void od_fill_dynamic_rsqrt_table(double *table, const int table_size,
+ const double start) {
+  int i;
+  for (i = 0; i < table_size; i++)
+    table[i] = od_rsqrt_table((int)(start + 2*i + 1));
+}
+
+/** Find the codepoint on the given PSphere closest to the desired
+ * vector. Double-precision PVQ search just to make sure our tests
+ * aren't limited by numerical accuracy.
+ *
+ * @param [in]      xcoeff  input vector to quantize (x in the math doc)
+ * @param [in]      n       number of dimensions
+ * @param [in]      k       number of pulses
+ * @param [out]     ypulse  optimal codevector found (y in the math doc)
+ * @param [out]     g2      multiplier for the distortion (typically squared
+ *                          gain units)
+ * @param [in] pvq_norm_lambda enc->pvq_norm_lambda for quantized RDO
+ * @param [in]      prev_k  number of pulses already in ypulse that we should
+ *                          reuse for the search (or 0 for a new search)
+ * @return                  cosine distance between x and y (between 0 and 1)
+ */
+double pvq_search_rdo_double_c(const od_val16 *xcoeff, int n, int k,
+ od_coeff *ypulse, double g2, double pvq_norm_lambda, int prev_k) {
+  int i, j;
+  double xy;
+  double yy;
+  /* TODO - This blows our 8kB stack space budget and should be fixed when
+   converting PVQ to fixed point. */
+  double x[MAXN];
+  double xx;
+  double lambda;
+  double norm_1;
+  int rdo_pulses;
+  double delta_rate;
+  xx = xy = yy = 0;
+  for (j = 0; j < n; j++) {
+    x[j] = fabs((float)xcoeff[j]);
+    xx += x[j]*x[j];
+  }
+  norm_1 = 1./sqrt(1e-30 + xx);
+  lambda = pvq_norm_lambda/(1e-30 + g2);
+  i = 0;
+  if (prev_k > 0 && prev_k <= k) {
+    /* We reuse pulses from a previous search so we don't have to search them
+       again. */
+    for (j = 0; j < n; j++) {
+      ypulse[j] = abs(ypulse[j]);
+      xy += x[j]*ypulse[j];
+      yy += ypulse[j]*ypulse[j];
+      i += ypulse[j];
+    }
+  }
+  else if (k > 2) {
+    double l1_norm;
+    double l1_inv;
+    l1_norm = 0;
+    for (j = 0; j < n; j++) l1_norm += x[j];
+    l1_inv = 1./OD_MAXF(l1_norm, 1e-100);
+    for (j = 0; j < n; j++) {
+      double tmp;
+      tmp = k*x[j]*l1_inv;
+      ypulse[j] = OD_MAXI(0, (int)floor(tmp));
+      xy += x[j]*ypulse[j];
+      yy += ypulse[j]*ypulse[j];
+      i += ypulse[j];
+    }
+  }
+  else OD_CLEAR(ypulse, n);
+
+  /* Only use RDO on the last few pulses. This not only saves CPU, but using
+     RDO on all pulses actually makes the results worse for reasons I don't
+     fully understand. */
+  rdo_pulses = 1 + k/4;
+  /* Rough assumption for now, the last position costs about 3 bits more than
+     the first. */
+  delta_rate = 3./n;
+  /* Search one pulse at a time */
+  for (; i < k - rdo_pulses; i++) {
+    int pos;
+    double best_xy;
+    double best_yy;
+    pos = 0;
+    best_xy = -10;
+    best_yy = 1;
+    for (j = 0; j < n; j++) {
+      double tmp_xy;
+      double tmp_yy;
+      tmp_xy = xy + x[j];
+      tmp_yy = yy + 2*ypulse[j] + 1;
+      tmp_xy *= tmp_xy;
+      if (j == 0 || tmp_xy*best_yy > best_xy*tmp_yy) {
+        best_xy = tmp_xy;
+        best_yy = tmp_yy;
+        pos = j;
+      }
+    }
+    xy = xy + x[pos];
+    yy = yy + 2*ypulse[pos] + 1;
+    ypulse[pos]++;
+  }
+  /* Search last pulses with RDO. Distortion is D = (x-y)^2 = x^2 - 2*x*y + y^2
+     and since x^2 and y^2 are constant, we just maximize x*y, plus a
+     lambda*rate term. Note that since x and y aren't normalized here,
+     we need to divide by sqrt(x^2)*sqrt(y^2). */
+  for (; i < k; i++) {
+    double rsqrt_table[4];
+    int rsqrt_table_size = 4;
+    int pos;
+    double best_cost;
+    pos = 0;
+    best_cost = -1e5;
+    /*Fill the small rsqrt lookup table with inputs relative to yy.
+      Specifically, the table of n values is filled with
+       rsqrt(yy + 1), rsqrt(yy + 2 + 1) .. rsqrt(yy + 2*(n-1) + 1).*/
+    od_fill_dynamic_rsqrt_table(rsqrt_table, rsqrt_table_size, yy);
+    for (j = 0; j < n; j++) {
+      double tmp_xy;
+      double tmp_yy;
+      tmp_xy = xy + x[j];
+      /*Calculate rsqrt(yy + 2*ypulse[j] + 1) using an optimized method.*/
+      tmp_yy = od_custom_rsqrt_dynamic_table(rsqrt_table, rsqrt_table_size,
+       yy, ypulse[j]);
+      tmp_xy = 2*tmp_xy*norm_1*tmp_yy - lambda*j*delta_rate;
+      if (j == 0 || tmp_xy > best_cost) {
+        best_cost = tmp_xy;
+        pos = j;
+      }
+    }
+    xy = xy + x[pos];
+    yy = yy + 2*ypulse[pos] + 1;
+    ypulse[pos]++;
+  }
+  for (i = 0; i < n; i++) {
+    if (xcoeff[i] < 0) ypulse[i] = -ypulse[i];
+  }
+  return xy/(1e-100 + sqrt(xx*yy));
+}
+
+/** Encodes the gain so that the return value increases with the
+ * distance |x-ref|, so that we can encode a zero when x=ref. The
+ * value x=0 is not covered because it is only allowed in the noref
+ * case.
+ *
+ * @param [in]      x      quantized gain to encode
+ * @param [in]      ref    quantized gain of the reference
+ * @return                 interleave-encoded quantized gain value
+ */
+static int neg_interleave(int x, int ref) {
+  if (x < ref) return -2*(x - ref) - 1;
+  else if (x < 2*ref) return 2*(x - ref);
+  else return x-1;
+}
+
+int od_vector_is_null(const od_coeff *x, int len) {
+  int i;
+  for (i = 0; i < len; i++) if (x[i]) return 0;
+  return 1;
+}
+
+static double od_pvq_rate(int qg, int icgr, int theta, int ts,
+  const od_adapt_ctx *adapt, const od_coeff *y0, int k, int n, int speed) {
+  double rate;
+  if (k == 0) rate = 0;
+  else if (speed > 0) {
+    int i;
+    int sum;
+    double f;
+    /* Compute "center of mass" of the pulse vector. */
+    sum = 0;
+    for (i = 0; i < n - (theta != -1); i++) sum += i*abs(y0[i]);
+    f = sum/(double)(k*n);
+    /* Estimates the number of bits it will cost to encode K pulses in
+       N dimensions based on hand-tuned fit for bitrate vs K, N and
+       "center of mass". */
+    rate = (1 + .4*f)*n*OD_LOG2(1 + OD_MAXF(0, log(n*2*(1*f + .025))*k/n)) + 3;
+  }
+  else {
+    aom_writer w;
+    od_pvq_codeword_ctx cd;
+    int tell;
+#if CONFIG_DAALA_EC
+    od_ec_enc_init(&w.ec, 1000);
+#else
+# error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+    OD_COPY(&cd, &adapt->pvq.pvq_codeword_ctx, 1);
+#if CONFIG_DAALA_EC
+    tell = od_ec_enc_tell_frac(&w.ec);
+#else
+# error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+    aom_encode_pvq_codeword(&w, &cd, y0, n - (theta != -1), k);
+#if CONFIG_DAALA_EC
+    rate = (od_ec_enc_tell_frac(&w.ec)-tell)/8.;
+    od_ec_enc_clear(&w.ec);
+#else
+# error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  }
+  if (qg > 0 && theta >= 0) {
+    /* Approximate cost of entropy-coding theta */
+    rate += .9*OD_LOG2(ts);
+    if (qg == icgr) rate -= .5;
+  }
+  return rate;
+}
+
+#define MAX_PVQ_ITEMS (20)
+/* This stores the information about a PVQ search candidate, so we can sort
+   based on K. */
+typedef struct {
+  int gain;
+  int k;
+  od_val32 qtheta;
+  int theta;
+  int ts;
+  od_val32 qcg;
+} pvq_search_item;
+
+int items_compare(pvq_search_item *a, pvq_search_item *b) {
+  /* Break ties in K with gain to ensure a stable sort.
+     Otherwise, the order depends on qsort implementation. */
+  return a->k == b->k ? a->gain - b->gain : a->k - b->k;
+}
+
+/** Perform PVQ quantization with prediction, trying several
+ * possible gains and angles. See draft-valin-videocodec-pvq and
+ * http://jmvalin.ca/slides/pvq.pdf for more details.
+ *
+ * @param [out]    out         coefficients after quantization
+ * @param [in]     x0          coefficients before quantization
+ * @param [in]     r0          reference, aka predicted coefficients
+ * @param [in]     n           number of dimensions
+ * @param [in]     q0          quantization step size
+ * @param [out]    y           pulse vector (i.e. selected PVQ codevector)
+ * @param [out]    itheta      angle between input and reference (-1 if noref)
+ * @param [out]    vk          total number of pulses
+ * @param [in]     beta        per-band activity masking beta param
+ * @param [out]    skip_diff   distortion cost of skipping this block
+ *                             (accumulated)
+ * @param [in]     is_keyframe whether we're encoding a keyframe
+ * @param [in]     pli         plane index
+ * @param [in]     adapt       probability adaptation context
+ * @param [in]     qm          QM with magnitude compensation
+ * @param [in]     qm_inv      Inverse of QM with magnitude compensation
+ * @param [in] pvq_norm_lambda enc->pvq_norm_lambda for quantized RDO
+ * @param [in]     speed       Make search faster by making approximations
+ * @return         gain        index of the quatized gain
+*/
+static int pvq_theta(od_coeff *out, const od_coeff *x0, const od_coeff *r0,
+    int n, int q0, od_coeff *y, int *itheta, int *vk,
+    od_val16 beta, double *skip_diff, int is_keyframe, int pli,
+    const od_adapt_ctx *adapt, const int16_t *qm, const int16_t *qm_inv,
+    double pvq_norm_lambda, int speed) {
+  od_val32 g;
+  od_val32 gr;
+  od_coeff y_tmp[MAXN + 3];
+  int i;
+  /* Number of pulses. */
+  int k;
+  /* Companded gain of x and reference, normalized to q. */
+  od_val32 cg;
+  od_val32 cgr;
+  int icgr;
+  int qg;
+  /* Best RDO cost (D + lamdba*R) so far. */
+  double best_cost;
+  double dist0;
+  /* Distortion (D) that corresponds to the best RDO cost. */
+  double best_dist;
+  double dist;
+  /* Sign of Householder reflection. */
+  int s;
+  /* Dimension on which Householder reflects. */
+  int m;
+  od_val32 theta;
+  double corr;
+  int best_k;
+  od_val32 best_qtheta;
+  od_val32 gain_offset;
+  int noref;
+  double skip_dist;
+  int cfl_enabled;
+  int skip;
+  double gain_weight;
+  od_val16 x16[MAXN];
+  od_val16 r16[MAXN];
+  int xshift;
+  int rshift;
+  /* Give more weight to gain error when calculating the total distortion. */
+  gain_weight = 1.0;
+  OD_ASSERT(n > 1);
+  corr = 0;
+#if !defined(OD_FLOAT_PVQ)
+  /* Shift needed to make x fit in 16 bits even after rotation.
+     This shift value is not normative (it can be changed without breaking
+     the bitstream) */
+  xshift = OD_MAXI(0, od_vector_log_mag(x0, n) - 15);
+  /* Shift needed to make the reference fit in 15 bits, so that the Householder
+     vector can fit in 16 bits.
+     This shift value *is* normative, and has to match the decoder. */
+  rshift = OD_MAXI(0, od_vector_log_mag(r0, n) - 14);
+#else
+  xshift = 0;
+  rshift = 0;
+#endif
+  for (i = 0; i < n; i++) {
+#if defined(OD_FLOAT_PVQ)
+    /*This is slightly different from the original float PVQ code,
+       where the qm was applied in the accumulation in od_pvq_compute_gain and
+       the vectors were od_coeffs, not od_val16 (i.e. double).*/
+    x16[i] = x0[i]*(double)qm[i]*OD_QM_SCALE_1;
+    r16[i] = r0[i]*(double)qm[i]*OD_QM_SCALE_1;
+#else
+    x16[i] = OD_SHR_ROUND(x0[i]*qm[i], OD_QM_SHIFT + xshift);
+    r16[i] = OD_SHR_ROUND(r0[i]*qm[i], OD_QM_SHIFT + rshift);
+#endif
+    corr += OD_MULT16_16(x16[i], r16[i]);
+  }
+  cfl_enabled = is_keyframe && pli != 0 && !OD_DISABLE_CFL;
+  cg  = od_pvq_compute_gain(x16, n, q0, &g, beta, xshift);
+  cgr = od_pvq_compute_gain(r16, n, q0, &gr, beta, rshift);
+  if (cfl_enabled) cgr = OD_CGAIN_SCALE;
+  /* gain_offset is meant to make sure one of the quantized gains has
+     exactly the same gain as the reference. */
+#if defined(OD_FLOAT_PVQ)
+  icgr = (int)floor(.5 + cgr);
+#else
+  icgr = OD_SHR_ROUND(cgr, OD_CGAIN_SHIFT);
+#endif
+  gain_offset = cgr - OD_SHL(icgr, OD_CGAIN_SHIFT);
+  /* Start search with null case: gain=0, no pulse. */
+  qg = 0;
+  dist = gain_weight*cg*cg*OD_CGAIN_SCALE_2;
+  best_dist = dist;
+  best_cost = dist + pvq_norm_lambda*od_pvq_rate(0, 0, -1, 0, adapt, NULL, 0,
+    n, speed);
+  noref = 1;
+  best_k = 0;
+  *itheta = -1;
+  OD_CLEAR(y, n);
+  best_qtheta = 0;
+  m = 0;
+  s = 1;
+  corr = corr/(1e-100 + g*(double)gr/OD_SHL(1, xshift + rshift));
+  corr = OD_MAXF(OD_MINF(corr, 1.), -1.);
+  if (is_keyframe) skip_dist = gain_weight*cg*cg*OD_CGAIN_SCALE_2;
+  else {
+    skip_dist = gain_weight*(cg - cgr)*(cg - cgr)
+     + cgr*(double)cg*(2 - 2*corr);
+    skip_dist *= OD_CGAIN_SCALE_2;
+  }
+  if (!is_keyframe) {
+    /* noref, gain=0 isn't allowed, but skip is allowed. */
+    od_val32 scgr;
+    scgr = OD_MAXF(0,gain_offset);
+    if (icgr == 0) {
+      best_dist = gain_weight*(cg - scgr)*(cg - scgr)
+       + scgr*(double)cg*(2 - 2*corr);
+      best_dist *= OD_CGAIN_SCALE_2;
+    }
+    best_cost = best_dist + pvq_norm_lambda*od_pvq_rate(0, icgr, 0, 0, adapt,
+     NULL, 0, n, speed);
+    best_qtheta = 0;
+    *itheta = 0;
+    noref = 0;
+  }
+  dist0 = best_dist;
+  if (n <= OD_MAX_PVQ_SIZE && !od_vector_is_null(r0, n) && corr > 0) {
+    od_val16 xr[MAXN];
+    int gain_bound;
+    int prev_k;
+    pvq_search_item items[MAX_PVQ_ITEMS];
+    int idx;
+    int nitems;
+    double cos_dist;
+    idx = 0;
+    gain_bound = OD_SHR(cg - gain_offset, OD_CGAIN_SHIFT);
+    /* Perform theta search only if prediction is useful. */
+    theta = OD_ROUND32(OD_THETA_SCALE*acos(corr));
+    m = od_compute_householder(r16, n, gr, &s, rshift);
+    od_apply_householder(xr, x16, r16, n);
+    prev_k = 0;
+    for (i = m; i < n - 1; i++) xr[i] = xr[i + 1];
+    /* Compute all candidate PVQ searches within a reasonable range of gain
+       and theta. */
+    for (i = OD_MAXI(1, gain_bound - 1); i <= gain_bound + 1; i++) {
+      int j;
+      od_val32 qcg;
+      int ts;
+      int theta_lower;
+      int theta_upper;
+      /* Quantized companded gain */
+      qcg = OD_SHL(i, OD_CGAIN_SHIFT) + gain_offset;
+      /* Set angular resolution (in ra) to match the encoded gain */
+      ts = od_pvq_compute_max_theta(qcg, beta);
+      theta_lower = OD_MAXI(0, (int)floor(.5 +
+       theta*OD_THETA_SCALE_1*2/M_PI*ts) - 2);
+      theta_upper = OD_MINI(ts - 1, (int)ceil(theta*OD_THETA_SCALE_1*2/M_PI*ts));
+      /* Include the angles within a reasonable range. */
+      for (j = theta_lower; j <= theta_upper; j++) {
+        od_val32 qtheta;
+        qtheta = od_pvq_compute_theta(j, ts);
+        k = od_pvq_compute_k(qcg, j, 0, n, beta);
+        items[idx].gain = i;
+        items[idx].theta = j;
+        items[idx].k = k;
+        items[idx].qcg = qcg;
+        items[idx].qtheta = qtheta;
+        items[idx].ts = ts;
+        idx++;
+        OD_ASSERT(idx < MAX_PVQ_ITEMS);
+      }
+    }
+    nitems = idx;
+    cos_dist = 0;
+    /* Sort PVQ search candidates in ascending order of pulses K so that
+       we can reuse all the previously searched pulses across searches. */
+    qsort(items, nitems, sizeof(items[0]),
+     (int (*)(const void *, const void *))items_compare);
+    /* Search for the best gain/theta in order. */
+    for (idx = 0; idx < nitems; idx++) {
+      int j;
+      od_val32 qcg;
+      int ts;
+      double cost;
+      double dist_theta;
+      double sin_prod;
+      od_val32 qtheta;
+      /* Quantized companded gain */
+      qcg = items[idx].qcg;
+      i = items[idx].gain;
+      j = items[idx].theta;
+      /* Set angular resolution (in ra) to match the encoded gain */
+      ts = items[idx].ts;
+      /* Search for the best angle within a reasonable range. */
+      qtheta = items[idx].qtheta;
+      k = items[idx].k;
+      /* Compute the minimal possible distortion by not taking the PVQ
+         cos_dist into account. */
+      dist_theta = 2 - 2.*od_pvq_cos(theta - qtheta)*OD_TRIG_SCALE_1;
+      dist = gain_weight*(qcg - cg)*(qcg - cg) + qcg*(double)cg*dist_theta;
+      dist *= OD_CGAIN_SCALE_2;
+      /* If we have no hope of beating skip (including a 1-bit worst-case
+         penalty), stop now. */
+      if (dist > dist0 + 1.0*pvq_norm_lambda && k != 0) continue;
+      sin_prod = od_pvq_sin(theta)*OD_TRIG_SCALE_1*od_pvq_sin(qtheta)*
+       OD_TRIG_SCALE_1;
+      /* PVQ search, using a gain of qcg*cg*sin(theta)*sin(qtheta) since
+         that's the factor by which cos_dist is multiplied to get the
+         distortion metric. */
+      if (k == 0) {
+        cos_dist = 0;
+        OD_CLEAR(y_tmp, n-1);
+      }
+      else if (k != prev_k) {
+        cos_dist = pvq_search_rdo_double(xr, n - 1, k, y_tmp,
+         qcg*(double)cg*sin_prod*OD_CGAIN_SCALE_2, pvq_norm_lambda, prev_k);
+      }
+      prev_k = k;
+      /* See Jmspeex' Journal of Dubious Theoretical Results. */
+      dist_theta = 2 - 2.*od_pvq_cos(theta - qtheta)*OD_TRIG_SCALE_1
+       + sin_prod*(2 - 2*cos_dist);
+      dist = gain_weight*(qcg - cg)*(qcg - cg) + qcg*(double)cg*dist_theta;
+      dist *= OD_CGAIN_SCALE_2;
+      /* Do approximate RDO. */
+      cost = dist + pvq_norm_lambda*od_pvq_rate(i, icgr, j, ts, adapt, y_tmp,
+       k, n, speed);
+      if (cost < best_cost) {
+        best_cost = cost;
+        best_dist = dist;
+        qg = i;
+        best_k = k;
+        best_qtheta = qtheta;
+        *itheta = j;
+        noref = 0;
+        OD_COPY(y, y_tmp, n - 1);
+      }
+    }
+  }
+  /* Don't bother with no-reference version if there's a reasonable
+     correlation. */
+  if (n <= OD_MAX_PVQ_SIZE && (corr < .5
+        || cg < (od_val32)(OD_SHL(2, OD_CGAIN_SHIFT)))) {
+    int gain_bound;
+    int prev_k;
+    gain_bound = OD_SHR(cg, OD_CGAIN_SHIFT);
+    prev_k = 0;
+    /* Search for the best gain (haven't determined reasonable range yet). */
+    for (i = OD_MAXI(1, gain_bound); i <= gain_bound + 1; i++) {
+      double cos_dist;
+      double cost;
+      od_val32 qcg;
+      qcg = OD_SHL(i, OD_CGAIN_SHIFT);
+      k = od_pvq_compute_k(qcg, -1, 1, n, beta);
+      /* Compute the minimal possible distortion by not taking the PVQ
+         cos_dist into account. */
+      dist = gain_weight*(qcg - cg)*(qcg - cg);
+      dist *= OD_CGAIN_SCALE_2;
+      if (dist > dist0 && k != 0) continue;
+      cos_dist = pvq_search_rdo_double(x16, n, k, y_tmp,
+       qcg*(double)cg*OD_CGAIN_SCALE_2, pvq_norm_lambda, prev_k);
+      prev_k = k;
+      /* See Jmspeex' Journal of Dubious Theoretical Results. */
+      dist = gain_weight*(qcg - cg)*(qcg - cg)
+       + qcg*(double)cg*(2 - 2*cos_dist);
+      dist *= OD_CGAIN_SCALE_2;
+      /* Do approximate RDO. */
+      cost = dist + pvq_norm_lambda*od_pvq_rate(i, 0, -1, 0, adapt, y_tmp, k,
+       n, speed);
+      if (cost <= best_cost) {
+        best_cost = cost;
+        best_dist = dist;
+        qg = i;
+        noref = 1;
+        best_k = k;
+        *itheta = -1;
+        OD_COPY(y, y_tmp, n);
+      }
+    }
+  }
+  k = best_k;
+  theta = best_qtheta;
+  skip = 0;
+  if (noref) {
+    if (qg == 0) skip = OD_PVQ_SKIP_ZERO;
+  }
+  else {
+    if (!is_keyframe && qg == 0) {
+      skip = (icgr ? OD_PVQ_SKIP_ZERO : OD_PVQ_SKIP_COPY);
+    }
+    if (qg == icgr && *itheta == 0 && !cfl_enabled) skip = OD_PVQ_SKIP_COPY;
+  }
+  /* Synthesize like the decoder would. */
+  if (skip) {
+    if (skip == OD_PVQ_SKIP_COPY) OD_COPY(out, r0, n);
+    else OD_CLEAR(out, n);
+  }
+  else {
+    if (noref) gain_offset = 0;
+    g = od_gain_expand(OD_SHL(qg, OD_CGAIN_SHIFT) + gain_offset, q0, beta);
+    od_pvq_synthesis_partial(out, y, r16, n, noref, g, theta, m, s,
+     qm_inv);
+  }
+  *vk = k;
+  *skip_diff += skip_dist - best_dist;
+  /* Encode gain differently depending on whether we use prediction or not.
+     Special encoding on inter frames where qg=0 is allowed for noref=0
+     but not noref=1.*/
+  if (is_keyframe) return noref ? qg : neg_interleave(qg, icgr);
+  else return noref ? qg - 1 : neg_interleave(qg + 1, icgr + 1);
+}
+
+/** Encodes a single vector of integers (eg, a partition within a
+ *  coefficient block) using PVQ
+ *
+ * @param [in,out] w          multi-symbol entropy encoder
+ * @param [in]     qg         quantized gain
+ * @param [in]     theta      quantized post-prediction theta
+ * @param [in]     in         coefficient vector to code
+ * @param [in]     n          number of coefficients in partition
+ * @param [in]     k          number of pulses in partition
+ * @param [in,out] model      entropy encoder state
+ * @param [in,out] adapt      adaptation context
+ * @param [in,out] exg        ExQ16 expectation of gain value
+ * @param [in,out] ext        ExQ16 expectation of theta value
+ * @param [in]     cdf_ctx    selects which cdf context to use
+ * @param [in]     is_keyframe whether we're encoding a keyframe
+ * @param [in]     code_skip  whether the "skip rest" flag is allowed
+ * @param [in]     skip_rest  when set, we skip all higher bands
+ * @param [in]     encode_flip whether we need to encode the CfL flip flag now
+ * @param [in]     flip       value of the CfL flip flag
+ */
+void pvq_encode_partition(aom_writer *w,
+                                 int qg,
+                                 int theta,
+                                 const od_coeff *in,
+                                 int n,
+                                 int k,
+                                 generic_encoder model[3],
+                                 od_adapt_ctx *adapt,
+                                 int *exg,
+                                 int *ext,
+                                 int cdf_ctx,
+                                 int is_keyframe,
+                                 int code_skip,
+                                 int skip_rest,
+                                 int encode_flip,
+                                 int flip) {
+  int noref;
+  int id;
+  noref = (theta == -1);
+  id = (qg > 0) + 2*OD_MINI(theta + 1,3) + 8*code_skip*skip_rest;
+  if (is_keyframe) {
+    OD_ASSERT(id != 8);
+    if (id >= 8) id--;
+  }
+  else {
+    OD_ASSERT(id != 10);
+    if (id >= 10) id--;
+  }
+  /* Jointly code gain, theta and noref for small values. Then we handle
+     larger gain and theta values. For noref, theta = -1. */
+  aom_write_symbol_pvq(w, id, &adapt->pvq.pvq_gaintheta_cdf[cdf_ctx][0],
+   8 + 7*code_skip);
+  if (encode_flip) {
+    /* We could eventually do some smarter entropy coding here, but it would
+       have to be good enough to overcome the overhead of the entropy coder.
+       An early attempt using a "toogle" flag with simple adaptation wasn't
+       worth the trouble. */
+    aom_write_bit(w, flip);
+  }
+  if (qg > 0) {
+    int tmp;
+    tmp = *exg;
+    generic_encode(w, &model[!noref], qg - 1, &tmp, 2);
+    OD_IIR_DIADIC(*exg, qg << 16, 2);
+  }
+  if (theta > 1) {
+    int tmp;
+    tmp = *ext;
+    generic_encode(w, &model[2], theta - 2, &tmp, 2);
+    OD_IIR_DIADIC(*ext, theta << 16, 2);
+  }
+  aom_encode_pvq_codeword(w, &adapt->pvq.pvq_codeword_ctx, in,
+   n - (theta != -1), k);
+}
+
+/** Quantizes a scalar with rate-distortion optimization (RDO)
+ * @param [in] x      unquantized value
+ * @param [in] q      quantization step size
+ * @param [in] delta0 rate increase for encoding a 1 instead of a 0
+ * @param [in] pvq_norm_lambda enc->pvq_norm_lambda for quantized RDO
+ * @retval quantized value
+ */
+int od_rdo_quant(od_coeff x, int q, double delta0, double pvq_norm_lambda) {
+  int n;
+  /* Optimal quantization threshold is 1/2 + lambda*delta_rate/2. See
+     Jmspeex' Journal of Dubious Theoretical Results for details. */
+  n = OD_DIV_R0(abs(x), q);
+  if ((double)abs(x)/q < (double)n/2 + pvq_norm_lambda*delta0/(2*n)) {
+    return 0;
+  }
+  else {
+    return OD_DIV_R0(x, q);
+  }
+}
+
+/** Encode a coefficient block (excepting DC) using PVQ
+ *
+ * @param [in,out] enc     daala encoder context
+ * @param [in]     ref     'reference' (prediction) vector
+ * @param [in]     in      coefficient block to quantize and encode
+ * @param [out]    out     quantized coefficient block
+ * @param [in]     q0      scale/quantizer
+ * @param [in]     pli     plane index
+ * @param [in]     bs      log of the block size minus two
+ * @param [in]     beta    per-band activity masking beta param
+ * @param [in]     is_keyframe whether we're encoding a keyframe
+ * @param [in]     qm      QM with magnitude compensation
+ * @param [in]     qm_inv  Inverse of QM with magnitude compensation
+ * @param [in]     speed   Make search faster by making approximations
+ * @param [in]     pvq_info If null, conisdered as RDO search mode
+ * @return         Returns block skip info indicating whether DC/AC are coded.
+ *                 bit0: DC is coded, bit1: AC is coded (1 means coded)
+ *
+ */
+PVQ_SKIP_TYPE od_pvq_encode(daala_enc_ctx *enc,
+                   od_coeff *ref,
+                   const od_coeff *in,
+                   od_coeff *out,
+                   int q_dc,
+                   int q_ac,
+                   int pli,
+                   int bs,
+                   const od_val16 *beta,
+                   int is_keyframe,
+                   const int16_t *qm,
+                   const int16_t *qm_inv,
+                   int speed,
+                   PVQ_INFO *pvq_info){
+  int theta[PVQ_MAX_PARTITIONS];
+  int qg[PVQ_MAX_PARTITIONS];
+  int k[PVQ_MAX_PARTITIONS];
+  od_coeff y[OD_TXSIZE_MAX*OD_TXSIZE_MAX];
+  int *exg;
+  int *ext;
+  int nb_bands;
+  int i;
+  const int *off;
+  int size[PVQ_MAX_PARTITIONS];
+  generic_encoder *model;
+  double skip_diff;
+  int tell;
+  uint16_t *skip_cdf;
+  od_rollback_buffer buf;
+  int dc_quant;
+  int flip;
+  int cfl_encoded;
+  int skip_rest;
+  int skip_dir;
+  int skip_theta_value;
+  const unsigned char *pvq_qm;
+  double dc_rate;
+  int use_masking;
+  PVQ_SKIP_TYPE ac_dc_coded;
+
+  aom_clear_system_state();
+
+  use_masking = enc->use_activity_masking;
+
+  if (use_masking)
+    pvq_qm = &enc->state.pvq_qm_q4[pli][0];
+  else
+    pvq_qm = 0;
+
+  exg = &enc->state.adapt->pvq.pvq_exg[pli][bs][0];
+  ext = enc->state.adapt->pvq.pvq_ext + bs*PVQ_MAX_PARTITIONS;
+  skip_cdf = enc->state.adapt->skip_cdf[2*bs + (pli != 0)];
+  model = enc->state.adapt->pvq.pvq_param_model;
+  nb_bands = OD_BAND_OFFSETS[bs][0];
+  off = &OD_BAND_OFFSETS[bs][1];
+
+  if (use_masking)
+    dc_quant = OD_MAXI(1, q_dc * pvq_qm[od_qm_get_index(bs, 0)] >> 4);
+  else
+    dc_quant = OD_MAXI(1, q_dc);
+
+  tell = 0;
+  for (i = 0; i < nb_bands; i++) size[i] = off[i+1] - off[i];
+  skip_diff = 0;
+  flip = 0;
+  /*If we are coding a chroma block of a keyframe, we are doing CfL.*/
+  if (pli != 0 && is_keyframe) {
+    od_val32 xy;
+    xy = 0;
+    /*Compute the dot-product of the first band of chroma with the luma ref.*/
+    for (i = off[0]; i < off[1]; i++) {
+#if defined(OD_FLOAT_PVQ)
+      xy += ref[i]*(double)qm[i]*OD_QM_SCALE_1*
+       (double)in[i]*(double)qm[i]*OD_QM_SCALE_1;
+#else
+      od_val32 rq;
+      od_val32 inq;
+      rq = ref[i]*qm[i];
+      inq = in[i]*qm[i];
+      xy += OD_SHR(rq*(int64_t)inq, OD_SHL(OD_QM_SHIFT + OD_CFL_FLIP_SHIFT,
+       1));
+#endif
+    }
+    /*If cos(theta) < 0, then |theta| > pi/2 and we should negate the ref.*/
+    if (xy < 0) {
+      flip = 1;
+      for(i = off[0]; i < off[nb_bands]; i++) ref[i] = -ref[i];
+    }
+  }
+  for (i = 0; i < nb_bands; i++) {
+    int q;
+
+    if (use_masking)
+      q = OD_MAXI(1, q_ac * pvq_qm[od_qm_get_index(bs, i + 1)] >> 4);
+    else
+      q = OD_MAXI(1, q_ac);
+
+    qg[i] = pvq_theta(out + off[i], in + off[i], ref + off[i], size[i],
+     q, y + off[i], &theta[i], &k[i], beta[i], &skip_diff, is_keyframe,
+     pli, enc->state.adapt, qm + off[i], qm_inv + off[i],
+     enc->pvq_norm_lambda, speed);
+  }
+  od_encode_checkpoint(enc, &buf);
+  if (is_keyframe) out[0] = 0;
+  else {
+    int n;
+    n = OD_DIV_R0(abs(in[0] - ref[0]), dc_quant);
+    if (n == 0) {
+      out[0] = 0;
+    } else {
+      int tell2;
+      od_rollback_buffer dc_buf;
+
+      dc_rate = -OD_LOG2((double)(skip_cdf[3] - skip_cdf[2])/
+       (double)(skip_cdf[2] - skip_cdf[1]));
+      dc_rate += 1;
+
+#if CONFIG_DAALA_EC
+      tell2 = od_ec_enc_tell_frac(&enc->w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+      od_encode_checkpoint(enc, &dc_buf);
+      generic_encode(&enc->w, &enc->state.adapt->model_dc[pli],
+       n - 1, &enc->state.adapt->ex_dc[pli][bs][0], 2);
+#if CONFIG_DAALA_EC
+      tell2 = od_ec_enc_tell_frac(&enc->w.ec) - tell2;
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+      dc_rate += tell2/8.0;
+      od_encode_rollback(enc, &dc_buf);
+
+      out[0] = od_rdo_quant(in[0] - ref[0], dc_quant, dc_rate,
+       enc->pvq_norm_lambda);
+    }
+  }
+#if CONFIG_DAALA_EC
+  tell = od_ec_enc_tell_frac(&enc->w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  /* Code as if we're not skipping. */
+  aom_write_symbol(&enc->w, 2 + (out[0] != 0), skip_cdf, 4);
+  ac_dc_coded = AC_CODED + (out[0] != 0);
+  cfl_encoded = 0;
+  skip_rest = 1;
+  skip_theta_value = is_keyframe ? -1 : 0;
+  for (i = 1; i < nb_bands; i++) {
+    if (theta[i] != skip_theta_value || qg[i]) skip_rest = 0;
+  }
+  skip_dir = 0;
+  if (nb_bands > 1) {
+    for (i = 0; i < 3; i++) {
+      int j;
+      int tmp;
+      tmp = 1;
+      // ToDo(yaowu): figure out better stop condition without gcc warning.
+      for (j = i + 1; j < nb_bands && j < PVQ_MAX_PARTITIONS; j += 3) {
+        if (theta[j] != skip_theta_value || qg[j]) tmp = 0;
+      }
+      skip_dir |= tmp << i;
+    }
+  }
+  if (theta[0] == skip_theta_value && qg[0] == 0 && skip_rest) nb_bands = 0;
+
+  /* NOTE: There was no other better place to put this function. */
+  if (pvq_info)
+    av1_store_pvq_enc_info(pvq_info, qg, theta, k, y, nb_bands, off, size,
+      skip_rest, skip_dir, bs);
+
+  for (i = 0; i < nb_bands; i++) {
+    int encode_flip;
+    /* Encode CFL flip bit just after the first time it's used. */
+    encode_flip = pli != 0 && is_keyframe && theta[i] != -1 && !cfl_encoded;
+    if (i == 0 || (!skip_rest && !(skip_dir & (1 << ((i - 1)%3))))) {
+      pvq_encode_partition(&enc->w, qg[i], theta[i], y + off[i],
+       size[i], k[i], model, enc->state.adapt, exg + i, ext + i,
+       (pli != 0)*OD_TXSIZES*PVQ_MAX_PARTITIONS + bs*PVQ_MAX_PARTITIONS + i,
+       is_keyframe, i == 0 && (i < nb_bands - 1), skip_rest, encode_flip, flip);
+    }
+    if (i == 0 && !skip_rest && bs > 0) {
+      aom_write_symbol(&enc->w, skip_dir,
+       &enc->state.adapt->pvq.pvq_skip_dir_cdf[(pli != 0) + 2*(bs - 1)][0], 7);
+    }
+    if (encode_flip) cfl_encoded = 1;
+  }
+#if CONFIG_DAALA_EC
+  tell = od_ec_enc_tell_frac(&enc->w.ec) - tell;
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+  /* Account for the rate of skipping the AC, based on the same DC decision
+     we made when trying to not skip AC. */
+  {
+    double skip_rate;
+    if (out[0] != 0) {
+      skip_rate = -OD_LOG2((skip_cdf[1] - skip_cdf[0])/
+     (double)skip_cdf[3]);
+    }
+    else {
+      skip_rate = -OD_LOG2(skip_cdf[0]/
+     (double)skip_cdf[3]);
+    }
+    tell -= (int)floor(.5+8*skip_rate);
+  }
+  if (nb_bands == 0 || skip_diff <= enc->pvq_norm_lambda/8*tell) {
+    if (is_keyframe) out[0] = 0;
+    else {
+      int n;
+      n = OD_DIV_R0(abs(in[0] - ref[0]), dc_quant);
+      if (n == 0) {
+        out[0] = 0;
+      } else {
+        int tell2;
+        od_rollback_buffer dc_buf;
+
+        dc_rate = -OD_LOG2((double)(skip_cdf[1] - skip_cdf[0])/
+         (double)skip_cdf[0]);
+        dc_rate += 1;
+
+#if CONFIG_DAALA_EC
+        tell2 = od_ec_enc_tell_frac(&enc->w.ec);
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+        od_encode_checkpoint(enc, &dc_buf);
+        generic_encode(&enc->w, &enc->state.adapt->model_dc[pli],
+         n - 1, &enc->state.adapt->ex_dc[pli][bs][0], 2);
+#if CONFIG_DAALA_EC
+        tell2 = od_ec_enc_tell_frac(&enc->w.ec) - tell2;
+#else
+#error "CONFIG_PVQ currently requires CONFIG_DAALA_EC."
+#endif
+        dc_rate += tell2/8.0;
+        od_encode_rollback(enc, &dc_buf);
+
+        out[0] = od_rdo_quant(in[0] - ref[0], dc_quant, dc_rate,
+         enc->pvq_norm_lambda);
+      }
+    }
+    /* We decide to skip, roll back everything as it was before. */
+    od_encode_rollback(enc, &buf);
+    aom_write_symbol(&enc->w, out[0] != 0, skip_cdf, 4);
+    ac_dc_coded = (out[0] != 0);
+    if (is_keyframe) for (i = 1; i < 1 << (2*bs + 4); i++) out[i] = 0;
+    else for (i = 1; i < 1 << (2*bs + 4); i++) out[i] = ref[i];
+  }
+  if (pvq_info)
+    pvq_info->ac_dc_coded = ac_dc_coded;
+  return ac_dc_coded;
+}
diff --git a/third_party/aom/av1/encoder/pvq_encoder.h b/third_party/aom/av1/encoder/pvq_encoder.h
new file mode 100644
index 0000000000..b84c8961b7
--- /dev/null
+++ b/third_party/aom/av1/encoder/pvq_encoder.h
@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2001-2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* clang-format off */
+
+#if !defined(_pvq_encoder_H)
+# define _pvq_encoder_H (1)
+# include "aom_dsp/bitwriter.h"
+# include "aom_dsp/entenc.h"
+# include "av1/common/blockd.h"
+# include "av1/common/pvq.h"
+# include "av1/encoder/encint.h"
+
+void aom_write_symbol_pvq(aom_writer *w, int symb, aom_cdf_prob *cdf,
+    int nsymbs);
+
+void aom_encode_band_pvq_splits(aom_writer *w, od_pvq_codeword_ctx *adapt,
+ const int *y, int n, int k, int level);
+
+void aom_laplace_encode_special(aom_writer *w, int x, unsigned decay);
+
+void pvq_encode_partition(aom_writer *w,
+                                 int qg,
+                                 int theta,
+                                 const od_coeff *in,
+                                 int n,
+                                 int k,
+                                 generic_encoder model[3],
+                                 od_adapt_ctx *adapt,
+                                 int *exg,
+                                 int *ext,
+                                 int cdf_ctx,
+                                 int is_keyframe,
+                                 int code_skip,
+                                 int skip_rest,
+                                 int encode_flip,
+                                 int flip);
+
+PVQ_SKIP_TYPE od_pvq_encode(daala_enc_ctx *enc, od_coeff *ref,
+    const od_coeff *in, od_coeff *out, int q_dc, int q_ac, int pli, int bs,
+    const od_val16 *beta, int is_keyframe,
+    const int16_t *qm, const int16_t *qm_inv, int speed,
+    PVQ_INFO *pvq_info);
+
+#endif
diff --git a/third_party/aom/av1/encoder/ransac.c b/third_party/aom/av1/encoder/ransac.c
new file mode 100644
index 0000000000..5d5dd75721
--- /dev/null
+++ b/third_party/aom/av1/encoder/ransac.c
@@ -0,0 +1,1210 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#define _POSIX_C_SOURCE 200112L  // rand_r()
+#include <memory.h>
+#include <math.h>
+#include <time.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
+
+#include "av1/encoder/ransac.h"
+
+#define MAX_MINPTS 4
+#define MAX_DEGENERATE_ITER 10
+#define MINPTS_MULTIPLIER 5
+
+#define INLIER_THRESHOLD 1.0
+#define MIN_TRIALS 20
+
+////////////////////////////////////////////////////////////////////////////////
+// ransac
+typedef int (*IsDegenerateFunc)(double *p);
+typedef void (*NormalizeFunc)(double *p, int np, double *T);
+typedef void (*DenormalizeFunc)(double *params, double *T1, double *T2);
+typedef int (*FindTransformationFunc)(int points, double *points1,
+                                      double *points2, double *params);
+typedef void (*ProjectPointsDoubleFunc)(double *mat, double *points,
+                                        double *proj, const int n,
+                                        const int stride_points,
+                                        const int stride_proj);
+
+static void project_points_double_translation(double *mat, double *points,
+                                              double *proj, const int n,
+                                              const int stride_points,
+                                              const int stride_proj) {
+  int i;
+  for (i = 0; i < n; ++i) {
+    const double x = *(points++), y = *(points++);
+    *(proj++) = x + mat[0];
+    *(proj++) = y + mat[1];
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+static void project_points_double_rotzoom(double *mat, double *points,
+                                          double *proj, const int n,
+                                          const int stride_points,
+                                          const int stride_proj) {
+  int i;
+  for (i = 0; i < n; ++i) {
+    const double x = *(points++), y = *(points++);
+    *(proj++) = mat[2] * x + mat[3] * y + mat[0];
+    *(proj++) = -mat[3] * x + mat[2] * y + mat[1];
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+static void project_points_double_affine(double *mat, double *points,
+                                         double *proj, const int n,
+                                         const int stride_points,
+                                         const int stride_proj) {
+  int i;
+  for (i = 0; i < n; ++i) {
+    const double x = *(points++), y = *(points++);
+    *(proj++) = mat[2] * x + mat[3] * y + mat[0];
+    *(proj++) = mat[4] * x + mat[5] * y + mat[1];
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+static void project_points_double_hortrapezoid(double *mat, double *points,
+                                               double *proj, const int n,
+                                               const int stride_points,
+                                               const int stride_proj) {
+  int i;
+  double x, y, Z, Z_inv;
+  for (i = 0; i < n; ++i) {
+    x = *(points++), y = *(points++);
+    Z_inv = mat[7] * y + 1;
+    assert(fabs(Z_inv) > 0.000001);
+    Z = 1. / Z_inv;
+    *(proj++) = (mat[2] * x + mat[3] * y + mat[0]) * Z;
+    *(proj++) = (mat[5] * y + mat[1]) * Z;
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+static void project_points_double_vertrapezoid(double *mat, double *points,
+                                               double *proj, const int n,
+                                               const int stride_points,
+                                               const int stride_proj) {
+  int i;
+  double x, y, Z, Z_inv;
+  for (i = 0; i < n; ++i) {
+    x = *(points++), y = *(points++);
+    Z_inv = mat[6] * x + 1;
+    assert(fabs(Z_inv) > 0.000001);
+    Z = 1. / Z_inv;
+    *(proj++) = (mat[2] * x + mat[0]) * Z;
+    *(proj++) = (mat[4] * x + mat[5] * y + mat[1]) * Z;
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+static void project_points_double_homography(double *mat, double *points,
+                                             double *proj, const int n,
+                                             const int stride_points,
+                                             const int stride_proj) {
+  int i;
+  double x, y, Z, Z_inv;
+  for (i = 0; i < n; ++i) {
+    x = *(points++), y = *(points++);
+    Z_inv = mat[6] * x + mat[7] * y + 1;
+    assert(fabs(Z_inv) > 0.000001);
+    Z = 1. / Z_inv;
+    *(proj++) = (mat[2] * x + mat[3] * y + mat[0]) * Z;
+    *(proj++) = (mat[4] * x + mat[5] * y + mat[1]) * Z;
+    points += stride_points - 2;
+    proj += stride_proj - 2;
+  }
+}
+
+///////////////////////////////////////////////////////////////////////////////
+// svdcmp
+// Adopted from Numerical Recipes in C
+
+static const double TINY_NEAR_ZERO = 1.0E-12;
+
+static INLINE double sign(double a, double b) {
+  return ((b) >= 0 ? fabs(a) : -fabs(a));
+}
+
+static INLINE double pythag(double a, double b) {
+  double ct;
+  const double absa = fabs(a);
+  const double absb = fabs(b);
+
+  if (absa > absb) {
+    ct = absb / absa;
+    return absa * sqrt(1.0 + ct * ct);
+  } else {
+    ct = absa / absb;
+    return (absb == 0) ? 0 : absb * sqrt(1.0 + ct * ct);
+  }
+}
+
+static void multiply_mat(const double *m1, const double *m2, double *res,
+                         const int m1_rows, const int inner_dim,
+                         const int m2_cols) {
+  double sum;
+
+  int row, col, inner;
+  for (row = 0; row < m1_rows; ++row) {
+    for (col = 0; col < m2_cols; ++col) {
+      sum = 0;
+      for (inner = 0; inner < inner_dim; ++inner)
+        sum += m1[row * inner_dim + inner] * m2[inner * m2_cols + col];
+      *(res++) = sum;
+    }
+  }
+}
+
+static int svdcmp(double **u, int m, int n, double w[], double **v) {
+  const int max_its = 30;
+  int flag, i, its, j, jj, k, l, nm;
+  double anorm, c, f, g, h, s, scale, x, y, z;
+  double *rv1 = (double *)aom_malloc(sizeof(*rv1) * (n + 1));
+  g = scale = anorm = 0.0;
+  for (i = 0; i < n; i++) {
+    l = i + 1;
+    rv1[i] = scale * g;
+    g = s = scale = 0.0;
+    if (i < m) {
+      for (k = i; k < m; k++) scale += fabs(u[k][i]);
+      if (scale != 0.) {
+        for (k = i; k < m; k++) {
+          u[k][i] /= scale;
+          s += u[k][i] * u[k][i];
+        }
+        f = u[i][i];
+        g = -sign(sqrt(s), f);
+        h = f * g - s;
+        u[i][i] = f - g;
+        for (j = l; j < n; j++) {
+          for (s = 0.0, k = i; k < m; k++) s += u[k][i] * u[k][j];
+          f = s / h;
+          for (k = i; k < m; k++) u[k][j] += f * u[k][i];
+        }
+        for (k = i; k < m; k++) u[k][i] *= scale;
+      }
+    }
+    w[i] = scale * g;
+    g = s = scale = 0.0;
+    if (i < m && i != n - 1) {
+      for (k = l; k < n; k++) scale += fabs(u[i][k]);
+      if (scale != 0.) {
+        for (k = l; k < n; k++) {
+          u[i][k] /= scale;
+          s += u[i][k] * u[i][k];
+        }
+        f = u[i][l];
+        g = -sign(sqrt(s), f);
+        h = f * g - s;
+        u[i][l] = f - g;
+        for (k = l; k < n; k++) rv1[k] = u[i][k] / h;
+        for (j = l; j < m; j++) {
+          for (s = 0.0, k = l; k < n; k++) s += u[j][k] * u[i][k];
+          for (k = l; k < n; k++) u[j][k] += s * rv1[k];
+        }
+        for (k = l; k < n; k++) u[i][k] *= scale;
+      }
+    }
+    anorm = fmax(anorm, (fabs(w[i]) + fabs(rv1[i])));
+  }
+
+  for (i = n - 1; i >= 0; i--) {
+    if (i < n - 1) {
+      if (g != 0.) {
+        for (j = l; j < n; j++) v[j][i] = (u[i][j] / u[i][l]) / g;
+        for (j = l; j < n; j++) {
+          for (s = 0.0, k = l; k < n; k++) s += u[i][k] * v[k][j];
+          for (k = l; k < n; k++) v[k][j] += s * v[k][i];
+        }
+      }
+      for (j = l; j < n; j++) v[i][j] = v[j][i] = 0.0;
+    }
+    v[i][i] = 1.0;
+    g = rv1[i];
+    l = i;
+  }
+  for (i = AOMMIN(m, n) - 1; i >= 0; i--) {
+    l = i + 1;
+    g = w[i];
+    for (j = l; j < n; j++) u[i][j] = 0.0;
+    if (g != 0.) {
+      g = 1.0 / g;
+      for (j = l; j < n; j++) {
+        for (s = 0.0, k = l; k < m; k++) s += u[k][i] * u[k][j];
+        f = (s / u[i][i]) * g;
+        for (k = i; k < m; k++) u[k][j] += f * u[k][i];
+      }
+      for (j = i; j < m; j++) u[j][i] *= g;
+    } else {
+      for (j = i; j < m; j++) u[j][i] = 0.0;
+    }
+    ++u[i][i];
+  }
+  for (k = n - 1; k >= 0; k--) {
+    for (its = 0; its < max_its; its++) {
+      flag = 1;
+      for (l = k; l >= 0; l--) {
+        nm = l - 1;
+        if ((double)(fabs(rv1[l]) + anorm) == anorm || nm < 0) {
+          flag = 0;
+          break;
+        }
+        if ((double)(fabs(w[nm]) + anorm) == anorm) break;
+      }
+      if (flag) {
+        c = 0.0;
+        s = 1.0;
+        for (i = l; i <= k; i++) {
+          f = s * rv1[i];
+          rv1[i] = c * rv1[i];
+          if ((double)(fabs(f) + anorm) == anorm) break;
+          g = w[i];
+          h = pythag(f, g);
+          w[i] = h;
+          h = 1.0 / h;
+          c = g * h;
+          s = -f * h;
+          for (j = 0; j < m; j++) {
+            y = u[j][nm];
+            z = u[j][i];
+            u[j][nm] = y * c + z * s;
+            u[j][i] = z * c - y * s;
+          }
+        }
+      }
+      z = w[k];
+      if (l == k) {
+        if (z < 0.0) {
+          w[k] = -z;
+          for (j = 0; j < n; j++) v[j][k] = -v[j][k];
+        }
+        break;
+      }
+      if (its == max_its - 1) {
+        aom_free(rv1);
+        return 1;
+      }
+      assert(k > 0);
+      x = w[l];
+      nm = k - 1;
+      y = w[nm];
+      g = rv1[nm];
+      h = rv1[k];
+      f = ((y - z) * (y + z) + (g - h) * (g + h)) / (2.0 * h * y);
+      g = pythag(f, 1.0);
+      f = ((x - z) * (x + z) + h * ((y / (f + sign(g, f))) - h)) / x;
+      c = s = 1.0;
+      for (j = l; j <= nm; j++) {
+        i = j + 1;
+        g = rv1[i];
+        y = w[i];
+        h = s * g;
+        g = c * g;
+        z = pythag(f, h);
+        rv1[j] = z;
+        c = f / z;
+        s = h / z;
+        f = x * c + g * s;
+        g = g * c - x * s;
+        h = y * s;
+        y *= c;
+        for (jj = 0; jj < n; jj++) {
+          x = v[jj][j];
+          z = v[jj][i];
+          v[jj][j] = x * c + z * s;
+          v[jj][i] = z * c - x * s;
+        }
+        z = pythag(f, h);
+        w[j] = z;
+        if (z != 0.) {
+          z = 1.0 / z;
+          c = f * z;
+          s = h * z;
+        }
+        f = c * g + s * y;
+        x = c * y - s * g;
+        for (jj = 0; jj < m; jj++) {
+          y = u[jj][j];
+          z = u[jj][i];
+          u[jj][j] = y * c + z * s;
+          u[jj][i] = z * c - y * s;
+        }
+      }
+      rv1[l] = 0.0;
+      rv1[k] = f;
+      w[k] = x;
+    }
+  }
+  aom_free(rv1);
+  return 0;
+}
+
+static int SVD(double *U, double *W, double *V, double *matx, int M, int N) {
+  // Assumes allocation for U is MxN
+  double **nrU = (double **)aom_malloc((M) * sizeof(*nrU));
+  double **nrV = (double **)aom_malloc((N) * sizeof(*nrV));
+  int problem, i;
+
+  problem = !(nrU && nrV);
+  if (!problem) {
+    for (i = 0; i < M; i++) {
+      nrU[i] = &U[i * N];
+    }
+    for (i = 0; i < N; i++) {
+      nrV[i] = &V[i * N];
+    }
+  } else {
+    if (nrU) aom_free(nrU);
+    if (nrV) aom_free(nrV);
+    return 1;
+  }
+
+  /* copy from given matx into nrU */
+  for (i = 0; i < M; i++) {
+    memcpy(&(nrU[i][0]), matx + N * i, N * sizeof(*matx));
+  }
+
+  /* HERE IT IS: do SVD */
+  if (svdcmp(nrU, M, N, W, nrV)) {
+    aom_free(nrU);
+    aom_free(nrV);
+    return 1;
+  }
+
+  /* aom_free Numerical Recipes arrays */
+  aom_free(nrU);
+  aom_free(nrV);
+
+  return 0;
+}
+
+int pseudo_inverse(double *inv, double *matx, const int M, const int N) {
+  double ans;
+  int i, j, k;
+  double *const U = (double *)aom_malloc(M * N * sizeof(*matx));
+  double *const W = (double *)aom_malloc(N * sizeof(*matx));
+  double *const V = (double *)aom_malloc(N * N * sizeof(*matx));
+
+  if (!(U && W && V)) {
+    return 1;
+  }
+  if (SVD(U, W, V, matx, M, N)) {
+    aom_free(U);
+    aom_free(W);
+    aom_free(V);
+    return 1;
+  }
+  for (i = 0; i < N; i++) {
+    if (fabs(W[i]) < TINY_NEAR_ZERO) {
+      aom_free(U);
+      aom_free(W);
+      aom_free(V);
+      return 1;
+    }
+  }
+
+  for (i = 0; i < N; i++) {
+    for (j = 0; j < M; j++) {
+      ans = 0;
+      for (k = 0; k < N; k++) {
+        ans += V[k + N * i] * U[k + N * j] / W[k];
+      }
+      inv[j + M * i] = ans;
+    }
+  }
+  aom_free(U);
+  aom_free(W);
+  aom_free(V);
+  return 0;
+}
+
+static void normalize_homography(double *pts, int n, double *T) {
+  double *p = pts;
+  double mean[2] = { 0, 0 };
+  double msqe = 0;
+  double scale;
+  int i;
+  for (i = 0; i < n; ++i, p += 2) {
+    mean[0] += p[0];
+    mean[1] += p[1];
+  }
+  mean[0] /= n;
+  mean[1] /= n;
+  for (p = pts, i = 0; i < n; ++i, p += 2) {
+    p[0] -= mean[0];
+    p[1] -= mean[1];
+    msqe += sqrt(p[0] * p[0] + p[1] * p[1]);
+  }
+  msqe /= n;
+  scale = (msqe == 0 ? 1.0 : sqrt(2) / msqe);
+  T[0] = scale;
+  T[1] = 0;
+  T[2] = -scale * mean[0];
+  T[3] = 0;
+  T[4] = scale;
+  T[5] = -scale * mean[1];
+  T[6] = 0;
+  T[7] = 0;
+  T[8] = 1;
+  for (p = pts, i = 0; i < n; ++i, p += 2) {
+    p[0] *= scale;
+    p[1] *= scale;
+  }
+}
+
+static void invnormalize_mat(double *T, double *iT) {
+  double is = 1.0 / T[0];
+  double m0 = -T[2] * is;
+  double m1 = -T[5] * is;
+  iT[0] = is;
+  iT[1] = 0;
+  iT[2] = m0;
+  iT[3] = 0;
+  iT[4] = is;
+  iT[5] = m1;
+  iT[6] = 0;
+  iT[7] = 0;
+  iT[8] = 1;
+}
+
+static void denormalize_homography(double *params, double *T1, double *T2) {
+  double iT2[9];
+  double params2[9];
+  invnormalize_mat(T2, iT2);
+  multiply_mat(params, T1, params2, 3, 3, 3);
+  multiply_mat(iT2, params2, params, 3, 3, 3);
+}
+
+static void denormalize_homography_reorder(double *params, double *T1,
+                                           double *T2) {
+  double params_denorm[MAX_PARAMDIM];
+  memcpy(params_denorm, params, sizeof(*params) * 8);
+  params_denorm[8] = 1.0;
+  denormalize_homography(params_denorm, T1, T2);
+  params[0] = params_denorm[2];
+  params[1] = params_denorm[5];
+  params[2] = params_denorm[0];
+  params[3] = params_denorm[1];
+  params[4] = params_denorm[3];
+  params[5] = params_denorm[4];
+  params[6] = params_denorm[6];
+  params[7] = params_denorm[7];
+}
+
+static void denormalize_affine_reorder(double *params, double *T1, double *T2) {
+  double params_denorm[MAX_PARAMDIM];
+  params_denorm[0] = params[0];
+  params_denorm[1] = params[1];
+  params_denorm[2] = params[4];
+  params_denorm[3] = params[2];
+  params_denorm[4] = params[3];
+  params_denorm[5] = params[5];
+  params_denorm[6] = params_denorm[7] = 0;
+  params_denorm[8] = 1;
+  denormalize_homography(params_denorm, T1, T2);
+  params[0] = params_denorm[2];
+  params[1] = params_denorm[5];
+  params[2] = params_denorm[0];
+  params[3] = params_denorm[1];
+  params[4] = params_denorm[3];
+  params[5] = params_denorm[4];
+  params[6] = params[7] = 0;
+}
+
+static void denormalize_rotzoom_reorder(double *params, double *T1,
+                                        double *T2) {
+  double params_denorm[MAX_PARAMDIM];
+  params_denorm[0] = params[0];
+  params_denorm[1] = params[1];
+  params_denorm[2] = params[2];
+  params_denorm[3] = -params[1];
+  params_denorm[4] = params[0];
+  params_denorm[5] = params[3];
+  params_denorm[6] = params_denorm[7] = 0;
+  params_denorm[8] = 1;
+  denormalize_homography(params_denorm, T1, T2);
+  params[0] = params_denorm[2];
+  params[1] = params_denorm[5];
+  params[2] = params_denorm[0];
+  params[3] = params_denorm[1];
+  params[4] = -params[3];
+  params[5] = params[2];
+  params[6] = params[7] = 0;
+}
+
+static void denormalize_translation_reorder(double *params, double *T1,
+                                            double *T2) {
+  double params_denorm[MAX_PARAMDIM];
+  params_denorm[0] = 1;
+  params_denorm[1] = 0;
+  params_denorm[2] = params[0];
+  params_denorm[3] = 0;
+  params_denorm[4] = 1;
+  params_denorm[5] = params[1];
+  params_denorm[6] = params_denorm[7] = 0;
+  params_denorm[8] = 1;
+  denormalize_homography(params_denorm, T1, T2);
+  params[0] = params_denorm[2];
+  params[1] = params_denorm[5];
+  params[2] = params[5] = 1;
+  params[3] = params[4] = 0;
+  params[6] = params[7] = 0;
+}
+
+static int find_translation(int np, double *pts1, double *pts2, double *mat) {
+  int i;
+  double sx, sy, dx, dy;
+  double sumx, sumy;
+
+  double T1[9], T2[9];
+  normalize_homography(pts1, np, T1);
+  normalize_homography(pts2, np, T2);
+
+  sumx = 0;
+  sumy = 0;
+  for (i = 0; i < np; ++i) {
+    dx = *(pts2++);
+    dy = *(pts2++);
+    sx = *(pts1++);
+    sy = *(pts1++);
+
+    sumx += dx - sx;
+    sumy += dy - sy;
+  }
+  mat[0] = sumx / np;
+  mat[1] = sumy / np;
+  denormalize_translation_reorder(mat, T1, T2);
+  return 0;
+}
+
+static int find_rotzoom(int np, double *pts1, double *pts2, double *mat) {
+  const int np2 = np * 2;
+  double *a = (double *)aom_malloc(sizeof(*a) * np2 * 9);
+  double *b = a + np2 * 4;
+  double *temp = b + np2;
+  int i;
+  double sx, sy, dx, dy;
+
+  double T1[9], T2[9];
+  normalize_homography(pts1, np, T1);
+  normalize_homography(pts2, np, T2);
+
+  for (i = 0; i < np; ++i) {
+    dx = *(pts2++);
+    dy = *(pts2++);
+    sx = *(pts1++);
+    sy = *(pts1++);
+
+    a[i * 2 * 4 + 0] = sx;
+    a[i * 2 * 4 + 1] = sy;
+    a[i * 2 * 4 + 2] = 1;
+    a[i * 2 * 4 + 3] = 0;
+    a[(i * 2 + 1) * 4 + 0] = sy;
+    a[(i * 2 + 1) * 4 + 1] = -sx;
+    a[(i * 2 + 1) * 4 + 2] = 0;
+    a[(i * 2 + 1) * 4 + 3] = 1;
+
+    b[2 * i] = dx;
+    b[2 * i + 1] = dy;
+  }
+  if (pseudo_inverse(temp, a, np2, 4)) {
+    aom_free(a);
+    return 1;
+  }
+  multiply_mat(temp, b, mat, 4, np2, 1);
+  denormalize_rotzoom_reorder(mat, T1, T2);
+  aom_free(a);
+  return 0;
+}
+
+static int find_affine(int np, double *pts1, double *pts2, double *mat) {
+  const int np2 = np * 2;
+  double *a = (double *)aom_malloc(sizeof(*a) * np2 * 13);
+  double *b = a + np2 * 6;
+  double *temp = b + np2;
+  int i;
+  double sx, sy, dx, dy;
+
+  double T1[9], T2[9];
+  normalize_homography(pts1, np, T1);
+  normalize_homography(pts2, np, T2);
+
+  for (i = 0; i < np; ++i) {
+    dx = *(pts2++);
+    dy = *(pts2++);
+    sx = *(pts1++);
+    sy = *(pts1++);
+
+    a[i * 2 * 6 + 0] = sx;
+    a[i * 2 * 6 + 1] = sy;
+    a[i * 2 * 6 + 2] = 0;
+    a[i * 2 * 6 + 3] = 0;
+    a[i * 2 * 6 + 4] = 1;
+    a[i * 2 * 6 + 5] = 0;
+    a[(i * 2 + 1) * 6 + 0] = 0;
+    a[(i * 2 + 1) * 6 + 1] = 0;
+    a[(i * 2 + 1) * 6 + 2] = sx;
+    a[(i * 2 + 1) * 6 + 3] = sy;
+    a[(i * 2 + 1) * 6 + 4] = 0;
+    a[(i * 2 + 1) * 6 + 5] = 1;
+
+    b[2 * i] = dx;
+    b[2 * i + 1] = dy;
+  }
+  if (pseudo_inverse(temp, a, np2, 6)) {
+    aom_free(a);
+    return 1;
+  }
+  multiply_mat(temp, b, mat, 6, np2, 1);
+  denormalize_affine_reorder(mat, T1, T2);
+  aom_free(a);
+  return 0;
+}
+
+static int find_vertrapezoid(int np, double *pts1, double *pts2, double *mat) {
+  const int np3 = np * 3;
+  double *a = (double *)aom_malloc(sizeof(*a) * np3 * 14);
+  double *U = a + np3 * 7;
+  double S[7], V[7 * 7], H[9];
+  int i, mini;
+  double sx, sy, dx, dy;
+  double T1[9], T2[9];
+
+  normalize_homography(pts1, np, T1);
+  normalize_homography(pts2, np, T2);
+
+  for (i = 0; i < np; ++i) {
+    dx = *(pts2++);
+    dy = *(pts2++);
+    sx = *(pts1++);
+    sy = *(pts1++);
+
+    a[i * 3 * 7 + 0] = a[i * 3 * 7 + 1] = 0;
+    a[i * 3 * 7 + 2] = -sx;
+    a[i * 3 * 7 + 3] = -sy;
+    a[i * 3 * 7 + 4] = -1;
+    a[i * 3 * 7 + 5] = dy * sx;
+    a[i * 3 * 7 + 6] = dy;
+
+    a[(i * 3 + 1) * 7 + 0] = sx;
+    a[(i * 3 + 1) * 7 + 1] = 1;
+    a[(i * 3 + 1) * 7 + 2] = a[(i * 3 + 1) * 7 + 3] = a[(i * 3 + 1) * 7 + 4] =
+        0;
+    a[(i * 3 + 1) * 7 + 5] = -dx * sx;
+    a[(i * 3 + 1) * 7 + 6] = -dx;
+
+    a[(i * 3 + 2) * 7 + 0] = -dy * sx;
+    a[(i * 3 + 2) * 7 + 1] = -dy;
+    a[(i * 3 + 2) * 7 + 2] = dx * sx;
+    a[(i * 3 + 2) * 7 + 3] = dx * sy;
+    a[(i * 3 + 2) * 7 + 4] = dx;
+    a[(i * 3 + 2) * 7 + 5] = a[(i * 3 + 2) * 7 + 6] = 0;
+  }
+  if (SVD(U, S, V, a, np3, 7)) {
+    aom_free(a);
+    return 1;
+  } else {
+    double minS = 1e12;
+    mini = -1;
+    for (i = 0; i < 7; ++i) {
+      if (S[i] < minS) {
+        minS = S[i];
+        mini = i;
+      }
+    }
+  }
+  H[1] = H[7] = 0;
+  for (i = 0; i < 1; i++) H[i] = V[i * 7 + mini];
+  for (; i < 6; i++) H[i + 1] = V[i * 7 + mini];
+  for (; i < 7; i++) H[i + 2] = V[i * 7 + mini];
+
+  denormalize_homography_reorder(H, T1, T2);
+  aom_free(a);
+  if (H[8] == 0.0) {
+    return 1;
+  } else {
+    // normalize
+    double f = 1.0 / H[8];
+    for (i = 0; i < 8; i++) mat[i] = f * H[i];
+  }
+  return 0;
+}
+
+static int find_hortrapezoid(int np, double *pts1, double *pts2, double *mat) {
+  const int np3 = np * 3;
+  double *a = (double *)aom_malloc(sizeof(*a) * np3 * 14);
+  double *U = a + np3 * 7;
+  double S[7], V[7 * 7], H[9];
+  int i, mini;
+  double sx, sy, dx, dy;
+  double T1[9], T2[9];
+
+  normalize_homography(pts1, np, T1);
+  normalize_homography(pts2, np, T2);
+
+  for (i = 0; i < np; ++i) {
+    dx = *(pts2++);
+    dy = *(pts2++);
+    sx = *(pts1++);
+    sy = *(pts1++);
+
+    a[i * 3 * 7 + 0] = a[i * 3 * 7 + 1] = a[i * 3 * 7 + 2] = 0;
+    a[i * 3 * 7 + 3] = -sy;
+    a[i * 3 * 7 + 4] = -1;
+    a[i * 3 * 7 + 5] = dy * sy;
+    a[i * 3 * 7 + 6] = dy;
+
+    a[(i * 3 + 1) * 7 + 0] = sx;
+    a[(i * 3 + 1) * 7 + 1] = sy;
+    a[(i * 3 + 1) * 7 + 2] = 1;
+    a[(i * 3 + 1) * 7 + 3] = a[(i * 3 + 1) * 7 + 4] = 0;
+    a[(i * 3 + 1) * 7 + 5] = -dx * sy;
+    a[(i * 3 + 1) * 7 + 6] = -dx;
+
+    a[(i * 3 + 2) * 7 + 0] = -dy * sx;
+    a[(i * 3 + 2) * 7 + 1] = -dy * sy;
+    a[(i * 3 + 2) * 7 + 2] = -dy;
+    a[(i * 3 + 2) * 7 + 3] = dx * sy;
+    a[(i * 3 + 2) * 7 + 4] = dx;
+    a[(i * 3 + 2) * 7 + 5] = a[(i * 3 + 2) * 7 + 6] = 0;
+  }
+
+  if (SVD(U, S, V, a, np3, 7)) {
+    aom_free(a);
+    return 1;
+  } else {
+    double minS = 1e12;
+    mini = -1;
+    for (i = 0; i < 7; ++i) {
+      if (S[i] < minS) {
+        minS = S[i];
+        mini = i;
+      }
+    }
+  }
+  H[3] = H[6] = 0;
+  for (i = 0; i < 3; i++) H[i] = V[i * 7 + mini];
+  for (; i < 5; i++) H[i + 1] = V[i * 7 + mini];
+  for (; i < 7; i++) H[i + 2] = V[i * 7 + mini];
+
+  denormalize_homography_reorder(H, T1, T2);
+  aom_free(a);
+  if (H[8] == 0.0) {
+    return 1;
+  } else {
+    // normalize
+    double f = 1.0 / H[8];
+    for (i = 0; i < 8; i++) mat[i] = f * H[i];
+  }
+  return 0;
+}
+
+static int find_homography(int np, double *pts1, double *pts2, double *mat) {
+  // Implemented from Peter Kovesi's normalized implementation
+  const int np3 = np * 3;
+  double *a = (double *)aom_malloc(sizeof(*a) * np3 * 18);
+  double *U = a + np3 * 9;
+  double S[9], V[9 * 9], H[9];
+  int i, mini;
+  double sx, sy, dx, dy;
+  double T1[9], T2[9];
+
+  normalize_homography(pts1, np, T1);
+  normalize_homography(pts2, np, T2);
+
+  for (i = 0; i < np; ++i) {
+    dx = *(pts2++);
+    dy = *(pts2++);
+    sx = *(pts1++);
+    sy = *(pts1++);
+
+    a[i * 3 * 9 + 0] = a[i * 3 * 9 + 1] = a[i * 3 * 9 + 2] = 0;
+    a[i * 3 * 9 + 3] = -sx;
+    a[i * 3 * 9 + 4] = -sy;
+    a[i * 3 * 9 + 5] = -1;
+    a[i * 3 * 9 + 6] = dy * sx;
+    a[i * 3 * 9 + 7] = dy * sy;
+    a[i * 3 * 9 + 8] = dy;
+
+    a[(i * 3 + 1) * 9 + 0] = sx;
+    a[(i * 3 + 1) * 9 + 1] = sy;
+    a[(i * 3 + 1) * 9 + 2] = 1;
+    a[(i * 3 + 1) * 9 + 3] = a[(i * 3 + 1) * 9 + 4] = a[(i * 3 + 1) * 9 + 5] =
+        0;
+    a[(i * 3 + 1) * 9 + 6] = -dx * sx;
+    a[(i * 3 + 1) * 9 + 7] = -dx * sy;
+    a[(i * 3 + 1) * 9 + 8] = -dx;
+
+    a[(i * 3 + 2) * 9 + 0] = -dy * sx;
+    a[(i * 3 + 2) * 9 + 1] = -dy * sy;
+    a[(i * 3 + 2) * 9 + 2] = -dy;
+    a[(i * 3 + 2) * 9 + 3] = dx * sx;
+    a[(i * 3 + 2) * 9 + 4] = dx * sy;
+    a[(i * 3 + 2) * 9 + 5] = dx;
+    a[(i * 3 + 2) * 9 + 6] = a[(i * 3 + 2) * 9 + 7] = a[(i * 3 + 2) * 9 + 8] =
+        0;
+  }
+
+  if (SVD(U, S, V, a, np3, 9)) {
+    aom_free(a);
+    return 1;
+  } else {
+    double minS = 1e12;
+    mini = -1;
+    for (i = 0; i < 9; ++i) {
+      if (S[i] < minS) {
+        minS = S[i];
+        mini = i;
+      }
+    }
+  }
+
+  for (i = 0; i < 9; i++) H[i] = V[i * 9 + mini];
+  denormalize_homography_reorder(H, T1, T2);
+  aom_free(a);
+  if (H[8] == 0.0) {
+    return 1;
+  } else {
+    // normalize
+    double f = 1.0 / H[8];
+    for (i = 0; i < 8; i++) mat[i] = f * H[i];
+  }
+  return 0;
+}
+
+static int get_rand_indices(int npoints, int minpts, int *indices,
+                            unsigned int *seed) {
+  int i, j;
+  int ptr = rand_r(seed) % npoints;
+  if (minpts > npoints) return 0;
+  indices[0] = ptr;
+  ptr = (ptr == npoints - 1 ? 0 : ptr + 1);
+  i = 1;
+  while (i < minpts) {
+    int index = rand_r(seed) % npoints;
+    while (index) {
+      ptr = (ptr == npoints - 1 ? 0 : ptr + 1);
+      for (j = 0; j < i; ++j) {
+        if (indices[j] == ptr) break;
+      }
+      if (j == i) index--;
+    }
+    indices[i++] = ptr;
+  }
+  return 1;
+}
+
+typedef struct {
+  int num_inliers;
+  double variance;
+  int *inlier_indices;
+} RANSAC_MOTION;
+
+// Return -1 if 'a' is a better motion, 1 if 'b' is better, 0 otherwise.
+static int compare_motions(const void *arg_a, const void *arg_b) {
+  const RANSAC_MOTION *motion_a = (RANSAC_MOTION *)arg_a;
+  const RANSAC_MOTION *motion_b = (RANSAC_MOTION *)arg_b;
+
+  if (motion_a->num_inliers > motion_b->num_inliers) return -1;
+  if (motion_a->num_inliers < motion_b->num_inliers) return 1;
+  if (motion_a->variance < motion_b->variance) return -1;
+  if (motion_a->variance > motion_b->variance) return 1;
+  return 0;
+}
+
+static int is_better_motion(const RANSAC_MOTION *motion_a,
+                            const RANSAC_MOTION *motion_b) {
+  return compare_motions(motion_a, motion_b) < 0;
+}
+
+static void copy_points_at_indices(double *dest, const double *src,
+                                   const int *indices, int num_points) {
+  for (int i = 0; i < num_points; ++i) {
+    const int index = indices[i];
+    dest[i * 2] = src[index * 2];
+    dest[i * 2 + 1] = src[index * 2 + 1];
+  }
+}
+
+static const double kInfiniteVariance = 1e12;
+
+static void clear_motion(RANSAC_MOTION *motion, int num_points) {
+  motion->num_inliers = 0;
+  motion->variance = kInfiniteVariance;
+  memset(motion->inlier_indices, 0,
+         sizeof(*motion->inlier_indices * num_points));
+}
+
+static int ransac(const int *matched_points, int npoints,
+                  int *num_inliers_by_motion, double *params_by_motion,
+                  int num_desired_motions, const int minpts,
+                  IsDegenerateFunc is_degenerate,
+                  FindTransformationFunc find_transformation,
+                  ProjectPointsDoubleFunc projectpoints) {
+  static const double PROBABILITY_REQUIRED = 0.9;
+  static const double EPS = 1e-12;
+
+  int N = 10000, trial_count = 0;
+  int i = 0;
+  int ret_val = 0;
+
+  unsigned int seed = (unsigned int)npoints;
+
+  int indices[MAX_MINPTS] = { 0 };
+
+  double *points1, *points2;
+  double *corners1, *corners2;
+  double *image1_coord;
+
+  // Store information for the num_desired_motions best transformations found
+  // and the worst motion among them, as well as the motion currently under
+  // consideration.
+  RANSAC_MOTION *motions, *worst_kept_motion = NULL;
+  RANSAC_MOTION current_motion;
+
+  // Store the parameters and the indices of the inlier points for the motion
+  // currently under consideration.
+  double params_this_motion[MAX_PARAMDIM];
+
+  double *cnp1, *cnp2;
+
+  if (npoints < minpts * MINPTS_MULTIPLIER || npoints == 0) {
+    return 1;
+  }
+
+  points1 = (double *)aom_malloc(sizeof(*points1) * npoints * 2);
+  points2 = (double *)aom_malloc(sizeof(*points2) * npoints * 2);
+  corners1 = (double *)aom_malloc(sizeof(*corners1) * npoints * 2);
+  corners2 = (double *)aom_malloc(sizeof(*corners2) * npoints * 2);
+  image1_coord = (double *)aom_malloc(sizeof(*image1_coord) * npoints * 2);
+
+  motions =
+      (RANSAC_MOTION *)aom_malloc(sizeof(RANSAC_MOTION) * num_desired_motions);
+  for (i = 0; i < num_desired_motions; ++i) {
+    motions[i].inlier_indices =
+        (int *)aom_malloc(sizeof(*motions->inlier_indices) * npoints);
+    clear_motion(motions + i, npoints);
+  }
+  current_motion.inlier_indices =
+      (int *)aom_malloc(sizeof(*current_motion.inlier_indices) * npoints);
+  clear_motion(&current_motion, npoints);
+
+  worst_kept_motion = motions;
+
+  if (!(points1 && points2 && corners1 && corners2 && image1_coord && motions &&
+        current_motion.inlier_indices)) {
+    ret_val = 1;
+    goto finish_ransac;
+  }
+
+  cnp1 = corners1;
+  cnp2 = corners2;
+  for (i = 0; i < npoints; ++i) {
+    *(cnp1++) = *(matched_points++);
+    *(cnp1++) = *(matched_points++);
+    *(cnp2++) = *(matched_points++);
+    *(cnp2++) = *(matched_points++);
+  }
+
+  while (N > trial_count) {
+    double sum_distance = 0.0;
+    double sum_distance_squared = 0.0;
+
+    clear_motion(&current_motion, npoints);
+
+    int degenerate = 1;
+    int num_degenerate_iter = 0;
+
+    while (degenerate) {
+      num_degenerate_iter++;
+      if (!get_rand_indices(npoints, minpts, indices, &seed)) {
+        ret_val = 1;
+        goto finish_ransac;
+      }
+
+      copy_points_at_indices(points1, corners1, indices, minpts);
+      copy_points_at_indices(points2, corners2, indices, minpts);
+
+      degenerate = is_degenerate(points1);
+      if (num_degenerate_iter > MAX_DEGENERATE_ITER) {
+        ret_val = 1;
+        goto finish_ransac;
+      }
+    }
+
+    if (find_transformation(minpts, points1, points2, params_this_motion)) {
+      trial_count++;
+      continue;
+    }
+
+    projectpoints(params_this_motion, corners1, image1_coord, npoints, 2, 2);
+
+    for (i = 0; i < npoints; ++i) {
+      double dx = image1_coord[i * 2] - corners2[i * 2];
+      double dy = image1_coord[i * 2 + 1] - corners2[i * 2 + 1];
+      double distance = sqrt(dx * dx + dy * dy);
+
+      if (distance < INLIER_THRESHOLD) {
+        current_motion.inlier_indices[current_motion.num_inliers++] = i;
+        sum_distance += distance;
+        sum_distance_squared += distance * distance;
+      }
+    }
+
+    if (current_motion.num_inliers >= worst_kept_motion->num_inliers &&
+        current_motion.num_inliers > 1) {
+      int temp;
+      double fracinliers, pNoOutliers, mean_distance;
+      mean_distance = sum_distance / ((double)current_motion.num_inliers);
+      current_motion.variance =
+          sum_distance_squared / ((double)current_motion.num_inliers - 1.0) -
+          mean_distance * mean_distance * ((double)current_motion.num_inliers) /
+              ((double)current_motion.num_inliers - 1.0);
+      if (is_better_motion(&current_motion, worst_kept_motion)) {
+        // This motion is better than the worst currently kept motion. Remember
+        // the inlier points and variance. The parameters for each kept motion
+        // will be recomputed later using only the inliers.
+        worst_kept_motion->num_inliers = current_motion.num_inliers;
+        worst_kept_motion->variance = current_motion.variance;
+        memcpy(worst_kept_motion->inlier_indices, current_motion.inlier_indices,
+               sizeof(*current_motion.inlier_indices) * npoints);
+
+        assert(npoints > 0);
+        fracinliers = (double)current_motion.num_inliers / (double)npoints;
+        pNoOutliers = 1 - pow(fracinliers, minpts);
+        pNoOutliers = fmax(EPS, pNoOutliers);
+        pNoOutliers = fmin(1 - EPS, pNoOutliers);
+        temp = (int)(log(1.0 - PROBABILITY_REQUIRED) / log(pNoOutliers));
+
+        if (temp > 0 && temp < N) {
+          N = AOMMAX(temp, MIN_TRIALS);
+        }
+
+        // Determine the new worst kept motion and its num_inliers and variance.
+        for (i = 0; i < num_desired_motions; ++i) {
+          if (is_better_motion(worst_kept_motion, &motions[i])) {
+            worst_kept_motion = &motions[i];
+          }
+        }
+      }
+    }
+    trial_count++;
+  }
+
+  // Sort the motions, best first.
+  qsort(motions, num_desired_motions, sizeof(RANSAC_MOTION), compare_motions);
+
+  // Recompute the motions using only the inliers.
+  for (i = 0; i < num_desired_motions; ++i) {
+    copy_points_at_indices(points1, corners1, motions[i].inlier_indices,
+                           motions[i].num_inliers);
+    copy_points_at_indices(points2, corners2, motions[i].inlier_indices,
+                           motions[i].num_inliers);
+
+    find_transformation(motions[i].num_inliers, points1, points2,
+                        params_by_motion + (MAX_PARAMDIM - 1) * i);
+    num_inliers_by_motion[i] = motions[i].num_inliers;
+  }
+
+finish_ransac:
+  aom_free(points1);
+  aom_free(points2);
+  aom_free(corners1);
+  aom_free(corners2);
+  aom_free(image1_coord);
+  aom_free(current_motion.inlier_indices);
+  for (i = 0; i < num_desired_motions; ++i) {
+    aom_free(motions[i].inlier_indices);
+  }
+  aom_free(motions);
+
+  return ret_val;
+}
+
+static int is_collinear3(double *p1, double *p2, double *p3) {
+  static const double collinear_eps = 1e-3;
+  const double v =
+      (p2[0] - p1[0]) * (p3[1] - p1[1]) - (p2[1] - p1[1]) * (p3[0] - p1[0]);
+  return fabs(v) < collinear_eps;
+}
+
+static int is_degenerate_translation(double *p) {
+  return (p[0] - p[2]) * (p[0] - p[2]) + (p[1] - p[3]) * (p[1] - p[3]) <= 2;
+}
+
+static int is_degenerate_affine(double *p) {
+  return is_collinear3(p, p + 2, p + 4);
+}
+
+static int is_degenerate_homography(double *p) {
+  return is_collinear3(p, p + 2, p + 4) || is_collinear3(p, p + 2, p + 6) ||
+         is_collinear3(p, p + 4, p + 6) || is_collinear3(p + 2, p + 4, p + 6);
+}
+
+int ransac_translation(int *matched_points, int npoints,
+                       int *num_inliers_by_motion, double *params_by_motion,
+                       int num_desired_motions) {
+  return ransac(matched_points, npoints, num_inliers_by_motion,
+                params_by_motion, num_desired_motions, 3,
+                is_degenerate_translation, find_translation,
+                project_points_double_translation);
+}
+
+int ransac_rotzoom(int *matched_points, int npoints, int *num_inliers_by_motion,
+                   double *params_by_motion, int num_desired_motions) {
+  return ransac(matched_points, npoints, num_inliers_by_motion,
+                params_by_motion, num_desired_motions, 3, is_degenerate_affine,
+                find_rotzoom, project_points_double_rotzoom);
+}
+
+int ransac_affine(int *matched_points, int npoints, int *num_inliers_by_motion,
+                  double *params_by_motion, int num_desired_motions) {
+  return ransac(matched_points, npoints, num_inliers_by_motion,
+                params_by_motion, num_desired_motions, 3, is_degenerate_affine,
+                find_affine, project_points_double_affine);
+}
+
+int ransac_homography(int *matched_points, int npoints,
+                      int *num_inliers_by_motion, double *params_by_motion,
+                      int num_desired_motions) {
+  return ransac(matched_points, npoints, num_inliers_by_motion,
+                params_by_motion, num_desired_motions, 4,
+                is_degenerate_homography, find_homography,
+                project_points_double_homography);
+}
+
+int ransac_hortrapezoid(int *matched_points, int npoints,
+                        int *num_inliers_by_motion, double *params_by_motion,
+                        int num_desired_motions) {
+  return ransac(matched_points, npoints, num_inliers_by_motion,
+                params_by_motion, num_desired_motions, 4,
+                is_degenerate_homography, find_hortrapezoid,
+                project_points_double_hortrapezoid);
+}
+
+int ransac_vertrapezoid(int *matched_points, int npoints,
+                        int *num_inliers_by_motion, double *params_by_motion,
+                        int num_desired_motions) {
+  return ransac(matched_points, npoints, num_inliers_by_motion,
+                params_by_motion, num_desired_motions, 4,
+                is_degenerate_homography, find_vertrapezoid,
+                project_points_double_vertrapezoid);
+}
diff --git a/third_party/aom/av1/encoder/ransac.h b/third_party/aom/av1/encoder/ransac.h
new file mode 100644
index 0000000000..f611add369
--- /dev/null
+++ b/third_party/aom/av1/encoder/ransac.h
@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_RANSAC_H_
+#define AV1_ENCODER_RANSAC_H_
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <math.h>
+#include <memory.h>
+
+#include "av1/common/warped_motion.h"
+
+typedef int (*RansacFunc)(int *matched_points, int npoints,
+                          int *num_inliers_by_motion, double *params_by_motion,
+                          int num_motions);
+
+/* Each of these functions fits a motion model from a set of
+   corresponding points in 2 frames using RANSAC. */
+int ransac_homography(int *matched_points, int npoints,
+                      int *num_inliers_by_motion, double *params_by_motion,
+                      int num_motions);
+int ransac_affine(int *matched_points, int npoints, int *num_inliers_by_motion,
+                  double *params_by_motion, int num_motions);
+int ransac_hortrapezoid(int *matched_points, int npoints,
+                        int *num_inliers_by_motion, double *params_by_motion,
+                        int num_motions);
+int ransac_vertrapezoid(int *matched_points, int npoints,
+                        int *num_inliers_by_motion, double *params_by_motion,
+                        int num_motions);
+int ransac_rotzoom(int *matched_points, int npoints, int *num_inliers_by_motion,
+                   double *params_by_motion, int num_motions);
+int ransac_translation(int *matched_points, int npoints,
+                       int *num_inliers_by_motion, double *params_by_motion,
+                       int num_motions);
+#endif  // AV1_ENCODER_RANSAC_H_
diff --git a/third_party/aom/av1/encoder/ratectrl.c b/third_party/aom/av1/encoder/ratectrl.c
new file mode 100644
index 0000000000..1f2ea3606f
--- /dev/null
+++ b/third_party/aom/av1/encoder/ratectrl.c
@@ -0,0 +1,1759 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/system_state.h"
+
+#include "av1/common/alloccommon.h"
+#include "av1/encoder/aq_cyclicrefresh.h"
+#include "av1/common/common.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/seg_common.h"
+
+#include "av1/encoder/encodemv.h"
+#include "av1/encoder/ratectrl.h"
+
+// Max rate target for 1080P and below encodes under normal circumstances
+// (1920 * 1080 / (16 * 16)) * MAX_MB_RATE bits per MB
+#define MAX_MB_RATE 250
+#define MAXRATE_1080P 2025000
+
+#define DEFAULT_KF_BOOST 2000
+#define DEFAULT_GF_BOOST 2000
+
+#define MIN_BPB_FACTOR 0.005
+#define MAX_BPB_FACTOR 50
+
+#define FRAME_OVERHEAD_BITS 200
+#if CONFIG_HIGHBITDEPTH
+#define ASSIGN_MINQ_TABLE(bit_depth, name)                   \
+  do {                                                       \
+    switch (bit_depth) {                                     \
+      case AOM_BITS_8: name = name##_8; break;               \
+      case AOM_BITS_10: name = name##_10; break;             \
+      case AOM_BITS_12: name = name##_12; break;             \
+      default:                                               \
+        assert(0 &&                                          \
+               "bit_depth should be AOM_BITS_8, AOM_BITS_10" \
+               " or AOM_BITS_12");                           \
+        name = NULL;                                         \
+    }                                                        \
+  } while (0)
+#else
+#define ASSIGN_MINQ_TABLE(bit_depth, name) \
+  do {                                     \
+    (void)bit_depth;                       \
+    name = name##_8;                       \
+  } while (0)
+#endif
+
+// Tables relating active max Q to active min Q
+static int kf_low_motion_minq_8[QINDEX_RANGE];
+static int kf_high_motion_minq_8[QINDEX_RANGE];
+static int arfgf_low_motion_minq_8[QINDEX_RANGE];
+static int arfgf_high_motion_minq_8[QINDEX_RANGE];
+static int inter_minq_8[QINDEX_RANGE];
+static int rtc_minq_8[QINDEX_RANGE];
+
+#if CONFIG_HIGHBITDEPTH
+static int kf_low_motion_minq_10[QINDEX_RANGE];
+static int kf_high_motion_minq_10[QINDEX_RANGE];
+static int arfgf_low_motion_minq_10[QINDEX_RANGE];
+static int arfgf_high_motion_minq_10[QINDEX_RANGE];
+static int inter_minq_10[QINDEX_RANGE];
+static int rtc_minq_10[QINDEX_RANGE];
+static int kf_low_motion_minq_12[QINDEX_RANGE];
+static int kf_high_motion_minq_12[QINDEX_RANGE];
+static int arfgf_low_motion_minq_12[QINDEX_RANGE];
+static int arfgf_high_motion_minq_12[QINDEX_RANGE];
+static int inter_minq_12[QINDEX_RANGE];
+static int rtc_minq_12[QINDEX_RANGE];
+#endif
+
+static int gf_high = 2000;
+static int gf_low = 400;
+static int kf_high = 5000;
+static int kf_low = 400;
+
+// Functions to compute the active minq lookup table entries based on a
+// formulaic approach to facilitate easier adjustment of the Q tables.
+// The formulae were derived from computing a 3rd order polynomial best
+// fit to the original data (after plotting real maxq vs minq (not q index))
+static int get_minq_index(double maxq, double x3, double x2, double x1,
+                          aom_bit_depth_t bit_depth) {
+  int i;
+  const double minqtarget = AOMMIN(((x3 * maxq + x2) * maxq + x1) * maxq, maxq);
+
+  // Special case handling to deal with the step from q2.0
+  // down to lossless mode represented by q 1.0.
+  if (minqtarget <= 2.0) return 0;
+
+  for (i = 0; i < QINDEX_RANGE; i++) {
+    if (minqtarget <= av1_convert_qindex_to_q(i, bit_depth)) return i;
+  }
+
+  return QINDEX_RANGE - 1;
+}
+
+static void init_minq_luts(int *kf_low_m, int *kf_high_m, int *arfgf_low,
+                           int *arfgf_high, int *inter, int *rtc,
+                           aom_bit_depth_t bit_depth) {
+  int i;
+  for (i = 0; i < QINDEX_RANGE; i++) {
+    const double maxq = av1_convert_qindex_to_q(i, bit_depth);
+    kf_low_m[i] = get_minq_index(maxq, 0.000001, -0.0004, 0.150, bit_depth);
+    kf_high_m[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.55, bit_depth);
+    arfgf_low[i] = get_minq_index(maxq, 0.0000015, -0.0009, 0.30, bit_depth);
+    arfgf_high[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.55, bit_depth);
+    inter[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.90, bit_depth);
+    rtc[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.70, bit_depth);
+  }
+}
+
+void av1_rc_init_minq_luts(void) {
+  init_minq_luts(kf_low_motion_minq_8, kf_high_motion_minq_8,
+                 arfgf_low_motion_minq_8, arfgf_high_motion_minq_8,
+                 inter_minq_8, rtc_minq_8, AOM_BITS_8);
+#if CONFIG_HIGHBITDEPTH
+  init_minq_luts(kf_low_motion_minq_10, kf_high_motion_minq_10,
+                 arfgf_low_motion_minq_10, arfgf_high_motion_minq_10,
+                 inter_minq_10, rtc_minq_10, AOM_BITS_10);
+  init_minq_luts(kf_low_motion_minq_12, kf_high_motion_minq_12,
+                 arfgf_low_motion_minq_12, arfgf_high_motion_minq_12,
+                 inter_minq_12, rtc_minq_12, AOM_BITS_12);
+#endif
+}
+
+// These functions use formulaic calculations to make playing with the
+// quantizer tables easier. If necessary they can be replaced by lookup
+// tables if and when things settle down in the experimental bitstream
+double av1_convert_qindex_to_q(int qindex, aom_bit_depth_t bit_depth) {
+// Convert the index to a real Q value (scaled down to match old Q values)
+#if CONFIG_HIGHBITDEPTH
+  switch (bit_depth) {
+    case AOM_BITS_8: return av1_ac_quant(qindex, 0, bit_depth) / 4.0;
+    case AOM_BITS_10: return av1_ac_quant(qindex, 0, bit_depth) / 16.0;
+    case AOM_BITS_12: return av1_ac_quant(qindex, 0, bit_depth) / 64.0;
+    default:
+      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
+      return -1.0;
+  }
+#else
+  return av1_ac_quant(qindex, 0, bit_depth) / 4.0;
+#endif
+}
+
+int av1_rc_bits_per_mb(FRAME_TYPE frame_type, int qindex,
+                       double correction_factor, aom_bit_depth_t bit_depth) {
+  const double q = av1_convert_qindex_to_q(qindex, bit_depth);
+  int enumerator = frame_type == KEY_FRAME ? 2700000 : 1800000;
+
+  assert(correction_factor <= MAX_BPB_FACTOR &&
+         correction_factor >= MIN_BPB_FACTOR);
+
+  // q based adjustment to baseline enumerator
+  enumerator += (int)(enumerator * q) >> 12;
+  return (int)(enumerator * correction_factor / q);
+}
+
+int av1_estimate_bits_at_q(FRAME_TYPE frame_type, int q, int mbs,
+                           double correction_factor,
+                           aom_bit_depth_t bit_depth) {
+  const int bpm =
+      (int)(av1_rc_bits_per_mb(frame_type, q, correction_factor, bit_depth));
+  return AOMMAX(FRAME_OVERHEAD_BITS,
+                (int)((uint64_t)bpm * mbs) >> BPER_MB_NORMBITS);
+}
+
+int av1_rc_clamp_pframe_target_size(const AV1_COMP *const cpi, int target) {
+  const RATE_CONTROL *rc = &cpi->rc;
+  const AV1EncoderConfig *oxcf = &cpi->oxcf;
+  const int min_frame_target =
+      AOMMAX(rc->min_frame_bandwidth, rc->avg_frame_bandwidth >> 5);
+// Clip the frame target to the minimum setup value.
+#if CONFIG_EXT_REFS
+  if (cpi->rc.is_src_frame_alt_ref) {
+#else
+  if (cpi->refresh_golden_frame && rc->is_src_frame_alt_ref) {
+#endif  // CONFIG_EXT_REFS
+    // If there is an active ARF at this location use the minimum
+    // bits on this frame even if it is a constructed arf.
+    // The active maximum quantizer insures that an appropriate
+    // number of bits will be spent if needed for constructed ARFs.
+    target = min_frame_target;
+  } else if (target < min_frame_target) {
+    target = min_frame_target;
+  }
+
+  // Clip the frame target to the maximum allowed value.
+  if (target > rc->max_frame_bandwidth) target = rc->max_frame_bandwidth;
+  if (oxcf->rc_max_inter_bitrate_pct) {
+    const int max_rate =
+        rc->avg_frame_bandwidth * oxcf->rc_max_inter_bitrate_pct / 100;
+    target = AOMMIN(target, max_rate);
+  }
+
+  return target;
+}
+
+int av1_rc_clamp_iframe_target_size(const AV1_COMP *const cpi, int target) {
+  const RATE_CONTROL *rc = &cpi->rc;
+  const AV1EncoderConfig *oxcf = &cpi->oxcf;
+  if (oxcf->rc_max_intra_bitrate_pct) {
+    const int max_rate =
+        rc->avg_frame_bandwidth * oxcf->rc_max_intra_bitrate_pct / 100;
+    target = AOMMIN(target, max_rate);
+  }
+  if (target > rc->max_frame_bandwidth) target = rc->max_frame_bandwidth;
+  return target;
+}
+
+// Update the buffer level: leaky bucket model.
+static void update_buffer_level(AV1_COMP *cpi, int encoded_frame_size) {
+  const AV1_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+
+// Non-viewable frames are a special case and are treated as pure overhead.
+#if CONFIG_EXT_REFS
+  // TODO(zoeliu): To further explore whether we should treat BWDREF_FRAME
+  //               differently, since it is a no-show frame.
+  if (!cm->show_frame && !rc->is_bwd_ref_frame)
+#else
+  if (!cm->show_frame)
+#endif  // CONFIG_EXT_REFS
+    rc->bits_off_target -= encoded_frame_size;
+  else
+    rc->bits_off_target += rc->avg_frame_bandwidth - encoded_frame_size;
+
+  // Clip the buffer level to the maximum specified buffer size.
+  rc->bits_off_target = AOMMIN(rc->bits_off_target, rc->maximum_buffer_size);
+  rc->buffer_level = rc->bits_off_target;
+}
+
+int av1_rc_get_default_min_gf_interval(int width, int height,
+                                       double framerate) {
+  // Assume we do not need any constraint lower than 4K 20 fps
+  static const double factor_safe = 3840 * 2160 * 20.0;
+  const double factor = width * height * framerate;
+  const int default_interval =
+      clamp((int)(framerate * 0.125), MIN_GF_INTERVAL, MAX_GF_INTERVAL);
+
+  if (factor <= factor_safe)
+    return default_interval;
+  else
+    return AOMMAX(default_interval,
+                  (int)(MIN_GF_INTERVAL * factor / factor_safe + 0.5));
+  // Note this logic makes:
+  // 4K24: 5
+  // 4K30: 6
+  // 4K60: 12
+}
+
+int av1_rc_get_default_max_gf_interval(double framerate, int min_gf_interval) {
+  int interval = AOMMIN(MAX_GF_INTERVAL, (int)(framerate * 0.75));
+  interval += (interval & 0x01);  // Round to even value
+  return AOMMAX(interval, min_gf_interval);
+}
+
+void av1_rc_init(const AV1EncoderConfig *oxcf, int pass, RATE_CONTROL *rc) {
+  int i;
+
+  if (pass == 0 && oxcf->rc_mode == AOM_CBR) {
+    rc->avg_frame_qindex[KEY_FRAME] = oxcf->worst_allowed_q;
+    rc->avg_frame_qindex[INTER_FRAME] = oxcf->worst_allowed_q;
+  } else {
+    rc->avg_frame_qindex[KEY_FRAME] =
+        (oxcf->worst_allowed_q + oxcf->best_allowed_q) / 2;
+    rc->avg_frame_qindex[INTER_FRAME] =
+        (oxcf->worst_allowed_q + oxcf->best_allowed_q) / 2;
+  }
+
+  rc->last_q[KEY_FRAME] = oxcf->best_allowed_q;
+  rc->last_q[INTER_FRAME] = oxcf->worst_allowed_q;
+
+  rc->buffer_level = rc->starting_buffer_level;
+  rc->bits_off_target = rc->starting_buffer_level;
+
+  rc->rolling_target_bits = rc->avg_frame_bandwidth;
+  rc->rolling_actual_bits = rc->avg_frame_bandwidth;
+  rc->long_rolling_target_bits = rc->avg_frame_bandwidth;
+  rc->long_rolling_actual_bits = rc->avg_frame_bandwidth;
+
+  rc->total_actual_bits = 0;
+  rc->total_target_bits = 0;
+  rc->total_target_vs_actual = 0;
+
+  rc->frames_since_key = 8;  // Sensible default for first frame.
+  rc->this_key_frame_forced = 0;
+  rc->next_key_frame_forced = 0;
+  rc->source_alt_ref_pending = 0;
+  rc->source_alt_ref_active = 0;
+
+  rc->frames_till_gf_update_due = 0;
+  rc->ni_av_qi = oxcf->worst_allowed_q;
+  rc->ni_tot_qi = 0;
+  rc->ni_frames = 0;
+
+  rc->tot_q = 0.0;
+  rc->avg_q = av1_convert_qindex_to_q(oxcf->worst_allowed_q, oxcf->bit_depth);
+
+  for (i = 0; i < RATE_FACTOR_LEVELS; ++i) {
+    rc->rate_correction_factors[i] = 1.0;
+  }
+
+  rc->min_gf_interval = oxcf->min_gf_interval;
+  rc->max_gf_interval = oxcf->max_gf_interval;
+  if (rc->min_gf_interval == 0)
+    rc->min_gf_interval = av1_rc_get_default_min_gf_interval(
+        oxcf->width, oxcf->height, oxcf->init_framerate);
+  if (rc->max_gf_interval == 0)
+    rc->max_gf_interval = av1_rc_get_default_max_gf_interval(
+        oxcf->init_framerate, rc->min_gf_interval);
+  rc->baseline_gf_interval = (rc->min_gf_interval + rc->max_gf_interval) / 2;
+}
+
+int av1_rc_drop_frame(AV1_COMP *cpi) {
+  const AV1EncoderConfig *oxcf = &cpi->oxcf;
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  if (!oxcf->drop_frames_water_mark) {
+    return 0;
+  } else {
+    if (rc->buffer_level < 0) {
+      // Always drop if buffer is below 0.
+      return 1;
+    } else {
+      // If buffer is below drop_mark, for now just drop every other frame
+      // (starting with the next frame) until it increases back over drop_mark.
+      int drop_mark =
+          (int)(oxcf->drop_frames_water_mark * rc->optimal_buffer_level / 100);
+      if ((rc->buffer_level > drop_mark) && (rc->decimation_factor > 0)) {
+        --rc->decimation_factor;
+      } else if (rc->buffer_level <= drop_mark && rc->decimation_factor == 0) {
+        rc->decimation_factor = 1;
+      }
+      if (rc->decimation_factor > 0) {
+        if (rc->decimation_count > 0) {
+          --rc->decimation_count;
+          return 1;
+        } else {
+          rc->decimation_count = rc->decimation_factor;
+          return 0;
+        }
+      } else {
+        rc->decimation_count = 0;
+        return 0;
+      }
+    }
+  }
+}
+
+static double get_rate_correction_factor(const AV1_COMP *cpi) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  double rcf;
+
+  if (cpi->common.frame_type == KEY_FRAME) {
+    rcf = rc->rate_correction_factors[KF_STD];
+  } else if (cpi->oxcf.pass == 2) {
+    RATE_FACTOR_LEVEL rf_lvl =
+        cpi->twopass.gf_group.rf_level[cpi->twopass.gf_group.index];
+    rcf = rc->rate_correction_factors[rf_lvl];
+  } else {
+    if ((cpi->refresh_alt_ref_frame || cpi->refresh_golden_frame) &&
+        !rc->is_src_frame_alt_ref &&
+        (cpi->oxcf.rc_mode != AOM_CBR || cpi->oxcf.gf_cbr_boost_pct > 20))
+      rcf = rc->rate_correction_factors[GF_ARF_STD];
+    else
+      rcf = rc->rate_correction_factors[INTER_NORMAL];
+  }
+  rcf *= rcf_mult[rc->frame_size_selector];
+  return fclamp(rcf, MIN_BPB_FACTOR, MAX_BPB_FACTOR);
+}
+
+static void set_rate_correction_factor(AV1_COMP *cpi, double factor) {
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  // Normalize RCF to account for the size-dependent scaling factor.
+  factor /= rcf_mult[cpi->rc.frame_size_selector];
+
+  factor = fclamp(factor, MIN_BPB_FACTOR, MAX_BPB_FACTOR);
+
+  if (cpi->common.frame_type == KEY_FRAME) {
+    rc->rate_correction_factors[KF_STD] = factor;
+  } else if (cpi->oxcf.pass == 2) {
+    RATE_FACTOR_LEVEL rf_lvl =
+        cpi->twopass.gf_group.rf_level[cpi->twopass.gf_group.index];
+    rc->rate_correction_factors[rf_lvl] = factor;
+  } else {
+    if ((cpi->refresh_alt_ref_frame || cpi->refresh_golden_frame) &&
+        !rc->is_src_frame_alt_ref &&
+        (cpi->oxcf.rc_mode != AOM_CBR || cpi->oxcf.gf_cbr_boost_pct > 20))
+      rc->rate_correction_factors[GF_ARF_STD] = factor;
+    else
+      rc->rate_correction_factors[INTER_NORMAL] = factor;
+  }
+}
+
+void av1_rc_update_rate_correction_factors(AV1_COMP *cpi) {
+  const AV1_COMMON *const cm = &cpi->common;
+  int correction_factor = 100;
+  double rate_correction_factor = get_rate_correction_factor(cpi);
+  double adjustment_limit;
+
+  int projected_size_based_on_q = 0;
+
+  // Do not update the rate factors for arf overlay frames.
+  if (cpi->rc.is_src_frame_alt_ref) return;
+
+  // Clear down mmx registers to allow floating point in what follows
+  aom_clear_system_state();
+
+  // Work out how big we would have expected the frame to be at this Q given
+  // the current correction factor.
+  // Stay in double to avoid int overflow when values are large
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cpi->common.seg.enabled) {
+    projected_size_based_on_q =
+        av1_cyclic_refresh_estimate_bits_at_q(cpi, rate_correction_factor);
+  } else {
+    projected_size_based_on_q =
+        av1_estimate_bits_at_q(cpi->common.frame_type, cm->base_qindex, cm->MBs,
+                               rate_correction_factor, cm->bit_depth);
+  }
+  // Work out a size correction factor.
+  if (projected_size_based_on_q > FRAME_OVERHEAD_BITS)
+    correction_factor = (int)((100 * (int64_t)cpi->rc.projected_frame_size) /
+                              projected_size_based_on_q);
+
+  // More heavily damped adjustment used if we have been oscillating either side
+  // of target.
+  if (correction_factor > 0) {
+    adjustment_limit =
+        0.25 + 0.5 * AOMMIN(1, fabs(log10(0.01 * correction_factor)));
+  } else {
+    adjustment_limit = 0.75;
+  }
+
+  cpi->rc.q_2_frame = cpi->rc.q_1_frame;
+  cpi->rc.q_1_frame = cm->base_qindex;
+  cpi->rc.rc_2_frame = cpi->rc.rc_1_frame;
+  if (correction_factor > 110)
+    cpi->rc.rc_1_frame = -1;
+  else if (correction_factor < 90)
+    cpi->rc.rc_1_frame = 1;
+  else
+    cpi->rc.rc_1_frame = 0;
+
+  if (correction_factor > 102) {
+    // We are not already at the worst allowable quality
+    correction_factor =
+        (int)(100 + ((correction_factor - 100) * adjustment_limit));
+    rate_correction_factor = (rate_correction_factor * correction_factor) / 100;
+    // Keep rate_correction_factor within limits
+    if (rate_correction_factor > MAX_BPB_FACTOR)
+      rate_correction_factor = MAX_BPB_FACTOR;
+  } else if (correction_factor < 99) {
+    // We are not already at the best allowable quality
+    correction_factor =
+        (int)(100 - ((100 - correction_factor) * adjustment_limit));
+    rate_correction_factor = (rate_correction_factor * correction_factor) / 100;
+
+    // Keep rate_correction_factor within limits
+    if (rate_correction_factor < MIN_BPB_FACTOR)
+      rate_correction_factor = MIN_BPB_FACTOR;
+  }
+
+  set_rate_correction_factor(cpi, rate_correction_factor);
+}
+
+int av1_rc_regulate_q(const AV1_COMP *cpi, int target_bits_per_frame,
+                      int active_best_quality, int active_worst_quality) {
+  const AV1_COMMON *const cm = &cpi->common;
+  int q = active_worst_quality;
+  int last_error = INT_MAX;
+  int i, target_bits_per_mb, bits_per_mb_at_this_q;
+  const double correction_factor = get_rate_correction_factor(cpi);
+
+  // Calculate required scaling factor based on target frame size and size of
+  // frame produced using previous Q.
+  target_bits_per_mb =
+      (int)((uint64_t)target_bits_per_frame << BPER_MB_NORMBITS) / cm->MBs;
+
+  i = active_best_quality;
+
+  do {
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
+      bits_per_mb_at_this_q =
+          (int)av1_cyclic_refresh_rc_bits_per_mb(cpi, i, correction_factor);
+    } else {
+      bits_per_mb_at_this_q = (int)av1_rc_bits_per_mb(
+          cm->frame_type, i, correction_factor, cm->bit_depth);
+    }
+
+    if (bits_per_mb_at_this_q <= target_bits_per_mb) {
+      if ((target_bits_per_mb - bits_per_mb_at_this_q) <= last_error)
+        q = i;
+      else
+        q = i - 1;
+
+      break;
+    } else {
+      last_error = bits_per_mb_at_this_q - target_bits_per_mb;
+    }
+  } while (++i <= active_worst_quality);
+
+  // In CBR mode, this makes sure q is between oscillating Qs to prevent
+  // resonance.
+  if (cpi->oxcf.rc_mode == AOM_CBR &&
+      (cpi->rc.rc_1_frame * cpi->rc.rc_2_frame == -1) &&
+      cpi->rc.q_1_frame != cpi->rc.q_2_frame) {
+    q = clamp(q, AOMMIN(cpi->rc.q_1_frame, cpi->rc.q_2_frame),
+              AOMMAX(cpi->rc.q_1_frame, cpi->rc.q_2_frame));
+  }
+  return q;
+}
+
+static int get_active_quality(int q, int gfu_boost, int low, int high,
+                              int *low_motion_minq, int *high_motion_minq) {
+  if (gfu_boost > high) {
+    return low_motion_minq[q];
+  } else if (gfu_boost < low) {
+    return high_motion_minq[q];
+  } else {
+    const int gap = high - low;
+    const int offset = high - gfu_boost;
+    const int qdiff = high_motion_minq[q] - low_motion_minq[q];
+    const int adjustment = ((offset * qdiff) + (gap >> 1)) / gap;
+    return low_motion_minq[q] + adjustment;
+  }
+}
+
+static int get_kf_active_quality(const RATE_CONTROL *const rc, int q,
+                                 aom_bit_depth_t bit_depth) {
+  int *kf_low_motion_minq;
+  int *kf_high_motion_minq;
+  ASSIGN_MINQ_TABLE(bit_depth, kf_low_motion_minq);
+  ASSIGN_MINQ_TABLE(bit_depth, kf_high_motion_minq);
+  return get_active_quality(q, rc->kf_boost, kf_low, kf_high,
+                            kf_low_motion_minq, kf_high_motion_minq);
+}
+
+static int get_gf_active_quality(const RATE_CONTROL *const rc, int q,
+                                 aom_bit_depth_t bit_depth) {
+  int *arfgf_low_motion_minq;
+  int *arfgf_high_motion_minq;
+  ASSIGN_MINQ_TABLE(bit_depth, arfgf_low_motion_minq);
+  ASSIGN_MINQ_TABLE(bit_depth, arfgf_high_motion_minq);
+  return get_active_quality(q, rc->gfu_boost, gf_low, gf_high,
+                            arfgf_low_motion_minq, arfgf_high_motion_minq);
+}
+
+static int calc_active_worst_quality_one_pass_vbr(const AV1_COMP *cpi) {
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const unsigned int curr_frame = cpi->common.current_video_frame;
+  int active_worst_quality;
+
+  if (cpi->common.frame_type == KEY_FRAME) {
+    active_worst_quality =
+        curr_frame == 0 ? rc->worst_quality : rc->last_q[KEY_FRAME] * 2;
+  } else {
+    if (!rc->is_src_frame_alt_ref &&
+        (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+      active_worst_quality = curr_frame == 1 ? rc->last_q[KEY_FRAME] * 5 / 4
+                                             : rc->last_q[INTER_FRAME];
+    } else {
+      active_worst_quality = curr_frame == 1 ? rc->last_q[KEY_FRAME] * 2
+                                             : rc->last_q[INTER_FRAME] * 2;
+    }
+  }
+  return AOMMIN(active_worst_quality, rc->worst_quality);
+}
+
+// Adjust active_worst_quality level based on buffer level.
+static int calc_active_worst_quality_one_pass_cbr(const AV1_COMP *cpi) {
+  // Adjust active_worst_quality: If buffer is above the optimal/target level,
+  // bring active_worst_quality down depending on fullness of buffer.
+  // If buffer is below the optimal level, let the active_worst_quality go from
+  // ambient Q (at buffer = optimal level) to worst_quality level
+  // (at buffer = critical level).
+  const AV1_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *rc = &cpi->rc;
+  // Buffer level below which we push active_worst to worst_quality.
+  int64_t critical_level = rc->optimal_buffer_level >> 3;
+  int64_t buff_lvl_step = 0;
+  int adjustment = 0;
+  int active_worst_quality;
+  int ambient_qp;
+  if (cm->frame_type == KEY_FRAME) return rc->worst_quality;
+  // For ambient_qp we use minimum of avg_frame_qindex[KEY_FRAME/INTER_FRAME]
+  // for the first few frames following key frame. These are both initialized
+  // to worst_quality and updated with (3/4, 1/4) average in postencode_update.
+  // So for first few frames following key, the qp of that key frame is weighted
+  // into the active_worst_quality setting.
+  ambient_qp = (cm->current_video_frame < 5)
+                   ? AOMMIN(rc->avg_frame_qindex[INTER_FRAME],
+                            rc->avg_frame_qindex[KEY_FRAME])
+                   : rc->avg_frame_qindex[INTER_FRAME];
+  active_worst_quality = AOMMIN(rc->worst_quality, ambient_qp * 5 / 4);
+  if (rc->buffer_level > rc->optimal_buffer_level) {
+    // Adjust down.
+    // Maximum limit for down adjustment, ~30%.
+    int max_adjustment_down = active_worst_quality / 3;
+    if (max_adjustment_down) {
+      buff_lvl_step = ((rc->maximum_buffer_size - rc->optimal_buffer_level) /
+                       max_adjustment_down);
+      if (buff_lvl_step)
+        adjustment = (int)((rc->buffer_level - rc->optimal_buffer_level) /
+                           buff_lvl_step);
+      active_worst_quality -= adjustment;
+    }
+  } else if (rc->buffer_level > critical_level) {
+    // Adjust up from ambient Q.
+    if (critical_level) {
+      buff_lvl_step = (rc->optimal_buffer_level - critical_level);
+      if (buff_lvl_step) {
+        adjustment = (int)((rc->worst_quality - ambient_qp) *
+                           (rc->optimal_buffer_level - rc->buffer_level) /
+                           buff_lvl_step);
+      }
+      active_worst_quality = ambient_qp + adjustment;
+    }
+  } else {
+    // Set to worst_quality if buffer is below critical level.
+    active_worst_quality = rc->worst_quality;
+  }
+  return active_worst_quality;
+}
+
+static int rc_pick_q_and_bounds_one_pass_cbr(const AV1_COMP *cpi,
+                                             int *bottom_index,
+                                             int *top_index) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  int active_best_quality;
+  int active_worst_quality = calc_active_worst_quality_one_pass_cbr(cpi);
+  int q;
+  int *rtc_minq;
+  ASSIGN_MINQ_TABLE(cm->bit_depth, rtc_minq);
+
+  if (frame_is_intra_only(cm)) {
+    active_best_quality = rc->best_quality;
+    // Handle the special case for key frames forced when we have reached
+    // the maximum key frame interval. Here force the Q to a range
+    // based on the ambient Q to reduce the risk of popping.
+    if (rc->this_key_frame_forced) {
+      int qindex = rc->last_boosted_qindex;
+      double last_boosted_q = av1_convert_qindex_to_q(qindex, cm->bit_depth);
+      int delta_qindex = av1_compute_qdelta(
+          rc, last_boosted_q, (last_boosted_q * 0.75), cm->bit_depth);
+      active_best_quality = AOMMAX(qindex + delta_qindex, rc->best_quality);
+    } else if (cm->current_video_frame > 0) {
+      // not first frame of one pass and kf_boost is set
+      double q_adj_factor = 1.0;
+      double q_val;
+
+      active_best_quality = get_kf_active_quality(
+          rc, rc->avg_frame_qindex[KEY_FRAME], cm->bit_depth);
+
+      // Allow somewhat lower kf minq with small image formats.
+      if ((cm->width * cm->height) <= (352 * 288)) {
+        q_adj_factor -= 0.25;
+      }
+
+      // Convert the adjustment factor to a qindex delta
+      // on active_best_quality.
+      q_val = av1_convert_qindex_to_q(active_best_quality, cm->bit_depth);
+      active_best_quality +=
+          av1_compute_qdelta(rc, q_val, q_val * q_adj_factor, cm->bit_depth);
+    }
+  } else if (!rc->is_src_frame_alt_ref &&
+             (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+    // Use the lower of active_worst_quality and recent
+    // average Q as basis for GF/ARF best Q limit unless last frame was
+    // a key frame.
+    if (rc->frames_since_key > 1 &&
+        rc->avg_frame_qindex[INTER_FRAME] < active_worst_quality) {
+      q = rc->avg_frame_qindex[INTER_FRAME];
+    } else {
+      q = active_worst_quality;
+    }
+    active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+  } else {
+    // Use the lower of active_worst_quality and recent/average Q.
+    if (cm->current_video_frame > 1) {
+      if (rc->avg_frame_qindex[INTER_FRAME] < active_worst_quality)
+        active_best_quality = rtc_minq[rc->avg_frame_qindex[INTER_FRAME]];
+      else
+        active_best_quality = rtc_minq[active_worst_quality];
+    } else {
+      if (rc->avg_frame_qindex[KEY_FRAME] < active_worst_quality)
+        active_best_quality = rtc_minq[rc->avg_frame_qindex[KEY_FRAME]];
+      else
+        active_best_quality = rtc_minq[active_worst_quality];
+    }
+  }
+
+  // Clip the active best and worst quality values to limits
+  active_best_quality =
+      clamp(active_best_quality, rc->best_quality, rc->worst_quality);
+  active_worst_quality =
+      clamp(active_worst_quality, active_best_quality, rc->worst_quality);
+
+  *top_index = active_worst_quality;
+  *bottom_index = active_best_quality;
+
+  // Limit Q range for the adaptive loop.
+  if (cm->frame_type == KEY_FRAME && !rc->this_key_frame_forced &&
+      !(cm->current_video_frame == 0)) {
+    int qdelta = 0;
+    aom_clear_system_state();
+    qdelta = av1_compute_qdelta_by_rate(
+        &cpi->rc, cm->frame_type, active_worst_quality, 2.0, cm->bit_depth);
+    *top_index = active_worst_quality + qdelta;
+    *top_index = AOMMAX(*top_index, *bottom_index);
+  }
+
+  // Special case code to try and match quality with forced key frames
+  if (cm->frame_type == KEY_FRAME && rc->this_key_frame_forced) {
+    q = rc->last_boosted_qindex;
+  } else {
+    q = av1_rc_regulate_q(cpi, rc->this_frame_target, active_best_quality,
+                          active_worst_quality);
+    if (q > *top_index) {
+      // Special case when we are targeting the max allowed rate
+      if (rc->this_frame_target >= rc->max_frame_bandwidth)
+        *top_index = q;
+      else
+        q = *top_index;
+    }
+  }
+
+  assert(*top_index <= rc->worst_quality && *top_index >= rc->best_quality);
+  assert(*bottom_index <= rc->worst_quality &&
+         *bottom_index >= rc->best_quality);
+  assert(q <= rc->worst_quality && q >= rc->best_quality);
+  return q;
+}
+
+static int get_active_cq_level(const RATE_CONTROL *rc,
+                               const AV1EncoderConfig *const oxcf) {
+  static const double cq_adjust_threshold = 0.1;
+  int active_cq_level = oxcf->cq_level;
+  if (oxcf->rc_mode == AOM_CQ && rc->total_target_bits > 0) {
+    const double x = (double)rc->total_actual_bits / rc->total_target_bits;
+    if (x < cq_adjust_threshold) {
+      active_cq_level = (int)(active_cq_level * x / cq_adjust_threshold);
+    }
+  }
+  return active_cq_level;
+}
+
+static int rc_pick_q_and_bounds_one_pass_vbr(const AV1_COMP *cpi,
+                                             int *bottom_index,
+                                             int *top_index) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  const int cq_level = get_active_cq_level(rc, oxcf);
+  int active_best_quality;
+  int active_worst_quality = calc_active_worst_quality_one_pass_vbr(cpi);
+  int q;
+  int *inter_minq;
+  ASSIGN_MINQ_TABLE(cm->bit_depth, inter_minq);
+
+  if (frame_is_intra_only(cm)) {
+    if (oxcf->rc_mode == AOM_Q) {
+      const int qindex = cq_level;
+      const double q_val = av1_convert_qindex_to_q(qindex, cm->bit_depth);
+      const int delta_qindex =
+          av1_compute_qdelta(rc, q_val, q_val * 0.25, cm->bit_depth);
+      active_best_quality = AOMMAX(qindex + delta_qindex, rc->best_quality);
+    } else if (rc->this_key_frame_forced) {
+      const int qindex = rc->last_boosted_qindex;
+      const double last_boosted_q =
+          av1_convert_qindex_to_q(qindex, cm->bit_depth);
+      const int delta_qindex = av1_compute_qdelta(
+          rc, last_boosted_q, last_boosted_q * 0.75, cm->bit_depth);
+      active_best_quality = AOMMAX(qindex + delta_qindex, rc->best_quality);
+    } else {  // not first frame of one pass and kf_boost is set
+      double q_adj_factor = 1.0;
+
+      active_best_quality = get_kf_active_quality(
+          rc, rc->avg_frame_qindex[KEY_FRAME], cm->bit_depth);
+
+      // Allow somewhat lower kf minq with small image formats.
+      if ((cm->width * cm->height) <= (352 * 288)) {
+        q_adj_factor -= 0.25;
+      }
+
+      // Convert the adjustment factor to a qindex delta on active_best_quality.
+      {
+        const double q_val =
+            av1_convert_qindex_to_q(active_best_quality, cm->bit_depth);
+        active_best_quality +=
+            av1_compute_qdelta(rc, q_val, q_val * q_adj_factor, cm->bit_depth);
+      }
+    }
+  } else if (!rc->is_src_frame_alt_ref &&
+             (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+    // Use the lower of active_worst_quality and recent
+    // average Q as basis for GF/ARF best Q limit unless last frame was
+    // a key frame.
+    q = (rc->frames_since_key > 1 &&
+         rc->avg_frame_qindex[INTER_FRAME] < active_worst_quality)
+            ? rc->avg_frame_qindex[INTER_FRAME]
+            : rc->avg_frame_qindex[KEY_FRAME];
+    // For constrained quality dont allow Q less than the cq level
+    if (oxcf->rc_mode == AOM_CQ) {
+      if (q < cq_level) q = cq_level;
+      active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+      // Constrained quality use slightly lower active best.
+      active_best_quality = active_best_quality * 15 / 16;
+    } else if (oxcf->rc_mode == AOM_Q) {
+      const int qindex = cq_level;
+      const double q_val = av1_convert_qindex_to_q(qindex, cm->bit_depth);
+      const int delta_qindex =
+          (cpi->refresh_alt_ref_frame)
+              ? av1_compute_qdelta(rc, q_val, q_val * 0.40, cm->bit_depth)
+              : av1_compute_qdelta(rc, q_val, q_val * 0.50, cm->bit_depth);
+      active_best_quality = AOMMAX(qindex + delta_qindex, rc->best_quality);
+    } else {
+      active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+    }
+  } else {
+    if (oxcf->rc_mode == AOM_Q) {
+      const int qindex = cq_level;
+      const double q_val = av1_convert_qindex_to_q(qindex, cm->bit_depth);
+      const double delta_rate[FIXED_GF_INTERVAL] = { 0.50, 1.0, 0.85, 1.0,
+                                                     0.70, 1.0, 0.85, 1.0 };
+      const int delta_qindex = av1_compute_qdelta(
+          rc, q_val,
+          q_val * delta_rate[cm->current_video_frame % FIXED_GF_INTERVAL],
+          cm->bit_depth);
+      active_best_quality = AOMMAX(qindex + delta_qindex, rc->best_quality);
+    } else {
+      // Use the lower of active_worst_quality and recent/average Q.
+      active_best_quality = (cm->current_video_frame > 1)
+                                ? inter_minq[rc->avg_frame_qindex[INTER_FRAME]]
+                                : inter_minq[rc->avg_frame_qindex[KEY_FRAME]];
+      // For the constrained quality mode we don't want
+      // q to fall below the cq level.
+      if ((oxcf->rc_mode == AOM_CQ) && (active_best_quality < cq_level)) {
+        active_best_quality = cq_level;
+      }
+    }
+  }
+
+  // Clip the active best and worst quality values to limits
+  active_best_quality =
+      clamp(active_best_quality, rc->best_quality, rc->worst_quality);
+  active_worst_quality =
+      clamp(active_worst_quality, active_best_quality, rc->worst_quality);
+
+  *top_index = active_worst_quality;
+  *bottom_index = active_best_quality;
+
+  // Limit Q range for the adaptive loop.
+  {
+    int qdelta = 0;
+    aom_clear_system_state();
+    if (cm->frame_type == KEY_FRAME && !rc->this_key_frame_forced &&
+        !(cm->current_video_frame == 0)) {
+      qdelta = av1_compute_qdelta_by_rate(
+          &cpi->rc, cm->frame_type, active_worst_quality, 2.0, cm->bit_depth);
+    } else if (!rc->is_src_frame_alt_ref &&
+               (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+      qdelta = av1_compute_qdelta_by_rate(
+          &cpi->rc, cm->frame_type, active_worst_quality, 1.75, cm->bit_depth);
+    }
+    *top_index = active_worst_quality + qdelta;
+    *top_index = AOMMAX(*top_index, *bottom_index);
+  }
+
+  if (oxcf->rc_mode == AOM_Q) {
+    q = active_best_quality;
+    // Special case code to try and match quality with forced key frames
+  } else if ((cm->frame_type == KEY_FRAME) && rc->this_key_frame_forced) {
+    q = rc->last_boosted_qindex;
+  } else {
+    q = av1_rc_regulate_q(cpi, rc->this_frame_target, active_best_quality,
+                          active_worst_quality);
+    if (q > *top_index) {
+      // Special case when we are targeting the max allowed rate
+      if (rc->this_frame_target >= rc->max_frame_bandwidth)
+        *top_index = q;
+      else
+        q = *top_index;
+    }
+  }
+
+  assert(*top_index <= rc->worst_quality && *top_index >= rc->best_quality);
+  assert(*bottom_index <= rc->worst_quality &&
+         *bottom_index >= rc->best_quality);
+  assert(q <= rc->worst_quality && q >= rc->best_quality);
+  return q;
+}
+
+int av1_frame_type_qdelta(const AV1_COMP *cpi, int rf_level, int q) {
+  static const double rate_factor_deltas[RATE_FACTOR_LEVELS] = {
+    1.00,  // INTER_NORMAL
+#if CONFIG_EXT_REFS
+    0.80,  // INTER_LOW
+    1.50,  // INTER_HIGH
+    1.25,  // GF_ARF_LOW
+#else
+    1.00,  // INTER_HIGH
+    1.50,  // GF_ARF_LOW
+#endif     // CONFIG_EXT_REFS
+    2.00,  // GF_ARF_STD
+    2.00,  // KF_STD
+  };
+  static const FRAME_TYPE frame_type[RATE_FACTOR_LEVELS] =
+#if CONFIG_EXT_REFS
+      { INTER_FRAME, INTER_FRAME, INTER_FRAME,
+        INTER_FRAME, INTER_FRAME, KEY_FRAME };
+#else
+      { INTER_FRAME, INTER_FRAME, INTER_FRAME, INTER_FRAME, KEY_FRAME };
+#endif  // CONFIG_EXT_REFS
+  const AV1_COMMON *const cm = &cpi->common;
+  int qdelta =
+      av1_compute_qdelta_by_rate(&cpi->rc, frame_type[rf_level], q,
+                                 rate_factor_deltas[rf_level], cm->bit_depth);
+  return qdelta;
+}
+
+#define STATIC_MOTION_THRESH 95
+static int rc_pick_q_and_bounds_two_pass(const AV1_COMP *cpi, int *bottom_index,
+                                         int *top_index) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  const GF_GROUP *gf_group = &cpi->twopass.gf_group;
+  const int cq_level = get_active_cq_level(rc, oxcf);
+  int active_best_quality;
+  int active_worst_quality = cpi->twopass.active_worst_quality;
+  int q;
+  int *inter_minq;
+  ASSIGN_MINQ_TABLE(cm->bit_depth, inter_minq);
+
+  if (frame_is_intra_only(cm)) {
+    // Handle the special case for key frames forced when we have reached
+    // the maximum key frame interval. Here force the Q to a range
+    // based on the ambient Q to reduce the risk of popping.
+    if (rc->this_key_frame_forced) {
+      double last_boosted_q;
+      int delta_qindex;
+      int qindex;
+
+      if (cpi->twopass.last_kfgroup_zeromotion_pct >= STATIC_MOTION_THRESH) {
+        qindex = AOMMIN(rc->last_kf_qindex, rc->last_boosted_qindex);
+        active_best_quality = qindex;
+        last_boosted_q = av1_convert_qindex_to_q(qindex, cm->bit_depth);
+        delta_qindex = av1_compute_qdelta(rc, last_boosted_q,
+                                          last_boosted_q * 1.25, cm->bit_depth);
+        active_worst_quality =
+            AOMMIN(qindex + delta_qindex, active_worst_quality);
+      } else {
+        qindex = rc->last_boosted_qindex;
+        last_boosted_q = av1_convert_qindex_to_q(qindex, cm->bit_depth);
+        delta_qindex = av1_compute_qdelta(rc, last_boosted_q,
+                                          last_boosted_q * 0.75, cm->bit_depth);
+        active_best_quality = AOMMAX(qindex + delta_qindex, rc->best_quality);
+      }
+    } else {
+      // Not forced keyframe.
+      double q_adj_factor = 1.0;
+      double q_val;
+
+      // Baseline value derived from cpi->active_worst_quality and kf boost.
+      active_best_quality =
+          get_kf_active_quality(rc, active_worst_quality, cm->bit_depth);
+
+      // Allow somewhat lower kf minq with small image formats.
+      if ((cm->width * cm->height) <= (352 * 288)) {
+        q_adj_factor -= 0.25;
+      }
+
+      // Make a further adjustment based on the kf zero motion measure.
+      q_adj_factor += 0.05 - (0.001 * (double)cpi->twopass.kf_zeromotion_pct);
+
+      // Convert the adjustment factor to a qindex delta
+      // on active_best_quality.
+      q_val = av1_convert_qindex_to_q(active_best_quality, cm->bit_depth);
+      active_best_quality +=
+          av1_compute_qdelta(rc, q_val, q_val * q_adj_factor, cm->bit_depth);
+    }
+  } else if (!rc->is_src_frame_alt_ref &&
+             (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+    // Use the lower of active_worst_quality and recent
+    // average Q as basis for GF/ARF best Q limit unless last frame was
+    // a key frame.
+    if (rc->frames_since_key > 1 &&
+        rc->avg_frame_qindex[INTER_FRAME] < active_worst_quality) {
+      q = rc->avg_frame_qindex[INTER_FRAME];
+    } else {
+      q = active_worst_quality;
+    }
+    // For constrained quality dont allow Q less than the cq level
+    if (oxcf->rc_mode == AOM_CQ) {
+      if (q < cq_level) q = cq_level;
+
+      active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+
+      // Constrained quality use slightly lower active best.
+      active_best_quality = active_best_quality * 15 / 16;
+
+    } else if (oxcf->rc_mode == AOM_Q) {
+      if (!cpi->refresh_alt_ref_frame) {
+        active_best_quality = cq_level;
+      } else {
+        active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+
+        // Modify best quality for second level arfs. For mode AOM_Q this
+        // becomes the baseline frame q.
+        if (gf_group->rf_level[gf_group->index] == GF_ARF_LOW)
+          active_best_quality = (active_best_quality + cq_level + 1) / 2;
+      }
+    } else {
+      active_best_quality = get_gf_active_quality(rc, q, cm->bit_depth);
+    }
+  } else {
+    if (oxcf->rc_mode == AOM_Q) {
+      active_best_quality = cq_level;
+    } else {
+      active_best_quality = inter_minq[active_worst_quality];
+
+      // For the constrained quality mode we don't want
+      // q to fall below the cq level.
+      if ((oxcf->rc_mode == AOM_CQ) && (active_best_quality < cq_level)) {
+        active_best_quality = cq_level;
+      }
+    }
+  }
+
+  // Extension to max or min Q if undershoot or overshoot is outside
+  // the permitted range.
+  if ((cpi->oxcf.rc_mode != AOM_Q) &&
+      (cpi->twopass.gf_zeromotion_pct < VLOW_MOTION_THRESHOLD)) {
+    if (frame_is_intra_only(cm) ||
+        (!rc->is_src_frame_alt_ref &&
+         (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame))) {
+      active_best_quality -=
+          (cpi->twopass.extend_minq + cpi->twopass.extend_minq_fast);
+      active_worst_quality += (cpi->twopass.extend_maxq / 2);
+    } else {
+      active_best_quality -=
+          (cpi->twopass.extend_minq + cpi->twopass.extend_minq_fast) / 2;
+      active_worst_quality += cpi->twopass.extend_maxq;
+    }
+  }
+
+  aom_clear_system_state();
+  // Static forced key frames Q restrictions dealt with elsewhere.
+  if (!(frame_is_intra_only(cm)) || !rc->this_key_frame_forced ||
+      (cpi->twopass.last_kfgroup_zeromotion_pct < STATIC_MOTION_THRESH)) {
+    int qdelta = av1_frame_type_qdelta(cpi, gf_group->rf_level[gf_group->index],
+                                       active_worst_quality);
+    active_worst_quality =
+        AOMMAX(active_worst_quality + qdelta, active_best_quality);
+  }
+
+  // Modify active_best_quality for downscaled normal frames.
+  if (rc->frame_size_selector != UNSCALED && !frame_is_kf_gf_arf(cpi)) {
+    int qdelta = av1_compute_qdelta_by_rate(
+        rc, cm->frame_type, active_best_quality, 2.0, cm->bit_depth);
+    active_best_quality =
+        AOMMAX(active_best_quality + qdelta, rc->best_quality);
+  }
+
+  active_best_quality =
+      clamp(active_best_quality, rc->best_quality, rc->worst_quality);
+  active_worst_quality =
+      clamp(active_worst_quality, active_best_quality, rc->worst_quality);
+
+  if (oxcf->rc_mode == AOM_Q) {
+    q = active_best_quality;
+    // Special case code to try and match quality with forced key frames.
+  } else if (frame_is_intra_only(cm) && rc->this_key_frame_forced) {
+    // If static since last kf use better of last boosted and last kf q.
+    if (cpi->twopass.last_kfgroup_zeromotion_pct >= STATIC_MOTION_THRESH) {
+      q = AOMMIN(rc->last_kf_qindex, rc->last_boosted_qindex);
+    } else {
+      q = rc->last_boosted_qindex;
+    }
+  } else {
+    q = av1_rc_regulate_q(cpi, rc->this_frame_target, active_best_quality,
+                          active_worst_quality);
+    if (q > active_worst_quality) {
+      // Special case when we are targeting the max allowed rate.
+      if (rc->this_frame_target >= rc->max_frame_bandwidth)
+        active_worst_quality = q;
+      else
+        q = active_worst_quality;
+    }
+  }
+  clamp(q, active_best_quality, active_worst_quality);
+
+  *top_index = active_worst_quality;
+  *bottom_index = active_best_quality;
+
+  assert(*top_index <= rc->worst_quality && *top_index >= rc->best_quality);
+  assert(*bottom_index <= rc->worst_quality &&
+         *bottom_index >= rc->best_quality);
+  assert(q <= rc->worst_quality && q >= rc->best_quality);
+  return q;
+}
+
+int av1_rc_pick_q_and_bounds(const AV1_COMP *cpi, int *bottom_index,
+                             int *top_index) {
+  int q;
+  if (cpi->oxcf.pass == 0) {
+    if (cpi->oxcf.rc_mode == AOM_CBR)
+      q = rc_pick_q_and_bounds_one_pass_cbr(cpi, bottom_index, top_index);
+    else
+      q = rc_pick_q_and_bounds_one_pass_vbr(cpi, bottom_index, top_index);
+  } else {
+    q = rc_pick_q_and_bounds_two_pass(cpi, bottom_index, top_index);
+  }
+
+  return q;
+}
+
+void av1_rc_compute_frame_size_bounds(const AV1_COMP *cpi, int frame_target,
+                                      int *frame_under_shoot_limit,
+                                      int *frame_over_shoot_limit) {
+  if (cpi->oxcf.rc_mode == AOM_Q) {
+    *frame_under_shoot_limit = 0;
+    *frame_over_shoot_limit = INT_MAX;
+  } else {
+    // For very small rate targets where the fractional adjustment
+    // may be tiny make sure there is at least a minimum range.
+    const int tolerance = (cpi->sf.recode_tolerance * frame_target) / 100;
+    *frame_under_shoot_limit = AOMMAX(frame_target - tolerance - 200, 0);
+    *frame_over_shoot_limit =
+        AOMMIN(frame_target + tolerance + 200, cpi->rc.max_frame_bandwidth);
+  }
+}
+
+void av1_rc_set_frame_target(AV1_COMP *cpi, int target) {
+  const AV1_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+
+  rc->this_frame_target = target;
+
+  // Modify frame size target when down-scaling.
+  if (cpi->oxcf.resize_mode == RESIZE_DYNAMIC &&
+      rc->frame_size_selector != UNSCALED)
+    rc->this_frame_target = (int)(rc->this_frame_target *
+                                  rate_thresh_mult[rc->frame_size_selector]);
+
+  // Target rate per SB64 (including partial SB64s.
+  rc->sb64_target_rate = (int)((int64_t)rc->this_frame_target * 64 * 64) /
+                         (cm->width * cm->height);
+}
+
+static void update_alt_ref_frame_stats(AV1_COMP *cpi) {
+  // this frame refreshes means next frames don't unless specified by user
+  RATE_CONTROL *const rc = &cpi->rc;
+  rc->frames_since_golden = 0;
+
+  // Mark the alt ref as done (setting to 0 means no further alt refs pending).
+  rc->source_alt_ref_pending = 0;
+
+  // Set the alternate reference frame active flag
+  rc->source_alt_ref_active = 1;
+}
+
+static void update_golden_frame_stats(AV1_COMP *cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+
+#if CONFIG_EXT_REFS
+  // Update the Golden frame usage counts.
+  // NOTE(weitinglin): If we use show_existing_frame for an OVERLAY frame,
+  //                   only the virtual indices for the reference frame will be
+  //                   updated and cpi->refresh_golden_frame will still be zero.
+  if (cpi->refresh_golden_frame || rc->is_src_frame_alt_ref) {
+#else
+  // Update the Golden frame usage counts.
+  if (cpi->refresh_golden_frame) {
+#endif  // CONFIG_EXT_REFS
+
+#if CONFIG_EXT_REFS
+    // We will not use internal overlay frames to replace the golden frame
+    if (!rc->is_src_frame_ext_arf)
+#endif  // CONFIG_EXT_REFS
+      // this frame refreshes means next frames don't unless specified by user
+      rc->frames_since_golden = 0;
+
+    // If we are not using alt ref in the up and coming group clear the arf
+    // active flag. In multi arf group case, if the index is not 0 then
+    // we are overlaying a mid group arf so should not reset the flag.
+    if (cpi->oxcf.pass == 2) {
+      if (!rc->source_alt_ref_pending && (cpi->twopass.gf_group.index == 0))
+        rc->source_alt_ref_active = 0;
+    } else if (!rc->source_alt_ref_pending) {
+      rc->source_alt_ref_active = 0;
+    }
+
+    // Decrement count down till next gf
+    if (rc->frames_till_gf_update_due > 0) rc->frames_till_gf_update_due--;
+
+  } else if (!cpi->refresh_alt_ref_frame) {
+    // Decrement count down till next gf
+    if (rc->frames_till_gf_update_due > 0) rc->frames_till_gf_update_due--;
+
+    rc->frames_since_golden++;
+  }
+}
+
+void av1_rc_postencode_update(AV1_COMP *cpi, uint64_t bytes_used) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  RATE_CONTROL *const rc = &cpi->rc;
+  const int qindex = cm->base_qindex;
+
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
+    av1_cyclic_refresh_postencode(cpi);
+  }
+
+  // Update rate control heuristics
+  rc->projected_frame_size = (int)(bytes_used << 3);
+
+  // Post encode loop adjustment of Q prediction.
+  av1_rc_update_rate_correction_factors(cpi);
+
+  // Keep a record of last Q and ambient average Q.
+  if (cm->frame_type == KEY_FRAME) {
+    rc->last_q[KEY_FRAME] = qindex;
+    rc->avg_frame_qindex[KEY_FRAME] =
+        ROUND_POWER_OF_TWO(3 * rc->avg_frame_qindex[KEY_FRAME] + qindex, 2);
+  } else {
+    if (!rc->is_src_frame_alt_ref &&
+        !(cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)) {
+      rc->last_q[INTER_FRAME] = qindex;
+      rc->avg_frame_qindex[INTER_FRAME] =
+          ROUND_POWER_OF_TWO(3 * rc->avg_frame_qindex[INTER_FRAME] + qindex, 2);
+      rc->ni_frames++;
+      rc->tot_q += av1_convert_qindex_to_q(qindex, cm->bit_depth);
+      rc->avg_q = rc->tot_q / rc->ni_frames;
+      // Calculate the average Q for normal inter frames (not key or GFU
+      // frames).
+      rc->ni_tot_qi += qindex;
+      rc->ni_av_qi = rc->ni_tot_qi / rc->ni_frames;
+    }
+  }
+
+  // Keep record of last boosted (KF/GF/ARF) Q value.
+  // If the current frame is coded at a lower Q then we also update it.
+  // If all mbs in this group are skipped only update if the Q value is
+  // better than that already stored.
+  // This is used to help set quality in forced key frames to reduce popping
+  if ((qindex < rc->last_boosted_qindex) || (cm->frame_type == KEY_FRAME) ||
+      (!rc->constrained_gf_group &&
+       (cpi->refresh_alt_ref_frame ||
+        (cpi->refresh_golden_frame && !rc->is_src_frame_alt_ref)))) {
+    rc->last_boosted_qindex = qindex;
+  }
+  if (cm->frame_type == KEY_FRAME) rc->last_kf_qindex = qindex;
+
+  update_buffer_level(cpi, rc->projected_frame_size);
+
+  // Rolling monitors of whether we are over or underspending used to help
+  // regulate min and Max Q in two pass.
+  if (cm->frame_type != KEY_FRAME) {
+    rc->rolling_target_bits = ROUND_POWER_OF_TWO(
+        rc->rolling_target_bits * 3 + rc->this_frame_target, 2);
+    rc->rolling_actual_bits = ROUND_POWER_OF_TWO(
+        rc->rolling_actual_bits * 3 + rc->projected_frame_size, 2);
+    rc->long_rolling_target_bits = ROUND_POWER_OF_TWO(
+        rc->long_rolling_target_bits * 31 + rc->this_frame_target, 5);
+    rc->long_rolling_actual_bits = ROUND_POWER_OF_TWO(
+        rc->long_rolling_actual_bits * 31 + rc->projected_frame_size, 5);
+  }
+
+  // Actual bits spent
+  rc->total_actual_bits += rc->projected_frame_size;
+#if CONFIG_EXT_REFS
+  rc->total_target_bits +=
+      (cm->show_frame || rc->is_bwd_ref_frame) ? rc->avg_frame_bandwidth : 0;
+#else
+  rc->total_target_bits += cm->show_frame ? rc->avg_frame_bandwidth : 0;
+#endif  // CONFIG_EXT_REFS
+
+  rc->total_target_vs_actual = rc->total_actual_bits - rc->total_target_bits;
+
+  if (is_altref_enabled(cpi) && cpi->refresh_alt_ref_frame &&
+      (cm->frame_type != KEY_FRAME))
+    // Update the alternate reference frame stats as appropriate.
+    update_alt_ref_frame_stats(cpi);
+  else
+    // Update the Golden frame stats as appropriate.
+    update_golden_frame_stats(cpi);
+
+  if (cm->frame_type == KEY_FRAME) rc->frames_since_key = 0;
+
+#if CONFIG_EXT_REFS
+  if (cm->show_frame || rc->is_bwd_ref_frame) {
+#else
+  if (cm->show_frame) {
+#endif  // CONFIG_EXT_REFS
+    rc->frames_since_key++;
+    rc->frames_to_key--;
+  }
+
+  // Trigger the resizing of the next frame if it is scaled.
+  if (oxcf->pass != 0) {
+    cpi->resize_pending =
+        rc->next_frame_size_selector != rc->frame_size_selector;
+    rc->frame_size_selector = rc->next_frame_size_selector;
+  }
+}
+
+void av1_rc_postencode_update_drop_frame(AV1_COMP *cpi) {
+  // Update buffer level with zero size, update frame counters, and return.
+  update_buffer_level(cpi, 0);
+  cpi->rc.frames_since_key++;
+  cpi->rc.frames_to_key--;
+  cpi->rc.rc_2_frame = 0;
+  cpi->rc.rc_1_frame = 0;
+}
+
+// Use this macro to turn on/off use of alt-refs in one-pass mode.
+#define USE_ALTREF_FOR_ONE_PASS 1
+
+static int calc_pframe_target_size_one_pass_vbr(const AV1_COMP *const cpi) {
+  static const int af_ratio = 10;
+  const RATE_CONTROL *const rc = &cpi->rc;
+  int target;
+#if USE_ALTREF_FOR_ONE_PASS
+  target =
+      (!rc->is_src_frame_alt_ref &&
+       (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame))
+          ? (rc->avg_frame_bandwidth * rc->baseline_gf_interval * af_ratio) /
+                (rc->baseline_gf_interval + af_ratio - 1)
+          : (rc->avg_frame_bandwidth * rc->baseline_gf_interval) /
+                (rc->baseline_gf_interval + af_ratio - 1);
+#else
+  target = rc->avg_frame_bandwidth;
+#endif
+  return av1_rc_clamp_pframe_target_size(cpi, target);
+}
+
+static int calc_iframe_target_size_one_pass_vbr(const AV1_COMP *const cpi) {
+  static const int kf_ratio = 25;
+  const RATE_CONTROL *rc = &cpi->rc;
+  const int target = rc->avg_frame_bandwidth * kf_ratio;
+  return av1_rc_clamp_iframe_target_size(cpi, target);
+}
+
+void av1_rc_get_one_pass_vbr_params(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int target;
+  // TODO(yaowu): replace the "auto_key && 0" below with proper decision logic.
+  if (!cpi->refresh_alt_ref_frame &&
+      (cm->current_video_frame == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY) ||
+       rc->frames_to_key == 0 || (cpi->oxcf.auto_key && 0))) {
+    cm->frame_type = KEY_FRAME;
+    rc->this_key_frame_forced =
+        cm->current_video_frame != 0 && rc->frames_to_key == 0;
+    rc->frames_to_key = cpi->oxcf.key_freq;
+    rc->kf_boost = DEFAULT_KF_BOOST;
+    rc->source_alt_ref_active = 0;
+  } else {
+    cm->frame_type = INTER_FRAME;
+  }
+  if (rc->frames_till_gf_update_due == 0) {
+    rc->baseline_gf_interval = (rc->min_gf_interval + rc->max_gf_interval) / 2;
+    rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+    // NOTE: frames_till_gf_update_due must be <= frames_to_key.
+    if (rc->frames_till_gf_update_due > rc->frames_to_key) {
+      rc->frames_till_gf_update_due = rc->frames_to_key;
+      rc->constrained_gf_group = 1;
+    } else {
+      rc->constrained_gf_group = 0;
+    }
+    cpi->refresh_golden_frame = 1;
+    rc->source_alt_ref_pending = USE_ALTREF_FOR_ONE_PASS;
+    rc->gfu_boost = DEFAULT_GF_BOOST;
+  }
+  if (cm->frame_type == KEY_FRAME)
+    target = calc_iframe_target_size_one_pass_vbr(cpi);
+  else
+    target = calc_pframe_target_size_one_pass_vbr(cpi);
+  av1_rc_set_frame_target(cpi, target);
+}
+
+static int calc_pframe_target_size_one_pass_cbr(const AV1_COMP *cpi) {
+  const AV1EncoderConfig *oxcf = &cpi->oxcf;
+  const RATE_CONTROL *rc = &cpi->rc;
+  const int64_t diff = rc->optimal_buffer_level - rc->buffer_level;
+  const int64_t one_pct_bits = 1 + rc->optimal_buffer_level / 100;
+  int min_frame_target =
+      AOMMAX(rc->avg_frame_bandwidth >> 4, FRAME_OVERHEAD_BITS);
+  int target;
+
+  if (oxcf->gf_cbr_boost_pct) {
+    const int af_ratio_pct = oxcf->gf_cbr_boost_pct + 100;
+    target = cpi->refresh_golden_frame
+                 ? (rc->avg_frame_bandwidth * rc->baseline_gf_interval *
+                    af_ratio_pct) /
+                       (rc->baseline_gf_interval * 100 + af_ratio_pct - 100)
+                 : (rc->avg_frame_bandwidth * rc->baseline_gf_interval * 100) /
+                       (rc->baseline_gf_interval * 100 + af_ratio_pct - 100);
+  } else {
+    target = rc->avg_frame_bandwidth;
+  }
+
+  if (diff > 0) {
+    // Lower the target bandwidth for this frame.
+    const int pct_low = (int)AOMMIN(diff / one_pct_bits, oxcf->under_shoot_pct);
+    target -= (target * pct_low) / 200;
+  } else if (diff < 0) {
+    // Increase the target bandwidth for this frame.
+    const int pct_high =
+        (int)AOMMIN(-diff / one_pct_bits, oxcf->over_shoot_pct);
+    target += (target * pct_high) / 200;
+  }
+  if (oxcf->rc_max_inter_bitrate_pct) {
+    const int max_rate =
+        rc->avg_frame_bandwidth * oxcf->rc_max_inter_bitrate_pct / 100;
+    target = AOMMIN(target, max_rate);
+  }
+  return AOMMAX(min_frame_target, target);
+}
+
+static int calc_iframe_target_size_one_pass_cbr(const AV1_COMP *cpi) {
+  const RATE_CONTROL *rc = &cpi->rc;
+  int target;
+  if (cpi->common.current_video_frame == 0) {
+    target = ((rc->starting_buffer_level / 2) > INT_MAX)
+                 ? INT_MAX
+                 : (int)(rc->starting_buffer_level / 2);
+  } else {
+    int kf_boost = 32;
+    double framerate = cpi->framerate;
+
+    kf_boost = AOMMAX(kf_boost, (int)(2 * framerate - 16));
+    if (rc->frames_since_key < framerate / 2) {
+      kf_boost = (int)(kf_boost * rc->frames_since_key / (framerate / 2));
+    }
+    target = ((16 + kf_boost) * rc->avg_frame_bandwidth) >> 4;
+  }
+  return av1_rc_clamp_iframe_target_size(cpi, target);
+}
+
+void av1_rc_get_one_pass_cbr_params(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int target;
+  // TODO(yaowu): replace the "auto_key && 0" below with proper decision logic.
+  if ((cm->current_video_frame == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY) ||
+       rc->frames_to_key == 0 || (cpi->oxcf.auto_key && 0))) {
+    cm->frame_type = KEY_FRAME;
+    rc->this_key_frame_forced =
+        cm->current_video_frame != 0 && rc->frames_to_key == 0;
+    rc->frames_to_key = cpi->oxcf.key_freq;
+    rc->kf_boost = DEFAULT_KF_BOOST;
+    rc->source_alt_ref_active = 0;
+  } else {
+    cm->frame_type = INTER_FRAME;
+  }
+  if (rc->frames_till_gf_update_due == 0) {
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
+      av1_cyclic_refresh_set_golden_update(cpi);
+    else
+      rc->baseline_gf_interval =
+          (rc->min_gf_interval + rc->max_gf_interval) / 2;
+    rc->frames_till_gf_update_due = rc->baseline_gf_interval;
+    // NOTE: frames_till_gf_update_due must be <= frames_to_key.
+    if (rc->frames_till_gf_update_due > rc->frames_to_key)
+      rc->frames_till_gf_update_due = rc->frames_to_key;
+    cpi->refresh_golden_frame = 1;
+    rc->gfu_boost = DEFAULT_GF_BOOST;
+  }
+
+  // Any update/change of global cyclic refresh parameters (amount/delta-qp)
+  // should be done here, before the frame qp is selected.
+  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
+    av1_cyclic_refresh_update_parameters(cpi);
+
+  if (cm->frame_type == KEY_FRAME)
+    target = calc_iframe_target_size_one_pass_cbr(cpi);
+  else
+    target = calc_pframe_target_size_one_pass_cbr(cpi);
+
+  av1_rc_set_frame_target(cpi, target);
+  if (cpi->oxcf.resize_mode == RESIZE_DYNAMIC)
+    cpi->resize_pending = av1_resize_one_pass_cbr(cpi);
+  else
+    cpi->resize_pending = 0;
+}
+
+int av1_compute_qdelta(const RATE_CONTROL *rc, double qstart, double qtarget,
+                       aom_bit_depth_t bit_depth) {
+  int start_index = rc->worst_quality;
+  int target_index = rc->worst_quality;
+  int i;
+
+  // Convert the average q value to an index.
+  for (i = rc->best_quality; i < rc->worst_quality; ++i) {
+    start_index = i;
+    if (av1_convert_qindex_to_q(i, bit_depth) >= qstart) break;
+  }
+
+  // Convert the q target to an index
+  for (i = rc->best_quality; i < rc->worst_quality; ++i) {
+    target_index = i;
+    if (av1_convert_qindex_to_q(i, bit_depth) >= qtarget) break;
+  }
+
+  return target_index - start_index;
+}
+
+int av1_compute_qdelta_by_rate(const RATE_CONTROL *rc, FRAME_TYPE frame_type,
+                               int qindex, double rate_target_ratio,
+                               aom_bit_depth_t bit_depth) {
+  int target_index = rc->worst_quality;
+  int i;
+
+  // Look up the current projected bits per block for the base index
+  const int base_bits_per_mb =
+      av1_rc_bits_per_mb(frame_type, qindex, 1.0, bit_depth);
+
+  // Find the target bits per mb based on the base value and given ratio.
+  const int target_bits_per_mb = (int)(rate_target_ratio * base_bits_per_mb);
+
+  // Convert the q target to an index
+  for (i = rc->best_quality; i < rc->worst_quality; ++i) {
+    if (av1_rc_bits_per_mb(frame_type, i, 1.0, bit_depth) <=
+        target_bits_per_mb) {
+      target_index = i;
+      break;
+    }
+  }
+  return target_index - qindex;
+}
+
+void av1_rc_set_gf_interval_range(const AV1_COMP *const cpi,
+                                  RATE_CONTROL *const rc) {
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+
+  // Special case code for 1 pass fixed Q mode tests
+  if ((oxcf->pass == 0) && (oxcf->rc_mode == AOM_Q)) {
+    rc->max_gf_interval = FIXED_GF_INTERVAL;
+    rc->min_gf_interval = FIXED_GF_INTERVAL;
+    rc->static_scene_max_gf_interval = FIXED_GF_INTERVAL;
+  } else {
+    // Set Maximum gf/arf interval
+    rc->max_gf_interval = oxcf->max_gf_interval;
+    rc->min_gf_interval = oxcf->min_gf_interval;
+    if (rc->min_gf_interval == 0)
+      rc->min_gf_interval = av1_rc_get_default_min_gf_interval(
+          oxcf->width, oxcf->height, cpi->framerate);
+    if (rc->max_gf_interval == 0)
+      rc->max_gf_interval = av1_rc_get_default_max_gf_interval(
+          cpi->framerate, rc->min_gf_interval);
+
+    // Extended interval for genuinely static scenes
+    rc->static_scene_max_gf_interval = MAX_LAG_BUFFERS * 2;
+
+    if (is_altref_enabled(cpi)) {
+      if (rc->static_scene_max_gf_interval > oxcf->lag_in_frames - 1)
+        rc->static_scene_max_gf_interval = oxcf->lag_in_frames - 1;
+    }
+
+    if (rc->max_gf_interval > rc->static_scene_max_gf_interval)
+      rc->max_gf_interval = rc->static_scene_max_gf_interval;
+
+    // Clamp min to max
+    rc->min_gf_interval = AOMMIN(rc->min_gf_interval, rc->max_gf_interval);
+  }
+}
+
+void av1_rc_update_framerate(AV1_COMP *cpi) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int vbr_max_bits;
+
+  rc->avg_frame_bandwidth = (int)(oxcf->target_bandwidth / cpi->framerate);
+  rc->min_frame_bandwidth =
+      (int)(rc->avg_frame_bandwidth * oxcf->two_pass_vbrmin_section / 100);
+
+  rc->min_frame_bandwidth =
+      AOMMAX(rc->min_frame_bandwidth, FRAME_OVERHEAD_BITS);
+
+  // A maximum bitrate for a frame is defined.
+  // The baseline for this aligns with HW implementations that
+  // can support decode of 1080P content up to a bitrate of MAX_MB_RATE bits
+  // per 16x16 MB (averaged over a frame). However this limit is extended if
+  // a very high rate is given on the command line or the the rate cannnot
+  // be acheived because of a user specificed max q (e.g. when the user
+  // specifies lossless encode.
+  vbr_max_bits =
+      (int)(((int64_t)rc->avg_frame_bandwidth * oxcf->two_pass_vbrmax_section) /
+            100);
+  rc->max_frame_bandwidth =
+      AOMMAX(AOMMAX((cm->MBs * MAX_MB_RATE), MAXRATE_1080P), vbr_max_bits);
+
+  av1_rc_set_gf_interval_range(cpi, rc);
+}
+
+#define VBR_PCT_ADJUSTMENT_LIMIT 50
+// For VBR...adjustment to the frame target based on error from previous frames
+static void vbr_rate_correction(AV1_COMP *cpi, int *this_frame_target) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  int64_t vbr_bits_off_target = rc->vbr_bits_off_target;
+  int max_delta;
+  double position_factor = 1.0;
+
+  // How far through the clip are we.
+  // This number is used to damp the per frame rate correction.
+  // Range 0 - 1.0
+  if (cpi->twopass.total_stats.count != 0.) {
+    position_factor = sqrt((double)cpi->common.current_video_frame /
+                           cpi->twopass.total_stats.count);
+  }
+  max_delta = (int)(position_factor *
+                    ((*this_frame_target * VBR_PCT_ADJUSTMENT_LIMIT) / 100));
+
+  // vbr_bits_off_target > 0 means we have extra bits to spend
+  if (vbr_bits_off_target > 0) {
+    *this_frame_target += (vbr_bits_off_target > max_delta)
+                              ? max_delta
+                              : (int)vbr_bits_off_target;
+  } else {
+    *this_frame_target -= (vbr_bits_off_target < -max_delta)
+                              ? max_delta
+                              : (int)-vbr_bits_off_target;
+  }
+
+  // Fast redistribution of bits arising from massive local undershoot.
+  // Dont do it for kf,arf,gf or overlay frames.
+  if (!frame_is_kf_gf_arf(cpi) && !rc->is_src_frame_alt_ref &&
+      rc->vbr_bits_off_target_fast) {
+    int one_frame_bits = AOMMAX(rc->avg_frame_bandwidth, *this_frame_target);
+    int fast_extra_bits;
+    fast_extra_bits = (int)AOMMIN(rc->vbr_bits_off_target_fast, one_frame_bits);
+    fast_extra_bits = (int)AOMMIN(
+        fast_extra_bits,
+        AOMMAX(one_frame_bits / 8, rc->vbr_bits_off_target_fast / 8));
+    *this_frame_target += (int)fast_extra_bits;
+    rc->vbr_bits_off_target_fast -= fast_extra_bits;
+  }
+}
+
+void av1_set_target_rate(AV1_COMP *cpi) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  int target_rate = rc->base_frame_target;
+
+  // Correction to rate target based on prior over or under shoot.
+  if (cpi->oxcf.rc_mode == AOM_VBR || cpi->oxcf.rc_mode == AOM_CQ)
+    vbr_rate_correction(cpi, &target_rate);
+  av1_rc_set_frame_target(cpi, target_rate);
+}
+
+// Check if we should resize, based on average QP from past x frames.
+// Only allow for resize at most one scale down for now, scaling factor is 2.
+int av1_resize_one_pass_cbr(AV1_COMP *cpi) {
+  const AV1_COMMON *const cm = &cpi->common;
+  RATE_CONTROL *const rc = &cpi->rc;
+  int resize_now = 0;
+  cpi->resize_scale_num = 1;
+  cpi->resize_scale_den = 1;
+  // Don't resize on key frame; reset the counters on key frame.
+  if (cm->frame_type == KEY_FRAME) {
+    cpi->resize_avg_qp = 0;
+    cpi->resize_count = 0;
+    return 0;
+  }
+  // Resize based on average buffer underflow and QP over some window.
+  // Ignore samples close to key frame, since QP is usually high after key.
+  if (cpi->rc.frames_since_key > 2 * cpi->framerate) {
+    const int window = (int)(5 * cpi->framerate);
+    cpi->resize_avg_qp += cm->base_qindex;
+    if (cpi->rc.buffer_level < (int)(30 * rc->optimal_buffer_level / 100))
+      ++cpi->resize_buffer_underflow;
+    ++cpi->resize_count;
+    // Check for resize action every "window" frames.
+    if (cpi->resize_count >= window) {
+      int avg_qp = cpi->resize_avg_qp / cpi->resize_count;
+      // Resize down if buffer level has underflowed sufficent amount in past
+      // window, and we are at original resolution.
+      // Resize back up if average QP is low, and we are currently in a resized
+      // down state.
+      if (cpi->resize_state == 0 &&
+          cpi->resize_buffer_underflow > (cpi->resize_count >> 2)) {
+        resize_now = 1;
+        cpi->resize_state = 1;
+      } else if (cpi->resize_state == 1 &&
+                 avg_qp < 40 * cpi->rc.worst_quality / 100) {
+        resize_now = -1;
+        cpi->resize_state = 0;
+      }
+      // Reset for next window measurement.
+      cpi->resize_avg_qp = 0;
+      cpi->resize_count = 0;
+      cpi->resize_buffer_underflow = 0;
+    }
+  }
+  // If decision is to resize, reset some quantities, and check is we should
+  // reduce rate correction factor,
+  if (resize_now != 0) {
+    int target_bits_per_frame;
+    int active_worst_quality;
+    int qindex;
+    int tot_scale_change;
+    // For now, resize is by 1/2 x 1/2.
+    cpi->resize_scale_num = 1;
+    cpi->resize_scale_den = 2;
+    tot_scale_change = (cpi->resize_scale_den * cpi->resize_scale_den) /
+                       (cpi->resize_scale_num * cpi->resize_scale_num);
+    // Reset buffer level to optimal, update target size.
+    rc->buffer_level = rc->optimal_buffer_level;
+    rc->bits_off_target = rc->optimal_buffer_level;
+    rc->this_frame_target = calc_pframe_target_size_one_pass_cbr(cpi);
+    // Reset cyclic refresh parameters.
+    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
+      av1_cyclic_refresh_reset_resize(cpi);
+    // Get the projected qindex, based on the scaled target frame size (scaled
+    // so target_bits_per_mb in av1_rc_regulate_q will be correct target).
+    target_bits_per_frame = (resize_now == 1)
+                                ? rc->this_frame_target * tot_scale_change
+                                : rc->this_frame_target / tot_scale_change;
+    active_worst_quality = calc_active_worst_quality_one_pass_cbr(cpi);
+    qindex = av1_rc_regulate_q(cpi, target_bits_per_frame, rc->best_quality,
+                               active_worst_quality);
+    // If resize is down, check if projected q index is close to worst_quality,
+    // and if so, reduce the rate correction factor (since likely can afford
+    // lower q for resized frame).
+    if (resize_now == 1 && qindex > 90 * cpi->rc.worst_quality / 100) {
+      rc->rate_correction_factors[INTER_NORMAL] *= 0.85;
+    }
+    // If resize is back up, check if projected q index is too much above the
+    // current base_qindex, and if so, reduce the rate correction factor
+    // (since prefer to keep q for resized frame at least close to previous q).
+    if (resize_now == -1 && qindex > 130 * cm->base_qindex / 100) {
+      rc->rate_correction_factors[INTER_NORMAL] *= 0.9;
+    }
+  }
+  return resize_now;
+}
diff --git a/third_party/aom/av1/encoder/ratectrl.h b/third_party/aom/av1/encoder/ratectrl.h
new file mode 100644
index 0000000000..93a9b49397
--- /dev/null
+++ b/third_party/aom/av1/encoder/ratectrl.h
@@ -0,0 +1,284 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_RATECTRL_H_
+#define AV1_ENCODER_RATECTRL_H_
+
+#include "aom/aom_codec.h"
+#include "aom/aom_integer.h"
+
+#include "av1/common/blockd.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Bits Per MB at different Q (Multiplied by 512)
+#define BPER_MB_NORMBITS 9
+
+#define MIN_GF_INTERVAL 4
+#define MAX_GF_INTERVAL 16
+#define FIXED_GF_INTERVAL 8  // Used in some testing modes only
+
+#if CONFIG_EXT_REFS
+typedef enum {
+  INTER_NORMAL = 0,
+  INTER_LOW = 1,
+  INTER_HIGH = 2,
+  GF_ARF_LOW = 3,
+  GF_ARF_STD = 4,
+  KF_STD = 5,
+  RATE_FACTOR_LEVELS = 6
+} RATE_FACTOR_LEVEL;
+#else
+typedef enum {
+  INTER_NORMAL = 0,
+  INTER_HIGH = 1,
+  GF_ARF_LOW = 2,
+  GF_ARF_STD = 3,
+  KF_STD = 4,
+  RATE_FACTOR_LEVELS = 5
+} RATE_FACTOR_LEVEL;
+#endif  // CONFIG_EXT_REFS
+
+// Internal frame scaling level.
+typedef enum {
+  UNSCALED = 0,     // Frame is unscaled.
+  SCALE_STEP1 = 1,  // First-level down-scaling.
+  FRAME_SCALE_STEPS
+} FRAME_SCALE_LEVEL;
+
+// Frame dimensions multiplier wrt the native frame size, in 1/16ths,
+// specified for the scale-up case.
+// e.g. 24 => 16/24 = 2/3 of native size. The restriction to 1/16th is
+// intended to match the capabilities of the normative scaling filters,
+// giving precedence to the up-scaling accuracy.
+static const int frame_scale_factor[FRAME_SCALE_STEPS] = { 16, 24 };
+
+// Multiplier of the target rate to be used as threshold for triggering scaling.
+static const double rate_thresh_mult[FRAME_SCALE_STEPS] = { 1.0, 2.0 };
+
+// Scale dependent Rate Correction Factor multipliers. Compensates for the
+// greater number of bits per pixel generated in down-scaled frames.
+static const double rcf_mult[FRAME_SCALE_STEPS] = { 1.0, 2.0 };
+
+typedef struct {
+  // Rate targetting variables
+  int base_frame_target;  // A baseline frame target before adjustment
+                          // for previous under or over shoot.
+  int this_frame_target;  // Actual frame target after rc adjustment.
+  int projected_frame_size;
+  int sb64_target_rate;
+  int last_q[FRAME_TYPES];  // Separate values for Intra/Inter
+  int last_boosted_qindex;  // Last boosted GF/KF/ARF q
+  int last_kf_qindex;       // Q index of the last key frame coded.
+
+  int gfu_boost;
+  int last_boost;
+  int kf_boost;
+
+  double rate_correction_factors[RATE_FACTOR_LEVELS];
+
+  int frames_since_golden;
+  int frames_till_gf_update_due;
+  int min_gf_interval;
+  int max_gf_interval;
+  int static_scene_max_gf_interval;
+  int baseline_gf_interval;
+  int constrained_gf_group;
+  int frames_to_key;
+  int frames_since_key;
+  int this_key_frame_forced;
+  int next_key_frame_forced;
+  int source_alt_ref_pending;
+  int source_alt_ref_active;
+  int is_src_frame_alt_ref;
+
+#if CONFIG_EXT_REFS
+  // Length of the bi-predictive frame group interval
+  int bipred_group_interval;
+
+  // NOTE: Different types of frames may have different bits allocated
+  //       accordingly, aiming to achieve the overall optimal RD performance.
+  int is_bwd_ref_frame;
+  int is_last_bipred_frame;
+  int is_bipred_frame;
+  int is_src_frame_ext_arf;
+#endif  // CONFIG_EXT_REFS
+
+  int avg_frame_bandwidth;  // Average frame size target for clip
+  int min_frame_bandwidth;  // Minimum allocation used for any frame
+  int max_frame_bandwidth;  // Maximum burst rate allowed for a frame.
+
+  int ni_av_qi;
+  int ni_tot_qi;
+  int ni_frames;
+  int avg_frame_qindex[FRAME_TYPES];
+  double tot_q;
+  double avg_q;
+
+  int64_t buffer_level;
+  int64_t bits_off_target;
+  int64_t vbr_bits_off_target;
+  int64_t vbr_bits_off_target_fast;
+
+  int decimation_factor;
+  int decimation_count;
+
+  int rolling_target_bits;
+  int rolling_actual_bits;
+
+  int long_rolling_target_bits;
+  int long_rolling_actual_bits;
+
+  int rate_error_estimate;
+
+  int64_t total_actual_bits;
+  int64_t total_target_bits;
+  int64_t total_target_vs_actual;
+
+  int worst_quality;
+  int best_quality;
+
+  int64_t starting_buffer_level;
+  int64_t optimal_buffer_level;
+  int64_t maximum_buffer_size;
+
+  // rate control history for last frame(1) and the frame before(2).
+  // -1: undershot
+  //  1: overshoot
+  //  0: not initialized.
+  int rc_1_frame;
+  int rc_2_frame;
+  int q_1_frame;
+  int q_2_frame;
+
+  // Auto frame-scaling variables.
+  FRAME_SCALE_LEVEL frame_size_selector;
+  FRAME_SCALE_LEVEL next_frame_size_selector;
+  int frame_width[FRAME_SCALE_STEPS];
+  int frame_height[FRAME_SCALE_STEPS];
+  int rf_level_maxq[RATE_FACTOR_LEVELS];
+} RATE_CONTROL;
+
+struct AV1_COMP;
+struct AV1EncoderConfig;
+
+void av1_rc_init(const struct AV1EncoderConfig *oxcf, int pass,
+                 RATE_CONTROL *rc);
+
+int av1_estimate_bits_at_q(FRAME_TYPE frame_kind, int q, int mbs,
+                           double correction_factor, aom_bit_depth_t bit_depth);
+
+double av1_convert_qindex_to_q(int qindex, aom_bit_depth_t bit_depth);
+
+void av1_rc_init_minq_luts(void);
+
+int av1_rc_get_default_min_gf_interval(int width, int height, double framerate);
+// Note av1_rc_get_default_max_gf_interval() requires the min_gf_interval to
+// be passed in to ensure that the max_gf_interval returned is at least as bis
+// as that.
+int av1_rc_get_default_max_gf_interval(double framerate, int min_frame_rate);
+
+// Generally at the high level, the following flow is expected
+// to be enforced for rate control:
+// First call per frame, one of:
+//   av1_rc_get_one_pass_vbr_params()
+//   av1_rc_get_one_pass_cbr_params()
+//   av1_rc_get_first_pass_params()
+//   av1_rc_get_second_pass_params()
+// depending on the usage to set the rate control encode parameters desired.
+//
+// Then, call encode_frame_to_data_rate() to perform the
+// actual encode. This function will in turn call encode_frame()
+// one or more times, followed by one of:
+//   av1_rc_postencode_update()
+//   av1_rc_postencode_update_drop_frame()
+//
+// The majority of rate control parameters are only expected
+// to be set in the av1_rc_get_..._params() functions and
+// updated during the av1_rc_postencode_update...() functions.
+// The only exceptions are av1_rc_drop_frame() and
+// av1_rc_update_rate_correction_factors() functions.
+
+// Functions to set parameters for encoding before the actual
+// encode_frame_to_data_rate() function.
+void av1_rc_get_one_pass_vbr_params(struct AV1_COMP *cpi);
+void av1_rc_get_one_pass_cbr_params(struct AV1_COMP *cpi);
+
+// Post encode update of the rate control parameters based
+// on bytes used
+void av1_rc_postencode_update(struct AV1_COMP *cpi, uint64_t bytes_used);
+// Post encode update of the rate control parameters for dropped frames
+void av1_rc_postencode_update_drop_frame(struct AV1_COMP *cpi);
+
+// Updates rate correction factors
+// Changes only the rate correction factors in the rate control structure.
+void av1_rc_update_rate_correction_factors(struct AV1_COMP *cpi);
+
+// Decide if we should drop this frame: For 1-pass CBR.
+// Changes only the decimation count in the rate control structure
+int av1_rc_drop_frame(struct AV1_COMP *cpi);
+
+// Computes frame size bounds.
+void av1_rc_compute_frame_size_bounds(const struct AV1_COMP *cpi,
+                                      int this_frame_target,
+                                      int *frame_under_shoot_limit,
+                                      int *frame_over_shoot_limit);
+
+// Picks q and q bounds given the target for bits
+int av1_rc_pick_q_and_bounds(const struct AV1_COMP *cpi, int *bottom_index,
+                             int *top_index);
+
+// Estimates q to achieve a target bits per frame
+int av1_rc_regulate_q(const struct AV1_COMP *cpi, int target_bits_per_frame,
+                      int active_best_quality, int active_worst_quality);
+
+// Estimates bits per mb for a given qindex and correction factor.
+int av1_rc_bits_per_mb(FRAME_TYPE frame_type, int qindex,
+                       double correction_factor, aom_bit_depth_t bit_depth);
+
+// Clamping utilities for bitrate targets for iframes and pframes.
+int av1_rc_clamp_iframe_target_size(const struct AV1_COMP *const cpi,
+                                    int target);
+int av1_rc_clamp_pframe_target_size(const struct AV1_COMP *const cpi,
+                                    int target);
+// Utility to set frame_target into the RATE_CONTROL structure
+// This function is called only from the av1_rc_get_..._params() functions.
+void av1_rc_set_frame_target(struct AV1_COMP *cpi, int target);
+
+// Computes a q delta (in "q index" terms) to get from a starting q value
+// to a target q value
+int av1_compute_qdelta(const RATE_CONTROL *rc, double qstart, double qtarget,
+                       aom_bit_depth_t bit_depth);
+
+// Computes a q delta (in "q index" terms) to get from a starting q value
+// to a value that should equate to the given rate ratio.
+int av1_compute_qdelta_by_rate(const RATE_CONTROL *rc, FRAME_TYPE frame_type,
+                               int qindex, double rate_target_ratio,
+                               aom_bit_depth_t bit_depth);
+
+int av1_frame_type_qdelta(const struct AV1_COMP *cpi, int rf_level, int q);
+
+void av1_rc_update_framerate(struct AV1_COMP *cpi);
+
+void av1_rc_set_gf_interval_range(const struct AV1_COMP *const cpi,
+                                  RATE_CONTROL *const rc);
+
+void av1_set_target_rate(struct AV1_COMP *cpi);
+
+int av1_resize_one_pass_cbr(struct AV1_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_RATECTRL_H_
diff --git a/third_party/aom/av1/encoder/ratectrl_xiph.c b/third_party/aom/av1/encoder/ratectrl_xiph.c
new file mode 100644
index 0000000000..b9f827528f
--- /dev/null
+++ b/third_party/aom/av1/encoder/ratectrl_xiph.c
@@ -0,0 +1,1244 @@
+/*
+ * Copyright (c) 2001-2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+#include "av1/common/odintrin.h"
+#include "av1/encoder/ratectrl_xiph.h"
+
+#define OD_Q57(v) ((int64_t)((uint64_t)(v) << 57))
+#define OD_F_Q45(v) ((int64_t)(((v) * ((int64_t)1 << 45))))
+#define OD_F_Q12(v) ((int32_t)(((v) * ((int32_t)1 << 12))))
+
+/*A rough lookup table for tan(x), 0 <= x < pi/2.
+  The values are Q12 fixed-point and spaced at 5 degree intervals.
+  These decisions are somewhat arbitrary, but sufficient for the 2nd order
+   Bessel follower below.
+  Values of x larger than 85 degrees are extrapolated from the last interval,
+   which is way off, but "good enough".*/
+static uint16_t OD_ROUGH_TAN_LOOKUP[18] = { 0,     358,   722,  1098, 1491,
+                                            1910,  2365,  2868, 3437, 4096,
+                                            4881,  5850,  7094, 8784, 11254,
+                                            15286, 23230, 46817 };
+
+/*alpha is Q24 in the range [0,0.5).
+  The return values is 5.12.*/
+static int od_warp_alpha(int alpha) {
+  int i;
+  int d;
+  int t0;
+  int t1;
+  i = alpha * 36 >> 24;
+  if (i >= 17) i = 16;
+  t0 = OD_ROUGH_TAN_LOOKUP[i];
+  t1 = OD_ROUGH_TAN_LOOKUP[i + 1];
+  d = alpha * 36 - (i << 24);
+  return (int)((((int64_t)t0 << 32) + ((t1 - t0) << 8) * (int64_t)d) >> 32);
+}
+
+static const int64_t OD_ATANH_LOG2[32] = {
+  0x32B803473F7AD0F4LL, 0x2F2A71BD4E25E916LL, 0x2E68B244BB93BA06LL,
+  0x2E39FB9198CE62E4LL, 0x2E2E683F68565C8FLL, 0x2E2B850BE2077FC1LL,
+  0x2E2ACC58FE7B78DBLL, 0x2E2A9E2DE52FD5F2LL, 0x2E2A92A338D53EECLL,
+  0x2E2A8FC08F5E19B6LL, 0x2E2A8F07E51A485ELL, 0x2E2A8ED9BA8AF388LL,
+  0x2E2A8ECE2FE7384ALL, 0x2E2A8ECB4D3E4B1ALL, 0x2E2A8ECA94940FE8LL,
+  0x2E2A8ECA6669811DLL, 0x2E2A8ECA5ADEDD6ALL, 0x2E2A8ECA57FC347ELL,
+  0x2E2A8ECA57438A43LL, 0x2E2A8ECA57155FB4LL, 0x2E2A8ECA5709D510LL,
+  0x2E2A8ECA5706F267LL, 0x2E2A8ECA570639BDLL, 0x2E2A8ECA57060B92LL,
+  0x2E2A8ECA57060008LL, 0x2E2A8ECA5705FD25LL, 0x2E2A8ECA5705FC6CLL,
+  0x2E2A8ECA5705FC3ELL, 0x2E2A8ECA5705FC33LL, 0x2E2A8ECA5705FC30LL,
+  0x2E2A8ECA5705FC2FLL, 0x2E2A8ECA5705FC2FLL
+};
+
+static int od_ilog64(int64_t v) {
+  static const unsigned char OD_DEBRUIJN_IDX64[64] = {
+    0,  1,  2,  7,  3,  13, 8,  19, 4,  25, 14, 28, 9,  34, 20, 40,
+    5,  17, 26, 38, 15, 46, 29, 48, 10, 31, 35, 54, 21, 50, 41, 57,
+    63, 6,  12, 18, 24, 27, 33, 39, 16, 37, 45, 47, 30, 53, 49, 56,
+    62, 11, 23, 32, 36, 44, 52, 55, 61, 22, 43, 51, 60, 42, 59, 58
+  };
+  int ret;
+  v |= v >> 1;
+  v |= v >> 2;
+  v |= v >> 4;
+  v |= v >> 8;
+  v |= v >> 16;
+  v |= v >> 32;
+  ret = (int)v & 1;
+  v = (v >> 1) + 1;
+  ret += OD_DEBRUIJN_IDX64[v * UINT64_C(0x218A392CD3D5DBF) >> 58 & 0x3F];
+  return ret;
+}
+
+/*Computes the binary exponential of logq57.
+  input: a log base 2 in Q57 format
+  output: a 64 bit integer in Q0 (no fraction) */
+static int64_t od_bexp64(int64_t logq57) {
+  int64_t w;
+  int64_t z;
+  int ipart;
+  ipart = (int)(logq57 >> 57);
+  if (ipart < 0) return 0;
+  if (ipart >= 63) return 0x7FFFFFFFFFFFFFFFLL;
+  z = logq57 - OD_Q57(ipart);
+  if (z) {
+    int64_t mask;
+    int64_t wlo;
+    int i;
+    /*C doesn't give us 64x64->128 muls, so we use CORDIC.
+      This is not particularly fast, but it's not being used in time-critical
+       code; it is very accurate.*/
+    /*z is the fractional part of the log in Q62 format.
+      We need 1 bit of headroom since the magnitude can get larger than 1
+       during the iteration, and a sign bit.*/
+    z <<= 5;
+    /*w is the exponential in Q61 format (since it also needs headroom and can
+       get as large as 2.0); we could get another bit if we dropped the sign,
+       but we'll recover that bit later anyway.
+      Ideally this should start out as
+        \lim_{n->\infty} 2^{61}/\product_{i=1}^n \sqrt{1-2^{-2i}}
+       but in order to guarantee convergence we have to repeat iterations 4,
+        13 (=3*4+1), and 40 (=3*13+1, etc.), so it winds up somewhat larger.*/
+    w = 0x26A3D0E401DD846DLL;
+    for (i = 0;; i++) {
+      mask = -(z < 0);
+      w += ((w >> (i + 1)) + mask) ^ mask;
+      z -= (OD_ATANH_LOG2[i] + mask) ^ mask;
+      /*Repeat iteration 4.*/
+      if (i >= 3) break;
+      z *= 2;
+    }
+    for (;; i++) {
+      mask = -(z < 0);
+      w += ((w >> (i + 1)) + mask) ^ mask;
+      z -= (OD_ATANH_LOG2[i] + mask) ^ mask;
+      /*Repeat iteration 13.*/
+      if (i >= 12) break;
+      z *= 2;
+    }
+    for (; i < 32; i++) {
+      mask = -(z < 0);
+      w += ((w >> (i + 1)) + mask) ^ mask;
+      z = (z - ((OD_ATANH_LOG2[i] + mask) ^ mask)) * 2;
+    }
+    wlo = 0;
+    /*Skip the remaining iterations unless we really require that much
+       precision.
+      We could have bailed out earlier for smaller iparts, but that would
+       require initializing w from a table, as the limit doesn't converge to
+       61-bit precision until n=30.*/
+    if (ipart > 30) {
+      /*For these iterations, we just update the low bits, as the high bits
+         can't possibly be affected.
+        OD_ATANH_LOG2 has also converged (it actually did so one iteration
+         earlier, but that's no reason for an extra special case).*/
+      for (;; i++) {
+        mask = -(z < 0);
+        wlo += ((w >> i) + mask) ^ mask;
+        z -= (OD_ATANH_LOG2[31] + mask) ^ mask;
+        /*Repeat iteration 40.*/
+        if (i >= 39) break;
+        z <<= 1;
+      }
+      for (; i < 61; i++) {
+        mask = -(z < 0);
+        wlo += ((w >> i) + mask) ^ mask;
+        z = (z - ((OD_ATANH_LOG2[31] + mask) ^ mask)) << 1;
+      }
+    }
+    w = (w << 1) + wlo;
+  } else {
+    w = (int64_t)1 << 62;
+  }
+  if (ipart < 62) {
+    w = ((w >> (61 - ipart)) + 1) >> 1;
+  }
+  return w;
+}
+
+/*Computes the binary log of w
+  input: a 64-bit integer in Q0 (no fraction)
+  output: a 64-bit log in Q57 */
+static int64_t od_blog64(int64_t w) {
+  int64_t z;
+  int ipart;
+  if (w <= 0) return -1;
+  ipart = od_ilog64(w) - 1;
+  if (ipart > 61) {
+    w >>= ipart - 61;
+  } else {
+    w <<= 61 - ipart;
+  }
+  z = 0;
+  if (w & (w - 1)) {
+    int64_t x;
+    int64_t y;
+    int64_t u;
+    int64_t mask;
+    int i;
+    /*C doesn't give us 64x64->128 muls, so we use CORDIC.
+      This is not particularly fast, but it's not being used in time-critical
+       code; it is very accurate.*/
+    /*z is the fractional part of the log in Q61 format.*/
+    /*x and y are the cosh() and sinh(), respectively, in Q61 format.
+      We are computing z = 2*atanh(y/x) = 2*atanh((w - 1)/(w + 1)).*/
+    x = w + ((int64_t)1 << 61);
+    y = w - ((int64_t)1 << 61);
+    for (i = 0; i < 4; i++) {
+      mask = -(y < 0);
+      z += ((OD_ATANH_LOG2[i] >> i) + mask) ^ mask;
+      u = x >> (i + 1);
+      x -= ((y >> (i + 1)) + mask) ^ mask;
+      y -= (u + mask) ^ mask;
+    }
+    /*Repeat iteration 4.*/
+    for (i--; i < 13; i++) {
+      mask = -(y < 0);
+      z += ((OD_ATANH_LOG2[i] >> i) + mask) ^ mask;
+      u = x >> (i + 1);
+      x -= ((y >> (i + 1)) + mask) ^ mask;
+      y -= (u + mask) ^ mask;
+    }
+    /*Repeat iteration 13.*/
+    for (i--; i < 32; i++) {
+      mask = -(y < 0);
+      z += ((OD_ATANH_LOG2[i] >> i) + mask) ^ mask;
+      u = x >> (i + 1);
+      x -= ((y >> (i + 1)) + mask) ^ mask;
+      y -= (u + mask) ^ mask;
+    }
+    /*OD_ATANH_LOG2 has converged.*/
+    for (; i < 40; i++) {
+      mask = -(y < 0);
+      z += ((OD_ATANH_LOG2[31] >> i) + mask) ^ mask;
+      u = x >> (i + 1);
+      x -= ((y >> (i + 1)) + mask) ^ mask;
+      y -= (u + mask) ^ mask;
+    }
+    /*Repeat iteration 40.*/
+    for (i--; i < 62; i++) {
+      mask = -(y < 0);
+      z += ((OD_ATANH_LOG2[31] >> i) + mask) ^ mask;
+      u = x >> (i + 1);
+      x -= ((y >> (i + 1)) + mask) ^ mask;
+      y -= (u + mask) ^ mask;
+    }
+    z = (z + 8) >> 4;
+  }
+  return OD_Q57(ipart) + z;
+}
+
+/*Convenience function converts Q57 value to a clamped 32-bit Q24 value
+  in: input in Q57 format.
+  Return: same number in Q24 */
+static int32_t od_q57_to_q24(int64_t in) {
+  int64_t ret;
+  ret = (in + ((int64_t)1 << 32)) >> 33;
+  /*0x80000000 is automatically converted to unsigned on 32-bit systems.
+    -0x7FFFFFFF-1 is needed to avoid "promoting" the whole expression to
+    unsigned.*/
+  return (int32_t)OD_CLAMPI(-0x7FFFFFFF - 1, ret, 0x7FFFFFFF);
+}
+
+/*Binary exponential of log_scale with 24-bit fractional precision and
+   saturation.
+  log_scale: A binary logarithm in Q57 format.
+  Return: The binary exponential in Q24 format, saturated to 2**31-1 if
+   log_scale was too large.*/
+static int32_t od_bexp64_q24(int64_t log_scale) {
+  if (log_scale < OD_Q57(8)) {
+    int64_t ret;
+    ret = od_bexp64(log_scale + OD_Q57(24));
+    return ret < 0x7FFFFFFF ? (int32_t)ret : 0x7FFFFFFF;
+  }
+  return 0x7FFFFFFF;
+}
+
+/*Re-initialize Bessel filter coefficients with the specified delay.
+  This does not alter the x/y state, but changes the reaction time of the
+   filter.
+  Altering the time constant of a reactive filter without alterning internal
+   state is something that has to be done carefuly, but our design operates at
+   high enough delays and with small enough time constant changes to make it
+   safe.*/
+static void od_iir_bessel2_reinit(od_iir_bessel2 *f, int delay) {
+  int alpha;
+  int64_t one48;
+  int64_t warp;
+  int64_t k1;
+  int64_t k2;
+  int64_t d;
+  int64_t a;
+  int64_t ik2;
+  int64_t b1;
+  int64_t b2;
+  /*This borrows some code from an unreleased version of Postfish.
+    See the recipe at http://unicorn.us.com/alex/2polefilters.html for details
+     on deriving the filter coefficients.*/
+  /*alpha is Q24*/
+  alpha = (1 << 24) / delay;
+  one48 = (int64_t)1 << 48;
+  /*warp is 7.12*/
+  warp = OD_MAXI(od_warp_alpha(alpha), 1);
+  /*k1 is 9.12*/
+  k1 = 3 * warp;
+  /*k2 is 16.24.*/
+  k2 = k1 * warp;
+  /*d is 16.15.*/
+  d = ((((1 << 12) + k1) << 12) + k2 + 256) >> 9;
+  /*a is 0.32, since d is larger than both 1.0 and k2.*/
+  a = (k2 << 23) / d;
+  /*ik2 is 25.24.*/
+  ik2 = one48 / k2;
+  /*b1 is Q56; in practice, the integer ranges between -2 and 2.*/
+  b1 = 2 * a * (ik2 - (1 << 24));
+  /*b2 is Q56; in practice, the integer ranges between -2 and 2.*/
+  b2 = (one48 << 8) - ((4 * a) << 24) - b1;
+  /*All of the filter parameters are Q24.*/
+  f->c[0] = (int32_t)((b1 + ((int64_t)1 << 31)) >> 32);
+  f->c[1] = (int32_t)((b2 + ((int64_t)1 << 31)) >> 32);
+  f->g = (int32_t)((a + 128) >> 8);
+}
+
+/*Initialize a 2nd order low-pass Bessel filter with the corresponding delay
+   and initial value.
+  value is Q24.*/
+static void od_iir_bessel2_init(od_iir_bessel2 *f, int delay, int32_t value) {
+  od_iir_bessel2_reinit(f, delay);
+  f->y[1] = f->y[0] = f->x[1] = f->x[0] = value;
+}
+
+static int64_t od_iir_bessel2_update(od_iir_bessel2 *f, int32_t x) {
+  int64_t c0;
+  int64_t c1;
+  int64_t g;
+  int64_t x0;
+  int64_t x1;
+  int64_t y0;
+  int64_t y1;
+  int64_t ya;
+  c0 = f->c[0];
+  c1 = f->c[1];
+  g = f->g;
+  x0 = f->x[0];
+  x1 = f->x[1];
+  y0 = f->y[0];
+  y1 = f->y[1];
+  ya = ((x + x0 * 2 + x1) * g + y0 * c0 + y1 * c1 + (1 << 23)) >> 24;
+  f->x[1] = (int32_t)x0;
+  f->x[0] = x;
+  f->y[1] = (int32_t)y0;
+  f->y[0] = (int32_t)ya;
+  return ya;
+}
+
+static void od_enc_rc_reset(od_rc_state *rc) {
+  int64_t npixels;
+  int64_t ibpp;
+  rc->bits_per_frame = (int64_t)(rc->target_bitrate / rc->framerate);
+  /*Insane framerates or frame sizes mean insane bitrates.
+    Let's not get carried away.*/
+  if (rc->bits_per_frame > 0x400000000000LL) {
+    rc->bits_per_frame = (int64_t)0x400000000000LL;
+  } else {
+    if (rc->bits_per_frame < 32) {
+      rc->bits_per_frame = 32;
+    }
+  }
+  rc->reservoir_frame_delay = OD_MAXI(rc->reservoir_frame_delay, 12);
+  rc->reservoir_max = rc->bits_per_frame * rc->reservoir_frame_delay;
+  /*Start with a buffer fullness and fullness target of 50% */
+  rc->reservoir_target = (rc->reservoir_max + 1) >> 1;
+  rc->reservoir_fullness = rc->reservoir_target;
+  /*Pick exponents and initial scales for quantizer selection.*/
+  npixels = rc->frame_width * (int64_t)rc->frame_height;
+  rc->log_npixels = od_blog64(npixels);
+  ibpp = npixels / rc->bits_per_frame;
+  /*All of these initial scale/exp values are from Theora, and have not yet
+     been adapted to Daala, so they're certainly wrong.
+    The B-frame values especially are simply copies of the P-frame values.*/
+  if (ibpp < 1) {
+    rc->exp[OD_I_FRAME] = 59;
+    rc->log_scale[OD_I_FRAME] = od_blog64(1997) - OD_Q57(OD_COEFF_SHIFT);
+  } else if (ibpp < 2) {
+    rc->exp[OD_I_FRAME] = 55;
+    rc->log_scale[OD_I_FRAME] = od_blog64(1604) - OD_Q57(OD_COEFF_SHIFT);
+  } else {
+    rc->exp[OD_I_FRAME] = 48;
+    rc->log_scale[OD_I_FRAME] = od_blog64(834) - OD_Q57(OD_COEFF_SHIFT);
+  }
+  if (ibpp < 4) {
+    rc->exp[OD_P_FRAME] = 100;
+    rc->log_scale[OD_P_FRAME] = od_blog64(2249) - OD_Q57(OD_COEFF_SHIFT);
+  } else if (ibpp < 8) {
+    rc->exp[OD_P_FRAME] = 95;
+    rc->log_scale[OD_P_FRAME] = od_blog64(1751) - OD_Q57(OD_COEFF_SHIFT);
+  } else {
+    rc->exp[OD_P_FRAME] = 73;
+    rc->log_scale[OD_P_FRAME] = od_blog64(1260) - OD_Q57(OD_COEFF_SHIFT);
+  }
+  /*Golden P-frames both use the same log_scale and exp modeling
+     values as regular P-frames and the same scale follower.
+    For convenience in the rate calculation code, we maintain a copy of
+    the scale and exp values in OD_GOLDEN_P_FRAME.*/
+  rc->exp[OD_GOLDEN_P_FRAME] = rc->exp[OD_P_FRAME];
+  rc->log_scale[OD_GOLDEN_P_FRAME] = rc->log_scale[OD_P_FRAME];
+  rc->exp[OD_ALTREF_P_FRAME] = rc->exp[OD_P_FRAME];
+  rc->log_scale[OD_ALTREF_P_FRAME] = rc->log_scale[OD_P_FRAME];
+  /*We clamp the actual I and B frame delays to a minimum of 10 to work within
+     the range of values where later incrementing the delay works as designed.
+    10 is not an exact choice, but rather a good working trade-off.*/
+  rc->inter_p_delay = 10;
+  rc->inter_delay_target = rc->reservoir_frame_delay >> 1;
+  memset(rc->frame_count, 0, sizeof(rc->frame_count));
+  /*Drop-frame tracking is concerned with more than just the basic three frame
+     types.
+    It needs to track boosted and cut subtypes (of which there is only one
+     right now, OD_GOLDEN_P_FRAME). */
+  rc->prev_drop_count[OD_I_FRAME] = 0;
+  rc->log_drop_scale[OD_I_FRAME] = OD_Q57(0);
+  rc->prev_drop_count[OD_P_FRAME] = 0;
+  rc->log_drop_scale[OD_P_FRAME] = OD_Q57(0);
+  rc->prev_drop_count[OD_GOLDEN_P_FRAME] = 0;
+  rc->log_drop_scale[OD_GOLDEN_P_FRAME] = OD_Q57(0);
+  rc->prev_drop_count[OD_ALTREF_P_FRAME] = 0;
+  rc->log_drop_scale[OD_ALTREF_P_FRAME] = OD_Q57(0);
+  /*Set up second order followers, initialized according to corresponding
+     time constants.*/
+  od_iir_bessel2_init(&rc->scalefilter[OD_I_FRAME], 4,
+                      od_q57_to_q24(rc->log_scale[OD_I_FRAME]));
+  od_iir_bessel2_init(&rc->scalefilter[OD_P_FRAME], rc->inter_p_delay,
+                      od_q57_to_q24(rc->log_scale[OD_P_FRAME]));
+  od_iir_bessel2_init(&rc->vfrfilter[OD_I_FRAME], 4,
+                      od_bexp64_q24(rc->log_drop_scale[OD_I_FRAME]));
+  od_iir_bessel2_init(&rc->vfrfilter[OD_P_FRAME], 4,
+                      od_bexp64_q24(rc->log_drop_scale[OD_P_FRAME]));
+  od_iir_bessel2_init(&rc->vfrfilter[OD_GOLDEN_P_FRAME], 4,
+                      od_bexp64_q24(rc->log_drop_scale[OD_GOLDEN_P_FRAME]));
+  od_iir_bessel2_init(&rc->vfrfilter[OD_ALTREF_P_FRAME], 4,
+                      od_bexp64_q24(rc->log_drop_scale[OD_ALTREF_P_FRAME]));
+}
+
+int od_enc_rc_resize(od_rc_state *rc) {
+  /*If encoding has not yet begun, reset the buffer state.*/
+  if (rc->cur_frame == 0) {
+    od_enc_rc_reset(rc);
+  } else {
+    int idt;
+    /*Otherwise, update the bounds on the buffer, but not the current
+       fullness.*/
+    rc->bits_per_frame = (int64_t)(rc->target_bitrate / rc->framerate);
+    /*Insane framerates or frame sizes mean insane bitrates.
+      Let's not get carried away.*/
+    if (rc->bits_per_frame > 0x400000000000LL) {
+      rc->bits_per_frame = (int64_t)0x400000000000LL;
+    } else {
+      if (rc->bits_per_frame < 32) {
+        rc->bits_per_frame = 32;
+      }
+    }
+    rc->reservoir_frame_delay = OD_MAXI(rc->reservoir_frame_delay, 12);
+    rc->reservoir_max = rc->bits_per_frame * rc->reservoir_frame_delay;
+    rc->reservoir_target =
+        ((rc->reservoir_max + 1) >> 1) +
+        ((rc->bits_per_frame + 2) >> 2) *
+            OD_MINI(rc->keyframe_rate, rc->reservoir_frame_delay);
+    /*Update the INTER-frame scale filter delay.
+      We jump to it immediately if we've already seen enough frames; otherwise
+       it is simply set as the new target.*/
+    rc->inter_delay_target = idt = OD_MAXI(rc->reservoir_frame_delay >> 1, 10);
+    if (idt < OD_MINI(rc->inter_p_delay, rc->frame_count[OD_P_FRAME])) {
+      od_iir_bessel2_init(&rc->scalefilter[OD_P_FRAME], idt,
+                          rc->scalefilter[OD_P_FRAME].y[0]);
+      rc->inter_p_delay = idt;
+    }
+  }
+  return 0;
+}
+
+int od_enc_rc_init(od_rc_state *rc, int64_t bitrate, int delay_ms) {
+  if (rc->framerate <= 0) return 1;
+  if (rc->target_bitrate > 0) {
+    /*State has already been initialized; rather than reinitialize,
+      adjust the buffering for the new target rate. */
+    rc->target_bitrate = bitrate;
+    return od_enc_rc_resize(rc);
+  }
+  rc->target_quantizer = 0;
+  rc->target_bitrate = bitrate;
+  rc->rate_bias = 0;
+  if (bitrate > 0) {
+    /* The buffer size is clamped between [12, 256], this interval is short
+       enough to
+       allow reaction, but long enough to allow looking into the next GOP
+       (avoiding
+       the case where the last frames before an I-frame get starved).
+       The 12 frame minimum gives us some chance to distribute bit estimation
+       errors in the worst case. The 256 frame maximum means we'll require 8-10
+       seconds
+       of pre-buffering at 24-30 fps, which is not unreasonable.*/
+    rc->reservoir_frame_delay =
+        (int)OD_MINI((delay_ms / 1000) * rc->framerate, 256);
+    rc->drop_frames = 1;
+    rc->cap_overflow = 1;
+    rc->cap_underflow = 0;
+    rc->twopass_state = 0;
+    od_enc_rc_reset(rc);
+  }
+  return 0;
+}
+
+/*Scale the number of frames by the number of expected drops/duplicates.*/
+static int od_rc_scale_drop(od_rc_state *rc, int frame_type, int nframes) {
+  if (rc->prev_drop_count[frame_type] > 0 ||
+      rc->log_drop_scale[frame_type] > OD_Q57(0)) {
+    int64_t dup_scale;
+    dup_scale = od_bexp64(((rc->log_drop_scale[frame_type] +
+                            od_blog64(rc->prev_drop_count[frame_type] + 1)) >>
+                           1) +
+                          OD_Q57(8));
+    if (dup_scale < nframes << 8) {
+      int dup_scalei;
+      dup_scalei = (int)dup_scale;
+      if (dup_scalei > 0) {
+        nframes = ((nframes << 8) + dup_scalei - 1) / dup_scalei;
+      }
+    } else {
+      nframes = !!nframes;
+    }
+  }
+  return nframes;
+}
+
+/*Closed form version of frame determination code.
+  Used by rate control to predict frame types and subtypes into the future.
+  No side effects, may be called any number of times.
+  Note that it ignores end-of-file conditions; one-pass planning *should*
+   ignore end-of-file. */
+int od_frame_type(od_rc_state *rc, int64_t coding_frame_count, int *is_golden,
+                  int *is_altref, int64_t *ip_count) {
+  int frame_type;
+  if (coding_frame_count == 0) {
+    *is_golden = 1;
+    *is_altref = 1;
+    *ip_count = 0;
+    frame_type = OD_I_FRAME;
+  } else {
+    int keyrate = rc->keyframe_rate;
+    if (rc->closed_gop) {
+      int ip_per_gop;
+      int gop_n;
+      int gop_i;
+      ip_per_gop = (keyrate - 1) / 2;
+      gop_n = coding_frame_count / keyrate;
+      gop_i = coding_frame_count - gop_n * keyrate;
+      *ip_count = gop_n * ip_per_gop + (gop_i > 0) + (gop_i - 1);
+      frame_type = gop_i == 0 ? OD_I_FRAME : OD_P_FRAME;
+    } else {
+      int ip_per_gop;
+      int gop_n;
+      int gop_i;
+      ip_per_gop = (keyrate);
+      gop_n = (coding_frame_count - 1) / keyrate;
+      gop_i = coding_frame_count - gop_n * keyrate - 1;
+      *ip_count = (coding_frame_count > 0) + gop_n * ip_per_gop + (gop_i);
+      frame_type = gop_i / 1 < ip_per_gop - 1 ? OD_P_FRAME : OD_I_FRAME;
+    }
+  }
+  *is_golden =
+      (*ip_count % rc->goldenframe_rate) == 0 || frame_type == OD_I_FRAME;
+  *is_altref = (*ip_count % rc->altref_rate) == 0 || frame_type == OD_I_FRAME;
+  return frame_type;
+}
+
+/*Count frames types forward from the current frame up to but not including
+   the last I-frame in reservoir_frame_delay.
+  If reservoir_frame_delay contains no I-frames (or the current frame is the
+   only I-frame), count all reservoir_frame_delay frames.
+  Returns the number of frames counted.
+  Right now, this implementation is simple, brute-force, and expensive.
+  It is also easy to understand and debug.
+  TODO: replace with a virtual FIFO that keeps running totals as
+   repeating the counting over-and-over will have a performance impact on
+   whole-file 2pass usage.*/
+static int frame_type_count(od_rc_state *rc, int nframes[OD_FRAME_NSUBTYPES]) {
+  int i;
+  int j;
+  int acc[OD_FRAME_NSUBTYPES];
+  int count;
+  int reservoir_frames;
+  int reservoir_frame_delay;
+  memset(nframes, 0, OD_FRAME_NSUBTYPES * sizeof(*nframes));
+  memset(acc, 0, sizeof(acc));
+  count = 0;
+  reservoir_frames = 0;
+#if 1
+  /*Go ahead and count past end-of-stream.
+    We won't nail the exact bitrate on short files that end with a partial
+     GOP, but we also won't [potentially] destroy the quality of the last few
+     frames in that same case when we suddenly find out the stream is ending
+     before the original planning horizon.*/
+  reservoir_frame_delay = rc->reservoir_frame_delay;
+#else
+  /*Don't count past the end of the stream (once we know where end-of-stream
+     is).*/
+  reservoir_frame_delay =
+      rc->end_of_input ? rc->input_size + 1 : rc->reservoir_frame_delay;
+#endif
+  for (i = 0; i < reservoir_frame_delay; i++) {
+    int frame_type;
+    int is_golden;
+    int is_altref;
+    int64_t dummy;
+    frame_type =
+        od_frame_type(rc, rc->cur_frame + i, &is_golden, &is_altref, &dummy);
+    switch (frame_type) {
+      case OD_I_FRAME: {
+        for (j = 0; j < OD_FRAME_NSUBTYPES; j++) nframes[j] += acc[j];
+        reservoir_frames += count;
+        memset(acc, 0, sizeof(acc));
+        acc[OD_I_FRAME] = 1;
+        count = 1;
+        break;
+      }
+      case OD_P_FRAME: {
+        if (is_golden) {
+          ++acc[OD_GOLDEN_P_FRAME];
+          ++count;
+        } else if (is_altref) {
+          ++acc[OD_ALTREF_P_FRAME];
+          ++count;
+        } else {
+          ++acc[OD_P_FRAME];
+          ++count;
+        }
+        break;
+      }
+    }
+  }
+  /*If there were no I-frames at all, or only the first frame was an I-frame,
+     the accumulators never flushed and still contain the counts for the
+     entire buffer.
+    In both these cases, we return these counts.
+    Otherwise, we discard what remains in the accumulators as they contain
+     the counts from and past the last I-frame.*/
+  if (reservoir_frames == 0) {
+    for (i = 0; i < OD_FRAME_NSUBTYPES; i++) nframes[i] = acc[i];
+    reservoir_frames += count;
+  }
+  return reservoir_frames;
+}
+
+static int convert_to_ac_quant(int q, int bit_depth) {
+  return lrint(av1_convert_qindex_to_q(q, bit_depth));
+}
+
+int od_enc_rc_select_quantizers_and_lambdas(od_rc_state *rc,
+                                            int is_golden_frame,
+                                            int is_altref_frame, int frame_type,
+                                            int *bottom_idx, int *top_idx) {
+  int frame_subtype;
+  int64_t log_cur_scale;
+  int lossy_quantizer_min;
+  int lossy_quantizer_max;
+  double mqp_i = OD_MQP_I;
+  double mqp_p = OD_MQP_P;
+  double mqp_gp = OD_MQP_GP;
+  double mqp_ap = OD_MQP_AP;
+  int reservoir_frames;
+  int nframes[OD_FRAME_NSUBTYPES];
+  int32_t mqp_Q12[OD_FRAME_NSUBTYPES];
+  int64_t dqp_Q45[OD_FRAME_NSUBTYPES];
+  /*Verify the closed-form frame type determination code matches what the
+     input queue set.*/
+  /*One pseudo-non-closed-form caveat:
+    Once we've seen end-of-input, the batched frame determination code
+     suppresses the last open-GOP's I-frame (since it would only be
+     useful for the next GOP, which doesn't exist).
+     Thus, don't check one the input queue is drained.*/
+  if (!rc->end_of_input) {
+    int closed_form_type;
+    int closed_form_golden;
+    int closed_form_altref;
+    int64_t closed_form_cur_frame;
+    closed_form_type =
+        od_frame_type(rc, rc->cur_frame, &closed_form_golden,
+                      &closed_form_altref, &closed_form_cur_frame);
+    OD_UNUSED(closed_form_type);
+    OD_UNUSED(is_altref_frame);
+    assert(closed_form_type == frame_type);
+    assert(closed_form_cur_frame == rc->cur_frame);
+    assert(closed_form_altref == is_altref_frame);
+    assert(closed_form_golden == is_golden_frame);
+  }
+
+  log_cur_scale = (int64_t)rc->scalefilter[frame_type].y[0] << 33;
+
+  /*Count the various types and classes of frames.*/
+  reservoir_frames = frame_type_count(rc, nframes);
+  nframes[OD_I_FRAME] = od_rc_scale_drop(rc, OD_I_FRAME, nframes[OD_I_FRAME]);
+  nframes[OD_P_FRAME] = od_rc_scale_drop(rc, OD_P_FRAME, nframes[OD_P_FRAME]);
+  nframes[OD_GOLDEN_P_FRAME] =
+      od_rc_scale_drop(rc, OD_GOLDEN_P_FRAME, nframes[OD_GOLDEN_P_FRAME]);
+  nframes[OD_ALTREF_P_FRAME] =
+      od_rc_scale_drop(rc, OD_ALTREF_P_FRAME, nframes[OD_ALTREF_P_FRAME]);
+
+  switch (rc->twopass_state) {
+    default: break;
+    case 1: {
+      /*Pass 1 mode: use a fixed qi value.*/
+      return rc->firstpass_quant;
+    } break;
+    case 2: {
+      int i;
+      int64_t scale_sum[OD_FRAME_NSUBTYPES];
+      int qti;
+      /*Pass 2 mode: we know exactly how much of each frame type there is in
+         the current buffer window, and have estimates for the scales.*/
+      for (i = 0; i < OD_FRAME_NSUBTYPES; i++) {
+        nframes[i] = rc->nframes[i];
+        nframes[i] = rc->nframes[i];
+        scale_sum[i] = rc->scale_sum[i];
+      }
+      /*If we're not using the same frame type as in pass 1 (because someone
+         changed the keyframe interval), remove that scale estimate.
+        We'll add in a replacement for the correct frame type below.*/
+      qti = rc->cur_metrics.frame_type;
+      if (qti != frame_type) {
+        nframes[qti]--;
+        scale_sum[qti] -= od_bexp64_q24(rc->cur_metrics.log_scale);
+      }
+      /*Compute log_scale estimates for each frame type from the pass-1 scales
+         we measured in the current window.*/
+      for (qti = 0; qti < OD_FRAME_NSUBTYPES; qti++) {
+        rc->log_scale[qti] = nframes[qti] > 0
+                                 ? od_blog64(scale_sum[qti]) -
+                                       od_blog64(nframes[qti]) - OD_Q57(24)
+                                 : -rc->log_npixels;
+      }
+      /*If we're not using the same frame type as in pass 1, add a scale
+         estimate for the corresponding frame using the current low-pass
+         filter value.
+        This is mostly to ensure we have a valid estimate even when pass 1 had
+         no frames of this type in the buffer window.
+        TODO: We could also plan ahead and figure out how many keyframes we'll
+         be forced to add in the current buffer window.*/
+      qti = rc->cur_metrics.frame_type;
+      if (qti != frame_type) {
+        int64_t scale;
+        scale = rc->log_scale[frame_type] < OD_Q57(23)
+                    ? od_bexp64(rc->log_scale[frame_type] + OD_Q57(24))
+                    : 0x7FFFFFFFFFFFLL;
+        scale *= nframes[frame_type];
+        nframes[frame_type]++;
+        scale += od_bexp64_q24(log_cur_scale >> 33);
+        rc->log_scale[frame_type] =
+            od_blog64(scale) - od_blog64(nframes[qti]) - OD_Q57(24);
+      } else {
+        log_cur_scale = (int64_t)rc->cur_metrics.log_scale << 33;
+      }
+    } break;
+  }
+
+  /*Quantizer selection sticks to the codable, lossy portion of the quantizer
+    range.*/
+  lossy_quantizer_min = convert_to_ac_quant(rc->minq, rc->bit_depth);
+  lossy_quantizer_max = convert_to_ac_quant(rc->maxq, rc->bit_depth);
+  frame_subtype = frame_type;
+  /*Stash quantizer modulation by frame type.*/
+  mqp_Q12[OD_I_FRAME] = OD_F_Q12(mqp_i);
+  mqp_Q12[OD_P_FRAME] = OD_F_Q12(mqp_p);
+  mqp_Q12[OD_GOLDEN_P_FRAME] = OD_F_Q12(mqp_gp);
+  mqp_Q12[OD_ALTREF_P_FRAME] = OD_F_Q12(mqp_ap);
+  dqp_Q45[OD_I_FRAME] = OD_F_Q45(OD_DQP_I);
+  dqp_Q45[OD_P_FRAME] = OD_F_Q45(OD_DQP_P);
+  dqp_Q45[OD_GOLDEN_P_FRAME] = OD_F_Q45(OD_DQP_GP);
+  dqp_Q45[OD_ALTREF_P_FRAME] = OD_F_Q45(OD_DQP_AP);
+  /*Is rate control active?*/
+  if (rc->target_bitrate <= 0) {
+    /*Rate control is not active; derive quantizer directly from
+      quality parameter and frame type. */
+    /*Can't use the OD_LOSSLESS macro, as it uses state.quantizer to intuit,
+      and we've not set it yet.*/
+    if (rc->quality == 0) {
+      /*Lossless coding requested.*/
+      rc->base_quantizer = 0;
+      rc->target_quantizer = 0;
+    } else {
+      int64_t log_quantizer;
+
+      /* Adjust the modulation constants using the last frame's quantizer. */
+      double mqp_delta = (255 - rc->target_quantizer) / 2000.0f;
+      mqp_i -= mqp_delta;
+      mqp_p += mqp_delta;
+      mqp_gp -= mqp_delta;
+      mqp_Q12[OD_I_FRAME] = OD_F_Q12(mqp_i);
+      mqp_Q12[OD_P_FRAME] = OD_F_Q12(mqp_p);
+      mqp_Q12[OD_GOLDEN_P_FRAME] = OD_F_Q12(mqp_gp);
+      mqp_Q12[OD_ALTREF_P_FRAME] = OD_F_Q12(mqp_ap);
+
+      if (rc->quality == -1) {
+        /*A quality of -1 means quality was unset; use a default.*/
+        rc->base_quantizer = convert_to_ac_quant(10, rc->bit_depth);
+      } else {
+        rc->base_quantizer = convert_to_ac_quant(rc->quality, rc->bit_depth);
+      }
+
+      if (rc->periodic_boosts && !is_golden_frame) {
+        int pattern_rate = (rc->goldenframe_rate >> 1);
+        int dist_to_golden = rc->cur_frame % pattern_rate;
+        int dist_away_golden = pattern_rate - dist_to_golden;
+        int boost = dist_to_golden;
+        if (dist_away_golden > dist_to_golden) boost = dist_away_golden;
+        boost -= pattern_rate;
+        boost *= (rc->base_quantizer) / OD_PERIODIC_BOOST_DIV;
+        rc->base_quantizer = rc->base_quantizer + boost;
+      }
+
+      /*As originally written, qp modulation is applied to the coded quantizer.
+        Because we now have and use a more precise target quantizer for various
+        calculation, that needs to be modulated as well.
+        Calculate what is, effectively, a fractional coded quantizer. */
+      /*Get the log2 quantizer in Q57 (normalized for coefficient shift).*/
+      log_quantizer = od_blog64(rc->base_quantizer) - OD_Q57(OD_COEFF_SHIFT);
+      /*log_quantizer to Q21.*/
+      log_quantizer >>= 36;
+      /*scale log quantizer, result is Q33.*/
+      log_quantizer *= OD_LOG_QUANTIZER_BASE_Q12;
+      /*Add Q33 offset to Q33 log_quantizer.*/
+      log_quantizer += OD_LOG_QUANTIZER_OFFSET_Q45 >> 12;
+      /*Modulate quantizer according to frame type; result is Q45.*/
+      log_quantizer *= mqp_Q12[frame_subtype];
+      /*Add Q45 boost/cut to Q45 fractional coded quantizer.*/
+      log_quantizer += dqp_Q45[frame_subtype];
+      /*Back to log2 quantizer in Q57.*/
+      log_quantizer = (log_quantizer - OD_LOG_QUANTIZER_OFFSET_Q45) *
+                          OD_LOG_QUANTIZER_EXP_Q12 +
+                      OD_Q57(OD_COEFF_SHIFT);
+      /*Convert Q57 log2 quantizer to unclamped linear target quantizer value.*/
+      rc->target_quantizer = od_bexp64(log_quantizer);
+    }
+  } else {
+    int clamp;
+    int64_t rate_bias;
+    int64_t rate_total;
+    int base_quantizer;
+    int64_t log_quantizer;
+    int qlo;
+    int qhi;
+    int i;
+    /*We clamp the allowed amount of qi change (after initialization).*/
+    clamp = rc->cur_frame > 0;
+    /*Figure out how to re-distribute bits so that we hit our fullness target
+       before the last keyframe in our current buffer window (after the current
+       frame), or the end of the buffer window, whichever comes first.*/
+    /*Single pass only right now.*/
+    /*If we've been missing our target, add a penalty term.*/
+    rate_bias = (rc->rate_bias / (rc->cur_frame + 1000)) * reservoir_frames;
+    /*rate_total is the total bits available over the next
+       reservoir_frames frames.*/
+    rate_total = rc->reservoir_fullness - rc->reservoir_target + rate_bias +
+                 reservoir_frames * rc->bits_per_frame;
+    /*Find a target quantizer that meets our rate target for the specific mix
+       of frame types we'll have over the next frame_delay frames.
+      We model the rate<->quantizer relationship as:
+       rate = scale*(quantizer**-exp)
+      In this case, we have our desired rate, an exponent selected in setup,
+       and a scale that's been measured over our frame history, so we're
+       solving for the quantizer.
+      Exponentiation with arbitrary exponents is expensive, so we work in
+       the binary log domain (binary exp and log aren't too bad):
+       rate = e2(log2_scale - log2_quantizer * exp)
+      There's no easy closed form solution, so we bisection search for it.*/
+    /*We do not currently allow rate control to select lossless encoding.*/
+    qlo = 1;
+    /*If there's a quality specified, it's used to select the
+       coarsest base quantizer we can select.
+      Otherwise we can use up to and including the coarsest codable
+       quantizer.*/
+    if (rc->quality > 0)
+      qhi = convert_to_ac_quant(rc->quality, rc->bit_depth);
+    else
+      qhi = lossy_quantizer_max;
+    base_quantizer = (qlo + qhi) >> 1;
+    while (qlo < qhi) {
+      volatile int64_t log_base_quantizer;
+      int64_t diff;
+      int64_t bits;
+      /*Count bits contributed by each frame type using the model.*/
+      bits = 0;
+      log_base_quantizer = od_blog64(base_quantizer);
+      for (i = 0; i < OD_FRAME_NSUBTYPES; i++) {
+        /*Modulate base quantizer by frame type.*/
+        /*Get the log2 quantizer in Q57 (normalized for coefficient shift).*/
+        log_quantizer = log_base_quantizer - OD_Q57(OD_COEFF_SHIFT);
+        /*log_quantizer to Q21.*/
+        log_quantizer >>= 36;
+        /*scale log quantizer, result is Q33.*/
+        log_quantizer *= OD_LOG_QUANTIZER_BASE_Q12;
+        /*Add Q33 offset to Q33 log_quantizer.*/
+        log_quantizer += OD_LOG_QUANTIZER_OFFSET_Q45 >> 12;
+        /*Modulate quantizer according to frame type; result is Q45.*/
+        log_quantizer *= mqp_Q12[i];
+        /*Add Q45 boost/cut to Q45 fractional coded quantizer.*/
+        log_quantizer += dqp_Q45[i];
+        /*Back to log2 quantizer in Q57.*/
+        log_quantizer = (log_quantizer - OD_LOG_QUANTIZER_OFFSET_Q45) *
+                            OD_LOG_QUANTIZER_EXP_Q12 +
+                        OD_Q57(OD_COEFF_SHIFT);
+        /*Clamp modulated quantizer values.*/
+        log_quantizer = OD_CLAMPI(od_blog64(lossy_quantizer_min), log_quantizer,
+                                  od_blog64(lossy_quantizer_max));
+        /* All the fields here are Q57 except for the exponent which is Q6.*/
+        bits += nframes[i] * od_bexp64(rc->log_scale[i] + rc->log_npixels -
+                                       (log_quantizer >> 6) * rc->exp[i]);
+      }
+      diff = bits - rate_total;
+      if (diff > 0) {
+        qlo = base_quantizer + 1;
+      } else if (diff < 0) {
+        qhi = base_quantizer - 1;
+      } else {
+        break;
+      }
+      base_quantizer = (qlo + qhi) >> 1;
+    }
+    /*If this was not one of the initial frames, limit the change in base
+       quantizer to within [0.8*Q,1.2*Q], where Q is the previous frame's
+       base quantizer.*/
+    if (clamp) {
+      base_quantizer = OD_CLAMPI((rc->base_quantizer * 0x0CCCD + 0x8000) >> 16,
+                                 base_quantizer,
+                                 (rc->base_quantizer * 0x13333 + 0x8000) >> 16);
+    }
+    /*Modulate chosen base quantizer to produce target quantizer.*/
+    log_quantizer = od_blog64(base_quantizer);
+    /*Get the log2 quantizer in Q57 (normalized for coefficient shift).*/
+    log_quantizer -= OD_Q57(OD_COEFF_SHIFT);
+    /*log_quantizer to Q21.*/
+    log_quantizer >>= 36;
+    /*scale log quantizer, result is Q33.*/
+    log_quantizer *= OD_LOG_QUANTIZER_BASE_Q12;
+    /*Add Q33 offset to Q33 log_quantizer.*/
+    log_quantizer += OD_LOG_QUANTIZER_OFFSET_Q45 >> 12;
+    /*Modulate quantizer according to frame type; result is Q45.*/
+    log_quantizer *= mqp_Q12[frame_subtype];
+    /*Add Q45 boost/cut to Q45 fractional coded quantizer.*/
+    log_quantizer += dqp_Q45[frame_subtype];
+    /*Back to log2 quantizer in Q57.*/
+    log_quantizer = (log_quantizer - OD_LOG_QUANTIZER_OFFSET_Q45) *
+                        OD_LOG_QUANTIZER_EXP_Q12 +
+                    OD_Q57(OD_COEFF_SHIFT);
+    /*Clamp modulated quantizer values.*/
+    log_quantizer = OD_CLAMPI(od_blog64(lossy_quantizer_min), log_quantizer,
+                              od_blog64(lossy_quantizer_max));
+    /*The above allocation looks only at the total rate we'll accumulate in
+       the next reservoir_frame_delay frames.
+      However we could overflow the bit reservoir on the very next frame, so
+       check for that here if we're not using a soft target.*/
+    if (rc->cap_overflow) {
+      int64_t margin;
+      int64_t soft_limit;
+      int64_t log_soft_limit;
+      int64_t log_scale_pixels;
+      int64_t exp;
+      int64_t log_qexp;
+      /*Allow 3% of the buffer for prediction error.
+        This should be plenty, and we don't mind if we go a bit over; we only
+         want to keep these bits from being completely wasted.*/
+      margin = (rc->reservoir_max + 31) >> 5;
+      /*We want to use at least this many bits next frame.*/
+      soft_limit = rc->reservoir_fullness + rc->bits_per_frame -
+                   (rc->reservoir_max - margin);
+      log_soft_limit = od_blog64(soft_limit);
+      /*If we're predicting we won't use that many bits...*/
+      log_scale_pixels = rc->log_scale[frame_subtype] + rc->log_npixels;
+      exp = rc->exp[frame_subtype];
+      log_qexp = (log_quantizer >> 6) * exp;
+      if (log_scale_pixels - log_qexp < log_soft_limit) {
+        /*Scale the adjustment based on how far into the margin we are.*/
+        log_qexp += ((log_scale_pixels - log_soft_limit - log_qexp) >> 32) *
+                    (OD_MINI(margin, soft_limit) << 32) / margin;
+        log_quantizer = (((log_qexp + (exp >> 1)) / exp) << 6);
+      }
+    }
+    /*We just checked we don't overflow the reservoir next frame, now check
+       we don't underflow and bust the budget (when not using a soft target).
+      Disabled when a quality bound is set; if we saturate quantizer to the
+       maximum possible size when we have a limiting max quality, the
+       resulting lambda can cause strange behavior.*/
+    if (rc->quality == -1) {
+      int64_t exp;
+      int64_t log_qexp;
+      int64_t log_scale_pixels;
+      int64_t log_hard_limit;
+      /*Compute the maximum number of bits we can use in the next frame.
+        Allow 50% of the rate for a single frame for prediction error.
+        This may not be enough for keyframes or sudden changes in
+         complexity.*/
+      log_hard_limit =
+          od_blog64(rc->reservoir_fullness + (rc->bits_per_frame >> 1));
+      /*If we're predicting we'll use more than this...*/
+      log_scale_pixels = rc->log_scale[frame_subtype] + rc->log_npixels;
+      exp = rc->exp[frame_subtype];
+      log_qexp = (log_quantizer >> 6) * exp;
+      if (log_scale_pixels - log_qexp > log_hard_limit) {
+        /*Force the target to hit our limit exactly.*/
+        log_qexp = log_scale_pixels - log_hard_limit;
+        log_quantizer = (log_qexp + (exp >> 1)) / exp << 6;
+        /*If that target is unreasonable, oh well; we'll have to drop.*/
+        log_quantizer = OD_MAXI(log_quantizer, od_blog64(lossy_quantizer_max));
+      }
+    }
+    /*Compute a final estimate of the number of bits we plan to use, update
+       the running rate bias measurement.*/
+    {
+      int64_t log_qexp;
+      int64_t log_scale_pixels;
+      log_scale_pixels = rc->log_scale[frame_subtype] + rc->log_npixels;
+      log_qexp = (log_quantizer >> 6) * rc->exp[frame_subtype];
+      rc->rate_bias += od_bexp64(log_scale_pixels - log_qexp);
+    }
+    rc->target_quantizer = od_bexp64(log_quantizer);
+    /*The various cappings and adjustments may have altered the log_quantizer
+       target significantly.
+      We can either update the base quantizer to be consistent with the
+       target or let it track separately.
+      Theora behavior effectively keeps them consistent, as it regenerates
+       the effective base quantizer from the target each frame rather than
+       saving both.
+      For Daala, it's easier to allow them to track separately.
+      For now, allow them to track separately and see how it behaves.*/
+    rc->base_quantizer = base_quantizer;
+  }
+  *bottom_idx = lossy_quantizer_min;
+  *top_idx = lossy_quantizer_max;
+  rc->target_quantizer = av1_qindex_from_ac(
+      OD_CLAMPI(lossy_quantizer_min, rc->target_quantizer, lossy_quantizer_max),
+      rc->bit_depth);
+  return rc->target_quantizer;
+}
+
+int od_enc_rc_update_state(od_rc_state *rc, int64_t bits, int is_golden_frame,
+                           int is_altref_frame, int frame_type, int droppable) {
+  int dropped;
+  dropped = 0;
+  /*Update rate control only if rate control is active.*/
+  if (rc->target_bitrate > 0) {
+    int64_t log_scale;
+    int frame_subtype;
+    frame_subtype = frame_type;
+    /*Track non-golden and golden P frame drops separately.*/
+    if (is_golden_frame && frame_type == OD_P_FRAME)
+      frame_subtype = OD_GOLDEN_P_FRAME;
+    else if (is_altref_frame && frame_type == OD_P_FRAME)
+      frame_subtype = OD_ALTREF_P_FRAME;
+    if (bits <= 0) {
+      /*We didn't code any blocks in this frame.*/
+      log_scale = OD_Q57(-64);
+      bits = 0;
+      ++rc->prev_drop_count[frame_subtype];
+    } else {
+      int64_t log_bits;
+      int64_t log_qexp;
+      /*Compute the estimated scale factor for this frame type.*/
+      log_bits = od_blog64(bits);
+      log_qexp = od_blog64(rc->target_quantizer);
+      log_qexp = (log_qexp >> 6) * (rc->exp[frame_type]);
+      log_scale = OD_MINI(log_bits - rc->log_npixels + log_qexp, OD_Q57(16));
+    }
+
+    switch (rc->twopass_state) {
+      case 1: {
+        int golden, altref;
+        int64_t ipc;
+        rc->cur_metrics.frame_type =
+            od_frame_type(rc, rc->cur_frame, &golden, &altref, &ipc);
+        /*Pass 1 mode: save the metrics for this frame.*/
+        rc->cur_metrics.log_scale = od_q57_to_q24(log_scale);
+      } break;
+      case 2: {
+        /*Pass 2 mode:*/
+        int m_frame_type = rc->cur_metrics.frame_type;
+        rc->nframes[m_frame_type]--;
+        rc->scale_sum[m_frame_type] -= od_bexp64_q24(rc->cur_metrics.log_scale);
+      } break;
+    }
+
+    if (bits > 0) {
+      od_iir_bessel2 *f;
+      /*If this is the first example of the given frame type we've
+         seen, we immediately replace the default scale factor guess
+         with the estimate we just computed using the first frame.*/
+      if (rc->frame_count[frame_type] == 0) {
+        f = rc->scalefilter + frame_type;
+        f->y[1] = f->y[0] = f->x[1] = f->x[0] = od_q57_to_q24(log_scale);
+        rc->log_scale[frame_type] = log_scale;
+      } else {
+        /*Lengthen the time constant for the inter filters as we collect more
+           frame statistics, until we reach our target.*/
+        if (frame_type != OD_I_FRAME &&
+            rc->inter_p_delay < rc->inter_delay_target &&
+            rc->frame_count[frame_type] >= rc->inter_p_delay) {
+          od_iir_bessel2_reinit(&rc->scalefilter[frame_type],
+                                ++rc->inter_p_delay);
+        }
+        /*Update the low-pass scale filter for this frame type
+           regardless of whether or not we drop this frame.*/
+        rc->log_scale[frame_type] =
+            od_iir_bessel2_update(rc->scalefilter + frame_type,
+                                  od_q57_to_q24(log_scale))
+            << 33;
+      }
+      /*If this frame busts our budget, it must be dropped.*/
+      if (droppable && rc->reservoir_fullness + rc->bits_per_frame < bits) {
+        ++rc->prev_drop_count[frame_subtype];
+        bits = 0;
+        dropped = 1;
+      } else {
+        uint32_t drop_count;
+        /*Update a low-pass filter to estimate the "real" frame rate taking
+           drops into account.
+          This is only done if the frame is coded, as it needs the final
+           count of dropped frames.*/
+        drop_count = rc->prev_drop_count[frame_subtype] + 1;
+        if (drop_count > 0x7F) {
+          drop_count = 0x7FFFFFFF;
+        } else {
+          drop_count <<= 24;
+        }
+        rc->log_drop_scale[frame_subtype] =
+            od_blog64(od_iir_bessel2_update(rc->vfrfilter + frame_subtype,
+                                            drop_count)) -
+            OD_Q57(24);
+        /*Zero the drop count for this frame.
+          It will be increased if we drop frames.*/
+        rc->prev_drop_count[frame_subtype] = 0;
+      }
+      /*Increment the frame count for filter adaptation purposes.*/
+      if (!rc->twopass_state) rc->frame_count[frame_type]++;
+    }
+    rc->reservoir_fullness += rc->bits_per_frame - bits;
+    /*If we're too quick filling the buffer and overflow is capped,
+      that rate is lost forever.*/
+    if (rc->cap_overflow && rc->reservoir_fullness > rc->reservoir_max) {
+      rc->reservoir_fullness = rc->reservoir_max;
+    }
+    /*If we're too quick draining the buffer and underflow is capped,
+      don't try to make up that rate later.*/
+    if (rc->cap_underflow && rc->reservoir_fullness < 0) {
+      rc->reservoir_fullness = 0;
+    }
+    /*Adjust the bias for the real bits we've used.*/
+    rc->rate_bias -= bits;
+  }
+  return dropped;
+}
+
+static INLINE void od_rc_buffer_val(od_rc_state *rc, int64_t val, int bytes) {
+  while (bytes-- > 0) {
+    rc->twopass_buffer[rc->twopass_buffer_bytes++] = (uint8_t)(val & 0xFF);
+    val >>= 8;
+  }
+}
+
+static INLINE int64_t od_rc_unbuffer_val(od_rc_state *rc, int bytes) {
+  int64_t ret = 0;
+  int shift = 0;
+  while (bytes-- > 0) {
+    ret |= ((int64_t)rc->twopass_buffer[rc->twopass_buffer_bytes++]) << shift;
+    shift += 8;
+  }
+  return ret;
+}
+
+int od_enc_rc_2pass_out(od_rc_state *rc, struct aom_codec_pkt_list *pkt_list,
+                        int summary) {
+  int i;
+  struct aom_codec_cx_pkt pkt;
+  rc->twopass_buffer = rc->firstpass_buffer;
+  rc->twopass_buffer_bytes = 0;
+  if (!rc->twopass_state) {
+    rc->twopass_state = 1;
+    for (i = 0; i < OD_FRAME_NSUBTYPES; i++) {
+      rc->frame_count[i] = 0;
+      rc->exp[i] = 0;
+      rc->scale_sum[i] = 0;
+    }
+  }
+  if (summary) {
+    od_rc_buffer_val(rc, OD_RC_2PASS_MAGIC, 4);
+    od_rc_buffer_val(rc, OD_RC_2PASS_VERSION, 1);
+    for (i = 0; i < OD_FRAME_NSUBTYPES; i++) {
+      od_rc_buffer_val(rc, rc->frame_count[i], 4);
+      od_rc_buffer_val(rc, rc->exp[i], 4);
+      od_rc_buffer_val(rc, rc->scale_sum[i], 8);
+    }
+  } else {
+    int frame_type = rc->cur_metrics.frame_type;
+    rc->scale_sum[frame_type] += od_bexp64_q24(rc->cur_metrics.log_scale);
+    rc->frame_count[frame_type]++;
+    od_rc_buffer_val(rc, rc->cur_metrics.frame_type, 1);
+    od_rc_buffer_val(rc, rc->cur_metrics.log_scale, 4);
+  }
+  pkt.data.twopass_stats.buf = rc->firstpass_buffer;
+  pkt.data.twopass_stats.sz = rc->twopass_buffer_bytes;
+  pkt.kind = AOM_CODEC_STATS_PKT;
+  aom_codec_pkt_list_add(pkt_list, &pkt);
+  return 0;
+}
+
+int od_enc_rc_2pass_in(od_rc_state *rc) {
+  /* Enable pass 2 mode if this is the first call. */
+  if (rc->twopass_state == 0) {
+    uint32_t i, total_frames = 0;
+
+    if (!rc->twopass_allframes_buf ||
+        rc->twopass_allframes_buf_size < OD_RC_2PASS_MIN)
+      return -1;
+
+    /* Find summary packet at the end */
+    rc->twopass_buffer = rc->twopass_allframes_buf;
+    rc->twopass_buffer +=
+        rc->twopass_allframes_buf_size - OD_RC_2PASS_SUMMARY_SZ;
+    rc->twopass_buffer_bytes = 0;
+
+    if (od_rc_unbuffer_val(rc, 4) != OD_RC_2PASS_MAGIC) return -1;
+    if (od_rc_unbuffer_val(rc, 1) != OD_RC_2PASS_VERSION) return -1;
+
+    for (i = 0; i < OD_FRAME_NSUBTYPES; i++) {
+      rc->frame_count[i] = od_rc_unbuffer_val(rc, 4);
+      rc->exp[i] = od_rc_unbuffer_val(rc, 4);
+      rc->scale_sum[i] = od_rc_unbuffer_val(rc, 8);
+      rc->nframes[i] = rc->frame_count[i];
+      total_frames += rc->frame_count[i];
+    }
+
+    if (total_frames < 1) return -1;
+
+    if (total_frames * OD_RC_2PASS_PACKET_SZ > rc->twopass_allframes_buf_size)
+      return -1;
+
+    od_enc_rc_reset(rc);
+
+    /* Everything looks ok */
+    rc->twopass_buffer = rc->twopass_allframes_buf;
+    rc->twopass_state = 2;
+    rc->twopass_buffer_bytes = 0;
+  }
+
+  rc->cur_metrics.frame_type = od_rc_unbuffer_val(rc, 1);
+  rc->cur_metrics.log_scale = od_rc_unbuffer_val(rc, 4);
+
+  return 0;
+}
diff --git a/third_party/aom/av1/encoder/ratectrl_xiph.h b/third_party/aom/av1/encoder/ratectrl_xiph.h
new file mode 100644
index 0000000000..a4a9052faa
--- /dev/null
+++ b/third_party/aom/av1/encoder/ratectrl_xiph.h
@@ -0,0 +1,200 @@
+/*
+ * Copyright (c) 2001-2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if !defined(_ratectrl_xiph_H)
+#define _ratectrl_xiph_H (1)
+
+#include "av1/encoder/ratectrl.h"
+#include "aom/internal/aom_codec_internal.h"
+
+/*Frame types.*/
+#define OD_I_FRAME (0)
+#define OD_P_FRAME (1)
+#define OD_GOLDEN_P_FRAME (2)
+#define OD_ALTREF_P_FRAME (3)
+
+#define OD_FRAME_NSUBTYPES (OD_ALTREF_P_FRAME + 1)
+
+/* Periodic boost (in between golden frames) strength - lower is more */
+#define OD_PERIODIC_BOOST_DIV (10)
+
+/* Constants for frame QP modulation <- tweak these
+ * Adjusts how the rate control system decides the quantizers per frame
+ * (sub)type */
+#define OD_MQP_I (0.98)
+#define OD_MQP_P (1.06)
+#define OD_MQP_GP (0.99)
+#define OD_MQP_AP (0.92)
+#define OD_DQP_I (-2)
+#define OD_DQP_P (0)
+#define OD_DQP_GP (-2)
+#define OD_DQP_AP (-2)
+
+/*Fractional_coded_quantizer ~=
+   log2(quantizer / (1 << OD_COEFF_SHIFT))*6.307 + 6.235*/
+/*Base/scale factor for linear quantizer to fractional coded quantizer
+   conversion (6.307 * 2^12) */
+#define OD_LOG_QUANTIZER_BASE_Q12 (0x0064EB)
+/*Inverse of above scale factor.*/
+#define OD_LOG_QUANTIZER_EXP_Q12 (0x000289)
+/*Offset for linear quantizer to fractional coded quantizer
+   conversion (6.235 * 2^45) */
+#define OD_LOG_QUANTIZER_OFFSET_Q45 (0x0000C7851EB851ECLL)
+
+#define OD_RC_2PASS_MAGIC (0x53015641) /* [A, V, 1, S] in little endian */
+#define OD_RC_2PASS_SUMMARY_SZ (4 + 1 + (4 + 4 + 8) * OD_FRAME_NSUBTYPES)
+#define OD_RC_2PASS_PACKET_SZ (1 + 4)
+#define OD_RC_2PASS_MIN (OD_RC_2PASS_PACKET_SZ + OD_RC_2PASS_SUMMARY_SZ)
+#define OD_RC_2PASS_VERSION (1)
+
+/*A 2nd order low-pass Bessel follower.
+  We use this for rate control because it has fast reaction time, but is
+   critically damped.*/
+typedef struct od_iir_bessel2 {
+  int32_t c[2];
+  int64_t g;
+  int32_t x[2];
+  int32_t y[2];
+} od_iir_bessel2;
+
+/* The 2-pass metrics associated with a single frame. */
+typedef struct od_frame_metrics {
+  /*The log base 2 of the scale factor for this frame in Q24 format.*/
+  int64_t log_scale;
+  /*The frame type from pass 1.*/
+  unsigned frame_type : 1;
+} od_frame_metrics;
+
+/*Rate control setup and working state information.*/
+typedef struct od_rc_state {
+  /* Image format */
+  int frame_width;
+  int frame_height;
+  int bit_depth;
+
+  /* Framerate */
+  double framerate;
+  /* Keyframe rate */
+  int keyframe_rate;
+  /* Golden frame period */
+  int goldenframe_rate;
+  /* Altref frame period */
+  int altref_rate;
+  /*The target bit-rate in bits per second.*/
+  int64_t target_bitrate;
+  /* Quality level for non-bitrate-targeting */
+  int quality;
+  /* Copied from oxcf->frame_periodic_boost */
+  int periodic_boosts;
+  /* Max Q */
+  int maxq;
+  /* Min Q */
+  int minq;
+  /* Quantizer to use for the first pass */
+  int firstpass_quant;
+
+  /* 2-pass metrics */
+  od_frame_metrics cur_metrics;
+
+  /* 2-pass state */
+  int64_t scale_sum[OD_FRAME_NSUBTYPES];
+  int nframes[OD_FRAME_NSUBTYPES];
+
+  /* 2-pass bytestream reader/writer context */
+  uint8_t *twopass_buffer;
+  int twopass_buffer_bytes;
+
+  /* Pass 1 stats packet storage */
+  uint8_t firstpass_buffer[OD_RC_2PASS_SUMMARY_SZ];
+
+  /* Every state packet from the first pass in a single buffer */
+  uint8_t *twopass_allframes_buf;
+  size_t twopass_allframes_buf_size;
+
+  /* Actual returned quantizer */
+  int target_quantizer;
+  /*The full-precision, unmodulated quantizer upon which
+    our modulated quantizers are based.*/
+  int base_quantizer;
+
+  /* Increments by 1 for each frame. */
+  int64_t cur_frame;
+
+  /* End of input flag */
+  int end_of_input;
+  /* Closed GOP flag */
+  int closed_gop;
+  /*The number of frames over which to distribute the reservoir usage.*/
+  int reservoir_frame_delay;
+  /*Will we drop frames to meet bitrate target?*/
+  unsigned char drop_frames;
+  /*Do we respect the maximum reservoir fullness?*/
+  unsigned char cap_overflow;
+  /*Can the reservoir go negative?*/
+  unsigned char cap_underflow;
+  /*Two-pass mode state.
+    0 => 1-pass encoding.
+    1 => 1st pass of 2-pass encoding.
+    2 => 2nd pass of 2-pass encoding.*/
+  int twopass_state;
+  /*The log of the number of pixels in a frame in Q57 format.*/
+  int64_t log_npixels;
+  /*The target average bits per frame.*/
+  int64_t bits_per_frame;
+  /*The current bit reservoir fullness (bits available to be used).*/
+  int64_t reservoir_fullness;
+  /*The target buffer fullness.
+    This is where we'd like to be by the last keyframe the appears in the next
+     buf_delay frames.*/
+  int64_t reservoir_target;
+  /*The maximum buffer fullness (total size of the buffer).*/
+  int64_t reservoir_max;
+  /*The log of estimated scale factor for the rate model in Q57 format.*/
+  int64_t log_scale[OD_FRAME_NSUBTYPES];
+  /*The exponent used in the rate model in Q8 format.*/
+  unsigned exp[OD_FRAME_NSUBTYPES];
+  /*The log of an estimated scale factor used to obtain the real framerate, for
+     VFR sources or, e.g., 12 fps content doubled to 24 fps, etc.*/
+  int64_t log_drop_scale[OD_FRAME_NSUBTYPES];
+  /*The total drop count from the previous frame.*/
+  uint32_t prev_drop_count[OD_FRAME_NSUBTYPES];
+  /*Second-order lowpass filters to track scale and VFR/drops.*/
+  od_iir_bessel2 scalefilter[OD_FRAME_NSUBTYPES];
+  od_iir_bessel2 vfrfilter[OD_FRAME_NSUBTYPES];
+  int frame_count[OD_FRAME_NSUBTYPES];
+  int inter_p_delay;
+  int inter_delay_target;
+  /*The total accumulated estimation bias.*/
+  int64_t rate_bias;
+} od_rc_state;
+
+int od_enc_rc_init(od_rc_state *rc, int64_t bitrate, int delay_ms);
+
+int od_enc_rc_select_quantizers_and_lambdas(od_rc_state *rc,
+                                            int is_golden_frame,
+                                            int is_altref_frame, int frame_type,
+                                            int *bottom_idx, int *top_idx);
+
+/* Returns 1 if the frame should be dropped */
+int od_enc_rc_update_state(od_rc_state *rc, int64_t bits, int is_golden_frame,
+                           int is_altref_frame, int frame_type, int droppable);
+
+int od_frame_type(od_rc_state *rc, int64_t coding_frame_count, int *is_golden,
+                  int *is_altref, int64_t *ip_count);
+
+int od_enc_rc_resize(od_rc_state *rc);
+
+int od_enc_rc_2pass_out(od_rc_state *rc, struct aom_codec_pkt_list *pkt_list,
+                        int summary);
+
+int od_enc_rc_2pass_in(od_rc_state *rc);
+
+#endif
diff --git a/third_party/aom/av1/encoder/rd.c b/third_party/aom/av1/encoder/rd.c
new file mode 100644
index 0000000000..f06e569e7f
--- /dev/null
+++ b/third_party/aom/av1/encoder/rd.c
@@ -0,0 +1,1204 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+
+#include "./av1_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/bitops.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/system_state.h"
+
+#include "av1/common/common.h"
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/mvref_common.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+#include "av1/common/seg_common.h"
+
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/cost.h"
+#include "av1/encoder/encodemb.h"
+#include "av1/encoder/encodemv.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/mcomp.h"
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/rd.h"
+#include "av1/encoder/tokenize.h"
+
+#define RD_THRESH_POW 1.25
+
+// Factor to weigh the rate for switchable interp filters.
+#define SWITCHABLE_INTERP_RATE_FACTOR 1
+
+// The baseline rd thresholds for breaking out of the rd loop for
+// certain modes are assumed to be based on 8x8 blocks.
+// This table is used to correct for block size.
+// The factors here are << 2 (2 = x0.5, 32 = x8 etc).
+static const uint8_t rd_thresh_block_size_factor[BLOCK_SIZES] = {
+#if CONFIG_CB4X4
+  2,  2,  2,
+#endif
+  2,  3,  3, 4, 6, 6, 8, 12, 12, 16, 24, 24, 32,
+#if CONFIG_EXT_PARTITION
+  48, 48, 64
+#endif  // CONFIG_EXT_PARTITION
+};
+
+static void fill_mode_costs(AV1_COMP *cpi) {
+  const FRAME_CONTEXT *const fc = cpi->common.fc;
+  int i, j;
+
+  for (i = 0; i < INTRA_MODES; ++i)
+    for (j = 0; j < INTRA_MODES; ++j)
+      av1_cost_tokens(cpi->y_mode_costs[i][j], av1_kf_y_mode_prob[i][j],
+                      av1_intra_mode_tree);
+
+  for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
+    av1_cost_tokens(cpi->mbmode_cost[i], fc->y_mode_prob[i],
+                    av1_intra_mode_tree);
+
+  for (i = 0; i < INTRA_MODES; ++i)
+    av1_cost_tokens(cpi->intra_uv_mode_cost[i], fc->uv_mode_prob[i],
+                    av1_intra_mode_tree);
+
+  for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
+    av1_cost_tokens(cpi->switchable_interp_costs[i],
+                    fc->switchable_interp_prob[i], av1_switchable_interp_tree);
+
+#if CONFIG_PALETTE
+  for (i = 0; i < PALETTE_BLOCK_SIZES; ++i) {
+    av1_cost_tokens(cpi->palette_y_size_cost[i],
+                    av1_default_palette_y_size_prob[i], av1_palette_size_tree);
+    av1_cost_tokens(cpi->palette_uv_size_cost[i],
+                    av1_default_palette_uv_size_prob[i], av1_palette_size_tree);
+  }
+
+  for (i = 0; i < PALETTE_SIZES; ++i) {
+    for (j = 0; j < PALETTE_COLOR_INDEX_CONTEXTS; ++j) {
+      av1_cost_tokens(cpi->palette_y_color_cost[i][j],
+                      av1_default_palette_y_color_index_prob[i][j],
+                      av1_palette_color_index_tree[i]);
+      av1_cost_tokens(cpi->palette_uv_color_cost[i][j],
+                      av1_default_palette_uv_color_index_prob[i][j],
+                      av1_palette_color_index_tree[i]);
+    }
+  }
+#endif  // CONFIG_PALETTE
+
+  for (i = 0; i < MAX_TX_DEPTH; ++i)
+    for (j = 0; j < TX_SIZE_CONTEXTS; ++j)
+      av1_cost_tokens(cpi->tx_size_cost[i][j], fc->tx_size_probs[i][j],
+                      av1_tx_size_tree[i]);
+
+#if CONFIG_EXT_TX
+  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+    int s;
+    for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
+      if (use_inter_ext_tx_for_txsize[s][i]) {
+        av1_cost_tokens(cpi->inter_tx_type_costs[s][i],
+                        fc->inter_ext_tx_prob[s][i], av1_ext_tx_inter_tree[s]);
+      }
+    }
+    for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
+      if (use_intra_ext_tx_for_txsize[s][i]) {
+        for (j = 0; j < INTRA_MODES; ++j)
+          av1_cost_tokens(cpi->intra_tx_type_costs[s][i][j],
+                          fc->intra_ext_tx_prob[s][i][j],
+                          av1_ext_tx_intra_tree[s]);
+      }
+    }
+  }
+#else
+  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+    for (j = 0; j < TX_TYPES; ++j)
+      av1_cost_tokens(cpi->intra_tx_type_costs[i][j],
+                      fc->intra_ext_tx_prob[i][j], av1_ext_tx_tree);
+  }
+  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
+    av1_cost_tokens(cpi->inter_tx_type_costs[i], fc->inter_ext_tx_prob[i],
+                    av1_ext_tx_tree);
+  }
+#endif  // CONFIG_EXT_TX
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+  for (i = 0; i < INTRA_FILTERS + 1; ++i)
+    av1_cost_tokens(cpi->intra_filter_cost[i], fc->intra_filter_probs[i],
+                    av1_intra_filter_tree);
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_LOOP_RESTORATION
+  av1_cost_tokens(cpi->switchable_restore_cost, fc->switchable_restore_prob,
+                  av1_switchable_restore_tree);
+#endif  // CONFIG_LOOP_RESTORATION
+#if CONFIG_GLOBAL_MOTION
+  av1_cost_tokens(cpi->gmtype_cost, fc->global_motion_types_prob,
+                  av1_global_motion_types_tree);
+#endif  // CONFIG_GLOBAL_MOTION
+}
+
+void av1_fill_token_costs(av1_coeff_cost *c,
+                          av1_coeff_probs_model (*p)[PLANE_TYPES]) {
+  int i, j, k, l;
+  TX_SIZE t;
+  for (t = 0; t < TX_SIZES; ++t)
+    for (i = 0; i < PLANE_TYPES; ++i)
+      for (j = 0; j < REF_TYPES; ++j)
+        for (k = 0; k < COEF_BANDS; ++k)
+          for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
+            aom_prob probs[ENTROPY_NODES];
+            av1_model_to_full_probs(p[t][i][j][k][l], probs);
+            av1_cost_tokens((int *)c[t][i][j][k][0][l], probs, av1_coef_tree);
+            av1_cost_tokens_skip((int *)c[t][i][j][k][1][l], probs,
+                                 av1_coef_tree);
+            assert(c[t][i][j][k][0][l][EOB_TOKEN] ==
+                   c[t][i][j][k][1][l][EOB_TOKEN]);
+          }
+}
+
+// Values are now correlated to quantizer.
+static int sad_per_bit16lut_8[QINDEX_RANGE];
+static int sad_per_bit4lut_8[QINDEX_RANGE];
+
+#if CONFIG_HIGHBITDEPTH
+static int sad_per_bit16lut_10[QINDEX_RANGE];
+static int sad_per_bit4lut_10[QINDEX_RANGE];
+static int sad_per_bit16lut_12[QINDEX_RANGE];
+static int sad_per_bit4lut_12[QINDEX_RANGE];
+#endif
+
+static void init_me_luts_bd(int *bit16lut, int *bit4lut, int range,
+                            aom_bit_depth_t bit_depth) {
+  int i;
+  // Initialize the sad lut tables using a formulaic calculation for now.
+  // This is to make it easier to resolve the impact of experimental changes
+  // to the quantizer tables.
+  for (i = 0; i < range; i++) {
+    const double q = av1_convert_qindex_to_q(i, bit_depth);
+    bit16lut[i] = (int)(0.0418 * q + 2.4107);
+    bit4lut[i] = (int)(0.063 * q + 2.742);
+  }
+}
+
+void av1_init_me_luts(void) {
+  init_me_luts_bd(sad_per_bit16lut_8, sad_per_bit4lut_8, QINDEX_RANGE,
+                  AOM_BITS_8);
+#if CONFIG_HIGHBITDEPTH
+  init_me_luts_bd(sad_per_bit16lut_10, sad_per_bit4lut_10, QINDEX_RANGE,
+                  AOM_BITS_10);
+  init_me_luts_bd(sad_per_bit16lut_12, sad_per_bit4lut_12, QINDEX_RANGE,
+                  AOM_BITS_12);
+#endif
+}
+
+static const int rd_boost_factor[16] = { 64, 32, 32, 32, 24, 16, 12, 12,
+                                         8,  8,  4,  4,  2,  2,  1,  0 };
+static const int rd_frame_type_factor[FRAME_UPDATE_TYPES] = {
+  128, 144, 128, 128, 144,
+#if CONFIG_EXT_REFS
+  // TODO(zoeliu): To adjust further following factor values.
+  128, 128, 128
+  // TODO(weitinglin): We should investigate if the values should be the same
+  //                   as the value used by OVERLAY frame
+  ,
+  144
+#endif  // CONFIG_EXT_REFS
+};
+
+int av1_compute_rd_mult(const AV1_COMP *cpi, int qindex) {
+  const int64_t q = av1_dc_quant(qindex, 0, cpi->common.bit_depth);
+#if CONFIG_HIGHBITDEPTH
+  int64_t rdmult = 0;
+  switch (cpi->common.bit_depth) {
+    case AOM_BITS_8: rdmult = 88 * q * q / 24; break;
+    case AOM_BITS_10: rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 4); break;
+    case AOM_BITS_12: rdmult = ROUND_POWER_OF_TWO(88 * q * q / 24, 8); break;
+    default:
+      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
+      return -1;
+  }
+#else
+  int64_t rdmult = 88 * q * q / 24;
+#endif  // CONFIG_HIGHBITDEPTH
+  if (cpi->oxcf.pass == 2 && (cpi->common.frame_type != KEY_FRAME)) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    const FRAME_UPDATE_TYPE frame_type = gf_group->update_type[gf_group->index];
+    const int boost_index = AOMMIN(15, (cpi->rc.gfu_boost / 100));
+
+    rdmult = (rdmult * rd_frame_type_factor[frame_type]) >> 7;
+    rdmult += ((rdmult * rd_boost_factor[boost_index]) >> 7);
+  }
+  if (rdmult < 1) rdmult = 1;
+  return (int)rdmult;
+}
+
+static int compute_rd_thresh_factor(int qindex, aom_bit_depth_t bit_depth) {
+  double q;
+#if CONFIG_HIGHBITDEPTH
+  switch (bit_depth) {
+    case AOM_BITS_8: q = av1_dc_quant(qindex, 0, AOM_BITS_8) / 4.0; break;
+    case AOM_BITS_10: q = av1_dc_quant(qindex, 0, AOM_BITS_10) / 16.0; break;
+    case AOM_BITS_12: q = av1_dc_quant(qindex, 0, AOM_BITS_12) / 64.0; break;
+    default:
+      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
+      return -1;
+  }
+#else
+  (void)bit_depth;
+  q = av1_dc_quant(qindex, 0, AOM_BITS_8) / 4.0;
+#endif  // CONFIG_HIGHBITDEPTH
+  // TODO(debargha): Adjust the function below.
+  return AOMMAX((int)(pow(q, RD_THRESH_POW) * 5.12), 8);
+}
+
+void av1_initialize_me_consts(const AV1_COMP *cpi, MACROBLOCK *x, int qindex) {
+#if CONFIG_HIGHBITDEPTH
+  switch (cpi->common.bit_depth) {
+    case AOM_BITS_8:
+      x->sadperbit16 = sad_per_bit16lut_8[qindex];
+      x->sadperbit4 = sad_per_bit4lut_8[qindex];
+      break;
+    case AOM_BITS_10:
+      x->sadperbit16 = sad_per_bit16lut_10[qindex];
+      x->sadperbit4 = sad_per_bit4lut_10[qindex];
+      break;
+    case AOM_BITS_12:
+      x->sadperbit16 = sad_per_bit16lut_12[qindex];
+      x->sadperbit4 = sad_per_bit4lut_12[qindex];
+      break;
+    default:
+      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
+  }
+#else
+  (void)cpi;
+  x->sadperbit16 = sad_per_bit16lut_8[qindex];
+  x->sadperbit4 = sad_per_bit4lut_8[qindex];
+#endif  // CONFIG_HIGHBITDEPTH
+}
+
+static void set_block_thresholds(const AV1_COMMON *cm, RD_OPT *rd) {
+  int i, bsize, segment_id;
+
+  for (segment_id = 0; segment_id < MAX_SEGMENTS; ++segment_id) {
+    const int qindex =
+        clamp(av1_get_qindex(&cm->seg, segment_id, cm->base_qindex) +
+                  cm->y_dc_delta_q,
+              0, MAXQ);
+    const int q = compute_rd_thresh_factor(qindex, cm->bit_depth);
+
+    for (bsize = 0; bsize < BLOCK_SIZES; ++bsize) {
+      // Threshold here seems unnecessarily harsh but fine given actual
+      // range of values used for cpi->sf.thresh_mult[].
+      const int t = q * rd_thresh_block_size_factor[bsize];
+      const int thresh_max = INT_MAX / t;
+
+#if CONFIG_CB4X4
+      for (i = 0; i < MAX_MODES; ++i)
+        rd->threshes[segment_id][bsize][i] = rd->thresh_mult[i] < thresh_max
+                                                 ? rd->thresh_mult[i] * t / 4
+                                                 : INT_MAX;
+#else
+      if (bsize >= BLOCK_8X8) {
+        for (i = 0; i < MAX_MODES; ++i)
+          rd->threshes[segment_id][bsize][i] = rd->thresh_mult[i] < thresh_max
+                                                   ? rd->thresh_mult[i] * t / 4
+                                                   : INT_MAX;
+      } else {
+        for (i = 0; i < MAX_REFS; ++i)
+          rd->threshes[segment_id][bsize][i] =
+              rd->thresh_mult_sub8x8[i] < thresh_max
+                  ? rd->thresh_mult_sub8x8[i] * t / 4
+                  : INT_MAX;
+      }
+#endif
+    }
+  }
+}
+
+#if CONFIG_REF_MV
+void av1_set_mvcost(MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame, int ref,
+                    int ref_mv_idx) {
+  MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
+  int8_t rf_type = av1_ref_frame_type(x->e_mbd.mi[0]->mbmi.ref_frame);
+  int nmv_ctx = av1_nmv_ctx(mbmi_ext->ref_mv_count[rf_type],
+                            mbmi_ext->ref_mv_stack[rf_type], ref, ref_mv_idx);
+  (void)ref_frame;
+  x->mvcost = x->mv_cost_stack[nmv_ctx];
+  x->nmvjointcost = x->nmv_vec_cost[nmv_ctx];
+  x->mvsadcost = x->mvcost;
+  x->nmvjointsadcost = x->nmvjointcost;
+}
+#endif
+
+void av1_initialize_rd_consts(AV1_COMP *cpi) {
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->td.mb;
+  RD_OPT *const rd = &cpi->rd;
+  int i;
+#if CONFIG_REF_MV
+  int nmv_ctx;
+#endif
+
+  aom_clear_system_state();
+
+  rd->RDDIV = RDDIV_BITS;  // In bits (to multiply D by 128).
+  rd->RDMULT = av1_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);
+
+  set_error_per_bit(x, rd->RDMULT);
+
+  set_block_thresholds(cm, rd);
+
+#if CONFIG_REF_MV
+  for (nmv_ctx = 0; nmv_ctx < NMV_CONTEXTS; ++nmv_ctx) {
+    av1_build_nmv_cost_table(
+        x->nmv_vec_cost[nmv_ctx],
+        cm->allow_high_precision_mv ? x->nmvcost_hp[nmv_ctx]
+                                    : x->nmvcost[nmv_ctx],
+        &cm->fc->nmvc[nmv_ctx], cm->allow_high_precision_mv);
+  }
+  x->mvcost = x->mv_cost_stack[0];
+  x->nmvjointcost = x->nmv_vec_cost[0];
+  x->mvsadcost = x->mvcost;
+  x->nmvjointsadcost = x->nmvjointcost;
+#else
+  av1_build_nmv_cost_table(
+      x->nmvjointcost, cm->allow_high_precision_mv ? x->nmvcost_hp : x->nmvcost,
+      &cm->fc->nmvc, cm->allow_high_precision_mv);
+#endif
+
+  if (cpi->oxcf.pass != 1) {
+    av1_fill_token_costs(x->token_costs, cm->fc->coef_probs);
+
+    if (cpi->sf.partition_search_type != VAR_BASED_PARTITION ||
+        cm->frame_type == KEY_FRAME) {
+#if CONFIG_EXT_PARTITION_TYPES
+      for (i = 0; i < PARTITION_PLOFFSET; ++i)
+        av1_cost_tokens(cpi->partition_cost[i], cm->fc->partition_prob[i],
+                        av1_partition_tree);
+      for (; i < PARTITION_CONTEXTS_PRIMARY; ++i)
+        av1_cost_tokens(cpi->partition_cost[i], cm->fc->partition_prob[i],
+                        av1_ext_partition_tree);
+#else
+      for (i = 0; i < PARTITION_CONTEXTS_PRIMARY; ++i)
+        av1_cost_tokens(cpi->partition_cost[i], cm->fc->partition_prob[i],
+                        av1_partition_tree);
+#endif  // CONFIG_EXT_PARTITION_TYPES
+#if CONFIG_UNPOISON_PARTITION_CTX
+      for (; i < PARTITION_CONTEXTS_PRIMARY + PARTITION_BLOCK_SIZES; ++i) {
+        aom_prob p = cm->fc->partition_prob[i][PARTITION_VERT];
+        assert(p > 0);
+        cpi->partition_cost[i][PARTITION_NONE] = INT_MAX;
+        cpi->partition_cost[i][PARTITION_HORZ] = INT_MAX;
+        cpi->partition_cost[i][PARTITION_VERT] = av1_cost_bit(p, 0);
+        cpi->partition_cost[i][PARTITION_SPLIT] = av1_cost_bit(p, 1);
+      }
+      for (; i < PARTITION_CONTEXTS_PRIMARY + 2 * PARTITION_BLOCK_SIZES; ++i) {
+        aom_prob p = cm->fc->partition_prob[i][PARTITION_HORZ];
+        assert(p > 0);
+        cpi->partition_cost[i][PARTITION_NONE] = INT_MAX;
+        cpi->partition_cost[i][PARTITION_HORZ] = av1_cost_bit(p, 0);
+        cpi->partition_cost[i][PARTITION_VERT] = INT_MAX;
+        cpi->partition_cost[i][PARTITION_SPLIT] = av1_cost_bit(p, 1);
+      }
+      cpi->partition_cost[PARTITION_CONTEXTS][PARTITION_NONE] = INT_MAX;
+      cpi->partition_cost[PARTITION_CONTEXTS][PARTITION_HORZ] = INT_MAX;
+      cpi->partition_cost[PARTITION_CONTEXTS][PARTITION_VERT] = INT_MAX;
+      cpi->partition_cost[PARTITION_CONTEXTS][PARTITION_SPLIT] = 0;
+#endif  // CONFIG_UNPOISON_PARTITION_CTX
+    }
+
+    fill_mode_costs(cpi);
+
+    if (!frame_is_intra_only(cm)) {
+#if CONFIG_REF_MV
+      for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i) {
+        cpi->newmv_mode_cost[i][0] = av1_cost_bit(cm->fc->newmv_prob[i], 0);
+        cpi->newmv_mode_cost[i][1] = av1_cost_bit(cm->fc->newmv_prob[i], 1);
+      }
+
+      for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i) {
+        cpi->zeromv_mode_cost[i][0] = av1_cost_bit(cm->fc->zeromv_prob[i], 0);
+        cpi->zeromv_mode_cost[i][1] = av1_cost_bit(cm->fc->zeromv_prob[i], 1);
+      }
+
+      for (i = 0; i < REFMV_MODE_CONTEXTS; ++i) {
+        cpi->refmv_mode_cost[i][0] = av1_cost_bit(cm->fc->refmv_prob[i], 0);
+        cpi->refmv_mode_cost[i][1] = av1_cost_bit(cm->fc->refmv_prob[i], 1);
+      }
+
+      for (i = 0; i < DRL_MODE_CONTEXTS; ++i) {
+        cpi->drl_mode_cost0[i][0] = av1_cost_bit(cm->fc->drl_prob[i], 0);
+        cpi->drl_mode_cost0[i][1] = av1_cost_bit(cm->fc->drl_prob[i], 1);
+      }
+#else
+      for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
+        av1_cost_tokens((int *)cpi->inter_mode_cost[i],
+                        cm->fc->inter_mode_probs[i], av1_inter_mode_tree);
+#endif  // CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+      for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
+        av1_cost_tokens((int *)cpi->inter_compound_mode_cost[i],
+                        cm->fc->inter_compound_mode_probs[i],
+                        av1_inter_compound_mode_tree);
+      for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
+        av1_cost_tokens((int *)cpi->interintra_mode_cost[i],
+                        cm->fc->interintra_mode_prob[i],
+                        av1_interintra_mode_tree);
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      for (i = BLOCK_8X8; i < BLOCK_SIZES; i++) {
+        av1_cost_tokens((int *)cpi->motion_mode_cost[i],
+                        cm->fc->motion_mode_prob[i], av1_motion_mode_tree);
+      }
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+      for (i = BLOCK_8X8; i < BLOCK_SIZES; i++) {
+        cpi->motion_mode_cost1[i][0] = av1_cost_bit(cm->fc->obmc_prob[i], 0);
+        cpi->motion_mode_cost1[i][1] = av1_cost_bit(cm->fc->obmc_prob[i], 1);
+      }
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    }
+  }
+}
+
+static void model_rd_norm(int xsq_q10, int *r_q10, int *d_q10) {
+  // NOTE: The tables below must be of the same size.
+
+  // The functions described below are sampled at the four most significant
+  // bits of x^2 + 8 / 256.
+
+  // Normalized rate:
+  // This table models the rate for a Laplacian source with given variance
+  // when quantized with a uniform quantizer with given stepsize. The
+  // closed form expression is:
+  // Rn(x) = H(sqrt(r)) + sqrt(r)*[1 + H(r)/(1 - r)],
+  // where r = exp(-sqrt(2) * x) and x = qpstep / sqrt(variance),
+  // and H(x) is the binary entropy function.
+  static const int rate_tab_q10[] = {
+    65536, 6086, 5574, 5275, 5063, 4899, 4764, 4651, 4553, 4389, 4255, 4142,
+    4044,  3958, 3881, 3811, 3748, 3635, 3538, 3453, 3376, 3307, 3244, 3186,
+    3133,  3037, 2952, 2877, 2809, 2747, 2690, 2638, 2589, 2501, 2423, 2353,
+    2290,  2232, 2179, 2130, 2084, 2001, 1928, 1862, 1802, 1748, 1698, 1651,
+    1608,  1530, 1460, 1398, 1342, 1290, 1243, 1199, 1159, 1086, 1021, 963,
+    911,   864,  821,  781,  745,  680,  623,  574,  530,  490,  455,  424,
+    395,   345,  304,  269,  239,  213,  190,  171,  154,  126,  104,  87,
+    73,    61,   52,   44,   38,   28,   21,   16,   12,   10,   8,    6,
+    5,     3,    2,    1,    1,    1,    0,    0,
+  };
+  // Normalized distortion:
+  // This table models the normalized distortion for a Laplacian source
+  // with given variance when quantized with a uniform quantizer
+  // with given stepsize. The closed form expression is:
+  // Dn(x) = 1 - 1/sqrt(2) * x / sinh(x/sqrt(2))
+  // where x = qpstep / sqrt(variance).
+  // Note the actual distortion is Dn * variance.
+  static const int dist_tab_q10[] = {
+    0,    0,    1,    1,    1,    2,    2,    2,    3,    3,    4,    5,
+    5,    6,    7,    7,    8,    9,    11,   12,   13,   15,   16,   17,
+    18,   21,   24,   26,   29,   31,   34,   36,   39,   44,   49,   54,
+    59,   64,   69,   73,   78,   88,   97,   106,  115,  124,  133,  142,
+    151,  167,  184,  200,  215,  231,  245,  260,  274,  301,  327,  351,
+    375,  397,  418,  439,  458,  495,  528,  559,  587,  613,  637,  659,
+    680,  717,  749,  777,  801,  823,  842,  859,  874,  899,  919,  936,
+    949,  960,  969,  977,  983,  994,  1001, 1006, 1010, 1013, 1015, 1017,
+    1018, 1020, 1022, 1022, 1023, 1023, 1023, 1024,
+  };
+  static const int xsq_iq_q10[] = {
+    0,      4,      8,      12,     16,     20,     24,     28,     32,
+    40,     48,     56,     64,     72,     80,     88,     96,     112,
+    128,    144,    160,    176,    192,    208,    224,    256,    288,
+    320,    352,    384,    416,    448,    480,    544,    608,    672,
+    736,    800,    864,    928,    992,    1120,   1248,   1376,   1504,
+    1632,   1760,   1888,   2016,   2272,   2528,   2784,   3040,   3296,
+    3552,   3808,   4064,   4576,   5088,   5600,   6112,   6624,   7136,
+    7648,   8160,   9184,   10208,  11232,  12256,  13280,  14304,  15328,
+    16352,  18400,  20448,  22496,  24544,  26592,  28640,  30688,  32736,
+    36832,  40928,  45024,  49120,  53216,  57312,  61408,  65504,  73696,
+    81888,  90080,  98272,  106464, 114656, 122848, 131040, 147424, 163808,
+    180192, 196576, 212960, 229344, 245728,
+  };
+  const int tmp = (xsq_q10 >> 2) + 8;
+  const int k = get_msb(tmp) - 3;
+  const int xq = (k << 3) + ((tmp >> k) & 0x7);
+  const int one_q10 = 1 << 10;
+  const int a_q10 = ((xsq_q10 - xsq_iq_q10[xq]) << 10) >> (2 + k);
+  const int b_q10 = one_q10 - a_q10;
+  *r_q10 = (rate_tab_q10[xq] * b_q10 + rate_tab_q10[xq + 1] * a_q10) >> 10;
+  *d_q10 = (dist_tab_q10[xq] * b_q10 + dist_tab_q10[xq + 1] * a_q10) >> 10;
+}
+
+void av1_model_rd_from_var_lapndz(int64_t var, unsigned int n_log2,
+                                  unsigned int qstep, int *rate,
+                                  int64_t *dist) {
+  // This function models the rate and distortion for a Laplacian
+  // source with given variance when quantized with a uniform quantizer
+  // with given stepsize. The closed form expressions are in:
+  // Hang and Chen, "Source Model for transform video coder and its
+  // application - Part I: Fundamental Theory", IEEE Trans. Circ.
+  // Sys. for Video Tech., April 1997.
+  if (var == 0) {
+    *rate = 0;
+    *dist = 0;
+  } else {
+    int d_q10, r_q10;
+    static const uint32_t MAX_XSQ_Q10 = 245727;
+    const uint64_t xsq_q10_64 =
+        (((uint64_t)qstep * qstep << (n_log2 + 10)) + (var >> 1)) / var;
+    const int xsq_q10 = (int)AOMMIN(xsq_q10_64, MAX_XSQ_Q10);
+    model_rd_norm(xsq_q10, &r_q10, &d_q10);
+    *rate = ROUND_POWER_OF_TWO(r_q10 << n_log2, 10 - AV1_PROB_COST_SHIFT);
+    *dist = (var * (int64_t)d_q10 + 512) >> 10;
+  }
+}
+
+static void get_entropy_contexts_plane(
+    BLOCK_SIZE plane_bsize, TX_SIZE tx_size, const struct macroblockd_plane *pd,
+    ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE],
+    ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE]) {
+  const int num_4x4_w = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+  const int num_4x4_h = block_size_high[plane_bsize] >> tx_size_high_log2[0];
+  const ENTROPY_CONTEXT *const above = pd->above_context;
+  const ENTROPY_CONTEXT *const left = pd->left_context;
+
+  int i;
+
+#if CONFIG_CB4X4
+  switch (tx_size) {
+    case TX_2X2:
+      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
+      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
+      break;
+    case TX_4X4:
+      for (i = 0; i < num_4x4_w; i += 2)
+        t_above[i] = !!*(const uint16_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 2)
+        t_left[i] = !!*(const uint16_t *)&left[i];
+      break;
+    case TX_8X8:
+      for (i = 0; i < num_4x4_w; i += 4)
+        t_above[i] = !!*(const uint32_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 4)
+        t_left[i] = !!*(const uint32_t *)&left[i];
+      break;
+    case TX_16X16:
+      for (i = 0; i < num_4x4_w; i += 8)
+        t_above[i] = !!*(const uint64_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 8)
+        t_left[i] = !!*(const uint64_t *)&left[i];
+      break;
+    case TX_32X32:
+      for (i = 0; i < num_4x4_w; i += 16)
+        t_above[i] =
+            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8]);
+      for (i = 0; i < num_4x4_h; i += 16)
+        t_left[i] =
+            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8]);
+      break;
+    case TX_4X8:
+      for (i = 0; i < num_4x4_w; i += 2)
+        t_above[i] = !!*(const uint16_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 4)
+        t_left[i] = !!*(const uint32_t *)&left[i];
+      break;
+    case TX_8X4:
+      for (i = 0; i < num_4x4_w; i += 4)
+        t_above[i] = !!*(const uint32_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 2)
+        t_left[i] = !!*(const uint16_t *)&left[i];
+      break;
+    case TX_8X16:
+      for (i = 0; i < num_4x4_w; i += 4)
+        t_above[i] = !!*(const uint32_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 8)
+        t_left[i] = !!*(const uint64_t *)&left[i];
+      break;
+    case TX_16X8:
+      for (i = 0; i < num_4x4_w; i += 8)
+        t_above[i] = !!*(const uint64_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 4)
+        t_left[i] = !!*(const uint32_t *)&left[i];
+      break;
+    case TX_16X32:
+      for (i = 0; i < num_4x4_w; i += 8)
+        t_above[i] = !!*(const uint64_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 16)
+        t_left[i] =
+            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8]);
+      break;
+    case TX_32X16:
+      for (i = 0; i < num_4x4_w; i += 16)
+        t_above[i] =
+            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8]);
+      for (i = 0; i < num_4x4_h; i += 8)
+        t_left[i] = !!*(const uint64_t *)&left[i];
+      break;
+
+    default: assert(0 && "Invalid transform size."); break;
+  }
+  return;
+#endif
+
+  switch (tx_size) {
+    case TX_4X4:
+      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
+      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
+      break;
+    case TX_8X8:
+      for (i = 0; i < num_4x4_w; i += 2)
+        t_above[i] = !!*(const uint16_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 2)
+        t_left[i] = !!*(const uint16_t *)&left[i];
+      break;
+    case TX_16X16:
+      for (i = 0; i < num_4x4_w; i += 4)
+        t_above[i] = !!*(const uint32_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 4)
+        t_left[i] = !!*(const uint32_t *)&left[i];
+      break;
+    case TX_32X32:
+      for (i = 0; i < num_4x4_w; i += 8)
+        t_above[i] = !!*(const uint64_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 8)
+        t_left[i] = !!*(const uint64_t *)&left[i];
+      break;
+#if CONFIG_TX64X64
+    case TX_64X64:
+      for (i = 0; i < num_4x4_w; i += 16)
+        t_above[i] =
+            !!(*(const uint64_t *)&above[i] | *(const uint64_t *)&above[i + 8]);
+      for (i = 0; i < num_4x4_h; i += 16)
+        t_left[i] =
+            !!(*(const uint64_t *)&left[i] | *(const uint64_t *)&left[i + 8]);
+      break;
+#endif  // CONFIG_TX64X64
+    case TX_4X8:
+      memcpy(t_above, above, sizeof(ENTROPY_CONTEXT) * num_4x4_w);
+      for (i = 0; i < num_4x4_h; i += 2)
+        t_left[i] = !!*(const uint16_t *)&left[i];
+      break;
+    case TX_8X4:
+      for (i = 0; i < num_4x4_w; i += 2)
+        t_above[i] = !!*(const uint16_t *)&above[i];
+      memcpy(t_left, left, sizeof(ENTROPY_CONTEXT) * num_4x4_h);
+      break;
+    case TX_8X16:
+      for (i = 0; i < num_4x4_w; i += 2)
+        t_above[i] = !!*(const uint16_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 4)
+        t_left[i] = !!*(const uint32_t *)&left[i];
+      break;
+    case TX_16X8:
+      for (i = 0; i < num_4x4_w; i += 4)
+        t_above[i] = !!*(const uint32_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 2)
+        t_left[i] = !!*(const uint16_t *)&left[i];
+      break;
+    case TX_16X32:
+      for (i = 0; i < num_4x4_w; i += 4)
+        t_above[i] = !!*(const uint32_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 8)
+        t_left[i] = !!*(const uint64_t *)&left[i];
+      break;
+    case TX_32X16:
+      for (i = 0; i < num_4x4_w; i += 8)
+        t_above[i] = !!*(const uint64_t *)&above[i];
+      for (i = 0; i < num_4x4_h; i += 4)
+        t_left[i] = !!*(const uint32_t *)&left[i];
+      break;
+    default: assert(0 && "Invalid transform size."); break;
+  }
+}
+
+void av1_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              const struct macroblockd_plane *pd,
+                              ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE],
+                              ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE]) {
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+  get_entropy_contexts_plane(plane_bsize, tx_size, pd, t_above, t_left);
+}
+
+void av1_mv_pred(const AV1_COMP *cpi, MACROBLOCK *x, uint8_t *ref_y_buffer,
+                 int ref_y_stride, int ref_frame, BLOCK_SIZE block_size) {
+  int i;
+  int zero_seen = 0;
+  int best_index = 0;
+  int best_sad = INT_MAX;
+  int this_sad = INT_MAX;
+  int max_mv = 0;
+  int near_same_nearest;
+  uint8_t *src_y_ptr = x->plane[0].src.buf;
+  uint8_t *ref_y_ptr;
+  const int num_mv_refs =
+      MAX_MV_REF_CANDIDATES +
+      (cpi->sf.adaptive_motion_search && block_size < x->max_partition_size);
+
+  MV pred_mv[3];
+  pred_mv[0] = x->mbmi_ext->ref_mvs[ref_frame][0].as_mv;
+  pred_mv[1] = x->mbmi_ext->ref_mvs[ref_frame][1].as_mv;
+  pred_mv[2] = x->pred_mv[ref_frame];
+  assert(num_mv_refs <= (int)(sizeof(pred_mv) / sizeof(pred_mv[0])));
+
+  near_same_nearest = x->mbmi_ext->ref_mvs[ref_frame][0].as_int ==
+                      x->mbmi_ext->ref_mvs[ref_frame][1].as_int;
+  // Get the sad for each candidate reference mv.
+  for (i = 0; i < num_mv_refs; ++i) {
+    const MV *this_mv = &pred_mv[i];
+    int fp_row, fp_col;
+
+    if (i == 1 && near_same_nearest) continue;
+    fp_row = (this_mv->row + 3 + (this_mv->row >= 0)) >> 3;
+    fp_col = (this_mv->col + 3 + (this_mv->col >= 0)) >> 3;
+    max_mv = AOMMAX(max_mv, AOMMAX(abs(this_mv->row), abs(this_mv->col)) >> 3);
+
+    if (fp_row == 0 && fp_col == 0 && zero_seen) continue;
+    zero_seen |= (fp_row == 0 && fp_col == 0);
+
+    ref_y_ptr = &ref_y_buffer[ref_y_stride * fp_row + fp_col];
+    // Find sad for current vector.
+    this_sad = cpi->fn_ptr[block_size].sdf(src_y_ptr, x->plane[0].src.stride,
+                                           ref_y_ptr, ref_y_stride);
+    // Note if it is the best so far.
+    if (this_sad < best_sad) {
+      best_sad = this_sad;
+      best_index = i;
+    }
+  }
+
+  // Note the index of the mv that worked best in the reference list.
+  x->mv_best_ref_index[ref_frame] = best_index;
+  x->max_mv_context[ref_frame] = max_mv;
+  x->pred_mv_sad[ref_frame] = best_sad;
+}
+
+void av1_setup_pred_block(const MACROBLOCKD *xd,
+                          struct buf_2d dst[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
+                          const struct scale_factors *scale,
+                          const struct scale_factors *scale_uv) {
+  int i;
+
+  dst[0].buf = src->y_buffer;
+  dst[0].stride = src->y_stride;
+  dst[1].buf = src->u_buffer;
+  dst[2].buf = src->v_buffer;
+  dst[1].stride = dst[2].stride = src->uv_stride;
+
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    setup_pred_plane(dst + i, xd->mi[0]->mbmi.sb_type, dst[i].buf,
+                     i ? src->uv_crop_width : src->y_crop_width,
+                     i ? src->uv_crop_height : src->y_crop_height,
+                     dst[i].stride, mi_row, mi_col, i ? scale_uv : scale,
+                     xd->plane[i].subsampling_x, xd->plane[i].subsampling_y);
+  }
+}
+
+int av1_raster_block_offset(BLOCK_SIZE plane_bsize, int raster_block,
+                            int stride) {
+  const int bw = b_width_log2_lookup[plane_bsize];
+  const int y = 4 * (raster_block >> bw);
+  const int x = 4 * (raster_block & ((1 << bw) - 1));
+  return y * stride + x;
+}
+
+int16_t *av1_raster_block_offset_int16(BLOCK_SIZE plane_bsize, int raster_block,
+                                       int16_t *base) {
+  const int stride = block_size_wide[plane_bsize];
+  return base + av1_raster_block_offset(plane_bsize, raster_block, stride);
+}
+
+YV12_BUFFER_CONFIG *av1_get_scaled_ref_frame(const AV1_COMP *cpi,
+                                             int ref_frame) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1];
+  const int ref_idx = get_ref_frame_buf_idx(cpi, ref_frame);
+  return (scaled_idx != ref_idx && scaled_idx != INVALID_IDX)
+             ? &cm->buffer_pool->frame_bufs[scaled_idx].buf
+             : NULL;
+}
+
+#if CONFIG_DUAL_FILTER
+int av1_get_switchable_rate(const AV1_COMP *cpi, const MACROBLOCKD *xd) {
+  const AV1_COMMON *const cm = &cpi->common;
+  if (cm->interp_filter == SWITCHABLE) {
+    const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+    int inter_filter_cost = 0;
+    int dir;
+
+    for (dir = 0; dir < 2; ++dir) {
+      if (has_subpel_mv_component(xd->mi[0], xd, dir) ||
+          (mbmi->ref_frame[1] > INTRA_FRAME &&
+           has_subpel_mv_component(xd->mi[0], xd, dir + 2))) {
+        const int ctx = av1_get_pred_context_switchable_interp(xd, dir);
+        inter_filter_cost +=
+            cpi->switchable_interp_costs[ctx][mbmi->interp_filter[dir]];
+      }
+    }
+    return SWITCHABLE_INTERP_RATE_FACTOR * inter_filter_cost;
+  } else {
+    return 0;
+  }
+}
+#else
+int av1_get_switchable_rate(const AV1_COMP *cpi, const MACROBLOCKD *xd) {
+  const AV1_COMMON *const cm = &cpi->common;
+  if (cm->interp_filter == SWITCHABLE) {
+    const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+    const int ctx = av1_get_pred_context_switchable_interp(xd);
+    return SWITCHABLE_INTERP_RATE_FACTOR *
+           cpi->switchable_interp_costs[ctx][mbmi->interp_filter];
+  }
+  return 0;
+}
+#endif
+
+void av1_set_rd_speed_thresholds(AV1_COMP *cpi) {
+  int i;
+  RD_OPT *const rd = &cpi->rd;
+  SPEED_FEATURES *const sf = &cpi->sf;
+
+  // Set baseline threshold values.
+  for (i = 0; i < MAX_MODES; ++i) rd->thresh_mult[i] = cpi->oxcf.mode == 0;
+
+  if (sf->adaptive_rd_thresh) {
+    rd->thresh_mult[THR_NEARESTMV] = 300;
+#if CONFIG_EXT_REFS
+    rd->thresh_mult[THR_NEARESTL2] = 300;
+    rd->thresh_mult[THR_NEARESTL3] = 300;
+    rd->thresh_mult[THR_NEARESTB] = 300;
+#endif  // CONFIG_EXT_REFS
+    rd->thresh_mult[THR_NEARESTA] = 300;
+    rd->thresh_mult[THR_NEARESTG] = 300;
+  } else {
+    rd->thresh_mult[THR_NEARESTMV] = 0;
+#if CONFIG_EXT_REFS
+    rd->thresh_mult[THR_NEARESTL2] = 0;
+    rd->thresh_mult[THR_NEARESTL3] = 0;
+    rd->thresh_mult[THR_NEARESTB] = 0;
+#endif  // CONFIG_EXT_REFS
+    rd->thresh_mult[THR_NEARESTA] = 0;
+    rd->thresh_mult[THR_NEARESTG] = 0;
+  }
+
+  rd->thresh_mult[THR_DC] += 1000;
+
+  rd->thresh_mult[THR_NEWMV] += 1000;
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_NEWL2] += 1000;
+  rd->thresh_mult[THR_NEWL3] += 1000;
+  rd->thresh_mult[THR_NEWB] += 1000;
+#endif  // CONFIG_EXT_REFS
+  rd->thresh_mult[THR_NEWA] += 1000;
+  rd->thresh_mult[THR_NEWG] += 1000;
+
+  rd->thresh_mult[THR_NEARMV] += 1000;
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_NEARL2] += 1000;
+  rd->thresh_mult[THR_NEARL3] += 1000;
+  rd->thresh_mult[THR_NEARB] += 1000;
+#endif  // CONFIG_EXT_REFS
+  rd->thresh_mult[THR_NEARA] += 1000;
+  rd->thresh_mult[THR_NEARG] += 1000;
+
+  rd->thresh_mult[THR_ZEROMV] += 2000;
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_ZEROL2] += 2000;
+  rd->thresh_mult[THR_ZEROL3] += 2000;
+  rd->thresh_mult[THR_ZEROB] += 2000;
+#endif  // CONFIG_EXT_REFS
+  rd->thresh_mult[THR_ZEROG] += 2000;
+  rd->thresh_mult[THR_ZEROA] += 2000;
+
+  rd->thresh_mult[THR_TM] += 1000;
+
+#if CONFIG_EXT_INTER
+
+  rd->thresh_mult[THR_COMP_NEAREST_NEARESTLA] += 1000;
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEAREST_NEARESTL2A] += 1000;
+  rd->thresh_mult[THR_COMP_NEAREST_NEARESTL3A] += 1000;
+#endif  // CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEAREST_NEARESTGA] += 1000;
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEAREST_NEARESTLB] += 1000;
+  rd->thresh_mult[THR_COMP_NEAREST_NEARESTL2B] += 1000;
+  rd->thresh_mult[THR_COMP_NEAREST_NEARESTL3B] += 1000;
+  rd->thresh_mult[THR_COMP_NEAREST_NEARESTGB] += 1000;
+#endif  // CONFIG_EXT_REFS
+
+#else  // CONFIG_EXT_INTER
+
+  rd->thresh_mult[THR_COMP_NEARESTLA] += 1000;
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEARESTL2A] += 1000;
+  rd->thresh_mult[THR_COMP_NEARESTL3A] += 1000;
+#endif  // CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEARESTGA] += 1000;
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEARESTLB] += 1000;
+  rd->thresh_mult[THR_COMP_NEARESTL2B] += 1000;
+  rd->thresh_mult[THR_COMP_NEARESTL3B] += 1000;
+  rd->thresh_mult[THR_COMP_NEARESTGB] += 1000;
+#endif  // CONFIG_EXT_REFS
+
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_EXT_INTER
+
+  rd->thresh_mult[THR_COMP_NEAREST_NEARLA] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARESTLA] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARLA] += 1200;
+  rd->thresh_mult[THR_COMP_NEAREST_NEWLA] += 1500;
+  rd->thresh_mult[THR_COMP_NEW_NEARESTLA] += 1500;
+  rd->thresh_mult[THR_COMP_NEAR_NEWLA] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEARLA] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEWLA] += 2000;
+  rd->thresh_mult[THR_COMP_ZERO_ZEROLA] += 2500;
+
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEAREST_NEARL2A] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARESTL2A] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARL2A] += 1200;
+  rd->thresh_mult[THR_COMP_NEAREST_NEWL2A] += 1500;
+  rd->thresh_mult[THR_COMP_NEW_NEARESTL2A] += 1500;
+  rd->thresh_mult[THR_COMP_NEAR_NEWL2A] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEARL2A] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEWL2A] += 2000;
+  rd->thresh_mult[THR_COMP_ZERO_ZEROL2A] += 2500;
+
+  rd->thresh_mult[THR_COMP_NEAREST_NEARL3A] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARESTL3A] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARL3A] += 1200;
+  rd->thresh_mult[THR_COMP_NEAREST_NEWL3A] += 1500;
+  rd->thresh_mult[THR_COMP_NEW_NEARESTL3A] += 1500;
+  rd->thresh_mult[THR_COMP_NEAR_NEWL3A] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEARL3A] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEWL3A] += 2000;
+  rd->thresh_mult[THR_COMP_ZERO_ZEROL3A] += 2500;
+#endif  // CONFIG_EXT_REFS
+
+  rd->thresh_mult[THR_COMP_NEAREST_NEARGA] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARESTGA] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARGA] += 1200;
+  rd->thresh_mult[THR_COMP_NEAREST_NEWGA] += 1500;
+  rd->thresh_mult[THR_COMP_NEW_NEARESTGA] += 1500;
+  rd->thresh_mult[THR_COMP_NEAR_NEWGA] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEARGA] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEWGA] += 2000;
+  rd->thresh_mult[THR_COMP_ZERO_ZEROGA] += 2500;
+
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEAREST_NEARLB] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARESTLB] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARLB] += 1200;
+  rd->thresh_mult[THR_COMP_NEAREST_NEWLB] += 1500;
+  rd->thresh_mult[THR_COMP_NEW_NEARESTLB] += 1500;
+  rd->thresh_mult[THR_COMP_NEAR_NEWLB] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEARLB] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEWLB] += 2000;
+  rd->thresh_mult[THR_COMP_ZERO_ZEROLB] += 2500;
+
+  rd->thresh_mult[THR_COMP_NEAREST_NEARL2B] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARESTL2B] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARL2B] += 1200;
+  rd->thresh_mult[THR_COMP_NEAREST_NEWL2B] += 1500;
+  rd->thresh_mult[THR_COMP_NEW_NEARESTL2B] += 1500;
+  rd->thresh_mult[THR_COMP_NEAR_NEWL2B] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEARL2B] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEWL2B] += 2000;
+  rd->thresh_mult[THR_COMP_ZERO_ZEROL2B] += 2500;
+
+  rd->thresh_mult[THR_COMP_NEAREST_NEARL3B] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARESTL3B] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARL3B] += 1200;
+  rd->thresh_mult[THR_COMP_NEAREST_NEWL3B] += 1500;
+  rd->thresh_mult[THR_COMP_NEW_NEARESTL3B] += 1500;
+  rd->thresh_mult[THR_COMP_NEAR_NEWL3B] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEARL3B] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEWL3B] += 2000;
+  rd->thresh_mult[THR_COMP_ZERO_ZEROL3B] += 2500;
+
+  rd->thresh_mult[THR_COMP_NEAREST_NEARGB] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARESTGB] += 1200;
+  rd->thresh_mult[THR_COMP_NEAR_NEARGB] += 1200;
+  rd->thresh_mult[THR_COMP_NEAREST_NEWGB] += 1500;
+  rd->thresh_mult[THR_COMP_NEW_NEARESTGB] += 1500;
+  rd->thresh_mult[THR_COMP_NEAR_NEWGB] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEARGB] += 1700;
+  rd->thresh_mult[THR_COMP_NEW_NEWGB] += 2000;
+  rd->thresh_mult[THR_COMP_ZERO_ZEROGB] += 2500;
+#endif  // CONFIG_EXT_REFS
+
+#else  // CONFIG_EXT_INTER
+
+  rd->thresh_mult[THR_COMP_NEARLA] += 1500;
+  rd->thresh_mult[THR_COMP_NEWLA] += 2000;
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEARL2A] += 1500;
+  rd->thresh_mult[THR_COMP_NEWL2A] += 2000;
+  rd->thresh_mult[THR_COMP_NEARL3A] += 1500;
+  rd->thresh_mult[THR_COMP_NEWL3A] += 2000;
+#endif  // CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEARGA] += 1500;
+  rd->thresh_mult[THR_COMP_NEWGA] += 2000;
+
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_NEARLB] += 1500;
+  rd->thresh_mult[THR_COMP_NEWLB] += 2000;
+  rd->thresh_mult[THR_COMP_NEARL2B] += 1500;
+  rd->thresh_mult[THR_COMP_NEWL2B] += 2000;
+  rd->thresh_mult[THR_COMP_NEARL3B] += 1500;
+  rd->thresh_mult[THR_COMP_NEWL3B] += 2000;
+  rd->thresh_mult[THR_COMP_NEARGB] += 1500;
+  rd->thresh_mult[THR_COMP_NEWGB] += 2000;
+#endif  // CONFIG_EXT_REFS
+
+  rd->thresh_mult[THR_COMP_ZEROLA] += 2500;
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_ZEROL2A] += 2500;
+  rd->thresh_mult[THR_COMP_ZEROL3A] += 2500;
+#endif  // CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_ZEROGA] += 2500;
+
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_ZEROLB] += 2500;
+  rd->thresh_mult[THR_COMP_ZEROL2B] += 2500;
+  rd->thresh_mult[THR_COMP_ZEROL3B] += 2500;
+  rd->thresh_mult[THR_COMP_ZEROGB] += 2500;
+#endif  // CONFIG_EXT_REFS
+
+#endif  // CONFIG_EXT_INTER
+
+  rd->thresh_mult[THR_H_PRED] += 2000;
+  rd->thresh_mult[THR_V_PRED] += 2000;
+  rd->thresh_mult[THR_D135_PRED] += 2500;
+  rd->thresh_mult[THR_D207_PRED] += 2500;
+  rd->thresh_mult[THR_D153_PRED] += 2500;
+  rd->thresh_mult[THR_D63_PRED] += 2500;
+  rd->thresh_mult[THR_D117_PRED] += 2500;
+  rd->thresh_mult[THR_D45_PRED] += 2500;
+
+#if CONFIG_EXT_INTER
+  rd->thresh_mult[THR_COMP_INTERINTRA_ZEROL] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARESTL] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARL] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEWL] += 2000;
+
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_INTERINTRA_ZEROL2] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARESTL2] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARL2] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEWL2] += 2000;
+
+  rd->thresh_mult[THR_COMP_INTERINTRA_ZEROL3] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARESTL3] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARL3] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEWL3] += 2000;
+#endif  // CONFIG_EXT_REFS
+
+  rd->thresh_mult[THR_COMP_INTERINTRA_ZEROG] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARESTG] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARG] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEWG] += 2000;
+
+#if CONFIG_EXT_REFS
+  rd->thresh_mult[THR_COMP_INTERINTRA_ZEROB] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARESTB] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARB] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEWB] += 2000;
+#endif  // CONFIG_EXT_REFS
+
+  rd->thresh_mult[THR_COMP_INTERINTRA_ZEROA] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARESTA] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEARA] += 1500;
+  rd->thresh_mult[THR_COMP_INTERINTRA_NEWA] += 2000;
+#endif  // CONFIG_EXT_INTER
+}
+
+void av1_set_rd_speed_thresholds_sub8x8(AV1_COMP *cpi) {
+  static const int thresh_mult[MAX_REFS] = {
+#if CONFIG_EXT_REFS
+    2500,
+    2500,
+    2500,
+    2500,
+    2500,
+    2500,
+    4500,
+    4500,
+    4500,
+    4500,
+    4500,
+    4500,
+    4500,
+    4500,
+    2500
+#else
+    2500,
+    2500,
+    2500,
+    4500,
+    4500,
+    2500
+#endif  // CONFIG_EXT_REFS
+  };
+  RD_OPT *const rd = &cpi->rd;
+  memcpy(rd->thresh_mult_sub8x8, thresh_mult, sizeof(thresh_mult));
+}
+
+void av1_update_rd_thresh_fact(const AV1_COMMON *const cm,
+                               int (*factor_buf)[MAX_MODES], int rd_thresh,
+                               int bsize, int best_mode_index) {
+  if (rd_thresh > 0) {
+#if CONFIG_CB4X4
+    const int top_mode = MAX_MODES;
+#else
+    const int top_mode = bsize < BLOCK_8X8 ? MAX_REFS : MAX_MODES;
+#endif
+    int mode;
+    for (mode = 0; mode < top_mode; ++mode) {
+      const BLOCK_SIZE min_size = AOMMAX(bsize - 1, BLOCK_4X4);
+      const BLOCK_SIZE max_size = AOMMIN(bsize + 2, (int)cm->sb_size);
+      BLOCK_SIZE bs;
+      for (bs = min_size; bs <= max_size; ++bs) {
+        int *const fact = &factor_buf[bs][mode];
+        if (mode == best_mode_index) {
+          *fact -= (*fact >> 4);
+        } else {
+          *fact = AOMMIN(*fact + RD_THRESH_INC, rd_thresh * RD_THRESH_MAX_FACT);
+        }
+      }
+    }
+  }
+}
+
+int av1_get_intra_cost_penalty(int qindex, int qdelta,
+                               aom_bit_depth_t bit_depth) {
+  const int q = av1_dc_quant(qindex, qdelta, bit_depth);
+#if CONFIG_HIGHBITDEPTH
+  switch (bit_depth) {
+    case AOM_BITS_8: return 20 * q;
+    case AOM_BITS_10: return 5 * q;
+    case AOM_BITS_12: return ROUND_POWER_OF_TWO(5 * q, 2);
+    default:
+      assert(0 && "bit_depth should be AOM_BITS_8, AOM_BITS_10 or AOM_BITS_12");
+      return -1;
+  }
+#else
+  return 20 * q;
+#endif  // CONFIG_HIGHBITDEPTH
+}
diff --git a/third_party/aom/av1/encoder/rd.h b/third_party/aom/av1/encoder/rd.h
new file mode 100644
index 0000000000..c0ac1f7e75
--- /dev/null
+++ b/third_party/aom/av1/encoder/rd.h
@@ -0,0 +1,505 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_RD_H_
+#define AV1_ENCODER_RD_H_
+
+#include <limits.h>
+
+#if CONFIG_ANS
+#include "aom_dsp/ans.h"
+#endif  // CONFIG_ANS
+#include "av1/common/blockd.h"
+
+#include "av1/encoder/block.h"
+#include "av1/encoder/context_tree.h"
+#include "av1/encoder/cost.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define RDDIV_BITS 7
+#define RD_EPB_SHIFT 6
+
+#define RDCOST(RM, DM, R, D) \
+  (ROUND_POWER_OF_TWO(((int64_t)R) * (RM), AV1_PROB_COST_SHIFT) + (D << DM))
+
+#define RDCOST_DBL(RM, DM, R, D)                                   \
+  (((((double)(R)) * (RM)) / (double)(1 << AV1_PROB_COST_SHIFT)) + \
+   ((double)(D) * (1 << (DM))))
+
+#define QIDX_SKIP_THRESH 115
+
+#define MV_COST_WEIGHT 108
+#define MV_COST_WEIGHT_SUB 120
+
+#define INVALID_MV 0x80008000
+
+#if CONFIG_EXT_REFS
+#define MAX_REFS 15
+#else
+#define MAX_REFS 6
+#endif  // CONFIG_EXT_REFS
+
+#define RD_THRESH_MAX_FACT 64
+#define RD_THRESH_INC 1
+
+// This enumerator type needs to be kept aligned with the mode order in
+// const MODE_DEFINITION av1_mode_order[MAX_MODES] used in the rd code.
+typedef enum {
+  THR_NEARESTMV,
+#if CONFIG_EXT_REFS
+  THR_NEARESTL2,
+  THR_NEARESTL3,
+  THR_NEARESTB,
+#endif  // CONFIG_EXT_REFS
+  THR_NEARESTA,
+  THR_NEARESTG,
+
+  THR_DC,
+
+  THR_NEWMV,
+#if CONFIG_EXT_REFS
+  THR_NEWL2,
+  THR_NEWL3,
+  THR_NEWB,
+#endif  // CONFIG_EXT_REFS
+  THR_NEWA,
+  THR_NEWG,
+
+  THR_NEARMV,
+#if CONFIG_EXT_REFS
+  THR_NEARL2,
+  THR_NEARL3,
+  THR_NEARB,
+#endif  // CONFIG_EXT_REFS
+  THR_NEARA,
+  THR_NEARG,
+
+  THR_ZEROMV,
+#if CONFIG_EXT_REFS
+  THR_ZEROL2,
+  THR_ZEROL3,
+  THR_ZEROB,
+#endif  // CONFIG_EXT_REFS
+  THR_ZEROG,
+  THR_ZEROA,
+
+#if CONFIG_EXT_INTER
+
+  THR_COMP_NEAREST_NEARESTLA,
+#if CONFIG_EXT_REFS
+  THR_COMP_NEAREST_NEARESTL2A,
+  THR_COMP_NEAREST_NEARESTL3A,
+#endif  // CONFIG_EXT_REFS
+  THR_COMP_NEAREST_NEARESTGA,
+#if CONFIG_EXT_REFS
+  THR_COMP_NEAREST_NEARESTLB,
+  THR_COMP_NEAREST_NEARESTL2B,
+  THR_COMP_NEAREST_NEARESTL3B,
+  THR_COMP_NEAREST_NEARESTGB,
+#endif  // CONFIG_EXT_REFS
+
+#else  // CONFIG_EXT_INTER
+
+  THR_COMP_NEARESTLA,
+#if CONFIG_EXT_REFS
+  THR_COMP_NEARESTL2A,
+  THR_COMP_NEARESTL3A,
+#endif  // CONFIG_EXT_REFS
+  THR_COMP_NEARESTGA,
+#if CONFIG_EXT_REFS
+  THR_COMP_NEARESTLB,
+  THR_COMP_NEARESTL2B,
+  THR_COMP_NEARESTL3B,
+  THR_COMP_NEARESTGB,
+#endif  // CONFIG_EXT_REFS
+
+#endif  // CONFIG_EXT_INTER
+
+  THR_TM,
+
+#if CONFIG_ALT_INTRA
+  THR_SMOOTH,
+#endif  // CONFIG_ALT_INTRA
+
+#if CONFIG_EXT_INTER
+
+  THR_COMP_NEAR_NEARESTLA,
+  THR_COMP_NEAREST_NEARLA,
+  THR_COMP_NEAR_NEARLA,
+  THR_COMP_NEW_NEARESTLA,
+  THR_COMP_NEAREST_NEWLA,
+  THR_COMP_NEW_NEARLA,
+  THR_COMP_NEAR_NEWLA,
+  THR_COMP_NEW_NEWLA,
+  THR_COMP_ZERO_ZEROLA,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_NEAR_NEARESTL2A,
+  THR_COMP_NEAREST_NEARL2A,
+  THR_COMP_NEAR_NEARL2A,
+  THR_COMP_NEW_NEARESTL2A,
+  THR_COMP_NEAREST_NEWL2A,
+  THR_COMP_NEW_NEARL2A,
+  THR_COMP_NEAR_NEWL2A,
+  THR_COMP_NEW_NEWL2A,
+  THR_COMP_ZERO_ZEROL2A,
+
+  THR_COMP_NEAR_NEARESTL3A,
+  THR_COMP_NEAREST_NEARL3A,
+  THR_COMP_NEAR_NEARL3A,
+  THR_COMP_NEW_NEARESTL3A,
+  THR_COMP_NEAREST_NEWL3A,
+  THR_COMP_NEW_NEARL3A,
+  THR_COMP_NEAR_NEWL3A,
+  THR_COMP_NEW_NEWL3A,
+  THR_COMP_ZERO_ZEROL3A,
+#endif  // CONFIG_EXT_REFS
+
+  THR_COMP_NEAR_NEARESTGA,
+  THR_COMP_NEAREST_NEARGA,
+  THR_COMP_NEAR_NEARGA,
+  THR_COMP_NEW_NEARESTGA,
+  THR_COMP_NEAREST_NEWGA,
+  THR_COMP_NEW_NEARGA,
+  THR_COMP_NEAR_NEWGA,
+  THR_COMP_NEW_NEWGA,
+  THR_COMP_ZERO_ZEROGA,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_NEAR_NEARESTLB,
+  THR_COMP_NEAREST_NEARLB,
+  THR_COMP_NEAR_NEARLB,
+  THR_COMP_NEW_NEARESTLB,
+  THR_COMP_NEAREST_NEWLB,
+  THR_COMP_NEW_NEARLB,
+  THR_COMP_NEAR_NEWLB,
+  THR_COMP_NEW_NEWLB,
+  THR_COMP_ZERO_ZEROLB,
+
+  THR_COMP_NEAR_NEARESTL2B,
+  THR_COMP_NEAREST_NEARL2B,
+  THR_COMP_NEAR_NEARL2B,
+  THR_COMP_NEW_NEARESTL2B,
+  THR_COMP_NEAREST_NEWL2B,
+  THR_COMP_NEW_NEARL2B,
+  THR_COMP_NEAR_NEWL2B,
+  THR_COMP_NEW_NEWL2B,
+  THR_COMP_ZERO_ZEROL2B,
+
+  THR_COMP_NEAR_NEARESTL3B,
+  THR_COMP_NEAREST_NEARL3B,
+  THR_COMP_NEAR_NEARL3B,
+  THR_COMP_NEW_NEARESTL3B,
+  THR_COMP_NEAREST_NEWL3B,
+  THR_COMP_NEW_NEARL3B,
+  THR_COMP_NEAR_NEWL3B,
+  THR_COMP_NEW_NEWL3B,
+  THR_COMP_ZERO_ZEROL3B,
+
+  THR_COMP_NEAR_NEARESTGB,
+  THR_COMP_NEAREST_NEARGB,
+  THR_COMP_NEAR_NEARGB,
+  THR_COMP_NEW_NEARESTGB,
+  THR_COMP_NEAREST_NEWGB,
+  THR_COMP_NEW_NEARGB,
+  THR_COMP_NEAR_NEWGB,
+  THR_COMP_NEW_NEWGB,
+  THR_COMP_ZERO_ZEROGB,
+#endif  // CONFIG_EXT_REFS
+
+#else  // CONFIG_EXT_INTER
+
+  THR_COMP_NEARLA,
+  THR_COMP_NEWLA,
+#if CONFIG_EXT_REFS
+  THR_COMP_NEARL2A,
+  THR_COMP_NEWL2A,
+  THR_COMP_NEARL3A,
+  THR_COMP_NEWL3A,
+#endif  // CONFIG_EXT_REFS
+  THR_COMP_NEARGA,
+  THR_COMP_NEWGA,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_NEARLB,
+  THR_COMP_NEWLB,
+  THR_COMP_NEARL2B,
+  THR_COMP_NEWL2B,
+  THR_COMP_NEARL3B,
+  THR_COMP_NEWL3B,
+  THR_COMP_NEARGB,
+  THR_COMP_NEWGB,
+#endif  // CONFIG_EXT_REFS
+
+  THR_COMP_ZEROLA,
+#if CONFIG_EXT_REFS
+  THR_COMP_ZEROL2A,
+  THR_COMP_ZEROL3A,
+#endif  // CONFIG_EXT_REFS
+  THR_COMP_ZEROGA,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_ZEROLB,
+  THR_COMP_ZEROL2B,
+  THR_COMP_ZEROL3B,
+  THR_COMP_ZEROGB,
+#endif  // CONFIG_EXT_REFS
+
+#endif  // CONFIG_EXT_INTER
+
+  THR_H_PRED,
+  THR_V_PRED,
+  THR_D135_PRED,
+  THR_D207_PRED,
+  THR_D153_PRED,
+  THR_D63_PRED,
+  THR_D117_PRED,
+  THR_D45_PRED,
+
+#if CONFIG_EXT_INTER
+  THR_COMP_INTERINTRA_ZEROL,
+  THR_COMP_INTERINTRA_NEARESTL,
+  THR_COMP_INTERINTRA_NEARL,
+  THR_COMP_INTERINTRA_NEWL,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_INTERINTRA_ZEROL2,
+  THR_COMP_INTERINTRA_NEARESTL2,
+  THR_COMP_INTERINTRA_NEARL2,
+  THR_COMP_INTERINTRA_NEWL2,
+
+  THR_COMP_INTERINTRA_ZEROL3,
+  THR_COMP_INTERINTRA_NEARESTL3,
+  THR_COMP_INTERINTRA_NEARL3,
+  THR_COMP_INTERINTRA_NEWL3,
+#endif  // CONFIG_EXT_REFS
+
+  THR_COMP_INTERINTRA_ZEROG,
+  THR_COMP_INTERINTRA_NEARESTG,
+  THR_COMP_INTERINTRA_NEARG,
+  THR_COMP_INTERINTRA_NEWG,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_INTERINTRA_ZEROB,
+  THR_COMP_INTERINTRA_NEARESTB,
+  THR_COMP_INTERINTRA_NEARB,
+  THR_COMP_INTERINTRA_NEWB,
+#endif  // CONFIG_EXT_REFS
+
+  THR_COMP_INTERINTRA_ZEROA,
+  THR_COMP_INTERINTRA_NEARESTA,
+  THR_COMP_INTERINTRA_NEARA,
+  THR_COMP_INTERINTRA_NEWA,
+#endif  // CONFIG_EXT_INTER
+  MAX_MODES
+} THR_MODES;
+
+typedef enum {
+  THR_LAST,
+#if CONFIG_EXT_REFS
+  THR_LAST2,
+  THR_LAST3,
+  THR_BWDR,
+#endif  // CONFIG_EXT_REFS
+  THR_GOLD,
+  THR_ALTR,
+
+  THR_COMP_LA,
+#if CONFIG_EXT_REFS
+  THR_COMP_L2A,
+  THR_COMP_L3A,
+#endif  // CONFIG_EXT_REFS
+  THR_COMP_GA,
+
+#if CONFIG_EXT_REFS
+  THR_COMP_LB,
+  THR_COMP_L2B,
+  THR_COMP_L3B,
+  THR_COMP_GB,
+#endif  // CONFIG_EXT_REFS
+
+  THR_INTRA,
+} THR_MODES_SUB8X8;
+
+typedef struct RD_OPT {
+  // Thresh_mult is used to set a threshold for the rd score. A higher value
+  // means that we will accept the best mode so far more often. This number
+  // is used in combination with the current block size, and thresh_freq_fact
+  // to pick a threshold.
+  int thresh_mult[MAX_MODES];
+  int thresh_mult_sub8x8[MAX_REFS];
+
+  int threshes[MAX_SEGMENTS][BLOCK_SIZES][MAX_MODES];
+
+  int64_t prediction_type_threshes[TOTAL_REFS_PER_FRAME][REFERENCE_MODES];
+
+  int RDMULT;
+  int RDDIV;
+} RD_OPT;
+
+static INLINE void av1_init_rd_stats(RD_STATS *rd_stats) {
+#if CONFIG_RD_DEBUG
+  int plane;
+#endif
+  rd_stats->rate = 0;
+  rd_stats->dist = 0;
+  rd_stats->rdcost = 0;
+  rd_stats->sse = 0;
+  rd_stats->skip = 1;
+#if CONFIG_RD_DEBUG
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    rd_stats->txb_coeff_cost[plane] = 0;
+#if CONFIG_VAR_TX
+    {
+      int r, c;
+      for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
+        for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c)
+          rd_stats->txb_coeff_cost_map[plane][r][c] = 0;
+    }
+#endif
+  }
+#endif
+}
+
+static INLINE void av1_invalid_rd_stats(RD_STATS *rd_stats) {
+#if CONFIG_RD_DEBUG
+  int plane;
+#endif
+  rd_stats->rate = INT_MAX;
+  rd_stats->dist = INT64_MAX;
+  rd_stats->rdcost = INT64_MAX;
+  rd_stats->sse = INT64_MAX;
+  rd_stats->skip = 0;
+#if CONFIG_RD_DEBUG
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    rd_stats->txb_coeff_cost[plane] = INT_MAX;
+#if CONFIG_VAR_TX
+    {
+      int r, c;
+      for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
+        for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c)
+          rd_stats->txb_coeff_cost_map[plane][r][c] = INT_MAX;
+    }
+#endif
+  }
+#endif
+}
+
+static INLINE void av1_merge_rd_stats(RD_STATS *rd_stats_dst,
+                                      const RD_STATS *rd_stats_src) {
+#if CONFIG_RD_DEBUG
+  int plane;
+#endif
+  rd_stats_dst->rate += rd_stats_src->rate;
+  rd_stats_dst->dist += rd_stats_src->dist;
+  rd_stats_dst->sse += rd_stats_src->sse;
+  rd_stats_dst->skip &= rd_stats_src->skip;
+#if CONFIG_RD_DEBUG
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    rd_stats_dst->txb_coeff_cost[plane] += rd_stats_src->txb_coeff_cost[plane];
+#if CONFIG_VAR_TX
+    {
+      // TODO(angiebird): optimize this part
+      int r, c;
+      int ref_txb_coeff_cost = 0;
+      for (r = 0; r < TXB_COEFF_COST_MAP_SIZE; ++r)
+        for (c = 0; c < TXB_COEFF_COST_MAP_SIZE; ++c) {
+          rd_stats_dst->txb_coeff_cost_map[plane][r][c] +=
+              rd_stats_src->txb_coeff_cost_map[plane][r][c];
+          ref_txb_coeff_cost += rd_stats_dst->txb_coeff_cost_map[plane][r][c];
+        }
+      assert(ref_txb_coeff_cost == rd_stats_dst->txb_coeff_cost[plane]);
+    }
+#endif
+  }
+#endif
+}
+
+struct TileInfo;
+struct TileDataEnc;
+struct AV1_COMP;
+struct macroblock;
+
+int av1_compute_rd_mult(const struct AV1_COMP *cpi, int qindex);
+
+void av1_initialize_rd_consts(struct AV1_COMP *cpi);
+
+void av1_initialize_me_consts(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                              int qindex);
+
+void av1_model_rd_from_var_lapndz(int64_t var, unsigned int n,
+                                  unsigned int qstep, int *rate, int64_t *dist);
+
+int av1_get_switchable_rate(const struct AV1_COMP *cpi, const MACROBLOCKD *xd);
+
+int av1_raster_block_offset(BLOCK_SIZE plane_bsize, int raster_block,
+                            int stride);
+
+int16_t *av1_raster_block_offset_int16(BLOCK_SIZE plane_bsize, int raster_block,
+                                       int16_t *base);
+
+YV12_BUFFER_CONFIG *av1_get_scaled_ref_frame(const struct AV1_COMP *cpi,
+                                             int ref_frame);
+
+void av1_init_me_luts(void);
+
+#if CONFIG_REF_MV
+void av1_set_mvcost(MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame, int ref,
+                    int ref_mv_idx);
+#endif
+
+void av1_get_entropy_contexts(BLOCK_SIZE bsize, TX_SIZE tx_size,
+                              const struct macroblockd_plane *pd,
+                              ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE],
+                              ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE]);
+
+void av1_set_rd_speed_thresholds(struct AV1_COMP *cpi);
+
+void av1_set_rd_speed_thresholds_sub8x8(struct AV1_COMP *cpi);
+
+void av1_update_rd_thresh_fact(const AV1_COMMON *const cm,
+                               int (*fact)[MAX_MODES], int rd_thresh, int bsize,
+                               int best_mode_index);
+
+void av1_fill_token_costs(av1_coeff_cost *c,
+                          av1_coeff_probs_model (*p)[PLANE_TYPES]);
+
+static INLINE int rd_less_than_thresh(int64_t best_rd, int thresh,
+                                      int thresh_fact) {
+  return best_rd < ((int64_t)thresh * thresh_fact >> 5) || thresh == INT_MAX;
+}
+
+void av1_mv_pred(const struct AV1_COMP *cpi, MACROBLOCK *x,
+                 uint8_t *ref_y_buffer, int ref_y_stride, int ref_frame,
+                 BLOCK_SIZE block_size);
+
+static INLINE void set_error_per_bit(MACROBLOCK *x, int rdmult) {
+  x->errorperbit = rdmult >> RD_EPB_SHIFT;
+  x->errorperbit += (x->errorperbit == 0);
+}
+
+void av1_setup_pred_block(const MACROBLOCKD *xd,
+                          struct buf_2d dst[MAX_MB_PLANE],
+                          const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
+                          const struct scale_factors *scale,
+                          const struct scale_factors *scale_uv);
+
+int av1_get_intra_cost_penalty(int qindex, int qdelta,
+                               aom_bit_depth_t bit_depth);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_RD_H_
diff --git a/third_party/aom/av1/encoder/rdopt.c b/third_party/aom/av1/encoder/rdopt.c
new file mode 100644
index 0000000000..a1096f782d
--- /dev/null
+++ b/third_party/aom/av1/encoder/rdopt.c
@@ -0,0 +1,12713 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/system_state.h"
+
+#include "av1/common/common.h"
+#include "av1/common/common_data.h"
+#include "av1/common/entropy.h"
+#include "av1/common/entropymode.h"
+#include "av1/common/idct.h"
+#include "av1/common/mvref_common.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/reconintra.h"
+#include "av1/common/scan.h"
+#include "av1/common/seg_common.h"
+#if CONFIG_LV_MAP
+#include "av1/common/txb_common.h"
+#endif
+#if CONFIG_WARPED_MOTION
+#include "av1/common/warped_motion.h"
+#endif  // CONFIG_WARPED_MOTION
+
+#include "av1/encoder/aq_variance.h"
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/cost.h"
+#include "av1/encoder/encodemb.h"
+#include "av1/encoder/encodemv.h"
+#include "av1/encoder/encoder.h"
+#if CONFIG_LV_MAP
+#include "av1/encoder/encodetxb.h"
+#endif
+#include "av1/encoder/hybrid_fwd_txfm.h"
+#include "av1/encoder/mcomp.h"
+#if CONFIG_PALETTE
+#include "av1/encoder/palette.h"
+#endif  // CONFIG_PALETTE
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/rd.h"
+#include "av1/encoder/rdopt.h"
+#include "av1/encoder/tokenize.h"
+#if CONFIG_PVQ
+#include "av1/encoder/pvq_encoder.h"
+#endif  // CONFIG_PVQ
+#if CONFIG_PVQ || CONFIG_DAALA_DIST
+#include "av1/common/pvq.h"
+#endif  // CONFIG_PVQ || CONFIG_DAALA_DIST
+#if CONFIG_DUAL_FILTER
+#define DUAL_FILTER_SET_SIZE (SWITCHABLE_FILTERS * SWITCHABLE_FILTERS)
+static const int filter_sets[DUAL_FILTER_SET_SIZE][2] = {
+  { 0, 0 }, { 0, 1 }, { 0, 2 }, { 0, 3 }, { 1, 0 }, { 1, 1 },
+  { 1, 2 }, { 1, 3 }, { 2, 0 }, { 2, 1 }, { 2, 2 }, { 2, 3 },
+  { 3, 0 }, { 3, 1 }, { 3, 2 }, { 3, 3 },
+};
+#endif  // CONFIG_DUAL_FILTER
+
+#if CONFIG_EXT_REFS
+
+#define LAST_FRAME_MODE_MASK                                      \
+  ((1 << INTRA_FRAME) | (1 << LAST2_FRAME) | (1 << LAST3_FRAME) | \
+   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
+#define LAST2_FRAME_MODE_MASK                                    \
+  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST3_FRAME) | \
+   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
+#define LAST3_FRAME_MODE_MASK                                    \
+  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
+   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
+#define GOLDEN_FRAME_MODE_MASK                                   \
+  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
+   (1 << LAST3_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
+#define BWDREF_FRAME_MODE_MASK                                   \
+  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
+   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME))
+#define ALTREF_FRAME_MODE_MASK                                   \
+  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
+   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME))
+
+#else
+
+#define LAST_FRAME_MODE_MASK \
+  ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
+#define GOLDEN_FRAME_MODE_MASK \
+  ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
+#define ALTREF_FRAME_MODE_MASK \
+  ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | (1 << INTRA_FRAME))
+
+#endif  // CONFIG_EXT_REFS
+
+#if CONFIG_EXT_REFS
+#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | (1 << BWDREF_FRAME) | 0x01)
+#else
+#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
+#endif  // CONFIG_EXT_REFS
+
+#define MIN_EARLY_TERM_INDEX 3
+#define NEW_MV_DISCOUNT_FACTOR 8
+
+#if CONFIG_EXT_INTRA
+#define ANGLE_SKIP_THRESH 10
+#define FILTER_FAST_SEARCH 1
+#endif  // CONFIG_EXT_INTRA
+
+const double ADST_FLIP_SVM[8] = { -6.6623, -2.8062, -3.2531, 3.1671,    // vert
+                                  -7.7051, -3.2234, -3.6193, 3.4533 };  // horz
+
+typedef struct {
+  PREDICTION_MODE mode;
+  MV_REFERENCE_FRAME ref_frame[2];
+} MODE_DEFINITION;
+
+typedef struct { MV_REFERENCE_FRAME ref_frame[2]; } REF_DEFINITION;
+
+struct rdcost_block_args {
+  const AV1_COMP *cpi;
+  MACROBLOCK *x;
+  ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE];
+  ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE];
+  RD_STATS rd_stats;
+  int64_t this_rd;
+  int64_t best_rd;
+  int exit_early;
+  int use_fast_coef_costing;
+};
+
+#define LAST_NEW_MV_INDEX 6
+static const MODE_DEFINITION av1_mode_order[MAX_MODES] = {
+  { NEARESTMV, { LAST_FRAME, NONE_FRAME } },
+#if CONFIG_EXT_REFS
+  { NEARESTMV, { LAST2_FRAME, NONE_FRAME } },
+  { NEARESTMV, { LAST3_FRAME, NONE_FRAME } },
+  { NEARESTMV, { BWDREF_FRAME, NONE_FRAME } },
+#endif  // CONFIG_EXT_REFS
+  { NEARESTMV, { ALTREF_FRAME, NONE_FRAME } },
+  { NEARESTMV, { GOLDEN_FRAME, NONE_FRAME } },
+
+  { DC_PRED, { INTRA_FRAME, NONE_FRAME } },
+
+  { NEWMV, { LAST_FRAME, NONE_FRAME } },
+#if CONFIG_EXT_REFS
+  { NEWMV, { LAST2_FRAME, NONE_FRAME } },
+  { NEWMV, { LAST3_FRAME, NONE_FRAME } },
+  { NEWMV, { BWDREF_FRAME, NONE_FRAME } },
+#endif  // CONFIG_EXT_REFS
+  { NEWMV, { ALTREF_FRAME, NONE_FRAME } },
+  { NEWMV, { GOLDEN_FRAME, NONE_FRAME } },
+
+  { NEARMV, { LAST_FRAME, NONE_FRAME } },
+#if CONFIG_EXT_REFS
+  { NEARMV, { LAST2_FRAME, NONE_FRAME } },
+  { NEARMV, { LAST3_FRAME, NONE_FRAME } },
+  { NEARMV, { BWDREF_FRAME, NONE_FRAME } },
+#endif  // CONFIG_EXT_REFS
+  { NEARMV, { ALTREF_FRAME, NONE_FRAME } },
+  { NEARMV, { GOLDEN_FRAME, NONE_FRAME } },
+
+  { ZEROMV, { LAST_FRAME, NONE_FRAME } },
+#if CONFIG_EXT_REFS
+  { ZEROMV, { LAST2_FRAME, NONE_FRAME } },
+  { ZEROMV, { LAST3_FRAME, NONE_FRAME } },
+  { ZEROMV, { BWDREF_FRAME, NONE_FRAME } },
+#endif  // CONFIG_EXT_REFS
+  { ZEROMV, { GOLDEN_FRAME, NONE_FRAME } },
+  { ZEROMV, { ALTREF_FRAME, NONE_FRAME } },
+
+// TODO(zoeliu): May need to reconsider the order on the modes to check
+
+#if CONFIG_EXT_INTER
+  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
+#if CONFIG_EXT_REFS
+  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+#if CONFIG_EXT_REFS
+  { NEAREST_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+
+#else  // CONFIG_EXT_INTER
+
+  { NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
+#if CONFIG_EXT_REFS
+  { NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+  { NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+#if CONFIG_EXT_REFS
+  { NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+#endif  // CONFIG_EXT_INTER
+
+  { TM_PRED, { INTRA_FRAME, NONE_FRAME } },
+
+#if CONFIG_ALT_INTRA
+  { SMOOTH_PRED, { INTRA_FRAME, NONE_FRAME } },
+#endif  // CONFIG_ALT_INTRA
+
+#if CONFIG_EXT_INTER
+  { NEAR_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
+  { NEAREST_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
+  { NEAR_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
+  { NEW_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
+  { NEAREST_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
+  { NEW_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
+  { NEAR_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
+  { NEW_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
+  { ZERO_ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
+
+#if CONFIG_EXT_REFS
+  { NEAR_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEAREST_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEAR_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEW_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEAREST_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEW_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEAR_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEW_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { ZERO_ZEROMV, { LAST2_FRAME, ALTREF_FRAME } },
+
+  { NEAR_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
+  { NEAREST_NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
+  { NEAR_NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
+  { NEW_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
+  { NEAREST_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
+  { NEW_NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
+  { NEAR_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
+  { NEW_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
+  { ZERO_ZEROMV, { LAST3_FRAME, ALTREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+
+  { NEAR_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+  { NEAREST_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+  { NEAR_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+  { NEW_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+  { NEAREST_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+  { NEW_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+  { NEAR_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+  { NEW_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+  { ZERO_ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+
+#if CONFIG_EXT_REFS
+  { NEAR_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEAR_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEW_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEW_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEAR_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEW_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
+  { ZERO_ZEROMV, { LAST_FRAME, BWDREF_FRAME } },
+
+  { NEAR_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEAR_NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEW_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEW_NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEAR_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEW_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { ZERO_ZEROMV, { LAST2_FRAME, BWDREF_FRAME } },
+
+  { NEAR_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEAR_NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEW_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEW_NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEAR_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEW_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { ZERO_ZEROMV, { LAST3_FRAME, BWDREF_FRAME } },
+
+  { NEAR_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+  { NEAR_NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+  { NEW_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+  { NEAREST_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+  { NEW_NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+  { NEAR_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+  { NEW_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+  { ZERO_ZEROMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+
+#else  // CONFIG_EXT_INTER
+
+  { NEARMV, { LAST_FRAME, ALTREF_FRAME } },
+  { NEWMV, { LAST_FRAME, ALTREF_FRAME } },
+#if CONFIG_EXT_REFS
+  { NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
+  { NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+  { NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+  { NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+
+#if CONFIG_EXT_REFS
+  { NEARMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEWMV, { LAST_FRAME, BWDREF_FRAME } },
+  { NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+  { NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+
+  { ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
+#if CONFIG_EXT_REFS
+  { ZEROMV, { LAST2_FRAME, ALTREF_FRAME } },
+  { ZEROMV, { LAST3_FRAME, ALTREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+  { ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
+
+#if CONFIG_EXT_REFS
+  { ZEROMV, { LAST_FRAME, BWDREF_FRAME } },
+  { ZEROMV, { LAST2_FRAME, BWDREF_FRAME } },
+  { ZEROMV, { LAST3_FRAME, BWDREF_FRAME } },
+  { ZEROMV, { GOLDEN_FRAME, BWDREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+
+#endif  // CONFIG_EXT_INTER
+
+  { H_PRED, { INTRA_FRAME, NONE_FRAME } },
+  { V_PRED, { INTRA_FRAME, NONE_FRAME } },
+  { D135_PRED, { INTRA_FRAME, NONE_FRAME } },
+  { D207_PRED, { INTRA_FRAME, NONE_FRAME } },
+  { D153_PRED, { INTRA_FRAME, NONE_FRAME } },
+  { D63_PRED, { INTRA_FRAME, NONE_FRAME } },
+  { D117_PRED, { INTRA_FRAME, NONE_FRAME } },
+  { D45_PRED, { INTRA_FRAME, NONE_FRAME } },
+
+#if CONFIG_EXT_INTER
+  { ZEROMV, { LAST_FRAME, INTRA_FRAME } },
+  { NEARESTMV, { LAST_FRAME, INTRA_FRAME } },
+  { NEARMV, { LAST_FRAME, INTRA_FRAME } },
+  { NEWMV, { LAST_FRAME, INTRA_FRAME } },
+
+#if CONFIG_EXT_REFS
+  { ZEROMV, { LAST2_FRAME, INTRA_FRAME } },
+  { NEARESTMV, { LAST2_FRAME, INTRA_FRAME } },
+  { NEARMV, { LAST2_FRAME, INTRA_FRAME } },
+  { NEWMV, { LAST2_FRAME, INTRA_FRAME } },
+
+  { ZEROMV, { LAST3_FRAME, INTRA_FRAME } },
+  { NEARESTMV, { LAST3_FRAME, INTRA_FRAME } },
+  { NEARMV, { LAST3_FRAME, INTRA_FRAME } },
+  { NEWMV, { LAST3_FRAME, INTRA_FRAME } },
+#endif  // CONFIG_EXT_REFS
+
+  { ZEROMV, { GOLDEN_FRAME, INTRA_FRAME } },
+  { NEARESTMV, { GOLDEN_FRAME, INTRA_FRAME } },
+  { NEARMV, { GOLDEN_FRAME, INTRA_FRAME } },
+  { NEWMV, { GOLDEN_FRAME, INTRA_FRAME } },
+
+#if CONFIG_EXT_REFS
+  { ZEROMV, { BWDREF_FRAME, INTRA_FRAME } },
+  { NEARESTMV, { BWDREF_FRAME, INTRA_FRAME } },
+  { NEARMV, { BWDREF_FRAME, INTRA_FRAME } },
+  { NEWMV, { BWDREF_FRAME, INTRA_FRAME } },
+#endif  // CONFIG_EXT_REFS
+
+  { ZEROMV, { ALTREF_FRAME, INTRA_FRAME } },
+  { NEARESTMV, { ALTREF_FRAME, INTRA_FRAME } },
+  { NEARMV, { ALTREF_FRAME, INTRA_FRAME } },
+  { NEWMV, { ALTREF_FRAME, INTRA_FRAME } },
+#endif  // CONFIG_EXT_INTER
+};
+
+static const REF_DEFINITION av1_ref_order[MAX_REFS] = {
+  { { LAST_FRAME, NONE_FRAME } },
+#if CONFIG_EXT_REFS
+  { { LAST2_FRAME, NONE_FRAME } },    { { LAST3_FRAME, NONE_FRAME } },
+  { { BWDREF_FRAME, NONE_FRAME } },
+#endif  // CONFIG_EXT_REFS
+  { { GOLDEN_FRAME, NONE_FRAME } },   { { ALTREF_FRAME, NONE_FRAME } },
+
+  { { LAST_FRAME, ALTREF_FRAME } },
+#if CONFIG_EXT_REFS
+  { { LAST2_FRAME, ALTREF_FRAME } },  { { LAST3_FRAME, ALTREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+  { { GOLDEN_FRAME, ALTREF_FRAME } },
+
+#if CONFIG_EXT_REFS
+  { { LAST_FRAME, BWDREF_FRAME } },   { { LAST2_FRAME, BWDREF_FRAME } },
+  { { LAST3_FRAME, BWDREF_FRAME } },  { { GOLDEN_FRAME, BWDREF_FRAME } },
+#endif  // CONFIG_EXT_REFS
+
+  { { INTRA_FRAME, NONE_FRAME } },
+};
+
+#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
+static INLINE int write_uniform_cost(int n, int v) {
+  const int l = get_unsigned_bits(n);
+  const int m = (1 << l) - n;
+  if (l == 0) return 0;
+  if (v < m)
+    return (l - 1) * av1_cost_bit(128, 0);
+  else
+    return l * av1_cost_bit(128, 0);
+}
+#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
+
+// constants for prune 1 and prune 2 decision boundaries
+#define FAST_EXT_TX_CORR_MID 0.0
+#define FAST_EXT_TX_EDST_MID 0.1
+#define FAST_EXT_TX_CORR_MARGIN 0.5
+#define FAST_EXT_TX_EDST_MARGIN 0.3
+
+static const TX_TYPE_1D vtx_tab[TX_TYPES] = {
+  DCT_1D,      ADST_1D, DCT_1D,      ADST_1D,
+#if CONFIG_EXT_TX
+  FLIPADST_1D, DCT_1D,  FLIPADST_1D, ADST_1D, FLIPADST_1D, IDTX_1D,
+  DCT_1D,      IDTX_1D, ADST_1D,     IDTX_1D, FLIPADST_1D, IDTX_1D,
+#endif  // CONFIG_EXT_TX
+};
+
+static const TX_TYPE_1D htx_tab[TX_TYPES] = {
+  DCT_1D,  DCT_1D,      ADST_1D,     ADST_1D,
+#if CONFIG_EXT_TX
+  DCT_1D,  FLIPADST_1D, FLIPADST_1D, FLIPADST_1D, ADST_1D, IDTX_1D,
+  IDTX_1D, DCT_1D,      IDTX_1D,     ADST_1D,     IDTX_1D, FLIPADST_1D,
+#endif  // CONFIG_EXT_TX
+};
+
+#if CONFIG_DAALA_DIST
+static int od_compute_var_4x4(od_coeff *x, int stride) {
+  int sum;
+  int s2;
+  int i;
+  sum = 0;
+  s2 = 0;
+  for (i = 0; i < 4; i++) {
+    int j;
+    for (j = 0; j < 4; j++) {
+      int t;
+
+      t = x[i * stride + j];
+      sum += t;
+      s2 += t * t;
+    }
+  }
+  // TODO(yushin) : Check wheter any changes are required for high bit depth.
+  return (s2 - (sum * sum >> 4)) >> 4;
+}
+
+/* OD_DIST_LP_MID controls the frequency weighting filter used for computing
+   the distortion. For a value X, the filter is [1 X 1]/(X + 2) and
+   is applied both horizontally and vertically. For X=5, the filter is
+   a good approximation for the OD_QM8_Q4_HVS quantization matrix. */
+#define OD_DIST_LP_MID (5)
+#define OD_DIST_LP_NORM (OD_DIST_LP_MID + 2)
+
+static double od_compute_dist_8x8(int qm, int use_activity_masking, od_coeff *x,
+                                  od_coeff *y, od_coeff *e_lp, int stride) {
+  double sum;
+  int min_var;
+  double mean_var;
+  double var_stat;
+  double activity;
+  double calibration;
+  int i;
+  int j;
+  double vardist;
+
+  vardist = 0;
+  OD_ASSERT(qm != OD_FLAT_QM);
+  (void)qm;
+#if 1
+  min_var = INT_MAX;
+  mean_var = 0;
+  for (i = 0; i < 3; i++) {
+    for (j = 0; j < 3; j++) {
+      int varx;
+      int vary;
+      varx = od_compute_var_4x4(x + 2 * i * stride + 2 * j, stride);
+      vary = od_compute_var_4x4(y + 2 * i * stride + 2 * j, stride);
+      min_var = OD_MINI(min_var, varx);
+      mean_var += 1. / (1 + varx);
+      /* The cast to (double) is to avoid an overflow before the sqrt.*/
+      vardist += varx - 2 * sqrt(varx * (double)vary) + vary;
+    }
+  }
+  /* We use a different variance statistic depending on whether activity
+     masking is used, since the harmonic mean appeared slghtly worse with
+     masking off. The calibration constant just ensures that we preserve the
+     rate compared to activity=1. */
+  if (use_activity_masking) {
+    calibration = 1.95;
+    var_stat = 9. / mean_var;
+  } else {
+    calibration = 1.62;
+    var_stat = min_var;
+  }
+  /* 1.62 is a calibration constant, 0.25 is a noise floor and 1/6 is the
+     activity masking constant. */
+  activity = calibration * pow(.25 + var_stat, -1. / 6);
+#else
+  activity = 1;
+#endif  // 1
+  sum = 0;
+  for (i = 0; i < 8; i++) {
+    for (j = 0; j < 8; j++)
+      sum += e_lp[i * stride + j] * (double)e_lp[i * stride + j];
+  }
+  /* Normalize the filter to unit DC response. */
+  sum *= 1. / (OD_DIST_LP_NORM * OD_DIST_LP_NORM * OD_DIST_LP_NORM *
+               OD_DIST_LP_NORM);
+  return activity * activity * (sum + vardist);
+}
+
+// Note : Inputs x and y are in a pixel domain
+static double od_compute_dist(int qm, int activity_masking, od_coeff *x,
+                              od_coeff *y, int bsize_w, int bsize_h,
+                              int qindex) {
+  int i;
+  double sum;
+  sum = 0;
+
+  assert(bsize_w >= 8 && bsize_h >= 8);
+
+  if (qm == OD_FLAT_QM) {
+    for (i = 0; i < bsize_w * bsize_h; i++) {
+      double tmp;
+      tmp = x[i] - y[i];
+      sum += tmp * tmp;
+    }
+  } else {
+    int j;
+    DECLARE_ALIGNED(16, od_coeff, e[MAX_TX_SQUARE]);
+    DECLARE_ALIGNED(16, od_coeff, tmp[MAX_TX_SQUARE]);
+    DECLARE_ALIGNED(16, od_coeff, e_lp[MAX_TX_SQUARE]);
+    int mid = OD_DIST_LP_MID;
+    for (i = 0; i < bsize_h; i++) {
+      for (j = 0; j < bsize_w; j++) {
+        e[i * bsize_w + j] = x[i * bsize_w + j] - y[i * bsize_w + j];
+      }
+    }
+    for (i = 0; i < bsize_h; i++) {
+      tmp[i * bsize_w] = mid * e[i * bsize_w] + 2 * e[i * bsize_w + 1];
+      tmp[i * bsize_w + bsize_w - 1] =
+          mid * e[i * bsize_w + bsize_w - 1] + 2 * e[i * bsize_w + bsize_w - 2];
+      for (j = 1; j < bsize_w - 1; j++) {
+        tmp[i * bsize_w + j] = mid * e[i * bsize_w + j] +
+                               e[i * bsize_w + j - 1] + e[i * bsize_w + j + 1];
+      }
+    }
+    for (j = 0; j < bsize_w; j++) {
+      e_lp[j] = mid * tmp[j] + 2 * tmp[bsize_w + j];
+      e_lp[(bsize_h - 1) * bsize_w + j] =
+          mid * tmp[(bsize_h - 1) * bsize_w + j] +
+          2 * tmp[(bsize_h - 2) * bsize_w + j];
+    }
+    for (i = 1; i < bsize_h - 1; i++) {
+      for (j = 0; j < bsize_w; j++) {
+        e_lp[i * bsize_w + j] = mid * tmp[i * bsize_w + j] +
+                                tmp[(i - 1) * bsize_w + j] +
+                                tmp[(i + 1) * bsize_w + j];
+      }
+    }
+    for (i = 0; i < bsize_h; i += 8) {
+      for (j = 0; j < bsize_w; j += 8) {
+        sum += od_compute_dist_8x8(qm, activity_masking, &x[i * bsize_w + j],
+                                   &y[i * bsize_w + j], &e_lp[i * bsize_w + j],
+                                   bsize_w);
+      }
+    }
+    /* Scale according to linear regression against SSE, for 8x8 blocks. */
+    if (activity_masking) {
+      sum *= 2.2 + (1.7 - 2.2) * (qindex - 99) / (210 - 99) +
+             (qindex < 99 ? 2.5 * (qindex - 99) / 99 * (qindex - 99) / 99 : 0);
+    } else {
+      sum *= qindex >= 128
+                 ? 1.4 + (0.9 - 1.4) * (qindex - 128) / (209 - 128)
+                 : qindex <= 43
+                       ? 1.5 + (2.0 - 1.5) * (qindex - 43) / (16 - 43)
+                       : 1.5 + (1.4 - 1.5) * (qindex - 43) / (128 - 43);
+    }
+  }
+  return sum;
+}
+
+static int64_t av1_daala_dist(const uint8_t *src, int src_stride,
+                              const uint8_t *dst, int dst_stride, int bsw,
+                              int bsh, int qm, int use_activity_masking,
+                              int qindex) {
+  int i, j;
+  int64_t d;
+  DECLARE_ALIGNED(16, od_coeff, orig[MAX_TX_SQUARE]);
+  DECLARE_ALIGNED(16, od_coeff, rec[MAX_TX_SQUARE]);
+
+  assert(qm == OD_HVS_QM);
+
+  for (j = 0; j < bsh; j++)
+    for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];
+
+  for (j = 0; j < bsh; j++)
+    for (i = 0; i < bsw; i++) rec[j * bsw + i] = dst[j * dst_stride + i];
+
+  d = (int64_t)od_compute_dist(qm, use_activity_masking, orig, rec, bsw, bsh,
+                               qindex);
+  return d;
+}
+#endif  // CONFIG_DAALA_DIST
+
+static void get_energy_distribution_fine(const AV1_COMP *cpi, BLOCK_SIZE bsize,
+                                         const uint8_t *src, int src_stride,
+                                         const uint8_t *dst, int dst_stride,
+                                         double *hordist, double *verdist) {
+  const int bw = block_size_wide[bsize];
+  const int bh = block_size_high[bsize];
+  unsigned int esq[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
+
+  const int f_index = bsize - BLOCK_16X16;
+  if (f_index < 0) {
+    const int w_shift = bw == 8 ? 1 : 2;
+    const int h_shift = bh == 8 ? 1 : 2;
+#if CONFIG_HIGHBITDEPTH
+    if (cpi->common.use_highbitdepth) {
+      const uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+      const uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
+      for (int i = 0; i < bh; ++i)
+        for (int j = 0; j < bw; ++j) {
+          const int index = (j >> w_shift) + ((i >> h_shift) << 2);
+          esq[index] +=
+              (src16[j + i * src_stride] - dst16[j + i * dst_stride]) *
+              (src16[j + i * src_stride] - dst16[j + i * dst_stride]);
+        }
+    } else {
+#endif  // CONFIG_HIGHBITDEPTH
+
+      for (int i = 0; i < bh; ++i)
+        for (int j = 0; j < bw; ++j) {
+          const int index = (j >> w_shift) + ((i >> h_shift) << 2);
+          esq[index] += (src[j + i * src_stride] - dst[j + i * dst_stride]) *
+                        (src[j + i * src_stride] - dst[j + i * dst_stride]);
+        }
+#if CONFIG_HIGHBITDEPTH
+    }
+#endif  // CONFIG_HIGHBITDEPTH
+  } else {
+    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[0]);
+    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
+                            &esq[1]);
+    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
+                            &esq[2]);
+    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
+                            dst_stride, &esq[3]);
+    src += bh / 4 * src_stride;
+    dst += bh / 4 * dst_stride;
+
+    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[4]);
+    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
+                            &esq[5]);
+    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
+                            &esq[6]);
+    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
+                            dst_stride, &esq[7]);
+    src += bh / 4 * src_stride;
+    dst += bh / 4 * dst_stride;
+
+    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[8]);
+    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
+                            &esq[9]);
+    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
+                            &esq[10]);
+    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
+                            dst_stride, &esq[11]);
+    src += bh / 4 * src_stride;
+    dst += bh / 4 * dst_stride;
+
+    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[12]);
+    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
+                            &esq[13]);
+    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
+                            &esq[14]);
+    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
+                            dst_stride, &esq[15]);
+  }
+
+  double total = (double)esq[0] + esq[1] + esq[2] + esq[3] + esq[4] + esq[5] +
+                 esq[6] + esq[7] + esq[8] + esq[9] + esq[10] + esq[11] +
+                 esq[12] + esq[13] + esq[14] + esq[15];
+  if (total > 0) {
+    const double e_recip = 1.0 / total;
+    hordist[0] = ((double)esq[0] + esq[4] + esq[8] + esq[12]) * e_recip;
+    hordist[1] = ((double)esq[1] + esq[5] + esq[9] + esq[13]) * e_recip;
+    hordist[2] = ((double)esq[2] + esq[6] + esq[10] + esq[14]) * e_recip;
+    verdist[0] = ((double)esq[0] + esq[1] + esq[2] + esq[3]) * e_recip;
+    verdist[1] = ((double)esq[4] + esq[5] + esq[6] + esq[7]) * e_recip;
+    verdist[2] = ((double)esq[8] + esq[9] + esq[10] + esq[11]) * e_recip;
+  } else {
+    hordist[0] = verdist[0] = 0.25;
+    hordist[1] = verdist[1] = 0.25;
+    hordist[2] = verdist[2] = 0.25;
+  }
+}
+
+static int adst_vs_flipadst(const AV1_COMP *cpi, BLOCK_SIZE bsize,
+                            const uint8_t *src, int src_stride,
+                            const uint8_t *dst, int dst_stride) {
+  int prune_bitmask = 0;
+  double svm_proj_h = 0, svm_proj_v = 0;
+  double hdist[3] = { 0, 0, 0 }, vdist[3] = { 0, 0, 0 };
+  get_energy_distribution_fine(cpi, bsize, src, src_stride, dst, dst_stride,
+                               hdist, vdist);
+
+  svm_proj_v = vdist[0] * ADST_FLIP_SVM[0] + vdist[1] * ADST_FLIP_SVM[1] +
+               vdist[2] * ADST_FLIP_SVM[2] + ADST_FLIP_SVM[3];
+  svm_proj_h = hdist[0] * ADST_FLIP_SVM[4] + hdist[1] * ADST_FLIP_SVM[5] +
+               hdist[2] * ADST_FLIP_SVM[6] + ADST_FLIP_SVM[7];
+  if (svm_proj_v > FAST_EXT_TX_EDST_MID + FAST_EXT_TX_EDST_MARGIN)
+    prune_bitmask |= 1 << FLIPADST_1D;
+  else if (svm_proj_v < FAST_EXT_TX_EDST_MID - FAST_EXT_TX_EDST_MARGIN)
+    prune_bitmask |= 1 << ADST_1D;
+
+  if (svm_proj_h > FAST_EXT_TX_EDST_MID + FAST_EXT_TX_EDST_MARGIN)
+    prune_bitmask |= 1 << (FLIPADST_1D + 8);
+  else if (svm_proj_h < FAST_EXT_TX_EDST_MID - FAST_EXT_TX_EDST_MARGIN)
+    prune_bitmask |= 1 << (ADST_1D + 8);
+
+  return prune_bitmask;
+}
+
+#if CONFIG_EXT_TX
+static void get_horver_correlation(const int16_t *diff, int stride, int w,
+                                   int h, double *hcorr, double *vcorr) {
+  // Returns hor/ver correlation coefficient
+  const int num = (h - 1) * (w - 1);
+  double num_r;
+  int i, j;
+  int64_t xy_sum = 0, xz_sum = 0;
+  int64_t x_sum = 0, y_sum = 0, z_sum = 0;
+  int64_t x2_sum = 0, y2_sum = 0, z2_sum = 0;
+  double x_var_n, y_var_n, z_var_n, xy_var_n, xz_var_n;
+  *hcorr = *vcorr = 1;
+
+  assert(num > 0);
+  num_r = 1.0 / num;
+  for (i = 1; i < h; ++i) {
+    for (j = 1; j < w; ++j) {
+      const int16_t x = diff[i * stride + j];
+      const int16_t y = diff[i * stride + j - 1];
+      const int16_t z = diff[(i - 1) * stride + j];
+      xy_sum += x * y;
+      xz_sum += x * z;
+      x_sum += x;
+      y_sum += y;
+      z_sum += z;
+      x2_sum += x * x;
+      y2_sum += y * y;
+      z2_sum += z * z;
+    }
+  }
+  x_var_n = x2_sum - (x_sum * x_sum) * num_r;
+  y_var_n = y2_sum - (y_sum * y_sum) * num_r;
+  z_var_n = z2_sum - (z_sum * z_sum) * num_r;
+  xy_var_n = xy_sum - (x_sum * y_sum) * num_r;
+  xz_var_n = xz_sum - (x_sum * z_sum) * num_r;
+  if (x_var_n > 0 && y_var_n > 0) {
+    *hcorr = xy_var_n / sqrt(x_var_n * y_var_n);
+    *hcorr = *hcorr < 0 ? 0 : *hcorr;
+  }
+  if (x_var_n > 0 && z_var_n > 0) {
+    *vcorr = xz_var_n / sqrt(x_var_n * z_var_n);
+    *vcorr = *vcorr < 0 ? 0 : *vcorr;
+  }
+}
+
+int dct_vs_idtx(const int16_t *diff, int stride, int w, int h) {
+  double hcorr, vcorr;
+  int prune_bitmask = 0;
+  get_horver_correlation(diff, stride, w, h, &hcorr, &vcorr);
+
+  if (vcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
+    prune_bitmask |= 1 << IDTX_1D;
+  else if (vcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
+    prune_bitmask |= 1 << DCT_1D;
+
+  if (hcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
+    prune_bitmask |= 1 << (IDTX_1D + 8);
+  else if (hcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
+    prune_bitmask |= 1 << (DCT_1D + 8);
+  return prune_bitmask;
+}
+
+// Performance drop: 0.5%, Speed improvement: 24%
+static int prune_two_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
+                             MACROBLOCK *x, const MACROBLOCKD *xd,
+                             int adst_flipadst, int dct_idtx) {
+  int prune = 0;
+
+  if (adst_flipadst) {
+    const struct macroblock_plane *const p = &x->plane[0];
+    const struct macroblockd_plane *const pd = &xd->plane[0];
+    prune |= adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride,
+                              pd->dst.buf, pd->dst.stride);
+  }
+  if (dct_idtx) {
+    av1_subtract_plane(x, bsize, 0);
+    const struct macroblock_plane *const p = &x->plane[0];
+    const int bw = 4 << (b_width_log2_lookup[bsize]);
+    const int bh = 4 << (b_height_log2_lookup[bsize]);
+    prune |= dct_vs_idtx(p->src_diff, bw, bw, bh);
+  }
+
+  return prune;
+}
+#endif  // CONFIG_EXT_TX
+
+// Performance drop: 0.3%, Speed improvement: 5%
+static int prune_one_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
+                             const MACROBLOCK *x, const MACROBLOCKD *xd) {
+  const struct macroblock_plane *const p = &x->plane[0];
+  const struct macroblockd_plane *const pd = &xd->plane[0];
+  return adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride, pd->dst.buf,
+                          pd->dst.stride);
+}
+
+static int prune_tx_types(const AV1_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
+                          const MACROBLOCKD *const xd, int tx_set) {
+#if CONFIG_EXT_TX
+  const int *tx_set_1D = ext_tx_used_inter_1D[tx_set];
+#else
+  const int tx_set_1D[TX_TYPES_1D] = { 0 };
+#endif  // CONFIG_EXT_TX
+
+  switch (cpi->sf.tx_type_search.prune_mode) {
+    case NO_PRUNE: return 0; break;
+    case PRUNE_ONE:
+      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D]))
+        return 0;
+      return prune_one_for_sby(cpi, bsize, x, xd);
+      break;
+#if CONFIG_EXT_TX
+    case PRUNE_TWO:
+      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D])) {
+        if (!(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D])) return 0;
+        return prune_two_for_sby(cpi, bsize, x, xd, 0, 1);
+      }
+      if ((tx_set >= 0) && !(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D]))
+        return prune_two_for_sby(cpi, bsize, x, xd, 1, 0);
+      return prune_two_for_sby(cpi, bsize, x, xd, 1, 1);
+      break;
+#endif  // CONFIG_EXT_TX
+  }
+  assert(0);
+  return 0;
+}
+
+static int do_tx_type_search(TX_TYPE tx_type, int prune) {
+// TODO(sarahparker) implement for non ext tx
+#if CONFIG_EXT_TX
+  return !(((prune >> vtx_tab[tx_type]) & 1) |
+           ((prune >> (htx_tab[tx_type] + 8)) & 1));
+#else
+  // temporary to avoid compiler warnings
+  (void)vtx_tab;
+  (void)htx_tab;
+  (void)tx_type;
+  (void)prune;
+  return 1;
+#endif  // CONFIG_EXT_TX
+}
+
+static void model_rd_from_sse(const AV1_COMP *const cpi,
+                              const MACROBLOCKD *const xd, BLOCK_SIZE bsize,
+                              int plane, int64_t sse, int *rate,
+                              int64_t *dist) {
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int dequant_shift =
+#if CONFIG_HIGHBITDEPTH
+      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
+#endif  // CONFIG_HIGHBITDEPTH
+                                                    3;
+
+  // Fast approximate the modelling function.
+  if (cpi->sf.simple_model_rd_from_var) {
+    const int64_t square_error = sse;
+    int quantizer = (pd->dequant[1] >> dequant_shift);
+
+    if (quantizer < 120)
+      *rate = (int)((square_error * (280 - quantizer)) >>
+                    (16 - AV1_PROB_COST_SHIFT));
+    else
+      *rate = 0;
+    *dist = (square_error * quantizer) >> 8;
+  } else {
+    av1_model_rd_from_var_lapndz(sse, num_pels_log2_lookup[bsize],
+                                 pd->dequant[1] >> dequant_shift, rate, dist);
+  }
+
+  *dist <<= 4;
+}
+
+static void model_rd_for_sb(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
+                            MACROBLOCK *x, MACROBLOCKD *xd, int plane_from,
+                            int plane_to, int *out_rate_sum,
+                            int64_t *out_dist_sum, int *skip_txfm_sb,
+                            int64_t *skip_sse_sb) {
+  // Note our transform coeffs are 8 times an orthogonal transform.
+  // Hence quantizer step is also 8 times. To get effective quantizer
+  // we need to divide by 8 before sending to modeling function.
+  int plane;
+  const int ref = xd->mi[0]->mbmi.ref_frame[0];
+
+  int64_t rate_sum = 0;
+  int64_t dist_sum = 0;
+  int64_t total_sse = 0;
+
+  x->pred_sse[ref] = 0;
+
+  for (plane = plane_from; plane <= plane_to; ++plane) {
+    struct macroblock_plane *const p = &x->plane[plane];
+    struct macroblockd_plane *const pd = &xd->plane[plane];
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+    const BLOCK_SIZE bs = AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#else
+    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
+#endif  // CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+
+    unsigned int sse;
+    int rate;
+    int64_t dist;
+
+#if CONFIG_CB4X4
+    if (x->skip_chroma_rd && plane) continue;
+#endif  // CONFIG_CB4X4
+
+    // TODO(geza): Write direct sse functions that do not compute
+    // variance as well.
+    cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
+                       &sse);
+
+    if (plane == 0) x->pred_sse[ref] = sse;
+
+    total_sse += sse;
+
+    model_rd_from_sse(cpi, xd, bs, plane, sse, &rate, &dist);
+
+    rate_sum += rate;
+    dist_sum += dist;
+  }
+
+  *skip_txfm_sb = total_sse == 0;
+  *skip_sse_sb = total_sse << 4;
+  *out_rate_sum = (int)rate_sum;
+  *out_dist_sum = dist_sum;
+}
+
+int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
+                          intptr_t block_size, int64_t *ssz) {
+  int i;
+  int64_t error = 0, sqcoeff = 0;
+
+  for (i = 0; i < block_size; i++) {
+    const int diff = coeff[i] - dqcoeff[i];
+    error += diff * diff;
+    sqcoeff += coeff[i] * coeff[i];
+  }
+
+  *ssz = sqcoeff;
+  return error;
+}
+
+int64_t av1_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff,
+                             int block_size) {
+  int i;
+  int64_t error = 0;
+
+  for (i = 0; i < block_size; i++) {
+    const int diff = coeff[i] - dqcoeff[i];
+    error += diff * diff;
+  }
+
+  return error;
+}
+
+#if CONFIG_HIGHBITDEPTH
+int64_t av1_highbd_block_error_c(const tran_low_t *coeff,
+                                 const tran_low_t *dqcoeff, intptr_t block_size,
+                                 int64_t *ssz, int bd) {
+  int i;
+  int64_t error = 0, sqcoeff = 0;
+  int shift = 2 * (bd - 8);
+  int rounding = shift > 0 ? 1 << (shift - 1) : 0;
+
+  for (i = 0; i < block_size; i++) {
+    const int64_t diff = coeff[i] - dqcoeff[i];
+    error += diff * diff;
+    sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
+  }
+  assert(error >= 0 && sqcoeff >= 0);
+  error = (error + rounding) >> shift;
+  sqcoeff = (sqcoeff + rounding) >> shift;
+
+  *ssz = sqcoeff;
+  return error;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_PVQ
+// Without PVQ, av1_block_error_c() return two kind of errors,
+// 1) reconstruction (i.e. decoded) error and
+// 2) Squared sum of transformed residue (i.e. 'coeff')
+// However, if PVQ is enabled, coeff does not keep the transformed residue
+// but instead a transformed original is kept.
+// Hence, new parameter ref vector (i.e. transformed predicted signal)
+// is required to derive the residue signal,
+// i.e. coeff - ref = residue (all transformed).
+
+#if CONFIG_HIGHBITDEPTH
+static int64_t av1_highbd_block_error2_c(const tran_low_t *coeff,
+                                         const tran_low_t *dqcoeff,
+                                         const tran_low_t *ref,
+                                         intptr_t block_size, int64_t *ssz,
+                                         int bd) {
+  int64_t error;
+  int64_t sqcoeff;
+  int shift = 2 * (bd - 8);
+  int rounding = shift > 0 ? 1 << (shift - 1) : 0;
+  // Use the existing sse codes for calculating distortion of decoded signal:
+  // i.e. (orig - decoded)^2
+  // For high bit depth, throw away ssz until a 32-bit version of
+  // av1_block_error_fp is written.
+  int64_t ssz_trash;
+  error = av1_block_error(coeff, dqcoeff, block_size, &ssz_trash);
+  // prediction residue^2 = (orig - ref)^2
+  sqcoeff = av1_block_error(coeff, ref, block_size, &ssz_trash);
+  error = (error + rounding) >> shift;
+  sqcoeff = (sqcoeff + rounding) >> shift;
+  *ssz = sqcoeff;
+  return error;
+}
+#else
+// TODO(yushin) : Since 4x4 case does not need ssz, better to refactor into
+// a separate function that does not do the extra computations for ssz.
+static int64_t av1_block_error2_c(const tran_low_t *coeff,
+                                  const tran_low_t *dqcoeff,
+                                  const tran_low_t *ref, intptr_t block_size,
+                                  int64_t *ssz) {
+  int64_t error;
+  // Use the existing sse codes for calculating distortion of decoded signal:
+  // i.e. (orig - decoded)^2
+  error = av1_block_error_fp(coeff, dqcoeff, block_size);
+  // prediction residue^2 = (orig - ref)^2
+  *ssz = av1_block_error_fp(coeff, ref, block_size);
+  return error;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_PVQ
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+/* The trailing '0' is a terminator which is used inside av1_cost_coeffs() to
+ * decide whether to include cost of a trailing EOB node or not (i.e. we
+ * can skip this if the last coefficient in this transform block, e.g. the
+ * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
+ * were non-zero). */
+#if !CONFIG_LV_MAP
+static int cost_coeffs(const AV1_COMMON *const cm, MACROBLOCK *x, int plane,
+                       int block, TX_SIZE tx_size, const SCAN_ORDER *scan_order,
+                       const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
+                       int use_fast_coef_costing) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const struct macroblock_plane *p = &x->plane[plane];
+  const struct macroblockd_plane *pd = &xd->plane[plane];
+  const PLANE_TYPE type = pd->plane_type;
+  const uint16_t *band_count = &band_count_table[tx_size][1];
+  const int eob = p->eobs[block];
+  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+  const int tx_size_ctx = txsize_sqr_map[tx_size];
+  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
+      x->token_costs[tx_size_ctx][type][is_inter_block(mbmi)];
+  uint8_t token_cache[MAX_TX_SQUARE];
+  int pt = combine_entropy_contexts(*a, *l);
+  int c, cost;
+  const int16_t *scan = scan_order->scan;
+  const int16_t *nb = scan_order->neighbors;
+#if CONFIG_NEW_TOKENSET
+  const int ref = is_inter_block(mbmi);
+  aom_prob *blockz_probs =
+      cm->fc->blockzero_probs[txsize_sqr_map[tx_size]][type][ref];
+
+#endif  // CONFIG_NEW_TOKENSET
+
+#if CONFIG_HIGHBITDEPTH
+  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
+#else
+  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, 8);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if !CONFIG_VAR_TX && !CONFIG_SUPERTX
+  // Check for consistency of tx_size with mode info
+  assert(tx_size == get_tx_size(plane, xd));
+#endif  // !CONFIG_VAR_TX && !CONFIG_SUPERTX
+  (void)cm;
+
+  if (eob == 0) {
+#if CONFIG_NEW_TOKENSET
+    // single eob token
+    cost = av1_cost_bit(blockz_probs[pt], 0);
+#else
+    cost = token_costs[0][0][pt][EOB_TOKEN];
+#endif  // CONFIG_NEW_TOKENSET
+  } else {
+    if (use_fast_coef_costing) {
+      int band_left = *band_count++;
+
+      // dc token
+      int v = qcoeff[0];
+      int16_t prev_t;
+      cost = av1_get_token_cost(v, &prev_t, cat6_bits);
+#if CONFIG_NEW_TOKENSET
+      cost += (*token_costs)[!prev_t][pt][prev_t];
+#else
+      cost += (*token_costs)[0][pt][prev_t];
+#endif
+
+      token_cache[0] = av1_pt_energy_class[prev_t];
+      ++token_costs;
+
+      // ac tokens
+      for (c = 1; c < eob; c++) {
+        const int rc = scan[c];
+        int16_t t;
+
+        v = qcoeff[rc];
+        cost += av1_get_token_cost(v, &t, cat6_bits);
+#if CONFIG_NEW_TOKENSET
+        cost += (*token_costs)[!t][!prev_t][t];
+#else
+        cost += (*token_costs)[!prev_t][!prev_t][t];
+#endif
+        prev_t = t;
+        if (!--band_left) {
+          band_left = *band_count++;
+          ++token_costs;
+        }
+      }
+
+      // eob token
+      if (band_left || CONFIG_NEW_TOKENSET)
+        cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
+
+    } else {  // !use_fast_coef_costing
+      int band_left = *band_count++;
+
+      // dc token
+      int v = qcoeff[0];
+      int16_t tok;
+#if !CONFIG_NEW_TOKENSET
+      unsigned int(*tok_cost_ptr)[COEFF_CONTEXTS][ENTROPY_TOKENS];
+#endif
+      cost = av1_get_token_cost(v, &tok, cat6_bits);
+#if CONFIG_NEW_TOKENSET
+      cost += (*token_costs)[!tok][pt][tok];
+#else
+      cost += (*token_costs)[0][pt][tok];
+#endif
+
+      token_cache[0] = av1_pt_energy_class[tok];
+      ++token_costs;
+
+#if !CONFIG_NEW_TOKENSET
+      tok_cost_ptr = &((*token_costs)[!tok]);
+#endif
+
+      // ac tokens
+      for (c = 1; c < eob; c++) {
+        const int rc = scan[c];
+
+        v = qcoeff[rc];
+        cost += av1_get_token_cost(v, &tok, cat6_bits);
+        pt = get_coef_context(nb, token_cache, c);
+#if CONFIG_NEW_TOKENSET
+        cost += (*token_costs)[!tok][pt][tok];
+#else
+        cost += (*tok_cost_ptr)[pt][tok];
+#endif
+        token_cache[rc] = av1_pt_energy_class[tok];
+        if (!--band_left) {
+          band_left = *band_count++;
+          ++token_costs;
+        }
+#if !CONFIG_NEW_TOKENSET
+        tok_cost_ptr = &((*token_costs)[!tok]);
+#endif
+      }
+
+      // eob token
+      if (band_left || CONFIG_NEW_TOKENSET) {
+        pt = get_coef_context(nb, token_cache, c);
+        cost += (*token_costs)[0][pt][EOB_TOKEN];
+      }
+    }
+  }
+
+  return cost;
+}
+#endif  // !CONFIG_LV_MAP
+
+int av1_cost_coeffs(const AV1_COMP *const cpi, MACROBLOCK *x, int plane,
+                    int block, TX_SIZE tx_size, const SCAN_ORDER *scan_order,
+                    const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
+                    int use_fast_coef_costing) {
+#if !CONFIG_LV_MAP
+  const AV1_COMMON *const cm = &cpi->common;
+  return cost_coeffs(cm, x, plane, block, tx_size, scan_order, a, l,
+                     use_fast_coef_costing);
+#else  // !CONFIG_LV_MAP
+  (void)scan_order;
+  (void)use_fast_coef_costing;
+  const MACROBLOCKD *xd = &x->e_mbd;
+  const MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const struct macroblockd_plane *pd = &xd->plane[plane];
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#else
+  const BLOCK_SIZE plane_bsize =
+      AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#endif  // CONFIG_CHROMA_2X2
+#else   // CONFIG_CB4X4
+  const BLOCK_SIZE plane_bsize =
+      get_plane_block_size(AOMMAX(BLOCK_8X8, bsize), pd);
+#endif  // CONFIG_CB4X4
+
+  TXB_CTX txb_ctx;
+  get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx);
+  return av1_cost_coeffs_txb(cpi, x, plane, block, &txb_ctx);
+#endif  // !CONFIG_LV_MAP
+}
+#endif  // !CONFIG_PVQ || CONFIG_VAR_TX
+
+// Get transform block visible dimensions cropped to the MI units.
+static void get_txb_dimensions(const MACROBLOCKD *xd, int plane,
+                               BLOCK_SIZE plane_bsize, int blk_row, int blk_col,
+                               BLOCK_SIZE tx_bsize, int *width, int *height,
+                               int *visible_width, int *visible_height) {
+  assert(tx_bsize <= plane_bsize);
+  int txb_height = block_size_high[tx_bsize];
+  int txb_width = block_size_wide[tx_bsize];
+  const int block_height = block_size_high[plane_bsize];
+  const int block_width = block_size_wide[plane_bsize];
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  // TODO(aconverse@google.com): Investigate using crop_width/height here rather
+  // than the MI size
+  const int block_rows =
+      (xd->mb_to_bottom_edge >= 0)
+          ? block_height
+          : (xd->mb_to_bottom_edge >> (3 + pd->subsampling_y)) + block_height;
+  const int block_cols =
+      (xd->mb_to_right_edge >= 0)
+          ? block_width
+          : (xd->mb_to_right_edge >> (3 + pd->subsampling_x)) + block_width;
+  const int tx_unit_size = tx_size_wide_log2[0];
+  if (width) *width = txb_width;
+  if (height) *height = txb_height;
+  *visible_width = clamp(block_cols - (blk_col << tx_unit_size), 0, txb_width);
+  *visible_height =
+      clamp(block_rows - (blk_row << tx_unit_size), 0, txb_height);
+}
+
+// Compute the pixel domain sum square error on all visible 4x4s in the
+// transform block.
+static unsigned pixel_sse(const AV1_COMP *const cpi, const MACROBLOCKD *xd,
+                          int plane, const uint8_t *src, const int src_stride,
+                          const uint8_t *dst, const int dst_stride, int blk_row,
+                          int blk_col, const BLOCK_SIZE plane_bsize,
+                          const BLOCK_SIZE tx_bsize) {
+  int txb_rows, txb_cols, visible_rows, visible_cols;
+  get_txb_dimensions(xd, plane, plane_bsize, blk_row, blk_col, tx_bsize,
+                     &txb_cols, &txb_rows, &visible_cols, &visible_rows);
+  assert(visible_rows > 0);
+  assert(visible_cols > 0);
+  if (txb_rows == visible_rows && txb_cols == visible_cols) {
+    unsigned sse;
+    cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &sse);
+    return sse;
+  }
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    uint64_t sse = aom_highbd_sse_odd_size(src, src_stride, dst, dst_stride,
+                                           visible_cols, visible_rows);
+    return (unsigned int)ROUND_POWER_OF_TWO(sse, (xd->bd - 8) * 2);
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  unsigned sse = aom_sse_odd_size(src, src_stride, dst, dst_stride,
+                                  visible_cols, visible_rows);
+  return sse;
+}
+
+// Compute the squares sum squares on all visible 4x4s in the transform block.
+static int64_t sum_squares_visible(const MACROBLOCKD *xd, int plane,
+                                   const int16_t *diff, const int diff_stride,
+                                   int blk_row, int blk_col,
+                                   const BLOCK_SIZE plane_bsize,
+                                   const BLOCK_SIZE tx_bsize) {
+  int visible_rows, visible_cols;
+  get_txb_dimensions(xd, plane, plane_bsize, blk_row, blk_col, tx_bsize, NULL,
+                     NULL, &visible_cols, &visible_rows);
+  return aom_sum_squares_2d_i16(diff, diff_stride, visible_cols, visible_rows);
+}
+
+void av1_dist_block(const AV1_COMP *cpi, MACROBLOCK *x, int plane,
+                    BLOCK_SIZE plane_bsize, int block, int blk_row, int blk_col,
+                    TX_SIZE tx_size, int64_t *out_dist, int64_t *out_sse,
+                    OUTPUT_STATUS output_status) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const struct macroblock_plane *const p = &x->plane[plane];
+#if CONFIG_DAALA_DIST
+  int qm = OD_HVS_QM;
+  int use_activity_masking = 0;
+#if CONFIG_PVQ
+  use_activity_masking = x->daala_enc.use_activity_masking;
+#endif  // CONFIG_PVQ
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+#else   // CONFIG_DAALA_DIST
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+#endif  // CONFIG_DAALA_DIST
+
+  if (cpi->sf.use_transform_domain_distortion && !CONFIG_DAALA_DIST) {
+    // Transform domain distortion computation is more efficient as it does
+    // not involve an inverse transform, but it is less accurate.
+    const int buffer_length = tx_size_2d[tx_size];
+    int64_t this_sse;
+    int shift = (MAX_TX_SCALE - av1_get_tx_scale(tx_size)) * 2;
+    tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
+    tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+#if CONFIG_PVQ
+    tran_low_t *ref_coeff = BLOCK_OFFSET(pd->pvq_ref_coeff, block);
+
+#if CONFIG_HIGHBITDEPTH
+    const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
+    *out_dist = av1_highbd_block_error2_c(coeff, dqcoeff, ref_coeff,
+                                          buffer_length, &this_sse, bd) >>
+                shift;
+#else
+    *out_dist = av1_block_error2_c(coeff, dqcoeff, ref_coeff, buffer_length,
+                                   &this_sse) >>
+                shift;
+#endif  // CONFIG_HIGHBITDEPTH
+#elif CONFIG_HIGHBITDEPTH
+    const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
+    *out_dist =
+        av1_highbd_block_error(coeff, dqcoeff, buffer_length, &this_sse, bd) >>
+        shift;
+#else
+    *out_dist =
+        av1_block_error(coeff, dqcoeff, buffer_length, &this_sse) >> shift;
+#endif  // CONFIG_PVQ
+    *out_sse = this_sse >> shift;
+  } else {
+    const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
+#if !CONFIG_PVQ || CONFIG_DAALA_DIST
+    const int bsw = block_size_wide[tx_bsize];
+    const int bsh = block_size_high[tx_bsize];
+#endif
+    const int src_stride = x->plane[plane].src.stride;
+    const int dst_stride = xd->plane[plane].dst.stride;
+    // Scale the transform block index to pixel unit.
+    const int src_idx = (blk_row * src_stride + blk_col)
+                        << tx_size_wide_log2[0];
+    const int dst_idx = (blk_row * dst_stride + blk_col)
+                        << tx_size_wide_log2[0];
+    const uint8_t *src = &x->plane[plane].src.buf[src_idx];
+    const uint8_t *dst = &xd->plane[plane].dst.buf[dst_idx];
+    const tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+    const uint16_t eob = p->eobs[block];
+
+    assert(cpi != NULL);
+    assert(tx_size_wide_log2[0] == tx_size_high_log2[0]);
+
+#if CONFIG_DAALA_DIST
+    if (plane == 0 && bsw >= 8 && bsh >= 8) {
+      if (output_status == OUTPUT_HAS_DECODED_PIXELS) {
+        const int pred_stride = block_size_wide[plane_bsize];
+        const int pred_idx = (blk_row * pred_stride + blk_col)
+                             << tx_size_wide_log2[0];
+        const int16_t *pred = &pd->pred[pred_idx];
+        int i, j;
+        DECLARE_ALIGNED(16, uint8_t, pred8[MAX_TX_SQUARE]);
+
+        for (j = 0; j < bsh; j++)
+          for (i = 0; i < bsw; i++)
+            pred8[j * bsw + i] = pred[j * pred_stride + i];
+        *out_sse = av1_daala_dist(src, src_stride, pred8, bsw, bsw, bsh, qm,
+                                  use_activity_masking, x->qindex);
+      } else {
+        *out_sse = av1_daala_dist(src, src_stride, dst, dst_stride, bsw, bsh,
+                                  qm, use_activity_masking, x->qindex);
+      }
+    } else
+#endif  // CONFIG_DAALA_DIST
+    {
+      const int diff_stride = block_size_wide[plane_bsize];
+      const int diff_idx = (blk_row * diff_stride + blk_col)
+                           << tx_size_wide_log2[0];
+      const int16_t *diff = &p->src_diff[diff_idx];
+      *out_sse = sum_squares_visible(xd, plane, diff, diff_stride, blk_row,
+                                     blk_col, plane_bsize, tx_bsize);
+#if CONFIG_HIGHBITDEPTH
+      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+        *out_sse = ROUND_POWER_OF_TWO(*out_sse, (xd->bd - 8) * 2);
+#endif  // CONFIG_HIGHBITDEPTH
+    }
+    *out_sse *= 16;
+
+    if (eob) {
+      if (output_status == OUTPUT_HAS_DECODED_PIXELS) {
+#if CONFIG_DAALA_DIST
+        if (plane == 0 && bsw >= 8 && bsh >= 8)
+          *out_dist = av1_daala_dist(src, src_stride, dst, dst_stride, bsw, bsh,
+                                     qm, use_activity_masking, x->qindex);
+        else
+#endif  // CONFIG_DAALA_DIST
+          *out_dist =
+              pixel_sse(cpi, xd, plane, src, src_stride, dst, dst_stride,
+                        blk_row, blk_col, plane_bsize, tx_bsize);
+      } else {
+#if CONFIG_HIGHBITDEPTH
+        uint8_t *recon;
+        DECLARE_ALIGNED(16, uint16_t, recon16[MAX_TX_SQUARE]);
+
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+          recon = CONVERT_TO_BYTEPTR(recon16);
+        else
+          recon = (uint8_t *)recon16;
+#else
+        DECLARE_ALIGNED(16, uint8_t, recon[MAX_TX_SQUARE]);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if !CONFIG_PVQ
+#if CONFIG_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          aom_highbd_convolve_copy(dst, dst_stride, recon, MAX_TX_SIZE, NULL, 0,
+                                   NULL, 0, bsw, bsh, xd->bd);
+        } else {
+#endif  // CONFIG_HIGHBITDEPTH
+          aom_convolve_copy(dst, dst_stride, recon, MAX_TX_SIZE, NULL, 0, NULL,
+                            0, bsw, bsh);
+#if CONFIG_HIGHBITDEPTH
+        }
+#endif  // CONFIG_HIGHBITDEPTH
+#else
+        (void)dst;
+#endif  // !CONFIG_PVQ
+
+        const PLANE_TYPE plane_type = get_plane_type(plane);
+        TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+
+        av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, recon,
+                                    MAX_TX_SIZE, eob);
+
+#if CONFIG_DAALA_DIST
+        if (plane == 0 && bsw >= 8 && bsh >= 8) {
+          *out_dist = av1_daala_dist(src, src_stride, recon, MAX_TX_SIZE, bsw,
+                                     bsh, qm, use_activity_masking, x->qindex);
+        } else {
+          if (plane == 0) {
+            // Save decoded pixels for inter block in pd->pred to avoid
+            // block_8x8_rd_txfm_daala_dist() need to produce them
+            // by calling av1_inverse_transform_block() again.
+            const int pred_stride = block_size_wide[plane_bsize];
+            const int pred_idx = (blk_row * pred_stride + blk_col)
+                                 << tx_size_wide_log2[0];
+            int16_t *pred = &pd->pred[pred_idx];
+            int i, j;
+
+            for (j = 0; j < bsh; j++)
+              for (i = 0; i < bsw; i++)
+                pred[j * pred_stride + i] = recon[j * MAX_TX_SIZE + i];
+          }
+#endif  // CONFIG_DAALA_DIST
+          *out_dist =
+              pixel_sse(cpi, xd, plane, src, src_stride, recon, MAX_TX_SIZE,
+                        blk_row, blk_col, plane_bsize, tx_bsize);
+#if CONFIG_DAALA_DIST
+        }
+#endif  // CONFIG_DAALA_DIST
+      }
+      *out_dist *= 16;
+    } else {
+      *out_dist = *out_sse;
+    }
+  }
+}
+
+static void block_rd_txfm(int plane, int block, int blk_row, int blk_col,
+                          BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
+  struct rdcost_block_args *args = arg;
+  MACROBLOCK *const x = args->x;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const AV1_COMP *cpi = args->cpi;
+  ENTROPY_CONTEXT *a = args->t_above + blk_col;
+  ENTROPY_CONTEXT *l = args->t_left + blk_row;
+#if !CONFIG_TXK_SEL
+  const AV1_COMMON *cm = &cpi->common;
+#endif
+  int64_t rd1, rd2, rd;
+  RD_STATS this_rd_stats;
+
+  assert(tx_size == get_tx_size(plane, xd));
+
+  av1_init_rd_stats(&this_rd_stats);
+
+  if (args->exit_early) return;
+
+  if (!is_inter_block(mbmi)) {
+    av1_predict_intra_block_facade(xd, plane, block, blk_col, blk_row, tx_size);
+    av1_subtract_txb(x, plane, plane_bsize, blk_col, blk_row, tx_size);
+  }
+
+#if !CONFIG_TXK_SEL
+  // full forward transform and quantization
+  const int coeff_ctx = combine_entropy_contexts(*a, *l);
+  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                  coeff_ctx, AV1_XFORM_QUANT_FP);
+  if (x->plane[plane].eobs[block] && !xd->lossless[mbmi->segment_id])
+    av1_optimize_b(cm, x, plane, block, tx_size, coeff_ctx);
+
+  if (!is_inter_block(mbmi)) {
+    struct macroblock_plane *const p = &x->plane[plane];
+    av1_inverse_transform_block_facade(xd, plane, block, blk_row, blk_col,
+                                       p->eobs[block]);
+    av1_dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
+                   tx_size, &this_rd_stats.dist, &this_rd_stats.sse,
+                   OUTPUT_HAS_DECODED_PIXELS);
+  } else {
+    av1_dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
+                   tx_size, &this_rd_stats.dist, &this_rd_stats.sse,
+                   OUTPUT_HAS_PREDICTED_PIXELS);
+  }
+#if CONFIG_CFL
+  if (plane == AOM_PLANE_Y && x->cfl_store_y) {
+    struct macroblockd_plane *const pd = &xd->plane[plane];
+    const int dst_stride = pd->dst.stride;
+    uint8_t *dst =
+        &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
+    cfl_store(xd->cfl, dst, dst_stride, blk_row, blk_col, tx_size);
+  }
+#endif
+  rd = RDCOST(x->rdmult, x->rddiv, 0, this_rd_stats.dist);
+  if (args->this_rd + rd > args->best_rd) {
+    args->exit_early = 1;
+    return;
+  }
+#if !CONFIG_PVQ
+  const PLANE_TYPE plane_type = get_plane_type(plane);
+  const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  this_rd_stats.rate =
+      av1_cost_coeffs(cpi, x, plane, block, tx_size, scan_order, a, l,
+                      args->use_fast_coef_costing);
+#else   // !CONFIG_PVQ
+  this_rd_stats.rate = x->rate;
+#endif  // !CONFIG_PVQ
+#else   // !CONFIG_TXK_SEL
+  av1_search_txk_type(cpi, x, plane, block, blk_row, blk_col, plane_bsize,
+                      tx_size, a, l, args->use_fast_coef_costing,
+                      &this_rd_stats);
+#endif  // !CONFIG_TXK_SEL
+
+#if !CONFIG_PVQ
+#if CONFIG_RD_DEBUG
+  av1_update_txb_coeff_cost(&this_rd_stats, plane, tx_size, blk_row, blk_col,
+                            this_rd_stats.rate);
+#endif  // CONFIG_RD_DEBUG
+  av1_set_txb_context(x, plane, block, tx_size, a, l);
+#endif  // !CONFIG_PVQ
+
+  rd1 = RDCOST(x->rdmult, x->rddiv, this_rd_stats.rate, this_rd_stats.dist);
+  rd2 = RDCOST(x->rdmult, x->rddiv, 0, this_rd_stats.sse);
+
+  // TODO(jingning): temporarily enabled only for luma component
+  rd = AOMMIN(rd1, rd2);
+
+#if CONFIG_DAALA_DIST
+  if (plane == 0 &&
+      (tx_size == TX_4X4 || tx_size == TX_4X8 || tx_size == TX_8X4)) {
+    this_rd_stats.dist = 0;
+    this_rd_stats.sse = 0;
+    rd = 0;
+    x->rate_4x4[block] = this_rd_stats.rate;
+  }
+#endif  // CONFIG_DAALA_DIST
+
+#if !CONFIG_PVQ
+  this_rd_stats.skip &= !x->plane[plane].eobs[block];
+#else
+  this_rd_stats.skip &= x->pvq_skip[plane];
+#endif  // !CONFIG_PVQ
+  av1_merge_rd_stats(&args->rd_stats, &this_rd_stats);
+
+  args->this_rd += rd;
+
+  if (args->this_rd > args->best_rd) {
+    args->exit_early = 1;
+    return;
+  }
+}
+
+#if CONFIG_DAALA_DIST
+static void block_8x8_rd_txfm_daala_dist(int plane, int block, int blk_row,
+                                         int blk_col, BLOCK_SIZE plane_bsize,
+                                         TX_SIZE tx_size, void *arg) {
+  struct rdcost_block_args *args = arg;
+  MACROBLOCK *const x = args->x;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int64_t rd, rd1, rd2;
+  RD_STATS this_rd_stats;
+  int qm = OD_HVS_QM;
+  int use_activity_masking = 0;
+
+  (void)tx_size;
+#if CONFIG_PVQ
+  use_activity_masking = x->daala_enc.use_activity_masking;
+#endif  // CONFIG_PVQ
+  av1_init_rd_stats(&this_rd_stats);
+
+  if (args->exit_early) return;
+
+  {
+    const struct macroblock_plane *const p = &x->plane[plane];
+    struct macroblockd_plane *const pd = &xd->plane[plane];
+
+    const int src_stride = p->src.stride;
+    const int dst_stride = pd->dst.stride;
+    const int diff_stride = block_size_wide[plane_bsize];
+
+    const uint8_t *src =
+        &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
+    const uint8_t *dst =
+        &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
+
+    unsigned int tmp1, tmp2;
+    int qindex = x->qindex;
+    const int pred_stride = block_size_wide[plane_bsize];
+    const int pred_idx = (blk_row * pred_stride + blk_col)
+                         << tx_size_wide_log2[0];
+    int16_t *pred = &pd->pred[pred_idx];
+    int i, j;
+    const int tx_blk_size = 8;
+
+    DECLARE_ALIGNED(16, uint8_t, pred8[8 * 8]);
+
+    for (j = 0; j < tx_blk_size; j++)
+      for (i = 0; i < tx_blk_size; i++)
+        pred8[j * tx_blk_size + i] = pred[j * diff_stride + i];
+
+    tmp1 = av1_daala_dist(src, src_stride, pred8, tx_blk_size, 8, 8, qm,
+                          use_activity_masking, qindex);
+    tmp2 = av1_daala_dist(src, src_stride, dst, dst_stride, 8, 8, qm,
+                          use_activity_masking, qindex);
+
+    if (!is_inter_block(mbmi)) {
+      this_rd_stats.sse = (int64_t)tmp1 * 16;
+      this_rd_stats.dist = (int64_t)tmp2 * 16;
+    } else {
+      // For inter mode, the decoded pixels are provided in pd->pred,
+      // while the predicted pixels are in dst.
+      this_rd_stats.sse = (int64_t)tmp2 * 16;
+      this_rd_stats.dist = (int64_t)tmp1 * 16;
+    }
+  }
+
+  rd = RDCOST(x->rdmult, x->rddiv, 0, this_rd_stats.dist);
+  if (args->this_rd + rd > args->best_rd) {
+    args->exit_early = 1;
+    return;
+  }
+
+  {
+    const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+    // The rate of the current 8x8 block is the sum of four 4x4 blocks in it.
+    this_rd_stats.rate = x->rate_4x4[block - max_blocks_wide - 1] +
+                         x->rate_4x4[block - max_blocks_wide] +
+                         x->rate_4x4[block - 1] + x->rate_4x4[block];
+  }
+  rd1 = RDCOST(x->rdmult, x->rddiv, this_rd_stats.rate, this_rd_stats.dist);
+  rd2 = RDCOST(x->rdmult, x->rddiv, 0, this_rd_stats.sse);
+  rd = AOMMIN(rd1, rd2);
+
+  args->rd_stats.dist += this_rd_stats.dist;
+  args->rd_stats.sse += this_rd_stats.sse;
+
+  args->this_rd += rd;
+
+  if (args->this_rd > args->best_rd) {
+    args->exit_early = 1;
+    return;
+  }
+}
+#endif  // CONFIG_DAALA_DIST
+
+static void txfm_rd_in_plane(MACROBLOCK *x, const AV1_COMP *cpi,
+                             RD_STATS *rd_stats, int64_t ref_best_rd, int plane,
+                             BLOCK_SIZE bsize, TX_SIZE tx_size,
+                             int use_fast_coef_casting) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  struct rdcost_block_args args;
+  av1_zero(args);
+  args.x = x;
+  args.cpi = cpi;
+  args.best_rd = ref_best_rd;
+  args.use_fast_coef_costing = use_fast_coef_casting;
+  av1_init_rd_stats(&args.rd_stats);
+
+  if (plane == 0) xd->mi[0]->mbmi.tx_size = tx_size;
+
+  av1_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
+
+#if CONFIG_DAALA_DIST
+  if (plane == 0 &&
+      (tx_size == TX_4X4 || tx_size == TX_4X8 || tx_size == TX_8X4))
+    av1_foreach_8x8_transformed_block_in_plane(
+        xd, bsize, plane, block_rd_txfm, block_8x8_rd_txfm_daala_dist, &args);
+  else
+#endif  // CONFIG_DAALA_DIST
+    av1_foreach_transformed_block_in_plane(xd, bsize, plane, block_rd_txfm,
+                                           &args);
+
+  if (args.exit_early) {
+    av1_invalid_rd_stats(rd_stats);
+  } else {
+    *rd_stats = args.rd_stats;
+  }
+}
+
+#if CONFIG_SUPERTX
+void av1_txfm_rd_in_plane_supertx(MACROBLOCK *x, const AV1_COMP *cpi, int *rate,
+                                  int64_t *distortion, int *skippable,
+                                  int64_t *sse, int64_t ref_best_rd, int plane,
+                                  BLOCK_SIZE bsize, TX_SIZE tx_size,
+                                  int use_fast_coef_casting) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  struct rdcost_block_args args;
+  av1_zero(args);
+  args.cpi = cpi;
+  args.x = x;
+  args.best_rd = ref_best_rd;
+  args.use_fast_coef_costing = use_fast_coef_casting;
+
+#if CONFIG_EXT_TX
+  assert(tx_size < TX_SIZES);
+#endif  // CONFIG_EXT_TX
+
+  if (plane == 0) xd->mi[0]->mbmi.tx_size = tx_size;
+
+  av1_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
+
+  block_rd_txfm(plane, 0, 0, 0, get_plane_block_size(bsize, pd), tx_size,
+                &args);
+
+  if (args.exit_early) {
+    *rate = INT_MAX;
+    *distortion = INT64_MAX;
+    *sse = INT64_MAX;
+    *skippable = 0;
+  } else {
+    *distortion = args.rd_stats.dist;
+    *rate = args.rd_stats.rate;
+    *sse = args.rd_stats.sse;
+    *skippable = !x->plane[plane].eobs[0];
+  }
+}
+#endif  // CONFIG_SUPERTX
+
+static int tx_size_cost(const AV1_COMP *const cpi, const MACROBLOCK *const x,
+                        BLOCK_SIZE bsize, TX_SIZE tx_size) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+
+  const int tx_select =
+      cm->tx_mode == TX_MODE_SELECT && mbmi->sb_type >= BLOCK_8X8;
+
+  if (tx_select) {
+    const int is_inter = is_inter_block(mbmi);
+    const int tx_size_cat = is_inter ? inter_tx_size_cat_lookup[bsize]
+                                     : intra_tx_size_cat_lookup[bsize];
+    const TX_SIZE coded_tx_size = txsize_sqr_up_map[tx_size];
+    const int depth = tx_size_to_depth(coded_tx_size);
+    const int tx_size_ctx = get_tx_size_context(xd);
+    const int r_tx_size = cpi->tx_size_cost[tx_size_cat][tx_size_ctx][depth];
+    return r_tx_size;
+  } else {
+    return 0;
+  }
+}
+
+// #TODO(angiebird): use this function whenever it's possible
+int av1_tx_type_cost(const AV1_COMP *cpi, const MACROBLOCKD *xd,
+                     BLOCK_SIZE bsize, int plane, TX_SIZE tx_size,
+                     TX_TYPE tx_type) {
+  if (plane > 0) return 0;
+
+  const MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const int is_inter = is_inter_block(mbmi);
+#if CONFIG_EXT_TX
+  const AV1_COMMON *cm = &cpi->common;
+  if (get_ext_tx_types(tx_size, bsize, is_inter, cm->reduced_tx_set_used) > 1 &&
+      !xd->lossless[xd->mi[0]->mbmi.segment_id]) {
+    const int ext_tx_set =
+        get_ext_tx_set(tx_size, bsize, is_inter, cm->reduced_tx_set_used);
+    if (is_inter) {
+      if (ext_tx_set > 0)
+        return cpi
+            ->inter_tx_type_costs[ext_tx_set][txsize_sqr_map[tx_size]][tx_type];
+    } else {
+      if (ext_tx_set > 0 && ALLOW_INTRA_EXT_TX)
+        return cpi->intra_tx_type_costs[ext_tx_set][txsize_sqr_map[tx_size]]
+                                       [mbmi->mode][tx_type];
+    }
+  }
+#else
+  (void)bsize;
+  if (tx_size < TX_32X32 && !xd->lossless[xd->mi[0]->mbmi.segment_id] &&
+      !FIXED_TX_TYPE) {
+    if (is_inter) {
+      return cpi->inter_tx_type_costs[tx_size][tx_type];
+    } else {
+      return cpi->intra_tx_type_costs[tx_size]
+                                     [intra_mode_to_tx_type_context[mbmi->mode]]
+                                     [tx_type];
+    }
+  }
+#endif  // CONFIG_EXT_TX
+  return 0;
+}
+static int64_t txfm_yrd(const AV1_COMP *const cpi, MACROBLOCK *x,
+                        RD_STATS *rd_stats, int64_t ref_best_rd, BLOCK_SIZE bs,
+                        TX_TYPE tx_type, int tx_size) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int64_t rd = INT64_MAX;
+  aom_prob skip_prob = av1_get_skip_prob(cm, xd);
+  int s0, s1;
+  const int is_inter = is_inter_block(mbmi);
+  const int tx_select =
+      cm->tx_mode == TX_MODE_SELECT && mbmi->sb_type >= BLOCK_8X8;
+
+  const int r_tx_size = tx_size_cost(cpi, x, bs, tx_size);
+
+  assert(skip_prob > 0);
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+  assert(IMPLIES(is_rect_tx(tx_size), is_rect_tx_allowed_bsize(bs)));
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+
+  s0 = av1_cost_bit(skip_prob, 0);
+  s1 = av1_cost_bit(skip_prob, 1);
+
+  mbmi->tx_type = tx_type;
+  mbmi->tx_size = tx_size;
+  txfm_rd_in_plane(x, cpi, rd_stats, ref_best_rd, 0, bs, tx_size,
+                   cpi->sf.use_fast_coef_costing);
+  if (rd_stats->rate == INT_MAX) return INT64_MAX;
+#if !CONFIG_TXK_SEL
+  int plane = 0;
+  rd_stats->rate += av1_tx_type_cost(cpi, xd, bs, plane, tx_size, tx_type);
+#endif
+
+  if (rd_stats->skip) {
+    if (is_inter) {
+      rd = RDCOST(x->rdmult, x->rddiv, s1, rd_stats->sse);
+    } else {
+      rd = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size * tx_select,
+                  rd_stats->sse);
+    }
+  } else {
+    rd = RDCOST(x->rdmult, x->rddiv,
+                rd_stats->rate + s0 + r_tx_size * tx_select, rd_stats->dist);
+  }
+
+  if (tx_select) rd_stats->rate += r_tx_size;
+
+  if (is_inter && !xd->lossless[xd->mi[0]->mbmi.segment_id] &&
+      !(rd_stats->skip))
+    rd = AOMMIN(rd, RDCOST(x->rdmult, x->rddiv, s1, rd_stats->sse));
+
+  return rd;
+}
+
+static int skip_txfm_search(const AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bs,
+                            TX_TYPE tx_type, TX_SIZE tx_size) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const TX_SIZE max_tx_size = max_txsize_lookup[bs];
+  const int is_inter = is_inter_block(mbmi);
+  int prune = 0;
+  if (is_inter && cpi->sf.tx_type_search.prune_mode > NO_PRUNE)
+    // passing -1 in for tx_type indicates that all 1D
+    // transforms should be considered for pruning
+    prune = prune_tx_types(cpi, bs, x, xd, -1);
+
+#if CONFIG_REF_MV
+  if (mbmi->ref_mv_idx > 0 && tx_type != DCT_DCT) return 1;
+#endif  // CONFIG_REF_MV
+  if (FIXED_TX_TYPE && tx_type != get_default_tx_type(0, xd, 0, tx_size))
+    return 1;
+  if (!is_inter && x->use_default_intra_tx_type &&
+      tx_type != get_default_tx_type(0, xd, 0, tx_size))
+    return 1;
+  if (is_inter && x->use_default_inter_tx_type &&
+      tx_type != get_default_tx_type(0, xd, 0, tx_size))
+    return 1;
+  if (max_tx_size >= TX_32X32 && tx_size == TX_4X4) return 1;
+#if CONFIG_EXT_TX
+  const AV1_COMMON *const cm = &cpi->common;
+  int ext_tx_set =
+      get_ext_tx_set(tx_size, bs, is_inter, cm->reduced_tx_set_used);
+  if (is_inter) {
+    if (!ext_tx_used_inter[ext_tx_set][tx_type]) return 1;
+    if (cpi->sf.tx_type_search.prune_mode > NO_PRUNE) {
+      if (!do_tx_type_search(tx_type, prune)) return 1;
+    }
+  } else {
+    if (!ALLOW_INTRA_EXT_TX && bs >= BLOCK_8X8) {
+      if (tx_type != intra_mode_to_tx_type_context[mbmi->mode]) return 1;
+    }
+    if (!ext_tx_used_intra[ext_tx_set][tx_type]) return 1;
+  }
+#else   // CONFIG_EXT_TX
+  if (tx_size >= TX_32X32 && tx_type != DCT_DCT) return 1;
+  if (is_inter && cpi->sf.tx_type_search.prune_mode > NO_PRUNE &&
+      !do_tx_type_search(tx_type, prune))
+    return 1;
+#endif  // CONFIG_EXT_TX
+  return 0;
+}
+
+#if CONFIG_EXT_INTER
+static int64_t estimate_yrd_for_sb(const AV1_COMP *const cpi, BLOCK_SIZE bs,
+                                   MACROBLOCK *x, int *r, int64_t *d, int *s,
+                                   int64_t *sse, int64_t ref_best_rd) {
+  RD_STATS rd_stats;
+  int64_t rd = txfm_yrd(cpi, x, &rd_stats, ref_best_rd, bs, DCT_DCT,
+                        max_txsize_lookup[bs]);
+  *r = rd_stats.rate;
+  *d = rd_stats.dist;
+  *s = rd_stats.skip;
+  *sse = rd_stats.sse;
+  return rd;
+}
+#endif  // CONFIG_EXT_INTER
+
+static void choose_largest_tx_size(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                   RD_STATS *rd_stats, int64_t ref_best_rd,
+                                   BLOCK_SIZE bs) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  TX_TYPE tx_type, best_tx_type = DCT_DCT;
+  int64_t this_rd, best_rd = INT64_MAX;
+  aom_prob skip_prob = av1_get_skip_prob(cm, xd);
+  int s0 = av1_cost_bit(skip_prob, 0);
+  int s1 = av1_cost_bit(skip_prob, 1);
+  const int is_inter = is_inter_block(mbmi);
+  int prune = 0;
+  const int plane = 0;
+#if CONFIG_EXT_TX
+  int ext_tx_set;
+#endif  // CONFIG_EXT_TX
+  av1_invalid_rd_stats(rd_stats);
+
+  mbmi->tx_size = tx_size_from_tx_mode(bs, cm->tx_mode, is_inter);
+#if CONFIG_VAR_TX
+  mbmi->min_tx_size = get_min_tx_size(mbmi->tx_size);
+#endif  // CONFIG_VAR_TX
+#if CONFIG_EXT_TX
+  ext_tx_set =
+      get_ext_tx_set(mbmi->tx_size, bs, is_inter, cm->reduced_tx_set_used);
+#endif  // CONFIG_EXT_TX
+
+  if (is_inter && cpi->sf.tx_type_search.prune_mode > NO_PRUNE)
+#if CONFIG_EXT_TX
+    prune = prune_tx_types(cpi, bs, x, xd, ext_tx_set);
+#else
+    prune = prune_tx_types(cpi, bs, x, xd, 0);
+#endif  // CONFIG_EXT_TX
+#if CONFIG_EXT_TX
+  if (get_ext_tx_types(mbmi->tx_size, bs, is_inter, cm->reduced_tx_set_used) >
+          1 &&
+      !xd->lossless[mbmi->segment_id]) {
+#if CONFIG_PVQ
+    od_rollback_buffer pre_buf, post_buf;
+
+    od_encode_checkpoint(&x->daala_enc, &pre_buf);
+    od_encode_checkpoint(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+
+    for (tx_type = DCT_DCT; tx_type < TX_TYPES; ++tx_type) {
+      RD_STATS this_rd_stats;
+      if (is_inter) {
+        if (x->use_default_inter_tx_type &&
+            tx_type != get_default_tx_type(0, xd, 0, mbmi->tx_size))
+          continue;
+        if (!ext_tx_used_inter[ext_tx_set][tx_type]) continue;
+        if (cpi->sf.tx_type_search.prune_mode > NO_PRUNE) {
+          if (!do_tx_type_search(tx_type, prune)) continue;
+        }
+      } else {
+        if (x->use_default_intra_tx_type &&
+            tx_type != get_default_tx_type(0, xd, 0, mbmi->tx_size))
+          continue;
+        if (!ALLOW_INTRA_EXT_TX && bs >= BLOCK_8X8) {
+          if (tx_type != intra_mode_to_tx_type_context[mbmi->mode]) continue;
+        }
+        if (!ext_tx_used_intra[ext_tx_set][tx_type]) continue;
+      }
+
+      mbmi->tx_type = tx_type;
+
+      txfm_rd_in_plane(x, cpi, &this_rd_stats, ref_best_rd, 0, bs,
+                       mbmi->tx_size, cpi->sf.use_fast_coef_costing);
+#if CONFIG_PVQ
+      od_encode_rollback(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+      if (this_rd_stats.rate == INT_MAX) continue;
+      av1_tx_type_cost(cpi, xd, bs, plane, mbmi->tx_size, tx_type);
+
+      if (this_rd_stats.skip)
+        this_rd = RDCOST(x->rdmult, x->rddiv, s1, this_rd_stats.sse);
+      else
+        this_rd = RDCOST(x->rdmult, x->rddiv, this_rd_stats.rate + s0,
+                         this_rd_stats.dist);
+      if (is_inter_block(mbmi) && !xd->lossless[mbmi->segment_id] &&
+          !this_rd_stats.skip)
+        this_rd =
+            AOMMIN(this_rd, RDCOST(x->rdmult, x->rddiv, s1, this_rd_stats.sse));
+
+      if (this_rd < best_rd) {
+        best_rd = this_rd;
+        best_tx_type = mbmi->tx_type;
+        *rd_stats = this_rd_stats;
+#if CONFIG_PVQ
+        od_encode_checkpoint(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+      }
+    }
+#if CONFIG_PVQ
+    od_encode_rollback(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+  } else {
+    mbmi->tx_type = DCT_DCT;
+    txfm_rd_in_plane(x, cpi, rd_stats, ref_best_rd, 0, bs, mbmi->tx_size,
+                     cpi->sf.use_fast_coef_costing);
+  }
+#else   // CONFIG_EXT_TX
+  if (mbmi->tx_size < TX_32X32 && !xd->lossless[mbmi->segment_id]) {
+    for (tx_type = 0; tx_type < TX_TYPES; ++tx_type) {
+      RD_STATS this_rd_stats;
+      if (!is_inter && x->use_default_intra_tx_type &&
+          tx_type != get_default_tx_type(0, xd, 0, mbmi->tx_size))
+        continue;
+      if (is_inter && x->use_default_inter_tx_type &&
+          tx_type != get_default_tx_type(0, xd, 0, mbmi->tx_size))
+        continue;
+      mbmi->tx_type = tx_type;
+      txfm_rd_in_plane(x, cpi, &this_rd_stats, ref_best_rd, 0, bs,
+                       mbmi->tx_size, cpi->sf.use_fast_coef_costing);
+      if (this_rd_stats.rate == INT_MAX) continue;
+
+      av1_tx_type_cost(cpi, xd, bs, plane, mbmi->tx_size, tx_type);
+      if (is_inter) {
+        if (cpi->sf.tx_type_search.prune_mode > NO_PRUNE &&
+            !do_tx_type_search(tx_type, prune))
+          continue;
+      }
+      if (this_rd_stats.skip)
+        this_rd = RDCOST(x->rdmult, x->rddiv, s1, this_rd_stats.sse);
+      else
+        this_rd = RDCOST(x->rdmult, x->rddiv, this_rd_stats.rate + s0,
+                         this_rd_stats.dist);
+      if (is_inter && !xd->lossless[mbmi->segment_id] && !this_rd_stats.skip)
+        this_rd =
+            AOMMIN(this_rd, RDCOST(x->rdmult, x->rddiv, s1, this_rd_stats.sse));
+
+      if (this_rd < best_rd) {
+        best_rd = this_rd;
+        best_tx_type = mbmi->tx_type;
+        *rd_stats = this_rd_stats;
+      }
+    }
+  } else {
+    mbmi->tx_type = DCT_DCT;
+    txfm_rd_in_plane(x, cpi, rd_stats, ref_best_rd, 0, bs, mbmi->tx_size,
+                     cpi->sf.use_fast_coef_costing);
+  }
+#endif  // CONFIG_EXT_TX
+  mbmi->tx_type = best_tx_type;
+}
+
+static void choose_smallest_tx_size(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                    RD_STATS *rd_stats, int64_t ref_best_rd,
+                                    BLOCK_SIZE bs) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+
+  mbmi->tx_size = TX_4X4;
+  mbmi->tx_type = DCT_DCT;
+#if CONFIG_VAR_TX
+  mbmi->min_tx_size = get_min_tx_size(TX_4X4);
+#endif  // CONFIG_VAR_TX
+
+  txfm_rd_in_plane(x, cpi, rd_stats, ref_best_rd, 0, bs, mbmi->tx_size,
+                   cpi->sf.use_fast_coef_costing);
+}
+
+#if CONFIG_TXK_SEL || CONFIG_VAR_TX
+static INLINE int bsize_to_num_blk(BLOCK_SIZE bsize) {
+  int num_blk = 1 << (num_pels_log2_lookup[bsize] - 2 * tx_size_wide_log2[0]);
+  return num_blk;
+}
+#endif  // CONFIG_TXK_SEL || CONFIG_VAR_TX
+
+static void choose_tx_size_type_from_rd(const AV1_COMP *const cpi,
+                                        MACROBLOCK *x, RD_STATS *rd_stats,
+                                        int64_t ref_best_rd, BLOCK_SIZE bs) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int64_t rd = INT64_MAX;
+  int n;
+  int start_tx, end_tx;
+  int64_t best_rd = INT64_MAX, last_rd = INT64_MAX;
+  const TX_SIZE max_tx_size = max_txsize_lookup[bs];
+  TX_SIZE best_tx_size = max_tx_size;
+  TX_TYPE best_tx_type = DCT_DCT;
+#if CONFIG_TXK_SEL
+  TX_TYPE best_txk_type[MAX_SB_SQUARE / (TX_SIZE_W_MIN * TX_SIZE_H_MIN)];
+  const int num_blk = bsize_to_num_blk(bs);
+#endif  // CONFIG_TXK_SEL
+  const int tx_select = cm->tx_mode == TX_MODE_SELECT;
+  const int is_inter = is_inter_block(mbmi);
+#if CONFIG_PVQ
+  od_rollback_buffer buf;
+  od_encode_checkpoint(&x->daala_enc, &buf);
+#endif  // CONFIG_PVQ
+
+  av1_invalid_rd_stats(rd_stats);
+
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+  int evaluate_rect_tx = 0;
+  if (tx_select) {
+    evaluate_rect_tx = is_rect_tx_allowed(xd, mbmi);
+  } else {
+    const TX_SIZE chosen_tx_size =
+        tx_size_from_tx_mode(bs, cm->tx_mode, is_inter);
+    evaluate_rect_tx = is_rect_tx(chosen_tx_size);
+    assert(IMPLIES(evaluate_rect_tx, is_rect_tx_allowed(xd, mbmi)));
+  }
+  if (evaluate_rect_tx) {
+    TX_TYPE tx_start = DCT_DCT;
+    TX_TYPE tx_end = TX_TYPES;
+#if CONFIG_TXK_SEL
+    // The tx_type becomes dummy when lv_map is on. The tx_type search will be
+    // performed in av1_search_txk_type()
+    tx_end = DCT_DCT + 1;
+#endif
+    TX_TYPE tx_type;
+    for (tx_type = tx_start; tx_type < tx_end; ++tx_type) {
+#if CONFIG_REF_MV
+      if (mbmi->ref_mv_idx > 0 && tx_type != DCT_DCT) continue;
+#endif  // CONFIG_REF_MV
+      const TX_SIZE rect_tx_size = max_txsize_rect_lookup[bs];
+      RD_STATS this_rd_stats;
+      int ext_tx_set =
+          get_ext_tx_set(rect_tx_size, bs, is_inter, cm->reduced_tx_set_used);
+      if ((is_inter && ext_tx_used_inter[ext_tx_set][tx_type]) ||
+          (!is_inter && ext_tx_used_intra[ext_tx_set][tx_type])) {
+        rd = txfm_yrd(cpi, x, &this_rd_stats, ref_best_rd, bs, tx_type,
+                      rect_tx_size);
+        if (rd < best_rd) {
+#if CONFIG_TXK_SEL
+          memcpy(best_txk_type, mbmi->txk_type,
+                 sizeof(best_txk_type[0]) * num_blk);
+#endif
+          best_tx_type = tx_type;
+          best_tx_size = rect_tx_size;
+          best_rd = rd;
+          *rd_stats = this_rd_stats;
+        }
+      }
+#if CONFIG_CB4X4 && !USE_TXTYPE_SEARCH_FOR_SUB8X8_IN_CB4X4
+      const int is_inter = is_inter_block(mbmi);
+      if (mbmi->sb_type < BLOCK_8X8 && is_inter) break;
+#endif  // CONFIG_CB4X4 && !USE_TXTYPE_SEARCH_FOR_SUB8X8_IN_CB4X4
+    }
+  }
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+
+  if (tx_select) {
+    start_tx = max_tx_size;
+    end_tx = (max_tx_size >= TX_32X32) ? TX_8X8 : TX_4X4;
+  } else {
+    const TX_SIZE chosen_tx_size =
+        tx_size_from_tx_mode(bs, cm->tx_mode, is_inter);
+    start_tx = chosen_tx_size;
+    end_tx = chosen_tx_size;
+  }
+
+  last_rd = INT64_MAX;
+  for (n = start_tx; n >= end_tx; --n) {
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+    if (is_rect_tx(n)) break;
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+    TX_TYPE tx_start = DCT_DCT;
+    TX_TYPE tx_end = TX_TYPES;
+#if CONFIG_TXK_SEL
+    // The tx_type becomes dummy when lv_map is on. The tx_type search will be
+    // performed in av1_search_txk_type()
+    tx_end = DCT_DCT + 1;
+#endif
+    TX_TYPE tx_type;
+    for (tx_type = tx_start; tx_type < tx_end; ++tx_type) {
+      RD_STATS this_rd_stats;
+      if (skip_txfm_search(cpi, x, bs, tx_type, n)) continue;
+      rd = txfm_yrd(cpi, x, &this_rd_stats, ref_best_rd, bs, tx_type, n);
+#if CONFIG_PVQ
+      od_encode_rollback(&x->daala_enc, &buf);
+#endif  // CONFIG_PVQ
+      // Early termination in transform size search.
+      if (cpi->sf.tx_size_search_breakout &&
+          (rd == INT64_MAX ||
+           (this_rd_stats.skip == 1 && tx_type != DCT_DCT && n < start_tx) ||
+           (n < (int)max_tx_size && rd > last_rd)))
+        break;
+
+      last_rd = rd;
+      if (rd < best_rd) {
+#if CONFIG_TXK_SEL
+        memcpy(best_txk_type, mbmi->txk_type,
+               sizeof(best_txk_type[0]) * num_blk);
+#endif
+        best_tx_type = tx_type;
+        best_tx_size = n;
+        best_rd = rd;
+        *rd_stats = this_rd_stats;
+      }
+#if CONFIG_CB4X4 && !USE_TXTYPE_SEARCH_FOR_SUB8X8_IN_CB4X4
+      const int is_inter = is_inter_block(mbmi);
+      if (mbmi->sb_type < BLOCK_8X8 && is_inter) break;
+#endif  // CONFIG_CB4X4 && !USE_TXTYPE_SEARCH_FOR_SUB8X8_IN_CB4X4
+    }
+  }
+  mbmi->tx_size = best_tx_size;
+  mbmi->tx_type = best_tx_type;
+#if CONFIG_TXK_SEL
+  memcpy(mbmi->txk_type, best_txk_type, sizeof(best_txk_type[0]) * num_blk);
+#endif
+
+#if CONFIG_VAR_TX
+  mbmi->min_tx_size = get_min_tx_size(mbmi->tx_size);
+#endif  // CONFIG_VAR_TX
+
+#if !CONFIG_EXT_TX
+  if (mbmi->tx_size >= TX_32X32) assert(mbmi->tx_type == DCT_DCT);
+#endif  // !CONFIG_EXT_TX
+#if CONFIG_PVQ
+  if (best_rd != INT64_MAX) {
+    txfm_yrd(cpi, x, rd_stats, ref_best_rd, bs, best_tx_type, best_tx_size);
+  }
+#endif  // CONFIG_PVQ
+}
+
+static void super_block_yrd(const AV1_COMP *const cpi, MACROBLOCK *x,
+                            RD_STATS *rd_stats, BLOCK_SIZE bs,
+                            int64_t ref_best_rd) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  av1_init_rd_stats(rd_stats);
+
+  assert(bs == xd->mi[0]->mbmi.sb_type);
+
+  if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
+    choose_smallest_tx_size(cpi, x, rd_stats, ref_best_rd, bs);
+  } else if (cpi->sf.tx_size_search_method == USE_LARGESTALL) {
+    choose_largest_tx_size(cpi, x, rd_stats, ref_best_rd, bs);
+  } else {
+    choose_tx_size_type_from_rd(cpi, x, rd_stats, ref_best_rd, bs);
+  }
+}
+
+static int conditional_skipintra(PREDICTION_MODE mode,
+                                 PREDICTION_MODE best_intra_mode) {
+  if (mode == D117_PRED && best_intra_mode != V_PRED &&
+      best_intra_mode != D135_PRED)
+    return 1;
+  if (mode == D63_PRED && best_intra_mode != V_PRED &&
+      best_intra_mode != D45_PRED)
+    return 1;
+  if (mode == D207_PRED && best_intra_mode != H_PRED &&
+      best_intra_mode != D45_PRED)
+    return 1;
+  if (mode == D153_PRED && best_intra_mode != H_PRED &&
+      best_intra_mode != D135_PRED)
+    return 1;
+  return 0;
+}
+
+// Model based RD estimation for luma intra blocks.
+static int64_t intra_model_yrd(const AV1_COMP *const cpi, MACROBLOCK *const x,
+                               BLOCK_SIZE bsize, int mode_cost) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  RD_STATS this_rd_stats;
+  int row, col;
+  int64_t temp_sse, this_rd;
+  const TX_SIZE tx_size = tx_size_from_tx_mode(bsize, cpi->common.tx_mode, 0);
+  const int stepr = tx_size_high_unit[tx_size];
+  const int stepc = tx_size_wide_unit[tx_size];
+  const int max_blocks_wide = max_block_wide(xd, bsize, 0);
+  const int max_blocks_high = max_block_high(xd, bsize, 0);
+  mbmi->tx_size = tx_size;
+  // Prediction.
+  const int step = stepr * stepc;
+  int block = 0;
+  for (row = 0; row < max_blocks_high; row += stepr) {
+    for (col = 0; col < max_blocks_wide; col += stepc) {
+      av1_predict_intra_block_facade(xd, 0, block, col, row, tx_size);
+      block += step;
+    }
+  }
+  // RD estimation.
+  model_rd_for_sb(cpi, bsize, x, xd, 0, 0, &this_rd_stats.rate,
+                  &this_rd_stats.dist, &this_rd_stats.skip, &temp_sse);
+#if CONFIG_EXT_INTRA
+  if (av1_is_directional_mode(mbmi->mode, bsize)) {
+    mode_cost += write_uniform_cost(2 * MAX_ANGLE_DELTA + 1,
+                                    MAX_ANGLE_DELTA + mbmi->angle_delta[0]);
+  }
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  if (mbmi->mode == DC_PRED) {
+    const aom_prob prob = cpi->common.fc->filter_intra_probs[0];
+    if (mbmi->filter_intra_mode_info.use_filter_intra_mode[0]) {
+      const int mode = mbmi->filter_intra_mode_info.filter_intra_mode[0];
+      mode_cost += (av1_cost_bit(prob, 1) +
+                    write_uniform_cost(FILTER_INTRA_MODES, mode));
+    } else {
+      mode_cost += av1_cost_bit(prob, 0);
+    }
+  }
+#endif  // CONFIG_FILTER_INTRA
+  this_rd = RDCOST(x->rdmult, x->rddiv, this_rd_stats.rate + mode_cost,
+                   this_rd_stats.dist);
+  return this_rd;
+}
+
+#if CONFIG_PALETTE
+// Extends 'color_map' array from 'orig_width x orig_height' to 'new_width x
+// new_height'. Extra rows and columns are filled in by copying last valid
+// row/column.
+static void extend_palette_color_map(uint8_t *const color_map, int orig_width,
+                                     int orig_height, int new_width,
+                                     int new_height) {
+  int j;
+  assert(new_width >= orig_width);
+  assert(new_height >= orig_height);
+  if (new_width == orig_width && new_height == orig_height) return;
+
+  for (j = orig_height - 1; j >= 0; --j) {
+    memmove(color_map + j * new_width, color_map + j * orig_width, orig_width);
+    // Copy last column to extra columns.
+    memset(color_map + j * new_width + orig_width,
+           color_map[j * new_width + orig_width - 1], new_width - orig_width);
+  }
+  // Copy last row to extra rows.
+  for (j = orig_height; j < new_height; ++j) {
+    memcpy(color_map + j * new_width, color_map + (orig_height - 1) * new_width,
+           new_width);
+  }
+}
+
+static int rd_pick_palette_intra_sby(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                     BLOCK_SIZE bsize, int palette_ctx,
+                                     int dc_mode_cost, MB_MODE_INFO *best_mbmi,
+                                     uint8_t *best_palette_color_map,
+                                     int64_t *best_rd, int64_t *best_model_rd,
+                                     int *rate, int *rate_tokenonly,
+                                     int64_t *distortion, int *skippable) {
+  int rate_overhead = 0;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mic = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mic->mbmi;
+  int this_rate, colors, n;
+  const int src_stride = x->plane[0].src.stride;
+  const uint8_t *const src = x->plane[0].src.buf;
+  uint8_t *const color_map = xd->plane[0].color_index_map;
+  int block_width, block_height, rows, cols;
+  av1_get_block_dimensions(bsize, 0, xd, &block_width, &block_height, &rows,
+                           &cols);
+
+  assert(cpi->common.allow_screen_content_tools);
+
+#if CONFIG_HIGHBITDEPTH
+  if (cpi->common.use_highbitdepth)
+    colors = av1_count_colors_highbd(src, src_stride, rows, cols,
+                                     cpi->common.bit_depth);
+  else
+#endif  // CONFIG_HIGHBITDEPTH
+    colors = av1_count_colors(src, src_stride, rows, cols);
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+#endif  // CONFIG_FILTER_INTRA
+
+  if (colors > 1 && colors <= 64) {
+    int r, c, i, j, k, palette_mode_cost;
+    const int max_itr = 50;
+    uint8_t color_order[PALETTE_MAX_SIZE];
+    float *const data = x->palette_buffer->kmeans_data_buf;
+    float centroids[PALETTE_MAX_SIZE];
+    float lb, ub, val;
+    RD_STATS tokenonly_rd_stats;
+    int64_t this_rd, this_model_rd;
+    PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+#if CONFIG_HIGHBITDEPTH
+    uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+    if (cpi->common.use_highbitdepth)
+      lb = ub = src16[0];
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      lb = ub = src[0];
+
+#if CONFIG_HIGHBITDEPTH
+    if (cpi->common.use_highbitdepth) {
+      for (r = 0; r < rows; ++r) {
+        for (c = 0; c < cols; ++c) {
+          val = src16[r * src_stride + c];
+          data[r * cols + c] = val;
+          if (val < lb)
+            lb = val;
+          else if (val > ub)
+            ub = val;
+        }
+      }
+    } else {
+#endif  // CONFIG_HIGHBITDEPTH
+      for (r = 0; r < rows; ++r) {
+        for (c = 0; c < cols; ++c) {
+          val = src[r * src_stride + c];
+          data[r * cols + c] = val;
+          if (val < lb)
+            lb = val;
+          else if (val > ub)
+            ub = val;
+        }
+      }
+#if CONFIG_HIGHBITDEPTH
+    }
+#endif  // CONFIG_HIGHBITDEPTH
+
+    mbmi->mode = DC_PRED;
+#if CONFIG_FILTER_INTRA
+    mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+#endif  // CONFIG_FILTER_INTRA
+
+    if (rows * cols > PALETTE_MAX_BLOCK_SIZE) return 0;
+
+    for (n = colors > PALETTE_MAX_SIZE ? PALETTE_MAX_SIZE : colors; n >= 2;
+         --n) {
+      for (i = 0; i < n; ++i)
+        centroids[i] = lb + (2 * i + 1) * (ub - lb) / n / 2;
+      av1_k_means(data, centroids, color_map, rows * cols, n, 1, max_itr);
+      k = av1_remove_duplicates(centroids, n);
+
+#if CONFIG_HIGHBITDEPTH
+      if (cpi->common.use_highbitdepth)
+        for (i = 0; i < k; ++i)
+          pmi->palette_colors[i] =
+              clip_pixel_highbd((int)centroids[i], cpi->common.bit_depth);
+      else
+#endif  // CONFIG_HIGHBITDEPTH
+        for (i = 0; i < k; ++i)
+          pmi->palette_colors[i] = clip_pixel((int)centroids[i]);
+      pmi->palette_size[0] = k;
+
+      av1_calc_indices(data, centroids, color_map, rows * cols, k, 1);
+      extend_palette_color_map(color_map, cols, rows, block_width,
+                               block_height);
+      palette_mode_cost =
+          dc_mode_cost +
+          cpi->palette_y_size_cost[bsize - BLOCK_8X8][k - PALETTE_MIN_SIZE] +
+          write_uniform_cost(k, color_map[0]) +
+          av1_cost_bit(
+              av1_default_palette_y_mode_prob[bsize - BLOCK_8X8][palette_ctx],
+              1);
+      palette_mode_cost += av1_palette_color_cost_y(pmi, cpi->common.bit_depth);
+      for (i = 0; i < rows; ++i) {
+        for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
+          int color_idx;
+          const int color_ctx = av1_get_palette_color_index_context(
+              color_map, block_width, i, j, k, color_order, &color_idx);
+          assert(color_idx >= 0 && color_idx < k);
+          palette_mode_cost += cpi->palette_y_color_cost[k - PALETTE_MIN_SIZE]
+                                                        [color_ctx][color_idx];
+        }
+      }
+      this_model_rd = intra_model_yrd(cpi, x, bsize, palette_mode_cost);
+      if (*best_model_rd != INT64_MAX &&
+          this_model_rd > *best_model_rd + (*best_model_rd >> 1))
+        continue;
+      if (this_model_rd < *best_model_rd) *best_model_rd = this_model_rd;
+      super_block_yrd(cpi, x, &tokenonly_rd_stats, bsize, *best_rd);
+      if (tokenonly_rd_stats.rate == INT_MAX) continue;
+      this_rate = tokenonly_rd_stats.rate + palette_mode_cost;
+      this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, tokenonly_rd_stats.dist);
+      if (!xd->lossless[mbmi->segment_id] && mbmi->sb_type >= BLOCK_8X8) {
+        tokenonly_rd_stats.rate -= tx_size_cost(cpi, x, bsize, mbmi->tx_size);
+      }
+      if (this_rd < *best_rd) {
+        *best_rd = this_rd;
+        memcpy(best_palette_color_map, color_map,
+               block_width * block_height * sizeof(color_map[0]));
+        *best_mbmi = *mbmi;
+        rate_overhead = this_rate - tokenonly_rd_stats.rate;
+        if (rate) *rate = this_rate;
+        if (rate_tokenonly) *rate_tokenonly = tokenonly_rd_stats.rate;
+        if (distortion) *distortion = tokenonly_rd_stats.dist;
+        if (skippable) *skippable = tokenonly_rd_stats.skip;
+      }
+    }
+  }
+
+  if (best_mbmi->palette_mode_info.palette_size[0] > 0) {
+    memcpy(color_map, best_palette_color_map,
+           rows * cols * sizeof(best_palette_color_map[0]));
+  }
+  *mbmi = *best_mbmi;
+  return rate_overhead;
+}
+#endif  // CONFIG_PALETTE
+
+static int64_t rd_pick_intra_sub_8x8_y_subblock_mode(
+    const AV1_COMP *const cpi, MACROBLOCK *x, int row, int col,
+    PREDICTION_MODE *best_mode, const int *bmode_costs, ENTROPY_CONTEXT *a,
+    ENTROPY_CONTEXT *l, int *bestrate, int *bestratey, int64_t *bestdistortion,
+    BLOCK_SIZE bsize, TX_SIZE tx_size, int *y_skip, int64_t rd_thresh) {
+  const AV1_COMMON *const cm = &cpi->common;
+  PREDICTION_MODE mode;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int64_t best_rd = rd_thresh;
+  struct macroblock_plane *p = &x->plane[0];
+  struct macroblockd_plane *pd = &xd->plane[0];
+  const int src_stride = p->src.stride;
+  const int dst_stride = pd->dst.stride;
+  const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
+  uint8_t *dst_init = &pd->dst.buf[row * 4 * dst_stride + col * 4];
+#if CONFIG_CB4X4
+  // TODO(jingning): This is a temporal change. The whole function should be
+  // out when cb4x4 is enabled.
+  ENTROPY_CONTEXT ta[4], tempa[4];
+  ENTROPY_CONTEXT tl[4], templ[4];
+#else
+  ENTROPY_CONTEXT ta[2], tempa[2];
+  ENTROPY_CONTEXT tl[2], templ[2];
+#endif  // CONFIG_CB4X4
+
+  const int pred_width_in_4x4_blocks = num_4x4_blocks_wide_lookup[bsize];
+  const int pred_height_in_4x4_blocks = num_4x4_blocks_high_lookup[bsize];
+  const int tx_width_unit = tx_size_wide_unit[tx_size];
+  const int tx_height_unit = tx_size_high_unit[tx_size];
+  const int pred_block_width = block_size_wide[bsize];
+  const int pred_block_height = block_size_high[bsize];
+  const int tx_width = tx_size_wide[tx_size];
+  const int tx_height = tx_size_high[tx_size];
+  const int pred_width_in_transform_blocks = pred_block_width / tx_width;
+  const int pred_height_in_transform_blocks = pred_block_height / tx_height;
+  int idx, idy;
+  int best_can_skip = 0;
+  uint8_t best_dst[8 * 8];
+#if CONFIG_HIGHBITDEPTH
+  uint16_t best_dst16[8 * 8];
+#endif  // CONFIG_HIGHBITDEPTH
+  const int is_lossless = xd->lossless[xd->mi[0]->mbmi.segment_id];
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+  const int sub_bsize = bsize;
+#else
+  const int sub_bsize = BLOCK_4X4;
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+
+#if CONFIG_PVQ
+  od_rollback_buffer pre_buf, post_buf;
+  od_encode_checkpoint(&x->daala_enc, &pre_buf);
+  od_encode_checkpoint(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+
+  assert(bsize < BLOCK_8X8);
+  assert(tx_width < 8 || tx_height < 8);
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+  if (is_lossless)
+    assert(tx_width == 4 && tx_height == 4);
+  else
+    assert(tx_width == pred_block_width && tx_height == pred_block_height);
+#else
+  assert(tx_width == 4 && tx_height == 4);
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+
+  memcpy(ta, a, pred_width_in_transform_blocks * sizeof(a[0]));
+  memcpy(tl, l, pred_height_in_transform_blocks * sizeof(l[0]));
+
+  xd->mi[0]->mbmi.tx_size = tx_size;
+
+#if CONFIG_PALETTE
+  xd->mi[0]->mbmi.palette_mode_info.palette_size[0] = 0;
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+#if CONFIG_PVQ
+    od_encode_checkpoint(&x->daala_enc, &pre_buf);
+#endif
+    for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
+      int64_t this_rd;
+      int ratey = 0;
+      int64_t distortion = 0;
+      int rate = bmode_costs[mode];
+      int can_skip = 1;
+
+      if (!(cpi->sf.intra_y_mode_mask[txsize_sqr_up_map[tx_size]] &
+            (1 << mode)))
+        continue;
+
+      // Only do the oblique modes if the best so far is
+      // one of the neighboring directional modes
+      if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
+        if (conditional_skipintra(mode, *best_mode)) continue;
+      }
+
+      memcpy(tempa, ta, pred_width_in_transform_blocks * sizeof(ta[0]));
+      memcpy(templ, tl, pred_height_in_transform_blocks * sizeof(tl[0]));
+
+      for (idy = 0; idy < pred_height_in_transform_blocks; ++idy) {
+        for (idx = 0; idx < pred_width_in_transform_blocks; ++idx) {
+          const int block_raster_idx = (row + idy) * 2 + (col + idx);
+          const int block =
+              av1_raster_order_to_block_index(tx_size, block_raster_idx);
+          const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
+          uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
+#if !CONFIG_PVQ
+          int16_t *const src_diff = av1_raster_block_offset_int16(
+              BLOCK_8X8, block_raster_idx, p->src_diff);
+#endif
+          int skip;
+          assert(block < 4);
+          assert(IMPLIES(tx_size == TX_4X8 || tx_size == TX_8X4,
+                         idx == 0 && idy == 0));
+          assert(IMPLIES(tx_size == TX_4X8 || tx_size == TX_8X4,
+                         block == 0 || block == 2));
+          xd->mi[0]->bmi[block_raster_idx].as_mode = mode;
+          av1_predict_intra_block(
+              xd, pd->width, pd->height, txsize_to_bsize[tx_size], mode, dst,
+              dst_stride, dst, dst_stride, col + idx, row + idy, 0);
+#if !CONFIG_PVQ
+          aom_highbd_subtract_block(tx_height, tx_width, src_diff, 8, src,
+                                    src_stride, dst, dst_stride, xd->bd);
+#endif
+          if (is_lossless) {
+            TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block, tx_size);
+            const SCAN_ORDER *scan_order = get_scan(cm, tx_size, tx_type, 0);
+            const int coeff_ctx =
+                combine_entropy_contexts(tempa[idx], templ[idy]);
+#if !CONFIG_PVQ
+            av1_xform_quant(cm, x, 0, block, row + idy, col + idx, BLOCK_8X8,
+                            tx_size, coeff_ctx, AV1_XFORM_QUANT_FP);
+            ratey += av1_cost_coeffs(cpi, x, 0, block, tx_size, scan_order,
+                                     tempa + idx, templ + idy,
+                                     cpi->sf.use_fast_coef_costing);
+            skip = (p->eobs[block] == 0);
+            can_skip &= skip;
+            tempa[idx] = !skip;
+            templ[idy] = !skip;
+#if CONFIG_EXT_TX
+            if (tx_size == TX_8X4) {
+              tempa[idx + 1] = tempa[idx];
+            } else if (tx_size == TX_4X8) {
+              templ[idy + 1] = templ[idy];
+            }
+#endif  // CONFIG_EXT_TX
+#else
+            (void)scan_order;
+
+            av1_xform_quant(cm, x, 0, block, row + idy, col + idx, BLOCK_8X8,
+                            tx_size, coeff_ctx, AV1_XFORM_QUANT_B);
+
+            ratey += x->rate;
+            skip = x->pvq_skip[0];
+            tempa[idx] = !skip;
+            templ[idy] = !skip;
+            can_skip &= skip;
+#endif
+            if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
+              goto next_highbd;
+#if CONFIG_PVQ
+            if (!skip)
+#endif
+              av1_inverse_transform_block(xd, BLOCK_OFFSET(pd->dqcoeff, block),
+                                          DCT_DCT, tx_size, dst, dst_stride,
+                                          p->eobs[block]);
+          } else {
+            int64_t dist;
+            unsigned int tmp;
+            TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block, tx_size);
+            const SCAN_ORDER *scan_order = get_scan(cm, tx_size, tx_type, 0);
+            const int coeff_ctx =
+                combine_entropy_contexts(tempa[idx], templ[idy]);
+#if !CONFIG_PVQ
+            av1_xform_quant(cm, x, 0, block, row + idy, col + idx, BLOCK_8X8,
+                            tx_size, coeff_ctx, AV1_XFORM_QUANT_FP);
+            av1_optimize_b(cm, x, 0, block, tx_size, coeff_ctx);
+            ratey += av1_cost_coeffs(cpi, x, 0, block, tx_size, scan_order,
+                                     tempa + idx, templ + idy,
+                                     cpi->sf.use_fast_coef_costing);
+            skip = (p->eobs[block] == 0);
+            can_skip &= skip;
+            tempa[idx] = !skip;
+            templ[idy] = !skip;
+#if CONFIG_EXT_TX
+            if (tx_size == TX_8X4) {
+              tempa[idx + 1] = tempa[idx];
+            } else if (tx_size == TX_4X8) {
+              templ[idy + 1] = templ[idy];
+            }
+#endif  // CONFIG_EXT_TX
+#else
+            (void)scan_order;
+
+            av1_xform_quant(cm, x, 0, block, row + idy, col + idx, BLOCK_8X8,
+                            tx_size, coeff_ctx, AV1_XFORM_QUANT_FP);
+            ratey += x->rate;
+            skip = x->pvq_skip[0];
+            tempa[idx] = !skip;
+            templ[idy] = !skip;
+            can_skip &= skip;
+#endif
+#if CONFIG_PVQ
+            if (!skip)
+#endif
+              av1_inverse_transform_block(xd, BLOCK_OFFSET(pd->dqcoeff, block),
+                                          tx_type, tx_size, dst, dst_stride,
+                                          p->eobs[block]);
+            cpi->fn_ptr[sub_bsize].vf(src, src_stride, dst, dst_stride, &tmp);
+            dist = (int64_t)tmp << 4;
+            distortion += dist;
+            if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
+              goto next_highbd;
+          }
+        }
+      }
+
+      rate += ratey;
+      this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
+
+      if (this_rd < best_rd) {
+        *bestrate = rate;
+        *bestratey = ratey;
+        *bestdistortion = distortion;
+        best_rd = this_rd;
+        best_can_skip = can_skip;
+        *best_mode = mode;
+        memcpy(a, tempa, pred_width_in_transform_blocks * sizeof(tempa[0]));
+        memcpy(l, templ, pred_height_in_transform_blocks * sizeof(templ[0]));
+#if CONFIG_PVQ
+        od_encode_checkpoint(&x->daala_enc, &post_buf);
+#endif
+        for (idy = 0; idy < pred_height_in_transform_blocks * 4; ++idy) {
+          memcpy(best_dst16 + idy * 8,
+                 CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
+                 pred_width_in_transform_blocks * 4 * sizeof(uint16_t));
+        }
+      }
+    next_highbd : {}
+#if CONFIG_PVQ
+      od_encode_rollback(&x->daala_enc, &pre_buf);
+#endif
+    }
+
+    if (best_rd >= rd_thresh) return best_rd;
+
+#if CONFIG_PVQ
+    od_encode_rollback(&x->daala_enc, &post_buf);
+#endif
+
+    if (y_skip) *y_skip &= best_can_skip;
+
+    for (idy = 0; idy < pred_height_in_transform_blocks * 4; ++idy) {
+      memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
+             best_dst16 + idy * 8,
+             pred_width_in_transform_blocks * 4 * sizeof(uint16_t));
+    }
+
+    return best_rd;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_PVQ
+  od_encode_checkpoint(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+
+  for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
+    int64_t this_rd;
+    int ratey = 0;
+    int64_t distortion = 0;
+    int rate = bmode_costs[mode];
+    int can_skip = 1;
+
+    if (!(cpi->sf.intra_y_mode_mask[txsize_sqr_up_map[tx_size]] &
+          (1 << mode))) {
+      continue;
+    }
+
+    // Only do the oblique modes if the best so far is
+    // one of the neighboring directional modes
+    if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
+      if (conditional_skipintra(mode, *best_mode)) continue;
+    }
+
+    memcpy(tempa, ta, pred_width_in_transform_blocks * sizeof(ta[0]));
+    memcpy(templ, tl, pred_height_in_transform_blocks * sizeof(tl[0]));
+
+    for (idy = 0; idy < pred_height_in_4x4_blocks; idy += tx_height_unit) {
+      for (idx = 0; idx < pred_width_in_4x4_blocks; idx += tx_width_unit) {
+        const int block_raster_idx = (row + idy) * 2 + (col + idx);
+        int block = av1_raster_order_to_block_index(tx_size, block_raster_idx);
+        const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
+        uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
+#if !CONFIG_PVQ
+        int16_t *const src_diff = av1_raster_block_offset_int16(
+            BLOCK_8X8, block_raster_idx, p->src_diff);
+#endif  // !CONFIG_PVQ
+        int skip;
+        assert(block < 4);
+        assert(IMPLIES(tx_size == TX_4X8 || tx_size == TX_8X4,
+                       idx == 0 && idy == 0));
+        assert(IMPLIES(tx_size == TX_4X8 || tx_size == TX_8X4,
+                       block == 0 || block == 2));
+        xd->mi[0]->bmi[block_raster_idx].as_mode = mode;
+        av1_predict_intra_block(xd, pd->width, pd->height,
+                                txsize_to_bsize[tx_size], mode, dst, dst_stride,
+                                dst, dst_stride,
+#if CONFIG_CB4X4
+                                2 * (col + idx), 2 * (row + idy),
+#else
+                                col + idx, row + idy,
+#endif  // CONFIG_CB4X4
+                                0);
+#if !CONFIG_PVQ
+        aom_subtract_block(tx_height, tx_width, src_diff, 8, src, src_stride,
+                           dst, dst_stride);
+#endif  // !CONFIG_PVQ
+
+        TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, block, tx_size);
+        const SCAN_ORDER *scan_order = get_scan(cm, tx_size, tx_type, 0);
+        const int coeff_ctx = combine_entropy_contexts(tempa[idx], templ[idy]);
+#if CONFIG_CB4X4
+        block = 4 * block;
+#endif  // CONFIG_CB4X4
+#if !CONFIG_PVQ
+        const AV1_XFORM_QUANT xform_quant =
+            is_lossless ? AV1_XFORM_QUANT_B : AV1_XFORM_QUANT_FP;
+        av1_xform_quant(cm, x, 0, block,
+#if CONFIG_CB4X4
+                        2 * (row + idy), 2 * (col + idx),
+#else
+                        row + idy, col + idx,
+#endif  // CONFIG_CB4X4
+                        BLOCK_8X8, tx_size, coeff_ctx, xform_quant);
+
+        if (!is_lossless) {
+          av1_optimize_b(cm, x, 0, block, tx_size, coeff_ctx);
+        }
+
+        ratey +=
+            av1_cost_coeffs(cpi, x, 0, block, tx_size, scan_order, tempa + idx,
+                            templ + idy, cpi->sf.use_fast_coef_costing);
+        skip = (p->eobs[block] == 0);
+        can_skip &= skip;
+        tempa[idx] = !skip;
+        templ[idy] = !skip;
+#if CONFIG_EXT_TX
+        if (tx_size == TX_8X4) {
+          tempa[idx + 1] = tempa[idx];
+        } else if (tx_size == TX_4X8) {
+          templ[idy + 1] = templ[idy];
+        }
+#endif  // CONFIG_EXT_TX
+#else
+        (void)scan_order;
+
+        av1_xform_quant(cm, x, 0, block,
+#if CONFIG_CB4X4
+                        2 * (row + idy), 2 * (col + idx),
+#else
+                        row + idy, col + idx,
+#endif  // CONFIG_CB4X4
+                        BLOCK_8X8, tx_size, coeff_ctx, AV1_XFORM_QUANT_FP);
+
+        ratey += x->rate;
+        skip = x->pvq_skip[0];
+        tempa[idx] = !skip;
+        templ[idy] = !skip;
+        can_skip &= skip;
+#endif  // !CONFIG_PVQ
+
+        if (!is_lossless) {  // To use the pixel domain distortion, we need to
+                             // calculate inverse txfm *before* calculating RD
+                             // cost. Compared to calculating the distortion in
+                             // the frequency domain, the overhead of encoding
+                             // effort is low.
+#if CONFIG_PVQ
+          if (!skip)
+#endif  // CONFIG_PVQ
+            av1_inverse_transform_block(xd, BLOCK_OFFSET(pd->dqcoeff, block),
+                                        tx_type, tx_size, dst, dst_stride,
+                                        p->eobs[block]);
+          unsigned int tmp;
+          cpi->fn_ptr[sub_bsize].vf(src, src_stride, dst, dst_stride, &tmp);
+          const int64_t dist = (int64_t)tmp << 4;
+          distortion += dist;
+        }
+
+        if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
+          goto next;
+
+        if (is_lossless) {  // Calculate inverse txfm *after* RD cost.
+#if CONFIG_PVQ
+          if (!skip)
+#endif  // CONFIG_PVQ
+            av1_inverse_transform_block(xd, BLOCK_OFFSET(pd->dqcoeff, block),
+                                        DCT_DCT, tx_size, dst, dst_stride,
+                                        p->eobs[block]);
+        }
+      }
+    }
+
+    rate += ratey;
+    this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
+
+    if (this_rd < best_rd) {
+      *bestrate = rate;
+      *bestratey = ratey;
+      *bestdistortion = distortion;
+      best_rd = this_rd;
+      best_can_skip = can_skip;
+      *best_mode = mode;
+      memcpy(a, tempa, pred_width_in_transform_blocks * sizeof(tempa[0]));
+      memcpy(l, templ, pred_height_in_transform_blocks * sizeof(templ[0]));
+#if CONFIG_PVQ
+      od_encode_checkpoint(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+      for (idy = 0; idy < pred_height_in_transform_blocks * 4; ++idy)
+        memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
+               pred_width_in_transform_blocks * 4);
+    }
+  next : {}
+#if CONFIG_PVQ
+    od_encode_rollback(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+  }     // mode decision loop
+
+  if (best_rd >= rd_thresh) return best_rd;
+
+#if CONFIG_PVQ
+  od_encode_rollback(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+
+  if (y_skip) *y_skip &= best_can_skip;
+
+  for (idy = 0; idy < pred_height_in_transform_blocks * 4; ++idy)
+    memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
+           pred_width_in_transform_blocks * 4);
+
+  return best_rd;
+}
+
+static int64_t rd_pick_intra_sub_8x8_y_mode(const AV1_COMP *const cpi,
+                                            MACROBLOCK *mb, int *rate,
+                                            int *rate_y, int64_t *distortion,
+                                            int *y_skip, int64_t best_rd) {
+  const MACROBLOCKD *const xd = &mb->e_mbd;
+  MODE_INFO *const mic = xd->mi[0];
+  const MODE_INFO *above_mi = xd->above_mi;
+  const MODE_INFO *left_mi = xd->left_mi;
+  MB_MODE_INFO *const mbmi = &mic->mbmi;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  const int pred_width_in_4x4_blocks = num_4x4_blocks_wide_lookup[bsize];
+  const int pred_height_in_4x4_blocks = num_4x4_blocks_high_lookup[bsize];
+  int idx, idy;
+  int cost = 0;
+  int64_t total_distortion = 0;
+  int tot_rate_y = 0;
+  int64_t total_rd = 0;
+  const int *bmode_costs = cpi->mbmode_cost[0];
+  const int is_lossless = xd->lossless[mbmi->segment_id];
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+  const TX_SIZE tx_size = is_lossless ? TX_4X4 : max_txsize_rect_lookup[bsize];
+#else
+  const TX_SIZE tx_size = TX_4X4;
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+
+#if CONFIG_EXT_INTRA
+#if CONFIG_INTRA_INTERP
+  mbmi->intra_filter = INTRA_FILTER_LINEAR;
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+#endif  // CONFIG_FILTER_INTRA
+
+  // TODO(any): Add search of the tx_type to improve rd performance at the
+  // expense of speed.
+  mbmi->tx_type = DCT_DCT;
+  mbmi->tx_size = tx_size;
+
+  if (y_skip) *y_skip = 1;
+
+  // Pick modes for each prediction sub-block (of size 4x4, 4x8, or 8x4) in this
+  // 8x8 coding block.
+  for (idy = 0; idy < 2; idy += pred_height_in_4x4_blocks) {
+    for (idx = 0; idx < 2; idx += pred_width_in_4x4_blocks) {
+      PREDICTION_MODE best_mode = DC_PRED;
+      int r = INT_MAX, ry = INT_MAX;
+      int64_t d = INT64_MAX, this_rd = INT64_MAX;
+      int j;
+      const int pred_block_idx = idy * 2 + idx;
+      if (cpi->common.frame_type == KEY_FRAME) {
+        const PREDICTION_MODE A =
+            av1_above_block_mode(mic, above_mi, pred_block_idx);
+        const PREDICTION_MODE L =
+            av1_left_block_mode(mic, left_mi, pred_block_idx);
+
+        bmode_costs = cpi->y_mode_costs[A][L];
+      }
+      this_rd = rd_pick_intra_sub_8x8_y_subblock_mode(
+          cpi, mb, idy, idx, &best_mode, bmode_costs,
+          xd->plane[0].above_context + idx, xd->plane[0].left_context + idy, &r,
+          &ry, &d, bsize, tx_size, y_skip, best_rd - total_rd);
+#if !CONFIG_DAALA_DIST
+      if (this_rd >= best_rd - total_rd) return INT64_MAX;
+#endif  // !CONFIG_DAALA_DIST
+      total_rd += this_rd;
+      cost += r;
+      total_distortion += d;
+      tot_rate_y += ry;
+
+      mic->bmi[pred_block_idx].as_mode = best_mode;
+      for (j = 1; j < pred_height_in_4x4_blocks; ++j)
+        mic->bmi[pred_block_idx + j * 2].as_mode = best_mode;
+      for (j = 1; j < pred_width_in_4x4_blocks; ++j)
+        mic->bmi[pred_block_idx + j].as_mode = best_mode;
+
+      if (total_rd >= best_rd) return INT64_MAX;
+    }
+  }
+  mbmi->mode = mic->bmi[3].as_mode;
+
+#if CONFIG_DAALA_DIST
+  {
+    const struct macroblock_plane *p = &mb->plane[0];
+    const struct macroblockd_plane *pd = &xd->plane[0];
+    const int src_stride = p->src.stride;
+    const int dst_stride = pd->dst.stride;
+    uint8_t *src = p->src.buf;
+    uint8_t *dst = pd->dst.buf;
+    int use_activity_masking = 0;
+    int qm = OD_HVS_QM;
+
+#if CONFIG_PVQ
+    use_activity_masking = mb->daala_enc.use_activity_masking;
+#endif  // CONFIG_PVQ
+    // Daala-defined distortion computed for the block of 8x8 pixels
+    total_distortion = av1_daala_dist(src, src_stride, dst, dst_stride, 8, 8,
+                                      qm, use_activity_masking, mb->qindex)
+                       << 4;
+  }
+#endif  // CONFIG_DAALA_DIST
+  // Add in the cost of the transform type
+  if (!is_lossless) {
+    int rate_tx_type = 0;
+#if CONFIG_EXT_TX
+    if (get_ext_tx_types(tx_size, bsize, 0, cpi->common.reduced_tx_set_used) >
+        1) {
+      const int eset =
+          get_ext_tx_set(tx_size, bsize, 0, cpi->common.reduced_tx_set_used);
+      rate_tx_type = cpi->intra_tx_type_costs[eset][txsize_sqr_map[tx_size]]
+                                             [mbmi->mode][mbmi->tx_type];
+    }
+#else
+    rate_tx_type =
+        cpi->intra_tx_type_costs[txsize_sqr_map[tx_size]]
+                                [intra_mode_to_tx_type_context[mbmi->mode]]
+                                [mbmi->tx_type];
+#endif  // CONFIG_EXT_TX
+    assert(mbmi->tx_size == tx_size);
+    cost += rate_tx_type;
+    tot_rate_y += rate_tx_type;
+  }
+
+  *rate = cost;
+  *rate_y = tot_rate_y;
+  *distortion = total_distortion;
+
+  return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
+}
+
+#if CONFIG_FILTER_INTRA
+// Return 1 if an filter intra mode is selected; return 0 otherwise.
+static int rd_pick_filter_intra_sby(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                    int *rate, int *rate_tokenonly,
+                                    int64_t *distortion, int *skippable,
+                                    BLOCK_SIZE bsize, int mode_cost,
+                                    int64_t *best_rd, int64_t *best_model_rd,
+                                    uint16_t skip_mask) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mic = xd->mi[0];
+  MB_MODE_INFO *mbmi = &mic->mbmi;
+  int filter_intra_selected_flag = 0;
+  FILTER_INTRA_MODE mode;
+  TX_SIZE best_tx_size = TX_4X4;
+  FILTER_INTRA_MODE_INFO filter_intra_mode_info;
+  TX_TYPE best_tx_type;
+
+  av1_zero(filter_intra_mode_info);
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 1;
+  mbmi->mode = DC_PRED;
+#if CONFIG_PALETTE
+  mbmi->palette_mode_info.palette_size[0] = 0;
+#endif  // CONFIG_PALETTE
+
+  for (mode = 0; mode < FILTER_INTRA_MODES; ++mode) {
+    int this_rate;
+    int64_t this_rd, this_model_rd;
+    RD_STATS tokenonly_rd_stats;
+    if (skip_mask & (1 << mode)) continue;
+    mbmi->filter_intra_mode_info.filter_intra_mode[0] = mode;
+    this_model_rd = intra_model_yrd(cpi, x, bsize, mode_cost);
+    if (*best_model_rd != INT64_MAX &&
+        this_model_rd > *best_model_rd + (*best_model_rd >> 1))
+      continue;
+    if (this_model_rd < *best_model_rd) *best_model_rd = this_model_rd;
+    super_block_yrd(cpi, x, &tokenonly_rd_stats, bsize, *best_rd);
+    if (tokenonly_rd_stats.rate == INT_MAX) continue;
+    this_rate = tokenonly_rd_stats.rate +
+                av1_cost_bit(cpi->common.fc->filter_intra_probs[0], 1) +
+                write_uniform_cost(FILTER_INTRA_MODES, mode) + mode_cost;
+    this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, tokenonly_rd_stats.dist);
+
+    if (this_rd < *best_rd) {
+      *best_rd = this_rd;
+      best_tx_size = mic->mbmi.tx_size;
+      filter_intra_mode_info = mbmi->filter_intra_mode_info;
+      best_tx_type = mic->mbmi.tx_type;
+      *rate = this_rate;
+      *rate_tokenonly = tokenonly_rd_stats.rate;
+      *distortion = tokenonly_rd_stats.dist;
+      *skippable = tokenonly_rd_stats.skip;
+      filter_intra_selected_flag = 1;
+    }
+  }
+
+  if (filter_intra_selected_flag) {
+    mbmi->mode = DC_PRED;
+    mbmi->tx_size = best_tx_size;
+    mbmi->filter_intra_mode_info.use_filter_intra_mode[0] =
+        filter_intra_mode_info.use_filter_intra_mode[0];
+    mbmi->filter_intra_mode_info.filter_intra_mode[0] =
+        filter_intra_mode_info.filter_intra_mode[0];
+    mbmi->tx_type = best_tx_type;
+    return 1;
+  } else {
+    return 0;
+  }
+}
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_EXT_INTRA
+// Run RD calculation with given luma intra prediction angle., and return
+// the RD cost. Update the best mode info. if the RD cost is the best so far.
+static int64_t calc_rd_given_intra_angle(
+    const AV1_COMP *const cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int mode_cost,
+    int64_t best_rd_in, int8_t angle_delta, int max_angle_delta, int *rate,
+    RD_STATS *rd_stats, int *best_angle_delta, TX_SIZE *best_tx_size,
+    TX_TYPE *best_tx_type,
+#if CONFIG_INTRA_INTERP
+    INTRA_FILTER *best_filter,
+#endif  // CONFIG_INTRA_INTERP
+    int64_t *best_rd, int64_t *best_model_rd) {
+  int this_rate;
+  RD_STATS tokenonly_rd_stats;
+  int64_t this_rd, this_model_rd;
+  MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
+
+  mbmi->angle_delta[0] = angle_delta;
+  this_model_rd = intra_model_yrd(cpi, x, bsize, mode_cost);
+  if (*best_model_rd != INT64_MAX &&
+      this_model_rd > *best_model_rd + (*best_model_rd >> 1))
+    return INT64_MAX;
+  if (this_model_rd < *best_model_rd) *best_model_rd = this_model_rd;
+  super_block_yrd(cpi, x, &tokenonly_rd_stats, bsize, best_rd_in);
+  if (tokenonly_rd_stats.rate == INT_MAX) return INT64_MAX;
+
+  this_rate = tokenonly_rd_stats.rate + mode_cost +
+              write_uniform_cost(2 * max_angle_delta + 1,
+                                 mbmi->angle_delta[0] + max_angle_delta);
+  this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, tokenonly_rd_stats.dist);
+
+  if (this_rd < *best_rd) {
+    *best_rd = this_rd;
+    *best_angle_delta = mbmi->angle_delta[0];
+    *best_tx_size = mbmi->tx_size;
+#if CONFIG_INTRA_INTERP
+    *best_filter = mbmi->intra_filter;
+#endif  // CONFIG_INTRA_INTERP
+    *best_tx_type = mbmi->tx_type;
+    *rate = this_rate;
+    rd_stats->rate = tokenonly_rd_stats.rate;
+    rd_stats->dist = tokenonly_rd_stats.dist;
+    rd_stats->skip = tokenonly_rd_stats.skip;
+  }
+  return this_rd;
+}
+
+// With given luma directional intra prediction mode, pick the best angle delta
+// Return the RD cost corresponding to the best angle delta.
+static int64_t rd_pick_intra_angle_sby(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                       int *rate, RD_STATS *rd_stats,
+                                       BLOCK_SIZE bsize, int mode_cost,
+                                       int64_t best_rd,
+                                       int64_t *best_model_rd) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mic = xd->mi[0];
+  MB_MODE_INFO *mbmi = &mic->mbmi;
+  int i, angle_delta, best_angle_delta = 0;
+  int first_try = 1;
+#if CONFIG_INTRA_INTERP
+  int p_angle;
+  const int intra_filter_ctx = av1_get_pred_context_intra_interp(xd);
+  INTRA_FILTER filter, best_filter = INTRA_FILTER_LINEAR;
+#endif  // CONFIG_INTRA_INTERP
+  int64_t this_rd, best_rd_in, rd_cost[2 * (MAX_ANGLE_DELTA + 2)];
+  TX_SIZE best_tx_size = mic->mbmi.tx_size;
+  TX_TYPE best_tx_type = mbmi->tx_type;
+
+  for (i = 0; i < 2 * (MAX_ANGLE_DELTA + 2); ++i) rd_cost[i] = INT64_MAX;
+
+  for (angle_delta = 0; angle_delta <= MAX_ANGLE_DELTA; angle_delta += 2) {
+#if CONFIG_INTRA_INTERP
+    for (filter = INTRA_FILTER_LINEAR; filter < INTRA_FILTERS; ++filter) {
+      if (FILTER_FAST_SEARCH && filter != INTRA_FILTER_LINEAR) continue;
+      mic->mbmi.intra_filter = filter;
+#endif  // CONFIG_INTRA_INTERP
+      for (i = 0; i < 2; ++i) {
+        best_rd_in = (best_rd == INT64_MAX)
+                         ? INT64_MAX
+                         : (best_rd + (best_rd >> (first_try ? 3 : 5)));
+        this_rd = calc_rd_given_intra_angle(
+            cpi, x, bsize,
+#if CONFIG_INTRA_INTERP
+            mode_cost + cpi->intra_filter_cost[intra_filter_ctx][filter],
+#else
+          mode_cost,
+#endif  // CONFIG_INTRA_INTERP
+            best_rd_in, (1 - 2 * i) * angle_delta, MAX_ANGLE_DELTA, rate,
+            rd_stats, &best_angle_delta, &best_tx_size, &best_tx_type,
+#if CONFIG_INTRA_INTERP
+            &best_filter,
+#endif  // CONFIG_INTRA_INTERP
+            &best_rd, best_model_rd);
+        rd_cost[2 * angle_delta + i] = this_rd;
+        if (first_try && this_rd == INT64_MAX) return best_rd;
+        first_try = 0;
+        if (angle_delta == 0) {
+          rd_cost[1] = this_rd;
+          break;
+        }
+      }
+#if CONFIG_INTRA_INTERP
+    }
+#endif  // CONFIG_INTRA_INTERP
+  }
+
+  assert(best_rd != INT64_MAX);
+  for (angle_delta = 1; angle_delta <= MAX_ANGLE_DELTA; angle_delta += 2) {
+    int64_t rd_thresh;
+#if CONFIG_INTRA_INTERP
+    for (filter = INTRA_FILTER_LINEAR; filter < INTRA_FILTERS; ++filter) {
+      if (FILTER_FAST_SEARCH && filter != INTRA_FILTER_LINEAR) continue;
+      mic->mbmi.intra_filter = filter;
+#endif  // CONFIG_INTRA_INTERP
+      for (i = 0; i < 2; ++i) {
+        int skip_search = 0;
+        rd_thresh = best_rd + (best_rd >> 5);
+        if (rd_cost[2 * (angle_delta + 1) + i] > rd_thresh &&
+            rd_cost[2 * (angle_delta - 1) + i] > rd_thresh)
+          skip_search = 1;
+        if (!skip_search) {
+          calc_rd_given_intra_angle(
+              cpi, x, bsize,
+#if CONFIG_INTRA_INTERP
+              mode_cost + cpi->intra_filter_cost[intra_filter_ctx][filter],
+#else
+            mode_cost,
+#endif  // CONFIG_INTRA_INTERP
+              best_rd, (1 - 2 * i) * angle_delta, MAX_ANGLE_DELTA, rate,
+              rd_stats, &best_angle_delta, &best_tx_size, &best_tx_type,
+#if CONFIG_INTRA_INTERP
+              &best_filter,
+#endif  // CONFIG_INTRA_INTERP
+              &best_rd, best_model_rd);
+        }
+      }
+#if CONFIG_INTRA_INTERP
+    }
+#endif  // CONFIG_INTRA_INTERP
+  }
+
+#if CONFIG_INTRA_INTERP
+  if (FILTER_FAST_SEARCH && rd_stats->rate < INT_MAX) {
+    p_angle = mode_to_angle_map[mbmi->mode] + best_angle_delta * ANGLE_STEP;
+    if (av1_is_intra_filter_switchable(p_angle)) {
+      for (filter = INTRA_FILTER_LINEAR + 1; filter < INTRA_FILTERS; ++filter) {
+        mic->mbmi.intra_filter = filter;
+        this_rd = calc_rd_given_intra_angle(
+            cpi, x, bsize,
+            mode_cost + cpi->intra_filter_cost[intra_filter_ctx][filter],
+            best_rd, best_angle_delta, MAX_ANGLE_DELTA, rate, rd_stats,
+            &best_angle_delta, &best_tx_size, &best_tx_type, &best_filter,
+            &best_rd, best_model_rd);
+      }
+    }
+  }
+#endif  // CONFIG_INTRA_INTERP
+
+  mbmi->tx_size = best_tx_size;
+  mbmi->angle_delta[0] = best_angle_delta;
+#if CONFIG_INTRA_INTERP
+  mic->mbmi.intra_filter = best_filter;
+#endif  // CONFIG_INTRA_INTERP
+  mbmi->tx_type = best_tx_type;
+  return best_rd;
+}
+
+// Indices are sign, integer, and fractional part of the gradient value
+static const uint8_t gradient_to_angle_bin[2][7][16] = {
+  {
+      { 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1 },
+      { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 },
+      { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 },
+      { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 },
+      { 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2 },
+      { 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2 },
+  },
+  {
+      { 6, 6, 6, 6, 5, 5, 5, 5, 5, 5, 5, 5, 4, 4, 4, 4 },
+      { 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 3, 3, 3, 3, 3, 3 },
+      { 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3 },
+      { 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3 },
+      { 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3 },
+      { 3, 3, 3, 3, 3, 3, 3, 2, 2, 2, 2, 2, 2, 2, 2, 2 },
+      { 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2 },
+  },
+};
+
+static const uint8_t mode_to_angle_bin[INTRA_MODES] = {
+  0, 2, 6, 0, 4, 3, 5, 7, 1, 0,
+};
+
+static void angle_estimation(const uint8_t *src, int src_stride, int rows,
+                             int cols, uint8_t *directional_mode_skip_mask) {
+  int i, r, c, index, dx, dy, temp, sn, remd, quot;
+  uint64_t hist[DIRECTIONAL_MODES];
+  uint64_t hist_sum = 0;
+
+  memset(hist, 0, DIRECTIONAL_MODES * sizeof(hist[0]));
+  src += src_stride;
+  for (r = 1; r < rows; ++r) {
+    for (c = 1; c < cols; ++c) {
+      dx = src[c] - src[c - 1];
+      dy = src[c] - src[c - src_stride];
+      temp = dx * dx + dy * dy;
+      if (dy == 0) {
+        index = 2;
+      } else {
+        sn = (dx > 0) ^ (dy > 0);
+        dx = abs(dx);
+        dy = abs(dy);
+        remd = dx % dy;
+        quot = dx / dy;
+        remd = remd * 16 / dy;
+        index = gradient_to_angle_bin[sn][AOMMIN(quot, 6)][AOMMIN(remd, 15)];
+      }
+      hist[index] += temp;
+    }
+    src += src_stride;
+  }
+
+  for (i = 0; i < DIRECTIONAL_MODES; ++i) hist_sum += hist[i];
+  for (i = 0; i < INTRA_MODES; ++i) {
+    if (i != DC_PRED && i != TM_PRED) {
+      const uint8_t angle_bin = mode_to_angle_bin[i];
+      uint64_t score = 2 * hist[angle_bin];
+      int weight = 2;
+      if (angle_bin > 0) {
+        score += hist[angle_bin - 1];
+        ++weight;
+      }
+      if (angle_bin < DIRECTIONAL_MODES - 1) {
+        score += hist[angle_bin + 1];
+        ++weight;
+      }
+      if (score * ANGLE_SKIP_THRESH < hist_sum * weight)
+        directional_mode_skip_mask[i] = 1;
+    }
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+static void highbd_angle_estimation(const uint8_t *src8, int src_stride,
+                                    int rows, int cols,
+                                    uint8_t *directional_mode_skip_mask) {
+  int i, r, c, index, dx, dy, temp, sn, remd, quot;
+  uint64_t hist[DIRECTIONAL_MODES];
+  uint64_t hist_sum = 0;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+
+  memset(hist, 0, DIRECTIONAL_MODES * sizeof(hist[0]));
+  src += src_stride;
+  for (r = 1; r < rows; ++r) {
+    for (c = 1; c < cols; ++c) {
+      dx = src[c] - src[c - 1];
+      dy = src[c] - src[c - src_stride];
+      temp = dx * dx + dy * dy;
+      if (dy == 0) {
+        index = 2;
+      } else {
+        sn = (dx > 0) ^ (dy > 0);
+        dx = abs(dx);
+        dy = abs(dy);
+        remd = dx % dy;
+        quot = dx / dy;
+        remd = remd * 16 / dy;
+        index = gradient_to_angle_bin[sn][AOMMIN(quot, 6)][AOMMIN(remd, 15)];
+      }
+      hist[index] += temp;
+    }
+    src += src_stride;
+  }
+
+  for (i = 0; i < DIRECTIONAL_MODES; ++i) hist_sum += hist[i];
+  for (i = 0; i < INTRA_MODES; ++i) {
+    if (i != DC_PRED && i != TM_PRED) {
+      const uint8_t angle_bin = mode_to_angle_bin[i];
+      uint64_t score = 2 * hist[angle_bin];
+      int weight = 2;
+      if (angle_bin > 0) {
+        score += hist[angle_bin - 1];
+        ++weight;
+      }
+      if (angle_bin < DIRECTIONAL_MODES - 1) {
+        score += hist[angle_bin + 1];
+        ++weight;
+      }
+      if (score * ANGLE_SKIP_THRESH < hist_sum * weight)
+        directional_mode_skip_mask[i] = 1;
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_EXT_INTRA
+
+// This function is used only for intra_only frames
+static int64_t rd_pick_intra_sby_mode(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                      int *rate, int *rate_tokenonly,
+                                      int64_t *distortion, int *skippable,
+                                      BLOCK_SIZE bsize, int64_t best_rd) {
+  uint8_t mode_idx;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mic = xd->mi[0];
+  MB_MODE_INFO *const mbmi = &mic->mbmi;
+  MB_MODE_INFO best_mbmi = *mbmi;
+  int64_t best_model_rd = INT64_MAX;
+#if CONFIG_EXT_INTRA
+  const int rows = block_size_high[bsize];
+  const int cols = block_size_wide[bsize];
+#if CONFIG_INTRA_INTERP
+  const int intra_filter_ctx = av1_get_pred_context_intra_interp(xd);
+#endif  // CONFIG_INTRA_INTERP
+  int is_directional_mode;
+  uint8_t directional_mode_skip_mask[INTRA_MODES];
+  const int src_stride = x->plane[0].src.stride;
+  const uint8_t *src = x->plane[0].src.buf;
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  int beat_best_rd = 0;
+  uint16_t filter_intra_mode_skip_mask = (1 << FILTER_INTRA_MODES) - 1;
+#endif  // CONFIG_FILTER_INTRA
+  const int *bmode_costs;
+#if CONFIG_PALETTE
+  PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+  uint8_t *best_palette_color_map =
+      cpi->common.allow_screen_content_tools
+          ? x->palette_buffer->best_palette_color_map
+          : NULL;
+  int palette_y_mode_ctx = 0;
+  const int try_palette =
+      cpi->common.allow_screen_content_tools && bsize >= BLOCK_8X8;
+#endif  // CONFIG_PALETTE
+  const MODE_INFO *above_mi = xd->above_mi;
+  const MODE_INFO *left_mi = xd->left_mi;
+  const PREDICTION_MODE A = av1_above_block_mode(mic, above_mi, 0);
+  const PREDICTION_MODE L = av1_left_block_mode(mic, left_mi, 0);
+  const PREDICTION_MODE FINAL_MODE_SEARCH = TM_PRED + 1;
+#if CONFIG_PVQ
+  od_rollback_buffer pre_buf, post_buf;
+
+  od_encode_checkpoint(&x->daala_enc, &pre_buf);
+  od_encode_checkpoint(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+  bmode_costs = cpi->y_mode_costs[A][L];
+
+#if CONFIG_EXT_INTRA
+  mbmi->angle_delta[0] = 0;
+  memset(directional_mode_skip_mask, 0,
+         sizeof(directional_mode_skip_mask[0]) * INTRA_MODES);
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+    highbd_angle_estimation(src, src_stride, rows, cols,
+                            directional_mode_skip_mask);
+  else
+#endif  // CONFIG_HIGHBITDEPTH
+    angle_estimation(src, src_stride, rows, cols, directional_mode_skip_mask);
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_PALETTE
+  pmi->palette_size[0] = 0;
+  if (above_mi)
+    palette_y_mode_ctx +=
+        (above_mi->mbmi.palette_mode_info.palette_size[0] > 0);
+  if (left_mi)
+    palette_y_mode_ctx += (left_mi->mbmi.palette_mode_info.palette_size[0] > 0);
+#endif  // CONFIG_PALETTE
+
+  if (cpi->sf.tx_type_search.fast_intra_tx_type_search)
+    x->use_default_intra_tx_type = 1;
+  else
+    x->use_default_intra_tx_type = 0;
+
+  /* Y Search for intra prediction mode */
+  for (mode_idx = DC_PRED; mode_idx <= FINAL_MODE_SEARCH; ++mode_idx) {
+    RD_STATS this_rd_stats;
+    int this_rate, this_rate_tokenonly, s;
+    int64_t this_distortion, this_rd, this_model_rd;
+    if (mode_idx == FINAL_MODE_SEARCH) {
+      if (x->use_default_intra_tx_type == 0) break;
+      mbmi->mode = best_mbmi.mode;
+      x->use_default_intra_tx_type = 0;
+    } else {
+      mbmi->mode = mode_idx;
+    }
+#if CONFIG_PVQ
+    od_encode_rollback(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+#if CONFIG_EXT_INTRA
+    mbmi->angle_delta[0] = 0;
+#endif  // CONFIG_EXT_INTRA
+    this_model_rd = intra_model_yrd(cpi, x, bsize, bmode_costs[mbmi->mode]);
+    if (best_model_rd != INT64_MAX &&
+        this_model_rd > best_model_rd + (best_model_rd >> 1))
+      continue;
+    if (this_model_rd < best_model_rd) best_model_rd = this_model_rd;
+#if CONFIG_EXT_INTRA
+    is_directional_mode = av1_is_directional_mode(mbmi->mode, bsize);
+    if (is_directional_mode && directional_mode_skip_mask[mbmi->mode]) continue;
+    if (is_directional_mode) {
+      this_rd_stats.rate = INT_MAX;
+      rd_pick_intra_angle_sby(cpi, x, &this_rate, &this_rd_stats, bsize,
+                              bmode_costs[mbmi->mode], best_rd, &best_model_rd);
+    } else {
+      super_block_yrd(cpi, x, &this_rd_stats, bsize, best_rd);
+    }
+#else
+    super_block_yrd(cpi, x, &this_rd_stats, bsize, best_rd);
+#endif  // CONFIG_EXT_INTRA
+    this_rate_tokenonly = this_rd_stats.rate;
+    this_distortion = this_rd_stats.dist;
+    s = this_rd_stats.skip;
+
+    if (this_rate_tokenonly == INT_MAX) continue;
+
+    this_rate = this_rate_tokenonly + bmode_costs[mbmi->mode];
+
+    if (!xd->lossless[mbmi->segment_id] && mbmi->sb_type >= BLOCK_8X8) {
+      // super_block_yrd above includes the cost of the tx_size in the
+      // tokenonly rate, but for intra blocks, tx_size is always coded
+      // (prediction granularity), so we account for it in the full rate,
+      // not the tokenonly rate.
+      this_rate_tokenonly -= tx_size_cost(cpi, x, bsize, mbmi->tx_size);
+    }
+#if CONFIG_PALETTE
+    if (try_palette && mbmi->mode == DC_PRED) {
+      this_rate +=
+          av1_cost_bit(av1_default_palette_y_mode_prob[bsize - BLOCK_8X8]
+                                                      [palette_y_mode_ctx],
+                       0);
+    }
+#endif  // CONFIG_PALETTE
+#if CONFIG_FILTER_INTRA
+    if (mbmi->mode == DC_PRED)
+      this_rate += av1_cost_bit(cpi->common.fc->filter_intra_probs[0], 0);
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_EXT_INTRA
+    if (is_directional_mode) {
+#if CONFIG_INTRA_INTERP
+      const int p_angle =
+          mode_to_angle_map[mbmi->mode] + mbmi->angle_delta[0] * ANGLE_STEP;
+      if (av1_is_intra_filter_switchable(p_angle))
+        this_rate +=
+            cpi->intra_filter_cost[intra_filter_ctx][mbmi->intra_filter];
+#endif  // CONFIG_INTRA_INTERP
+      this_rate += write_uniform_cost(2 * MAX_ANGLE_DELTA + 1,
+                                      MAX_ANGLE_DELTA + mbmi->angle_delta[0]);
+    }
+#endif  // CONFIG_EXT_INTRA
+    this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
+#if CONFIG_FILTER_INTRA
+    if (best_rd == INT64_MAX || this_rd - best_rd < (best_rd >> 4)) {
+      filter_intra_mode_skip_mask ^= (1 << mbmi->mode);
+    }
+#endif  // CONFIG_FILTER_INTRA
+
+    if (this_rd < best_rd) {
+      best_mbmi = *mbmi;
+      best_rd = this_rd;
+#if CONFIG_FILTER_INTRA
+      beat_best_rd = 1;
+#endif  // CONFIG_FILTER_INTRA
+      *rate = this_rate;
+      *rate_tokenonly = this_rate_tokenonly;
+      *distortion = this_distortion;
+      *skippable = s;
+#if CONFIG_PVQ
+      od_encode_checkpoint(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+    }
+  }
+
+#if CONFIG_PVQ
+  od_encode_rollback(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+
+#if CONFIG_CFL
+  // Perform one extra txfm_rd_in_plane() call, this time with the best value so
+  // we can store reconstructed luma values
+  RD_STATS this_rd_stats;
+  x->cfl_store_y = 1;
+  txfm_rd_in_plane(x, cpi, &this_rd_stats, INT64_MAX, 0, bsize,
+                   mic->mbmi.tx_size, cpi->sf.use_fast_coef_costing);
+  x->cfl_store_y = 0;
+#endif
+
+#if CONFIG_PALETTE
+  if (try_palette) {
+    rd_pick_palette_intra_sby(cpi, x, bsize, palette_y_mode_ctx,
+                              bmode_costs[DC_PRED], &best_mbmi,
+                              best_palette_color_map, &best_rd, &best_model_rd,
+                              rate, rate_tokenonly, distortion, skippable);
+  }
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_FILTER_INTRA
+  if (beat_best_rd) {
+    if (rd_pick_filter_intra_sby(cpi, x, rate, rate_tokenonly, distortion,
+                                 skippable, bsize, bmode_costs[DC_PRED],
+                                 &best_rd, &best_model_rd,
+                                 filter_intra_mode_skip_mask)) {
+      best_mbmi = *mbmi;
+    }
+  }
+#endif  // CONFIG_FILTER_INTRA
+
+  *mbmi = best_mbmi;
+  return best_rd;
+}
+
+// Return value 0: early termination triggered, no valid rd cost available;
+//              1: rd cost values are valid.
+static int super_block_uvrd(const AV1_COMP *const cpi, MACROBLOCK *x,
+                            RD_STATS *rd_stats, BLOCK_SIZE bsize,
+                            int64_t ref_best_rd) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const TX_SIZE uv_tx_size = get_uv_tx_size(mbmi, &xd->plane[1]);
+  int plane;
+  int is_cost_valid = 1;
+  av1_init_rd_stats(rd_stats);
+
+  if (ref_best_rd < 0) is_cost_valid = 0;
+
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+  if (x->skip_chroma_rd) return is_cost_valid;
+
+  bsize = scale_chroma_bsize(bsize, xd->plane[1].subsampling_x,
+                             xd->plane[1].subsampling_y);
+#endif  // CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+
+#if !CONFIG_PVQ
+  if (is_inter_block(mbmi) && is_cost_valid) {
+    for (plane = 1; plane < MAX_MB_PLANE; ++plane)
+      av1_subtract_plane(x, bsize, plane);
+  }
+#endif  // !CONFIG_PVQ
+
+  if (is_cost_valid) {
+    for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
+      RD_STATS pn_rd_stats;
+      txfm_rd_in_plane(x, cpi, &pn_rd_stats, ref_best_rd, plane, bsize,
+                       uv_tx_size, cpi->sf.use_fast_coef_costing);
+      if (pn_rd_stats.rate == INT_MAX) {
+        is_cost_valid = 0;
+        break;
+      }
+      av1_merge_rd_stats(rd_stats, &pn_rd_stats);
+      if (RDCOST(x->rdmult, x->rddiv, rd_stats->rate, rd_stats->dist) >
+              ref_best_rd &&
+          RDCOST(x->rdmult, x->rddiv, 0, rd_stats->sse) > ref_best_rd) {
+        is_cost_valid = 0;
+        break;
+      }
+    }
+  }
+
+  if (!is_cost_valid) {
+    // reset cost value
+    av1_invalid_rd_stats(rd_stats);
+  }
+
+  return is_cost_valid;
+}
+
+#if CONFIG_VAR_TX
+// FIXME crop these calls
+static uint64_t sum_squares_2d(const int16_t *diff, int diff_stride,
+                               TX_SIZE tx_size) {
+  return aom_sum_squares_2d_i16(diff, diff_stride, tx_size_wide[tx_size],
+                                tx_size_high[tx_size]);
+}
+
+void av1_tx_block_rd_b(const AV1_COMP *cpi, MACROBLOCK *x, TX_SIZE tx_size,
+                       int blk_row, int blk_col, int plane, int block,
+                       int plane_bsize, const ENTROPY_CONTEXT *a,
+                       const ENTROPY_CONTEXT *l, RD_STATS *rd_stats) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *xd = &x->e_mbd;
+  const struct macroblock_plane *const p = &x->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  int64_t tmp;
+  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+  PLANE_TYPE plane_type = get_plane_type(plane);
+  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(&xd->mi[0]->mbmi));
+  BLOCK_SIZE txm_bsize = txsize_to_bsize[tx_size];
+  int bh = block_size_high[txm_bsize];
+  int bw = block_size_wide[txm_bsize];
+  int txb_h = tx_size_high_unit[tx_size];
+  int txb_w = tx_size_wide_unit[tx_size];
+
+  int src_stride = p->src.stride;
+  uint8_t *src =
+      &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
+  uint8_t *dst =
+      &pd->dst
+           .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, rec_buffer16[MAX_TX_SQUARE]);
+  uint8_t *rec_buffer;
+#else
+  DECLARE_ALIGNED(16, uint8_t, rec_buffer[MAX_TX_SQUARE]);
+#endif  // CONFIG_HIGHBITDEPTH
+  int max_blocks_high = block_size_high[plane_bsize];
+  int max_blocks_wide = block_size_wide[plane_bsize];
+  const int diff_stride = max_blocks_wide;
+  const int16_t *diff =
+      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
+  int txb_coeff_cost;
+
+  assert(tx_size < TX_SIZES_ALL);
+
+  if (xd->mb_to_bottom_edge < 0)
+    max_blocks_high += xd->mb_to_bottom_edge >> (3 + pd->subsampling_y);
+  if (xd->mb_to_right_edge < 0)
+    max_blocks_wide += xd->mb_to_right_edge >> (3 + pd->subsampling_x);
+
+  max_blocks_high >>= tx_size_wide_log2[0];
+  max_blocks_wide >>= tx_size_wide_log2[0];
+
+  int coeff_ctx = get_entropy_context(tx_size, a, l);
+
+  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
+                  coeff_ctx, AV1_XFORM_QUANT_FP);
+
+  av1_optimize_b(cm, x, plane, block, tx_size, coeff_ctx);
+
+// TODO(any): Use av1_dist_block to compute distortion
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    rec_buffer = CONVERT_TO_BYTEPTR(rec_buffer16);
+    aom_highbd_convolve_copy(dst, pd->dst.stride, rec_buffer, MAX_TX_SIZE, NULL,
+                             0, NULL, 0, bw, bh, xd->bd);
+  } else {
+    rec_buffer = (uint8_t *)rec_buffer16;
+    aom_convolve_copy(dst, pd->dst.stride, rec_buffer, MAX_TX_SIZE, NULL, 0,
+                      NULL, 0, bw, bh);
+  }
+#else
+  aom_convolve_copy(dst, pd->dst.stride, rec_buffer, MAX_TX_SIZE, NULL, 0, NULL,
+                    0, bw, bh);
+#endif  // CONFIG_HIGHBITDEPTH
+
+  if (blk_row + txb_h > max_blocks_high || blk_col + txb_w > max_blocks_wide) {
+    int idx, idy;
+    int blocks_height = AOMMIN(txb_h, max_blocks_high - blk_row);
+    int blocks_width = AOMMIN(txb_w, max_blocks_wide - blk_col);
+    tmp = 0;
+    for (idy = 0; idy < blocks_height; ++idy) {
+      for (idx = 0; idx < blocks_width; ++idx) {
+        const int16_t *d =
+            diff + ((idy * diff_stride + idx) << tx_size_wide_log2[0]);
+        tmp += sum_squares_2d(d, diff_stride, 0);
+      }
+    }
+  } else {
+    tmp = sum_squares_2d(diff, diff_stride, tx_size);
+  }
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+    tmp = ROUND_POWER_OF_TWO(tmp, (xd->bd - 8) * 2);
+#endif  // CONFIG_HIGHBITDEPTH
+  rd_stats->sse += tmp * 16;
+  const int eob = p->eobs[block];
+
+  av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, rec_buffer,
+                              MAX_TX_SIZE, eob);
+  if (eob > 0) {
+    if (txb_w + blk_col > max_blocks_wide ||
+        txb_h + blk_row > max_blocks_high) {
+      int idx, idy;
+      unsigned int this_dist;
+      int blocks_height = AOMMIN(txb_h, max_blocks_high - blk_row);
+      int blocks_width = AOMMIN(txb_w, max_blocks_wide - blk_col);
+      tmp = 0;
+      for (idy = 0; idy < blocks_height; ++idy) {
+        for (idx = 0; idx < blocks_width; ++idx) {
+          uint8_t *const s =
+              src + ((idy * src_stride + idx) << tx_size_wide_log2[0]);
+          uint8_t *const r =
+              rec_buffer + ((idy * MAX_TX_SIZE + idx) << tx_size_wide_log2[0]);
+          cpi->fn_ptr[0].vf(s, src_stride, r, MAX_TX_SIZE, &this_dist);
+          tmp += this_dist;
+        }
+      }
+    } else {
+      uint32_t this_dist;
+      cpi->fn_ptr[txm_bsize].vf(src, src_stride, rec_buffer, MAX_TX_SIZE,
+                                &this_dist);
+      tmp = this_dist;
+    }
+  }
+  rd_stats->dist += tmp * 16;
+  txb_coeff_cost =
+      av1_cost_coeffs(cpi, x, plane, block, tx_size, scan_order, a, l, 0);
+  rd_stats->rate += txb_coeff_cost;
+  rd_stats->skip &= (eob == 0);
+
+#if CONFIG_RD_DEBUG
+  av1_update_txb_coeff_cost(rd_stats, plane, tx_size, blk_row, blk_col,
+                            txb_coeff_cost);
+#endif  // CONFIG_RD_DEBUG
+}
+
+static void select_tx_block(const AV1_COMP *cpi, MACROBLOCK *x, int blk_row,
+                            int blk_col, int plane, int block, int block32,
+                            TX_SIZE tx_size, int depth, BLOCK_SIZE plane_bsize,
+                            ENTROPY_CONTEXT *ta, ENTROPY_CONTEXT *tl,
+                            TXFM_CONTEXT *tx_above, TXFM_CONTEXT *tx_left,
+                            RD_STATS *rd_stats, int64_t ref_best_rd,
+                            int *is_cost_valid, RD_STATS *rd_stats_stack) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  struct macroblock_plane *const p = &x->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const int tx_row = blk_row >> (1 - pd->subsampling_y);
+  const int tx_col = blk_col >> (1 - pd->subsampling_x);
+  TX_SIZE(*const inter_tx_size)
+  [MAX_MIB_SIZE] =
+      (TX_SIZE(*)[MAX_MIB_SIZE]) & mbmi->inter_tx_size[tx_row][tx_col];
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+  const int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+  int64_t this_rd = INT64_MAX;
+  ENTROPY_CONTEXT *pta = ta + blk_col;
+  ENTROPY_CONTEXT *ptl = tl + blk_row;
+  int coeff_ctx, i;
+  int ctx =
+      txfm_partition_context(tx_above + (blk_col >> 1),
+                             tx_left + (blk_row >> 1), mbmi->sb_type, tx_size);
+  int64_t sum_rd = INT64_MAX;
+  int tmp_eob = 0;
+  int zero_blk_rate;
+  RD_STATS sum_rd_stats;
+  const int tx_size_ctx = txsize_sqr_map[tx_size];
+
+  av1_init_rd_stats(&sum_rd_stats);
+
+  assert(tx_size < TX_SIZES_ALL);
+
+  if (ref_best_rd < 0) {
+    *is_cost_valid = 0;
+    return;
+  }
+
+  coeff_ctx = get_entropy_context(tx_size, pta, ptl);
+
+  av1_init_rd_stats(rd_stats);
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  zero_blk_rate = x->token_costs[tx_size_ctx][pd->plane_type][1][0][0]
+                                [coeff_ctx][EOB_TOKEN];
+
+  if (cpi->common.tx_mode == TX_MODE_SELECT || tx_size == TX_4X4) {
+    inter_tx_size[0][0] = tx_size;
+
+    if (tx_size == TX_32X32 && mbmi->tx_type != DCT_DCT &&
+        rd_stats_stack[block32].rate != INT_MAX) {
+      *rd_stats = rd_stats_stack[block32];
+      p->eobs[block] = !rd_stats->skip;
+      x->blk_skip[plane][blk_row * bw + blk_col] = rd_stats->skip;
+    } else {
+      av1_tx_block_rd_b(cpi, x, tx_size, blk_row, blk_col, plane, block,
+                        plane_bsize, pta, ptl, rd_stats);
+      if (tx_size == TX_32X32) {
+        rd_stats_stack[block32] = *rd_stats;
+      }
+    }
+
+    if ((RDCOST(x->rdmult, x->rddiv, rd_stats->rate, rd_stats->dist) >=
+             RDCOST(x->rdmult, x->rddiv, zero_blk_rate, rd_stats->sse) ||
+         rd_stats->skip == 1) &&
+        !xd->lossless[mbmi->segment_id]) {
+#if CONFIG_RD_DEBUG
+      av1_update_txb_coeff_cost(rd_stats, plane, tx_size, blk_row, blk_col,
+                                zero_blk_rate - rd_stats->rate);
+#endif  // CONFIG_RD_DEBUG
+      rd_stats->rate = zero_blk_rate;
+      rd_stats->dist = rd_stats->sse;
+      rd_stats->skip = 1;
+      x->blk_skip[plane][blk_row * bw + blk_col] = 1;
+      p->eobs[block] = 0;
+    } else {
+      x->blk_skip[plane][blk_row * bw + blk_col] = 0;
+      rd_stats->skip = 0;
+    }
+
+    if (tx_size > TX_4X4 && depth < MAX_VARTX_DEPTH)
+      rd_stats->rate +=
+          av1_cost_bit(cpi->common.fc->txfm_partition_prob[ctx], 0);
+    this_rd = RDCOST(x->rdmult, x->rddiv, rd_stats->rate, rd_stats->dist);
+    tmp_eob = p->eobs[block];
+  }
+
+  if (tx_size > TX_4X4 && depth < MAX_VARTX_DEPTH) {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int sub_step = tx_size_wide_unit[sub_txs] * tx_size_high_unit[sub_txs];
+    RD_STATS this_rd_stats;
+    int this_cost_valid = 1;
+    int64_t tmp_rd = 0;
+
+    sum_rd_stats.rate =
+        av1_cost_bit(cpi->common.fc->txfm_partition_prob[ctx], 1);
+
+    assert(tx_size < TX_SIZES_ALL);
+
+    for (i = 0; i < 4 && this_cost_valid; ++i) {
+      int offsetr = blk_row + (i >> 1) * bsl;
+      int offsetc = blk_col + (i & 0x01) * bsl;
+
+      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
+
+      select_tx_block(cpi, x, offsetr, offsetc, plane, block, block32, sub_txs,
+                      depth + 1, plane_bsize, ta, tl, tx_above, tx_left,
+                      &this_rd_stats, ref_best_rd - tmp_rd, &this_cost_valid,
+                      rd_stats_stack);
+
+      av1_merge_rd_stats(&sum_rd_stats, &this_rd_stats);
+
+      tmp_rd =
+          RDCOST(x->rdmult, x->rddiv, sum_rd_stats.rate, sum_rd_stats.dist);
+      if (this_rd < tmp_rd) break;
+      block += sub_step;
+    }
+    if (this_cost_valid) sum_rd = tmp_rd;
+  }
+
+  if (this_rd < sum_rd) {
+    int idx, idy;
+    for (i = 0; i < tx_size_wide_unit[tx_size]; ++i) pta[i] = !(tmp_eob == 0);
+    for (i = 0; i < tx_size_high_unit[tx_size]; ++i) ptl[i] = !(tmp_eob == 0);
+    txfm_partition_update(tx_above + (blk_col >> 1), tx_left + (blk_row >> 1),
+                          tx_size, tx_size);
+    inter_tx_size[0][0] = tx_size;
+    for (idy = 0; idy < tx_size_high_unit[tx_size] / 2; ++idy)
+      for (idx = 0; idx < tx_size_wide_unit[tx_size] / 2; ++idx)
+        inter_tx_size[idy][idx] = tx_size;
+    mbmi->tx_size = tx_size;
+    if (this_rd == INT64_MAX) *is_cost_valid = 0;
+    x->blk_skip[plane][blk_row * bw + blk_col] = rd_stats->skip;
+  } else {
+    *rd_stats = sum_rd_stats;
+    if (sum_rd == INT64_MAX) *is_cost_valid = 0;
+  }
+}
+
+static void inter_block_yrd(const AV1_COMP *cpi, MACROBLOCK *x,
+                            RD_STATS *rd_stats, BLOCK_SIZE bsize,
+                            int64_t ref_best_rd, RD_STATS *rd_stats_stack) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  int is_cost_valid = 1;
+  int64_t this_rd = 0;
+
+  if (ref_best_rd < 0) is_cost_valid = 0;
+
+  av1_init_rd_stats(rd_stats);
+
+  if (is_cost_valid) {
+    const struct macroblockd_plane *const pd = &xd->plane[0];
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+    const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+    const int mi_height = block_size_high[plane_bsize] >> tx_size_high_log2[0];
+    const TX_SIZE max_tx_size = max_txsize_rect_lookup[plane_bsize];
+    const int bh = tx_size_high_unit[max_tx_size];
+    const int bw = tx_size_wide_unit[max_tx_size];
+    int idx, idy;
+    int block = 0;
+    int block32 = 0;
+    int step = tx_size_wide_unit[max_tx_size] * tx_size_high_unit[max_tx_size];
+    ENTROPY_CONTEXT ctxa[2 * MAX_MIB_SIZE];
+    ENTROPY_CONTEXT ctxl[2 * MAX_MIB_SIZE];
+    TXFM_CONTEXT tx_above[MAX_MIB_SIZE];
+    TXFM_CONTEXT tx_left[MAX_MIB_SIZE];
+
+    RD_STATS pn_rd_stats;
+    av1_init_rd_stats(&pn_rd_stats);
+
+    av1_get_entropy_contexts(bsize, 0, pd, ctxa, ctxl);
+    memcpy(tx_above, xd->above_txfm_context,
+           sizeof(TXFM_CONTEXT) * (mi_width >> 1));
+    memcpy(tx_left, xd->left_txfm_context,
+           sizeof(TXFM_CONTEXT) * (mi_height >> 1));
+
+    for (idy = 0; idy < mi_height; idy += bh) {
+      for (idx = 0; idx < mi_width; idx += bw) {
+        select_tx_block(cpi, x, idy, idx, 0, block, block32, max_tx_size,
+                        mi_height != mi_width, plane_bsize, ctxa, ctxl,
+                        tx_above, tx_left, &pn_rd_stats, ref_best_rd - this_rd,
+                        &is_cost_valid, rd_stats_stack);
+        av1_merge_rd_stats(rd_stats, &pn_rd_stats);
+        this_rd += AOMMIN(
+            RDCOST(x->rdmult, x->rddiv, pn_rd_stats.rate, pn_rd_stats.dist),
+            RDCOST(x->rdmult, x->rddiv, 0, pn_rd_stats.sse));
+        block += step;
+        ++block32;
+      }
+    }
+  }
+
+  this_rd = AOMMIN(RDCOST(x->rdmult, x->rddiv, rd_stats->rate, rd_stats->dist),
+                   RDCOST(x->rdmult, x->rddiv, 0, rd_stats->sse));
+  if (this_rd > ref_best_rd) is_cost_valid = 0;
+
+  if (!is_cost_valid) {
+    // reset cost value
+    av1_invalid_rd_stats(rd_stats);
+  }
+}
+
+static int64_t select_tx_size_fix_type(const AV1_COMP *cpi, MACROBLOCK *x,
+                                       RD_STATS *rd_stats, BLOCK_SIZE bsize,
+                                       int64_t ref_best_rd, TX_TYPE tx_type,
+                                       RD_STATS *rd_stats_stack) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int is_inter = is_inter_block(mbmi);
+  aom_prob skip_prob = av1_get_skip_prob(cm, xd);
+  int s0 = av1_cost_bit(skip_prob, 0);
+  int s1 = av1_cost_bit(skip_prob, 1);
+  int64_t rd;
+  int row, col;
+  const int max_blocks_high = max_block_high(xd, bsize, 0);
+  const int max_blocks_wide = max_block_wide(xd, bsize, 0);
+
+  mbmi->tx_type = tx_type;
+  mbmi->min_tx_size = TX_SIZES_ALL;
+  inter_block_yrd(cpi, x, rd_stats, bsize, ref_best_rd, rd_stats_stack);
+
+  if (rd_stats->rate == INT_MAX) return INT64_MAX;
+
+  for (row = 0; row < max_blocks_high / 2; ++row)
+    for (col = 0; col < max_blocks_wide / 2; ++col)
+      mbmi->min_tx_size = AOMMIN(
+          mbmi->min_tx_size, get_min_tx_size(mbmi->inter_tx_size[row][col]));
+
+#if CONFIG_EXT_TX
+  if (get_ext_tx_types(mbmi->min_tx_size, bsize, is_inter,
+                       cm->reduced_tx_set_used) > 1 &&
+      !xd->lossless[xd->mi[0]->mbmi.segment_id]) {
+    const int ext_tx_set = get_ext_tx_set(mbmi->min_tx_size, bsize, is_inter,
+                                          cm->reduced_tx_set_used);
+    if (is_inter) {
+      if (ext_tx_set > 0)
+        rd_stats->rate +=
+            cpi->inter_tx_type_costs[ext_tx_set]
+                                    [txsize_sqr_map[mbmi->min_tx_size]]
+                                    [mbmi->tx_type];
+    } else {
+      if (ext_tx_set > 0 && ALLOW_INTRA_EXT_TX)
+        rd_stats->rate +=
+            cpi->intra_tx_type_costs[ext_tx_set][mbmi->min_tx_size][mbmi->mode]
+                                    [mbmi->tx_type];
+    }
+  }
+#else   // CONFIG_EXT_TX
+  if (mbmi->min_tx_size < TX_32X32 && !xd->lossless[xd->mi[0]->mbmi.segment_id])
+    rd_stats->rate +=
+        cpi->inter_tx_type_costs[mbmi->min_tx_size][mbmi->tx_type];
+#endif  // CONFIG_EXT_TX
+
+  if (rd_stats->skip)
+    rd = RDCOST(x->rdmult, x->rddiv, s1, rd_stats->sse);
+  else
+    rd = RDCOST(x->rdmult, x->rddiv, rd_stats->rate + s0, rd_stats->dist);
+
+  if (is_inter && !xd->lossless[xd->mi[0]->mbmi.segment_id] &&
+      !(rd_stats->skip))
+    rd = AOMMIN(rd, RDCOST(x->rdmult, x->rddiv, s1, rd_stats->sse));
+
+  return rd;
+}
+
+static void select_tx_type_yrd(const AV1_COMP *cpi, MACROBLOCK *x,
+                               RD_STATS *rd_stats, BLOCK_SIZE bsize,
+                               int64_t ref_best_rd) {
+  const AV1_COMMON *cm = &cpi->common;
+  const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int64_t rd = INT64_MAX;
+  int64_t best_rd = INT64_MAX;
+  TX_TYPE tx_type, best_tx_type = DCT_DCT;
+  const int is_inter = is_inter_block(mbmi);
+  TX_SIZE best_tx_size[MAX_MIB_SIZE][MAX_MIB_SIZE];
+  TX_SIZE best_tx = max_txsize_lookup[bsize];
+  TX_SIZE best_min_tx_size = TX_SIZES_ALL;
+  uint8_t best_blk_skip[MAX_MIB_SIZE * MAX_MIB_SIZE * 8];
+  const int n4 = bsize_to_num_blk(bsize);
+  int idx, idy;
+  int prune = 0;
+  const int count32 =
+      1 << (2 * (cm->mib_size_log2 - mi_width_log2_lookup[BLOCK_32X32]));
+#if CONFIG_EXT_PARTITION
+  RD_STATS rd_stats_stack[16];
+#else
+  RD_STATS rd_stats_stack[4];
+#endif  // CONFIG_EXT_PARTITION
+#if CONFIG_EXT_TX
+  const int ext_tx_set =
+      get_ext_tx_set(max_tx_size, bsize, is_inter, cm->reduced_tx_set_used);
+#endif  // CONFIG_EXT_TX
+
+  if (is_inter && cpi->sf.tx_type_search.prune_mode > NO_PRUNE)
+#if CONFIG_EXT_TX
+    prune = prune_tx_types(cpi, bsize, x, xd, ext_tx_set);
+#else
+    prune = prune_tx_types(cpi, bsize, x, xd, 0);
+#endif  // CONFIG_EXT_TX
+
+  av1_invalid_rd_stats(rd_stats);
+
+  for (idx = 0; idx < count32; ++idx)
+    av1_invalid_rd_stats(&rd_stats_stack[idx]);
+
+  for (tx_type = DCT_DCT; tx_type < TX_TYPES; ++tx_type) {
+    RD_STATS this_rd_stats;
+    av1_init_rd_stats(&this_rd_stats);
+#if CONFIG_EXT_TX
+    if (is_inter) {
+      if (!ext_tx_used_inter[ext_tx_set][tx_type]) continue;
+      if (cpi->sf.tx_type_search.prune_mode > NO_PRUNE) {
+        if (!do_tx_type_search(tx_type, prune)) continue;
+      }
+    } else {
+      if (!ALLOW_INTRA_EXT_TX && bsize >= BLOCK_8X8) {
+        if (tx_type != intra_mode_to_tx_type_context[mbmi->mode]) continue;
+      }
+      if (!ext_tx_used_intra[ext_tx_set][tx_type]) continue;
+    }
+#else   // CONFIG_EXT_TX
+    if (is_inter && cpi->sf.tx_type_search.prune_mode > NO_PRUNE &&
+        !do_tx_type_search(tx_type, prune))
+      continue;
+#endif  // CONFIG_EXT_TX
+    if (is_inter && x->use_default_inter_tx_type &&
+        tx_type != get_default_tx_type(0, xd, 0, max_tx_size))
+      continue;
+
+    if (xd->lossless[mbmi->segment_id])
+      if (tx_type != DCT_DCT) continue;
+
+    rd = select_tx_size_fix_type(cpi, x, &this_rd_stats, bsize, ref_best_rd,
+                                 tx_type, rd_stats_stack);
+
+    if (rd < best_rd) {
+      best_rd = rd;
+      *rd_stats = this_rd_stats;
+      best_tx_type = mbmi->tx_type;
+      best_tx = mbmi->tx_size;
+      best_min_tx_size = mbmi->min_tx_size;
+      memcpy(best_blk_skip, x->blk_skip[0], sizeof(best_blk_skip[0]) * n4);
+      for (idy = 0; idy < xd->n8_h; ++idy)
+        for (idx = 0; idx < xd->n8_w; ++idx)
+          best_tx_size[idy][idx] = mbmi->inter_tx_size[idy][idx];
+    }
+  }
+
+  mbmi->tx_type = best_tx_type;
+  for (idy = 0; idy < xd->n8_h; ++idy)
+    for (idx = 0; idx < xd->n8_w; ++idx)
+      mbmi->inter_tx_size[idy][idx] = best_tx_size[idy][idx];
+  mbmi->tx_size = best_tx;
+  mbmi->min_tx_size = best_min_tx_size;
+  memcpy(x->blk_skip[0], best_blk_skip, sizeof(best_blk_skip[0]) * n4);
+}
+
+static void tx_block_rd(const AV1_COMP *cpi, MACROBLOCK *x, int blk_row,
+                        int blk_col, int plane, int block, TX_SIZE tx_size,
+                        BLOCK_SIZE plane_bsize, ENTROPY_CONTEXT *above_ctx,
+                        ENTROPY_CONTEXT *left_ctx, RD_STATS *rd_stats) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  struct macroblock_plane *const p = &x->plane[plane];
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
+  const int tx_row = blk_row >> (1 - pd->subsampling_y);
+  const int tx_col = blk_col >> (1 - pd->subsampling_x);
+  TX_SIZE plane_tx_size;
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+
+  assert(tx_size < TX_SIZES_ALL);
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  plane_tx_size =
+      plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
+            : mbmi->inter_tx_size[tx_row][tx_col];
+
+  if (tx_size == plane_tx_size) {
+    int i;
+    ENTROPY_CONTEXT *ta = above_ctx + blk_col;
+    ENTROPY_CONTEXT *tl = left_ctx + blk_row;
+    av1_tx_block_rd_b(cpi, x, tx_size, blk_row, blk_col, plane, block,
+                      plane_bsize, ta, tl, rd_stats);
+
+    for (i = 0; i < tx_size_wide_unit[tx_size]; ++i)
+      ta[i] = !(p->eobs[block] == 0);
+    for (i = 0; i < tx_size_high_unit[tx_size]; ++i)
+      tl[i] = !(p->eobs[block] == 0);
+  } else {
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int step = tx_size_wide_unit[sub_txs] * tx_size_high_unit[sub_txs];
+    int i;
+
+    assert(bsl > 0);
+
+    for (i = 0; i < 4; ++i) {
+      int offsetr = blk_row + (i >> 1) * bsl;
+      int offsetc = blk_col + (i & 0x01) * bsl;
+
+      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
+
+      tx_block_rd(cpi, x, offsetr, offsetc, plane, block, sub_txs, plane_bsize,
+                  above_ctx, left_ctx, rd_stats);
+      block += step;
+    }
+  }
+}
+
+// Return value 0: early termination triggered, no valid rd cost available;
+//              1: rd cost values are valid.
+static int inter_block_uvrd(const AV1_COMP *cpi, MACROBLOCK *x,
+                            RD_STATS *rd_stats, BLOCK_SIZE bsize,
+                            int64_t ref_best_rd) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int plane;
+  int is_cost_valid = 1;
+  int64_t this_rd;
+
+  if (ref_best_rd < 0) is_cost_valid = 0;
+
+  av1_init_rd_stats(rd_stats);
+
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+  if (x->skip_chroma_rd) return is_cost_valid;
+  bsize = AOMMAX(BLOCK_8X8, bsize);
+#endif  // CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+  if (is_rect_tx(mbmi->tx_size)) {
+    return super_block_uvrd(cpi, x, rd_stats, bsize, ref_best_rd);
+  }
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+
+  if (is_inter_block(mbmi) && is_cost_valid) {
+    for (plane = 1; plane < MAX_MB_PLANE; ++plane)
+      av1_subtract_plane(x, bsize, plane);
+  }
+
+  for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
+    const struct macroblockd_plane *const pd = &xd->plane[plane];
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+    const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+    const int mi_height = block_size_high[plane_bsize] >> tx_size_high_log2[0];
+    const TX_SIZE max_tx_size = max_txsize_rect_lookup[plane_bsize];
+    const int bh = tx_size_high_unit[max_tx_size];
+    const int bw = tx_size_wide_unit[max_tx_size];
+    int idx, idy;
+    int block = 0;
+    const int step = bh * bw;
+    ENTROPY_CONTEXT ta[2 * MAX_MIB_SIZE];
+    ENTROPY_CONTEXT tl[2 * MAX_MIB_SIZE];
+    RD_STATS pn_rd_stats;
+    av1_init_rd_stats(&pn_rd_stats);
+
+    av1_get_entropy_contexts(bsize, 0, pd, ta, tl);
+
+    for (idy = 0; idy < mi_height; idy += bh) {
+      for (idx = 0; idx < mi_width; idx += bw) {
+        tx_block_rd(cpi, x, idy, idx, plane, block, max_tx_size, plane_bsize,
+                    ta, tl, &pn_rd_stats);
+        block += step;
+      }
+    }
+
+    if (pn_rd_stats.rate == INT_MAX) {
+      is_cost_valid = 0;
+      break;
+    }
+
+    av1_merge_rd_stats(rd_stats, &pn_rd_stats);
+
+    this_rd =
+        AOMMIN(RDCOST(x->rdmult, x->rddiv, rd_stats->rate, rd_stats->dist),
+               RDCOST(x->rdmult, x->rddiv, 0, rd_stats->sse));
+
+    if (this_rd > ref_best_rd) {
+      is_cost_valid = 0;
+      break;
+    }
+  }
+
+  if (!is_cost_valid) {
+    // reset cost value
+    av1_invalid_rd_stats(rd_stats);
+  }
+
+  return is_cost_valid;
+}
+#endif  // CONFIG_VAR_TX
+
+#if CONFIG_PALETTE
+static void rd_pick_palette_intra_sbuv(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                       int dc_mode_cost,
+                                       uint8_t *best_palette_color_map,
+                                       MB_MODE_INFO *const best_mbmi,
+                                       int64_t *best_rd, int *rate,
+                                       int *rate_tokenonly, int64_t *distortion,
+                                       int *skippable) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  int this_rate;
+  int64_t this_rd;
+  int colors_u, colors_v, colors;
+  const int src_stride = x->plane[1].src.stride;
+  const uint8_t *const src_u = x->plane[1].src.buf;
+  const uint8_t *const src_v = x->plane[2].src.buf;
+  uint8_t *const color_map = xd->plane[1].color_index_map;
+  RD_STATS tokenonly_rd_stats;
+  int plane_block_width, plane_block_height, rows, cols;
+  av1_get_block_dimensions(bsize, 1, xd, &plane_block_width,
+                           &plane_block_height, &rows, &cols);
+  if (rows * cols > PALETTE_MAX_BLOCK_SIZE) return;
+
+  mbmi->uv_mode = DC_PRED;
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_HIGHBITDEPTH
+  if (cpi->common.use_highbitdepth) {
+    colors_u = av1_count_colors_highbd(src_u, src_stride, rows, cols,
+                                       cpi->common.bit_depth);
+    colors_v = av1_count_colors_highbd(src_v, src_stride, rows, cols,
+                                       cpi->common.bit_depth);
+  } else {
+#endif  // CONFIG_HIGHBITDEPTH
+    colors_u = av1_count_colors(src_u, src_stride, rows, cols);
+    colors_v = av1_count_colors(src_v, src_stride, rows, cols);
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  colors = colors_u > colors_v ? colors_u : colors_v;
+  if (colors > 1 && colors <= 64) {
+    int r, c, n, i, j;
+    const int max_itr = 50;
+    uint8_t color_order[PALETTE_MAX_SIZE];
+    float lb_u, ub_u, val_u;
+    float lb_v, ub_v, val_v;
+    float *const data = x->palette_buffer->kmeans_data_buf;
+    float centroids[2 * PALETTE_MAX_SIZE];
+
+#if CONFIG_HIGHBITDEPTH
+    uint16_t *src_u16 = CONVERT_TO_SHORTPTR(src_u);
+    uint16_t *src_v16 = CONVERT_TO_SHORTPTR(src_v);
+    if (cpi->common.use_highbitdepth) {
+      lb_u = src_u16[0];
+      ub_u = src_u16[0];
+      lb_v = src_v16[0];
+      ub_v = src_v16[0];
+    } else {
+#endif  // CONFIG_HIGHBITDEPTH
+      lb_u = src_u[0];
+      ub_u = src_u[0];
+      lb_v = src_v[0];
+      ub_v = src_v[0];
+#if CONFIG_HIGHBITDEPTH
+    }
+#endif  // CONFIG_HIGHBITDEPTH
+
+    for (r = 0; r < rows; ++r) {
+      for (c = 0; c < cols; ++c) {
+#if CONFIG_HIGHBITDEPTH
+        if (cpi->common.use_highbitdepth) {
+          val_u = src_u16[r * src_stride + c];
+          val_v = src_v16[r * src_stride + c];
+          data[(r * cols + c) * 2] = val_u;
+          data[(r * cols + c) * 2 + 1] = val_v;
+        } else {
+#endif  // CONFIG_HIGHBITDEPTH
+          val_u = src_u[r * src_stride + c];
+          val_v = src_v[r * src_stride + c];
+          data[(r * cols + c) * 2] = val_u;
+          data[(r * cols + c) * 2 + 1] = val_v;
+#if CONFIG_HIGHBITDEPTH
+        }
+#endif  // CONFIG_HIGHBITDEPTH
+        if (val_u < lb_u)
+          lb_u = val_u;
+        else if (val_u > ub_u)
+          ub_u = val_u;
+        if (val_v < lb_v)
+          lb_v = val_v;
+        else if (val_v > ub_v)
+          ub_v = val_v;
+      }
+    }
+
+    for (n = colors > PALETTE_MAX_SIZE ? PALETTE_MAX_SIZE : colors; n >= 2;
+         --n) {
+      for (i = 0; i < n; ++i) {
+        centroids[i * 2] = lb_u + (2 * i + 1) * (ub_u - lb_u) / n / 2;
+        centroids[i * 2 + 1] = lb_v + (2 * i + 1) * (ub_v - lb_v) / n / 2;
+      }
+      av1_k_means(data, centroids, color_map, rows * cols, n, 2, max_itr);
+#if CONFIG_PALETTE_DELTA_ENCODING
+      // Sort the U channel colors in ascending order.
+      for (i = 0; i < 2 * (n - 1); i += 2) {
+        int min_idx = i;
+        float min_val = centroids[i];
+        for (j = i + 2; j < 2 * n; j += 2)
+          if (centroids[j] < min_val) min_val = centroids[j], min_idx = j;
+        if (min_idx != i) {
+          float temp_u = centroids[i], temp_v = centroids[i + 1];
+          centroids[i] = centroids[min_idx];
+          centroids[i + 1] = centroids[min_idx + 1];
+          centroids[min_idx] = temp_u, centroids[min_idx + 1] = temp_v;
+        }
+      }
+      av1_calc_indices(data, centroids, color_map, rows * cols, n, 2);
+#endif  // CONFIG_PALETTE_DELTA_ENCODING
+      extend_palette_color_map(color_map, cols, rows, plane_block_width,
+                               plane_block_height);
+      pmi->palette_size[1] = n;
+      for (i = 1; i < 3; ++i) {
+        for (j = 0; j < n; ++j) {
+#if CONFIG_HIGHBITDEPTH
+          if (cpi->common.use_highbitdepth)
+            pmi->palette_colors[i * PALETTE_MAX_SIZE + j] = clip_pixel_highbd(
+                (int)centroids[j * 2 + i - 1], cpi->common.bit_depth);
+          else
+#endif  // CONFIG_HIGHBITDEPTH
+            pmi->palette_colors[i * PALETTE_MAX_SIZE + j] =
+                clip_pixel((int)centroids[j * 2 + i - 1]);
+        }
+      }
+
+      super_block_uvrd(cpi, x, &tokenonly_rd_stats, bsize, *best_rd);
+      if (tokenonly_rd_stats.rate == INT_MAX) continue;
+      this_rate =
+          tokenonly_rd_stats.rate + dc_mode_cost +
+          cpi->palette_uv_size_cost[bsize - BLOCK_8X8][n - PALETTE_MIN_SIZE] +
+          write_uniform_cost(n, color_map[0]) +
+          av1_cost_bit(
+              av1_default_palette_uv_mode_prob[pmi->palette_size[0] > 0], 1);
+      this_rate += av1_palette_color_cost_uv(pmi, cpi->common.bit_depth);
+      for (i = 0; i < rows; ++i) {
+        for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
+          int color_idx;
+          const int color_ctx = av1_get_palette_color_index_context(
+              color_map, plane_block_width, i, j, n, color_order, &color_idx);
+          assert(color_idx >= 0 && color_idx < n);
+          this_rate += cpi->palette_uv_color_cost[n - PALETTE_MIN_SIZE]
+                                                 [color_ctx][color_idx];
+        }
+      }
+
+      this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, tokenonly_rd_stats.dist);
+      if (this_rd < *best_rd) {
+        *best_rd = this_rd;
+        *best_mbmi = *mbmi;
+        memcpy(best_palette_color_map, color_map,
+               plane_block_width * plane_block_height *
+                   sizeof(best_palette_color_map[0]));
+        *rate = this_rate;
+        *distortion = tokenonly_rd_stats.dist;
+        *rate_tokenonly = tokenonly_rd_stats.rate;
+        *skippable = tokenonly_rd_stats.skip;
+      }
+    }
+  }
+  if (best_mbmi->palette_mode_info.palette_size[1] > 0) {
+    memcpy(color_map, best_palette_color_map,
+           rows * cols * sizeof(best_palette_color_map[0]));
+  }
+}
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_FILTER_INTRA
+// Return 1 if an filter intra mode is selected; return 0 otherwise.
+static int rd_pick_filter_intra_sbuv(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                     int *rate, int *rate_tokenonly,
+                                     int64_t *distortion, int *skippable,
+                                     BLOCK_SIZE bsize, int64_t *best_rd) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  int filter_intra_selected_flag = 0;
+  int this_rate;
+  int64_t this_rd;
+  FILTER_INTRA_MODE mode;
+  FILTER_INTRA_MODE_INFO filter_intra_mode_info;
+  RD_STATS tokenonly_rd_stats;
+
+  av1_zero(filter_intra_mode_info);
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 1;
+  mbmi->uv_mode = DC_PRED;
+#if CONFIG_PALETTE
+  mbmi->palette_mode_info.palette_size[1] = 0;
+#endif  // CONFIG_PALETTE
+
+  for (mode = 0; mode < FILTER_INTRA_MODES; ++mode) {
+    mbmi->filter_intra_mode_info.filter_intra_mode[1] = mode;
+    if (!super_block_uvrd(cpi, x, &tokenonly_rd_stats, bsize, *best_rd))
+      continue;
+
+    this_rate = tokenonly_rd_stats.rate +
+                av1_cost_bit(cpi->common.fc->filter_intra_probs[1], 1) +
+                cpi->intra_uv_mode_cost[mbmi->mode][mbmi->uv_mode] +
+                write_uniform_cost(FILTER_INTRA_MODES, mode);
+    this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, tokenonly_rd_stats.dist);
+    if (this_rd < *best_rd) {
+      *best_rd = this_rd;
+      *rate = this_rate;
+      *rate_tokenonly = tokenonly_rd_stats.rate;
+      *distortion = tokenonly_rd_stats.dist;
+      *skippable = tokenonly_rd_stats.skip;
+      filter_intra_mode_info = mbmi->filter_intra_mode_info;
+      filter_intra_selected_flag = 1;
+    }
+  }
+
+  if (filter_intra_selected_flag) {
+    mbmi->uv_mode = DC_PRED;
+    mbmi->filter_intra_mode_info.use_filter_intra_mode[1] =
+        filter_intra_mode_info.use_filter_intra_mode[1];
+    mbmi->filter_intra_mode_info.filter_intra_mode[1] =
+        filter_intra_mode_info.filter_intra_mode[1];
+    return 1;
+  } else {
+    return 0;
+  }
+}
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_EXT_INTRA
+// Run RD calculation with given chroma intra prediction angle., and return
+// the RD cost. Update the best mode info. if the RD cost is the best so far.
+static int64_t pick_intra_angle_routine_sbuv(
+    const AV1_COMP *const cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
+    int rate_overhead, int64_t best_rd_in, int *rate, RD_STATS *rd_stats,
+    int *best_angle_delta, int64_t *best_rd) {
+  MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
+  int this_rate;
+  int64_t this_rd;
+  RD_STATS tokenonly_rd_stats;
+
+  if (!super_block_uvrd(cpi, x, &tokenonly_rd_stats, bsize, best_rd_in))
+    return INT64_MAX;
+  this_rate = tokenonly_rd_stats.rate + rate_overhead;
+  this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, tokenonly_rd_stats.dist);
+  if (this_rd < *best_rd) {
+    *best_rd = this_rd;
+    *best_angle_delta = mbmi->angle_delta[1];
+    *rate = this_rate;
+    rd_stats->rate = tokenonly_rd_stats.rate;
+    rd_stats->dist = tokenonly_rd_stats.dist;
+    rd_stats->skip = tokenonly_rd_stats.skip;
+  }
+  return this_rd;
+}
+
+// With given chroma directional intra prediction mode, pick the best angle
+// delta. Return true if a RD cost that is smaller than the input one is found.
+static int rd_pick_intra_angle_sbuv(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                    BLOCK_SIZE bsize, int rate_overhead,
+                                    int64_t best_rd, int *rate,
+                                    RD_STATS *rd_stats) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  int i, angle_delta, best_angle_delta = 0;
+  int64_t this_rd, best_rd_in, rd_cost[2 * (MAX_ANGLE_DELTA + 2)];
+
+  rd_stats->rate = INT_MAX;
+  rd_stats->skip = 0;
+  rd_stats->dist = INT64_MAX;
+  for (i = 0; i < 2 * (MAX_ANGLE_DELTA + 2); ++i) rd_cost[i] = INT64_MAX;
+
+  for (angle_delta = 0; angle_delta <= MAX_ANGLE_DELTA; angle_delta += 2) {
+    for (i = 0; i < 2; ++i) {
+      best_rd_in = (best_rd == INT64_MAX)
+                       ? INT64_MAX
+                       : (best_rd + (best_rd >> ((angle_delta == 0) ? 3 : 5)));
+      mbmi->angle_delta[1] = (1 - 2 * i) * angle_delta;
+      this_rd = pick_intra_angle_routine_sbuv(cpi, x, bsize, rate_overhead,
+                                              best_rd_in, rate, rd_stats,
+                                              &best_angle_delta, &best_rd);
+      rd_cost[2 * angle_delta + i] = this_rd;
+      if (angle_delta == 0) {
+        if (this_rd == INT64_MAX) return 0;
+        rd_cost[1] = this_rd;
+        break;
+      }
+    }
+  }
+
+  assert(best_rd != INT64_MAX);
+  for (angle_delta = 1; angle_delta <= MAX_ANGLE_DELTA; angle_delta += 2) {
+    int64_t rd_thresh;
+    for (i = 0; i < 2; ++i) {
+      int skip_search = 0;
+      rd_thresh = best_rd + (best_rd >> 5);
+      if (rd_cost[2 * (angle_delta + 1) + i] > rd_thresh &&
+          rd_cost[2 * (angle_delta - 1) + i] > rd_thresh)
+        skip_search = 1;
+      if (!skip_search) {
+        mbmi->angle_delta[1] = (1 - 2 * i) * angle_delta;
+        pick_intra_angle_routine_sbuv(cpi, x, bsize, rate_overhead, best_rd,
+                                      rate, rd_stats, &best_angle_delta,
+                                      &best_rd);
+      }
+    }
+  }
+
+  mbmi->angle_delta[1] = best_angle_delta;
+  return rd_stats->rate != INT_MAX;
+}
+#endif  // CONFIG_EXT_INTRA
+
+static int64_t rd_pick_intra_sbuv_mode(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                       int *rate, int *rate_tokenonly,
+                                       int64_t *distortion, int *skippable,
+                                       BLOCK_SIZE bsize, TX_SIZE max_tx_size) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO best_mbmi = *mbmi;
+  PREDICTION_MODE mode;
+  int64_t best_rd = INT64_MAX, this_rd;
+  int this_rate;
+  RD_STATS tokenonly_rd_stats;
+#if CONFIG_PVQ
+  od_rollback_buffer buf;
+  od_encode_checkpoint(&x->daala_enc, &buf);
+#endif  // CONFIG_PVQ
+#if CONFIG_PALETTE
+  PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+  uint8_t *best_palette_color_map = NULL;
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_PALETTE
+  pmi->palette_size[1] = 0;
+#endif  // CONFIG_PALETTE
+  for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
+#if CONFIG_EXT_INTRA
+    const int is_directional_mode =
+        av1_is_directional_mode(mode, mbmi->sb_type);
+#endif  // CONFIG_EXT_INTRA
+    if (!(cpi->sf.intra_uv_mode_mask[txsize_sqr_up_map[max_tx_size]] &
+          (1 << mode)))
+      continue;
+
+    mbmi->uv_mode = mode;
+#if CONFIG_EXT_INTRA
+    mbmi->angle_delta[1] = 0;
+    if (is_directional_mode) {
+      const int rate_overhead = cpi->intra_uv_mode_cost[mbmi->mode][mode] +
+                                write_uniform_cost(2 * MAX_ANGLE_DELTA + 1, 0);
+      if (!rd_pick_intra_angle_sbuv(cpi, x, bsize, rate_overhead, best_rd,
+                                    &this_rate, &tokenonly_rd_stats))
+        continue;
+    } else {
+#endif  // CONFIG_EXT_INTRA
+      if (!super_block_uvrd(cpi, x, &tokenonly_rd_stats, bsize, best_rd)) {
+#if CONFIG_PVQ
+        od_encode_rollback(&x->daala_enc, &buf);
+#endif  // CONFIG_PVQ
+        continue;
+      }
+#if CONFIG_EXT_INTRA
+    }
+#endif  // CONFIG_EXT_INTRA
+    this_rate =
+        tokenonly_rd_stats.rate + cpi->intra_uv_mode_cost[mbmi->mode][mode];
+
+#if CONFIG_EXT_INTRA
+    if (is_directional_mode) {
+      this_rate += write_uniform_cost(2 * MAX_ANGLE_DELTA + 1,
+                                      MAX_ANGLE_DELTA + mbmi->angle_delta[1]);
+    }
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+    if (mbmi->sb_type >= BLOCK_8X8 && mode == DC_PRED)
+      this_rate += av1_cost_bit(cpi->common.fc->filter_intra_probs[1], 0);
+#endif  // CONFIG_FILTER_INTRA
+#if CONFIG_PALETTE
+    if (cpi->common.allow_screen_content_tools && mbmi->sb_type >= BLOCK_8X8 &&
+        mode == DC_PRED)
+      this_rate += av1_cost_bit(
+          av1_default_palette_uv_mode_prob[pmi->palette_size[0] > 0], 0);
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_PVQ
+    od_encode_rollback(&x->daala_enc, &buf);
+#endif  // CONFIG_PVQ
+    this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, tokenonly_rd_stats.dist);
+
+    if (this_rd < best_rd) {
+      best_mbmi = *mbmi;
+      best_rd = this_rd;
+      *rate = this_rate;
+      *rate_tokenonly = tokenonly_rd_stats.rate;
+      *distortion = tokenonly_rd_stats.dist;
+      *skippable = tokenonly_rd_stats.skip;
+    }
+  }
+
+#if CONFIG_PALETTE
+  if (cpi->common.allow_screen_content_tools && mbmi->sb_type >= BLOCK_8X8) {
+    best_palette_color_map = x->palette_buffer->best_palette_color_map;
+    rd_pick_palette_intra_sbuv(cpi, x,
+                               cpi->intra_uv_mode_cost[mbmi->mode][DC_PRED],
+                               best_palette_color_map, &best_mbmi, &best_rd,
+                               rate, rate_tokenonly, distortion, skippable);
+  }
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_FILTER_INTRA
+  if (mbmi->sb_type >= BLOCK_8X8) {
+    if (rd_pick_filter_intra_sbuv(cpi, x, rate, rate_tokenonly, distortion,
+                                  skippable, bsize, &best_rd))
+      best_mbmi = *mbmi;
+  }
+#endif  // CONFIG_FILTER_INTRA
+
+  *mbmi = best_mbmi;
+  // Make sure we actually chose a mode
+  assert(best_rd < INT64_MAX);
+  return best_rd;
+}
+
+static void choose_intra_uv_mode(const AV1_COMP *const cpi, MACROBLOCK *const x,
+                                 PICK_MODE_CONTEXT *ctx, BLOCK_SIZE bsize,
+                                 TX_SIZE max_tx_size, int *rate_uv,
+                                 int *rate_uv_tokenonly, int64_t *dist_uv,
+                                 int *skip_uv, PREDICTION_MODE *mode_uv) {
+  // Use an estimated rd for uv_intra based on DC_PRED if the
+  // appropriate speed flag is set.
+  (void)ctx;
+#if CONFIG_CB4X4
+#if CONFIG_CHROMA_2X2
+  rd_pick_intra_sbuv_mode(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
+                          bsize, max_tx_size);
+#else
+  max_tx_size = AOMMAX(max_tx_size, TX_4X4);
+  if (x->skip_chroma_rd) {
+    *rate_uv = 0;
+    *rate_uv_tokenonly = 0;
+    *dist_uv = 0;
+    *skip_uv = 1;
+    *mode_uv = DC_PRED;
+    return;
+  }
+  BLOCK_SIZE bs = scale_chroma_bsize(bsize, x->e_mbd.plane[1].subsampling_x,
+                                     x->e_mbd.plane[1].subsampling_y);
+  rd_pick_intra_sbuv_mode(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
+                          bs, max_tx_size);
+#endif  // CONFIG_CHROMA_2X2
+#else
+  rd_pick_intra_sbuv_mode(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
+                          bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize, max_tx_size);
+#endif  // CONFIG_CB4X4
+  *mode_uv = x->e_mbd.mi[0]->mbmi.uv_mode;
+}
+
+static int cost_mv_ref(const AV1_COMP *const cpi, PREDICTION_MODE mode,
+                       int16_t mode_context) {
+#if CONFIG_EXT_INTER
+  if (is_inter_compound_mode(mode)) {
+    return cpi
+        ->inter_compound_mode_cost[mode_context][INTER_COMPOUND_OFFSET(mode)];
+  }
+#endif
+
+#if CONFIG_REF_MV
+  int mode_cost = 0;
+  int16_t mode_ctx = mode_context & NEWMV_CTX_MASK;
+  int16_t is_all_zero_mv = mode_context & (1 << ALL_ZERO_FLAG_OFFSET);
+
+  assert(is_inter_mode(mode));
+
+  if (mode == NEWMV) {
+    mode_cost = cpi->newmv_mode_cost[mode_ctx][0];
+    return mode_cost;
+  } else {
+    mode_cost = cpi->newmv_mode_cost[mode_ctx][1];
+    mode_ctx = (mode_context >> ZEROMV_OFFSET) & ZEROMV_CTX_MASK;
+
+    if (is_all_zero_mv) return mode_cost;
+
+    if (mode == ZEROMV) {
+      mode_cost += cpi->zeromv_mode_cost[mode_ctx][0];
+      return mode_cost;
+    } else {
+      mode_cost += cpi->zeromv_mode_cost[mode_ctx][1];
+      mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK;
+
+      if (mode_context & (1 << SKIP_NEARESTMV_OFFSET)) mode_ctx = 6;
+      if (mode_context & (1 << SKIP_NEARMV_OFFSET)) mode_ctx = 7;
+      if (mode_context & (1 << SKIP_NEARESTMV_SUB8X8_OFFSET)) mode_ctx = 8;
+
+      mode_cost += cpi->refmv_mode_cost[mode_ctx][mode != NEARESTMV];
+      return mode_cost;
+    }
+  }
+#else
+  assert(is_inter_mode(mode));
+  return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
+#endif  // CONFIG_REF_MV
+}
+
+#if CONFIG_EXT_INTER
+static int get_interinter_compound_type_bits(BLOCK_SIZE bsize,
+                                             COMPOUND_TYPE comp_type) {
+  (void)bsize;
+  switch (comp_type) {
+    case COMPOUND_AVERAGE: return 0;
+#if CONFIG_WEDGE
+    case COMPOUND_WEDGE: return get_interinter_wedge_bits(bsize);
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+    case COMPOUND_SEG: return 1;
+#endif  // CONFIG_COMPOUND_SEGMENT
+    default: assert(0); return 0;
+  }
+}
+#endif  // CONFIG_EXT_INTER
+
+static int set_and_cost_bmi_mvs(
+    const AV1_COMP *const cpi, MACROBLOCK *x, MACROBLOCKD *xd, int i,
+    PREDICTION_MODE mode, int_mv this_mv[2],
+    int_mv frame_mv[MB_MODE_COUNT][TOTAL_REFS_PER_FRAME],
+    int_mv seg_mvs[TOTAL_REFS_PER_FRAME],
+#if CONFIG_EXT_INTER
+    int_mv compound_seg_newmvs[2],
+#endif  // CONFIG_EXT_INTER
+    int_mv *best_ref_mv[2], const int *mvjcost, int *mvcost[2], int mi_row,
+    int mi_col) {
+  MODE_INFO *const mic = xd->mi[0];
+  const MB_MODE_INFO *const mbmi = &mic->mbmi;
+  const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  int thismvcost = 0;
+  int idx, idy;
+  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
+  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
+  const int is_compound = has_second_ref(mbmi);
+  int mode_ctx;
+  (void)mi_row;
+  (void)mi_col;
+
+  switch (mode) {
+    case NEWMV: this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
+#if CONFIG_EXT_INTER
+      if (!cpi->common.allow_high_precision_mv)
+        lower_mv_precision(&this_mv[0].as_mv, 0);
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_REF_MV
+      for (idx = 0; idx < 1 + is_compound; ++idx) {
+        this_mv[idx] = seg_mvs[mbmi->ref_frame[idx]];
+        av1_set_mvcost(x, mbmi->ref_frame[idx], idx, mbmi->ref_mv_idx);
+        thismvcost +=
+            av1_mv_bit_cost(&this_mv[idx].as_mv, &best_ref_mv[idx]->as_mv,
+                            x->nmvjointcost, x->mvcost, MV_COST_WEIGHT_SUB);
+      }
+      (void)mvjcost;
+      (void)mvcost;
+#else
+      thismvcost += av1_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
+                                    mvjcost, mvcost, MV_COST_WEIGHT_SUB);
+#if !CONFIG_EXT_INTER
+      if (is_compound) {
+        this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
+        thismvcost += av1_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
+                                      mvjcost, mvcost, MV_COST_WEIGHT_SUB);
+      }
+#endif  // !CONFIG_EXT_INTER
+#endif  // CONFIG_REF_MV
+      break;
+    case NEARMV:
+    case NEARESTMV:
+      this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
+      if (is_compound)
+        this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
+      break;
+    case ZEROMV: {
+      int ref;
+      for (ref = 0; ref < 1 + is_compound; ++ref) {
+#if CONFIG_GLOBAL_MOTION
+        this_mv[ref].as_int =
+            gm_get_motion_vector(
+                &cpi->common.global_motion[mbmi->ref_frame[ref]],
+                cpi->common.allow_high_precision_mv, mbmi->sb_type, mi_col,
+                mi_row, i)
+                .as_int;
+#else
+        this_mv[ref].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+      }
+      break;
+    }
+#if CONFIG_EXT_INTER
+    case NEW_NEWMV:
+      if (compound_seg_newmvs[0].as_int == INVALID_MV ||
+          compound_seg_newmvs[1].as_int == INVALID_MV) {
+        this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
+        this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
+      } else {
+        this_mv[0].as_int = compound_seg_newmvs[0].as_int;
+        this_mv[1].as_int = compound_seg_newmvs[1].as_int;
+      }
+      if (!cpi->common.allow_high_precision_mv)
+        lower_mv_precision(&this_mv[0].as_mv, 0);
+      if (!cpi->common.allow_high_precision_mv)
+        lower_mv_precision(&this_mv[1].as_mv, 0);
+#if CONFIG_REF_MV
+      av1_set_mvcost(x, mbmi->ref_frame[0], 0, mbmi->ref_mv_idx);
+#endif
+      thismvcost += av1_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
+                                    mvjcost, mvcost, MV_COST_WEIGHT_SUB);
+#if CONFIG_REF_MV
+      av1_set_mvcost(x, mbmi->ref_frame[1], 1, mbmi->ref_mv_idx);
+#endif
+      thismvcost += av1_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
+                                    mvjcost, mvcost, MV_COST_WEIGHT_SUB);
+      break;
+    case NEW_NEARMV:
+    case NEW_NEARESTMV:
+      this_mv[0].as_int = seg_mvs[mbmi->ref_frame[0]].as_int;
+      if (!cpi->common.allow_high_precision_mv)
+        lower_mv_precision(&this_mv[0].as_mv, 0);
+#if CONFIG_REF_MV
+      av1_set_mvcost(x, mbmi->ref_frame[0], 0, mbmi->ref_mv_idx);
+#endif
+      thismvcost += av1_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
+                                    mvjcost, mvcost, MV_COST_WEIGHT_SUB);
+      this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
+      break;
+    case NEAR_NEWMV:
+    case NEAREST_NEWMV:
+      this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
+      this_mv[1].as_int = seg_mvs[mbmi->ref_frame[1]].as_int;
+      if (!cpi->common.allow_high_precision_mv)
+        lower_mv_precision(&this_mv[1].as_mv, 0);
+#if CONFIG_REF_MV
+      av1_set_mvcost(x, mbmi->ref_frame[1], 1, mbmi->ref_mv_idx);
+#endif
+      thismvcost += av1_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
+                                    mvjcost, mvcost, MV_COST_WEIGHT_SUB);
+      break;
+    case NEAREST_NEARMV:
+    case NEAR_NEARESTMV:
+    case NEAREST_NEARESTMV:
+    case NEAR_NEARMV:
+      this_mv[0].as_int = frame_mv[mode][mbmi->ref_frame[0]].as_int;
+      this_mv[1].as_int = frame_mv[mode][mbmi->ref_frame[1]].as_int;
+      break;
+    case ZERO_ZEROMV:
+#if CONFIG_GLOBAL_MOTION
+      this_mv[0].as_int =
+          gm_get_motion_vector(&cpi->common.global_motion[mbmi->ref_frame[0]],
+                               cpi->common.allow_high_precision_mv,
+                               mbmi->sb_type, mi_col, mi_row, i)
+              .as_int;
+      this_mv[1].as_int =
+          gm_get_motion_vector(&cpi->common.global_motion[mbmi->ref_frame[1]],
+                               cpi->common.allow_high_precision_mv,
+                               mbmi->sb_type, mi_col, mi_row, i)
+              .as_int;
+#else
+      this_mv[0].as_int = 0;
+      this_mv[1].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+      break;
+#endif  // CONFIG_EXT_INTER
+    default: break;
+  }
+
+  mic->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
+  if (is_compound) mic->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
+
+  mic->bmi[i].as_mode = mode;
+
+#if CONFIG_REF_MV
+  if (mode == NEWMV) {
+    mic->bmi[i].pred_mv[0].as_int =
+        mbmi_ext->ref_mvs[mbmi->ref_frame[0]][0].as_int;
+    if (is_compound)
+      mic->bmi[i].pred_mv[1].as_int =
+          mbmi_ext->ref_mvs[mbmi->ref_frame[1]][0].as_int;
+  } else {
+    mic->bmi[i].pred_mv[0].as_int = this_mv[0].as_int;
+    if (is_compound) mic->bmi[i].pred_mv[1].as_int = this_mv[1].as_int;
+  }
+#endif  // CONFIG_REF_MV
+
+  for (idy = 0; idy < num_4x4_blocks_high; ++idy)
+    for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
+      memmove(&mic->bmi[i + idy * 2 + idx], &mic->bmi[i], sizeof(mic->bmi[i]));
+
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+  if (is_compound)
+    mode_ctx = mbmi_ext->compound_mode_context[mbmi->ref_frame[0]];
+  else
+#endif  // CONFIG_EXT_INTER
+    mode_ctx = av1_mode_context_analyzer(mbmi_ext->mode_context,
+                                         mbmi->ref_frame, mbmi->sb_type, i);
+#else  // CONFIG_REF_MV
+  mode_ctx = mbmi_ext->mode_context[mbmi->ref_frame[0]];
+#endif  // CONFIG_REF_MV
+  return cost_mv_ref(cpi, mode, mode_ctx) + thismvcost;
+}
+
+static int64_t encode_inter_mb_segment_sub8x8(
+    const AV1_COMP *const cpi, MACROBLOCK *x, int64_t best_yrd, int i,
+    int *labelyrate, int64_t *distortion, int64_t *sse, ENTROPY_CONTEXT *ta,
+    ENTROPY_CONTEXT *tl, int ir, int ic, int mi_row, int mi_col) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *xd = &x->e_mbd;
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  struct macroblock_plane *const p = &x->plane[0];
+  MODE_INFO *const mi = xd->mi[0];
+  const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
+  const int txb_width = max_block_wide(xd, plane_bsize, 0);
+  const int txb_height = max_block_high(xd, plane_bsize, 0);
+  const int width = block_size_wide[plane_bsize];
+  const int height = block_size_high[plane_bsize];
+  int idx, idy;
+  const uint8_t *const src =
+      &p->src.buf[av1_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
+  uint8_t *const dst =
+      &pd->dst.buf[av1_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
+  int64_t thisdistortion = 0, thissse = 0;
+  int thisrate = 0;
+  TX_SIZE tx_size = mi->mbmi.tx_size;
+  TX_TYPE tx_type = get_tx_type(PLANE_TYPE_Y, xd, i, tx_size);
+  const int num_4x4_w = tx_size_wide_unit[tx_size];
+  const int num_4x4_h = tx_size_high_unit[tx_size];
+#if !CONFIG_PVQ
+  const SCAN_ORDER *scan_order = get_scan(cm, tx_size, tx_type, 1);
+#else
+  (void)cpi;
+  (void)ta;
+  (void)tl;
+  (void)tx_type;
+#endif  // !CONFIG_PVQ
+
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+  assert(IMPLIES(xd->lossless[mi->mbmi.segment_id], tx_size == TX_4X4));
+  assert(IMPLIES(!xd->lossless[mi->mbmi.segment_id],
+                 tx_size == max_txsize_rect_lookup[mi->mbmi.sb_type]));
+#else
+  assert(tx_size == TX_4X4);
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+
+  assert(tx_type == DCT_DCT);
+
+  av1_build_inter_predictor_sub8x8(xd, 0, i, ir, ic, mi_row, mi_col);
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    aom_highbd_subtract_block(
+        height, width, av1_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
+        8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
+  } else {
+    aom_subtract_block(height, width,
+                       av1_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
+                       8, src, p->src.stride, dst, pd->dst.stride);
+  }
+#else
+  aom_subtract_block(height, width,
+                     av1_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
+                     8, src, p->src.stride, dst, pd->dst.stride);
+#endif  // CONFIG_HIGHBITDEPTH
+
+  for (idy = 0; idy < txb_height; idy += num_4x4_h) {
+    for (idx = 0; idx < txb_width; idx += num_4x4_w) {
+      int64_t dist, ssz, rd, rd1, rd2;
+      int coeff_ctx;
+      const int k = i + (idy * 2 + idx);
+      const int block = av1_raster_order_to_block_index(tx_size, k);
+      assert(IMPLIES(tx_size == TX_4X8 || tx_size == TX_8X4,
+                     idx == 0 && idy == 0));
+      coeff_ctx = combine_entropy_contexts(*(ta + (k & 1)), *(tl + (k >> 1)));
+      av1_xform_quant(cm, x, 0, block, idy + (i >> 1), idx + (i & 0x01),
+                      BLOCK_8X8, tx_size, coeff_ctx, AV1_XFORM_QUANT_FP);
+      if (xd->lossless[xd->mi[0]->mbmi.segment_id] == 0)
+        av1_optimize_b(cm, x, 0, block, tx_size, coeff_ctx);
+      av1_dist_block(cpi, x, 0, BLOCK_8X8, block, idy + (i >> 1),
+                     idx + (i & 0x1), tx_size, &dist, &ssz,
+                     OUTPUT_HAS_PREDICTED_PIXELS);
+      thisdistortion += dist;
+      thissse += ssz;
+#if !CONFIG_PVQ
+      thisrate +=
+          av1_cost_coeffs(cpi, x, 0, block, tx_size, scan_order, (ta + (k & 1)),
+                          (tl + (k >> 1)), cpi->sf.use_fast_coef_costing);
+      *(ta + (k & 1)) = !(p->eobs[block] == 0);
+      *(tl + (k >> 1)) = !(p->eobs[block] == 0);
+#else
+      thisrate += x->rate;
+#endif  // !CONFIG_PVQ
+#if CONFIG_EXT_TX
+      if (tx_size == TX_8X4) {
+        *(ta + (k & 1) + 1) = *(ta + (k & 1));
+      }
+      if (tx_size == TX_4X8) {
+        *(tl + (k >> 1) + 1) = *(tl + (k >> 1));
+      }
+#endif  // CONFIG_EXT_TX
+      rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion);
+      rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse);
+      rd = AOMMIN(rd1, rd2);
+      if (rd >= best_yrd) return INT64_MAX;
+    }
+  }
+
+  *distortion = thisdistortion;
+  *labelyrate = thisrate;
+  *sse = thissse;
+
+  return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
+}
+
+typedef struct {
+  int eobs;
+  int brate;
+  int byrate;
+  int64_t bdist;
+  int64_t bsse;
+  int64_t brdcost;
+  int_mv mvs[2];
+#if CONFIG_REF_MV
+  int_mv pred_mv[2];
+#endif  // CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+  int_mv ref_mv[2];
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_CB4X4
+  ENTROPY_CONTEXT ta[4];
+  ENTROPY_CONTEXT tl[4];
+#else
+  ENTROPY_CONTEXT ta[2];
+  ENTROPY_CONTEXT tl[2];
+#endif  // CONFIG_CB4X4
+} SEG_RDSTAT;
+
+typedef struct {
+  int_mv *ref_mv[2];
+  int_mv mvp;
+
+  int64_t segment_rd;
+  int r;
+  int64_t d;
+  int64_t sse;
+  int segment_yrate;
+  PREDICTION_MODE modes[4];
+#if CONFIG_EXT_INTER
+  SEG_RDSTAT rdstat[4][INTER_MODES + INTER_COMPOUND_MODES];
+#else
+  SEG_RDSTAT rdstat[4][INTER_MODES];
+#endif  // CONFIG_EXT_INTER
+  int mvthresh;
+} BEST_SEG_INFO;
+
+static INLINE int mv_check_bounds(const MvLimits *mv_limits, const MV *mv) {
+  return (mv->row >> 3) < mv_limits->row_min ||
+         (mv->row >> 3) > mv_limits->row_max ||
+         (mv->col >> 3) < mv_limits->col_min ||
+         (mv->col >> 3) > mv_limits->col_max;
+}
+
+static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
+  MB_MODE_INFO *const mbmi = &x->e_mbd.mi[0]->mbmi;
+  struct macroblock_plane *const p = &x->plane[0];
+  struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
+
+  p->src.buf =
+      &p->src.buf[av1_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
+  assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
+  pd->pre[0].buf =
+      &pd->pre[0].buf[av1_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
+  if (has_second_ref(mbmi))
+    pd->pre[1].buf =
+        &pd->pre[1]
+             .buf[av1_raster_block_offset(BLOCK_8X8, i, pd->pre[1].stride)];
+}
+
+static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
+                                  struct buf_2d orig_pre[2]) {
+  MB_MODE_INFO *mbmi = &x->e_mbd.mi[0]->mbmi;
+  x->plane[0].src = orig_src;
+  x->e_mbd.plane[0].pre[0] = orig_pre[0];
+  if (has_second_ref(mbmi)) x->e_mbd.plane[0].pre[1] = orig_pre[1];
+}
+
+// Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
+// TODO(aconverse): Find out if this is still productive then clean up or remove
+static int check_best_zero_mv(
+    const AV1_COMP *const cpi, const int16_t mode_context[TOTAL_REFS_PER_FRAME],
+#if CONFIG_REF_MV && CONFIG_EXT_INTER
+    const int16_t compound_mode_context[TOTAL_REFS_PER_FRAME],
+#endif  // CONFIG_REF_MV && CONFIG_EXT_INTER
+    int_mv frame_mv[MB_MODE_COUNT][TOTAL_REFS_PER_FRAME], int this_mode,
+    const MV_REFERENCE_FRAME ref_frames[2], const BLOCK_SIZE bsize, int block,
+    int mi_row, int mi_col) {
+  int_mv zeromv[2];
+  int comp_pred_mode = ref_frames[1] > INTRA_FRAME;
+  int cur_frm;
+  (void)mi_row;
+  (void)mi_col;
+  for (cur_frm = 0; cur_frm < 1 + comp_pred_mode; cur_frm++) {
+#if CONFIG_GLOBAL_MOTION
+    if (this_mode == ZEROMV
+#if CONFIG_EXT_INTER
+        || this_mode == ZERO_ZEROMV
+#endif  // CONFIG_EXT_INTER
+        )
+      zeromv[cur_frm].as_int =
+          gm_get_motion_vector(&cpi->common.global_motion[ref_frames[cur_frm]],
+                               cpi->common.allow_high_precision_mv, bsize,
+                               mi_col, mi_row, block)
+              .as_int;
+    else
+#endif  // CONFIG_GLOBAL_MOTION
+      zeromv[cur_frm].as_int = 0;
+  }
+#if !CONFIG_EXT_INTER
+  assert(ref_frames[1] != INTRA_FRAME);  // Just sanity check
+#endif                                   // !CONFIG_EXT_INTER
+  if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
+      frame_mv[this_mode][ref_frames[0]].as_int == zeromv[0].as_int &&
+      (ref_frames[1] <= INTRA_FRAME ||
+       frame_mv[this_mode][ref_frames[1]].as_int == zeromv[1].as_int)) {
+#if CONFIG_REF_MV
+    int16_t rfc =
+        av1_mode_context_analyzer(mode_context, ref_frames, bsize, block);
+#else
+    int16_t rfc = mode_context[ref_frames[0]];
+#endif  // CONFIG_REF_MV
+    int c1 = cost_mv_ref(cpi, NEARMV, rfc);
+    int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
+    int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
+
+#if !CONFIG_REF_MV
+    (void)bsize;
+    (void)block;
+#endif  // !CONFIG_REF_MV
+
+    if (this_mode == NEARMV) {
+      if (c1 > c3) return 0;
+    } else if (this_mode == NEARESTMV) {
+      if (c2 > c3) return 0;
+    } else {
+      assert(this_mode == ZEROMV);
+      if (ref_frames[1] <= INTRA_FRAME) {
+        if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
+            (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
+          return 0;
+      } else {
+        if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
+             frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
+            (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
+             frame_mv[NEARMV][ref_frames[1]].as_int == 0))
+          return 0;
+      }
+    }
+  }
+#if CONFIG_EXT_INTER
+  else if ((this_mode == NEAREST_NEARESTMV || this_mode == NEAREST_NEARMV ||
+            this_mode == NEAR_NEARESTMV || this_mode == NEAR_NEARMV ||
+            this_mode == ZERO_ZEROMV) &&
+           frame_mv[this_mode][ref_frames[0]].as_int == zeromv[0].as_int &&
+           frame_mv[this_mode][ref_frames[1]].as_int == zeromv[1].as_int) {
+#if CONFIG_REF_MV
+    int16_t rfc = compound_mode_context[ref_frames[0]];
+#else
+    int16_t rfc = mode_context[ref_frames[0]];
+#endif  // CONFIG_REF_MV
+    int c1 = cost_mv_ref(cpi, NEAREST_NEARMV, rfc);
+    int c2 = cost_mv_ref(cpi, NEAREST_NEARESTMV, rfc);
+    int c3 = cost_mv_ref(cpi, ZERO_ZEROMV, rfc);
+    int c4 = cost_mv_ref(cpi, NEAR_NEARESTMV, rfc);
+    int c5 = cost_mv_ref(cpi, NEAR_NEARMV, rfc);
+
+    if (this_mode == NEAREST_NEARMV) {
+      if (c1 > c3) return 0;
+    } else if (this_mode == NEAREST_NEARESTMV) {
+      if (c2 > c3) return 0;
+    } else if (this_mode == NEAR_NEARESTMV) {
+      if (c4 > c3) return 0;
+    } else if (this_mode == NEAR_NEARMV) {
+      if (c5 > c3) return 0;
+    } else {
+      assert(this_mode == ZERO_ZEROMV);
+      if ((c3 >= c2 && frame_mv[NEAREST_NEARESTMV][ref_frames[0]].as_int == 0 &&
+           frame_mv[NEAREST_NEARESTMV][ref_frames[1]].as_int == 0) ||
+          (c3 >= c1 && frame_mv[NEAREST_NEARMV][ref_frames[0]].as_int == 0 &&
+           frame_mv[NEAREST_NEARMV][ref_frames[1]].as_int == 0) ||
+          (c3 >= c5 && frame_mv[NEAR_NEARMV][ref_frames[0]].as_int == 0 &&
+           frame_mv[NEAR_NEARMV][ref_frames[1]].as_int == 0) ||
+          (c3 >= c4 && frame_mv[NEAR_NEARESTMV][ref_frames[0]].as_int == 0 &&
+           frame_mv[NEAR_NEARESTMV][ref_frames[1]].as_int == 0))
+        return 0;
+    }
+  }
+#endif  // CONFIG_EXT_INTER
+  return 1;
+}
+
+static void joint_motion_search(const AV1_COMP *cpi, MACROBLOCK *x,
+                                BLOCK_SIZE bsize, int_mv *frame_mv, int mi_row,
+                                int mi_col,
+#if CONFIG_EXT_INTER
+                                int_mv *ref_mv_sub8x8[2],
+#endif  // CONFIG_EXT_INTER
+                                int *rate_mv, const int block) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const int pw = block_size_wide[bsize];
+  const int ph = block_size_high[bsize];
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  // This function should only ever be called for compound modes
+  assert(has_second_ref(mbmi));
+  const int refs[2] = { mbmi->ref_frame[0], mbmi->ref_frame[1] };
+  int_mv ref_mv[2];
+  int ite, ref;
+#if CONFIG_DUAL_FILTER
+  InterpFilter interp_filter[4] = {
+    mbmi->interp_filter[0], mbmi->interp_filter[1], mbmi->interp_filter[2],
+    mbmi->interp_filter[3],
+  };
+#else
+  const InterpFilter interp_filter = mbmi->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+  struct scale_factors sf;
+  struct macroblockd_plane *const pd = &xd->plane[0];
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+  // ic and ir are the 4x4 coordiantes of the sub8x8 at index "block"
+  const int ic = block & 1;
+  const int ir = (block - ic) >> 1;
+  const int p_col = ((mi_col * MI_SIZE) >> pd->subsampling_x) + 4 * ic;
+  const int p_row = ((mi_row * MI_SIZE) >> pd->subsampling_y) + 4 * ir;
+#if CONFIG_GLOBAL_MOTION
+  int is_global[2];
+  for (ref = 0; ref < 2; ++ref) {
+    WarpedMotionParams *const wm =
+        &xd->global_motion[xd->mi[0]->mbmi.ref_frame[ref]];
+    is_global[ref] = is_global_mv_block(xd->mi[0], block, wm->wmtype);
+  }
+#endif  // CONFIG_GLOBAL_MOTION
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+
+  // Do joint motion search in compound mode to get more accurate mv.
+  struct buf_2d backup_yv12[2][MAX_MB_PLANE];
+  int last_besterr[2] = { INT_MAX, INT_MAX };
+  const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
+    av1_get_scaled_ref_frame(cpi, refs[0]),
+    av1_get_scaled_ref_frame(cpi, refs[1])
+  };
+
+// Prediction buffer from second frame.
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[MAX_SB_SQUARE]);
+  uint8_t *second_pred;
+#else
+  DECLARE_ALIGNED(16, uint8_t, second_pred[MAX_SB_SQUARE]);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_EXT_INTER && CONFIG_CB4X4
+  (void)ref_mv_sub8x8;
+#endif  // CONFIG_EXT_INTER && CONFIG_CB4X4
+
+  for (ref = 0; ref < 2; ++ref) {
+#if CONFIG_EXT_INTER && !CONFIG_CB4X4
+    if (bsize < BLOCK_8X8 && ref_mv_sub8x8 != NULL)
+      ref_mv[ref].as_int = ref_mv_sub8x8[ref]->as_int;
+    else
+#endif  // CONFIG_EXT_INTER && !CONFIG_CB4X4
+      ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
+
+    if (scaled_ref_frame[ref]) {
+      int i;
+      // Swap out the reference frame for a version that's been scaled to
+      // match the resolution of the current frame, allowing the existing
+      // motion search code to be used without additional modifications.
+      for (i = 0; i < MAX_MB_PLANE; i++)
+        backup_yv12[ref][i] = xd->plane[i].pre[ref];
+      av1_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
+                           NULL);
+    }
+  }
+
+// Since we have scaled the reference frames to match the size of the current
+// frame we must use a unit scaling factor during mode selection.
+#if CONFIG_HIGHBITDEPTH
+  av1_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
+                                    cm->height, cm->use_highbitdepth);
+#else
+  av1_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
+                                    cm->height);
+#endif  // CONFIG_HIGHBITDEPTH
+
+  // Allow joint search multiple times iteratively for each reference frame
+  // and break out of the search loop if it couldn't find a better mv.
+  for (ite = 0; ite < 4; ite++) {
+    struct buf_2d ref_yv12[2];
+    int bestsme = INT_MAX;
+    int sadpb = x->sadperbit16;
+    MV *const best_mv = &x->best_mv.as_mv;
+    int search_range = 3;
+
+    MvLimits tmp_mv_limits = x->mv_limits;
+    int id = ite % 2;  // Even iterations search in the first reference frame,
+                       // odd iterations search in the second. The predictor
+                       // found for the 'other' reference frame is factored in.
+    const int plane = 0;
+    ConvolveParams conv_params = get_conv_params(0, plane);
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+    WarpTypesAllowed warp_types;
+#if CONFIG_GLOBAL_MOTION
+    warp_types.global_warp_allowed = is_global[!id];
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_WARPED_MOTION
+    warp_types.local_warp_allowed = mbmi->motion_mode == WARPED_CAUSAL;
+#endif  // CONFIG_WARPED_MOTION
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+
+    // Initialized here because of compiler problem in Visual Studio.
+    ref_yv12[0] = xd->plane[plane].pre[0];
+    ref_yv12[1] = xd->plane[plane].pre[1];
+
+#if CONFIG_DUAL_FILTER
+    // reload the filter types
+    interp_filter[0] =
+        (id == 0) ? mbmi->interp_filter[2] : mbmi->interp_filter[0];
+    interp_filter[1] =
+        (id == 0) ? mbmi->interp_filter[3] : mbmi->interp_filter[1];
+#endif  // CONFIG_DUAL_FILTER
+
+// Get the prediction block from the 'other' reference frame.
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
+      av1_highbd_build_inter_predictor(
+          ref_yv12[!id].buf, ref_yv12[!id].stride, second_pred, pw,
+          &frame_mv[refs[!id]].as_mv, &sf, pw, ph, 0, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+          &warp_types, p_col, p_row,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+          plane, MV_PRECISION_Q3, mi_col * MI_SIZE, mi_row * MI_SIZE, xd);
+    } else {
+      second_pred = (uint8_t *)second_pred_alloc_16;
+#endif  // CONFIG_HIGHBITDEPTH
+      av1_build_inter_predictor(
+          ref_yv12[!id].buf, ref_yv12[!id].stride, second_pred, pw,
+          &frame_mv[refs[!id]].as_mv, &sf, pw, ph, &conv_params, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+          &warp_types, p_col, p_row, plane, !id,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+          MV_PRECISION_Q3, mi_col * MI_SIZE, mi_row * MI_SIZE, xd);
+#if CONFIG_HIGHBITDEPTH
+    }
+#endif  // CONFIG_HIGHBITDEPTH
+
+    // Do compound motion search on the current reference frame.
+    if (id) xd->plane[plane].pre[0] = ref_yv12[id];
+    av1_set_mv_search_range(&x->mv_limits, &ref_mv[id].as_mv);
+
+    // Use the mv result from the single mode as mv predictor.
+    *best_mv = frame_mv[refs[id]].as_mv;
+
+    best_mv->col >>= 3;
+    best_mv->row >>= 3;
+
+#if CONFIG_REF_MV
+    av1_set_mvcost(x, refs[id], id, mbmi->ref_mv_idx);
+#endif  // CONFIG_REF_MV
+
+    // Small-range full-pixel motion search.
+    bestsme =
+        av1_refining_search_8p_c(x, sadpb, search_range, &cpi->fn_ptr[bsize],
+                                 &ref_mv[id].as_mv, second_pred);
+    if (bestsme < INT_MAX)
+      bestsme = av1_get_mvpred_av_var(x, best_mv, &ref_mv[id].as_mv,
+                                      second_pred, &cpi->fn_ptr[bsize], 1);
+
+    x->mv_limits = tmp_mv_limits;
+
+    if (bestsme < INT_MAX) {
+      int dis; /* TODO: use dis in distortion calculation later. */
+      unsigned int sse;
+      if (cpi->sf.use_upsampled_references) {
+        // Use up-sampled reference frames.
+        struct buf_2d backup_pred = pd->pre[0];
+        const YV12_BUFFER_CONFIG *upsampled_ref =
+            get_upsampled_ref(cpi, refs[id]);
+
+        // Set pred for Y plane
+        setup_pred_plane(&pd->pre[0], bsize, upsampled_ref->y_buffer,
+                         upsampled_ref->y_crop_width,
+                         upsampled_ref->y_crop_height, upsampled_ref->y_stride,
+                         (mi_row << 3), (mi_col << 3), NULL, pd->subsampling_x,
+                         pd->subsampling_y);
+
+// If bsize < BLOCK_8X8, adjust pred pointer for this block
+#if !CONFIG_CB4X4
+        if (bsize < BLOCK_8X8)
+          pd->pre[0].buf =
+              &pd->pre[0].buf[(av1_raster_block_offset(BLOCK_8X8, block,
+                                                       pd->pre[0].stride))
+                              << 3];
+#endif  // !CONFIG_CB4X4
+
+        bestsme = cpi->find_fractional_mv_step(
+            x, &ref_mv[id].as_mv, cpi->common.allow_high_precision_mv,
+            x->errorperbit, &cpi->fn_ptr[bsize], 0,
+            cpi->sf.mv.subpel_iters_per_step, NULL, x->nmvjointcost, x->mvcost,
+            &dis, &sse, second_pred, pw, ph, 1);
+
+        // Restore the reference frames.
+        pd->pre[0] = backup_pred;
+      } else {
+        (void)block;
+        bestsme = cpi->find_fractional_mv_step(
+            x, &ref_mv[id].as_mv, cpi->common.allow_high_precision_mv,
+            x->errorperbit, &cpi->fn_ptr[bsize], 0,
+            cpi->sf.mv.subpel_iters_per_step, NULL, x->nmvjointcost, x->mvcost,
+            &dis, &sse, second_pred, pw, ph, 0);
+      }
+    }
+
+    // Restore the pointer to the first (possibly scaled) prediction buffer.
+    if (id) xd->plane[plane].pre[0] = ref_yv12[0];
+
+    if (bestsme < last_besterr[id]) {
+      frame_mv[refs[id]].as_mv = *best_mv;
+      last_besterr[id] = bestsme;
+    } else {
+      break;
+    }
+  }
+
+  *rate_mv = 0;
+
+  for (ref = 0; ref < 2; ++ref) {
+    if (scaled_ref_frame[ref]) {
+      // Restore the prediction frame pointers to their unscaled versions.
+      int i;
+      for (i = 0; i < MAX_MB_PLANE; i++)
+        xd->plane[i].pre[ref] = backup_yv12[ref][i];
+    }
+#if CONFIG_REF_MV
+    av1_set_mvcost(x, refs[ref], ref, mbmi->ref_mv_idx);
+#endif  // CONFIG_REF_MV
+#if CONFIG_EXT_INTER && !CONFIG_CB4X4
+    if (bsize >= BLOCK_8X8)
+#endif  // CONFIG_EXT_INTER && !CONFIG_CB4X4
+      *rate_mv += av1_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
+                                  &x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
+                                  x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+#if CONFIG_EXT_INTER && !CONFIG_CB4X4
+    else
+      *rate_mv += av1_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
+                                  &ref_mv_sub8x8[ref]->as_mv, x->nmvjointcost,
+                                  x->mvcost, MV_COST_WEIGHT);
+#endif  // CONFIG_EXT_INTER && !CONFIG_CB4X4
+  }
+}
+
+#if CONFIG_REF_MV && !CONFIG_EXT_INTER
+static void update_mv_search_and_seg_mvs(
+    int *const run_mv_search, int_mv *const seg_mvs, int has_second_rf,
+    const MV_REFERENCE_FRAME *const ref_frame,
+    const SEG_RDSTAT *const ref_rdstat, int_mv *const bsi_ref_mv[2]) {
+  if (has_second_rf) {
+    if (seg_mvs[ref_frame[0]].as_int == ref_rdstat->mvs[0].as_int &&
+        ref_rdstat->mvs[0].as_int != INVALID_MV)
+      if (bsi_ref_mv[0]->as_int == ref_rdstat->pred_mv[0].as_int)
+        --*run_mv_search;
+
+    if (seg_mvs[ref_frame[1]].as_int == ref_rdstat->mvs[1].as_int &&
+        ref_rdstat->mvs[1].as_int != INVALID_MV)
+      if (bsi_ref_mv[1]->as_int == ref_rdstat->pred_mv[1].as_int)
+        --*run_mv_search;
+  } else {
+    if (bsi_ref_mv[0]->as_int == ref_rdstat->pred_mv[0].as_int &&
+        ref_rdstat->mvs[0].as_int != INVALID_MV) {
+      *run_mv_search = 0;
+      seg_mvs[ref_frame[0]].as_int = ref_rdstat->mvs[0].as_int;
+    }
+  }
+}
+#endif  // CONFIG_REF_MV && !CONFIG_EXT_INTER
+
+static int64_t rd_pick_inter_best_sub8x8_mode(
+    const AV1_COMP *const cpi, MACROBLOCK *x, int_mv *best_ref_mv,
+    int_mv *second_best_ref_mv, int64_t best_rd, int *returntotrate,
+    int *returnyrate, int64_t *returndistortion, int *skippable, int64_t *psse,
+    int mvthresh, int_mv seg_mvs[4][TOTAL_REFS_PER_FRAME],
+#if CONFIG_EXT_INTER
+    int_mv compound_seg_newmvs[4][2],
+#endif  // CONFIG_EXT_INTER
+    BEST_SEG_INFO *bsi_buf, int filter_idx, int mi_row, int mi_col) {
+  BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
+#if CONFIG_REF_MV
+  int_mv tmp_ref_mv[2];
+#endif  // CONFIG_REF_MV
+  MACROBLOCKD *xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  MB_MODE_INFO *mbmi = &mi->mbmi;
+  int mode_idx;
+  int k, br = 0, idx, idy;
+  int64_t bd = 0, block_sse = 0;
+  PREDICTION_MODE this_mode;
+  const AV1_COMMON *cm = &cpi->common;
+  struct macroblock_plane *const p = &x->plane[0];
+  struct macroblockd_plane *const pd = &xd->plane[0];
+  const int label_count = 4;
+  int64_t this_segment_rd = 0;
+  int label_mv_thresh;
+  int segmentyrate = 0;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
+  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
+#if CONFIG_CB4X4
+  ENTROPY_CONTEXT t_above[4], t_left[4];
+#else
+  ENTROPY_CONTEXT t_above[2], t_left[2];
+#endif  // CONFIG_CB4X4
+  int subpelmv = 1, have_ref = 0;
+  const int has_second_rf = has_second_ref(mbmi);
+  const int inter_mode_mask = cpi->sf.inter_mode_mask[bsize];
+  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+#if CONFIG_PVQ
+  od_rollback_buffer pre_buf;
+
+  od_encode_checkpoint(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+#if CONFIG_EXT_TX && CONFIG_RECT_TX
+  mbmi->tx_size =
+      xd->lossless[mbmi->segment_id] ? TX_4X4 : max_txsize_rect_lookup[bsize];
+#else
+  mbmi->tx_size = TX_4X4;
+#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX
+
+  av1_zero(*bsi);
+
+  bsi->segment_rd = best_rd;
+  bsi->ref_mv[0] = best_ref_mv;
+  bsi->ref_mv[1] = second_best_ref_mv;
+  bsi->mvp.as_int = best_ref_mv->as_int;
+  bsi->mvthresh = mvthresh;
+
+  for (idx = 0; idx < 4; ++idx) bsi->modes[idx] = ZEROMV;
+
+#if CONFIG_REF_MV
+  for (idx = 0; idx < 4; ++idx) {
+    for (k = NEARESTMV; k <= NEWMV; ++k) {
+      bsi->rdstat[idx][INTER_OFFSET(k)].pred_mv[0].as_int = INVALID_MV;
+      bsi->rdstat[idx][INTER_OFFSET(k)].pred_mv[1].as_int = INVALID_MV;
+
+      bsi->rdstat[idx][INTER_OFFSET(k)].mvs[0].as_int = INVALID_MV;
+      bsi->rdstat[idx][INTER_OFFSET(k)].mvs[1].as_int = INVALID_MV;
+    }
+  }
+#endif  // CONFIG_REF_MV
+
+  memcpy(t_above, pd->above_context, sizeof(t_above));
+  memcpy(t_left, pd->left_context, sizeof(t_left));
+
+  // 64 makes this threshold really big effectively
+  // making it so that we very rarely check mvs on
+  // segments.   setting this to 1 would make mv thresh
+  // roughly equal to what it is for macroblocks
+  label_mv_thresh = 1 * bsi->mvthresh / label_count;
+
+  // Segmentation method overheads
+  for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
+    for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
+      // TODO(jingning,rbultje): rewrite the rate-distortion optimization
+      // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
+      int_mv mode_mv[MB_MODE_COUNT][2];
+      int_mv frame_mv[MB_MODE_COUNT][TOTAL_REFS_PER_FRAME];
+      PREDICTION_MODE mode_selected = ZEROMV;
+      int64_t new_best_rd = INT64_MAX;
+      const int index = idy * 2 + idx;
+      int ref;
+#if CONFIG_REF_MV
+      CANDIDATE_MV ref_mv_stack[2][MAX_REF_MV_STACK_SIZE];
+      uint8_t ref_mv_count[2];
+#endif  // CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+      int_mv ref_mvs_sub8x8[2][2];
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_PVQ
+      od_rollback_buffer idx_buf, post_buf;
+      od_encode_checkpoint(&x->daala_enc, &idx_buf);
+      od_encode_checkpoint(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+
+      for (ref = 0; ref < 1 + has_second_rf; ++ref) {
+        const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
+#if CONFIG_EXT_INTER
+        int_mv mv_ref_list[MAX_MV_REF_CANDIDATES];
+        av1_update_mv_context(cm, xd, mi, frame, mv_ref_list, index, mi_row,
+                              mi_col, NULL);
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_GLOBAL_MOTION
+        frame_mv[ZEROMV][frame].as_int =
+            gm_get_motion_vector(&cm->global_motion[frame],
+                                 cm->allow_high_precision_mv, mbmi->sb_type,
+                                 mi_col, mi_row, index)
+                .as_int;
+#else   // CONFIG_GLOBAL_MOTION
+        frame_mv[ZEROMV][frame].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+        av1_append_sub8x8_mvs_for_idx(cm, xd, index, ref, mi_row, mi_col,
+#if CONFIG_REF_MV
+                                      ref_mv_stack[ref], &ref_mv_count[ref],
+#endif  // CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+                                      mv_ref_list,
+#endif  // CONFIG_EXT_INTER
+                                      &frame_mv[NEARESTMV][frame],
+                                      &frame_mv[NEARMV][frame]);
+
+#if CONFIG_REF_MV
+        tmp_ref_mv[ref] = frame_mv[NEARESTMV][mbmi->ref_frame[ref]];
+        lower_mv_precision(&tmp_ref_mv[ref].as_mv, cm->allow_high_precision_mv);
+        bsi->ref_mv[ref] = &tmp_ref_mv[ref];
+        mbmi_ext->ref_mvs[frame][0] = tmp_ref_mv[ref];
+#endif  // CONFIG_REF_MV
+
+#if CONFIG_EXT_INTER
+        mv_ref_list[0].as_int = frame_mv[NEARESTMV][frame].as_int;
+        mv_ref_list[1].as_int = frame_mv[NEARMV][frame].as_int;
+        av1_find_best_ref_mvs(cm->allow_high_precision_mv, mv_ref_list,
+                              &ref_mvs_sub8x8[0][ref], &ref_mvs_sub8x8[1][ref]);
+
+        if (has_second_rf) {
+#if CONFIG_GLOBAL_MOTION
+          frame_mv[ZERO_ZEROMV][frame].as_int =
+              gm_get_motion_vector(&cm->global_motion[frame],
+                                   cm->allow_high_precision_mv, mbmi->sb_type,
+                                   mi_col, mi_row, index)
+                  .as_int;
+#else
+          frame_mv[ZERO_ZEROMV][frame].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+          frame_mv[NEAREST_NEARESTMV][frame].as_int =
+              frame_mv[NEARESTMV][frame].as_int;
+
+          if (ref == 0) {
+            frame_mv[NEAREST_NEARMV][frame].as_int =
+                frame_mv[NEARESTMV][frame].as_int;
+            frame_mv[NEAR_NEARESTMV][frame].as_int =
+                frame_mv[NEARMV][frame].as_int;
+            frame_mv[NEAREST_NEWMV][frame].as_int =
+                frame_mv[NEARESTMV][frame].as_int;
+            frame_mv[NEAR_NEWMV][frame].as_int = frame_mv[NEARMV][frame].as_int;
+            frame_mv[NEAR_NEARMV][frame].as_int =
+                frame_mv[NEARMV][frame].as_int;
+          } else if (ref == 1) {
+            frame_mv[NEAREST_NEARMV][frame].as_int =
+                frame_mv[NEARMV][frame].as_int;
+            frame_mv[NEAR_NEARESTMV][frame].as_int =
+                frame_mv[NEARESTMV][frame].as_int;
+            frame_mv[NEW_NEARESTMV][frame].as_int =
+                frame_mv[NEARESTMV][frame].as_int;
+            frame_mv[NEW_NEARMV][frame].as_int = frame_mv[NEARMV][frame].as_int;
+            frame_mv[NEAR_NEARMV][frame].as_int =
+                frame_mv[NEARMV][frame].as_int;
+          }
+        }
+#endif  // CONFIG_EXT_INTER
+      }
+
+// search for the best motion vector on this segment
+#if CONFIG_EXT_INTER
+      for (this_mode = (has_second_rf ? NEAREST_NEARESTMV : NEARESTMV);
+           this_mode <= (has_second_rf ? NEW_NEWMV : NEWMV); ++this_mode)
+#else
+      for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode)
+#endif  // CONFIG_EXT_INTER
+      {
+        const struct buf_2d orig_src = x->plane[0].src;
+        struct buf_2d orig_pre[2];
+        // This flag controls if the motion estimation will kick off. When it
+        // is set to a non-zero value, the encoder will force motion estimation.
+        int run_mv_search = 0;
+
+        mode_idx = INTER_OFFSET(this_mode);
+#if CONFIG_EXT_INTER
+        for (ref = 0; ref < 1 + has_second_rf; ++ref)
+          bsi->ref_mv[ref]->as_int = ref_mvs_sub8x8[0][ref].as_int;
+#endif  // CONFIG_EXT_INTER
+        bsi->rdstat[index][mode_idx].brdcost = INT64_MAX;
+        if (!(inter_mode_mask & (1 << this_mode))) continue;
+
+#if CONFIG_REF_MV
+        run_mv_search = 2;
+#if !CONFIG_EXT_INTER
+        if (filter_idx > 0 && this_mode == NEWMV) {
+          const BEST_SEG_INFO *ref_bsi = bsi_buf;
+          const SEG_RDSTAT *ref_rdstat = &ref_bsi->rdstat[index][mode_idx];
+
+          update_mv_search_and_seg_mvs(&run_mv_search, seg_mvs[index],
+                                       has_second_rf, mbmi->ref_frame,
+                                       ref_rdstat, bsi->ref_mv);
+
+          if (run_mv_search != 0 && filter_idx > 1) {
+            ref_bsi = bsi_buf + 1;
+            ref_rdstat = &ref_bsi->rdstat[index][mode_idx];
+            run_mv_search = 2;
+            update_mv_search_and_seg_mvs(&run_mv_search, seg_mvs[index],
+                                         has_second_rf, mbmi->ref_frame,
+                                         ref_rdstat, bsi->ref_mv);
+          }
+        }
+#endif  // !CONFIG_EXT_INTER
+#endif  // CONFIG_REF_MV
+
+#if CONFIG_GLOBAL_MOTION
+        if (cm->global_motion[mbmi->ref_frame[0]].wmtype == IDENTITY &&
+            (!has_second_rf ||
+             cm->global_motion[mbmi->ref_frame[1]].wmtype == IDENTITY))
+#endif  // CONFIG_GLOBAL_MOTION
+
+          if (!check_best_zero_mv(cpi, mbmi_ext->mode_context,
+#if CONFIG_REF_MV && CONFIG_EXT_INTER
+                                  mbmi_ext->compound_mode_context,
+#endif  // CONFIG_REF_MV && CONFIG_EXT_INTER
+                                  frame_mv, this_mode, mbmi->ref_frame, bsize,
+                                  index, mi_row, mi_col))
+            continue;
+
+        memcpy(orig_pre, pd->pre, sizeof(orig_pre));
+        memcpy(bsi->rdstat[index][mode_idx].ta, t_above,
+               sizeof(bsi->rdstat[index][mode_idx].ta));
+        memcpy(bsi->rdstat[index][mode_idx].tl, t_left,
+               sizeof(bsi->rdstat[index][mode_idx].tl));
+#if CONFIG_PVQ
+        od_encode_rollback(&x->daala_enc, &idx_buf);
+#endif  // CONFIG_PVQ
+
+        // motion search for newmv (single predictor case only)
+        if (!has_second_rf &&
+#if CONFIG_EXT_INTER
+            have_newmv_in_inter_mode(this_mode) &&
+            (seg_mvs[index][mbmi->ref_frame[0]].as_int == INVALID_MV)
+#else
+            this_mode == NEWMV &&
+            (seg_mvs[index][mbmi->ref_frame[0]].as_int == INVALID_MV ||
+             run_mv_search)
+#endif  // CONFIG_EXT_INTER
+                ) {
+          int step_param = 0;
+          int bestsme = INT_MAX;
+          int sadpb = x->sadperbit4;
+          MV mvp_full;
+          int max_mv;
+          int cost_list[5];
+          MvLimits tmp_mv_limits = x->mv_limits;
+
+          /* Is the best so far sufficiently good that we cant justify doing
+           * and new motion search. */
+          if (new_best_rd < label_mv_thresh) break;
+
+#if CONFIG_EXT_INTER
+          bsi->mvp.as_int = bsi->ref_mv[0]->as_int;
+#else
+// use previous block's result as next block's MV predictor.
+#if !CONFIG_REF_MV
+          if (index > 0) {
+            bsi->mvp.as_int = mi->bmi[index - 1].as_mv[0].as_int;
+            if (index == 2)
+              bsi->mvp.as_int = mi->bmi[index - 2].as_mv[0].as_int;
+          }
+#endif  // !CONFIG_REF_MV
+#endif  // CONFIG_EXT_INTER
+          max_mv = (index == 0) ? (int)x->max_mv_context[mbmi->ref_frame[0]]
+                                : AOMMAX(abs(bsi->mvp.as_mv.row),
+                                         abs(bsi->mvp.as_mv.col)) >>
+                                      3;
+
+          if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
+            // Take wtd average of the step_params based on the last frame's
+            // max mv magnitude and the best ref mvs of the current block for
+            // the given reference.
+            step_param =
+                (av1_init_search_range(max_mv) + cpi->mv_step_param) / 2;
+          } else {
+            step_param = cpi->mv_step_param;
+          }
+
+#if CONFIG_REF_MV
+          mvp_full.row = bsi->ref_mv[0]->as_mv.row >> 3;
+          mvp_full.col = bsi->ref_mv[0]->as_mv.col >> 3;
+#else
+          mvp_full.row = bsi->mvp.as_mv.row >> 3;
+          mvp_full.col = bsi->mvp.as_mv.col >> 3;
+#endif  // CONFIG_REF_MV
+
+          if (cpi->sf.adaptive_motion_search) {
+            mvp_full.row = x->pred_mv[mbmi->ref_frame[0]].row >> 3;
+            mvp_full.col = x->pred_mv[mbmi->ref_frame[0]].col >> 3;
+            step_param = AOMMAX(step_param, 8);
+          }
+
+          // adjust src pointer for this block
+          mi_buf_shift(x, index);
+
+          av1_set_mv_search_range(&x->mv_limits, &bsi->ref_mv[0]->as_mv);
+
+          x->best_mv.as_int = x->second_best_mv.as_int = INVALID_MV;
+
+#if CONFIG_REF_MV
+          av1_set_mvcost(x, mbmi->ref_frame[0], 0, mbmi->ref_mv_idx);
+#endif  // CONFIG_REF_MV
+          bestsme = av1_full_pixel_search(
+              cpi, x, bsize, &mvp_full, step_param, sadpb,
+              cpi->sf.mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
+              &bsi->ref_mv[0]->as_mv, INT_MAX, 1);
+
+          x->mv_limits = tmp_mv_limits;
+
+          if (bestsme < INT_MAX) {
+            int distortion;
+            if (cpi->sf.use_upsampled_references) {
+              int best_mv_var;
+              const int try_second =
+                  x->second_best_mv.as_int != INVALID_MV &&
+                  x->second_best_mv.as_int != x->best_mv.as_int;
+              const int pw = block_size_wide[bsize];
+              const int ph = block_size_high[bsize];
+              // Use up-sampled reference frames.
+              struct buf_2d backup_pred = pd->pre[0];
+              const YV12_BUFFER_CONFIG *upsampled_ref =
+                  get_upsampled_ref(cpi, mbmi->ref_frame[0]);
+
+              // Set pred for Y plane
+              setup_pred_plane(
+                  &pd->pre[0], bsize, upsampled_ref->y_buffer,
+                  upsampled_ref->y_crop_width, upsampled_ref->y_crop_height,
+                  upsampled_ref->y_stride, (mi_row << 3), (mi_col << 3), NULL,
+                  pd->subsampling_x, pd->subsampling_y);
+
+              // adjust pred pointer for this block
+              pd->pre[0].buf =
+                  &pd->pre[0].buf[(av1_raster_block_offset(BLOCK_8X8, index,
+                                                           pd->pre[0].stride))
+                                  << 3];
+
+              best_mv_var = cpi->find_fractional_mv_step(
+                  x, &bsi->ref_mv[0]->as_mv, cm->allow_high_precision_mv,
+                  x->errorperbit, &cpi->fn_ptr[bsize],
+                  cpi->sf.mv.subpel_force_stop,
+                  cpi->sf.mv.subpel_iters_per_step,
+                  cond_cost_list(cpi, cost_list), x->nmvjointcost, x->mvcost,
+                  &distortion, &x->pred_sse[mbmi->ref_frame[0]], NULL, pw, ph,
+                  1);
+
+              if (try_second) {
+                int this_var;
+                MV best_mv = x->best_mv.as_mv;
+                const MV ref_mv = bsi->ref_mv[0]->as_mv;
+                const int minc =
+                    AOMMAX(x->mv_limits.col_min * 8, ref_mv.col - MV_MAX);
+                const int maxc =
+                    AOMMIN(x->mv_limits.col_max * 8, ref_mv.col + MV_MAX);
+                const int minr =
+                    AOMMAX(x->mv_limits.row_min * 8, ref_mv.row - MV_MAX);
+                const int maxr =
+                    AOMMIN(x->mv_limits.row_max * 8, ref_mv.row + MV_MAX);
+
+                x->best_mv = x->second_best_mv;
+                if (x->best_mv.as_mv.row * 8 <= maxr &&
+                    x->best_mv.as_mv.row * 8 >= minr &&
+                    x->best_mv.as_mv.col * 8 <= maxc &&
+                    x->best_mv.as_mv.col * 8 >= minc) {
+                  this_var = cpi->find_fractional_mv_step(
+                      x, &bsi->ref_mv[0]->as_mv, cm->allow_high_precision_mv,
+                      x->errorperbit, &cpi->fn_ptr[bsize],
+                      cpi->sf.mv.subpel_force_stop,
+                      cpi->sf.mv.subpel_iters_per_step,
+                      cond_cost_list(cpi, cost_list), x->nmvjointcost,
+                      x->mvcost, &distortion, &x->pred_sse[mbmi->ref_frame[0]],
+                      NULL, pw, ph, 1);
+                  if (this_var < best_mv_var) best_mv = x->best_mv.as_mv;
+                  x->best_mv.as_mv = best_mv;
+                }
+              }
+
+              // Restore the reference frames.
+              pd->pre[0] = backup_pred;
+            } else {
+              cpi->find_fractional_mv_step(
+                  x, &bsi->ref_mv[0]->as_mv, cm->allow_high_precision_mv,
+                  x->errorperbit, &cpi->fn_ptr[bsize],
+                  cpi->sf.mv.subpel_force_stop,
+                  cpi->sf.mv.subpel_iters_per_step,
+                  cond_cost_list(cpi, cost_list), x->nmvjointcost, x->mvcost,
+                  &distortion, &x->pred_sse[mbmi->ref_frame[0]], NULL, 0, 0, 0);
+            }
+
+// save motion search result for use in compound prediction
+#if CONFIG_EXT_INTER
+            seg_mvs[index][mbmi->ref_frame[0]].as_mv = x->best_mv.as_mv;
+#else
+            seg_mvs[index][mbmi->ref_frame[0]].as_mv = x->best_mv.as_mv;
+#endif  // CONFIG_EXT_INTER
+          }
+
+          if (cpi->sf.adaptive_motion_search)
+            x->pred_mv[mbmi->ref_frame[0]] = x->best_mv.as_mv;
+
+#if CONFIG_EXT_INTER
+          mode_mv[this_mode][0] = x->best_mv;
+#else
+          mode_mv[NEWMV][0] = x->best_mv;
+#endif  // CONFIG_EXT_INTER
+
+          // restore src pointers
+          mi_buf_restore(x, orig_src, orig_pre);
+        }
+
+        if (has_second_rf) {
+#if CONFIG_EXT_INTER
+          if (seg_mvs[index][mbmi->ref_frame[1]].as_int == INVALID_MV ||
+              seg_mvs[index][mbmi->ref_frame[0]].as_int == INVALID_MV)
+#else
+          if (seg_mvs[index][mbmi->ref_frame[1]].as_int == INVALID_MV ||
+              seg_mvs[index][mbmi->ref_frame[0]].as_int == INVALID_MV)
+#endif  // CONFIG_EXT_INTER
+            continue;
+        }
+
+#if CONFIG_DUAL_FILTER
+        (void)run_mv_search;
+#endif  // CONFIG_DUAL_FILTER
+
+        if (has_second_rf &&
+#if CONFIG_EXT_INTER
+            this_mode == NEW_NEWMV &&
+#else
+            this_mode == NEWMV &&
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_DUAL_FILTER
+            (mbmi->interp_filter[0] == EIGHTTAP_REGULAR || run_mv_search))
+#else
+            (mbmi->interp_filter == EIGHTTAP_REGULAR || run_mv_search))
+#endif  // CONFIG_DUAL_FILTER
+        {
+          // adjust src pointers
+          mi_buf_shift(x, index);
+          if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
+            int rate_mv;
+            frame_mv[this_mode][mbmi->ref_frame[0]].as_int =
+                seg_mvs[index][mbmi->ref_frame[0]].as_int;
+            frame_mv[this_mode][mbmi->ref_frame[1]].as_int =
+                seg_mvs[index][mbmi->ref_frame[1]].as_int;
+            joint_motion_search(cpi, x, bsize, frame_mv[this_mode], mi_row,
+                                mi_col,
+#if CONFIG_EXT_INTER
+                                bsi->ref_mv,
+#endif  // CONFIG_EXT_INTER
+                                &rate_mv, index);
+#if CONFIG_EXT_INTER
+            compound_seg_newmvs[index][0].as_int =
+                frame_mv[this_mode][mbmi->ref_frame[0]].as_int;
+            compound_seg_newmvs[index][1].as_int =
+                frame_mv[this_mode][mbmi->ref_frame[1]].as_int;
+#else
+            seg_mvs[index][mbmi->ref_frame[0]].as_int =
+                frame_mv[this_mode][mbmi->ref_frame[0]].as_int;
+            seg_mvs[index][mbmi->ref_frame[1]].as_int =
+                frame_mv[this_mode][mbmi->ref_frame[1]].as_int;
+#endif  // CONFIG_EXT_INTER
+          }
+          // restore src pointers
+          mi_buf_restore(x, orig_src, orig_pre);
+        }
+
+        bsi->rdstat[index][mode_idx].brate = set_and_cost_bmi_mvs(
+            cpi, x, xd, index, this_mode, mode_mv[this_mode], frame_mv,
+            seg_mvs[index],
+#if CONFIG_EXT_INTER
+            compound_seg_newmvs[index],
+#endif  // CONFIG_EXT_INTER
+            bsi->ref_mv, x->nmvjointcost, x->mvcost, mi_row, mi_col);
+
+        for (ref = 0; ref < 1 + has_second_rf; ++ref) {
+          bsi->rdstat[index][mode_idx].mvs[ref].as_int =
+              mode_mv[this_mode][ref].as_int;
+          if (num_4x4_blocks_wide > 1)
+            bsi->rdstat[index + 1][mode_idx].mvs[ref].as_int =
+                mode_mv[this_mode][ref].as_int;
+          if (num_4x4_blocks_high > 1)
+            bsi->rdstat[index + 2][mode_idx].mvs[ref].as_int =
+                mode_mv[this_mode][ref].as_int;
+#if CONFIG_REF_MV
+          bsi->rdstat[index][mode_idx].pred_mv[ref].as_int =
+              mi->bmi[index].pred_mv[ref].as_int;
+          if (num_4x4_blocks_wide > 1)
+            bsi->rdstat[index + 1][mode_idx].pred_mv[ref].as_int =
+                mi->bmi[index].pred_mv[ref].as_int;
+          if (num_4x4_blocks_high > 1)
+            bsi->rdstat[index + 2][mode_idx].pred_mv[ref].as_int =
+                mi->bmi[index].pred_mv[ref].as_int;
+#endif  // CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+          bsi->rdstat[index][mode_idx].ref_mv[ref].as_int =
+              bsi->ref_mv[ref]->as_int;
+          if (num_4x4_blocks_wide > 1)
+            bsi->rdstat[index + 1][mode_idx].ref_mv[ref].as_int =
+                bsi->ref_mv[ref]->as_int;
+          if (num_4x4_blocks_high > 1)
+            bsi->rdstat[index + 2][mode_idx].ref_mv[ref].as_int =
+                bsi->ref_mv[ref]->as_int;
+#endif  // CONFIG_EXT_INTER
+        }
+
+        // Trap vectors that reach beyond the UMV borders
+        if (mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][0].as_mv) ||
+            (has_second_rf &&
+             mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][1].as_mv)))
+          continue;
+
+        if (filter_idx > 0) {
+          BEST_SEG_INFO *ref_bsi = bsi_buf;
+          subpelmv = 0;
+          have_ref = 1;
+
+          for (ref = 0; ref < 1 + has_second_rf; ++ref) {
+            subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
+#if CONFIG_EXT_INTER
+            if (have_newmv_in_inter_mode(this_mode))
+              have_ref &=
+                  ((mode_mv[this_mode][ref].as_int ==
+                    ref_bsi->rdstat[index][mode_idx].mvs[ref].as_int) &&
+                   (bsi->ref_mv[ref]->as_int ==
+                    ref_bsi->rdstat[index][mode_idx].ref_mv[ref].as_int));
+            else
+#endif  // CONFIG_EXT_INTER
+              have_ref &= mode_mv[this_mode][ref].as_int ==
+                          ref_bsi->rdstat[index][mode_idx].mvs[ref].as_int;
+          }
+
+          have_ref &= ref_bsi->rdstat[index][mode_idx].brate > 0;
+
+          if (filter_idx > 1 && !subpelmv && !have_ref) {
+            ref_bsi = bsi_buf + 1;
+            have_ref = 1;
+            for (ref = 0; ref < 1 + has_second_rf; ++ref)
+#if CONFIG_EXT_INTER
+              if (have_newmv_in_inter_mode(this_mode))
+                have_ref &=
+                    ((mode_mv[this_mode][ref].as_int ==
+                      ref_bsi->rdstat[index][mode_idx].mvs[ref].as_int) &&
+                     (bsi->ref_mv[ref]->as_int ==
+                      ref_bsi->rdstat[index][mode_idx].ref_mv[ref].as_int));
+              else
+#endif  // CONFIG_EXT_INTER
+                have_ref &= mode_mv[this_mode][ref].as_int ==
+                            ref_bsi->rdstat[index][mode_idx].mvs[ref].as_int;
+
+            have_ref &= ref_bsi->rdstat[index][mode_idx].brate > 0;
+          }
+
+          if (!subpelmv && have_ref &&
+              ref_bsi->rdstat[index][mode_idx].brdcost < INT64_MAX) {
+#if CONFIG_REF_MV
+            bsi->rdstat[index][mode_idx].byrate =
+                ref_bsi->rdstat[index][mode_idx].byrate;
+            bsi->rdstat[index][mode_idx].bdist =
+                ref_bsi->rdstat[index][mode_idx].bdist;
+            bsi->rdstat[index][mode_idx].bsse =
+                ref_bsi->rdstat[index][mode_idx].bsse;
+            bsi->rdstat[index][mode_idx].brate +=
+                ref_bsi->rdstat[index][mode_idx].byrate;
+            bsi->rdstat[index][mode_idx].eobs =
+                ref_bsi->rdstat[index][mode_idx].eobs;
+
+            bsi->rdstat[index][mode_idx].brdcost =
+                RDCOST(x->rdmult, x->rddiv, bsi->rdstat[index][mode_idx].brate,
+                       bsi->rdstat[index][mode_idx].bdist);
+
+            memcpy(bsi->rdstat[index][mode_idx].ta,
+                   ref_bsi->rdstat[index][mode_idx].ta,
+                   sizeof(bsi->rdstat[index][mode_idx].ta));
+            memcpy(bsi->rdstat[index][mode_idx].tl,
+                   ref_bsi->rdstat[index][mode_idx].tl,
+                   sizeof(bsi->rdstat[index][mode_idx].tl));
+#else
+            memcpy(&bsi->rdstat[index][mode_idx],
+                   &ref_bsi->rdstat[index][mode_idx], sizeof(SEG_RDSTAT));
+#endif  // CONFIG_REF_MV
+            if (num_4x4_blocks_wide > 1)
+              bsi->rdstat[index + 1][mode_idx].eobs =
+                  ref_bsi->rdstat[index + 1][mode_idx].eobs;
+            if (num_4x4_blocks_high > 1)
+              bsi->rdstat[index + 2][mode_idx].eobs =
+                  ref_bsi->rdstat[index + 2][mode_idx].eobs;
+
+            if (bsi->rdstat[index][mode_idx].brdcost < new_best_rd) {
+#if CONFIG_REF_MV
+              // If the NEWMV mode is using the same motion vector as the
+              // NEARESTMV mode, skip the rest rate-distortion calculations
+              // and use the inferred motion vector modes.
+              if (this_mode == NEWMV) {
+                if (has_second_rf) {
+                  if (bsi->rdstat[index][mode_idx].mvs[0].as_int ==
+                          bsi->ref_mv[0]->as_int &&
+                      bsi->rdstat[index][mode_idx].mvs[1].as_int ==
+                          bsi->ref_mv[1]->as_int)
+                    continue;
+                } else {
+                  if (bsi->rdstat[index][mode_idx].mvs[0].as_int ==
+                      bsi->ref_mv[0]->as_int)
+                    continue;
+                }
+              }
+#endif  // CONFIG_REF_MV
+              mode_selected = this_mode;
+              new_best_rd = bsi->rdstat[index][mode_idx].brdcost;
+#if CONFIG_PVQ
+              od_encode_checkpoint(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+            }
+            continue;
+          }
+        }
+
+        bsi->rdstat[index][mode_idx].brdcost = encode_inter_mb_segment_sub8x8(
+            cpi, x, bsi->segment_rd - this_segment_rd, index,
+            &bsi->rdstat[index][mode_idx].byrate,
+            &bsi->rdstat[index][mode_idx].bdist,
+            &bsi->rdstat[index][mode_idx].bsse, bsi->rdstat[index][mode_idx].ta,
+            bsi->rdstat[index][mode_idx].tl, idy, idx, mi_row, mi_col);
+
+        if (bsi->rdstat[index][mode_idx].brdcost < INT64_MAX) {
+          bsi->rdstat[index][mode_idx].brdcost += RDCOST(
+              x->rdmult, x->rddiv, bsi->rdstat[index][mode_idx].brate, 0);
+          bsi->rdstat[index][mode_idx].brate +=
+              bsi->rdstat[index][mode_idx].byrate;
+          bsi->rdstat[index][mode_idx].eobs = p->eobs[index];
+          if (num_4x4_blocks_wide > 1)
+            bsi->rdstat[index + 1][mode_idx].eobs = p->eobs[index + 1];
+          if (num_4x4_blocks_high > 1)
+            bsi->rdstat[index + 2][mode_idx].eobs = p->eobs[index + 2];
+        }
+
+        if (bsi->rdstat[index][mode_idx].brdcost < new_best_rd) {
+#if CONFIG_REF_MV
+          // If the NEWMV mode is using the same motion vector as the
+          // NEARESTMV mode, skip the rest rate-distortion calculations
+          // and use the inferred motion vector modes.
+          if (this_mode == NEWMV) {
+            if (has_second_rf) {
+              if (bsi->rdstat[index][mode_idx].mvs[0].as_int ==
+                      bsi->ref_mv[0]->as_int &&
+                  bsi->rdstat[index][mode_idx].mvs[1].as_int ==
+                      bsi->ref_mv[1]->as_int)
+                continue;
+            } else {
+              if (bsi->rdstat[index][mode_idx].mvs[0].as_int ==
+                  bsi->ref_mv[0]->as_int)
+                continue;
+            }
+          }
+#endif  // CONFIG_REF_MV
+          mode_selected = this_mode;
+          new_best_rd = bsi->rdstat[index][mode_idx].brdcost;
+
+#if CONFIG_PVQ
+          od_encode_checkpoint(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+        }
+      } /*for each 4x4 mode*/
+
+      if (new_best_rd == INT64_MAX) {
+        int iy, midx;
+        for (iy = index + 1; iy < 4; ++iy)
+#if CONFIG_EXT_INTER
+          for (midx = 0; midx < INTER_MODES + INTER_COMPOUND_MODES; ++midx)
+#else
+          for (midx = 0; midx < INTER_MODES; ++midx)
+#endif  // CONFIG_EXT_INTER
+            bsi->rdstat[iy][midx].brdcost = INT64_MAX;
+        bsi->segment_rd = INT64_MAX;
+#if CONFIG_PVQ
+        od_encode_rollback(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+        return INT64_MAX;
+      }
+
+      mode_idx = INTER_OFFSET(mode_selected);
+      memcpy(t_above, bsi->rdstat[index][mode_idx].ta, sizeof(t_above));
+      memcpy(t_left, bsi->rdstat[index][mode_idx].tl, sizeof(t_left));
+#if CONFIG_PVQ
+      od_encode_rollback(&x->daala_enc, &post_buf);
+#endif  // CONFIG_PVQ
+
+#if CONFIG_EXT_INTER
+      bsi->ref_mv[0]->as_int = bsi->rdstat[index][mode_idx].ref_mv[0].as_int;
+      if (has_second_rf)
+        bsi->ref_mv[1]->as_int = bsi->rdstat[index][mode_idx].ref_mv[1].as_int;
+#endif  // CONFIG_EXT_INTER
+      set_and_cost_bmi_mvs(cpi, x, xd, index, mode_selected,
+                           mode_mv[mode_selected], frame_mv, seg_mvs[index],
+#if CONFIG_EXT_INTER
+                           compound_seg_newmvs[index],
+#endif  // CONFIG_EXT_INTER
+                           bsi->ref_mv, x->nmvjointcost, x->mvcost, mi_row,
+                           mi_col);
+
+      br += bsi->rdstat[index][mode_idx].brate;
+      bd += bsi->rdstat[index][mode_idx].bdist;
+      block_sse += bsi->rdstat[index][mode_idx].bsse;
+      segmentyrate += bsi->rdstat[index][mode_idx].byrate;
+      this_segment_rd += bsi->rdstat[index][mode_idx].brdcost;
+
+      if (this_segment_rd > bsi->segment_rd) {
+        int iy, midx;
+        for (iy = index + 1; iy < 4; ++iy)
+#if CONFIG_EXT_INTER
+          for (midx = 0; midx < INTER_MODES + INTER_COMPOUND_MODES; ++midx)
+#else
+          for (midx = 0; midx < INTER_MODES; ++midx)
+#endif  // CONFIG_EXT_INTER
+            bsi->rdstat[iy][midx].brdcost = INT64_MAX;
+        bsi->segment_rd = INT64_MAX;
+#if CONFIG_PVQ
+        od_encode_rollback(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+        return INT64_MAX;
+      }
+    }
+  } /* for each label */
+#if CONFIG_PVQ
+  od_encode_rollback(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+
+  bsi->r = br;
+  bsi->d = bd;
+  bsi->segment_yrate = segmentyrate;
+  bsi->segment_rd = this_segment_rd;
+  bsi->sse = block_sse;
+
+  // update the coding decisions
+  for (k = 0; k < 4; ++k) bsi->modes[k] = mi->bmi[k].as_mode;
+
+#if CONFIG_DAALA_DIST
+  // Compute prediction (i.e. skip) and decoded distortion by daala-distortion.
+  {
+    const int src_stride = p->src.stride;
+    const int dst_stride = pd->dst.stride;
+    uint8_t *src = p->src.buf;
+    uint8_t *dst = pd->dst.buf;
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
+    const int use_activity_masking = 0;
+    const int qm = OD_HVS_QM;
+    const int bsw = block_size_wide[plane_bsize];
+    const int bsh = block_size_high[plane_bsize];
+    int64_t rd1, rd2;
+    int64_t daala_sse, daala_dist;
+    TX_SIZE tx_size = mbmi->tx_size;
+
+#if CONFIG_HIGHBITDEPTH
+    uint8_t *recon_8x8;
+    DECLARE_ALIGNED(16, uint16_t, recon16[8 * 8]);
+
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+      recon_8x8 = CONVERT_TO_BYTEPTR(recon16);
+    else
+      recon_8x8 = (uint8_t *)recon16;
+#else
+    DECLARE_ALIGNED(16, uint8_t, recon_8x8[8 * 8]);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_PVQ
+    use_activity_masking = x->daala_enc.use_activity_masking;
+#endif  // CONFIG_PVQ
+
+    // For each of sub8x8 prediction block in a 8x8 block
+    for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
+      for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
+        int i = idy * 2 + idx;
+        const uint8_t *const src_sub8x8 =
+            src + av1_raster_block_offset(BLOCK_8X8, i, p->src.stride);
+        uint8_t *const dst_sub8x8 =
+            dst + av1_raster_block_offset(BLOCK_8X8, i, pd->dst.stride);
+        uint8_t *recon_sub8x8 = recon_8x8 + (idy * 8 + idx) * 4;
+        const int txb_width = max_block_wide(xd, plane_bsize, 0);
+        const int txb_height = max_block_high(xd, plane_bsize, 0);
+        int idx_, idy_;
+
+        av1_build_inter_predictor_sub8x8(xd, 0, i, idy, idx, mi_row, mi_col);
+#if CONFIG_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          aom_highbd_subtract_block(
+              height, width,
+              av1_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff), 8,
+              src_sub8x8, p->src.stride, dst_sub8x8, pd->dst.stride, xd->bd);
+        } else {
+          aom_subtract_block(
+              height, width,
+              av1_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff), 8,
+              src_sub8x8, p->src.stride, dst_sub8x8, pd->dst.stride);
+        }
+#else
+        aom_subtract_block(
+            bsh, bsw, av1_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
+            8, src_sub8x8, p->src.stride, dst_sub8x8, pd->dst.stride);
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          aom_highbd_convolve_copy(dst_sub8x8, dst_stride, recon_sub8x8, 8,
+                                   NULL, 0, NULL, 0, bsw, bsh, xd->bd);
+        } else {
+#endif  // CONFIG_HIGHBITDEPTH
+          aom_convolve_copy(dst_sub8x8, dst_stride, recon_sub8x8, 8, NULL, 0,
+                            NULL, 0, bsw, bsh);
+#if CONFIG_HIGHBITDEPTH
+        }
+#endif  // CONFIG_HIGHBITDEPTH
+
+        // To get decoded pixels, do 4x4 xform and quant for each 4x4 block
+        // in a sub8x8 prediction block. In case remaining parts of
+        // sub8x8 inter mode rdo assume pd->dst stores predicted pixels,
+        // use local buffer to store decoded pixels.
+        for (idy_ = 0; idy_ < txb_height; idy_++) {
+          for (idx_ = 0; idx_ < txb_width; idx_++) {
+            int coeff_ctx = 0;
+            const tran_low_t *dqcoeff;
+            uint16_t eob;
+            const PLANE_TYPE plane_type = PLANE_TYPE_Y;
+            uint8_t *recon_4x4 = recon_sub8x8 + (idy_ * 8 + idx_) * 4;
+            const int block_raster_idx = (idy + idy_) * 2 + (idx + idx_);
+            const int block =
+                av1_raster_order_to_block_index(tx_size, block_raster_idx);
+            TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
+
+            dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
+            av1_xform_quant(cm, x, 0, block, idy + idy_, idx + idx_, BLOCK_8X8,
+                            tx_size, coeff_ctx, AV1_XFORM_QUANT_FP);
+            if (xd->lossless[xd->mi[0]->mbmi.segment_id] == 0)
+              av1_optimize_b(cm, x, 0, block, tx_size, coeff_ctx);
+
+            eob = p->eobs[block];
+            av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size,
+                                        recon_4x4, 8, eob);
+          }
+        }
+      }
+    }
+    // Compute daala-distortion for a 8x8 block
+    daala_sse = av1_daala_dist(src, src_stride, pd->dst.buf, dst_stride, 8, 8,
+                               qm, use_activity_masking, x->qindex)
+                << 4;
+
+    daala_dist = av1_daala_dist(src, src_stride, recon_8x8, 8, 8, 8, qm,
+                                use_activity_masking, x->qindex)
+                 << 4;
+
+    bsi->sse = daala_sse;
+    bsi->d = daala_dist;
+
+    rd1 = RDCOST(x->rdmult, x->rddiv, bsi->r, bsi->d);
+    rd2 = RDCOST(x->rdmult, x->rddiv, 0, bsi->sse);
+    bsi->segment_rd = AOMMIN(rd1, rd2);
+  }
+#endif  // CONFIG_DAALA_DIST
+
+  if (bsi->segment_rd > best_rd) return INT64_MAX;
+  /* set it to the best */
+  for (idx = 0; idx < 4; idx++) {
+    mode_idx = INTER_OFFSET(bsi->modes[idx]);
+    mi->bmi[idx].as_mv[0].as_int = bsi->rdstat[idx][mode_idx].mvs[0].as_int;
+    if (has_second_ref(mbmi))
+      mi->bmi[idx].as_mv[1].as_int = bsi->rdstat[idx][mode_idx].mvs[1].as_int;
+#if CONFIG_REF_MV
+    mi->bmi[idx].pred_mv[0] = bsi->rdstat[idx][mode_idx].pred_mv[0];
+    if (has_second_ref(mbmi))
+      mi->bmi[idx].pred_mv[1] = bsi->rdstat[idx][mode_idx].pred_mv[1];
+#endif  // CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+    mi->bmi[idx].ref_mv[0].as_int = bsi->rdstat[idx][mode_idx].ref_mv[0].as_int;
+    if (has_second_rf)
+      mi->bmi[idx].ref_mv[1].as_int =
+          bsi->rdstat[idx][mode_idx].ref_mv[1].as_int;
+#endif  // CONFIG_EXT_INTER
+    x->plane[0].eobs[idx] = bsi->rdstat[idx][mode_idx].eobs;
+    mi->bmi[idx].as_mode = bsi->modes[idx];
+  }
+
+  /*
+   * used to set mbmi->mv.as_int
+   */
+  *returntotrate = bsi->r;
+  *returndistortion = bsi->d;
+  *returnyrate = bsi->segment_yrate;
+  *skippable = av1_is_skippable_in_plane(x, BLOCK_8X8, 0);
+  *psse = bsi->sse;
+  mbmi->mode = bsi->modes[3];
+
+  return bsi->segment_rd;
+}
+
+static void estimate_ref_frame_costs(const AV1_COMMON *cm,
+                                     const MACROBLOCKD *xd, int segment_id,
+                                     unsigned int *ref_costs_single,
+                                     unsigned int *ref_costs_comp,
+                                     aom_prob *comp_mode_p) {
+  int seg_ref_active =
+      segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME);
+  if (seg_ref_active) {
+    memset(ref_costs_single, 0,
+           TOTAL_REFS_PER_FRAME * sizeof(*ref_costs_single));
+    memset(ref_costs_comp, 0, TOTAL_REFS_PER_FRAME * sizeof(*ref_costs_comp));
+    *comp_mode_p = 128;
+  } else {
+    aom_prob intra_inter_p = av1_get_intra_inter_prob(cm, xd);
+    aom_prob comp_inter_p = 128;
+
+    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+      comp_inter_p = av1_get_reference_mode_prob(cm, xd);
+      *comp_mode_p = comp_inter_p;
+    } else {
+      *comp_mode_p = 128;
+    }
+
+    ref_costs_single[INTRA_FRAME] = av1_cost_bit(intra_inter_p, 0);
+
+    if (cm->reference_mode != COMPOUND_REFERENCE) {
+      aom_prob ref_single_p1 = av1_get_pred_prob_single_ref_p1(cm, xd);
+      aom_prob ref_single_p2 = av1_get_pred_prob_single_ref_p2(cm, xd);
+#if CONFIG_EXT_REFS
+      aom_prob ref_single_p3 = av1_get_pred_prob_single_ref_p3(cm, xd);
+      aom_prob ref_single_p4 = av1_get_pred_prob_single_ref_p4(cm, xd);
+      aom_prob ref_single_p5 = av1_get_pred_prob_single_ref_p5(cm, xd);
+#endif  // CONFIG_EXT_REFS
+
+      unsigned int base_cost = av1_cost_bit(intra_inter_p, 1);
+
+      ref_costs_single[LAST_FRAME] =
+#if CONFIG_EXT_REFS
+          ref_costs_single[LAST2_FRAME] = ref_costs_single[LAST3_FRAME] =
+              ref_costs_single[BWDREF_FRAME] =
+#endif  // CONFIG_EXT_REFS
+                  ref_costs_single[GOLDEN_FRAME] =
+                      ref_costs_single[ALTREF_FRAME] = base_cost;
+
+#if CONFIG_EXT_REFS
+      ref_costs_single[LAST_FRAME] += av1_cost_bit(ref_single_p1, 0);
+      ref_costs_single[LAST2_FRAME] += av1_cost_bit(ref_single_p1, 0);
+      ref_costs_single[LAST3_FRAME] += av1_cost_bit(ref_single_p1, 0);
+      ref_costs_single[GOLDEN_FRAME] += av1_cost_bit(ref_single_p1, 0);
+      ref_costs_single[BWDREF_FRAME] += av1_cost_bit(ref_single_p1, 1);
+      ref_costs_single[ALTREF_FRAME] += av1_cost_bit(ref_single_p1, 1);
+
+      ref_costs_single[LAST_FRAME] += av1_cost_bit(ref_single_p3, 0);
+      ref_costs_single[LAST2_FRAME] += av1_cost_bit(ref_single_p3, 0);
+      ref_costs_single[LAST3_FRAME] += av1_cost_bit(ref_single_p3, 1);
+      ref_costs_single[GOLDEN_FRAME] += av1_cost_bit(ref_single_p3, 1);
+
+      ref_costs_single[BWDREF_FRAME] += av1_cost_bit(ref_single_p2, 0);
+      ref_costs_single[ALTREF_FRAME] += av1_cost_bit(ref_single_p2, 1);
+
+      ref_costs_single[LAST_FRAME] += av1_cost_bit(ref_single_p4, 0);
+      ref_costs_single[LAST2_FRAME] += av1_cost_bit(ref_single_p4, 1);
+
+      ref_costs_single[LAST3_FRAME] += av1_cost_bit(ref_single_p5, 0);
+      ref_costs_single[GOLDEN_FRAME] += av1_cost_bit(ref_single_p5, 1);
+#else
+      ref_costs_single[LAST_FRAME] += av1_cost_bit(ref_single_p1, 0);
+      ref_costs_single[GOLDEN_FRAME] += av1_cost_bit(ref_single_p1, 1);
+      ref_costs_single[ALTREF_FRAME] += av1_cost_bit(ref_single_p1, 1);
+
+      ref_costs_single[GOLDEN_FRAME] += av1_cost_bit(ref_single_p2, 0);
+      ref_costs_single[ALTREF_FRAME] += av1_cost_bit(ref_single_p2, 1);
+#endif  // CONFIG_EXT_REFS
+    } else {
+      ref_costs_single[LAST_FRAME] = 512;
+#if CONFIG_EXT_REFS
+      ref_costs_single[LAST2_FRAME] = 512;
+      ref_costs_single[LAST3_FRAME] = 512;
+      ref_costs_single[BWDREF_FRAME] = 512;
+#endif  // CONFIG_EXT_REFS
+      ref_costs_single[GOLDEN_FRAME] = 512;
+      ref_costs_single[ALTREF_FRAME] = 512;
+    }
+
+    if (cm->reference_mode != SINGLE_REFERENCE) {
+      aom_prob ref_comp_p = av1_get_pred_prob_comp_ref_p(cm, xd);
+#if CONFIG_EXT_REFS
+      aom_prob ref_comp_p1 = av1_get_pred_prob_comp_ref_p1(cm, xd);
+      aom_prob ref_comp_p2 = av1_get_pred_prob_comp_ref_p2(cm, xd);
+      aom_prob bwdref_comp_p = av1_get_pred_prob_comp_bwdref_p(cm, xd);
+#endif  // CONFIG_EXT_REFS
+
+      unsigned int base_cost = av1_cost_bit(intra_inter_p, 1);
+
+      ref_costs_comp[LAST_FRAME] =
+#if CONFIG_EXT_REFS
+          ref_costs_comp[LAST2_FRAME] = ref_costs_comp[LAST3_FRAME] =
+#endif  // CONFIG_EXT_REFS
+              ref_costs_comp[GOLDEN_FRAME] = base_cost;
+
+#if CONFIG_EXT_REFS
+      ref_costs_comp[BWDREF_FRAME] = ref_costs_comp[ALTREF_FRAME] = 0;
+#endif  // CONFIG_EXT_REFS
+
+#if CONFIG_EXT_REFS
+      ref_costs_comp[LAST_FRAME] += av1_cost_bit(ref_comp_p, 0);
+      ref_costs_comp[LAST2_FRAME] += av1_cost_bit(ref_comp_p, 0);
+      ref_costs_comp[LAST3_FRAME] += av1_cost_bit(ref_comp_p, 1);
+      ref_costs_comp[GOLDEN_FRAME] += av1_cost_bit(ref_comp_p, 1);
+
+      ref_costs_comp[LAST_FRAME] += av1_cost_bit(ref_comp_p1, 1);
+      ref_costs_comp[LAST2_FRAME] += av1_cost_bit(ref_comp_p1, 0);
+
+      ref_costs_comp[LAST3_FRAME] += av1_cost_bit(ref_comp_p2, 0);
+      ref_costs_comp[GOLDEN_FRAME] += av1_cost_bit(ref_comp_p2, 1);
+
+      // NOTE(zoeliu): BWDREF and ALTREF each add an extra cost by coding 1
+      //               more bit.
+      ref_costs_comp[BWDREF_FRAME] += av1_cost_bit(bwdref_comp_p, 0);
+      ref_costs_comp[ALTREF_FRAME] += av1_cost_bit(bwdref_comp_p, 1);
+#else
+      ref_costs_comp[LAST_FRAME] += av1_cost_bit(ref_comp_p, 0);
+      ref_costs_comp[GOLDEN_FRAME] += av1_cost_bit(ref_comp_p, 1);
+#endif  // CONFIG_EXT_REFS
+    } else {
+      ref_costs_comp[LAST_FRAME] = 512;
+#if CONFIG_EXT_REFS
+      ref_costs_comp[LAST2_FRAME] = 512;
+      ref_costs_comp[LAST3_FRAME] = 512;
+      ref_costs_comp[BWDREF_FRAME] = 512;
+      ref_costs_comp[ALTREF_FRAME] = 512;
+#endif  // CONFIG_EXT_REFS
+      ref_costs_comp[GOLDEN_FRAME] = 512;
+    }
+  }
+}
+
+static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
+                                 int mode_index,
+                                 int64_t comp_pred_diff[REFERENCE_MODES],
+                                 int skippable) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+
+  // Take a snapshot of the coding context so it can be
+  // restored if we decide to encode this way
+  ctx->skip = x->skip;
+  ctx->skippable = skippable;
+  ctx->best_mode_index = mode_index;
+  ctx->mic = *xd->mi[0];
+  ctx->mbmi_ext = *x->mbmi_ext;
+  ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
+  ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
+  ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
+}
+
+static void setup_buffer_inter(
+    const AV1_COMP *const cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
+    BLOCK_SIZE block_size, int mi_row, int mi_col,
+    int_mv frame_nearest_mv[TOTAL_REFS_PER_FRAME],
+    int_mv frame_near_mv[TOTAL_REFS_PER_FRAME],
+    struct buf_2d yv12_mb[TOTAL_REFS_PER_FRAME][MAX_MB_PLANE]) {
+  const AV1_COMMON *cm = &cpi->common;
+  const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MODE_INFO *const mi = xd->mi[0];
+  int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
+  const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
+  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+
+  assert(yv12 != NULL);
+
+  // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
+  // use the UV scaling factors.
+  av1_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
+
+  // Gets an initial list of candidate vectors from neighbours and orders them
+  av1_find_mv_refs(
+      cm, xd, mi, ref_frame,
+#if CONFIG_REF_MV
+      &mbmi_ext->ref_mv_count[ref_frame], mbmi_ext->ref_mv_stack[ref_frame],
+#if CONFIG_EXT_INTER
+      mbmi_ext->compound_mode_context,
+#endif  // CONFIG_EXT_INTER
+#endif  // CONFIG_REF_MV
+      candidates, mi_row, mi_col, NULL, NULL, mbmi_ext->mode_context);
+
+  // Candidate refinement carried out at encoder and decoder
+  av1_find_best_ref_mvs(cm->allow_high_precision_mv, candidates,
+                        &frame_nearest_mv[ref_frame],
+                        &frame_near_mv[ref_frame]);
+
+// Further refinement that is encode side only to test the top few candidates
+// in full and choose the best as the centre point for subsequent searches.
+// The current implementation doesn't support scaling.
+#if CONFIG_CB4X4
+  av1_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
+              block_size);
+#else
+  if (!av1_is_scaled(sf) && block_size >= BLOCK_8X8)
+    av1_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
+                block_size);
+#endif  // CONFIG_CB4X4
+}
+
+static void single_motion_search(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                 BLOCK_SIZE bsize, int mi_row, int mi_col,
+#if CONFIG_EXT_INTER
+                                 int ref_idx,
+#endif  // CONFIG_EXT_INTER
+                                 int *rate_mv) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  const AV1_COMMON *cm = &cpi->common;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0, 0, 0, 0 } };
+  int bestsme = INT_MAX;
+  int step_param;
+  int sadpb = x->sadperbit16;
+  MV mvp_full;
+#if CONFIG_EXT_INTER
+  int ref = mbmi->ref_frame[ref_idx];
+#else
+  int ref = mbmi->ref_frame[0];
+  int ref_idx = 0;
+#endif  // CONFIG_EXT_INTER
+  MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
+
+  MvLimits tmp_mv_limits = x->mv_limits;
+  int cost_list[5];
+
+  const YV12_BUFFER_CONFIG *scaled_ref_frame =
+      av1_get_scaled_ref_frame(cpi, ref);
+
+  MV pred_mv[3];
+  pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
+  pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
+  pred_mv[2] = x->pred_mv[ref];
+
+  if (scaled_ref_frame) {
+    int i;
+    // Swap out the reference frame for a version that's been scaled to
+    // match the resolution of the current frame, allowing the existing
+    // motion search code to be used without additional modifications.
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      backup_yv12[i] = xd->plane[i].pre[ref_idx];
+
+    av1_setup_pre_planes(xd, ref_idx, scaled_ref_frame, mi_row, mi_col, NULL);
+  }
+
+  av1_set_mv_search_range(&x->mv_limits, &ref_mv);
+
+#if CONFIG_REF_MV
+  av1_set_mvcost(x, ref, ref_idx, mbmi->ref_mv_idx);
+#endif  // CONFIG_REF_MV
+
+  // Work out the size of the first step in the mv step search.
+  // 0 here is maximum length first step. 1 is AOMMAX >> 1 etc.
+  if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
+    // Take wtd average of the step_params based on the last frame's
+    // max mv magnitude and that based on the best ref mvs of the current
+    // block for the given reference.
+    step_param =
+        (av1_init_search_range(x->max_mv_context[ref]) + cpi->mv_step_param) /
+        2;
+  } else {
+    step_param = cpi->mv_step_param;
+  }
+
+  if (cpi->sf.adaptive_motion_search && bsize < cm->sb_size) {
+    int boffset =
+        2 * (b_width_log2_lookup[cm->sb_size] -
+             AOMMIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
+    step_param = AOMMAX(step_param, boffset);
+  }
+
+  if (cpi->sf.adaptive_motion_search) {
+    int bwl = b_width_log2_lookup[bsize];
+    int bhl = b_height_log2_lookup[bsize];
+    int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
+
+    if (tlevel < 5) step_param += 2;
+
+    // prev_mv_sad is not setup for dynamically scaled frames.
+    if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
+      int i;
+      for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
+        if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
+          x->pred_mv[ref].row = 0;
+          x->pred_mv[ref].col = 0;
+          x->best_mv.as_int = INVALID_MV;
+
+          if (scaled_ref_frame) {
+            int j;
+            for (j = 0; j < MAX_MB_PLANE; ++j)
+              xd->plane[j].pre[ref_idx] = backup_yv12[j];
+          }
+          return;
+        }
+      }
+    }
+  }
+
+  av1_set_mv_search_range(&x->mv_limits, &ref_mv);
+
+#if CONFIG_MOTION_VAR
+  if (mbmi->motion_mode != SIMPLE_TRANSLATION)
+    mvp_full = mbmi->mv[0].as_mv;
+  else
+#endif  // CONFIG_MOTION_VAR
+    mvp_full = pred_mv[x->mv_best_ref_index[ref]];
+
+  mvp_full.col >>= 3;
+  mvp_full.row >>= 3;
+
+  x->best_mv.as_int = x->second_best_mv.as_int = INVALID_MV;
+
+#if CONFIG_MOTION_VAR
+  switch (mbmi->motion_mode) {
+    case SIMPLE_TRANSLATION:
+#endif  // CONFIG_MOTION_VAR
+      bestsme = av1_full_pixel_search(cpi, x, bsize, &mvp_full, step_param,
+                                      sadpb, cond_cost_list(cpi, cost_list),
+                                      &ref_mv, INT_MAX, 1);
+#if CONFIG_MOTION_VAR
+      break;
+    case OBMC_CAUSAL:
+      bestsme = av1_obmc_full_pixel_diamond(
+          cpi, x, &mvp_full, step_param, sadpb,
+          MAX_MVSEARCH_STEPS - 1 - step_param, 1, &cpi->fn_ptr[bsize], &ref_mv,
+          &(x->best_mv.as_mv), 0);
+      break;
+    default: assert("Invalid motion mode!\n");
+  }
+#endif  // CONFIG_MOTION_VAR
+
+  x->mv_limits = tmp_mv_limits;
+
+  if (bestsme < INT_MAX) {
+    int dis; /* TODO: use dis in distortion calculation later. */
+#if CONFIG_MOTION_VAR
+    switch (mbmi->motion_mode) {
+      case SIMPLE_TRANSLATION:
+#endif  // CONFIG_MOTION_VAR
+        if (cpi->sf.use_upsampled_references) {
+          int best_mv_var;
+          const int try_second = x->second_best_mv.as_int != INVALID_MV &&
+                                 x->second_best_mv.as_int != x->best_mv.as_int;
+          const int pw = block_size_wide[bsize];
+          const int ph = block_size_high[bsize];
+          // Use up-sampled reference frames.
+          struct macroblockd_plane *const pd = &xd->plane[0];
+          struct buf_2d backup_pred = pd->pre[ref_idx];
+          const YV12_BUFFER_CONFIG *upsampled_ref = get_upsampled_ref(cpi, ref);
+
+          // Set pred for Y plane
+          setup_pred_plane(
+              &pd->pre[ref_idx], bsize, upsampled_ref->y_buffer,
+              upsampled_ref->y_crop_width, upsampled_ref->y_crop_height,
+              upsampled_ref->y_stride, (mi_row << 3), (mi_col << 3), NULL,
+              pd->subsampling_x, pd->subsampling_y);
+
+          best_mv_var = cpi->find_fractional_mv_step(
+              x, &ref_mv, cm->allow_high_precision_mv, x->errorperbit,
+              &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
+              cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
+              x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, pw, ph,
+              1);
+
+          if (try_second) {
+            const int minc =
+                AOMMAX(x->mv_limits.col_min * 8, ref_mv.col - MV_MAX);
+            const int maxc =
+                AOMMIN(x->mv_limits.col_max * 8, ref_mv.col + MV_MAX);
+            const int minr =
+                AOMMAX(x->mv_limits.row_min * 8, ref_mv.row - MV_MAX);
+            const int maxr =
+                AOMMIN(x->mv_limits.row_max * 8, ref_mv.row + MV_MAX);
+            int this_var;
+            MV best_mv = x->best_mv.as_mv;
+
+            x->best_mv = x->second_best_mv;
+            if (x->best_mv.as_mv.row * 8 <= maxr &&
+                x->best_mv.as_mv.row * 8 >= minr &&
+                x->best_mv.as_mv.col * 8 <= maxc &&
+                x->best_mv.as_mv.col * 8 >= minc) {
+              this_var = cpi->find_fractional_mv_step(
+                  x, &ref_mv, cm->allow_high_precision_mv, x->errorperbit,
+                  &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
+                  cpi->sf.mv.subpel_iters_per_step,
+                  cond_cost_list(cpi, cost_list), x->nmvjointcost, x->mvcost,
+                  &dis, &x->pred_sse[ref], NULL, pw, ph, 1);
+              if (this_var < best_mv_var) best_mv = x->best_mv.as_mv;
+              x->best_mv.as_mv = best_mv;
+            }
+          }
+
+          // Restore the reference frames.
+          pd->pre[ref_idx] = backup_pred;
+        } else {
+          cpi->find_fractional_mv_step(
+              x, &ref_mv, cm->allow_high_precision_mv, x->errorperbit,
+              &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
+              cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
+              x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0,
+              0);
+        }
+#if CONFIG_MOTION_VAR
+        break;
+      case OBMC_CAUSAL:
+        av1_find_best_obmc_sub_pixel_tree_up(
+            cpi, x, mi_row, mi_col, &x->best_mv.as_mv, &ref_mv,
+            cm->allow_high_precision_mv, x->errorperbit, &cpi->fn_ptr[bsize],
+            cpi->sf.mv.subpel_force_stop, cpi->sf.mv.subpel_iters_per_step,
+            x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], 0,
+            cpi->sf.use_upsampled_references);
+        break;
+      default: assert("Invalid motion mode!\n");
+    }
+#endif  // CONFIG_MOTION_VAR
+  }
+  *rate_mv = av1_mv_bit_cost(&x->best_mv.as_mv, &ref_mv, x->nmvjointcost,
+                             x->mvcost, MV_COST_WEIGHT);
+
+#if CONFIG_MOTION_VAR
+  if (cpi->sf.adaptive_motion_search && mbmi->motion_mode == SIMPLE_TRANSLATION)
+#else
+  if (cpi->sf.adaptive_motion_search)
+#endif  // CONFIG_MOTION_VAR
+    x->pred_mv[ref] = x->best_mv.as_mv;
+
+  if (scaled_ref_frame) {
+    int i;
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      xd->plane[i].pre[ref_idx] = backup_yv12[i];
+  }
+}
+
+static INLINE void restore_dst_buf(MACROBLOCKD *xd, BUFFER_SET dst) {
+  int i;
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    xd->plane[i].dst.buf = dst.plane[i];
+    xd->plane[i].dst.stride = dst.stride[i];
+  }
+}
+
+#if CONFIG_EXT_INTER
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+static void do_masked_motion_search(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                    const uint8_t *mask, int mask_stride,
+                                    BLOCK_SIZE bsize, int mi_row, int mi_col,
+                                    int_mv *tmp_mv, int *rate_mv, int ref_idx) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  const AV1_COMMON *cm = &cpi->common;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0, 0, 0, 0 } };
+  int bestsme = INT_MAX;
+  int step_param;
+  int sadpb = x->sadperbit16;
+  MV mvp_full;
+  int ref = mbmi->ref_frame[ref_idx];
+  MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
+
+  MvLimits tmp_mv_limits = x->mv_limits;
+
+  const YV12_BUFFER_CONFIG *scaled_ref_frame =
+      av1_get_scaled_ref_frame(cpi, ref);
+  int i;
+
+  MV pred_mv[3];
+  pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
+  pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
+  pred_mv[2] = x->pred_mv[ref];
+
+#if CONFIG_REF_MV
+  av1_set_mvcost(x, ref, ref_idx, mbmi->ref_mv_idx);
+#endif  // CONFIG_REF_MV
+
+  if (scaled_ref_frame) {
+    // Swap out the reference frame for a version that's been scaled to
+    // match the resolution of the current frame, allowing the existing
+    // motion search code to be used without additional modifications.
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      backup_yv12[i] = xd->plane[i].pre[ref_idx];
+
+    av1_setup_pre_planes(xd, ref_idx, scaled_ref_frame, mi_row, mi_col, NULL);
+  }
+
+  av1_set_mv_search_range(&x->mv_limits, &ref_mv);
+
+  // Work out the size of the first step in the mv step search.
+  // 0 here is maximum length first step. 1 is MAX >> 1 etc.
+  if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
+    // Take wtd average of the step_params based on the last frame's
+    // max mv magnitude and that based on the best ref mvs of the current
+    // block for the given reference.
+    step_param =
+        (av1_init_search_range(x->max_mv_context[ref]) + cpi->mv_step_param) /
+        2;
+  } else {
+    step_param = cpi->mv_step_param;
+  }
+
+  // TODO(debargha): is show_frame needed here?
+  if (cpi->sf.adaptive_motion_search && bsize < cm->sb_size && cm->show_frame) {
+    int boffset =
+        2 * (b_width_log2_lookup[cm->sb_size] -
+             AOMMIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
+    step_param = AOMMAX(step_param, boffset);
+  }
+
+  if (cpi->sf.adaptive_motion_search) {
+    int bwl = b_width_log2_lookup[bsize];
+    int bhl = b_height_log2_lookup[bsize];
+    int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
+
+    if (tlevel < 5) step_param += 2;
+
+    // prev_mv_sad is not setup for dynamically scaled frames.
+    if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
+      for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
+        if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
+          x->pred_mv[ref].row = 0;
+          x->pred_mv[ref].col = 0;
+          tmp_mv->as_int = INVALID_MV;
+
+          if (scaled_ref_frame) {
+            int j;
+            for (j = 0; j < MAX_MB_PLANE; ++j)
+              xd->plane[j].pre[ref_idx] = backup_yv12[j];
+          }
+          return;
+        }
+      }
+    }
+  }
+
+  mvp_full = pred_mv[x->mv_best_ref_index[ref]];
+
+  mvp_full.col >>= 3;
+  mvp_full.row >>= 3;
+
+  bestsme = av1_masked_full_pixel_diamond(
+      cpi, x, mask, mask_stride, &mvp_full, step_param, sadpb,
+      MAX_MVSEARCH_STEPS - 1 - step_param, 1, &cpi->fn_ptr[bsize], &ref_mv,
+      &tmp_mv->as_mv, ref_idx);
+
+  x->mv_limits = tmp_mv_limits;
+
+  if (bestsme < INT_MAX) {
+    int dis; /* TODO: use dis in distortion calculation later. */
+    av1_find_best_masked_sub_pixel_tree_up(
+        cpi, x, mask, mask_stride, mi_row, mi_col, &tmp_mv->as_mv, &ref_mv,
+        cm->allow_high_precision_mv, x->errorperbit, &cpi->fn_ptr[bsize],
+        cpi->sf.mv.subpel_force_stop, cpi->sf.mv.subpel_iters_per_step,
+        x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], ref_idx,
+        cpi->sf.use_upsampled_references);
+  }
+  *rate_mv = av1_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
+                             x->mvcost, MV_COST_WEIGHT);
+
+  if (cpi->sf.adaptive_motion_search && cm->show_frame)
+    x->pred_mv[ref] = tmp_mv->as_mv;
+
+  if (scaled_ref_frame) {
+    for (i = 0; i < MAX_MB_PLANE; i++)
+      xd->plane[i].pre[ref_idx] = backup_yv12[i];
+  }
+}
+
+static void do_masked_motion_search_indexed(
+    const AV1_COMP *const cpi, MACROBLOCK *x,
+    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE bsize,
+    int mi_row, int mi_col, int_mv *tmp_mv, int *rate_mv, int which) {
+  // NOTE: which values: 0 - 0 only, 1 - 1 only, 2 - both
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  BLOCK_SIZE sb_type = mbmi->sb_type;
+  const uint8_t *mask;
+  const int mask_stride = block_size_wide[bsize];
+
+  mask = av1_get_compound_type_mask(comp_data, sb_type);
+
+  if (which == 0 || which == 2)
+    do_masked_motion_search(cpi, x, mask, mask_stride, bsize, mi_row, mi_col,
+                            &tmp_mv[0], &rate_mv[0], 0);
+
+  if (which == 1 || which == 2) {
+// get the negative mask
+#if CONFIG_COMPOUND_SEGMENT
+    uint8_t inv_mask_buf[2 * MAX_SB_SQUARE];
+    const int h = block_size_high[bsize];
+    mask = av1_get_compound_type_mask_inverse(
+        comp_data, inv_mask_buf, h, mask_stride, mask_stride, sb_type);
+#else
+    mask = av1_get_compound_type_mask_inverse(comp_data, sb_type);
+#endif  // CONFIG_COMPOUND_SEGMENT
+    do_masked_motion_search(cpi, x, mask, mask_stride, bsize, mi_row, mi_col,
+                            &tmp_mv[1], &rate_mv[1], 1);
+  }
+}
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#endif  // CONFIG_EXT_INTER
+
+// In some situations we want to discount tha pparent cost of a new motion
+// vector. Where there is a subtle motion field and especially where there is
+// low spatial complexity then it can be hard to cover the cost of a new motion
+// vector in a single block, even if that motion vector reduces distortion.
+// However, once established that vector may be usable through the nearest and
+// near mv modes to reduce distortion in subsequent blocks and also improve
+// visual quality.
+static int discount_newmv_test(const AV1_COMP *const cpi, int this_mode,
+                               int_mv this_mv,
+                               int_mv (*mode_mv)[TOTAL_REFS_PER_FRAME],
+                               int ref_frame) {
+  return (!cpi->rc.is_src_frame_alt_ref && (this_mode == NEWMV) &&
+          (this_mv.as_int != 0) &&
+          ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
+           (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
+          ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
+           (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
+}
+
+#define LEFT_TOP_MARGIN ((AOM_BORDER_IN_PIXELS - AOM_INTERP_EXTEND) << 3)
+#define RIGHT_BOTTOM_MARGIN ((AOM_BORDER_IN_PIXELS - AOM_INTERP_EXTEND) << 3)
+
+// TODO(jingning): this mv clamping function should be block size dependent.
+static INLINE void clamp_mv2(MV *mv, const MACROBLOCKD *xd) {
+  clamp_mv(mv, xd->mb_to_left_edge - LEFT_TOP_MARGIN,
+           xd->mb_to_right_edge + RIGHT_BOTTOM_MARGIN,
+           xd->mb_to_top_edge - LEFT_TOP_MARGIN,
+           xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN);
+}
+
+#if CONFIG_EXT_INTER
+#if CONFIG_WEDGE
+static int estimate_wedge_sign(const AV1_COMP *cpi, const MACROBLOCK *x,
+                               const BLOCK_SIZE bsize, const uint8_t *pred0,
+                               int stride0, const uint8_t *pred1, int stride1) {
+  const struct macroblock_plane *const p = &x->plane[0];
+  const uint8_t *src = p->src.buf;
+  int src_stride = p->src.stride;
+  const int f_index = bsize - BLOCK_8X8;
+  const int bw = block_size_wide[bsize];
+  const int bh = block_size_high[bsize];
+  uint32_t esq[2][4], var;
+  int64_t tl, br;
+
+#if CONFIG_HIGHBITDEPTH
+  if (x->e_mbd.cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    pred0 = CONVERT_TO_BYTEPTR(pred0);
+    pred1 = CONVERT_TO_BYTEPTR(pred1);
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+
+  var = cpi->fn_ptr[f_index].vf(src, src_stride, pred0, stride0, &esq[0][0]);
+  var = cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, pred0 + bw / 2,
+                                stride0, &esq[0][1]);
+  var = cpi->fn_ptr[f_index].vf(src + bh / 2 * src_stride, src_stride,
+                                pred0 + bh / 2 * stride0, stride0, &esq[0][2]);
+  var = cpi->fn_ptr[f_index].vf(src + bh / 2 * src_stride + bw / 2, src_stride,
+                                pred0 + bh / 2 * stride0 + bw / 2, stride0,
+                                &esq[0][3]);
+  var = cpi->fn_ptr[f_index].vf(src, src_stride, pred1, stride1, &esq[1][0]);
+  var = cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, pred1 + bw / 2,
+                                stride1, &esq[1][1]);
+  var = cpi->fn_ptr[f_index].vf(src + bh / 2 * src_stride, src_stride,
+                                pred1 + bh / 2 * stride1, stride0, &esq[1][2]);
+  var = cpi->fn_ptr[f_index].vf(src + bh / 2 * src_stride + bw / 2, src_stride,
+                                pred1 + bh / 2 * stride1 + bw / 2, stride0,
+                                &esq[1][3]);
+  (void)var;
+
+  tl = (int64_t)(esq[0][0] + esq[0][1] + esq[0][2]) -
+       (int64_t)(esq[1][0] + esq[1][1] + esq[1][2]);
+  br = (int64_t)(esq[1][3] + esq[1][1] + esq[1][2]) -
+       (int64_t)(esq[0][3] + esq[0][1] + esq[0][2]);
+  return (tl + br > 0);
+}
+#endif  // CONFIG_WEDGE
+#endif  // CONFIG_EXT_INTER
+
+#if !CONFIG_DUAL_FILTER
+static InterpFilter predict_interp_filter(
+    const AV1_COMP *cpi, const MACROBLOCK *x, const BLOCK_SIZE bsize,
+    const int mi_row, const int mi_col,
+    InterpFilter (*single_filter)[TOTAL_REFS_PER_FRAME]) {
+  InterpFilter best_filter = SWITCHABLE;
+  const AV1_COMMON *cm = &cpi->common;
+  const MACROBLOCKD *xd = &x->e_mbd;
+  int bsl = mi_width_log2_lookup[bsize];
+  int pred_filter_search =
+      cpi->sf.cb_pred_filter_search
+          ? (((mi_row + mi_col) >> bsl) +
+             get_chessboard_index(cm->current_video_frame)) &
+                0x1
+          : 0;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  const int is_comp_pred = has_second_ref(mbmi);
+  const int this_mode = mbmi->mode;
+  int refs[2] = { mbmi->ref_frame[0],
+                  (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
+  if (pred_filter_search) {
+    InterpFilter af = SWITCHABLE, lf = SWITCHABLE;
+    if (xd->up_available) af = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
+    if (xd->left_available) lf = xd->mi[-1]->mbmi.interp_filter;
+
+#if CONFIG_EXT_INTER
+    if ((this_mode != NEWMV && this_mode != NEW_NEWMV) || (af == lf))
+#else
+    if ((this_mode != NEWMV) || (af == lf))
+#endif  // CONFIG_EXT_INTER
+      best_filter = af;
+  }
+  if (is_comp_pred) {
+    if (cpi->sf.adaptive_mode_search) {
+#if CONFIG_EXT_INTER
+      switch (this_mode) {
+        case NEAREST_NEARESTMV:
+          if (single_filter[NEARESTMV][refs[0]] ==
+              single_filter[NEARESTMV][refs[1]])
+            best_filter = single_filter[NEARESTMV][refs[0]];
+          break;
+        case NEAREST_NEARMV:
+          if (single_filter[NEARESTMV][refs[0]] ==
+              single_filter[NEARMV][refs[1]])
+            best_filter = single_filter[NEARESTMV][refs[0]];
+          break;
+        case NEAR_NEARESTMV:
+          if (single_filter[NEARMV][refs[0]] ==
+              single_filter[NEARESTMV][refs[1]])
+            best_filter = single_filter[NEARMV][refs[0]];
+          break;
+        case NEAR_NEARMV:
+          if (single_filter[NEARMV][refs[0]] == single_filter[NEARMV][refs[1]])
+            best_filter = single_filter[NEARMV][refs[0]];
+          break;
+        case ZERO_ZEROMV:
+          if (single_filter[ZEROMV][refs[0]] == single_filter[ZEROMV][refs[1]])
+            best_filter = single_filter[ZEROMV][refs[0]];
+          break;
+        case NEW_NEWMV:
+          if (single_filter[NEWMV][refs[0]] == single_filter[NEWMV][refs[1]])
+            best_filter = single_filter[NEWMV][refs[0]];
+          break;
+        case NEAREST_NEWMV:
+          if (single_filter[NEARESTMV][refs[0]] ==
+              single_filter[NEWMV][refs[1]])
+            best_filter = single_filter[NEARESTMV][refs[0]];
+          break;
+        case NEAR_NEWMV:
+          if (single_filter[NEARMV][refs[0]] == single_filter[NEWMV][refs[1]])
+            best_filter = single_filter[NEARMV][refs[0]];
+          break;
+        case NEW_NEARESTMV:
+          if (single_filter[NEWMV][refs[0]] ==
+              single_filter[NEARESTMV][refs[1]])
+            best_filter = single_filter[NEWMV][refs[0]];
+          break;
+        case NEW_NEARMV:
+          if (single_filter[NEWMV][refs[0]] == single_filter[NEARMV][refs[1]])
+            best_filter = single_filter[NEWMV][refs[0]];
+          break;
+        default:
+          if (single_filter[this_mode][refs[0]] ==
+              single_filter[this_mode][refs[1]])
+            best_filter = single_filter[this_mode][refs[0]];
+          break;
+      }
+#else
+      if (single_filter[this_mode][refs[0]] ==
+          single_filter[this_mode][refs[1]])
+        best_filter = single_filter[this_mode][refs[0]];
+#endif  // CONFIG_EXT_INTER
+    }
+  }
+  if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
+    best_filter = EIGHTTAP_REGULAR;
+  }
+  return best_filter;
+}
+#endif  // !CONFIG_DUAL_FILTER
+
+#if CONFIG_EXT_INTER
+// Choose the best wedge index and sign
+#if CONFIG_WEDGE
+static int64_t pick_wedge(const AV1_COMP *const cpi, const MACROBLOCK *const x,
+                          const BLOCK_SIZE bsize, const uint8_t *const p0,
+                          const uint8_t *const p1, int *const best_wedge_sign,
+                          int *const best_wedge_index) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const src = &x->plane[0].src;
+  const int bw = block_size_wide[bsize];
+  const int bh = block_size_high[bsize];
+  const int N = bw * bh;
+  int rate;
+  int64_t dist;
+  int64_t rd, best_rd = INT64_MAX;
+  int wedge_index;
+  int wedge_sign;
+  int wedge_types = (1 << get_wedge_bits_lookup(bsize));
+  const uint8_t *mask;
+  uint64_t sse;
+#if CONFIG_HIGHBITDEPTH
+  const int hbd = xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH;
+  const int bd_round = hbd ? (xd->bd - 8) * 2 : 0;
+#else
+  const int bd_round = 0;
+#endif  // CONFIG_HIGHBITDEPTH
+
+  DECLARE_ALIGNED(32, int16_t, r0[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, r1[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, d10[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, ds[MAX_SB_SQUARE]);
+
+  int64_t sign_limit;
+
+#if CONFIG_HIGHBITDEPTH
+  if (hbd) {
+    aom_highbd_subtract_block(bh, bw, r0, bw, src->buf, src->stride,
+                              CONVERT_TO_BYTEPTR(p0), bw, xd->bd);
+    aom_highbd_subtract_block(bh, bw, r1, bw, src->buf, src->stride,
+                              CONVERT_TO_BYTEPTR(p1), bw, xd->bd);
+    aom_highbd_subtract_block(bh, bw, d10, bw, CONVERT_TO_BYTEPTR(p1), bw,
+                              CONVERT_TO_BYTEPTR(p0), bw, xd->bd);
+  } else  // NOLINT
+#endif    // CONFIG_HIGHBITDEPTH
+  {
+    aom_subtract_block(bh, bw, r0, bw, src->buf, src->stride, p0, bw);
+    aom_subtract_block(bh, bw, r1, bw, src->buf, src->stride, p1, bw);
+    aom_subtract_block(bh, bw, d10, bw, p1, bw, p0, bw);
+  }
+
+  sign_limit = ((int64_t)aom_sum_squares_i16(r0, N) -
+                (int64_t)aom_sum_squares_i16(r1, N)) *
+               (1 << WEDGE_WEIGHT_BITS) / 2;
+
+  if (N < 64)
+    av1_wedge_compute_delta_squares_c(ds, r0, r1, N);
+  else
+    av1_wedge_compute_delta_squares(ds, r0, r1, N);
+
+  for (wedge_index = 0; wedge_index < wedge_types; ++wedge_index) {
+    mask = av1_get_contiguous_soft_mask(wedge_index, 0, bsize);
+
+    // TODO(jingning): Make sse2 functions support N = 16 case
+    if (N < 64)
+      wedge_sign = av1_wedge_sign_from_residuals_c(ds, mask, N, sign_limit);
+    else
+      wedge_sign = av1_wedge_sign_from_residuals(ds, mask, N, sign_limit);
+
+    mask = av1_get_contiguous_soft_mask(wedge_index, wedge_sign, bsize);
+    if (N < 64)
+      sse = av1_wedge_sse_from_residuals_c(r1, d10, mask, N);
+    else
+      sse = av1_wedge_sse_from_residuals(r1, d10, mask, N);
+    sse = ROUND_POWER_OF_TWO(sse, bd_round);
+
+    model_rd_from_sse(cpi, xd, bsize, 0, sse, &rate, &dist);
+    rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
+
+    if (rd < best_rd) {
+      *best_wedge_index = wedge_index;
+      *best_wedge_sign = wedge_sign;
+      best_rd = rd;
+    }
+  }
+
+  return best_rd;
+}
+
+// Choose the best wedge index the specified sign
+static int64_t pick_wedge_fixed_sign(
+    const AV1_COMP *const cpi, const MACROBLOCK *const x,
+    const BLOCK_SIZE bsize, const uint8_t *const p0, const uint8_t *const p1,
+    const int wedge_sign, int *const best_wedge_index) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const struct buf_2d *const src = &x->plane[0].src;
+  const int bw = block_size_wide[bsize];
+  const int bh = block_size_high[bsize];
+  const int N = bw * bh;
+  int rate;
+  int64_t dist;
+  int64_t rd, best_rd = INT64_MAX;
+  int wedge_index;
+  int wedge_types = (1 << get_wedge_bits_lookup(bsize));
+  const uint8_t *mask;
+  uint64_t sse;
+#if CONFIG_HIGHBITDEPTH
+  const int hbd = xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH;
+  const int bd_round = hbd ? (xd->bd - 8) * 2 : 0;
+#else
+  const int bd_round = 0;
+#endif  // CONFIG_HIGHBITDEPTH
+
+  DECLARE_ALIGNED(32, int16_t, r1[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, d10[MAX_SB_SQUARE]);
+
+#if CONFIG_HIGHBITDEPTH
+  if (hbd) {
+    aom_highbd_subtract_block(bh, bw, r1, bw, src->buf, src->stride,
+                              CONVERT_TO_BYTEPTR(p1), bw, xd->bd);
+    aom_highbd_subtract_block(bh, bw, d10, bw, CONVERT_TO_BYTEPTR(p1), bw,
+                              CONVERT_TO_BYTEPTR(p0), bw, xd->bd);
+  } else  // NOLINT
+#endif    // CONFIG_HIGHBITDEPTH
+  {
+    aom_subtract_block(bh, bw, r1, bw, src->buf, src->stride, p1, bw);
+    aom_subtract_block(bh, bw, d10, bw, p1, bw, p0, bw);
+  }
+
+  for (wedge_index = 0; wedge_index < wedge_types; ++wedge_index) {
+    mask = av1_get_contiguous_soft_mask(wedge_index, wedge_sign, bsize);
+    if (N < 64)
+      sse = av1_wedge_sse_from_residuals_c(r1, d10, mask, N);
+    else
+      sse = av1_wedge_sse_from_residuals(r1, d10, mask, N);
+    sse = ROUND_POWER_OF_TWO(sse, bd_round);
+
+    model_rd_from_sse(cpi, xd, bsize, 0, sse, &rate, &dist);
+    rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
+
+    if (rd < best_rd) {
+      *best_wedge_index = wedge_index;
+      best_rd = rd;
+    }
+  }
+
+  return best_rd;
+}
+
+static int64_t pick_interinter_wedge(const AV1_COMP *const cpi,
+                                     MACROBLOCK *const x,
+                                     const BLOCK_SIZE bsize,
+                                     const uint8_t *const p0,
+                                     const uint8_t *const p1) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int bw = block_size_wide[bsize];
+
+  int64_t rd;
+  int wedge_index = -1;
+  int wedge_sign = 0;
+
+  assert(is_interinter_compound_used(COMPOUND_WEDGE, bsize));
+
+  if (cpi->sf.fast_wedge_sign_estimate) {
+    wedge_sign = estimate_wedge_sign(cpi, x, bsize, p0, bw, p1, bw);
+    rd = pick_wedge_fixed_sign(cpi, x, bsize, p0, p1, wedge_sign, &wedge_index);
+  } else {
+    rd = pick_wedge(cpi, x, bsize, p0, p1, &wedge_sign, &wedge_index);
+  }
+
+  mbmi->wedge_sign = wedge_sign;
+  mbmi->wedge_index = wedge_index;
+  return rd;
+}
+#endif  // CONFIG_WEDGE
+
+#if CONFIG_COMPOUND_SEGMENT
+static int64_t pick_interinter_seg(const AV1_COMP *const cpi,
+                                   MACROBLOCK *const x, const BLOCK_SIZE bsize,
+                                   const uint8_t *const p0,
+                                   const uint8_t *const p1) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const struct buf_2d *const src = &x->plane[0].src;
+  const int bw = block_size_wide[bsize];
+  const int bh = block_size_high[bsize];
+  const int N = bw * bh;
+  int rate;
+  uint64_t sse;
+  int64_t dist;
+  int64_t rd0;
+  SEG_MASK_TYPE cur_mask_type;
+  int64_t best_rd = INT64_MAX;
+  SEG_MASK_TYPE best_mask_type = 0;
+#if CONFIG_HIGHBITDEPTH
+  const int hbd = xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH;
+  const int bd_round = hbd ? (xd->bd - 8) * 2 : 0;
+#else
+  const int bd_round = 0;
+#endif  // CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(32, int16_t, r0[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, r1[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, d10[MAX_SB_SQUARE]);
+
+#if CONFIG_HIGHBITDEPTH
+  if (hbd) {
+    aom_highbd_subtract_block(bh, bw, r0, bw, src->buf, src->stride,
+                              CONVERT_TO_BYTEPTR(p0), bw, xd->bd);
+    aom_highbd_subtract_block(bh, bw, r1, bw, src->buf, src->stride,
+                              CONVERT_TO_BYTEPTR(p1), bw, xd->bd);
+    aom_highbd_subtract_block(bh, bw, d10, bw, CONVERT_TO_BYTEPTR(p1), bw,
+                              CONVERT_TO_BYTEPTR(p0), bw, xd->bd);
+  } else  // NOLINT
+#endif    // CONFIG_HIGHBITDEPTH
+  {
+    aom_subtract_block(bh, bw, r0, bw, src->buf, src->stride, p0, bw);
+    aom_subtract_block(bh, bw, r1, bw, src->buf, src->stride, p1, bw);
+    aom_subtract_block(bh, bw, d10, bw, p1, bw, p0, bw);
+  }
+
+  // try each mask type and its inverse
+  for (cur_mask_type = 0; cur_mask_type < SEG_MASK_TYPES; cur_mask_type++) {
+// build mask and inverse
+#if CONFIG_HIGHBITDEPTH
+    if (hbd)
+      build_compound_seg_mask_highbd(
+          xd->seg_mask, cur_mask_type, CONVERT_TO_BYTEPTR(p0), bw,
+          CONVERT_TO_BYTEPTR(p1), bw, bsize, bh, bw, xd->bd);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      build_compound_seg_mask(xd->seg_mask, cur_mask_type, p0, bw, p1, bw,
+                              bsize, bh, bw);
+
+    // compute rd for mask
+    sse = av1_wedge_sse_from_residuals(r1, d10, xd->seg_mask, N);
+    sse = ROUND_POWER_OF_TWO(sse, bd_round);
+
+    model_rd_from_sse(cpi, xd, bsize, 0, sse, &rate, &dist);
+    rd0 = RDCOST(x->rdmult, x->rddiv, rate, dist);
+
+    if (rd0 < best_rd) {
+      best_mask_type = cur_mask_type;
+      best_rd = rd0;
+    }
+  }
+
+  // make final mask
+  mbmi->mask_type = best_mask_type;
+#if CONFIG_HIGHBITDEPTH
+  if (hbd)
+    build_compound_seg_mask_highbd(
+        xd->seg_mask, mbmi->mask_type, CONVERT_TO_BYTEPTR(p0), bw,
+        CONVERT_TO_BYTEPTR(p1), bw, bsize, bh, bw, xd->bd);
+  else
+#endif  // CONFIG_HIGHBITDEPTH
+    build_compound_seg_mask(xd->seg_mask, mbmi->mask_type, p0, bw, p1, bw,
+                            bsize, bh, bw);
+
+  return best_rd;
+}
+#endif  // CONFIG_COMPOUND_SEGMENT
+
+#if CONFIG_WEDGE && CONFIG_INTERINTRA
+static int64_t pick_interintra_wedge(const AV1_COMP *const cpi,
+                                     const MACROBLOCK *const x,
+                                     const BLOCK_SIZE bsize,
+                                     const uint8_t *const p0,
+                                     const uint8_t *const p1) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+
+  int64_t rd;
+  int wedge_index = -1;
+
+  assert(is_interintra_wedge_used(bsize));
+
+  rd = pick_wedge_fixed_sign(cpi, x, bsize, p0, p1, 0, &wedge_index);
+
+  mbmi->interintra_wedge_sign = 0;
+  mbmi->interintra_wedge_index = wedge_index;
+  return rd;
+}
+#endif  // CONFIG_WEDGE && CONFIG_INTERINTRA
+
+#if CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+static int64_t pick_interinter_mask(const AV1_COMP *const cpi, MACROBLOCK *x,
+                                    const BLOCK_SIZE bsize,
+                                    const uint8_t *const p0,
+                                    const uint8_t *const p1) {
+  const COMPOUND_TYPE compound_type =
+      x->e_mbd.mi[0]->mbmi.interinter_compound_type;
+  switch (compound_type) {
+#if CONFIG_WEDGE
+    case COMPOUND_WEDGE: return pick_interinter_wedge(cpi, x, bsize, p0, p1);
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+    case COMPOUND_SEG: return pick_interinter_seg(cpi, x, bsize, p0, p1);
+#endif  // CONFIG_COMPOUND_SEGMENT
+    default: assert(0); return 0;
+  }
+}
+
+static int interinter_compound_motion_search(const AV1_COMP *const cpi,
+                                             MACROBLOCK *x,
+                                             const BLOCK_SIZE bsize,
+                                             const int this_mode, int mi_row,
+                                             int mi_col) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int_mv tmp_mv[2];
+  int rate_mvs[2], tmp_rate_mv = 0;
+  const INTERINTER_COMPOUND_DATA compound_data = {
+#if CONFIG_WEDGE
+    mbmi->wedge_index,
+    mbmi->wedge_sign,
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+    mbmi->mask_type,
+    xd->seg_mask,
+#endif  // CONFIG_COMPOUND_SEGMENT
+    mbmi->interinter_compound_type
+  };
+  if (this_mode == NEW_NEWMV) {
+    do_masked_motion_search_indexed(cpi, x, &compound_data, bsize, mi_row,
+                                    mi_col, tmp_mv, rate_mvs, 2);
+    tmp_rate_mv = rate_mvs[0] + rate_mvs[1];
+    mbmi->mv[0].as_int = tmp_mv[0].as_int;
+    mbmi->mv[1].as_int = tmp_mv[1].as_int;
+  } else if (this_mode == NEW_NEARESTMV || this_mode == NEW_NEARMV) {
+    do_masked_motion_search_indexed(cpi, x, &compound_data, bsize, mi_row,
+                                    mi_col, tmp_mv, rate_mvs, 0);
+    tmp_rate_mv = rate_mvs[0];
+    mbmi->mv[0].as_int = tmp_mv[0].as_int;
+  } else if (this_mode == NEAREST_NEWMV || this_mode == NEAR_NEWMV) {
+    do_masked_motion_search_indexed(cpi, x, &compound_data, bsize, mi_row,
+                                    mi_col, tmp_mv, rate_mvs, 1);
+    tmp_rate_mv = rate_mvs[1];
+    mbmi->mv[1].as_int = tmp_mv[1].as_int;
+  }
+  return tmp_rate_mv;
+}
+
+static int64_t build_and_cost_compound_type(
+    const AV1_COMP *const cpi, MACROBLOCK *x, const int_mv *const cur_mv,
+    const BLOCK_SIZE bsize, const int this_mode, int rs2, int rate_mv,
+    BUFFER_SET *ctx, int *out_rate_mv, uint8_t **preds0, uint8_t **preds1,
+    int *strides, int mi_row, int mi_col) {
+  MACROBLOCKD *xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int rate_sum;
+  int64_t dist_sum;
+  int64_t best_rd_cur = INT64_MAX;
+  int64_t rd = INT64_MAX;
+  int tmp_skip_txfm_sb;
+  int64_t tmp_skip_sse_sb;
+  const COMPOUND_TYPE compound_type = mbmi->interinter_compound_type;
+
+  best_rd_cur = pick_interinter_mask(cpi, x, bsize, *preds0, *preds1);
+  best_rd_cur += RDCOST(x->rdmult, x->rddiv, rs2 + rate_mv, 0);
+
+  if (have_newmv_in_inter_mode(this_mode) &&
+      use_masked_motion_search(compound_type)) {
+    *out_rate_mv = interinter_compound_motion_search(cpi, x, bsize, this_mode,
+                                                     mi_row, mi_col);
+    av1_build_inter_predictors_sby(xd, mi_row, mi_col, ctx, bsize);
+    model_rd_for_sb(cpi, bsize, x, xd, 0, 0, &rate_sum, &dist_sum,
+                    &tmp_skip_txfm_sb, &tmp_skip_sse_sb);
+    rd = RDCOST(x->rdmult, x->rddiv, rs2 + *out_rate_mv + rate_sum, dist_sum);
+    if (rd >= best_rd_cur) {
+      mbmi->mv[0].as_int = cur_mv[0].as_int;
+      mbmi->mv[1].as_int = cur_mv[1].as_int;
+      *out_rate_mv = rate_mv;
+      av1_build_wedge_inter_predictor_from_buf(xd, bsize, 0, 0,
+#if CONFIG_SUPERTX
+                                               0, 0,
+#endif  // CONFIG_SUPERTX
+                                               preds0, strides, preds1,
+                                               strides);
+    }
+    av1_subtract_plane(x, bsize, 0);
+    rd = estimate_yrd_for_sb(cpi, bsize, x, &rate_sum, &dist_sum,
+                             &tmp_skip_txfm_sb, &tmp_skip_sse_sb, INT64_MAX);
+    if (rd != INT64_MAX)
+      rd = RDCOST(x->rdmult, x->rddiv, rs2 + *out_rate_mv + rate_sum, dist_sum);
+    best_rd_cur = rd;
+
+  } else {
+    av1_build_wedge_inter_predictor_from_buf(xd, bsize, 0, 0,
+#if CONFIG_SUPERTX
+                                             0, 0,
+#endif  // CONFIG_SUPERTX
+                                             preds0, strides, preds1, strides);
+    av1_subtract_plane(x, bsize, 0);
+    rd = estimate_yrd_for_sb(cpi, bsize, x, &rate_sum, &dist_sum,
+                             &tmp_skip_txfm_sb, &tmp_skip_sse_sb, INT64_MAX);
+    if (rd != INT64_MAX)
+      rd = RDCOST(x->rdmult, x->rddiv, rs2 + rate_mv + rate_sum, dist_sum);
+    best_rd_cur = rd;
+  }
+  return best_rd_cur;
+}
+#endif  // CONFIG_COMPOUND_SEGMENT || CONFIG_WEDGE
+#endif  // CONFIG_EXT_INTER
+
+typedef struct {
+#if CONFIG_MOTION_VAR
+  // Inter prediction buffers and respective strides
+  uint8_t *above_pred_buf[MAX_MB_PLANE];
+  int above_pred_stride[MAX_MB_PLANE];
+  uint8_t *left_pred_buf[MAX_MB_PLANE];
+  int left_pred_stride[MAX_MB_PLANE];
+#endif  // CONFIG_MOTION_VAR
+  int_mv *single_newmv;
+#if CONFIG_EXT_INTER
+  // Pointer to array of motion vectors to use for each ref and their rates
+  // Should point to first of 2 arrays in 2D array
+  int *single_newmv_rate;
+  // Pointers costs of compound inter-intra and inter-inter predictions
+  int *compmode_interintra_cost;
+  int *compmode_interinter_cost;
+  // Pointer to array of predicted rate-distortion
+  // Should point to first of 2 arrays in 2D array
+  int64_t (*modelled_rd)[TOTAL_REFS_PER_FRAME];
+#endif  // CONFIG_EXT_INTER
+  InterpFilter single_filter[MB_MODE_COUNT][TOTAL_REFS_PER_FRAME];
+} HandleInterModeArgs;
+
+static int64_t handle_newmv(const AV1_COMP *const cpi, MACROBLOCK *const x,
+                            const BLOCK_SIZE bsize,
+                            int_mv (*const mode_mv)[TOTAL_REFS_PER_FRAME],
+                            const int mi_row, const int mi_col,
+                            int *const rate_mv, int_mv *const single_newmv,
+                            HandleInterModeArgs *const args) {
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  const int is_comp_pred = has_second_ref(mbmi);
+  const PREDICTION_MODE this_mode = mbmi->mode;
+#if CONFIG_EXT_INTER
+  const int is_comp_interintra_pred = (mbmi->ref_frame[1] == INTRA_FRAME);
+#endif  // CONFIG_EXT_INTER
+  int_mv *const frame_mv = mode_mv[this_mode];
+  const int refs[2] = { mbmi->ref_frame[0],
+                        mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1] };
+  int i;
+
+  (void)args;
+
+  if (is_comp_pred) {
+#if CONFIG_EXT_INTER
+    for (i = 0; i < 2; ++i) {
+      single_newmv[refs[i]].as_int = args->single_newmv[refs[i]].as_int;
+    }
+
+    if (this_mode == NEW_NEWMV) {
+      frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
+      frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
+
+      if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
+        joint_motion_search(cpi, x, bsize, frame_mv, mi_row, mi_col, NULL,
+                            rate_mv, 0);
+      } else {
+        *rate_mv = 0;
+        for (i = 0; i < 2; ++i) {
+#if CONFIG_REF_MV
+          av1_set_mvcost(x, refs[i], i, mbmi->ref_mv_idx);
+#endif  // CONFIG_REF_MV
+          *rate_mv += av1_mv_bit_cost(
+              &frame_mv[refs[i]].as_mv, &mbmi_ext->ref_mvs[refs[i]][0].as_mv,
+              x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+        }
+      }
+    } else if (this_mode == NEAREST_NEWMV || this_mode == NEAR_NEWMV) {
+      frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
+#if CONFIG_REF_MV
+      av1_set_mvcost(x, refs[1], 1, mbmi->ref_mv_idx);
+#endif  // CONFIG_REF_MV
+      *rate_mv = av1_mv_bit_cost(&frame_mv[refs[1]].as_mv,
+                                 &mbmi_ext->ref_mvs[refs[1]][0].as_mv,
+                                 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+    } else {
+      assert(this_mode == NEW_NEARESTMV || this_mode == NEW_NEARMV);
+      frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
+#if CONFIG_REF_MV
+      av1_set_mvcost(x, refs[0], 0, mbmi->ref_mv_idx);
+#endif  // CONFIG_REF_MV
+      *rate_mv = av1_mv_bit_cost(&frame_mv[refs[0]].as_mv,
+                                 &mbmi_ext->ref_mvs[refs[0]][0].as_mv,
+                                 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+    }
+#else
+    // Initialize mv using single prediction mode result.
+    frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
+    frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
+
+    if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
+      joint_motion_search(cpi, x, bsize, frame_mv, mi_row, mi_col, rate_mv, 0);
+    } else {
+      *rate_mv = 0;
+      for (i = 0; i < 2; ++i) {
+#if CONFIG_REF_MV
+        av1_set_mvcost(x, refs[i], i, mbmi->ref_mv_idx);
+#endif  // CONFIG_REF_MV
+        *rate_mv += av1_mv_bit_cost(&frame_mv[refs[i]].as_mv,
+                                    &mbmi_ext->ref_mvs[refs[i]][0].as_mv,
+                                    x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
+      }
+    }
+#endif  // CONFIG_EXT_INTER
+  } else {
+#if CONFIG_EXT_INTER
+    if (is_comp_interintra_pred) {
+      x->best_mv = args->single_newmv[refs[0]];
+      *rate_mv = args->single_newmv_rate[refs[0]];
+    } else {
+      single_motion_search(cpi, x, bsize, mi_row, mi_col, 0, rate_mv);
+      args->single_newmv[refs[0]] = x->best_mv;
+      args->single_newmv_rate[refs[0]] = *rate_mv;
+    }
+#else
+    single_motion_search(cpi, x, bsize, mi_row, mi_col, rate_mv);
+    single_newmv[refs[0]] = x->best_mv;
+#endif  // CONFIG_EXT_INTER
+
+    if (x->best_mv.as_int == INVALID_MV) return INT64_MAX;
+
+    frame_mv[refs[0]] = x->best_mv;
+    xd->mi[0]->bmi[0].as_mv[0] = x->best_mv;
+
+    // Estimate the rate implications of a new mv but discount this
+    // under certain circumstances where we want to help initiate a weak
+    // motion field, where the distortion gain for a single block may not
+    // be enough to overcome the cost of a new mv.
+    if (discount_newmv_test(cpi, this_mode, x->best_mv, mode_mv, refs[0])) {
+      *rate_mv = AOMMAX(*rate_mv / NEW_MV_DISCOUNT_FACTOR, 1);
+    }
+  }
+
+  return 0;
+}
+
+int64_t interpolation_filter_search(
+    MACROBLOCK *const x, const AV1_COMP *const cpi, BLOCK_SIZE bsize,
+    int mi_row, int mi_col, const BUFFER_SET *const tmp_dst,
+    BUFFER_SET *const orig_dst,
+    InterpFilter (*const single_filter)[TOTAL_REFS_PER_FRAME],
+    int64_t *const rd, int *const switchable_rate, int *const skip_txfm_sb,
+    int64_t *const skip_sse_sb) {
+  const AV1_COMMON *cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  int i;
+  int tmp_rate;
+  int64_t tmp_dist;
+
+  (void)single_filter;
+
+  InterpFilter assign_filter = SWITCHABLE;
+
+  if (cm->interp_filter == SWITCHABLE) {
+#if !CONFIG_DUAL_FILTER
+    assign_filter = av1_is_interp_needed(xd)
+                        ? predict_interp_filter(cpi, x, bsize, mi_row, mi_col,
+                                                single_filter)
+                        : cm->interp_filter;
+#endif  // !CONFIG_DUAL_FILTER
+  } else {
+    assign_filter = cm->interp_filter;
+  }
+
+  set_default_interp_filters(mbmi, assign_filter);
+
+  *switchable_rate = av1_get_switchable_rate(cpi, xd);
+  av1_build_inter_predictors_sb(xd, mi_row, mi_col, orig_dst, bsize);
+  model_rd_for_sb(cpi, bsize, x, xd, 0, MAX_MB_PLANE - 1, &tmp_rate, &tmp_dist,
+                  skip_txfm_sb, skip_sse_sb);
+  *rd = RDCOST(x->rdmult, x->rddiv, *switchable_rate + tmp_rate, tmp_dist);
+
+  if (assign_filter == SWITCHABLE) {
+    // do interp_filter search
+    if (av1_is_interp_needed(xd) && av1_is_interp_search_needed(xd)) {
+#if CONFIG_DUAL_FILTER
+      const int filter_set_size = DUAL_FILTER_SET_SIZE;
+#else
+      const int filter_set_size = SWITCHABLE_FILTERS;
+#endif  // CONFIG_DUAL_FILTER
+      int best_in_temp = 0;
+#if CONFIG_DUAL_FILTER
+      InterpFilter best_filter[4];
+      av1_copy(best_filter, mbmi->interp_filter);
+#else
+      InterpFilter best_filter = mbmi->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+      restore_dst_buf(xd, *tmp_dst);
+      // EIGHTTAP_REGULAR mode is calculated beforehand
+      for (i = 1; i < filter_set_size; ++i) {
+        int tmp_skip_sb = 0;
+        int64_t tmp_skip_sse = INT64_MAX;
+        int tmp_rs;
+        int64_t tmp_rd;
+#if CONFIG_DUAL_FILTER
+        mbmi->interp_filter[0] = filter_sets[i][0];
+        mbmi->interp_filter[1] = filter_sets[i][1];
+        mbmi->interp_filter[2] = filter_sets[i][0];
+        mbmi->interp_filter[3] = filter_sets[i][1];
+#else
+        mbmi->interp_filter = (InterpFilter)i;
+#endif  // CONFIG_DUAL_FILTER
+        tmp_rs = av1_get_switchable_rate(cpi, xd);
+        av1_build_inter_predictors_sb(xd, mi_row, mi_col, orig_dst, bsize);
+        model_rd_for_sb(cpi, bsize, x, xd, 0, MAX_MB_PLANE - 1, &tmp_rate,
+                        &tmp_dist, &tmp_skip_sb, &tmp_skip_sse);
+        tmp_rd = RDCOST(x->rdmult, x->rddiv, tmp_rs + tmp_rate, tmp_dist);
+
+        if (tmp_rd < *rd) {
+          *rd = tmp_rd;
+          *switchable_rate = av1_get_switchable_rate(cpi, xd);
+#if CONFIG_DUAL_FILTER
+          av1_copy(best_filter, mbmi->interp_filter);
+#else
+          best_filter = mbmi->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+          *skip_txfm_sb = tmp_skip_sb;
+          *skip_sse_sb = tmp_skip_sse;
+          best_in_temp = !best_in_temp;
+          if (best_in_temp) {
+            restore_dst_buf(xd, *orig_dst);
+          } else {
+            restore_dst_buf(xd, *tmp_dst);
+          }
+        }
+      }
+      if (best_in_temp) {
+        restore_dst_buf(xd, *tmp_dst);
+      } else {
+        restore_dst_buf(xd, *orig_dst);
+      }
+#if CONFIG_DUAL_FILTER
+      av1_copy(mbmi->interp_filter, best_filter);
+#else
+      mbmi->interp_filter = best_filter;
+#endif  // CONFIG_DUAL_FILTER
+    } else {
+#if CONFIG_DUAL_FILTER
+      for (i = 0; i < 4; ++i)
+        assert(mbmi->interp_filter[i] == EIGHTTAP_REGULAR);
+#else
+      assert(mbmi->interp_filter == EIGHTTAP_REGULAR);
+#endif  // CONFIG_DUAL_FILTER
+    }
+  }
+
+  return 0;
+}
+
+// TODO(afergs): Refactor the MBMI references in here - there's four
+// TODO(afergs): Refactor optional args - add them to a struct or remove
+static int64_t motion_mode_rd(
+    const AV1_COMP *const cpi, MACROBLOCK *const x, BLOCK_SIZE bsize,
+    RD_STATS *rd_stats, RD_STATS *rd_stats_y, RD_STATS *rd_stats_uv,
+    int *disable_skip, int_mv (*mode_mv)[TOTAL_REFS_PER_FRAME], int mi_row,
+    int mi_col, HandleInterModeArgs *const args, const int64_t ref_best_rd,
+    const int *refs, int rate_mv,
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+#if CONFIG_EXT_INTER
+    int rate2_bmc_nocoeff, MB_MODE_INFO *best_bmc_mbmi,
+#if CONFIG_MOTION_VAR
+    int rate_mv_bmc,
+#endif  // CONFIG_MOTION_VAR
+#endif  // CONFIG_EXT_INTER
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    int rs, int *skip_txfm_sb, int64_t *skip_sse_sb, BUFFER_SET *orig_dst) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  MB_MODE_INFO *mbmi = &mi->mbmi;
+  const int is_comp_pred = has_second_ref(mbmi);
+  const PREDICTION_MODE this_mode = mbmi->mode;
+
+  (void)mode_mv;
+  (void)mi_row;
+  (void)mi_col;
+  (void)args;
+  (void)refs;
+  (void)rate_mv;
+  (void)is_comp_pred;
+  (void)this_mode;
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  MOTION_MODE motion_mode, last_motion_mode_allowed;
+  int rate2_nocoeff = 0, best_xskip, best_disable_skip = 0;
+  RD_STATS best_rd_stats, best_rd_stats_y, best_rd_stats_uv;
+  MB_MODE_INFO base_mbmi, best_mbmi;
+#if CONFIG_VAR_TX
+  uint8_t best_blk_skip[MAX_MB_PLANE][MAX_MIB_SIZE * MAX_MIB_SIZE * 4];
+#endif  // CONFIG_VAR_TX
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_WARPED_MOTION
+  int pts[SAMPLES_ARRAY_SIZE], pts_inref[SAMPLES_ARRAY_SIZE];
+#endif  // CONFIG_WARPED_MOTION
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  av1_invalid_rd_stats(&best_rd_stats);
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+  if (cm->interp_filter == SWITCHABLE) rd_stats->rate += rs;
+#if CONFIG_WARPED_MOTION
+  aom_clear_system_state();
+  mbmi->num_proj_ref[0] = findSamples(cm, xd, mi_row, mi_col, pts, pts_inref);
+#if CONFIG_EXT_INTER
+  best_bmc_mbmi->num_proj_ref[0] = mbmi->num_proj_ref[0];
+#endif  // CONFIG_EXT_INTER
+#endif  // CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  rate2_nocoeff = rd_stats->rate;
+  last_motion_mode_allowed = motion_mode_allowed(
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      0, xd->global_motion,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+      mi);
+  base_mbmi = *mbmi;
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  int64_t best_rd = INT64_MAX;
+  for (motion_mode = SIMPLE_TRANSLATION;
+       motion_mode <= last_motion_mode_allowed; motion_mode++) {
+    int64_t tmp_rd = INT64_MAX;
+    int tmp_rate;
+    int64_t tmp_dist;
+#if CONFIG_EXT_INTER
+    int tmp_rate2 =
+        motion_mode != SIMPLE_TRANSLATION ? rate2_bmc_nocoeff : rate2_nocoeff;
+#else
+    int tmp_rate2 = rate2_nocoeff;
+#endif  // CONFIG_EXT_INTER
+
+    *mbmi = base_mbmi;
+    mbmi->motion_mode = motion_mode;
+#if CONFIG_MOTION_VAR
+    if (mbmi->motion_mode == OBMC_CAUSAL) {
+#if CONFIG_EXT_INTER
+      *mbmi = *best_bmc_mbmi;
+      mbmi->motion_mode = OBMC_CAUSAL;
+#endif  // CONFIG_EXT_INTER
+      if (!is_comp_pred && have_newmv_in_inter_mode(this_mode)) {
+        int tmp_rate_mv = 0;
+
+        single_motion_search(cpi, x, bsize, mi_row, mi_col,
+#if CONFIG_EXT_INTER
+                             0,
+#endif  // CONFIG_EXT_INTER
+                             &tmp_rate_mv);
+        mbmi->mv[0].as_int = x->best_mv.as_int;
+        if (discount_newmv_test(cpi, this_mode, mbmi->mv[0], mode_mv,
+                                refs[0])) {
+          tmp_rate_mv = AOMMAX((tmp_rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
+        }
+#if CONFIG_EXT_INTER
+        tmp_rate2 = rate2_bmc_nocoeff - rate_mv_bmc + tmp_rate_mv;
+#else
+        tmp_rate2 = rate2_nocoeff - rate_mv + tmp_rate_mv;
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_DUAL_FILTER
+        if (!has_subpel_mv_component(xd->mi[0], xd, 0))
+          mbmi->interp_filter[0] = EIGHTTAP_REGULAR;
+        if (!has_subpel_mv_component(xd->mi[0], xd, 1))
+          mbmi->interp_filter[1] = EIGHTTAP_REGULAR;
+#endif  // CONFIG_DUAL_FILTER
+        av1_build_inter_predictors_sb(xd, mi_row, mi_col, orig_dst, bsize);
+#if CONFIG_EXT_INTER
+      } else {
+        av1_build_inter_predictors_sb(xd, mi_row, mi_col, orig_dst, bsize);
+#endif  // CONFIG_EXT_INTER
+      }
+      av1_build_obmc_inter_prediction(
+          cm, xd, mi_row, mi_col, args->above_pred_buf, args->above_pred_stride,
+          args->left_pred_buf, args->left_pred_stride);
+      model_rd_for_sb(cpi, bsize, x, xd, 0, MAX_MB_PLANE - 1, &tmp_rate,
+                      &tmp_dist, skip_txfm_sb, skip_sse_sb);
+    }
+#endif  // CONFIG_MOTION_VAR
+
+#if CONFIG_WARPED_MOTION
+    if (mbmi->motion_mode == WARPED_CAUSAL) {
+#if CONFIG_EXT_INTER
+      *mbmi = *best_bmc_mbmi;
+      mbmi->motion_mode = WARPED_CAUSAL;
+#endif  // CONFIG_EXT_INTER
+      mbmi->wm_params[0].wmtype = DEFAULT_WMTYPE;
+#if CONFIG_DUAL_FILTER
+      for (int dir = 0; dir < 4; ++dir)
+        mbmi->interp_filter[dir] = cm->interp_filter == SWITCHABLE
+                                       ? EIGHTTAP_REGULAR
+                                       : cm->interp_filter;
+#else
+      mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP_REGULAR
+                                                            : cm->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+
+      if (find_projection(mbmi->num_proj_ref[0], pts, pts_inref, bsize,
+                          mbmi->mv[0].as_mv.row, mbmi->mv[0].as_mv.col,
+                          &mbmi->wm_params[0], mi_row, mi_col) == 0) {
+        av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL, bsize);
+        model_rd_for_sb(cpi, bsize, x, xd, 0, MAX_MB_PLANE - 1, &tmp_rate,
+                        &tmp_dist, skip_txfm_sb, skip_sse_sb);
+      } else {
+        continue;
+      }
+    }
+#endif  // CONFIG_WARPED_MOTION
+    x->skip = 0;
+
+    rd_stats->dist = 0;
+    rd_stats->sse = 0;
+    rd_stats->skip = 1;
+    rd_stats->rate = tmp_rate2;
+    if (last_motion_mode_allowed > SIMPLE_TRANSLATION) {
+#if CONFIG_WARPED_MOTION && CONFIG_MOTION_VAR
+      if (last_motion_mode_allowed == WARPED_CAUSAL)
+#endif  // CONFIG_WARPED_MOTION && CONFIG_MOTION_VAR
+        rd_stats->rate += cpi->motion_mode_cost[bsize][mbmi->motion_mode];
+#if CONFIG_WARPED_MOTION && CONFIG_MOTION_VAR
+      else
+        rd_stats->rate += cpi->motion_mode_cost1[bsize][mbmi->motion_mode];
+#endif  // CONFIG_WARPED_MOTION && CONFIG_MOTION_VAR
+    }
+#if CONFIG_WARPED_MOTION
+    if (mbmi->motion_mode == WARPED_CAUSAL) {
+      rd_stats->rate -= rs;
+    }
+#endif  // CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    if (!*skip_txfm_sb) {
+      int64_t rdcosty = INT64_MAX;
+      int is_cost_valid_uv = 0;
+
+      // cost and distortion
+      av1_subtract_plane(x, bsize, 0);
+#if CONFIG_VAR_TX
+      if (cm->tx_mode == TX_MODE_SELECT && !xd->lossless[mbmi->segment_id]) {
+        select_tx_type_yrd(cpi, x, rd_stats_y, bsize, ref_best_rd);
+      } else {
+        int idx, idy;
+        super_block_yrd(cpi, x, rd_stats_y, bsize, ref_best_rd);
+        for (idy = 0; idy < xd->n8_h; ++idy)
+          for (idx = 0; idx < xd->n8_w; ++idx)
+            mbmi->inter_tx_size[idy][idx] = mbmi->tx_size;
+        memset(x->blk_skip[0], rd_stats_y->skip,
+               sizeof(uint8_t) * xd->n8_h * xd->n8_w * 4);
+      }
+#else
+    /* clang-format off */
+      super_block_yrd(cpi, x, rd_stats_y, bsize, ref_best_rd);
+/* clang-format on */
+#endif  // CONFIG_VAR_TX
+
+      if (rd_stats_y->rate == INT_MAX) {
+        av1_invalid_rd_stats(rd_stats);
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+        if (mbmi->motion_mode != SIMPLE_TRANSLATION) {
+          continue;
+        } else {
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+          restore_dst_buf(xd, *orig_dst);
+          return INT64_MAX;
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+        }
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      }
+
+      av1_merge_rd_stats(rd_stats, rd_stats_y);
+
+      rdcosty = RDCOST(x->rdmult, x->rddiv, rd_stats->rate, rd_stats->dist);
+      rdcosty = AOMMIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, rd_stats->sse));
+/* clang-format off */
+#if CONFIG_VAR_TX
+      is_cost_valid_uv =
+          inter_block_uvrd(cpi, x, rd_stats_uv, bsize, ref_best_rd - rdcosty);
+#else
+      is_cost_valid_uv =
+          super_block_uvrd(cpi, x, rd_stats_uv, bsize, ref_best_rd - rdcosty);
+#endif  // CONFIG_VAR_TX
+      if (!is_cost_valid_uv) {
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+        continue;
+#else
+        restore_dst_buf(xd, *orig_dst);
+        return INT64_MAX;
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      }
+      /* clang-format on */
+      av1_merge_rd_stats(rd_stats, rd_stats_uv);
+#if CONFIG_RD_DEBUG
+      // record transform block coefficient cost
+      // TODO(angiebird): So far rd_debug tool only detects discrepancy of
+      // coefficient cost. Therefore, it is fine to copy rd_stats into mbmi
+      // here because we already collect the coefficient cost. Move this part to
+      // other place when we need to compare non-coefficient cost.
+      mbmi->rd_stats = *rd_stats;
+#endif  // CONFIG_RD_DEBUG
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      if (rd_stats->skip) {
+        rd_stats->rate -= rd_stats_uv->rate + rd_stats_y->rate;
+        rd_stats_y->rate = 0;
+        rd_stats_uv->rate = 0;
+        rd_stats->rate += av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+        mbmi->skip = 0;
+        // here mbmi->skip temporarily plays a role as what this_skip2 does
+      } else if (!xd->lossless[mbmi->segment_id] &&
+                 (RDCOST(x->rdmult, x->rddiv,
+                         rd_stats_y->rate + rd_stats_uv->rate +
+                             av1_cost_bit(av1_get_skip_prob(cm, xd), 0),
+                         rd_stats->dist) >=
+                  RDCOST(x->rdmult, x->rddiv,
+                         av1_cost_bit(av1_get_skip_prob(cm, xd), 1),
+                         rd_stats->sse))) {
+        rd_stats->rate -= rd_stats_uv->rate + rd_stats_y->rate;
+        rd_stats->rate += av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+        rd_stats->dist = rd_stats->sse;
+        rd_stats_y->rate = 0;
+        rd_stats_uv->rate = 0;
+        mbmi->skip = 1;
+      } else {
+        rd_stats->rate += av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+        mbmi->skip = 0;
+      }
+      *disable_skip = 0;
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    } else {
+      x->skip = 1;
+      *disable_skip = 1;
+      mbmi->tx_size = tx_size_from_tx_mode(bsize, cm->tx_mode, 1);
+
+// The cost of skip bit needs to be added.
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      mbmi->skip = 0;
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      rd_stats->rate += av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+
+      rd_stats->dist = *skip_sse_sb;
+      rd_stats->sse = *skip_sse_sb;
+      rd_stats_y->rate = 0;
+      rd_stats_uv->rate = 0;
+      rd_stats->skip = 1;
+    }
+
+#if CONFIG_GLOBAL_MOTION
+    if (this_mode == ZEROMV
+#if CONFIG_EXT_INTER
+        || this_mode == ZERO_ZEROMV
+#endif  // CONFIG_EXT_INTER
+        ) {
+      if (is_nontrans_global_motion(xd)) {
+        rd_stats->rate -= rs;
+#if CONFIG_DUAL_FILTER
+        mbmi->interp_filter[0] = cm->interp_filter == SWITCHABLE
+                                     ? EIGHTTAP_REGULAR
+                                     : cm->interp_filter;
+        mbmi->interp_filter[1] = cm->interp_filter == SWITCHABLE
+                                     ? EIGHTTAP_REGULAR
+                                     : cm->interp_filter;
+#else
+        mbmi->interp_filter = cm->interp_filter == SWITCHABLE
+                                  ? EIGHTTAP_REGULAR
+                                  : cm->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+      }
+    }
+#endif  // CONFIG_GLOBAL_MOTION
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    tmp_rd = RDCOST(x->rdmult, x->rddiv, rd_stats->rate, rd_stats->dist);
+    if (mbmi->motion_mode == SIMPLE_TRANSLATION || (tmp_rd < best_rd)) {
+      best_mbmi = *mbmi;
+      best_rd = tmp_rd;
+      best_rd_stats = *rd_stats;
+      best_rd_stats_y = *rd_stats_y;
+      best_rd_stats_uv = *rd_stats_uv;
+#if CONFIG_VAR_TX
+      for (int i = 0; i < MAX_MB_PLANE; ++i)
+        memcpy(best_blk_skip[i], x->blk_skip[i],
+               sizeof(uint8_t) * xd->n8_h * xd->n8_w * 4);
+#endif  // CONFIG_VAR_TX
+      best_xskip = x->skip;
+      best_disable_skip = *disable_skip;
+    }
+  }
+
+  if (best_rd == INT64_MAX) {
+    av1_invalid_rd_stats(rd_stats);
+    restore_dst_buf(xd, *orig_dst);
+    return INT64_MAX;
+  }
+  *mbmi = best_mbmi;
+  *rd_stats = best_rd_stats;
+  *rd_stats_y = best_rd_stats_y;
+  *rd_stats_uv = best_rd_stats_uv;
+#if CONFIG_VAR_TX
+  for (int i = 0; i < MAX_MB_PLANE; ++i)
+    memcpy(x->blk_skip[i], best_blk_skip[i],
+           sizeof(uint8_t) * xd->n8_h * xd->n8_w * 4);
+#endif  // CONFIG_VAR_TX
+  x->skip = best_xskip;
+  *disable_skip = best_disable_skip;
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+  restore_dst_buf(xd, *orig_dst);
+  return 0;
+}
+
+static int64_t handle_inter_mode(
+    const AV1_COMP *const cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
+    RD_STATS *rd_stats, RD_STATS *rd_stats_y, RD_STATS *rd_stats_uv,
+    int *disable_skip, int_mv (*mode_mv)[TOTAL_REFS_PER_FRAME], int mi_row,
+    int mi_col, HandleInterModeArgs *args, const int64_t ref_best_rd) {
+  const AV1_COMMON *cm = &cpi->common;
+  (void)cm;
+  MACROBLOCKD *xd = &x->e_mbd;
+  MODE_INFO *mi = xd->mi[0];
+  MB_MODE_INFO *mbmi = &mi->mbmi;
+  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  const int is_comp_pred = has_second_ref(mbmi);
+  const int this_mode = mbmi->mode;
+  int_mv *frame_mv = mode_mv[this_mode];
+  int i;
+  int refs[2] = { mbmi->ref_frame[0],
+                  (mbmi->ref_frame[1] < 0 ? 0 : mbmi->ref_frame[1]) };
+  int_mv cur_mv[2];
+  int rate_mv = 0;
+#if CONFIG_EXT_INTER
+  int pred_exists = 1;
+  const int bw = block_size_wide[bsize];
+  int_mv single_newmv[TOTAL_REFS_PER_FRAME];
+#if CONFIG_INTERINTRA
+  const unsigned int *const interintra_mode_cost =
+      cpi->interintra_mode_cost[size_group_lookup[bsize]];
+#endif  // CONFIG_INTERINTRA
+  const int is_comp_interintra_pred = (mbmi->ref_frame[1] == INTRA_FRAME);
+#if CONFIG_REF_MV
+  uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+#endif  // CONFIG_REF_MV
+#else
+  int_mv *const single_newmv = args->single_newmv;
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf_[2 * MAX_MB_PLANE * MAX_SB_SQUARE]);
+#else
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf_[MAX_MB_PLANE * MAX_SB_SQUARE]);
+#endif  // CONFIG_HIGHBITDEPTH
+  uint8_t *tmp_buf;
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+#if CONFIG_EXT_INTER
+  int rate2_bmc_nocoeff;
+  MB_MODE_INFO best_bmc_mbmi;
+#if CONFIG_MOTION_VAR
+  int rate_mv_bmc;
+#endif  // CONFIG_MOTION_VAR
+#endif  // CONFIG_EXT_INTER
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  int64_t rd = INT64_MAX;
+  BUFFER_SET orig_dst, tmp_dst;
+  int rs = 0;
+
+  int skip_txfm_sb = 0;
+  int64_t skip_sse_sb = INT64_MAX;
+  int16_t mode_ctx;
+
+#if CONFIG_EXT_INTER
+  *args->compmode_interintra_cost = 0;
+  mbmi->use_wedge_interintra = 0;
+  *args->compmode_interinter_cost = 0;
+  mbmi->interinter_compound_type = COMPOUND_AVERAGE;
+
+  // is_comp_interintra_pred implies !is_comp_pred
+  assert(!is_comp_interintra_pred || (!is_comp_pred));
+  // is_comp_interintra_pred implies is_interintra_allowed(mbmi->sb_type)
+  assert(!is_comp_interintra_pred || is_interintra_allowed(mbmi));
+#endif  // CONFIG_EXT_INTER
+
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+  if (is_comp_pred)
+    mode_ctx = mbmi_ext->compound_mode_context[refs[0]];
+  else
+#endif  // CONFIG_EXT_INTER
+    mode_ctx = av1_mode_context_analyzer(mbmi_ext->mode_context,
+                                         mbmi->ref_frame, bsize, -1);
+#else   // CONFIG_REF_MV
+  mode_ctx = mbmi_ext->mode_context[refs[0]];
+#endif  // CONFIG_REF_MV
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+    tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf_);
+  else
+#endif  // CONFIG_HIGHBITDEPTH
+    tmp_buf = tmp_buf_;
+  // Make sure that we didn't leave the plane destination buffers set
+  // to tmp_buf at the end of the last iteration
+  assert(xd->plane[0].dst.buf != tmp_buf);
+
+#if CONFIG_WARPED_MOTION
+  mbmi->num_proj_ref[0] = 0;
+  mbmi->num_proj_ref[1] = 0;
+#endif  // CONFIG_WARPED_MOTION
+
+  if (is_comp_pred) {
+    if (frame_mv[refs[0]].as_int == INVALID_MV ||
+        frame_mv[refs[1]].as_int == INVALID_MV)
+      return INT64_MAX;
+  }
+
+  mbmi->motion_mode = SIMPLE_TRANSLATION;
+  if (have_newmv_in_inter_mode(this_mode)) {
+    const int64_t ret_val = handle_newmv(cpi, x, bsize, mode_mv, mi_row, mi_col,
+                                         &rate_mv, single_newmv, args);
+    if (ret_val != 0)
+      return ret_val;
+    else
+      rd_stats->rate += rate_mv;
+  }
+  for (i = 0; i < is_comp_pred + 1; ++i) {
+    cur_mv[i] = frame_mv[refs[i]];
+    // Clip "next_nearest" so that it does not extend to far out of image
+    if (this_mode != NEWMV) clamp_mv2(&cur_mv[i].as_mv, xd);
+    if (mv_check_bounds(&x->mv_limits, &cur_mv[i].as_mv)) return INT64_MAX;
+    mbmi->mv[i].as_int = cur_mv[i].as_int;
+  }
+
+#if CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+  if (this_mode == NEAREST_NEARESTMV)
+#else
+  if (this_mode == NEARESTMV && is_comp_pred)
+#endif  // CONFIG_EXT_INTER
+  {
+#if !CONFIG_EXT_INTER
+    uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+#endif  // !CONFIG_EXT_INTER
+    if (mbmi_ext->ref_mv_count[ref_frame_type] > 0) {
+      cur_mv[0] = mbmi_ext->ref_mv_stack[ref_frame_type][0].this_mv;
+      cur_mv[1] = mbmi_ext->ref_mv_stack[ref_frame_type][0].comp_mv;
+
+      for (i = 0; i < 2; ++i) {
+        clamp_mv2(&cur_mv[i].as_mv, xd);
+        if (mv_check_bounds(&x->mv_limits, &cur_mv[i].as_mv)) return INT64_MAX;
+        mbmi->mv[i].as_int = cur_mv[i].as_int;
+      }
+    }
+  }
+
+#if CONFIG_EXT_INTER
+  if (mbmi_ext->ref_mv_count[ref_frame_type] > 0) {
+    if (this_mode == NEAREST_NEWMV || this_mode == NEAREST_NEARMV) {
+      cur_mv[0] = mbmi_ext->ref_mv_stack[ref_frame_type][0].this_mv;
+
+      lower_mv_precision(&cur_mv[0].as_mv, cm->allow_high_precision_mv);
+      clamp_mv2(&cur_mv[0].as_mv, xd);
+      if (mv_check_bounds(&x->mv_limits, &cur_mv[0].as_mv)) return INT64_MAX;
+      mbmi->mv[0].as_int = cur_mv[0].as_int;
+    }
+
+    if (this_mode == NEW_NEARESTMV || this_mode == NEAR_NEARESTMV) {
+      cur_mv[1] = mbmi_ext->ref_mv_stack[ref_frame_type][0].comp_mv;
+
+      lower_mv_precision(&cur_mv[1].as_mv, cm->allow_high_precision_mv);
+      clamp_mv2(&cur_mv[1].as_mv, xd);
+      if (mv_check_bounds(&x->mv_limits, &cur_mv[1].as_mv)) return INT64_MAX;
+      mbmi->mv[1].as_int = cur_mv[1].as_int;
+    }
+  }
+
+  if (mbmi_ext->ref_mv_count[ref_frame_type] > 1) {
+    int ref_mv_idx = mbmi->ref_mv_idx + 1;
+    if (this_mode == NEAR_NEWMV || this_mode == NEAR_NEARESTMV ||
+        this_mode == NEAR_NEARMV) {
+      cur_mv[0] = mbmi_ext->ref_mv_stack[ref_frame_type][ref_mv_idx].this_mv;
+
+      lower_mv_precision(&cur_mv[0].as_mv, cm->allow_high_precision_mv);
+      clamp_mv2(&cur_mv[0].as_mv, xd);
+      if (mv_check_bounds(&x->mv_limits, &cur_mv[0].as_mv)) return INT64_MAX;
+      mbmi->mv[0].as_int = cur_mv[0].as_int;
+    }
+
+    if (this_mode == NEW_NEARMV || this_mode == NEAREST_NEARMV ||
+        this_mode == NEAR_NEARMV) {
+      cur_mv[1] = mbmi_ext->ref_mv_stack[ref_frame_type][ref_mv_idx].comp_mv;
+
+      lower_mv_precision(&cur_mv[1].as_mv, cm->allow_high_precision_mv);
+      clamp_mv2(&cur_mv[1].as_mv, xd);
+      if (mv_check_bounds(&x->mv_limits, &cur_mv[1].as_mv)) return INT64_MAX;
+      mbmi->mv[1].as_int = cur_mv[1].as_int;
+    }
+  }
+#else
+  if (this_mode == NEARMV && is_comp_pred) {
+    uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+    if (mbmi_ext->ref_mv_count[ref_frame_type] > 1) {
+      int ref_mv_idx = mbmi->ref_mv_idx + 1;
+      cur_mv[0] = mbmi_ext->ref_mv_stack[ref_frame_type][ref_mv_idx].this_mv;
+      cur_mv[1] = mbmi_ext->ref_mv_stack[ref_frame_type][ref_mv_idx].comp_mv;
+
+      for (i = 0; i < 2; ++i) {
+        clamp_mv2(&cur_mv[i].as_mv, xd);
+        if (mv_check_bounds(&x->mv_limits, &cur_mv[i].as_mv)) return INT64_MAX;
+        mbmi->mv[i].as_int = cur_mv[i].as_int;
+      }
+    }
+  }
+#endif  // CONFIG_EXT_INTER
+#endif  // CONFIG_REF_MV
+
+  // do first prediction into the destination buffer. Do the next
+  // prediction into a temporary buffer. Then keep track of which one
+  // of these currently holds the best predictor, and use the other
+  // one for future predictions. In the end, copy from tmp_buf to
+  // dst if necessary.
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    tmp_dst.plane[i] = tmp_buf + i * MAX_SB_SQUARE;
+    tmp_dst.stride[i] = MAX_SB_SIZE;
+  }
+  for (i = 0; i < MAX_MB_PLANE; i++) {
+    orig_dst.plane[i] = xd->plane[i].dst.buf;
+    orig_dst.stride[i] = xd->plane[i].dst.stride;
+  }
+
+  // We don't include the cost of the second reference here, because there
+  // are only three options: Last/Golden, ARF/Last or Golden/ARF, or in other
+  // words if you present them in that order, the second one is always known
+  // if the first is known.
+  //
+  // Under some circumstances we discount the cost of new mv mode to encourage
+  // initiation of a motion field.
+  if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]], mode_mv,
+                          refs[0])) {
+#if CONFIG_EXT_INTER
+    rd_stats->rate +=
+        AOMMIN(cost_mv_ref(cpi, this_mode, mode_ctx),
+               cost_mv_ref(cpi, is_comp_pred ? NEAREST_NEARESTMV : NEARESTMV,
+                           mode_ctx));
+#else
+    rd_stats->rate += AOMMIN(cost_mv_ref(cpi, this_mode, mode_ctx),
+                             cost_mv_ref(cpi, NEARESTMV, mode_ctx));
+#endif  // CONFIG_REF_MV && CONFIG_EXT_INTER
+  } else {
+    rd_stats->rate += cost_mv_ref(cpi, this_mode, mode_ctx);
+  }
+
+  if (RDCOST(x->rdmult, x->rddiv, rd_stats->rate, 0) > ref_best_rd &&
+#if CONFIG_EXT_INTER
+      mbmi->mode != NEARESTMV && mbmi->mode != NEAREST_NEARESTMV
+#else
+      mbmi->mode != NEARESTMV
+#endif  // CONFIG_EXT_INTER
+      )
+    return INT64_MAX;
+
+  int64_t ret_val = interpolation_filter_search(
+      x, cpi, bsize, mi_row, mi_col, &tmp_dst, &orig_dst, args->single_filter,
+      &rd, &rs, &skip_txfm_sb, &skip_sse_sb);
+  if (ret_val != 0) return ret_val;
+
+#if CONFIG_EXT_INTER
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+  best_bmc_mbmi = *mbmi;
+  rate2_bmc_nocoeff = rd_stats->rate;
+  if (cm->interp_filter == SWITCHABLE) rate2_bmc_nocoeff += rs;
+#if CONFIG_MOTION_VAR
+  rate_mv_bmc = rate_mv;
+#endif  // CONFIG_MOTION_VAR
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+
+  if (is_comp_pred) {
+    int rate_sum, rs2;
+    int64_t dist_sum;
+    int64_t best_rd_compound = INT64_MAX, best_rd_cur = INT64_MAX;
+    INTERINTER_COMPOUND_DATA best_compound_data;
+    int_mv best_mv[2];
+    int best_tmp_rate_mv = rate_mv;
+    int tmp_skip_txfm_sb;
+    int64_t tmp_skip_sse_sb;
+    int compound_type_cost[COMPOUND_TYPES];
+    uint8_t pred0[2 * MAX_SB_SQUARE];
+    uint8_t pred1[2 * MAX_SB_SQUARE];
+    uint8_t *preds0[1] = { pred0 };
+    uint8_t *preds1[1] = { pred1 };
+    int strides[1] = { bw };
+    int tmp_rate_mv;
+    int masked_compound_used = is_any_masked_compound_used(bsize);
+    COMPOUND_TYPE cur_type;
+
+    best_mv[0].as_int = cur_mv[0].as_int;
+    best_mv[1].as_int = cur_mv[1].as_int;
+    memset(&best_compound_data, 0, sizeof(best_compound_data));
+#if CONFIG_COMPOUND_SEGMENT
+    uint8_t tmp_mask_buf[2 * MAX_SB_SQUARE];
+    best_compound_data.seg_mask = tmp_mask_buf;
+#endif  // CONFIG_COMPOUND_SEGMENT
+    av1_cost_tokens(compound_type_cost, cm->fc->compound_type_prob[bsize],
+                    av1_compound_type_tree);
+
+    if (masked_compound_used) {
+      av1_cost_tokens(compound_type_cost, cm->fc->compound_type_prob[bsize],
+                      av1_compound_type_tree);
+      // get inter predictors to use for masked compound modes
+      av1_build_inter_predictors_for_planes_single_buf(
+          xd, bsize, 0, 0, mi_row, mi_col, 0, preds0, strides);
+      av1_build_inter_predictors_for_planes_single_buf(
+          xd, bsize, 0, 0, mi_row, mi_col, 1, preds1, strides);
+    }
+
+    for (cur_type = COMPOUND_AVERAGE; cur_type < COMPOUND_TYPES; cur_type++) {
+      if (!is_interinter_compound_used(cur_type, bsize)) break;
+      tmp_rate_mv = rate_mv;
+      best_rd_cur = INT64_MAX;
+      mbmi->interinter_compound_type = cur_type;
+      rs2 = av1_cost_literal(get_interinter_compound_type_bits(
+                bsize, mbmi->interinter_compound_type)) +
+            (masked_compound_used
+                 ? compound_type_cost[mbmi->interinter_compound_type]
+                 : 0);
+
+      switch (cur_type) {
+        case COMPOUND_AVERAGE:
+          av1_build_inter_predictors_sby(xd, mi_row, mi_col, &orig_dst, bsize);
+          av1_subtract_plane(x, bsize, 0);
+          rd = estimate_yrd_for_sb(cpi, bsize, x, &rate_sum, &dist_sum,
+                                   &tmp_skip_txfm_sb, &tmp_skip_sse_sb,
+                                   INT64_MAX);
+          if (rd != INT64_MAX)
+            best_rd_cur =
+                RDCOST(x->rdmult, x->rddiv, rs2 + rate_mv + rate_sum, dist_sum);
+          best_rd_compound = best_rd_cur;
+          break;
+#if CONFIG_WEDGE
+        case COMPOUND_WEDGE:
+          if (x->source_variance > cpi->sf.disable_wedge_search_var_thresh &&
+              best_rd_compound / 3 < ref_best_rd) {
+            best_rd_cur = build_and_cost_compound_type(
+                cpi, x, cur_mv, bsize, this_mode, rs2, rate_mv, &orig_dst,
+                &tmp_rate_mv, preds0, preds1, strides, mi_row, mi_col);
+          }
+          break;
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+        case COMPOUND_SEG:
+          if (x->source_variance > cpi->sf.disable_wedge_search_var_thresh &&
+              best_rd_compound / 3 < ref_best_rd) {
+            best_rd_cur = build_and_cost_compound_type(
+                cpi, x, cur_mv, bsize, this_mode, rs2, rate_mv, &orig_dst,
+                &tmp_rate_mv, preds0, preds1, strides, mi_row, mi_col);
+          }
+          break;
+#endif  // CONFIG_COMPOUND_SEGMENT
+        default: assert(0); return 0;
+      }
+
+      if (best_rd_cur < best_rd_compound) {
+        best_rd_compound = best_rd_cur;
+#if CONFIG_WEDGE
+        best_compound_data.wedge_index = mbmi->wedge_index;
+        best_compound_data.wedge_sign = mbmi->wedge_sign;
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+        best_compound_data.mask_type = mbmi->mask_type;
+        memcpy(best_compound_data.seg_mask, xd->seg_mask,
+               2 * MAX_SB_SQUARE * sizeof(uint8_t));
+#endif  // CONFIG_COMPOUND_SEGMENT
+        best_compound_data.interinter_compound_type =
+            mbmi->interinter_compound_type;
+        if (have_newmv_in_inter_mode(this_mode)) {
+          if (use_masked_motion_search(cur_type)) {
+            best_tmp_rate_mv = tmp_rate_mv;
+            best_mv[0].as_int = mbmi->mv[0].as_int;
+            best_mv[1].as_int = mbmi->mv[1].as_int;
+          } else {
+            best_mv[0].as_int = cur_mv[0].as_int;
+            best_mv[1].as_int = cur_mv[1].as_int;
+          }
+        }
+      }
+      // reset to original mvs for next iteration
+      mbmi->mv[0].as_int = cur_mv[0].as_int;
+      mbmi->mv[1].as_int = cur_mv[1].as_int;
+    }
+#if CONFIG_WEDGE
+    mbmi->wedge_index = best_compound_data.wedge_index;
+    mbmi->wedge_sign = best_compound_data.wedge_sign;
+#endif  // CONFIG_WEDGE
+#if CONFIG_COMPOUND_SEGMENT
+    mbmi->mask_type = best_compound_data.mask_type;
+    memcpy(xd->seg_mask, best_compound_data.seg_mask,
+           2 * MAX_SB_SQUARE * sizeof(uint8_t));
+#endif  // CONFIG_COMPOUND_SEGMENT
+    mbmi->interinter_compound_type =
+        best_compound_data.interinter_compound_type;
+    if (have_newmv_in_inter_mode(this_mode)) {
+      mbmi->mv[0].as_int = best_mv[0].as_int;
+      mbmi->mv[1].as_int = best_mv[1].as_int;
+      xd->mi[0]->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int;
+      xd->mi[0]->bmi[0].as_mv[1].as_int = mbmi->mv[1].as_int;
+      if (use_masked_motion_search(mbmi->interinter_compound_type)) {
+        rd_stats->rate += best_tmp_rate_mv - rate_mv;
+        rate_mv = best_tmp_rate_mv;
+      }
+    }
+
+    if (ref_best_rd < INT64_MAX && best_rd_compound / 3 > ref_best_rd) {
+      restore_dst_buf(xd, orig_dst);
+      return INT64_MAX;
+    }
+
+    pred_exists = 0;
+
+    *args->compmode_interinter_cost =
+        av1_cost_literal(get_interinter_compound_type_bits(
+            bsize, mbmi->interinter_compound_type)) +
+        (masked_compound_used
+             ? compound_type_cost[mbmi->interinter_compound_type]
+             : 0);
+  }
+
+#if CONFIG_INTERINTRA
+  if (is_comp_interintra_pred) {
+    INTERINTRA_MODE best_interintra_mode = II_DC_PRED;
+    int64_t best_interintra_rd = INT64_MAX;
+    int rmode, rate_sum;
+    int64_t dist_sum;
+    int j;
+    int tmp_rate_mv = 0;
+    int tmp_skip_txfm_sb;
+    int64_t tmp_skip_sse_sb;
+    DECLARE_ALIGNED(16, uint8_t, intrapred_[2 * MAX_SB_SQUARE]);
+    uint8_t *intrapred;
+
+#if CONFIG_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+      intrapred = CONVERT_TO_BYTEPTR(intrapred_);
+    else
+#endif  // CONFIG_HIGHBITDEPTH
+      intrapred = intrapred_;
+
+    mbmi->ref_frame[1] = NONE_FRAME;
+    for (j = 0; j < MAX_MB_PLANE; j++) {
+      xd->plane[j].dst.buf = tmp_buf + j * MAX_SB_SQUARE;
+      xd->plane[j].dst.stride = bw;
+    }
+    av1_build_inter_predictors_sby(xd, mi_row, mi_col, &orig_dst, bsize);
+    restore_dst_buf(xd, orig_dst);
+    mbmi->ref_frame[1] = INTRA_FRAME;
+    mbmi->use_wedge_interintra = 0;
+
+    for (j = 0; j < INTERINTRA_MODES; ++j) {
+      mbmi->interintra_mode = (INTERINTRA_MODE)j;
+      rmode = interintra_mode_cost[mbmi->interintra_mode];
+      av1_build_intra_predictors_for_interintra(xd, bsize, 0, &orig_dst,
+                                                intrapred, bw);
+      av1_combine_interintra(xd, bsize, 0, tmp_buf, bw, intrapred, bw);
+      model_rd_for_sb(cpi, bsize, x, xd, 0, 0, &rate_sum, &dist_sum,
+                      &tmp_skip_txfm_sb, &tmp_skip_sse_sb);
+      rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate_mv + rate_sum, dist_sum);
+      if (rd < best_interintra_rd) {
+        best_interintra_rd = rd;
+        best_interintra_mode = mbmi->interintra_mode;
+      }
+    }
+    mbmi->interintra_mode = best_interintra_mode;
+    rmode = interintra_mode_cost[mbmi->interintra_mode];
+    av1_build_intra_predictors_for_interintra(xd, bsize, 0, &orig_dst,
+                                              intrapred, bw);
+    av1_combine_interintra(xd, bsize, 0, tmp_buf, bw, intrapred, bw);
+    av1_subtract_plane(x, bsize, 0);
+    rd = estimate_yrd_for_sb(cpi, bsize, x, &rate_sum, &dist_sum,
+                             &tmp_skip_txfm_sb, &tmp_skip_sse_sb, INT64_MAX);
+    if (rd != INT64_MAX)
+      rd = RDCOST(x->rdmult, x->rddiv, rate_mv + rmode + rate_sum, dist_sum);
+    best_interintra_rd = rd;
+
+    if (ref_best_rd < INT64_MAX && best_interintra_rd > 2 * ref_best_rd) {
+      // Don't need to call restore_dst_buf here
+      return INT64_MAX;
+    }
+#if CONFIG_WEDGE
+    if (is_interintra_wedge_used(bsize)) {
+      int64_t best_interintra_rd_nowedge = INT64_MAX;
+      int64_t best_interintra_rd_wedge = INT64_MAX;
+      int_mv tmp_mv;
+      int rwedge = av1_cost_bit(cm->fc->wedge_interintra_prob[bsize], 0);
+      if (rd != INT64_MAX)
+        rd = RDCOST(x->rdmult, x->rddiv, rmode + rate_mv + rwedge + rate_sum,
+                    dist_sum);
+      best_interintra_rd_nowedge = rd;
+
+      // Disable wedge search if source variance is small
+      if (x->source_variance > cpi->sf.disable_wedge_search_var_thresh) {
+        mbmi->use_wedge_interintra = 1;
+
+        rwedge = av1_cost_literal(get_interintra_wedge_bits(bsize)) +
+                 av1_cost_bit(cm->fc->wedge_interintra_prob[bsize], 1);
+
+        best_interintra_rd_wedge =
+            pick_interintra_wedge(cpi, x, bsize, intrapred_, tmp_buf_);
+
+        best_interintra_rd_wedge +=
+            RDCOST(x->rdmult, x->rddiv, rmode + rate_mv + rwedge, 0);
+        // Refine motion vector.
+        if (have_newmv_in_inter_mode(this_mode)) {
+          // get negative of mask
+          const uint8_t *mask = av1_get_contiguous_soft_mask(
+              mbmi->interintra_wedge_index, 1, bsize);
+          do_masked_motion_search(cpi, x, mask, bw, bsize, mi_row, mi_col,
+                                  &tmp_mv, &tmp_rate_mv, 0);
+          mbmi->mv[0].as_int = tmp_mv.as_int;
+          av1_build_inter_predictors_sby(xd, mi_row, mi_col, &orig_dst, bsize);
+          model_rd_for_sb(cpi, bsize, x, xd, 0, 0, &rate_sum, &dist_sum,
+                          &tmp_skip_txfm_sb, &tmp_skip_sse_sb);
+          rd = RDCOST(x->rdmult, x->rddiv,
+                      rmode + tmp_rate_mv + rwedge + rate_sum, dist_sum);
+          if (rd < best_interintra_rd_wedge) {
+            best_interintra_rd_wedge = rd;
+          } else {
+            tmp_mv.as_int = cur_mv[0].as_int;
+            tmp_rate_mv = rate_mv;
+          }
+        } else {
+          tmp_mv.as_int = cur_mv[0].as_int;
+          tmp_rate_mv = rate_mv;
+          av1_combine_interintra(xd, bsize, 0, tmp_buf, bw, intrapred, bw);
+        }
+        // Evaluate closer to true rd
+        av1_subtract_plane(x, bsize, 0);
+        rd =
+            estimate_yrd_for_sb(cpi, bsize, x, &rate_sum, &dist_sum,
+                                &tmp_skip_txfm_sb, &tmp_skip_sse_sb, INT64_MAX);
+        if (rd != INT64_MAX)
+          rd = RDCOST(x->rdmult, x->rddiv,
+                      rmode + tmp_rate_mv + rwedge + rate_sum, dist_sum);
+        best_interintra_rd_wedge = rd;
+        if (best_interintra_rd_wedge < best_interintra_rd_nowedge) {
+          mbmi->use_wedge_interintra = 1;
+          best_interintra_rd = best_interintra_rd_wedge;
+          mbmi->mv[0].as_int = tmp_mv.as_int;
+          rd_stats->rate += tmp_rate_mv - rate_mv;
+          rate_mv = tmp_rate_mv;
+        } else {
+          mbmi->use_wedge_interintra = 0;
+          best_interintra_rd = best_interintra_rd_nowedge;
+          mbmi->mv[0].as_int = cur_mv[0].as_int;
+        }
+      } else {
+        mbmi->use_wedge_interintra = 0;
+        best_interintra_rd = best_interintra_rd_nowedge;
+      }
+    }
+#endif  // CONFIG_WEDGE
+
+    pred_exists = 0;
+    *args->compmode_interintra_cost =
+        av1_cost_bit(cm->fc->interintra_prob[size_group_lookup[bsize]], 1);
+    *args->compmode_interintra_cost +=
+        interintra_mode_cost[mbmi->interintra_mode];
+    if (is_interintra_wedge_used(bsize)) {
+      *args->compmode_interintra_cost += av1_cost_bit(
+          cm->fc->wedge_interintra_prob[bsize], mbmi->use_wedge_interintra);
+      if (mbmi->use_wedge_interintra) {
+        *args->compmode_interintra_cost +=
+            av1_cost_literal(get_interintra_wedge_bits(bsize));
+      }
+    }
+  } else if (is_interintra_allowed(mbmi)) {
+    *args->compmode_interintra_cost =
+        av1_cost_bit(cm->fc->interintra_prob[size_group_lookup[bsize]], 0);
+  }
+#endif  // CONFIG_INTERINTRA
+
+  if (pred_exists == 0) {
+    int tmp_rate;
+    int64_t tmp_dist;
+    av1_build_inter_predictors_sb(xd, mi_row, mi_col, &orig_dst, bsize);
+    model_rd_for_sb(cpi, bsize, x, xd, 0, MAX_MB_PLANE - 1, &tmp_rate,
+                    &tmp_dist, &skip_txfm_sb, &skip_sse_sb);
+    rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
+  }
+#endif  // CONFIG_EXT_INTER
+
+  if (!is_comp_pred)
+#if CONFIG_DUAL_FILTER
+    args->single_filter[this_mode][refs[0]] = mbmi->interp_filter[0];
+#else
+    args->single_filter[this_mode][refs[0]] = mbmi->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+
+#if CONFIG_EXT_INTER
+  if (args->modelled_rd != NULL) {
+    if (is_comp_pred) {
+      const int mode0 = compound_ref0_mode(this_mode);
+      const int mode1 = compound_ref1_mode(this_mode);
+      const int64_t mrd = AOMMIN(args->modelled_rd[mode0][refs[0]],
+                                 args->modelled_rd[mode1][refs[1]]);
+      if (rd / 4 * 3 > mrd && ref_best_rd < INT64_MAX) {
+        restore_dst_buf(xd, orig_dst);
+        return INT64_MAX;
+      }
+    } else if (!is_comp_interintra_pred) {
+      args->modelled_rd[this_mode][refs[0]] = rd;
+    }
+  }
+#endif  // CONFIG_EXT_INTER
+
+  if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
+    // if current pred_error modeled rd is substantially more than the best
+    // so far, do not bother doing full rd
+    if (rd / 2 > ref_best_rd) {
+      restore_dst_buf(xd, orig_dst);
+      return INT64_MAX;
+    }
+  }
+
+  ret_val = motion_mode_rd(cpi, x, bsize, rd_stats, rd_stats_y, rd_stats_uv,
+                           disable_skip, mode_mv, mi_row, mi_col, args,
+                           ref_best_rd, refs, rate_mv,
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+#if CONFIG_EXT_INTER
+                           rate2_bmc_nocoeff, &best_bmc_mbmi,
+#if CONFIG_MOTION_VAR
+                           rate_mv_bmc,
+#endif  // CONFIG_MOTION_VAR
+#endif  // CONFIG_EXT_INTER
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+                           rs, &skip_txfm_sb, &skip_sse_sb, &orig_dst);
+  if (ret_val != 0) return ret_val;
+
+  return 0;  // The rate-distortion cost will be re-calculated by caller.
+}
+
+#if CONFIG_INTRABC
+static int64_t rd_pick_intrabc_mode_sb(const AV1_COMP *cpi, MACROBLOCK *x,
+                                       RD_STATS *rd_cost, BLOCK_SIZE bsize,
+                                       int64_t best_rd) {
+  const AV1_COMMON *const cm = &cpi->common;
+  if (bsize < BLOCK_8X8 || !cm->allow_screen_content_tools) return INT64_MAX;
+
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const TileInfo *tile = &xd->tile;
+  MODE_INFO *const mi = xd->mi[0];
+  const int mi_row = -xd->mb_to_top_edge / (8 * MI_SIZE);
+  const int mi_col = -xd->mb_to_left_edge / (8 * MI_SIZE);
+  const int w = block_size_wide[bsize];
+  const int h = block_size_high[bsize];
+  const int sb_row = mi_row / MAX_MIB_SIZE;
+
+  int_mv dv_ref;
+  av1_find_ref_dv(&dv_ref, mi_row, mi_col);
+
+  const MvLimits tmp_mv_limits = x->mv_limits;
+
+  // TODO(aconverse@google.com): Handle same row DV.
+  x->mv_limits.col_min = (tile->mi_col_start - mi_col) * MI_SIZE;
+  x->mv_limits.col_max = (tile->mi_col_end - mi_col) * MI_SIZE - w;
+  x->mv_limits.row_min = (tile->mi_row_start - mi_row) * MI_SIZE;
+  x->mv_limits.row_max = (sb_row * MAX_MIB_SIZE - mi_row) * MI_SIZE - h;
+  assert(x->mv_limits.col_min >= tmp_mv_limits.col_min);
+  assert(x->mv_limits.col_max <= tmp_mv_limits.col_max);
+  assert(x->mv_limits.row_min >= tmp_mv_limits.row_min);
+  assert(x->mv_limits.row_max <= tmp_mv_limits.row_max);
+  av1_set_mv_search_range(&x->mv_limits, &dv_ref.as_mv);
+
+  if (x->mv_limits.col_max < x->mv_limits.col_min ||
+      x->mv_limits.row_max < x->mv_limits.row_min) {
+    x->mv_limits = tmp_mv_limits;
+    return INT64_MAX;
+  }
+
+  struct buf_2d yv12_mb[MAX_MB_PLANE];
+  av1_setup_pred_block(xd, yv12_mb, xd->cur_buf, mi_row, mi_col, NULL, NULL);
+  for (int i = 0; i < MAX_MB_PLANE; ++i) {
+    xd->plane[i].pre[0] = yv12_mb[i];
+  }
+
+  int step_param = cpi->mv_step_param;
+  MV mvp_full = dv_ref.as_mv;
+  mvp_full.col >>= 3;
+  mvp_full.row >>= 3;
+  int sadpb = x->sadperbit16;
+  int cost_list[5];
+  int bestsme = av1_full_pixel_search(cpi, x, bsize, &mvp_full, step_param,
+                                      sadpb, cond_cost_list(cpi, cost_list),
+                                      &dv_ref.as_mv, INT_MAX, 1);
+
+  x->mv_limits = tmp_mv_limits;
+  if (bestsme == INT_MAX) return INT64_MAX;
+  mvp_full = x->best_mv.as_mv;
+  MV dv = {.row = mvp_full.row * 8, .col = mvp_full.col * 8 };
+  if (mv_check_bounds(&x->mv_limits, &dv)) return INT64_MAX;
+  if (!is_dv_valid(dv, tile, mi_row, mi_col, bsize)) return INT64_MAX;
+  MB_MODE_INFO *mbmi = &mi->mbmi;
+  MB_MODE_INFO best_mbmi = *mbmi;
+  RD_STATS best_rdcost = *rd_cost;
+  int best_skip = x->skip;
+#if CONFIG_PALETTE
+  memset(&mbmi->palette_mode_info, 0, sizeof(mbmi->palette_mode_info));
+#endif
+  mbmi->use_intrabc = 1;
+  mbmi->mode = DC_PRED;
+  mbmi->uv_mode = DC_PRED;
+  mbmi->mv[0].as_mv = dv;
+#if CONFIG_DUAL_FILTER
+  for (int idx = 0; idx < 4; ++idx) mbmi->interp_filter[idx] = BILINEAR;
+#else
+  mbmi->interp_filter = BILINEAR;
+#endif
+  mbmi->skip = 0;
+  x->skip = 0;
+  av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL, bsize);
+
+  int rate_mv = av1_mv_bit_cost(&dv, &dv_ref.as_mv, x->nmvjointcost, x->mvcost,
+                                MV_COST_WEIGHT);
+  const PREDICTION_MODE A = av1_above_block_mode(mi, xd->above_mi, 0);
+  const PREDICTION_MODE L = av1_left_block_mode(mi, xd->left_mi, 0);
+  const int rate_mode =
+      cpi->y_mode_costs[A][L][DC_PRED] + av1_cost_bit(INTRABC_PROB, 1);
+
+  RD_STATS rd_stats, rd_stats_uv;
+  av1_subtract_plane(x, bsize, 0);
+  super_block_yrd(cpi, x, &rd_stats, bsize, INT64_MAX);
+  super_block_uvrd(cpi, x, &rd_stats_uv, bsize, INT64_MAX);
+  av1_merge_rd_stats(&rd_stats, &rd_stats_uv);
+#if CONFIG_RD_DEBUG
+  mbmi->rd_stats = rd_stats;
+#endif
+
+  const aom_prob skip_prob = av1_get_skip_prob(cm, xd);
+
+  RD_STATS rdc_noskip;
+  av1_init_rd_stats(&rdc_noskip);
+  rdc_noskip.rate =
+      rate_mode + rate_mv + rd_stats.rate + av1_cost_bit(skip_prob, 0);
+  rdc_noskip.dist = rd_stats.dist;
+  rdc_noskip.rdcost =
+      RDCOST(x->rdmult, x->rddiv, rdc_noskip.rate, rdc_noskip.dist);
+  if (rdc_noskip.rdcost < best_rd) {
+    best_rd = rdc_noskip.rdcost;
+    best_mbmi = *mbmi;
+    best_skip = x->skip;
+    best_rdcost = rdc_noskip;
+  }
+
+  x->skip = 1;
+  mbmi->skip = 1;
+  RD_STATS rdc_skip;
+  av1_init_rd_stats(&rdc_skip);
+  rdc_skip.rate = rate_mode + rate_mv + av1_cost_bit(skip_prob, 1);
+  rdc_skip.dist = rd_stats.sse;
+  rdc_skip.rdcost = RDCOST(x->rdmult, x->rddiv, rdc_skip.rate, rdc_skip.dist);
+  if (rdc_skip.rdcost < best_rd) {
+    best_rd = rdc_skip.rdcost;
+    best_mbmi = *mbmi;
+    best_skip = x->skip;
+    best_rdcost = rdc_skip;
+  }
+  *mbmi = best_mbmi;
+  *rd_cost = best_rdcost;
+  x->skip = best_skip;
+  return best_rd;
+}
+#endif  // CONFIG_INTRABC
+
+void av1_rd_pick_intra_mode_sb(const AV1_COMP *cpi, MACROBLOCK *x,
+                               RD_STATS *rd_cost, BLOCK_SIZE bsize,
+                               PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblockd_plane *const pd = xd->plane;
+  int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
+  int y_skip = 0, uv_skip = 0;
+  int64_t dist_y = 0, dist_uv = 0;
+  TX_SIZE max_uv_tx_size;
+  const int unify_bsize = CONFIG_CB4X4;
+
+  ctx->skip = 0;
+  xd->mi[0]->mbmi.ref_frame[0] = INTRA_FRAME;
+  xd->mi[0]->mbmi.ref_frame[1] = NONE_FRAME;
+#if CONFIG_INTRABC
+  xd->mi[0]->mbmi.use_intrabc = 0;
+#endif  // CONFIG_INTRABC
+
+  const int64_t intra_yrd =
+      (bsize >= BLOCK_8X8 || unify_bsize)
+          ? rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly, &dist_y,
+                                   &y_skip, bsize, best_rd)
+          : rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
+                                         &dist_y, &y_skip, best_rd);
+
+  if (intra_yrd < best_rd) {
+    max_uv_tx_size = uv_txsize_lookup[bsize][xd->mi[0]->mbmi.tx_size]
+                                     [pd[1].subsampling_x][pd[1].subsampling_y];
+
+#if CONFIG_CB4X4
+#if !CONFIG_CHROMA_2X2
+    max_uv_tx_size = AOMMAX(max_uv_tx_size, TX_4X4);
+#endif  // !CONFIG_CHROMA_2X2
+    if (!x->skip_chroma_rd)
+      rd_pick_intra_sbuv_mode(cpi, x, &rate_uv, &rate_uv_tokenonly, &dist_uv,
+                              &uv_skip, bsize, max_uv_tx_size);
+#else
+    rd_pick_intra_sbuv_mode(cpi, x, &rate_uv, &rate_uv_tokenonly, &dist_uv,
+                            &uv_skip, AOMMAX(BLOCK_8X8, bsize), max_uv_tx_size);
+#endif  // CONFIG_CB4X4
+
+    if (y_skip && uv_skip) {
+      rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
+                      av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+      rd_cost->dist = dist_y + dist_uv;
+    } else {
+      rd_cost->rate =
+          rate_y + rate_uv + av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+      rd_cost->dist = dist_y + dist_uv;
+    }
+    rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
+  } else {
+    rd_cost->rate = INT_MAX;
+  }
+
+#if CONFIG_INTRABC
+  if (rd_cost->rate != INT_MAX && rd_cost->rdcost < best_rd)
+    best_rd = rd_cost->rdcost;
+  if (rd_pick_intrabc_mode_sb(cpi, x, rd_cost, bsize, best_rd) < best_rd) {
+    ctx->skip = x->skip;  // FIXME where is the proper place to set this?!
+    assert(rd_cost->rate != INT_MAX);
+    rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
+  }
+#endif
+  if (rd_cost->rate == INT_MAX) return;
+
+  ctx->mic = *xd->mi[0];
+  ctx->mbmi_ext = *x->mbmi_ext;
+}
+
+// Do we have an internal image edge (e.g. formatting bars).
+int av1_internal_image_edge(const AV1_COMP *cpi) {
+  return (cpi->oxcf.pass == 2) &&
+         ((cpi->twopass.this_frame_stats.inactive_zone_rows > 0) ||
+          (cpi->twopass.this_frame_stats.inactive_zone_cols > 0));
+}
+
+// Checks to see if a super block is on a horizontal image edge.
+// In most cases this is the "real" edge unless there are formatting
+// bars embedded in the stream.
+int av1_active_h_edge(const AV1_COMP *cpi, int mi_row, int mi_step) {
+  int top_edge = 0;
+  int bottom_edge = cpi->common.mi_rows;
+  int is_active_h_edge = 0;
+
+  // For two pass account for any formatting bars detected.
+  if (cpi->oxcf.pass == 2) {
+    const TWO_PASS *const twopass = &cpi->twopass;
+
+    // The inactive region is specified in MBs not mi units.
+    // The image edge is in the following MB row.
+    top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
+
+    bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
+    bottom_edge = AOMMAX(top_edge, bottom_edge);
+  }
+
+  if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
+      ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
+    is_active_h_edge = 1;
+  }
+  return is_active_h_edge;
+}
+
+// Checks to see if a super block is on a vertical image edge.
+// In most cases this is the "real" edge unless there are formatting
+// bars embedded in the stream.
+int av1_active_v_edge(const AV1_COMP *cpi, int mi_col, int mi_step) {
+  int left_edge = 0;
+  int right_edge = cpi->common.mi_cols;
+  int is_active_v_edge = 0;
+
+  // For two pass account for any formatting bars detected.
+  if (cpi->oxcf.pass == 2) {
+    const TWO_PASS *const twopass = &cpi->twopass;
+
+    // The inactive region is specified in MBs not mi units.
+    // The image edge is in the following MB row.
+    left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
+
+    right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
+    right_edge = AOMMAX(left_edge, right_edge);
+  }
+
+  if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
+      ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
+    is_active_v_edge = 1;
+  }
+  return is_active_v_edge;
+}
+
+// Checks to see if a super block is at the edge of the active image.
+// In most cases this is the "real" edge unless there are formatting
+// bars embedded in the stream.
+int av1_active_edge_sb(const AV1_COMP *cpi, int mi_row, int mi_col) {
+  return av1_active_h_edge(cpi, mi_row, cpi->common.mib_size) ||
+         av1_active_v_edge(cpi, mi_col, cpi->common.mib_size);
+}
+
+#if CONFIG_PALETTE
+static void restore_uv_color_map(const AV1_COMP *const cpi, MACROBLOCK *x) {
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+  const BLOCK_SIZE bsize = mbmi->sb_type;
+  int src_stride = x->plane[1].src.stride;
+  const uint8_t *const src_u = x->plane[1].src.buf;
+  const uint8_t *const src_v = x->plane[2].src.buf;
+  float *const data = x->palette_buffer->kmeans_data_buf;
+  float centroids[2 * PALETTE_MAX_SIZE];
+  uint8_t *const color_map = xd->plane[1].color_index_map;
+  int r, c;
+#if CONFIG_HIGHBITDEPTH
+  const uint16_t *const src_u16 = CONVERT_TO_SHORTPTR(src_u);
+  const uint16_t *const src_v16 = CONVERT_TO_SHORTPTR(src_v);
+#endif  // CONFIG_HIGHBITDEPTH
+  int plane_block_width, plane_block_height, rows, cols;
+  av1_get_block_dimensions(bsize, 1, xd, &plane_block_width,
+                           &plane_block_height, &rows, &cols);
+  (void)cpi;
+
+  for (r = 0; r < rows; ++r) {
+    for (c = 0; c < cols; ++c) {
+#if CONFIG_HIGHBITDEPTH
+      if (cpi->common.use_highbitdepth) {
+        data[(r * cols + c) * 2] = src_u16[r * src_stride + c];
+        data[(r * cols + c) * 2 + 1] = src_v16[r * src_stride + c];
+      } else {
+#endif  // CONFIG_HIGHBITDEPTH
+        data[(r * cols + c) * 2] = src_u[r * src_stride + c];
+        data[(r * cols + c) * 2 + 1] = src_v[r * src_stride + c];
+#if CONFIG_HIGHBITDEPTH
+      }
+#endif  // CONFIG_HIGHBITDEPTH
+    }
+  }
+
+  for (r = 1; r < 3; ++r) {
+    for (c = 0; c < pmi->palette_size[1]; ++c) {
+      centroids[c * 2 + r - 1] = pmi->palette_colors[r * PALETTE_MAX_SIZE + c];
+    }
+  }
+
+  av1_calc_indices(data, centroids, color_map, rows * cols,
+                   pmi->palette_size[1], 2);
+  extend_palette_color_map(color_map, cols, rows, plane_block_width,
+                           plane_block_height);
+}
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_FILTER_INTRA
+static void pick_filter_intra_interframe(
+    const AV1_COMP *cpi, MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
+    BLOCK_SIZE bsize, int mi_row, int mi_col, int *rate_uv_intra,
+    int *rate_uv_tokenonly, int64_t *dist_uv, int *skip_uv,
+    PREDICTION_MODE *mode_uv, FILTER_INTRA_MODE_INFO *filter_intra_mode_info_uv,
+#if CONFIG_EXT_INTRA
+    int8_t *uv_angle_delta,
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_PALETTE
+    PALETTE_MODE_INFO *pmi_uv, int palette_ctx,
+#endif  // CONFIG_PALETTE
+    int skip_mask, unsigned int *ref_costs_single, int64_t *best_rd,
+    int64_t *best_intra_rd, PREDICTION_MODE *best_intra_mode,
+    int *best_mode_index, int *best_skip2, int *best_mode_skippable,
+#if CONFIG_SUPERTX
+    int *returnrate_nocoef,
+#endif  // CONFIG_SUPERTX
+    int64_t *best_pred_rd, MB_MODE_INFO *best_mbmode, RD_STATS *rd_cost) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+#if CONFIG_PALETTE
+  PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+#endif  // CONFIG_PALETTE
+  int rate2 = 0, rate_y = INT_MAX, skippable = 0, rate_uv, rate_dummy, i;
+  int dc_mode_index;
+  const int *const intra_mode_cost = cpi->mbmode_cost[size_group_lookup[bsize]];
+  int64_t distortion2 = 0, distortion_y = 0, this_rd = *best_rd;
+  int64_t distortion_uv, model_rd = INT64_MAX;
+  TX_SIZE uv_tx;
+
+  for (i = 0; i < MAX_MODES; ++i)
+    if (av1_mode_order[i].mode == DC_PRED &&
+        av1_mode_order[i].ref_frame[0] == INTRA_FRAME)
+      break;
+  dc_mode_index = i;
+  assert(i < MAX_MODES);
+
+  // TODO(huisu): use skip_mask for further speedup.
+  (void)skip_mask;
+  mbmi->mode = DC_PRED;
+  mbmi->uv_mode = DC_PRED;
+  mbmi->ref_frame[0] = INTRA_FRAME;
+  mbmi->ref_frame[1] = NONE_FRAME;
+  if (!rd_pick_filter_intra_sby(cpi, x, &rate_dummy, &rate_y, &distortion_y,
+                                &skippable, bsize, intra_mode_cost[mbmi->mode],
+                                &this_rd, &model_rd, 0)) {
+    return;
+  }
+  if (rate_y == INT_MAX) return;
+
+  uv_tx = uv_txsize_lookup[bsize][mbmi->tx_size][xd->plane[1].subsampling_x]
+                          [xd->plane[1].subsampling_y];
+  if (rate_uv_intra[uv_tx] == INT_MAX) {
+    choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx, &rate_uv_intra[uv_tx],
+                         &rate_uv_tokenonly[uv_tx], &dist_uv[uv_tx],
+                         &skip_uv[uv_tx], &mode_uv[uv_tx]);
+#if CONFIG_PALETTE
+    if (cm->allow_screen_content_tools) pmi_uv[uv_tx] = *pmi;
+#endif  // CONFIG_PALETTE
+    filter_intra_mode_info_uv[uv_tx] = mbmi->filter_intra_mode_info;
+#if CONFIG_EXT_INTRA
+    uv_angle_delta[uv_tx] = mbmi->angle_delta[1];
+#endif  // CONFIG_EXT_INTRA
+  }
+
+  rate_uv = rate_uv_tokenonly[uv_tx];
+  distortion_uv = dist_uv[uv_tx];
+  skippable = skippable && skip_uv[uv_tx];
+  mbmi->uv_mode = mode_uv[uv_tx];
+#if CONFIG_PALETTE
+  if (cm->allow_screen_content_tools) {
+    pmi->palette_size[1] = pmi_uv[uv_tx].palette_size[1];
+    memcpy(pmi->palette_colors + PALETTE_MAX_SIZE,
+           pmi_uv[uv_tx].palette_colors + PALETTE_MAX_SIZE,
+           2 * PALETTE_MAX_SIZE * sizeof(pmi->palette_colors[0]));
+  }
+#endif  // CONFIG_PALETTE
+#if CONFIG_EXT_INTRA
+  mbmi->angle_delta[1] = uv_angle_delta[uv_tx];
+#endif  // CONFIG_EXT_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[1] =
+      filter_intra_mode_info_uv[uv_tx].use_filter_intra_mode[1];
+  if (filter_intra_mode_info_uv[uv_tx].use_filter_intra_mode[1]) {
+    mbmi->filter_intra_mode_info.filter_intra_mode[1] =
+        filter_intra_mode_info_uv[uv_tx].filter_intra_mode[1];
+  }
+
+  rate2 = rate_y + intra_mode_cost[mbmi->mode] + rate_uv +
+          cpi->intra_uv_mode_cost[mbmi->mode][mbmi->uv_mode];
+#if CONFIG_PALETTE
+  if (cpi->common.allow_screen_content_tools && mbmi->mode == DC_PRED &&
+      bsize >= BLOCK_8X8)
+    rate2 += av1_cost_bit(
+        av1_default_palette_y_mode_prob[bsize - BLOCK_8X8][palette_ctx], 0);
+#endif  // CONFIG_PALETTE
+
+  if (!xd->lossless[mbmi->segment_id]) {
+    // super_block_yrd above includes the cost of the tx_size in the
+    // tokenonly rate, but for intra blocks, tx_size is always coded
+    // (prediction granularity), so we account for it in the full rate,
+    // not the tokenonly rate.
+    rate_y -= tx_size_cost(cpi, x, bsize, mbmi->tx_size);
+  }
+
+  rate2 += av1_cost_bit(cm->fc->filter_intra_probs[0],
+                        mbmi->filter_intra_mode_info.use_filter_intra_mode[0]);
+  rate2 += write_uniform_cost(
+      FILTER_INTRA_MODES, mbmi->filter_intra_mode_info.filter_intra_mode[0]);
+#if CONFIG_EXT_INTRA
+  if (av1_is_directional_mode(mbmi->uv_mode, bsize)) {
+    rate2 += write_uniform_cost(2 * MAX_ANGLE_DELTA + 1,
+                                MAX_ANGLE_DELTA + mbmi->angle_delta[1]);
+  }
+#endif  // CONFIG_EXT_INTRA
+  if (mbmi->mode == DC_PRED) {
+    rate2 +=
+        av1_cost_bit(cpi->common.fc->filter_intra_probs[1],
+                     mbmi->filter_intra_mode_info.use_filter_intra_mode[1]);
+    if (mbmi->filter_intra_mode_info.use_filter_intra_mode[1])
+      rate2 +=
+          write_uniform_cost(FILTER_INTRA_MODES,
+                             mbmi->filter_intra_mode_info.filter_intra_mode[1]);
+  }
+  distortion2 = distortion_y + distortion_uv;
+  av1_encode_intra_block_plane((AV1_COMMON *)cm, x, bsize, 0, 0, mi_row,
+                               mi_col);
+
+  rate2 += ref_costs_single[INTRA_FRAME];
+
+  if (skippable) {
+    rate2 -= (rate_y + rate_uv);
+    rate_y = 0;
+    rate_uv = 0;
+    rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+  } else {
+    rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+  }
+  this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+
+  if (this_rd < *best_intra_rd) {
+    *best_intra_rd = this_rd;
+    *best_intra_mode = mbmi->mode;
+  }
+  for (i = 0; i < REFERENCE_MODES; ++i)
+    best_pred_rd[i] = AOMMIN(best_pred_rd[i], this_rd);
+
+  if (this_rd < *best_rd) {
+    *best_mode_index = dc_mode_index;
+    mbmi->mv[0].as_int = 0;
+    rd_cost->rate = rate2;
+#if CONFIG_SUPERTX
+    if (x->skip)
+      *returnrate_nocoef = rate2;
+    else
+      *returnrate_nocoef = rate2 - rate_y - rate_uv;
+    *returnrate_nocoef -= av1_cost_bit(av1_get_skip_prob(cm, xd), skippable);
+    *returnrate_nocoef -= av1_cost_bit(av1_get_intra_inter_prob(cm, xd),
+                                       mbmi->ref_frame[0] != INTRA_FRAME);
+#endif  // CONFIG_SUPERTX
+    rd_cost->dist = distortion2;
+    rd_cost->rdcost = this_rd;
+    *best_rd = this_rd;
+    *best_mbmode = *mbmi;
+    *best_skip2 = 0;
+    *best_mode_skippable = skippable;
+  }
+}
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_MOTION_VAR
+static void calc_target_weighted_pred(const AV1_COMMON *cm, const MACROBLOCK *x,
+                                      const MACROBLOCKD *xd, int mi_row,
+                                      int mi_col, const uint8_t *above,
+                                      int above_stride, const uint8_t *left,
+                                      int left_stride);
+#endif  // CONFIG_MOTION_VAR
+
+void av1_rd_pick_inter_mode_sb(const AV1_COMP *cpi, TileDataEnc *tile_data,
+                               MACROBLOCK *x, int mi_row, int mi_col,
+                               RD_STATS *rd_cost,
+#if CONFIG_SUPERTX
+                               int *returnrate_nocoef,
+#endif  // CONFIG_SUPERTX
+                               BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                               int64_t best_rd_so_far) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const RD_OPT *const rd_opt = &cpi->rd;
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+#if CONFIG_PALETTE
+  const int try_palette =
+      cpi->common.allow_screen_content_tools && bsize >= BLOCK_8X8;
+  PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+#endif  // CONFIG_PALETTE
+  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
+  const struct segmentation *const seg = &cm->seg;
+  PREDICTION_MODE this_mode;
+  MV_REFERENCE_FRAME ref_frame, second_ref_frame;
+  unsigned char segment_id = mbmi->segment_id;
+  int comp_pred, i, k;
+  int_mv frame_mv[MB_MODE_COUNT][TOTAL_REFS_PER_FRAME];
+  struct buf_2d yv12_mb[TOTAL_REFS_PER_FRAME][MAX_MB_PLANE];
+  int_mv single_newmv[TOTAL_REFS_PER_FRAME] = { { 0 } };
+#if CONFIG_EXT_INTER
+  int single_newmv_rate[TOTAL_REFS_PER_FRAME] = { 0 };
+  int64_t modelled_rd[MB_MODE_COUNT][TOTAL_REFS_PER_FRAME];
+#endif  // CONFIG_EXT_INTER
+  static const int flag_list[TOTAL_REFS_PER_FRAME] = {
+    0,
+    AOM_LAST_FLAG,
+#if CONFIG_EXT_REFS
+    AOM_LAST2_FLAG,
+    AOM_LAST3_FLAG,
+#endif  // CONFIG_EXT_REFS
+    AOM_GOLD_FLAG,
+#if CONFIG_EXT_REFS
+    AOM_BWD_FLAG,
+#endif  // CONFIG_EXT_REFS
+    AOM_ALT_FLAG
+  };
+  int64_t best_rd = best_rd_so_far;
+  int best_rate_y = INT_MAX, best_rate_uv = INT_MAX;
+  int64_t best_pred_diff[REFERENCE_MODES];
+  int64_t best_pred_rd[REFERENCE_MODES];
+  MB_MODE_INFO best_mbmode;
+#if CONFIG_REF_MV
+  int rate_skip0 = av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+  int rate_skip1 = av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+#endif  // CONFIG_REF_MV
+  int best_mode_skippable = 0;
+  int midx, best_mode_index = -1;
+  unsigned int ref_costs_single[TOTAL_REFS_PER_FRAME];
+  unsigned int ref_costs_comp[TOTAL_REFS_PER_FRAME];
+  aom_prob comp_mode_p;
+  int64_t best_intra_rd = INT64_MAX;
+  unsigned int best_pred_sse = UINT_MAX;
+  PREDICTION_MODE best_intra_mode = DC_PRED;
+  int rate_uv_intra[TX_SIZES_ALL], rate_uv_tokenonly[TX_SIZES_ALL];
+  int64_t dist_uvs[TX_SIZES_ALL];
+  int skip_uvs[TX_SIZES_ALL];
+  PREDICTION_MODE mode_uv[TX_SIZES_ALL];
+#if CONFIG_PALETTE
+  PALETTE_MODE_INFO pmi_uv[TX_SIZES_ALL];
+#endif  // CONFIG_PALETTE
+#if CONFIG_EXT_INTRA
+  int8_t uv_angle_delta[TX_SIZES_ALL];
+  int is_directional_mode, angle_stats_ready = 0;
+  uint8_t directional_mode_skip_mask[INTRA_MODES];
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+  int8_t dc_skipped = 1;
+  FILTER_INTRA_MODE_INFO filter_intra_mode_info_uv[TX_SIZES_ALL];
+#endif  // CONFIG_FILTER_INTRA
+  const int intra_cost_penalty = av1_get_intra_cost_penalty(
+      cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
+  const int *const intra_mode_cost = cpi->mbmode_cost[size_group_lookup[bsize]];
+  int best_skip2 = 0;
+  uint8_t ref_frame_skip_mask[2] = { 0 };
+#if CONFIG_EXT_INTER
+  uint32_t mode_skip_mask[TOTAL_REFS_PER_FRAME] = { 0 };
+  MV_REFERENCE_FRAME best_single_inter_ref = LAST_FRAME;
+  int64_t best_single_inter_rd = INT64_MAX;
+#else
+  uint16_t mode_skip_mask[TOTAL_REFS_PER_FRAME] = { 0 };
+#endif  // CONFIG_EXT_INTER
+  int mode_skip_start = sf->mode_skip_start + 1;
+  const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
+  const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
+  int64_t mode_threshold[MAX_MODES];
+  int *mode_map = tile_data->mode_map[bsize];
+  const int mode_search_skip_flags = sf->mode_search_skip_flags;
+#if CONFIG_PVQ
+  od_rollback_buffer pre_buf;
+#endif  // CONFIG_PVQ
+
+  HandleInterModeArgs args = {
+#if CONFIG_MOTION_VAR
+    { NULL },
+    { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE },
+    { NULL },
+    { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE },
+#endif  // CONFIG_MOTION_VAR
+#if CONFIG_EXT_INTER
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+    NULL,
+#else   // CONFIG_EXT_INTER
+    NULL,
+#endif  // CONFIG_EXT_INTER
+    { { 0 } },
+  };
+
+#if CONFIG_PALETTE || CONFIG_EXT_INTRA
+  const int rows = block_size_high[bsize];
+  const int cols = block_size_wide[bsize];
+#endif  // CONFIG_PALETTE || CONFIG_EXT_INTRA
+#if CONFIG_PALETTE
+  int palette_ctx = 0;
+  const MODE_INFO *above_mi = xd->above_mi;
+  const MODE_INFO *left_mi = xd->left_mi;
+#endif  // CONFIG_PALETTE
+#if CONFIG_MOTION_VAR
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf1[2 * MAX_MB_PLANE * MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf2[2 * MAX_MB_PLANE * MAX_SB_SQUARE]);
+#else
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf1[MAX_MB_PLANE * MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(16, uint8_t, tmp_buf2[MAX_MB_PLANE * MAX_SB_SQUARE]);
+#endif  // CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, int32_t, weighted_src_buf[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(16, int32_t, mask2d_buf[MAX_SB_SQUARE]);
+  int dst_width1[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_width2[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_height1[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+  int dst_height2[MAX_MB_PLANE] = { MAX_SB_SIZE, MAX_SB_SIZE, MAX_SB_SIZE };
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    int len = sizeof(uint16_t);
+    args.above_pred_buf[0] = CONVERT_TO_BYTEPTR(tmp_buf1);
+    args.above_pred_buf[1] = CONVERT_TO_BYTEPTR(tmp_buf1 + MAX_SB_SQUARE * len);
+    args.above_pred_buf[2] =
+        CONVERT_TO_BYTEPTR(tmp_buf1 + 2 * MAX_SB_SQUARE * len);
+    args.left_pred_buf[0] = CONVERT_TO_BYTEPTR(tmp_buf2);
+    args.left_pred_buf[1] = CONVERT_TO_BYTEPTR(tmp_buf2 + MAX_SB_SQUARE * len);
+    args.left_pred_buf[2] =
+        CONVERT_TO_BYTEPTR(tmp_buf2 + 2 * MAX_SB_SQUARE * len);
+  } else {
+#endif  // CONFIG_HIGHBITDEPTH
+    args.above_pred_buf[0] = tmp_buf1;
+    args.above_pred_buf[1] = tmp_buf1 + MAX_SB_SQUARE;
+    args.above_pred_buf[2] = tmp_buf1 + 2 * MAX_SB_SQUARE;
+    args.left_pred_buf[0] = tmp_buf2;
+    args.left_pred_buf[1] = tmp_buf2 + MAX_SB_SQUARE;
+    args.left_pred_buf[2] = tmp_buf2 + 2 * MAX_SB_SQUARE;
+#if CONFIG_HIGHBITDEPTH
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // CONFIG_MOTION_VAR
+
+  av1_zero(best_mbmode);
+
+#if CONFIG_PALETTE
+  av1_zero(pmi_uv);
+  if (try_palette) {
+    if (above_mi)
+      palette_ctx += (above_mi->mbmi.palette_mode_info.palette_size[0] > 0);
+    if (left_mi)
+      palette_ctx += (left_mi->mbmi.palette_mode_info.palette_size[0] > 0);
+  }
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_EXT_INTRA
+  memset(directional_mode_skip_mask, 0,
+         sizeof(directional_mode_skip_mask[0]) * INTRA_MODES);
+#endif  // CONFIG_EXT_INTRA
+
+  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
+                           &comp_mode_p);
+
+  for (i = 0; i < REFERENCE_MODES; ++i) best_pred_rd[i] = INT64_MAX;
+  for (i = 0; i < TX_SIZES_ALL; i++) rate_uv_intra[i] = INT_MAX;
+  for (i = 0; i < TOTAL_REFS_PER_FRAME; ++i) x->pred_sse[i] = INT_MAX;
+  for (i = 0; i < MB_MODE_COUNT; ++i) {
+    for (k = 0; k < TOTAL_REFS_PER_FRAME; ++k) {
+      args.single_filter[i][k] = SWITCHABLE;
+    }
+  }
+
+  rd_cost->rate = INT_MAX;
+#if CONFIG_SUPERTX
+  *returnrate_nocoef = INT_MAX;
+#endif  // CONFIG_SUPERTX
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    x->pred_mv_sad[ref_frame] = INT_MAX;
+    x->mbmi_ext->mode_context[ref_frame] = 0;
+#if CONFIG_REF_MV && CONFIG_EXT_INTER
+    x->mbmi_ext->compound_mode_context[ref_frame] = 0;
+#endif  // CONFIG_REF_MV && CONFIG_EXT_INTER
+    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
+      assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
+      setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
+                         frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
+    }
+    frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
+#if CONFIG_GLOBAL_MOTION
+    frame_mv[ZEROMV][ref_frame].as_int =
+        gm_get_motion_vector(&cm->global_motion[ref_frame],
+                             cm->allow_high_precision_mv, bsize, mi_col, mi_row,
+                             0)
+            .as_int;
+#else   // CONFIG_GLOBAL_MOTION
+    frame_mv[ZEROMV][ref_frame].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_EXT_INTER
+    frame_mv[NEW_NEWMV][ref_frame].as_int = INVALID_MV;
+#if CONFIG_GLOBAL_MOTION
+    frame_mv[ZERO_ZEROMV][ref_frame].as_int =
+        gm_get_motion_vector(&cm->global_motion[ref_frame],
+                             cm->allow_high_precision_mv, bsize, mi_col, mi_row,
+                             0)
+            .as_int;
+#else   // CONFIG_GLOBAL_MOTION
+    frame_mv[ZERO_ZEROMV][ref_frame].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+#endif  // CONFIG_EXT_INTER
+  }
+
+#if CONFIG_REF_MV
+  for (; ref_frame < MODE_CTX_REF_FRAMES; ++ref_frame) {
+    MODE_INFO *const mi = xd->mi[0];
+    int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
+    x->mbmi_ext->mode_context[ref_frame] = 0;
+    av1_find_mv_refs(cm, xd, mi, ref_frame, &mbmi_ext->ref_mv_count[ref_frame],
+                     mbmi_ext->ref_mv_stack[ref_frame],
+#if CONFIG_EXT_INTER
+                     mbmi_ext->compound_mode_context,
+#endif  // CONFIG_EXT_INTER
+                     candidates, mi_row, mi_col, NULL, NULL,
+                     mbmi_ext->mode_context);
+    if (mbmi_ext->ref_mv_count[ref_frame] < 2) {
+      MV_REFERENCE_FRAME rf[2];
+      av1_set_ref_frame(rf, ref_frame);
+      if (mbmi_ext->ref_mvs[rf[0]][0].as_int !=
+              frame_mv[ZEROMV][rf[0]].as_int ||
+          mbmi_ext->ref_mvs[rf[0]][1].as_int !=
+              frame_mv[ZEROMV][rf[0]].as_int ||
+          mbmi_ext->ref_mvs[rf[1]][0].as_int !=
+              frame_mv[ZEROMV][rf[1]].as_int ||
+          mbmi_ext->ref_mvs[rf[1]][1].as_int != frame_mv[ZEROMV][rf[1]].as_int)
+        mbmi_ext->mode_context[ref_frame] &= ~(1 << ALL_ZERO_FLAG_OFFSET);
+    }
+  }
+#endif  // CONFIG_REF_MV
+
+#if CONFIG_MOTION_VAR
+  av1_count_overlappable_neighbors(cm, xd, mi_row, mi_col);
+  if (check_num_overlappable_neighbors(mbmi) &&
+      is_motion_variation_allowed_bsize(bsize)) {
+    av1_build_prediction_by_above_preds(cm, xd, mi_row, mi_col,
+                                        args.above_pred_buf, dst_width1,
+                                        dst_height1, args.above_pred_stride);
+    av1_build_prediction_by_left_preds(cm, xd, mi_row, mi_col,
+                                       args.left_pred_buf, dst_width2,
+                                       dst_height2, args.left_pred_stride);
+    av1_setup_dst_planes(xd->plane, bsize, get_frame_new_buffer(cm), mi_row,
+                         mi_col);
+    x->mask_buf = mask2d_buf;
+    x->wsrc_buf = weighted_src_buf;
+    calc_target_weighted_pred(cm, x, xd, mi_row, mi_col, args.above_pred_buf[0],
+                              args.above_pred_stride[0], args.left_pred_buf[0],
+                              args.left_pred_stride[0]);
+  }
+#endif  // CONFIG_MOTION_VAR
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
+    if (!(cpi->ref_frame_flags & flag_list[ref_frame])) {
+// Skip checking missing references in both single and compound reference
+// modes. Note that a mode will be skipped iff both reference frames
+// are masked out.
+#if CONFIG_EXT_REFS
+      if (ref_frame == BWDREF_FRAME || ref_frame == ALTREF_FRAME) {
+        ref_frame_skip_mask[0] |= (1 << ref_frame);
+        ref_frame_skip_mask[1] |= ((1 << ref_frame) | 0x01);
+      } else {
+#endif  // CONFIG_EXT_REFS
+        ref_frame_skip_mask[0] |= (1 << ref_frame);
+        ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+#if CONFIG_EXT_REFS
+      }
+#endif  // CONFIG_EXT_REFS
+    } else {
+      for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
+        // Skip fixed mv modes for poor references
+        if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
+          mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
+          break;
+        }
+      }
+    }
+    // If the segment reference frame feature is enabled....
+    // then do nothing if the current ref frame is not allowed..
+    if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
+        get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
+      ref_frame_skip_mask[0] |= (1 << ref_frame);
+      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+    }
+  }
+
+  // Disable this drop out case if the ref frame
+  // segment level feature is enabled for this segment. This is to
+  // prevent the possibility that we end up unable to pick any mode.
+  if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
+    // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
+    // unless ARNR filtering is enabled in which case we want
+    // an unfiltered alternative. We allow near/nearest as well
+    // because they may result in zero-zero MVs but be cheaper.
+    if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
+      int_mv zeromv;
+      ref_frame_skip_mask[0] = (1 << LAST_FRAME) |
+#if CONFIG_EXT_REFS
+                               (1 << LAST2_FRAME) | (1 << LAST3_FRAME) |
+                               (1 << BWDREF_FRAME) |
+#endif  // CONFIG_EXT_REFS
+                               (1 << GOLDEN_FRAME);
+      ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
+      // TODO(zoeliu): To further explore whether following needs to be done for
+      //               BWDREF_FRAME as well.
+      mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
+#if CONFIG_GLOBAL_MOTION
+      zeromv.as_int = gm_get_motion_vector(&cm->global_motion[ALTREF_FRAME],
+                                           cm->allow_high_precision_mv, bsize,
+                                           mi_col, mi_row, 0)
+                          .as_int;
+#else
+      zeromv.as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+      if (frame_mv[NEARMV][ALTREF_FRAME].as_int != zeromv.as_int)
+        mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
+      if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != zeromv.as_int)
+        mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
+#if CONFIG_EXT_INTER
+      if (frame_mv[NEAREST_NEARESTMV][ALTREF_FRAME].as_int != zeromv.as_int)
+        mode_skip_mask[ALTREF_FRAME] |= (1 << NEAREST_NEARESTMV);
+      if (frame_mv[NEAREST_NEARMV][ALTREF_FRAME].as_int != zeromv.as_int)
+        mode_skip_mask[ALTREF_FRAME] |= (1 << NEAREST_NEARMV);
+      if (frame_mv[NEAR_NEARESTMV][ALTREF_FRAME].as_int != zeromv.as_int)
+        mode_skip_mask[ALTREF_FRAME] |= (1 << NEAR_NEARESTMV);
+      if (frame_mv[NEAR_NEARMV][ALTREF_FRAME].as_int != zeromv.as_int)
+        mode_skip_mask[ALTREF_FRAME] |= (1 << NEAR_NEARMV);
+#endif  // CONFIG_EXT_INTER
+    }
+  }
+
+  if (cpi->rc.is_src_frame_alt_ref) {
+    if (sf->alt_ref_search_fp) {
+      assert(cpi->ref_frame_flags & flag_list[ALTREF_FRAME]);
+      mode_skip_mask[ALTREF_FRAME] = 0;
+      ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME);
+      ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
+    }
+  }
+
+  if (sf->alt_ref_search_fp)
+    if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
+      if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
+        mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
+
+  if (sf->adaptive_mode_search) {
+    if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
+        cpi->rc.frames_since_golden >= 3)
+      if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
+        mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
+  }
+
+  if (bsize > sf->max_intra_bsize) {
+    ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
+    ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
+  }
+
+  mode_skip_mask[INTRA_FRAME] |=
+      ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
+
+  for (i = 0; i <= LAST_NEW_MV_INDEX; ++i) mode_threshold[i] = 0;
+  for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
+    mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
+
+  midx = sf->schedule_mode_search ? mode_skip_start : 0;
+  while (midx > 4) {
+    uint8_t end_pos = 0;
+    for (i = 5; i < midx; ++i) {
+      if (mode_threshold[mode_map[i - 1]] > mode_threshold[mode_map[i]]) {
+        uint8_t tmp = mode_map[i];
+        mode_map[i] = mode_map[i - 1];
+        mode_map[i - 1] = tmp;
+        end_pos = i;
+      }
+    }
+    midx = end_pos;
+  }
+
+  if (cpi->sf.tx_type_search.fast_intra_tx_type_search)
+    x->use_default_intra_tx_type = 1;
+  else
+    x->use_default_intra_tx_type = 0;
+
+  if (cpi->sf.tx_type_search.fast_inter_tx_type_search)
+    x->use_default_inter_tx_type = 1;
+  else
+    x->use_default_inter_tx_type = 0;
+#if CONFIG_PVQ
+  od_encode_checkpoint(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+#if CONFIG_EXT_INTER
+  for (i = 0; i < MB_MODE_COUNT; ++i)
+    for (ref_frame = 0; ref_frame < TOTAL_REFS_PER_FRAME; ++ref_frame)
+      modelled_rd[i][ref_frame] = INT64_MAX;
+#endif  // CONFIG_EXT_INTER
+
+  for (midx = 0; midx < MAX_MODES; ++midx) {
+    int mode_index;
+    int mode_excluded = 0;
+    int64_t this_rd = INT64_MAX;
+    int disable_skip = 0;
+    int compmode_cost = 0;
+#if CONFIG_EXT_INTER
+    int compmode_interintra_cost = 0;
+    int compmode_interinter_cost = 0;
+#endif  // CONFIG_EXT_INTER
+    int rate2 = 0, rate_y = 0, rate_uv = 0;
+    int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
+    int skippable = 0;
+    int this_skip2 = 0;
+    int64_t total_sse = INT64_MAX;
+#if CONFIG_REF_MV
+    uint8_t ref_frame_type;
+#endif  // CONFIG_REF_MV
+#if CONFIG_PVQ
+    od_encode_rollback(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+    mode_index = mode_map[midx];
+    this_mode = av1_mode_order[mode_index].mode;
+    ref_frame = av1_mode_order[mode_index].ref_frame[0];
+    second_ref_frame = av1_mode_order[mode_index].ref_frame[1];
+#if CONFIG_REF_MV
+    mbmi->ref_mv_idx = 0;
+#endif  // CONFIG_REF_MV
+
+#if CONFIG_EXT_INTER
+    if (ref_frame > INTRA_FRAME && second_ref_frame == INTRA_FRAME) {
+      // Mode must by compatible
+      if (!is_interintra_allowed_mode(this_mode)) continue;
+      if (!is_interintra_allowed_bsize(bsize)) continue;
+    }
+
+    if (is_inter_compound_mode(this_mode)) {
+      frame_mv[this_mode][ref_frame].as_int =
+          frame_mv[compound_ref0_mode(this_mode)][ref_frame].as_int;
+      frame_mv[this_mode][second_ref_frame].as_int =
+          frame_mv[compound_ref1_mode(this_mode)][second_ref_frame].as_int;
+    }
+#endif  // CONFIG_EXT_INTER
+
+    // Look at the reference frame of the best mode so far and set the
+    // skip mask to look at a subset of the remaining modes.
+    if (midx == mode_skip_start && best_mode_index >= 0) {
+      switch (best_mbmode.ref_frame[0]) {
+        case INTRA_FRAME: break;
+        case LAST_FRAME:
+          ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK;
+          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+          break;
+#if CONFIG_EXT_REFS
+        case LAST2_FRAME:
+          ref_frame_skip_mask[0] |= LAST2_FRAME_MODE_MASK;
+          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+          break;
+        case LAST3_FRAME:
+          ref_frame_skip_mask[0] |= LAST3_FRAME_MODE_MASK;
+          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+          break;
+#endif  // CONFIG_EXT_REFS
+        case GOLDEN_FRAME:
+          ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
+          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+          break;
+#if CONFIG_EXT_REFS
+        case BWDREF_FRAME:
+          ref_frame_skip_mask[0] |= BWDREF_FRAME_MODE_MASK;
+          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+          break;
+#endif  // CONFIG_EXT_REFS
+        case ALTREF_FRAME: ref_frame_skip_mask[0] |= ALTREF_FRAME_MODE_MASK;
+#if CONFIG_EXT_REFS
+          ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+#endif  // CONFIG_EXT_REFS
+          break;
+        case NONE_FRAME:
+        case TOTAL_REFS_PER_FRAME:
+          assert(0 && "Invalid Reference frame");
+          break;
+      }
+    }
+
+    if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
+        (ref_frame_skip_mask[1] & (1 << AOMMAX(0, second_ref_frame))))
+      continue;
+
+    if (mode_skip_mask[ref_frame] & (1 << this_mode)) continue;
+
+    // Test best rd so far against threshold for trying this mode.
+    if (best_mode_skippable && sf->schedule_mode_search)
+      mode_threshold[mode_index] <<= 1;
+
+    if (best_rd < mode_threshold[mode_index]) continue;
+
+    // This is only used in motion vector unit test.
+    if (cpi->oxcf.motion_vector_unit_test && ref_frame == INTRA_FRAME) continue;
+
+#if CONFIG_LOWDELAY_COMPOUND  // Changes LL bitstream
+#if CONFIG_EXT_REFS
+    if (cpi->oxcf.pass == 0) {
+      // Complexity-compression trade-offs
+      // if (ref_frame == ALTREF_FRAME) continue;
+      // if (ref_frame == BWDREF_FRAME) continue;
+      if (second_ref_frame == ALTREF_FRAME) continue;
+      // if (second_ref_frame == BWDREF_FRAME) continue;
+    }
+#endif
+#endif
+    comp_pred = second_ref_frame > INTRA_FRAME;
+    if (comp_pred) {
+      if (!cpi->allow_comp_inter_inter) continue;
+
+      // Skip compound inter modes if ARF is not available.
+      if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
+
+      // Do not allow compound prediction if the segment level reference frame
+      // feature is in use as in this case there can only be one reference.
+      if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) continue;
+
+      if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
+          best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
+        continue;
+
+      mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
+    } else {
+      if (ref_frame != INTRA_FRAME)
+        mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
+    }
+
+    if (ref_frame == INTRA_FRAME) {
+      if (sf->adaptive_mode_search)
+        if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
+          continue;
+
+      if (this_mode != DC_PRED) {
+        // Disable intra modes other than DC_PRED for blocks with low variance
+        // Threshold for intra skipping based on source variance
+        // TODO(debargha): Specialize the threshold for super block sizes
+        const unsigned int skip_intra_var_thresh = 64;
+        if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
+            x->source_variance < skip_intra_var_thresh)
+          continue;
+        // Only search the oblique modes if the best so far is
+        // one of the neighboring directional modes
+        if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
+            (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
+          if (best_mode_index >= 0 && best_mbmode.ref_frame[0] > INTRA_FRAME)
+            continue;
+        }
+        if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
+          if (conditional_skipintra(this_mode, best_intra_mode)) continue;
+        }
+      }
+#if CONFIG_GLOBAL_MOTION
+    } else if (cm->global_motion[ref_frame].wmtype == IDENTITY &&
+               (!comp_pred ||
+                cm->global_motion[second_ref_frame].wmtype == IDENTITY)) {
+#else   // CONFIG_GLOBAL_MOTION
+    } else {
+#endif  // CONFIG_GLOBAL_MOTION
+      const MV_REFERENCE_FRAME ref_frames[2] = { ref_frame, second_ref_frame };
+      if (!check_best_zero_mv(cpi, mbmi_ext->mode_context,
+#if CONFIG_REF_MV && CONFIG_EXT_INTER
+                              mbmi_ext->compound_mode_context,
+#endif  // CONFIG_REF_MV && CONFIG_EXT_INTER
+                              frame_mv, this_mode, ref_frames, bsize, -1,
+                              mi_row, mi_col))
+        continue;
+    }
+
+    mbmi->mode = this_mode;
+    mbmi->uv_mode = DC_PRED;
+    mbmi->ref_frame[0] = ref_frame;
+    mbmi->ref_frame[1] = second_ref_frame;
+#if CONFIG_PALETTE
+    pmi->palette_size[0] = 0;
+    pmi->palette_size[1] = 0;
+#endif  // CONFIG_PALETTE
+#if CONFIG_FILTER_INTRA
+    mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+    mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+#endif  // CONFIG_FILTER_INTRA
+        // Evaluate all sub-pel filters irrespective of whether we can use
+        // them for this frame.
+
+    set_default_interp_filters(mbmi, cm->interp_filter);
+
+    mbmi->mv[0].as_int = mbmi->mv[1].as_int = 0;
+    mbmi->motion_mode = SIMPLE_TRANSLATION;
+
+    x->skip = 0;
+    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
+
+    // Select prediction reference frames.
+    for (i = 0; i < MAX_MB_PLANE; i++) {
+      xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
+      if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
+    }
+
+#if CONFIG_EXT_INTER
+    mbmi->interintra_mode = (INTERINTRA_MODE)(II_DC_PRED - 1);
+#endif  // CONFIG_EXT_INTER
+
+    if (ref_frame == INTRA_FRAME) {
+      RD_STATS rd_stats_y;
+      TX_SIZE uv_tx;
+      struct macroblockd_plane *const pd = &xd->plane[1];
+#if CONFIG_EXT_INTRA
+      is_directional_mode = av1_is_directional_mode(mbmi->mode, bsize);
+      if (is_directional_mode) {
+        int rate_dummy;
+        int64_t model_rd = INT64_MAX;
+        if (!angle_stats_ready) {
+          const int src_stride = x->plane[0].src.stride;
+          const uint8_t *src = x->plane[0].src.buf;
+#if CONFIG_HIGHBITDEPTH
+          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
+            highbd_angle_estimation(src, src_stride, rows, cols,
+                                    directional_mode_skip_mask);
+          else
+#endif  // CONFIG_HIGHBITDEPTH
+            angle_estimation(src, src_stride, rows, cols,
+                             directional_mode_skip_mask);
+          angle_stats_ready = 1;
+        }
+        if (directional_mode_skip_mask[mbmi->mode]) continue;
+        rd_stats_y.rate = INT_MAX;
+        rd_pick_intra_angle_sby(cpi, x, &rate_dummy, &rd_stats_y, bsize,
+                                intra_mode_cost[mbmi->mode], best_rd,
+                                &model_rd);
+      } else {
+        mbmi->angle_delta[0] = 0;
+        super_block_yrd(cpi, x, &rd_stats_y, bsize, best_rd);
+      }
+#else
+      super_block_yrd(cpi, x, &rd_stats_y, bsize, best_rd);
+#endif  // CONFIG_EXT_INTRA
+      rate_y = rd_stats_y.rate;
+      distortion_y = rd_stats_y.dist;
+      skippable = rd_stats_y.skip;
+
+      if (rate_y == INT_MAX) continue;
+
+#if CONFIG_FILTER_INTRA
+      if (mbmi->mode == DC_PRED) dc_skipped = 0;
+#endif  // CONFIG_FILTER_INTRA
+
+      uv_tx = uv_txsize_lookup[bsize][mbmi->tx_size][pd->subsampling_x]
+                              [pd->subsampling_y];
+      if (rate_uv_intra[uv_tx] == INT_MAX) {
+        choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx, &rate_uv_intra[uv_tx],
+                             &rate_uv_tokenonly[uv_tx], &dist_uvs[uv_tx],
+                             &skip_uvs[uv_tx], &mode_uv[uv_tx]);
+#if CONFIG_PALETTE
+        if (try_palette) pmi_uv[uv_tx] = *pmi;
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_EXT_INTRA
+        uv_angle_delta[uv_tx] = mbmi->angle_delta[1];
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+        filter_intra_mode_info_uv[uv_tx] = mbmi->filter_intra_mode_info;
+#endif  // CONFIG_FILTER_INTRA
+      }
+
+      rate_uv = rate_uv_tokenonly[uv_tx];
+      distortion_uv = dist_uvs[uv_tx];
+      skippable = skippable && skip_uvs[uv_tx];
+      mbmi->uv_mode = mode_uv[uv_tx];
+#if CONFIG_PALETTE
+      if (try_palette) {
+        pmi->palette_size[1] = pmi_uv[uv_tx].palette_size[1];
+        memcpy(pmi->palette_colors + PALETTE_MAX_SIZE,
+               pmi_uv[uv_tx].palette_colors + PALETTE_MAX_SIZE,
+               2 * PALETTE_MAX_SIZE * sizeof(pmi->palette_colors[0]));
+      }
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_EXT_INTRA
+      mbmi->angle_delta[1] = uv_angle_delta[uv_tx];
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+      mbmi->filter_intra_mode_info.use_filter_intra_mode[1] =
+          filter_intra_mode_info_uv[uv_tx].use_filter_intra_mode[1];
+      if (filter_intra_mode_info_uv[uv_tx].use_filter_intra_mode[1]) {
+        mbmi->filter_intra_mode_info.filter_intra_mode[1] =
+            filter_intra_mode_info_uv[uv_tx].filter_intra_mode[1];
+      }
+#endif  // CONFIG_FILTER_INTRA
+
+#if CONFIG_CB4X4
+      rate2 = rate_y + intra_mode_cost[mbmi->mode];
+      if (!x->skip_chroma_rd)
+        rate2 += rate_uv + cpi->intra_uv_mode_cost[mbmi->mode][mbmi->uv_mode];
+#else
+      rate2 = rate_y + intra_mode_cost[mbmi->mode] + rate_uv +
+              cpi->intra_uv_mode_cost[mbmi->mode][mbmi->uv_mode];
+#endif  // CONFIG_CB4X4
+
+#if CONFIG_PALETTE
+      if (try_palette && mbmi->mode == DC_PRED) {
+        rate2 += av1_cost_bit(
+            av1_default_palette_y_mode_prob[bsize - BLOCK_8X8][palette_ctx], 0);
+      }
+#endif  // CONFIG_PALETTE
+
+      if (!xd->lossless[mbmi->segment_id] && bsize >= BLOCK_8X8) {
+        // super_block_yrd above includes the cost of the tx_size in the
+        // tokenonly rate, but for intra blocks, tx_size is always coded
+        // (prediction granularity), so we account for it in the full rate,
+        // not the tokenonly rate.
+        rate_y -= tx_size_cost(cpi, x, bsize, mbmi->tx_size);
+      }
+#if CONFIG_EXT_INTRA
+      if (is_directional_mode) {
+#if CONFIG_INTRA_INTERP
+        const int intra_filter_ctx = av1_get_pred_context_intra_interp(xd);
+        const int p_angle =
+            mode_to_angle_map[mbmi->mode] + mbmi->angle_delta[0] * ANGLE_STEP;
+        if (av1_is_intra_filter_switchable(p_angle))
+          rate2 += cpi->intra_filter_cost[intra_filter_ctx][mbmi->intra_filter];
+#endif  // CONFIG_INTRA_INTERP
+        rate2 += write_uniform_cost(2 * MAX_ANGLE_DELTA + 1,
+                                    MAX_ANGLE_DELTA + mbmi->angle_delta[0]);
+      }
+      if (mbmi->uv_mode != DC_PRED && mbmi->uv_mode != TM_PRED) {
+        rate2 += write_uniform_cost(2 * MAX_ANGLE_DELTA + 1,
+                                    MAX_ANGLE_DELTA + mbmi->angle_delta[1]);
+      }
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+      if (mbmi->mode == DC_PRED) {
+        rate2 +=
+            av1_cost_bit(cm->fc->filter_intra_probs[0],
+                         mbmi->filter_intra_mode_info.use_filter_intra_mode[0]);
+        if (mbmi->filter_intra_mode_info.use_filter_intra_mode[0]) {
+          rate2 += write_uniform_cost(
+              FILTER_INTRA_MODES,
+              mbmi->filter_intra_mode_info.filter_intra_mode[0]);
+        }
+      }
+      if (mbmi->uv_mode == DC_PRED) {
+        rate2 +=
+            av1_cost_bit(cpi->common.fc->filter_intra_probs[1],
+                         mbmi->filter_intra_mode_info.use_filter_intra_mode[1]);
+        if (mbmi->filter_intra_mode_info.use_filter_intra_mode[1])
+          rate2 += write_uniform_cost(
+              FILTER_INTRA_MODES,
+              mbmi->filter_intra_mode_info.filter_intra_mode[1]);
+      }
+#endif  // CONFIG_FILTER_INTRA
+      if (mbmi->mode != DC_PRED && mbmi->mode != TM_PRED)
+        rate2 += intra_cost_penalty;
+      distortion2 = distortion_y + distortion_uv;
+    } else {
+#if CONFIG_REF_MV
+      int_mv backup_ref_mv[2];
+
+#if !SUB8X8_COMP_REF
+      if (bsize < BLOCK_8X8 && mbmi->ref_frame[1] > INTRA_FRAME) continue;
+#endif  // !SUB8X8_COMP_REF
+
+      backup_ref_mv[0] = mbmi_ext->ref_mvs[ref_frame][0];
+      if (comp_pred) backup_ref_mv[1] = mbmi_ext->ref_mvs[second_ref_frame][0];
+#endif  // CONFIG_REF_MV
+#if CONFIG_EXT_INTER
+      if (second_ref_frame == INTRA_FRAME) {
+        if (best_single_inter_ref != ref_frame) continue;
+        mbmi->interintra_mode = intra_to_interintra_mode[best_intra_mode];
+// TODO(debargha|geza.lore):
+// Should we use ext_intra modes for interintra?
+#if CONFIG_EXT_INTRA
+        mbmi->angle_delta[0] = 0;
+        mbmi->angle_delta[1] = 0;
+#if CONFIG_INTRA_INTERP
+        mbmi->intra_filter = INTRA_FILTER_LINEAR;
+#endif  // CONFIG_INTRA_INTERP
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+        mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+        mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+#endif  // CONFIG_FILTER_INTRA
+      }
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_REF_MV
+      mbmi->ref_mv_idx = 0;
+      ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
+
+#if CONFIG_EXT_INTER
+      if (comp_pred) {
+        if (mbmi_ext->ref_mv_count[ref_frame_type] > 1) {
+          int ref_mv_idx = 0;
+          // Special case: NEAR_NEWMV and NEW_NEARMV modes use
+          // 1 + mbmi->ref_mv_idx (like NEARMV) instead of
+          // mbmi->ref_mv_idx (like NEWMV)
+          if (mbmi->mode == NEAR_NEWMV || mbmi->mode == NEW_NEARMV)
+            ref_mv_idx = 1;
+
+          if (compound_ref0_mode(mbmi->mode) == NEWMV) {
+            int_mv this_mv =
+                mbmi_ext->ref_mv_stack[ref_frame_type][ref_mv_idx].this_mv;
+            clamp_mv_ref(&this_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                         xd->n8_h << MI_SIZE_LOG2, xd);
+            mbmi_ext->ref_mvs[mbmi->ref_frame[0]][0] = this_mv;
+          }
+          if (compound_ref1_mode(mbmi->mode) == NEWMV) {
+            int_mv this_mv =
+                mbmi_ext->ref_mv_stack[ref_frame_type][ref_mv_idx].comp_mv;
+            clamp_mv_ref(&this_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                         xd->n8_h << MI_SIZE_LOG2, xd);
+            mbmi_ext->ref_mvs[mbmi->ref_frame[1]][0] = this_mv;
+          }
+        }
+      } else {
+#endif  // CONFIG_EXT_INTER
+        if (mbmi->mode == NEWMV && mbmi_ext->ref_mv_count[ref_frame_type] > 1) {
+          int ref;
+          for (ref = 0; ref < 1 + comp_pred; ++ref) {
+            int_mv this_mv =
+                (ref == 0) ? mbmi_ext->ref_mv_stack[ref_frame_type][0].this_mv
+                           : mbmi_ext->ref_mv_stack[ref_frame_type][0].comp_mv;
+            clamp_mv_ref(&this_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                         xd->n8_h << MI_SIZE_LOG2, xd);
+            mbmi_ext->ref_mvs[mbmi->ref_frame[ref]][0] = this_mv;
+          }
+        }
+#if CONFIG_EXT_INTER
+      }
+#endif  // CONFIG_EXT_INTER
+#endif  // CONFIG_REF_MV
+      {
+        RD_STATS rd_stats, rd_stats_y, rd_stats_uv;
+        av1_init_rd_stats(&rd_stats);
+        rd_stats.rate = rate2;
+
+        // Point to variables that are maintained between loop iterations
+        args.single_newmv = single_newmv;
+#if CONFIG_EXT_INTER
+        args.single_newmv_rate = single_newmv_rate;
+        args.compmode_interintra_cost = &compmode_interintra_cost;
+        args.compmode_interinter_cost = &compmode_interinter_cost;
+        args.modelled_rd = modelled_rd;
+#endif  // CONFIG_EXT_INTER
+        this_rd = handle_inter_mode(cpi, x, bsize, &rd_stats, &rd_stats_y,
+                                    &rd_stats_uv, &disable_skip, frame_mv,
+                                    mi_row, mi_col, &args, best_rd);
+// Prevent pointers from escaping local scope
+#if CONFIG_EXT_INTER
+        args.compmode_interintra_cost = NULL;
+        args.compmode_interinter_cost = NULL;
+#endif  // CONFIG_EXT_INTER
+
+        rate2 = rd_stats.rate;
+        skippable = rd_stats.skip;
+        distortion2 = rd_stats.dist;
+        total_sse = rd_stats.sse;
+        rate_y = rd_stats_y.rate;
+        rate_uv = rd_stats_uv.rate;
+      }
+
+#if CONFIG_REF_MV
+// TODO(jingning): This needs some refactoring to improve code quality
+// and reduce redundant steps.
+#if CONFIG_EXT_INTER
+      if ((have_nearmv_in_inter_mode(mbmi->mode) &&
+           mbmi_ext->ref_mv_count[ref_frame_type] > 2) ||
+          ((mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV) &&
+           mbmi_ext->ref_mv_count[ref_frame_type] > 1)) {
+#else
+      if ((mbmi->mode == NEARMV &&
+           mbmi_ext->ref_mv_count[ref_frame_type] > 2) ||
+          (mbmi->mode == NEWMV && mbmi_ext->ref_mv_count[ref_frame_type] > 1)) {
+#endif
+        int_mv backup_mv = frame_mv[NEARMV][ref_frame];
+        MB_MODE_INFO backup_mbmi = *mbmi;
+        int backup_skip = x->skip;
+        int64_t tmp_ref_rd = this_rd;
+        int ref_idx;
+
+// TODO(jingning): This should be deprecated shortly.
+#if CONFIG_EXT_INTER
+        int idx_offset = have_nearmv_in_inter_mode(mbmi->mode) ? 1 : 0;
+#else
+        int idx_offset = (mbmi->mode == NEARMV) ? 1 : 0;
+#endif  // CONFIG_EXT_INTER
+        int ref_set =
+            AOMMIN(2, mbmi_ext->ref_mv_count[ref_frame_type] - 1 - idx_offset);
+
+        uint8_t drl_ctx =
+            av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx_offset);
+        // Dummy
+        int_mv backup_fmv[2];
+        backup_fmv[0] = frame_mv[NEWMV][ref_frame];
+        if (comp_pred) backup_fmv[1] = frame_mv[NEWMV][second_ref_frame];
+
+        rate2 += (rate2 < INT_MAX ? cpi->drl_mode_cost0[drl_ctx][0] : 0);
+
+        if (this_rd < INT64_MAX) {
+          if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
+              RDCOST(x->rdmult, x->rddiv, 0, total_sse))
+            tmp_ref_rd =
+                RDCOST(x->rdmult, x->rddiv,
+                       rate2 + av1_cost_bit(av1_get_skip_prob(cm, xd), 0),
+                       distortion2);
+          else
+            tmp_ref_rd =
+                RDCOST(x->rdmult, x->rddiv,
+                       rate2 + av1_cost_bit(av1_get_skip_prob(cm, xd), 1) -
+                           rate_y - rate_uv,
+                       total_sse);
+        }
+#if CONFIG_VAR_TX
+        for (i = 0; i < MAX_MB_PLANE; ++i)
+          memcpy(x->blk_skip_drl[i], x->blk_skip[i],
+                 sizeof(uint8_t) * ctx->num_4x4_blk);
+#endif  // CONFIG_VAR_TX
+
+        for (ref_idx = 0; ref_idx < ref_set; ++ref_idx) {
+          int64_t tmp_alt_rd = INT64_MAX;
+          int dummy_disable_skip = 0;
+          int ref;
+          int_mv cur_mv;
+          RD_STATS tmp_rd_stats, tmp_rd_stats_y, tmp_rd_stats_uv;
+#if CONFIG_EXT_INTER
+          int tmp_compmode_interintra_cost = 0;
+          int tmp_compmode_interinter_cost = 0;
+#endif  // CONFIG_EXT_INTER
+
+          av1_invalid_rd_stats(&tmp_rd_stats);
+          x->skip = 0;
+
+          mbmi->ref_mv_idx = 1 + ref_idx;
+
+#if CONFIG_EXT_INTER
+          if (comp_pred) {
+            int ref_mv_idx = mbmi->ref_mv_idx;
+            // Special case: NEAR_NEWMV and NEW_NEARMV modes use
+            // 1 + mbmi->ref_mv_idx (like NEARMV) instead of
+            // mbmi->ref_mv_idx (like NEWMV)
+            if (mbmi->mode == NEAR_NEWMV || mbmi->mode == NEW_NEARMV)
+              ref_mv_idx = 1 + mbmi->ref_mv_idx;
+
+            if (compound_ref0_mode(mbmi->mode) == NEWMV) {
+              int_mv this_mv =
+                  mbmi_ext->ref_mv_stack[ref_frame_type][ref_mv_idx].this_mv;
+              clamp_mv_ref(&this_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                           xd->n8_h << MI_SIZE_LOG2, xd);
+              mbmi_ext->ref_mvs[mbmi->ref_frame[0]][0] = this_mv;
+            } else if (compound_ref0_mode(mbmi->mode) == NEARESTMV) {
+              int_mv this_mv =
+                  mbmi_ext->ref_mv_stack[ref_frame_type][0].this_mv;
+              clamp_mv_ref(&this_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                           xd->n8_h << MI_SIZE_LOG2, xd);
+              mbmi_ext->ref_mvs[mbmi->ref_frame[0]][0] = this_mv;
+            }
+
+            if (compound_ref1_mode(mbmi->mode) == NEWMV) {
+              int_mv this_mv =
+                  mbmi_ext->ref_mv_stack[ref_frame_type][ref_mv_idx].comp_mv;
+              clamp_mv_ref(&this_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                           xd->n8_h << MI_SIZE_LOG2, xd);
+              mbmi_ext->ref_mvs[mbmi->ref_frame[1]][0] = this_mv;
+            } else if (compound_ref1_mode(mbmi->mode) == NEARESTMV) {
+              int_mv this_mv =
+                  mbmi_ext->ref_mv_stack[ref_frame_type][0].comp_mv;
+              clamp_mv_ref(&this_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                           xd->n8_h << MI_SIZE_LOG2, xd);
+              mbmi_ext->ref_mvs[mbmi->ref_frame[1]][0] = this_mv;
+            }
+          } else {
+#endif  // CONFIG_EXT_INTER
+            for (ref = 0; ref < 1 + comp_pred; ++ref) {
+              int_mv this_mv =
+                  (ref == 0)
+                      ? mbmi_ext->ref_mv_stack[ref_frame_type][mbmi->ref_mv_idx]
+                            .this_mv
+                      : mbmi_ext->ref_mv_stack[ref_frame_type][mbmi->ref_mv_idx]
+                            .comp_mv;
+              clamp_mv_ref(&this_mv.as_mv, xd->n8_w << MI_SIZE_LOG2,
+                           xd->n8_h << MI_SIZE_LOG2, xd);
+              mbmi_ext->ref_mvs[mbmi->ref_frame[ref]][0] = this_mv;
+            }
+#if CONFIG_EXT_INTER
+          }
+#endif
+
+          cur_mv =
+              mbmi_ext->ref_mv_stack[ref_frame][mbmi->ref_mv_idx + idx_offset]
+                  .this_mv;
+          clamp_mv2(&cur_mv.as_mv, xd);
+
+          if (!mv_check_bounds(&x->mv_limits, &cur_mv.as_mv)) {
+            int_mv dummy_single_newmv[TOTAL_REFS_PER_FRAME] = { { 0 } };
+#if CONFIG_EXT_INTER
+            int dummy_single_newmv_rate[TOTAL_REFS_PER_FRAME] = { 0 };
+#endif  // CONFIG_EXT_INTER
+
+            frame_mv[NEARMV][ref_frame] = cur_mv;
+            av1_init_rd_stats(&tmp_rd_stats);
+
+            // Point to variables that are not maintained between iterations
+            args.single_newmv = dummy_single_newmv;
+#if CONFIG_EXT_INTER
+            args.single_newmv_rate = dummy_single_newmv_rate;
+            args.compmode_interintra_cost = &tmp_compmode_interintra_cost;
+            args.compmode_interinter_cost = &tmp_compmode_interinter_cost;
+            args.modelled_rd = NULL;
+#endif  // CONFIG_EXT_INTER
+            tmp_alt_rd = handle_inter_mode(
+                cpi, x, bsize, &tmp_rd_stats, &tmp_rd_stats_y, &tmp_rd_stats_uv,
+                &dummy_disable_skip, frame_mv, mi_row, mi_col, &args, best_rd);
+            // Prevent pointers from escaping local scope
+            args.single_newmv = NULL;
+#if CONFIG_EXT_INTER
+            args.single_newmv_rate = NULL;
+            args.compmode_interintra_cost = NULL;
+            args.compmode_interinter_cost = NULL;
+#endif  // CONFIG_EXT_INTER
+          }
+
+          for (i = 0; i < mbmi->ref_mv_idx; ++i) {
+            uint8_t drl1_ctx = 0;
+            drl1_ctx = av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type],
+                                   i + idx_offset);
+            tmp_rd_stats.rate +=
+                (tmp_rd_stats.rate < INT_MAX ? cpi->drl_mode_cost0[drl1_ctx][1]
+                                             : 0);
+          }
+
+          if (mbmi_ext->ref_mv_count[ref_frame_type] >
+                  mbmi->ref_mv_idx + idx_offset + 1 &&
+              ref_idx < ref_set - 1) {
+            uint8_t drl1_ctx =
+                av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type],
+                            mbmi->ref_mv_idx + idx_offset);
+            tmp_rd_stats.rate +=
+                (tmp_rd_stats.rate < INT_MAX ? cpi->drl_mode_cost0[drl1_ctx][0]
+                                             : 0);
+          }
+
+          if (tmp_alt_rd < INT64_MAX) {
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+            tmp_alt_rd = RDCOST(x->rdmult, x->rddiv, tmp_rd_stats.rate,
+                                tmp_rd_stats.dist);
+#else
+            if (RDCOST(x->rdmult, x->rddiv,
+                       tmp_rd_stats_y.rate + tmp_rd_stats_uv.rate,
+                       tmp_rd_stats.dist) <
+                RDCOST(x->rdmult, x->rddiv, 0, tmp_rd_stats.sse))
+              tmp_alt_rd =
+                  RDCOST(x->rdmult, x->rddiv,
+                         tmp_rd_stats.rate +
+                             av1_cost_bit(av1_get_skip_prob(cm, xd), 0),
+                         tmp_rd_stats.dist);
+            else
+              tmp_alt_rd =
+                  RDCOST(x->rdmult, x->rddiv,
+                         tmp_rd_stats.rate +
+                             av1_cost_bit(av1_get_skip_prob(cm, xd), 1) -
+                             tmp_rd_stats_y.rate - tmp_rd_stats_uv.rate,
+                         tmp_rd_stats.sse);
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+          }
+
+          if (tmp_ref_rd > tmp_alt_rd) {
+            rate2 = tmp_rd_stats.rate;
+            disable_skip = dummy_disable_skip;
+            distortion2 = tmp_rd_stats.dist;
+            skippable = tmp_rd_stats.skip;
+            rate_y = tmp_rd_stats_y.rate;
+            rate_uv = tmp_rd_stats_uv.rate;
+            total_sse = tmp_rd_stats.sse;
+            this_rd = tmp_alt_rd;
+            tmp_ref_rd = tmp_alt_rd;
+            backup_mbmi = *mbmi;
+            backup_skip = x->skip;
+#if CONFIG_VAR_TX
+            for (i = 0; i < MAX_MB_PLANE; ++i)
+              memcpy(x->blk_skip_drl[i], x->blk_skip[i],
+                     sizeof(uint8_t) * ctx->num_4x4_blk);
+#endif  // CONFIG_VAR_TX
+#if CONFIG_EXT_INTER
+            compmode_interintra_cost = tmp_compmode_interintra_cost;
+            compmode_interinter_cost = tmp_compmode_interinter_cost;
+#endif  // CONFIG_EXT_INTER
+          } else {
+            *mbmi = backup_mbmi;
+            x->skip = backup_skip;
+          }
+        }
+
+        frame_mv[NEARMV][ref_frame] = backup_mv;
+        frame_mv[NEWMV][ref_frame] = backup_fmv[0];
+        if (comp_pred) frame_mv[NEWMV][second_ref_frame] = backup_fmv[1];
+#if CONFIG_VAR_TX
+        for (i = 0; i < MAX_MB_PLANE; ++i)
+          memcpy(x->blk_skip[i], x->blk_skip_drl[i],
+                 sizeof(uint8_t) * ctx->num_4x4_blk);
+#endif  // CONFIG_VAR_TX
+      }
+      mbmi_ext->ref_mvs[ref_frame][0] = backup_ref_mv[0];
+      if (comp_pred) mbmi_ext->ref_mvs[second_ref_frame][0] = backup_ref_mv[1];
+#endif  // CONFIG_REF_MV
+
+      if (this_rd == INT64_MAX) continue;
+
+#if SUB8X8_COMP_REF
+      compmode_cost = av1_cost_bit(comp_mode_p, comp_pred);
+#else
+      if (mbmi->sb_type >= BLOCK_8X8)
+        compmode_cost = av1_cost_bit(comp_mode_p, comp_pred);
+#endif  // SUB8X8_COMP_REF
+
+      if (cm->reference_mode == REFERENCE_MODE_SELECT) rate2 += compmode_cost;
+    }
+
+#if CONFIG_EXT_INTER
+    rate2 += compmode_interintra_cost;
+    if (cm->reference_mode != SINGLE_REFERENCE && comp_pred)
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      if (mbmi->motion_mode == SIMPLE_TRANSLATION)
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+        rate2 += compmode_interinter_cost;
+#endif  // CONFIG_EXT_INTER
+
+    // Estimate the reference frame signaling cost and add it
+    // to the rolling cost variable.
+    if (comp_pred) {
+      rate2 += ref_costs_comp[ref_frame];
+#if CONFIG_EXT_REFS
+      rate2 += ref_costs_comp[second_ref_frame];
+#endif  // CONFIG_EXT_REFS
+    } else {
+      rate2 += ref_costs_single[ref_frame];
+    }
+
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    if (ref_frame == INTRA_FRAME) {
+#else
+    if (!disable_skip) {
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+      if (skippable) {
+        // Back out the coefficient coding costs
+        rate2 -= (rate_y + rate_uv);
+        rate_y = 0;
+        rate_uv = 0;
+        // Cost the skip mb case
+        rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+      } else if (ref_frame != INTRA_FRAME && !xd->lossless[mbmi->segment_id]) {
+#if CONFIG_REF_MV
+        if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv + rate_skip0,
+                   distortion2) <
+            RDCOST(x->rdmult, x->rddiv, rate_skip1, total_sse)) {
+#else
+        if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
+            RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
+#endif  // CONFIG_REF_MV
+          // Add in the cost of the no skip flag.
+          rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+        } else {
+          // FIXME(rbultje) make this work for splitmv also
+          rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+          distortion2 = total_sse;
+          assert(total_sse >= 0);
+          rate2 -= (rate_y + rate_uv);
+          this_skip2 = 1;
+          rate_y = 0;
+          rate_uv = 0;
+        }
+      } else {
+        // Add in the cost of the no skip flag.
+        rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+      }
+
+      // Calculate the final RD estimate for this mode.
+      this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    } else {
+      this_skip2 = mbmi->skip;
+      this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+      if (this_skip2) {
+        rate_y = 0;
+        rate_uv = 0;
+      }
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+    }
+
+    if (ref_frame == INTRA_FRAME) {
+      // Keep record of best intra rd
+      if (this_rd < best_intra_rd) {
+        best_intra_rd = this_rd;
+        best_intra_mode = mbmi->mode;
+      }
+#if CONFIG_EXT_INTER
+    } else if (second_ref_frame == NONE_FRAME) {
+      if (this_rd < best_single_inter_rd) {
+        best_single_inter_rd = this_rd;
+        best_single_inter_ref = mbmi->ref_frame[0];
+      }
+#endif  // CONFIG_EXT_INTER
+    }
+
+    if (!disable_skip && ref_frame == INTRA_FRAME) {
+      for (i = 0; i < REFERENCE_MODES; ++i)
+        best_pred_rd[i] = AOMMIN(best_pred_rd[i], this_rd);
+    }
+
+    // Did this mode help.. i.e. is it the new best mode
+    if (this_rd < best_rd || x->skip) {
+      if (!mode_excluded) {
+        // Note index of best mode so far
+        best_mode_index = mode_index;
+
+        if (ref_frame == INTRA_FRAME) {
+          /* required for left and above block mv */
+          mbmi->mv[0].as_int = 0;
+        } else {
+          best_pred_sse = x->pred_sse[ref_frame];
+        }
+
+        rd_cost->rate = rate2;
+#if CONFIG_SUPERTX
+        if (x->skip)
+          *returnrate_nocoef = rate2;
+        else
+          *returnrate_nocoef = rate2 - rate_y - rate_uv;
+        *returnrate_nocoef -= av1_cost_bit(
+            av1_get_skip_prob(cm, xd), disable_skip || skippable || this_skip2);
+        *returnrate_nocoef -= av1_cost_bit(av1_get_intra_inter_prob(cm, xd),
+                                           mbmi->ref_frame[0] != INTRA_FRAME);
+#if CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+#if CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+        MODE_INFO *const mi = xd->mi[0];
+        const MOTION_MODE motion_allowed = motion_mode_allowed(
+#if CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+            0, xd->global_motion,
+#endif  // CONFIG_GLOBAL_MOTION && SEPARATE_GLOBAL_MOTION
+            mi);
+        if (motion_allowed == WARPED_CAUSAL)
+          *returnrate_nocoef -= cpi->motion_mode_cost[bsize][mbmi->motion_mode];
+        else if (motion_allowed == OBMC_CAUSAL)
+          *returnrate_nocoef -=
+              cpi->motion_mode_cost1[bsize][mbmi->motion_mode];
+#else
+        *returnrate_nocoef -= cpi->motion_mode_cost[bsize][mbmi->motion_mode];
+#endif  // CONFIG_MOTION_VAR && CONFIG_WARPED_MOTION
+#endif  // CONFIG_MOTION_VAR || CONFIG_WARPED_MOTION
+#endif  // CONFIG_SUPERTX
+        rd_cost->dist = distortion2;
+        rd_cost->rdcost = this_rd;
+        best_rd = this_rd;
+        best_mbmode = *mbmi;
+        best_skip2 = this_skip2;
+        best_mode_skippable = skippable;
+        best_rate_y = rate_y + av1_cost_bit(av1_get_skip_prob(cm, xd),
+                                            this_skip2 || skippable);
+        best_rate_uv = rate_uv;
+
+#if CONFIG_VAR_TX
+        for (i = 0; i < MAX_MB_PLANE; ++i)
+          memcpy(ctx->blk_skip[i], x->blk_skip[i],
+                 sizeof(uint8_t) * ctx->num_4x4_blk);
+#endif  // CONFIG_VAR_TX
+      }
+    }
+
+    /* keep record of best compound/single-only prediction */
+    if (!disable_skip && ref_frame != INTRA_FRAME) {
+      int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
+
+      if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+        single_rate = rate2 - compmode_cost;
+        hybrid_rate = rate2;
+      } else {
+        single_rate = rate2;
+        hybrid_rate = rate2 + compmode_cost;
+      }
+
+      single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
+      hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
+
+      if (!comp_pred) {
+        if (single_rd < best_pred_rd[SINGLE_REFERENCE])
+          best_pred_rd[SINGLE_REFERENCE] = single_rd;
+      } else {
+        if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
+          best_pred_rd[COMPOUND_REFERENCE] = single_rd;
+      }
+      if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
+        best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
+    }
+
+    if (x->skip && !comp_pred) break;
+  }
+
+  if (xd->lossless[mbmi->segment_id] == 0 && best_mode_index >= 0 &&
+      ((sf->tx_type_search.fast_inter_tx_type_search == 1 &&
+        is_inter_mode(best_mbmode.mode)) ||
+       (sf->tx_type_search.fast_intra_tx_type_search == 1 &&
+        !is_inter_mode(best_mbmode.mode)))) {
+    int skip_blk = 0;
+    RD_STATS rd_stats_y, rd_stats_uv;
+
+    x->use_default_inter_tx_type = 0;
+    x->use_default_intra_tx_type = 0;
+
+    *mbmi = best_mbmode;
+
+    set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
+
+    // Select prediction reference frames.
+    for (i = 0; i < MAX_MB_PLANE; i++) {
+      xd->plane[i].pre[0] = yv12_mb[mbmi->ref_frame[0]][i];
+      if (has_second_ref(mbmi))
+        xd->plane[i].pre[1] = yv12_mb[mbmi->ref_frame[1]][i];
+    }
+
+    if (is_inter_mode(mbmi->mode)) {
+      av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL, bsize);
+#if CONFIG_MOTION_VAR
+      if (mbmi->motion_mode == OBMC_CAUSAL) {
+        av1_build_obmc_inter_prediction(
+            cm, xd, mi_row, mi_col, args.above_pred_buf, args.above_pred_stride,
+            args.left_pred_buf, args.left_pred_stride);
+      }
+#endif  // CONFIG_MOTION_VAR
+      av1_subtract_plane(x, bsize, 0);
+#if CONFIG_VAR_TX
+      if (cm->tx_mode == TX_MODE_SELECT || xd->lossless[mbmi->segment_id]) {
+        select_tx_type_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
+      } else {
+        int idx, idy;
+        super_block_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
+        for (idy = 0; idy < xd->n8_h; ++idy)
+          for (idx = 0; idx < xd->n8_w; ++idx)
+            mbmi->inter_tx_size[idy][idx] = mbmi->tx_size;
+        memset(x->blk_skip[0], rd_stats_y.skip,
+               sizeof(uint8_t) * xd->n8_h * xd->n8_w * 4);
+      }
+
+      inter_block_uvrd(cpi, x, &rd_stats_uv, bsize, INT64_MAX);
+#else
+      super_block_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
+      super_block_uvrd(cpi, x, &rd_stats_uv, bsize, INT64_MAX);
+#endif  // CONFIG_VAR_TX
+    } else {
+      super_block_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
+      super_block_uvrd(cpi, x, &rd_stats_uv, bsize, INT64_MAX);
+    }
+
+    if (RDCOST(x->rdmult, x->rddiv, rd_stats_y.rate + rd_stats_uv.rate,
+               (rd_stats_y.dist + rd_stats_uv.dist)) >
+        RDCOST(x->rdmult, x->rddiv, 0, (rd_stats_y.sse + rd_stats_uv.sse))) {
+      skip_blk = 1;
+      rd_stats_y.rate = av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+      rd_stats_uv.rate = 0;
+      rd_stats_y.dist = rd_stats_y.sse;
+      rd_stats_uv.dist = rd_stats_uv.sse;
+    } else {
+      skip_blk = 0;
+      rd_stats_y.rate += av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+    }
+
+    if (RDCOST(x->rdmult, x->rddiv, best_rate_y + best_rate_uv, rd_cost->dist) >
+        RDCOST(x->rdmult, x->rddiv, rd_stats_y.rate + rd_stats_uv.rate,
+               (rd_stats_y.dist + rd_stats_uv.dist))) {
+#if CONFIG_VAR_TX
+      int idx, idy;
+#endif  // CONFIG_VAR_TX
+      best_mbmode.tx_type = mbmi->tx_type;
+      best_mbmode.tx_size = mbmi->tx_size;
+#if CONFIG_VAR_TX
+      for (idy = 0; idy < xd->n8_h; ++idy)
+        for (idx = 0; idx < xd->n8_w; ++idx)
+          best_mbmode.inter_tx_size[idy][idx] = mbmi->inter_tx_size[idy][idx];
+
+      for (i = 0; i < MAX_MB_PLANE; ++i)
+        memcpy(ctx->blk_skip[i], x->blk_skip[i],
+               sizeof(uint8_t) * ctx->num_4x4_blk);
+
+      best_mbmode.min_tx_size = mbmi->min_tx_size;
+#endif  // CONFIG_VAR_TX
+      rd_cost->rate +=
+          (rd_stats_y.rate + rd_stats_uv.rate - best_rate_y - best_rate_uv);
+      rd_cost->dist = rd_stats_y.dist + rd_stats_uv.dist;
+      rd_cost->rdcost =
+          RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
+      best_skip2 = skip_blk;
+    }
+  }
+
+#if CONFIG_PALETTE
+  // Only try palette mode when the best mode so far is an intra mode.
+  if (try_palette && !is_inter_mode(best_mbmode.mode)) {
+    int rate2 = 0;
+#if CONFIG_SUPERTX
+    int best_rate_nocoef;
+#endif  // CONFIG_SUPERTX
+    int64_t distortion2 = 0, best_rd_palette = best_rd, this_rd,
+            best_model_rd_palette = INT64_MAX;
+    int skippable = 0, rate_overhead_palette = 0;
+    RD_STATS rd_stats_y;
+    TX_SIZE uv_tx;
+    uint8_t *const best_palette_color_map =
+        x->palette_buffer->best_palette_color_map;
+    uint8_t *const color_map = xd->plane[0].color_index_map;
+    MB_MODE_INFO best_mbmi_palette = best_mbmode;
+
+    mbmi->mode = DC_PRED;
+    mbmi->uv_mode = DC_PRED;
+    mbmi->ref_frame[0] = INTRA_FRAME;
+    mbmi->ref_frame[1] = NONE_FRAME;
+    rate_overhead_palette = rd_pick_palette_intra_sby(
+        cpi, x, bsize, palette_ctx, intra_mode_cost[DC_PRED],
+        &best_mbmi_palette, best_palette_color_map, &best_rd_palette,
+        &best_model_rd_palette, NULL, NULL, NULL, NULL);
+    if (pmi->palette_size[0] == 0) goto PALETTE_EXIT;
+    memcpy(color_map, best_palette_color_map,
+           rows * cols * sizeof(best_palette_color_map[0]));
+    super_block_yrd(cpi, x, &rd_stats_y, bsize, best_rd);
+    if (rd_stats_y.rate == INT_MAX) goto PALETTE_EXIT;
+    uv_tx = uv_txsize_lookup[bsize][mbmi->tx_size][xd->plane[1].subsampling_x]
+                            [xd->plane[1].subsampling_y];
+    if (rate_uv_intra[uv_tx] == INT_MAX) {
+      choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx, &rate_uv_intra[uv_tx],
+                           &rate_uv_tokenonly[uv_tx], &dist_uvs[uv_tx],
+                           &skip_uvs[uv_tx], &mode_uv[uv_tx]);
+      pmi_uv[uv_tx] = *pmi;
+#if CONFIG_EXT_INTRA
+      uv_angle_delta[uv_tx] = mbmi->angle_delta[1];
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+      filter_intra_mode_info_uv[uv_tx] = mbmi->filter_intra_mode_info;
+#endif  // CONFIG_FILTER_INTRA
+    }
+    mbmi->uv_mode = mode_uv[uv_tx];
+    pmi->palette_size[1] = pmi_uv[uv_tx].palette_size[1];
+    if (pmi->palette_size[1] > 0) {
+      memcpy(pmi->palette_colors + PALETTE_MAX_SIZE,
+             pmi_uv[uv_tx].palette_colors + PALETTE_MAX_SIZE,
+             2 * PALETTE_MAX_SIZE * sizeof(pmi->palette_colors[0]));
+    }
+#if CONFIG_EXT_INTRA
+    mbmi->angle_delta[1] = uv_angle_delta[uv_tx];
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_FILTER_INTRA
+    mbmi->filter_intra_mode_info.use_filter_intra_mode[1] =
+        filter_intra_mode_info_uv[uv_tx].use_filter_intra_mode[1];
+    if (filter_intra_mode_info_uv[uv_tx].use_filter_intra_mode[1]) {
+      mbmi->filter_intra_mode_info.filter_intra_mode[1] =
+          filter_intra_mode_info_uv[uv_tx].filter_intra_mode[1];
+    }
+#endif  // CONFIG_FILTER_INTRA
+    skippable = rd_stats_y.skip && skip_uvs[uv_tx];
+    distortion2 = rd_stats_y.dist + dist_uvs[uv_tx];
+    rate2 = rd_stats_y.rate + rate_overhead_palette + rate_uv_intra[uv_tx];
+    rate2 += ref_costs_single[INTRA_FRAME];
+
+    if (skippable) {
+      rate2 -= (rd_stats_y.rate + rate_uv_tokenonly[uv_tx]);
+#if CONFIG_SUPERTX
+      best_rate_nocoef = rate2;
+#endif  // CONFIG_SUPERTX
+      rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+    } else {
+#if CONFIG_SUPERTX
+      best_rate_nocoef = rate2 - (rd_stats_y.rate + rate_uv_tokenonly[uv_tx]);
+#endif  // CONFIG_SUPERTX
+      rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+    }
+    this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+    if (this_rd < best_rd) {
+      best_mode_index = 3;
+      mbmi->mv[0].as_int = 0;
+      rd_cost->rate = rate2;
+#if CONFIG_SUPERTX
+      *returnrate_nocoef = best_rate_nocoef;
+#endif  // CONFIG_SUPERTX
+      rd_cost->dist = distortion2;
+      rd_cost->rdcost = this_rd;
+      best_rd = this_rd;
+      best_mbmode = *mbmi;
+      best_skip2 = 0;
+      best_mode_skippable = skippable;
+    }
+  }
+PALETTE_EXIT:
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_FILTER_INTRA
+  // TODO(huisu): filter-intra is turned off in lossless mode for now to
+  // avoid a unit test failure
+  if (!xd->lossless[mbmi->segment_id] &&
+#if CONFIG_PALETTE
+      pmi->palette_size[0] == 0 &&
+#endif  // CONFIG_PALETTE
+      !dc_skipped && best_mode_index >= 0 &&
+      best_intra_rd < (best_rd + (best_rd >> 3))) {
+    pick_filter_intra_interframe(
+        cpi, x, ctx, bsize, mi_row, mi_col, rate_uv_intra, rate_uv_tokenonly,
+        dist_uvs, skip_uvs, mode_uv, filter_intra_mode_info_uv,
+#if CONFIG_EXT_INTRA
+        uv_angle_delta,
+#endif  // CONFIG_EXT_INTRA
+#if CONFIG_PALETTE
+        pmi_uv, palette_ctx,
+#endif  // CONFIG_PALETTE
+        0, ref_costs_single, &best_rd, &best_intra_rd, &best_intra_mode,
+        &best_mode_index, &best_skip2, &best_mode_skippable,
+#if CONFIG_SUPERTX
+        returnrate_nocoef,
+#endif  // CONFIG_SUPERTX
+        best_pred_rd, &best_mbmode, rd_cost);
+  }
+#endif  // CONFIG_FILTER_INTRA
+
+  // The inter modes' rate costs are not calculated precisely in some cases.
+  // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
+  // ZEROMV. Here, checks are added for those cases, and the mode decisions
+  // are corrected.
+  if (best_mbmode.mode == NEWMV
+#if CONFIG_EXT_INTER
+      || best_mbmode.mode == NEW_NEWMV
+#endif  // CONFIG_EXT_INTER
+      ) {
+    const MV_REFERENCE_FRAME refs[2] = { best_mbmode.ref_frame[0],
+                                         best_mbmode.ref_frame[1] };
+    int comp_pred_mode = refs[1] > INTRA_FRAME;
+    int_mv zeromv[2];
+#if CONFIG_REF_MV
+    const uint8_t rf_type = av1_ref_frame_type(best_mbmode.ref_frame);
+#endif  // CONFIG_REF_MV
+#if CONFIG_GLOBAL_MOTION
+    zeromv[0].as_int = gm_get_motion_vector(&cm->global_motion[refs[0]],
+                                            cm->allow_high_precision_mv, bsize,
+                                            mi_col, mi_row, 0)
+                           .as_int;
+    zeromv[1].as_int = comp_pred_mode
+                           ? gm_get_motion_vector(&cm->global_motion[refs[1]],
+                                                  cm->allow_high_precision_mv,
+                                                  bsize, mi_col, mi_row, 0)
+                                 .as_int
+                           : 0;
+#else
+    zeromv[0].as_int = 0;
+    zeromv[1].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+#if CONFIG_REF_MV
+    if (!comp_pred_mode) {
+      int ref_set = (mbmi_ext->ref_mv_count[rf_type] >= 2)
+                        ? AOMMIN(2, mbmi_ext->ref_mv_count[rf_type] - 2)
+                        : INT_MAX;
+
+      for (i = 0; i <= ref_set && ref_set != INT_MAX; ++i) {
+        int_mv cur_mv = mbmi_ext->ref_mv_stack[rf_type][i + 1].this_mv;
+        if (cur_mv.as_int == best_mbmode.mv[0].as_int) {
+          best_mbmode.mode = NEARMV;
+          best_mbmode.ref_mv_idx = i;
+        }
+      }
+
+      if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int)
+        best_mbmode.mode = NEARESTMV;
+      else if (best_mbmode.mv[0].as_int == zeromv[0].as_int)
+        best_mbmode.mode = ZEROMV;
+    } else {
+      int_mv nearestmv[2];
+      int_mv nearmv[2];
+
+#if CONFIG_EXT_INTER
+      if (mbmi_ext->ref_mv_count[rf_type] > 1) {
+        nearmv[0] = mbmi_ext->ref_mv_stack[rf_type][1].this_mv;
+        nearmv[1] = mbmi_ext->ref_mv_stack[rf_type][1].comp_mv;
+      } else {
+        nearmv[0] = frame_mv[NEARMV][refs[0]];
+        nearmv[1] = frame_mv[NEARMV][refs[1]];
+      }
+#else
+      int ref_set = (mbmi_ext->ref_mv_count[rf_type] >= 2)
+                        ? AOMMIN(2, mbmi_ext->ref_mv_count[rf_type] - 2)
+                        : INT_MAX;
+
+      for (i = 0; i <= ref_set && ref_set != INT_MAX; ++i) {
+        nearmv[0] = mbmi_ext->ref_mv_stack[rf_type][i + 1].this_mv;
+        nearmv[1] = mbmi_ext->ref_mv_stack[rf_type][i + 1].comp_mv;
+
+        if (nearmv[0].as_int == best_mbmode.mv[0].as_int &&
+            nearmv[1].as_int == best_mbmode.mv[1].as_int) {
+          best_mbmode.mode = NEARMV;
+          best_mbmode.ref_mv_idx = i;
+        }
+      }
+#endif  // CONFIG_EXT_INTER
+      if (mbmi_ext->ref_mv_count[rf_type] >= 1) {
+        nearestmv[0] = mbmi_ext->ref_mv_stack[rf_type][0].this_mv;
+        nearestmv[1] = mbmi_ext->ref_mv_stack[rf_type][0].comp_mv;
+      } else {
+        nearestmv[0] = frame_mv[NEARESTMV][refs[0]];
+        nearestmv[1] = frame_mv[NEARESTMV][refs[1]];
+      }
+
+      if (nearestmv[0].as_int == best_mbmode.mv[0].as_int &&
+          nearestmv[1].as_int == best_mbmode.mv[1].as_int) {
+#if CONFIG_EXT_INTER
+        best_mbmode.mode = NEAREST_NEARESTMV;
+      } else {
+        int ref_set = (mbmi_ext->ref_mv_count[rf_type] >= 2)
+                          ? AOMMIN(2, mbmi_ext->ref_mv_count[rf_type] - 2)
+                          : INT_MAX;
+
+        for (i = 0; i <= ref_set && ref_set != INT_MAX; ++i) {
+          nearmv[0] = mbmi_ext->ref_mv_stack[rf_type][i + 1].this_mv;
+          nearmv[1] = mbmi_ext->ref_mv_stack[rf_type][i + 1].comp_mv;
+
+          // Try switching to the NEAR_NEAREST type modes first
+          if (nearestmv[0].as_int == best_mbmode.mv[0].as_int &&
+              nearmv[1].as_int == best_mbmode.mv[1].as_int) {
+            best_mbmode.mode = NEAREST_NEARMV;
+            best_mbmode.ref_mv_idx = i;
+          } else if (nearmv[0].as_int == best_mbmode.mv[0].as_int &&
+                     nearestmv[1].as_int == best_mbmode.mv[1].as_int) {
+            best_mbmode.mode = NEAR_NEARESTMV;
+            best_mbmode.ref_mv_idx = i;
+          } else if (nearmv[0].as_int == best_mbmode.mv[0].as_int &&
+                     nearmv[1].as_int == best_mbmode.mv[1].as_int) {
+            best_mbmode.mode = NEAR_NEARMV;
+            best_mbmode.ref_mv_idx = i;
+          }
+        }
+
+        if (best_mbmode.mode == NEW_NEWMV &&
+            best_mbmode.mv[0].as_int == zeromv[0].as_int &&
+            best_mbmode.mv[1].as_int == zeromv[1].as_int)
+          best_mbmode.mode = ZERO_ZEROMV;
+      }
+#else
+        best_mbmode.mode = NEARESTMV;
+      } else if (best_mbmode.mv[0].as_int == zeromv[0].as_int &&
+                 best_mbmode.mv[1].as_int == zeromv[1].as_int) {
+        best_mbmode.mode = ZEROMV;
+      }
+#endif  // CONFIG_EXT_INTER
+    }
+#else
+#if CONFIG_EXT_INTER
+    if (!comp_pred_mode) {
+#endif  // CONFIG_EXT_INTER
+      if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
+          ((comp_pred_mode &&
+            frame_mv[NEARESTMV][refs[1]].as_int == best_mbmode.mv[1].as_int) ||
+           !comp_pred_mode))
+        best_mbmode.mode = NEARESTMV;
+      else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
+               ((comp_pred_mode &&
+                 frame_mv[NEARMV][refs[1]].as_int ==
+                     best_mbmode.mv[1].as_int) ||
+                !comp_pred_mode))
+        best_mbmode.mode = NEARMV;
+      else if (best_mbmode.mv[0].as_int == zeromv[0].as_int &&
+               ((comp_pred_mode &&
+                 best_mbmode.mv[1].as_int == zeromv[1].as_int) ||
+                !comp_pred_mode))
+        best_mbmode.mode = ZEROMV;
+#if CONFIG_EXT_INTER
+    } else {
+#if CONFIG_GLOBAL_MOTION
+      zeromv[0].as_int = gm_get_motion_vector(&cm->global_motion[refs[0]],
+                                              cm->allow_high_precision_mv,
+                                              bsize, mi_col, mi_row, 0)
+                             .as_int;
+      zeromv[1].as_int = comp_pred_mode
+                             ? gm_get_motion_vector(&cm->global_motion[refs[1]],
+                                                    cm->allow_high_precision_mv,
+                                                    bsize, mi_col, mi_row, 0)
+                                   .as_int
+                             : 0;
+#else
+      zeromv[0].as_int = 0;
+      zeromv[1].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+      if (frame_mv[NEAREST_NEARESTMV][refs[0]].as_int ==
+              best_mbmode.mv[0].as_int &&
+          frame_mv[NEAREST_NEARESTMV][refs[1]].as_int ==
+              best_mbmode.mv[1].as_int)
+        best_mbmode.mode = NEAREST_NEARESTMV;
+      else if (frame_mv[NEAREST_NEARMV][refs[0]].as_int ==
+                   best_mbmode.mv[0].as_int &&
+               frame_mv[NEAREST_NEARMV][refs[1]].as_int ==
+                   best_mbmode.mv[1].as_int)
+        best_mbmode.mode = NEAREST_NEARMV;
+      else if (frame_mv[NEAR_NEARESTMV][refs[0]].as_int ==
+                   best_mbmode.mv[0].as_int &&
+               frame_mv[NEAR_NEARESTMV][refs[1]].as_int ==
+                   best_mbmode.mv[1].as_int)
+        best_mbmode.mode = NEAR_NEARESTMV;
+      else if (frame_mv[NEAR_NEARMV][refs[0]].as_int ==
+                   best_mbmode.mv[0].as_int &&
+               frame_mv[NEAR_NEARMV][refs[1]].as_int ==
+                   best_mbmode.mv[1].as_int)
+        best_mbmode.mode = NEAR_NEARMV;
+      else if (best_mbmode.mv[0].as_int == zeromv[0].as_int &&
+               best_mbmode.mv[1].as_int == zeromv[1].as_int)
+        best_mbmode.mode = ZERO_ZEROMV;
+    }
+#endif  // CONFIG_EXT_INTER
+#endif  // CONFIG_REF_MV
+  }
+
+#if CONFIG_REF_MV
+  // Make sure that the ref_mv_idx is only nonzero when we're
+  // using a mode which can support ref_mv_idx
+  if (best_mbmode.ref_mv_idx != 0 &&
+#if CONFIG_EXT_INTER
+      !(best_mbmode.mode == NEWMV || best_mbmode.mode == NEW_NEWMV ||
+        have_nearmv_in_inter_mode(best_mbmode.mode))) {
+#else
+      !(best_mbmode.mode == NEARMV || best_mbmode.mode == NEWMV)) {
+#endif
+    best_mbmode.ref_mv_idx = 0;
+  }
+
+  {
+    int8_t ref_frame_type = av1_ref_frame_type(best_mbmode.ref_frame);
+    int16_t mode_ctx = mbmi_ext->mode_context[ref_frame_type];
+    if (mode_ctx & (1 << ALL_ZERO_FLAG_OFFSET)) {
+      int_mv zeromv[2];
+#if CONFIG_GLOBAL_MOTION
+      const MV_REFERENCE_FRAME refs[2] = { best_mbmode.ref_frame[0],
+                                           best_mbmode.ref_frame[1] };
+      zeromv[0].as_int = gm_get_motion_vector(&cm->global_motion[refs[0]],
+                                              cm->allow_high_precision_mv,
+                                              bsize, mi_col, mi_row, 0)
+                             .as_int;
+      zeromv[1].as_int = (refs[1] != NONE_FRAME)
+                             ? gm_get_motion_vector(&cm->global_motion[refs[1]],
+                                                    cm->allow_high_precision_mv,
+                                                    bsize, mi_col, mi_row, 0)
+                                   .as_int
+                             : 0;
+      lower_mv_precision(&zeromv[0].as_mv, cm->allow_high_precision_mv);
+      lower_mv_precision(&zeromv[1].as_mv, cm->allow_high_precision_mv);
+#else
+      zeromv[0].as_int = zeromv[1].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+      if (best_mbmode.ref_frame[0] > INTRA_FRAME &&
+          best_mbmode.mv[0].as_int == zeromv[0].as_int &&
+#if CONFIG_EXT_INTER
+          (best_mbmode.ref_frame[1] <= INTRA_FRAME)
+#else
+          (best_mbmode.ref_frame[1] == NONE_FRAME ||
+           best_mbmode.mv[1].as_int == zeromv[1].as_int)
+#endif  // CONFIG_EXT_INTER
+              ) {
+        best_mbmode.mode = ZEROMV;
+      }
+    }
+  }
+#endif  // CONFIG_REF_MV
+
+  if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
+    rd_cost->rate = INT_MAX;
+    rd_cost->rdcost = INT64_MAX;
+    return;
+  }
+
+#if CONFIG_DUAL_FILTER
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == best_mbmode.interp_filter[0]) ||
+         !is_inter_block(&best_mbmode));
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == best_mbmode.interp_filter[1]) ||
+         !is_inter_block(&best_mbmode));
+  if (best_mbmode.ref_frame[1] > INTRA_FRAME) {
+    assert((cm->interp_filter == SWITCHABLE) ||
+           (cm->interp_filter == best_mbmode.interp_filter[2]) ||
+           !is_inter_block(&best_mbmode));
+    assert((cm->interp_filter == SWITCHABLE) ||
+           (cm->interp_filter == best_mbmode.interp_filter[3]) ||
+           !is_inter_block(&best_mbmode));
+  }
+#else
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == best_mbmode.interp_filter) ||
+         !is_inter_block(&best_mbmode));
+#endif  // CONFIG_DUAL_FILTER
+
+  if (!cpi->rc.is_src_frame_alt_ref)
+    av1_update_rd_thresh_fact(cm, tile_data->thresh_freq_fact,
+                              sf->adaptive_rd_thresh, bsize, best_mode_index);
+
+  // macroblock modes
+  *mbmi = best_mbmode;
+  x->skip |= best_skip2;
+
+// Note: this section is needed since the mode may have been forced to
+// ZEROMV by the all-zero mode handling of ref-mv.
+#if CONFIG_GLOBAL_MOTION
+  if (mbmi->mode == ZEROMV
+#if CONFIG_EXT_INTER
+      || mbmi->mode == ZERO_ZEROMV
+#endif  // CONFIG_EXT_INTER
+      ) {
+#if CONFIG_WARPED_MOTION || CONFIG_MOTION_VAR
+    // Correct the motion mode for ZEROMV
+    const MOTION_MODE last_motion_mode_allowed = motion_mode_allowed(
+#if SEPARATE_GLOBAL_MOTION
+        0, xd->global_motion,
+#endif  // SEPARATE_GLOBAL_MOTION
+        xd->mi[0]);
+    if (mbmi->motion_mode > last_motion_mode_allowed)
+      mbmi->motion_mode = last_motion_mode_allowed;
+#endif  // CONFIG_WARPED_MOTION || CONFIG_MOTION_VAR
+
+    // Correct the interpolation filter for ZEROMV
+    if (is_nontrans_global_motion(xd)) {
+#if CONFIG_DUAL_FILTER
+      mbmi->interp_filter[0] = cm->interp_filter == SWITCHABLE
+                                   ? EIGHTTAP_REGULAR
+                                   : cm->interp_filter;
+      mbmi->interp_filter[1] = cm->interp_filter == SWITCHABLE
+                                   ? EIGHTTAP_REGULAR
+                                   : cm->interp_filter;
+#else
+      mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP_REGULAR
+                                                            : cm->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+    }
+  }
+#endif  // CONFIG_GLOBAL_MOTION
+
+#if CONFIG_REF_MV
+  for (i = 0; i < 1 + has_second_ref(mbmi); ++i) {
+    if (mbmi->mode != NEWMV)
+      mbmi->pred_mv[i].as_int = mbmi->mv[i].as_int;
+    else
+      mbmi->pred_mv[i].as_int = mbmi_ext->ref_mvs[mbmi->ref_frame[i]][0].as_int;
+  }
+#endif  // CONFIG_REF_MV
+
+  for (i = 0; i < REFERENCE_MODES; ++i) {
+    if (best_pred_rd[i] == INT64_MAX)
+      best_pred_diff[i] = INT_MIN;
+    else
+      best_pred_diff[i] = best_rd - best_pred_rd[i];
+  }
+
+  x->skip |= best_mode_skippable;
+
+  assert(best_mode_index >= 0);
+
+  store_coding_context(x, ctx, best_mode_index, best_pred_diff,
+                       best_mode_skippable);
+
+#if CONFIG_PALETTE
+  if (cm->allow_screen_content_tools && pmi->palette_size[1] > 0) {
+    restore_uv_color_map(cpi, x);
+  }
+#endif  // CONFIG_PALETTE
+}
+
+void av1_rd_pick_inter_mode_sb_seg_skip(const AV1_COMP *cpi,
+                                        TileDataEnc *tile_data, MACROBLOCK *x,
+                                        int mi_row, int mi_col,
+                                        RD_STATS *rd_cost, BLOCK_SIZE bsize,
+                                        PICK_MODE_CONTEXT *ctx,
+                                        int64_t best_rd_so_far) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  unsigned char segment_id = mbmi->segment_id;
+  const int comp_pred = 0;
+  int i;
+  int64_t best_pred_diff[REFERENCE_MODES];
+  unsigned int ref_costs_single[TOTAL_REFS_PER_FRAME];
+  unsigned int ref_costs_comp[TOTAL_REFS_PER_FRAME];
+  aom_prob comp_mode_p;
+  InterpFilter best_filter = SWITCHABLE;
+  int64_t this_rd = INT64_MAX;
+  int rate2 = 0;
+  const int64_t distortion2 = 0;
+  (void)mi_row;
+  (void)mi_col;
+
+  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
+                           &comp_mode_p);
+
+  for (i = 0; i < TOTAL_REFS_PER_FRAME; ++i) x->pred_sse[i] = INT_MAX;
+  for (i = LAST_FRAME; i < TOTAL_REFS_PER_FRAME; ++i)
+    x->pred_mv_sad[i] = INT_MAX;
+
+  rd_cost->rate = INT_MAX;
+
+  assert(segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
+
+#if CONFIG_PALETTE
+  mbmi->palette_mode_info.palette_size[0] = 0;
+  mbmi->palette_mode_info.palette_size[1] = 0;
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+#endif  // CONFIG_FILTER_INTRA
+  mbmi->mode = ZEROMV;
+  mbmi->motion_mode = SIMPLE_TRANSLATION;
+  mbmi->uv_mode = DC_PRED;
+  mbmi->ref_frame[0] = LAST_FRAME;
+  mbmi->ref_frame[1] = NONE_FRAME;
+#if CONFIG_GLOBAL_MOTION
+  mbmi->mv[0].as_int =
+      gm_get_motion_vector(&cm->global_motion[mbmi->ref_frame[0]],
+                           cm->allow_high_precision_mv, bsize, mi_col, mi_row,
+                           0)
+          .as_int;
+#else   // CONFIG_GLOBAL_MOTION
+  mbmi->mv[0].as_int = 0;
+#endif  // CONFIG_GLOBAL_MOTION
+  mbmi->tx_size = max_txsize_lookup[bsize];
+  x->skip = 1;
+
+#if CONFIG_REF_MV
+  mbmi->ref_mv_idx = 0;
+  mbmi->pred_mv[0].as_int = 0;
+#endif  // CONFIG_REF_MV
+
+  mbmi->motion_mode = SIMPLE_TRANSLATION;
+#if CONFIG_MOTION_VAR
+  av1_count_overlappable_neighbors(cm, xd, mi_row, mi_col);
+#endif
+#if CONFIG_WARPED_MOTION
+  if (is_motion_variation_allowed_bsize(bsize) && !has_second_ref(mbmi)) {
+    int pts[SAMPLES_ARRAY_SIZE], pts_inref[SAMPLES_ARRAY_SIZE];
+    mbmi->num_proj_ref[0] = findSamples(cm, xd, mi_row, mi_col, pts, pts_inref);
+  }
+#endif
+
+  set_default_interp_filters(mbmi, cm->interp_filter);
+
+  if (cm->interp_filter != SWITCHABLE) {
+    best_filter = cm->interp_filter;
+  } else {
+    best_filter = EIGHTTAP_REGULAR;
+    if (av1_is_interp_needed(xd) && av1_is_interp_search_needed(xd) &&
+        x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
+      int rs;
+      int best_rs = INT_MAX;
+      for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
+#if CONFIG_DUAL_FILTER
+        int k;
+        for (k = 0; k < 4; ++k) mbmi->interp_filter[k] = i;
+#else
+        mbmi->interp_filter = i;
+#endif  // CONFIG_DUAL_FILTER
+        rs = av1_get_switchable_rate(cpi, xd);
+        if (rs < best_rs) {
+          best_rs = rs;
+#if CONFIG_DUAL_FILTER
+          best_filter = mbmi->interp_filter[0];
+#else
+          best_filter = mbmi->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+        }
+      }
+    }
+  }
+// Set the appropriate filter
+#if CONFIG_DUAL_FILTER
+  for (i = 0; i < 4; ++i) mbmi->interp_filter[i] = best_filter;
+#else
+  mbmi->interp_filter = best_filter;
+#endif  // CONFIG_DUAL_FILTER
+  rate2 += av1_get_switchable_rate(cpi, xd);
+
+  if (cm->reference_mode == REFERENCE_MODE_SELECT)
+    rate2 += av1_cost_bit(comp_mode_p, comp_pred);
+
+  // Estimate the reference frame signaling cost and add it
+  // to the rolling cost variable.
+  rate2 += ref_costs_single[LAST_FRAME];
+  this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+
+  rd_cost->rate = rate2;
+  rd_cost->dist = distortion2;
+  rd_cost->rdcost = this_rd;
+
+  if (this_rd >= best_rd_so_far) {
+    rd_cost->rate = INT_MAX;
+    rd_cost->rdcost = INT64_MAX;
+    return;
+  }
+
+#if CONFIG_DUAL_FILTER
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == mbmi->interp_filter[0]));
+#else
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == mbmi->interp_filter));
+#endif  // CONFIG_DUAL_FILTER
+
+  av1_update_rd_thresh_fact(cm, tile_data->thresh_freq_fact,
+                            cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
+
+  av1_zero(best_pred_diff);
+
+  store_coding_context(x, ctx, THR_ZEROMV, best_pred_diff, 0);
+}
+
+void av1_rd_pick_inter_mode_sub8x8(const struct AV1_COMP *cpi,
+                                   TileDataEnc *tile_data, struct macroblock *x,
+                                   int mi_row, int mi_col,
+                                   struct RD_STATS *rd_cost,
+#if CONFIG_SUPERTX
+                                   int *returnrate_nocoef,
+#endif  // CONFIG_SUPERTX
+                                   BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                                   int64_t best_rd_so_far) {
+  const AV1_COMMON *const cm = &cpi->common;
+  const RD_OPT *const rd_opt = &cpi->rd;
+  const SPEED_FEATURES *const sf = &cpi->sf;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const struct segmentation *const seg = &cm->seg;
+  MV_REFERENCE_FRAME ref_frame, second_ref_frame;
+  unsigned char segment_id = mbmi->segment_id;
+  int comp_pred, i;
+  int_mv frame_mv[MB_MODE_COUNT][TOTAL_REFS_PER_FRAME];
+  struct buf_2d yv12_mb[TOTAL_REFS_PER_FRAME][MAX_MB_PLANE];
+  static const int flag_list[TOTAL_REFS_PER_FRAME] = {
+    0,
+    AOM_LAST_FLAG,
+#if CONFIG_EXT_REFS
+    AOM_LAST2_FLAG,
+    AOM_LAST3_FLAG,
+#endif  // CONFIG_EXT_REFS
+    AOM_GOLD_FLAG,
+#if CONFIG_EXT_REFS
+    AOM_BWD_FLAG,
+#endif  // CONFIG_EXT_REFS
+    AOM_ALT_FLAG
+  };
+  int64_t best_rd = best_rd_so_far;
+  int64_t best_yrd = best_rd_so_far;  // FIXME(rbultje) more precise
+  int64_t best_pred_diff[REFERENCE_MODES];
+  int64_t best_pred_rd[REFERENCE_MODES];
+  MB_MODE_INFO best_mbmode;
+  int ref_index, best_ref_index = 0;
+  unsigned int ref_costs_single[TOTAL_REFS_PER_FRAME];
+  unsigned int ref_costs_comp[TOTAL_REFS_PER_FRAME];
+  aom_prob comp_mode_p;
+#if CONFIG_DUAL_FILTER
+  InterpFilter tmp_best_filter[4] = { 0 };
+#else
+  InterpFilter tmp_best_filter = SWITCHABLE;
+#endif  // CONFIG_DUAL_FILTER
+  int rate_uv_intra, rate_uv_tokenonly = INT_MAX;
+  int64_t dist_uv = INT64_MAX;
+  int skip_uv;
+  PREDICTION_MODE mode_uv = DC_PRED;
+  const int intra_cost_penalty = av1_get_intra_cost_penalty(
+      cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
+  int_mv seg_mvs[4][TOTAL_REFS_PER_FRAME];
+  b_mode_info best_bmodes[4];
+  int best_skip2 = 0;
+  int ref_frame_skip_mask[2] = { 0 };
+  int internal_active_edge =
+      av1_active_edge_sb(cpi, mi_row, mi_col) && av1_internal_image_edge(cpi);
+#if CONFIG_PVQ
+  od_rollback_buffer pre_buf;
+
+  od_encode_checkpoint(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+
+#if CONFIG_SUPERTX
+  best_rd_so_far = INT64_MAX;
+  best_rd = best_rd_so_far;
+  best_yrd = best_rd_so_far;
+#endif  // CONFIG_SUPERTX
+  av1_zero(best_mbmode);
+
+#if CONFIG_FILTER_INTRA
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[0] = 0;
+  mbmi->filter_intra_mode_info.use_filter_intra_mode[1] = 0;
+#endif  // CONFIG_FILTER_INTRA
+  mbmi->motion_mode = SIMPLE_TRANSLATION;
+#if CONFIG_EXT_INTER
+  mbmi->interinter_compound_type = COMPOUND_AVERAGE;
+  mbmi->use_wedge_interintra = 0;
+#endif  // CONFIG_EXT_INTER
+#if CONFIG_WARPED_MOTION
+  mbmi->num_proj_ref[0] = 0;
+  mbmi->num_proj_ref[1] = 0;
+#endif  // CONFIG_WARPED_MOTION
+
+  for (i = 0; i < 4; i++) {
+    int j;
+    for (j = 0; j < TOTAL_REFS_PER_FRAME; j++)
+      seg_mvs[i][j].as_int = INVALID_MV;
+  }
+
+  estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
+                           &comp_mode_p);
+
+  for (i = 0; i < REFERENCE_MODES; ++i) best_pred_rd[i] = INT64_MAX;
+  rate_uv_intra = INT_MAX;
+
+  rd_cost->rate = INT_MAX;
+#if CONFIG_SUPERTX
+  *returnrate_nocoef = INT_MAX;
+#endif  // CONFIG_SUPERTX
+
+  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
+    x->mbmi_ext->mode_context[ref_frame] = 0;
+#if CONFIG_REF_MV && CONFIG_EXT_INTER
+    x->mbmi_ext->compound_mode_context[ref_frame] = 0;
+#endif  // CONFIG_REF_MV && CONFIG_EXT_INTER
+    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
+      setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
+                         frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
+    } else {
+      ref_frame_skip_mask[0] |= (1 << ref_frame);
+      ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+    }
+    frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
+#if CONFIG_EXT_INTER
+#endif  // CONFIG_EXT_INTER
+    frame_mv[ZEROMV][ref_frame].as_int = 0;
+  }
+
+#if CONFIG_PALETTE
+  mbmi->palette_mode_info.palette_size[0] = 0;
+  mbmi->palette_mode_info.palette_size[1] = 0;
+#endif  // CONFIG_PALETTE
+
+  for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
+    int mode_excluded = 0;
+    int64_t this_rd = INT64_MAX;
+    int disable_skip = 0;
+    int compmode_cost = 0;
+    int rate2 = 0, rate_y = 0, rate_uv = 0;
+    int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
+    int skippable = 0;
+    int this_skip2 = 0;
+    int64_t total_sse = INT_MAX;
+
+#if CONFIG_PVQ
+    od_encode_rollback(&x->daala_enc, &pre_buf);
+#endif  // CONFIG_PVQ
+
+    ref_frame = av1_ref_order[ref_index].ref_frame[0];
+    second_ref_frame = av1_ref_order[ref_index].ref_frame[1];
+
+#if CONFIG_REF_MV
+    mbmi->ref_mv_idx = 0;
+#endif  // CONFIG_REF_MV
+
+    // Look at the reference frame of the best mode so far and set the
+    // skip mask to look at a subset of the remaining modes.
+    if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
+      if (ref_index == 3) {
+        switch (best_mbmode.ref_frame[0]) {
+          case INTRA_FRAME: break;
+          case LAST_FRAME:
+            ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) |
+#if CONFIG_EXT_REFS
+                                      (1 << LAST2_FRAME) | (1 << LAST3_FRAME) |
+                                      (1 << BWDREF_FRAME) |
+#endif  // CONFIG_EXT_REFS
+                                      (1 << ALTREF_FRAME);
+            ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+            break;
+#if CONFIG_EXT_REFS
+          case LAST2_FRAME:
+            ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << LAST3_FRAME) |
+                                      (1 << GOLDEN_FRAME) |
+                                      (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME);
+            ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+            break;
+          case LAST3_FRAME:
+            ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << LAST2_FRAME) |
+                                      (1 << GOLDEN_FRAME) |
+                                      (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME);
+            ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+            break;
+#endif  // CONFIG_EXT_REFS
+          case GOLDEN_FRAME:
+            ref_frame_skip_mask[0] |= (1 << LAST_FRAME) |
+#if CONFIG_EXT_REFS
+                                      (1 << LAST2_FRAME) | (1 << LAST3_FRAME) |
+                                      (1 << BWDREF_FRAME) |
+#endif  // CONFIG_EXT_REFS
+                                      (1 << ALTREF_FRAME);
+            ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
+            break;
+#if CONFIG_EXT_REFS
+          case BWDREF_FRAME:
+            ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << LAST2_FRAME) |
+                                      (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) |
+                                      (1 << ALTREF_FRAME);
+            ref_frame_skip_mask[1] |= (1 << ALTREF_FRAME) | 0x01;
+            break;
+#endif  // CONFIG_EXT_REFS
+          case ALTREF_FRAME:
+            ref_frame_skip_mask[0] |= (1 << LAST_FRAME) |
+#if CONFIG_EXT_REFS
+                                      (1 << LAST2_FRAME) | (1 << LAST3_FRAME) |
+                                      (1 << BWDREF_FRAME) |
+#endif  // CONFIG_EXT_REFS
+                                      (1 << GOLDEN_FRAME);
+#if CONFIG_EXT_REFS
+            ref_frame_skip_mask[1] |= (1 << BWDREF_FRAME) | 0x01;
+#endif  // CONFIG_EXT_REFS
+            break;
+          case NONE_FRAME:
+          case TOTAL_REFS_PER_FRAME:
+            assert(0 && "Invalid Reference frame");
+            break;
+        }
+      }
+    }
+
+    if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
+        (ref_frame_skip_mask[1] & (1 << AOMMAX(0, second_ref_frame))))
+      continue;
+
+    // Test best rd so far against threshold for trying this mode.
+    if (!internal_active_edge &&
+        rd_less_than_thresh(best_rd,
+                            rd_opt->threshes[segment_id][bsize][ref_index],
+                            tile_data->thresh_freq_fact[bsize][ref_index]))
+      continue;
+
+    // This is only used in motion vector unit test.
+    if (cpi->oxcf.motion_vector_unit_test && ref_frame == INTRA_FRAME) continue;
+
+#if CONFIG_LOWDELAY_COMPOUND  // Changes LL bitstream
+#if CONFIG_EXT_REFS
+    if (cpi->oxcf.pass == 0) {
+      // Complexity-compression trade-offs
+      // if (ref_frame == ALTREF_FRAME) continue;
+      // if (ref_frame == BWDREF_FRAME) continue;
+      if (second_ref_frame == ALTREF_FRAME) continue;
+      // if (second_ref_frame == BWDREF_FRAME) continue;
+    }
+#endif
+#endif
+    comp_pred = second_ref_frame > INTRA_FRAME;
+    if (comp_pred) {
+      if (!cpi->allow_comp_inter_inter) continue;
+      if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
+      // Do not allow compound prediction if the segment level reference frame
+      // feature is in use as in this case there can only be one reference.
+      if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) continue;
+
+      if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
+          best_mbmode.ref_frame[0] == INTRA_FRAME)
+        continue;
+    }
+
+    // TODO(jingning, jkoleszar): scaling reference frame not supported for
+    // sub8x8 blocks.
+    if (ref_frame > INTRA_FRAME &&
+        av1_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
+      continue;
+
+    if (second_ref_frame > INTRA_FRAME &&
+        av1_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf))
+      continue;
+
+    if (comp_pred)
+      mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
+    else if (ref_frame != INTRA_FRAME)
+      mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
+
+    // If the segment reference frame feature is enabled....
+    // then do nothing if the current ref frame is not allowed..
+    if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
+        get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
+      continue;
+      // Disable this drop out case if the ref frame
+      // segment level feature is enabled for this segment. This is to
+      // prevent the possibility that we end up unable to pick any mode.
+    } else if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
+      // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
+      // unless ARNR filtering is enabled in which case we want
+      // an unfiltered alternative. We allow near/nearest as well
+      // because they may result in zero-zero MVs but be cheaper.
+      if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
+        continue;
+    }
+
+    mbmi->tx_size = TX_4X4;
+    mbmi->uv_mode = DC_PRED;
+    mbmi->ref_frame[0] = ref_frame;
+    mbmi->ref_frame[1] = second_ref_frame;
+// Evaluate all sub-pel filters irrespective of whether we can use
+// them for this frame.
+#if CONFIG_DUAL_FILTER
+    for (i = 0; i < 4; ++i)
+      mbmi->interp_filter[i] = cm->interp_filter == SWITCHABLE
+                                   ? EIGHTTAP_REGULAR
+                                   : cm->interp_filter;
+#else
+    mbmi->interp_filter =
+        cm->interp_filter == SWITCHABLE ? EIGHTTAP_REGULAR : cm->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+    x->skip = 0;
+    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
+
+    // Select prediction reference frames.
+    for (i = 0; i < MAX_MB_PLANE; i++) {
+      xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
+      if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
+    }
+
+#if CONFIG_VAR_TX
+    mbmi->inter_tx_size[0][0] = mbmi->tx_size;
+    mbmi->min_tx_size = get_min_tx_size(mbmi->tx_size);
+#endif  // CONFIG_VAR_TX
+
+    if (ref_frame == INTRA_FRAME) {
+      int rate;
+      if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y, &distortion_y,
+                                       NULL, best_rd) >= best_rd)
+        continue;
+      rate2 += rate;
+      rate2 += intra_cost_penalty;
+      distortion2 += distortion_y;
+
+      if (rate_uv_intra == INT_MAX) {
+        choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4, &rate_uv_intra,
+                             &rate_uv_tokenonly, &dist_uv, &skip_uv, &mode_uv);
+      }
+      rate2 += rate_uv_intra;
+      rate_uv = rate_uv_tokenonly;
+      distortion2 += dist_uv;
+      distortion_uv = dist_uv;
+      mbmi->uv_mode = mode_uv;
+    } else {
+      int rate;
+      int64_t distortion;
+      int64_t this_rd_thresh;
+      int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
+      int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
+      int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
+      int tmp_best_skippable = 0;
+      int switchable_filter_index;
+      int_mv *second_ref =
+          comp_pred ? &x->mbmi_ext->ref_mvs[second_ref_frame][0] : NULL;
+      b_mode_info tmp_best_bmodes[16];  // Should this be 4 ?
+      MB_MODE_INFO tmp_best_mbmode;
+#if CONFIG_DUAL_FILTER
+      BEST_SEG_INFO bsi[DUAL_FILTER_SET_SIZE];
+#else
+      BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
+#endif  // CONFIG_DUAL_FILTER
+      int pred_exists = 0;
+      int uv_skippable;
+#if CONFIG_EXT_INTER
+      int_mv compound_seg_newmvs[4][2];
+      for (i = 0; i < 4; i++) {
+        compound_seg_newmvs[i][0].as_int = INVALID_MV;
+        compound_seg_newmvs[i][1].as_int = INVALID_MV;
+      }
+#endif  // CONFIG_EXT_INTER
+
+      this_rd_thresh = (ref_frame == LAST_FRAME)
+                           ? rd_opt->threshes[segment_id][bsize][THR_LAST]
+                           : rd_opt->threshes[segment_id][bsize][THR_ALTR];
+#if CONFIG_EXT_REFS
+      this_rd_thresh = (ref_frame == LAST2_FRAME)
+                           ? rd_opt->threshes[segment_id][bsize][THR_LAST2]
+                           : this_rd_thresh;
+      this_rd_thresh = (ref_frame == LAST3_FRAME)
+                           ? rd_opt->threshes[segment_id][bsize][THR_LAST3]
+                           : this_rd_thresh;
+      this_rd_thresh = (ref_frame == BWDREF_FRAME)
+                           ? rd_opt->threshes[segment_id][bsize][THR_BWDR]
+                           : this_rd_thresh;
+#endif  // CONFIG_EXT_REFS
+      this_rd_thresh = (ref_frame == GOLDEN_FRAME)
+                           ? rd_opt->threshes[segment_id][bsize][THR_GOLD]
+                           : this_rd_thresh;
+
+      // TODO(any): Add search of the tx_type to improve rd performance at the
+      // expense of speed.
+      mbmi->tx_type = DCT_DCT;
+
+      if (cm->interp_filter != BILINEAR) {
+#if CONFIG_DUAL_FILTER
+        tmp_best_filter[0] = EIGHTTAP_REGULAR;
+        tmp_best_filter[1] = EIGHTTAP_REGULAR;
+        tmp_best_filter[2] = EIGHTTAP_REGULAR;
+        tmp_best_filter[3] = EIGHTTAP_REGULAR;
+#else
+        tmp_best_filter = EIGHTTAP_REGULAR;
+#endif  // CONFIG_DUAL_FILTER
+        if (x->source_variance < sf->disable_filter_search_var_thresh) {
+#if CONFIG_DUAL_FILTER
+          tmp_best_filter[0] = EIGHTTAP_REGULAR;
+#else
+          tmp_best_filter = EIGHTTAP_REGULAR;
+#endif  // CONFIG_DUAL_FILTER
+        } else if (sf->adaptive_pred_interp_filter == 1 &&
+                   ctx->pred_interp_filter < SWITCHABLE) {
+#if CONFIG_DUAL_FILTER
+          tmp_best_filter[0] = ctx->pred_interp_filter;
+#else
+          tmp_best_filter = ctx->pred_interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+        } else if (sf->adaptive_pred_interp_filter == 2) {
+#if CONFIG_DUAL_FILTER
+          tmp_best_filter[0] = ctx->pred_interp_filter < SWITCHABLE
+                                   ? ctx->pred_interp_filter
+                                   : 0;
+#else
+          tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE
+                                ? ctx->pred_interp_filter
+                                : 0;
+#endif  // CONFIG_DUAL_FILTER
+        } else {
+#if CONFIG_DUAL_FILTER
+          const int filter_set_size = DUAL_FILTER_SET_SIZE;
+#else
+          const int filter_set_size = SWITCHABLE_FILTERS;
+#endif  // CONFIG_DUAL_FILTER
+          for (switchable_filter_index = 0;
+               switchable_filter_index < filter_set_size;
+               ++switchable_filter_index) {
+            int newbest, rs;
+            int64_t rs_rd;
+            MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
+#if CONFIG_DUAL_FILTER
+            mbmi->interp_filter[0] = filter_sets[switchable_filter_index][0];
+            mbmi->interp_filter[1] = filter_sets[switchable_filter_index][1];
+            mbmi->interp_filter[2] = filter_sets[switchable_filter_index][0];
+            mbmi->interp_filter[3] = filter_sets[switchable_filter_index][1];
+#else
+            mbmi->interp_filter = switchable_filter_index;
+#endif  // CONFIG_DUAL_FILTER
+            tmp_rd = rd_pick_inter_best_sub8x8_mode(
+                cpi, x, &mbmi_ext->ref_mvs[ref_frame][0], second_ref, best_yrd,
+                &rate, &rate_y, &distortion, &skippable, &total_sse,
+                (int)this_rd_thresh, seg_mvs,
+#if CONFIG_EXT_INTER
+                compound_seg_newmvs,
+#endif  // CONFIG_EXT_INTER
+                bsi, switchable_filter_index, mi_row, mi_col);
+            if (tmp_rd == INT64_MAX) continue;
+            rs = av1_get_switchable_rate(cpi, xd);
+            rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
+            if (cm->interp_filter == SWITCHABLE) tmp_rd += rs_rd;
+
+            newbest = (tmp_rd < tmp_best_rd);
+            if (newbest) {
+#if CONFIG_DUAL_FILTER
+              tmp_best_filter[0] = mbmi->interp_filter[0];
+              tmp_best_filter[1] = mbmi->interp_filter[1];
+              tmp_best_filter[2] = mbmi->interp_filter[2];
+              tmp_best_filter[3] = mbmi->interp_filter[3];
+#else
+              tmp_best_filter = mbmi->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+              tmp_best_rd = tmp_rd;
+            }
+            if ((newbest && cm->interp_filter == SWITCHABLE) ||
+                (
+#if CONFIG_DUAL_FILTER
+                    mbmi->interp_filter[0] == cm->interp_filter
+#else
+                    mbmi->interp_filter == cm->interp_filter
+#endif  // CONFIG_DUAL_FILTER
+                    && cm->interp_filter != SWITCHABLE)) {
+              tmp_best_rdu = tmp_rd;
+              tmp_best_rate = rate;
+              tmp_best_ratey = rate_y;
+              tmp_best_distortion = distortion;
+              tmp_best_sse = total_sse;
+              tmp_best_skippable = skippable;
+              tmp_best_mbmode = *mbmi;
+              for (i = 0; i < 4; i++) {
+                tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
+              }
+              pred_exists = 1;
+            }
+          }  // switchable_filter_index loop
+        }
+      }
+
+      if (tmp_best_rdu == INT64_MAX && pred_exists) continue;
+
+#if CONFIG_DUAL_FILTER
+      mbmi->interp_filter[0] =
+          (cm->interp_filter == SWITCHABLE ? tmp_best_filter[0]
+                                           : cm->interp_filter);
+      mbmi->interp_filter[1] =
+          (cm->interp_filter == SWITCHABLE ? tmp_best_filter[1]
+                                           : cm->interp_filter);
+      mbmi->interp_filter[2] =
+          (cm->interp_filter == SWITCHABLE ? tmp_best_filter[2]
+                                           : cm->interp_filter);
+      mbmi->interp_filter[3] =
+          (cm->interp_filter == SWITCHABLE ? tmp_best_filter[3]
+                                           : cm->interp_filter);
+#else
+      mbmi->interp_filter =
+          (cm->interp_filter == SWITCHABLE ? tmp_best_filter
+                                           : cm->interp_filter);
+#endif  // CONFIG_DUAL_FILTER
+
+      if (!pred_exists) {
+        // Handles the special case when a filter that is not in the
+        // switchable list (bilinear) is indicated at the frame level
+        tmp_rd = rd_pick_inter_best_sub8x8_mode(
+            cpi, x, &x->mbmi_ext->ref_mvs[ref_frame][0], second_ref, best_yrd,
+            &rate, &rate_y, &distortion, &skippable, &total_sse,
+            (int)this_rd_thresh, seg_mvs,
+#if CONFIG_EXT_INTER
+            compound_seg_newmvs,
+#endif  // CONFIG_EXT_INTER
+            bsi, 0, mi_row, mi_col);
+        if (tmp_rd == INT64_MAX) continue;
+      } else {
+        total_sse = tmp_best_sse;
+        rate = tmp_best_rate;
+        rate_y = tmp_best_ratey;
+        distortion = tmp_best_distortion;
+        skippable = tmp_best_skippable;
+        *mbmi = tmp_best_mbmode;
+        for (i = 0; i < 4; i++) xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
+      }
+      // Add in the cost of the transform type
+      if (!xd->lossless[mbmi->segment_id]) {
+        int rate_tx_type = 0;
+#if CONFIG_EXT_TX
+        if (get_ext_tx_types(mbmi->tx_size, bsize, 1, cm->reduced_tx_set_used) >
+            1) {
+          const int eset =
+              get_ext_tx_set(mbmi->tx_size, bsize, 1, cm->reduced_tx_set_used);
+          rate_tx_type =
+              cpi->inter_tx_type_costs[eset][mbmi->tx_size][mbmi->tx_type];
+        }
+#else
+        if (mbmi->tx_size < TX_32X32) {
+          rate_tx_type = cpi->inter_tx_type_costs[mbmi->tx_size][mbmi->tx_type];
+        }
+#endif  // CONFIG_EXT_TX
+        rate += rate_tx_type;
+        rate_y += rate_tx_type;
+      }
+
+      rate2 += rate;
+      distortion2 += distortion;
+
+      if (cm->interp_filter == SWITCHABLE)
+        rate2 += av1_get_switchable_rate(cpi, xd);
+
+      if (!mode_excluded)
+        mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
+                                  : cm->reference_mode == COMPOUND_REFERENCE;
+
+      compmode_cost = av1_cost_bit(comp_mode_p, comp_pred);
+
+      tmp_best_rdu =
+          best_rd - AOMMIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
+                           RDCOST(x->rdmult, x->rddiv, 0, total_sse));
+
+      if (tmp_best_rdu > 0) {
+        // If even the 'Y' rd value of split is higher than best so far
+        // then dont bother looking at UV
+        int is_cost_valid_uv;
+        RD_STATS rd_stats_uv;
+        av1_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col, NULL,
+                                        BLOCK_8X8);
+#if CONFIG_VAR_TX
+        is_cost_valid_uv =
+            inter_block_uvrd(cpi, x, &rd_stats_uv, BLOCK_8X8, tmp_best_rdu);
+#else
+        is_cost_valid_uv =
+            super_block_uvrd(cpi, x, &rd_stats_uv, BLOCK_8X8, tmp_best_rdu);
+#endif  // CONFIG_VAR_TX
+        rate_uv = rd_stats_uv.rate;
+        distortion_uv = rd_stats_uv.dist;
+        uv_skippable = rd_stats_uv.skip;
+        uv_sse = rd_stats_uv.sse;
+
+        if (!is_cost_valid_uv) continue;
+        rate2 += rate_uv;
+        distortion2 += distortion_uv;
+        skippable = skippable && uv_skippable;
+        total_sse += uv_sse;
+      } else {
+        continue;
+      }
+    }
+
+    if (cm->reference_mode == REFERENCE_MODE_SELECT) rate2 += compmode_cost;
+
+    // Estimate the reference frame signaling cost and add it
+    // to the rolling cost variable.
+    if (second_ref_frame > INTRA_FRAME) {
+      rate2 += ref_costs_comp[ref_frame];
+#if CONFIG_EXT_REFS
+      rate2 += ref_costs_comp[second_ref_frame];
+#endif  // CONFIG_EXT_REFS
+    } else {
+      rate2 += ref_costs_single[ref_frame];
+    }
+
+    if (!disable_skip) {
+      // Skip is never coded at the segment level for sub8x8 blocks and instead
+      // always coded in the bitstream at the mode info level.
+
+      if (ref_frame != INTRA_FRAME && !xd->lossless[mbmi->segment_id]) {
+        if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv, distortion2) <
+            RDCOST(x->rdmult, x->rddiv, 0, total_sse)) {
+          // Add in the cost of the no skip flag.
+          rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+        } else {
+          // FIXME(rbultje) make this work for splitmv also
+          rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+          distortion2 = total_sse;
+          assert(total_sse >= 0);
+          rate2 -= (rate_y + rate_uv);
+          rate_y = 0;
+          rate_uv = 0;
+          this_skip2 = 1;
+        }
+      } else {
+        // Add in the cost of the no skip flag.
+        rate2 += av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+      }
+
+      // Calculate the final RD estimate for this mode.
+      this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
+    }
+
+    if (!disable_skip && ref_frame == INTRA_FRAME) {
+      for (i = 0; i < REFERENCE_MODES; ++i)
+        best_pred_rd[i] = AOMMIN(best_pred_rd[i], this_rd);
+    }
+
+    // Did this mode help.. i.e. is it the new best mode
+    if (this_rd < best_rd || x->skip) {
+      if (!mode_excluded) {
+        // Note index of best mode so far
+        best_ref_index = ref_index;
+
+        if (ref_frame == INTRA_FRAME) {
+          /* required for left and above block mv */
+          mbmi->mv[0].as_int = 0;
+        }
+
+        rd_cost->rate = rate2;
+#if CONFIG_SUPERTX
+        *returnrate_nocoef = rate2 - rate_y - rate_uv;
+        if (!disable_skip)
+          *returnrate_nocoef -=
+              av1_cost_bit(av1_get_skip_prob(cm, xd), this_skip2);
+        *returnrate_nocoef -= av1_cost_bit(av1_get_intra_inter_prob(cm, xd),
+                                           mbmi->ref_frame[0] != INTRA_FRAME);
+        assert(*returnrate_nocoef > 0);
+#endif  // CONFIG_SUPERTX
+        rd_cost->dist = distortion2;
+        rd_cost->rdcost = this_rd;
+        best_rd = this_rd;
+        best_yrd =
+            best_rd - RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
+        best_mbmode = *mbmi;
+        best_skip2 = this_skip2;
+
+#if CONFIG_VAR_TX
+        for (i = 0; i < MAX_MB_PLANE; ++i)
+          memset(ctx->blk_skip[i], 0, sizeof(uint8_t) * ctx->num_4x4_blk);
+#endif  // CONFIG_VAR_TX
+
+        for (i = 0; i < 4; i++) best_bmodes[i] = xd->mi[0]->bmi[i];
+      }
+    }
+
+    /* keep record of best compound/single-only prediction */
+    if (!disable_skip && ref_frame != INTRA_FRAME) {
+      int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
+
+      if (cm->reference_mode == REFERENCE_MODE_SELECT) {
+        single_rate = rate2 - compmode_cost;
+        hybrid_rate = rate2;
+      } else {
+        single_rate = rate2;
+        hybrid_rate = rate2 + compmode_cost;
+      }
+
+      single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
+      hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
+
+      if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
+        best_pred_rd[SINGLE_REFERENCE] = single_rd;
+      else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
+        best_pred_rd[COMPOUND_REFERENCE] = single_rd;
+
+      if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
+        best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
+    }
+
+    if (x->skip && !comp_pred) break;
+  }
+
+  if (best_rd >= best_rd_so_far) {
+    rd_cost->rate = INT_MAX;
+    rd_cost->rdcost = INT64_MAX;
+#if CONFIG_SUPERTX
+    *returnrate_nocoef = INT_MAX;
+#endif  // CONFIG_SUPERTX
+    return;
+  }
+
+  if (best_rd == INT64_MAX) {
+    rd_cost->rate = INT_MAX;
+    rd_cost->dist = INT64_MAX;
+    rd_cost->rdcost = INT64_MAX;
+#if CONFIG_SUPERTX
+    *returnrate_nocoef = INT_MAX;
+#endif  // CONFIG_SUPERTX
+    return;
+  }
+
+#if CONFIG_DUAL_FILTER
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == best_mbmode.interp_filter[0]) ||
+         !is_inter_block(&best_mbmode));
+#else
+  assert((cm->interp_filter == SWITCHABLE) ||
+         (cm->interp_filter == best_mbmode.interp_filter) ||
+         !is_inter_block(&best_mbmode));
+#endif  // CONFIG_DUAL_FILTER
+
+  av1_update_rd_thresh_fact(cm, tile_data->thresh_freq_fact,
+                            sf->adaptive_rd_thresh, bsize, best_ref_index);
+
+  // macroblock modes
+  *mbmi = best_mbmode;
+#if CONFIG_VAR_TX
+  mbmi->inter_tx_size[0][0] = mbmi->tx_size;
+#endif  // CONFIG_VAR_TX
+
+  x->skip |= best_skip2;
+  if (!is_inter_block(&best_mbmode)) {
+    for (i = 0; i < 4; i++) xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
+  } else {
+    for (i = 0; i < 4; ++i)
+      memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
+
+#if CONFIG_REF_MV
+    mbmi->pred_mv[0].as_int = xd->mi[0]->bmi[3].pred_mv[0].as_int;
+    mbmi->pred_mv[1].as_int = xd->mi[0]->bmi[3].pred_mv[1].as_int;
+#endif  // CONFIG_REF_MV
+    mbmi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
+    mbmi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
+  }
+
+// Note: this section is needed since the mode may have been forced to ZEROMV
+#if CONFIG_GLOBAL_MOTION
+  if (mbmi->mode == ZEROMV
+#if CONFIG_EXT_INTER
+      || mbmi->mode == ZERO_ZEROMV
+#endif  // CONFIG_EXT_INTER
+      ) {
+    if (is_nontrans_global_motion(xd)) {
+#if CONFIG_DUAL_FILTER
+      mbmi->interp_filter[0] = cm->interp_filter == SWITCHABLE
+                                   ? EIGHTTAP_REGULAR
+                                   : cm->interp_filter;
+      mbmi->interp_filter[1] = cm->interp_filter == SWITCHABLE
+                                   ? EIGHTTAP_REGULAR
+                                   : cm->interp_filter;
+#else
+      mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP_REGULAR
+                                                            : cm->interp_filter;
+#endif  // CONFIG_DUAL_FILTER
+    }
+  }
+#endif  // CONFIG_GLOBAL_MOTION
+
+  for (i = 0; i < REFERENCE_MODES; ++i) {
+    if (best_pred_rd[i] == INT64_MAX)
+      best_pred_diff[i] = INT_MIN;
+    else
+      best_pred_diff[i] = best_rd - best_pred_rd[i];
+  }
+
+  store_coding_context(x, ctx, best_ref_index, best_pred_diff, 0);
+}
+
+#if CONFIG_MOTION_VAR
+// This function has a structure similar to av1_build_obmc_inter_prediction
+//
+// The OBMC predictor is computed as:
+//
+//  PObmc(x,y) =
+//    AOM_BLEND_A64(Mh(x),
+//                  AOM_BLEND_A64(Mv(y), P(x,y), PAbove(x,y)),
+//                  PLeft(x, y))
+//
+// Scaling up by AOM_BLEND_A64_MAX_ALPHA ** 2 and omitting the intermediate
+// rounding, this can be written as:
+//
+//  AOM_BLEND_A64_MAX_ALPHA * AOM_BLEND_A64_MAX_ALPHA * Pobmc(x,y) =
+//    Mh(x) * Mv(y) * P(x,y) +
+//      Mh(x) * Cv(y) * Pabove(x,y) +
+//      AOM_BLEND_A64_MAX_ALPHA * Ch(x) * PLeft(x, y)
+//
+// Where :
+//
+//  Cv(y) = AOM_BLEND_A64_MAX_ALPHA - Mv(y)
+//  Ch(y) = AOM_BLEND_A64_MAX_ALPHA - Mh(y)
+//
+// This function computes 'wsrc' and 'mask' as:
+//
+//  wsrc(x, y) =
+//    AOM_BLEND_A64_MAX_ALPHA * AOM_BLEND_A64_MAX_ALPHA * src(x, y) -
+//      Mh(x) * Cv(y) * Pabove(x,y) +
+//      AOM_BLEND_A64_MAX_ALPHA * Ch(x) * PLeft(x, y)
+//
+//  mask(x, y) = Mh(x) * Mv(y)
+//
+// These can then be used to efficiently approximate the error for any
+// predictor P in the context of the provided neighbouring predictors by
+// computing:
+//
+//  error(x, y) =
+//    wsrc(x, y) - mask(x, y) * P(x, y) / (AOM_BLEND_A64_MAX_ALPHA ** 2)
+//
+static void calc_target_weighted_pred(const AV1_COMMON *cm, const MACROBLOCK *x,
+                                      const MACROBLOCKD *xd, int mi_row,
+                                      int mi_col, const uint8_t *above,
+                                      int above_stride, const uint8_t *left,
+                                      int left_stride) {
+  const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+  int row, col, i;
+  const int bw = xd->n8_w << MI_SIZE_LOG2;
+  const int bh = xd->n8_h << MI_SIZE_LOG2;
+  int32_t *mask_buf = x->mask_buf;
+  int32_t *wsrc_buf = x->wsrc_buf;
+  const int wsrc_stride = bw;
+  const int mask_stride = bw;
+  const int src_scale = AOM_BLEND_A64_MAX_ALPHA * AOM_BLEND_A64_MAX_ALPHA;
+#if CONFIG_HIGHBITDEPTH
+  const int is_hbd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0;
+#else
+  const int is_hbd = 0;
+#endif  // CONFIG_HIGHBITDEPTH
+
+  // plane 0 should not be subsampled
+  assert(xd->plane[0].subsampling_x == 0);
+  assert(xd->plane[0].subsampling_y == 0);
+
+  av1_zero_array(wsrc_buf, bw * bh);
+  for (i = 0; i < bw * bh; ++i) mask_buf[i] = AOM_BLEND_A64_MAX_ALPHA;
+
+  // handle above row
+  if (xd->up_available) {
+    const int overlap = num_4x4_blocks_high_lookup[bsize] * 2;
+    const int miw = AOMMIN(xd->n8_w, cm->mi_cols - mi_col);
+    const int mi_row_offset = -1;
+    const uint8_t *const mask1d = av1_get_obmc_mask(overlap);
+    const int neighbor_limit = max_neighbor_obmc[b_width_log2_lookup[bsize]];
+    int neighbor_count = 0;
+
+    assert(miw > 0);
+
+    i = 0;
+    do {  // for each mi in the above row
+      const int mi_col_offset = i;
+      const MB_MODE_INFO *const above_mbmi =
+          &xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride]->mbmi;
+      const BLOCK_SIZE a_bsize = above_mbmi->sb_type;
+      const int mi_step = AOMMIN(xd->n8_w, mi_size_wide[a_bsize]);
+      const int neighbor_bw = mi_step * MI_SIZE;
+
+      if (is_neighbor_overlappable(above_mbmi)) {
+        if (!CONFIG_CB4X4 && (a_bsize == BLOCK_4X4 || a_bsize == BLOCK_4X8))
+          neighbor_count += 2;
+        else
+          neighbor_count++;
+        if (neighbor_count > neighbor_limit) break;
+
+        const int tmp_stride = above_stride;
+        int32_t *wsrc = wsrc_buf + (i * MI_SIZE);
+        int32_t *mask = mask_buf + (i * MI_SIZE);
+
+        if (!is_hbd) {
+          const uint8_t *tmp = above;
+
+          for (row = 0; row < overlap; ++row) {
+            const uint8_t m0 = mask1d[row];
+            const uint8_t m1 = AOM_BLEND_A64_MAX_ALPHA - m0;
+            for (col = 0; col < neighbor_bw; ++col) {
+              wsrc[col] = m1 * tmp[col];
+              mask[col] = m0;
+            }
+            wsrc += wsrc_stride;
+            mask += mask_stride;
+            tmp += tmp_stride;
+          }
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          const uint16_t *tmp = CONVERT_TO_SHORTPTR(above);
+
+          for (row = 0; row < overlap; ++row) {
+            const uint8_t m0 = mask1d[row];
+            const uint8_t m1 = AOM_BLEND_A64_MAX_ALPHA - m0;
+            for (col = 0; col < neighbor_bw; ++col) {
+              wsrc[col] = m1 * tmp[col];
+              mask[col] = m0;
+            }
+            wsrc += wsrc_stride;
+            mask += mask_stride;
+            tmp += tmp_stride;
+          }
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+
+      above += neighbor_bw;
+      i += mi_step;
+    } while (i < miw);
+  }
+
+  for (i = 0; i < bw * bh; ++i) {
+    wsrc_buf[i] *= AOM_BLEND_A64_MAX_ALPHA;
+    mask_buf[i] *= AOM_BLEND_A64_MAX_ALPHA;
+  }
+
+  // handle left column
+  if (xd->left_available) {
+    const int overlap = num_4x4_blocks_wide_lookup[bsize] * 2;
+    const int mih = AOMMIN(xd->n8_h, cm->mi_rows - mi_row);
+    const int mi_col_offset = -1;
+    const uint8_t *const mask1d = av1_get_obmc_mask(overlap);
+    const int neighbor_limit = max_neighbor_obmc[b_height_log2_lookup[bsize]];
+    int neighbor_count = 0;
+
+    assert(mih > 0);
+
+    i = 0;
+    do {  // for each mi in the left column
+      const int mi_row_offset = i;
+      const MB_MODE_INFO *const left_mbmi =
+          &xd->mi[mi_col_offset + mi_row_offset * xd->mi_stride]->mbmi;
+      const BLOCK_SIZE l_bsize = left_mbmi->sb_type;
+      const int mi_step = AOMMIN(xd->n8_h, mi_size_high[l_bsize]);
+      const int neighbor_bh = mi_step * MI_SIZE;
+
+      if (is_neighbor_overlappable(left_mbmi)) {
+        if (!CONFIG_CB4X4 && (l_bsize == BLOCK_4X4 || l_bsize == BLOCK_8X4))
+          neighbor_count += 2;
+        else
+          neighbor_count++;
+        if (neighbor_count > neighbor_limit) break;
+
+        const int tmp_stride = left_stride;
+        int32_t *wsrc = wsrc_buf + (i * MI_SIZE * wsrc_stride);
+        int32_t *mask = mask_buf + (i * MI_SIZE * mask_stride);
+
+        if (!is_hbd) {
+          const uint8_t *tmp = left;
+
+          for (row = 0; row < neighbor_bh; ++row) {
+            for (col = 0; col < overlap; ++col) {
+              const uint8_t m0 = mask1d[col];
+              const uint8_t m1 = AOM_BLEND_A64_MAX_ALPHA - m0;
+              wsrc[col] = (wsrc[col] >> AOM_BLEND_A64_ROUND_BITS) * m0 +
+                          (tmp[col] << AOM_BLEND_A64_ROUND_BITS) * m1;
+              mask[col] = (mask[col] >> AOM_BLEND_A64_ROUND_BITS) * m0;
+            }
+            wsrc += wsrc_stride;
+            mask += mask_stride;
+            tmp += tmp_stride;
+          }
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          const uint16_t *tmp = CONVERT_TO_SHORTPTR(left);
+
+          for (row = 0; row < neighbor_bh; ++row) {
+            for (col = 0; col < overlap; ++col) {
+              const uint8_t m0 = mask1d[col];
+              const uint8_t m1 = AOM_BLEND_A64_MAX_ALPHA - m0;
+              wsrc[col] = (wsrc[col] >> AOM_BLEND_A64_ROUND_BITS) * m0 +
+                          (tmp[col] << AOM_BLEND_A64_ROUND_BITS) * m1;
+              mask[col] = (mask[col] >> AOM_BLEND_A64_ROUND_BITS) * m0;
+            }
+            wsrc += wsrc_stride;
+            mask += mask_stride;
+            tmp += tmp_stride;
+          }
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+
+      left += neighbor_bh * left_stride;
+      i += mi_step;
+    } while (i < mih);
+  }
+
+  if (!is_hbd) {
+    const uint8_t *src = x->plane[0].src.buf;
+
+    for (row = 0; row < bh; ++row) {
+      for (col = 0; col < bw; ++col) {
+        wsrc_buf[col] = src[col] * src_scale - wsrc_buf[col];
+      }
+      wsrc_buf += wsrc_stride;
+      src += x->plane[0].src.stride;
+    }
+#if CONFIG_HIGHBITDEPTH
+  } else {
+    const uint16_t *src = CONVERT_TO_SHORTPTR(x->plane[0].src.buf);
+
+    for (row = 0; row < bh; ++row) {
+      for (col = 0; col < bw; ++col) {
+        wsrc_buf[col] = src[col] * src_scale - wsrc_buf[col];
+      }
+      wsrc_buf += wsrc_stride;
+      src += x->plane[0].src.stride;
+    }
+#endif  // CONFIG_HIGHBITDEPTH
+  }
+}
+
+#if CONFIG_NCOBMC
+void av1_check_ncobmc_rd(const struct AV1_COMP *cpi, struct macroblock *x,
+                         int mi_row, int mi_col) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  MB_MODE_INFO backup_mbmi;
+  BLOCK_SIZE bsize = mbmi->sb_type;
+  int ref, skip_blk, backup_skip = x->skip;
+  int64_t rd_causal;
+  RD_STATS rd_stats_y, rd_stats_uv;
+  int rate_skip0 = av1_cost_bit(av1_get_skip_prob(cm, xd), 0);
+  int rate_skip1 = av1_cost_bit(av1_get_skip_prob(cm, xd), 1);
+
+  // Recompute the best causal predictor and rd
+  mbmi->motion_mode = SIMPLE_TRANSLATION;
+  set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
+  for (ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) {
+    YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, mbmi->ref_frame[ref]);
+    assert(cfg != NULL);
+    av1_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
+                         &xd->block_refs[ref]->sf);
+  }
+  av1_setup_dst_planes(x->e_mbd.plane, bsize,
+                       get_frame_new_buffer(&cpi->common), mi_row, mi_col);
+
+  av1_build_inter_predictors_sb(xd, mi_row, mi_col, NULL, bsize);
+
+  av1_subtract_plane(x, bsize, 0);
+  super_block_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
+  super_block_uvrd(cpi, x, &rd_stats_uv, bsize, INT64_MAX);
+  assert(rd_stats_y.rate != INT_MAX && rd_stats_uv.rate != INT_MAX);
+  if (rd_stats_y.skip && rd_stats_uv.skip) {
+    rd_stats_y.rate = rate_skip1;
+    rd_stats_uv.rate = 0;
+    rd_stats_y.dist = rd_stats_y.sse;
+    rd_stats_uv.dist = rd_stats_uv.sse;
+    skip_blk = 0;
+  } else if (RDCOST(x->rdmult, x->rddiv,
+                    (rd_stats_y.rate + rd_stats_uv.rate + rate_skip0),
+                    (rd_stats_y.dist + rd_stats_uv.dist)) >
+             RDCOST(x->rdmult, x->rddiv, rate_skip1,
+                    (rd_stats_y.sse + rd_stats_uv.sse))) {
+    rd_stats_y.rate = rate_skip1;
+    rd_stats_uv.rate = 0;
+    rd_stats_y.dist = rd_stats_y.sse;
+    rd_stats_uv.dist = rd_stats_uv.sse;
+    skip_blk = 1;
+  } else {
+    rd_stats_y.rate += rate_skip0;
+    skip_blk = 0;
+  }
+  backup_skip = skip_blk;
+  backup_mbmi = *mbmi;
+  rd_causal = RDCOST(x->rdmult, x->rddiv, (rd_stats_y.rate + rd_stats_uv.rate),
+                     (rd_stats_y.dist + rd_stats_uv.dist));
+  rd_causal += RDCOST(x->rdmult, x->rddiv,
+                      av1_cost_bit(cm->fc->motion_mode_prob[bsize][0], 0), 0);
+
+  // Check non-causal mode
+  mbmi->motion_mode = OBMC_CAUSAL;
+  av1_build_ncobmc_inter_predictors_sb(cm, xd, mi_row, mi_col);
+
+  av1_subtract_plane(x, bsize, 0);
+  super_block_yrd(cpi, x, &rd_stats_y, bsize, INT64_MAX);
+  super_block_uvrd(cpi, x, &rd_stats_uv, bsize, INT64_MAX);
+  assert(rd_stats_y.rate != INT_MAX && rd_stats_uv.rate != INT_MAX);
+  if (rd_stats_y.skip && rd_stats_uv.skip) {
+    rd_stats_y.rate = rate_skip1;
+    rd_stats_uv.rate = 0;
+    rd_stats_y.dist = rd_stats_y.sse;
+    rd_stats_uv.dist = rd_stats_uv.sse;
+    skip_blk = 0;
+  } else if (RDCOST(x->rdmult, x->rddiv,
+                    (rd_stats_y.rate + rd_stats_uv.rate + rate_skip0),
+                    (rd_stats_y.dist + rd_stats_uv.dist)) >
+             RDCOST(x->rdmult, x->rddiv, rate_skip1,
+                    (rd_stats_y.sse + rd_stats_uv.sse))) {
+    rd_stats_y.rate = rate_skip1;
+    rd_stats_uv.rate = 0;
+    rd_stats_y.dist = rd_stats_y.sse;
+    rd_stats_uv.dist = rd_stats_uv.sse;
+    skip_blk = 1;
+  } else {
+    rd_stats_y.rate += rate_skip0;
+    skip_blk = 0;
+  }
+
+  if (rd_causal >
+      RDCOST(x->rdmult, x->rddiv,
+             rd_stats_y.rate + rd_stats_uv.rate +
+                 av1_cost_bit(cm->fc->motion_mode_prob[bsize][0], 1),
+             (rd_stats_y.dist + rd_stats_uv.dist))) {
+    x->skip = skip_blk;
+  } else {
+    *mbmi = backup_mbmi;
+    x->skip = backup_skip;
+  }
+}
+#endif  // CONFIG_NCOBMC
+#endif  // CONFIG_MOTION_VAR
diff --git a/third_party/aom/av1/encoder/rdopt.h b/third_party/aom/av1/encoder/rdopt.h
new file mode 100644
index 0000000000..a7053b2897
--- /dev/null
+++ b/third_party/aom/av1/encoder/rdopt.h
@@ -0,0 +1,142 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_RDOPT_H_
+#define AV1_ENCODER_RDOPT_H_
+
+#include "av1/common/blockd.h"
+
+#include "av1/encoder/block.h"
+#include "av1/encoder/context_tree.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct TileInfo;
+struct AV1_COMP;
+struct macroblock;
+struct RD_STATS;
+
+#if CONFIG_RD_DEBUG
+static INLINE void av1_update_txb_coeff_cost(RD_STATS *rd_stats, int plane,
+                                             TX_SIZE tx_size, int blk_row,
+                                             int blk_col, int txb_coeff_cost) {
+  (void)blk_row;
+  (void)blk_col;
+  (void)tx_size;
+  rd_stats->txb_coeff_cost[plane] += txb_coeff_cost;
+
+#if CONFIG_VAR_TX
+  {
+    const int txb_h = tx_size_high_unit[tx_size];
+    const int txb_w = tx_size_wide_unit[tx_size];
+    int idx, idy;
+    for (idy = 0; idy < txb_h; ++idy)
+      for (idx = 0; idx < txb_w; ++idx)
+        rd_stats->txb_coeff_cost_map[plane][blk_row + idy][blk_col + idx] = 0;
+
+    rd_stats->txb_coeff_cost_map[plane][blk_row][blk_col] = txb_coeff_cost;
+  }
+  assert(blk_row < TXB_COEFF_COST_MAP_SIZE);
+  assert(blk_col < TXB_COEFF_COST_MAP_SIZE);
+#endif
+}
+#endif
+
+typedef enum OUTPUT_STATUS {
+  OUTPUT_HAS_PREDICTED_PIXELS,
+  OUTPUT_HAS_DECODED_PIXELS
+} OUTPUT_STATUS;
+
+void av1_dist_block(const AV1_COMP *cpi, MACROBLOCK *x, int plane,
+                    BLOCK_SIZE plane_bsize, int block, int blk_row, int blk_col,
+                    TX_SIZE tx_size, int64_t *out_dist, int64_t *out_sse,
+                    OUTPUT_STATUS output_status);
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+int av1_cost_coeffs(const AV1_COMP *const cpi, MACROBLOCK *x, int plane,
+                    int block, TX_SIZE tx_size, const SCAN_ORDER *scan_order,
+                    const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
+                    int use_fast_coef_costing);
+#endif
+void av1_rd_pick_intra_mode_sb(const struct AV1_COMP *cpi, struct macroblock *x,
+                               struct RD_STATS *rd_cost, BLOCK_SIZE bsize,
+                               PICK_MODE_CONTEXT *ctx, int64_t best_rd);
+
+unsigned int av1_get_sby_perpixel_variance(const AV1_COMP *cpi,
+                                           const struct buf_2d *ref,
+                                           BLOCK_SIZE bs);
+#if CONFIG_HIGHBITDEPTH
+unsigned int av1_high_get_sby_perpixel_variance(const AV1_COMP *cpi,
+                                                const struct buf_2d *ref,
+                                                BLOCK_SIZE bs, int bd);
+#endif
+
+void av1_rd_pick_inter_mode_sb(const struct AV1_COMP *cpi,
+                               struct TileDataEnc *tile_data,
+                               struct macroblock *x, int mi_row, int mi_col,
+                               struct RD_STATS *rd_cost,
+#if CONFIG_SUPERTX
+                               int *returnrate_nocoef,
+#endif  // CONFIG_SUPERTX
+                               BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                               int64_t best_rd_so_far);
+
+void av1_rd_pick_inter_mode_sb_seg_skip(
+    const struct AV1_COMP *cpi, struct TileDataEnc *tile_data,
+    struct macroblock *x, int mi_row, int mi_col, struct RD_STATS *rd_cost,
+    BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx, int64_t best_rd_so_far);
+
+int av1_internal_image_edge(const struct AV1_COMP *cpi);
+int av1_active_h_edge(const struct AV1_COMP *cpi, int mi_row, int mi_step);
+int av1_active_v_edge(const struct AV1_COMP *cpi, int mi_col, int mi_step);
+int av1_active_edge_sb(const struct AV1_COMP *cpi, int mi_row, int mi_col);
+
+void av1_rd_pick_inter_mode_sub8x8(const struct AV1_COMP *cpi,
+                                   struct TileDataEnc *tile_data,
+                                   struct macroblock *x, int mi_row, int mi_col,
+                                   struct RD_STATS *rd_cost,
+#if CONFIG_SUPERTX
+                                   int *returnrate_nocoef,
+#endif  // CONFIG_SUPERTX
+                                   BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
+                                   int64_t best_rd_so_far);
+
+#if CONFIG_MOTION_VAR && CONFIG_NCOBMC
+void av1_check_ncobmc_rd(const struct AV1_COMP *cpi, struct macroblock *x,
+                         int mi_row, int mi_col);
+#endif  // CONFIG_MOTION_VAR && CONFIG_NCOBMC
+
+#if CONFIG_SUPERTX
+#if CONFIG_VAR_TX
+void av1_tx_block_rd_b(const AV1_COMP *cpi, MACROBLOCK *x, TX_SIZE tx_size,
+                       int blk_row, int blk_col, int plane, int block,
+                       int plane_bsize, const ENTROPY_CONTEXT *a,
+                       const ENTROPY_CONTEXT *l, RD_STATS *rd_stats);
+#endif
+
+void av1_txfm_rd_in_plane_supertx(MACROBLOCK *x, const AV1_COMP *cpi, int *rate,
+                                  int64_t *distortion, int *skippable,
+                                  int64_t *sse, int64_t ref_best_rd, int plane,
+                                  BLOCK_SIZE bsize, TX_SIZE tx_size,
+                                  int use_fast_coef_casting);
+#endif  // CONFIG_SUPERTX
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+int av1_tx_type_cost(const AV1_COMP *cpi, const MACROBLOCKD *xd,
+                     BLOCK_SIZE bsize, int plane, TX_SIZE tx_size,
+                     TX_TYPE tx_type);
+
+#endif  // AV1_ENCODER_RDOPT_H_
diff --git a/third_party/aom/av1/encoder/segmentation.c b/third_party/aom/av1/encoder/segmentation.c
new file mode 100644
index 0000000000..b581a61d0a
--- /dev/null
+++ b/third_party/aom/av1/encoder/segmentation.c
@@ -0,0 +1,394 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+
+#include "aom_mem/aom_mem.h"
+
+#include "av1/common/pred_common.h"
+#include "av1/common/tile_common.h"
+
+#include "av1/encoder/cost.h"
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/subexp.h"
+
+void av1_enable_segmentation(struct segmentation *seg) {
+  seg->enabled = 1;
+  seg->update_map = 1;
+  seg->update_data = 1;
+}
+
+void av1_disable_segmentation(struct segmentation *seg) {
+  seg->enabled = 0;
+  seg->update_map = 0;
+  seg->update_data = 0;
+}
+
+void av1_set_segment_data(struct segmentation *seg, signed char *feature_data,
+                          unsigned char abs_delta) {
+  seg->abs_delta = abs_delta;
+
+  memcpy(seg->feature_data, feature_data, sizeof(seg->feature_data));
+}
+void av1_disable_segfeature(struct segmentation *seg, int segment_id,
+                            SEG_LVL_FEATURES feature_id) {
+  seg->feature_mask[segment_id] &= ~(1 << feature_id);
+}
+
+void av1_clear_segdata(struct segmentation *seg, int segment_id,
+                       SEG_LVL_FEATURES feature_id) {
+  seg->feature_data[segment_id][feature_id] = 0;
+}
+
+// Based on set of segment counts calculate a probability tree
+static void calc_segtree_probs(unsigned *segcounts,
+                               aom_prob *segment_tree_probs,
+                               const aom_prob *cur_tree_probs,
+                               const int probwt) {
+  // Work out probabilities of each segment
+  const unsigned cc[4] = { segcounts[0] + segcounts[1],
+                           segcounts[2] + segcounts[3],
+                           segcounts[4] + segcounts[5],
+                           segcounts[6] + segcounts[7] };
+  const unsigned ccc[2] = { cc[0] + cc[1], cc[2] + cc[3] };
+  int i;
+
+  segment_tree_probs[0] = get_binary_prob(ccc[0], ccc[1]);
+  segment_tree_probs[1] = get_binary_prob(cc[0], cc[1]);
+  segment_tree_probs[2] = get_binary_prob(cc[2], cc[3]);
+  segment_tree_probs[3] = get_binary_prob(segcounts[0], segcounts[1]);
+  segment_tree_probs[4] = get_binary_prob(segcounts[2], segcounts[3]);
+  segment_tree_probs[5] = get_binary_prob(segcounts[4], segcounts[5]);
+  segment_tree_probs[6] = get_binary_prob(segcounts[6], segcounts[7]);
+
+  for (i = 0; i < 7; i++) {
+    const unsigned *ct =
+        i == 0 ? ccc : i < 3 ? cc + (i & 2) : segcounts + (i - 3) * 2;
+    av1_prob_diff_update_savings_search(ct, cur_tree_probs[i],
+                                        &segment_tree_probs[i],
+                                        DIFF_UPDATE_PROB, probwt);
+  }
+}
+
+// Based on set of segment counts and probabilities calculate a cost estimate
+static int cost_segmap(unsigned *segcounts, aom_prob *probs) {
+  const int c01 = segcounts[0] + segcounts[1];
+  const int c23 = segcounts[2] + segcounts[3];
+  const int c45 = segcounts[4] + segcounts[5];
+  const int c67 = segcounts[6] + segcounts[7];
+  const int c0123 = c01 + c23;
+  const int c4567 = c45 + c67;
+
+  // Cost the top node of the tree
+  int cost = c0123 * av1_cost_zero(probs[0]) + c4567 * av1_cost_one(probs[0]);
+
+  // Cost subsequent levels
+  if (c0123 > 0) {
+    cost += c01 * av1_cost_zero(probs[1]) + c23 * av1_cost_one(probs[1]);
+
+    if (c01 > 0)
+      cost += segcounts[0] * av1_cost_zero(probs[3]) +
+              segcounts[1] * av1_cost_one(probs[3]);
+    if (c23 > 0)
+      cost += segcounts[2] * av1_cost_zero(probs[4]) +
+              segcounts[3] * av1_cost_one(probs[4]);
+  }
+
+  if (c4567 > 0) {
+    cost += c45 * av1_cost_zero(probs[2]) + c67 * av1_cost_one(probs[2]);
+
+    if (c45 > 0)
+      cost += segcounts[4] * av1_cost_zero(probs[5]) +
+              segcounts[5] * av1_cost_one(probs[5]);
+    if (c67 > 0)
+      cost += segcounts[6] * av1_cost_zero(probs[6]) +
+              segcounts[7] * av1_cost_one(probs[6]);
+  }
+
+  return cost;
+}
+
+static void count_segs(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                       const TileInfo *tile, MODE_INFO **mi,
+                       unsigned *no_pred_segcounts,
+                       unsigned (*temporal_predictor_count)[2],
+                       unsigned *t_unpred_seg_counts, int bw, int bh,
+                       int mi_row, int mi_col) {
+  int segment_id;
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  xd->mi = mi;
+  segment_id = xd->mi[0]->mbmi.segment_id;
+
+  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw,
+#if CONFIG_DEPENDENT_HORZTILES
+                 cm->dependent_horz_tiles,
+#endif  // CONFIG_DEPENDENT_HORZTILES
+                 cm->mi_rows, cm->mi_cols);
+
+  // Count the number of hits on each segment with no prediction
+  no_pred_segcounts[segment_id]++;
+
+  // Temporal prediction not allowed on key frames
+  if (cm->frame_type != KEY_FRAME) {
+    const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
+    // Test to see if the segment id matches the predicted value.
+    const int pred_segment_id =
+        get_segment_id(cm, cm->last_frame_seg_map, bsize, mi_row, mi_col);
+    const int pred_flag = pred_segment_id == segment_id;
+    const int pred_context = av1_get_pred_context_seg_id(xd);
+
+    // Store the prediction status for this mb and update counts
+    // as appropriate
+    xd->mi[0]->mbmi.seg_id_predicted = pred_flag;
+    temporal_predictor_count[pred_context][pred_flag]++;
+
+    // Update the "unpredicted" segment count
+    if (!pred_flag) t_unpred_seg_counts[segment_id]++;
+  }
+}
+
+static void count_segs_sb(const AV1_COMMON *cm, MACROBLOCKD *xd,
+                          const TileInfo *tile, MODE_INFO **mi,
+                          unsigned *no_pred_segcounts,
+                          unsigned (*temporal_predictor_count)[2],
+                          unsigned *t_unpred_seg_counts, int mi_row, int mi_col,
+                          BLOCK_SIZE bsize) {
+  const int mis = cm->mi_stride;
+  const int bs = mi_size_wide[bsize], hbs = bs / 2;
+#if CONFIG_EXT_PARTITION_TYPES
+  PARTITION_TYPE partition;
+#else
+  int bw, bh;
+#endif  // CONFIG_EXT_PARTITION_TYPES
+
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+#if CONFIG_EXT_PARTITION_TYPES
+  if (bsize == BLOCK_8X8)
+    partition = PARTITION_NONE;
+  else
+    partition = get_partition(cm, mi_row, mi_col, bsize);
+  switch (partition) {
+    case PARTITION_NONE:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, bs, bs, mi_row, mi_col);
+      break;
+    case PARTITION_HORZ:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
+                 mi_row + hbs, mi_col);
+      break;
+    case PARTITION_VERT:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, bs, mi_row,
+                 mi_col + hbs);
+      break;
+    case PARTITION_HORZ_A:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, hbs, hbs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row, mi_col + hbs);
+      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
+                 mi_row + hbs, mi_col);
+      break;
+    case PARTITION_HORZ_B:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row + hbs, mi_col);
+      count_segs(cm, xd, tile, mi + hbs + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row + hbs, mi_col + hbs);
+      break;
+    case PARTITION_VERT_A:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, hbs, hbs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row + hbs, mi_col);
+      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, bs, mi_row,
+                 mi_col + hbs);
+      break;
+    case PARTITION_VERT_B:
+      count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+                 t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
+      count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row, mi_col + hbs);
+      count_segs(cm, xd, tile, mi + hbs + hbs * mis, no_pred_segcounts,
+                 temporal_predictor_count, t_unpred_seg_counts, hbs, hbs,
+                 mi_row + hbs, mi_col + hbs);
+      break;
+    case PARTITION_SPLIT: {
+      const BLOCK_SIZE subsize = subsize_lookup[PARTITION_SPLIT][bsize];
+      int n;
+
+      assert(num_8x8_blocks_wide_lookup[mi[0]->mbmi.sb_type] < bs &&
+             num_8x8_blocks_high_lookup[mi[0]->mbmi.sb_type] < bs);
+
+      for (n = 0; n < 4; n++) {
+        const int mi_dc = hbs * (n & 1);
+        const int mi_dr = hbs * (n >> 1);
+
+        count_segs_sb(cm, xd, tile, &mi[mi_dr * mis + mi_dc], no_pred_segcounts,
+                      temporal_predictor_count, t_unpred_seg_counts,
+                      mi_row + mi_dr, mi_col + mi_dc, subsize);
+      }
+    } break;
+    default: assert(0);
+  }
+#else
+  bw = mi_size_wide[mi[0]->mbmi.sb_type];
+  bh = mi_size_high[mi[0]->mbmi.sb_type];
+
+  if (bw == bs && bh == bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, bs, bs, mi_row, mi_col);
+  } else if (bw == bs && bh < bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, bs, hbs, mi_row, mi_col);
+    count_segs(cm, xd, tile, mi + hbs * mis, no_pred_segcounts,
+               temporal_predictor_count, t_unpred_seg_counts, bs, hbs,
+               mi_row + hbs, mi_col);
+  } else if (bw < bs && bh == bs) {
+    count_segs(cm, xd, tile, mi, no_pred_segcounts, temporal_predictor_count,
+               t_unpred_seg_counts, hbs, bs, mi_row, mi_col);
+    count_segs(cm, xd, tile, mi + hbs, no_pred_segcounts,
+               temporal_predictor_count, t_unpred_seg_counts, hbs, bs, mi_row,
+               mi_col + hbs);
+  } else {
+    const BLOCK_SIZE subsize = subsize_lookup[PARTITION_SPLIT][bsize];
+    int n;
+
+    assert(bw < bs && bh < bs);
+
+    for (n = 0; n < 4; n++) {
+      const int mi_dc = hbs * (n & 1);
+      const int mi_dr = hbs * (n >> 1);
+
+      count_segs_sb(cm, xd, tile, &mi[mi_dr * mis + mi_dc], no_pred_segcounts,
+                    temporal_predictor_count, t_unpred_seg_counts,
+                    mi_row + mi_dr, mi_col + mi_dc, subsize);
+    }
+  }
+#endif  // CONFIG_EXT_PARTITION_TYPES
+}
+
+void av1_choose_segmap_coding_method(AV1_COMMON *cm, MACROBLOCKD *xd) {
+  struct segmentation *seg = &cm->seg;
+  struct segmentation_probs *segp = &cm->fc->seg;
+
+  int no_pred_cost;
+  int t_pred_cost = INT_MAX;
+
+  int i, tile_col, tile_row, mi_row, mi_col;
+#if CONFIG_TILE_GROUPS
+  const int probwt = cm->num_tg;
+#else
+  const int probwt = 1;
+#endif
+
+  unsigned(*temporal_predictor_count)[2] = cm->counts.seg.pred;
+  unsigned *no_pred_segcounts = cm->counts.seg.tree_total;
+  unsigned *t_unpred_seg_counts = cm->counts.seg.tree_mispred;
+
+  aom_prob no_pred_tree[SEG_TREE_PROBS];
+  aom_prob t_pred_tree[SEG_TREE_PROBS];
+  aom_prob t_nopred_prob[PREDICTION_PROBS];
+
+  (void)xd;
+
+  // We are about to recompute all the segment counts, so zero the accumulators.
+  av1_zero(cm->counts.seg);
+
+  // First of all generate stats regarding how well the last segment map
+  // predicts this one
+  for (tile_row = 0; tile_row < cm->tile_rows; tile_row++) {
+    TileInfo tile_info;
+    av1_tile_set_row(&tile_info, cm, tile_row);
+    for (tile_col = 0; tile_col < cm->tile_cols; tile_col++) {
+      MODE_INFO **mi_ptr;
+      av1_tile_set_col(&tile_info, cm, tile_col);
+#if CONFIG_TILE_GROUPS && CONFIG_DEPENDENT_HORZTILES
+      av1_tile_set_tg_boundary(&tile_info, cm, tile_row, tile_col);
+#endif
+      mi_ptr = cm->mi_grid_visible + tile_info.mi_row_start * cm->mi_stride +
+               tile_info.mi_col_start;
+      for (mi_row = tile_info.mi_row_start; mi_row < tile_info.mi_row_end;
+           mi_row += cm->mib_size, mi_ptr += cm->mib_size * cm->mi_stride) {
+        MODE_INFO **mi = mi_ptr;
+        for (mi_col = tile_info.mi_col_start; mi_col < tile_info.mi_col_end;
+             mi_col += cm->mib_size, mi += cm->mib_size) {
+          count_segs_sb(cm, xd, &tile_info, mi, no_pred_segcounts,
+                        temporal_predictor_count, t_unpred_seg_counts, mi_row,
+                        mi_col, cm->sb_size);
+        }
+      }
+    }
+  }
+
+  // Work out probability tree for coding segments without prediction
+  // and the cost.
+  calc_segtree_probs(no_pred_segcounts, no_pred_tree, segp->tree_probs, probwt);
+  no_pred_cost = cost_segmap(no_pred_segcounts, no_pred_tree);
+
+  // Key frames cannot use temporal prediction
+  if (!frame_is_intra_only(cm) && !cm->error_resilient_mode) {
+    // Work out probability tree for coding those segments not
+    // predicted using the temporal method and the cost.
+    calc_segtree_probs(t_unpred_seg_counts, t_pred_tree, segp->tree_probs,
+                       probwt);
+    t_pred_cost = cost_segmap(t_unpred_seg_counts, t_pred_tree);
+
+    // Add in the cost of the signaling for each prediction context.
+    for (i = 0; i < PREDICTION_PROBS; i++) {
+      const int count0 = temporal_predictor_count[i][0];
+      const int count1 = temporal_predictor_count[i][1];
+
+      t_nopred_prob[i] = get_binary_prob(count0, count1);
+      av1_prob_diff_update_savings_search(
+          temporal_predictor_count[i], segp->pred_probs[i], &t_nopred_prob[i],
+          DIFF_UPDATE_PROB, probwt);
+
+      // Add in the predictor signaling cost
+      t_pred_cost += count0 * av1_cost_zero(t_nopred_prob[i]) +
+                     count1 * av1_cost_one(t_nopred_prob[i]);
+    }
+  }
+
+  // Now choose which coding method to use.
+  if (t_pred_cost < no_pred_cost) {
+    assert(!cm->error_resilient_mode);
+    seg->temporal_update = 1;
+  } else {
+    seg->temporal_update = 0;
+  }
+}
+
+void av1_reset_segment_features(AV1_COMMON *cm) {
+  struct segmentation *seg = &cm->seg;
+
+  // Set up default state for MB feature flags
+  seg->enabled = 0;
+  seg->update_map = 0;
+  seg->update_data = 0;
+  av1_clearall_segfeatures(seg);
+}
diff --git a/third_party/aom/av1/encoder/segmentation.h b/third_party/aom/av1/encoder/segmentation.h
new file mode 100644
index 0000000000..c1491ca2af
--- /dev/null
+++ b/third_party/aom/av1/encoder/segmentation.h
@@ -0,0 +1,51 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_SEGMENTATION_H_
+#define AV1_ENCODER_SEGMENTATION_H_
+
+#include "av1/common/blockd.h"
+#include "av1/encoder/encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_enable_segmentation(struct segmentation *seg);
+void av1_disable_segmentation(struct segmentation *seg);
+
+void av1_disable_segfeature(struct segmentation *seg, int segment_id,
+                            SEG_LVL_FEATURES feature_id);
+void av1_clear_segdata(struct segmentation *seg, int segment_id,
+                       SEG_LVL_FEATURES feature_id);
+
+// The values given for each segment can be either deltas (from the default
+// value chosen for the frame) or absolute values.
+//
+// Valid range for abs values is (0-127 for MB_LVL_ALT_Q), (0-63 for
+// SEGMENT_ALT_LF)
+// Valid range for delta values are (+/-127 for MB_LVL_ALT_Q), (+/-63 for
+// SEGMENT_ALT_LF)
+//
+// abs_delta = SEGMENT_DELTADATA (deltas) abs_delta = SEGMENT_ABSDATA (use
+// the absolute values given).
+void av1_set_segment_data(struct segmentation *seg, signed char *feature_data,
+                          unsigned char abs_delta);
+
+void av1_choose_segmap_coding_method(AV1_COMMON *cm, MACROBLOCKD *xd);
+
+void av1_reset_segment_features(AV1_COMMON *cm);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_SEGMENTATION_H_
diff --git a/third_party/aom/av1/encoder/speed_features.c b/third_party/aom/av1/encoder/speed_features.c
new file mode 100644
index 0000000000..20c96761b7
--- /dev/null
+++ b/third_party/aom/av1/encoder/speed_features.c
@@ -0,0 +1,506 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <limits.h>
+
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/speed_features.h"
+#include "av1/encoder/rdopt.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+
+#define MAX_MESH_SPEED 5  // Max speed setting for mesh motion method
+static MESH_PATTERN
+    good_quality_mesh_patterns[MAX_MESH_SPEED + 1][MAX_MESH_STEP] = {
+      { { 64, 8 }, { 28, 4 }, { 15, 1 }, { 7, 1 } },
+      { { 64, 8 }, { 28, 4 }, { 15, 1 }, { 7, 1 } },
+      { { 64, 8 }, { 14, 2 }, { 7, 1 }, { 7, 1 } },
+      { { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
+      { { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
+      { { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
+    };
+static unsigned char good_quality_max_mesh_pct[MAX_MESH_SPEED + 1] = {
+  50, 25, 15, 5, 1, 1
+};
+
+#if CONFIG_INTRABC
+// TODO(aconverse@google.com): These settings are pretty relaxed, tune them for
+// each speed setting
+static MESH_PATTERN intrabc_mesh_patterns[MAX_MESH_SPEED + 1][MAX_MESH_STEP] = {
+  { { 64, 1 }, { 64, 1 }, { 0, 0 }, { 0, 0 } },
+  { { 64, 1 }, { 64, 1 }, { 0, 0 }, { 0, 0 } },
+  { { 64, 1 }, { 64, 1 }, { 0, 0 }, { 0, 0 } },
+  { { 64, 4 }, { 16, 1 }, { 0, 0 }, { 0, 0 } },
+  { { 64, 4 }, { 16, 1 }, { 0, 0 }, { 0, 0 } },
+  { { 64, 4 }, { 16, 1 }, { 0, 0 }, { 0, 0 } },
+};
+static uint8_t intrabc_max_mesh_pct[MAX_MESH_SPEED + 1] = { 100, 100, 100,
+                                                            25,  25,  10 };
+#endif
+
+// Intra only frames, golden frames (except alt ref overlays) and
+// alt ref frames tend to be coded at a higher than ambient quality
+static int frame_is_boosted(const AV1_COMP *cpi) {
+  return frame_is_kf_gf_arf(cpi);
+}
+
+// Sets a partition size down to which the auto partition code will always
+// search (can go lower), based on the image dimensions. The logic here
+// is that the extent to which ringing artefacts are offensive, depends
+// partly on the screen area that over which they propogate. Propogation is
+// limited by transform block size but the screen area take up by a given block
+// size will be larger for a small image format stretched to full screen.
+static BLOCK_SIZE set_partition_min_limit(AV1_COMMON *const cm) {
+  unsigned int screen_area = (cm->width * cm->height);
+
+  // Select block size based on image format size.
+  if (screen_area < 1280 * 720) {
+    // Formats smaller in area than 720P
+    return BLOCK_4X4;
+  } else if (screen_area < 1920 * 1080) {
+    // Format >= 720P and < 1080P
+    return BLOCK_8X8;
+  } else {
+    // Formats 1080P and up
+    return BLOCK_16X16;
+  }
+}
+
+static void set_good_speed_feature_framesize_dependent(AV1_COMP *cpi,
+                                                       SPEED_FEATURES *sf,
+                                                       int speed) {
+  AV1_COMMON *const cm = &cpi->common;
+
+  if (speed >= 1) {
+    if (AOMMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask =
+          cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
+      sf->partition_search_breakout_dist_thr = (1 << 23);
+    } else {
+      sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
+      sf->partition_search_breakout_dist_thr = (1 << 21);
+    }
+  }
+
+  if (speed >= 2) {
+    if (AOMMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask =
+          cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
+      sf->adaptive_pred_interp_filter = 0;
+      sf->partition_search_breakout_dist_thr = (1 << 24);
+      sf->partition_search_breakout_rate_thr = 120;
+    } else {
+      sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
+      sf->partition_search_breakout_dist_thr = (1 << 22);
+      sf->partition_search_breakout_rate_thr = 100;
+    }
+    sf->rd_auto_partition_min_limit = set_partition_min_limit(cm);
+  }
+
+  if (speed >= 3) {
+    if (AOMMIN(cm->width, cm->height) >= 720) {
+      sf->disable_split_mask = DISABLE_ALL_SPLIT;
+      sf->schedule_mode_search = cm->base_qindex < 220 ? 1 : 0;
+      sf->partition_search_breakout_dist_thr = (1 << 25);
+      sf->partition_search_breakout_rate_thr = 200;
+    } else {
+      sf->max_intra_bsize = BLOCK_32X32;
+      sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT;
+      sf->schedule_mode_search = cm->base_qindex < 175 ? 1 : 0;
+      sf->partition_search_breakout_dist_thr = (1 << 23);
+      sf->partition_search_breakout_rate_thr = 120;
+    }
+  }
+
+  // If this is a two pass clip that fits the criteria for animated or
+  // graphics content then reset disable_split_mask for speeds 1-4.
+  // Also if the image edge is internal to the coded area.
+  if ((speed >= 1) && (cpi->oxcf.pass == 2) &&
+      ((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ||
+       (av1_internal_image_edge(cpi)))) {
+    sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
+  }
+
+  if (speed >= 4) {
+    if (AOMMIN(cm->width, cm->height) >= 720) {
+      sf->partition_search_breakout_dist_thr = (1 << 26);
+    } else {
+      sf->partition_search_breakout_dist_thr = (1 << 24);
+    }
+    sf->disable_split_mask = DISABLE_ALL_SPLIT;
+  }
+}
+
+static void set_good_speed_feature(AV1_COMP *cpi, AV1_COMMON *cm,
+                                   SPEED_FEATURES *sf, int speed) {
+  const int boosted = frame_is_boosted(cpi);
+
+  if (speed >= 1) {
+    sf->tx_type_search.fast_intra_tx_type_search = 1;
+    sf->tx_type_search.fast_inter_tx_type_search = 1;
+  }
+
+  if (speed >= 2) {
+    if ((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ||
+        av1_internal_image_edge(cpi)) {
+      sf->use_square_partition_only = !frame_is_boosted(cpi);
+    } else {
+      sf->use_square_partition_only = !frame_is_intra_only(cm);
+    }
+
+    sf->less_rectangular_check = 1;
+
+    sf->use_rd_breakout = 1;
+    sf->adaptive_motion_search = 1;
+    sf->mv.auto_mv_step_size = 1;
+    sf->adaptive_rd_thresh = 1;
+    sf->mv.subpel_iters_per_step = 1;
+    sf->mode_skip_start = 10;
+    sf->adaptive_pred_interp_filter = 1;
+
+    sf->recode_loop = ALLOW_RECODE_KFARFGF;
+#if CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_64X64] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_64X64] = INTRA_DC_H_V;
+#endif  // CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
+    sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
+    sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
+
+    sf->tx_size_search_breakout = 1;
+    sf->partition_search_breakout_rate_thr = 80;
+    sf->tx_type_search.prune_mode = PRUNE_ONE;
+    // Use transform domain distortion.
+    // Note var-tx expt always uses pixel domain distortion.
+    sf->use_transform_domain_distortion = 1;
+#if CONFIG_EXT_INTER
+    sf->disable_wedge_search_var_thresh = 100;
+    sf->fast_wedge_sign_estimate = 1;
+#endif  // CONFIG_EXT_INTER
+  }
+
+  if (speed >= 3) {
+    sf->tx_size_search_method =
+        frame_is_boosted(cpi) ? USE_FULL_RD : USE_LARGESTALL;
+    sf->mode_search_skip_flags =
+        (cm->frame_type == KEY_FRAME)
+            ? 0
+            : FLAG_SKIP_INTRA_DIRMISMATCH | FLAG_SKIP_INTRA_BESTINTER |
+                  FLAG_SKIP_COMP_BESTINTRA | FLAG_SKIP_INTRA_LOWVAR;
+    sf->disable_filter_search_var_thresh = 100;
+    sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
+    sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
+    sf->allow_partition_search_skip = 1;
+    sf->use_upsampled_references = 0;
+    sf->adaptive_rd_thresh = 2;
+#if CONFIG_EXT_TX
+    sf->tx_type_search.prune_mode = PRUNE_TWO;
+#endif
+  }
+
+  if (speed >= 4) {
+    sf->use_square_partition_only = !frame_is_intra_only(cm);
+    sf->tx_size_search_method =
+        frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL;
+    sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED;
+    sf->adaptive_pred_interp_filter = 0;
+    sf->adaptive_mode_search = 1;
+    sf->cb_partition_search = !boosted;
+    sf->cb_pred_filter_search = 1;
+    sf->alt_ref_search_fp = 1;
+    sf->recode_loop = ALLOW_RECODE_KFMAXBW;
+    sf->adaptive_rd_thresh = 3;
+    sf->mode_skip_start = 6;
+#if CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_64X64] = INTRA_DC;
+    sf->intra_uv_mode_mask[TX_64X64] = INTRA_DC;
+#endif  // CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
+    sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC;
+    sf->adaptive_interp_filter_search = 1;
+  }
+
+  if (speed >= 5) {
+    sf->use_square_partition_only = 1;
+    sf->tx_size_search_method = USE_LARGESTALL;
+    sf->mv.search_method = BIGDIA;
+    sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
+    sf->adaptive_rd_thresh = 4;
+    if (cm->frame_type != KEY_FRAME)
+      sf->mode_search_skip_flags |= FLAG_EARLY_TERMINATE;
+    sf->disable_filter_search_var_thresh = 200;
+    sf->use_fast_coef_updates = ONE_LOOP_REDUCED;
+    sf->use_fast_coef_costing = 1;
+    sf->partition_search_breakout_rate_thr = 300;
+  }
+
+  if (speed >= 6) {
+    int i;
+    sf->optimize_coefficients = 0;
+    sf->mv.search_method = HEX;
+    sf->disable_filter_search_var_thresh = 500;
+    for (i = 0; i < TX_SIZES; ++i) {
+      sf->intra_y_mode_mask[i] = INTRA_DC;
+      sf->intra_uv_mode_mask[i] = INTRA_DC;
+    }
+    sf->partition_search_breakout_rate_thr = 500;
+    sf->mv.reduce_first_step_size = 1;
+    sf->simple_model_rd_from_var = 1;
+  }
+  if (speed >= 7) {
+    const int is_keyframe = cm->frame_type == KEY_FRAME;
+    const int frames_since_key = is_keyframe ? 0 : cpi->rc.frames_since_key;
+    sf->default_max_partition_size = BLOCK_32X32;
+    sf->default_min_partition_size = BLOCK_8X8;
+#if CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_64X64] = INTRA_DC;
+#endif  // CONFIG_TX64X64
+    sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
+    sf->frame_parameter_update = 0;
+    sf->mv.search_method = FAST_HEX;
+    sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEAR_NEW;
+    sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST;
+#if CONFIG_EXT_PARTITION
+    sf->inter_mode_mask[BLOCK_64X128] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_128X64] = INTER_NEAREST;
+    sf->inter_mode_mask[BLOCK_128X128] = INTER_NEAREST;
+#endif  // CONFIG_EXT_PARTITION
+    sf->partition_search_type = REFERENCE_PARTITION;
+    sf->default_min_partition_size = BLOCK_8X8;
+    sf->reuse_inter_pred_sby = 1;
+    sf->force_frame_boost =
+        is_keyframe ||
+        (frames_since_key % (sf->last_partitioning_redo_frequency << 1) == 1);
+    sf->max_delta_qindex = is_keyframe ? 20 : 15;
+    sf->coeff_prob_appx_step = 4;
+    sf->mode_search_skip_flags |= FLAG_SKIP_INTRA_DIRMISMATCH;
+  }
+}
+
+void av1_set_speed_features_framesize_dependent(AV1_COMP *cpi) {
+  SPEED_FEATURES *const sf = &cpi->sf;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  AV1_COMMON *const cm = &cpi->common;
+  RD_OPT *const rd = &cpi->rd;
+  int i;
+
+// Limit memory usage for high resolutions
+#if CONFIG_EXT_REFS
+  // TODO(zoeliu): Temporary solution to resolve the insufficient RAM issue for
+  //               ext-refs. Need to work with @yunqingwang to have a more
+  //               effective solution.
+  if (AOMMIN(cm->width, cm->height) > 720) {
+    // Turn off the use of upsampled references for HD resolution
+    sf->use_upsampled_references = 0;
+  } else if ((AOMMIN(cm->width, cm->height) > 540) &&
+             (oxcf->profile != PROFILE_0)) {
+    sf->use_upsampled_references = 0;
+  }
+#else
+  if (AOMMIN(cm->width, cm->height) > 1080) {
+    sf->use_upsampled_references = 0;
+  } else if ((AOMMIN(cm->width, cm->height) > 720) &&
+             (oxcf->profile != PROFILE_0)) {
+    sf->use_upsampled_references = 0;
+  }
+#endif  // CONFIG_EXT_REFS
+
+  if (oxcf->mode == GOOD) {
+    set_good_speed_feature_framesize_dependent(cpi, sf, oxcf->speed);
+  }
+
+  if (sf->disable_split_mask == DISABLE_ALL_SPLIT) {
+    sf->adaptive_pred_interp_filter = 0;
+  }
+
+  // Check for masked out split cases.
+  for (i = 0; i < MAX_REFS; ++i) {
+    if (sf->disable_split_mask & (1 << i)) {
+      rd->thresh_mult_sub8x8[i] = INT_MAX;
+    }
+  }
+
+  // This is only used in motion vector unit test.
+  if (cpi->oxcf.motion_vector_unit_test == 1)
+    cpi->find_fractional_mv_step = av1_return_max_sub_pixel_mv;
+  else if (cpi->oxcf.motion_vector_unit_test == 2)
+    cpi->find_fractional_mv_step = av1_return_min_sub_pixel_mv;
+}
+
+void av1_set_speed_features_framesize_independent(AV1_COMP *cpi) {
+  SPEED_FEATURES *const sf = &cpi->sf;
+  AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &cpi->td.mb;
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  int i;
+
+  // best quality defaults
+  sf->frame_parameter_update = 1;
+  sf->mv.search_method = NSTEP;
+  sf->recode_loop = ALLOW_RECODE;
+  sf->mv.subpel_search_method = SUBPEL_TREE;
+  sf->mv.subpel_iters_per_step = 2;
+  sf->mv.subpel_force_stop = 0;
+  sf->optimize_coefficients = !is_lossless_requested(&cpi->oxcf);
+  sf->mv.reduce_first_step_size = 0;
+  sf->coeff_prob_appx_step = 1;
+  sf->mv.auto_mv_step_size = 0;
+  sf->mv.fullpel_search_step_param = 6;
+  sf->comp_inter_joint_search_thresh = BLOCK_4X4;
+  sf->adaptive_rd_thresh = 0;
+  sf->tx_size_search_method = USE_FULL_RD;
+  sf->adaptive_motion_search = 0;
+  sf->adaptive_pred_interp_filter = 0;
+  sf->adaptive_mode_search = 0;
+  sf->cb_pred_filter_search = 0;
+  sf->cb_partition_search = 0;
+  sf->alt_ref_search_fp = 0;
+  sf->partition_search_type = SEARCH_PARTITION;
+  sf->tx_type_search.prune_mode = NO_PRUNE;
+  sf->tx_type_search.fast_intra_tx_type_search = 0;
+  sf->tx_type_search.fast_inter_tx_type_search = 0;
+  sf->less_rectangular_check = 0;
+  sf->use_square_partition_only = 0;
+  sf->auto_min_max_partition_size = NOT_IN_USE;
+  sf->rd_auto_partition_min_limit = BLOCK_4X4;
+  sf->default_max_partition_size = BLOCK_LARGEST;
+  sf->default_min_partition_size = BLOCK_4X4;
+  sf->adjust_partitioning_from_last_frame = 0;
+  sf->last_partitioning_redo_frequency = 4;
+  sf->disable_split_mask = 0;
+  sf->mode_search_skip_flags = 0;
+  sf->force_frame_boost = 0;
+  sf->max_delta_qindex = 0;
+  sf->disable_filter_search_var_thresh = 0;
+  sf->adaptive_interp_filter_search = 0;
+  sf->allow_partition_search_skip = 0;
+  sf->use_upsampled_references = 1;
+#if CONFIG_EXT_INTER
+  sf->disable_wedge_search_var_thresh = 0;
+  sf->fast_wedge_sign_estimate = 0;
+#endif  // CONFIG_EXT_INTER
+
+  for (i = 0; i < TX_SIZES; i++) {
+    sf->intra_y_mode_mask[i] = INTRA_ALL;
+    sf->intra_uv_mode_mask[i] = INTRA_ALL;
+  }
+  sf->use_rd_breakout = 0;
+  sf->lpf_pick = LPF_PICK_FROM_FULL_IMAGE;
+  sf->use_fast_coef_updates = TWO_LOOP;
+  sf->use_fast_coef_costing = 0;
+  sf->mode_skip_start = MAX_MODES;  // Mode index at which mode skip mask set
+  sf->schedule_mode_search = 0;
+  for (i = 0; i < BLOCK_SIZES; ++i) sf->inter_mode_mask[i] = INTER_ALL;
+  sf->max_intra_bsize = BLOCK_LARGEST;
+  sf->reuse_inter_pred_sby = 0;
+  // This setting only takes effect when partition_search_type is set
+  // to FIXED_PARTITION.
+  sf->always_this_block_size = BLOCK_16X16;
+  sf->search_type_check_frequency = 50;
+  // Recode loop tolerance %.
+  sf->recode_tolerance = 25;
+  sf->default_interp_filter = SWITCHABLE;
+  sf->tx_size_search_breakout = 0;
+  sf->partition_search_breakout_dist_thr = 0;
+  sf->partition_search_breakout_rate_thr = 0;
+  sf->simple_model_rd_from_var = 0;
+
+  // Set this at the appropriate speed levels
+  sf->use_transform_domain_distortion = 0;
+
+  if (oxcf->mode == GOOD
+#if CONFIG_XIPHRC
+      || oxcf->pass == 1
+#endif
+      )
+    set_good_speed_feature(cpi, cm, sf, oxcf->speed);
+
+  // sf->partition_search_breakout_dist_thr is set assuming max 64x64
+  // blocks. Normalise this if the blocks are bigger.
+  if (MAX_SB_SIZE_LOG2 > 6) {
+    sf->partition_search_breakout_dist_thr <<= 2 * (MAX_SB_SIZE_LOG2 - 6);
+  }
+
+  cpi->full_search_sad = av1_full_search_sad;
+  cpi->diamond_search_sad = av1_diamond_search_sad;
+
+  sf->allow_exhaustive_searches = 1;
+  int speed = (oxcf->speed > MAX_MESH_SPEED) ? MAX_MESH_SPEED : oxcf->speed;
+  if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION)
+    sf->exhaustive_searches_thresh = (1 << 24);
+  else
+    sf->exhaustive_searches_thresh = (1 << 25);
+  sf->max_exaustive_pct = good_quality_max_mesh_pct[speed];
+  if (speed > 0)
+    sf->exhaustive_searches_thresh = sf->exhaustive_searches_thresh << 1;
+
+  for (i = 0; i < MAX_MESH_STEP; ++i) {
+    sf->mesh_patterns[i].range = good_quality_mesh_patterns[speed][i].range;
+    sf->mesh_patterns[i].interval =
+        good_quality_mesh_patterns[speed][i].interval;
+  }
+#if CONFIG_INTRABC
+  if ((frame_is_intra_only(cm) && cm->allow_screen_content_tools) &&
+      (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION ||
+       cpi->oxcf.content == AOM_CONTENT_SCREEN)) {
+    for (i = 0; i < MAX_MESH_STEP; ++i) {
+      sf->mesh_patterns[i].range = intrabc_mesh_patterns[speed][i].range;
+      sf->mesh_patterns[i].interval = intrabc_mesh_patterns[speed][i].interval;
+    }
+    sf->max_exaustive_pct = intrabc_max_mesh_pct[speed];
+  }
+#endif  // CONFIG_INTRABC
+
+#if !CONFIG_XIPHRC
+  // Slow quant, dct and trellis not worthwhile for first pass
+  // so make sure they are always turned off.
+  if (oxcf->pass == 1) sf->optimize_coefficients = 0;
+#endif
+
+  // No recode for 1 pass.
+  if (oxcf->pass == 0) {
+    sf->recode_loop = DISALLOW_RECODE;
+    sf->optimize_coefficients = 0;
+  }
+
+  if (sf->mv.subpel_search_method == SUBPEL_TREE) {
+    cpi->find_fractional_mv_step = av1_find_best_sub_pixel_tree;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED) {
+    cpi->find_fractional_mv_step = av1_find_best_sub_pixel_tree_pruned;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_MORE) {
+    cpi->find_fractional_mv_step = av1_find_best_sub_pixel_tree_pruned_more;
+  } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_EVENMORE) {
+    cpi->find_fractional_mv_step = av1_find_best_sub_pixel_tree_pruned_evenmore;
+  }
+
+#if !CONFIG_AOM_QM
+  x->optimize = sf->optimize_coefficients == 1 && oxcf->pass != 1;
+#else
+  // FIXME: trellis not very efficient for quantisation matrices
+  x->optimize = 0;
+#endif
+
+  x->min_partition_size = sf->default_min_partition_size;
+  x->max_partition_size = sf->default_max_partition_size;
+
+  if (!cpi->oxcf.frame_periodic_boost) {
+    sf->max_delta_qindex = 0;
+  }
+
+  // This is only used in motion vector unit test.
+  if (cpi->oxcf.motion_vector_unit_test == 1)
+    cpi->find_fractional_mv_step = av1_return_max_sub_pixel_mv;
+  else if (cpi->oxcf.motion_vector_unit_test == 2)
+    cpi->find_fractional_mv_step = av1_return_min_sub_pixel_mv;
+}
diff --git a/third_party/aom/av1/encoder/speed_features.h b/third_party/aom/av1/encoder/speed_features.h
new file mode 100644
index 0000000000..af54a1a9ae
--- /dev/null
+++ b/third_party/aom/av1/encoder/speed_features.h
@@ -0,0 +1,484 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_SPEED_FEATURES_H_
+#define AV1_ENCODER_SPEED_FEATURES_H_
+
+#include "av1/common/enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+enum {
+  INTRA_ALL = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED) | (1 << D45_PRED) |
+              (1 << D135_PRED) | (1 << D117_PRED) | (1 << D153_PRED) |
+              (1 << D207_PRED) | (1 << D63_PRED) |
+#if CONFIG_ALT_INTRA
+              (1 << SMOOTH_PRED) |
+#endif  // CONFIG_ALT_INTRA
+              (1 << TM_PRED),
+  INTRA_DC = (1 << DC_PRED),
+  INTRA_DC_TM = (1 << DC_PRED) | (1 << TM_PRED),
+  INTRA_DC_H_V = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
+  INTRA_DC_TM_H_V =
+      (1 << DC_PRED) | (1 << TM_PRED) | (1 << V_PRED) | (1 << H_PRED)
+};
+
+#if CONFIG_EXT_INTER
+enum {
+  INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) | (1 << NEWMV) |
+              (1 << NEAREST_NEARESTMV) | (1 << NEAR_NEARMV) |
+              (1 << NEAREST_NEARMV) | (1 << NEAR_NEARESTMV) | (1 << NEW_NEWMV) |
+              (1 << NEAREST_NEWMV) | (1 << NEAR_NEWMV) | (1 << NEW_NEARMV) |
+              (1 << NEW_NEARESTMV) | (1 << ZERO_ZEROMV),
+  INTER_NEAREST = (1 << NEARESTMV) | (1 << NEAREST_NEARESTMV) |
+                  (1 << NEAREST_NEARMV) | (1 << NEAR_NEARESTMV) |
+                  (1 << NEW_NEARESTMV) | (1 << NEAREST_NEWMV),
+  INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV) |
+                      (1 << NEAREST_NEARESTMV) | (1 << NEW_NEWMV) |
+                      (1 << NEAR_NEARESTMV) | (1 << NEAREST_NEARMV) |
+                      (1 << NEW_NEARESTMV) | (1 << NEAREST_NEWMV) |
+                      (1 << NEW_NEARMV) | (1 << NEAR_NEWMV),
+  INTER_NEAREST_ZERO = (1 << NEARESTMV) | (1 << ZEROMV) |
+                       (1 << NEAREST_NEARESTMV) | (1 << ZERO_ZEROMV) |
+                       (1 << NEAREST_NEARMV) | (1 << NEAR_NEARESTMV) |
+                       (1 << NEAREST_NEWMV) | (1 << NEW_NEARESTMV),
+  INTER_NEAREST_NEW_ZERO =
+      (1 << NEARESTMV) | (1 << ZEROMV) | (1 << NEWMV) |
+      (1 << NEAREST_NEARESTMV) | (1 << ZERO_ZEROMV) | (1 << NEW_NEWMV) |
+      (1 << NEAREST_NEARMV) | (1 << NEAR_NEARESTMV) | (1 << NEW_NEARESTMV) |
+      (1 << NEAREST_NEWMV) | (1 << NEW_NEARMV) | (1 << NEAR_NEWMV),
+  INTER_NEAREST_NEAR_NEW =
+      (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV) |
+      (1 << NEAREST_NEARESTMV) | (1 << NEW_NEWMV) | (1 << NEAREST_NEARMV) |
+      (1 << NEAR_NEARESTMV) | (1 << NEW_NEARESTMV) | (1 << NEAREST_NEWMV) |
+      (1 << NEW_NEARMV) | (1 << NEAR_NEWMV) | (1 << NEAR_NEARMV),
+  INTER_NEAREST_NEAR_ZERO =
+      (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) |
+      (1 << NEAREST_NEARESTMV) | (1 << ZERO_ZEROMV) | (1 << NEAREST_NEARMV) |
+      (1 << NEAR_NEARESTMV) | (1 << NEAREST_NEWMV) | (1 << NEW_NEARESTMV) |
+      (1 << NEW_NEARMV) | (1 << NEAR_NEWMV) | (1 << NEAR_NEARMV),
+};
+#else
+enum {
+  INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) | (1 << NEWMV),
+  INTER_NEAREST = (1 << NEARESTMV),
+  INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV),
+  INTER_NEAREST_ZERO = (1 << NEARESTMV) | (1 << ZEROMV),
+  INTER_NEAREST_NEW_ZERO = (1 << NEARESTMV) | (1 << ZEROMV) | (1 << NEWMV),
+  INTER_NEAREST_NEAR_NEW = (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV),
+  INTER_NEAREST_NEAR_ZERO = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV),
+};
+#endif  // CONFIG_EXT_INTER
+
+enum {
+  DISABLE_ALL_INTER_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
+                            (1 << THR_ALTR) | (1 << THR_GOLD) | (1 << THR_LAST),
+
+  DISABLE_ALL_SPLIT = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
+
+  DISABLE_COMPOUND_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
+
+  LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
+                              (1 << THR_ALTR) | (1 << THR_GOLD)
+};
+
+typedef enum {
+  DIAMOND = 0,
+  NSTEP = 1,
+  HEX = 2,
+  BIGDIA = 3,
+  SQUARE = 4,
+  FAST_HEX = 5,
+  FAST_DIAMOND = 6
+} SEARCH_METHODS;
+
+typedef enum {
+  // No recode.
+  DISALLOW_RECODE = 0,
+  // Allow recode for KF and exceeding maximum frame bandwidth.
+  ALLOW_RECODE_KFMAXBW = 1,
+  // Allow recode only for KF/ARF/GF frames.
+  ALLOW_RECODE_KFARFGF = 2,
+  // Allow recode for all frames based on bitrate constraints.
+  ALLOW_RECODE = 3,
+} RECODE_LOOP_TYPE;
+
+typedef enum {
+  SUBPEL_TREE = 0,
+  SUBPEL_TREE_PRUNED = 1,           // Prunes 1/2-pel searches
+  SUBPEL_TREE_PRUNED_MORE = 2,      // Prunes 1/2-pel searches more aggressively
+  SUBPEL_TREE_PRUNED_EVENMORE = 3,  // Prunes 1/2- and 1/4-pel searches
+  // Other methods to come
+} SUBPEL_SEARCH_METHODS;
+
+typedef enum {
+  NO_MOTION_THRESHOLD = 0,
+  LOW_MOTION_THRESHOLD = 7
+} MOTION_THRESHOLD;
+
+typedef enum {
+  USE_FULL_RD = 0,
+  USE_LARGESTALL,
+  USE_TX_8X8
+} TX_SIZE_SEARCH_METHOD;
+
+typedef enum {
+  NOT_IN_USE = 0,
+  RELAXED_NEIGHBORING_MIN_MAX = 1
+} AUTO_MIN_MAX_MODE;
+
+typedef enum {
+  // Try the full image with different values.
+  LPF_PICK_FROM_FULL_IMAGE,
+  // Try a small portion of the image with different values.
+  LPF_PICK_FROM_SUBIMAGE,
+  // Estimate the level based on quantizer and frame type
+  LPF_PICK_FROM_Q,
+  // Pick 0 to disable LPF if LPF was enabled last frame
+  LPF_PICK_MINIMAL_LPF
+} LPF_PICK_METHOD;
+
+typedef enum {
+  // Terminate search early based on distortion so far compared to
+  // qp step, distortion in the neighborhood of the frame, etc.
+  FLAG_EARLY_TERMINATE = 1 << 0,
+
+  // Skips comp inter modes if the best so far is an intra mode.
+  FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
+
+  // Skips oblique intra modes if the best so far is an inter mode.
+  FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
+
+  // Skips oblique intra modes  at angles 27, 63, 117, 153 if the best
+  // intra so far is not one of the neighboring directions.
+  FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
+
+  // Skips intra modes other than DC_PRED if the source variance is small
+  FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
+} MODE_SEARCH_SKIP_LOGIC;
+
+typedef enum {
+  FLAG_SKIP_EIGHTTAP_REGULAR = 1 << EIGHTTAP_REGULAR,
+  FLAG_SKIP_EIGHTTAP_SMOOTH = 1 << EIGHTTAP_SMOOTH,
+  FLAG_SKIP_MULTITAP_SHARP = 1 << MULTITAP_SHARP,
+} INTERP_FILTER_MASK;
+
+typedef enum {
+  NO_PRUNE = 0,
+  // eliminates one tx type in vertical and horizontal direction
+  PRUNE_ONE = 1,
+#if CONFIG_EXT_TX
+  // eliminates two tx types in each direction
+  PRUNE_TWO = 2,
+#endif
+} TX_TYPE_PRUNE_MODE;
+
+typedef struct {
+  TX_TYPE_PRUNE_MODE prune_mode;
+  int fast_intra_tx_type_search;
+  int fast_inter_tx_type_search;
+} TX_TYPE_SEARCH;
+
+typedef enum {
+  // Search partitions using RD criterion
+  SEARCH_PARTITION,
+
+  // Always use a fixed size partition
+  FIXED_PARTITION,
+
+  REFERENCE_PARTITION,
+
+  // Use an arbitrary partitioning scheme based on source variance within
+  // a 64X64 SB
+  VAR_BASED_PARTITION,
+
+  // Use non-fixed partitions based on source variance
+  SOURCE_VAR_BASED_PARTITION
+} PARTITION_SEARCH_TYPE;
+
+typedef enum {
+  // Does a dry run to see if any of the contexts need to be updated or not,
+  // before the final run.
+  TWO_LOOP = 0,
+
+  // No dry run, also only half the coef contexts and bands are updated.
+  // The rest are not updated at all.
+  ONE_LOOP_REDUCED = 1
+} FAST_COEFF_UPDATE;
+
+typedef struct MV_SPEED_FEATURES {
+  // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
+  SEARCH_METHODS search_method;
+
+  // This parameter controls which step in the n-step process we start at.
+  // It's changed adaptively based on circumstances.
+  int reduce_first_step_size;
+
+  // If this is set to 1, we limit the motion search range to 2 times the
+  // largest motion vector found in the last frame.
+  int auto_mv_step_size;
+
+  // Subpel_search_method can only be subpel_tree which does a subpixel
+  // logarithmic search that keeps stepping at 1/2 pixel units until
+  // you stop getting a gain, and then goes on to 1/4 and repeats
+  // the same process. Along the way it skips many diagonals.
+  SUBPEL_SEARCH_METHODS subpel_search_method;
+
+  // Maximum number of steps in logarithmic subpel search before giving up.
+  int subpel_iters_per_step;
+
+  // Control when to stop subpel search
+  int subpel_force_stop;
+
+  // This variable sets the step_param used in full pel motion search.
+  int fullpel_search_step_param;
+} MV_SPEED_FEATURES;
+
+#define MAX_MESH_STEP 4
+
+typedef struct MESH_PATTERN {
+  int range;
+  int interval;
+} MESH_PATTERN;
+
+typedef struct SPEED_FEATURES {
+  MV_SPEED_FEATURES mv;
+
+  // Frame level coding parameter update
+  int frame_parameter_update;
+
+  RECODE_LOOP_TYPE recode_loop;
+
+  // Trellis (dynamic programming) optimization of quantized values (+1, 0).
+  int optimize_coefficients;
+
+  // Always set to 0. If on it enables 0 cost background transmission
+  // (except for the initial transmission of the segmentation). The feature is
+  // disabled because the addition of very large block sizes make the
+  // backgrounds very to cheap to encode, and the segmentation we have
+  // adds overhead.
+  int static_segmentation;
+
+  // If 1 we iterate finding a best reference for 2 ref frames together - via
+  // a log search that iterates 4 times (check around mv for last for best
+  // error of combined predictor then check around mv for alt). If 0 we
+  // we just use the best motion vector found for each frame by itself.
+  BLOCK_SIZE comp_inter_joint_search_thresh;
+
+  // This variable is used to cap the maximum number of times we skip testing a
+  // mode to be evaluated. A high value means we will be faster.
+  int adaptive_rd_thresh;
+
+  // Coefficient probability model approximation step size
+  int coeff_prob_appx_step;
+
+  // The threshold is to determine how slow the motino is, it is used when
+  // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
+  MOTION_THRESHOLD lf_motion_threshold;
+
+  // Determine which method we use to determine transform size. We can choose
+  // between options like full rd, largest for prediction size, largest
+  // for intra and model coefs for the rest.
+  TX_SIZE_SEARCH_METHOD tx_size_search_method;
+
+  // After looking at the first set of modes (set by index here), skip
+  // checking modes for reference frames that don't match the reference frame
+  // of the best so far.
+  int mode_skip_start;
+
+  PARTITION_SEARCH_TYPE partition_search_type;
+
+  TX_TYPE_SEARCH tx_type_search;
+
+  // Used if partition_search_type = FIXED_SIZE_PARTITION
+  BLOCK_SIZE always_this_block_size;
+
+  // Skip rectangular partition test when partition type none gives better
+  // rd than partition type split.
+  int less_rectangular_check;
+
+  // Disable testing non square partitions. (eg 16x32)
+  int use_square_partition_only;
+
+  // Sets min and max partition sizes for this superblock based on the
+  // same superblock in last encoded frame, and the left and above neighbor.
+  AUTO_MIN_MAX_MODE auto_min_max_partition_size;
+  // Ensures the rd based auto partition search will always
+  // go down at least to the specified level.
+  BLOCK_SIZE rd_auto_partition_min_limit;
+
+  // Min and max partition size we enable (block_size) as per auto
+  // min max, but also used by adjust partitioning, and pick_partitioning.
+  BLOCK_SIZE default_min_partition_size;
+  BLOCK_SIZE default_max_partition_size;
+
+  // Whether or not we allow partitions one smaller or one greater than the last
+  // frame's partitioning. Only used if use_lastframe_partitioning is set.
+  int adjust_partitioning_from_last_frame;
+
+  // How frequently we re do the partitioning from scratch. Only used if
+  // use_lastframe_partitioning is set.
+  int last_partitioning_redo_frequency;
+
+  // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable
+  // it always, to allow it for only Last frame and Intra, disable it for all
+  // inter modes or to enable it always.
+  int disable_split_mask;
+
+  // TODO(jingning): combine the related motion search speed features
+  // This allows us to use motion search at other sizes as a starting
+  // point for this motion search and limits the search range around it.
+  int adaptive_motion_search;
+
+  // Flag for allowing some use of exhaustive searches;
+  int allow_exhaustive_searches;
+
+  // Threshold for allowing exhaistive motion search.
+  int exhaustive_searches_thresh;
+
+  // Maximum number of exhaustive searches for a frame.
+  int max_exaustive_pct;
+
+  // Pattern to be used for any exhaustive mesh searches.
+  MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
+
+  int schedule_mode_search;
+
+  // Allows sub 8x8 modes to use the prediction filter that was determined
+  // best for 8x8 mode. If set to 0 we always re check all the filters for
+  // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter
+  // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
+  int adaptive_pred_interp_filter;
+
+  // Adaptive prediction mode search
+  int adaptive_mode_search;
+
+  // Chessboard pattern prediction filter type search
+  int cb_pred_filter_search;
+
+  int cb_partition_search;
+
+  int alt_ref_search_fp;
+
+  // Use finer quantizer in every other few frames that run variable block
+  // partition type search.
+  int force_frame_boost;
+
+  // Maximally allowed base quantization index fluctuation.
+  int max_delta_qindex;
+
+  // Implements various heuristics to skip searching modes
+  // The heuristics selected are based on  flags
+  // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
+  unsigned int mode_search_skip_flags;
+
+  // A source variance threshold below which filter search is disabled
+  // Choose a very large value (UINT_MAX) to use 8-tap always
+  unsigned int disable_filter_search_var_thresh;
+
+#if CONFIG_EXT_INTER
+  // A source variance threshold below which wedge search is disabled
+  unsigned int disable_wedge_search_var_thresh;
+
+  // Whether fast wedge sign estimate is used
+  int fast_wedge_sign_estimate;
+#endif  // CONFIG_EXT_INTER
+
+  // These bit masks allow you to enable or disable intra modes for each
+  // transform size separately.
+  int intra_y_mode_mask[TX_SIZES];
+  int intra_uv_mode_mask[TX_SIZES];
+
+  // These bit masks allow you to enable or disable intra modes for each
+  // prediction block size separately.
+  int intra_y_mode_bsize_mask[BLOCK_SIZES];
+
+  // This variable enables an early break out of mode testing if the model for
+  // rd built from the prediction signal indicates a value that's much
+  // higher than the best rd we've seen so far.
+  int use_rd_breakout;
+
+  // This feature controls how the loop filter level is determined.
+  LPF_PICK_METHOD lpf_pick;
+
+  // This feature limits the number of coefficients updates we actually do
+  // by only looking at counts from 1/2 the bands.
+  FAST_COEFF_UPDATE use_fast_coef_updates;
+
+  // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
+  // modes are used in order from LSB to MSB for each BLOCK_SIZE.
+  int inter_mode_mask[BLOCK_SIZES];
+
+  // This feature controls whether we do the expensive context update and
+  // calculation in the rd coefficient costing loop.
+  int use_fast_coef_costing;
+
+  // This feature controls the tolerence vs target used in deciding whether to
+  // recode a frame. It has no meaning if recode is disabled.
+  int recode_tolerance;
+
+  // This variable controls the maximum block size where intra blocks can be
+  // used in inter frames.
+  // TODO(aconverse): Fold this into one of the other many mode skips
+  BLOCK_SIZE max_intra_bsize;
+
+  // The frequency that we check if SOURCE_VAR_BASED_PARTITION or
+  // FIXED_PARTITION search type should be used.
+  int search_type_check_frequency;
+
+  // When partition is pre-set, the inter prediction result from pick_inter_mode
+  // can be reused in final block encoding process. It is enabled only for real-
+  // time mode speed 6.
+  int reuse_inter_pred_sby;
+
+  // default interp filter choice
+  InterpFilter default_interp_filter;
+
+  // Early termination in transform size search, which only applies while
+  // tx_size_search_method is USE_FULL_RD.
+  int tx_size_search_breakout;
+
+  // adaptive interp_filter search to allow skip of certain filter types.
+  int adaptive_interp_filter_search;
+
+  // mask for skip evaluation of certain interp_filter type.
+  INTERP_FILTER_MASK interp_filter_search_mask;
+
+  // Partition search early breakout thresholds.
+  int64_t partition_search_breakout_dist_thr;
+  int partition_search_breakout_rate_thr;
+
+  // Allow skipping partition search for still image frame
+  int allow_partition_search_skip;
+
+  // Fast approximation of av1_model_rd_from_var_lapndz
+  int simple_model_rd_from_var;
+
+  // Do sub-pixel search in up-sampled reference frames
+  int use_upsampled_references;
+
+  // Whether to compute distortion in the image domain (slower but
+  // more accurate), or in the transform domain (faster but less acurate).
+  int use_transform_domain_distortion;
+} SPEED_FEATURES;
+
+struct AV1_COMP;
+
+void av1_set_speed_features_framesize_independent(struct AV1_COMP *cpi);
+void av1_set_speed_features_framesize_dependent(struct AV1_COMP *cpi);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_SPEED_FEATURES_H_
diff --git a/third_party/aom/av1/encoder/subexp.c b/third_party/aom/av1/encoder/subexp.c
new file mode 100644
index 0000000000..8960d33414
--- /dev/null
+++ b/third_party/aom/av1/encoder/subexp.c
@@ -0,0 +1,282 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "aom_dsp/bitwriter.h"
+
+#include "av1/common/common.h"
+#include "av1/common/entropy.h"
+#include "av1/encoder/cost.h"
+#include "av1/encoder/subexp.h"
+
+static const uint8_t update_bits[255] = {
+  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  5,  6,  6,  6,
+  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  6,  8,  8,  8,  8,  8,  8,
+  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+  8,  8,  8,  8,  8,  8,  8,  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
+  10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11, 11,
+  11, 11, 11, 11, 11, 11, 11, 0,
+};
+#define MIN_DELP_BITS 5
+
+static int recenter_nonneg(int v, int m) {
+  if (v > (m << 1))
+    return v;
+  else if (v >= m)
+    return ((v - m) << 1);
+  else
+    return ((m - v) << 1) - 1;
+}
+
+static int remap_prob(int v, int m) {
+  int i;
+  static const uint8_t map_table[MAX_PROB - 1] = {
+    // generated by:
+    //   map_table[j] = split_index(j, MAX_PROB - 1, MODULUS_PARAM);
+    20,  21,  22,  23,  24,  25,  0,   26,  27,  28,  29,  30,  31,  32,  33,
+    34,  35,  36,  37,  1,   38,  39,  40,  41,  42,  43,  44,  45,  46,  47,
+    48,  49,  2,   50,  51,  52,  53,  54,  55,  56,  57,  58,  59,  60,  61,
+    3,   62,  63,  64,  65,  66,  67,  68,  69,  70,  71,  72,  73,  4,   74,
+    75,  76,  77,  78,  79,  80,  81,  82,  83,  84,  85,  5,   86,  87,  88,
+    89,  90,  91,  92,  93,  94,  95,  96,  97,  6,   98,  99,  100, 101, 102,
+    103, 104, 105, 106, 107, 108, 109, 7,   110, 111, 112, 113, 114, 115, 116,
+    117, 118, 119, 120, 121, 8,   122, 123, 124, 125, 126, 127, 128, 129, 130,
+    131, 132, 133, 9,   134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144,
+    145, 10,  146, 147, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 11,
+    158, 159, 160, 161, 162, 163, 164, 165, 166, 167, 168, 169, 12,  170, 171,
+    172, 173, 174, 175, 176, 177, 178, 179, 180, 181, 13,  182, 183, 184, 185,
+    186, 187, 188, 189, 190, 191, 192, 193, 14,  194, 195, 196, 197, 198, 199,
+    200, 201, 202, 203, 204, 205, 15,  206, 207, 208, 209, 210, 211, 212, 213,
+    214, 215, 216, 217, 16,  218, 219, 220, 221, 222, 223, 224, 225, 226, 227,
+    228, 229, 17,  230, 231, 232, 233, 234, 235, 236, 237, 238, 239, 240, 241,
+    18,  242, 243, 244, 245, 246, 247, 248, 249, 250, 251, 252, 253, 19,
+  };
+  v--;
+  m--;
+  if ((m << 1) <= MAX_PROB)
+    i = recenter_nonneg(v, m) - 1;
+  else
+    i = recenter_nonneg(MAX_PROB - 1 - v, MAX_PROB - 1 - m) - 1;
+
+  i = map_table[i];
+  return i;
+}
+
+static int prob_diff_update_cost(aom_prob newp, aom_prob oldp) {
+  int delp = remap_prob(newp, oldp);
+  return update_bits[delp] << AV1_PROB_COST_SHIFT;
+}
+
+static void encode_uniform(aom_writer *w, int v) {
+  const int l = 8;
+  const int m = (1 << l) - 190;
+  if (v < m) {
+    aom_write_literal(w, v, l - 1);
+  } else {
+    aom_write_literal(w, m + ((v - m) >> 1), l - 1);
+    aom_write_literal(w, (v - m) & 1, 1);
+  }
+}
+
+static INLINE int write_bit_gte(aom_writer *w, int word, int test) {
+  aom_write_literal(w, word >= test, 1);
+  return word >= test;
+}
+
+static void encode_term_subexp(aom_writer *w, int word) {
+  if (!write_bit_gte(w, word, 16)) {
+    aom_write_literal(w, word, 4);
+  } else if (!write_bit_gte(w, word, 32)) {
+    aom_write_literal(w, word - 16, 4);
+  } else if (!write_bit_gte(w, word, 64)) {
+    aom_write_literal(w, word - 32, 5);
+  } else {
+    encode_uniform(w, word - 64);
+  }
+}
+
+void av1_write_prob_diff_update(aom_writer *w, aom_prob newp, aom_prob oldp) {
+  const int delp = remap_prob(newp, oldp);
+  encode_term_subexp(w, delp);
+}
+
+int av1_prob_diff_update_savings_search(const unsigned int *ct, aom_prob oldp,
+                                        aom_prob *bestp, aom_prob upd,
+                                        int probwt) {
+  const uint32_t old_b = cost_branch256(ct, oldp);
+  int bestsavings = 0;
+  aom_prob newp, bestnewp = oldp;
+  const int step = *bestp > oldp ? -1 : 1;
+  const int upd_cost = av1_cost_one(upd) - av1_cost_zero(upd);
+
+  if (old_b > (uint32_t)upd_cost + (MIN_DELP_BITS << AV1_PROB_COST_SHIFT)) {
+    for (newp = *bestp; newp != oldp; newp += step) {
+      const int new_b = cost_branch256(ct, newp);
+      const int update_b = prob_diff_update_cost(newp, oldp) + upd_cost;
+      const int savings = (int)((int64_t)old_b - new_b - update_b * probwt);
+      if (savings > bestsavings) {
+        bestsavings = savings;
+        bestnewp = newp;
+      }
+    }
+  }
+  *bestp = bestnewp;
+  return bestsavings;
+}
+
+int av1_prob_diff_update_savings_search_model(const unsigned int *ct,
+                                              const aom_prob oldp,
+                                              aom_prob *bestp, aom_prob upd,
+                                              int stepsize, int probwt) {
+  int i, old_b, new_b, update_b, savings, bestsavings;
+  int newp;
+  const int step_sign = *bestp > oldp ? -1 : 1;
+  const int step = stepsize * step_sign;
+  const int upd_cost = av1_cost_one(upd) - av1_cost_zero(upd);
+  const aom_prob *newplist, *oldplist;
+  aom_prob bestnewp;
+  oldplist = av1_pareto8_full[oldp - 1];
+  old_b = cost_branch256(ct + 2 * PIVOT_NODE, oldp);
+  for (i = UNCONSTRAINED_NODES; i < ENTROPY_NODES; ++i)
+    old_b += cost_branch256(ct + 2 * i, oldplist[i - UNCONSTRAINED_NODES]);
+
+  bestsavings = 0;
+  bestnewp = oldp;
+
+  assert(stepsize > 0);
+
+  if (old_b > upd_cost + (MIN_DELP_BITS << AV1_PROB_COST_SHIFT)) {
+    for (newp = *bestp; (newp - oldp) * step_sign < 0; newp += step) {
+      if (newp < 1 || newp > 255) continue;
+      newplist = av1_pareto8_full[newp - 1];
+      new_b = cost_branch256(ct + 2 * PIVOT_NODE, newp);
+      for (i = UNCONSTRAINED_NODES; i < ENTROPY_NODES; ++i)
+        new_b += cost_branch256(ct + 2 * i, newplist[i - UNCONSTRAINED_NODES]);
+      update_b = prob_diff_update_cost(newp, oldp) + upd_cost;
+      savings = old_b - new_b - update_b * probwt;
+      if (savings > bestsavings) {
+        bestsavings = savings;
+        bestnewp = newp;
+      }
+    }
+  }
+
+  *bestp = bestnewp;
+  return bestsavings;
+}
+
+#if CONFIG_SUBFRAME_PROB_UPDATE
+static int get_cost(unsigned int ct[][2], aom_prob p, int n) {
+  int i, p0 = p;
+  unsigned int total_ct[2] = { 0, 0 };
+  int cost = 0;
+
+  for (i = 0; i <= n; ++i) {
+    cost += cost_branch256(ct[i], p);
+    total_ct[0] += ct[i][0];
+    total_ct[1] += ct[i][1];
+    if (i < n)
+      p = av1_merge_probs(p0, total_ct, COEF_COUNT_SAT, COEF_MAX_UPDATE_FACTOR);
+  }
+  return cost;
+}
+
+int av1_prob_update_search_subframe(unsigned int ct[][2], aom_prob oldp,
+                                    aom_prob *bestp, aom_prob upd, int n) {
+  const int old_b = get_cost(ct, oldp, n);
+  int bestsavings = 0;
+  const int upd_cost = av1_cost_one(upd) - av1_cost_zero(upd);
+  aom_prob newp, bestnewp = oldp;
+  const int step = *bestp > oldp ? -1 : 1;
+
+  for (newp = *bestp; newp != oldp; newp += step) {
+    const int new_b = get_cost(ct, newp, n);
+    const int update_b = prob_diff_update_cost(newp, oldp) + upd_cost;
+    const int savings = old_b - new_b - update_b;
+    if (savings > bestsavings) {
+      bestsavings = savings;
+      bestnewp = newp;
+    }
+  }
+  *bestp = bestnewp;
+  return bestsavings;
+}
+
+int av1_prob_update_search_model_subframe(
+    unsigned int ct[ENTROPY_NODES][COEF_PROBS_BUFS][2], const aom_prob *oldp,
+    aom_prob *bestp, aom_prob upd, int stepsize, int n) {
+  int i, old_b, new_b, update_b, savings, bestsavings;
+  int newp;
+  const int step_sign = *bestp > oldp[PIVOT_NODE] ? -1 : 1;
+  const int step = stepsize * step_sign;
+  const int upd_cost = av1_cost_one(upd) - av1_cost_zero(upd);
+  aom_prob bestnewp, newplist[ENTROPY_NODES], oldplist[ENTROPY_NODES];
+  av1_model_to_full_probs(oldp, oldplist);
+  memcpy(newplist, oldp, sizeof(aom_prob) * UNCONSTRAINED_NODES);
+  for (i = UNCONSTRAINED_NODES, old_b = 0; i < ENTROPY_NODES; ++i)
+    old_b += get_cost(ct[i], oldplist[i], n);
+  old_b += get_cost(ct[PIVOT_NODE], oldplist[PIVOT_NODE], n);
+
+  bestsavings = 0;
+  bestnewp = oldp[PIVOT_NODE];
+
+  assert(stepsize > 0);
+
+  for (newp = *bestp; (newp - oldp[PIVOT_NODE]) * step_sign < 0; newp += step) {
+    if (newp < 1 || newp > 255) continue;
+    newplist[PIVOT_NODE] = newp;
+    av1_model_to_full_probs(newplist, newplist);
+    for (i = UNCONSTRAINED_NODES, new_b = 0; i < ENTROPY_NODES; ++i)
+      new_b += get_cost(ct[i], newplist[i], n);
+    new_b += get_cost(ct[PIVOT_NODE], newplist[PIVOT_NODE], n);
+    update_b = prob_diff_update_cost(newp, oldp[PIVOT_NODE]) + upd_cost;
+    savings = old_b - new_b - update_b;
+    if (savings > bestsavings) {
+      bestsavings = savings;
+      bestnewp = newp;
+    }
+  }
+
+  *bestp = bestnewp;
+  return bestsavings;
+}
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+
+void av1_cond_prob_diff_update(aom_writer *w, aom_prob *oldp,
+                               const unsigned int ct[2], int probwt) {
+  const aom_prob upd = DIFF_UPDATE_PROB;
+  aom_prob newp = get_binary_prob(ct[0], ct[1]);
+  const int savings =
+      av1_prob_diff_update_savings_search(ct, *oldp, &newp, upd, probwt);
+  assert(newp >= 1);
+  if (savings > 0) {
+    aom_write(w, 1, upd);
+    av1_write_prob_diff_update(w, newp, *oldp);
+    *oldp = newp;
+  } else {
+    aom_write(w, 0, upd);
+  }
+}
+
+int av1_cond_prob_diff_update_savings(aom_prob *oldp, const unsigned int ct[2],
+                                      int probwt) {
+  const aom_prob upd = DIFF_UPDATE_PROB;
+  aom_prob newp = get_binary_prob(ct[0], ct[1]);
+  const int savings =
+      av1_prob_diff_update_savings_search(ct, *oldp, &newp, upd, probwt);
+  return savings;
+}
diff --git a/third_party/aom/av1/encoder/subexp.h b/third_party/aom/av1/encoder/subexp.h
new file mode 100644
index 0000000000..049265cb88
--- /dev/null
+++ b/third_party/aom/av1/encoder/subexp.h
@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_SUBEXP_H_
+#define AV1_ENCODER_SUBEXP_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "aom_dsp/bitwriter.h"
+#include "aom_dsp/prob.h"
+
+void av1_write_prob_diff_update(aom_writer *w, aom_prob newp, aom_prob oldpm);
+
+void av1_cond_prob_diff_update(aom_writer *w, aom_prob *oldp,
+                               const unsigned int ct[2], int probwt);
+
+int av1_prob_diff_update_savings_search(const unsigned int *ct, aom_prob oldp,
+                                        aom_prob *bestp, aom_prob upd,
+                                        int probwt);
+
+int av1_prob_diff_update_savings_search_model(const unsigned int *ct,
+                                              const aom_prob oldp,
+                                              aom_prob *bestp, aom_prob upd,
+                                              int stepsize, int probwt);
+
+int av1_cond_prob_diff_update_savings(aom_prob *oldp, const unsigned int ct[2],
+                                      int probwt);
+#if CONFIG_SUBFRAME_PROB_UPDATE
+int av1_prob_update_search_subframe(unsigned int ct[][2], aom_prob oldp,
+                                    aom_prob *bestp, aom_prob upd, int n);
+int av1_prob_update_search_model_subframe(
+    unsigned int ct[ENTROPY_NODES][COEF_PROBS_BUFS][2], const aom_prob *oldp,
+    aom_prob *bestp, aom_prob upd, int stepsize, int n);
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_SUBEXP_H_
diff --git a/third_party/aom/av1/encoder/temporal_filter.c b/third_party/aom/av1/encoder/temporal_filter.c
new file mode 100644
index 0000000000..de962fe84d
--- /dev/null
+++ b/third_party/aom/av1/encoder/temporal_filter.c
@@ -0,0 +1,719 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <limits.h>
+
+#include "./aom_config.h"
+#include "av1/common/alloccommon.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/reconinter.h"
+#include "av1/common/odintrin.h"
+#include "av1/encoder/av1_quantize.h"
+#include "av1/encoder/extend.h"
+#include "av1/encoder/firstpass.h"
+#include "av1/encoder/mcomp.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/ratectrl.h"
+#include "av1/encoder/segmentation.h"
+#include "av1/encoder/temporal_filter.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/aom_timer.h"
+#include "aom_scale/aom_scale.h"
+
+static void temporal_filter_predictors_mb_c(
+    MACROBLOCKD *xd, uint8_t *y_mb_ptr, uint8_t *u_mb_ptr, uint8_t *v_mb_ptr,
+    int stride, int uv_block_width, int uv_block_height, int mv_row, int mv_col,
+    uint8_t *pred, struct scale_factors *scale, int x, int y) {
+  const int which_mv = 0;
+  const MV mv = { mv_row, mv_col };
+  enum mv_precision mv_precision_uv;
+  int uv_stride;
+  // TODO(angiebird): change plane setting accordingly
+  ConvolveParams conv_params = get_conv_params(which_mv, 0);
+
+#if USE_TEMPORALFILTER_12TAP
+#if CONFIG_DUAL_FILTER
+  const InterpFilter interp_filter[4] = { TEMPORALFILTER_12TAP,
+                                          TEMPORALFILTER_12TAP,
+                                          TEMPORALFILTER_12TAP,
+                                          TEMPORALFILTER_12TAP };
+#else
+  const InterpFilter interp_filter = TEMPORALFILTER_12TAP;
+#endif
+  (void)xd;
+#else
+  const InterpFilter interp_filter = xd->mi[0]->mbmi.interp_filter;
+#endif  // USE_TEMPORALFILTER_12TAP
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+  WarpTypesAllowed warp_types;
+  memset(&warp_types, 0, sizeof(WarpTypesAllowed));
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+
+  if (uv_block_width == 8) {
+    uv_stride = (stride + 1) >> 1;
+    mv_precision_uv = MV_PRECISION_Q4;
+  } else {
+    uv_stride = stride;
+    mv_precision_uv = MV_PRECISION_Q3;
+  }
+
+#if CONFIG_HIGHBITDEPTH
+  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    av1_highbd_build_inter_predictor(y_mb_ptr, stride, &pred[0], 16, &mv, scale,
+                                     16, 16, which_mv, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     &warp_types, x, y,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     0, MV_PRECISION_Q3, x, y, xd);
+
+    av1_highbd_build_inter_predictor(u_mb_ptr, uv_stride, &pred[256],
+                                     uv_block_width, &mv, scale, uv_block_width,
+                                     uv_block_height, which_mv, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     &warp_types, x, y,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     1, mv_precision_uv, x, y, xd);
+
+    av1_highbd_build_inter_predictor(v_mb_ptr, uv_stride, &pred[512],
+                                     uv_block_width, &mv, scale, uv_block_width,
+                                     uv_block_height, which_mv, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     &warp_types, x, y,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                                     2, mv_precision_uv, x, y, xd);
+    return;
+  }
+#endif  // CONFIG_HIGHBITDEPTH
+  av1_build_inter_predictor(y_mb_ptr, stride, &pred[0], 16, &mv, scale, 16, 16,
+                            &conv_params, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            &warp_types, x, y, 0, 0,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            MV_PRECISION_Q3, x, y, xd);
+
+  av1_build_inter_predictor(u_mb_ptr, uv_stride, &pred[256], uv_block_width,
+                            &mv, scale, uv_block_width, uv_block_height,
+                            &conv_params, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            &warp_types, x, y, 1, 0,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            mv_precision_uv, x, y, xd);
+
+  av1_build_inter_predictor(v_mb_ptr, uv_stride, &pred[512], uv_block_width,
+                            &mv, scale, uv_block_width, uv_block_height,
+                            &conv_params, interp_filter,
+#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            &warp_types, x, y, 2, 0,
+#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
+                            mv_precision_uv, x, y, xd);
+}
+
+void av1_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride,
+                                 uint8_t *frame2, unsigned int block_width,
+                                 unsigned int block_height, int strength,
+                                 int filter_weight, unsigned int *accumulator,
+                                 uint16_t *count) {
+  unsigned int i, j, k;
+  int modifier;
+  int byte = 0;
+  const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
+
+  for (i = 0, k = 0; i < block_height; i++) {
+    for (j = 0; j < block_width; j++, k++) {
+      int pixel_value = *frame2;
+
+      // non-local mean approach
+      int diff_sse[9] = { 0 };
+      int idx, idy, index = 0;
+
+      for (idy = -1; idy <= 1; ++idy) {
+        for (idx = -1; idx <= 1; ++idx) {
+          int row = (int)i + idy;
+          int col = (int)j + idx;
+
+          if (row >= 0 && row < (int)block_height && col >= 0 &&
+              col < (int)block_width) {
+            int diff = frame1[byte + idy * (int)stride + idx] -
+                       frame2[idy * (int)block_width + idx];
+            diff_sse[index] = diff * diff;
+            ++index;
+          }
+        }
+      }
+
+      assert(index > 0);
+
+      modifier = 0;
+      for (idx = 0; idx < 9; ++idx) modifier += diff_sse[idx];
+
+      modifier *= 3;
+      modifier /= index;
+
+      ++frame2;
+
+      modifier += rounding;
+      modifier >>= strength;
+
+      if (modifier > 16) modifier = 16;
+
+      modifier = 16 - modifier;
+      modifier *= filter_weight;
+
+      count[k] += modifier;
+      accumulator[k] += modifier * pixel_value;
+
+      byte++;
+    }
+
+    byte += stride - block_width;
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH
+void av1_highbd_temporal_filter_apply_c(
+    uint8_t *frame1_8, unsigned int stride, uint8_t *frame2_8,
+    unsigned int block_width, unsigned int block_height, int strength,
+    int filter_weight, unsigned int *accumulator, uint16_t *count) {
+  uint16_t *frame1 = CONVERT_TO_SHORTPTR(frame1_8);
+  uint16_t *frame2 = CONVERT_TO_SHORTPTR(frame2_8);
+  unsigned int i, j, k;
+  int modifier;
+  int byte = 0;
+  const int rounding = strength > 0 ? 1 << (strength - 1) : 0;
+
+  for (i = 0, k = 0; i < block_height; i++) {
+    for (j = 0; j < block_width; j++, k++) {
+      int pixel_value = *frame2;
+
+      // non-local mean approach
+      int diff_sse[9] = { 0 };
+      int idx, idy, index = 0;
+
+      for (idy = -1; idy <= 1; ++idy) {
+        for (idx = -1; idx <= 1; ++idx) {
+          int row = (int)i + idy;
+          int col = (int)j + idx;
+
+          if (row >= 0 && row < (int)block_height && col >= 0 &&
+              col < (int)block_width) {
+            int diff = frame1[byte + idy * (int)stride + idx] -
+                       frame2[idy * (int)block_width + idx];
+            diff_sse[index] = diff * diff;
+            ++index;
+          }
+        }
+      }
+
+      assert(index > 0);
+
+      modifier = 0;
+      for (idx = 0; idx < 9; ++idx) modifier += diff_sse[idx];
+
+      modifier *= 3;
+      modifier /= index;
+
+      ++frame2;
+
+      modifier += rounding;
+      modifier >>= strength;
+
+      if (modifier > 16) modifier = 16;
+
+      modifier = 16 - modifier;
+      modifier *= filter_weight;
+
+      count[k] += modifier;
+      accumulator[k] += modifier * pixel_value;
+
+      byte++;
+    }
+
+    byte += stride - block_width;
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+static int temporal_filter_find_matching_mb_c(AV1_COMP *cpi,
+                                              uint8_t *arf_frame_buf,
+                                              uint8_t *frame_ptr_buf,
+                                              int stride) {
+  MACROBLOCK *const x = &cpi->td.mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  const MV_SPEED_FEATURES *const mv_sf = &cpi->sf.mv;
+  int step_param;
+  int sadpb = x->sadperbit16;
+  int bestsme = INT_MAX;
+  int distortion;
+  unsigned int sse;
+  int cost_list[5];
+  MvLimits tmp_mv_limits = x->mv_limits;
+
+  MV best_ref_mv1 = { 0, 0 };
+  MV best_ref_mv1_full; /* full-pixel value of best_ref_mv1 */
+
+  // Save input state
+  struct buf_2d src = x->plane[0].src;
+  struct buf_2d pre = xd->plane[0].pre[0];
+
+  best_ref_mv1_full.col = best_ref_mv1.col >> 3;
+  best_ref_mv1_full.row = best_ref_mv1.row >> 3;
+
+  // Setup frame pointers
+  x->plane[0].src.buf = arf_frame_buf;
+  x->plane[0].src.stride = stride;
+  xd->plane[0].pre[0].buf = frame_ptr_buf;
+  xd->plane[0].pre[0].stride = stride;
+
+  step_param = mv_sf->reduce_first_step_size;
+  step_param = AOMMIN(step_param, MAX_MVSEARCH_STEPS - 2);
+
+  av1_set_mv_search_range(&x->mv_limits, &best_ref_mv1);
+
+#if CONFIG_REF_MV
+  x->mvcost = x->mv_cost_stack[0];
+  x->nmvjointcost = x->nmv_vec_cost[0];
+  x->mvsadcost = x->mvcost;
+  x->nmvjointsadcost = x->nmvjointcost;
+#endif
+
+  // Ignore mv costing by sending NULL pointer instead of cost arrays
+  av1_hex_search(x, &best_ref_mv1_full, step_param, sadpb, 1,
+                 cond_cost_list(cpi, cost_list), &cpi->fn_ptr[BLOCK_16X16], 0,
+                 &best_ref_mv1);
+
+  x->mv_limits = tmp_mv_limits;
+
+  // Ignore mv costing by sending NULL pointer instead of cost array
+  bestsme = cpi->find_fractional_mv_step(
+      x, &best_ref_mv1, cpi->common.allow_high_precision_mv, x->errorperbit,
+      &cpi->fn_ptr[BLOCK_16X16], 0, mv_sf->subpel_iters_per_step,
+      cond_cost_list(cpi, cost_list), NULL, NULL, &distortion, &sse, NULL, 0, 0,
+      0);
+
+  x->e_mbd.mi[0]->bmi[0].as_mv[0] = x->best_mv;
+
+  // Restore input state
+  x->plane[0].src = src;
+  xd->plane[0].pre[0] = pre;
+
+  return bestsme;
+}
+
+static void temporal_filter_iterate_c(AV1_COMP *cpi,
+                                      YV12_BUFFER_CONFIG **frames,
+                                      int frame_count, int alt_ref_index,
+                                      int strength,
+                                      struct scale_factors *scale) {
+  int byte;
+  int frame;
+  int mb_col, mb_row;
+  unsigned int filter_weight;
+  int mb_cols = (frames[alt_ref_index]->y_crop_width + 15) >> 4;
+  int mb_rows = (frames[alt_ref_index]->y_crop_height + 15) >> 4;
+  int mb_y_offset = 0;
+  int mb_uv_offset = 0;
+  DECLARE_ALIGNED(16, unsigned int, accumulator[16 * 16 * 3]);
+  DECLARE_ALIGNED(16, uint16_t, count[16 * 16 * 3]);
+  MACROBLOCKD *mbd = &cpi->td.mb.e_mbd;
+  YV12_BUFFER_CONFIG *f = frames[alt_ref_index];
+  uint8_t *dst1, *dst2;
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, predictor16[16 * 16 * 3]);
+  DECLARE_ALIGNED(16, uint8_t, predictor8[16 * 16 * 3]);
+  uint8_t *predictor;
+#else
+  DECLARE_ALIGNED(16, uint8_t, predictor[16 * 16 * 3]);
+#endif
+  const int mb_uv_height = 16 >> mbd->plane[1].subsampling_y;
+  const int mb_uv_width = 16 >> mbd->plane[1].subsampling_x;
+
+  // Save input state
+  uint8_t *input_buffer[MAX_MB_PLANE];
+  int i;
+#if CONFIG_HIGHBITDEPTH
+  if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+    predictor = CONVERT_TO_BYTEPTR(predictor16);
+  } else {
+    predictor = predictor8;
+  }
+#endif
+
+  for (i = 0; i < MAX_MB_PLANE; i++) input_buffer[i] = mbd->plane[i].pre[0].buf;
+
+  for (mb_row = 0; mb_row < mb_rows; mb_row++) {
+    // Source frames are extended to 16 pixels. This is different than
+    //  L/A/G reference frames that have a border of 32 (AV1ENCBORDERINPIXELS)
+    // A 6/8 tap filter is used for motion search.  This requires 2 pixels
+    //  before and 3 pixels after.  So the largest Y mv on a border would
+    //  then be 16 - AOM_INTERP_EXTEND. The UV blocks are half the size of the
+    //  Y and therefore only extended by 8.  The largest mv that a UV block
+    //  can support is 8 - AOM_INTERP_EXTEND.  A UV mv is half of a Y mv.
+    //  (16 - AOM_INTERP_EXTEND) >> 1 which is greater than
+    //  8 - AOM_INTERP_EXTEND.
+    // To keep the mv in play for both Y and UV planes the max that it
+    //  can be on a border is therefore 16 - (2*AOM_INTERP_EXTEND+1).
+    cpi->td.mb.mv_limits.row_min =
+        -((mb_row * 16) + (17 - 2 * AOM_INTERP_EXTEND));
+    cpi->td.mb.mv_limits.row_max =
+        ((mb_rows - 1 - mb_row) * 16) + (17 - 2 * AOM_INTERP_EXTEND);
+
+    for (mb_col = 0; mb_col < mb_cols; mb_col++) {
+      int j, k;
+      int stride;
+
+      memset(accumulator, 0, 16 * 16 * 3 * sizeof(accumulator[0]));
+      memset(count, 0, 16 * 16 * 3 * sizeof(count[0]));
+
+      cpi->td.mb.mv_limits.col_min =
+          -((mb_col * 16) + (17 - 2 * AOM_INTERP_EXTEND));
+      cpi->td.mb.mv_limits.col_max =
+          ((mb_cols - 1 - mb_col) * 16) + (17 - 2 * AOM_INTERP_EXTEND);
+
+      for (frame = 0; frame < frame_count; frame++) {
+        const int thresh_low = 10000;
+        const int thresh_high = 20000;
+
+        if (frames[frame] == NULL) continue;
+
+        mbd->mi[0]->bmi[0].as_mv[0].as_mv.row = 0;
+        mbd->mi[0]->bmi[0].as_mv[0].as_mv.col = 0;
+
+        if (frame == alt_ref_index) {
+          filter_weight = 2;
+        } else {
+          // Find best match in this frame by MC
+          int err = temporal_filter_find_matching_mb_c(
+              cpi, frames[alt_ref_index]->y_buffer + mb_y_offset,
+              frames[frame]->y_buffer + mb_y_offset, frames[frame]->y_stride);
+
+          // Assign higher weight to matching MB if it's error
+          // score is lower. If not applying MC default behavior
+          // is to weight all MBs equal.
+          filter_weight = err < thresh_low ? 2 : err < thresh_high ? 1 : 0;
+        }
+
+        if (filter_weight != 0) {
+          // Construct the predictors
+          temporal_filter_predictors_mb_c(
+              mbd, frames[frame]->y_buffer + mb_y_offset,
+              frames[frame]->u_buffer + mb_uv_offset,
+              frames[frame]->v_buffer + mb_uv_offset, frames[frame]->y_stride,
+              mb_uv_width, mb_uv_height, mbd->mi[0]->bmi[0].as_mv[0].as_mv.row,
+              mbd->mi[0]->bmi[0].as_mv[0].as_mv.col, predictor, scale,
+              mb_col * 16, mb_row * 16);
+
+#if CONFIG_HIGHBITDEPTH
+          if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+            int adj_strength = strength + 2 * (mbd->bd - 8);
+            // Apply the filter (YUV)
+            av1_highbd_temporal_filter_apply(
+                f->y_buffer + mb_y_offset, f->y_stride, predictor, 16, 16,
+                adj_strength, filter_weight, accumulator, count);
+            av1_highbd_temporal_filter_apply(
+                f->u_buffer + mb_uv_offset, f->uv_stride, predictor + 256,
+                mb_uv_width, mb_uv_height, adj_strength, filter_weight,
+                accumulator + 256, count + 256);
+            av1_highbd_temporal_filter_apply(
+                f->v_buffer + mb_uv_offset, f->uv_stride, predictor + 512,
+                mb_uv_width, mb_uv_height, adj_strength, filter_weight,
+                accumulator + 512, count + 512);
+          } else {
+            // Apply the filter (YUV)
+            av1_temporal_filter_apply_c(f->y_buffer + mb_y_offset, f->y_stride,
+                                        predictor, 16, 16, strength,
+                                        filter_weight, accumulator, count);
+            av1_temporal_filter_apply_c(
+                f->u_buffer + mb_uv_offset, f->uv_stride, predictor + 256,
+                mb_uv_width, mb_uv_height, strength, filter_weight,
+                accumulator + 256, count + 256);
+            av1_temporal_filter_apply_c(
+                f->v_buffer + mb_uv_offset, f->uv_stride, predictor + 512,
+                mb_uv_width, mb_uv_height, strength, filter_weight,
+                accumulator + 512, count + 512);
+          }
+#else
+          // Apply the filter (YUV)
+          av1_temporal_filter_apply_c(f->y_buffer + mb_y_offset, f->y_stride,
+                                      predictor, 16, 16, strength,
+                                      filter_weight, accumulator, count);
+          av1_temporal_filter_apply_c(f->u_buffer + mb_uv_offset, f->uv_stride,
+                                      predictor + 256, mb_uv_width,
+                                      mb_uv_height, strength, filter_weight,
+                                      accumulator + 256, count + 256);
+          av1_temporal_filter_apply_c(f->v_buffer + mb_uv_offset, f->uv_stride,
+                                      predictor + 512, mb_uv_width,
+                                      mb_uv_height, strength, filter_weight,
+                                      accumulator + 512, count + 512);
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+
+#if CONFIG_HIGHBITDEPTH
+      if (mbd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+        uint16_t *dst1_16;
+        uint16_t *dst2_16;
+        // Normalize filter output to produce AltRef frame
+        dst1 = cpi->alt_ref_buffer.y_buffer;
+        dst1_16 = CONVERT_TO_SHORTPTR(dst1);
+        stride = cpi->alt_ref_buffer.y_stride;
+        byte = mb_y_offset;
+        for (i = 0, k = 0; i < 16; i++) {
+          for (j = 0; j < 16; j++, k++) {
+            dst1_16[byte] =
+                (uint16_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+            // move to next pixel
+            byte++;
+          }
+
+          byte += stride - 16;
+        }
+
+        dst1 = cpi->alt_ref_buffer.u_buffer;
+        dst2 = cpi->alt_ref_buffer.v_buffer;
+        dst1_16 = CONVERT_TO_SHORTPTR(dst1);
+        dst2_16 = CONVERT_TO_SHORTPTR(dst2);
+        stride = cpi->alt_ref_buffer.uv_stride;
+        byte = mb_uv_offset;
+        for (i = 0, k = 256; i < mb_uv_height; i++) {
+          for (j = 0; j < mb_uv_width; j++, k++) {
+            int m = k + 256;
+
+            // U
+            dst1_16[byte] =
+                (uint16_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+            // V
+            dst2_16[byte] =
+                (uint16_t)OD_DIVU(accumulator[m] + (count[m] >> 1), count[m]);
+
+            // move to next pixel
+            byte++;
+          }
+
+          byte += stride - mb_uv_width;
+        }
+      } else {
+        // Normalize filter output to produce AltRef frame
+        dst1 = cpi->alt_ref_buffer.y_buffer;
+        stride = cpi->alt_ref_buffer.y_stride;
+        byte = mb_y_offset;
+        for (i = 0, k = 0; i < 16; i++) {
+          for (j = 0; j < 16; j++, k++) {
+            dst1[byte] =
+                (uint8_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+            // move to next pixel
+            byte++;
+          }
+          byte += stride - 16;
+        }
+
+        dst1 = cpi->alt_ref_buffer.u_buffer;
+        dst2 = cpi->alt_ref_buffer.v_buffer;
+        stride = cpi->alt_ref_buffer.uv_stride;
+        byte = mb_uv_offset;
+        for (i = 0, k = 256; i < mb_uv_height; i++) {
+          for (j = 0; j < mb_uv_width; j++, k++) {
+            int m = k + 256;
+
+            // U
+            dst1[byte] =
+                (uint8_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+            // V
+            dst2[byte] =
+                (uint8_t)OD_DIVU(accumulator[m] + (count[m] >> 1), count[m]);
+
+            // move to next pixel
+            byte++;
+          }
+          byte += stride - mb_uv_width;
+        }
+      }
+#else
+      // Normalize filter output to produce AltRef frame
+      dst1 = cpi->alt_ref_buffer.y_buffer;
+      stride = cpi->alt_ref_buffer.y_stride;
+      byte = mb_y_offset;
+      for (i = 0, k = 0; i < 16; i++) {
+        for (j = 0; j < 16; j++, k++) {
+          dst1[byte] =
+              (uint8_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+          // move to next pixel
+          byte++;
+        }
+        byte += stride - 16;
+      }
+
+      dst1 = cpi->alt_ref_buffer.u_buffer;
+      dst2 = cpi->alt_ref_buffer.v_buffer;
+      stride = cpi->alt_ref_buffer.uv_stride;
+      byte = mb_uv_offset;
+      for (i = 0, k = 256; i < mb_uv_height; i++) {
+        for (j = 0; j < mb_uv_width; j++, k++) {
+          int m = k + 256;
+
+          // U
+          dst1[byte] =
+              (uint8_t)OD_DIVU(accumulator[k] + (count[k] >> 1), count[k]);
+
+          // V
+          dst2[byte] =
+              (uint8_t)OD_DIVU(accumulator[m] + (count[m] >> 1), count[m]);
+
+          // move to next pixel
+          byte++;
+        }
+        byte += stride - mb_uv_width;
+      }
+#endif  // CONFIG_HIGHBITDEPTH
+      mb_y_offset += 16;
+      mb_uv_offset += mb_uv_width;
+    }
+    mb_y_offset += 16 * (f->y_stride - mb_cols);
+    mb_uv_offset += mb_uv_height * f->uv_stride - mb_uv_width * mb_cols;
+  }
+
+  // Restore input state
+  for (i = 0; i < MAX_MB_PLANE; i++) mbd->plane[i].pre[0].buf = input_buffer[i];
+}
+
+// Apply buffer limits and context specific adjustments to arnr filter.
+static void adjust_arnr_filter(AV1_COMP *cpi, int distance, int group_boost,
+                               int *arnr_frames, int *arnr_strength) {
+  const AV1EncoderConfig *const oxcf = &cpi->oxcf;
+  const int frames_after_arf =
+      av1_lookahead_depth(cpi->lookahead) - distance - 1;
+  int frames_fwd = (cpi->oxcf.arnr_max_frames - 1) >> 1;
+  int frames_bwd;
+  int q, frames, strength;
+
+  // Define the forward and backwards filter limits for this arnr group.
+  if (frames_fwd > frames_after_arf) frames_fwd = frames_after_arf;
+  if (frames_fwd > distance) frames_fwd = distance;
+
+  frames_bwd = frames_fwd;
+
+  // For even length filter there is one more frame backward
+  // than forward: e.g. len=6 ==> bbbAff, len=7 ==> bbbAfff.
+  if (frames_bwd < distance) frames_bwd += (oxcf->arnr_max_frames + 1) & 0x1;
+
+  // Set the baseline active filter size.
+  frames = frames_bwd + 1 + frames_fwd;
+
+  // Adjust the strength based on active max q.
+  if (cpi->common.current_video_frame > 1)
+    q = ((int)av1_convert_qindex_to_q(cpi->rc.avg_frame_qindex[INTER_FRAME],
+                                      cpi->common.bit_depth));
+  else
+    q = ((int)av1_convert_qindex_to_q(cpi->rc.avg_frame_qindex[KEY_FRAME],
+                                      cpi->common.bit_depth));
+  if (q > 16) {
+    strength = oxcf->arnr_strength;
+  } else {
+    strength = oxcf->arnr_strength - ((16 - q) / 2);
+    if (strength < 0) strength = 0;
+  }
+
+  // Adjust number of frames in filter and strength based on gf boost level.
+  if (frames > group_boost / 150) {
+    frames = group_boost / 150;
+    frames += !(frames & 1);
+  }
+
+  if (strength > group_boost / 300) {
+    strength = group_boost / 300;
+  }
+
+  // Adjustments for second level arf in multi arf case.
+  if (cpi->oxcf.pass == 2 && cpi->multi_arf_allowed) {
+    const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+    if (gf_group->rf_level[gf_group->index] != GF_ARF_STD) {
+      strength >>= 1;
+    }
+  }
+
+  *arnr_frames = frames;
+  *arnr_strength = strength;
+}
+
+void av1_temporal_filter(AV1_COMP *cpi, int distance) {
+  RATE_CONTROL *const rc = &cpi->rc;
+  int frame;
+  int frames_to_blur;
+  int start_frame;
+  int strength;
+  int frames_to_blur_backward;
+  int frames_to_blur_forward;
+  struct scale_factors sf;
+  YV12_BUFFER_CONFIG *frames[MAX_LAG_BUFFERS] = { NULL };
+#if CONFIG_EXT_REFS
+  const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
+#endif
+
+  // Apply context specific adjustments to the arnr filter parameters.
+  adjust_arnr_filter(cpi, distance, rc->gfu_boost, &frames_to_blur, &strength);
+// TODO(weitinglin): Currently, we enforce the filtering strength on
+//                   extra ARFs' to be zeros. We should investigate in which
+//                   case it is more beneficial to use non-zero strength
+//                   filtering.
+#if CONFIG_EXT_REFS
+  if (gf_group->rf_level[gf_group->index] == GF_ARF_LOW) {
+    strength = 0;
+    frames_to_blur = 1;
+  }
+#endif
+
+#if CONFIG_EXT_REFS
+  if (strength == 0 && frames_to_blur == 1) {
+    cpi->is_arf_filter_off[gf_group->arf_update_idx[gf_group->index]] = 1;
+  } else {
+    cpi->is_arf_filter_off[gf_group->arf_update_idx[gf_group->index]] = 0;
+  }
+#endif
+
+  frames_to_blur_backward = (frames_to_blur / 2);
+  frames_to_blur_forward = ((frames_to_blur - 1) / 2);
+  start_frame = distance + frames_to_blur_forward;
+
+  // Setup frame pointers, NULL indicates frame not included in filter.
+  for (frame = 0; frame < frames_to_blur; ++frame) {
+    const int which_buffer = start_frame - frame;
+    struct lookahead_entry *buf =
+        av1_lookahead_peek(cpi->lookahead, which_buffer);
+    frames[frames_to_blur - 1 - frame] = &buf->img;
+  }
+
+  if (frames_to_blur > 0) {
+// Setup scaling factors. Scaling on each of the arnr frames is not
+// supported.
+// ARF is produced at the native frame size and resized when coded.
+#if CONFIG_HIGHBITDEPTH
+    av1_setup_scale_factors_for_frame(
+        &sf, frames[0]->y_crop_width, frames[0]->y_crop_height,
+        frames[0]->y_crop_width, frames[0]->y_crop_height,
+        cpi->common.use_highbitdepth);
+#else
+    av1_setup_scale_factors_for_frame(
+        &sf, frames[0]->y_crop_width, frames[0]->y_crop_height,
+        frames[0]->y_crop_width, frames[0]->y_crop_height);
+#endif  // CONFIG_HIGHBITDEPTH
+  }
+
+  temporal_filter_iterate_c(cpi, frames, frames_to_blur,
+                            frames_to_blur_backward, strength, &sf);
+}
diff --git a/third_party/aom/av1/encoder/temporal_filter.h b/third_party/aom/av1/encoder/temporal_filter.h
new file mode 100644
index 0000000000..bc0863a638
--- /dev/null
+++ b/third_party/aom/av1/encoder/temporal_filter.h
@@ -0,0 +1,25 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_TEMPORAL_FILTER_H_
+#define AV1_ENCODER_TEMPORAL_FILTER_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_temporal_filter(AV1_COMP *cpi, int distance);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_TEMPORAL_FILTER_H_
diff --git a/third_party/aom/av1/encoder/tokenize.c b/third_party/aom/av1/encoder/tokenize.c
new file mode 100644
index 0000000000..f48493bf89
--- /dev/null
+++ b/third_party/aom/av1/encoder/tokenize.c
@@ -0,0 +1,887 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "aom_mem/aom_mem.h"
+
+#include "av1/common/entropy.h"
+#include "av1/common/pred_common.h"
+#include "av1/common/scan.h"
+#include "av1/common/seg_common.h"
+
+#include "av1/encoder/cost.h"
+#include "av1/encoder/encoder.h"
+#include "av1/encoder/rdopt.h"
+#include "av1/encoder/tokenize.h"
+
+static const TOKENVALUE dct_cat_lt_10_value_tokens[] = {
+  { 9, 63 }, { 9, 61 }, { 9, 59 }, { 9, 57 }, { 9, 55 }, { 9, 53 }, { 9, 51 },
+  { 9, 49 }, { 9, 47 }, { 9, 45 }, { 9, 43 }, { 9, 41 }, { 9, 39 }, { 9, 37 },
+  { 9, 35 }, { 9, 33 }, { 9, 31 }, { 9, 29 }, { 9, 27 }, { 9, 25 }, { 9, 23 },
+  { 9, 21 }, { 9, 19 }, { 9, 17 }, { 9, 15 }, { 9, 13 }, { 9, 11 }, { 9, 9 },
+  { 9, 7 },  { 9, 5 },  { 9, 3 },  { 9, 1 },  { 8, 31 }, { 8, 29 }, { 8, 27 },
+  { 8, 25 }, { 8, 23 }, { 8, 21 }, { 8, 19 }, { 8, 17 }, { 8, 15 }, { 8, 13 },
+  { 8, 11 }, { 8, 9 },  { 8, 7 },  { 8, 5 },  { 8, 3 },  { 8, 1 },  { 7, 15 },
+  { 7, 13 }, { 7, 11 }, { 7, 9 },  { 7, 7 },  { 7, 5 },  { 7, 3 },  { 7, 1 },
+  { 6, 7 },  { 6, 5 },  { 6, 3 },  { 6, 1 },  { 5, 3 },  { 5, 1 },  { 4, 1 },
+  { 3, 1 },  { 2, 1 },  { 1, 1 },  { 0, 0 },  { 1, 0 },  { 2, 0 },  { 3, 0 },
+  { 4, 0 },  { 5, 0 },  { 5, 2 },  { 6, 0 },  { 6, 2 },  { 6, 4 },  { 6, 6 },
+  { 7, 0 },  { 7, 2 },  { 7, 4 },  { 7, 6 },  { 7, 8 },  { 7, 10 }, { 7, 12 },
+  { 7, 14 }, { 8, 0 },  { 8, 2 },  { 8, 4 },  { 8, 6 },  { 8, 8 },  { 8, 10 },
+  { 8, 12 }, { 8, 14 }, { 8, 16 }, { 8, 18 }, { 8, 20 }, { 8, 22 }, { 8, 24 },
+  { 8, 26 }, { 8, 28 }, { 8, 30 }, { 9, 0 },  { 9, 2 },  { 9, 4 },  { 9, 6 },
+  { 9, 8 },  { 9, 10 }, { 9, 12 }, { 9, 14 }, { 9, 16 }, { 9, 18 }, { 9, 20 },
+  { 9, 22 }, { 9, 24 }, { 9, 26 }, { 9, 28 }, { 9, 30 }, { 9, 32 }, { 9, 34 },
+  { 9, 36 }, { 9, 38 }, { 9, 40 }, { 9, 42 }, { 9, 44 }, { 9, 46 }, { 9, 48 },
+  { 9, 50 }, { 9, 52 }, { 9, 54 }, { 9, 56 }, { 9, 58 }, { 9, 60 }, { 9, 62 }
+};
+const TOKENVALUE *av1_dct_cat_lt_10_value_tokens =
+    dct_cat_lt_10_value_tokens +
+    (sizeof(dct_cat_lt_10_value_tokens) / sizeof(*dct_cat_lt_10_value_tokens)) /
+        2;
+// The corresponding costs of the extrabits for the tokens in the above table
+// are stored in the table below. The values are obtained from looking up the
+// entry for the specified extrabits in the table corresponding to the token
+// (as defined in cost element av1_extra_bits)
+// e.g. {9, 63} maps to cat5_cost[63 >> 1], {1, 1} maps to sign_cost[1 >> 1]
+static const int dct_cat_lt_10_value_cost[] = {
+  3773, 3750, 3704, 3681, 3623, 3600, 3554, 3531, 3432, 3409, 3363, 3340, 3282,
+  3259, 3213, 3190, 3136, 3113, 3067, 3044, 2986, 2963, 2917, 2894, 2795, 2772,
+  2726, 2703, 2645, 2622, 2576, 2553, 3197, 3116, 3058, 2977, 2881, 2800, 2742,
+  2661, 2615, 2534, 2476, 2395, 2299, 2218, 2160, 2079, 2566, 2427, 2334, 2195,
+  2023, 1884, 1791, 1652, 1893, 1696, 1453, 1256, 1229, 864,  512,  512,  512,
+  512,  0,    512,  512,  512,  512,  864,  1229, 1256, 1453, 1696, 1893, 1652,
+  1791, 1884, 2023, 2195, 2334, 2427, 2566, 2079, 2160, 2218, 2299, 2395, 2476,
+  2534, 2615, 2661, 2742, 2800, 2881, 2977, 3058, 3116, 3197, 2553, 2576, 2622,
+  2645, 2703, 2726, 2772, 2795, 2894, 2917, 2963, 2986, 3044, 3067, 3113, 3136,
+  3190, 3213, 3259, 3282, 3340, 3363, 3409, 3432, 3531, 3554, 3600, 3623, 3681,
+  3704, 3750, 3773,
+};
+const int *av1_dct_cat_lt_10_value_cost =
+    dct_cat_lt_10_value_cost +
+    (sizeof(dct_cat_lt_10_value_cost) / sizeof(*dct_cat_lt_10_value_cost)) / 2;
+
+// Array indices are identical to previously-existing CONTEXT_NODE indices
+/* clang-format off */
+const aom_tree_index av1_coef_tree[TREE_SIZE(ENTROPY_TOKENS)] = {
+  -EOB_TOKEN, 2,                       // 0  = EOB
+  -ZERO_TOKEN, 4,                      // 1  = ZERO
+  -ONE_TOKEN, 6,                       // 2  = ONE
+  8, 12,                               // 3  = LOW_VAL
+  -TWO_TOKEN, 10,                      // 4  = TWO
+  -THREE_TOKEN, -FOUR_TOKEN,           // 5  = THREE
+  14, 16,                              // 6  = HIGH_LOW
+  -CATEGORY1_TOKEN, -CATEGORY2_TOKEN,  // 7  = CAT_ONE
+  18, 20,                              // 8  = CAT_THREEFOUR
+  -CATEGORY3_TOKEN, -CATEGORY4_TOKEN,  // 9  = CAT_THREE
+  -CATEGORY5_TOKEN, -CATEGORY6_TOKEN   // 10 = CAT_FIVE
+};
+/* clang-format on */
+
+static const int16_t zero_cost[] = { 0 };
+static const int16_t sign_cost[1] = { 512 };
+static const int16_t cat1_cost[1 << 1] = { 864, 1229 };
+static const int16_t cat2_cost[1 << 2] = { 1256, 1453, 1696, 1893 };
+static const int16_t cat3_cost[1 << 3] = { 1652, 1791, 1884, 2023,
+                                           2195, 2334, 2427, 2566 };
+static const int16_t cat4_cost[1 << 4] = { 2079, 2160, 2218, 2299, 2395, 2476,
+                                           2534, 2615, 2661, 2742, 2800, 2881,
+                                           2977, 3058, 3116, 3197 };
+static const int16_t cat5_cost[1 << 5] = {
+  2553, 2576, 2622, 2645, 2703, 2726, 2772, 2795, 2894, 2917, 2963,
+  2986, 3044, 3067, 3113, 3136, 3190, 3213, 3259, 3282, 3340, 3363,
+  3409, 3432, 3531, 3554, 3600, 3623, 3681, 3704, 3750, 3773
+};
+const int16_t av1_cat6_low_cost[256] = {
+  3378, 3390, 3401, 3413, 3435, 3447, 3458, 3470, 3517, 3529, 3540, 3552, 3574,
+  3586, 3597, 3609, 3671, 3683, 3694, 3706, 3728, 3740, 3751, 3763, 3810, 3822,
+  3833, 3845, 3867, 3879, 3890, 3902, 3973, 3985, 3996, 4008, 4030, 4042, 4053,
+  4065, 4112, 4124, 4135, 4147, 4169, 4181, 4192, 4204, 4266, 4278, 4289, 4301,
+  4323, 4335, 4346, 4358, 4405, 4417, 4428, 4440, 4462, 4474, 4485, 4497, 4253,
+  4265, 4276, 4288, 4310, 4322, 4333, 4345, 4392, 4404, 4415, 4427, 4449, 4461,
+  4472, 4484, 4546, 4558, 4569, 4581, 4603, 4615, 4626, 4638, 4685, 4697, 4708,
+  4720, 4742, 4754, 4765, 4777, 4848, 4860, 4871, 4883, 4905, 4917, 4928, 4940,
+  4987, 4999, 5010, 5022, 5044, 5056, 5067, 5079, 5141, 5153, 5164, 5176, 5198,
+  5210, 5221, 5233, 5280, 5292, 5303, 5315, 5337, 5349, 5360, 5372, 4988, 5000,
+  5011, 5023, 5045, 5057, 5068, 5080, 5127, 5139, 5150, 5162, 5184, 5196, 5207,
+  5219, 5281, 5293, 5304, 5316, 5338, 5350, 5361, 5373, 5420, 5432, 5443, 5455,
+  5477, 5489, 5500, 5512, 5583, 5595, 5606, 5618, 5640, 5652, 5663, 5675, 5722,
+  5734, 5745, 5757, 5779, 5791, 5802, 5814, 5876, 5888, 5899, 5911, 5933, 5945,
+  5956, 5968, 6015, 6027, 6038, 6050, 6072, 6084, 6095, 6107, 5863, 5875, 5886,
+  5898, 5920, 5932, 5943, 5955, 6002, 6014, 6025, 6037, 6059, 6071, 6082, 6094,
+  6156, 6168, 6179, 6191, 6213, 6225, 6236, 6248, 6295, 6307, 6318, 6330, 6352,
+  6364, 6375, 6387, 6458, 6470, 6481, 6493, 6515, 6527, 6538, 6550, 6597, 6609,
+  6620, 6632, 6654, 6666, 6677, 6689, 6751, 6763, 6774, 6786, 6808, 6820, 6831,
+  6843, 6890, 6902, 6913, 6925, 6947, 6959, 6970, 6982
+};
+const int av1_cat6_high_cost[CAT6_HIGH_COST_ENTRIES] = {
+  100,   2263,  2739,  4902,  3160,  5323,  5799,  7962,  3678,  5841,  6317,
+  8480,  6738,  8901,  9377,  11540, 3678,  5841,  6317,  8480,  6738,  8901,
+  9377,  11540, 7256,  9419,  9895,  12058, 10316, 12479, 12955, 15118, 3678,
+  5841,  6317,  8480,  6738,  8901,  9377,  11540, 7256,  9419,  9895,  12058,
+  10316, 12479, 12955, 15118, 7256,  9419,  9895,  12058, 10316, 12479, 12955,
+  15118, 10834, 12997, 13473, 15636, 13894, 16057, 16533, 18696,
+#if CONFIG_HIGHBITDEPTH
+  4193,  6356,  6832,  8995,  7253,  9416,  9892,  12055, 7771,  9934,  10410,
+  12573, 10831, 12994, 13470, 15633, 7771,  9934,  10410, 12573, 10831, 12994,
+  13470, 15633, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 7771,
+  9934,  10410, 12573, 10831, 12994, 13470, 15633, 11349, 13512, 13988, 16151,
+  14409, 16572, 17048, 19211, 11349, 13512, 13988, 16151, 14409, 16572, 17048,
+  19211, 14927, 17090, 17566, 19729, 17987, 20150, 20626, 22789, 4193,  6356,
+  6832,  8995,  7253,  9416,  9892,  12055, 7771,  9934,  10410, 12573, 10831,
+  12994, 13470, 15633, 7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633,
+  11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 7771,  9934,  10410,
+  12573, 10831, 12994, 13470, 15633, 11349, 13512, 13988, 16151, 14409, 16572,
+  17048, 19211, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 14927,
+  17090, 17566, 19729, 17987, 20150, 20626, 22789, 8286,  10449, 10925, 13088,
+  11346, 13509, 13985, 16148, 11864, 14027, 14503, 16666, 14924, 17087, 17563,
+  19726, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605,
+  18081, 20244, 18502, 20665, 21141, 23304, 11864, 14027, 14503, 16666, 14924,
+  17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304,
+  15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 19020, 21183, 21659,
+  23822, 22080, 24243, 24719, 26882, 4193,  6356,  6832,  8995,  7253,  9416,
+  9892,  12055, 7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633, 7771,
+  9934,  10410, 12573, 10831, 12994, 13470, 15633, 11349, 13512, 13988, 16151,
+  14409, 16572, 17048, 19211, 7771,  9934,  10410, 12573, 10831, 12994, 13470,
+  15633, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 11349, 13512,
+  13988, 16151, 14409, 16572, 17048, 19211, 14927, 17090, 17566, 19729, 17987,
+  20150, 20626, 22789, 8286,  10449, 10925, 13088, 11346, 13509, 13985, 16148,
+  11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 11864, 14027, 14503,
+  16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665,
+  21141, 23304, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442,
+  17605, 18081, 20244, 18502, 20665, 21141, 23304, 15442, 17605, 18081, 20244,
+  18502, 20665, 21141, 23304, 19020, 21183, 21659, 23822, 22080, 24243, 24719,
+  26882, 8286,  10449, 10925, 13088, 11346, 13509, 13985, 16148, 11864, 14027,
+  14503, 16666, 14924, 17087, 17563, 19726, 11864, 14027, 14503, 16666, 14924,
+  17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304,
+  11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081,
+  20244, 18502, 20665, 21141, 23304, 15442, 17605, 18081, 20244, 18502, 20665,
+  21141, 23304, 19020, 21183, 21659, 23822, 22080, 24243, 24719, 26882, 12379,
+  14542, 15018, 17181, 15439, 17602, 18078, 20241, 15957, 18120, 18596, 20759,
+  19017, 21180, 21656, 23819, 15957, 18120, 18596, 20759, 19017, 21180, 21656,
+  23819, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 15957, 18120,
+  18596, 20759, 19017, 21180, 21656, 23819, 19535, 21698, 22174, 24337, 22595,
+  24758, 25234, 27397, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397,
+  23113, 25276, 25752, 27915, 26173, 28336, 28812, 30975, 4193,  6356,  6832,
+  8995,  7253,  9416,  9892,  12055, 7771,  9934,  10410, 12573, 10831, 12994,
+  13470, 15633, 7771,  9934,  10410, 12573, 10831, 12994, 13470, 15633, 11349,
+  13512, 13988, 16151, 14409, 16572, 17048, 19211, 7771,  9934,  10410, 12573,
+  10831, 12994, 13470, 15633, 11349, 13512, 13988, 16151, 14409, 16572, 17048,
+  19211, 11349, 13512, 13988, 16151, 14409, 16572, 17048, 19211, 14927, 17090,
+  17566, 19729, 17987, 20150, 20626, 22789, 8286,  10449, 10925, 13088, 11346,
+  13509, 13985, 16148, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726,
+  11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081,
+  20244, 18502, 20665, 21141, 23304, 11864, 14027, 14503, 16666, 14924, 17087,
+  17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 15442,
+  17605, 18081, 20244, 18502, 20665, 21141, 23304, 19020, 21183, 21659, 23822,
+  22080, 24243, 24719, 26882, 8286,  10449, 10925, 13088, 11346, 13509, 13985,
+  16148, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726, 11864, 14027,
+  14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244, 18502,
+  20665, 21141, 23304, 11864, 14027, 14503, 16666, 14924, 17087, 17563, 19726,
+  15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 15442, 17605, 18081,
+  20244, 18502, 20665, 21141, 23304, 19020, 21183, 21659, 23822, 22080, 24243,
+  24719, 26882, 12379, 14542, 15018, 17181, 15439, 17602, 18078, 20241, 15957,
+  18120, 18596, 20759, 19017, 21180, 21656, 23819, 15957, 18120, 18596, 20759,
+  19017, 21180, 21656, 23819, 19535, 21698, 22174, 24337, 22595, 24758, 25234,
+  27397, 15957, 18120, 18596, 20759, 19017, 21180, 21656, 23819, 19535, 21698,
+  22174, 24337, 22595, 24758, 25234, 27397, 19535, 21698, 22174, 24337, 22595,
+  24758, 25234, 27397, 23113, 25276, 25752, 27915, 26173, 28336, 28812, 30975,
+  8286,  10449, 10925, 13088, 11346, 13509, 13985, 16148, 11864, 14027, 14503,
+  16666, 14924, 17087, 17563, 19726, 11864, 14027, 14503, 16666, 14924, 17087,
+  17563, 19726, 15442, 17605, 18081, 20244, 18502, 20665, 21141, 23304, 11864,
+  14027, 14503, 16666, 14924, 17087, 17563, 19726, 15442, 17605, 18081, 20244,
+  18502, 20665, 21141, 23304, 15442, 17605, 18081, 20244, 18502, 20665, 21141,
+  23304, 19020, 21183, 21659, 23822, 22080, 24243, 24719, 26882, 12379, 14542,
+  15018, 17181, 15439, 17602, 18078, 20241, 15957, 18120, 18596, 20759, 19017,
+  21180, 21656, 23819, 15957, 18120, 18596, 20759, 19017, 21180, 21656, 23819,
+  19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 15957, 18120, 18596,
+  20759, 19017, 21180, 21656, 23819, 19535, 21698, 22174, 24337, 22595, 24758,
+  25234, 27397, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 23113,
+  25276, 25752, 27915, 26173, 28336, 28812, 30975, 12379, 14542, 15018, 17181,
+  15439, 17602, 18078, 20241, 15957, 18120, 18596, 20759, 19017, 21180, 21656,
+  23819, 15957, 18120, 18596, 20759, 19017, 21180, 21656, 23819, 19535, 21698,
+  22174, 24337, 22595, 24758, 25234, 27397, 15957, 18120, 18596, 20759, 19017,
+  21180, 21656, 23819, 19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397,
+  19535, 21698, 22174, 24337, 22595, 24758, 25234, 27397, 23113, 25276, 25752,
+  27915, 26173, 28336, 28812, 30975, 16472, 18635, 19111, 21274, 19532, 21695,
+  22171, 24334, 20050, 22213, 22689, 24852, 23110, 25273, 25749, 27912, 20050,
+  22213, 22689, 24852, 23110, 25273, 25749, 27912, 23628, 25791, 26267, 28430,
+  26688, 28851, 29327, 31490, 20050, 22213, 22689, 24852, 23110, 25273, 25749,
+  27912, 23628, 25791, 26267, 28430, 26688, 28851, 29327, 31490, 23628, 25791,
+  26267, 28430, 26688, 28851, 29327, 31490, 27206, 29369, 29845, 32008, 30266,
+  32429, 32905, 35068
+#endif
+};
+
+const uint8_t av1_cat6_skipped_bits_discount[8] = {
+  0, 3, 6, 9, 12, 18, 24, 30
+};
+
+#if CONFIG_NEW_MULTISYMBOL
+const av1_extra_bit av1_extra_bits[ENTROPY_TOKENS] = {
+  { 0, 0, 0, zero_cost },                        // ZERO_TOKEN
+  { 0, 0, 1, sign_cost },                        // ONE_TOKEN
+  { 0, 0, 2, sign_cost },                        // TWO_TOKEN
+  { 0, 0, 3, sign_cost },                        // THREE_TOKEN
+  { 0, 0, 4, sign_cost },                        // FOUR_TOKEN
+  { av1_cat1_cdf, 1, CAT1_MIN_VAL, cat1_cost },  // CATEGORY1_TOKEN
+  { av1_cat2_cdf, 2, CAT2_MIN_VAL, cat2_cost },  // CATEGORY2_TOKEN
+  { av1_cat3_cdf, 3, CAT3_MIN_VAL, cat3_cost },  // CATEGORY3_TOKEN
+  { av1_cat4_cdf, 4, CAT4_MIN_VAL, cat4_cost },  // CATEGORY4_TOKEN
+  { av1_cat5_cdf, 5, CAT5_MIN_VAL, cat5_cost },  // CATEGORY5_TOKEN
+  { av1_cat6_cdf, 18, CAT6_MIN_VAL, 0 },         // CATEGORY6_TOKEN
+  { 0, 0, 0, zero_cost }                         // EOB_TOKEN
+};
+#else
+const av1_extra_bit av1_extra_bits[ENTROPY_TOKENS] = {
+  { 0, 0, 0, zero_cost },                         // ZERO_TOKEN
+  { 0, 0, 1, sign_cost },                         // ONE_TOKEN
+  { 0, 0, 2, sign_cost },                         // TWO_TOKEN
+  { 0, 0, 3, sign_cost },                         // THREE_TOKEN
+  { 0, 0, 4, sign_cost },                         // FOUR_TOKEN
+  { av1_cat1_prob, 1, CAT1_MIN_VAL, cat1_cost },  // CATEGORY1_TOKEN
+  { av1_cat2_prob, 2, CAT2_MIN_VAL, cat2_cost },  // CATEGORY2_TOKEN
+  { av1_cat3_prob, 3, CAT3_MIN_VAL, cat3_cost },  // CATEGORY3_TOKEN
+  { av1_cat4_prob, 4, CAT4_MIN_VAL, cat4_cost },  // CATEGORY4_TOKEN
+  { av1_cat5_prob, 5, CAT5_MIN_VAL, cat5_cost },  // CATEGORY5_TOKEN
+  { av1_cat6_prob, 18, CAT6_MIN_VAL, 0 },         // CATEGORY6_TOKEN
+  { 0, 0, 0, zero_cost }                          // EOB_TOKEN
+};
+#endif
+
+#if !CONFIG_EC_MULTISYMBOL
+const struct av1_token av1_coef_encodings[ENTROPY_TOKENS] = {
+  { 2, 2 },  { 6, 3 },   { 28, 5 },  { 58, 6 },  { 59, 6 },  { 60, 6 },
+  { 61, 6 }, { 124, 7 }, { 125, 7 }, { 126, 7 }, { 127, 7 }, { 0, 1 }
+};
+#endif  // !CONFIG_EC_MULTISYMBOL
+
+struct tokenize_b_args {
+  const AV1_COMP *cpi;
+  ThreadData *td;
+  TOKENEXTRA **tp;
+  int this_rate;
+};
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+static void cost_coeffs_b(int plane, int block, int blk_row, int blk_col,
+                          BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
+  struct tokenize_b_args *const args = arg;
+  const AV1_COMP *const cpi = args->cpi;
+  const AV1_COMMON *cm = &cpi->common;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  const PLANE_TYPE type = pd->plane_type;
+  const int ref = is_inter_block(mbmi);
+  const TX_TYPE tx_type = get_tx_type(type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, ref);
+  const int rate = av1_cost_coeffs(cpi, x, plane, block, tx_size, scan_order,
+                                   pd->above_context + blk_col,
+                                   pd->left_context + blk_row, 0);
+  args->this_rate += rate;
+  (void)plane_bsize;
+  av1_set_contexts(xd, pd, plane, tx_size, p->eobs[block] > 0, blk_col,
+                   blk_row);
+}
+
+static void set_entropy_context_b(int plane, int block, int blk_row,
+                                  int blk_col, BLOCK_SIZE plane_bsize,
+                                  TX_SIZE tx_size, void *arg) {
+  struct tokenize_b_args *const args = arg;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  (void)plane_bsize;
+  av1_set_contexts(xd, pd, plane, tx_size, p->eobs[block] > 0, blk_col,
+                   blk_row);
+}
+
+#if CONFIG_NEW_TOKENSET
+static INLINE void add_token(TOKENEXTRA **t,
+                             aom_cdf_prob (*tail_cdf)[CDF_SIZE(ENTROPY_TOKENS)],
+                             aom_cdf_prob (*head_cdf)[CDF_SIZE(ENTROPY_TOKENS)],
+                             int eob_val, int first_val, int32_t extra,
+                             uint8_t token) {
+  (*t)->token = token;
+  (*t)->extra = extra;
+  (*t)->tail_cdf = tail_cdf;
+  (*t)->head_cdf = head_cdf;
+  (*t)->eob_val = eob_val;
+  (*t)->first_val = first_val;
+  (*t)++;
+}
+
+#else  // CONFIG_NEW_TOKENSET
+static INLINE void add_token(
+    TOKENEXTRA **t, const aom_prob *context_tree,
+#if CONFIG_EC_MULTISYMBOL
+    aom_cdf_prob (*token_cdf)[CDF_SIZE(ENTROPY_TOKENS)],
+#endif  // CONFIG_EC_MULTISYMBOL
+    int32_t extra, uint8_t token, uint8_t skip_eob_node, unsigned int *counts) {
+  (*t)->token = token;
+  (*t)->extra = extra;
+  (*t)->context_tree = context_tree;
+#if CONFIG_EC_MULTISYMBOL
+  (*t)->token_cdf = token_cdf;
+#endif  // CONFIG_EC_MULTISYMBOL
+  (*t)->skip_eob_node = skip_eob_node;
+  (*t)++;
+  ++counts[token];
+}
+#endif  // CONFIG_NEW_TOKENSET
+#endif  // !CONFIG_PVQ || CONFIG_VAR_TX
+
+#if CONFIG_PALETTE
+void av1_tokenize_palette_sb(const AV1_COMP *cpi,
+                             const struct ThreadData *const td, int plane,
+                             TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                             int *rate) {
+  const MACROBLOCK *const x = &td->mb;
+  const MACROBLOCKD *const xd = &x->e_mbd;
+  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const uint8_t *const color_map = xd->plane[plane].color_index_map;
+  const PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
+  const int n = pmi->palette_size[plane];
+  int i, j;
+  int this_rate = 0;
+  uint8_t color_order[PALETTE_MAX_SIZE];
+  const aom_prob(
+      *const probs)[PALETTE_COLOR_INDEX_CONTEXTS][PALETTE_COLORS - 1] =
+      plane == 0 ? av1_default_palette_y_color_index_prob
+                 : av1_default_palette_uv_color_index_prob;
+  int plane_block_width, rows, cols;
+  av1_get_block_dimensions(bsize, plane, xd, &plane_block_width, NULL, &rows,
+                           &cols);
+  assert(plane == 0 || plane == 1);
+
+#if CONFIG_PALETTE_THROUGHPUT
+  int k;
+  for (k = 1; k < rows + cols - 1; ++k) {
+    for (j = AOMMIN(k, cols - 1); j >= AOMMAX(0, k - rows + 1); --j) {
+      i = k - j;
+#else
+  for (i = 0; i < rows; ++i) {
+    for (j = (i == 0 ? 1 : 0); j < cols; ++j) {
+#endif  // CONFIG_PALETTE_THROUGHPUT
+      int color_new_idx;
+      const int color_ctx = av1_get_palette_color_index_context(
+          color_map, plane_block_width, i, j, n, color_order, &color_new_idx);
+      assert(color_new_idx >= 0 && color_new_idx < n);
+      if (dry_run == DRY_RUN_COSTCOEFFS)
+        this_rate += cpi->palette_y_color_cost[n - PALETTE_MIN_SIZE][color_ctx]
+                                              [color_new_idx];
+      (*t)->token = color_new_idx;
+      (*t)->context_tree = probs[n - PALETTE_MIN_SIZE][color_ctx];
+      (*t)->skip_eob_node = 0;
+      ++(*t);
+    }
+  }
+  if (rate) *rate += this_rate;
+}
+#endif  // CONFIG_PALETTE
+
+#if CONFIG_PVQ
+static void add_pvq_block(AV1_COMMON *const cm, MACROBLOCK *const x,
+                          PVQ_INFO *pvq) {
+  PVQ_QUEUE *q = x->pvq_q;
+  if (q->curr_pos >= q->buf_len) {
+    int new_buf_len = 2 * q->buf_len + 1;
+    PVQ_INFO *new_buf;
+    CHECK_MEM_ERROR(cm, new_buf, aom_malloc(new_buf_len * sizeof(PVQ_INFO)));
+    memcpy(new_buf, q->buf, q->buf_len * sizeof(PVQ_INFO));
+    aom_free(q->buf);
+    q->buf = new_buf;
+    q->buf_len = new_buf_len;
+  }
+  OD_COPY(q->buf + q->curr_pos, pvq, 1);
+  ++q->curr_pos;
+}
+
+// NOTE: This does not actually generate tokens, instead we store the encoding
+// decisions made for PVQ in a queue that we will read from when
+// actually writing the bitstream in write_modes_b
+static void tokenize_pvq(int plane, int block, int blk_row, int blk_col,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
+  struct tokenize_b_args *const args = arg;
+  const AV1_COMP *cpi = args->cpi;
+  const AV1_COMMON *const cm = &cpi->common;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  PVQ_INFO *pvq_info;
+
+  (void)block;
+  (void)blk_row;
+  (void)blk_col;
+  (void)plane_bsize;
+  (void)tx_size;
+
+  assert(block < MAX_PVQ_BLOCKS_IN_SB);
+  pvq_info = &x->pvq[block][plane];
+  add_pvq_block((AV1_COMMON * const)cm, x, pvq_info);
+}
+#endif  // CONFIG_PVQ
+
+static void tokenize_b(int plane, int block, int blk_row, int blk_col,
+                       BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
+#if !CONFIG_PVQ
+  struct tokenize_b_args *const args = arg;
+  const AV1_COMP *cpi = args->cpi;
+  const AV1_COMMON *const cm = &cpi->common;
+  ThreadData *const td = args->td;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  TOKENEXTRA **tp = args->tp;
+  uint8_t token_cache[MAX_TX_SQUARE];
+  struct macroblock_plane *p = &x->plane[plane];
+  struct macroblockd_plane *pd = &xd->plane[plane];
+  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
+  int pt; /* near block/prev token context index */
+  int c;
+  TOKENEXTRA *t = *tp; /* store tokens starting here */
+  const int eob = p->eobs[block];
+  const PLANE_TYPE type = pd->plane_type;
+  const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
+#if CONFIG_SUPERTX
+  const int segment_id = AOMMIN(mbmi->segment_id, mbmi->segment_id_supertx);
+#else
+  const int segment_id = mbmi->segment_id;
+#endif  // CONFIG_SUEPRTX
+  const int16_t *scan, *nb;
+  const TX_TYPE tx_type = get_tx_type(type, xd, block, tx_size);
+  const SCAN_ORDER *const scan_order =
+      get_scan(cm, tx_size, tx_type, is_inter_block(mbmi));
+  const int ref = is_inter_block(mbmi);
+  unsigned int(*const counts)[COEFF_CONTEXTS][ENTROPY_TOKENS] =
+      td->rd_counts.coef_counts[txsize_sqr_map[tx_size]][type][ref];
+#if !CONFIG_NEW_TOKENSET
+#if CONFIG_SUBFRAME_PROB_UPDATE
+  const aom_prob(*coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
+      cpi->subframe_stats.coef_probs_buf[cpi->common.coef_probs_update_idx]
+                                        [txsize_sqr_map[tx_size]][type][ref];
+#else
+  aom_prob(*const coef_probs)[COEFF_CONTEXTS][UNCONSTRAINED_NODES] =
+      cpi->common.fc->coef_probs[txsize_sqr_map[tx_size]][type][ref];
+#endif  // CONFIG_SUBFRAME_PROB_UPDATE
+#endif  // !CONFIG_NEW_TOKENSET
+#if CONFIG_EC_ADAPT
+  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
+#elif CONFIG_EC_MULTISYMBOL
+  FRAME_CONTEXT *ec_ctx = cpi->common.fc;
+#endif
+#if CONFIG_NEW_TOKENSET
+  aom_cdf_prob(
+      *const coef_head_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_head_cdfs[txsize_sqr_map[tx_size]][type][ref];
+  aom_cdf_prob(
+      *const coef_tail_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_tail_cdfs[txsize_sqr_map[tx_size]][type][ref];
+  unsigned int(*const blockz_count)[2] =
+      td->counts->blockz_count[txsize_sqr_map[tx_size]][type][ref];
+  int eob_val;
+  int first_val = 1;
+#else
+#if CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob(*const coef_cdfs)[COEFF_CONTEXTS][CDF_SIZE(ENTROPY_TOKENS)] =
+      ec_ctx->coef_cdfs[txsize_sqr_map[tx_size]][type][ref];
+#endif
+  int skip_eob = 0;
+#endif
+  const int seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
+  unsigned int(*const eob_branch)[COEFF_CONTEXTS] =
+      td->counts->eob_branch[txsize_sqr_map[tx_size]][type][ref];
+  const uint8_t *const band = get_band_translate(tx_size);
+  int16_t token;
+  EXTRABIT extra;
+  (void)plane_bsize;
+  pt = get_entropy_context(tx_size, pd->above_context + blk_col,
+                           pd->left_context + blk_row);
+  scan = scan_order->scan;
+  nb = scan_order->neighbors;
+  c = 0;
+
+#if CONFIG_NEW_TOKENSET
+  if (eob == 0)
+    add_token(&t, &coef_tail_cdfs[band[c]][pt], &coef_head_cdfs[band[c]][pt], 1,
+              1, 0, BLOCK_Z_TOKEN);
+
+  ++blockz_count[pt][eob != 0];
+
+  while (c < eob) {
+    int v = qcoeff[scan[c]];
+    first_val = (c == 0);
+
+    if (!v) {
+      add_token(&t, &coef_tail_cdfs[band[c]][pt], &coef_head_cdfs[band[c]][pt],
+                0, first_val, 0, ZERO_TOKEN);
+      ++counts[band[c]][pt][ZERO_TOKEN];
+      token_cache[scan[c]] = 0;
+    } else {
+      eob_val =
+          (c + 1 == eob) ? (c + 1 == seg_eob ? LAST_EOB : EARLY_EOB) : NO_EOB;
+
+      av1_get_token_extra(v, &token, &extra);
+
+      add_token(&t, &coef_tail_cdfs[band[c]][pt], &coef_head_cdfs[band[c]][pt],
+                eob_val, first_val, extra, (uint8_t)token);
+
+      if (eob_val != LAST_EOB) {
+        ++counts[band[c]][pt][token];
+        ++eob_branch[band[c]][pt];
+        counts[band[c]][pt][EOB_TOKEN] += eob_val != NO_EOB;
+      }
+
+      token_cache[scan[c]] = av1_pt_energy_class[token];
+    }
+    ++c;
+    pt = get_coef_context(nb, token_cache, AOMMIN(c, eob - 1));
+  }
+#else
+  while (c < eob) {
+    const int v = qcoeff[scan[c]];
+    eob_branch[band[c]][pt] += !skip_eob;
+
+    av1_get_token_extra(v, &token, &extra);
+
+    add_token(&t, coef_probs[band[c]][pt],
+#if CONFIG_EC_MULTISYMBOL
+              &coef_cdfs[band[c]][pt],
+#endif
+              extra, (uint8_t)token, (uint8_t)skip_eob, counts[band[c]][pt]);
+
+    token_cache[scan[c]] = av1_pt_energy_class[token];
+    ++c;
+    pt = get_coef_context(nb, token_cache, c);
+    skip_eob = (token == ZERO_TOKEN);
+  }
+  if (c < seg_eob) {
+    add_token(&t, coef_probs[band[c]][pt],
+#if CONFIG_EC_MULTISYMBOL
+              NULL,
+#endif
+              0, EOB_TOKEN, 0, counts[band[c]][pt]);
+    ++eob_branch[band[c]][pt];
+  }
+#endif  // CONFIG_NEW_TOKENSET
+
+#if CONFIG_COEF_INTERLEAVE
+  t->token = EOSB_TOKEN;
+  t++;
+#endif
+
+  *tp = t;
+
+#if CONFIG_ADAPT_SCAN
+  // Since dqcoeff is not available here, we pass qcoeff into
+  // av1_update_scan_count_facade(). The update behavior should be the same
+  // because av1_update_scan_count_facade() only cares if coefficients are zero
+  // or not.
+  av1_update_scan_count_facade((AV1_COMMON *)cm, td->counts, tx_size, tx_type,
+                               qcoeff, c);
+#endif
+
+  av1_set_contexts(xd, pd, plane, tx_size, c > 0, blk_col, blk_row);
+#else   // !CONFIG_PVQ
+  tokenize_pvq(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
+#endif  // !CONFIG_PVQ
+}
+
+struct is_skippable_args {
+  uint16_t *eobs;
+  int *skippable;
+};
+static void is_skippable(int plane, int block, int blk_row, int blk_col,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *argv) {
+  struct is_skippable_args *args = argv;
+  (void)plane;
+  (void)plane_bsize;
+  (void)tx_size;
+  (void)blk_row;
+  (void)blk_col;
+  args->skippable[0] &= (!args->eobs[block]);
+}
+
+// TODO(yaowu): rewrite and optimize this function to remove the usage of
+//              av1_foreach_transform_block() and simplify is_skippable().
+int av1_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
+  int result = 1;
+  struct is_skippable_args args = { x->plane[plane].eobs, &result };
+  av1_foreach_transformed_block_in_plane(&x->e_mbd, bsize, plane, is_skippable,
+                                         &args);
+  return result;
+}
+
+#if CONFIG_VAR_TX
+void tokenize_vartx(ThreadData *td, TOKENEXTRA **t, RUN_TYPE dry_run,
+                    TX_SIZE tx_size, BLOCK_SIZE plane_bsize, int blk_row,
+                    int blk_col, int block, int plane, void *arg) {
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const struct macroblockd_plane *const pd = &xd->plane[plane];
+  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
+  const int tx_row = blk_row >> (1 - pd->subsampling_y);
+  const int tx_col = blk_col >> (1 - pd->subsampling_x);
+  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
+  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
+  TX_SIZE plane_tx_size;
+
+  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
+
+  plane_tx_size =
+      plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
+            : mbmi->inter_tx_size[tx_row][tx_col];
+
+  if (tx_size == plane_tx_size) {
+    plane_bsize = get_plane_block_size(mbmi->sb_type, pd);
+    if (!dry_run)
+      tokenize_b(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
+    else if (dry_run == DRY_RUN_NORMAL)
+      set_entropy_context_b(plane, block, blk_row, blk_col, plane_bsize,
+                            tx_size, arg);
+    else if (dry_run == DRY_RUN_COSTCOEFFS)
+      cost_coeffs_b(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
+  } else {
+    // Half the block size in transform block unit.
+    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
+    const int bsl = tx_size_wide_unit[sub_txs];
+    int i;
+
+    assert(bsl > 0);
+
+    for (i = 0; i < 4; ++i) {
+      const int offsetr = blk_row + ((i >> 1) * bsl);
+      const int offsetc = blk_col + ((i & 0x01) * bsl);
+
+      int step = tx_size_wide_unit[sub_txs] * tx_size_high_unit[sub_txs];
+
+      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
+
+      tokenize_vartx(td, t, dry_run, sub_txs, plane_bsize, offsetr, offsetc,
+                     block, plane, arg);
+      block += step;
+    }
+  }
+}
+
+void av1_tokenize_sb_vartx(const AV1_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
+                           RUN_TYPE dry_run, int mi_row, int mi_col,
+                           BLOCK_SIZE bsize, int *rate) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  TOKENEXTRA *t_backup = *t;
+  const int ctx = av1_get_skip_context(xd);
+  const int skip_inc =
+      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP);
+  struct tokenize_b_args arg = { cpi, td, t, 0 };
+  int plane;
+  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
+
+  if (mbmi->skip) {
+    if (!dry_run) td->counts->skip[ctx][1] += skip_inc;
+    reset_skip_context(xd, bsize);
+    if (dry_run) *t = t_backup;
+    return;
+  }
+
+  if (!dry_run)
+    td->counts->skip[ctx][0] += skip_inc;
+  else
+    *t = t_backup;
+
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4
+    if (!is_chroma_reference(mi_row, mi_col, bsize,
+                             xd->plane[plane].subsampling_x,
+                             xd->plane[plane].subsampling_y)) {
+#if !CONFIG_PVQ
+      if (!dry_run) {
+        (*t)->token = EOSB_TOKEN;
+        (*t)++;
+      }
+#endif
+      continue;
+    }
+#endif
+    const struct macroblockd_plane *const pd = &xd->plane[plane];
+#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
+    const BLOCK_SIZE plane_bsize =
+        AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
+#else
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
+#endif
+    const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
+    const int mi_height = block_size_high[plane_bsize] >> tx_size_wide_log2[0];
+    const TX_SIZE max_tx_size = get_vartx_max_txsize(mbmi, plane_bsize);
+    const BLOCK_SIZE txb_size = txsize_to_bsize[max_tx_size];
+    int bw = block_size_wide[txb_size] >> tx_size_wide_log2[0];
+    int bh = block_size_high[txb_size] >> tx_size_wide_log2[0];
+    int idx, idy;
+    int block = 0;
+    int step = tx_size_wide_unit[max_tx_size] * tx_size_high_unit[max_tx_size];
+    for (idy = 0; idy < mi_height; idy += bh) {
+      for (idx = 0; idx < mi_width; idx += bw) {
+        tokenize_vartx(td, t, dry_run, max_tx_size, plane_bsize, idy, idx,
+                       block, plane, &arg);
+        block += step;
+      }
+    }
+
+    if (!dry_run) {
+      (*t)->token = EOSB_TOKEN;
+      (*t)++;
+    }
+  }
+  if (rate) *rate += arg.this_rate;
+}
+#endif  // CONFIG_VAR_TX
+
+void av1_tokenize_sb(const AV1_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
+                     RUN_TYPE dry_run, BLOCK_SIZE bsize, int *rate,
+                     const int mi_row, const int mi_col) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCK *const x = &td->mb;
+  MACROBLOCKD *const xd = &x->e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  const int ctx = av1_get_skip_context(xd);
+  const int skip_inc =
+      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP);
+  struct tokenize_b_args arg = { cpi, td, t, 0 };
+  if (mbmi->skip) {
+    if (!dry_run) td->counts->skip[ctx][1] += skip_inc;
+    reset_skip_context(xd, bsize);
+    return;
+  }
+
+  if (!dry_run) {
+#if CONFIG_COEF_INTERLEAVE
+    td->counts->skip[ctx][0] += skip_inc;
+    av1_foreach_transformed_block_interleave(xd, bsize, tokenize_b, &arg);
+#else
+    int plane;
+
+    td->counts->skip[ctx][0] += skip_inc;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4
+      if (!is_chroma_reference(mi_row, mi_col, bsize,
+                               xd->plane[plane].subsampling_x,
+                               xd->plane[plane].subsampling_y)) {
+#if !CONFIG_PVQ
+        (*t)->token = EOSB_TOKEN;
+        (*t)++;
+#endif
+        continue;
+      }
+#else
+      (void)mi_row;
+      (void)mi_col;
+#endif
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane, tokenize_b,
+                                             &arg);
+#if !CONFIG_PVQ
+      (*t)->token = EOSB_TOKEN;
+      (*t)++;
+#endif  // !CONFIG_PVQ
+    }
+#endif
+  }
+#if !CONFIG_PVQ
+  else if (dry_run == DRY_RUN_NORMAL) {
+    int plane;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4
+      if (!is_chroma_reference(mi_row, mi_col, bsize,
+                               xd->plane[plane].subsampling_x,
+                               xd->plane[plane].subsampling_y))
+        continue;
+#else
+      (void)mi_row;
+      (void)mi_col;
+#endif
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane,
+                                             set_entropy_context_b, &arg);
+    }
+  } else if (dry_run == DRY_RUN_COSTCOEFFS) {
+    int plane;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+#if CONFIG_CB4X4
+      if (!is_chroma_reference(mi_row, mi_col, bsize,
+                               xd->plane[plane].subsampling_x,
+                               xd->plane[plane].subsampling_y))
+        continue;
+#else
+      (void)mi_row;
+      (void)mi_col;
+#endif
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane, cost_coeffs_b,
+                                             &arg);
+    }
+  }
+#endif  // !CONFIG_PVQ
+
+  if (rate) *rate += arg.this_rate;
+}
+
+#if CONFIG_SUPERTX
+void av1_tokenize_sb_supertx(const AV1_COMP *cpi, ThreadData *td,
+                             TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                             int *rate) {
+  const AV1_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &td->mb.e_mbd;
+  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
+  TOKENEXTRA *t_backup = *t;
+  const int ctx = av1_get_skip_context(xd);
+  const int skip_inc =
+      !segfeature_active(&cm->seg, mbmi->segment_id_supertx, SEG_LVL_SKIP);
+  struct tokenize_b_args arg = { cpi, td, t, 0 };
+  if (mbmi->skip) {
+    if (!dry_run) td->counts->skip[ctx][1] += skip_inc;
+    reset_skip_context(xd, bsize);
+    if (dry_run) *t = t_backup;
+    return;
+  }
+
+  if (!dry_run) {
+    int plane;
+    td->counts->skip[ctx][0] += skip_inc;
+
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane, tokenize_b,
+                                             &arg);
+      (*t)->token = EOSB_TOKEN;
+      (*t)++;
+    }
+  } else if (dry_run == DRY_RUN_NORMAL) {
+    int plane;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane)
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane,
+                                             set_entropy_context_b, &arg);
+    *t = t_backup;
+  } else if (dry_run == DRY_RUN_COSTCOEFFS) {
+    int plane;
+    for (plane = 0; plane < MAX_MB_PLANE; ++plane)
+      av1_foreach_transformed_block_in_plane(xd, bsize, plane, cost_coeffs_b,
+                                             &arg);
+  }
+  if (rate) *rate += arg.this_rate;
+}
+#endif  // CONFIG_SUPERTX
diff --git a/third_party/aom/av1/encoder/tokenize.h b/third_party/aom/av1/encoder/tokenize.h
new file mode 100644
index 0000000000..3928111d6f
--- /dev/null
+++ b/third_party/aom/av1/encoder/tokenize.h
@@ -0,0 +1,151 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_TOKENIZE_H_
+#define AV1_ENCODER_TOKENIZE_H_
+
+#include "av1/common/entropy.h"
+
+#include "av1/encoder/block.h"
+#include "av1/encoder/treewriter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define EOSB_TOKEN 127  // Not signalled, encoder only
+
+#if CONFIG_HIGHBITDEPTH
+typedef int32_t EXTRABIT;
+#else
+typedef int16_t EXTRABIT;
+#endif
+
+typedef struct {
+  int16_t token;
+  EXTRABIT extra;
+} TOKENVALUE;
+
+typedef struct {
+#if CONFIG_NEW_TOKENSET
+  aom_cdf_prob (*tail_cdf)[CDF_SIZE(ENTROPY_TOKENS)];
+  aom_cdf_prob (*head_cdf)[CDF_SIZE(ENTROPY_TOKENS)];
+  int eob_val;
+  int first_val;
+#elif CONFIG_EC_MULTISYMBOL
+  aom_cdf_prob (*token_cdf)[CDF_SIZE(ENTROPY_TOKENS)];
+#endif
+  const aom_prob *context_tree;
+  EXTRABIT extra;
+  uint8_t token;
+  uint8_t skip_eob_node;
+} TOKENEXTRA;
+
+extern const aom_tree_index av1_coef_tree[];
+extern const aom_tree_index av1_coef_con_tree[];
+#if !CONFIG_EC_MULTISYMBOL
+extern const struct av1_token av1_coef_encodings[];
+#endif  // !CONFIG_EC_MULTISYMBOL
+
+int av1_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
+
+struct AV1_COMP;
+struct ThreadData;
+
+typedef enum {
+  OUTPUT_ENABLED = 0,
+  DRY_RUN_NORMAL,
+  DRY_RUN_COSTCOEFFS,
+} RUN_TYPE;
+
+// Note in all the tokenize functions rate if non NULL is incremented
+// with the coefficient token cost only if dry_run = DRY_RUN_COSTCOEFS,
+// otherwise rate is not incremented.
+#if CONFIG_VAR_TX
+void av1_tokenize_sb_vartx(const struct AV1_COMP *cpi, struct ThreadData *td,
+                           TOKENEXTRA **t, RUN_TYPE dry_run, int mi_row,
+                           int mi_col, BLOCK_SIZE bsize, int *rate);
+#endif
+#if CONFIG_PALETTE
+void av1_tokenize_palette_sb(const struct AV1_COMP *cpi,
+                             const struct ThreadData *const td, int plane,
+                             TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                             int *rate);
+#endif  // CONFIG_PALETTE
+void av1_tokenize_sb(const struct AV1_COMP *cpi, struct ThreadData *td,
+                     TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                     int *rate, const int mi_row, const int mi_col);
+#if CONFIG_SUPERTX
+void av1_tokenize_sb_supertx(const struct AV1_COMP *cpi, struct ThreadData *td,
+                             TOKENEXTRA **t, RUN_TYPE dry_run, BLOCK_SIZE bsize,
+                             int *rate);
+#endif
+
+extern const int16_t *av1_dct_value_cost_ptr;
+/* TODO: The Token field should be broken out into a separate char array to
+ *  improve cache locality, since it's needed for costing when the rest of the
+ *  fields are not.
+ */
+extern const TOKENVALUE *av1_dct_value_tokens_ptr;
+extern const TOKENVALUE *av1_dct_cat_lt_10_value_tokens;
+extern const int *av1_dct_cat_lt_10_value_cost;
+extern const int16_t av1_cat6_low_cost[256];
+#if CONFIG_HIGHBITDEPTH
+#define CAT6_HIGH_COST_ENTRIES 1024
+#else
+#define CAT6_HIGH_COST_ENTRIES 64
+#endif
+extern const int av1_cat6_high_cost[CAT6_HIGH_COST_ENTRIES];
+extern const uint8_t av1_cat6_skipped_bits_discount[8];
+
+static INLINE void av1_get_token_extra(int v, int16_t *token, EXTRABIT *extra) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL) {
+    *token = CATEGORY6_TOKEN;
+    if (v >= CAT6_MIN_VAL)
+      *extra = 2 * v - 2 * CAT6_MIN_VAL;
+    else
+      *extra = -2 * v - 2 * CAT6_MIN_VAL + 1;
+    return;
+  }
+  *token = av1_dct_cat_lt_10_value_tokens[v].token;
+  *extra = av1_dct_cat_lt_10_value_tokens[v].extra;
+}
+static INLINE int16_t av1_get_token(int v) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL) return 10;
+  return av1_dct_cat_lt_10_value_tokens[v].token;
+}
+
+static INLINE int av1_get_token_cost(int v, int16_t *token, int cat6_bits) {
+  if (v >= CAT6_MIN_VAL || v <= -CAT6_MIN_VAL) {
+    EXTRABIT extrabits;
+    *token = CATEGORY6_TOKEN;
+    extrabits = abs(v) - CAT6_MIN_VAL;
+    return av1_cat6_low_cost[extrabits & 0xff] +
+           av1_cat6_high_cost[extrabits >> 8] -
+           av1_cat6_skipped_bits_discount[18 - cat6_bits];
+  }
+  *token = av1_dct_cat_lt_10_value_tokens[v].token;
+  return av1_dct_cat_lt_10_value_cost[v];
+}
+
+#if !CONFIG_PVQ || CONFIG_VAR_TX
+static INLINE int get_tx_eob(const struct segmentation *seg, int segment_id,
+                             TX_SIZE tx_size) {
+  const int eob_max = tx_size_2d[tx_size];
+  return segfeature_active(seg, segment_id, SEG_LVL_SKIP) ? 0 : eob_max;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_TOKENIZE_H_
diff --git a/third_party/aom/av1/encoder/treewriter.c b/third_party/aom/av1/encoder/treewriter.c
new file mode 100644
index 0000000000..50be72413b
--- /dev/null
+++ b/third_party/aom/av1/encoder/treewriter.c
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/encoder/treewriter.h"
+
+static void tree2tok(struct av1_token *tokens, const aom_tree_index *tree,
+                     int i, int v, int l) {
+  v += v;
+  ++l;
+
+  do {
+    const aom_tree_index j = tree[i++];
+    if (j <= 0) {
+      tokens[-j].value = v;
+      tokens[-j].len = l;
+    } else {
+      tree2tok(tokens, tree, j, v, l);
+    }
+  } while (++v & 1);
+}
+
+void av1_tokens_from_tree(struct av1_token *tokens,
+                          const aom_tree_index *tree) {
+  tree2tok(tokens, tree, 0, 0, 0);
+}
+
+static unsigned int convert_distribution(unsigned int i, aom_tree tree,
+                                         unsigned int branch_ct[][2],
+                                         const unsigned int num_events[]) {
+  unsigned int left, right;
+
+  if (tree[i] <= 0)
+    left = num_events[-tree[i]];
+  else
+    left = convert_distribution(tree[i], tree, branch_ct, num_events);
+
+  if (tree[i + 1] <= 0)
+    right = num_events[-tree[i + 1]];
+  else
+    right = convert_distribution(tree[i + 1], tree, branch_ct, num_events);
+
+  branch_ct[i >> 1][0] = left;
+  branch_ct[i >> 1][1] = right;
+  return left + right;
+}
+
+void av1_tree_probs_from_distribution(aom_tree tree,
+                                      unsigned int branch_ct[/* n-1 */][2],
+                                      const unsigned int num_events[/* n */]) {
+  convert_distribution(0, tree, branch_ct, num_events);
+}
diff --git a/third_party/aom/av1/encoder/treewriter.h b/third_party/aom/av1/encoder/treewriter.h
new file mode 100644
index 0000000000..9a4cb86cb2
--- /dev/null
+++ b/third_party/aom/av1/encoder/treewriter.h
@@ -0,0 +1,42 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_TREEWRITER_H_
+#define AV1_ENCODER_TREEWRITER_H_
+
+#include "aom_dsp/bitwriter.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void av1_tree_probs_from_distribution(aom_tree tree,
+                                      unsigned int branch_ct[/* n - 1 */][2],
+                                      const unsigned int num_events[/* n */]);
+
+struct av1_token {
+  int value;
+  int len;
+};
+
+void av1_tokens_from_tree(struct av1_token *, const aom_tree_index *);
+
+static INLINE void av1_write_token(aom_writer *w, const aom_tree_index *tree,
+                                   const aom_prob *probs,
+                                   const struct av1_token *token) {
+  aom_write_tree(w, tree, probs, token->value, token->len, 0);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AV1_ENCODER_TREEWRITER_H_
diff --git a/third_party/aom/av1/encoder/variance_tree.c b/third_party/aom/av1/encoder/variance_tree.c
new file mode 100644
index 0000000000..9384cd78ef
--- /dev/null
+++ b/third_party/aom/av1/encoder/variance_tree.c
@@ -0,0 +1,61 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/encoder/variance_tree.h"
+#include "av1/encoder/encoder.h"
+
+void av1_setup_var_tree(struct AV1Common *cm, ThreadData *td) {
+  int i, j;
+#if CONFIG_EXT_PARTITION
+  const int leaf_nodes = 1024;
+  const int tree_nodes = 1024 + 256 + 64 + 16 + 4 + 1;
+#else
+  const int leaf_nodes = 256;
+  const int tree_nodes = 256 + 64 + 16 + 4 + 1;
+#endif  // CONFIG_EXT_PARTITION
+  int index = 0;
+  VAR_TREE *this_var;
+  int nodes;
+
+  aom_free(td->var_tree);
+  CHECK_MEM_ERROR(cm, td->var_tree,
+                  aom_calloc(tree_nodes, sizeof(*td->var_tree)));
+
+  this_var = &td->var_tree[0];
+
+  // Sets up all the leaf nodes in the tree.
+  for (index = 0; index < leaf_nodes; ++index) {
+    VAR_TREE *const leaf = &td->var_tree[index];
+    leaf->split[0] = NULL;
+  }
+
+  // Each node has 4 leaf nodes, fill in the child pointers
+  // from leafs to the root.
+  for (nodes = leaf_nodes >> 2; nodes > 0; nodes >>= 2) {
+    for (i = 0; i < nodes; ++i, ++index) {
+      VAR_TREE *const node = &td->var_tree[index];
+      for (j = 0; j < 4; j++) node->split[j] = this_var++;
+    }
+  }
+
+  // Set up the root node for the largest superblock size
+  i = MAX_MIB_SIZE_LOG2 - MIN_MIB_SIZE_LOG2;
+  td->var_root[i] = &td->var_tree[tree_nodes - 1];
+  // Set up the root nodes for the rest of the possible superblock sizes
+  while (--i >= 0) {
+    td->var_root[i] = td->var_root[i + 1]->split[0];
+  }
+}
+
+void av1_free_var_tree(ThreadData *td) {
+  aom_free(td->var_tree);
+  td->var_tree = NULL;
+}
diff --git a/third_party/aom/av1/encoder/variance_tree.h b/third_party/aom/av1/encoder/variance_tree.h
new file mode 100644
index 0000000000..a9f27302e9
--- /dev/null
+++ b/third_party/aom/av1/encoder/variance_tree.h
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_ENCODER_VARIANCE_TREE_H_
+#define AV1_ENCODER_VARIANCE_TREE_H_
+
+#include <assert.h>
+
+#include "./aom_config.h"
+
+#include "aom/aom_integer.h"
+
+#include "av1/common/enums.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AV1Common;
+struct ThreadData;
+
+typedef struct {
+  int64_t sum_square_error;
+  int64_t sum_error;
+  int log2_count;
+  int variance;
+} VAR;
+
+typedef struct {
+  VAR none;
+  VAR horz[2];
+  VAR vert[2];
+} partition_variance;
+
+typedef struct VAR_TREE {
+  int force_split;
+  partition_variance variances;
+  struct VAR_TREE *split[4];
+  BLOCK_SIZE bsize;
+  const uint8_t *src;
+  const uint8_t *ref;
+  int src_stride;
+  int ref_stride;
+  int width;
+  int height;
+#if CONFIG_HIGHBITDEPTH
+  int highbd;
+#endif  // CONFIG_HIGHBITDEPTH
+} VAR_TREE;
+
+void av1_setup_var_tree(struct AV1Common *cm, struct ThreadData *td);
+void av1_free_var_tree(struct ThreadData *td);
+
+// Set variance values given sum square error, sum error, count.
+static INLINE void fill_variance(int64_t s2, int64_t s, int c, VAR *v) {
+  v->sum_square_error = s2;
+  v->sum_error = s;
+  v->log2_count = c;
+  v->variance =
+      (int)(256 * (v->sum_square_error -
+                   ((v->sum_error * v->sum_error) >> v->log2_count)) >>
+            v->log2_count);
+}
+
+static INLINE void sum_2_variances(const VAR *a, const VAR *b, VAR *r) {
+  assert(a->log2_count == b->log2_count);
+  fill_variance(a->sum_square_error + b->sum_square_error,
+                a->sum_error + b->sum_error, a->log2_count + 1, r);
+}
+
+static INLINE void fill_variance_node(VAR_TREE *vt) {
+  sum_2_variances(&vt->split[0]->variances.none, &vt->split[1]->variances.none,
+                  &vt->variances.horz[0]);
+  sum_2_variances(&vt->split[2]->variances.none, &vt->split[3]->variances.none,
+                  &vt->variances.horz[1]);
+  sum_2_variances(&vt->split[0]->variances.none, &vt->split[2]->variances.none,
+                  &vt->variances.vert[0]);
+  sum_2_variances(&vt->split[1]->variances.none, &vt->split[3]->variances.none,
+                  &vt->variances.vert[1]);
+  sum_2_variances(&vt->variances.vert[0], &vt->variances.vert[1],
+                  &vt->variances.none);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif /* AV1_ENCODER_VARIANCE_TREE_H_ */
diff --git a/third_party/aom/av1/encoder/wedge_utils.c b/third_party/aom/av1/encoder/wedge_utils.c
new file mode 100644
index 0000000000..e6edbb6af0
--- /dev/null
+++ b/third_party/aom/av1/encoder/wedge_utils.c
@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+
+#include "aom_ports/mem.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+
+#include "av1/common/reconinter.h"
+
+#define MAX_MASK_VALUE (1 << WEDGE_WEIGHT_BITS)
+
+/**
+ * Computes SSE of a compound predictor constructed from 2 fundamental
+ * predictors p0 and p1 using blending with mask.
+ *
+ * r1:  Residuals of p1.
+ *      (source - p1)
+ * d:   Difference of p1 and p0.
+ *      (p1 - p0)
+ * m:   The blending mask
+ * N:   Number of pixels
+ *
+ * 'r1', 'd', and 'm' are contiguous.
+ *
+ * Computes:
+ *  Sum((MAX_MASK_VALUE*r1 + mask*d)**2), which is equivalent to:
+ *  Sum((mask*r0 + (MAX_MASK_VALUE-mask)*r1)**2),
+ *    where r0 is (source - p0), and r1 is (source - p1), which is in turn
+ *    is equivalent to:
+ *  Sum((source*MAX_MASK_VALUE - (mask*p0 + (MAX_MASK_VALUE-mask)*p1))**2),
+ *    which is the SSE of the residuals of the compound predictor scaled up by
+ *    MAX_MASK_VALUE**2.
+ *
+ * Note that we clamp the partial term in the loop to 16 bits signed. This is
+ * to facilitate equivalent SIMD implementation. It should have no effect if
+ * residuals are within 16 - WEDGE_WEIGHT_BITS (=10) signed, which always
+ * holds for 8 bit input, and on real input, it should hold practically always,
+ * as residuals are expected to be small.
+ */
+uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d,
+                                        const uint8_t *m, int N) {
+  uint64_t csse = 0;
+  int i;
+
+  for (i = 0; i < N; i++) {
+    int32_t t = MAX_MASK_VALUE * r1[i] + m[i] * d[i];
+    t = clamp(t, INT16_MIN, INT16_MAX);
+    csse += t * t;
+  }
+  return ROUND_POWER_OF_TWO(csse, 2 * WEDGE_WEIGHT_BITS);
+}
+
+/**
+ * Choose the mask sign for a compound predictor.
+ *
+ * ds:    Difference of the squares of the residuals.
+ *        r0**2 - r1**2
+ * m:     The blending mask
+ * N:     Number of pixels
+ * limit: Pre-computed threshold value.
+ *        MAX_MASK_VALUE/2 * (sum(r0**2) - sum(r1**2))
+ *
+ * 'ds' and 'm' are contiguous.
+ *
+ * Returns true if the negated mask has lower SSE compared to the positive
+ * mask. Computation is based on:
+ *  Sum((mask*r0 + (MAX_MASK_VALUE-mask)*r1)**2)
+ *                                     >
+ *                                Sum(((MAX_MASK_VALUE-mask)*r0 + mask*r1)**2)
+ *
+ *  which can be simplified to:
+ *
+ *  Sum(mask*(r0**2 - r1**2)) > MAX_MASK_VALUE/2 * (sum(r0**2) - sum(r1**2))
+ *
+ *  The right hand side does not depend on the mask, and needs to be passed as
+ *  the 'limit' parameter.
+ *
+ *  After pre-computing (r0**2 - r1**2), which is passed in as 'ds', the left
+ *  hand side is simply a scalar product between an int16_t and uint8_t vector.
+ *
+ *  Note that for efficiency, ds is stored on 16 bits. Real input residuals
+ *  being small, this should not cause a noticeable issue.
+ */
+int av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N,
+                                    int64_t limit) {
+  int64_t acc = 0;
+
+  do {
+    acc += *ds++ * *m++;
+  } while (--N);
+
+  return acc > limit;
+}
+
+/**
+ * Compute the element-wise difference of the squares of 2 arrays.
+ *
+ * d: Difference of the squares of the inputs: a**2 - b**2
+ * a: First input array
+ * b: Second input array
+ * N: Number of elements
+ *
+ * 'd', 'a', and 'b' are contiguous.
+ *
+ * The result is saturated to signed 16 bits.
+ */
+void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a,
+                                       const int16_t *b, int N) {
+  int i;
+
+  for (i = 0; i < N; i++)
+    d[i] = clamp(a[i] * a[i] - b[i] * b[i], INT16_MIN, INT16_MAX);
+}
diff --git a/third_party/aom/av1/encoder/x86/av1_highbd_quantize_sse4.c b/third_party/aom/av1/encoder/x86/av1_highbd_quantize_sse4.c
new file mode 100644
index 0000000000..fa5626002f
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/av1_highbd_quantize_sse4.c
@@ -0,0 +1,193 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <smmintrin.h>
+#include <stdint.h>
+
+#include "./av1_rtcd.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+// Coefficient quantization phase 1
+// param[0-2] : rounding/quan/dequan constants
+static INLINE void quantize_coeff_phase1(__m128i *coeff, const __m128i *param,
+                                         const int shift, const int scale,
+                                         __m128i *qcoeff, __m128i *dquan,
+                                         __m128i *sign) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi32(1);
+
+  *sign = _mm_cmplt_epi32(*coeff, zero);
+  *sign = _mm_or_si128(*sign, one);
+  *coeff = _mm_abs_epi32(*coeff);
+
+  qcoeff[0] = _mm_add_epi32(*coeff, param[0]);
+  qcoeff[1] = _mm_unpackhi_epi32(qcoeff[0], zero);
+  qcoeff[0] = _mm_unpacklo_epi32(qcoeff[0], zero);
+
+  qcoeff[0] = _mm_mul_epi32(qcoeff[0], param[1]);
+  qcoeff[0] = _mm_srli_epi64(qcoeff[0], shift);
+  dquan[0] = _mm_mul_epi32(qcoeff[0], param[2]);
+  dquan[0] = _mm_srli_epi64(dquan[0], scale);
+}
+
+// Coefficient quantization phase 2
+static INLINE void quantize_coeff_phase2(__m128i *qcoeff, __m128i *dquan,
+                                         const __m128i *sign,
+                                         const __m128i *param, const int shift,
+                                         const int scale, tran_low_t *qAddr,
+                                         tran_low_t *dqAddr) {
+  __m128i mask0L = _mm_set_epi32(-1, -1, 0, 0);
+  __m128i mask0H = _mm_set_epi32(0, 0, -1, -1);
+
+  qcoeff[1] = _mm_mul_epi32(qcoeff[1], param[1]);
+  qcoeff[1] = _mm_srli_epi64(qcoeff[1], shift);
+  dquan[1] = _mm_mul_epi32(qcoeff[1], param[2]);
+  dquan[1] = _mm_srli_epi64(dquan[1], scale);
+
+  // combine L&H
+  qcoeff[0] = _mm_shuffle_epi32(qcoeff[0], 0xd8);
+  qcoeff[1] = _mm_shuffle_epi32(qcoeff[1], 0x8d);
+
+  qcoeff[0] = _mm_and_si128(qcoeff[0], mask0H);
+  qcoeff[1] = _mm_and_si128(qcoeff[1], mask0L);
+
+  dquan[0] = _mm_shuffle_epi32(dquan[0], 0xd8);
+  dquan[1] = _mm_shuffle_epi32(dquan[1], 0x8d);
+
+  dquan[0] = _mm_and_si128(dquan[0], mask0H);
+  dquan[1] = _mm_and_si128(dquan[1], mask0L);
+
+  qcoeff[0] = _mm_or_si128(qcoeff[0], qcoeff[1]);
+  dquan[0] = _mm_or_si128(dquan[0], dquan[1]);
+
+  qcoeff[0] = _mm_sign_epi32(qcoeff[0], *sign);
+  dquan[0] = _mm_sign_epi32(dquan[0], *sign);
+
+  _mm_storeu_si128((__m128i *)qAddr, qcoeff[0]);
+  _mm_storeu_si128((__m128i *)dqAddr, dquan[0]);
+}
+
+static INLINE void find_eob(tran_low_t *qcoeff_ptr, const int16_t *iscan,
+                            __m128i *eob) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i mask, iscanIdx;
+  const __m128i q0 = _mm_loadu_si128((__m128i const *)qcoeff_ptr);
+  const __m128i q1 = _mm_loadu_si128((__m128i const *)(qcoeff_ptr + 4));
+  __m128i nz_flag0 = _mm_cmpeq_epi32(q0, zero);
+  __m128i nz_flag1 = _mm_cmpeq_epi32(q1, zero);
+
+  nz_flag0 = _mm_cmpeq_epi32(nz_flag0, zero);
+  nz_flag1 = _mm_cmpeq_epi32(nz_flag1, zero);
+
+  mask = _mm_packs_epi32(nz_flag0, nz_flag1);
+  iscanIdx = _mm_loadu_si128((__m128i const *)iscan);
+  iscanIdx = _mm_sub_epi16(iscanIdx, mask);
+  iscanIdx = _mm_and_si128(iscanIdx, mask);
+  *eob = _mm_max_epi16(*eob, iscanIdx);
+}
+
+static INLINE uint16_t get_accumulated_eob(__m128i *eob) {
+  __m128i eob_shuffled;
+  uint16_t eobValue;
+  eob_shuffled = _mm_shuffle_epi32(*eob, 0xe);
+  *eob = _mm_max_epi16(*eob, eob_shuffled);
+  eob_shuffled = _mm_shufflelo_epi16(*eob, 0xe);
+  *eob = _mm_max_epi16(*eob, eob_shuffled);
+  eob_shuffled = _mm_shufflelo_epi16(*eob, 0x1);
+  *eob = _mm_max_epi16(*eob, eob_shuffled);
+  eobValue = _mm_extract_epi16(*eob, 0);
+  return eobValue;
+}
+
+void av1_highbd_quantize_fp_sse4_1(
+    const tran_low_t *coeff_ptr, intptr_t count, int skip_block,
+    const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan, int log_scale) {
+  __m128i coeff[2], qcoeff[2], dequant[2], qparam[3], coeff_sign;
+  __m128i eob = _mm_setzero_si128();
+  const tran_low_t *src = coeff_ptr;
+  tran_low_t *quanAddr = qcoeff_ptr;
+  tran_low_t *dquanAddr = dqcoeff_ptr;
+  const int shift = 16 - log_scale;
+  const int coeff_stride = 4;
+  const int quan_stride = coeff_stride;
+  (void)skip_block;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)scan;
+
+  memset(quanAddr, 0, count * sizeof(quanAddr[0]));
+  memset(dquanAddr, 0, count * sizeof(dquanAddr[0]));
+
+  if (!skip_block) {
+    coeff[0] = _mm_loadu_si128((__m128i const *)src);
+
+    qparam[0] =
+        _mm_set_epi32(round_ptr[1], round_ptr[1], round_ptr[1], round_ptr[0]);
+    qparam[1] = _mm_set_epi64x(quant_ptr[1], quant_ptr[0]);
+    qparam[2] = _mm_set_epi64x(dequant_ptr[1], dequant_ptr[0]);
+
+    // DC and first 3 AC
+    quantize_coeff_phase1(&coeff[0], qparam, shift, log_scale, qcoeff, dequant,
+                          &coeff_sign);
+
+    // update round/quan/dquan for AC
+    qparam[0] = _mm_unpackhi_epi64(qparam[0], qparam[0]);
+    qparam[1] = _mm_set_epi64x(quant_ptr[1], quant_ptr[1]);
+    qparam[2] = _mm_set_epi64x(dequant_ptr[1], dequant_ptr[1]);
+
+    quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift,
+                          log_scale, quanAddr, dquanAddr);
+
+    // next 4 AC
+    coeff[1] = _mm_loadu_si128((__m128i const *)(src + coeff_stride));
+    quantize_coeff_phase1(&coeff[1], qparam, shift, log_scale, qcoeff, dequant,
+                          &coeff_sign);
+    quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift,
+                          log_scale, quanAddr + quan_stride,
+                          dquanAddr + quan_stride);
+
+    find_eob(quanAddr, iscan, &eob);
+
+    count -= 8;
+
+    // loop for the rest of AC
+    while (count > 0) {
+      src += coeff_stride << 1;
+      quanAddr += quan_stride << 1;
+      dquanAddr += quan_stride << 1;
+      iscan += quan_stride << 1;
+
+      coeff[0] = _mm_loadu_si128((__m128i const *)src);
+      coeff[1] = _mm_loadu_si128((__m128i const *)(src + coeff_stride));
+
+      quantize_coeff_phase1(&coeff[0], qparam, shift, log_scale, qcoeff,
+                            dequant, &coeff_sign);
+      quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift,
+                            log_scale, quanAddr, dquanAddr);
+
+      quantize_coeff_phase1(&coeff[1], qparam, shift, log_scale, qcoeff,
+                            dequant, &coeff_sign);
+      quantize_coeff_phase2(qcoeff, dequant, &coeff_sign, qparam, shift,
+                            log_scale, quanAddr + quan_stride,
+                            dquanAddr + quan_stride);
+
+      find_eob(quanAddr, iscan, &eob);
+
+      count -= 8;
+    }
+    *eob_ptr = get_accumulated_eob(&eob);
+  } else {
+    *eob_ptr = 0;
+  }
+}
diff --git a/third_party/aom/av1/encoder/x86/av1_quantize_sse2.c b/third_party/aom/av1/encoder/x86/av1_quantize_sse2.c
new file mode 100644
index 0000000000..f9c95b6cb2
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/av1_quantize_sse2.c
@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>
+#include <xmmintrin.h>
+
+#include "./av1_rtcd.h"
+#include "aom/aom_integer.h"
+
+void av1_quantize_fp_sse2(const int16_t *coeff_ptr, intptr_t n_coeffs,
+                          int skip_block, const int16_t *zbin_ptr,
+                          const int16_t *round_ptr, const int16_t *quant_ptr,
+                          const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                          int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                          uint16_t *eob_ptr, const int16_t *scan_ptr,
+                          const int16_t *iscan_ptr) {
+  __m128i zero;
+  __m128i thr;
+  int16_t nzflag;
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+
+  coeff_ptr += n_coeffs;
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i *)round_ptr);
+        quant = _mm_load_si128((const __m128i *)quant_ptr);
+        dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = _mm_load_si128((const __m128i *)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i *)(coeff_ptr + n_coeffs) + 1);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    thr = _mm_srai_epi16(dequant, 1);
+
+    // AC only loop
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        coeff0 = _mm_load_si128((const __m128i *)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i *)(coeff_ptr + n_coeffs) + 1);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+                 _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
+
+        if (nzflag) {
+          qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+          qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+          qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+          qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+          // Reinsert signs
+          qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+          qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+          qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+          qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+          coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+          coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+        } else {
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero);
+
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero);
+        }
+      }
+
+      if (nzflag) {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
diff --git a/third_party/aom/av1/encoder/x86/av1_quantize_ssse3_x86_64.asm b/third_party/aom/av1/encoder/x86/av1_quantize_ssse3_x86_64.asm
new file mode 100644
index 0000000000..ad4ae274e2
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/av1_quantize_ssse3_x86_64.asm
@@ -0,0 +1,204 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%define private_prefix av1
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_1: times 8 dw 1
+
+SECTION .text
+
+%macro QUANTIZE_FP 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, \
+                                eob, scan, iscan
+  cmp                    dword skipm, 0
+  jne .blank
+
+  ; actual quantize loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, dequantmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+%ifidn %1, fp_32x32
+  pcmpeqw                         m5, m5
+  psrlw                           m5, 15
+  paddw                           m1, m5
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  mova                            m3, [r2q]                ; m3 = dequant
+  mov                             r3, qcoeffmp
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+%ifidn %1, fp_32x32
+  psllw                           m2, 1
+%endif
+  pxor                            m5, m5                   ; m5 = dedicated zero
+
+  lea                         coeffq, [  coeffq+ncoeffq*2]
+  lea                            r5q, [  r5q+ncoeffq*2]
+  lea                            r3q, [ r3q+ncoeffq*2]
+  lea                            r4q, [r4q+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpeqw                         m7, m7
+
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  mova            [r3q+ncoeffq*2+ 0], m8
+  mova            [r3q+ncoeffq*2+16], m13
+%ifidn %1, fp_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; r4[i] = r3[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; r4[i] = r3[i] * q
+%ifidn %1, fp_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+  psrlw                           m0, m3, 2
+%else
+  psrlw                           m0, m3, 1
+%endif
+  mova            [r4q+ncoeffq*2+ 0], m8
+  mova            [r4q+ncoeffq*2+16], m13
+  pcmpeqw                         m8, m5                   ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  r5q+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  r5q+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m7                   ; m11 = scan[i] + 1
+  pandn                           m8, m6                   ; m8 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jz .accumulate_eob
+
+.ac_only_loop:
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+
+  pcmpgtw                         m7, m6,  m0
+  pcmpgtw                        m12, m11, m0
+  pmovmskb                       r6d, m7
+  pmovmskb                       r2d, m12
+
+  or                              r6, r2
+  jz .skip_iter
+
+  pcmpeqw                         m7, m7
+
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  mova            [r3q+ncoeffq*2+ 0], m14
+  mova            [r3q+ncoeffq*2+16], m13
+%ifidn %1, fp_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; r4[i] = r3[i] * q
+  pmullw                         m13, m3                   ; r4[i] = r3[i] * q
+%ifidn %1, fp_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+  mova            [r4q+ncoeffq*2+ 0], m14
+  mova            [r4q+ncoeffq*2+16], m13
+  pcmpeqw                        m14, m5                   ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  r5q+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  r5q+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m7                   ; m11 = scan[i] + 1
+  pandn                          m14, m6                   ; m14 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+  jmp .accumulate_eob
+.skip_iter:
+  mova            [r3q+ncoeffq*2+ 0], m5
+  mova            [r3q+ncoeffq*2+16], m5
+  mova            [r4q+ncoeffq*2+ 0], m5
+  mova            [r4q+ncoeffq*2+16], m5
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+.accumulate_eob:
+  ; horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  pextrw                          r6, m8, 0
+  mov                           [r2], r6
+  RET
+
+  ; skip-block, i.e. just write all zeroes
+.blank:
+  mov                             r0, dqcoeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, qcoeffmp
+  mov                             r3, eobmp
+
+  lea                            r0q, [r0q+ncoeffq*2]
+  lea                            r2q, [r2q+ncoeffq*2]
+  neg                        ncoeffq
+  pxor                            m7, m7
+.blank_loop:
+  mova            [r0q+ncoeffq*2+ 0], m7
+  mova            [r0q+ncoeffq*2+16], m7
+  mova            [r2q+ncoeffq*2+ 0], m7
+  mova            [r2q+ncoeffq*2+16], m7
+  add                        ncoeffq, mmsize
+  jl .blank_loop
+  mov                     word [r3q], 0
+  RET
+%endmacro
+
+INIT_XMM ssse3
+QUANTIZE_FP fp, 7
+QUANTIZE_FP fp_32x32, 7
diff --git a/third_party/aom/av1/encoder/x86/av1_ssim_opt_x86_64.asm b/third_party/aom/av1/encoder/x86/av1_ssim_opt_x86_64.asm
new file mode 100644
index 0000000000..dcc697ba30
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/av1_ssim_opt_x86_64.asm
@@ -0,0 +1,219 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%include "aom_ports/x86_abi_support.asm"
+
+; tabulate_ssim - sums sum_s,sum_r,sum_sq_s,sum_sq_r, sum_sxr
+%macro TABULATE_SSIM 0
+        paddusw         xmm15, xmm3  ; sum_s
+        paddusw         xmm14, xmm4  ; sum_r
+        movdqa          xmm1, xmm3
+        pmaddwd         xmm1, xmm1
+        paddd           xmm13, xmm1 ; sum_sq_s
+        movdqa          xmm2, xmm4
+        pmaddwd         xmm2, xmm2
+        paddd           xmm12, xmm2 ; sum_sq_r
+        pmaddwd         xmm3, xmm4
+        paddd           xmm11, xmm3  ; sum_sxr
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_Q 1
+        movdqa          xmm2,%1
+        punpckldq       %1,xmm0
+        punpckhdq       xmm2,xmm0
+        paddq           %1,xmm2
+        movdqa          xmm2,%1
+        punpcklqdq      %1,xmm0
+        punpckhqdq      xmm2,xmm0
+        paddq           %1,xmm2
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_W 1
+        movdqa          xmm1, %1
+        punpcklwd       %1,xmm0
+        punpckhwd       xmm1,xmm0
+        paddd           %1, xmm1
+        SUM_ACROSS_Q    %1
+%endmacro
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    unsigned long *sum_s,
+;    unsigned long *sum_r,
+;    unsigned long *sum_sq_s,
+;    unsigned long *sum_sq_r,
+;    unsigned long *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+global sym(av1_ssim_parms_16x16_sse2) PRIVATE
+sym(av1_ssim_parms_16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 16      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movdqu          xmm5, [rsi]
+    movdqu          xmm6, [rdi]
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpckhbw       xmm3, xmm0 ; high_s
+    punpckhbw       xmm4, xmm0 ; high_r
+
+    TABULATE_SSIM
+
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    unsigned long *sum_s,
+;    unsigned long *sum_r,
+;    unsigned long *sum_sq_s,
+;    unsigned long *sum_sq_r,
+;    unsigned long *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+global sym(av1_ssim_parms_8x8_sse2) PRIVATE
+sym(av1_ssim_parms_8x8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 8      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movq            xmm3, [rsi]
+    movq            xmm4, [rdi]
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/av1/encoder/x86/dct_intrin_sse2.c b/third_party/aom/av1/encoder/x86/dct_intrin_sse2.c
new file mode 100644
index 0000000000..37c4b0d888
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/dct_intrin_sse2.c
@@ -0,0 +1,3884 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>  // SSE2
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_dsp/x86/fwd_txfm_sse2.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+#include "aom_ports/mem.h"
+
+static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in,
+                                   int stride, int flipud, int fliplr) {
+  const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
+  const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
+  __m128i mask;
+
+  if (!flipud) {
+    in[0] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+    in[1] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+    in[2] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
+    in[3] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
+  } else {
+    in[0] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
+    in[1] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
+    in[2] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+    in[3] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+  }
+
+  if (fliplr) {
+    in[0] = _mm_shufflelo_epi16(in[0], 0x1b);
+    in[1] = _mm_shufflelo_epi16(in[1], 0x1b);
+    in[2] = _mm_shufflelo_epi16(in[2], 0x1b);
+    in[3] = _mm_shufflelo_epi16(in[3], 0x1b);
+  }
+
+  in[0] = _mm_slli_epi16(in[0], 4);
+  in[1] = _mm_slli_epi16(in[1], 4);
+  in[2] = _mm_slli_epi16(in[2], 4);
+  in[3] = _mm_slli_epi16(in[3], 4);
+
+  mask = _mm_cmpeq_epi16(in[0], k__nonzero_bias_a);
+  in[0] = _mm_add_epi16(in[0], mask);
+  in[0] = _mm_add_epi16(in[0], k__nonzero_bias_b);
+}
+
+static INLINE void write_buffer_4x4(tran_low_t *output, __m128i *res) {
+  const __m128i kOne = _mm_set1_epi16(1);
+  __m128i in01 = _mm_unpacklo_epi64(res[0], res[1]);
+  __m128i in23 = _mm_unpacklo_epi64(res[2], res[3]);
+  __m128i out01 = _mm_add_epi16(in01, kOne);
+  __m128i out23 = _mm_add_epi16(in23, kOne);
+  out01 = _mm_srai_epi16(out01, 2);
+  out23 = _mm_srai_epi16(out23, 2);
+  store_output(&out01, (output + 0 * 8));
+  store_output(&out23, (output + 1 * 8));
+}
+
+static INLINE void transpose_4x4(__m128i *res) {
+  // Combine and transpose
+  // 00 01 02 03 20 21 22 23
+  // 10 11 12 13 30 31 32 33
+  const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
+  const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
+
+  // 00 10 01 11 02 12 03 13
+  // 20 30 21 31 22 32 23 33
+  res[0] = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  res[2] = _mm_unpackhi_epi32(tr0_0, tr0_1);
+
+  // 00 10 20 30 01 11 21 31
+  // 02 12 22 32 03 13 23 33
+  // only use the first 4 16-bit integers
+  res[1] = _mm_unpackhi_epi64(res[0], res[0]);
+  res[3] = _mm_unpackhi_epi64(res[2], res[2]);
+}
+
+static void fdct4_sse2(__m128i *in) {
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i u[4], v[4];
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpacklo_epi16(in[3], in[2]);
+
+  v[0] = _mm_add_epi16(u[0], u[1]);
+  v[1] = _mm_sub_epi16(u[0], u[1]);
+
+  u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);  // 0
+  u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16);  // 2
+  u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24);  // 1
+  u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08);  // 3
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[1]);
+  in[1] = _mm_packs_epi32(u[2], u[3]);
+  transpose_4x4(in);
+}
+
+static void fadst4_sse2(__m128i *in) {
+  const __m128i k__sinpi_p01_p02 = pair_set_epi16(sinpi_1_9, sinpi_2_9);
+  const __m128i k__sinpi_p04_m01 = pair_set_epi16(sinpi_4_9, -sinpi_1_9);
+  const __m128i k__sinpi_p03_p04 = pair_set_epi16(sinpi_3_9, sinpi_4_9);
+  const __m128i k__sinpi_m03_p02 = pair_set_epi16(-sinpi_3_9, sinpi_2_9);
+  const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8];
+  __m128i in7 = _mm_add_epi16(in[0], in[1]);
+
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpacklo_epi16(in[2], in[3]);
+  u[2] = _mm_unpacklo_epi16(in7, kZero);
+  u[3] = _mm_unpacklo_epi16(in[2], kZero);
+  u[4] = _mm_unpacklo_epi16(in[3], kZero);
+
+  v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02);  // s0 + s2
+  v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04);  // s4 + s5
+  v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x1
+  v[3] = _mm_madd_epi16(u[0], k__sinpi_p04_m01);  // s1 - s3
+  v[4] = _mm_madd_epi16(u[1], k__sinpi_m03_p02);  // -s4 + s6
+  v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s4
+  v[6] = _mm_madd_epi16(u[4], k__sinpi_p03_p03);
+
+  u[0] = _mm_add_epi32(v[0], v[1]);
+  u[1] = _mm_sub_epi32(v[2], v[6]);
+  u[2] = _mm_add_epi32(v[3], v[4]);
+  u[3] = _mm_sub_epi32(u[2], u[0]);
+  u[4] = _mm_slli_epi32(v[5], 2);
+  u[5] = _mm_sub_epi32(u[4], v[5]);
+  u[6] = _mm_add_epi32(u[3], u[5]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[2]);
+  in[1] = _mm_packs_epi32(u[1], u[3]);
+  transpose_4x4(in);
+}
+
+#if CONFIG_EXT_TX
+static void fidtx4_sse2(__m128i *in) {
+  const __m128i k__zero_epi16 = _mm_set1_epi16((int16_t)0);
+  const __m128i k__sqrt2_epi16 = _mm_set1_epi16((int16_t)Sqrt2);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i v0, v1, v2, v3;
+  __m128i u0, u1, u2, u3;
+
+  v0 = _mm_unpacklo_epi16(in[0], k__zero_epi16);
+  v1 = _mm_unpacklo_epi16(in[1], k__zero_epi16);
+  v2 = _mm_unpacklo_epi16(in[2], k__zero_epi16);
+  v3 = _mm_unpacklo_epi16(in[3], k__zero_epi16);
+
+  u0 = _mm_madd_epi16(v0, k__sqrt2_epi16);
+  u1 = _mm_madd_epi16(v1, k__sqrt2_epi16);
+  u2 = _mm_madd_epi16(v2, k__sqrt2_epi16);
+  u3 = _mm_madd_epi16(v3, k__sqrt2_epi16);
+
+  v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u0, u2);
+  in[1] = _mm_packs_epi32(u1, u3);
+  transpose_4x4(in);
+}
+#endif  // CONFIG_EXT_TX
+
+void av1_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride,
+                     int tx_type) {
+  __m128i in[4];
+
+  switch (tx_type) {
+    case DCT_DCT: aom_fdct4x4_sse2(input, output, stride); break;
+    case ADST_DCT:
+      load_buffer_4x4(input, in, stride, 0, 0);
+      fadst4_sse2(in);
+      fdct4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case DCT_ADST:
+      load_buffer_4x4(input, in, stride, 0, 0);
+      fdct4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case ADST_ADST:
+      load_buffer_4x4(input, in, stride, 0, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_4x4(input, in, stride, 1, 0);
+      fadst4_sse2(in);
+      fdct4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_4x4(input, in, stride, 0, 1);
+      fdct4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_4x4(input, in, stride, 1, 1);
+      fadst4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_4x4(input, in, stride, 0, 1);
+      fadst4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_4x4(input, in, stride, 1, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case IDTX:
+      load_buffer_4x4(input, in, stride, 0, 0);
+      fidtx4_sse2(in);
+      fidtx4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case V_DCT:
+      load_buffer_4x4(input, in, stride, 0, 0);
+      fdct4_sse2(in);
+      fidtx4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case H_DCT:
+      load_buffer_4x4(input, in, stride, 0, 0);
+      fidtx4_sse2(in);
+      fdct4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case V_ADST:
+      load_buffer_4x4(input, in, stride, 0, 0);
+      fadst4_sse2(in);
+      fidtx4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case H_ADST:
+      load_buffer_4x4(input, in, stride, 0, 0);
+      fidtx4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case V_FLIPADST:
+      load_buffer_4x4(input, in, stride, 1, 0);
+      fadst4_sse2(in);
+      fidtx4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case H_FLIPADST:
+      load_buffer_4x4(input, in, stride, 0, 1);
+      fidtx4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+}
+
+void av1_fdct8x8_quant_sse2(const int16_t *input, int stride,
+                            int16_t *coeff_ptr, intptr_t n_coeffs,
+                            int skip_block, const int16_t *zbin_ptr,
+                            const int16_t *round_ptr, const int16_t *quant_ptr,
+                            const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
+                            int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                            uint16_t *eob_ptr, const int16_t *scan_ptr,
+                            const int16_t *iscan_ptr) {
+  __m128i zero;
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // Load input
+  __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  __m128i *in[8];
+  int index = 0;
+
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)coeff_ptr;
+
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  in[0] = &in0;
+  in[1] = &in1;
+  in[2] = &in2;
+  in[3] = &in3;
+  in[4] = &in4;
+  in[5] = &in5;
+  in[6] = &in6;
+  in[7] = &in7;
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = _mm_add_epi16(in0, in7);
+    const __m128i q1 = _mm_add_epi16(in1, in6);
+    const __m128i q2 = _mm_add_epi16(in2, in5);
+    const __m128i q3 = _mm_add_epi16(in3, in4);
+    const __m128i q4 = _mm_sub_epi16(in3, in4);
+    const __m128i q5 = _mm_sub_epi16(in2, in5);
+    const __m128i q6 = _mm_sub_epi16(in1, in6);
+    const __m128i q7 = _mm_sub_epi16(in0, in7);
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = _mm_add_epi16(q0, q3);
+      const __m128i r1 = _mm_add_epi16(q1, q2);
+      const __m128i r2 = _mm_sub_epi16(q1, q2);
+      const __m128i r3 = _mm_sub_epi16(q0, q3);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+      const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+      const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+      const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+      // dct_const_round_shift
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+      res0 = _mm_packs_epi32(w0, w1);
+      res4 = _mm_packs_epi32(w2, w3);
+      res2 = _mm_packs_epi32(w4, w5);
+      res6 = _mm_packs_epi32(w6, w7);
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
+      const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
+      const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
+      const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
+      const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
+      const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
+      // dct_const_round_shift
+      const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
+      const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
+      const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
+      const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
+      const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
+      const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
+      const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
+      const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
+      // Combine
+      const __m128i r0 = _mm_packs_epi32(s0, s1);
+      const __m128i r1 = _mm_packs_epi32(s2, s3);
+      // Add/subtract
+      const __m128i x0 = _mm_add_epi16(q4, r0);
+      const __m128i x1 = _mm_sub_epi16(q4, r0);
+      const __m128i x2 = _mm_sub_epi16(q7, r1);
+      const __m128i x3 = _mm_add_epi16(q7, r1);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+      const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+      const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+      const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+      // dct_const_round_shift
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+      res1 = _mm_packs_epi32(w0, w1);
+      res7 = _mm_packs_epi32(w2, w3);
+      res5 = _mm_packs_epi32(w4, w5);
+      res3 = _mm_packs_epi32(w6, w7);
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+  }
+
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i *)round_ptr);
+        quant = _mm_load_si128((const __m128i *)quant_ptr);
+        dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = *in[0];
+        coeff1 = *in[1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // AC only loop
+    index = 2;
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
+        coeff0 = *in[index];
+        coeff1 = *in[index + 1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+      index += 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
+
+// load 8x8 array
+static INLINE void load_buffer_8x8(const int16_t *input, __m128i *in,
+                                   int stride, int flipud, int fliplr) {
+  if (!flipud) {
+    in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride));
+    in[1] = _mm_load_si128((const __m128i *)(input + 1 * stride));
+    in[2] = _mm_load_si128((const __m128i *)(input + 2 * stride));
+    in[3] = _mm_load_si128((const __m128i *)(input + 3 * stride));
+    in[4] = _mm_load_si128((const __m128i *)(input + 4 * stride));
+    in[5] = _mm_load_si128((const __m128i *)(input + 5 * stride));
+    in[6] = _mm_load_si128((const __m128i *)(input + 6 * stride));
+    in[7] = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  } else {
+    in[0] = _mm_load_si128((const __m128i *)(input + 7 * stride));
+    in[1] = _mm_load_si128((const __m128i *)(input + 6 * stride));
+    in[2] = _mm_load_si128((const __m128i *)(input + 5 * stride));
+    in[3] = _mm_load_si128((const __m128i *)(input + 4 * stride));
+    in[4] = _mm_load_si128((const __m128i *)(input + 3 * stride));
+    in[5] = _mm_load_si128((const __m128i *)(input + 2 * stride));
+    in[6] = _mm_load_si128((const __m128i *)(input + 1 * stride));
+    in[7] = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  }
+
+  if (fliplr) {
+    in[0] = mm_reverse_epi16(in[0]);
+    in[1] = mm_reverse_epi16(in[1]);
+    in[2] = mm_reverse_epi16(in[2]);
+    in[3] = mm_reverse_epi16(in[3]);
+    in[4] = mm_reverse_epi16(in[4]);
+    in[5] = mm_reverse_epi16(in[5]);
+    in[6] = mm_reverse_epi16(in[6]);
+    in[7] = mm_reverse_epi16(in[7]);
+  }
+
+  in[0] = _mm_slli_epi16(in[0], 2);
+  in[1] = _mm_slli_epi16(in[1], 2);
+  in[2] = _mm_slli_epi16(in[2], 2);
+  in[3] = _mm_slli_epi16(in[3], 2);
+  in[4] = _mm_slli_epi16(in[4], 2);
+  in[5] = _mm_slli_epi16(in[5], 2);
+  in[6] = _mm_slli_epi16(in[6], 2);
+  in[7] = _mm_slli_epi16(in[7], 2);
+}
+
+// right shift and rounding
+static INLINE void right_shift_8x8(__m128i *res, const int bit) {
+  __m128i sign0 = _mm_srai_epi16(res[0], 15);
+  __m128i sign1 = _mm_srai_epi16(res[1], 15);
+  __m128i sign2 = _mm_srai_epi16(res[2], 15);
+  __m128i sign3 = _mm_srai_epi16(res[3], 15);
+  __m128i sign4 = _mm_srai_epi16(res[4], 15);
+  __m128i sign5 = _mm_srai_epi16(res[5], 15);
+  __m128i sign6 = _mm_srai_epi16(res[6], 15);
+  __m128i sign7 = _mm_srai_epi16(res[7], 15);
+
+  if (bit == 2) {
+    const __m128i const_rounding = _mm_set1_epi16(1);
+    res[0] = _mm_adds_epi16(res[0], const_rounding);
+    res[1] = _mm_adds_epi16(res[1], const_rounding);
+    res[2] = _mm_adds_epi16(res[2], const_rounding);
+    res[3] = _mm_adds_epi16(res[3], const_rounding);
+    res[4] = _mm_adds_epi16(res[4], const_rounding);
+    res[5] = _mm_adds_epi16(res[5], const_rounding);
+    res[6] = _mm_adds_epi16(res[6], const_rounding);
+    res[7] = _mm_adds_epi16(res[7], const_rounding);
+  }
+
+  res[0] = _mm_sub_epi16(res[0], sign0);
+  res[1] = _mm_sub_epi16(res[1], sign1);
+  res[2] = _mm_sub_epi16(res[2], sign2);
+  res[3] = _mm_sub_epi16(res[3], sign3);
+  res[4] = _mm_sub_epi16(res[4], sign4);
+  res[5] = _mm_sub_epi16(res[5], sign5);
+  res[6] = _mm_sub_epi16(res[6], sign6);
+  res[7] = _mm_sub_epi16(res[7], sign7);
+
+  if (bit == 1) {
+    res[0] = _mm_srai_epi16(res[0], 1);
+    res[1] = _mm_srai_epi16(res[1], 1);
+    res[2] = _mm_srai_epi16(res[2], 1);
+    res[3] = _mm_srai_epi16(res[3], 1);
+    res[4] = _mm_srai_epi16(res[4], 1);
+    res[5] = _mm_srai_epi16(res[5], 1);
+    res[6] = _mm_srai_epi16(res[6], 1);
+    res[7] = _mm_srai_epi16(res[7], 1);
+  } else {
+    res[0] = _mm_srai_epi16(res[0], 2);
+    res[1] = _mm_srai_epi16(res[1], 2);
+    res[2] = _mm_srai_epi16(res[2], 2);
+    res[3] = _mm_srai_epi16(res[3], 2);
+    res[4] = _mm_srai_epi16(res[4], 2);
+    res[5] = _mm_srai_epi16(res[5], 2);
+    res[6] = _mm_srai_epi16(res[6], 2);
+    res[7] = _mm_srai_epi16(res[7], 2);
+  }
+}
+
+// write 8x8 array
+static INLINE void write_buffer_8x8(tran_low_t *output, __m128i *res,
+                                    int stride) {
+  store_output(&res[0], (output + 0 * stride));
+  store_output(&res[1], (output + 1 * stride));
+  store_output(&res[2], (output + 2 * stride));
+  store_output(&res[3], (output + 3 * stride));
+  store_output(&res[4], (output + 4 * stride));
+  store_output(&res[5], (output + 5 * stride));
+  store_output(&res[6], (output + 6 * stride));
+  store_output(&res[7], (output + 7 * stride));
+}
+
+// perform in-place transpose
+static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
+  // 00 10 01 11 02 12 03 13
+  // 20 30 21 31 22 32 23 33
+  // 04 14 05 15 06 16 07 17
+  // 24 34 25 35 26 36 27 37
+  // 40 50 41 51 42 52 43 53
+  // 60 70 61 71 62 72 63 73
+  // 44 54 45 55 46 56 47 57
+  // 64 74 65 75 66 76 67 77
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+  // 00 10 20 30 01 11 21 31
+  // 40 50 60 70 41 51 61 71
+  // 02 12 22 32 03 13 23 33
+  // 42 52 62 72 43 53 63 73
+  // 04 14 24 34 05 15 25 35
+  // 44 54 64 74 45 55 65 75
+  // 06 16 26 36 07 17 27 37
+  // 46 56 66 76 47 57 67 77
+  res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
+  res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
+  res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
+  res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
+  res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
+  res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
+  res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
+  res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
+  // 00 10 20 30 40 50 60 70
+  // 01 11 21 31 41 51 61 71
+  // 02 12 22 32 42 52 62 72
+  // 03 13 23 33 43 53 63 73
+  // 04 14 24 34 44 54 64 74
+  // 05 15 25 35 45 55 65 75
+  // 06 16 26 36 46 56 66 76
+  // 07 17 27 37 47 57 67 77
+}
+
+static void fdct8_sse2(__m128i *in) {
+  // constants
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+
+  // stage 1
+  s0 = _mm_add_epi16(in[0], in[7]);
+  s1 = _mm_add_epi16(in[1], in[6]);
+  s2 = _mm_add_epi16(in[2], in[5]);
+  s3 = _mm_add_epi16(in[3], in[4]);
+  s4 = _mm_sub_epi16(in[3], in[4]);
+  s5 = _mm_sub_epi16(in[2], in[5]);
+  s6 = _mm_sub_epi16(in[1], in[6]);
+  s7 = _mm_sub_epi16(in[0], in[7]);
+
+  u0 = _mm_add_epi16(s0, s3);
+  u1 = _mm_add_epi16(s1, s2);
+  u2 = _mm_sub_epi16(s1, s2);
+  u3 = _mm_sub_epi16(s0, s3);
+  // interleave and perform butterfly multiplication/addition
+  v0 = _mm_unpacklo_epi16(u0, u1);
+  v1 = _mm_unpackhi_epi16(u0, u1);
+  v2 = _mm_unpacklo_epi16(u2, u3);
+  v3 = _mm_unpackhi_epi16(u2, u3);
+
+  u0 = _mm_madd_epi16(v0, k__cospi_p16_p16);
+  u1 = _mm_madd_epi16(v1, k__cospi_p16_p16);
+  u2 = _mm_madd_epi16(v0, k__cospi_p16_m16);
+  u3 = _mm_madd_epi16(v1, k__cospi_p16_m16);
+  u4 = _mm_madd_epi16(v2, k__cospi_p24_p08);
+  u5 = _mm_madd_epi16(v3, k__cospi_p24_p08);
+  u6 = _mm_madd_epi16(v2, k__cospi_m08_p24);
+  u7 = _mm_madd_epi16(v3, k__cospi_m08_p24);
+
+  // shift and rounding
+  v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u0, u1);
+  in[2] = _mm_packs_epi32(u4, u5);
+  in[4] = _mm_packs_epi32(u2, u3);
+  in[6] = _mm_packs_epi32(u6, u7);
+
+  // stage 2
+  // interleave and perform butterfly multiplication/addition
+  u0 = _mm_unpacklo_epi16(s6, s5);
+  u1 = _mm_unpackhi_epi16(s6, s5);
+  v0 = _mm_madd_epi16(u0, k__cospi_p16_m16);
+  v1 = _mm_madd_epi16(u1, k__cospi_p16_m16);
+  v2 = _mm_madd_epi16(u0, k__cospi_p16_p16);
+  v3 = _mm_madd_epi16(u1, k__cospi_p16_p16);
+
+  // shift and rounding
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+
+  u0 = _mm_packs_epi32(v0, v1);
+  u1 = _mm_packs_epi32(v2, v3);
+
+  // stage 3
+  s0 = _mm_add_epi16(s4, u0);
+  s1 = _mm_sub_epi16(s4, u0);
+  s2 = _mm_sub_epi16(s7, u1);
+  s3 = _mm_add_epi16(s7, u1);
+
+  // stage 4
+  u0 = _mm_unpacklo_epi16(s0, s3);
+  u1 = _mm_unpackhi_epi16(s0, s3);
+  u2 = _mm_unpacklo_epi16(s1, s2);
+  u3 = _mm_unpackhi_epi16(s1, s2);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p28_p04);
+  v1 = _mm_madd_epi16(u1, k__cospi_p28_p04);
+  v2 = _mm_madd_epi16(u2, k__cospi_p12_p20);
+  v3 = _mm_madd_epi16(u3, k__cospi_p12_p20);
+  v4 = _mm_madd_epi16(u2, k__cospi_m20_p12);
+  v5 = _mm_madd_epi16(u3, k__cospi_m20_p12);
+  v6 = _mm_madd_epi16(u0, k__cospi_m04_p28);
+  v7 = _mm_madd_epi16(u1, k__cospi_m04_p28);
+
+  // shift and rounding
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  in[1] = _mm_packs_epi32(v0, v1);
+  in[3] = _mm_packs_epi32(v4, v5);
+  in[5] = _mm_packs_epi32(v2, v3);
+  in[7] = _mm_packs_epi32(v6, v7);
+
+  // transpose
+  array_transpose_8x8(in, in);
+}
+
+static void fadst8_sse2(__m128i *in) {
+  // Constants
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__const_0 = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+
+  // properly aligned for butterfly input
+  in0 = in[7];
+  in1 = in[0];
+  in2 = in[5];
+  in3 = in[2];
+  in4 = in[3];
+  in5 = in[4];
+  in6 = in[1];
+  in7 = in[6];
+
+  // column transformation
+  // stage 1
+  // interleave and multiply/add into 32-bit integer
+  s0 = _mm_unpacklo_epi16(in0, in1);
+  s1 = _mm_unpackhi_epi16(in0, in1);
+  s2 = _mm_unpacklo_epi16(in2, in3);
+  s3 = _mm_unpackhi_epi16(in2, in3);
+  s4 = _mm_unpacklo_epi16(in4, in5);
+  s5 = _mm_unpackhi_epi16(in4, in5);
+  s6 = _mm_unpacklo_epi16(in6, in7);
+  s7 = _mm_unpackhi_epi16(in6, in7);
+
+  u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
+  u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
+  u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
+  u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
+  u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
+  u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
+  u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
+  u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
+  u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
+  u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
+  u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
+  u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
+  u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
+  u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
+  u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
+  u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
+
+  // addition
+  w0 = _mm_add_epi32(u0, u8);
+  w1 = _mm_add_epi32(u1, u9);
+  w2 = _mm_add_epi32(u2, u10);
+  w3 = _mm_add_epi32(u3, u11);
+  w4 = _mm_add_epi32(u4, u12);
+  w5 = _mm_add_epi32(u5, u13);
+  w6 = _mm_add_epi32(u6, u14);
+  w7 = _mm_add_epi32(u7, u15);
+  w8 = _mm_sub_epi32(u0, u8);
+  w9 = _mm_sub_epi32(u1, u9);
+  w10 = _mm_sub_epi32(u2, u10);
+  w11 = _mm_sub_epi32(u3, u11);
+  w12 = _mm_sub_epi32(u4, u12);
+  w13 = _mm_sub_epi32(u5, u13);
+  w14 = _mm_sub_epi32(u6, u14);
+  w15 = _mm_sub_epi32(u7, u15);
+
+  // shift and rounding
+  v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
+  v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
+  v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
+  v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
+  v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
+  v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
+  v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
+  v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
+
+  u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
+  u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
+  u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
+  u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
+  u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
+  u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
+  u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
+  u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
+
+  // back to 16-bit and pack 8 integers into __m128i
+  v0 = _mm_add_epi32(w0, w4);
+  v1 = _mm_add_epi32(w1, w5);
+  v2 = _mm_add_epi32(w2, w6);
+  v3 = _mm_add_epi32(w3, w7);
+  v4 = _mm_sub_epi32(w0, w4);
+  v5 = _mm_sub_epi32(w1, w5);
+  v6 = _mm_sub_epi32(w2, w6);
+  v7 = _mm_sub_epi32(w3, w7);
+
+  w0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  w1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  w2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  w3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  w4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  w5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  w6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  w7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(w0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(w1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(w2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(w3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(w4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(w5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(w6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(w7, DCT_CONST_BITS);
+
+  in[4] = _mm_packs_epi32(u8, u9);
+  in[5] = _mm_packs_epi32(u10, u11);
+  in[6] = _mm_packs_epi32(u12, u13);
+  in[7] = _mm_packs_epi32(u14, u15);
+
+  // stage 2
+  s0 = _mm_packs_epi32(v0, v1);
+  s1 = _mm_packs_epi32(v2, v3);
+  s2 = _mm_packs_epi32(v4, v5);
+  s3 = _mm_packs_epi32(v6, v7);
+
+  u0 = _mm_unpacklo_epi16(in[4], in[5]);
+  u1 = _mm_unpackhi_epi16(in[4], in[5]);
+  u2 = _mm_unpacklo_epi16(in[6], in[7]);
+  u3 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
+  v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
+  v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
+  v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
+  v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
+  v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
+  v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
+  v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
+
+  w0 = _mm_add_epi32(v0, v4);
+  w1 = _mm_add_epi32(v1, v5);
+  w2 = _mm_add_epi32(v2, v6);
+  w3 = _mm_add_epi32(v3, v7);
+  w4 = _mm_sub_epi32(v0, v4);
+  w5 = _mm_sub_epi32(v1, v5);
+  w6 = _mm_sub_epi32(v2, v6);
+  w7 = _mm_sub_epi32(v3, v7);
+
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  // back to 16-bit intergers
+  s4 = _mm_packs_epi32(u0, u1);
+  s5 = _mm_packs_epi32(u2, u3);
+  s6 = _mm_packs_epi32(u4, u5);
+  s7 = _mm_packs_epi32(u6, u7);
+
+  // stage 3
+  u0 = _mm_unpacklo_epi16(s2, s3);
+  u1 = _mm_unpackhi_epi16(s2, s3);
+  u2 = _mm_unpacklo_epi16(s6, s7);
+  u3 = _mm_unpackhi_epi16(s6, s7);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
+  v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
+  v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
+  v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
+  v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
+  v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
+  v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
+  v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
+
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  s2 = _mm_packs_epi32(v0, v1);
+  s3 = _mm_packs_epi32(v2, v3);
+  s6 = _mm_packs_epi32(v4, v5);
+  s7 = _mm_packs_epi32(v6, v7);
+
+  // FIXME(jingning): do subtract using bit inversion?
+  in[0] = s0;
+  in[1] = _mm_sub_epi16(k__const_0, s4);
+  in[2] = s6;
+  in[3] = _mm_sub_epi16(k__const_0, s2);
+  in[4] = s3;
+  in[5] = _mm_sub_epi16(k__const_0, s7);
+  in[6] = s5;
+  in[7] = _mm_sub_epi16(k__const_0, s1);
+
+  // transpose
+  array_transpose_8x8(in, in);
+}
+
+#if CONFIG_EXT_TX
+static void fidtx8_sse2(__m128i *in) {
+  in[0] = _mm_slli_epi16(in[0], 1);
+  in[1] = _mm_slli_epi16(in[1], 1);
+  in[2] = _mm_slli_epi16(in[2], 1);
+  in[3] = _mm_slli_epi16(in[3], 1);
+  in[4] = _mm_slli_epi16(in[4], 1);
+  in[5] = _mm_slli_epi16(in[5], 1);
+  in[6] = _mm_slli_epi16(in[6], 1);
+  in[7] = _mm_slli_epi16(in[7], 1);
+
+  array_transpose_8x8(in, in);
+}
+#endif  // CONFIG_EXT_TX
+
+void av1_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride,
+                     int tx_type) {
+  __m128i in[8];
+
+  switch (tx_type) {
+    case DCT_DCT: aom_fdct8x8_sse2(input, output, stride); break;
+    case ADST_DCT:
+      load_buffer_8x8(input, in, stride, 0, 0);
+      fadst8_sse2(in);
+      fdct8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case DCT_ADST:
+      load_buffer_8x8(input, in, stride, 0, 0);
+      fdct8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case ADST_ADST:
+      load_buffer_8x8(input, in, stride, 0, 0);
+      fadst8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_8x8(input, in, stride, 1, 0);
+      fadst8_sse2(in);
+      fdct8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_8x8(input, in, stride, 0, 1);
+      fdct8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_8x8(input, in, stride, 1, 1);
+      fadst8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_8x8(input, in, stride, 0, 1);
+      fadst8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_8x8(input, in, stride, 1, 0);
+      fadst8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case IDTX:
+      load_buffer_8x8(input, in, stride, 0, 0);
+      fidtx8_sse2(in);
+      fidtx8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case V_DCT:
+      load_buffer_8x8(input, in, stride, 0, 0);
+      fdct8_sse2(in);
+      fidtx8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case H_DCT:
+      load_buffer_8x8(input, in, stride, 0, 0);
+      fidtx8_sse2(in);
+      fdct8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case V_ADST:
+      load_buffer_8x8(input, in, stride, 0, 0);
+      fadst8_sse2(in);
+      fidtx8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case H_ADST:
+      load_buffer_8x8(input, in, stride, 0, 0);
+      fidtx8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case V_FLIPADST:
+      load_buffer_8x8(input, in, stride, 1, 0);
+      fadst8_sse2(in);
+      fidtx8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case H_FLIPADST:
+      load_buffer_8x8(input, in, stride, 0, 1);
+      fidtx8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+}
+
+static INLINE void load_buffer_16x16(const int16_t *input, __m128i *in0,
+                                     __m128i *in1, int stride, int flipud,
+                                     int fliplr) {
+  // Load 4 8x8 blocks
+  const int16_t *topL = input;
+  const int16_t *topR = input + 8;
+  const int16_t *botL = input + 8 * stride;
+  const int16_t *botR = input + 8 * stride + 8;
+
+  const int16_t *tmp;
+
+  if (flipud) {
+    // Swap left columns
+    tmp = topL;
+    topL = botL;
+    botL = tmp;
+    // Swap right columns
+    tmp = topR;
+    topR = botR;
+    botR = tmp;
+  }
+
+  if (fliplr) {
+    // Swap top rows
+    tmp = topL;
+    topL = topR;
+    topR = tmp;
+    // Swap bottom rows
+    tmp = botL;
+    botL = botR;
+    botR = tmp;
+  }
+
+  // load first 8 columns
+  load_buffer_8x8(topL, in0, stride, flipud, fliplr);
+  load_buffer_8x8(botL, in0 + 8, stride, flipud, fliplr);
+
+  // load second 8 columns
+  load_buffer_8x8(topR, in1, stride, flipud, fliplr);
+  load_buffer_8x8(botR, in1 + 8, stride, flipud, fliplr);
+}
+
+static INLINE void write_buffer_16x16(tran_low_t *output, __m128i *in0,
+                                      __m128i *in1, int stride) {
+  // write first 8 columns
+  write_buffer_8x8(output, in0, stride);
+  write_buffer_8x8(output + 8 * stride, in0 + 8, stride);
+  // write second 8 columns
+  output += 8;
+  write_buffer_8x8(output, in1, stride);
+  write_buffer_8x8(output + 8 * stride, in1 + 8, stride);
+}
+
+static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
+  __m128i tbuf[8];
+  array_transpose_8x8(res0, res0);
+  array_transpose_8x8(res1, tbuf);
+  array_transpose_8x8(res0 + 8, res1);
+  array_transpose_8x8(res1 + 8, res1 + 8);
+
+  res0[8] = tbuf[0];
+  res0[9] = tbuf[1];
+  res0[10] = tbuf[2];
+  res0[11] = tbuf[3];
+  res0[12] = tbuf[4];
+  res0[13] = tbuf[5];
+  res0[14] = tbuf[6];
+  res0[15] = tbuf[7];
+}
+
+static INLINE void right_shift_16x16(__m128i *res0, __m128i *res1) {
+  // perform rounding operations
+  right_shift_8x8(res0, 2);
+  right_shift_8x8(res0 + 8, 2);
+  right_shift_8x8(res1, 2);
+  right_shift_8x8(res1 + 8, 2);
+}
+
+static void fdct16_8col(__m128i *in) {
+  // perform 16x16 1-D DCT for 8 columns
+  __m128i i[8], s[8], p[8], t[8], u[16], v[16];
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
+  const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
+  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
+  const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
+  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  // stage 1
+  i[0] = _mm_add_epi16(in[0], in[15]);
+  i[1] = _mm_add_epi16(in[1], in[14]);
+  i[2] = _mm_add_epi16(in[2], in[13]);
+  i[3] = _mm_add_epi16(in[3], in[12]);
+  i[4] = _mm_add_epi16(in[4], in[11]);
+  i[5] = _mm_add_epi16(in[5], in[10]);
+  i[6] = _mm_add_epi16(in[6], in[9]);
+  i[7] = _mm_add_epi16(in[7], in[8]);
+
+  s[0] = _mm_sub_epi16(in[7], in[8]);
+  s[1] = _mm_sub_epi16(in[6], in[9]);
+  s[2] = _mm_sub_epi16(in[5], in[10]);
+  s[3] = _mm_sub_epi16(in[4], in[11]);
+  s[4] = _mm_sub_epi16(in[3], in[12]);
+  s[5] = _mm_sub_epi16(in[2], in[13]);
+  s[6] = _mm_sub_epi16(in[1], in[14]);
+  s[7] = _mm_sub_epi16(in[0], in[15]);
+
+  p[0] = _mm_add_epi16(i[0], i[7]);
+  p[1] = _mm_add_epi16(i[1], i[6]);
+  p[2] = _mm_add_epi16(i[2], i[5]);
+  p[3] = _mm_add_epi16(i[3], i[4]);
+  p[4] = _mm_sub_epi16(i[3], i[4]);
+  p[5] = _mm_sub_epi16(i[2], i[5]);
+  p[6] = _mm_sub_epi16(i[1], i[6]);
+  p[7] = _mm_sub_epi16(i[0], i[7]);
+
+  u[0] = _mm_add_epi16(p[0], p[3]);
+  u[1] = _mm_add_epi16(p[1], p[2]);
+  u[2] = _mm_sub_epi16(p[1], p[2]);
+  u[3] = _mm_sub_epi16(p[0], p[3]);
+
+  v[0] = _mm_unpacklo_epi16(u[0], u[1]);
+  v[1] = _mm_unpackhi_epi16(u[0], u[1]);
+  v[2] = _mm_unpacklo_epi16(u[2], u[3]);
+  v[3] = _mm_unpackhi_epi16(u[2], u[3]);
+
+  u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);
+  u[1] = _mm_madd_epi16(v[1], k__cospi_p16_p16);
+  u[2] = _mm_madd_epi16(v[0], k__cospi_p16_m16);
+  u[3] = _mm_madd_epi16(v[1], k__cospi_p16_m16);
+  u[4] = _mm_madd_epi16(v[2], k__cospi_p24_p08);
+  u[5] = _mm_madd_epi16(v[3], k__cospi_p24_p08);
+  u[6] = _mm_madd_epi16(v[2], k__cospi_m08_p24);
+  u[7] = _mm_madd_epi16(v[3], k__cospi_m08_p24);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[1]);
+  in[4] = _mm_packs_epi32(u[4], u[5]);
+  in[8] = _mm_packs_epi32(u[2], u[3]);
+  in[12] = _mm_packs_epi32(u[6], u[7]);
+
+  u[0] = _mm_unpacklo_epi16(p[5], p[6]);
+  u[1] = _mm_unpackhi_epi16(p[5], p[6]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+
+  u[0] = _mm_packs_epi32(v[0], v[1]);
+  u[1] = _mm_packs_epi32(v[2], v[3]);
+
+  t[0] = _mm_add_epi16(p[4], u[0]);
+  t[1] = _mm_sub_epi16(p[4], u[0]);
+  t[2] = _mm_sub_epi16(p[7], u[1]);
+  t[3] = _mm_add_epi16(p[7], u[1]);
+
+  u[0] = _mm_unpacklo_epi16(t[0], t[3]);
+  u[1] = _mm_unpackhi_epi16(t[0], t[3]);
+  u[2] = _mm_unpacklo_epi16(t[1], t[2]);
+  u[3] = _mm_unpackhi_epi16(t[1], t[2]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p28_p04);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p28_p04);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_p12_p20);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_p12_p20);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m20_p12);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m20_p12);
+  v[6] = _mm_madd_epi16(u[0], k__cospi_m04_p28);
+  v[7] = _mm_madd_epi16(u[1], k__cospi_m04_p28);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  in[2] = _mm_packs_epi32(v[0], v[1]);
+  in[6] = _mm_packs_epi32(v[4], v[5]);
+  in[10] = _mm_packs_epi32(v[2], v[3]);
+  in[14] = _mm_packs_epi32(v[6], v[7]);
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[2], s[5]);
+  u[1] = _mm_unpackhi_epi16(s[2], s[5]);
+  u[2] = _mm_unpacklo_epi16(s[3], s[4]);
+  u[3] = _mm_unpackhi_epi16(s[3], s[4]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+  v[6] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[7] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[2] = _mm_packs_epi32(v[0], v[1]);
+  t[3] = _mm_packs_epi32(v[2], v[3]);
+  t[4] = _mm_packs_epi32(v[4], v[5]);
+  t[5] = _mm_packs_epi32(v[6], v[7]);
+
+  // stage 3
+  p[0] = _mm_add_epi16(s[0], t[3]);
+  p[1] = _mm_add_epi16(s[1], t[2]);
+  p[2] = _mm_sub_epi16(s[1], t[2]);
+  p[3] = _mm_sub_epi16(s[0], t[3]);
+  p[4] = _mm_sub_epi16(s[7], t[4]);
+  p[5] = _mm_sub_epi16(s[6], t[5]);
+  p[6] = _mm_add_epi16(s[6], t[5]);
+  p[7] = _mm_add_epi16(s[7], t[4]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(p[1], p[6]);
+  u[1] = _mm_unpackhi_epi16(p[1], p[6]);
+  u[2] = _mm_unpacklo_epi16(p[2], p[5]);
+  u[3] = _mm_unpackhi_epi16(p[2], p[5]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m08_p24);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m08_p24);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_m24_m08);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_m24_m08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m08_p24);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m08_p24);
+  v[6] = _mm_madd_epi16(u[0], k__cospi_p24_p08);
+  v[7] = _mm_madd_epi16(u[1], k__cospi_p24_p08);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[1] = _mm_packs_epi32(v[0], v[1]);
+  t[2] = _mm_packs_epi32(v[2], v[3]);
+  t[5] = _mm_packs_epi32(v[4], v[5]);
+  t[6] = _mm_packs_epi32(v[6], v[7]);
+
+  // stage 5
+  s[0] = _mm_add_epi16(p[0], t[1]);
+  s[1] = _mm_sub_epi16(p[0], t[1]);
+  s[2] = _mm_sub_epi16(p[3], t[2]);
+  s[3] = _mm_add_epi16(p[3], t[2]);
+  s[4] = _mm_add_epi16(p[4], t[5]);
+  s[5] = _mm_sub_epi16(p[4], t[5]);
+  s[6] = _mm_sub_epi16(p[7], t[6]);
+  s[7] = _mm_add_epi16(p[7], t[6]);
+
+  // stage 6
+  u[0] = _mm_unpacklo_epi16(s[0], s[7]);
+  u[1] = _mm_unpackhi_epi16(s[0], s[7]);
+  u[2] = _mm_unpacklo_epi16(s[1], s[6]);
+  u[3] = _mm_unpackhi_epi16(s[1], s[6]);
+  u[4] = _mm_unpacklo_epi16(s[2], s[5]);
+  u[5] = _mm_unpackhi_epi16(s[2], s[5]);
+  u[6] = _mm_unpacklo_epi16(s[3], s[4]);
+  u[7] = _mm_unpackhi_epi16(s[3], s[4]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p30_p02);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p30_p02);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_p14_p18);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_p14_p18);
+  v[4] = _mm_madd_epi16(u[4], k__cospi_p22_p10);
+  v[5] = _mm_madd_epi16(u[5], k__cospi_p22_p10);
+  v[6] = _mm_madd_epi16(u[6], k__cospi_p06_p26);
+  v[7] = _mm_madd_epi16(u[7], k__cospi_p06_p26);
+  v[8] = _mm_madd_epi16(u[6], k__cospi_m26_p06);
+  v[9] = _mm_madd_epi16(u[7], k__cospi_m26_p06);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_m10_p22);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_m10_p22);
+  v[12] = _mm_madd_epi16(u[2], k__cospi_m18_p14);
+  v[13] = _mm_madd_epi16(u[3], k__cospi_m18_p14);
+  v[14] = _mm_madd_epi16(u[0], k__cospi_m02_p30);
+  v[15] = _mm_madd_epi16(u[1], k__cospi_m02_p30);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  in[1] = _mm_packs_epi32(v[0], v[1]);
+  in[9] = _mm_packs_epi32(v[2], v[3]);
+  in[5] = _mm_packs_epi32(v[4], v[5]);
+  in[13] = _mm_packs_epi32(v[6], v[7]);
+  in[3] = _mm_packs_epi32(v[8], v[9]);
+  in[11] = _mm_packs_epi32(v[10], v[11]);
+  in[7] = _mm_packs_epi32(v[12], v[13]);
+  in[15] = _mm_packs_epi32(v[14], v[15]);
+}
+
+static void fadst16_8col(__m128i *in) {
+  // perform 16x16 1-D ADST for 8 columns
+  __m128i s[16], x[16], u[32], v[32];
+  const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
+  const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kZero = _mm_set1_epi16(0);
+
+  u[0] = _mm_unpacklo_epi16(in[15], in[0]);
+  u[1] = _mm_unpackhi_epi16(in[15], in[0]);
+  u[2] = _mm_unpacklo_epi16(in[13], in[2]);
+  u[3] = _mm_unpackhi_epi16(in[13], in[2]);
+  u[4] = _mm_unpacklo_epi16(in[11], in[4]);
+  u[5] = _mm_unpackhi_epi16(in[11], in[4]);
+  u[6] = _mm_unpacklo_epi16(in[9], in[6]);
+  u[7] = _mm_unpackhi_epi16(in[9], in[6]);
+  u[8] = _mm_unpacklo_epi16(in[7], in[8]);
+  u[9] = _mm_unpackhi_epi16(in[7], in[8]);
+  u[10] = _mm_unpacklo_epi16(in[5], in[10]);
+  u[11] = _mm_unpackhi_epi16(in[5], in[10]);
+  u[12] = _mm_unpacklo_epi16(in[3], in[12]);
+  u[13] = _mm_unpackhi_epi16(in[3], in[12]);
+  u[14] = _mm_unpacklo_epi16(in[1], in[14]);
+  u[15] = _mm_unpackhi_epi16(in[1], in[14]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
+  v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
+  v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
+  v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
+  v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
+  v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
+  v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
+  v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
+  v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
+  v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
+  v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
+  v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
+  v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
+  v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
+  v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
+  v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
+  v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
+
+  u[0] = _mm_add_epi32(v[0], v[16]);
+  u[1] = _mm_add_epi32(v[1], v[17]);
+  u[2] = _mm_add_epi32(v[2], v[18]);
+  u[3] = _mm_add_epi32(v[3], v[19]);
+  u[4] = _mm_add_epi32(v[4], v[20]);
+  u[5] = _mm_add_epi32(v[5], v[21]);
+  u[6] = _mm_add_epi32(v[6], v[22]);
+  u[7] = _mm_add_epi32(v[7], v[23]);
+  u[8] = _mm_add_epi32(v[8], v[24]);
+  u[9] = _mm_add_epi32(v[9], v[25]);
+  u[10] = _mm_add_epi32(v[10], v[26]);
+  u[11] = _mm_add_epi32(v[11], v[27]);
+  u[12] = _mm_add_epi32(v[12], v[28]);
+  u[13] = _mm_add_epi32(v[13], v[29]);
+  u[14] = _mm_add_epi32(v[14], v[30]);
+  u[15] = _mm_add_epi32(v[15], v[31]);
+  u[16] = _mm_sub_epi32(v[0], v[16]);
+  u[17] = _mm_sub_epi32(v[1], v[17]);
+  u[18] = _mm_sub_epi32(v[2], v[18]);
+  u[19] = _mm_sub_epi32(v[3], v[19]);
+  u[20] = _mm_sub_epi32(v[4], v[20]);
+  u[21] = _mm_sub_epi32(v[5], v[21]);
+  u[22] = _mm_sub_epi32(v[6], v[22]);
+  u[23] = _mm_sub_epi32(v[7], v[23]);
+  u[24] = _mm_sub_epi32(v[8], v[24]);
+  u[25] = _mm_sub_epi32(v[9], v[25]);
+  u[26] = _mm_sub_epi32(v[10], v[26]);
+  u[27] = _mm_sub_epi32(v[11], v[27]);
+  u[28] = _mm_sub_epi32(v[12], v[28]);
+  u[29] = _mm_sub_epi32(v[13], v[29]);
+  u[30] = _mm_sub_epi32(v[14], v[30]);
+  u[31] = _mm_sub_epi32(v[15], v[31]);
+
+  v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
+  v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
+  v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
+  v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
+  v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
+  v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
+  v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
+  v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
+  v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
+  v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
+  v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
+  v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
+  v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
+  v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
+  v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
+  v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
+
+  u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
+  u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
+  u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
+  u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
+  u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
+  u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
+  u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
+  u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
+  u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
+  u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
+  u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
+  u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
+  u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
+  u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
+  u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
+  u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
+
+  v[0] = _mm_add_epi32(u[0], u[8]);
+  v[1] = _mm_add_epi32(u[1], u[9]);
+  v[2] = _mm_add_epi32(u[2], u[10]);
+  v[3] = _mm_add_epi32(u[3], u[11]);
+  v[4] = _mm_add_epi32(u[4], u[12]);
+  v[5] = _mm_add_epi32(u[5], u[13]);
+  v[6] = _mm_add_epi32(u[6], u[14]);
+  v[7] = _mm_add_epi32(u[7], u[15]);
+
+  v[16] = _mm_add_epi32(v[0], v[4]);
+  v[17] = _mm_add_epi32(v[1], v[5]);
+  v[18] = _mm_add_epi32(v[2], v[6]);
+  v[19] = _mm_add_epi32(v[3], v[7]);
+  v[20] = _mm_sub_epi32(v[0], v[4]);
+  v[21] = _mm_sub_epi32(v[1], v[5]);
+  v[22] = _mm_sub_epi32(v[2], v[6]);
+  v[23] = _mm_sub_epi32(v[3], v[7]);
+  v[16] = _mm_add_epi32(v[16], k__DCT_CONST_ROUNDING);
+  v[17] = _mm_add_epi32(v[17], k__DCT_CONST_ROUNDING);
+  v[18] = _mm_add_epi32(v[18], k__DCT_CONST_ROUNDING);
+  v[19] = _mm_add_epi32(v[19], k__DCT_CONST_ROUNDING);
+  v[20] = _mm_add_epi32(v[20], k__DCT_CONST_ROUNDING);
+  v[21] = _mm_add_epi32(v[21], k__DCT_CONST_ROUNDING);
+  v[22] = _mm_add_epi32(v[22], k__DCT_CONST_ROUNDING);
+  v[23] = _mm_add_epi32(v[23], k__DCT_CONST_ROUNDING);
+  v[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
+  v[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
+  v[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
+  v[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
+  v[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
+  v[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
+  v[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
+  v[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
+  s[0] = _mm_packs_epi32(v[16], v[17]);
+  s[1] = _mm_packs_epi32(v[18], v[19]);
+  s[2] = _mm_packs_epi32(v[20], v[21]);
+  s[3] = _mm_packs_epi32(v[22], v[23]);
+
+  v[8] = _mm_sub_epi32(u[0], u[8]);
+  v[9] = _mm_sub_epi32(u[1], u[9]);
+  v[10] = _mm_sub_epi32(u[2], u[10]);
+  v[11] = _mm_sub_epi32(u[3], u[11]);
+  v[12] = _mm_sub_epi32(u[4], u[12]);
+  v[13] = _mm_sub_epi32(u[5], u[13]);
+  v[14] = _mm_sub_epi32(u[6], u[14]);
+  v[15] = _mm_sub_epi32(u[7], u[15]);
+
+  v[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+  s[4] = _mm_packs_epi32(v[8], v[9]);
+  s[5] = _mm_packs_epi32(v[10], v[11]);
+  s[6] = _mm_packs_epi32(v[12], v[13]);
+  s[7] = _mm_packs_epi32(v[14], v[15]);
+  //
+
+  s[8] = _mm_packs_epi32(u[16], u[17]);
+  s[9] = _mm_packs_epi32(u[18], u[19]);
+  s[10] = _mm_packs_epi32(u[20], u[21]);
+  s[11] = _mm_packs_epi32(u[22], u[23]);
+  s[12] = _mm_packs_epi32(u[24], u[25]);
+  s[13] = _mm_packs_epi32(u[26], u[27]);
+  s[14] = _mm_packs_epi32(u[28], u[29]);
+  s[15] = _mm_packs_epi32(u[30], u[31]);
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[9]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[9]);
+  u[2] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[3] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[4] = _mm_unpacklo_epi16(s[12], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[12], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], v[8]);
+  u[1] = _mm_add_epi32(v[1], v[9]);
+  u[2] = _mm_add_epi32(v[2], v[10]);
+  u[3] = _mm_add_epi32(v[3], v[11]);
+  u[4] = _mm_add_epi32(v[4], v[12]);
+  u[5] = _mm_add_epi32(v[5], v[13]);
+  u[6] = _mm_add_epi32(v[6], v[14]);
+  u[7] = _mm_add_epi32(v[7], v[15]);
+  u[8] = _mm_sub_epi32(v[0], v[8]);
+  u[9] = _mm_sub_epi32(v[1], v[9]);
+  u[10] = _mm_sub_epi32(v[2], v[10]);
+  u[11] = _mm_sub_epi32(v[3], v[11]);
+  u[12] = _mm_sub_epi32(v[4], v[12]);
+  u[13] = _mm_sub_epi32(v[5], v[13]);
+  u[14] = _mm_sub_epi32(v[6], v[14]);
+  u[15] = _mm_sub_epi32(v[7], v[15]);
+
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+  v[8] = _mm_add_epi32(u[0], u[4]);
+  v[9] = _mm_add_epi32(u[1], u[5]);
+  v[10] = _mm_add_epi32(u[2], u[6]);
+  v[11] = _mm_add_epi32(u[3], u[7]);
+  v[12] = _mm_sub_epi32(u[0], u[4]);
+  v[13] = _mm_sub_epi32(u[1], u[5]);
+  v[14] = _mm_sub_epi32(u[2], u[6]);
+  v[15] = _mm_sub_epi32(u[3], u[7]);
+
+  v[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+  s[8] = _mm_packs_epi32(v[8], v[9]);
+  s[9] = _mm_packs_epi32(v[10], v[11]);
+  s[10] = _mm_packs_epi32(v[12], v[13]);
+  s[11] = _mm_packs_epi32(v[14], v[15]);
+
+  x[12] = _mm_packs_epi32(u[8], u[9]);
+  x[13] = _mm_packs_epi32(u[10], u[11]);
+  x[14] = _mm_packs_epi32(u[12], u[13]);
+  x[15] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  u[0] = _mm_unpacklo_epi16(s[4], s[5]);
+  u[1] = _mm_unpackhi_epi16(s[4], s[5]);
+  u[2] = _mm_unpacklo_epi16(s[6], s[7]);
+  u[3] = _mm_unpackhi_epi16(s[6], s[7]);
+  u[4] = _mm_unpacklo_epi16(x[12], x[13]);
+  u[5] = _mm_unpackhi_epi16(x[12], x[13]);
+  u[6] = _mm_unpacklo_epi16(x[14], x[15]);
+  u[7] = _mm_unpackhi_epi16(x[14], x[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], v[4]);
+  u[1] = _mm_add_epi32(v[1], v[5]);
+  u[2] = _mm_add_epi32(v[2], v[6]);
+  u[3] = _mm_add_epi32(v[3], v[7]);
+  u[4] = _mm_sub_epi32(v[0], v[4]);
+  u[5] = _mm_sub_epi32(v[1], v[5]);
+  u[6] = _mm_sub_epi32(v[2], v[6]);
+  u[7] = _mm_sub_epi32(v[3], v[7]);
+  u[8] = _mm_add_epi32(v[8], v[12]);
+  u[9] = _mm_add_epi32(v[9], v[13]);
+  u[10] = _mm_add_epi32(v[10], v[14]);
+  u[11] = _mm_add_epi32(v[11], v[15]);
+  u[12] = _mm_sub_epi32(v[8], v[12]);
+  u[13] = _mm_sub_epi32(v[9], v[13]);
+  u[14] = _mm_sub_epi32(v[10], v[14]);
+  u[15] = _mm_sub_epi32(v[11], v[15]);
+
+  u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[4] = _mm_packs_epi32(v[0], v[1]);
+  s[5] = _mm_packs_epi32(v[2], v[3]);
+  s[6] = _mm_packs_epi32(v[4], v[5]);
+  s[7] = _mm_packs_epi32(v[6], v[7]);
+
+  s[12] = _mm_packs_epi32(v[8], v[9]);
+  s[13] = _mm_packs_epi32(v[10], v[11]);
+  s[14] = _mm_packs_epi32(v[12], v[13]);
+  s[15] = _mm_packs_epi32(v[14], v[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(s[2], s[3]);
+  u[1] = _mm_unpackhi_epi16(s[2], s[3]);
+  u[2] = _mm_unpacklo_epi16(s[6], s[7]);
+  u[3] = _mm_unpackhi_epi16(s[6], s[7]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  in[0] = s[0];
+  in[1] = _mm_sub_epi16(kZero, s[8]);
+  in[2] = s[12];
+  in[3] = _mm_sub_epi16(kZero, s[4]);
+  in[4] = _mm_packs_epi32(v[4], v[5]);
+  in[5] = _mm_packs_epi32(v[12], v[13]);
+  in[6] = _mm_packs_epi32(v[8], v[9]);
+  in[7] = _mm_packs_epi32(v[0], v[1]);
+  in[8] = _mm_packs_epi32(v[2], v[3]);
+  in[9] = _mm_packs_epi32(v[10], v[11]);
+  in[10] = _mm_packs_epi32(v[14], v[15]);
+  in[11] = _mm_packs_epi32(v[6], v[7]);
+  in[12] = s[5];
+  in[13] = _mm_sub_epi16(kZero, s[13]);
+  in[14] = s[9];
+  in[15] = _mm_sub_epi16(kZero, s[1]);
+}
+
+static void fdct16_sse2(__m128i *in0, __m128i *in1) {
+  fdct16_8col(in0);
+  fdct16_8col(in1);
+  array_transpose_16x16(in0, in1);
+}
+
+static void fadst16_sse2(__m128i *in0, __m128i *in1) {
+  fadst16_8col(in0);
+  fadst16_8col(in1);
+  array_transpose_16x16(in0, in1);
+}
+
+#if CONFIG_EXT_TX
+static void fidtx16_sse2(__m128i *in0, __m128i *in1) {
+  idtx16_8col(in0);
+  idtx16_8col(in1);
+  array_transpose_16x16(in0, in1);
+}
+#endif  // CONFIG_EXT_TX
+
+void av1_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride,
+                       int tx_type) {
+  __m128i in0[16], in1[16];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      load_buffer_16x16(input, in0, in1, stride, 0, 0);
+      fdct16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fdct16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case ADST_DCT:
+      load_buffer_16x16(input, in0, in1, stride, 0, 0);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fdct16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case DCT_ADST:
+      load_buffer_16x16(input, in0, in1, stride, 0, 0);
+      fdct16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case ADST_ADST:
+      load_buffer_16x16(input, in0, in1, stride, 0, 0);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_16x16(input, in0, in1, stride, 1, 0);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fdct16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_16x16(input, in0, in1, stride, 0, 1);
+      fdct16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_16x16(input, in0, in1, stride, 1, 1);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_16x16(input, in0, in1, stride, 0, 1);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_16x16(input, in0, in1, stride, 1, 0);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case IDTX:
+      load_buffer_16x16(input, in0, in1, stride, 0, 0);
+      fidtx16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fidtx16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case V_DCT:
+      load_buffer_16x16(input, in0, in1, stride, 0, 0);
+      fdct16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fidtx16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case H_DCT:
+      load_buffer_16x16(input, in0, in1, stride, 0, 0);
+      fidtx16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fdct16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case V_ADST:
+      load_buffer_16x16(input, in0, in1, stride, 0, 0);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fidtx16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case H_ADST:
+      load_buffer_16x16(input, in0, in1, stride, 0, 0);
+      fidtx16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case V_FLIPADST:
+      load_buffer_16x16(input, in0, in1, stride, 1, 0);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fidtx16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case H_FLIPADST:
+      load_buffer_16x16(input, in0, in1, stride, 0, 1);
+      fidtx16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+}
+
+static INLINE void prepare_4x8_row_first(__m128i *in) {
+  in[0] = _mm_unpacklo_epi64(in[0], in[2]);
+  in[1] = _mm_unpacklo_epi64(in[1], in[3]);
+  transpose_4x4(in);
+  in[4] = _mm_unpacklo_epi64(in[4], in[6]);
+  in[5] = _mm_unpacklo_epi64(in[5], in[7]);
+  transpose_4x4(in + 4);
+}
+
+// Load input into the left-hand half of in (ie, into lanes 0..3 of
+// each element of in). The right hand half (lanes 4..7) should be
+// treated as being filled with "don't care" values.
+static INLINE void load_buffer_4x8(const int16_t *input, __m128i *in,
+                                   int stride, int flipud, int fliplr) {
+  const int shift = 2;
+  if (!flipud) {
+    in[0] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+    in[1] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+    in[2] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
+    in[3] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
+    in[4] = _mm_loadl_epi64((const __m128i *)(input + 4 * stride));
+    in[5] = _mm_loadl_epi64((const __m128i *)(input + 5 * stride));
+    in[6] = _mm_loadl_epi64((const __m128i *)(input + 6 * stride));
+    in[7] = _mm_loadl_epi64((const __m128i *)(input + 7 * stride));
+  } else {
+    in[0] = _mm_loadl_epi64((const __m128i *)(input + 7 * stride));
+    in[1] = _mm_loadl_epi64((const __m128i *)(input + 6 * stride));
+    in[2] = _mm_loadl_epi64((const __m128i *)(input + 5 * stride));
+    in[3] = _mm_loadl_epi64((const __m128i *)(input + 4 * stride));
+    in[4] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
+    in[5] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
+    in[6] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+    in[7] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+  }
+
+  if (fliplr) {
+    in[0] = _mm_shufflelo_epi16(in[0], 0x1b);
+    in[1] = _mm_shufflelo_epi16(in[1], 0x1b);
+    in[2] = _mm_shufflelo_epi16(in[2], 0x1b);
+    in[3] = _mm_shufflelo_epi16(in[3], 0x1b);
+    in[4] = _mm_shufflelo_epi16(in[4], 0x1b);
+    in[5] = _mm_shufflelo_epi16(in[5], 0x1b);
+    in[6] = _mm_shufflelo_epi16(in[6], 0x1b);
+    in[7] = _mm_shufflelo_epi16(in[7], 0x1b);
+  }
+
+  in[0] = _mm_slli_epi16(in[0], shift);
+  in[1] = _mm_slli_epi16(in[1], shift);
+  in[2] = _mm_slli_epi16(in[2], shift);
+  in[3] = _mm_slli_epi16(in[3], shift);
+  in[4] = _mm_slli_epi16(in[4], shift);
+  in[5] = _mm_slli_epi16(in[5], shift);
+  in[6] = _mm_slli_epi16(in[6], shift);
+  in[7] = _mm_slli_epi16(in[7], shift);
+
+  scale_sqrt2_8x4(in);
+  scale_sqrt2_8x4(in + 4);
+  prepare_4x8_row_first(in);
+}
+
+static INLINE void write_buffer_4x8(tran_low_t *output, __m128i *res) {
+  __m128i in01, in23, in45, in67, sign01, sign23, sign45, sign67;
+  const int shift = 1;
+
+  // revert the 8x8 txfm's transpose
+  array_transpose_8x8(res, res);
+
+  in01 = _mm_unpacklo_epi64(res[0], res[1]);
+  in23 = _mm_unpacklo_epi64(res[2], res[3]);
+  in45 = _mm_unpacklo_epi64(res[4], res[5]);
+  in67 = _mm_unpacklo_epi64(res[6], res[7]);
+
+  sign01 = _mm_srai_epi16(in01, 15);
+  sign23 = _mm_srai_epi16(in23, 15);
+  sign45 = _mm_srai_epi16(in45, 15);
+  sign67 = _mm_srai_epi16(in67, 15);
+
+  in01 = _mm_sub_epi16(in01, sign01);
+  in23 = _mm_sub_epi16(in23, sign23);
+  in45 = _mm_sub_epi16(in45, sign45);
+  in67 = _mm_sub_epi16(in67, sign67);
+
+  in01 = _mm_srai_epi16(in01, shift);
+  in23 = _mm_srai_epi16(in23, shift);
+  in45 = _mm_srai_epi16(in45, shift);
+  in67 = _mm_srai_epi16(in67, shift);
+
+  store_output(&in01, (output + 0 * 8));
+  store_output(&in23, (output + 1 * 8));
+  store_output(&in45, (output + 2 * 8));
+  store_output(&in67, (output + 3 * 8));
+}
+
+void av1_fht4x8_sse2(const int16_t *input, tran_low_t *output, int stride,
+                     int tx_type) {
+  __m128i in[8];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      load_buffer_4x8(input, in, stride, 0, 0);
+      fdct4_sse2(in);
+      fdct4_sse2(in + 4);
+      fdct8_sse2(in);
+      break;
+    case ADST_DCT:
+      load_buffer_4x8(input, in, stride, 0, 0);
+      fdct4_sse2(in);
+      fdct4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case DCT_ADST:
+      load_buffer_4x8(input, in, stride, 0, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fdct8_sse2(in);
+      break;
+    case ADST_ADST:
+      load_buffer_4x8(input, in, stride, 0, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_4x8(input, in, stride, 1, 0);
+      fdct4_sse2(in);
+      fdct4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_4x8(input, in, stride, 0, 1);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fdct8_sse2(in);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_4x8(input, in, stride, 1, 1);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_4x8(input, in, stride, 0, 1);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_4x8(input, in, stride, 1, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case IDTX:
+      load_buffer_4x8(input, in, stride, 0, 0);
+      fidtx4_sse2(in);
+      fidtx4_sse2(in + 4);
+      fidtx8_sse2(in);
+      break;
+    case V_DCT:
+      load_buffer_4x8(input, in, stride, 0, 0);
+      fidtx4_sse2(in);
+      fidtx4_sse2(in + 4);
+      fdct8_sse2(in);
+      break;
+    case H_DCT:
+      load_buffer_4x8(input, in, stride, 0, 0);
+      fdct4_sse2(in);
+      fdct4_sse2(in + 4);
+      fidtx8_sse2(in);
+      break;
+    case V_ADST:
+      load_buffer_4x8(input, in, stride, 0, 0);
+      fidtx4_sse2(in);
+      fidtx4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case H_ADST:
+      load_buffer_4x8(input, in, stride, 0, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fidtx8_sse2(in);
+      break;
+    case V_FLIPADST:
+      load_buffer_4x8(input, in, stride, 1, 0);
+      fidtx4_sse2(in);
+      fidtx4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case H_FLIPADST:
+      load_buffer_4x8(input, in, stride, 0, 1);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fidtx8_sse2(in);
+      break;
+#endif
+    default: assert(0); break;
+  }
+  write_buffer_4x8(output, in);
+}
+
+// Load input into the left-hand half of in (ie, into lanes 0..3 of
+// each element of in). The right hand half (lanes 4..7) should be
+// treated as being filled with "don't care" values.
+// The input is split horizontally into two 4x4
+// chunks 'l' and 'r'. Then 'l' is stored in the top-left 4x4
+// block of 'in' and 'r' is stored in the bottom-left block.
+// This is to allow us to reuse 4x4 transforms.
+static INLINE void load_buffer_8x4(const int16_t *input, __m128i *in,
+                                   int stride, int flipud, int fliplr) {
+  const int shift = 2;
+  if (!flipud) {
+    in[0] = _mm_loadu_si128((const __m128i *)(input + 0 * stride));
+    in[1] = _mm_loadu_si128((const __m128i *)(input + 1 * stride));
+    in[2] = _mm_loadu_si128((const __m128i *)(input + 2 * stride));
+    in[3] = _mm_loadu_si128((const __m128i *)(input + 3 * stride));
+  } else {
+    in[0] = _mm_loadu_si128((const __m128i *)(input + 3 * stride));
+    in[1] = _mm_loadu_si128((const __m128i *)(input + 2 * stride));
+    in[2] = _mm_loadu_si128((const __m128i *)(input + 1 * stride));
+    in[3] = _mm_loadu_si128((const __m128i *)(input + 0 * stride));
+  }
+
+  if (fliplr) {
+    in[0] = mm_reverse_epi16(in[0]);
+    in[1] = mm_reverse_epi16(in[1]);
+    in[2] = mm_reverse_epi16(in[2]);
+    in[3] = mm_reverse_epi16(in[3]);
+  }
+
+  in[0] = _mm_slli_epi16(in[0], shift);
+  in[1] = _mm_slli_epi16(in[1], shift);
+  in[2] = _mm_slli_epi16(in[2], shift);
+  in[3] = _mm_slli_epi16(in[3], shift);
+
+  scale_sqrt2_8x4(in);
+
+  in[4] = _mm_shuffle_epi32(in[0], 0xe);
+  in[5] = _mm_shuffle_epi32(in[1], 0xe);
+  in[6] = _mm_shuffle_epi32(in[2], 0xe);
+  in[7] = _mm_shuffle_epi32(in[3], 0xe);
+}
+
+static INLINE void write_buffer_8x4(tran_low_t *output, __m128i *res) {
+  __m128i out0, out1, out2, out3, sign0, sign1, sign2, sign3;
+  const int shift = 1;
+  sign0 = _mm_srai_epi16(res[0], 15);
+  sign1 = _mm_srai_epi16(res[1], 15);
+  sign2 = _mm_srai_epi16(res[2], 15);
+  sign3 = _mm_srai_epi16(res[3], 15);
+
+  out0 = _mm_sub_epi16(res[0], sign0);
+  out1 = _mm_sub_epi16(res[1], sign1);
+  out2 = _mm_sub_epi16(res[2], sign2);
+  out3 = _mm_sub_epi16(res[3], sign3);
+
+  out0 = _mm_srai_epi16(out0, shift);
+  out1 = _mm_srai_epi16(out1, shift);
+  out2 = _mm_srai_epi16(out2, shift);
+  out3 = _mm_srai_epi16(out3, shift);
+
+  store_output(&out0, (output + 0 * 8));
+  store_output(&out1, (output + 1 * 8));
+  store_output(&out2, (output + 2 * 8));
+  store_output(&out3, (output + 3 * 8));
+}
+
+void av1_fht8x4_sse2(const int16_t *input, tran_low_t *output, int stride,
+                     int tx_type) {
+  __m128i in[8];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      load_buffer_8x4(input, in, stride, 0, 0);
+      fdct4_sse2(in);
+      fdct4_sse2(in + 4);
+      fdct8_sse2(in);
+      break;
+    case ADST_DCT:
+      load_buffer_8x4(input, in, stride, 0, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fdct8_sse2(in);
+      break;
+    case DCT_ADST:
+      load_buffer_8x4(input, in, stride, 0, 0);
+      fdct4_sse2(in);
+      fdct4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case ADST_ADST:
+      load_buffer_8x4(input, in, stride, 0, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_8x4(input, in, stride, 1, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fdct8_sse2(in);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_8x4(input, in, stride, 0, 1);
+      fdct4_sse2(in);
+      fdct4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_8x4(input, in, stride, 1, 1);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_8x4(input, in, stride, 0, 1);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_8x4(input, in, stride, 1, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case IDTX:
+      load_buffer_8x4(input, in, stride, 0, 0);
+      fidtx4_sse2(in);
+      fidtx4_sse2(in + 4);
+      fidtx8_sse2(in);
+      break;
+    case V_DCT:
+      load_buffer_8x4(input, in, stride, 0, 0);
+      fdct4_sse2(in);
+      fdct4_sse2(in + 4);
+      fidtx8_sse2(in);
+      break;
+    case H_DCT:
+      load_buffer_8x4(input, in, stride, 0, 0);
+      fidtx4_sse2(in);
+      fidtx4_sse2(in + 4);
+      fdct8_sse2(in);
+      break;
+    case V_ADST:
+      load_buffer_8x4(input, in, stride, 0, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fidtx8_sse2(in);
+      break;
+    case H_ADST:
+      load_buffer_8x4(input, in, stride, 0, 0);
+      fidtx4_sse2(in);
+      fidtx4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+    case V_FLIPADST:
+      load_buffer_8x4(input, in, stride, 1, 0);
+      fadst4_sse2(in);
+      fadst4_sse2(in + 4);
+      fidtx8_sse2(in);
+      break;
+    case H_FLIPADST:
+      load_buffer_8x4(input, in, stride, 0, 1);
+      fidtx4_sse2(in);
+      fidtx4_sse2(in + 4);
+      fadst8_sse2(in);
+      break;
+#endif
+    default: assert(0); break;
+  }
+  write_buffer_8x4(output, in);
+}
+
+static INLINE void load_buffer_8x16(const int16_t *input, __m128i *in,
+                                    int stride, int flipud, int fliplr) {
+  // Load 2 8x8 blocks
+  const int16_t *t = input;
+  const int16_t *b = input + 8 * stride;
+
+  if (flipud) {
+    const int16_t *const tmp = t;
+    t = b;
+    b = tmp;
+  }
+
+  load_buffer_8x8(t, in, stride, flipud, fliplr);
+  scale_sqrt2_8x8(in);
+  load_buffer_8x8(b, in + 8, stride, flipud, fliplr);
+  scale_sqrt2_8x8(in + 8);
+}
+
+static INLINE void round_power_of_two_signed(__m128i *x, int n) {
+  const __m128i rounding = _mm_set1_epi16((1 << n) >> 1);
+  const __m128i sign = _mm_srai_epi16(*x, 15);
+  const __m128i res = _mm_add_epi16(_mm_add_epi16(*x, rounding), sign);
+  *x = _mm_srai_epi16(res, n);
+}
+
+static void row_8x16_rounding(__m128i *in, int bits) {
+  int i;
+  for (i = 0; i < 16; i++) {
+    round_power_of_two_signed(&in[i], bits);
+  }
+}
+
+void av1_fht8x16_sse2(const int16_t *input, tran_low_t *output, int stride,
+                      int tx_type) {
+  __m128i in[16];
+
+  __m128i *const t = in;      // Alias to top 8x8 sub block
+  __m128i *const b = in + 8;  // Alias to bottom 8x8 sub block
+
+  switch (tx_type) {
+    case DCT_DCT:
+      load_buffer_8x16(input, in, stride, 0, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fdct8_sse2(t);
+      fdct8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fdct16_8col(in);
+      break;
+    case ADST_DCT:
+      load_buffer_8x16(input, in, stride, 0, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fdct8_sse2(t);
+      fdct8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case DCT_ADST:
+      load_buffer_8x16(input, in, stride, 0, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fadst8_sse2(t);
+      fadst8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fdct16_8col(in);
+      break;
+    case ADST_ADST:
+      load_buffer_8x16(input, in, stride, 0, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fadst8_sse2(t);
+      fadst8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_8x16(input, in, stride, 1, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fdct8_sse2(t);
+      fdct8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_8x16(input, in, stride, 0, 1);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fadst8_sse2(t);
+      fadst8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fdct16_8col(in);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_8x16(input, in, stride, 1, 1);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fadst8_sse2(t);
+      fadst8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_8x16(input, in, stride, 0, 1);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fadst8_sse2(t);
+      fadst8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_8x16(input, in, stride, 1, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fadst8_sse2(t);
+      fadst8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case IDTX:
+      load_buffer_8x16(input, in, stride, 0, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fidtx8_sse2(t);
+      fidtx8_sse2(b);
+      row_8x16_rounding(in, 2);
+      idtx16_8col(in);
+      break;
+    case V_DCT:
+      load_buffer_8x16(input, in, stride, 0, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fidtx8_sse2(t);
+      fidtx8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fdct16_8col(in);
+      break;
+    case H_DCT:
+      load_buffer_8x16(input, in, stride, 0, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fdct8_sse2(t);
+      fdct8_sse2(b);
+      row_8x16_rounding(in, 2);
+      idtx16_8col(in);
+      break;
+    case V_ADST:
+      load_buffer_8x16(input, in, stride, 0, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fidtx8_sse2(t);
+      fidtx8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case H_ADST:
+      load_buffer_8x16(input, in, stride, 0, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fadst8_sse2(t);
+      fadst8_sse2(b);
+      row_8x16_rounding(in, 2);
+      idtx16_8col(in);
+      break;
+    case V_FLIPADST:
+      load_buffer_8x16(input, in, stride, 1, 0);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fidtx8_sse2(t);
+      fidtx8_sse2(b);
+      row_8x16_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case H_FLIPADST:
+      load_buffer_8x16(input, in, stride, 0, 1);
+      array_transpose_8x8(t, t);
+      array_transpose_8x8(b, b);
+      fadst8_sse2(t);
+      fadst8_sse2(b);
+      row_8x16_rounding(in, 2);
+      idtx16_8col(in);
+      break;
+#endif
+    default: assert(0); break;
+  }
+  write_buffer_8x8(output, t, 8);
+  write_buffer_8x8(output + 64, b, 8);
+}
+
+static INLINE void load_buffer_16x8(const int16_t *input, __m128i *in,
+                                    int stride, int flipud, int fliplr) {
+  // Load 2 8x8 blocks
+  const int16_t *l = input;
+  const int16_t *r = input + 8;
+
+  if (fliplr) {
+    const int16_t *const tmp = l;
+    l = r;
+    r = tmp;
+  }
+
+  // load first 8 columns
+  load_buffer_8x8(l, in, stride, flipud, fliplr);
+  scale_sqrt2_8x8(in);
+  load_buffer_8x8(r, in + 8, stride, flipud, fliplr);
+  scale_sqrt2_8x8(in + 8);
+}
+
+#define col_16x8_rounding row_8x16_rounding
+
+void av1_fht16x8_sse2(const int16_t *input, tran_low_t *output, int stride,
+                      int tx_type) {
+  __m128i in[16];
+
+  __m128i *const l = in;      // Alias to left 8x8 sub block
+  __m128i *const r = in + 8;  // Alias to right 8x8 sub block, which we store
+                              // in the second half of the array
+
+  switch (tx_type) {
+    case DCT_DCT:
+      load_buffer_16x8(input, in, stride, 0, 0);
+      fdct8_sse2(l);
+      fdct8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fdct16_8col(in);
+      break;
+    case ADST_DCT:
+      load_buffer_16x8(input, in, stride, 0, 0);
+      fadst8_sse2(l);
+      fadst8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fdct16_8col(in);
+      break;
+    case DCT_ADST:
+      load_buffer_16x8(input, in, stride, 0, 0);
+      fdct8_sse2(l);
+      fdct8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case ADST_ADST:
+      load_buffer_16x8(input, in, stride, 0, 0);
+      fadst8_sse2(l);
+      fadst8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_16x8(input, in, stride, 1, 0);
+      fadst8_sse2(l);
+      fadst8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fdct16_8col(in);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_16x8(input, in, stride, 0, 1);
+      fdct8_sse2(l);
+      fdct8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_16x8(input, in, stride, 1, 1);
+      fadst8_sse2(l);
+      fadst8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_16x8(input, in, stride, 0, 1);
+      fadst8_sse2(l);
+      fadst8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_16x8(input, in, stride, 1, 0);
+      fadst8_sse2(l);
+      fadst8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case IDTX:
+      load_buffer_16x8(input, in, stride, 0, 0);
+      fidtx8_sse2(l);
+      fidtx8_sse2(r);
+      col_16x8_rounding(in, 2);
+      idtx16_8col(in);
+      break;
+    case V_DCT:
+      load_buffer_16x8(input, in, stride, 0, 0);
+      fdct8_sse2(l);
+      fdct8_sse2(r);
+      col_16x8_rounding(in, 2);
+      idtx16_8col(in);
+      break;
+    case H_DCT:
+      load_buffer_16x8(input, in, stride, 0, 0);
+      fidtx8_sse2(l);
+      fidtx8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fdct16_8col(in);
+      break;
+    case V_ADST:
+      load_buffer_16x8(input, in, stride, 0, 0);
+      fadst8_sse2(l);
+      fadst8_sse2(r);
+      col_16x8_rounding(in, 2);
+      idtx16_8col(in);
+      break;
+    case H_ADST:
+      load_buffer_16x8(input, in, stride, 0, 0);
+      fidtx8_sse2(l);
+      fidtx8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+    case V_FLIPADST:
+      load_buffer_16x8(input, in, stride, 1, 0);
+      fadst8_sse2(l);
+      fadst8_sse2(r);
+      col_16x8_rounding(in, 2);
+      idtx16_8col(in);
+      break;
+    case H_FLIPADST:
+      load_buffer_16x8(input, in, stride, 0, 1);
+      fidtx8_sse2(l);
+      fidtx8_sse2(r);
+      col_16x8_rounding(in, 2);
+      fadst16_8col(in);
+      break;
+#endif
+    default: assert(0); break;
+  }
+  array_transpose_8x8(l, l);
+  array_transpose_8x8(r, r);
+  write_buffer_8x8(output, l, 16);
+  write_buffer_8x8(output + 8, r, 16);
+}
+
+// Note: The 16-column 32-element transforms expect their input to be
+// split up into a 2x2 grid of 8x16 blocks
+static INLINE void fdct32_16col(__m128i *tl, __m128i *tr, __m128i *bl,
+                                __m128i *br) {
+  fdct32_8col(tl, bl);
+  fdct32_8col(tr, br);
+  array_transpose_16x16(tl, tr);
+  array_transpose_16x16(bl, br);
+}
+
+#if CONFIG_EXT_TX
+static INLINE void fidtx32_16col(__m128i *tl, __m128i *tr, __m128i *bl,
+                                 __m128i *br) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    tl[i] = _mm_slli_epi16(tl[i], 2);
+    tr[i] = _mm_slli_epi16(tr[i], 2);
+    bl[i] = _mm_slli_epi16(bl[i], 2);
+    br[i] = _mm_slli_epi16(br[i], 2);
+  }
+  array_transpose_16x16(tl, tr);
+  array_transpose_16x16(bl, br);
+}
+#endif
+
+static INLINE void load_buffer_16x32(const int16_t *input, __m128i *intl,
+                                     __m128i *intr, __m128i *inbl,
+                                     __m128i *inbr, int stride, int flipud,
+                                     int fliplr) {
+  int i;
+  if (flipud) {
+    input = input + 31 * stride;
+    stride = -stride;
+  }
+
+  for (i = 0; i < 16; ++i) {
+    intl[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + i * stride + 0)), 2);
+    intr[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + i * stride + 8)), 2);
+    inbl[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + (i + 16) * stride + 0)), 2);
+    inbr[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + (i + 16) * stride + 8)), 2);
+  }
+
+  if (fliplr) {
+    __m128i tmp;
+    for (i = 0; i < 16; ++i) {
+      tmp = intl[i];
+      intl[i] = mm_reverse_epi16(intr[i]);
+      intr[i] = mm_reverse_epi16(tmp);
+      tmp = inbl[i];
+      inbl[i] = mm_reverse_epi16(inbr[i]);
+      inbr[i] = mm_reverse_epi16(tmp);
+    }
+  }
+
+  scale_sqrt2_8x16(intl);
+  scale_sqrt2_8x16(intr);
+  scale_sqrt2_8x16(inbl);
+  scale_sqrt2_8x16(inbr);
+}
+
+static INLINE void write_buffer_16x32(tran_low_t *output, __m128i *restl,
+                                      __m128i *restr, __m128i *resbl,
+                                      __m128i *resbr) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    store_output(&restl[i], output + i * 16 + 0);
+    store_output(&restr[i], output + i * 16 + 8);
+    store_output(&resbl[i], output + (i + 16) * 16 + 0);
+    store_output(&resbr[i], output + (i + 16) * 16 + 8);
+  }
+}
+
+static INLINE void round_signed_8x8(__m128i *in, const int bit) {
+  const __m128i rounding = _mm_set1_epi16((1 << bit) >> 1);
+  __m128i sign0 = _mm_srai_epi16(in[0], 15);
+  __m128i sign1 = _mm_srai_epi16(in[1], 15);
+  __m128i sign2 = _mm_srai_epi16(in[2], 15);
+  __m128i sign3 = _mm_srai_epi16(in[3], 15);
+  __m128i sign4 = _mm_srai_epi16(in[4], 15);
+  __m128i sign5 = _mm_srai_epi16(in[5], 15);
+  __m128i sign6 = _mm_srai_epi16(in[6], 15);
+  __m128i sign7 = _mm_srai_epi16(in[7], 15);
+
+  in[0] = _mm_add_epi16(_mm_add_epi16(in[0], rounding), sign0);
+  in[1] = _mm_add_epi16(_mm_add_epi16(in[1], rounding), sign1);
+  in[2] = _mm_add_epi16(_mm_add_epi16(in[2], rounding), sign2);
+  in[3] = _mm_add_epi16(_mm_add_epi16(in[3], rounding), sign3);
+  in[4] = _mm_add_epi16(_mm_add_epi16(in[4], rounding), sign4);
+  in[5] = _mm_add_epi16(_mm_add_epi16(in[5], rounding), sign5);
+  in[6] = _mm_add_epi16(_mm_add_epi16(in[6], rounding), sign6);
+  in[7] = _mm_add_epi16(_mm_add_epi16(in[7], rounding), sign7);
+
+  in[0] = _mm_srai_epi16(in[0], bit);
+  in[1] = _mm_srai_epi16(in[1], bit);
+  in[2] = _mm_srai_epi16(in[2], bit);
+  in[3] = _mm_srai_epi16(in[3], bit);
+  in[4] = _mm_srai_epi16(in[4], bit);
+  in[5] = _mm_srai_epi16(in[5], bit);
+  in[6] = _mm_srai_epi16(in[6], bit);
+  in[7] = _mm_srai_epi16(in[7], bit);
+}
+
+static INLINE void round_signed_16x16(__m128i *in0, __m128i *in1) {
+  const int bit = 4;
+  round_signed_8x8(in0, bit);
+  round_signed_8x8(in0 + 8, bit);
+  round_signed_8x8(in1, bit);
+  round_signed_8x8(in1 + 8, bit);
+}
+
+// Note:
+//  suffix "t" indicates the transpose operation comes first
+static void fdct16t_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  fdct16_8col(in0);
+  fdct16_8col(in1);
+}
+
+static void fadst16t_sse2(__m128i *in0, __m128i *in1) {
+  array_transpose_16x16(in0, in1);
+  fadst16_8col(in0);
+  fadst16_8col(in1);
+}
+
+static INLINE void fdct32t_16col(__m128i *tl, __m128i *tr, __m128i *bl,
+                                 __m128i *br) {
+  array_transpose_16x16(tl, tr);
+  array_transpose_16x16(bl, br);
+  fdct32_8col(tl, bl);
+  fdct32_8col(tr, br);
+}
+
+typedef enum transpose_indicator_ {
+  transpose,
+  no_transpose,
+} transpose_indicator;
+
+static INLINE void fhalfright32_16col(__m128i *tl, __m128i *tr, __m128i *bl,
+                                      __m128i *br, transpose_indicator t) {
+  __m128i tmpl[16], tmpr[16];
+  int i;
+
+  // Copy the bottom half of the input to temporary storage
+  for (i = 0; i < 16; ++i) {
+    tmpl[i] = bl[i];
+    tmpr[i] = br[i];
+  }
+
+  // Generate the bottom half of the output
+  for (i = 0; i < 16; ++i) {
+    bl[i] = _mm_slli_epi16(tl[i], 2);
+    br[i] = _mm_slli_epi16(tr[i], 2);
+  }
+  array_transpose_16x16(bl, br);
+
+  // Copy the temporary storage back to the top half of the input
+  for (i = 0; i < 16; ++i) {
+    tl[i] = tmpl[i];
+    tr[i] = tmpr[i];
+  }
+
+  // Generate the top half of the output
+  scale_sqrt2_8x16(tl);
+  scale_sqrt2_8x16(tr);
+  if (t == transpose)
+    fdct16t_sse2(tl, tr);
+  else
+    fdct16_sse2(tl, tr);
+}
+
+// Note on data layout, for both this and the 32x16 transforms:
+// So that we can reuse the 16-element transforms easily,
+// we want to split the input into 8x16 blocks.
+// For 16x32, this means the input is a 2x2 grid of such blocks.
+// For 32x16, it means the input is a 4x1 grid.
+void av1_fht16x32_sse2(const int16_t *input, tran_low_t *output, int stride,
+                       int tx_type) {
+  __m128i intl[16], intr[16], inbl[16], inbr[16];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 0);
+      fdct16t_sse2(intl, intr);
+      fdct16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fdct32t_16col(intl, intr, inbl, inbr);
+      break;
+    case ADST_DCT:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 0);
+      fdct16t_sse2(intl, intr);
+      fdct16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fhalfright32_16col(intl, intr, inbl, inbr, transpose);
+      break;
+    case DCT_ADST:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 0);
+      fadst16t_sse2(intl, intr);
+      fadst16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fdct32t_16col(intl, intr, inbl, inbr);
+      break;
+    case ADST_ADST:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 0);
+      fadst16t_sse2(intl, intr);
+      fadst16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fhalfright32_16col(intl, intr, inbl, inbr, transpose);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 1, 0);
+      fdct16t_sse2(intl, intr);
+      fdct16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fhalfright32_16col(intl, intr, inbl, inbr, transpose);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 1);
+      fadst16t_sse2(intl, intr);
+      fadst16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fdct32t_16col(intl, intr, inbl, inbr);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 1, 1);
+      fadst16t_sse2(intl, intr);
+      fadst16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fhalfright32_16col(intl, intr, inbl, inbr, transpose);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 1);
+      fadst16t_sse2(intl, intr);
+      fadst16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fhalfright32_16col(intl, intr, inbl, inbr, transpose);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 1, 0);
+      fadst16t_sse2(intl, intr);
+      fadst16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fhalfright32_16col(intl, intr, inbl, inbr, transpose);
+      break;
+    case IDTX:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 0);
+      fidtx16_sse2(intl, intr);
+      fidtx16_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fidtx32_16col(intl, intr, inbl, inbr);
+      break;
+    case V_DCT:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 0);
+      fidtx16_sse2(intl, intr);
+      fidtx16_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fdct32t_16col(intl, intr, inbl, inbr);
+      break;
+    case H_DCT:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 0);
+      fdct16t_sse2(intl, intr);
+      fdct16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fidtx32_16col(intl, intr, inbl, inbr);
+      break;
+    case V_ADST:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 0);
+      fidtx16_sse2(intl, intr);
+      fidtx16_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fhalfright32_16col(intl, intr, inbl, inbr, transpose);
+      break;
+    case H_ADST:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 0);
+      fadst16t_sse2(intl, intr);
+      fadst16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fidtx32_16col(intl, intr, inbl, inbr);
+      break;
+    case V_FLIPADST:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 1, 0);
+      fidtx16_sse2(intl, intr);
+      fidtx16_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fhalfright32_16col(intl, intr, inbl, inbr, transpose);
+      break;
+    case H_FLIPADST:
+      load_buffer_16x32(input, intl, intr, inbl, inbr, stride, 0, 1);
+      fadst16t_sse2(intl, intr);
+      fadst16t_sse2(inbl, inbr);
+      round_signed_16x16(intl, intr);
+      round_signed_16x16(inbl, inbr);
+      fidtx32_16col(intl, intr, inbl, inbr);
+      break;
+#endif
+    default: assert(0); break;
+  }
+  write_buffer_16x32(output, intl, intr, inbl, inbr);
+}
+
+static INLINE void load_buffer_32x16(const int16_t *input, __m128i *in0,
+                                     __m128i *in1, __m128i *in2, __m128i *in3,
+                                     int stride, int flipud, int fliplr) {
+  int i;
+  if (flipud) {
+    input += 15 * stride;
+    stride = -stride;
+  }
+
+  for (i = 0; i < 16; ++i) {
+    in0[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + i * stride + 0)), 2);
+    in1[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + i * stride + 8)), 2);
+    in2[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + i * stride + 16)), 2);
+    in3[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + i * stride + 24)), 2);
+  }
+
+  if (fliplr) {
+    for (i = 0; i < 16; ++i) {
+      __m128i tmp1 = in0[i];
+      __m128i tmp2 = in1[i];
+      in0[i] = mm_reverse_epi16(in3[i]);
+      in1[i] = mm_reverse_epi16(in2[i]);
+      in2[i] = mm_reverse_epi16(tmp2);
+      in3[i] = mm_reverse_epi16(tmp1);
+    }
+  }
+
+  scale_sqrt2_8x16(in0);
+  scale_sqrt2_8x16(in1);
+  scale_sqrt2_8x16(in2);
+  scale_sqrt2_8x16(in3);
+}
+
+static INLINE void write_buffer_32x16(tran_low_t *output, __m128i *res0,
+                                      __m128i *res1, __m128i *res2,
+                                      __m128i *res3) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    store_output(&res0[i], output + i * 32 + 0);
+    store_output(&res1[i], output + i * 32 + 8);
+    store_output(&res2[i], output + i * 32 + 16);
+    store_output(&res3[i], output + i * 32 + 24);
+  }
+}
+
+void av1_fht32x16_sse2(const int16_t *input, tran_low_t *output, int stride,
+                       int tx_type) {
+  __m128i in0[16], in1[16], in2[16], in3[16];
+
+  load_buffer_32x16(input, in0, in1, in2, in3, stride, 0, 0);
+  switch (tx_type) {
+    case DCT_DCT:
+      fdct16_sse2(in0, in1);
+      fdct16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fdct32_16col(in0, in1, in2, in3);
+      break;
+    case ADST_DCT:
+      fadst16_sse2(in0, in1);
+      fadst16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fdct32_16col(in0, in1, in2, in3);
+      break;
+    case DCT_ADST:
+      fdct16_sse2(in0, in1);
+      fdct16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fhalfright32_16col(in0, in1, in2, in3, no_transpose);
+      break;
+    case ADST_ADST:
+      fadst16_sse2(in0, in1);
+      fadst16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fhalfright32_16col(in0, in1, in2, in3, no_transpose);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 1, 0);
+      fadst16_sse2(in0, in1);
+      fadst16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fdct32_16col(in0, in1, in2, in3);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 0, 1);
+      fdct16_sse2(in0, in1);
+      fdct16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fhalfright32_16col(in0, in1, in2, in3, no_transpose);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 1, 1);
+      fadst16_sse2(in0, in1);
+      fadst16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fhalfright32_16col(in0, in1, in2, in3, no_transpose);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 0, 1);
+      fadst16_sse2(in0, in1);
+      fadst16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fhalfright32_16col(in0, in1, in2, in3, no_transpose);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 1, 0);
+      fadst16_sse2(in0, in1);
+      fadst16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fhalfright32_16col(in0, in1, in2, in3, no_transpose);
+      break;
+    case IDTX:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 0, 0);
+      fidtx16_sse2(in0, in1);
+      fidtx16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fidtx32_16col(in0, in1, in2, in3);
+      break;
+    case V_DCT:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 0, 0);
+      fdct16_sse2(in0, in1);
+      fdct16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fidtx32_16col(in0, in1, in2, in3);
+      break;
+    case H_DCT:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 0, 0);
+      fidtx16_sse2(in0, in1);
+      fidtx16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fdct32_16col(in0, in1, in2, in3);
+      break;
+    case V_ADST:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 0, 0);
+      fadst16_sse2(in0, in1);
+      fadst16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fidtx32_16col(in0, in1, in2, in3);
+      break;
+    case H_ADST:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 0, 0);
+      fidtx16_sse2(in0, in1);
+      fidtx16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fhalfright32_16col(in0, in1, in2, in3, no_transpose);
+      break;
+    case V_FLIPADST:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 1, 0);
+      fadst16_sse2(in0, in1);
+      fadst16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fidtx32_16col(in0, in1, in2, in3);
+      break;
+    case H_FLIPADST:
+      load_buffer_32x16(input, in0, in1, in2, in3, stride, 0, 1);
+      fidtx16_sse2(in0, in1);
+      fidtx16_sse2(in2, in3);
+      round_signed_16x16(in0, in1);
+      round_signed_16x16(in2, in3);
+      fhalfright32_16col(in0, in1, in2, in3, no_transpose);
+      break;
+#endif
+    default: assert(0); break;
+  }
+  write_buffer_32x16(output, in0, in1, in2, in3);
+}
+
+// Note:
+// 32x32 hybrid fwd txfm
+//  4x2 grids of 8x16 block. Each block is represented by __m128i in[16]
+static INLINE void load_buffer_32x32(const int16_t *input,
+                                     __m128i *in0 /*in0[32]*/,
+                                     __m128i *in1 /*in1[32]*/,
+                                     __m128i *in2 /*in2[32]*/,
+                                     __m128i *in3 /*in3[32]*/, int stride,
+                                     int flipud, int fliplr) {
+  if (flipud) {
+    input += 31 * stride;
+    stride = -stride;
+  }
+
+  int i;
+  for (i = 0; i < 32; ++i) {
+    in0[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + i * stride + 0)), 2);
+    in1[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + i * stride + 8)), 2);
+    in2[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + i * stride + 16)), 2);
+    in3[i] = _mm_slli_epi16(
+        _mm_load_si128((const __m128i *)(input + i * stride + 24)), 2);
+  }
+
+  if (fliplr) {
+    for (i = 0; i < 32; ++i) {
+      __m128i tmp1 = in0[i];
+      __m128i tmp2 = in1[i];
+      in0[i] = mm_reverse_epi16(in3[i]);
+      in1[i] = mm_reverse_epi16(in2[i]);
+      in2[i] = mm_reverse_epi16(tmp2);
+      in3[i] = mm_reverse_epi16(tmp1);
+    }
+  }
+}
+
+static INLINE void swap_16x16(__m128i *b0l /*b0l[16]*/,
+                              __m128i *b0r /*b0r[16]*/,
+                              __m128i *b1l /*b1l[16]*/,
+                              __m128i *b1r /*b1r[16]*/) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    __m128i tmp0 = b1l[i];
+    __m128i tmp1 = b1r[i];
+    b1l[i] = b0l[i];
+    b1r[i] = b0r[i];
+    b0l[i] = tmp0;
+    b0r[i] = tmp1;
+  }
+}
+
+static INLINE void fdct32(__m128i *in0, __m128i *in1, __m128i *in2,
+                          __m128i *in3) {
+  fdct32_8col(in0, &in0[16]);
+  fdct32_8col(in1, &in1[16]);
+  fdct32_8col(in2, &in2[16]);
+  fdct32_8col(in3, &in3[16]);
+
+  array_transpose_16x16(in0, in1);
+  array_transpose_16x16(&in0[16], &in1[16]);
+  array_transpose_16x16(in2, in3);
+  array_transpose_16x16(&in2[16], &in3[16]);
+
+  swap_16x16(&in0[16], &in1[16], in2, in3);
+}
+
+static INLINE void fhalfright32(__m128i *in0, __m128i *in1, __m128i *in2,
+                                __m128i *in3) {
+  fhalfright32_16col(in0, in1, &in0[16], &in1[16], no_transpose);
+  fhalfright32_16col(in2, in3, &in2[16], &in3[16], no_transpose);
+  swap_16x16(&in0[16], &in1[16], in2, in3);
+}
+
+#if CONFIG_EXT_TX
+static INLINE void fidtx32(__m128i *in0, __m128i *in1, __m128i *in2,
+                           __m128i *in3) {
+  fidtx32_16col(in0, in1, &in0[16], &in1[16]);
+  fidtx32_16col(in2, in3, &in2[16], &in3[16]);
+  swap_16x16(&in0[16], &in1[16], in2, in3);
+}
+#endif
+
+static INLINE void round_signed_32x32(__m128i *in0, __m128i *in1, __m128i *in2,
+                                      __m128i *in3) {
+  round_signed_16x16(in0, in1);
+  round_signed_16x16(&in0[16], &in1[16]);
+  round_signed_16x16(in2, in3);
+  round_signed_16x16(&in2[16], &in3[16]);
+}
+
+static INLINE void write_buffer_32x32(__m128i *in0, __m128i *in1, __m128i *in2,
+                                      __m128i *in3, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 32; ++i) {
+    store_output(&in0[i], output + i * 32 + 0);
+    store_output(&in1[i], output + i * 32 + 8);
+    store_output(&in2[i], output + i * 32 + 16);
+    store_output(&in3[i], output + i * 32 + 24);
+  }
+}
+
+void av1_fht32x32_sse2(const int16_t *input, tran_low_t *output, int stride,
+                       int tx_type) {
+  __m128i in0[32], in1[32], in2[32], in3[32];
+
+  load_buffer_32x32(input, in0, in1, in2, in3, stride, 0, 0);
+  switch (tx_type) {
+    case DCT_DCT:
+      fdct32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fdct32(in0, in1, in2, in3);
+      break;
+    case ADST_DCT:
+      fhalfright32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fdct32(in0, in1, in2, in3);
+      break;
+    case DCT_ADST:
+      fdct32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fhalfright32(in0, in1, in2, in3);
+      break;
+    case ADST_ADST:
+      fhalfright32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fhalfright32(in0, in1, in2, in3);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_32x32(input, in0, in1, in2, in3, stride, 1, 0);
+      fhalfright32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fdct32(in0, in1, in2, in3);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_32x32(input, in0, in1, in2, in3, stride, 0, 1);
+      fdct32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fhalfright32(in0, in1, in2, in3);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_32x32(input, in0, in1, in2, in3, stride, 1, 1);
+      fhalfright32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fhalfright32(in0, in1, in2, in3);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_32x32(input, in0, in1, in2, in3, stride, 0, 1);
+      fhalfright32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fhalfright32(in0, in1, in2, in3);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_32x32(input, in0, in1, in2, in3, stride, 1, 0);
+      fhalfright32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fhalfright32(in0, in1, in2, in3);
+      break;
+    case IDTX:
+      fidtx32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fidtx32(in0, in1, in2, in3);
+      break;
+    case V_DCT:
+      fdct32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fidtx32(in0, in1, in2, in3);
+      break;
+    case H_DCT:
+      fidtx32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fdct32(in0, in1, in2, in3);
+      break;
+    case V_ADST:
+      fhalfright32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fidtx32(in0, in1, in2, in3);
+      break;
+    case H_ADST:
+      fidtx32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fhalfright32(in0, in1, in2, in3);
+      break;
+    case V_FLIPADST:
+      load_buffer_32x32(input, in0, in1, in2, in3, stride, 1, 0);
+      fhalfright32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fidtx32(in0, in1, in2, in3);
+      break;
+    case H_FLIPADST:
+      load_buffer_32x32(input, in0, in1, in2, in3, stride, 0, 1);
+      fidtx32(in0, in1, in2, in3);
+      round_signed_32x32(in0, in1, in2, in3);
+      fhalfright32(in0, in1, in2, in3);
+      break;
+#endif
+    default: assert(0);
+  }
+  write_buffer_32x32(in0, in1, in2, in3, output);
+}
diff --git a/third_party/aom/av1/encoder/x86/dct_sse2.asm b/third_party/aom/av1/encoder/x86/dct_sse2.asm
new file mode 100644
index 0000000000..a99db3d6ed
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/dct_sse2.asm
@@ -0,0 +1,87 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+%define private_prefix av1
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro TRANSFORM_COLS 0
+  paddw           m0,        m1
+  movq            m4,        m0
+  psubw           m3,        m2
+  psubw           m4,        m3
+  psraw           m4,        1
+  movq            m5,        m4
+  psubw           m5,        m1 ;b1
+  psubw           m4,        m2 ;c1
+  psubw           m0,        m4
+  paddw           m3,        m5
+                                ; m0 a0
+  SWAP            1,         4  ; m1 c1
+  SWAP            2,         3  ; m2 d1
+  SWAP            3,         5  ; m3 b1
+%endmacro
+
+%macro TRANSPOSE_4X4 0
+                                ; 00 01 02 03
+                                ; 10 11 12 13
+                                ; 20 21 22 23
+                                ; 30 31 32 33
+  punpcklwd       m0,        m1 ; 00 10 01 11  02 12 03 13
+  punpcklwd       m2,        m3 ; 20 30 21 31  22 32 23 33
+  mova            m1,        m0
+  punpckldq       m0,        m2 ; 00 10 20 30  01 11 21 31
+  punpckhdq       m1,        m2 ; 02 12 22 32  03 13 23 33
+%endmacro
+
+INIT_XMM sse2
+cglobal fwht4x4, 3, 4, 8, input, output, stride
+  lea             r3q,       [inputq + strideq*4]
+  movq            m0,        [inputq] ;a1
+  movq            m1,        [inputq + strideq*2] ;b1
+  movq            m2,        [r3q] ;c1
+  movq            m3,        [r3q + strideq*2] ;d1
+
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+  SWAP            1,         2
+  psrldq          m1,        m0, 8
+  psrldq          m3,        m2, 8
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+
+  psllw           m0,        2
+  psllw           m1,        2
+
+%if CONFIG_HIGHBITDEPTH
+  ; sign extension
+  mova            m2,             m0
+  mova            m3,             m1
+  punpcklwd       m0,             m0
+  punpcklwd       m1,             m1
+  punpckhwd       m2,             m2
+  punpckhwd       m3,             m3
+  psrad           m0,             16
+  psrad           m1,             16
+  psrad           m2,             16
+  psrad           m3,             16
+  mova            [outputq],      m0
+  mova            [outputq + 16], m2
+  mova            [outputq + 32], m1
+  mova            [outputq + 48], m3
+%else
+  mova            [outputq],      m0
+  mova            [outputq + 16], m1
+%endif
+
+  RET
diff --git a/third_party/aom/av1/encoder/x86/dct_ssse3.c b/third_party/aom/av1/encoder/x86/dct_ssse3.c
new file mode 100644
index 0000000000..717a99af8f
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/dct_ssse3.c
@@ -0,0 +1,469 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#if defined(_MSC_VER) && _MSC_VER <= 1500
+// Need to include math.h before calling tmmintrin.h/intrin.h
+// in certain versions of MSVS.
+#include <math.h>
+#endif
+#include <tmmintrin.h>  // SSSE3
+
+#include "./av1_rtcd.h"
+#include "aom_dsp/x86/inv_txfm_sse2.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+
+void av1_fdct8x8_quant_ssse3(
+    const int16_t *input, int stride, int16_t *coeff_ptr, intptr_t n_coeffs,
+    int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr,
+    const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
+    int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+    uint16_t *eob_ptr, const int16_t *scan_ptr, const int16_t *iscan_ptr) {
+  __m128i zero;
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__dual_p16_p16 = dual_set_epi16(23170, 23170);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // Load input
+  __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  __m128i *in[8];
+  int index = 0;
+
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)coeff_ptr;
+
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  in[0] = &in0;
+  in[1] = &in1;
+  in[2] = &in2;
+  in[3] = &in3;
+  in[4] = &in4;
+  in[5] = &in5;
+  in[6] = &in6;
+  in[7] = &in7;
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = _mm_add_epi16(in0, in7);
+    const __m128i q1 = _mm_add_epi16(in1, in6);
+    const __m128i q2 = _mm_add_epi16(in2, in5);
+    const __m128i q3 = _mm_add_epi16(in3, in4);
+    const __m128i q4 = _mm_sub_epi16(in3, in4);
+    const __m128i q5 = _mm_sub_epi16(in2, in5);
+    const __m128i q6 = _mm_sub_epi16(in1, in6);
+    const __m128i q7 = _mm_sub_epi16(in0, in7);
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = _mm_add_epi16(q0, q3);
+      const __m128i r1 = _mm_add_epi16(q1, q2);
+      const __m128i r2 = _mm_sub_epi16(q1, q2);
+      const __m128i r3 = _mm_sub_epi16(q0, q3);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+      const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+      const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+      const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+      // dct_const_round_shift
+
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+
+      res0 = _mm_packs_epi32(w0, w1);
+      res4 = _mm_packs_epi32(w2, w3);
+      res2 = _mm_packs_epi32(w4, w5);
+      res6 = _mm_packs_epi32(w6, w7);
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_sub_epi16(q6, q5);
+      const __m128i d1 = _mm_add_epi16(q6, q5);
+      const __m128i r0 = _mm_mulhrs_epi16(d0, k__dual_p16_p16);
+      const __m128i r1 = _mm_mulhrs_epi16(d1, k__dual_p16_p16);
+
+      // Add/subtract
+      const __m128i x0 = _mm_add_epi16(q4, r0);
+      const __m128i x1 = _mm_sub_epi16(q4, r0);
+      const __m128i x2 = _mm_sub_epi16(q7, r1);
+      const __m128i x3 = _mm_add_epi16(q7, r1);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+      const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+      const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+      const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+      // dct_const_round_shift
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+      res1 = _mm_packs_epi32(w0, w1);
+      res7 = _mm_packs_epi32(w2, w3);
+      res5 = _mm_packs_epi32(w4, w5);
+      res3 = _mm_packs_epi32(w6, w7);
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+  }
+
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant, thr;
+    int16_t nzflag;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i *)round_ptr);
+        quant = _mm_load_si128((const __m128i *)quant_ptr);
+        dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = *in[0];
+        coeff1 = *in[1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // AC only loop
+    index = 2;
+    thr = _mm_srai_epi16(dequant, 1);
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
+        coeff0 = *in[index];
+        coeff1 = *in[index + 1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+                 _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
+
+        if (nzflag) {
+          qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+          qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+          qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+          qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+          // Reinsert signs
+          qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+          qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+          qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+          qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+          coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+          coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+        } else {
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero);
+
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero);
+        }
+      }
+
+      if (nzflag) {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+      index += 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
diff --git a/third_party/aom/av1/encoder/x86/error_intrin_avx2.c b/third_party/aom/av1/encoder/x86/error_intrin_avx2.c
new file mode 100644
index 0000000000..ae733a1ce8
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/error_intrin_avx2.c
@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "./av1_rtcd.h"
+#include "aom/aom_integer.h"
+
+int64_t av1_block_error_avx2(const int16_t *coeff, const int16_t *dqcoeff,
+                             intptr_t block_size, int64_t *ssz) {
+  __m256i sse_reg, ssz_reg, coeff_reg, dqcoeff_reg;
+  __m256i exp_dqcoeff_lo, exp_dqcoeff_hi, exp_coeff_lo, exp_coeff_hi;
+  __m256i sse_reg_64hi, ssz_reg_64hi;
+  __m128i sse_reg128, ssz_reg128;
+  int64_t sse;
+  int i;
+  const __m256i zero_reg = _mm256_set1_epi16(0);
+
+  // init sse and ssz registerd to zero
+  sse_reg = _mm256_set1_epi16(0);
+  ssz_reg = _mm256_set1_epi16(0);
+
+  for (i = 0; i < block_size; i += 16) {
+    // load 32 bytes from coeff and dqcoeff
+    coeff_reg = _mm256_loadu_si256((const __m256i *)(coeff + i));
+    dqcoeff_reg = _mm256_loadu_si256((const __m256i *)(dqcoeff + i));
+    // dqcoeff - coeff
+    dqcoeff_reg = _mm256_sub_epi16(dqcoeff_reg, coeff_reg);
+    // madd (dqcoeff - coeff)
+    dqcoeff_reg = _mm256_madd_epi16(dqcoeff_reg, dqcoeff_reg);
+    // madd coeff
+    coeff_reg = _mm256_madd_epi16(coeff_reg, coeff_reg);
+    // expand each double word of madd (dqcoeff - coeff) to quad word
+    exp_dqcoeff_lo = _mm256_unpacklo_epi32(dqcoeff_reg, zero_reg);
+    exp_dqcoeff_hi = _mm256_unpackhi_epi32(dqcoeff_reg, zero_reg);
+    // expand each double word of madd (coeff) to quad word
+    exp_coeff_lo = _mm256_unpacklo_epi32(coeff_reg, zero_reg);
+    exp_coeff_hi = _mm256_unpackhi_epi32(coeff_reg, zero_reg);
+    // add each quad word of madd (dqcoeff - coeff) and madd (coeff)
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_lo);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_lo);
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_hi);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_hi);
+  }
+  // save the higher 64 bit of each 128 bit lane
+  sse_reg_64hi = _mm256_srli_si256(sse_reg, 8);
+  ssz_reg_64hi = _mm256_srli_si256(ssz_reg, 8);
+  // add the higher 64 bit to the low 64 bit
+  sse_reg = _mm256_add_epi64(sse_reg, sse_reg_64hi);
+  ssz_reg = _mm256_add_epi64(ssz_reg, ssz_reg_64hi);
+
+  // add each 64 bit from each of the 128 bit lane of the 256 bit
+  sse_reg128 = _mm_add_epi64(_mm256_castsi256_si128(sse_reg),
+                             _mm256_extractf128_si256(sse_reg, 1));
+
+  ssz_reg128 = _mm_add_epi64(_mm256_castsi256_si128(ssz_reg),
+                             _mm256_extractf128_si256(ssz_reg, 1));
+
+  // store the results
+  _mm_storel_epi64((__m128i *)(&sse), sse_reg128);
+
+  _mm_storel_epi64((__m128i *)(ssz), ssz_reg128);
+  _mm256_zeroupper();
+  return sse;
+}
diff --git a/third_party/aom/av1/encoder/x86/error_sse2.asm b/third_party/aom/av1/encoder/x86/error_sse2.asm
new file mode 100644
index 0000000000..4680f1fabd
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/error_sse2.asm
@@ -0,0 +1,125 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+%define private_prefix av1
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; int64_t av1_block_error(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size,
+;                         int64_t *ssz)
+
+INIT_XMM sse2
+cglobal block_error, 3, 3, 8, uqc, dqc, size, ssz
+  pxor      m4, m4                 ; sse accumulator
+  pxor      m6, m6                 ; ssz accumulator
+  pxor      m5, m5                 ; dedicated zero register
+  lea     uqcq, [uqcq+sizeq*2]
+  lea     dqcq, [dqcq+sizeq*2]
+  neg    sizeq
+.loop:
+  mova      m2, [uqcq+sizeq*2]
+  mova      m0, [dqcq+sizeq*2]
+  mova      m3, [uqcq+sizeq*2+mmsize]
+  mova      m1, [dqcq+sizeq*2+mmsize]
+  psubw     m0, m2
+  psubw     m1, m3
+  ; individual errors are max. 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit)
+  pmaddwd   m0, m0
+  pmaddwd   m1, m1
+  pmaddwd   m2, m2
+  pmaddwd   m3, m3
+  ; accumulate in 64bit
+  punpckldq m7, m0, m5
+  punpckhdq m0, m5
+  paddq     m4, m7
+  punpckldq m7, m1, m5
+  paddq     m4, m0
+  punpckhdq m1, m5
+  paddq     m4, m7
+  punpckldq m7, m2, m5
+  paddq     m4, m1
+  punpckhdq m2, m5
+  paddq     m6, m7
+  punpckldq m7, m3, m5
+  paddq     m6, m2
+  punpckhdq m3, m5
+  paddq     m6, m7
+  paddq     m6, m3
+  add    sizeq, mmsize
+  jl .loop
+
+  ; accumulate horizontally and store in return value
+  movhlps   m5, m4
+  movhlps   m7, m6
+  paddq     m4, m5
+  paddq     m6, m7
+%if ARCH_X86_64
+  movq    rax, m4
+  movq [sszq], m6
+%else
+  mov     eax, sszm
+  pshufd   m5, m4, 0x1
+  movq  [eax], m6
+  movd    eax, m4
+  movd    edx, m5
+%endif
+  RET
+
+; Compute the sum of squared difference between two int16_t vectors.
+; int64_t av1_block_error_fp(int16_t *coeff, int16_t *dqcoeff,
+;                            intptr_t block_size)
+
+INIT_XMM sse2
+cglobal block_error_fp, 3, 3, 6, uqc, dqc, size
+  pxor      m4, m4                 ; sse accumulator
+  pxor      m5, m5                 ; dedicated zero register
+  lea     uqcq, [uqcq+sizeq*2]
+  lea     dqcq, [dqcq+sizeq*2]
+  neg    sizeq
+.loop:
+  mova      m2, [uqcq+sizeq*2]
+  mova      m0, [dqcq+sizeq*2]
+  mova      m3, [uqcq+sizeq*2+mmsize]
+  mova      m1, [dqcq+sizeq*2+mmsize]
+  psubw     m0, m2
+  psubw     m1, m3
+  ; individual errors are max. 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit)
+  pmaddwd   m0, m0
+  pmaddwd   m1, m1
+  ; accumulate in 64bit
+  punpckldq m3, m0, m5
+  punpckhdq m0, m5
+  paddq     m4, m3
+  punpckldq m3, m1, m5
+  paddq     m4, m0
+  punpckhdq m1, m5
+  paddq     m4, m3
+  paddq     m4, m1
+  add    sizeq, mmsize
+  jl .loop
+
+  ; accumulate horizontally and store in return value
+  movhlps   m5, m4
+  paddq     m4, m5
+%if ARCH_X86_64
+  movq    rax, m4
+%else
+  pshufd   m5, m4, 0x1
+  movd    eax, m4
+  movd    edx, m5
+%endif
+  RET
diff --git a/third_party/aom/av1/encoder/x86/highbd_block_error_intrin_sse2.c b/third_party/aom/av1/encoder/x86/highbd_block_error_intrin_sse2.c
new file mode 100644
index 0000000000..777304ace7
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/highbd_block_error_intrin_sse2.c
@@ -0,0 +1,72 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <emmintrin.h>
+#include <stdio.h>
+
+#include "av1/common/common.h"
+
+int64_t av1_highbd_block_error_sse2(tran_low_t *coeff, tran_low_t *dqcoeff,
+                                    intptr_t block_size, int64_t *ssz,
+                                    int bps) {
+  int i, j, test;
+  uint32_t temp[4];
+  __m128i max, min, cmp0, cmp1, cmp2, cmp3;
+  int64_t error = 0, sqcoeff = 0;
+  const int shift = 2 * (bps - 8);
+  const int rounding = shift > 0 ? 1 << (shift - 1) : 0;
+
+  for (i = 0; i < block_size; i += 8) {
+    // Load the data into xmm registers
+    __m128i mm_coeff = _mm_load_si128((__m128i *)(coeff + i));
+    __m128i mm_coeff2 = _mm_load_si128((__m128i *)(coeff + i + 4));
+    __m128i mm_dqcoeff = _mm_load_si128((__m128i *)(dqcoeff + i));
+    __m128i mm_dqcoeff2 = _mm_load_si128((__m128i *)(dqcoeff + i + 4));
+    // Check if any values require more than 15 bit
+    max = _mm_set1_epi32(0x3fff);
+    min = _mm_set1_epi32(0xffffc000);
+    cmp0 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff, max),
+                         _mm_cmplt_epi32(mm_coeff, min));
+    cmp1 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff2, max),
+                         _mm_cmplt_epi32(mm_coeff2, min));
+    cmp2 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff, max),
+                         _mm_cmplt_epi32(mm_dqcoeff, min));
+    cmp3 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff2, max),
+                         _mm_cmplt_epi32(mm_dqcoeff2, min));
+    test = _mm_movemask_epi8(
+        _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3)));
+
+    if (!test) {
+      __m128i mm_diff, error_sse2, sqcoeff_sse2;
+      mm_coeff = _mm_packs_epi32(mm_coeff, mm_coeff2);
+      mm_dqcoeff = _mm_packs_epi32(mm_dqcoeff, mm_dqcoeff2);
+      mm_diff = _mm_sub_epi16(mm_coeff, mm_dqcoeff);
+      error_sse2 = _mm_madd_epi16(mm_diff, mm_diff);
+      sqcoeff_sse2 = _mm_madd_epi16(mm_coeff, mm_coeff);
+      _mm_storeu_si128((__m128i *)temp, error_sse2);
+      error = error + temp[0] + temp[1] + temp[2] + temp[3];
+      _mm_storeu_si128((__m128i *)temp, sqcoeff_sse2);
+      sqcoeff += temp[0] + temp[1] + temp[2] + temp[3];
+    } else {
+      for (j = 0; j < 8; j++) {
+        const int64_t diff = coeff[i + j] - dqcoeff[i + j];
+        error += diff * diff;
+        sqcoeff += (int64_t)coeff[i + j] * (int64_t)coeff[i + j];
+      }
+    }
+  }
+  assert(error >= 0 && sqcoeff >= 0);
+  error = (error + rounding) >> shift;
+  sqcoeff = (sqcoeff + rounding) >> shift;
+
+  *ssz = sqcoeff;
+  return error;
+}
diff --git a/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c b/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c
new file mode 100644
index 0000000000..f201a29aaa
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c
@@ -0,0 +1,1895 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <assert.h>
+#include <smmintrin.h> /* SSE4.1 */
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "av1/common/av1_fwd_txfm2d_cfg.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/x86/highbd_txfm_utility_sse4.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+#include "aom_ports/mem.h"
+
+static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in,
+                                   int stride, int flipud, int fliplr,
+                                   int shift) {
+  if (!flipud) {
+    in[0] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+    in[1] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+    in[2] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
+    in[3] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
+  } else {
+    in[0] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
+    in[1] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
+    in[2] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+    in[3] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+  }
+
+  if (fliplr) {
+    in[0] = _mm_shufflelo_epi16(in[0], 0x1b);
+    in[1] = _mm_shufflelo_epi16(in[1], 0x1b);
+    in[2] = _mm_shufflelo_epi16(in[2], 0x1b);
+    in[3] = _mm_shufflelo_epi16(in[3], 0x1b);
+  }
+
+  in[0] = _mm_cvtepi16_epi32(in[0]);
+  in[1] = _mm_cvtepi16_epi32(in[1]);
+  in[2] = _mm_cvtepi16_epi32(in[2]);
+  in[3] = _mm_cvtepi16_epi32(in[3]);
+
+  in[0] = _mm_slli_epi32(in[0], shift);
+  in[1] = _mm_slli_epi32(in[1], shift);
+  in[2] = _mm_slli_epi32(in[2], shift);
+  in[3] = _mm_slli_epi32(in[3], shift);
+}
+
+// We only use stage-2 bit;
+// shift[0] is used in load_buffer_4x4()
+// shift[1] is used in txfm_func_col()
+// shift[2] is used in txfm_func_row()
+static void fdct4x4_sse4_1(__m128i *in, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  __m128i s0, s1, s2, s3;
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3;
+
+  s0 = _mm_add_epi32(in[0], in[3]);
+  s1 = _mm_add_epi32(in[1], in[2]);
+  s2 = _mm_sub_epi32(in[1], in[2]);
+  s3 = _mm_sub_epi32(in[0], in[3]);
+
+  // btf_32_sse4_1_type0(cospi32, cospi32, s[01], u[02], bit);
+  u0 = _mm_mullo_epi32(s0, cospi32);
+  u1 = _mm_mullo_epi32(s1, cospi32);
+  u2 = _mm_add_epi32(u0, u1);
+  v0 = _mm_sub_epi32(u0, u1);
+
+  u3 = _mm_add_epi32(u2, rnding);
+  v1 = _mm_add_epi32(v0, rnding);
+
+  u0 = _mm_srai_epi32(u3, bit);
+  u2 = _mm_srai_epi32(v1, bit);
+
+  // btf_32_sse4_1_type1(cospi48, cospi16, s[23], u[13], bit);
+  v0 = _mm_mullo_epi32(s2, cospi48);
+  v1 = _mm_mullo_epi32(s3, cospi16);
+  v2 = _mm_add_epi32(v0, v1);
+
+  v3 = _mm_add_epi32(v2, rnding);
+  u1 = _mm_srai_epi32(v3, bit);
+
+  v0 = _mm_mullo_epi32(s2, cospi16);
+  v1 = _mm_mullo_epi32(s3, cospi48);
+  v2 = _mm_sub_epi32(v1, v0);
+
+  v3 = _mm_add_epi32(v2, rnding);
+  u3 = _mm_srai_epi32(v3, bit);
+
+  // Note: shift[1] and shift[2] are zeros
+
+  // Transpose 4x4 32-bit
+  v0 = _mm_unpacklo_epi32(u0, u1);
+  v1 = _mm_unpackhi_epi32(u0, u1);
+  v2 = _mm_unpacklo_epi32(u2, u3);
+  v3 = _mm_unpackhi_epi32(u2, u3);
+
+  in[0] = _mm_unpacklo_epi64(v0, v2);
+  in[1] = _mm_unpackhi_epi64(v0, v2);
+  in[2] = _mm_unpacklo_epi64(v1, v3);
+  in[3] = _mm_unpackhi_epi64(v1, v3);
+}
+
+static INLINE void write_buffer_4x4(__m128i *res, tran_low_t *output) {
+  _mm_store_si128((__m128i *)(output + 0 * 4), res[0]);
+  _mm_store_si128((__m128i *)(output + 1 * 4), res[1]);
+  _mm_store_si128((__m128i *)(output + 2 * 4), res[2]);
+  _mm_store_si128((__m128i *)(output + 3 * 4), res[3]);
+}
+
+// Note:
+//  We implement av1_fwd_txfm2d_4x4(). This function is kept here since
+//  av1_highbd_fht4x4_c() is not removed yet
+void av1_highbd_fht4x4_sse4_1(const int16_t *input, tran_low_t *output,
+                              int stride, int tx_type) {
+  (void)input;
+  (void)output;
+  (void)stride;
+  (void)tx_type;
+  assert(0);
+}
+
+static void fadst4x4_sse4_1(__m128i *in, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const __m128i kZero = _mm_setzero_si128();
+  __m128i s0, s1, s2, s3;
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3;
+
+  // stage 0
+  // stage 1
+  // stage 2
+  u0 = _mm_mullo_epi32(in[3], cospi8);
+  u1 = _mm_mullo_epi32(in[0], cospi56);
+  u2 = _mm_add_epi32(u0, u1);
+  s0 = _mm_add_epi32(u2, rnding);
+  s0 = _mm_srai_epi32(s0, bit);
+
+  v0 = _mm_mullo_epi32(in[3], cospi56);
+  v1 = _mm_mullo_epi32(in[0], cospi8);
+  v2 = _mm_sub_epi32(v0, v1);
+  s1 = _mm_add_epi32(v2, rnding);
+  s1 = _mm_srai_epi32(s1, bit);
+
+  u0 = _mm_mullo_epi32(in[1], cospi40);
+  u1 = _mm_mullo_epi32(in[2], cospi24);
+  u2 = _mm_add_epi32(u0, u1);
+  s2 = _mm_add_epi32(u2, rnding);
+  s2 = _mm_srai_epi32(s2, bit);
+
+  v0 = _mm_mullo_epi32(in[1], cospi24);
+  v1 = _mm_mullo_epi32(in[2], cospi40);
+  v2 = _mm_sub_epi32(v0, v1);
+  s3 = _mm_add_epi32(v2, rnding);
+  s3 = _mm_srai_epi32(s3, bit);
+
+  // stage 3
+  u0 = _mm_add_epi32(s0, s2);
+  u2 = _mm_sub_epi32(s0, s2);
+  u1 = _mm_add_epi32(s1, s3);
+  u3 = _mm_sub_epi32(s1, s3);
+
+  // stage 4
+  v0 = _mm_mullo_epi32(u2, cospi32);
+  v1 = _mm_mullo_epi32(u3, cospi32);
+  v2 = _mm_add_epi32(v0, v1);
+  s2 = _mm_add_epi32(v2, rnding);
+  u2 = _mm_srai_epi32(s2, bit);
+
+  v2 = _mm_sub_epi32(v0, v1);
+  s3 = _mm_add_epi32(v2, rnding);
+  u3 = _mm_srai_epi32(s3, bit);
+
+  // u0, u1, u2, u3
+  u2 = _mm_sub_epi32(kZero, u2);
+  u1 = _mm_sub_epi32(kZero, u1);
+
+  // u0, u2, u3, u1
+  // Transpose 4x4 32-bit
+  v0 = _mm_unpacklo_epi32(u0, u2);
+  v1 = _mm_unpackhi_epi32(u0, u2);
+  v2 = _mm_unpacklo_epi32(u3, u1);
+  v3 = _mm_unpackhi_epi32(u3, u1);
+
+  in[0] = _mm_unpacklo_epi64(v0, v2);
+  in[1] = _mm_unpackhi_epi64(v0, v2);
+  in[2] = _mm_unpacklo_epi64(v1, v3);
+  in[3] = _mm_unpackhi_epi64(v1, v3);
+}
+
+void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *coeff,
+                               int input_stride, int tx_type, int bd) {
+  __m128i in[4];
+  const TXFM_2D_CFG *cfg = NULL;
+
+  switch (tx_type) {
+    case DCT_DCT:
+      cfg = &fwd_txfm_2d_cfg_dct_dct_4;
+      load_buffer_4x4(input, in, input_stride, 0, 0, cfg->shift[0]);
+      fdct4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      fdct4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      write_buffer_4x4(in, coeff);
+      break;
+    case ADST_DCT:
+      cfg = &fwd_txfm_2d_cfg_adst_dct_4;
+      load_buffer_4x4(input, in, input_stride, 0, 0, cfg->shift[0]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      fdct4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      write_buffer_4x4(in, coeff);
+      break;
+    case DCT_ADST:
+      cfg = &fwd_txfm_2d_cfg_dct_adst_4;
+      load_buffer_4x4(input, in, input_stride, 0, 0, cfg->shift[0]);
+      fdct4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      write_buffer_4x4(in, coeff);
+      break;
+    case ADST_ADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_4;
+      load_buffer_4x4(input, in, input_stride, 0, 0, cfg->shift[0]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      write_buffer_4x4(in, coeff);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      cfg = &fwd_txfm_2d_cfg_adst_dct_4;
+      load_buffer_4x4(input, in, input_stride, 1, 0, cfg->shift[0]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      fdct4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      write_buffer_4x4(in, coeff);
+      break;
+    case DCT_FLIPADST:
+      cfg = &fwd_txfm_2d_cfg_dct_adst_4;
+      load_buffer_4x4(input, in, input_stride, 0, 1, cfg->shift[0]);
+      fdct4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      write_buffer_4x4(in, coeff);
+      break;
+    case FLIPADST_FLIPADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_4;
+      load_buffer_4x4(input, in, input_stride, 1, 1, cfg->shift[0]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      write_buffer_4x4(in, coeff);
+      break;
+    case ADST_FLIPADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_4;
+      load_buffer_4x4(input, in, input_stride, 0, 1, cfg->shift[0]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      write_buffer_4x4(in, coeff);
+      break;
+    case FLIPADST_ADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_4;
+      load_buffer_4x4(input, in, input_stride, 1, 0, cfg->shift[0]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_col[2]);
+      fadst4x4_sse4_1(in, cfg->cos_bit_row[2]);
+      write_buffer_4x4(in, coeff);
+      break;
+#endif
+    default: assert(0);
+  }
+  (void)bd;
+}
+
+static INLINE void load_buffer_8x8(const int16_t *input, __m128i *in,
+                                   int stride, int flipud, int fliplr,
+                                   int shift) {
+  __m128i u;
+  if (!flipud) {
+    in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride));
+    in[1] = _mm_load_si128((const __m128i *)(input + 1 * stride));
+    in[2] = _mm_load_si128((const __m128i *)(input + 2 * stride));
+    in[3] = _mm_load_si128((const __m128i *)(input + 3 * stride));
+    in[4] = _mm_load_si128((const __m128i *)(input + 4 * stride));
+    in[5] = _mm_load_si128((const __m128i *)(input + 5 * stride));
+    in[6] = _mm_load_si128((const __m128i *)(input + 6 * stride));
+    in[7] = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  } else {
+    in[0] = _mm_load_si128((const __m128i *)(input + 7 * stride));
+    in[1] = _mm_load_si128((const __m128i *)(input + 6 * stride));
+    in[2] = _mm_load_si128((const __m128i *)(input + 5 * stride));
+    in[3] = _mm_load_si128((const __m128i *)(input + 4 * stride));
+    in[4] = _mm_load_si128((const __m128i *)(input + 3 * stride));
+    in[5] = _mm_load_si128((const __m128i *)(input + 2 * stride));
+    in[6] = _mm_load_si128((const __m128i *)(input + 1 * stride));
+    in[7] = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  }
+
+  if (fliplr) {
+    in[0] = mm_reverse_epi16(in[0]);
+    in[1] = mm_reverse_epi16(in[1]);
+    in[2] = mm_reverse_epi16(in[2]);
+    in[3] = mm_reverse_epi16(in[3]);
+    in[4] = mm_reverse_epi16(in[4]);
+    in[5] = mm_reverse_epi16(in[5]);
+    in[6] = mm_reverse_epi16(in[6]);
+    in[7] = mm_reverse_epi16(in[7]);
+  }
+
+  u = _mm_unpackhi_epi64(in[4], in[4]);
+  in[8] = _mm_cvtepi16_epi32(in[4]);
+  in[9] = _mm_cvtepi16_epi32(u);
+
+  u = _mm_unpackhi_epi64(in[5], in[5]);
+  in[10] = _mm_cvtepi16_epi32(in[5]);
+  in[11] = _mm_cvtepi16_epi32(u);
+
+  u = _mm_unpackhi_epi64(in[6], in[6]);
+  in[12] = _mm_cvtepi16_epi32(in[6]);
+  in[13] = _mm_cvtepi16_epi32(u);
+
+  u = _mm_unpackhi_epi64(in[7], in[7]);
+  in[14] = _mm_cvtepi16_epi32(in[7]);
+  in[15] = _mm_cvtepi16_epi32(u);
+
+  u = _mm_unpackhi_epi64(in[3], in[3]);
+  in[6] = _mm_cvtepi16_epi32(in[3]);
+  in[7] = _mm_cvtepi16_epi32(u);
+
+  u = _mm_unpackhi_epi64(in[2], in[2]);
+  in[4] = _mm_cvtepi16_epi32(in[2]);
+  in[5] = _mm_cvtepi16_epi32(u);
+
+  u = _mm_unpackhi_epi64(in[1], in[1]);
+  in[2] = _mm_cvtepi16_epi32(in[1]);
+  in[3] = _mm_cvtepi16_epi32(u);
+
+  u = _mm_unpackhi_epi64(in[0], in[0]);
+  in[0] = _mm_cvtepi16_epi32(in[0]);
+  in[1] = _mm_cvtepi16_epi32(u);
+
+  in[0] = _mm_slli_epi32(in[0], shift);
+  in[1] = _mm_slli_epi32(in[1], shift);
+  in[2] = _mm_slli_epi32(in[2], shift);
+  in[3] = _mm_slli_epi32(in[3], shift);
+  in[4] = _mm_slli_epi32(in[4], shift);
+  in[5] = _mm_slli_epi32(in[5], shift);
+  in[6] = _mm_slli_epi32(in[6], shift);
+  in[7] = _mm_slli_epi32(in[7], shift);
+
+  in[8] = _mm_slli_epi32(in[8], shift);
+  in[9] = _mm_slli_epi32(in[9], shift);
+  in[10] = _mm_slli_epi32(in[10], shift);
+  in[11] = _mm_slli_epi32(in[11], shift);
+  in[12] = _mm_slli_epi32(in[12], shift);
+  in[13] = _mm_slli_epi32(in[13], shift);
+  in[14] = _mm_slli_epi32(in[14], shift);
+  in[15] = _mm_slli_epi32(in[15], shift);
+}
+
+static INLINE void col_txfm_8x8_rounding(__m128i *in, int shift) {
+  const __m128i rounding = _mm_set1_epi32(1 << (shift - 1));
+
+  in[0] = _mm_add_epi32(in[0], rounding);
+  in[1] = _mm_add_epi32(in[1], rounding);
+  in[2] = _mm_add_epi32(in[2], rounding);
+  in[3] = _mm_add_epi32(in[3], rounding);
+  in[4] = _mm_add_epi32(in[4], rounding);
+  in[5] = _mm_add_epi32(in[5], rounding);
+  in[6] = _mm_add_epi32(in[6], rounding);
+  in[7] = _mm_add_epi32(in[7], rounding);
+  in[8] = _mm_add_epi32(in[8], rounding);
+  in[9] = _mm_add_epi32(in[9], rounding);
+  in[10] = _mm_add_epi32(in[10], rounding);
+  in[11] = _mm_add_epi32(in[11], rounding);
+  in[12] = _mm_add_epi32(in[12], rounding);
+  in[13] = _mm_add_epi32(in[13], rounding);
+  in[14] = _mm_add_epi32(in[14], rounding);
+  in[15] = _mm_add_epi32(in[15], rounding);
+
+  in[0] = _mm_srai_epi32(in[0], shift);
+  in[1] = _mm_srai_epi32(in[1], shift);
+  in[2] = _mm_srai_epi32(in[2], shift);
+  in[3] = _mm_srai_epi32(in[3], shift);
+  in[4] = _mm_srai_epi32(in[4], shift);
+  in[5] = _mm_srai_epi32(in[5], shift);
+  in[6] = _mm_srai_epi32(in[6], shift);
+  in[7] = _mm_srai_epi32(in[7], shift);
+  in[8] = _mm_srai_epi32(in[8], shift);
+  in[9] = _mm_srai_epi32(in[9], shift);
+  in[10] = _mm_srai_epi32(in[10], shift);
+  in[11] = _mm_srai_epi32(in[11], shift);
+  in[12] = _mm_srai_epi32(in[12], shift);
+  in[13] = _mm_srai_epi32(in[13], shift);
+  in[14] = _mm_srai_epi32(in[14], shift);
+  in[15] = _mm_srai_epi32(in[15], shift);
+}
+
+static INLINE void write_buffer_8x8(const __m128i *res, tran_low_t *output) {
+  _mm_store_si128((__m128i *)(output + 0 * 4), res[0]);
+  _mm_store_si128((__m128i *)(output + 1 * 4), res[1]);
+  _mm_store_si128((__m128i *)(output + 2 * 4), res[2]);
+  _mm_store_si128((__m128i *)(output + 3 * 4), res[3]);
+
+  _mm_store_si128((__m128i *)(output + 4 * 4), res[4]);
+  _mm_store_si128((__m128i *)(output + 5 * 4), res[5]);
+  _mm_store_si128((__m128i *)(output + 6 * 4), res[6]);
+  _mm_store_si128((__m128i *)(output + 7 * 4), res[7]);
+
+  _mm_store_si128((__m128i *)(output + 8 * 4), res[8]);
+  _mm_store_si128((__m128i *)(output + 9 * 4), res[9]);
+  _mm_store_si128((__m128i *)(output + 10 * 4), res[10]);
+  _mm_store_si128((__m128i *)(output + 11 * 4), res[11]);
+
+  _mm_store_si128((__m128i *)(output + 12 * 4), res[12]);
+  _mm_store_si128((__m128i *)(output + 13 * 4), res[13]);
+  _mm_store_si128((__m128i *)(output + 14 * 4), res[14]);
+  _mm_store_si128((__m128i *)(output + 15 * 4), res[15]);
+}
+
+static void fdct8x8_sse4_1(__m128i *in, __m128i *out, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  __m128i u[8], v[8];
+
+  // Even 8 points 0, 2, ..., 14
+  // stage 0
+  // stage 1
+  u[0] = _mm_add_epi32(in[0], in[14]);
+  v[7] = _mm_sub_epi32(in[0], in[14]);  // v[7]
+  u[1] = _mm_add_epi32(in[2], in[12]);
+  u[6] = _mm_sub_epi32(in[2], in[12]);
+  u[2] = _mm_add_epi32(in[4], in[10]);
+  u[5] = _mm_sub_epi32(in[4], in[10]);
+  u[3] = _mm_add_epi32(in[6], in[8]);
+  v[4] = _mm_sub_epi32(in[6], in[8]);  // v[4]
+
+  // stage 2
+  v[0] = _mm_add_epi32(u[0], u[3]);
+  v[3] = _mm_sub_epi32(u[0], u[3]);
+  v[1] = _mm_add_epi32(u[1], u[2]);
+  v[2] = _mm_sub_epi32(u[1], u[2]);
+
+  v[5] = _mm_mullo_epi32(u[5], cospim32);
+  v[6] = _mm_mullo_epi32(u[6], cospi32);
+  v[5] = _mm_add_epi32(v[5], v[6]);
+  v[5] = _mm_add_epi32(v[5], rnding);
+  v[5] = _mm_srai_epi32(v[5], bit);
+
+  u[0] = _mm_mullo_epi32(u[5], cospi32);
+  v[6] = _mm_mullo_epi32(u[6], cospim32);
+  v[6] = _mm_sub_epi32(u[0], v[6]);
+  v[6] = _mm_add_epi32(v[6], rnding);
+  v[6] = _mm_srai_epi32(v[6], bit);
+
+  // stage 3
+  // type 0
+  v[0] = _mm_mullo_epi32(v[0], cospi32);
+  v[1] = _mm_mullo_epi32(v[1], cospi32);
+  u[0] = _mm_add_epi32(v[0], v[1]);
+  u[0] = _mm_add_epi32(u[0], rnding);
+  u[0] = _mm_srai_epi32(u[0], bit);
+
+  u[1] = _mm_sub_epi32(v[0], v[1]);
+  u[1] = _mm_add_epi32(u[1], rnding);
+  u[1] = _mm_srai_epi32(u[1], bit);
+
+  // type 1
+  v[0] = _mm_mullo_epi32(v[2], cospi48);
+  v[1] = _mm_mullo_epi32(v[3], cospi16);
+  u[2] = _mm_add_epi32(v[0], v[1]);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  v[0] = _mm_mullo_epi32(v[2], cospi16);
+  v[1] = _mm_mullo_epi32(v[3], cospi48);
+  u[3] = _mm_sub_epi32(v[1], v[0]);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  u[4] = _mm_add_epi32(v[4], v[5]);
+  u[5] = _mm_sub_epi32(v[4], v[5]);
+  u[6] = _mm_sub_epi32(v[7], v[6]);
+  u[7] = _mm_add_epi32(v[7], v[6]);
+
+  // stage 4
+  // stage 5
+  v[0] = _mm_mullo_epi32(u[4], cospi56);
+  v[1] = _mm_mullo_epi32(u[7], cospi8);
+  v[0] = _mm_add_epi32(v[0], v[1]);
+  v[0] = _mm_add_epi32(v[0], rnding);
+  out[2] = _mm_srai_epi32(v[0], bit);  // buf0[4]
+
+  v[0] = _mm_mullo_epi32(u[4], cospi8);
+  v[1] = _mm_mullo_epi32(u[7], cospi56);
+  v[0] = _mm_sub_epi32(v[1], v[0]);
+  v[0] = _mm_add_epi32(v[0], rnding);
+  out[14] = _mm_srai_epi32(v[0], bit);  // buf0[7]
+
+  v[0] = _mm_mullo_epi32(u[5], cospi24);
+  v[1] = _mm_mullo_epi32(u[6], cospi40);
+  v[0] = _mm_add_epi32(v[0], v[1]);
+  v[0] = _mm_add_epi32(v[0], rnding);
+  out[10] = _mm_srai_epi32(v[0], bit);  // buf0[5]
+
+  v[0] = _mm_mullo_epi32(u[5], cospi40);
+  v[1] = _mm_mullo_epi32(u[6], cospi24);
+  v[0] = _mm_sub_epi32(v[1], v[0]);
+  v[0] = _mm_add_epi32(v[0], rnding);
+  out[6] = _mm_srai_epi32(v[0], bit);  // buf0[6]
+
+  out[0] = u[0];   // buf0[0]
+  out[8] = u[1];   // buf0[1]
+  out[4] = u[2];   // buf0[2]
+  out[12] = u[3];  // buf0[3]
+
+  // Odd 8 points: 1, 3, ..., 15
+  // stage 0
+  // stage 1
+  u[0] = _mm_add_epi32(in[1], in[15]);
+  v[7] = _mm_sub_epi32(in[1], in[15]);  // v[7]
+  u[1] = _mm_add_epi32(in[3], in[13]);
+  u[6] = _mm_sub_epi32(in[3], in[13]);
+  u[2] = _mm_add_epi32(in[5], in[11]);
+  u[5] = _mm_sub_epi32(in[5], in[11]);
+  u[3] = _mm_add_epi32(in[7], in[9]);
+  v[4] = _mm_sub_epi32(in[7], in[9]);  // v[4]
+
+  // stage 2
+  v[0] = _mm_add_epi32(u[0], u[3]);
+  v[3] = _mm_sub_epi32(u[0], u[3]);
+  v[1] = _mm_add_epi32(u[1], u[2]);
+  v[2] = _mm_sub_epi32(u[1], u[2]);
+
+  v[5] = _mm_mullo_epi32(u[5], cospim32);
+  v[6] = _mm_mullo_epi32(u[6], cospi32);
+  v[5] = _mm_add_epi32(v[5], v[6]);
+  v[5] = _mm_add_epi32(v[5], rnding);
+  v[5] = _mm_srai_epi32(v[5], bit);
+
+  u[0] = _mm_mullo_epi32(u[5], cospi32);
+  v[6] = _mm_mullo_epi32(u[6], cospim32);
+  v[6] = _mm_sub_epi32(u[0], v[6]);
+  v[6] = _mm_add_epi32(v[6], rnding);
+  v[6] = _mm_srai_epi32(v[6], bit);
+
+  // stage 3
+  // type 0
+  v[0] = _mm_mullo_epi32(v[0], cospi32);
+  v[1] = _mm_mullo_epi32(v[1], cospi32);
+  u[0] = _mm_add_epi32(v[0], v[1]);
+  u[0] = _mm_add_epi32(u[0], rnding);
+  u[0] = _mm_srai_epi32(u[0], bit);
+
+  u[1] = _mm_sub_epi32(v[0], v[1]);
+  u[1] = _mm_add_epi32(u[1], rnding);
+  u[1] = _mm_srai_epi32(u[1], bit);
+
+  // type 1
+  v[0] = _mm_mullo_epi32(v[2], cospi48);
+  v[1] = _mm_mullo_epi32(v[3], cospi16);
+  u[2] = _mm_add_epi32(v[0], v[1]);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  v[0] = _mm_mullo_epi32(v[2], cospi16);
+  v[1] = _mm_mullo_epi32(v[3], cospi48);
+  u[3] = _mm_sub_epi32(v[1], v[0]);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  u[4] = _mm_add_epi32(v[4], v[5]);
+  u[5] = _mm_sub_epi32(v[4], v[5]);
+  u[6] = _mm_sub_epi32(v[7], v[6]);
+  u[7] = _mm_add_epi32(v[7], v[6]);
+
+  // stage 4
+  // stage 5
+  v[0] = _mm_mullo_epi32(u[4], cospi56);
+  v[1] = _mm_mullo_epi32(u[7], cospi8);
+  v[0] = _mm_add_epi32(v[0], v[1]);
+  v[0] = _mm_add_epi32(v[0], rnding);
+  out[3] = _mm_srai_epi32(v[0], bit);  // buf0[4]
+
+  v[0] = _mm_mullo_epi32(u[4], cospi8);
+  v[1] = _mm_mullo_epi32(u[7], cospi56);
+  v[0] = _mm_sub_epi32(v[1], v[0]);
+  v[0] = _mm_add_epi32(v[0], rnding);
+  out[15] = _mm_srai_epi32(v[0], bit);  // buf0[7]
+
+  v[0] = _mm_mullo_epi32(u[5], cospi24);
+  v[1] = _mm_mullo_epi32(u[6], cospi40);
+  v[0] = _mm_add_epi32(v[0], v[1]);
+  v[0] = _mm_add_epi32(v[0], rnding);
+  out[11] = _mm_srai_epi32(v[0], bit);  // buf0[5]
+
+  v[0] = _mm_mullo_epi32(u[5], cospi40);
+  v[1] = _mm_mullo_epi32(u[6], cospi24);
+  v[0] = _mm_sub_epi32(v[1], v[0]);
+  v[0] = _mm_add_epi32(v[0], rnding);
+  out[7] = _mm_srai_epi32(v[0], bit);  // buf0[6]
+
+  out[1] = u[0];   // buf0[0]
+  out[9] = u[1];   // buf0[1]
+  out[5] = u[2];   // buf0[2]
+  out[13] = u[3];  // buf0[3]
+}
+
+static void fadst8x8_sse4_1(__m128i *in, __m128i *out, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const __m128i kZero = _mm_setzero_si128();
+  __m128i u[8], v[8], x;
+
+  // Even 8 points: 0, 2, ..., 14
+  // stage 0
+  // stage 1
+  // stage 2
+  // (1)
+  u[0] = _mm_mullo_epi32(in[14], cospi4);
+  x = _mm_mullo_epi32(in[0], cospi60);
+  u[0] = _mm_add_epi32(u[0], x);
+  u[0] = _mm_add_epi32(u[0], rnding);
+  u[0] = _mm_srai_epi32(u[0], bit);
+
+  u[1] = _mm_mullo_epi32(in[14], cospi60);
+  x = _mm_mullo_epi32(in[0], cospi4);
+  u[1] = _mm_sub_epi32(u[1], x);
+  u[1] = _mm_add_epi32(u[1], rnding);
+  u[1] = _mm_srai_epi32(u[1], bit);
+
+  // (2)
+  u[2] = _mm_mullo_epi32(in[10], cospi20);
+  x = _mm_mullo_epi32(in[4], cospi44);
+  u[2] = _mm_add_epi32(u[2], x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_mullo_epi32(in[10], cospi44);
+  x = _mm_mullo_epi32(in[4], cospi20);
+  u[3] = _mm_sub_epi32(u[3], x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  // (3)
+  u[4] = _mm_mullo_epi32(in[6], cospi36);
+  x = _mm_mullo_epi32(in[8], cospi28);
+  u[4] = _mm_add_epi32(u[4], x);
+  u[4] = _mm_add_epi32(u[4], rnding);
+  u[4] = _mm_srai_epi32(u[4], bit);
+
+  u[5] = _mm_mullo_epi32(in[6], cospi28);
+  x = _mm_mullo_epi32(in[8], cospi36);
+  u[5] = _mm_sub_epi32(u[5], x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  // (4)
+  u[6] = _mm_mullo_epi32(in[2], cospi52);
+  x = _mm_mullo_epi32(in[12], cospi12);
+  u[6] = _mm_add_epi32(u[6], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_mullo_epi32(in[2], cospi12);
+  x = _mm_mullo_epi32(in[12], cospi52);
+  u[7] = _mm_sub_epi32(u[7], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 3
+  v[0] = _mm_add_epi32(u[0], u[4]);
+  v[4] = _mm_sub_epi32(u[0], u[4]);
+  v[1] = _mm_add_epi32(u[1], u[5]);
+  v[5] = _mm_sub_epi32(u[1], u[5]);
+  v[2] = _mm_add_epi32(u[2], u[6]);
+  v[6] = _mm_sub_epi32(u[2], u[6]);
+  v[3] = _mm_add_epi32(u[3], u[7]);
+  v[7] = _mm_sub_epi32(u[3], u[7]);
+
+  // stage 4
+  u[0] = v[0];
+  u[1] = v[1];
+  u[2] = v[2];
+  u[3] = v[3];
+
+  u[4] = _mm_mullo_epi32(v[4], cospi16);
+  x = _mm_mullo_epi32(v[5], cospi48);
+  u[4] = _mm_add_epi32(u[4], x);
+  u[4] = _mm_add_epi32(u[4], rnding);
+  u[4] = _mm_srai_epi32(u[4], bit);
+
+  u[5] = _mm_mullo_epi32(v[4], cospi48);
+  x = _mm_mullo_epi32(v[5], cospi16);
+  u[5] = _mm_sub_epi32(u[5], x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  u[6] = _mm_mullo_epi32(v[6], cospim48);
+  x = _mm_mullo_epi32(v[7], cospi16);
+  u[6] = _mm_add_epi32(u[6], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_mullo_epi32(v[6], cospi16);
+  x = _mm_mullo_epi32(v[7], cospim48);
+  u[7] = _mm_sub_epi32(u[7], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 5
+  v[0] = _mm_add_epi32(u[0], u[2]);
+  v[2] = _mm_sub_epi32(u[0], u[2]);
+  v[1] = _mm_add_epi32(u[1], u[3]);
+  v[3] = _mm_sub_epi32(u[1], u[3]);
+  v[4] = _mm_add_epi32(u[4], u[6]);
+  v[6] = _mm_sub_epi32(u[4], u[6]);
+  v[5] = _mm_add_epi32(u[5], u[7]);
+  v[7] = _mm_sub_epi32(u[5], u[7]);
+
+  // stage 6
+  u[0] = v[0];
+  u[1] = v[1];
+  u[4] = v[4];
+  u[5] = v[5];
+
+  v[0] = _mm_mullo_epi32(v[2], cospi32);
+  x = _mm_mullo_epi32(v[3], cospi32);
+  u[2] = _mm_add_epi32(v[0], x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_sub_epi32(v[0], x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  v[0] = _mm_mullo_epi32(v[6], cospi32);
+  x = _mm_mullo_epi32(v[7], cospi32);
+  u[6] = _mm_add_epi32(v[0], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_sub_epi32(v[0], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 7
+  out[0] = u[0];
+  out[2] = _mm_sub_epi32(kZero, u[4]);
+  out[4] = u[6];
+  out[6] = _mm_sub_epi32(kZero, u[2]);
+  out[8] = u[3];
+  out[10] = _mm_sub_epi32(kZero, u[7]);
+  out[12] = u[5];
+  out[14] = _mm_sub_epi32(kZero, u[1]);
+
+  // Odd 8 points: 1, 3, ..., 15
+  // stage 0
+  // stage 1
+  // stage 2
+  // (1)
+  u[0] = _mm_mullo_epi32(in[15], cospi4);
+  x = _mm_mullo_epi32(in[1], cospi60);
+  u[0] = _mm_add_epi32(u[0], x);
+  u[0] = _mm_add_epi32(u[0], rnding);
+  u[0] = _mm_srai_epi32(u[0], bit);
+
+  u[1] = _mm_mullo_epi32(in[15], cospi60);
+  x = _mm_mullo_epi32(in[1], cospi4);
+  u[1] = _mm_sub_epi32(u[1], x);
+  u[1] = _mm_add_epi32(u[1], rnding);
+  u[1] = _mm_srai_epi32(u[1], bit);
+
+  // (2)
+  u[2] = _mm_mullo_epi32(in[11], cospi20);
+  x = _mm_mullo_epi32(in[5], cospi44);
+  u[2] = _mm_add_epi32(u[2], x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_mullo_epi32(in[11], cospi44);
+  x = _mm_mullo_epi32(in[5], cospi20);
+  u[3] = _mm_sub_epi32(u[3], x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  // (3)
+  u[4] = _mm_mullo_epi32(in[7], cospi36);
+  x = _mm_mullo_epi32(in[9], cospi28);
+  u[4] = _mm_add_epi32(u[4], x);
+  u[4] = _mm_add_epi32(u[4], rnding);
+  u[4] = _mm_srai_epi32(u[4], bit);
+
+  u[5] = _mm_mullo_epi32(in[7], cospi28);
+  x = _mm_mullo_epi32(in[9], cospi36);
+  u[5] = _mm_sub_epi32(u[5], x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  // (4)
+  u[6] = _mm_mullo_epi32(in[3], cospi52);
+  x = _mm_mullo_epi32(in[13], cospi12);
+  u[6] = _mm_add_epi32(u[6], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_mullo_epi32(in[3], cospi12);
+  x = _mm_mullo_epi32(in[13], cospi52);
+  u[7] = _mm_sub_epi32(u[7], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 3
+  v[0] = _mm_add_epi32(u[0], u[4]);
+  v[4] = _mm_sub_epi32(u[0], u[4]);
+  v[1] = _mm_add_epi32(u[1], u[5]);
+  v[5] = _mm_sub_epi32(u[1], u[5]);
+  v[2] = _mm_add_epi32(u[2], u[6]);
+  v[6] = _mm_sub_epi32(u[2], u[6]);
+  v[3] = _mm_add_epi32(u[3], u[7]);
+  v[7] = _mm_sub_epi32(u[3], u[7]);
+
+  // stage 4
+  u[0] = v[0];
+  u[1] = v[1];
+  u[2] = v[2];
+  u[3] = v[3];
+
+  u[4] = _mm_mullo_epi32(v[4], cospi16);
+  x = _mm_mullo_epi32(v[5], cospi48);
+  u[4] = _mm_add_epi32(u[4], x);
+  u[4] = _mm_add_epi32(u[4], rnding);
+  u[4] = _mm_srai_epi32(u[4], bit);
+
+  u[5] = _mm_mullo_epi32(v[4], cospi48);
+  x = _mm_mullo_epi32(v[5], cospi16);
+  u[5] = _mm_sub_epi32(u[5], x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  u[6] = _mm_mullo_epi32(v[6], cospim48);
+  x = _mm_mullo_epi32(v[7], cospi16);
+  u[6] = _mm_add_epi32(u[6], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_mullo_epi32(v[6], cospi16);
+  x = _mm_mullo_epi32(v[7], cospim48);
+  u[7] = _mm_sub_epi32(u[7], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 5
+  v[0] = _mm_add_epi32(u[0], u[2]);
+  v[2] = _mm_sub_epi32(u[0], u[2]);
+  v[1] = _mm_add_epi32(u[1], u[3]);
+  v[3] = _mm_sub_epi32(u[1], u[3]);
+  v[4] = _mm_add_epi32(u[4], u[6]);
+  v[6] = _mm_sub_epi32(u[4], u[6]);
+  v[5] = _mm_add_epi32(u[5], u[7]);
+  v[7] = _mm_sub_epi32(u[5], u[7]);
+
+  // stage 6
+  u[0] = v[0];
+  u[1] = v[1];
+  u[4] = v[4];
+  u[5] = v[5];
+
+  v[0] = _mm_mullo_epi32(v[2], cospi32);
+  x = _mm_mullo_epi32(v[3], cospi32);
+  u[2] = _mm_add_epi32(v[0], x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_sub_epi32(v[0], x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  v[0] = _mm_mullo_epi32(v[6], cospi32);
+  x = _mm_mullo_epi32(v[7], cospi32);
+  u[6] = _mm_add_epi32(v[0], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_sub_epi32(v[0], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 7
+  out[1] = u[0];
+  out[3] = _mm_sub_epi32(kZero, u[4]);
+  out[5] = u[6];
+  out[7] = _mm_sub_epi32(kZero, u[2]);
+  out[9] = u[3];
+  out[11] = _mm_sub_epi32(kZero, u[7]);
+  out[13] = u[5];
+  out[15] = _mm_sub_epi32(kZero, u[1]);
+}
+
+void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *coeff, int stride,
+                               int tx_type, int bd) {
+  __m128i in[16], out[16];
+  const TXFM_2D_CFG *cfg = NULL;
+
+  switch (tx_type) {
+    case DCT_DCT:
+      cfg = &fwd_txfm_2d_cfg_dct_dct_8;
+      load_buffer_8x8(input, in, stride, 0, 0, cfg->shift[0]);
+      fdct8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+      col_txfm_8x8_rounding(out, -cfg->shift[1]);
+      transpose_8x8(out, in);
+      fdct8x8_sse4_1(in, out, cfg->cos_bit_row[2]);
+      transpose_8x8(out, in);
+      write_buffer_8x8(in, coeff);
+      break;
+    case ADST_DCT:
+      cfg = &fwd_txfm_2d_cfg_adst_dct_8;
+      load_buffer_8x8(input, in, stride, 0, 0, cfg->shift[0]);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+      col_txfm_8x8_rounding(out, -cfg->shift[1]);
+      transpose_8x8(out, in);
+      fdct8x8_sse4_1(in, out, cfg->cos_bit_row[2]);
+      transpose_8x8(out, in);
+      write_buffer_8x8(in, coeff);
+      break;
+    case DCT_ADST:
+      cfg = &fwd_txfm_2d_cfg_dct_adst_8;
+      load_buffer_8x8(input, in, stride, 0, 0, cfg->shift[0]);
+      fdct8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+      col_txfm_8x8_rounding(out, -cfg->shift[1]);
+      transpose_8x8(out, in);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_row[2]);
+      transpose_8x8(out, in);
+      write_buffer_8x8(in, coeff);
+      break;
+    case ADST_ADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_8;
+      load_buffer_8x8(input, in, stride, 0, 0, cfg->shift[0]);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+      col_txfm_8x8_rounding(out, -cfg->shift[1]);
+      transpose_8x8(out, in);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_row[2]);
+      transpose_8x8(out, in);
+      write_buffer_8x8(in, coeff);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      cfg = &fwd_txfm_2d_cfg_adst_dct_8;
+      load_buffer_8x8(input, in, stride, 1, 0, cfg->shift[0]);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+      col_txfm_8x8_rounding(out, -cfg->shift[1]);
+      transpose_8x8(out, in);
+      fdct8x8_sse4_1(in, out, cfg->cos_bit_row[2]);
+      transpose_8x8(out, in);
+      write_buffer_8x8(in, coeff);
+      break;
+    case DCT_FLIPADST:
+      cfg = &fwd_txfm_2d_cfg_dct_adst_8;
+      load_buffer_8x8(input, in, stride, 0, 1, cfg->shift[0]);
+      fdct8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+      col_txfm_8x8_rounding(out, -cfg->shift[1]);
+      transpose_8x8(out, in);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_row[2]);
+      transpose_8x8(out, in);
+      write_buffer_8x8(in, coeff);
+      break;
+    case FLIPADST_FLIPADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_8;
+      load_buffer_8x8(input, in, stride, 1, 1, cfg->shift[0]);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+      col_txfm_8x8_rounding(out, -cfg->shift[1]);
+      transpose_8x8(out, in);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_row[2]);
+      transpose_8x8(out, in);
+      write_buffer_8x8(in, coeff);
+      break;
+    case ADST_FLIPADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_8;
+      load_buffer_8x8(input, in, stride, 0, 1, cfg->shift[0]);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+      col_txfm_8x8_rounding(out, -cfg->shift[1]);
+      transpose_8x8(out, in);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_row[2]);
+      transpose_8x8(out, in);
+      write_buffer_8x8(in, coeff);
+      break;
+    case FLIPADST_ADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_8;
+      load_buffer_8x8(input, in, stride, 1, 0, cfg->shift[0]);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_col[2]);
+      col_txfm_8x8_rounding(out, -cfg->shift[1]);
+      transpose_8x8(out, in);
+      fadst8x8_sse4_1(in, out, cfg->cos_bit_row[2]);
+      transpose_8x8(out, in);
+      write_buffer_8x8(in, coeff);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+  (void)bd;
+}
+
+// Hybrid Transform 16x16
+
+static INLINE void convert_8x8_to_16x16(const __m128i *in, __m128i *out) {
+  int row_index = 0;
+  int dst_index = 0;
+  int src_index = 0;
+
+  // row 0, 1, .., 7
+  do {
+    out[dst_index] = in[src_index];
+    out[dst_index + 1] = in[src_index + 1];
+    out[dst_index + 2] = in[src_index + 16];
+    out[dst_index + 3] = in[src_index + 17];
+    dst_index += 4;
+    src_index += 2;
+    row_index += 1;
+  } while (row_index < 8);
+
+  // row 8, 9, ..., 15
+  src_index += 16;
+  do {
+    out[dst_index] = in[src_index];
+    out[dst_index + 1] = in[src_index + 1];
+    out[dst_index + 2] = in[src_index + 16];
+    out[dst_index + 3] = in[src_index + 17];
+    dst_index += 4;
+    src_index += 2;
+    row_index += 1;
+  } while (row_index < 16);
+}
+
+static INLINE void load_buffer_16x16(const int16_t *input, __m128i *out,
+                                     int stride, int flipud, int fliplr,
+                                     int shift) {
+  __m128i in[64];
+  // Load 4 8x8 blocks
+  const int16_t *topL = input;
+  const int16_t *topR = input + 8;
+  const int16_t *botL = input + 8 * stride;
+  const int16_t *botR = input + 8 * stride + 8;
+
+  const int16_t *tmp;
+
+  if (flipud) {
+    // Swap left columns
+    tmp = topL;
+    topL = botL;
+    botL = tmp;
+    // Swap right columns
+    tmp = topR;
+    topR = botR;
+    botR = tmp;
+  }
+
+  if (fliplr) {
+    // Swap top rows
+    tmp = topL;
+    topL = topR;
+    topR = tmp;
+    // Swap bottom rows
+    tmp = botL;
+    botL = botR;
+    botR = tmp;
+  }
+
+  // load first 8 columns
+  load_buffer_8x8(topL, &in[0], stride, flipud, fliplr, shift);
+  load_buffer_8x8(botL, &in[32], stride, flipud, fliplr, shift);
+
+  // load second 8 columns
+  load_buffer_8x8(topR, &in[16], stride, flipud, fliplr, shift);
+  load_buffer_8x8(botR, &in[48], stride, flipud, fliplr, shift);
+
+  convert_8x8_to_16x16(in, out);
+}
+
+static void fdct16x16_sse4_1(__m128i *in, __m128i *out, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  __m128i u[16], v[16], x;
+  const int col_num = 4;
+  int col;
+
+  // Calculate the column 0, 1, 2, 3
+  for (col = 0; col < col_num; ++col) {
+    // stage 0
+    // stage 1
+    u[0] = _mm_add_epi32(in[0 * col_num + col], in[15 * col_num + col]);
+    u[15] = _mm_sub_epi32(in[0 * col_num + col], in[15 * col_num + col]);
+    u[1] = _mm_add_epi32(in[1 * col_num + col], in[14 * col_num + col]);
+    u[14] = _mm_sub_epi32(in[1 * col_num + col], in[14 * col_num + col]);
+    u[2] = _mm_add_epi32(in[2 * col_num + col], in[13 * col_num + col]);
+    u[13] = _mm_sub_epi32(in[2 * col_num + col], in[13 * col_num + col]);
+    u[3] = _mm_add_epi32(in[3 * col_num + col], in[12 * col_num + col]);
+    u[12] = _mm_sub_epi32(in[3 * col_num + col], in[12 * col_num + col]);
+    u[4] = _mm_add_epi32(in[4 * col_num + col], in[11 * col_num + col]);
+    u[11] = _mm_sub_epi32(in[4 * col_num + col], in[11 * col_num + col]);
+    u[5] = _mm_add_epi32(in[5 * col_num + col], in[10 * col_num + col]);
+    u[10] = _mm_sub_epi32(in[5 * col_num + col], in[10 * col_num + col]);
+    u[6] = _mm_add_epi32(in[6 * col_num + col], in[9 * col_num + col]);
+    u[9] = _mm_sub_epi32(in[6 * col_num + col], in[9 * col_num + col]);
+    u[7] = _mm_add_epi32(in[7 * col_num + col], in[8 * col_num + col]);
+    u[8] = _mm_sub_epi32(in[7 * col_num + col], in[8 * col_num + col]);
+
+    // stage 2
+    v[0] = _mm_add_epi32(u[0], u[7]);
+    v[7] = _mm_sub_epi32(u[0], u[7]);
+    v[1] = _mm_add_epi32(u[1], u[6]);
+    v[6] = _mm_sub_epi32(u[1], u[6]);
+    v[2] = _mm_add_epi32(u[2], u[5]);
+    v[5] = _mm_sub_epi32(u[2], u[5]);
+    v[3] = _mm_add_epi32(u[3], u[4]);
+    v[4] = _mm_sub_epi32(u[3], u[4]);
+    v[8] = u[8];
+    v[9] = u[9];
+
+    v[10] = _mm_mullo_epi32(u[10], cospim32);
+    x = _mm_mullo_epi32(u[13], cospi32);
+    v[10] = _mm_add_epi32(v[10], x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[13] = _mm_mullo_epi32(u[10], cospi32);
+    x = _mm_mullo_epi32(u[13], cospim32);
+    v[13] = _mm_sub_epi32(v[13], x);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    v[11] = _mm_mullo_epi32(u[11], cospim32);
+    x = _mm_mullo_epi32(u[12], cospi32);
+    v[11] = _mm_add_epi32(v[11], x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = _mm_mullo_epi32(u[11], cospi32);
+    x = _mm_mullo_epi32(u[12], cospim32);
+    v[12] = _mm_sub_epi32(v[12], x);
+    v[12] = _mm_add_epi32(v[12], rnding);
+    v[12] = _mm_srai_epi32(v[12], bit);
+    v[14] = u[14];
+    v[15] = u[15];
+
+    // stage 3
+    u[0] = _mm_add_epi32(v[0], v[3]);
+    u[3] = _mm_sub_epi32(v[0], v[3]);
+    u[1] = _mm_add_epi32(v[1], v[2]);
+    u[2] = _mm_sub_epi32(v[1], v[2]);
+    u[4] = v[4];
+
+    u[5] = _mm_mullo_epi32(v[5], cospim32);
+    x = _mm_mullo_epi32(v[6], cospi32);
+    u[5] = _mm_add_epi32(u[5], x);
+    u[5] = _mm_add_epi32(u[5], rnding);
+    u[5] = _mm_srai_epi32(u[5], bit);
+
+    u[6] = _mm_mullo_epi32(v[5], cospi32);
+    x = _mm_mullo_epi32(v[6], cospim32);
+    u[6] = _mm_sub_epi32(u[6], x);
+    u[6] = _mm_add_epi32(u[6], rnding);
+    u[6] = _mm_srai_epi32(u[6], bit);
+
+    u[7] = v[7];
+    u[8] = _mm_add_epi32(v[8], v[11]);
+    u[11] = _mm_sub_epi32(v[8], v[11]);
+    u[9] = _mm_add_epi32(v[9], v[10]);
+    u[10] = _mm_sub_epi32(v[9], v[10]);
+    u[12] = _mm_sub_epi32(v[15], v[12]);
+    u[15] = _mm_add_epi32(v[15], v[12]);
+    u[13] = _mm_sub_epi32(v[14], v[13]);
+    u[14] = _mm_add_epi32(v[14], v[13]);
+
+    // stage 4
+    u[0] = _mm_mullo_epi32(u[0], cospi32);
+    u[1] = _mm_mullo_epi32(u[1], cospi32);
+    v[0] = _mm_add_epi32(u[0], u[1]);
+    v[0] = _mm_add_epi32(v[0], rnding);
+    v[0] = _mm_srai_epi32(v[0], bit);
+
+    v[1] = _mm_sub_epi32(u[0], u[1]);
+    v[1] = _mm_add_epi32(v[1], rnding);
+    v[1] = _mm_srai_epi32(v[1], bit);
+
+    v[2] = _mm_mullo_epi32(u[2], cospi48);
+    x = _mm_mullo_epi32(u[3], cospi16);
+    v[2] = _mm_add_epi32(v[2], x);
+    v[2] = _mm_add_epi32(v[2], rnding);
+    v[2] = _mm_srai_epi32(v[2], bit);
+
+    v[3] = _mm_mullo_epi32(u[2], cospi16);
+    x = _mm_mullo_epi32(u[3], cospi48);
+    v[3] = _mm_sub_epi32(x, v[3]);
+    v[3] = _mm_add_epi32(v[3], rnding);
+    v[3] = _mm_srai_epi32(v[3], bit);
+
+    v[4] = _mm_add_epi32(u[4], u[5]);
+    v[5] = _mm_sub_epi32(u[4], u[5]);
+    v[6] = _mm_sub_epi32(u[7], u[6]);
+    v[7] = _mm_add_epi32(u[7], u[6]);
+    v[8] = u[8];
+
+    v[9] = _mm_mullo_epi32(u[9], cospim16);
+    x = _mm_mullo_epi32(u[14], cospi48);
+    v[9] = _mm_add_epi32(v[9], x);
+    v[9] = _mm_add_epi32(v[9], rnding);
+    v[9] = _mm_srai_epi32(v[9], bit);
+
+    v[14] = _mm_mullo_epi32(u[9], cospi48);
+    x = _mm_mullo_epi32(u[14], cospim16);
+    v[14] = _mm_sub_epi32(v[14], x);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[10] = _mm_mullo_epi32(u[10], cospim48);
+    x = _mm_mullo_epi32(u[13], cospim16);
+    v[10] = _mm_add_epi32(v[10], x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[13] = _mm_mullo_epi32(u[10], cospim16);
+    x = _mm_mullo_epi32(u[13], cospim48);
+    v[13] = _mm_sub_epi32(v[13], x);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    v[11] = u[11];
+    v[12] = u[12];
+    v[15] = u[15];
+
+    // stage 5
+    u[0] = v[0];
+    u[1] = v[1];
+    u[2] = v[2];
+    u[3] = v[3];
+
+    u[4] = _mm_mullo_epi32(v[4], cospi56);
+    x = _mm_mullo_epi32(v[7], cospi8);
+    u[4] = _mm_add_epi32(u[4], x);
+    u[4] = _mm_add_epi32(u[4], rnding);
+    u[4] = _mm_srai_epi32(u[4], bit);
+
+    u[7] = _mm_mullo_epi32(v[4], cospi8);
+    x = _mm_mullo_epi32(v[7], cospi56);
+    u[7] = _mm_sub_epi32(x, u[7]);
+    u[7] = _mm_add_epi32(u[7], rnding);
+    u[7] = _mm_srai_epi32(u[7], bit);
+
+    u[5] = _mm_mullo_epi32(v[5], cospi24);
+    x = _mm_mullo_epi32(v[6], cospi40);
+    u[5] = _mm_add_epi32(u[5], x);
+    u[5] = _mm_add_epi32(u[5], rnding);
+    u[5] = _mm_srai_epi32(u[5], bit);
+
+    u[6] = _mm_mullo_epi32(v[5], cospi40);
+    x = _mm_mullo_epi32(v[6], cospi24);
+    u[6] = _mm_sub_epi32(x, u[6]);
+    u[6] = _mm_add_epi32(u[6], rnding);
+    u[6] = _mm_srai_epi32(u[6], bit);
+
+    u[8] = _mm_add_epi32(v[8], v[9]);
+    u[9] = _mm_sub_epi32(v[8], v[9]);
+    u[10] = _mm_sub_epi32(v[11], v[10]);
+    u[11] = _mm_add_epi32(v[11], v[10]);
+    u[12] = _mm_add_epi32(v[12], v[13]);
+    u[13] = _mm_sub_epi32(v[12], v[13]);
+    u[14] = _mm_sub_epi32(v[15], v[14]);
+    u[15] = _mm_add_epi32(v[15], v[14]);
+
+    // stage 6
+    v[0] = u[0];
+    v[1] = u[1];
+    v[2] = u[2];
+    v[3] = u[3];
+    v[4] = u[4];
+    v[5] = u[5];
+    v[6] = u[6];
+    v[7] = u[7];
+
+    v[8] = _mm_mullo_epi32(u[8], cospi60);
+    x = _mm_mullo_epi32(u[15], cospi4);
+    v[8] = _mm_add_epi32(v[8], x);
+    v[8] = _mm_add_epi32(v[8], rnding);
+    v[8] = _mm_srai_epi32(v[8], bit);
+
+    v[15] = _mm_mullo_epi32(u[8], cospi4);
+    x = _mm_mullo_epi32(u[15], cospi60);
+    v[15] = _mm_sub_epi32(x, v[15]);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    v[9] = _mm_mullo_epi32(u[9], cospi28);
+    x = _mm_mullo_epi32(u[14], cospi36);
+    v[9] = _mm_add_epi32(v[9], x);
+    v[9] = _mm_add_epi32(v[9], rnding);
+    v[9] = _mm_srai_epi32(v[9], bit);
+
+    v[14] = _mm_mullo_epi32(u[9], cospi36);
+    x = _mm_mullo_epi32(u[14], cospi28);
+    v[14] = _mm_sub_epi32(x, v[14]);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[10] = _mm_mullo_epi32(u[10], cospi44);
+    x = _mm_mullo_epi32(u[13], cospi20);
+    v[10] = _mm_add_epi32(v[10], x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[13] = _mm_mullo_epi32(u[10], cospi20);
+    x = _mm_mullo_epi32(u[13], cospi44);
+    v[13] = _mm_sub_epi32(x, v[13]);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    v[11] = _mm_mullo_epi32(u[11], cospi12);
+    x = _mm_mullo_epi32(u[12], cospi52);
+    v[11] = _mm_add_epi32(v[11], x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = _mm_mullo_epi32(u[11], cospi52);
+    x = _mm_mullo_epi32(u[12], cospi12);
+    v[12] = _mm_sub_epi32(x, v[12]);
+    v[12] = _mm_add_epi32(v[12], rnding);
+    v[12] = _mm_srai_epi32(v[12], bit);
+
+    out[0 * col_num + col] = v[0];
+    out[1 * col_num + col] = v[8];
+    out[2 * col_num + col] = v[4];
+    out[3 * col_num + col] = v[12];
+    out[4 * col_num + col] = v[2];
+    out[5 * col_num + col] = v[10];
+    out[6 * col_num + col] = v[6];
+    out[7 * col_num + col] = v[14];
+    out[8 * col_num + col] = v[1];
+    out[9 * col_num + col] = v[9];
+    out[10 * col_num + col] = v[5];
+    out[11 * col_num + col] = v[13];
+    out[12 * col_num + col] = v[3];
+    out[13 * col_num + col] = v[11];
+    out[14 * col_num + col] = v[7];
+    out[15 * col_num + col] = v[15];
+  }
+}
+
+static void fadst16x16_sse4_1(__m128i *in, __m128i *out, int bit) {
+  const int32_t *cospi = cospi_arr[bit - cos_bit_min];
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+  const __m128i cospi10 = _mm_set1_epi32(cospi[10]);
+  const __m128i cospi54 = _mm_set1_epi32(cospi[54]);
+  const __m128i cospi18 = _mm_set1_epi32(cospi[18]);
+  const __m128i cospi46 = _mm_set1_epi32(cospi[46]);
+  const __m128i cospi26 = _mm_set1_epi32(cospi[26]);
+  const __m128i cospi38 = _mm_set1_epi32(cospi[38]);
+  const __m128i cospi34 = _mm_set1_epi32(cospi[34]);
+  const __m128i cospi30 = _mm_set1_epi32(cospi[30]);
+  const __m128i cospi42 = _mm_set1_epi32(cospi[42]);
+  const __m128i cospi22 = _mm_set1_epi32(cospi[22]);
+  const __m128i cospi50 = _mm_set1_epi32(cospi[50]);
+  const __m128i cospi14 = _mm_set1_epi32(cospi[14]);
+  const __m128i cospi58 = _mm_set1_epi32(cospi[58]);
+  const __m128i cospi6 = _mm_set1_epi32(cospi[6]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospim56 = _mm_set1_epi32(-cospi[56]);
+  const __m128i cospim24 = _mm_set1_epi32(-cospi[24]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  __m128i u[16], v[16], x, y;
+  const int col_num = 4;
+  int col;
+
+  // Calculate the column 0, 1, 2, 3
+  for (col = 0; col < col_num; ++col) {
+    // stage 0
+    // stage 1
+    // stage 2
+    v[0] = _mm_mullo_epi32(in[15 * col_num + col], cospi2);
+    x = _mm_mullo_epi32(in[0 * col_num + col], cospi62);
+    v[0] = _mm_add_epi32(v[0], x);
+    v[0] = _mm_add_epi32(v[0], rnding);
+    v[0] = _mm_srai_epi32(v[0], bit);
+
+    v[1] = _mm_mullo_epi32(in[15 * col_num + col], cospi62);
+    x = _mm_mullo_epi32(in[0 * col_num + col], cospi2);
+    v[1] = _mm_sub_epi32(v[1], x);
+    v[1] = _mm_add_epi32(v[1], rnding);
+    v[1] = _mm_srai_epi32(v[1], bit);
+
+    v[2] = _mm_mullo_epi32(in[13 * col_num + col], cospi10);
+    x = _mm_mullo_epi32(in[2 * col_num + col], cospi54);
+    v[2] = _mm_add_epi32(v[2], x);
+    v[2] = _mm_add_epi32(v[2], rnding);
+    v[2] = _mm_srai_epi32(v[2], bit);
+
+    v[3] = _mm_mullo_epi32(in[13 * col_num + col], cospi54);
+    x = _mm_mullo_epi32(in[2 * col_num + col], cospi10);
+    v[3] = _mm_sub_epi32(v[3], x);
+    v[3] = _mm_add_epi32(v[3], rnding);
+    v[3] = _mm_srai_epi32(v[3], bit);
+
+    v[4] = _mm_mullo_epi32(in[11 * col_num + col], cospi18);
+    x = _mm_mullo_epi32(in[4 * col_num + col], cospi46);
+    v[4] = _mm_add_epi32(v[4], x);
+    v[4] = _mm_add_epi32(v[4], rnding);
+    v[4] = _mm_srai_epi32(v[4], bit);
+
+    v[5] = _mm_mullo_epi32(in[11 * col_num + col], cospi46);
+    x = _mm_mullo_epi32(in[4 * col_num + col], cospi18);
+    v[5] = _mm_sub_epi32(v[5], x);
+    v[5] = _mm_add_epi32(v[5], rnding);
+    v[5] = _mm_srai_epi32(v[5], bit);
+
+    v[6] = _mm_mullo_epi32(in[9 * col_num + col], cospi26);
+    x = _mm_mullo_epi32(in[6 * col_num + col], cospi38);
+    v[6] = _mm_add_epi32(v[6], x);
+    v[6] = _mm_add_epi32(v[6], rnding);
+    v[6] = _mm_srai_epi32(v[6], bit);
+
+    v[7] = _mm_mullo_epi32(in[9 * col_num + col], cospi38);
+    x = _mm_mullo_epi32(in[6 * col_num + col], cospi26);
+    v[7] = _mm_sub_epi32(v[7], x);
+    v[7] = _mm_add_epi32(v[7], rnding);
+    v[7] = _mm_srai_epi32(v[7], bit);
+
+    v[8] = _mm_mullo_epi32(in[7 * col_num + col], cospi34);
+    x = _mm_mullo_epi32(in[8 * col_num + col], cospi30);
+    v[8] = _mm_add_epi32(v[8], x);
+    v[8] = _mm_add_epi32(v[8], rnding);
+    v[8] = _mm_srai_epi32(v[8], bit);
+
+    v[9] = _mm_mullo_epi32(in[7 * col_num + col], cospi30);
+    x = _mm_mullo_epi32(in[8 * col_num + col], cospi34);
+    v[9] = _mm_sub_epi32(v[9], x);
+    v[9] = _mm_add_epi32(v[9], rnding);
+    v[9] = _mm_srai_epi32(v[9], bit);
+
+    v[10] = _mm_mullo_epi32(in[5 * col_num + col], cospi42);
+    x = _mm_mullo_epi32(in[10 * col_num + col], cospi22);
+    v[10] = _mm_add_epi32(v[10], x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[11] = _mm_mullo_epi32(in[5 * col_num + col], cospi22);
+    x = _mm_mullo_epi32(in[10 * col_num + col], cospi42);
+    v[11] = _mm_sub_epi32(v[11], x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = _mm_mullo_epi32(in[3 * col_num + col], cospi50);
+    x = _mm_mullo_epi32(in[12 * col_num + col], cospi14);
+    v[12] = _mm_add_epi32(v[12], x);
+    v[12] = _mm_add_epi32(v[12], rnding);
+    v[12] = _mm_srai_epi32(v[12], bit);
+
+    v[13] = _mm_mullo_epi32(in[3 * col_num + col], cospi14);
+    x = _mm_mullo_epi32(in[12 * col_num + col], cospi50);
+    v[13] = _mm_sub_epi32(v[13], x);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    v[14] = _mm_mullo_epi32(in[1 * col_num + col], cospi58);
+    x = _mm_mullo_epi32(in[14 * col_num + col], cospi6);
+    v[14] = _mm_add_epi32(v[14], x);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[15] = _mm_mullo_epi32(in[1 * col_num + col], cospi6);
+    x = _mm_mullo_epi32(in[14 * col_num + col], cospi58);
+    v[15] = _mm_sub_epi32(v[15], x);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    // stage 3
+    u[0] = _mm_add_epi32(v[0], v[8]);
+    u[8] = _mm_sub_epi32(v[0], v[8]);
+    u[1] = _mm_add_epi32(v[1], v[9]);
+    u[9] = _mm_sub_epi32(v[1], v[9]);
+    u[2] = _mm_add_epi32(v[2], v[10]);
+    u[10] = _mm_sub_epi32(v[2], v[10]);
+    u[3] = _mm_add_epi32(v[3], v[11]);
+    u[11] = _mm_sub_epi32(v[3], v[11]);
+    u[4] = _mm_add_epi32(v[4], v[12]);
+    u[12] = _mm_sub_epi32(v[4], v[12]);
+    u[5] = _mm_add_epi32(v[5], v[13]);
+    u[13] = _mm_sub_epi32(v[5], v[13]);
+    u[6] = _mm_add_epi32(v[6], v[14]);
+    u[14] = _mm_sub_epi32(v[6], v[14]);
+    u[7] = _mm_add_epi32(v[7], v[15]);
+    u[15] = _mm_sub_epi32(v[7], v[15]);
+
+    // stage 4
+    v[0] = u[0];
+    v[1] = u[1];
+    v[2] = u[2];
+    v[3] = u[3];
+    v[4] = u[4];
+    v[5] = u[5];
+    v[6] = u[6];
+    v[7] = u[7];
+
+    v[8] = _mm_mullo_epi32(u[8], cospi8);
+    x = _mm_mullo_epi32(u[9], cospi56);
+    v[8] = _mm_add_epi32(v[8], x);
+    v[8] = _mm_add_epi32(v[8], rnding);
+    v[8] = _mm_srai_epi32(v[8], bit);
+
+    v[9] = _mm_mullo_epi32(u[8], cospi56);
+    x = _mm_mullo_epi32(u[9], cospi8);
+    v[9] = _mm_sub_epi32(v[9], x);
+    v[9] = _mm_add_epi32(v[9], rnding);
+    v[9] = _mm_srai_epi32(v[9], bit);
+
+    v[10] = _mm_mullo_epi32(u[10], cospi40);
+    x = _mm_mullo_epi32(u[11], cospi24);
+    v[10] = _mm_add_epi32(v[10], x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[11] = _mm_mullo_epi32(u[10], cospi24);
+    x = _mm_mullo_epi32(u[11], cospi40);
+    v[11] = _mm_sub_epi32(v[11], x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = _mm_mullo_epi32(u[12], cospim56);
+    x = _mm_mullo_epi32(u[13], cospi8);
+    v[12] = _mm_add_epi32(v[12], x);
+    v[12] = _mm_add_epi32(v[12], rnding);
+    v[12] = _mm_srai_epi32(v[12], bit);
+
+    v[13] = _mm_mullo_epi32(u[12], cospi8);
+    x = _mm_mullo_epi32(u[13], cospim56);
+    v[13] = _mm_sub_epi32(v[13], x);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    v[14] = _mm_mullo_epi32(u[14], cospim24);
+    x = _mm_mullo_epi32(u[15], cospi40);
+    v[14] = _mm_add_epi32(v[14], x);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[15] = _mm_mullo_epi32(u[14], cospi40);
+    x = _mm_mullo_epi32(u[15], cospim24);
+    v[15] = _mm_sub_epi32(v[15], x);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    // stage 5
+    u[0] = _mm_add_epi32(v[0], v[4]);
+    u[4] = _mm_sub_epi32(v[0], v[4]);
+    u[1] = _mm_add_epi32(v[1], v[5]);
+    u[5] = _mm_sub_epi32(v[1], v[5]);
+    u[2] = _mm_add_epi32(v[2], v[6]);
+    u[6] = _mm_sub_epi32(v[2], v[6]);
+    u[3] = _mm_add_epi32(v[3], v[7]);
+    u[7] = _mm_sub_epi32(v[3], v[7]);
+    u[8] = _mm_add_epi32(v[8], v[12]);
+    u[12] = _mm_sub_epi32(v[8], v[12]);
+    u[9] = _mm_add_epi32(v[9], v[13]);
+    u[13] = _mm_sub_epi32(v[9], v[13]);
+    u[10] = _mm_add_epi32(v[10], v[14]);
+    u[14] = _mm_sub_epi32(v[10], v[14]);
+    u[11] = _mm_add_epi32(v[11], v[15]);
+    u[15] = _mm_sub_epi32(v[11], v[15]);
+
+    // stage 6
+    v[0] = u[0];
+    v[1] = u[1];
+    v[2] = u[2];
+    v[3] = u[3];
+
+    v[4] = _mm_mullo_epi32(u[4], cospi16);
+    x = _mm_mullo_epi32(u[5], cospi48);
+    v[4] = _mm_add_epi32(v[4], x);
+    v[4] = _mm_add_epi32(v[4], rnding);
+    v[4] = _mm_srai_epi32(v[4], bit);
+
+    v[5] = _mm_mullo_epi32(u[4], cospi48);
+    x = _mm_mullo_epi32(u[5], cospi16);
+    v[5] = _mm_sub_epi32(v[5], x);
+    v[5] = _mm_add_epi32(v[5], rnding);
+    v[5] = _mm_srai_epi32(v[5], bit);
+
+    v[6] = _mm_mullo_epi32(u[6], cospim48);
+    x = _mm_mullo_epi32(u[7], cospi16);
+    v[6] = _mm_add_epi32(v[6], x);
+    v[6] = _mm_add_epi32(v[6], rnding);
+    v[6] = _mm_srai_epi32(v[6], bit);
+
+    v[7] = _mm_mullo_epi32(u[6], cospi16);
+    x = _mm_mullo_epi32(u[7], cospim48);
+    v[7] = _mm_sub_epi32(v[7], x);
+    v[7] = _mm_add_epi32(v[7], rnding);
+    v[7] = _mm_srai_epi32(v[7], bit);
+
+    v[8] = u[8];
+    v[9] = u[9];
+    v[10] = u[10];
+    v[11] = u[11];
+
+    v[12] = _mm_mullo_epi32(u[12], cospi16);
+    x = _mm_mullo_epi32(u[13], cospi48);
+    v[12] = _mm_add_epi32(v[12], x);
+    v[12] = _mm_add_epi32(v[12], rnding);
+    v[12] = _mm_srai_epi32(v[12], bit);
+
+    v[13] = _mm_mullo_epi32(u[12], cospi48);
+    x = _mm_mullo_epi32(u[13], cospi16);
+    v[13] = _mm_sub_epi32(v[13], x);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    v[14] = _mm_mullo_epi32(u[14], cospim48);
+    x = _mm_mullo_epi32(u[15], cospi16);
+    v[14] = _mm_add_epi32(v[14], x);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[15] = _mm_mullo_epi32(u[14], cospi16);
+    x = _mm_mullo_epi32(u[15], cospim48);
+    v[15] = _mm_sub_epi32(v[15], x);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    // stage 7
+    u[0] = _mm_add_epi32(v[0], v[2]);
+    u[2] = _mm_sub_epi32(v[0], v[2]);
+    u[1] = _mm_add_epi32(v[1], v[3]);
+    u[3] = _mm_sub_epi32(v[1], v[3]);
+    u[4] = _mm_add_epi32(v[4], v[6]);
+    u[6] = _mm_sub_epi32(v[4], v[6]);
+    u[5] = _mm_add_epi32(v[5], v[7]);
+    u[7] = _mm_sub_epi32(v[5], v[7]);
+    u[8] = _mm_add_epi32(v[8], v[10]);
+    u[10] = _mm_sub_epi32(v[8], v[10]);
+    u[9] = _mm_add_epi32(v[9], v[11]);
+    u[11] = _mm_sub_epi32(v[9], v[11]);
+    u[12] = _mm_add_epi32(v[12], v[14]);
+    u[14] = _mm_sub_epi32(v[12], v[14]);
+    u[13] = _mm_add_epi32(v[13], v[15]);
+    u[15] = _mm_sub_epi32(v[13], v[15]);
+
+    // stage 8
+    v[0] = u[0];
+    v[1] = u[1];
+
+    y = _mm_mullo_epi32(u[2], cospi32);
+    x = _mm_mullo_epi32(u[3], cospi32);
+    v[2] = _mm_add_epi32(y, x);
+    v[2] = _mm_add_epi32(v[2], rnding);
+    v[2] = _mm_srai_epi32(v[2], bit);
+
+    v[3] = _mm_sub_epi32(y, x);
+    v[3] = _mm_add_epi32(v[3], rnding);
+    v[3] = _mm_srai_epi32(v[3], bit);
+
+    v[4] = u[4];
+    v[5] = u[5];
+
+    y = _mm_mullo_epi32(u[6], cospi32);
+    x = _mm_mullo_epi32(u[7], cospi32);
+    v[6] = _mm_add_epi32(y, x);
+    v[6] = _mm_add_epi32(v[6], rnding);
+    v[6] = _mm_srai_epi32(v[6], bit);
+
+    v[7] = _mm_sub_epi32(y, x);
+    v[7] = _mm_add_epi32(v[7], rnding);
+    v[7] = _mm_srai_epi32(v[7], bit);
+
+    v[8] = u[8];
+    v[9] = u[9];
+
+    y = _mm_mullo_epi32(u[10], cospi32);
+    x = _mm_mullo_epi32(u[11], cospi32);
+    v[10] = _mm_add_epi32(y, x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[11] = _mm_sub_epi32(y, x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = u[12];
+    v[13] = u[13];
+
+    y = _mm_mullo_epi32(u[14], cospi32);
+    x = _mm_mullo_epi32(u[15], cospi32);
+    v[14] = _mm_add_epi32(y, x);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[15] = _mm_sub_epi32(y, x);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    // stage 9
+    out[0 * col_num + col] = v[0];
+    out[1 * col_num + col] = _mm_sub_epi32(_mm_set1_epi32(0), v[8]);
+    out[2 * col_num + col] = v[12];
+    out[3 * col_num + col] = _mm_sub_epi32(_mm_set1_epi32(0), v[4]);
+    out[4 * col_num + col] = v[6];
+    out[5 * col_num + col] = _mm_sub_epi32(_mm_set1_epi32(0), v[14]);
+    out[6 * col_num + col] = v[10];
+    out[7 * col_num + col] = _mm_sub_epi32(_mm_set1_epi32(0), v[2]);
+    out[8 * col_num + col] = v[3];
+    out[9 * col_num + col] = _mm_sub_epi32(_mm_set1_epi32(0), v[11]);
+    out[10 * col_num + col] = v[15];
+    out[11 * col_num + col] = _mm_sub_epi32(_mm_set1_epi32(0), v[7]);
+    out[12 * col_num + col] = v[5];
+    out[13 * col_num + col] = _mm_sub_epi32(_mm_set1_epi32(0), v[13]);
+    out[14 * col_num + col] = v[9];
+    out[15 * col_num + col] = _mm_sub_epi32(_mm_set1_epi32(0), v[1]);
+  }
+}
+
+static void col_txfm_16x16_rounding(__m128i *in, int shift) {
+  // Note:
+  //  We split 16x16 rounding into 4 sections of 8x8 rounding,
+  //  instead of 4 columns
+  col_txfm_8x8_rounding(&in[0], shift);
+  col_txfm_8x8_rounding(&in[16], shift);
+  col_txfm_8x8_rounding(&in[32], shift);
+  col_txfm_8x8_rounding(&in[48], shift);
+}
+
+static void write_buffer_16x16(const __m128i *in, tran_low_t *output) {
+  const int size_8x8 = 16 * 4;
+  write_buffer_8x8(&in[0], output);
+  output += size_8x8;
+  write_buffer_8x8(&in[16], output);
+  output += size_8x8;
+  write_buffer_8x8(&in[32], output);
+  output += size_8x8;
+  write_buffer_8x8(&in[48], output);
+}
+
+void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *coeff,
+                                 int stride, int tx_type, int bd) {
+  __m128i in[64], out[64];
+  const TXFM_2D_CFG *cfg = NULL;
+
+  switch (tx_type) {
+    case DCT_DCT:
+      cfg = &fwd_txfm_2d_cfg_dct_dct_16;
+      load_buffer_16x16(input, in, stride, 0, 0, cfg->shift[0]);
+      fdct16x16_sse4_1(in, out, cfg->cos_bit_col[0]);
+      col_txfm_16x16_rounding(out, -cfg->shift[1]);
+      transpose_16x16(out, in);
+      fdct16x16_sse4_1(in, out, cfg->cos_bit_row[0]);
+      transpose_16x16(out, in);
+      write_buffer_16x16(in, coeff);
+      break;
+    case ADST_DCT:
+      cfg = &fwd_txfm_2d_cfg_adst_dct_16;
+      load_buffer_16x16(input, in, stride, 0, 0, cfg->shift[0]);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_col[0]);
+      col_txfm_16x16_rounding(out, -cfg->shift[1]);
+      transpose_16x16(out, in);
+      fdct16x16_sse4_1(in, out, cfg->cos_bit_row[0]);
+      transpose_16x16(out, in);
+      write_buffer_16x16(in, coeff);
+      break;
+    case DCT_ADST:
+      cfg = &fwd_txfm_2d_cfg_dct_adst_16;
+      load_buffer_16x16(input, in, stride, 0, 0, cfg->shift[0]);
+      fdct16x16_sse4_1(in, out, cfg->cos_bit_col[0]);
+      col_txfm_16x16_rounding(out, -cfg->shift[1]);
+      transpose_16x16(out, in);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_row[0]);
+      transpose_16x16(out, in);
+      write_buffer_16x16(in, coeff);
+      break;
+    case ADST_ADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_16;
+      load_buffer_16x16(input, in, stride, 0, 0, cfg->shift[0]);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_col[0]);
+      col_txfm_16x16_rounding(out, -cfg->shift[1]);
+      transpose_16x16(out, in);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_row[0]);
+      transpose_16x16(out, in);
+      write_buffer_16x16(in, coeff);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      cfg = &fwd_txfm_2d_cfg_adst_dct_16;
+      load_buffer_16x16(input, in, stride, 1, 0, cfg->shift[0]);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_col[0]);
+      col_txfm_16x16_rounding(out, -cfg->shift[1]);
+      transpose_16x16(out, in);
+      fdct16x16_sse4_1(in, out, cfg->cos_bit_row[0]);
+      transpose_16x16(out, in);
+      write_buffer_16x16(in, coeff);
+      break;
+    case DCT_FLIPADST:
+      cfg = &fwd_txfm_2d_cfg_dct_adst_16;
+      load_buffer_16x16(input, in, stride, 0, 1, cfg->shift[0]);
+      fdct16x16_sse4_1(in, out, cfg->cos_bit_col[0]);
+      col_txfm_16x16_rounding(out, -cfg->shift[1]);
+      transpose_16x16(out, in);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_row[0]);
+      transpose_16x16(out, in);
+      write_buffer_16x16(in, coeff);
+      break;
+    case FLIPADST_FLIPADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_16;
+      load_buffer_16x16(input, in, stride, 1, 1, cfg->shift[0]);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_col[0]);
+      col_txfm_16x16_rounding(out, -cfg->shift[1]);
+      transpose_16x16(out, in);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_row[0]);
+      transpose_16x16(out, in);
+      write_buffer_16x16(in, coeff);
+      break;
+    case ADST_FLIPADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_16;
+      load_buffer_16x16(input, in, stride, 0, 1, cfg->shift[0]);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_col[0]);
+      col_txfm_16x16_rounding(out, -cfg->shift[1]);
+      transpose_16x16(out, in);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_row[0]);
+      transpose_16x16(out, in);
+      write_buffer_16x16(in, coeff);
+      break;
+    case FLIPADST_ADST:
+      cfg = &fwd_txfm_2d_cfg_adst_adst_16;
+      load_buffer_16x16(input, in, stride, 1, 0, cfg->shift[0]);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_col[0]);
+      col_txfm_16x16_rounding(out, -cfg->shift[1]);
+      transpose_16x16(out, in);
+      fadst16x16_sse4_1(in, out, cfg->cos_bit_row[0]);
+      transpose_16x16(out, in);
+      write_buffer_16x16(in, coeff);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0);
+  }
+  (void)bd;
+}
diff --git a/third_party/aom/av1/encoder/x86/hybrid_fwd_txfm_avx2.c b/third_party/aom/av1/encoder/x86/hybrid_fwd_txfm_avx2.c
new file mode 100644
index 0000000000..198e4e4c4c
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/hybrid_fwd_txfm_avx2.c
@@ -0,0 +1,1678 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>  // avx2
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "aom_dsp/x86/fwd_txfm_avx2.h"
+#include "aom_dsp/txfm_common.h"
+#include "aom_dsp/x86/txfm_common_avx2.h"
+
+static int32_t get_16x16_sum(const int16_t *input, int stride) {
+  __m256i r0, r1, r2, r3, u0, u1;
+  __m256i zero = _mm256_setzero_si256();
+  __m256i sum = _mm256_setzero_si256();
+  const int16_t *blockBound = input + (stride << 4);
+  __m128i v0, v1;
+
+  while (input < blockBound) {
+    r0 = _mm256_loadu_si256((__m256i const *)input);
+    r1 = _mm256_loadu_si256((__m256i const *)(input + stride));
+    r2 = _mm256_loadu_si256((__m256i const *)(input + 2 * stride));
+    r3 = _mm256_loadu_si256((__m256i const *)(input + 3 * stride));
+
+    u0 = _mm256_add_epi16(r0, r1);
+    u1 = _mm256_add_epi16(r2, r3);
+    sum = _mm256_add_epi16(sum, u0);
+    sum = _mm256_add_epi16(sum, u1);
+
+    input += stride << 2;
+  }
+
+  // unpack 16 int16_t into 2x8 int32_t
+  u0 = _mm256_unpacklo_epi16(zero, sum);
+  u1 = _mm256_unpackhi_epi16(zero, sum);
+  u0 = _mm256_srai_epi32(u0, 16);
+  u1 = _mm256_srai_epi32(u1, 16);
+  sum = _mm256_add_epi32(u0, u1);
+
+  u0 = _mm256_srli_si256(sum, 8);
+  u1 = _mm256_add_epi32(sum, u0);
+
+  v0 = _mm_add_epi32(_mm256_extracti128_si256(u1, 1),
+                     _mm256_castsi256_si128(u1));
+  v1 = _mm_srli_si128(v0, 4);
+  v0 = _mm_add_epi32(v0, v1);
+  return (int32_t)_mm_extract_epi32(v0, 0);
+}
+
+void aom_fdct16x16_1_avx2(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  int32_t dc = get_16x16_sum(input, stride);
+  output[0] = (tran_low_t)(dc >> 1);
+  _mm256_zeroupper();
+}
+
+static INLINE void load_buffer_16x16(const int16_t *input, int stride,
+                                     int flipud, int fliplr, __m256i *in) {
+  if (!flipud) {
+    in[0] = _mm256_loadu_si256((const __m256i *)(input + 0 * stride));
+    in[1] = _mm256_loadu_si256((const __m256i *)(input + 1 * stride));
+    in[2] = _mm256_loadu_si256((const __m256i *)(input + 2 * stride));
+    in[3] = _mm256_loadu_si256((const __m256i *)(input + 3 * stride));
+    in[4] = _mm256_loadu_si256((const __m256i *)(input + 4 * stride));
+    in[5] = _mm256_loadu_si256((const __m256i *)(input + 5 * stride));
+    in[6] = _mm256_loadu_si256((const __m256i *)(input + 6 * stride));
+    in[7] = _mm256_loadu_si256((const __m256i *)(input + 7 * stride));
+    in[8] = _mm256_loadu_si256((const __m256i *)(input + 8 * stride));
+    in[9] = _mm256_loadu_si256((const __m256i *)(input + 9 * stride));
+    in[10] = _mm256_loadu_si256((const __m256i *)(input + 10 * stride));
+    in[11] = _mm256_loadu_si256((const __m256i *)(input + 11 * stride));
+    in[12] = _mm256_loadu_si256((const __m256i *)(input + 12 * stride));
+    in[13] = _mm256_loadu_si256((const __m256i *)(input + 13 * stride));
+    in[14] = _mm256_loadu_si256((const __m256i *)(input + 14 * stride));
+    in[15] = _mm256_loadu_si256((const __m256i *)(input + 15 * stride));
+  } else {
+    in[0] = _mm256_loadu_si256((const __m256i *)(input + 15 * stride));
+    in[1] = _mm256_loadu_si256((const __m256i *)(input + 14 * stride));
+    in[2] = _mm256_loadu_si256((const __m256i *)(input + 13 * stride));
+    in[3] = _mm256_loadu_si256((const __m256i *)(input + 12 * stride));
+    in[4] = _mm256_loadu_si256((const __m256i *)(input + 11 * stride));
+    in[5] = _mm256_loadu_si256((const __m256i *)(input + 10 * stride));
+    in[6] = _mm256_loadu_si256((const __m256i *)(input + 9 * stride));
+    in[7] = _mm256_loadu_si256((const __m256i *)(input + 8 * stride));
+    in[8] = _mm256_loadu_si256((const __m256i *)(input + 7 * stride));
+    in[9] = _mm256_loadu_si256((const __m256i *)(input + 6 * stride));
+    in[10] = _mm256_loadu_si256((const __m256i *)(input + 5 * stride));
+    in[11] = _mm256_loadu_si256((const __m256i *)(input + 4 * stride));
+    in[12] = _mm256_loadu_si256((const __m256i *)(input + 3 * stride));
+    in[13] = _mm256_loadu_si256((const __m256i *)(input + 2 * stride));
+    in[14] = _mm256_loadu_si256((const __m256i *)(input + 1 * stride));
+    in[15] = _mm256_loadu_si256((const __m256i *)(input + 0 * stride));
+  }
+
+  if (fliplr) {
+    mm256_reverse_epi16(&in[0]);
+    mm256_reverse_epi16(&in[1]);
+    mm256_reverse_epi16(&in[2]);
+    mm256_reverse_epi16(&in[3]);
+    mm256_reverse_epi16(&in[4]);
+    mm256_reverse_epi16(&in[5]);
+    mm256_reverse_epi16(&in[6]);
+    mm256_reverse_epi16(&in[7]);
+    mm256_reverse_epi16(&in[8]);
+    mm256_reverse_epi16(&in[9]);
+    mm256_reverse_epi16(&in[10]);
+    mm256_reverse_epi16(&in[11]);
+    mm256_reverse_epi16(&in[12]);
+    mm256_reverse_epi16(&in[13]);
+    mm256_reverse_epi16(&in[14]);
+    mm256_reverse_epi16(&in[15]);
+  }
+
+  in[0] = _mm256_slli_epi16(in[0], 2);
+  in[1] = _mm256_slli_epi16(in[1], 2);
+  in[2] = _mm256_slli_epi16(in[2], 2);
+  in[3] = _mm256_slli_epi16(in[3], 2);
+  in[4] = _mm256_slli_epi16(in[4], 2);
+  in[5] = _mm256_slli_epi16(in[5], 2);
+  in[6] = _mm256_slli_epi16(in[6], 2);
+  in[7] = _mm256_slli_epi16(in[7], 2);
+  in[8] = _mm256_slli_epi16(in[8], 2);
+  in[9] = _mm256_slli_epi16(in[9], 2);
+  in[10] = _mm256_slli_epi16(in[10], 2);
+  in[11] = _mm256_slli_epi16(in[11], 2);
+  in[12] = _mm256_slli_epi16(in[12], 2);
+  in[13] = _mm256_slli_epi16(in[13], 2);
+  in[14] = _mm256_slli_epi16(in[14], 2);
+  in[15] = _mm256_slli_epi16(in[15], 2);
+}
+
+static INLINE void write_buffer_16x16(const __m256i *in, tran_low_t *output) {
+  int i;
+  for (i = 0; i < 16; ++i) {
+    storeu_output_avx2(&in[i], output + (i << 4));
+  }
+}
+
+static void right_shift_16x16(__m256i *in) {
+  const __m256i one = _mm256_set1_epi16(1);
+  __m256i s0 = _mm256_srai_epi16(in[0], 15);
+  __m256i s1 = _mm256_srai_epi16(in[1], 15);
+  __m256i s2 = _mm256_srai_epi16(in[2], 15);
+  __m256i s3 = _mm256_srai_epi16(in[3], 15);
+  __m256i s4 = _mm256_srai_epi16(in[4], 15);
+  __m256i s5 = _mm256_srai_epi16(in[5], 15);
+  __m256i s6 = _mm256_srai_epi16(in[6], 15);
+  __m256i s7 = _mm256_srai_epi16(in[7], 15);
+  __m256i s8 = _mm256_srai_epi16(in[8], 15);
+  __m256i s9 = _mm256_srai_epi16(in[9], 15);
+  __m256i s10 = _mm256_srai_epi16(in[10], 15);
+  __m256i s11 = _mm256_srai_epi16(in[11], 15);
+  __m256i s12 = _mm256_srai_epi16(in[12], 15);
+  __m256i s13 = _mm256_srai_epi16(in[13], 15);
+  __m256i s14 = _mm256_srai_epi16(in[14], 15);
+  __m256i s15 = _mm256_srai_epi16(in[15], 15);
+
+  in[0] = _mm256_add_epi16(in[0], one);
+  in[1] = _mm256_add_epi16(in[1], one);
+  in[2] = _mm256_add_epi16(in[2], one);
+  in[3] = _mm256_add_epi16(in[3], one);
+  in[4] = _mm256_add_epi16(in[4], one);
+  in[5] = _mm256_add_epi16(in[5], one);
+  in[6] = _mm256_add_epi16(in[6], one);
+  in[7] = _mm256_add_epi16(in[7], one);
+  in[8] = _mm256_add_epi16(in[8], one);
+  in[9] = _mm256_add_epi16(in[9], one);
+  in[10] = _mm256_add_epi16(in[10], one);
+  in[11] = _mm256_add_epi16(in[11], one);
+  in[12] = _mm256_add_epi16(in[12], one);
+  in[13] = _mm256_add_epi16(in[13], one);
+  in[14] = _mm256_add_epi16(in[14], one);
+  in[15] = _mm256_add_epi16(in[15], one);
+
+  in[0] = _mm256_sub_epi16(in[0], s0);
+  in[1] = _mm256_sub_epi16(in[1], s1);
+  in[2] = _mm256_sub_epi16(in[2], s2);
+  in[3] = _mm256_sub_epi16(in[3], s3);
+  in[4] = _mm256_sub_epi16(in[4], s4);
+  in[5] = _mm256_sub_epi16(in[5], s5);
+  in[6] = _mm256_sub_epi16(in[6], s6);
+  in[7] = _mm256_sub_epi16(in[7], s7);
+  in[8] = _mm256_sub_epi16(in[8], s8);
+  in[9] = _mm256_sub_epi16(in[9], s9);
+  in[10] = _mm256_sub_epi16(in[10], s10);
+  in[11] = _mm256_sub_epi16(in[11], s11);
+  in[12] = _mm256_sub_epi16(in[12], s12);
+  in[13] = _mm256_sub_epi16(in[13], s13);
+  in[14] = _mm256_sub_epi16(in[14], s14);
+  in[15] = _mm256_sub_epi16(in[15], s15);
+
+  in[0] = _mm256_srai_epi16(in[0], 2);
+  in[1] = _mm256_srai_epi16(in[1], 2);
+  in[2] = _mm256_srai_epi16(in[2], 2);
+  in[3] = _mm256_srai_epi16(in[3], 2);
+  in[4] = _mm256_srai_epi16(in[4], 2);
+  in[5] = _mm256_srai_epi16(in[5], 2);
+  in[6] = _mm256_srai_epi16(in[6], 2);
+  in[7] = _mm256_srai_epi16(in[7], 2);
+  in[8] = _mm256_srai_epi16(in[8], 2);
+  in[9] = _mm256_srai_epi16(in[9], 2);
+  in[10] = _mm256_srai_epi16(in[10], 2);
+  in[11] = _mm256_srai_epi16(in[11], 2);
+  in[12] = _mm256_srai_epi16(in[12], 2);
+  in[13] = _mm256_srai_epi16(in[13], 2);
+  in[14] = _mm256_srai_epi16(in[14], 2);
+  in[15] = _mm256_srai_epi16(in[15], 2);
+}
+
+static void fdct16_avx2(__m256i *in) {
+  // sequence: cospi_L_H = pairs(L, H) and L first
+  const __m256i cospi_p16_m16 = pair256_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m256i cospi_p16_p16 = pair256_set_epi16(cospi_16_64, cospi_16_64);
+  const __m256i cospi_p24_p08 = pair256_set_epi16(cospi_24_64, cospi_8_64);
+  const __m256i cospi_m08_p24 = pair256_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m256i cospi_m24_m08 = pair256_set_epi16(-cospi_24_64, -cospi_8_64);
+
+  const __m256i cospi_p28_p04 = pair256_set_epi16(cospi_28_64, cospi_4_64);
+  const __m256i cospi_m04_p28 = pair256_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m256i cospi_p12_p20 = pair256_set_epi16(cospi_12_64, cospi_20_64);
+  const __m256i cospi_m20_p12 = pair256_set_epi16(-cospi_20_64, cospi_12_64);
+
+  const __m256i cospi_p30_p02 = pair256_set_epi16(cospi_30_64, cospi_2_64);
+  const __m256i cospi_m02_p30 = pair256_set_epi16(-cospi_2_64, cospi_30_64);
+
+  const __m256i cospi_p14_p18 = pair256_set_epi16(cospi_14_64, cospi_18_64);
+  const __m256i cospi_m18_p14 = pair256_set_epi16(-cospi_18_64, cospi_14_64);
+
+  const __m256i cospi_p22_p10 = pair256_set_epi16(cospi_22_64, cospi_10_64);
+  const __m256i cospi_m10_p22 = pair256_set_epi16(-cospi_10_64, cospi_22_64);
+
+  const __m256i cospi_p06_p26 = pair256_set_epi16(cospi_6_64, cospi_26_64);
+  const __m256i cospi_m26_p06 = pair256_set_epi16(-cospi_26_64, cospi_6_64);
+
+  __m256i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m256i s0, s1, s2, s3, s4, s5, s6, s7;
+  __m256i t0, t1, t2, t3, t4, t5, t6, t7;
+  __m256i v0, v1, v2, v3;
+  __m256i x0, x1;
+
+  // 0, 4, 8, 12
+  u0 = _mm256_add_epi16(in[0], in[15]);
+  u1 = _mm256_add_epi16(in[1], in[14]);
+  u2 = _mm256_add_epi16(in[2], in[13]);
+  u3 = _mm256_add_epi16(in[3], in[12]);
+  u4 = _mm256_add_epi16(in[4], in[11]);
+  u5 = _mm256_add_epi16(in[5], in[10]);
+  u6 = _mm256_add_epi16(in[6], in[9]);
+  u7 = _mm256_add_epi16(in[7], in[8]);
+
+  s0 = _mm256_add_epi16(u0, u7);
+  s1 = _mm256_add_epi16(u1, u6);
+  s2 = _mm256_add_epi16(u2, u5);
+  s3 = _mm256_add_epi16(u3, u4);
+
+  // 0, 8
+  v0 = _mm256_add_epi16(s0, s3);
+  v1 = _mm256_add_epi16(s1, s2);
+
+  x0 = _mm256_unpacklo_epi16(v0, v1);
+  x1 = _mm256_unpackhi_epi16(v0, v1);
+
+  t0 = butter_fly(x0, x1, cospi_p16_p16);
+  t1 = butter_fly(x0, x1, cospi_p16_m16);
+
+  // 4, 12
+  v0 = _mm256_sub_epi16(s1, s2);
+  v1 = _mm256_sub_epi16(s0, s3);
+
+  x0 = _mm256_unpacklo_epi16(v0, v1);
+  x1 = _mm256_unpackhi_epi16(v0, v1);
+
+  t2 = butter_fly(x0, x1, cospi_p24_p08);
+  t3 = butter_fly(x0, x1, cospi_m08_p24);
+
+  // 2, 6, 10, 14
+  s0 = _mm256_sub_epi16(u3, u4);
+  s1 = _mm256_sub_epi16(u2, u5);
+  s2 = _mm256_sub_epi16(u1, u6);
+  s3 = _mm256_sub_epi16(u0, u7);
+
+  v0 = s0;  // output[4]
+  v3 = s3;  // output[7]
+
+  x0 = _mm256_unpacklo_epi16(s2, s1);
+  x1 = _mm256_unpackhi_epi16(s2, s1);
+
+  v2 = butter_fly(x0, x1, cospi_p16_p16);  // output[5]
+  v1 = butter_fly(x0, x1, cospi_p16_m16);  // output[6]
+
+  s0 = _mm256_add_epi16(v0, v1);  // step[4]
+  s1 = _mm256_sub_epi16(v0, v1);  // step[5]
+  s2 = _mm256_sub_epi16(v3, v2);  // step[6]
+  s3 = _mm256_add_epi16(v3, v2);  // step[7]
+
+  // 2, 14
+  x0 = _mm256_unpacklo_epi16(s0, s3);
+  x1 = _mm256_unpackhi_epi16(s0, s3);
+
+  t4 = butter_fly(x0, x1, cospi_p28_p04);
+  t5 = butter_fly(x0, x1, cospi_m04_p28);
+
+  // 10, 6
+  x0 = _mm256_unpacklo_epi16(s1, s2);
+  x1 = _mm256_unpackhi_epi16(s1, s2);
+  t6 = butter_fly(x0, x1, cospi_p12_p20);
+  t7 = butter_fly(x0, x1, cospi_m20_p12);
+
+  // 1, 3, 5, 7, 9, 11, 13, 15
+  s0 = _mm256_sub_epi16(in[7], in[8]);  // step[8]
+  s1 = _mm256_sub_epi16(in[6], in[9]);  // step[9]
+  u2 = _mm256_sub_epi16(in[5], in[10]);
+  u3 = _mm256_sub_epi16(in[4], in[11]);
+  u4 = _mm256_sub_epi16(in[3], in[12]);
+  u5 = _mm256_sub_epi16(in[2], in[13]);
+  s6 = _mm256_sub_epi16(in[1], in[14]);  // step[14]
+  s7 = _mm256_sub_epi16(in[0], in[15]);  // step[15]
+
+  in[0] = t0;
+  in[8] = t1;
+  in[4] = t2;
+  in[12] = t3;
+  in[2] = t4;
+  in[14] = t5;
+  in[10] = t6;
+  in[6] = t7;
+
+  x0 = _mm256_unpacklo_epi16(u5, u2);
+  x1 = _mm256_unpackhi_epi16(u5, u2);
+
+  s2 = butter_fly(x0, x1, cospi_p16_p16);  // step[13]
+  s5 = butter_fly(x0, x1, cospi_p16_m16);  // step[10]
+
+  x0 = _mm256_unpacklo_epi16(u4, u3);
+  x1 = _mm256_unpackhi_epi16(u4, u3);
+
+  s3 = butter_fly(x0, x1, cospi_p16_p16);  // step[12]
+  s4 = butter_fly(x0, x1, cospi_p16_m16);  // step[11]
+
+  u0 = _mm256_add_epi16(s0, s4);  // output[8]
+  u1 = _mm256_add_epi16(s1, s5);
+  u2 = _mm256_sub_epi16(s1, s5);
+  u3 = _mm256_sub_epi16(s0, s4);
+  u4 = _mm256_sub_epi16(s7, s3);
+  u5 = _mm256_sub_epi16(s6, s2);
+  u6 = _mm256_add_epi16(s6, s2);
+  u7 = _mm256_add_epi16(s7, s3);
+
+  // stage 4
+  s0 = u0;
+  s3 = u3;
+  s4 = u4;
+  s7 = u7;
+
+  x0 = _mm256_unpacklo_epi16(u1, u6);
+  x1 = _mm256_unpackhi_epi16(u1, u6);
+
+  s1 = butter_fly(x0, x1, cospi_m08_p24);
+  s6 = butter_fly(x0, x1, cospi_p24_p08);
+
+  x0 = _mm256_unpacklo_epi16(u2, u5);
+  x1 = _mm256_unpackhi_epi16(u2, u5);
+
+  s2 = butter_fly(x0, x1, cospi_m24_m08);
+  s5 = butter_fly(x0, x1, cospi_m08_p24);
+
+  // stage 5
+  u0 = _mm256_add_epi16(s0, s1);
+  u1 = _mm256_sub_epi16(s0, s1);
+  u2 = _mm256_sub_epi16(s3, s2);
+  u3 = _mm256_add_epi16(s3, s2);
+  u4 = _mm256_add_epi16(s4, s5);
+  u5 = _mm256_sub_epi16(s4, s5);
+  u6 = _mm256_sub_epi16(s7, s6);
+  u7 = _mm256_add_epi16(s7, s6);
+
+  // stage 6
+  x0 = _mm256_unpacklo_epi16(u0, u7);
+  x1 = _mm256_unpackhi_epi16(u0, u7);
+  in[1] = butter_fly(x0, x1, cospi_p30_p02);
+  in[15] = butter_fly(x0, x1, cospi_m02_p30);
+
+  x0 = _mm256_unpacklo_epi16(u1, u6);
+  x1 = _mm256_unpackhi_epi16(u1, u6);
+  in[9] = butter_fly(x0, x1, cospi_p14_p18);
+  in[7] = butter_fly(x0, x1, cospi_m18_p14);
+
+  x0 = _mm256_unpacklo_epi16(u2, u5);
+  x1 = _mm256_unpackhi_epi16(u2, u5);
+  in[5] = butter_fly(x0, x1, cospi_p22_p10);
+  in[11] = butter_fly(x0, x1, cospi_m10_p22);
+
+  x0 = _mm256_unpacklo_epi16(u3, u4);
+  x1 = _mm256_unpackhi_epi16(u3, u4);
+  in[13] = butter_fly(x0, x1, cospi_p06_p26);
+  in[3] = butter_fly(x0, x1, cospi_m26_p06);
+}
+
+void fadst16_avx2(__m256i *in) {
+  const __m256i cospi_p01_p31 = pair256_set_epi16(cospi_1_64, cospi_31_64);
+  const __m256i cospi_p31_m01 = pair256_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m256i cospi_p05_p27 = pair256_set_epi16(cospi_5_64, cospi_27_64);
+  const __m256i cospi_p27_m05 = pair256_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m256i cospi_p09_p23 = pair256_set_epi16(cospi_9_64, cospi_23_64);
+  const __m256i cospi_p23_m09 = pair256_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m256i cospi_p13_p19 = pair256_set_epi16(cospi_13_64, cospi_19_64);
+  const __m256i cospi_p19_m13 = pair256_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m256i cospi_p17_p15 = pair256_set_epi16(cospi_17_64, cospi_15_64);
+  const __m256i cospi_p15_m17 = pair256_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m256i cospi_p21_p11 = pair256_set_epi16(cospi_21_64, cospi_11_64);
+  const __m256i cospi_p11_m21 = pair256_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m256i cospi_p25_p07 = pair256_set_epi16(cospi_25_64, cospi_7_64);
+  const __m256i cospi_p07_m25 = pair256_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m256i cospi_p29_p03 = pair256_set_epi16(cospi_29_64, cospi_3_64);
+  const __m256i cospi_p03_m29 = pair256_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m256i cospi_p04_p28 = pair256_set_epi16(cospi_4_64, cospi_28_64);
+  const __m256i cospi_p28_m04 = pair256_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m256i cospi_p20_p12 = pair256_set_epi16(cospi_20_64, cospi_12_64);
+  const __m256i cospi_p12_m20 = pair256_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m256i cospi_m28_p04 = pair256_set_epi16(-cospi_28_64, cospi_4_64);
+  const __m256i cospi_m12_p20 = pair256_set_epi16(-cospi_12_64, cospi_20_64);
+  const __m256i cospi_p08_p24 = pair256_set_epi16(cospi_8_64, cospi_24_64);
+  const __m256i cospi_p24_m08 = pair256_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m256i cospi_m24_p08 = pair256_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m256i cospi_m16_m16 = _mm256_set1_epi16((int16_t)-cospi_16_64);
+  const __m256i cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64);
+  const __m256i cospi_p16_m16 = pair256_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m256i cospi_m16_p16 = pair256_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m256i zero = _mm256_setzero_si256();
+  const __m256i dct_rounding = _mm256_set1_epi32(DCT_CONST_ROUNDING);
+  __m256i s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14, s15;
+  __m256i x0, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, x15;
+  __m256i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
+  __m256i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
+  __m256i y0, y1;
+
+  // stage 1, s takes low 256 bits; x takes high 256 bits
+  y0 = _mm256_unpacklo_epi16(in[15], in[0]);
+  y1 = _mm256_unpackhi_epi16(in[15], in[0]);
+  s0 = _mm256_madd_epi16(y0, cospi_p01_p31);
+  x0 = _mm256_madd_epi16(y1, cospi_p01_p31);
+  s1 = _mm256_madd_epi16(y0, cospi_p31_m01);
+  x1 = _mm256_madd_epi16(y1, cospi_p31_m01);
+
+  y0 = _mm256_unpacklo_epi16(in[13], in[2]);
+  y1 = _mm256_unpackhi_epi16(in[13], in[2]);
+  s2 = _mm256_madd_epi16(y0, cospi_p05_p27);
+  x2 = _mm256_madd_epi16(y1, cospi_p05_p27);
+  s3 = _mm256_madd_epi16(y0, cospi_p27_m05);
+  x3 = _mm256_madd_epi16(y1, cospi_p27_m05);
+
+  y0 = _mm256_unpacklo_epi16(in[11], in[4]);
+  y1 = _mm256_unpackhi_epi16(in[11], in[4]);
+  s4 = _mm256_madd_epi16(y0, cospi_p09_p23);
+  x4 = _mm256_madd_epi16(y1, cospi_p09_p23);
+  s5 = _mm256_madd_epi16(y0, cospi_p23_m09);
+  x5 = _mm256_madd_epi16(y1, cospi_p23_m09);
+
+  y0 = _mm256_unpacklo_epi16(in[9], in[6]);
+  y1 = _mm256_unpackhi_epi16(in[9], in[6]);
+  s6 = _mm256_madd_epi16(y0, cospi_p13_p19);
+  x6 = _mm256_madd_epi16(y1, cospi_p13_p19);
+  s7 = _mm256_madd_epi16(y0, cospi_p19_m13);
+  x7 = _mm256_madd_epi16(y1, cospi_p19_m13);
+
+  y0 = _mm256_unpacklo_epi16(in[7], in[8]);
+  y1 = _mm256_unpackhi_epi16(in[7], in[8]);
+  s8 = _mm256_madd_epi16(y0, cospi_p17_p15);
+  x8 = _mm256_madd_epi16(y1, cospi_p17_p15);
+  s9 = _mm256_madd_epi16(y0, cospi_p15_m17);
+  x9 = _mm256_madd_epi16(y1, cospi_p15_m17);
+
+  y0 = _mm256_unpacklo_epi16(in[5], in[10]);
+  y1 = _mm256_unpackhi_epi16(in[5], in[10]);
+  s10 = _mm256_madd_epi16(y0, cospi_p21_p11);
+  x10 = _mm256_madd_epi16(y1, cospi_p21_p11);
+  s11 = _mm256_madd_epi16(y0, cospi_p11_m21);
+  x11 = _mm256_madd_epi16(y1, cospi_p11_m21);
+
+  y0 = _mm256_unpacklo_epi16(in[3], in[12]);
+  y1 = _mm256_unpackhi_epi16(in[3], in[12]);
+  s12 = _mm256_madd_epi16(y0, cospi_p25_p07);
+  x12 = _mm256_madd_epi16(y1, cospi_p25_p07);
+  s13 = _mm256_madd_epi16(y0, cospi_p07_m25);
+  x13 = _mm256_madd_epi16(y1, cospi_p07_m25);
+
+  y0 = _mm256_unpacklo_epi16(in[1], in[14]);
+  y1 = _mm256_unpackhi_epi16(in[1], in[14]);
+  s14 = _mm256_madd_epi16(y0, cospi_p29_p03);
+  x14 = _mm256_madd_epi16(y1, cospi_p29_p03);
+  s15 = _mm256_madd_epi16(y0, cospi_p03_m29);
+  x15 = _mm256_madd_epi16(y1, cospi_p03_m29);
+
+  // u takes low 256 bits; v takes high 256 bits
+  u0 = _mm256_add_epi32(s0, s8);
+  u1 = _mm256_add_epi32(s1, s9);
+  u2 = _mm256_add_epi32(s2, s10);
+  u3 = _mm256_add_epi32(s3, s11);
+  u4 = _mm256_add_epi32(s4, s12);
+  u5 = _mm256_add_epi32(s5, s13);
+  u6 = _mm256_add_epi32(s6, s14);
+  u7 = _mm256_add_epi32(s7, s15);
+
+  u8 = _mm256_sub_epi32(s0, s8);
+  u9 = _mm256_sub_epi32(s1, s9);
+  u10 = _mm256_sub_epi32(s2, s10);
+  u11 = _mm256_sub_epi32(s3, s11);
+  u12 = _mm256_sub_epi32(s4, s12);
+  u13 = _mm256_sub_epi32(s5, s13);
+  u14 = _mm256_sub_epi32(s6, s14);
+  u15 = _mm256_sub_epi32(s7, s15);
+
+  v0 = _mm256_add_epi32(x0, x8);
+  v1 = _mm256_add_epi32(x1, x9);
+  v2 = _mm256_add_epi32(x2, x10);
+  v3 = _mm256_add_epi32(x3, x11);
+  v4 = _mm256_add_epi32(x4, x12);
+  v5 = _mm256_add_epi32(x5, x13);
+  v6 = _mm256_add_epi32(x6, x14);
+  v7 = _mm256_add_epi32(x7, x15);
+
+  v8 = _mm256_sub_epi32(x0, x8);
+  v9 = _mm256_sub_epi32(x1, x9);
+  v10 = _mm256_sub_epi32(x2, x10);
+  v11 = _mm256_sub_epi32(x3, x11);
+  v12 = _mm256_sub_epi32(x4, x12);
+  v13 = _mm256_sub_epi32(x5, x13);
+  v14 = _mm256_sub_epi32(x6, x14);
+  v15 = _mm256_sub_epi32(x7, x15);
+
+  // low 256 bits rounding
+  u8 = _mm256_add_epi32(u8, dct_rounding);
+  u9 = _mm256_add_epi32(u9, dct_rounding);
+  u10 = _mm256_add_epi32(u10, dct_rounding);
+  u11 = _mm256_add_epi32(u11, dct_rounding);
+  u12 = _mm256_add_epi32(u12, dct_rounding);
+  u13 = _mm256_add_epi32(u13, dct_rounding);
+  u14 = _mm256_add_epi32(u14, dct_rounding);
+  u15 = _mm256_add_epi32(u15, dct_rounding);
+
+  u8 = _mm256_srai_epi32(u8, DCT_CONST_BITS);
+  u9 = _mm256_srai_epi32(u9, DCT_CONST_BITS);
+  u10 = _mm256_srai_epi32(u10, DCT_CONST_BITS);
+  u11 = _mm256_srai_epi32(u11, DCT_CONST_BITS);
+  u12 = _mm256_srai_epi32(u12, DCT_CONST_BITS);
+  u13 = _mm256_srai_epi32(u13, DCT_CONST_BITS);
+  u14 = _mm256_srai_epi32(u14, DCT_CONST_BITS);
+  u15 = _mm256_srai_epi32(u15, DCT_CONST_BITS);
+
+  // high 256 bits rounding
+  v8 = _mm256_add_epi32(v8, dct_rounding);
+  v9 = _mm256_add_epi32(v9, dct_rounding);
+  v10 = _mm256_add_epi32(v10, dct_rounding);
+  v11 = _mm256_add_epi32(v11, dct_rounding);
+  v12 = _mm256_add_epi32(v12, dct_rounding);
+  v13 = _mm256_add_epi32(v13, dct_rounding);
+  v14 = _mm256_add_epi32(v14, dct_rounding);
+  v15 = _mm256_add_epi32(v15, dct_rounding);
+
+  v8 = _mm256_srai_epi32(v8, DCT_CONST_BITS);
+  v9 = _mm256_srai_epi32(v9, DCT_CONST_BITS);
+  v10 = _mm256_srai_epi32(v10, DCT_CONST_BITS);
+  v11 = _mm256_srai_epi32(v11, DCT_CONST_BITS);
+  v12 = _mm256_srai_epi32(v12, DCT_CONST_BITS);
+  v13 = _mm256_srai_epi32(v13, DCT_CONST_BITS);
+  v14 = _mm256_srai_epi32(v14, DCT_CONST_BITS);
+  v15 = _mm256_srai_epi32(v15, DCT_CONST_BITS);
+
+  // Saturation pack 32-bit to 16-bit
+  x8 = _mm256_packs_epi32(u8, v8);
+  x9 = _mm256_packs_epi32(u9, v9);
+  x10 = _mm256_packs_epi32(u10, v10);
+  x11 = _mm256_packs_epi32(u11, v11);
+  x12 = _mm256_packs_epi32(u12, v12);
+  x13 = _mm256_packs_epi32(u13, v13);
+  x14 = _mm256_packs_epi32(u14, v14);
+  x15 = _mm256_packs_epi32(u15, v15);
+
+  // stage 2
+  y0 = _mm256_unpacklo_epi16(x8, x9);
+  y1 = _mm256_unpackhi_epi16(x8, x9);
+  s8 = _mm256_madd_epi16(y0, cospi_p04_p28);
+  x8 = _mm256_madd_epi16(y1, cospi_p04_p28);
+  s9 = _mm256_madd_epi16(y0, cospi_p28_m04);
+  x9 = _mm256_madd_epi16(y1, cospi_p28_m04);
+
+  y0 = _mm256_unpacklo_epi16(x10, x11);
+  y1 = _mm256_unpackhi_epi16(x10, x11);
+  s10 = _mm256_madd_epi16(y0, cospi_p20_p12);
+  x10 = _mm256_madd_epi16(y1, cospi_p20_p12);
+  s11 = _mm256_madd_epi16(y0, cospi_p12_m20);
+  x11 = _mm256_madd_epi16(y1, cospi_p12_m20);
+
+  y0 = _mm256_unpacklo_epi16(x12, x13);
+  y1 = _mm256_unpackhi_epi16(x12, x13);
+  s12 = _mm256_madd_epi16(y0, cospi_m28_p04);
+  x12 = _mm256_madd_epi16(y1, cospi_m28_p04);
+  s13 = _mm256_madd_epi16(y0, cospi_p04_p28);
+  x13 = _mm256_madd_epi16(y1, cospi_p04_p28);
+
+  y0 = _mm256_unpacklo_epi16(x14, x15);
+  y1 = _mm256_unpackhi_epi16(x14, x15);
+  s14 = _mm256_madd_epi16(y0, cospi_m12_p20);
+  x14 = _mm256_madd_epi16(y1, cospi_m12_p20);
+  s15 = _mm256_madd_epi16(y0, cospi_p20_p12);
+  x15 = _mm256_madd_epi16(y1, cospi_p20_p12);
+
+  x0 = _mm256_add_epi32(u0, u4);
+  s0 = _mm256_add_epi32(v0, v4);
+  x1 = _mm256_add_epi32(u1, u5);
+  s1 = _mm256_add_epi32(v1, v5);
+  x2 = _mm256_add_epi32(u2, u6);
+  s2 = _mm256_add_epi32(v2, v6);
+  x3 = _mm256_add_epi32(u3, u7);
+  s3 = _mm256_add_epi32(v3, v7);
+
+  v8 = _mm256_sub_epi32(u0, u4);
+  v9 = _mm256_sub_epi32(v0, v4);
+  v10 = _mm256_sub_epi32(u1, u5);
+  v11 = _mm256_sub_epi32(v1, v5);
+  v12 = _mm256_sub_epi32(u2, u6);
+  v13 = _mm256_sub_epi32(v2, v6);
+  v14 = _mm256_sub_epi32(u3, u7);
+  v15 = _mm256_sub_epi32(v3, v7);
+
+  v8 = _mm256_add_epi32(v8, dct_rounding);
+  v9 = _mm256_add_epi32(v9, dct_rounding);
+  v10 = _mm256_add_epi32(v10, dct_rounding);
+  v11 = _mm256_add_epi32(v11, dct_rounding);
+  v12 = _mm256_add_epi32(v12, dct_rounding);
+  v13 = _mm256_add_epi32(v13, dct_rounding);
+  v14 = _mm256_add_epi32(v14, dct_rounding);
+  v15 = _mm256_add_epi32(v15, dct_rounding);
+
+  v8 = _mm256_srai_epi32(v8, DCT_CONST_BITS);
+  v9 = _mm256_srai_epi32(v9, DCT_CONST_BITS);
+  v10 = _mm256_srai_epi32(v10, DCT_CONST_BITS);
+  v11 = _mm256_srai_epi32(v11, DCT_CONST_BITS);
+  v12 = _mm256_srai_epi32(v12, DCT_CONST_BITS);
+  v13 = _mm256_srai_epi32(v13, DCT_CONST_BITS);
+  v14 = _mm256_srai_epi32(v14, DCT_CONST_BITS);
+  v15 = _mm256_srai_epi32(v15, DCT_CONST_BITS);
+
+  x4 = _mm256_packs_epi32(v8, v9);
+  x5 = _mm256_packs_epi32(v10, v11);
+  x6 = _mm256_packs_epi32(v12, v13);
+  x7 = _mm256_packs_epi32(v14, v15);
+
+  u8 = _mm256_add_epi32(s8, s12);
+  u9 = _mm256_add_epi32(s9, s13);
+  u10 = _mm256_add_epi32(s10, s14);
+  u11 = _mm256_add_epi32(s11, s15);
+  u12 = _mm256_sub_epi32(s8, s12);
+  u13 = _mm256_sub_epi32(s9, s13);
+  u14 = _mm256_sub_epi32(s10, s14);
+  u15 = _mm256_sub_epi32(s11, s15);
+
+  v8 = _mm256_add_epi32(x8, x12);
+  v9 = _mm256_add_epi32(x9, x13);
+  v10 = _mm256_add_epi32(x10, x14);
+  v11 = _mm256_add_epi32(x11, x15);
+  v12 = _mm256_sub_epi32(x8, x12);
+  v13 = _mm256_sub_epi32(x9, x13);
+  v14 = _mm256_sub_epi32(x10, x14);
+  v15 = _mm256_sub_epi32(x11, x15);
+
+  u12 = _mm256_add_epi32(u12, dct_rounding);
+  u13 = _mm256_add_epi32(u13, dct_rounding);
+  u14 = _mm256_add_epi32(u14, dct_rounding);
+  u15 = _mm256_add_epi32(u15, dct_rounding);
+
+  u12 = _mm256_srai_epi32(u12, DCT_CONST_BITS);
+  u13 = _mm256_srai_epi32(u13, DCT_CONST_BITS);
+  u14 = _mm256_srai_epi32(u14, DCT_CONST_BITS);
+  u15 = _mm256_srai_epi32(u15, DCT_CONST_BITS);
+
+  v12 = _mm256_add_epi32(v12, dct_rounding);
+  v13 = _mm256_add_epi32(v13, dct_rounding);
+  v14 = _mm256_add_epi32(v14, dct_rounding);
+  v15 = _mm256_add_epi32(v15, dct_rounding);
+
+  v12 = _mm256_srai_epi32(v12, DCT_CONST_BITS);
+  v13 = _mm256_srai_epi32(v13, DCT_CONST_BITS);
+  v14 = _mm256_srai_epi32(v14, DCT_CONST_BITS);
+  v15 = _mm256_srai_epi32(v15, DCT_CONST_BITS);
+
+  x12 = _mm256_packs_epi32(u12, v12);
+  x13 = _mm256_packs_epi32(u13, v13);
+  x14 = _mm256_packs_epi32(u14, v14);
+  x15 = _mm256_packs_epi32(u15, v15);
+
+  // stage 3
+  y0 = _mm256_unpacklo_epi16(x4, x5);
+  y1 = _mm256_unpackhi_epi16(x4, x5);
+  s4 = _mm256_madd_epi16(y0, cospi_p08_p24);
+  x4 = _mm256_madd_epi16(y1, cospi_p08_p24);
+  s5 = _mm256_madd_epi16(y0, cospi_p24_m08);
+  x5 = _mm256_madd_epi16(y1, cospi_p24_m08);
+
+  y0 = _mm256_unpacklo_epi16(x6, x7);
+  y1 = _mm256_unpackhi_epi16(x6, x7);
+  s6 = _mm256_madd_epi16(y0, cospi_m24_p08);
+  x6 = _mm256_madd_epi16(y1, cospi_m24_p08);
+  s7 = _mm256_madd_epi16(y0, cospi_p08_p24);
+  x7 = _mm256_madd_epi16(y1, cospi_p08_p24);
+
+  y0 = _mm256_unpacklo_epi16(x12, x13);
+  y1 = _mm256_unpackhi_epi16(x12, x13);
+  s12 = _mm256_madd_epi16(y0, cospi_p08_p24);
+  x12 = _mm256_madd_epi16(y1, cospi_p08_p24);
+  s13 = _mm256_madd_epi16(y0, cospi_p24_m08);
+  x13 = _mm256_madd_epi16(y1, cospi_p24_m08);
+
+  y0 = _mm256_unpacklo_epi16(x14, x15);
+  y1 = _mm256_unpackhi_epi16(x14, x15);
+  s14 = _mm256_madd_epi16(y0, cospi_m24_p08);
+  x14 = _mm256_madd_epi16(y1, cospi_m24_p08);
+  s15 = _mm256_madd_epi16(y0, cospi_p08_p24);
+  x15 = _mm256_madd_epi16(y1, cospi_p08_p24);
+
+  u0 = _mm256_add_epi32(x0, x2);
+  v0 = _mm256_add_epi32(s0, s2);
+  u1 = _mm256_add_epi32(x1, x3);
+  v1 = _mm256_add_epi32(s1, s3);
+  u2 = _mm256_sub_epi32(x0, x2);
+  v2 = _mm256_sub_epi32(s0, s2);
+  u3 = _mm256_sub_epi32(x1, x3);
+  v3 = _mm256_sub_epi32(s1, s3);
+
+  u0 = _mm256_add_epi32(u0, dct_rounding);
+  v0 = _mm256_add_epi32(v0, dct_rounding);
+  u1 = _mm256_add_epi32(u1, dct_rounding);
+  v1 = _mm256_add_epi32(v1, dct_rounding);
+  u2 = _mm256_add_epi32(u2, dct_rounding);
+  v2 = _mm256_add_epi32(v2, dct_rounding);
+  u3 = _mm256_add_epi32(u3, dct_rounding);
+  v3 = _mm256_add_epi32(v3, dct_rounding);
+
+  u0 = _mm256_srai_epi32(u0, DCT_CONST_BITS);
+  v0 = _mm256_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm256_srai_epi32(u1, DCT_CONST_BITS);
+  v1 = _mm256_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm256_srai_epi32(u2, DCT_CONST_BITS);
+  v2 = _mm256_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm256_srai_epi32(u3, DCT_CONST_BITS);
+  v3 = _mm256_srai_epi32(v3, DCT_CONST_BITS);
+
+  in[0] = _mm256_packs_epi32(u0, v0);
+  x1 = _mm256_packs_epi32(u1, v1);
+  x2 = _mm256_packs_epi32(u2, v2);
+  x3 = _mm256_packs_epi32(u3, v3);
+
+  // Rounding on s4 + s6, s5 + s7, s4 - s6, s5 - s7
+  u4 = _mm256_add_epi32(s4, s6);
+  u5 = _mm256_add_epi32(s5, s7);
+  u6 = _mm256_sub_epi32(s4, s6);
+  u7 = _mm256_sub_epi32(s5, s7);
+
+  v4 = _mm256_add_epi32(x4, x6);
+  v5 = _mm256_add_epi32(x5, x7);
+  v6 = _mm256_sub_epi32(x4, x6);
+  v7 = _mm256_sub_epi32(x5, x7);
+
+  u4 = _mm256_add_epi32(u4, dct_rounding);
+  u5 = _mm256_add_epi32(u5, dct_rounding);
+  u6 = _mm256_add_epi32(u6, dct_rounding);
+  u7 = _mm256_add_epi32(u7, dct_rounding);
+
+  u4 = _mm256_srai_epi32(u4, DCT_CONST_BITS);
+  u5 = _mm256_srai_epi32(u5, DCT_CONST_BITS);
+  u6 = _mm256_srai_epi32(u6, DCT_CONST_BITS);
+  u7 = _mm256_srai_epi32(u7, DCT_CONST_BITS);
+
+  v4 = _mm256_add_epi32(v4, dct_rounding);
+  v5 = _mm256_add_epi32(v5, dct_rounding);
+  v6 = _mm256_add_epi32(v6, dct_rounding);
+  v7 = _mm256_add_epi32(v7, dct_rounding);
+
+  v4 = _mm256_srai_epi32(v4, DCT_CONST_BITS);
+  v5 = _mm256_srai_epi32(v5, DCT_CONST_BITS);
+  v6 = _mm256_srai_epi32(v6, DCT_CONST_BITS);
+  v7 = _mm256_srai_epi32(v7, DCT_CONST_BITS);
+
+  x4 = _mm256_packs_epi32(u4, v4);
+  in[12] = _mm256_packs_epi32(u5, v5);
+  x6 = _mm256_packs_epi32(u6, v6);
+  x7 = _mm256_packs_epi32(u7, v7);
+
+  u0 = _mm256_add_epi32(u8, u10);
+  v0 = _mm256_add_epi32(v8, v10);
+  u1 = _mm256_add_epi32(u9, u11);
+  v1 = _mm256_add_epi32(v9, v11);
+  u2 = _mm256_sub_epi32(u8, u10);
+  v2 = _mm256_sub_epi32(v8, v10);
+  u3 = _mm256_sub_epi32(u9, u11);
+  v3 = _mm256_sub_epi32(v9, v11);
+
+  u0 = _mm256_add_epi32(u0, dct_rounding);
+  v0 = _mm256_add_epi32(v0, dct_rounding);
+  u1 = _mm256_add_epi32(u1, dct_rounding);
+  v1 = _mm256_add_epi32(v1, dct_rounding);
+  u2 = _mm256_add_epi32(u2, dct_rounding);
+  v2 = _mm256_add_epi32(v2, dct_rounding);
+  u3 = _mm256_add_epi32(u3, dct_rounding);
+  v3 = _mm256_add_epi32(v3, dct_rounding);
+
+  u0 = _mm256_srai_epi32(u0, DCT_CONST_BITS);
+  v0 = _mm256_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm256_srai_epi32(u1, DCT_CONST_BITS);
+  v1 = _mm256_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm256_srai_epi32(u2, DCT_CONST_BITS);
+  v2 = _mm256_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm256_srai_epi32(u3, DCT_CONST_BITS);
+  v3 = _mm256_srai_epi32(v3, DCT_CONST_BITS);
+
+  x8 = _mm256_packs_epi32(u0, v0);
+  in[14] = _mm256_packs_epi32(u1, v1);
+  x10 = _mm256_packs_epi32(u2, v2);
+  x11 = _mm256_packs_epi32(u3, v3);
+
+  // Rounding on s12 + s14, s13 + s15, s12 - s14, s13 - s15
+  u12 = _mm256_add_epi32(s12, s14);
+  u13 = _mm256_add_epi32(s13, s15);
+  u14 = _mm256_sub_epi32(s12, s14);
+  u15 = _mm256_sub_epi32(s13, s15);
+
+  v12 = _mm256_add_epi32(x12, x14);
+  v13 = _mm256_add_epi32(x13, x15);
+  v14 = _mm256_sub_epi32(x12, x14);
+  v15 = _mm256_sub_epi32(x13, x15);
+
+  u12 = _mm256_add_epi32(u12, dct_rounding);
+  u13 = _mm256_add_epi32(u13, dct_rounding);
+  u14 = _mm256_add_epi32(u14, dct_rounding);
+  u15 = _mm256_add_epi32(u15, dct_rounding);
+
+  u12 = _mm256_srai_epi32(u12, DCT_CONST_BITS);
+  u13 = _mm256_srai_epi32(u13, DCT_CONST_BITS);
+  u14 = _mm256_srai_epi32(u14, DCT_CONST_BITS);
+  u15 = _mm256_srai_epi32(u15, DCT_CONST_BITS);
+
+  v12 = _mm256_add_epi32(v12, dct_rounding);
+  v13 = _mm256_add_epi32(v13, dct_rounding);
+  v14 = _mm256_add_epi32(v14, dct_rounding);
+  v15 = _mm256_add_epi32(v15, dct_rounding);
+
+  v12 = _mm256_srai_epi32(v12, DCT_CONST_BITS);
+  v13 = _mm256_srai_epi32(v13, DCT_CONST_BITS);
+  v14 = _mm256_srai_epi32(v14, DCT_CONST_BITS);
+  v15 = _mm256_srai_epi32(v15, DCT_CONST_BITS);
+
+  x12 = _mm256_packs_epi32(u12, v12);
+  x13 = _mm256_packs_epi32(u13, v13);
+  x14 = _mm256_packs_epi32(u14, v14);
+  x15 = _mm256_packs_epi32(u15, v15);
+  in[2] = x12;
+
+  // stage 4
+  y0 = _mm256_unpacklo_epi16(x2, x3);
+  y1 = _mm256_unpackhi_epi16(x2, x3);
+  s2 = _mm256_madd_epi16(y0, cospi_m16_m16);
+  x2 = _mm256_madd_epi16(y1, cospi_m16_m16);
+  s3 = _mm256_madd_epi16(y0, cospi_p16_m16);
+  x3 = _mm256_madd_epi16(y1, cospi_p16_m16);
+
+  y0 = _mm256_unpacklo_epi16(x6, x7);
+  y1 = _mm256_unpackhi_epi16(x6, x7);
+  s6 = _mm256_madd_epi16(y0, cospi_p16_p16);
+  x6 = _mm256_madd_epi16(y1, cospi_p16_p16);
+  s7 = _mm256_madd_epi16(y0, cospi_m16_p16);
+  x7 = _mm256_madd_epi16(y1, cospi_m16_p16);
+
+  y0 = _mm256_unpacklo_epi16(x10, x11);
+  y1 = _mm256_unpackhi_epi16(x10, x11);
+  s10 = _mm256_madd_epi16(y0, cospi_p16_p16);
+  x10 = _mm256_madd_epi16(y1, cospi_p16_p16);
+  s11 = _mm256_madd_epi16(y0, cospi_m16_p16);
+  x11 = _mm256_madd_epi16(y1, cospi_m16_p16);
+
+  y0 = _mm256_unpacklo_epi16(x14, x15);
+  y1 = _mm256_unpackhi_epi16(x14, x15);
+  s14 = _mm256_madd_epi16(y0, cospi_m16_m16);
+  x14 = _mm256_madd_epi16(y1, cospi_m16_m16);
+  s15 = _mm256_madd_epi16(y0, cospi_p16_m16);
+  x15 = _mm256_madd_epi16(y1, cospi_p16_m16);
+
+  // Rounding
+  u2 = _mm256_add_epi32(s2, dct_rounding);
+  u3 = _mm256_add_epi32(s3, dct_rounding);
+  u6 = _mm256_add_epi32(s6, dct_rounding);
+  u7 = _mm256_add_epi32(s7, dct_rounding);
+
+  u10 = _mm256_add_epi32(s10, dct_rounding);
+  u11 = _mm256_add_epi32(s11, dct_rounding);
+  u14 = _mm256_add_epi32(s14, dct_rounding);
+  u15 = _mm256_add_epi32(s15, dct_rounding);
+
+  u2 = _mm256_srai_epi32(u2, DCT_CONST_BITS);
+  u3 = _mm256_srai_epi32(u3, DCT_CONST_BITS);
+  u6 = _mm256_srai_epi32(u6, DCT_CONST_BITS);
+  u7 = _mm256_srai_epi32(u7, DCT_CONST_BITS);
+
+  u10 = _mm256_srai_epi32(u10, DCT_CONST_BITS);
+  u11 = _mm256_srai_epi32(u11, DCT_CONST_BITS);
+  u14 = _mm256_srai_epi32(u14, DCT_CONST_BITS);
+  u15 = _mm256_srai_epi32(u15, DCT_CONST_BITS);
+
+  v2 = _mm256_add_epi32(x2, dct_rounding);
+  v3 = _mm256_add_epi32(x3, dct_rounding);
+  v6 = _mm256_add_epi32(x6, dct_rounding);
+  v7 = _mm256_add_epi32(x7, dct_rounding);
+
+  v10 = _mm256_add_epi32(x10, dct_rounding);
+  v11 = _mm256_add_epi32(x11, dct_rounding);
+  v14 = _mm256_add_epi32(x14, dct_rounding);
+  v15 = _mm256_add_epi32(x15, dct_rounding);
+
+  v2 = _mm256_srai_epi32(v2, DCT_CONST_BITS);
+  v3 = _mm256_srai_epi32(v3, DCT_CONST_BITS);
+  v6 = _mm256_srai_epi32(v6, DCT_CONST_BITS);
+  v7 = _mm256_srai_epi32(v7, DCT_CONST_BITS);
+
+  v10 = _mm256_srai_epi32(v10, DCT_CONST_BITS);
+  v11 = _mm256_srai_epi32(v11, DCT_CONST_BITS);
+  v14 = _mm256_srai_epi32(v14, DCT_CONST_BITS);
+  v15 = _mm256_srai_epi32(v15, DCT_CONST_BITS);
+
+  in[7] = _mm256_packs_epi32(u2, v2);
+  in[8] = _mm256_packs_epi32(u3, v3);
+
+  in[4] = _mm256_packs_epi32(u6, v6);
+  in[11] = _mm256_packs_epi32(u7, v7);
+
+  in[6] = _mm256_packs_epi32(u10, v10);
+  in[9] = _mm256_packs_epi32(u11, v11);
+
+  in[5] = _mm256_packs_epi32(u14, v14);
+  in[10] = _mm256_packs_epi32(u15, v15);
+
+  in[1] = _mm256_sub_epi16(zero, x8);
+  in[3] = _mm256_sub_epi16(zero, x4);
+  in[13] = _mm256_sub_epi16(zero, x13);
+  in[15] = _mm256_sub_epi16(zero, x1);
+}
+
+#if CONFIG_EXT_TX
+static void fidtx16_avx2(__m256i *in) { txfm_scaling16_avx2(Sqrt2, in); }
+#endif
+
+void av1_fht16x16_avx2(const int16_t *input, tran_low_t *output, int stride,
+                       int tx_type) {
+  __m256i in[16];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      load_buffer_16x16(input, stride, 0, 0, in);
+      fdct16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fdct16_avx2(in);
+      break;
+    case ADST_DCT:
+      load_buffer_16x16(input, stride, 0, 0, in);
+      fadst16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fdct16_avx2(in);
+      break;
+    case DCT_ADST:
+      load_buffer_16x16(input, stride, 0, 0, in);
+      fdct16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fadst16_avx2(in);
+      break;
+    case ADST_ADST:
+      load_buffer_16x16(input, stride, 0, 0, in);
+      fadst16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fadst16_avx2(in);
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      load_buffer_16x16(input, stride, 1, 0, in);
+      fadst16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fdct16_avx2(in);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_16x16(input, stride, 0, 1, in);
+      fdct16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fadst16_avx2(in);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_16x16(input, stride, 1, 1, in);
+      fadst16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fadst16_avx2(in);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_16x16(input, stride, 0, 1, in);
+      fadst16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fadst16_avx2(in);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_16x16(input, stride, 1, 0, in);
+      fadst16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fadst16_avx2(in);
+      break;
+    case IDTX:
+      load_buffer_16x16(input, stride, 0, 0, in);
+      fidtx16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fidtx16_avx2(in);
+      break;
+    case V_DCT:
+      load_buffer_16x16(input, stride, 0, 0, in);
+      fdct16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fidtx16_avx2(in);
+      break;
+    case H_DCT:
+      load_buffer_16x16(input, stride, 0, 0, in);
+      fidtx16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fdct16_avx2(in);
+      break;
+    case V_ADST:
+      load_buffer_16x16(input, stride, 0, 0, in);
+      fadst16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fidtx16_avx2(in);
+      break;
+    case H_ADST:
+      load_buffer_16x16(input, stride, 0, 0, in);
+      fidtx16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fadst16_avx2(in);
+      break;
+    case V_FLIPADST:
+      load_buffer_16x16(input, stride, 1, 0, in);
+      fadst16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fidtx16_avx2(in);
+      break;
+    case H_FLIPADST:
+      load_buffer_16x16(input, stride, 0, 1, in);
+      fidtx16_avx2(in);
+      mm256_transpose_16x16(in);
+      right_shift_16x16(in);
+      fadst16_avx2(in);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+  mm256_transpose_16x16(in);
+  write_buffer_16x16(in, output);
+  _mm256_zeroupper();
+}
+
+void aom_fdct32x32_1_avx2(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  // left and upper corner
+  int32_t sum = get_16x16_sum(input, stride);
+  // right and upper corner
+  sum += get_16x16_sum(input + 16, stride);
+  // left and lower corner
+  sum += get_16x16_sum(input + (stride << 4), stride);
+  // right and lower corner
+  sum += get_16x16_sum(input + (stride << 4) + 16, stride);
+
+  sum >>= 3;
+  output[0] = (tran_low_t)sum;
+  _mm256_zeroupper();
+}
+
+static void mm256_vectors_swap(__m256i *a0, __m256i *a1, const int size) {
+  int i = 0;
+  __m256i temp;
+  while (i < size) {
+    temp = a0[i];
+    a0[i] = a1[i];
+    a1[i] = temp;
+    i++;
+  }
+}
+
+static void mm256_transpose_32x32(__m256i *in0, __m256i *in1) {
+  mm256_transpose_16x16(in0);
+  mm256_transpose_16x16(&in0[16]);
+  mm256_transpose_16x16(in1);
+  mm256_transpose_16x16(&in1[16]);
+  mm256_vectors_swap(&in0[16], in1, 16);
+}
+
+static void prepare_16x16_even(const __m256i *in, __m256i *even) {
+  even[0] = _mm256_add_epi16(in[0], in[31]);
+  even[1] = _mm256_add_epi16(in[1], in[30]);
+  even[2] = _mm256_add_epi16(in[2], in[29]);
+  even[3] = _mm256_add_epi16(in[3], in[28]);
+  even[4] = _mm256_add_epi16(in[4], in[27]);
+  even[5] = _mm256_add_epi16(in[5], in[26]);
+  even[6] = _mm256_add_epi16(in[6], in[25]);
+  even[7] = _mm256_add_epi16(in[7], in[24]);
+  even[8] = _mm256_add_epi16(in[8], in[23]);
+  even[9] = _mm256_add_epi16(in[9], in[22]);
+  even[10] = _mm256_add_epi16(in[10], in[21]);
+  even[11] = _mm256_add_epi16(in[11], in[20]);
+  even[12] = _mm256_add_epi16(in[12], in[19]);
+  even[13] = _mm256_add_epi16(in[13], in[18]);
+  even[14] = _mm256_add_epi16(in[14], in[17]);
+  even[15] = _mm256_add_epi16(in[15], in[16]);
+}
+
+static void prepare_16x16_odd(const __m256i *in, __m256i *odd) {
+  odd[0] = _mm256_sub_epi16(in[15], in[16]);
+  odd[1] = _mm256_sub_epi16(in[14], in[17]);
+  odd[2] = _mm256_sub_epi16(in[13], in[18]);
+  odd[3] = _mm256_sub_epi16(in[12], in[19]);
+  odd[4] = _mm256_sub_epi16(in[11], in[20]);
+  odd[5] = _mm256_sub_epi16(in[10], in[21]);
+  odd[6] = _mm256_sub_epi16(in[9], in[22]);
+  odd[7] = _mm256_sub_epi16(in[8], in[23]);
+  odd[8] = _mm256_sub_epi16(in[7], in[24]);
+  odd[9] = _mm256_sub_epi16(in[6], in[25]);
+  odd[10] = _mm256_sub_epi16(in[5], in[26]);
+  odd[11] = _mm256_sub_epi16(in[4], in[27]);
+  odd[12] = _mm256_sub_epi16(in[3], in[28]);
+  odd[13] = _mm256_sub_epi16(in[2], in[29]);
+  odd[14] = _mm256_sub_epi16(in[1], in[30]);
+  odd[15] = _mm256_sub_epi16(in[0], in[31]);
+}
+
+static void collect_16col(const __m256i *even, const __m256i *odd,
+                          __m256i *out) {
+  // fdct16_avx2() already maps the output
+  out[0] = even[0];
+  out[2] = even[1];
+  out[4] = even[2];
+  out[6] = even[3];
+  out[8] = even[4];
+  out[10] = even[5];
+  out[12] = even[6];
+  out[14] = even[7];
+  out[16] = even[8];
+  out[18] = even[9];
+  out[20] = even[10];
+  out[22] = even[11];
+  out[24] = even[12];
+  out[26] = even[13];
+  out[28] = even[14];
+  out[30] = even[15];
+
+  out[1] = odd[0];
+  out[17] = odd[1];
+  out[9] = odd[2];
+  out[25] = odd[3];
+  out[5] = odd[4];
+  out[21] = odd[5];
+  out[13] = odd[6];
+  out[29] = odd[7];
+  out[3] = odd[8];
+  out[19] = odd[9];
+  out[11] = odd[10];
+  out[27] = odd[11];
+  out[7] = odd[12];
+  out[23] = odd[13];
+  out[15] = odd[14];
+  out[31] = odd[15];
+}
+
+static void collect_coeffs(const __m256i *first_16col_even,
+                           const __m256i *first_16col_odd,
+                           const __m256i *second_16col_even,
+                           const __m256i *second_16col_odd, __m256i *in0,
+                           __m256i *in1) {
+  collect_16col(first_16col_even, first_16col_odd, in0);
+  collect_16col(second_16col_even, second_16col_odd, in1);
+}
+
+static void fdct16_odd_avx2(__m256i *in) {
+  // sequence: cospi_L_H = pairs(L, H) and L first
+  const __m256i cospi_p16_p16 = pair256_set_epi16(cospi_16_64, cospi_16_64);
+  const __m256i cospi_m16_p16 = pair256_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m256i cospi_m08_p24 = pair256_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m256i cospi_p24_p08 = pair256_set_epi16(cospi_24_64, cospi_8_64);
+  const __m256i cospi_m24_m08 = pair256_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m256i cospi_m04_p28 = pair256_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m256i cospi_p28_p04 = pair256_set_epi16(cospi_28_64, cospi_4_64);
+  const __m256i cospi_m28_m04 = pair256_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m256i cospi_m20_p12 = pair256_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m256i cospi_p12_p20 = pair256_set_epi16(cospi_12_64, cospi_20_64);
+  const __m256i cospi_m12_m20 = pair256_set_epi16(-cospi_12_64, -cospi_20_64);
+
+  const __m256i cospi_p31_p01 = pair256_set_epi16(cospi_31_64, cospi_1_64);
+  const __m256i cospi_m01_p31 = pair256_set_epi16(-cospi_1_64, cospi_31_64);
+  const __m256i cospi_p15_p17 = pair256_set_epi16(cospi_15_64, cospi_17_64);
+  const __m256i cospi_m17_p15 = pair256_set_epi16(-cospi_17_64, cospi_15_64);
+  const __m256i cospi_p23_p09 = pair256_set_epi16(cospi_23_64, cospi_9_64);
+  const __m256i cospi_m09_p23 = pair256_set_epi16(-cospi_9_64, cospi_23_64);
+  const __m256i cospi_p07_p25 = pair256_set_epi16(cospi_7_64, cospi_25_64);
+  const __m256i cospi_m25_p07 = pair256_set_epi16(-cospi_25_64, cospi_7_64);
+  const __m256i cospi_p27_p05 = pair256_set_epi16(cospi_27_64, cospi_5_64);
+  const __m256i cospi_m05_p27 = pair256_set_epi16(-cospi_5_64, cospi_27_64);
+  const __m256i cospi_p11_p21 = pair256_set_epi16(cospi_11_64, cospi_21_64);
+  const __m256i cospi_m21_p11 = pair256_set_epi16(-cospi_21_64, cospi_11_64);
+  const __m256i cospi_p19_p13 = pair256_set_epi16(cospi_19_64, cospi_13_64);
+  const __m256i cospi_m13_p19 = pair256_set_epi16(-cospi_13_64, cospi_19_64);
+  const __m256i cospi_p03_p29 = pair256_set_epi16(cospi_3_64, cospi_29_64);
+  const __m256i cospi_m29_p03 = pair256_set_epi16(-cospi_29_64, cospi_3_64);
+
+  __m256i x0, x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, x15;
+  __m256i y0, y1, y2, y3, y4, y5, y6, y7, y8, y9, y10, y11, y12, y13, y14, y15;
+  __m256i u0, u1;
+
+  // stage 1 is in prepare_16x16_odd()
+
+  // stage 2
+  y0 = in[0];
+  y1 = in[1];
+  y2 = in[2];
+  y3 = in[3];
+
+  u0 = _mm256_unpacklo_epi16(in[4], in[11]);
+  u1 = _mm256_unpackhi_epi16(in[4], in[11]);
+  y4 = butter_fly(u0, u1, cospi_m16_p16);
+  y11 = butter_fly(u0, u1, cospi_p16_p16);
+
+  u0 = _mm256_unpacklo_epi16(in[5], in[10]);
+  u1 = _mm256_unpackhi_epi16(in[5], in[10]);
+  y5 = butter_fly(u0, u1, cospi_m16_p16);
+  y10 = butter_fly(u0, u1, cospi_p16_p16);
+
+  u0 = _mm256_unpacklo_epi16(in[6], in[9]);
+  u1 = _mm256_unpackhi_epi16(in[6], in[9]);
+  y6 = butter_fly(u0, u1, cospi_m16_p16);
+  y9 = butter_fly(u0, u1, cospi_p16_p16);
+
+  u0 = _mm256_unpacklo_epi16(in[7], in[8]);
+  u1 = _mm256_unpackhi_epi16(in[7], in[8]);
+  y7 = butter_fly(u0, u1, cospi_m16_p16);
+  y8 = butter_fly(u0, u1, cospi_p16_p16);
+
+  y12 = in[12];
+  y13 = in[13];
+  y14 = in[14];
+  y15 = in[15];
+
+  // stage 3
+  x0 = _mm256_add_epi16(y0, y7);
+  x1 = _mm256_add_epi16(y1, y6);
+  x2 = _mm256_add_epi16(y2, y5);
+  x3 = _mm256_add_epi16(y3, y4);
+  x4 = _mm256_sub_epi16(y3, y4);
+  x5 = _mm256_sub_epi16(y2, y5);
+  x6 = _mm256_sub_epi16(y1, y6);
+  x7 = _mm256_sub_epi16(y0, y7);
+  x8 = _mm256_sub_epi16(y15, y8);
+  x9 = _mm256_sub_epi16(y14, y9);
+  x10 = _mm256_sub_epi16(y13, y10);
+  x11 = _mm256_sub_epi16(y12, y11);
+  x12 = _mm256_add_epi16(y12, y11);
+  x13 = _mm256_add_epi16(y13, y10);
+  x14 = _mm256_add_epi16(y14, y9);
+  x15 = _mm256_add_epi16(y15, y8);
+
+  // stage 4
+  y0 = x0;
+  y1 = x1;
+  y6 = x6;
+  y7 = x7;
+  y8 = x8;
+  y9 = x9;
+  y14 = x14;
+  y15 = x15;
+
+  u0 = _mm256_unpacklo_epi16(x2, x13);
+  u1 = _mm256_unpackhi_epi16(x2, x13);
+  y2 = butter_fly(u0, u1, cospi_m08_p24);
+  y13 = butter_fly(u0, u1, cospi_p24_p08);
+
+  u0 = _mm256_unpacklo_epi16(x3, x12);
+  u1 = _mm256_unpackhi_epi16(x3, x12);
+  y3 = butter_fly(u0, u1, cospi_m08_p24);
+  y12 = butter_fly(u0, u1, cospi_p24_p08);
+
+  u0 = _mm256_unpacklo_epi16(x4, x11);
+  u1 = _mm256_unpackhi_epi16(x4, x11);
+  y4 = butter_fly(u0, u1, cospi_m24_m08);
+  y11 = butter_fly(u0, u1, cospi_m08_p24);
+
+  u0 = _mm256_unpacklo_epi16(x5, x10);
+  u1 = _mm256_unpackhi_epi16(x5, x10);
+  y5 = butter_fly(u0, u1, cospi_m24_m08);
+  y10 = butter_fly(u0, u1, cospi_m08_p24);
+
+  // stage 5
+  x0 = _mm256_add_epi16(y0, y3);
+  x1 = _mm256_add_epi16(y1, y2);
+  x2 = _mm256_sub_epi16(y1, y2);
+  x3 = _mm256_sub_epi16(y0, y3);
+  x4 = _mm256_sub_epi16(y7, y4);
+  x5 = _mm256_sub_epi16(y6, y5);
+  x6 = _mm256_add_epi16(y6, y5);
+  x7 = _mm256_add_epi16(y7, y4);
+
+  x8 = _mm256_add_epi16(y8, y11);
+  x9 = _mm256_add_epi16(y9, y10);
+  x10 = _mm256_sub_epi16(y9, y10);
+  x11 = _mm256_sub_epi16(y8, y11);
+  x12 = _mm256_sub_epi16(y15, y12);
+  x13 = _mm256_sub_epi16(y14, y13);
+  x14 = _mm256_add_epi16(y14, y13);
+  x15 = _mm256_add_epi16(y15, y12);
+
+  // stage 6
+  y0 = x0;
+  y3 = x3;
+  y4 = x4;
+  y7 = x7;
+  y8 = x8;
+  y11 = x11;
+  y12 = x12;
+  y15 = x15;
+
+  u0 = _mm256_unpacklo_epi16(x1, x14);
+  u1 = _mm256_unpackhi_epi16(x1, x14);
+  y1 = butter_fly(u0, u1, cospi_m04_p28);
+  y14 = butter_fly(u0, u1, cospi_p28_p04);
+
+  u0 = _mm256_unpacklo_epi16(x2, x13);
+  u1 = _mm256_unpackhi_epi16(x2, x13);
+  y2 = butter_fly(u0, u1, cospi_m28_m04);
+  y13 = butter_fly(u0, u1, cospi_m04_p28);
+
+  u0 = _mm256_unpacklo_epi16(x5, x10);
+  u1 = _mm256_unpackhi_epi16(x5, x10);
+  y5 = butter_fly(u0, u1, cospi_m20_p12);
+  y10 = butter_fly(u0, u1, cospi_p12_p20);
+
+  u0 = _mm256_unpacklo_epi16(x6, x9);
+  u1 = _mm256_unpackhi_epi16(x6, x9);
+  y6 = butter_fly(u0, u1, cospi_m12_m20);
+  y9 = butter_fly(u0, u1, cospi_m20_p12);
+
+  // stage 7
+  x0 = _mm256_add_epi16(y0, y1);
+  x1 = _mm256_sub_epi16(y0, y1);
+  x2 = _mm256_sub_epi16(y3, y2);
+  x3 = _mm256_add_epi16(y3, y2);
+  x4 = _mm256_add_epi16(y4, y5);
+  x5 = _mm256_sub_epi16(y4, y5);
+  x6 = _mm256_sub_epi16(y7, y6);
+  x7 = _mm256_add_epi16(y7, y6);
+
+  x8 = _mm256_add_epi16(y8, y9);
+  x9 = _mm256_sub_epi16(y8, y9);
+  x10 = _mm256_sub_epi16(y11, y10);
+  x11 = _mm256_add_epi16(y11, y10);
+  x12 = _mm256_add_epi16(y12, y13);
+  x13 = _mm256_sub_epi16(y12, y13);
+  x14 = _mm256_sub_epi16(y15, y14);
+  x15 = _mm256_add_epi16(y15, y14);
+
+  // stage 8
+  u0 = _mm256_unpacklo_epi16(x0, x15);
+  u1 = _mm256_unpackhi_epi16(x0, x15);
+  in[0] = butter_fly(u0, u1, cospi_p31_p01);
+  in[15] = butter_fly(u0, u1, cospi_m01_p31);
+
+  u0 = _mm256_unpacklo_epi16(x1, x14);
+  u1 = _mm256_unpackhi_epi16(x1, x14);
+  in[1] = butter_fly(u0, u1, cospi_p15_p17);
+  in[14] = butter_fly(u0, u1, cospi_m17_p15);
+
+  u0 = _mm256_unpacklo_epi16(x2, x13);
+  u1 = _mm256_unpackhi_epi16(x2, x13);
+  in[2] = butter_fly(u0, u1, cospi_p23_p09);
+  in[13] = butter_fly(u0, u1, cospi_m09_p23);
+
+  u0 = _mm256_unpacklo_epi16(x3, x12);
+  u1 = _mm256_unpackhi_epi16(x3, x12);
+  in[3] = butter_fly(u0, u1, cospi_p07_p25);
+  in[12] = butter_fly(u0, u1, cospi_m25_p07);
+
+  u0 = _mm256_unpacklo_epi16(x4, x11);
+  u1 = _mm256_unpackhi_epi16(x4, x11);
+  in[4] = butter_fly(u0, u1, cospi_p27_p05);
+  in[11] = butter_fly(u0, u1, cospi_m05_p27);
+
+  u0 = _mm256_unpacklo_epi16(x5, x10);
+  u1 = _mm256_unpackhi_epi16(x5, x10);
+  in[5] = butter_fly(u0, u1, cospi_p11_p21);
+  in[10] = butter_fly(u0, u1, cospi_m21_p11);
+
+  u0 = _mm256_unpacklo_epi16(x6, x9);
+  u1 = _mm256_unpackhi_epi16(x6, x9);
+  in[6] = butter_fly(u0, u1, cospi_p19_p13);
+  in[9] = butter_fly(u0, u1, cospi_m13_p19);
+
+  u0 = _mm256_unpacklo_epi16(x7, x8);
+  u1 = _mm256_unpackhi_epi16(x7, x8);
+  in[7] = butter_fly(u0, u1, cospi_p03_p29);
+  in[8] = butter_fly(u0, u1, cospi_m29_p03);
+}
+
+static void fdct32_avx2(__m256i *in0, __m256i *in1) {
+  __m256i even0[16], even1[16], odd0[16], odd1[16];
+  prepare_16x16_even(in0, even0);
+  fdct16_avx2(even0);
+
+  prepare_16x16_odd(in0, odd0);
+  fdct16_odd_avx2(odd0);
+
+  prepare_16x16_even(in1, even1);
+  fdct16_avx2(even1);
+
+  prepare_16x16_odd(in1, odd1);
+  fdct16_odd_avx2(odd1);
+
+  collect_coeffs(even0, odd0, even1, odd1, in0, in1);
+
+  mm256_transpose_32x32(in0, in1);
+}
+
+static INLINE void write_buffer_32x32(const __m256i *in0, const __m256i *in1,
+                                      tran_low_t *output) {
+  int i = 0;
+  const int stride = 32;
+  tran_low_t *coeff = output;
+  while (i < 32) {
+    storeu_output_avx2(&in0[i], coeff);
+    storeu_output_avx2(&in1[i], coeff + 16);
+    coeff += stride;
+    i += 1;
+  }
+}
+
+#if CONFIG_EXT_TX
+static void fhalfright32_16col_avx2(__m256i *in) {
+  int i = 0;
+  const __m256i zero = _mm256_setzero_si256();
+  const __m256i sqrt2 = _mm256_set1_epi16(Sqrt2);
+  const __m256i dct_rounding = _mm256_set1_epi32(DCT_CONST_ROUNDING);
+  __m256i x0, x1;
+
+  while (i < 16) {
+    in[i] = _mm256_slli_epi16(in[i], 2);
+    x0 = _mm256_unpacklo_epi16(in[i + 16], zero);
+    x1 = _mm256_unpackhi_epi16(in[i + 16], zero);
+    x0 = _mm256_madd_epi16(x0, sqrt2);
+    x1 = _mm256_madd_epi16(x1, sqrt2);
+    x0 = _mm256_add_epi32(x0, dct_rounding);
+    x1 = _mm256_add_epi32(x1, dct_rounding);
+    x0 = _mm256_srai_epi32(x0, DCT_CONST_BITS);
+    x1 = _mm256_srai_epi32(x1, DCT_CONST_BITS);
+    in[i + 16] = _mm256_packs_epi32(x0, x1);
+    i += 1;
+  }
+  fdct16_avx2(&in[16]);
+}
+
+static void fhalfright32_avx2(__m256i *in0, __m256i *in1) {
+  fhalfright32_16col_avx2(in0);
+  fhalfright32_16col_avx2(in1);
+  mm256_vectors_swap(in0, &in0[16], 16);
+  mm256_vectors_swap(in1, &in1[16], 16);
+  mm256_transpose_32x32(in0, in1);
+}
+#endif  // CONFIG_EXT_TX
+
+static INLINE void load_buffer_32x32(const int16_t *input, int stride,
+                                     int flipud, int fliplr, __m256i *in0,
+                                     __m256i *in1) {
+  // Load 4 16x16 blocks
+  const int16_t *topL = input;
+  const int16_t *topR = input + 16;
+  const int16_t *botL = input + 16 * stride;
+  const int16_t *botR = input + 16 * stride + 16;
+
+  const int16_t *tmp;
+
+  if (flipud) {
+    // Swap left columns
+    tmp = topL;
+    topL = botL;
+    botL = tmp;
+    // Swap right columns
+    tmp = topR;
+    topR = botR;
+    botR = tmp;
+  }
+
+  if (fliplr) {
+    // Swap top rows
+    tmp = topL;
+    topL = topR;
+    topR = tmp;
+    // Swap bottom rows
+    tmp = botL;
+    botL = botR;
+    botR = tmp;
+  }
+
+  // load first 16 columns
+  load_buffer_16x16(topL, stride, flipud, fliplr, in0);
+  load_buffer_16x16(botL, stride, flipud, fliplr, in0 + 16);
+
+  // load second 16 columns
+  load_buffer_16x16(topR, stride, flipud, fliplr, in1);
+  load_buffer_16x16(botR, stride, flipud, fliplr, in1 + 16);
+}
+
+static INLINE void right_shift_32x32_16col(int bit, __m256i *in) {
+  int i = 0;
+  const __m256i rounding = _mm256_set1_epi16((1 << bit) >> 1);
+  __m256i sign;
+  while (i < 32) {
+    sign = _mm256_srai_epi16(in[i], 15);
+    in[i] = _mm256_add_epi16(in[i], rounding);
+    in[i] = _mm256_add_epi16(in[i], sign);
+    in[i] = _mm256_srai_epi16(in[i], bit);
+    i += 1;
+  }
+}
+
+// Positive rounding
+static INLINE void right_shift_32x32(__m256i *in0, __m256i *in1) {
+  const int bit = 4;
+  right_shift_32x32_16col(bit, in0);
+  right_shift_32x32_16col(bit, in1);
+}
+
+#if CONFIG_EXT_TX
+static void fidtx32_avx2(__m256i *in0, __m256i *in1) {
+  int i = 0;
+  while (i < 32) {
+    in0[i] = _mm256_slli_epi16(in0[i], 2);
+    in1[i] = _mm256_slli_epi16(in1[i], 2);
+    i += 1;
+  }
+  mm256_transpose_32x32(in0, in1);
+}
+#endif
+
+void av1_fht32x32_avx2(const int16_t *input, tran_low_t *output, int stride,
+                       int tx_type) {
+  __m256i in0[32];  // left 32 columns
+  __m256i in1[32];  // right 32 columns
+
+  switch (tx_type) {
+    case DCT_DCT:
+      load_buffer_32x32(input, stride, 0, 0, in0, in1);
+      fdct32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fdct32_avx2(in0, in1);
+      break;
+#if CONFIG_EXT_TX
+    case ADST_DCT:
+      load_buffer_32x32(input, stride, 0, 0, in0, in1);
+      fhalfright32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fdct32_avx2(in0, in1);
+      break;
+    case DCT_ADST:
+      load_buffer_32x32(input, stride, 0, 0, in0, in1);
+      fdct32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fhalfright32_avx2(in0, in1);
+      break;
+    case ADST_ADST:
+      load_buffer_32x32(input, stride, 0, 0, in0, in1);
+      fhalfright32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fhalfright32_avx2(in0, in1);
+      break;
+    case FLIPADST_DCT:
+      load_buffer_32x32(input, stride, 1, 0, in0, in1);
+      fhalfright32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fdct32_avx2(in0, in1);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_32x32(input, stride, 0, 1, in0, in1);
+      fdct32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fhalfright32_avx2(in0, in1);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_32x32(input, stride, 1, 1, in0, in1);
+      fhalfright32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fhalfright32_avx2(in0, in1);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_32x32(input, stride, 0, 1, in0, in1);
+      fhalfright32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fhalfright32_avx2(in0, in1);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_32x32(input, stride, 1, 0, in0, in1);
+      fhalfright32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fhalfright32_avx2(in0, in1);
+      break;
+    case IDTX:
+      load_buffer_32x32(input, stride, 0, 0, in0, in1);
+      fidtx32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fidtx32_avx2(in0, in1);
+      break;
+    case V_DCT:
+      load_buffer_32x32(input, stride, 0, 0, in0, in1);
+      fdct32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fidtx32_avx2(in0, in1);
+      break;
+    case H_DCT:
+      load_buffer_32x32(input, stride, 0, 0, in0, in1);
+      fidtx32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fdct32_avx2(in0, in1);
+      break;
+    case V_ADST:
+      load_buffer_32x32(input, stride, 0, 0, in0, in1);
+      fhalfright32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fidtx32_avx2(in0, in1);
+      break;
+    case H_ADST:
+      load_buffer_32x32(input, stride, 0, 0, in0, in1);
+      fidtx32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fhalfright32_avx2(in0, in1);
+      break;
+    case V_FLIPADST:
+      load_buffer_32x32(input, stride, 1, 0, in0, in1);
+      fhalfright32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fidtx32_avx2(in0, in1);
+      break;
+    case H_FLIPADST:
+      load_buffer_32x32(input, stride, 0, 1, in0, in1);
+      fidtx32_avx2(in0, in1);
+      right_shift_32x32(in0, in1);
+      fhalfright32_avx2(in0, in1);
+      break;
+#endif  // CONFIG_EXT_TX
+    default: assert(0); break;
+  }
+  write_buffer_32x32(in0, in1, output);
+  _mm256_zeroupper();
+}
diff --git a/third_party/aom/av1/encoder/x86/temporal_filter_apply_sse2.asm b/third_party/aom/av1/encoder/x86/temporal_filter_apply_sse2.asm
new file mode 100644
index 0000000000..7186b6b924
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/temporal_filter_apply_sse2.asm
@@ -0,0 +1,215 @@
+;
+; Copyright (c) 2016, Alliance for Open Media. All rights reserved
+;
+; This source code is subject to the terms of the BSD 2 Clause License and
+; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+; was not distributed with this source code in the LICENSE file, you can
+; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+; Media Patent License 1.0 was not distributed with this source code in the
+; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+;
+
+;
+
+
+%include "aom_ports/x86_abi_support.asm"
+
+; void av1_temporal_filter_apply_sse2 | arg
+;  (unsigned char  *frame1,           |  0
+;   unsigned int    stride,           |  1
+;   unsigned char  *frame2,           |  2
+;   unsigned int    block_width,      |  3
+;   unsigned int    block_height,     |  4
+;   int             strength,         |  5
+;   int             filter_weight,    |  6
+;   unsigned int   *accumulator,      |  7
+;   unsigned short *count)            |  8
+global sym(av1_temporal_filter_apply_sse2) PRIVATE
+sym(av1_temporal_filter_apply_sse2):
+
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ALIGN_STACK 16, rax
+    %define block_width    0
+    %define block_height  16
+    %define strength      32
+    %define filter_weight 48
+    %define rounding_bit  64
+    %define rbp_backup    80
+    %define stack_size    96
+    sub         rsp,           stack_size
+    mov         [rsp + rbp_backup], rbp
+    ; end prolog
+
+        mov         edx,            arg(3)
+        mov         [rsp + block_width], rdx
+        mov         edx,            arg(4)
+        mov         [rsp + block_height], rdx
+        movd        xmm6,           arg(5)
+        movdqa      [rsp + strength], xmm6 ; where strength is used, all 16 bytes are read
+
+        ; calculate the rounding bit outside the loop
+        ; 0x8000 >> (16 - strength)
+        mov         rdx,            16
+        sub         rdx,            arg(5) ; 16 - strength
+        movq        xmm4,           rdx    ; can't use rdx w/ shift
+        movdqa      xmm5,           [GLOBAL(_const_top_bit)]
+        psrlw       xmm5,           xmm4
+        movdqa      [rsp + rounding_bit], xmm5
+
+        mov         rsi,            arg(0) ; src/frame1
+        mov         rdx,            arg(2) ; predictor frame
+        mov         rdi,            arg(7) ; accumulator
+        mov         rax,            arg(8) ; count
+
+        ; dup the filter weight and store for later
+        movd        xmm0,           arg(6) ; filter_weight
+        pshuflw     xmm0,           xmm0, 0
+        punpcklwd   xmm0,           xmm0
+        movdqa      [rsp + filter_weight], xmm0
+
+        mov         rbp,            arg(1) ; stride
+        pxor        xmm7,           xmm7   ; zero for extraction
+
+        mov         rcx,            [rsp + block_width]
+        imul        rcx,            [rsp + block_height]
+        add         rcx,            rdx
+        cmp         dword ptr [rsp + block_width], 8
+        jne         .temporal_filter_apply_load_16
+
+.temporal_filter_apply_load_8:
+        movq        xmm0,           [rsi]  ; first row
+        lea         rsi,            [rsi + rbp] ; += stride
+        punpcklbw   xmm0,           xmm7   ; src[ 0- 7]
+        movq        xmm1,           [rsi]  ; second row
+        lea         rsi,            [rsi + rbp] ; += stride
+        punpcklbw   xmm1,           xmm7   ; src[ 8-15]
+        jmp         .temporal_filter_apply_load_finished
+
+.temporal_filter_apply_load_16:
+        movdqa      xmm0,           [rsi]  ; src (frame1)
+        lea         rsi,            [rsi + rbp] ; += stride
+        movdqa      xmm1,           xmm0
+        punpcklbw   xmm0,           xmm7   ; src[ 0- 7]
+        punpckhbw   xmm1,           xmm7   ; src[ 8-15]
+
+.temporal_filter_apply_load_finished:
+        movdqa      xmm2,           [rdx]  ; predictor (frame2)
+        movdqa      xmm3,           xmm2
+        punpcklbw   xmm2,           xmm7   ; pred[ 0- 7]
+        punpckhbw   xmm3,           xmm7   ; pred[ 8-15]
+
+        ; modifier = src_byte - pixel_value
+        psubw       xmm0,           xmm2   ; src - pred[ 0- 7]
+        psubw       xmm1,           xmm3   ; src - pred[ 8-15]
+
+        ; modifier *= modifier
+        pmullw      xmm0,           xmm0   ; modifer[ 0- 7]^2
+        pmullw      xmm1,           xmm1   ; modifer[ 8-15]^2
+
+        ; modifier *= 3
+        pmullw      xmm0,           [GLOBAL(_const_3w)]
+        pmullw      xmm1,           [GLOBAL(_const_3w)]
+
+        ; modifer += 0x8000 >> (16 - strength)
+        paddw       xmm0,           [rsp + rounding_bit]
+        paddw       xmm1,           [rsp + rounding_bit]
+
+        ; modifier >>= strength
+        psrlw       xmm0,           [rsp + strength]
+        psrlw       xmm1,           [rsp + strength]
+
+        ; modifier = 16 - modifier
+        ; saturation takes care of modifier > 16
+        movdqa      xmm3,           [GLOBAL(_const_16w)]
+        movdqa      xmm2,           [GLOBAL(_const_16w)]
+        psubusw     xmm3,           xmm1
+        psubusw     xmm2,           xmm0
+
+        ; modifier *= filter_weight
+        pmullw      xmm2,           [rsp + filter_weight]
+        pmullw      xmm3,           [rsp + filter_weight]
+
+        ; count
+        movdqa      xmm4,           [rax]
+        movdqa      xmm5,           [rax+16]
+        ; += modifier
+        paddw       xmm4,           xmm2
+        paddw       xmm5,           xmm3
+        ; write back
+        movdqa      [rax],          xmm4
+        movdqa      [rax+16],       xmm5
+        lea         rax,            [rax + 16*2] ; count += 16*(sizeof(short))
+
+        ; load and extract the predictor up to shorts
+        pxor        xmm7,           xmm7
+        movdqa      xmm0,           [rdx]
+        lea         rdx,            [rdx + 16*1] ; pred += 16*(sizeof(char))
+        movdqa      xmm1,           xmm0
+        punpcklbw   xmm0,           xmm7   ; pred[ 0- 7]
+        punpckhbw   xmm1,           xmm7   ; pred[ 8-15]
+
+        ; modifier *= pixel_value
+        pmullw      xmm0,           xmm2
+        pmullw      xmm1,           xmm3
+
+        ; expand to double words
+        movdqa      xmm2,           xmm0
+        punpcklwd   xmm0,           xmm7   ; [ 0- 3]
+        punpckhwd   xmm2,           xmm7   ; [ 4- 7]
+        movdqa      xmm3,           xmm1
+        punpcklwd   xmm1,           xmm7   ; [ 8-11]
+        punpckhwd   xmm3,           xmm7   ; [12-15]
+
+        ; accumulator
+        movdqa      xmm4,           [rdi]
+        movdqa      xmm5,           [rdi+16]
+        movdqa      xmm6,           [rdi+32]
+        movdqa      xmm7,           [rdi+48]
+        ; += modifier
+        paddd       xmm4,           xmm0
+        paddd       xmm5,           xmm2
+        paddd       xmm6,           xmm1
+        paddd       xmm7,           xmm3
+        ; write back
+        movdqa      [rdi],          xmm4
+        movdqa      [rdi+16],       xmm5
+        movdqa      [rdi+32],       xmm6
+        movdqa      [rdi+48],       xmm7
+        lea         rdi,            [rdi + 16*4] ; accumulator += 16*(sizeof(int))
+
+        cmp         rdx,            rcx
+        je          .temporal_filter_apply_epilog
+        pxor        xmm7,           xmm7   ; zero for extraction
+        cmp         dword ptr [rsp + block_width], 16
+        je          .temporal_filter_apply_load_16
+        jmp         .temporal_filter_apply_load_8
+
+.temporal_filter_apply_epilog:
+    ; begin epilog
+    mov         rbp,            [rsp + rbp_backup]
+    add         rsp,            stack_size
+    pop         rsp
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+_const_3w:
+    times 8 dw 3
+align 16
+_const_top_bit:
+    times 8 dw 1<<15
+align 16
+_const_16w:
+    times 8 dw 16
diff --git a/third_party/aom/av1/encoder/x86/wedge_utils_sse2.c b/third_party/aom/av1/encoder/x86/wedge_utils_sse2.c
new file mode 100644
index 0000000000..bf233ca4d9
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/wedge_utils_sse2.c
@@ -0,0 +1,254 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <immintrin.h>
+
+#include "aom_dsp/x86/synonyms.h"
+
+#include "aom/aom_integer.h"
+
+#include "av1/common/reconinter.h"
+
+#define MAX_MASK_VALUE (1 << WEDGE_WEIGHT_BITS)
+
+/**
+ * See av1_wedge_sse_from_residuals_c
+ */
+uint64_t av1_wedge_sse_from_residuals_sse2(const int16_t *r1, const int16_t *d,
+                                           const uint8_t *m, int N) {
+  int n = -N;
+  int n8 = n + 8;
+
+  uint64_t csse;
+
+  const __m128i v_mask_max_w = _mm_set1_epi16(MAX_MASK_VALUE);
+  const __m128i v_zext_q = _mm_set_epi32(0, 0xffffffff, 0, 0xffffffff);
+
+  __m128i v_acc0_q = _mm_setzero_si128();
+
+  assert(N % 64 == 0);
+
+  r1 += N;
+  d += N;
+  m += N;
+
+  do {
+    const __m128i v_r0_w = xx_load_128(r1 + n);
+    const __m128i v_r1_w = xx_load_128(r1 + n8);
+    const __m128i v_d0_w = xx_load_128(d + n);
+    const __m128i v_d1_w = xx_load_128(d + n8);
+    const __m128i v_m01_b = xx_load_128(m + n);
+
+    const __m128i v_rd0l_w = _mm_unpacklo_epi16(v_d0_w, v_r0_w);
+    const __m128i v_rd0h_w = _mm_unpackhi_epi16(v_d0_w, v_r0_w);
+    const __m128i v_rd1l_w = _mm_unpacklo_epi16(v_d1_w, v_r1_w);
+    const __m128i v_rd1h_w = _mm_unpackhi_epi16(v_d1_w, v_r1_w);
+    const __m128i v_m0_w = _mm_unpacklo_epi8(v_m01_b, _mm_setzero_si128());
+    const __m128i v_m1_w = _mm_unpackhi_epi8(v_m01_b, _mm_setzero_si128());
+
+    const __m128i v_m0l_w = _mm_unpacklo_epi16(v_m0_w, v_mask_max_w);
+    const __m128i v_m0h_w = _mm_unpackhi_epi16(v_m0_w, v_mask_max_w);
+    const __m128i v_m1l_w = _mm_unpacklo_epi16(v_m1_w, v_mask_max_w);
+    const __m128i v_m1h_w = _mm_unpackhi_epi16(v_m1_w, v_mask_max_w);
+
+    const __m128i v_t0l_d = _mm_madd_epi16(v_rd0l_w, v_m0l_w);
+    const __m128i v_t0h_d = _mm_madd_epi16(v_rd0h_w, v_m0h_w);
+    const __m128i v_t1l_d = _mm_madd_epi16(v_rd1l_w, v_m1l_w);
+    const __m128i v_t1h_d = _mm_madd_epi16(v_rd1h_w, v_m1h_w);
+
+    const __m128i v_t0_w = _mm_packs_epi32(v_t0l_d, v_t0h_d);
+    const __m128i v_t1_w = _mm_packs_epi32(v_t1l_d, v_t1h_d);
+
+    const __m128i v_sq0_d = _mm_madd_epi16(v_t0_w, v_t0_w);
+    const __m128i v_sq1_d = _mm_madd_epi16(v_t1_w, v_t1_w);
+
+    const __m128i v_sum0_q = _mm_add_epi64(_mm_and_si128(v_sq0_d, v_zext_q),
+                                           _mm_srli_epi64(v_sq0_d, 32));
+    const __m128i v_sum1_q = _mm_add_epi64(_mm_and_si128(v_sq1_d, v_zext_q),
+                                           _mm_srli_epi64(v_sq1_d, 32));
+
+    v_acc0_q = _mm_add_epi64(v_acc0_q, v_sum0_q);
+    v_acc0_q = _mm_add_epi64(v_acc0_q, v_sum1_q);
+
+    n8 += 16;
+    n += 16;
+  } while (n);
+
+  v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_srli_si128(v_acc0_q, 8));
+
+#if ARCH_X86_64
+  csse = (uint64_t)_mm_cvtsi128_si64(v_acc0_q);
+#else
+  xx_storel_64(&csse, v_acc0_q);
+#endif
+
+  return ROUND_POWER_OF_TWO(csse, 2 * WEDGE_WEIGHT_BITS);
+}
+
+/**
+ * See av1_wedge_sign_from_residuals_c
+ */
+int av1_wedge_sign_from_residuals_sse2(const int16_t *ds, const uint8_t *m,
+                                       int N, int64_t limit) {
+  int64_t acc;
+
+  __m128i v_sign_d;
+  __m128i v_acc0_d = _mm_setzero_si128();
+  __m128i v_acc1_d = _mm_setzero_si128();
+  __m128i v_acc_q;
+
+  // Input size limited to 8192 by the use of 32 bit accumulators and m
+  // being between [0, 64]. Overflow might happen at larger sizes,
+  // though it is practically impossible on real video input.
+  assert(N < 8192);
+  assert(N % 64 == 0);
+
+  do {
+    const __m128i v_m01_b = xx_load_128(m);
+    const __m128i v_m23_b = xx_load_128(m + 16);
+    const __m128i v_m45_b = xx_load_128(m + 32);
+    const __m128i v_m67_b = xx_load_128(m + 48);
+
+    const __m128i v_d0_w = xx_load_128(ds);
+    const __m128i v_d1_w = xx_load_128(ds + 8);
+    const __m128i v_d2_w = xx_load_128(ds + 16);
+    const __m128i v_d3_w = xx_load_128(ds + 24);
+    const __m128i v_d4_w = xx_load_128(ds + 32);
+    const __m128i v_d5_w = xx_load_128(ds + 40);
+    const __m128i v_d6_w = xx_load_128(ds + 48);
+    const __m128i v_d7_w = xx_load_128(ds + 56);
+
+    const __m128i v_m0_w = _mm_unpacklo_epi8(v_m01_b, _mm_setzero_si128());
+    const __m128i v_m1_w = _mm_unpackhi_epi8(v_m01_b, _mm_setzero_si128());
+    const __m128i v_m2_w = _mm_unpacklo_epi8(v_m23_b, _mm_setzero_si128());
+    const __m128i v_m3_w = _mm_unpackhi_epi8(v_m23_b, _mm_setzero_si128());
+    const __m128i v_m4_w = _mm_unpacklo_epi8(v_m45_b, _mm_setzero_si128());
+    const __m128i v_m5_w = _mm_unpackhi_epi8(v_m45_b, _mm_setzero_si128());
+    const __m128i v_m6_w = _mm_unpacklo_epi8(v_m67_b, _mm_setzero_si128());
+    const __m128i v_m7_w = _mm_unpackhi_epi8(v_m67_b, _mm_setzero_si128());
+
+    const __m128i v_p0_d = _mm_madd_epi16(v_d0_w, v_m0_w);
+    const __m128i v_p1_d = _mm_madd_epi16(v_d1_w, v_m1_w);
+    const __m128i v_p2_d = _mm_madd_epi16(v_d2_w, v_m2_w);
+    const __m128i v_p3_d = _mm_madd_epi16(v_d3_w, v_m3_w);
+    const __m128i v_p4_d = _mm_madd_epi16(v_d4_w, v_m4_w);
+    const __m128i v_p5_d = _mm_madd_epi16(v_d5_w, v_m5_w);
+    const __m128i v_p6_d = _mm_madd_epi16(v_d6_w, v_m6_w);
+    const __m128i v_p7_d = _mm_madd_epi16(v_d7_w, v_m7_w);
+
+    const __m128i v_p01_d = _mm_add_epi32(v_p0_d, v_p1_d);
+    const __m128i v_p23_d = _mm_add_epi32(v_p2_d, v_p3_d);
+    const __m128i v_p45_d = _mm_add_epi32(v_p4_d, v_p5_d);
+    const __m128i v_p67_d = _mm_add_epi32(v_p6_d, v_p7_d);
+
+    const __m128i v_p0123_d = _mm_add_epi32(v_p01_d, v_p23_d);
+    const __m128i v_p4567_d = _mm_add_epi32(v_p45_d, v_p67_d);
+
+    v_acc0_d = _mm_add_epi32(v_acc0_d, v_p0123_d);
+    v_acc1_d = _mm_add_epi32(v_acc1_d, v_p4567_d);
+
+    ds += 64;
+    m += 64;
+
+    N -= 64;
+  } while (N);
+
+  v_sign_d = _mm_cmplt_epi32(v_acc0_d, _mm_setzero_si128());
+  v_acc0_d = _mm_add_epi64(_mm_unpacklo_epi32(v_acc0_d, v_sign_d),
+                           _mm_unpackhi_epi32(v_acc0_d, v_sign_d));
+
+  v_sign_d = _mm_cmplt_epi32(v_acc1_d, _mm_setzero_si128());
+  v_acc1_d = _mm_add_epi64(_mm_unpacklo_epi32(v_acc1_d, v_sign_d),
+                           _mm_unpackhi_epi32(v_acc1_d, v_sign_d));
+
+  v_acc_q = _mm_add_epi64(v_acc0_d, v_acc1_d);
+
+  v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_si128(v_acc_q, 8));
+
+#if ARCH_X86_64
+  acc = (uint64_t)_mm_cvtsi128_si64(v_acc_q);
+#else
+  xx_storel_64(&acc, v_acc_q);
+#endif
+
+  return acc > limit;
+}
+
+// Negate under mask
+static INLINE __m128i negm_epi16(__m128i v_v_w, __m128i v_mask_w) {
+  return _mm_sub_epi16(_mm_xor_si128(v_v_w, v_mask_w), v_mask_w);
+}
+
+/**
+ * av1_wedge_compute_delta_squares_c
+ */
+void av1_wedge_compute_delta_squares_sse2(int16_t *d, const int16_t *a,
+                                          const int16_t *b, int N) {
+  const __m128i v_neg_w =
+      _mm_set_epi16(0xffff, 0, 0xffff, 0, 0xffff, 0, 0xffff, 0);
+
+  assert(N % 64 == 0);
+
+  do {
+    const __m128i v_a0_w = xx_load_128(a);
+    const __m128i v_b0_w = xx_load_128(b);
+    const __m128i v_a1_w = xx_load_128(a + 8);
+    const __m128i v_b1_w = xx_load_128(b + 8);
+    const __m128i v_a2_w = xx_load_128(a + 16);
+    const __m128i v_b2_w = xx_load_128(b + 16);
+    const __m128i v_a3_w = xx_load_128(a + 24);
+    const __m128i v_b3_w = xx_load_128(b + 24);
+
+    const __m128i v_ab0l_w = _mm_unpacklo_epi16(v_a0_w, v_b0_w);
+    const __m128i v_ab0h_w = _mm_unpackhi_epi16(v_a0_w, v_b0_w);
+    const __m128i v_ab1l_w = _mm_unpacklo_epi16(v_a1_w, v_b1_w);
+    const __m128i v_ab1h_w = _mm_unpackhi_epi16(v_a1_w, v_b1_w);
+    const __m128i v_ab2l_w = _mm_unpacklo_epi16(v_a2_w, v_b2_w);
+    const __m128i v_ab2h_w = _mm_unpackhi_epi16(v_a2_w, v_b2_w);
+    const __m128i v_ab3l_w = _mm_unpacklo_epi16(v_a3_w, v_b3_w);
+    const __m128i v_ab3h_w = _mm_unpackhi_epi16(v_a3_w, v_b3_w);
+
+    // Negate top word of pairs
+    const __m128i v_abl0n_w = negm_epi16(v_ab0l_w, v_neg_w);
+    const __m128i v_abh0n_w = negm_epi16(v_ab0h_w, v_neg_w);
+    const __m128i v_abl1n_w = negm_epi16(v_ab1l_w, v_neg_w);
+    const __m128i v_abh1n_w = negm_epi16(v_ab1h_w, v_neg_w);
+    const __m128i v_abl2n_w = negm_epi16(v_ab2l_w, v_neg_w);
+    const __m128i v_abh2n_w = negm_epi16(v_ab2h_w, v_neg_w);
+    const __m128i v_abl3n_w = negm_epi16(v_ab3l_w, v_neg_w);
+    const __m128i v_abh3n_w = negm_epi16(v_ab3h_w, v_neg_w);
+
+    const __m128i v_r0l_w = _mm_madd_epi16(v_ab0l_w, v_abl0n_w);
+    const __m128i v_r0h_w = _mm_madd_epi16(v_ab0h_w, v_abh0n_w);
+    const __m128i v_r1l_w = _mm_madd_epi16(v_ab1l_w, v_abl1n_w);
+    const __m128i v_r1h_w = _mm_madd_epi16(v_ab1h_w, v_abh1n_w);
+    const __m128i v_r2l_w = _mm_madd_epi16(v_ab2l_w, v_abl2n_w);
+    const __m128i v_r2h_w = _mm_madd_epi16(v_ab2h_w, v_abh2n_w);
+    const __m128i v_r3l_w = _mm_madd_epi16(v_ab3l_w, v_abl3n_w);
+    const __m128i v_r3h_w = _mm_madd_epi16(v_ab3h_w, v_abh3n_w);
+
+    const __m128i v_r0_w = _mm_packs_epi32(v_r0l_w, v_r0h_w);
+    const __m128i v_r1_w = _mm_packs_epi32(v_r1l_w, v_r1h_w);
+    const __m128i v_r2_w = _mm_packs_epi32(v_r2l_w, v_r2h_w);
+    const __m128i v_r3_w = _mm_packs_epi32(v_r3l_w, v_r3h_w);
+
+    xx_store_128(d, v_r0_w);
+    xx_store_128(d + 8, v_r1_w);
+    xx_store_128(d + 16, v_r2_w);
+    xx_store_128(d + 24, v_r3_w);
+
+    a += 32;
+    b += 32;
+    d += 32;
+    N -= 32;
+  } while (N);
+}
diff --git a/third_party/aom/av1/exports_dec b/third_party/aom/av1/exports_dec
new file mode 100644
index 0000000000..05860e8c0a
--- /dev/null
+++ b/third_party/aom/av1/exports_dec
@@ -0,0 +1,2 @@
+data aom_codec_av1_dx_algo
+text aom_codec_av1_dx
diff --git a/third_party/aom/av1/exports_enc b/third_party/aom/av1/exports_enc
new file mode 100644
index 0000000000..dc4a9eae79
--- /dev/null
+++ b/third_party/aom/av1/exports_enc
@@ -0,0 +1,2 @@
+data aom_codec_av1_cx_algo
+text aom_codec_av1_cx
diff --git a/third_party/aom/build/cmake/aom_config.c.cmake b/third_party/aom/build/cmake/aom_config.c.cmake
new file mode 100644
index 0000000000..70bf950371
--- /dev/null
+++ b/third_party/aom/build/cmake/aom_config.c.cmake
@@ -0,0 +1,15 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "aom/aom_codec.h"
+static const char* const cfg = "${AOM_CMAKE_CONFIG}";
+static const char* const aom_git_hash = "${AOM_GIT_HASH}";
+const char *aom_codec_build_config(void) {return cfg;}
+const char *aom_codec_git_hash(void) {return aom_git_hash;}
diff --git a/third_party/aom/build/cmake/aom_config_defaults.cmake b/third_party/aom/build/cmake/aom_config_defaults.cmake
new file mode 100644
index 0000000000..5c2bc5801f
--- /dev/null
+++ b/third_party/aom/build/cmake/aom_config_defaults.cmake
@@ -0,0 +1,167 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+# Defaults for every libaom configuration variable. Here we add all libaom
+# config variables to the cmake variable cache, but omit the FORCE parameter to
+# allow users to specify values when executing cmake to generate build files.
+# Values here are used only if not set by the user.
+set(RESTRICT "" CACHE STRING "Sets RESTRICT value for current target.")
+set(INLINE "" CACHE STRING "Sets INLINE value for current target.")
+set(ARCH_ARM 0 CACHE BOOL "Enables ARM architecture.")
+set(ARCH_MIPS 0 CACHE BOOL "Enables MIPS architecture.")
+set(ARCH_X86 0 CACHE BOOL "Enables X86 architecture.")
+set(ARCH_X86_64 0 CACHE BOOL "Enables X86_64 architecture.")
+set(HAVE_EDSP 0 CACHE BOOL "Enables EDSP optimizations.")
+set(HAVE_MEDIA 0 CACHE BOOL "Enables MEDIA optimizations.")
+set(HAVE_NEON 0 CACHE BOOL "Enables NEON intrinsics optimizations.")
+set(HAVE_NEON_ASM 0 CACHE BOOL "Enables NEON assembly optimizations.")
+set(HAVE_MIPS32 0 CACHE BOOL "Enables MIPS32 optimizations.")
+set(HAVE_DSPR2 0 CACHE BOOL "Enables DSPR2 optimizations.")
+set(HAVE_MSA 0 CACHE BOOL "Enables MSA optimizations.")
+set(HAVE_MIPS64 0 CACHE BOOL "Enables MIPS64 optimizations. ")
+set(HAVE_MMX 0 CACHE BOOL "Enables MMX optimizations. ")
+set(HAVE_SSE 0 CACHE BOOL "Enables SSE optimizations.")
+set(HAVE_SSE2 0 CACHE BOOL "Enables SSE2 optimizations.")
+set(HAVE_SSE3 0 CACHE BOOL "Enables SSE3 optimizations.")
+set(HAVE_SSSE3 0 CACHE BOOL "Enables SSSE3 optimizations.")
+set(HAVE_SSE4_1 0 CACHE BOOL "Enables SSE 4.1 optimizations.")
+set(HAVE_AVX 0 CACHE BOOL "Enables AVX optimizations.")
+set(HAVE_AVX2 0 CACHE BOOL "Enables AVX2 optimizations.")
+set(HAVE_AOM_PORTS 0 CACHE BOOL "Internal flag, deprecated.")
+set(HAVE_FEXCEPT 0 CACHE BOOL "Internal flag, GNU fenv.h present for target.")
+set(HAVE_PTHREAD_H 0 CACHE BOOL "Internal flag, target pthread support.")
+set(HAVE_UNISTD_H 0 CACHE BOOL "Internal flag, unistd.h present for target.")
+set(HAVE_WXWIDGETS 0 CACHE BOOL "WxWidgets present.")
+set(CONFIG_DEPENDENCY_TRACKING 1 CACHE BOOL "Internal flag.")
+set(CONFIG_EXTERNAL_BUILD 0 CACHE BOOL "Internal flag.")
+set(CONFIG_INSTALL_DOCS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_INSTALL_BINS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_INSTALL_LIBS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_INSTALL_SRCS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_DEBUG 0 CACHE BOOL "Internal flag.")
+set(CONFIG_GPROF 0 CACHE BOOL "Internal flag.")
+set(CONFIG_GCOV 0 CACHE BOOL "Internal flag.")
+set(CONFIG_RVCT 0 CACHE BOOL "Internal flag.")
+set(CONFIG_GCC 0 CACHE BOOL "Internal flag.")
+set(CONFIG_MSVS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_PIC 0 CACHE BOOL "Internal flag.")
+set(CONFIG_BIG_ENDIAN 0 CACHE BOOL "Internal flag.")
+set(CONFIG_CODEC_SRCS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_DEBUG_LIBS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_RUNTIME_CPU_DETECT 1 CACHE BOOL "Internal flag.")
+set(CONFIG_POSTPROC 1 CACHE BOOL "Internal flag.")
+set(CONFIG_MULTITHREAD 1 CACHE BOOL "Internal flag.")
+set(CONFIG_INTERNAL_STATS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_AV1_ENCODER 1 CACHE BOOL "Enable AV1 encoder.")
+set(CONFIG_AV1_DECODER 1 CACHE BOOL "Enable AV1 decoder.")
+set(CONFIG_AV1 1 CACHE BOOL "Internal flag.")
+set(CONFIG_ENCODERS 1 CACHE BOOL "Enable encoding.")
+set(CONFIG_DECODERS 1 CACHE BOOL "Enable decoding.")
+set(CONFIG_STATIC_MSVCRT 0 CACHE BOOL "Internal flag.")
+set(CONFIG_SPATIAL_RESAMPLING 1 CACHE BOOL "Internal flag.")
+set(CONFIG_REALTIME_ONLY 0 CACHE BOOL "Internal flag.")
+set(CONFIG_ONTHEFLY_BITPACKING 0 CACHE BOOL "Internal flag.")
+set(CONFIG_ERROR_CONCEALMENT 0 CACHE BOOL "Internal flag.")
+set(CONFIG_SHARED 0 CACHE BOOL "Internal flag.")
+set(CONFIG_STATIC 1 CACHE BOOL "Internal flag.")
+set(CONFIG_SMALL 0 CACHE BOOL "Internal flag.")
+set(CONFIG_POSTPROC_VISUALIZER 0 CACHE BOOL "Internal flag.")
+set(CONFIG_OS_SUPPORT 0 CACHE BOOL "Internal flag.")
+set(CONFIG_UNIT_TESTS 1 CACHE BOOL "Internal flag.")
+set(CONFIG_WEBM_IO 1 CACHE BOOL "Enables WebM support.")
+set(CONFIG_LIBYUV 1 CACHE BOOL "Enables libyuv scaling and conversion support.")
+set(CONFIG_ACCOUNTING 0 CACHE BOOL "Enables bit accounting.")
+set(CONFIG_INSPECTION 0 CACHE BOOL "Enables bitstream inspection.")
+set(CONFIG_DECODE_PERF_TESTS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_ENCODE_PERF_TESTS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_COEFFICIENT_RANGE_CHECKING 0 CACHE BOOL "Internal flag.")
+set(CONFIG_LOWBITDEPTH 1 CACHE BOOL "Internal flag.")
+set(CONFIG_HIGHBITDEPTH 0 CACHE BOOL "Internal flag.")
+set(CONFIG_EXPERIMENTAL 0 CACHE BOOL "Internal flag.")
+set(CONFIG_SIZE_LIMIT 0 CACHE BOOL "Internal flag.")
+set(CONFIG_AOM_QM 0 CACHE BOOL "Internal flag.")
+set(CONFIG_FP_MB_STATS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_CDEF 1 CACHE BOOL "Internal flag.")
+set(CONFIG_VAR_TX 0 CACHE BOOL "Internal flag.")
+set(CONFIG_RECT_TX 1 CACHE BOOL "Internal flag.")
+set(CONFIG_REF_MV 1 CACHE BOOL "Internal flag.")
+set(CONFIG_TPL_MV 0 CACHE BOOL "Internal flag.")
+set(CONFIG_DUAL_FILTER 1 CACHE BOOL "Internal flag.")
+set(CONFIG_CONVOLVE_ROUND 0 CACHE BOOL "Internal flag.")
+set(CONFIG_COMPOUND_ROUND 0 CACHE BOOL "Internal flag.")
+set(CONFIG_EXT_TX 0 CACHE BOOL "Internal flag.")
+set(CONFIG_TX64X64 0 CACHE BOOL "Internal flag.")
+set(CONFIG_SUB8X8_MC 0 CACHE BOOL "Internal flag.")
+set(CONFIG_EXT_INTRA 1 CACHE BOOL "Internal flag.")
+set(CONFIG_INTRA_INTERP 0 CACHE BOOL "Internal flag.")
+set(CONFIG_FILTER_INTRA 0 CACHE BOOL "Internal flag.")
+set(CONFIG_INTRABC 0 CACHE BOOL "Internal flag.")
+set(CONFIG_EXT_INTER 0 CACHE BOOL "Internal flag.")
+set(CONFIG_INTERINTRA 0 CACHE BOOL "Internal flag.")
+set(CONFIG_WEDGE 0 CACHE BOOL "Internal flag.")
+set(CONFIG_COMPOUND_SEGMENT 0 CACHE BOOL "Internal flag.")
+set(CONFIG_EXT_REFS 1 CACHE BOOL "Internal flag.")
+set(CONFIG_GLOBAL_MOTION 0 CACHE BOOL "Internal flag.")
+set(CONFIG_NEW_QUANT 0 CACHE BOOL "Internal flag.")
+set(CONFIG_SUPERTX 0 CACHE BOOL "Internal flag.")
+set(CONFIG_ANS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_EC_MULTISYMBOL 1 CACHE BOOL "Internal flag.")
+set(CONFIG_NEW_TOKENSET 1 CACHE BOOL "Internal flag.")
+set(CONFIG_LOOP_RESTORATION 0 CACHE BOOL "Internal flag.")
+set(CONFIG_EXT_PARTITION 0 CACHE BOOL "Internal flag.")
+set(CONFIG_EXT_PARTITION_TYPES 0 CACHE BOOL "Internal flag.")
+set(CONFIG_UNPOISON_PARTITION_CTX 0 CACHE BOOL "Internal flag.")
+set(CONFIG_EXT_TILE 0 CACHE BOOL "Internal flag.")
+set(CONFIG_MOTION_VAR 0 CACHE BOOL "Internal flag.")
+set(CONFIG_NCOBMC 0 CACHE BOOL "Internal flag.")
+set(CONFIG_WARPED_MOTION 0 CACHE BOOL "Internal flag.")
+set(CONFIG_Q_ADAPT_PROBS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_SUBFRAME_PROB_UPDATE 0 CACHE BOOL "Internal flag.")
+set(CONFIG_BITSTREAM_DEBUG 0 CACHE BOOL "Internal flag.")
+set(CONFIG_ALT_INTRA 1 CACHE BOOL "Internal flag.")
+set(CONFIG_PALETTE 1 CACHE BOOL "Internal flag.")
+set(CONFIG_PALETTE_DELTA_ENCODING 0 CACHE BOOL "Internal flag.")
+set(CONFIG_DAALA_EC 1 CACHE BOOL "Internal flag.")
+set(CONFIG_RAWBITS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_EC_SMALLMUL 0 CACHE BOOL "Internal flag.")
+set(CONFIG_PVQ 0 CACHE BOOL "Internal flag.")
+set(CONFIG_CFL 0 CACHE BOOL "Internal flag.")
+set(CONFIG_XIPHRC 0 CACHE BOOL "Internal flag.")
+set(CONFIG_CB4X4 1 CACHE BOOL "Internal flag.")
+set(CONFIG_CHROMA_2X2 0 CACHE BOOL "Internal flag.")
+set(CONFIG_FRAME_SIZE 0 CACHE BOOL "Internal flag.")
+set(CONFIG_DELTA_Q 1 CACHE BOOL "Internal flag.")
+set(CONFIG_EXT_DELTA_Q 0 CACHE BOOL "Internal flag.")
+set(CONFIG_ADAPT_SCAN 0 CACHE BOOL "Internal flag.")
+set(CONFIG_FILTER_7BIT 1 CACHE BOOL "Internal flag.")
+set(CONFIG_PARALLEL_DEBLOCKING 0 CACHE BOOL "Internal flag.")
+set(CONFIG_PARALLEL_DEBLOCKING_15TAP 0 CACHE BOOL "Internal flag.")
+set(CONFIG_LOOPFILTERING_ACROSS_TILES 0 CACHE BOOL "Internal flag.")
+set(CONFIG_TILE_GROUPS 1 CACHE BOOL "Internal flag.")
+set(CONFIG_EC_ADAPT 1 CACHE BOOL "Internal flag.")
+set(CONFIG_TEMPMV_SIGNALING 0 CACHE BOOL "Internal flag.")
+set(CONFIG_RD_DEBUG 0 CACHE BOOL "Internal flag.")
+set(CONFIG_REFERENCE_BUFFER 1 CACHE BOOL "Internal flag.")
+set(CONFIG_COEF_INTERLEAVE 0 CACHE BOOL "Internal flag.")
+set(CONFIG_ENTROPY_STATS 0 CACHE BOOL "Internal flag.")
+set(CONFIG_MASKED_TX 0 CACHE BOOL "Internal flag.")
+set(CONFIG_DEPENDENT_HORZTILES 0 CACHE BOOL "Internal flag.")
+set(CONFIG_DAALA_DIST 0 CACHE BOOL "Internal flag.")
+set(CONFIG_TRIPRED 0 CACHE BOOL "Internal flag.")
+set(CONFIG_PALETTE_THROUGHPUT 1 CACHE BOOL "Internal flag.")
+set(CONFIG_REF_ADAPT 0 CACHE BOOL "Internal flag.")
+set(CONFIG_LV_MAP 0 CACHE BOOL "Internal flag.")
+set(CONFIG_TXK_SEL 0 CACHE BOOL "Internal flag.")
+set(CONFIG_MV_COMPRESS 1 CACHE BOOL "Internal flag.")
+set(CONFIG_FRAME_SUPERRES 0 CACHE BOOL "Internal flag.")
+set(CONFIG_NEW_MULTISYMBOL 0 CACHE BOOL "Internal flag.")
+set(CONFIG_COMPOUND_SINGLEREF 0 CACHE BOOL "Internal flag.")
+set(CONFIG_ANALYZER 0 CACHE BOOL "Internal flag.")
diff --git a/third_party/aom/build/cmake/aom_configure.cmake b/third_party/aom/build/cmake/aom_configure.cmake
new file mode 100644
index 0000000000..3c9402d0bd
--- /dev/null
+++ b/third_party/aom/build/cmake/aom_configure.cmake
@@ -0,0 +1,306 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+include(FindGit)
+include(FindPerl)
+include(FindThreads)
+include(FindwxWidgets)
+
+set(AOM_SUPPORTED_CPU_TARGETS
+    "arm64 armv7 armv7s generic mips32 mips64 x86 x86_64")
+
+# Generate the user config settings. This must occur before include of
+# aom_config_defaults.cmake (because it turns every config variable into a cache
+# variable with its own help string).
+get_cmake_property(cmake_cache_vars CACHE_VARIABLES)
+foreach(cache_var ${cmake_cache_vars})
+  get_property(cache_var_helpstring CACHE ${cache_var} PROPERTY HELPSTRING)
+  set(cmdline_helpstring  "No help, variable specified on the command line.")
+  if("${cache_var_helpstring}" STREQUAL "${cmdline_helpstring}")
+    set(AOM_CMAKE_CONFIG "${AOM_CMAKE_CONFIG} -D${cache_var}=${${cache_var}}")
+  endif ()
+endforeach()
+string(STRIP "${AOM_CMAKE_CONFIG}" AOM_CMAKE_CONFIG)
+
+include("${AOM_ROOT}/build/cmake/aom_config_defaults.cmake")
+include("${AOM_ROOT}/build/cmake/aom_optimization.cmake")
+include("${AOM_ROOT}/build/cmake/compiler_flags.cmake")
+include("${AOM_ROOT}/build/cmake/compiler_tests.cmake")
+
+# Build a list of all configurable variables.
+get_cmake_property(cmake_cache_vars CACHE_VARIABLES)
+foreach (var ${cmake_cache_vars})
+  if ("${var}" MATCHES "^CONFIG_")
+    list(APPEND AOM_CONFIG_VARS ${var})
+  endif ()
+endforeach ()
+
+# Detect target CPU.
+if (NOT AOM_TARGET_CPU)
+  if ("${CMAKE_SYSTEM_PROCESSOR}" STREQUAL "AMD64" OR
+      "${CMAKE_SYSTEM_PROCESSOR}" STREQUAL "x86_64")
+    if (${CMAKE_SIZEOF_VOID_P} EQUAL 4)
+      set(AOM_TARGET_CPU "x86")
+    elseif (${CMAKE_SIZEOF_VOID_P} EQUAL 8)
+      set(AOM_TARGET_CPU "x86_64")
+    else ()
+      message(FATAL_ERROR
+              "--- Unexpected pointer size (${CMAKE_SIZEOF_VOID_P}) for\n"
+              "      CMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME}\n"
+              "      CMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}\n"
+              "      CMAKE_GENERATOR=${CMAKE_GENERATOR}\n")
+    endif ()
+  elseif ("${CMAKE_SYSTEM_PROCESSOR}" STREQUAL "i386" OR
+          "${CMAKE_SYSTEM_PROCESSOR}" STREQUAL "x86")
+    set(AOM_TARGET_CPU "x86")
+  elseif ("${CMAKE_SYSTEM_PROCESSOR}" MATCHES "^arm" OR
+          "${CMAKE_SYSTEM_PROCESSOR}" MATCHES "^mips")
+    set(AOM_TARGET_CPU "${CMAKE_SYSTEM_PROCESSOR}")
+  endif ()
+endif ()
+
+if (CMAKE_TOOLCHAIN_FILE)
+  # Add toolchain file to config string.
+  set(toolchain_string "-DCMAKE_TOOLCHAIN_FILE=\\\"${CMAKE_TOOLCHAIN_FILE}\\\"")
+  set(AOM_CMAKE_CONFIG "${toolchain_string} ${AOM_CMAKE_CONFIG}")
+else ()
+  # Add detected CPU to the config string.
+  set(AOM_CMAKE_CONFIG "-DAOM_TARGET_CPU=${AOM_TARGET_CPU} ${AOM_CMAKE_CONFIG}")
+endif ()
+set(AOM_CMAKE_CONFIG "-G \\\"${CMAKE_GENERATOR}\\\" ${AOM_CMAKE_CONFIG}")
+string(STRIP "${AOM_CMAKE_CONFIG}" AOM_CMAKE_CONFIG)
+
+message("--- aom_configure: Detected CPU: ${AOM_TARGET_CPU}")
+set(AOM_TARGET_SYSTEM ${CMAKE_SYSTEM_NAME})
+
+if (NOT "${AOM_SUPPORTED_CPU_TARGETS}" MATCHES "${AOM_TARGET_CPU}")
+  message(FATAL_ERROR "No RTCD support for ${AOM_TARGET_CPU}. Create it, or "
+          "add -DAOM_TARGET_CPU=generic to your cmake command line for a "
+          "generic build of libaom and tools.")
+endif ()
+
+if ("${AOM_TARGET_CPU}" STREQUAL "x86" OR "${AOM_TARGET_CPU}" STREQUAL "x86_64")
+  # TODO(tomfinegan): Support nasm at least as well as the existing build
+  # system.
+  find_program(AS_EXECUTABLE yasm $ENV{YASM_PATH})
+  if (NOT AS_EXECUTABLE)
+    message(FATAL_ERROR "Unable to find yasm. To build without optimizations, "
+            "add -DAOM_TARGET_CPU=generic to your cmake command line.")
+  endif ()
+  get_asm_obj_format("objformat")
+  set(AOM_AS_FLAGS -f ${objformat} ${AOM_AS_FLAGS})
+  string(STRIP "${AOM_AS_FLAGS}" AOM_AS_FLAGS)
+elseif ("${AOM_TARGET_CPU}" MATCHES "arm")
+  if ("${AOM_TARGET_SYSTEM}" STREQUAL "Darwin")
+    set(AS_EXECUTABLE as)
+    set(AOM_AS_FLAGS -arch ${AOM_TARGET_CPU} -isysroot ${CMAKE_OSX_SYSROOT})
+  elseif ("${AOM_TARGET_SYSTEM}" STREQUAL "Linux")
+    # arm linux assembler settings controlled by
+    # build/cmake/toolchains/arm*-linux*.cmake
+  endif ()
+  if (NOT AS_EXECUTABLE)
+    message(FATAL_ERROR
+            "Unknown assembler for: ${AOM_TARGET_CPU}-${AOM_TARGET_SYSTEM}")
+  endif ()
+
+  string(STRIP "${AOM_AS_FLAGS}" AOM_AS_FLAGS)
+endif ()
+
+include("${AOM_ROOT}/build/cmake/cpu.cmake")
+
+# Test compiler flags.
+if (MSVC)
+  add_compiler_flag_if_supported("/W3")
+  # Disable MSVC warnings that suggest making code non-portable.
+  add_compiler_flag_if_supported("/wd4996")
+  if (ENABLE_WERROR)
+    add_compiler_flag_if_supported("/WX")
+  endif ()
+else ()
+  require_c_flag("-std=c99" YES)
+  add_compiler_flag_if_supported("-Wall")
+  add_compiler_flag_if_supported("-Wdisabled-optimization")
+  add_compiler_flag_if_supported("-Wextra")
+  add_compiler_flag_if_supported("-Wfloat-conversion")
+  add_compiler_flag_if_supported("-Wimplicit-function-declaration")
+  add_compiler_flag_if_supported("-Wpointer-arith")
+  add_compiler_flag_if_supported("-Wsign-compare")
+  add_compiler_flag_if_supported("-Wtype-limits")
+  add_compiler_flag_if_supported("-Wuninitialized")
+  add_compiler_flag_if_supported("-Wunused")
+  add_compiler_flag_if_supported("-Wvla")
+  # TODO(jzern): this could be added as a cxx flags for test/*.cc only,
+  # avoiding third_party.
+  add_c_flag_if_supported("-Wshorten-64-to-32")
+
+  # Add -Wshadow only for C files to avoid massive gtest warning spam.
+  add_c_flag_if_supported("-Wshadow")
+
+  # Add -Wundef only for C files to avoid massive gtest warning spam.
+  add_c_flag_if_supported("-Wundef")
+
+  if (ENABLE_WERROR)
+    add_compiler_flag_if_supported("-Werror")
+  endif ()
+  # Flag(s) added here negate CMake defaults and produce build output similar
+  # to the existing configure/make build system.
+  add_compiler_flag_if_supported("-Wno-unused-function")
+
+  if (CMAKE_C_COMPILER_ID MATCHES "GNU\|Clang")
+    set(CONFIG_GCC 1)
+  endif ()
+
+  if ("${CMAKE_BUILD_TYPE}" MATCHES "Rel")
+    add_compiler_flag_if_supported("-U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=0")
+  endif ()
+  add_compiler_flag_if_supported("-D_LARGEFILE_SOURCE")
+  add_compiler_flag_if_supported("-D_FILE_OFFSET_BITS=64")
+endif ()
+
+if ("${AOM_TARGET_SYSTEM}" MATCHES "Darwin\|Linux\|Windows")
+  set(CONFIG_OS_SUPPORT 1)
+endif ()
+
+# Test compiler support.
+aom_get_inline("INLINE")
+
+# TODO(tomfinegan): aom_ports_check is legacy; HAVE_AOM_PORTS is not used
+# anywhere in the aom sources. To be removed after parity with the legacy
+# build system stops being important.
+aom_check_source_compiles("aom_ports_check"
+                          "#include \"${AOM_ROOT}/aom/aom_integer.h\""
+                          HAVE_AOM_PORTS)
+aom_check_source_compiles("pthread_check" "#include <pthread.h>" HAVE_PTHREAD_H)
+aom_check_source_compiles("unistd_check" "#include <unistd.h>" HAVE_UNISTD_H)
+
+if (CONFIG_ANALYZER)
+  find_package(wxWidgets REQUIRED adv base core)
+  include(${wxWidgets_USE_FILE})
+
+  if (NOT CONFIG_INSPECTION)
+    set(CONFIG_INSPECTION 1)
+    message(WARNING
+            "--- Enabled CONFIG_INSPECTION, required for CONFIG_ANALYZER.")
+  endif ()
+endif ()
+
+if (CONFIG_ANS AND CONFIG_DAALA_EC)
+  message(FATAL_ERROR
+          "CONFIG_ANS and CONFIG_DAALA_EC cannot be enabled together.")
+endif ()
+
+if (NOT MSVC)
+  aom_push_var(CMAKE_REQUIRED_LIBRARIES "m")
+  aom_check_c_compiles("fenv_check"
+                       "#define _GNU_SOURCE
+                        #include <fenv.h>
+                        void unused(void) {
+                          (void)feenableexcept(FE_DIVBYZERO | FE_INVALID);
+                        }" HAVE_FEXCEPT)
+  aom_pop_var(CMAKE_REQUIRED_LIBRARIES)
+endif()
+
+# Generate aom_config templates.
+set(aom_config_asm_template "${AOM_CONFIG_DIR}/aom_config.asm.cmake")
+set(aom_config_h_template "${AOM_CONFIG_DIR}/aom_config.h.cmake")
+execute_process(COMMAND ${CMAKE_COMMAND}
+  -DAOM_CONFIG_DIR=${AOM_CONFIG_DIR}
+  -DAOM_ROOT=${AOM_ROOT}
+  -P "${AOM_ROOT}/build/cmake/generate_aom_config_templates.cmake")
+
+# Generate aom_config.{asm,h}.
+configure_file("${aom_config_asm_template}" "${AOM_CONFIG_DIR}/aom_config.asm")
+configure_file("${aom_config_h_template}" "${AOM_CONFIG_DIR}/aom_config.h")
+
+# Read the current git hash.
+find_package(Git)
+set(AOM_GIT_DESCRIPTION)
+set(AOM_GIT_HASH)
+if (GIT_FOUND)
+  # TODO(tomfinegan): Add build rule so users don't have to re-run cmake to
+  # create accurately versioned cmake builds.
+  execute_process(COMMAND ${GIT_EXECUTABLE}
+                  --git-dir=${AOM_ROOT}/.git rev-parse HEAD
+                  OUTPUT_VARIABLE AOM_GIT_HASH)
+  execute_process(COMMAND ${GIT_EXECUTABLE} --git-dir=${AOM_ROOT}/.git describe
+                  OUTPUT_VARIABLE AOM_GIT_DESCRIPTION ERROR_QUIET)
+  # Consume the newline at the end of the git output.
+  string(STRIP "${AOM_GIT_HASH}" AOM_GIT_HASH)
+  string(STRIP "${AOM_GIT_DESCRIPTION}" AOM_GIT_DESCRIPTION)
+endif ()
+
+configure_file("${AOM_ROOT}/build/cmake/aom_config.c.cmake"
+               "${AOM_CONFIG_DIR}/aom_config.c")
+
+# Find Perl and generate the RTCD sources.
+find_package(Perl)
+if (NOT PERL_FOUND)
+  message(FATAL_ERROR "Perl is required to build libaom.")
+endif ()
+configure_file(
+  "${AOM_ROOT}/build/cmake/rtcd_config.cmake"
+  "${AOM_CONFIG_DIR}/${AOM_TARGET_CPU}_rtcd_config.rtcd")
+
+set(AOM_RTCD_CONFIG_FILE_LIST
+    "${AOM_ROOT}/aom_dsp/aom_dsp_rtcd_defs.pl"
+    "${AOM_ROOT}/aom_scale/aom_scale_rtcd.pl"
+    "${AOM_ROOT}/av1/common/av1_rtcd_defs.pl")
+set(AOM_RTCD_HEADER_FILE_LIST
+    "${AOM_CONFIG_DIR}/aom_dsp_rtcd.h"
+    "${AOM_CONFIG_DIR}/aom_scale_rtcd.h"
+    "${AOM_CONFIG_DIR}/av1_rtcd.h")
+set(AOM_RTCD_SOURCE_FILE_LIST
+    "${AOM_ROOT}/aom_dsp/aom_dsp_rtcd.c"
+    "${AOM_ROOT}/aom_scale/aom_scale_rtcd.c"
+    "${AOM_ROOT}/av1/common/av1_rtcd.c")
+set(AOM_RTCD_SYMBOL_LIST aom_dsp_rtcd aom_scale_rtcd av1_rtcd)
+list(LENGTH AOM_RTCD_SYMBOL_LIST AOM_RTCD_CUSTOM_COMMAND_COUNT)
+math(EXPR AOM_RTCD_CUSTOM_COMMAND_COUNT "${AOM_RTCD_CUSTOM_COMMAND_COUNT} - 1")
+
+foreach(NUM RANGE ${AOM_RTCD_CUSTOM_COMMAND_COUNT})
+  list(GET AOM_RTCD_CONFIG_FILE_LIST ${NUM} AOM_RTCD_CONFIG_FILE)
+  list(GET AOM_RTCD_HEADER_FILE_LIST ${NUM} AOM_RTCD_HEADER_FILE)
+  list(GET AOM_RTCD_SOURCE_FILE_LIST ${NUM} AOM_RTCD_SOURCE_FILE)
+  list(GET AOM_RTCD_SYMBOL_LIST ${NUM} AOM_RTCD_SYMBOL)
+  execute_process(
+    COMMAND ${PERL_EXECUTABLE} "${AOM_ROOT}/build/make/rtcd.pl"
+      --arch=${AOM_TARGET_CPU} --sym=${AOM_RTCD_SYMBOL} ${AOM_RTCD_FLAGS}
+      --config=${AOM_CONFIG_DIR}/${AOM_TARGET_CPU}_rtcd_config.rtcd
+      ${AOM_RTCD_CONFIG_FILE}
+    OUTPUT_FILE ${AOM_RTCD_HEADER_FILE})
+endforeach()
+
+function (add_rtcd_build_step config output source symbol)
+  add_custom_command(
+    OUTPUT ${output}
+    COMMAND ${PERL_EXECUTABLE}
+    ARGS "${AOM_ROOT}/build/make/rtcd.pl"
+      --arch=${AOM_TARGET_CPU}
+      --sym=${symbol}
+      ${AOM_RTCD_FLAGS}
+      --config=${AOM_CONFIG_DIR}/${AOM_TARGET_CPU}_rtcd_config.rtcd
+      ${config}
+      > ${output}
+    DEPENDS ${config}
+    COMMENT "Generating ${output}"
+    WORKING_DIRECTORY ${AOM_CONFIG_DIR}
+    VERBATIM)
+  set_property(SOURCE ${source} PROPERTY OBJECT_DEPENDS ${output})
+  set_property(SOURCE ${output} PROPERTY GENERATED)
+endfunction ()
+
+# Generate aom_version.h.
+if ("${AOM_GIT_DESCRIPTION}" STREQUAL "")
+  set(AOM_GIT_DESCRIPTION "${AOM_ROOT}/CHANGELOG")
+endif ()
+execute_process(
+  COMMAND ${PERL_EXECUTABLE} "${AOM_ROOT}/build/cmake/aom_version.pl"
+  --version_data=${AOM_GIT_DESCRIPTION}
+  --version_filename=${AOM_CONFIG_DIR}/aom_version.h)
diff --git a/third_party/aom/build/cmake/aom_optimization.cmake b/third_party/aom/build/cmake/aom_optimization.cmake
new file mode 100644
index 0000000000..e2b0ba07e3
--- /dev/null
+++ b/third_party/aom/build/cmake/aom_optimization.cmake
@@ -0,0 +1,204 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_AOM_OPTIMIZATION_CMAKE_)
+set(AOM_BUILD_CMAKE_AOM_OPTIMIZATION_CMAKE_ 1)
+
+# Translate $flag to one which MSVC understands, and write the new flag to the
+# variable named by $translated_flag (or unset it, when MSVC needs no flag).
+function (get_msvc_intrinsic_flag flag translated_flag)
+  if ("${flag}" STREQUAL "-mavx")
+    set(${translated_flag} "/arch:AVX" PARENT_SCOPE)
+  elseif ("${flag}" STREQUAL "-mavx2")
+    set(${translated_flag} "/arch:AVX2" PARENT_SCOPE)
+  else ()
+    # MSVC does not need flags for intrinsics flavors other than AVX/AVX2.
+    unset(${translated_flag} PARENT_SCOPE)
+  endif ()
+endfunction ()
+
+# Adds an object library target. Terminates generation if $flag is not supported
+# by the current compiler. $flag is the intrinsics flag required by the current
+# compiler, and is added to the compile flags for all sources in $sources.
+# $opt_name is used to name the target. $target_to_update is made
+# dependent upon the created target.
+#
+# Note: the libaom target is always updated because OBJECT libraries have rules
+# that disallow the direct addition of .o files to them as dependencies. Static
+# libraries do not have this limitation.
+function (add_intrinsics_object_library flag opt_name target_to_update sources)
+  set(target_name ${target_to_update}_${opt_name}_intrinsics)
+  add_library(${target_name} OBJECT ${${sources}})
+
+  if (MSVC)
+    get_msvc_intrinsic_flag(${flag} "flag")
+  endif ()
+
+  if (flag)
+    target_compile_options(${target_name} PUBLIC ${flag})
+  endif ()
+
+  target_sources(aom PUBLIC $<TARGET_OBJECTS:${target_name}>)
+
+  # Add the new lib target to the global list of aom library targets.
+  list(APPEND AOM_LIB_TARGETS ${target_name})
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} PARENT_SCOPE)
+endfunction ()
+
+# Adds sources in list named by $sources to $target and adds $flag to the
+# compile flags for each source file.
+function (add_intrinsics_source_to_target flag target sources)
+  target_sources(${target} PUBLIC ${${sources}})
+  if (MSVC)
+    get_msvc_intrinsic_flag(${flag} "flag")
+  endif ()
+  if (flag)
+    foreach (source ${${sources}})
+      set_property(SOURCE ${source} APPEND PROPERTY COMPILE_FLAGS ${flag})
+    endforeach ()
+  endif ()
+endfunction ()
+
+# Writes object format for the current target to the var named by $out_format,
+# or terminates the build when the object format for the current target is
+# unknown.
+function (get_asm_obj_format out_format)
+  if ("${AOM_TARGET_CPU}" STREQUAL "x86_64")
+    if ("${AOM_TARGET_SYSTEM}" STREQUAL "Darwin")
+      set(objformat "macho64")
+    elseif ("${AOM_TARGET_SYSTEM}" STREQUAL "Linux")
+      set(objformat "elf64")
+    elseif ("${AOM_TARGET_SYSTEM}" STREQUAL "Windows")
+      set(objformat "win64")
+    else ()
+      message(FATAL_ERROR "Unknown obj format: ${AOM_TARGET_SYSTEM}")
+    endif ()
+  elseif ("${AOM_TARGET_CPU}" STREQUAL "x86")
+    if ("${AOM_TARGET_SYSTEM}" STREQUAL "Darwin")
+      set(objformat "macho32")
+    elseif ("${AOM_TARGET_SYSTEM}" STREQUAL "Linux")
+      set(objformat "elf32")
+    elseif ("${AOM_TARGET_SYSTEM}" STREQUAL "Windows")
+      set(objformat "win32")
+    else ()
+      message(FATAL_ERROR "Unknown obj format: ${AOM_TARGET_SYSTEM}")
+    endif ()
+  else ()
+    message(FATAL_ERROR
+            "Unknown obj format: ${AOM_TARGET_CPU}-${AOM_TARGET_SYSTEM}")
+  endif ()
+
+  set(${out_format} ${objformat} PARENT_SCOPE)
+endfunction ()
+
+# Adds library target named $lib_name for ASM files in variable named by
+# $asm_sources. Builds an output directory path from $lib_name. Links $lib_name
+# into $dependent_target. Generates a dummy C file with a dummy function to
+# ensure that all cmake generators can determine the linker language, and that
+# build tools don't complain that an object exposes no symbols.
+function (add_asm_library lib_name asm_sources dependent_target)
+  set(asm_lib_obj_dir "${AOM_CONFIG_DIR}/asm_objects/${lib_name}")
+  if (NOT EXISTS "${asm_lib_obj_dir}")
+    file(MAKE_DIRECTORY "${asm_lib_obj_dir}")
+  endif ()
+
+  # TODO(tomfinegan): If cmake ever allows addition of .o files to OBJECT lib
+  # targets, make this OBJECT instead of STATIC to hide the target from
+  # consumers of the AOM cmake build.
+  add_library(${lib_name} STATIC ${${asm_sources}})
+
+  foreach (asm_source ${${asm_sources}})
+    get_filename_component(asm_source_name "${asm_source}" NAME)
+    set(asm_object "${asm_lib_obj_dir}/${asm_source_name}.o")
+    add_custom_command(OUTPUT "${asm_object}"
+                       COMMAND ${AS_EXECUTABLE}
+                       ARGS ${AOM_AS_FLAGS}
+                            -I${AOM_ROOT} -I${AOM_CONFIG_DIR}
+                            -o "${asm_object}" "${asm_source}"
+                       DEPENDS "${asm_source}"
+                       COMMENT "Building ASM object ${asm_object}"
+                       WORKING_DIRECTORY "${AOM_CONFIG_DIR}"
+                       VERBATIM)
+    target_sources(${lib_name} PRIVATE "${asm_object}")
+  endforeach ()
+
+  # The above created a target containing only ASM sources. Cmake needs help
+  # here to determine the linker language. Add a dummy C file to force the
+  # linker language to C. We don't bother with setting the LINKER_LANGUAGE
+  # property on the library target because not all generators obey it (looking
+  # at you, xcode generator).
+  set(dummy_c_file "${AOM_CONFIG_DIR}/${lib_name}_dummy.c")
+  file(WRITE "${dummy_c_file}"
+       "// Generated file. DO NOT EDIT!\n"
+       "// ${lib_name} needs C file to force link language, ignore me.\n"
+       "void ${lib_name}_dummy_function(void) {}\n")
+  target_sources(${lib_name} PUBLIC ${dummy_c_file})
+
+  target_link_libraries(${dependent_target} PRIVATE ${lib_name})
+
+  # Add the new lib target to the global list of aom library targets.
+  list(APPEND AOM_LIB_TARGETS ${lib_name})
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} PARENT_SCOPE)
+endfunction ()
+
+# Converts asm sources in $asm_sources using $AOM_ADS2GAS and calls
+# add_asm_library() to create a library from the converted sources. At
+# generation time the converted sources are created, and a custom rule is added
+# to ensure the sources are reconverted when the original asm source is updated.
+# See add_asm_library() for more information.
+function (add_gas_asm_library lib_name asm_sources dependent_target)
+  set(asm_converted_source_dir "${AOM_CONFIG_DIR}/asm_gas/${lib_name}")
+  if (NOT EXISTS "${asm_converted_source_dir}")
+    file(MAKE_DIRECTORY "${asm_converted_source_dir}")
+  endif ()
+
+  # Create the converted version of each assembly source at generation time.
+  unset(gas_target_sources)
+  foreach (neon_asm_source ${${asm_sources}})
+    get_filename_component(output_asm_source "${neon_asm_source}" NAME)
+    set(output_asm_source "${asm_converted_source_dir}/${output_asm_source}")
+    set(output_asm_source "${output_asm_source}.${AOM_GAS_EXT}")
+    execute_process(COMMAND "${PERL_EXECUTABLE}" "${AOM_ADS2GAS}"
+                    INPUT_FILE "${neon_asm_source}"
+                    OUTPUT_FILE "${output_asm_source}")
+    list(APPEND gas_target_sources "${output_asm_source}")
+  endforeach ()
+
+  add_asm_library("${lib_name}" "gas_target_sources" "${dependent_target}")
+
+  # For each of the converted sources, create a custom rule that will regenerate
+  # the converted source when its input is touched.
+  list(LENGTH gas_target_sources num_asm_files)
+  math(EXPR num_asm_files "${num_asm_files} - 1")
+  foreach(NUM RANGE ${num_asm_files})
+    list(GET ${asm_sources} ${NUM} neon_asm_source)
+    list(GET gas_target_sources ${NUM} gas_asm_source)
+
+    # Grab only the filename for the custom command output to keep build output
+    # reasonably sane.
+    get_filename_component(neon_name "${neon_asm_source}" NAME)
+    get_filename_component(gas_name "${gas_asm_source}" NAME)
+
+    add_custom_command(
+        OUTPUT "${gas_asm_source}"
+        COMMAND ${PERL_EXECUTABLE}
+        ARGS "${AOM_ADS2GAS}" < "${neon_asm_source}" > "${gas_asm_source}"
+        DEPENDS "${neon_asm_source}"
+        COMMENT "ads2gas conversion ${neon_name} -> ${gas_name}"
+        WORKING_DIRECTORY "${AOM_CONFIG_DIR}"
+        VERBATIM)
+  endforeach ()
+
+  # Update the sources list passed in to include the converted asm source files.
+  list(APPEND asm_sources ${gas_target_sources})
+  set(${asm_sources} ${${asm_sources}} PARENT_SCOPE)
+endfunction ()
+
+endif ()  # AOM_BUILD_CMAKE_AOM_OPTIMIZATION_CMAKE_
diff --git a/third_party/aom/build/cmake/aom_version.pl b/third_party/aom/build/cmake/aom_version.pl
new file mode 100755
index 0000000000..3412feebdf
--- /dev/null
+++ b/third_party/aom/build/cmake/aom_version.pl
@@ -0,0 +1,93 @@
+#!/usr/bin/env perl
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+use strict;
+use warnings;
+use 5.010;
+use Getopt::Long;
+
+my $version_data;
+my $version_filename;
+GetOptions('version_data=s' => \$version_data,
+           'version_filename=s' => \$version_filename) or
+  die("Invalid arg(s): $!");
+
+if (!defined $version_data || length($version_data) == 0 ||
+    !defined $version_filename || length($version_filename) == 0) {
+  die("--version_data and --version_filename are required.");
+}
+
+# Determine if $version_data is a filename or a git tag/description.
+my $version_string;
+if (-r $version_data) {
+  # $version_data is the path to the CHANGELOG. Parse the most recent version.
+  my $changelog_filename = $version_data;
+  open(my $changelog_file, '<', $changelog_filename) or
+    die("Unable to open CHANGELOG @ $changelog_filename: $!.");
+
+  while (my $line = <$changelog_file>) {
+    my @split_line = split(" ", $line, 3);
+    next if @split_line < 2;
+    $version_string = $split_line[1];
+    last if substr($version_string, 0, 1) eq "v";
+  }
+  close($changelog_file);
+} else {
+  # $version_data is either a tag name or a full git description, one of:
+  # tagName OR tagName-commitsSinceTag-shortCommitHash
+  # In either case we want the first element of the array returned by split.
+  $version_string = (split("-", $version_data))[0];
+}
+
+if (substr($version_string, 0, 1) eq "v") {
+  $version_string = substr($version_string, 1);
+}
+
+my @version_components = split('\.', $version_string, 4);
+my $version_major = $version_components[0];
+my $version_minor = $version_components[1];
+my $version_patch = $version_components[2];
+
+my $version_extra = "";
+if (@version_components > 3) {
+  $version_extra = $version_components[3];
+}
+
+open(my $version_file, '>', $version_filename) or
+  die("Cannot open $version_filename: $!");
+
+my $version_packed = "((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))";
+my $year = (localtime)[5] + 1900;
+my $lic_block = << "EOF";
+/*
+ * Copyright (c) $year, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+EOF
+
+select $version_file;
+print << "EOF";
+$lic_block
+#define VERSION_MAJOR $version_major
+#define VERSION_MINOR $version_minor
+#define VERSION_PATCH $version_patch
+#define VERSION_EXTRA \"$version_extra\"
+#define VERSION_PACKED $version_packed
+#define VERSION_STRING_NOSP \"v$version_string\"
+#define VERSION_STRING \" v$version_string\"
+EOF
+close($version_file);
diff --git a/third_party/aom/build/cmake/compiler_flags.cmake b/third_party/aom/build/cmake/compiler_flags.cmake
new file mode 100644
index 0000000000..beb217abcb
--- /dev/null
+++ b/third_party/aom/build/cmake/compiler_flags.cmake
@@ -0,0 +1,218 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+include(CheckCCompilerFlag)
+include(CheckCXXCompilerFlag)
+
+# Strings used to cache failed C/CXX flags.
+set(AOM_FAILED_C_FLAGS)
+set(AOM_FAILED_CXX_FLAGS)
+
+# Checks C compiler for support of $c_flag. Adds $c_flag to $CMAKE_C_FLAGS when
+# the compile test passes. Caches $c_flag in $AOM_FAILED_C_FLAGS when the test
+# fails.
+function (add_c_flag_if_supported c_flag)
+  unset(C_FLAG_FOUND CACHE)
+  string(FIND "${CMAKE_C_FLAGS}" "${c_flag}" C_FLAG_FOUND)
+  unset(C_FLAG_FAILED CACHE)
+  string(FIND "${AOM_FAILED_C_FLAGS}" "${c_flag}" C_FLAG_FAILED)
+
+  if (${C_FLAG_FOUND} EQUAL -1 AND ${C_FLAG_FAILED} EQUAL -1)
+    unset(C_FLAG_SUPPORTED CACHE)
+    message("Checking C compiler flag support for: " ${c_flag})
+    check_c_compiler_flag("${c_flag}" C_FLAG_SUPPORTED)
+    if (C_FLAG_SUPPORTED)
+      set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${c_flag}" CACHE STRING "" FORCE)
+    else ()
+      set(AOM_FAILED_C_FLAGS "${AOM_FAILED_C_FLAGS} ${c_flag}" CACHE STRING ""
+          FORCE)
+    endif ()
+  endif ()
+endfunction ()
+
+# Checks C++ compiler for support of $cxx_flag. Adds $cxx_flag to
+# $CMAKE_CXX_FLAGS when the compile test passes. Caches $c_flag in
+# $AOM_FAILED_CXX_FLAGS when the test fails.
+function (add_cxx_flag_if_supported cxx_flag)
+  unset(CXX_FLAG_FOUND CACHE)
+  string(FIND "${CMAKE_CXX_FLAGS}" "${cxx_flag}" CXX_FLAG_FOUND)
+  unset(CXX_FLAG_FAILED CACHE)
+  string(FIND "${AOM_FAILED_CXX_FLAGS}" "${cxx_flag}" CXX_FLAG_FAILED)
+
+  if (${CXX_FLAG_FOUND} EQUAL -1 AND ${CXX_FLAG_FAILED} EQUAL -1)
+    unset(CXX_FLAG_SUPPORTED CACHE)
+    message("Checking CXX compiler flag support for: " ${cxx_flag})
+    check_cxx_compiler_flag("${cxx_flag}" CXX_FLAG_SUPPORTED)
+    if (CXX_FLAG_SUPPORTED)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${cxx_flag}" CACHE STRING ""
+          FORCE)
+    else()
+      set(AOM_FAILED_CXX_FLAGS "${AOM_FAILED_CXX_FLAGS} ${cxx_flag}" CACHE
+          STRING "" FORCE)
+    endif ()
+  endif ()
+endfunction ()
+
+# Convenience method for adding a flag to both the C and C++ compiler command
+# lines.
+function (add_compiler_flag_if_supported flag)
+  add_c_flag_if_supported(${flag})
+  add_cxx_flag_if_supported(${flag})
+endfunction ()
+
+# Checks C compiler for support of $c_flag and terminates generation when
+# support is not present.
+function (require_c_flag c_flag update_c_flags)
+  unset(C_FLAG_FOUND CACHE)
+  string(FIND "${CMAKE_C_FLAGS}" "${c_flag}" C_FLAG_FOUND)
+
+  if (${C_FLAG_FOUND} EQUAL -1)
+    unset(HAVE_C_FLAG CACHE)
+    message("Checking C compiler flag support for: " ${c_flag})
+    check_c_compiler_flag("${c_flag}" HAVE_C_FLAG)
+    if (NOT HAVE_C_FLAG)
+      message(FATAL_ERROR
+              "${PROJECT_NAME} requires support for C flag: ${c_flag}.")
+    endif ()
+    if (update_c_flags)
+      set(CMAKE_C_FLAGS "${c_flag} ${CMAKE_C_FLAGS}" CACHE STRING "" FORCE)
+    endif ()
+  endif ()
+endfunction ()
+
+# Checks CXX compiler for support of $cxx_flag and terminates generation when
+# support is not present.
+function (require_cxx_flag cxx_flag update_cxx_flags)
+  unset(CXX_FLAG_FOUND CACHE)
+  string(FIND "${CMAKE_CXX_FLAGS}" "${cxx_flag}" CXX_FLAG_FOUND)
+
+  if (${CXX_FLAG_FOUND} EQUAL -1)
+    unset(HAVE_CXX_FLAG CACHE)
+    message("Checking CXX compiler flag support for: " ${cxx_flag})
+    check_cxx_compiler_flag("${cxx_flag}" HAVE_CXX_FLAG)
+    if (NOT HAVE_CXX_FLAG)
+      message(FATAL_ERROR
+              "${PROJECT_NAME} requires support for CXX flag: ${cxx_flag}.")
+    endif ()
+    if (update_cxx_flags)
+      set(CMAKE_CXX_FLAGS "${cxx_flag} ${CMAKE_CXX_FLAGS}" CACHE STRING ""
+          FORCE)
+    endif ()
+  endif ()
+endfunction ()
+
+# Checks for support of $flag by both the C and CXX compilers. Terminates
+# generation when support is not present in both compilers.
+function (require_flag flag update_cmake_flags)
+  require_c_flag(${flag} ${update_cmake_flags})
+  require_cxx_flag(${flag} ${update_cmake_flags})
+endfunction ()
+
+# Checks only non-MSVC targets for support of $c_flag and terminates generation
+# when support is not present.
+function (require_c_flag_nomsvc c_flag update_c_flags)
+  if (NOT MSVC)
+    require_c_flag(${c_flag} ${update_c_flags})
+  endif ()
+endfunction ()
+
+# Checks only non-MSVC targets for support of $cxx_flag and terminates
+# generation when support is not present.
+function (require_cxx_flag_nomsvc cxx_flag update_cxx_flags)
+  if (NOT MSVC)
+    require_cxx_flag(${cxx_flag} ${update_cxx_flags})
+  endif ()
+endfunction ()
+
+# Checks only non-MSVC targets for support of $flag by both the C and CXX
+# compilers. Terminates generation when support is not present in both
+# compilers.
+function (require_flag_nomsvc flag update_cmake_flags)
+  require_c_flag_nomsvc(${flag} ${update_cmake_flags})
+  require_cxx_flag_nomsvc(${flag} ${update_cmake_flags})
+endfunction ()
+
+# Adds $preproc_def to C compiler command line (as -D$preproc_def) if not
+# already present.
+function (add_c_preproc_definition preproc_def)
+  unset(PREPROC_DEF_FOUND CACHE)
+  string(FIND "${CMAKE_C_FLAGS}" "${preproc_def}" PREPROC_DEF_FOUND)
+
+  if (${PREPROC_DEF_FOUND} EQUAL -1)
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -D${preproc_def}" CACHE STRING ""
+        FORCE)
+  endif ()
+endfunction ()
+
+# Adds $preproc_def to CXX compiler command line (as -D$preproc_def) if not
+# already present.
+function (add_cxx_preproc_definition preproc_def)
+  unset(PREPROC_DEF_FOUND CACHE)
+  string(FIND "${CMAKE_CXX_FLAGS}" "${preproc_def}" PREPROC_DEF_FOUND)
+
+  if (${PREPROC_DEF_FOUND} EQUAL -1)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D${preproc_def}" CACHE STRING ""
+        FORCE)
+  endif ()
+endfunction ()
+
+# Adds $preproc_def to C and CXX compiler command line (as -D$preproc_def) if
+# not already present.
+function (add_preproc_definition preproc_def)
+  add_c_preproc_definition(${preproc_def})
+  add_cxx_preproc_definition(${preproc_def})
+endfunction ()
+
+# Adds $flag to assembler command line.
+function (append_as_flag flag)
+  unset(AS_FLAG_FOUND CACHE)
+  string(FIND "${AOM_AS_FLAGS}" "${flag}" AS_FLAG_FOUND)
+
+  if (${AS_FLAG_FOUND} EQUAL -1)
+    set(AOM_AS_FLAGS "${AOM_AS_FLAGS} ${flag}" CACHE STRING "" FORCE)
+  endif ()
+endfunction ()
+
+# Adds $flag to the C compiler command line.
+function (append_c_flag flag)
+  unset(C_FLAG_FOUND CACHE)
+  string(FIND "${CMAKE_C_FLAGS}" "${flag}" C_FLAG_FOUND)
+
+  if (${C_FLAG_FOUND} EQUAL -1)
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${flag}" CACHE STRING "" FORCE)
+  endif ()
+endfunction ()
+
+# Adds $flag to the CXX compiler command line.
+function (append_cxx_flag flag)
+  unset(CXX_FLAG_FOUND CACHE)
+  string(FIND "${CMAKE_CXX_FLAGS}" "${flag}" CXX_FLAG_FOUND)
+
+  if (${CXX_FLAG_FOUND} EQUAL -1)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${flag}" CACHE STRING "" FORCE)
+  endif ()
+endfunction ()
+
+# Adds $flag to the C and CXX compiler command lines.
+function (append_compiler_flag flag)
+  append_c_flag(${flag})
+  append_cxx_flag(${flag})
+endfunction ()
+
+# Adds $flag to the executable linker command line.
+function (append_exe_linker_flag flag)
+  unset(LINKER_FLAG_FOUND CACHE)
+  string(FIND "${CMAKE_EXE_LINKER_FLAGS}" "${flag}" LINKER_FLAG_FOUND)
+
+  if (${LINKER_FLAG_FOUND} EQUAL -1)
+    set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${flag}" CACHE STRING
+        "" FORCE)
+  endif ()
+endfunction ()
diff --git a/third_party/aom/build/cmake/compiler_tests.cmake b/third_party/aom/build/cmake/compiler_tests.cmake
new file mode 100644
index 0000000000..e763597a20
--- /dev/null
+++ b/third_party/aom/build/cmake/compiler_tests.cmake
@@ -0,0 +1,133 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+include(CheckCSourceCompiles)
+include(CheckCXXSourceCompiles)
+
+# The basic main() function used in all compile tests.
+set(AOM_C_MAIN "\nint main(void) { return 0; }")
+set(AOM_CXX_MAIN "\nint main() { return 0; }")
+
+# Strings containing the names of passed and failed tests.
+set(AOM_C_PASSED_TESTS)
+set(AOM_C_FAILED_TESTS)
+set(AOM_CXX_PASSED_TESTS)
+set(AOM_CXX_FAILED_TESTS)
+
+function(aom_push_var var new_value)
+  set(SAVED_${var} ${var} PARENT_SCOPE)
+  set(${var} ${new_value} PARENT_SCOPE)
+endfunction ()
+
+function(aom_pop_var var)
+  set(var ${SAVED_${var}} PARENT_SCOPE)
+  unset(SAVED_${var} PARENT_SCOPE)
+endfunction ()
+
+# Confirms $test_source compiles and stores $test_name in one of
+# $AOM_C_PASSED_TESTS or $AOM_C_FAILED_TESTS depending on out come. When the
+# test passes $result_var is set to 1. When it fails $result_var is unset.
+# The test is not run if the test name is found in either of the passed or
+# failed test variables.
+function(aom_check_c_compiles test_name test_source result_var)
+  unset(C_TEST_PASSED CACHE)
+  unset(C_TEST_FAILED CACHE)
+  string(FIND "${AOM_C_PASSED_TESTS}" "${test_name}" C_TEST_PASSED)
+  string(FIND "${AOM_C_FAILED_TESTS}" "${test_name}" C_TEST_FAILED)
+  if (${C_TEST_PASSED} EQUAL -1 AND ${C_TEST_FAILED} EQUAL -1)
+    unset(C_TEST_COMPILED CACHE)
+    message("Running C compiler test: ${test_name}")
+    check_c_source_compiles("${test_source} ${AOM_C_MAIN}" C_TEST_COMPILED)
+    set(${result_var} ${C_TEST_COMPILED} PARENT_SCOPE)
+
+    if (C_TEST_COMPILED)
+      set(AOM_C_PASSED_TESTS "${AOM_C_PASSED_TESTS} ${test_name}" CACHE STRING
+          "" FORCE)
+    else ()
+      set(AOM_C_FAILED_TESTS "${AOM_C_FAILED_TESTS} ${test_name}" CACHE STRING
+          "" FORCE)
+      message("C Compiler test ${test_name} failed.")
+    endif ()
+  elseif (NOT ${C_TEST_PASSED} EQUAL -1)
+    set(${result_var} 1 PARENT_SCOPE)
+  else ()  # ${C_TEST_FAILED} NOT EQUAL -1
+    unset(${result_var} PARENT_SCOPE)
+  endif ()
+endfunction ()
+
+# Confirms $test_source compiles and stores $test_name in one of
+# $AOM_CXX_PASSED_TESTS or $AOM_CXX_FAILED_TESTS depending on out come. When the
+# test passes $result_var is set to 1. When it fails $result_var is unset.
+# The test is not run if the test name is found in either of the passed or
+# failed test variables.
+function(aom_check_cxx_compiles test_name test_source result_var)
+  unset(CXX_TEST_PASSED CACHE)
+  unset(CXX_TEST_FAILED CACHE)
+  string(FIND "${AOM_CXX_PASSED_TESTS}" "${test_name}" CXX_TEST_PASSED)
+  string(FIND "${AOM_CXX_FAILED_TESTS}" "${test_name}" CXX_TEST_FAILED)
+  if (${CXX_TEST_PASSED} EQUAL -1 AND ${CXX_TEST_FAILED} EQUAL -1)
+    unset(CXX_TEST_COMPILED CACHE)
+    message("Running CXX compiler test: ${test_name}")
+    check_cxx_source_compiles("${test_source} ${AOM_CXX_MAIN}"
+                              CXX_TEST_COMPILED)
+    set(${result_var} ${CXX_TEST_COMPILED} PARENT_SCOPE)
+
+    if (CXX_TEST_COMPILED)
+      set(AOM_CXX_PASSED_TESTS "${AOM_CXX_PASSED_TESTS} ${test_name}" CACHE
+          STRING "" FORCE)
+    else ()
+      set(AOM_CXX_FAILED_TESTS "${AOM_CXX_FAILED_TESTS} ${test_name}" CACHE
+          STRING "" FORCE)
+      message("CXX Compiler test ${test_name} failed.")
+    endif ()
+  elseif (NOT ${CXX_TEST_PASSED} EQUAL -1)
+    set(${result_var} 1 PARENT_SCOPE)
+  else ()  # ${CXX_TEST_FAILED} NOT EQUAL -1
+    unset(${result_var} PARENT_SCOPE)
+  endif ()
+endfunction ()
+
+# Convenience function that confirms $test_source compiles as C and C++.
+# $result_var is set to 1 when both tests are successful, and 0 when one or both
+# tests fail.
+# Note: This function is intended to be used to write to result variables that are
+# expanded via configure_file(). $result_var is set to 1 or 0 to allow direct
+# usage of the value in generated source files.
+function(aom_check_source_compiles test_name test_source result_var)
+  unset(C_PASSED)
+  unset(CXX_PASSED)
+  aom_check_c_compiles(${test_name} ${test_source} C_PASSED)
+  aom_check_cxx_compiles(${test_name} ${test_source} CXX_PASSED)
+  if (C_PASSED AND CXX_PASSED)
+    set(${result_var} 1 PARENT_SCOPE)
+  else ()
+    set(${result_var} 0 PARENT_SCOPE)
+  endif ()
+endfunction ()
+
+# When inline support is detected for the current compiler the supported
+# inlining keyword is written to $result in caller scope.
+function (aom_get_inline result)
+  aom_check_source_compiles("inline_check_1"
+                            "static inline void function(void) {}"
+                            HAVE_INLINE_1)
+  if (HAVE_INLINE_1 EQUAL 1)
+    set(${result} "inline" PARENT_SCOPE)
+    return()
+  endif ()
+
+  # Check __inline.
+  aom_check_source_compiles("inline_check_2"
+                            "static __inline void function(void) {}"
+                            HAVE_INLINE_2)
+  if (HAVE_INLINE_2 EQUAL 1)
+    set(${result} "__inline" PARENT_SCOPE)
+  endif ()
+endfunction ()
diff --git a/third_party/aom/build/cmake/cpu.cmake b/third_party/aom/build/cmake/cpu.cmake
new file mode 100644
index 0000000000..5d0b1a6e8a
--- /dev/null
+++ b/third_party/aom/build/cmake/cpu.cmake
@@ -0,0 +1,72 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if ("${AOM_TARGET_CPU}" STREQUAL "arm64")
+  set(ARCH_ARM 1)
+  set(HAVE_NEON 1)
+  set(RTCD_ARCH_ARM "yes")
+  set(RTCD_HAVE_NEON "yes")
+elseif ("${AOM_TARGET_CPU}" MATCHES "^armv7")
+  set(ARCH_ARM 1)
+  set(HAVE_NEON 1)
+  set(HAVE_NEON_ASM 1)
+  set(RTCD_ARCH_ARM "yes")
+  set(RTCD_HAVE_NEON "yes")
+  set(RTCD_HAVE_NEON_ASM "yes")
+elseif ("${AOM_TARGET_CPU}" MATCHES "^mips")
+  set(ARCH_MIPS 1)
+
+  if ("${AOM_TARGET_CPU}" STREQUAL "mips32")
+    set(HAVE_MIPS32 1)
+  elseif ("${AOM_TARGET_CPU}" STREQUAL "mips64")
+    set(HAVE_MIPS64 1)
+  endif ()
+
+  set(RTCD_ARCH_MIPS "yes")
+
+  if (HAVE_DSPR2)
+    set(RTCD_HAVE_DSPR2 "yes")
+  endif ()
+
+  if (HAVE_MSA)
+    set(RTCD_HAVE_MSA "yes")
+  endif ()
+elseif ("${AOM_TARGET_CPU}" MATCHES "^x86")
+  if ("${AOM_TARGET_CPU}" STREQUAL "x86")
+    set(ARCH_X86 1)
+    set(RTCD_ARCH_X86 "yes")
+  elseif ("${AOM_TARGET_CPU}" STREQUAL "x86_64")
+    set(ARCH_X86_64 1)
+    set(RTCD_ARCH_X86_64 "yes")
+  endif ()
+
+  set(HAVE_MMX 1)
+  set(HAVE_SSE 1)
+  set(HAVE_SSE2 1)
+  set(HAVE_SSE3 1)
+  set(HAVE_SSSE3 1)
+  set(HAVE_SSE4_1 1)
+  set(HAVE_AVX 1)
+  set(HAVE_AVX2 1)
+  set(RTCD_HAVE_MMX "yes")
+  set(RTCD_HAVE_SSE "yes")
+  set(RTCD_HAVE_SSE2 "yes")
+  set(RTCD_HAVE_SSE3 "yes")
+  set(RTCD_HAVE_SSSE3 "yes")
+  set(RTCD_HAVE_SSE4_1 "yes")
+  set(RTCD_HAVE_AVX "yes")
+  set(RTCD_HAVE_AVX2 "yes")
+endif ()
+
+foreach (config_var ${AOM_CONFIG_VARS})
+  if (${${config_var}})
+    set(RTCD_${config_var} yes)
+  endif ()
+endforeach ()
diff --git a/third_party/aom/build/cmake/generate_aom_config_templates.cmake b/third_party/aom/build/cmake/generate_aom_config_templates.cmake
new file mode 100644
index 0000000000..aea4253bbe
--- /dev/null
+++ b/third_party/aom/build/cmake/generate_aom_config_templates.cmake
@@ -0,0 +1,84 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+cmake_minimum_required(VERSION 3.5)
+
+string(TIMESTAMP year "%Y")
+set(asm_file_header_block
+"\;
+\; Copyright (c) ${year}, Alliance for Open Media. All rights reserved
+\;
+\; This source code is subject to the terms of the BSD 2 Clause License and
+\; the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+\; was not distributed with this source code in the LICENSE file, you can
+\; obtain it at www.aomedia.org/license/software. If the Alliance for Open
+\; Media Patent License 1.0 was not distributed with this source code in the
+\; PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+\;
+")
+set(h_file_header_block
+"/*
+ * Copyright (c) ${year}, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+\#ifndef AOM_CONFIG_H_
+\#define AOM_CONFIG_H_
+")
+
+# Terminates cmake execution when $var_name is an empty string, or the variable
+# name it contains does not expand to an existing directory.
+function (check_directory_var var_name)
+  if ("${var_name}" STREQUAL "")
+    message(FATAL_ERROR "The CMake variable ${var_name} must be defined.")
+  endif ()
+
+  if (NOT EXISTS "${${var_name}}")
+    message(FATAL_ERROR "${${var_name}} (${var_name}) missing.")
+  endif ()
+endfunction ()
+
+check_directory_var(AOM_CONFIG_DIR)
+check_directory_var(AOM_ROOT)
+
+set(AOM_DEFAULTS "${AOM_ROOT}/build/cmake/aom_config_defaults.cmake")
+if (NOT EXISTS "${AOM_DEFAULTS}")
+  message(FATAL_ERROR
+          "Configuration default values file (${AOM_DEFAULTS}) missing.")
+endif ()
+
+include("${AOM_ROOT}/build/cmake/aom_config_defaults.cmake")
+get_cmake_property(cmake_cache_vars CACHE_VARIABLES)
+
+set(aom_config_h_template "${AOM_CONFIG_DIR}/aom_config.h.cmake")
+file(WRITE "${aom_config_h_template}" ${h_file_header_block})
+foreach(cache_var ${cmake_cache_vars})
+  if (NOT "${cache_var}" MATCHES "AOM_CONFIG_DIR\|AOM_ROOT\|^CMAKE_")
+    file(APPEND
+         "${aom_config_h_template}" "\#define ${cache_var} \${${cache_var}}\n")
+  endif ()
+endforeach()
+file(APPEND "${aom_config_h_template}" "\#endif  /* AOM_CONFIG_H_ */")
+
+set(aom_asm_config_template "${AOM_CONFIG_DIR}/aom_config.asm.cmake")
+file(WRITE "${aom_asm_config_template}" ${asm_file_header_block})
+foreach(cache_var ${cmake_cache_vars})
+  if (NOT "${cache_var}" MATCHES
+      "AOM_CONFIG_DIR\|AOM_ROOT\|^CMAKE_\|INLINE\|RESTRICT")
+    file(APPEND "${aom_asm_config_template}"
+         "${cache_var} equ \${${cache_var}}\n")
+  endif ()
+endforeach ()
+
diff --git a/third_party/aom/build/cmake/msvc_runtime.cmake b/third_party/aom/build/cmake/msvc_runtime.cmake
new file mode 100644
index 0000000000..0327217b4d
--- /dev/null
+++ b/third_party/aom/build/cmake/msvc_runtime.cmake
@@ -0,0 +1,26 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+if (MSVC)
+  # CMake defaults to producing code linked to the DLL MSVC runtime. That will
+  # not work with googletest, and isn't what we want anyway.
+  if (NOT "${MSVC_RUNTIME}" STREQUAL "dll")
+    foreach (flag_var
+             CMAKE_C_FLAGS CMAKE_C_FLAGS_DEBUG CMAKE_C_FLAGS_RELEASE
+             CMAKE_C_FLAGS_MINSIZEREL CMAKE_C_FLAGS_RELWITHDEBINFO
+             CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
+             CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
+      if (${flag_var} MATCHES "/MD")
+        string(REGEX REPLACE "/MD" "/MT" ${flag_var} "${${flag_var}}")
+      endif (${flag_var} MATCHES "/MD")
+    endforeach (flag_var)
+  endif ()
+endif ()
diff --git a/third_party/aom/build/cmake/rtcd_config.cmake b/third_party/aom/build/cmake/rtcd_config.cmake
new file mode 100644
index 0000000000..cdea3452a7
--- /dev/null
+++ b/third_party/aom/build/cmake/rtcd_config.cmake
@@ -0,0 +1,137 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+ARCH_ARM=${RTCD_ARCH_ARM}
+ARCH_MIPS=${RTCD_ARCH_MIPS}
+ARCH_X86=${RTCD_ARCH_X86}
+ARCH_X86_64=${RTCD_ARCH_X86_64}
+HAVE_NEON=${RTCD_HAVE_NEON}
+HAVE_NEON_ASM=${RTCD_HAVE_NEON_ASM}
+HAVE_MIPS32=${RTCD_HAVE_MIPS32}
+HAVE_DSPR2=${RTCD_HAVE_DSPR2}
+HAVE_MSA=${RTCD_HAVE_MSA}
+HAVE_MIPS64=${RTCD_HAVE_MIPS64}
+HAVE_MMX=${RTCD_HAVE_MMX}
+HAVE_SSE=${RTCD_HAVE_SSE}
+HAVE_SSE2=${RTCD_HAVE_SSE2}
+HAVE_SSE3=${RTCD_HAVE_SSE3}
+HAVE_SSSE3=${RTCD_HAVE_SSSE3}
+HAVE_SSE4_1=${RTCD_HAVE_SSE4_1}
+HAVE_AVX=${RTCD_HAVE_AVX}
+HAVE_AVX2=${RTCD_HAVE_AVX2}
+CONFIG_ACCOUNTING=${RTCD_CONFIG_ACCOUNTING}
+CONFIG_INSPECTION=${RTCD_CONFIG_INSPECTION}
+CONFIG_ADAPT_SCAN=${RTCD_CONFIG_ADAPT_SCAN}
+CONFIG_ALT_INTRA=${RTCD_CONFIG_ALT_INTRA}
+CONFIG_ANS=${RTCD_CONFIG_ANS}
+CONFIG_HIGHBITDEPTH=${RTCD_CONFIG_HIGHBITDEPTH}
+CONFIG_AOM_QM=${RTCD_CONFIG_AOM_QM}
+CONFIG_AV1=${RTCD_CONFIG_AV1}
+CONFIG_AV1_DECODER=${RTCD_CONFIG_AV1_DECODER}
+CONFIG_AV1_ENCODER=${RTCD_CONFIG_AV1_ENCODER}
+CONFIG_BIG_ENDIAN=${RTCD_CONFIG_BIG_ENDIAN}
+CONFIG_BITSTREAM_DEBUG=${RTCD_CONFIG_BITSTREAM_DEBUG}
+CONFIG_CB4X4=${RTCD_CONFIG_CB4X4}
+CONFIG_CDEF=${RTCD_CONFIG_CDEF}
+CONFIG_CHROMA_2X2=${RTCD_CONFIG_CHROMA_2X2}
+CONFIG_CODEC_SRCS=${RTCD_CONFIG_CODEC_SRCS}
+CONFIG_COEFFICIENT_RANGE_CHECKING=${RTCD_CONFIG_COEFFICIENT_RANGE_CHECKING}
+CONFIG_COEF_INTERLEAVE=${RTCD_CONFIG_COEF_INTERLEAVE}
+CONFIG_COMPOUND_SEGMENT=${RTCD_CONFIG_COMPOUND_SEGMENT}
+CONFIG_CONVOLVE_ROUND=${RTCD_CONFIG_CONVOLVE_ROUND}
+CONFIG_DAALA_DIST=${RTCD_CONFIG_DAALA_DIST}
+CONFIG_DAALA_EC=${RTCD_CONFIG_DAALA_EC}
+CONFIG_DEBUG=${RTCD_CONFIG_DEBUG}
+CONFIG_DEBUG_LIBS=${RTCD_CONFIG_DEBUG_LIBS}
+CONFIG_DECODERS=${RTCD_CONFIG_DECODERS}
+CONFIG_DECODE_PERF_TESTS=${RTCD_CONFIG_DECODE_PERF_TESTS}
+CONFIG_DELTA_Q=${RTCD_CONFIG_DELTA_Q}
+CONFIG_DEPENDENCY_TRACKING=${RTCD_CONFIG_DEPENDENCY_TRACKING}
+CONFIG_DEPENDENT_HORZTILES=${RTCD_CONFIG_DEPENDENT_HORZTILES}
+CONFIG_DUAL_FILTER=${RTCD_CONFIG_DUAL_FILTER}
+CONFIG_EC_ADAPT=${RTCD_CONFIG_EC_ADAPT}
+CONFIG_EC_MULTISYMBOL=${RTCD_CONFIG_EC_MULTISYMBOL}
+CONFIG_ENCODERS=${RTCD_CONFIG_ENCODERS}
+CONFIG_ENCODE_PERF_TESTS=${RTCD_CONFIG_ENCODE_PERF_TESTS}
+CONFIG_ENTROPY_STATS=${RTCD_CONFIG_ENTROPY_STATS}
+CONFIG_ERROR_CONCEALMENT=${RTCD_CONFIG_ERROR_CONCEALMENT}
+CONFIG_EXPERIMENTAL=${RTCD_CONFIG_EXPERIMENTAL}
+CONFIG_EXTERNAL_BUILD=${RTCD_CONFIG_EXTERNAL_BUILD}
+CONFIG_EXT_INTER=${RTCD_CONFIG_EXT_INTER}
+CONFIG_EXT_INTRA=${RTCD_CONFIG_EXT_INTRA}
+CONFIG_EXT_PARTITION=${RTCD_CONFIG_EXT_PARTITION}
+CONFIG_EXT_PARTITION_TYPES=${RTCD_CONFIG_EXT_PARTITION_TYPES}
+CONFIG_EXT_REFS=${RTCD_CONFIG_EXT_REFS}
+CONFIG_EXT_TILE=${RTCD_CONFIG_EXT_TILE}
+CONFIG_EXT_TX=${RTCD_CONFIG_EXT_TX}
+CONFIG_FILTER_7BIT=${RTCD_CONFIG_FILTER_7BIT}
+CONFIG_FILTER_INTRA=${RTCD_CONFIG_FILTER_INTRA}
+CONFIG_FP_MB_STATS=${RTCD_CONFIG_FP_MB_STATS}
+CONFIG_FRAME_SIZE=${RTCD_CONFIG_FRAME_SIZE}
+CONFIG_GCC=${RTCD_CONFIG_GCC}
+CONFIG_GCOV=${RTCD_CONFIG_GCOV}
+CONFIG_GLOBAL_MOTION=${RTCD_CONFIG_GLOBAL_MOTION}
+CONFIG_GPROF=${RTCD_CONFIG_GPROF}
+CONFIG_INSTALL_BINS=${RTCD_CONFIG_INSTALL_BINS}
+CONFIG_INSTALL_DOCS=${RTCD_CONFIG_INSTALL_DOCS}
+CONFIG_INSTALL_LIBS=${RTCD_CONFIG_INSTALL_LIBS}
+CONFIG_INSTALL_SRCS=${RTCD_CONFIG_INSTALL_SRCS}
+CONFIG_INTERNAL_STATS=${RTCD_CONFIG_INTERNAL_STATS}
+CONFIG_INTRA_INTERP=${RTCD_CONFIG_INTRA_INTERP}
+CONFIG_LIBYUV=${RTCD_CONFIG_LIBYUV}
+CONFIG_LOOPFILTERING_ACROSS_TILES=${RTCD_CONFIG_LOOPFILTERING_ACROSS_TILES}
+CONFIG_LOOP_RESTORATION=${RTCD_CONFIG_LOOP_RESTORATION}
+CONFIG_LOWBITDEPTH=${RTCD_CONFIG_LOWBITDEPTH}
+CONFIG_LV_MAP=${RTCD_CONFIG_LV_MAP}
+CONFIG_MASKED_TX=${RTCD_CONFIG_MASKED_TX}
+CONFIG_MOTION_VAR=${RTCD_CONFIG_MOTION_VAR}
+CONFIG_MSVS=${RTCD_CONFIG_MSVS}
+CONFIG_MULTITHREAD=${RTCD_CONFIG_MULTITHREAD}
+CONFIG_MV_COMPRESS=${RTCD_CONFIG_MV_COMPRESS}
+CONFIG_NCOBMC=${RTCD_CONFIG_NCOBMC}
+CONFIG_NEW_QUANT=${RTCD_CONFIG_NEW_QUANT}
+CONFIG_NEW_TOKENSET=${RTCD_CONFIG_NEW_TOKENSET}
+CONFIG_ONTHEFLY_BITPACKING=${RTCD_CONFIG_ONTHEFLY_BITPACKING}
+CONFIG_OS_SUPPORT=${RTCD_CONFIG_OS_SUPPORT}
+CONFIG_PALETTE=${RTCD_CONFIG_PALETTE}
+CONFIG_PALETTE_THROUGHPUT=${RTCD_CONFIG_PALETTE_THROUGHPUT}
+CONFIG_PARALLEL_DEBLOCKING=${RTCD_CONFIG_PARALLEL_DEBLOCKING}
+CONFIG_PIC=${RTCD_CONFIG_PIC}
+CONFIG_POSTPROC=${RTCD_CONFIG_POSTPROC}
+CONFIG_POSTPROC_VISUALIZER=${RTCD_CONFIG_POSTPROC_VISUALIZER}
+CONFIG_PVQ=${RTCD_CONFIG_PVQ}
+CONFIG_RAWBITS=${RTCD_CONFIG_RAWBITS}
+CONFIG_RD_DEBUG=${RTCD_CONFIG_RD_DEBUG}
+CONFIG_REALTIME_ONLY=${RTCD_CONFIG_REALTIME_ONLY}
+CONFIG_RECT_TX=${RTCD_CONFIG_RECT_TX}
+CONFIG_REFERENCE_BUFFER=${RTCD_CONFIG_REFERENCE_BUFFER}
+CONFIG_REF_ADAPT=${RTCD_CONFIG_REF_ADAPT}
+CONFIG_REF_MV=${RTCD_CONFIG_REF_MV}
+CONFIG_RUNTIME_CPU_DETECT=${RTCD_CONFIG_RUNTIME_CPU_DETECT}
+CONFIG_RVCT=${RTCD_CONFIG_RVCT}
+CONFIG_SHARED=${RTCD_CONFIG_SHARED}
+CONFIG_SIZE_LIMIT=${RTCD_CONFIG_SIZE_LIMIT}
+CONFIG_SMALL=${RTCD_CONFIG_SMALL}
+CONFIG_SPATIAL_RESAMPLING=${RTCD_CONFIG_SPATIAL_RESAMPLING}
+CONFIG_STATIC=${RTCD_CONFIG_STATIC}
+CONFIG_STATIC_MSVCRT=${RTCD_CONFIG_STATIC_MSVCRT}
+CONFIG_SUB8X8_MC=${RTCD_CONFIG_SUB8X8_MC}
+CONFIG_SUPERTX=${RTCD_CONFIG_SUPERTX}
+CONFIG_TEMPMV_SIGNALING=${RTCD_CONFIG_TEMPMV_SIGNALING}
+CONFIG_TILE_GROUPS=${RTCD_CONFIG_TILE_GROUPS}
+CONFIG_TPL_MV=${RTCD_CONFIG_TPL_MV}
+CONFIG_TRIPRED=${RTCD_CONFIG_TRIPRED}
+CONFIG_TX64X64=${RTCD_CONFIG_TX64X64}
+CONFIG_UNIT_TESTS=${RTCD_CONFIG_UNIT_TESTS}
+CONFIG_UNPOISON_PARTITION_CTX=${RTCD_CONFIG_UNPOISON_PARTITION_CTX}
+CONFIG_VAR_TX=${RTCD_CONFIG_VAR_TX}
+CONFIG_WARPED_MOTION=${RTCD_CONFIG_WARPED_MOTION}
+CONFIG_WEBM_IO=${RTCD_CONFIG_WEBM_IO}
+CONFIG_XIPHRC=${RTCD_CONFIG_XIPHRC}
diff --git a/third_party/aom/build/cmake/toolchains/arm-ios-common.cmake b/third_party/aom/build/cmake/toolchains/arm-ios-common.cmake
new file mode 100644
index 0000000000..8317ae2725
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/arm-ios-common.cmake
@@ -0,0 +1,31 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_ARM_IOS_COMMON_CMAKE_)
+set(AOM_BUILD_CMAKE_ARM_IOS_COMMON_CMAKE_ 1)
+
+set(CMAKE_SYSTEM_NAME "Darwin")
+set(CMAKE_OSX_SYSROOT iphoneos)
+set(CMAKE_C_COMPILER clang)
+set(CMAKE_C_COMPILER_ARG1 "-arch ${CMAKE_SYSTEM_PROCESSOR}")
+set(CMAKE_CXX_COMPILER clang++)
+set(CMAKE_CXX_COMPILER_ARG1 "-arch ${CMAKE_SYSTEM_PROCESSOR}")
+
+# Assembler sources must be converted for ARM iOS targets.
+set(AOM_ADS2GAS_REQUIRED 1)
+set(AOM_ADS2GAS "${CMAKE_CURRENT_SOURCE_DIR}/build/make/ads2gas_apple.pl")
+set(AOM_GAS_EXT "S")
+
+# No runtime cpu detect for arm*-ios targets.
+set(CONFIG_RUNTIME_CPU_DETECT 0 CACHE BOOL "")
+
+# TODO(tomfinegan): Handle bit code embedding.
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_ARM_IOS_COMMON_CMAKE_
diff --git a/third_party/aom/build/cmake/toolchains/arm64-ios.cmake b/third_party/aom/build/cmake/toolchains/arm64-ios.cmake
new file mode 100644
index 0000000000..434809db91
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/arm64-ios.cmake
@@ -0,0 +1,24 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_ARM64_IOS_CMAKE_)
+set(AOM_BUILD_CMAKE_TOOLCHAINS_ARM64_IOS_CMAKE_ 1)
+
+if (XCODE)
+  # TODO(tomfinegan): Handle arm builds in Xcode.
+  message(FATAL_ERROR "This toolchain does not support Xcode.")
+endif ()
+
+set(CMAKE_SYSTEM_PROCESSOR "arm64")
+set(CMAKE_OSX_ARCHITECTURES "arm64")
+
+include("${CMAKE_CURRENT_LIST_DIR}/arm-ios-common.cmake")
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_ARM64_IOS_CMAKE_
diff --git a/third_party/aom/build/cmake/toolchains/arm64-linux-gcc.cmake b/third_party/aom/build/cmake/toolchains/arm64-linux-gcc.cmake
new file mode 100644
index 0000000000..b8efe6be66
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/arm64-linux-gcc.cmake
@@ -0,0 +1,35 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_ARM64_LINUX_GCC_CMAKE_)
+set(AOM_BUILD_CMAKE_TOOLCHAINS_ARM64_LINUX_GCC_CMAKE_ 1)
+
+set(CMAKE_SYSTEM_NAME "Linux")
+
+if ("${CROSS}" STREQUAL "")
+  # Default the cross compiler prefix to something known to work.
+  set(CROSS aarch64-linux-gnu-)
+endif ()
+
+set(CMAKE_C_COMPILER ${CROSS}gcc)
+set(CMAKE_CXX_COMPILER ${CROSS}g++)
+set(AS_EXECUTABLE ${CROSS}as)
+set(CMAKE_C_COMPILER_ARG1 "-march=armv8-a")
+set(CMAKE_CXX_COMPILER_ARG1 "-march=armv8-a")
+set(AOM_AS_FLAGS "-march=armv8-a")
+set(CMAKE_SYSTEM_PROCESSOR "arm64")
+
+# No intrinsics flag required for arm64-linux-gcc.
+set(AOM_NEON_INTRIN_FLAG "")
+
+# No runtime cpu detect for arm64-linux-gcc.
+set(CONFIG_RUNTIME_CPU_DETECT 0 CACHE BOOL "")
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_ARM64_LINUX_GCC_CMAKE_
diff --git a/third_party/aom/build/cmake/toolchains/armv7-ios.cmake b/third_party/aom/build/cmake/toolchains/armv7-ios.cmake
new file mode 100644
index 0000000000..bcd37a06d7
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/armv7-ios.cmake
@@ -0,0 +1,34 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_ARMV7_IOS_CMAKE_)
+set(AOM_BUILD_CMAKE_TOOLCHAINS_ARMV7_IOS_CMAKE_ 1)
+
+if (XCODE)
+  # TODO(tomfinegan): Handle arm builds in Xcode.
+  message(FATAL_ERROR "This toolchain does not support Xcode.")
+endif ()
+
+set(CMAKE_SYSTEM_PROCESSOR "armv7")
+set(CMAKE_OSX_ARCHITECTURES "armv7")
+
+include("${CMAKE_CURRENT_LIST_DIR}/arm-ios-common.cmake")
+
+# No intrinsics flag required for armv7s-ios.
+set(AOM_NEON_INTRIN_FLAG "")
+
+# No runtime cpu detect for armv7s-ios.
+set(CONFIG_RUNTIME_CPU_DETECT 0 CACHE BOOL "")
+
+# RTCD generation requires --disable-media for armv7s-ios.
+set(AOM_RTCD_FLAGS ${AOM_RTCD_FLAGS} --disable-media)
+string(STRIP AOM_RTCD_FLAGS ${AOM_RTCD_FLAGS})
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_ARMV7_IOS_CMAKE_
diff --git a/third_party/aom/build/cmake/toolchains/armv7-linux-gcc.cmake b/third_party/aom/build/cmake/toolchains/armv7-linux-gcc.cmake
new file mode 100644
index 0000000000..eedfda4641
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/armv7-linux-gcc.cmake
@@ -0,0 +1,48 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_ARMV7_LINUX_GCC_CMAKE_)
+set(AOM_BUILD_CMAKE_TOOLCHAINS_ARMV7_LINUX_GCC_CMAKE_ 1)
+
+set(CMAKE_SYSTEM_NAME "Linux")
+
+if ("${CROSS}" STREQUAL "")
+  # Default the cross compiler prefix to something known to work.
+  set(CROSS arm-linux-gnueabihf-)
+endif ()
+
+if (NOT ${CROSS} MATCHES hf-$)
+  set(AOM_EXTRA_TOOLCHAIN_FLAGS "-mfloat-abi=softfp")
+endif ()
+
+set(CMAKE_C_COMPILER ${CROSS}gcc)
+set(CMAKE_CXX_COMPILER ${CROSS}g++)
+set(AS_EXECUTABLE ${CROSS}as)
+set(CMAKE_C_COMPILER_ARG1
+    "-march=armv7-a -mfpu=neon ${AOM_EXTRA_TOOLCHAIN_FLAGS}")
+set(CMAKE_CXX_COMPILER_ARG1
+    "-march=armv7-a -mfpu=neon ${AOM_EXTRA_TOOLCHAIN_FLAGS}")
+set(AOM_AS_FLAGS
+    --defsym ARCHITECTURE=7 -march=armv7-a -mfpu=neon
+    ${AOM_EXTRA_TOOLCHAIN_FLAGS})
+set(CMAKE_SYSTEM_PROCESSOR "armv7")
+
+# No intrinsics flag required for armv7-linux-gcc.
+set(AOM_NEON_INTRIN_FLAG "")
+
+# Assembler sources must be converted for armv7-linux-gcc targets.
+set(AOM_ADS2GAS_REQUIRED 1)
+set(AOM_ADS2GAS "${CMAKE_CURRENT_SOURCE_DIR}/build/make/ads2gas.pl")
+set(AOM_GAS_EXT "S")
+
+# No runtime cpu detect for armv7-linux-gcc.
+set(CONFIG_RUNTIME_CPU_DETECT 0 CACHE BOOL "")
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_ARMV7_LINUX_GCC_CMAKE_
diff --git a/third_party/aom/build/cmake/toolchains/armv7s-ios.cmake b/third_party/aom/build/cmake/toolchains/armv7s-ios.cmake
new file mode 100644
index 0000000000..08a0a37ee7
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/armv7s-ios.cmake
@@ -0,0 +1,34 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_ARMV7S_IOS_CMAKE_)
+set(AOM_BUILD_CMAKE_TOOLCHAINS_ARMV7S_IOS_CMAKE_ 1)
+
+if (XCODE)
+  # TODO(tomfinegan): Handle arm builds in Xcode.
+  message(FATAL_ERROR "This toolchain does not support Xcode.")
+endif ()
+
+set(CMAKE_SYSTEM_PROCESSOR "armv7s")
+set(CMAKE_OSX_ARCHITECTURES "armv7s")
+
+include("${CMAKE_CURRENT_LIST_DIR}/arm-ios-common.cmake")
+
+# No intrinsics flag required for armv7s-ios.
+set(AOM_NEON_INTRIN_FLAG "")
+
+# No runtime cpu detect for armv7s-ios.
+set(CONFIG_RUNTIME_CPU_DETECT 0 CACHE BOOL "")
+
+# RTCD generation requires --disable-media for armv7s-ios.
+set(AOM_RTCD_FLAGS ${AOM_RTCD_FLAGS} --disable-media)
+string(STRIP AOM_RTCD_FLAGS ${AOM_RTCD_FLAGS})
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_ARMV7S_IOS_CMAKE_
diff --git a/third_party/aom/build/cmake/toolchains/ios-simulator-common.cmake b/third_party/aom/build/cmake/toolchains/ios-simulator-common.cmake
new file mode 100644
index 0000000000..7a28e329c6
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/ios-simulator-common.cmake
@@ -0,0 +1,23 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_IOS_SIMULATOR_COMMON_CMAKE_)
+set(AOM_BUILD_CMAKE_IOS_SIMULATOR_COMMON_CMAKE_ 1)
+
+set(CMAKE_SYSTEM_NAME "Darwin")
+set(CMAKE_OSX_SYSROOT iphonesimulator)
+set(CMAKE_C_COMPILER clang)
+set(CMAKE_C_COMPILER_ARG1 "-arch ${CMAKE_SYSTEM_PROCESSOR}")
+set(CMAKE_CXX_COMPILER clang++)
+set(CMAKE_CXX_COMPILER_ARG1 "-arch ${CMAKE_SYSTEM_PROCESSOR}")
+
+# TODO(tomfinegan): Handle bit code embedding.
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_IOS_SIMULATOR_COMMON_CMAKE_
diff --git a/third_party/aom/build/cmake/toolchains/mips32-linux-gcc.cmake b/third_party/aom/build/cmake/toolchains/mips32-linux-gcc.cmake
new file mode 100644
index 0000000000..a55c41115f
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/mips32-linux-gcc.cmake
@@ -0,0 +1,71 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_MIPS32_LINUX_GCC_CMAKE_)
+set(AOM_BUILD_CMAKE_TOOLCHAINS_MIPS32_LINUX_GCC_CMAKE_ 1)
+
+set(CMAKE_SYSTEM_NAME "Linux")
+
+if (ENABLE_DSPR2 AND ENABLE_MSA)
+  message(FATAL_ERROR "ENABLE_DSPR2 and ENABLE_MSA cannot be combined.")
+endif ()
+
+if (ENABLE_DSPR2)
+  set(HAVE_DSPR2 1 CACHE BOOL "" FORCE)
+
+  if ("${CROSS}" STREQUAL "")
+    # Default the cross compiler prefix to something known to work.
+    set(CROSS mips-linux-gnu-)
+  endif ()
+
+  set(MIPS_CFLAGS "-mdspr2")
+  set(MIPS_CXXFLAGS "-mdspr2")
+elseif (ENABLE_MSA)
+  set(HAVE_MSA 1 CACHE BOOL "" FORCE)
+
+  if ("${CROSS}" STREQUAL "")
+    # Default the cross compiler prefix to something known to work.
+    set(CROSS mips-mti-linux-gnu-)
+  endif ()
+
+  set(MIPS_CFLAGS "-mmsa")
+  set(MIPS_CXXFLAGS "-mmsa")
+endif ()
+
+if ("${CROSS}" STREQUAL "")
+  # TODO(tomfinegan): Make it possible to turn this off. The $CROSS prefix
+  # won't be desired on a mips host.
+  # Default cross compiler prefix to something that might work for an
+  # unoptimized build.
+  set(CROSS mips-linux-gnu-)
+endif ()
+
+if ("${MIPS_CPU}" STREQUAL "")
+  set(MIPS_CFLAGS "${MIPS_CFLAGS} -mips32r2")
+  set(MIPS_CXXFLAGS "${MIPS_CXXFLAGS} -mips32r2")
+elseif ("${MIPS_CPU}" STREQUAL "p5600")
+  set(P56_FLAGS
+      "-mips32r5 -mload-store-pairs -msched-weight -mhard-float -mfp64")
+  set(MIPS_CFLAGS "${MIPS_CFLAGS} ${P56_FLAGS}")
+  set(MIPS_CXXFLAGS "${MIPS_CXXFLAGS} ${P56_FLAGS}")
+  set(CMAKE_EXE_LINKER_FLAGS "-mfp64 ${CMAKE_EXE_LINKER_FLAGS}")
+endif ()
+
+set(CMAKE_C_COMPILER ${CROSS}gcc)
+set(CMAKE_CXX_COMPILER ${CROSS}g++)
+set(AS_EXECUTABLE ${CROSS}as)
+set(CMAKE_C_COMPILER_ARG1 "-EL ${MIPS_CFLAGS}")
+set(CMAKE_CXX_COMPILER_ARG1 "-EL ${MIPS_CXXFLAGS}")
+set(CMAKE_SYSTEM_PROCESSOR "mips32")
+
+# No runtime cpu detect for mips32-linux-gcc.
+set(CONFIG_RUNTIME_CPU_DETECT 0 CACHE BOOL "")
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_MIPS32_LINUX_GCC_CMAKE_
diff --git a/third_party/aom/build/cmake/toolchains/mips64-linux-gcc.cmake b/third_party/aom/build/cmake/toolchains/mips64-linux-gcc.cmake
new file mode 100644
index 0000000000..28b1582cc3
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/mips64-linux-gcc.cmake
@@ -0,0 +1,48 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_MIPS64_LINUX_GCC_CMAKE_)
+set(AOM_BUILD_CMAKE_TOOLCHAINS_MIPS64_LINUX_GCC_CMAKE_ 1)
+
+set(CMAKE_SYSTEM_NAME "Linux")
+
+if ("${CROSS}" STREQUAL "")
+  # TODO(tomfinegan): Make it possible to turn this off. The $CROSS prefix
+  # won't be desired on a mips host.
+  # Default the cross compiler prefix to something known to work.
+  set(CROSS mips-img-linux-gnu-)
+endif ()
+
+if (ENABLE_MSA)
+  set(HAVE_MSA 1 CACHE BOOL "" FORCE)
+  set(MIPS_CFLAGS "-mmsa")
+  set(MIPS_CXXFLAGS "-mmsa")
+endif ()
+
+if ("${MIPS_CPU}" STREQUAL "i6400" OR "${MIPS_CPU}" STREQUAL "p6600")
+  set(MIPS_CPU_FLAGS "-mips64r6 -mabi=64 -mload-store-pairs -msched-weight")
+  set(MIPS_CPU_FLAGS "${MIPS_CPU_FLAGS} -mhard-float -mfp64")
+  set(MIPS_CFLAGS "${MIPS_CFLAGS} ${MIPS_CPU_FLAGS}")
+  set(MIPS_CXXFLAGS "${MIPS_CXXFLAGS} ${MIPS_CPU_FLAGS}")
+  set(CMAKE_EXE_LINKER_FLAGS
+      "-mips64r6 -mabi64 -mfp64 ${CMAKE_EXE_LINKER_FLAGS}")
+endif ()
+
+set(CMAKE_C_COMPILER ${CROSS}gcc)
+set(CMAKE_CXX_COMPILER ${CROSS}g++)
+set(AS_EXECUTABLE ${CROSS}as)
+set(CMAKE_C_COMPILER_ARG1 "-EL ${MIPS_CFLAGS}")
+set(CMAKE_CXX_COMPILER_ARG1 "-EL ${MIPS_CXXFLAGS}")
+set(CMAKE_SYSTEM_PROCESSOR "mips64")
+
+# No runtime cpu detect for mips64-linux-gcc.
+set(CONFIG_RUNTIME_CPU_DETECT 0 CACHE BOOL "")
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_MIPS64_LINUX_GCC_CMAKE_
diff --git a/third_party/aom/build/cmake/toolchains/x86-ios-simulator.cmake b/third_party/aom/build/cmake/toolchains/x86-ios-simulator.cmake
new file mode 100644
index 0000000000..4e4ebc0348
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/x86-ios-simulator.cmake
@@ -0,0 +1,27 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_X86_IOS_SIMULATOR_CMAKE_)
+set(AOM_BUILD_CMAKE_TOOLCHAINS_X86_IOS_SIMULATOR_CMAKE_ 1)
+
+if (XCODE)
+  # TODO(tomfinegan): Handle ios sim builds in Xcode.
+  message(FATAL_ERROR "This toolchain does not support Xcode.")
+endif ()
+
+set(CMAKE_SYSTEM_PROCESSOR "i386")
+set(CMAKE_OSX_ARCHITECTURES "i386")
+
+# Avoid noisy PIC/PIE warnings.
+set(CONFIG_PIC 1 CACHE BOOL "")
+
+include("${CMAKE_CURRENT_LIST_DIR}/ios-simulator-common.cmake")
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_X86_IOS_SIMULATOR_CMAKE_
diff --git a/third_party/aom/build/cmake/toolchains/x86-linux.cmake b/third_party/aom/build/cmake/toolchains/x86-linux.cmake
new file mode 100644
index 0000000000..077c8f325b
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/x86-linux.cmake
@@ -0,0 +1,14 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+set(CMAKE_SYSTEM_PROCESSOR "x86")
+set(CMAKE_SYSTEM_NAME "Linux")
+set(CMAKE_C_COMPILER_ARG1 "-m32")
+set(CMAKE_CXX_COMPILER_ARG1 "-m32")
diff --git a/third_party/aom/build/cmake/toolchains/x86-macos.cmake b/third_party/aom/build/cmake/toolchains/x86-macos.cmake
new file mode 100644
index 0000000000..22d0171a71
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/x86-macos.cmake
@@ -0,0 +1,18 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+set(CMAKE_SYSTEM_PROCESSOR "x86")
+set(CMAKE_SYSTEM_NAME "Darwin")
+set(CMAKE_OSX_ARCHITECTURES "i386")
+set(CMAKE_C_COMPILER_ARG1 "-arch i386")
+set(CMAKE_CXX_COMPILER_ARG1 "-arch i386")
+
+# Apple tools always complain in 32 bit mode without PIC.
+set(CONFIG_PIC 1 CACHE STRING "")
diff --git a/third_party/aom/build/cmake/toolchains/x86_64-ios-simulator.cmake b/third_party/aom/build/cmake/toolchains/x86_64-ios-simulator.cmake
new file mode 100644
index 0000000000..884540a9d2
--- /dev/null
+++ b/third_party/aom/build/cmake/toolchains/x86_64-ios-simulator.cmake
@@ -0,0 +1,24 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_BUILD_CMAKE_TOOLCHAINS_X86_64_IOS_SIMULATOR_CMAKE_)
+set(AOM_BUILD_CMAKE_TOOLCHAINS_X86_64_IOS_SIMULATOR_CMAKE_ 1)
+
+if (XCODE)
+  # TODO(tomfinegan): Handle ios sim builds in Xcode.
+  message(FATAL_ERROR "This toolchain does not support Xcode.")
+endif ()
+
+set(CMAKE_SYSTEM_PROCESSOR "x86_64")
+set(CMAKE_OSX_ARCHITECTURES "x86_64")
+
+include("${CMAKE_CURRENT_LIST_DIR}/ios-simulator-common.cmake")
+
+endif ()  # AOM_BUILD_CMAKE_TOOLCHAINS_X86_64_IOS_SIMULATOR_CMAKE_
diff --git a/third_party/aom/build/make/Android.mk b/third_party/aom/build/make/Android.mk
new file mode 100644
index 0000000000..6757b1f595
--- /dev/null
+++ b/third_party/aom/build/make/Android.mk
@@ -0,0 +1,201 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+#
+# This file is to be used for compiling libaom for Android using the NDK.
+# In an Android project place a libaom checkout in the jni directory.
+# Run the configure script from the jni directory.  Base libaom
+# encoder/decoder configuration will look similar to:
+# ./libaom/configure --target=armv7-android-gcc --disable-examples \
+#                    --sdk-path=/opt/android-ndk-r6b/
+#
+# When targeting Android, realtime-only is enabled by default.  This can
+# be overridden by adding the command line flag:
+#  --disable-realtime-only
+#
+# This will create .mk files that contain variables that contain the
+# source files to compile.
+#
+# Place an Android.mk file in the jni directory that references the
+# Android.mk file in the libaom directory:
+# LOCAL_PATH := $(call my-dir)
+# include $(CLEAR_VARS)
+# include jni/libaom/build/make/Android.mk
+#
+# There are currently two TARGET_ARCH_ABI targets for ARM.
+# armeabi and armeabi-v7a.  armeabi-v7a is selected by creating an
+# Application.mk in the jni directory that contains:
+# APP_ABI := armeabi-v7a
+#
+# By default libaom will detect at runtime the existance of NEON extension.
+# For this we import the 'cpufeatures' module from the NDK sources.
+# libaom can also be configured without this runtime detection method.
+# Configuring with --disable-runtime-cpu-detect will assume presence of NEON.
+# Configuring with --disable-runtime-cpu-detect --disable-neon \
+#     --disable-neon-asm
+# will remove any NEON dependency.
+
+# To change to building armeabi, run ./libaom/configure again, but with
+# --target=armv6-android-gcc and modify the Application.mk file to
+# set APP_ABI := armeabi
+#
+# Running ndk-build will build libaom and include it in your project.
+#
+
+CONFIG_DIR := $(LOCAL_PATH)/
+LIBAOM_PATH := $(LOCAL_PATH)/libaom
+ASM_CNV_PATH_LOCAL := $(TARGET_ARCH_ABI)/ads2gas
+ASM_CNV_PATH := $(LOCAL_PATH)/$(ASM_CNV_PATH_LOCAL)
+
+# Use the makefiles generated by upstream configure to determine which files to
+# build. Also set any architecture-specific flags.
+ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+  include $(CONFIG_DIR)libs-armv7-android-gcc.mk
+  LOCAL_ARM_MODE := arm
+else ifeq  ($(TARGET_ARCH_ABI),armeabi)
+  include $(CONFIG_DIR)libs-armv6-android-gcc.mk
+  LOCAL_ARM_MODE := arm
+else ifeq  ($(TARGET_ARCH_ABI),arm64-v8a)
+  include $(CONFIG_DIR)libs-armv8-android-gcc.mk
+  LOCAL_ARM_MODE := arm
+else ifeq ($(TARGET_ARCH_ABI),x86)
+  include $(CONFIG_DIR)libs-x86-android-gcc.mk
+else ifeq ($(TARGET_ARCH_ABI),x86_64)
+  include $(CONFIG_DIR)libs-x86_64-android-gcc.mk
+else ifeq ($(TARGET_ARCH_ABI),mips)
+  include $(CONFIG_DIR)libs-mips-android-gcc.mk
+else
+  $(error Not a supported TARGET_ARCH_ABI: $(TARGET_ARCH_ABI))
+endif
+
+# Rule that is normally in Makefile created by libaom
+# configure.  Used to filter out source files based on configuration.
+enabled=$(filter-out $($(1)-no),$($(1)-yes))
+
+# Override the relative path that is defined by the libaom
+# configure process
+SRC_PATH_BARE := $(LIBAOM_PATH)
+
+# Include the list of files to be built
+include $(LIBAOM_PATH)/libs.mk
+
+# Optimise the code. May want to revisit this setting in the future.
+LOCAL_CFLAGS := -O3
+
+# For x86, include the source code in the search path so it will find files
+# like x86inc.asm and x86_abi_support.asm
+LOCAL_ASMFLAGS := -I$(LIBAOM_PATH)
+
+.PRECIOUS: %.asm.s
+$(ASM_CNV_PATH)/libaom/%.asm.s: $(LIBAOM_PATH)/%.asm
+	@mkdir -p $(dir $@)
+	@$(CONFIG_DIR)$(ASM_CONVERSION) <$< > $@
+
+# For building *_rtcd.h, which have rules in libs.mk
+TGT_ISA:=$(word 1, $(subst -, ,$(TOOLCHAIN)))
+target := libs
+
+LOCAL_SRC_FILES += aom_config.c
+
+# Remove duplicate entries
+CODEC_SRCS_UNIQUE = $(sort $(CODEC_SRCS))
+
+# Pull out C files.  aom_config.c is in the immediate directory and
+# so it does not need libaom/ prefixed like the rest of the source files.
+# The neon files with intrinsics need to have .neon appended so the proper
+# flags are applied.
+CODEC_SRCS_C = $(filter %.c, $(CODEC_SRCS_UNIQUE))
+LOCAL_NEON_SRCS_C = $(filter %_neon.c, $(CODEC_SRCS_C))
+LOCAL_CODEC_SRCS_C = $(filter-out aom_config.c %_neon.c, $(CODEC_SRCS_C))
+
+LOCAL_SRC_FILES += $(foreach file, $(LOCAL_CODEC_SRCS_C), libaom/$(file))
+ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+  LOCAL_SRC_FILES += $(foreach file, $(LOCAL_NEON_SRCS_C), libaom/$(file).neon)
+else # If there are neon sources then we are building for arm64 and do not need to specify .neon
+  LOCAL_SRC_FILES += $(foreach file, $(LOCAL_NEON_SRCS_C), libaom/$(file))
+endif
+
+# Pull out assembly files, splitting NEON from the rest.  This is
+# done to specify that the NEON assembly files use NEON assembler flags.
+# x86 assembly matches %.asm, arm matches %.asm.s
+
+# x86:
+
+CODEC_SRCS_ASM_X86 = $(filter %.asm, $(CODEC_SRCS_UNIQUE))
+LOCAL_SRC_FILES += $(foreach file, $(CODEC_SRCS_ASM_X86), libaom/$(file))
+
+# arm:
+CODEC_SRCS_ASM_ARM_ALL = $(filter %.asm.s, $(CODEC_SRCS_UNIQUE))
+CODEC_SRCS_ASM_ARM = $(foreach v, \
+                     $(CODEC_SRCS_ASM_ARM_ALL), \
+                     $(if $(findstring neon,$(v)),,$(v)))
+CODEC_SRCS_ASM_ADS2GAS = $(patsubst %.s, \
+                         $(ASM_CNV_PATH_LOCAL)/libaom/%.s, \
+                         $(CODEC_SRCS_ASM_ARM))
+LOCAL_SRC_FILES += $(CODEC_SRCS_ASM_ADS2GAS)
+
+ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
+  CODEC_SRCS_ASM_NEON = $(foreach v, \
+                        $(CODEC_SRCS_ASM_ARM_ALL),\
+                        $(if $(findstring neon,$(v)),$(v),))
+  CODEC_SRCS_ASM_NEON_ADS2GAS = $(patsubst %.s, \
+                                $(ASM_CNV_PATH_LOCAL)/libaom/%.s, \
+                                $(CODEC_SRCS_ASM_NEON))
+  LOCAL_SRC_FILES += $(patsubst %.s, \
+                     %.s.neon, \
+                     $(CODEC_SRCS_ASM_NEON_ADS2GAS))
+endif
+
+LOCAL_CFLAGS += \
+    -DHAVE_CONFIG_H=aom_config.h \
+    -I$(LIBAOM_PATH) \
+    -I$(ASM_CNV_PATH)
+
+LOCAL_MODULE := libaom
+
+ifeq ($(CONFIG_RUNTIME_CPU_DETECT),yes)
+  LOCAL_STATIC_LIBRARIES := cpufeatures
+endif
+
+# Add a dependency to force generation of the RTCD files.
+define rtcd_dep_template
+rtcd_dep_template_SRCS := $(addprefix $(LOCAL_PATH)/, $(LOCAL_SRC_FILES))
+rtcd_dep_template_SRCS := $$(rtcd_dep_template_SRCS:.neon=)
+ifeq ($(CONFIG_AV1), yes)
+$$(rtcd_dep_template_SRCS): av1_rtcd.h
+endif
+$$(rtcd_dep_template_SRCS): aom_scale_rtcd.h
+$$(rtcd_dep_template_SRCS): aom_dsp_rtcd.h
+
+ifneq ($(findstring $(TARGET_ARCH_ABI),x86 x86_64),)
+$$(rtcd_dep_template_SRCS): aom_config.asm
+endif
+endef
+
+$(eval $(call rtcd_dep_template))
+
+.PHONY: clean
+clean:
+	@echo "Clean: ads2gas files [$(TARGET_ARCH_ABI)]"
+	@$(RM) $(CODEC_SRCS_ASM_ADS2GAS) $(CODEC_SRCS_ASM_NEON_ADS2GAS)
+	@$(RM) -r $(ASM_CNV_PATH)
+	@$(RM) $(CLEAN-OBJS)
+
+ifeq ($(ENABLE_SHARED),1)
+  include $(BUILD_SHARED_LIBRARY)
+else
+  include $(BUILD_STATIC_LIBRARY)
+endif
+
+ifeq ($(CONFIG_RUNTIME_CPU_DETECT),yes)
+$(call import-module,cpufeatures)
+endif
diff --git a/third_party/aom/build/make/Makefile b/third_party/aom/build/make/Makefile
new file mode 100644
index 0000000000..0b869db0a1
--- /dev/null
+++ b/third_party/aom/build/make/Makefile
@@ -0,0 +1,466 @@
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+include config.mk
+quiet?=true
+ifeq ($(target),)
+# If a target wasn't specified, invoke for all enabled targets.
+.DEFAULT:
+	@for t in $(ALL_TARGETS); do \
+	     $(MAKE) --no-print-directory target=$$t $(MAKECMDGOALS) || exit $$?;\
+        done
+all: .DEFAULT
+clean:: .DEFAULT
+exampletest: .DEFAULT
+install:: .DEFAULT
+test:: .DEFAULT
+test-no-data-check:: .DEFAULT
+testdata:: .DEFAULT
+utiltest: .DEFAULT
+exampletest-no-data-check utiltest-no-data-check: .DEFAULT
+test_%: .DEFAULT ;
+
+# Note: md5sum is not installed on OS X, but openssl is. Openssl may not be
+# installed on cygwin, so we need to autodetect here.
+md5sum := $(firstword $(wildcard \
+          $(foreach e,md5sum openssl,\
+          $(foreach p,$(subst :, ,$(PATH)),$(p)/$(e)*))\
+          ))
+md5sum := $(if $(filter %openssl,$(md5sum)),$(md5sum) dgst -md5,$(md5sum))
+
+TGT_CC:=$(word 3, $(subst -, ,$(TOOLCHAIN)))
+dist:
+	@for t in $(ALL_TARGETS); do \
+	     $(MAKE) --no-print-directory target=$$t $(MAKECMDGOALS) || exit $$?;\
+        done
+        # Run configure for the user with the current toolchain.
+	@if [ -d "$(DIST_DIR)/src" ]; then \
+            mkdir -p "$(DIST_DIR)/build"; \
+            cd "$(DIST_DIR)/build"; \
+            echo "Rerunning configure $(CONFIGURE_ARGS)"; \
+            ../src/configure $(CONFIGURE_ARGS); \
+            $(if $(filter vs%,$(TGT_CC)),make NO_LAUNCH_DEVENV=1;) \
+        fi
+	@if [ -d "$(DIST_DIR)" ]; then \
+            echo "    [MD5SUM] $(DIST_DIR)"; \
+	    cd $(DIST_DIR) && \
+	    $(md5sum) `find . -name md5sums.txt -prune -o -type f -print` \
+                | sed -e 's/MD5(\(.*\))= \([0-9a-f]\{32\}\)/\2  \1/' \
+                > md5sums.txt;\
+        fi
+endif
+
+# Since we invoke make recursively for multiple targets we need to include the
+# .mk file for the correct target, but only when $(target) is non-empty.
+ifneq ($(target),)
+include $(target)-$(TOOLCHAIN).mk
+endif
+BUILD_ROOT?=.
+VPATH=$(SRC_PATH_BARE)
+CFLAGS+=-I$(BUILD_PFX)$(BUILD_ROOT) -I$(SRC_PATH)
+CXXFLAGS+=-I$(BUILD_PFX)$(BUILD_ROOT) -I$(SRC_PATH)
+ASFLAGS+=-I$(BUILD_PFX)$(BUILD_ROOT)/ -I$(SRC_PATH)/
+DIST_DIR?=dist
+HOSTCC?=gcc
+TGT_ISA:=$(word 1, $(subst -, ,$(TOOLCHAIN)))
+TGT_OS:=$(word 2, $(subst -, ,$(TOOLCHAIN)))
+TGT_CC:=$(word 3, $(subst -, ,$(TOOLCHAIN)))
+quiet:=$(if $(or $(verbose), $(V)),, yes)
+qexec=$(if $(quiet),@)
+
+# Cancel built-in implicit rules
+%: %.o
+%.asm:
+%.a:
+%: %.cc
+
+#
+# Common rules"
+#
+.PHONY: all
+all:
+
+.PHONY: clean
+clean::
+	rm -f $(OBJS-yes) $(OBJS-yes:.o=.d) $(OBJS-yes:.asm.s.o=.asm.s)
+	rm -f $(CLEAN-OBJS)
+
+.PHONY: clean
+distclean: clean
+	if [ -z "$(target)" ]; then \
+      rm -f Makefile; \
+      rm -f config.log config.mk; \
+      rm -f aom_config.[hc] aom_config.asm; \
+    else \
+      rm -f $(target)-$(TOOLCHAIN).mk; \
+    fi
+
+.PHONY: dist
+dist:
+.PHONY: exampletest
+exampletest:
+.PHONY: install
+install::
+.PHONY: test
+test::
+.PHONY: testdata
+testdata::
+.PHONY: utiltest
+utiltest:
+.PHONY: test-no-data-check exampletest-no-data-check utiltest-no-data-check
+test-no-data-check::
+exampletest-no-data-check utiltest-no-data-check:
+
+# Force to realign stack always on OS/2
+ifeq ($(TOOLCHAIN), x86-os2-gcc)
+CFLAGS += -mstackrealign
+endif
+
+$(BUILD_PFX)%_mmx.c.d: CFLAGS += -mmmx
+$(BUILD_PFX)%_mmx.c.o: CFLAGS += -mmmx
+$(BUILD_PFX)%_sse2.c.d: CFLAGS += -msse2
+$(BUILD_PFX)%_sse2.c.o: CFLAGS += -msse2
+$(BUILD_PFX)%_sse3.c.d: CFLAGS += -msse3
+$(BUILD_PFX)%_sse3.c.o: CFLAGS += -msse3
+$(BUILD_PFX)%_ssse3.c.d: CFLAGS += -mssse3
+$(BUILD_PFX)%_ssse3.c.o: CFLAGS += -mssse3
+$(BUILD_PFX)%_sse4.c.d: CFLAGS += -msse4.1
+$(BUILD_PFX)%_sse4.c.o: CFLAGS += -msse4.1
+$(BUILD_PFX)%_avx.c.d: CFLAGS += -mavx
+$(BUILD_PFX)%_avx.c.o: CFLAGS += -mavx
+$(BUILD_PFX)%_avx2.c.d: CFLAGS += -mavx2
+$(BUILD_PFX)%_avx2.c.o: CFLAGS += -mavx2
+$(BUILD_PFX)%_mmx.cc.d: CXXFLAGS += -mmmx
+$(BUILD_PFX)%_mmx.cc.o: CXXFLAGS += -mmmx
+$(BUILD_PFX)%_sse2.cc.d: CXXFLAGS += -msse2
+$(BUILD_PFX)%_sse2.cc.o: CXXFLAGS += -msse2
+$(BUILD_PFX)%_sse3.cc.d: CXXFLAGS += -msse3
+$(BUILD_PFX)%_sse3.cc.o: CXXFLAGS += -msse3
+$(BUILD_PFX)%_ssse3.cc.d: CXXFLAGS += -mssse3
+$(BUILD_PFX)%_ssse3.cc.o: CXXFLAGS += -mssse3
+$(BUILD_PFX)%_sse4.cc.d: CXXFLAGS += -msse4.1
+$(BUILD_PFX)%_sse4.cc.o: CXXFLAGS += -msse4.1
+$(BUILD_PFX)%_avx.cc.d: CXXFLAGS += -mavx
+$(BUILD_PFX)%_avx.cc.o: CXXFLAGS += -mavx
+$(BUILD_PFX)%_avx2.cc.d: CXXFLAGS += -mavx2
+$(BUILD_PFX)%_avx2.cc.o: CXXFLAGS += -mavx2
+
+$(BUILD_PFX)%.c.d: %.c
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(CC) $(INTERNAL_CFLAGS) $(CFLAGS) -M $< | $(fmt_deps) > $@
+
+$(BUILD_PFX)%.c.o: %.c
+	$(if $(quiet),@echo "    [CC] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(CC) $(INTERNAL_CFLAGS) $(CFLAGS) -c -o $@ $<
+
+$(BUILD_PFX)%.cc.d: %.cc
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -M $< | $(fmt_deps) > $@
+
+$(BUILD_PFX)%.cc.o: %.cc
+	$(if $(quiet),@echo "    [CXX] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -c -o $@ $<
+
+$(BUILD_PFX)%.cpp.d: %.cpp
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -M $< | $(fmt_deps) > $@
+
+$(BUILD_PFX)%.cpp.o: %.cpp
+	$(if $(quiet),@echo "    [CXX] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(CXX) $(INTERNAL_CFLAGS) $(CXXFLAGS) -c -o $@ $<
+
+$(BUILD_PFX)%.asm.d: %.asm
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(SRC_PATH_BARE)/build/make/gen_asm_deps.sh \
+            --build-pfx=$(BUILD_PFX) --depfile=$@ $(ASFLAGS) $< > $@
+
+$(BUILD_PFX)%.asm.o: %.asm
+	$(if $(quiet),@echo "    [AS] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(AS) $(ASFLAGS) -o $@ $<
+
+$(BUILD_PFX)%.s.d: %.s
+	$(if $(quiet),@echo "    [DEP] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(SRC_PATH_BARE)/build/make/gen_asm_deps.sh \
+            --build-pfx=$(BUILD_PFX) --depfile=$@ $(ASFLAGS) $< > $@
+
+$(BUILD_PFX)%.s.o: %.s
+	$(if $(quiet),@echo "    [AS] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(AS) $(ASFLAGS) -o $@ $<
+
+.PRECIOUS: %.c.S
+%.c.S: CFLAGS += -DINLINE_ASM
+$(BUILD_PFX)%.c.S: %.c
+	$(if $(quiet),@echo "    [GEN] $@")
+	$(qexec)$(if $(CONFIG_DEPENDENCY_TRACKING),,mkdir -p $(dir $@))
+	$(qexec)$(CC) -S $(CFLAGS) -o $@ $<
+
+.PRECIOUS: %.asm.s
+$(BUILD_PFX)%.asm.s: %.asm
+	$(if $(quiet),@echo "    [ASM CONVERSION] $@")
+	$(qexec)mkdir -p $(dir $@)
+	$(qexec)$(ASM_CONVERSION) <$< >$@
+
+# If we're in debug mode, pretend we don't have GNU strip, to fall back to
+# the copy implementation
+HAVE_GNU_STRIP := $(if $(CONFIG_DEBUG),,$(HAVE_GNU_STRIP))
+ifeq ($(HAVE_GNU_STRIP),yes)
+# Older binutils strip global symbols not needed for relocation processing
+# when given --strip-unneeded. Using nm and awk to identify globals and
+# keep them caused command line length issues under mingw and segfaults in
+# test_libaom were observed under OS/2: simply use --strip-debug.
+%.a: %_g.a
+	$(if $(quiet),@echo "    [STRIP] $@ < $<")
+	$(qexec)$(STRIP) --strip-debug \
+          -o $@ $<
+else
+%.a: %_g.a
+	$(if $(quiet),@echo "    [CP] $@ < $<")
+	$(qexec)cp $< $@
+endif
+
+#
+# Utility functions
+#
+pairmap=$(if $(strip $(2)),\
+    $(call $(1),$(word 1,$(2)),$(word 2,$(2)))\
+    $(call pairmap,$(1),$(wordlist 3,$(words $(2)),$(2)))\
+)
+
+enabled=$(filter-out $($(1)-no),$($(1)-yes))
+cond_enabled=$(if $(filter yes,$($(1))), $(call enabled,$(2)))
+
+find_file1=$(word 1,$(wildcard $(subst //,/,$(addsuffix /$(1),$(2)))))
+find_file=$(foreach f,$(1),$(call find_file1,$(strip $(f)),$(strip $(2))) )
+obj_pats=.c=.c.o $(AS_SFX)=$(AS_SFX).o .cc=.cc.o .cpp=.cpp.o
+objs=$(addprefix $(BUILD_PFX),$(foreach p,$(obj_pats),$(filter %.o,$(1:$(p))) ))
+
+install_map_templates=$(eval $(call install_map_template,$(1),$(2)))
+
+not=$(subst yes,no,$(1))
+
+ifeq ($(CONFIG_MSVS),yes)
+lib_file_name=$(1).lib
+else
+lib_file_name=lib$(1).a
+endif
+#
+# Rule Templates
+#
+define linker_template
+$(1): $(filter-out -%,$(2))
+$(1):
+	$(if $(quiet),@echo    "    [LD] $$@")
+	$(qexec)$$(LD) $$(strip $$(INTERNAL_LDFLAGS) $$(LDFLAGS) -o $$@ $(2) $(3) $$(extralibs))
+endef
+define linkerxx_template
+$(1): $(filter-out -%,$(2))
+$(1):
+	$(if $(quiet),@echo    "    [LD] $$@")
+	$(qexec)$$(CXX) $$(strip $$(INTERNAL_LDFLAGS) $$(LDFLAGS) -o $$@ $(2) $(3) $$(extralibs))
+endef
+# make-3.80 has a bug with expanding large input strings to the eval function,
+# which was triggered in some cases by the following component of
+# linker_template:
+#   $(1): $$(call find_file, $(patsubst -l%,lib%.a,$(filter -l%,$(2))),\
+#                           $$(patsubst -L%,%,$$(filter -L%,$$(LDFLAGS) $(2))))
+# This may be useful to revisit in the future (it tries to locate libraries
+# in a search path and add them as prerequisites
+
+define install_map_template
+$(DIST_DIR)/$(1): $(2)
+	$(if $(quiet),@echo "    [INSTALL] $$@")
+	$(qexec)mkdir -p $$(dir $$@)
+	$(qexec)cp -p $$< $$@
+endef
+
+define archive_template
+# Not using a pattern rule here because we don't want to generate empty
+# archives when they are listed as a dependency in files not responsible
+# for creating them.
+$(1):
+	$(if $(quiet),@echo "    [AR] $$@")
+	$(qexec)$$(AR) $$(ARFLAGS) $$@ $$^
+endef
+
+define so_template
+# Not using a pattern rule here because we don't want to generate empty
+# archives when they are listed as a dependency in files not responsible
+# for creating them.
+#
+# This needs further abstraction for dealing with non-GNU linkers.
+$(1):
+	$(if $(quiet),@echo "    [LD] $$@")
+	$(qexec)$$(LD) -shared $$(LDFLAGS) \
+            -Wl,--no-undefined -Wl,-soname,$$(SONAME) \
+            -Wl,--version-script,$$(EXPORTS_FILE) -o $$@ \
+            $$(filter %.o,$$^) $$(extralibs)
+endef
+
+define dl_template
+# Not using a pattern rule here because we don't want to generate empty
+# archives when they are listed as a dependency in files not responsible
+# for creating them.
+$(1):
+	$(if $(quiet),@echo "    [LD] $$@")
+	$(qexec)$$(LD) -dynamiclib $$(LDFLAGS) \
+	    -exported_symbols_list $$(EXPORTS_FILE) \
+        -Wl,-headerpad_max_install_names,-compatibility_version,1.0,-current_version,$$(VERSION_MAJOR) \
+        -o $$@ \
+        $$(filter %.o,$$^) $$(extralibs)
+endef
+
+define dll_template
+# Not using a pattern rule here because we don't want to generate empty
+# archives when they are listed as a dependency in files not responsible
+# for creating them.
+$(1):
+	$(if $(quiet),@echo "    [LD] $$@")
+	$(qexec)$$(LD) -Zdll $$(LDFLAGS) \
+        -o $$@ \
+        $$(filter %.o,$$^) $$(extralibs) $$(EXPORTS_FILE)
+endef
+
+
+#
+# Get current configuration
+#
+ifneq ($(target),)
+include $(SRC_PATH_BARE)/$(target:-$(TOOLCHAIN)=).mk
+endif
+
+skip_deps := $(filter %clean,$(MAKECMDGOALS))
+skip_deps += $(findstring testdata,$(MAKECMDGOALS))
+ifeq ($(strip $(skip_deps)),)
+  ifeq ($(CONFIG_DEPENDENCY_TRACKING),yes)
+    # Older versions of make don't like -include directives with no arguments
+    ifneq ($(filter %.d,$(OBJS-yes:.o=.d)),)
+      -include $(filter %.d,$(OBJS-yes:.o=.d))
+    endif
+  endif
+endif
+
+#
+# Configuration dependent rules
+#
+$(call pairmap,install_map_templates,$(INSTALL_MAPS))
+
+DOCS=$(call cond_enabled,CONFIG_INSTALL_DOCS,DOCS)
+.docs: $(DOCS)
+	@touch $@
+
+INSTALL-DOCS=$(call cond_enabled,CONFIG_INSTALL_DOCS,INSTALL-DOCS)
+ifeq ($(MAKECMDGOALS),dist)
+INSTALL-DOCS+=$(call cond_enabled,CONFIG_INSTALL_DOCS,DIST-DOCS)
+endif
+.install-docs: .docs $(addprefix $(DIST_DIR)/,$(INSTALL-DOCS))
+	@touch $@
+
+clean::
+	rm -f .docs .install-docs $(DOCS)
+
+BINS=$(call enabled,BINS)
+.bins: $(BINS)
+	@touch $@
+
+INSTALL-BINS=$(call cond_enabled,CONFIG_INSTALL_BINS,INSTALL-BINS)
+ifeq ($(MAKECMDGOALS),dist)
+INSTALL-BINS+=$(call cond_enabled,CONFIG_INSTALL_BINS,DIST-BINS)
+endif
+.install-bins: .bins $(addprefix $(DIST_DIR)/,$(INSTALL-BINS))
+	@touch $@
+
+clean::
+	rm -f .bins .install-bins $(BINS)
+
+LIBS=$(call enabled,LIBS)
+.libs: $(LIBS)
+	@touch $@
+$(foreach lib,$(filter %_g.a,$(LIBS)),$(eval $(call archive_template,$(lib))))
+$(foreach lib,$(filter %so.$(SO_VERSION_MAJOR).$(SO_VERSION_MINOR).$(SO_VERSION_PATCH),$(LIBS)),$(eval $(call so_template,$(lib))))
+$(foreach lib,$(filter %$(SO_VERSION_MAJOR).dylib,$(LIBS)),$(eval $(call dl_template,$(lib))))
+$(foreach lib,$(filter %$(SO_VERSION_MAJOR).dll,$(LIBS)),$(eval $(call dll_template,$(lib))))
+
+INSTALL-LIBS=$(call cond_enabled,CONFIG_INSTALL_LIBS,INSTALL-LIBS)
+ifeq ($(MAKECMDGOALS),dist)
+INSTALL-LIBS+=$(call cond_enabled,CONFIG_INSTALL_LIBS,DIST-LIBS)
+endif
+.install-libs: .libs $(addprefix $(DIST_DIR)/,$(INSTALL-LIBS))
+	@touch $@
+
+clean::
+	rm -f .libs .install-libs $(LIBS)
+
+ifeq ($(CONFIG_EXTERNAL_BUILD),yes)
+PROJECTS=$(call enabled,PROJECTS)
+.projects: $(PROJECTS)
+	@touch $@
+
+INSTALL-PROJECTS=$(call cond_enabled,CONFIG_INSTALL_PROJECTS,INSTALL-PROJECTS)
+ifeq ($(MAKECMDGOALS),dist)
+INSTALL-PROJECTS+=$(call cond_enabled,CONFIG_INSTALL_PROJECTS,DIST-PROJECTS)
+endif
+.install-projects: .projects $(addprefix $(DIST_DIR)/,$(INSTALL-PROJECTS))
+	@touch $@
+
+clean::
+	rm -f .projects .install-projects $(PROJECTS)
+endif
+
+# If there are any source files to be distributed, then include the build
+# system too.
+ifneq ($(call enabled,DIST-SRCS),)
+    DIST-SRCS-yes            += configure
+    DIST-SRCS-yes            += build/make/configure.sh
+    DIST-SRCS-yes            += build/make/gen_asm_deps.sh
+    DIST-SRCS-yes            += build/make/Makefile
+    DIST-SRCS-$(CONFIG_MSVS)  += build/make/gen_msvs_def.sh
+    DIST-SRCS-$(CONFIG_MSVS)  += build/make/gen_msvs_sln.sh
+    DIST-SRCS-$(CONFIG_MSVS)  += build/make/gen_msvs_vcxproj.sh
+    DIST-SRCS-$(CONFIG_MSVS)  += build/make/msvs_common.sh
+    DIST-SRCS-$(CONFIG_RVCT) += build/make/armlink_adapter.sh
+    DIST-SRCS-$(ARCH_ARM)    += build/make/ads2gas.pl
+    DIST-SRCS-$(ARCH_ARM)    += build/make/ads2gas_apple.pl
+    DIST-SRCS-$(ARCH_ARM)    += build/make/ads2armasm_ms.pl
+    DIST-SRCS-$(ARCH_ARM)    += build/make/thumb.pm
+    DIST-SRCS-yes            += $(target:-$(TOOLCHAIN)=).mk
+endif
+INSTALL-SRCS := $(call cond_enabled,CONFIG_INSTALL_SRCS,INSTALL-SRCS)
+ifeq ($(MAKECMDGOALS),dist)
+INSTALL-SRCS += $(call cond_enabled,CONFIG_INSTALL_SRCS,DIST-SRCS)
+endif
+.install-srcs: $(addprefix $(DIST_DIR)/src/,$(INSTALL-SRCS))
+	@touch $@
+
+clean::
+	rm -f .install-srcs
+
+ifeq ($(CONFIG_EXTERNAL_BUILD),yes)
+    BUILD_TARGETS += .projects
+    INSTALL_TARGETS += .install-projects
+endif
+BUILD_TARGETS += .docs .libs .bins
+INSTALL_TARGETS += .install-docs .install-srcs .install-libs .install-bins
+all: $(BUILD_TARGETS)
+install:: $(INSTALL_TARGETS)
+dist: $(INSTALL_TARGETS)
+test::
+
+.SUFFIXES:  # Delete default suffix rules
diff --git a/third_party/aom/build/make/ads2armasm_ms.pl b/third_party/aom/build/make/ads2armasm_ms.pl
new file mode 100755
index 0000000000..8568a2dadf
--- /dev/null
+++ b/third_party/aom/build/make/ads2armasm_ms.pl
@@ -0,0 +1,39 @@
+#!/usr/bin/env perl
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+use FindBin;
+use lib $FindBin::Bin;
+use thumb;
+
+print "; This file was created from a .asm file\n";
+print ";  using the ads2armasm_ms.pl script.\n";
+
+while (<STDIN>)
+{
+    undef $comment;
+    undef $line;
+
+    s/REQUIRE8//;
+    s/PRESERVE8//;
+    s/^\s*ARM\s*$//;
+    s/AREA\s+\|\|(.*)\|\|/AREA |$1|/;
+    s/qsubaddx/qsax/i;
+    s/qaddsubx/qasx/i;
+
+    thumb::FixThumbInstructions($_, 1);
+
+    s/ldrneb/ldrbne/i;
+    s/ldrneh/ldrhne/i;
+    s/^(\s*)ENDP.*/$&\n$1ALIGN 4/;
+
+    print;
+}
+
diff --git a/third_party/aom/build/make/ads2gas.pl b/third_party/aom/build/make/ads2gas.pl
new file mode 100755
index 0000000000..adf45a3c91
--- /dev/null
+++ b/third_party/aom/build/make/ads2gas.pl
@@ -0,0 +1,236 @@
+#!/usr/bin/env perl
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+# ads2gas.pl
+# Author: Eric Fung (efung (at) acm.org)
+#
+# Convert ARM Developer Suite 1.0.1 syntax assembly source to GNU as format
+#
+# Usage: cat inputfile | perl ads2gas.pl > outputfile
+#
+
+use FindBin;
+use lib $FindBin::Bin;
+use thumb;
+
+my $thumb = 0;
+
+foreach my $arg (@ARGV) {
+    $thumb = 1 if ($arg eq "-thumb");
+}
+
+print "@ This file was created from a .asm file\n";
+print "@  using the ads2gas.pl script.\n";
+print "\t.equ DO1STROUNDING, 0\n";
+if ($thumb) {
+    print "\t.syntax unified\n";
+    print "\t.thumb\n";
+}
+
+# Stack of procedure names.
+@proc_stack = ();
+
+while (<STDIN>)
+{
+    undef $comment;
+    undef $line;
+    $comment_char = ";";
+    $comment_sub = "@";
+
+    # Handle comments.
+    if (/$comment_char/)
+    {
+      $comment = "";
+      ($line, $comment) = /(.*?)$comment_char(.*)/;
+      $_ = $line;
+    }
+
+    # Load and store alignment
+    s/@/,:/g;
+
+    # Hexadecimal constants prefaced by 0x
+    s/#&/#0x/g;
+
+    # Convert :OR: to |
+    s/:OR:/ | /g;
+
+    # Convert :AND: to &
+    s/:AND:/ & /g;
+
+    # Convert :NOT: to ~
+    s/:NOT:/ ~ /g;
+
+    # Convert :SHL: to <<
+    s/:SHL:/ << /g;
+
+    # Convert :SHR: to >>
+    s/:SHR:/ >> /g;
+
+    # Convert ELSE to .else
+    s/\bELSE\b/.else/g;
+
+    # Convert ENDIF to .endif
+    s/\bENDIF\b/.endif/g;
+
+    # Convert ELSEIF to .elseif
+    s/\bELSEIF\b/.elseif/g;
+
+    # Convert LTORG to .ltorg
+    s/\bLTORG\b/.ltorg/g;
+
+    # Convert endfunc to nothing.
+    s/\bendfunc\b//ig;
+
+    # Convert FUNCTION to nothing.
+    s/\bFUNCTION\b//g;
+    s/\bfunction\b//g;
+
+    s/\bENTRY\b//g;
+    s/\bMSARMASM\b/0/g;
+    s/^\s+end\s+$//g;
+
+    # Convert IF :DEF:to .if
+    # gcc doesn't have the ability to do a conditional
+    # if defined variable that is set by IF :DEF: on
+    # armasm, so convert it to a normal .if and then
+    # make sure to define a value elesewhere
+    if (s/\bIF :DEF:\b/.if /g)
+    {
+        s/=/==/g;
+    }
+
+    # Convert IF to .if
+    if (s/\bIF\b/.if/g)
+    {
+        s/=+/==/g;
+    }
+
+    # Convert INCLUDE to .INCLUDE "file"
+    s/INCLUDE(\s*)(.*)$/.include $1\"$2\"/;
+
+    # Code directive (ARM vs Thumb)
+    s/CODE([0-9][0-9])/.code $1/;
+
+    # No AREA required
+    # But ALIGNs in AREA must be obeyed
+    s/^\s*AREA.*ALIGN=([0-9])$/.text\n.p2align $1/;
+    # If no ALIGN, strip the AREA and align to 4 bytes
+    s/^\s*AREA.*$/.text\n.p2align 2/;
+
+    # DCD to .word
+    # This one is for incoming symbols
+    s/DCD\s+\|(\w*)\|/.long $1/;
+
+    # DCW to .short
+    s/DCW\s+\|(\w*)\|/.short $1/;
+    s/DCW(.*)/.short $1/;
+
+    # Constants defined in scope
+    s/DCD(.*)/.long $1/;
+    s/DCB(.*)/.byte $1/;
+
+    # RN to .req
+    if (s/RN\s+([Rr]\d+|lr)/.req $1/)
+    {
+        print;
+        print "$comment_sub$comment\n" if defined $comment;
+        next;
+    }
+
+    # Make function visible to linker, and make additional symbol with
+    # prepended underscore
+    s/EXPORT\s+\|([\$\w]*)\|/.global $1 \n\t.type $1, function/;
+    s/IMPORT\s+\|([\$\w]*)\|/.global $1/;
+
+    s/EXPORT\s+([\$\w]*)/.global $1/;
+    s/export\s+([\$\w]*)/.global $1/;
+
+    # No vertical bars required; make additional symbol with prepended
+    # underscore
+    s/^\|(\$?\w+)\|/_$1\n\t$1:/g;
+
+    # Labels need trailing colon
+#   s/^(\w+)/$1:/ if !/EQU/;
+    # put the colon at the end of the line in the macro
+    s/^([a-zA-Z_0-9\$]+)/$1:/ if !/EQU/;
+
+    # ALIGN directive
+    s/\bALIGN\b/.balign/g;
+
+    if ($thumb) {
+        # ARM code - we force everything to thumb with the declaration in the header
+        s/\sARM//g;
+    } else {
+        # ARM code
+        s/\sARM/.arm/g;
+    }
+
+    # push/pop
+    s/(push\s+)(r\d+)/stmdb sp\!, \{$2\}/g;
+    s/(pop\s+)(r\d+)/ldmia sp\!, \{$2\}/g;
+
+    # NEON code
+    s/(vld1.\d+\s+)(q\d+)/$1\{$2\}/g;
+    s/(vtbl.\d+\s+[^,]+),([^,]+)/$1,\{$2\}/g;
+
+    if ($thumb) {
+        thumb::FixThumbInstructions($_, 0);
+    }
+
+    # eabi_attributes numerical equivalents can be found in the
+    # "ARM IHI 0045C" document.
+
+    # REQUIRE8 Stack is required to be 8-byte aligned
+    s/\sREQUIRE8/.eabi_attribute 24, 1 \@Tag_ABI_align_needed/g;
+
+    # PRESERVE8 Stack 8-byte align is preserved
+    s/\sPRESERVE8/.eabi_attribute 25, 1 \@Tag_ABI_align_preserved/g;
+
+    # Use PROC and ENDP to give the symbols a .size directive.
+    # This makes them show up properly in debugging tools like gdb and valgrind.
+    if (/\bPROC\b/)
+    {
+        my $proc;
+        /^_([\.0-9A-Z_a-z]\w+)\b/;
+        $proc = $1;
+        push(@proc_stack, $proc) if ($proc);
+        s/\bPROC\b/@ $&/;
+    }
+    if (/\bENDP\b/)
+    {
+        my $proc;
+        s/\bENDP\b/@ $&/;
+        $proc = pop(@proc_stack);
+        $_ = "\t.size $proc, .-$proc".$_ if ($proc);
+    }
+
+    # EQU directive
+    s/(\S+\s+)EQU(\s+\S+)/.equ $1, $2/;
+
+    # Begin macro definition
+    if (/\bMACRO\b/) {
+        $_ = <STDIN>;
+        s/^/.macro/;
+        s/\$//g;                # remove formal param reference
+        s/;/@/g;                # change comment characters
+    }
+
+    # For macros, use \ to reference formal params
+    s/\$/\\/g;                  # End macro definition
+    s/\bMEND\b/.endm/;              # No need to tell it where to stop assembling
+    next if /^\s*END\s*$/;
+    print;
+    print "$comment_sub$comment\n" if defined $comment;
+}
+
+# Mark that this object doesn't need an executable stack.
+printf ("\t.section\t.note.GNU-stack,\"\",\%\%progbits\n");
diff --git a/third_party/aom/build/make/ads2gas_apple.pl b/third_party/aom/build/make/ads2gas_apple.pl
new file mode 100755
index 0000000000..31ec91d567
--- /dev/null
+++ b/third_party/aom/build/make/ads2gas_apple.pl
@@ -0,0 +1,235 @@
+#!/usr/bin/env perl
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+# ads2gas_apple.pl
+# Author: Eric Fung (efung (at) acm.org)
+#
+# Convert ARM Developer Suite 1.0.1 syntax assembly source to GNU as format
+#
+# Usage: cat inputfile | perl ads2gas_apple.pl > outputfile
+#
+
+my $chromium = 0;
+
+foreach my $arg (@ARGV) {
+    $chromium = 1 if ($arg eq "-chromium");
+}
+
+print "@ This file was created from a .asm file\n";
+print "@  using the ads2gas_apple.pl script.\n\n";
+print "\t.set WIDE_REFERENCE, 0\n";
+print "\t.set ARCHITECTURE, 5\n";
+print "\t.set DO1STROUNDING, 0\n";
+
+my %register_aliases;
+my %macro_aliases;
+
+my @mapping_list = ("\$0", "\$1", "\$2", "\$3", "\$4", "\$5", "\$6", "\$7", "\$8", "\$9");
+
+my @incoming_array;
+
+my @imported_functions;
+
+# Perl trim function to remove whitespace from the start and end of the string
+sub trim($)
+{
+    my $string = shift;
+    $string =~ s/^\s+//;
+    $string =~ s/\s+$//;
+    return $string;
+}
+
+while (<STDIN>)
+{
+    # Load and store alignment
+    s/@/,:/g;
+
+    # Comment character
+    s/;/ @/g;
+
+    # Hexadecimal constants prefaced by 0x
+    s/#&/#0x/g;
+
+    # Convert :OR: to |
+    s/:OR:/ | /g;
+
+    # Convert :AND: to &
+    s/:AND:/ & /g;
+
+    # Convert :NOT: to ~
+    s/:NOT:/ ~ /g;
+
+    # Convert :SHL: to <<
+    s/:SHL:/ << /g;
+
+    # Convert :SHR: to >>
+    s/:SHR:/ >> /g;
+
+    # Convert ELSE to .else
+    s/\bELSE\b/.else/g;
+
+    # Convert ENDIF to .endif
+    s/\bENDIF\b/.endif/g;
+
+    # Convert ELSEIF to .elseif
+    s/\bELSEIF\b/.elseif/g;
+
+    # Convert LTORG to .ltorg
+    s/\bLTORG\b/.ltorg/g;
+
+    # Convert IF :DEF:to .if
+    # gcc doesn't have the ability to do a conditional
+    # if defined variable that is set by IF :DEF: on
+    # armasm, so convert it to a normal .if and then
+    # make sure to define a value elesewhere
+    if (s/\bIF :DEF:\b/.if /g)
+    {
+        s/=/==/g;
+    }
+
+    # Convert IF to .if
+    if (s/\bIF\b/.if/g)
+    {
+        s/=/==/g;
+    }
+
+    # Convert INCLUDE to .INCLUDE "file"
+    s/INCLUDE(\s*)(.*)$/.include $1\"$2\"/;
+
+    # Code directive (ARM vs Thumb)
+    s/CODE([0-9][0-9])/.code $1/;
+
+    # No AREA required
+    # But ALIGNs in AREA must be obeyed
+    s/^\s*AREA.*ALIGN=([0-9])$/.text\n.p2align $1/;
+    # If no ALIGN, strip the AREA and align to 4 bytes
+    s/^\s*AREA.*$/.text\n.p2align 2/;
+
+    # DCD to .word
+    # This one is for incoming symbols
+    s/DCD\s+\|(\w*)\|/.long $1/;
+
+    # DCW to .short
+    s/DCW\s+\|(\w*)\|/.short $1/;
+    s/DCW(.*)/.short $1/;
+
+    # Constants defined in scope
+    s/DCD(.*)/.long $1/;
+    s/DCB(.*)/.byte $1/;
+
+    # Build a hash of all the register - alias pairs.
+    if (s/(.*)RN(.*)/$1 .req $2/g)
+    {
+        $register_aliases{trim($1)} = trim($2);
+        next;
+    }
+
+    while (($key, $value) = each(%register_aliases))
+    {
+        s/\b$key\b/$value/g;
+    }
+
+    # Make function visible to linker, and make additional symbol with
+    # prepended underscore
+    s/EXPORT\s+\|([\$\w]*)\|/.globl _$1\n\t.globl $1/;
+
+    # Prepend imported functions with _
+    if (s/IMPORT\s+\|([\$\w]*)\|/.globl $1/)
+    {
+        $function = trim($1);
+        push(@imported_functions, $function);
+    }
+
+    foreach $function (@imported_functions)
+    {
+        s/$function/_$function/;
+    }
+
+    # No vertical bars required; make additional symbol with prepended
+    # underscore
+    s/^\|(\$?\w+)\|/_$1\n\t$1:/g;
+
+    # Labels need trailing colon
+#   s/^(\w+)/$1:/ if !/EQU/;
+    # put the colon at the end of the line in the macro
+    s/^([a-zA-Z_0-9\$]+)/$1:/ if !/EQU/;
+
+    # ALIGN directive
+    s/\bALIGN\b/.balign/g;
+
+    # Strip ARM
+    s/\sARM/@ ARM/g;
+
+    # Strip REQUIRE8
+    #s/\sREQUIRE8/@ REQUIRE8/g;
+    s/\sREQUIRE8/@ /g;
+
+    # Strip PRESERVE8
+    s/\sPRESERVE8/@ PRESERVE8/g;
+
+    # Strip PROC and ENDPROC
+    s/\bPROC\b/@/g;
+    s/\bENDP\b/@/g;
+
+    # EQU directive
+    s/(.*)EQU(.*)/.set $1, $2/;
+
+    # Begin macro definition
+    if (/\bMACRO\b/)
+    {
+        # Process next line down, which will be the macro definition
+        $_ = <STDIN>;
+
+        $trimmed = trim($_);
+
+        # remove commas that are separating list
+        $trimmed =~ s/,//g;
+
+        # string to array
+        @incoming_array = split(/\s+/, $trimmed);
+
+        print ".macro @incoming_array[0]\n";
+
+        # remove the first element, as that is the name of the macro
+        shift (@incoming_array);
+
+        @macro_aliases{@incoming_array} = @mapping_list;
+
+        next;
+    }
+
+    while (($key, $value) = each(%macro_aliases))
+    {
+        $key =~ s/\$/\\\$/;
+        s/$key\b/$value/g;
+    }
+
+    # For macros, use \ to reference formal params
+#   s/\$/\\/g;                  # End macro definition
+    s/\bMEND\b/.endm/;              # No need to tell it where to stop assembling
+    next if /^\s*END\s*$/;
+
+    # Clang used by Chromium differs slightly from clang in XCode in what it
+    # will accept in the assembly.
+    if ($chromium) {
+        s/qsubaddx/qsax/i;
+        s/qaddsubx/qasx/i;
+        s/ldrneb/ldrbne/i;
+        s/ldrneh/ldrhne/i;
+        s/(vqshrun\.s16 .*, \#)0$/${1}8/i;
+
+        # http://llvm.org/bugs/show_bug.cgi?id=16022
+        s/\.include/#include/;
+    }
+
+    print;
+}
diff --git a/third_party/aom/build/make/armlink_adapter.sh b/third_party/aom/build/make/armlink_adapter.sh
new file mode 100755
index 0000000000..85c6c96c13
--- /dev/null
+++ b/third_party/aom/build/make/armlink_adapter.sh
@@ -0,0 +1,53 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+verbose=0
+set -- $*
+for i; do
+    if [ "$i" = "-o" ]; then
+        on_of=1
+    elif [ "$i" = "-v" ]; then
+        verbose=1
+    elif [ "$i" = "-g" ]; then
+        args="${args} --debug"
+    elif [ "$on_of" = "1" ]; then
+        outfile=$i
+        on_of=0
+    elif [ -f "$i" ]; then
+        infiles="$infiles $i"
+    elif [ "${i#-l}" != "$i" ]; then
+        libs="$libs ${i#-l}"
+    elif [ "${i#-L}" != "$i" ]; then
+        libpaths="${libpaths} ${i#-L}"
+    else
+        args="${args} ${i}"
+    fi
+    shift
+done
+
+# Absolutize library file names
+for f in $libs; do
+    found=0
+    for d in $libpaths; do
+        [ -f "$d/$f" ] && infiles="$infiles $d/$f" && found=1 && break
+        [ -f "$d/lib${f}.so" ] && infiles="$infiles $d/lib${f}.so" && found=1 && break
+        [ -f "$d/lib${f}.a" ] && infiles="$infiles $d/lib${f}.a" && found=1 && break
+    done
+    [ $found -eq 0 ] && infiles="$infiles $f"
+done
+for d in $libpaths; do
+    [ -n "$libsearchpath" ] && libsearchpath="${libsearchpath},"
+    libsearchpath="${libsearchpath}$d"
+done
+
+cmd="armlink $args --userlibpath=$libsearchpath --output=$outfile $infiles"
+[ $verbose -eq 1 ] && echo $cmd
+$cmd
diff --git a/third_party/aom/build/make/configure.sh b/third_party/aom/build/make/configure.sh
new file mode 100644
index 0000000000..4ece17aee4
--- /dev/null
+++ b/third_party/aom/build/make/configure.sh
@@ -0,0 +1,1556 @@
+#!/bin/sh
+##
+##  configure.sh
+##
+##  This script is sourced by the main configure script and contains
+##  utility functions and other common bits that aren't strictly libaom
+##  related.
+##
+##  This build system is based in part on the FFmpeg configure script.
+##
+
+
+#
+# Logging / Output Functions
+#
+die_unknown(){
+  echo "Unknown option \"$1\"."
+  echo "See $0 --help for available options."
+  clean_temp_files
+  exit 1
+}
+
+die() {
+  echo "$@"
+  echo
+  echo "Configuration failed. This could reflect a misconfiguration of your"
+  echo "toolchains, improper options selected, or another problem. If you"
+  echo "don't see any useful error messages above, the next step is to look"
+  echo "at the configure error log file ($logfile) to determine what"
+  echo "configure was trying to do when it died."
+  clean_temp_files
+  exit 1
+}
+
+log(){
+  echo "$@" >>$logfile
+}
+
+log_file(){
+  log BEGIN $1
+  cat -n $1 >>$logfile
+  log END $1
+}
+
+log_echo() {
+  echo "$@"
+  log "$@"
+}
+
+fwrite () {
+  outfile=$1
+  shift
+  echo "$@" >> ${outfile}
+}
+
+show_help_pre(){
+  for opt in ${CMDLINE_SELECT}; do
+    opt2=`echo $opt | sed -e 's;_;-;g'`
+    if enabled $opt; then
+      eval "toggle_${opt}=\"--disable-${opt2}\""
+    else
+      eval "toggle_${opt}=\"--enable-${opt2} \""
+    fi
+  done
+
+  cat <<EOF
+Usage: configure [options]
+Options:
+
+Build options:
+  --help                      print this message
+  --log=yes|no|FILE           file configure log is written to [config.log]
+  --target=TARGET             target platform tuple [generic-gnu]
+  --cpu=CPU                   optimize for a specific cpu rather than a family
+  --extra-cflags=ECFLAGS      add ECFLAGS to CFLAGS [$CFLAGS]
+  --extra-cxxflags=ECXXFLAGS  add ECXXFLAGS to CXXFLAGS [$CXXFLAGS]
+  ${toggle_extra_warnings}    emit harmless warnings (always non-fatal)
+  ${toggle_werror}            treat warnings as errors, if possible
+                              (not available with all compilers)
+  ${toggle_optimizations}     turn on/off compiler optimization flags
+  ${toggle_pic}               turn on/off Position Independent Code
+  ${toggle_ccache}            turn on/off compiler cache
+  ${toggle_debug}             enable/disable debug mode
+  ${toggle_gprof}             enable/disable gprof profiling instrumentation
+  ${toggle_gcov}              enable/disable gcov coverage instrumentation
+  ${toggle_thumb}             enable/disable building arm assembly in thumb mode
+  ${toggle_dependency_tracking}
+                              disable to speed up one-time build
+
+Install options:
+  ${toggle_install_docs}      control whether docs are installed
+  ${toggle_install_bins}      control whether binaries are installed
+  ${toggle_install_libs}      control whether libraries are installed
+  ${toggle_install_srcs}      control whether sources are installed
+
+
+EOF
+}
+
+show_help_post(){
+  cat <<EOF
+
+
+NOTES:
+    Object files are built at the place where configure is launched.
+
+    All boolean options can be negated. The default value is the opposite
+    of that shown above. If the option --disable-foo is listed, then
+    the default value for foo is enabled.
+
+Supported targets:
+EOF
+  show_targets ${all_platforms}
+  echo
+  exit 1
+}
+
+show_targets() {
+  while [ -n "$*" ]; do
+    if [ "${1%%-*}" = "${2%%-*}" ]; then
+      if [ "${2%%-*}" = "${3%%-*}" ]; then
+        printf "    %-24s %-24s %-24s\n" "$1" "$2" "$3"
+        shift; shift; shift
+      else
+        printf "    %-24s %-24s\n" "$1" "$2"
+        shift; shift
+      fi
+    else
+      printf "    %-24s\n" "$1"
+      shift
+    fi
+  done
+}
+
+show_help() {
+  show_help_pre
+  show_help_post
+}
+
+#
+# List Processing Functions
+#
+set_all(){
+  value=$1
+  shift
+  for var in $*; do
+    eval $var=$value
+  done
+}
+
+is_in(){
+  value=$1
+  shift
+  for var in $*; do
+    [ $var = $value ] && return 0
+  done
+  return 1
+}
+
+add_cflags() {
+  CFLAGS="${CFLAGS} $@"
+  CXXFLAGS="${CXXFLAGS} $@"
+}
+
+add_cflags_only() {
+  CFLAGS="${CFLAGS} $@"
+}
+
+add_cxxflags_only() {
+  CXXFLAGS="${CXXFLAGS} $@"
+}
+
+add_ldflags() {
+  LDFLAGS="${LDFLAGS} $@"
+}
+
+add_asflags() {
+  ASFLAGS="${ASFLAGS} $@"
+}
+
+add_extralibs() {
+  extralibs="${extralibs} $@"
+}
+
+#
+# Boolean Manipulation Functions
+#
+
+enable_feature(){
+  set_all yes $*
+}
+
+disable_feature(){
+  set_all no $*
+}
+
+enabled(){
+  eval test "x\$$1" = "xyes"
+}
+
+disabled(){
+  eval test "x\$$1" = "xno"
+}
+
+enable_codec(){
+  enabled "${1}" || echo "  enabling ${1}"
+  enable_feature "${1}"
+
+  is_in "${1}" av1 && enable_feature "${1}_encoder" "${1}_decoder"
+}
+
+disable_codec(){
+  disabled "${1}" || echo "  disabling ${1}"
+  disable_feature "${1}"
+
+  is_in "${1}" av1 && disable_feature "${1}_encoder" "${1}_decoder"
+}
+
+# Iterates through positional parameters, checks to confirm the parameter has
+# not been explicitly (force) disabled, and enables the setting controlled by
+# the parameter when the setting is not disabled.
+# Note: Does NOT alter RTCD generation options ($RTCD_OPTIONS).
+soft_enable() {
+  for var in $*; do
+    if ! disabled $var; then
+      enabled $var || log_echo "  enabling $var"
+      enable_feature $var
+    fi
+  done
+}
+
+# Iterates through positional parameters, checks to confirm the parameter has
+# not been explicitly (force) enabled, and disables the setting controlled by
+# the parameter when the setting is not enabled.
+# Note: Does NOT alter RTCD generation options ($RTCD_OPTIONS).
+soft_disable() {
+  for var in $*; do
+    if ! enabled $var; then
+      disabled $var || log_echo "  disabling $var"
+      disable_feature $var
+    fi
+  done
+}
+
+#
+# Text Processing Functions
+#
+toupper(){
+  echo "$@" | tr abcdefghijklmnopqrstuvwxyz ABCDEFGHIJKLMNOPQRSTUVWXYZ
+}
+
+tolower(){
+  echo "$@" | tr ABCDEFGHIJKLMNOPQRSTUVWXYZ abcdefghijklmnopqrstuvwxyz
+}
+
+#
+# Temporary File Functions
+#
+source_path=${0%/*}
+enable_feature source_path_used
+if [ -z "$source_path" ] || [ "$source_path" = "." ]; then
+  source_path="`pwd`"
+  disable_feature source_path_used
+fi
+
+if test ! -z "$TMPDIR" ; then
+  TMPDIRx="${TMPDIR}"
+elif test ! -z "$TEMPDIR" ; then
+  TMPDIRx="${TEMPDIR}"
+else
+  TMPDIRx="/tmp"
+fi
+RAND=$(awk 'BEGIN { srand(); printf "%d\n",(rand() * 32768)}')
+TMP_H="${TMPDIRx}/aom-conf-$$-${RAND}.h"
+TMP_C="${TMPDIRx}/aom-conf-$$-${RAND}.c"
+TMP_CC="${TMPDIRx}/aom-conf-$$-${RAND}.cc"
+TMP_O="${TMPDIRx}/aom-conf-$$-${RAND}.o"
+TMP_X="${TMPDIRx}/aom-conf-$$-${RAND}.x"
+TMP_ASM="${TMPDIRx}/aom-conf-$$-${RAND}.asm"
+
+clean_temp_files() {
+  rm -f ${TMP_C} ${TMP_CC} ${TMP_H} ${TMP_O} ${TMP_X} ${TMP_ASM}
+  enabled gcov && rm -f ${TMP_C%.c}.gcno ${TMP_CC%.cc}.gcno
+}
+
+#
+# Toolchain Check Functions
+#
+check_cmd() {
+  enabled external_build && return
+  log "$@"
+  "$@" >>${logfile} 2>&1
+}
+
+check_cc() {
+  log check_cc "$@"
+  cat >${TMP_C}
+  log_file ${TMP_C}
+  check_cmd ${CC} ${CFLAGS} "$@" -c -o ${TMP_O} ${TMP_C}
+}
+
+check_cxx() {
+  log check_cxx "$@"
+  cat >${TMP_CC}
+  log_file ${TMP_CC}
+  check_cmd ${CXX} ${CXXFLAGS} "$@" -c -o ${TMP_O} ${TMP_CC}
+}
+
+check_cpp() {
+  log check_cpp "$@"
+  cat > ${TMP_C}
+  log_file ${TMP_C}
+  check_cmd ${CC} ${CFLAGS} "$@" -E -o ${TMP_O} ${TMP_C}
+}
+
+check_ld() {
+  log check_ld "$@"
+  check_cc $@ \
+    && check_cmd ${LD} ${LDFLAGS} "$@" -o ${TMP_X} ${TMP_O} ${extralibs}
+}
+
+check_header(){
+  log check_header "$@"
+  header=$1
+  shift
+  var=`echo $header | sed 's/[^A-Za-z0-9_]/_/g'`
+  disable_feature $var
+  check_cpp "$@" <<EOF && enable_feature $var
+#include "$header"
+int x;
+EOF
+}
+
+check_cflags() {
+ log check_cflags "$@"
+ check_cc -Werror "$@" <<EOF
+int x;
+EOF
+}
+
+check_cxxflags() {
+  log check_cxxflags "$@"
+
+  # Catch CFLAGS that trigger CXX warnings
+  case "$CXX" in
+    *c++-analyzer|*clang++|*g++*)
+      check_cxx -Werror "$@" <<EOF
+int x;
+EOF
+      ;;
+    *)
+      check_cxx -Werror "$@" <<EOF
+int x;
+EOF
+      ;;
+    esac
+}
+
+check_add_cflags() {
+  check_cxxflags "$@" && add_cxxflags_only "$@"
+  check_cflags "$@" && add_cflags_only "$@"
+}
+
+check_add_cxxflags() {
+  check_cxxflags "$@" && add_cxxflags_only "$@"
+}
+
+check_add_asflags() {
+  log add_asflags "$@"
+  add_asflags "$@"
+}
+
+check_add_ldflags() {
+  log add_ldflags "$@"
+  add_ldflags "$@"
+}
+
+check_asm_align() {
+  log check_asm_align "$@"
+  cat >${TMP_ASM} <<EOF
+section .rodata
+align 16
+EOF
+  log_file ${TMP_ASM}
+  check_cmd ${AS} ${ASFLAGS} -o ${TMP_O} ${TMP_ASM}
+  readelf -WS ${TMP_O} >${TMP_X}
+  log_file ${TMP_X}
+  if ! grep -q '\.rodata .* 16$' ${TMP_X}; then
+    die "${AS} ${ASFLAGS} does not support section alignment (nasm <=2.08?)"
+  fi
+}
+
+# tests for -m$1 toggling the feature given in $2. If $2 is empty $1 is used.
+check_gcc_machine_option() {
+  opt="$1"
+  feature="$2"
+  [ -n "$feature" ] || feature="$opt"
+
+  if enabled gcc && ! disabled "$feature" && ! check_cflags "-m$opt"; then
+    RTCD_OPTIONS="${RTCD_OPTIONS}--disable-$feature "
+  else
+    soft_enable "$feature"
+  fi
+}
+
+write_common_config_banner() {
+  print_webm_license config.mk "##" ""
+  echo '# This file automatically generated by configure. Do not edit!' >> config.mk
+  echo "TOOLCHAIN := ${toolchain}" >> config.mk
+
+  case ${toolchain} in
+    *-linux-rvct)
+      echo "ALT_LIBC := ${alt_libc}" >> config.mk
+      ;;
+  esac
+}
+
+write_common_config_targets() {
+  for t in ${all_targets}; do
+    if enabled ${t}; then
+      if enabled child; then
+        fwrite config.mk "ALL_TARGETS += ${t}-${toolchain}"
+      else
+        fwrite config.mk "ALL_TARGETS += ${t}"
+      fi
+    fi
+    true;
+  done
+  true
+}
+
+write_common_target_config_mk() {
+  saved_CC="${CC}"
+  saved_CXX="${CXX}"
+  enabled ccache && CC="ccache ${CC}"
+  enabled ccache && CXX="ccache ${CXX}"
+  print_webm_license $1 "##" ""
+
+  cat >> $1 << EOF
+# This file automatically generated by configure. Do not edit!
+SRC_PATH="$source_path"
+SRC_PATH_BARE=$source_path
+BUILD_PFX=${BUILD_PFX}
+TOOLCHAIN=${toolchain}
+ASM_CONVERSION=${asm_conversion_cmd:-${source_path}/build/make/ads2gas.pl}
+GEN_VCPROJ=${gen_vcproj_cmd}
+MSVS_ARCH_DIR=${msvs_arch_dir}
+
+CC=${CC}
+CXX=${CXX}
+AR=${AR}
+LD=${LD}
+AS=${AS}
+STRIP=${STRIP}
+NM=${NM}
+
+CFLAGS  = ${CFLAGS}
+CXXFLAGS  = ${CXXFLAGS}
+ARFLAGS = crs\$(if \$(quiet),,v)
+LDFLAGS = ${LDFLAGS}
+ASFLAGS = ${ASFLAGS}
+extralibs = ${extralibs}
+AS_SFX    = ${AS_SFX:-.asm}
+EXE_SFX   = ${EXE_SFX}
+VCPROJ_SFX = ${VCPROJ_SFX}
+RTCD_OPTIONS = ${RTCD_OPTIONS}
+WX_CXXFLAGS = ${WX_CXXFLAGS}
+WX_LDFLAGS = ${WX_LDFLAGS}
+EOF
+
+  if enabled rvct; then cat >> $1 << EOF
+fmt_deps = sed -e 's;^__image.axf;\${@:.d=.o} \$@;' #hide
+EOF
+  else cat >> $1 << EOF
+fmt_deps = sed -e 's;^\([a-zA-Z0-9_]*\)\.o;\${@:.d=.o} \$@;'
+EOF
+  fi
+
+  print_config_mk ARCH   "${1}" ${ARCH_LIST}
+  print_config_mk HAVE   "${1}" ${HAVE_LIST}
+  print_config_mk CONFIG "${1}" ${CONFIG_LIST}
+  print_config_mk HAVE   "${1}" gnu_strip
+
+  enabled msvs && echo "CONFIG_VS_VERSION=${vs_version}" >> "${1}"
+
+  CC="${saved_CC}"
+  CXX="${saved_CXX}"
+}
+
+write_common_target_config_h() {
+  print_webm_license ${TMP_H} "/*" " */"
+  cat >> ${TMP_H} << EOF
+/* This file automatically generated by configure. Do not edit! */
+#ifndef AOM_CONFIG_H
+#define AOM_CONFIG_H
+#define RESTRICT    ${RESTRICT}
+#define INLINE      ${INLINE}
+EOF
+  print_config_h ARCH   "${TMP_H}" ${ARCH_LIST}
+  print_config_h HAVE   "${TMP_H}" ${HAVE_LIST}
+  print_config_h CONFIG "${TMP_H}" ${CONFIG_LIST}
+  print_config_vars_h   "${TMP_H}" ${VAR_LIST}
+  echo "#endif /* AOM_CONFIG_H */" >> ${TMP_H}
+  mkdir -p `dirname "$1"`
+  cmp "$1" ${TMP_H} >/dev/null 2>&1 || mv ${TMP_H} "$1"
+}
+
+process_common_cmdline() {
+  for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+      --child)
+        enable_feature child
+        ;;
+      --log*)
+        logging="$optval"
+        if ! disabled logging ; then
+          enabled logging || logfile="$logging"
+        else
+          logfile=/dev/null
+        fi
+        ;;
+      --target=*)
+        toolchain="${toolchain:-${optval}}"
+        ;;
+      --force-target=*)
+        toolchain="${toolchain:-${optval}}"
+        enable_feature force_toolchain
+        ;;
+      --cpu=*)
+        tune_cpu="$optval"
+        ;;
+      --extra-cflags=*)
+        extra_cflags="${optval}"
+        ;;
+      --extra-cxxflags=*)
+        extra_cxxflags="${optval}"
+        ;;
+      --enable-?*|--disable-?*)
+        eval `echo "$opt" | sed 's/--/action=/;s/-/ option=/;s/-/_/g'`
+        if is_in ${option} ${ARCH_EXT_LIST}; then
+          [ $action = "disable" ] && RTCD_OPTIONS="${RTCD_OPTIONS}--disable-${option} "
+        elif [ $action = "disable" ] && ! disabled $option ; then
+          is_in ${option} ${CMDLINE_SELECT} || die_unknown $opt
+          log_echo "  disabling $option"
+        elif [ $action = "enable" ] && ! enabled $option ; then
+          is_in ${option} ${CMDLINE_SELECT} || die_unknown $opt
+          log_echo "  enabling $option"
+        fi
+        ${action}_feature $option
+        ;;
+      --require-?*)
+        eval `echo "$opt" | sed 's/--/action=/;s/-/ option=/;s/-/_/g'`
+        if is_in ${option} ${ARCH_EXT_LIST}; then
+            RTCD_OPTIONS="${RTCD_OPTIONS}${opt} "
+        else
+            die_unknown $opt
+        fi
+        ;;
+      --force-enable-?*|--force-disable-?*)
+        eval `echo "$opt" | sed 's/--force-/action=/;s/-/ option=/;s/-/_/g'`
+        ${action}_feature $option
+        ;;
+      --libc=*)
+        [ -d "${optval}" ] || die "Not a directory: ${optval}"
+        disable_feature builtin_libc
+        alt_libc="${optval}"
+        ;;
+      --as=*)
+        [ "${optval}" = yasm ] || [ "${optval}" = nasm ] \
+          || [ "${optval}" = auto ] \
+          || die "Must be yasm, nasm or auto: ${optval}"
+        alt_as="${optval}"
+        ;;
+      --size-limit=*)
+        w="${optval%%x*}"
+        h="${optval##*x}"
+        VAR_LIST="DECODE_WIDTH_LIMIT ${w} DECODE_HEIGHT_LIMIT ${h}"
+        [ ${w} -gt 0 ] && [ ${h} -gt 0 ] || die "Invalid size-limit: too small."
+        [ ${w} -lt 65536 ] && [ ${h} -lt 65536 ] \
+            || die "Invalid size-limit: too big."
+        enable_feature size_limit
+        ;;
+      --prefix=*)
+        prefix="${optval}"
+        ;;
+      --libdir=*)
+        libdir="${optval}"
+        ;;
+      --sdk-path=*)
+        [ -d "${optval}" ] || die "Not a directory: ${optval}"
+        sdk_path="${optval}"
+        ;;
+      --libc|--as|--prefix|--libdir|--sdk-path)
+        die "Option ${opt} requires argument"
+        ;;
+      --help|-h)
+        show_help
+        ;;
+      *)
+        die_unknown $opt
+        ;;
+    esac
+  done
+}
+
+process_cmdline() {
+  for opt do
+    optval="${opt#*=}"
+    case "$opt" in
+      *)
+        process_common_cmdline $opt
+        ;;
+    esac
+  done
+}
+
+post_process_common_cmdline() {
+  prefix="${prefix:-/usr/local}"
+  prefix="${prefix%/}"
+  libdir="${libdir:-${prefix}/lib}"
+  libdir="${libdir%/}"
+  if [ "${libdir#${prefix}}" = "${libdir}" ]; then
+    die "Libdir ${libdir} must be a subdirectory of ${prefix}"
+  fi
+}
+
+post_process_cmdline() {
+  true;
+}
+
+setup_gnu_toolchain() {
+  CC=${CC:-${CROSS}gcc}
+  CXX=${CXX:-${CROSS}g++}
+  AR=${AR:-${CROSS}ar}
+  LD=${LD:-${CROSS}${link_with_cc:-ld}}
+  AS=${AS:-${CROSS}as}
+  STRIP=${STRIP:-${CROSS}strip}
+  NM=${NM:-${CROSS}nm}
+  AS_SFX=.s
+  EXE_SFX=
+}
+
+# Reliably find the newest available Darwin SDKs. (Older versions of
+# xcrun don't support --show-sdk-path.)
+show_darwin_sdk_path() {
+  xcrun --sdk $1 --show-sdk-path 2>/dev/null ||
+    xcodebuild -sdk $1 -version Path 2>/dev/null
+}
+
+# Print the major version number of the Darwin SDK specified by $1.
+show_darwin_sdk_major_version() {
+  xcrun --sdk $1 --show-sdk-version 2>/dev/null | cut -d. -f1
+}
+
+# Print the Xcode version.
+show_xcode_version() {
+  xcodebuild -version | head -n1 | cut -d' ' -f2
+}
+
+# Fails when Xcode version is less than 6.3.
+check_xcode_minimum_version() {
+  xcode_major=$(show_xcode_version | cut -f1 -d.)
+  xcode_minor=$(show_xcode_version | cut -f2 -d.)
+  xcode_min_major=6
+  xcode_min_minor=3
+  if [ ${xcode_major} -lt ${xcode_min_major} ]; then
+    return 1
+  fi
+  if [ ${xcode_major} -eq ${xcode_min_major} ] \
+    && [ ${xcode_minor} -lt ${xcode_min_minor} ]; then
+    return 1
+  fi
+}
+
+process_common_toolchain() {
+  case "$toolchain" in
+    *-vs*) ;;
+    *) add_cflags_only -std=c99 ;;
+  esac
+
+  if [ -z "$toolchain" ]; then
+    gcctarget="${CHOST:-$(gcc -dumpmachine 2> /dev/null)}"
+
+    # detect tgt_isa
+    case "$gcctarget" in
+      aarch64*)
+        tgt_isa=arm64
+        ;;
+      armv6*)
+        tgt_isa=armv6
+        ;;
+      armv7*-hardfloat* | armv7*-gnueabihf | arm-*-gnueabihf)
+        tgt_isa=armv7
+        float_abi=hard
+        ;;
+      armv7*)
+        tgt_isa=armv7
+        float_abi=softfp
+        ;;
+      *x86_64*|*amd64*)
+        tgt_isa=x86_64
+        ;;
+      *i[3456]86*)
+        tgt_isa=x86
+        ;;
+      *sparc*)
+        tgt_isa=sparc
+        ;;
+    esac
+
+    # detect tgt_os
+    case "$gcctarget" in
+      *darwin10*)
+        tgt_isa=x86_64
+        tgt_os=darwin10
+        ;;
+      *darwin11*)
+        tgt_isa=x86_64
+        tgt_os=darwin11
+        ;;
+      *darwin12*)
+        tgt_isa=x86_64
+        tgt_os=darwin12
+        ;;
+      *darwin13*)
+        tgt_isa=x86_64
+        tgt_os=darwin13
+        ;;
+      *darwin14*)
+        tgt_isa=x86_64
+        tgt_os=darwin14
+        ;;
+      *darwin15*)
+        tgt_isa=x86_64
+        tgt_os=darwin15
+        ;;
+       *darwin16*)
+        tgt_isa=x86_64
+        tgt_os=darwin16
+        ;;
+      x86_64*mingw32*)
+        tgt_os=win64
+        ;;
+      *mingw32*|*cygwin*)
+        [ -z "$tgt_isa" ] && tgt_isa=x86
+        tgt_os=win32
+        ;;
+      *linux*|*bsd*)
+        tgt_os=linux
+        ;;
+      *solaris2.10)
+        tgt_os=solaris
+        ;;
+      *os2*)
+        tgt_os=os2
+        ;;
+    esac
+
+    if [ -n "$tgt_isa" ] && [ -n "$tgt_os" ]; then
+      toolchain=${tgt_isa}-${tgt_os}-gcc
+    fi
+  fi
+
+  toolchain=${toolchain:-generic-gnu}
+
+  is_in ${toolchain} ${all_platforms} || enabled force_toolchain \
+    || die "Unrecognized toolchain '${toolchain}'"
+
+  enabled child || log_echo "Configuring for target '${toolchain}'"
+
+  #
+  # Set up toolchain variables
+  #
+  tgt_isa=$(echo ${toolchain} | awk 'BEGIN{FS="-"}{print $1}')
+  tgt_os=$(echo ${toolchain} | awk 'BEGIN{FS="-"}{print $2}')
+  tgt_cc=$(echo ${toolchain} | awk 'BEGIN{FS="-"}{print $3}')
+
+  # Mark the specific ISA requested as enabled
+  soft_enable ${tgt_isa}
+  enable_feature ${tgt_os}
+  enable_feature ${tgt_cc}
+
+  # Enable the architecture family
+  case ${tgt_isa} in
+    arm*)
+      enable_feature arm
+      ;;
+    mips*)
+      enable_feature mips
+      ;;
+  esac
+
+  # PIC is probably what we want when building shared libs
+  enabled shared && soft_enable pic
+
+  # Minimum iOS version for all target platforms (darwin and iphonesimulator).
+  # Shared library framework builds are only possible on iOS 8 and later.
+  if enabled shared; then
+    IOS_VERSION_OPTIONS="--enable-shared"
+    IOS_VERSION_MIN="8.0"
+  else
+    IOS_VERSION_OPTIONS=""
+    IOS_VERSION_MIN="6.0"
+  fi
+
+  # Handle darwin variants. Newer SDKs allow targeting older
+  # platforms, so use the newest one available.
+  case ${toolchain} in
+    arm*-darwin*)
+      add_cflags "-miphoneos-version-min=${IOS_VERSION_MIN}"
+      iphoneos_sdk_dir="$(show_darwin_sdk_path iphoneos)"
+      if [ -d "${iphoneos_sdk_dir}" ]; then
+        add_cflags  "-isysroot ${iphoneos_sdk_dir}"
+        add_ldflags "-isysroot ${iphoneos_sdk_dir}"
+      fi
+      ;;
+    x86*-darwin*)
+      osx_sdk_dir="$(show_darwin_sdk_path macosx)"
+      if [ -d "${osx_sdk_dir}" ]; then
+        add_cflags  "-isysroot ${osx_sdk_dir}"
+        add_ldflags "-isysroot ${osx_sdk_dir}"
+      fi
+      ;;
+  esac
+
+  case ${toolchain} in
+    *-darwin8-*)
+      add_cflags  "-mmacosx-version-min=10.4"
+      add_ldflags "-mmacosx-version-min=10.4"
+      ;;
+    *-darwin9-*)
+      add_cflags  "-mmacosx-version-min=10.5"
+      add_ldflags "-mmacosx-version-min=10.5"
+      ;;
+    *-darwin10-*)
+      add_cflags  "-mmacosx-version-min=10.6"
+      add_ldflags "-mmacosx-version-min=10.6"
+      ;;
+    *-darwin11-*)
+      add_cflags  "-mmacosx-version-min=10.7"
+      add_ldflags "-mmacosx-version-min=10.7"
+      ;;
+    *-darwin12-*)
+      add_cflags  "-mmacosx-version-min=10.8"
+      add_ldflags "-mmacosx-version-min=10.8"
+      ;;
+    *-darwin13-*)
+      add_cflags  "-mmacosx-version-min=10.9"
+      add_ldflags "-mmacosx-version-min=10.9"
+      ;;
+    *-darwin14-*)
+      add_cflags  "-mmacosx-version-min=10.10"
+      add_ldflags "-mmacosx-version-min=10.10"
+      ;;
+    *-darwin15-*)
+      add_cflags  "-mmacosx-version-min=10.11"
+      add_ldflags "-mmacosx-version-min=10.11"
+      ;;
+    *-darwin16-*)
+      add_cflags  "-mmacosx-version-min=10.12"
+      add_ldflags "-mmacosx-version-min=10.12"
+      ;;
+    *-iphonesimulator-*)
+      add_cflags  "-miphoneos-version-min=${IOS_VERSION_MIN}"
+      add_ldflags "-miphoneos-version-min=${IOS_VERSION_MIN}"
+      iossim_sdk_dir="$(show_darwin_sdk_path iphonesimulator)"
+      if [ -d "${iossim_sdk_dir}" ]; then
+        add_cflags  "-isysroot ${iossim_sdk_dir}"
+        add_ldflags "-isysroot ${iossim_sdk_dir}"
+      fi
+      ;;
+  esac
+
+  # Handle Solaris variants. Solaris 10 needs -lposix4
+  case ${toolchain} in
+    sparc-solaris-*)
+      add_extralibs -lposix4
+      ;;
+    *-solaris-*)
+      add_extralibs -lposix4
+      ;;
+  esac
+
+  # Process ARM architecture variants
+  case ${toolchain} in
+    arm*)
+      # on arm, isa versions are supersets
+      case ${tgt_isa} in
+        arm64|armv8)
+          soft_enable neon
+          ;;
+        armv7|armv7s)
+          soft_enable neon
+          # Only enable neon_asm when neon is also enabled.
+          enabled neon && soft_enable neon_asm
+          # If someone tries to force it through, die.
+          if disabled neon && enabled neon_asm; then
+            die "Disabling neon while keeping neon-asm is not supported"
+          fi
+          case ${toolchain} in
+            # Apple iOS SDKs no longer support armv6 as of the version 9
+            # release (coincides with release of Xcode 7). Only enable media
+            # when using earlier SDK releases.
+            *-darwin*)
+              if [ "$(show_darwin_sdk_major_version iphoneos)" -lt 9 ]; then
+                soft_enable media
+              else
+                soft_disable media
+                RTCD_OPTIONS="${RTCD_OPTIONS}--disable-media "
+              fi
+              ;;
+            *)
+              soft_enable media
+              ;;
+          esac
+          ;;
+        armv6)
+          case ${toolchain} in
+            *-darwin*)
+              if [ "$(show_darwin_sdk_major_version iphoneos)" -lt 9 ]; then
+                soft_enable media
+              else
+                die "Your iOS SDK does not support armv6."
+              fi
+              ;;
+            *)
+              soft_enable media
+              ;;
+          esac
+          ;;
+      esac
+
+      asm_conversion_cmd="cat"
+
+      case ${tgt_cc} in
+        gcc)
+          link_with_cc=gcc
+          setup_gnu_toolchain
+          arch_int=${tgt_isa##armv}
+          arch_int=${arch_int%%te}
+          check_add_asflags --defsym ARCHITECTURE=${arch_int}
+          tune_cflags="-mtune="
+          if [ ${tgt_isa} = "armv7" ] || [ ${tgt_isa} = "armv7s" ]; then
+            if [ -z "${float_abi}" ]; then
+              check_cpp <<EOF && float_abi=hard || float_abi=softfp
+#ifndef __ARM_PCS_VFP
+#error "not hardfp"
+#endif
+EOF
+            fi
+            check_add_cflags  -march=armv7-a -mfloat-abi=${float_abi}
+            check_add_asflags -march=armv7-a -mfloat-abi=${float_abi}
+
+            if enabled neon || enabled neon_asm; then
+              check_add_cflags -mfpu=neon #-ftree-vectorize
+              check_add_asflags -mfpu=neon
+            fi
+          elif [ ${tgt_isa} = "arm64" ] || [ ${tgt_isa} = "armv8" ]; then
+            check_add_cflags -march=armv8-a
+            check_add_asflags -march=armv8-a
+          else
+            check_add_cflags -march=${tgt_isa}
+            check_add_asflags -march=${tgt_isa}
+          fi
+
+          enabled debug && add_asflags -g
+          asm_conversion_cmd="${source_path}/build/make/ads2gas.pl"
+          if enabled thumb; then
+            asm_conversion_cmd="$asm_conversion_cmd -thumb"
+            check_add_cflags -mthumb
+            check_add_asflags -mthumb -mimplicit-it=always
+          fi
+          ;;
+        vs*)
+          asm_conversion_cmd="${source_path}/build/make/ads2armasm_ms.pl"
+          AS_SFX=.s
+          msvs_arch_dir=arm-msvs
+          disable_feature multithread
+          disable_feature unit_tests
+          vs_version=${tgt_cc##vs}
+          # MSVC 2013 doesn't allow doing plain .exe projects for ARM,
+          # only "AppContainerApplication" which requires an AppxManifest.
+          # Therefore disable the examples, just build the library.
+          disable_feature examples
+          ;;
+        rvct)
+          CC=armcc
+          AR=armar
+          AS=armasm
+          LD="${source_path}/build/make/armlink_adapter.sh"
+          STRIP=arm-none-linux-gnueabi-strip
+          NM=arm-none-linux-gnueabi-nm
+          tune_cflags="--cpu="
+          tune_asflags="--cpu="
+          if [ -z "${tune_cpu}" ]; then
+            if [ ${tgt_isa} = "armv7" ]; then
+              if enabled neon || enabled neon_asm
+              then
+                check_add_cflags --fpu=softvfp+vfpv3
+                check_add_asflags --fpu=softvfp+vfpv3
+              fi
+              check_add_cflags --cpu=Cortex-A8
+              check_add_asflags --cpu=Cortex-A8
+            else
+              check_add_cflags --cpu=${tgt_isa##armv}
+              check_add_asflags --cpu=${tgt_isa##armv}
+            fi
+          fi
+          arch_int=${tgt_isa##armv}
+          arch_int=${arch_int%%te}
+          check_add_asflags --pd "\"ARCHITECTURE SETA ${arch_int}\""
+          enabled debug && add_asflags -g
+          add_cflags --gnu
+          add_cflags --enum_is_int
+          add_cflags --wchar32
+          ;;
+      esac
+
+      case ${tgt_os} in
+        none*)
+          disable_feature multithread
+          disable_feature os_support
+          ;;
+
+        android*)
+          if [ -z "${sdk_path}" ]; then
+            die "Must specify --sdk-path for Android builds."
+          fi
+
+          SDK_PATH=${sdk_path}
+          COMPILER_LOCATION=`find "${SDK_PATH}" \
+                             -name "arm-linux-androideabi-gcc*" -print -quit`
+          TOOLCHAIN_PATH=${COMPILER_LOCATION%/*}/arm-linux-androideabi-
+          CC=${TOOLCHAIN_PATH}gcc
+          CXX=${TOOLCHAIN_PATH}g++
+          AR=${TOOLCHAIN_PATH}ar
+          LD=${TOOLCHAIN_PATH}gcc
+          AS=${TOOLCHAIN_PATH}as
+          STRIP=${TOOLCHAIN_PATH}strip
+          NM=${TOOLCHAIN_PATH}nm
+
+          if [ -z "${alt_libc}" ]; then
+            alt_libc=`find "${SDK_PATH}" -name arch-arm -print | \
+              awk '{n = split($0,a,"/"); \
+                split(a[n-1],b,"-"); \
+                print $0 " " b[2]}' | \
+                sort -g -k 2 | \
+                awk '{ print $1 }' | tail -1`
+          fi
+
+          if [ -d "${alt_libc}" ]; then
+            add_cflags "--sysroot=${alt_libc}"
+            add_ldflags "--sysroot=${alt_libc}"
+          fi
+
+          # linker flag that routes around a CPU bug in some
+          # Cortex-A8 implementations (NDK Dev Guide)
+          add_ldflags "-Wl,--fix-cortex-a8"
+
+          enable_feature pic
+          soft_enable realtime_only
+          if [ ${tgt_isa} = "armv7" ]; then
+            soft_enable runtime_cpu_detect
+          fi
+          if enabled runtime_cpu_detect; then
+            add_cflags "-I${SDK_PATH}/sources/android/cpufeatures"
+          fi
+          ;;
+
+        darwin*)
+          XCRUN_FIND="xcrun --sdk iphoneos --find"
+          CXX="$(${XCRUN_FIND} clang++)"
+          CC="$(${XCRUN_FIND} clang)"
+          AR="$(${XCRUN_FIND} ar)"
+          AS="$(${XCRUN_FIND} as)"
+          STRIP="$(${XCRUN_FIND} strip)"
+          NM="$(${XCRUN_FIND} nm)"
+          RANLIB="$(${XCRUN_FIND} ranlib)"
+          AS_SFX=.s
+          LD="${CXX:-$(${XCRUN_FIND} ld)}"
+
+          # ASFLAGS is written here instead of using check_add_asflags
+          # because we need to overwrite all of ASFLAGS and purge the
+          # options that were put in above
+          ASFLAGS="-arch ${tgt_isa} -g"
+
+          add_cflags -arch ${tgt_isa}
+          add_ldflags -arch ${tgt_isa}
+
+          alt_libc="$(show_darwin_sdk_path iphoneos)"
+          if [ -d "${alt_libc}" ]; then
+            add_cflags -isysroot ${alt_libc}
+          fi
+
+          if [ "${LD}" = "${CXX}" ]; then
+            add_ldflags -miphoneos-version-min="${IOS_VERSION_MIN}"
+          else
+            add_ldflags -ios_version_min "${IOS_VERSION_MIN}"
+          fi
+
+          for d in lib usr/lib usr/lib/system; do
+            try_dir="${alt_libc}/${d}"
+            [ -d "${try_dir}" ] && add_ldflags -L"${try_dir}"
+          done
+
+          case ${tgt_isa} in
+            armv7|armv7s|armv8|arm64)
+              if enabled neon && ! check_xcode_minimum_version; then
+                soft_disable neon
+                log_echo "  neon disabled: upgrade Xcode (need v6.3+)."
+                if enabled neon_asm; then
+                  soft_disable neon_asm
+                  log_echo "  neon_asm disabled: upgrade Xcode (need v6.3+)."
+                fi
+              fi
+              ;;
+          esac
+
+          asm_conversion_cmd="${source_path}/build/make/ads2gas_apple.pl"
+
+          if [ "$(show_darwin_sdk_major_version iphoneos)" -gt 8 ]; then
+            check_add_cflags -fembed-bitcode
+            check_add_asflags -fembed-bitcode
+            check_add_ldflags -fembed-bitcode
+          fi
+          ;;
+
+        linux*)
+          enable_feature linux
+          if enabled rvct; then
+            # Check if we have CodeSourcery GCC in PATH. Needed for
+            # libraries
+            which arm-none-linux-gnueabi-gcc 2>&- || \
+              die "Couldn't find CodeSourcery GCC from PATH"
+
+            # Use armcc as a linker to enable translation of
+            # some gcc specific options such as -lm and -lpthread.
+            LD="armcc --translate_gcc"
+
+            # create configuration file (uses path to CodeSourcery GCC)
+            armcc --arm_linux_configure --arm_linux_config_file=arm_linux.cfg
+
+            add_cflags --arm_linux_paths --arm_linux_config_file=arm_linux.cfg
+            add_asflags --no_hide_all --apcs=/interwork
+            add_ldflags --arm_linux_paths --arm_linux_config_file=arm_linux.cfg
+            enabled pic && add_cflags --apcs=/fpic
+            enabled pic && add_asflags --apcs=/fpic
+            enabled shared && add_cflags --shared
+          fi
+          ;;
+      esac
+      ;;
+    mips*)
+      link_with_cc=gcc
+      setup_gnu_toolchain
+      tune_cflags="-mtune="
+      if enabled dspr2; then
+        check_add_cflags -mips32r2 -mdspr2
+      fi
+
+      if enabled runtime_cpu_detect; then
+        disable_feature runtime_cpu_detect
+      fi
+
+      if [ -n "${tune_cpu}" ]; then
+        case ${tune_cpu} in
+          p5600)
+            check_add_cflags -mips32r5 -mload-store-pairs
+            check_add_cflags -msched-weight -mhard-float -mfp64
+            check_add_asflags -mips32r5 -mhard-float -mfp64
+            check_add_ldflags -mfp64
+            ;;
+          i6400|p6600)
+            check_add_cflags -mips64r6 -mabi=64 -msched-weight
+            check_add_cflags  -mload-store-pairs -mhard-float -mfp64
+            check_add_asflags -mips64r6 -mabi=64 -mhard-float -mfp64
+            check_add_ldflags -mips64r6 -mabi=64 -mfp64
+            ;;
+        esac
+
+        if enabled msa; then
+          add_cflags -mmsa
+          add_asflags -mmsa
+          add_ldflags -mmsa
+        fi
+      fi
+
+      check_add_cflags -march=${tgt_isa}
+      check_add_asflags -march=${tgt_isa}
+      check_add_asflags -KPIC
+      ;;
+    x86*)
+      case  ${tgt_os} in
+        win*)
+          enabled gcc && add_cflags -fno-common
+          ;;
+        solaris*)
+          CC=${CC:-${CROSS}gcc}
+          CXX=${CXX:-${CROSS}g++}
+          LD=${LD:-${CROSS}gcc}
+          CROSS=${CROSS-g}
+          ;;
+        os2)
+          disable_feature pic
+          AS=${AS:-nasm}
+          add_ldflags -Zhigh-mem
+          ;;
+      esac
+
+      AS="${alt_as:-${AS:-auto}}"
+      case  ${tgt_cc} in
+        icc*)
+          CC=${CC:-icc}
+          LD=${LD:-icc}
+          setup_gnu_toolchain
+          add_cflags -use-msasm  # remove -use-msasm too?
+          # add -no-intel-extensions to suppress warning #10237
+          # refer to http://software.intel.com/en-us/forums/topic/280199
+          add_ldflags -i-static -no-intel-extensions
+          enabled x86_64 && add_cflags -ipo -static -O3 -no-prec-div
+          enabled x86_64 && AR=xiar
+          case ${tune_cpu} in
+            atom*)
+              tune_cflags="-x"
+              tune_cpu="SSE3_ATOM"
+              ;;
+            *)
+              tune_cflags="-march="
+              ;;
+          esac
+          ;;
+        gcc*)
+          link_with_cc=gcc
+          tune_cflags="-march="
+          setup_gnu_toolchain
+          #for 32 bit x86 builds, -O3 did not turn on this flag
+          enabled optimizations && disabled gprof && check_add_cflags -fomit-frame-pointer
+          ;;
+        vs*)
+          # When building with Microsoft Visual Studio the assembler is
+          # invoked directly. Checking at configure time is unnecessary.
+          # Skip the check by setting AS arbitrarily
+          AS=msvs
+          msvs_arch_dir=x86-msvs
+          vc_version=${tgt_cc##vs}
+          ;;
+      esac
+
+      bits=32
+      enabled x86_64 && bits=64
+      check_cpp <<EOF && bits=x32
+#if !defined(__ILP32__) || !defined(__x86_64__)
+#error "not x32"
+#endif
+EOF
+      case ${tgt_cc} in
+        gcc*)
+          add_cflags -m${bits}
+          add_ldflags -m${bits}
+          ;;
+      esac
+
+      soft_enable runtime_cpu_detect
+      # We can't use 'check_cflags' until the compiler is configured and CC is
+      # populated.
+      for ext in ${ARCH_EXT_LIST_X86}; do
+        # disable higher order extensions to simplify asm dependencies
+        if [ "$disable_exts" = "yes" ]; then
+          if ! disabled $ext; then
+            RTCD_OPTIONS="${RTCD_OPTIONS}--disable-${ext} "
+            disable_feature $ext
+          fi
+        elif disabled $ext; then
+          disable_exts="yes"
+        else
+          # use the shortened version for the flag: sse4_1 -> sse4
+          check_gcc_machine_option ${ext%_*} $ext
+        fi
+      done
+
+      if enabled external_build; then
+        log_echo "  skipping assembler detection"
+      else
+        case "${AS}" in
+          auto|"")
+            which nasm >/dev/null 2>&1 && AS=nasm
+            which yasm >/dev/null 2>&1 && AS=yasm
+            if [ "${AS}" = nasm ] ; then
+              # Apple ships version 0.98 of nasm through at least Xcode 6. Revisit
+              # this check if they start shipping a compatible version.
+              apple=`nasm -v | grep "Apple"`
+              [ -n "${apple}" ] \
+                && echo "Unsupported version of nasm: ${apple}" \
+                && AS=""
+            fi
+            [ "${AS}" = auto ] || [ -z "${AS}" ] \
+              && die "Neither yasm nor nasm have been found." \
+                     "See the prerequisites section in the README for more info."
+            ;;
+        esac
+        log_echo "  using $AS"
+      fi
+      [ "${AS##*/}" = nasm ] && add_asflags -Ox
+      AS_SFX=.asm
+      case  ${tgt_os} in
+        win32)
+          add_asflags -f win32
+          enabled debug && add_asflags -g cv8
+          EXE_SFX=.exe
+          ;;
+        win64)
+          add_asflags -f x64
+          enabled debug && add_asflags -g cv8
+          EXE_SFX=.exe
+          ;;
+        linux*|solaris*|android*)
+          add_asflags -f elf${bits}
+          enabled debug && [ "${AS}" = yasm ] && add_asflags -g dwarf2
+          enabled debug && [ "${AS}" = nasm ] && add_asflags -g
+          [ "${AS##*/}" = nasm ] && check_asm_align
+          ;;
+        darwin*)
+          add_asflags -f macho${bits}
+          enabled x86 && darwin_arch="-arch i386" || darwin_arch="-arch x86_64"
+          add_cflags  ${darwin_arch}
+          add_ldflags ${darwin_arch}
+          # -mdynamic-no-pic is still a bit of voodoo -- it was required at
+          # one time, but does not seem to be now, and it breaks some of the
+          # code that still relies on inline assembly.
+          # enabled icc && ! enabled pic && add_cflags -fno-pic -mdynamic-no-pic
+          enabled icc && ! enabled pic && add_cflags -fno-pic
+          ;;
+        iphonesimulator)
+          add_asflags -f macho${bits}
+          enabled x86 && sim_arch="-arch i386" || sim_arch="-arch x86_64"
+          add_cflags  ${sim_arch}
+          add_ldflags ${sim_arch}
+
+          if [ "$(show_darwin_sdk_major_version iphonesimulator)" -gt 8 ]; then
+            # yasm v1.3.0 doesn't know what -fembed-bitcode means, so turning it
+            # on is pointless (unless building a C-only lib). Warn the user, but
+            # do nothing here.
+            log "Warning: Bitcode embed disabled for simulator targets."
+          fi
+          ;;
+        os2)
+          add_asflags -f aout
+          enabled debug && add_asflags -g
+          EXE_SFX=.exe
+          ;;
+        *)
+          log "Warning: Unknown os $tgt_os while setting up $AS flags"
+          ;;
+      esac
+      ;;
+    *-gcc|generic-gnu)
+      link_with_cc=gcc
+      enable_feature gcc
+      setup_gnu_toolchain
+      ;;
+  esac
+
+  # Try to enable CPU specific tuning
+  if [ -n "${tune_cpu}" ]; then
+    if [ -n "${tune_cflags}" ]; then
+      check_add_cflags ${tune_cflags}${tune_cpu} || \
+        die "Requested CPU '${tune_cpu}' not supported by compiler"
+    fi
+    if [ -n "${tune_asflags}" ]; then
+      check_add_asflags ${tune_asflags}${tune_cpu} || \
+        die "Requested CPU '${tune_cpu}' not supported by assembler"
+    fi
+    if [ -z "${tune_cflags}${tune_asflags}" ]; then
+      log_echo "Warning: CPU tuning not supported by this toolchain"
+    fi
+  fi
+
+  if enabled debug; then
+    check_add_cflags -g && check_add_ldflags -g
+  else
+    check_add_cflags -DNDEBUG
+  fi
+
+  enabled gprof && check_add_cflags -pg && check_add_ldflags -pg
+  enabled gcov &&
+    check_add_cflags -fprofile-arcs -ftest-coverage &&
+    check_add_ldflags -fprofile-arcs -ftest-coverage
+
+  if enabled optimizations; then
+    if enabled rvct; then
+      enabled small && check_add_cflags -Ospace || check_add_cflags -Otime
+    else
+      enabled small && check_add_cflags -O2 ||  check_add_cflags -O3
+    fi
+  fi
+
+  # Position Independent Code (PIC) support, for building relocatable
+  # shared objects
+  enabled gcc && enabled pic && check_add_cflags -fPIC
+
+  # Work around longjmp interception on glibc >= 2.11, to improve binary
+  # compatibility. See http://code.google.com/p/webm/issues/detail?id=166
+  enabled linux && check_add_cflags -U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=0
+
+  # Check for strip utility variant
+  ${STRIP} -V 2>/dev/null | grep GNU >/dev/null && enable_feature gnu_strip
+
+  # Try to determine target endianness
+  check_cc <<EOF
+unsigned int e = 'O'<<24 | '2'<<16 | 'B'<<8 | 'E';
+EOF
+    [ -f "${TMP_O}" ] && od -A n -t x1 "${TMP_O}" | tr -d '\n' |
+        grep '4f *32 *42 *45' >/dev/null 2>&1 && enable_feature big_endian
+
+    # Try to find which inline keywords are supported
+    check_cc <<EOF && INLINE="inline"
+static inline function() {}
+EOF
+
+  # Almost every platform uses pthreads.
+  if enabled multithread; then
+    case ${toolchain} in
+      *-win*-vs*)
+        ;;
+      *-android-gcc)
+        ;;
+      *)
+        check_header pthread.h && add_extralibs -lpthread
+        ;;
+    esac
+  fi
+
+  # only for MIPS platforms
+  case ${toolchain} in
+    mips*)
+      if enabled big_endian; then
+        if enabled dspr2; then
+          echo "dspr2 optimizations are available only for little endian platforms"
+          disable_feature dspr2
+        fi
+        if enabled msa; then
+          echo "msa optimizations are available only for little endian platforms"
+          disable_feature msa
+        fi
+      fi
+      ;;
+  esac
+
+  # glibc needs these
+  if enabled linux || [ "$toolchain" = "generic-gnu" ]; then
+    add_cflags -D_LARGEFILE_SOURCE
+    add_cflags -D_FILE_OFFSET_BITS=64
+  fi
+}
+
+process_toolchain() {
+  process_common_toolchain
+}
+
+print_config_mk() {
+  saved_prefix="${prefix}"
+  prefix=$1
+  makefile=$2
+  shift 2
+  for cfg; do
+    if enabled $cfg; then
+      upname="`toupper $cfg`"
+      echo "${prefix}_${upname}=yes" >> $makefile
+    fi
+  done
+  prefix="${saved_prefix}"
+}
+
+print_config_h() {
+  saved_prefix="${prefix}"
+  prefix=$1
+  header=$2
+  shift 2
+  for cfg; do
+    upname="`toupper $cfg`"
+    if enabled $cfg; then
+      echo "#define ${prefix}_${upname} 1" >> $header
+    else
+      echo "#define ${prefix}_${upname} 0" >> $header
+    fi
+  done
+  prefix="${saved_prefix}"
+}
+
+print_config_vars_h() {
+  header=$1
+  shift
+  while [ $# -gt 0 ]; do
+    upname="`toupper $1`"
+    echo "#define ${upname} $2" >> $header
+    shift 2
+  done
+}
+
+print_webm_license() {
+  saved_prefix="${prefix}"
+  destination=$1
+  prefix="$2"
+  suffix="$3"
+  shift 3
+  cat <<EOF > ${destination}
+${prefix} Copyright (c) 2016, Alliance for Open Media. All rights reserved.${suffix}
+${prefix} ${suffix}
+${prefix} This source code is subject to the terms of the BSD 2 Clause License and${suffix}
+${prefix} the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License${suffix}
+${prefix} was not distributed with this source code in the LICENSE file, you can${suffix}
+${prefix} obtain it at www.aomedia.org/license/software. If the Alliance for Open${suffix}
+${prefix} Media Patent License 1.0 was not distributed with this source code in the${suffix}
+${prefix} PATENTS file, you can obtain it at www.aomedia.org/license/patent.${suffix}
+EOF
+  prefix="${saved_prefix}"
+}
+
+process_targets() {
+  true;
+}
+
+process_detect() {
+  true;
+}
+
+enable_feature logging
+logfile="config.log"
+self=$0
+process() {
+  cmdline_args="$@"
+  process_cmdline "$@"
+  if enabled child; then
+    echo "# ${self} $@" >> ${logfile}
+  else
+    echo "# ${self} $@" > ${logfile}
+  fi
+  post_process_common_cmdline
+  post_process_cmdline
+  process_toolchain
+  process_detect
+  process_targets
+
+  OOT_INSTALLS="${OOT_INSTALLS}"
+  if enabled source_path_used; then
+  # Prepare the PWD for building.
+  for f in ${OOT_INSTALLS}; do
+    install -D "${source_path}/$f" "$f"
+  done
+  fi
+  cp "${source_path}/build/make/Makefile" .
+
+  clean_temp_files
+  true
+}
diff --git a/third_party/aom/build/make/gen_asm_deps.sh b/third_party/aom/build/make/gen_asm_deps.sh
new file mode 100755
index 0000000000..c867cc2bf5
--- /dev/null
+++ b/third_party/aom/build/make/gen_asm_deps.sh
@@ -0,0 +1,65 @@
+#!/bin/sh
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+self=$0
+show_help() {
+    echo "usage: $self [options] <srcfile>"
+    echo
+    echo "Generate Makefile dependency information from assembly code source"
+    echo
+    exit 1
+}
+die_unknown(){
+    echo "Unknown option \"$1\"."
+    echo "See $0 --help for available options."
+    exit 1
+}
+for opt do
+    optval="${opt#*=}"
+    case "$opt" in
+    --build-pfx=*) pfx="${optval}"
+    ;;
+    --depfile=*) out="${optval}"
+    ;;
+    -I*) raw_inc_paths="${raw_inc_paths} ${opt}"
+         inc_path="${inc_path} ${opt#-I}"
+    ;;
+    -h|--help) show_help
+    ;;
+    *) [ -f "$opt" ] && srcfile="$opt"
+    ;;
+    esac
+done
+
+[ -n "$srcfile" ] || show_help
+sfx=${sfx:-asm}
+includes=$(LC_ALL=C egrep -i "include +\"?[a-z0-9_/]+\.${sfx}" $srcfile |
+           perl -p -e "s;.*?([a-z0-9_/]+.${sfx}).*;\1;")
+#" restore editor state
+for inc in ${includes}; do
+    found_inc_path=
+    for idir in ${inc_path}; do
+        [ -f "${idir}/${inc}" ] && found_inc_path="${idir}" && break
+    done
+    if [ -f `dirname $srcfile`/$inc ]; then
+        # Handle include files in the same directory as the source
+        $self --build-pfx=$pfx --depfile=$out ${raw_inc_paths} `dirname $srcfile`/$inc
+    elif [ -n "${found_inc_path}" ]; then
+        # Handle include files on the include path
+        $self --build-pfx=$pfx --depfile=$out ${raw_inc_paths} "${found_inc_path}/$inc"
+    else
+        # Handle generated includes in the build root (which may not exist yet)
+        echo ${out} ${out%d}o: "${pfx}${inc}"
+    fi
+done
+echo ${out} ${out%d}o: $srcfile
diff --git a/third_party/aom/build/make/gen_msvs_def.sh b/third_party/aom/build/make/gen_msvs_def.sh
new file mode 100755
index 0000000000..dbb2674ac5
--- /dev/null
+++ b/third_party/aom/build/make/gen_msvs_def.sh
@@ -0,0 +1,82 @@
+#!/bin/bash
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+self=$0
+self_basename=${self##*/}
+EOL=$'\n'
+
+show_help() {
+    cat <<EOF
+Usage: ${self_basename} [options] file1 [file2 ...]
+
+This script generates a MSVC module definition file containing a list of symbols
+to export from a DLL. Source files are technically bash scripts (and thus may
+use #comment syntax) but in general, take the form of a list of symbols:
+
+  <kind> symbol1 [symbol2, symbol3, ...]
+
+where <kind> is either 'text' or 'data'
+
+
+Options:
+    --help                      Print this message
+    --out=filename              Write output to a file [stdout]
+    --name=project_name         Name of the library (required)
+EOF
+    exit 1
+}
+
+die() {
+    echo "${self_basename}: $@"
+    exit 1
+}
+
+die_unknown(){
+    echo "Unknown option \"$1\"."
+    echo "See ${self_basename} --help for available options."
+    exit 1
+}
+
+text() {
+    for sym in "$@"; do
+        echo "  $sym" >> ${outfile}
+    done
+}
+
+data() {
+    for sym in "$@"; do
+        printf "  %-40s DATA\n" "$sym" >> ${outfile}
+    done
+}
+
+# Process command line
+for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+    --help|-h) show_help
+    ;;
+    --out=*) outfile="$optval"
+    ;;
+    --name=*) name="${optval}"
+    ;;
+     -*) die_unknown $opt
+    ;;
+    *) file_list[${#file_list[@]}]="$opt"
+    esac
+done
+outfile=${outfile:-/dev/stdout}
+[ -n "$name" ] || die "Library name (--name) must be specified!"
+
+echo "LIBRARY ${name}" > ${outfile}
+echo "EXPORTS" >> ${outfile}
+for f in "${file_list[@]}"; do
+    . $f
+done
diff --git a/third_party/aom/build/make/gen_msvs_sln.sh b/third_party/aom/build/make/gen_msvs_sln.sh
new file mode 100755
index 0000000000..77c68f1da2
--- /dev/null
+++ b/third_party/aom/build/make/gen_msvs_sln.sh
@@ -0,0 +1,254 @@
+#!/bin/bash
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+self=$0
+self_basename=${self##*/}
+EOL=$'\n'
+EOLDOS=$'\r'
+
+show_help() {
+    cat <<EOF
+Usage: ${self_basename} [options] file1 [file2 ...]
+
+This script generates a Visual Studio solution file from a list of project
+files.
+
+Options:
+    --help                      Print this message
+    --out=outfile               Redirect output to a file
+    --ver=version               Version (12,14) of visual studio to generate for
+    --target=isa-os-cc          Target specifier
+EOF
+    exit 1
+}
+
+die() {
+    echo "${self_basename}: $@" >&2
+    [ -f "${outfile}" ] && rm -f ${outfile}{,.mk}
+    exit 1
+}
+
+die_unknown(){
+    echo "Unknown option \"$1\"." >&2
+    echo "See ${self_basename} --help for available options." >&2
+    [ -f "${outfile}" ] && rm -f ${outfile}{,.mk}
+    exit 1
+}
+
+indent1=$'\t'
+indent=""
+indent_push() {
+    indent="${indent}${indent1}"
+}
+indent_pop() {
+    indent="${indent%${indent1}}"
+}
+
+parse_project() {
+    local file=$1
+    local name=`grep RootNamespace "$file" | sed 's,.*<.*>\(.*\)</.*>.*,\1,'`
+    local guid=`grep ProjectGuid "$file" | sed 's,.*<.*>\(.*\)</.*>.*,\1,'`
+
+    # save the project GUID to a varaible, normalizing to the basename of the
+    # vcxproj file without the extension
+    local var
+    var=${file##*/}
+    var=${var%%.${sfx}}
+    eval "${var}_file=\"$1\""
+    eval "${var}_name=$name"
+    eval "${var}_guid=$guid"
+
+    cur_config_list=`grep -B1 'Label="Configuration"' $file |
+        grep Condition | cut -d\' -f4`
+    new_config_list=$(for i in $config_list $cur_config_list; do
+        echo $i
+    done | sort | uniq)
+    if [ "$config_list" != "" ] && [ "$config_list" != "$new_config_list" ]; then
+        mixed_platforms=1
+    fi
+    config_list="$new_config_list"
+    eval "${var}_config_list=\"$cur_config_list\""
+    proj_list="${proj_list} ${var}"
+}
+
+process_project() {
+    eval "local file=\${$1_file}"
+    eval "local name=\${$1_name}"
+    eval "local guid=\${$1_guid}"
+
+    # save the project GUID to a varaible, normalizing to the basename of the
+    # vcproj file without the extension
+    local var
+    var=${file##*/}
+    var=${var%%.${sfx}}
+    eval "${var}_guid=$guid"
+
+    echo "Project(\"{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}\") = \"$name\", \"$file\", \"$guid\""
+    echo "EndProject"
+}
+
+process_global() {
+    echo "Global"
+    indent_push
+
+    #
+    # Solution Configuration Platforms
+    #
+    echo "${indent}GlobalSection(SolutionConfigurationPlatforms) = preSolution"
+    indent_push
+    IFS_bak=${IFS}
+    IFS=$'\r'$'\n'
+    if [ "$mixed_platforms" != "" ]; then
+        config_list="
+Release|Mixed Platforms
+Debug|Mixed Platforms"
+    fi
+    for config in ${config_list}; do
+        echo "${indent}$config = $config"
+    done
+    IFS=${IFS_bak}
+    indent_pop
+    echo "${indent}EndGlobalSection"
+
+    #
+    # Project Configuration Platforms
+    #
+    echo "${indent}GlobalSection(ProjectConfigurationPlatforms) = postSolution"
+    indent_push
+    for proj in ${proj_list}; do
+        eval "local proj_guid=\${${proj}_guid}"
+        eval "local proj_config_list=\${${proj}_config_list}"
+        IFS=$'\r'$'\n'
+        for config in ${proj_config_list}; do
+            if [ "$mixed_platforms" != "" ]; then
+                local c=${config%%|*}
+                echo "${indent}${proj_guid}.${c}|Mixed Platforms.ActiveCfg = ${config}"
+                echo "${indent}${proj_guid}.${c}|Mixed Platforms.Build.0 = ${config}"
+            else
+                echo "${indent}${proj_guid}.${config}.ActiveCfg = ${config}"
+                echo "${indent}${proj_guid}.${config}.Build.0 = ${config}"
+            fi
+
+        done
+        IFS=${IFS_bak}
+    done
+    indent_pop
+    echo "${indent}EndGlobalSection"
+
+    #
+    # Solution Properties
+    #
+    echo "${indent}GlobalSection(SolutionProperties) = preSolution"
+    indent_push
+    echo "${indent}HideSolutionNode = FALSE"
+    indent_pop
+    echo "${indent}EndGlobalSection"
+
+    indent_pop
+    echo "EndGlobal"
+}
+
+process_makefile() {
+    IFS_bak=${IFS}
+    IFS=$'\r'$'\n'
+    local TAB=$'\t'
+    cat <<EOF
+MSBUILD_TOOL := msbuild.exe
+found_devenv := \$(shell which \$(MSBUILD_TOOL) >/dev/null 2>&1 && echo yes)
+.nodevenv.once:
+${TAB}@echo "  * \$(MSBUILD_TOOL) not found in path."
+${TAB}@echo "  * "
+${TAB}@echo "  * You will have to build all configurations manually using the"
+${TAB}@echo "  * Visual Studio IDE. To allow make to build them automatically,"
+${TAB}@echo "  * add the Common7/IDE directory of your Visual Studio"
+${TAB}@echo "  * installation to your path, eg:"
+${TAB}@echo "  *   C:\Program Files\Microsoft Visual Studio 12.0\Common7\IDE"
+${TAB}@echo "  * "
+${TAB}@touch \$@
+CLEAN-OBJS += \$(if \$(found_devenv),,.nodevenv.once)
+
+EOF
+
+    for sln_config in ${config_list}; do
+        local config=${sln_config%%|*}
+        local platform=${sln_config##*|}
+        local nows_sln_config=`echo $sln_config | sed -e 's/[^a-zA-Z0-9]/_/g'`
+        cat <<EOF
+BUILD_TARGETS += \$(if \$(NO_LAUNCH_DEVENV),,$nows_sln_config)
+clean::
+${TAB}rm -rf "$platform"/"$config"
+.PHONY: $nows_sln_config
+ifneq (\$(found_devenv),)
+$nows_sln_config: $outfile
+${TAB}\$(MSBUILD_TOOL) $outfile -m -t:Build \\
+${TAB}${TAB}-p:Configuration="$config" -p:Platform="$platform"
+else
+$nows_sln_config: $outfile .nodevenv.once
+${TAB}@echo "  * Skipping build of $sln_config (\$(MSBUILD_TOOL) not in path)."
+${TAB}@echo "  * "
+endif
+
+EOF
+    done
+    IFS=${IFS_bak}
+}
+
+# Process command line
+outfile=/dev/stdout
+for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+    --help|-h) show_help
+    ;;
+    --out=*) outfile="${optval}"; mkoutfile="${optval}".mk
+    ;;
+    --dep=*) eval "${optval%%:*}_deps=\"\${${optval%%:*}_deps} ${optval##*:}\""
+    ;;
+    --ver=*) vs_ver="$optval"
+             case $optval in
+             12|14)
+             ;;
+             *) die Unrecognized Visual Studio Version in $opt
+             ;;
+             esac
+    ;;
+    --target=*) target="${optval}"
+    ;;
+    -*) die_unknown $opt
+    ;;
+    *) file_list[${#file_list[@]}]="$opt"
+    esac
+done
+outfile=${outfile:-/dev/stdout}
+mkoutfile=${mkoutfile:-/dev/stdout}
+case "${vs_ver:-12}" in
+    12) sln_vers="12.00"
+       sln_vers_str="Visual Studio 2013"
+    ;;
+    14) sln_vers="14.00"
+       sln_vers_str="Visual Studio 2015"
+    ;;
+esac
+sfx=vcxproj
+
+for f in "${file_list[@]}"; do
+    parse_project $f
+done
+cat  >${outfile} <<EOF
+Microsoft Visual Studio Solution File, Format Version $sln_vers${EOLDOS}
+# $sln_vers_str${EOLDOS}
+EOF
+for proj in ${proj_list}; do
+    process_project $proj >>${outfile}
+done
+process_global >>${outfile}
+process_makefile >${mkoutfile}
diff --git a/third_party/aom/build/make/gen_msvs_vcxproj.sh b/third_party/aom/build/make/gen_msvs_vcxproj.sh
new file mode 100755
index 0000000000..7543bda172
--- /dev/null
+++ b/third_party/aom/build/make/gen_msvs_vcxproj.sh
@@ -0,0 +1,474 @@
+#!/bin/bash
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+self=$0
+self_basename=${self##*/}
+self_dirname=$(dirname "$0")
+
+. "$self_dirname/msvs_common.sh"|| exit 127
+
+show_help() {
+    cat <<EOF
+Usage: ${self_basename} --name=projname [options] file1 [file2 ...]
+
+This script generates a Visual Studio project file from a list of source
+code files.
+
+Options:
+    --help                      Print this message
+    --exe                       Generate a project for building an Application
+    --lib                       Generate a project for creating a static library
+    --dll                       Generate a project for creating a dll
+    --static-crt                Use the static C runtime (/MT)
+    --enable-werror             Treat warnings as errors (/WX)
+    --target=isa-os-cc          Target specifier (required)
+    --out=filename              Write output to a file [stdout]
+    --name=project_name         Name of the project (required)
+    --proj-guid=GUID            GUID to use for the project
+    --module-def=filename       File containing export definitions (for DLLs)
+    --ver=version               Version (12,14) of visual studio to generate for
+    --src-path-bare=dir         Path to root of source tree
+    -Ipath/to/include           Additional include directories
+    -DFLAG[=value]              Preprocessor macros to define
+    -Lpath/to/lib               Additional library search paths
+    -llibname                   Library to link against
+EOF
+    exit 1
+}
+
+tag_content() {
+    local tag=$1
+    local content=$2
+    shift
+    shift
+    if [ $# -ne 0 ]; then
+        echo "${indent}<${tag}"
+        indent_push
+        tag_attributes "$@"
+        echo "${indent}>${content}</${tag}>"
+        indent_pop
+    else
+        echo "${indent}<${tag}>${content}</${tag}>"
+    fi
+}
+
+generate_filter() {
+    local name=$1
+    local pats=$2
+    local file_list_sz
+    local i
+    local f
+    local saveIFS="$IFS"
+    local pack
+    echo "generating filter '$name' from ${#file_list[@]} files" >&2
+    IFS=*
+
+    file_list_sz=${#file_list[@]}
+    for i in ${!file_list[@]}; do
+        f=${file_list[i]}
+        for pat in ${pats//;/$IFS}; do
+            if [ "${f##*.}" == "$pat" ]; then
+                unset file_list[i]
+
+                objf=$(echo ${f%.*}.obj \
+                       | sed -e "s,$src_path_bare,," \
+                             -e 's/^[\./]\+//g' -e 's,[:/ ],_,g')
+
+                if ([ "$pat" == "asm" ] || [ "$pat" == "s" ]) && $asm_use_custom_step; then
+                    # Avoid object file name collisions, i.e. aom_config.c and
+                    # aom_config.asm produce the same object file without
+                    # this additional suffix.
+                    objf=${objf%.obj}_asm.obj
+                    open_tag CustomBuild \
+                        Include="$f"
+                    for plat in "${platforms[@]}"; do
+                        for cfg in Debug Release; do
+                            tag_content Message "Assembling %(Filename)%(Extension)" \
+                                Condition="'\$(Configuration)|\$(Platform)'=='$cfg|$plat'"
+                            tag_content Command "$(eval echo \$asm_${cfg}_cmdline) -o \$(IntDir)$objf" \
+                                Condition="'\$(Configuration)|\$(Platform)'=='$cfg|$plat'"
+                            tag_content Outputs "\$(IntDir)$objf" \
+                                Condition="'\$(Configuration)|\$(Platform)'=='$cfg|$plat'"
+                        done
+                    done
+                    close_tag CustomBuild
+                elif [ "$pat" == "c" ] || \
+                     [ "$pat" == "cc" ] || [ "$pat" == "cpp" ]; then
+                    open_tag ClCompile \
+                        Include="$f"
+                    # Separate file names with Condition?
+                    tag_content ObjectFileName "\$(IntDir)$objf"
+                    # Check for AVX and turn it on to avoid warnings.
+                    if [[ $f =~ avx.?\.c$ ]]; then
+                        tag_content AdditionalOptions "/arch:AVX"
+                    fi
+                    close_tag ClCompile
+                elif [ "$pat" == "h" ] ; then
+                    tag ClInclude \
+                        Include="$f"
+                elif [ "$pat" == "vcxproj" ] ; then
+                    open_tag ProjectReference \
+                        Include="$f"
+                    depguid=`grep ProjectGuid "$f" | sed 's,.*<.*>\(.*\)</.*>.*,\1,'`
+                    tag_content Project "$depguid"
+                    tag_content ReferenceOutputAssembly false
+                    close_tag ProjectReference
+                else
+                    tag None \
+                        Include="$f"
+                fi
+
+                break
+            fi
+        done
+    done
+
+    IFS="$saveIFS"
+}
+
+# Process command line
+unset target
+for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+        --help|-h) show_help
+        ;;
+        --target=*) target="${optval}"
+        ;;
+        --out=*) outfile="$optval"
+        ;;
+        --name=*) name="${optval}"
+        ;;
+        --proj-guid=*) guid="${optval}"
+        ;;
+        --module-def=*) module_def="${optval}"
+        ;;
+        --exe) proj_kind="exe"
+        ;;
+        --dll) proj_kind="dll"
+        ;;
+        --lib) proj_kind="lib"
+        ;;
+        --src-path-bare=*)
+            src_path_bare=$(fix_path "$optval")
+            src_path_bare=${src_path_bare%/}
+        ;;
+        --static-crt) use_static_runtime=true
+        ;;
+        --enable-werror) werror=true
+        ;;
+        --ver=*)
+            vs_ver="$optval"
+            case "$optval" in
+                12|14)
+                ;;
+                *) die Unrecognized Visual Studio Version in $opt
+                ;;
+            esac
+        ;;
+        -I*)
+            opt=${opt##-I}
+            opt=$(fix_path "$opt")
+            opt="${opt%/}"
+            incs="${incs}${incs:+;}&quot;${opt}&quot;"
+            yasmincs="${yasmincs} -I&quot;${opt}&quot;"
+        ;;
+        -D*) defines="${defines}${defines:+;}${opt##-D}"
+        ;;
+        -L*) # fudge . to $(OutDir)
+            if [ "${opt##-L}" == "." ]; then
+                libdirs="${libdirs}${libdirs:+;}&quot;\$(OutDir)&quot;"
+            else
+                 # Also try directories for this platform/configuration
+                 opt=${opt##-L}
+                 opt=$(fix_path "$opt")
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}&quot;"
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}/\$(PlatformName)/\$(Configuration)&quot;"
+                 libdirs="${libdirs}${libdirs:+;}&quot;${opt}/\$(PlatformName)&quot;"
+            fi
+        ;;
+        -l*) libs="${libs}${libs:+ }${opt##-l}.lib"
+        ;;
+        -*) die_unknown $opt
+        ;;
+        *)
+            # The paths in file_list are fixed outside of the loop.
+            file_list[${#file_list[@]}]="$opt"
+            case "$opt" in
+                 *.asm|*.s) uses_asm=true
+                 ;;
+            esac
+        ;;
+    esac
+done
+
+# Make one call to fix_path for file_list to improve performance.
+fix_file_list file_list
+
+outfile=${outfile:-/dev/stdout}
+guid=${guid:-`generate_uuid`}
+asm_use_custom_step=false
+uses_asm=${uses_asm:-false}
+case "${vs_ver:-12}" in
+    12|14)
+       asm_use_custom_step=$uses_asm
+    ;;
+esac
+
+[ -n "$name" ] || die "Project name (--name) must be specified!"
+[ -n "$target" ] || die "Target (--target) must be specified!"
+
+if ${use_static_runtime:-false}; then
+    release_runtime=MultiThreaded
+    debug_runtime=MultiThreadedDebug
+    lib_sfx=mt
+else
+    release_runtime=MultiThreadedDLL
+    debug_runtime=MultiThreadedDebugDLL
+    lib_sfx=md
+fi
+
+# Calculate debug lib names: If a lib ends in ${lib_sfx}.lib, then rename
+# it to ${lib_sfx}d.lib. This precludes linking to release libs from a
+# debug exe, so this may need to be refactored later.
+for lib in ${libs}; do
+    if [ "$lib" != "${lib%${lib_sfx}.lib}" ]; then
+        lib=${lib%.lib}d.lib
+    fi
+    debug_libs="${debug_libs}${debug_libs:+ }${lib}"
+done
+debug_libs=${debug_libs// /;}
+libs=${libs// /;}
+
+
+# List of all platforms supported for this target
+case "$target" in
+    x86_64*)
+        platforms[0]="x64"
+        asm_Debug_cmdline="yasm -Xvc -g cv8 -f win64 ${yasmincs} &quot;%(FullPath)&quot;"
+        asm_Release_cmdline="yasm -Xvc -f win64 ${yasmincs} &quot;%(FullPath)&quot;"
+    ;;
+    x86*)
+        platforms[0]="Win32"
+        asm_Debug_cmdline="yasm -Xvc -g cv8 -f win32 ${yasmincs} &quot;%(FullPath)&quot;"
+        asm_Release_cmdline="yasm -Xvc -f win32 ${yasmincs} &quot;%(FullPath)&quot;"
+    ;;
+    arm*)
+        platforms[0]="ARM"
+        asm_Debug_cmdline="armasm -nologo -oldit &quot;%(FullPath)&quot;"
+        asm_Release_cmdline="armasm -nologo -oldit &quot;%(FullPath)&quot;"
+    ;;
+    *) die "Unsupported target $target!"
+    ;;
+esac
+
+generate_vcxproj() {
+    echo "<?xml version=\"1.0\" encoding=\"utf-8\"?>"
+    open_tag Project \
+        DefaultTargets="Build" \
+        ToolsVersion="4.0" \
+        xmlns="http://schemas.microsoft.com/developer/msbuild/2003" \
+
+    open_tag ItemGroup \
+        Label="ProjectConfigurations"
+    for plat in "${platforms[@]}"; do
+        for config in Debug Release; do
+            open_tag ProjectConfiguration \
+                Include="$config|$plat"
+            tag_content Configuration $config
+            tag_content Platform $plat
+            close_tag ProjectConfiguration
+        done
+    done
+    close_tag ItemGroup
+
+    open_tag PropertyGroup \
+        Label="Globals"
+        tag_content ProjectGuid "{${guid}}"
+        tag_content RootNamespace ${name}
+        tag_content Keyword ManagedCProj
+        if [ "${platforms[0]}" = "ARM" ]; then
+            tag_content AppContainerApplication true
+            # The application type can be one of "Windows Store",
+            # "Windows Phone" or "Windows Phone Silverlight". The
+            # actual value doesn't matter from the libaom point of view,
+            # since a static library built for one works on the others.
+            # The PlatformToolset field needs to be set in sync with this;
+            # for Windows Store and Windows Phone Silverlight it should be
+            # v120 while it should be v120_wp81 if the type is Windows Phone.
+            tag_content ApplicationType "Windows Store"
+            tag_content ApplicationTypeRevision 8.1
+        fi
+    close_tag PropertyGroup
+
+    tag Import \
+        Project="\$(VCTargetsPath)\\Microsoft.Cpp.Default.props"
+
+    for plat in "${platforms[@]}"; do
+        for config in Release Debug; do
+            open_tag PropertyGroup \
+                Condition="'\$(Configuration)|\$(Platform)'=='$config|$plat'" \
+                Label="Configuration"
+            if [ "$proj_kind" = "exe" ]; then
+                tag_content ConfigurationType Application
+            elif [ "$proj_kind" = "dll" ]; then
+                tag_content ConfigurationType DynamicLibrary
+            else
+                tag_content ConfigurationType StaticLibrary
+            fi
+            if [ "$vs_ver" = "12" ]; then
+                # Setting a PlatformToolset indicating windows phone isn't
+                # enough to build code for arm with MSVC 2013, one strictly
+                # has to enable AppContainerApplication as well.
+                tag_content PlatformToolset v120
+            fi
+            if [ "$vs_ver" = "14" ]; then
+                tag_content PlatformToolset v140
+            fi
+            tag_content CharacterSet Unicode
+            if [ "$config" = "Release" ]; then
+                tag_content WholeProgramOptimization true
+            fi
+            close_tag PropertyGroup
+        done
+    done
+
+    tag Import \
+        Project="\$(VCTargetsPath)\\Microsoft.Cpp.props"
+
+    open_tag ImportGroup \
+        Label="PropertySheets"
+        tag Import \
+            Project="\$(UserRootDir)\\Microsoft.Cpp.\$(Platform).user.props" \
+            Condition="exists('\$(UserRootDir)\\Microsoft.Cpp.\$(Platform).user.props')" \
+            Label="LocalAppDataPlatform"
+    close_tag ImportGroup
+
+    tag PropertyGroup \
+        Label="UserMacros"
+
+    for plat in "${platforms[@]}"; do
+        plat_no_ws=`echo $plat | sed 's/[^A-Za-z0-9_]/_/g'`
+        for config in Debug Release; do
+            open_tag PropertyGroup \
+                Condition="'\$(Configuration)|\$(Platform)'=='$config|$plat'"
+            tag_content OutDir "\$(SolutionDir)$plat_no_ws\\\$(Configuration)\\"
+            tag_content IntDir "$plat_no_ws\\\$(Configuration)\\${name}\\"
+            if [ "$proj_kind" == "lib" ]; then
+              if [ "$config" == "Debug" ]; then
+                config_suffix=d
+              else
+                config_suffix=""
+              fi
+              tag_content TargetName "${name}${lib_sfx}${config_suffix}"
+            fi
+            close_tag PropertyGroup
+        done
+    done
+
+    for plat in "${platforms[@]}"; do
+        for config in Debug Release; do
+            open_tag ItemDefinitionGroup \
+                Condition="'\$(Configuration)|\$(Platform)'=='$config|$plat'"
+            if [ "$name" == "aom" ]; then
+                hostplat=$plat
+                if [ "$hostplat" == "ARM" ]; then
+                    hostplat=Win32
+                fi
+            fi
+            open_tag ClCompile
+            if [ "$config" = "Debug" ]; then
+                opt=Disabled
+                runtime=$debug_runtime
+                curlibs=$debug_libs
+                debug=_DEBUG
+            else
+                opt=MaxSpeed
+                runtime=$release_runtime
+                curlibs=$libs
+                tag_content FavorSizeOrSpeed Speed
+                debug=NDEBUG
+            fi
+            extradefines=";$defines"
+            tag_content Optimization $opt
+            tag_content AdditionalIncludeDirectories "$incs;%(AdditionalIncludeDirectories)"
+            tag_content PreprocessorDefinitions "WIN32;$debug;_CRT_SECURE_NO_WARNINGS;_CRT_SECURE_NO_DEPRECATE$extradefines;%(PreprocessorDefinitions)"
+            tag_content RuntimeLibrary $runtime
+            tag_content WarningLevel Level3
+            if ${werror:-false}; then
+                tag_content TreatWarningAsError true
+            fi
+            # We need to override the defaults for these settings
+            # if AppContainerApplication is set.
+            tag_content CompileAsWinRT false
+            tag_content PrecompiledHeader NotUsing
+            tag_content SDLCheck false
+            close_tag ClCompile
+            case "$proj_kind" in
+            exe)
+                open_tag Link
+                tag_content GenerateDebugInformation true
+                # Console is the default normally, but if
+                # AppContainerApplication is set, we need to override it.
+                tag_content SubSystem Console
+                close_tag Link
+                ;;
+            dll)
+                open_tag Link
+                tag_content GenerateDebugInformation true
+                tag_content ModuleDefinitionFile $module_def
+                close_tag Link
+                ;;
+            lib)
+                ;;
+            esac
+            close_tag ItemDefinitionGroup
+        done
+
+    done
+
+    open_tag ItemGroup
+    generate_filter "Source Files"   "c;cc;cpp;def;odl;idl;hpj;bat;asm;asmx;s"
+    close_tag ItemGroup
+    open_tag ItemGroup
+    generate_filter "Header Files"   "h;hm;inl;inc;xsd"
+    close_tag ItemGroup
+    open_tag ItemGroup
+    generate_filter "Build Files"    "mk"
+    close_tag ItemGroup
+    open_tag ItemGroup
+    generate_filter "References"     "vcxproj"
+    close_tag ItemGroup
+
+    tag Import \
+        Project="\$(VCTargetsPath)\\Microsoft.Cpp.targets"
+
+    open_tag ImportGroup \
+        Label="ExtensionTargets"
+    close_tag ImportGroup
+
+    close_tag Project
+
+    # This must be done from within the {} subshell
+    echo "Ignored files list (${#file_list[@]} items) is:" >&2
+    for f in "${file_list[@]}"; do
+        echo "    $f" >&2
+    done
+}
+
+# This regexp doesn't catch most of the strings in the vcxproj format,
+# since they're like <tag>path</tag> instead of <tag attr="path" />
+# as previously. It still seems to work ok despite this.
+generate_vcxproj |
+    sed  -e '/"/s;\([^ "]\)/;\1\\;g' |
+    sed  -e '/xmlns/s;\\;/;g' > ${outfile}
+
+exit
diff --git a/third_party/aom/build/make/ios-Info.plist b/third_party/aom/build/make/ios-Info.plist
new file mode 100644
index 0000000000..300e3e310d
--- /dev/null
+++ b/third_party/aom/build/make/ios-Info.plist
@@ -0,0 +1,37 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+	<key>CFBundleDevelopmentRegion</key>
+	<string>en</string>
+	<key>CFBundleExecutable</key>
+	<string>AOM</string>
+	<key>CFBundleIdentifier</key>
+	<string>org.webmproject.AOM</string>
+	<key>CFBundleInfoDictionaryVersion</key>
+	<string>6.0</string>
+	<key>CFBundleName</key>
+	<string>AOM</string>
+	<key>CFBundlePackageType</key>
+	<string>FMWK</string>
+	<key>CFBundleShortVersionString</key>
+	<string>${VERSION}</string>
+	<key>CFBundleSignature</key>
+	<string>????</string>
+	<key>CFBundleSupportedPlatforms</key>
+	<array>
+		<string>iPhoneOS</string>
+	</array>
+	<key>CFBundleVersion</key>
+	<string>${VERSION}</string>
+	<key>MinimumOSVersion</key>
+	<string>${IOS_VERSION_MIN}</string>
+	<key>UIDeviceFamily</key>
+	<array>
+		<integer>1</integer>
+		<integer>2</integer>
+	</array>
+	<key>AOMFullVersion</key>
+	<string>${FULLVERSION}</string>
+</dict>
+</plist>
diff --git a/third_party/aom/build/make/iosbuild.sh b/third_party/aom/build/make/iosbuild.sh
new file mode 100755
index 0000000000..ca8214b622
--- /dev/null
+++ b/third_party/aom/build/make/iosbuild.sh
@@ -0,0 +1,383 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This script generates 'AOM.framework'. An iOS app can encode and decode AVx
+## video by including 'AOM.framework'.
+##
+## Run iosbuild.sh to create 'AOM.framework' in the current directory.
+##
+set -e
+devnull='> /dev/null 2>&1'
+
+BUILD_ROOT="_iosbuild"
+CONFIGURE_ARGS="--disable-docs
+                --disable-examples
+                --disable-libyuv
+                --disable-unit-tests"
+DIST_DIR="_dist"
+FRAMEWORK_DIR="AOM.framework"
+FRAMEWORK_LIB="AOM.framework/AOM"
+HEADER_DIR="${FRAMEWORK_DIR}/Headers/aom"
+SCRIPT_DIR=$(dirname "$0")
+LIBAOM_SOURCE_DIR=$(cd ${SCRIPT_DIR}/../..; pwd)
+LIPO=$(xcrun -sdk iphoneos${SDK} -find lipo)
+ORIG_PWD="$(pwd)"
+ARM_TARGETS="arm64-darwin-gcc
+             armv7-darwin-gcc
+             armv7s-darwin-gcc"
+SIM_TARGETS="x86-iphonesimulator-gcc
+             x86_64-iphonesimulator-gcc"
+OSX_TARGETS="x86-darwin16-gcc
+             x86_64-darwin16-gcc"
+TARGETS="${ARM_TARGETS} ${SIM_TARGETS}"
+
+# Configures for the target specified by $1, and invokes make with the dist
+# target using $ as the distribution output directory.
+build_target() {
+  local target="$1"
+  local old_pwd="$(pwd)"
+  local target_specific_flags=""
+
+  vlog "***Building target: ${target}***"
+
+  case "${target}" in
+    x86-*)
+      target_specific_flags="--enable-pic"
+      vlog "Enabled PIC for ${target}"
+      ;;
+  esac
+
+  mkdir "${target}"
+  cd "${target}"
+  eval "${LIBAOM_SOURCE_DIR}/configure" --target="${target}" \
+    ${CONFIGURE_ARGS} ${EXTRA_CONFIGURE_ARGS} ${target_specific_flags} \
+    ${devnull}
+  export DIST_DIR
+  eval make dist ${devnull}
+  cd "${old_pwd}"
+
+  vlog "***Done building target: ${target}***"
+}
+
+# Returns the preprocessor symbol for the target specified by $1.
+target_to_preproc_symbol() {
+  target="$1"
+  case "${target}" in
+    arm64-*)
+      echo "__aarch64__"
+      ;;
+    armv7-*)
+      echo "__ARM_ARCH_7A__"
+      ;;
+    armv7s-*)
+      echo "__ARM_ARCH_7S__"
+      ;;
+    x86-*)
+      echo "__i386__"
+      ;;
+    x86_64-*)
+      echo "__x86_64__"
+      ;;
+    *)
+      echo "#error ${target} unknown/unsupported"
+      return 1
+      ;;
+  esac
+}
+
+# Create a aom_config.h shim that, based on preprocessor settings for the
+# current target CPU, includes the real aom_config.h for the current target.
+# $1 is the list of targets.
+create_aom_framework_config_shim() {
+  local targets="$1"
+  local config_file="${HEADER_DIR}/aom_config.h"
+  local preproc_symbol=""
+  local target=""
+  local include_guard="AOM_FRAMEWORK_HEADERS_AOM_AOM_CONFIG_H_"
+
+  local file_header="/*
+ *  Copyright (c) $(date +%Y), Alliance for Open Media. All rights reserved.
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+/* GENERATED FILE: DO NOT EDIT! */
+
+#ifndef ${include_guard}
+#define ${include_guard}
+
+#if defined"
+
+  printf "%s" "${file_header}" > "${config_file}"
+  for target in ${targets}; do
+    preproc_symbol=$(target_to_preproc_symbol "${target}")
+    printf " ${preproc_symbol}\n" >> "${config_file}"
+    printf "#define AOM_FRAMEWORK_TARGET \"${target}\"\n" >> "${config_file}"
+    printf "#include \"AOM/aom/${target}/aom_config.h\"\n" >> "${config_file}"
+    printf "#elif defined" >> "${config_file}"
+    mkdir "${HEADER_DIR}/${target}"
+    cp -p "${BUILD_ROOT}/${target}/aom_config.h" "${HEADER_DIR}/${target}"
+  done
+
+  # Consume the last line of output from the loop: We don't want it.
+  sed -i '' -e '$d' "${config_file}"
+
+  printf "#endif\n\n" >> "${config_file}"
+  printf "#endif  // ${include_guard}" >> "${config_file}"
+}
+
+# Verifies that $FRAMEWORK_LIB fat library contains requested builds.
+verify_framework_targets() {
+  local requested_cpus=""
+  local cpu=""
+
+  # Extract CPU from full target name.
+  for target; do
+    cpu="${target%%-*}"
+    if [ "${cpu}" = "x86" ]; then
+      # lipo -info outputs i386 for libaom x86 targets.
+      cpu="i386"
+    fi
+    requested_cpus="${requested_cpus}${cpu} "
+  done
+
+  # Get target CPUs present in framework library.
+  local targets_built=$(${LIPO} -info ${FRAMEWORK_LIB})
+
+  # $LIPO -info outputs a string like the following:
+  #   Architectures in the fat file: $FRAMEWORK_LIB <architectures>
+  # Capture only the architecture strings.
+  targets_built=${targets_built##*: }
+
+  # Sort CPU strings to make the next step a simple string compare.
+  local actual=$(echo ${targets_built} | tr " " "\n" | sort | tr "\n" " ")
+  local requested=$(echo ${requested_cpus} | tr " " "\n" | sort | tr "\n" " ")
+
+  vlog "Requested ${FRAMEWORK_LIB} CPUs: ${requested}"
+  vlog "Actual ${FRAMEWORK_LIB} CPUs: ${actual}"
+
+  if [ "${requested}" != "${actual}" ]; then
+    elog "Actual ${FRAMEWORK_LIB} targets do not match requested target list."
+    elog "  Requested target CPUs: ${requested}"
+    elog "  Actual target CPUs: ${actual}"
+    return 1
+  fi
+}
+
+# Configures and builds each target specified by $1, and then builds
+# AOM.framework.
+build_framework() {
+  local lib_list=""
+  local targets="$1"
+  local target=""
+  local target_dist_dir=""
+
+  # Clean up from previous build(s).
+  rm -rf "${BUILD_ROOT}" "${FRAMEWORK_DIR}"
+
+  # Create output dirs.
+  mkdir -p "${BUILD_ROOT}"
+  mkdir -p "${HEADER_DIR}"
+
+  cd "${BUILD_ROOT}"
+
+  for target in ${targets}; do
+    build_target "${target}"
+    target_dist_dir="${BUILD_ROOT}/${target}/${DIST_DIR}"
+    if [ "${ENABLE_SHARED}" = "yes" ]; then
+      local suffix="dylib"
+    else
+      local suffix="a"
+    fi
+    lib_list="${lib_list} ${target_dist_dir}/lib/libaom.${suffix}"
+  done
+
+  cd "${ORIG_PWD}"
+
+  # The basic libaom API includes are all the same; just grab the most recent
+  # set.
+  cp -p "${target_dist_dir}"/include/aom/* "${HEADER_DIR}"
+
+  # Build the fat library.
+  ${LIPO} -create ${lib_list} -output ${FRAMEWORK_DIR}/AOM
+
+  # Create the aom_config.h shim that allows usage of aom_config.h from
+  # within AOM.framework.
+  create_aom_framework_config_shim "${targets}"
+
+  # Copy in aom_version.h.
+  cp -p "${BUILD_ROOT}/${target}/aom_version.h" "${HEADER_DIR}"
+
+  if [ "${ENABLE_SHARED}" = "yes" ]; then
+    # Adjust the dylib's name so dynamic linking in apps works as expected.
+    install_name_tool -id '@rpath/AOM.framework/AOM' ${FRAMEWORK_DIR}/AOM
+
+    # Copy in Info.plist.
+    cat "${SCRIPT_DIR}/ios-Info.plist" \
+      | sed "s/\${FULLVERSION}/${FULLVERSION}/g" \
+      | sed "s/\${VERSION}/${VERSION}/g" \
+      | sed "s/\${IOS_VERSION_MIN}/${IOS_VERSION_MIN}/g" \
+      > "${FRAMEWORK_DIR}/Info.plist"
+  fi
+
+  # Confirm AOM.framework/AOM contains the targets requested.
+  verify_framework_targets ${targets}
+
+  vlog "Created fat library ${FRAMEWORK_LIB} containing:"
+  for lib in ${lib_list}; do
+    vlog "  $(echo ${lib} | awk -F / '{print $2, $NF}')"
+  done
+}
+
+# Trap function. Cleans up the subtree used to build all targets contained in
+# $TARGETS.
+cleanup() {
+  local readonly res=$?
+  cd "${ORIG_PWD}"
+
+  if [ $res -ne 0 ]; then
+    elog "build exited with error ($res)"
+  fi
+
+  if [ "${PRESERVE_BUILD_OUTPUT}" != "yes" ]; then
+    rm -rf "${BUILD_ROOT}"
+  fi
+}
+
+print_list() {
+  local indent="$1"
+  shift
+  local list="$@"
+  for entry in ${list}; do
+    echo "${indent}${entry}"
+  done
+}
+
+iosbuild_usage() {
+cat << EOF
+  Usage: ${0##*/} [arguments]
+    --help: Display this message and exit.
+    --enable-shared: Build a dynamic framework for use on iOS 8 or later.
+    --extra-configure-args <args>: Extra args to pass when configuring libaom.
+    --macosx: Uses darwin16 targets instead of iphonesimulator targets for x86
+              and x86_64. Allows linking to framework when builds target MacOSX
+              instead of iOS.
+    --preserve-build-output: Do not delete the build directory.
+    --show-build-output: Show output from each library build.
+    --targets <targets>: Override default target list. Defaults:
+$(print_list "        " ${TARGETS})
+    --test-link: Confirms all targets can be linked. Functionally identical to
+                 passing --enable-examples via --extra-configure-args.
+    --verbose: Output information about the environment and each stage of the
+               build.
+EOF
+}
+
+elog() {
+  echo "${0##*/} failed because: $@" 1>&2
+}
+
+vlog() {
+  if [ "${VERBOSE}" = "yes" ]; then
+    echo "$@"
+  fi
+}
+
+trap cleanup EXIT
+
+# Parse the command line.
+while [ -n "$1" ]; do
+  case "$1" in
+    --extra-configure-args)
+      EXTRA_CONFIGURE_ARGS="$2"
+      shift
+      ;;
+    --help)
+      iosbuild_usage
+      exit
+      ;;
+    --enable-shared)
+      ENABLE_SHARED=yes
+      ;;
+    --preserve-build-output)
+      PRESERVE_BUILD_OUTPUT=yes
+      ;;
+    --show-build-output)
+      devnull=
+      ;;
+    --test-link)
+      EXTRA_CONFIGURE_ARGS="${EXTRA_CONFIGURE_ARGS} --enable-examples"
+      ;;
+    --targets)
+      TARGETS="$2"
+      shift
+      ;;
+    --macosx)
+      TARGETS="${ARM_TARGETS} ${OSX_TARGETS}"
+      ;;
+    --verbose)
+      VERBOSE=yes
+      ;;
+    *)
+      iosbuild_usage
+      exit 1
+      ;;
+  esac
+  shift
+done
+
+if [ "${ENABLE_SHARED}" = "yes" ]; then
+  CONFIGURE_ARGS="--enable-shared ${CONFIGURE_ARGS}"
+fi
+
+FULLVERSION=$("${SCRIPT_DIR}"/version.sh --bare "${LIBAOM_SOURCE_DIR}")
+VERSION=$(echo "${FULLVERSION}" | sed -E 's/^v([0-9]+\.[0-9]+\.[0-9]+).*$/\1/')
+
+if [ "$ENABLE_SHARED" = "yes" ]; then
+  IOS_VERSION_OPTIONS="--enable-shared"
+  IOS_VERSION_MIN="8.0"
+else
+  IOS_VERSION_OPTIONS=""
+  IOS_VERSION_MIN="6.0"
+fi
+
+if [ "${VERBOSE}" = "yes" ]; then
+cat << EOF
+  BUILD_ROOT=${BUILD_ROOT}
+  DIST_DIR=${DIST_DIR}
+  CONFIGURE_ARGS=${CONFIGURE_ARGS}
+  EXTRA_CONFIGURE_ARGS=${EXTRA_CONFIGURE_ARGS}
+  FRAMEWORK_DIR=${FRAMEWORK_DIR}
+  FRAMEWORK_LIB=${FRAMEWORK_LIB}
+  HEADER_DIR=${HEADER_DIR}
+  LIBAOM_SOURCE_DIR=${LIBAOM_SOURCE_DIR}
+  LIPO=${LIPO}
+  MAKEFLAGS=${MAKEFLAGS}
+  ORIG_PWD=${ORIG_PWD}
+  PRESERVE_BUILD_OUTPUT=${PRESERVE_BUILD_OUTPUT}
+  TARGETS="$(print_list "" ${TARGETS})"
+  ENABLE_SHARED=${ENABLE_SHARED}
+  OSX_TARGETS="${OSX_TARGETS}"
+  SIM_TARGETS="${SIM_TARGETS}"
+  SCRIPT_DIR="${SCRIPT_DIR}"
+  FULLVERSION="${FULLVERSION}"
+  VERSION="${VERSION}"
+  IOS_VERSION_MIN="${IOS_VERSION_MIN}"
+EOF
+fi
+
+build_framework "${TARGETS}"
+echo "Successfully built '${FRAMEWORK_DIR}' for:"
+print_list "" ${TARGETS}
diff --git a/third_party/aom/build/make/msvs_common.sh b/third_party/aom/build/make/msvs_common.sh
new file mode 100644
index 0000000000..2df27df8de
--- /dev/null
+++ b/third_party/aom/build/make/msvs_common.sh
@@ -0,0 +1,114 @@
+#!/bin/bash
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+if [ "$(uname -o 2>/dev/null)" = "Cygwin" ] \
+   && cygpath --help >/dev/null 2>&1; then
+    FIXPATH='cygpath -m'
+else
+    FIXPATH='echo_path'
+fi
+
+die() {
+    echo "${self_basename}: $@" >&2
+    exit 1
+}
+
+die_unknown(){
+    echo "Unknown option \"$1\"." >&2
+    echo "See ${self_basename} --help for available options." >&2
+    exit 1
+}
+
+echo_path() {
+    for path; do
+        echo "$path"
+    done
+}
+
+# Output one, possibly changed based on the system, path per line.
+fix_path() {
+    $FIXPATH "$@"
+}
+
+# Corrects the paths in file_list in one pass for efficiency.
+# $1 is the name of the array to be modified.
+fix_file_list() {
+    declare -n array_ref=$1
+    files=$(fix_path "${array_ref[@]}")
+    local IFS=$'\n'
+    array_ref=($files)
+}
+
+generate_uuid() {
+    local hex="0123456789ABCDEF"
+    local i
+    local uuid=""
+    local j
+    #93995380-89BD-4b04-88EB-625FBE52EBFB
+    for ((i=0; i<32; i++)); do
+        (( j = $RANDOM % 16 ))
+        uuid="${uuid}${hex:$j:1}"
+    done
+    echo "${uuid:0:8}-${uuid:8:4}-${uuid:12:4}-${uuid:16:4}-${uuid:20:12}"
+}
+
+indent1="    "
+indent=""
+indent_push() {
+    indent="${indent}${indent1}"
+}
+indent_pop() {
+    indent="${indent%${indent1}}"
+}
+
+tag_attributes() {
+    for opt in "$@"; do
+        optval="${opt#*=}"
+        [ -n "${optval}" ] ||
+            die "Missing attribute value in '$opt' while generating $tag tag"
+        echo "${indent}${opt%%=*}=\"${optval}\""
+    done
+}
+
+open_tag() {
+    local tag=$1
+    shift
+    if [ $# -ne 0 ]; then
+        echo "${indent}<${tag}"
+        indent_push
+        tag_attributes "$@"
+        echo "${indent}>"
+    else
+        echo "${indent}<${tag}>"
+        indent_push
+    fi
+}
+
+close_tag() {
+    local tag=$1
+    indent_pop
+    echo "${indent}</${tag}>"
+}
+
+tag() {
+    local tag=$1
+    shift
+    if [ $# -ne 0 ]; then
+        echo "${indent}<${tag}"
+        indent_push
+        tag_attributes "$@"
+        indent_pop
+        echo "${indent}/>"
+    else
+        echo "${indent}<${tag}/>"
+    fi
+}
+
diff --git a/third_party/aom/build/make/rtcd.pl b/third_party/aom/build/make/rtcd.pl
new file mode 100755
index 0000000000..354ae51766
--- /dev/null
+++ b/third_party/aom/build/make/rtcd.pl
@@ -0,0 +1,424 @@
+#!/usr/bin/env perl
+
+no strict 'refs';
+use warnings;
+use Getopt::Long;
+Getopt::Long::Configure("auto_help") if $Getopt::Long::VERSION > 2.32;
+
+my %ALL_FUNCS = ();
+my @ALL_ARCHS;
+my @ALL_FORWARD_DECLS;
+my @REQUIRES;
+
+my %opts = ();
+my %disabled = ();
+my %required = ();
+
+my @argv;
+foreach (@ARGV) {
+  $disabled{$1} = 1, next if /--disable-(.*)/;
+  $required{$1} = 1, next if /--require-(.*)/;
+  push @argv, $_;
+}
+
+# NB: use GetOptions() instead of GetOptionsFromArray() for compatibility.
+@ARGV = @argv;
+GetOptions(
+  \%opts,
+  'arch=s',
+  'sym=s',
+  'config=s',
+);
+
+foreach my $opt (qw/arch config/) {
+  if (!defined($opts{$opt})) {
+    warn "--$opt is required!\n";
+    Getopt::Long::HelpMessage('-exit' => 1);
+  }
+}
+
+foreach my $defs_file (@ARGV) {
+  if (!-f $defs_file) {
+    warn "$defs_file: $!\n";
+    Getopt::Long::HelpMessage('-exit' => 1);
+  }
+}
+
+open CONFIG_FILE, $opts{config} or
+  die "Error opening config file '$opts{config}': $!\n";
+
+my %config = ();
+while (<CONFIG_FILE>) {
+  next if !/^(?:CONFIG_|HAVE_)/;
+  chomp;
+  s/\r$//;
+  my @pair = split /=/;
+  $config{$pair[0]} = $pair[1];
+}
+close CONFIG_FILE;
+
+#
+# Routines for the RTCD DSL to call
+#
+sub aom_config($) {
+  return (defined $config{$_[0]}) ? $config{$_[0]} : "";
+}
+
+sub specialize {
+  if (@_ <= 1) {
+    die "'specialize' must be called with a function name and at least one ",
+        "architecture ('C' is implied): \n@_\n";
+  }
+  my $fn=$_[0];
+  shift;
+  foreach my $opt (@_) {
+    eval "\$${fn}_${opt}=${fn}_${opt}";
+  }
+}
+
+sub add_proto {
+  my $fn = splice(@_, -2, 1);
+  $ALL_FUNCS{$fn} = \@_;
+  specialize $fn, "c";
+}
+
+sub require {
+  foreach my $fn (keys %ALL_FUNCS) {
+    foreach my $opt (@_) {
+      my $ofn = eval "\$${fn}_${opt}";
+      next if !$ofn;
+
+      # if we already have a default, then we can disable it, as we know
+      # we can do better.
+      my $best = eval "\$${fn}_default";
+      if ($best) {
+        my $best_ofn = eval "\$${best}";
+        if ($best_ofn && "$best_ofn" ne "$ofn") {
+          eval "\$${best}_link = 'false'";
+        }
+      }
+      eval "\$${fn}_default=${fn}_${opt}";
+      eval "\$${fn}_${opt}_link='true'";
+    }
+  }
+}
+
+sub forward_decls {
+  push @ALL_FORWARD_DECLS, @_;
+}
+
+#
+# Include the user's directives
+#
+foreach my $f (@ARGV) {
+  open FILE, "<", $f or die "cannot open $f: $!\n";
+  my $contents = join('', <FILE>);
+  close FILE;
+  eval $contents or warn "eval failed: $@\n";
+}
+
+#
+# Process the directives according to the command line
+#
+sub process_forward_decls() {
+  foreach (@ALL_FORWARD_DECLS) {
+    $_->();
+  }
+}
+
+sub determine_indirection {
+  aom_config("CONFIG_RUNTIME_CPU_DETECT") eq "yes" or &require(@ALL_ARCHS);
+  foreach my $fn (keys %ALL_FUNCS) {
+    my $n = "";
+    my @val = @{$ALL_FUNCS{$fn}};
+    my $args = pop @val;
+    my $rtyp = "@val";
+    my $dfn = eval "\$${fn}_default";
+    $dfn = eval "\$${dfn}";
+    foreach my $opt (@_) {
+      my $ofn = eval "\$${fn}_${opt}";
+      next if !$ofn;
+      my $link = eval "\$${fn}_${opt}_link";
+      next if $link && $link eq "false";
+      $n .= "x";
+    }
+    if ($n eq "x") {
+      eval "\$${fn}_indirect = 'false'";
+    } else {
+      eval "\$${fn}_indirect = 'true'";
+    }
+  }
+}
+
+sub declare_function_pointers {
+  foreach my $fn (sort keys %ALL_FUNCS) {
+    my @val = @{$ALL_FUNCS{$fn}};
+    my $args = pop @val;
+    my $rtyp = "@val";
+    my $dfn = eval "\$${fn}_default";
+    $dfn = eval "\$${dfn}";
+    foreach my $opt (@_) {
+      my $ofn = eval "\$${fn}_${opt}";
+      next if !$ofn;
+      print "$rtyp ${ofn}($args);\n";
+    }
+    if (eval "\$${fn}_indirect" eq "false") {
+      print "#define ${fn} ${dfn}\n";
+    } else {
+      print "RTCD_EXTERN $rtyp (*${fn})($args);\n";
+    }
+    print "\n";
+  }
+}
+
+sub set_function_pointers {
+  foreach my $fn (sort keys %ALL_FUNCS) {
+    my @val = @{$ALL_FUNCS{$fn}};
+    my $args = pop @val;
+    my $rtyp = "@val";
+    my $dfn = eval "\$${fn}_default";
+    $dfn = eval "\$${dfn}";
+    if (eval "\$${fn}_indirect" eq "true") {
+      print "    $fn = $dfn;\n";
+      foreach my $opt (@_) {
+        my $ofn = eval "\$${fn}_${opt}";
+        next if !$ofn;
+        next if "$ofn" eq "$dfn";
+        my $link = eval "\$${fn}_${opt}_link";
+        next if $link && $link eq "false";
+        my $cond = eval "\$have_${opt}";
+        print "    if (${cond}) $fn = $ofn;\n"
+      }
+    }
+  }
+}
+
+sub filter {
+  my @filtered;
+  foreach (@_) { push @filtered, $_ unless $disabled{$_}; }
+  return @filtered;
+}
+
+#
+# Helper functions for generating the arch specific RTCD files
+#
+sub common_top() {
+  my $include_guard = uc($opts{sym})."_H_";
+  print <<EOF;
+#ifndef ${include_guard}
+#define ${include_guard}
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+EOF
+
+process_forward_decls();
+print <<EOF;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+EOF
+declare_function_pointers("c", @ALL_ARCHS);
+
+print <<EOF;
+void $opts{sym}(void);
+
+EOF
+}
+
+sub common_bottom() {
+  print <<EOF;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
+EOF
+}
+
+sub x86() {
+  determine_indirection("c", @ALL_ARCHS);
+
+  # Assign the helper variable for each enabled extension
+  foreach my $opt (@ALL_ARCHS) {
+    my $opt_uc = uc $opt;
+    eval "\$have_${opt}=\"flags & HAS_${opt_uc}\"";
+  }
+
+  common_top;
+  print <<EOF;
+#ifdef RTCD_C
+#include "aom_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+EOF
+
+  set_function_pointers("c", @ALL_ARCHS);
+
+  print <<EOF;
+}
+#endif
+EOF
+  common_bottom;
+}
+
+sub arm() {
+  determine_indirection("c", @ALL_ARCHS);
+
+  # Assign the helper variable for each enabled extension
+  foreach my $opt (@ALL_ARCHS) {
+    my $opt_uc = uc $opt;
+    # Enable neon assembly based on HAVE_NEON logic instead of adding new
+    # HAVE_NEON_ASM logic
+    if ($opt eq 'neon_asm') { $opt_uc = 'NEON' }
+    eval "\$have_${opt}=\"flags & HAS_${opt_uc}\"";
+  }
+
+  common_top;
+  print <<EOF;
+#include "aom_config.h"
+
+#ifdef RTCD_C
+#include "aom_ports/arm.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = arm_cpu_caps();
+
+    (void)flags;
+
+EOF
+
+  set_function_pointers("c", @ALL_ARCHS);
+
+  print <<EOF;
+}
+#endif
+EOF
+  common_bottom;
+}
+
+sub mips() {
+  determine_indirection("c", @ALL_ARCHS);
+  common_top;
+
+  print <<EOF;
+#include "aom_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+EOF
+
+  set_function_pointers("c", @ALL_ARCHS);
+
+  print <<EOF;
+#if HAVE_DSPR2
+void aom_dsputil_static_init();
+aom_dsputil_static_init();
+#endif
+}
+#endif
+EOF
+  common_bottom;
+}
+
+sub unoptimized() {
+  determine_indirection "c";
+  common_top;
+  print <<EOF;
+#include "aom_config.h"
+
+#ifdef RTCD_C
+static void setup_rtcd_internal(void)
+{
+EOF
+
+  set_function_pointers "c";
+
+  print <<EOF;
+}
+#endif
+EOF
+  common_bottom;
+}
+
+#
+# Main Driver
+#
+
+&require("c");
+if ($opts{arch} eq 'x86') {
+  @ALL_ARCHS = filter(qw/mmx sse sse2 sse3 ssse3 sse4_1 avx avx2/);
+  x86;
+} elsif ($opts{arch} eq 'x86_64') {
+  @ALL_ARCHS = filter(qw/mmx sse sse2 sse3 ssse3 sse4_1 avx avx2/);
+  @REQUIRES = filter(keys %required ? keys %required : qw/mmx sse sse2/);
+  &require(@REQUIRES);
+  x86;
+} elsif ($opts{arch} eq 'mips32' || $opts{arch} eq 'mips64') {
+  @ALL_ARCHS = filter("$opts{arch}");
+  open CONFIG_FILE, $opts{config} or
+    die "Error opening config file '$opts{config}': $!\n";
+  while (<CONFIG_FILE>) {
+    if (/HAVE_DSPR2=yes/) {
+      @ALL_ARCHS = filter("$opts{arch}", qw/dspr2/);
+      last;
+    }
+    if (/HAVE_MSA=yes/) {
+      @ALL_ARCHS = filter("$opts{arch}", qw/msa/);
+      last;
+    }
+  }
+  close CONFIG_FILE;
+  mips;
+} elsif ($opts{arch} eq 'armv6') {
+  @ALL_ARCHS = filter(qw/media/);
+  arm;
+} elsif ($opts{arch} =~ /armv7\w?/) {
+  @ALL_ARCHS = filter(qw/media neon_asm neon/);
+  @REQUIRES = filter(keys %required ? keys %required : qw/media/);
+  &require(@REQUIRES);
+  arm;
+} elsif ($opts{arch} eq 'armv8' || $opts{arch} eq 'arm64' ) {
+  @ALL_ARCHS = filter(qw/neon/);
+  arm;
+} else {
+  unoptimized;
+}
+
+__END__
+
+=head1 NAME
+
+rtcd -
+
+=head1 SYNOPSIS
+
+Usage: rtcd.pl [options] FILE
+
+See 'perldoc rtcd.pl' for more details.
+
+=head1 DESCRIPTION
+
+Reads the Run Time CPU Detections definitions from FILE and generates a
+C header file on stdout.
+
+=head1 OPTIONS
+
+Options:
+  --arch=ARCH       Architecture to generate defs for (required)
+  --disable-EXT     Disable support for EXT extensions
+  --require-EXT     Require support for EXT extensions
+  --sym=SYMBOL      Unique symbol to use for RTCD initialization function
+  --config=FILE     File with CONFIG_FOO=yes lines to parse
diff --git a/third_party/aom/build/make/thumb.pm b/third_party/aom/build/make/thumb.pm
new file mode 100644
index 0000000000..8248694e94
--- /dev/null
+++ b/third_party/aom/build/make/thumb.pm
@@ -0,0 +1,71 @@
+#!/usr/bin/env perl
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+package thumb;
+
+sub FixThumbInstructions($$)
+{
+    my $short_branches = $_[1];
+    my $branch_shift_offset = $short_branches ? 1 : 0;
+
+    # Write additions with shifts, such as "add r10, r11, lsl #8",
+    # in three operand form, "add r10, r10, r11, lsl #8".
+    s/(add\s+)(r\d+),\s*(r\d+),\s*(lsl #\d+)/$1$2, $2, $3, $4/g;
+
+    # Convert additions with a non-constant shift into a sequence
+    # with left shift, addition and a right shift (to restore the
+    # register to the original value). Currently the right shift
+    # isn't necessary in the code base since the values in these
+    # registers aren't used, but doing the shift for consistency.
+    # This converts instructions such as "add r12, r12, r5, lsl r4"
+    # into the sequence "lsl r5, r4", "add r12, r12, r5", "lsr r5, r4".
+    s/^(\s*)(add)(\s+)(r\d+),\s*(r\d+),\s*(r\d+),\s*lsl (r\d+)/$1lsl$3$6, $7\n$1$2$3$4, $5, $6\n$1lsr$3$6, $7/g;
+
+    # Convert loads with right shifts in the indexing into a
+    # sequence of an add, load and sub. This converts
+    # "ldrb r4, [r9, lr, asr #1]" into "add r9, r9, lr, asr #1",
+    # "ldrb r9, [r9]", "sub r9, r9, lr, asr #1".
+    s/^(\s*)(ldrb)(\s+)(r\d+),\s*\[(\w+),\s*(\w+),\s*(asr #\d+)\]/$1add $3$5, $5, $6, $7\n$1$2$3$4, [$5]\n$1sub $3$5, $5, $6, $7/g;
+
+    # Convert register indexing with writeback into a separate add
+    # instruction. This converts "ldrb r12, [r1, r2]!" into
+    # "ldrb r12, [r1, r2]", "add r1, r1, r2".
+    s/^(\s*)(ldrb)(\s+)(r\d+),\s*\[(\w+),\s*(\w+)\]!/$1$2$3$4, [$5, $6]\n$1add $3$5, $6/g;
+
+    # Convert negative register indexing into separate sub/add instructions.
+    # This converts "ldrne r4, [src, -pstep, lsl #1]" into
+    # "subne src, src, pstep, lsl #1", "ldrne r4, [src]",
+    # "addne src, src, pstep, lsl #1". In a couple of cases where
+    # this is used, it's used for two subsequent load instructions,
+    # where a hand-written version of it could merge two subsequent
+    # add and sub instructions.
+    s/^(\s*)((ldr|str|pld)(ne)?)(\s+)(r\d+,\s*)?\[(\w+), -([^\]]+)\]/$1sub$4$5$7, $7, $8\n$1$2$5$6\[$7\]\n$1add$4$5$7, $7, $8/g;
+
+    # Convert register post indexing to a separate add instruction.
+    # This converts "ldrneb r9, [r0], r2" into "ldrneb r9, [r0]",
+    # "addne r0, r0, r2".
+    s/^(\s*)((ldr|str)(ne)?[bhd]?)(\s+)(\w+),(\s*\w+,)?\s*\[(\w+)\],\s*(\w+)/$1$2$5$6,$7 [$8]\n$1add$4$5$8, $8, $9/g;
+
+    # Convert a conditional addition to the pc register into a series of
+    # instructions. This converts "addlt pc, pc, r3, lsl #2" into
+    # "itttt lt", "movlt.n r12, pc", "addlt.w r12, #12",
+    # "addlt.w r12, r12, r3, lsl #2", "movlt.n pc, r12".
+    # This assumes that r12 is free at this point.
+    s/^(\s*)addlt(\s+)pc,\s*pc,\s*(\w+),\s*lsl\s*#(\d+)/$1itttt$2lt\n$1movlt.n$2r12, pc\n$1addlt.w$2r12, #12\n$1addlt.w$2r12, r12, $3, lsl #($4-$branch_shift_offset)\n$1movlt.n$2pc, r12/g;
+
+    # Convert "mov pc, lr" into "bx lr", since the former only works
+    # for switching from arm to thumb (and only in armv7), but not
+    # from thumb to arm.
+    s/mov(\s*)pc\s*,\s*lr/bx$1lr/g;
+}
+
+1;
diff --git a/third_party/aom/build/make/version.sh b/third_party/aom/build/make/version.sh
new file mode 100755
index 0000000000..c0eef9f580
--- /dev/null
+++ b/third_party/aom/build/make/version.sh
@@ -0,0 +1,77 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+
+for opt in "$@"; do
+    optval="${opt#*=}"
+    case "$opt" in
+    --bare) bare=true ;;
+    *) break ;;
+    esac
+    shift
+done
+source_path=${1:-.}
+out_file=${2}
+id=${3:-VERSION_STRING}
+
+git_version_id=""
+if [ -e "${source_path}/.git" ]; then
+    # Source Path is a git working copy. Check for local modifications.
+    # Note that git submodules may have a file as .git, not a directory.
+    export GIT_DIR="${source_path}/.git"
+    git_version_id=`git describe --match=v[0-9]* 2>/dev/null`
+fi
+
+changelog_version=""
+for p in "${source_path}" "${source_path}/.."; do
+    if [ -z "$git_version_id" -a -f "${p}/CHANGELOG" ]; then
+        changelog_version=`head -n1 "${p}/CHANGELOG" | awk '{print $2}'`
+        changelog_version="${changelog_version}"
+        break
+    fi
+done
+version_str="${changelog_version}${git_version_id}"
+bare_version=${version_str#v}
+major_version=${bare_version%%.*}
+bare_version=${bare_version#*.}
+minor_version=${bare_version%%.*}
+bare_version=${bare_version#*.}
+patch_version=${bare_version%%-*}
+bare_version=${bare_version#${patch_version}}
+extra_version=${bare_version##-}
+
+#since they'll be used as integers below make sure they are or force to 0
+for v in major_version minor_version patch_version; do
+    if eval echo \$$v |grep -E -q '[^[:digit:]]'; then
+        eval $v=0
+    fi
+done
+
+if [ ${bare} ]; then
+    echo "${changelog_version}${git_version_id}" > $$.tmp
+else
+    cat<<EOF>$$.tmp
+#define VERSION_MAJOR  $major_version
+#define VERSION_MINOR  $minor_version
+#define VERSION_PATCH  $patch_version
+#define VERSION_EXTRA  "$extra_version"
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define ${id}_NOSP "${version_str}"
+#define ${id}      " ${version_str}"
+EOF
+fi
+if [ -n "$out_file" ]; then
+diff $$.tmp ${out_file} >/dev/null 2>&1 || cat $$.tmp > ${out_file}
+else
+cat $$.tmp
+fi
+rm $$.tmp
diff --git a/third_party/aom/codereview.settings b/third_party/aom/codereview.settings
new file mode 100644
index 0000000000..0f36904869
--- /dev/null
+++ b/third_party/aom/codereview.settings
@@ -0,0 +1,5 @@
+# This file is used by gcl to get repository specific information.
+GERRIT_HOST: aomedia-review.googlesource.com
+GERRIT_PORT: 29418
+GERRIT_SQUASH_UPLOADS: false
+CODE_REVIEW_SERVER: aomedia-review.googlesource.com
diff --git a/third_party/aom/configure b/third_party/aom/configure
new file mode 100755
index 0000000000..83ec996ce8
--- /dev/null
+++ b/third_party/aom/configure
@@ -0,0 +1,865 @@
+#!/bin/sh
+##
+##  configure
+##
+##  This script is the front-end to the build system. It provides a similar
+##  interface to standard configure scripts with some extra bits for dealing
+##  with toolchains that differ from the standard POSIX interface and
+##  for extracting subsets of the source tree. In theory, reusable parts
+##  of this script were intended to live in build/make/configure.sh,
+##  but in practice, the line is pretty blurry.
+##
+##  This build system is based in part on the FFmpeg configure script.
+##
+
+#source_path="`dirname \"$0\"`"
+source_path=${0%/*}
+. "${source_path}/build/make/configure.sh"
+
+show_help(){
+    show_help_pre
+    cat << EOF
+Advanced options:
+  ${toggle_libs}                  libraries
+  ${toggle_examples}              examples
+  ${toggle_analyzer}              analyzer
+  ${toggle_docs}                  documentation
+  ${toggle_unit_tests}            unit tests
+  ${toggle_decode_perf_tests}     build decoder perf tests with unit tests
+  ${toggle_encode_perf_tests}     build encoder perf tests with unit tests
+  --cpu=CPU                       tune for the specified CPU (ARM: cortex-a8, X86: sse3)
+  --libc=PATH                     path to alternate libc
+  --size-limit=WxH                max size to allow in the decoder
+  --as={yasm|nasm|auto}           use specified assembler [auto, yasm preferred]
+  --sdk-path=PATH                 path to root of sdk (android builds only)
+  ${toggle_codec_srcs}            in/exclude codec library source code
+  ${toggle_debug_libs}            in/exclude debug version of libraries
+  ${toggle_static_msvcrt}         use static MSVCRT (VS builds only)
+  ${toggle_highbitdepth}          enable 16-bit generic pixel pipeline (used by high bitdepth profiles)
+  ${toggle_lowbitdepth}           enable 8-bit optimized pixel pipeline
+  ${toggle_av1}                  AV1 codec support
+  ${toggle_internal_stats}        output of encoder internal stats for debug, if supported (encoders)
+  ${toggle_postproc}              postprocessing
+  ${toggle_multithread}           multithreaded encoding and decoding
+  ${toggle_spatial_resampling}    spatial sampling (scaling) support
+  ${toggle_realtime_only}         enable this option while building for real-time encoding
+  ${toggle_onthefly_bitpacking}   enable on-the-fly bitpacking in real-time encoding
+  ${toggle_error_concealment}     enable this option to get a decoder which is able to conceal losses
+  ${toggle_coefficient_range_checking}
+                                  enable decoder to check if intermediate
+                                  transform coefficients are in valid range
+  ${toggle_runtime_cpu_detect}    runtime cpu detection
+  ${toggle_shared}                shared library support
+  ${toggle_static}                static library support
+  ${toggle_small}                 favor smaller size over speed
+  ${toggle_postproc_visualizer}   macro block / block level visualizers
+  ${toggle_webm_io}               enable input from and output to WebM container
+  ${toggle_libyuv}                enable libyuv
+  ${toggle_accounting}            enable bit accounting
+  ${toggle_inspection}            enable bitstream inspection
+
+Codecs:
+  Codecs can be selectively enabled or disabled individually, or by family:
+      --disable-<codec>
+  is equivalent to:
+      --disable-<codec>-encoder
+      --disable-<codec>-decoder
+
+  Codecs available in this distribution:
+EOF
+#restore editor state '
+
+    family="";
+    last_family="";
+    c="";
+    str="";
+    for c in ${CODECS}; do
+        family=${c%_*}
+        if [ "${family}" != "${last_family}" ]; then
+            [ -z "${str}" ] || echo "${str}"
+            str="$(printf '    %10s:' ${family})"
+        fi
+        str="${str} $(printf '%10s' ${c#*_})"
+        last_family=${family}
+    done
+    echo "${str}"
+    show_help_post
+}
+
+##
+## BEGIN APPLICATION SPECIFIC CONFIGURATION
+##
+
+# all_platforms is a list of all supported target platforms. Maintain
+# alphabetically by architecture, generic-gnu last.
+all_platforms="${all_platforms} arm64-darwin-gcc"
+all_platforms="${all_platforms} arm64-linux-gcc"
+all_platforms="${all_platforms} armv6-linux-rvct"
+all_platforms="${all_platforms} armv6-linux-gcc"
+all_platforms="${all_platforms} armv6-none-rvct"
+all_platforms="${all_platforms} armv7-android-gcc"   #neon Cortex-A8
+all_platforms="${all_platforms} armv7-darwin-gcc"    #neon Cortex-A8
+all_platforms="${all_platforms} armv7-linux-rvct"    #neon Cortex-A8
+all_platforms="${all_platforms} armv7-linux-gcc"     #neon Cortex-A8
+all_platforms="${all_platforms} armv7-none-rvct"     #neon Cortex-A8
+all_platforms="${all_platforms} armv7-win32-vs12"
+all_platforms="${all_platforms} armv7-win32-vs14"
+all_platforms="${all_platforms} armv7s-darwin-gcc"
+all_platforms="${all_platforms} armv8-linux-gcc"
+all_platforms="${all_platforms} mips32-linux-gcc"
+all_platforms="${all_platforms} mips64-linux-gcc"
+all_platforms="${all_platforms} sparc-solaris-gcc"
+all_platforms="${all_platforms} x86-android-gcc"
+all_platforms="${all_platforms} x86-darwin8-gcc"
+all_platforms="${all_platforms} x86-darwin8-icc"
+all_platforms="${all_platforms} x86-darwin9-gcc"
+all_platforms="${all_platforms} x86-darwin9-icc"
+all_platforms="${all_platforms} x86-darwin10-gcc"
+all_platforms="${all_platforms} x86-darwin11-gcc"
+all_platforms="${all_platforms} x86-darwin12-gcc"
+all_platforms="${all_platforms} x86-darwin13-gcc"
+all_platforms="${all_platforms} x86-darwin14-gcc"
+all_platforms="${all_platforms} x86-darwin15-gcc"
+all_platforms="${all_platforms} x86-darwin16-gcc"
+all_platforms="${all_platforms} x86-iphonesimulator-gcc"
+all_platforms="${all_platforms} x86-linux-gcc"
+all_platforms="${all_platforms} x86-linux-icc"
+all_platforms="${all_platforms} x86-os2-gcc"
+all_platforms="${all_platforms} x86-solaris-gcc"
+all_platforms="${all_platforms} x86-win32-gcc"
+all_platforms="${all_platforms} x86-win32-vs12"
+all_platforms="${all_platforms} x86-win32-vs14"
+all_platforms="${all_platforms} x86_64-android-gcc"
+all_platforms="${all_platforms} x86_64-darwin9-gcc"
+all_platforms="${all_platforms} x86_64-darwin10-gcc"
+all_platforms="${all_platforms} x86_64-darwin11-gcc"
+all_platforms="${all_platforms} x86_64-darwin12-gcc"
+all_platforms="${all_platforms} x86_64-darwin13-gcc"
+all_platforms="${all_platforms} x86_64-darwin14-gcc"
+all_platforms="${all_platforms} x86_64-darwin15-gcc"
+all_platforms="${all_platforms} x86_64-darwin16-gcc"
+all_platforms="${all_platforms} x86_64-iphonesimulator-gcc"
+all_platforms="${all_platforms} x86_64-linux-gcc"
+all_platforms="${all_platforms} x86_64-linux-icc"
+all_platforms="${all_platforms} x86_64-solaris-gcc"
+all_platforms="${all_platforms} x86_64-win64-gcc"
+all_platforms="${all_platforms} x86_64-win64-vs12"
+all_platforms="${all_platforms} x86_64-win64-vs14"
+all_platforms="${all_platforms} generic-gnu"
+
+# all_targets is a list of all targets that can be configured
+# note that these should be in dependency order for now.
+all_targets="libs examples docs"
+
+# all targets available are enabled, by default.
+for t in ${all_targets}; do
+    [ -f "${source_path}/${t}.mk" ] && enable_feature ${t}
+done
+
+if ! perl --version >/dev/null; then
+    die "Perl is required to build"
+fi
+
+
+if [ "`cd \"${source_path}\" && pwd`" != "`pwd`" ]; then
+  # test to see if source_path already configured
+  if [ -f "${source_path}/aom_config.h" ]; then
+    die "source directory already configured; run 'make distclean' there first"
+  fi
+fi
+
+# check installed doxygen version
+doxy_version=$(doxygen --version 2>/dev/null)
+doxy_major=${doxy_version%%.*}
+if [ ${doxy_major:-0} -ge 1 ]; then
+    doxy_version=${doxy_version#*.}
+    doxy_minor=${doxy_version%%.*}
+    doxy_patch=${doxy_version##*.}
+
+    [ $doxy_major -gt 1 ] && enable_feature doxygen
+    [ $doxy_minor -gt 5 ] && enable_feature doxygen
+    [ $doxy_minor -eq 5 ] && [ $doxy_patch -ge 3 ] && enable_feature doxygen
+fi
+
+# disable codecs when their source directory does not exist
+[ -d "${source_path}/av1" ] || disable_codec av1
+
+# install everything except the sources, by default. sources will have
+# to be enabled when doing dist builds, since that's no longer a common
+# case.
+enabled doxygen && enable_feature install_docs
+enable_feature install_bins
+enable_feature install_libs
+
+enable_feature static
+enable_feature optimizations
+enable_feature dependency_tracking
+enable_feature spatial_resampling
+enable_feature multithread
+enable_feature os_support
+
+CODECS="
+    av1_encoder
+    av1_decoder
+"
+CODEC_FAMILIES="
+    av1
+"
+
+ARCH_LIST="
+    arm
+    mips
+    x86
+    x86_64
+"
+ARCH_EXT_LIST_X86="
+    mmx
+    sse
+    sse2
+    sse3
+    ssse3
+    sse4_1
+    avx
+    avx2
+"
+ARCH_EXT_LIST="
+    edsp
+    media
+    neon
+    neon_asm
+
+    mips32
+    dspr2
+    msa
+    mips64
+
+    ${ARCH_EXT_LIST_X86}
+"
+HAVE_LIST="
+    ${ARCH_EXT_LIST}
+    aom_ports
+    fexcept
+    pthread_h
+    unistd_h
+    wxwidgets
+"
+EXPERIMENT_LIST="
+    fp_mb_stats
+    cdef
+    var_tx
+    rect_tx
+    ref_mv
+    tpl_mv
+    dual_filter
+    convolve_round
+    compound_round
+    ext_tx
+    tx64x64
+    sub8x8_mc
+    ext_intra
+    intra_interp
+    filter_intra
+    intrabc
+    ext_inter
+    interintra
+    wedge
+    compound_segment
+    ext_refs
+    global_motion
+    new_quant
+    supertx
+    ans
+    ec_multisymbol
+    new_tokenset
+    loop_restoration
+    ext_partition
+    ext_partition_types
+    unpoison_partition_ctx
+    ext_tile
+    motion_var
+    ncobmc
+    warped_motion
+    q_adapt_probs
+    subframe_prob_update
+    bitstream_debug
+    alt_intra
+    palette
+    palette_delta_encoding
+    daala_ec
+    rawbits
+    ec_smallmul
+    pvq
+    cfl
+    xiphrc
+    cb4x4
+    chroma_2x2
+    chroma_sub8x8
+    frame_size
+    delta_q
+    ext_delta_q
+    adapt_scan
+    filter_7bit
+    parallel_deblocking
+    parallel_deblocking_15tap
+    loopfiltering_across_tiles
+    tile_groups
+    ec_adapt
+    tempmv_signaling
+    rd_debug
+    reference_buffer
+    coef_interleave
+    entropy_stats
+    masked_tx
+    dependent_horztiles
+    daala_dist
+    tripred
+    palette_throughput
+    ref_adapt
+    lv_map
+    txk_sel
+    mv_compress
+    frame_superres
+    new_multisymbol
+    compound_singleref
+    aom_qm
+    lowdelay_compound
+"
+CONFIG_LIST="
+    dependency_tracking
+    external_build
+    install_docs
+    install_bins
+    install_libs
+    install_srcs
+    debug
+    gprof
+    gcov
+    rvct
+    gcc
+    msvs
+    pic
+    big_endian
+
+    codec_srcs
+    debug_libs
+
+    runtime_cpu_detect
+    postproc
+    multithread
+    internal_stats
+    ${CODECS}
+    ${CODEC_FAMILIES}
+    encoders
+    decoders
+    static_msvcrt
+    spatial_resampling
+    realtime_only
+    onthefly_bitpacking
+    error_concealment
+    shared
+    static
+    small
+    postproc_visualizer
+    os_support
+    unit_tests
+    webm_io
+    libyuv
+    accounting
+    inspection
+    decode_perf_tests
+    encode_perf_tests
+    coefficient_range_checking
+    lowbitdepth
+    highbitdepth
+    experimental
+    size_limit
+    ${EXPERIMENT_LIST}
+    analyzer
+"
+CMDLINE_SELECT="
+    dependency_tracking
+    external_build
+    extra_warnings
+    werror
+    install_docs
+    install_bins
+    install_libs
+    install_srcs
+    debug
+    gprof
+    gcov
+    pic
+    optimizations
+    ccache
+    runtime_cpu_detect
+    thumb
+
+    libs
+    examples
+    analyzer
+    docs
+    libc
+    as
+    size_limit
+    codec_srcs
+    debug_libs
+
+    postproc
+    multithread
+    internal_stats
+    ${CODECS}
+    ${CODEC_FAMILIES}
+    static_msvcrt
+    spatial_resampling
+    realtime_only
+    onthefly_bitpacking
+    error_concealment
+    shared
+    static
+    small
+    postproc_visualizer
+    unit_tests
+    webm_io
+    libyuv
+    accounting
+    inspection
+    decode_perf_tests
+    encode_perf_tests
+    coefficient_range_checking
+    lowbitdepth
+    aom_highbitdepth
+    highbitdepth
+    experimental
+"
+
+process_cmdline() {
+    for opt do
+        optval="${opt#*=}"
+        case "$opt" in
+        --disable-codecs)
+          for c in ${CODEC_FAMILIES}; do disable_codec $c; done
+          ;;
+        --enable-?*|--disable-?*)
+        eval `echo "$opt" | sed 's/--/action=/;s/-/ option=/;s/-/_/g'`
+        if is_in ${option} ${EXPERIMENT_LIST}; then
+            if enabled experimental; then
+                ${action}_feature $option
+            else
+                log_echo "Ignoring $opt -- not in experimental mode."
+            fi
+        elif is_in ${option} "${CODECS} ${CODEC_FAMILIES}"; then
+            ${action}_codec ${option}
+        else
+            process_common_cmdline $opt
+        fi
+        ;;
+        *) process_common_cmdline "$opt"
+        ;;
+        esac
+    done
+}
+
+post_process_cmdline() {
+    c=""
+
+    # Enable all detected codecs, if they haven't been disabled
+    for c in ${CODECS}; do soft_enable $c; done
+
+    # Enable the codec family if any component of that family is enabled
+    for c in ${CODECS}; do
+        enabled $c && enable_feature ${c%_*}
+    done
+
+    # Set the {en,de}coders variable if any algorithm in that class is enabled
+    for c in ${CODECS}; do
+        enabled ${c} && enable_feature ${c##*_}s
+    done
+
+    # Enable adopted experiments by default
+    soft_enable ref_mv
+    soft_enable cb4x4
+    soft_enable filter_7bit
+    soft_enable reference_buffer
+    soft_enable delta_q
+    soft_enable tile_groups
+    soft_enable rect_tx
+    soft_enable global_motion
+    soft_enable cdef
+    soft_enable ec_adapt
+    soft_enable new_tokenset
+    soft_enable ext_intra
+    soft_enable mv_compress
+    soft_enable ext_refs
+    soft_enable dual_filter
+    soft_enable motion_var
+    soft_enable warped_motion
+    soft_enable ext_delta_q
+
+    # Backwards/jenkins compatibility with --enable-aom-highbitdepth
+    enabled aom_highbitdepth && enable_feature highbitdepth
+
+    # Enable low-bitdepth pixel pipeline by default
+    soft_enable lowbitdepth
+
+    ! enabled ans && soft_enable daala_ec
+    soft_enable ec_multisymbol
+    soft_enable palette
+    soft_enable alt_intra
+    soft_enable palette_throughput
+
+    # Fix up experiment dependencies
+    enabled pvq && enable_feature ec_adapt
+    enabled ec_adapt && enable_feature ec_multisymbol
+    enabled new_tokenset && enable_feature ec_multisymbol
+    enabled new_multisymbol && enable_feature ec_multisymbol
+    enabled ec_multisymbol && ! enabled ans && soft_enable daala_ec
+    enabled ec_multisymbol && ! enabled daala_ec && soft_enable ans
+    enabled palette_throughput && soft_enable palette
+    enabled ext_delta_q && soft_enable delta_q
+    enabled txk_sel && soft_enable lv_map
+    enabled compound_round && soft_enable convolve_round
+    if enabled rawbits && ! enabled daala_ec; then
+      log_echo "rawbits requires daala_ec, so disabling rawbits"
+      disable_feature rawbits
+    fi
+    if enabled ec_smallmul && ! enabled daala_ec; then
+      log_echo "ec_smallmul requires daala_ec, so disabling ec_smallmul"
+      disable_feature ec_smallmul
+    fi
+    if enabled ext_tile; then
+      log_echo "ext_tile not compatible with reference_buffer, so"
+      log_echo "disabling reference_buffer"
+      disable_feature reference_buffer
+    fi
+    if enabled ext_tile; then
+      log_echo "ext_tile not compatible with tile_groups, so"
+      log_echo "disabling tile_groups"
+      disable_feature tile_groups
+    fi
+    # Enable accounting and inspection when building the analyzer
+    if enabled analyzer; then
+      soft_enable accounting
+      soft_enable inspection
+    fi
+}
+
+process_targets() {
+    enabled child || write_common_config_banner
+    write_common_target_config_h ${BUILD_PFX}aom_config.h
+    write_common_config_targets
+
+    # Calculate the default distribution name, based on the enabled features
+    cf=""
+    DIST_DIR=aom
+    for cf in $CODEC_FAMILIES; do
+        if enabled ${cf}_encoder && enabled ${cf}_decoder; then
+            DIST_DIR="${DIST_DIR}-${cf}"
+        elif enabled ${cf}_encoder; then
+            DIST_DIR="${DIST_DIR}-${cf}cx"
+        elif enabled ${cf}_decoder; then
+            DIST_DIR="${DIST_DIR}-${cf}dx"
+        fi
+    done
+    enabled debug_libs && DIST_DIR="${DIST_DIR}-debug"
+    enabled codec_srcs && DIST_DIR="${DIST_DIR}-src"
+    ! enabled postproc && DIST_DIR="${DIST_DIR}-nopost"
+    ! enabled multithread && DIST_DIR="${DIST_DIR}-nomt"
+    ! enabled install_docs && DIST_DIR="${DIST_DIR}-nodocs"
+    DIST_DIR="${DIST_DIR}-${tgt_isa}-${tgt_os}"
+    case "${tgt_os}" in
+    win*) enabled static_msvcrt && DIST_DIR="${DIST_DIR}mt" || DIST_DIR="${DIST_DIR}md"
+          DIST_DIR="${DIST_DIR}-${tgt_cc}"
+          ;;
+    esac
+    if [ -f "${source_path}/build/make/version.sh" ]; then
+        ver=`"$source_path/build/make/version.sh" --bare "$source_path"`
+        DIST_DIR="${DIST_DIR}-${ver}"
+        VERSION_STRING=${ver}
+        ver=${ver%%-*}
+        VERSION_PATCH=${ver##*.}
+        ver=${ver%.*}
+        VERSION_MINOR=${ver##*.}
+        ver=${ver#v}
+        VERSION_MAJOR=${ver%.*}
+    fi
+    enabled child || cat <<EOF >> config.mk
+
+PREFIX=${prefix}
+ifeq (\$(MAKECMDGOALS),dist)
+DIST_DIR?=${DIST_DIR}
+else
+DIST_DIR?=\$(DESTDIR)${prefix}
+endif
+LIBSUBDIR=${libdir##${prefix}/}
+
+VERSION_STRING=${VERSION_STRING}
+
+VERSION_MAJOR=${VERSION_MAJOR}
+VERSION_MINOR=${VERSION_MINOR}
+VERSION_PATCH=${VERSION_PATCH}
+
+CONFIGURE_ARGS=${CONFIGURE_ARGS}
+EOF
+    enabled child || echo "CONFIGURE_ARGS?=${CONFIGURE_ARGS}" >> config.mk
+
+    #
+    # Write makefiles for all enabled targets
+    #
+    for tgt in libs examples docs solution; do
+        tgt_fn="$tgt-$toolchain.mk"
+
+        if enabled $tgt; then
+            echo "Creating makefiles for ${toolchain} ${tgt}"
+            write_common_target_config_mk $tgt_fn ${BUILD_PFX}aom_config.h
+            #write_${tgt}_config
+        fi
+    done
+
+}
+
+process_detect() {
+    if enabled shared; then
+        # Can only build shared libs on a subset of platforms. Doing this check
+        # here rather than at option parse time because the target auto-detect
+        # magic happens after the command line has been parsed.
+        case "${tgt_os}" in
+        linux|os2|darwin*|iphonesimulator*)
+            # Supported platforms
+            ;;
+        *)
+            if enabled gnu; then
+                echo "--enable-shared is only supported on ELF; assuming this is OK"
+            else
+                die "--enable-shared only supported on ELF, OS/2, and Darwin for now"
+            fi
+            ;;
+        esac
+    fi
+    if [ -z "$CC" ] || enabled external_build; then
+        echo "Bypassing toolchain for environment detection."
+        enable_feature external_build
+        check_header() {
+            log fake_check_header "$@"
+            header=$1
+            shift
+            var=`echo $header | sed 's/[^A-Za-z0-9_]/_/g'`
+            disable_feature $var
+            # Headers common to all environments
+            case $header in
+                stdio.h)
+                    true;
+                ;;
+                *)
+                    result=false
+                    for d in "$@"; do
+                        [ -f "${d##-I}/$header" ] && result=true && break
+                    done
+                    ${result:-true}
+            esac && enable_feature $var
+
+            # Specialize windows and POSIX environments.
+            case $toolchain in
+                *-win*-*)
+                    # Don't check for any headers in Windows builds.
+                    false
+                ;;
+                *)
+                    case $header in
+                        pthread.h) true;;
+                        unistd.h) true;;
+                        *) false;;
+                    esac && enable_feature $var
+            esac
+            enabled $var
+        }
+        check_ld() {
+            true
+        }
+    fi
+    check_header stdio.h || die "Unable to invoke compiler: ${CC} ${CFLAGS}"
+    check_ld <<EOF || die "Toolchain is unable to link executables"
+int main(void) {return 0;}
+EOF
+    # check system headers
+    check_header pthread.h
+    check_header unistd.h # for sysconf(3) and friends.
+
+    check_header aom/aom_integer.h -I${source_path} && enable_feature aom_ports
+
+    check_ld <<EOF && enable_feature fexcept
+#define _GNU_SOURCE
+#include <fenv.h>
+int main(void) { (void)feenableexcept(FE_DIVBYZERO | FE_INVALID); return 0; }
+EOF
+}
+
+process_toolchain() {
+    process_common_toolchain
+
+    # Enable some useful compiler flags
+    if enabled gcc; then
+        enabled werror && check_add_cflags -Werror
+        check_add_cflags -Wall
+        check_add_cflags -Wdisabled-optimization
+        check_add_cflags -Wfloat-conversion
+        check_add_cflags -Wpointer-arith
+        check_add_cflags -Wtype-limits
+        check_add_cflags -Wvla
+        check_add_cflags -Wimplicit-function-declaration
+        check_add_cflags -Wuninitialized
+        check_add_cflags -Wunused
+        check_add_cflags -Wsign-compare
+        check_add_cflags -Wlogical-op
+        check_add_cflags -Wstack-usage=320000
+        # Enabling the following warning (in combination with -Wunused above)
+        # for C++ generates errors in third_party code including googletest and
+        # libyuv. So enable it only for C code.
+        check_cflags "-Wextra" && add_cflags_only "-Wextra"
+        # Enabling the following warning for C++ generates some useless warnings
+        # about some function parameters shadowing class member function names.
+        # So, only enable this warning for C code.
+        check_cflags "-Wshadow" && add_cflags_only "-Wshadow"
+        if enabled mips || [ -z "${INLINE}" ]; then
+          enabled extra_warnings || check_add_cflags -Wno-unused-function
+        fi
+        # gtest makes heavy use of undefined pre-processor symbols
+        check_cflags "-Wundef" && add_cflags_only "-Wundef"
+        # Avoid this warning for third_party C++ sources. Some reorganization
+        # would be needed to apply this only to test/*.cc.
+        check_cflags -Wshorten-64-to-32 && add_cflags_only -Wshorten-64-to-32
+    fi
+
+    if enabled icc; then
+        enabled werror && check_add_cflags -Werror
+        check_add_cflags -Wall
+        check_add_cflags -Wpointer-arith
+
+        # ICC has a number of floating point optimizations that we disable
+        # in favor of deterministic output WRT to other compilers
+        add_cflags -fp-model precise
+    fi
+
+    if enabled analyzer; then
+      soft_enable wxwidgets
+      if ! wx-config --version > /dev/null; then
+        die "Couldn't find wx-config"
+      fi
+
+      add_cxxflags_only $(wx-config --cppflags)
+      add_extralibs $(wx-config --libs)
+    fi
+
+    # Enable extra, harmless warnings. These might provide additional insight
+    # to what the compiler is doing and why, but in general, but they shouldn't
+    # be treated as fatal, even if we're treating warnings as errors.
+    GCC_EXTRA_WARNINGS="
+        -Wdisabled-optimization
+        -Winline
+    "
+    enabled gcc && EXTRA_WARNINGS="${GCC_EXTRA_WARNINGS}"
+    RVCT_EXTRA_WARNINGS="
+        --remarks
+    "
+    enabled rvct && EXTRA_WARNINGS="${RVCT_EXTRA_WARNINGS}"
+    if enabled extra_warnings; then
+        for w in ${EXTRA_WARNINGS}; do
+            check_add_cflags ${w}
+            enabled gcc && enabled werror && check_add_cflags -Wno-error=${w}
+        done
+    fi
+
+    # ccache only really works on gcc toolchains
+    enabled gcc || soft_disable ccache
+
+    # Enable the postbuild target if building for visual studio.
+    case "$tgt_cc" in
+        vs*) enable_feature msvs
+             enable_feature solution
+             vs_version=${tgt_cc##vs}
+             VCPROJ_SFX=vcxproj
+             gen_vcproj_cmd=${source_path}/build/make/gen_msvs_vcxproj.sh
+             enabled werror && gen_vcproj_cmd="${gen_vcproj_cmd} --enable-werror"
+             all_targets="${all_targets} solution"
+             INLINE="__forceinline"
+        ;;
+    esac
+
+    # Other toolchain specific defaults
+    case $toolchain in x86*) soft_enable postproc;; esac
+
+    if enabled postproc_visualizer; then
+        enabled postproc || die "postproc_visualizer requires postproc to be enabled"
+    fi
+
+    # Enable unit tests by default if we have a working C++ compiler.
+    case "$toolchain" in
+        *-vs*)
+            soft_enable unit_tests
+            soft_enable webm_io
+            soft_enable libyuv
+        ;;
+        *-android-*)
+            soft_enable webm_io
+            soft_enable libyuv
+            # GTestLog must be modified to use Android logging utilities.
+        ;;
+        *-darwin-*)
+            # iOS/ARM builds do not work with gtest. This does not match
+            # x86 targets.
+        ;;
+        *-iphonesimulator-*)
+            soft_enable webm_io
+            soft_enable libyuv
+        ;;
+        *-win*)
+            # Some mingw toolchains don't have pthread available by default.
+            # Treat these more like visual studio where threading in gtest
+            # would be disabled for the same reason.
+            check_cxx "$@" <<EOF && soft_enable unit_tests
+int z;
+EOF
+            check_cxx "$@" <<EOF && soft_enable webm_io
+int z;
+EOF
+            check_cxx "$@" <<EOF && soft_enable libyuv
+int z;
+EOF
+        ;;
+        *)
+            enabled pthread_h && check_cxx "$@" <<EOF && soft_enable unit_tests
+int z;
+EOF
+            check_cxx "$@" <<EOF && soft_enable webm_io
+int z;
+EOF
+            check_cxx "$@" <<EOF && soft_enable libyuv
+int z;
+EOF
+        ;;
+    esac
+    # libwebm needs to be linked with C++ standard library
+    enabled webm_io && LD=${CXX}
+
+    # append any user defined extra cflags
+    if [ -n "${extra_cflags}" ] ; then
+        check_add_cflags ${extra_cflags} || \
+        die "Requested extra CFLAGS '${extra_cflags}' not supported by compiler"
+    fi
+    if [ -n "${extra_cxxflags}" ]; then
+        check_add_cxxflags ${extra_cxxflags} || \
+        die "Requested extra CXXFLAGS '${extra_cxxflags}' not supported by compiler"
+    fi
+}
+
+
+##
+## END APPLICATION SPECIFIC CONFIGURATION
+##
+CONFIGURE_ARGS="$@"
+process "$@"
+print_webm_license ${BUILD_PFX}aom_config.c "/*" " */"
+cat <<EOF >> ${BUILD_PFX}aom_config.c
+#include "aom/aom_codec.h"
+static const char* const cfg = "$CONFIGURE_ARGS";
+const char *aom_codec_build_config(void) {return cfg;}
+EOF
diff --git a/third_party/aom/docs.mk b/third_party/aom/docs.mk
new file mode 100644
index 0000000000..0dfc65b75f
--- /dev/null
+++ b/third_party/aom/docs.mk
@@ -0,0 +1,50 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+
+INSTALL_MAPS += docs/%    docs/%
+INSTALL_MAPS += src/%     %
+INSTALL_MAPS += %         %
+
+# Static documentation authored in doxygen
+CODEC_DOX :=    mainpage.dox \
+		keywords.dox \
+		usage.dox \
+		usage_cx.dox \
+		usage_dx.dox \
+
+# Other doxy files sourced in Markdown
+TXT_DOX = $(call enabled,TXT_DOX)
+
+EXAMPLE_PATH += $(SRC_PATH_BARE) #for CHANGELOG, README, etc
+EXAMPLE_PATH += $(SRC_PATH_BARE)/examples
+
+doxyfile: $(if $(findstring examples, $(ALL_TARGETS)),examples.doxy)
+doxyfile: libs.doxy_template libs.doxy
+	@echo "    [CREATE] $@"
+	@cat $^ > $@
+	@echo "STRIP_FROM_PATH += $(SRC_PATH_BARE) $(BUILD_ROOT)" >> $@
+	@echo "INPUT += $(addprefix $(SRC_PATH_BARE)/,$(CODEC_DOX))" >> $@;
+	@echo "INPUT += $(TXT_DOX)" >> $@;
+	@echo "EXAMPLE_PATH += $(EXAMPLE_PATH)" >> $@
+
+CLEAN-OBJS += doxyfile $(wildcard docs/html/*)
+docs/html/index.html: doxyfile $(CODEC_DOX) $(TXT_DOX)
+	@echo "    [DOXYGEN] $<"
+	@doxygen $<
+DOCS-yes += docs/html/index.html
+
+DIST-DOCS-yes = $(wildcard docs/html/*)
+DIST-DOCS-$(CONFIG_CODEC_SRCS) += $(addprefix src/,$(CODEC_DOX))
+DIST-DOCS-$(CONFIG_CODEC_SRCS) += src/libs.doxy_template
+DIST-DOCS-yes                  += CHANGELOG
+DIST-DOCS-yes                  += README
diff --git a/third_party/aom/examples.mk b/third_party/aom/examples.mk
new file mode 100644
index 0000000000..95206924ec
--- /dev/null
+++ b/third_party/aom/examples.mk
@@ -0,0 +1,383 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+LIBYUV_SRCS +=  third_party/libyuv/include/libyuv/basic_types.h  \
+                third_party/libyuv/include/libyuv/convert.h \
+                third_party/libyuv/include/libyuv/convert_argb.h \
+                third_party/libyuv/include/libyuv/convert_from.h \
+                third_party/libyuv/include/libyuv/cpu_id.h  \
+                third_party/libyuv/include/libyuv/planar_functions.h  \
+                third_party/libyuv/include/libyuv/rotate.h  \
+                third_party/libyuv/include/libyuv/row.h  \
+                third_party/libyuv/include/libyuv/scale.h  \
+                third_party/libyuv/include/libyuv/scale_row.h  \
+                third_party/libyuv/source/cpu_id.cc \
+                third_party/libyuv/source/planar_functions.cc \
+                third_party/libyuv/source/row_any.cc \
+                third_party/libyuv/source/row_common.cc \
+                third_party/libyuv/source/row_gcc.cc \
+                third_party/libyuv/source/row_mips.cc \
+                third_party/libyuv/source/row_neon.cc \
+                third_party/libyuv/source/row_neon64.cc \
+                third_party/libyuv/source/row_win.cc \
+                third_party/libyuv/source/scale.cc \
+                third_party/libyuv/source/scale_any.cc \
+                third_party/libyuv/source/scale_common.cc \
+                third_party/libyuv/source/scale_gcc.cc \
+                third_party/libyuv/source/scale_mips.cc \
+                third_party/libyuv/source/scale_neon.cc \
+                third_party/libyuv/source/scale_neon64.cc \
+                third_party/libyuv/source/scale_win.cc \
+
+LIBWEBM_COMMON_SRCS += third_party/libwebm/common/hdr_util.cc \
+                       third_party/libwebm/common/hdr_util.h \
+                       third_party/libwebm/common/webmids.h
+
+LIBWEBM_MUXER_SRCS += third_party/libwebm/mkvmuxer/mkvmuxer.cc \
+                      third_party/libwebm/mkvmuxer/mkvmuxerutil.cc \
+                      third_party/libwebm/mkvmuxer/mkvwriter.cc \
+                      third_party/libwebm/mkvmuxer/mkvmuxer.h \
+                      third_party/libwebm/mkvmuxer/mkvmuxertypes.h \
+                      third_party/libwebm/mkvmuxer/mkvmuxerutil.h \
+                      third_party/libwebm/mkvparser/mkvparser.h \
+                      third_party/libwebm/mkvmuxer/mkvwriter.h
+
+LIBWEBM_PARSER_SRCS = third_party/libwebm/mkvparser/mkvparser.cc \
+                      third_party/libwebm/mkvparser/mkvreader.cc \
+                      third_party/libwebm/mkvparser/mkvparser.h \
+                      third_party/libwebm/mkvparser/mkvreader.h
+
+# Add compile flags and include path for libwebm sources.
+ifeq ($(CONFIG_WEBM_IO),yes)
+  CXXFLAGS     += -D__STDC_CONSTANT_MACROS -D__STDC_LIMIT_MACROS
+  INC_PATH-yes += $(SRC_PATH_BARE)/third_party/libwebm
+endif
+
+# List of examples to build. UTILS are tools meant for distribution
+# while EXAMPLES demonstrate specific portions of the API.
+UTILS-$(CONFIG_DECODERS)    += aomdec.c
+aomdec.SRCS                 += md5_utils.c md5_utils.h
+aomdec.SRCS                 += aom_ports/mem_ops.h
+aomdec.SRCS                 += aom_ports/mem_ops_aligned.h
+aomdec.SRCS                 += aom_ports/msvc.h
+aomdec.SRCS                 += aom_ports/aom_timer.h
+aomdec.SRCS                 += aom/aom_integer.h
+aomdec.SRCS                 += args.c args.h
+aomdec.SRCS                 += ivfdec.c ivfdec.h
+aomdec.SRCS                 += tools_common.c tools_common.h
+aomdec.SRCS                 += y4menc.c y4menc.h
+ifeq ($(CONFIG_LIBYUV),yes)
+  aomdec.SRCS                 += $(LIBYUV_SRCS)
+endif
+ifeq ($(CONFIG_WEBM_IO),yes)
+  aomdec.SRCS                 += $(LIBWEBM_COMMON_SRCS)
+  aomdec.SRCS                 += $(LIBWEBM_MUXER_SRCS)
+  aomdec.SRCS                 += $(LIBWEBM_PARSER_SRCS)
+  aomdec.SRCS                 += webmdec.cc webmdec.h
+endif
+aomdec.GUID                  = BA5FE66F-38DD-E034-F542-B1578C5FB950
+aomdec.DESCRIPTION           = Full featured decoder
+UTILS-$(CONFIG_ENCODERS)    += aomenc.c
+aomenc.SRCS                 += args.c args.h y4minput.c y4minput.h aomenc.h
+aomenc.SRCS                 += ivfdec.c ivfdec.h
+aomenc.SRCS                 += ivfenc.c ivfenc.h
+aomenc.SRCS                 += rate_hist.c rate_hist.h
+aomenc.SRCS                 += tools_common.c tools_common.h
+aomenc.SRCS                 += examples/encoder_util.h examples/encoder_util.c
+aomenc.SRCS                 += warnings.c warnings.h
+aomenc.SRCS                 += aom_ports/mem_ops.h
+aomenc.SRCS                 += aom_ports/mem_ops_aligned.h
+aomenc.SRCS                 += aom_ports/msvc.h
+aomenc.SRCS                 += aom_ports/aom_timer.h
+aomenc.SRCS                 += aomstats.c aomstats.h
+ifeq ($(CONFIG_LIBYUV),yes)
+  aomenc.SRCS                 += $(LIBYUV_SRCS)
+endif
+ifeq ($(CONFIG_WEBM_IO),yes)
+  aomenc.SRCS                 += $(LIBWEBM_COMMON_SRCS)
+  aomenc.SRCS                 += $(LIBWEBM_MUXER_SRCS)
+  aomenc.SRCS                 += $(LIBWEBM_PARSER_SRCS)
+  aomenc.SRCS                 += webmenc.cc webmenc.h
+endif
+aomenc.GUID                  = 548DEC74-7A15-4B2B-AFC3-AA102E7C25C1
+aomenc.DESCRIPTION           = Full featured encoder
+
+ifeq ($(CONFIG_ANALYZER),yes)
+  EXAMPLES-$(CONFIG_DECODERS)        += analyzer.cc
+  analyzer.GUID                       = 83827a8c-e3c3-4b19-8832-0cfc206c4496
+  analyzer.SRCS                      += ivfdec.h ivfdec.c
+  analyzer.SRCS                      += av1/decoder/inspection.h
+  analyzer.SRCS                      += av1/decoder/inspection.c
+  analyzer.SRCS                      += video_reader.h video_reader.c
+  analyzer.SRCS                      += tools_common.h tools_common.c
+  analyzer.DESCRIPTION                = Bitstream analyzer
+endif
+
+ifeq ($(CONFIG_INSPECTION),yes)
+EXAMPLES-$(CONFIG_DECODERS) += inspect.c
+inspect.GUID                 = FA46A420-3356-441F-B0FD-60AA1345C181
+inspect.SRCS                += ivfdec.h ivfdec.c
+inspect.SRCS                += args.c args.h
+inspect.SRCS                += tools_common.h tools_common.c
+inspect.SRCS                += video_common.h
+inspect.SRCS                += video_reader.h video_reader.c
+inspect.SRCS                += aom_ports/mem_ops.h
+inspect.SRCS                += aom_ports/mem_ops_aligned.h
+inspect.SRCS                += aom_ports/msvc.h
+inspect.DESCRIPTION          = Dump inspection data
+endif
+
+EXAMPLES-$(CONFIG_DECODERS)        += simple_decoder.c
+simple_decoder.GUID                 = D3BBF1E9-2427-450D-BBFF-B2843C1D44CC
+simple_decoder.SRCS                += ivfdec.h ivfdec.c
+simple_decoder.SRCS                += tools_common.h tools_common.c
+simple_decoder.SRCS                += video_common.h
+simple_decoder.SRCS                += video_reader.h video_reader.c
+simple_decoder.SRCS                += aom_ports/mem_ops.h
+simple_decoder.SRCS                += aom_ports/mem_ops_aligned.h
+simple_decoder.SRCS                += aom_ports/msvc.h
+simple_decoder.DESCRIPTION          = Simplified decoder loop
+EXAMPLES-$(CONFIG_DECODERS)        += decode_to_md5.c
+decode_to_md5.SRCS                 += md5_utils.h md5_utils.c
+decode_to_md5.SRCS                 += ivfdec.h ivfdec.c
+decode_to_md5.SRCS                 += tools_common.h tools_common.c
+decode_to_md5.SRCS                 += video_common.h
+decode_to_md5.SRCS                 += video_reader.h video_reader.c
+decode_to_md5.SRCS                 += aom_ports/mem_ops.h
+decode_to_md5.SRCS                 += aom_ports/mem_ops_aligned.h
+decode_to_md5.SRCS                 += aom_ports/msvc.h
+decode_to_md5.GUID                  = 59120B9B-2735-4BFE-B022-146CA340FE42
+decode_to_md5.DESCRIPTION           = Frame by frame MD5 checksum
+EXAMPLES-$(CONFIG_ENCODERS)     += simple_encoder.c
+simple_encoder.SRCS             += ivfenc.h ivfenc.c
+simple_encoder.SRCS             += tools_common.h tools_common.c
+simple_encoder.SRCS             += video_common.h
+simple_encoder.SRCS             += video_writer.h video_writer.c
+simple_encoder.SRCS             += aom_ports/msvc.h
+simple_encoder.GUID              = 4607D299-8A71-4D2C-9B1D-071899B6FBFD
+simple_encoder.DESCRIPTION       = Simplified encoder loop
+EXAMPLES-$(CONFIG_AV1_ENCODER)  += lossless_encoder.c
+lossless_encoder.SRCS           += ivfenc.h ivfenc.c
+lossless_encoder.SRCS           += tools_common.h tools_common.c
+lossless_encoder.SRCS           += video_common.h
+lossless_encoder.SRCS           += video_writer.h video_writer.c
+lossless_encoder.SRCS           += aom_ports/msvc.h
+lossless_encoder.GUID            = B63C7C88-5348-46DC-A5A6-CC151EF93366
+lossless_encoder.DESCRIPTION     = Simplified lossless encoder
+EXAMPLES-$(CONFIG_ENCODERS)     += twopass_encoder.c
+twopass_encoder.SRCS            += ivfenc.h ivfenc.c
+twopass_encoder.SRCS            += tools_common.h tools_common.c
+twopass_encoder.SRCS            += video_common.h
+twopass_encoder.SRCS            += video_writer.h video_writer.c
+twopass_encoder.SRCS            += aom_ports/msvc.h
+twopass_encoder.GUID             = 73494FA6-4AF9-4763-8FBB-265C92402FD8
+twopass_encoder.DESCRIPTION      = Two-pass encoder loop
+EXAMPLES-$(CONFIG_DECODERS)     += decode_with_drops.c
+decode_with_drops.SRCS          += ivfdec.h ivfdec.c
+decode_with_drops.SRCS          += tools_common.h tools_common.c
+decode_with_drops.SRCS          += video_common.h
+decode_with_drops.SRCS          += video_reader.h video_reader.c
+decode_with_drops.SRCS          += aom_ports/mem_ops.h
+decode_with_drops.SRCS          += aom_ports/mem_ops_aligned.h
+decode_with_drops.SRCS          += aom_ports/msvc.h
+decode_with_drops.GUID           = CE5C53C4-8DDA-438A-86ED-0DDD3CDB8D26
+decode_with_drops.DESCRIPTION    = Drops frames while decoding
+EXAMPLES-$(CONFIG_ENCODERS)        += set_maps.c
+set_maps.SRCS                      += ivfenc.h ivfenc.c
+set_maps.SRCS                      += tools_common.h tools_common.c
+set_maps.SRCS                      += video_common.h
+set_maps.SRCS                      += video_writer.h video_writer.c
+set_maps.SRCS                      += aom_ports/msvc.h
+set_maps.GUID                       = ECB2D24D-98B8-4015-A465-A4AF3DCC145F
+set_maps.DESCRIPTION                = Set active and ROI maps
+ifeq ($(CONFIG_ENCODERS),yes)
+ifeq ($(CONFIG_DECODERS),yes)
+EXAMPLES-$(CONFIG_ENCODERS)        += aom_cx_set_ref.c
+aom_cx_set_ref.SRCS                += ivfenc.h ivfenc.c
+aom_cx_set_ref.SRCS                += tools_common.h tools_common.c
+aom_cx_set_ref.SRCS                += examples/encoder_util.h
+aom_cx_set_ref.SRCS                += examples/encoder_util.c
+aom_cx_set_ref.SRCS                += video_common.h
+aom_cx_set_ref.SRCS                += video_writer.h video_writer.c
+aom_cx_set_ref.SRCS                += aom_ports/msvc.h
+aom_cx_set_ref.GUID                 = C5E31F7F-96F6-48BD-BD3E-10EBF6E8057A
+aom_cx_set_ref.DESCRIPTION          = AV1 set encoder reference frame
+endif
+endif
+
+# Handle extra library flags depending on codec configuration
+
+# We should not link to math library (libm) on RVCT
+# when building for bare-metal targets
+ifeq ($(CONFIG_OS_SUPPORT), yes)
+CODEC_EXTRA_LIBS-$(CONFIG_AV1)            += m
+else
+    ifeq ($(CONFIG_GCC), yes)
+    CODEC_EXTRA_LIBS-$(CONFIG_AV1)        += m
+    endif
+endif
+#
+# End of specified files. The rest of the build rules should happen
+# automagically from here.
+#
+
+
+# Examples need different flags based on whether we're building
+# from an installed tree or a version controlled tree. Determine
+# the proper paths.
+ifeq ($(HAVE_ALT_TREE_LAYOUT),yes)
+    LIB_PATH-yes := $(SRC_PATH_BARE)/../lib
+    INC_PATH-yes := $(SRC_PATH_BARE)/../include
+else
+    LIB_PATH-yes                     += $(if $(BUILD_PFX),$(BUILD_PFX),.)
+    INC_PATH-$(CONFIG_AV1_DECODER)   += $(SRC_PATH_BARE)/av1
+    INC_PATH-$(CONFIG_AV1_ENCODER)   += $(SRC_PATH_BARE)/av1
+endif
+INC_PATH-$(CONFIG_LIBYUV) += $(SRC_PATH_BARE)/third_party/libyuv/include
+LIB_PATH := $(call enabled,LIB_PATH)
+INC_PATH := $(call enabled,INC_PATH)
+INTERNAL_CFLAGS = $(addprefix -I,$(INC_PATH))
+INTERNAL_LDFLAGS += $(addprefix -L,$(LIB_PATH))
+
+
+# Expand list of selected examples to build (as specified above)
+UTILS           = $(call enabled,UTILS)
+EXAMPLES        = $(addprefix examples/,$(call enabled,EXAMPLES))
+ALL_EXAMPLES    = $(UTILS) $(EXAMPLES)
+UTIL_SRCS       = $(foreach ex,$(UTILS),$($(ex:.c=).SRCS) $($(ex:.cc=).SRCS))
+ALL_SRCS        = $(foreach ex, $(ALL_EXAMPLES),  \
+                      $($(notdir $(ex:.c=)).SRCS) \
+                      $($(notdir $(ex:.cc=)).SRCS))
+CODEC_EXTRA_LIBS=$(sort $(call enabled,CODEC_EXTRA_LIBS))
+
+
+# Expand all example sources into a variable containing all sources
+# for that example (not just them main one specified in UTILS/EXAMPLES)
+# and add this file to the list (for MSVS workspace generation)
+EXAMPLES_C = $(filter-out %.cc, $(ALL_EXAMPLES))
+$(foreach ex,$(EXAMPLES_C), \
+    $(eval $(notdir $(ex:.c=)).SRCS += $(ex) examples.mk))
+EXAMPLES_CXX = $(filter-out %.c, $(ALL_EXAMPLES))
+$(foreach ex,$(EXAMPLES_CXX), \
+    $(eval $(notdir $(ex:.cc=)).SRCS += $(ex) examples.mk))
+
+# Create build/install dependencies for all examples. The common case
+# is handled here. The MSVS case is handled below.
+NOT_MSVS = $(if $(CONFIG_MSVS),,yes)
+DIST-BINS-$(NOT_MSVS)      += $(addprefix bin/,$(EXAMPLES_C:.c=$(EXE_SFX)))
+DIST-BINS-$(NOT_MSVS)      += $(addprefix bin/,$(EXAMPLES_CXX:.cc=$(EXE_SFX)))
+INSTALL-BINS-$(NOT_MSVS)   += $(addprefix bin/,$(UTILS:.c=$(EXE_SFX)))
+DIST-SRCS-yes              += $(ALL_SRCS)
+INSTALL-SRCS-yes           += $(UTIL_SRCS)
+OBJS-$(NOT_MSVS)           += $(call objs,$(ALL_SRCS))
+BINS-$(NOT_MSVS)           += $(addprefix $(BUILD_PFX), \
+                                  $(EXAMPLES_C:.c=$(EXE_SFX)))
+BINS-$(NOT_MSVS)           += $(addprefix $(BUILD_PFX), \
+                                  $(EXAMPLES_CXX:.cc=$(EXE_SFX)))
+
+# Instantiate linker template for all examples.
+CODEC_LIB=$(if $(CONFIG_DEBUG_LIBS),aom_g,aom)
+ifneq ($(filter darwin%,$(TGT_OS)),)
+SHARED_LIB_SUF=.dylib
+else
+ifneq ($(filter os2%,$(TGT_OS)),)
+SHARED_LIB_SUF=_dll.a
+else
+SHARED_LIB_SUF=.so
+endif
+endif
+CODEC_LIB_SUF=$(if $(CONFIG_SHARED),$(SHARED_LIB_SUF),.a)
+$(foreach bin,$(BINS-yes),\
+    $(eval $(bin):$(LIB_PATH)/lib$(CODEC_LIB)$(CODEC_LIB_SUF))\
+    $(eval $(call linker_template,$(bin),\
+        $(call objs,$($(notdir $(bin:$(EXE_SFX)=)).SRCS)) \
+        -l$(CODEC_LIB) $(addprefix -l,$(CODEC_EXTRA_LIBS))\
+        )))
+
+# The following pairs define a mapping of locations in the distribution
+# tree to locations in the source/build trees.
+INSTALL_MAPS += src/%.c   %.c
+INSTALL_MAPS += src/%     $(SRC_PATH_BARE)/%
+INSTALL_MAPS += bin/%     %
+INSTALL_MAPS += %         %
+
+
+# Set up additional MSVS environment
+ifeq ($(CONFIG_MSVS),yes)
+CODEC_LIB=$(if $(CONFIG_SHARED),aom,$(if $(CONFIG_STATIC_MSVCRT),aommt,aommd))
+# This variable uses deferred expansion intentionally, since the results of
+# $(wildcard) may change during the course of the Make.
+VS_PLATFORMS = $(foreach d,$(wildcard */Release/$(CODEC_LIB).lib),$(word 1,$(subst /, ,$(d))))
+INSTALL_MAPS += $(foreach p,$(VS_PLATFORMS),bin/$(p)/%  $(p)/Release/%)
+endif
+
+# Build Visual Studio Projects. We use a template here to instantiate
+# explicit rules rather than using an implicit rule because we want to
+# leverage make's VPATH searching rather than specifying the paths on
+# each file in ALL_EXAMPLES. This has the unfortunate side effect that
+# touching the source files trigger a rebuild of the project files
+# even though there is no real dependency there (the dependency is on
+# the makefiles). We may want to revisit this.
+define vcproj_template
+$(1): $($(1:.$(VCPROJ_SFX)=).SRCS) aom.$(VCPROJ_SFX)
+	$(if $(quiet),@echo "    [vcproj] $$@")
+	$(qexec)$$(GEN_VCPROJ)\
+            --exe\
+            --target=$$(TOOLCHAIN)\
+            --name=$$(@:.$(VCPROJ_SFX)=)\
+            --ver=$$(CONFIG_VS_VERSION)\
+            --proj-guid=$$($$(@:.$(VCPROJ_SFX)=).GUID)\
+            --src-path-bare="$(SRC_PATH_BARE)" \
+            $$(if $$(CONFIG_STATIC_MSVCRT),--static-crt) \
+            --out=$$@ $$(INTERNAL_CFLAGS) $$(CFLAGS) \
+            $$(INTERNAL_LDFLAGS) $$(LDFLAGS) -l$$(CODEC_LIB) $$^
+endef
+ALL_EXAMPLES_BASENAME := $(notdir $(ALL_EXAMPLES))
+PROJECTS-$(CONFIG_MSVS) += $(ALL_EXAMPLES_BASENAME:.c=.$(VCPROJ_SFX))
+INSTALL-BINS-$(CONFIG_MSVS) += $(foreach p,$(VS_PLATFORMS),\
+                               $(addprefix bin/$(p)/,$(ALL_EXAMPLES_BASENAME:.c=.exe)))
+$(foreach proj,$(call enabled,PROJECTS),\
+    $(eval $(call vcproj_template,$(proj))))
+
+#
+# Documentation Rules
+#
+%.dox: %.c
+	@echo "    [DOXY] $@"
+	@mkdir -p $(dir $@)
+	@echo "/*!\page example_$(@F:.dox=) $(@F:.dox=)" > $@
+	@echo "   \includelineno $(<F)" >> $@
+	@echo "*/" >> $@
+
+samples.dox: examples.mk
+	@echo "    [DOXY] $@"
+	@echo "/*!\page samples Sample Code" > $@
+	@echo "    This SDK includes a number of sample applications."\
+	      "Each sample documents a feature of the SDK in both prose"\
+	      "and the associated C code."\
+	      "The following samples are included: ">>$@
+	@$(foreach ex,$(sort $(notdir $(EXAMPLES:.c=))),\
+	   echo "     - \subpage example_$(ex) $($(ex).DESCRIPTION)" >> $@;)
+	@echo >> $@
+	@echo "    In addition, the SDK contains a number of utilities."\
+              "Since these utilities are built upon the concepts described"\
+              "in the sample code listed above, they are not documented in"\
+              "pieces like the samples are. Their source is included here"\
+              "for reference. The following utilities are included:" >> $@
+	@$(foreach ex,$(sort $(UTILS:.c=)),\
+	   echo "     - \subpage example_$(ex) $($(ex).DESCRIPTION)" >> $@;)
+	@echo "*/" >> $@
+
+CLEAN-OBJS += examples.doxy samples.dox $(ALL_EXAMPLES:.c=.dox)
+DOCS-yes += examples.doxy samples.dox
+examples.doxy: samples.dox $(ALL_EXAMPLES:.c=.dox)
+	@echo "INPUT += $^" > $@
diff --git a/third_party/aom/examples/analyzer.cc b/third_party/aom/examples/analyzer.cc
new file mode 100644
index 0000000000..591aaf25e7
--- /dev/null
+++ b/third_party/aom/examples/analyzer.cc
@@ -0,0 +1,695 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <wx/wx.h>
+#include <wx/aboutdlg.h>
+#include <wx/cmdline.h>
+#include <wx/dcbuffer.h>
+#include "./tools_common.h"
+#include "./video_reader.h"
+#include "aom/aom_decoder.h"
+#include "aom/aomdx.h"
+#include "av1/decoder/accounting.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/decoder/inspection.h"
+
+#define OD_SIGNMASK(a) (-((a) < 0))
+#define OD_FLIPSIGNI(a, b) (((a) + OD_SIGNMASK(b)) ^ OD_SIGNMASK(b))
+#define OD_DIV_ROUND(x, y) (((x) + OD_FLIPSIGNI((y) >> 1, x)) / (y))
+
+enum {
+  OD_LUMA_MASK = 1 << 0,
+  OD_CB_MASK = 1 << 1,
+  OD_CR_MASK = 1 << 2,
+  OD_ALL_MASK = OD_LUMA_MASK | OD_CB_MASK | OD_CR_MASK
+};
+
+class AV1Decoder {
+ private:
+  FILE *input;
+  wxString path;
+
+  AvxVideoReader *reader;
+  const AvxVideoInfo *info;
+  const AvxInterface *decoder;
+
+  insp_frame_data frame_data;
+
+  aom_codec_ctx_t codec;
+  bool show_padding;
+
+ public:
+  aom_image_t *image;
+  int frame;
+
+  int plane_mask;
+
+  AV1Decoder();
+  ~AV1Decoder();
+
+  bool open(const wxString &path);
+  void close();
+  bool step();
+
+  int getWidthPadding() const;
+  int getHeightPadding() const;
+  void togglePadding();
+  int getWidth() const;
+  int getHeight() const;
+
+  bool getAccountingStruct(Accounting **acct);
+  bool setInspectionCallback();
+
+  static void inspect(void *decoder, void *data);
+};
+
+AV1Decoder::AV1Decoder()
+    : reader(NULL), info(NULL), decoder(NULL), show_padding(false), image(NULL),
+      frame(0) {}
+
+AV1Decoder::~AV1Decoder() {}
+
+void AV1Decoder::togglePadding() { show_padding = !show_padding; }
+
+bool AV1Decoder::open(const wxString &path) {
+  reader = aom_video_reader_open(path.mb_str());
+  if (!reader) {
+    fprintf(stderr, "Failed to open %s for reading.", path.mb_str().data());
+    return false;
+  }
+  this->path = path;
+  info = aom_video_reader_get_info(reader);
+  decoder = get_aom_decoder_by_fourcc(info->codec_fourcc);
+  if (!decoder) {
+    fprintf(stderr, "Unknown input codec.");
+    return false;
+  }
+  printf("Using %s\n", aom_codec_iface_name(decoder->codec_interface()));
+  if (aom_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0)) {
+    fprintf(stderr, "Failed to initialize decoder.");
+    return false;
+  }
+  ifd_init(&frame_data, info->frame_width, info->frame_height);
+  setInspectionCallback();
+  return true;
+}
+
+void AV1Decoder::close() {}
+
+bool AV1Decoder::step() {
+  if (aom_video_reader_read_frame(reader)) {
+    size_t frame_size;
+    const unsigned char *frame_data;
+    frame_data = aom_video_reader_get_frame(reader, &frame_size);
+    if (aom_codec_decode(&codec, frame_data, frame_size, NULL, 0)) {
+      fprintf(stderr, "Failed to decode frame.");
+      return false;
+    } else {
+      aom_codec_iter_t iter = NULL;
+      image = aom_codec_get_frame(&codec, &iter);
+      if (image != NULL) {
+        frame++;
+        return true;
+      }
+      return false;
+    }
+  }
+  return false;
+}
+
+int AV1Decoder::getWidth() const {
+  return info->frame_width + 2 * getWidthPadding();
+}
+
+int AV1Decoder::getWidthPadding() const {
+  return show_padding
+             ? AOMMAX(info->frame_width + 16,
+                      ALIGN_POWER_OF_TWO(info->frame_width, 6)) -
+                   info->frame_width
+             : 0;
+}
+
+int AV1Decoder::getHeight() const {
+  return info->frame_height + 2 * getHeightPadding();
+}
+
+int AV1Decoder::getHeightPadding() const {
+  return show_padding
+             ? AOMMAX(info->frame_height + 16,
+                      ALIGN_POWER_OF_TWO(info->frame_height, 6)) -
+                   info->frame_height
+             : 0;
+}
+
+bool AV1Decoder::getAccountingStruct(Accounting **accounting) {
+  return aom_codec_control(&codec, AV1_GET_ACCOUNTING, accounting) ==
+         AOM_CODEC_OK;
+}
+
+bool AV1Decoder::setInspectionCallback() {
+  aom_inspect_init ii;
+  ii.inspect_cb = AV1Decoder::inspect;
+  ii.inspect_ctx = (void *)this;
+  return aom_codec_control(&codec, AV1_SET_INSPECTION_CALLBACK, &ii) ==
+         AOM_CODEC_OK;
+}
+
+void AV1Decoder::inspect(void *pbi, void *data) {
+  AV1Decoder *decoder = (AV1Decoder *)data;
+  ifd_inspect(&decoder->frame_data, pbi);
+}
+
+#define MIN_ZOOM (1)
+#define MAX_ZOOM (4)
+
+class AnalyzerPanel : public wxPanel {
+  DECLARE_EVENT_TABLE()
+
+ private:
+  AV1Decoder decoder;
+  const wxString path;
+
+  int zoom;
+  unsigned char *pixels;
+
+  const bool bit_accounting;
+  double *bpp_q3;
+
+  int plane_mask;
+
+  // The display size is the decode size, scaled by the zoom.
+  int getDisplayWidth() const;
+  int getDisplayHeight() const;
+
+  bool updateDisplaySize();
+
+  void computeBitsPerPixel();
+
+ public:
+  AnalyzerPanel(wxWindow *parent, const wxString &path,
+                const bool bit_accounting);
+  ~AnalyzerPanel();
+
+  bool open(const wxString &path);
+  void close();
+  void render();
+  void togglePadding();
+  bool nextFrame();
+  void refresh();
+
+  int getZoom() const;
+  bool setZoom(int zoom);
+
+  void setShowPlane(bool show_plane, int mask);
+
+  void onPaint(wxPaintEvent &event);  // NOLINT
+};
+
+BEGIN_EVENT_TABLE(AnalyzerPanel, wxPanel)
+EVT_PAINT(AnalyzerPanel::onPaint)
+END_EVENT_TABLE()
+
+AnalyzerPanel::AnalyzerPanel(wxWindow *parent, const wxString &path,
+                             const bool bit_accounting)
+    : wxPanel(parent), path(path), zoom(0), pixels(NULL),
+      bit_accounting(bit_accounting), bpp_q3(NULL), plane_mask(OD_ALL_MASK) {}
+
+AnalyzerPanel::~AnalyzerPanel() { close(); }
+
+void AnalyzerPanel::setShowPlane(bool show_plane, int mask) {
+  if (show_plane) {
+    plane_mask |= mask;
+  } else {
+    plane_mask &= ~mask;
+  }
+}
+
+void AnalyzerPanel::render() {
+  aom_image_t *img = decoder.image;
+  int y_stride = img->stride[0];
+  int cb_stride = img->stride[1];
+  int cr_stride = img->stride[2];
+  int p_stride = 3 * getDisplayWidth();
+  unsigned char *y_row = img->planes[0];
+  unsigned char *cb_row = img->planes[1];
+  unsigned char *cr_row = img->planes[2];
+  unsigned char *p_row = pixels;
+  int y_width_padding = decoder.getWidthPadding();
+  int cb_width_padding = y_width_padding >> 1;
+  int cr_width_padding = y_width_padding >> 1;
+  int y_height_padding = decoder.getHeightPadding();
+  int cb_height_padding = y_height_padding >> 1;
+  int cr_height_padding = y_height_padding >> 1;
+  for (int j = 0; j < decoder.getHeight(); j++) {
+    unsigned char *y = y_row - y_stride * y_height_padding;
+    unsigned char *cb = cb_row - cb_stride * cb_height_padding;
+    unsigned char *cr = cr_row - cr_stride * cr_height_padding;
+    unsigned char *p = p_row;
+    for (int i = 0; i < decoder.getWidth(); i++) {
+      int64_t yval;
+      int64_t cbval;
+      int64_t crval;
+      int pmask;
+      unsigned rval;
+      unsigned gval;
+      unsigned bval;
+      yval = *(y - y_width_padding);
+      cbval = *(cb - cb_width_padding);
+      crval = *(cr - cr_width_padding);
+      pmask = plane_mask;
+      if (pmask & OD_LUMA_MASK) {
+        yval -= 16;
+      } else {
+        yval = 128;
+      }
+      cbval = ((pmask & OD_CB_MASK) >> 1) * (cbval - 128);
+      crval = ((pmask & OD_CR_MASK) >> 2) * (crval - 128);
+      /*This is intentionally slow and very accurate.*/
+      rval = OD_CLAMPI(0, (int32_t)OD_DIV_ROUND(
+                              2916394880000LL * yval + 4490222169144LL * crval,
+                              9745792000LL),
+                       65535);
+      gval = OD_CLAMPI(0, (int32_t)OD_DIV_ROUND(2916394880000LL * yval -
+                                                    534117096223LL * cbval -
+                                                    1334761232047LL * crval,
+                                                9745792000LL),
+                       65535);
+      bval = OD_CLAMPI(0, (int32_t)OD_DIV_ROUND(
+                              2916394880000LL * yval + 5290866304968LL * cbval,
+                              9745792000LL),
+                       65535);
+      unsigned char *px_row = p;
+      for (int v = 0; v < zoom; v++) {
+        unsigned char *px = px_row;
+        for (int u = 0; u < zoom; u++) {
+          *(px + 0) = (unsigned char)(rval >> 8);
+          *(px + 1) = (unsigned char)(gval >> 8);
+          *(px + 2) = (unsigned char)(bval >> 8);
+          px += 3;
+        }
+        px_row += p_stride;
+      }
+      int dc = ((y - y_row) & 1) | (1 - img->x_chroma_shift);
+      y++;
+      cb += dc;
+      cr += dc;
+      p += zoom * 3;
+    }
+    int dc = -((j & 1) | (1 - img->y_chroma_shift));
+    y_row += y_stride;
+    cb_row += dc & cb_stride;
+    cr_row += dc & cr_stride;
+    p_row += zoom * p_stride;
+  }
+}
+
+void AnalyzerPanel::computeBitsPerPixel() {
+  Accounting *acct;
+  double bpp_total;
+  int totals_q3[MAX_SYMBOL_TYPES] = { 0 };
+  int sym_count[MAX_SYMBOL_TYPES] = { 0 };
+  decoder.getAccountingStruct(&acct);
+  for (int j = 0; j < decoder.getHeight(); j++) {
+    for (int i = 0; i < decoder.getWidth(); i++) {
+      bpp_q3[j * decoder.getWidth() + i] = 0.0;
+    }
+  }
+  bpp_total = 0;
+  for (int i = 0; i < acct->syms.num_syms; i++) {
+    AccountingSymbol *s;
+    s = &acct->syms.syms[i];
+    totals_q3[s->id] += s->bits;
+    sym_count[s->id] += s->samples;
+  }
+  printf("=== Frame: %-3i ===\n", decoder.frame - 1);
+  for (int i = 0; i < acct->syms.dictionary.num_strs; i++) {
+    if (totals_q3[i]) {
+      printf("%30s = %10.3f (%f bit/symbol)\n", acct->syms.dictionary.strs[i],
+             (float)totals_q3[i] / 8, (float)totals_q3[i] / 8 / sym_count[i]);
+    }
+  }
+  printf("\n");
+}
+
+void AnalyzerPanel::togglePadding() {
+  decoder.togglePadding();
+  updateDisplaySize();
+}
+
+bool AnalyzerPanel::nextFrame() {
+  if (decoder.step()) {
+    refresh();
+    return true;
+  }
+  return false;
+}
+
+void AnalyzerPanel::refresh() {
+  if (bit_accounting) {
+    computeBitsPerPixel();
+  }
+  render();
+}
+
+int AnalyzerPanel::getDisplayWidth() const { return zoom * decoder.getWidth(); }
+
+int AnalyzerPanel::getDisplayHeight() const {
+  return zoom * decoder.getHeight();
+}
+
+bool AnalyzerPanel::updateDisplaySize() {
+  unsigned char *p = (unsigned char *)malloc(
+      sizeof(*p) * 3 * getDisplayWidth() * getDisplayHeight());
+  if (p == NULL) {
+    return false;
+  }
+  free(pixels);
+  pixels = p;
+  SetSize(getDisplayWidth(), getDisplayHeight());
+  return true;
+}
+
+bool AnalyzerPanel::open(const wxString &path) {
+  if (!decoder.open(path)) {
+    return false;
+  }
+  if (!setZoom(MIN_ZOOM)) {
+    return false;
+  }
+  if (bit_accounting) {
+    bpp_q3 = (double *)malloc(sizeof(*bpp_q3) * decoder.getWidth() *
+                              decoder.getHeight());
+    if (bpp_q3 == NULL) {
+      fprintf(stderr, "Could not allocate memory for bit accounting\n");
+      close();
+      return false;
+    }
+  }
+  if (!nextFrame()) {
+    close();
+    return false;
+  }
+  SetFocus();
+  return true;
+}
+
+void AnalyzerPanel::close() {
+  decoder.close();
+  free(pixels);
+  pixels = NULL;
+  free(bpp_q3);
+  bpp_q3 = NULL;
+}
+
+int AnalyzerPanel::getZoom() const { return zoom; }
+
+bool AnalyzerPanel::setZoom(int z) {
+  if (z <= MAX_ZOOM && z >= MIN_ZOOM && zoom != z) {
+    int old_zoom = zoom;
+    zoom = z;
+    if (!updateDisplaySize()) {
+      zoom = old_zoom;
+      return false;
+    }
+    return true;
+  }
+  return false;
+}
+
+void AnalyzerPanel::onPaint(wxPaintEvent &) {
+  wxBitmap bmp(wxImage(getDisplayWidth(), getDisplayHeight(), pixels, true));
+  wxBufferedPaintDC dc(this, bmp);
+}
+
+class AnalyzerFrame : public wxFrame {
+  DECLARE_EVENT_TABLE()
+
+ private:
+  AnalyzerPanel *panel;
+  const bool bit_accounting;
+
+  wxMenu *fileMenu;
+  wxMenu *viewMenu;
+  wxMenu *playbackMenu;
+
+ public:
+  AnalyzerFrame(const bool bit_accounting);  // NOLINT
+
+  void onOpen(wxCommandEvent &event);   // NOLINT
+  void onClose(wxCommandEvent &event);  // NOLINT
+  void onQuit(wxCommandEvent &event);   // NOLINT
+
+  void onTogglePadding(wxCommandEvent &event);  // NOLINT
+  void onZoomIn(wxCommandEvent &event);         // NOLINT
+  void onZoomOut(wxCommandEvent &event);        // NOLINT
+  void onActualSize(wxCommandEvent &event);     // NOLINT
+
+  void onToggleViewMenuCheckBox(wxCommandEvent &event);          // NOLINT
+  void onResetAndToggleViewMenuCheckBox(wxCommandEvent &event);  // NOLINT
+
+  void onNextFrame(wxCommandEvent &event);  // NOLINT
+  void onGotoFrame(wxCommandEvent &event);  // NOLINT
+  void onRestart(wxCommandEvent &event);    // NOLINT
+
+  void onAbout(wxCommandEvent &event);  // NOLINT
+
+  bool open(const wxString &path);
+  bool setZoom(int zoom);
+  void updateViewMenu();
+};
+
+enum {
+  wxID_NEXT_FRAME = 6000,
+  wxID_SHOW_Y,
+  wxID_SHOW_U,
+  wxID_SHOW_V,
+  wxID_GOTO_FRAME,
+  wxID_RESTART,
+  wxID_ACTUAL_SIZE,
+  wxID_PADDING
+};
+
+BEGIN_EVENT_TABLE(AnalyzerFrame, wxFrame)
+EVT_MENU(wxID_OPEN, AnalyzerFrame::onOpen)
+EVT_MENU(wxID_CLOSE, AnalyzerFrame::onClose)
+EVT_MENU(wxID_EXIT, AnalyzerFrame::onQuit)
+EVT_MENU(wxID_PADDING, AnalyzerFrame::onTogglePadding)
+EVT_MENU(wxID_ZOOM_IN, AnalyzerFrame::onZoomIn)
+EVT_MENU(wxID_ZOOM_OUT, AnalyzerFrame::onZoomOut)
+EVT_MENU(wxID_ACTUAL_SIZE, AnalyzerFrame::onActualSize)
+EVT_MENU(wxID_SHOW_Y, AnalyzerFrame::onResetAndToggleViewMenuCheckBox)
+EVT_MENU(wxID_SHOW_U, AnalyzerFrame::onResetAndToggleViewMenuCheckBox)
+EVT_MENU(wxID_SHOW_V, AnalyzerFrame::onResetAndToggleViewMenuCheckBox)
+EVT_MENU(wxID_NEXT_FRAME, AnalyzerFrame::onNextFrame)
+EVT_MENU(wxID_GOTO_FRAME, AnalyzerFrame::onGotoFrame)
+EVT_MENU(wxID_RESTART, AnalyzerFrame::onRestart)
+EVT_MENU(wxID_ABOUT, AnalyzerFrame::onAbout)
+END_EVENT_TABLE()
+
+AnalyzerFrame::AnalyzerFrame(const bool bit_accounting)
+    : wxFrame(NULL, wxID_ANY, _("AV1 Stream Analyzer"), wxDefaultPosition,
+              wxDefaultSize, wxDEFAULT_FRAME_STYLE),
+      panel(NULL), bit_accounting(bit_accounting) {
+  wxMenuBar *mb = new wxMenuBar();
+
+  fileMenu = new wxMenu();
+  fileMenu->Append(wxID_OPEN, _("&Open...\tCtrl-O"), _("Open daala file"));
+  fileMenu->Append(wxID_CLOSE, _("&Close\tCtrl-W"), _("Close daala file"));
+  fileMenu->Enable(wxID_CLOSE, false);
+  fileMenu->Append(wxID_EXIT, _("E&xit\tCtrl-Q"), _("Quit this program"));
+  mb->Append(fileMenu, _("&File"));
+
+  wxAcceleratorEntry entries[2];
+  entries[0].Set(wxACCEL_CTRL, (int)'=', wxID_ZOOM_IN);
+  entries[1].Set(wxACCEL_CTRL | wxACCEL_SHIFT, (int)'-', wxID_ZOOM_OUT);
+  wxAcceleratorTable accel(2, entries);
+  this->SetAcceleratorTable(accel);
+
+  viewMenu = new wxMenu();
+  +viewMenu->Append(wxID_PADDING, _("Toggle padding\tCtrl-p"),
+                    _("Show padding"));
+  viewMenu->Append(wxID_ZOOM_IN, _("Zoom-In\tCtrl-+"), _("Double image size"));
+  viewMenu->Append(wxID_ZOOM_OUT, _("Zoom-Out\tCtrl--"), _("Half image size"));
+  viewMenu->Append(wxID_ACTUAL_SIZE, _("Actual size\tCtrl-0"),
+                   _("Actual size of the frame"));
+  viewMenu->AppendSeparator();
+  viewMenu->AppendCheckItem(wxID_SHOW_Y, _("&Y plane\tCtrl-Y"),
+                            _("Show Y plane"));
+  viewMenu->AppendCheckItem(wxID_SHOW_U, _("&U plane\tCtrl-U"),
+                            _("Show U plane"));
+  viewMenu->AppendCheckItem(wxID_SHOW_V, _("&V plane\tCtrl-V"),
+                            _("Show V plane"));
+  mb->Append(viewMenu, _("&View"));
+
+  playbackMenu = new wxMenu();
+  playbackMenu->Append(wxID_NEXT_FRAME, _("Next frame\tCtrl-."),
+                       _("Go to next frame"));
+  /*playbackMenu->Append(wxID_RESTART, _("&Restart\tCtrl-R"),
+                       _("Set video to frame 0"));
+  playbackMenu->Append(wxID_GOTO_FRAME, _("Jump to Frame\tCtrl-J"),
+                       _("Go to frame number"));*/
+  mb->Append(playbackMenu, _("&Playback"));
+
+  wxMenu *helpMenu = new wxMenu();
+  helpMenu->Append(wxID_ABOUT, _("&About...\tF1"), _("Show about dialog"));
+  mb->Append(helpMenu, _("&Help"));
+
+  SetMenuBar(mb);
+
+  CreateStatusBar(1);
+}
+
+void AnalyzerFrame::onOpen(wxCommandEvent &WXUNUSED(event)) {
+  wxFileDialog openFileDialog(this, _("Open file"), wxEmptyString,
+                              wxEmptyString, _("AV1 files (*.ivf)|*.ivf"),
+                              wxFD_OPEN | wxFD_FILE_MUST_EXIST);
+  if (openFileDialog.ShowModal() != wxID_CANCEL) {
+    open(openFileDialog.GetPath());
+  }
+}
+
+void AnalyzerFrame::onClose(wxCommandEvent &WXUNUSED(event)) {}
+
+void AnalyzerFrame::onQuit(wxCommandEvent &WXUNUSED(event)) { Close(true); }
+
+void AnalyzerFrame::onTogglePadding(wxCommandEvent &WXUNUSED(event)) {
+  panel->togglePadding();
+  SetClientSize(panel->GetSize());
+  panel->render();
+  panel->Refresh();
+}
+
+void AnalyzerFrame::onZoomIn(wxCommandEvent &WXUNUSED(event)) {
+  setZoom(panel->getZoom() + 1);
+}
+
+void AnalyzerFrame::onZoomOut(wxCommandEvent &WXUNUSED(event)) {
+  setZoom(panel->getZoom() - 1);
+}
+
+void AnalyzerFrame::onActualSize(wxCommandEvent &WXUNUSED(event)) {
+  setZoom(MIN_ZOOM);
+}
+
+void AnalyzerFrame::onToggleViewMenuCheckBox(wxCommandEvent &event) {  // NOLINT
+  GetMenuBar()->Check(event.GetId(), event.IsChecked());
+  updateViewMenu();
+}
+
+void AnalyzerFrame::onResetAndToggleViewMenuCheckBox(
+    wxCommandEvent &event) {  // NOLINT
+  int id = event.GetId();
+  if (id != wxID_SHOW_Y && id != wxID_SHOW_U && id != wxID_SHOW_V) {
+    GetMenuBar()->Check(wxID_SHOW_Y, true);
+    GetMenuBar()->Check(wxID_SHOW_U, true);
+    GetMenuBar()->Check(wxID_SHOW_V, true);
+  }
+  onToggleViewMenuCheckBox(event);
+}
+
+void AnalyzerFrame::onNextFrame(wxCommandEvent &WXUNUSED(event)) {
+  panel->nextFrame();
+  panel->Refresh(false);
+}
+
+void AnalyzerFrame::onGotoFrame(wxCommandEvent &WXUNUSED(event)) {}
+
+void AnalyzerFrame::onRestart(wxCommandEvent &WXUNUSED(event)) {}
+
+void AnalyzerFrame::onAbout(wxCommandEvent &WXUNUSED(event)) {
+  wxAboutDialogInfo info;
+  info.SetName(_("AV1 Bitstream Analyzer"));
+  info.SetVersion(_("0.1-beta"));
+  info.SetDescription(
+      _("This program implements a bitstream analyzer for AV1"));
+  info.SetCopyright(
+      wxT("(C) 2017 Alliance for Open Media <negge@mozilla.com>"));
+  wxAboutBox(info);
+}
+
+bool AnalyzerFrame::open(const wxString &path) {
+  panel = new AnalyzerPanel(this, path, bit_accounting);
+  if (panel->open(path)) {
+    SetClientSize(panel->GetSize());
+    return true;
+  } else {
+    delete panel;
+    return false;
+  }
+}
+
+bool AnalyzerFrame::setZoom(int zoom) {
+  if (panel->setZoom(zoom)) {
+    GetMenuBar()->Enable(wxID_ACTUAL_SIZE, zoom != MIN_ZOOM);
+    GetMenuBar()->Enable(wxID_ZOOM_IN, zoom != MAX_ZOOM);
+    GetMenuBar()->Enable(wxID_ZOOM_OUT, zoom != MIN_ZOOM);
+    SetClientSize(panel->GetSize());
+    panel->render();
+    panel->Refresh();
+    return true;
+  }
+  return false;
+}
+
+void AnalyzerFrame::updateViewMenu() {
+  panel->setShowPlane(GetMenuBar()->IsChecked(wxID_SHOW_Y), OD_LUMA_MASK);
+  panel->setShowPlane(GetMenuBar()->IsChecked(wxID_SHOW_U), OD_CB_MASK);
+  panel->setShowPlane(GetMenuBar()->IsChecked(wxID_SHOW_V), OD_CR_MASK);
+  SetClientSize(panel->GetSize());
+  panel->render();
+  panel->Refresh(false);
+}
+
+class Analyzer : public wxApp {
+ private:
+  AnalyzerFrame *frame;
+
+ public:
+  void OnInitCmdLine(wxCmdLineParser &parser);    // NOLINT
+  bool OnCmdLineParsed(wxCmdLineParser &parser);  // NOLINT
+};
+
+static const wxCmdLineEntryDesc CMD_LINE_DESC[] = {
+  { wxCMD_LINE_SWITCH, _("h"), _("help"), _("Display this help and exit."),
+    wxCMD_LINE_VAL_NONE, wxCMD_LINE_OPTION_HELP },
+  { wxCMD_LINE_SWITCH, _("a"), _("bit-accounting"), _("Enable bit accounting"),
+    wxCMD_LINE_VAL_NONE, wxCMD_LINE_PARAM_OPTIONAL },
+  { wxCMD_LINE_PARAM, NULL, NULL, _("input.ivf"), wxCMD_LINE_VAL_STRING,
+    wxCMD_LINE_PARAM_OPTIONAL },
+  { wxCMD_LINE_NONE }
+};
+
+void Analyzer::OnInitCmdLine(wxCmdLineParser &parser) {  // NOLINT
+  parser.SetDesc(CMD_LINE_DESC);
+  parser.SetSwitchChars(_("-"));
+}
+
+bool Analyzer::OnCmdLineParsed(wxCmdLineParser &parser) {  // NOLINT
+  bool bit_accounting = parser.Found(_("a"));
+  if (bit_accounting && !CONFIG_ACCOUNTING) {
+    fprintf(stderr,
+            "Bit accounting support not found.  "
+            "Recompile with:\n./configure --enable-accounting\n");
+    return false;
+  }
+  frame = new AnalyzerFrame(parser.Found(_("a")));
+  frame->Show();
+  if (parser.GetParamCount() > 0) {
+    return frame->open(parser.GetParam(0));
+  }
+  return true;
+}
+
+void usage_exit(void) {
+  fprintf(stderr, "uhh\n");
+  exit(EXIT_FAILURE);
+}
+
+IMPLEMENT_APP(Analyzer)
diff --git a/third_party/aom/examples/aom_cx_set_ref.c b/third_party/aom/examples/aom_cx_set_ref.c
new file mode 100644
index 0000000000..ff24fa14a3
--- /dev/null
+++ b/third_party/aom/examples/aom_cx_set_ref.c
@@ -0,0 +1,323 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// AV1 Set Reference Frame
+// ============================
+//
+// This is an example demonstrating how to overwrite the AV1 encoder's
+// internal reference frame. In the sample we set the last frame to the
+// current frame. This technique could be used to bounce between two cameras.
+//
+// The decoder would also have to set the reference frame to the same value
+// on the same frame, or the video will become corrupt. The 'test_decode'
+// variable is set to 1 in this example that tests if the encoder and decoder
+// results are matching.
+//
+// Usage
+// -----
+// This example encodes a raw video. And the last argument passed in specifies
+// the frame number to update the reference frame on. For example, run
+// examples/aom_cx_set_ref av1 352 288 in.yuv out.ivf 4 30
+// The parameter is parsed as follows:
+//
+//
+// Extra Variables
+// ---------------
+// This example maintains the frame number passed on the command line
+// in the `update_frame_num` variable.
+//
+//
+// Configuration
+// -------------
+//
+// The reference frame is updated on the frame specified on the command
+// line.
+//
+// Observing The Effects
+// ---------------------
+// The encoder and decoder results should be matching when the same reference
+// frame setting operation is done in both encoder and decoder. Otherwise,
+// the encoder/decoder mismatch would be seen.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aomcx.h"
+#include "aom/aom_decoder.h"
+#include "aom/aom_encoder.h"
+#include "examples/encoder_util.h"
+#include "./tools_common.h"
+#include "./video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit() {
+  fprintf(stderr,
+          "Usage: %s <codec> <width> <height> <infile> <outfile> "
+          "<frame> <limit(optional)>\n",
+          exec_name);
+  exit(EXIT_FAILURE);
+}
+
+static void testing_decode(aom_codec_ctx_t *encoder, aom_codec_ctx_t *decoder,
+                           unsigned int frame_out, int *mismatch_seen) {
+  aom_image_t enc_img, dec_img;
+  struct av1_ref_frame ref_enc, ref_dec;
+
+  if (*mismatch_seen) return;
+
+  ref_enc.idx = 0;
+  ref_dec.idx = 0;
+  if (aom_codec_control(encoder, AV1_GET_REFERENCE, &ref_enc))
+    die_codec(encoder, "Failed to get encoder reference frame");
+  enc_img = ref_enc.img;
+  if (aom_codec_control(decoder, AV1_GET_REFERENCE, &ref_dec))
+    die_codec(decoder, "Failed to get decoder reference frame");
+  dec_img = ref_dec.img;
+
+  if (!aom_compare_img(&enc_img, &dec_img)) {
+    int y[4], u[4], v[4];
+
+    *mismatch_seen = 1;
+
+    aom_find_mismatch(&enc_img, &dec_img, y, u, v);
+    printf(
+        "Encode/decode mismatch on frame %d at"
+        " Y[%d, %d] {%d/%d},"
+        " U[%d, %d] {%d/%d},"
+        " V[%d, %d] {%d/%d}",
+        frame_out, y[0], y[1], y[2], y[3], u[0], u[1], u[2], u[3], v[0], v[1],
+        v[2], v[3]);
+  }
+
+  aom_img_free(&enc_img);
+  aom_img_free(&dec_img);
+}
+
+static int encode_frame(aom_codec_ctx_t *ecodec, aom_image_t *img,
+                        unsigned int frame_in, AvxVideoWriter *writer,
+                        int test_decode, aom_codec_ctx_t *dcodec,
+                        unsigned int *frame_out, int *mismatch_seen) {
+  int got_pkts = 0;
+  aom_codec_iter_t iter = NULL;
+  const aom_codec_cx_pkt_t *pkt = NULL;
+  int got_data;
+  const aom_codec_err_t res =
+      aom_codec_encode(ecodec, img, frame_in, 1, 0, AOM_DL_GOOD_QUALITY);
+  if (res != AOM_CODEC_OK) die_codec(ecodec, "Failed to encode frame");
+
+  got_data = 0;
+
+  while ((pkt = aom_codec_get_cx_data(ecodec, &iter)) != NULL) {
+    got_pkts = 1;
+
+    if (pkt->kind == AOM_CODEC_CX_FRAME_PKT) {
+      const int keyframe = (pkt->data.frame.flags & AOM_FRAME_IS_KEY) != 0;
+
+      if (!(pkt->data.frame.flags & AOM_FRAME_IS_FRAGMENT)) {
+        *frame_out += 1;
+      }
+
+      if (!aom_video_writer_write_frame(writer, pkt->data.frame.buf,
+                                        pkt->data.frame.sz,
+                                        pkt->data.frame.pts)) {
+        die_codec(ecodec, "Failed to write compressed frame");
+      }
+      printf(keyframe ? "K" : ".");
+      fflush(stdout);
+      got_data = 1;
+
+      // Decode 1 frame.
+      if (test_decode) {
+        if (aom_codec_decode(dcodec, pkt->data.frame.buf,
+                             (unsigned int)pkt->data.frame.sz, NULL, 0))
+          die_codec(dcodec, "Failed to decode frame.");
+      }
+    }
+  }
+
+  // Mismatch checking
+  if (got_data && test_decode) {
+    testing_decode(ecodec, dcodec, *frame_out, mismatch_seen);
+  }
+
+  return got_pkts;
+}
+
+int main(int argc, char **argv) {
+  FILE *infile = NULL;
+  // Encoder
+  aom_codec_ctx_t ecodec;
+  aom_codec_enc_cfg_t cfg;
+  unsigned int frame_in = 0;
+  aom_image_t raw;
+  aom_codec_err_t res;
+  AvxVideoInfo info;
+  AvxVideoWriter *writer = NULL;
+  const AvxInterface *encoder = NULL;
+
+  // Test encoder/decoder mismatch.
+  int test_decode = 1;
+  // Decoder
+  aom_codec_ctx_t dcodec;
+  unsigned int frame_out = 0;
+
+  // The frame number to set reference frame on
+  unsigned int update_frame_num = 0;
+  int mismatch_seen = 0;
+
+  const int fps = 30;
+  const int bitrate = 500;
+
+  const char *codec_arg = NULL;
+  const char *width_arg = NULL;
+  const char *height_arg = NULL;
+  const char *infile_arg = NULL;
+  const char *outfile_arg = NULL;
+  const char *update_frame_num_arg = NULL;
+  unsigned int limit = 0;
+  exec_name = argv[0];
+
+  // Clear explicitly, as simply assigning "{ 0 }" generates
+  // "missing-field-initializers" warning in some compilers.
+  memset(&ecodec, 0, sizeof(ecodec));
+  memset(&cfg, 0, sizeof(cfg));
+  memset(&info, 0, sizeof(info));
+
+  if (argc < 7) die("Invalid number of arguments");
+
+  codec_arg = argv[1];
+  width_arg = argv[2];
+  height_arg = argv[3];
+  infile_arg = argv[4];
+  outfile_arg = argv[5];
+  update_frame_num_arg = argv[6];
+
+  encoder = get_aom_encoder_by_name(codec_arg);
+  if (!encoder) die("Unsupported codec.");
+
+  update_frame_num = (unsigned int)strtoul(update_frame_num_arg, NULL, 0);
+  // In AV1, the reference buffers (cm->buffer_pool->frame_bufs[i].buf) are
+  // allocated while calling aom_codec_encode(), thus, setting reference for
+  // 1st frame isn't supported.
+  if (update_frame_num <= 1) {
+    die("Couldn't parse frame number '%s'\n", update_frame_num_arg);
+  }
+
+  if (argc > 7) {
+    limit = (unsigned int)strtoul(argv[7], NULL, 0);
+    if (update_frame_num > limit)
+      die("Update frame number couldn't larger than limit\n");
+  }
+
+  info.codec_fourcc = encoder->fourcc;
+  info.frame_width = (int)strtol(width_arg, NULL, 0);
+  info.frame_height = (int)strtol(height_arg, NULL, 0);
+  info.time_base.numerator = 1;
+  info.time_base.denominator = fps;
+
+  if (info.frame_width <= 0 || info.frame_height <= 0 ||
+      (info.frame_width % 2) != 0 || (info.frame_height % 2) != 0) {
+    die("Invalid frame size: %dx%d", info.frame_width, info.frame_height);
+  }
+
+  if (!aom_img_alloc(&raw, AOM_IMG_FMT_I420, info.frame_width,
+                     info.frame_height, 1)) {
+    die("Failed to allocate image.");
+  }
+
+  printf("Using %s\n", aom_codec_iface_name(encoder->codec_interface()));
+
+  res = aom_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res) die_codec(&ecodec, "Failed to get default codec config.");
+
+  cfg.g_w = info.frame_width;
+  cfg.g_h = info.frame_height;
+  cfg.g_timebase.num = info.time_base.numerator;
+  cfg.g_timebase.den = info.time_base.denominator;
+  cfg.rc_target_bitrate = bitrate;
+  cfg.g_lag_in_frames = 3;
+
+  writer = aom_video_writer_open(outfile_arg, kContainerIVF, &info);
+  if (!writer) die("Failed to open %s for writing.", outfile_arg);
+
+  if (!(infile = fopen(infile_arg, "rb")))
+    die("Failed to open %s for reading.", infile_arg);
+
+  if (aom_codec_enc_init(&ecodec, encoder->codec_interface(), &cfg, 0))
+    die_codec(&ecodec, "Failed to initialize encoder");
+
+  // Disable alt_ref.
+  if (aom_codec_control(&ecodec, AOME_SET_ENABLEAUTOALTREF, 0))
+    die_codec(&ecodec, "Failed to set enable auto alt ref");
+
+  if (test_decode) {
+    const AvxInterface *decoder = get_aom_decoder_by_name(codec_arg);
+    if (aom_codec_dec_init(&dcodec, decoder->codec_interface(), NULL, 0))
+      die_codec(&dcodec, "Failed to initialize decoder.");
+  }
+
+  // Encode frames.
+  while (aom_img_read(&raw, infile)) {
+    if (limit && frame_in >= limit) break;
+    if (update_frame_num > 1 && frame_out + 1 == update_frame_num) {
+      aom_ref_frame_t ref;
+      ref.frame_type = AOM_LAST_FRAME;
+      ref.img = raw;
+      // Set reference frame in encoder.
+      if (aom_codec_control(&ecodec, AOM_SET_REFERENCE, &ref))
+        die_codec(&ecodec, "Failed to set reference frame");
+      printf(" <SET_REF>");
+
+      // If set_reference in decoder is commented out, the enc/dec mismatch
+      // would be seen.
+      if (test_decode) {
+        if (aom_codec_control(&dcodec, AOM_SET_REFERENCE, &ref))
+          die_codec(&dcodec, "Failed to set reference frame");
+      }
+    }
+
+    encode_frame(&ecodec, &raw, frame_in, writer, test_decode, &dcodec,
+                 &frame_out, &mismatch_seen);
+    frame_in++;
+    if (mismatch_seen) break;
+  }
+
+  // Flush encoder.
+  if (!mismatch_seen)
+    while (encode_frame(&ecodec, NULL, frame_in, writer, test_decode, &dcodec,
+                        &frame_out, &mismatch_seen)) {
+    }
+
+  printf("\n");
+  fclose(infile);
+  printf("Processed %d frames.\n", frame_out);
+
+  if (test_decode) {
+    if (!mismatch_seen)
+      printf("Encoder/decoder results are matching.\n");
+    else
+      printf("Encoder/decoder results are NOT matching.\n");
+  }
+
+  if (test_decode)
+    if (aom_codec_destroy(&dcodec))
+      die_codec(&dcodec, "Failed to destroy decoder");
+
+  aom_img_free(&raw);
+  if (aom_codec_destroy(&ecodec))
+    die_codec(&ecodec, "Failed to destroy encoder.");
+
+  aom_video_writer_close(writer);
+
+  return EXIT_SUCCESS;
+}
diff --git a/third_party/aom/examples/decode_to_md5.c b/third_party/aom/examples/decode_to_md5.c
new file mode 100644
index 0000000000..5ab2532096
--- /dev/null
+++ b/third_party/aom/examples/decode_to_md5.c
@@ -0,0 +1,133 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// Frame-by-frame MD5 Checksum
+// ===========================
+//
+// This example builds upon the simple decoder loop to show how checksums
+// of the decoded output can be generated. These are used for validating
+// decoder implementations against the reference implementation, for example.
+//
+// MD5 algorithm
+// -------------
+// The Message-Digest 5 (MD5) is a well known hash function. We have provided
+// an implementation derived from the RSA Data Security, Inc. MD5 Message-Digest
+// Algorithm for your use. Our implmentation only changes the interface of this
+// reference code. You must include the `md5_utils.h` header for access to these
+// functions.
+//
+// Processing The Decoded Data
+// ---------------------------
+// Each row of the image is passed to the MD5 accumulator. First the Y plane
+// is processed, then U, then V. It is important to honor the image's `stride`
+// values.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aomdx.h"
+#include "aom/aom_decoder.h"
+
+#include "../md5_utils.h"
+#include "../tools_common.h"
+#include "../video_reader.h"
+#include "./aom_config.h"
+
+static void get_image_md5(const aom_image_t *img, unsigned char digest[16]) {
+  int plane, y;
+  MD5Context md5;
+
+  MD5Init(&md5);
+
+  for (plane = 0; plane < 3; ++plane) {
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int w = plane ? (img->d_w + 1) >> 1 : img->d_w;
+    const int h = plane ? (img->d_h + 1) >> 1 : img->d_h;
+
+    for (y = 0; y < h; ++y) {
+      MD5Update(&md5, buf, w);
+      buf += stride;
+    }
+  }
+
+  MD5Final(digest, &md5);
+}
+
+static void print_md5(FILE *stream, unsigned char digest[16]) {
+  int i;
+
+  for (i = 0; i < 16; ++i) fprintf(stream, "%02x", digest[i]);
+}
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <infile> <outfile>\n", exec_name);
+  exit(EXIT_FAILURE);
+}
+
+int main(int argc, char **argv) {
+  int frame_cnt = 0;
+  FILE *outfile = NULL;
+  aom_codec_ctx_t codec;
+  AvxVideoReader *reader = NULL;
+  const AvxVideoInfo *info = NULL;
+  const AvxInterface *decoder = NULL;
+
+  exec_name = argv[0];
+
+  if (argc != 3) die("Invalid number of arguments.");
+
+  reader = aom_video_reader_open(argv[1]);
+  if (!reader) die("Failed to open %s for reading.", argv[1]);
+
+  if (!(outfile = fopen(argv[2], "wb")))
+    die("Failed to open %s for writing.", argv[2]);
+
+  info = aom_video_reader_get_info(reader);
+
+  decoder = get_aom_decoder_by_fourcc(info->codec_fourcc);
+  if (!decoder) die("Unknown input codec.");
+
+  printf("Using %s\n", aom_codec_iface_name(decoder->codec_interface()));
+
+  if (aom_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0))
+    die_codec(&codec, "Failed to initialize decoder");
+
+  while (aom_video_reader_read_frame(reader)) {
+    aom_codec_iter_t iter = NULL;
+    aom_image_t *img = NULL;
+    size_t frame_size = 0;
+    const unsigned char *frame =
+        aom_video_reader_get_frame(reader, &frame_size);
+    if (aom_codec_decode(&codec, frame, (unsigned int)frame_size, NULL, 0))
+      die_codec(&codec, "Failed to decode frame");
+
+    while ((img = aom_codec_get_frame(&codec, &iter)) != NULL) {
+      unsigned char digest[16];
+
+      get_image_md5(img, digest);
+      print_md5(outfile, digest);
+      fprintf(outfile, "  img-%dx%d-%04d.i420\n", img->d_w, img->d_h,
+              ++frame_cnt);
+    }
+  }
+
+  printf("Processed %d frames.\n", frame_cnt);
+  if (aom_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec.");
+
+  aom_video_reader_close(reader);
+
+  fclose(outfile);
+  return EXIT_SUCCESS;
+}
diff --git a/third_party/aom/examples/decode_with_drops.c b/third_party/aom/examples/decode_with_drops.c
new file mode 100644
index 0000000000..45e0fb027e
--- /dev/null
+++ b/third_party/aom/examples/decode_with_drops.c
@@ -0,0 +1,149 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// Decode With Drops Example
+// =========================
+//
+// This is an example utility which drops a series of frames, as specified
+// on the command line. This is useful for observing the error recovery
+// features of the codec.
+//
+// Usage
+// -----
+// This example adds a single argument to the `simple_decoder` example,
+// which specifies the range or pattern of frames to drop. The parameter is
+// parsed as follows:
+//
+// Dropping A Range Of Frames
+// --------------------------
+// To drop a range of frames, specify the starting frame and the ending
+// frame to drop, separated by a dash. The following command will drop
+// frames 5 through 10 (base 1).
+//
+//  $ ./decode_with_drops in.ivf out.i420 5-10
+//
+//
+// Dropping A Pattern Of Frames
+// ----------------------------
+// To drop a pattern of frames, specify the number of frames to drop and
+// the number of frames after which to repeat the pattern, separated by
+// a forward-slash. The following command will drop 3 of 7 frames.
+// Specifically, it will decode 4 frames, then drop 3 frames, and then
+// repeat.
+//
+//  $ ./decode_with_drops in.ivf out.i420 3/7
+//
+//
+// Extra Variables
+// ---------------
+// This example maintains the pattern passed on the command line in the
+// `n`, `m`, and `is_range` variables:
+//
+//
+// Making The Drop Decision
+// ------------------------
+// The example decides whether to drop the frame based on the current
+// frame number, immediately before decoding the frame.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aomdx.h"
+#include "aom/aom_decoder.h"
+
+#include "../tools_common.h"
+#include "../video_reader.h"
+#include "./aom_config.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <infile> <outfile> <N-M|N/M>\n", exec_name);
+  exit(EXIT_FAILURE);
+}
+
+int main(int argc, char **argv) {
+  int frame_cnt = 0;
+  FILE *outfile = NULL;
+  aom_codec_ctx_t codec;
+  const AvxInterface *decoder = NULL;
+  AvxVideoReader *reader = NULL;
+  const AvxVideoInfo *info = NULL;
+  int n = 0;
+  int m = 0;
+  int is_range = 0;
+  char *nptr = NULL;
+
+  exec_name = argv[0];
+
+  if (argc != 4) die("Invalid number of arguments.");
+
+  reader = aom_video_reader_open(argv[1]);
+  if (!reader) die("Failed to open %s for reading.", argv[1]);
+
+  if (!(outfile = fopen(argv[2], "wb")))
+    die("Failed to open %s for writing.", argv[2]);
+
+  n = (int)strtol(argv[3], &nptr, 0);
+  m = (int)strtol(nptr + 1, NULL, 0);
+  is_range = (*nptr == '-');
+  if (!n || !m || (*nptr != '-' && *nptr != '/'))
+    die("Couldn't parse pattern %s.\n", argv[3]);
+
+  info = aom_video_reader_get_info(reader);
+
+  decoder = get_aom_decoder_by_fourcc(info->codec_fourcc);
+  if (!decoder) die("Unknown input codec.");
+
+  printf("Using %s\n", aom_codec_iface_name(decoder->codec_interface()));
+
+  if (aom_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0))
+    die_codec(&codec, "Failed to initialize decoder.");
+
+  while (aom_video_reader_read_frame(reader)) {
+    aom_codec_iter_t iter = NULL;
+    aom_image_t *img = NULL;
+    size_t frame_size = 0;
+    int skip;
+    const unsigned char *frame =
+        aom_video_reader_get_frame(reader, &frame_size);
+    if (aom_codec_decode(&codec, frame, (unsigned int)frame_size, NULL, 0))
+      die_codec(&codec, "Failed to decode frame.");
+
+    ++frame_cnt;
+
+    skip = (is_range && frame_cnt >= n && frame_cnt <= m) ||
+           (!is_range && m - (frame_cnt - 1) % m <= n);
+
+    if (!skip) {
+      putc('.', stdout);
+
+      while ((img = aom_codec_get_frame(&codec, &iter)) != NULL)
+        aom_img_write(img, outfile);
+    } else {
+      putc('X', stdout);
+    }
+
+    fflush(stdout);
+  }
+
+  printf("Processed %d frames.\n", frame_cnt);
+  if (aom_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec.");
+
+  printf("Play: ffplay -f rawvideo -pix_fmt yuv420p -s %dx%d %s\n",
+         info->frame_width, info->frame_height, argv[2]);
+
+  aom_video_reader_close(reader);
+  fclose(outfile);
+
+  return EXIT_SUCCESS;
+}
diff --git a/third_party/aom/examples/encoder_util.c b/third_party/aom/examples/encoder_util.c
new file mode 100644
index 0000000000..1aa3a7eef4
--- /dev/null
+++ b/third_party/aom/examples/encoder_util.c
@@ -0,0 +1,139 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// Utility functions used by encoder binaries.
+
+#include <assert.h>
+#include <string.h>
+
+#include "./encoder_util.h"
+#include "aom/aom_integer.h"
+
+#define mmin(a, b) ((a) < (b) ? (a) : (b))
+
+static void find_mismatch_plane(const aom_image_t *const img1,
+                                const aom_image_t *const img2, int plane,
+                                int use_highbitdepth, int loc[4]) {
+  const unsigned char *const p1 = img1->planes[plane];
+  const int p1_stride = img1->stride[plane] >> use_highbitdepth;
+  const unsigned char *const p2 = img2->planes[plane];
+  const int p2_stride = img2->stride[plane] >> use_highbitdepth;
+  const uint32_t bsize = 64;
+  const int is_y_plane = (plane == AOM_PLANE_Y);
+  const uint32_t bsizex = is_y_plane ? bsize : bsize >> img1->x_chroma_shift;
+  const uint32_t bsizey = is_y_plane ? bsize : bsize >> img1->y_chroma_shift;
+  const uint32_t c_w =
+      is_y_plane ? img1->d_w
+                 : (img1->d_w + img1->x_chroma_shift) >> img1->x_chroma_shift;
+  const uint32_t c_h =
+      is_y_plane ? img1->d_h
+                 : (img1->d_h + img1->y_chroma_shift) >> img1->y_chroma_shift;
+  assert(img1->d_w == img2->d_w && img1->d_h == img2->d_h);
+  assert(img1->x_chroma_shift == img2->x_chroma_shift &&
+         img1->y_chroma_shift == img2->y_chroma_shift);
+  loc[0] = loc[1] = loc[2] = loc[3] = -1;
+  int match = 1;
+  uint32_t i, j;
+  for (i = 0; match && i < c_h; i += bsizey) {
+    for (j = 0; match && j < c_w; j += bsizex) {
+      const int si =
+          is_y_plane ? mmin(i + bsizey, c_h) - i : mmin(i + bsizey, c_h - i);
+      const int sj =
+          is_y_plane ? mmin(j + bsizex, c_w) - j : mmin(j + bsizex, c_w - j);
+      int k, l;
+      for (k = 0; match && k < si; ++k) {
+        for (l = 0; match && l < sj; ++l) {
+          const int row = i + k;
+          const int col = j + l;
+          const int offset1 = row * p1_stride + col;
+          const int offset2 = row * p2_stride + col;
+          const int val1 = use_highbitdepth
+                               ? p1[2 * offset1] | (p1[2 * offset1 + 1] << 8)
+                               : p1[offset1];
+          const int val2 = use_highbitdepth
+                               ? p2[2 * offset2] | (p2[2 * offset2 + 1] << 8)
+                               : p2[offset2];
+          if (val1 != val2) {
+            loc[0] = row;
+            loc[1] = col;
+            loc[2] = val1;
+            loc[3] = val2;
+            match = 0;
+            break;
+          }
+        }
+      }
+    }
+  }
+}
+
+static void find_mismatch_helper(const aom_image_t *const img1,
+                                 const aom_image_t *const img2,
+                                 int use_highbitdepth, int yloc[4], int uloc[4],
+                                 int vloc[4]) {
+#if !CONFIG_HIGHBITDEPTH
+  assert(!use_highbitdepth);
+#endif  // !CONFIG_HIGHBITDEPTH
+  find_mismatch_plane(img1, img2, AOM_PLANE_Y, use_highbitdepth, yloc);
+  find_mismatch_plane(img1, img2, AOM_PLANE_U, use_highbitdepth, uloc);
+  find_mismatch_plane(img1, img2, AOM_PLANE_V, use_highbitdepth, vloc);
+}
+
+#if CONFIG_HIGHBITDEPTH
+void aom_find_mismatch_high(const aom_image_t *const img1,
+                            const aom_image_t *const img2, int yloc[4],
+                            int uloc[4], int vloc[4]) {
+  find_mismatch_helper(img1, img2, 1, yloc, uloc, vloc);
+}
+#endif
+
+void aom_find_mismatch(const aom_image_t *const img1,
+                       const aom_image_t *const img2, int yloc[4], int uloc[4],
+                       int vloc[4]) {
+  find_mismatch_helper(img1, img2, 0, yloc, uloc, vloc);
+}
+
+int aom_compare_img(const aom_image_t *const img1,
+                    const aom_image_t *const img2) {
+  uint32_t l_w = img1->d_w;
+  uint32_t c_w = (img1->d_w + img1->x_chroma_shift) >> img1->x_chroma_shift;
+  const uint32_t c_h =
+      (img1->d_h + img1->y_chroma_shift) >> img1->y_chroma_shift;
+  uint32_t i;
+  int match = 1;
+
+  match &= (img1->fmt == img2->fmt);
+  match &= (img1->d_w == img2->d_w);
+  match &= (img1->d_h == img2->d_h);
+#if CONFIG_HIGHBITDEPTH
+  if (img1->fmt & AOM_IMG_FMT_HIGHBITDEPTH) {
+    l_w *= 2;
+    c_w *= 2;
+  }
+#endif
+
+  for (i = 0; i < img1->d_h; ++i)
+    match &= (memcmp(img1->planes[AOM_PLANE_Y] + i * img1->stride[AOM_PLANE_Y],
+                     img2->planes[AOM_PLANE_Y] + i * img2->stride[AOM_PLANE_Y],
+                     l_w) == 0);
+
+  for (i = 0; i < c_h; ++i)
+    match &= (memcmp(img1->planes[AOM_PLANE_U] + i * img1->stride[AOM_PLANE_U],
+                     img2->planes[AOM_PLANE_U] + i * img2->stride[AOM_PLANE_U],
+                     c_w) == 0);
+
+  for (i = 0; i < c_h; ++i)
+    match &= (memcmp(img1->planes[AOM_PLANE_V] + i * img1->stride[AOM_PLANE_V],
+                     img2->planes[AOM_PLANE_V] + i * img2->stride[AOM_PLANE_V],
+                     c_w) == 0);
+
+  return match;
+}
diff --git a/third_party/aom/examples/encoder_util.h b/third_party/aom/examples/encoder_util.h
new file mode 100644
index 0000000000..38deef03d4
--- /dev/null
+++ b/third_party/aom/examples/encoder_util.h
@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// Utility functions used by encoder binaries.
+
+#ifndef EXAMPLES_ENCODER_UTIL_H_
+#define EXAMPLES_ENCODER_UTIL_H_
+
+#include "./aom_config.h"
+#include "aom/aom_image.h"
+
+// Returns mismatch location (?loc[0],?loc[1]) and the values at that location
+// in img1 (?loc[2]) and img2 (?loc[3]).
+#if CONFIG_HIGHBITDEPTH
+void aom_find_mismatch_high(const aom_image_t *const img1,
+                            const aom_image_t *const img2, int yloc[4],
+                            int uloc[4], int vloc[4]);
+#endif  // CONFIG_HIGHBITDEPTH
+
+void aom_find_mismatch(const aom_image_t *const img1,
+                       const aom_image_t *const img2, int yloc[4], int uloc[4],
+                       int vloc[4]);
+
+// Returns 1 if the two images match.
+int aom_compare_img(const aom_image_t *const img1,
+                    const aom_image_t *const img2);
+
+#endif  // EXAMPLES_ENCODER_UTIL_H_
diff --git a/third_party/aom/examples/inspect.c b/third_party/aom/examples/inspect.c
new file mode 100644
index 0000000000..345c0884d1
--- /dev/null
+++ b/third_party/aom/examples/inspect.c
@@ -0,0 +1,678 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// Inspect Decoder
+// ================
+//
+// This is a simple decoder loop that writes JSON stats to stdout. This tool
+// can also be compiled with Emscripten and used as a library.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "./args.h"
+#ifdef __EMSCRIPTEN__
+#include <emscripten.h>
+#else
+#define EMSCRIPTEN_KEEPALIVE
+#endif
+
+#include "aom/aom_decoder.h"
+#include "aom/aomdx.h"
+
+#include "../tools_common.h"
+#include "../video_reader.h"
+#include "./aom_config.h"
+// #include "av1/av1_dx_iface.c"
+#include "../av1/common/onyxc_int.h"
+#if CONFIG_ACCOUNTING
+#include "../av1/decoder/accounting.h"
+#endif
+#include "../av1/decoder/inspection.h"
+
+#include "../video_common.h"
+
+// Max JSON buffer size.
+const int MAX_BUFFER = 1024 * 1024 * 32;
+
+typedef enum {
+  ACCOUNTING_LAYER = 1,
+  BLOCK_SIZE_LAYER = 1 << 1,
+  TRANSFORM_SIZE_LAYER = 1 << 2,
+  TRANSFORM_TYPE_LAYER = 1 << 3,
+  MODE_LAYER = 1 << 4,
+  SKIP_LAYER = 1 << 5,
+  FILTER_LAYER = 1 << 6,
+  CDEF_LAYER = 1 << 7,
+  REFERENCE_FRAME_LAYER = 1 << 8,
+  MOTION_VECTORS_LAYER = 1 << 9,
+  UV_MODE_LAYER = 1 << 10,
+  ALL_LAYERS = (1 << 11) - 1
+} LayerType;
+
+static LayerType layers = 0;
+
+static int stop_after = 0;
+static int compress = 0;
+
+static const arg_def_t limit_arg =
+    ARG_DEF(NULL, "limit", 1, "Stop decoding after n frames");
+static const arg_def_t dump_all_arg = ARG_DEF("A", "all", 0, "Dump All");
+static const arg_def_t compress_arg =
+    ARG_DEF("x", "compress", 0, "Compress JSON using RLE");
+static const arg_def_t dump_accounting_arg =
+    ARG_DEF("a", "accounting", 0, "Dump Accounting");
+static const arg_def_t dump_block_size_arg =
+    ARG_DEF("bs", "blockSize", 0, "Dump Block Size");
+static const arg_def_t dump_motion_vectors_arg =
+    ARG_DEF("mv", "motionVectors", 0, "Dump Motion Vectors");
+static const arg_def_t dump_transform_size_arg =
+    ARG_DEF("ts", "transformSize", 0, "Dump Transform Size");
+static const arg_def_t dump_transform_type_arg =
+    ARG_DEF("tt", "transformType", 0, "Dump Transform Type");
+static const arg_def_t dump_mode_arg = ARG_DEF("m", "mode", 0, "Dump Mode");
+static const arg_def_t dump_uv_mode_arg =
+    ARG_DEF("uvm", "uv_mode", 0, "Dump UV Intra Prediction Modes");
+static const arg_def_t dump_skip_arg = ARG_DEF("s", "skip", 0, "Dump Skip");
+static const arg_def_t dump_filter_arg =
+    ARG_DEF("f", "filter", 0, "Dump Filter");
+static const arg_def_t dump_cdef_arg = ARG_DEF("c", "cdef", 0, "Dump CDEF");
+static const arg_def_t dump_reference_frame_arg =
+    ARG_DEF("r", "referenceFrame", 0, "Dump Reference Frame");
+static const arg_def_t usage_arg = ARG_DEF("h", "help", 0, "Help");
+
+static const arg_def_t *main_args[] = { &limit_arg,
+                                        &dump_all_arg,
+                                        &compress_arg,
+#if CONFIG_ACCOUNTING
+                                        &dump_accounting_arg,
+#endif
+                                        &dump_block_size_arg,
+                                        &dump_transform_size_arg,
+                                        &dump_transform_type_arg,
+                                        &dump_mode_arg,
+                                        &dump_uv_mode_arg,
+                                        &dump_skip_arg,
+                                        &dump_filter_arg,
+#if CONFIG_CDEF
+                                        &dump_cdef_arg,
+#endif
+                                        &dump_reference_frame_arg,
+                                        &dump_motion_vectors_arg,
+                                        &usage_arg,
+                                        NULL };
+#define ENUM(name) \
+  { #name, name }
+#define LAST_ENUM \
+  { NULL, 0 }
+typedef struct map_entry {
+  const char *name;
+  int value;
+} map_entry;
+
+const map_entry refs_map[] = { ENUM(INTRA_FRAME),  ENUM(LAST_FRAME),
+#if CONFIG_EXT_REFS
+                               ENUM(LAST2_FRAME),  ENUM(LAST3_FRAME),
+                               ENUM(GOLDEN_FRAME), ENUM(BWDREF_FRAME),
+                               ENUM(ALTREF_FRAME),
+#else
+                               ENUM(GOLDEN_FRAME), ENUM(ALTREF_FRAME),
+#endif
+                               LAST_ENUM };
+
+const map_entry block_size_map[] = {
+#if CONFIG_CB4X4
+  ENUM(BLOCK_2X2),    ENUM(BLOCK_2X4),    ENUM(BLOCK_4X2),
+#endif
+  ENUM(BLOCK_4X4),    ENUM(BLOCK_4X8),    ENUM(BLOCK_8X4),
+  ENUM(BLOCK_8X8),    ENUM(BLOCK_8X16),   ENUM(BLOCK_16X8),
+  ENUM(BLOCK_16X16),  ENUM(BLOCK_16X32),  ENUM(BLOCK_32X16),
+  ENUM(BLOCK_32X32),  ENUM(BLOCK_32X64),  ENUM(BLOCK_64X32),
+  ENUM(BLOCK_64X64),
+#if CONFIG_EXT_PARTITION
+  ENUM(BLOCK_64X128), ENUM(BLOCK_128X64), ENUM(BLOCK_128X128),
+#endif
+  LAST_ENUM
+};
+
+const map_entry tx_size_map[] = {
+#if CONFIG_CB4X4
+  ENUM(TX_2X2),
+#endif
+  ENUM(TX_4X4),   ENUM(TX_8X8),   ENUM(TX_16X16), ENUM(TX_32X32),
+#if CONFIG_TX64X64
+  ENUM(TX_64X64),
+#endif
+  ENUM(TX_4X8),   ENUM(TX_8X4),   ENUM(TX_8X16),  ENUM(TX_16X8),
+  ENUM(TX_16X32), ENUM(TX_32X16), ENUM(TX_4X16),  ENUM(TX_16X4),
+  ENUM(TX_8X32),  ENUM(TX_32X8),  LAST_ENUM
+};
+
+const map_entry tx_type_map[] = { ENUM(DCT_DCT),
+                                  ENUM(ADST_DCT),
+                                  ENUM(DCT_ADST),
+                                  ENUM(ADST_ADST),
+#if CONFIG_EXT_TX
+                                  ENUM(FLIPADST_DCT),
+                                  ENUM(DCT_FLIPADST),
+                                  ENUM(FLIPADST_FLIPADST),
+                                  ENUM(ADST_FLIPADST),
+                                  ENUM(FLIPADST_ADST),
+                                  ENUM(IDTX),
+                                  ENUM(V_DCT),
+                                  ENUM(H_DCT),
+                                  ENUM(V_ADST),
+                                  ENUM(H_ADST),
+                                  ENUM(V_FLIPADST),
+                                  ENUM(H_FLIPADST),
+#endif
+                                  LAST_ENUM };
+
+const map_entry prediction_mode_map[] = {
+  ENUM(DC_PRED),        ENUM(V_PRED),
+  ENUM(H_PRED),         ENUM(D45_PRED),
+  ENUM(D135_PRED),      ENUM(D117_PRED),
+  ENUM(D153_PRED),      ENUM(D207_PRED),
+  ENUM(D63_PRED),
+#if CONFIG_ALT_INTRA
+  ENUM(SMOOTH_PRED),
+#endif
+  ENUM(TM_PRED),        ENUM(NEARESTMV),
+  ENUM(NEARMV),         ENUM(ZEROMV),
+  ENUM(NEWMV),
+#if CONFIG_EXT_INTER
+  ENUM(NEWFROMNEARMV),  ENUM(NEAREST_NEARESTMV),
+  ENUM(NEAREST_NEARMV), ENUM(NEAR_NEARESTMV),
+  ENUM(NEAR_NEARMV),    ENUM(NEAREST_NEWMV),
+  ENUM(NEW_NEARESTMV),  ENUM(NEAR_NEWMV),
+  ENUM(NEW_NEARMV),     ENUM(ZERO_ZEROMV),
+  ENUM(NEW_NEWMV),
+#endif
+  ENUM(INTRA_INVALID),  LAST_ENUM
+};
+
+#define NO_SKIP 0
+#define SKIP 1
+
+const map_entry skip_map[] = { ENUM(SKIP), ENUM(NO_SKIP), LAST_ENUM };
+
+const map_entry config_map[] = { ENUM(MI_SIZE), LAST_ENUM };
+
+static const char *exec_name;
+
+insp_frame_data frame_data;
+int frame_count = 0;
+int decoded_frame_count = 0;
+aom_codec_ctx_t codec;
+AvxVideoReader *reader = NULL;
+const AvxVideoInfo *info = NULL;
+aom_image_t *img = NULL;
+
+void on_frame_decoded_dump(char *json) {
+#ifdef __EMSCRIPTEN__
+  EM_ASM_({ Module.on_frame_decoded_json($0); }, json);
+#else
+  printf("%s", json);
+#endif
+}
+
+// Writing out the JSON buffer using snprintf is very slow, especially when
+// compiled with emscripten, these functions speed things up quite a bit.
+int put_str(char *buffer, const char *str) {
+  int i;
+  for (i = 0; str[i] != '\0'; i++) {
+    buffer[i] = str[i];
+  }
+  return i;
+}
+
+int put_num(char *buffer, char prefix, int num, char suffix) {
+  int i = 0;
+  char *buf = buffer;
+  int is_neg = 0;
+  if (prefix) {
+    buf[i++] = prefix;
+  }
+  if (num == 0) {
+    buf[i++] = '0';
+  } else {
+    if (num < 0) {
+      num = -num;
+      is_neg = 1;
+    }
+    int s = i;
+    while (num != 0) {
+      buf[i++] = '0' + (num % 10);
+      num = num / 10;
+    }
+    if (is_neg) {
+      buf[i++] = '-';
+    }
+    int e = i - 1;
+    while (s < e) {
+      int t = buf[s];
+      buf[s] = buf[e];
+      buf[e] = t;
+      s++;
+      e--;
+    }
+  }
+  if (suffix) {
+    buf[i++] = suffix;
+  }
+  return i;
+}
+
+int put_map(char *buffer, const map_entry *map) {
+  char *buf = buffer;
+  const map_entry *entry = map;
+  while (entry->name != NULL) {
+    *(buf++) = '"';
+    buf += put_str(buf, entry->name);
+    *(buf++) = '"';
+    buf += put_num(buf, ':', entry->value, 0);
+    entry++;
+    if (entry->name != NULL) {
+      *(buf++) = ',';
+    }
+  }
+  return buf - buffer;
+}
+
+int put_reference_frame(char *buffer) {
+  const int mi_rows = frame_data.mi_rows;
+  const int mi_cols = frame_data.mi_cols;
+  char *buf = buffer;
+  int r, c, t;
+  buf += put_str(buf, "  \"referenceFrameMap\": {");
+  buf += put_map(buf, refs_map);
+  buf += put_str(buf, "},\n");
+  buf += put_str(buf, "  \"referenceFrame\": [");
+  for (r = 0; r < mi_rows; ++r) {
+    *(buf++) = '[';
+    for (c = 0; c < mi_cols; ++c) {
+      insp_mi_data *mi = &frame_data.mi_grid[r * mi_cols + c];
+      buf += put_num(buf, '[', mi->ref_frame[0], 0);
+      buf += put_num(buf, ',', mi->ref_frame[1], ']');
+      if (compress) {  // RLE
+        for (t = c + 1; t < mi_cols; ++t) {
+          insp_mi_data *next_mi = &frame_data.mi_grid[r * mi_cols + t];
+          if (mi->ref_frame[0] != next_mi->ref_frame[0] ||
+              mi->ref_frame[1] != next_mi->ref_frame[1]) {
+            break;
+          }
+        }
+        if (t - c > 1) {
+          *(buf++) = ',';
+          buf += put_num(buf, '[', t - c - 1, ']');
+          c = t - 1;
+        }
+      }
+      if (c < mi_cols - 1) *(buf++) = ',';
+    }
+    *(buf++) = ']';
+    if (r < mi_rows - 1) *(buf++) = ',';
+  }
+  buf += put_str(buf, "],\n");
+  return buf - buffer;
+}
+
+int put_motion_vectors(char *buffer) {
+  const int mi_rows = frame_data.mi_rows;
+  const int mi_cols = frame_data.mi_cols;
+  char *buf = buffer;
+  int r, c, t;
+  buf += put_str(buf, "  \"motionVectors\": [");
+  for (r = 0; r < mi_rows; ++r) {
+    *(buf++) = '[';
+    for (c = 0; c < mi_cols; ++c) {
+      insp_mi_data *mi = &frame_data.mi_grid[r * mi_cols + c];
+      buf += put_num(buf, '[', mi->mv[0].col, 0);
+      buf += put_num(buf, ',', mi->mv[0].row, 0);
+      buf += put_num(buf, ',', mi->mv[1].col, 0);
+      buf += put_num(buf, ',', mi->mv[1].row, ']');
+      if (compress) {  // RLE
+        for (t = c + 1; t < mi_cols; ++t) {
+          insp_mi_data *next_mi = &frame_data.mi_grid[r * mi_cols + t];
+          if (mi->mv[0].col != next_mi->mv[0].col ||
+              mi->mv[0].row != next_mi->mv[0].row ||
+              mi->mv[1].col != next_mi->mv[1].col ||
+              mi->mv[1].row != next_mi->mv[1].row) {
+            break;
+          }
+        }
+        if (t - c > 1) {
+          *(buf++) = ',';
+          buf += put_num(buf, '[', t - c - 1, ']');
+          c = t - 1;
+        }
+      }
+      if (c < mi_cols - 1) *(buf++) = ',';
+    }
+    *(buf++) = ']';
+    if (r < mi_rows - 1) *(buf++) = ',';
+  }
+  buf += put_str(buf, "],\n");
+  return buf - buffer;
+}
+
+int put_block_info(char *buffer, const map_entry *map, const char *name,
+                   size_t offset) {
+  const int mi_rows = frame_data.mi_rows;
+  const int mi_cols = frame_data.mi_cols;
+  char *buf = buffer;
+  int r, c, t, v;
+  if (map) {
+    buf += snprintf(buf, MAX_BUFFER, "  \"%sMap\": {", name);
+    buf += put_map(buf, map);
+    buf += put_str(buf, "},\n");
+  }
+  buf += snprintf(buf, MAX_BUFFER, "  \"%s\": [", name);
+  for (r = 0; r < mi_rows; ++r) {
+    *(buf++) = '[';
+    for (c = 0; c < mi_cols; ++c) {
+      insp_mi_data *curr_mi = &frame_data.mi_grid[r * mi_cols + c];
+      v = *(((int8_t *)curr_mi) + offset);
+      buf += put_num(buf, 0, v, 0);
+      if (compress) {  // RLE
+        for (t = c + 1; t < mi_cols; ++t) {
+          insp_mi_data *next_mi = &frame_data.mi_grid[r * mi_cols + t];
+          if (v != *(((int8_t *)next_mi) + offset)) {
+            break;
+          }
+        }
+        if (t - c > 1) {
+          *(buf++) = ',';
+          buf += put_num(buf, '[', t - c - 1, ']');
+          c = t - 1;
+        }
+      }
+      if (c < mi_cols - 1) *(buf++) = ',';
+    }
+    *(buf++) = ']';
+    if (r < mi_rows - 1) *(buf++) = ',';
+  }
+  buf += put_str(buf, "],\n");
+  return buf - buffer;
+}
+
+#if CONFIG_ACCOUNTING
+int put_accounting(char *buffer) {
+  char *buf = buffer;
+  int i;
+  const Accounting *accounting = frame_data.accounting;
+  if (accounting == NULL) {
+    printf("XXX\n");
+    return 0;
+  }
+  const int num_syms = accounting->syms.num_syms;
+  const int num_strs = accounting->syms.dictionary.num_strs;
+  buf += put_str(buf, "  \"symbolsMap\": [");
+  for (i = 0; i < num_strs; i++) {
+    buf += snprintf(buf, MAX_BUFFER, "\"%s\"",
+                    accounting->syms.dictionary.strs[i]);
+    if (i < num_strs - 1) *(buf++) = ',';
+  }
+  buf += put_str(buf, "],\n");
+  buf += put_str(buf, "  \"symbols\": [\n    ");
+  AccountingSymbolContext context;
+  context.x = -2;
+  context.y = -2;
+  AccountingSymbol *sym;
+  for (i = 0; i < num_syms; i++) {
+    sym = &accounting->syms.syms[i];
+    if (memcmp(&context, &sym->context, sizeof(AccountingSymbolContext)) != 0) {
+      buf += put_num(buf, '[', sym->context.x, 0);
+      buf += put_num(buf, ',', sym->context.y, ']');
+    } else {
+      buf += put_num(buf, '[', sym->id, 0);
+      buf += put_num(buf, ',', sym->bits, 0);
+      buf += put_num(buf, ',', sym->samples, ']');
+    }
+    context = sym->context;
+    if (i < num_syms - 1) *(buf++) = ',';
+  }
+  buf += put_str(buf, "],\n");
+  return buf - buffer;
+}
+#endif
+
+void inspect(void *pbi, void *data) {
+  /* Fetch frame data. */
+  ifd_inspect(&frame_data, pbi);
+  (void)data;
+  // We allocate enough space and hope we don't write out of bounds. Totally
+  // unsafe but this speeds things up, especially when compiled to Javascript.
+  char *buffer = aom_malloc(MAX_BUFFER);
+  char *buf = buffer;
+  buf += put_str(buf, "{\n");
+  if (layers & BLOCK_SIZE_LAYER) {
+    buf += put_block_info(buf, block_size_map, "blockSize",
+                          offsetof(insp_mi_data, sb_type));
+  }
+  if (layers & TRANSFORM_SIZE_LAYER) {
+    buf += put_block_info(buf, tx_size_map, "transformSize",
+                          offsetof(insp_mi_data, tx_size));
+  }
+  if (layers & TRANSFORM_TYPE_LAYER) {
+    buf += put_block_info(buf, tx_type_map, "transformType",
+                          offsetof(insp_mi_data, tx_type));
+  }
+  if (layers & MODE_LAYER) {
+    buf += put_block_info(buf, prediction_mode_map, "mode",
+                          offsetof(insp_mi_data, mode));
+  }
+  if (layers & UV_MODE_LAYER) {
+    buf += put_block_info(buf, prediction_mode_map, "uv_mode",
+                          offsetof(insp_mi_data, uv_mode));
+  }
+  if (layers & SKIP_LAYER) {
+    buf += put_block_info(buf, skip_map, "skip", offsetof(insp_mi_data, skip));
+  }
+  if (layers & FILTER_LAYER) {
+    buf += put_block_info(buf, NULL, "filter", offsetof(insp_mi_data, filter));
+  }
+#if CONFIG_CDEF
+  if (layers & CDEF_LAYER) {
+    buf += put_block_info(buf, NULL, "cdef_level",
+                          offsetof(insp_mi_data, cdef_level));
+    buf += put_block_info(buf, NULL, "cdef_strength",
+                          offsetof(insp_mi_data, cdef_strength));
+  }
+#endif
+  if (layers & MOTION_VECTORS_LAYER) {
+    buf += put_motion_vectors(buf);
+  }
+  if (layers & REFERENCE_FRAME_LAYER) {
+    buf += put_reference_frame(buf);
+  }
+#if CONFIG_ACCOUNTING
+  if (layers & ACCOUNTING_LAYER) {
+    buf += put_accounting(buf);
+  }
+#endif
+  buf += snprintf(buf, MAX_BUFFER, "  \"frame\": %d,\n", decoded_frame_count);
+  buf += snprintf(buf, MAX_BUFFER, "  \"showFrame\": %d,\n",
+                  frame_data.show_frame);
+  buf += snprintf(buf, MAX_BUFFER, "  \"frameType\": %d,\n",
+                  frame_data.frame_type);
+  buf += snprintf(buf, MAX_BUFFER, "  \"baseQIndex\": %d,\n",
+                  frame_data.base_qindex);
+  buf += snprintf(buf, MAX_BUFFER, "  \"tileCols\": %d,\n",
+                  frame_data.tile_mi_cols);
+  buf += snprintf(buf, MAX_BUFFER, "  \"tileRows\": %d,\n",
+                  frame_data.tile_mi_rows);
+  buf += put_str(buf, "  \"config\": {");
+  buf += put_map(buf, config_map);
+  buf += put_str(buf, "},\n");
+  buf += snprintf(buf, MAX_BUFFER, "  \"configString\": \"%s\"\n",
+                  aom_codec_build_config());
+  decoded_frame_count++;
+  buf += put_str(buf, "},\n");
+  *(buf++) = 0;
+  on_frame_decoded_dump(buffer);
+  aom_free(buffer);
+}
+
+void ifd_init_cb() {
+  aom_inspect_init ii;
+  ii.inspect_cb = inspect;
+  ii.inspect_ctx = NULL;
+  aom_codec_control(&codec, AV1_SET_INSPECTION_CALLBACK, &ii);
+}
+
+EMSCRIPTEN_KEEPALIVE
+int open_file(char *file) {
+  if (file == NULL) {
+    // The JS analyzer puts the .ivf file at this location.
+    file = "/tmp/input.ivf";
+  }
+  reader = aom_video_reader_open(file);
+  if (!reader) die("Failed to open %s for reading.", file);
+  info = aom_video_reader_get_info(reader);
+  const AvxInterface *decoder = get_aom_decoder_by_fourcc(info->codec_fourcc);
+  if (!decoder) die("Unknown input codec.");
+  fprintf(stderr, "Using %s\n",
+          aom_codec_iface_name(decoder->codec_interface()));
+  if (aom_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0))
+    die_codec(&codec, "Failed to initialize decoder.");
+  ifd_init(&frame_data, info->frame_width, info->frame_height);
+  ifd_init_cb();
+  return EXIT_SUCCESS;
+}
+
+EMSCRIPTEN_KEEPALIVE
+int read_frame() {
+  if (!aom_video_reader_read_frame(reader)) return EXIT_FAILURE;
+  img = NULL;
+  aom_codec_iter_t iter = NULL;
+  size_t frame_size = 0;
+  const unsigned char *frame = aom_video_reader_get_frame(reader, &frame_size);
+  if (aom_codec_decode(&codec, frame, (unsigned int)frame_size, NULL, 0) !=
+      AOM_CODEC_OK) {
+    die_codec(&codec, "Failed to decode frame.");
+  }
+  img = aom_codec_get_frame(&codec, &iter);
+  if (img == NULL) {
+    return EXIT_FAILURE;
+  }
+  ++frame_count;
+  return EXIT_SUCCESS;
+}
+
+EMSCRIPTEN_KEEPALIVE
+const char *get_aom_codec_build_config() { return aom_codec_build_config(); }
+
+EMSCRIPTEN_KEEPALIVE
+int get_bit_depth() { return img->bit_depth; }
+
+EMSCRIPTEN_KEEPALIVE
+unsigned char *get_plane(int plane) { return img->planes[plane]; }
+
+EMSCRIPTEN_KEEPALIVE
+int get_plane_stride(int plane) { return img->stride[plane]; }
+
+EMSCRIPTEN_KEEPALIVE
+int get_plane_width(int plane) { return aom_img_plane_width(img, plane); }
+
+EMSCRIPTEN_KEEPALIVE
+int get_plane_height(int plane) { return aom_img_plane_height(img, plane); }
+
+EMSCRIPTEN_KEEPALIVE
+int get_frame_width() { return info->frame_width; }
+
+EMSCRIPTEN_KEEPALIVE
+int get_frame_height() { return info->frame_height; }
+
+static void parse_args(char **argv) {
+  char **argi, **argj;
+  struct arg arg;
+  (void)dump_accounting_arg;
+  (void)dump_cdef_arg;
+  for (argi = argj = argv; (*argj = *argi); argi += arg.argv_step) {
+    arg.argv_step = 1;
+    if (arg_match(&arg, &dump_block_size_arg, argi)) layers |= BLOCK_SIZE_LAYER;
+#if CONFIG_ACCOUNTING
+    else if (arg_match(&arg, &dump_accounting_arg, argi))
+      layers |= ACCOUNTING_LAYER;
+#endif
+    else if (arg_match(&arg, &dump_transform_size_arg, argi))
+      layers |= TRANSFORM_SIZE_LAYER;
+    else if (arg_match(&arg, &dump_transform_type_arg, argi))
+      layers |= TRANSFORM_TYPE_LAYER;
+    else if (arg_match(&arg, &dump_mode_arg, argi))
+      layers |= MODE_LAYER;
+    else if (arg_match(&arg, &dump_uv_mode_arg, argi))
+      layers |= UV_MODE_LAYER;
+    else if (arg_match(&arg, &dump_skip_arg, argi))
+      layers |= SKIP_LAYER;
+    else if (arg_match(&arg, &dump_filter_arg, argi))
+      layers |= FILTER_LAYER;
+#if CONFIG_CDEF
+    else if (arg_match(&arg, &dump_cdef_arg, argi))
+      layers |= CDEF_LAYER;
+#endif
+    else if (arg_match(&arg, &dump_reference_frame_arg, argi))
+      layers |= REFERENCE_FRAME_LAYER;
+    else if (arg_match(&arg, &dump_motion_vectors_arg, argi))
+      layers |= MOTION_VECTORS_LAYER;
+    else if (arg_match(&arg, &dump_all_arg, argi))
+      layers |= ALL_LAYERS;
+    else if (arg_match(&arg, &compress_arg, argi))
+      compress = 1;
+    else if (arg_match(&arg, &usage_arg, argi))
+      usage_exit();
+    else if (arg_match(&arg, &limit_arg, argi))
+      stop_after = arg_parse_uint(&arg);
+    else
+      argj++;
+  }
+}
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s src_filename <options>\n", exec_name);
+  fprintf(stderr, "\nOptions:\n");
+  arg_show_usage(stderr, main_args);
+  exit(EXIT_FAILURE);
+}
+
+EMSCRIPTEN_KEEPALIVE
+int main(int argc, char **argv) {
+  exec_name = argv[0];
+  parse_args(argv);
+  if (argc >= 2) {
+    open_file(argv[1]);
+    printf("[\n");
+    while (1) {
+      if (stop_after && (decoded_frame_count >= stop_after)) break;
+      if (read_frame()) break;
+    }
+    printf("null\n");
+    printf("]");
+  } else {
+    usage_exit();
+  }
+}
+
+EMSCRIPTEN_KEEPALIVE
+void quit() {
+  if (aom_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec");
+  aom_video_reader_close(reader);
+}
+
+EMSCRIPTEN_KEEPALIVE
+void set_layers(LayerType v) { layers = v; }
+
+EMSCRIPTEN_KEEPALIVE
+void set_compress(int v) { compress = v; }
diff --git a/third_party/aom/examples/lossless_encoder.c b/third_party/aom/examples/lossless_encoder.c
new file mode 100644
index 0000000000..32ab18a163
--- /dev/null
+++ b/third_party/aom/examples/lossless_encoder.c
@@ -0,0 +1,139 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aom_encoder.h"
+#include "aom/aomcx.h"
+
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr,
+          "lossless_encoder: Example demonstrating lossless "
+          "encoding feature. Supports raw input only.\n");
+  fprintf(stderr, "Usage: %s <width> <height> <infile> <outfile>\n", exec_name);
+  exit(EXIT_FAILURE);
+}
+
+static int encode_frame(aom_codec_ctx_t *codec, aom_image_t *img,
+                        int frame_index, int flags, AvxVideoWriter *writer) {
+  int got_pkts = 0;
+  aom_codec_iter_t iter = NULL;
+  const aom_codec_cx_pkt_t *pkt = NULL;
+  const aom_codec_err_t res =
+      aom_codec_encode(codec, img, frame_index, 1, flags, AOM_DL_GOOD_QUALITY);
+  if (res != AOM_CODEC_OK) die_codec(codec, "Failed to encode frame");
+
+  while ((pkt = aom_codec_get_cx_data(codec, &iter)) != NULL) {
+    got_pkts = 1;
+
+    if (pkt->kind == AOM_CODEC_CX_FRAME_PKT) {
+      const int keyframe = (pkt->data.frame.flags & AOM_FRAME_IS_KEY) != 0;
+      if (!aom_video_writer_write_frame(writer, pkt->data.frame.buf,
+                                        pkt->data.frame.sz,
+                                        pkt->data.frame.pts)) {
+        die_codec(codec, "Failed to write compressed frame");
+      }
+      printf(keyframe ? "K" : ".");
+      fflush(stdout);
+    }
+  }
+
+  return got_pkts;
+}
+
+int main(int argc, char **argv) {
+  FILE *infile = NULL;
+  aom_codec_ctx_t codec;
+  aom_codec_enc_cfg_t cfg;
+  int frame_count = 0;
+  aom_image_t raw;
+  aom_codec_err_t res;
+  AvxVideoInfo info;
+  AvxVideoWriter *writer = NULL;
+  const AvxInterface *encoder = NULL;
+  const int fps = 30;
+
+  exec_name = argv[0];
+
+  // Clear explicitly, as simply assigning "{ 0 }" generates
+  // "missing-field-initializers" warning in some compilers.
+  memset(&info, 0, sizeof(info));
+
+  if (argc < 5) die("Invalid number of arguments");
+
+  encoder = get_aom_encoder_by_name("av1");
+  if (!encoder) die("Unsupported codec.");
+
+  info.codec_fourcc = encoder->fourcc;
+  info.frame_width = (int)strtol(argv[1], NULL, 0);
+  info.frame_height = (int)strtol(argv[2], NULL, 0);
+  info.time_base.numerator = 1;
+  info.time_base.denominator = fps;
+
+  if (info.frame_width <= 0 || info.frame_height <= 0 ||
+      (info.frame_width % 2) != 0 || (info.frame_height % 2) != 0) {
+    die("Invalid frame size: %dx%d", info.frame_width, info.frame_height);
+  }
+
+  if (!aom_img_alloc(&raw, AOM_IMG_FMT_I420, info.frame_width,
+                     info.frame_height, 1)) {
+    die("Failed to allocate image.");
+  }
+
+  printf("Using %s\n", aom_codec_iface_name(encoder->codec_interface()));
+
+  res = aom_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res) die_codec(&codec, "Failed to get default codec config.");
+
+  cfg.g_w = info.frame_width;
+  cfg.g_h = info.frame_height;
+  cfg.g_timebase.num = info.time_base.numerator;
+  cfg.g_timebase.den = info.time_base.denominator;
+
+  writer = aom_video_writer_open(argv[4], kContainerIVF, &info);
+  if (!writer) die("Failed to open %s for writing.", argv[4]);
+
+  if (!(infile = fopen(argv[3], "rb")))
+    die("Failed to open %s for reading.", argv[3]);
+
+  if (aom_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  if (aom_codec_control_(&codec, AV1E_SET_LOSSLESS, 1))
+    die_codec(&codec, "Failed to use lossless mode");
+
+  // Encode frames.
+  while (aom_img_read(&raw, infile)) {
+    encode_frame(&codec, &raw, frame_count++, 0, writer);
+  }
+
+  // Flush encoder.
+  while (encode_frame(&codec, NULL, -1, 0, writer)) {
+  }
+
+  printf("\n");
+  fclose(infile);
+  printf("Processed %d frames.\n", frame_count);
+
+  aom_img_free(&raw);
+  if (aom_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec.");
+
+  aom_video_writer_close(writer);
+
+  return EXIT_SUCCESS;
+}
diff --git a/third_party/aom/examples/resize_util.c b/third_party/aom/examples/resize_util.c
new file mode 100644
index 0000000000..5485691a8b
--- /dev/null
+++ b/third_party/aom/examples/resize_util.c
@@ -0,0 +1,124 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "../tools_common.h"
+#include "../av1/encoder/av1_resize.h"
+
+static const char *exec_name = NULL;
+
+static void usage() {
+  printf("Usage:\n");
+  printf("%s <input_yuv> <width>x<height> <target_width>x<target_height> ",
+         exec_name);
+  printf("<output_yuv> [<frames>]\n");
+}
+
+void usage_exit(void) {
+  usage();
+  exit(EXIT_FAILURE);
+}
+
+static int parse_dim(char *v, int *width, int *height) {
+  char *x = strchr(v, 'x');
+  if (x == NULL) x = strchr(v, 'X');
+  if (x == NULL) return 0;
+  *width = atoi(v);
+  *height = atoi(&x[1]);
+  if (*width <= 0 || *height <= 0)
+    return 0;
+  else
+    return 1;
+}
+
+int main(int argc, char *argv[]) {
+  char *fin, *fout;
+  FILE *fpin, *fpout;
+  uint8_t *inbuf, *outbuf;
+  uint8_t *inbuf_u, *outbuf_u;
+  uint8_t *inbuf_v, *outbuf_v;
+  int f, frames;
+  int width, height, target_width, target_height;
+
+  exec_name = argv[0];
+
+  if (argc < 5) {
+    printf("Incorrect parameters:\n");
+    usage();
+    return 1;
+  }
+
+  fin = argv[1];
+  fout = argv[4];
+  if (!parse_dim(argv[2], &width, &height)) {
+    printf("Incorrect parameters: %s\n", argv[2]);
+    usage();
+    return 1;
+  }
+  if (!parse_dim(argv[3], &target_width, &target_height)) {
+    printf("Incorrect parameters: %s\n", argv[3]);
+    usage();
+    return 1;
+  }
+
+  fpin = fopen(fin, "rb");
+  if (fpin == NULL) {
+    printf("Can't open file %s to read\n", fin);
+    usage();
+    return 1;
+  }
+  fpout = fopen(fout, "wb");
+  if (fpout == NULL) {
+    printf("Can't open file %s to write\n", fout);
+    usage();
+    return 1;
+  }
+  if (argc >= 6)
+    frames = atoi(argv[5]);
+  else
+    frames = INT_MAX;
+
+  printf("Input size:  %dx%d\n", width, height);
+  printf("Target size: %dx%d, Frames: ", target_width, target_height);
+  if (frames == INT_MAX)
+    printf("All\n");
+  else
+    printf("%d\n", frames);
+
+  inbuf = (uint8_t *)malloc(width * height * 3 / 2);
+  outbuf = (uint8_t *)malloc(target_width * target_height * 3 / 2);
+  inbuf_u = inbuf + width * height;
+  inbuf_v = inbuf_u + width * height / 4;
+  outbuf_u = outbuf + target_width * target_height;
+  outbuf_v = outbuf_u + target_width * target_height / 4;
+  f = 0;
+  while (f < frames) {
+    if (fread(inbuf, width * height * 3 / 2, 1, fpin) != 1) break;
+    av1_resize_frame420(inbuf, width, inbuf_u, inbuf_v, width / 2, height,
+                        width, outbuf, target_width, outbuf_u, outbuf_v,
+                        target_width / 2, target_height, target_width);
+    fwrite(outbuf, target_width * target_height * 3 / 2, 1, fpout);
+    f++;
+  }
+  printf("%d frames processed\n", f);
+  fclose(fpin);
+  fclose(fpout);
+
+  free(inbuf);
+  free(outbuf);
+  return 0;
+}
diff --git a/third_party/aom/examples/set_maps.c b/third_party/aom/examples/set_maps.c
new file mode 100644
index 0000000000..e88cd426f1
--- /dev/null
+++ b/third_party/aom/examples/set_maps.c
@@ -0,0 +1,210 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// AOM Set Active and ROI Maps
+// ===========================
+//
+// This is an example demonstrating how to control the AOM encoder's
+// ROI and Active maps.
+//
+// ROI (Reigon of Interest) maps are a way for the application to assign
+// each macroblock in the image to a region, and then set quantizer and
+// filtering parameters on that image.
+//
+// Active maps are a way for the application to specify on a
+// macroblock-by-macroblock basis whether there is any activity in that
+// macroblock.
+//
+//
+// Configuration
+// -------------
+// An ROI map is set on frame 22. If the width of the image in macroblocks
+// is evenly divisble by 4, then the output will appear to have distinct
+// columns, where the quantizer, loopfilter, and static threshold differ
+// from column to column.
+//
+// An active map is set on frame 33. If the width of the image in macroblocks
+// is evenly divisble by 4, then the output will appear to have distinct
+// columns, where one column will have motion and the next will not.
+//
+// The active map is cleared on frame 44.
+//
+// Observing The Effects
+// ---------------------
+// Use the `simple_decoder` example to decode this sample, and observe
+// the change in the image at frames 22, 33, and 44.
+
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aomcx.h"
+#include "aom/aom_encoder.h"
+
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <codec> <width> <height> <infile> <outfile>\n",
+          exec_name);
+  exit(EXIT_FAILURE);
+}
+
+static void set_active_map(const aom_codec_enc_cfg_t *cfg,
+                           aom_codec_ctx_t *codec) {
+  unsigned int i;
+  aom_active_map_t map = { 0, 0, 0 };
+
+  map.rows = (cfg->g_h + 15) / 16;
+  map.cols = (cfg->g_w + 15) / 16;
+
+  map.active_map = (uint8_t *)malloc(map.rows * map.cols);
+  for (i = 0; i < map.rows * map.cols; ++i) map.active_map[i] = i % 2;
+
+  if (aom_codec_control(codec, AOME_SET_ACTIVEMAP, &map))
+    die_codec(codec, "Failed to set active map");
+
+  free(map.active_map);
+}
+
+static void unset_active_map(const aom_codec_enc_cfg_t *cfg,
+                             aom_codec_ctx_t *codec) {
+  aom_active_map_t map = { 0, 0, 0 };
+
+  map.rows = (cfg->g_h + 15) / 16;
+  map.cols = (cfg->g_w + 15) / 16;
+  map.active_map = NULL;
+
+  if (aom_codec_control(codec, AOME_SET_ACTIVEMAP, &map))
+    die_codec(codec, "Failed to set active map");
+}
+
+static int encode_frame(aom_codec_ctx_t *codec, aom_image_t *img,
+                        int frame_index, AvxVideoWriter *writer) {
+  int got_pkts = 0;
+  aom_codec_iter_t iter = NULL;
+  const aom_codec_cx_pkt_t *pkt = NULL;
+  const aom_codec_err_t res =
+      aom_codec_encode(codec, img, frame_index, 1, 0, AOM_DL_GOOD_QUALITY);
+  if (res != AOM_CODEC_OK) die_codec(codec, "Failed to encode frame");
+
+  while ((pkt = aom_codec_get_cx_data(codec, &iter)) != NULL) {
+    got_pkts = 1;
+
+    if (pkt->kind == AOM_CODEC_CX_FRAME_PKT) {
+      const int keyframe = (pkt->data.frame.flags & AOM_FRAME_IS_KEY) != 0;
+      if (!aom_video_writer_write_frame(writer, pkt->data.frame.buf,
+                                        pkt->data.frame.sz,
+                                        pkt->data.frame.pts)) {
+        die_codec(codec, "Failed to write compressed frame");
+      }
+
+      printf(keyframe ? "K" : ".");
+      fflush(stdout);
+    }
+  }
+
+  return got_pkts;
+}
+
+int main(int argc, char **argv) {
+  FILE *infile = NULL;
+  aom_codec_ctx_t codec;
+  aom_codec_enc_cfg_t cfg;
+  int frame_count = 0;
+  const int limit = 30;
+  aom_image_t raw;
+  aom_codec_err_t res;
+  AvxVideoInfo info;
+  AvxVideoWriter *writer = NULL;
+  const AvxInterface *encoder = NULL;
+  const int fps = 2;  // TODO(dkovalev) add command line argument
+  const double bits_per_pixel_per_frame = 0.067;
+
+  exec_name = argv[0];
+  if (argc != 6) die("Invalid number of arguments");
+
+  memset(&info, 0, sizeof(info));
+
+  encoder = get_aom_encoder_by_name(argv[1]);
+  if (encoder == NULL) {
+    die("Unsupported codec.");
+  }
+  assert(encoder != NULL);
+  info.codec_fourcc = encoder->fourcc;
+  info.frame_width = (int)strtol(argv[2], NULL, 0);
+  info.frame_height = (int)strtol(argv[3], NULL, 0);
+  info.time_base.numerator = 1;
+  info.time_base.denominator = fps;
+
+  if (info.frame_width <= 0 || info.frame_height <= 0 ||
+      (info.frame_width % 2) != 0 || (info.frame_height % 2) != 0) {
+    die("Invalid frame size: %dx%d", info.frame_width, info.frame_height);
+  }
+
+  if (!aom_img_alloc(&raw, AOM_IMG_FMT_I420, info.frame_width,
+                     info.frame_height, 1)) {
+    die("Failed to allocate image.");
+  }
+
+  printf("Using %s\n", aom_codec_iface_name(encoder->codec_interface()));
+
+  res = aom_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res) die_codec(&codec, "Failed to get default codec config.");
+
+  cfg.g_w = info.frame_width;
+  cfg.g_h = info.frame_height;
+  cfg.g_timebase.num = info.time_base.numerator;
+  cfg.g_timebase.den = info.time_base.denominator;
+  cfg.rc_target_bitrate =
+      (unsigned int)(bits_per_pixel_per_frame * cfg.g_w * cfg.g_h * fps / 1000);
+  cfg.g_lag_in_frames = 0;
+
+  writer = aom_video_writer_open(argv[5], kContainerIVF, &info);
+  if (!writer) die("Failed to open %s for writing.", argv[5]);
+
+  if (!(infile = fopen(argv[4], "rb")))
+    die("Failed to open %s for reading.", argv[4]);
+
+  if (aom_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  // Encode frames.
+  while (aom_img_read(&raw, infile) && frame_count < limit) {
+    ++frame_count;
+
+    if (frame_count == 11) {
+      set_active_map(&cfg, &codec);
+    } else if (frame_count == 22) {
+      unset_active_map(&cfg, &codec);
+    }
+
+    encode_frame(&codec, &raw, frame_count, writer);
+  }
+
+  // Flush encoder.
+  while (encode_frame(&codec, NULL, -1, writer)) {
+  }
+
+  printf("\n");
+  fclose(infile);
+  printf("Processed %d frames.\n", frame_count);
+
+  aom_img_free(&raw);
+  if (aom_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec.");
+
+  aom_video_writer_close(writer);
+
+  return EXIT_SUCCESS;
+}
diff --git a/third_party/aom/examples/simple_decoder.c b/third_party/aom/examples/simple_decoder.c
new file mode 100644
index 0000000000..33a8945395
--- /dev/null
+++ b/third_party/aom/examples/simple_decoder.c
@@ -0,0 +1,150 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// Simple Decoder
+// ==============
+//
+// This is an example of a simple decoder loop. It takes an input file
+// containing the compressed data (in IVF format), passes it through the
+// decoder, and writes the decompressed frames to disk. Other decoder
+// examples build upon this one.
+//
+// The details of the IVF format have been elided from this example for
+// simplicity of presentation, as IVF files will not generally be used by
+// your application. In general, an IVF file consists of a file header,
+// followed by a variable number of frames. Each frame consists of a frame
+// header followed by a variable length payload. The length of the payload
+// is specified in the first four bytes of the frame header. The payload is
+// the raw compressed data.
+//
+// Standard Includes
+// -----------------
+// For decoders, you only have to include `aom_decoder.h` and then any
+// header files for the specific codecs you use. In this case, we're using
+// aom.
+//
+// Initializing The Codec
+// ----------------------
+// The libaom decoder is initialized by the call to aom_codec_dec_init().
+// Determining the codec interface to use is handled by AvxVideoReader and the
+// functions prefixed with aom_video_reader_. Discussion of those functions is
+// beyond the scope of this example, but the main gist is to open the input file
+// and parse just enough of it to determine if it's a AVx file and which AVx
+// codec is contained within the file.
+// Note the NULL pointer passed to aom_codec_dec_init(). We do that in this
+// example because we want the algorithm to determine the stream configuration
+// (width/height) and allocate memory automatically.
+//
+// Decoding A Frame
+// ----------------
+// Once the frame has been read into memory, it is decoded using the
+// `aom_codec_decode` function. The call takes a pointer to the data
+// (`frame`) and the length of the data (`frame_size`). No application data
+// is associated with the frame in this example, so the `user_priv`
+// parameter is NULL. The `deadline` parameter is left at zero for this
+// example. This parameter is generally only used when doing adaptive post
+// processing.
+//
+// Codecs may produce a variable number of output frames for every call to
+// `aom_codec_decode`. These frames are retrieved by the
+// `aom_codec_get_frame` iterator function. The iterator variable `iter` is
+// initialized to NULL each time `aom_codec_decode` is called.
+// `aom_codec_get_frame` is called in a loop, returning a pointer to a
+// decoded image or NULL to indicate the end of list.
+//
+// Processing The Decoded Data
+// ---------------------------
+// In this example, we simply write the encoded data to disk. It is
+// important to honor the image's `stride` values.
+//
+// Cleanup
+// -------
+// The `aom_codec_destroy` call frees any memory allocated by the codec.
+//
+// Error Handling
+// --------------
+// This example does not special case any error return codes. If there was
+// an error, a descriptive message is printed and the program exits. With
+// few exceptions, aom_codec functions return an enumerated error status,
+// with the value `0` indicating success.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aom_decoder.h"
+
+#include "../tools_common.h"
+#include "../video_reader.h"
+#include "./aom_config.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr, "Usage: %s <infile> <outfile>\n", exec_name);
+  exit(EXIT_FAILURE);
+}
+
+int main(int argc, char **argv) {
+  int frame_cnt = 0;
+  FILE *outfile = NULL;
+  aom_codec_ctx_t codec;
+  AvxVideoReader *reader = NULL;
+  const AvxInterface *decoder = NULL;
+  const AvxVideoInfo *info = NULL;
+
+  exec_name = argv[0];
+
+  if (argc != 3) die("Invalid number of arguments.");
+
+  reader = aom_video_reader_open(argv[1]);
+  if (!reader) die("Failed to open %s for reading.", argv[1]);
+
+  if (!(outfile = fopen(argv[2], "wb")))
+    die("Failed to open %s for writing.", argv[2]);
+
+  info = aom_video_reader_get_info(reader);
+
+  decoder = get_aom_decoder_by_fourcc(info->codec_fourcc);
+  if (!decoder) die("Unknown input codec.");
+
+  printf("Using %s\n", aom_codec_iface_name(decoder->codec_interface()));
+
+  if (aom_codec_dec_init(&codec, decoder->codec_interface(), NULL, 0))
+    die_codec(&codec, "Failed to initialize decoder.");
+
+  while (aom_video_reader_read_frame(reader)) {
+    aom_codec_iter_t iter = NULL;
+    aom_image_t *img = NULL;
+    size_t frame_size = 0;
+    const unsigned char *frame =
+        aom_video_reader_get_frame(reader, &frame_size);
+    if (aom_codec_decode(&codec, frame, (unsigned int)frame_size, NULL, 0))
+      die_codec(&codec, "Failed to decode frame.");
+
+    while ((img = aom_codec_get_frame(&codec, &iter)) != NULL) {
+      aom_img_write(img, outfile);
+      ++frame_cnt;
+    }
+  }
+
+  printf("Processed %d frames.\n", frame_cnt);
+  if (aom_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec");
+
+  printf("Play: ffplay -f rawvideo -pix_fmt yuv420p -s %dx%d %s\n",
+         info->frame_width, info->frame_height, argv[2]);
+
+  aom_video_reader_close(reader);
+
+  fclose(outfile);
+
+  return EXIT_SUCCESS;
+}
diff --git a/third_party/aom/examples/simple_encoder.c b/third_party/aom/examples/simple_encoder.c
new file mode 100644
index 0000000000..996f6dacf1
--- /dev/null
+++ b/third_party/aom/examples/simple_encoder.c
@@ -0,0 +1,250 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// Simple Encoder
+// ==============
+//
+// This is an example of a simple encoder loop. It takes an input file in
+// YV12 format, passes it through the encoder, and writes the compressed
+// frames to disk in IVF format. Other decoder examples build upon this
+// one.
+//
+// The details of the IVF format have been elided from this example for
+// simplicity of presentation, as IVF files will not generally be used by
+// your application. In general, an IVF file consists of a file header,
+// followed by a variable number of frames. Each frame consists of a frame
+// header followed by a variable length payload. The length of the payload
+// is specified in the first four bytes of the frame header. The payload is
+// the raw compressed data.
+//
+// Standard Includes
+// -----------------
+// For encoders, you only have to include `aom_encoder.h` and then any
+// header files for the specific codecs you use. In this case, we're using
+// aom.
+//
+// Getting The Default Configuration
+// ---------------------------------
+// Encoders have the notion of "usage profiles." For example, an encoder
+// may want to publish default configurations for both a video
+// conferencing application and a best quality offline encoder. These
+// obviously have very different default settings. Consult the
+// documentation for your codec to see if it provides any default
+// configurations. All codecs provide a default configuration, number 0,
+// which is valid for material in the vacinity of QCIF/QVGA.
+//
+// Updating The Configuration
+// ---------------------------------
+// Almost all applications will want to update the default configuration
+// with settings specific to their usage. Here we set the width and height
+// of the video file to that specified on the command line. We also scale
+// the default bitrate based on the ratio between the default resolution
+// and the resolution specified on the command line.
+//
+// Initializing The Codec
+// ----------------------
+// The encoder is initialized by the following code.
+//
+// Encoding A Frame
+// ----------------
+// The frame is read as a continuous block (size width * height * 3 / 2)
+// from the input file. If a frame was read (the input file has not hit
+// EOF) then the frame is passed to the encoder. Otherwise, a NULL
+// is passed, indicating the End-Of-Stream condition to the encoder. The
+// `frame_cnt` is reused as the presentation time stamp (PTS) and each
+// frame is shown for one frame-time in duration. The flags parameter is
+// unused in this example.
+
+// Forced Keyframes
+// ----------------
+// Keyframes can be forced by setting the AOM_EFLAG_FORCE_KF bit of the
+// flags passed to `aom_codec_control()`. In this example, we force a
+// keyframe every <keyframe-interval> frames. Note, the output stream can
+// contain additional keyframes beyond those that have been forced using the
+// AOM_EFLAG_FORCE_KF flag because of automatic keyframe placement by the
+// encoder.
+//
+// Processing The Encoded Data
+// ---------------------------
+// Each packet of type `AOM_CODEC_CX_FRAME_PKT` contains the encoded data
+// for this frame. We write a IVF frame header, followed by the raw data.
+//
+// Cleanup
+// -------
+// The `aom_codec_destroy` call frees any memory allocated by the codec.
+//
+// Error Handling
+// --------------
+// This example does not special case any error return codes. If there was
+// an error, a descriptive message is printed and the program exits. With
+// few exeptions, aom_codec functions return an enumerated error status,
+// with the value `0` indicating success.
+//
+// Error Resiliency Features
+// -------------------------
+// Error resiliency is controlled by the g_error_resilient member of the
+// configuration structure. Use the `decode_with_drops` example to decode with
+// frames 5-10 dropped. Compare the output for a file encoded with this example
+// versus one encoded with the `simple_encoder` example.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aom_encoder.h"
+
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr,
+          "Usage: %s <codec> <width> <height> <infile> <outfile> "
+          "<keyframe-interval> <error-resilient> <frames to encode>\n"
+          "See comments in simple_encoder.c for more information.\n",
+          exec_name);
+  exit(EXIT_FAILURE);
+}
+
+static int encode_frame(aom_codec_ctx_t *codec, aom_image_t *img,
+                        int frame_index, int flags, AvxVideoWriter *writer) {
+  int got_pkts = 0;
+  aom_codec_iter_t iter = NULL;
+  const aom_codec_cx_pkt_t *pkt = NULL;
+  const aom_codec_err_t res =
+      aom_codec_encode(codec, img, frame_index, 1, flags, AOM_DL_GOOD_QUALITY);
+  if (res != AOM_CODEC_OK) die_codec(codec, "Failed to encode frame");
+
+  while ((pkt = aom_codec_get_cx_data(codec, &iter)) != NULL) {
+    got_pkts = 1;
+
+    if (pkt->kind == AOM_CODEC_CX_FRAME_PKT) {
+      const int keyframe = (pkt->data.frame.flags & AOM_FRAME_IS_KEY) != 0;
+      if (!aom_video_writer_write_frame(writer, pkt->data.frame.buf,
+                                        pkt->data.frame.sz,
+                                        pkt->data.frame.pts)) {
+        die_codec(codec, "Failed to write compressed frame");
+      }
+      printf(keyframe ? "K" : ".");
+      fflush(stdout);
+    }
+  }
+
+  return got_pkts;
+}
+
+// TODO(tomfinegan): Improve command line parsing and add args for bitrate/fps.
+int main(int argc, char **argv) {
+  FILE *infile = NULL;
+  aom_codec_ctx_t codec;
+  aom_codec_enc_cfg_t cfg;
+  int frame_count = 0;
+  aom_image_t raw;
+  aom_codec_err_t res;
+  AvxVideoInfo info;
+  AvxVideoWriter *writer = NULL;
+  const AvxInterface *encoder = NULL;
+  const int fps = 30;
+  const int bitrate = 200;
+  int keyframe_interval = 0;
+  int max_frames = 0;
+  int frames_encoded = 0;
+  const char *codec_arg = NULL;
+  const char *width_arg = NULL;
+  const char *height_arg = NULL;
+  const char *infile_arg = NULL;
+  const char *outfile_arg = NULL;
+  const char *keyframe_interval_arg = NULL;
+
+  exec_name = argv[0];
+
+  // Clear explicitly, as simply assigning "{ 0 }" generates
+  // "missing-field-initializers" warning in some compilers.
+  memset(&info, 0, sizeof(info));
+
+  if (argc != 9) die("Invalid number of arguments");
+
+  codec_arg = argv[1];
+  width_arg = argv[2];
+  height_arg = argv[3];
+  infile_arg = argv[4];
+  outfile_arg = argv[5];
+  keyframe_interval_arg = argv[6];
+  max_frames = (int)strtol(argv[8], NULL, 0);
+
+  encoder = get_aom_encoder_by_name(codec_arg);
+  if (!encoder) die("Unsupported codec.");
+
+  info.codec_fourcc = encoder->fourcc;
+  info.frame_width = (int)strtol(width_arg, NULL, 0);
+  info.frame_height = (int)strtol(height_arg, NULL, 0);
+  info.time_base.numerator = 1;
+  info.time_base.denominator = fps;
+
+  if (info.frame_width <= 0 || info.frame_height <= 0 ||
+      (info.frame_width % 2) != 0 || (info.frame_height % 2) != 0) {
+    die("Invalid frame size: %dx%d", info.frame_width, info.frame_height);
+  }
+
+  if (!aom_img_alloc(&raw, AOM_IMG_FMT_I420, info.frame_width,
+                     info.frame_height, 1)) {
+    die("Failed to allocate image.");
+  }
+
+  keyframe_interval = (int)strtol(keyframe_interval_arg, NULL, 0);
+  if (keyframe_interval < 0) die("Invalid keyframe interval value.");
+
+  printf("Using %s\n", aom_codec_iface_name(encoder->codec_interface()));
+
+  res = aom_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res) die_codec(&codec, "Failed to get default codec config.");
+
+  cfg.g_w = info.frame_width;
+  cfg.g_h = info.frame_height;
+  cfg.g_timebase.num = info.time_base.numerator;
+  cfg.g_timebase.den = info.time_base.denominator;
+  cfg.rc_target_bitrate = bitrate;
+  cfg.g_error_resilient = (aom_codec_er_flags_t)strtoul(argv[7], NULL, 0);
+
+  writer = aom_video_writer_open(outfile_arg, kContainerIVF, &info);
+  if (!writer) die("Failed to open %s for writing.", outfile_arg);
+
+  if (!(infile = fopen(infile_arg, "rb")))
+    die("Failed to open %s for reading.", infile_arg);
+
+  if (aom_codec_enc_init(&codec, encoder->codec_interface(), &cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  // Encode frames.
+  while (aom_img_read(&raw, infile)) {
+    int flags = 0;
+    if (keyframe_interval > 0 && frame_count % keyframe_interval == 0)
+      flags |= AOM_EFLAG_FORCE_KF;
+    encode_frame(&codec, &raw, frame_count++, flags, writer);
+    frames_encoded++;
+    if (max_frames > 0 && frames_encoded >= max_frames) break;
+  }
+
+  // Flush encoder.
+  while (encode_frame(&codec, NULL, -1, 0, writer)) continue;
+
+  printf("\n");
+  fclose(infile);
+  printf("Processed %d frames.\n", frame_count);
+
+  aom_img_free(&raw);
+  if (aom_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec.");
+
+  aom_video_writer_close(writer);
+
+  return EXIT_SUCCESS;
+}
diff --git a/third_party/aom/examples/twopass_encoder.c b/third_party/aom/examples/twopass_encoder.c
new file mode 100644
index 0000000000..e767bb5d72
--- /dev/null
+++ b/third_party/aom/examples/twopass_encoder.c
@@ -0,0 +1,255 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// Two Pass Encoder
+// ================
+//
+// This is an example of a two pass encoder loop. It takes an input file in
+// YV12 format, passes it through the encoder twice, and writes the compressed
+// frames to disk in IVF format. It builds upon the simple_encoder example.
+//
+// Twopass Variables
+// -----------------
+// Twopass mode needs to track the current pass number and the buffer of
+// statistics packets.
+//
+// Updating The Configuration
+// ---------------------------------
+// In two pass mode, the configuration has to be updated on each pass. The
+// statistics buffer is passed on the last pass.
+//
+// Encoding A Frame
+// ----------------
+// Encoding a frame in two pass mode is identical to the simple encoder
+// example.
+//
+// Processing Statistics Packets
+// -----------------------------
+// Each packet of type `AOM_CODEC_CX_FRAME_PKT` contains the encoded data
+// for this frame. We write a IVF frame header, followed by the raw data.
+//
+//
+// Pass Progress Reporting
+// -----------------------------
+// It's sometimes helpful to see when each pass completes.
+//
+//
+// Clean-up
+// -----------------------------
+// Destruction of the encoder instance must be done on each pass. The
+// raw image should be destroyed at the end as usual.
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aom_encoder.h"
+
+#include "../tools_common.h"
+#include "../video_writer.h"
+
+static const char *exec_name;
+
+void usage_exit(void) {
+  fprintf(stderr,
+          "Usage: %s <codec> <width> <height> <infile> <outfile> "
+          "<limit(optional)>\n",
+          exec_name);
+  exit(EXIT_FAILURE);
+}
+
+static int get_frame_stats(aom_codec_ctx_t *ctx, const aom_image_t *img,
+                           aom_codec_pts_t pts, unsigned int duration,
+                           aom_enc_frame_flags_t flags, unsigned int deadline,
+                           aom_fixed_buf_t *stats) {
+  int got_pkts = 0;
+  aom_codec_iter_t iter = NULL;
+  const aom_codec_cx_pkt_t *pkt = NULL;
+  const aom_codec_err_t res =
+      aom_codec_encode(ctx, img, pts, duration, flags, deadline);
+  if (res != AOM_CODEC_OK) die_codec(ctx, "Failed to get frame stats.");
+
+  while ((pkt = aom_codec_get_cx_data(ctx, &iter)) != NULL) {
+    got_pkts = 1;
+
+    if (pkt->kind == AOM_CODEC_STATS_PKT) {
+      const uint8_t *const pkt_buf = pkt->data.twopass_stats.buf;
+      const size_t pkt_size = pkt->data.twopass_stats.sz;
+      stats->buf = realloc(stats->buf, stats->sz + pkt_size);
+      memcpy((uint8_t *)stats->buf + stats->sz, pkt_buf, pkt_size);
+      stats->sz += pkt_size;
+    }
+  }
+
+  return got_pkts;
+}
+
+static int encode_frame(aom_codec_ctx_t *ctx, const aom_image_t *img,
+                        aom_codec_pts_t pts, unsigned int duration,
+                        aom_enc_frame_flags_t flags, unsigned int deadline,
+                        AvxVideoWriter *writer) {
+  int got_pkts = 0;
+  aom_codec_iter_t iter = NULL;
+  const aom_codec_cx_pkt_t *pkt = NULL;
+  const aom_codec_err_t res =
+      aom_codec_encode(ctx, img, pts, duration, flags, deadline);
+  if (res != AOM_CODEC_OK) die_codec(ctx, "Failed to encode frame.");
+
+  while ((pkt = aom_codec_get_cx_data(ctx, &iter)) != NULL) {
+    got_pkts = 1;
+    if (pkt->kind == AOM_CODEC_CX_FRAME_PKT) {
+      const int keyframe = (pkt->data.frame.flags & AOM_FRAME_IS_KEY) != 0;
+
+      if (!aom_video_writer_write_frame(writer, pkt->data.frame.buf,
+                                        pkt->data.frame.sz,
+                                        pkt->data.frame.pts))
+        die_codec(ctx, "Failed to write compressed frame.");
+      printf(keyframe ? "K" : ".");
+      fflush(stdout);
+    }
+  }
+
+  return got_pkts;
+}
+
+static aom_fixed_buf_t pass0(aom_image_t *raw, FILE *infile,
+                             const AvxInterface *encoder,
+                             const aom_codec_enc_cfg_t *cfg, int limit) {
+  aom_codec_ctx_t codec;
+  int frame_count = 0;
+  aom_fixed_buf_t stats = { NULL, 0 };
+
+  if (aom_codec_enc_init(&codec, encoder->codec_interface(), cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  // Calculate frame statistics.
+  while (aom_img_read(raw, infile) && frame_count < limit) {
+    ++frame_count;
+    get_frame_stats(&codec, raw, frame_count, 1, 0, AOM_DL_GOOD_QUALITY,
+                    &stats);
+  }
+
+  // Flush encoder.
+  while (get_frame_stats(&codec, NULL, frame_count, 1, 0, AOM_DL_GOOD_QUALITY,
+                         &stats)) {
+  }
+
+  printf("Pass 0 complete. Processed %d frames.\n", frame_count);
+  if (aom_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec.");
+
+  return stats;
+}
+
+static void pass1(aom_image_t *raw, FILE *infile, const char *outfile_name,
+                  const AvxInterface *encoder, const aom_codec_enc_cfg_t *cfg,
+                  int limit) {
+  AvxVideoInfo info = { encoder->fourcc,
+                        cfg->g_w,
+                        cfg->g_h,
+                        { cfg->g_timebase.num, cfg->g_timebase.den } };
+  AvxVideoWriter *writer = NULL;
+  aom_codec_ctx_t codec;
+  int frame_count = 0;
+
+  writer = aom_video_writer_open(outfile_name, kContainerIVF, &info);
+  if (!writer) die("Failed to open %s for writing", outfile_name);
+
+  if (aom_codec_enc_init(&codec, encoder->codec_interface(), cfg, 0))
+    die_codec(&codec, "Failed to initialize encoder");
+
+  // Encode frames.
+  while (aom_img_read(raw, infile) && frame_count < limit) {
+    ++frame_count;
+    encode_frame(&codec, raw, frame_count, 1, 0, AOM_DL_GOOD_QUALITY, writer);
+  }
+
+  // Flush encoder.
+  while (encode_frame(&codec, NULL, -1, 1, 0, AOM_DL_GOOD_QUALITY, writer)) {
+  }
+
+  printf("\n");
+
+  if (aom_codec_destroy(&codec)) die_codec(&codec, "Failed to destroy codec.");
+
+  aom_video_writer_close(writer);
+
+  printf("Pass 1 complete. Processed %d frames.\n", frame_count);
+}
+
+int main(int argc, char **argv) {
+  FILE *infile = NULL;
+  int w, h;
+  aom_codec_ctx_t codec;
+  aom_codec_enc_cfg_t cfg;
+  aom_image_t raw;
+  aom_codec_err_t res;
+  aom_fixed_buf_t stats;
+
+  const AvxInterface *encoder = NULL;
+  const int fps = 30;       // TODO(dkovalev) add command line argument
+  const int bitrate = 200;  // kbit/s TODO(dkovalev) add command line argument
+  const char *const codec_arg = argv[1];
+  const char *const width_arg = argv[2];
+  const char *const height_arg = argv[3];
+  const char *const infile_arg = argv[4];
+  const char *const outfile_arg = argv[5];
+  int limit = 0;
+  exec_name = argv[0];
+
+  if (argc < 6) die("Invalid number of arguments");
+
+  if (argc > 6) limit = (int)strtol(argv[6], NULL, 0);
+
+  if (limit == 0) limit = 100;
+
+  encoder = get_aom_encoder_by_name(codec_arg);
+  if (!encoder) die("Unsupported codec.");
+
+  w = (int)strtol(width_arg, NULL, 0);
+  h = (int)strtol(height_arg, NULL, 0);
+
+  if (w <= 0 || h <= 0 || (w % 2) != 0 || (h % 2) != 0)
+    die("Invalid frame size: %dx%d", w, h);
+
+  if (!aom_img_alloc(&raw, AOM_IMG_FMT_I420, w, h, 1))
+    die("Failed to allocate image", w, h);
+
+  printf("Using %s\n", aom_codec_iface_name(encoder->codec_interface()));
+
+  // Configuration
+  res = aom_codec_enc_config_default(encoder->codec_interface(), &cfg, 0);
+  if (res) die_codec(&codec, "Failed to get default codec config.");
+
+  cfg.g_w = w;
+  cfg.g_h = h;
+  cfg.g_timebase.num = 1;
+  cfg.g_timebase.den = fps;
+  cfg.rc_target_bitrate = bitrate;
+
+  if (!(infile = fopen(infile_arg, "rb")))
+    die("Failed to open %s for reading", infile_arg);
+
+  // Pass 0
+  cfg.g_pass = AOM_RC_FIRST_PASS;
+  stats = pass0(&raw, infile, encoder, &cfg, limit);
+
+  // Pass 1
+  rewind(infile);
+  cfg.g_pass = AOM_RC_LAST_PASS;
+  cfg.rc_twopass_stats_in = stats;
+  pass1(&raw, infile, outfile_arg, encoder, &cfg, limit);
+  free(stats.buf);
+
+  aom_img_free(&raw);
+  fclose(infile);
+
+  return EXIT_SUCCESS;
+}
diff --git a/third_party/aom/ivfdec.c b/third_party/aom/ivfdec.c
new file mode 100644
index 0000000000..fc11b95440
--- /dev/null
+++ b/third_party/aom/ivfdec.c
@@ -0,0 +1,102 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom_ports/mem_ops.h"
+
+#include "./ivfdec.h"
+
+static const char *IVF_SIGNATURE = "DKIF";
+
+static void fix_framerate(int *num, int *den) {
+  if (*den <= 0 || *den >= 1000000000 || *num <= 0 || *num >= 1000) {
+    // framerate seems to be invalid, just default to 30fps.
+    *num = 30;
+    *den = 1;
+  }
+}
+
+int file_is_ivf(struct AvxInputContext *input_ctx) {
+  char raw_hdr[32];
+  int is_ivf = 0;
+
+  if (fread(raw_hdr, 1, 32, input_ctx->file) == 32) {
+    if (memcmp(IVF_SIGNATURE, raw_hdr, 4) == 0) {
+      is_ivf = 1;
+
+      if (mem_get_le16(raw_hdr + 4) != 0) {
+        fprintf(stderr,
+                "Error: Unrecognized IVF version! This file may not"
+                " decode properly.");
+      }
+
+      input_ctx->fourcc = mem_get_le32(raw_hdr + 8);
+      input_ctx->width = mem_get_le16(raw_hdr + 12);
+      input_ctx->height = mem_get_le16(raw_hdr + 14);
+      input_ctx->framerate.numerator = mem_get_le32(raw_hdr + 16);
+      input_ctx->framerate.denominator = mem_get_le32(raw_hdr + 20);
+      fix_framerate(&input_ctx->framerate.numerator,
+                    &input_ctx->framerate.denominator);
+    }
+  }
+
+  if (!is_ivf) {
+    rewind(input_ctx->file);
+    input_ctx->detect.buf_read = 0;
+  } else {
+    input_ctx->detect.position = 4;
+  }
+  return is_ivf;
+}
+
+int ivf_read_frame(FILE *infile, uint8_t **buffer, size_t *bytes_read,
+                   size_t *buffer_size) {
+  char raw_header[IVF_FRAME_HDR_SZ] = { 0 };
+  size_t frame_size = 0;
+
+  if (fread(raw_header, IVF_FRAME_HDR_SZ, 1, infile) != 1) {
+    if (!feof(infile)) warn("Failed to read frame size\n");
+  } else {
+    frame_size = mem_get_le32(raw_header);
+
+    if (frame_size > 256 * 1024 * 1024) {
+      warn("Read invalid frame size (%u)\n", (unsigned int)frame_size);
+      frame_size = 0;
+    }
+
+    if (frame_size > *buffer_size) {
+      uint8_t *new_buffer = (uint8_t *)realloc(*buffer, 2 * frame_size);
+
+      if (new_buffer) {
+        *buffer = new_buffer;
+        *buffer_size = 2 * frame_size;
+      } else {
+        warn("Failed to allocate compressed data buffer\n");
+        frame_size = 0;
+      }
+    }
+  }
+
+  if (!feof(infile)) {
+    if (fread(*buffer, 1, frame_size, infile) != frame_size) {
+      warn("Failed to read full frame\n");
+      return 1;
+    }
+
+    *bytes_read = frame_size;
+    return 0;
+  }
+
+  return 1;
+}
diff --git a/third_party/aom/ivfdec.h b/third_party/aom/ivfdec.h
new file mode 100644
index 0000000000..36a6fb84e6
--- /dev/null
+++ b/third_party/aom/ivfdec.h
@@ -0,0 +1,29 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef IVFDEC_H_
+#define IVFDEC_H_
+
+#include "./tools_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int file_is_ivf(struct AvxInputContext *input);
+
+int ivf_read_frame(FILE *infile, uint8_t **buffer, size_t *bytes_read,
+                   size_t *buffer_size);
+
+#ifdef __cplusplus
+} /* extern "C" */
+#endif
+
+#endif  // IVFDEC_H_
diff --git a/third_party/aom/ivfenc.c b/third_party/aom/ivfenc.c
new file mode 100644
index 0000000000..80f4d14e39
--- /dev/null
+++ b/third_party/aom/ivfenc.c
@@ -0,0 +1,52 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./ivfenc.h"
+
+#include "aom/aom_encoder.h"
+#include "aom_ports/mem_ops.h"
+
+void ivf_write_file_header(FILE *outfile, const struct aom_codec_enc_cfg *cfg,
+                           unsigned int fourcc, int frame_cnt) {
+  char header[32];
+
+  header[0] = 'D';
+  header[1] = 'K';
+  header[2] = 'I';
+  header[3] = 'F';
+  mem_put_le16(header + 4, 0);                     // version
+  mem_put_le16(header + 6, 32);                    // header size
+  mem_put_le32(header + 8, fourcc);                // fourcc
+  mem_put_le16(header + 12, cfg->g_w);             // width
+  mem_put_le16(header + 14, cfg->g_h);             // height
+  mem_put_le32(header + 16, cfg->g_timebase.den);  // rate
+  mem_put_le32(header + 20, cfg->g_timebase.num);  // scale
+  mem_put_le32(header + 24, frame_cnt);            // length
+  mem_put_le32(header + 28, 0);                    // unused
+
+  fwrite(header, 1, 32, outfile);
+}
+
+void ivf_write_frame_header(FILE *outfile, int64_t pts, size_t frame_size) {
+  char header[12];
+
+  mem_put_le32(header, (int)frame_size);
+  mem_put_le32(header + 4, (int)(pts & 0xFFFFFFFF));
+  mem_put_le32(header + 8, (int)(pts >> 32));
+  fwrite(header, 1, 12, outfile);
+}
+
+void ivf_write_frame_size(FILE *outfile, size_t frame_size) {
+  char header[4];
+
+  mem_put_le32(header, (int)frame_size);
+  fwrite(header, 1, 4, outfile);
+}
diff --git a/third_party/aom/ivfenc.h b/third_party/aom/ivfenc.h
new file mode 100644
index 0000000000..62b3a91501
--- /dev/null
+++ b/third_party/aom/ivfenc.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef IVFENC_H_
+#define IVFENC_H_
+
+#include "./tools_common.h"
+
+struct aom_codec_enc_cfg;
+struct aom_codec_cx_pkt;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ivf_write_file_header(FILE *outfile, const struct aom_codec_enc_cfg *cfg,
+                           uint32_t fourcc, int frame_cnt);
+
+void ivf_write_frame_header(FILE *outfile, int64_t pts, size_t frame_size);
+
+void ivf_write_frame_size(FILE *outfile, size_t frame_size);
+
+#ifdef __cplusplus
+} /* extern "C" */
+#endif
+
+#endif  // IVFENC_H_
diff --git a/third_party/aom/keywords.dox b/third_party/aom/keywords.dox
new file mode 100644
index 0000000000..56f5368900
--- /dev/null
+++ b/third_party/aom/keywords.dox
@@ -0,0 +1,51 @@
+/*!\page rfc2119 RFC2119 Keywords
+
+      The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL
+      NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED",  "MAY", and
+      "OPTIONAL" in this document are to be interpreted as described in
+      <a href="http://www.ietf.org/rfc/rfc2119.txt">RFC 2119.</a>
+
+Specifically, the following definitions are used:
+
+\section MUST
+\anchor REQUIRED
+\anchor SHALL
+   This word, or the terms "REQUIRED" or "SHALL", mean that the
+   definition is an absolute requirement of the specification.
+
+\section MUSTNOT MUST NOT
+\anchor SHALLNOT
+   This phrase, or the phrase "SHALL NOT", mean that the
+   definition is an absolute prohibition of the specification.
+
+\section SHOULD
+\anchor RECOMMENDED
+   This word, or the adjective "RECOMMENDED", mean that there
+   may exist valid reasons in particular circumstances to ignore a
+   particular item, but the full implications must be understood and
+   carefully weighed before choosing a different course.
+
+\section SHOULDNOT SHOULD NOT
+\anchor NOTRECOMMENDED
+   This phrase, or the phrase "NOT RECOMMENDED" mean that
+   there may exist valid reasons in particular circumstances when the
+   particular behavior is acceptable or even useful, but the full
+   implications should be understood and the case carefully weighed
+   before implementing any behavior described with this label.
+
+\section MAY
+\anchor OPTIONAL
+   This word, or the adjective "OPTIONAL", mean that an item is
+   truly optional.  One vendor may choose to include the item because a
+   particular marketplace requires it or because the vendor feels that
+   it enhances the product while another vendor may omit the same item.
+   An implementation which does not include a particular option \ref MUST be
+   prepared to interoperate with another implementation which does
+   include the option, though perhaps with reduced functionality. In the
+   same vein an implementation which does include a particular option
+   \ref MUST be prepared to interoperate with another implementation which
+   does not include the option (except, of course, for the feature the
+   option provides.)
+
+
+*/
diff --git a/third_party/aom/libs.doxy_template b/third_party/aom/libs.doxy_template
new file mode 100644
index 0000000000..23400b4299
--- /dev/null
+++ b/third_party/aom/libs.doxy_template
@@ -0,0 +1,1295 @@
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+# Doxyfile 1.5.4
+
+# This file describes the settings to be used by the documentation system
+# doxygen (www.doxygen.org) for a project
+#
+# All text after a hash (#) is considered a comment and will be ignored
+# The format is:
+#       TAG = value [value, ...]
+# For lists items can also be appended using:
+#       TAG += value [value, ...]
+# Values that contain spaces should be placed between quotes (" ")
+
+#---------------------------------------------------------------------------
+# Project related configuration options
+#---------------------------------------------------------------------------
+
+# This tag specifies the encoding used for all characters in the config file that
+# follow. The default is UTF-8 which is also the encoding used for all text before
+# the first occurrence of this tag. Doxygen uses libiconv (or the iconv built into
+# libc) for the transcoding. See http://www.gnu.org/software/libiconv for the list of
+# possible encodings.
+
+DOXYFILE_ENCODING      = UTF-8
+
+# The PROJECT_NAME tag is a single word (or a sequence of words surrounded
+# by quotes) that should identify the project.
+
+PROJECT_NAME           = "AOMedia Codec SDK"
+
+# The OUTPUT_DIRECTORY tag is used to specify the (relative or absolute)
+# base path where the generated documentation will be put.
+# If a relative path is entered, it will be relative to the location
+# where doxygen was started. If left blank the current directory will be used.
+
+OUTPUT_DIRECTORY       = docs
+
+# If the CREATE_SUBDIRS tag is set to YES, then doxygen will create
+# 4096 sub-directories (in 2 levels) under the output directory of each output
+# format and will distribute the generated files over these directories.
+# Enabling this option can be useful when feeding doxygen a huge amount of
+# source files, where putting all generated files in the same directory would
+# otherwise cause performance problems for the file system.
+
+CREATE_SUBDIRS         = NO
+
+# The OUTPUT_LANGUAGE tag is used to specify the language in which all
+# documentation generated by doxygen is written. Doxygen will use this
+# information to generate all constant output in the proper language.
+# The default language is English, other supported languages are:
+# Afrikaans, Arabic, Brazilian, Catalan, Chinese, Chinese-Traditional,
+# Croatian, Czech, Danish, Dutch, Finnish, French, German, Greek, Hungarian,
+# Italian, Japanese, Japanese-en (Japanese with English messages), Korean,
+# Korean-en, Lithuanian, Norwegian, Polish, Portuguese, Romanian, Russian,
+# Serbian, Slovak, Slovene, Spanish, Swedish, and Ukrainian.
+
+OUTPUT_LANGUAGE        = English
+
+# If the BRIEF_MEMBER_DESC tag is set to YES (the default) Doxygen will
+# include brief member descriptions after the members that are listed in
+# the file and class documentation (similar to java_doc).
+# Set to NO to disable this.
+
+BRIEF_MEMBER_DESC      = YES
+
+# If the REPEAT_BRIEF tag is set to YES (the default) Doxygen will prepend
+# the brief description of a member or function before the detailed description.
+# Note: if both HIDE_UNDOC_MEMBERS and BRIEF_MEMBER_DESC are set to NO, the
+# brief descriptions will be completely suppressed.
+
+REPEAT_BRIEF           = YES
+
+# This tag implements a quasi-intelligent brief description abbreviator
+# that is used to form the text in various listings. Each string
+# in this list, if found as the leading text of the brief description, will be
+# stripped from the text and the result after processing the whole list, is
+# used as the annotated text. Otherwise, the brief description is used as-is.
+# If left blank, the following values are used ("$name" is automatically
+# replaced with the name of the entity): "The $name class" "The $name widget"
+# "The $name file" "is" "provides" "specifies" "contains"
+# "represents" "a" "an" "the"
+
+ABBREVIATE_BRIEF       =
+
+# If the ALWAYS_DETAILED_SEC and REPEAT_BRIEF tags are both set to YES then
+# Doxygen will generate a detailed section even if there is only a brief
+# description.
+
+ALWAYS_DETAILED_SEC    = NO
+
+# If the INLINE_INHERITED_MEMB tag is set to YES, doxygen will show all
+# inherited members of a class in the documentation of that class as if those
+# members were ordinary class members. Constructors, destructors and assignment
+# operators of the base classes will not be shown.
+
+INLINE_INHERITED_MEMB  = NO
+
+# If the FULL_PATH_NAMES tag is set to YES then Doxygen will prepend the full
+# path before files name in the file list and in the header files. If set
+# to NO the shortest path that makes the file name unique will be used.
+
+FULL_PATH_NAMES        = YES
+
+# If the FULL_PATH_NAMES tag is set to YES then the STRIP_FROM_PATH tag
+# can be used to strip a user-defined part of the path. Stripping is
+# only done if one of the specified strings matches the left-hand part of
+# the path. The tag can be used to show relative paths in the file list.
+# If left blank the directory from which doxygen is run is used as the
+# path to strip.
+
+STRIP_FROM_PATH        =
+
+# The STRIP_FROM_INC_PATH tag can be used to strip a user-defined part of
+# the path mentioned in the documentation of a class, which tells
+# the reader which header file to include in order to use a class.
+# If left blank only the name of the header file containing the class
+# definition is used. Otherwise one should specify the include paths that
+# are normally passed to the compiler using the -I flag.
+
+STRIP_FROM_INC_PATH    =
+
+# If the SHORT_NAMES tag is set to YES, doxygen will generate much shorter
+# (but less readable) file names. This can be useful is your file systems
+# doesn't support long names like on DOS, Mac, or CD-ROM.
+
+SHORT_NAMES            = NO
+
+# If the JAVADOC_AUTOBRIEF tag is set to YES then Doxygen
+# will interpret the first line (until the first dot) of a java_doc-style
+# comment as the brief description. If set to NO, the java_doc
+# comments will behave just like regular Qt-style comments
+# (thus requiring an explicit @brief command for a brief description.)
+
+JAVADOC_AUTOBRIEF      = NO
+
+# If the QT_AUTOBRIEF tag is set to YES then Doxygen will
+# interpret the first line (until the first dot) of a Qt-style
+# comment as the brief description. If set to NO, the comments
+# will behave just like regular Qt-style comments (thus requiring
+# an explicit \brief command for a brief description.)
+
+QT_AUTOBRIEF           = NO
+
+# The MULTILINE_CPP_IS_BRIEF tag can be set to YES to make Doxygen
+# treat a multi-line C++ special comment block (i.e. a block of //! or ///
+# comments) as a brief description. This used to be the default behaviour.
+# The new default is to treat a multi-line C++ comment block as a detailed
+# description. Set this tag to YES if you prefer the old behaviour instead.
+
+MULTILINE_CPP_IS_BRIEF = NO
+
+# If the INHERIT_DOCS tag is set to YES (the default) then an undocumented
+# member inherits the documentation from any documented member that it
+# re-implements.
+
+INHERIT_DOCS           = YES
+
+# If the SEPARATE_MEMBER_PAGES tag is set to YES, then doxygen will produce
+# a new page for each member. If set to NO, the documentation of a member will
+# be part of the file/class/namespace that contains it.
+
+SEPARATE_MEMBER_PAGES  = NO
+
+# The TAB_SIZE tag can be used to set the number of spaces in a tab.
+# Doxygen uses this value to replace tabs by spaces in code fragments.
+
+TAB_SIZE               = 4
+
+# This tag can be used to specify a number of aliases that acts
+# as commands in the documentation. An alias has the form "name=value".
+# For example adding "sideeffect=\par Side Effects:\n" will allow you to
+# put the command \sideeffect (or @sideeffect) in the documentation, which
+# will result in a user-defined paragraph with heading "Side Effects:".
+# You can put \n's in the value part of an alias to insert newlines.
+
+ALIASES                =
+
+# Set the OPTIMIZE_OUTPUT_FOR_C tag to YES if your project consists of C
+# sources only. Doxygen will then generate output that is more tailored for C.
+# For instance, some of the names that are used will be different. The list
+# of all members will be omitted, etc.
+
+OPTIMIZE_OUTPUT_FOR_C  = YES
+
+# Set the OPTIMIZE_OUTPUT_JAVA tag to YES if your project consists of Java
+# sources only. Doxygen will then generate output that is more tailored for Java.
+# For instance, namespaces will be presented as packages, qualified scopes
+# will look different, etc.
+
+OPTIMIZE_OUTPUT_JAVA   = NO
+
+# If you use STL classes (i.e. std::string, std::vector, etc.) but do not want to
+# include (a tag file for) the STL sources as input, then you should
+# set this tag to YES in order to let doxygen match functions declarations and
+# definitions whose arguments contain STL classes (e.g. func(std::string); v.s.
+# func(std::string) {}). This also make the inheritance and collaboration
+# diagrams that involve STL classes more complete and accurate.
+
+BUILTIN_STL_SUPPORT    = NO
+
+# If you use Microsoft's C++/CLI language, you should set this option to YES to
+# enable parsing support.
+
+CPP_CLI_SUPPORT        = NO
+
+# Set the SIP_SUPPORT tag to YES if your project consists of sip sources only.
+# Doxygen will parse them like normal C++ but will assume all classes use public
+# instead of private inheritance when no explicit protection keyword is present.
+
+SIP_SUPPORT            = NO
+
+# If member grouping is used in the documentation and the DISTRIBUTE_GROUP_DOC
+# tag is set to YES, then doxygen will reuse the documentation of the first
+# member in the group (if any) for the other members of the group. By default
+# all members of a group must be documented explicitly.
+
+DISTRIBUTE_GROUP_DOC   = NO
+
+# Set the SUBGROUPING tag to YES (the defqault) to allow class member groups of
+# the same type (for instance a group of public functions) to be put as a
+# subgroup of that type (e.g. under the Public Functions section). Set it to
+# NO to prevent subgrouping. Alternatively, this can be done per class using
+# the \nosubgrouping command.
+
+SUBGROUPING            = YES
+
+# When TYPEDEF_HIDES_STRUCT is enabled, a typedef of a struct (or union) is
+# documented as struct with the name of the typedef. So
+# typedef struct type_s {} type_t, will appear in the documentation as a struct
+# with name type_t. When disabled the typedef will appear as a member of a file,
+# namespace, or class. And the struct will be named type_s. This can typically
+# be useful for C code where the coding convention is that all structs are
+# typedef'ed and only the typedef is referenced never the struct's name.
+
+TYPEDEF_HIDES_STRUCT   = NO
+
+#---------------------------------------------------------------------------
+# Build related configuration options
+#---------------------------------------------------------------------------
+
+# If the EXTRACT_ALL tag is set to YES doxygen will assume all entities in
+# documentation are documented, even if no documentation was available.
+# Private class members and static file members will be hidden unless
+# the EXTRACT_PRIVATE and EXTRACT_STATIC tags are set to YES
+
+EXTRACT_ALL            = NO
+
+# If the EXTRACT_PRIVATE tag is set to YES all private members of a class
+# will be included in the documentation.
+
+EXTRACT_PRIVATE        = NO
+
+# If the EXTRACT_STATIC tag is set to YES all static members of a file
+# will be included in the documentation.
+
+EXTRACT_STATIC         = NO
+
+# If the EXTRACT_LOCAL_CLASSES tag is set to YES classes (and structs)
+# defined locally in source files will be included in the documentation.
+# If set to NO only classes defined in header files are included.
+
+EXTRACT_LOCAL_CLASSES  = YES
+
+# This flag is only useful for Objective-C code. When set to YES local
+# methods, which are defined in the implementation section but not in
+# the interface are included in the documentation.
+# If set to NO (the default) only methods in the interface are included.
+
+EXTRACT_LOCAL_METHODS  = NO
+
+# If this flag is set to YES, the members of anonymous namespaces will be extracted
+# and appear in the documentation as a namespace called 'anonymous_namespace{file}',
+# where file will be replaced with the base name of the file that contains the anonymous
+# namespace. By default anonymous namespace are hidden.
+
+EXTRACT_ANON_NSPACES   = NO
+
+# If the HIDE_UNDOC_MEMBERS tag is set to YES, Doxygen will hide all
+# undocumented members of documented classes, files or namespaces.
+# If set to NO (the default) these members will be included in the
+# various overviews, but no documentation section is generated.
+# This option has no effect if EXTRACT_ALL is enabled.
+
+HIDE_UNDOC_MEMBERS     = NO
+
+# If the HIDE_UNDOC_CLASSES tag is set to YES, Doxygen will hide all
+# undocumented classes that are normally visible in the class hierarchy.
+# If set to NO (the default) these classes will be included in the various
+# overviews. This option has no effect if EXTRACT_ALL is enabled.
+
+HIDE_UNDOC_CLASSES     = NO
+
+# If the HIDE_FRIEND_COMPOUNDS tag is set to YES, Doxygen will hide all
+# friend (class|struct|union) declarations.
+# If set to NO (the default) these declarations will be included in the
+# documentation.
+
+HIDE_FRIEND_COMPOUNDS  = NO
+
+# If the HIDE_IN_BODY_DOCS tag is set to YES, Doxygen will hide any
+# documentation blocks found inside the body of a function.
+# If set to NO (the default) these blocks will be appended to the
+# function's detailed documentation block.
+
+HIDE_IN_BODY_DOCS      = NO
+
+# The INTERNAL_DOCS tag determines if documentation
+# that is typed after a \internal command is included. If the tag is set
+# to NO (the default) then the documentation will be excluded.
+# Set it to YES to include the internal documentation.
+
+INTERNAL_DOCS          = NO
+
+# If the CASE_SENSE_NAMES tag is set to NO then Doxygen will only generate
+# file names in lower-case letters. If set to YES upper-case letters are also
+# allowed. This is useful if you have classes or files whose names only differ
+# in case and if your file system supports case sensitive file names. Windows
+# and Mac users are advised to set this option to NO.
+
+CASE_SENSE_NAMES       = YES
+
+# If the HIDE_SCOPE_NAMES tag is set to NO (the default) then Doxygen
+# will show members with their full class and namespace scopes in the
+# documentation. If set to YES the scope will be hidden.
+
+HIDE_SCOPE_NAMES       = NO
+
+# If the SHOW_INCLUDE_FILES tag is set to YES (the default) then Doxygen
+# will put a list of the files that are included by a file in the documentation
+# of that file.
+
+SHOW_INCLUDE_FILES     = YES
+
+# If the INLINE_INFO tag is set to YES (the default) then a tag [inline]
+# is inserted in the documentation for inline members.
+
+INLINE_INFO            = YES
+
+# If the SORT_MEMBER_DOCS tag is set to YES (the default) then doxygen
+# will sort the (detailed) documentation of file and class members
+# alphabetically by member name. If set to NO the members will appear in
+# declaration order.
+
+SORT_MEMBER_DOCS       = NO
+
+# If the SORT_BRIEF_DOCS tag is set to YES then doxygen will sort the
+# brief documentation of file, namespace and class members alphabetically
+# by member name. If set to NO (the default) the members will appear in
+# declaration order.
+
+SORT_BRIEF_DOCS        = NO
+
+# If the SORT_BY_SCOPE_NAME tag is set to YES, the class list will be
+# sorted by fully-qualified names, including namespaces. If set to
+# NO (the default), the class list will be sorted only by class name,
+# not including the namespace part.
+# Note: This option is not very useful if HIDE_SCOPE_NAMES is set to YES.
+# Note: This option applies only to the class list, not to the
+# alphabetical list.
+
+SORT_BY_SCOPE_NAME     = NO
+
+# The GENERATE_TODOLIST tag can be used to enable (YES) or
+# disable (NO) the todo list. This list is created by putting \todo
+# commands in the documentation.
+
+GENERATE_TODOLIST      = YES
+
+# The GENERATE_TESTLIST tag can be used to enable (YES) or
+# disable (NO) the test list. This list is created by putting \test
+# commands in the documentation.
+
+GENERATE_TESTLIST      = YES
+
+# The GENERATE_BUGLIST tag can be used to enable (YES) or
+# disable (NO) the bug list. This list is created by putting \bug
+# commands in the documentation.
+
+GENERATE_BUGLIST       = YES
+
+# The GENERATE_DEPRECATEDLIST tag can be used to enable (YES) or
+# disable (NO) the deprecated list. This list is created by putting
+# \deprecated commands in the documentation.
+
+GENERATE_DEPRECATEDLIST= YES
+
+# The ENABLED_SECTIONS tag can be used to enable conditional
+# documentation sections, marked by \if sectionname ... \endif.
+
+ENABLED_SECTIONS       =
+
+# The MAX_INITIALIZER_LINES tag determines the maximum number of lines
+# the initial value of a variable or define consists of for it to appear in
+# the documentation. If the initializer consists of more lines than specified
+# here it will be hidden. Use a value of 0 to hide initializers completely.
+# The appearance of the initializer of individual variables and defines in the
+# documentation can be controlled using \showinitializer or \hideinitializer
+# command in the documentation regardless of this setting.
+
+MAX_INITIALIZER_LINES  = 30
+
+# Set the SHOW_USED_FILES tag to NO to disable the list of files generated
+# at the bottom of the documentation of classes and structs. If set to YES the
+# list will mention the files that were used to generate the documentation.
+
+SHOW_USED_FILES        = YES
+
+# The FILE_VERSION_FILTER tag can be used to specify a program or script that
+# doxygen should invoke to get the current version for each file (typically from the
+# version control system). Doxygen will invoke the program by executing (via
+# popen()) the command <command> <input-file>, where <command> is the value of
+# the FILE_VERSION_FILTER tag, and <input-file> is the name of an input file
+# provided by doxygen. Whatever the program writes to standard output
+# is used as the file version. See the manual for examples.
+
+FILE_VERSION_FILTER    =
+
+#---------------------------------------------------------------------------
+# configuration options related to warning and progress messages
+#---------------------------------------------------------------------------
+
+# The QUIET tag can be used to turn on/off the messages that are generated
+# by doxygen. Possible values are YES and NO. If left blank NO is used.
+
+QUIET                  = YES
+
+# The WARNINGS tag can be used to turn on/off the warning messages that are
+# generated by doxygen. Possible values are YES and NO. If left blank
+# NO is used.
+
+WARNINGS               = YES
+
+# If WARN_IF_UNDOCUMENTED is set to YES, then doxygen will generate warnings
+# for undocumented members. If EXTRACT_ALL is set to YES then this flag will
+# automatically be disabled.
+
+WARN_IF_UNDOCUMENTED   = YES
+
+# If WARN_IF_DOC_ERROR is set to YES, doxygen will generate warnings for
+# potential errors in the documentation, such as not documenting some
+# parameters in a documented function, or documenting parameters that
+# don't exist or using markup commands wrongly.
+
+WARN_IF_DOC_ERROR      = YES
+
+# This WARN_NO_PARAMDOC option can be abled to get warnings for
+# functions that are documented, but have no documentation for their parameters
+# or return value. If set to NO (the default) doxygen will only warn about
+# wrong or incomplete parameter documentation, but not about the absence of
+# documentation.
+
+WARN_NO_PARAMDOC       = NO
+
+# The WARN_FORMAT tag determines the format of the warning messages that
+# doxygen can produce. The string should contain the $file, $line, and $text
+# tags, which will be replaced by the file and line number from which the
+# warning originated and the warning text. Optionally the format may contain
+# $version, which will be replaced by the version of the file (if it could
+# be obtained via FILE_VERSION_FILTER)
+
+WARN_FORMAT            = "$file:$line: $text"
+
+# The WARN_LOGFILE tag can be used to specify a file to which warning
+# and error messages should be written. If left blank the output is written
+# to stderr.
+
+WARN_LOGFILE           =
+
+#---------------------------------------------------------------------------
+# configuration options related to the input files
+#---------------------------------------------------------------------------
+
+# The INPUT tag can be used to specify the files and/or directories that contain
+# documented source files. You may enter file names like "myfile.cpp" or
+# directories like "/usr/src/myproject". Separate the files or directories
+# with spaces.
+
+INPUT =
+
+# This tag can be used to specify the character encoding of the source files that
+# doxygen parses. Internally doxygen uses the UTF-8 encoding, which is also the default
+# input encoding. Doxygen uses libiconv (or the iconv built into libc) for the transcoding.
+# See http://www.gnu.org/software/libiconv for the list of possible encodings.
+
+INPUT_ENCODING         = UTF-8
+
+# If the value of the INPUT tag contains directories, you can use the
+# FILE_PATTERNS tag to specify one or more wildcard pattern (like *.cpp
+# and *.h) to filter out the source-files in the directories. If left
+# blank the following patterns are tested:
+# *.c *.cc *.cxx *.cpp *.c++ *.java *.ii *.ixx *.ipp *.i++ *.inl *.h *.hh *.hxx
+# *.hpp *.h++ *.idl *.odl *.cs *.php *.php3 *.inc *.m *.mm *.py *.f90
+
+FILE_PATTERNS          =
+
+# The RECURSIVE tag can be used to turn specify whether or not subdirectories
+# should be searched for input files as well. Possible values are YES and NO.
+# If left blank NO is used.
+
+RECURSIVE              = NO
+
+# The EXCLUDE tag can be used to specify files and/or directories that should
+# excluded from the INPUT source files. This way you can easily exclude a
+# subdirectory from a directory tree whose root is specified with the INPUT tag.
+
+EXCLUDE                =
+
+# The EXCLUDE_SYMLINKS tag can be used select whether or not files or
+# directories that are symbolic links (a Unix filesystem feature) are excluded
+# from the input.
+
+EXCLUDE_SYMLINKS       = NO
+
+# If the value of the INPUT tag contains directories, you can use the
+# EXCLUDE_PATTERNS tag to specify one or more wildcard patterns to exclude
+# certain files from those directories. Note that the wildcards are matched
+# against the file with absolute path, so to exclude all test directories
+# for example use the pattern */test/*
+
+EXCLUDE_PATTERNS       =
+
+# The EXCLUDE_SYMBOLS tag can be used to specify one or more symbol names
+# (namespaces, classes, functions, etc.) that should be excluded from the output.
+# The symbol name can be a fully qualified name, a word, or if the wildcard * is used,
+# a substring. Examples: ANamespace, AClass, AClass::ANamespace, ANamespace::*Test
+
+EXCLUDE_SYMBOLS        =
+
+# The EXAMPLE_PATH tag can be used to specify one or more files or
+# directories that contain example code fragments that are included (see
+# the \include command).
+
+EXAMPLE_PATH           =
+
+# If the value of the EXAMPLE_PATH tag contains directories, you can use the
+# EXAMPLE_PATTERNS tag to specify one or more wildcard pattern (like *.cpp
+# and *.h) to filter out the source-files in the directories. If left
+# blank all files are included.
+
+EXAMPLE_PATTERNS       =
+
+# If the EXAMPLE_RECURSIVE tag is set to YES then subdirectories will be
+# searched for input files to be used with the \include or \dontinclude
+# commands irrespective of the value of the RECURSIVE tag.
+# Possible values are YES and NO. If left blank NO is used.
+
+EXAMPLE_RECURSIVE      = NO
+
+# The IMAGE_PATH tag can be used to specify one or more files or
+# directories that contain image that are included in the documentation (see
+# the \image command).
+
+IMAGE_PATH             =
+
+# The INPUT_FILTER tag can be used to specify a program that doxygen should
+# invoke to filter for each input file. Doxygen will invoke the filter program
+# by executing (via popen()) the command <filter> <input-file>, where <filter>
+# is the value of the INPUT_FILTER tag, and <input-file> is the name of an
+# input file. Doxygen will then use the output that the filter program writes
+# to standard output.  If FILTER_PATTERNS is specified, this tag will be
+# ignored.
+
+INPUT_FILTER           =
+
+# The FILTER_PATTERNS tag can be used to specify filters on a per file pattern
+# basis.  Doxygen will compare the file name with each pattern and apply the
+# filter if there is a match.  The filters are a list of the form:
+# pattern=filter (like *.cpp=my_cpp_filter). See INPUT_FILTER for further
+# info on how filters are used. If FILTER_PATTERNS is empty, INPUT_FILTER
+# is applied to all files.
+
+FILTER_PATTERNS        =
+
+# If the FILTER_SOURCE_FILES tag is set to YES, the input filter (if set using
+# INPUT_FILTER) will be used to filter the input files when producing source
+# files to browse (i.e. when SOURCE_BROWSER is set to YES).
+
+FILTER_SOURCE_FILES    = NO
+
+#---------------------------------------------------------------------------
+# configuration options related to source browsing
+#---------------------------------------------------------------------------
+
+# If the SOURCE_BROWSER tag is set to YES then a list of source files will
+# be generated. Documented entities will be cross-referenced with these sources.
+# Note: To get rid of all source code in the generated output, make sure also
+# VERBATIM_HEADERS is set to NO. If you have enabled CALL_GRAPH or CALLER_GRAPH
+# then you must also enable this option. If you don't then doxygen will produce
+# a warning and turn it on anyway
+
+SOURCE_BROWSER         = NO
+
+# Setting the INLINE_SOURCES tag to YES will include the body
+# of functions and classes directly in the documentation.
+
+INLINE_SOURCES         = NO
+
+# Setting the STRIP_CODE_COMMENTS tag to YES (the default) will instruct
+# doxygen to hide any special comment blocks from generated source code
+# fragments. Normal C and C++ comments will always remain visible.
+
+STRIP_CODE_COMMENTS    = YES
+
+# If the REFERENCED_BY_RELATION tag is set to YES (the default)
+# then for each documented function all documented
+# functions referencing it will be listed.
+
+REFERENCED_BY_RELATION = YES
+
+# If the REFERENCES_RELATION tag is set to YES (the default)
+# then for each documented function all documented entities
+# called/used by that function will be listed.
+
+REFERENCES_RELATION    = YES
+
+# If the REFERENCES_LINK_SOURCE tag is set to YES (the default)
+# and SOURCE_BROWSER tag is set to YES, then the hyperlinks from
+# functions in REFERENCES_RELATION and REFERENCED_BY_RELATION lists will
+# link to the source code.  Otherwise they will link to the documentstion.
+
+REFERENCES_LINK_SOURCE = YES
+
+# If the USE_HTAGS tag is set to YES then the references to source code
+# will point to the HTML generated by the htags(1) tool instead of doxygen
+# built-in source browser. The htags tool is part of GNU's global source
+# tagging system (see http://www.gnu.org/software/global/global.html). You
+# will need version 4.8.6 or higher.
+
+USE_HTAGS              = NO
+
+# If the VERBATIM_HEADERS tag is set to YES (the default) then Doxygen
+# will generate a verbatim copy of the header file for each class for
+# which an include is specified. Set to NO to disable this.
+
+VERBATIM_HEADERS       = YES
+
+#---------------------------------------------------------------------------
+# configuration options related to the alphabetical class index
+#---------------------------------------------------------------------------
+
+# If the ALPHABETICAL_INDEX tag is set to YES, an alphabetical index
+# of all compounds will be generated. Enable this if the project
+# contains a lot of classes, structs, unions or interfaces.
+
+ALPHABETICAL_INDEX     = NO
+
+# If the alphabetical index is enabled (see ALPHABETICAL_INDEX) then
+# the COLS_IN_ALPHA_INDEX tag can be used to specify the number of columns
+# in which this list will be split (can be a number in the range [1..20])
+
+COLS_IN_ALPHA_INDEX    = 5
+
+# In case all classes in a project start with a common prefix, all
+# classes will be put under the same header in the alphabetical index.
+# The IGNORE_PREFIX tag can be used to specify one or more prefixes that
+# should be ignored while generating the index headers.
+
+IGNORE_PREFIX          =
+
+#---------------------------------------------------------------------------
+# configuration options related to the HTML output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_HTML tag is set to YES (the default) Doxygen will
+# generate HTML output.
+
+GENERATE_HTML          = YES
+
+# The HTML_OUTPUT tag is used to specify where the HTML docs will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be
+# put in front of it. If left blank `html' will be used as the default path.
+
+HTML_OUTPUT            = html
+
+# The HTML_FILE_EXTENSION tag can be used to specify the file extension for
+# each generated HTML page (for example: .htm,.php,.asp). If it is left blank
+# doxygen will generate files with .html extension.
+
+HTML_FILE_EXTENSION    = .html
+
+# The HTML_HEADER tag can be used to specify a personal HTML header for
+# each generated HTML page. If it is left blank doxygen will generate a
+# standard header.
+
+HTML_HEADER            =
+
+# The HTML_FOOTER tag can be used to specify a personal HTML footer for
+# each generated HTML page. If it is left blank doxygen will generate a
+# standard footer.
+
+HTML_FOOTER            =
+
+# The HTML_STYLESHEET tag can be used to specify a user-defined cascading
+# style sheet that is used by each HTML page. It can be used to
+# fine-tune the look of the HTML output. If the tag is left blank doxygen
+# will generate a default style sheet. Note that doxygen will try to copy
+# the style sheet file to the HTML output directory, so don't put your own
+# stylesheet in the HTML output directory as well, or it will be erased!
+
+HTML_STYLESHEET        =
+
+# If the GENERATE_HTMLHELP tag is set to YES, additional index files
+# will be generated that can be used as input for tools like the
+# Microsoft HTML help workshop to generate a compressed HTML help file (.chm)
+# of the generated HTML documentation.
+
+GENERATE_HTMLHELP      = NO
+
+# If the HTML_DYNAMIC_SECTIONS tag is set to YES then the generated HTML
+# documentation will contain sections that can be hidden and shown after the
+# page has loaded. For this to work a browser that supports
+# java_script and DHTML is required (for instance Mozilla 1.0+, Firefox
+# Netscape 6.0+, Internet explorer 5.0+, Konqueror, or Safari).
+
+HTML_DYNAMIC_SECTIONS  = NO
+
+# If the GENERATE_HTMLHELP tag is set to YES, the CHM_FILE tag can
+# be used to specify the file name of the resulting .chm file. You
+# can add a path in front of the file if the result should not be
+# written to the html output directory.
+
+CHM_FILE               =
+
+# If the GENERATE_HTMLHELP tag is set to YES, the HHC_LOCATION tag can
+# be used to specify the location (absolute path including file name) of
+# the HTML help compiler (hhc.exe). If non-empty doxygen will try to run
+# the HTML help compiler on the generated index.hhp.
+
+HHC_LOCATION           =
+
+# If the GENERATE_HTMLHELP tag is set to YES, the GENERATE_CHI flag
+# controls if a separate .chi index file is generated (YES) or that
+# it should be included in the master .chm file (NO).
+
+GENERATE_CHI           = NO
+
+# If the GENERATE_HTMLHELP tag is set to YES, the BINARY_TOC flag
+# controls whether a binary table of contents is generated (YES) or a
+# normal table of contents (NO) in the .chm file.
+
+BINARY_TOC             = NO
+
+# The TOC_EXPAND flag can be set to YES to add extra items for group members
+# to the contents of the HTML help documentation and to the tree view.
+
+TOC_EXPAND             = NO
+
+# The DISABLE_INDEX tag can be used to turn on/off the condensed index at
+# top of each HTML page. The value NO (the default) enables the index and
+# the value YES disables it.
+
+DISABLE_INDEX          = NO
+
+# This tag can be used to set the number of enum values (range [1..20])
+# that doxygen will group on one line in the generated HTML documentation.
+
+ENUM_VALUES_PER_LINE   = 4
+
+# If the GENERATE_TREEVIEW tag is set to YES, a side panel will be
+# generated containing a tree-like index structure (just like the one that
+# is generated for HTML Help). For this to work a browser that supports
+# java_script, DHTML, CSS and frames is required (for instance Mozilla 1.0+,
+# Netscape 6.0+, Internet explorer 5.0+, or Konqueror). Windows users are
+# probably better off using the HTML help feature.
+
+GENERATE_TREEVIEW      = NO
+
+# If the treeview is enabled (see GENERATE_TREEVIEW) then this tag can be
+# used to set the initial width (in pixels) of the frame in which the tree
+# is shown.
+
+TREEVIEW_WIDTH         = 250
+
+#---------------------------------------------------------------------------
+# configuration options related to the la_te_x output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_LATEX tag is set to YES (the default) Doxygen will
+# generate Latex output.
+
+GENERATE_LATEX         = YES
+
+# The LATEX_OUTPUT tag is used to specify where the la_te_x docs will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be
+# put in front of it. If left blank `latex' will be used as the default path.
+
+LATEX_OUTPUT           = latex
+
+# The LATEX_CMD_NAME tag can be used to specify the la_te_x command name to be
+# invoked. If left blank `latex' will be used as the default command name.
+
+LATEX_CMD_NAME         = latex
+
+# The MAKEINDEX_CMD_NAME tag can be used to specify the command name to
+# generate index for la_te_x. If left blank `makeindex' will be used as the
+# default command name.
+
+MAKEINDEX_CMD_NAME     = makeindex
+
+# If the COMPACT_LATEX tag is set to YES Doxygen generates more compact
+# la_te_x documents. This may be useful for small projects and may help to
+# save some trees in general.
+
+COMPACT_LATEX          = YES
+
+# The PAPER_TYPE tag can be used to set the paper type that is used
+# by the printer. Possible values are: a4, a4wide, letter, legal and
+# executive. If left blank a4wide will be used.
+
+PAPER_TYPE             = letter
+
+# The EXTRA_PACKAGES tag can be to specify one or more names of la_te_x
+# packages that should be included in the la_te_x output.
+
+EXTRA_PACKAGES         =
+
+# The LATEX_HEADER tag can be used to specify a personal la_te_x header for
+# the generated latex document. The header should contain everything until
+# the first chapter. If it is left blank doxygen will generate a
+# standard header. Notice: only use this tag if you know what you are doing!
+
+LATEX_HEADER           =
+
+# If the PDF_HYPERLINKS tag is set to YES, the la_te_x that is generated
+# is prepared for conversion to pdf (using ps2pdf). The pdf file will
+# contain links (just like the HTML output) instead of page references
+# This makes the output suitable for online browsing using a pdf viewer.
+
+PDF_HYPERLINKS         = YES
+
+# If the USE_PDFLATEX tag is set to YES, pdflatex will be used instead of
+# plain latex in the generated Makefile. Set this option to YES to get a
+# higher quality PDF documentation.
+
+USE_PDFLATEX           = YES
+
+# If the LATEX_BATCHMODE tag is set to YES, doxygen will add the \\batchmode.
+# command to the generated la_te_x files. This will instruct la_te_x to keep
+# running if errors occur, instead of asking the user for help.
+# This option is also used when generating formulas in HTML.
+
+LATEX_BATCHMODE        = NO
+
+# If LATEX_HIDE_INDICES is set to YES then doxygen will not
+# include the index chapters (such as File Index, Compound Index, etc.)
+# in the output.
+
+LATEX_HIDE_INDICES     = NO
+
+#---------------------------------------------------------------------------
+# configuration options related to the RTF output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_RTF tag is set to YES Doxygen will generate RTF output
+# The RTF output is optimized for Word 97 and may not look very pretty with
+# other RTF readers or editors.
+
+GENERATE_RTF           = NO
+
+# The RTF_OUTPUT tag is used to specify where the RTF docs will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be
+# put in front of it. If left blank `rtf' will be used as the default path.
+
+RTF_OUTPUT             = rtf
+
+# If the COMPACT_RTF tag is set to YES Doxygen generates more compact
+# RTF documents. This may be useful for small projects and may help to
+# save some trees in general.
+
+COMPACT_RTF            = NO
+
+# If the RTF_HYPERLINKS tag is set to YES, the RTF that is generated
+# will contain hyperlink fields. The RTF file will
+# contain links (just like the HTML output) instead of page references.
+# This makes the output suitable for online browsing using WORD or other
+# programs which support those fields.
+# Note: wordpad (write) and others do not support links.
+
+RTF_HYPERLINKS         = NO
+
+# Load stylesheet definitions from file. Syntax is similar to doxygen's
+# config file, i.e. a series of assignments. You only have to provide
+# replacements, missing definitions are set to their default value.
+
+RTF_STYLESHEET_FILE    =
+
+# Set optional variables used in the generation of an rtf document.
+# Syntax is similar to doxygen's config file.
+
+RTF_EXTENSIONS_FILE    =
+
+#---------------------------------------------------------------------------
+# configuration options related to the man page output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_MAN tag is set to YES (the default) Doxygen will
+# generate man pages
+
+GENERATE_MAN           = NO
+
+# The MAN_OUTPUT tag is used to specify where the man pages will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be
+# put in front of it. If left blank `man' will be used as the default path.
+
+MAN_OUTPUT             = man
+
+# The MAN_EXTENSION tag determines the extension that is added to
+# the generated man pages (default is the subroutine's section .3)
+
+MAN_EXTENSION          = .3
+
+# If the MAN_LINKS tag is set to YES and Doxygen generates man output,
+# then it will generate one additional man file for each entity
+# documented in the real man page(s). These additional files
+# only source the real man page, but without them the man command
+# would be unable to find the correct page. The default is NO.
+
+MAN_LINKS              = YES
+
+#---------------------------------------------------------------------------
+# configuration options related to the XML output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_XML tag is set to YES Doxygen will
+# generate an XML file that captures the structure of
+# the code including all documentation.
+
+GENERATE_XML           = NO
+
+# The XML_OUTPUT tag is used to specify where the XML pages will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be
+# put in front of it. If left blank `xml' will be used as the default path.
+
+XML_OUTPUT             = xml
+
+# The XML_SCHEMA tag can be used to specify an XML schema,
+# which can be used by a validating XML parser to check the
+# syntax of the XML files.
+
+XML_SCHEMA             =
+
+# The XML_DTD tag can be used to specify an XML DTD,
+# which can be used by a validating XML parser to check the
+# syntax of the XML files.
+
+XML_DTD                =
+
+# If the XML_PROGRAMLISTING tag is set to YES Doxygen will
+# dump the program listings (including syntax highlighting
+# and cross-referencing information) to the XML output. Note that
+# enabling this will significantly increase the size of the XML output.
+
+XML_PROGRAMLISTING     = YES
+
+#---------------------------------------------------------------------------
+# configuration options for the auto_gen Definitions output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_AUTOGEN_DEF tag is set to YES Doxygen will
+# generate an auto_gen Definitions (see autogen.sf.net) file
+# that captures the structure of the code including all
+# documentation. Note that this feature is still experimental
+# and incomplete at the moment.
+
+GENERATE_AUTOGEN_DEF   = NO
+
+#---------------------------------------------------------------------------
+# configuration options related to the Perl module output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_PERLMOD tag is set to YES Doxygen will
+# generate a Perl module file that captures the structure of
+# the code including all documentation. Note that this
+# feature is still experimental and incomplete at the
+# moment.
+
+GENERATE_PERLMOD       = NO
+
+# If the PERLMOD_LATEX tag is set to YES Doxygen will generate
+# the necessary Makefile rules, Perl scripts and la_te_x code to be able
+# to generate PDF and DVI output from the Perl module output.
+
+PERLMOD_LATEX          = NO
+
+# If the PERLMOD_PRETTY tag is set to YES the Perl module output will be
+# nicely formatted so it can be parsed by a human reader.  This is useful
+# if you want to understand what is going on.  On the other hand, if this
+# tag is set to NO the size of the Perl module output will be much smaller
+# and Perl will parse it just the same.
+
+PERLMOD_PRETTY         = YES
+
+# The names of the make variables in the generated doxyrules.make file
+# are prefixed with the string contained in PERLMOD_MAKEVAR_PREFIX.
+# This is useful so different doxyrules.make files included by the same
+# Makefile don't overwrite each other's variables.
+
+PERLMOD_MAKEVAR_PREFIX =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the preprocessor
+#---------------------------------------------------------------------------
+
+# If the ENABLE_PREPROCESSING tag is set to YES (the default) Doxygen will
+# evaluate all C-preprocessor directives found in the sources and include
+# files.
+
+ENABLE_PREPROCESSING   = YES
+
+# If the MACRO_EXPANSION tag is set to YES Doxygen will expand all macro
+# names in the source code. If set to NO (the default) only conditional
+# compilation will be performed. Macro expansion can be done in a controlled
+# way by setting EXPAND_ONLY_PREDEF to YES.
+
+MACRO_EXPANSION        = YES
+
+# If the EXPAND_ONLY_PREDEF and MACRO_EXPANSION tags are both set to YES
+# then the macro expansion is limited to the macros specified with the
+# PREDEFINED and EXPAND_AS_DEFINED tags.
+
+EXPAND_ONLY_PREDEF     = NO
+
+# If the SEARCH_INCLUDES tag is set to YES (the default) the includes files
+# in the INCLUDE_PATH (see below) will be search if a #include is found.
+
+SEARCH_INCLUDES        = YES
+
+# The INCLUDE_PATH tag can be used to specify one or more directories that
+# contain include files that are not input files but should be processed by
+# the preprocessor.
+
+INCLUDE_PATH           =
+
+# You can use the INCLUDE_FILE_PATTERNS tag to specify one or more wildcard
+# patterns (like *.h and *.hpp) to filter out the header-files in the
+# directories. If left blank, the patterns specified with FILE_PATTERNS will
+# be used.
+
+INCLUDE_FILE_PATTERNS  = *.h
+
+# The PREDEFINED tag can be used to specify one or more macro names that
+# are defined before the preprocessor is started (similar to the -D option of
+# gcc). The argument of the tag is a list of macros of the form: name
+# or name=definition (no spaces). If the definition and the = are
+# omitted =1 is assumed. To prevent a macro definition from being
+# undefined via #undef or recursively expanded use the := operator
+# instead of the = operator.
+
+PREDEFINED             =
+
+# If the MACRO_EXPANSION and EXPAND_ONLY_PREDEF tags are set to YES then
+# this tag can be used to specify a list of macro names that should be expanded.
+# The macro definition that is found in the sources will be used.
+# Use the PREDEFINED tag if you want to use a different macro definition.
+
+EXPAND_AS_DEFINED      =
+
+# If the SKIP_FUNCTION_MACROS tag is set to YES (the default) then
+# doxygen's preprocessor will remove all function-like macros that are alone
+# on a line, have an all uppercase name, and do not end with a semicolon. Such
+# function macros are typically used for boiler-plate code, and will confuse
+# the parser if not removed.
+
+SKIP_FUNCTION_MACROS   = YES
+
+#---------------------------------------------------------------------------
+# Configuration::additions related to external references
+#---------------------------------------------------------------------------
+
+# The TAGFILES option can be used to specify one or more tagfiles.
+# Optionally an initial location of the external documentation
+# can be added for each tagfile. The format of a tag file without
+# this location is as follows:
+#   TAGFILES = file1 file2 ...
+# Adding location for the tag files is done as follows:
+#   TAGFILES = file1=loc1 "file2 = loc2" ...
+# where "loc1" and "loc2" can be relative or absolute paths or
+# URLs. If a location is present for each tag, the installdox tool
+# does not have to be run to correct the links.
+# Note that each tag file must have a unique name
+# (where the name does NOT include the path)
+# If a tag file is not located in the directory in which doxygen
+# is run, you must also specify the path to the tagfile here.
+
+TAGFILES               =
+
+# When a file name is specified after GENERATE_TAGFILE, doxygen will create
+# a tag file that is based on the input files it reads.
+
+GENERATE_TAGFILE       =
+
+# If the ALLEXTERNALS tag is set to YES all external classes will be listed
+# in the class index. If set to NO only the inherited external classes
+# will be listed.
+
+ALLEXTERNALS           = NO
+
+# If the EXTERNAL_GROUPS tag is set to YES all external groups will be listed
+# in the modules index. If set to NO, only the current project's groups will
+# be listed.
+
+EXTERNAL_GROUPS        = YES
+
+# The PERL_PATH should be the absolute path and name of the perl script
+# interpreter (i.e. the result of `which perl').
+
+PERL_PATH              = /usr/bin/perl
+
+#---------------------------------------------------------------------------
+# Configuration options related to the dot tool
+#---------------------------------------------------------------------------
+
+# If the CLASS_DIAGRAMS tag is set to YES (the default) Doxygen will
+# generate a inheritance diagram (in HTML, RTF and la_te_x) for classes with base
+# or super classes. Setting the tag to NO turns the diagrams off. Note that
+# this option is superseded by the HAVE_DOT option below. This is only a
+# fallback. It is recommended to install and use dot, since it yields more
+# powerful graphs.
+
+CLASS_DIAGRAMS         = YES
+
+# You can define message sequence charts within doxygen comments using the \msc
+# command. Doxygen will then run the mscgen tool (see http://www.mcternan.me.uk/mscgen/) to
+# produce the chart and insert it in the documentation. The MSCGEN_PATH tag allows you to
+# specify the directory where the mscgen tool resides. If left empty the tool is assumed to
+# be found in the default search path.
+
+MSCGEN_PATH            =
+
+# If set to YES, the inheritance and collaboration graphs will hide
+# inheritance and usage relations if the target is undocumented
+# or is not a class.
+
+HIDE_UNDOC_RELATIONS   = YES
+
+# If you set the HAVE_DOT tag to YES then doxygen will assume the dot tool is
+# available from the path. This tool is part of Graphviz, a graph visualization
+# toolkit from AT&T and Lucent Bell Labs. The other options in this section
+# have no effect if this option is set to NO (the default)
+
+HAVE_DOT               = NO
+
+# If the CLASS_GRAPH and HAVE_DOT tags are set to YES then doxygen
+# will generate a graph for each documented class showing the direct and
+# indirect inheritance relations. Setting this tag to YES will force the
+# the CLASS_DIAGRAMS tag to NO.
+
+CLASS_GRAPH            = YES
+
+# If the COLLABORATION_GRAPH and HAVE_DOT tags are set to YES then doxygen
+# will generate a graph for each documented class showing the direct and
+# indirect implementation dependencies (inheritance, containment, and
+# class references variables) of the class with other documented classes.
+
+COLLABORATION_GRAPH    = YES
+
+# If the GROUP_GRAPHS and HAVE_DOT tags are set to YES then doxygen
+# will generate a graph for groups, showing the direct groups dependencies
+
+GROUP_GRAPHS           = YES
+
+# If the UML_LOOK tag is set to YES doxygen will generate inheritance and
+# collaboration diagrams in a style similar to the OMG's Unified Modeling
+# Language.
+
+UML_LOOK               = NO
+
+# If set to YES, the inheritance and collaboration graphs will show the
+# relations between templates and their instances.
+
+TEMPLATE_RELATIONS     = NO
+
+# If the ENABLE_PREPROCESSING, SEARCH_INCLUDES, INCLUDE_GRAPH, and HAVE_DOT
+# tags are set to YES then doxygen will generate a graph for each documented
+# file showing the direct and indirect include dependencies of the file with
+# other documented files.
+
+INCLUDE_GRAPH          = YES
+
+# If the ENABLE_PREPROCESSING, SEARCH_INCLUDES, INCLUDED_BY_GRAPH, and
+# HAVE_DOT tags are set to YES then doxygen will generate a graph for each
+# documented header file showing the documented files that directly or
+# indirectly include this file.
+
+INCLUDED_BY_GRAPH      = YES
+
+# If the CALL_GRAPH, SOURCE_BROWSER and HAVE_DOT tags are set to YES then doxygen will
+# generate a call dependency graph for every global function or class method.
+# Note that enabling this option will significantly increase the time of a run.
+# So in most cases it will be better to enable call graphs for selected
+# functions only using the \callgraph command.
+
+CALL_GRAPH             = NO
+
+# If the CALLER_GRAPH, SOURCE_BROWSER and HAVE_DOT tags are set to YES then doxygen will
+# generate a caller dependency graph for every global function or class method.
+# Note that enabling this option will significantly increase the time of a run.
+# So in most cases it will be better to enable caller graphs for selected
+# functions only using the \callergraph command.
+
+CALLER_GRAPH           = NO
+
+# If the GRAPHICAL_HIERARCHY and HAVE_DOT tags are set to YES then doxygen
+# will graphical hierarchy of all classes instead of a textual one.
+
+GRAPHICAL_HIERARCHY    = YES
+
+# If the DIRECTORY_GRAPH, SHOW_DIRECTORIES and HAVE_DOT tags are set to YES
+# then doxygen will show the dependencies a directory has on other directories
+# in a graphical way. The dependency relations are determined by the #include
+# relations between the files in the directories.
+
+DIRECTORY_GRAPH        = YES
+
+# The DOT_IMAGE_FORMAT tag can be used to set the image format of the images
+# generated by dot. Possible values are png, jpg, or gif
+# If left blank png will be used.
+
+DOT_IMAGE_FORMAT       = png
+
+# The tag DOT_PATH can be used to specify the path where the dot tool can be
+# found. If left blank, it is assumed the dot tool can be found in the path.
+
+DOT_PATH               =
+
+# The DOTFILE_DIRS tag can be used to specify one or more directories that
+# contain dot files that are included in the documentation (see the
+# \dotfile command).
+
+DOTFILE_DIRS           =
+
+# The MAX_DOT_GRAPH_MAX_NODES tag can be used to set the maximum number of
+# nodes that will be shown in the graph. If the number of nodes in a graph
+# becomes larger than this value, doxygen will truncate the graph, which is
+# visualized by representing a node as a red box. Note that doxygen if the number
+# of direct children of the root node in a graph is already larger than
+# MAX_DOT_GRAPH_NOTES then the graph will not be shown at all. Also note
+# that the size of a graph can be further restricted by MAX_DOT_GRAPH_DEPTH.
+
+DOT_GRAPH_MAX_NODES    = 50
+
+# The MAX_DOT_GRAPH_DEPTH tag can be used to set the maximum depth of the
+# graphs generated by dot. A depth value of 3 means that only nodes reachable
+# from the root by following a path via at most 3 edges will be shown. Nodes
+# that lay further from the root node will be omitted. Note that setting this
+# option to 1 or 2 may greatly reduce the computation time needed for large
+# code bases. Also note that the size of a graph can be further restricted by
+# DOT_GRAPH_MAX_NODES. Using a depth of 0 means no depth restriction.
+
+MAX_DOT_GRAPH_DEPTH    = 0
+
+# Set the DOT_TRANSPARENT tag to YES to generate images with a transparent
+# background. This is disabled by default, which results in a white background.
+# Warning: Depending on the platform used, enabling this option may lead to
+# badly anti-aliased labels on the edges of a graph (i.e. they become hard to
+# read).
+
+DOT_TRANSPARENT        = YES
+
+# Set the DOT_MULTI_TARGETS tag to YES allow dot to generate multiple output
+# files in one run (i.e. multiple -o and -T options on the command line). This
+# makes dot run faster, but since only newer versions of dot (>1.8.10)
+# support this, this feature is disabled by default.
+
+DOT_MULTI_TARGETS      = NO
+
+# If the GENERATE_LEGEND tag is set to YES (the default) Doxygen will
+# generate a legend page explaining the meaning of the various boxes and
+# arrows in the dot generated graphs.
+
+GENERATE_LEGEND        = YES
+
+# If the DOT_CLEANUP tag is set to YES (the default) Doxygen will
+# remove the intermediate dot files that are used to generate
+# the various graphs.
+
+DOT_CLEANUP            = YES
+
+#---------------------------------------------------------------------------
+# Configuration::additions related to the search engine
+#---------------------------------------------------------------------------
+
+# The SEARCHENGINE tag specifies whether or not a search engine should be
+# used. If set to NO the values of all tags below this one will be ignored.
+
+SEARCHENGINE           = NO
diff --git a/third_party/aom/libs.mk b/third_party/aom/libs.mk
new file mode 100644
index 0000000000..4f2e5cab55
--- /dev/null
+++ b/third_party/aom/libs.mk
@@ -0,0 +1,591 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+# ARM assembly files are written in RVCT-style. We use some make magic to
+# filter those files to allow GCC compilation
+ifeq ($(ARCH_ARM),yes)
+  ASM:=$(if $(filter yes,$(CONFIG_GCC)$(CONFIG_MSVS)),.asm.s,.asm)
+else
+  ASM:=.asm
+endif
+
+#
+# Rule to generate runtime cpu detection files
+#
+define rtcd_h_template
+$$(BUILD_PFX)$(1).h: $$(SRC_PATH_BARE)/$(2)
+	@echo "    [CREATE] $$@"
+	$$(qexec)$$(SRC_PATH_BARE)/build/make/rtcd.pl --arch=$$(TGT_ISA) \
+          --sym=$(1) \
+          --config=$$(CONFIG_DIR)$$(target)-$$(TOOLCHAIN).mk \
+          $$(RTCD_OPTIONS) $$^ > $$@
+CLEAN-OBJS += $$(BUILD_PFX)$(1).h
+RTCD += $$(BUILD_PFX)$(1).h
+endef
+
+CODEC_SRCS-yes += CHANGELOG
+CODEC_SRCS-yes += libs.mk
+
+include $(SRC_PATH_BARE)/aom/aom_codec.mk
+CODEC_SRCS-yes += $(addprefix aom/,$(call enabled,API_SRCS))
+CODEC_DOC_SRCS += $(addprefix aom/,$(call enabled,API_DOC_SRCS))
+
+include $(SRC_PATH_BARE)/aom_mem/aom_mem.mk
+CODEC_SRCS-yes += $(addprefix aom_mem/,$(call enabled,MEM_SRCS))
+
+include $(SRC_PATH_BARE)/aom_scale/aom_scale.mk
+CODEC_SRCS-yes += $(addprefix aom_scale/,$(call enabled,SCALE_SRCS))
+
+include $(SRC_PATH_BARE)/aom_ports/aom_ports.mk
+CODEC_SRCS-yes += $(addprefix aom_ports/,$(call enabled,PORTS_SRCS))
+
+include $(SRC_PATH_BARE)/aom_dsp/aom_dsp.mk
+CODEC_SRCS-yes += $(addprefix aom_dsp/,$(call enabled,DSP_SRCS))
+
+include $(SRC_PATH_BARE)/aom_util/aom_util.mk
+CODEC_SRCS-yes += $(addprefix aom_util/,$(call enabled,UTIL_SRCS))
+
+#  AV1 make file
+ifeq ($(CONFIG_AV1),yes)
+  AV1_PREFIX=av1/
+  include $(SRC_PATH_BARE)/$(AV1_PREFIX)av1_common.mk
+endif
+
+ifeq ($(CONFIG_AV1_ENCODER),yes)
+  AV1_PREFIX=av1/
+  include $(SRC_PATH_BARE)/$(AV1_PREFIX)av1_cx.mk
+  CODEC_SRCS-yes += $(addprefix $(AV1_PREFIX),$(call enabled,AV1_CX_SRCS))
+  CODEC_EXPORTS-yes += $(addprefix $(AV1_PREFIX),$(AV1_CX_EXPORTS))
+  CODEC_SRCS-yes += $(AV1_PREFIX)av1_cx.mk aom/aom.h aom/aomcx.h
+  INSTALL-LIBS-yes += include/aom/aom.h include/aom/aomcx.h
+  INSTALL_MAPS += include/aom/% $(SRC_PATH_BARE)/$(AV1_PREFIX)/%
+  CODEC_DOC_SRCS += aom/aom.h aom/aomcx.h
+  CODEC_DOC_SECTIONS += av1 av1_encoder
+endif
+
+ifeq ($(CONFIG_AV1_DECODER),yes)
+  AV1_PREFIX=av1/
+  include $(SRC_PATH_BARE)/$(AV1_PREFIX)av1_dx.mk
+  CODEC_SRCS-yes += $(addprefix $(AV1_PREFIX),$(call enabled,AV1_DX_SRCS))
+  CODEC_EXPORTS-yes += $(addprefix $(AV1_PREFIX),$(AV1_DX_EXPORTS))
+  CODEC_SRCS-yes += $(AV1_PREFIX)av1_dx.mk aom/aom.h aom/aomdx.h
+  INSTALL-LIBS-yes += include/aom/aom.h include/aom/aomdx.h
+  INSTALL_MAPS += include/aom/% $(SRC_PATH_BARE)/$(AV1_PREFIX)/%
+  CODEC_DOC_SRCS += aom/aom.h aom/aomdx.h
+  CODEC_DOC_SECTIONS += av1 av1_decoder
+endif
+
+AV1_PREFIX=av1/
+$(BUILD_PFX)$(AV1_PREFIX)%.c.o: CFLAGS += -Wextra
+
+ifeq ($(CONFIG_ENCODERS),yes)
+  CODEC_DOC_SECTIONS += encoder
+endif
+ifeq ($(CONFIG_DECODERS),yes)
+  CODEC_DOC_SECTIONS += decoder
+endif
+
+
+ifeq ($(CONFIG_MSVS),yes)
+CODEC_LIB=$(if $(CONFIG_STATIC_MSVCRT),aommt,aommd)
+GTEST_LIB=$(if $(CONFIG_STATIC_MSVCRT),gtestmt,gtestmd)
+# This variable uses deferred expansion intentionally, since the results of
+# $(wildcard) may change during the course of the Make.
+VS_PLATFORMS = $(foreach d,$(wildcard */Release/$(CODEC_LIB).lib),$(word 1,$(subst /, ,$(d))))
+endif
+
+# The following pairs define a mapping of locations in the distribution
+# tree to locations in the source/build trees.
+INSTALL_MAPS += include/aom/% $(SRC_PATH_BARE)/aom/%
+INSTALL_MAPS += include/aom/% $(SRC_PATH_BARE)/aom_ports/%
+INSTALL_MAPS += $(LIBSUBDIR)/%     %
+INSTALL_MAPS += src/%     $(SRC_PATH_BARE)/%
+ifeq ($(CONFIG_MSVS),yes)
+INSTALL_MAPS += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/%  $(p)/Release/%)
+INSTALL_MAPS += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/%  $(p)/Debug/%)
+endif
+
+CODEC_SRCS-yes += build/make/version.sh
+CODEC_SRCS-yes += build/make/rtcd.pl
+CODEC_SRCS-yes += aom_ports/emmintrin_compat.h
+CODEC_SRCS-yes += aom_ports/mem_ops.h
+CODEC_SRCS-yes += aom_ports/mem_ops_aligned.h
+CODEC_SRCS-yes += aom_ports/aom_once.h
+CODEC_SRCS-yes += $(BUILD_PFX)aom_config.c
+INSTALL-SRCS-no += $(BUILD_PFX)aom_config.c
+ifeq ($(ARCH_X86)$(ARCH_X86_64),yes)
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += third_party/x86inc/x86inc.asm
+endif
+CODEC_EXPORTS-yes += aom/exports_com
+CODEC_EXPORTS-$(CONFIG_ENCODERS) += aom/exports_enc
+CODEC_EXPORTS-$(CONFIG_DECODERS) += aom/exports_dec
+
+INSTALL-LIBS-yes += include/aom/aom_codec.h
+INSTALL-LIBS-yes += include/aom/aom_frame_buffer.h
+INSTALL-LIBS-yes += include/aom/aom_image.h
+INSTALL-LIBS-yes += include/aom/aom_integer.h
+INSTALL-LIBS-$(CONFIG_DECODERS) += include/aom/aom_decoder.h
+INSTALL-LIBS-$(CONFIG_ENCODERS) += include/aom/aom_encoder.h
+ifeq ($(CONFIG_EXTERNAL_BUILD),yes)
+ifeq ($(CONFIG_MSVS),yes)
+INSTALL-LIBS-yes                  += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/$(CODEC_LIB).lib)
+INSTALL-LIBS-$(CONFIG_DEBUG_LIBS) += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/$(CODEC_LIB)d.lib)
+INSTALL-LIBS-$(CONFIG_SHARED) += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/aom.dll)
+INSTALL-LIBS-$(CONFIG_SHARED) += $(foreach p,$(VS_PLATFORMS),$(LIBSUBDIR)/$(p)/aom.exp)
+endif
+else
+INSTALL-LIBS-$(CONFIG_STATIC) += $(LIBSUBDIR)/libaom.a
+INSTALL-LIBS-$(CONFIG_DEBUG_LIBS) += $(LIBSUBDIR)/libaom_g.a
+endif
+
+CODEC_SRCS=$(call enabled,CODEC_SRCS)
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += $(CODEC_SRCS)
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += $(call enabled,CODEC_EXPORTS)
+
+
+# Generate a list of all enabled sources, in particular for exporting to gyp
+# based build systems.
+libaom_srcs.txt:
+	@echo "    [CREATE] $@"
+	@echo $(CODEC_SRCS) | xargs -n1 echo | LC_ALL=C sort -u > $@
+CLEAN-OBJS += libaom_srcs.txt
+
+
+ifeq ($(CONFIG_EXTERNAL_BUILD),yes)
+ifeq ($(CONFIG_MSVS),yes)
+
+aom.def: $(call enabled,CODEC_EXPORTS)
+	@echo "    [CREATE] $@"
+	$(qexec)$(SRC_PATH_BARE)/build/make/gen_msvs_def.sh\
+            --name=aom\
+            --out=$@ $^
+CLEAN-OBJS += aom.def
+
+# Assembly files that are included, but don't define symbols themselves.
+# Filtered out to avoid Visual Studio build warnings.
+ASM_INCLUDES := \
+    third_party/x86inc/x86inc.asm \
+    aom_config.asm \
+    aom_ports/x86_abi_support.asm \
+
+aom.$(VCPROJ_SFX): $(CODEC_SRCS) aom.def
+	@echo "    [CREATE] $@"
+	$(qexec)$(GEN_VCPROJ) \
+            $(if $(CONFIG_SHARED),--dll,--lib) \
+            --target=$(TOOLCHAIN) \
+            $(if $(CONFIG_STATIC_MSVCRT),--static-crt) \
+            --name=aom \
+            --proj-guid=DCE19DAF-69AC-46DB-B14A-39F0FAA5DB74 \
+            --module-def=aom.def \
+            --ver=$(CONFIG_VS_VERSION) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
+            --out=$@ $(CFLAGS) \
+            $(filter-out $(addprefix %, $(ASM_INCLUDES)), $^) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
+
+PROJECTS-yes += aom.$(VCPROJ_SFX)
+
+aom.$(VCPROJ_SFX): aom_config.asm
+aom.$(VCPROJ_SFX): $(RTCD)
+
+endif
+else
+LIBAOM_OBJS=$(call objs,$(CODEC_SRCS))
+OBJS-yes += $(LIBAOM_OBJS)
+LIBS-$(if yes,$(CONFIG_STATIC)) += $(BUILD_PFX)libaom.a $(BUILD_PFX)libaom_g.a
+$(BUILD_PFX)libaom_g.a: $(LIBAOM_OBJS)
+
+SO_VERSION_MAJOR := 3
+SO_VERSION_MINOR := 0
+SO_VERSION_PATCH := 0
+ifeq ($(filter darwin%,$(TGT_OS)),$(TGT_OS))
+LIBAOM_SO               := libaom.$(SO_VERSION_MAJOR).dylib
+SHARED_LIB_SUF          := .dylib
+EXPORT_FILE             := libaom.syms
+LIBAOM_SO_SYMLINKS      := $(addprefix $(LIBSUBDIR)/, \
+                             libaom.dylib  )
+else
+ifeq ($(filter iphonesimulator%,$(TGT_OS)),$(TGT_OS))
+LIBAOM_SO               := libaom.$(SO_VERSION_MAJOR).dylib
+SHARED_LIB_SUF          := .dylib
+EXPORT_FILE             := libaom.syms
+LIBAOM_SO_SYMLINKS      := $(addprefix $(LIBSUBDIR)/, libaom.dylib)
+else
+ifeq ($(filter os2%,$(TGT_OS)),$(TGT_OS))
+LIBAOM_SO               := libaom$(SO_VERSION_MAJOR).dll
+SHARED_LIB_SUF          := _dll.a
+EXPORT_FILE             := libaom.def
+LIBAOM_SO_SYMLINKS      :=
+LIBAOM_SO_IMPLIB        := libaom_dll.a
+else
+LIBAOM_SO               := libaom.so.$(SO_VERSION_MAJOR).$(SO_VERSION_MINOR).$(SO_VERSION_PATCH)
+SHARED_LIB_SUF          := .so
+EXPORT_FILE             := libaom.ver
+LIBAOM_SO_SYMLINKS      := $(addprefix $(LIBSUBDIR)/, \
+                             libaom.so libaom.so.$(SO_VERSION_MAJOR) \
+                             libaom.so.$(SO_VERSION_MAJOR).$(SO_VERSION_MINOR))
+endif
+endif
+endif
+
+LIBS-$(CONFIG_SHARED) += $(BUILD_PFX)$(LIBAOM_SO)\
+                           $(notdir $(LIBAOM_SO_SYMLINKS)) \
+                           $(if $(LIBAOM_SO_IMPLIB), $(BUILD_PFX)$(LIBAOM_SO_IMPLIB))
+$(BUILD_PFX)$(LIBAOM_SO): $(LIBAOM_OBJS) $(EXPORT_FILE)
+$(BUILD_PFX)$(LIBAOM_SO): extralibs += -lm
+$(BUILD_PFX)$(LIBAOM_SO): SONAME = libaom.so.$(SO_VERSION_MAJOR)
+$(BUILD_PFX)$(LIBAOM_SO): EXPORTS_FILE = $(EXPORT_FILE)
+
+libaom.ver: $(call enabled,CODEC_EXPORTS)
+	@echo "    [CREATE] $@"
+	$(qexec)echo "{ global:" > $@
+	$(qexec)for f in $?; do awk '{print $$2";"}' < $$f >>$@; done
+	$(qexec)echo "local: *; };" >> $@
+CLEAN-OBJS += libaom.ver
+
+libaom.syms: $(call enabled,CODEC_EXPORTS)
+	@echo "    [CREATE] $@"
+	$(qexec)awk '{print "_"$$2}' $^ >$@
+CLEAN-OBJS += libaom.syms
+
+libaom.def: $(call enabled,CODEC_EXPORTS)
+	@echo "    [CREATE] $@"
+	$(qexec)echo LIBRARY $(LIBAOM_SO:.dll=) INITINSTANCE TERMINSTANCE > $@
+	$(qexec)echo "DATA MULTIPLE NONSHARED" >> $@
+	$(qexec)echo "EXPORTS" >> $@
+	$(qexec)awk '{print "_"$$2}' $^ >>$@
+CLEAN-OBJS += libaom.def
+
+libaom_dll.a: $(LIBAOM_SO)
+	@echo "    [IMPLIB] $@"
+	$(qexec)emximp -o $@ $<
+CLEAN-OBJS += libaom_dll.a
+
+define libaom_symlink_template
+$(1): $(2)
+	@echo "    [LN]     $(2) $$@"
+	$(qexec)mkdir -p $$(dir $$@)
+	$(qexec)ln -sf $(2) $$@
+endef
+
+$(eval $(call libaom_symlink_template,\
+    $(addprefix $(BUILD_PFX),$(notdir $(LIBAOM_SO_SYMLINKS))),\
+    $(BUILD_PFX)$(LIBAOM_SO)))
+$(eval $(call libaom_symlink_template,\
+    $(addprefix $(DIST_DIR)/,$(LIBAOM_SO_SYMLINKS)),\
+    $(LIBAOM_SO)))
+
+
+INSTALL-LIBS-$(CONFIG_SHARED) += $(LIBAOM_SO_SYMLINKS)
+INSTALL-LIBS-$(CONFIG_SHARED) += $(LIBSUBDIR)/$(LIBAOM_SO)
+INSTALL-LIBS-$(CONFIG_SHARED) += $(if $(LIBAOM_SO_IMPLIB),$(LIBSUBDIR)/$(LIBAOM_SO_IMPLIB))
+
+
+LIBS-yes += aom.pc
+aom.pc: config.mk libs.mk
+	@echo "    [CREATE] $@"
+	$(qexec)echo '# pkg-config file from libaom $(VERSION_STRING)' > $@
+	$(qexec)echo 'prefix=$(PREFIX)' >> $@
+	$(qexec)echo 'exec_prefix=$${prefix}' >> $@
+	$(qexec)echo 'libdir=$${prefix}/$(LIBSUBDIR)' >> $@
+	$(qexec)echo 'includedir=$${prefix}/include' >> $@
+	$(qexec)echo '' >> $@
+	$(qexec)echo 'Name: aom' >> $@
+	$(qexec)echo 'Description: WebM Project AVx codec implementation' >> $@
+	$(qexec)echo 'Version: $(VERSION_MAJOR).$(VERSION_MINOR).$(VERSION_PATCH)' >> $@
+	$(qexec)echo 'Requires:' >> $@
+	$(qexec)echo 'Conflicts:' >> $@
+	$(qexec)echo 'Libs: -L$${libdir} -laom -lm' >> $@
+ifeq ($(HAVE_PTHREAD_H),yes)
+	$(qexec)echo 'Libs.private: -lm -lpthread' >> $@
+else
+	$(qexec)echo 'Libs.private: -lm' >> $@
+endif
+	$(qexec)echo 'Cflags: -I$${includedir}' >> $@
+INSTALL-LIBS-yes += $(LIBSUBDIR)/pkgconfig/aom.pc
+INSTALL_MAPS += $(LIBSUBDIR)/pkgconfig/%.pc %.pc
+CLEAN-OBJS += aom.pc
+endif
+
+#
+# Rule to make assembler configuration file from C configuration file
+#
+ifeq ($(ARCH_X86)$(ARCH_X86_64),yes)
+# YASM
+$(BUILD_PFX)aom_config.asm: $(BUILD_PFX)aom_config.h
+	@echo "    [CREATE] $@"
+	@egrep "#define [A-Z0-9_]+ [01]" $< \
+	    | awk '{print $$2 " equ " $$3}' > $@
+else
+ADS2GAS=$(if $(filter yes,$(CONFIG_GCC)),| $(ASM_CONVERSION))
+$(BUILD_PFX)aom_config.asm: $(BUILD_PFX)aom_config.h
+	@echo "    [CREATE] $@"
+	@egrep "#define [A-Z0-9_]+ [01]" $< \
+	    | awk '{print $$2 " EQU " $$3}' $(ADS2GAS) > $@
+	@echo "        END" $(ADS2GAS) >> $@
+CLEAN-OBJS += $(BUILD_PFX)aom_config.asm
+endif
+
+#
+# Add assembler dependencies for configuration.
+#
+$(filter %.s.o,$(OBJS-yes)):     $(BUILD_PFX)aom_config.asm
+$(filter %$(ASM).o,$(OBJS-yes)): $(BUILD_PFX)aom_config.asm
+
+
+$(shell $(SRC_PATH_BARE)/build/make/version.sh "$(SRC_PATH_BARE)" $(BUILD_PFX)aom_version.h)
+CLEAN-OBJS += $(BUILD_PFX)aom_version.h
+
+#
+# Add include path for libwebm sources.
+#
+ifeq ($(CONFIG_WEBM_IO),yes)
+  CXXFLAGS += -I$(SRC_PATH_BARE)/third_party/libwebm
+endif
+
+##
+## libaom test directives
+##
+ifeq ($(CONFIG_UNIT_TESTS),yes)
+LIBAOM_TEST_DATA_PATH ?= .
+
+include $(SRC_PATH_BARE)/test/test.mk
+LIBAOM_TEST_SRCS=$(addprefix test/,$(call enabled,LIBAOM_TEST_SRCS))
+LIBAOM_TEST_BIN=./test_libaom$(EXE_SFX)
+LIBAOM_TEST_DATA=$(addprefix $(LIBAOM_TEST_DATA_PATH)/,\
+                     $(call enabled,LIBAOM_TEST_DATA))
+libaom_test_data_url=https://storage.googleapis.com/downloads.webmproject.org/test_data/libvpx/$(1)
+
+TEST_INTRA_PRED_SPEED_BIN=./test_intra_pred_speed$(EXE_SFX)
+TEST_INTRA_PRED_SPEED_SRCS=$(addprefix test/,$(call enabled,TEST_INTRA_PRED_SPEED_SRCS))
+TEST_INTRA_PRED_SPEED_OBJS := $(sort $(call objs,$(TEST_INTRA_PRED_SPEED_SRCS)))
+
+libaom_test_srcs.txt:
+	@echo "    [CREATE] $@"
+	@echo $(LIBAOM_TEST_SRCS) | xargs -n1 echo | LC_ALL=C sort -u > $@
+CLEAN-OBJS += libaom_test_srcs.txt
+
+$(LIBAOM_TEST_DATA): $(SRC_PATH_BARE)/test/test-data.sha1
+	@echo "    [DOWNLOAD] $@"
+	$(qexec)trap 'rm -f $@' INT TERM &&\
+            curl -L -o $@ $(call libaom_test_data_url,$(@F))
+
+testdata:: $(LIBAOM_TEST_DATA)
+	$(qexec)[ -x "$$(which sha1sum)" ] && sha1sum=sha1sum;\
+          [ -x "$$(which shasum)" ] && sha1sum=shasum;\
+          [ -x "$$(which sha1)" ] && sha1sum=sha1;\
+          if [ -n "$${sha1sum}" ]; then\
+            set -e;\
+            echo "Checking test data:";\
+            for f in $(call enabled,LIBAOM_TEST_DATA); do\
+                grep $$f $(SRC_PATH_BARE)/test/test-data.sha1 |\
+                    (cd $(LIBAOM_TEST_DATA_PATH); $${sha1sum} -c);\
+            done; \
+        else\
+            echo "Skipping test data integrity check, sha1sum not found.";\
+        fi
+
+ifeq ($(CONFIG_EXTERNAL_BUILD),yes)
+ifeq ($(CONFIG_MSVS),yes)
+gtest.$(VCPROJ_SFX): $(SRC_PATH_BARE)/third_party/googletest/src/googletest/src/gtest-all.cc
+	@echo "    [CREATE] $@"
+	$(qexec)$(GEN_VCPROJ) \
+            --lib \
+            --target=$(TOOLCHAIN) \
+            $(if $(CONFIG_STATIC_MSVCRT),--static-crt) \
+            --name=gtest \
+            --proj-guid=EC00E1EC-AF68-4D92-A255-181690D1C9B1 \
+            --ver=$(CONFIG_VS_VERSION) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
+            -D_VARIADIC_MAX=10 \
+            --out=gtest.$(VCPROJ_SFX) \
+	    $(SRC_PATH_BARE)/third_party/googletest/src/googletest/src/gtest-all.cc \
+            -I. -I"$(SRC_PATH_BARE)/third_party/googletest/src/googletest/include" \
+	    -I"$(SRC_PATH_BARE)/third_party/googletest/src/googletest"
+
+PROJECTS-$(CONFIG_MSVS) += gtest.$(VCPROJ_SFX)
+
+test_libaom.$(VCPROJ_SFX): $(LIBAOM_TEST_SRCS) aom.$(VCPROJ_SFX) gtest.$(VCPROJ_SFX)
+	@echo "    [CREATE] $@"
+	$(qexec)$(GEN_VCPROJ) \
+            --exe \
+            --target=$(TOOLCHAIN) \
+            --name=test_libaom \
+            -D_VARIADIC_MAX=10 \
+            --proj-guid=CD837F5F-52D8-4314-A370-895D614166A7 \
+            --ver=$(CONFIG_VS_VERSION) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
+            $(if $(CONFIG_STATIC_MSVCRT),--static-crt) \
+            --out=$@ $(INTERNAL_CFLAGS) $(CFLAGS) \
+            -I. -I"$(SRC_PATH_BARE)/third_party/googletest/src/googletest/include" \
+            $(if $(CONFIG_WEBM_IO),-I"$(SRC_PATH_BARE)/third_party/libwebm") \
+            -L. -l$(CODEC_LIB) -l$(GTEST_LIB) $^
+
+PROJECTS-$(CONFIG_MSVS) += test_libaom.$(VCPROJ_SFX)
+
+LIBAOM_TEST_BIN := $(addprefix $(TGT_OS:win64=x64)/Release/,$(notdir $(LIBAOM_TEST_BIN)))
+
+ifneq ($(strip $(TEST_INTRA_PRED_SPEED_OBJS)),)
+PROJECTS-$(CONFIG_MSVS) += test_intra_pred_speed.$(VCPROJ_SFX)
+test_intra_pred_speed.$(VCPROJ_SFX): $(TEST_INTRA_PRED_SPEED_SRCS) aom.$(VCPROJ_SFX) gtest.$(VCPROJ_SFX)
+	@echo "    [CREATE] $@"
+	$(qexec)$(GEN_VCPROJ) \
+            --exe \
+            --target=$(TOOLCHAIN) \
+            --name=test_intra_pred_speed \
+            -D_VARIADIC_MAX=10 \
+            --proj-guid=CD837F5F-52D8-4314-A370-895D614166A7 \
+            --ver=$(CONFIG_VS_VERSION) \
+            --src-path-bare="$(SRC_PATH_BARE)" \
+            $(if $(CONFIG_STATIC_MSVCRT),--static-crt) \
+            --out=$@ $(INTERNAL_CFLAGS) $(CFLAGS) \
+            -I. -I"$(SRC_PATH_BARE)/third_party/googletest/src/googletest/include" \
+            -L. -l$(CODEC_LIB) -l$(GTEST_LIB) $^
+endif  # TEST_INTRA_PRED_SPEED
+endif
+else
+
+include $(SRC_PATH_BARE)/third_party/googletest/gtest.mk
+GTEST_SRCS := $(addprefix third_party/,$(call enabled,GTEST_SRCS))
+GTEST_OBJS=$(call objs,$(GTEST_SRCS))
+ifeq ($(filter win%,$(TGT_OS)),$(TGT_OS))
+# Disabling pthreads globally will cause issues on darwin and possibly elsewhere
+$(GTEST_OBJS) $(GTEST_OBJS:.o=.d): CXXFLAGS += -DGTEST_HAS_PTHREAD=0
+endif
+GTEST_INCLUDES := -I$(SRC_PATH_BARE)/third_party/googletest/src/googletest
+GTEST_INCLUDES += -I$(SRC_PATH_BARE)/third_party/googletest/src/googletest/include
+$(GTEST_OBJS) $(GTEST_OBJS:.o=.d): CXXFLAGS += $(GTEST_INCLUDES)
+OBJS-yes += $(GTEST_OBJS)
+LIBS-yes += $(BUILD_PFX)libgtest.a $(BUILD_PFX)libgtest_g.a
+$(BUILD_PFX)libgtest_g.a: $(GTEST_OBJS)
+
+LIBAOM_TEST_OBJS=$(sort $(call objs,$(LIBAOM_TEST_SRCS)))
+$(LIBAOM_TEST_OBJS) $(LIBAOM_TEST_OBJS:.o=.d): CXXFLAGS += $(GTEST_INCLUDES)
+OBJS-yes += $(LIBAOM_TEST_OBJS)
+BINS-yes += $(LIBAOM_TEST_BIN)
+
+CODEC_LIB=$(if $(CONFIG_DEBUG_LIBS),aom_g,aom)
+CODEC_LIB_SUF=$(if $(CONFIG_SHARED),$(SHARED_LIB_SUF),.a)
+TEST_LIBS := lib$(CODEC_LIB)$(CODEC_LIB_SUF) libgtest.a
+$(LIBAOM_TEST_BIN): $(TEST_LIBS)
+$(eval $(call linkerxx_template,$(LIBAOM_TEST_BIN), \
+              $(LIBAOM_TEST_OBJS) \
+              -L. -laom -lgtest $(extralibs) -lm))
+
+ifneq ($(strip $(TEST_INTRA_PRED_SPEED_OBJS)),)
+$(TEST_INTRA_PRED_SPEED_OBJS) $(TEST_INTRA_PRED_SPEED_OBJS:.o=.d): CXXFLAGS += $(GTEST_INCLUDES)
+OBJS-yes += $(TEST_INTRA_PRED_SPEED_OBJS)
+BINS-yes += $(TEST_INTRA_PRED_SPEED_BIN)
+
+$(TEST_INTRA_PRED_SPEED_BIN): $(TEST_LIBS)
+$(eval $(call linkerxx_template,$(TEST_INTRA_PRED_SPEED_BIN), \
+              $(TEST_INTRA_PRED_SPEED_OBJS) \
+              -L. -laom -lgtest $(extralibs) -lm))
+endif  # TEST_INTRA_PRED_SPEED
+
+endif  # CONFIG_UNIT_TESTS
+
+# Install test sources only if codec source is included
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += $(patsubst $(SRC_PATH_BARE)/%,%,\
+    $(shell find $(SRC_PATH_BARE)/third_party/googletest -type f))
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += $(LIBAOM_TEST_SRCS)
+INSTALL-SRCS-$(CONFIG_CODEC_SRCS) += $(TEST_INTRA_PRED_SPEED_SRCS)
+
+define test_shard_template
+test:: test_shard.$(1)
+test-no-data-check:: test_shard_ndc.$(1)
+test_shard.$(1) test_shard_ndc.$(1): $(LIBAOM_TEST_BIN)
+	@set -e; \
+	 export GTEST_SHARD_INDEX=$(1); \
+	 export GTEST_TOTAL_SHARDS=$(2); \
+	 $(LIBAOM_TEST_BIN)
+test_shard.$(1): testdata
+.PHONY: test_shard.$(1)
+endef
+
+NUM_SHARDS := 10
+SHARDS := 0 1 2 3 4 5 6 7 8 9
+$(foreach s,$(SHARDS),$(eval $(call test_shard_template,$(s),$(NUM_SHARDS))))
+
+endif
+
+##
+## documentation directives
+##
+CLEAN-OBJS += libs.doxy
+DOCS-yes += libs.doxy
+libs.doxy: $(CODEC_DOC_SRCS)
+	@echo "    [CREATE] $@"
+	@rm -f $@
+	@echo "INPUT += $^" >> $@
+	@echo "INCLUDE_PATH += ." >> $@;
+	@echo "ENABLED_SECTIONS += $(sort $(CODEC_DOC_SECTIONS))" >> $@
+
+## Generate rtcd.h for all objects
+ifeq ($(CONFIG_DEPENDENCY_TRACKING),yes)
+$(OBJS-yes:.o=.d): $(RTCD)
+else
+$(OBJS-yes): $(RTCD)
+endif
+
+## Update the global src list
+SRCS += $(CODEC_SRCS) $(LIBAOM_TEST_SRCS) $(GTEST_SRCS)
+
+##
+## aomdec/aomenc tests.
+##
+ifeq ($(CONFIG_UNIT_TESTS),yes)
+TEST_BIN_PATH = .
+ifeq ($(CONFIG_MSVS),yes)
+# MSVC will build both Debug and Release configurations of tools in a
+# sub directory named for the current target. Assume the user wants to
+# run the Release tools, and assign TEST_BIN_PATH accordingly.
+# TODO(tomfinegan): Is this adequate for ARM?
+# TODO(tomfinegan): Support running the debug versions of tools?
+TEST_BIN_PATH := $(addsuffix /$(TGT_OS:win64=x64)/Release, $(TEST_BIN_PATH))
+endif
+utiltest utiltest-no-data-check:
+	$(qexec)$(SRC_PATH_BARE)/test/aomdec.sh \
+		--test-data-path $(LIBAOM_TEST_DATA_PATH) \
+		--bin-path $(TEST_BIN_PATH)
+	$(qexec)$(SRC_PATH_BARE)/test/aomenc.sh \
+		--test-data-path $(LIBAOM_TEST_DATA_PATH) \
+		--bin-path $(TEST_BIN_PATH)
+utiltest: testdata
+else
+utiltest utiltest-no-data-check:
+	@echo Unit tests must be enabled to make the utiltest target.
+endif
+
+##
+## Example tests.
+##
+ifeq ($(CONFIG_UNIT_TESTS),yes)
+# All non-MSVC targets output example targets in a sub dir named examples.
+EXAMPLES_BIN_PATH = examples
+ifeq ($(CONFIG_MSVS),yes)
+# MSVC will build both Debug and Release configurations of the examples in a
+# sub directory named for the current target. Assume the user wants to
+# run the Release tools, and assign EXAMPLES_BIN_PATH accordingly.
+# TODO(tomfinegan): Is this adequate for ARM?
+# TODO(tomfinegan): Support running the debug versions of tools?
+EXAMPLES_BIN_PATH := $(TGT_OS:win64=x64)/Release
+endif
+exampletest exampletest-no-data-check: examples
+	$(qexec)$(SRC_PATH_BARE)/test/examples.sh \
+		--test-data-path $(LIBAOM_TEST_DATA_PATH) \
+		--bin-path $(EXAMPLES_BIN_PATH)
+exampletest: testdata
+else
+exampletest exampletest-no-data-check:
+	@echo Unit tests must be enabled to make the exampletest target.
+endif
diff --git a/third_party/aom/mainpage.dox b/third_party/aom/mainpage.dox
new file mode 100644
index 0000000000..9a82f43603
--- /dev/null
+++ b/third_party/aom/mainpage.dox
@@ -0,0 +1,52 @@
+/*!\mainpage AMedia Codec SDK
+
+  \section main_contents Page Contents
+  - \ref main_intro
+  - \ref main_startpoints
+  - \ref main_support
+
+  \section main_intro Introduction
+  Welcome to the AMedia Codec SDK. This SDK allows you to integrate your
+  applications with the AOM and AV1 video codecs.
+
+  This distribution of the AOMedia Codec SDK includes the following support:
+
+  \if aom_encoder
+  - \ref aom_encoder
+  \endif
+  \if aom_decoder
+  - \ref aom_decoder
+  \endif
+
+
+  \section main_startpoints Starting Points
+  - Consult the \ref changelog for a complete list of improvements in this
+    release.
+  - The \ref readme contains instructions on recompiling the sample applications.
+  - Read the \ref usage "usage" for a narrative on codec usage.
+  - Read the \ref samples "sample code" for examples of how to interact with the
+    codec.
+  - \ref codec reference
+  \if encoder
+  - \ref encoder reference
+  \endif
+  \if decoder
+  - \ref decoder reference
+  \endif
+
+  \section main_support Support Options & FAQ
+  The AOMedia project is an open source project supported by its community. For
+  questions about this SDK, please mail the apps-devel@webmproject.org list.
+  To contribute, see http://www.webmproject.org/code/contribute and mail
+  codec-devel@webmproject.org.
+*/
+
+/*!\page changelog CHANGELOG
+   \verbinclude CHANGELOG
+*/
+
+/*!\page readme README
+   \verbinclude README
+*/
+
+/*!\defgroup codecs Supported Codecs */
diff --git a/third_party/aom/md5_utils.c b/third_party/aom/md5_utils.c
new file mode 100644
index 0000000000..34012b205b
--- /dev/null
+++ b/third_party/aom/md5_utils.c
@@ -0,0 +1,249 @@
+/*
+ * This code implements the MD5 message-digest algorithm.
+ * The algorithm is due to Ron Rivest.  This code was
+ * written by Colin Plumb in 1993, no copyright is claimed.
+ * This code is in the public domain; do with it what you wish.
+ *
+ * Equivalent code is available from RSA Data Security, Inc.
+ * This code has been tested against that, and is equivalent,
+ * except that you don't need to include two pages of legalese
+ * with every copy.
+ *
+ * To compute the message digest of a chunk of bytes, declare an
+ * MD5Context structure, pass it to MD5Init, call MD5Update as
+ * needed on buffers full of bytes, and then call MD5Final, which
+ * will fill a supplied 16-byte array with the digest.
+ *
+ * Changed so as no longer to depend on Colin Plumb's `usual.h' header
+ * definitions
+ *  - Ian Jackson <ian@chiark.greenend.org.uk>.
+ * Still in the public domain.
+ */
+
+#include <string.h> /* for memcpy() */
+
+#include "md5_utils.h"
+
+static void byteSwap(UWORD32 *buf, unsigned words) {
+  md5byte *p;
+
+  /* Only swap bytes for big endian machines */
+  int i = 1;
+
+  if (*(char *)&i == 1) return;
+
+  p = (md5byte *)buf;
+
+  do {
+    *buf++ = (UWORD32)((unsigned)p[3] << 8 | p[2]) << 16 |
+             ((unsigned)p[1] << 8 | p[0]);
+    p += 4;
+  } while (--words);
+}
+
+/*
+ * Start MD5 accumulation.  Set bit count to 0 and buffer to mysterious
+ * initialization constants.
+ */
+void MD5Init(struct MD5Context *ctx) {
+  ctx->buf[0] = 0x67452301;
+  ctx->buf[1] = 0xefcdab89;
+  ctx->buf[2] = 0x98badcfe;
+  ctx->buf[3] = 0x10325476;
+
+  ctx->bytes[0] = 0;
+  ctx->bytes[1] = 0;
+}
+
+/*
+ * Update context to reflect the concatenation of another buffer full
+ * of bytes.
+ */
+void MD5Update(struct MD5Context *ctx, md5byte const *buf, unsigned len) {
+  UWORD32 t;
+
+  /* Update byte count */
+
+  t = ctx->bytes[0];
+
+  if ((ctx->bytes[0] = t + len) < t)
+    ctx->bytes[1]++; /* Carry from low to high */
+
+  t = 64 - (t & 0x3f); /* Space available in ctx->in (at least 1) */
+
+  if (t > len) {
+    memcpy((md5byte *)ctx->in + 64 - t, buf, len);
+    return;
+  }
+
+  /* First chunk is an odd size */
+  memcpy((md5byte *)ctx->in + 64 - t, buf, t);
+  byteSwap(ctx->in, 16);
+  MD5Transform(ctx->buf, ctx->in);
+  buf += t;
+  len -= t;
+
+  /* Process data in 64-byte chunks */
+  while (len >= 64) {
+    memcpy(ctx->in, buf, 64);
+    byteSwap(ctx->in, 16);
+    MD5Transform(ctx->buf, ctx->in);
+    buf += 64;
+    len -= 64;
+  }
+
+  /* Handle any remaining bytes of data. */
+  memcpy(ctx->in, buf, len);
+}
+
+/*
+ * Final wrapup - pad to 64-byte boundary with the bit pattern
+ * 1 0* (64-bit count of bits processed, MSB-first)
+ */
+void MD5Final(md5byte digest[16], struct MD5Context *ctx) {
+  int count = ctx->bytes[0] & 0x3f; /* Number of bytes in ctx->in */
+  md5byte *p = (md5byte *)ctx->in + count;
+
+  /* Set the first char of padding to 0x80.  There is always room. */
+  *p++ = 0x80;
+
+  /* Bytes of padding needed to make 56 bytes (-8..55) */
+  count = 56 - 1 - count;
+
+  if (count < 0) { /* Padding forces an extra block */
+    memset(p, 0, count + 8);
+    byteSwap(ctx->in, 16);
+    MD5Transform(ctx->buf, ctx->in);
+    p = (md5byte *)ctx->in;
+    count = 56;
+  }
+
+  memset(p, 0, count);
+  byteSwap(ctx->in, 14);
+
+  /* Append length in bits and transform */
+  ctx->in[14] = ctx->bytes[0] << 3;
+  ctx->in[15] = ctx->bytes[1] << 3 | ctx->bytes[0] >> 29;
+  MD5Transform(ctx->buf, ctx->in);
+
+  byteSwap(ctx->buf, 4);
+  memcpy(digest, ctx->buf, 16);
+  memset(ctx, 0, sizeof(*ctx)); /* In case it's sensitive */
+}
+
+#ifndef ASM_MD5
+
+/* The four core functions - F1 is optimized somewhat */
+
+/* #define F1(x, y, z) (x & y | ~x & z) */
+#define F1(x, y, z) (z ^ (x & (y ^ z)))
+#define F2(x, y, z) F1(z, x, y)
+#define F3(x, y, z) (x ^ y ^ z)
+#define F4(x, y, z) (y ^ (x | ~z))
+
+/* This is the central step in the MD5 algorithm. */
+#define MD5STEP(f, w, x, y, z, in, s) \
+  (w += f(x, y, z) + in, w = (w << s | w >> (32 - s)) + x)
+
+#if defined(__clang__) && defined(__has_attribute)
+#if __has_attribute(no_sanitize)
+#define AOM_NO_UNSIGNED_OVERFLOW_CHECK \
+  __attribute__((no_sanitize("unsigned-integer-overflow")))
+#endif
+#endif
+
+#ifndef AOM_NO_UNSIGNED_OVERFLOW_CHECK
+#define AOM_NO_UNSIGNED_OVERFLOW_CHECK
+#endif
+
+/*
+ * The core of the MD5 algorithm, this alters an existing MD5 hash to
+ * reflect the addition of 16 longwords of new data.  MD5Update blocks
+ * the data and converts bytes into longwords for this routine.
+ */
+AOM_NO_UNSIGNED_OVERFLOW_CHECK void MD5Transform(UWORD32 buf[4],
+                                                 UWORD32 const in[16]) {
+  register UWORD32 a, b, c, d;
+
+  a = buf[0];
+  b = buf[1];
+  c = buf[2];
+  d = buf[3];
+
+  MD5STEP(F1, a, b, c, d, in[0] + 0xd76aa478, 7);
+  MD5STEP(F1, d, a, b, c, in[1] + 0xe8c7b756, 12);
+  MD5STEP(F1, c, d, a, b, in[2] + 0x242070db, 17);
+  MD5STEP(F1, b, c, d, a, in[3] + 0xc1bdceee, 22);
+  MD5STEP(F1, a, b, c, d, in[4] + 0xf57c0faf, 7);
+  MD5STEP(F1, d, a, b, c, in[5] + 0x4787c62a, 12);
+  MD5STEP(F1, c, d, a, b, in[6] + 0xa8304613, 17);
+  MD5STEP(F1, b, c, d, a, in[7] + 0xfd469501, 22);
+  MD5STEP(F1, a, b, c, d, in[8] + 0x698098d8, 7);
+  MD5STEP(F1, d, a, b, c, in[9] + 0x8b44f7af, 12);
+  MD5STEP(F1, c, d, a, b, in[10] + 0xffff5bb1, 17);
+  MD5STEP(F1, b, c, d, a, in[11] + 0x895cd7be, 22);
+  MD5STEP(F1, a, b, c, d, in[12] + 0x6b901122, 7);
+  MD5STEP(F1, d, a, b, c, in[13] + 0xfd987193, 12);
+  MD5STEP(F1, c, d, a, b, in[14] + 0xa679438e, 17);
+  MD5STEP(F1, b, c, d, a, in[15] + 0x49b40821, 22);
+
+  MD5STEP(F2, a, b, c, d, in[1] + 0xf61e2562, 5);
+  MD5STEP(F2, d, a, b, c, in[6] + 0xc040b340, 9);
+  MD5STEP(F2, c, d, a, b, in[11] + 0x265e5a51, 14);
+  MD5STEP(F2, b, c, d, a, in[0] + 0xe9b6c7aa, 20);
+  MD5STEP(F2, a, b, c, d, in[5] + 0xd62f105d, 5);
+  MD5STEP(F2, d, a, b, c, in[10] + 0x02441453, 9);
+  MD5STEP(F2, c, d, a, b, in[15] + 0xd8a1e681, 14);
+  MD5STEP(F2, b, c, d, a, in[4] + 0xe7d3fbc8, 20);
+  MD5STEP(F2, a, b, c, d, in[9] + 0x21e1cde6, 5);
+  MD5STEP(F2, d, a, b, c, in[14] + 0xc33707d6, 9);
+  MD5STEP(F2, c, d, a, b, in[3] + 0xf4d50d87, 14);
+  MD5STEP(F2, b, c, d, a, in[8] + 0x455a14ed, 20);
+  MD5STEP(F2, a, b, c, d, in[13] + 0xa9e3e905, 5);
+  MD5STEP(F2, d, a, b, c, in[2] + 0xfcefa3f8, 9);
+  MD5STEP(F2, c, d, a, b, in[7] + 0x676f02d9, 14);
+  MD5STEP(F2, b, c, d, a, in[12] + 0x8d2a4c8a, 20);
+
+  MD5STEP(F3, a, b, c, d, in[5] + 0xfffa3942, 4);
+  MD5STEP(F3, d, a, b, c, in[8] + 0x8771f681, 11);
+  MD5STEP(F3, c, d, a, b, in[11] + 0x6d9d6122, 16);
+  MD5STEP(F3, b, c, d, a, in[14] + 0xfde5380c, 23);
+  MD5STEP(F3, a, b, c, d, in[1] + 0xa4beea44, 4);
+  MD5STEP(F3, d, a, b, c, in[4] + 0x4bdecfa9, 11);
+  MD5STEP(F3, c, d, a, b, in[7] + 0xf6bb4b60, 16);
+  MD5STEP(F3, b, c, d, a, in[10] + 0xbebfbc70, 23);
+  MD5STEP(F3, a, b, c, d, in[13] + 0x289b7ec6, 4);
+  MD5STEP(F3, d, a, b, c, in[0] + 0xeaa127fa, 11);
+  MD5STEP(F3, c, d, a, b, in[3] + 0xd4ef3085, 16);
+  MD5STEP(F3, b, c, d, a, in[6] + 0x04881d05, 23);
+  MD5STEP(F3, a, b, c, d, in[9] + 0xd9d4d039, 4);
+  MD5STEP(F3, d, a, b, c, in[12] + 0xe6db99e5, 11);
+  MD5STEP(F3, c, d, a, b, in[15] + 0x1fa27cf8, 16);
+  MD5STEP(F3, b, c, d, a, in[2] + 0xc4ac5665, 23);
+
+  MD5STEP(F4, a, b, c, d, in[0] + 0xf4292244, 6);
+  MD5STEP(F4, d, a, b, c, in[7] + 0x432aff97, 10);
+  MD5STEP(F4, c, d, a, b, in[14] + 0xab9423a7, 15);
+  MD5STEP(F4, b, c, d, a, in[5] + 0xfc93a039, 21);
+  MD5STEP(F4, a, b, c, d, in[12] + 0x655b59c3, 6);
+  MD5STEP(F4, d, a, b, c, in[3] + 0x8f0ccc92, 10);
+  MD5STEP(F4, c, d, a, b, in[10] + 0xffeff47d, 15);
+  MD5STEP(F4, b, c, d, a, in[1] + 0x85845dd1, 21);
+  MD5STEP(F4, a, b, c, d, in[8] + 0x6fa87e4f, 6);
+  MD5STEP(F4, d, a, b, c, in[15] + 0xfe2ce6e0, 10);
+  MD5STEP(F4, c, d, a, b, in[6] + 0xa3014314, 15);
+  MD5STEP(F4, b, c, d, a, in[13] + 0x4e0811a1, 21);
+  MD5STEP(F4, a, b, c, d, in[4] + 0xf7537e82, 6);
+  MD5STEP(F4, d, a, b, c, in[11] + 0xbd3af235, 10);
+  MD5STEP(F4, c, d, a, b, in[2] + 0x2ad7d2bb, 15);
+  MD5STEP(F4, b, c, d, a, in[9] + 0xeb86d391, 21);
+
+  buf[0] += a;
+  buf[1] += b;
+  buf[2] += c;
+  buf[3] += d;
+}
+
+#undef AOM_NO_UNSIGNED_OVERFLOW_CHECK
+
+#endif
diff --git a/third_party/aom/md5_utils.h b/third_party/aom/md5_utils.h
new file mode 100644
index 0000000000..bd4991b3ad
--- /dev/null
+++ b/third_party/aom/md5_utils.h
@@ -0,0 +1,49 @@
+/*
+ * This is the header file for the MD5 message-digest algorithm.
+ * The algorithm is due to Ron Rivest.  This code was
+ * written by Colin Plumb in 1993, no copyright is claimed.
+ * This code is in the public domain; do with it what you wish.
+ *
+ * Equivalent code is available from RSA Data Security, Inc.
+ * This code has been tested against that, and is equivalent,
+ * except that you don't need to include two pages of legalese
+ * with every copy.
+ *
+ * To compute the message digest of a chunk of bytes, declare an
+ * MD5Context structure, pass it to MD5Init, call MD5Update as
+ * needed on buffers full of bytes, and then call MD5Final, which
+ * will fill a supplied 16-byte array with the digest.
+ *
+ * Changed so as no longer to depend on Colin Plumb's `usual.h'
+ * header definitions
+ *  - Ian Jackson <ian@chiark.greenend.org.uk>.
+ * Still in the public domain.
+ */
+
+#ifndef MD5_UTILS_H_
+#define MD5_UTILS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define md5byte unsigned char
+#define UWORD32 unsigned int
+
+typedef struct MD5Context MD5Context;
+struct MD5Context {
+  UWORD32 buf[4];
+  UWORD32 bytes[2];
+  UWORD32 in[16];
+};
+
+void MD5Init(struct MD5Context *context);
+void MD5Update(struct MD5Context *context, md5byte const *buf, unsigned len);
+void MD5Final(unsigned char digest[16], struct MD5Context *context);
+void MD5Transform(UWORD32 buf[4], UWORD32 const in[16]);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // MD5_UTILS_H_
diff --git a/third_party/aom/rate_hist.c b/third_party/aom/rate_hist.c
new file mode 100644
index 0000000000..ffc7b8997e
--- /dev/null
+++ b/third_party/aom/rate_hist.c
@@ -0,0 +1,271 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <stdlib.h>
+#include <limits.h>
+#include <stdio.h>
+#include <math.h>
+
+#include "./rate_hist.h"
+
+#define RATE_BINS 100
+#define HIST_BAR_MAX 40
+
+struct hist_bucket {
+  int low;
+  int high;
+  int count;
+};
+
+struct rate_hist {
+  int64_t *pts;
+  int *sz;
+  int samples;
+  int frames;
+  struct hist_bucket bucket[RATE_BINS];
+  int total;
+};
+
+struct rate_hist *init_rate_histogram(const aom_codec_enc_cfg_t *cfg,
+                                      const aom_rational_t *fps) {
+  int i;
+  struct rate_hist *hist = malloc(sizeof(*hist));
+
+  // Determine the number of samples in the buffer. Use the file's framerate
+  // to determine the number of frames in rc_buf_sz milliseconds, with an
+  // adjustment (5/4) to account for alt-refs
+  hist->samples = cfg->rc_buf_sz * 5 / 4 * fps->num / fps->den / 1000;
+
+  // prevent division by zero
+  if (hist->samples == 0) hist->samples = 1;
+
+  hist->frames = 0;
+  hist->total = 0;
+
+  hist->pts = calloc(hist->samples, sizeof(*hist->pts));
+  hist->sz = calloc(hist->samples, sizeof(*hist->sz));
+  for (i = 0; i < RATE_BINS; i++) {
+    hist->bucket[i].low = INT_MAX;
+    hist->bucket[i].high = 0;
+    hist->bucket[i].count = 0;
+  }
+
+  return hist;
+}
+
+void destroy_rate_histogram(struct rate_hist *hist) {
+  if (hist) {
+    free(hist->pts);
+    free(hist->sz);
+    free(hist);
+  }
+}
+
+void update_rate_histogram(struct rate_hist *hist,
+                           const aom_codec_enc_cfg_t *cfg,
+                           const aom_codec_cx_pkt_t *pkt) {
+  int i;
+  int64_t then = 0;
+  int64_t avg_bitrate = 0;
+  int64_t sum_sz = 0;
+  const int64_t now = pkt->data.frame.pts * 1000 *
+                      (uint64_t)cfg->g_timebase.num /
+                      (uint64_t)cfg->g_timebase.den;
+
+  int idx = hist->frames++ % hist->samples;
+  hist->pts[idx] = now;
+  hist->sz[idx] = (int)pkt->data.frame.sz;
+
+  if (now < cfg->rc_buf_initial_sz) return;
+
+  if (!cfg->rc_target_bitrate) return;
+
+  then = now;
+
+  /* Sum the size over the past rc_buf_sz ms */
+  for (i = hist->frames; i > 0 && hist->frames - i < hist->samples; i--) {
+    const int i_idx = (i - 1) % hist->samples;
+
+    then = hist->pts[i_idx];
+    if (now - then > cfg->rc_buf_sz) break;
+    sum_sz += hist->sz[i_idx];
+  }
+
+  if (now == then) return;
+
+  avg_bitrate = sum_sz * 8 * 1000 / (now - then);
+  idx = (int)(avg_bitrate * (RATE_BINS / 2) / (cfg->rc_target_bitrate * 1000));
+  if (idx < 0) idx = 0;
+  if (idx > RATE_BINS - 1) idx = RATE_BINS - 1;
+  if (hist->bucket[idx].low > avg_bitrate)
+    hist->bucket[idx].low = (int)avg_bitrate;
+  if (hist->bucket[idx].high < avg_bitrate)
+    hist->bucket[idx].high = (int)avg_bitrate;
+  hist->bucket[idx].count++;
+  hist->total++;
+}
+
+static int merge_hist_buckets(struct hist_bucket *bucket, int max_buckets,
+                              int *num_buckets) {
+  int small_bucket = 0, merge_bucket = INT_MAX, big_bucket = 0;
+  int buckets = *num_buckets;
+  int i;
+
+  /* Find the extrema for this list of buckets */
+  big_bucket = small_bucket = 0;
+  for (i = 0; i < buckets; i++) {
+    if (bucket[i].count < bucket[small_bucket].count) small_bucket = i;
+    if (bucket[i].count > bucket[big_bucket].count) big_bucket = i;
+  }
+
+  /* If we have too many buckets, merge the smallest with an adjacent
+   * bucket.
+   */
+  while (buckets > max_buckets) {
+    int last_bucket = buckets - 1;
+
+    /* merge the small bucket with an adjacent one. */
+    if (small_bucket == 0)
+      merge_bucket = 1;
+    else if (small_bucket == last_bucket)
+      merge_bucket = last_bucket - 1;
+    else if (bucket[small_bucket - 1].count < bucket[small_bucket + 1].count)
+      merge_bucket = small_bucket - 1;
+    else
+      merge_bucket = small_bucket + 1;
+
+    assert(abs(merge_bucket - small_bucket) <= 1);
+    assert(small_bucket < buckets);
+    assert(big_bucket < buckets);
+    assert(merge_bucket < buckets);
+
+    if (merge_bucket < small_bucket) {
+      bucket[merge_bucket].high = bucket[small_bucket].high;
+      bucket[merge_bucket].count += bucket[small_bucket].count;
+    } else {
+      bucket[small_bucket].high = bucket[merge_bucket].high;
+      bucket[small_bucket].count += bucket[merge_bucket].count;
+      merge_bucket = small_bucket;
+    }
+
+    assert(bucket[merge_bucket].low != bucket[merge_bucket].high);
+
+    buckets--;
+
+    /* Remove the merge_bucket from the list, and find the new small
+     * and big buckets while we're at it
+     */
+    big_bucket = small_bucket = 0;
+    for (i = 0; i < buckets; i++) {
+      if (i > merge_bucket) bucket[i] = bucket[i + 1];
+
+      if (bucket[i].count < bucket[small_bucket].count) small_bucket = i;
+      if (bucket[i].count > bucket[big_bucket].count) big_bucket = i;
+    }
+  }
+
+  *num_buckets = buckets;
+  return bucket[big_bucket].count;
+}
+
+static void show_histogram(const struct hist_bucket *bucket, int buckets,
+                           int total, int scale) {
+  const char *pat1, *pat2;
+  int i;
+
+  switch ((int)(log(bucket[buckets - 1].high) / log(10)) + 1) {
+    case 1:
+    case 2:
+      pat1 = "%4d %2s: ";
+      pat2 = "%4d-%2d: ";
+      break;
+    case 3:
+      pat1 = "%5d %3s: ";
+      pat2 = "%5d-%3d: ";
+      break;
+    case 4:
+      pat1 = "%6d %4s: ";
+      pat2 = "%6d-%4d: ";
+      break;
+    case 5:
+      pat1 = "%7d %5s: ";
+      pat2 = "%7d-%5d: ";
+      break;
+    case 6:
+      pat1 = "%8d %6s: ";
+      pat2 = "%8d-%6d: ";
+      break;
+    case 7:
+      pat1 = "%9d %7s: ";
+      pat2 = "%9d-%7d: ";
+      break;
+    default:
+      pat1 = "%12d %10s: ";
+      pat2 = "%12d-%10d: ";
+      break;
+  }
+
+  for (i = 0; i < buckets; i++) {
+    int len;
+    int j;
+    float pct;
+
+    pct = (float)(100.0 * bucket[i].count / total);
+    len = HIST_BAR_MAX * bucket[i].count / scale;
+    if (len < 1) len = 1;
+    assert(len <= HIST_BAR_MAX);
+
+    if (bucket[i].low == bucket[i].high)
+      fprintf(stderr, pat1, bucket[i].low, "");
+    else
+      fprintf(stderr, pat2, bucket[i].low, bucket[i].high);
+
+    for (j = 0; j < HIST_BAR_MAX; j++) fprintf(stderr, j < len ? "=" : " ");
+    fprintf(stderr, "\t%5d (%6.2f%%)\n", bucket[i].count, pct);
+  }
+}
+
+void show_q_histogram(const int counts[64], int max_buckets) {
+  struct hist_bucket bucket[64];
+  int buckets = 0;
+  int total = 0;
+  int scale;
+  int i;
+
+  for (i = 0; i < 64; i++) {
+    if (counts[i]) {
+      bucket[buckets].low = bucket[buckets].high = i;
+      bucket[buckets].count = counts[i];
+      buckets++;
+      total += counts[i];
+    }
+  }
+
+  fprintf(stderr, "\nQuantizer Selection:\n");
+  scale = merge_hist_buckets(bucket, max_buckets, &buckets);
+  show_histogram(bucket, buckets, total, scale);
+}
+
+void show_rate_histogram(struct rate_hist *hist, const aom_codec_enc_cfg_t *cfg,
+                         int max_buckets) {
+  int i, scale;
+  int buckets = 0;
+
+  for (i = 0; i < RATE_BINS; i++) {
+    if (hist->bucket[i].low == INT_MAX) continue;
+    hist->bucket[buckets++] = hist->bucket[i];
+  }
+
+  fprintf(stderr, "\nRate (over %dms window):\n", cfg->rc_buf_sz);
+  scale = merge_hist_buckets(hist->bucket, max_buckets, &buckets);
+  show_histogram(hist->bucket, buckets, hist->total, scale);
+}
diff --git a/third_party/aom/rate_hist.h b/third_party/aom/rate_hist.h
new file mode 100644
index 0000000000..e6aa149aef
--- /dev/null
+++ b/third_party/aom/rate_hist.h
@@ -0,0 +1,41 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef RATE_HIST_H_
+#define RATE_HIST_H_
+
+#include "aom/aom_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct rate_hist;
+
+struct rate_hist *init_rate_histogram(const aom_codec_enc_cfg_t *cfg,
+                                      const aom_rational_t *fps);
+
+void destroy_rate_histogram(struct rate_hist *hist);
+
+void update_rate_histogram(struct rate_hist *hist,
+                           const aom_codec_enc_cfg_t *cfg,
+                           const aom_codec_cx_pkt_t *pkt);
+
+void show_q_histogram(const int counts[64], int max_buckets);
+
+void show_rate_histogram(struct rate_hist *hist, const aom_codec_enc_cfg_t *cfg,
+                         int max_buckets);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // RATE_HIST_H_
diff --git a/third_party/aom/solution.mk b/third_party/aom/solution.mk
new file mode 100644
index 0000000000..caa8bc17bf
--- /dev/null
+++ b/third_party/aom/solution.mk
@@ -0,0 +1,33 @@
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+
+# libaom reverse dependencies (targets that depend on libaom)
+AOM_NONDEPS=$(addsuffix .$(VCPROJ_SFX),aom gtest)
+AOM_RDEPS=$(foreach vcp,\
+              $(filter-out $(AOM_NONDEPS),$^), --dep=$(vcp:.$(VCPROJ_SFX)=):aom)
+
+aom.sln: $(wildcard *.$(VCPROJ_SFX))
+	@echo "    [CREATE] $@"
+	$(SRC_PATH_BARE)/build/make/gen_msvs_sln.sh \
+            $(if $(filter aom.$(VCPROJ_SFX),$^),$(AOM_RDEPS)) \
+            --dep=test_libaom:gtest \
+            --ver=$(CONFIG_VS_VERSION)\
+            --out=$@ $^
+aom.sln.mk: aom.sln
+	@true
+
+PROJECTS-yes += aom.sln aom.sln.mk
+-include aom.sln.mk
+
+# Always install this file, as it is an unconditional post-build rule.
+INSTALL_MAPS += src/%     $(SRC_PATH_BARE)/%
+INSTALL-SRCS-yes            += $(target).mk
diff --git a/third_party/aom/test/accounting_test.cc b/third_party/aom/test/accounting_test.cc
new file mode 100644
index 0000000000..e8387d0dc0
--- /dev/null
+++ b/third_party/aom/test/accounting_test.cc
@@ -0,0 +1,78 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/acm_random.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/bitreader.h"
+#include "aom_dsp/bitwriter.h"
+
+using libaom_test::ACMRandom;
+
+TEST(AV1, TestAccounting) {
+  const int kBufferSize = 10000;
+  const int kSymbols = 1024;
+  aom_writer bw;
+  uint8_t bw_buffer[kBufferSize];
+  aom_start_encode(&bw, bw_buffer);
+  for (int i = 0; i < kSymbols; i++) {
+    aom_write(&bw, 0, 32);
+    aom_write(&bw, 0, 32);
+    aom_write(&bw, 0, 32);
+  }
+  aom_stop_encode(&bw);
+  aom_reader br;
+#if CONFIG_ANS && ANS_MAX_SYMBOLS
+  br.window_size = 1 << 16;
+#endif
+  aom_reader_init(&br, bw_buffer, bw.pos, NULL, NULL);
+
+  Accounting accounting;
+  aom_accounting_init(&accounting);
+  br.accounting = &accounting;
+  for (int i = 0; i < kSymbols; i++) {
+    aom_read(&br, 32, "A");
+  }
+  // Consecutive symbols that are the same are coalesced.
+  GTEST_ASSERT_EQ(accounting.syms.num_syms, 1);
+  GTEST_ASSERT_EQ(accounting.syms.syms[0].samples, (unsigned int)kSymbols);
+
+  aom_accounting_reset(&accounting);
+  GTEST_ASSERT_EQ(accounting.syms.num_syms, 0);
+
+  // Should record 2 * kSymbols accounting symbols.
+  aom_reader_init(&br, bw_buffer, bw.pos, NULL, NULL);
+  br.accounting = &accounting;
+  for (int i = 0; i < kSymbols; i++) {
+    aom_read(&br, 32, "A");
+    aom_read(&br, 32, "B");
+    aom_read(&br, 32, "B");
+  }
+  GTEST_ASSERT_EQ(accounting.syms.num_syms, kSymbols * 2);
+  uint32_t tell_frac = aom_reader_tell_frac(&br);
+  for (int i = 0; i < accounting.syms.num_syms; i++) {
+    tell_frac -= accounting.syms.syms[i].bits;
+  }
+  GTEST_ASSERT_EQ(tell_frac, 0U);
+
+  GTEST_ASSERT_EQ(aom_accounting_dictionary_lookup(&accounting, "A"),
+                  aom_accounting_dictionary_lookup(&accounting, "A"));
+
+  // Check for collisions. The current aom_accounting_hash function returns
+  // the same hash code for AB and BA.
+  GTEST_ASSERT_NE(aom_accounting_dictionary_lookup(&accounting, "AB"),
+                  aom_accounting_dictionary_lookup(&accounting, "BA"));
+}
diff --git a/third_party/aom/test/acm_random.h b/third_party/aom/test/acm_random.h
new file mode 100644
index 0000000000..4842345ff6
--- /dev/null
+++ b/third_party/aom/test/acm_random.h
@@ -0,0 +1,71 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef TEST_ACM_RANDOM_H_
+#define TEST_ACM_RANDOM_H_
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "aom/aom_integer.h"
+
+namespace libaom_test {
+
+class ACMRandom {
+ public:
+  ACMRandom() : random_(DeterministicSeed()) {}
+
+  explicit ACMRandom(int seed) : random_(seed) {}
+
+  void Reset(int seed) { random_.Reseed(seed); }
+
+  uint32_t Rand31(void) {
+    return random_.Generate(testing::internal::Random::kMaxRange);
+  }
+
+  uint16_t Rand16(void) {
+    const uint32_t value =
+        random_.Generate(testing::internal::Random::kMaxRange);
+    return (value >> 15) & 0xffff;
+  }
+
+  int16_t Rand9Signed(void) {
+    // Use 9 bits: values between 255 (0x0FF) and -256 (0x100).
+    const uint32_t value = random_.Generate(512);
+    return static_cast<int16_t>(value) - 256;
+  }
+
+  uint8_t Rand8(void) {
+    const uint32_t value =
+        random_.Generate(testing::internal::Random::kMaxRange);
+    // There's a bit more entropy in the upper bits of this implementation.
+    return (value >> 23) & 0xff;
+  }
+
+  uint8_t Rand8Extremes(void) {
+    // Returns a random value near 0 or near 255, to better exercise
+    // saturation behavior.
+    const uint8_t r = Rand8();
+    return r < 128 ? r << 4 : r >> 4;
+  }
+
+  int PseudoUniform(int range) { return random_.Generate(range); }
+
+  int operator()(int n) { return PseudoUniform(n); }
+
+  static int DeterministicSeed(void) { return 0xbaba; }
+
+ private:
+  testing::internal::Random random_;
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_ACM_RANDOM_H_
diff --git a/third_party/aom/test/active_map_refresh_test.cc b/third_party/aom/test/active_map_refresh_test.cc
new file mode 100644
index 0000000000..7ee86e7e67
--- /dev/null
+++ b/third_party/aom/test/active_map_refresh_test.cc
@@ -0,0 +1,129 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <algorithm>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+
+namespace {
+
+// Check if any pixel in a 16x16 macroblock varies between frames.
+int CheckMb(const aom_image_t &current, const aom_image_t &previous, int mb_r,
+            int mb_c) {
+  for (int plane = 0; plane < 3; plane++) {
+    int r = 16 * mb_r;
+    int c0 = 16 * mb_c;
+    int r_top = std::min(r + 16, static_cast<int>(current.d_h));
+    int c_top = std::min(c0 + 16, static_cast<int>(current.d_w));
+    r = std::max(r, 0);
+    c0 = std::max(c0, 0);
+    if (plane > 0 && current.x_chroma_shift) {
+      c_top = (c_top + 1) >> 1;
+      c0 >>= 1;
+    }
+    if (plane > 0 && current.y_chroma_shift) {
+      r_top = (r_top + 1) >> 1;
+      r >>= 1;
+    }
+    for (; r < r_top; ++r) {
+      for (int c = c0; c < c_top; ++c) {
+        if (current.planes[plane][current.stride[plane] * r + c] !=
+            previous.planes[plane][previous.stride[plane] * r + c])
+          return 1;
+      }
+    }
+  }
+  return 0;
+}
+
+void GenerateMap(int mb_rows, int mb_cols, const aom_image_t &current,
+                 const aom_image_t &previous, uint8_t *map) {
+  for (int mb_r = 0; mb_r < mb_rows; ++mb_r) {
+    for (int mb_c = 0; mb_c < mb_cols; ++mb_c) {
+      map[mb_r * mb_cols + mb_c] = CheckMb(current, previous, mb_r, mb_c);
+    }
+  }
+}
+
+const int kAqModeCyclicRefresh = 3;
+
+class ActiveMapRefreshTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<libaom_test::TestMode, int> {
+ protected:
+  ActiveMapRefreshTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~ActiveMapRefreshTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    cpu_used_ = GET_PARAM(2);
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    ::libaom_test::Y4mVideoSource *y4m_video =
+        static_cast<libaom_test::Y4mVideoSource *>(video);
+    if (video->frame() == 1) {
+      encoder->Control(AOME_SET_CPUUSED, cpu_used_);
+      encoder->Control(AV1E_SET_AQ_MODE, kAqModeCyclicRefresh);
+    } else if (video->frame() >= 2 && video->img()) {
+      aom_image_t *current = video->img();
+      aom_image_t *previous = y4m_holder_->img();
+      ASSERT_TRUE(previous != NULL);
+      aom_active_map_t map = aom_active_map_t();
+      const int width = static_cast<int>(current->d_w);
+      const int height = static_cast<int>(current->d_h);
+      const int mb_width = (width + 15) / 16;
+      const int mb_height = (height + 15) / 16;
+      uint8_t *active_map = new uint8_t[mb_width * mb_height];
+      GenerateMap(mb_height, mb_width, *current, *previous, active_map);
+      map.cols = mb_width;
+      map.rows = mb_height;
+      map.active_map = active_map;
+      encoder->Control(AOME_SET_ACTIVEMAP, &map);
+      delete[] active_map;
+    }
+    if (video->img()) {
+      y4m_video->SwapBuffers(y4m_holder_);
+    }
+  }
+
+  int cpu_used_;
+  ::libaom_test::Y4mVideoSource *y4m_holder_;
+};
+
+TEST_P(ActiveMapRefreshTest, Test) {
+  cfg_.g_lag_in_frames = 0;
+  cfg_.g_profile = 1;
+  cfg_.rc_target_bitrate = 600;
+  cfg_.rc_resize_allowed = 0;
+  cfg_.rc_min_quantizer = 8;
+  cfg_.rc_max_quantizer = 30;
+  cfg_.g_pass = AOM_RC_ONE_PASS;
+  cfg_.rc_end_usage = AOM_CBR;
+  cfg_.kf_max_dist = 90000;
+
+  ::libaom_test::Y4mVideoSource video("desktop_credits.y4m", 0, 10);
+  ::libaom_test::Y4mVideoSource video_holder("desktop_credits.y4m", 0, 10);
+  video_holder.Begin();
+  y4m_holder_ = &video_holder;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+AV1_INSTANTIATE_TEST_CASE(ActiveMapRefreshTest,
+                          ::testing::Values(::libaom_test::kRealTime),
+                          ::testing::Range(5, 6));
+}  // namespace
diff --git a/third_party/aom/test/active_map_test.cc b/third_party/aom/test/active_map_test.cc
new file mode 100644
index 0000000000..a926b0faf3
--- /dev/null
+++ b/third_party/aom/test/active_map_test.cc
@@ -0,0 +1,103 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <climits>
+#include <vector>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+class ActiveMapTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<libaom_test::TestMode, int> {
+ protected:
+  static const int kWidth = 208;
+  static const int kHeight = 144;
+
+  ActiveMapTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~ActiveMapTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    cpu_used_ = GET_PARAM(2);
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AOME_SET_CPUUSED, cpu_used_);
+    } else if (video->frame() == 3) {
+      aom_active_map_t map = aom_active_map_t();
+      /* clang-format off */
+      uint8_t active_map[9 * 13] = {
+        1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0,
+        1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0,
+        1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0,
+        1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 0,
+        0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 1,
+        0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 1, 0, 1,
+        0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 1, 0, 1,
+        0, 0, 0, 0, 0, 0, 1, 1, 0, 1, 0, 1, 1,
+        1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 0,
+      };
+      /* clang-format on */
+      map.cols = (kWidth + 15) / 16;
+      map.rows = (kHeight + 15) / 16;
+      ASSERT_EQ(map.cols, 13u);
+      ASSERT_EQ(map.rows, 9u);
+      map.active_map = active_map;
+      encoder->Control(AOME_SET_ACTIVEMAP, &map);
+    } else if (video->frame() == 15) {
+      aom_active_map_t map = aom_active_map_t();
+      map.cols = (kWidth + 15) / 16;
+      map.rows = (kHeight + 15) / 16;
+      map.active_map = NULL;
+      encoder->Control(AOME_SET_ACTIVEMAP, &map);
+    }
+  }
+
+  void DoTest() {
+    // Validate that this non multiple of 64 wide clip encodes
+    cfg_.g_lag_in_frames = 0;
+    cfg_.rc_target_bitrate = 400;
+    cfg_.rc_resize_allowed = 0;
+    cfg_.g_pass = AOM_RC_ONE_PASS;
+    cfg_.rc_end_usage = AOM_CBR;
+    cfg_.kf_max_dist = 90000;
+    ::libaom_test::I420VideoSource video("hantro_odd.yuv", kWidth, kHeight, 30,
+                                         1, 0, 20);
+
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  }
+
+  int cpu_used_;
+};
+
+TEST_P(ActiveMapTest, Test) { DoTest(); }
+
+class ActiveMapTestLarge : public ActiveMapTest {};
+
+TEST_P(ActiveMapTestLarge, Test) { DoTest(); }
+
+AV1_INSTANTIATE_TEST_CASE(ActiveMapTestLarge,
+                          ::testing::Values(::libaom_test::kRealTime),
+                          ::testing::Range(0, 5));
+
+AV1_INSTANTIATE_TEST_CASE(ActiveMapTest,
+                          ::testing::Values(::libaom_test::kRealTime),
+                          ::testing::Range(5, 9));
+
+}  // namespace
diff --git a/third_party/aom/test/altref_test.cc b/third_party/aom/test/altref_test.cc
new file mode 100644
index 0000000000..6dd8b5186f
--- /dev/null
+++ b/third_party/aom/test/altref_test.cc
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+namespace {
+
+class AltRefForcedKeyTestLarge
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<libaom_test::TestMode, int> {
+ protected:
+  AltRefForcedKeyTestLarge()
+      : EncoderTest(GET_PARAM(0)), encoding_mode_(GET_PARAM(1)),
+        cpu_used_(GET_PARAM(2)), forced_kf_frame_num_(1), frame_num_(0) {}
+  virtual ~AltRefForcedKeyTestLarge() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+    cfg_.rc_end_usage = AOM_VBR;
+    cfg_.g_threads = 0;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      encoder->Control(AOME_SET_CPUUSED, cpu_used_);
+      encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+#if CONFIG_AV1_ENCODER
+      // override test default for tile columns if necessary.
+      if (GET_PARAM(0) == &libaom_test::kAV1) {
+        encoder->Control(AV1E_SET_TILE_COLUMNS, 6);
+      }
+#endif
+    }
+    frame_flags_ =
+        (video->frame() == forced_kf_frame_num_) ? AOM_EFLAG_FORCE_KF : 0;
+  }
+
+  virtual void FramePktHook(const aom_codec_cx_pkt_t *pkt) {
+    if (frame_num_ == forced_kf_frame_num_) {
+      ASSERT_TRUE(!!(pkt->data.frame.flags & AOM_FRAME_IS_KEY))
+          << "Frame #" << frame_num_ << " isn't a keyframe!";
+    }
+    ++frame_num_;
+  }
+
+  ::libaom_test::TestMode encoding_mode_;
+  int cpu_used_;
+  unsigned int forced_kf_frame_num_;
+  unsigned int frame_num_;
+};
+
+TEST_P(AltRefForcedKeyTestLarge, Frame1IsKey) {
+  const aom_rational timebase = { 1, 30 };
+  const int lag_values[] = { 3, 15, 25, -1 };
+
+  forced_kf_frame_num_ = 1;
+  for (int i = 0; lag_values[i] != -1; ++i) {
+    frame_num_ = 0;
+    cfg_.g_lag_in_frames = lag_values[i];
+    libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       timebase.den, timebase.num, 0, 30);
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  }
+}
+
+TEST_P(AltRefForcedKeyTestLarge, ForcedFrameIsKey) {
+  const aom_rational timebase = { 1, 30 };
+  const int lag_values[] = { 3, 15, 25, -1 };
+
+  for (int i = 0; lag_values[i] != -1; ++i) {
+    frame_num_ = 0;
+    forced_kf_frame_num_ = lag_values[i] - 1;
+    cfg_.g_lag_in_frames = lag_values[i];
+    libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       timebase.den, timebase.num, 0, 30);
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  }
+}
+
+AV1_INSTANTIATE_TEST_CASE(AltRefForcedKeyTestLarge,
+                          ::testing::Values(::libaom_test::kOnePassGood),
+                          ::testing::Range(0, 9));
+
+}  // namespace
diff --git a/third_party/aom/test/android/Android.mk b/third_party/aom/test/android/Android.mk
new file mode 100644
index 0000000000..74f9d7cbaf
--- /dev/null
+++ b/third_party/aom/test/android/Android.mk
@@ -0,0 +1,58 @@
+#
+# Copyright (c) 2016, Alliance for Open Media. All rights reserved
+#
+# This source code is subject to the terms of the BSD 2 Clause License and
+# the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+# was not distributed with this source code in the LICENSE file, you can
+# obtain it at www.aomedia.org/license/software. If the Alliance for Open
+# Media Patent License 1.0 was not distributed with this source code in the
+# PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+#
+# This make file builds aom_test app for android.
+# The test app itself runs on the command line through adb shell
+# The paths are really messed up as the libaom make file
+# expects to be made from a parent directory.
+CUR_WD := $(call my-dir)
+BINDINGS_DIR := $(CUR_WD)/../../..
+LOCAL_PATH := $(CUR_WD)/../../..
+
+#libwebm
+include $(CLEAR_VARS)
+include $(BINDINGS_DIR)/libaom/third_party/libwebm/Android.mk
+LOCAL_PATH := $(CUR_WD)/../../..
+
+#libaom
+include $(CLEAR_VARS)
+LOCAL_STATIC_LIBRARIES := libwebm
+include $(BINDINGS_DIR)/libaom/build/make/Android.mk
+LOCAL_PATH := $(CUR_WD)/../..
+
+#libgtest
+include $(CLEAR_VARS)
+LOCAL_ARM_MODE := arm
+LOCAL_CPP_EXTENSION := .cc
+LOCAL_MODULE := gtest
+LOCAL_C_INCLUDES := $(LOCAL_PATH)/third_party/googletest/src/googletest/src
+LOCAL_C_INCLUDES += $(LOCAL_PATH)/third_party/googletest/src/googletest/include
+LOCAL_SRC_FILES := ./third_party/googletest/src/googletest/src/gtest-all.cc
+include $(BUILD_STATIC_LIBRARY)
+
+#libaom_test
+include $(CLEAR_VARS)
+LOCAL_ARM_MODE := arm
+LOCAL_MODULE := libaom_test
+LOCAL_STATIC_LIBRARIES := gtest libwebm
+
+ifeq ($(ENABLE_SHARED),1)
+  LOCAL_SHARED_LIBRARIES := aom
+else
+  LOCAL_STATIC_LIBRARIES += aom
+endif
+
+include $(LOCAL_PATH)/test/test.mk
+LOCAL_C_INCLUDES := $(BINDINGS_DIR)
+FILTERED_SRC := $(sort $(filter %.cc %.c, $(LIBAOM_TEST_SRCS-yes)))
+LOCAL_SRC_FILES := $(addprefix ./test/, $(FILTERED_SRC))
+# some test files depend on *_rtcd.h, ensure they're generated first.
+$(eval $(call rtcd_dep_template))
+include $(BUILD_EXECUTABLE)
diff --git a/third_party/aom/test/android/README b/third_party/aom/test/android/README
new file mode 100644
index 0000000000..35c8297386
--- /dev/null
+++ b/third_party/aom/test/android/README
@@ -0,0 +1,32 @@
+Android.mk will build aom unittests on android.
+1) Configure libaom from the parent directory:
+./libaom/configure --target=armv7-android-gcc --enable-external-build \
+  --enable-postproc --disable-install-srcs --enable-multi-res-encoding \
+  --enable-temporal-denoising --disable-unit-tests --disable-install-docs \
+  --disable-examples --disable-runtime-cpu-detect --sdk-path=$NDK
+
+2) From the parent directory, invoke ndk-build:
+NDK_PROJECT_PATH=. ndk-build APP_BUILD_SCRIPT=./libaom/test/android/Android.mk \
+  APP_ABI=armeabi-v7a APP_PLATFORM=android-18 APP_OPTIM=release \
+  APP_STL=gnustl_static
+
+Note: Both adb and ndk-build are available prebuilt at:
+  https://chromium.googlesource.com/android_tools
+
+3) Run get_files.py to download the test files:
+python get_files.py -i /path/to/test-data.sha1 -o /path/to/put/files \
+  -u http://downloads.webmproject.org/test_data/libaom
+
+4) Transfer files to device using adb. Ensure you have proper permissions for
+the target
+
+adb push /path/to/test_files /data/local/tmp
+adb push /path/to/built_libs /data/local/tmp
+
+NOTE: Built_libs defaults to parent_dir/libs/armeabi-v7a
+
+5) Run tests:
+adb shell
+(on device)
+cd /data/local/tmp
+LD_LIBRARY_PATH=. ./aom_test
diff --git a/third_party/aom/test/android/get_files.py b/third_party/aom/test/android/get_files.py
new file mode 100644
index 0000000000..bdae9a315e
--- /dev/null
+++ b/third_party/aom/test/android/get_files.py
@@ -0,0 +1,120 @@
+#
+# Copyright (c) 2016, Alliance for Open Media. All rights reserved
+#
+# This source code is subject to the terms of the BSD 2 Clause License and
+# the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+# was not distributed with this source code in the LICENSE file, you can
+# obtain it at www.aomedia.org/license/software. If the Alliance for Open
+# Media Patent License 1.0 was not distributed with this source code in the
+# PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+#
+# This simple script pulls test files from the webm homepage
+# It is intelligent enough to only pull files if
+#   1) File / test_data folder does not exist
+#   2) SHA mismatch
+
+import pycurl
+import csv
+import hashlib
+import re
+import os.path
+import time
+import itertools
+import sys
+import getopt
+
+#globals
+url = ''
+file_list_path = ''
+local_resource_path = ''
+
+# Helper functions:
+# A simple function which returns the sha hash of a file in hex
+def get_file_sha(filename):
+  try:
+    sha_hash = hashlib.sha1()
+    with open(filename, 'rb') as file:
+      buf = file.read(HASH_CHUNK)
+      while len(buf) > 0:
+        sha_hash.update(buf)
+        buf = file.read(HASH_CHUNK)
+      return sha_hash.hexdigest()
+  except IOError:
+    print "Error reading " + filename
+
+# Downloads a file from a url, and then checks the sha against the passed
+# in sha
+def download_and_check_sha(url, filename, sha):
+  path = os.path.join(local_resource_path, filename)
+  fp = open(path, "wb")
+  curl = pycurl.Curl()
+  curl.setopt(pycurl.URL, url + "/" + filename)
+  curl.setopt(pycurl.WRITEDATA, fp)
+  curl.perform()
+  curl.close()
+  fp.close()
+  return get_file_sha(path) == sha
+
+#constants
+ftp_retries = 3
+
+SHA_COL = 0
+NAME_COL = 1
+EXPECTED_COL = 2
+HASH_CHUNK = 65536
+
+# Main script
+try:
+  opts, args = \
+      getopt.getopt(sys.argv[1:], \
+                    "u:i:o:", ["url=", "input_csv=", "output_dir="])
+except:
+  print 'get_files.py -u <url> -i <input_csv> -o <output_dir>'
+  sys.exit(2)
+
+for opt, arg in opts:
+  if opt == '-u':
+    url = arg
+  elif opt in ("-i", "--input_csv"):
+    file_list_path = os.path.join(arg)
+  elif opt in ("-o", "--output_dir"):
+    local_resource_path = os.path.join(arg)
+
+if len(sys.argv) != 7:
+  print "Expects two paths and a url!"
+  exit(1)
+
+if not os.path.isdir(local_resource_path):
+  os.makedirs(local_resource_path)
+
+file_list_csv = open(file_list_path, "rb")
+
+# Our 'csv' file uses multiple spaces as a delimiter, python's
+# csv class only uses single character delimiters, so we convert them below
+file_list_reader = csv.reader((re.sub(' +', ' ', line) \
+    for line in file_list_csv), delimiter = ' ')
+
+file_shas = []
+file_names = []
+
+for row in file_list_reader:
+  if len(row) != EXPECTED_COL:
+      continue
+  file_shas.append(row[SHA_COL])
+  file_names.append(row[NAME_COL])
+
+file_list_csv.close()
+
+# Download files, only if they don't already exist and have correct shas
+for filename, sha in itertools.izip(file_names, file_shas):
+  path = os.path.join(local_resource_path, filename)
+  if os.path.isfile(path) \
+      and get_file_sha(path) == sha:
+    print path + ' exists, skipping'
+    continue
+  for retry in range(0, ftp_retries):
+    print "Downloading " + path
+    if not download_and_check_sha(url, filename, sha):
+      print "Sha does not match, retrying..."
+    else:
+      break
diff --git a/third_party/aom/test/android/scrape_gtest_log.py b/third_party/aom/test/android/scrape_gtest_log.py
new file mode 100644
index 0000000000..e0c929a5d5
--- /dev/null
+++ b/third_party/aom/test/android/scrape_gtest_log.py
@@ -0,0 +1,60 @@
+#
+# Copyright (c) 2016, Alliance for Open Media. All rights reserved
+#
+# This source code is subject to the terms of the BSD 2 Clause License and
+# the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+# was not distributed with this source code in the LICENSE file, you can
+# obtain it at www.aomedia.org/license/software. If the Alliance for Open
+# Media Patent License 1.0 was not distributed with this source code in the
+# PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+#
+
+"""Standalone script which parses a gtest log for json.
+
+Json is returned returns as an array.  This script is used by the libaom
+waterfall to gather json results mixed in with gtest logs.  This is
+dubious software engineering.
+"""
+
+import getopt
+import json
+import os
+import re
+import sys
+
+
+def main():
+  if len(sys.argv) != 3:
+    print "Expects a file to write json to!"
+    exit(1)
+
+  try:
+    opts, _ = \
+        getopt.getopt(sys.argv[1:], \
+                      'o:', ['output-json='])
+  except getopt.GetOptError:
+    print 'scrape_gtest_log.py -o <output_json>'
+    sys.exit(2)
+
+  output_json = ''
+  for opt, arg in opts:
+    if opt in ('-o', '--output-json'):
+      output_json = os.path.join(arg)
+
+  blob = sys.stdin.read()
+  json_string = '[' + ','.join('{' + x + '}' for x in
+                               re.findall(r'{([^}]*.?)}', blob)) + ']'
+  print blob
+
+  output = json.dumps(json.loads(json_string), indent=4, sort_keys=True)
+  print output
+
+  path = os.path.dirname(output_json)
+  if path and not os.path.exists(path):
+    os.makedirs(path)
+
+  outfile = open(output_json, 'w')
+  outfile.write(output)
+
+if __name__ == '__main__':
+  sys.exit(main())
diff --git a/third_party/aom/test/ans_codec_test.cc b/third_party/aom/test/ans_codec_test.cc
new file mode 100644
index 0000000000..a1b25fbdac
--- /dev/null
+++ b/third_party/aom/test/ans_codec_test.cc
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "aom_dsp/ans.h"
+#include "av1/av1_dx_iface.c"
+
+// A note on ANS_MAX_SYMBOLS == 0:
+// Fused gtest doesn't work with EXPECT_FATAL_FAILURE [1]. Just run with a
+// single iteration and don't try to check the window size if we are unwindowed.
+// [1] https://github.com/google/googletest/issues/356
+
+namespace {
+
+const char kTestVideoName[] = "niklas_1280_720_30.y4m";
+const int kTestVideoFrames = 10;
+
+class AnsCodecTest : public ::libaom_test::EncoderTest,
+                     public ::libaom_test::CodecTestWithParam<int> {
+ protected:
+  AnsCodecTest()
+      : EncoderTest(GET_PARAM(0)), ans_window_size_log2_(GET_PARAM(1)) {}
+
+  virtual ~AnsCodecTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(::libaom_test::kOnePassGood);
+    cfg_.g_lag_in_frames = 25;
+    cfg_.rc_end_usage = AOM_CQ;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+#if ANS_MAX_SYMBOLS
+      encoder->Control(AV1E_SET_ANS_WINDOW_SIZE_LOG2, ans_window_size_log2_);
+#endif
+      // Try to push a high symbol count through the codec
+      encoder->Control(AOME_SET_CQ_LEVEL, 8);
+      encoder->Control(AOME_SET_CPUUSED, 2);
+      encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+      encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
+      encoder->Control(AOME_SET_ARNR_STRENGTH, 5);
+      encoder->Control(AV1E_SET_TILE_COLUMNS, 0);
+      encoder->Control(AV1E_SET_TILE_ROWS, 0);
+    }
+  }
+
+  virtual bool HandleDecodeResult(const aom_codec_err_t res_dec,
+                                  libaom_test::Decoder *decoder) {
+    aom_codec_ctx_t *const av1_decoder = decoder->GetDecoder();
+#if ANS_MAX_SYMBOLS
+    aom_codec_alg_priv_t *const priv =
+        reinterpret_cast<aom_codec_alg_priv_t *>(av1_decoder->priv);
+    FrameWorkerData *const worker_data =
+        reinterpret_cast<FrameWorkerData *>(priv->frame_workers[0].data1);
+    AV1_COMMON *const common = &worker_data->pbi->common;
+
+    EXPECT_EQ(ans_window_size_log2_, common->ans_window_size_log2);
+#endif
+
+    EXPECT_EQ(AOM_CODEC_OK, res_dec) << decoder->DecodeError();
+    return AOM_CODEC_OK == res_dec;
+  }
+
+ private:
+  int ans_window_size_log2_;
+};
+
+TEST_P(AnsCodecTest, BitstreamParms) {
+  testing::internal::scoped_ptr<libaom_test::VideoSource> video(
+      new libaom_test::Y4mVideoSource(kTestVideoName, 0, kTestVideoFrames));
+  ASSERT_TRUE(video.get() != NULL);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video.get()));
+}
+
+#if ANS_MAX_SYMBOLS
+AV1_INSTANTIATE_TEST_CASE(AnsCodecTest, ::testing::Range(8, 24));
+#else
+AV1_INSTANTIATE_TEST_CASE(AnsCodecTest, ::testing::Range(0, 1));
+#endif
+}  // namespace
diff --git a/third_party/aom/test/ans_test.cc b/third_party/aom/test/ans_test.cc
new file mode 100644
index 0000000000..a553a9e842
--- /dev/null
+++ b/third_party/aom/test/ans_test.cc
@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+#include <ctime>
+#include <utility>
+#include <vector>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/acm_random.h"
+#include "aom_dsp/ansreader.h"
+#include "aom_dsp/buf_ans.h"
+
+namespace {
+typedef std::vector<std::pair<uint8_t, bool> > PvVec;
+
+const int kPrintStats = 0;
+// Use a small buffer size to exercise ANS window spills or buffer growth
+const int kBufAnsSize = 1 << 8;
+
+PvVec abs_encode_build_vals(int iters) {
+  PvVec ret;
+  libaom_test::ACMRandom gen(0x30317076);
+  double entropy = 0;
+  for (int i = 0; i < iters; ++i) {
+    uint8_t p;
+    do {
+      p = gen.Rand8();
+    } while (p == 0);  // zero is not a valid coding probability
+    bool b = gen.Rand8() < p;
+    ret.push_back(std::make_pair(static_cast<uint8_t>(p), b));
+    if (kPrintStats) {
+      double d = p / 256.;
+      entropy += -d * log2(d) - (1 - d) * log2(1 - d);
+    }
+  }
+  if (kPrintStats) printf("entropy %f\n", entropy);
+  return ret;
+}
+
+bool check_rabs(const PvVec &pv_vec, uint8_t *buf) {
+  BufAnsCoder a;
+  aom_buf_ans_alloc(&a, NULL, kBufAnsSize);
+  buf_ans_write_init(&a, buf);
+
+  std::clock_t start = std::clock();
+  for (PvVec::const_iterator it = pv_vec.begin(); it != pv_vec.end(); ++it) {
+    buf_rabs_write(&a, it->second, 256 - it->first);
+  }
+  aom_buf_ans_flush(&a);
+  std::clock_t enc_time = std::clock() - start;
+  int offset = buf_ans_write_end(&a);
+  aom_buf_ans_free(&a);
+  bool okay = true;
+  AnsDecoder d;
+#if ANS_MAX_SYMBOLS
+  d.window_size = kBufAnsSize;
+#endif
+  if (ans_read_init(&d, buf, offset)) return false;
+  start = std::clock();
+  for (PvVec::const_iterator it = pv_vec.begin(); it != pv_vec.end(); ++it) {
+    okay = okay && (rabs_read(&d, 256 - it->first) != 0) == it->second;
+  }
+  std::clock_t dec_time = std::clock() - start;
+  if (!okay) return false;
+  if (kPrintStats)
+    printf("uABS size %d enc_time %f dec_time %f\n", offset,
+           static_cast<float>(enc_time) / CLOCKS_PER_SEC,
+           static_cast<float>(dec_time) / CLOCKS_PER_SEC);
+  return ans_read_end(&d) != 0;
+}
+
+const aom_cdf_prob spareto65[] = { 8320, 6018, 4402, 3254, 4259,
+                                   3919, 2057, 492,  45,   2 };
+
+const int kRansSymbols =
+    static_cast<int>(sizeof(spareto65) / sizeof(spareto65[0]));
+
+struct rans_sym {
+  aom_cdf_prob prob;
+  aom_cdf_prob cum_prob;  // not-inclusive
+};
+
+std::vector<int> ans_encode_build_vals(rans_sym *const tab, int iters) {
+  aom_cdf_prob sum = 0;
+  for (int i = 0; i < kRansSymbols; ++i) {
+    tab[i].cum_prob = sum;
+    tab[i].prob = spareto65[i];
+    sum += spareto65[i];
+  }
+  std::vector<int> p_to_sym;
+  for (int i = 0; i < kRansSymbols; ++i) {
+    p_to_sym.insert(p_to_sym.end(), tab[i].prob, i);
+  }
+  assert(p_to_sym.size() == RANS_PRECISION);
+  std::vector<int> ret;
+  libaom_test::ACMRandom gen(18543637);
+  for (int i = 0; i < iters; ++i) {
+    int sym =
+        p_to_sym[((gen.Rand8() << 8) + gen.Rand8()) & (RANS_PRECISION - 1)];
+    ret.push_back(sym);
+  }
+  return ret;
+}
+
+void rans_build_dec_tab(const struct rans_sym sym_tab[],
+                        aom_cdf_prob *dec_tab) {
+  unsigned int sum = 0;
+  for (int i = 0; sum < RANS_PRECISION; ++i) {
+    dec_tab[i] = sum += sym_tab[i].prob;
+  }
+}
+
+bool check_rans(const std::vector<int> &sym_vec, const rans_sym *const tab,
+                uint8_t *buf) {
+  BufAnsCoder a;
+  aom_buf_ans_alloc(&a, NULL, kBufAnsSize);
+  buf_ans_write_init(&a, buf);
+  aom_cdf_prob dec_tab[kRansSymbols];
+  rans_build_dec_tab(tab, dec_tab);
+
+  std::clock_t start = std::clock();
+  for (std::vector<int>::const_iterator it = sym_vec.begin();
+       it != sym_vec.end(); ++it) {
+    buf_rans_write(&a, tab[*it].cum_prob, tab[*it].prob);
+  }
+  aom_buf_ans_flush(&a);
+  std::clock_t enc_time = std::clock() - start;
+  int offset = buf_ans_write_end(&a);
+  aom_buf_ans_free(&a);
+  bool okay = true;
+  AnsDecoder d;
+#if ANS_MAX_SYMBOLS
+  d.window_size = kBufAnsSize;
+#endif
+  if (ans_read_init(&d, buf, offset)) return false;
+  start = std::clock();
+  for (std::vector<int>::const_iterator it = sym_vec.begin();
+       it != sym_vec.end(); ++it) {
+    okay &= rans_read(&d, dec_tab) == *it;
+  }
+  std::clock_t dec_time = std::clock() - start;
+  if (!okay) return false;
+  if (kPrintStats)
+    printf("rANS size %d enc_time %f dec_time %f\n", offset,
+           static_cast<float>(enc_time) / CLOCKS_PER_SEC,
+           static_cast<float>(dec_time) / CLOCKS_PER_SEC);
+  return ans_read_end(&d) != 0;
+}
+
+class AbsTestFix : public ::testing::Test {
+ protected:
+  static void SetUpTestCase() { pv_vec_ = abs_encode_build_vals(kNumBools); }
+  virtual void SetUp() { buf_ = new uint8_t[kNumBools / 8]; }
+  virtual void TearDown() { delete[] buf_; }
+  static const int kNumBools = 100000000;
+  static PvVec pv_vec_;
+  uint8_t *buf_;
+};
+PvVec AbsTestFix::pv_vec_;
+
+class AnsTestFix : public ::testing::Test {
+ protected:
+  static void SetUpTestCase() {
+    sym_vec_ = ans_encode_build_vals(rans_sym_tab_, kNumSyms);
+  }
+  virtual void SetUp() { buf_ = new uint8_t[kNumSyms / 2]; }
+  virtual void TearDown() { delete[] buf_; }
+  static const int kNumSyms = 25000000;
+  static std::vector<int> sym_vec_;
+  static rans_sym rans_sym_tab_[kRansSymbols];
+  uint8_t *buf_;
+};
+std::vector<int> AnsTestFix::sym_vec_;
+rans_sym AnsTestFix::rans_sym_tab_[kRansSymbols];
+
+TEST_F(AbsTestFix, Rabs) { EXPECT_TRUE(check_rabs(pv_vec_, buf_)); }
+TEST_F(AnsTestFix, Rans) {
+  EXPECT_TRUE(check_rans(sym_vec_, rans_sym_tab_, buf_));
+}
+TEST(AnsTest, FinalStateSerialization) {
+  for (unsigned i = L_BASE; i < L_BASE * IO_BASE; ++i) {
+    uint8_t buf[8];
+    AnsCoder c;
+    ans_write_init(&c, buf);
+    c.state = i;
+    const int written_size = ans_write_end(&c);
+    ASSERT_LT(static_cast<size_t>(written_size), sizeof(buf));
+    AnsDecoder d;
+#if ANS_MAX_SYMBOLS
+    // There is no real data window here because no symbols are sent through
+    // ans (only synthetic states), so use a dummy value
+    d.window_size = 1024;
+#endif
+    const int read_init_status = ans_read_init(&d, buf, written_size);
+    EXPECT_EQ(read_init_status, 0);
+    EXPECT_EQ(d.state, i);
+  }
+}
+}  // namespace
diff --git a/third_party/aom/test/aomcx_set_ref.sh b/third_party/aom/test/aomcx_set_ref.sh
new file mode 100755
index 0000000000..f51b73c58e
--- /dev/null
+++ b/third_party/aom/test/aomcx_set_ref.sh
@@ -0,0 +1,58 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This file tests the libaom aom_cx_set_ref example. To add new tests to this
+## file, do the following:
+##   1. Write a shell function (this is your test).
+##   2. Add the function to aom_cx_set_ref_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+aom_cx_set_ref_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libaom test data must exist in LIBAOM_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs aom_cx_set_ref and updates the reference frame before encoding frame 90.
+# $1 is the codec name, which aom_cx_set_ref does not support at present: It's
+# currently used only to name the output file.
+# TODO(tomfinegan): Pass the codec param once the example is updated to support
+# AV1.
+aom_set_ref() {
+  local encoder="${LIBAOM_BIN_PATH}/aom_cx_set_ref${AOM_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file="${AOM_TEST_OUTPUT_DIR}/aom_cx_set_ref_${codec}.ivf"
+  local ref_frame_num=4
+  local limit=10
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${AOM_TEST_PREFIX}" "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" \
+      "${ref_frame_num}" "${limit}" ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+aom_cx_set_ref_av1() {
+  if [ "$(av1_encode_available)" = "yes" ]; then
+    aom_set_ref av1 || return 1
+  fi
+}
+
+aom_cx_set_ref_tests="aom_cx_set_ref_av1"
+
+run_tests aom_cx_set_ref_verify_environment "${aom_cx_set_ref_tests}"
+
diff --git a/third_party/aom/test/aomdec.sh b/third_party/aom/test/aomdec.sh
new file mode 100755
index 0000000000..28901ed1bd
--- /dev/null
+++ b/third_party/aom/test/aomdec.sh
@@ -0,0 +1,124 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This file tests aomdec. To add new tests to this file, do the following:
+##   1. Write a shell function (this is your test).
+##   2. Add the function to aomdec_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available.
+aomdec_verify_environment() {
+  if [ "$(av1_encode_available)" != "yes" ] ; then
+    if [ ! -e "${AV1_WEBM_FILE}" ] || \
+      [ ! -e "${AV1_FPM_WEBM_FILE}" ] || \
+      [ ! -e "${AV1_LT_50_FRAMES_WEBM_FILE}" ] ; then
+      elog "Libaom test data must exist in LIBAOM_TEST_DATA_PATH."
+      return 1
+    fi
+  fi
+  if [ -z "$(aom_tool_path aomdec)" ]; then
+    elog "aomdec not found. It must exist in LIBAOM_BIN_PATH or its parent."
+    return 1
+  fi
+}
+
+# Wrapper function for running aomdec with pipe input. Requires that
+# LIBAOM_BIN_PATH points to the directory containing aomdec. $1 is used as the
+# input file path and shifted away. All remaining parameters are passed through
+# to aomdec.
+aomdec_pipe() {
+  local readonly input="$1"
+  shift
+  if [ ! -e "${input}" ]; then
+    local file="${AOM_TEST_OUTPUT_DIR}/test_encode.ivf"
+    encode_yuv_raw_input_av1 "${file}" --ivf
+  else
+    local file="${input}"
+  fi
+  cat "${file}" | aomdec - "$@" ${devnull}
+}
+
+
+# Wrapper function for running aomdec. Requires that LIBAOM_BIN_PATH points to
+# the directory containing aomdec. $1 one is used as the input file path and
+# shifted away. All remaining parameters are passed through to aomdec.
+aomdec() {
+  local readonly decoder="$(aom_tool_path aomdec)"
+  local readonly input="$1"
+  shift
+  eval "${AOM_TEST_PREFIX}" "${decoder}" "$input" "$@" ${devnull}
+}
+
+aomdec_can_decode_av1() {
+  if [ "$(av1_decode_available)" = "yes" ]; then
+    echo yes
+  fi
+}
+
+aomdec_aom_ivf_pipe_input() {
+  if [ "$(aomdec_can_decode_av1)" = "yes" ]; then
+    aomdec_pipe "${AOM_IVF_FILE}" --summary --noblit
+  fi
+}
+
+aomdec_av1_webm() {
+  if [ "$(aomdec_can_decode_av1)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    if [ ! -e "${AV1_WEBM_FILE}" ]; then
+      local file="${AOM_TEST_OUTPUT_DIR}/test_encode.webm"
+      encode_yuv_raw_input_av1 "${file}"
+    else
+      aomdec "${AV1_WEBM_FILE}" --summary --noblit
+    fi
+  fi
+}
+
+aomdec_av1_webm_frame_parallel() {
+  if [ "$(aomdec_can_decode_av1)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local file
+    if [ ! -e "${AV1_WEBM_FILE}" ]; then
+      file="${AOM_TEST_OUTPUT_DIR}/test_encode.webm"
+      encode_yuv_raw_input_av1 "${file}" "--ivf --error-resilient=1 "
+    else
+      file="${AV1_FPM_WEBM_FILE}"
+    fi
+    for threads in 2 3 4 5 6 7 8; do
+      aomdec "${file}" --summary --noblit --threads=$threads \
+        --frame-parallel
+    done
+  fi
+}
+
+# TODO(vigneshv): Enable or remove this test and associated code.
+DISABLED_aomdec_av1_webm_less_than_50_frames() {
+  # ensure that reaching eof in webm_guess_framerate doesn't result in invalid
+  # frames in actual webm_read_frame calls.
+  if [ "$(aomdec_can_decode_av1)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly decoder="$(aom_tool_path aomdec)"
+    local readonly expected=10
+    local readonly num_frames=$(${AOM_TEST_PREFIX} "${decoder}" \
+      "${AV1_LT_50_FRAMES_WEBM_FILE}" --summary --noblit 2>&1 \
+      | awk '/^[0-9]+ decoded frames/ { print $1 }')
+    if [ "$num_frames" -ne "$expected" ]; then
+      elog "Output frames ($num_frames) != expected ($expected)"
+      return 1
+    fi
+  fi
+}
+
+aomdec_tests="aomdec_av1_webm
+              aomdec_av1_webm_frame_parallel
+              aomdec_aom_ivf_pipe_input
+              DISABLED_aomdec_av1_webm_less_than_50_frames"
+
+run_tests aomdec_verify_environment "${aomdec_tests}"
diff --git a/third_party/aom/test/aomenc.sh b/third_party/aom/test/aomenc.sh
new file mode 100755
index 0000000000..57a4c28a5a
--- /dev/null
+++ b/third_party/aom/test/aomenc.sh
@@ -0,0 +1,241 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This file tests aomenc using hantro_collage_w352h288.yuv as input. To add
+## new tests to this file, do the following:
+##   1. Write a shell function (this is your test).
+##   2. Add the function to aomenc_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+readonly TEST_FRAMES=5
+
+# Environment check: Make sure input is available.
+aomenc_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    elog "The file ${YUV_RAW_INPUT##*/} must exist in LIBAOM_TEST_DATA_PATH."
+    return 1
+  fi
+  if [ "$(aomenc_can_encode_av1)" = "yes" ]; then
+    if [ ! -e "${Y4M_NOSQ_PAR_INPUT}" ]; then
+      elog "The file ${Y4M_NOSQ_PAR_INPUT##*/} must exist in"
+      elog "LIBAOM_TEST_DATA_PATH."
+      return 1
+    fi
+  fi
+  if [ -z "$(aom_tool_path aomenc)" ]; then
+    elog "aomenc not found. It must exist in LIBAOM_BIN_PATH or its parent."
+    return 1
+  fi
+}
+
+aomenc_can_encode_av1() {
+  if [ "$(av1_encode_available)" = "yes" ]; then
+    echo yes
+  fi
+}
+
+aomenc_can_encode_av1() {
+  if [ "$(av1_encode_available)" = "yes" ]; then
+    echo yes
+  fi
+}
+
+# Utilities that echo aomenc input file parameters.
+y4m_input_non_square_par() {
+  echo ""${Y4M_NOSQ_PAR_INPUT}""
+}
+
+y4m_input_720p() {
+  echo ""${Y4M_720P_INPUT}""
+}
+
+# Echo default aomenc real time encoding params. $1 is the codec, which defaults
+# to av1 if unspecified.
+aomenc_rt_params() {
+  local readonly codec="${1:-av1}"
+  echo "--codec=${codec}
+    --buf-initial-sz=500
+    --buf-optimal-sz=600
+    --buf-sz=1000
+    --cpu-used=-6
+    --end-usage=cbr
+    --error-resilient=1
+    --kf-max-dist=90000
+    --lag-in-frames=0
+    --max-intra-rate=300
+    --max-q=56
+    --min-q=2
+    --noise-sensitivity=0
+    --overshoot-pct=50
+    --passes=1
+    --profile=0
+    --resize-allowed=0
+    --rt
+    --static-thresh=0
+    --undershoot-pct=50"
+}
+
+# Wrapper function for running aomenc with pipe input. Requires that
+# LIBAOM_BIN_PATH points to the directory containing aomenc. $1 is used as the
+# input file path and shifted away. All remaining parameters are passed through
+# to aomenc.
+aomenc_pipe() {
+  local readonly encoder="$(aom_tool_path aomenc)"
+  local readonly input="$1"
+  shift
+  cat "${input}" | eval "${AOM_TEST_PREFIX}" "${encoder}" - \
+    --test-decode=fatal \
+    "$@" ${devnull}
+}
+
+# Wrapper function for running aomenc. Requires that LIBAOM_BIN_PATH points to
+# the directory containing aomenc. $1 one is used as the input file path and
+# shifted away. All remaining parameters are passed through to aomenc.
+aomenc() {
+  local readonly encoder="$(aom_tool_path aomenc)"
+  local readonly input="$1"
+  shift
+  eval "${AOM_TEST_PREFIX}" "${encoder}" "${input}" \
+    --test-decode=fatal \
+    "$@" ${devnull}
+}
+
+aomenc_av1_ivf() {
+  if [ "$(aomenc_can_encode_av1)" = "yes" ]; then
+    local readonly output="${AOM_TEST_OUTPUT_DIR}/av1.ivf"
+    aomenc $(yuv_raw_input) \
+      --codec=av1 \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+aomenc_av1_webm() {
+  if [ "$(aomenc_can_encode_av1)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${AOM_TEST_OUTPUT_DIR}/av1.webm"
+    aomenc $(yuv_raw_input) \
+      --codec=av1 \
+      --limit="${TEST_FRAMES}" \
+      --output="${output}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+aomenc_av1_webm_2pass() {
+  if [ "$(aomenc_can_encode_av1)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${AOM_TEST_OUTPUT_DIR}/av1.webm"
+    aomenc $(yuv_raw_input) \
+      --codec=av1 \
+      --limit="${TEST_FRAMES}" \
+      --output="${output}" \
+      --passes=2
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+aomenc_av1_ivf_lossless() {
+  if [ "$(aomenc_can_encode_av1)" = "yes" ]; then
+    local readonly output="${AOM_TEST_OUTPUT_DIR}/av1_lossless.ivf"
+    aomenc $(yuv_raw_input) \
+      --codec=av1 \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}" \
+      --lossless=1
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+aomenc_av1_ivf_minq0_maxq0() {
+  if [ "$(aomenc_can_encode_av1)" = "yes" ]; then
+    local readonly output="${AOM_TEST_OUTPUT_DIR}/av1_lossless_minq0_maxq0.ivf"
+    aomenc $(yuv_raw_input) \
+      --codec=av1 \
+      --limit="${TEST_FRAMES}" \
+      --ivf \
+      --output="${output}" \
+      --min-q=0 \
+      --max-q=0
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+aomenc_av1_webm_lag5_frames10() {
+  if [ "$(aomenc_can_encode_av1)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly lag_total_frames=10
+    local readonly lag_frames=5
+    local readonly output="${AOM_TEST_OUTPUT_DIR}/av1_lag5_frames10.webm"
+    aomenc $(yuv_raw_input) \
+      --codec=av1 \
+      --limit="${lag_total_frames}" \
+      --lag-in-frames="${lag_frames}" \
+      --output="${output}" \
+      --passes=2 \
+      --auto-alt-ref=1
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+# TODO(fgalligan): Test that DisplayWidth is different than video width.
+aomenc_av1_webm_non_square_par() {
+  if [ "$(aomenc_can_encode_av1)" = "yes" ] && \
+     [ "$(webm_io_available)" = "yes" ]; then
+    local readonly output="${AOM_TEST_OUTPUT_DIR}/av1_non_square_par.webm"
+    aomenc $(y4m_input_non_square_par) \
+      --codec=av1 \
+      --limit="${TEST_FRAMES}" \
+      --output="${output}"
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+aomenc_tests="aomenc_av1_ivf
+              aomenc_av1_webm
+              aomenc_av1_webm_2pass
+              aomenc_av1_ivf_lossless
+              aomenc_av1_ivf_minq0_maxq0
+              aomenc_av1_webm_lag5_frames10
+              aomenc_av1_webm_non_square_par"
+
+run_tests aomenc_verify_environment "${aomenc_tests}"
diff --git a/third_party/aom/test/aq_segment_test.cc b/third_party/aom/test/aq_segment_test.cc
new file mode 100644
index 0000000000..5dc93ec793
--- /dev/null
+++ b/third_party/aom/test/aq_segment_test.cc
@@ -0,0 +1,129 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "./aom_config.h"
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+class AqSegmentTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<libaom_test::TestMode, int> {
+ protected:
+  AqSegmentTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~AqSegmentTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    set_cpu_used_ = GET_PARAM(2);
+    aq_mode_ = 0;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AOME_SET_CPUUSED, set_cpu_used_);
+      encoder->Control(AV1E_SET_AQ_MODE, aq_mode_);
+#if CONFIG_EXT_DELTA_Q
+      encoder->Control(AV1E_SET_DELTAQ_MODE, deltaq_mode_);
+#endif
+      encoder->Control(AOME_SET_MAX_INTRA_BITRATE_PCT, 100);
+    }
+  }
+
+  void DoTest(int aq_mode) {
+    aq_mode_ = aq_mode;
+#if CONFIG_EXT_DELTA_Q
+    deltaq_mode_ = 0;
+#endif
+    cfg_.kf_max_dist = 12;
+    cfg_.rc_min_quantizer = 8;
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_end_usage = AOM_CBR;
+    cfg_.g_lag_in_frames = 6;
+    cfg_.rc_buf_initial_sz = 500;
+    cfg_.rc_buf_optimal_sz = 500;
+    cfg_.rc_buf_sz = 1000;
+    cfg_.rc_target_bitrate = 300;
+    ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352,
+                                         288, 30, 1, 0, 15);
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  }
+
+  int set_cpu_used_;
+  int aq_mode_;
+#if CONFIG_EXT_DELTA_Q
+  int deltaq_mode_;
+#endif
+};
+
+// Validate that this AQ segmentation mode (AQ=1, variance_ap)
+// encodes and decodes without a mismatch.
+TEST_P(AqSegmentTest, TestNoMisMatchAQ1) { DoTest(1); }
+
+// Validate that this AQ segmentation mode (AQ=2, complexity_aq)
+// encodes and decodes without a mismatch.
+TEST_P(AqSegmentTest, TestNoMisMatchAQ2) { DoTest(2); }
+
+// Validate that this AQ segmentation mode (AQ=3, cyclic_refresh_aq)
+// encodes and decodes without a mismatch.
+TEST_P(AqSegmentTest, TestNoMisMatchAQ3) { DoTest(3); }
+
+class AqSegmentTestLarge : public AqSegmentTest {};
+
+TEST_P(AqSegmentTestLarge, TestNoMisMatchAQ1) { DoTest(1); }
+
+TEST_P(AqSegmentTestLarge, TestNoMisMatchAQ2) { DoTest(2); }
+
+TEST_P(AqSegmentTestLarge, TestNoMisMatchAQ3) { DoTest(3); }
+
+#if CONFIG_DELTA_Q & !CONFIG_EXT_DELTA_Q
+// Validate that this AQ mode (AQ=4, delta q)
+// encodes and decodes without a mismatch.
+TEST_P(AqSegmentTest, TestNoMisMatchAQ4) {
+  cfg_.rc_end_usage = AOM_CQ;
+  aq_mode_ = 4;
+
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 100);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+#endif
+
+#if CONFIG_EXT_DELTA_Q
+// Validate that this delta q mode
+// encodes and decodes without a mismatch.
+TEST_P(AqSegmentTest, TestNoMisMatchExtDeltaQ) {
+  cfg_.rc_end_usage = AOM_CQ;
+  aq_mode_ = 0;
+  deltaq_mode_ = 2;
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 100);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+#endif
+
+AV1_INSTANTIATE_TEST_CASE(AqSegmentTest,
+                          ::testing::Values(::libaom_test::kRealTime,
+                                            ::libaom_test::kOnePassGood),
+                          ::testing::Range(5, 9));
+AV1_INSTANTIATE_TEST_CASE(AqSegmentTestLarge,
+                          ::testing::Values(::libaom_test::kRealTime,
+                                            ::libaom_test::kOnePassGood),
+                          ::testing::Range(3, 5));
+}  // namespace
diff --git a/third_party/aom/test/arf_freq_test.cc b/third_party/aom/test/arf_freq_test.cc
new file mode 100644
index 0000000000..bef58b3e85
--- /dev/null
+++ b/third_party/aom/test/arf_freq_test.cc
@@ -0,0 +1,232 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "test/yuv_video_source.h"
+#include "av1/encoder/ratectrl.h"
+
+namespace {
+
+const unsigned int kFrames = 100;
+const int kBitrate = 500;
+
+#define ARF_NOT_SEEN 1000001
+#define ARF_SEEN_ONCE 1000000
+
+typedef struct {
+  const char *filename;
+  unsigned int width;
+  unsigned int height;
+  unsigned int framerate_num;
+  unsigned int framerate_den;
+  unsigned int input_bit_depth;
+  aom_img_fmt fmt;
+  aom_bit_depth_t bit_depth;
+  unsigned int profile;
+} TestVideoParam;
+
+typedef struct {
+  libaom_test::TestMode mode;
+  int cpu_used;
+} TestEncodeParam;
+
+const TestVideoParam kTestVectors[] = {
+  // artificially increase framerate to trigger default check
+  { "hantro_collage_w352h288.yuv", 352, 288, 5000, 1, 8, AOM_IMG_FMT_I420,
+    AOM_BITS_8, 0 },
+  { "hantro_collage_w352h288.yuv", 352, 288, 30, 1, 8, AOM_IMG_FMT_I420,
+    AOM_BITS_8, 0 },
+  { "rush_hour_444.y4m", 352, 288, 30, 1, 8, AOM_IMG_FMT_I444, AOM_BITS_8, 1 },
+#if CONFIG_HIGHBITDEPTH
+// Add list of profile 2/3 test videos here ...
+#endif  // CONFIG_HIGHBITDEPTH
+};
+
+const TestEncodeParam kEncodeVectors[] = {
+  { ::libaom_test::kOnePassGood, 2 }, { ::libaom_test::kOnePassGood, 5 },
+  { ::libaom_test::kTwoPassGood, 1 }, { ::libaom_test::kTwoPassGood, 2 },
+  { ::libaom_test::kTwoPassGood, 5 }, { ::libaom_test::kRealTime, 5 },
+};
+
+const int kMinArfVectors[] = {
+  // NOTE: 0 refers to the default built-in logic in:
+  //       av1_rc_get_default_min_gf_interval(...)
+  0, 4, 8, 12, 15
+};
+
+int is_extension_y4m(const char *filename) {
+  const char *dot = strrchr(filename, '.');
+  if (!dot || dot == filename)
+    return 0;
+  else
+    return !strcmp(dot, ".y4m");
+}
+
+class ArfFreqTestLarge
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith3Params<TestVideoParam,
+                                                 TestEncodeParam, int> {
+ protected:
+  ArfFreqTestLarge()
+      : EncoderTest(GET_PARAM(0)), test_video_param_(GET_PARAM(1)),
+        test_encode_param_(GET_PARAM(2)), min_arf_requested_(GET_PARAM(3)) {}
+
+  virtual ~ArfFreqTestLarge() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(test_encode_param_.mode);
+    if (test_encode_param_.mode != ::libaom_test::kRealTime) {
+      cfg_.g_lag_in_frames = 25;
+      cfg_.rc_end_usage = AOM_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = AOM_CBR;
+      cfg_.rc_buf_sz = 1000;
+      cfg_.rc_buf_initial_sz = 500;
+      cfg_.rc_buf_optimal_sz = 600;
+    }
+    dec_cfg_.threads = 4;
+  }
+
+  virtual void BeginPassHook(unsigned int) {
+    min_run_ = ARF_NOT_SEEN;
+    run_of_visible_frames_ = 0;
+  }
+
+  int GetNumFramesInPkt(const aom_codec_cx_pkt_t *pkt) {
+    const uint8_t *buffer = reinterpret_cast<uint8_t *>(pkt->data.frame.buf);
+    const uint8_t marker = buffer[pkt->data.frame.sz - 1];
+    const int mag = ((marker >> 3) & 3) + 1;
+    int frames = (marker & 0x7) + 1;
+    const unsigned int index_sz = 2 + mag * frames;
+    // Check for superframe or not.
+    // Assume superframe has only one visible frame, the rest being
+    // invisible. If superframe index is not found, then there is only
+    // one frame.
+    if (!((marker & 0xe0) == 0xc0 && pkt->data.frame.sz >= index_sz &&
+          buffer[pkt->data.frame.sz - index_sz] == marker)) {
+      frames = 1;
+    }
+    return frames;
+  }
+
+  virtual void FramePktHook(const aom_codec_cx_pkt_t *pkt) {
+    if (pkt->kind != AOM_CODEC_CX_FRAME_PKT) return;
+    const int frames = GetNumFramesInPkt(pkt);
+    if (frames == 1) {
+      run_of_visible_frames_++;
+    } else if (frames == 2) {
+      if (min_run_ == ARF_NOT_SEEN) {
+        min_run_ = ARF_SEEN_ONCE;
+      } else if (min_run_ == ARF_SEEN_ONCE ||
+                 run_of_visible_frames_ < min_run_) {
+        min_run_ = run_of_visible_frames_;
+      }
+      run_of_visible_frames_ = 1;
+    } else {
+      min_run_ = 0;
+      run_of_visible_frames_ = 1;
+    }
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      encoder->Control(AV1E_SET_FRAME_PARALLEL_DECODING, 1);
+      encoder->Control(AV1E_SET_TILE_COLUMNS, 4);
+      encoder->Control(AOME_SET_CPUUSED, test_encode_param_.cpu_used);
+      encoder->Control(AV1E_SET_MIN_GF_INTERVAL, min_arf_requested_);
+      if (test_encode_param_.mode != ::libaom_test::kRealTime) {
+        encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(AOME_SET_ARNR_STRENGTH, 5);
+      }
+    }
+  }
+
+  int GetMinVisibleRun() const { return min_run_; }
+
+  int GetMinArfDistanceRequested() const {
+    if (min_arf_requested_)
+      return min_arf_requested_;
+    else
+      return av1_rc_get_default_min_gf_interval(
+          test_video_param_.width, test_video_param_.height,
+          (double)test_video_param_.framerate_num /
+              test_video_param_.framerate_den);
+  }
+
+  TestVideoParam test_video_param_;
+  TestEncodeParam test_encode_param_;
+
+ private:
+  int min_arf_requested_;
+  int min_run_;
+  int run_of_visible_frames_;
+};
+
+TEST_P(ArfFreqTestLarge, MinArfFreqTest) {
+  cfg_.rc_target_bitrate = kBitrate;
+  cfg_.g_error_resilient = 0;
+  cfg_.g_profile = test_video_param_.profile;
+  cfg_.g_input_bit_depth = test_video_param_.input_bit_depth;
+  cfg_.g_bit_depth = test_video_param_.bit_depth;
+  init_flags_ = AOM_CODEC_USE_PSNR;
+  if (cfg_.g_bit_depth > 8) init_flags_ |= AOM_CODEC_USE_HIGHBITDEPTH;
+
+  testing::internal::scoped_ptr<libaom_test::VideoSource> video;
+  if (is_extension_y4m(test_video_param_.filename)) {
+    video.reset(new libaom_test::Y4mVideoSource(test_video_param_.filename, 0,
+                                                kFrames));
+  } else {
+    video.reset(new libaom_test::YUVVideoSource(
+        test_video_param_.filename, test_video_param_.fmt,
+        test_video_param_.width, test_video_param_.height,
+        test_video_param_.framerate_num, test_video_param_.framerate_den, 0,
+        kFrames));
+  }
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video.get()));
+  const int min_run = GetMinVisibleRun();
+  const int min_arf_dist_requested = GetMinArfDistanceRequested();
+  if (min_run != ARF_NOT_SEEN && min_run != ARF_SEEN_ONCE) {
+    const int min_arf_dist = min_run + 1;
+    EXPECT_GE(min_arf_dist, min_arf_dist_requested);
+  }
+}
+
+#if CONFIG_HIGHBITDEPTH || CONFIG_EXT_REFS
+#if CONFIG_AV1_ENCODER
+// TODO(angiebird): 25-29 fail in high bitdepth mode.
+// TODO(zoeliu): This ArfFreqTest does not work with BWDREF_FRAME, as
+// BWDREF_FRAME is also a non-show frame, and the minimum run between two
+// consecutive BWDREF_FRAME's may vary between 1 and any arbitrary positive
+// number as long as it does not exceed the gf_group interval.
+INSTANTIATE_TEST_CASE_P(
+    DISABLED_AV1, ArfFreqTestLarge,
+    ::testing::Combine(
+        ::testing::Values(
+            static_cast<const libaom_test::CodecFactory *>(&libaom_test::kAV1)),
+        ::testing::ValuesIn(kTestVectors), ::testing::ValuesIn(kEncodeVectors),
+        ::testing::ValuesIn(kMinArfVectors)));
+#endif  // CONFIG_AV1_ENCODER
+#else
+AV1_INSTANTIATE_TEST_CASE(ArfFreqTestLarge, ::testing::ValuesIn(kTestVectors),
+                          ::testing::ValuesIn(kEncodeVectors),
+                          ::testing::ValuesIn(kMinArfVectors));
+#endif  // CONFIG_HIGHBITDEPTH || CONFIG_EXT_REFS
+}  // namespace
diff --git a/third_party/aom/test/av1_convolve_optimz_test.cc b/third_party/aom/test/av1_convolve_optimz_test.cc
new file mode 100644
index 0000000000..fd0f6dbce2
--- /dev/null
+++ b/third_party/aom/test/av1_convolve_optimz_test.cc
@@ -0,0 +1,405 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+namespace {
+
+using std::tr1::tuple;
+using libaom_test::ACMRandom;
+
+typedef void (*ConvInit)();
+typedef void (*conv_filter_t)(const uint8_t *, int, uint8_t *, int, int, int,
+                              const InterpFilterParams, int, int,
+                              ConvolveParams *);
+#if CONFIG_HIGHBITDEPTH
+typedef void (*hbd_conv_filter_t)(const uint16_t *, int, uint16_t *, int, int,
+                                  int, const InterpFilterParams, int, int, int,
+                                  int);
+#endif
+
+// Test parameter list:
+//  <convolve_horiz_func, convolve_vert_func,
+//  <width, height>, filter_params, subpel_x_q4, avg>
+typedef tuple<int, int> BlockDimension;
+typedef tuple<ConvInit, conv_filter_t, conv_filter_t, BlockDimension,
+              InterpFilter, int, int>
+    ConvParams;
+#if CONFIG_HIGHBITDEPTH
+// Test parameter list:
+//  <convolve_horiz_func, convolve_vert_func,
+//  <width, height>, filter_params, subpel_x_q4, avg, bit_dpeth>
+typedef tuple<ConvInit, hbd_conv_filter_t, hbd_conv_filter_t, BlockDimension,
+              InterpFilter, int, int, int>
+    HbdConvParams;
+#endif
+
+// Note:
+//  src_ and src_ref_ have special boundary requirement
+//  dst_ and dst_ref_ don't
+const size_t maxWidth = 256;
+const size_t maxHeight = 256;
+const size_t maxBlockSize = maxWidth * maxHeight;
+const int horizOffset = 32;
+const int vertiOffset = 32;
+const int stride = 128;
+const int x_step_q4 = 16;
+
+class AV1ConvolveOptimzTest : public ::testing::TestWithParam<ConvParams> {
+ public:
+  virtual ~AV1ConvolveOptimzTest() {}
+  virtual void SetUp() {
+    ConvInit conv_init = GET_PARAM(0);
+    conv_init();
+    conv_horiz_ = GET_PARAM(1);
+    conv_vert_ = GET_PARAM(2);
+    BlockDimension block = GET_PARAM(3);
+    width_ = std::tr1::get<0>(block);
+    height_ = std::tr1::get<1>(block);
+    filter_ = GET_PARAM(4);
+    subpel_ = GET_PARAM(5);
+    int ref = GET_PARAM(6);
+    const int plane = 0;
+    conv_params_ = get_conv_params(ref, plane);
+
+    alloc_ = new uint8_t[maxBlockSize * 4];
+    src_ = alloc_ + (vertiOffset * maxWidth);
+    src_ += horizOffset;
+    src_ref_ = src_ + maxBlockSize;
+
+    dst_ = alloc_ + 2 * maxBlockSize;
+    dst_ref_ = alloc_ + 3 * maxBlockSize;
+  }
+
+  virtual void TearDown() {
+    delete[] alloc_;
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void RunHorizFilterBitExactCheck();
+  void RunVertFilterBitExactCheck();
+
+ private:
+  void PrepFilterBuffer();
+  void DiffFilterBuffer();
+  conv_filter_t conv_horiz_;
+  conv_filter_t conv_vert_;
+  uint8_t *alloc_;
+  uint8_t *src_;
+  uint8_t *dst_;
+  uint8_t *src_ref_;
+  uint8_t *dst_ref_;
+  int width_;
+  int height_;
+  InterpFilter filter_;
+  int subpel_;
+  ConvolveParams conv_params_;
+};
+
+void AV1ConvolveOptimzTest::PrepFilterBuffer() {
+  int r, c;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+  memset(alloc_, 0, 4 * maxBlockSize * sizeof(alloc_[0]));
+
+  uint8_t *src_ptr = src_;
+  uint8_t *dst_ptr = dst_;
+  uint8_t *src_ref_ptr = src_ref_;
+  uint8_t *dst_ref_ptr = dst_ref_;
+
+  for (r = 0; r < height_; ++r) {
+    for (c = 0; c < width_; ++c) {
+      src_ptr[c] = rnd.Rand8();
+      src_ref_ptr[c] = src_ptr[c];
+      dst_ptr[c] = rnd.Rand8();
+      dst_ref_ptr[c] = dst_ptr[c];
+    }
+    src_ptr += stride;
+    src_ref_ptr += stride;
+    dst_ptr += stride;
+    dst_ref_ptr += stride;
+  }
+}
+
+void AV1ConvolveOptimzTest::DiffFilterBuffer() {
+  int r, c;
+  const uint8_t *dst_ptr = dst_;
+  const uint8_t *dst_ref_ptr = dst_ref_;
+  for (r = 0; r < height_; ++r) {
+    for (c = 0; c < width_; ++c) {
+      EXPECT_EQ((uint8_t)dst_ref_ptr[c], (uint8_t)dst_ptr[c])
+          << "Error at row: " << r << " col: " << c << " "
+          << "w = " << width_ << " "
+          << "h = " << height_ << " "
+          << "filter group index = " << filter_ << " "
+          << "filter index = " << subpel_;
+    }
+    dst_ptr += stride;
+    dst_ref_ptr += stride;
+  }
+}
+
+void AV1ConvolveOptimzTest::RunHorizFilterBitExactCheck() {
+  PrepFilterBuffer();
+
+  InterpFilterParams filter_params = av1_get_interp_filter_params(filter_);
+
+  av1_convolve_horiz_c(src_ref_, stride, dst_ref_, stride, width_, height_,
+                       filter_params, subpel_, x_step_q4, &conv_params_);
+
+  conv_horiz_(src_, stride, dst_, stride, width_, height_, filter_params,
+              subpel_, x_step_q4, &conv_params_);
+
+  DiffFilterBuffer();
+
+  // Note:
+  // Here we need calculate a height which is different from the specified one
+  // and test again.
+  int intermediate_height =
+      (((height_ - 1) * 16 + subpel_) >> SUBPEL_BITS) + filter_params.taps;
+  PrepFilterBuffer();
+
+  av1_convolve_horiz_c(src_ref_, stride, dst_ref_, stride, width_,
+                       intermediate_height, filter_params, subpel_, x_step_q4,
+                       &conv_params_);
+
+  conv_horiz_(src_, stride, dst_, stride, width_, intermediate_height,
+              filter_params, subpel_, x_step_q4, &conv_params_);
+
+  DiffFilterBuffer();
+}
+
+void AV1ConvolveOptimzTest::RunVertFilterBitExactCheck() {
+  PrepFilterBuffer();
+
+  InterpFilterParams filter_params = av1_get_interp_filter_params(filter_);
+
+  av1_convolve_vert_c(src_ref_, stride, dst_ref_, stride, width_, height_,
+                      filter_params, subpel_, x_step_q4, &conv_params_);
+
+  conv_vert_(src_, stride, dst_, stride, width_, height_, filter_params,
+             subpel_, x_step_q4, &conv_params_);
+
+  DiffFilterBuffer();
+}
+
+TEST_P(AV1ConvolveOptimzTest, HorizBitExactCheck) {
+  RunHorizFilterBitExactCheck();
+}
+TEST_P(AV1ConvolveOptimzTest, VerticalBitExactCheck) {
+  RunVertFilterBitExactCheck();
+}
+
+using std::tr1::make_tuple;
+
+#if (HAVE_SSSE3 || HAVE_SSE4_1) && CONFIG_DUAL_FILTER
+const BlockDimension kBlockDim[] = {
+  make_tuple(2, 2),    make_tuple(2, 4),    make_tuple(4, 4),
+  make_tuple(4, 8),    make_tuple(8, 4),    make_tuple(8, 8),
+  make_tuple(8, 16),   make_tuple(16, 8),   make_tuple(16, 16),
+  make_tuple(16, 32),  make_tuple(32, 16),  make_tuple(32, 32),
+  make_tuple(32, 64),  make_tuple(64, 32),  make_tuple(64, 64),
+  make_tuple(64, 128), make_tuple(128, 64), make_tuple(128, 128),
+};
+
+// 10/12-tap filters
+const InterpFilter kFilter[] = { FILTER_REGULAR_UV, BILINEAR, MULTITAP_SHARP };
+
+const int kSubpelQ4[] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
+
+const int kAvg[] = { 0, 1 };
+#endif
+
+#if HAVE_SSSE3 && CONFIG_DUAL_FILTER
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, AV1ConvolveOptimzTest,
+    ::testing::Combine(::testing::Values(av1_lowbd_convolve_init_ssse3),
+                       ::testing::Values(av1_convolve_horiz_ssse3),
+                       ::testing::Values(av1_convolve_vert_ssse3),
+                       ::testing::ValuesIn(kBlockDim),
+                       ::testing::ValuesIn(kFilter),
+                       ::testing::ValuesIn(kSubpelQ4),
+                       ::testing::ValuesIn(kAvg)));
+#endif  // HAVE_SSSE3 && CONFIG_DUAL_FILTER
+
+#if CONFIG_HIGHBITDEPTH
+typedef ::testing::TestWithParam<HbdConvParams> TestWithHbdConvParams;
+class AV1HbdConvolveOptimzTest : public TestWithHbdConvParams {
+ public:
+  virtual ~AV1HbdConvolveOptimzTest() {}
+  virtual void SetUp() {
+    ConvInit conv_init = GET_PARAM(0);
+    conv_init();
+    conv_horiz_ = GET_PARAM(1);
+    conv_vert_ = GET_PARAM(2);
+    BlockDimension block = GET_PARAM(3);
+    width_ = std::tr1::get<0>(block);
+    height_ = std::tr1::get<1>(block);
+    filter_ = GET_PARAM(4);
+    subpel_ = GET_PARAM(5);
+    avg_ = GET_PARAM(6);
+    bit_depth_ = GET_PARAM(7);
+
+    alloc_ = new uint16_t[maxBlockSize * 4];
+    src_ = alloc_ + (vertiOffset * maxWidth);
+    src_ += horizOffset;
+    src_ref_ = src_ + maxBlockSize;
+
+    dst_ = alloc_ + 2 * maxBlockSize;
+    dst_ref_ = alloc_ + 3 * maxBlockSize;
+  }
+
+  virtual void TearDown() {
+    delete[] alloc_;
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void RunHorizFilterBitExactCheck();
+  void RunVertFilterBitExactCheck();
+
+ private:
+  void PrepFilterBuffer();
+  void DiffFilterBuffer();
+  hbd_conv_filter_t conv_horiz_;
+  hbd_conv_filter_t conv_vert_;
+  uint16_t *alloc_;
+  uint16_t *src_;
+  uint16_t *dst_;
+  uint16_t *src_ref_;
+  uint16_t *dst_ref_;
+  int width_;
+  int height_;
+  InterpFilter filter_;
+  int subpel_;
+  int avg_;
+  int bit_depth_;
+};
+
+void AV1HbdConvolveOptimzTest::PrepFilterBuffer() {
+  int r, c;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+  memset(alloc_, 0, 4 * maxBlockSize * sizeof(alloc_[0]));
+
+  uint16_t *src_ptr = src_;
+  uint16_t *dst_ptr = dst_;
+  uint16_t *dst_ref_ptr = dst_ref_;
+  uint16_t hbd_mask = (1 << bit_depth_) - 1;
+
+  for (r = 0; r < height_; ++r) {
+    for (c = 0; c < width_; ++c) {
+      src_ptr[c] = rnd.Rand16() & hbd_mask;
+      dst_ptr[c] = rnd.Rand16() & hbd_mask;
+      dst_ref_ptr[c] = dst_ptr[c];
+    }
+    src_ptr += stride;
+    dst_ptr += stride;
+    dst_ref_ptr += stride;
+  }
+}
+
+void AV1HbdConvolveOptimzTest::DiffFilterBuffer() {
+  int r, c;
+  const uint16_t *dst_ptr = dst_;
+  const uint16_t *dst_ref_ptr = dst_ref_;
+  for (r = 0; r < height_; ++r) {
+    for (c = 0; c < width_; ++c) {
+      EXPECT_EQ((uint16_t)dst_ref_ptr[c], (uint16_t)dst_ptr[c])
+          << "Error at row: " << r << " col: " << c << " "
+          << "w = " << width_ << " "
+          << "h = " << height_ << " "
+          << "filter group index = " << filter_ << " "
+          << "filter index = " << subpel_ << " "
+          << "bit depth = " << bit_depth_;
+    }
+    dst_ptr += stride;
+    dst_ref_ptr += stride;
+  }
+}
+
+void AV1HbdConvolveOptimzTest::RunHorizFilterBitExactCheck() {
+  PrepFilterBuffer();
+
+  InterpFilterParams filter_params = av1_get_interp_filter_params(filter_);
+
+  av1_highbd_convolve_horiz_c(src_, stride, dst_ref_, stride, width_, height_,
+                              filter_params, subpel_, x_step_q4, avg_,
+                              bit_depth_);
+
+  conv_horiz_(src_, stride, dst_, stride, width_, height_, filter_params,
+              subpel_, x_step_q4, avg_, bit_depth_);
+
+  DiffFilterBuffer();
+
+  // Note:
+  // Here we need calculate a height which is different from the specified one
+  // and test again.
+  int intermediate_height =
+      (((height_ - 1) * 16 + subpel_) >> SUBPEL_BITS) + filter_params.taps;
+  PrepFilterBuffer();
+
+  av1_highbd_convolve_horiz_c(src_, stride, dst_ref_, stride, width_,
+                              intermediate_height, filter_params, subpel_,
+                              x_step_q4, avg_, bit_depth_);
+
+  conv_horiz_(src_, stride, dst_, stride, width_, intermediate_height,
+              filter_params, subpel_, x_step_q4, avg_, bit_depth_);
+
+  DiffFilterBuffer();
+}
+
+void AV1HbdConvolveOptimzTest::RunVertFilterBitExactCheck() {
+  PrepFilterBuffer();
+
+  InterpFilterParams filter_params = av1_get_interp_filter_params(filter_);
+
+  av1_highbd_convolve_vert_c(src_, stride, dst_ref_, stride, width_, height_,
+                             filter_params, subpel_, x_step_q4, avg_,
+                             bit_depth_);
+
+  conv_vert_(src_, stride, dst_, stride, width_, height_, filter_params,
+             subpel_, x_step_q4, avg_, bit_depth_);
+
+  DiffFilterBuffer();
+}
+
+TEST_P(AV1HbdConvolveOptimzTest, HorizBitExactCheck) {
+  RunHorizFilterBitExactCheck();
+}
+TEST_P(AV1HbdConvolveOptimzTest, VertBitExactCheck) {
+  RunVertFilterBitExactCheck();
+}
+
+#if HAVE_SSE4_1 && CONFIG_DUAL_FILTER
+
+const int kBitdepth[] = { 10, 12 };
+
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, AV1HbdConvolveOptimzTest,
+    ::testing::Combine(::testing::Values(av1_highbd_convolve_init_sse4_1),
+                       ::testing::Values(av1_highbd_convolve_horiz_sse4_1),
+                       ::testing::Values(av1_highbd_convolve_vert_sse4_1),
+                       ::testing::ValuesIn(kBlockDim),
+                       ::testing::ValuesIn(kFilter),
+                       ::testing::ValuesIn(kSubpelQ4),
+                       ::testing::ValuesIn(kAvg),
+                       ::testing::ValuesIn(kBitdepth)));
+#endif  // HAVE_SSE4_1 && CONFIG_DUAL_FILTER
+#endif  // CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/av1_convolve_test.cc b/third_party/aom/test/av1_convolve_test.cc
new file mode 100644
index 0000000000..02ac8e7bb9
--- /dev/null
+++ b/third_party/aom/test/av1_convolve_test.cc
@@ -0,0 +1,522 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <algorithm>
+#include <vector>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+#include "av1/common/filter.h"
+#include "av1/common/convolve.h"
+#include "test/acm_random.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+using std::tr1::tuple;
+static void filter_block1d_horiz_c(const uint8_t *src_ptr, int src_stride,
+                                   const int16_t *filter, int tap,
+                                   uint8_t *dst_ptr, int dst_stride, int w,
+                                   int h) {
+  src_ptr -= tap / 2 - 1;
+  for (int r = 0; r < h; ++r) {
+    for (int c = 0; c < w; ++c) {
+      int sum = 0;
+      for (int i = 0; i < tap; ++i) {
+        sum += src_ptr[c + i] * filter[i];
+      }
+      dst_ptr[c] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void filter_block1d_vert_c(const uint8_t *src_ptr, int src_stride,
+                                  const int16_t *filter, int tap,
+                                  uint8_t *dst_ptr, int dst_stride, int w,
+                                  int h) {
+  src_ptr -= (tap / 2 - 1) * src_stride;
+  for (int r = 0; r < h; ++r) {
+    for (int c = 0; c < w; ++c) {
+      int sum = 0;
+      for (int i = 0; i < tap; ++i) {
+        sum += src_ptr[c + i * src_stride] * filter[i];
+      }
+      dst_ptr[c] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+    }
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static int match(const uint8_t *out, int out_stride, const uint8_t *ref_out,
+                 int ref_out_stride, int w, int h) {
+  for (int r = 0; r < h; ++r) {
+    for (int c = 0; c < w; ++c) {
+      if (out[r * out_stride + c] != ref_out[r * ref_out_stride + c]) return 0;
+    }
+  }
+  return 1;
+}
+
+typedef void (*ConvolveFunc)(const uint8_t *src, int src_stride, uint8_t *dst,
+                             int dst_stride, int w, int h,
+                             const InterpFilterParams filter_params,
+                             const int subpel_q4, int step_q4,
+                             ConvolveParams *conv_params);
+
+struct ConvolveFunctions {
+  ConvolveFunctions(ConvolveFunc hf, ConvolveFunc vf) : hf_(hf), vf_(vf) {}
+  ConvolveFunc hf_;
+  ConvolveFunc vf_;
+};
+
+typedef tuple<ConvolveFunctions *, InterpFilter /*filter_x*/,
+              InterpFilter /*filter_y*/>
+    ConvolveParam;
+
+class Av1ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
+ public:
+  virtual void SetUp() {
+    rnd_(ACMRandom::DeterministicSeed());
+    cfs_ = GET_PARAM(0);
+    interp_filter_ls_[0] = GET_PARAM(2);
+    interp_filter_ls_[2] = interp_filter_ls_[0];
+    interp_filter_ls_[1] = GET_PARAM(1);
+    interp_filter_ls_[3] = interp_filter_ls_[1];
+  }
+  virtual void TearDown() {
+    while (buf_ls_.size() > 0) {
+      uint8_t *buf = buf_ls_.back();
+      aom_free(buf);
+      buf_ls_.pop_back();
+    }
+  }
+  virtual uint8_t *add_input(int w, int h, int *stride) {
+    uint8_t *buf =
+        reinterpret_cast<uint8_t *>(aom_memalign(kDataAlignment, kBufferSize));
+    buf_ls_.push_back(buf);
+    *stride = w + MAX_FILTER_TAP - 1;
+    int offset = MAX_FILTER_TAP / 2 - 1;
+    for (int r = 0; r < h + MAX_FILTER_TAP - 1; ++r) {
+      for (int c = 0; c < w + MAX_FILTER_TAP - 1; ++c) {
+        buf[r * (*stride) + c] = rnd_.Rand8();
+      }
+    }
+    return buf + offset * (*stride) + offset;
+  }
+  virtual uint8_t *add_output(int w, int /*h*/, int *stride) {
+    uint8_t *buf =
+        reinterpret_cast<uint8_t *>(aom_memalign(kDataAlignment, kBufferSize));
+    buf_ls_.push_back(buf);
+    *stride = w;
+    return buf;
+  }
+  virtual void random_init_buf(uint8_t *buf, int w, int h, int stride) {
+    for (int r = 0; r < h; ++r) {
+      for (int c = 0; c < w; ++c) {
+        buf[r * stride + c] = rnd_.Rand8();
+      }
+    }
+  }
+
+ protected:
+  static const int kDataAlignment = 16;
+  static const int kOuterBlockSize = MAX_SB_SIZE + MAX_FILTER_TAP - 1;
+  static const int kBufferSize = kOuterBlockSize * kOuterBlockSize;
+  std::vector<uint8_t *> buf_ls_;
+  InterpFilter interp_filter_ls_[4];
+  ConvolveFunctions *cfs_;
+  ACMRandom rnd_;
+};
+
+int bsize_ls[] = { 1, 2, 4, 8, 16, 32, 64, 3, 7, 15, 31, 63 };
+int bsize_num = sizeof(bsize_ls) / sizeof(bsize_ls[0]);
+
+TEST_P(Av1ConvolveTest, av1_convolve_vert) {
+  const int y_step_q4 = 16;
+  ConvolveParams conv_params = get_conv_params(0, 0);
+
+  int in_stride, out_stride, ref_out_stride, avg_out_stride, ref_avg_out_stride;
+  uint8_t *in = add_input(MAX_SB_SIZE, MAX_SB_SIZE, &in_stride);
+  uint8_t *out = add_output(MAX_SB_SIZE, MAX_SB_SIZE, &out_stride);
+  uint8_t *ref_out = add_output(MAX_SB_SIZE, MAX_SB_SIZE, &ref_out_stride);
+  uint8_t *avg_out = add_output(MAX_SB_SIZE, MAX_SB_SIZE, &avg_out_stride);
+  uint8_t *ref_avg_out =
+      add_output(MAX_SB_SIZE, MAX_SB_SIZE, &ref_avg_out_stride);
+  for (int hb_idx = 0; hb_idx < bsize_num; ++hb_idx) {
+    for (int vb_idx = 0; vb_idx < bsize_num; ++vb_idx) {
+      int w = bsize_ls[hb_idx];
+      int h = bsize_ls[vb_idx];
+      for (int subpel_y_q4 = 0; subpel_y_q4 < SUBPEL_SHIFTS; ++subpel_y_q4) {
+        InterpFilter filter_y = interp_filter_ls_[0];
+        InterpFilterParams param_vert = av1_get_interp_filter_params(filter_y);
+        const int16_t *filter_vert =
+            av1_get_interp_filter_subpel_kernel(param_vert, subpel_y_q4);
+
+        filter_block1d_vert_c(in, in_stride, filter_vert, param_vert.taps,
+                              ref_out, ref_out_stride, w, h);
+
+        conv_params.ref = 0;
+        cfs_->vf_(in, in_stride, out, out_stride, w, h, param_vert, subpel_y_q4,
+                  y_step_q4, &conv_params);
+        EXPECT_EQ(match(out, out_stride, ref_out, ref_out_stride, w, h), 1)
+            << " hb_idx " << hb_idx << " vb_idx " << vb_idx << " filter_y "
+            << filter_y << " subpel_y_q4 " << subpel_y_q4;
+
+        random_init_buf(avg_out, w, h, avg_out_stride);
+        for (int r = 0; r < h; ++r) {
+          for (int c = 0; c < w; ++c) {
+            ref_avg_out[r * ref_avg_out_stride + c] = ROUND_POWER_OF_TWO(
+                avg_out[r * avg_out_stride + c] + out[r * out_stride + c], 1);
+          }
+        }
+        conv_params.ref = 1;
+        cfs_->vf_(in, in_stride, avg_out, avg_out_stride, w, h, param_vert,
+                  subpel_y_q4, y_step_q4, &conv_params);
+        EXPECT_EQ(match(avg_out, avg_out_stride, ref_avg_out,
+                        ref_avg_out_stride, w, h),
+                  1)
+            << " hb_idx " << hb_idx << " vb_idx " << vb_idx << " filter_y "
+            << filter_y << " subpel_y_q4 " << subpel_y_q4;
+      }
+    }
+  }
+};
+
+TEST_P(Av1ConvolveTest, av1_convolve_horiz) {
+  const int x_step_q4 = 16;
+  ConvolveParams conv_params = get_conv_params(0, 0);
+
+  int in_stride, out_stride, ref_out_stride, avg_out_stride, ref_avg_out_stride;
+  uint8_t *in = add_input(MAX_SB_SIZE, MAX_SB_SIZE, &in_stride);
+  uint8_t *out = add_output(MAX_SB_SIZE, MAX_SB_SIZE, &out_stride);
+  uint8_t *ref_out = add_output(MAX_SB_SIZE, MAX_SB_SIZE, &ref_out_stride);
+  uint8_t *avg_out = add_output(MAX_SB_SIZE, MAX_SB_SIZE, &avg_out_stride);
+  uint8_t *ref_avg_out =
+      add_output(MAX_SB_SIZE, MAX_SB_SIZE, &ref_avg_out_stride);
+  for (int hb_idx = 0; hb_idx < bsize_num; ++hb_idx) {
+    for (int vb_idx = 0; vb_idx < bsize_num; ++vb_idx) {
+      int w = bsize_ls[hb_idx];
+      int h = bsize_ls[vb_idx];
+      for (int subpel_x_q4 = 0; subpel_x_q4 < SUBPEL_SHIFTS; ++subpel_x_q4) {
+        InterpFilter filter_x = interp_filter_ls_[1];
+        InterpFilterParams param_horiz = av1_get_interp_filter_params(filter_x);
+        const int16_t *filter_horiz =
+            av1_get_interp_filter_subpel_kernel(param_horiz, subpel_x_q4);
+
+        filter_block1d_horiz_c(in, in_stride, filter_horiz, param_horiz.taps,
+                               ref_out, ref_out_stride, w, h);
+
+        conv_params.ref = 0;
+        cfs_->hf_(in, in_stride, out, out_stride, w, h, param_horiz,
+                  subpel_x_q4, x_step_q4, &conv_params);
+        EXPECT_EQ(match(out, out_stride, ref_out, ref_out_stride, w, h), 1)
+            << " hb_idx " << hb_idx << " vb_idx " << vb_idx << " filter_x "
+            << filter_x << " subpel_x_q4 " << subpel_x_q4;
+
+        random_init_buf(avg_out, w, h, avg_out_stride);
+        for (int r = 0; r < h; ++r) {
+          for (int c = 0; c < w; ++c) {
+            ref_avg_out[r * ref_avg_out_stride + c] = ROUND_POWER_OF_TWO(
+                avg_out[r * avg_out_stride + c] + out[r * out_stride + c], 1);
+          }
+        }
+        conv_params.ref = 1;
+        cfs_->hf_(in, in_stride, avg_out, avg_out_stride, w, h, param_horiz,
+                  subpel_x_q4, x_step_q4, &conv_params);
+        EXPECT_EQ(match(avg_out, avg_out_stride, ref_avg_out,
+                        ref_avg_out_stride, w, h),
+                  1)
+            << "hb_idx " << hb_idx << "vb_idx" << vb_idx << " filter_x "
+            << filter_x << "subpel_x_q4 " << subpel_x_q4;
+      }
+    }
+  }
+};
+
+ConvolveFunctions convolve_functions_c(av1_convolve_horiz_c,
+                                       av1_convolve_vert_c);
+
+InterpFilter filter_ls[] = { EIGHTTAP_REGULAR, EIGHTTAP_SMOOTH,
+                             MULTITAP_SHARP };
+
+INSTANTIATE_TEST_CASE_P(
+    C, Av1ConvolveTest,
+    ::testing::Combine(::testing::Values(&convolve_functions_c),
+                       ::testing::ValuesIn(filter_ls),
+                       ::testing::ValuesIn(filter_ls)));
+
+#if CONFIG_HIGHBITDEPTH
+TEST(AV1ConvolveTest, av1_highbd_convolve) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+#if CONFIG_DUAL_FILTER
+  InterpFilter interp_filter[4] = { EIGHTTAP_REGULAR, EIGHTTAP_REGULAR,
+                                    EIGHTTAP_REGULAR, EIGHTTAP_REGULAR };
+  InterpFilterParams filter_params =
+      av1_get_interp_filter_params(interp_filter[0]);
+#else
+  InterpFilter interp_filter = EIGHTTAP_REGULAR;
+  InterpFilterParams filter_params =
+      av1_get_interp_filter_params(interp_filter);
+#endif
+  int filter_size = filter_params.taps;
+  int filter_center = filter_size / 2 - 1;
+  uint16_t src[12 * 12];
+  int src_stride = filter_size;
+  uint16_t dst[1] = { 0 };
+  int dst_stride = 1;
+  int x_step_q4 = 16;
+  int y_step_q4 = 16;
+  int avg = 0;
+  int bd = 10;
+  int w = 1;
+  int h = 1;
+
+  int subpel_x_q4;
+  int subpel_y_q4;
+
+  for (int i = 0; i < filter_size * filter_size; i++) {
+    src[i] = rnd.Rand16() % (1 << bd);
+  }
+
+  for (subpel_x_q4 = 0; subpel_x_q4 < SUBPEL_SHIFTS; subpel_x_q4++) {
+    for (subpel_y_q4 = 0; subpel_y_q4 < SUBPEL_SHIFTS; subpel_y_q4++) {
+      av1_highbd_convolve(
+          CONVERT_TO_BYTEPTR(src + src_stride * filter_center + filter_center),
+          src_stride, CONVERT_TO_BYTEPTR(dst), dst_stride, w, h, interp_filter,
+          subpel_x_q4, x_step_q4, subpel_y_q4, y_step_q4, avg, bd);
+
+      const int16_t *x_filter =
+          av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
+      const int16_t *y_filter =
+          av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
+
+      int temp[12];
+      int dst_ref = 0;
+      for (int r = 0; r < filter_size; r++) {
+        temp[r] = 0;
+        for (int c = 0; c < filter_size; c++) {
+          temp[r] += x_filter[c] * src[r * filter_size + c];
+        }
+        temp[r] =
+            clip_pixel_highbd(ROUND_POWER_OF_TWO(temp[r], FILTER_BITS), bd);
+        dst_ref += temp[r] * y_filter[r];
+      }
+      dst_ref = clip_pixel_highbd(ROUND_POWER_OF_TWO(dst_ref, FILTER_BITS), bd);
+      EXPECT_EQ(dst[0], dst_ref);
+    }
+  }
+}
+
+TEST(AV1ConvolveTest, av1_highbd_convolve_avg) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+#if CONFIG_DUAL_FILTER
+  InterpFilter interp_filter[4] = { EIGHTTAP_REGULAR, EIGHTTAP_REGULAR,
+                                    EIGHTTAP_REGULAR, EIGHTTAP_REGULAR };
+  InterpFilterParams filter_params =
+      av1_get_interp_filter_params(interp_filter[0]);
+#else
+  InterpFilter interp_filter = EIGHTTAP_REGULAR;
+  InterpFilterParams filter_params =
+      av1_get_interp_filter_params(interp_filter);
+#endif
+  int filter_size = filter_params.taps;
+  int filter_center = filter_size / 2 - 1;
+  uint16_t src0[12 * 12];
+  uint16_t src1[12 * 12];
+  int src_stride = filter_size;
+  uint16_t dst0[1] = { 0 };
+  uint16_t dst1[1] = { 0 };
+  uint16_t dst[1] = { 0 };
+  int dst_stride = 1;
+  int x_step_q4 = 16;
+  int y_step_q4 = 16;
+  int avg = 0;
+  int bd = 10;
+
+  int w = 1;
+  int h = 1;
+
+  int subpel_x_q4;
+  int subpel_y_q4;
+
+  for (int i = 0; i < filter_size * filter_size; i++) {
+    src0[i] = rnd.Rand16() % (1 << bd);
+    src1[i] = rnd.Rand16() % (1 << bd);
+  }
+
+  for (subpel_x_q4 = 0; subpel_x_q4 < SUBPEL_SHIFTS; subpel_x_q4++) {
+    for (subpel_y_q4 = 0; subpel_y_q4 < SUBPEL_SHIFTS; subpel_y_q4++) {
+      int offset = filter_size * filter_center + filter_center;
+
+      avg = 0;
+      av1_highbd_convolve(CONVERT_TO_BYTEPTR(src0 + offset), src_stride,
+                          CONVERT_TO_BYTEPTR(dst0), dst_stride, w, h,
+                          interp_filter, subpel_x_q4, x_step_q4, subpel_y_q4,
+                          y_step_q4, avg, bd);
+      avg = 0;
+      av1_highbd_convolve(CONVERT_TO_BYTEPTR(src1 + offset), src_stride,
+                          CONVERT_TO_BYTEPTR(dst1), dst_stride, w, h,
+                          interp_filter, subpel_x_q4, x_step_q4, subpel_y_q4,
+                          y_step_q4, avg, bd);
+
+      avg = 0;
+      av1_highbd_convolve(CONVERT_TO_BYTEPTR(src0 + offset), src_stride,
+                          CONVERT_TO_BYTEPTR(dst), dst_stride, w, h,
+                          interp_filter, subpel_x_q4, x_step_q4, subpel_y_q4,
+                          y_step_q4, avg, bd);
+      avg = 1;
+      av1_highbd_convolve(CONVERT_TO_BYTEPTR(src1 + offset), src_stride,
+                          CONVERT_TO_BYTEPTR(dst), dst_stride, w, h,
+                          interp_filter, subpel_x_q4, x_step_q4, subpel_y_q4,
+                          y_step_q4, avg, bd);
+
+      EXPECT_EQ(dst[0], ROUND_POWER_OF_TWO(dst0[0] + dst1[0], 1));
+    }
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#define CONVOLVE_SPEED_TEST 0
+#if CONVOLVE_SPEED_TEST
+#define highbd_convolve_speed(func, block_size, frame_size)                  \
+  TEST(AV1ConvolveTest, func##_speed_##block_size##_##frame_size) {          \
+    ACMRandom rnd(ACMRandom::DeterministicSeed());                           \
+    InterpFilter interp_filter = EIGHTTAP;                                   \
+    InterpFilterParams filter_params =                                       \
+        av1_get_interp_filter_params(interp_filter);                         \
+    int filter_size = filter_params.tap;                                     \
+    int filter_center = filter_size / 2 - 1;                                 \
+    DECLARE_ALIGNED(16, uint16_t,                                            \
+                    src[(frame_size + 7) * (frame_size + 7)]) = { 0 };       \
+    int src_stride = frame_size + 7;                                         \
+    DECLARE_ALIGNED(16, uint16_t, dst[frame_size * frame_size]) = { 0 };     \
+    int dst_stride = frame_size;                                             \
+    int x_step_q4 = 16;                                                      \
+    int y_step_q4 = 16;                                                      \
+    int subpel_x_q4 = 8;                                                     \
+    int subpel_y_q4 = 6;                                                     \
+    int bd = 10;                                                             \
+                                                                             \
+    int w = block_size;                                                      \
+    int h = block_size;                                                      \
+                                                                             \
+    const int16_t *filter_x =                                                \
+        av1_get_interp_filter_kernel(filter_params, subpel_x_q4);            \
+    const int16_t *filter_y =                                                \
+        av1_get_interp_filter_kernel(filter_params, subpel_y_q4);            \
+                                                                             \
+    for (int i = 0; i < src_stride * src_stride; i++) {                      \
+      src[i] = rnd.Rand16() % (1 << bd);                                     \
+    }                                                                        \
+                                                                             \
+    int offset = filter_center * src_stride + filter_center;                 \
+    int row_offset = 0;                                                      \
+    int col_offset = 0;                                                      \
+    for (int i = 0; i < 100000; i++) {                                       \
+      int src_total_offset = offset + col_offset * src_stride + row_offset;  \
+      int dst_total_offset = col_offset * dst_stride + row_offset;           \
+      func(CONVERT_TO_BYTEPTR(src + src_total_offset), src_stride,           \
+           CONVERT_TO_BYTEPTR(dst + dst_total_offset), dst_stride, filter_x, \
+           x_step_q4, filter_y, y_step_q4, w, h, bd);                        \
+      if (offset + w + w < frame_size) {                                     \
+        row_offset += w;                                                     \
+      } else {                                                               \
+        row_offset = 0;                                                      \
+        col_offset += h;                                                     \
+      }                                                                      \
+      if (col_offset + h >= frame_size) {                                    \
+        col_offset = 0;                                                      \
+      }                                                                      \
+    }                                                                        \
+  }
+
+#define lowbd_convolve_speed(func, block_size, frame_size)                  \
+  TEST(AV1ConvolveTest, func##_speed_l_##block_size##_##frame_size) {       \
+    ACMRandom rnd(ACMRandom::DeterministicSeed());                          \
+    InterpFilter interp_filter = EIGHTTAP;                                  \
+    InterpFilterParams filter_params =                                      \
+        av1_get_interp_filter_params(interp_filter);                        \
+    int filter_size = filter_params.tap;                                    \
+    int filter_center = filter_size / 2 - 1;                                \
+    DECLARE_ALIGNED(16, uint8_t, src[(frame_size + 7) * (frame_size + 7)]); \
+    int src_stride = frame_size + 7;                                        \
+    DECLARE_ALIGNED(16, uint8_t, dst[frame_size * frame_size]);             \
+    int dst_stride = frame_size;                                            \
+    int x_step_q4 = 16;                                                     \
+    int y_step_q4 = 16;                                                     \
+    int subpel_x_q4 = 8;                                                    \
+    int subpel_y_q4 = 6;                                                    \
+    int bd = 8;                                                             \
+                                                                            \
+    int w = block_size;                                                     \
+    int h = block_size;                                                     \
+                                                                            \
+    const int16_t *filter_x =                                               \
+        av1_get_interp_filter_kernel(filter_params, subpel_x_q4);           \
+    const int16_t *filter_y =                                               \
+        av1_get_interp_filter_kernel(filter_params, subpel_y_q4);           \
+                                                                            \
+    for (int i = 0; i < src_stride * src_stride; i++) {                     \
+      src[i] = rnd.Rand16() % (1 << bd);                                    \
+    }                                                                       \
+                                                                            \
+    int offset = filter_center * src_stride + filter_center;                \
+    int row_offset = 0;                                                     \
+    int col_offset = 0;                                                     \
+    for (int i = 0; i < 100000; i++) {                                      \
+      func(src + offset, src_stride, dst, dst_stride, filter_x, x_step_q4,  \
+           filter_y, y_step_q4, w, h);                                      \
+      if (offset + w + w < frame_size) {                                    \
+        row_offset += w;                                                    \
+      } else {                                                              \
+        row_offset = 0;                                                     \
+        col_offset += h;                                                    \
+      }                                                                     \
+      if (col_offset + h >= frame_size) {                                   \
+        col_offset = 0;                                                     \
+      }                                                                     \
+    }                                                                       \
+  }
+
+// This experiment shows that when frame size is 64x64
+// aom_highbd_convolve8_sse2 and aom_convolve8_sse2's speed are similar.
+// However when frame size becomes 1024x1024
+// aom_highbd_convolve8_sse2 is around 50% slower than aom_convolve8_sse2
+// we think the bottleneck is from memory IO
+highbd_convolve_speed(aom_highbd_convolve8_sse2, 8, 64);
+highbd_convolve_speed(aom_highbd_convolve8_sse2, 16, 64);
+highbd_convolve_speed(aom_highbd_convolve8_sse2, 32, 64);
+highbd_convolve_speed(aom_highbd_convolve8_sse2, 64, 64);
+
+lowbd_convolve_speed(aom_convolve8_sse2, 8, 64);
+lowbd_convolve_speed(aom_convolve8_sse2, 16, 64);
+lowbd_convolve_speed(aom_convolve8_sse2, 32, 64);
+lowbd_convolve_speed(aom_convolve8_sse2, 64, 64);
+
+highbd_convolve_speed(aom_highbd_convolve8_sse2, 8, 1024);
+highbd_convolve_speed(aom_highbd_convolve8_sse2, 16, 1024);
+highbd_convolve_speed(aom_highbd_convolve8_sse2, 32, 1024);
+highbd_convolve_speed(aom_highbd_convolve8_sse2, 64, 1024);
+
+lowbd_convolve_speed(aom_convolve8_sse2, 8, 1024);
+lowbd_convolve_speed(aom_convolve8_sse2, 16, 1024);
+lowbd_convolve_speed(aom_convolve8_sse2, 32, 1024);
+lowbd_convolve_speed(aom_convolve8_sse2, 64, 1024);
+#endif  // CONVOLVE_SPEED_TEST
+}  // namespace
diff --git a/third_party/aom/test/av1_dct_test.cc b/third_party/aom/test/av1_dct_test.cc
new file mode 100644
index 0000000000..691cc8b794
--- /dev/null
+++ b/third_party/aom/test/av1_dct_test.cc
@@ -0,0 +1,108 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <new>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/acm_random.h"
+#include "test/util.h"
+#include "./aom_config.h"
+#include "aom_ports/msvc.h"
+
+#undef CONFIG_COEFFICIENT_RANGE_CHECKING
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 1
+#define AV1_DCT_GTEST
+#include "av1/encoder/dct.c"
+
+using libaom_test::ACMRandom;
+
+namespace {
+void reference_dct_1d(const double *in, double *out, int size) {
+  const double kInvSqrt2 = 0.707106781186547524400844362104;
+  for (int k = 0; k < size; ++k) {
+    out[k] = 0;
+    for (int n = 0; n < size; ++n) {
+      out[k] += in[n] * cos(PI * (2 * n + 1) * k / (2 * size));
+    }
+    if (k == 0) out[k] = out[k] * kInvSqrt2;
+  }
+}
+
+typedef void (*FdctFuncRef)(const double *in, double *out, int size);
+typedef void (*IdctFuncRef)(const double *in, double *out, int size);
+typedef void (*FdctFunc)(const tran_low_t *in, tran_low_t *out);
+typedef void (*IdctFunc)(const tran_low_t *in, tran_low_t *out);
+
+class TransTestBase {
+ public:
+  virtual ~TransTestBase() {}
+
+ protected:
+  void RunFwdAccuracyCheck() {
+    tran_low_t *input = new tran_low_t[txfm_size_];
+    tran_low_t *output = new tran_low_t[txfm_size_];
+    double *ref_input = new double[txfm_size_];
+    double *ref_output = new double[txfm_size_];
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 5000;
+    for (int ti = 0; ti < count_test_block; ++ti) {
+      for (int ni = 0; ni < txfm_size_; ++ni) {
+        input[ni] = rnd.Rand8() - rnd.Rand8();
+        ref_input[ni] = static_cast<double>(input[ni]);
+      }
+
+      fwd_txfm_(input, output);
+      fwd_txfm_ref_(ref_input, ref_output, txfm_size_);
+
+      for (int ni = 0; ni < txfm_size_; ++ni) {
+        EXPECT_LE(
+            abs(output[ni] - static_cast<tran_low_t>(round(ref_output[ni]))),
+            max_error_);
+      }
+    }
+
+    delete[] input;
+    delete[] output;
+    delete[] ref_input;
+    delete[] ref_output;
+  }
+
+  double max_error_;
+  int txfm_size_;
+  FdctFunc fwd_txfm_;
+  FdctFuncRef fwd_txfm_ref_;
+};
+
+typedef std::tr1::tuple<FdctFunc, FdctFuncRef, int, int> FdctParam;
+class AV1FwdTxfm : public TransTestBase,
+                   public ::testing::TestWithParam<FdctParam> {
+ public:
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    fwd_txfm_ref_ = GET_PARAM(1);
+    txfm_size_ = GET_PARAM(2);
+    max_error_ = GET_PARAM(3);
+  }
+  virtual void TearDown() {}
+};
+
+TEST_P(AV1FwdTxfm, RunFwdAccuracyCheck) { RunFwdAccuracyCheck(); }
+
+INSTANTIATE_TEST_CASE_P(
+    C, AV1FwdTxfm,
+    ::testing::Values(FdctParam(&fdct4, &reference_dct_1d, 4, 1),
+                      FdctParam(&fdct8, &reference_dct_1d, 8, 1),
+                      FdctParam(&fdct16, &reference_dct_1d, 16, 2),
+                      FdctParam(&fdct32, &reference_dct_1d, 32, 3)));
+}  // namespace
diff --git a/third_party/aom/test/av1_ext_tile_test.cc b/third_party/aom/test/av1_ext_tile_test.cc
new file mode 100644
index 0000000000..f96447965d
--- /dev/null
+++ b/third_party/aom/test/av1_ext_tile_test.cc
@@ -0,0 +1,200 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <string>
+#include <vector>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+
+namespace {
+// The number of frames to be encoded/decoded
+const int kLimit = 8;
+// Skip 1 frame to check the frame decoding independency.
+const int kSkip = 5;
+const int kTileSize = 1;
+const int kTIleSizeInPixels = (kTileSize << 6);
+// Fake width and height so that they can be multiples of the tile size.
+const int kImgWidth = 704;
+const int kImgHeight = 576;
+
+// This test tests "tile_encoding_mode = TILE_VR" case. The TILE_NORMAL case is
+// tested by the tile_independence test.
+class AV1ExtTileTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<libaom_test::TestMode, int> {
+ protected:
+  AV1ExtTileTest()
+      : EncoderTest(GET_PARAM(0)), encoding_mode_(GET_PARAM(1)),
+        set_cpu_used_(GET_PARAM(2)) {
+    init_flags_ = AOM_CODEC_USE_PSNR;
+    aom_codec_dec_cfg_t cfg = aom_codec_dec_cfg_t();
+    cfg.w = kImgWidth;
+    cfg.h = kImgHeight;
+
+    decoder_ = codec_->CreateDecoder(cfg, 0);
+    decoder_->Control(AV1_SET_DECODE_TILE_ROW, -1);
+    decoder_->Control(AV1_SET_DECODE_TILE_COL, -1);
+
+    // Allocate buffer to store tile image.
+    aom_img_alloc(&tile_img_, AOM_IMG_FMT_I420, kImgWidth, kImgHeight, 32);
+
+    md5_.clear();
+    tile_md5_.clear();
+  }
+
+  virtual ~AV1ExtTileTest() {
+    aom_img_free(&tile_img_);
+    delete decoder_;
+  }
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+
+    cfg_.g_lag_in_frames = 0;
+    cfg_.rc_end_usage = AOM_VBR;
+    cfg_.g_error_resilient = 1;
+
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_min_quantizer = 0;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      // Encode setting
+      encoder->Control(AOME_SET_CPUUSED, set_cpu_used_);
+      encoder->Control(AOME_SET_ENABLEAUTOALTREF, 0);
+      encoder->Control(AV1E_SET_FRAME_PARALLEL_DECODING, 1);
+
+      // The tile size is 64x64.
+      encoder->Control(AV1E_SET_TILE_COLUMNS, kTileSize);
+      encoder->Control(AV1E_SET_TILE_ROWS, kTileSize);
+      encoder->Control(AV1E_SET_TILE_ENCODING_MODE, 1);  // TILE_VR
+#if CONFIG_EXT_PARTITION
+      // Always use 64x64 max partition.
+      encoder->Control(AV1E_SET_SUPERBLOCK_SIZE, AOM_SUPERBLOCK_SIZE_64X64);
+#endif
+    }
+
+    if (video->frame() == 1) {
+      frame_flags_ =
+          AOM_EFLAG_NO_UPD_LAST | AOM_EFLAG_NO_UPD_GF | AOM_EFLAG_NO_UPD_ARF;
+    }
+  }
+
+  virtual void DecompressedFrameHook(const aom_image_t &img,
+                                     aom_codec_pts_t pts) {
+    // Skip 1 already decoded frame to be consistent with the decoder in this
+    // test.
+    if (pts == (aom_codec_pts_t)kSkip) return;
+
+    // Calculate MD5 as the reference.
+    ::libaom_test::MD5 md5_res;
+    md5_res.Add(&img);
+    md5_.push_back(md5_res.Get());
+  }
+
+  virtual void FramePktHook(const aom_codec_cx_pkt_t *pkt) {
+    // Skip decoding 1 frame.
+    if (pkt->data.frame.pts == (aom_codec_pts_t)kSkip) return;
+
+    bool IsLastFrame = (pkt->data.frame.pts == (aom_codec_pts_t)(kLimit - 1));
+
+    // Decode the first (kLimit - 1) frames as whole frame, and decode the last
+    // frame in single tiles.
+    for (int r = 0; r < kImgHeight / kTIleSizeInPixels; ++r) {
+      for (int c = 0; c < kImgWidth / kTIleSizeInPixels; ++c) {
+        if (!IsLastFrame) {
+          decoder_->Control(AV1_SET_DECODE_TILE_ROW, -1);
+          decoder_->Control(AV1_SET_DECODE_TILE_COL, -1);
+        } else {
+          decoder_->Control(AV1_SET_DECODE_TILE_ROW, r);
+          decoder_->Control(AV1_SET_DECODE_TILE_COL, c);
+        }
+
+        const aom_codec_err_t res = decoder_->DecodeFrame(
+            reinterpret_cast<uint8_t *>(pkt->data.frame.buf),
+            pkt->data.frame.sz);
+        if (res != AOM_CODEC_OK) {
+          abort_ = true;
+          ASSERT_EQ(AOM_CODEC_OK, res);
+        }
+        const aom_image_t *img = decoder_->GetDxData().Next();
+
+        if (!IsLastFrame) {
+          if (img) {
+            ::libaom_test::MD5 md5_res;
+            md5_res.Add(img);
+            tile_md5_.push_back(md5_res.Get());
+          }
+          break;
+        }
+
+        const int kMaxMBPlane = 3;
+        for (int plane = 0; plane < kMaxMBPlane; ++plane) {
+          const int shift = (plane == 0) ? 0 : 1;
+          int tile_height = kTIleSizeInPixels >> shift;
+          int tile_width = kTIleSizeInPixels >> shift;
+
+          for (int tr = 0; tr < tile_height; ++tr) {
+            memcpy(tile_img_.planes[plane] +
+                       tile_img_.stride[plane] * (r * tile_height + tr) +
+                       c * tile_width,
+                   img->planes[plane] + img->stride[plane] * tr, tile_width);
+          }
+        }
+      }
+
+      if (!IsLastFrame) break;
+    }
+
+    if (IsLastFrame) {
+      ::libaom_test::MD5 md5_res;
+      md5_res.Add(&tile_img_);
+      tile_md5_.push_back(md5_res.Get());
+    }
+  }
+
+  ::libaom_test::TestMode encoding_mode_;
+  int set_cpu_used_;
+  ::libaom_test::Decoder *decoder_;
+  aom_image_t tile_img_;
+  std::vector<std::string> md5_;
+  std::vector<std::string> tile_md5_;
+};
+
+TEST_P(AV1ExtTileTest, DecoderResultTest) {
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", kImgWidth,
+                                       kImgHeight, 30, 1, 0, kLimit);
+  cfg_.rc_target_bitrate = 500;
+  cfg_.g_error_resilient = AOM_ERROR_RESILIENT_DEFAULT;
+  cfg_.g_lag_in_frames = 0;
+  cfg_.g_threads = 1;
+
+  // Tile encoding
+  init_flags_ = AOM_CODEC_USE_PSNR;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  // Compare to check if two vectors are equal.
+  ASSERT_EQ(md5_, tile_md5_);
+}
+
+AV1_INSTANTIATE_TEST_CASE(
+    // Now only test 2-pass mode.
+    AV1ExtTileTest, ::testing::Values(::libaom_test::kTwoPassGood),
+    ::testing::Range(0, 4));
+}  // namespace
diff --git a/third_party/aom/test/av1_fht16x16_test.cc b/third_party/aom/test/av1_fht16x16_test.cc
new file mode 100644
index 0000000000..e1032ef241
--- /dev/null
+++ b/third_party/aom/test/av1_fht16x16_test.cc
@@ -0,0 +1,275 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+#include "aom_ports/mem.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+using std::tr1::tuple;
+using libaom_test::FhtFunc;
+typedef tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht16x16Param;
+
+void fht16x16_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht16x16_c(in, out, stride, tx_type);
+}
+
+void iht16x16_ref(const tran_low_t *in, uint8_t *dest, int stride,
+                  int tx_type) {
+  av1_iht16x16_256_add_c(in, dest, stride, tx_type);
+}
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*IHbdHtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                           int tx_type, int bd);
+typedef void (*HbdHtFunc)(const int16_t *input, int32_t *output, int stride,
+                          int tx_type, int bd);
+
+// Target optimized function, tx_type, bit depth
+typedef tuple<HbdHtFunc, int, int> HighbdHt16x16Param;
+
+void highbd_fht16x16_ref(const int16_t *in, int32_t *out, int stride,
+                         int tx_type, int bd) {
+  av1_fwd_txfm2d_16x16_c(in, out, stride, tx_type, bd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+class AV1Trans16x16HT : public libaom_test::TransformTestBase,
+                        public ::testing::TestWithParam<Ht16x16Param> {
+ public:
+  virtual ~AV1Trans16x16HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 16;
+    height_ = 16;
+    fwd_txfm_ref = fht16x16_ref;
+    inv_txfm_ref = iht16x16_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(AV1Trans16x16HT, MemCheck) { RunMemCheck(); }
+TEST_P(AV1Trans16x16HT, AccuracyCheck) { RunAccuracyCheck(1, 0.001); }
+TEST_P(AV1Trans16x16HT, InvAccuracyCheck) { RunInvAccuracyCheck(1); }
+TEST_P(AV1Trans16x16HT, CoeffCheck) { RunCoeffCheck(); }
+TEST_P(AV1Trans16x16HT, InvCoeffCheck) { RunInvCoeffCheck(); }
+
+#if CONFIG_HIGHBITDEPTH
+class AV1HighbdTrans16x16HT
+    : public ::testing::TestWithParam<HighbdHt16x16Param> {
+ public:
+  virtual ~AV1HighbdTrans16x16HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    fwd_txfm_ref_ = highbd_fht16x16_ref;
+    tx_type_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = 256;
+
+    input_ = reinterpret_cast<int16_t *>(
+        aom_memalign(32, sizeof(int16_t) * num_coeffs_));
+    output_ = reinterpret_cast<int32_t *>(
+        aom_memalign(32, sizeof(int32_t) * num_coeffs_));
+    output_ref_ = reinterpret_cast<int32_t *>(
+        aom_memalign(32, sizeof(int32_t) * num_coeffs_));
+  }
+
+  virtual void TearDown() {
+    aom_free(input_);
+    aom_free(output_);
+    aom_free(output_ref_);
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void RunBitexactCheck();
+
+ private:
+  HbdHtFunc fwd_txfm_;
+  HbdHtFunc fwd_txfm_ref_;
+  int tx_type_;
+  int bit_depth_;
+  int mask_;
+  int num_coeffs_;
+  int16_t *input_;
+  int32_t *output_;
+  int32_t *output_ref_;
+};
+
+void AV1HighbdTrans16x16HT::RunBitexactCheck() {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int i, j;
+  const int stride = 16;
+  const int num_tests = 1000;
+
+  for (i = 0; i < num_tests; ++i) {
+    for (j = 0; j < num_coeffs_; ++j) {
+      input_[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+    }
+
+    fwd_txfm_ref_(input_, output_ref_, stride, tx_type_, bit_depth_);
+    ASM_REGISTER_STATE_CHECK(
+        fwd_txfm_(input_, output_, stride, tx_type_, bit_depth_));
+
+    for (j = 0; j < num_coeffs_; ++j) {
+      EXPECT_EQ(output_ref_[j], output_[j])
+          << "Not bit-exact result at index: " << j << " at test block: " << i;
+    }
+  }
+}
+
+TEST_P(AV1HighbdTrans16x16HT, HighbdCoeffCheck) { RunBitexactCheck(); }
+#endif  // CONFIG_HIGHBITDEPTH
+
+using std::tr1::make_tuple;
+
+#if HAVE_SSE2
+const Ht16x16Param kArrayHt16x16Param_sse2[] = {
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 0, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 1, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 2, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 3, AOM_BITS_8,
+             256),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 4, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 5, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 6, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 7, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 8, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 9, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 10, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 11, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 12, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 13, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 14, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2, 15, AOM_BITS_8,
+             256)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE2, AV1Trans16x16HT,
+                        ::testing::ValuesIn(kArrayHt16x16Param_sse2));
+#endif  // HAVE_SSE2
+
+#if HAVE_AVX2
+const Ht16x16Param kArrayHt16x16Param_avx2[] = {
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 0, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 1, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 2, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 3, AOM_BITS_8,
+             256),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 4, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 5, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 6, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 7, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 8, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 9, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 10, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 11, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 12, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 13, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 14, AOM_BITS_8,
+             256),
+  make_tuple(&av1_fht16x16_avx2, &av1_iht16x16_256_add_avx2, 15, AOM_BITS_8,
+             256)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(AVX2, AV1Trans16x16HT,
+                        ::testing::ValuesIn(kArrayHt16x16Param_avx2));
+#endif  // HAVE_AVX2
+
+#if HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+const HighbdHt16x16Param kArrayHBDHt16x16Param_sse4_1[] = {
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 0, 10),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 0, 12),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 1, 10),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 1, 12),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 2, 10),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 2, 12),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 3, 10),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 3, 12),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 4, 10),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 4, 12),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 5, 10),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 5, 12),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 6, 10),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 6, 12),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 7, 10),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 7, 12),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 8, 10),
+  make_tuple(&av1_fwd_txfm2d_16x16_sse4_1, 8, 12),
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE4_1, AV1HighbdTrans16x16HT,
+                        ::testing::ValuesIn(kArrayHBDHt16x16Param_sse4_1));
+#endif  // HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+
+}  // namespace
diff --git a/third_party/aom/test/av1_fht16x32_test.cc b/third_party/aom/test/av1_fht16x32_test.cc
new file mode 100644
index 0000000000..43d0253273
--- /dev/null
+++ b/third_party/aom/test/av1_fht16x32_test.cc
@@ -0,0 +1,144 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+using std::tr1::tuple;
+using libaom_test::FhtFunc;
+typedef tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht16x32Param;
+
+void fht16x32_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht16x32_c(in, out, stride, tx_type);
+}
+
+void iht16x32_ref(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_iht16x32_512_add_c(in, out, stride, tx_type);
+}
+
+class AV1Trans16x32HT : public libaom_test::TransformTestBase,
+                        public ::testing::TestWithParam<Ht16x32Param> {
+ public:
+  virtual ~AV1Trans16x32HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 16;
+    height_ = 32;
+    fwd_txfm_ref = fht16x32_ref;
+    inv_txfm_ref = iht16x32_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(AV1Trans16x32HT, AccuracyCheck) { RunAccuracyCheck(4, 0.2); }
+TEST_P(AV1Trans16x32HT, CoeffCheck) { RunCoeffCheck(); }
+TEST_P(AV1Trans16x32HT, MemCheck) { RunMemCheck(); }
+TEST_P(AV1Trans16x32HT, InvCoeffCheck) { RunInvCoeffCheck(); }
+TEST_P(AV1Trans16x32HT, InvAccuracyCheck) { RunInvAccuracyCheck(4); }
+
+using std::tr1::make_tuple;
+const Ht16x32Param kArrayHt16x32Param_c[] = {
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 0, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 1, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 2, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 3, AOM_BITS_8, 512),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 4, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 5, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 6, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 7, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 8, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 9, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 10, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 11, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 12, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 13, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 14, AOM_BITS_8, 512),
+  make_tuple(&av1_fht16x32_c, &av1_iht16x32_512_add_c, 15, AOM_BITS_8, 512)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(C, AV1Trans16x32HT,
+                        ::testing::ValuesIn(kArrayHt16x32Param_c));
+
+#if HAVE_SSE2
+const Ht16x32Param kArrayHt16x32Param_sse2[] = {
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 0, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 1, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 2, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 3, AOM_BITS_8,
+             512),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 4, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 5, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 6, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 7, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 8, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 9, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 10, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 11, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 12, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 13, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 14, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht16x32_sse2, &av1_iht16x32_512_add_sse2, 15, AOM_BITS_8,
+             512)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE2, AV1Trans16x32HT,
+                        ::testing::ValuesIn(kArrayHt16x32Param_sse2));
+#endif  // HAVE_SSE2
+
+}  // namespace
diff --git a/third_party/aom/test/av1_fht16x8_test.cc b/third_party/aom/test/av1_fht16x8_test.cc
new file mode 100644
index 0000000000..d99bec5eb3
--- /dev/null
+++ b/third_party/aom/test/av1_fht16x8_test.cc
@@ -0,0 +1,129 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+using std::tr1::tuple;
+using libaom_test::FhtFunc;
+typedef tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht16x8Param;
+
+void fht16x8_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht16x8_c(in, out, stride, tx_type);
+}
+
+void iht16x8_ref(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_iht16x8_128_add_c(in, out, stride, tx_type);
+}
+
+class AV1Trans16x8HT : public libaom_test::TransformTestBase,
+                       public ::testing::TestWithParam<Ht16x8Param> {
+ public:
+  virtual ~AV1Trans16x8HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 16;
+    height_ = 8;
+    inv_txfm_ref = iht16x8_ref;
+    fwd_txfm_ref = fht16x8_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(AV1Trans16x8HT, AccuracyCheck) { RunAccuracyCheck(1, 0.001); }
+TEST_P(AV1Trans16x8HT, CoeffCheck) { RunCoeffCheck(); }
+TEST_P(AV1Trans16x8HT, MemCheck) { RunMemCheck(); }
+TEST_P(AV1Trans16x8HT, InvCoeffCheck) { RunInvCoeffCheck(); }
+TEST_P(AV1Trans16x8HT, InvAccuracyCheck) { RunInvAccuracyCheck(1); }
+
+using std::tr1::make_tuple;
+
+const Ht16x8Param kArrayHt16x8Param_c[] = {
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 0, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 1, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 2, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 3, AOM_BITS_8, 128),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 4, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 5, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 6, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 7, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 8, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 9, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 10, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 11, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 12, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 13, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 14, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_c, &av1_iht16x8_128_add_c, 15, AOM_BITS_8, 128)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(C, AV1Trans16x8HT,
+                        ::testing::ValuesIn(kArrayHt16x8Param_c));
+
+#if HAVE_SSE2
+const Ht16x8Param kArrayHt16x8Param_sse2[] = {
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 0, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 1, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 2, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 3, AOM_BITS_8, 128),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 4, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 5, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 6, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 7, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 8, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 9, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 10, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 11, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 12, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 13, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 14, AOM_BITS_8, 128),
+  make_tuple(&av1_fht16x8_sse2, &av1_iht16x8_128_add_sse2, 15, AOM_BITS_8, 128)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE2, AV1Trans16x8HT,
+                        ::testing::ValuesIn(kArrayHt16x8Param_sse2));
+#endif  // HAVE_SSE2
+
+}  // namespace
diff --git a/third_party/aom/test/av1_fht32x16_test.cc b/third_party/aom/test/av1_fht32x16_test.cc
new file mode 100644
index 0000000000..e38283f865
--- /dev/null
+++ b/third_party/aom/test/av1_fht32x16_test.cc
@@ -0,0 +1,144 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+using std::tr1::tuple;
+using libaom_test::FhtFunc;
+typedef tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht32x16Param;
+
+void fht32x16_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht32x16_c(in, out, stride, tx_type);
+}
+
+void iht32x16_ref(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_iht32x16_512_add_c(in, out, stride, tx_type);
+}
+
+class AV1Trans32x16HT : public libaom_test::TransformTestBase,
+                        public ::testing::TestWithParam<Ht32x16Param> {
+ public:
+  virtual ~AV1Trans32x16HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 32;
+    height_ = 16;
+    fwd_txfm_ref = fht32x16_ref;
+    inv_txfm_ref = iht32x16_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(AV1Trans32x16HT, MemCheck) { RunMemCheck(); }
+TEST_P(AV1Trans32x16HT, AccuracyCheck) { RunAccuracyCheck(4, 0.2); }
+TEST_P(AV1Trans32x16HT, CoeffCheck) { RunCoeffCheck(); }
+TEST_P(AV1Trans32x16HT, InvCoeffCheck) { RunInvCoeffCheck(); }
+TEST_P(AV1Trans32x16HT, InvAccuracyCheck) { RunInvAccuracyCheck(4); }
+
+using std::tr1::make_tuple;
+const Ht32x16Param kArrayHt32x16Param_c[] = {
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 0, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 1, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 2, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 3, AOM_BITS_8, 512),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 4, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 5, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 6, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 7, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 8, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 9, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 10, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 11, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 12, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 13, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 14, AOM_BITS_8, 512),
+  make_tuple(&av1_fht32x16_c, &av1_iht32x16_512_add_c, 15, AOM_BITS_8, 512)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(C, AV1Trans32x16HT,
+                        ::testing::ValuesIn(kArrayHt32x16Param_c));
+
+#if HAVE_SSE2
+const Ht32x16Param kArrayHt32x16Param_sse2[] = {
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 0, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 1, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 2, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 3, AOM_BITS_8,
+             512),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 4, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 5, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 6, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 7, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 8, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 9, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 10, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 11, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 12, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 13, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 14, AOM_BITS_8,
+             512),
+  make_tuple(&av1_fht32x16_sse2, &av1_iht32x16_512_add_sse2, 15, AOM_BITS_8,
+             512)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE2, AV1Trans32x16HT,
+                        ::testing::ValuesIn(kArrayHt32x16Param_sse2));
+#endif  // HAVE_SSE2
+
+}  // namespace
diff --git a/third_party/aom/test/av1_fht4x4_test.cc b/third_party/aom/test/av1_fht4x4_test.cc
new file mode 100644
index 0000000000..42837d3a4c
--- /dev/null
+++ b/third_party/aom/test/av1_fht4x4_test.cc
@@ -0,0 +1,222 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+#include "aom_ports/mem.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+using std::tr1::tuple;
+using libaom_test::FhtFunc;
+typedef tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht4x4Param;
+
+void fht4x4_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht4x4_c(in, out, stride, tx_type);
+}
+
+void iht4x4_ref(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_iht4x4_16_add_c(in, out, stride, tx_type);
+}
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*IhighbdHtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                              int tx_type, int bd);
+typedef void (*HBDFhtFunc)(const int16_t *input, int32_t *output, int stride,
+                           int tx_type, int bd);
+
+// HighbdHt4x4Param argument list:
+// <Target optimized function, tx_type, bit depth>
+typedef tuple<HBDFhtFunc, int, int> HighbdHt4x4Param;
+
+void highbe_fht4x4_ref(const int16_t *in, int32_t *out, int stride, int tx_type,
+                       int bd) {
+  av1_fwd_txfm2d_4x4_c(in, out, stride, tx_type, bd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+class AV1Trans4x4HT : public libaom_test::TransformTestBase,
+                      public ::testing::TestWithParam<Ht4x4Param> {
+ public:
+  virtual ~AV1Trans4x4HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 4;
+    height_ = 4;
+    fwd_txfm_ref = fht4x4_ref;
+    inv_txfm_ref = iht4x4_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(AV1Trans4x4HT, MemCheck) { RunMemCheck(); }
+TEST_P(AV1Trans4x4HT, CoeffCheck) { RunCoeffCheck(); }
+// Note:
+//  TODO(luoyi): Add tx_type, 9-15 for inverse transform.
+//  Need cleanup since same tests may be done in fdct4x4_test.cc
+// TEST_P(AV1Trans4x4HT, AccuracyCheck) { RunAccuracyCheck(0); }
+// TEST_P(AV1Trans4x4HT, InvAccuracyCheck) { RunInvAccuracyCheck(0); }
+// TEST_P(AV1Trans4x4HT, InvCoeffCheck) { RunInvCoeffCheck(); }
+
+#if CONFIG_HIGHBITDEPTH
+class AV1HighbdTrans4x4HT : public ::testing::TestWithParam<HighbdHt4x4Param> {
+ public:
+  virtual ~AV1HighbdTrans4x4HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    fwd_txfm_ref_ = highbe_fht4x4_ref;
+    tx_type_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = 16;
+
+    input_ = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(int16_t) * num_coeffs_));
+    output_ = reinterpret_cast<int32_t *>(
+        aom_memalign(16, sizeof(int32_t) * num_coeffs_));
+    output_ref_ = reinterpret_cast<int32_t *>(
+        aom_memalign(16, sizeof(int32_t) * num_coeffs_));
+  }
+
+  virtual void TearDown() {
+    aom_free(input_);
+    aom_free(output_);
+    aom_free(output_ref_);
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void RunBitexactCheck();
+
+ private:
+  HBDFhtFunc fwd_txfm_;
+  HBDFhtFunc fwd_txfm_ref_;
+  int tx_type_;
+  int bit_depth_;
+  int mask_;
+  int num_coeffs_;
+  int16_t *input_;
+  int32_t *output_;
+  int32_t *output_ref_;
+};
+
+void AV1HighbdTrans4x4HT::RunBitexactCheck() {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int i, j;
+  const int stride = 4;
+  const int num_tests = 1000;
+  const int num_coeffs = 16;
+
+  for (i = 0; i < num_tests; ++i) {
+    for (j = 0; j < num_coeffs; ++j) {
+      input_[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+    }
+
+    fwd_txfm_ref_(input_, output_ref_, stride, tx_type_, bit_depth_);
+    fwd_txfm_(input_, output_, stride, tx_type_, bit_depth_);
+
+    for (j = 0; j < num_coeffs; ++j) {
+      EXPECT_EQ(output_[j], output_ref_[j])
+          << "Not bit-exact result at index: " << j << " at test block: " << i;
+    }
+  }
+}
+
+TEST_P(AV1HighbdTrans4x4HT, HighbdCoeffCheck) { RunBitexactCheck(); }
+#endif  // CONFIG_HIGHBITDEPTH
+
+using std::tr1::make_tuple;
+
+#if HAVE_SSE2
+const Ht4x4Param kArrayHt4x4Param_sse2[] = {
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 0, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 1, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 2, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 3, AOM_BITS_8, 16),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 4, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 5, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 6, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 7, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 8, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 9, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 10, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 11, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 12, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 13, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 14, AOM_BITS_8, 16),
+  make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 15, AOM_BITS_8, 16)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE2, AV1Trans4x4HT,
+                        ::testing::ValuesIn(kArrayHt4x4Param_sse2));
+#endif  // HAVE_SSE2
+
+#if HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+const HighbdHt4x4Param kArrayHighbdHt4x4Param[] = {
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 0, 10),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 0, 12),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 1, 10),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 1, 12),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 2, 10),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 2, 12),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 3, 10),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 3, 12),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 4, 10),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 4, 12),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 5, 10),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 5, 12),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 6, 10),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 6, 12),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 7, 10),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 7, 12),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 8, 10),
+  make_tuple(&av1_fwd_txfm2d_4x4_sse4_1, 8, 12),
+#endif  // CONFIG_EXT_TX
+};
+
+INSTANTIATE_TEST_CASE_P(SSE4_1, AV1HighbdTrans4x4HT,
+                        ::testing::ValuesIn(kArrayHighbdHt4x4Param));
+
+#endif  // HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+
+}  // namespace
diff --git a/third_party/aom/test/av1_fht4x8_test.cc b/third_party/aom/test/av1_fht4x8_test.cc
new file mode 100644
index 0000000000..a899c87391
--- /dev/null
+++ b/third_party/aom/test/av1_fht4x8_test.cc
@@ -0,0 +1,129 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+using std::tr1::tuple;
+using libaom_test::FhtFunc;
+typedef tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht4x8Param;
+
+void fht4x8_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht4x8_c(in, out, stride, tx_type);
+}
+
+void iht4x8_ref(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_iht4x8_32_add_c(in, out, stride, tx_type);
+}
+
+class AV1Trans4x8HT : public libaom_test::TransformTestBase,
+                      public ::testing::TestWithParam<Ht4x8Param> {
+ public:
+  virtual ~AV1Trans4x8HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 4;
+    height_ = 8;
+    fwd_txfm_ref = fht4x8_ref;
+    inv_txfm_ref = iht4x8_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(AV1Trans4x8HT, AccuracyCheck) { RunAccuracyCheck(0, 0.00001); }
+TEST_P(AV1Trans4x8HT, CoeffCheck) { RunCoeffCheck(); }
+TEST_P(AV1Trans4x8HT, MemCheck) { RunMemCheck(); }
+TEST_P(AV1Trans4x8HT, InvCoeffCheck) { RunInvCoeffCheck(); }
+TEST_P(AV1Trans4x8HT, InvAccuracyCheck) { RunInvAccuracyCheck(0); }
+
+using std::tr1::make_tuple;
+
+const Ht4x8Param kArrayHt4x8Param_c[] = {
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 0, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 1, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 2, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 3, AOM_BITS_8, 32),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 4, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 5, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 6, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 7, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 8, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 9, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 10, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 11, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 12, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 13, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 14, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_c, &av1_iht4x8_32_add_c, 15, AOM_BITS_8, 32)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(C, AV1Trans4x8HT,
+                        ::testing::ValuesIn(kArrayHt4x8Param_c));
+
+#if HAVE_SSE2
+const Ht4x8Param kArrayHt4x8Param_sse2[] = {
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 0, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 1, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 2, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 3, AOM_BITS_8, 32),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 4, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 5, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 6, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 7, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 8, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 9, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 10, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 11, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 12, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 13, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 14, AOM_BITS_8, 32),
+  make_tuple(&av1_fht4x8_sse2, &av1_iht4x8_32_add_sse2, 15, AOM_BITS_8, 32)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE2, AV1Trans4x8HT,
+                        ::testing::ValuesIn(kArrayHt4x8Param_sse2));
+#endif  // HAVE_SSE2
+
+}  // namespace
diff --git a/third_party/aom/test/av1_fht8x16_test.cc b/third_party/aom/test/av1_fht8x16_test.cc
new file mode 100644
index 0000000000..ace9a8f47d
--- /dev/null
+++ b/third_party/aom/test/av1_fht8x16_test.cc
@@ -0,0 +1,128 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+using std::tr1::tuple;
+using libaom_test::FhtFunc;
+typedef tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht8x16Param;
+
+void fht8x16_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht8x16_c(in, out, stride, tx_type);
+}
+
+void iht8x16_ref(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_iht8x16_128_add_c(in, out, stride, tx_type);
+}
+
+class AV1Trans8x16HT : public libaom_test::TransformTestBase,
+                       public ::testing::TestWithParam<Ht8x16Param> {
+ public:
+  virtual ~AV1Trans8x16HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 8;
+    height_ = 16;
+    inv_txfm_ref = iht8x16_ref;
+    fwd_txfm_ref = fht8x16_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(AV1Trans8x16HT, AccuracyCheck) { RunAccuracyCheck(1, 0.001); }
+TEST_P(AV1Trans8x16HT, MemCheck) { RunMemCheck(); }
+TEST_P(AV1Trans8x16HT, CoeffCheck) { RunCoeffCheck(); }
+TEST_P(AV1Trans8x16HT, InvCoeffCheck) { RunInvCoeffCheck(); }
+TEST_P(AV1Trans8x16HT, InvAccuracyCheck) { RunInvAccuracyCheck(1); }
+
+using std::tr1::make_tuple;
+
+const Ht8x16Param kArrayHt8x16Param_c[] = {
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 0, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 1, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 2, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 3, AOM_BITS_8, 128),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 4, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 5, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 6, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 7, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 8, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 9, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 10, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 11, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 12, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 13, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 14, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_c, &av1_iht8x16_128_add_c, 15, AOM_BITS_8, 128)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(C, AV1Trans8x16HT,
+                        ::testing::ValuesIn(kArrayHt8x16Param_c));
+
+#if HAVE_SSE2
+const Ht8x16Param kArrayHt8x16Param_sse2[] = {
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 0, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 1, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 2, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 3, AOM_BITS_8, 128),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 4, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 5, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 6, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 7, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 8, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 9, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 10, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 11, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 12, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 13, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 14, AOM_BITS_8, 128),
+  make_tuple(&av1_fht8x16_sse2, &av1_iht8x16_128_add_sse2, 15, AOM_BITS_8, 128)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE2, AV1Trans8x16HT,
+                        ::testing::ValuesIn(kArrayHt8x16Param_sse2));
+#endif  // HAVE_SSE2
+
+}  // namespace
diff --git a/third_party/aom/test/av1_fht8x4_test.cc b/third_party/aom/test/av1_fht8x4_test.cc
new file mode 100644
index 0000000000..9bf4ff6475
--- /dev/null
+++ b/third_party/aom/test/av1_fht8x4_test.cc
@@ -0,0 +1,128 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+using std::tr1::tuple;
+using libaom_test::FhtFunc;
+typedef tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht8x4Param;
+
+void fht8x4_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht8x4_c(in, out, stride, tx_type);
+}
+
+void iht8x4_ref(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_iht8x4_32_add_c(in, out, stride, tx_type);
+}
+
+class AV1Trans8x4HT : public libaom_test::TransformTestBase,
+                      public ::testing::TestWithParam<Ht8x4Param> {
+ public:
+  virtual ~AV1Trans8x4HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 8;
+    height_ = 4;
+    fwd_txfm_ref = fht8x4_ref;
+    inv_txfm_ref = iht8x4_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(AV1Trans8x4HT, AccuracyCheck) { RunAccuracyCheck(0, 0.00001); }
+TEST_P(AV1Trans8x4HT, CoeffCheck) { RunCoeffCheck(); }
+TEST_P(AV1Trans8x4HT, MemCheck) { RunMemCheck(); }
+TEST_P(AV1Trans8x4HT, InvCoeffCheck) { RunInvCoeffCheck(); }
+TEST_P(AV1Trans8x4HT, InvAccuracyCheck) { RunInvAccuracyCheck(0); }
+
+using std::tr1::make_tuple;
+
+const Ht8x4Param kArrayHt8x4Param_c[] = {
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 0, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 1, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 2, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 3, AOM_BITS_8, 32),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 4, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 5, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 6, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 7, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 8, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 9, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 10, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 11, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 12, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 13, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 14, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_c, &av1_iht8x4_32_add_c, 15, AOM_BITS_8, 32)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(C, AV1Trans8x4HT,
+                        ::testing::ValuesIn(kArrayHt8x4Param_c));
+
+#if HAVE_SSE2
+const Ht8x4Param kArrayHt8x4Param_sse2[] = {
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 0, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 1, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 2, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 3, AOM_BITS_8, 32),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 4, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 5, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 6, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 7, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 8, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 9, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 10, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 11, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 12, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 13, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 14, AOM_BITS_8, 32),
+  make_tuple(&av1_fht8x4_sse2, &av1_iht8x4_32_add_sse2, 15, AOM_BITS_8, 32)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE2, AV1Trans8x4HT,
+                        ::testing::ValuesIn(kArrayHt8x4Param_sse2));
+#endif  // HAVE_SSE2
+
+}  // namespace
diff --git a/third_party/aom/test/av1_fht8x8_test.cc b/third_party/aom/test/av1_fht8x8_test.cc
new file mode 100644
index 0000000000..99cff1014f
--- /dev/null
+++ b/third_party/aom/test/av1_fht8x8_test.cc
@@ -0,0 +1,220 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+#include "aom_ports/mem.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+
+using libaom_test::FhtFunc;
+using std::tr1::tuple;
+typedef tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht8x8Param;
+
+void fht8x8_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht8x8_c(in, out, stride, tx_type);
+}
+
+void iht8x8_ref(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_iht8x8_64_add_c(in, out, stride, tx_type);
+}
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*IHbdHtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                           int tx_type, int bd);
+typedef void (*HbdHtFunc)(const int16_t *input, int32_t *output, int stride,
+                          int tx_type, int bd);
+// Target optimized function, tx_type, bit depth
+typedef tuple<HbdHtFunc, int, int> HighbdHt8x8Param;
+
+void highbd_fht8x8_ref(const int16_t *in, int32_t *out, int stride, int tx_type,
+                       int bd) {
+  av1_fwd_txfm2d_8x8_c(in, out, stride, tx_type, bd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+class AV1Trans8x8HT : public libaom_test::TransformTestBase,
+                      public ::testing::TestWithParam<Ht8x8Param> {
+ public:
+  virtual ~AV1Trans8x8HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 8;
+    height_ = 8;
+    fwd_txfm_ref = fht8x8_ref;
+    inv_txfm_ref = iht8x8_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(AV1Trans8x8HT, MemCheck) { RunMemCheck(); }
+TEST_P(AV1Trans8x8HT, CoeffCheck) { RunCoeffCheck(); }
+// Note:
+//  TODO(luoyi): Add tx_type, 9-15 for inverse transform.
+//  Need cleanup since same tests may be done in fdct8x8_test.cc
+// TEST_P(AV1Trans8x8HT, AccuracyCheck) { RunAccuracyCheck(0); }
+// TEST_P(AV1Trans8x8HT, InvAccuracyCheck) { RunInvAccuracyCheck(0); }
+// TEST_P(AV1Trans8x8HT, InvCoeffCheck) { RunInvCoeffCheck(); }
+
+#if CONFIG_HIGHBITDEPTH
+class AV1HighbdTrans8x8HT : public ::testing::TestWithParam<HighbdHt8x8Param> {
+ public:
+  virtual ~AV1HighbdTrans8x8HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    fwd_txfm_ref_ = highbd_fht8x8_ref;
+    tx_type_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = 64;
+
+    input_ = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(int16_t) * num_coeffs_));
+    output_ = reinterpret_cast<int32_t *>(
+        aom_memalign(16, sizeof(int32_t) * num_coeffs_));
+    output_ref_ = reinterpret_cast<int32_t *>(
+        aom_memalign(16, sizeof(int32_t) * num_coeffs_));
+  }
+
+  virtual void TearDown() {
+    aom_free(input_);
+    aom_free(output_);
+    aom_free(output_ref_);
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void RunBitexactCheck();
+
+ private:
+  HbdHtFunc fwd_txfm_;
+  HbdHtFunc fwd_txfm_ref_;
+  int tx_type_;
+  int bit_depth_;
+  int mask_;
+  int num_coeffs_;
+  int16_t *input_;
+  int32_t *output_;
+  int32_t *output_ref_;
+};
+
+void AV1HighbdTrans8x8HT::RunBitexactCheck() {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int i, j;
+  const int stride = 8;
+  const int num_tests = 1000;
+  const int num_coeffs = 64;
+
+  for (i = 0; i < num_tests; ++i) {
+    for (j = 0; j < num_coeffs; ++j) {
+      input_[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+    }
+
+    fwd_txfm_ref_(input_, output_ref_, stride, tx_type_, bit_depth_);
+    ASM_REGISTER_STATE_CHECK(
+        fwd_txfm_(input_, output_, stride, tx_type_, bit_depth_));
+
+    for (j = 0; j < num_coeffs; ++j) {
+      EXPECT_EQ(output_ref_[j], output_[j])
+          << "Not bit-exact result at index: " << j << " at test block: " << i;
+    }
+  }
+}
+
+TEST_P(AV1HighbdTrans8x8HT, HighbdCoeffCheck) { RunBitexactCheck(); }
+#endif  // CONFIG_HIGHBITDEPTH
+
+using std::tr1::make_tuple;
+
+#if HAVE_SSE2
+const Ht8x8Param kArrayHt8x8Param_sse2[] = {
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 0, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 1, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 2, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 3, AOM_BITS_8, 64),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 4, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 5, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 6, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 7, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 8, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 9, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 10, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 11, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 12, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 13, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 14, AOM_BITS_8, 64),
+  make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 15, AOM_BITS_8, 64)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE2, AV1Trans8x8HT,
+                        ::testing::ValuesIn(kArrayHt8x8Param_sse2));
+#endif  // HAVE_SSE2
+
+#if HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+const HighbdHt8x8Param kArrayHBDHt8x8Param_sse4_1[] = {
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 0, 10),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 0, 12),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 1, 10),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 1, 12),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 2, 10),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 2, 12),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 3, 10),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 3, 12),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 4, 10),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 4, 12),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 5, 10),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 5, 12),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 6, 10),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 6, 12),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 7, 10),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 7, 12),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 8, 10),
+  make_tuple(&av1_fwd_txfm2d_8x8_sse4_1, 8, 12),
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE4_1, AV1HighbdTrans8x8HT,
+                        ::testing::ValuesIn(kArrayHBDHt8x8Param_sse4_1));
+#endif  // HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+
+}  // namespace
diff --git a/third_party/aom/test/av1_fwd_txfm1d_test.cc b/third_party/aom/test/av1_fwd_txfm1d_test.cc
new file mode 100644
index 0000000000..a9b3f8e406
--- /dev/null
+++ b/third_party/aom/test/av1_fwd_txfm1d_test.cc
@@ -0,0 +1,133 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/av1_fwd_txfm1d.h"
+#include "test/av1_txfm_test.h"
+
+using libaom_test::ACMRandom;
+using libaom_test::input_base;
+using libaom_test::reference_hybrid_1d;
+using libaom_test::TYPE_TXFM;
+using libaom_test::TYPE_DCT;
+using libaom_test::TYPE_ADST;
+
+namespace {
+const int txfm_type_num = 2;
+const TYPE_TXFM txfm_type_ls[2] = { TYPE_DCT, TYPE_ADST };
+
+const int txfm_size_num = 5;
+const int txfm_size_ls[5] = { 4, 8, 16, 32, 64 };
+
+const TxfmFunc fwd_txfm_func_ls[2][5] = {
+#if CONFIG_TX64X64
+  { av1_fdct4_new, av1_fdct8_new, av1_fdct16_new, av1_fdct32_new,
+    av1_fdct64_new },
+#else
+  { av1_fdct4_new, av1_fdct8_new, av1_fdct16_new, av1_fdct32_new, NULL },
+#endif
+  { av1_fadst4_new, av1_fadst8_new, av1_fadst16_new, av1_fadst32_new, NULL }
+};
+
+// the maximum stage number of fwd/inv 1d dct/adst txfm is 12
+const int8_t cos_bit[12] = { 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14 };
+const int8_t range_bit[12] = { 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32 };
+
+TEST(av1_fwd_txfm1d, round_shift) {
+  EXPECT_EQ(round_shift(7, 1), 4);
+  EXPECT_EQ(round_shift(-7, 1), -3);
+
+  EXPECT_EQ(round_shift(7, 2), 2);
+  EXPECT_EQ(round_shift(-7, 2), -2);
+
+  EXPECT_EQ(round_shift(8, 2), 2);
+  EXPECT_EQ(round_shift(-8, 2), -2);
+}
+
+TEST(av1_fwd_txfm1d, get_max_bit) {
+  int max_bit = get_max_bit(8);
+  EXPECT_EQ(max_bit, 3);
+}
+
+TEST(av1_fwd_txfm1d, cospi_arr) {
+  for (int i = 0; i < 7; i++) {
+    for (int j = 0; j < 64; j++) {
+      EXPECT_EQ(cospi_arr[i][j],
+                (int32_t)round(cos(M_PI * j / 128) * (1 << (cos_bit_min + i))));
+    }
+  }
+}
+
+TEST(av1_fwd_txfm1d, clamp_block) {
+  int16_t block[5][5] = { { 7, -5, 6, -3, 9 },
+                          { 7, -5, 6, -3, 9 },
+                          { 7, -5, 6, -3, 9 },
+                          { 7, -5, 6, -3, 9 },
+                          { 7, -5, 6, -3, 9 } };
+
+  int16_t ref_block[5][5] = { { 7, -5, 6, -3, 9 },
+                              { 7, -5, 6, -3, 9 },
+                              { 7, -4, 2, -3, 9 },
+                              { 7, -4, 2, -3, 9 },
+                              { 7, -4, 2, -3, 9 } };
+
+  int row = 2;
+  int col = 1;
+  int block_size = 3;
+  int stride = 5;
+  clamp_block(block[row] + col, block_size, stride, -4, 2);
+  for (int r = 0; r < stride; r++) {
+    for (int c = 0; c < stride; c++) {
+      EXPECT_EQ(block[r][c], ref_block[r][c]);
+    }
+  }
+}
+
+TEST(av1_fwd_txfm1d, accuracy) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  for (int si = 0; si < txfm_size_num; ++si) {
+    int txfm_size = txfm_size_ls[si];
+    int32_t *input = new int32_t[txfm_size];
+    int32_t *output = new int32_t[txfm_size];
+    double *ref_input = new double[txfm_size];
+    double *ref_output = new double[txfm_size];
+
+    for (int ti = 0; ti < txfm_type_num; ++ti) {
+      TYPE_TXFM txfm_type = txfm_type_ls[ti];
+      TxfmFunc fwd_txfm_func = fwd_txfm_func_ls[ti][si];
+      int max_error = 7;
+
+      const int count_test_block = 5000;
+      if (fwd_txfm_func != NULL) {
+        for (int ti = 0; ti < count_test_block; ++ti) {
+          for (int ni = 0; ni < txfm_size; ++ni) {
+            input[ni] = rnd.Rand16() % input_base - rnd.Rand16() % input_base;
+            ref_input[ni] = static_cast<double>(input[ni]);
+          }
+
+          fwd_txfm_func(input, output, cos_bit, range_bit);
+          reference_hybrid_1d(ref_input, ref_output, txfm_size, txfm_type);
+
+          for (int ni = 0; ni < txfm_size; ++ni) {
+            EXPECT_LE(
+                abs(output[ni] - static_cast<int32_t>(round(ref_output[ni]))),
+                max_error);
+          }
+        }
+      }
+    }
+
+    delete[] input;
+    delete[] output;
+    delete[] ref_input;
+    delete[] ref_output;
+  }
+}
+}  // namespace
diff --git a/third_party/aom/test/av1_fwd_txfm2d_test.cc b/third_party/aom/test/av1_fwd_txfm2d_test.cc
new file mode 100644
index 0000000000..25cf5ad53b
--- /dev/null
+++ b/third_party/aom/test/av1_fwd_txfm2d_test.cc
@@ -0,0 +1,179 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "test/acm_random.h"
+#include "test/util.h"
+#include "test/av1_txfm_test.h"
+#include "av1/common/av1_txfm.h"
+#include "./av1_rtcd.h"
+
+using libaom_test::ACMRandom;
+using libaom_test::input_base;
+using libaom_test::bd;
+using libaom_test::compute_avg_abs_error;
+using libaom_test::Fwd_Txfm2d_Func;
+using libaom_test::TYPE_TXFM;
+
+namespace {
+#if CONFIG_HIGHBITDEPTH
+// tx_type_, tx_size_, max_error_, max_avg_error_
+typedef std::tr1::tuple<TX_TYPE, TX_SIZE, double, double> AV1FwdTxfm2dParam;
+
+class AV1FwdTxfm2d : public ::testing::TestWithParam<AV1FwdTxfm2dParam> {
+ public:
+  virtual void SetUp() {
+    tx_type_ = GET_PARAM(0);
+    tx_size_ = GET_PARAM(1);
+    max_error_ = GET_PARAM(2);
+    max_avg_error_ = GET_PARAM(3);
+    count_ = 500;
+    TXFM_2D_FLIP_CFG fwd_txfm_flip_cfg =
+        av1_get_fwd_txfm_cfg(tx_type_, tx_size_);
+    const TXFM_2D_CFG *fwd_txfm_cfg = fwd_txfm_flip_cfg.cfg;
+    int amplify_bit = fwd_txfm_cfg->shift[0] + fwd_txfm_cfg->shift[1] +
+                      fwd_txfm_cfg->shift[2];
+    ud_flip_ = fwd_txfm_flip_cfg.ud_flip;
+    lr_flip_ = fwd_txfm_flip_cfg.lr_flip;
+    amplify_factor_ =
+        amplify_bit >= 0 ? (1 << amplify_bit) : (1.0 / (1 << -amplify_bit));
+
+    fwd_txfm_ = libaom_test::fwd_txfm_func_ls[tx_size_];
+    txfm1d_size_ = libaom_test::get_txfm1d_size(tx_size_);
+    txfm2d_size_ = txfm1d_size_ * txfm1d_size_;
+    get_txfm1d_type(tx_type_, &type0_, &type1_);
+    input_ = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(input_[0]) * txfm2d_size_));
+    output_ = reinterpret_cast<int32_t *>(
+        aom_memalign(16, sizeof(output_[0]) * txfm2d_size_));
+    ref_input_ = reinterpret_cast<double *>(
+        aom_memalign(16, sizeof(ref_input_[0]) * txfm2d_size_));
+    ref_output_ = reinterpret_cast<double *>(
+        aom_memalign(16, sizeof(ref_output_[0]) * txfm2d_size_));
+  }
+
+  void RunFwdAccuracyCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    double avg_abs_error = 0;
+    for (int ci = 0; ci < count_; ci++) {
+      for (int ni = 0; ni < txfm2d_size_; ++ni) {
+        input_[ni] = rnd.Rand16() % input_base;
+        ref_input_[ni] = static_cast<double>(input_[ni]);
+        output_[ni] = 0;
+        ref_output_[ni] = 0;
+      }
+
+      fwd_txfm_(input_, output_, txfm1d_size_, tx_type_, bd);
+
+      if (lr_flip_ && ud_flip_)
+        libaom_test::fliplrud(ref_input_, txfm1d_size_, txfm1d_size_);
+      else if (lr_flip_)
+        libaom_test::fliplr(ref_input_, txfm1d_size_, txfm1d_size_);
+      else if (ud_flip_)
+        libaom_test::flipud(ref_input_, txfm1d_size_, txfm1d_size_);
+
+      reference_hybrid_2d(ref_input_, ref_output_, txfm1d_size_, type0_,
+                          type1_);
+
+      for (int ni = 0; ni < txfm2d_size_; ++ni) {
+        ref_output_[ni] = round(ref_output_[ni] * amplify_factor_);
+        EXPECT_GE(max_error_,
+                  fabs(output_[ni] - ref_output_[ni]) / amplify_factor_);
+      }
+      avg_abs_error += compute_avg_abs_error<int32_t, double>(
+          output_, ref_output_, txfm2d_size_);
+    }
+
+    avg_abs_error /= amplify_factor_;
+    avg_abs_error /= count_;
+    // max_abs_avg_error comes from upper bound of avg_abs_error
+    // printf("type0: %d type1: %d txfm_size: %d accuracy_avg_abs_error:
+    // %f\n", type0_, type1_, txfm1d_size_, avg_abs_error);
+    EXPECT_GE(max_avg_error_, avg_abs_error);
+  }
+
+  virtual void TearDown() {
+    aom_free(input_);
+    aom_free(output_);
+    aom_free(ref_input_);
+    aom_free(ref_output_);
+  }
+
+ private:
+  double max_error_;
+  double max_avg_error_;
+  int count_;
+  double amplify_factor_;
+  TX_TYPE tx_type_;
+  TX_SIZE tx_size_;
+  int txfm1d_size_;
+  int txfm2d_size_;
+  Fwd_Txfm2d_Func fwd_txfm_;
+  TYPE_TXFM type0_;
+  TYPE_TXFM type1_;
+  int16_t *input_;
+  int32_t *output_;
+  double *ref_input_;
+  double *ref_output_;
+  int ud_flip_;  // flip upside down
+  int lr_flip_;  // flip left to right
+};
+
+TEST_P(AV1FwdTxfm2d, RunFwdAccuracyCheck) { RunFwdAccuracyCheck(); }
+const AV1FwdTxfm2dParam av1_fwd_txfm2d_param_c[] = {
+#if CONFIG_EXT_TX
+  AV1FwdTxfm2dParam(FLIPADST_DCT, TX_4X4, 2, 0.2),
+  AV1FwdTxfm2dParam(DCT_FLIPADST, TX_4X4, 2, 0.2),
+  AV1FwdTxfm2dParam(FLIPADST_FLIPADST, TX_4X4, 2, 0.2),
+  AV1FwdTxfm2dParam(ADST_FLIPADST, TX_4X4, 2, 0.2),
+  AV1FwdTxfm2dParam(FLIPADST_ADST, TX_4X4, 2, 0.2),
+  AV1FwdTxfm2dParam(FLIPADST_DCT, TX_8X8, 5, 0.6),
+  AV1FwdTxfm2dParam(DCT_FLIPADST, TX_8X8, 5, 0.6),
+  AV1FwdTxfm2dParam(FLIPADST_FLIPADST, TX_8X8, 5, 0.6),
+  AV1FwdTxfm2dParam(ADST_FLIPADST, TX_8X8, 5, 0.6),
+  AV1FwdTxfm2dParam(FLIPADST_ADST, TX_8X8, 5, 0.6),
+  AV1FwdTxfm2dParam(FLIPADST_DCT, TX_16X16, 11, 1.5),
+  AV1FwdTxfm2dParam(DCT_FLIPADST, TX_16X16, 11, 1.5),
+  AV1FwdTxfm2dParam(FLIPADST_FLIPADST, TX_16X16, 11, 1.5),
+  AV1FwdTxfm2dParam(ADST_FLIPADST, TX_16X16, 11, 1.5),
+  AV1FwdTxfm2dParam(FLIPADST_ADST, TX_16X16, 11, 1.5),
+  AV1FwdTxfm2dParam(FLIPADST_DCT, TX_32X32, 70, 7),
+  AV1FwdTxfm2dParam(DCT_FLIPADST, TX_32X32, 70, 7),
+  AV1FwdTxfm2dParam(FLIPADST_FLIPADST, TX_32X32, 70, 7),
+  AV1FwdTxfm2dParam(ADST_FLIPADST, TX_32X32, 70, 7),
+  AV1FwdTxfm2dParam(FLIPADST_ADST, TX_32X32, 70, 7),
+#endif
+  AV1FwdTxfm2dParam(DCT_DCT, TX_4X4, 2, 0.2),
+  AV1FwdTxfm2dParam(ADST_DCT, TX_4X4, 2, 0.2),
+  AV1FwdTxfm2dParam(DCT_ADST, TX_4X4, 2, 0.2),
+  AV1FwdTxfm2dParam(ADST_ADST, TX_4X4, 2, 0.2),
+  AV1FwdTxfm2dParam(DCT_DCT, TX_8X8, 5, 0.6),
+  AV1FwdTxfm2dParam(ADST_DCT, TX_8X8, 5, 0.6),
+  AV1FwdTxfm2dParam(DCT_ADST, TX_8X8, 5, 0.6),
+  AV1FwdTxfm2dParam(ADST_ADST, TX_8X8, 5, 0.6),
+  AV1FwdTxfm2dParam(DCT_DCT, TX_16X16, 11, 1.5),
+  AV1FwdTxfm2dParam(ADST_DCT, TX_16X16, 11, 1.5),
+  AV1FwdTxfm2dParam(DCT_ADST, TX_16X16, 11, 1.5),
+  AV1FwdTxfm2dParam(ADST_ADST, TX_16X16, 11, 1.5),
+  AV1FwdTxfm2dParam(DCT_DCT, TX_32X32, 70, 7),
+  AV1FwdTxfm2dParam(ADST_DCT, TX_32X32, 70, 7),
+  AV1FwdTxfm2dParam(DCT_ADST, TX_32X32, 70, 7),
+  AV1FwdTxfm2dParam(ADST_ADST, TX_32X32, 70, 7)
+};
+
+INSTANTIATE_TEST_CASE_P(C, AV1FwdTxfm2d,
+                        ::testing::ValuesIn(av1_fwd_txfm2d_param_c));
+
+#endif  // CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/av1_highbd_iht_test.cc b/third_party/aom/test/av1_highbd_iht_test.cc
new file mode 100644
index 0000000000..3b263638f8
--- /dev/null
+++ b/third_party/aom/test/av1_highbd_iht_test.cc
@@ -0,0 +1,236 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "av1/common/enums.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_ports/mem.h"
+
+namespace {
+
+using std::tr1::tuple;
+using libaom_test::ACMRandom;
+
+typedef void (*HbdHtFunc)(const int16_t *input, int32_t *output, int stride,
+                          int tx_type, int bd);
+
+typedef void (*IHbdHtFunc)(const int32_t *coeff, uint16_t *output, int stride,
+                           int tx_type, int bd);
+
+// Test parameter argument list:
+//   <transform reference function,
+//    optimized inverse transform function,
+//    inverse transform reference function,
+//    num_coeffs,
+//    tx_type,
+//    bit_depth>
+typedef tuple<HbdHtFunc, IHbdHtFunc, IHbdHtFunc, int, int, int> IHbdHtParam;
+
+class AV1HighbdInvHTNxN : public ::testing::TestWithParam<IHbdHtParam> {
+ public:
+  virtual ~AV1HighbdInvHTNxN() {}
+
+  virtual void SetUp() {
+    txfm_ref_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    inv_txfm_ref_ = GET_PARAM(2);
+    num_coeffs_ = GET_PARAM(3);
+    tx_type_ = GET_PARAM(4);
+    bit_depth_ = GET_PARAM(5);
+
+    input_ = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(input_[0]) * num_coeffs_));
+
+    // Note:
+    // Inverse transform input buffer is 32-byte aligned
+    // Refer to <root>/av1/encoder/context_tree.c, function,
+    // void alloc_mode_context().
+    coeffs_ = reinterpret_cast<int32_t *>(
+        aom_memalign(32, sizeof(coeffs_[0]) * num_coeffs_));
+    output_ = reinterpret_cast<uint16_t *>(
+        aom_memalign(32, sizeof(output_[0]) * num_coeffs_));
+    output_ref_ = reinterpret_cast<uint16_t *>(
+        aom_memalign(32, sizeof(output_ref_[0]) * num_coeffs_));
+  }
+
+  virtual void TearDown() {
+    aom_free(input_);
+    aom_free(coeffs_);
+    aom_free(output_);
+    aom_free(output_ref_);
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void RunBitexactCheck();
+
+ private:
+  int GetStride() const {
+    if (16 == num_coeffs_) {
+      return 4;
+    } else if (64 == num_coeffs_) {
+      return 8;
+    } else if (256 == num_coeffs_) {
+      return 16;
+    } else if (1024 == num_coeffs_) {
+      return 32;
+    } else {
+      return 0;
+    }
+  }
+
+  HbdHtFunc txfm_ref_;
+  IHbdHtFunc inv_txfm_;
+  IHbdHtFunc inv_txfm_ref_;
+  int num_coeffs_;
+  int tx_type_;
+  int bit_depth_;
+
+  int16_t *input_;
+  int32_t *coeffs_;
+  uint16_t *output_;
+  uint16_t *output_ref_;
+};
+
+void AV1HighbdInvHTNxN::RunBitexactCheck() {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int stride = GetStride();
+  const int num_tests = 20000;
+  const uint16_t mask = (1 << bit_depth_) - 1;
+
+  for (int i = 0; i < num_tests; ++i) {
+    for (int j = 0; j < num_coeffs_; ++j) {
+      input_[j] = (rnd.Rand16() & mask) - (rnd.Rand16() & mask);
+      output_ref_[j] = rnd.Rand16() & mask;
+      output_[j] = output_ref_[j];
+    }
+
+    txfm_ref_(input_, coeffs_, stride, tx_type_, bit_depth_);
+    inv_txfm_ref_(coeffs_, output_ref_, stride, tx_type_, bit_depth_);
+    ASM_REGISTER_STATE_CHECK(
+        inv_txfm_(coeffs_, output_, stride, tx_type_, bit_depth_));
+
+    for (int j = 0; j < num_coeffs_; ++j) {
+      EXPECT_EQ(output_ref_[j], output_[j])
+          << "Not bit-exact result at index: " << j << " At test block: " << i;
+    }
+  }
+}
+
+TEST_P(AV1HighbdInvHTNxN, InvTransResultCheck) { RunBitexactCheck(); }
+
+using std::tr1::make_tuple;
+
+#if HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+#define PARAM_LIST_4X4                                   \
+  &av1_fwd_txfm2d_4x4_c, &av1_inv_txfm2d_add_4x4_sse4_1, \
+      &av1_inv_txfm2d_add_4x4_c, 16
+
+#define PARAM_LIST_8X8                                   \
+  &av1_fwd_txfm2d_8x8_c, &av1_inv_txfm2d_add_8x8_sse4_1, \
+      &av1_inv_txfm2d_add_8x8_c, 64
+
+#define PARAM_LIST_16X16                                     \
+  &av1_fwd_txfm2d_16x16_c, &av1_inv_txfm2d_add_16x16_sse4_1, \
+      &av1_inv_txfm2d_add_16x16_c, 256
+
+const IHbdHtParam kArrayIhtParam[] = {
+  // 16x16
+  make_tuple(PARAM_LIST_16X16, DCT_DCT, 10),
+  make_tuple(PARAM_LIST_16X16, DCT_DCT, 12),
+  make_tuple(PARAM_LIST_16X16, ADST_DCT, 10),
+  make_tuple(PARAM_LIST_16X16, ADST_DCT, 12),
+  make_tuple(PARAM_LIST_16X16, DCT_ADST, 10),
+  make_tuple(PARAM_LIST_16X16, DCT_ADST, 12),
+  make_tuple(PARAM_LIST_16X16, ADST_ADST, 10),
+  make_tuple(PARAM_LIST_16X16, ADST_ADST, 12),
+#if CONFIG_EXT_TX
+  make_tuple(PARAM_LIST_16X16, FLIPADST_DCT, 10),
+  make_tuple(PARAM_LIST_16X16, FLIPADST_DCT, 12),
+  make_tuple(PARAM_LIST_16X16, DCT_FLIPADST, 10),
+  make_tuple(PARAM_LIST_16X16, DCT_FLIPADST, 12),
+  make_tuple(PARAM_LIST_16X16, FLIPADST_FLIPADST, 10),
+  make_tuple(PARAM_LIST_16X16, FLIPADST_FLIPADST, 12),
+  make_tuple(PARAM_LIST_16X16, ADST_FLIPADST, 10),
+  make_tuple(PARAM_LIST_16X16, ADST_FLIPADST, 12),
+  make_tuple(PARAM_LIST_16X16, FLIPADST_ADST, 10),
+  make_tuple(PARAM_LIST_16X16, FLIPADST_ADST, 12),
+#endif
+  // 8x8
+  make_tuple(PARAM_LIST_8X8, DCT_DCT, 10),
+  make_tuple(PARAM_LIST_8X8, DCT_DCT, 12),
+  make_tuple(PARAM_LIST_8X8, ADST_DCT, 10),
+  make_tuple(PARAM_LIST_8X8, ADST_DCT, 12),
+  make_tuple(PARAM_LIST_8X8, DCT_ADST, 10),
+  make_tuple(PARAM_LIST_8X8, DCT_ADST, 12),
+  make_tuple(PARAM_LIST_8X8, ADST_ADST, 10),
+  make_tuple(PARAM_LIST_8X8, ADST_ADST, 12),
+#if CONFIG_EXT_TX
+  make_tuple(PARAM_LIST_8X8, FLIPADST_DCT, 10),
+  make_tuple(PARAM_LIST_8X8, FLIPADST_DCT, 12),
+  make_tuple(PARAM_LIST_8X8, DCT_FLIPADST, 10),
+  make_tuple(PARAM_LIST_8X8, DCT_FLIPADST, 12),
+  make_tuple(PARAM_LIST_8X8, FLIPADST_FLIPADST, 10),
+  make_tuple(PARAM_LIST_8X8, FLIPADST_FLIPADST, 12),
+  make_tuple(PARAM_LIST_8X8, ADST_FLIPADST, 10),
+  make_tuple(PARAM_LIST_8X8, ADST_FLIPADST, 12),
+  make_tuple(PARAM_LIST_8X8, FLIPADST_ADST, 10),
+  make_tuple(PARAM_LIST_8X8, FLIPADST_ADST, 12),
+#endif
+  // 4x4
+  make_tuple(PARAM_LIST_4X4, DCT_DCT, 10),
+  make_tuple(PARAM_LIST_4X4, DCT_DCT, 12),
+  make_tuple(PARAM_LIST_4X4, ADST_DCT, 10),
+  make_tuple(PARAM_LIST_4X4, ADST_DCT, 12),
+  make_tuple(PARAM_LIST_4X4, DCT_ADST, 10),
+  make_tuple(PARAM_LIST_4X4, DCT_ADST, 12),
+  make_tuple(PARAM_LIST_4X4, ADST_ADST, 10),
+  make_tuple(PARAM_LIST_4X4, ADST_ADST, 12),
+#if CONFIG_EXT_TX
+  make_tuple(PARAM_LIST_4X4, FLIPADST_DCT, 10),
+  make_tuple(PARAM_LIST_4X4, FLIPADST_DCT, 12),
+  make_tuple(PARAM_LIST_4X4, DCT_FLIPADST, 10),
+  make_tuple(PARAM_LIST_4X4, DCT_FLIPADST, 12),
+  make_tuple(PARAM_LIST_4X4, FLIPADST_FLIPADST, 10),
+  make_tuple(PARAM_LIST_4X4, FLIPADST_FLIPADST, 12),
+  make_tuple(PARAM_LIST_4X4, ADST_FLIPADST, 10),
+  make_tuple(PARAM_LIST_4X4, ADST_FLIPADST, 12),
+  make_tuple(PARAM_LIST_4X4, FLIPADST_ADST, 10),
+  make_tuple(PARAM_LIST_4X4, FLIPADST_ADST, 12),
+#endif
+};
+
+INSTANTIATE_TEST_CASE_P(SSE4_1, AV1HighbdInvHTNxN,
+                        ::testing::ValuesIn(kArrayIhtParam));
+#endif  // HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+
+#if HAVE_AVX2 && CONFIG_HIGHBITDEPTH
+#define PARAM_LIST_32X32                                   \
+  &av1_fwd_txfm2d_32x32_c, &av1_inv_txfm2d_add_32x32_avx2, \
+      &av1_inv_txfm2d_add_32x32_c, 1024
+
+const IHbdHtParam kArrayIhtParam32x32[] = {
+  // 32x32
+  make_tuple(PARAM_LIST_32X32, DCT_DCT, 10),
+  make_tuple(PARAM_LIST_32X32, DCT_DCT, 12),
+};
+
+INSTANTIATE_TEST_CASE_P(AVX2, AV1HighbdInvHTNxN,
+                        ::testing::ValuesIn(kArrayIhtParam32x32));
+
+#endif  // HAVE_AVX2 && CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/av1_inv_txfm1d_test.cc b/third_party/aom/test/av1_inv_txfm1d_test.cc
new file mode 100644
index 0000000000..9cf33a2fd4
--- /dev/null
+++ b/third_party/aom/test/av1_inv_txfm1d_test.cc
@@ -0,0 +1,87 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "test/av1_txfm_test.h"
+#include "av1/common/av1_fwd_txfm1d.h"
+#include "av1/common/av1_inv_txfm1d.h"
+
+using libaom_test::ACMRandom;
+using libaom_test::input_base;
+
+namespace {
+const int txfm_type_num = 2;
+const int txfm_size_ls[5] = { 4, 8, 16, 32, 64 };
+
+const TxfmFunc fwd_txfm_func_ls[][2] = {
+  { av1_fdct4_new, av1_fadst4_new },
+  { av1_fdct8_new, av1_fadst8_new },
+  { av1_fdct16_new, av1_fadst16_new },
+  { av1_fdct32_new, av1_fadst32_new },
+#if CONFIG_TX64X64
+  { av1_fdct64_new, NULL },
+#endif
+};
+
+const TxfmFunc inv_txfm_func_ls[][2] = {
+  { av1_idct4_new, av1_iadst4_new },
+  { av1_idct8_new, av1_iadst8_new },
+  { av1_idct16_new, av1_iadst16_new },
+  { av1_idct32_new, av1_iadst32_new },
+#if CONFIG_TX64X64
+  { av1_idct64_new, NULL },
+#endif
+};
+
+// the maximum stage number of fwd/inv 1d dct/adst txfm is 12
+const int8_t cos_bit[12] = { 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14, 14 };
+const int8_t range_bit[12] = { 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32, 32 };
+
+#define ARRAY_SIZE(x) (int)(sizeof(x) / sizeof(x[0]))
+
+TEST(av1_inv_txfm1d, round_trip) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  for (int si = 0; si < ARRAY_SIZE(fwd_txfm_func_ls); ++si) {
+    int txfm_size = txfm_size_ls[si];
+
+    for (int ti = 0; ti < txfm_type_num; ++ti) {
+      TxfmFunc fwd_txfm_func = fwd_txfm_func_ls[si][ti];
+      TxfmFunc inv_txfm_func = inv_txfm_func_ls[si][ti];
+      int max_error = 2;
+
+      if (!fwd_txfm_func) continue;
+
+      const int count_test_block = 5000;
+      for (int ci = 0; ci < count_test_block; ++ci) {
+        int32_t input[64];
+        int32_t output[64];
+        int32_t round_trip_output[64];
+
+        assert(txfm_size <= ARRAY_SIZE(input));
+
+        for (int ni = 0; ni < txfm_size; ++ni) {
+          input[ni] = rnd.Rand16() % input_base - rnd.Rand16() % input_base;
+        }
+
+        fwd_txfm_func(input, output, cos_bit, range_bit);
+        inv_txfm_func(output, round_trip_output, cos_bit, range_bit);
+
+        for (int ni = 0; ni < txfm_size; ++ni) {
+          int node_err =
+              abs(input[ni] - round_shift(round_trip_output[ni],
+                                          get_max_bit(txfm_size) - 1));
+          EXPECT_LE(node_err, max_error);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace
diff --git a/third_party/aom/test/av1_inv_txfm2d_test.cc b/third_party/aom/test/av1_inv_txfm2d_test.cc
new file mode 100644
index 0000000000..bb2743af1b
--- /dev/null
+++ b/third_party/aom/test/av1_inv_txfm2d_test.cc
@@ -0,0 +1,158 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "./av1_rtcd.h"
+#include "test/acm_random.h"
+#include "test/util.h"
+#include "test/av1_txfm_test.h"
+#include "av1/common/av1_inv_txfm2d_cfg.h"
+
+using libaom_test::ACMRandom;
+using libaom_test::input_base;
+using libaom_test::bd;
+using libaom_test::compute_avg_abs_error;
+using libaom_test::Fwd_Txfm2d_Func;
+using libaom_test::Inv_Txfm2d_Func;
+
+namespace {
+
+#if CONFIG_HIGHBITDEPTH
+// AV1InvTxfm2dParam argument list:
+// tx_type_, tx_size_, max_error_, max_avg_error_
+typedef std::tr1::tuple<TX_TYPE, TX_SIZE, int, double> AV1InvTxfm2dParam;
+
+class AV1InvTxfm2d : public ::testing::TestWithParam<AV1InvTxfm2dParam> {
+ public:
+  virtual void SetUp() {
+    tx_type_ = GET_PARAM(0);
+    tx_size_ = GET_PARAM(1);
+    max_error_ = GET_PARAM(2);
+    max_avg_error_ = GET_PARAM(3);
+    txfm1d_size_ = libaom_test::get_txfm1d_size(tx_size_);
+    txfm2d_size_ = txfm1d_size_ * txfm1d_size_;
+    count_ = 500;
+
+    input_ = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(int16_t) * txfm2d_size_));
+    ref_input_ = reinterpret_cast<uint16_t *>(
+        aom_memalign(16, sizeof(uint16_t) * txfm2d_size_));
+    output_ = reinterpret_cast<int32_t *>(
+        aom_memalign(16, sizeof(int32_t) * txfm2d_size_));
+  }
+
+  void RunRoundtripCheck() {
+    const Fwd_Txfm2d_Func fwd_txfm_func =
+        libaom_test::fwd_txfm_func_ls[tx_size_];
+    const Inv_Txfm2d_Func inv_txfm_func =
+        libaom_test::inv_txfm_func_ls[tx_size_];
+    double avg_abs_error = 0;
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    for (int ci = 0; ci < count_; ci++) {
+      for (int ni = 0; ni < txfm2d_size_; ++ni) {
+        if (ci == 0) {
+          int extreme_input = input_base - 1;
+          input_[ni] = extreme_input;  // extreme case
+          ref_input_[ni] = 0;
+        } else {
+          input_[ni] = rnd.Rand16() % input_base;
+          ref_input_[ni] = 0;
+        }
+      }
+
+      fwd_txfm_func(input_, output_, txfm1d_size_, tx_type_, bd);
+      inv_txfm_func(output_, ref_input_, txfm1d_size_, tx_type_, bd);
+
+      for (int ni = 0; ni < txfm2d_size_; ++ni) {
+        EXPECT_GE(max_error_, abs(input_[ni] - ref_input_[ni]));
+      }
+      avg_abs_error += compute_avg_abs_error<int16_t, uint16_t>(
+          input_, ref_input_, txfm2d_size_);
+    }
+
+    avg_abs_error /= count_;
+    // max_abs_avg_error comes from upper bound of
+    // printf("txfm1d_size: %d accuracy_avg_abs_error: %f\n",
+    // txfm1d_size_, avg_abs_error);
+    EXPECT_GE(max_avg_error_, avg_abs_error);
+  }
+
+  virtual void TearDown() {
+    aom_free(input_);
+    aom_free(output_);
+    aom_free(ref_input_);
+  }
+
+ private:
+  int count_;
+  int max_error_;
+  double max_avg_error_;
+  TX_TYPE tx_type_;
+  TX_SIZE tx_size_;
+  int txfm1d_size_;
+  int txfm2d_size_;
+  int16_t *input_;
+  uint16_t *ref_input_;
+  int32_t *output_;
+};
+
+TEST_P(AV1InvTxfm2d, RunRoundtripCheck) { RunRoundtripCheck(); }
+
+const AV1InvTxfm2dParam av1_inv_txfm2d_param[] = {
+#if CONFIG_EXT_TX
+  AV1InvTxfm2dParam(FLIPADST_DCT, TX_4X4, 2, 0.002),
+  AV1InvTxfm2dParam(DCT_FLIPADST, TX_4X4, 2, 0.002),
+  AV1InvTxfm2dParam(FLIPADST_FLIPADST, TX_4X4, 2, 0.002),
+  AV1InvTxfm2dParam(ADST_FLIPADST, TX_4X4, 2, 0.002),
+  AV1InvTxfm2dParam(FLIPADST_ADST, TX_4X4, 2, 0.002),
+  AV1InvTxfm2dParam(FLIPADST_DCT, TX_8X8, 2, 0.02),
+  AV1InvTxfm2dParam(DCT_FLIPADST, TX_8X8, 2, 0.02),
+  AV1InvTxfm2dParam(FLIPADST_FLIPADST, TX_8X8, 2, 0.02),
+  AV1InvTxfm2dParam(ADST_FLIPADST, TX_8X8, 2, 0.02),
+  AV1InvTxfm2dParam(FLIPADST_ADST, TX_8X8, 2, 0.02),
+  AV1InvTxfm2dParam(FLIPADST_DCT, TX_16X16, 2, 0.04),
+  AV1InvTxfm2dParam(DCT_FLIPADST, TX_16X16, 2, 0.04),
+  AV1InvTxfm2dParam(FLIPADST_FLIPADST, TX_16X16, 11, 0.04),
+  AV1InvTxfm2dParam(ADST_FLIPADST, TX_16X16, 2, 0.04),
+  AV1InvTxfm2dParam(FLIPADST_ADST, TX_16X16, 2, 0.04),
+  AV1InvTxfm2dParam(FLIPADST_DCT, TX_32X32, 4, 0.4),
+  AV1InvTxfm2dParam(DCT_FLIPADST, TX_32X32, 4, 0.4),
+  AV1InvTxfm2dParam(FLIPADST_FLIPADST, TX_32X32, 4, 0.4),
+  AV1InvTxfm2dParam(ADST_FLIPADST, TX_32X32, 4, 0.4),
+  AV1InvTxfm2dParam(FLIPADST_ADST, TX_32X32, 4, 0.4),
+#endif
+  AV1InvTxfm2dParam(DCT_DCT, TX_4X4, 2, 0.002),
+  AV1InvTxfm2dParam(ADST_DCT, TX_4X4, 2, 0.002),
+  AV1InvTxfm2dParam(DCT_ADST, TX_4X4, 2, 0.002),
+  AV1InvTxfm2dParam(ADST_ADST, TX_4X4, 2, 0.002),
+  AV1InvTxfm2dParam(DCT_DCT, TX_8X8, 2, 0.02),
+  AV1InvTxfm2dParam(ADST_DCT, TX_8X8, 2, 0.02),
+  AV1InvTxfm2dParam(DCT_ADST, TX_8X8, 2, 0.02),
+  AV1InvTxfm2dParam(ADST_ADST, TX_8X8, 2, 0.02),
+  AV1InvTxfm2dParam(DCT_DCT, TX_16X16, 2, 0.04),
+  AV1InvTxfm2dParam(ADST_DCT, TX_16X16, 2, 0.04),
+  AV1InvTxfm2dParam(DCT_ADST, TX_16X16, 2, 0.04),
+  AV1InvTxfm2dParam(ADST_ADST, TX_16X16, 2, 0.04),
+  AV1InvTxfm2dParam(DCT_DCT, TX_32X32, 4, 0.4),
+  AV1InvTxfm2dParam(ADST_DCT, TX_32X32, 4, 0.4),
+  AV1InvTxfm2dParam(DCT_ADST, TX_32X32, 4, 0.4),
+  AV1InvTxfm2dParam(ADST_ADST, TX_32X32, 4, 0.4)
+};
+
+INSTANTIATE_TEST_CASE_P(C, AV1InvTxfm2d,
+                        ::testing::ValuesIn(av1_inv_txfm2d_param));
+
+#endif  // CONFIG_HIGHBITDEPTH
+
+}  // namespace
diff --git a/third_party/aom/test/av1_inv_txfm_test.cc b/third_party/aom/test/av1_inv_txfm_test.cc
new file mode 100644
index 0000000000..af3fee8726
--- /dev/null
+++ b/third_party/aom/test/av1_inv_txfm_test.cc
@@ -0,0 +1,282 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "av1/common/blockd.h"
+#include "av1/common/scan.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/inv_txfm.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+const double kInvSqrt2 = 0.707106781186547524400844362104;
+
+void reference_idct_1d(const double *in, double *out, int size) {
+  for (int n = 0; n < size; ++n) {
+    out[n] = 0;
+    for (int k = 0; k < size; ++k) {
+      if (k == 0)
+        out[n] += kInvSqrt2 * in[k] * cos(PI * (2 * n + 1) * k / (2 * size));
+      else
+        out[n] += in[k] * cos(PI * (2 * n + 1) * k / (2 * size));
+    }
+  }
+}
+
+typedef void (*IdctFuncRef)(const double *in, double *out, int size);
+typedef void (*IdctFunc)(const tran_low_t *in, tran_low_t *out);
+
+class TransTestBase {
+ public:
+  virtual ~TransTestBase() {}
+
+ protected:
+  void RunInvAccuracyCheck() {
+    tran_low_t *input = new tran_low_t[txfm_size_];
+    tran_low_t *output = new tran_low_t[txfm_size_];
+    double *ref_input = new double[txfm_size_];
+    double *ref_output = new double[txfm_size_];
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 5000;
+    for (int ti = 0; ti < count_test_block; ++ti) {
+      for (int ni = 0; ni < txfm_size_; ++ni) {
+        input[ni] = rnd.Rand8() - rnd.Rand8();
+        ref_input[ni] = static_cast<double>(input[ni]);
+      }
+
+      fwd_txfm_(input, output);
+      fwd_txfm_ref_(ref_input, ref_output, txfm_size_);
+
+      for (int ni = 0; ni < txfm_size_; ++ni) {
+        EXPECT_LE(
+            abs(output[ni] - static_cast<tran_low_t>(round(ref_output[ni]))),
+            max_error_);
+      }
+    }
+
+    delete[] input;
+    delete[] output;
+    delete[] ref_input;
+    delete[] ref_output;
+  }
+
+  double max_error_;
+  int txfm_size_;
+  IdctFunc fwd_txfm_;
+  IdctFuncRef fwd_txfm_ref_;
+};
+
+typedef std::tr1::tuple<IdctFunc, IdctFuncRef, int, int> IdctParam;
+class AV1InvTxfm : public TransTestBase,
+                   public ::testing::TestWithParam<IdctParam> {
+ public:
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    fwd_txfm_ref_ = GET_PARAM(1);
+    txfm_size_ = GET_PARAM(2);
+    max_error_ = GET_PARAM(3);
+  }
+  virtual void TearDown() {}
+};
+
+TEST_P(AV1InvTxfm, RunInvAccuracyCheck) { RunInvAccuracyCheck(); }
+
+INSTANTIATE_TEST_CASE_P(
+    C, AV1InvTxfm,
+    ::testing::Values(IdctParam(&aom_idct4_c, &reference_idct_1d, 4, 1),
+                      IdctParam(&aom_idct8_c, &reference_idct_1d, 8, 2),
+                      IdctParam(&aom_idct16_c, &reference_idct_1d, 16, 4),
+                      IdctParam(&aom_idct32_c, &reference_idct_1d, 32, 6)));
+
+#if CONFIG_AV1_ENCODER
+typedef void (*FwdTxfmFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*InvTxfmFunc)(const tran_low_t *in, uint8_t *out, int stride);
+typedef std::tr1::tuple<FwdTxfmFunc, InvTxfmFunc, InvTxfmFunc, TX_SIZE, int>
+    PartialInvTxfmParam;
+#if !CONFIG_ADAPT_SCAN
+const int kMaxNumCoeffs = 1024;
+#endif
+class AV1PartialIDctTest
+    : public ::testing::TestWithParam<PartialInvTxfmParam> {
+ public:
+  virtual ~AV1PartialIDctTest() {}
+  virtual void SetUp() {
+    ftxfm_ = GET_PARAM(0);
+    full_itxfm_ = GET_PARAM(1);
+    partial_itxfm_ = GET_PARAM(2);
+    tx_size_ = GET_PARAM(3);
+    last_nonzero_ = GET_PARAM(4);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  int last_nonzero_;
+  TX_SIZE tx_size_;
+  FwdTxfmFunc ftxfm_;
+  InvTxfmFunc full_itxfm_;
+  InvTxfmFunc partial_itxfm_;
+};
+
+#if !CONFIG_ADAPT_SCAN
+TEST_P(AV1PartialIDctTest, RunQuantCheck) {
+  int size;
+  switch (tx_size_) {
+    case TX_4X4: size = 4; break;
+    case TX_8X8: size = 8; break;
+    case TX_16X16: size = 16; break;
+    case TX_32X32: size = 32; break;
+    default: FAIL() << "Wrong Size!"; break;
+  }
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block2[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst2[kMaxNumCoeffs]);
+
+  const int count_test_block = 1000;
+  const int block_size = size * size;
+
+  DECLARE_ALIGNED(16, int16_t, input_extreme_block[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kMaxNumCoeffs]);
+
+  int max_error = 0;
+  for (int m = 0; m < count_test_block; ++m) {
+    // clear out destination buffer
+    memset(dst1, 0, sizeof(*dst1) * block_size);
+    memset(dst2, 0, sizeof(*dst2) * block_size);
+    memset(test_coef_block1, 0, sizeof(*test_coef_block1) * block_size);
+    memset(test_coef_block2, 0, sizeof(*test_coef_block2) * block_size);
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+    for (int n = 0; n < count_test_block; ++n) {
+      // Initialize a test block with input range [-255, 255].
+      if (n == 0) {
+        for (int j = 0; j < block_size; ++j) input_extreme_block[j] = 255;
+      } else if (n == 1) {
+        for (int j = 0; j < block_size; ++j) input_extreme_block[j] = -255;
+      } else {
+        for (int j = 0; j < block_size; ++j) {
+          input_extreme_block[j] = rnd.Rand8() % 2 ? 255 : -255;
+        }
+      }
+
+      ftxfm_(input_extreme_block, output_ref_block, size);
+
+      // quantization with maximum allowed step sizes
+      test_coef_block1[0] = (output_ref_block[0] / 1336) * 1336;
+      for (int j = 1; j < last_nonzero_; ++j)
+        test_coef_block1[get_scan((const AV1_COMMON *)NULL, tx_size_, DCT_DCT,
+                                  0)
+                             ->scan[j]] = (output_ref_block[j] / 1828) * 1828;
+    }
+
+    ASM_REGISTER_STATE_CHECK(full_itxfm_(test_coef_block1, dst1, size));
+    ASM_REGISTER_STATE_CHECK(partial_itxfm_(test_coef_block1, dst2, size));
+
+    for (int j = 0; j < block_size; ++j) {
+      const int diff = dst1[j] - dst2[j];
+      const int error = diff * diff;
+      if (max_error < error) max_error = error;
+    }
+  }
+
+  EXPECT_EQ(0, max_error)
+      << "Error: partial inverse transform produces different results";
+}
+
+TEST_P(AV1PartialIDctTest, ResultsMatch) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int size;
+  switch (tx_size_) {
+    case TX_4X4: size = 4; break;
+    case TX_8X8: size = 8; break;
+    case TX_16X16: size = 16; break;
+    case TX_32X32: size = 32; break;
+    default: FAIL() << "Wrong Size!"; break;
+  }
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, test_coef_block2[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst1[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst2[kMaxNumCoeffs]);
+  const int count_test_block = 1000;
+  const int max_coeff = 32766 / 4;
+  const int block_size = size * size;
+  int max_error = 0;
+  for (int i = 0; i < count_test_block; ++i) {
+    // clear out destination buffer
+    memset(dst1, 0, sizeof(*dst1) * block_size);
+    memset(dst2, 0, sizeof(*dst2) * block_size);
+    memset(test_coef_block1, 0, sizeof(*test_coef_block1) * block_size);
+    memset(test_coef_block2, 0, sizeof(*test_coef_block2) * block_size);
+    int max_energy_leftover = max_coeff * max_coeff;
+    for (int j = 0; j < last_nonzero_; ++j) {
+      int16_t coef = static_cast<int16_t>(sqrt(1.0 * max_energy_leftover) *
+                                          (rnd.Rand16() - 32768) / 65536);
+      max_energy_leftover -= coef * coef;
+      if (max_energy_leftover < 0) {
+        max_energy_leftover = 0;
+        coef = 0;
+      }
+      test_coef_block1[get_scan((const AV1_COMMON *)NULL, tx_size_, DCT_DCT, 0)
+                           ->scan[j]] = coef;
+    }
+
+    memcpy(test_coef_block2, test_coef_block1,
+           sizeof(*test_coef_block2) * block_size);
+
+    ASM_REGISTER_STATE_CHECK(full_itxfm_(test_coef_block1, dst1, size));
+    ASM_REGISTER_STATE_CHECK(partial_itxfm_(test_coef_block2, dst2, size));
+
+    for (int j = 0; j < block_size; ++j) {
+      const int diff = dst1[j] - dst2[j];
+      const int error = diff * diff;
+      if (max_error < error) max_error = error;
+    }
+  }
+
+  EXPECT_EQ(0, max_error)
+      << "Error: partial inverse transform produces different results";
+}
+#endif
+using std::tr1::make_tuple;
+
+INSTANTIATE_TEST_CASE_P(
+    C, AV1PartialIDctTest,
+    ::testing::Values(make_tuple(&aom_fdct32x32_c, &aom_idct32x32_1024_add_c,
+                                 &aom_idct32x32_34_add_c, TX_32X32, 34),
+                      make_tuple(&aom_fdct32x32_c, &aom_idct32x32_1024_add_c,
+                                 &aom_idct32x32_1_add_c, TX_32X32, 1),
+                      make_tuple(&aom_fdct16x16_c, &aom_idct16x16_256_add_c,
+                                 &aom_idct16x16_10_add_c, TX_16X16, 10),
+                      make_tuple(&aom_fdct16x16_c, &aom_idct16x16_256_add_c,
+                                 &aom_idct16x16_1_add_c, TX_16X16, 1),
+                      make_tuple(&aom_fdct8x8_c, &aom_idct8x8_64_add_c,
+                                 &aom_idct8x8_12_add_c, TX_8X8, 12),
+                      make_tuple(&aom_fdct8x8_c, &aom_idct8x8_64_add_c,
+                                 &aom_idct8x8_1_add_c, TX_8X8, 1),
+                      make_tuple(&aom_fdct4x4_c, &aom_idct4x4_16_add_c,
+                                 &aom_idct4x4_1_add_c, TX_4X4, 1)));
+#endif  // CONFIG_AV1_ENCODER
+}  // namespace
diff --git a/third_party/aom/test/av1_quantize_test.cc b/third_party/aom/test/av1_quantize_test.cc
new file mode 100644
index 0000000000..b5d1531f5c
--- /dev/null
+++ b/third_party/aom/test/av1_quantize_test.cc
@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <stdlib.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "av1/common/scan.h"
+
+namespace {
+
+typedef void (*QuantizeFpFunc)(
+    const tran_low_t *coeff_ptr, intptr_t count, int skip_block,
+    const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan, int log_scale);
+
+struct QuantizeFuncParams {
+  QuantizeFuncParams(QuantizeFpFunc qF = NULL, QuantizeFpFunc qRefF = NULL,
+                     int count = 16)
+      : qFunc(qF), qFuncRef(qRefF), coeffCount(count) {}
+  QuantizeFpFunc qFunc;
+  QuantizeFpFunc qFuncRef;
+  int coeffCount;
+};
+
+using libaom_test::ACMRandom;
+
+const int numTests = 1000;
+const int maxSize = 1024;
+const int roundFactorRange = 127;
+const int dequantRange = 32768;
+const int coeffRange = (1 << 20) - 1;
+
+class AV1QuantizeTest : public ::testing::TestWithParam<QuantizeFuncParams> {
+ public:
+  void RunQuantizeTest() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    DECLARE_ALIGNED(16, tran_low_t, coeff_ptr[maxSize]);
+    DECLARE_ALIGNED(16, int16_t, zbin_ptr[2]);
+    DECLARE_ALIGNED(16, int16_t, round_ptr[2]);
+    DECLARE_ALIGNED(16, int16_t, quant_ptr[2]);
+    DECLARE_ALIGNED(16, int16_t, quant_shift_ptr[2]);
+    DECLARE_ALIGNED(16, tran_low_t, qcoeff_ptr[maxSize]);
+    DECLARE_ALIGNED(16, tran_low_t, dqcoeff_ptr[maxSize]);
+    DECLARE_ALIGNED(16, tran_low_t, ref_qcoeff_ptr[maxSize]);
+    DECLARE_ALIGNED(16, tran_low_t, ref_dqcoeff_ptr[maxSize]);
+    DECLARE_ALIGNED(16, int16_t, dequant_ptr[2]);
+    uint16_t eob;
+    uint16_t ref_eob;
+    int err_count_total = 0;
+    int first_failure = -1;
+    int skip_block = 0;
+    int count = params_.coeffCount;
+    const TX_SIZE txSize = getTxSize(count);
+    int log_scale = (txSize == TX_32X32);
+    QuantizeFpFunc quanFunc = params_.qFunc;
+    QuantizeFpFunc quanFuncRef = params_.qFuncRef;
+
+    const SCAN_ORDER scanOrder = av1_default_scan_orders[txSize];
+    for (int i = 0; i < numTests; i++) {
+      int err_count = 0;
+      ref_eob = eob = -1;
+      for (int j = 0; j < count; j++) {
+        coeff_ptr[j] = rnd(coeffRange);
+      }
+
+      for (int j = 0; j < 2; j++) {
+        zbin_ptr[j] = rnd.Rand16();
+        quant_shift_ptr[j] = rnd.Rand16();
+        // int16_t positive
+        dequant_ptr[j] = abs(rnd(dequantRange));
+        quant_ptr[j] = (1 << 16) / dequant_ptr[j];
+        round_ptr[j] = (abs(rnd(roundFactorRange)) * dequant_ptr[j]) >> 7;
+      }
+
+      quanFuncRef(coeff_ptr, count, skip_block, zbin_ptr, round_ptr, quant_ptr,
+                  quant_shift_ptr, ref_qcoeff_ptr, ref_dqcoeff_ptr, dequant_ptr,
+                  &ref_eob, scanOrder.scan, scanOrder.iscan, log_scale);
+
+      ASM_REGISTER_STATE_CHECK(
+          quanFunc(coeff_ptr, count, skip_block, zbin_ptr, round_ptr, quant_ptr,
+                   quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr, &eob,
+                   scanOrder.scan, scanOrder.iscan, log_scale));
+
+      for (int j = 0; j < count; ++j) {
+        err_count += (ref_qcoeff_ptr[j] != qcoeff_ptr[j]) |
+                     (ref_dqcoeff_ptr[j] != dqcoeff_ptr[j]);
+        EXPECT_EQ(ref_qcoeff_ptr[j], qcoeff_ptr[j]) << "qcoeff error: i = " << i
+                                                    << " j = " << j << "\n";
+        EXPECT_EQ(ref_dqcoeff_ptr[j], dqcoeff_ptr[j])
+            << "dqcoeff error: i = " << i << " j = " << j << "\n";
+      }
+      EXPECT_EQ(ref_eob, eob) << "eob error: "
+                              << "i = " << i << "\n";
+      err_count += (ref_eob != eob);
+      if (err_count && !err_count_total) {
+        first_failure = i;
+      }
+      err_count_total += err_count;
+    }
+    EXPECT_EQ(0, err_count_total)
+        << "Error: Quantization Test, C output doesn't match SSE2 output. "
+        << "First failed at test case " << first_failure;
+  }
+
+  void RunEobTest() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    DECLARE_ALIGNED(16, tran_low_t, coeff_ptr[maxSize]);
+    DECLARE_ALIGNED(16, int16_t, zbin_ptr[2]);
+    DECLARE_ALIGNED(16, int16_t, round_ptr[2]);
+    DECLARE_ALIGNED(16, int16_t, quant_ptr[2]);
+    DECLARE_ALIGNED(16, int16_t, quant_shift_ptr[2]);
+    DECLARE_ALIGNED(16, tran_low_t, qcoeff_ptr[maxSize]);
+    DECLARE_ALIGNED(16, tran_low_t, dqcoeff_ptr[maxSize]);
+    DECLARE_ALIGNED(16, tran_low_t, ref_qcoeff_ptr[maxSize]);
+    DECLARE_ALIGNED(16, tran_low_t, ref_dqcoeff_ptr[maxSize]);
+    DECLARE_ALIGNED(16, int16_t, dequant_ptr[2]);
+    uint16_t eob;
+    uint16_t ref_eob;
+    int skip_block = 0;
+    int count = params_.coeffCount;
+    const TX_SIZE txSize = getTxSize(count);
+    int log_scale = (txSize == TX_32X32);
+    QuantizeFpFunc quanFunc = params_.qFunc;
+    QuantizeFpFunc quanFuncRef = params_.qFuncRef;
+    const SCAN_ORDER scanOrder = av1_default_scan_orders[txSize];
+
+    for (int i = 0; i < numTests; i++) {
+      ref_eob = eob = -1;
+      for (int j = 0; j < count; j++) {
+        coeff_ptr[j] = 0;
+      }
+
+      coeff_ptr[rnd(count)] = rnd(coeffRange);
+      coeff_ptr[rnd(count)] = rnd(coeffRange);
+      coeff_ptr[rnd(count)] = rnd(coeffRange);
+
+      for (int j = 0; j < 2; j++) {
+        zbin_ptr[j] = rnd.Rand16();
+        quant_shift_ptr[j] = rnd.Rand16();
+        // int16_t positive
+        dequant_ptr[j] = abs(rnd(dequantRange));
+        quant_ptr[j] = (1 << 16) / dequant_ptr[j];
+        round_ptr[j] = (abs(rnd(roundFactorRange)) * dequant_ptr[j]) >> 7;
+      }
+
+      quanFuncRef(coeff_ptr, count, skip_block, zbin_ptr, round_ptr, quant_ptr,
+                  quant_shift_ptr, ref_qcoeff_ptr, ref_dqcoeff_ptr, dequant_ptr,
+                  &ref_eob, scanOrder.scan, scanOrder.iscan, log_scale);
+
+      ASM_REGISTER_STATE_CHECK(
+          quanFunc(coeff_ptr, count, skip_block, zbin_ptr, round_ptr, quant_ptr,
+                   quant_shift_ptr, qcoeff_ptr, dqcoeff_ptr, dequant_ptr, &eob,
+                   scanOrder.scan, scanOrder.iscan, log_scale));
+      EXPECT_EQ(ref_eob, eob) << "eob error: "
+                              << "i = " << i << "\n";
+    }
+  }
+
+  virtual void SetUp() { params_ = GetParam(); }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+  virtual ~AV1QuantizeTest() {}
+
+ private:
+  TX_SIZE getTxSize(int count) {
+    switch (count) {
+      case 16: return TX_4X4;
+      case 64: return TX_8X8;
+      case 256: return TX_16X16;
+      case 1024: return TX_32X32;
+      default: return TX_4X4;
+    }
+  }
+
+  QuantizeFuncParams params_;
+};
+
+TEST_P(AV1QuantizeTest, BitExactCheck) { RunQuantizeTest(); }
+TEST_P(AV1QuantizeTest, EobVerify) { RunEobTest(); }
+
+#if HAVE_SSE4_1
+#if !CONFIG_AOM_QM
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, AV1QuantizeTest,
+    ::testing::Values(QuantizeFuncParams(&av1_highbd_quantize_fp_sse4_1,
+                                         &av1_highbd_quantize_fp_c, 16),
+                      QuantizeFuncParams(&av1_highbd_quantize_fp_sse4_1,
+                                         &av1_highbd_quantize_fp_c, 64),
+                      QuantizeFuncParams(&av1_highbd_quantize_fp_sse4_1,
+                                         &av1_highbd_quantize_fp_c, 256),
+                      QuantizeFuncParams(&av1_highbd_quantize_fp_sse4_1,
+                                         &av1_highbd_quantize_fp_c, 1024)));
+#endif  // !CONFIG_AOM_QM
+#endif  // HAVE_SSE4_1
+}  // namespace
diff --git a/third_party/aom/test/av1_txfm_test.cc b/third_party/aom/test/av1_txfm_test.cc
new file mode 100644
index 0000000000..1e473b3044
--- /dev/null
+++ b/third_party/aom/test/av1_txfm_test.cc
@@ -0,0 +1,164 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdio.h>
+#include "test/av1_txfm_test.h"
+
+namespace libaom_test {
+
+int get_txfm1d_size(TX_SIZE tx_size) { return tx_size_wide[tx_size]; }
+
+void get_txfm1d_type(TX_TYPE txfm2d_type, TYPE_TXFM *type0, TYPE_TXFM *type1) {
+  switch (txfm2d_type) {
+    case DCT_DCT:
+      *type0 = TYPE_DCT;
+      *type1 = TYPE_DCT;
+      break;
+    case ADST_DCT:
+      *type0 = TYPE_ADST;
+      *type1 = TYPE_DCT;
+      break;
+    case DCT_ADST:
+      *type0 = TYPE_DCT;
+      *type1 = TYPE_ADST;
+      break;
+    case ADST_ADST:
+      *type0 = TYPE_ADST;
+      *type1 = TYPE_ADST;
+      break;
+#if CONFIG_EXT_TX
+    case FLIPADST_DCT:
+      *type0 = TYPE_ADST;
+      *type1 = TYPE_DCT;
+      break;
+    case DCT_FLIPADST:
+      *type0 = TYPE_DCT;
+      *type1 = TYPE_ADST;
+      break;
+    case FLIPADST_FLIPADST:
+      *type0 = TYPE_ADST;
+      *type1 = TYPE_ADST;
+      break;
+    case ADST_FLIPADST:
+      *type0 = TYPE_ADST;
+      *type1 = TYPE_ADST;
+      break;
+    case FLIPADST_ADST:
+      *type0 = TYPE_ADST;
+      *type1 = TYPE_ADST;
+      break;
+#endif  // CONFIG_EXT_TX
+    default:
+      *type0 = TYPE_DCT;
+      *type1 = TYPE_DCT;
+      assert(0);
+      break;
+  }
+}
+
+double invSqrt2 = 1 / pow(2, 0.5);
+
+void reference_dct_1d(const double *in, double *out, int size) {
+  for (int k = 0; k < size; ++k) {
+    out[k] = 0;
+    for (int n = 0; n < size; ++n) {
+      out[k] += in[n] * cos(M_PI * (2 * n + 1) * k / (2 * size));
+    }
+    if (k == 0) out[k] = out[k] * invSqrt2;
+  }
+}
+
+void reference_adst_1d(const double *in, double *out, int size) {
+  for (int k = 0; k < size; ++k) {
+    out[k] = 0;
+    for (int n = 0; n < size; ++n) {
+      out[k] += in[n] * sin(M_PI * (2 * n + 1) * (2 * k + 1) / (4 * size));
+    }
+  }
+}
+
+void reference_hybrid_1d(double *in, double *out, int size, int type) {
+  if (type == TYPE_DCT)
+    reference_dct_1d(in, out, size);
+  else
+    reference_adst_1d(in, out, size);
+}
+
+void reference_hybrid_2d(double *in, double *out, int size, int type0,
+                         int type1) {
+  double *tempOut = new double[size * size];
+
+  for (int r = 0; r < size; r++) {
+    // out ->tempOut
+    for (int c = 0; c < size; c++) {
+      tempOut[r * size + c] = in[c * size + r];
+    }
+  }
+
+  // dct each row: in -> out
+  for (int r = 0; r < size; r++) {
+    reference_hybrid_1d(tempOut + r * size, out + r * size, size, type0);
+  }
+
+  for (int r = 0; r < size; r++) {
+    // out ->tempOut
+    for (int c = 0; c < size; c++) {
+      tempOut[r * size + c] = out[c * size + r];
+    }
+  }
+
+  for (int r = 0; r < size; r++) {
+    reference_hybrid_1d(tempOut + r * size, out + r * size, size, type1);
+  }
+  delete[] tempOut;
+}
+
+template <typename Type>
+void fliplr(Type *dest, int stride, int length) {
+  int i, j;
+  for (i = 0; i < length; ++i) {
+    for (j = 0; j < length / 2; ++j) {
+      const Type tmp = dest[i * stride + j];
+      dest[i * stride + j] = dest[i * stride + length - 1 - j];
+      dest[i * stride + length - 1 - j] = tmp;
+    }
+  }
+}
+
+template <typename Type>
+void flipud(Type *dest, int stride, int length) {
+  int i, j;
+  for (j = 0; j < length; ++j) {
+    for (i = 0; i < length / 2; ++i) {
+      const Type tmp = dest[i * stride + j];
+      dest[i * stride + j] = dest[(length - 1 - i) * stride + j];
+      dest[(length - 1 - i) * stride + j] = tmp;
+    }
+  }
+}
+
+template <typename Type>
+void fliplrud(Type *dest, int stride, int length) {
+  int i, j;
+  for (i = 0; i < length / 2; ++i) {
+    for (j = 0; j < length; ++j) {
+      const Type tmp = dest[i * stride + j];
+      dest[i * stride + j] = dest[(length - 1 - i) * stride + length - 1 - j];
+      dest[(length - 1 - i) * stride + length - 1 - j] = tmp;
+    }
+  }
+}
+
+template void fliplr<double>(double *dest, int stride, int length);
+template void flipud<double>(double *dest, int stride, int length);
+template void fliplrud<double>(double *dest, int stride, int length);
+
+}  // namespace libaom_test
diff --git a/third_party/aom/test/av1_txfm_test.h b/third_party/aom/test/av1_txfm_test.h
new file mode 100644
index 0000000000..70f971d090
--- /dev/null
+++ b/third_party/aom/test/av1_txfm_test.h
@@ -0,0 +1,100 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef AV1_TXFM_TEST_H_
+#define AV1_TXFM_TEST_H_
+
+#include <stdio.h>
+#include <stdlib.h>
+#ifdef _MSC_VER
+#define _USE_MATH_DEFINES
+#endif
+#include <math.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/acm_random.h"
+#include "av1/common/enums.h"
+#include "av1/common/av1_txfm.h"
+#include "./av1_rtcd.h"
+
+namespace libaom_test {
+typedef enum {
+  TYPE_DCT = 0,
+  TYPE_ADST,
+  TYPE_IDCT,
+  TYPE_IADST,
+  TYPE_LAST
+} TYPE_TXFM;
+
+int get_txfm1d_size(TX_SIZE tx_size);
+
+void get_txfm1d_type(TX_TYPE txfm2d_type, TYPE_TXFM *type0, TYPE_TXFM *type1);
+
+void reference_dct_1d(const double *in, double *out, int size);
+
+void reference_adst_1d(const double *in, double *out, int size);
+
+void reference_hybrid_1d(double *in, double *out, int size, int type);
+
+void reference_hybrid_2d(double *in, double *out, int size, int type0,
+                         int type1);
+template <typename Type1, typename Type2>
+static double compute_avg_abs_error(const Type1 *a, const Type2 *b,
+                                    const int size) {
+  double error = 0;
+  for (int i = 0; i < size; i++) {
+    error += fabs(static_cast<double>(a[i]) - static_cast<double>(b[i]));
+  }
+  error = error / size;
+  return error;
+}
+
+template <typename Type>
+void fliplr(Type *dest, int stride, int length);
+
+template <typename Type>
+void flipud(Type *dest, int stride, int length);
+
+template <typename Type>
+void fliplrud(Type *dest, int stride, int length);
+
+typedef void (*TxfmFunc)(const int32_t *in, int32_t *out, const int8_t *cos_bit,
+                         const int8_t *range_bit);
+
+typedef void (*Fwd_Txfm2d_Func)(const int16_t *, int32_t *, int, int, int);
+typedef void (*Inv_Txfm2d_Func)(const int32_t *, uint16_t *, int, int, int);
+
+static const int bd = 10;
+static const int input_base = (1 << bd);
+
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_AV1_ENCODER
+static const Fwd_Txfm2d_Func fwd_txfm_func_ls[TX_SIZES] = {
+#if CONFIG_CB4X4
+  NULL,
+#endif
+  av1_fwd_txfm2d_4x4_c, av1_fwd_txfm2d_8x8_c, av1_fwd_txfm2d_16x16_c,
+  av1_fwd_txfm2d_32x32_c
+};
+#endif
+
+static const Inv_Txfm2d_Func inv_txfm_func_ls[TX_SIZES] = {
+#if CONFIG_CB4X4
+  NULL,
+#endif
+  av1_inv_txfm2d_add_4x4_c, av1_inv_txfm2d_add_8x8_c,
+  av1_inv_txfm2d_add_16x16_c, av1_inv_txfm2d_add_32x32_c
+};
+#endif  // CONFIG_HIGHBITDEPTH
+
+}  // namespace libaom_test
+#endif  // AV1_TXFM_TEST_H_
diff --git a/third_party/aom/test/av1_wedge_utils_test.cc b/third_party/aom/test/av1_wedge_utils_test.cc
new file mode 100644
index 0000000000..d4b560fc10
--- /dev/null
+++ b/third_party/aom/test/av1_wedge_utils_test.cc
@@ -0,0 +1,383 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "./av1_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+
+#include "av1/common/enums.h"
+
+#include "test/acm_random.h"
+#include "test/function_equivalence_test.h"
+#include "test/register_state_check.h"
+
+#define WEDGE_WEIGHT_BITS 6
+#define MAX_MASK_VALUE (1 << (WEDGE_WEIGHT_BITS))
+
+using libaom_test::ACMRandom;
+using libaom_test::FunctionEquivalenceTest;
+
+namespace {
+
+static const int16_t kInt13Max = (1 << 12) - 1;
+
+//////////////////////////////////////////////////////////////////////////////
+// av1_wedge_sse_from_residuals - functionality
+//////////////////////////////////////////////////////////////////////////////
+
+class WedgeUtilsSSEFuncTest : public testing::Test {
+ protected:
+  WedgeUtilsSSEFuncTest() : rng_(ACMRandom::DeterministicSeed()) {}
+
+  static const int kIterations = 1000;
+
+  ACMRandom rng_;
+};
+
+static void equiv_blend_residuals(int16_t *r, const int16_t *r0,
+                                  const int16_t *r1, const uint8_t *m, int N) {
+  for (int i = 0; i < N; i++) {
+    const int32_t m0 = m[i];
+    const int32_t m1 = MAX_MASK_VALUE - m0;
+    const int16_t R = m0 * r0[i] + m1 * r1[i];
+    // Note that this rounding is designed to match the result
+    // you would get when actually blending the 2 predictors and computing
+    // the residuals.
+    r[i] = ROUND_POWER_OF_TWO(R - 1, WEDGE_WEIGHT_BITS);
+  }
+}
+
+static uint64_t equiv_sse_from_residuals(const int16_t *r0, const int16_t *r1,
+                                         const uint8_t *m, int N) {
+  uint64_t acc = 0;
+  for (int i = 0; i < N; i++) {
+    const int32_t m0 = m[i];
+    const int32_t m1 = MAX_MASK_VALUE - m0;
+    const int16_t R = m0 * r0[i] + m1 * r1[i];
+    const int32_t r = ROUND_POWER_OF_TWO(R - 1, WEDGE_WEIGHT_BITS);
+    acc += r * r;
+  }
+  return acc;
+}
+
+TEST_F(WedgeUtilsSSEFuncTest, ResidualBlendingEquiv) {
+  DECLARE_ALIGNED(32, uint8_t, s[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, uint8_t, p0[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, uint8_t, p1[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, uint8_t, p[MAX_SB_SQUARE]);
+
+  DECLARE_ALIGNED(32, int16_t, r0[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, r1[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, r_ref[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, r_tst[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, uint8_t, m[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      s[i] = rng_.Rand8();
+      m[i] = rng_(MAX_MASK_VALUE + 1);
+    }
+
+    const int w = 1 << (rng_(MAX_SB_SIZE_LOG2 + 1 - 3) + 3);
+    const int h = 1 << (rng_(MAX_SB_SIZE_LOG2 + 1 - 3) + 3);
+    const int N = w * h;
+
+    for (int j = 0; j < N; j++) {
+      p0[j] = clamp(s[j] + rng_(33) - 16, 0, UINT8_MAX);
+      p1[j] = clamp(s[j] + rng_(33) - 16, 0, UINT8_MAX);
+    }
+
+    aom_blend_a64_mask(p, w, p0, w, p1, w, m, w, h, w, 0, 0);
+
+    aom_subtract_block(h, w, r0, w, s, w, p0, w);
+    aom_subtract_block(h, w, r1, w, s, w, p1, w);
+
+    aom_subtract_block(h, w, r_ref, w, s, w, p, w);
+    equiv_blend_residuals(r_tst, r0, r1, m, N);
+
+    for (int i = 0; i < N; ++i) ASSERT_EQ(r_ref[i], r_tst[i]);
+
+    uint64_t ref_sse = aom_sum_squares_i16(r_ref, N);
+    uint64_t tst_sse = equiv_sse_from_residuals(r0, r1, m, N);
+
+    ASSERT_EQ(ref_sse, tst_sse);
+  }
+}
+
+static uint64_t sse_from_residuals(const int16_t *r0, const int16_t *r1,
+                                   const uint8_t *m, int N) {
+  uint64_t acc = 0;
+  for (int i = 0; i < N; i++) {
+    const int32_t m0 = m[i];
+    const int32_t m1 = MAX_MASK_VALUE - m0;
+    const int32_t r = m0 * r0[i] + m1 * r1[i];
+    acc += r * r;
+  }
+  return ROUND_POWER_OF_TWO(acc, 2 * WEDGE_WEIGHT_BITS);
+}
+
+TEST_F(WedgeUtilsSSEFuncTest, ResidualBlendingMethod) {
+  DECLARE_ALIGNED(32, int16_t, r0[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, r1[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, d[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, uint8_t, m[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      r1[i] = rng_(2 * INT8_MAX - 2 * INT8_MIN + 1) + 2 * INT8_MIN;
+      d[i] = rng_(2 * INT8_MAX - 2 * INT8_MIN + 1) + 2 * INT8_MIN;
+      m[i] = rng_(MAX_MASK_VALUE + 1);
+    }
+
+    const int N = 64 * (rng_(MAX_SB_SQUARE / 64) + 1);
+
+    for (int i = 0; i < N; i++) r0[i] = r1[i] + d[i];
+
+    const uint64_t ref_res = sse_from_residuals(r0, r1, m, N);
+    const uint64_t tst_res = av1_wedge_sse_from_residuals(r1, d, m, N);
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// av1_wedge_sse_from_residuals - optimizations
+//////////////////////////////////////////////////////////////////////////////
+
+typedef uint64_t (*FSSE)(const int16_t *r1, const int16_t *d, const uint8_t *m,
+                         int N);
+typedef libaom_test::FuncParam<FSSE> TestFuncsFSSE;
+
+class WedgeUtilsSSEOptTest : public FunctionEquivalenceTest<FSSE> {
+ protected:
+  static const int kIterations = 10000;
+};
+
+TEST_P(WedgeUtilsSSEOptTest, RandomValues) {
+  DECLARE_ALIGNED(32, int16_t, r1[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, d[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, uint8_t, m[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      r1[i] = rng_(2 * kInt13Max + 1) - kInt13Max;
+      d[i] = rng_(2 * kInt13Max + 1) - kInt13Max;
+      m[i] = rng_(MAX_MASK_VALUE + 1);
+    }
+
+    const int N = 64 * (rng_(MAX_SB_SQUARE / 64) + 1);
+
+    const uint64_t ref_res = params_.ref_func(r1, d, m, N);
+    uint64_t tst_res;
+    ASM_REGISTER_STATE_CHECK(tst_res = params_.tst_func(r1, d, m, N));
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+TEST_P(WedgeUtilsSSEOptTest, ExtremeValues) {
+  DECLARE_ALIGNED(32, int16_t, r1[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, d[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, uint8_t, m[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    if (rng_(2)) {
+      for (int i = 0; i < MAX_SB_SQUARE; ++i) r1[i] = kInt13Max;
+    } else {
+      for (int i = 0; i < MAX_SB_SQUARE; ++i) r1[i] = -kInt13Max;
+    }
+
+    if (rng_(2)) {
+      for (int i = 0; i < MAX_SB_SQUARE; ++i) d[i] = kInt13Max;
+    } else {
+      for (int i = 0; i < MAX_SB_SQUARE; ++i) d[i] = -kInt13Max;
+    }
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) m[i] = MAX_MASK_VALUE;
+
+    const int N = 64 * (rng_(MAX_SB_SQUARE / 64) + 1);
+
+    const uint64_t ref_res = params_.ref_func(r1, d, m, N);
+    uint64_t tst_res;
+    ASM_REGISTER_STATE_CHECK(tst_res = params_.tst_func(r1, d, m, N));
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, WedgeUtilsSSEOptTest,
+    ::testing::Values(TestFuncsFSSE(av1_wedge_sse_from_residuals_c,
+                                    av1_wedge_sse_from_residuals_sse2)));
+
+#endif  // HAVE_SSE2
+
+//////////////////////////////////////////////////////////////////////////////
+// av1_wedge_sign_from_residuals
+//////////////////////////////////////////////////////////////////////////////
+
+typedef int (*FSign)(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
+typedef libaom_test::FuncParam<FSign> TestFuncsFSign;
+
+class WedgeUtilsSignOptTest : public FunctionEquivalenceTest<FSign> {
+ protected:
+  static const int kIterations = 10000;
+  static const int kMaxSize = 8196;  // Size limited by SIMD implementation.
+};
+
+TEST_P(WedgeUtilsSignOptTest, RandomValues) {
+  DECLARE_ALIGNED(32, int16_t, r0[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, r1[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, ds[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, uint8_t, m[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      r0[i] = rng_(2 * kInt13Max + 1) - kInt13Max;
+      r1[i] = rng_(2 * kInt13Max + 1) - kInt13Max;
+      m[i] = rng_(MAX_MASK_VALUE + 1);
+    }
+
+    const int maxN = AOMMIN(kMaxSize, MAX_SB_SQUARE);
+    const int N = 64 * (rng_(maxN / 64 - 1) + 1);
+
+    int64_t limit;
+    limit = (int64_t)aom_sum_squares_i16(r0, N);
+    limit -= (int64_t)aom_sum_squares_i16(r1, N);
+    limit *= (1 << WEDGE_WEIGHT_BITS) / 2;
+
+    for (int i = 0; i < N; i++)
+      ds[i] = clamp(r0[i] * r0[i] - r1[i] * r1[i], INT16_MIN, INT16_MAX);
+
+    const int ref_res = params_.ref_func(ds, m, N, limit);
+    int tst_res;
+    ASM_REGISTER_STATE_CHECK(tst_res = params_.tst_func(ds, m, N, limit));
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+TEST_P(WedgeUtilsSignOptTest, ExtremeValues) {
+  DECLARE_ALIGNED(32, int16_t, r0[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, r1[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, ds[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, uint8_t, m[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    switch (rng_(4)) {
+      case 0:
+        for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+          r0[i] = 0;
+          r1[i] = kInt13Max;
+        }
+        break;
+      case 1:
+        for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+          r0[i] = kInt13Max;
+          r1[i] = 0;
+        }
+        break;
+      case 2:
+        for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+          r0[i] = 0;
+          r1[i] = -kInt13Max;
+        }
+        break;
+      default:
+        for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+          r0[i] = -kInt13Max;
+          r1[i] = 0;
+        }
+        break;
+    }
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) m[i] = MAX_MASK_VALUE;
+
+    const int maxN = AOMMIN(kMaxSize, MAX_SB_SQUARE);
+    const int N = 64 * (rng_(maxN / 64 - 1) + 1);
+
+    int64_t limit;
+    limit = (int64_t)aom_sum_squares_i16(r0, N);
+    limit -= (int64_t)aom_sum_squares_i16(r1, N);
+    limit *= (1 << WEDGE_WEIGHT_BITS) / 2;
+
+    for (int i = 0; i < N; i++)
+      ds[i] = clamp(r0[i] * r0[i] - r1[i] * r1[i], INT16_MIN, INT16_MAX);
+
+    const int ref_res = params_.ref_func(ds, m, N, limit);
+    int tst_res;
+    ASM_REGISTER_STATE_CHECK(tst_res = params_.tst_func(ds, m, N, limit));
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+#if HAVE_SSE2
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, WedgeUtilsSignOptTest,
+    ::testing::Values(TestFuncsFSign(av1_wedge_sign_from_residuals_c,
+                                     av1_wedge_sign_from_residuals_sse2)));
+
+#endif  // HAVE_SSE2
+
+//////////////////////////////////////////////////////////////////////////////
+// av1_wedge_compute_delta_squares
+//////////////////////////////////////////////////////////////////////////////
+
+typedef void (*FDS)(int16_t *d, const int16_t *a, const int16_t *b, int N);
+typedef libaom_test::FuncParam<FDS> TestFuncsFDS;
+
+class WedgeUtilsDeltaSquaresOptTest : public FunctionEquivalenceTest<FDS> {
+ protected:
+  static const int kIterations = 10000;
+};
+
+TEST_P(WedgeUtilsDeltaSquaresOptTest, RandomValues) {
+  DECLARE_ALIGNED(32, int16_t, a[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, b[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, d_ref[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int16_t, d_tst[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      a[i] = rng_.Rand16();
+      b[i] = rng_(2 * INT16_MAX + 1) - INT16_MAX;
+    }
+
+    const int N = 64 * (rng_(MAX_SB_SQUARE / 64) + 1);
+
+    memset(&d_ref, INT16_MAX, sizeof(d_ref));
+    memset(&d_tst, INT16_MAX, sizeof(d_tst));
+
+    params_.ref_func(d_ref, a, b, N);
+    ASM_REGISTER_STATE_CHECK(params_.tst_func(d_tst, a, b, N));
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) ASSERT_EQ(d_ref[i], d_tst[i]);
+  }
+}
+
+#if HAVE_SSE2
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, WedgeUtilsDeltaSquaresOptTest,
+    ::testing::Values(TestFuncsFDS(av1_wedge_compute_delta_squares_c,
+                                   av1_wedge_compute_delta_squares_sse2)));
+
+#endif  // HAVE_SSE2
+
+}  // namespace
diff --git a/third_party/aom/test/avg_test.cc b/third_party/aom/test/avg_test.cc
new file mode 100644
index 0000000000..b040f6a34c
--- /dev/null
+++ b/third_party/aom/test/avg_test.cc
@@ -0,0 +1,396 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <limits.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "aom_mem/aom_mem.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+class AverageTestBase : public ::testing::Test {
+ public:
+  AverageTestBase(int width, int height) : width_(width), height_(height) {}
+
+  static void SetUpTestCase() {
+    source_data_ = reinterpret_cast<uint8_t *>(
+        aom_memalign(kDataAlignment, kDataBlockSize));
+  }
+
+  static void TearDownTestCase() {
+    aom_free(source_data_);
+    source_data_ = NULL;
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  // Handle blocks up to 4 blocks 64x64 with stride up to 128
+  static const int kDataAlignment = 16;
+  static const int kDataBlockSize = 64 * 128;
+
+  virtual void SetUp() {
+    source_stride_ = (width_ + 31) & ~31;
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+  }
+
+  // Sum Pixels
+  static unsigned int ReferenceAverage8x8(const uint8_t *source, int pitch) {
+    unsigned int average = 0;
+    for (int h = 0; h < 8; ++h)
+      for (int w = 0; w < 8; ++w) average += source[h * pitch + w];
+    return ((average + 32) >> 6);
+  }
+
+  static unsigned int ReferenceAverage4x4(const uint8_t *source, int pitch) {
+    unsigned int average = 0;
+    for (int h = 0; h < 4; ++h)
+      for (int w = 0; w < 4; ++w) average += source[h * pitch + w];
+    return ((average + 8) >> 4);
+  }
+
+  void FillConstant(uint8_t fill_constant) {
+    for (int i = 0; i < width_ * height_; ++i) {
+      source_data_[i] = fill_constant;
+    }
+  }
+
+  void FillRandom() {
+    for (int i = 0; i < width_ * height_; ++i) {
+      source_data_[i] = rnd_.Rand8();
+    }
+  }
+
+  int width_, height_;
+  static uint8_t *source_data_;
+  int source_stride_;
+
+  ACMRandom rnd_;
+};
+typedef unsigned int (*AverageFunction)(const uint8_t *s, int pitch);
+
+typedef std::tr1::tuple<int, int, int, int, AverageFunction> AvgFunc;
+
+class AverageTest : public AverageTestBase,
+                    public ::testing::WithParamInterface<AvgFunc> {
+ public:
+  AverageTest() : AverageTestBase(GET_PARAM(0), GET_PARAM(1)) {}
+
+ protected:
+  void CheckAverages() {
+    const int block_size = GET_PARAM(3);
+    unsigned int expected = 0;
+    if (block_size == 8) {
+      expected =
+          ReferenceAverage8x8(source_data_ + GET_PARAM(2), source_stride_);
+    } else if (block_size == 4) {
+      expected =
+          ReferenceAverage4x4(source_data_ + GET_PARAM(2), source_stride_);
+    }
+
+    ASM_REGISTER_STATE_CHECK(
+        GET_PARAM(4)(source_data_ + GET_PARAM(2), source_stride_));
+    unsigned int actual =
+        GET_PARAM(4)(source_data_ + GET_PARAM(2), source_stride_);
+
+    EXPECT_EQ(expected, actual);
+  }
+};
+
+typedef void (*IntProRowFunc)(int16_t hbuf[16], uint8_t const *ref,
+                              const int ref_stride, const int height);
+
+typedef std::tr1::tuple<int, IntProRowFunc, IntProRowFunc> IntProRowParam;
+
+class IntProRowTest : public AverageTestBase,
+                      public ::testing::WithParamInterface<IntProRowParam> {
+ public:
+  IntProRowTest()
+      : AverageTestBase(16, GET_PARAM(0)), hbuf_asm_(NULL), hbuf_c_(NULL) {
+    asm_func_ = GET_PARAM(1);
+    c_func_ = GET_PARAM(2);
+  }
+
+ protected:
+  virtual void SetUp() {
+    hbuf_asm_ = reinterpret_cast<int16_t *>(
+        aom_memalign(kDataAlignment, sizeof(*hbuf_asm_) * 16));
+    hbuf_c_ = reinterpret_cast<int16_t *>(
+        aom_memalign(kDataAlignment, sizeof(*hbuf_c_) * 16));
+  }
+
+  virtual void TearDown() {
+    aom_free(hbuf_c_);
+    hbuf_c_ = NULL;
+    aom_free(hbuf_asm_);
+    hbuf_asm_ = NULL;
+  }
+
+  void RunComparison() {
+    ASM_REGISTER_STATE_CHECK(c_func_(hbuf_c_, source_data_, 0, height_));
+    ASM_REGISTER_STATE_CHECK(asm_func_(hbuf_asm_, source_data_, 0, height_));
+    EXPECT_EQ(0, memcmp(hbuf_c_, hbuf_asm_, sizeof(*hbuf_c_) * 16))
+        << "Output mismatch";
+  }
+
+ private:
+  IntProRowFunc asm_func_;
+  IntProRowFunc c_func_;
+  int16_t *hbuf_asm_;
+  int16_t *hbuf_c_;
+};
+
+typedef int16_t (*IntProColFunc)(uint8_t const *ref, const int width);
+
+typedef std::tr1::tuple<int, IntProColFunc, IntProColFunc> IntProColParam;
+
+class IntProColTest : public AverageTestBase,
+                      public ::testing::WithParamInterface<IntProColParam> {
+ public:
+  IntProColTest() : AverageTestBase(GET_PARAM(0), 1), sum_asm_(0), sum_c_(0) {
+    asm_func_ = GET_PARAM(1);
+    c_func_ = GET_PARAM(2);
+  }
+
+ protected:
+  void RunComparison() {
+    ASM_REGISTER_STATE_CHECK(sum_c_ = c_func_(source_data_, width_));
+    ASM_REGISTER_STATE_CHECK(sum_asm_ = asm_func_(source_data_, width_));
+    EXPECT_EQ(sum_c_, sum_asm_) << "Output mismatch";
+  }
+
+ private:
+  IntProColFunc asm_func_;
+  IntProColFunc c_func_;
+  int16_t sum_asm_;
+  int16_t sum_c_;
+};
+
+typedef int (*SatdFunc)(const int16_t *coeffs, int length);
+typedef std::tr1::tuple<int, SatdFunc> SatdTestParam;
+
+class SatdTest : public ::testing::Test,
+                 public ::testing::WithParamInterface<SatdTestParam> {
+ protected:
+  virtual void SetUp() {
+    satd_size_ = GET_PARAM(0);
+    satd_func_ = GET_PARAM(1);
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+    src_ = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(*src_) * satd_size_));
+    ASSERT_TRUE(src_ != NULL);
+  }
+
+  virtual void TearDown() {
+    libaom_test::ClearSystemState();
+    aom_free(src_);
+  }
+
+  void FillConstant(const int16_t val) {
+    for (int i = 0; i < satd_size_; ++i) src_[i] = val;
+  }
+
+  void FillRandom() {
+    for (int i = 0; i < satd_size_; ++i) src_[i] = rnd_.Rand16();
+  }
+
+  void Check(int expected) {
+    int total;
+    ASM_REGISTER_STATE_CHECK(total = satd_func_(src_, satd_size_));
+    EXPECT_EQ(expected, total);
+  }
+
+  int satd_size_;
+
+ private:
+  int16_t *src_;
+  SatdFunc satd_func_;
+  ACMRandom rnd_;
+};
+
+uint8_t *AverageTestBase::source_data_ = NULL;
+
+TEST_P(AverageTest, MinValue) {
+  FillConstant(0);
+  CheckAverages();
+}
+
+TEST_P(AverageTest, MaxValue) {
+  FillConstant(255);
+  CheckAverages();
+}
+
+TEST_P(AverageTest, Random) {
+  // The reference frame, but not the source frame, may be unaligned for
+  // certain types of searches.
+  for (int i = 0; i < 1000; i++) {
+    FillRandom();
+    CheckAverages();
+  }
+}
+
+TEST_P(IntProRowTest, MinValue) {
+  FillConstant(0);
+  RunComparison();
+}
+
+TEST_P(IntProRowTest, MaxValue) {
+  FillConstant(255);
+  RunComparison();
+}
+
+TEST_P(IntProRowTest, Random) {
+  FillRandom();
+  RunComparison();
+}
+
+TEST_P(IntProColTest, MinValue) {
+  FillConstant(0);
+  RunComparison();
+}
+
+TEST_P(IntProColTest, MaxValue) {
+  FillConstant(255);
+  RunComparison();
+}
+
+TEST_P(IntProColTest, Random) {
+  FillRandom();
+  RunComparison();
+}
+
+TEST_P(SatdTest, MinValue) {
+  const int kMin = -32640;
+  const int expected = -kMin * satd_size_;
+  FillConstant(kMin);
+  Check(expected);
+}
+
+TEST_P(SatdTest, MaxValue) {
+  const int kMax = 32640;
+  const int expected = kMax * satd_size_;
+  FillConstant(kMax);
+  Check(expected);
+}
+
+TEST_P(SatdTest, Random) {
+  int expected;
+  switch (satd_size_) {
+    case 16: expected = 205298; break;
+    case 64: expected = 1113950; break;
+    case 256: expected = 4268415; break;
+    case 1024: expected = 16954082; break;
+    default:
+      FAIL() << "Invalid satd size (" << satd_size_
+             << ") valid: 16/64/256/1024";
+  }
+  FillRandom();
+  Check(expected);
+}
+
+using std::tr1::make_tuple;
+
+INSTANTIATE_TEST_CASE_P(
+    C, AverageTest,
+    ::testing::Values(make_tuple(16, 16, 1, 8, &aom_avg_8x8_c),
+                      make_tuple(16, 16, 1, 4, &aom_avg_4x4_c)));
+
+INSTANTIATE_TEST_CASE_P(C, SatdTest,
+                        ::testing::Values(make_tuple(16, &aom_satd_c),
+                                          make_tuple(64, &aom_satd_c),
+                                          make_tuple(256, &aom_satd_c),
+                                          make_tuple(1024, &aom_satd_c)));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, AverageTest,
+    ::testing::Values(make_tuple(16, 16, 0, 8, &aom_avg_8x8_sse2),
+                      make_tuple(16, 16, 5, 8, &aom_avg_8x8_sse2),
+                      make_tuple(32, 32, 15, 8, &aom_avg_8x8_sse2),
+                      make_tuple(16, 16, 0, 4, &aom_avg_4x4_sse2),
+                      make_tuple(16, 16, 5, 4, &aom_avg_4x4_sse2),
+                      make_tuple(32, 32, 15, 4, &aom_avg_4x4_sse2)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, IntProRowTest,
+    ::testing::Values(make_tuple(16, &aom_int_pro_row_sse2, &aom_int_pro_row_c),
+                      make_tuple(32, &aom_int_pro_row_sse2, &aom_int_pro_row_c),
+                      make_tuple(64, &aom_int_pro_row_sse2,
+                                 &aom_int_pro_row_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, IntProColTest,
+    ::testing::Values(make_tuple(16, &aom_int_pro_col_sse2, &aom_int_pro_col_c),
+                      make_tuple(32, &aom_int_pro_col_sse2, &aom_int_pro_col_c),
+                      make_tuple(64, &aom_int_pro_col_sse2,
+                                 &aom_int_pro_col_c)));
+
+INSTANTIATE_TEST_CASE_P(SSE2, SatdTest,
+                        ::testing::Values(make_tuple(16, &aom_satd_sse2),
+                                          make_tuple(64, &aom_satd_sse2),
+                                          make_tuple(256, &aom_satd_sse2),
+                                          make_tuple(1024, &aom_satd_sse2)));
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+    NEON, AverageTest,
+    ::testing::Values(make_tuple(16, 16, 0, 8, &aom_avg_8x8_neon),
+                      make_tuple(16, 16, 5, 8, &aom_avg_8x8_neon),
+                      make_tuple(32, 32, 15, 8, &aom_avg_8x8_neon),
+                      make_tuple(16, 16, 0, 4, &aom_avg_4x4_neon),
+                      make_tuple(16, 16, 5, 4, &aom_avg_4x4_neon),
+                      make_tuple(32, 32, 15, 4, &aom_avg_4x4_neon)));
+
+INSTANTIATE_TEST_CASE_P(
+    NEON, IntProRowTest,
+    ::testing::Values(make_tuple(16, &aom_int_pro_row_neon, &aom_int_pro_row_c),
+                      make_tuple(32, &aom_int_pro_row_neon, &aom_int_pro_row_c),
+                      make_tuple(64, &aom_int_pro_row_neon,
+                                 &aom_int_pro_row_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    NEON, IntProColTest,
+    ::testing::Values(make_tuple(16, &aom_int_pro_col_neon, &aom_int_pro_col_c),
+                      make_tuple(32, &aom_int_pro_col_neon, &aom_int_pro_col_c),
+                      make_tuple(64, &aom_int_pro_col_neon,
+                                 &aom_int_pro_col_c)));
+
+INSTANTIATE_TEST_CASE_P(NEON, SatdTest,
+                        ::testing::Values(make_tuple(16, &aom_satd_neon),
+                                          make_tuple(64, &aom_satd_neon),
+                                          make_tuple(256, &aom_satd_neon),
+                                          make_tuple(1024, &aom_satd_neon)));
+#endif
+
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(
+    MSA, AverageTest,
+    ::testing::Values(make_tuple(16, 16, 0, 8, &aom_avg_8x8_msa),
+                      make_tuple(16, 16, 5, 8, &aom_avg_8x8_msa),
+                      make_tuple(32, 32, 15, 8, &aom_avg_8x8_msa),
+                      make_tuple(16, 16, 0, 4, &aom_avg_4x4_msa),
+                      make_tuple(16, 16, 5, 4, &aom_avg_4x4_msa),
+                      make_tuple(32, 32, 15, 4, &aom_avg_4x4_msa)));
+#endif
+
+}  // namespace
diff --git a/third_party/aom/test/binary_codes_test.cc b/third_party/aom/test/binary_codes_test.cc
new file mode 100644
index 0000000000..385ec76871
--- /dev/null
+++ b/third_party/aom/test/binary_codes_test.cc
@@ -0,0 +1,130 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/acm_random.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/bitreader.h"
+#include "aom_dsp/bitwriter.h"
+#include "aom_dsp/binary_codes_reader.h"
+#include "aom_dsp/binary_codes_writer.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+
+// Test for Bilevel code with reference
+TEST(AV1, TestPrimitiveRefbilivel) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int kBufferSize = 65536;
+  aom_writer bw;
+  uint8_t bw_buffer[kBufferSize];
+  const uint16_t kRanges = 8;
+  const uint16_t kNearRanges = 8;
+  const uint16_t kReferences = 8;
+  const uint16_t kValues = 16;
+  const uint16_t range_vals[kRanges] = { 1, 13, 64, 120, 230, 420, 1100, 8000 };
+  uint16_t enc_values[kRanges][kNearRanges][kReferences][kValues][4];
+  aom_start_encode(&bw, bw_buffer);
+  for (int n = 0; n < kRanges; ++n) {
+    const uint16_t range = range_vals[n];
+    for (int p = 0; p < kNearRanges; ++p) {
+      const uint16_t near_range = 1 + rnd(range);
+      for (int r = 0; r < kReferences; ++r) {
+        const uint16_t ref = rnd(range);
+        for (int v = 0; v < kValues; ++v) {
+          const uint16_t value = rnd(range);
+          enc_values[n][p][r][v][0] = range;
+          enc_values[n][p][r][v][1] = near_range;
+          enc_values[n][p][r][v][2] = ref;
+          enc_values[n][p][r][v][3] = value;
+          aom_write_primitive_refbilevel(&bw, range, near_range, ref, value);
+        }
+      }
+    }
+  }
+  aom_stop_encode(&bw);
+  aom_reader br;
+  aom_reader_init(&br, bw_buffer, bw.pos, NULL, NULL);
+  GTEST_ASSERT_GE(aom_reader_tell(&br), 0u);
+  GTEST_ASSERT_LE(aom_reader_tell(&br), 1u);
+  for (int n = 0; n < kRanges; ++n) {
+    for (int p = 0; p < kNearRanges; ++p) {
+      for (int r = 0; r < kReferences; ++r) {
+        for (int v = 0; v < kValues; ++v) {
+          const uint16_t range = enc_values[n][p][r][v][0];
+          const uint16_t near_range = enc_values[n][p][r][v][1];
+          const uint16_t ref = enc_values[n][p][r][v][2];
+          const uint16_t value =
+              aom_read_primitive_refbilevel(&br, range, near_range, ref);
+          GTEST_ASSERT_EQ(value, enc_values[n][p][r][v][3]);
+        }
+      }
+    }
+  }
+}
+
+// Test for Finite subexponential code with reference
+TEST(AV1, TestPrimitiveRefsubexpfin) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int kBufferSize = 65536;
+  aom_writer bw;
+  uint8_t bw_buffer[kBufferSize];
+  const uint16_t kRanges = 8;
+  const uint16_t kSubexpParams = 6;
+  const uint16_t kReferences = 8;
+  const uint16_t kValues = 16;
+  uint16_t enc_values[kRanges][kSubexpParams][kReferences][kValues][4];
+  const uint16_t range_vals[kRanges] = { 1, 13, 64, 120, 230, 420, 1100, 8000 };
+  aom_start_encode(&bw, bw_buffer);
+  for (int n = 0; n < kRanges; ++n) {
+    const uint16_t range = range_vals[n];
+    for (int k = 0; k < kSubexpParams; ++k) {
+      for (int r = 0; r < kReferences; ++r) {
+        const uint16_t ref = rnd(range);
+        for (int v = 0; v < kValues; ++v) {
+          const uint16_t value = rnd(range);
+          enc_values[n][k][r][v][0] = range;
+          enc_values[n][k][r][v][1] = k;
+          enc_values[n][k][r][v][2] = ref;
+          enc_values[n][k][r][v][3] = value;
+          aom_write_primitive_refsubexpfin(&bw, range, k, ref, value);
+        }
+      }
+    }
+  }
+  aom_stop_encode(&bw);
+  aom_reader br;
+  aom_reader_init(&br, bw_buffer, bw.pos, NULL, NULL);
+  GTEST_ASSERT_GE(aom_reader_tell(&br), 0u);
+  GTEST_ASSERT_LE(aom_reader_tell(&br), 1u);
+  for (int n = 0; n < kRanges; ++n) {
+    for (int k = 0; k < kSubexpParams; ++k) {
+      for (int r = 0; r < kReferences; ++r) {
+        for (int v = 0; v < kValues; ++v) {
+          const uint16_t range = enc_values[n][k][r][v][0];
+          assert(k == enc_values[n][k][r][v][1]);
+          const uint16_t ref = enc_values[n][k][r][v][2];
+          const uint16_t value =
+              aom_read_primitive_refsubexpfin(&br, range, k, ref);
+          GTEST_ASSERT_EQ(value, enc_values[n][k][r][v][3]);
+        }
+      }
+    }
+  }
+}
+// TODO(debargha): Adds tests for other primitives
+}  // namespace
diff --git a/third_party/aom/test/blend_a64_mask_1d_test.cc b/third_party/aom/test/blend_a64_mask_1d_test.cc
new file mode 100644
index 0000000000..66e741a74d
--- /dev/null
+++ b/third_party/aom/test/blend_a64_mask_1d_test.cc
@@ -0,0 +1,334 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/register_state_check.h"
+#include "test/function_equivalence_test.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+#include "./av1_rtcd.h"
+
+#include "av1/common/enums.h"
+
+#include "aom_dsp/blend.h"
+
+using libaom_test::FunctionEquivalenceTest;
+
+namespace {
+
+template <typename F, typename T>
+class BlendA64Mask1DTest : public FunctionEquivalenceTest<F> {
+ public:
+  static const int kIterations = 10000;
+  static const int kMaxWidth = MAX_SB_SIZE * 5;  // * 5 to cover longer strides
+  static const int kMaxHeight = MAX_SB_SIZE;
+  static const int kBufSize = kMaxWidth * kMaxHeight;
+  static const int kMaxMaskWidth = 2 * MAX_SB_SIZE;
+  static const int kMaxMaskSize = kMaxMaskWidth;
+
+  virtual ~BlendA64Mask1DTest() {}
+
+  virtual void Execute(const T *p_src0, const T *p_src1) = 0;
+
+  void Common() {
+    w_ = 1 << this->rng_(MAX_SB_SIZE_LOG2 + 1);
+    h_ = 1 << this->rng_(MAX_SB_SIZE_LOG2 + 1);
+
+    dst_offset_ = this->rng_(33);
+    dst_stride_ = this->rng_(kMaxWidth + 1 - w_) + w_;
+
+    src0_offset_ = this->rng_(33);
+    src0_stride_ = this->rng_(kMaxWidth + 1 - w_) + w_;
+
+    src1_offset_ = this->rng_(33);
+    src1_stride_ = this->rng_(kMaxWidth + 1 - w_) + w_;
+
+    T *p_src0;
+    T *p_src1;
+
+    switch (this->rng_(3)) {
+      case 0:  // Separate sources
+        p_src0 = src0_;
+        p_src1 = src1_;
+        break;
+      case 1:  // src0 == dst
+        p_src0 = dst_tst_;
+        src0_stride_ = dst_stride_;
+        src0_offset_ = dst_offset_;
+        p_src1 = src1_;
+        break;
+      case 2:  // src1 == dst
+        p_src0 = src0_;
+        p_src1 = dst_tst_;
+        src1_stride_ = dst_stride_;
+        src1_offset_ = dst_offset_;
+        break;
+      default: FAIL();
+    }
+
+    Execute(p_src0, p_src1);
+
+    for (int r = 0; r < h_; ++r) {
+      for (int c = 0; c < w_; ++c) {
+        ASSERT_EQ(dst_ref_[dst_offset_ + r * dst_stride_ + c],
+                  dst_tst_[dst_offset_ + r * dst_stride_ + c]);
+      }
+    }
+  }
+
+  T dst_ref_[kBufSize];
+  T dst_tst_[kBufSize];
+  uint32_t dst_stride_;
+  uint32_t dst_offset_;
+
+  T src0_[kBufSize];
+  uint32_t src0_stride_;
+  uint32_t src0_offset_;
+
+  T src1_[kBufSize];
+  uint32_t src1_stride_;
+  uint32_t src1_offset_;
+
+  uint8_t mask_[kMaxMaskSize];
+
+  int w_;
+  int h_;
+};
+
+//////////////////////////////////////////////////////////////////////////////
+// 8 bit version
+//////////////////////////////////////////////////////////////////////////////
+
+typedef void (*F8B)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+                    uint32_t src0_stride, const uint8_t *src1,
+                    uint32_t src1_stride, const uint8_t *mask, int h, int w);
+typedef libaom_test::FuncParam<F8B> TestFuncs;
+
+class BlendA64Mask1DTest8B : public BlendA64Mask1DTest<F8B, uint8_t> {
+ protected:
+  void Execute(const uint8_t *p_src0, const uint8_t *p_src1) {
+    params_.ref_func(dst_ref_ + dst_offset_, dst_stride_, p_src0 + src0_offset_,
+                     src0_stride_, p_src1 + src1_offset_, src1_stride_, mask_,
+                     h_, w_);
+    ASM_REGISTER_STATE_CHECK(params_.tst_func(
+        dst_tst_ + dst_offset_, dst_stride_, p_src0 + src0_offset_,
+        src0_stride_, p_src1 + src1_offset_, src1_stride_, mask_, h_, w_));
+  }
+};
+
+TEST_P(BlendA64Mask1DTest8B, RandomValues) {
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    for (int i = 0; i < kBufSize; ++i) {
+      dst_ref_[i] = rng_.Rand8();
+      dst_tst_[i] = rng_.Rand8();
+
+      src0_[i] = rng_.Rand8();
+      src1_[i] = rng_.Rand8();
+    }
+
+    for (int i = 0; i < kMaxMaskSize; ++i)
+      mask_[i] = rng_(AOM_BLEND_A64_MAX_ALPHA + 1);
+
+    Common();
+  }
+}
+
+TEST_P(BlendA64Mask1DTest8B, ExtremeValues) {
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    for (int i = 0; i < kBufSize; ++i) {
+      dst_ref_[i] = rng_(2) + 254;
+      dst_tst_[i] = rng_(2) + 254;
+      src0_[i] = rng_(2) + 254;
+      src1_[i] = rng_(2) + 254;
+    }
+
+    for (int i = 0; i < kMaxMaskSize; ++i)
+      mask_[i] = rng_(2) + AOM_BLEND_A64_MAX_ALPHA - 1;
+
+    Common();
+  }
+}
+
+static void blend_a64_hmask_ref(uint8_t *dst, uint32_t dst_stride,
+                                const uint8_t *src0, uint32_t src0_stride,
+                                const uint8_t *src1, uint32_t src1_stride,
+                                const uint8_t *mask, int h, int w) {
+  uint8_t mask2d[BlendA64Mask1DTest8B::kMaxMaskSize]
+                [BlendA64Mask1DTest8B::kMaxMaskSize];
+
+  for (int row = 0; row < h; ++row)
+    for (int col = 0; col < w; ++col) mask2d[row][col] = mask[col];
+
+  aom_blend_a64_mask_c(dst, dst_stride, src0, src0_stride, src1, src1_stride,
+                       &mask2d[0][0], BlendA64Mask1DTest8B::kMaxMaskSize, h, w,
+                       0, 0);
+}
+
+static void blend_a64_vmask_ref(uint8_t *dst, uint32_t dst_stride,
+                                const uint8_t *src0, uint32_t src0_stride,
+                                const uint8_t *src1, uint32_t src1_stride,
+                                const uint8_t *mask, int h, int w) {
+  uint8_t mask2d[BlendA64Mask1DTest8B::kMaxMaskSize]
+                [BlendA64Mask1DTest8B::kMaxMaskSize];
+
+  for (int row = 0; row < h; ++row)
+    for (int col = 0; col < w; ++col) mask2d[row][col] = mask[row];
+
+  aom_blend_a64_mask_c(dst, dst_stride, src0, src0_stride, src1, src1_stride,
+                       &mask2d[0][0], BlendA64Mask1DTest8B::kMaxMaskSize, h, w,
+                       0, 0);
+}
+
+INSTANTIATE_TEST_CASE_P(
+    C, BlendA64Mask1DTest8B,
+    ::testing::Values(TestFuncs(blend_a64_hmask_ref, aom_blend_a64_hmask_c),
+                      TestFuncs(blend_a64_vmask_ref, aom_blend_a64_vmask_c)));
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, BlendA64Mask1DTest8B,
+    ::testing::Values(
+        TestFuncs(blend_a64_hmask_ref, aom_blend_a64_hmask_sse4_1),
+        TestFuncs(blend_a64_vmask_ref, aom_blend_a64_vmask_sse4_1)));
+#endif  // HAVE_SSE4_1
+
+#if CONFIG_HIGHBITDEPTH
+//////////////////////////////////////////////////////////////////////////////
+// High bit-depth version
+//////////////////////////////////////////////////////////////////////////////
+
+typedef void (*FHBD)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+                     uint32_t src0_stride, const uint8_t *src1,
+                     uint32_t src1_stride, const uint8_t *mask, int h, int w,
+                     int bd);
+typedef libaom_test::FuncParam<FHBD> TestFuncsHBD;
+
+class BlendA64Mask1DTestHBD : public BlendA64Mask1DTest<FHBD, uint16_t> {
+ protected:
+  void Execute(const uint16_t *p_src0, const uint16_t *p_src1) {
+    params_.ref_func(CONVERT_TO_BYTEPTR(dst_ref_ + dst_offset_), dst_stride_,
+                     CONVERT_TO_BYTEPTR(p_src0 + src0_offset_), src0_stride_,
+                     CONVERT_TO_BYTEPTR(p_src1 + src1_offset_), src1_stride_,
+                     mask_, h_, w_, bit_depth_);
+    ASM_REGISTER_STATE_CHECK(params_.tst_func(
+        CONVERT_TO_BYTEPTR(dst_tst_ + dst_offset_), dst_stride_,
+        CONVERT_TO_BYTEPTR(p_src0 + src0_offset_), src0_stride_,
+        CONVERT_TO_BYTEPTR(p_src1 + src1_offset_), src1_stride_, mask_, h_, w_,
+        bit_depth_));
+  }
+
+  int bit_depth_;
+};
+
+TEST_P(BlendA64Mask1DTestHBD, RandomValues) {
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    switch (rng_(3)) {
+      case 0: bit_depth_ = 8; break;
+      case 1: bit_depth_ = 10; break;
+      default: bit_depth_ = 12; break;
+    }
+
+    const int hi = 1 << bit_depth_;
+
+    for (int i = 0; i < kBufSize; ++i) {
+      dst_ref_[i] = rng_(hi);
+      dst_tst_[i] = rng_(hi);
+      src0_[i] = rng_(hi);
+      src1_[i] = rng_(hi);
+    }
+
+    for (int i = 0; i < kMaxMaskSize; ++i)
+      mask_[i] = rng_(AOM_BLEND_A64_MAX_ALPHA + 1);
+
+    Common();
+  }
+}
+
+TEST_P(BlendA64Mask1DTestHBD, ExtremeValues) {
+  for (int iter = 0; iter < 1000 && !HasFatalFailure(); ++iter) {
+    switch (rng_(3)) {
+      case 0: bit_depth_ = 8; break;
+      case 1: bit_depth_ = 10; break;
+      default: bit_depth_ = 12; break;
+    }
+
+    const int hi = 1 << bit_depth_;
+    const int lo = hi - 2;
+
+    for (int i = 0; i < kBufSize; ++i) {
+      dst_ref_[i] = rng_(hi - lo) + lo;
+      dst_tst_[i] = rng_(hi - lo) + lo;
+      src0_[i] = rng_(hi - lo) + lo;
+      src1_[i] = rng_(hi - lo) + lo;
+    }
+
+    for (int i = 0; i < kMaxMaskSize; ++i)
+      mask_[i] = rng_(2) + AOM_BLEND_A64_MAX_ALPHA - 1;
+
+    Common();
+  }
+}
+
+static void highbd_blend_a64_hmask_ref(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, int h, int w, int bd) {
+  uint8_t mask2d[BlendA64Mask1DTestHBD::kMaxMaskSize]
+                [BlendA64Mask1DTestHBD::kMaxMaskSize];
+
+  for (int row = 0; row < h; ++row)
+    for (int col = 0; col < w; ++col) mask2d[row][col] = mask[col];
+
+  aom_highbd_blend_a64_mask_c(
+      dst, dst_stride, src0, src0_stride, src1, src1_stride, &mask2d[0][0],
+      BlendA64Mask1DTestHBD::kMaxMaskSize, h, w, 0, 0, bd);
+}
+
+static void highbd_blend_a64_vmask_ref(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, int h, int w, int bd) {
+  uint8_t mask2d[BlendA64Mask1DTestHBD::kMaxMaskSize]
+                [BlendA64Mask1DTestHBD::kMaxMaskSize];
+
+  for (int row = 0; row < h; ++row)
+    for (int col = 0; col < w; ++col) mask2d[row][col] = mask[row];
+
+  aom_highbd_blend_a64_mask_c(
+      dst, dst_stride, src0, src0_stride, src1, src1_stride, &mask2d[0][0],
+      BlendA64Mask1DTestHBD::kMaxMaskSize, h, w, 0, 0, bd);
+}
+
+INSTANTIATE_TEST_CASE_P(
+    C, BlendA64Mask1DTestHBD,
+    ::testing::Values(TestFuncsHBD(highbd_blend_a64_hmask_ref,
+                                   aom_highbd_blend_a64_hmask_c),
+                      TestFuncsHBD(highbd_blend_a64_vmask_ref,
+                                   aom_highbd_blend_a64_vmask_c)));
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, BlendA64Mask1DTestHBD,
+    ::testing::Values(TestFuncsHBD(highbd_blend_a64_hmask_ref,
+                                   aom_highbd_blend_a64_hmask_sse4_1),
+                      TestFuncsHBD(highbd_blend_a64_vmask_ref,
+                                   aom_highbd_blend_a64_vmask_sse4_1)));
+#endif  // HAVE_SSE4_1
+
+#endif  // CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/blend_a64_mask_test.cc b/third_party/aom/test/blend_a64_mask_test.cc
new file mode 100644
index 0000000000..fef124d347
--- /dev/null
+++ b/third_party/aom/test/blend_a64_mask_test.cc
@@ -0,0 +1,270 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/register_state_check.h"
+#include "test/function_equivalence_test.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+#include "./av1_rtcd.h"
+
+#include "av1/common/enums.h"
+
+#include "aom_dsp/blend.h"
+
+using libaom_test::FunctionEquivalenceTest;
+
+namespace {
+
+template <typename F, typename T>
+class BlendA64MaskTest : public FunctionEquivalenceTest<F> {
+ protected:
+  static const int kIterations = 10000;
+  static const int kMaxWidth = MAX_SB_SIZE * 5;  // * 5 to cover longer strides
+  static const int kMaxHeight = MAX_SB_SIZE;
+  static const int kBufSize = kMaxWidth * kMaxHeight;
+  static const int kMaxMaskWidth = 2 * MAX_SB_SIZE;
+  static const int kMaxMaskSize = kMaxMaskWidth * kMaxMaskWidth;
+
+  virtual ~BlendA64MaskTest() {}
+
+  virtual void Execute(const T *p_src0, const T *p_src1) = 0;
+
+  void Common() {
+    w_ = 1 << this->rng_(MAX_SB_SIZE_LOG2 + 1);
+    h_ = 1 << this->rng_(MAX_SB_SIZE_LOG2 + 1);
+
+    subx_ = this->rng_(2);
+    suby_ = this->rng_(2);
+
+    dst_offset_ = this->rng_(33);
+    dst_stride_ = this->rng_(kMaxWidth + 1 - w_) + w_;
+
+    src0_offset_ = this->rng_(33);
+    src0_stride_ = this->rng_(kMaxWidth + 1 - w_) + w_;
+
+    src1_offset_ = this->rng_(33);
+    src1_stride_ = this->rng_(kMaxWidth + 1 - w_) + w_;
+
+    mask_stride_ =
+        this->rng_(kMaxWidth + 1 - w_ * (subx_ ? 2 : 1)) + w_ * (subx_ ? 2 : 1);
+
+    T *p_src0;
+    T *p_src1;
+
+    switch (this->rng_(3)) {
+      case 0:  // Separate sources
+        p_src0 = src0_;
+        p_src1 = src1_;
+        break;
+      case 1:  // src0 == dst
+        p_src0 = dst_tst_;
+        src0_stride_ = dst_stride_;
+        src0_offset_ = dst_offset_;
+        p_src1 = src1_;
+        break;
+      case 2:  // src1 == dst
+        p_src0 = src0_;
+        p_src1 = dst_tst_;
+        src1_stride_ = dst_stride_;
+        src1_offset_ = dst_offset_;
+        break;
+      default: FAIL();
+    }
+
+    Execute(p_src0, p_src1);
+
+    for (int r = 0; r < h_; ++r) {
+      for (int c = 0; c < w_; ++c) {
+        ASSERT_EQ(dst_ref_[dst_offset_ + r * dst_stride_ + c],
+                  dst_tst_[dst_offset_ + r * dst_stride_ + c]);
+      }
+    }
+  }
+
+  T dst_ref_[kBufSize];
+  T dst_tst_[kBufSize];
+  uint32_t dst_stride_;
+  uint32_t dst_offset_;
+
+  T src0_[kBufSize];
+  uint32_t src0_stride_;
+  uint32_t src0_offset_;
+
+  T src1_[kBufSize];
+  uint32_t src1_stride_;
+  uint32_t src1_offset_;
+
+  uint8_t mask_[kMaxMaskSize];
+  size_t mask_stride_;
+
+  int w_;
+  int h_;
+
+  int suby_;
+  int subx_;
+};
+
+//////////////////////////////////////////////////////////////////////////////
+// 8 bit version
+//////////////////////////////////////////////////////////////////////////////
+
+typedef void (*F8B)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+                    uint32_t src0_stride, const uint8_t *src1,
+                    uint32_t src1_stride, const uint8_t *mask,
+                    uint32_t mask_stride, int h, int w, int suby, int subx);
+typedef libaom_test::FuncParam<F8B> TestFuncs;
+
+class BlendA64MaskTest8B : public BlendA64MaskTest<F8B, uint8_t> {
+ protected:
+  void Execute(const uint8_t *p_src0, const uint8_t *p_src1) {
+    params_.ref_func(dst_ref_ + dst_offset_, dst_stride_, p_src0 + src0_offset_,
+                     src0_stride_, p_src1 + src1_offset_, src1_stride_, mask_,
+                     kMaxMaskWidth, h_, w_, suby_, subx_);
+    ASM_REGISTER_STATE_CHECK(params_.tst_func(
+        dst_tst_ + dst_offset_, dst_stride_, p_src0 + src0_offset_,
+        src0_stride_, p_src1 + src1_offset_, src1_stride_, mask_, kMaxMaskWidth,
+        h_, w_, suby_, subx_));
+  }
+};
+
+TEST_P(BlendA64MaskTest8B, RandomValues) {
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    for (int i = 0; i < kBufSize; ++i) {
+      dst_ref_[i] = rng_.Rand8();
+      dst_tst_[i] = rng_.Rand8();
+
+      src0_[i] = rng_.Rand8();
+      src1_[i] = rng_.Rand8();
+    }
+
+    for (int i = 0; i < kMaxMaskSize; ++i)
+      mask_[i] = rng_(AOM_BLEND_A64_MAX_ALPHA + 1);
+
+    Common();
+  }
+}
+
+TEST_P(BlendA64MaskTest8B, ExtremeValues) {
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    for (int i = 0; i < kBufSize; ++i) {
+      dst_ref_[i] = rng_(2) + 254;
+      dst_tst_[i] = rng_(2) + 254;
+      src0_[i] = rng_(2) + 254;
+      src1_[i] = rng_(2) + 254;
+    }
+
+    for (int i = 0; i < kMaxMaskSize; ++i)
+      mask_[i] = rng_(2) + AOM_BLEND_A64_MAX_ALPHA - 1;
+
+    Common();
+  }
+}
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(SSE4_1, BlendA64MaskTest8B,
+                        ::testing::Values(TestFuncs(
+                            aom_blend_a64_mask_c, aom_blend_a64_mask_sse4_1)));
+#endif  // HAVE_SSE4_1
+
+#if CONFIG_HIGHBITDEPTH
+//////////////////////////////////////////////////////////////////////////////
+// High bit-depth version
+//////////////////////////////////////////////////////////////////////////////
+
+typedef void (*FHBD)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+                     uint32_t src0_stride, const uint8_t *src1,
+                     uint32_t src1_stride, const uint8_t *mask,
+                     uint32_t mask_stride, int h, int w, int suby, int subx,
+                     int bd);
+typedef libaom_test::FuncParam<FHBD> TestFuncsHBD;
+
+class BlendA64MaskTestHBD : public BlendA64MaskTest<FHBD, uint16_t> {
+ protected:
+  void Execute(const uint16_t *p_src0, const uint16_t *p_src1) {
+    params_.ref_func(CONVERT_TO_BYTEPTR(dst_ref_ + dst_offset_), dst_stride_,
+                     CONVERT_TO_BYTEPTR(p_src0 + src0_offset_), src0_stride_,
+                     CONVERT_TO_BYTEPTR(p_src1 + src1_offset_), src1_stride_,
+                     mask_, kMaxMaskWidth, h_, w_, suby_, subx_, bit_depth_);
+    ASM_REGISTER_STATE_CHECK(params_.tst_func(
+        CONVERT_TO_BYTEPTR(dst_tst_ + dst_offset_), dst_stride_,
+        CONVERT_TO_BYTEPTR(p_src0 + src0_offset_), src0_stride_,
+        CONVERT_TO_BYTEPTR(p_src1 + src1_offset_), src1_stride_, mask_,
+        kMaxMaskWidth, h_, w_, suby_, subx_, bit_depth_));
+  }
+
+  int bit_depth_;
+};
+
+TEST_P(BlendA64MaskTestHBD, RandomValues) {
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    switch (rng_(3)) {
+      case 0: bit_depth_ = 8; break;
+      case 1: bit_depth_ = 10; break;
+      default: bit_depth_ = 12; break;
+    }
+
+    const int hi = 1 << bit_depth_;
+
+    for (int i = 0; i < kBufSize; ++i) {
+      dst_ref_[i] = rng_(hi);
+      dst_tst_[i] = rng_(hi);
+      src0_[i] = rng_(hi);
+      src1_[i] = rng_(hi);
+    }
+
+    for (int i = 0; i < kMaxMaskSize; ++i)
+      mask_[i] = rng_(AOM_BLEND_A64_MAX_ALPHA + 1);
+
+    Common();
+  }
+}
+
+TEST_P(BlendA64MaskTestHBD, ExtremeValues) {
+  for (int iter = 0; iter < 1000 && !HasFatalFailure(); ++iter) {
+    switch (rng_(3)) {
+      case 0: bit_depth_ = 8; break;
+      case 1: bit_depth_ = 10; break;
+      default: bit_depth_ = 12; break;
+    }
+
+    const int hi = 1 << bit_depth_;
+    const int lo = hi - 2;
+
+    for (int i = 0; i < kBufSize; ++i) {
+      dst_ref_[i] = rng_(hi - lo) + lo;
+      dst_tst_[i] = rng_(hi - lo) + lo;
+      src0_[i] = rng_(hi - lo) + lo;
+      src1_[i] = rng_(hi - lo) + lo;
+    }
+
+    for (int i = 0; i < kMaxMaskSize; ++i)
+      mask_[i] = rng_(2) + AOM_BLEND_A64_MAX_ALPHA - 1;
+
+    Common();
+  }
+}
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, BlendA64MaskTestHBD,
+    ::testing::Values(TestFuncsHBD(aom_highbd_blend_a64_mask_c,
+                                   aom_highbd_blend_a64_mask_sse4_1)));
+#endif  // HAVE_SSE4_1
+#endif  // CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/boolcoder_test.cc b/third_party/aom/test/boolcoder_test.cc
new file mode 100644
index 0000000000..4d9d7aaf44
--- /dev/null
+++ b/third_party/aom/test/boolcoder_test.cc
@@ -0,0 +1,143 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/acm_random.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/bitreader.h"
+#include "aom_dsp/bitwriter.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+const int num_tests = 10;
+}  // namespace
+
+TEST(AV1, TestBitIO) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  for (int n = 0; n < num_tests; ++n) {
+    for (int method = 0; method <= 7; ++method) {  // we generate various proba
+      const int kBitsToTest = 1000;
+      uint8_t probas[kBitsToTest];
+
+      for (int i = 0; i < kBitsToTest; ++i) {
+        const int parity = i & 1;
+        /* clang-format off */
+        probas[i] =
+          (method == 0) ? 0 : (method == 1) ? 255 :
+          (method == 2) ? 128 :
+          (method == 3) ? rnd.Rand8() :
+          (method == 4) ? (parity ? 0 : 255) :
+            // alternate between low and high proba:
+            (method == 5) ? (parity ? rnd(128) : 255 - rnd(128)) :
+            (method == 6) ?
+            (parity ? rnd(64) : 255 - rnd(64)) :
+            (parity ? rnd(32) : 255 - rnd(32));
+        /* clang-format on */
+      }
+      for (int bit_method = 0; bit_method <= 3; ++bit_method) {
+        const int random_seed = 6432;
+        const int kBufferSize = 10000;
+        ACMRandom bit_rnd(random_seed);
+        aom_writer bw;
+        uint8_t bw_buffer[kBufferSize];
+        aom_start_encode(&bw, bw_buffer);
+
+        int bit = (bit_method == 0) ? 0 : (bit_method == 1) ? 1 : 0;
+        for (int i = 0; i < kBitsToTest; ++i) {
+          if (bit_method == 2) {
+            bit = (i & 1);
+          } else if (bit_method == 3) {
+            bit = bit_rnd(2);
+          }
+          aom_write(&bw, bit, static_cast<int>(probas[i]));
+        }
+
+        aom_stop_encode(&bw);
+
+#if !CONFIG_DAALA_EC
+        // First bit should be zero
+        GTEST_ASSERT_EQ(bw_buffer[0] & 0x80, 0);
+#endif
+
+        aom_reader br;
+        aom_reader_init(&br, bw_buffer, bw.pos, NULL, NULL);
+        bit_rnd.Reset(random_seed);
+        for (int i = 0; i < kBitsToTest; ++i) {
+          if (bit_method == 2) {
+            bit = (i & 1);
+          } else if (bit_method == 3) {
+            bit = bit_rnd(2);
+          }
+          GTEST_ASSERT_EQ(aom_read(&br, probas[i], NULL), bit)
+              << "pos: " << i << " / " << kBitsToTest
+              << " bit_method: " << bit_method << " method: " << method;
+        }
+      }
+    }
+  }
+}
+
+#if CONFIG_DAALA_EC
+#define FRAC_DIFF_TOTAL_ERROR 0.07
+#else
+#define FRAC_DIFF_TOTAL_ERROR 0.2
+#endif
+
+TEST(AV1, TestTell) {
+  const int kBufferSize = 10000;
+  aom_writer bw;
+  uint8_t bw_buffer[kBufferSize];
+  const int kSymbols = 1024;
+  // Coders are noisier at low probabilities, so we start at p = 4.
+  for (int p = 4; p < 256; p++) {
+    double probability = p / 256.;
+    aom_start_encode(&bw, bw_buffer);
+    for (int i = 0; i < kSymbols; i++) {
+      aom_write(&bw, 0, p);
+    }
+    aom_stop_encode(&bw);
+    aom_reader br;
+    aom_reader_init(&br, bw_buffer, bw.pos, NULL, NULL);
+    uint32_t last_tell = aom_reader_tell(&br);
+    uint32_t last_tell_frac = aom_reader_tell_frac(&br);
+    double frac_diff_total = 0;
+    GTEST_ASSERT_GE(aom_reader_tell(&br), 0u);
+    GTEST_ASSERT_LE(aom_reader_tell(&br), 1u);
+    for (int i = 0; i < kSymbols; i++) {
+      aom_read(&br, p, NULL);
+      uint32_t tell = aom_reader_tell(&br);
+      uint32_t tell_frac = aom_reader_tell_frac(&br);
+      GTEST_ASSERT_GE(tell, last_tell) << "tell: " << tell
+                                       << ", last_tell: " << last_tell;
+      GTEST_ASSERT_GE(tell_frac, last_tell_frac)
+          << "tell_frac: " << tell_frac
+          << ", last_tell_frac: " << last_tell_frac;
+      // Frac tell should round up to tell.
+      GTEST_ASSERT_EQ(tell, (tell_frac + 7) >> 3);
+      last_tell = tell;
+      frac_diff_total +=
+          fabs(((tell_frac - last_tell_frac) / 8.0) + log2(probability));
+      last_tell_frac = tell_frac;
+    }
+    const uint32_t expected = (uint32_t)(-kSymbols * log2(probability));
+    // Last tell should be close to the expected value.
+    GTEST_ASSERT_LE(last_tell, expected + 20) << " last_tell: " << last_tell;
+    // The average frac_diff error should be pretty small.
+    GTEST_ASSERT_LE(frac_diff_total / kSymbols, FRAC_DIFF_TOTAL_ERROR)
+        << " frac_diff_total: " << frac_diff_total;
+  }
+}
diff --git a/third_party/aom/test/borders_test.cc b/third_party/aom/test/borders_test.cc
new file mode 100644
index 0000000000..076f914044
--- /dev/null
+++ b/third_party/aom/test/borders_test.cc
@@ -0,0 +1,85 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <climits>
+#include <vector>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+class BordersTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWithParam<libaom_test::TestMode> {
+ protected:
+  BordersTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~BordersTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AOME_SET_CPUUSED, 1);
+      encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+      encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
+      encoder->Control(AOME_SET_ARNR_STRENGTH, 5);
+    }
+  }
+
+  virtual void FramePktHook(const aom_codec_cx_pkt_t *pkt) {
+    if (pkt->data.frame.flags & AOM_FRAME_IS_KEY) {
+    }
+  }
+};
+
+TEST_P(BordersTest, TestEncodeHighBitrate) {
+  // Validate that this non multiple of 64 wide clip encodes and decodes
+  // without a mismatch when passing in a very low max q.  This pushes
+  // the encoder to producing lots of big partitions which will likely
+  // extend into the border and test the border condition.
+  cfg_.g_lag_in_frames = 25;
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.rc_max_quantizer = 10;
+
+  ::libaom_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       40);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+TEST_P(BordersTest, TestLowBitrate) {
+  // Validate that this clip encodes and decodes without a mismatch
+  // when passing in a very high min q.  This pushes the encoder to producing
+  // lots of small partitions which might will test the other condition.
+
+  cfg_.g_lag_in_frames = 25;
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 200;
+  cfg_.rc_min_quantizer = 40;
+
+  ::libaom_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       40);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+AV1_INSTANTIATE_TEST_CASE(BordersTest,
+                          ::testing::Values(::libaom_test::kTwoPassGood));
+}  // namespace
diff --git a/third_party/aom/test/clear_system_state.h b/third_party/aom/test/clear_system_state.h
new file mode 100644
index 0000000000..4f3c1eed07
--- /dev/null
+++ b/third_party/aom/test/clear_system_state.h
@@ -0,0 +1,30 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef TEST_CLEAR_SYSTEM_STATE_H_
+#define TEST_CLEAR_SYSTEM_STATE_H_
+
+#include "./aom_config.h"
+#if ARCH_X86 || ARCH_X86_64
+#include "aom_ports/x86.h"
+#endif
+
+namespace libaom_test {
+
+// Reset system to a known state. This function should be used for all non-API
+// test cases.
+inline void ClearSystemState() {
+#if ARCH_X86 || ARCH_X86_64
+  aom_reset_mmx_state();
+#endif
+}
+
+}  // namespace libaom_test
+#endif  // TEST_CLEAR_SYSTEM_STATE_H_
diff --git a/third_party/aom/test/clpf_test.cc b/third_party/aom/test/clpf_test.cc
new file mode 100644
index 0000000000..2c0f8cf7f3
--- /dev/null
+++ b/third_party/aom/test/clpf_test.cc
@@ -0,0 +1,437 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <cstdlib>
+#include <string>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "aom_ports/aom_timer.h"
+#include "av1/common/od_dering.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+
+typedef void (*clpf_block_t)(uint8_t *dst, const uint16_t *src, int dstride,
+                             int sstride, int sizex, int sizey,
+                             unsigned int strength, unsigned int bitdepth);
+
+typedef std::tr1::tuple<clpf_block_t, clpf_block_t, int, int>
+    clpf_block_param_t;
+
+class CDEFClpfBlockTest : public ::testing::TestWithParam<clpf_block_param_t> {
+ public:
+  virtual ~CDEFClpfBlockTest() {}
+  virtual void SetUp() {
+    clpf = GET_PARAM(0);
+    ref_clpf = GET_PARAM(1);
+    sizex = GET_PARAM(2);
+    sizey = GET_PARAM(3);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  int sizex;
+  int sizey;
+  clpf_block_t clpf;
+  clpf_block_t ref_clpf;
+};
+
+typedef CDEFClpfBlockTest CDEFClpfSpeedTest;
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*clpf_block_hbd_t)(uint16_t *dst, const uint16_t *src,
+                                 int dstride, int sstride, int sizex, int sizey,
+                                 unsigned int strength, unsigned int bitdepth);
+
+typedef std::tr1::tuple<clpf_block_hbd_t, clpf_block_hbd_t, int, int>
+    clpf_block_hbd_param_t;
+
+class CDEFClpfBlockHbdTest
+    : public ::testing::TestWithParam<clpf_block_hbd_param_t> {
+ public:
+  virtual ~CDEFClpfBlockHbdTest() {}
+  virtual void SetUp() {
+    clpf = GET_PARAM(0);
+    ref_clpf = GET_PARAM(1);
+    sizex = GET_PARAM(2);
+    sizey = GET_PARAM(3);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  int sizex;
+  int sizey;
+  clpf_block_hbd_t clpf;
+  clpf_block_hbd_t ref_clpf;
+};
+
+typedef CDEFClpfBlockHbdTest ClpfHbdSpeedTest;
+#endif
+
+template <typename pixel>
+void test_clpf(int w, int h, unsigned int depth, unsigned int iterations,
+               void (*clpf)(pixel *dst, const uint16_t *src, int dstride,
+                            int sstride, int sizex, int sizey,
+                            unsigned int strength, unsigned int bitdepth),
+               void (*ref_clpf)(pixel *dst, const uint16_t *src, int dstride,
+                                int sstride, int sizex, int sizey,
+                                unsigned int strength, unsigned int bitdepth)) {
+  const int size = 24;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint16_t, s[size * size]);
+  DECLARE_ALIGNED(16, pixel, d[size * size]);
+  DECLARE_ALIGNED(16, pixel, ref_d[size * size]);
+  memset(ref_d, 0, size * size * sizeof(*ref_d));
+  memset(d, 0, size * size * sizeof(*d));
+
+  int error = 0, pos = 0, xpos = 8, ypos = 8;
+  unsigned int strength = 0, bits, level, count, damp = 0, boundary = 0;
+
+  assert(size >= w + 16 && size >= h + 16);
+  assert(depth >= 8);
+
+  // Test every combination of:
+  // * Input with up to <depth> bits of noise
+  // * Noise level around every value from 0 to (1<<depth)-1
+  // * All strengths
+  // * All dampings
+  // * Boundaries
+  // If clpf and ref_clpf are the same, we're just testing speed
+  for (boundary = 0; boundary < 16; boundary++) {
+    for (count = 0; count < iterations; count++) {
+      for (level = 0; level < (1U << depth) && !error;
+           level += (1 + 4 * !!boundary) << (depth - 8)) {
+        for (bits = 1; bits <= depth && !error; bits++) {
+          for (damp = 4 + depth - 8; damp < depth - 1 && !error; damp++) {
+            for (int i = 0; i < size * size; i++)
+              s[i] = clamp((rnd.Rand16() & ((1 << bits) - 1)) + level, 0,
+                           (1 << depth) - 1);
+            if (boundary) {
+              if (boundary & 1) {  // Left
+                for (int i = 0; i < size; i++)
+                  for (int j = 0; j < xpos; j++)
+                    s[i * size + j] = OD_DERING_VERY_LARGE;
+              }
+              if (boundary & 2) {  // Right
+                for (int i = 0; i < size; i++)
+                  for (int j = xpos + w; j < size; j++)
+                    s[i * size + j] = OD_DERING_VERY_LARGE;
+              }
+              if (boundary & 4) {  // Above
+                for (int i = 0; i < ypos; i++)
+                  for (int j = 0; j < size; j++)
+                    s[i * size + j] = OD_DERING_VERY_LARGE;
+              }
+              if (boundary & 8) {  // Below
+                for (int i = ypos + h; i < size; i++)
+                  for (int j = 0; j < size; j++)
+                    s[i * size + j] = OD_DERING_VERY_LARGE;
+              }
+            }
+            for (strength = depth - 8; strength < depth - 5 && !error;
+                 strength += !error) {
+              ref_clpf(ref_d + ypos * size + xpos, s + ypos * size + xpos, size,
+                       size, w, h, 1 << strength, damp);
+              if (clpf != ref_clpf)
+                ASM_REGISTER_STATE_CHECK(clpf(d + ypos * size + xpos,
+                                              s + ypos * size + xpos, size,
+                                              size, w, h, 1 << strength, damp));
+              if (ref_clpf != clpf) {
+                for (pos = 0; pos < size * size && !error; pos++) {
+                  error = ref_d[pos] != d[pos];
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  pos--;
+  EXPECT_EQ(0, error)
+      << "Error: CDEFClpfBlockTest, SIMD and C mismatch." << std::endl
+      << "First error at " << pos % size << "," << pos / size << " ("
+      << (int16_t)ref_d[pos] << " != " << (int16_t)d[pos] << ") " << std::endl
+      << "strength: " << (1 << strength) << std::endl
+      << "damping: " << damp << std::endl
+      << "depth: " << depth << std::endl
+      << "boundary: " << boundary << std::endl
+      << "w: " << w << std::endl
+      << "h: " << h << std::endl
+      << "A=" << (pos > 2 * size ? (int16_t)s[pos - 2 * size] : -1) << std::endl
+      << "B=" << (pos > size ? (int16_t)s[pos - size] : -1) << std::endl
+      << "C=" << (pos % size - 2 >= 0 ? (int16_t)s[pos - 2] : -1) << std::endl
+      << "D=" << (pos % size - 1 >= 0 ? (int16_t)s[pos - 1] : -1) << std::endl
+      << "X=" << (int16_t)s[pos] << std::endl
+      << "E=" << (pos % size + 1 < size ? (int16_t)s[pos + 1] : -1) << std::endl
+      << "F=" << (pos % size + 2 < size ? (int16_t)s[pos + 2] : -1) << std::endl
+      << "G=" << (pos + size < size * size ? (int16_t)s[pos + size] : -1)
+      << std::endl
+      << "H="
+      << (pos + 2 * size < size * size ? (int16_t)s[pos + 2 * size] : -1)
+      << std::endl;
+}
+
+template <typename pixel>
+void test_clpf_speed(int w, int h, unsigned int depth, unsigned int iterations,
+                     void (*clpf)(pixel *dst, const uint16_t *src, int dstride,
+                                  int sstride, int sizex, int sizey,
+                                  unsigned int strength, unsigned int bitdepth),
+                     void (*ref_clpf)(pixel *dst, const uint16_t *src,
+                                      int dstride, int sstride, int sizex,
+                                      int sizey, unsigned int strength,
+                                      unsigned int bitdepth)) {
+  aom_usec_timer ref_timer;
+  aom_usec_timer timer;
+
+  aom_usec_timer_start(&ref_timer);
+  test_clpf(w, h, depth, iterations, ref_clpf, ref_clpf);
+  aom_usec_timer_mark(&ref_timer);
+  int ref_elapsed_time = (int)aom_usec_timer_elapsed(&ref_timer);
+
+  aom_usec_timer_start(&timer);
+  test_clpf(w, h, depth, iterations, clpf, clpf);
+  aom_usec_timer_mark(&timer);
+  int elapsed_time = (int)aom_usec_timer_elapsed(&timer);
+
+#if 0
+  std::cout << "[          ] C time = " << ref_elapsed_time / 1000
+            << " ms, SIMD time = " << elapsed_time / 1000 << " ms" << std::endl;
+#endif
+
+  EXPECT_GT(ref_elapsed_time, elapsed_time)
+      << "Error: CDEFClpfSpeedTest, SIMD slower than C." << std::endl
+      << "C time: " << ref_elapsed_time << " us" << std::endl
+      << "SIMD time: " << elapsed_time << " us" << std::endl;
+}
+
+TEST_P(CDEFClpfBlockTest, TestSIMDNoMismatch) {
+  test_clpf(sizex, sizey, 8, 1, clpf, ref_clpf);
+}
+
+TEST_P(CDEFClpfSpeedTest, DISABLED_TestSpeed) {
+  test_clpf_speed(sizex, sizey, 8, 16, clpf, ref_clpf);
+}
+
+#if CONFIG_HIGHBITDEPTH
+TEST_P(CDEFClpfBlockHbdTest, TestSIMDNoMismatch) {
+  test_clpf(sizex, sizey, 12, 1, clpf, ref_clpf);
+}
+
+TEST_P(ClpfHbdSpeedTest, DISABLED_TestSpeed) {
+  test_clpf_speed(sizex, sizey, 12, 4, clpf, ref_clpf);
+}
+#endif
+
+using std::tr1::make_tuple;
+
+// VS compiling for 32 bit targets does not support vector types in
+// structs as arguments, which makes the v256 type of the intrinsics
+// hard to support, so optimizations for this target are disabled.
+#if defined(_WIN64) || !defined(_MSC_VER) || defined(__clang__)
+// Test all supported architectures and block sizes
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, CDEFClpfBlockTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_sse2, &aom_clpf_block_c, 8, 8),
+        make_tuple(&aom_clpf_block_sse2, &aom_clpf_block_c, 8, 4),
+        make_tuple(&aom_clpf_block_sse2, &aom_clpf_block_c, 4, 8),
+        make_tuple(&aom_clpf_block_sse2, &aom_clpf_block_c, 4, 4),
+        make_tuple(&aom_clpf_hblock_sse2, &aom_clpf_hblock_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_sse2, &aom_clpf_hblock_c, 8, 4),
+        make_tuple(&aom_clpf_hblock_sse2, &aom_clpf_hblock_c, 4, 8),
+        make_tuple(&aom_clpf_hblock_sse2, &aom_clpf_hblock_c, 4, 4)));
+#endif
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, CDEFClpfBlockTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_ssse3, &aom_clpf_block_c, 8, 8),
+        make_tuple(&aom_clpf_block_ssse3, &aom_clpf_block_c, 8, 4),
+        make_tuple(&aom_clpf_block_ssse3, &aom_clpf_block_c, 4, 8),
+        make_tuple(&aom_clpf_block_ssse3, &aom_clpf_block_c, 4, 4),
+        make_tuple(&aom_clpf_hblock_ssse3, &aom_clpf_hblock_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_ssse3, &aom_clpf_hblock_c, 8, 4),
+        make_tuple(&aom_clpf_hblock_ssse3, &aom_clpf_hblock_c, 4, 8),
+        make_tuple(&aom_clpf_hblock_ssse3, &aom_clpf_hblock_c, 4, 4)));
+#endif
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, CDEFClpfBlockTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_sse4_1, &aom_clpf_block_c, 8, 8),
+        make_tuple(&aom_clpf_block_sse4_1, &aom_clpf_block_c, 8, 4),
+        make_tuple(&aom_clpf_block_sse4_1, &aom_clpf_block_c, 4, 8),
+        make_tuple(&aom_clpf_block_sse4_1, &aom_clpf_block_c, 4, 4),
+        make_tuple(&aom_clpf_hblock_sse4_1, &aom_clpf_hblock_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_sse4_1, &aom_clpf_hblock_c, 8, 4),
+        make_tuple(&aom_clpf_hblock_sse4_1, &aom_clpf_hblock_c, 4, 8),
+        make_tuple(&aom_clpf_hblock_sse4_1, &aom_clpf_hblock_c, 4, 4)));
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+    NEON, CDEFClpfBlockTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_neon, &aom_clpf_block_c, 8, 8),
+        make_tuple(&aom_clpf_block_neon, &aom_clpf_block_c, 8, 4),
+        make_tuple(&aom_clpf_block_neon, &aom_clpf_block_c, 4, 8),
+        make_tuple(&aom_clpf_block_neon, &aom_clpf_block_c, 4, 4),
+        make_tuple(&aom_clpf_hblock_neon, &aom_clpf_hblock_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_neon, &aom_clpf_hblock_c, 8, 4),
+        make_tuple(&aom_clpf_hblock_neon, &aom_clpf_hblock_c, 4, 8),
+        make_tuple(&aom_clpf_hblock_neon, &aom_clpf_hblock_c, 4, 4)));
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, CDEFClpfBlockHbdTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_hbd_sse2, &aom_clpf_block_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_block_hbd_sse2, &aom_clpf_block_hbd_c, 8, 4),
+        make_tuple(&aom_clpf_block_hbd_sse2, &aom_clpf_block_hbd_c, 4, 8),
+        make_tuple(&aom_clpf_block_hbd_sse2, &aom_clpf_block_hbd_c, 4, 4),
+        make_tuple(&aom_clpf_hblock_hbd_sse2, &aom_clpf_hblock_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_hbd_sse2, &aom_clpf_hblock_hbd_c, 8, 4),
+        make_tuple(&aom_clpf_hblock_hbd_sse2, &aom_clpf_hblock_hbd_c, 4, 8),
+        make_tuple(&aom_clpf_hblock_hbd_sse2, &aom_clpf_hblock_hbd_c, 4, 4)));
+#endif
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, CDEFClpfBlockHbdTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_hbd_ssse3, &aom_clpf_block_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_block_hbd_ssse3, &aom_clpf_block_hbd_c, 8, 4),
+        make_tuple(&aom_clpf_block_hbd_ssse3, &aom_clpf_block_hbd_c, 4, 8),
+        make_tuple(&aom_clpf_block_hbd_ssse3, &aom_clpf_block_hbd_c, 4, 4),
+        make_tuple(&aom_clpf_hblock_hbd_ssse3, &aom_clpf_hblock_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_hbd_ssse3, &aom_clpf_hblock_hbd_c, 8, 4),
+        make_tuple(&aom_clpf_hblock_hbd_ssse3, &aom_clpf_hblock_hbd_c, 4, 8),
+        make_tuple(&aom_clpf_hblock_hbd_ssse3, &aom_clpf_hblock_hbd_c, 4, 4)));
+#endif
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, CDEFClpfBlockHbdTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_hbd_sse4_1, &aom_clpf_block_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_block_hbd_sse4_1, &aom_clpf_block_hbd_c, 8, 4),
+        make_tuple(&aom_clpf_block_hbd_sse4_1, &aom_clpf_block_hbd_c, 4, 8),
+        make_tuple(&aom_clpf_block_hbd_sse4_1, &aom_clpf_block_hbd_c, 4, 4),
+        make_tuple(&aom_clpf_hblock_hbd_sse4_1, &aom_clpf_hblock_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_hbd_sse4_1, &aom_clpf_hblock_hbd_c, 8, 4),
+        make_tuple(&aom_clpf_hblock_hbd_sse4_1, &aom_clpf_hblock_hbd_c, 4, 8),
+        make_tuple(&aom_clpf_hblock_hbd_sse4_1, &aom_clpf_hblock_hbd_c, 4, 4)));
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+    NEON, CDEFClpfBlockHbdTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_hbd_neon, &aom_clpf_block_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_block_hbd_neon, &aom_clpf_block_hbd_c, 8, 4),
+        make_tuple(&aom_clpf_block_hbd_neon, &aom_clpf_block_hbd_c, 4, 8),
+        make_tuple(&aom_clpf_block_hbd_neon, &aom_clpf_block_hbd_c, 4, 4),
+        make_tuple(&aom_clpf_hblock_hbd_neon, &aom_clpf_hblock_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_hbd_neon, &aom_clpf_hblock_hbd_c, 8, 4),
+        make_tuple(&aom_clpf_hblock_hbd_neon, &aom_clpf_hblock_hbd_c, 4, 8),
+        make_tuple(&aom_clpf_hblock_hbd_neon, &aom_clpf_hblock_hbd_c, 4, 4)));
+#endif
+#endif  // CONFIG_HIGHBITDEPTH
+
+// Test speed for all supported architectures
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, CDEFClpfSpeedTest,
+    ::testing::Values(make_tuple(&aom_clpf_block_sse2, &aom_clpf_block_c, 8, 8),
+                      make_tuple(&aom_clpf_hblock_sse2, &aom_clpf_hblock_c, 8,
+                                 8)));
+#endif
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(SSSE3, CDEFClpfSpeedTest,
+                        ::testing::Values(make_tuple(&aom_clpf_block_ssse3,
+                                                     &aom_clpf_block_c, 8, 8),
+                                          make_tuple(&aom_clpf_hblock_ssse3,
+                                                     &aom_clpf_hblock_c, 8,
+                                                     8)));
+#endif
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(SSE4_1, CDEFClpfSpeedTest,
+                        ::testing::Values(make_tuple(&aom_clpf_block_sse4_1,
+                                                     &aom_clpf_block_c, 8, 8),
+                                          make_tuple(&aom_clpf_hblock_sse4_1,
+                                                     &aom_clpf_hblock_c, 8,
+                                                     8)));
+
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+    NEON, CDEFClpfSpeedTest,
+    ::testing::Values(make_tuple(&aom_clpf_block_neon, &aom_clpf_block_c, 8, 8),
+                      make_tuple(&aom_clpf_hblock_neon, &aom_clpf_hblock_c, 8,
+                                 8)));
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, ClpfHbdSpeedTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_hbd_sse2, &aom_clpf_block_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_hbd_sse2, &aom_clpf_hblock_hbd_c, 8, 8)));
+#endif
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, ClpfHbdSpeedTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_hbd_ssse3, &aom_clpf_block_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_hbd_ssse3, &aom_clpf_hblock_hbd_c, 8, 8)));
+#endif
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, ClpfHbdSpeedTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_hbd_sse4_1, &aom_clpf_block_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_hbd_sse4_1, &aom_clpf_hblock_hbd_c, 8, 8)));
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+    NEON, ClpfHbdSpeedTest,
+    ::testing::Values(
+        make_tuple(&aom_clpf_block_hbd_neon, &aom_clpf_block_hbd_c, 8, 8),
+        make_tuple(&aom_clpf_hblock_hbd_neon, &aom_clpf_hblock_hbd_c, 8, 8)));
+#endif
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // defined(_WIN64) || !defined(_MSC_VER)
+
+}  // namespace
diff --git a/third_party/aom/test/codec_factory.h b/third_party/aom/test/codec_factory.h
new file mode 100644
index 0000000000..d2f20b832d
--- /dev/null
+++ b/third_party/aom/test/codec_factory.h
@@ -0,0 +1,166 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef TEST_CODEC_FACTORY_H_
+#define TEST_CODEC_FACTORY_H_
+
+#include "./aom_config.h"
+#include "aom/aom_decoder.h"
+#include "aom/aom_encoder.h"
+#if CONFIG_AV1_ENCODER
+#include "aom/aomcx.h"
+#endif
+#if CONFIG_AV1_DECODER
+#include "aom/aomdx.h"
+#endif
+
+#include "test/decode_test_driver.h"
+#include "test/encode_test_driver.h"
+namespace libaom_test {
+
+const int kCodecFactoryParam = 0;
+
+class CodecFactory {
+ public:
+  CodecFactory() {}
+
+  virtual ~CodecFactory() {}
+
+  virtual Decoder *CreateDecoder(aom_codec_dec_cfg_t cfg) const = 0;
+
+  virtual Decoder *CreateDecoder(aom_codec_dec_cfg_t cfg,
+                                 const aom_codec_flags_t flags) const = 0;
+
+  virtual Encoder *CreateEncoder(aom_codec_enc_cfg_t cfg,
+                                 unsigned long deadline,
+                                 const unsigned long init_flags,
+                                 TwopassStatsStore *stats) const = 0;
+
+  virtual aom_codec_err_t DefaultEncoderConfig(aom_codec_enc_cfg_t *cfg,
+                                               int usage) const = 0;
+};
+
+/* Provide CodecTestWith<n>Params classes for a variable number of parameters
+ * to avoid having to include a pointer to the CodecFactory in every test
+ * definition.
+ */
+template <class T1>
+class CodecTestWithParam
+    : public ::testing::TestWithParam<
+          std::tr1::tuple<const libaom_test::CodecFactory *, T1> > {};
+
+template <class T1, class T2>
+class CodecTestWith2Params
+    : public ::testing::TestWithParam<
+          std::tr1::tuple<const libaom_test::CodecFactory *, T1, T2> > {};
+
+template <class T1, class T2, class T3>
+class CodecTestWith3Params
+    : public ::testing::TestWithParam<
+          std::tr1::tuple<const libaom_test::CodecFactory *, T1, T2, T3> > {};
+
+/*
+ * AV1 Codec Definitions
+ */
+#if CONFIG_AV1
+class AV1Decoder : public Decoder {
+ public:
+  explicit AV1Decoder(aom_codec_dec_cfg_t cfg) : Decoder(cfg) {}
+
+  AV1Decoder(aom_codec_dec_cfg_t cfg, const aom_codec_flags_t flag)
+      : Decoder(cfg, flag) {}
+
+ protected:
+  virtual aom_codec_iface_t *CodecInterface() const {
+#if CONFIG_AV1_DECODER
+    return &aom_codec_av1_dx_algo;
+#else
+    return NULL;
+#endif
+  }
+};
+
+class AV1Encoder : public Encoder {
+ public:
+  AV1Encoder(aom_codec_enc_cfg_t cfg, unsigned long deadline,
+             const unsigned long init_flags, TwopassStatsStore *stats)
+      : Encoder(cfg, deadline, init_flags, stats) {}
+
+ protected:
+  virtual aom_codec_iface_t *CodecInterface() const {
+#if CONFIG_AV1_ENCODER
+    return &aom_codec_av1_cx_algo;
+#else
+    return NULL;
+#endif
+  }
+};
+
+class AV1CodecFactory : public CodecFactory {
+ public:
+  AV1CodecFactory() : CodecFactory() {}
+
+  virtual Decoder *CreateDecoder(aom_codec_dec_cfg_t cfg) const {
+    return CreateDecoder(cfg, 0);
+  }
+
+  virtual Decoder *CreateDecoder(aom_codec_dec_cfg_t cfg,
+                                 const aom_codec_flags_t flags) const {
+#if CONFIG_AV1_DECODER
+    return new AV1Decoder(cfg, flags);
+#else
+    (void)cfg;
+    (void)flags;
+    return NULL;
+#endif
+  }
+
+  virtual Encoder *CreateEncoder(aom_codec_enc_cfg_t cfg,
+                                 unsigned long deadline,
+                                 const unsigned long init_flags,
+                                 TwopassStatsStore *stats) const {
+#if CONFIG_AV1_ENCODER
+    return new AV1Encoder(cfg, deadline, init_flags, stats);
+#else
+    (void)cfg;
+    (void)deadline;
+    (void)init_flags;
+    (void)stats;
+    return NULL;
+#endif
+  }
+
+  virtual aom_codec_err_t DefaultEncoderConfig(aom_codec_enc_cfg_t *cfg,
+                                               int usage) const {
+#if CONFIG_AV1_ENCODER
+    return aom_codec_enc_config_default(&aom_codec_av1_cx_algo, cfg, usage);
+#else
+    (void)cfg;
+    (void)usage;
+    return AOM_CODEC_INCAPABLE;
+#endif
+  }
+};
+
+const libaom_test::AV1CodecFactory kAV1;
+
+#define AV1_INSTANTIATE_TEST_CASE(test, ...)                                \
+  INSTANTIATE_TEST_CASE_P(                                                  \
+      AV1, test,                                                            \
+      ::testing::Combine(                                                   \
+          ::testing::Values(static_cast<const libaom_test::CodecFactory *>( \
+              &libaom_test::kAV1)),                                         \
+          __VA_ARGS__))
+#else
+#define AV1_INSTANTIATE_TEST_CASE(test, ...)
+#endif  // CONFIG_AV1
+
+}  // namespace libaom_test
+#endif  // TEST_CODEC_FACTORY_H_
diff --git a/third_party/aom/test/convolve_test.cc b/third_party/aom/test/convolve_test.cc
new file mode 100644
index 0000000000..a84ef4ec8e
--- /dev/null
+++ b/third_party/aom/test/convolve_test.cc
@@ -0,0 +1,1345 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/aom_timer.h"
+#include "av1/common/filter.h"
+
+namespace {
+
+static const unsigned int kMaxDimension = MAX_SB_SIZE;
+
+typedef void (*ConvolveFunc)(const uint8_t *src, ptrdiff_t src_stride,
+                             uint8_t *dst, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int filter_x_stride,
+                             const int16_t *filter_y, int filter_y_stride,
+                             int w, int h);
+
+struct ConvolveFunctions {
+  ConvolveFunctions(ConvolveFunc copy, ConvolveFunc avg, ConvolveFunc h8,
+                    ConvolveFunc h8_avg, ConvolveFunc v8, ConvolveFunc v8_avg,
+                    ConvolveFunc hv8, ConvolveFunc hv8_avg, ConvolveFunc sh8,
+                    ConvolveFunc sh8_avg, ConvolveFunc sv8,
+                    ConvolveFunc sv8_avg, ConvolveFunc shv8,
+                    ConvolveFunc shv8_avg, int bd)
+      : copy_(copy), avg_(avg), h8_(h8), v8_(v8), hv8_(hv8), h8_avg_(h8_avg),
+        v8_avg_(v8_avg), hv8_avg_(hv8_avg), sh8_(sh8), sv8_(sv8), shv8_(shv8),
+        sh8_avg_(sh8_avg), sv8_avg_(sv8_avg), shv8_avg_(shv8_avg),
+        use_highbd_(bd) {}
+
+  ConvolveFunc copy_;
+  ConvolveFunc avg_;
+  ConvolveFunc h8_;
+  ConvolveFunc v8_;
+  ConvolveFunc hv8_;
+  ConvolveFunc h8_avg_;
+  ConvolveFunc v8_avg_;
+  ConvolveFunc hv8_avg_;
+  ConvolveFunc sh8_;       // scaled horiz
+  ConvolveFunc sv8_;       // scaled vert
+  ConvolveFunc shv8_;      // scaled horiz/vert
+  ConvolveFunc sh8_avg_;   // scaled avg horiz
+  ConvolveFunc sv8_avg_;   // scaled avg vert
+  ConvolveFunc shv8_avg_;  // scaled avg horiz/vert
+  int use_highbd_;  // 0 if high bitdepth not used, else the actual bit depth.
+};
+
+typedef std::tr1::tuple<int, int, const ConvolveFunctions *> ConvolveParam;
+
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+#define ALL_SIZES(convolve_fn)                                            \
+  make_tuple(128, 64, &convolve_fn), make_tuple(64, 128, &convolve_fn),   \
+      make_tuple(128, 128, &convolve_fn), make_tuple(4, 4, &convolve_fn), \
+      make_tuple(8, 4, &convolve_fn), make_tuple(4, 8, &convolve_fn),     \
+      make_tuple(8, 8, &convolve_fn), make_tuple(16, 8, &convolve_fn),    \
+      make_tuple(8, 16, &convolve_fn), make_tuple(16, 16, &convolve_fn),  \
+      make_tuple(32, 16, &convolve_fn), make_tuple(16, 32, &convolve_fn), \
+      make_tuple(32, 32, &convolve_fn), make_tuple(64, 32, &convolve_fn), \
+      make_tuple(32, 64, &convolve_fn), make_tuple(64, 64, &convolve_fn)
+#else
+#define ALL_SIZES(convolve_fn)                                            \
+  make_tuple(4, 4, &convolve_fn), make_tuple(8, 4, &convolve_fn),         \
+      make_tuple(4, 8, &convolve_fn), make_tuple(8, 8, &convolve_fn),     \
+      make_tuple(16, 8, &convolve_fn), make_tuple(8, 16, &convolve_fn),   \
+      make_tuple(16, 16, &convolve_fn), make_tuple(32, 16, &convolve_fn), \
+      make_tuple(16, 32, &convolve_fn), make_tuple(32, 32, &convolve_fn), \
+      make_tuple(64, 32, &convolve_fn), make_tuple(32, 64, &convolve_fn), \
+      make_tuple(64, 64, &convolve_fn)
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+
+// Reference 8-tap subpixel filter, slightly modified to fit into this test.
+#define AV1_FILTER_WEIGHT 128
+#define AV1_FILTER_SHIFT 7
+uint8_t clip_pixel(int x) { return x < 0 ? 0 : x > 255 ? 255 : x; }
+
+void filter_block2d_8_c(const uint8_t *src_ptr, unsigned int src_stride,
+                        const int16_t *HFilter, const int16_t *VFilter,
+                        uint8_t *dst_ptr, unsigned int dst_stride,
+                        unsigned int output_width, unsigned int output_height) {
+  // Between passes, we use an intermediate buffer whose height is extended to
+  // have enough horizontally filtered values as input for the vertical pass.
+  // This buffer is allocated to be big enough for the largest block type we
+  // support.
+  const int kInterp_Extend = 4;
+  const unsigned int intermediate_height =
+      (kInterp_Extend - 1) + output_height + kInterp_Extend;
+  unsigned int i, j;
+
+  assert(intermediate_height > 7);
+
+  // Size of intermediate_buffer is max_intermediate_height * filter_max_width,
+  // where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height
+  //                                 + kInterp_Extend
+  //                               = 3 + 16 + 4
+  //                               = 23
+  // and filter_max_width          = 16
+  //
+  uint8_t intermediate_buffer[(kMaxDimension + 8) * kMaxDimension];
+  const int intermediate_next_stride =
+      1 - static_cast<int>(intermediate_height * output_width);
+
+  // Horizontal pass (src -> transposed intermediate).
+  uint8_t *output_ptr = intermediate_buffer;
+  const int src_next_row_stride = src_stride - output_width;
+  src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1);
+  for (i = 0; i < intermediate_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      // Apply filter...
+      const int temp = (src_ptr[0] * HFilter[0]) + (src_ptr[1] * HFilter[1]) +
+                       (src_ptr[2] * HFilter[2]) + (src_ptr[3] * HFilter[3]) +
+                       (src_ptr[4] * HFilter[4]) + (src_ptr[5] * HFilter[5]) +
+                       (src_ptr[6] * HFilter[6]) + (src_ptr[7] * HFilter[7]) +
+                       (AV1_FILTER_WEIGHT >> 1);  // Rounding
+
+      // Normalize back to 0-255...
+      *output_ptr = clip_pixel(temp >> AV1_FILTER_SHIFT);
+      ++src_ptr;
+      output_ptr += intermediate_height;
+    }
+    src_ptr += src_next_row_stride;
+    output_ptr += intermediate_next_stride;
+  }
+
+  // Vertical pass (transposed intermediate -> dst).
+  src_ptr = intermediate_buffer;
+  const int dst_next_row_stride = dst_stride - output_width;
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      // Apply filter...
+      const int temp = (src_ptr[0] * VFilter[0]) + (src_ptr[1] * VFilter[1]) +
+                       (src_ptr[2] * VFilter[2]) + (src_ptr[3] * VFilter[3]) +
+                       (src_ptr[4] * VFilter[4]) + (src_ptr[5] * VFilter[5]) +
+                       (src_ptr[6] * VFilter[6]) + (src_ptr[7] * VFilter[7]) +
+                       (AV1_FILTER_WEIGHT >> 1);  // Rounding
+
+      // Normalize back to 0-255...
+      *dst_ptr++ = clip_pixel(temp >> AV1_FILTER_SHIFT);
+      src_ptr += intermediate_height;
+    }
+    src_ptr += intermediate_next_stride;
+    dst_ptr += dst_next_row_stride;
+  }
+}
+
+void block2d_average_c(uint8_t *src, unsigned int src_stride,
+                       uint8_t *output_ptr, unsigned int output_stride,
+                       unsigned int output_width, unsigned int output_height) {
+  unsigned int i, j;
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      output_ptr[j] = (output_ptr[j] + src[i * src_stride + j] + 1) >> 1;
+    }
+    output_ptr += output_stride;
+  }
+}
+
+void filter_average_block2d_8_c(const uint8_t *src_ptr,
+                                const unsigned int src_stride,
+                                const int16_t *HFilter, const int16_t *VFilter,
+                                uint8_t *dst_ptr, unsigned int dst_stride,
+                                unsigned int output_width,
+                                unsigned int output_height) {
+  uint8_t tmp[kMaxDimension * kMaxDimension];
+
+  assert(output_width <= kMaxDimension);
+  assert(output_height <= kMaxDimension);
+  filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, kMaxDimension,
+                     output_width, output_height);
+  block2d_average_c(tmp, kMaxDimension, dst_ptr, dst_stride, output_width,
+                    output_height);
+}
+
+#if CONFIG_HIGHBITDEPTH
+void highbd_filter_block2d_8_c(const uint16_t *src_ptr,
+                               const unsigned int src_stride,
+                               const int16_t *HFilter, const int16_t *VFilter,
+                               uint16_t *dst_ptr, unsigned int dst_stride,
+                               unsigned int output_width,
+                               unsigned int output_height, int bd) {
+  // Between passes, we use an intermediate buffer whose height is extended to
+  // have enough horizontally filtered values as input for the vertical pass.
+  // This buffer is allocated to be big enough for the largest block type we
+  // support.
+  const int kInterp_Extend = 4;
+  const unsigned int intermediate_height =
+      (kInterp_Extend - 1) + output_height + kInterp_Extend;
+
+  /* Size of intermediate_buffer is max_intermediate_height * filter_max_width,
+   * where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height
+   *                                 + kInterp_Extend
+   *                               = 3 + 16 + 4
+   *                               = 23
+   * and filter_max_width = 16
+   */
+  uint16_t intermediate_buffer[(kMaxDimension + 8) * kMaxDimension];
+  const int intermediate_next_stride =
+      1 - static_cast<int>(intermediate_height * output_width);
+
+  // Horizontal pass (src -> transposed intermediate).
+  {
+    uint16_t *output_ptr = intermediate_buffer;
+    const int src_next_row_stride = src_stride - output_width;
+    unsigned int i, j;
+    src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1);
+    for (i = 0; i < intermediate_height; ++i) {
+      for (j = 0; j < output_width; ++j) {
+        // Apply filter...
+        const int temp = (src_ptr[0] * HFilter[0]) + (src_ptr[1] * HFilter[1]) +
+                         (src_ptr[2] * HFilter[2]) + (src_ptr[3] * HFilter[3]) +
+                         (src_ptr[4] * HFilter[4]) + (src_ptr[5] * HFilter[5]) +
+                         (src_ptr[6] * HFilter[6]) + (src_ptr[7] * HFilter[7]) +
+                         (AV1_FILTER_WEIGHT >> 1);  // Rounding
+
+        // Normalize back to 0-255...
+        *output_ptr = clip_pixel_highbd(temp >> AV1_FILTER_SHIFT, bd);
+        ++src_ptr;
+        output_ptr += intermediate_height;
+      }
+      src_ptr += src_next_row_stride;
+      output_ptr += intermediate_next_stride;
+    }
+  }
+
+  // Vertical pass (transposed intermediate -> dst).
+  {
+    const uint16_t *interm_ptr = intermediate_buffer;
+    const int dst_next_row_stride = dst_stride - output_width;
+    unsigned int i, j;
+    for (i = 0; i < output_height; ++i) {
+      for (j = 0; j < output_width; ++j) {
+        // Apply filter...
+        const int temp =
+            (interm_ptr[0] * VFilter[0]) + (interm_ptr[1] * VFilter[1]) +
+            (interm_ptr[2] * VFilter[2]) + (interm_ptr[3] * VFilter[3]) +
+            (interm_ptr[4] * VFilter[4]) + (interm_ptr[5] * VFilter[5]) +
+            (interm_ptr[6] * VFilter[6]) + (interm_ptr[7] * VFilter[7]) +
+            (AV1_FILTER_WEIGHT >> 1);  // Rounding
+
+        // Normalize back to 0-255...
+        *dst_ptr++ = clip_pixel_highbd(temp >> AV1_FILTER_SHIFT, bd);
+        interm_ptr += intermediate_height;
+      }
+      interm_ptr += intermediate_next_stride;
+      dst_ptr += dst_next_row_stride;
+    }
+  }
+}
+
+void highbd_block2d_average_c(uint16_t *src, unsigned int src_stride,
+                              uint16_t *output_ptr, unsigned int output_stride,
+                              unsigned int output_width,
+                              unsigned int output_height) {
+  unsigned int i, j;
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; ++j) {
+      output_ptr[j] = (output_ptr[j] + src[i * src_stride + j] + 1) >> 1;
+    }
+    output_ptr += output_stride;
+  }
+}
+
+void highbd_filter_average_block2d_8_c(
+    const uint16_t *src_ptr, unsigned int src_stride, const int16_t *HFilter,
+    const int16_t *VFilter, uint16_t *dst_ptr, unsigned int dst_stride,
+    unsigned int output_width, unsigned int output_height, int bd) {
+  uint16_t tmp[kMaxDimension * kMaxDimension];
+
+  assert(output_width <= kMaxDimension);
+  assert(output_height <= kMaxDimension);
+  highbd_filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp,
+                            kMaxDimension, output_width, output_height, bd);
+  highbd_block2d_average_c(tmp, kMaxDimension, dst_ptr, dst_stride,
+                           output_width, output_height);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
+ public:
+  static void SetUpTestCase() {
+    // Force input_ to be unaligned, output to be 16 byte aligned.
+    input_ = reinterpret_cast<uint8_t *>(
+                 aom_memalign(kDataAlignment, kInputBufferSize + 1)) +
+             1;
+    output_ = reinterpret_cast<uint8_t *>(
+        aom_memalign(kDataAlignment, kOutputBufferSize));
+    output_ref_ = reinterpret_cast<uint8_t *>(
+        aom_memalign(kDataAlignment, kOutputBufferSize));
+#if CONFIG_HIGHBITDEPTH
+    input16_ = reinterpret_cast<uint16_t *>(aom_memalign(
+                   kDataAlignment, (kInputBufferSize + 1) * sizeof(uint16_t))) +
+               1;
+    output16_ = reinterpret_cast<uint16_t *>(
+        aom_memalign(kDataAlignment, (kOutputBufferSize) * sizeof(uint16_t)));
+    output16_ref_ = reinterpret_cast<uint16_t *>(
+        aom_memalign(kDataAlignment, (kOutputBufferSize) * sizeof(uint16_t)));
+#endif
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+  static void TearDownTestCase() {
+    aom_free(input_ - 1);
+    input_ = NULL;
+    aom_free(output_);
+    output_ = NULL;
+    aom_free(output_ref_);
+    output_ref_ = NULL;
+#if CONFIG_HIGHBITDEPTH
+    aom_free(input16_ - 1);
+    input16_ = NULL;
+    aom_free(output16_);
+    output16_ = NULL;
+    aom_free(output16_ref_);
+    output16_ref_ = NULL;
+#endif
+  }
+
+ protected:
+  static const int kDataAlignment = 16;
+  static const int kOuterBlockSize = 4 * kMaxDimension;
+  static const int kInputStride = kOuterBlockSize;
+  static const int kOutputStride = kOuterBlockSize;
+  static const int kInputBufferSize = kOuterBlockSize * kOuterBlockSize;
+  static const int kOutputBufferSize = kOuterBlockSize * kOuterBlockSize;
+
+  int Width() const { return GET_PARAM(0); }
+  int Height() const { return GET_PARAM(1); }
+  int BorderLeft() const {
+    const int center = (kOuterBlockSize - Width()) / 2;
+    return (center + (kDataAlignment - 1)) & ~(kDataAlignment - 1);
+  }
+  int BorderTop() const { return (kOuterBlockSize - Height()) / 2; }
+
+  bool IsIndexInBorder(int i) {
+    return (i < BorderTop() * kOuterBlockSize ||
+            i >= (BorderTop() + Height()) * kOuterBlockSize ||
+            i % kOuterBlockSize < BorderLeft() ||
+            i % kOuterBlockSize >= (BorderLeft() + Width()));
+  }
+
+  virtual void SetUp() {
+    UUT_ = GET_PARAM(2);
+#if CONFIG_HIGHBITDEPTH
+    if (UUT_->use_highbd_ != 0)
+      mask_ = (1 << UUT_->use_highbd_) - 1;
+    else
+      mask_ = 255;
+#endif
+    /* Set up guard blocks for an inner block centered in the outer block */
+    for (int i = 0; i < kOutputBufferSize; ++i) {
+      if (IsIndexInBorder(i))
+        output_[i] = 255;
+      else
+        output_[i] = 0;
+    }
+
+    ::libaom_test::ACMRandom prng;
+    for (int i = 0; i < kInputBufferSize; ++i) {
+      if (i & 1) {
+        input_[i] = 255;
+#if CONFIG_HIGHBITDEPTH
+        input16_[i] = mask_;
+#endif
+      } else {
+        input_[i] = prng.Rand8Extremes();
+#if CONFIG_HIGHBITDEPTH
+        input16_[i] = prng.Rand16() & mask_;
+#endif
+      }
+    }
+  }
+
+  void SetConstantInput(int value) {
+    memset(input_, value, kInputBufferSize);
+#if CONFIG_HIGHBITDEPTH
+    aom_memset16(input16_, value, kInputBufferSize);
+#endif
+  }
+
+  void CopyOutputToRef() {
+    memcpy(output_ref_, output_, kOutputBufferSize);
+#if CONFIG_HIGHBITDEPTH
+    // Copy 16-bit pixels values. The effective number of bytes is double.
+    memcpy(output16_ref_, output16_, sizeof(output16_[0]) * kOutputBufferSize);
+#endif
+  }
+
+  void CheckGuardBlocks() {
+    for (int i = 0; i < kOutputBufferSize; ++i) {
+      if (IsIndexInBorder(i)) EXPECT_EQ(255, output_[i]);
+    }
+  }
+
+  uint8_t *input() const {
+    const int offset = BorderTop() * kOuterBlockSize + BorderLeft();
+#if CONFIG_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      return input_ + offset;
+    } else {
+      return CONVERT_TO_BYTEPTR(input16_) + offset;
+    }
+#else
+    return input_ + offset;
+#endif
+  }
+
+  uint8_t *output() const {
+    const int offset = BorderTop() * kOuterBlockSize + BorderLeft();
+#if CONFIG_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      return output_ + offset;
+    } else {
+      return CONVERT_TO_BYTEPTR(output16_) + offset;
+    }
+#else
+    return output_ + offset;
+#endif
+  }
+
+  uint8_t *output_ref() const {
+    const int offset = BorderTop() * kOuterBlockSize + BorderLeft();
+#if CONFIG_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      return output_ref_ + offset;
+    } else {
+      return CONVERT_TO_BYTEPTR(output16_ref_) + offset;
+    }
+#else
+    return output_ref_ + offset;
+#endif
+  }
+
+  uint16_t lookup(uint8_t *list, int index) const {
+#if CONFIG_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      return list[index];
+    } else {
+      return CONVERT_TO_SHORTPTR(list)[index];
+    }
+#else
+    return list[index];
+#endif
+  }
+
+  void assign_val(uint8_t *list, int index, uint16_t val) const {
+#if CONFIG_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      list[index] = (uint8_t)val;
+    } else {
+      CONVERT_TO_SHORTPTR(list)[index] = val;
+    }
+#else
+    list[index] = (uint8_t)val;
+#endif
+  }
+
+  void wrapper_filter_average_block2d_8_c(
+      const uint8_t *src_ptr, unsigned int src_stride, const int16_t *HFilter,
+      const int16_t *VFilter, uint8_t *dst_ptr, unsigned int dst_stride,
+      unsigned int output_width, unsigned int output_height) {
+#if CONFIG_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      filter_average_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, dst_ptr,
+                                 dst_stride, output_width, output_height);
+    } else {
+      highbd_filter_average_block2d_8_c(
+          CONVERT_TO_SHORTPTR(src_ptr), src_stride, HFilter, VFilter,
+          CONVERT_TO_SHORTPTR(dst_ptr), dst_stride, output_width, output_height,
+          UUT_->use_highbd_);
+    }
+#else
+    filter_average_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, dst_ptr,
+                               dst_stride, output_width, output_height);
+#endif
+  }
+
+  void wrapper_filter_block2d_8_c(
+      const uint8_t *src_ptr, unsigned int src_stride, const int16_t *HFilter,
+      const int16_t *VFilter, uint8_t *dst_ptr, unsigned int dst_stride,
+      unsigned int output_width, unsigned int output_height) {
+#if CONFIG_HIGHBITDEPTH
+    if (UUT_->use_highbd_ == 0) {
+      filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, dst_ptr,
+                         dst_stride, output_width, output_height);
+    } else {
+      highbd_filter_block2d_8_c(CONVERT_TO_SHORTPTR(src_ptr), src_stride,
+                                HFilter, VFilter, CONVERT_TO_SHORTPTR(dst_ptr),
+                                dst_stride, output_width, output_height,
+                                UUT_->use_highbd_);
+    }
+#else
+    filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, dst_ptr,
+                       dst_stride, output_width, output_height);
+#endif
+  }
+
+  const ConvolveFunctions *UUT_;
+  static uint8_t *input_;
+  static uint8_t *output_;
+  static uint8_t *output_ref_;
+#if CONFIG_HIGHBITDEPTH
+  static uint16_t *input16_;
+  static uint16_t *output16_;
+  static uint16_t *output16_ref_;
+  int mask_;
+#endif
+};
+
+uint8_t *ConvolveTest::input_ = NULL;
+uint8_t *ConvolveTest::output_ = NULL;
+uint8_t *ConvolveTest::output_ref_ = NULL;
+#if CONFIG_HIGHBITDEPTH
+uint16_t *ConvolveTest::input16_ = NULL;
+uint16_t *ConvolveTest::output16_ = NULL;
+uint16_t *ConvolveTest::output16_ref_ = NULL;
+#endif
+
+TEST_P(ConvolveTest, GuardBlocks) { CheckGuardBlocks(); }
+
+TEST_P(ConvolveTest, Copy) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+
+  ASM_REGISTER_STATE_CHECK(UUT_->copy_(in, kInputStride, out, kOutputStride,
+                                       NULL, 0, NULL, 0, Width(), Height()));
+
+  CheckGuardBlocks();
+
+  for (int y = 0; y < Height(); ++y)
+    for (int x = 0; x < Width(); ++x)
+      ASSERT_EQ(lookup(out, y * kOutputStride + x),
+                lookup(in, y * kInputStride + x))
+          << "(" << x << "," << y << ")";
+}
+
+TEST_P(ConvolveTest, Avg) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+  uint8_t *const out_ref = output_ref();
+  CopyOutputToRef();
+
+  ASM_REGISTER_STATE_CHECK(UUT_->avg_(in, kInputStride, out, kOutputStride,
+                                      NULL, 0, NULL, 0, Width(), Height()));
+
+  CheckGuardBlocks();
+
+  for (int y = 0; y < Height(); ++y)
+    for (int x = 0; x < Width(); ++x)
+      ASSERT_EQ(lookup(out, y * kOutputStride + x),
+                ROUND_POWER_OF_TWO(lookup(in, y * kInputStride + x) +
+                                       lookup(out_ref, y * kOutputStride + x),
+                                   1))
+          << "(" << x << "," << y << ")";
+}
+
+TEST_P(ConvolveTest, CopyHoriz) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+  DECLARE_ALIGNED(256, const int16_t,
+                  filter8[8]) = { 0, 0, 0, 128, 0, 0, 0, 0 };
+
+  ASM_REGISTER_STATE_CHECK(UUT_->sh8_(in, kInputStride, out, kOutputStride,
+                                      filter8, 16, filter8, 16, Width(),
+                                      Height()));
+
+  CheckGuardBlocks();
+
+  for (int y = 0; y < Height(); ++y)
+    for (int x = 0; x < Width(); ++x)
+      ASSERT_EQ(lookup(out, y * kOutputStride + x),
+                lookup(in, y * kInputStride + x))
+          << "(" << x << "," << y << ")";
+}
+
+TEST_P(ConvolveTest, CopyVert) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+  DECLARE_ALIGNED(256, const int16_t,
+                  filter8[8]) = { 0, 0, 0, 128, 0, 0, 0, 0 };
+
+  ASM_REGISTER_STATE_CHECK(UUT_->sv8_(in, kInputStride, out, kOutputStride,
+                                      filter8, 16, filter8, 16, Width(),
+                                      Height()));
+
+  CheckGuardBlocks();
+
+  for (int y = 0; y < Height(); ++y)
+    for (int x = 0; x < Width(); ++x)
+      ASSERT_EQ(lookup(out, y * kOutputStride + x),
+                lookup(in, y * kInputStride + x))
+          << "(" << x << "," << y << ")";
+}
+
+TEST_P(ConvolveTest, Copy2D) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+  DECLARE_ALIGNED(256, const int16_t,
+                  filter8[8]) = { 0, 0, 0, 128, 0, 0, 0, 0 };
+
+  ASM_REGISTER_STATE_CHECK(UUT_->shv8_(in, kInputStride, out, kOutputStride,
+                                       filter8, 16, filter8, 16, Width(),
+                                       Height()));
+
+  CheckGuardBlocks();
+
+  for (int y = 0; y < Height(); ++y)
+    for (int x = 0; x < Width(); ++x)
+      ASSERT_EQ(lookup(out, y * kOutputStride + x),
+                lookup(in, y * kInputStride + x))
+          << "(" << x << "," << y << ")";
+}
+
+const int kNumFilterBanks = SWITCHABLE_FILTERS;
+const int kNumFilters = 16;
+
+TEST(ConvolveTest, FiltersWontSaturateWhenAddedPairwise) {
+  for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
+    const InterpFilter filter = (InterpFilter)filter_bank;
+    const InterpKernel *filters =
+        (const InterpKernel *)av1_get_interp_filter_kernel(filter);
+#if CONFIG_DUAL_FILTER
+    const InterpFilterParams filter_params =
+        av1_get_interp_filter_params(filter);
+    if (filter_params.taps != SUBPEL_TAPS) continue;
+#endif
+    for (int i = 0; i < kNumFilters; i++) {
+      const int p0 = filters[i][0] + filters[i][1];
+      const int p1 = filters[i][2] + filters[i][3];
+      const int p2 = filters[i][4] + filters[i][5];
+      const int p3 = filters[i][6] + filters[i][7];
+      EXPECT_LE(p0, 128);
+      EXPECT_LE(p1, 128);
+      EXPECT_LE(p2, 128);
+      EXPECT_LE(p3, 128);
+      EXPECT_LE(p0 + p3, 128);
+      EXPECT_LE(p0 + p3 + p1, 128);
+      EXPECT_LE(p0 + p3 + p1 + p2, 128);
+      EXPECT_EQ(p0 + p1 + p2 + p3, 128);
+    }
+  }
+}
+
+const int16_t kInvalidFilter[8] = { 0 };
+
+TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+#if CONFIG_HIGHBITDEPTH
+  uint8_t ref8[kOutputStride * kMaxDimension];
+  uint16_t ref16[kOutputStride * kMaxDimension];
+  uint8_t *ref;
+  if (UUT_->use_highbd_ == 0) {
+    ref = ref8;
+  } else {
+    ref = CONVERT_TO_BYTEPTR(ref16);
+  }
+#else
+  uint8_t ref[kOutputStride * kMaxDimension];
+#endif
+
+  for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
+    const InterpFilter filter = (InterpFilter)filter_bank;
+    const InterpKernel *filters =
+        (const InterpKernel *)av1_get_interp_filter_kernel(filter);
+#if CONFIG_DUAL_FILTER
+    const InterpFilterParams filter_params =
+        av1_get_interp_filter_params(filter);
+    if (filter_params.taps != SUBPEL_TAPS) continue;
+#endif
+
+    for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
+      for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
+        wrapper_filter_block2d_8_c(in, kInputStride, filters[filter_x],
+                                   filters[filter_y], ref, kOutputStride,
+                                   Width(), Height());
+
+        if (filter_x && filter_y)
+          ASM_REGISTER_STATE_CHECK(UUT_->hv8_(
+              in, kInputStride, out, kOutputStride, filters[filter_x], 16,
+              filters[filter_y], 16, Width(), Height()));
+        else if (filter_y)
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->v8_(in, kInputStride, out, kOutputStride, kInvalidFilter,
+                        16, filters[filter_y], 16, Width(), Height()));
+        else if (filter_x)
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->h8_(in, kInputStride, out, kOutputStride, filters[filter_x],
+                        16, kInvalidFilter, 16, Width(), Height()));
+        else
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->copy_(in, kInputStride, out, kOutputStride, kInvalidFilter,
+                          0, kInvalidFilter, 0, Width(), Height()));
+
+        CheckGuardBlocks();
+
+        for (int y = 0; y < Height(); ++y)
+          for (int x = 0; x < Width(); ++x)
+            ASSERT_EQ(lookup(ref, y * kOutputStride + x),
+                      lookup(out, y * kOutputStride + x))
+                << "mismatch at (" << x << "," << y << "), "
+                << "filters (" << filter_bank << "," << filter_x << ","
+                << filter_y << ")";
+      }
+    }
+  }
+}
+
+TEST_P(ConvolveTest, MatchesReferenceAveragingSubpixelFilter) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+#if CONFIG_HIGHBITDEPTH
+  uint8_t ref8[kOutputStride * kMaxDimension];
+  uint16_t ref16[kOutputStride * kMaxDimension];
+  uint8_t *ref;
+  if (UUT_->use_highbd_ == 0) {
+    ref = ref8;
+  } else {
+    ref = CONVERT_TO_BYTEPTR(ref16);
+  }
+#else
+  uint8_t ref[kOutputStride * kMaxDimension];
+#endif
+
+  // Populate ref and out with some random data
+  ::libaom_test::ACMRandom prng;
+  for (int y = 0; y < Height(); ++y) {
+    for (int x = 0; x < Width(); ++x) {
+      uint16_t r;
+#if CONFIG_HIGHBITDEPTH
+      if (UUT_->use_highbd_ == 0 || UUT_->use_highbd_ == 8) {
+        r = prng.Rand8Extremes();
+      } else {
+        r = prng.Rand16() & mask_;
+      }
+#else
+      r = prng.Rand8Extremes();
+#endif
+
+      assign_val(out, y * kOutputStride + x, r);
+      assign_val(ref, y * kOutputStride + x, r);
+    }
+  }
+
+  for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
+    const InterpFilter filter = (InterpFilter)filter_bank;
+    const InterpKernel *filters =
+        (const InterpKernel *)av1_get_interp_filter_kernel(filter);
+#if CONFIG_DUAL_FILTER
+    const InterpFilterParams filter_params =
+        av1_get_interp_filter_params(filter);
+    if (filter_params.taps != SUBPEL_TAPS) continue;
+#endif
+
+    for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
+      for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
+        wrapper_filter_average_block2d_8_c(in, kInputStride, filters[filter_x],
+                                           filters[filter_y], ref,
+                                           kOutputStride, Width(), Height());
+
+        if (filter_x && filter_y)
+          ASM_REGISTER_STATE_CHECK(UUT_->hv8_avg_(
+              in, kInputStride, out, kOutputStride, filters[filter_x], 16,
+              filters[filter_y], 16, Width(), Height()));
+        else if (filter_y)
+          ASM_REGISTER_STATE_CHECK(UUT_->v8_avg_(
+              in, kInputStride, out, kOutputStride, kInvalidFilter, 16,
+              filters[filter_y], 16, Width(), Height()));
+        else if (filter_x)
+          ASM_REGISTER_STATE_CHECK(UUT_->h8_avg_(
+              in, kInputStride, out, kOutputStride, filters[filter_x], 16,
+              kInvalidFilter, 16, Width(), Height()));
+        else
+          ASM_REGISTER_STATE_CHECK(
+              UUT_->avg_(in, kInputStride, out, kOutputStride, kInvalidFilter,
+                         0, kInvalidFilter, 0, Width(), Height()));
+
+        CheckGuardBlocks();
+
+        for (int y = 0; y < Height(); ++y)
+          for (int x = 0; x < Width(); ++x)
+            ASSERT_EQ(lookup(ref, y * kOutputStride + x),
+                      lookup(out, y * kOutputStride + x))
+                << "mismatch at (" << x << "," << y << "), "
+                << "filters (" << filter_bank << "," << filter_x << ","
+                << filter_y << ")";
+      }
+    }
+  }
+}
+
+TEST_P(ConvolveTest, FilterExtremes) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+#if CONFIG_HIGHBITDEPTH
+  uint8_t ref8[kOutputStride * kMaxDimension];
+  uint16_t ref16[kOutputStride * kMaxDimension];
+  uint8_t *ref;
+  if (UUT_->use_highbd_ == 0) {
+    ref = ref8;
+  } else {
+    ref = CONVERT_TO_BYTEPTR(ref16);
+  }
+#else
+  uint8_t ref[kOutputStride * kMaxDimension];
+#endif
+
+  // Populate ref and out with some random data
+  ::libaom_test::ACMRandom prng;
+  for (int y = 0; y < Height(); ++y) {
+    for (int x = 0; x < Width(); ++x) {
+      uint16_t r;
+#if CONFIG_HIGHBITDEPTH
+      if (UUT_->use_highbd_ == 0 || UUT_->use_highbd_ == 8) {
+        r = prng.Rand8Extremes();
+      } else {
+        r = prng.Rand16() & mask_;
+      }
+#else
+      r = prng.Rand8Extremes();
+#endif
+      assign_val(out, y * kOutputStride + x, r);
+      assign_val(ref, y * kOutputStride + x, r);
+    }
+  }
+
+  for (int axis = 0; axis < 2; axis++) {
+    int seed_val = 0;
+    while (seed_val < 256) {
+      for (int y = 0; y < 8; ++y) {
+        for (int x = 0; x < 8; ++x) {
+#if CONFIG_HIGHBITDEPTH
+          assign_val(in, y * kOutputStride + x - SUBPEL_TAPS / 2 + 1,
+                     ((seed_val >> (axis ? y : x)) & 1) * mask_);
+#else
+          assign_val(in, y * kOutputStride + x - SUBPEL_TAPS / 2 + 1,
+                     ((seed_val >> (axis ? y : x)) & 1) * 255);
+#endif
+          if (axis) seed_val++;
+        }
+        if (axis)
+          seed_val -= 8;
+        else
+          seed_val++;
+      }
+      if (axis) seed_val += 8;
+
+      for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
+        const InterpFilter filter = (InterpFilter)filter_bank;
+        const InterpKernel *filters =
+            (const InterpKernel *)av1_get_interp_filter_kernel(filter);
+#if CONFIG_DUAL_FILTER
+        const InterpFilterParams filter_params =
+            av1_get_interp_filter_params(filter);
+        if (filter_params.taps != SUBPEL_TAPS) continue;
+#endif
+        for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
+          for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
+            wrapper_filter_block2d_8_c(in, kInputStride, filters[filter_x],
+                                       filters[filter_y], ref, kOutputStride,
+                                       Width(), Height());
+            if (filter_x && filter_y)
+              ASM_REGISTER_STATE_CHECK(UUT_->hv8_(
+                  in, kInputStride, out, kOutputStride, filters[filter_x], 16,
+                  filters[filter_y], 16, Width(), Height()));
+            else if (filter_y)
+              ASM_REGISTER_STATE_CHECK(UUT_->v8_(
+                  in, kInputStride, out, kOutputStride, kInvalidFilter, 16,
+                  filters[filter_y], 16, Width(), Height()));
+            else if (filter_x)
+              ASM_REGISTER_STATE_CHECK(UUT_->h8_(
+                  in, kInputStride, out, kOutputStride, filters[filter_x], 16,
+                  kInvalidFilter, 16, Width(), Height()));
+            else
+              ASM_REGISTER_STATE_CHECK(UUT_->copy_(
+                  in, kInputStride, out, kOutputStride, kInvalidFilter, 0,
+                  kInvalidFilter, 0, Width(), Height()));
+
+            for (int y = 0; y < Height(); ++y)
+              for (int x = 0; x < Width(); ++x)
+                ASSERT_EQ(lookup(ref, y * kOutputStride + x),
+                          lookup(out, y * kOutputStride + x))
+                    << "mismatch at (" << x << "," << y << "), "
+                    << "filters (" << filter_bank << "," << filter_x << ","
+                    << filter_y << ")";
+          }
+        }
+      }
+    }
+  }
+}
+
+/* This test exercises that enough rows and columns are filtered with every
+   possible initial fractional positions and scaling steps. */
+TEST_P(ConvolveTest, CheckScalingFiltering) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+  const InterpKernel *const eighttap =
+      (const InterpKernel *)av1_get_interp_filter_kernel(EIGHTTAP_REGULAR);
+
+  SetConstantInput(127);
+
+  for (int frac = 0; frac < 16; ++frac) {
+    for (int step = 1; step <= 32; ++step) {
+      /* Test the horizontal and vertical filters in combination. */
+      ASM_REGISTER_STATE_CHECK(UUT_->shv8_(in, kInputStride, out, kOutputStride,
+                                           eighttap[frac], step, eighttap[frac],
+                                           step, Width(), Height()));
+
+      CheckGuardBlocks();
+
+      for (int y = 0; y < Height(); ++y) {
+        for (int x = 0; x < Width(); ++x) {
+          ASSERT_EQ(lookup(in, y * kInputStride + x),
+                    lookup(out, y * kOutputStride + x))
+              << "x == " << x << ", y == " << y << ", frac == " << frac
+              << ", step == " << step;
+        }
+      }
+    }
+  }
+}
+
+TEST_P(ConvolveTest, DISABLED_Copy_Speed) {
+  const uint8_t *const in = input();
+  uint8_t *const out = output();
+  const int kNumTests = 5000000;
+  const int width = Width();
+  const int height = Height();
+  aom_usec_timer timer;
+
+  aom_usec_timer_start(&timer);
+  for (int n = 0; n < kNumTests; ++n) {
+    UUT_->copy_(in, kInputStride, out, kOutputStride, NULL, 0, NULL, 0, width,
+                height);
+  }
+  aom_usec_timer_mark(&timer);
+
+  const int elapsed_time = static_cast<int>(aom_usec_timer_elapsed(&timer));
+  printf("convolve_copy_%dx%d_%d: %d us\n", width, height,
+         UUT_->use_highbd_ ? UUT_->use_highbd_ : 8, elapsed_time);
+}
+
+TEST_P(ConvolveTest, DISABLED_Avg_Speed) {
+  const uint8_t *const in = input();
+  uint8_t *const out = output();
+  const int kNumTests = 5000000;
+  const int width = Width();
+  const int height = Height();
+  aom_usec_timer timer;
+
+  aom_usec_timer_start(&timer);
+  for (int n = 0; n < kNumTests; ++n) {
+    UUT_->avg_(in, kInputStride, out, kOutputStride, NULL, 0, NULL, 0, width,
+               height);
+  }
+  aom_usec_timer_mark(&timer);
+
+  const int elapsed_time = static_cast<int>(aom_usec_timer_elapsed(&timer));
+  printf("convolve_avg_%dx%d_%d: %d us\n", width, height,
+         UUT_->use_highbd_ ? UUT_->use_highbd_ : 8, elapsed_time);
+}
+
+TEST_P(ConvolveTest, DISABLED_Speed) {
+  uint8_t *const in = input();
+  uint8_t *const out = output();
+#if CONFIG_HIGHBITDEPTH
+  uint8_t ref8[kOutputStride * kMaxDimension];
+  uint16_t ref16[kOutputStride * kMaxDimension];
+  uint8_t *ref;
+  if (UUT_->use_highbd_ == 0) {
+    ref = ref8;
+  } else {
+    ref = CONVERT_TO_BYTEPTR(ref16);
+  }
+#else
+  uint8_t ref[kOutputStride * kMaxDimension];
+#endif
+
+  // Populate ref and out with some random data
+  ::libaom_test::ACMRandom prng;
+  for (int y = 0; y < Height(); ++y) {
+    for (int x = 0; x < Width(); ++x) {
+      uint16_t r;
+#if CONFIG_HIGHBITDEPTH
+      if (UUT_->use_highbd_ == 0 || UUT_->use_highbd_ == 8) {
+        r = prng.Rand8Extremes();
+      } else {
+        r = prng.Rand16() & mask_;
+      }
+#else
+      r = prng.Rand8Extremes();
+#endif
+
+      assign_val(out, y * kOutputStride + x, r);
+      assign_val(ref, y * kOutputStride + x, r);
+    }
+  }
+
+  const InterpFilter filter = (InterpFilter)1;
+  const InterpKernel *filters =
+      (const InterpKernel *)av1_get_interp_filter_kernel(filter);
+  wrapper_filter_average_block2d_8_c(in, kInputStride, filters[1], filters[1],
+                                     out, kOutputStride, Width(), Height());
+
+  aom_usec_timer timer;
+  int tests_num = 1000;
+
+  aom_usec_timer_start(&timer);
+  while (tests_num > 0) {
+    for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
+      const InterpFilter filter = (InterpFilter)filter_bank;
+      const InterpKernel *filters =
+          (const InterpKernel *)av1_get_interp_filter_kernel(filter);
+#if CONFIG_DUAL_FILTER
+      const InterpFilterParams filter_params =
+          av1_get_interp_filter_params(filter);
+      if (filter_params.taps != SUBPEL_TAPS) continue;
+#endif
+
+      for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
+        for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
+          if (filter_x && filter_y)
+            ASM_REGISTER_STATE_CHECK(UUT_->hv8_(
+                in, kInputStride, out, kOutputStride, filters[filter_x], 16,
+                filters[filter_y], 16, Width(), Height()));
+          if (filter_y)
+            ASM_REGISTER_STATE_CHECK(
+                UUT_->v8_(in, kInputStride, out, kOutputStride, kInvalidFilter,
+                          16, filters[filter_y], 16, Width(), Height()));
+          else if (filter_x)
+            ASM_REGISTER_STATE_CHECK(UUT_->h8_(
+                in, kInputStride, out, kOutputStride, filters[filter_x], 16,
+                kInvalidFilter, 16, Width(), Height()));
+        }
+      }
+    }
+    tests_num--;
+  }
+  aom_usec_timer_mark(&timer);
+
+  const int elapsed_time =
+      static_cast<int>(aom_usec_timer_elapsed(&timer) / 1000);
+  printf("%dx%d (bitdepth %d) time: %5d ms\n", Width(), Height(),
+         UUT_->use_highbd_, elapsed_time);
+}
+
+using std::tr1::make_tuple;
+
+#if CONFIG_HIGHBITDEPTH
+#define WRAP(func, bd)                                                       \
+  void wrap_##func##_##bd(                                                   \
+      const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,                \
+      ptrdiff_t dst_stride, const int16_t *filter_x, int filter_x_stride,    \
+      const int16_t *filter_y, int filter_y_stride, int w, int h) {          \
+    aom_highbd_##func(src, src_stride, dst, dst_stride, filter_x,            \
+                      filter_x_stride, filter_y, filter_y_stride, w, h, bd); \
+  }
+#if HAVE_SSE2 && ARCH_X86_64
+WRAP(convolve_copy_sse2, 8)
+WRAP(convolve_avg_sse2, 8)
+WRAP(convolve_copy_sse2, 10)
+WRAP(convolve_avg_sse2, 10)
+WRAP(convolve_copy_sse2, 12)
+WRAP(convolve_avg_sse2, 12)
+WRAP(convolve8_horiz_sse2, 8)
+WRAP(convolve8_avg_horiz_sse2, 8)
+WRAP(convolve8_vert_sse2, 8)
+WRAP(convolve8_avg_vert_sse2, 8)
+WRAP(convolve8_sse2, 8)
+WRAP(convolve8_avg_sse2, 8)
+WRAP(convolve8_horiz_sse2, 10)
+WRAP(convolve8_avg_horiz_sse2, 10)
+WRAP(convolve8_vert_sse2, 10)
+WRAP(convolve8_avg_vert_sse2, 10)
+WRAP(convolve8_sse2, 10)
+WRAP(convolve8_avg_sse2, 10)
+WRAP(convolve8_horiz_sse2, 12)
+WRAP(convolve8_avg_horiz_sse2, 12)
+WRAP(convolve8_vert_sse2, 12)
+WRAP(convolve8_avg_vert_sse2, 12)
+WRAP(convolve8_sse2, 12)
+WRAP(convolve8_avg_sse2, 12)
+#endif  // HAVE_SSE2 && ARCH_X86_64
+
+WRAP(convolve_copy_c, 8)
+WRAP(convolve_avg_c, 8)
+WRAP(convolve8_horiz_c, 8)
+WRAP(convolve8_avg_horiz_c, 8)
+WRAP(convolve8_vert_c, 8)
+WRAP(convolve8_avg_vert_c, 8)
+WRAP(convolve8_c, 8)
+WRAP(convolve8_avg_c, 8)
+WRAP(convolve_copy_c, 10)
+WRAP(convolve_avg_c, 10)
+WRAP(convolve8_horiz_c, 10)
+WRAP(convolve8_avg_horiz_c, 10)
+WRAP(convolve8_vert_c, 10)
+WRAP(convolve8_avg_vert_c, 10)
+WRAP(convolve8_c, 10)
+WRAP(convolve8_avg_c, 10)
+WRAP(convolve_copy_c, 12)
+WRAP(convolve_avg_c, 12)
+WRAP(convolve8_horiz_c, 12)
+WRAP(convolve8_avg_horiz_c, 12)
+WRAP(convolve8_vert_c, 12)
+WRAP(convolve8_avg_vert_c, 12)
+WRAP(convolve8_c, 12)
+WRAP(convolve8_avg_c, 12)
+
+#if HAVE_AVX2
+WRAP(convolve_copy_avx2, 8)
+WRAP(convolve_avg_avx2, 8)
+WRAP(convolve8_horiz_avx2, 8)
+WRAP(convolve8_avg_horiz_avx2, 8)
+WRAP(convolve8_vert_avx2, 8)
+WRAP(convolve8_avg_vert_avx2, 8)
+WRAP(convolve8_avx2, 8)
+WRAP(convolve8_avg_avx2, 8)
+
+WRAP(convolve_copy_avx2, 10)
+WRAP(convolve_avg_avx2, 10)
+WRAP(convolve8_avx2, 10)
+WRAP(convolve8_horiz_avx2, 10)
+WRAP(convolve8_vert_avx2, 10)
+WRAP(convolve8_avg_avx2, 10)
+WRAP(convolve8_avg_horiz_avx2, 10)
+WRAP(convolve8_avg_vert_avx2, 10)
+
+WRAP(convolve_copy_avx2, 12)
+WRAP(convolve_avg_avx2, 12)
+WRAP(convolve8_avx2, 12)
+WRAP(convolve8_horiz_avx2, 12)
+WRAP(convolve8_vert_avx2, 12)
+WRAP(convolve8_avg_avx2, 12)
+WRAP(convolve8_avg_horiz_avx2, 12)
+WRAP(convolve8_avg_vert_avx2, 12)
+#endif  // HAVE_AVX2
+
+#undef WRAP
+
+const ConvolveFunctions convolve8_c(
+    wrap_convolve_copy_c_8, wrap_convolve_avg_c_8, wrap_convolve8_horiz_c_8,
+    wrap_convolve8_avg_horiz_c_8, wrap_convolve8_vert_c_8,
+    wrap_convolve8_avg_vert_c_8, wrap_convolve8_c_8, wrap_convolve8_avg_c_8,
+    wrap_convolve8_horiz_c_8, wrap_convolve8_avg_horiz_c_8,
+    wrap_convolve8_vert_c_8, wrap_convolve8_avg_vert_c_8, wrap_convolve8_c_8,
+    wrap_convolve8_avg_c_8, 8);
+const ConvolveFunctions convolve10_c(
+    wrap_convolve_copy_c_10, wrap_convolve_avg_c_10, wrap_convolve8_horiz_c_10,
+    wrap_convolve8_avg_horiz_c_10, wrap_convolve8_vert_c_10,
+    wrap_convolve8_avg_vert_c_10, wrap_convolve8_c_10, wrap_convolve8_avg_c_10,
+    wrap_convolve8_horiz_c_10, wrap_convolve8_avg_horiz_c_10,
+    wrap_convolve8_vert_c_10, wrap_convolve8_avg_vert_c_10, wrap_convolve8_c_10,
+    wrap_convolve8_avg_c_10, 10);
+const ConvolveFunctions convolve12_c(
+    wrap_convolve_copy_c_12, wrap_convolve_avg_c_12, wrap_convolve8_horiz_c_12,
+    wrap_convolve8_avg_horiz_c_12, wrap_convolve8_vert_c_12,
+    wrap_convolve8_avg_vert_c_12, wrap_convolve8_c_12, wrap_convolve8_avg_c_12,
+    wrap_convolve8_horiz_c_12, wrap_convolve8_avg_horiz_c_12,
+    wrap_convolve8_vert_c_12, wrap_convolve8_avg_vert_c_12, wrap_convolve8_c_12,
+    wrap_convolve8_avg_c_12, 12);
+const ConvolveParam kArrayConvolve_c[] = {
+  ALL_SIZES(convolve8_c), ALL_SIZES(convolve10_c), ALL_SIZES(convolve12_c)
+};
+
+#else
+const ConvolveFunctions convolve8_c(
+    aom_convolve_copy_c, aom_convolve_avg_c, aom_convolve8_horiz_c,
+    aom_convolve8_avg_horiz_c, aom_convolve8_vert_c, aom_convolve8_avg_vert_c,
+    aom_convolve8_c, aom_convolve8_avg_c, aom_scaled_horiz_c,
+    aom_scaled_avg_horiz_c, aom_scaled_vert_c, aom_scaled_avg_vert_c,
+    aom_scaled_2d_c, aom_scaled_avg_2d_c, 0);
+const ConvolveParam kArrayConvolve_c[] = { ALL_SIZES(convolve8_c) };
+#endif
+INSTANTIATE_TEST_CASE_P(C, ConvolveTest, ::testing::ValuesIn(kArrayConvolve_c));
+
+#if HAVE_SSE2 && ARCH_X86_64
+#if CONFIG_HIGHBITDEPTH
+const ConvolveFunctions convolve8_sse2(
+    wrap_convolve_copy_sse2_8, wrap_convolve_avg_sse2_8,
+    wrap_convolve8_horiz_sse2_8, wrap_convolve8_avg_horiz_sse2_8,
+    wrap_convolve8_vert_sse2_8, wrap_convolve8_avg_vert_sse2_8,
+    wrap_convolve8_sse2_8, wrap_convolve8_avg_sse2_8,
+    wrap_convolve8_horiz_sse2_8, wrap_convolve8_avg_horiz_sse2_8,
+    wrap_convolve8_vert_sse2_8, wrap_convolve8_avg_vert_sse2_8,
+    wrap_convolve8_sse2_8, wrap_convolve8_avg_sse2_8, 8);
+const ConvolveFunctions convolve10_sse2(
+    wrap_convolve_copy_sse2_10, wrap_convolve_avg_sse2_10,
+    wrap_convolve8_horiz_sse2_10, wrap_convolve8_avg_horiz_sse2_10,
+    wrap_convolve8_vert_sse2_10, wrap_convolve8_avg_vert_sse2_10,
+    wrap_convolve8_sse2_10, wrap_convolve8_avg_sse2_10,
+    wrap_convolve8_horiz_sse2_10, wrap_convolve8_avg_horiz_sse2_10,
+    wrap_convolve8_vert_sse2_10, wrap_convolve8_avg_vert_sse2_10,
+    wrap_convolve8_sse2_10, wrap_convolve8_avg_sse2_10, 10);
+const ConvolveFunctions convolve12_sse2(
+    wrap_convolve_copy_sse2_12, wrap_convolve_avg_sse2_12,
+    wrap_convolve8_horiz_sse2_12, wrap_convolve8_avg_horiz_sse2_12,
+    wrap_convolve8_vert_sse2_12, wrap_convolve8_avg_vert_sse2_12,
+    wrap_convolve8_sse2_12, wrap_convolve8_avg_sse2_12,
+    wrap_convolve8_horiz_sse2_12, wrap_convolve8_avg_horiz_sse2_12,
+    wrap_convolve8_vert_sse2_12, wrap_convolve8_avg_vert_sse2_12,
+    wrap_convolve8_sse2_12, wrap_convolve8_avg_sse2_12, 12);
+const ConvolveParam kArrayConvolve_sse2[] = { ALL_SIZES(convolve8_sse2),
+                                              ALL_SIZES(convolve10_sse2),
+                                              ALL_SIZES(convolve12_sse2) };
+#else
+const ConvolveFunctions convolve8_sse2(
+    aom_convolve_copy_sse2, aom_convolve_avg_sse2, aom_convolve8_horiz_sse2,
+    aom_convolve8_avg_horiz_sse2, aom_convolve8_vert_sse2,
+    aom_convolve8_avg_vert_sse2, aom_convolve8_sse2, aom_convolve8_avg_sse2,
+    aom_scaled_horiz_c, aom_scaled_avg_horiz_c, aom_scaled_vert_c,
+    aom_scaled_avg_vert_c, aom_scaled_2d_c, aom_scaled_avg_2d_c, 0);
+
+const ConvolveParam kArrayConvolve_sse2[] = { ALL_SIZES(convolve8_sse2) };
+#endif  // CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(SSE2, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve_sse2));
+#endif
+
+#if HAVE_SSSE3
+const ConvolveFunctions convolve8_ssse3(
+    aom_convolve_copy_c, aom_convolve_avg_c, aom_convolve8_horiz_ssse3,
+    aom_convolve8_avg_horiz_ssse3, aom_convolve8_vert_ssse3,
+    aom_convolve8_avg_vert_ssse3, aom_convolve8_ssse3, aom_convolve8_avg_ssse3,
+    aom_scaled_horiz_c, aom_scaled_avg_horiz_c, aom_scaled_vert_c,
+    aom_scaled_avg_vert_c, aom_scaled_2d_ssse3, aom_scaled_avg_2d_c, 0);
+
+const ConvolveParam kArrayConvolve8_ssse3[] = { ALL_SIZES(convolve8_ssse3) };
+INSTANTIATE_TEST_CASE_P(SSSE3, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve8_ssse3));
+#endif
+
+#if HAVE_AVX2
+#if CONFIG_HIGHBITDEPTH
+const ConvolveFunctions convolve8_avx2(
+    wrap_convolve_copy_avx2_8, wrap_convolve_avg_avx2_8,
+    wrap_convolve8_horiz_avx2_8, wrap_convolve8_avg_horiz_avx2_8,
+    wrap_convolve8_vert_avx2_8, wrap_convolve8_avg_vert_avx2_8,
+    wrap_convolve8_avx2_8, wrap_convolve8_avg_avx2_8, wrap_convolve8_horiz_c_8,
+    wrap_convolve8_avg_horiz_c_8, wrap_convolve8_vert_c_8,
+    wrap_convolve8_avg_vert_c_8, wrap_convolve8_c_8, wrap_convolve8_avg_c_8, 8);
+const ConvolveFunctions convolve10_avx2(
+    wrap_convolve_copy_avx2_10, wrap_convolve_avg_avx2_10,
+    wrap_convolve8_horiz_avx2_10, wrap_convolve8_avg_horiz_avx2_10,
+    wrap_convolve8_vert_avx2_10, wrap_convolve8_avg_vert_avx2_10,
+    wrap_convolve8_avx2_10, wrap_convolve8_avg_avx2_10,
+    wrap_convolve8_horiz_c_10, wrap_convolve8_avg_horiz_c_10,
+    wrap_convolve8_vert_c_10, wrap_convolve8_avg_vert_c_10, wrap_convolve8_c_10,
+    wrap_convolve8_avg_c_10, 10);
+const ConvolveFunctions convolve12_avx2(
+    wrap_convolve_copy_avx2_12, wrap_convolve_avg_avx2_12,
+    wrap_convolve8_horiz_avx2_12, wrap_convolve8_avg_horiz_avx2_12,
+    wrap_convolve8_vert_avx2_12, wrap_convolve8_avg_vert_avx2_12,
+    wrap_convolve8_avx2_12, wrap_convolve8_avg_avx2_12,
+    wrap_convolve8_horiz_c_12, wrap_convolve8_avg_horiz_c_12,
+    wrap_convolve8_vert_c_12, wrap_convolve8_avg_vert_c_12, wrap_convolve8_c_12,
+    wrap_convolve8_avg_c_12, 12);
+const ConvolveParam kArrayConvolve8_avx2[] = { ALL_SIZES(convolve8_avx2),
+                                               ALL_SIZES(convolve10_avx2),
+                                               ALL_SIZES(convolve12_avx2) };
+#else
+const ConvolveFunctions convolve8_avx2(
+    aom_convolve_copy_c, aom_convolve_avg_c, aom_convolve8_horiz_avx2,
+    aom_convolve8_avg_horiz_ssse3, aom_convolve8_vert_avx2,
+    aom_convolve8_avg_vert_ssse3, aom_convolve8_avx2, aom_convolve8_avg_ssse3,
+    aom_scaled_horiz_c, aom_scaled_avg_horiz_c, aom_scaled_vert_c,
+    aom_scaled_avg_vert_c, aom_scaled_2d_c, aom_scaled_avg_2d_c, 0);
+
+const ConvolveParam kArrayConvolve8_avx2[] = { ALL_SIZES(convolve8_avx2) };
+#endif  // CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(AVX2, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve8_avx2));
+#endif  // HAVE_AVX2
+
+// TODO(any): Make NEON versions support 128x128 128x64 64x128 block sizes
+#if HAVE_NEON && !(CONFIG_AV1 && CONFIG_EXT_PARTITION)
+#if HAVE_NEON_ASM
+const ConvolveFunctions convolve8_neon(
+    aom_convolve_copy_neon, aom_convolve_avg_neon, aom_convolve8_horiz_neon,
+    aom_convolve8_avg_horiz_neon, aom_convolve8_vert_neon,
+    aom_convolve8_avg_vert_neon, aom_convolve8_neon, aom_convolve8_avg_neon,
+    aom_scaled_horiz_c, aom_scaled_avg_horiz_c, aom_scaled_vert_c,
+    aom_scaled_avg_vert_c, aom_scaled_2d_c, aom_scaled_avg_2d_c, 0);
+#else   // HAVE_NEON
+const ConvolveFunctions convolve8_neon(
+    aom_convolve_copy_neon, aom_convolve_avg_neon, aom_convolve8_horiz_neon,
+    aom_convolve8_avg_horiz_neon, aom_convolve8_vert_neon,
+    aom_convolve8_avg_vert_neon, aom_convolve8_neon, aom_convolve8_avg_neon,
+    aom_scaled_horiz_c, aom_scaled_avg_horiz_c, aom_scaled_vert_c,
+    aom_scaled_avg_vert_c, aom_scaled_2d_c, aom_scaled_avg_2d_c, 0);
+#endif  // HAVE_NEON_ASM
+
+const ConvolveParam kArrayConvolve8_neon[] = { ALL_SIZES(convolve8_neon) };
+INSTANTIATE_TEST_CASE_P(NEON, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve8_neon));
+#endif  // HAVE_NEON
+
+// TODO(any): Make DSPR2 versions support 128x128 128x64 64x128 block sizes
+#if HAVE_DSPR2 && !(CONFIG_AV1 && CONFIG_EXT_PARTITION)
+const ConvolveFunctions convolve8_dspr2(
+    aom_convolve_copy_dspr2, aom_convolve_avg_dspr2, aom_convolve8_horiz_dspr2,
+    aom_convolve8_avg_horiz_dspr2, aom_convolve8_vert_dspr2,
+    aom_convolve8_avg_vert_dspr2, aom_convolve8_dspr2, aom_convolve8_avg_dspr2,
+    aom_scaled_horiz_c, aom_scaled_avg_horiz_c, aom_scaled_vert_c,
+    aom_scaled_avg_vert_c, aom_scaled_2d_c, aom_scaled_avg_2d_c, 0);
+
+const ConvolveParam kArrayConvolve8_dspr2[] = { ALL_SIZES(convolve8_dspr2) };
+INSTANTIATE_TEST_CASE_P(DSPR2, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve8_dspr2));
+#endif  // HAVE_DSPR2
+
+// TODO(any): Make MSA versions support 128x128 128x64 64x128 block sizes
+#if HAVE_MSA && !(CONFIG_AV1 && CONFIG_EXT_PARTITION)
+const ConvolveFunctions convolve8_msa(
+    aom_convolve_copy_msa, aom_convolve_avg_msa, aom_convolve8_horiz_msa,
+    aom_convolve8_avg_horiz_msa, aom_convolve8_vert_msa,
+    aom_convolve8_avg_vert_msa, aom_convolve8_msa, aom_convolve8_avg_msa,
+    aom_scaled_horiz_c, aom_scaled_avg_horiz_c, aom_scaled_vert_c,
+    aom_scaled_avg_vert_c, aom_scaled_2d_c, aom_scaled_avg_2d_c, 0);
+
+const ConvolveParam kArrayConvolve8_msa[] = { ALL_SIZES(convolve8_msa) };
+INSTANTIATE_TEST_CASE_P(MSA, ConvolveTest,
+                        ::testing::ValuesIn(kArrayConvolve8_msa));
+#endif  // HAVE_MSA
+}  // namespace
diff --git a/third_party/aom/test/cpu_speed_test.cc b/third_party/aom/test/cpu_speed_test.cc
new file mode 100644
index 0000000000..9b79664627
--- /dev/null
+++ b/third_party/aom/test/cpu_speed_test.cc
@@ -0,0 +1,180 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+
+namespace {
+
+const int kMaxPSNR = 100;
+
+class CpuSpeedTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<libaom_test::TestMode, int> {
+ protected:
+  CpuSpeedTest()
+      : EncoderTest(GET_PARAM(0)), encoding_mode_(GET_PARAM(1)),
+        set_cpu_used_(GET_PARAM(2)), min_psnr_(kMaxPSNR),
+        tune_content_(AOM_CONTENT_DEFAULT) {}
+  virtual ~CpuSpeedTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+    if (encoding_mode_ != ::libaom_test::kRealTime) {
+      cfg_.g_lag_in_frames = 25;
+      cfg_.rc_end_usage = AOM_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = AOM_CBR;
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) { min_psnr_ = kMaxPSNR; }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AOME_SET_CPUUSED, set_cpu_used_);
+      encoder->Control(AV1E_SET_TUNE_CONTENT, tune_content_);
+      if (encoding_mode_ != ::libaom_test::kRealTime) {
+        encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(AOME_SET_ARNR_STRENGTH, 5);
+      }
+    }
+  }
+
+  virtual void PSNRPktHook(const aom_codec_cx_pkt_t *pkt) {
+    if (pkt->data.psnr.psnr[0] < min_psnr_) min_psnr_ = pkt->data.psnr.psnr[0];
+  }
+
+  void TestQ0();
+  void TestScreencastQ0();
+  void TestTuneScreen();
+  void TestEncodeHighBitrate();
+  void TestLowBitrate();
+
+  ::libaom_test::TestMode encoding_mode_;
+  int set_cpu_used_;
+  double min_psnr_;
+  int tune_content_;
+};
+
+void CpuSpeedTest::TestQ0() {
+  // Validate that this non multiple of 64 wide clip encodes and decodes
+  // without a mismatch when passing in a very low max q.  This pushes
+  // the encoder to producing lots of big partitions which will likely
+  // extend into the border and test the border condition.
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 400;
+  cfg_.rc_max_quantizer = 0;
+  cfg_.rc_min_quantizer = 0;
+
+  ::libaom_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       10);
+
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_GE(min_psnr_, kMaxPSNR);
+}
+
+void CpuSpeedTest::TestScreencastQ0() {
+  ::libaom_test::Y4mVideoSource video("screendata.y4m", 0, 10);
+  cfg_.g_timebase = video.timebase();
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 400;
+  cfg_.rc_max_quantizer = 0;
+  cfg_.rc_min_quantizer = 0;
+
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_GE(min_psnr_, kMaxPSNR);
+}
+
+void CpuSpeedTest::TestTuneScreen() {
+  ::libaom_test::Y4mVideoSource video("screendata.y4m", 0, 10);
+  cfg_.g_timebase = video.timebase();
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_minsection_pct = 2000;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_min_quantizer = 0;
+  tune_content_ = AOM_CONTENT_SCREEN;
+
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+void CpuSpeedTest::TestEncodeHighBitrate() {
+  // Validate that this non multiple of 64 wide clip encodes and decodes
+  // without a mismatch when passing in a very low max q.  This pushes
+  // the encoder to producing lots of big partitions which will likely
+  // extend into the border and test the border condition.
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 12000;
+  cfg_.rc_max_quantizer = 10;
+  cfg_.rc_min_quantizer = 0;
+
+  ::libaom_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       10);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+void CpuSpeedTest::TestLowBitrate() {
+  // Validate that this clip encodes and decodes without a mismatch
+  // when passing in a very high min q.  This pushes the encoder to producing
+  // lots of small partitions which might will test the other condition.
+  cfg_.rc_2pass_vbr_minsection_pct = 5;
+  cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+  cfg_.rc_target_bitrate = 200;
+  cfg_.rc_min_quantizer = 40;
+
+  ::libaom_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       10);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+TEST_P(CpuSpeedTest, TestQ0) { TestQ0(); }
+TEST_P(CpuSpeedTest, TestScreencastQ0) { TestScreencastQ0(); }
+TEST_P(CpuSpeedTest, TestTuneScreen) { TestTuneScreen(); }
+TEST_P(CpuSpeedTest, TestEncodeHighBitrate) { TestEncodeHighBitrate(); }
+TEST_P(CpuSpeedTest, TestLowBitrate) { TestLowBitrate(); }
+
+class CpuSpeedTestLarge : public CpuSpeedTest {};
+
+TEST_P(CpuSpeedTestLarge, TestQ0) { TestQ0(); }
+TEST_P(CpuSpeedTestLarge, TestScreencastQ0) { TestScreencastQ0(); }
+TEST_P(CpuSpeedTestLarge, TestTuneScreen) { TestTuneScreen(); }
+TEST_P(CpuSpeedTestLarge, TestEncodeHighBitrate) { TestEncodeHighBitrate(); }
+TEST_P(CpuSpeedTestLarge, TestLowBitrate) { TestLowBitrate(); }
+
+AV1_INSTANTIATE_TEST_CASE(CpuSpeedTest,
+                          ::testing::Values(::libaom_test::kTwoPassGood,
+                                            ::libaom_test::kOnePassGood),
+                          ::testing::Range(1, 3));
+AV1_INSTANTIATE_TEST_CASE(CpuSpeedTestLarge,
+                          ::testing::Values(::libaom_test::kTwoPassGood,
+                                            ::libaom_test::kOnePassGood),
+                          ::testing::Range(0, 1));
+}  // namespace
diff --git a/third_party/aom/test/datarate_test.cc b/third_party/aom/test/datarate_test.cc
new file mode 100644
index 0000000000..48be4a46df
--- /dev/null
+++ b/third_party/aom/test/datarate_test.cc
@@ -0,0 +1,253 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "./aom_config.h"
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "aom/aom_codec.h"
+
+namespace {
+
+class DatarateTestLarge
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<libaom_test::TestMode, int> {
+ public:
+  DatarateTestLarge() : EncoderTest(GET_PARAM(0)) {}
+
+ protected:
+  virtual ~DatarateTestLarge() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    set_cpu_used_ = GET_PARAM(2);
+    ResetModel();
+  }
+
+  virtual void ResetModel() {
+    last_pts_ = 0;
+    bits_in_buffer_model_ = cfg_.rc_target_bitrate * cfg_.rc_buf_initial_sz;
+    frame_number_ = 0;
+    tot_frame_number_ = 0;
+    first_drop_ = 0;
+    num_drops_ = 0;
+    // Denoiser is off by default.
+    denoiser_on_ = 0;
+    bits_total_ = 0;
+    denoiser_offon_test_ = 0;
+    denoiser_offon_period_ = -1;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 0) encoder->Control(AOME_SET_CPUUSED, set_cpu_used_);
+
+    if (denoiser_offon_test_) {
+      ASSERT_GT(denoiser_offon_period_, 0)
+          << "denoiser_offon_period_ is not positive.";
+      if ((video->frame() + 1) % denoiser_offon_period_ == 0) {
+        // Flip denoiser_on_ periodically
+        denoiser_on_ ^= 1;
+      }
+    }
+
+    encoder->Control(AV1E_SET_NOISE_SENSITIVITY, denoiser_on_);
+
+    const aom_rational_t tb = video->timebase();
+    timebase_ = static_cast<double>(tb.num) / tb.den;
+    duration_ = 0;
+  }
+
+  virtual void FramePktHook(const aom_codec_cx_pkt_t *pkt) {
+    // Time since last timestamp = duration.
+    aom_codec_pts_t duration = pkt->data.frame.pts - last_pts_;
+
+    if (duration > 1) {
+      // If first drop not set and we have a drop set it to this time.
+      if (!first_drop_) first_drop_ = last_pts_ + 1;
+      // Update the number of frame drops.
+      num_drops_ += static_cast<int>(duration - 1);
+      // Update counter for total number of frames (#frames input to encoder).
+      // Needed for setting the proper layer_id below.
+      tot_frame_number_ += static_cast<int>(duration - 1);
+    }
+
+    // Add to the buffer the bits we'd expect from a constant bitrate server.
+    bits_in_buffer_model_ += static_cast<int64_t>(
+        duration * timebase_ * cfg_.rc_target_bitrate * 1000);
+
+    // Buffer should not go negative.
+    ASSERT_GE(bits_in_buffer_model_, 0) << "Buffer Underrun at frame "
+                                        << pkt->data.frame.pts;
+
+    const size_t frame_size_in_bits = pkt->data.frame.sz * 8;
+
+    // Update the total encoded bits.
+    bits_total_ += frame_size_in_bits;
+
+    // Update the most recent pts.
+    last_pts_ = pkt->data.frame.pts;
+    ++frame_number_;
+    ++tot_frame_number_;
+  }
+
+  virtual void EndPassHook(void) {
+    duration_ = (last_pts_ + 1) * timebase_;
+    // Effective file datarate:
+    effective_datarate_ = (bits_total_ / 1000.0) / duration_;
+  }
+
+  aom_codec_pts_t last_pts_;
+  double timebase_;
+  int frame_number_;      // Counter for number of non-dropped/encoded frames.
+  int tot_frame_number_;  // Counter for total number of input frames.
+  int64_t bits_total_;
+  double duration_;
+  double effective_datarate_;
+  int set_cpu_used_;
+  int64_t bits_in_buffer_model_;
+  aom_codec_pts_t first_drop_;
+  int num_drops_;
+  int denoiser_on_;
+  int denoiser_offon_test_;
+  int denoiser_offon_period_;
+};
+
+// Check basic rate targeting for VBR mode.
+TEST_P(DatarateTestLarge, BasicRateTargetingVBR) {
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.g_error_resilient = 0;
+  cfg_.rc_end_usage = AOM_VBR;
+  cfg_.g_lag_in_frames = 0;
+
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 140);
+  for (int i = 400; i <= 800; i += 400) {
+    cfg_.rc_target_bitrate = i;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(effective_datarate_, cfg_.rc_target_bitrate * 0.75)
+        << " The datarate for the file is lower than target by too much!";
+    ASSERT_LE(effective_datarate_, cfg_.rc_target_bitrate * 1.25)
+        << " The datarate for the file is greater than target by too much!";
+  }
+}
+
+// Check basic rate targeting for CBR,
+TEST_P(DatarateTestLarge, BasicRateTargeting) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = AOM_CBR;
+  cfg_.g_lag_in_frames = 0;
+
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 140);
+  for (int i = 150; i < 800; i += 400) {
+    cfg_.rc_target_bitrate = i;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(effective_datarate_, cfg_.rc_target_bitrate * 0.85)
+        << " The datarate for the file is lower than target by too much!";
+    ASSERT_LE(effective_datarate_, cfg_.rc_target_bitrate * 1.15)
+        << " The datarate for the file is greater than target by too much!";
+  }
+}
+
+// Check basic rate targeting for CBR.
+TEST_P(DatarateTestLarge, BasicRateTargeting444) {
+  ::libaom_test::Y4mVideoSource video("rush_hour_444.y4m", 0, 140);
+
+  cfg_.g_profile = 1;
+  cfg_.g_timebase = video.timebase();
+
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_dropframe_thresh = 1;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 63;
+  cfg_.rc_end_usage = AOM_CBR;
+
+  for (int i = 250; i < 900; i += 400) {
+    cfg_.rc_target_bitrate = i;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(static_cast<double>(cfg_.rc_target_bitrate),
+              effective_datarate_ * 0.85)
+        << " The datarate for the file exceeds the target by too much!";
+    ASSERT_LE(static_cast<double>(cfg_.rc_target_bitrate),
+              effective_datarate_ * 1.15)
+        << " The datarate for the file missed the target!"
+        << cfg_.rc_target_bitrate << " " << effective_datarate_;
+  }
+}
+
+// Check that (1) the first dropped frame gets earlier and earlier
+// as the drop frame threshold is increased, and (2) that the total number of
+// frame drops does not decrease as we increase frame drop threshold.
+// Use a lower qp-max to force some frame drops.
+TEST_P(DatarateTestLarge, ChangingDropFrameThresh) {
+  cfg_.rc_buf_initial_sz = 500;
+  cfg_.rc_buf_optimal_sz = 500;
+  cfg_.rc_buf_sz = 1000;
+  cfg_.rc_undershoot_pct = 20;
+  cfg_.rc_undershoot_pct = 20;
+  cfg_.rc_dropframe_thresh = 10;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 50;
+  cfg_.rc_end_usage = AOM_CBR;
+  cfg_.rc_target_bitrate = 200;
+  cfg_.g_lag_in_frames = 0;
+  // TODO(marpan): Investigate datarate target failures with a smaller keyframe
+  // interval (128).
+  cfg_.kf_max_dist = 9999;
+
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 100);
+
+  const int kDropFrameThreshTestStep = 30;
+  aom_codec_pts_t last_drop = 140;
+  int last_num_drops = 0;
+  for (int i = 40; i < 100; i += kDropFrameThreshTestStep) {
+    cfg_.rc_dropframe_thresh = i;
+    ResetModel();
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    ASSERT_GE(effective_datarate_, cfg_.rc_target_bitrate * 0.85)
+        << " The datarate for the file is lower than target by too much!";
+    ASSERT_LE(effective_datarate_, cfg_.rc_target_bitrate * 1.15)
+        << " The datarate for the file is greater than target by too much!";
+    ASSERT_LE(first_drop_, last_drop)
+        << " The first dropped frame for drop_thresh " << i
+        << " > first dropped frame for drop_thresh "
+        << i - kDropFrameThreshTestStep;
+    ASSERT_GE(num_drops_, last_num_drops * 0.85)
+        << " The number of dropped frames for drop_thresh " << i
+        << " < number of dropped frames for drop_thresh "
+        << i - kDropFrameThreshTestStep;
+    last_drop = first_drop_;
+    last_num_drops = num_drops_;
+  }
+}
+
+AV1_INSTANTIATE_TEST_CASE(DatarateTestLarge,
+                          ::testing::Values(::libaom_test::kOnePassGood,
+                                            ::libaom_test::kRealTime),
+                          ::testing::Range(2, 9, 2));
+}  // namespace
diff --git a/third_party/aom/test/dct16x16_test.cc b/third_party/aom/test/dct16x16_test.cc
new file mode 100644
index 0000000000..89263ce89e
--- /dev/null
+++ b/third_party/aom/test/dct16x16_test.cc
@@ -0,0 +1,876 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "av1/common/entropy.h"
+#include "av1/common/scan.h"
+#include "aom/aom_codec.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/msvc.h"  // for round()
+
+using libaom_test::ACMRandom;
+
+namespace {
+
+const int kNumCoeffs = 256;
+const double C1 = 0.995184726672197;
+const double C2 = 0.98078528040323;
+const double C3 = 0.956940335732209;
+const double C4 = 0.923879532511287;
+const double C5 = 0.881921264348355;
+const double C6 = 0.831469612302545;
+const double C7 = 0.773010453362737;
+const double C8 = 0.707106781186548;
+const double C9 = 0.634393284163646;
+const double C10 = 0.555570233019602;
+const double C11 = 0.471396736825998;
+const double C12 = 0.38268343236509;
+const double C13 = 0.290284677254462;
+const double C14 = 0.195090322016128;
+const double C15 = 0.098017140329561;
+
+void butterfly_16x16_dct_1d(double input[16], double output[16]) {
+  double step[16];
+  double intermediate[16];
+  double temp1, temp2;
+
+  // step 1
+  step[0] = input[0] + input[15];
+  step[1] = input[1] + input[14];
+  step[2] = input[2] + input[13];
+  step[3] = input[3] + input[12];
+  step[4] = input[4] + input[11];
+  step[5] = input[5] + input[10];
+  step[6] = input[6] + input[9];
+  step[7] = input[7] + input[8];
+  step[8] = input[7] - input[8];
+  step[9] = input[6] - input[9];
+  step[10] = input[5] - input[10];
+  step[11] = input[4] - input[11];
+  step[12] = input[3] - input[12];
+  step[13] = input[2] - input[13];
+  step[14] = input[1] - input[14];
+  step[15] = input[0] - input[15];
+
+  // step 2
+  output[0] = step[0] + step[7];
+  output[1] = step[1] + step[6];
+  output[2] = step[2] + step[5];
+  output[3] = step[3] + step[4];
+  output[4] = step[3] - step[4];
+  output[5] = step[2] - step[5];
+  output[6] = step[1] - step[6];
+  output[7] = step[0] - step[7];
+
+  temp1 = step[8] * C7;
+  temp2 = step[15] * C9;
+  output[8] = temp1 + temp2;
+
+  temp1 = step[9] * C11;
+  temp2 = step[14] * C5;
+  output[9] = temp1 - temp2;
+
+  temp1 = step[10] * C3;
+  temp2 = step[13] * C13;
+  output[10] = temp1 + temp2;
+
+  temp1 = step[11] * C15;
+  temp2 = step[12] * C1;
+  output[11] = temp1 - temp2;
+
+  temp1 = step[11] * C1;
+  temp2 = step[12] * C15;
+  output[12] = temp2 + temp1;
+
+  temp1 = step[10] * C13;
+  temp2 = step[13] * C3;
+  output[13] = temp2 - temp1;
+
+  temp1 = step[9] * C5;
+  temp2 = step[14] * C11;
+  output[14] = temp2 + temp1;
+
+  temp1 = step[8] * C9;
+  temp2 = step[15] * C7;
+  output[15] = temp2 - temp1;
+
+  // step 3
+  step[0] = output[0] + output[3];
+  step[1] = output[1] + output[2];
+  step[2] = output[1] - output[2];
+  step[3] = output[0] - output[3];
+
+  temp1 = output[4] * C14;
+  temp2 = output[7] * C2;
+  step[4] = temp1 + temp2;
+
+  temp1 = output[5] * C10;
+  temp2 = output[6] * C6;
+  step[5] = temp1 + temp2;
+
+  temp1 = output[5] * C6;
+  temp2 = output[6] * C10;
+  step[6] = temp2 - temp1;
+
+  temp1 = output[4] * C2;
+  temp2 = output[7] * C14;
+  step[7] = temp2 - temp1;
+
+  step[8] = output[8] + output[11];
+  step[9] = output[9] + output[10];
+  step[10] = output[9] - output[10];
+  step[11] = output[8] - output[11];
+
+  step[12] = output[12] + output[15];
+  step[13] = output[13] + output[14];
+  step[14] = output[13] - output[14];
+  step[15] = output[12] - output[15];
+
+  // step 4
+  output[0] = (step[0] + step[1]);
+  output[8] = (step[0] - step[1]);
+
+  temp1 = step[2] * C12;
+  temp2 = step[3] * C4;
+  temp1 = temp1 + temp2;
+  output[4] = 2 * (temp1 * C8);
+
+  temp1 = step[2] * C4;
+  temp2 = step[3] * C12;
+  temp1 = temp2 - temp1;
+  output[12] = 2 * (temp1 * C8);
+
+  output[2] = 2 * ((step[4] + step[5]) * C8);
+  output[14] = 2 * ((step[7] - step[6]) * C8);
+
+  temp1 = step[4] - step[5];
+  temp2 = step[6] + step[7];
+  output[6] = (temp1 + temp2);
+  output[10] = (temp1 - temp2);
+
+  intermediate[8] = step[8] + step[14];
+  intermediate[9] = step[9] + step[15];
+
+  temp1 = intermediate[8] * C12;
+  temp2 = intermediate[9] * C4;
+  temp1 = temp1 - temp2;
+  output[3] = 2 * (temp1 * C8);
+
+  temp1 = intermediate[8] * C4;
+  temp2 = intermediate[9] * C12;
+  temp1 = temp2 + temp1;
+  output[13] = 2 * (temp1 * C8);
+
+  output[9] = 2 * ((step[10] + step[11]) * C8);
+
+  intermediate[11] = step[10] - step[11];
+  intermediate[12] = step[12] + step[13];
+  intermediate[13] = step[12] - step[13];
+  intermediate[14] = step[8] - step[14];
+  intermediate[15] = step[9] - step[15];
+
+  output[15] = (intermediate[11] + intermediate[12]);
+  output[1] = -(intermediate[11] - intermediate[12]);
+
+  output[7] = 2 * (intermediate[13] * C8);
+
+  temp1 = intermediate[14] * C12;
+  temp2 = intermediate[15] * C4;
+  temp1 = temp1 - temp2;
+  output[11] = -2 * (temp1 * C8);
+
+  temp1 = intermediate[14] * C4;
+  temp2 = intermediate[15] * C12;
+  temp1 = temp2 + temp1;
+  output[5] = 2 * (temp1 * C8);
+}
+
+void reference_16x16_dct_2d(int16_t input[256], double output[256]) {
+  // First transform columns
+  for (int i = 0; i < 16; ++i) {
+    double temp_in[16], temp_out[16];
+    for (int j = 0; j < 16; ++j) temp_in[j] = input[j * 16 + i];
+    butterfly_16x16_dct_1d(temp_in, temp_out);
+    for (int j = 0; j < 16; ++j) output[j * 16 + i] = temp_out[j];
+  }
+  // Then transform rows
+  for (int i = 0; i < 16; ++i) {
+    double temp_in[16], temp_out[16];
+    for (int j = 0; j < 16; ++j) temp_in[j] = output[j + i * 16];
+    butterfly_16x16_dct_1d(temp_in, temp_out);
+    // Scale by some magic number
+    for (int j = 0; j < 16; ++j) output[j + i * 16] = temp_out[j] / 2;
+  }
+}
+
+typedef void (*FdctFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*IdctFunc)(const tran_low_t *in, uint8_t *out, int stride);
+typedef void (*FhtFunc)(const int16_t *in, tran_low_t *out, int stride,
+                        int tx_type);
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+
+typedef std::tr1::tuple<FdctFunc, IdctFunc, int, aom_bit_depth_t> Dct16x16Param;
+typedef std::tr1::tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t> Ht16x16Param;
+typedef std::tr1::tuple<IdctFunc, IdctFunc, int, aom_bit_depth_t>
+    Idct16x16Param;
+
+void fdct16x16_ref(const int16_t *in, tran_low_t *out, int stride,
+                   int /*tx_type*/) {
+  aom_fdct16x16_c(in, out, stride);
+}
+
+void idct16x16_ref(const tran_low_t *in, uint8_t *dest, int stride,
+                   int /*tx_type*/) {
+  aom_idct16x16_256_add_c(in, dest, stride);
+}
+
+void fht16x16_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht16x16_c(in, out, stride, tx_type);
+}
+
+void iht16x16_ref(const tran_low_t *in, uint8_t *dest, int stride,
+                  int tx_type) {
+  av1_iht16x16_256_add_c(in, dest, stride, tx_type);
+}
+
+#if CONFIG_HIGHBITDEPTH
+void iht16x16_10(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_highbd_iht16x16_256_add_c(in, out, stride, tx_type, 10);
+}
+
+void iht16x16_12(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_highbd_iht16x16_256_add_c(in, out, stride, tx_type, 12);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+class Trans16x16TestBase {
+ public:
+  virtual ~Trans16x16TestBase() {}
+
+ protected:
+  virtual void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) = 0;
+
+  virtual void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) = 0;
+
+  void RunAccuracyCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    uint32_t max_error = 0;
+    int64_t total_error = 0;
+    const int count_test_block = 10000;
+    for (int i = 0; i < count_test_block; ++i) {
+      DECLARE_ALIGNED(16, int16_t, test_input_block[kNumCoeffs]);
+      DECLARE_ALIGNED(16, tran_low_t, test_temp_block[kNumCoeffs]);
+      DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+      DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_HIGHBITDEPTH
+      DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+      DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+#endif
+
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (bit_depth_ == AOM_BITS_8) {
+          src[j] = rnd.Rand8();
+          dst[j] = rnd.Rand8();
+          test_input_block[j] = src[j] - dst[j];
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand16() & mask_;
+          dst16[j] = rnd.Rand16() & mask_;
+          test_input_block[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(test_input_block, test_temp_block, pitch_));
+      if (bit_depth_ == AOM_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(test_temp_block, dst, pitch_));
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(test_temp_block, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_HIGHBITDEPTH
+        const int32_t diff =
+            bit_depth_ == AOM_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int32_t diff = dst[j] - src[j];
+#endif
+        const uint32_t error = diff * diff;
+        if (max_error < error) max_error = error;
+        total_error += error;
+      }
+    }
+
+    EXPECT_GE(1u << 2 * (bit_depth_ - 8), max_error)
+        << "Error: 16x16 FHT/IHT has an individual round trip error > 1";
+
+    EXPECT_GE(count_test_block << 2 * (bit_depth_ - 8), total_error)
+        << "Error: 16x16 FHT/IHT has average round trip error > 1 per block";
+  }
+
+  void RunCoeffCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, input_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]);
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j)
+        input_block[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+
+      fwd_txfm_ref(input_block, output_ref_block, pitch_, tx_type_);
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(input_block, output_block, pitch_));
+
+      // The minimum quant value is 4.
+      for (int j = 0; j < kNumCoeffs; ++j)
+        EXPECT_EQ(output_block[j], output_ref_block[j]);
+    }
+  }
+
+  void RunMemCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, input_extreme_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]);
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        input_extreme_block[j] = rnd.Rand8() % 2 ? mask_ : -mask_;
+      }
+      if (i == 0) {
+        for (int j = 0; j < kNumCoeffs; ++j) input_extreme_block[j] = mask_;
+      } else if (i == 1) {
+        for (int j = 0; j < kNumCoeffs; ++j) input_extreme_block[j] = -mask_;
+      }
+
+      fwd_txfm_ref(input_extreme_block, output_ref_block, pitch_, tx_type_);
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(input_extreme_block, output_block, pitch_));
+
+      // The minimum quant value is 4.
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        EXPECT_EQ(output_block[j], output_ref_block[j]);
+        EXPECT_GE(4 * DCT_MAX_VALUE << (bit_depth_ - 8), abs(output_block[j]))
+            << "Error: 16x16 FDCT has coefficient larger than 4*DCT_MAX_VALUE";
+      }
+    }
+  }
+
+  void RunQuantCheck(int dc_thred, int ac_thred) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 100000;
+    DECLARE_ALIGNED(16, int16_t, input_extreme_block[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, ref[kNumCoeffs]);
+#if CONFIG_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, ref16[kNumCoeffs]);
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        input_extreme_block[j] = rnd.Rand8() % 2 ? mask_ : -mask_;
+      }
+      if (i == 0)
+        for (int j = 0; j < kNumCoeffs; ++j) input_extreme_block[j] = mask_;
+      if (i == 1)
+        for (int j = 0; j < kNumCoeffs; ++j) input_extreme_block[j] = -mask_;
+
+      fwd_txfm_ref(input_extreme_block, output_ref_block, pitch_, tx_type_);
+
+      // clear reconstructed pixel buffers
+      memset(dst, 0, kNumCoeffs * sizeof(uint8_t));
+      memset(ref, 0, kNumCoeffs * sizeof(uint8_t));
+#if CONFIG_HIGHBITDEPTH
+      memset(dst16, 0, kNumCoeffs * sizeof(uint16_t));
+      memset(ref16, 0, kNumCoeffs * sizeof(uint16_t));
+#endif
+
+      // quantization with maximum allowed step sizes
+      output_ref_block[0] = (output_ref_block[0] / dc_thred) * dc_thred;
+      for (int j = 1; j < kNumCoeffs; ++j)
+        output_ref_block[j] = (output_ref_block[j] / ac_thred) * ac_thred;
+      if (bit_depth_ == AOM_BITS_8) {
+        inv_txfm_ref(output_ref_block, ref, pitch_, tx_type_);
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(output_ref_block, dst, pitch_));
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        inv_txfm_ref(output_ref_block, CONVERT_TO_BYTEPTR(ref16), pitch_,
+                     tx_type_);
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(output_ref_block, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+      if (bit_depth_ == AOM_BITS_8) {
+        for (int j = 0; j < kNumCoeffs; ++j) EXPECT_EQ(ref[j], dst[j]);
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        for (int j = 0; j < kNumCoeffs; ++j) EXPECT_EQ(ref16[j], dst16[j]);
+#endif
+      }
+    }
+  }
+
+  void RunInvAccuracyCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, in[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+#endif  // CONFIG_HIGHBITDEPTH
+
+    for (int i = 0; i < count_test_block; ++i) {
+      double out_r[kNumCoeffs];
+
+      // Initialize a test block with input range [-255, 255].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (bit_depth_ == AOM_BITS_8) {
+          src[j] = rnd.Rand8();
+          dst[j] = rnd.Rand8();
+          in[j] = src[j] - dst[j];
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand16() & mask_;
+          dst16[j] = rnd.Rand16() & mask_;
+          in[j] = src16[j] - dst16[j];
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+
+      reference_16x16_dct_2d(in, out_r);
+      for (int j = 0; j < kNumCoeffs; ++j)
+        coeff[j] = static_cast<tran_low_t>(round(out_r[j]));
+
+      if (bit_depth_ == AOM_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, 16));
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(coeff, CONVERT_TO_BYTEPTR(dst16), 16));
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == AOM_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int diff = dst[j] - src[j];
+#endif  // CONFIG_HIGHBITDEPTH
+        const uint32_t error = diff * diff;
+        EXPECT_GE(1u, error) << "Error: 16x16 IDCT has error " << error
+                             << " at index " << j;
+      }
+    }
+  }
+
+  void CompareInvReference(IdctFunc ref_txfm, int thresh) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 10000;
+    const int eob = 10;
+    const int16_t *scan = av1_default_scan_orders[TX_16X16].scan;
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, ref[kNumCoeffs]);
+#if CONFIG_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, ref16[kNumCoeffs]);
+#endif  // CONFIG_HIGHBITDEPTH
+
+    for (int i = 0; i < count_test_block; ++i) {
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (j < eob) {
+          // Random values less than the threshold, either positive or negative
+          coeff[scan[j]] = rnd(thresh) * (1 - 2 * (i % 2));
+        } else {
+          coeff[scan[j]] = 0;
+        }
+        if (bit_depth_ == AOM_BITS_8) {
+          dst[j] = 0;
+          ref[j] = 0;
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          dst16[j] = 0;
+          ref16[j] = 0;
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+      if (bit_depth_ == AOM_BITS_8) {
+        ref_txfm(coeff, ref, pitch_);
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, pitch_));
+      } else {
+#if CONFIG_HIGHBITDEPTH
+        ref_txfm(coeff, CONVERT_TO_BYTEPTR(ref16), pitch_);
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(coeff, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == AOM_BITS_8 ? dst[j] - ref[j] : dst16[j] - ref16[j];
+#else
+        const int diff = dst[j] - ref[j];
+#endif  // CONFIG_HIGHBITDEPTH
+        const uint32_t error = diff * diff;
+        EXPECT_EQ(0u, error) << "Error: 16x16 IDCT Comparison has error "
+                             << error << " at index " << j;
+      }
+    }
+  }
+
+  int pitch_;
+  int tx_type_;
+  aom_bit_depth_t bit_depth_;
+  int mask_;
+  FhtFunc fwd_txfm_ref;
+  IhtFunc inv_txfm_ref;
+};
+
+class Trans16x16DCT : public Trans16x16TestBase,
+                      public ::testing::TestWithParam<Dct16x16Param> {
+ public:
+  virtual ~Trans16x16DCT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    bit_depth_ = GET_PARAM(3);
+    pitch_ = 16;
+    fwd_txfm_ref = fdct16x16_ref;
+    inv_txfm_ref = idct16x16_ref;
+    mask_ = (1 << bit_depth_) - 1;
+    inv_txfm_ref = idct16x16_ref;
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride);
+  }
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
+};
+
+TEST_P(Trans16x16DCT, AccuracyCheck) { RunAccuracyCheck(); }
+
+TEST_P(Trans16x16DCT, CoeffCheck) { RunCoeffCheck(); }
+
+TEST_P(Trans16x16DCT, MemCheck) { RunMemCheck(); }
+
+TEST_P(Trans16x16DCT, QuantCheck) {
+  // Use maximally allowed quantization step sizes for DC and AC
+  // coefficients respectively.
+  RunQuantCheck(1336, 1828);
+}
+
+TEST_P(Trans16x16DCT, InvAccuracyCheck) { RunInvAccuracyCheck(); }
+
+class Trans16x16HT : public Trans16x16TestBase,
+                     public ::testing::TestWithParam<Ht16x16Param> {
+ public:
+  virtual ~Trans16x16HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    bit_depth_ = GET_PARAM(3);
+    pitch_ = 16;
+    fwd_txfm_ref = fht16x16_ref;
+    inv_txfm_ref = iht16x16_ref;
+    mask_ = (1 << bit_depth_) - 1;
+#if CONFIG_HIGHBITDEPTH
+    switch (bit_depth_) {
+      case AOM_BITS_10: inv_txfm_ref = iht16x16_10; break;
+      case AOM_BITS_12: inv_txfm_ref = iht16x16_12; break;
+      default: inv_txfm_ref = iht16x16_ref; break;
+    }
+#else
+    inv_txfm_ref = iht16x16_ref;
+#endif
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(Trans16x16HT, AccuracyCheck) { RunAccuracyCheck(); }
+
+TEST_P(Trans16x16HT, CoeffCheck) { RunCoeffCheck(); }
+
+TEST_P(Trans16x16HT, MemCheck) { RunMemCheck(); }
+
+TEST_P(Trans16x16HT, QuantCheck) {
+  // The encoder skips any non-DC intra prediction modes,
+  // when the quantization step size goes beyond 988.
+  RunQuantCheck(429, 729);
+}
+
+class InvTrans16x16DCT : public Trans16x16TestBase,
+                         public ::testing::TestWithParam<Idct16x16Param> {
+ public:
+  virtual ~InvTrans16x16DCT() {}
+
+  virtual void SetUp() {
+    ref_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    thresh_ = GET_PARAM(2);
+    bit_depth_ = GET_PARAM(3);
+    pitch_ = 16;
+    mask_ = (1 << bit_depth_) - 1;
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(int16_t * /*in*/, tran_low_t * /*out*/, int /*stride*/) {}
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  IdctFunc ref_txfm_;
+  IdctFunc inv_txfm_;
+  int thresh_;
+};
+
+TEST_P(InvTrans16x16DCT, CompareReference) {
+  CompareInvReference(ref_txfm_, thresh_);
+}
+
+class PartialTrans16x16Test : public ::testing::TestWithParam<
+                                  std::tr1::tuple<FdctFunc, aom_bit_depth_t> > {
+ public:
+  virtual ~PartialTrans16x16Test() {}
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    bit_depth_ = GET_PARAM(1);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  aom_bit_depth_t bit_depth_;
+  FdctFunc fwd_txfm_;
+};
+
+TEST_P(PartialTrans16x16Test, Extremes) {
+#if CONFIG_HIGHBITDEPTH
+  const int16_t maxval =
+      static_cast<int16_t>(clip_pixel_highbd(1 << 30, bit_depth_));
+#else
+  const int16_t maxval = 255;
+#endif
+  const int minval = -maxval;
+  DECLARE_ALIGNED(16, int16_t, input[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output[kNumCoeffs]);
+
+  for (int i = 0; i < kNumCoeffs; ++i) input[i] = maxval;
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 16));
+  EXPECT_EQ((maxval * kNumCoeffs) >> 1, output[0]);
+
+  for (int i = 0; i < kNumCoeffs; ++i) input[i] = minval;
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 16));
+  EXPECT_EQ((minval * kNumCoeffs) >> 1, output[0]);
+}
+
+TEST_P(PartialTrans16x16Test, Random) {
+#if CONFIG_HIGHBITDEPTH
+  const int16_t maxval =
+      static_cast<int16_t>(clip_pixel_highbd(1 << 30, bit_depth_));
+#else
+  const int16_t maxval = 255;
+#endif
+  DECLARE_ALIGNED(16, int16_t, input[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output[kNumCoeffs]);
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+  int sum = 0;
+  for (int i = 0; i < kNumCoeffs; ++i) {
+    const int val = (i & 1) ? -rnd(maxval + 1) : rnd(maxval + 1);
+    input[i] = val;
+    sum += val;
+  }
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 16));
+  EXPECT_EQ(sum >> 1, output[0]);
+}
+
+using std::tr1::make_tuple;
+
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(C, Trans16x16DCT,
+                        ::testing::Values(make_tuple(&aom_fdct16x16_c,
+                                                     &aom_idct16x16_256_add_c,
+                                                     0, AOM_BITS_8)));
+#else
+INSTANTIATE_TEST_CASE_P(C, Trans16x16DCT,
+                        ::testing::Values(make_tuple(&aom_fdct16x16_c,
+                                                     &aom_idct16x16_256_add_c,
+                                                     0, AOM_BITS_8)));
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&av1_highbd_fht16x16_c, &iht16x16_10, 0, AOM_BITS_10),
+        make_tuple(&av1_highbd_fht16x16_c, &iht16x16_10, 1, AOM_BITS_10),
+        make_tuple(&av1_highbd_fht16x16_c, &iht16x16_10, 2, AOM_BITS_10),
+        make_tuple(&av1_highbd_fht16x16_c, &iht16x16_10, 3, AOM_BITS_10),
+        make_tuple(&av1_highbd_fht16x16_c, &iht16x16_12, 0, AOM_BITS_12),
+        make_tuple(&av1_highbd_fht16x16_c, &iht16x16_12, 1, AOM_BITS_12),
+        make_tuple(&av1_highbd_fht16x16_c, &iht16x16_12, 2, AOM_BITS_12),
+        make_tuple(&av1_highbd_fht16x16_c, &iht16x16_12, 3, AOM_BITS_12),
+        make_tuple(&av1_fht16x16_c, &av1_iht16x16_256_add_c, 0, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_c, &av1_iht16x16_256_add_c, 1, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_c, &av1_iht16x16_256_add_c, 2, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_c, &av1_iht16x16_256_add_c, 3, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    C, PartialTrans16x16Test,
+    ::testing::Values(make_tuple(&aom_highbd_fdct16x16_1_c, AOM_BITS_8),
+                      make_tuple(&aom_highbd_fdct16x16_1_c, AOM_BITS_10),
+                      make_tuple(&aom_highbd_fdct16x16_1_c, AOM_BITS_12)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&av1_fht16x16_c, &av1_iht16x16_256_add_c, 0, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_c, &av1_iht16x16_256_add_c, 1, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_c, &av1_iht16x16_256_add_c, 2, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_c, &av1_iht16x16_256_add_c, 3, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(C, PartialTrans16x16Test,
+                        ::testing::Values(make_tuple(&aom_fdct16x16_1_c,
+                                                     AOM_BITS_8)));
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if HAVE_NEON_ASM && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    NEON, Trans16x16DCT,
+    ::testing::Values(make_tuple(&aom_fdct16x16_c, &aom_idct16x16_256_add_neon,
+                                 0, AOM_BITS_8)));
+#endif
+
+#if HAVE_SSE2 && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans16x16DCT,
+    ::testing::Values(make_tuple(&aom_fdct16x16_sse2,
+                                 &aom_idct16x16_256_add_sse2, 0, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans16x16HT,
+    ::testing::Values(make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2,
+                                 0, AOM_BITS_8),
+                      make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2,
+                                 1, AOM_BITS_8),
+                      make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2,
+                                 2, AOM_BITS_8),
+                      make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_sse2,
+                                 3, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(SSE2, PartialTrans16x16Test,
+                        ::testing::Values(make_tuple(&aom_fdct16x16_1_sse2,
+                                                     AOM_BITS_8)));
+#endif  // HAVE_SSE2 && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_AVX2 && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(AVX2, PartialTrans16x16Test,
+                        ::testing::Values(make_tuple(&aom_fdct16x16_1_avx2,
+                                                     AOM_BITS_8)));
+#endif  // HAVE_AVX2 && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_SSE2 && CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(SSE2, Trans16x16DCT,
+                        ::testing::Values(make_tuple(&aom_fdct16x16_sse2,
+                                                     &aom_idct16x16_256_add_c,
+                                                     0, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_c, 0, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_c, 1, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_c, 2, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_sse2, &av1_iht16x16_256_add_c, 3,
+                   AOM_BITS_8)));
+// TODO(luoyi):
+// For this test case, we should test function: aom_highbd_fdct16x16_1_sse2.
+// However this function is not available yet. if we mistakely test
+// aom_fdct16x16_1_sse2, it could only pass AOM_BITS_8/AOM_BITS_10 but not
+// AOM_BITS_12.
+INSTANTIATE_TEST_CASE_P(SSE2, PartialTrans16x16Test,
+                        ::testing::Values(make_tuple(&aom_fdct16x16_1_sse2,
+                                                     AOM_BITS_8)));
+#endif  // HAVE_SSE2 && CONFIG_HIGHBITDEPTH
+
+#if HAVE_MSA && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(MSA, Trans16x16DCT,
+                        ::testing::Values(make_tuple(&aom_fdct16x16_msa,
+                                                     &aom_idct16x16_256_add_msa,
+                                                     0, AOM_BITS_8)));
+#if !CONFIG_EXT_TX
+// TODO(yaowu): re-enable this after msa versions are updated to match C.
+INSTANTIATE_TEST_CASE_P(
+    DISABLED_MSA, Trans16x16HT,
+    ::testing::Values(
+        make_tuple(&av1_fht16x16_msa, &av1_iht16x16_256_add_msa, 0, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_msa, &av1_iht16x16_256_add_msa, 1, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_msa, &av1_iht16x16_256_add_msa, 2, AOM_BITS_8),
+        make_tuple(&av1_fht16x16_msa, &av1_iht16x16_256_add_msa, 3,
+                   AOM_BITS_8)));
+#endif  // !CONFIG_EXT_TX
+INSTANTIATE_TEST_CASE_P(MSA, PartialTrans16x16Test,
+                        ::testing::Values(make_tuple(&aom_fdct16x16_1_msa,
+                                                     AOM_BITS_8)));
+#endif  // HAVE_MSA && !CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/dct32x32_test.cc b/third_party/aom/test/dct32x32_test.cc
new file mode 100644
index 0000000000..7c1db6501b
--- /dev/null
+++ b/third_party/aom/test/dct32x32_test.cc
@@ -0,0 +1,438 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "av1/common/entropy.h"
+#include "aom/aom_codec.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/msvc.h"  // for round()
+
+using libaom_test::ACMRandom;
+
+namespace {
+
+const int kNumCoeffs = 1024;
+const double kPi = 3.141592653589793238462643383279502884;
+void reference_32x32_dct_1d(const double in[32], double out[32]) {
+  const double kInvSqrt2 = 0.707106781186547524400844362104;
+  for (int k = 0; k < 32; k++) {
+    out[k] = 0.0;
+    for (int n = 0; n < 32; n++)
+      out[k] += in[n] * cos(kPi * (2 * n + 1) * k / 64.0);
+    if (k == 0) out[k] = out[k] * kInvSqrt2;
+  }
+}
+
+void reference_32x32_dct_2d(const int16_t input[kNumCoeffs],
+                            double output[kNumCoeffs]) {
+  // First transform columns
+  for (int i = 0; i < 32; ++i) {
+    double temp_in[32], temp_out[32];
+    for (int j = 0; j < 32; ++j) temp_in[j] = input[j * 32 + i];
+    reference_32x32_dct_1d(temp_in, temp_out);
+    for (int j = 0; j < 32; ++j) output[j * 32 + i] = temp_out[j];
+  }
+  // Then transform rows
+  for (int i = 0; i < 32; ++i) {
+    double temp_in[32], temp_out[32];
+    for (int j = 0; j < 32; ++j) temp_in[j] = output[j + i * 32];
+    reference_32x32_dct_1d(temp_in, temp_out);
+    // Scale by some magic number
+    for (int j = 0; j < 32; ++j) output[j + i * 32] = temp_out[j] / 4;
+  }
+}
+
+typedef void (*FwdTxfmFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*InvTxfmFunc)(const tran_low_t *in, uint8_t *out, int stride);
+
+typedef std::tr1::tuple<FwdTxfmFunc, InvTxfmFunc, int, aom_bit_depth_t>
+    Trans32x32Param;
+
+class Trans32x32Test : public ::testing::TestWithParam<Trans32x32Param> {
+ public:
+  virtual ~Trans32x32Test() {}
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    version_ = GET_PARAM(2);  // 0: high precision forward transform
+                              // 1: low precision version for rd loop
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  int version_;
+  aom_bit_depth_t bit_depth_;
+  int mask_;
+  FwdTxfmFunc fwd_txfm_;
+  InvTxfmFunc inv_txfm_;
+};
+
+TEST_P(Trans32x32Test, AccuracyCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  uint32_t max_error = 0;
+  int64_t total_error = 0;
+  const int count_test_block = 10000;
+  DECLARE_ALIGNED(16, int16_t, test_input_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, test_temp_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+#endif
+
+  for (int i = 0; i < count_test_block; ++i) {
+    // Initialize a test block with input range [-mask_, mask_].
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      if (bit_depth_ == AOM_BITS_8) {
+        src[j] = rnd.Rand8();
+        dst[j] = rnd.Rand8();
+        test_input_block[j] = src[j] - dst[j];
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        src16[j] = rnd.Rand16() & mask_;
+        dst16[j] = rnd.Rand16() & mask_;
+        test_input_block[j] = src16[j] - dst16[j];
+#endif
+      }
+    }
+
+    ASM_REGISTER_STATE_CHECK(fwd_txfm_(test_input_block, test_temp_block, 32));
+    if (bit_depth_ == AOM_BITS_8) {
+      ASM_REGISTER_STATE_CHECK(inv_txfm_(test_temp_block, dst, 32));
+#if CONFIG_HIGHBITDEPTH
+    } else {
+      ASM_REGISTER_STATE_CHECK(
+          inv_txfm_(test_temp_block, CONVERT_TO_BYTEPTR(dst16), 32));
+#endif
+    }
+
+    for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_HIGHBITDEPTH
+      const int32_t diff =
+          bit_depth_ == AOM_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+      const int32_t diff = dst[j] - src[j];
+#endif
+      const uint32_t error = diff * diff;
+      if (max_error < error) max_error = error;
+      total_error += error;
+    }
+  }
+
+  if (version_ == 1) {
+    max_error /= 2;
+    total_error /= 45;
+  }
+
+  EXPECT_GE(1u << 2 * (bit_depth_ - 8), max_error)
+      << "Error: 32x32 FDCT/IDCT has an individual round-trip error > 1";
+
+  EXPECT_GE(count_test_block << 2 * (bit_depth_ - 8), total_error)
+      << "Error: 32x32 FDCT/IDCT has average round-trip error > 1 per block";
+}
+
+TEST_P(Trans32x32Test, CoeffCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 1000;
+
+  DECLARE_ALIGNED(16, int16_t, input_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]);
+
+  for (int i = 0; i < count_test_block; ++i) {
+    for (int j = 0; j < kNumCoeffs; ++j)
+      input_block[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+
+    const int stride = 32;
+    aom_fdct32x32_c(input_block, output_ref_block, stride);
+    ASM_REGISTER_STATE_CHECK(fwd_txfm_(input_block, output_block, stride));
+
+    if (version_ == 0) {
+      for (int j = 0; j < kNumCoeffs; ++j)
+        EXPECT_EQ(output_block[j], output_ref_block[j])
+            << "Error: 32x32 FDCT versions have mismatched coefficients";
+    } else {
+      for (int j = 0; j < kNumCoeffs; ++j)
+        EXPECT_GE(6, abs(output_block[j] - output_ref_block[j]))
+            << "Error: 32x32 FDCT rd has mismatched coefficients";
+    }
+  }
+}
+
+TEST_P(Trans32x32Test, MemCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 2000;
+
+  DECLARE_ALIGNED(16, int16_t, input_extreme_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_block[kNumCoeffs]);
+
+  for (int i = 0; i < count_test_block; ++i) {
+    // Initialize a test block with input range [-mask_, mask_].
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      input_extreme_block[j] = rnd.Rand8() & 1 ? mask_ : -mask_;
+    }
+    if (i == 0) {
+      for (int j = 0; j < kNumCoeffs; ++j) input_extreme_block[j] = mask_;
+    } else if (i == 1) {
+      for (int j = 0; j < kNumCoeffs; ++j) input_extreme_block[j] = -mask_;
+    }
+
+    const int stride = 32;
+    aom_fdct32x32_c(input_extreme_block, output_ref_block, stride);
+    ASM_REGISTER_STATE_CHECK(
+        fwd_txfm_(input_extreme_block, output_block, stride));
+
+    // The minimum quant value is 4.
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      if (version_ == 0) {
+        EXPECT_EQ(output_block[j], output_ref_block[j])
+            << "Error: 32x32 FDCT versions have mismatched coefficients";
+      } else {
+        EXPECT_GE(6, abs(output_block[j] - output_ref_block[j]))
+            << "Error: 32x32 FDCT rd has mismatched coefficients";
+      }
+      EXPECT_GE(4 * DCT_MAX_VALUE << (bit_depth_ - 8), abs(output_ref_block[j]))
+          << "Error: 32x32 FDCT C has coefficient larger than 4*DCT_MAX_VALUE";
+      EXPECT_GE(4 * DCT_MAX_VALUE << (bit_depth_ - 8), abs(output_block[j]))
+          << "Error: 32x32 FDCT has coefficient larger than "
+          << "4*DCT_MAX_VALUE";
+    }
+  }
+}
+
+TEST_P(Trans32x32Test, InverseAccuracy) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 1000;
+  DECLARE_ALIGNED(16, int16_t, in[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+#endif
+
+  for (int i = 0; i < count_test_block; ++i) {
+    double out_r[kNumCoeffs];
+
+    // Initialize a test block with input range [-255, 255]
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      if (bit_depth_ == AOM_BITS_8) {
+        src[j] = rnd.Rand8();
+        dst[j] = rnd.Rand8();
+        in[j] = src[j] - dst[j];
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        src16[j] = rnd.Rand16() & mask_;
+        dst16[j] = rnd.Rand16() & mask_;
+        in[j] = src16[j] - dst16[j];
+#endif
+      }
+    }
+
+    reference_32x32_dct_2d(in, out_r);
+    for (int j = 0; j < kNumCoeffs; ++j)
+      coeff[j] = static_cast<tran_low_t>(round(out_r[j]));
+    if (bit_depth_ == AOM_BITS_8) {
+      ASM_REGISTER_STATE_CHECK(inv_txfm_(coeff, dst, 32));
+#if CONFIG_HIGHBITDEPTH
+    } else {
+      ASM_REGISTER_STATE_CHECK(inv_txfm_(coeff, CONVERT_TO_BYTEPTR(dst16), 32));
+#endif
+    }
+    for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_HIGHBITDEPTH
+      const int diff =
+          bit_depth_ == AOM_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+      const int diff = dst[j] - src[j];
+#endif
+      const int error = diff * diff;
+      EXPECT_GE(1, error) << "Error: 32x32 IDCT has error " << error
+                          << " at index " << j;
+    }
+  }
+}
+
+class PartialTrans32x32Test
+    : public ::testing::TestWithParam<
+          std::tr1::tuple<FwdTxfmFunc, aom_bit_depth_t> > {
+ public:
+  virtual ~PartialTrans32x32Test() {}
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    bit_depth_ = GET_PARAM(1);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  aom_bit_depth_t bit_depth_;
+  FwdTxfmFunc fwd_txfm_;
+};
+
+TEST_P(PartialTrans32x32Test, Extremes) {
+#if CONFIG_HIGHBITDEPTH
+  const int16_t maxval =
+      static_cast<int16_t>(clip_pixel_highbd(1 << 30, bit_depth_));
+#else
+  const int16_t maxval = 255;
+#endif
+  const int minval = -maxval;
+  DECLARE_ALIGNED(16, int16_t, input[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output[kNumCoeffs]);
+
+  for (int i = 0; i < kNumCoeffs; ++i) input[i] = maxval;
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 32));
+  EXPECT_EQ((maxval * kNumCoeffs) >> 3, output[0]);
+
+  for (int i = 0; i < kNumCoeffs; ++i) input[i] = minval;
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 32));
+  EXPECT_EQ((minval * kNumCoeffs) >> 3, output[0]);
+}
+
+TEST_P(PartialTrans32x32Test, Random) {
+#if CONFIG_HIGHBITDEPTH
+  const int16_t maxval =
+      static_cast<int16_t>(clip_pixel_highbd(1 << 30, bit_depth_));
+#else
+  const int16_t maxval = 255;
+#endif
+  DECLARE_ALIGNED(16, int16_t, input[kNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output[kNumCoeffs]);
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+  int sum = 0;
+  for (int i = 0; i < kNumCoeffs; ++i) {
+    const int val = (i & 1) ? -rnd(maxval + 1) : rnd(maxval + 1);
+    input[i] = val;
+    sum += val;
+  }
+  output[0] = 0;
+  ASM_REGISTER_STATE_CHECK(fwd_txfm_(input, output, 32));
+  EXPECT_EQ(sum >> 3, output[0]);
+}
+
+using std::tr1::make_tuple;
+
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans32x32Test,
+    ::testing::Values(make_tuple(&aom_fdct32x32_c, &aom_idct32x32_1024_add_c, 0,
+                                 AOM_BITS_8),
+                      make_tuple(&aom_fdct32x32_rd_c, &aom_idct32x32_1024_add_c,
+                                 1, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    C, PartialTrans32x32Test,
+    ::testing::Values(make_tuple(&aom_highbd_fdct32x32_1_c, AOM_BITS_8),
+                      make_tuple(&aom_highbd_fdct32x32_1_c, AOM_BITS_10),
+                      make_tuple(&aom_highbd_fdct32x32_1_c, AOM_BITS_12)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, Trans32x32Test,
+    ::testing::Values(make_tuple(&aom_fdct32x32_c, &aom_idct32x32_1024_add_c, 0,
+                                 AOM_BITS_8),
+                      make_tuple(&aom_fdct32x32_rd_c, &aom_idct32x32_1024_add_c,
+                                 1, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(C, PartialTrans32x32Test,
+                        ::testing::Values(make_tuple(&aom_fdct32x32_1_c,
+                                                     AOM_BITS_8)));
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if HAVE_NEON && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    NEON, Trans32x32Test,
+    ::testing::Values(make_tuple(&aom_fdct32x32_c, &aom_idct32x32_1024_add_neon,
+                                 0, AOM_BITS_8),
+                      make_tuple(&aom_fdct32x32_rd_c,
+                                 &aom_idct32x32_1024_add_neon, 1, AOM_BITS_8)));
+#endif  // HAVE_NEON && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_SSE2 && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans32x32Test,
+    ::testing::Values(make_tuple(&aom_fdct32x32_sse2,
+                                 &aom_idct32x32_1024_add_sse2, 0, AOM_BITS_8),
+                      make_tuple(&aom_fdct32x32_rd_sse2,
+                                 &aom_idct32x32_1024_add_sse2, 1, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(SSE2, PartialTrans32x32Test,
+                        ::testing::Values(make_tuple(&aom_fdct32x32_1_sse2,
+                                                     AOM_BITS_8)));
+#endif  // HAVE_SSE2 && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_AVX2 && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(AVX2, PartialTrans32x32Test,
+                        ::testing::Values(make_tuple(&aom_fdct32x32_1_avx2,
+                                                     AOM_BITS_8)));
+#endif  // HAVE_AVX2 && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_SSE2 && CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans32x32Test,
+    ::testing::Values(make_tuple(&aom_fdct32x32_sse2, &aom_idct32x32_1024_add_c,
+                                 0, AOM_BITS_8),
+                      make_tuple(&aom_fdct32x32_rd_sse2,
+                                 &aom_idct32x32_1024_add_c, 1, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(SSE2, PartialTrans32x32Test,
+                        ::testing::Values(make_tuple(&aom_fdct32x32_1_sse2,
+                                                     AOM_BITS_8)));
+#endif  // HAVE_SSE2 && CONFIG_HIGHBITDEPTH
+
+#if HAVE_AVX2 && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    AVX2, Trans32x32Test,
+    ::testing::Values(make_tuple(&aom_fdct32x32_avx2,
+                                 &aom_idct32x32_1024_add_sse2, 0, AOM_BITS_8),
+                      make_tuple(&aom_fdct32x32_rd_avx2,
+                                 &aom_idct32x32_1024_add_sse2, 1, AOM_BITS_8)));
+#endif  // HAVE_AVX2 && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_AVX2 && CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    AVX2, Trans32x32Test,
+    ::testing::Values(make_tuple(&aom_fdct32x32_avx2,
+                                 &aom_idct32x32_1024_add_sse2, 0, AOM_BITS_8),
+                      make_tuple(&aom_fdct32x32_rd_avx2,
+                                 &aom_idct32x32_1024_add_sse2, 1, AOM_BITS_8)));
+#endif  // HAVE_AVX2 && CONFIG_HIGHBITDEPTH
+
+#if HAVE_MSA && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    MSA, Trans32x32Test,
+    ::testing::Values(make_tuple(&aom_fdct32x32_msa,
+                                 &aom_idct32x32_1024_add_msa, 0, AOM_BITS_8),
+                      make_tuple(&aom_fdct32x32_rd_msa,
+                                 &aom_idct32x32_1024_add_msa, 1, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(MSA, PartialTrans32x32Test,
+                        ::testing::Values(make_tuple(&aom_fdct32x32_1_msa,
+                                                     AOM_BITS_8)));
+#endif  // HAVE_MSA && !CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/decode_api_test.cc b/third_party/aom/test/decode_api_test.cc
new file mode 100644
index 0000000000..6bd72a45d4
--- /dev/null
+++ b/third_party/aom/test/decode_api_test.cc
@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "test/ivf_video_source.h"
+#include "aom/aomdx.h"
+#include "aom/aom_decoder.h"
+
+namespace {
+
+#define NELEMENTS(x) static_cast<int>(sizeof(x) / sizeof(x[0]))
+
+TEST(DecodeAPI, InvalidParams) {
+  static const aom_codec_iface_t *kCodecs[] = {
+#if CONFIG_AV1_DECODER
+    &aom_codec_av1_dx_algo,
+#endif
+  };
+  uint8_t buf[1] = { 0 };
+  aom_codec_ctx_t dec;
+
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_dec_init(NULL, NULL, NULL, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_dec_init(&dec, NULL, NULL, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_decode(NULL, NULL, 0, NULL, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_decode(NULL, buf, 0, NULL, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM,
+            aom_codec_decode(NULL, buf, NELEMENTS(buf), NULL, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM,
+            aom_codec_decode(NULL, NULL, NELEMENTS(buf), NULL, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_destroy(NULL));
+  EXPECT_TRUE(aom_codec_error(NULL) != NULL);
+
+  for (int i = 0; i < NELEMENTS(kCodecs); ++i) {
+    EXPECT_EQ(AOM_CODEC_INVALID_PARAM,
+              aom_codec_dec_init(NULL, kCodecs[i], NULL, 0));
+
+    EXPECT_EQ(AOM_CODEC_OK, aom_codec_dec_init(&dec, kCodecs[i], NULL, 0));
+    EXPECT_EQ(AOM_CODEC_UNSUP_BITSTREAM,
+              aom_codec_decode(&dec, buf, NELEMENTS(buf), NULL, 0));
+    EXPECT_EQ(AOM_CODEC_INVALID_PARAM,
+              aom_codec_decode(&dec, NULL, NELEMENTS(buf), NULL, 0));
+    EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_decode(&dec, buf, 0, NULL, 0));
+
+    EXPECT_EQ(AOM_CODEC_OK, aom_codec_destroy(&dec));
+  }
+}
+
+}  // namespace
diff --git a/third_party/aom/test/decode_perf_test.cc b/third_party/aom/test/decode_perf_test.cc
new file mode 100644
index 0000000000..ede4f8849d
--- /dev/null
+++ b/third_party/aom/test/decode_perf_test.cc
@@ -0,0 +1,243 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <string>
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/ivf_video_source.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#include "test/webm_video_source.h"
+#include "aom_ports/aom_timer.h"
+#include "./ivfenc.h"
+#include "./aom_version.h"
+
+using std::tr1::make_tuple;
+
+namespace {
+
+#define VIDEO_NAME 0
+#define THREADS 1
+
+const int kMaxPsnr = 100;
+const double kUsecsInSec = 1000000.0;
+const char kNewEncodeOutputFile[] = "new_encode.ivf";
+
+/*
+ DecodePerfTest takes a tuple of filename + number of threads to decode with
+ */
+typedef std::tr1::tuple<const char *, unsigned> DecodePerfParam;
+
+// TODO(jimbankoski): Add actual test vectors here when available.
+// const DecodePerfParam kAV1DecodePerfVectors[] = {};
+
+/*
+ In order to reflect real world performance as much as possible, Perf tests
+ *DO NOT* do any correctness checks. Please run them alongside correctness
+ tests to ensure proper codec integrity. Furthermore, in this test we
+ deliberately limit the amount of system calls we make to avoid OS
+ preemption.
+
+ TODO(joshualitt) create a more detailed perf measurement test to collect
+   power/temp/min max frame decode times/etc
+ */
+
+class DecodePerfTest : public ::testing::TestWithParam<DecodePerfParam> {};
+
+TEST_P(DecodePerfTest, PerfTest) {
+  const char *const video_name = GET_PARAM(VIDEO_NAME);
+  const unsigned threads = GET_PARAM(THREADS);
+
+  libaom_test::WebMVideoSource video(video_name);
+  video.Init();
+
+  aom_codec_dec_cfg_t cfg = aom_codec_dec_cfg_t();
+  cfg.threads = threads;
+  libaom_test::AV1Decoder decoder(cfg, 0);
+
+  aom_usec_timer t;
+  aom_usec_timer_start(&t);
+
+  for (video.Begin(); video.cxdata() != NULL; video.Next()) {
+    decoder.DecodeFrame(video.cxdata(), video.frame_size());
+  }
+
+  aom_usec_timer_mark(&t);
+  const double elapsed_secs = double(aom_usec_timer_elapsed(&t)) / kUsecsInSec;
+  const unsigned frames = video.frame_number();
+  const double fps = double(frames) / elapsed_secs;
+
+  printf("{\n");
+  printf("\t\"type\" : \"decode_perf_test\",\n");
+  printf("\t\"version\" : \"%s\",\n", VERSION_STRING_NOSP);
+  printf("\t\"videoName\" : \"%s\",\n", video_name);
+  printf("\t\"threadCount\" : %u,\n", threads);
+  printf("\t\"decodeTimeSecs\" : %f,\n", elapsed_secs);
+  printf("\t\"totalFrames\" : %u,\n", frames);
+  printf("\t\"framesPerSecond\" : %f\n", fps);
+  printf("}\n");
+}
+
+// TODO(jimbankoski): Enabled when we have actual AV1 Decode vectors.
+// INSTANTIATE_TEST_CASE_P(AV1, DecodePerfTest,
+//                        ::testing::ValuesIn(kAV1DecodePerfVectors));
+
+class AV1NewEncodeDecodePerfTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWithParam<libaom_test::TestMode> {
+ protected:
+  AV1NewEncodeDecodePerfTest()
+      : EncoderTest(GET_PARAM(0)), encoding_mode_(GET_PARAM(1)), speed_(0),
+        outfile_(0), out_frames_(0) {}
+
+  virtual ~AV1NewEncodeDecodePerfTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+
+    cfg_.g_lag_in_frames = 25;
+    cfg_.rc_min_quantizer = 2;
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_dropframe_thresh = 0;
+    cfg_.rc_undershoot_pct = 50;
+    cfg_.rc_overshoot_pct = 50;
+    cfg_.rc_buf_sz = 1000;
+    cfg_.rc_buf_initial_sz = 500;
+    cfg_.rc_buf_optimal_sz = 600;
+    cfg_.rc_resize_allowed = 0;
+    cfg_.rc_end_usage = AOM_VBR;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AOME_SET_CPUUSED, speed_);
+      encoder->Control(AV1E_SET_FRAME_PARALLEL_DECODING, 1);
+      encoder->Control(AV1E_SET_TILE_COLUMNS, 2);
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    const std::string data_path = getenv("LIBAOM_TEST_DATA_PATH");
+    const std::string path_to_source = data_path + "/" + kNewEncodeOutputFile;
+    outfile_ = fopen(path_to_source.c_str(), "wb");
+    ASSERT_TRUE(outfile_ != NULL);
+  }
+
+  virtual void EndPassHook() {
+    if (outfile_ != NULL) {
+      if (!fseek(outfile_, 0, SEEK_SET))
+        ivf_write_file_header(outfile_, &cfg_, AV1_FOURCC, out_frames_);
+      fclose(outfile_);
+      outfile_ = NULL;
+    }
+  }
+
+  virtual void FramePktHook(const aom_codec_cx_pkt_t *pkt) {
+    ++out_frames_;
+
+    // Write initial file header if first frame.
+    if (pkt->data.frame.pts == 0)
+      ivf_write_file_header(outfile_, &cfg_, AV1_FOURCC, out_frames_);
+
+    // Write frame header and data.
+    ivf_write_frame_header(outfile_, out_frames_, pkt->data.frame.sz);
+    ASSERT_EQ(fwrite(pkt->data.frame.buf, 1, pkt->data.frame.sz, outfile_),
+              pkt->data.frame.sz);
+  }
+
+  virtual bool DoDecode() { return false; }
+
+  void set_speed(unsigned int speed) { speed_ = speed; }
+
+ private:
+  libaom_test::TestMode encoding_mode_;
+  uint32_t speed_;
+  FILE *outfile_;
+  uint32_t out_frames_;
+};
+
+struct EncodePerfTestVideo {
+  EncodePerfTestVideo(const char *name_, uint32_t width_, uint32_t height_,
+                      uint32_t bitrate_, int frames_)
+      : name(name_), width(width_), height(height_), bitrate(bitrate_),
+        frames(frames_) {}
+  const char *name;
+  uint32_t width;
+  uint32_t height;
+  uint32_t bitrate;
+  int frames;
+};
+
+const EncodePerfTestVideo kAV1EncodePerfTestVectors[] = {
+  EncodePerfTestVideo("niklas_1280_720_30.yuv", 1280, 720, 600, 470),
+};
+
+TEST_P(AV1NewEncodeDecodePerfTest, PerfTest) {
+  SetUp();
+
+  // TODO(JBB): Make this work by going through the set of given files.
+  const int i = 0;
+  const aom_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = kAV1EncodePerfTestVectors[i].bitrate;
+
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  const char *video_name = kAV1EncodePerfTestVectors[i].name;
+  libaom_test::I420VideoSource video(
+      video_name, kAV1EncodePerfTestVectors[i].width,
+      kAV1EncodePerfTestVectors[i].height, timebase.den, timebase.num, 0,
+      kAV1EncodePerfTestVectors[i].frames);
+  set_speed(2);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  const uint32_t threads = 4;
+
+  libaom_test::IVFVideoSource decode_video(kNewEncodeOutputFile);
+  decode_video.Init();
+
+  aom_codec_dec_cfg_t cfg = aom_codec_dec_cfg_t();
+  cfg.threads = threads;
+  libaom_test::AV1Decoder decoder(cfg, 0);
+
+  aom_usec_timer t;
+  aom_usec_timer_start(&t);
+
+  for (decode_video.Begin(); decode_video.cxdata() != NULL;
+       decode_video.Next()) {
+    decoder.DecodeFrame(decode_video.cxdata(), decode_video.frame_size());
+  }
+
+  aom_usec_timer_mark(&t);
+  const double elapsed_secs =
+      static_cast<double>(aom_usec_timer_elapsed(&t)) / kUsecsInSec;
+  const unsigned decode_frames = decode_video.frame_number();
+  const double fps = static_cast<double>(decode_frames) / elapsed_secs;
+
+  printf("{\n");
+  printf("\t\"type\" : \"decode_perf_test\",\n");
+  printf("\t\"version\" : \"%s\",\n", VERSION_STRING_NOSP);
+  printf("\t\"videoName\" : \"%s\",\n", kNewEncodeOutputFile);
+  printf("\t\"threadCount\" : %u,\n", threads);
+  printf("\t\"decodeTimeSecs\" : %f,\n", elapsed_secs);
+  printf("\t\"totalFrames\" : %u,\n", decode_frames);
+  printf("\t\"framesPerSecond\" : %f\n", fps);
+  printf("}\n");
+}
+
+AV1_INSTANTIATE_TEST_CASE(AV1NewEncodeDecodePerfTest,
+                          ::testing::Values(::libaom_test::kTwoPassGood));
+}  // namespace
diff --git a/third_party/aom/test/decode_test_driver.cc b/third_party/aom/test/decode_test_driver.cc
new file mode 100644
index 0000000000..35c28eafd5
--- /dev/null
+++ b/third_party/aom/test/decode_test_driver.cc
@@ -0,0 +1,124 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/register_state_check.h"
+#include "test/video_source.h"
+
+namespace libaom_test {
+
+const char kVP8Name[] = "WebM Project VP8";
+const char kAV1Name[] = "AOMedia Project AV1 Decoder";
+
+aom_codec_err_t Decoder::PeekStream(const uint8_t *cxdata, size_t size,
+                                    aom_codec_stream_info_t *stream_info) {
+  return aom_codec_peek_stream_info(
+      CodecInterface(), cxdata, static_cast<unsigned int>(size), stream_info);
+}
+
+aom_codec_err_t Decoder::DecodeFrame(const uint8_t *cxdata, size_t size) {
+  return DecodeFrame(cxdata, size, NULL);
+}
+
+aom_codec_err_t Decoder::DecodeFrame(const uint8_t *cxdata, size_t size,
+                                     void *user_priv) {
+  aom_codec_err_t res_dec;
+  InitOnce();
+  API_REGISTER_STATE_CHECK(
+      res_dec = aom_codec_decode(
+          &decoder_, cxdata, static_cast<unsigned int>(size), user_priv, 0));
+  return res_dec;
+}
+
+bool Decoder::IsVP8() const {
+  const char *codec_name = GetDecoderName();
+  return strncmp(kVP8Name, codec_name, sizeof(kVP8Name) - 1) == 0;
+}
+
+bool Decoder::IsAV1() const {
+  const char *codec_name = GetDecoderName();
+  return strncmp(kAV1Name, codec_name, sizeof(kAV1Name) - 1) == 0;
+}
+
+void DecoderTest::HandlePeekResult(Decoder *const decoder,
+                                   CompressedVideoSource *video,
+                                   const aom_codec_err_t res_peek) {
+  const bool is_vp8 = decoder->IsVP8();
+  if (is_vp8) {
+    /* Vp8's implementation of PeekStream returns an error if the frame you
+     * pass it is not a keyframe, so we only expect AOM_CODEC_OK on the first
+     * frame, which must be a keyframe. */
+    if (video->frame_number() == 0)
+      ASSERT_EQ(AOM_CODEC_OK, res_peek) << "Peek return failed: "
+                                        << aom_codec_err_to_string(res_peek);
+  } else {
+    /* The Av1 implementation of PeekStream returns an error only if the
+     * data passed to it isn't a valid Av1 chunk. */
+    ASSERT_EQ(AOM_CODEC_OK, res_peek) << "Peek return failed: "
+                                      << aom_codec_err_to_string(res_peek);
+  }
+}
+
+void DecoderTest::RunLoop(CompressedVideoSource *video,
+                          const aom_codec_dec_cfg_t &dec_cfg) {
+  Decoder *const decoder = codec_->CreateDecoder(dec_cfg, flags_);
+  ASSERT_TRUE(decoder != NULL);
+  bool end_of_file = false;
+
+  // Decode frames.
+  for (video->Begin(); !::testing::Test::HasFailure() && !end_of_file;
+       video->Next()) {
+    PreDecodeFrameHook(*video, decoder);
+
+    aom_codec_stream_info_t stream_info;
+    stream_info.sz = sizeof(stream_info);
+
+    if (video->cxdata() != NULL) {
+      const aom_codec_err_t res_peek = decoder->PeekStream(
+          video->cxdata(), video->frame_size(), &stream_info);
+      HandlePeekResult(decoder, video, res_peek);
+      ASSERT_FALSE(::testing::Test::HasFailure());
+
+      aom_codec_err_t res_dec =
+          decoder->DecodeFrame(video->cxdata(), video->frame_size());
+      if (!HandleDecodeResult(res_dec, decoder)) break;
+    } else {
+      // Signal end of the file to the decoder.
+      const aom_codec_err_t res_dec = decoder->DecodeFrame(NULL, 0);
+      ASSERT_EQ(AOM_CODEC_OK, res_dec) << decoder->DecodeError();
+      end_of_file = true;
+    }
+
+    DxDataIterator dec_iter = decoder->GetDxData();
+    const aom_image_t *img = NULL;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next()))
+      DecompressedFrameHook(*img, video->frame_number());
+  }
+  delete decoder;
+}
+
+void DecoderTest::RunLoop(CompressedVideoSource *video) {
+  aom_codec_dec_cfg_t dec_cfg = aom_codec_dec_cfg_t();
+  RunLoop(video, dec_cfg);
+}
+
+void DecoderTest::set_cfg(const aom_codec_dec_cfg_t &dec_cfg) {
+  memcpy(&cfg_, &dec_cfg, sizeof(cfg_));
+}
+
+void DecoderTest::set_flags(const aom_codec_flags_t flags) { flags_ = flags; }
+
+}  // namespace libaom_test
diff --git a/third_party/aom/test/decode_test_driver.h b/third_party/aom/test/decode_test_driver.h
new file mode 100644
index 0000000000..e7deb389c9
--- /dev/null
+++ b/third_party/aom/test/decode_test_driver.h
@@ -0,0 +1,164 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef TEST_DECODE_TEST_DRIVER_H_
+#define TEST_DECODE_TEST_DRIVER_H_
+#include <cstring>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "./aom_config.h"
+#include "aom/aom_decoder.h"
+
+namespace libaom_test {
+
+class CodecFactory;
+class CompressedVideoSource;
+
+// Provides an object to handle decoding output
+class DxDataIterator {
+ public:
+  explicit DxDataIterator(aom_codec_ctx_t *decoder)
+      : decoder_(decoder), iter_(NULL) {}
+
+  const aom_image_t *Next() { return aom_codec_get_frame(decoder_, &iter_); }
+
+ private:
+  aom_codec_ctx_t *decoder_;
+  aom_codec_iter_t iter_;
+};
+
+// Provides a simplified interface to manage one video decoding.
+// Similar to Encoder class, the exact services should be added
+// as more tests are added.
+class Decoder {
+ public:
+  explicit Decoder(aom_codec_dec_cfg_t cfg)
+      : cfg_(cfg), flags_(0), init_done_(false) {
+    memset(&decoder_, 0, sizeof(decoder_));
+  }
+
+  Decoder(aom_codec_dec_cfg_t cfg, const aom_codec_flags_t flag)
+      : cfg_(cfg), flags_(flag), init_done_(false) {
+    memset(&decoder_, 0, sizeof(decoder_));
+  }
+
+  virtual ~Decoder() { aom_codec_destroy(&decoder_); }
+
+  aom_codec_err_t PeekStream(const uint8_t *cxdata, size_t size,
+                             aom_codec_stream_info_t *stream_info);
+
+  aom_codec_err_t DecodeFrame(const uint8_t *cxdata, size_t size);
+
+  aom_codec_err_t DecodeFrame(const uint8_t *cxdata, size_t size,
+                              void *user_priv);
+
+  DxDataIterator GetDxData() { return DxDataIterator(&decoder_); }
+
+  void Control(int ctrl_id, int arg) { Control(ctrl_id, arg, AOM_CODEC_OK); }
+
+  void Control(int ctrl_id, const void *arg) {
+    InitOnce();
+    const aom_codec_err_t res = aom_codec_control_(&decoder_, ctrl_id, arg);
+    ASSERT_EQ(AOM_CODEC_OK, res) << DecodeError();
+  }
+
+  void Control(int ctrl_id, int arg, aom_codec_err_t expected_value) {
+    InitOnce();
+    const aom_codec_err_t res = aom_codec_control_(&decoder_, ctrl_id, arg);
+    ASSERT_EQ(expected_value, res) << DecodeError();
+  }
+
+  const char *DecodeError() {
+    const char *detail = aom_codec_error_detail(&decoder_);
+    return detail ? detail : aom_codec_error(&decoder_);
+  }
+
+  // Passes the external frame buffer information to libaom.
+  aom_codec_err_t SetFrameBufferFunctions(
+      aom_get_frame_buffer_cb_fn_t cb_get,
+      aom_release_frame_buffer_cb_fn_t cb_release, void *user_priv) {
+    InitOnce();
+    return aom_codec_set_frame_buffer_functions(&decoder_, cb_get, cb_release,
+                                                user_priv);
+  }
+
+  const char *GetDecoderName() const {
+    return aom_codec_iface_name(CodecInterface());
+  }
+
+  bool IsVP8() const;
+
+  bool IsAV1() const;
+
+  aom_codec_ctx_t *GetDecoder() { return &decoder_; }
+
+ protected:
+  virtual aom_codec_iface_t *CodecInterface() const = 0;
+
+  void InitOnce() {
+    if (!init_done_) {
+      const aom_codec_err_t res =
+          aom_codec_dec_init(&decoder_, CodecInterface(), &cfg_, flags_);
+      ASSERT_EQ(AOM_CODEC_OK, res) << DecodeError();
+      init_done_ = true;
+    }
+  }
+
+  aom_codec_ctx_t decoder_;
+  aom_codec_dec_cfg_t cfg_;
+  aom_codec_flags_t flags_;
+  bool init_done_;
+};
+
+// Common test functionality for all Decoder tests.
+class DecoderTest {
+ public:
+  // Main decoding loop
+  virtual void RunLoop(CompressedVideoSource *video);
+  virtual void RunLoop(CompressedVideoSource *video,
+                       const aom_codec_dec_cfg_t &dec_cfg);
+
+  virtual void set_cfg(const aom_codec_dec_cfg_t &dec_cfg);
+  virtual void set_flags(const aom_codec_flags_t flags);
+
+  // Hook to be called before decompressing every frame.
+  virtual void PreDecodeFrameHook(const CompressedVideoSource & /*video*/,
+                                  Decoder * /*decoder*/) {}
+
+  // Hook to be called to handle decode result. Return true to continue.
+  virtual bool HandleDecodeResult(const aom_codec_err_t res_dec,
+                                  Decoder *decoder) {
+    EXPECT_EQ(AOM_CODEC_OK, res_dec) << decoder->DecodeError();
+    return AOM_CODEC_OK == res_dec;
+  }
+
+  // Hook to be called on every decompressed frame.
+  virtual void DecompressedFrameHook(const aom_image_t & /*img*/,
+                                     const unsigned int /*frame_number*/) {}
+
+  // Hook to be called on peek result
+  virtual void HandlePeekResult(Decoder *const decoder,
+                                CompressedVideoSource *video,
+                                const aom_codec_err_t res_peek);
+
+ protected:
+  explicit DecoderTest(const CodecFactory *codec)
+      : codec_(codec), cfg_(), flags_(0) {}
+
+  virtual ~DecoderTest() {}
+
+  const CodecFactory *codec_;
+  aom_codec_dec_cfg_t cfg_;
+  aom_codec_flags_t flags_;
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_DECODE_TEST_DRIVER_H_
diff --git a/third_party/aom/test/decode_to_md5.sh b/third_party/aom/test/decode_to_md5.sh
new file mode 100755
index 0000000000..44c9f5f052
--- /dev/null
+++ b/third_party/aom/test/decode_to_md5.sh
@@ -0,0 +1,67 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This file tests the libaom decode_to_md5 example. To add new tests to this
+## file, do the following:
+##   1. Write a shell function (this is your test).
+##   2. Add the function to decode_to_md5_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+#   $AOM_IVF_FILE and $AV1_IVF_FILE are required.
+decode_to_md5_verify_environment() {
+  if [ "$(av1_encode_available)" != "yes" ] && [ ! -e "${AV1_IVF_FILE}" ]; then
+    return 1
+  fi
+}
+
+# Runs decode_to_md5 on $1 and captures the md5 sum for the final frame. $2 is
+# interpreted as codec name and used solely to name the output file. $3 is the
+# expected md5 sum: It must match that of the final frame.
+decode_to_md5() {
+  local decoder="${LIBAOM_BIN_PATH}/decode_to_md5${AOM_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local expected_md5="$3"
+  local output_file="${AOM_TEST_OUTPUT_DIR}/decode_to_md5_${codec}"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${AOM_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+
+  local md5_last_frame="$(tail -n1 "${output_file}" | awk '{print $1}')"
+  local actual_md5="$(echo "${md5_last_frame}" | awk '{print $1}')"
+  [ "${actual_md5}" = "${expected_md5}" ] || return 1
+}
+
+decode_to_md5_av1() {
+  # expected MD5 sum for the last frame.
+  local expected_md5="26d3ef1d60754a1f6acb603c3763efbe"
+  local file="${AV1_IVF_FILE}"
+
+  if [ "$(av1_decode_available)" = "yes" ]; then
+    if [ ! -e "${AV1_IVF_FILE}" ]; then
+      file="${AOM_TEST_OUTPUT_DIR}/test_encode.ivf"
+      encode_yuv_raw_input_av1 "${file}" --ivf
+    fi
+    decode_to_md5 "${file}" "av1" "${expected_md5}"
+  fi
+}
+
+decode_to_md5_tests="decode_to_md5_av1"
+
+run_tests decode_to_md5_verify_environment "${decode_to_md5_tests}"
diff --git a/third_party/aom/test/decode_with_drops.sh b/third_party/aom/test/decode_with_drops.sh
new file mode 100755
index 0000000000..5978312f20
--- /dev/null
+++ b/third_party/aom/test/decode_with_drops.sh
@@ -0,0 +1,67 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This file tests the libaom decode_with_drops example. To add new tests to
+## this file, do the following:
+##   1. Write a shell function (this is your test).
+##   2. Add the function to decode_with_drops_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+#   $AOM_IVF_FILE and $AV1_IVF_FILE are required.
+decode_with_drops_verify_environment() {
+  if [ "$(av1_encode_available)" != "yes" ] && [ ! -e "${AV1_IVF_FILE}" ]; then
+    return 1
+  fi
+}
+
+# Runs decode_with_drops on $1, $2 is interpreted as codec name and used solely
+# to name the output file. $3 is the drop mode, and is passed directly to
+# decode_with_drops.
+decode_with_drops() {
+  local decoder="${LIBAOM_BIN_PATH}/decode_with_drops${AOM_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local output_file="${AOM_TEST_OUTPUT_DIR}/decode_with_drops_${codec}"
+  local drop_mode="$3"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${AOM_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      "${drop_mode}" ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+
+# Decodes $AV1_IVF_FILE while dropping frames, twice: once in sequence mode,
+# and once in pattern mode.
+decode_with_drops_av1() {
+  if [ "$(av1_decode_available)" = "yes" ]; then
+    local file="${AV1_IVF_FILE}"
+    if [ ! -e "${AV1_IVF_FILE}" ]; then
+      file="${AOM_TEST_OUTPUT_DIR}/test_encode.ivf"
+      encode_yuv_raw_input_av1 "${file}" --ivf
+    fi
+    # Drop frames 2 and 3.
+    decode_with_drops "${file}" "av1" "2-3"
+
+    # Test pattern mode: Drop 3 of every 4 frames.
+    decode_with_drops "${file}" "av1" "3/4"
+  fi
+}
+
+decode_with_drops_tests="decode_with_drops_av1"
+
+run_tests decode_with_drops_verify_environment "${decode_with_drops_tests}"
diff --git a/third_party/aom/test/dering_test.cc b/third_party/aom/test/dering_test.cc
new file mode 100644
index 0000000000..195a60ff81
--- /dev/null
+++ b/third_party/aom/test/dering_test.cc
@@ -0,0 +1,388 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <cstdlib>
+#include <string>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "aom_ports/aom_timer.h"
+#include "av1/common/od_dering.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+
+typedef std::tr1::tuple<od_filter_dering_direction_func,
+                        od_filter_dering_direction_func, int>
+    dering_dir_param_t;
+
+class CDEFDeringDirTest : public ::testing::TestWithParam<dering_dir_param_t> {
+ public:
+  virtual ~CDEFDeringDirTest() {}
+  virtual void SetUp() {
+    dering = GET_PARAM(0);
+    ref_dering = GET_PARAM(1);
+    bsize = GET_PARAM(2);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  int bsize;
+  od_filter_dering_direction_func dering;
+  od_filter_dering_direction_func ref_dering;
+};
+
+typedef CDEFDeringDirTest CDEFDeringSpeedTest;
+
+void test_dering(int bsize, int iterations,
+                 od_filter_dering_direction_func dering,
+                 od_filter_dering_direction_func ref_dering) {
+  const int size = 8;
+  const int ysize = size + 2 * OD_FILT_VBORDER;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint16_t, s[ysize * OD_FILT_BSTRIDE]);
+  DECLARE_ALIGNED(16, static uint16_t, d[size * size]);
+  DECLARE_ALIGNED(16, static uint16_t, ref_d[size * size]);
+  memset(ref_d, 0, sizeof(ref_d));
+  memset(d, 0, sizeof(d));
+
+  int error = 0, threshold = 0, dir;
+  int boundary, damping, depth, bits, level, count,
+      errdepth = 0, errthreshold = 0, errboundary = 0, errdamping = 0;
+  unsigned int pos = 0;
+
+  for (boundary = 0; boundary < 16; boundary++) {
+    for (depth = 8; depth <= 12; depth += 2) {
+      for (damping = 5 + depth - 8; damping < 7 + depth - 8; damping++) {
+        for (count = 0; count < iterations; count++) {
+          for (level = 0; level < (1 << depth) && !error;
+               level += (1 + 4 * !!boundary) << (depth - 8)) {
+            for (bits = 1; bits <= depth && !error; bits++) {
+              for (unsigned int i = 0; i < sizeof(s) / sizeof(*s); i++)
+                s[i] = clamp((rnd.Rand16() & ((1 << bits) - 1)) + level, 0,
+                             (1 << depth) - 1);
+              if (boundary) {
+                if (boundary & 1) {  // Left
+                  for (int i = 0; i < ysize; i++)
+                    for (int j = 0; j < OD_FILT_HBORDER; j++)
+                      s[i * OD_FILT_BSTRIDE + j] = OD_DERING_VERY_LARGE;
+                }
+                if (boundary & 2) {  // Right
+                  for (int i = 0; i < ysize; i++)
+                    for (int j = OD_FILT_HBORDER + size; j < OD_FILT_BSTRIDE;
+                         j++)
+                      s[i * OD_FILT_BSTRIDE + j] = OD_DERING_VERY_LARGE;
+                }
+                if (boundary & 4) {  // Above
+                  for (int i = 0; i < OD_FILT_VBORDER; i++)
+                    for (int j = 0; j < OD_FILT_BSTRIDE; j++)
+                      s[i * OD_FILT_BSTRIDE + j] = OD_DERING_VERY_LARGE;
+                }
+                if (boundary & 8) {  // Below
+                  for (int i = OD_FILT_VBORDER + size; i < ysize; i++)
+                    for (int j = 0; j < OD_FILT_BSTRIDE; j++)
+                      s[i * OD_FILT_BSTRIDE + j] = OD_DERING_VERY_LARGE;
+                }
+              }
+              for (dir = 0; dir < 8; dir++) {
+                for (threshold = 0; threshold < 64 << (depth - 8) && !error;
+                     threshold += (1 + 4 * !!boundary) << (depth - 8)) {
+                  ref_dering(ref_d, size, s + OD_FILT_HBORDER +
+                                              OD_FILT_VBORDER * OD_FILT_BSTRIDE,
+                             threshold, dir, damping);
+                  // If dering and ref_dering are the same, we're just testing
+                  // speed
+                  if (dering != ref_dering)
+                    ASM_REGISTER_STATE_CHECK(dering(
+                        d, size,
+                        s + OD_FILT_HBORDER + OD_FILT_VBORDER * OD_FILT_BSTRIDE,
+                        threshold, dir, damping));
+                  if (ref_dering != dering) {
+                    for (pos = 0; pos < sizeof(d) / sizeof(*d) && !error;
+                         pos++) {
+                      error = ref_d[pos] != d[pos];
+                      errdepth = depth;
+                      errthreshold = threshold;
+                      errboundary = boundary;
+                      errdamping = damping;
+                    }
+                  }
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  pos--;
+  EXPECT_EQ(0, error) << "Error: CDEFDeringDirTest, SIMD and C mismatch."
+                      << std::endl
+                      << "First error at " << pos % size << "," << pos / size
+                      << " (" << (int16_t)ref_d[pos] << " : " << (int16_t)d[pos]
+                      << ") " << std::endl
+                      << "threshold: " << errthreshold << std::endl
+                      << "damping: " << errdamping << std::endl
+                      << "depth: " << errdepth << std::endl
+                      << "size: " << bsize << std::endl
+                      << "boundary: " << errboundary << std::endl
+                      << std::endl;
+}
+
+void test_dering_speed(int bsize, int iterations,
+                       od_filter_dering_direction_func dering,
+                       od_filter_dering_direction_func ref_dering) {
+  aom_usec_timer ref_timer;
+  aom_usec_timer timer;
+
+  aom_usec_timer_start(&ref_timer);
+  test_dering(bsize, iterations, ref_dering, ref_dering);
+  aom_usec_timer_mark(&ref_timer);
+  int ref_elapsed_time = (int)aom_usec_timer_elapsed(&ref_timer);
+
+  aom_usec_timer_start(&timer);
+  test_dering(bsize, iterations, dering, dering);
+  aom_usec_timer_mark(&timer);
+  int elapsed_time = (int)aom_usec_timer_elapsed(&timer);
+
+#if 0
+  std::cout << "[          ] C time = " << ref_elapsed_time / 1000
+            << " ms, SIMD time = " << elapsed_time / 1000 << " ms" << std::endl;
+#endif
+
+  EXPECT_GT(ref_elapsed_time, elapsed_time)
+      << "Error: CDEFDeringSpeedTest, SIMD slower than C." << std::endl
+      << "C time: " << ref_elapsed_time << " us" << std::endl
+      << "SIMD time: " << elapsed_time << " us" << std::endl;
+}
+
+typedef int (*find_dir_t)(const od_dering_in *img, int stride, int32_t *var,
+                          int coeff_shift);
+
+typedef std::tr1::tuple<find_dir_t, find_dir_t> find_dir_param_t;
+
+class CDEFDeringFindDirTest
+    : public ::testing::TestWithParam<find_dir_param_t> {
+ public:
+  virtual ~CDEFDeringFindDirTest() {}
+  virtual void SetUp() {
+    finddir = GET_PARAM(0);
+    ref_finddir = GET_PARAM(1);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  find_dir_t finddir;
+  find_dir_t ref_finddir;
+};
+
+typedef CDEFDeringFindDirTest CDEFDeringFindDirSpeedTest;
+
+void test_finddir(int (*finddir)(const od_dering_in *img, int stride,
+                                 int32_t *var, int coeff_shift),
+                  int (*ref_finddir)(const od_dering_in *img, int stride,
+                                     int32_t *var, int coeff_shift)) {
+  const int size = 8;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint16_t, s[size * size]);
+
+  int error = 0;
+  int depth, bits, level, count, errdepth = 0;
+  int ref_res = 0, res = 0;
+  int32_t ref_var = 0, var = 0;
+
+  for (depth = 8; depth <= 12 && !error; depth += 2) {
+    for (count = 0; count < 512 && !error; count++) {
+      for (level = 0; level < (1 << depth) && !error;
+           level += 1 << (depth - 8)) {
+        for (bits = 1; bits <= depth && !error; bits++) {
+          for (unsigned int i = 0; i < sizeof(s) / sizeof(*s); i++)
+            s[i] = clamp((rnd.Rand16() & ((1 << bits) - 1)) + level, 0,
+                         (1 << depth) - 1);
+          for (int c = 0; c < 1 + 9 * (finddir == ref_finddir); c++)
+            ref_res = ref_finddir(s, size, &ref_var, depth - 8);
+          if (finddir != ref_finddir)
+            ASM_REGISTER_STATE_CHECK(res = finddir(s, size, &var, depth - 8));
+          if (ref_finddir != finddir) {
+            if (res != ref_res || var != ref_var) error = 1;
+            errdepth = depth;
+          }
+        }
+      }
+    }
+  }
+
+  EXPECT_EQ(0, error) << "Error: CDEFDeringFindDirTest, SIMD and C mismatch."
+                      << std::endl
+                      << "return: " << res << " : " << ref_res << std::endl
+                      << "var: " << var << " : " << ref_var << std::endl
+                      << "depth: " << errdepth << std::endl
+                      << std::endl;
+}
+
+void test_finddir_speed(int (*finddir)(const od_dering_in *img, int stride,
+                                       int32_t *var, int coeff_shift),
+                        int (*ref_finddir)(const od_dering_in *img, int stride,
+                                           int32_t *var, int coeff_shift)) {
+  aom_usec_timer ref_timer;
+  aom_usec_timer timer;
+
+  aom_usec_timer_start(&ref_timer);
+  test_finddir(ref_finddir, ref_finddir);
+  aom_usec_timer_mark(&ref_timer);
+  int ref_elapsed_time = (int)aom_usec_timer_elapsed(&ref_timer);
+
+  aom_usec_timer_start(&timer);
+  test_finddir(finddir, finddir);
+  aom_usec_timer_mark(&timer);
+  int elapsed_time = (int)aom_usec_timer_elapsed(&timer);
+
+#if 0
+  std::cout << "[          ] C time = " << ref_elapsed_time / 1000
+            << " ms, SIMD time = " << elapsed_time / 1000 << " ms" << std::endl;
+#endif
+
+  EXPECT_GT(ref_elapsed_time, elapsed_time)
+      << "Error: CDEFDeringFindDirSpeedTest, SIMD slower than C." << std::endl
+      << "C time: " << ref_elapsed_time << " us" << std::endl
+      << "SIMD time: " << elapsed_time << " us" << std::endl;
+}
+
+TEST_P(CDEFDeringDirTest, TestSIMDNoMismatch) {
+  test_dering(bsize, 1, dering, ref_dering);
+}
+
+TEST_P(CDEFDeringSpeedTest, DISABLED_TestSpeed) {
+  test_dering_speed(bsize, 4, dering, ref_dering);
+}
+
+TEST_P(CDEFDeringFindDirTest, TestSIMDNoMismatch) {
+  test_finddir(finddir, ref_finddir);
+}
+
+TEST_P(CDEFDeringFindDirSpeedTest, DISABLED_TestSpeed) {
+  test_finddir_speed(finddir, ref_finddir);
+}
+
+using std::tr1::make_tuple;
+
+// VS compiling for 32 bit targets does not support vector types in
+// structs as arguments, which makes the v256 type of the intrinsics
+// hard to support, so optimizations for this target are disabled.
+#if defined(_WIN64) || !defined(_MSC_VER) || defined(__clang__)
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, CDEFDeringDirTest,
+    ::testing::Values(make_tuple(&od_filter_dering_direction_4x4_sse2,
+                                 &od_filter_dering_direction_4x4_c, 4),
+                      make_tuple(&od_filter_dering_direction_8x8_sse2,
+                                 &od_filter_dering_direction_8x8_c, 8)));
+INSTANTIATE_TEST_CASE_P(SSE2, CDEFDeringFindDirTest,
+                        ::testing::Values(make_tuple(&od_dir_find8_sse2,
+                                                     &od_dir_find8_c)));
+#endif
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, CDEFDeringDirTest,
+    ::testing::Values(make_tuple(&od_filter_dering_direction_4x4_ssse3,
+                                 &od_filter_dering_direction_4x4_c, 4),
+                      make_tuple(&od_filter_dering_direction_8x8_ssse3,
+                                 &od_filter_dering_direction_8x8_c, 8)));
+INSTANTIATE_TEST_CASE_P(SSSE3, CDEFDeringFindDirTest,
+                        ::testing::Values(make_tuple(&od_dir_find8_ssse3,
+                                                     &od_dir_find8_c)));
+#endif
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, CDEFDeringDirTest,
+    ::testing::Values(make_tuple(&od_filter_dering_direction_4x4_sse4_1,
+                                 &od_filter_dering_direction_4x4_c, 4),
+                      make_tuple(&od_filter_dering_direction_8x8_sse4_1,
+                                 &od_filter_dering_direction_8x8_c, 8)));
+INSTANTIATE_TEST_CASE_P(SSE4_1, CDEFDeringFindDirTest,
+                        ::testing::Values(make_tuple(&od_dir_find8_sse4_1,
+                                                     &od_dir_find8_c)));
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+    NEON, CDEFDeringDirTest,
+    ::testing::Values(make_tuple(&od_filter_dering_direction_4x4_neon,
+                                 &od_filter_dering_direction_4x4_c, 4),
+                      make_tuple(&od_filter_dering_direction_8x8_neon,
+                                 &od_filter_dering_direction_8x8_c, 8)));
+INSTANTIATE_TEST_CASE_P(NEON, CDEFDeringFindDirTest,
+                        ::testing::Values(make_tuple(&od_dir_find8_neon,
+                                                     &od_dir_find8_c)));
+#endif
+
+// Test speed for all supported architectures
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, CDEFDeringSpeedTest,
+    ::testing::Values(make_tuple(&od_filter_dering_direction_4x4_sse2,
+                                 &od_filter_dering_direction_4x4_c, 4),
+                      make_tuple(&od_filter_dering_direction_8x8_sse2,
+                                 &od_filter_dering_direction_8x8_c, 8)));
+INSTANTIATE_TEST_CASE_P(SSE2, CDEFDeringFindDirSpeedTest,
+                        ::testing::Values(make_tuple(&od_dir_find8_sse2,
+                                                     &od_dir_find8_c)));
+#endif
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, CDEFDeringSpeedTest,
+    ::testing::Values(make_tuple(&od_filter_dering_direction_4x4_ssse3,
+                                 &od_filter_dering_direction_4x4_c, 4),
+                      make_tuple(&od_filter_dering_direction_8x8_ssse3,
+                                 &od_filter_dering_direction_8x8_c, 8)));
+INSTANTIATE_TEST_CASE_P(SSSE3, CDEFDeringFindDirSpeedTest,
+                        ::testing::Values(make_tuple(&od_dir_find8_ssse3,
+                                                     &od_dir_find8_c)));
+#endif
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, CDEFDeringSpeedTest,
+    ::testing::Values(make_tuple(&od_filter_dering_direction_4x4_sse4_1,
+                                 &od_filter_dering_direction_4x4_c, 4),
+                      make_tuple(&od_filter_dering_direction_8x8_sse4_1,
+                                 &od_filter_dering_direction_8x8_c, 8)));
+INSTANTIATE_TEST_CASE_P(SSE4_1, CDEFDeringFindDirSpeedTest,
+                        ::testing::Values(make_tuple(&od_dir_find8_sse4_1,
+                                                     &od_dir_find8_c)));
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+    NEON, CDEFDeringSpeedTest,
+    ::testing::Values(make_tuple(&od_filter_dering_direction_4x4_neon,
+                                 &od_filter_dering_direction_4x4_c, 4),
+                      make_tuple(&od_filter_dering_direction_8x8_neon,
+                                 &od_filter_dering_direction_8x8_c, 8)));
+INSTANTIATE_TEST_CASE_P(NEON, CDEFDeringFindDirSpeedTest,
+                        ::testing::Values(make_tuple(&od_dir_find8_neon,
+                                                     &od_dir_find8_c)));
+#endif
+
+#endif  // defined(_WIN64) || !defined(_MSC_VER)
+}  // namespace
diff --git a/third_party/aom/test/divu_small_test.cc b/third_party/aom/test/divu_small_test.cc
new file mode 100644
index 0000000000..064f8ee454
--- /dev/null
+++ b/third_party/aom/test/divu_small_test.cc
@@ -0,0 +1,41 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/acm_random.h"
+#include "av1/common/odintrin.h"
+
+using libaom_test::ACMRandom;
+
+TEST(Daala, TestDIVUuptoMAX) {
+  for (int d = 1; d <= OD_DIVU_DMAX; d++) {
+    for (uint32_t x = 1; x <= 1000000; x++) {
+      GTEST_ASSERT_EQ(x / d, OD_DIVU_SMALL(x, d))
+          << "x=" << x << " d=" << d << " x/d=" << (x / d)
+          << " != " << OD_DIVU_SMALL(x, d);
+    }
+  }
+}
+
+TEST(Daala, TestDIVUrandI31) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  for (int d = 1; d < OD_DIVU_DMAX; d++) {
+    for (int i = 0; i < 1000000; i++) {
+      uint32_t x = rnd.Rand31();
+      GTEST_ASSERT_EQ(x / d, OD_DIVU_SMALL(x, d))
+          << "x=" << x << " d=" << d << " x/d=" << (x / d)
+          << " != " << OD_DIVU_SMALL(x, d);
+    }
+  }
+}
diff --git a/third_party/aom/test/encode_api_test.cc b/third_party/aom/test/encode_api_test.cc
new file mode 100644
index 0000000000..14e43c847f
--- /dev/null
+++ b/third_party/aom/test/encode_api_test.cc
@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "aom/aomcx.h"
+#include "aom/aom_encoder.h"
+
+namespace {
+
+#define NELEMENTS(x) static_cast<int>(sizeof(x) / sizeof(x[0]))
+
+TEST(EncodeAPI, InvalidParams) {
+  static const aom_codec_iface_t *kCodecs[] = {
+#if CONFIG_AV1_ENCODER
+    &aom_codec_av1_cx_algo,
+#endif
+  };
+  uint8_t buf[1] = { 0 };
+  aom_image_t img;
+  aom_codec_ctx_t enc;
+  aom_codec_enc_cfg_t cfg;
+
+  EXPECT_EQ(&img, aom_img_wrap(&img, AOM_IMG_FMT_I420, 1, 1, 1, buf));
+
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_enc_init(NULL, NULL, NULL, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_enc_init(&enc, NULL, NULL, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_encode(NULL, NULL, 0, 0, 0, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_encode(NULL, &img, 0, 0, 0, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM, aom_codec_destroy(NULL));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM,
+            aom_codec_enc_config_default(NULL, NULL, 0));
+  EXPECT_EQ(AOM_CODEC_INVALID_PARAM,
+            aom_codec_enc_config_default(NULL, &cfg, 0));
+  EXPECT_TRUE(aom_codec_error(NULL) != NULL);
+
+  for (int i = 0; i < NELEMENTS(kCodecs); ++i) {
+    SCOPED_TRACE(aom_codec_iface_name(kCodecs[i]));
+    EXPECT_EQ(AOM_CODEC_INVALID_PARAM,
+              aom_codec_enc_init(NULL, kCodecs[i], NULL, 0));
+    EXPECT_EQ(AOM_CODEC_INVALID_PARAM,
+              aom_codec_enc_init(&enc, kCodecs[i], NULL, 0));
+    EXPECT_EQ(AOM_CODEC_INVALID_PARAM,
+              aom_codec_enc_config_default(kCodecs[i], &cfg, 1));
+
+    EXPECT_EQ(AOM_CODEC_OK, aom_codec_enc_config_default(kCodecs[i], &cfg, 0));
+    EXPECT_EQ(AOM_CODEC_OK, aom_codec_enc_init(&enc, kCodecs[i], &cfg, 0));
+    EXPECT_EQ(AOM_CODEC_OK, aom_codec_encode(&enc, NULL, 0, 0, 0, 0));
+
+    EXPECT_EQ(AOM_CODEC_OK, aom_codec_destroy(&enc));
+  }
+}
+
+}  // namespace
diff --git a/third_party/aom/test/encode_perf_test.cc b/third_party/aom/test/encode_perf_test.cc
new file mode 100644
index 0000000000..e2a4f2b71c
--- /dev/null
+++ b/third_party/aom/test/encode_perf_test.cc
@@ -0,0 +1,189 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <string>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "./aom_config.h"
+#include "./aom_version.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "aom_ports/aom_timer.h"
+
+namespace {
+
+const int kMaxPsnr = 100;
+const double kUsecsInSec = 1000000.0;
+
+struct EncodePerfTestVideo {
+  EncodePerfTestVideo(const char *name_, uint32_t width_, uint32_t height_,
+                      uint32_t bitrate_, int frames_)
+      : name(name_), width(width_), height(height_), bitrate(bitrate_),
+        frames(frames_) {}
+  const char *name;
+  uint32_t width;
+  uint32_t height;
+  uint32_t bitrate;
+  int frames;
+};
+
+const EncodePerfTestVideo kAV1EncodePerfTestVectors[] = {
+  EncodePerfTestVideo("desktop_640_360_30.yuv", 640, 360, 200, 2484),
+  EncodePerfTestVideo("kirland_640_480_30.yuv", 640, 480, 200, 300),
+  EncodePerfTestVideo("macmarcomoving_640_480_30.yuv", 640, 480, 200, 987),
+  EncodePerfTestVideo("macmarcostationary_640_480_30.yuv", 640, 480, 200, 718),
+  EncodePerfTestVideo("niklas_640_480_30.yuv", 640, 480, 200, 471),
+  EncodePerfTestVideo("tacomanarrows_640_480_30.yuv", 640, 480, 200, 300),
+  EncodePerfTestVideo("tacomasmallcameramovement_640_480_30.yuv", 640, 480, 200,
+                      300),
+  EncodePerfTestVideo("thaloundeskmtg_640_480_30.yuv", 640, 480, 200, 300),
+  EncodePerfTestVideo("niklas_1280_720_30.yuv", 1280, 720, 600, 470),
+};
+
+const int kEncodePerfTestSpeeds[] = { 5, 6, 7, 8 };
+const int kEncodePerfTestThreads[] = { 1, 2, 4 };
+
+#define NELEMENTS(x) (sizeof((x)) / sizeof((x)[0]))
+
+class AV1EncodePerfTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWithParam<libaom_test::TestMode> {
+ protected:
+  AV1EncodePerfTest()
+      : EncoderTest(GET_PARAM(0)), min_psnr_(kMaxPsnr), nframes_(0),
+        encoding_mode_(GET_PARAM(1)), speed_(0), threads_(1) {}
+
+  virtual ~AV1EncodePerfTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+
+    cfg_.g_lag_in_frames = 0;
+    cfg_.rc_min_quantizer = 2;
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_dropframe_thresh = 0;
+    cfg_.rc_undershoot_pct = 50;
+    cfg_.rc_overshoot_pct = 50;
+    cfg_.rc_buf_sz = 1000;
+    cfg_.rc_buf_initial_sz = 500;
+    cfg_.rc_buf_optimal_sz = 600;
+    cfg_.rc_resize_allowed = 0;
+    cfg_.rc_end_usage = AOM_CBR;
+    cfg_.g_error_resilient = 1;
+    cfg_.g_threads = threads_;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      const int log2_tile_columns = 3;
+      encoder->Control(AOME_SET_CPUUSED, speed_);
+      encoder->Control(AV1E_SET_TILE_COLUMNS, log2_tile_columns);
+      encoder->Control(AV1E_SET_FRAME_PARALLEL_DECODING, 1);
+      encoder->Control(AOME_SET_ENABLEAUTOALTREF, 0);
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    min_psnr_ = kMaxPsnr;
+    nframes_ = 0;
+  }
+
+  virtual void PSNRPktHook(const aom_codec_cx_pkt_t *pkt) {
+    if (pkt->data.psnr.psnr[0] < min_psnr_) {
+      min_psnr_ = pkt->data.psnr.psnr[0];
+    }
+  }
+
+  // for performance reasons don't decode
+  virtual bool DoDecode() { return 0; }
+
+  double min_psnr() const { return min_psnr_; }
+
+  void set_speed(unsigned int speed) { speed_ = speed; }
+
+  void set_threads(unsigned int threads) { threads_ = threads; }
+
+ private:
+  double min_psnr_;
+  unsigned int nframes_;
+  libaom_test::TestMode encoding_mode_;
+  unsigned speed_;
+  unsigned int threads_;
+};
+
+TEST_P(AV1EncodePerfTest, PerfTest) {
+  for (size_t i = 0; i < NELEMENTS(kAV1EncodePerfTestVectors); ++i) {
+    for (size_t j = 0; j < NELEMENTS(kEncodePerfTestSpeeds); ++j) {
+      for (size_t k = 0; k < NELEMENTS(kEncodePerfTestThreads); ++k) {
+        if (kAV1EncodePerfTestVectors[i].width < 512 &&
+            kEncodePerfTestThreads[k] > 1)
+          continue;
+        else if (kAV1EncodePerfTestVectors[i].width < 1024 &&
+                 kEncodePerfTestThreads[k] > 2)
+          continue;
+
+        set_threads(kEncodePerfTestThreads[k]);
+        SetUp();
+
+        const aom_rational timebase = { 33333333, 1000000000 };
+        cfg_.g_timebase = timebase;
+        cfg_.rc_target_bitrate = kAV1EncodePerfTestVectors[i].bitrate;
+
+        init_flags_ = AOM_CODEC_USE_PSNR;
+
+        const unsigned frames = kAV1EncodePerfTestVectors[i].frames;
+        const char *video_name = kAV1EncodePerfTestVectors[i].name;
+        libaom_test::I420VideoSource video(
+            video_name, kAV1EncodePerfTestVectors[i].width,
+            kAV1EncodePerfTestVectors[i].height, timebase.den, timebase.num, 0,
+            kAV1EncodePerfTestVectors[i].frames);
+        set_speed(kEncodePerfTestSpeeds[j]);
+
+        aom_usec_timer t;
+        aom_usec_timer_start(&t);
+
+        ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+        aom_usec_timer_mark(&t);
+        const double elapsed_secs = aom_usec_timer_elapsed(&t) / kUsecsInSec;
+        const double fps = frames / elapsed_secs;
+        const double minimum_psnr = min_psnr();
+        std::string display_name(video_name);
+        if (kEncodePerfTestThreads[k] > 1) {
+          char thread_count[32];
+          snprintf(thread_count, sizeof(thread_count), "_t-%d",
+                   kEncodePerfTestThreads[k]);
+          display_name += thread_count;
+        }
+
+        printf("{\n");
+        printf("\t\"type\" : \"encode_perf_test\",\n");
+        printf("\t\"version\" : \"%s\",\n", VERSION_STRING_NOSP);
+        printf("\t\"videoName\" : \"%s\",\n", display_name.c_str());
+        printf("\t\"encodeTimeSecs\" : %f,\n", elapsed_secs);
+        printf("\t\"totalFrames\" : %u,\n", frames);
+        printf("\t\"framesPerSecond\" : %f,\n", fps);
+        printf("\t\"minPsnr\" : %f,\n", minimum_psnr);
+        printf("\t\"speed\" : %d,\n", kEncodePerfTestSpeeds[j]);
+        printf("\t\"threads\" : %d\n", kEncodePerfTestThreads[k]);
+        printf("}\n");
+      }
+    }
+  }
+}
+
+AV1_INSTANTIATE_TEST_CASE(AV1EncodePerfTest,
+                          ::testing::Values(::libaom_test::kRealTime));
+}  // namespace
diff --git a/third_party/aom/test/encode_test_driver.cc b/third_party/aom/test/encode_test_driver.cc
new file mode 100644
index 0000000000..80f155ab26
--- /dev/null
+++ b/third_party/aom/test/encode_test_driver.cc
@@ -0,0 +1,323 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <string>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "aom_ports/mem.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/encode_test_driver.h"
+#include "test/register_state_check.h"
+#include "test/video_source.h"
+
+namespace libaom_test {
+void Encoder::InitEncoder(VideoSource *video) {
+  aom_codec_err_t res;
+  const aom_image_t *img = video->img();
+
+  if (video->img() && !encoder_.priv) {
+    cfg_.g_w = img->d_w;
+    cfg_.g_h = img->d_h;
+    cfg_.g_timebase = video->timebase();
+    cfg_.rc_twopass_stats_in = stats_->buf();
+
+    res = aom_codec_enc_init(&encoder_, CodecInterface(), &cfg_, init_flags_);
+    ASSERT_EQ(AOM_CODEC_OK, res) << EncoderError();
+
+#if CONFIG_AV1_ENCODER
+    if (CodecInterface() == &aom_codec_av1_cx_algo) {
+// Default to 1 tile column for AV1. With CONFIG_EXT_TILE, the
+// default is already the largest possible tile size
+#if !CONFIG_EXT_TILE
+      const int log2_tile_columns = 0;
+      res = aom_codec_control_(&encoder_, AV1E_SET_TILE_COLUMNS,
+                               log2_tile_columns);
+      ASSERT_EQ(AOM_CODEC_OK, res) << EncoderError();
+#endif  // !CONFIG_EXT_TILE
+    } else
+#endif
+    {
+    }
+  }
+}
+
+void Encoder::EncodeFrame(VideoSource *video, const unsigned long frame_flags) {
+  if (video->img())
+    EncodeFrameInternal(*video, frame_flags);
+  else
+    Flush();
+
+  // Handle twopass stats
+  CxDataIterator iter = GetCxData();
+
+  while (const aom_codec_cx_pkt_t *pkt = iter.Next()) {
+    if (pkt->kind != AOM_CODEC_STATS_PKT) continue;
+
+    stats_->Append(*pkt);
+  }
+}
+
+void Encoder::EncodeFrameInternal(const VideoSource &video,
+                                  const unsigned long frame_flags) {
+  aom_codec_err_t res;
+  const aom_image_t *img = video.img();
+
+  // Handle frame resizing
+  if (cfg_.g_w != img->d_w || cfg_.g_h != img->d_h) {
+    cfg_.g_w = img->d_w;
+    cfg_.g_h = img->d_h;
+    res = aom_codec_enc_config_set(&encoder_, &cfg_);
+    ASSERT_EQ(AOM_CODEC_OK, res) << EncoderError();
+  }
+
+  // Encode the frame
+  API_REGISTER_STATE_CHECK(res = aom_codec_encode(&encoder_, img, video.pts(),
+                                                  video.duration(), frame_flags,
+                                                  deadline_));
+  ASSERT_EQ(AOM_CODEC_OK, res) << EncoderError();
+}
+
+void Encoder::Flush() {
+  const aom_codec_err_t res =
+      aom_codec_encode(&encoder_, NULL, 0, 0, 0, deadline_);
+  if (!encoder_.priv)
+    ASSERT_EQ(AOM_CODEC_ERROR, res) << EncoderError();
+  else
+    ASSERT_EQ(AOM_CODEC_OK, res) << EncoderError();
+}
+
+void EncoderTest::InitializeConfig() {
+  const aom_codec_err_t res = codec_->DefaultEncoderConfig(&cfg_, 0);
+  dec_cfg_ = aom_codec_dec_cfg_t();
+  ASSERT_EQ(AOM_CODEC_OK, res);
+}
+
+void EncoderTest::SetMode(TestMode mode) {
+  switch (mode) {
+    case kOnePassGood:
+    case kTwoPassGood: deadline_ = AOM_DL_GOOD_QUALITY; break;
+    case kRealTime:
+      deadline_ = AOM_DL_GOOD_QUALITY;
+      cfg_.g_lag_in_frames = 0;
+      break;
+    default: ASSERT_TRUE(false) << "Unexpected mode " << mode;
+  }
+  mode_ = mode;
+  if (mode == kTwoPassGood)
+    passes_ = 2;
+  else
+    passes_ = 1;
+}
+
+static bool compare_plane(const uint8_t *const buf1, int stride1,
+                          const uint8_t *const buf2, int stride2, int w, int h,
+                          int *const mismatch_row, int *const mismatch_col,
+                          int *const mismatch_pix1, int *const mismatch_pix2) {
+  int r, c;
+
+  for (r = 0; r < h; ++r) {
+    for (c = 0; c < w; ++c) {
+      const int pix1 = buf1[r * stride1 + c];
+      const int pix2 = buf2[r * stride2 + c];
+
+      if (pix1 != pix2) {
+        if (mismatch_row != NULL) *mismatch_row = r;
+        if (mismatch_col != NULL) *mismatch_col = c;
+        if (mismatch_pix1 != NULL) *mismatch_pix1 = pix1;
+        if (mismatch_pix2 != NULL) *mismatch_pix2 = pix2;
+        return false;
+      }
+    }
+  }
+
+  return true;
+}
+
+// The function should return "true" most of the time, therefore no early
+// break-out is implemented within the match checking process.
+static bool compare_img(const aom_image_t *img1, const aom_image_t *img2,
+                        int *const mismatch_row, int *const mismatch_col,
+                        int *const mismatch_plane, int *const mismatch_pix1,
+                        int *const mismatch_pix2) {
+  const unsigned int w_y = img1->d_w;
+  const unsigned int h_y = img1->d_h;
+  const unsigned int w_uv = ROUND_POWER_OF_TWO(w_y, img1->x_chroma_shift);
+  const unsigned int h_uv = ROUND_POWER_OF_TWO(h_y, img1->y_chroma_shift);
+
+  if (img1->fmt != img2->fmt || img1->cs != img2->cs ||
+      img1->d_w != img2->d_w || img1->d_h != img2->d_h) {
+    if (mismatch_row != NULL) *mismatch_row = -1;
+    if (mismatch_col != NULL) *mismatch_col = -1;
+    return false;
+  }
+
+  if (!compare_plane(img1->planes[AOM_PLANE_Y], img1->stride[AOM_PLANE_Y],
+                     img2->planes[AOM_PLANE_Y], img2->stride[AOM_PLANE_Y], w_y,
+                     h_y, mismatch_row, mismatch_col, mismatch_pix1,
+                     mismatch_pix2)) {
+    if (mismatch_plane != NULL) *mismatch_plane = AOM_PLANE_Y;
+    return false;
+  }
+
+  if (!compare_plane(img1->planes[AOM_PLANE_U], img1->stride[AOM_PLANE_U],
+                     img2->planes[AOM_PLANE_U], img2->stride[AOM_PLANE_U], w_uv,
+                     h_uv, mismatch_row, mismatch_col, mismatch_pix1,
+                     mismatch_pix2)) {
+    if (mismatch_plane != NULL) *mismatch_plane = AOM_PLANE_U;
+    return false;
+  }
+
+  if (!compare_plane(img1->planes[AOM_PLANE_V], img1->stride[AOM_PLANE_V],
+                     img2->planes[AOM_PLANE_V], img2->stride[AOM_PLANE_V], w_uv,
+                     h_uv, mismatch_row, mismatch_col, mismatch_pix1,
+                     mismatch_pix2)) {
+    if (mismatch_plane != NULL) *mismatch_plane = AOM_PLANE_U;
+    return false;
+  }
+
+  return true;
+}
+
+void EncoderTest::MismatchHook(const aom_image_t *img_enc,
+                               const aom_image_t *img_dec) {
+  int mismatch_row = 0;
+  int mismatch_col = 0;
+  int mismatch_plane = 0;
+  int mismatch_pix_enc = 0;
+  int mismatch_pix_dec = 0;
+
+  ASSERT_FALSE(compare_img(img_enc, img_dec, &mismatch_row, &mismatch_col,
+                           &mismatch_plane, &mismatch_pix_enc,
+                           &mismatch_pix_dec));
+
+  GTEST_FAIL() << "Encode/Decode mismatch found:" << std::endl
+               << "  pixel value enc/dec: " << mismatch_pix_enc << "/"
+               << mismatch_pix_dec << std::endl
+               << "                plane: " << mismatch_plane << std::endl
+               << "              row/col: " << mismatch_row << "/"
+               << mismatch_col << std::endl;
+}
+
+void EncoderTest::RunLoop(VideoSource *video) {
+  aom_codec_dec_cfg_t dec_cfg = aom_codec_dec_cfg_t();
+
+  stats_.Reset();
+
+  ASSERT_TRUE(passes_ == 1 || passes_ == 2);
+  for (unsigned int pass = 0; pass < passes_; pass++) {
+    last_pts_ = 0;
+
+    if (passes_ == 1)
+      cfg_.g_pass = AOM_RC_ONE_PASS;
+    else if (pass == 0)
+      cfg_.g_pass = AOM_RC_FIRST_PASS;
+    else
+      cfg_.g_pass = AOM_RC_LAST_PASS;
+
+    BeginPassHook(pass);
+    testing::internal::scoped_ptr<Encoder> encoder(
+        codec_->CreateEncoder(cfg_, deadline_, init_flags_, &stats_));
+    ASSERT_TRUE(encoder.get() != NULL);
+
+    ASSERT_NO_FATAL_FAILURE(video->Begin());
+    encoder->InitEncoder(video);
+
+    if (mode_ == kRealTime) {
+      encoder->Control(AOME_SET_ENABLEAUTOALTREF, 0);
+    }
+
+    ASSERT_FALSE(::testing::Test::HasFatalFailure());
+
+    unsigned long dec_init_flags = 0;  // NOLINT
+    // Use fragment decoder if encoder outputs partitions.
+    // NOTE: fragment decoder and partition encoder are only supported by VP8.
+    if (init_flags_ & AOM_CODEC_USE_OUTPUT_PARTITION)
+      dec_init_flags |= AOM_CODEC_USE_INPUT_FRAGMENTS;
+    testing::internal::scoped_ptr<Decoder> decoder(
+        codec_->CreateDecoder(dec_cfg, dec_init_flags));
+#if CONFIG_AV1 && CONFIG_EXT_TILE
+    if (decoder->IsAV1()) {
+      // Set dec_cfg.tile_row = -1 and dec_cfg.tile_col = -1 so that the whole
+      // frame is decoded.
+      decoder->Control(AV1_SET_DECODE_TILE_ROW, -1);
+      decoder->Control(AV1_SET_DECODE_TILE_COL, -1);
+    }
+#endif
+
+    bool again;
+    for (again = true; again; video->Next()) {
+      again = (video->img() != NULL);
+
+      PreEncodeFrameHook(video);
+      PreEncodeFrameHook(video, encoder.get());
+      encoder->EncodeFrame(video, frame_flags_);
+
+      CxDataIterator iter = encoder->GetCxData();
+
+      bool has_cxdata = false;
+      bool has_dxdata = false;
+      while (const aom_codec_cx_pkt_t *pkt = iter.Next()) {
+        pkt = MutateEncoderOutputHook(pkt);
+        again = true;
+        switch (pkt->kind) {
+          case AOM_CODEC_CX_FRAME_PKT:
+            has_cxdata = true;
+            if (decoder.get() != NULL && DoDecode()) {
+              aom_codec_err_t res_dec = decoder->DecodeFrame(
+                  (const uint8_t *)pkt->data.frame.buf, pkt->data.frame.sz);
+
+              if (!HandleDecodeResult(res_dec, decoder.get())) break;
+
+              has_dxdata = true;
+            }
+            ASSERT_GE(pkt->data.frame.pts, last_pts_);
+            last_pts_ = pkt->data.frame.pts;
+            FramePktHook(pkt);
+            break;
+
+          case AOM_CODEC_PSNR_PKT: PSNRPktHook(pkt); break;
+
+          default: break;
+        }
+      }
+
+      // Flush the decoder when there are no more fragments.
+      if ((init_flags_ & AOM_CODEC_USE_OUTPUT_PARTITION) && has_dxdata) {
+        const aom_codec_err_t res_dec = decoder->DecodeFrame(NULL, 0);
+        if (!HandleDecodeResult(res_dec, decoder.get())) break;
+      }
+
+      if (has_dxdata && has_cxdata) {
+        const aom_image_t *img_enc = encoder->GetPreviewFrame();
+        DxDataIterator dec_iter = decoder->GetDxData();
+        const aom_image_t *img_dec = dec_iter.Next();
+        if (img_enc && img_dec) {
+          const bool res =
+              compare_img(img_enc, img_dec, NULL, NULL, NULL, NULL, NULL);
+          if (!res) {  // Mismatch
+            MismatchHook(img_enc, img_dec);
+          }
+        }
+        if (img_dec) DecompressedFrameHook(*img_dec, video->pts());
+      }
+      if (!Continue()) break;
+    }
+
+    EndPassHook();
+
+    if (!Continue()) break;
+  }
+}
+
+}  // namespace libaom_test
diff --git a/third_party/aom/test/encode_test_driver.h b/third_party/aom/test/encode_test_driver.h
new file mode 100644
index 0000000000..91027b4f6d
--- /dev/null
+++ b/third_party/aom/test/encode_test_driver.h
@@ -0,0 +1,247 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef TEST_ENCODE_TEST_DRIVER_H_
+#define TEST_ENCODE_TEST_DRIVER_H_
+
+#include <string>
+#include <vector>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#if CONFIG_AV1_ENCODER
+#include "aom/aomcx.h"
+#endif
+#include "aom/aom_encoder.h"
+
+namespace libaom_test {
+
+class CodecFactory;
+class VideoSource;
+
+enum TestMode { kRealTime, kOnePassGood, kTwoPassGood };
+#define ALL_TEST_MODES                                                     \
+  ::testing::Values(::libaom_test::kRealTime, ::libaom_test::kOnePassGood, \
+                    ::libaom_test::kTwoPassGood)
+
+#define ONE_PASS_TEST_MODES \
+  ::testing::Values(::libaom_test::kRealTime, ::libaom_test::kOnePassGood)
+
+#define TWO_PASS_TEST_MODES ::testing::Values(::libaom_test::kTwoPassGood)
+
+// Provides an object to handle the libaom get_cx_data() iteration pattern
+class CxDataIterator {
+ public:
+  explicit CxDataIterator(aom_codec_ctx_t *encoder)
+      : encoder_(encoder), iter_(NULL) {}
+
+  const aom_codec_cx_pkt_t *Next() {
+    return aom_codec_get_cx_data(encoder_, &iter_);
+  }
+
+ private:
+  aom_codec_ctx_t *encoder_;
+  aom_codec_iter_t iter_;
+};
+
+// Implements an in-memory store for libaom twopass statistics
+class TwopassStatsStore {
+ public:
+  void Append(const aom_codec_cx_pkt_t &pkt) {
+    buffer_.append(reinterpret_cast<char *>(pkt.data.twopass_stats.buf),
+                   pkt.data.twopass_stats.sz);
+  }
+
+  aom_fixed_buf_t buf() {
+    const aom_fixed_buf_t buf = { &buffer_[0], buffer_.size() };
+    return buf;
+  }
+
+  void Reset() { buffer_.clear(); }
+
+ protected:
+  std::string buffer_;
+};
+
+// Provides a simplified interface to manage one video encoding pass, given
+// a configuration and video source.
+//
+// TODO(jkoleszar): The exact services it provides and the appropriate
+// level of abstraction will be fleshed out as more tests are written.
+class Encoder {
+ public:
+  Encoder(aom_codec_enc_cfg_t cfg, unsigned long deadline,
+          const unsigned long init_flags, TwopassStatsStore *stats)
+      : cfg_(cfg), deadline_(deadline), init_flags_(init_flags), stats_(stats) {
+    memset(&encoder_, 0, sizeof(encoder_));
+  }
+
+  virtual ~Encoder() { aom_codec_destroy(&encoder_); }
+
+  CxDataIterator GetCxData() { return CxDataIterator(&encoder_); }
+
+  void InitEncoder(VideoSource *video);
+
+  const aom_image_t *GetPreviewFrame() {
+    return aom_codec_get_preview_frame(&encoder_);
+  }
+  // This is a thin wrapper around aom_codec_encode(), so refer to
+  // aom_encoder.h for its semantics.
+  void EncodeFrame(VideoSource *video, const unsigned long frame_flags);
+
+  // Convenience wrapper for EncodeFrame()
+  void EncodeFrame(VideoSource *video) { EncodeFrame(video, 0); }
+
+  void Control(int ctrl_id, int arg) {
+    const aom_codec_err_t res = aom_codec_control_(&encoder_, ctrl_id, arg);
+    ASSERT_EQ(AOM_CODEC_OK, res) << EncoderError();
+  }
+
+  void Control(int ctrl_id, int *arg) {
+    const aom_codec_err_t res = aom_codec_control_(&encoder_, ctrl_id, arg);
+    ASSERT_EQ(AOM_CODEC_OK, res) << EncoderError();
+  }
+
+  void Control(int ctrl_id, struct aom_scaling_mode *arg) {
+    const aom_codec_err_t res = aom_codec_control_(&encoder_, ctrl_id, arg);
+    ASSERT_EQ(AOM_CODEC_OK, res) << EncoderError();
+  }
+
+#if CONFIG_AV1_ENCODER
+  void Control(int ctrl_id, aom_active_map_t *arg) {
+    const aom_codec_err_t res = aom_codec_control_(&encoder_, ctrl_id, arg);
+    ASSERT_EQ(AOM_CODEC_OK, res) << EncoderError();
+  }
+#endif
+
+  void Config(const aom_codec_enc_cfg_t *cfg) {
+    const aom_codec_err_t res = aom_codec_enc_config_set(&encoder_, cfg);
+    ASSERT_EQ(AOM_CODEC_OK, res) << EncoderError();
+    cfg_ = *cfg;
+  }
+
+  void set_deadline(unsigned long deadline) { deadline_ = deadline; }
+
+ protected:
+  virtual aom_codec_iface_t *CodecInterface() const = 0;
+
+  const char *EncoderError() {
+    const char *detail = aom_codec_error_detail(&encoder_);
+    return detail ? detail : aom_codec_error(&encoder_);
+  }
+
+  // Encode an image
+  void EncodeFrameInternal(const VideoSource &video,
+                           const unsigned long frame_flags);
+
+  // Flush the encoder on EOS
+  void Flush();
+
+  aom_codec_ctx_t encoder_;
+  aom_codec_enc_cfg_t cfg_;
+  unsigned long deadline_;
+  unsigned long init_flags_;
+  TwopassStatsStore *stats_;
+};
+
+// Common test functionality for all Encoder tests.
+//
+// This class is a mixin which provides the main loop common to all
+// encoder tests. It provides hooks which can be overridden by subclasses
+// to implement each test's specific behavior, while centralizing the bulk
+// of the boilerplate. Note that it doesn't inherit the gtest testing
+// classes directly, so that tests can be parameterized differently.
+class EncoderTest {
+ protected:
+  explicit EncoderTest(const CodecFactory *codec)
+      : codec_(codec), abort_(false), init_flags_(0), frame_flags_(0),
+        last_pts_(0), mode_(kRealTime) {
+    // Default to 1 thread.
+    cfg_.g_threads = 1;
+  }
+
+  virtual ~EncoderTest() {}
+
+  // Initialize the cfg_ member with the default configuration.
+  void InitializeConfig();
+
+  // Map the TestMode enum to the deadline_ and passes_ variables.
+  void SetMode(TestMode mode);
+
+  // Set encoder flag.
+  void set_init_flags(unsigned long flag) {  // NOLINT(runtime/int)
+    init_flags_ = flag;
+  }
+
+  // Main loop
+  virtual void RunLoop(VideoSource *video);
+
+  // Hook to be called at the beginning of a pass.
+  virtual void BeginPassHook(unsigned int /*pass*/) {}
+
+  // Hook to be called at the end of a pass.
+  virtual void EndPassHook() {}
+
+  // Hook to be called before encoding a frame.
+  virtual void PreEncodeFrameHook(VideoSource * /*video*/) {}
+  virtual void PreEncodeFrameHook(VideoSource * /*video*/,
+                                  Encoder * /*encoder*/) {}
+
+  // Hook to be called on every compressed data packet.
+  virtual void FramePktHook(const aom_codec_cx_pkt_t * /*pkt*/) {}
+
+  // Hook to be called on every PSNR packet.
+  virtual void PSNRPktHook(const aom_codec_cx_pkt_t * /*pkt*/) {}
+
+  // Hook to determine whether the encode loop should continue.
+  virtual bool Continue() const {
+    return !(::testing::Test::HasFatalFailure() || abort_);
+  }
+
+  const CodecFactory *codec_;
+  // Hook to determine whether to decode frame after encoding
+  virtual bool DoDecode() const { return 1; }
+
+  // Hook to handle encode/decode mismatch
+  virtual void MismatchHook(const aom_image_t *img1, const aom_image_t *img2);
+
+  // Hook to be called on every decompressed frame.
+  virtual void DecompressedFrameHook(const aom_image_t & /*img*/,
+                                     aom_codec_pts_t /*pts*/) {}
+
+  // Hook to be called to handle decode result. Return true to continue.
+  virtual bool HandleDecodeResult(const aom_codec_err_t res_dec,
+                                  Decoder *decoder) {
+    EXPECT_EQ(AOM_CODEC_OK, res_dec) << decoder->DecodeError();
+    return AOM_CODEC_OK == res_dec;
+  }
+
+  // Hook that can modify the encoder's output data
+  virtual const aom_codec_cx_pkt_t *MutateEncoderOutputHook(
+      const aom_codec_cx_pkt_t *pkt) {
+    return pkt;
+  }
+
+  bool abort_;
+  aom_codec_enc_cfg_t cfg_;
+  aom_codec_dec_cfg_t dec_cfg_;
+  unsigned int passes_;
+  unsigned long deadline_;
+  TwopassStatsStore stats_;
+  unsigned long init_flags_;
+  unsigned long frame_flags_;
+  aom_codec_pts_t last_pts_;
+  TestMode mode_;
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_ENCODE_TEST_DRIVER_H_
diff --git a/third_party/aom/test/encoder_parms_get_to_decoder.cc b/third_party/aom/test/encoder_parms_get_to_decoder.cc
new file mode 100644
index 0000000000..ca6a24ebea
--- /dev/null
+++ b/third_party/aom/test/encoder_parms_get_to_decoder.cc
@@ -0,0 +1,147 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "av1/av1_dx_iface.c"
+
+namespace {
+
+const int kCpuUsed = 2;
+
+struct EncodePerfTestVideo {
+  const char *name;
+  uint32_t width;
+  uint32_t height;
+  uint32_t bitrate;
+  int frames;
+};
+
+const EncodePerfTestVideo kAV1EncodePerfTestVectors[] = {
+  { "niklas_1280_720_30.y4m", 1280, 720, 600, 10 },
+};
+
+struct EncodeParameters {
+  int32_t tile_rows;
+  int32_t tile_cols;
+  int32_t lossless;
+  int32_t error_resilient;
+  int32_t frame_parallel;
+  aom_color_range_t color_range;
+  aom_color_space_t cs;
+  int render_size[2];
+  // TODO(JBB): quantizers / bitrate
+};
+
+const EncodeParameters kAV1EncodeParameterSet[] = {
+  { 0, 0, 0, 1, 0, AOM_CR_STUDIO_RANGE, AOM_CS_BT_601, { 0, 0 } },
+  { 0, 0, 0, 0, 0, AOM_CR_FULL_RANGE, AOM_CS_BT_709, { 0, 0 } },
+  { 0, 0, 1, 0, 0, AOM_CR_FULL_RANGE, AOM_CS_BT_2020, { 0, 0 } },
+  { 0, 2, 0, 0, 1, AOM_CR_STUDIO_RANGE, AOM_CS_UNKNOWN, { 640, 480 } },
+  // TODO(JBB): Test profiles (requires more work).
+};
+
+class AvxEncoderParmsGetToDecoder
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<EncodeParameters,
+                                                 EncodePerfTestVideo> {
+ protected:
+  AvxEncoderParmsGetToDecoder()
+      : EncoderTest(GET_PARAM(0)), encode_parms(GET_PARAM(1)) {}
+
+  virtual ~AvxEncoderParmsGetToDecoder() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(::libaom_test::kTwoPassGood);
+    cfg_.g_lag_in_frames = 25;
+    cfg_.g_error_resilient = encode_parms.error_resilient;
+    dec_cfg_.threads = 4;
+    test_video_ = GET_PARAM(2);
+    cfg_.rc_target_bitrate = test_video_.bitrate;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AV1E_SET_COLOR_SPACE, encode_parms.cs);
+      encoder->Control(AV1E_SET_COLOR_RANGE, encode_parms.color_range);
+      encoder->Control(AV1E_SET_LOSSLESS, encode_parms.lossless);
+      encoder->Control(AV1E_SET_FRAME_PARALLEL_DECODING,
+                       encode_parms.frame_parallel);
+      encoder->Control(AV1E_SET_TILE_ROWS, encode_parms.tile_rows);
+      encoder->Control(AV1E_SET_TILE_COLUMNS, encode_parms.tile_cols);
+      encoder->Control(AOME_SET_CPUUSED, kCpuUsed);
+      encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+      encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
+      encoder->Control(AOME_SET_ARNR_STRENGTH, 5);
+      if (encode_parms.render_size[0] > 0 && encode_parms.render_size[1] > 0)
+        encoder->Control(AV1E_SET_RENDER_SIZE, encode_parms.render_size);
+    }
+  }
+
+  virtual bool HandleDecodeResult(const aom_codec_err_t res_dec,
+                                  libaom_test::Decoder *decoder) {
+    aom_codec_ctx_t *const av1_decoder = decoder->GetDecoder();
+    aom_codec_alg_priv_t *const priv =
+        reinterpret_cast<aom_codec_alg_priv_t *>(av1_decoder->priv);
+    FrameWorkerData *const worker_data =
+        reinterpret_cast<FrameWorkerData *>(priv->frame_workers[0].data1);
+    AV1_COMMON *const common = &worker_data->pbi->common;
+
+    if (encode_parms.lossless) {
+      EXPECT_EQ(0, common->base_qindex);
+      EXPECT_EQ(0, common->y_dc_delta_q);
+      EXPECT_EQ(0, common->uv_dc_delta_q);
+      EXPECT_EQ(0, common->uv_ac_delta_q);
+      EXPECT_EQ(ONLY_4X4, common->tx_mode);
+    }
+    EXPECT_EQ(encode_parms.error_resilient, common->error_resilient_mode);
+    if (encode_parms.error_resilient) {
+      EXPECT_EQ(0, common->use_prev_frame_mvs);
+    }
+    EXPECT_EQ(encode_parms.color_range, common->color_range);
+    EXPECT_EQ(encode_parms.cs, common->color_space);
+    if (encode_parms.render_size[0] > 0 && encode_parms.render_size[1] > 0) {
+      EXPECT_EQ(encode_parms.render_size[0], common->render_width);
+      EXPECT_EQ(encode_parms.render_size[1], common->render_height);
+    }
+    EXPECT_EQ(encode_parms.tile_cols, common->log2_tile_cols);
+    EXPECT_EQ(encode_parms.tile_rows, common->log2_tile_rows);
+
+    EXPECT_EQ(AOM_CODEC_OK, res_dec) << decoder->DecodeError();
+    return AOM_CODEC_OK == res_dec;
+  }
+
+  EncodePerfTestVideo test_video_;
+
+ private:
+  EncodeParameters encode_parms;
+};
+
+TEST_P(AvxEncoderParmsGetToDecoder, BitstreamParms) {
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  testing::internal::scoped_ptr<libaom_test::VideoSource> video(
+      new libaom_test::Y4mVideoSource(test_video_.name, 0, test_video_.frames));
+  ASSERT_TRUE(video.get() != NULL);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video.get()));
+}
+
+AV1_INSTANTIATE_TEST_CASE(AvxEncoderParmsGetToDecoder,
+                          ::testing::ValuesIn(kAV1EncodeParameterSet),
+                          ::testing::ValuesIn(kAV1EncodePerfTestVectors));
+}  // namespace
diff --git a/third_party/aom/test/end_to_end_test.cc b/third_party/aom/test/end_to_end_test.cc
new file mode 100644
index 0000000000..0c8cbe274c
--- /dev/null
+++ b/third_party/aom/test/end_to_end_test.cc
@@ -0,0 +1,226 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+#include "test/yuv_video_source.h"
+
+namespace {
+
+const unsigned int kWidth = 160;
+const unsigned int kHeight = 90;
+const unsigned int kFramerate = 50;
+const unsigned int kFrames = 10;
+const int kBitrate = 500;
+// List of psnr thresholds for speed settings 0-7 and 5 encoding modes
+const double kPsnrThreshold[][5] = {
+// Note:
+// AV1 HBD average PSNR is slightly lower than AV1.
+// We make two cases here to enable the testing and
+// guard picture quality.
+#if CONFIG_AV1_ENCODER && CONFIG_HIGHBITDEPTH
+  { 36.0, 37.0, 37.0, 37.0, 37.0 }, { 31.0, 36.0, 36.0, 36.0, 36.0 },
+  { 31.0, 35.0, 35.0, 35.0, 35.0 }, { 31.0, 34.0, 34.0, 34.0, 34.0 },
+  { 31.0, 33.0, 33.0, 33.0, 33.0 }, { 31.0, 32.0, 32.0, 32.0, 32.0 },
+  { 30.0, 31.0, 31.0, 31.0, 31.0 }, { 29.0, 30.0, 30.0, 30.0, 30.0 },
+#else
+  { 36.0, 37.0, 37.0, 37.0, 37.0 }, { 35.0, 36.0, 36.0, 36.0, 36.0 },
+  { 34.0, 35.0, 35.0, 35.0, 35.0 }, { 33.0, 34.0, 34.0, 34.0, 34.0 },
+  { 32.0, 33.0, 33.0, 33.0, 33.0 }, { 31.0, 32.0, 32.0, 32.0, 32.0 },
+  { 30.0, 31.0, 31.0, 31.0, 31.0 }, { 29.0, 30.0, 30.0, 30.0, 30.0 },
+#endif  // CONFIG_HIGHBITDEPTH && CONFIG_AV1_ENCODER
+};
+
+typedef struct {
+  const char *filename;
+  unsigned int input_bit_depth;
+  aom_img_fmt fmt;
+  aom_bit_depth_t bit_depth;
+  unsigned int profile;
+} TestVideoParam;
+
+const TestVideoParam kTestVectors[] = {
+  { "park_joy_90p_8_420.y4m", 8, AOM_IMG_FMT_I420, AOM_BITS_8, 0 },
+  { "park_joy_90p_8_422.y4m", 8, AOM_IMG_FMT_I422, AOM_BITS_8, 1 },
+  { "park_joy_90p_8_444.y4m", 8, AOM_IMG_FMT_I444, AOM_BITS_8, 1 },
+  { "park_joy_90p_8_440.yuv", 8, AOM_IMG_FMT_I440, AOM_BITS_8, 1 },
+#if CONFIG_HIGHBITDEPTH
+  { "park_joy_90p_10_420.y4m", 10, AOM_IMG_FMT_I42016, AOM_BITS_10, 2 },
+  { "park_joy_90p_10_422.y4m", 10, AOM_IMG_FMT_I42216, AOM_BITS_10, 3 },
+  { "park_joy_90p_10_444.y4m", 10, AOM_IMG_FMT_I44416, AOM_BITS_10, 3 },
+  { "park_joy_90p_10_440.yuv", 10, AOM_IMG_FMT_I44016, AOM_BITS_10, 3 },
+  { "park_joy_90p_12_420.y4m", 12, AOM_IMG_FMT_I42016, AOM_BITS_12, 2 },
+  { "park_joy_90p_12_422.y4m", 12, AOM_IMG_FMT_I42216, AOM_BITS_12, 3 },
+  { "park_joy_90p_12_444.y4m", 12, AOM_IMG_FMT_I44416, AOM_BITS_12, 3 },
+  { "park_joy_90p_12_440.yuv", 12, AOM_IMG_FMT_I44016, AOM_BITS_12, 3 },
+#endif  // CONFIG_HIGHBITDEPTH
+};
+
+// Encoding modes tested
+const libaom_test::TestMode kEncodingModeVectors[] = {
+  ::libaom_test::kTwoPassGood, ::libaom_test::kOnePassGood,
+  ::libaom_test::kRealTime,
+};
+
+// Speed settings tested
+const int kCpuUsedVectors[] = { 1, 2, 3, 5, 6 };
+
+int is_extension_y4m(const char *filename) {
+  const char *dot = strrchr(filename, '.');
+  if (!dot || dot == filename)
+    return 0;
+  else
+    return !strcmp(dot, ".y4m");
+}
+
+class EndToEndTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith3Params<libaom_test::TestMode,
+                                                 TestVideoParam, int> {
+ protected:
+  EndToEndTest()
+      : EncoderTest(GET_PARAM(0)), test_video_param_(GET_PARAM(2)),
+        cpu_used_(GET_PARAM(3)), psnr_(0.0), nframes_(0),
+        encoding_mode_(GET_PARAM(1)) {}
+
+  virtual ~EndToEndTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+    if (encoding_mode_ != ::libaom_test::kRealTime) {
+      cfg_.g_lag_in_frames = 5;
+      cfg_.rc_end_usage = AOM_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = AOM_CBR;
+      cfg_.rc_buf_sz = 1000;
+      cfg_.rc_buf_initial_sz = 500;
+      cfg_.rc_buf_optimal_sz = 600;
+    }
+    dec_cfg_.threads = 4;
+  }
+
+  virtual void BeginPassHook(unsigned int) {
+    psnr_ = 0.0;
+    nframes_ = 0;
+  }
+
+  virtual void PSNRPktHook(const aom_codec_cx_pkt_t *pkt) {
+    psnr_ += pkt->data.psnr.psnr[0];
+    nframes_++;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AV1E_SET_FRAME_PARALLEL_DECODING, 1);
+      encoder->Control(AV1E_SET_TILE_COLUMNS, 4);
+      encoder->Control(AOME_SET_CPUUSED, cpu_used_);
+#if CONFIG_PALETTE
+      // Test screen coding tools at cpu_used = 1 && encoding mode is two-pass.
+      if (cpu_used_ == 1 && encoding_mode_ == ::libaom_test::kTwoPassGood)
+        encoder->Control(AV1E_SET_TUNE_CONTENT, AOM_CONTENT_SCREEN);
+      else
+        encoder->Control(AV1E_SET_TUNE_CONTENT, AOM_CONTENT_DEFAULT);
+#endif  // CONFIG_PALETTE
+      if (encoding_mode_ != ::libaom_test::kRealTime) {
+        encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(AOME_SET_ARNR_STRENGTH, 5);
+      }
+    }
+  }
+
+  double GetAveragePsnr() const {
+    if (nframes_) return psnr_ / nframes_;
+    return 0.0;
+  }
+
+  double GetPsnrThreshold() {
+    return kPsnrThreshold[cpu_used_][encoding_mode_];
+  }
+
+  TestVideoParam test_video_param_;
+  int cpu_used_;
+
+ private:
+  double psnr_;
+  unsigned int nframes_;
+  libaom_test::TestMode encoding_mode_;
+};
+
+class EndToEndTestLarge : public EndToEndTest {};
+
+TEST_P(EndToEndTestLarge, EndtoEndPSNRTest) {
+  cfg_.rc_target_bitrate = kBitrate;
+  cfg_.g_error_resilient = 0;
+  cfg_.g_profile = test_video_param_.profile;
+  cfg_.g_input_bit_depth = test_video_param_.input_bit_depth;
+  cfg_.g_bit_depth = test_video_param_.bit_depth;
+  init_flags_ = AOM_CODEC_USE_PSNR;
+  if (cfg_.g_bit_depth > 8) init_flags_ |= AOM_CODEC_USE_HIGHBITDEPTH;
+
+  testing::internal::scoped_ptr<libaom_test::VideoSource> video;
+  if (is_extension_y4m(test_video_param_.filename)) {
+    video.reset(new libaom_test::Y4mVideoSource(test_video_param_.filename, 0,
+                                                kFrames));
+  } else {
+    video.reset(new libaom_test::YUVVideoSource(
+        test_video_param_.filename, test_video_param_.fmt, kWidth, kHeight,
+        kFramerate, 1, 0, kFrames));
+  }
+  ASSERT_TRUE(video.get() != NULL);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video.get()));
+  const double psnr = GetAveragePsnr();
+  EXPECT_GT(psnr, GetPsnrThreshold());
+}
+
+TEST_P(EndToEndTest, EndtoEndPSNRTest) {
+  cfg_.rc_target_bitrate = kBitrate;
+  cfg_.g_error_resilient = 0;
+  cfg_.g_profile = test_video_param_.profile;
+  cfg_.g_input_bit_depth = test_video_param_.input_bit_depth;
+  cfg_.g_bit_depth = test_video_param_.bit_depth;
+  init_flags_ = AOM_CODEC_USE_PSNR;
+  if (cfg_.g_bit_depth > 8) init_flags_ |= AOM_CODEC_USE_HIGHBITDEPTH;
+
+  testing::internal::scoped_ptr<libaom_test::VideoSource> video;
+  if (is_extension_y4m(test_video_param_.filename)) {
+    video.reset(new libaom_test::Y4mVideoSource(test_video_param_.filename, 0,
+                                                kFrames));
+  } else {
+    video.reset(new libaom_test::YUVVideoSource(
+        test_video_param_.filename, test_video_param_.fmt, kWidth, kHeight,
+        kFramerate, 1, 0, kFrames));
+  }
+  ASSERT_TRUE(video.get() != NULL);
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video.get()));
+  const double psnr = GetAveragePsnr();
+  EXPECT_GT(psnr, GetPsnrThreshold());
+}
+
+AV1_INSTANTIATE_TEST_CASE(EndToEndTestLarge,
+                          ::testing::ValuesIn(kEncodingModeVectors),
+                          ::testing::ValuesIn(kTestVectors),
+                          ::testing::ValuesIn(kCpuUsedVectors));
+
+AV1_INSTANTIATE_TEST_CASE(EndToEndTest,
+                          ::testing::Values(kEncodingModeVectors[0]),
+                          ::testing::Values(kTestVectors[2]),  // 444
+                          ::testing::Values(kCpuUsedVectors[2]));
+}  // namespace
diff --git a/third_party/aom/test/error_block_test.cc b/third_party/aom/test/error_block_test.cc
new file mode 100644
index 0000000000..227065fa90
--- /dev/null
+++ b/third_party/aom/test/error_block_test.cc
@@ -0,0 +1,173 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <cmath>
+#include <cstdlib>
+#include <string>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./av1_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "av1/common/entropy.h"
+#include "aom/aom_codec.h"
+#include "aom/aom_integer.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+#if CONFIG_HIGHBITDEPTH
+const int kNumIterations = 1000;
+
+typedef int64_t (*ErrorBlockFunc)(const tran_low_t *coeff,
+                                  const tran_low_t *dqcoeff,
+                                  intptr_t block_size, int64_t *ssz, int bps);
+
+typedef std::tr1::tuple<ErrorBlockFunc, ErrorBlockFunc, aom_bit_depth_t>
+    ErrorBlockParam;
+
+class ErrorBlockTest : public ::testing::TestWithParam<ErrorBlockParam> {
+ public:
+  virtual ~ErrorBlockTest() {}
+  virtual void SetUp() {
+    error_block_op_ = GET_PARAM(0);
+    ref_error_block_op_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  aom_bit_depth_t bit_depth_;
+  ErrorBlockFunc error_block_op_;
+  ErrorBlockFunc ref_error_block_op_;
+};
+
+TEST_P(ErrorBlockTest, OperationCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, tran_low_t, coeff[4096]);
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[4096]);
+  int err_count_total = 0;
+  int first_failure = -1;
+  intptr_t block_size;
+  int64_t ssz;
+  int64_t ret;
+  int64_t ref_ssz;
+  int64_t ref_ret;
+  const int msb = bit_depth_ + 8 - 1;
+  for (int i = 0; i < kNumIterations; ++i) {
+    int err_count = 0;
+    block_size = 16 << (i % 9);  // All block sizes from 4x4, 8x4 ..64x64
+    for (int j = 0; j < block_size; j++) {
+      // coeff and dqcoeff will always have at least the same sign, and this
+      // can be used for optimization, so generate test input precisely.
+      if (rnd(2)) {
+        // Positive number
+        coeff[j] = rnd(1 << msb);
+        dqcoeff[j] = rnd(1 << msb);
+      } else {
+        // Negative number
+        coeff[j] = -rnd(1 << msb);
+        dqcoeff[j] = -rnd(1 << msb);
+      }
+    }
+    ref_ret =
+        ref_error_block_op_(coeff, dqcoeff, block_size, &ref_ssz, bit_depth_);
+    ASM_REGISTER_STATE_CHECK(
+        ret = error_block_op_(coeff, dqcoeff, block_size, &ssz, bit_depth_));
+    err_count += (ref_ret != ret) | (ref_ssz != ssz);
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Error Block Test, C output doesn't match optimized output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(ErrorBlockTest, ExtremeValues) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, tran_low_t, coeff[4096]);
+  DECLARE_ALIGNED(16, tran_low_t, dqcoeff[4096]);
+  int err_count_total = 0;
+  int first_failure = -1;
+  intptr_t block_size;
+  int64_t ssz;
+  int64_t ret;
+  int64_t ref_ssz;
+  int64_t ref_ret;
+  const int msb = bit_depth_ + 8 - 1;
+  int max_val = ((1 << msb) - 1);
+  for (int i = 0; i < kNumIterations; ++i) {
+    int err_count = 0;
+    int k = (i / 9) % 9;
+
+    // Change the maximum coeff value, to test different bit boundaries
+    if (k == 8 && (i % 9) == 0) {
+      max_val >>= 1;
+    }
+    block_size = 16 << (i % 9);  // All block sizes from 4x4, 8x4 ..64x64
+    for (int j = 0; j < block_size; j++) {
+      if (k < 4) {
+        // Test at positive maximum values
+        coeff[j] = k % 2 ? max_val : 0;
+        dqcoeff[j] = (k >> 1) % 2 ? max_val : 0;
+      } else if (k < 8) {
+        // Test at negative maximum values
+        coeff[j] = k % 2 ? -max_val : 0;
+        dqcoeff[j] = (k >> 1) % 2 ? -max_val : 0;
+      } else {
+        if (rnd(2)) {
+          // Positive number
+          coeff[j] = rnd(1 << 14);
+          dqcoeff[j] = rnd(1 << 14);
+        } else {
+          // Negative number
+          coeff[j] = -rnd(1 << 14);
+          dqcoeff[j] = -rnd(1 << 14);
+        }
+      }
+    }
+    ref_ret =
+        ref_error_block_op_(coeff, dqcoeff, block_size, &ref_ssz, bit_depth_);
+    ASM_REGISTER_STATE_CHECK(
+        ret = error_block_op_(coeff, dqcoeff, block_size, &ssz, bit_depth_));
+    err_count += (ref_ret != ret) | (ref_ssz != ssz);
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Error Block Test, C output doesn't match optimized output. "
+      << "First failed at test case " << first_failure;
+}
+
+#if HAVE_SSE2 || HAVE_AVX
+using std::tr1::make_tuple;
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, ErrorBlockTest,
+    ::testing::Values(make_tuple(&av1_highbd_block_error_sse2,
+                                 &av1_highbd_block_error_c, AOM_BITS_10),
+                      make_tuple(&av1_highbd_block_error_sse2,
+                                 &av1_highbd_block_error_c, AOM_BITS_12),
+                      make_tuple(&av1_highbd_block_error_sse2,
+                                 &av1_highbd_block_error_c, AOM_BITS_8)));
+#endif  // HAVE_SSE2
+
+#endif  // CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/error_resilience_test.cc b/third_party/aom/test/error_resilience_test.cc
new file mode 100644
index 0000000000..63f10012f8
--- /dev/null
+++ b/third_party/aom/test/error_resilience_test.cc
@@ -0,0 +1,235 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+const int kMaxErrorFrames = 12;
+const int kMaxDroppableFrames = 12;
+
+class ErrorResilienceTestLarge
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWithParam<libaom_test::TestMode> {
+ protected:
+  ErrorResilienceTestLarge()
+      : EncoderTest(GET_PARAM(0)), psnr_(0.0), nframes_(0), mismatch_psnr_(0.0),
+        mismatch_nframes_(0), encoding_mode_(GET_PARAM(1)) {
+    Reset();
+  }
+
+  virtual ~ErrorResilienceTestLarge() {}
+
+  void Reset() {
+    error_nframes_ = 0;
+    droppable_nframes_ = 0;
+    pattern_switch_ = 0;
+  }
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    psnr_ = 0.0;
+    nframes_ = 0;
+    mismatch_psnr_ = 0.0;
+    mismatch_nframes_ = 0;
+  }
+
+  virtual void PSNRPktHook(const aom_codec_cx_pkt_t *pkt) {
+    psnr_ += pkt->data.psnr.psnr[0];
+    nframes_++;
+  }
+
+  virtual void PreEncodeFrameHook(libaom_test::VideoSource *video) {
+    frame_flags_ &=
+        ~(AOM_EFLAG_NO_UPD_LAST | AOM_EFLAG_NO_UPD_GF | AOM_EFLAG_NO_UPD_ARF);
+    if (droppable_nframes_ > 0 &&
+        (cfg_.g_pass == AOM_RC_LAST_PASS || cfg_.g_pass == AOM_RC_ONE_PASS)) {
+      for (unsigned int i = 0; i < droppable_nframes_; ++i) {
+        if (droppable_frames_[i] == video->frame()) {
+          std::cout << "Encoding droppable frame: " << droppable_frames_[i]
+                    << "\n";
+          frame_flags_ |= (AOM_EFLAG_NO_UPD_LAST | AOM_EFLAG_NO_UPD_GF |
+                           AOM_EFLAG_NO_UPD_ARF);
+          return;
+        }
+      }
+    }
+  }
+
+  double GetAveragePsnr() const {
+    if (nframes_) return psnr_ / nframes_;
+    return 0.0;
+  }
+
+  double GetAverageMismatchPsnr() const {
+    if (mismatch_nframes_) return mismatch_psnr_ / mismatch_nframes_;
+    return 0.0;
+  }
+
+  virtual bool DoDecode() const {
+    if (error_nframes_ > 0 &&
+        (cfg_.g_pass == AOM_RC_LAST_PASS || cfg_.g_pass == AOM_RC_ONE_PASS)) {
+      for (unsigned int i = 0; i < error_nframes_; ++i) {
+        if (error_frames_[i] == nframes_ - 1) {
+          std::cout << "             Skipping decoding frame: "
+                    << error_frames_[i] << "\n";
+          return 0;
+        }
+      }
+    }
+    return 1;
+  }
+
+  virtual void MismatchHook(const aom_image_t *img1, const aom_image_t *img2) {
+    double mismatch_psnr = compute_psnr(img1, img2);
+    mismatch_psnr_ += mismatch_psnr;
+    ++mismatch_nframes_;
+    // std::cout << "Mismatch frame psnr: " << mismatch_psnr << "\n";
+    ::libaom_test::EncoderTest::MismatchHook(img1, img2);
+  }
+
+  void SetErrorFrames(int num, unsigned int *list) {
+    if (num > kMaxErrorFrames)
+      num = kMaxErrorFrames;
+    else if (num < 0)
+      num = 0;
+    error_nframes_ = num;
+    for (unsigned int i = 0; i < error_nframes_; ++i)
+      error_frames_[i] = list[i];
+  }
+
+  void SetDroppableFrames(int num, unsigned int *list) {
+    if (num > kMaxDroppableFrames)
+      num = kMaxDroppableFrames;
+    else if (num < 0)
+      num = 0;
+    droppable_nframes_ = num;
+    for (unsigned int i = 0; i < droppable_nframes_; ++i)
+      droppable_frames_[i] = list[i];
+  }
+
+  unsigned int GetMismatchFrames() { return mismatch_nframes_; }
+
+  void SetPatternSwitch(int frame_switch) { pattern_switch_ = frame_switch; }
+
+ private:
+  double psnr_;
+  unsigned int nframes_;
+  unsigned int error_nframes_;
+  unsigned int droppable_nframes_;
+  unsigned int pattern_switch_;
+  double mismatch_psnr_;
+  unsigned int mismatch_nframes_;
+  unsigned int error_frames_[kMaxErrorFrames];
+  unsigned int droppable_frames_[kMaxDroppableFrames];
+  libaom_test::TestMode encoding_mode_;
+};
+
+TEST_P(ErrorResilienceTestLarge, OnVersusOff) {
+  const aom_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.g_lag_in_frames = 10;
+
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 30);
+
+  // Error resilient mode OFF.
+  cfg_.g_error_resilient = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_resilience_off = GetAveragePsnr();
+  EXPECT_GT(psnr_resilience_off, 25.0);
+
+  // Error resilient mode ON.
+  cfg_.g_error_resilient = 1;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_resilience_on = GetAveragePsnr();
+  EXPECT_GT(psnr_resilience_on, 25.0);
+
+  // Test that turning on error resilient mode hurts by 10% at most.
+  if (psnr_resilience_off > 0.0) {
+    const double psnr_ratio = psnr_resilience_on / psnr_resilience_off;
+    EXPECT_GE(psnr_ratio, 0.9);
+    EXPECT_LE(psnr_ratio, 1.1);
+  }
+}
+
+// Check for successful decoding and no encoder/decoder mismatch
+// if we lose (i.e., drop before decoding) a set of droppable
+// frames (i.e., frames that don't update any reference buffers).
+// Check both isolated and consecutive loss.
+TEST_P(ErrorResilienceTestLarge, DropFramesWithoutRecovery) {
+  const aom_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 500;
+  // FIXME(debargha): Fix this to work for any lag.
+  // Currently this test only works for lag = 0
+  cfg_.g_lag_in_frames = 0;
+
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 40);
+
+  // Error resilient mode ON.
+  cfg_.g_error_resilient = 1;
+  cfg_.kf_mode = AOM_KF_DISABLED;
+
+  // Set an arbitrary set of error frames same as droppable frames.
+  // In addition to isolated loss/drop, add a long consecutive series
+  // (of size 9) of dropped frames.
+  unsigned int num_droppable_frames = 11;
+  unsigned int droppable_frame_list[] = { 5,  16, 22, 23, 24, 25,
+                                          26, 27, 28, 29, 30 };
+  SetDroppableFrames(num_droppable_frames, droppable_frame_list);
+  SetErrorFrames(num_droppable_frames, droppable_frame_list);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  // Test that no mismatches have been found
+  std::cout << "             Mismatch frames: " << GetMismatchFrames() << "\n";
+  EXPECT_EQ(GetMismatchFrames(), (unsigned int)0);
+
+  // Reset previously set of error/droppable frames.
+  Reset();
+
+#if 0
+  // TODO(jkoleszar): This test is disabled for the time being as too
+  // sensitive. It's not clear how to set a reasonable threshold for
+  // this behavior.
+
+  // Now set an arbitrary set of error frames that are non-droppable
+  unsigned int num_error_frames = 3;
+  unsigned int error_frame_list[] = {3, 10, 20};
+  SetErrorFrames(num_error_frames, error_frame_list);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  // Test that dropping an arbitrary set of inter frames does not hurt too much
+  // Note the Average Mismatch PSNR is the average of the PSNR between
+  // decoded frame and encoder's version of the same frame for all frames
+  // with mismatch.
+  const double psnr_resilience_mismatch = GetAverageMismatchPsnr();
+  std::cout << "             Mismatch PSNR: "
+            << psnr_resilience_mismatch << "\n";
+  EXPECT_GT(psnr_resilience_mismatch, 20.0);
+#endif
+}
+
+AV1_INSTANTIATE_TEST_CASE(ErrorResilienceTestLarge, ONE_PASS_TEST_MODES);
+}  // namespace
diff --git a/third_party/aom/test/ethread_test.cc b/third_party/aom/test/ethread_test.cc
new file mode 100644
index 0000000000..5b519f8feb
--- /dev/null
+++ b/third_party/aom/test/ethread_test.cc
@@ -0,0 +1,188 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <string>
+#include <vector>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+
+namespace {
+class AVxEncoderThreadTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<libaom_test::TestMode, int> {
+ protected:
+  AVxEncoderThreadTest()
+      : EncoderTest(GET_PARAM(0)), encoder_initialized_(false),
+        encoding_mode_(GET_PARAM(1)), set_cpu_used_(GET_PARAM(2)) {
+    init_flags_ = AOM_CODEC_USE_PSNR;
+    aom_codec_dec_cfg_t cfg = aom_codec_dec_cfg_t();
+    cfg.w = 1280;
+    cfg.h = 720;
+    decoder_ = codec_->CreateDecoder(cfg, 0);
+#if CONFIG_AV1 && CONFIG_EXT_TILE
+    if (decoder_->IsAV1()) {
+      decoder_->Control(AV1_SET_DECODE_TILE_ROW, -1);
+      decoder_->Control(AV1_SET_DECODE_TILE_COL, -1);
+    }
+#endif
+
+    size_enc_.clear();
+    md5_dec_.clear();
+    md5_enc_.clear();
+  }
+  virtual ~AVxEncoderThreadTest() { delete decoder_; }
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+
+    if (encoding_mode_ != ::libaom_test::kRealTime) {
+      cfg_.g_lag_in_frames = 3;
+      cfg_.rc_end_usage = AOM_VBR;
+      cfg_.rc_2pass_vbr_minsection_pct = 5;
+      cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = AOM_CBR;
+      cfg_.g_error_resilient = 1;
+    }
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_min_quantizer = 0;
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    encoder_initialized_ = false;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource * /*video*/,
+                                  ::libaom_test::Encoder *encoder) {
+    if (!encoder_initialized_) {
+#if CONFIG_AV1 && CONFIG_EXT_TILE
+      encoder->Control(AV1E_SET_TILE_COLUMNS, 1);
+      if (codec_ == &libaom_test::kAV1) {
+        // TODO(geza): Start using multiple tile rows when the multi-threaded
+        // encoder can handle them
+        encoder->Control(AV1E_SET_TILE_ROWS, 32);
+      } else {
+        encoder->Control(AV1E_SET_TILE_ROWS, 0);
+      }
+#else
+      // Encode 4 tile columns.
+      encoder->Control(AV1E_SET_TILE_COLUMNS, 2);
+      encoder->Control(AV1E_SET_TILE_ROWS, 0);
+#endif  // CONFIG_AV1 && CONFIG_EXT_TILE
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+      encoder->Control(AV1E_SET_TILE_LOOPFILTER, 0);
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+      encoder->Control(AOME_SET_CPUUSED, set_cpu_used_);
+      if (encoding_mode_ != ::libaom_test::kRealTime) {
+        encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(AOME_SET_ARNR_STRENGTH, 5);
+        encoder->Control(AV1E_SET_FRAME_PARALLEL_DECODING, 0);
+      } else {
+        encoder->Control(AOME_SET_ENABLEAUTOALTREF, 0);
+        encoder->Control(AV1E_SET_AQ_MODE, 3);
+      }
+      encoder_initialized_ = true;
+    }
+  }
+
+  virtual void FramePktHook(const aom_codec_cx_pkt_t *pkt) {
+    size_enc_.push_back(pkt->data.frame.sz);
+
+    ::libaom_test::MD5 md5_enc;
+    md5_enc.Add(reinterpret_cast<uint8_t *>(pkt->data.frame.buf),
+                pkt->data.frame.sz);
+    md5_enc_.push_back(md5_enc.Get());
+
+    const aom_codec_err_t res = decoder_->DecodeFrame(
+        reinterpret_cast<uint8_t *>(pkt->data.frame.buf), pkt->data.frame.sz);
+    if (res != AOM_CODEC_OK) {
+      abort_ = true;
+      ASSERT_EQ(AOM_CODEC_OK, res);
+    }
+    const aom_image_t *img = decoder_->GetDxData().Next();
+
+    if (img) {
+      ::libaom_test::MD5 md5_res;
+      md5_res.Add(img);
+      md5_dec_.push_back(md5_res.Get());
+    }
+  }
+
+  void DoTest() {
+    ::libaom_test::Y4mVideoSource video("niklas_1280_720_30.y4m", 15, 18);
+    cfg_.rc_target_bitrate = 1000;
+
+    // Encode using single thread.
+    cfg_.g_threads = 1;
+    init_flags_ = AOM_CODEC_USE_PSNR;
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    std::vector<size_t> single_thr_size_enc;
+    std::vector<std::string> single_thr_md5_enc;
+    std::vector<std::string> single_thr_md5_dec;
+    single_thr_size_enc = size_enc_;
+    single_thr_md5_enc = md5_enc_;
+    single_thr_md5_dec = md5_dec_;
+    size_enc_.clear();
+    md5_enc_.clear();
+    md5_dec_.clear();
+
+    // Encode using multiple threads.
+    cfg_.g_threads = 4;
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+    std::vector<size_t> multi_thr_size_enc;
+    std::vector<std::string> multi_thr_md5_enc;
+    std::vector<std::string> multi_thr_md5_dec;
+    multi_thr_size_enc = size_enc_;
+    multi_thr_md5_enc = md5_enc_;
+    multi_thr_md5_dec = md5_dec_;
+    size_enc_.clear();
+    md5_enc_.clear();
+    md5_dec_.clear();
+
+    // Check that the vectors are equal.
+    ASSERT_EQ(single_thr_size_enc, multi_thr_size_enc);
+    ASSERT_EQ(single_thr_md5_enc, multi_thr_md5_enc);
+    ASSERT_EQ(single_thr_md5_dec, multi_thr_md5_dec);
+  }
+
+  bool encoder_initialized_;
+  ::libaom_test::TestMode encoding_mode_;
+  int set_cpu_used_;
+  ::libaom_test::Decoder *decoder_;
+  std::vector<size_t> size_enc_;
+  std::vector<std::string> md5_enc_;
+  std::vector<std::string> md5_dec_;
+};
+
+TEST_P(AVxEncoderThreadTest, EncoderResultTest) { DoTest(); }
+
+class AVxEncoderThreadTestLarge : public AVxEncoderThreadTest {};
+
+TEST_P(AVxEncoderThreadTestLarge, EncoderResultTest) { DoTest(); }
+
+// For AV1, only test speed 0 to 3.
+AV1_INSTANTIATE_TEST_CASE(AVxEncoderThreadTest,
+                          ::testing::Values(::libaom_test::kTwoPassGood,
+                                            ::libaom_test::kOnePassGood),
+                          ::testing::Range(2, 4));
+
+AV1_INSTANTIATE_TEST_CASE(AVxEncoderThreadTestLarge,
+                          ::testing::Values(::libaom_test::kTwoPassGood,
+                                            ::libaom_test::kOnePassGood),
+                          ::testing::Range(0, 2));
+}  // namespace
diff --git a/third_party/aom/test/examples.sh b/third_party/aom/test/examples.sh
new file mode 100755
index 0000000000..d3152be7db
--- /dev/null
+++ b/third_party/aom/test/examples.sh
@@ -0,0 +1,29 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This file runs all of the tests for the libaom examples.
+##
+. $(dirname $0)/tools_common.sh
+
+example_tests=$(ls $(dirname $0)/*.sh)
+
+# List of script names to exclude.
+exclude_list="examples tools_common decode_to_md5"
+
+# Filter out the scripts in $exclude_list.
+for word in ${exclude_list}; do
+  example_tests=$(filter_strings "${example_tests}" "${word}" exclude)
+done
+
+for test in ${example_tests}; do
+  # Source each test script so that exporting variables can be avoided.
+  AOM_TEST_NAME="$(basename ${test%.*})"
+  . "${test}"
+done
diff --git a/third_party/aom/test/fdct4x4_test.cc b/third_party/aom/test/fdct4x4_test.cc
new file mode 100644
index 0000000000..ed265e84f2
--- /dev/null
+++ b/third_party/aom/test/fdct4x4_test.cc
@@ -0,0 +1,344 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+#include "av1/common/entropy.h"
+#include "aom/aom_codec.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*FdctFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*IdctFunc)(const tran_low_t *in, uint8_t *out, int stride);
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+using libaom_test::FhtFunc;
+
+typedef std::tr1::tuple<FdctFunc, IdctFunc, int, aom_bit_depth_t, int>
+    Dct4x4Param;
+typedef std::tr1::tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht4x4Param;
+
+void fdct4x4_ref(const int16_t *in, tran_low_t *out, int stride,
+                 int /*tx_type*/) {
+  aom_fdct4x4_c(in, out, stride);
+}
+
+void fht4x4_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht4x4_c(in, out, stride, tx_type);
+}
+
+void fwht4x4_ref(const int16_t *in, tran_low_t *out, int stride,
+                 int /*tx_type*/) {
+  av1_fwht4x4_c(in, out, stride);
+}
+
+#if CONFIG_HIGHBITDEPTH
+void idct4x4_10(const tran_low_t *in, uint8_t *out, int stride) {
+  aom_highbd_idct4x4_16_add_c(in, out, stride, 10);
+}
+
+void idct4x4_12(const tran_low_t *in, uint8_t *out, int stride) {
+  aom_highbd_idct4x4_16_add_c(in, out, stride, 12);
+}
+
+void iht4x4_10(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_highbd_iht4x4_16_add_c(in, out, stride, tx_type, 10);
+}
+
+void iht4x4_12(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_highbd_iht4x4_16_add_c(in, out, stride, tx_type, 12);
+}
+
+void iwht4x4_10(const tran_low_t *in, uint8_t *out, int stride) {
+  aom_highbd_iwht4x4_16_add_c(in, out, stride, 10);
+}
+
+void iwht4x4_12(const tran_low_t *in, uint8_t *out, int stride) {
+  aom_highbd_iwht4x4_16_add_c(in, out, stride, 12);
+}
+
+#if HAVE_SSE2
+void idct4x4_10_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  aom_highbd_idct4x4_16_add_sse2(in, out, stride, 10);
+}
+
+void idct4x4_12_sse2(const tran_low_t *in, uint8_t *out, int stride) {
+  aom_highbd_idct4x4_16_add_sse2(in, out, stride, 12);
+}
+#endif  // HAVE_SSE2
+#endif  // CONFIG_HIGHBITDEPTH
+
+class Trans4x4DCT : public libaom_test::TransformTestBase,
+                    public ::testing::TestWithParam<Dct4x4Param> {
+ public:
+  virtual ~Trans4x4DCT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 4;
+    height_ = 4;
+    fwd_txfm_ref = fdct4x4_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride);
+  }
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
+};
+
+TEST_P(Trans4x4DCT, AccuracyCheck) { RunAccuracyCheck(0, 0.00001); }
+
+TEST_P(Trans4x4DCT, CoeffCheck) { RunCoeffCheck(); }
+
+TEST_P(Trans4x4DCT, MemCheck) { RunMemCheck(); }
+
+TEST_P(Trans4x4DCT, InvAccuracyCheck) { RunInvAccuracyCheck(1); }
+
+class Trans4x4HT : public libaom_test::TransformTestBase,
+                   public ::testing::TestWithParam<Ht4x4Param> {
+ public:
+  virtual ~Trans4x4HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 4;
+    height_ = 4;
+    fwd_txfm_ref = fht4x4_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(Trans4x4HT, AccuracyCheck) { RunAccuracyCheck(1, 0.005); }
+
+TEST_P(Trans4x4HT, CoeffCheck) { RunCoeffCheck(); }
+
+TEST_P(Trans4x4HT, MemCheck) { RunMemCheck(); }
+
+TEST_P(Trans4x4HT, InvAccuracyCheck) { RunInvAccuracyCheck(1); }
+
+class Trans4x4WHT : public libaom_test::TransformTestBase,
+                    public ::testing::TestWithParam<Dct4x4Param> {
+ public:
+  virtual ~Trans4x4WHT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 4;
+    height_ = 4;
+    fwd_txfm_ref = fwht4x4_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride);
+  }
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
+};
+
+TEST_P(Trans4x4WHT, AccuracyCheck) { RunAccuracyCheck(0, 0.00001); }
+
+TEST_P(Trans4x4WHT, CoeffCheck) { RunCoeffCheck(); }
+
+TEST_P(Trans4x4WHT, MemCheck) { RunMemCheck(); }
+
+TEST_P(Trans4x4WHT, InvAccuracyCheck) { RunInvAccuracyCheck(0); }
+using std::tr1::make_tuple;
+
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4DCT,
+    ::testing::Values(
+        make_tuple(&aom_highbd_fdct4x4_c, &idct4x4_10, 0, AOM_BITS_10, 16),
+        make_tuple(&aom_highbd_fdct4x4_c, &idct4x4_12, 0, AOM_BITS_12, 16),
+        make_tuple(&aom_fdct4x4_c, &aom_idct4x4_16_add_c, 0, AOM_BITS_8, 16)));
+#else
+INSTANTIATE_TEST_CASE_P(C, Trans4x4DCT,
+                        ::testing::Values(make_tuple(&aom_fdct4x4_c,
+                                                     &aom_idct4x4_16_add_c, 0,
+                                                     AOM_BITS_8, 16)));
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&av1_highbd_fht4x4_c, &iht4x4_10, 0, AOM_BITS_10, 16),
+        make_tuple(&av1_highbd_fht4x4_c, &iht4x4_10, 1, AOM_BITS_10, 16),
+        make_tuple(&av1_highbd_fht4x4_c, &iht4x4_10, 2, AOM_BITS_10, 16),
+        make_tuple(&av1_highbd_fht4x4_c, &iht4x4_10, 3, AOM_BITS_10, 16),
+        make_tuple(&av1_highbd_fht4x4_c, &iht4x4_12, 0, AOM_BITS_12, 16),
+        make_tuple(&av1_highbd_fht4x4_c, &iht4x4_12, 1, AOM_BITS_12, 16),
+        make_tuple(&av1_highbd_fht4x4_c, &iht4x4_12, 2, AOM_BITS_12, 16),
+        make_tuple(&av1_highbd_fht4x4_c, &iht4x4_12, 3, AOM_BITS_12, 16),
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_c, 0, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_c, 1, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_c, 2, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_c, 3, AOM_BITS_8, 16)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_c, 0, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_c, 1, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_c, 2, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_c, 3, AOM_BITS_8, 16)));
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, Trans4x4WHT,
+    ::testing::Values(
+        make_tuple(&av1_highbd_fwht4x4_c, &iwht4x4_10, 0, AOM_BITS_10, 16),
+        make_tuple(&av1_highbd_fwht4x4_c, &iwht4x4_12, 0, AOM_BITS_12, 16),
+        make_tuple(&av1_fwht4x4_c, &aom_iwht4x4_16_add_c, 0, AOM_BITS_8, 16)));
+#else
+INSTANTIATE_TEST_CASE_P(C, Trans4x4WHT,
+                        ::testing::Values(make_tuple(&av1_fwht4x4_c,
+                                                     &aom_iwht4x4_16_add_c, 0,
+                                                     AOM_BITS_8, 16)));
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if HAVE_NEON_ASM && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(NEON, Trans4x4DCT,
+                        ::testing::Values(make_tuple(&aom_fdct4x4_c,
+                                                     &aom_idct4x4_16_add_neon,
+                                                     0, AOM_BITS_8, 16)));
+#endif  // HAVE_NEON_ASM && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_NEON && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    NEON, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_neon, 0, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_neon, 1, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_neon, 2, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_c, &av1_iht4x4_16_add_neon, 3, AOM_BITS_8, 16)));
+#endif  // HAVE_NEON && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans4x4WHT,
+    ::testing::Values(make_tuple(&av1_fwht4x4_c, &aom_iwht4x4_16_add_c, 0,
+                                 AOM_BITS_8, 16),
+                      make_tuple(&av1_fwht4x4_c, &aom_iwht4x4_16_add_sse2, 0,
+                                 AOM_BITS_8, 16)));
+#endif
+
+#if HAVE_SSE2 && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(SSE2, Trans4x4DCT,
+                        ::testing::Values(make_tuple(&aom_fdct4x4_sse2,
+                                                     &aom_idct4x4_16_add_sse2,
+                                                     0, AOM_BITS_8, 16)));
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans4x4HT,
+    ::testing::Values(make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 0,
+                                 AOM_BITS_8, 16),
+                      make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 1,
+                                 AOM_BITS_8, 16),
+                      make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 2,
+                                 AOM_BITS_8, 16),
+                      make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_sse2, 3,
+                                 AOM_BITS_8, 16)));
+#endif  // HAVE_SSE2 && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_SSE2 && CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans4x4DCT,
+    ::testing::Values(
+        make_tuple(&aom_highbd_fdct4x4_c, &idct4x4_10_sse2, 0, AOM_BITS_10, 16),
+        make_tuple(&aom_highbd_fdct4x4_sse2, &idct4x4_10_sse2, 0, AOM_BITS_10,
+                   16),
+        make_tuple(&aom_highbd_fdct4x4_c, &idct4x4_12_sse2, 0, AOM_BITS_12, 16),
+        make_tuple(&aom_highbd_fdct4x4_sse2, &idct4x4_12_sse2, 0, AOM_BITS_12,
+                   16),
+        make_tuple(&aom_fdct4x4_sse2, &aom_idct4x4_16_add_c, 0, AOM_BITS_8,
+                   16)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_c, 0, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_c, 1, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_c, 2, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_sse2, &av1_iht4x4_16_add_c, 3, AOM_BITS_8, 16)));
+#endif  // HAVE_SSE2 && CONFIG_HIGHBITDEPTH
+
+#if HAVE_MSA && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(MSA, Trans4x4DCT,
+                        ::testing::Values(make_tuple(&aom_fdct4x4_msa,
+                                                     &aom_idct4x4_16_add_msa, 0,
+                                                     AOM_BITS_8, 16)));
+#if !CONFIG_EXT_TX
+INSTANTIATE_TEST_CASE_P(
+    MSA, Trans4x4HT,
+    ::testing::Values(
+        make_tuple(&av1_fht4x4_msa, &av1_iht4x4_16_add_msa, 0, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_msa, &av1_iht4x4_16_add_msa, 1, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_msa, &av1_iht4x4_16_add_msa, 2, AOM_BITS_8, 16),
+        make_tuple(&av1_fht4x4_msa, &av1_iht4x4_16_add_msa, 3, AOM_BITS_8,
+                   16)));
+#endif  // !CONFIG_EXT_TX
+#endif  // HAVE_MSA && !CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/fdct8x8_test.cc b/third_party/aom/test/fdct8x8_test.cc
new file mode 100644
index 0000000000..0e86c70aae
--- /dev/null
+++ b/third_party/aom/test/fdct8x8_test.cc
@@ -0,0 +1,699 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "av1/common/entropy.h"
+#include "av1/common/scan.h"
+#include "aom/aom_codec.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+
+const int kNumCoeffs = 64;
+const double kPi = 3.141592653589793238462643383279502884;
+
+const int kSignBiasMaxDiff255 = 1500;
+const int kSignBiasMaxDiff15 = 10000;
+
+typedef void (*FdctFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*IdctFunc)(const tran_low_t *in, uint8_t *out, int stride);
+typedef void (*FhtFunc)(const int16_t *in, tran_low_t *out, int stride,
+                        int tx_type);
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+
+typedef std::tr1::tuple<FdctFunc, IdctFunc, int, aom_bit_depth_t> Dct8x8Param;
+typedef std::tr1::tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t> Ht8x8Param;
+typedef std::tr1::tuple<IdctFunc, IdctFunc, int, aom_bit_depth_t> Idct8x8Param;
+
+void reference_8x8_dct_1d(const double in[8], double out[8]) {
+  const double kInvSqrt2 = 0.707106781186547524400844362104;
+  for (int k = 0; k < 8; k++) {
+    out[k] = 0.0;
+    for (int n = 0; n < 8; n++)
+      out[k] += in[n] * cos(kPi * (2 * n + 1) * k / 16.0);
+    if (k == 0) out[k] = out[k] * kInvSqrt2;
+  }
+}
+
+void reference_8x8_dct_2d(const int16_t input[kNumCoeffs],
+                          double output[kNumCoeffs]) {
+  // First transform columns
+  for (int i = 0; i < 8; ++i) {
+    double temp_in[8], temp_out[8];
+    for (int j = 0; j < 8; ++j) temp_in[j] = input[j * 8 + i];
+    reference_8x8_dct_1d(temp_in, temp_out);
+    for (int j = 0; j < 8; ++j) output[j * 8 + i] = temp_out[j];
+  }
+  // Then transform rows
+  for (int i = 0; i < 8; ++i) {
+    double temp_in[8], temp_out[8];
+    for (int j = 0; j < 8; ++j) temp_in[j] = output[j + i * 8];
+    reference_8x8_dct_1d(temp_in, temp_out);
+    // Scale by some magic number
+    for (int j = 0; j < 8; ++j) output[j + i * 8] = temp_out[j] * 2;
+  }
+}
+
+void fdct8x8_ref(const int16_t *in, tran_low_t *out, int stride,
+                 int /*tx_type*/) {
+  aom_fdct8x8_c(in, out, stride);
+}
+
+void fht8x8_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht8x8_c(in, out, stride, tx_type);
+}
+
+#if CONFIG_HIGHBITDEPTH
+void iht8x8_10(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_highbd_iht8x8_64_add_c(in, out, stride, tx_type, 10);
+}
+
+void iht8x8_12(const tran_low_t *in, uint8_t *out, int stride, int tx_type) {
+  av1_highbd_iht8x8_64_add_c(in, out, stride, tx_type, 12);
+}
+
+#endif  // CONFIG_HIGHBITDEPTH
+
+class FwdTrans8x8TestBase {
+ public:
+  virtual ~FwdTrans8x8TestBase() {}
+
+ protected:
+  virtual void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) = 0;
+  virtual void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) = 0;
+
+  void RunSignBiasCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    DECLARE_ALIGNED(16, int16_t, test_input_block[64]);
+    DECLARE_ALIGNED(16, tran_low_t, test_output_block[64]);
+    int count_sign_block[64][2];
+    const int count_test_block = 100000;
+
+    memset(count_sign_block, 0, sizeof(count_sign_block));
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-255, 255].
+      for (int j = 0; j < 64; ++j)
+        test_input_block[j] = ((rnd.Rand16() >> (16 - bit_depth_)) & mask_) -
+                              ((rnd.Rand16() >> (16 - bit_depth_)) & mask_);
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(test_input_block, test_output_block, pitch_));
+
+      for (int j = 0; j < 64; ++j) {
+        if (test_output_block[j] < 0)
+          ++count_sign_block[j][0];
+        else if (test_output_block[j] > 0)
+          ++count_sign_block[j][1];
+      }
+    }
+
+    for (int j = 0; j < 64; ++j) {
+      const int diff = abs(count_sign_block[j][0] - count_sign_block[j][1]);
+      const int max_diff = kSignBiasMaxDiff255;
+      EXPECT_LT(diff, max_diff << (bit_depth_ - 8))
+          << "Error: 8x8 FDCT/FHT has a sign bias > "
+          << 1. * max_diff / count_test_block * 100 << "%"
+          << " for input range [-255, 255] at index " << j
+          << " count0: " << count_sign_block[j][0]
+          << " count1: " << count_sign_block[j][1] << " diff: " << diff;
+    }
+
+    memset(count_sign_block, 0, sizeof(count_sign_block));
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_ / 16, mask_ / 16].
+      for (int j = 0; j < 64; ++j)
+        test_input_block[j] =
+            ((rnd.Rand16() & mask_) >> 4) - ((rnd.Rand16() & mask_) >> 4);
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(test_input_block, test_output_block, pitch_));
+
+      for (int j = 0; j < 64; ++j) {
+        if (test_output_block[j] < 0)
+          ++count_sign_block[j][0];
+        else if (test_output_block[j] > 0)
+          ++count_sign_block[j][1];
+      }
+    }
+
+    for (int j = 0; j < 64; ++j) {
+      const int diff = abs(count_sign_block[j][0] - count_sign_block[j][1]);
+      const int max_diff = kSignBiasMaxDiff15;
+      EXPECT_LT(diff, max_diff << (bit_depth_ - 8))
+          << "Error: 8x8 FDCT/FHT has a sign bias > "
+          << 1. * max_diff / count_test_block * 100 << "%"
+          << " for input range [-15, 15] at index " << j
+          << " count0: " << count_sign_block[j][0]
+          << " count1: " << count_sign_block[j][1] << " diff: " << diff;
+    }
+  }
+
+  void RunRoundTripErrorCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    int max_error = 0;
+    int total_error = 0;
+    const int count_test_block = 100000;
+    DECLARE_ALIGNED(16, int16_t, test_input_block[64]);
+    DECLARE_ALIGNED(16, tran_low_t, test_temp_block[64]);
+    DECLARE_ALIGNED(16, uint8_t, dst[64]);
+    DECLARE_ALIGNED(16, uint8_t, src[64]);
+#if CONFIG_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[64]);
+    DECLARE_ALIGNED(16, uint16_t, src16[64]);
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < 64; ++j) {
+        if (bit_depth_ == AOM_BITS_8) {
+          src[j] = rnd.Rand8();
+          dst[j] = rnd.Rand8();
+          test_input_block[j] = src[j] - dst[j];
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand16() & mask_;
+          dst16[j] = rnd.Rand16() & mask_;
+          test_input_block[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(test_input_block, test_temp_block, pitch_));
+      for (int j = 0; j < 64; ++j) {
+        if (test_temp_block[j] > 0) {
+          test_temp_block[j] += 2;
+          test_temp_block[j] /= 4;
+          test_temp_block[j] *= 4;
+        } else {
+          test_temp_block[j] -= 2;
+          test_temp_block[j] /= 4;
+          test_temp_block[j] *= 4;
+        }
+      }
+      if (bit_depth_ == AOM_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(test_temp_block, dst, pitch_));
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(test_temp_block, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < 64; ++j) {
+#if CONFIG_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == AOM_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int diff = dst[j] - src[j];
+#endif
+        const int error = diff * diff;
+        if (max_error < error) max_error = error;
+        total_error += error;
+      }
+    }
+
+    EXPECT_GE(1 << 2 * (bit_depth_ - 8), max_error)
+        << "Error: 8x8 FDCT/IDCT or FHT/IHT has an individual"
+        << " roundtrip error > 1";
+
+    EXPECT_GE((count_test_block << 2 * (bit_depth_ - 8)) / 5, total_error)
+        << "Error: 8x8 FDCT/IDCT or FHT/IHT has average roundtrip "
+        << "error > 1/5 per block";
+  }
+
+  void RunExtremalCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    int max_error = 0;
+    int total_error = 0;
+    int total_coeff_error = 0;
+    const int count_test_block = 100000;
+    DECLARE_ALIGNED(16, int16_t, test_input_block[64]);
+    DECLARE_ALIGNED(16, tran_low_t, test_temp_block[64]);
+    DECLARE_ALIGNED(16, tran_low_t, ref_temp_block[64]);
+    DECLARE_ALIGNED(16, uint8_t, dst[64]);
+    DECLARE_ALIGNED(16, uint8_t, src[64]);
+#if CONFIG_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[64]);
+    DECLARE_ALIGNED(16, uint16_t, src16[64]);
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < 64; ++j) {
+        if (bit_depth_ == AOM_BITS_8) {
+          if (i == 0) {
+            src[j] = 255;
+            dst[j] = 0;
+          } else if (i == 1) {
+            src[j] = 0;
+            dst[j] = 255;
+          } else {
+            src[j] = rnd.Rand8() % 2 ? 255 : 0;
+            dst[j] = rnd.Rand8() % 2 ? 255 : 0;
+          }
+          test_input_block[j] = src[j] - dst[j];
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          if (i == 0) {
+            src16[j] = mask_;
+            dst16[j] = 0;
+          } else if (i == 1) {
+            src16[j] = 0;
+            dst16[j] = mask_;
+          } else {
+            src16[j] = rnd.Rand8() % 2 ? mask_ : 0;
+            dst16[j] = rnd.Rand8() % 2 ? mask_ : 0;
+          }
+          test_input_block[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(test_input_block, test_temp_block, pitch_));
+      ASM_REGISTER_STATE_CHECK(
+          fwd_txfm_ref(test_input_block, ref_temp_block, pitch_, tx_type_));
+      if (bit_depth_ == AOM_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(test_temp_block, dst, pitch_));
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(test_temp_block, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < 64; ++j) {
+#if CONFIG_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == AOM_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int diff = dst[j] - src[j];
+#endif
+        const int error = diff * diff;
+        if (max_error < error) max_error = error;
+        total_error += error;
+
+        const int coeff_diff = test_temp_block[j] - ref_temp_block[j];
+        total_coeff_error += abs(coeff_diff);
+      }
+
+      EXPECT_GE(1 << 2 * (bit_depth_ - 8), max_error)
+          << "Error: Extremal 8x8 FDCT/IDCT or FHT/IHT has"
+          << "an individual roundtrip error > 1";
+
+      EXPECT_GE((count_test_block << 2 * (bit_depth_ - 8)) / 5, total_error)
+          << "Error: Extremal 8x8 FDCT/IDCT or FHT/IHT has average"
+          << " roundtrip error > 1/5 per block";
+
+      EXPECT_EQ(0, total_coeff_error)
+          << "Error: Extremal 8x8 FDCT/FHT has"
+          << "overflow issues in the intermediate steps > 1";
+    }
+  }
+
+  void RunInvAccuracyCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, in[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, src[kNumCoeffs]);
+#if CONFIG_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, src16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      double out_r[kNumCoeffs];
+
+      // Initialize a test block with input range [-255, 255].
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (bit_depth_ == AOM_BITS_8) {
+          src[j] = rnd.Rand8() % 2 ? 255 : 0;
+          dst[j] = src[j] > 0 ? 0 : 255;
+          in[j] = src[j] - dst[j];
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand8() % 2 ? mask_ : 0;
+          dst16[j] = src16[j] > 0 ? 0 : mask_;
+          in[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      reference_8x8_dct_2d(in, out_r);
+      for (int j = 0; j < kNumCoeffs; ++j)
+        coeff[j] = static_cast<tran_low_t>(round(out_r[j]));
+
+      if (bit_depth_ == AOM_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, pitch_));
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(coeff, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == AOM_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int diff = dst[j] - src[j];
+#endif
+        const uint32_t error = diff * diff;
+        EXPECT_GE(1u << 2 * (bit_depth_ - 8), error)
+            << "Error: 8x8 IDCT has error " << error << " at index " << j;
+      }
+    }
+  }
+
+  void RunFwdAccuracyCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+    DECLARE_ALIGNED(16, int16_t, in[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, coeff_r[kNumCoeffs]);
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+
+    for (int i = 0; i < count_test_block; ++i) {
+      double out_r[kNumCoeffs];
+
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < kNumCoeffs; ++j)
+        in[j] = rnd.Rand8() % 2 == 0 ? mask_ : -mask_;
+
+      RunFwdTxfm(in, coeff, pitch_);
+      reference_8x8_dct_2d(in, out_r);
+      for (int j = 0; j < kNumCoeffs; ++j)
+        coeff_r[j] = static_cast<tran_low_t>(round(out_r[j]));
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        const int32_t diff = coeff[j] - coeff_r[j];
+        const uint32_t error = diff * diff;
+        EXPECT_GE(9u << 2 * (bit_depth_ - 8), error)
+            << "Error: 8x8 DCT has error " << error << " at index " << j;
+      }
+    }
+  }
+
+  void CompareInvReference(IdctFunc ref_txfm, int thresh) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 10000;
+    const int eob = 12;
+    DECLARE_ALIGNED(16, tran_low_t, coeff[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, dst[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint8_t, ref[kNumCoeffs]);
+#if CONFIG_HIGHBITDEPTH
+    DECLARE_ALIGNED(16, uint16_t, dst16[kNumCoeffs]);
+    DECLARE_ALIGNED(16, uint16_t, ref16[kNumCoeffs]);
+#endif
+    const int16_t *scan = av1_default_scan_orders[TX_8X8].scan;
+
+    for (int i = 0; i < count_test_block; ++i) {
+      for (int j = 0; j < kNumCoeffs; ++j) {
+        if (j < eob) {
+          // Random values less than the threshold, either positive or negative
+          coeff[scan[j]] = rnd(thresh) * (1 - 2 * (i % 2));
+        } else {
+          coeff[scan[j]] = 0;
+        }
+        if (bit_depth_ == AOM_BITS_8) {
+          dst[j] = 0;
+          ref[j] = 0;
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          dst16[j] = 0;
+          ref16[j] = 0;
+#endif
+        }
+      }
+      if (bit_depth_ == AOM_BITS_8) {
+        ref_txfm(coeff, ref, pitch_);
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, pitch_));
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        ref_txfm(coeff, CONVERT_TO_BYTEPTR(ref16), pitch_);
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(coeff, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < kNumCoeffs; ++j) {
+#if CONFIG_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == AOM_BITS_8 ? dst[j] - ref[j] : dst16[j] - ref16[j];
+#else
+        const int diff = dst[j] - ref[j];
+#endif
+        const uint32_t error = diff * diff;
+        EXPECT_EQ(0u, error) << "Error: 8x8 IDCT has error " << error
+                             << " at index " << j;
+      }
+    }
+  }
+  int pitch_;
+  int tx_type_;
+  FhtFunc fwd_txfm_ref;
+  aom_bit_depth_t bit_depth_;
+  int mask_;
+};
+
+class FwdTrans8x8DCT : public FwdTrans8x8TestBase,
+                       public ::testing::TestWithParam<Dct8x8Param> {
+ public:
+  virtual ~FwdTrans8x8DCT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 8;
+    fwd_txfm_ref = fdct8x8_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride);
+  }
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+
+  FdctFunc fwd_txfm_;
+  IdctFunc inv_txfm_;
+};
+
+TEST_P(FwdTrans8x8DCT, SignBiasCheck) { RunSignBiasCheck(); }
+
+TEST_P(FwdTrans8x8DCT, RoundTripErrorCheck) { RunRoundTripErrorCheck(); }
+
+TEST_P(FwdTrans8x8DCT, ExtremalCheck) { RunExtremalCheck(); }
+
+TEST_P(FwdTrans8x8DCT, FwdAccuracyCheck) { RunFwdAccuracyCheck(); }
+
+TEST_P(FwdTrans8x8DCT, InvAccuracyCheck) { RunInvAccuracyCheck(); }
+
+class FwdTrans8x8HT : public FwdTrans8x8TestBase,
+                      public ::testing::TestWithParam<Ht8x8Param> {
+ public:
+  virtual ~FwdTrans8x8HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 8;
+    fwd_txfm_ref = fht8x8_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(FwdTrans8x8HT, SignBiasCheck) { RunSignBiasCheck(); }
+
+TEST_P(FwdTrans8x8HT, RoundTripErrorCheck) { RunRoundTripErrorCheck(); }
+
+TEST_P(FwdTrans8x8HT, ExtremalCheck) { RunExtremalCheck(); }
+
+class InvTrans8x8DCT : public FwdTrans8x8TestBase,
+                       public ::testing::TestWithParam<Idct8x8Param> {
+ public:
+  virtual ~InvTrans8x8DCT() {}
+
+  virtual void SetUp() {
+    ref_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    thresh_ = GET_PARAM(2);
+    pitch_ = 8;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunInvTxfm(tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride);
+  }
+  void RunFwdTxfm(int16_t * /*out*/, tran_low_t * /*dst*/, int /*stride*/) {}
+
+  IdctFunc ref_txfm_;
+  IdctFunc inv_txfm_;
+  int thresh_;
+};
+
+TEST_P(InvTrans8x8DCT, CompareReference) {
+  CompareInvReference(ref_txfm_, thresh_);
+}
+
+using std::tr1::make_tuple;
+
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(C, FwdTrans8x8DCT,
+                        ::testing::Values(make_tuple(&aom_fdct8x8_c,
+                                                     &aom_idct8x8_64_add_c, 0,
+                                                     AOM_BITS_8)));
+#else
+INSTANTIATE_TEST_CASE_P(C, FwdTrans8x8DCT,
+                        ::testing::Values(make_tuple(&aom_fdct8x8_c,
+                                                     &aom_idct8x8_64_add_c, 0,
+                                                     AOM_BITS_8)));
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    C, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_c, 0, AOM_BITS_8),
+        make_tuple(&av1_highbd_fht8x8_c, &iht8x8_10, 0, AOM_BITS_10),
+        make_tuple(&av1_highbd_fht8x8_c, &iht8x8_10, 1, AOM_BITS_10),
+        make_tuple(&av1_highbd_fht8x8_c, &iht8x8_10, 2, AOM_BITS_10),
+        make_tuple(&av1_highbd_fht8x8_c, &iht8x8_10, 3, AOM_BITS_10),
+        make_tuple(&av1_highbd_fht8x8_c, &iht8x8_12, 0, AOM_BITS_12),
+        make_tuple(&av1_highbd_fht8x8_c, &iht8x8_12, 1, AOM_BITS_12),
+        make_tuple(&av1_highbd_fht8x8_c, &iht8x8_12, 2, AOM_BITS_12),
+        make_tuple(&av1_highbd_fht8x8_c, &iht8x8_12, 3, AOM_BITS_12),
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_c, 1, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_c, 2, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_c, 3, AOM_BITS_8)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    C, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_c, 0, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_c, 1, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_c, 2, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_c, 3, AOM_BITS_8)));
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if HAVE_NEON_ASM && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(NEON, FwdTrans8x8DCT,
+                        ::testing::Values(make_tuple(&aom_fdct8x8_neon,
+                                                     &aom_idct8x8_64_add_neon,
+                                                     0, AOM_BITS_8)));
+#endif  // HAVE_NEON_ASM && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_NEON && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    NEON, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_neon, 0, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_neon, 1, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_neon, 2, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_c, &av1_iht8x8_64_add_neon, 3, AOM_BITS_8)));
+#endif  // HAVE_NEON && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_SSE2 && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(SSE2, FwdTrans8x8DCT,
+                        ::testing::Values(make_tuple(&aom_fdct8x8_sse2,
+                                                     &aom_idct8x8_64_add_sse2,
+                                                     0, AOM_BITS_8)));
+INSTANTIATE_TEST_CASE_P(
+    SSE2, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 0, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 1, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 2, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_sse2, 3, AOM_BITS_8)));
+#endif  // HAVE_SSE2 && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_SSE2 && CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(SSE2, FwdTrans8x8DCT,
+                        ::testing::Values(make_tuple(&aom_fdct8x8_sse2,
+                                                     &aom_idct8x8_64_add_c, 0,
+                                                     AOM_BITS_8)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_c, 0, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_c, 1, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_c, 2, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_sse2, &av1_iht8x8_64_add_c, 3, AOM_BITS_8)));
+
+#endif  // HAVE_SSE2 && CONFIG_HIGHBITDEPTH
+
+#if HAVE_SSSE3 && ARCH_X86_64
+INSTANTIATE_TEST_CASE_P(SSSE3, FwdTrans8x8DCT,
+                        ::testing::Values(make_tuple(&aom_fdct8x8_ssse3,
+                                                     &aom_idct8x8_64_add_ssse3,
+                                                     0, AOM_BITS_8)));
+#endif
+
+#if HAVE_MSA && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(MSA, FwdTrans8x8DCT,
+                        ::testing::Values(make_tuple(&aom_fdct8x8_msa,
+                                                     &aom_idct8x8_64_add_msa, 0,
+                                                     AOM_BITS_8)));
+#if !CONFIG_EXT_TX
+INSTANTIATE_TEST_CASE_P(
+    MSA, FwdTrans8x8HT,
+    ::testing::Values(
+        make_tuple(&av1_fht8x8_msa, &av1_iht8x8_64_add_msa, 0, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_msa, &av1_iht8x8_64_add_msa, 1, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_msa, &av1_iht8x8_64_add_msa, 2, AOM_BITS_8),
+        make_tuple(&av1_fht8x8_msa, &av1_iht8x8_64_add_msa, 3, AOM_BITS_8)));
+#endif  // !CONFIG_EXT_TX
+#endif  // HAVE_MSA && !CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/fht32x32_test.cc b/third_party/aom/test/fht32x32_test.cc
new file mode 100644
index 0000000000..56ac597c03
--- /dev/null
+++ b/third_party/aom/test/fht32x32_test.cc
@@ -0,0 +1,216 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/transform_test_base.h"
+#include "test/util.h"
+#include "aom_ports/mem.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+using std::tr1::tuple;
+using libaom_test::FhtFunc;
+typedef tuple<FhtFunc, IhtFunc, int, aom_bit_depth_t, int> Ht32x32Param;
+
+void fht32x32_ref(const int16_t *in, tran_low_t *out, int stride, int tx_type) {
+  av1_fht32x32_c(in, out, stride, tx_type);
+}
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*IHbdHtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                           int tx_type, int bd);
+typedef void (*HbdHtFunc)(const int16_t *input, int32_t *output, int stride,
+                          int tx_type, int bd);
+
+// Target optimized function, tx_type, bit depth
+typedef tuple<HbdHtFunc, int, int> HighbdHt32x32Param;
+
+void highbd_fht32x32_ref(const int16_t *in, int32_t *out, int stride,
+                         int tx_type, int bd) {
+  av1_fwd_txfm2d_32x32_c(in, out, stride, tx_type, bd);
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if HAVE_AVX2
+void dummy_inv_txfm(const tran_low_t *in, uint8_t *out, int stride,
+                    int tx_type) {
+  (void)in;
+  (void)out;
+  (void)stride;
+  (void)tx_type;
+}
+#endif
+
+class AV1Trans32x32HT : public libaom_test::TransformTestBase,
+                        public ::testing::TestWithParam<Ht32x32Param> {
+ public:
+  virtual ~AV1Trans32x32HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    inv_txfm_ = GET_PARAM(1);
+    tx_type_ = GET_PARAM(2);
+    pitch_ = 32;
+    height_ = 32;
+    fwd_txfm_ref = fht32x32_ref;
+    bit_depth_ = GET_PARAM(3);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = GET_PARAM(4);
+  }
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) {
+    fwd_txfm_(in, out, stride, tx_type_);
+  }
+
+  void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) {
+    inv_txfm_(out, dst, stride, tx_type_);
+  }
+
+  FhtFunc fwd_txfm_;
+  IhtFunc inv_txfm_;
+};
+
+TEST_P(AV1Trans32x32HT, CoeffCheck) { RunCoeffCheck(); }
+TEST_P(AV1Trans32x32HT, MemCheck) { RunMemCheck(); }
+
+#if CONFIG_HIGHBITDEPTH
+class AV1HighbdTrans32x32HT
+    : public ::testing::TestWithParam<HighbdHt32x32Param> {
+ public:
+  virtual ~AV1HighbdTrans32x32HT() {}
+
+  virtual void SetUp() {
+    fwd_txfm_ = GET_PARAM(0);
+    fwd_txfm_ref_ = highbd_fht32x32_ref;
+    tx_type_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+    mask_ = (1 << bit_depth_) - 1;
+    num_coeffs_ = 1024;
+
+    input_ = reinterpret_cast<int16_t *>(
+        aom_memalign(32, sizeof(int16_t) * num_coeffs_));
+    output_ = reinterpret_cast<int32_t *>(
+        aom_memalign(32, sizeof(int32_t) * num_coeffs_));
+    output_ref_ = reinterpret_cast<int32_t *>(
+        aom_memalign(32, sizeof(int32_t) * num_coeffs_));
+  }
+
+  virtual void TearDown() {
+    aom_free(input_);
+    aom_free(output_);
+    aom_free(output_ref_);
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void RunBitexactCheck();
+
+ private:
+  HbdHtFunc fwd_txfm_;
+  HbdHtFunc fwd_txfm_ref_;
+  int tx_type_;
+  int bit_depth_;
+  int mask_;
+  int num_coeffs_;
+  int16_t *input_;
+  int32_t *output_;
+  int32_t *output_ref_;
+};
+
+void AV1HighbdTrans32x32HT::RunBitexactCheck() {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  int i, j;
+  const int stride = 32;
+  const int num_tests = 1000;
+
+  for (i = 0; i < num_tests; ++i) {
+    for (j = 0; j < num_coeffs_; ++j) {
+      input_[j] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+    }
+
+    fwd_txfm_ref_(input_, output_ref_, stride, tx_type_, bit_depth_);
+    ASM_REGISTER_STATE_CHECK(
+        fwd_txfm_(input_, output_, stride, tx_type_, bit_depth_));
+
+    for (j = 0; j < num_coeffs_; ++j) {
+      EXPECT_EQ(output_ref_[j], output_[j])
+          << "Not bit-exact result at index: " << j << " at test block: " << i;
+    }
+  }
+}
+
+TEST_P(AV1HighbdTrans32x32HT, HighbdCoeffCheck) { RunBitexactCheck(); }
+#endif  // CONFIG_HIGHBITDEPTH
+
+using std::tr1::make_tuple;
+
+#if HAVE_SSE2
+const Ht32x32Param kArrayHt32x32Param_sse2[] = {
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 0, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 1, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 2, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 3, AOM_BITS_8, 1024),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 4, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 5, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 6, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 7, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 8, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 9, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 10, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 11, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 12, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 13, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 14, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_sse2, &dummy_inv_txfm, 15, AOM_BITS_8, 1024)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(SSE2, AV1Trans32x32HT,
+                        ::testing::ValuesIn(kArrayHt32x32Param_sse2));
+#endif  // HAVE_SSE2
+
+#if HAVE_AVX2
+const Ht32x32Param kArrayHt32x32Param_avx2[] = {
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 0, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 1, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 2, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 3, AOM_BITS_8, 1024),
+#if CONFIG_EXT_TX
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 4, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 5, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 6, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 7, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 8, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 9, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 10, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 11, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 12, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 13, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 14, AOM_BITS_8, 1024),
+  make_tuple(&av1_fht32x32_avx2, &dummy_inv_txfm, 15, AOM_BITS_8, 1024)
+#endif  // CONFIG_EXT_TX
+};
+INSTANTIATE_TEST_CASE_P(AVX2, AV1Trans32x32HT,
+                        ::testing::ValuesIn(kArrayHt32x32Param_avx2));
+#endif  // HAVE_AVX2
+}  // namespace
diff --git a/third_party/aom/test/filterintra_predictors_test.cc b/third_party/aom/test/filterintra_predictors_test.cc
new file mode 100644
index 0000000000..5c6b56d144
--- /dev/null
+++ b/third_party/aom/test/filterintra_predictors_test.cc
@@ -0,0 +1,331 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "av1/common/enums.h"
+
+namespace {
+
+using std::tr1::tuple;
+using libaom_test::ACMRandom;
+
+typedef void (*Predictor)(uint8_t *dst, ptrdiff_t stride, int bs,
+                          const uint8_t *above, const uint8_t *left);
+
+// Note:
+//  Test parameter list:
+//  Reference predictor, optimized predictor, prediction mode, block size
+//
+typedef tuple<Predictor, Predictor, int> PredFuncMode;
+typedef tuple<PredFuncMode, int> PredParams;
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*HbdPredictor)(uint16_t *dst, ptrdiff_t stride, int bs,
+                             const uint16_t *above, const uint16_t *left,
+                             int bd);
+
+// Note:
+//  Test parameter list:
+//  Reference predictor, optimized predictor, prediction mode, block size,
+//  bit depth
+//
+typedef tuple<HbdPredictor, HbdPredictor, int> HbdPredFuncMode;
+typedef tuple<HbdPredFuncMode, int, int> HbdPredParams;
+#endif
+
+const int MaxBlkSize = 32;
+
+// By default, disable speed test
+#define PREDICTORS_SPEED_TEST (0)
+
+#if PREDICTORS_SPEED_TEST
+const int MaxTestNum = 100000;
+#else
+const int MaxTestNum = 100;
+#endif
+
+class AV1FilterIntraPredOptimzTest
+    : public ::testing::TestWithParam<PredParams> {
+ public:
+  virtual ~AV1FilterIntraPredOptimzTest() {}
+  virtual void SetUp() {
+    PredFuncMode funcMode = GET_PARAM(0);
+    predFuncRef_ = std::tr1::get<0>(funcMode);
+    predFunc_ = std::tr1::get<1>(funcMode);
+    mode_ = std::tr1::get<2>(funcMode);
+    blockSize_ = GET_PARAM(1);
+
+    alloc_ = new uint8_t[3 * MaxBlkSize + 2];
+    predRef_ = new uint8_t[MaxBlkSize * MaxBlkSize];
+    pred_ = new uint8_t[MaxBlkSize * MaxBlkSize];
+  }
+
+  virtual void TearDown() {
+    delete[] alloc_;
+    delete[] predRef_;
+    delete[] pred_;
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void RunTest() const {
+    int tstIndex = 0;
+    int stride = blockSize_;
+    uint8_t *left = alloc_;
+    uint8_t *above = alloc_ + MaxBlkSize + 1;
+    while (tstIndex < MaxTestNum) {
+      PrepareBuffer();
+      predFuncRef_(predRef_, stride, blockSize_, &above[1], left);
+      ASM_REGISTER_STATE_CHECK(
+          predFunc_(pred_, stride, blockSize_, &above[1], left));
+      DiffPred(tstIndex);
+      tstIndex += 1;
+    }
+  }
+
+  void RunSpeedTestC() const {
+    int tstIndex = 0;
+    int stride = blockSize_;
+    uint8_t *left = alloc_;
+    uint8_t *above = alloc_ + MaxBlkSize + 1;
+    PrepareBuffer();
+    while (tstIndex < MaxTestNum) {
+      predFuncRef_(predRef_, stride, blockSize_, &above[1], left);
+      tstIndex += 1;
+    }
+  }
+
+  void RunSpeedTestSSE() const {
+    int tstIndex = 0;
+    int stride = blockSize_;
+    uint8_t *left = alloc_;
+    uint8_t *above = alloc_ + MaxBlkSize + 1;
+    PrepareBuffer();
+    while (tstIndex < MaxTestNum) {
+      predFunc_(predRef_, stride, blockSize_, &above[1], left);
+      tstIndex += 1;
+    }
+  }
+
+ private:
+  void PrepareBuffer() const {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    int i = 0;
+    while (i < (3 * MaxBlkSize + 2)) {
+      alloc_[i] = rnd.Rand8();
+      i += 1;
+    }
+  }
+
+  void DiffPred(int testNum) const {
+    int i = 0;
+    while (i < blockSize_ * blockSize_) {
+      EXPECT_EQ(predRef_[i], pred_[i]) << "Error at position: " << i << " "
+                                       << "Block size: " << blockSize_ << " "
+                                       << "Test number: " << testNum;
+      i += 1;
+    }
+  }
+
+  Predictor predFunc_;
+  Predictor predFuncRef_;
+  int mode_;
+  int blockSize_;
+  uint8_t *alloc_;
+  uint8_t *pred_;
+  uint8_t *predRef_;
+};
+
+#if CONFIG_HIGHBITDEPTH
+class AV1HbdFilterIntraPredOptimzTest
+    : public ::testing::TestWithParam<HbdPredParams> {
+ public:
+  virtual ~AV1HbdFilterIntraPredOptimzTest() {}
+  virtual void SetUp() {
+    HbdPredFuncMode funcMode = GET_PARAM(0);
+    predFuncRef_ = std::tr1::get<0>(funcMode);
+    predFunc_ = std::tr1::get<1>(funcMode);
+    mode_ = std::tr1::get<2>(funcMode);
+    blockSize_ = GET_PARAM(1);
+    bd_ = GET_PARAM(2);
+
+    alloc_ = new uint16_t[3 * MaxBlkSize + 2];
+    predRef_ = new uint16_t[MaxBlkSize * MaxBlkSize];
+    pred_ = new uint16_t[MaxBlkSize * MaxBlkSize];
+  }
+
+  virtual void TearDown() {
+    delete[] alloc_;
+    delete[] predRef_;
+    delete[] pred_;
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void RunTest() const {
+    int tstIndex = 0;
+    int stride = blockSize_;
+    uint16_t *left = alloc_;
+    uint16_t *above = alloc_ + MaxBlkSize + 1;
+    while (tstIndex < MaxTestNum) {
+      PrepareBuffer();
+      predFuncRef_(predRef_, stride, blockSize_, &above[1], left, bd_);
+      ASM_REGISTER_STATE_CHECK(
+          predFunc_(pred_, stride, blockSize_, &above[1], left, bd_));
+      DiffPred(tstIndex);
+      tstIndex += 1;
+    }
+  }
+
+  void RunSpeedTestC() const {
+    int tstIndex = 0;
+    int stride = blockSize_;
+    uint16_t *left = alloc_;
+    uint16_t *above = alloc_ + MaxBlkSize + 1;
+    PrepareBuffer();
+    while (tstIndex < MaxTestNum) {
+      predFuncRef_(predRef_, stride, blockSize_, &above[1], left, bd_);
+      tstIndex += 1;
+    }
+  }
+
+  void RunSpeedTestSSE() const {
+    int tstIndex = 0;
+    int stride = blockSize_;
+    uint16_t *left = alloc_;
+    uint16_t *above = alloc_ + MaxBlkSize + 1;
+    PrepareBuffer();
+    while (tstIndex < MaxTestNum) {
+      predFunc_(predRef_, stride, blockSize_, &above[1], left, bd_);
+      tstIndex += 1;
+    }
+  }
+
+ private:
+  void PrepareBuffer() const {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    int i = 0;
+    while (i < (3 * MaxBlkSize + 2)) {
+      alloc_[i] = rnd.Rand16() & ((1 << bd_) - 1);
+      i += 1;
+    }
+  }
+
+  void DiffPred(int testNum) const {
+    int i = 0;
+    while (i < blockSize_ * blockSize_) {
+      EXPECT_EQ(predRef_[i], pred_[i]) << "Error at position: " << i << " "
+                                       << "Block size: " << blockSize_ << " "
+                                       << "Bit depth: " << bd_ << " "
+                                       << "Test number: " << testNum;
+      i += 1;
+    }
+  }
+
+  HbdPredictor predFunc_;
+  HbdPredictor predFuncRef_;
+  int mode_;
+  int blockSize_;
+  int bd_;
+  uint16_t *alloc_;
+  uint16_t *pred_;
+  uint16_t *predRef_;
+};
+#endif  // CONFIG_HIGHBITDEPTH
+
+TEST_P(AV1FilterIntraPredOptimzTest, BitExactCheck) { RunTest(); }
+
+#if PREDICTORS_SPEED_TEST
+TEST_P(AV1FilterIntraPredOptimzTest, SpeedCheckC) { RunSpeedTestC(); }
+
+TEST_P(AV1FilterIntraPredOptimzTest, SpeedCheckSSE) { RunSpeedTestSSE(); }
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+TEST_P(AV1HbdFilterIntraPredOptimzTest, BitExactCheck) { RunTest(); }
+
+#if PREDICTORS_SPEED_TEST
+TEST_P(AV1HbdFilterIntraPredOptimzTest, SpeedCheckC) { RunSpeedTestC(); }
+
+TEST_P(AV1HbdFilterIntraPredOptimzTest, SpeedCheckSSE) { RunSpeedTestSSE(); }
+#endif  // PREDICTORS_SPEED_TEST
+#endif  // CONFIG_HIGHBITDEPTH
+
+using std::tr1::make_tuple;
+
+const PredFuncMode kPredFuncMdArray[] = {
+  make_tuple(av1_dc_filter_predictor_c, av1_dc_filter_predictor_sse4_1,
+             DC_PRED),
+  make_tuple(av1_v_filter_predictor_c, av1_v_filter_predictor_sse4_1, V_PRED),
+  make_tuple(av1_h_filter_predictor_c, av1_h_filter_predictor_sse4_1, H_PRED),
+  make_tuple(av1_d45_filter_predictor_c, av1_d45_filter_predictor_sse4_1,
+             D45_PRED),
+  make_tuple(av1_d135_filter_predictor_c, av1_d135_filter_predictor_sse4_1,
+             D135_PRED),
+  make_tuple(av1_d117_filter_predictor_c, av1_d117_filter_predictor_sse4_1,
+             D117_PRED),
+  make_tuple(av1_d153_filter_predictor_c, av1_d153_filter_predictor_sse4_1,
+             D153_PRED),
+  make_tuple(av1_d207_filter_predictor_c, av1_d207_filter_predictor_sse4_1,
+             D207_PRED),
+  make_tuple(av1_d63_filter_predictor_c, av1_d63_filter_predictor_sse4_1,
+             D63_PRED),
+  make_tuple(av1_tm_filter_predictor_c, av1_tm_filter_predictor_sse4_1,
+             TM_PRED),
+};
+
+const int kBlkSize[] = { 4, 8, 16, 32 };
+
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, AV1FilterIntraPredOptimzTest,
+    ::testing::Combine(::testing::ValuesIn(kPredFuncMdArray),
+                       ::testing::ValuesIn(kBlkSize)));
+
+#if CONFIG_HIGHBITDEPTH
+const HbdPredFuncMode kHbdPredFuncMdArray[] = {
+  make_tuple(av1_highbd_dc_filter_predictor_c,
+             av1_highbd_dc_filter_predictor_sse4_1, DC_PRED),
+  make_tuple(av1_highbd_v_filter_predictor_c,
+             av1_highbd_v_filter_predictor_sse4_1, V_PRED),
+  make_tuple(av1_highbd_h_filter_predictor_c,
+             av1_highbd_h_filter_predictor_sse4_1, H_PRED),
+  make_tuple(av1_highbd_d45_filter_predictor_c,
+             av1_highbd_d45_filter_predictor_sse4_1, D45_PRED),
+  make_tuple(av1_highbd_d135_filter_predictor_c,
+             av1_highbd_d135_filter_predictor_sse4_1, D135_PRED),
+  make_tuple(av1_highbd_d117_filter_predictor_c,
+             av1_highbd_d117_filter_predictor_sse4_1, D117_PRED),
+  make_tuple(av1_highbd_d153_filter_predictor_c,
+             av1_highbd_d153_filter_predictor_sse4_1, D153_PRED),
+  make_tuple(av1_highbd_d207_filter_predictor_c,
+             av1_highbd_d207_filter_predictor_sse4_1, D207_PRED),
+  make_tuple(av1_highbd_d63_filter_predictor_c,
+             av1_highbd_d63_filter_predictor_sse4_1, D63_PRED),
+  make_tuple(av1_highbd_tm_filter_predictor_c,
+             av1_highbd_tm_filter_predictor_sse4_1, TM_PRED),
+};
+
+const int kBd[] = { 10, 12 };
+
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, AV1HbdFilterIntraPredOptimzTest,
+    ::testing::Combine(::testing::ValuesIn(kHbdPredFuncMdArray),
+                       ::testing::ValuesIn(kBlkSize),
+                       ::testing::ValuesIn(kBd)));
+#endif  // CONFIG_HIGHBITDEPTH
+
+}  // namespace
diff --git a/third_party/aom/test/frame_size_tests.cc b/third_party/aom/test/frame_size_tests.cc
new file mode 100644
index 0000000000..73cc9c0755
--- /dev/null
+++ b/third_party/aom/test/frame_size_tests.cc
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/video_source.h"
+
+namespace {
+
+class AV1FrameSizeTests : public ::libaom_test::EncoderTest,
+                          public ::testing::Test {
+ protected:
+  AV1FrameSizeTests()
+      : EncoderTest(&::libaom_test::kAV1), expected_res_(AOM_CODEC_OK) {}
+  virtual ~AV1FrameSizeTests() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(::libaom_test::kRealTime);
+  }
+
+  virtual bool HandleDecodeResult(const aom_codec_err_t res_dec,
+                                  libaom_test::Decoder *decoder) {
+    EXPECT_EQ(expected_res_, res_dec) << decoder->DecodeError();
+    return !::testing::Test::HasFailure();
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AOME_SET_CPUUSED, 7);
+      encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+      encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
+      encoder->Control(AOME_SET_ARNR_STRENGTH, 5);
+    }
+  }
+
+  int expected_res_;
+};
+
+TEST_F(AV1FrameSizeTests, TestInvalidSizes) {
+  ::libaom_test::RandomVideoSource video;
+
+#if CONFIG_SIZE_LIMIT
+  video.SetSize(DECODE_WIDTH_LIMIT + 16, DECODE_HEIGHT_LIMIT + 16);
+  video.set_limit(2);
+  expected_res_ = AOM_CODEC_CORRUPT_FRAME;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+#endif
+}
+
+TEST_F(AV1FrameSizeTests, LargeValidSizes) {
+  ::libaom_test::RandomVideoSource video;
+
+#if CONFIG_SIZE_LIMIT
+  video.SetSize(DECODE_WIDTH_LIMIT, DECODE_HEIGHT_LIMIT);
+  video.set_limit(2);
+  expected_res_ = AOM_CODEC_OK;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+#else
+  // This test produces a pretty large single frame allocation,  (roughly
+  // 25 megabits). The encoder allocates a good number of these frames
+  // one for each lag in frames (for 2 pass), and then one for each possible
+  // reference buffer (8) - we can end up with up to 30 buffers of roughly this
+  // size or almost 1 gig of memory.
+  // In total the allocations will exceed 2GiB which may cause a failure with
+  // non-64 bit platforms, use a smaller size in that case.
+  if (sizeof(void *) < 8)
+    video.SetSize(2560, 1440);
+  else
+    video.SetSize(4096, 4096);
+
+  video.set_limit(2);
+  expected_res_ = AOM_CODEC_OK;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+#endif
+}
+
+TEST_F(AV1FrameSizeTests, OneByOneVideo) {
+  ::libaom_test::RandomVideoSource video;
+
+  video.SetSize(1, 1);
+  video.set_limit(2);
+  expected_res_ = AOM_CODEC_OK;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+#undef ONE_BY_ONE_VIDEO_NAME
+}  // namespace
diff --git a/third_party/aom/test/function_equivalence_test.h b/third_party/aom/test/function_equivalence_test.h
new file mode 100644
index 0000000000..4b22c74a2b
--- /dev/null
+++ b/third_party/aom/test/function_equivalence_test.h
@@ -0,0 +1,69 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef TEST_FUNCTION_EQUIVALENCE_TEST_H_
+#define TEST_FUNCTION_EQUIVALENCE_TEST_H_
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/util.h"
+
+using libaom_test::ACMRandom;
+
+namespace libaom_test {
+// Base class for tests that compare 2 implementations of the same function
+// for equivalence. The template parameter should be pointer to a function
+// that is being tested.
+//
+// The test takes a 3-parameters encapsulating struct 'FuncParam', containing:
+//   - Pointer to reference function
+//   - Pointer to tested function
+//   - Integer bit depth (default to 0).
+//
+// These values are then accessible in the tests as member of params_:
+// params_.ref_func, params_.tst_func, and params_.bit_depth.
+//
+
+template <typename T>
+struct FuncParam {
+  FuncParam(T ref = NULL, T tst = NULL, int bit_depth = 0)
+      : ref_func(ref), tst_func(tst), bit_depth(bit_depth) {}
+  T ref_func;
+  T tst_func;
+  int bit_depth;
+};
+
+template <typename T>
+std::ostream &operator<<(std::ostream &os, const FuncParam<T> &p) {
+  return os << "bit_depth:" << p.bit_depth
+            << " function:" << reinterpret_cast<const void *>(p.ref_func)
+            << " function:" << reinterpret_cast<const void *>(p.tst_func);
+}
+
+template <typename T>
+class FunctionEquivalenceTest : public ::testing::TestWithParam<FuncParam<T> > {
+ public:
+  FunctionEquivalenceTest() : rng_(ACMRandom::DeterministicSeed()) {}
+
+  virtual ~FunctionEquivalenceTest() {}
+
+  virtual void SetUp() { params_ = this->GetParam(); }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  ACMRandom rng_;
+  FuncParam<T> params_;
+};
+
+}  // namespace libaom_test
+#endif  // TEST_FUNCTION_EQUIVALENCE_TEST_H_
diff --git a/third_party/aom/test/hadamard_test.cc b/third_party/aom/test/hadamard_test.cc
new file mode 100644
index 0000000000..db5cb74747
--- /dev/null
+++ b/third_party/aom/test/hadamard_test.cc
@@ -0,0 +1,221 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <algorithm>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_dsp_rtcd.h"
+
+#include "test/acm_random.h"
+#include "test/register_state_check.h"
+
+namespace {
+
+using ::libaom_test::ACMRandom;
+
+typedef void (*HadamardFunc)(const int16_t *a, int a_stride, int16_t *b);
+
+void hadamard_loop(const int16_t *a, int a_stride, int16_t *out) {
+  int16_t b[8];
+  for (int i = 0; i < 8; i += 2) {
+    b[i + 0] = a[i * a_stride] + a[(i + 1) * a_stride];
+    b[i + 1] = a[i * a_stride] - a[(i + 1) * a_stride];
+  }
+  int16_t c[8];
+  for (int i = 0; i < 8; i += 4) {
+    c[i + 0] = b[i + 0] + b[i + 2];
+    c[i + 1] = b[i + 1] + b[i + 3];
+    c[i + 2] = b[i + 0] - b[i + 2];
+    c[i + 3] = b[i + 1] - b[i + 3];
+  }
+  out[0] = c[0] + c[4];
+  out[7] = c[1] + c[5];
+  out[3] = c[2] + c[6];
+  out[4] = c[3] + c[7];
+  out[2] = c[0] - c[4];
+  out[6] = c[1] - c[5];
+  out[1] = c[2] - c[6];
+  out[5] = c[3] - c[7];
+}
+
+void reference_hadamard8x8(const int16_t *a, int a_stride, int16_t *b) {
+  int16_t buf[64];
+  for (int i = 0; i < 8; ++i) {
+    hadamard_loop(a + i, a_stride, buf + i * 8);
+  }
+
+  for (int i = 0; i < 8; ++i) {
+    hadamard_loop(buf + i, 8, b + i * 8);
+  }
+}
+
+void reference_hadamard16x16(const int16_t *a, int a_stride, int16_t *b) {
+  /* The source is a 16x16 block. The destination is rearranged to 8x32.
+   * Input is 9 bit. */
+  reference_hadamard8x8(a + 0 + 0 * a_stride, a_stride, b + 0);
+  reference_hadamard8x8(a + 8 + 0 * a_stride, a_stride, b + 64);
+  reference_hadamard8x8(a + 0 + 8 * a_stride, a_stride, b + 128);
+  reference_hadamard8x8(a + 8 + 8 * a_stride, a_stride, b + 192);
+
+  /* Overlay the 8x8 blocks and combine. */
+  for (int i = 0; i < 64; ++i) {
+    /* 8x8 steps the range up to 15 bits. */
+    const int16_t a0 = b[0];
+    const int16_t a1 = b[64];
+    const int16_t a2 = b[128];
+    const int16_t a3 = b[192];
+
+    /* Prevent the result from escaping int16_t. */
+    const int16_t b0 = (a0 + a1) >> 1;
+    const int16_t b1 = (a0 - a1) >> 1;
+    const int16_t b2 = (a2 + a3) >> 1;
+    const int16_t b3 = (a2 - a3) >> 1;
+
+    /* Store a 16 bit value. */
+    b[0] = b0 + b2;
+    b[64] = b1 + b3;
+    b[128] = b0 - b2;
+    b[192] = b1 - b3;
+
+    ++b;
+  }
+}
+
+class HadamardTestBase : public ::testing::TestWithParam<HadamardFunc> {
+ public:
+  virtual void SetUp() {
+    h_func_ = GetParam();
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+  }
+
+ protected:
+  HadamardFunc h_func_;
+  ACMRandom rnd_;
+};
+
+class Hadamard8x8Test : public HadamardTestBase {};
+
+TEST_P(Hadamard8x8Test, CompareReferenceRandom) {
+  DECLARE_ALIGNED(16, int16_t, a[64]);
+  DECLARE_ALIGNED(16, int16_t, b[64]);
+  int16_t b_ref[64];
+  for (int i = 0; i < 64; ++i) {
+    a[i] = rnd_.Rand9Signed();
+  }
+  memset(b, 0, sizeof(b));
+  memset(b_ref, 0, sizeof(b_ref));
+
+  reference_hadamard8x8(a, 8, b_ref);
+  ASM_REGISTER_STATE_CHECK(h_func_(a, 8, b));
+
+  // The order of the output is not important. Sort before checking.
+  std::sort(b, b + 64);
+  std::sort(b_ref, b_ref + 64);
+  EXPECT_EQ(0, memcmp(b, b_ref, sizeof(b)));
+}
+
+TEST_P(Hadamard8x8Test, VaryStride) {
+  DECLARE_ALIGNED(16, int16_t, a[64 * 8]);
+  DECLARE_ALIGNED(16, int16_t, b[64]);
+  int16_t b_ref[64];
+  for (int i = 0; i < 64 * 8; ++i) {
+    a[i] = rnd_.Rand9Signed();
+  }
+
+  for (int i = 8; i < 64; i += 8) {
+    memset(b, 0, sizeof(b));
+    memset(b_ref, 0, sizeof(b_ref));
+
+    reference_hadamard8x8(a, i, b_ref);
+    ASM_REGISTER_STATE_CHECK(h_func_(a, i, b));
+
+    // The order of the output is not important. Sort before checking.
+    std::sort(b, b + 64);
+    std::sort(b_ref, b_ref + 64);
+    EXPECT_EQ(0, memcmp(b, b_ref, sizeof(b)));
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(C, Hadamard8x8Test,
+                        ::testing::Values(&aom_hadamard_8x8_c));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, Hadamard8x8Test,
+                        ::testing::Values(&aom_hadamard_8x8_sse2));
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3 && ARCH_X86_64
+INSTANTIATE_TEST_CASE_P(SSSE3, Hadamard8x8Test,
+                        ::testing::Values(&aom_hadamard_8x8_ssse3));
+#endif  // HAVE_SSSE3 && ARCH_X86_64
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, Hadamard8x8Test,
+                        ::testing::Values(&aom_hadamard_8x8_neon));
+#endif  // HAVE_NEON
+
+class Hadamard16x16Test : public HadamardTestBase {};
+
+TEST_P(Hadamard16x16Test, CompareReferenceRandom) {
+  DECLARE_ALIGNED(16, int16_t, a[16 * 16]);
+  DECLARE_ALIGNED(16, int16_t, b[16 * 16]);
+  int16_t b_ref[16 * 16];
+  for (int i = 0; i < 16 * 16; ++i) {
+    a[i] = rnd_.Rand9Signed();
+  }
+  memset(b, 0, sizeof(b));
+  memset(b_ref, 0, sizeof(b_ref));
+
+  reference_hadamard16x16(a, 16, b_ref);
+  ASM_REGISTER_STATE_CHECK(h_func_(a, 16, b));
+
+  // The order of the output is not important. Sort before checking.
+  std::sort(b, b + 16 * 16);
+  std::sort(b_ref, b_ref + 16 * 16);
+  EXPECT_EQ(0, memcmp(b, b_ref, sizeof(b)));
+}
+
+TEST_P(Hadamard16x16Test, VaryStride) {
+  DECLARE_ALIGNED(16, int16_t, a[16 * 16 * 8]);
+  DECLARE_ALIGNED(16, int16_t, b[16 * 16]);
+  int16_t b_ref[16 * 16];
+  for (int i = 0; i < 16 * 16 * 8; ++i) {
+    a[i] = rnd_.Rand9Signed();
+  }
+
+  for (int i = 8; i < 64; i += 8) {
+    memset(b, 0, sizeof(b));
+    memset(b_ref, 0, sizeof(b_ref));
+
+    reference_hadamard16x16(a, i, b_ref);
+    ASM_REGISTER_STATE_CHECK(h_func_(a, i, b));
+
+    // The order of the output is not important. Sort before checking.
+    std::sort(b, b + 16 * 16);
+    std::sort(b_ref, b_ref + 16 * 16);
+    EXPECT_EQ(0, memcmp(b, b_ref, sizeof(b)));
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(C, Hadamard16x16Test,
+                        ::testing::Values(&aom_hadamard_16x16_c));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, Hadamard16x16Test,
+                        ::testing::Values(&aom_hadamard_16x16_sse2));
+#endif  // HAVE_SSE2
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, Hadamard16x16Test,
+                        ::testing::Values(&aom_hadamard_16x16_neon));
+#endif  // HAVE_NEON
+}  // namespace
diff --git a/third_party/aom/test/hbd_metrics_test.cc b/third_party/aom/test/hbd_metrics_test.cc
new file mode 100644
index 0000000000..4def53b215
--- /dev/null
+++ b/third_party/aom/test/hbd_metrics_test.cc
@@ -0,0 +1,237 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <new>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/acm_random.h"
+#include "test/util.h"
+#include "./aom_config.h"
+#include "aom_dsp/psnr.h"
+#include "aom_dsp/ssim.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/msvc.h"
+#include "aom_scale/yv12config.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+
+typedef double (*LBDMetricFunc)(const YV12_BUFFER_CONFIG *source,
+                                const YV12_BUFFER_CONFIG *dest);
+typedef double (*HBDMetricFunc)(const YV12_BUFFER_CONFIG *source,
+                                const YV12_BUFFER_CONFIG *dest, uint32_t in_bd,
+                                uint32_t bd);
+
+double compute_hbd_psnr(const YV12_BUFFER_CONFIG *source,
+                        const YV12_BUFFER_CONFIG *dest, uint32_t in_bd,
+                        uint32_t bd) {
+  PSNR_STATS psnr;
+  aom_calc_highbd_psnr(source, dest, &psnr, bd, in_bd);
+  return psnr.psnr[0];
+}
+
+double compute_psnr(const YV12_BUFFER_CONFIG *source,
+                    const YV12_BUFFER_CONFIG *dest) {
+  PSNR_STATS psnr;
+  aom_calc_psnr(source, dest, &psnr);
+  return psnr.psnr[0];
+}
+
+double compute_hbd_psnrhvs(const YV12_BUFFER_CONFIG *source,
+                           const YV12_BUFFER_CONFIG *dest, uint32_t in_bd,
+                           uint32_t bd) {
+  double tempy, tempu, tempv;
+  return aom_psnrhvs(source, dest, &tempy, &tempu, &tempv, bd, in_bd);
+}
+
+double compute_psnrhvs(const YV12_BUFFER_CONFIG *source,
+                       const YV12_BUFFER_CONFIG *dest) {
+  double tempy, tempu, tempv;
+  return aom_psnrhvs(source, dest, &tempy, &tempu, &tempv, 8, 8);
+}
+
+double compute_hbd_fastssim(const YV12_BUFFER_CONFIG *source,
+                            const YV12_BUFFER_CONFIG *dest, uint32_t in_bd,
+                            uint32_t bd) {
+  double tempy, tempu, tempv;
+  return aom_calc_fastssim(source, dest, &tempy, &tempu, &tempv, bd, in_bd);
+}
+
+double compute_fastssim(const YV12_BUFFER_CONFIG *source,
+                        const YV12_BUFFER_CONFIG *dest) {
+  double tempy, tempu, tempv;
+  return aom_calc_fastssim(source, dest, &tempy, &tempu, &tempv, 8, 8);
+}
+
+double compute_hbd_aomssim(const YV12_BUFFER_CONFIG *source,
+                           const YV12_BUFFER_CONFIG *dest, uint32_t in_bd,
+                           uint32_t bd) {
+  double ssim, weight;
+  ssim = aom_highbd_calc_ssim(source, dest, &weight, bd, in_bd);
+  return 100 * pow(ssim / weight, 8.0);
+}
+
+double compute_aomssim(const YV12_BUFFER_CONFIG *source,
+                       const YV12_BUFFER_CONFIG *dest) {
+  double ssim, weight;
+  ssim = aom_calc_ssim(source, dest, &weight);
+  return 100 * pow(ssim / weight, 8.0);
+}
+
+class HBDMetricsTestBase {
+ public:
+  virtual ~HBDMetricsTestBase() {}
+
+ protected:
+  void RunAccuracyCheck() {
+    const int width = 1920;
+    const int height = 1080;
+    size_t i = 0;
+    const uint8_t kPixFiller = 128;
+    YV12_BUFFER_CONFIG lbd_src, lbd_dst;
+    YV12_BUFFER_CONFIG hbd_src, hbd_dst;
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    double lbd_db, hbd_db;
+
+    memset(&lbd_src, 0, sizeof(lbd_src));
+    memset(&lbd_dst, 0, sizeof(lbd_dst));
+    memset(&hbd_src, 0, sizeof(hbd_src));
+    memset(&hbd_dst, 0, sizeof(hbd_dst));
+
+    aom_alloc_frame_buffer(&lbd_src, width, height, 1, 1, 0, 32, 16);
+    aom_alloc_frame_buffer(&lbd_dst, width, height, 1, 1, 0, 32, 16);
+    aom_alloc_frame_buffer(&hbd_src, width, height, 1, 1, 1, 32, 16);
+    aom_alloc_frame_buffer(&hbd_dst, width, height, 1, 1, 1, 32, 16);
+
+    memset(lbd_src.buffer_alloc, kPixFiller, lbd_src.buffer_alloc_sz);
+    while (i < lbd_src.buffer_alloc_sz) {
+      uint16_t spel, dpel;
+      spel = lbd_src.buffer_alloc[i];
+      // Create some distortion for dst buffer.
+      dpel = rnd.Rand8();
+      lbd_dst.buffer_alloc[i] = (uint8_t)dpel;
+      ((uint16_t *)(hbd_src.buffer_alloc))[i] = spel << (bit_depth_ - 8);
+      ((uint16_t *)(hbd_dst.buffer_alloc))[i] = dpel << (bit_depth_ - 8);
+      i++;
+    }
+
+    lbd_db = lbd_metric_(&lbd_src, &lbd_dst);
+    hbd_db = hbd_metric_(&hbd_src, &hbd_dst, input_bit_depth_, bit_depth_);
+    EXPECT_LE(fabs(lbd_db - hbd_db), threshold_);
+
+    i = 0;
+    while (i < lbd_src.buffer_alloc_sz) {
+      uint16_t dpel;
+      // Create some small distortion for dst buffer.
+      dpel = 120 + (rnd.Rand8() >> 4);
+      lbd_dst.buffer_alloc[i] = (uint8_t)dpel;
+      ((uint16_t *)(hbd_dst.buffer_alloc))[i] = dpel << (bit_depth_ - 8);
+      i++;
+    }
+
+    lbd_db = lbd_metric_(&lbd_src, &lbd_dst);
+    hbd_db = hbd_metric_(&hbd_src, &hbd_dst, input_bit_depth_, bit_depth_);
+    EXPECT_LE(fabs(lbd_db - hbd_db), threshold_);
+
+    i = 0;
+    while (i < lbd_src.buffer_alloc_sz) {
+      uint16_t dpel;
+      // Create some small distortion for dst buffer.
+      dpel = 126 + (rnd.Rand8() >> 6);
+      lbd_dst.buffer_alloc[i] = (uint8_t)dpel;
+      ((uint16_t *)(hbd_dst.buffer_alloc))[i] = dpel << (bit_depth_ - 8);
+      i++;
+    }
+
+    lbd_db = lbd_metric_(&lbd_src, &lbd_dst);
+    hbd_db = hbd_metric_(&hbd_src, &hbd_dst, input_bit_depth_, bit_depth_);
+    EXPECT_LE(fabs(lbd_db - hbd_db), threshold_);
+
+    aom_free_frame_buffer(&lbd_src);
+    aom_free_frame_buffer(&lbd_dst);
+    aom_free_frame_buffer(&hbd_src);
+    aom_free_frame_buffer(&hbd_dst);
+  }
+
+  int input_bit_depth_;
+  int bit_depth_;
+  double threshold_;
+  LBDMetricFunc lbd_metric_;
+  HBDMetricFunc hbd_metric_;
+};
+
+typedef std::tr1::tuple<LBDMetricFunc, HBDMetricFunc, int, int, double>
+    MetricTestTParam;
+class HBDMetricsTest : public HBDMetricsTestBase,
+                       public ::testing::TestWithParam<MetricTestTParam> {
+ public:
+  virtual void SetUp() {
+    lbd_metric_ = GET_PARAM(0);
+    hbd_metric_ = GET_PARAM(1);
+    input_bit_depth_ = GET_PARAM(2);
+    bit_depth_ = GET_PARAM(3);
+    threshold_ = GET_PARAM(4);
+  }
+  virtual void TearDown() {}
+};
+
+TEST_P(HBDMetricsTest, RunAccuracyCheck) { RunAccuracyCheck(); }
+
+// Allow small variation due to floating point operations.
+static const double kSsim_thresh = 0.001;
+// Allow some additional errors accumulated in floating point operations.
+static const double kFSsim_thresh = 0.03;
+// Allow some extra variation due to rounding error accumulated in dct.
+static const double kPhvs_thresh = 0.3;
+
+INSTANTIATE_TEST_CASE_P(
+    AOMSSIM, HBDMetricsTest,
+    ::testing::Values(MetricTestTParam(&compute_aomssim, &compute_hbd_aomssim,
+                                       8, 10, kSsim_thresh),
+                      MetricTestTParam(&compute_aomssim, &compute_hbd_aomssim,
+                                       10, 10, kPhvs_thresh),
+                      MetricTestTParam(&compute_aomssim, &compute_hbd_aomssim,
+                                       8, 12, kSsim_thresh),
+                      MetricTestTParam(&compute_aomssim, &compute_hbd_aomssim,
+                                       12, 12, kPhvs_thresh)));
+INSTANTIATE_TEST_CASE_P(
+    FASTSSIM, HBDMetricsTest,
+    ::testing::Values(MetricTestTParam(&compute_fastssim, &compute_hbd_fastssim,
+                                       8, 10, kFSsim_thresh),
+                      MetricTestTParam(&compute_fastssim, &compute_hbd_fastssim,
+                                       10, 10, kFSsim_thresh),
+                      MetricTestTParam(&compute_fastssim, &compute_hbd_fastssim,
+                                       8, 12, kFSsim_thresh),
+                      MetricTestTParam(&compute_fastssim, &compute_hbd_fastssim,
+                                       12, 12, kFSsim_thresh)));
+INSTANTIATE_TEST_CASE_P(
+    PSNRHVS, HBDMetricsTest,
+    ::testing::Values(MetricTestTParam(&compute_psnrhvs, &compute_hbd_psnrhvs,
+                                       8, 10, kPhvs_thresh),
+                      MetricTestTParam(&compute_psnrhvs, &compute_hbd_psnrhvs,
+                                       10, 10, kPhvs_thresh),
+                      MetricTestTParam(&compute_psnrhvs, &compute_hbd_psnrhvs,
+                                       8, 12, kPhvs_thresh),
+                      MetricTestTParam(&compute_psnrhvs, &compute_hbd_psnrhvs,
+                                       12, 12, kPhvs_thresh)));
+INSTANTIATE_TEST_CASE_P(
+    PSNR, HBDMetricsTest,
+    ::testing::Values(
+        MetricTestTParam(&compute_psnr, &compute_hbd_psnr, 8, 10, kPhvs_thresh),
+        MetricTestTParam(&compute_psnr, &compute_hbd_psnr, 10, 10,
+                         kPhvs_thresh),
+        MetricTestTParam(&compute_psnr, &compute_hbd_psnr, 8, 12, kPhvs_thresh),
+        MetricTestTParam(&compute_psnr, &compute_hbd_psnr, 12, 12,
+                         kPhvs_thresh)));
+}  // namespace
diff --git a/third_party/aom/test/i420_video_source.h b/third_party/aom/test/i420_video_source.h
new file mode 100644
index 0000000000..0825296d73
--- /dev/null
+++ b/third_party/aom/test/i420_video_source.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef TEST_I420_VIDEO_SOURCE_H_
+#define TEST_I420_VIDEO_SOURCE_H_
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+
+#include "test/yuv_video_source.h"
+
+namespace libaom_test {
+
+// This class extends VideoSource to allow parsing of raw yv12
+// so that we can do actual file encodes.
+class I420VideoSource : public YUVVideoSource {
+ public:
+  I420VideoSource(const std::string &file_name, unsigned int width,
+                  unsigned int height, int rate_numerator, int rate_denominator,
+                  unsigned int start, int limit)
+      : YUVVideoSource(file_name, AOM_IMG_FMT_I420, width, height,
+                       rate_numerator, rate_denominator, start, limit) {}
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_I420_VIDEO_SOURCE_H_
diff --git a/third_party/aom/test/idct8x8_test.cc b/third_party/aom/test/idct8x8_test.cc
new file mode 100644
index 0000000000..f99a4075f8
--- /dev/null
+++ b/third_party/aom/test/idct8x8_test.cc
@@ -0,0 +1,86 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/msvc.h"  // for round()
+
+using libaom_test::ACMRandom;
+
+namespace {
+
+void reference_dct_1d(double input[8], double output[8]) {
+  const double kPi = 3.141592653589793238462643383279502884;
+  const double kInvSqrt2 = 0.707106781186547524400844362104;
+  for (int k = 0; k < 8; k++) {
+    output[k] = 0.0;
+    for (int n = 0; n < 8; n++)
+      output[k] += input[n] * cos(kPi * (2 * n + 1) * k / 16.0);
+    if (k == 0) output[k] = output[k] * kInvSqrt2;
+  }
+}
+
+void reference_dct_2d(int16_t input[64], double output[64]) {
+  // First transform columns
+  for (int i = 0; i < 8; ++i) {
+    double temp_in[8], temp_out[8];
+    for (int j = 0; j < 8; ++j) temp_in[j] = input[j * 8 + i];
+    reference_dct_1d(temp_in, temp_out);
+    for (int j = 0; j < 8; ++j) output[j * 8 + i] = temp_out[j];
+  }
+  // Then transform rows
+  for (int i = 0; i < 8; ++i) {
+    double temp_in[8], temp_out[8];
+    for (int j = 0; j < 8; ++j) temp_in[j] = output[j + i * 8];
+    reference_dct_1d(temp_in, temp_out);
+    for (int j = 0; j < 8; ++j) output[j + i * 8] = temp_out[j];
+  }
+  // Scale by some magic number
+  for (int i = 0; i < 64; ++i) output[i] *= 2;
+}
+
+TEST(AV1Idct8x8Test, AccuracyCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = 10000;
+  for (int i = 0; i < count_test_block; ++i) {
+    int16_t input[64];
+    tran_low_t coeff[64];
+    double output_r[64];
+    uint8_t dst[64], src[64];
+
+    for (int j = 0; j < 64; ++j) {
+      src[j] = rnd.Rand8();
+      dst[j] = rnd.Rand8();
+    }
+    // Initialize a test block with input range [-255, 255].
+    for (int j = 0; j < 64; ++j) input[j] = src[j] - dst[j];
+
+    reference_dct_2d(input, output_r);
+    for (int j = 0; j < 64; ++j)
+      coeff[j] = static_cast<tran_low_t>(round(output_r[j]));
+    aom_idct8x8_64_add_c(coeff, dst, 8);
+    for (int j = 0; j < 64; ++j) {
+      const int diff = dst[j] - src[j];
+      const int error = diff * diff;
+      EXPECT_GE(1, error) << "Error: 8x8 FDCT/IDCT has error " << error
+                          << " at index " << j;
+    }
+  }
+}
+
+}  // namespace
diff --git a/third_party/aom/test/idct_test.cc b/third_party/aom/test/idct_test.cc
new file mode 100644
index 0000000000..a880a91825
--- /dev/null
+++ b/third_party/aom/test/idct_test.cc
@@ -0,0 +1,122 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "./aom_config.h"
+#include "./aom_rtcd.h"
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "aom/aom_integer.h"
+
+typedef void (*IdctFunc)(int16_t *input, unsigned char *pred_ptr,
+                         int pred_stride, unsigned char *dst_ptr,
+                         int dst_stride);
+namespace {
+class IDCTTest : public ::testing::TestWithParam<IdctFunc> {
+ protected:
+  virtual void SetUp() {
+    int i;
+
+    UUT = GetParam();
+    memset(input, 0, sizeof(input));
+    /* Set up guard blocks */
+    for (i = 0; i < 256; i++) output[i] = ((i & 0xF) < 4 && (i < 64)) ? 0 : -1;
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+  IdctFunc UUT;
+  int16_t input[16];
+  unsigned char output[256];
+  unsigned char predict[256];
+};
+
+TEST_P(IDCTTest, TestGuardBlocks) {
+  int i;
+
+  for (i = 0; i < 256; i++)
+    if ((i & 0xF) < 4 && i < 64)
+      EXPECT_EQ(0, output[i]) << i;
+    else
+      EXPECT_EQ(255, output[i]);
+}
+
+TEST_P(IDCTTest, TestAllZeros) {
+  int i;
+
+  ASM_REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
+
+  for (i = 0; i < 256; i++)
+    if ((i & 0xF) < 4 && i < 64)
+      EXPECT_EQ(0, output[i]) << "i==" << i;
+    else
+      EXPECT_EQ(255, output[i]) << "i==" << i;
+}
+
+TEST_P(IDCTTest, TestAllOnes) {
+  int i;
+
+  input[0] = 4;
+  ASM_REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
+
+  for (i = 0; i < 256; i++)
+    if ((i & 0xF) < 4 && i < 64)
+      EXPECT_EQ(1, output[i]) << "i==" << i;
+    else
+      EXPECT_EQ(255, output[i]) << "i==" << i;
+}
+
+TEST_P(IDCTTest, TestAddOne) {
+  int i;
+
+  for (i = 0; i < 256; i++) predict[i] = i;
+  input[0] = 4;
+  ASM_REGISTER_STATE_CHECK(UUT(input, predict, 16, output, 16));
+
+  for (i = 0; i < 256; i++)
+    if ((i & 0xF) < 4 && i < 64)
+      EXPECT_EQ(i + 1, output[i]) << "i==" << i;
+    else
+      EXPECT_EQ(255, output[i]) << "i==" << i;
+}
+
+TEST_P(IDCTTest, TestWithData) {
+  int i;
+
+  for (i = 0; i < 16; i++) input[i] = i;
+
+  ASM_REGISTER_STATE_CHECK(UUT(input, output, 16, output, 16));
+
+  for (i = 0; i < 256; i++)
+    if ((i & 0xF) > 3 || i > 63)
+      EXPECT_EQ(255, output[i]) << "i==" << i;
+    else if (i == 0)
+      EXPECT_EQ(11, output[i]) << "i==" << i;
+    else if (i == 34)
+      EXPECT_EQ(1, output[i]) << "i==" << i;
+    else if (i == 2 || i == 17 || i == 32)
+      EXPECT_EQ(3, output[i]) << "i==" << i;
+    else
+      EXPECT_EQ(0, output[i]) << "i==" << i;
+}
+
+INSTANTIATE_TEST_CASE_P(C, IDCTTest, ::testing::Values(aom_short_idct4x4llm_c));
+#if HAVE_MMX
+INSTANTIATE_TEST_CASE_P(MMX, IDCTTest,
+                        ::testing::Values(aom_short_idct4x4llm_mmx));
+#endif
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(MSA, IDCTTest,
+                        ::testing::Values(aom_short_idct4x4llm_msa));
+#endif
+}
diff --git a/third_party/aom/test/intrabc_test.cc b/third_party/aom/test/intrabc_test.cc
new file mode 100644
index 0000000000..84cfa5c485
--- /dev/null
+++ b/third_party/aom/test/intrabc_test.cc
@@ -0,0 +1,157 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "av1/common/enums.h"
+#include "av1/common/mv.h"
+#include "av1/common/mvref_common.h"
+#include "av1/common/tile_common.h"
+
+namespace {
+TEST(IntrabcTest, DvValidation) {
+  struct DvTestCase {
+    MV dv;
+    int mi_row_offset;
+    int mi_col_offset;
+    BLOCK_SIZE bsize;
+    bool valid;
+  };
+  const int kSubPelScale = 8;
+  const int kTileMaxMibWidth = 8;
+  const DvTestCase kDvCases[] = {
+#if CONFIG_EXT_PARTITION
+    { { 0, 0 }, 0, 0, BLOCK_128X128, false },
+#endif
+    { { 0, 0 }, 0, 0, BLOCK_64X64, false },
+    { { 0, 0 }, 0, 0, BLOCK_32X32, false },
+    { { 0, 0 }, 0, 0, BLOCK_16X16, false },
+    { { 0, 0 }, 0, 0, BLOCK_8X8, false },
+    { { 0, 0 }, 0, 0, BLOCK_4X4, false },
+    { { -MAX_SB_SIZE * kSubPelScale, -MAX_SB_SIZE * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_16X16,
+      true },
+    { { 0, -MAX_SB_SIZE * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_16X16,
+      true },
+    { { -MAX_SB_SIZE * kSubPelScale, 0 },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_16X16,
+      true },
+    { { MAX_SB_SIZE * kSubPelScale, 0 },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_16X16,
+      false },
+    { { 0, MAX_SB_SIZE * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_16X16,
+      false },
+    { { -32 * kSubPelScale, -32 * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_32X32,
+      true },
+    { { -32 * kSubPelScale, -32 * kSubPelScale },
+      32 / MI_SIZE,
+      32 / MI_SIZE,
+      BLOCK_32X32,
+      false },
+    { { -32 * kSubPelScale - kSubPelScale / 2, -32 * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_32X32,
+      false },
+    { { -33 * kSubPelScale, -32 * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_32X32,
+      true },
+    { { -32 * kSubPelScale, -32 * kSubPelScale - kSubPelScale / 2 },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_32X32,
+      false },
+    { { -32 * kSubPelScale, -33 * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_32X32,
+      true },
+    { { -MAX_SB_SIZE * kSubPelScale, -MAX_SB_SIZE * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_LARGEST,
+      true },
+    { { -(MAX_SB_SIZE + 1) * kSubPelScale, -MAX_SB_SIZE * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_LARGEST,
+      false },
+    { { -MAX_SB_SIZE * kSubPelScale, -(MAX_SB_SIZE + 1) * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_LARGEST,
+      false },
+    { { -(MAX_SB_SIZE - 1) * kSubPelScale, -MAX_SB_SIZE * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_LARGEST,
+      true },
+    { { -MAX_SB_SIZE * kSubPelScale, -(MAX_SB_SIZE - 1) * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_LARGEST,
+      true },
+    { { -(MAX_SB_SIZE - 1) * kSubPelScale, -(MAX_SB_SIZE - 1) * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_LARGEST,
+      false },
+    { { -MAX_SB_SIZE * kSubPelScale, MAX_SB_SIZE * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_LARGEST,
+      true },
+    { { -MAX_SB_SIZE * kSubPelScale,
+        (kTileMaxMibWidth - 2) * MAX_SB_SIZE * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_LARGEST,
+      true },
+    { { -MAX_SB_SIZE * kSubPelScale,
+        ((kTileMaxMibWidth - 2) * MAX_SB_SIZE + 1) * kSubPelScale },
+      MAX_SB_SIZE / MI_SIZE,
+      MAX_SB_SIZE / MI_SIZE,
+      BLOCK_LARGEST,
+      false },
+  };
+  TileInfo tile;
+  tile.mi_row_start = 8 * MAX_MIB_SIZE;
+  tile.mi_row_end = 16 * MAX_MIB_SIZE;
+  tile.mi_col_start = 24 * MAX_MIB_SIZE;
+  tile.mi_col_end = tile.mi_col_start + kTileMaxMibWidth * MAX_MIB_SIZE;
+  for (int i = 0; i < static_cast<int>(GTEST_ARRAY_SIZE_(kDvCases)); ++i) {
+    EXPECT_EQ(kDvCases[i].valid,
+              is_dv_valid(kDvCases[i].dv, &tile,
+                          tile.mi_row_start + kDvCases[i].mi_row_offset,
+                          tile.mi_col_start + kDvCases[i].mi_col_offset,
+                          kDvCases[i].bsize))
+        << "DvCases[" << i << "]";
+  }
+}
+}  // namespace
diff --git a/third_party/aom/test/intrapred_test.cc b/third_party/aom/test/intrapred_test.cc
new file mode 100644
index 0000000000..4efed57b6b
--- /dev/null
+++ b/third_party/aom/test/intrapred_test.cc
@@ -0,0 +1,231 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <string>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "av1/common/blockd.h"
+#include "av1/common/pred_common.h"
+#include "aom_mem/aom_mem.h"
+
+namespace {
+
+using libaom_test::ACMRandom;
+
+const int count_test_block = 100000;
+
+typedef void (*IntraPred)(uint16_t *dst, ptrdiff_t stride,
+                          const uint16_t *above, const uint16_t *left, int bps);
+
+struct IntraPredFunc {
+  IntraPredFunc(IntraPred pred = NULL, IntraPred ref = NULL,
+                int block_size_value = 0, int bit_depth_value = 0)
+      : pred_fn(pred), ref_fn(ref), block_size(block_size_value),
+        bit_depth(bit_depth_value) {}
+
+  IntraPred pred_fn;
+  IntraPred ref_fn;
+  int block_size;
+  int bit_depth;
+};
+
+class AV1IntraPredTest : public ::testing::TestWithParam<IntraPredFunc> {
+ public:
+  void RunTest(uint16_t *left_col, uint16_t *above_data, uint16_t *dst,
+               uint16_t *ref_dst) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int block_size = params_.block_size;
+    above_row_ = above_data + 16;
+    left_col_ = left_col;
+    dst_ = dst;
+    ref_dst_ = ref_dst;
+    int error_count = 0;
+    for (int i = 0; i < count_test_block; ++i) {
+      // Fill edges with random data, try first with saturated values.
+      for (int x = -1; x <= block_size * 2; x++) {
+        if (i == 0) {
+          above_row_[x] = mask_;
+        } else {
+          above_row_[x] = rnd.Rand16() & mask_;
+        }
+      }
+      for (int y = 0; y < block_size; y++) {
+        if (i == 0) {
+          left_col_[y] = mask_;
+        } else {
+          left_col_[y] = rnd.Rand16() & mask_;
+        }
+      }
+      Predict();
+      CheckPrediction(i, &error_count);
+    }
+    ASSERT_EQ(0, error_count);
+  }
+
+ protected:
+  virtual void SetUp() {
+    params_ = GetParam();
+    stride_ = params_.block_size * 3;
+    mask_ = (1 << params_.bit_depth) - 1;
+  }
+
+  void Predict() {
+    const int bit_depth = params_.bit_depth;
+    params_.ref_fn(ref_dst_, stride_, above_row_, left_col_, bit_depth);
+    ASM_REGISTER_STATE_CHECK(
+        params_.pred_fn(dst_, stride_, above_row_, left_col_, bit_depth));
+  }
+
+  void CheckPrediction(int test_case_number, int *error_count) const {
+    // For each pixel ensure that the calculated value is the same as reference.
+    const int block_size = params_.block_size;
+    for (int y = 0; y < block_size; y++) {
+      for (int x = 0; x < block_size; x++) {
+        *error_count += ref_dst_[x + y * stride_] != dst_[x + y * stride_];
+        if (*error_count == 1) {
+          ASSERT_EQ(ref_dst_[x + y * stride_], dst_[x + y * stride_])
+              << " Failed on Test Case Number " << test_case_number;
+        }
+      }
+    }
+  }
+
+  uint16_t *above_row_;
+  uint16_t *left_col_;
+  uint16_t *dst_;
+  uint16_t *ref_dst_;
+  ptrdiff_t stride_;
+  int mask_;
+
+  IntraPredFunc params_;
+};
+
+TEST_P(AV1IntraPredTest, IntraPredTests) {
+  // max block size is 32
+  DECLARE_ALIGNED(16, uint16_t, left_col[2 * 32]);
+  DECLARE_ALIGNED(16, uint16_t, above_data[2 * 32 + 32]);
+  DECLARE_ALIGNED(16, uint16_t, dst[3 * 32 * 32]);
+  DECLARE_ALIGNED(16, uint16_t, ref_dst[3 * 32 * 32]);
+  RunTest(left_col, above_data, dst, ref_dst);
+}
+
+#if HAVE_SSE2
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE2_TO_C_8, AV1IntraPredTest,
+    ::testing::Values(IntraPredFunc(&aom_highbd_dc_predictor_32x32_sse2,
+                                    &aom_highbd_dc_predictor_32x32_c, 32, 8),
+#if !CONFIG_ALT_INTRA
+                      IntraPredFunc(&aom_highbd_tm_predictor_16x16_sse2,
+                                    &aom_highbd_tm_predictor_16x16_c, 16, 8),
+                      IntraPredFunc(&aom_highbd_tm_predictor_32x32_sse2,
+                                    &aom_highbd_tm_predictor_32x32_c, 32, 8),
+#endif  // !CONFIG_ALT_INTRA
+
+                      IntraPredFunc(&aom_highbd_dc_predictor_4x4_sse2,
+                                    &aom_highbd_dc_predictor_4x4_c, 4, 8),
+                      IntraPredFunc(&aom_highbd_dc_predictor_8x8_sse2,
+                                    &aom_highbd_dc_predictor_8x8_c, 8, 8),
+                      IntraPredFunc(&aom_highbd_dc_predictor_16x16_sse2,
+                                    &aom_highbd_dc_predictor_16x16_c, 16, 8),
+                      IntraPredFunc(&aom_highbd_v_predictor_4x4_sse2,
+                                    &aom_highbd_v_predictor_4x4_c, 4, 8),
+                      IntraPredFunc(&aom_highbd_v_predictor_8x8_sse2,
+                                    &aom_highbd_v_predictor_8x8_c, 8, 8),
+                      IntraPredFunc(&aom_highbd_v_predictor_16x16_sse2,
+                                    &aom_highbd_v_predictor_16x16_c, 16, 8),
+                      IntraPredFunc(&aom_highbd_v_predictor_32x32_sse2,
+                                    &aom_highbd_v_predictor_32x32_c, 32, 8)
+#if !CONFIG_ALT_INTRA
+                          ,
+                      IntraPredFunc(&aom_highbd_tm_predictor_4x4_sse2,
+                                    &aom_highbd_tm_predictor_4x4_c, 4, 8),
+                      IntraPredFunc(&aom_highbd_tm_predictor_8x8_sse2,
+                                    &aom_highbd_tm_predictor_8x8_c, 8, 8)
+#endif  // !CONFIG_ALT_INTRA
+                          ));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2_TO_C_10, AV1IntraPredTest,
+    ::testing::Values(IntraPredFunc(&aom_highbd_dc_predictor_32x32_sse2,
+                                    &aom_highbd_dc_predictor_32x32_c, 32, 10),
+#if !CONFIG_ALT_INTRA
+                      IntraPredFunc(&aom_highbd_tm_predictor_16x16_sse2,
+                                    &aom_highbd_tm_predictor_16x16_c, 16, 10),
+                      IntraPredFunc(&aom_highbd_tm_predictor_32x32_sse2,
+                                    &aom_highbd_tm_predictor_32x32_c, 32, 10),
+#endif  // !CONFIG_ALT_INTRA
+                      IntraPredFunc(&aom_highbd_dc_predictor_4x4_sse2,
+                                    &aom_highbd_dc_predictor_4x4_c, 4, 10),
+                      IntraPredFunc(&aom_highbd_dc_predictor_8x8_sse2,
+                                    &aom_highbd_dc_predictor_8x8_c, 8, 10),
+                      IntraPredFunc(&aom_highbd_dc_predictor_16x16_sse2,
+                                    &aom_highbd_dc_predictor_16x16_c, 16, 10),
+                      IntraPredFunc(&aom_highbd_v_predictor_4x4_sse2,
+                                    &aom_highbd_v_predictor_4x4_c, 4, 10),
+                      IntraPredFunc(&aom_highbd_v_predictor_8x8_sse2,
+                                    &aom_highbd_v_predictor_8x8_c, 8, 10),
+                      IntraPredFunc(&aom_highbd_v_predictor_16x16_sse2,
+                                    &aom_highbd_v_predictor_16x16_c, 16, 10),
+                      IntraPredFunc(&aom_highbd_v_predictor_32x32_sse2,
+                                    &aom_highbd_v_predictor_32x32_c, 32, 10)
+#if !CONFIG_ALT_INTRA
+                          ,
+                      IntraPredFunc(&aom_highbd_tm_predictor_4x4_sse2,
+                                    &aom_highbd_tm_predictor_4x4_c, 4, 10),
+                      IntraPredFunc(&aom_highbd_tm_predictor_8x8_sse2,
+                                    &aom_highbd_tm_predictor_8x8_c, 8, 10)
+#endif  // !CONFIG_ALT_INTRA
+                          ));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2_TO_C_12, AV1IntraPredTest,
+    ::testing::Values(IntraPredFunc(&aom_highbd_dc_predictor_32x32_sse2,
+                                    &aom_highbd_dc_predictor_32x32_c, 32, 12),
+#if !CONFIG_ALT_INTRA
+                      IntraPredFunc(&aom_highbd_tm_predictor_16x16_sse2,
+                                    &aom_highbd_tm_predictor_16x16_c, 16, 12),
+                      IntraPredFunc(&aom_highbd_tm_predictor_32x32_sse2,
+                                    &aom_highbd_tm_predictor_32x32_c, 32, 12),
+#endif  // !CONFIG_ALT_INTRA
+                      IntraPredFunc(&aom_highbd_dc_predictor_4x4_sse2,
+                                    &aom_highbd_dc_predictor_4x4_c, 4, 12),
+                      IntraPredFunc(&aom_highbd_dc_predictor_8x8_sse2,
+                                    &aom_highbd_dc_predictor_8x8_c, 8, 12),
+                      IntraPredFunc(&aom_highbd_dc_predictor_16x16_sse2,
+                                    &aom_highbd_dc_predictor_16x16_c, 16, 12),
+                      IntraPredFunc(&aom_highbd_v_predictor_4x4_sse2,
+                                    &aom_highbd_v_predictor_4x4_c, 4, 12),
+                      IntraPredFunc(&aom_highbd_v_predictor_8x8_sse2,
+                                    &aom_highbd_v_predictor_8x8_c, 8, 12),
+                      IntraPredFunc(&aom_highbd_v_predictor_16x16_sse2,
+                                    &aom_highbd_v_predictor_16x16_c, 16, 12),
+                      IntraPredFunc(&aom_highbd_v_predictor_32x32_sse2,
+                                    &aom_highbd_v_predictor_32x32_c, 32, 12)
+#if !CONFIG_ALT_INTRA
+                          ,
+                      IntraPredFunc(&aom_highbd_tm_predictor_4x4_sse2,
+                                    &aom_highbd_tm_predictor_4x4_c, 4, 12),
+                      IntraPredFunc(&aom_highbd_tm_predictor_8x8_sse2,
+                                    &aom_highbd_tm_predictor_8x8_c, 8, 12)
+#endif  // !CONFIG_ALT_INTRA
+                          ));
+
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // HAVE_SSE2
+}  // namespace
diff --git a/third_party/aom/test/ivf_video_source.h b/third_party/aom/test/ivf_video_source.h
new file mode 100644
index 0000000000..0d3e9f9cb4
--- /dev/null
+++ b/third_party/aom/test/ivf_video_source.h
@@ -0,0 +1,107 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef TEST_IVF_VIDEO_SOURCE_H_
+#define TEST_IVF_VIDEO_SOURCE_H_
+#include <cstdio>
+#include <cstdlib>
+#include <new>
+#include <string>
+#include "test/video_source.h"
+
+namespace libaom_test {
+const unsigned int kCodeBufferSize = 256 * 1024;
+const unsigned int kIvfFileHdrSize = 32;
+const unsigned int kIvfFrameHdrSize = 12;
+
+static unsigned int MemGetLe32(const uint8_t *mem) {
+  return (mem[3] << 24) | (mem[2] << 16) | (mem[1] << 8) | (mem[0]);
+}
+
+// This class extends VideoSource to allow parsing of ivf files,
+// so that we can do actual file decodes.
+class IVFVideoSource : public CompressedVideoSource {
+ public:
+  explicit IVFVideoSource(const std::string &file_name)
+      : file_name_(file_name), input_file_(NULL), compressed_frame_buf_(NULL),
+        frame_sz_(0), frame_(0), end_of_file_(false) {}
+
+  virtual ~IVFVideoSource() {
+    delete[] compressed_frame_buf_;
+
+    if (input_file_) fclose(input_file_);
+  }
+
+  virtual void Init() {
+    // Allocate a buffer for read in the compressed video frame.
+    compressed_frame_buf_ = new uint8_t[libaom_test::kCodeBufferSize];
+    ASSERT_TRUE(compressed_frame_buf_ != NULL)
+        << "Allocate frame buffer failed";
+  }
+
+  virtual void Begin() {
+    input_file_ = OpenTestDataFile(file_name_);
+    ASSERT_TRUE(input_file_ != NULL) << "Input file open failed. Filename: "
+                                     << file_name_;
+
+    // Read file header
+    uint8_t file_hdr[kIvfFileHdrSize];
+    ASSERT_EQ(kIvfFileHdrSize, fread(file_hdr, 1, kIvfFileHdrSize, input_file_))
+        << "File header read failed.";
+    // Check file header
+    ASSERT_TRUE(file_hdr[0] == 'D' && file_hdr[1] == 'K' &&
+                file_hdr[2] == 'I' && file_hdr[3] == 'F')
+        << "Input is not an IVF file.";
+
+    FillFrame();
+  }
+
+  virtual void Next() {
+    ++frame_;
+    FillFrame();
+  }
+
+  void FillFrame() {
+    ASSERT_TRUE(input_file_ != NULL);
+    uint8_t frame_hdr[kIvfFrameHdrSize];
+    // Check frame header and read a frame from input_file.
+    if (fread(frame_hdr, 1, kIvfFrameHdrSize, input_file_) !=
+        kIvfFrameHdrSize) {
+      end_of_file_ = true;
+    } else {
+      end_of_file_ = false;
+
+      frame_sz_ = MemGetLe32(frame_hdr);
+      ASSERT_LE(frame_sz_, kCodeBufferSize)
+          << "Frame is too big for allocated code buffer";
+      ASSERT_EQ(frame_sz_,
+                fread(compressed_frame_buf_, 1, frame_sz_, input_file_))
+          << "Failed to read complete frame";
+    }
+  }
+
+  virtual const uint8_t *cxdata() const {
+    return end_of_file_ ? NULL : compressed_frame_buf_;
+  }
+  virtual size_t frame_size() const { return frame_sz_; }
+  virtual unsigned int frame_number() const { return frame_; }
+
+ protected:
+  std::string file_name_;
+  FILE *input_file_;
+  uint8_t *compressed_frame_buf_;
+  size_t frame_sz_;
+  unsigned int frame_;
+  bool end_of_file_;
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_IVF_VIDEO_SOURCE_H_
diff --git a/third_party/aom/test/level_test.cc b/third_party/aom/test/level_test.cc
new file mode 100644
index 0000000000..1049d4901e
--- /dev/null
+++ b/third_party/aom/test/level_test.cc
@@ -0,0 +1,117 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+class LevelTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<libaom_test::TestMode, int> {
+ protected:
+  LevelTest()
+      : EncoderTest(GET_PARAM(0)), encoding_mode_(GET_PARAM(1)),
+        cpu_used_(GET_PARAM(2)), min_gf_internal_(24), target_level_(0),
+        level_(0) {}
+  virtual ~LevelTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+    if (encoding_mode_ != ::libaom_test::kRealTime) {
+      cfg_.g_lag_in_frames = 25;
+      cfg_.rc_end_usage = AOM_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = AOM_CBR;
+    }
+    cfg_.rc_2pass_vbr_minsection_pct = 5;
+    cfg_.rc_2pass_vbr_maxsection_pct = 2000;
+    cfg_.rc_target_bitrate = 400;
+    cfg_.rc_max_quantizer = 63;
+    cfg_.rc_min_quantizer = 0;
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      encoder->Control(AOME_SET_CPUUSED, cpu_used_);
+      encoder->Control(AV1E_SET_TARGET_LEVEL, target_level_);
+      encoder->Control(AV1E_SET_MIN_GF_INTERVAL, min_gf_internal_);
+      if (encoding_mode_ != ::libaom_test::kRealTime) {
+        encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(AOME_SET_ARNR_STRENGTH, 5);
+      }
+    }
+    encoder->Control(AV1E_GET_LEVEL, &level_);
+    ASSERT_LE(level_, 51);
+    ASSERT_GE(level_, 0);
+  }
+
+  ::libaom_test::TestMode encoding_mode_;
+  int cpu_used_;
+  int min_gf_internal_;
+  int target_level_;
+  int level_;
+};
+
+// Test for keeping level stats only
+TEST_P(LevelTest, TestTargetLevel0) {
+  ::libaom_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       40);
+  target_level_ = 0;
+  min_gf_internal_ = 4;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  ASSERT_EQ(11, level_);
+
+  cfg_.rc_target_bitrate = 1600;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  ASSERT_EQ(20, level_);
+}
+
+// Test for level control being turned off
+TEST_P(LevelTest, TestTargetLevel255) {
+  ::libaom_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0,
+                                       30);
+  target_level_ = 255;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+TEST_P(LevelTest, TestTargetLevelApi) {
+  ::libaom_test::I420VideoSource video("hantro_odd.yuv", 208, 144, 30, 1, 0, 1);
+  static const aom_codec_iface_t *codec = &aom_codec_av1_cx_algo;
+  aom_codec_ctx_t enc;
+  aom_codec_enc_cfg_t cfg;
+  EXPECT_EQ(AOM_CODEC_OK, aom_codec_enc_config_default(codec, &cfg, 0));
+  EXPECT_EQ(AOM_CODEC_OK, aom_codec_enc_init(&enc, codec, &cfg, 0));
+  for (int level = 0; level <= 256; ++level) {
+    if (level == 10 || level == 11 || level == 20 || level == 21 ||
+        level == 30 || level == 31 || level == 40 || level == 41 ||
+        level == 50 || level == 51 || level == 52 || level == 60 ||
+        level == 61 || level == 62 || level == 0 || level == 255)
+      EXPECT_EQ(AOM_CODEC_OK,
+                aom_codec_control(&enc, AV1E_SET_TARGET_LEVEL, level));
+    else
+      EXPECT_EQ(AOM_CODEC_INVALID_PARAM,
+                aom_codec_control(&enc, AV1E_SET_TARGET_LEVEL, level));
+  }
+  EXPECT_EQ(AOM_CODEC_OK, aom_codec_destroy(&enc));
+}
+
+AV1_INSTANTIATE_TEST_CASE(LevelTest,
+                          ::testing::Values(::libaom_test::kTwoPassGood,
+                                            ::libaom_test::kOnePassGood),
+                          ::testing::Range(0, 9));
+}  // namespace
diff --git a/third_party/aom/test/lossless_test.cc b/third_party/aom/test/lossless_test.cc
new file mode 100644
index 0000000000..5c5b32d932
--- /dev/null
+++ b/third_party/aom/test/lossless_test.cc
@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+
+namespace {
+
+const int kMaxPsnr = 100;
+
+class LosslessTestLarge
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWithParam<libaom_test::TestMode> {
+ protected:
+  LosslessTestLarge()
+      : EncoderTest(GET_PARAM(0)), psnr_(kMaxPsnr), nframes_(0),
+        encoding_mode_(GET_PARAM(1)) {}
+
+  virtual ~LosslessTestLarge() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      // Only call Control if quantizer > 0 to verify that using quantizer
+      // alone will activate lossless
+      if (cfg_.rc_max_quantizer > 0 || cfg_.rc_min_quantizer > 0) {
+        encoder->Control(AV1E_SET_LOSSLESS, 1);
+      }
+    }
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    psnr_ = kMaxPsnr;
+    nframes_ = 0;
+  }
+
+  virtual void PSNRPktHook(const aom_codec_cx_pkt_t *pkt) {
+    if (pkt->data.psnr.psnr[0] < psnr_) psnr_ = pkt->data.psnr.psnr[0];
+  }
+
+  double GetMinPsnr() const { return psnr_; }
+
+ private:
+  double psnr_;
+  unsigned int nframes_;
+  libaom_test::TestMode encoding_mode_;
+};
+
+TEST_P(LosslessTestLarge, TestLossLessEncoding) {
+  const aom_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.g_lag_in_frames = 25;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 0;
+
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  // intentionally changed the dimension for better testing coverage
+  libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 5);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_lossless = GetMinPsnr();
+  EXPECT_GE(psnr_lossless, kMaxPsnr);
+}
+
+TEST_P(LosslessTestLarge, TestLossLessEncoding444) {
+  libaom_test::Y4mVideoSource video("rush_hour_444.y4m", 0, 5);
+
+  cfg_.g_profile = 1;
+  cfg_.g_timebase = video.timebase();
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.g_lag_in_frames = 25;
+  cfg_.rc_min_quantizer = 0;
+  cfg_.rc_max_quantizer = 0;
+
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_lossless = GetMinPsnr();
+  EXPECT_GE(psnr_lossless, kMaxPsnr);
+}
+
+TEST_P(LosslessTestLarge, TestLossLessEncodingCtrl) {
+  const aom_rational timebase = { 33333333, 1000000000 };
+  cfg_.g_timebase = timebase;
+  cfg_.rc_target_bitrate = 2000;
+  cfg_.g_lag_in_frames = 25;
+  // Intentionally set Q > 0, to make sure control can be used to activate
+  // lossless
+  cfg_.rc_min_quantizer = 10;
+  cfg_.rc_max_quantizer = 20;
+
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                     timebase.den, timebase.num, 0, 5);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  const double psnr_lossless = GetMinPsnr();
+  EXPECT_GE(psnr_lossless, kMaxPsnr);
+}
+
+AV1_INSTANTIATE_TEST_CASE(LosslessTestLarge,
+                          ::testing::Values(::libaom_test::kOnePassGood,
+                                            ::libaom_test::kTwoPassGood));
+}  // namespace
diff --git a/third_party/aom/test/lpf_8_test.cc b/third_party/aom/test/lpf_8_test.cc
new file mode 100644
index 0000000000..cee0d3b814
--- /dev/null
+++ b/third_party/aom/test/lpf_8_test.cc
@@ -0,0 +1,624 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <cmath>
+#include <cstdlib>
+#include <string>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "av1/common/av1_loopfilter.h"
+#include "av1/common/entropy.h"
+#include "aom/aom_integer.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+// Horizontally and Vertically need 32x32: 8  Coeffs preceeding filtered section
+//                                         16 Coefs within filtered section
+//                                         8  Coeffs following filtered section
+const int kNumCoeffs = 1024;
+
+const int number_of_iterations = 10000;
+
+#if CONFIG_HIGHBITDEPTH
+typedef void (*loop_op_t)(uint16_t *s, int p, const uint8_t *blimit,
+                          const uint8_t *limit, const uint8_t *thresh, int bd);
+typedef void (*dual_loop_op_t)(uint16_t *s, int p, const uint8_t *blimit0,
+                               const uint8_t *limit0, const uint8_t *thresh0,
+                               const uint8_t *blimit1, const uint8_t *limit1,
+                               const uint8_t *thresh1, int bd);
+#else
+typedef void (*loop_op_t)(uint8_t *s, int p, const uint8_t *blimit,
+                          const uint8_t *limit, const uint8_t *thresh);
+typedef void (*dual_loop_op_t)(uint8_t *s, int p, const uint8_t *blimit0,
+                               const uint8_t *limit0, const uint8_t *thresh0,
+                               const uint8_t *blimit1, const uint8_t *limit1,
+                               const uint8_t *thresh1);
+#endif  // CONFIG_HIGHBITDEPTH
+
+typedef std::tr1::tuple<loop_op_t, loop_op_t, int> loop8_param_t;
+typedef std::tr1::tuple<dual_loop_op_t, dual_loop_op_t, int> dualloop8_param_t;
+
+class Loop8Test6Param : public ::testing::TestWithParam<loop8_param_t> {
+ public:
+  virtual ~Loop8Test6Param() {}
+  virtual void SetUp() {
+    loopfilter_op_ = GET_PARAM(0);
+    ref_loopfilter_op_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  int bit_depth_;
+  int mask_;
+  loop_op_t loopfilter_op_;
+  loop_op_t ref_loopfilter_op_;
+};
+
+class Loop8Test9Param : public ::testing::TestWithParam<dualloop8_param_t> {
+ public:
+  virtual ~Loop8Test9Param() {}
+  virtual void SetUp() {
+    loopfilter_op_ = GET_PARAM(0);
+    ref_loopfilter_op_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+    mask_ = (1 << bit_depth_) - 1;
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  int bit_depth_;
+  int mask_;
+  dual_loop_op_t loopfilter_op_;
+  dual_loop_op_t ref_loopfilter_op_;
+};
+
+TEST_P(Loop8Test6Param, OperationCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = number_of_iterations;
+#if CONFIG_HIGHBITDEPTH
+  int32_t bd = bit_depth_;
+  DECLARE_ALIGNED(16, uint16_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, ref_s[kNumCoeffs]);
+#else
+  DECLARE_ALIGNED(8, uint8_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(8, uint8_t, ref_s[kNumCoeffs]);
+#endif  // CONFIG_HIGHBITDEPTH
+  int err_count_total = 0;
+  int first_failure = -1;
+  for (int i = 0; i < count_test_block; ++i) {
+    int err_count = 0;
+    uint8_t tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    blimit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                    tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    limit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                   tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t,
+                    thresh[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                    tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    int32_t p = kNumCoeffs / 32;
+
+    uint16_t tmp_s[kNumCoeffs];
+    int j = 0;
+    while (j < kNumCoeffs) {
+      uint8_t val = rnd.Rand8();
+      if (val & 0x80) {  // 50% chance to choose a new value.
+        tmp_s[j] = rnd.Rand16();
+        j++;
+      } else {  // 50% chance to repeat previous value in row X times
+        int k = 0;
+        while (k++ < ((val & 0x1f) + 1) && j < kNumCoeffs) {
+          if (j < 1) {
+            tmp_s[j] = rnd.Rand16();
+          } else if (val & 0x20) {  // Increment by an value within the limit
+            tmp_s[j] = (tmp_s[j - 1] + (*limit - 1));
+          } else {  // Decrement by an value within the limit
+            tmp_s[j] = (tmp_s[j - 1] - (*limit - 1));
+          }
+          j++;
+        }
+      }
+    }
+    for (j = 0; j < kNumCoeffs; j++) {
+      if (i % 2) {
+        s[j] = tmp_s[j] & mask_;
+      } else {
+        s[j] = tmp_s[p * (j % p) + j / p] & mask_;
+      }
+      ref_s[j] = s[j];
+    }
+#if CONFIG_HIGHBITDEPTH
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit, limit, thresh, bd);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh, bd));
+#else
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit, limit, thresh);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh));
+#endif  // CONFIG_HIGHBITDEPTH
+
+    for (j = 0; j < kNumCoeffs; ++j) {
+      err_count += ref_s[j] != s[j];
+    }
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Loop8Test6Param, C output doesn't match SSE2 "
+         "loopfilter output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(Loop8Test6Param, ValueCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = number_of_iterations;
+#if CONFIG_HIGHBITDEPTH
+  const int32_t bd = bit_depth_;
+  DECLARE_ALIGNED(16, uint16_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, ref_s[kNumCoeffs]);
+#else
+  DECLARE_ALIGNED(8, uint8_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(8, uint8_t, ref_s[kNumCoeffs]);
+#endif  // CONFIG_HIGHBITDEPTH
+  int err_count_total = 0;
+  int first_failure = -1;
+
+  // NOTE: The code in av1_loopfilter.c:update_sharpness computes mblim as a
+  // function of sharpness_lvl and the loopfilter lvl as:
+  // block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));
+  // ...
+  // memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
+  //        SIMD_WIDTH);
+  // This means that the largest value for mblim will occur when sharpness_lvl
+  // is equal to 0, and lvl is equal to its greatest value (MAX_LOOP_FILTER).
+  // In this case block_inside_limit will be equal to MAX_LOOP_FILTER and
+  // therefore mblim will be equal to (2 * (lvl + 2) + block_inside_limit) =
+  // 2 * (MAX_LOOP_FILTER + 2) + MAX_LOOP_FILTER = 3 * MAX_LOOP_FILTER + 4
+
+  for (int i = 0; i < count_test_block; ++i) {
+    int err_count = 0;
+    uint8_t tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    blimit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                    tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    limit[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                   tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t,
+                    thresh[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                    tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    int32_t p = kNumCoeffs / 32;
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      s[j] = rnd.Rand16() & mask_;
+      ref_s[j] = s[j];
+    }
+#if CONFIG_HIGHBITDEPTH
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit, limit, thresh, bd);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh, bd));
+#else
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit, limit, thresh);
+    ASM_REGISTER_STATE_CHECK(
+        loopfilter_op_(s + 8 + p * 8, p, blimit, limit, thresh));
+#endif  // CONFIG_HIGHBITDEPTH
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      err_count += ref_s[j] != s[j];
+    }
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Loop8Test6Param, C output doesn't match SSE2 "
+         "loopfilter output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(Loop8Test9Param, OperationCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = number_of_iterations;
+#if CONFIG_HIGHBITDEPTH
+  const int32_t bd = bit_depth_;
+  DECLARE_ALIGNED(16, uint16_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, ref_s[kNumCoeffs]);
+#else
+  DECLARE_ALIGNED(8, uint8_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(8, uint8_t, ref_s[kNumCoeffs]);
+#endif  // CONFIG_HIGHBITDEPTH
+  int err_count_total = 0;
+  int first_failure = -1;
+  for (int i = 0; i < count_test_block; ++i) {
+    int err_count = 0;
+    uint8_t tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    blimit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                     tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    limit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                    tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t,
+                    thresh0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                     tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    blimit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                     tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    limit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                    tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t,
+                    thresh1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                     tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    int32_t p = kNumCoeffs / 32;
+    uint16_t tmp_s[kNumCoeffs];
+    int j = 0;
+    const uint8_t limit = *limit0 < *limit1 ? *limit0 : *limit1;
+    while (j < kNumCoeffs) {
+      uint8_t val = rnd.Rand8();
+      if (val & 0x80) {  // 50% chance to choose a new value.
+        tmp_s[j] = rnd.Rand16();
+        j++;
+      } else {  // 50% chance to repeat previous value in row X times.
+        int k = 0;
+        while (k++ < ((val & 0x1f) + 1) && j < kNumCoeffs) {
+          if (j < 1) {
+            tmp_s[j] = rnd.Rand16();
+          } else if (val & 0x20) {  // Increment by a value within the limit.
+            tmp_s[j] = (tmp_s[j - 1] + (limit - 1));
+          } else {  // Decrement by an value within the limit.
+            tmp_s[j] = (tmp_s[j - 1] - (limit - 1));
+          }
+          j++;
+        }
+      }
+    }
+    for (j = 0; j < kNumCoeffs; j++) {
+      if (i % 2) {
+        s[j] = tmp_s[j] & mask_;
+      } else {
+        s[j] = tmp_s[p * (j % p) + j / p] & mask_;
+      }
+      ref_s[j] = s[j];
+    }
+#if CONFIG_HIGHBITDEPTH
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1,
+                       limit1, thresh1, bd);
+    ASM_REGISTER_STATE_CHECK(loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0,
+                                            thresh0, blimit1, limit1, thresh1,
+                                            bd));
+#else
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1,
+                       limit1, thresh1);
+    ASM_REGISTER_STATE_CHECK(loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0,
+                                            thresh0, blimit1, limit1, thresh1));
+#endif  // CONFIG_HIGHBITDEPTH
+    for (j = 0; j < kNumCoeffs; ++j) {
+      err_count += ref_s[j] != s[j];
+    }
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Loop8Test9Param, C output doesn't match SSE2 "
+         "loopfilter output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(Loop8Test9Param, ValueCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  const int count_test_block = number_of_iterations;
+#if CONFIG_HIGHBITDEPTH
+  DECLARE_ALIGNED(16, uint16_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(16, uint16_t, ref_s[kNumCoeffs]);
+#else
+  DECLARE_ALIGNED(8, uint8_t, s[kNumCoeffs]);
+  DECLARE_ALIGNED(8, uint8_t, ref_s[kNumCoeffs]);
+#endif  // CONFIG_HIGHBITDEPTH
+  int err_count_total = 0;
+  int first_failure = -1;
+  for (int i = 0; i < count_test_block; ++i) {
+    int err_count = 0;
+    uint8_t tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    blimit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                     tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    limit0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                    tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t,
+                    thresh0[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                     tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = static_cast<uint8_t>(rnd(3 * MAX_LOOP_FILTER + 4));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    blimit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                     tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = static_cast<uint8_t>(rnd(MAX_LOOP_FILTER));
+    DECLARE_ALIGNED(16, const uint8_t,
+                    limit1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                    tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    tmp = rnd.Rand8();
+    DECLARE_ALIGNED(16, const uint8_t,
+                    thresh1[16]) = { tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp,
+                                     tmp, tmp, tmp, tmp, tmp, tmp, tmp, tmp };
+    int32_t p = kNumCoeffs / 32;  // TODO(pdlf) can we have non-square here?
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      s[j] = rnd.Rand16() & mask_;
+      ref_s[j] = s[j];
+    }
+#if CONFIG_HIGHBITDEPTH
+    const int32_t bd = bit_depth_;
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1,
+                       limit1, thresh1, bd);
+    ASM_REGISTER_STATE_CHECK(loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0,
+                                            thresh0, blimit1, limit1, thresh1,
+                                            bd));
+#else
+    ref_loopfilter_op_(ref_s + 8 + p * 8, p, blimit0, limit0, thresh0, blimit1,
+                       limit1, thresh1);
+    ASM_REGISTER_STATE_CHECK(loopfilter_op_(s + 8 + p * 8, p, blimit0, limit0,
+                                            thresh0, blimit1, limit1, thresh1));
+#endif  // CONFIG_HIGHBITDEPTH
+    for (int j = 0; j < kNumCoeffs; ++j) {
+      err_count += ref_s[j] != s[j];
+    }
+    if (err_count && !err_count_total) {
+      first_failure = i;
+    }
+    err_count_total += err_count;
+  }
+  EXPECT_EQ(0, err_count_total)
+      << "Error: Loop8Test9Param, C output doesn't match SSE2"
+         "loopfilter output. "
+      << "First failed at test case " << first_failure;
+}
+
+using std::tr1::make_tuple;
+
+#if HAVE_SSE2
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Loop8Test6Param,
+    ::testing::Values(make_tuple(&aom_highbd_lpf_horizontal_4_sse2,
+                                 &aom_highbd_lpf_horizontal_4_c, 8),
+                      make_tuple(&aom_highbd_lpf_vertical_4_sse2,
+                                 &aom_highbd_lpf_vertical_4_c, 8),
+                      make_tuple(&aom_highbd_lpf_horizontal_8_sse2,
+                                 &aom_highbd_lpf_horizontal_8_c, 8),
+                      make_tuple(&aom_highbd_lpf_horizontal_edge_8_sse2,
+                                 &aom_highbd_lpf_horizontal_edge_8_c, 8),
+                      make_tuple(&aom_highbd_lpf_horizontal_edge_16_sse2,
+                                 &aom_highbd_lpf_horizontal_edge_16_c, 8),
+                      make_tuple(&aom_highbd_lpf_vertical_8_sse2,
+                                 &aom_highbd_lpf_vertical_8_c, 8),
+                      make_tuple(&aom_highbd_lpf_vertical_16_sse2,
+                                 &aom_highbd_lpf_vertical_16_c, 8),
+                      make_tuple(&aom_highbd_lpf_horizontal_4_sse2,
+                                 &aom_highbd_lpf_horizontal_4_c, 10),
+                      make_tuple(&aom_highbd_lpf_vertical_4_sse2,
+                                 &aom_highbd_lpf_vertical_4_c, 10),
+                      make_tuple(&aom_highbd_lpf_horizontal_8_sse2,
+                                 &aom_highbd_lpf_horizontal_8_c, 10),
+                      make_tuple(&aom_highbd_lpf_horizontal_edge_8_sse2,
+                                 &aom_highbd_lpf_horizontal_edge_8_c, 10),
+                      make_tuple(&aom_highbd_lpf_horizontal_edge_16_sse2,
+                                 &aom_highbd_lpf_horizontal_edge_16_c, 10),
+                      make_tuple(&aom_highbd_lpf_vertical_8_sse2,
+                                 &aom_highbd_lpf_vertical_8_c, 10),
+                      make_tuple(&aom_highbd_lpf_vertical_16_sse2,
+                                 &aom_highbd_lpf_vertical_16_c, 10),
+                      make_tuple(&aom_highbd_lpf_horizontal_4_sse2,
+                                 &aom_highbd_lpf_horizontal_4_c, 12),
+                      make_tuple(&aom_highbd_lpf_vertical_4_sse2,
+                                 &aom_highbd_lpf_vertical_4_c, 12),
+                      make_tuple(&aom_highbd_lpf_horizontal_8_sse2,
+                                 &aom_highbd_lpf_horizontal_8_c, 12),
+                      make_tuple(&aom_highbd_lpf_horizontal_edge_8_sse2,
+                                 &aom_highbd_lpf_horizontal_edge_8_c, 12),
+                      make_tuple(&aom_highbd_lpf_horizontal_edge_16_sse2,
+                                 &aom_highbd_lpf_horizontal_edge_16_c, 12),
+                      make_tuple(&aom_highbd_lpf_vertical_8_sse2,
+                                 &aom_highbd_lpf_vertical_8_c, 12),
+                      make_tuple(&aom_highbd_lpf_vertical_16_sse2,
+                                 &aom_highbd_lpf_vertical_16_c, 12),
+                      make_tuple(&aom_highbd_lpf_vertical_16_dual_sse2,
+                                 &aom_highbd_lpf_vertical_16_dual_c, 8),
+                      make_tuple(&aom_highbd_lpf_vertical_16_dual_sse2,
+                                 &aom_highbd_lpf_vertical_16_dual_c, 10),
+                      make_tuple(&aom_highbd_lpf_vertical_16_dual_sse2,
+                                 &aom_highbd_lpf_vertical_16_dual_c, 12)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Loop8Test6Param,
+    ::testing::Values(
+        make_tuple(&aom_lpf_horizontal_4_sse2, &aom_lpf_horizontal_4_c, 8),
+        make_tuple(&aom_lpf_horizontal_8_sse2, &aom_lpf_horizontal_8_c, 8),
+        make_tuple(&aom_lpf_horizontal_edge_8_sse2,
+                   &aom_lpf_horizontal_edge_8_c, 8),
+        make_tuple(&aom_lpf_horizontal_edge_16_sse2,
+                   &aom_lpf_horizontal_edge_16_c, 8),
+        make_tuple(&aom_lpf_vertical_4_sse2, &aom_lpf_vertical_4_c, 8),
+        make_tuple(&aom_lpf_vertical_8_sse2, &aom_lpf_vertical_8_c, 8),
+        make_tuple(&aom_lpf_vertical_16_sse2, &aom_lpf_vertical_16_c, 8),
+        make_tuple(&aom_lpf_vertical_16_dual_sse2, &aom_lpf_vertical_16_dual_c,
+                   8)));
+#endif  // CONFIG_HIGHBITDEPTH
+#endif
+
+#if HAVE_AVX2 && (!CONFIG_HIGHBITDEPTH)
+INSTANTIATE_TEST_CASE_P(
+    AVX2, Loop8Test6Param,
+    ::testing::Values(make_tuple(&aom_lpf_horizontal_edge_8_avx2,
+                                 &aom_lpf_horizontal_edge_8_c, 8),
+                      make_tuple(&aom_lpf_horizontal_edge_16_avx2,
+                                 &aom_lpf_horizontal_edge_16_c, 8)));
+#endif
+
+#if HAVE_SSE2
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Loop8Test9Param,
+    ::testing::Values(make_tuple(&aom_highbd_lpf_horizontal_4_dual_sse2,
+                                 &aom_highbd_lpf_horizontal_4_dual_c, 8),
+                      make_tuple(&aom_highbd_lpf_horizontal_8_dual_sse2,
+                                 &aom_highbd_lpf_horizontal_8_dual_c, 8),
+                      make_tuple(&aom_highbd_lpf_vertical_4_dual_sse2,
+                                 &aom_highbd_lpf_vertical_4_dual_c, 8),
+                      make_tuple(&aom_highbd_lpf_vertical_8_dual_sse2,
+                                 &aom_highbd_lpf_vertical_8_dual_c, 8),
+                      make_tuple(&aom_highbd_lpf_horizontal_4_dual_sse2,
+                                 &aom_highbd_lpf_horizontal_4_dual_c, 10),
+                      make_tuple(&aom_highbd_lpf_horizontal_8_dual_sse2,
+                                 &aom_highbd_lpf_horizontal_8_dual_c, 10),
+                      make_tuple(&aom_highbd_lpf_vertical_4_dual_sse2,
+                                 &aom_highbd_lpf_vertical_4_dual_c, 10),
+                      make_tuple(&aom_highbd_lpf_vertical_8_dual_sse2,
+                                 &aom_highbd_lpf_vertical_8_dual_c, 10),
+                      make_tuple(&aom_highbd_lpf_horizontal_4_dual_sse2,
+                                 &aom_highbd_lpf_horizontal_4_dual_c, 12),
+                      make_tuple(&aom_highbd_lpf_horizontal_8_dual_sse2,
+                                 &aom_highbd_lpf_horizontal_8_dual_c, 12),
+                      make_tuple(&aom_highbd_lpf_vertical_4_dual_sse2,
+                                 &aom_highbd_lpf_vertical_4_dual_c, 12),
+                      make_tuple(&aom_highbd_lpf_vertical_8_dual_sse2,
+                                 &aom_highbd_lpf_vertical_8_dual_c, 12)));
+#else
+INSTANTIATE_TEST_CASE_P(
+    SSE2, Loop8Test9Param,
+    ::testing::Values(make_tuple(&aom_lpf_horizontal_4_dual_sse2,
+                                 &aom_lpf_horizontal_4_dual_c, 8),
+                      make_tuple(&aom_lpf_horizontal_8_dual_sse2,
+                                 &aom_lpf_horizontal_8_dual_c, 8),
+                      make_tuple(&aom_lpf_vertical_4_dual_sse2,
+                                 &aom_lpf_vertical_4_dual_c, 8),
+                      make_tuple(&aom_lpf_vertical_8_dual_sse2,
+                                 &aom_lpf_vertical_8_dual_c, 8)));
+#endif  // CONFIG_HIGHBITDEPTH
+#endif
+
+#if HAVE_NEON
+#if CONFIG_HIGHBITDEPTH
+// No neon high bitdepth functions.
+#else
+INSTANTIATE_TEST_CASE_P(
+    NEON, Loop8Test6Param,
+    ::testing::Values(
+#if HAVE_NEON_ASM
+        // Using #if inside the macro is unsupported on MSVS but the tests are
+        // not
+        // currently built for MSVS with ARM and NEON.
+        make_tuple(&aom_lpf_horizontal_edge_8_neon,
+                   &aom_lpf_horizontal_edge_8_c, 8),
+        make_tuple(&aom_lpf_horizontal_edge_16_neon,
+                   &aom_lpf_horizontal_edge_16_c, 8),
+        make_tuple(&aom_lpf_vertical_16_neon, &aom_lpf_vertical_16_c, 8),
+        make_tuple(&aom_lpf_vertical_16_dual_neon, &aom_lpf_vertical_16_dual_c,
+                   8),
+#endif  // HAVE_NEON_ASM
+        make_tuple(&aom_lpf_horizontal_8_neon, &aom_lpf_horizontal_8_c, 8),
+        make_tuple(&aom_lpf_vertical_8_neon, &aom_lpf_vertical_8_c, 8),
+        make_tuple(&aom_lpf_horizontal_4_neon, &aom_lpf_horizontal_4_c, 8),
+        make_tuple(&aom_lpf_vertical_4_neon, &aom_lpf_vertical_4_c, 8)));
+INSTANTIATE_TEST_CASE_P(NEON, Loop8Test9Param,
+                        ::testing::Values(
+#if HAVE_NEON_ASM
+                            make_tuple(&aom_lpf_horizontal_8_dual_neon,
+                                       &aom_lpf_horizontal_8_dual_c, 8),
+                            make_tuple(&aom_lpf_vertical_8_dual_neon,
+                                       &aom_lpf_vertical_8_dual_c, 8),
+#endif  // HAVE_NEON_ASM
+                            make_tuple(&aom_lpf_horizontal_4_dual_neon,
+                                       &aom_lpf_horizontal_4_dual_c, 8),
+                            make_tuple(&aom_lpf_vertical_4_dual_neon,
+                                       &aom_lpf_vertical_4_dual_c, 8)));
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // HAVE_NEON
+
+#if HAVE_DSPR2 && !CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    DSPR2, Loop8Test6Param,
+    ::testing::Values(
+        make_tuple(&aom_lpf_horizontal_4_dspr2, &aom_lpf_horizontal_4_c, 8),
+        make_tuple(&aom_lpf_horizontal_8_dspr2, &aom_lpf_horizontal_8_c, 8),
+        make_tuple(&aom_lpf_horizontal_edge_8, &aom_lpf_horizontal_edge_8, 8),
+        make_tuple(&aom_lpf_horizontal_edge_16, &aom_lpf_horizontal_edge_16, 8),
+        make_tuple(&aom_lpf_vertical_4_dspr2, &aom_lpf_vertical_4_c, 8),
+        make_tuple(&aom_lpf_vertical_8_dspr2, &aom_lpf_vertical_8_c, 8),
+        make_tuple(&aom_lpf_vertical_16_dspr2, &aom_lpf_vertical_16_c, 8),
+        make_tuple(&aom_lpf_vertical_16_dual_dspr2, &aom_lpf_vertical_16_dual_c,
+                   8)));
+
+INSTANTIATE_TEST_CASE_P(
+    DSPR2, Loop8Test9Param,
+    ::testing::Values(make_tuple(&aom_lpf_horizontal_4_dual_dspr2,
+                                 &aom_lpf_horizontal_4_dual_c, 8),
+                      make_tuple(&aom_lpf_horizontal_8_dual_dspr2,
+                                 &aom_lpf_horizontal_8_dual_c, 8),
+                      make_tuple(&aom_lpf_vertical_4_dual_dspr2,
+                                 &aom_lpf_vertical_4_dual_c, 8),
+                      make_tuple(&aom_lpf_vertical_8_dual_dspr2,
+                                 &aom_lpf_vertical_8_dual_c, 8)));
+#endif  // HAVE_DSPR2 && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_MSA && (!CONFIG_HIGHBITDEPTH)
+INSTANTIATE_TEST_CASE_P(
+    MSA, Loop8Test6Param,
+    ::testing::Values(
+        make_tuple(&aom_lpf_horizontal_4_msa, &aom_lpf_horizontal_4_c, 8),
+        make_tuple(&aom_lpf_horizontal_8_msa, &aom_lpf_horizontal_8_c, 8),
+        make_tuple(&aom_lpf_horizontal_edge_8_msa, &aom_lpf_horizontal_edge_8_c,
+                   8),
+        make_tuple(&aom_lpf_horizontal_edge_16_msa,
+                   &aom_lpf_horizontal_edge_16_c, 8),
+        make_tuple(&aom_lpf_vertical_4_msa, &aom_lpf_vertical_4_c, 8),
+        make_tuple(&aom_lpf_vertical_8_msa, &aom_lpf_vertical_8_c, 8),
+        make_tuple(&aom_lpf_vertical_16_msa, &aom_lpf_vertical_16_c, 8)));
+
+INSTANTIATE_TEST_CASE_P(
+    MSA, Loop8Test9Param,
+    ::testing::Values(make_tuple(&aom_lpf_horizontal_4_dual_msa,
+                                 &aom_lpf_horizontal_4_dual_c, 8),
+                      make_tuple(&aom_lpf_horizontal_8_dual_msa,
+                                 &aom_lpf_horizontal_8_dual_c, 8),
+                      make_tuple(&aom_lpf_vertical_4_dual_msa,
+                                 &aom_lpf_vertical_4_dual_c, 8),
+                      make_tuple(&aom_lpf_vertical_8_dual_msa,
+                                 &aom_lpf_vertical_8_dual_c, 8)));
+#endif  // HAVE_MSA && (!CONFIG_HIGHBITDEPTH)
+
+}  // namespace
diff --git a/third_party/aom/test/masked_sad_test.cc b/third_party/aom/test/masked_sad_test.cc
new file mode 100644
index 0000000000..53f85eef7d
--- /dev/null
+++ b/third_party/aom/test/masked_sad_test.cc
@@ -0,0 +1,206 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+const int number_of_iterations = 500;
+
+typedef unsigned int (*MaskedSADFunc)(const uint8_t *a, int a_stride,
+                                      const uint8_t *b, int b_stride,
+                                      const uint8_t *m, int m_stride);
+typedef std::tr1::tuple<MaskedSADFunc, MaskedSADFunc> MaskedSADParam;
+
+class MaskedSADTest : public ::testing::TestWithParam<MaskedSADParam> {
+ public:
+  virtual ~MaskedSADTest() {}
+  virtual void SetUp() {
+    maskedSAD_op_ = GET_PARAM(0);
+    ref_maskedSAD_op_ = GET_PARAM(1);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  MaskedSADFunc maskedSAD_op_;
+  MaskedSADFunc ref_maskedSAD_op_;
+};
+
+TEST_P(MaskedSADTest, OperationCheck) {
+  unsigned int ref_ret, ret;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint8_t, src_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint8_t, ref_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint8_t, msk_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  int err_count = 0;
+  int first_failure = -1;
+  int src_stride = MAX_SB_SIZE;
+  int ref_stride = MAX_SB_SIZE;
+  int msk_stride = MAX_SB_SIZE;
+  for (int i = 0; i < number_of_iterations; ++i) {
+    for (int j = 0; j < MAX_SB_SIZE * MAX_SB_SIZE; j++) {
+      src_ptr[j] = rnd.Rand8();
+      ref_ptr[j] = rnd.Rand8();
+      msk_ptr[j] = ((rnd.Rand8() & 0x7f) > 64) ? rnd.Rand8() & 0x3f : 64;
+      assert(msk_ptr[j] <= 64);
+    }
+
+    ref_ret = ref_maskedSAD_op_(src_ptr, src_stride, ref_ptr, ref_stride,
+                                msk_ptr, msk_stride);
+    ASM_REGISTER_STATE_CHECK(ret = maskedSAD_op_(src_ptr, src_stride, ref_ptr,
+                                                 ref_stride, msk_ptr,
+                                                 msk_stride));
+    if (ret != ref_ret) {
+      err_count++;
+      if (first_failure == -1) first_failure = i;
+    }
+  }
+  EXPECT_EQ(0, err_count)
+      << "Error: Masked SAD Test, C output doesn't match SSSE3 output. "
+      << "First failed at test case " << first_failure;
+}
+
+#if CONFIG_HIGHBITDEPTH
+typedef unsigned int (*HighbdMaskedSADFunc)(const uint8_t *a, int a_stride,
+                                            const uint8_t *b, int b_stride,
+                                            const uint8_t *m, int m_stride);
+typedef std::tr1::tuple<HighbdMaskedSADFunc, HighbdMaskedSADFunc>
+    HighbdMaskedSADParam;
+
+class HighbdMaskedSADTest
+    : public ::testing::TestWithParam<HighbdMaskedSADParam> {
+ public:
+  virtual ~HighbdMaskedSADTest() {}
+  virtual void SetUp() {
+    maskedSAD_op_ = GET_PARAM(0);
+    ref_maskedSAD_op_ = GET_PARAM(1);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  HighbdMaskedSADFunc maskedSAD_op_;
+  HighbdMaskedSADFunc ref_maskedSAD_op_;
+};
+
+TEST_P(HighbdMaskedSADTest, OperationCheck) {
+  unsigned int ref_ret, ret;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint16_t, src_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint16_t, ref_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint8_t, msk_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  uint8_t *src8_ptr = CONVERT_TO_BYTEPTR(src_ptr);
+  uint8_t *ref8_ptr = CONVERT_TO_BYTEPTR(ref_ptr);
+  int err_count = 0;
+  int first_failure = -1;
+  int src_stride = MAX_SB_SIZE;
+  int ref_stride = MAX_SB_SIZE;
+  int msk_stride = MAX_SB_SIZE;
+  for (int i = 0; i < number_of_iterations; ++i) {
+    for (int j = 0; j < MAX_SB_SIZE * MAX_SB_SIZE; j++) {
+      src_ptr[j] = rnd.Rand16() & 0xfff;
+      ref_ptr[j] = rnd.Rand16() & 0xfff;
+      msk_ptr[j] = ((rnd.Rand8() & 0x7f) > 64) ? rnd.Rand8() & 0x3f : 64;
+    }
+
+    ref_ret = ref_maskedSAD_op_(src8_ptr, src_stride, ref8_ptr, ref_stride,
+                                msk_ptr, msk_stride);
+    ASM_REGISTER_STATE_CHECK(ret = maskedSAD_op_(src8_ptr, src_stride, ref8_ptr,
+                                                 ref_stride, msk_ptr,
+                                                 msk_stride));
+    if (ret != ref_ret) {
+      err_count++;
+      if (first_failure == -1) first_failure = i;
+    }
+  }
+  EXPECT_EQ(0, err_count)
+      << "Error: High BD Masked SAD Test, C output doesn't match SSSE3 output. "
+      << "First failed at test case " << first_failure;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+using std::tr1::make_tuple;
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3_C_COMPARE, MaskedSADTest,
+    ::testing::Values(
+#if CONFIG_EXT_PARTITION
+        make_tuple(&aom_masked_sad128x128_ssse3, &aom_masked_sad128x128_c),
+        make_tuple(&aom_masked_sad128x64_ssse3, &aom_masked_sad128x64_c),
+        make_tuple(&aom_masked_sad64x128_ssse3, &aom_masked_sad64x128_c),
+#endif  // CONFIG_EXT_PARTITION
+        make_tuple(&aom_masked_sad64x64_ssse3, &aom_masked_sad64x64_c),
+        make_tuple(&aom_masked_sad64x32_ssse3, &aom_masked_sad64x32_c),
+        make_tuple(&aom_masked_sad32x64_ssse3, &aom_masked_sad32x64_c),
+        make_tuple(&aom_masked_sad32x32_ssse3, &aom_masked_sad32x32_c),
+        make_tuple(&aom_masked_sad32x16_ssse3, &aom_masked_sad32x16_c),
+        make_tuple(&aom_masked_sad16x32_ssse3, &aom_masked_sad16x32_c),
+        make_tuple(&aom_masked_sad16x16_ssse3, &aom_masked_sad16x16_c),
+        make_tuple(&aom_masked_sad16x8_ssse3, &aom_masked_sad16x8_c),
+        make_tuple(&aom_masked_sad8x16_ssse3, &aom_masked_sad8x16_c),
+        make_tuple(&aom_masked_sad8x8_ssse3, &aom_masked_sad8x8_c),
+        make_tuple(&aom_masked_sad8x4_ssse3, &aom_masked_sad8x4_c),
+        make_tuple(&aom_masked_sad4x8_ssse3, &aom_masked_sad4x8_c),
+        make_tuple(&aom_masked_sad4x4_ssse3, &aom_masked_sad4x4_c)));
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(SSSE3_C_COMPARE, HighbdMaskedSADTest,
+                        ::testing::Values(
+#if CONFIG_EXT_PARTITION
+                            make_tuple(&aom_highbd_masked_sad128x128_ssse3,
+                                       &aom_highbd_masked_sad128x128_c),
+                            make_tuple(&aom_highbd_masked_sad128x64_ssse3,
+                                       &aom_highbd_masked_sad128x64_c),
+                            make_tuple(&aom_highbd_masked_sad64x128_ssse3,
+                                       &aom_highbd_masked_sad64x128_c),
+#endif  // CONFIG_EXT_PARTITION
+                            make_tuple(&aom_highbd_masked_sad64x64_ssse3,
+                                       &aom_highbd_masked_sad64x64_c),
+                            make_tuple(&aom_highbd_masked_sad64x32_ssse3,
+                                       &aom_highbd_masked_sad64x32_c),
+                            make_tuple(&aom_highbd_masked_sad32x64_ssse3,
+                                       &aom_highbd_masked_sad32x64_c),
+                            make_tuple(&aom_highbd_masked_sad32x32_ssse3,
+                                       &aom_highbd_masked_sad32x32_c),
+                            make_tuple(&aom_highbd_masked_sad32x16_ssse3,
+                                       &aom_highbd_masked_sad32x16_c),
+                            make_tuple(&aom_highbd_masked_sad16x32_ssse3,
+                                       &aom_highbd_masked_sad16x32_c),
+                            make_tuple(&aom_highbd_masked_sad16x16_ssse3,
+                                       &aom_highbd_masked_sad16x16_c),
+                            make_tuple(&aom_highbd_masked_sad16x8_ssse3,
+                                       &aom_highbd_masked_sad16x8_c),
+                            make_tuple(&aom_highbd_masked_sad8x16_ssse3,
+                                       &aom_highbd_masked_sad8x16_c),
+                            make_tuple(&aom_highbd_masked_sad8x8_ssse3,
+                                       &aom_highbd_masked_sad8x8_c),
+                            make_tuple(&aom_highbd_masked_sad8x4_ssse3,
+                                       &aom_highbd_masked_sad8x4_c),
+                            make_tuple(&aom_highbd_masked_sad4x8_ssse3,
+                                       &aom_highbd_masked_sad4x8_c),
+                            make_tuple(&aom_highbd_masked_sad4x4_ssse3,
+                                       &aom_highbd_masked_sad4x4_c)));
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // HAVE_SSSE3
+}  // namespace
diff --git a/third_party/aom/test/masked_variance_test.cc b/third_party/aom/test/masked_variance_test.cc
new file mode 100644
index 0000000000..65e852aeab
--- /dev/null
+++ b/third_party/aom/test/masked_variance_test.cc
@@ -0,0 +1,790 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_codec.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_mem/aom_mem.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+const int number_of_iterations = 500;
+
+typedef unsigned int (*MaskedVarianceFunc)(const uint8_t *a, int a_stride,
+                                           const uint8_t *b, int b_stride,
+                                           const uint8_t *m, int m_stride,
+                                           unsigned int *sse);
+
+typedef std::tr1::tuple<MaskedVarianceFunc, MaskedVarianceFunc>
+    MaskedVarianceParam;
+
+class MaskedVarianceTest
+    : public ::testing::TestWithParam<MaskedVarianceParam> {
+ public:
+  virtual ~MaskedVarianceTest() {}
+  virtual void SetUp() {
+    opt_func_ = GET_PARAM(0);
+    ref_func_ = GET_PARAM(1);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  MaskedVarianceFunc opt_func_;
+  MaskedVarianceFunc ref_func_;
+};
+
+TEST_P(MaskedVarianceTest, OperationCheck) {
+  unsigned int ref_ret, opt_ret;
+  unsigned int ref_sse, opt_sse;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint8_t, src_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint8_t, ref_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint8_t, msk_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  int err_count = 0;
+  int first_failure = -1;
+  int src_stride = MAX_SB_SIZE;
+  int ref_stride = MAX_SB_SIZE;
+  int msk_stride = MAX_SB_SIZE;
+
+  for (int i = 0; i < number_of_iterations; ++i) {
+    for (int j = 0; j < MAX_SB_SIZE * MAX_SB_SIZE; j++) {
+      src_ptr[j] = rnd.Rand8();
+      ref_ptr[j] = rnd.Rand8();
+      msk_ptr[j] = rnd(65);
+    }
+
+    ref_ret = ref_func_(src_ptr, src_stride, ref_ptr, ref_stride, msk_ptr,
+                        msk_stride, &ref_sse);
+    ASM_REGISTER_STATE_CHECK(opt_ret = opt_func_(src_ptr, src_stride, ref_ptr,
+                                                 ref_stride, msk_ptr,
+                                                 msk_stride, &opt_sse));
+
+    if (opt_ret != ref_ret || opt_sse != ref_sse) {
+      err_count++;
+      if (first_failure == -1) first_failure = i;
+    }
+  }
+
+  EXPECT_EQ(0, err_count) << "Error: Masked Variance Test OperationCheck,"
+                          << "C output doesn't match SSSE3 output. "
+                          << "First failed at test case " << first_failure;
+}
+
+TEST_P(MaskedVarianceTest, ExtremeValues) {
+  unsigned int ref_ret, opt_ret;
+  unsigned int ref_sse, opt_sse;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint8_t, src_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint8_t, ref_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint8_t, msk_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  int err_count = 0;
+  int first_failure = -1;
+  int src_stride = MAX_SB_SIZE;
+  int ref_stride = MAX_SB_SIZE;
+  int msk_stride = MAX_SB_SIZE;
+
+  for (int i = 0; i < 8; ++i) {
+    memset(src_ptr, (i & 0x1) ? 255 : 0, MAX_SB_SIZE * MAX_SB_SIZE);
+    memset(ref_ptr, (i & 0x2) ? 255 : 0, MAX_SB_SIZE * MAX_SB_SIZE);
+    memset(msk_ptr, (i & 0x4) ? 64 : 0, MAX_SB_SIZE * MAX_SB_SIZE);
+
+    ref_ret = ref_func_(src_ptr, src_stride, ref_ptr, ref_stride, msk_ptr,
+                        msk_stride, &ref_sse);
+    ASM_REGISTER_STATE_CHECK(opt_ret = opt_func_(src_ptr, src_stride, ref_ptr,
+                                                 ref_stride, msk_ptr,
+                                                 msk_stride, &opt_sse));
+
+    if (opt_ret != ref_ret || opt_sse != ref_sse) {
+      err_count++;
+      if (first_failure == -1) first_failure = i;
+    }
+  }
+
+  EXPECT_EQ(0, err_count) << "Error: Masked Variance Test ExtremeValues,"
+                          << "C output doesn't match SSSE3 output. "
+                          << "First failed at test case " << first_failure;
+}
+
+typedef unsigned int (*MaskedSubPixelVarianceFunc)(
+    const uint8_t *a, int a_stride, int xoffset, int yoffset, const uint8_t *b,
+    int b_stride, const uint8_t *m, int m_stride, unsigned int *sse);
+
+typedef std::tr1::tuple<MaskedSubPixelVarianceFunc, MaskedSubPixelVarianceFunc>
+    MaskedSubPixelVarianceParam;
+
+class MaskedSubPixelVarianceTest
+    : public ::testing::TestWithParam<MaskedSubPixelVarianceParam> {
+ public:
+  virtual ~MaskedSubPixelVarianceTest() {}
+  virtual void SetUp() {
+    opt_func_ = GET_PARAM(0);
+    ref_func_ = GET_PARAM(1);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  MaskedSubPixelVarianceFunc opt_func_;
+  MaskedSubPixelVarianceFunc ref_func_;
+};
+
+TEST_P(MaskedSubPixelVarianceTest, OperationCheck) {
+  unsigned int ref_ret, opt_ret;
+  unsigned int ref_sse, opt_sse;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint8_t, src_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  DECLARE_ALIGNED(16, uint8_t, ref_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  DECLARE_ALIGNED(16, uint8_t, msk_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  int err_count = 0;
+  int first_failure = -1;
+  int src_stride = (MAX_SB_SIZE + 1);
+  int ref_stride = (MAX_SB_SIZE + 1);
+  int msk_stride = (MAX_SB_SIZE + 1);
+  int xoffset;
+  int yoffset;
+
+  for (int i = 0; i < number_of_iterations; ++i) {
+    int xoffsets[] = { 0, 4, rnd(BIL_SUBPEL_SHIFTS) };
+    int yoffsets[] = { 0, 4, rnd(BIL_SUBPEL_SHIFTS) };
+    for (int j = 0; j < (MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1); j++) {
+      src_ptr[j] = rnd.Rand8();
+      ref_ptr[j] = rnd.Rand8();
+      msk_ptr[j] = rnd(65);
+    }
+    for (int k = 0; k < 3; k++) {
+      xoffset = xoffsets[k];
+      for (int l = 0; l < 3; l++) {
+        xoffset = xoffsets[k];
+        yoffset = yoffsets[l];
+
+        ref_ret = ref_func_(src_ptr, src_stride, xoffset, yoffset, ref_ptr,
+                            ref_stride, msk_ptr, msk_stride, &ref_sse);
+        ASM_REGISTER_STATE_CHECK(
+            opt_ret = opt_func_(src_ptr, src_stride, xoffset, yoffset, ref_ptr,
+                                ref_stride, msk_ptr, msk_stride, &opt_sse));
+
+        if (opt_ret != ref_ret || opt_sse != ref_sse) {
+          err_count++;
+          if (first_failure == -1) first_failure = i;
+        }
+      }
+    }
+  }
+
+  EXPECT_EQ(0, err_count)
+      << "Error: Masked Sub Pixel Variance Test OperationCheck,"
+      << "C output doesn't match SSSE3 output. "
+      << "First failed at test case " << first_failure;
+}
+
+TEST_P(MaskedSubPixelVarianceTest, ExtremeValues) {
+  unsigned int ref_ret, opt_ret;
+  unsigned int ref_sse, opt_sse;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint8_t, src_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  DECLARE_ALIGNED(16, uint8_t, ref_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  DECLARE_ALIGNED(16, uint8_t, msk_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  int first_failure_x = -1;
+  int first_failure_y = -1;
+  int err_count = 0;
+  int first_failure = -1;
+  int src_stride = (MAX_SB_SIZE + 1);
+  int ref_stride = (MAX_SB_SIZE + 1);
+  int msk_stride = (MAX_SB_SIZE + 1);
+
+  for (int xoffset = 0; xoffset < BIL_SUBPEL_SHIFTS; xoffset++) {
+    for (int yoffset = 0; yoffset < BIL_SUBPEL_SHIFTS; yoffset++) {
+      for (int i = 0; i < 8; ++i) {
+        memset(src_ptr, (i & 0x1) ? 255 : 0,
+               (MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1));
+        memset(ref_ptr, (i & 0x2) ? 255 : 0,
+               (MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1));
+        memset(msk_ptr, (i & 0x4) ? 64 : 0,
+               (MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1));
+
+        ref_ret = ref_func_(src_ptr, src_stride, xoffset, yoffset, ref_ptr,
+                            ref_stride, msk_ptr, msk_stride, &ref_sse);
+        ASM_REGISTER_STATE_CHECK(
+            opt_ret = opt_func_(src_ptr, src_stride, xoffset, yoffset, ref_ptr,
+                                ref_stride, msk_ptr, msk_stride, &opt_sse));
+
+        if (opt_ret != ref_ret || opt_sse != ref_sse) {
+          err_count++;
+          if (first_failure == -1) {
+            first_failure = i;
+            first_failure_x = xoffset;
+            first_failure_y = yoffset;
+          }
+        }
+      }
+    }
+  }
+
+  EXPECT_EQ(0, err_count) << "Error: Masked Variance Test ExtremeValues,"
+                          << "C output doesn't match SSSE3 output. "
+                          << "First failed at test case " << first_failure
+                          << " x_offset = " << first_failure_x
+                          << " y_offset = " << first_failure_y;
+}
+
+#if CONFIG_HIGHBITDEPTH
+typedef std::tr1::tuple<MaskedVarianceFunc, MaskedVarianceFunc, aom_bit_depth_t>
+    HighbdMaskedVarianceParam;
+
+class HighbdMaskedVarianceTest
+    : public ::testing::TestWithParam<HighbdMaskedVarianceParam> {
+ public:
+  virtual ~HighbdMaskedVarianceTest() {}
+  virtual void SetUp() {
+    opt_func_ = GET_PARAM(0);
+    ref_func_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  MaskedVarianceFunc opt_func_;
+  MaskedVarianceFunc ref_func_;
+  aom_bit_depth_t bit_depth_;
+};
+
+TEST_P(HighbdMaskedVarianceTest, OperationCheck) {
+  unsigned int ref_ret, opt_ret;
+  unsigned int ref_sse, opt_sse;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint16_t, src_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint16_t, ref_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint8_t, msk_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  uint8_t *src8_ptr = CONVERT_TO_BYTEPTR(src_ptr);
+  uint8_t *ref8_ptr = CONVERT_TO_BYTEPTR(ref_ptr);
+  int err_count = 0;
+  int first_failure = -1;
+  int src_stride = MAX_SB_SIZE;
+  int ref_stride = MAX_SB_SIZE;
+  int msk_stride = MAX_SB_SIZE;
+
+  for (int i = 0; i < number_of_iterations; ++i) {
+    for (int j = 0; j < MAX_SB_SIZE * MAX_SB_SIZE; j++) {
+      src_ptr[j] = rnd.Rand16() & ((1 << bit_depth_) - 1);
+      ref_ptr[j] = rnd.Rand16() & ((1 << bit_depth_) - 1);
+      msk_ptr[j] = rnd(65);
+    }
+
+    ref_ret = ref_func_(src8_ptr, src_stride, ref8_ptr, ref_stride, msk_ptr,
+                        msk_stride, &ref_sse);
+    ASM_REGISTER_STATE_CHECK(opt_ret = opt_func_(src8_ptr, src_stride, ref8_ptr,
+                                                 ref_stride, msk_ptr,
+                                                 msk_stride, &opt_sse));
+
+    if (opt_ret != ref_ret || opt_sse != ref_sse) {
+      err_count++;
+      if (first_failure == -1) first_failure = i;
+    }
+  }
+
+  EXPECT_EQ(0, err_count) << "Error: Masked Variance Test OperationCheck,"
+                          << "C output doesn't match SSSE3 output. "
+                          << "First failed at test case " << first_failure;
+}
+
+TEST_P(HighbdMaskedVarianceTest, ExtremeValues) {
+  unsigned int ref_ret, opt_ret;
+  unsigned int ref_sse, opt_sse;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint16_t, src_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint16_t, ref_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  DECLARE_ALIGNED(16, uint8_t, msk_ptr[MAX_SB_SIZE * MAX_SB_SIZE]);
+  uint8_t *src8_ptr = CONVERT_TO_BYTEPTR(src_ptr);
+  uint8_t *ref8_ptr = CONVERT_TO_BYTEPTR(ref_ptr);
+  int err_count = 0;
+  int first_failure = -1;
+  int src_stride = MAX_SB_SIZE;
+  int ref_stride = MAX_SB_SIZE;
+  int msk_stride = MAX_SB_SIZE;
+
+  for (int i = 0; i < 8; ++i) {
+    aom_memset16(src_ptr, (i & 0x1) ? ((1 << bit_depth_) - 1) : 0,
+                 MAX_SB_SIZE * MAX_SB_SIZE);
+    aom_memset16(ref_ptr, (i & 0x2) ? ((1 << bit_depth_) - 1) : 0,
+                 MAX_SB_SIZE * MAX_SB_SIZE);
+    memset(msk_ptr, (i & 0x4) ? 64 : 0, MAX_SB_SIZE * MAX_SB_SIZE);
+
+    ref_ret = ref_func_(src8_ptr, src_stride, ref8_ptr, ref_stride, msk_ptr,
+                        msk_stride, &ref_sse);
+    ASM_REGISTER_STATE_CHECK(opt_ret = opt_func_(src8_ptr, src_stride, ref8_ptr,
+                                                 ref_stride, msk_ptr,
+                                                 msk_stride, &opt_sse));
+
+    if (opt_ret != ref_ret || opt_sse != ref_sse) {
+      err_count++;
+      if (first_failure == -1) first_failure = i;
+    }
+  }
+
+  EXPECT_EQ(0, err_count) << "Error: Masked Variance Test ExtremeValues,"
+                          << "C output doesn't match SSSE3 output. "
+                          << "First failed at test case " << first_failure;
+}
+
+typedef std::tr1::tuple<MaskedSubPixelVarianceFunc, MaskedSubPixelVarianceFunc,
+                        aom_bit_depth_t>
+    HighbdMaskedSubPixelVarianceParam;
+
+class HighbdMaskedSubPixelVarianceTest
+    : public ::testing::TestWithParam<HighbdMaskedSubPixelVarianceParam> {
+ public:
+  virtual ~HighbdMaskedSubPixelVarianceTest() {}
+  virtual void SetUp() {
+    opt_func_ = GET_PARAM(0);
+    ref_func_ = GET_PARAM(1);
+    bit_depth_ = GET_PARAM(2);
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  MaskedSubPixelVarianceFunc opt_func_;
+  MaskedSubPixelVarianceFunc ref_func_;
+  aom_bit_depth_t bit_depth_;
+};
+
+TEST_P(HighbdMaskedSubPixelVarianceTest, OperationCheck) {
+  unsigned int ref_ret, opt_ret;
+  unsigned int ref_sse, opt_sse;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint16_t, src_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  DECLARE_ALIGNED(16, uint16_t, ref_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  DECLARE_ALIGNED(16, uint8_t, msk_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  uint8_t *src8_ptr = CONVERT_TO_BYTEPTR(src_ptr);
+  uint8_t *ref8_ptr = CONVERT_TO_BYTEPTR(ref_ptr);
+  int err_count = 0;
+  int first_failure = -1;
+  int first_failure_x = -1;
+  int first_failure_y = -1;
+  int src_stride = (MAX_SB_SIZE + 1);
+  int ref_stride = (MAX_SB_SIZE + 1);
+  int msk_stride = (MAX_SB_SIZE + 1);
+  int xoffset, yoffset;
+
+  for (int i = 0; i < number_of_iterations; ++i) {
+    for (xoffset = 0; xoffset < BIL_SUBPEL_SHIFTS; xoffset++) {
+      for (yoffset = 0; yoffset < BIL_SUBPEL_SHIFTS; yoffset++) {
+        for (int j = 0; j < (MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1); j++) {
+          src_ptr[j] = rnd.Rand16() & ((1 << bit_depth_) - 1);
+          ref_ptr[j] = rnd.Rand16() & ((1 << bit_depth_) - 1);
+          msk_ptr[j] = rnd(65);
+        }
+
+        ref_ret = ref_func_(src8_ptr, src_stride, xoffset, yoffset, ref8_ptr,
+                            ref_stride, msk_ptr, msk_stride, &ref_sse);
+        ASM_REGISTER_STATE_CHECK(opt_ret =
+                                     opt_func_(src8_ptr, src_stride, xoffset,
+                                               yoffset, ref8_ptr, ref_stride,
+                                               msk_ptr, msk_stride, &opt_sse));
+
+        if (opt_ret != ref_ret || opt_sse != ref_sse) {
+          err_count++;
+          if (first_failure == -1) {
+            first_failure = i;
+            first_failure_x = xoffset;
+            first_failure_y = yoffset;
+          }
+        }
+      }
+    }
+  }
+
+  EXPECT_EQ(0, err_count)
+      << "Error: Masked Sub Pixel Variance Test OperationCheck,"
+      << "C output doesn't match SSSE3 output. "
+      << "First failed at test case " << first_failure
+      << " x_offset = " << first_failure_x << " y_offset = " << first_failure_y;
+}
+
+TEST_P(HighbdMaskedSubPixelVarianceTest, ExtremeValues) {
+  unsigned int ref_ret, opt_ret;
+  unsigned int ref_sse, opt_sse;
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, uint16_t, src_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  DECLARE_ALIGNED(16, uint16_t, ref_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  DECLARE_ALIGNED(16, uint8_t, msk_ptr[(MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1)]);
+  uint8_t *src8_ptr = CONVERT_TO_BYTEPTR(src_ptr);
+  uint8_t *ref8_ptr = CONVERT_TO_BYTEPTR(ref_ptr);
+  int first_failure_x = -1;
+  int first_failure_y = -1;
+  int err_count = 0;
+  int first_failure = -1;
+  int src_stride = (MAX_SB_SIZE + 1);
+  int ref_stride = (MAX_SB_SIZE + 1);
+  int msk_stride = (MAX_SB_SIZE + 1);
+
+  for (int xoffset = 0; xoffset < BIL_SUBPEL_SHIFTS; xoffset++) {
+    for (int yoffset = 0; yoffset < BIL_SUBPEL_SHIFTS; yoffset++) {
+      for (int i = 0; i < 8; ++i) {
+        aom_memset16(src_ptr, (i & 0x1) ? ((1 << bit_depth_) - 1) : 0,
+                     (MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1));
+        aom_memset16(ref_ptr, (i & 0x2) ? ((1 << bit_depth_) - 1) : 0,
+                     (MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1));
+        memset(msk_ptr, (i & 0x4) ? 64 : 0,
+               (MAX_SB_SIZE + 1) * (MAX_SB_SIZE + 1));
+
+        ref_ret = ref_func_(src8_ptr, src_stride, xoffset, yoffset, ref8_ptr,
+                            ref_stride, msk_ptr, msk_stride, &ref_sse);
+        ASM_REGISTER_STATE_CHECK(opt_ret =
+                                     opt_func_(src8_ptr, src_stride, xoffset,
+                                               yoffset, ref8_ptr, ref_stride,
+                                               msk_ptr, msk_stride, &opt_sse));
+
+        if (opt_ret != ref_ret || opt_sse != ref_sse) {
+          err_count++;
+          if (first_failure == -1) {
+            first_failure = i;
+            first_failure_x = xoffset;
+            first_failure_y = yoffset;
+          }
+        }
+      }
+    }
+  }
+
+  EXPECT_EQ(0, err_count) << "Error: Masked Variance Test ExtremeValues,"
+                          << "C output doesn't match SSSE3 output. "
+                          << "First failed at test case " << first_failure
+                          << " x_offset = " << first_failure_x
+                          << " y_offset = " << first_failure_y;
+}
+#endif  // CONFIG_HIGHBITDEPTH
+
+using std::tr1::make_tuple;
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3_C_COMPARE, MaskedVarianceTest,
+    ::testing::Values(
+#if CONFIG_EXT_PARTITION
+        make_tuple(&aom_masked_variance128x128_ssse3,
+                   &aom_masked_variance128x128_c),
+        make_tuple(&aom_masked_variance128x64_ssse3,
+                   &aom_masked_variance128x64_c),
+        make_tuple(&aom_masked_variance64x128_ssse3,
+                   &aom_masked_variance64x128_c),
+#endif  // CONFIG_EXT_PARTITION
+        make_tuple(&aom_masked_variance64x64_ssse3,
+                   &aom_masked_variance64x64_c),
+        make_tuple(&aom_masked_variance64x32_ssse3,
+                   &aom_masked_variance64x32_c),
+        make_tuple(&aom_masked_variance32x64_ssse3,
+                   &aom_masked_variance32x64_c),
+        make_tuple(&aom_masked_variance32x32_ssse3,
+                   &aom_masked_variance32x32_c),
+        make_tuple(&aom_masked_variance32x16_ssse3,
+                   &aom_masked_variance32x16_c),
+        make_tuple(&aom_masked_variance16x32_ssse3,
+                   &aom_masked_variance16x32_c),
+        make_tuple(&aom_masked_variance16x16_ssse3,
+                   &aom_masked_variance16x16_c),
+        make_tuple(&aom_masked_variance16x8_ssse3, &aom_masked_variance16x8_c),
+        make_tuple(&aom_masked_variance8x16_ssse3, &aom_masked_variance8x16_c),
+        make_tuple(&aom_masked_variance8x8_ssse3, &aom_masked_variance8x8_c),
+        make_tuple(&aom_masked_variance8x4_ssse3, &aom_masked_variance8x4_c),
+        make_tuple(&aom_masked_variance4x8_ssse3, &aom_masked_variance4x8_c),
+        make_tuple(&aom_masked_variance4x4_ssse3, &aom_masked_variance4x4_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSSE3_C_COMPARE, MaskedSubPixelVarianceTest,
+    ::testing::Values(
+#if CONFIG_EXT_PARTITION
+        make_tuple(&aom_masked_sub_pixel_variance128x128_ssse3,
+                   &aom_masked_sub_pixel_variance128x128_c),
+        make_tuple(&aom_masked_sub_pixel_variance128x64_ssse3,
+                   &aom_masked_sub_pixel_variance128x64_c),
+        make_tuple(&aom_masked_sub_pixel_variance64x128_ssse3,
+                   &aom_masked_sub_pixel_variance64x128_c),
+#endif  // CONFIG_EXT_PARTITION
+        make_tuple(&aom_masked_sub_pixel_variance64x64_ssse3,
+                   &aom_masked_sub_pixel_variance64x64_c),
+        make_tuple(&aom_masked_sub_pixel_variance64x32_ssse3,
+                   &aom_masked_sub_pixel_variance64x32_c),
+        make_tuple(&aom_masked_sub_pixel_variance32x64_ssse3,
+                   &aom_masked_sub_pixel_variance32x64_c),
+        make_tuple(&aom_masked_sub_pixel_variance32x32_ssse3,
+                   &aom_masked_sub_pixel_variance32x32_c),
+        make_tuple(&aom_masked_sub_pixel_variance32x16_ssse3,
+                   &aom_masked_sub_pixel_variance32x16_c),
+        make_tuple(&aom_masked_sub_pixel_variance16x32_ssse3,
+                   &aom_masked_sub_pixel_variance16x32_c),
+        make_tuple(&aom_masked_sub_pixel_variance16x16_ssse3,
+                   &aom_masked_sub_pixel_variance16x16_c),
+        make_tuple(&aom_masked_sub_pixel_variance16x8_ssse3,
+                   &aom_masked_sub_pixel_variance16x8_c),
+        make_tuple(&aom_masked_sub_pixel_variance8x16_ssse3,
+                   &aom_masked_sub_pixel_variance8x16_c),
+        make_tuple(&aom_masked_sub_pixel_variance8x8_ssse3,
+                   &aom_masked_sub_pixel_variance8x8_c),
+        make_tuple(&aom_masked_sub_pixel_variance8x4_ssse3,
+                   &aom_masked_sub_pixel_variance8x4_c),
+        make_tuple(&aom_masked_sub_pixel_variance4x8_ssse3,
+                   &aom_masked_sub_pixel_variance4x8_c),
+        make_tuple(&aom_masked_sub_pixel_variance4x4_ssse3,
+                   &aom_masked_sub_pixel_variance4x4_c)));
+
+#if CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSSE3_C_COMPARE, HighbdMaskedVarianceTest,
+    ::testing::Values(
+#if CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_masked_variance128x128_ssse3,
+                   &aom_highbd_masked_variance128x128_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance128x64_ssse3,
+                   &aom_highbd_masked_variance128x64_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance64x128_ssse3,
+                   &aom_highbd_masked_variance64x128_c, AOM_BITS_8),
+#endif  // CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_masked_variance64x64_ssse3,
+                   &aom_highbd_masked_variance64x64_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance64x32_ssse3,
+                   &aom_highbd_masked_variance64x32_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance32x64_ssse3,
+                   &aom_highbd_masked_variance32x64_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance32x32_ssse3,
+                   &aom_highbd_masked_variance32x32_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance32x16_ssse3,
+                   &aom_highbd_masked_variance32x16_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance16x32_ssse3,
+                   &aom_highbd_masked_variance16x32_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance16x16_ssse3,
+                   &aom_highbd_masked_variance16x16_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance16x8_ssse3,
+                   &aom_highbd_masked_variance16x8_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance8x16_ssse3,
+                   &aom_highbd_masked_variance8x16_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance8x8_ssse3,
+                   &aom_highbd_masked_variance8x8_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance8x4_ssse3,
+                   &aom_highbd_masked_variance8x4_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance4x8_ssse3,
+                   &aom_highbd_masked_variance4x8_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_variance4x4_ssse3,
+                   &aom_highbd_masked_variance4x4_c, AOM_BITS_8),
+#if CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_10_masked_variance128x128_ssse3,
+                   &aom_highbd_10_masked_variance128x128_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance128x64_ssse3,
+                   &aom_highbd_10_masked_variance128x64_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance64x128_ssse3,
+                   &aom_highbd_10_masked_variance64x128_c, AOM_BITS_10),
+#endif  // CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_10_masked_variance64x64_ssse3,
+                   &aom_highbd_10_masked_variance64x64_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance64x32_ssse3,
+                   &aom_highbd_10_masked_variance64x32_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance32x64_ssse3,
+                   &aom_highbd_10_masked_variance32x64_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance32x32_ssse3,
+                   &aom_highbd_10_masked_variance32x32_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance32x16_ssse3,
+                   &aom_highbd_10_masked_variance32x16_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance16x32_ssse3,
+                   &aom_highbd_10_masked_variance16x32_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance16x16_ssse3,
+                   &aom_highbd_10_masked_variance16x16_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance16x8_ssse3,
+                   &aom_highbd_10_masked_variance16x8_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance8x16_ssse3,
+                   &aom_highbd_10_masked_variance8x16_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance8x8_ssse3,
+                   &aom_highbd_10_masked_variance8x8_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance8x4_ssse3,
+                   &aom_highbd_10_masked_variance8x4_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance4x8_ssse3,
+                   &aom_highbd_10_masked_variance4x8_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_variance4x4_ssse3,
+                   &aom_highbd_10_masked_variance4x4_c, AOM_BITS_10),
+#if CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_12_masked_variance128x128_ssse3,
+                   &aom_highbd_12_masked_variance128x128_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance128x64_ssse3,
+                   &aom_highbd_12_masked_variance128x64_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance64x128_ssse3,
+                   &aom_highbd_12_masked_variance64x128_c, AOM_BITS_12),
+#endif  // CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_12_masked_variance64x64_ssse3,
+                   &aom_highbd_12_masked_variance64x64_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance64x32_ssse3,
+                   &aom_highbd_12_masked_variance64x32_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance32x64_ssse3,
+                   &aom_highbd_12_masked_variance32x64_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance32x32_ssse3,
+                   &aom_highbd_12_masked_variance32x32_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance32x16_ssse3,
+                   &aom_highbd_12_masked_variance32x16_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance16x32_ssse3,
+                   &aom_highbd_12_masked_variance16x32_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance16x16_ssse3,
+                   &aom_highbd_12_masked_variance16x16_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance16x8_ssse3,
+                   &aom_highbd_12_masked_variance16x8_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance8x16_ssse3,
+                   &aom_highbd_12_masked_variance8x16_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance8x8_ssse3,
+                   &aom_highbd_12_masked_variance8x8_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance8x4_ssse3,
+                   &aom_highbd_12_masked_variance8x4_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance4x8_ssse3,
+                   &aom_highbd_12_masked_variance4x8_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_variance4x4_ssse3,
+                   &aom_highbd_12_masked_variance4x4_c, AOM_BITS_12)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSSE3_C_COMPARE, HighbdMaskedSubPixelVarianceTest,
+    ::testing::Values(
+#if CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_masked_sub_pixel_variance128x128_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance128x128_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance128x64_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance128x64_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance64x128_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance64x128_c, AOM_BITS_8),
+#endif  // CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_masked_sub_pixel_variance64x64_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance64x64_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance64x32_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance64x32_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance32x64_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance32x64_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance32x32_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance32x32_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance32x16_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance32x16_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance16x32_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance16x32_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance16x16_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance16x16_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance16x8_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance16x8_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance8x16_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance8x16_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance8x8_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance8x8_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance8x4_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance8x4_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance4x8_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance4x8_c, AOM_BITS_8),
+        make_tuple(&aom_highbd_masked_sub_pixel_variance4x4_ssse3,
+                   &aom_highbd_masked_sub_pixel_variance4x4_c, AOM_BITS_8),
+#if CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance128x128_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance128x128_c,
+                   AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance128x64_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance128x64_c,
+                   AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance64x128_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance64x128_c,
+                   AOM_BITS_10),
+#endif  // CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance64x64_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance64x64_c,
+                   AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance64x32_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance64x32_c,
+                   AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance32x64_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance32x64_c,
+                   AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance32x32_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance32x32_c,
+                   AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance32x16_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance32x16_c,
+                   AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance16x32_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance16x32_c,
+                   AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance16x16_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance16x16_c,
+                   AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance16x8_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance16x8_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance8x16_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance8x16_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance8x8_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance8x8_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance8x4_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance8x4_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance4x8_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance4x8_c, AOM_BITS_10),
+        make_tuple(&aom_highbd_10_masked_sub_pixel_variance4x4_ssse3,
+                   &aom_highbd_10_masked_sub_pixel_variance4x4_c, AOM_BITS_10),
+#if CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance128x128_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance128x128_c,
+                   AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance128x64_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance128x64_c,
+                   AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance64x128_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance64x128_c,
+                   AOM_BITS_12),
+#endif  // CONFIG_EXT_PARTITION
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance64x64_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance64x64_c,
+                   AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance64x32_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance64x32_c,
+                   AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance32x64_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance32x64_c,
+                   AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance32x32_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance32x32_c,
+                   AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance32x16_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance32x16_c,
+                   AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance16x32_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance16x32_c,
+                   AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance16x16_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance16x16_c,
+                   AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance16x8_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance16x8_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance8x16_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance8x16_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance8x8_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance8x8_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance8x4_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance8x4_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance4x8_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance4x8_c, AOM_BITS_12),
+        make_tuple(&aom_highbd_12_masked_sub_pixel_variance4x4_ssse3,
+                   &aom_highbd_12_masked_sub_pixel_variance4x4_c,
+                   AOM_BITS_12)));
+#endif  // CONFIG_HIGHBITDEPTH
+
+#endif  // HAVE_SSSE3
+}  // namespace
diff --git a/third_party/aom/test/md5_helper.h b/third_party/aom/test/md5_helper.h
new file mode 100644
index 0000000000..8c9d4f706f
--- /dev/null
+++ b/third_party/aom/test/md5_helper.h
@@ -0,0 +1,76 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef TEST_MD5_HELPER_H_
+#define TEST_MD5_HELPER_H_
+
+#include "./md5_utils.h"
+#include "aom/aom_decoder.h"
+
+namespace libaom_test {
+class MD5 {
+ public:
+  MD5() { MD5Init(&md5_); }
+
+  void Add(const aom_image_t *img) {
+    for (int plane = 0; plane < 3; ++plane) {
+      const uint8_t *buf = img->planes[plane];
+      // Calculate the width and height to do the md5 check. For the chroma
+      // plane, we never want to round down and thus skip a pixel so if
+      // we are shifting by 1 (chroma_shift) we add 1 before doing the shift.
+      // This works only for chroma_shift of 0 and 1.
+      const int bytes_per_sample =
+          (img->fmt & AOM_IMG_FMT_HIGHBITDEPTH) ? 2 : 1;
+      const int h =
+          plane ? (img->d_h + img->y_chroma_shift) >> img->y_chroma_shift
+                : img->d_h;
+      const int w =
+          (plane ? (img->d_w + img->x_chroma_shift) >> img->x_chroma_shift
+                 : img->d_w) *
+          bytes_per_sample;
+
+      for (int y = 0; y < h; ++y) {
+        MD5Update(&md5_, buf, w);
+        buf += img->stride[plane];
+      }
+    }
+  }
+
+  void Add(const uint8_t *data, size_t size) {
+    MD5Update(&md5_, data, static_cast<uint32_t>(size));
+  }
+
+  const char *Get(void) {
+    static const char hex[16] = {
+      '0', '1', '2', '3', '4', '5', '6', '7',
+      '8', '9', 'a', 'b', 'c', 'd', 'e', 'f',
+    };
+    uint8_t tmp[16];
+    MD5Context ctx_tmp = md5_;
+
+    MD5Final(tmp, &ctx_tmp);
+    for (int i = 0; i < 16; i++) {
+      res_[i * 2 + 0] = hex[tmp[i] >> 4];
+      res_[i * 2 + 1] = hex[tmp[i] & 0xf];
+    }
+    res_[32] = 0;
+
+    return res_;
+  }
+
+ protected:
+  char res_[33];
+  MD5Context md5_;
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_MD5_HELPER_H_
diff --git a/third_party/aom/test/minmax_test.cc b/third_party/aom/test/minmax_test.cc
new file mode 100644
index 0000000000..f825291927
--- /dev/null
+++ b/third_party/aom/test/minmax_test.cc
@@ -0,0 +1,131 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+#include "test/acm_random.h"
+#include "test/register_state_check.h"
+
+namespace {
+
+using ::libaom_test::ACMRandom;
+
+typedef void (*MinMaxFunc)(const uint8_t *a, int a_stride, const uint8_t *b,
+                           int b_stride, int *min, int *max);
+
+class MinMaxTest : public ::testing::TestWithParam<MinMaxFunc> {
+ public:
+  virtual void SetUp() {
+    mm_func_ = GetParam();
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+  }
+
+ protected:
+  MinMaxFunc mm_func_;
+  ACMRandom rnd_;
+};
+
+void reference_minmax(const uint8_t *a, int a_stride, const uint8_t *b,
+                      int b_stride, int *min_ret, int *max_ret) {
+  int min = 255;
+  int max = 0;
+  for (int i = 0; i < 8; i++) {
+    for (int j = 0; j < 8; j++) {
+      const int diff = abs(a[i * a_stride + j] - b[i * b_stride + j]);
+      if (min > diff) min = diff;
+      if (max < diff) max = diff;
+    }
+  }
+
+  *min_ret = min;
+  *max_ret = max;
+}
+
+TEST_P(MinMaxTest, MinValue) {
+  for (int i = 0; i < 64; i++) {
+    uint8_t a[64], b[64];
+    memset(a, 0, sizeof(a));
+    memset(b, 255, sizeof(b));
+    b[i] = i;  // Set a minimum difference of i.
+
+    int min, max;
+    ASM_REGISTER_STATE_CHECK(mm_func_(a, 8, b, 8, &min, &max));
+    EXPECT_EQ(255, max);
+    EXPECT_EQ(i, min);
+  }
+}
+
+TEST_P(MinMaxTest, MaxValue) {
+  for (int i = 0; i < 64; i++) {
+    uint8_t a[64], b[64];
+    memset(a, 0, sizeof(a));
+    memset(b, 0, sizeof(b));
+    b[i] = i;  // Set a maximum difference of i.
+
+    int min, max;
+    ASM_REGISTER_STATE_CHECK(mm_func_(a, 8, b, 8, &min, &max));
+    EXPECT_EQ(i, max);
+    EXPECT_EQ(0, min);
+  }
+}
+
+TEST_P(MinMaxTest, CompareReference) {
+  uint8_t a[64], b[64];
+  for (int j = 0; j < 64; j++) {
+    a[j] = rnd_.Rand8();
+    b[j] = rnd_.Rand8();
+  }
+
+  int min_ref, max_ref, min, max;
+  reference_minmax(a, 8, b, 8, &min_ref, &max_ref);
+  ASM_REGISTER_STATE_CHECK(mm_func_(a, 8, b, 8, &min, &max));
+  EXPECT_EQ(max_ref, max);
+  EXPECT_EQ(min_ref, min);
+}
+
+TEST_P(MinMaxTest, CompareReferenceAndVaryStride) {
+  uint8_t a[8 * 64], b[8 * 64];
+  for (int i = 0; i < 8 * 64; i++) {
+    a[i] = rnd_.Rand8();
+    b[i] = rnd_.Rand8();
+  }
+  for (int a_stride = 8; a_stride <= 64; a_stride += 8) {
+    for (int b_stride = 8; b_stride <= 64; b_stride += 8) {
+      int min_ref, max_ref, min, max;
+      reference_minmax(a, a_stride, b, b_stride, &min_ref, &max_ref);
+      ASM_REGISTER_STATE_CHECK(mm_func_(a, a_stride, b, b_stride, &min, &max));
+      EXPECT_EQ(max_ref, max) << "when a_stride = " << a_stride
+                              << " and b_stride = " << b_stride;
+      EXPECT_EQ(min_ref, min) << "when a_stride = " << a_stride
+                              << " and b_stride = " << b_stride;
+    }
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(C, MinMaxTest, ::testing::Values(&aom_minmax_8x8_c));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, MinMaxTest,
+                        ::testing::Values(&aom_minmax_8x8_sse2));
+#endif
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, MinMaxTest,
+                        ::testing::Values(&aom_minmax_8x8_neon));
+#endif
+
+}  // namespace
diff --git a/third_party/aom/test/motion_vector_test.cc b/third_party/aom/test/motion_vector_test.cc
new file mode 100644
index 0000000000..403a8f1a79
--- /dev/null
+++ b/third_party/aom/test/motion_vector_test.cc
@@ -0,0 +1,115 @@
+/*
+ * Copyright (c) 2017, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/yuv_video_source.h"
+
+namespace {
+#if defined(__has_feature)
+#if __has_feature(address_sanitizer)
+#define BUILDING_WITH_ASAN
+#endif
+#endif
+
+#define MAX_EXTREME_MV 1
+#define MIN_EXTREME_MV 2
+
+// Encoding modes
+const libaom_test::TestMode kEncodingModeVectors[] = {
+  ::libaom_test::kTwoPassGood, ::libaom_test::kOnePassGood,
+};
+
+// Encoding speeds
+const int kCpuUsedVectors[] = { 0, 1, 2, 3, 4, 5, 6, 7 };
+
+// MV test modes: 1 - always use maximum MV; 2 - always use minimum MV.
+const int kMVTestModes[] = { MAX_EXTREME_MV, MIN_EXTREME_MV };
+
+class MotionVectorTestLarge
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith3Params<libaom_test::TestMode, int,
+                                                 int> {
+ protected:
+  MotionVectorTestLarge()
+      : EncoderTest(GET_PARAM(0)), encoding_mode_(GET_PARAM(1)),
+        cpu_used_(GET_PARAM(2)), mv_test_mode_(GET_PARAM(3)) {}
+
+  virtual ~MotionVectorTestLarge() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(encoding_mode_);
+    if (encoding_mode_ != ::libaom_test::kRealTime) {
+      cfg_.g_lag_in_frames = 3;
+      cfg_.rc_end_usage = AOM_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = AOM_CBR;
+      cfg_.rc_buf_sz = 1000;
+      cfg_.rc_buf_initial_sz = 500;
+      cfg_.rc_buf_optimal_sz = 600;
+    }
+  }
+
+  virtual void PreEncodeFrameHook(::libaom_test::VideoSource *video,
+                                  ::libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AOME_SET_CPUUSED, cpu_used_);
+      encoder->Control(AV1E_ENABLE_MOTION_VECTOR_UNIT_TEST, mv_test_mode_);
+      if (encoding_mode_ != ::libaom_test::kRealTime) {
+        encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(AOME_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(AOME_SET_ARNR_STRENGTH, 5);
+      }
+    }
+  }
+
+  libaom_test::TestMode encoding_mode_;
+  int cpu_used_;
+  int mv_test_mode_;
+};
+
+TEST_P(MotionVectorTestLarge, OverallTest) {
+  int width = 3840;
+  int height = 2160;
+
+#ifdef BUILDING_WITH_ASAN
+  // On the 32-bit system, if using 4k test clip, an "out of memory" error
+  // occurs because of the AddressSanitizer instrumentation memory overhead.
+  // Here, reduce the test clip's resolution while testing on 32-bit system
+  // and AddressSanitizer is enabled.
+  if (sizeof(void *) == 4) {
+    width = 2048;
+    height = 1080;
+  }
+#endif
+
+  cfg_.rc_target_bitrate = 24000;
+  cfg_.g_profile = 0;
+  init_flags_ = AOM_CODEC_USE_PSNR;
+
+  testing::internal::scoped_ptr<libaom_test::VideoSource> video;
+  video.reset(new libaom_test::YUVVideoSource(
+      "niklas_640_480_30.yuv", AOM_IMG_FMT_I420, width, height, 30, 1, 0, 5));
+
+  ASSERT_TRUE(video.get() != NULL);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video.get()));
+}
+
+AV1_INSTANTIATE_TEST_CASE(MotionVectorTestLarge,
+                          ::testing::ValuesIn(kEncodingModeVectors),
+                          ::testing::ValuesIn(kCpuUsedVectors),
+                          ::testing::ValuesIn(kMVTestModes));
+}  // namespace
diff --git a/third_party/aom/test/obmc_sad_test.cc b/third_party/aom/test/obmc_sad_test.cc
new file mode 100644
index 0000000000..219c5d8109
--- /dev/null
+++ b/third_party/aom/test/obmc_sad_test.cc
@@ -0,0 +1,199 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "test/function_equivalence_test.h"
+#include "test/register_state_check.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+#define MAX_SB_SQUARE (MAX_SB_SIZE * MAX_SB_SIZE)
+
+using libaom_test::FunctionEquivalenceTest;
+
+namespace {
+
+static const int kIterations = 1000;
+static const int kMaskMax = 64;
+
+typedef unsigned int (*ObmcSadF)(const uint8_t *pre, int pre_stride,
+                                 const int32_t *wsrc, const int32_t *mask);
+typedef libaom_test::FuncParam<ObmcSadF> TestFuncs;
+
+////////////////////////////////////////////////////////////////////////////////
+// 8 bit
+////////////////////////////////////////////////////////////////////////////////
+
+class ObmcSadTest : public FunctionEquivalenceTest<ObmcSadF> {};
+
+TEST_P(ObmcSadTest, RandomValues) {
+  DECLARE_ALIGNED(32, uint8_t, pre[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, wsrc[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, mask[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    const int pre_stride = rng_(MAX_SB_SIZE + 1);
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      pre[i] = rng_.Rand8();
+      wsrc[i] = rng_.Rand8() * rng_(kMaskMax * kMaskMax + 1);
+      mask[i] = rng_(kMaskMax * kMaskMax + 1);
+    }
+
+    const unsigned int ref_res = params_.ref_func(pre, pre_stride, wsrc, mask);
+    unsigned int tst_res;
+    ASM_REGISTER_STATE_CHECK(tst_res =
+                                 params_.tst_func(pre, pre_stride, wsrc, mask));
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+TEST_P(ObmcSadTest, ExtremeValues) {
+  DECLARE_ALIGNED(32, uint8_t, pre[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, wsrc[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, mask[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < MAX_SB_SIZE && !HasFatalFailure(); ++iter) {
+    const int pre_stride = iter;
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      pre[i] = UINT8_MAX;
+      wsrc[i] = UINT8_MAX * kMaskMax * kMaskMax;
+      mask[i] = kMaskMax * kMaskMax;
+    }
+
+    const unsigned int ref_res = params_.ref_func(pre, pre_stride, wsrc, mask);
+    unsigned int tst_res;
+    ASM_REGISTER_STATE_CHECK(tst_res =
+                                 params_.tst_func(pre, pre_stride, wsrc, mask));
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+#if HAVE_SSE4_1
+#if CONFIG_MOTION_VAR
+const ObmcSadTest::ParamType sse4_functions[] = {
+#if CONFIG_EXT_PARTITION
+  TestFuncs(aom_obmc_sad128x128_c, aom_obmc_sad128x128_sse4_1),
+  TestFuncs(aom_obmc_sad128x64_c, aom_obmc_sad128x64_sse4_1),
+  TestFuncs(aom_obmc_sad64x128_c, aom_obmc_sad64x128_sse4_1),
+#endif  // CONFIG_EXT_PARTITION
+  TestFuncs(aom_obmc_sad64x64_c, aom_obmc_sad64x64_sse4_1),
+  TestFuncs(aom_obmc_sad64x32_c, aom_obmc_sad64x32_sse4_1),
+  TestFuncs(aom_obmc_sad32x64_c, aom_obmc_sad32x64_sse4_1),
+  TestFuncs(aom_obmc_sad32x32_c, aom_obmc_sad32x32_sse4_1),
+  TestFuncs(aom_obmc_sad32x16_c, aom_obmc_sad32x16_sse4_1),
+  TestFuncs(aom_obmc_sad16x32_c, aom_obmc_sad16x32_sse4_1),
+  TestFuncs(aom_obmc_sad16x16_c, aom_obmc_sad16x16_sse4_1),
+  TestFuncs(aom_obmc_sad16x8_c, aom_obmc_sad16x8_sse4_1),
+  TestFuncs(aom_obmc_sad8x16_c, aom_obmc_sad8x16_sse4_1),
+  TestFuncs(aom_obmc_sad8x8_c, aom_obmc_sad8x8_sse4_1),
+  TestFuncs(aom_obmc_sad8x4_c, aom_obmc_sad8x4_sse4_1),
+  TestFuncs(aom_obmc_sad4x8_c, aom_obmc_sad4x8_sse4_1),
+  TestFuncs(aom_obmc_sad4x4_c, aom_obmc_sad4x4_sse4_1)
+};
+
+INSTANTIATE_TEST_CASE_P(SSE4_1, ObmcSadTest,
+                        ::testing::ValuesIn(sse4_functions));
+#endif  // CONFIG_MOTION_VAR
+#endif  // HAVE_SSE4_1
+
+////////////////////////////////////////////////////////////////////////////////
+// High bit-depth
+////////////////////////////////////////////////////////////////////////////////
+
+#if CONFIG_HIGHBITDEPTH
+class ObmcSadHBDTest : public FunctionEquivalenceTest<ObmcSadF> {};
+
+TEST_P(ObmcSadHBDTest, RandomValues) {
+  DECLARE_ALIGNED(32, uint16_t, pre[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, wsrc[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, mask[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    const int pre_stride = rng_(MAX_SB_SIZE + 1);
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      pre[i] = rng_(1 << 12);
+      wsrc[i] = rng_(1 << 12) * rng_(kMaskMax * kMaskMax + 1);
+      mask[i] = rng_(kMaskMax * kMaskMax + 1);
+    }
+
+    const unsigned int ref_res =
+        params_.ref_func(CONVERT_TO_BYTEPTR(pre), pre_stride, wsrc, mask);
+    unsigned int tst_res;
+    ASM_REGISTER_STATE_CHECK(
+        tst_res =
+            params_.tst_func(CONVERT_TO_BYTEPTR(pre), pre_stride, wsrc, mask));
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+TEST_P(ObmcSadHBDTest, ExtremeValues) {
+  DECLARE_ALIGNED(32, uint16_t, pre[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, wsrc[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, mask[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < MAX_SB_SIZE && !HasFatalFailure(); ++iter) {
+    const int pre_stride = iter;
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      pre[i] = (1 << 12) - 1;
+      wsrc[i] = ((1 << 12) - 1) * kMaskMax * kMaskMax;
+      mask[i] = kMaskMax * kMaskMax;
+    }
+
+    const unsigned int ref_res =
+        params_.ref_func(CONVERT_TO_BYTEPTR(pre), pre_stride, wsrc, mask);
+    unsigned int tst_res;
+    ASM_REGISTER_STATE_CHECK(
+        tst_res =
+            params_.tst_func(CONVERT_TO_BYTEPTR(pre), pre_stride, wsrc, mask));
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+#if HAVE_SSE4_1
+#if CONFIG_MOTION_VAR
+ObmcSadHBDTest::ParamType sse4_functions_hbd[] = {
+#if CONFIG_EXT_PARTITION
+  TestFuncs(aom_highbd_obmc_sad128x128_c, aom_highbd_obmc_sad128x128_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad128x64_c, aom_highbd_obmc_sad128x64_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad64x128_c, aom_highbd_obmc_sad64x128_sse4_1),
+#endif  // CONFIG_EXT_PARTITION
+  TestFuncs(aom_highbd_obmc_sad64x64_c, aom_highbd_obmc_sad64x64_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad64x32_c, aom_highbd_obmc_sad64x32_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad32x64_c, aom_highbd_obmc_sad32x64_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad32x32_c, aom_highbd_obmc_sad32x32_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad32x16_c, aom_highbd_obmc_sad32x16_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad16x32_c, aom_highbd_obmc_sad16x32_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad16x16_c, aom_highbd_obmc_sad16x16_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad16x8_c, aom_highbd_obmc_sad16x8_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad8x16_c, aom_highbd_obmc_sad8x16_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad8x8_c, aom_highbd_obmc_sad8x8_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad8x4_c, aom_highbd_obmc_sad8x4_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad4x8_c, aom_highbd_obmc_sad4x8_sse4_1),
+  TestFuncs(aom_highbd_obmc_sad4x4_c, aom_highbd_obmc_sad4x4_sse4_1)
+};
+
+INSTANTIATE_TEST_CASE_P(SSE4_1, ObmcSadHBDTest,
+                        ::testing::ValuesIn(sse4_functions_hbd));
+#endif  // CONFIG_MOTION_VAR
+#endif  // HAVE_SSE4_1
+#endif  // CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/obmc_variance_test.cc b/third_party/aom/test/obmc_variance_test.cc
new file mode 100644
index 0000000000..1b30645a52
--- /dev/null
+++ b/third_party/aom/test/obmc_variance_test.cc
@@ -0,0 +1,297 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/acm_random.h"
+
+#include "test/function_equivalence_test.h"
+#include "test/register_state_check.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+
+#define MAX_SB_SQUARE (MAX_SB_SIZE * MAX_SB_SIZE)
+
+using libaom_test::ACMRandom;
+using libaom_test::FunctionEquivalenceTest;
+
+namespace {
+
+static const int kIterations = 1000;
+static const int kMaskMax = 64;
+
+typedef unsigned int (*ObmcVarF)(const uint8_t *pre, int pre_stride,
+                                 const int32_t *wsrc, const int32_t *mask,
+                                 unsigned int *sse);
+typedef libaom_test::FuncParam<ObmcVarF> TestFuncs;
+
+////////////////////////////////////////////////////////////////////////////////
+// 8 bit
+////////////////////////////////////////////////////////////////////////////////
+
+class ObmcVarianceTest : public FunctionEquivalenceTest<ObmcVarF> {};
+
+TEST_P(ObmcVarianceTest, RandomValues) {
+  DECLARE_ALIGNED(32, uint8_t, pre[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, wsrc[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, mask[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    const int pre_stride = this->rng_(MAX_SB_SIZE + 1);
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      pre[i] = this->rng_.Rand8();
+      wsrc[i] = this->rng_.Rand8() * this->rng_(kMaskMax * kMaskMax + 1);
+      mask[i] = this->rng_(kMaskMax * kMaskMax + 1);
+    }
+
+    unsigned int ref_sse, tst_sse;
+    const unsigned int ref_res =
+        params_.ref_func(pre, pre_stride, wsrc, mask, &ref_sse);
+    unsigned int tst_res;
+    ASM_REGISTER_STATE_CHECK(
+        tst_res = params_.tst_func(pre, pre_stride, wsrc, mask, &tst_sse));
+
+    ASSERT_EQ(ref_res, tst_res);
+    ASSERT_EQ(ref_sse, tst_sse);
+  }
+}
+
+TEST_P(ObmcVarianceTest, ExtremeValues) {
+  DECLARE_ALIGNED(32, uint8_t, pre[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, wsrc[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, mask[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < MAX_SB_SIZE && !HasFatalFailure(); ++iter) {
+    const int pre_stride = iter;
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      pre[i] = UINT8_MAX;
+      wsrc[i] = UINT8_MAX * kMaskMax * kMaskMax;
+      mask[i] = kMaskMax * kMaskMax;
+    }
+
+    unsigned int ref_sse, tst_sse;
+    const unsigned int ref_res =
+        params_.ref_func(pre, pre_stride, wsrc, mask, &ref_sse);
+    unsigned int tst_res;
+    ASM_REGISTER_STATE_CHECK(
+        tst_res = params_.tst_func(pre, pre_stride, wsrc, mask, &tst_sse));
+
+    ASSERT_EQ(ref_res, tst_res);
+    ASSERT_EQ(ref_sse, tst_sse);
+  }
+}
+
+#if HAVE_SSE4_1
+#if CONFIG_MOTION_VAR
+const ObmcVarianceTest::ParamType sse4_functions[] = {
+#if CONFIG_EXT_PARTITION
+  TestFuncs(aom_obmc_variance128x128_c, aom_obmc_variance128x128_sse4_1),
+  TestFuncs(aom_obmc_variance128x64_c, aom_obmc_variance128x64_sse4_1),
+  TestFuncs(aom_obmc_variance64x128_c, aom_obmc_variance64x128_sse4_1),
+#endif  // CONFIG_EXT_PARTITION
+  TestFuncs(aom_obmc_variance64x64_c, aom_obmc_variance64x64_sse4_1),
+  TestFuncs(aom_obmc_variance64x32_c, aom_obmc_variance64x32_sse4_1),
+  TestFuncs(aom_obmc_variance32x64_c, aom_obmc_variance32x64_sse4_1),
+  TestFuncs(aom_obmc_variance32x32_c, aom_obmc_variance32x32_sse4_1),
+  TestFuncs(aom_obmc_variance32x16_c, aom_obmc_variance32x16_sse4_1),
+  TestFuncs(aom_obmc_variance16x32_c, aom_obmc_variance16x32_sse4_1),
+  TestFuncs(aom_obmc_variance16x16_c, aom_obmc_variance16x16_sse4_1),
+  TestFuncs(aom_obmc_variance16x8_c, aom_obmc_variance16x8_sse4_1),
+  TestFuncs(aom_obmc_variance8x16_c, aom_obmc_variance8x16_sse4_1),
+  TestFuncs(aom_obmc_variance8x8_c, aom_obmc_variance8x8_sse4_1),
+  TestFuncs(aom_obmc_variance8x4_c, aom_obmc_variance8x4_sse4_1),
+  TestFuncs(aom_obmc_variance4x8_c, aom_obmc_variance4x8_sse4_1),
+  TestFuncs(aom_obmc_variance4x4_c, aom_obmc_variance4x4_sse4_1)
+};
+
+INSTANTIATE_TEST_CASE_P(SSE4_1, ObmcVarianceTest,
+                        ::testing::ValuesIn(sse4_functions));
+#endif  // CONFIG_MOTION_VAR
+#endif  // HAVE_SSE4_1
+
+////////////////////////////////////////////////////////////////////////////////
+// High bit-depth
+////////////////////////////////////////////////////////////////////////////////
+
+#if CONFIG_HIGHBITDEPTH
+class ObmcVarianceHBDTest : public FunctionEquivalenceTest<ObmcVarF> {};
+
+TEST_P(ObmcVarianceHBDTest, RandomValues) {
+  DECLARE_ALIGNED(32, uint16_t, pre[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, wsrc[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, mask[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    const int pre_stride = this->rng_(MAX_SB_SIZE + 1);
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      pre[i] = this->rng_(1 << params_.bit_depth);
+      wsrc[i] = this->rng_(1 << params_.bit_depth) *
+                this->rng_(kMaskMax * kMaskMax + 1);
+      mask[i] = this->rng_(kMaskMax * kMaskMax + 1);
+    }
+
+    unsigned int ref_sse, tst_sse;
+    const unsigned int ref_res = params_.ref_func(
+        CONVERT_TO_BYTEPTR(pre), pre_stride, wsrc, mask, &ref_sse);
+    unsigned int tst_res;
+    ASM_REGISTER_STATE_CHECK(tst_res = params_.tst_func(CONVERT_TO_BYTEPTR(pre),
+                                                        pre_stride, wsrc, mask,
+                                                        &tst_sse));
+
+    ASSERT_EQ(ref_res, tst_res);
+    ASSERT_EQ(ref_sse, tst_sse);
+  }
+}
+
+TEST_P(ObmcVarianceHBDTest, ExtremeValues) {
+  DECLARE_ALIGNED(32, uint16_t, pre[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, wsrc[MAX_SB_SQUARE]);
+  DECLARE_ALIGNED(32, int32_t, mask[MAX_SB_SQUARE]);
+
+  for (int iter = 0; iter < MAX_SB_SIZE && !HasFatalFailure(); ++iter) {
+    const int pre_stride = iter;
+
+    for (int i = 0; i < MAX_SB_SQUARE; ++i) {
+      pre[i] = (1 << params_.bit_depth) - 1;
+      wsrc[i] = ((1 << params_.bit_depth) - 1) * kMaskMax * kMaskMax;
+      mask[i] = kMaskMax * kMaskMax;
+    }
+
+    unsigned int ref_sse, tst_sse;
+    const unsigned int ref_res = params_.ref_func(
+        CONVERT_TO_BYTEPTR(pre), pre_stride, wsrc, mask, &ref_sse);
+    unsigned int tst_res;
+    ASM_REGISTER_STATE_CHECK(tst_res = params_.tst_func(CONVERT_TO_BYTEPTR(pre),
+                                                        pre_stride, wsrc, mask,
+                                                        &tst_sse));
+
+    ASSERT_EQ(ref_res, tst_res);
+    ASSERT_EQ(ref_sse, tst_sse);
+  }
+}
+
+#if HAVE_SSE4_1
+#if CONFIG_MOTION_VAR
+ObmcVarianceHBDTest::ParamType sse4_functions_hbd[] = {
+#if CONFIG_EXT_PARTITION
+  TestFuncs(aom_highbd_obmc_variance128x128_c,
+            aom_highbd_obmc_variance128x128_sse4_1, 8),
+  TestFuncs(aom_highbd_obmc_variance128x64_c,
+            aom_highbd_obmc_variance128x64_sse4_1, 8),
+  TestFuncs(aom_highbd_obmc_variance64x128_c,
+            aom_highbd_obmc_variance64x128_sse4_1, 8),
+#endif  // CONFIG_EXT_PARTITION
+  TestFuncs(aom_highbd_obmc_variance64x64_c,
+            aom_highbd_obmc_variance64x64_sse4_1, 8),
+  TestFuncs(aom_highbd_obmc_variance64x32_c,
+            aom_highbd_obmc_variance64x32_sse4_1, 8),
+  TestFuncs(aom_highbd_obmc_variance32x64_c,
+            aom_highbd_obmc_variance32x64_sse4_1, 8),
+  TestFuncs(aom_highbd_obmc_variance32x32_c,
+            aom_highbd_obmc_variance32x32_sse4_1, 8),
+  TestFuncs(aom_highbd_obmc_variance32x16_c,
+            aom_highbd_obmc_variance32x16_sse4_1, 8),
+  TestFuncs(aom_highbd_obmc_variance16x32_c,
+            aom_highbd_obmc_variance16x32_sse4_1, 8),
+  TestFuncs(aom_highbd_obmc_variance16x16_c,
+            aom_highbd_obmc_variance16x16_sse4_1, 8),
+  TestFuncs(aom_highbd_obmc_variance16x8_c, aom_highbd_obmc_variance16x8_sse4_1,
+            8),
+  TestFuncs(aom_highbd_obmc_variance8x16_c, aom_highbd_obmc_variance8x16_sse4_1,
+            8),
+  TestFuncs(aom_highbd_obmc_variance8x8_c, aom_highbd_obmc_variance8x8_sse4_1,
+            8),
+  TestFuncs(aom_highbd_obmc_variance8x4_c, aom_highbd_obmc_variance8x4_sse4_1,
+            8),
+  TestFuncs(aom_highbd_obmc_variance4x8_c, aom_highbd_obmc_variance4x8_sse4_1,
+            8),
+  TestFuncs(aom_highbd_obmc_variance4x4_c, aom_highbd_obmc_variance4x4_sse4_1,
+            8),
+#if CONFIG_EXT_PARTITION
+  TestFuncs(aom_highbd_10_obmc_variance128x128_c,
+            aom_highbd_10_obmc_variance128x128_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance128x64_c,
+            aom_highbd_10_obmc_variance128x64_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance64x128_c,
+            aom_highbd_10_obmc_variance64x128_sse4_1, 10),
+#endif  // CONFIG_EXT_PARTITION
+  TestFuncs(aom_highbd_10_obmc_variance64x64_c,
+            aom_highbd_10_obmc_variance64x64_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance64x32_c,
+            aom_highbd_10_obmc_variance64x32_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance32x64_c,
+            aom_highbd_10_obmc_variance32x64_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance32x32_c,
+            aom_highbd_10_obmc_variance32x32_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance32x16_c,
+            aom_highbd_10_obmc_variance32x16_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance16x32_c,
+            aom_highbd_10_obmc_variance16x32_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance16x16_c,
+            aom_highbd_10_obmc_variance16x16_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance16x8_c,
+            aom_highbd_10_obmc_variance16x8_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance8x16_c,
+            aom_highbd_10_obmc_variance8x16_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance8x8_c,
+            aom_highbd_10_obmc_variance8x8_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance8x4_c,
+            aom_highbd_10_obmc_variance8x4_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance4x8_c,
+            aom_highbd_10_obmc_variance4x8_sse4_1, 10),
+  TestFuncs(aom_highbd_10_obmc_variance4x4_c,
+            aom_highbd_10_obmc_variance4x4_sse4_1, 10),
+#if CONFIG_EXT_PARTITION
+  TestFuncs(aom_highbd_12_obmc_variance128x128_c,
+            aom_highbd_12_obmc_variance128x128_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance128x64_c,
+            aom_highbd_12_obmc_variance128x64_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance64x128_c,
+            aom_highbd_12_obmc_variance64x128_sse4_1, 12),
+#endif  // CONFIG_EXT_PARTITION
+  TestFuncs(aom_highbd_12_obmc_variance64x64_c,
+            aom_highbd_12_obmc_variance64x64_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance64x32_c,
+            aom_highbd_12_obmc_variance64x32_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance32x64_c,
+            aom_highbd_12_obmc_variance32x64_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance32x32_c,
+            aom_highbd_12_obmc_variance32x32_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance32x16_c,
+            aom_highbd_12_obmc_variance32x16_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance16x32_c,
+            aom_highbd_12_obmc_variance16x32_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance16x16_c,
+            aom_highbd_12_obmc_variance16x16_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance16x8_c,
+            aom_highbd_12_obmc_variance16x8_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance8x16_c,
+            aom_highbd_12_obmc_variance8x16_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance8x8_c,
+            aom_highbd_12_obmc_variance8x8_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance8x4_c,
+            aom_highbd_12_obmc_variance8x4_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance4x8_c,
+            aom_highbd_12_obmc_variance4x8_sse4_1, 12),
+  TestFuncs(aom_highbd_12_obmc_variance4x4_c,
+            aom_highbd_12_obmc_variance4x4_sse4_1, 12)
+};
+
+INSTANTIATE_TEST_CASE_P(SSE4_1, ObmcVarianceHBDTest,
+                        ::testing::ValuesIn(sse4_functions_hbd));
+#endif  // CONFIG_MOTION_VAR
+#endif  // HAVE_SSE4_1
+#endif  // CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/partial_idct_test.cc b/third_party/aom/test/partial_idct_test.cc
new file mode 100644
index 0000000000..0899b60c32
--- /dev/null
+++ b/third_party/aom/test/partial_idct_test.cc
@@ -0,0 +1,485 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <math.h>
+#include <stdlib.h>
+#include <string.h>
+#include <limits>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "av1/common/blockd.h"
+#include "av1/common/scan.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/aom_timer.h"
+
+using libaom_test::ACMRandom;
+
+namespace {
+typedef void (*FwdTxfmFunc)(const int16_t *in, tran_low_t *out, int stride);
+typedef void (*InvTxfmFunc)(const tran_low_t *in, uint8_t *out, int stride);
+typedef void (*InvTxfmWithBdFunc)(const tran_low_t *in, uint8_t *out,
+                                  int stride, int bd);
+
+template <InvTxfmFunc fn>
+void wrapper(const tran_low_t *in, uint8_t *out, int stride, int bd) {
+  (void)bd;
+  fn(in, out, stride);
+}
+
+#if CONFIG_HIGHBITDEPTH
+template <InvTxfmWithBdFunc fn>
+void highbd_wrapper(const tran_low_t *in, uint8_t *out, int stride, int bd) {
+  fn(in, CONVERT_TO_BYTEPTR(out), stride, bd);
+}
+#endif
+
+typedef std::tr1::tuple<FwdTxfmFunc, InvTxfmWithBdFunc, InvTxfmWithBdFunc,
+                        TX_SIZE, int, int, int>
+    PartialInvTxfmParam;
+const int kMaxNumCoeffs = 1024;
+const int kCountTestBlock = 1000;
+
+class PartialIDctTest : public ::testing::TestWithParam<PartialInvTxfmParam> {
+ public:
+  virtual ~PartialIDctTest() {}
+  virtual void SetUp() {
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+    ftxfm_ = GET_PARAM(0);
+    full_itxfm_ = GET_PARAM(1);
+    partial_itxfm_ = GET_PARAM(2);
+    tx_size_ = GET_PARAM(3);
+    last_nonzero_ = GET_PARAM(4);
+    bit_depth_ = GET_PARAM(5);
+    pixel_size_ = GET_PARAM(6);
+    mask_ = (1 << bit_depth_) - 1;
+
+    switch (tx_size_) {
+      case TX_4X4: size_ = 4; break;
+      case TX_8X8: size_ = 8; break;
+      case TX_16X16: size_ = 16; break;
+      case TX_32X32: size_ = 32; break;
+      default: FAIL() << "Wrong Size!"; break;
+    }
+
+    // Randomize stride_ to a value less than or equal to 1024
+    stride_ = rnd_(1024) + 1;
+    if (stride_ < size_) {
+      stride_ = size_;
+    }
+    // Align stride_ to 16 if it's bigger than 16.
+    if (stride_ > 16) {
+      stride_ &= ~15;
+    }
+
+    input_block_size_ = size_ * size_;
+    output_block_size_ = size_ * stride_;
+
+    input_block_ = reinterpret_cast<tran_low_t *>(
+        aom_memalign(16, sizeof(*input_block_) * input_block_size_));
+    output_block_ = reinterpret_cast<uint8_t *>(
+        aom_memalign(16, pixel_size_ * output_block_size_));
+    output_block_ref_ = reinterpret_cast<uint8_t *>(
+        aom_memalign(16, pixel_size_ * output_block_size_));
+  }
+
+  virtual void TearDown() {
+    aom_free(input_block_);
+    input_block_ = NULL;
+    aom_free(output_block_);
+    output_block_ = NULL;
+    aom_free(output_block_ref_);
+    output_block_ref_ = NULL;
+    libaom_test::ClearSystemState();
+  }
+
+  void InitMem() {
+    memset(input_block_, 0, sizeof(*input_block_) * input_block_size_);
+    if (pixel_size_ == 1) {
+      for (int j = 0; j < output_block_size_; ++j) {
+        output_block_[j] = output_block_ref_[j] = rnd_.Rand16() & mask_;
+      }
+    } else {
+      ASSERT_EQ(2, pixel_size_);
+      uint16_t *const output = reinterpret_cast<uint16_t *>(output_block_);
+      uint16_t *const output_ref =
+          reinterpret_cast<uint16_t *>(output_block_ref_);
+      for (int j = 0; j < output_block_size_; ++j) {
+        output[j] = output_ref[j] = rnd_.Rand16() & mask_;
+      }
+    }
+  }
+
+  void InitInput() {
+    const int max_coeff = 32766 / 4;
+    int max_energy_leftover = max_coeff * max_coeff;
+    for (int j = 0; j < last_nonzero_; ++j) {
+      int16_t coeff = static_cast<int16_t>(sqrt(1.0 * max_energy_leftover) *
+                                           (rnd_.Rand16() - 32768) / 65536);
+      max_energy_leftover -= coeff * coeff;
+      if (max_energy_leftover < 0) {
+        max_energy_leftover = 0;
+        coeff = 0;
+      }
+      input_block_[av1_default_scan_orders[tx_size_].scan[j]] = coeff;
+    }
+  }
+
+ protected:
+  int last_nonzero_;
+  TX_SIZE tx_size_;
+  tran_low_t *input_block_;
+  uint8_t *output_block_;
+  uint8_t *output_block_ref_;
+  int size_;
+  int stride_;
+  int pixel_size_;
+  int input_block_size_;
+  int output_block_size_;
+  int bit_depth_;
+  int mask_;
+  FwdTxfmFunc ftxfm_;
+  InvTxfmWithBdFunc full_itxfm_;
+  InvTxfmWithBdFunc partial_itxfm_;
+  ACMRandom rnd_;
+};
+
+TEST_P(PartialIDctTest, RunQuantCheck) {
+  DECLARE_ALIGNED(16, int16_t, input_extreme_block[kMaxNumCoeffs]);
+  DECLARE_ALIGNED(16, tran_low_t, output_ref_block[kMaxNumCoeffs]);
+
+  InitMem();
+  for (int i = 0; i < kCountTestBlock; ++i) {
+    // Initialize a test block with input range [-mask_, mask_].
+    if (i == 0) {
+      for (int k = 0; k < input_block_size_; ++k) {
+        input_extreme_block[k] = mask_;
+      }
+    } else if (i == 1) {
+      for (int k = 0; k < input_block_size_; ++k) {
+        input_extreme_block[k] = -mask_;
+      }
+    } else {
+      for (int k = 0; k < input_block_size_; ++k) {
+        input_extreme_block[k] = rnd_.Rand8() % 2 ? mask_ : -mask_;
+      }
+    }
+
+    ftxfm_(input_extreme_block, output_ref_block, size_);
+
+    // quantization with minimum allowed step sizes
+    input_block_[0] = (output_ref_block[0] / 4) * 4;
+    for (int k = 1; k < last_nonzero_; ++k) {
+      const int pos = av1_default_scan_orders[tx_size_].scan[k];
+      input_block_[pos] = (output_ref_block[pos] / 4) * 4;
+    }
+
+    ASM_REGISTER_STATE_CHECK(
+        full_itxfm_(input_block_, output_block_ref_, stride_, bit_depth_));
+    ASM_REGISTER_STATE_CHECK(
+        partial_itxfm_(input_block_, output_block_, stride_, bit_depth_));
+    ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
+                        pixel_size_ * output_block_size_))
+        << "Error: partial inverse transform produces different results";
+  }
+}
+
+TEST_P(PartialIDctTest, ResultsMatch) {
+  for (int i = 0; i < kCountTestBlock; ++i) {
+    InitMem();
+    InitInput();
+
+    ASM_REGISTER_STATE_CHECK(
+        full_itxfm_(input_block_, output_block_ref_, stride_, bit_depth_));
+    ASM_REGISTER_STATE_CHECK(
+        partial_itxfm_(input_block_, output_block_, stride_, bit_depth_));
+    ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
+                        pixel_size_ * output_block_size_))
+        << "Error: partial inverse transform produces different results";
+  }
+}
+
+TEST_P(PartialIDctTest, AddOutputBlock) {
+  for (int i = 0; i < kCountTestBlock; ++i) {
+    InitMem();
+    for (int j = 0; j < last_nonzero_; ++j) {
+      input_block_[av1_default_scan_orders[tx_size_].scan[j]] = 10;
+    }
+
+    ASM_REGISTER_STATE_CHECK(
+        full_itxfm_(input_block_, output_block_ref_, stride_, bit_depth_));
+    ASM_REGISTER_STATE_CHECK(
+        partial_itxfm_(input_block_, output_block_, stride_, bit_depth_));
+    ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
+                        pixel_size_ * output_block_size_))
+        << "Error: Transform results are not correctly added to output.";
+  }
+}
+
+TEST_P(PartialIDctTest, SingleExtremeCoeff) {
+  const int16_t max_coeff = std::numeric_limits<int16_t>::max();
+  const int16_t min_coeff = std::numeric_limits<int16_t>::min();
+  for (int i = 0; i < last_nonzero_; ++i) {
+    memset(input_block_, 0, sizeof(*input_block_) * input_block_size_);
+    // Run once for min and once for max.
+    for (int j = 0; j < 2; ++j) {
+      const int coeff = j ? min_coeff : max_coeff;
+
+      memset(output_block_, 0, pixel_size_ * output_block_size_);
+      memset(output_block_ref_, 0, pixel_size_ * output_block_size_);
+      input_block_[av1_default_scan_orders[tx_size_].scan[i]] = coeff;
+
+      ASM_REGISTER_STATE_CHECK(
+          full_itxfm_(input_block_, output_block_ref_, stride_, bit_depth_));
+      ASM_REGISTER_STATE_CHECK(
+          partial_itxfm_(input_block_, output_block_, stride_, bit_depth_));
+      ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
+                          pixel_size_ * output_block_size_))
+          << "Error: Fails with single coeff of " << coeff << " at " << i
+          << ".";
+    }
+  }
+}
+
+TEST_P(PartialIDctTest, DISABLED_Speed) {
+  // Keep runtime stable with transform size.
+  const int kCountSpeedTestBlock = 500000000 / input_block_size_;
+  InitMem();
+  InitInput();
+
+  for (int i = 0; i < kCountSpeedTestBlock; ++i) {
+    ASM_REGISTER_STATE_CHECK(
+        full_itxfm_(input_block_, output_block_ref_, stride_, bit_depth_));
+  }
+  aom_usec_timer timer;
+  aom_usec_timer_start(&timer);
+  for (int i = 0; i < kCountSpeedTestBlock; ++i) {
+    partial_itxfm_(input_block_, output_block_, stride_, bit_depth_);
+  }
+  libaom_test::ClearSystemState();
+  aom_usec_timer_mark(&timer);
+  const int elapsed_time =
+      static_cast<int>(aom_usec_timer_elapsed(&timer) / 1000);
+  printf("idct%dx%d_%d (bitdepth %d) time: %5d ms\n", size_, size_,
+         last_nonzero_, bit_depth_, elapsed_time);
+
+  ASSERT_EQ(0, memcmp(output_block_ref_, output_block_,
+                      pixel_size_ * output_block_size_))
+      << "Error: partial inverse transform produces different results";
+}
+
+using std::tr1::make_tuple;
+
+const PartialInvTxfmParam c_partial_idct_tests[] = {
+#if CONFIG_HIGHBITDEPTH
+  make_tuple(&aom_highbd_fdct4x4_c,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_c>,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_c>, TX_4X4, 16, 8, 2),
+  make_tuple(&aom_highbd_fdct4x4_c,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_c>,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_c>, TX_4X4, 16, 10, 2),
+  make_tuple(&aom_highbd_fdct4x4_c,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_c>,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_c>, TX_4X4, 16, 12, 2),
+  make_tuple(&aom_highbd_fdct4x4_c,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_c>,
+             &highbd_wrapper<aom_highbd_idct4x4_1_add_c>, TX_4X4, 1, 8, 2),
+  make_tuple(&aom_highbd_fdct4x4_c,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_c>,
+             &highbd_wrapper<aom_highbd_idct4x4_1_add_c>, TX_4X4, 1, 10, 2),
+  make_tuple(&aom_highbd_fdct4x4_c,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_c>,
+             &highbd_wrapper<aom_highbd_idct4x4_1_add_c>, TX_4X4, 1, 12, 2),
+#endif  // CONFIG_HIGHBITDEPTH
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1024_add_c>, TX_32X32, 1024, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_135_add_c>, TX_32X32, 135, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_34_add_c>, TX_32X32, 34, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1_add_c>, TX_32X32, 1, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_256_add_c>, TX_16X16, 256, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_38_add_c>, TX_16X16, 38, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_10_add_c>, TX_16X16, 10, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_1_add_c>, TX_16X16, 1, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_64_add_c>, TX_8X8, 64, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_12_add_c>, TX_8X8, 12, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_1_add_c>, TX_8X8, 1, 8, 1),
+  make_tuple(&aom_fdct4x4_c, &wrapper<aom_idct4x4_16_add_c>,
+             &wrapper<aom_idct4x4_16_add_c>, TX_4X4, 16, 8, 1),
+  make_tuple(&aom_fdct4x4_c, &wrapper<aom_idct4x4_16_add_c>,
+             &wrapper<aom_idct4x4_1_add_c>, TX_4X4, 1, 8, 1)
+};
+
+INSTANTIATE_TEST_CASE_P(C, PartialIDctTest,
+                        ::testing::ValuesIn(c_partial_idct_tests));
+
+#if HAVE_NEON && !CONFIG_HIGHBITDEPTH
+const PartialInvTxfmParam neon_partial_idct_tests[] = {
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1_add_neon>, TX_32X32, 1, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_10_add_neon>, TX_16X16, 10, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_1_add_neon>, TX_16X16, 1, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_12_add_neon>, TX_8X8, 12, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_1_add_neon>, TX_8X8, 1, 8, 1),
+  make_tuple(&aom_fdct4x4_c, &wrapper<aom_idct4x4_16_add_c>,
+             &wrapper<aom_idct4x4_1_add_neon>, TX_4X4, 1, 8, 1)
+};
+
+INSTANTIATE_TEST_CASE_P(NEON, PartialIDctTest,
+                        ::testing::ValuesIn(neon_partial_idct_tests));
+#endif  // HAVE_NEON && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_SSE2
+const PartialInvTxfmParam sse2_partial_idct_tests[] = {
+#if CONFIG_HIGHBITDEPTH
+  make_tuple(&aom_highbd_fdct4x4_c,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_c>,
+             &highbd_wrapper<aom_highbd_idct4x4_16_add_sse2>, TX_4X4, 16, 8, 2),
+  make_tuple(
+      &aom_highbd_fdct4x4_c, &highbd_wrapper<aom_highbd_idct4x4_16_add_c>,
+      &highbd_wrapper<aom_highbd_idct4x4_16_add_sse2>, TX_4X4, 16, 10, 2),
+  make_tuple(
+      &aom_highbd_fdct4x4_c, &highbd_wrapper<aom_highbd_idct4x4_16_add_c>,
+      &highbd_wrapper<aom_highbd_idct4x4_16_add_sse2>, TX_4X4, 16, 12, 2),
+#endif  // CONFIG_HIGHBITDEPTH
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1024_add_sse2>, TX_32X32, 1024, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1024_add_sse2>, TX_32X32, 135, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_34_add_sse2>, TX_32X32, 34, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1_add_sse2>, TX_32X32, 1, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_256_add_sse2>, TX_16X16, 256, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_10_add_sse2>, TX_16X16, 10, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_1_add_sse2>, TX_16X16, 1, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_64_add_sse2>, TX_8X8, 64, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_12_add_sse2>, TX_8X8, 12, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_1_add_sse2>, TX_8X8, 1, 8, 1),
+  make_tuple(&aom_fdct4x4_c, &wrapper<aom_idct4x4_16_add_c>,
+             &wrapper<aom_idct4x4_16_add_sse2>, TX_4X4, 16, 8, 1),
+  make_tuple(&aom_fdct4x4_c, &wrapper<aom_idct4x4_16_add_c>,
+             &wrapper<aom_idct4x4_1_add_sse2>, TX_4X4, 1, 8, 1)
+};
+
+INSTANTIATE_TEST_CASE_P(SSE2, PartialIDctTest,
+                        ::testing::ValuesIn(sse2_partial_idct_tests));
+
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3
+const PartialInvTxfmParam ssse3_partial_idct_tests[] = {
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1024_add_ssse3>, TX_32X32, 1024, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_135_add_ssse3>, TX_32X32, 135, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_34_add_ssse3>, TX_32X32, 34, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_64_add_ssse3>, TX_8X8, 64, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_12_add_ssse3>, TX_8X8, 12, 8, 1)
+};
+
+INSTANTIATE_TEST_CASE_P(SSSE3, PartialIDctTest,
+                        ::testing::ValuesIn(ssse3_partial_idct_tests));
+#endif  // HAVE_SSSE3
+
+#if HAVE_DSPR2 && !CONFIG_HIGHBITDEPTH
+const PartialInvTxfmParam dspr2_partial_idct_tests[] = {
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1024_add_dspr2>, TX_32X32, 1024, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1024_add_dspr2>, TX_32X32, 135, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_34_add_dspr2>, TX_32X32, 34, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1_add_dspr2>, TX_32X32, 1, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_256_add_dspr2>, TX_16X16, 256, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_10_add_dspr2>, TX_16X16, 10, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_1_add_dspr2>, TX_16X16, 1, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_64_add_dspr2>, TX_8X8, 64, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_12_add_dspr2>, TX_8X8, 12, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_1_add_dspr2>, TX_8X8, 1, 8, 1),
+  make_tuple(&aom_fdct4x4_c, &wrapper<aom_idct4x4_16_add_c>,
+             &wrapper<aom_idct4x4_16_add_dspr2>, TX_4X4, 16, 8, 1),
+  make_tuple(&aom_fdct4x4_c, &wrapper<aom_idct4x4_16_add_c>,
+             &wrapper<aom_idct4x4_1_add_dspr2>, TX_4X4, 1, 8, 1)
+};
+
+INSTANTIATE_TEST_CASE_P(DSPR2, PartialIDctTest,
+                        ::testing::ValuesIn(dspr2_partial_idct_tests));
+#endif  // HAVE_DSPR2 && !CONFIG_HIGHBITDEPTH
+
+#if HAVE_MSA && !CONFIG_HIGHBITDEPTH
+const PartialInvTxfmParam msa_partial_idct_tests[] = {
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1024_add_msa>, TX_32X32, 1024, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1024_add_msa>, TX_32X32, 135, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_34_add_msa>, TX_32X32, 34, 8, 1),
+  make_tuple(&aom_fdct32x32_c, &wrapper<aom_idct32x32_1024_add_c>,
+             &wrapper<aom_idct32x32_1_add_msa>, TX_32X32, 1, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_256_add_msa>, TX_16X16, 256, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_10_add_msa>, TX_16X16, 10, 8, 1),
+  make_tuple(&aom_fdct16x16_c, &wrapper<aom_idct16x16_256_add_c>,
+             &wrapper<aom_idct16x16_1_add_msa>, TX_16X16, 1, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_64_add_msa>, TX_8X8, 64, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_12_add_msa>, TX_8X8, 12, 8, 1),
+  make_tuple(&aom_fdct8x8_c, &wrapper<aom_idct8x8_64_add_c>,
+             &wrapper<aom_idct8x8_1_add_msa>, TX_8X8, 1, 8, 1),
+  make_tuple(&aom_fdct4x4_c, &wrapper<aom_idct4x4_16_add_c>,
+             &wrapper<aom_idct4x4_16_add_msa>, TX_4X4, 16, 8, 1),
+  make_tuple(&aom_fdct4x4_c, &wrapper<aom_idct4x4_16_add_c>,
+             &wrapper<aom_idct4x4_1_add_msa>, TX_4X4, 1, 8, 1)
+};
+
+INSTANTIATE_TEST_CASE_P(MSA, PartialIDctTest,
+                        ::testing::ValuesIn(msa_partial_idct_tests));
+#endif  // HAVE_MSA && !CONFIG_HIGHBITDEPTH
+
+}  // namespace
diff --git a/third_party/aom/test/quantize_test.cc b/third_party/aom/test/quantize_test.cc
new file mode 100644
index 0000000000..4f61484a26
--- /dev/null
+++ b/third_party/aom/test/quantize_test.cc
@@ -0,0 +1,205 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "vp8/common/blockd.h"
+#include "vp8/common/onyx.h"
+#include "vp8/encoder/block.h"
+#include "vp8/encoder/onyx_int.h"
+#include "vp8/encoder/quantize.h"
+#include "aom/aom_integer.h"
+#include "aom_mem/aom_mem.h"
+
+namespace {
+#if !CONFIG_AOM_QM
+
+const int kNumBlocks = 25;
+const int kNumBlockEntries = 16;
+
+typedef void (*VP8Quantize)(BLOCK *b, BLOCKD *d);
+
+typedef std::tr1::tuple<VP8Quantize, VP8Quantize> VP8QuantizeParam;
+
+using libaom_test::ACMRandom;
+using std::tr1::make_tuple;
+
+// Create and populate a VP8_COMP instance which has a complete set of
+// quantization inputs as well as a second MACROBLOCKD for output.
+class QuantizeTestBase {
+ public:
+  virtual ~QuantizeTestBase() {
+    vp8_remove_compressor(&vp8_comp_);
+    vp8_comp_ = NULL;
+    aom_free(macroblockd_dst_);
+    macroblockd_dst_ = NULL;
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void SetupCompressor() {
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+
+    // The full configuration is necessary to generate the quantization tables.
+    VP8_CONFIG vp8_config;
+    memset(&vp8_config, 0, sizeof(vp8_config));
+
+    vp8_comp_ = vp8_create_compressor(&vp8_config);
+
+    // Set the tables based on a quantizer of 0.
+    vp8_set_quantizer(vp8_comp_, 0);
+
+    // Set up all the block/blockd pointers for the mb in vp8_comp_.
+    vp8cx_frame_init_quantizer(vp8_comp_);
+
+    // Copy macroblockd from the reference to get pre-set-up dequant values.
+    macroblockd_dst_ = reinterpret_cast<MACROBLOCKD *>(
+        aom_memalign(32, sizeof(*macroblockd_dst_)));
+    memcpy(macroblockd_dst_, &vp8_comp_->mb.e_mbd, sizeof(*macroblockd_dst_));
+    // Fix block pointers - currently they point to the blocks in the reference
+    // structure.
+    vp8_setup_block_dptrs(macroblockd_dst_);
+  }
+
+  void UpdateQuantizer(int q) {
+    vp8_set_quantizer(vp8_comp_, q);
+
+    memcpy(macroblockd_dst_, &vp8_comp_->mb.e_mbd, sizeof(*macroblockd_dst_));
+    vp8_setup_block_dptrs(macroblockd_dst_);
+  }
+
+  void FillCoeffConstant(int16_t c) {
+    for (int i = 0; i < kNumBlocks * kNumBlockEntries; ++i) {
+      vp8_comp_->mb.coeff[i] = c;
+    }
+  }
+
+  void FillCoeffRandom() {
+    for (int i = 0; i < kNumBlocks * kNumBlockEntries; ++i) {
+      vp8_comp_->mb.coeff[i] = rnd_.Rand8();
+    }
+  }
+
+  void CheckOutput() {
+    EXPECT_EQ(0, memcmp(vp8_comp_->mb.e_mbd.qcoeff, macroblockd_dst_->qcoeff,
+                        sizeof(*macroblockd_dst_->qcoeff) * kNumBlocks *
+                            kNumBlockEntries))
+        << "qcoeff mismatch";
+    EXPECT_EQ(0, memcmp(vp8_comp_->mb.e_mbd.dqcoeff, macroblockd_dst_->dqcoeff,
+                        sizeof(*macroblockd_dst_->dqcoeff) * kNumBlocks *
+                            kNumBlockEntries))
+        << "dqcoeff mismatch";
+    EXPECT_EQ(0, memcmp(vp8_comp_->mb.e_mbd.eobs, macroblockd_dst_->eobs,
+                        sizeof(*macroblockd_dst_->eobs) * kNumBlocks))
+        << "eobs mismatch";
+  }
+
+  VP8_COMP *vp8_comp_;
+  MACROBLOCKD *macroblockd_dst_;
+
+ private:
+  ACMRandom rnd_;
+};
+
+class QuantizeTest : public QuantizeTestBase,
+                     public ::testing::TestWithParam<VP8QuantizeParam> {
+ protected:
+  virtual void SetUp() {
+    SetupCompressor();
+    asm_quant_ = GET_PARAM(0);
+    c_quant_ = GET_PARAM(1);
+  }
+
+  void RunComparison() {
+    for (int i = 0; i < kNumBlocks; ++i) {
+      ASM_REGISTER_STATE_CHECK(
+          c_quant_(&vp8_comp_->mb.block[i], &vp8_comp_->mb.e_mbd.block[i]));
+      ASM_REGISTER_STATE_CHECK(
+          asm_quant_(&vp8_comp_->mb.block[i], &macroblockd_dst_->block[i]));
+    }
+
+    CheckOutput();
+  }
+
+ private:
+  VP8Quantize asm_quant_;
+  VP8Quantize c_quant_;
+};
+
+TEST_P(QuantizeTest, TestZeroInput) {
+  FillCoeffConstant(0);
+  RunComparison();
+}
+
+TEST_P(QuantizeTest, TestLargeNegativeInput) {
+  FillCoeffConstant(0);
+  // Generate a qcoeff which contains 512/-512 (0x0100/0xFE00) to catch issues
+  // like BUG=883 where the constant being compared was incorrectly initialized.
+  vp8_comp_->mb.coeff[0] = -8191;
+  RunComparison();
+}
+
+TEST_P(QuantizeTest, TestRandomInput) {
+  FillCoeffRandom();
+  RunComparison();
+}
+
+TEST_P(QuantizeTest, TestMultipleQ) {
+  for (int q = 0; q < QINDEX_RANGE; ++q) {
+    UpdateQuantizer(q);
+    FillCoeffRandom();
+    RunComparison();
+  }
+}
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, QuantizeTest,
+    ::testing::Values(
+        make_tuple(&vp8_fast_quantize_b_sse2, &vp8_fast_quantize_b_c),
+        make_tuple(&vp8_regular_quantize_b_sse2, &vp8_regular_quantize_b_c)));
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(SSSE3, QuantizeTest,
+                        ::testing::Values(make_tuple(&vp8_fast_quantize_b_ssse3,
+                                                     &vp8_fast_quantize_b_c)));
+#endif  // HAVE_SSSE3
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, QuantizeTest,
+    ::testing::Values(make_tuple(&vp8_regular_quantize_b_sse4_1,
+                                 &vp8_regular_quantize_b_c)));
+#endif  // HAVE_SSE4_1
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, QuantizeTest,
+                        ::testing::Values(make_tuple(&vp8_fast_quantize_b_neon,
+                                                     &vp8_fast_quantize_b_c)));
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(
+    MSA, QuantizeTest,
+    ::testing::Values(
+        make_tuple(&vp8_fast_quantize_b_msa, &vp8_fast_quantize_b_c),
+        make_tuple(&vp8_regular_quantize_b_msa, &vp8_regular_quantize_b_c)));
+#endif  // HAVE_MSA
+#endif  // CONFIG_AOM_QM
+}  // namespace
diff --git a/third_party/aom/test/realtime_test.cc b/third_party/aom/test/realtime_test.cc
new file mode 100644
index 0000000000..ffe4a3146b
--- /dev/null
+++ b/third_party/aom/test/realtime_test.cc
@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/util.h"
+#include "test/video_source.h"
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+namespace {
+
+const int kVideoSourceWidth = 320;
+const int kVideoSourceHeight = 240;
+const int kFramesToEncode = 2;
+
+class RealtimeTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWithParam<libaom_test::TestMode> {
+ protected:
+  RealtimeTest() : EncoderTest(GET_PARAM(0)), frame_packets_(0) {}
+  virtual ~RealtimeTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    cfg_.g_lag_in_frames = 0;
+    SetMode(::libaom_test::kRealTime);
+  }
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+    // TODO(tomfinegan): We're changing the pass value here to make sure
+    // we get frames when real time mode is combined with |g_pass| set to
+    // AOM_RC_FIRST_PASS. This is necessary because EncoderTest::RunLoop() sets
+    // the pass value based on the mode passed into EncoderTest::SetMode(),
+    // which overrides the one specified in SetUp() above.
+    cfg_.g_pass = AOM_RC_FIRST_PASS;
+  }
+  virtual void FramePktHook(const aom_codec_cx_pkt_t * /*pkt*/) {
+    frame_packets_++;
+  }
+
+  int frame_packets_;
+};
+
+TEST_P(RealtimeTest, RealtimeFirstPassProducesFrames) {
+  ::libaom_test::RandomVideoSource video;
+  video.SetSize(kVideoSourceWidth, kVideoSourceHeight);
+  video.set_limit(kFramesToEncode);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+  EXPECT_EQ(kFramesToEncode, frame_packets_);
+}
+
+AV1_INSTANTIATE_TEST_CASE(RealtimeTest,
+                          ::testing::Values(::libaom_test::kRealTime));
+
+}  // namespace
diff --git a/third_party/aom/test/register_state_check.h b/third_party/aom/test/register_state_check.h
new file mode 100644
index 0000000000..3308201739
--- /dev/null
+++ b/third_party/aom/test/register_state_check.h
@@ -0,0 +1,196 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef TEST_REGISTER_STATE_CHECK_H_
+#define TEST_REGISTER_STATE_CHECK_H_
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "./aom_config.h"
+#include "aom/aom_integer.h"
+
+// ASM_REGISTER_STATE_CHECK(asm_function)
+//   Minimally validates the environment pre & post function execution. This
+//   variant should be used with assembly functions which are not expected to
+//   fully restore the system state. See platform implementations of
+//   RegisterStateCheck for details.
+//
+// API_REGISTER_STATE_CHECK(api_function)
+//   Performs all the checks done by ASM_REGISTER_STATE_CHECK() and any
+//   additional checks to ensure the environment is in a consistent state pre &
+//   post function execution. This variant should be used with API functions.
+//   See platform implementations of RegisterStateCheckXXX for details.
+//
+
+#if defined(_WIN64)
+
+#undef NOMINMAX
+#define NOMINMAX
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#include <winnt.h>
+
+inline bool operator==(const M128A &lhs, const M128A &rhs) {
+  return (lhs.Low == rhs.Low && lhs.High == rhs.High);
+}
+
+namespace libaom_test {
+
+// Compares the state of xmm[6-15] at construction with their state at
+// destruction. These registers should be preserved by the callee on
+// Windows x64.
+class RegisterStateCheck {
+ public:
+  RegisterStateCheck() { initialized_ = StoreRegisters(&pre_context_); }
+  ~RegisterStateCheck() { EXPECT_TRUE(Check()); }
+
+ private:
+  static bool StoreRegisters(CONTEXT *const context) {
+    const HANDLE this_thread = GetCurrentThread();
+    EXPECT_TRUE(this_thread != NULL);
+    context->ContextFlags = CONTEXT_FLOATING_POINT;
+    const bool context_saved = GetThreadContext(this_thread, context) == TRUE;
+    EXPECT_TRUE(context_saved) << "GetLastError: " << GetLastError();
+    return context_saved;
+  }
+
+  // Compares the register state. Returns true if the states match.
+  bool Check() const {
+    if (!initialized_) return false;
+    CONTEXT post_context;
+    if (!StoreRegisters(&post_context)) return false;
+
+    const M128A *xmm_pre = &pre_context_.Xmm6;
+    const M128A *xmm_post = &post_context.Xmm6;
+    for (int i = 6; i <= 15; ++i) {
+      EXPECT_EQ(*xmm_pre, *xmm_post) << "xmm" << i << " has been modified!";
+      ++xmm_pre;
+      ++xmm_post;
+    }
+    return !testing::Test::HasNonfatalFailure();
+  }
+
+  bool initialized_;
+  CONTEXT pre_context_;
+};
+
+#define ASM_REGISTER_STATE_CHECK(statement)    \
+  do {                                         \
+    libaom_test::RegisterStateCheck reg_check; \
+    statement;                                 \
+  } while (false)
+
+}  // namespace libaom_test
+
+#elif defined(CONFIG_SHARED) && defined(HAVE_NEON_ASM) && !CONFIG_SHARED && \
+    HAVE_NEON_ASM && CONFIG_AV1
+
+extern "C" {
+// Save the d8-d15 registers into store.
+void aom_push_neon(int64_t *store);
+}
+
+namespace libaom_test {
+
+// Compares the state of d8-d15 at construction with their state at
+// destruction. These registers should be preserved by the callee on
+// arm platform.
+class RegisterStateCheck {
+ public:
+  RegisterStateCheck() { initialized_ = StoreRegisters(pre_store_); }
+  ~RegisterStateCheck() { EXPECT_TRUE(Check()); }
+
+ private:
+  static bool StoreRegisters(int64_t store[8]) {
+    aom_push_neon(store);
+    return true;
+  }
+
+  // Compares the register state. Returns true if the states match.
+  bool Check() const {
+    if (!initialized_) return false;
+    int64_t post_store[8];
+    aom_push_neon(post_store);
+    for (int i = 0; i < 8; ++i) {
+      EXPECT_EQ(pre_store_[i], post_store[i]) << "d" << i + 8
+                                              << " has been modified";
+    }
+    return !testing::Test::HasNonfatalFailure();
+  }
+
+  bool initialized_;
+  int64_t pre_store_[8];
+};
+
+#define ASM_REGISTER_STATE_CHECK(statement)    \
+  do {                                         \
+    libaom_test::RegisterStateCheck reg_check; \
+    statement;                                 \
+  } while (false)
+
+}  // namespace libaom_test
+
+#else
+
+namespace libaom_test {
+
+class RegisterStateCheck {};
+#define ASM_REGISTER_STATE_CHECK(statement) statement
+
+}  // namespace libaom_test
+
+#endif  // _WIN64
+
+#if ARCH_X86 || ARCH_X86_64
+#if defined(__GNUC__)
+
+namespace libaom_test {
+
+// Checks the FPU tag word pre/post execution to ensure emms has been called.
+class RegisterStateCheckMMX {
+ public:
+  RegisterStateCheckMMX() {
+    __asm__ volatile("fstenv %0" : "=rm"(pre_fpu_env_));
+  }
+  ~RegisterStateCheckMMX() { EXPECT_TRUE(Check()); }
+
+ private:
+  // Checks the FPU tag word pre/post execution, returning false if not cleared
+  // to 0xffff.
+  bool Check() const {
+    EXPECT_EQ(0xffff, pre_fpu_env_[4])
+        << "FPU was in an inconsistent state prior to call";
+
+    uint16_t post_fpu_env[14];
+    __asm__ volatile("fstenv %0" : "=rm"(post_fpu_env));
+    EXPECT_EQ(0xffff, post_fpu_env[4])
+        << "FPU was left in an inconsistent state after call";
+    return !testing::Test::HasNonfatalFailure();
+  }
+
+  uint16_t pre_fpu_env_[14];
+};
+
+#define API_REGISTER_STATE_CHECK(statement)       \
+  do {                                            \
+    libaom_test::RegisterStateCheckMMX reg_check; \
+    ASM_REGISTER_STATE_CHECK(statement);          \
+  } while (false)
+
+}  // namespace libaom_test
+
+#endif  // __GNUC__
+#endif  // ARCH_X86 || ARCH_X86_64
+
+#ifndef API_REGISTER_STATE_CHECK
+#define API_REGISTER_STATE_CHECK ASM_REGISTER_STATE_CHECK
+#endif
+
+#endif  // TEST_REGISTER_STATE_CHECK_H_
diff --git a/third_party/aom/test/resize_test.cc b/third_party/aom/test/resize_test.cc
new file mode 100644
index 0000000000..994b30117b
--- /dev/null
+++ b/third_party/aom/test/resize_test.cc
@@ -0,0 +1,717 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <climits>
+#include <vector>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/video_source.h"
+#include "test/util.h"
+
+// Enable(1) or Disable(0) writing of the compressed bitstream.
+#define WRITE_COMPRESSED_STREAM 0
+
+namespace {
+
+#if WRITE_COMPRESSED_STREAM
+static void mem_put_le16(char *const mem, unsigned int val) {
+  mem[0] = val;
+  mem[1] = val >> 8;
+}
+
+static void mem_put_le32(char *const mem, unsigned int val) {
+  mem[0] = val;
+  mem[1] = val >> 8;
+  mem[2] = val >> 16;
+  mem[3] = val >> 24;
+}
+
+static void write_ivf_file_header(const aom_codec_enc_cfg_t *const cfg,
+                                  int frame_cnt, FILE *const outfile) {
+  char header[32];
+
+  header[0] = 'D';
+  header[1] = 'K';
+  header[2] = 'I';
+  header[3] = 'F';
+  mem_put_le16(header + 4, 0);                    /* version */
+  mem_put_le16(header + 6, 32);                   /* headersize */
+  mem_put_le32(header + 8, 0x30395056);           /* fourcc (av1) */
+  mem_put_le16(header + 12, cfg->g_w);            /* width */
+  mem_put_le16(header + 14, cfg->g_h);            /* height */
+  mem_put_le32(header + 16, cfg->g_timebase.den); /* rate */
+  mem_put_le32(header + 20, cfg->g_timebase.num); /* scale */
+  mem_put_le32(header + 24, frame_cnt);           /* length */
+  mem_put_le32(header + 28, 0);                   /* unused */
+
+  (void)fwrite(header, 1, 32, outfile);
+}
+
+static void write_ivf_frame_size(FILE *const outfile, const size_t size) {
+  char header[4];
+  mem_put_le32(header, static_cast<unsigned int>(size));
+  (void)fwrite(header, 1, 4, outfile);
+}
+
+static void write_ivf_frame_header(const aom_codec_cx_pkt_t *const pkt,
+                                   FILE *const outfile) {
+  char header[12];
+  aom_codec_pts_t pts;
+
+  if (pkt->kind != AOM_CODEC_CX_FRAME_PKT) return;
+
+  pts = pkt->data.frame.pts;
+  mem_put_le32(header, static_cast<unsigned int>(pkt->data.frame.sz));
+  mem_put_le32(header + 4, pts & 0xFFFFFFFF);
+  mem_put_le32(header + 8, pts >> 32);
+
+  (void)fwrite(header, 1, 12, outfile);
+}
+#endif  // WRITE_COMPRESSED_STREAM
+
+const unsigned int kInitialWidth = 320;
+const unsigned int kInitialHeight = 240;
+
+struct FrameInfo {
+  FrameInfo(aom_codec_pts_t _pts, unsigned int _w, unsigned int _h)
+      : pts(_pts), w(_w), h(_h) {}
+
+  aom_codec_pts_t pts;
+  unsigned int w;
+  unsigned int h;
+};
+
+void ScaleForFrameNumber(unsigned int frame, unsigned int initial_w,
+                         unsigned int initial_h, unsigned int *w,
+                         unsigned int *h, int flag_codec) {
+  if (frame < 10) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 20) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 30) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 40) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 50) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 60) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 70) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 80) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 90) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 100) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 110) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 120) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 130) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 140) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 150) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 160) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 170) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 180) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 190) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 200) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 210) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 220) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 230) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  if (frame < 240) {
+    *w = initial_w * 3 / 4;
+    *h = initial_h * 3 / 4;
+    return;
+  }
+  if (frame < 250) {
+    *w = initial_w / 2;
+    *h = initial_h / 2;
+    return;
+  }
+  if (frame < 260) {
+    *w = initial_w;
+    *h = initial_h;
+    return;
+  }
+  // Go down very low.
+  if (frame < 270) {
+    *w = initial_w / 4;
+    *h = initial_h / 4;
+    return;
+  }
+  if (flag_codec == 1) {
+    // Cases that only works for AV1.
+    // For AV1: Swap width and height of original.
+    if (frame < 320) {
+      *w = initial_h;
+      *h = initial_w;
+      return;
+    }
+  }
+  *w = initial_w;
+  *h = initial_h;
+}
+
+class ResizingVideoSource : public ::libaom_test::DummyVideoSource {
+ public:
+  ResizingVideoSource() {
+    SetSize(kInitialWidth, kInitialHeight);
+    limit_ = 350;
+  }
+  int flag_codec_;
+  virtual ~ResizingVideoSource() {}
+
+ protected:
+  virtual void Next() {
+    ++frame_;
+    unsigned int width;
+    unsigned int height;
+    ScaleForFrameNumber(frame_, kInitialWidth, kInitialHeight, &width, &height,
+                        flag_codec_);
+    SetSize(width, height);
+    FillFrame();
+  }
+};
+
+class ResizeTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWithParam<libaom_test::TestMode> {
+ protected:
+  ResizeTest() : EncoderTest(GET_PARAM(0)) {}
+
+  virtual ~ResizeTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+  }
+
+  virtual void DecompressedFrameHook(const aom_image_t &img,
+                                     aom_codec_pts_t pts) {
+    frame_info_list_.push_back(FrameInfo(pts, img.d_w, img.d_h));
+  }
+
+  std::vector<FrameInfo> frame_info_list_;
+};
+
+TEST_P(ResizeTest, TestExternalResizeWorks) {
+  ResizingVideoSource video;
+  video.flag_codec_ = 0;
+  cfg_.g_lag_in_frames = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  for (std::vector<FrameInfo>::const_iterator info = frame_info_list_.begin();
+       info != frame_info_list_.end(); ++info) {
+    const unsigned int frame = static_cast<unsigned>(info->pts);
+    unsigned int expected_w;
+    unsigned int expected_h;
+    ScaleForFrameNumber(frame, kInitialWidth, kInitialHeight, &expected_w,
+                        &expected_h, 0);
+    EXPECT_EQ(expected_w, info->w) << "Frame " << frame
+                                   << " had unexpected width";
+    EXPECT_EQ(expected_h, info->h) << "Frame " << frame
+                                   << " had unexpected height";
+  }
+}
+
+const unsigned int kStepDownFrame = 3;
+const unsigned int kStepUpFrame = 6;
+
+class ResizeInternalTest : public ResizeTest {
+ protected:
+#if WRITE_COMPRESSED_STREAM
+  ResizeInternalTest()
+      : ResizeTest(), frame0_psnr_(0.0), outfile_(NULL), out_frames_(0) {}
+#else
+  ResizeInternalTest() : ResizeTest(), frame0_psnr_(0.0) {}
+#endif
+
+  virtual ~ResizeInternalTest() {}
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+#if WRITE_COMPRESSED_STREAM
+    outfile_ = fopen("av10-2-05-resize.ivf", "wb");
+#endif
+  }
+
+  virtual void EndPassHook() {
+#if WRITE_COMPRESSED_STREAM
+    if (outfile_) {
+      if (!fseek(outfile_, 0, SEEK_SET))
+        write_ivf_file_header(&cfg_, out_frames_, outfile_);
+      fclose(outfile_);
+      outfile_ = NULL;
+    }
+#endif
+  }
+
+  virtual void PreEncodeFrameHook(libaom_test::VideoSource *video,
+                                  libaom_test::Encoder *encoder) {
+    if (change_config_) {
+      int new_q = 60;
+      if (video->frame() == 0) {
+        struct aom_scaling_mode mode = { AOME_ONETWO, AOME_ONETWO };
+        encoder->Control(AOME_SET_SCALEMODE, &mode);
+      }
+      if (video->frame() == 1) {
+        struct aom_scaling_mode mode = { AOME_NORMAL, AOME_NORMAL };
+        encoder->Control(AOME_SET_SCALEMODE, &mode);
+        cfg_.rc_min_quantizer = cfg_.rc_max_quantizer = new_q;
+        encoder->Config(&cfg_);
+      }
+    } else {
+      if (video->frame() == kStepDownFrame) {
+        struct aom_scaling_mode mode = { AOME_FOURFIVE, AOME_THREEFIVE };
+        encoder->Control(AOME_SET_SCALEMODE, &mode);
+      }
+      if (video->frame() == kStepUpFrame) {
+        struct aom_scaling_mode mode = { AOME_NORMAL, AOME_NORMAL };
+        encoder->Control(AOME_SET_SCALEMODE, &mode);
+      }
+    }
+  }
+
+  virtual void PSNRPktHook(const aom_codec_cx_pkt_t *pkt) {
+    if (frame0_psnr_ == 0.) frame0_psnr_ = pkt->data.psnr.psnr[0];
+    EXPECT_NEAR(pkt->data.psnr.psnr[0], frame0_psnr_, 2.0);
+  }
+
+#if WRITE_COMPRESSED_STREAM
+  virtual void FramePktHook(const aom_codec_cx_pkt_t *pkt) {
+    ++out_frames_;
+
+    // Write initial file header if first frame.
+    if (pkt->data.frame.pts == 0) write_ivf_file_header(&cfg_, 0, outfile_);
+
+    // Write frame header and data.
+    write_ivf_frame_header(pkt, outfile_);
+    (void)fwrite(pkt->data.frame.buf, 1, pkt->data.frame.sz, outfile_);
+  }
+#endif
+
+  double frame0_psnr_;
+  bool change_config_;
+#if WRITE_COMPRESSED_STREAM
+  FILE *outfile_;
+  unsigned int out_frames_;
+#endif
+};
+
+TEST_P(ResizeInternalTest, TestInternalResizeWorks) {
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 10);
+  init_flags_ = AOM_CODEC_USE_PSNR;
+  change_config_ = false;
+
+  // q picked such that initial keyframe on this clip is ~30dB PSNR
+  cfg_.rc_min_quantizer = cfg_.rc_max_quantizer = 48;
+
+  // If the number of frames being encoded is smaller than g_lag_in_frames
+  // the encoded frame is unavailable using the current API. Comparing
+  // frames to detect mismatch would then not be possible. Set
+  // g_lag_in_frames = 0 to get around this.
+  cfg_.g_lag_in_frames = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  for (std::vector<FrameInfo>::const_iterator info = frame_info_list_.begin();
+       info != frame_info_list_.end(); ++info) {
+    const aom_codec_pts_t pts = info->pts;
+    if (pts >= kStepDownFrame && pts < kStepUpFrame) {
+      ASSERT_EQ(282U, info->w) << "Frame " << pts << " had unexpected width";
+      ASSERT_EQ(173U, info->h) << "Frame " << pts << " had unexpected height";
+    } else {
+      EXPECT_EQ(352U, info->w) << "Frame " << pts << " had unexpected width";
+      EXPECT_EQ(288U, info->h) << "Frame " << pts << " had unexpected height";
+    }
+  }
+}
+
+TEST_P(ResizeInternalTest, TestInternalResizeChangeConfig) {
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 10);
+  cfg_.g_w = 352;
+  cfg_.g_h = 288;
+  change_config_ = true;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+class ResizeRealtimeTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<libaom_test::TestMode, int> {
+ protected:
+  ResizeRealtimeTest() : EncoderTest(GET_PARAM(0)) {}
+  virtual ~ResizeRealtimeTest() {}
+
+  virtual void PreEncodeFrameHook(libaom_test::VideoSource *video,
+                                  libaom_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      encoder->Control(AV1E_SET_AQ_MODE, 3);
+      encoder->Control(AOME_SET_CPUUSED, set_cpu_used_);
+    }
+
+    if (change_bitrate_ && video->frame() == 120) {
+      change_bitrate_ = false;
+      cfg_.rc_target_bitrate = 500;
+      encoder->Config(&cfg_);
+    }
+  }
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(GET_PARAM(1));
+    set_cpu_used_ = GET_PARAM(2);
+  }
+
+  virtual void DecompressedFrameHook(const aom_image_t &img,
+                                     aom_codec_pts_t pts) {
+    frame_info_list_.push_back(FrameInfo(pts, img.d_w, img.d_h));
+  }
+
+  virtual void MismatchHook(const aom_image_t *img1, const aom_image_t *img2) {
+    double mismatch_psnr = compute_psnr(img1, img2);
+    mismatch_psnr_ += mismatch_psnr;
+    ++mismatch_nframes_;
+  }
+
+  unsigned int GetMismatchFrames() { return mismatch_nframes_; }
+
+  void DefaultConfig() {
+    cfg_.rc_buf_initial_sz = 500;
+    cfg_.rc_buf_optimal_sz = 600;
+    cfg_.rc_buf_sz = 1000;
+    cfg_.rc_min_quantizer = 2;
+    cfg_.rc_max_quantizer = 56;
+    cfg_.rc_undershoot_pct = 50;
+    cfg_.rc_overshoot_pct = 50;
+    cfg_.rc_end_usage = AOM_CBR;
+    cfg_.kf_mode = AOM_KF_AUTO;
+    cfg_.g_lag_in_frames = 0;
+    cfg_.kf_min_dist = cfg_.kf_max_dist = 3000;
+    // Enable dropped frames.
+    cfg_.rc_dropframe_thresh = 1;
+    // Enable error_resilience mode.
+    cfg_.g_error_resilient = 1;
+    // Enable dynamic resizing.
+    cfg_.rc_resize_allowed = 1;
+    // Run at low bitrate.
+    cfg_.rc_target_bitrate = 200;
+  }
+
+  std::vector<FrameInfo> frame_info_list_;
+  int set_cpu_used_;
+  bool change_bitrate_;
+  double mismatch_psnr_;
+  int mismatch_nframes_;
+};
+
+TEST_P(ResizeRealtimeTest, TestExternalResizeWorks) {
+  ResizingVideoSource video;
+  video.flag_codec_ = 1;
+  DefaultConfig();
+  // Disable internal resize for this test.
+  cfg_.rc_resize_allowed = 0;
+  change_bitrate_ = false;
+  mismatch_psnr_ = 0.0;
+  mismatch_nframes_ = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  for (std::vector<FrameInfo>::const_iterator info = frame_info_list_.begin();
+       info != frame_info_list_.end(); ++info) {
+    const unsigned int frame = static_cast<unsigned>(info->pts);
+    unsigned int expected_w;
+    unsigned int expected_h;
+    ScaleForFrameNumber(frame, kInitialWidth, kInitialHeight, &expected_w,
+                        &expected_h, 1);
+    EXPECT_EQ(expected_w, info->w) << "Frame " << frame
+                                   << " had unexpected width";
+    EXPECT_EQ(expected_h, info->h) << "Frame " << frame
+                                   << " had unexpected height";
+    EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+  }
+}
+
+// Verify the dynamic resizer behavior for real time, 1 pass CBR mode.
+// Run at low bitrate, with resize_allowed = 1, and verify that we get
+// one resize down event.
+TEST_P(ResizeRealtimeTest, TestInternalResizeDown) {
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 299);
+  DefaultConfig();
+  cfg_.g_w = 352;
+  cfg_.g_h = 288;
+  change_bitrate_ = false;
+  mismatch_psnr_ = 0.0;
+  mismatch_nframes_ = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  unsigned int last_w = cfg_.g_w;
+  unsigned int last_h = cfg_.g_h;
+  int resize_count = 0;
+  for (std::vector<FrameInfo>::const_iterator info = frame_info_list_.begin();
+       info != frame_info_list_.end(); ++info) {
+    if (info->w != last_w || info->h != last_h) {
+      // Verify that resize down occurs.
+      ASSERT_LT(info->w, last_w);
+      ASSERT_LT(info->h, last_h);
+      last_w = info->w;
+      last_h = info->h;
+      resize_count++;
+    }
+  }
+
+#if CONFIG_AV1_DECODER
+  // Verify that we get 1 resize down event in this test.
+  ASSERT_EQ(1, resize_count) << "Resizing should occur.";
+  EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+#else
+  printf("Warning: AV1 decoder unavailable, unable to check resize count!\n");
+#endif
+}
+
+// Verify the dynamic resizer behavior for real time, 1 pass CBR mode.
+// Start at low target bitrate, raise the bitrate in the middle of the clip,
+// scaling-up should occur after bitrate changed.
+TEST_P(ResizeRealtimeTest, TestInternalResizeDownUpChangeBitRate) {
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 359);
+  DefaultConfig();
+  cfg_.g_w = 352;
+  cfg_.g_h = 288;
+  change_bitrate_ = true;
+  mismatch_psnr_ = 0.0;
+  mismatch_nframes_ = 0;
+  // Disable dropped frames.
+  cfg_.rc_dropframe_thresh = 0;
+  // Starting bitrate low.
+  cfg_.rc_target_bitrate = 80;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+  unsigned int last_w = cfg_.g_w;
+  unsigned int last_h = cfg_.g_h;
+  int resize_count = 0;
+  for (std::vector<FrameInfo>::const_iterator info = frame_info_list_.begin();
+       info != frame_info_list_.end(); ++info) {
+    if (info->w != last_w || info->h != last_h) {
+      resize_count++;
+      if (resize_count == 1) {
+        // Verify that resize down occurs.
+        ASSERT_LT(info->w, last_w);
+        ASSERT_LT(info->h, last_h);
+      } else if (resize_count == 2) {
+        // Verify that resize up occurs.
+        ASSERT_GT(info->w, last_w);
+        ASSERT_GT(info->h, last_h);
+      }
+      last_w = info->w;
+      last_h = info->h;
+    }
+  }
+
+#if CONFIG_AV1_DECODER
+  // Verify that we get 2 resize events in this test.
+  ASSERT_EQ(resize_count, 2) << "Resizing should occur twice.";
+  EXPECT_EQ(static_cast<unsigned int>(0), GetMismatchFrames());
+#else
+  printf("Warning: AV1 decoder unavailable, unable to check resize count!\n");
+#endif
+}
+
+aom_img_fmt_t CspForFrameNumber(int frame) {
+  if (frame < 10) return AOM_IMG_FMT_I420;
+  if (frame < 20) return AOM_IMG_FMT_I444;
+  return AOM_IMG_FMT_I420;
+}
+
+class ResizeCspTest : public ResizeTest {
+ protected:
+#if WRITE_COMPRESSED_STREAM
+  ResizeCspTest()
+      : ResizeTest(), frame0_psnr_(0.0), outfile_(NULL), out_frames_(0) {}
+#else
+  ResizeCspTest() : ResizeTest(), frame0_psnr_(0.0) {}
+#endif
+
+  virtual ~ResizeCspTest() {}
+
+  virtual void BeginPassHook(unsigned int /*pass*/) {
+#if WRITE_COMPRESSED_STREAM
+    outfile_ = fopen("av11-2-05-cspchape.ivf", "wb");
+#endif
+  }
+
+  virtual void EndPassHook() {
+#if WRITE_COMPRESSED_STREAM
+    if (outfile_) {
+      if (!fseek(outfile_, 0, SEEK_SET))
+        write_ivf_file_header(&cfg_, out_frames_, outfile_);
+      fclose(outfile_);
+      outfile_ = NULL;
+    }
+#endif
+  }
+
+  virtual void PreEncodeFrameHook(libaom_test::VideoSource *video,
+                                  libaom_test::Encoder *encoder) {
+    if (CspForFrameNumber(video->frame()) != AOM_IMG_FMT_I420 &&
+        cfg_.g_profile != 1) {
+      cfg_.g_profile = 1;
+      encoder->Config(&cfg_);
+    }
+    if (CspForFrameNumber(video->frame()) == AOM_IMG_FMT_I420 &&
+        cfg_.g_profile != 0) {
+      cfg_.g_profile = 0;
+      encoder->Config(&cfg_);
+    }
+  }
+
+  virtual void PSNRPktHook(const aom_codec_cx_pkt_t *pkt) {
+    if (frame0_psnr_ == 0.) frame0_psnr_ = pkt->data.psnr.psnr[0];
+    EXPECT_NEAR(pkt->data.psnr.psnr[0], frame0_psnr_, 2.0);
+  }
+
+#if WRITE_COMPRESSED_STREAM
+  virtual void FramePktHook(const aom_codec_cx_pkt_t *pkt) {
+    ++out_frames_;
+
+    // Write initial file header if first frame.
+    if (pkt->data.frame.pts == 0) write_ivf_file_header(&cfg_, 0, outfile_);
+
+    // Write frame header and data.
+    write_ivf_frame_header(pkt, outfile_);
+    (void)fwrite(pkt->data.frame.buf, 1, pkt->data.frame.sz, outfile_);
+  }
+#endif
+
+  double frame0_psnr_;
+#if WRITE_COMPRESSED_STREAM
+  FILE *outfile_;
+  unsigned int out_frames_;
+#endif
+};
+
+class ResizingCspVideoSource : public ::libaom_test::DummyVideoSource {
+ public:
+  ResizingCspVideoSource() {
+    SetSize(kInitialWidth, kInitialHeight);
+    limit_ = 30;
+  }
+
+  virtual ~ResizingCspVideoSource() {}
+
+ protected:
+  virtual void Next() {
+    ++frame_;
+    SetImageFormat(CspForFrameNumber(frame_));
+    FillFrame();
+  }
+};
+
+TEST_P(ResizeCspTest, TestResizeCspWorks) {
+  ResizingCspVideoSource video;
+  init_flags_ = AOM_CODEC_USE_PSNR;
+  cfg_.rc_min_quantizer = cfg_.rc_max_quantizer = 48;
+  cfg_.g_lag_in_frames = 0;
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+}
+
+AV1_INSTANTIATE_TEST_CASE(ResizeTest,
+                          ::testing::Values(::libaom_test::kRealTime));
+AV1_INSTANTIATE_TEST_CASE(ResizeInternalTest,
+                          ::testing::Values(::libaom_test::kOnePassBest));
+AV1_INSTANTIATE_TEST_CASE(ResizeRealtimeTest,
+                          ::testing::Values(::libaom_test::kRealTime),
+                          ::testing::Range(5, 9));
+AV1_INSTANTIATE_TEST_CASE(ResizeCspTest,
+                          ::testing::Values(::libaom_test::kRealTime));
+}  // namespace
diff --git a/third_party/aom/test/sad_test.cc b/third_party/aom/test/sad_test.cc
new file mode 100644
index 0000000000..c3b5dac422
--- /dev/null
+++ b/third_party/aom/test/sad_test.cc
@@ -0,0 +1,1172 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <string.h>
+#include <limits.h>
+#include <stdio.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "aom/aom_codec.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+typedef unsigned int (*SadMxNFunc)(const uint8_t *src_ptr, int src_stride,
+                                   const uint8_t *ref_ptr, int ref_stride);
+typedef std::tr1::tuple<int, int, SadMxNFunc, int> SadMxNParam;
+
+typedef uint32_t (*SadMxNAvgFunc)(const uint8_t *src_ptr, int src_stride,
+                                  const uint8_t *ref_ptr, int ref_stride,
+                                  const uint8_t *second_pred);
+typedef std::tr1::tuple<int, int, SadMxNAvgFunc, int> SadMxNAvgParam;
+
+typedef void (*SadMxNx4Func)(const uint8_t *src_ptr, int src_stride,
+                             const uint8_t *const ref_ptr[], int ref_stride,
+                             uint32_t *sad_array);
+typedef std::tr1::tuple<int, int, SadMxNx4Func, int> SadMxNx4Param;
+
+using libaom_test::ACMRandom;
+
+namespace {
+class SADTestBase : public ::testing::Test {
+ public:
+  SADTestBase(int width, int height, int bit_depth)
+      : width_(width), height_(height), bd_(bit_depth) {}
+
+  static void SetUpTestCase() {
+    source_data8_ = reinterpret_cast<uint8_t *>(
+        aom_memalign(kDataAlignment, kDataBlockSize));
+    reference_data8_ = reinterpret_cast<uint8_t *>(
+        aom_memalign(kDataAlignment, kDataBufferSize));
+    second_pred8_ =
+        reinterpret_cast<uint8_t *>(aom_memalign(kDataAlignment, 128 * 128));
+    source_data16_ = reinterpret_cast<uint16_t *>(
+        aom_memalign(kDataAlignment, kDataBlockSize * sizeof(uint16_t)));
+    reference_data16_ = reinterpret_cast<uint16_t *>(
+        aom_memalign(kDataAlignment, kDataBufferSize * sizeof(uint16_t)));
+    second_pred16_ = reinterpret_cast<uint16_t *>(
+        aom_memalign(kDataAlignment, 128 * 128 * sizeof(uint16_t)));
+  }
+
+  static void TearDownTestCase() {
+    aom_free(source_data8_);
+    source_data8_ = NULL;
+    aom_free(reference_data8_);
+    reference_data8_ = NULL;
+    aom_free(second_pred8_);
+    second_pred8_ = NULL;
+    aom_free(source_data16_);
+    source_data16_ = NULL;
+    aom_free(reference_data16_);
+    reference_data16_ = NULL;
+    aom_free(second_pred16_);
+    second_pred16_ = NULL;
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  // Handle up to 4 128x128 blocks, with stride up to 256
+  static const int kDataAlignment = 16;
+  static const int kDataBlockSize = 128 * 256;
+  static const int kDataBufferSize = 4 * kDataBlockSize;
+
+  virtual void SetUp() {
+    if (bd_ == -1) {
+      use_high_bit_depth_ = false;
+      bit_depth_ = AOM_BITS_8;
+      source_data_ = source_data8_;
+      reference_data_ = reference_data8_;
+      second_pred_ = second_pred8_;
+#if CONFIG_HIGHBITDEPTH
+    } else {
+      use_high_bit_depth_ = true;
+      bit_depth_ = static_cast<aom_bit_depth_t>(bd_);
+      source_data_ = CONVERT_TO_BYTEPTR(source_data16_);
+      reference_data_ = CONVERT_TO_BYTEPTR(reference_data16_);
+      second_pred_ = CONVERT_TO_BYTEPTR(second_pred16_);
+#endif  // CONFIG_HIGHBITDEPTH
+    }
+    mask_ = (1 << bit_depth_) - 1;
+    source_stride_ = (width_ + 31) & ~31;
+    reference_stride_ = width_ * 2;
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+  }
+
+  virtual uint8_t *GetReference(int block_idx) {
+#if CONFIG_HIGHBITDEPTH
+    if (use_high_bit_depth_)
+      return CONVERT_TO_BYTEPTR(CONVERT_TO_SHORTPTR(reference_data_) +
+                                block_idx * kDataBlockSize);
+#endif  // CONFIG_HIGHBITDEPTH
+    return reference_data_ + block_idx * kDataBlockSize;
+  }
+
+  // Sum of Absolute Differences. Given two blocks, calculate the absolute
+  // difference between two pixels in the same relative location; accumulate.
+  unsigned int ReferenceSAD(int block_idx) {
+    unsigned int sad = 0;
+    const uint8_t *const reference8 = GetReference(block_idx);
+    const uint8_t *const source8 = source_data_;
+#if CONFIG_HIGHBITDEPTH
+    const uint16_t *const reference16 =
+        CONVERT_TO_SHORTPTR(GetReference(block_idx));
+    const uint16_t *const source16 = CONVERT_TO_SHORTPTR(source_data_);
+#endif  // CONFIG_HIGHBITDEPTH
+    for (int h = 0; h < height_; ++h) {
+      for (int w = 0; w < width_; ++w) {
+        if (!use_high_bit_depth_) {
+          sad += abs(source8[h * source_stride_ + w] -
+                     reference8[h * reference_stride_ + w]);
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          sad += abs(source16[h * source_stride_ + w] -
+                     reference16[h * reference_stride_ + w]);
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+    }
+    return sad;
+  }
+
+  // Sum of Absolute Differences Average. Given two blocks, and a prediction
+  // calculate the absolute difference between one pixel and average of the
+  // corresponding and predicted pixels; accumulate.
+  unsigned int ReferenceSADavg(int block_idx) {
+    unsigned int sad = 0;
+    const uint8_t *const reference8 = GetReference(block_idx);
+    const uint8_t *const source8 = source_data_;
+    const uint8_t *const second_pred8 = second_pred_;
+#if CONFIG_HIGHBITDEPTH
+    const uint16_t *const reference16 =
+        CONVERT_TO_SHORTPTR(GetReference(block_idx));
+    const uint16_t *const source16 = CONVERT_TO_SHORTPTR(source_data_);
+    const uint16_t *const second_pred16 = CONVERT_TO_SHORTPTR(second_pred_);
+#endif  // CONFIG_HIGHBITDEPTH
+    for (int h = 0; h < height_; ++h) {
+      for (int w = 0; w < width_; ++w) {
+        if (!use_high_bit_depth_) {
+          const int tmp = second_pred8[h * width_ + w] +
+                          reference8[h * reference_stride_ + w];
+          const uint8_t comp_pred = ROUND_POWER_OF_TWO(tmp, 1);
+          sad += abs(source8[h * source_stride_ + w] - comp_pred);
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          const int tmp = second_pred16[h * width_ + w] +
+                          reference16[h * reference_stride_ + w];
+          const uint16_t comp_pred = ROUND_POWER_OF_TWO(tmp, 1);
+          sad += abs(source16[h * source_stride_ + w] - comp_pred);
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+    }
+    return sad;
+  }
+
+  void FillConstant(uint8_t *data, int stride, uint16_t fill_constant) {
+    uint8_t *data8 = data;
+#if CONFIG_HIGHBITDEPTH
+    uint16_t *data16 = CONVERT_TO_SHORTPTR(data);
+#endif  // CONFIG_HIGHBITDEPTH
+    for (int h = 0; h < height_; ++h) {
+      for (int w = 0; w < width_; ++w) {
+        if (!use_high_bit_depth_) {
+          data8[h * stride + w] = static_cast<uint8_t>(fill_constant);
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          data16[h * stride + w] = fill_constant;
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+    }
+  }
+
+  void FillRandom(uint8_t *data, int stride) {
+    uint8_t *data8 = data;
+#if CONFIG_HIGHBITDEPTH
+    uint16_t *data16 = CONVERT_TO_SHORTPTR(data);
+#endif  // CONFIG_HIGHBITDEPTH
+    for (int h = 0; h < height_; ++h) {
+      for (int w = 0; w < width_; ++w) {
+        if (!use_high_bit_depth_) {
+          data8[h * stride + w] = rnd_.Rand8();
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          data16[h * stride + w] = rnd_.Rand16() & mask_;
+#endif  // CONFIG_HIGHBITDEPTH
+        }
+      }
+    }
+  }
+
+  int width_, height_, mask_, bd_;
+  aom_bit_depth_t bit_depth_;
+  static uint8_t *source_data_;
+  static uint8_t *reference_data_;
+  static uint8_t *second_pred_;
+  int source_stride_;
+  bool use_high_bit_depth_;
+  static uint8_t *source_data8_;
+  static uint8_t *reference_data8_;
+  static uint8_t *second_pred8_;
+  static uint16_t *source_data16_;
+  static uint16_t *reference_data16_;
+  static uint16_t *second_pred16_;
+  int reference_stride_;
+
+  ACMRandom rnd_;
+};
+
+class SADx4Test : public SADTestBase,
+                  public ::testing::WithParamInterface<SadMxNx4Param> {
+ public:
+  SADx4Test() : SADTestBase(GET_PARAM(0), GET_PARAM(1), GET_PARAM(3)) {}
+
+ protected:
+  void SADs(unsigned int *results) {
+    const uint8_t *references[] = { GetReference(0), GetReference(1),
+                                    GetReference(2), GetReference(3) };
+
+    ASM_REGISTER_STATE_CHECK(GET_PARAM(2)(
+        source_data_, source_stride_, references, reference_stride_, results));
+  }
+
+  void CheckSADs() {
+    unsigned int reference_sad, exp_sad[4];
+
+    SADs(exp_sad);
+    for (int block = 0; block < 4; ++block) {
+      reference_sad = ReferenceSAD(block);
+
+      EXPECT_EQ(reference_sad, exp_sad[block]) << "block " << block;
+    }
+  }
+};
+
+class SADTest : public SADTestBase,
+                public ::testing::WithParamInterface<SadMxNParam> {
+ public:
+  SADTest() : SADTestBase(GET_PARAM(0), GET_PARAM(1), GET_PARAM(3)) {}
+
+ protected:
+  unsigned int SAD(int block_idx) {
+    unsigned int ret;
+    const uint8_t *const reference = GetReference(block_idx);
+
+    ASM_REGISTER_STATE_CHECK(ret = GET_PARAM(2)(source_data_, source_stride_,
+                                                reference, reference_stride_));
+    return ret;
+  }
+
+  void CheckSAD() {
+    const unsigned int reference_sad = ReferenceSAD(0);
+    const unsigned int exp_sad = SAD(0);
+
+    ASSERT_EQ(reference_sad, exp_sad);
+  }
+
+  void SpeedSAD() {
+    int test_count = 20000000;
+    while (test_count > 0) {
+      SAD(0);
+      test_count -= 1;
+    }
+  }
+};
+
+class SADavgTest : public SADTestBase,
+                   public ::testing::WithParamInterface<SadMxNAvgParam> {
+ public:
+  SADavgTest() : SADTestBase(GET_PARAM(0), GET_PARAM(1), GET_PARAM(3)) {}
+
+ protected:
+  unsigned int SAD_avg(int block_idx) {
+    unsigned int ret;
+    const uint8_t *const reference = GetReference(block_idx);
+
+    ASM_REGISTER_STATE_CHECK(ret = GET_PARAM(2)(source_data_, source_stride_,
+                                                reference, reference_stride_,
+                                                second_pred_));
+    return ret;
+  }
+
+  void CheckSAD() {
+    const unsigned int reference_sad = ReferenceSADavg(0);
+    const unsigned int exp_sad = SAD_avg(0);
+
+    ASSERT_EQ(reference_sad, exp_sad);
+  }
+};
+
+uint8_t *SADTestBase::source_data_ = NULL;
+uint8_t *SADTestBase::reference_data_ = NULL;
+uint8_t *SADTestBase::second_pred_ = NULL;
+uint8_t *SADTestBase::source_data8_ = NULL;
+uint8_t *SADTestBase::reference_data8_ = NULL;
+uint8_t *SADTestBase::second_pred8_ = NULL;
+uint16_t *SADTestBase::source_data16_ = NULL;
+uint16_t *SADTestBase::reference_data16_ = NULL;
+uint16_t *SADTestBase::second_pred16_ = NULL;
+
+TEST_P(SADTest, MaxRef) {
+  FillConstant(source_data_, source_stride_, 0);
+  FillConstant(reference_data_, reference_stride_, mask_);
+  CheckSAD();
+}
+
+TEST_P(SADTest, MaxSrc) {
+  FillConstant(source_data_, source_stride_, mask_);
+  FillConstant(reference_data_, reference_stride_, 0);
+  CheckSAD();
+}
+
+TEST_P(SADTest, ShortRef) {
+  const int tmp_stride = reference_stride_;
+  reference_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADTest, UnalignedRef) {
+  // The reference frame, but not the source frame, may be unaligned for
+  // certain types of searches.
+  const int tmp_stride = reference_stride_;
+  reference_stride_ -= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  CheckSAD();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADTest, ShortSrc) {
+  const int tmp_stride = source_stride_;
+  source_stride_ >>= 1;
+  int test_count = 2000;
+  while (test_count > 0) {
+    FillRandom(source_data_, source_stride_);
+    FillRandom(reference_data_, reference_stride_);
+    CheckSAD();
+    test_count -= 1;
+  }
+  source_stride_ = tmp_stride;
+}
+
+#define SPEED_TEST (0)
+#if SPEED_TEST
+TEST_P(SADTest, Speed) {
+  const int tmp_stride = source_stride_;
+  source_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  SpeedSAD();
+  source_stride_ = tmp_stride;
+}
+#endif
+
+TEST_P(SADavgTest, MaxRef) {
+  FillConstant(source_data_, source_stride_, 0);
+  FillConstant(reference_data_, reference_stride_, mask_);
+  FillConstant(second_pred_, width_, 0);
+  CheckSAD();
+}
+TEST_P(SADavgTest, MaxSrc) {
+  FillConstant(source_data_, source_stride_, mask_);
+  FillConstant(reference_data_, reference_stride_, 0);
+  FillConstant(second_pred_, width_, 0);
+  CheckSAD();
+}
+
+TEST_P(SADavgTest, ShortRef) {
+  const int tmp_stride = reference_stride_;
+  reference_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  FillRandom(second_pred_, width_);
+  CheckSAD();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADavgTest, UnalignedRef) {
+  // The reference frame, but not the source frame, may be unaligned for
+  // certain types of searches.
+  const int tmp_stride = reference_stride_;
+  reference_stride_ -= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(reference_data_, reference_stride_);
+  FillRandom(second_pred_, width_);
+  CheckSAD();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADavgTest, ShortSrc) {
+  const int tmp_stride = source_stride_;
+  source_stride_ >>= 1;
+  int test_count = 2000;
+  while (test_count > 0) {
+    FillRandom(source_data_, source_stride_);
+    FillRandom(reference_data_, reference_stride_);
+    FillRandom(second_pred_, width_);
+    CheckSAD();
+    test_count -= 1;
+  }
+  source_stride_ = tmp_stride;
+}
+
+TEST_P(SADx4Test, MaxRef) {
+  FillConstant(source_data_, source_stride_, 0);
+  FillConstant(GetReference(0), reference_stride_, mask_);
+  FillConstant(GetReference(1), reference_stride_, mask_);
+  FillConstant(GetReference(2), reference_stride_, mask_);
+  FillConstant(GetReference(3), reference_stride_, mask_);
+  CheckSADs();
+}
+
+TEST_P(SADx4Test, MaxSrc) {
+  FillConstant(source_data_, source_stride_, mask_);
+  FillConstant(GetReference(0), reference_stride_, 0);
+  FillConstant(GetReference(1), reference_stride_, 0);
+  FillConstant(GetReference(2), reference_stride_, 0);
+  FillConstant(GetReference(3), reference_stride_, 0);
+  CheckSADs();
+}
+
+TEST_P(SADx4Test, ShortRef) {
+  int tmp_stride = reference_stride_;
+  reference_stride_ >>= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(GetReference(0), reference_stride_);
+  FillRandom(GetReference(1), reference_stride_);
+  FillRandom(GetReference(2), reference_stride_);
+  FillRandom(GetReference(3), reference_stride_);
+  CheckSADs();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADx4Test, UnalignedRef) {
+  // The reference frame, but not the source frame, may be unaligned for
+  // certain types of searches.
+  int tmp_stride = reference_stride_;
+  reference_stride_ -= 1;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(GetReference(0), reference_stride_);
+  FillRandom(GetReference(1), reference_stride_);
+  FillRandom(GetReference(2), reference_stride_);
+  FillRandom(GetReference(3), reference_stride_);
+  CheckSADs();
+  reference_stride_ = tmp_stride;
+}
+
+TEST_P(SADx4Test, ShortSrc) {
+  int tmp_stride = source_stride_;
+  source_stride_ >>= 1;
+  int test_count = 1000;
+  while (test_count > 0) {
+    FillRandom(source_data_, source_stride_);
+    FillRandom(GetReference(0), reference_stride_);
+    FillRandom(GetReference(1), reference_stride_);
+    FillRandom(GetReference(2), reference_stride_);
+    FillRandom(GetReference(3), reference_stride_);
+    CheckSADs();
+    test_count -= 1;
+  }
+  source_stride_ = tmp_stride;
+}
+
+TEST_P(SADx4Test, SrcAlignedByWidth) {
+  uint8_t *tmp_source_data = source_data_;
+  source_data_ += width_;
+  FillRandom(source_data_, source_stride_);
+  FillRandom(GetReference(0), reference_stride_);
+  FillRandom(GetReference(1), reference_stride_);
+  FillRandom(GetReference(2), reference_stride_);
+  FillRandom(GetReference(3), reference_stride_);
+  CheckSADs();
+  source_data_ = tmp_source_data;
+}
+
+using std::tr1::make_tuple;
+
+//------------------------------------------------------------------------------
+// C functions
+const SadMxNParam c_tests[] = {
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_sad128x128_c, -1),
+  make_tuple(128, 64, &aom_sad128x64_c, -1),
+  make_tuple(64, 128, &aom_sad64x128_c, -1),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_sad64x64_c, -1),
+  make_tuple(64, 32, &aom_sad64x32_c, -1),
+  make_tuple(32, 64, &aom_sad32x64_c, -1),
+  make_tuple(32, 32, &aom_sad32x32_c, -1),
+  make_tuple(32, 16, &aom_sad32x16_c, -1),
+  make_tuple(16, 32, &aom_sad16x32_c, -1),
+  make_tuple(16, 16, &aom_sad16x16_c, -1),
+  make_tuple(16, 8, &aom_sad16x8_c, -1),
+  make_tuple(8, 16, &aom_sad8x16_c, -1),
+  make_tuple(8, 8, &aom_sad8x8_c, -1),
+  make_tuple(8, 4, &aom_sad8x4_c, -1),
+  make_tuple(4, 8, &aom_sad4x8_c, -1),
+  make_tuple(4, 4, &aom_sad4x4_c, -1),
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128_c, 8),
+  make_tuple(128, 64, &aom_highbd_sad128x64_c, 8),
+  make_tuple(64, 128, &aom_highbd_sad64x128_c, 8),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_highbd_sad64x64_c, 8),
+  make_tuple(64, 32, &aom_highbd_sad64x32_c, 8),
+  make_tuple(32, 64, &aom_highbd_sad32x64_c, 8),
+  make_tuple(32, 32, &aom_highbd_sad32x32_c, 8),
+  make_tuple(32, 16, &aom_highbd_sad32x16_c, 8),
+  make_tuple(16, 32, &aom_highbd_sad16x32_c, 8),
+  make_tuple(16, 16, &aom_highbd_sad16x16_c, 8),
+  make_tuple(16, 8, &aom_highbd_sad16x8_c, 8),
+  make_tuple(8, 16, &aom_highbd_sad8x16_c, 8),
+  make_tuple(8, 8, &aom_highbd_sad8x8_c, 8),
+  make_tuple(8, 4, &aom_highbd_sad8x4_c, 8),
+  make_tuple(4, 8, &aom_highbd_sad4x8_c, 8),
+  make_tuple(4, 4, &aom_highbd_sad4x4_c, 8),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128_c, 10),
+  make_tuple(128, 64, &aom_highbd_sad128x64_c, 10),
+  make_tuple(64, 128, &aom_highbd_sad64x128_c, 10),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_highbd_sad64x64_c, 10),
+  make_tuple(64, 32, &aom_highbd_sad64x32_c, 10),
+  make_tuple(32, 64, &aom_highbd_sad32x64_c, 10),
+  make_tuple(32, 32, &aom_highbd_sad32x32_c, 10),
+  make_tuple(32, 16, &aom_highbd_sad32x16_c, 10),
+  make_tuple(16, 32, &aom_highbd_sad16x32_c, 10),
+  make_tuple(16, 16, &aom_highbd_sad16x16_c, 10),
+  make_tuple(16, 8, &aom_highbd_sad16x8_c, 10),
+  make_tuple(8, 16, &aom_highbd_sad8x16_c, 10),
+  make_tuple(8, 8, &aom_highbd_sad8x8_c, 10),
+  make_tuple(8, 4, &aom_highbd_sad8x4_c, 10),
+  make_tuple(4, 8, &aom_highbd_sad4x8_c, 10),
+  make_tuple(4, 4, &aom_highbd_sad4x4_c, 10),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128_c, 12),
+  make_tuple(128, 64, &aom_highbd_sad128x64_c, 12),
+  make_tuple(64, 128, &aom_highbd_sad64x128_c, 12),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_highbd_sad64x64_c, 12),
+  make_tuple(64, 32, &aom_highbd_sad64x32_c, 12),
+  make_tuple(32, 64, &aom_highbd_sad32x64_c, 12),
+  make_tuple(32, 32, &aom_highbd_sad32x32_c, 12),
+  make_tuple(32, 16, &aom_highbd_sad32x16_c, 12),
+  make_tuple(16, 32, &aom_highbd_sad16x32_c, 12),
+  make_tuple(16, 16, &aom_highbd_sad16x16_c, 12),
+  make_tuple(16, 8, &aom_highbd_sad16x8_c, 12),
+  make_tuple(8, 16, &aom_highbd_sad8x16_c, 12),
+  make_tuple(8, 8, &aom_highbd_sad8x8_c, 12),
+  make_tuple(8, 4, &aom_highbd_sad8x4_c, 12),
+  make_tuple(4, 8, &aom_highbd_sad4x8_c, 12),
+  make_tuple(4, 4, &aom_highbd_sad4x4_c, 12),
+#endif  // CONFIG_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(C, SADTest, ::testing::ValuesIn(c_tests));
+
+const SadMxNAvgParam avg_c_tests[] = {
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_sad128x128_avg_c, -1),
+  make_tuple(128, 64, &aom_sad128x64_avg_c, -1),
+  make_tuple(64, 128, &aom_sad64x128_avg_c, -1),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_sad64x64_avg_c, -1),
+  make_tuple(64, 32, &aom_sad64x32_avg_c, -1),
+  make_tuple(32, 64, &aom_sad32x64_avg_c, -1),
+  make_tuple(32, 32, &aom_sad32x32_avg_c, -1),
+  make_tuple(32, 16, &aom_sad32x16_avg_c, -1),
+  make_tuple(16, 32, &aom_sad16x32_avg_c, -1),
+  make_tuple(16, 16, &aom_sad16x16_avg_c, -1),
+  make_tuple(16, 8, &aom_sad16x8_avg_c, -1),
+  make_tuple(8, 16, &aom_sad8x16_avg_c, -1),
+  make_tuple(8, 8, &aom_sad8x8_avg_c, -1),
+  make_tuple(8, 4, &aom_sad8x4_avg_c, -1),
+  make_tuple(4, 8, &aom_sad4x8_avg_c, -1),
+  make_tuple(4, 4, &aom_sad4x4_avg_c, -1),
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128_avg_c, 8),
+  make_tuple(128, 64, &aom_highbd_sad128x64_avg_c, 8),
+  make_tuple(64, 128, &aom_highbd_sad64x128_avg_c, 8),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_highbd_sad64x64_avg_c, 8),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avg_c, 8),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avg_c, 8),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avg_c, 8),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avg_c, 8),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avg_c, 8),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avg_c, 8),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avg_c, 8),
+  make_tuple(8, 16, &aom_highbd_sad8x16_avg_c, 8),
+  make_tuple(8, 8, &aom_highbd_sad8x8_avg_c, 8),
+  make_tuple(8, 4, &aom_highbd_sad8x4_avg_c, 8),
+  make_tuple(4, 8, &aom_highbd_sad4x8_avg_c, 8),
+  make_tuple(4, 4, &aom_highbd_sad4x4_avg_c, 8),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128_avg_c, 10),
+  make_tuple(128, 64, &aom_highbd_sad128x64_avg_c, 10),
+  make_tuple(64, 128, &aom_highbd_sad64x128_avg_c, 10),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_highbd_sad64x64_avg_c, 10),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avg_c, 10),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avg_c, 10),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avg_c, 10),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avg_c, 10),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avg_c, 10),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avg_c, 10),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avg_c, 10),
+  make_tuple(8, 16, &aom_highbd_sad8x16_avg_c, 10),
+  make_tuple(8, 8, &aom_highbd_sad8x8_avg_c, 10),
+  make_tuple(8, 4, &aom_highbd_sad8x4_avg_c, 10),
+  make_tuple(4, 8, &aom_highbd_sad4x8_avg_c, 10),
+  make_tuple(4, 4, &aom_highbd_sad4x4_avg_c, 10),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128_avg_c, 12),
+  make_tuple(128, 64, &aom_highbd_sad128x64_avg_c, 12),
+  make_tuple(64, 128, &aom_highbd_sad64x128_avg_c, 12),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_highbd_sad64x64_avg_c, 12),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avg_c, 12),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avg_c, 12),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avg_c, 12),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avg_c, 12),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avg_c, 12),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avg_c, 12),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avg_c, 12),
+  make_tuple(8, 16, &aom_highbd_sad8x16_avg_c, 12),
+  make_tuple(8, 8, &aom_highbd_sad8x8_avg_c, 12),
+  make_tuple(8, 4, &aom_highbd_sad8x4_avg_c, 12),
+  make_tuple(4, 8, &aom_highbd_sad4x8_avg_c, 12),
+  make_tuple(4, 4, &aom_highbd_sad4x4_avg_c, 12),
+#endif  // CONFIG_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(C, SADavgTest, ::testing::ValuesIn(avg_c_tests));
+
+const SadMxNx4Param x4d_c_tests[] = {
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_sad128x128x4d_c, -1),
+  make_tuple(128, 64, &aom_sad128x64x4d_c, -1),
+  make_tuple(64, 128, &aom_sad64x128x4d_c, -1),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_sad64x64x4d_c, -1),
+  make_tuple(64, 32, &aom_sad64x32x4d_c, -1),
+  make_tuple(32, 64, &aom_sad32x64x4d_c, -1),
+  make_tuple(32, 32, &aom_sad32x32x4d_c, -1),
+  make_tuple(32, 16, &aom_sad32x16x4d_c, -1),
+  make_tuple(16, 32, &aom_sad16x32x4d_c, -1),
+  make_tuple(16, 16, &aom_sad16x16x4d_c, -1),
+  make_tuple(16, 8, &aom_sad16x8x4d_c, -1),
+  make_tuple(8, 16, &aom_sad8x16x4d_c, -1),
+  make_tuple(8, 8, &aom_sad8x8x4d_c, -1),
+  make_tuple(8, 4, &aom_sad8x4x4d_c, -1),
+  make_tuple(4, 8, &aom_sad4x8x4d_c, -1),
+  make_tuple(4, 4, &aom_sad4x4x4d_c, -1),
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128x4d_c, 8),
+  make_tuple(128, 64, &aom_highbd_sad128x64x4d_c, 8),
+  make_tuple(64, 128, &aom_highbd_sad64x128x4d_c, 8),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_highbd_sad64x64x4d_c, 8),
+  make_tuple(64, 32, &aom_highbd_sad64x32x4d_c, 8),
+  make_tuple(32, 64, &aom_highbd_sad32x64x4d_c, 8),
+  make_tuple(32, 32, &aom_highbd_sad32x32x4d_c, 8),
+  make_tuple(32, 16, &aom_highbd_sad32x16x4d_c, 8),
+  make_tuple(16, 32, &aom_highbd_sad16x32x4d_c, 8),
+  make_tuple(16, 16, &aom_highbd_sad16x16x4d_c, 8),
+  make_tuple(16, 8, &aom_highbd_sad16x8x4d_c, 8),
+  make_tuple(8, 16, &aom_highbd_sad8x16x4d_c, 8),
+  make_tuple(8, 8, &aom_highbd_sad8x8x4d_c, 8),
+  make_tuple(8, 4, &aom_highbd_sad8x4x4d_c, 8),
+  make_tuple(4, 8, &aom_highbd_sad4x8x4d_c, 8),
+  make_tuple(4, 4, &aom_highbd_sad4x4x4d_c, 8),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128x4d_c, 10),
+  make_tuple(128, 64, &aom_highbd_sad128x64x4d_c, 10),
+  make_tuple(64, 128, &aom_highbd_sad64x128x4d_c, 10),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_highbd_sad64x64x4d_c, 10),
+  make_tuple(64, 32, &aom_highbd_sad64x32x4d_c, 10),
+  make_tuple(32, 64, &aom_highbd_sad32x64x4d_c, 10),
+  make_tuple(32, 32, &aom_highbd_sad32x32x4d_c, 10),
+  make_tuple(32, 16, &aom_highbd_sad32x16x4d_c, 10),
+  make_tuple(16, 32, &aom_highbd_sad16x32x4d_c, 10),
+  make_tuple(16, 16, &aom_highbd_sad16x16x4d_c, 10),
+  make_tuple(16, 8, &aom_highbd_sad16x8x4d_c, 10),
+  make_tuple(8, 16, &aom_highbd_sad8x16x4d_c, 10),
+  make_tuple(8, 8, &aom_highbd_sad8x8x4d_c, 10),
+  make_tuple(8, 4, &aom_highbd_sad8x4x4d_c, 10),
+  make_tuple(4, 8, &aom_highbd_sad4x8x4d_c, 10),
+  make_tuple(4, 4, &aom_highbd_sad4x4x4d_c, 10),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128x4d_c, 12),
+  make_tuple(128, 64, &aom_highbd_sad128x64x4d_c, 12),
+  make_tuple(64, 128, &aom_highbd_sad64x128x4d_c, 12),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_highbd_sad64x64x4d_c, 12),
+  make_tuple(64, 32, &aom_highbd_sad64x32x4d_c, 12),
+  make_tuple(32, 64, &aom_highbd_sad32x64x4d_c, 12),
+  make_tuple(32, 32, &aom_highbd_sad32x32x4d_c, 12),
+  make_tuple(32, 16, &aom_highbd_sad32x16x4d_c, 12),
+  make_tuple(16, 32, &aom_highbd_sad16x32x4d_c, 12),
+  make_tuple(16, 16, &aom_highbd_sad16x16x4d_c, 12),
+  make_tuple(16, 8, &aom_highbd_sad16x8x4d_c, 12),
+  make_tuple(8, 16, &aom_highbd_sad8x16x4d_c, 12),
+  make_tuple(8, 8, &aom_highbd_sad8x8x4d_c, 12),
+  make_tuple(8, 4, &aom_highbd_sad8x4x4d_c, 12),
+  make_tuple(4, 8, &aom_highbd_sad4x8x4d_c, 12),
+  make_tuple(4, 4, &aom_highbd_sad4x4x4d_c, 12),
+#endif  // CONFIG_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(C, SADx4Test, ::testing::ValuesIn(x4d_c_tests));
+
+//------------------------------------------------------------------------------
+// ARM functions
+#if HAVE_MEDIA
+const SadMxNParam media_tests[] = {
+  make_tuple(16, 16, &aom_sad16x16_media, -1),
+};
+INSTANTIATE_TEST_CASE_P(MEDIA, SADTest, ::testing::ValuesIn(media_tests));
+#endif  // HAVE_MEDIA
+
+#if HAVE_NEON
+const SadMxNParam neon_tests[] = {
+  make_tuple(64, 64, &aom_sad64x64_neon, -1),
+  make_tuple(32, 32, &aom_sad32x32_neon, -1),
+  make_tuple(16, 16, &aom_sad16x16_neon, -1),
+  make_tuple(16, 8, &aom_sad16x8_neon, -1),
+  make_tuple(8, 16, &aom_sad8x16_neon, -1),
+  make_tuple(8, 8, &aom_sad8x8_neon, -1),
+  make_tuple(4, 4, &aom_sad4x4_neon, -1),
+};
+INSTANTIATE_TEST_CASE_P(NEON, SADTest, ::testing::ValuesIn(neon_tests));
+
+const SadMxNx4Param x4d_neon_tests[] = {
+  make_tuple(64, 64, &aom_sad64x64x4d_neon, -1),
+  make_tuple(32, 32, &aom_sad32x32x4d_neon, -1),
+  make_tuple(16, 16, &aom_sad16x16x4d_neon, -1),
+};
+INSTANTIATE_TEST_CASE_P(NEON, SADx4Test, ::testing::ValuesIn(x4d_neon_tests));
+#endif  // HAVE_NEON
+
+//------------------------------------------------------------------------------
+// x86 functions
+#if HAVE_SSE2
+const SadMxNParam sse2_tests[] = {
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_sad128x128_sse2, -1),
+  make_tuple(128, 64, &aom_sad128x64_sse2, -1),
+  make_tuple(64, 128, &aom_sad64x128_sse2, -1),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_sad64x64_sse2, -1),
+  make_tuple(64, 32, &aom_sad64x32_sse2, -1),
+  make_tuple(32, 64, &aom_sad32x64_sse2, -1),
+  make_tuple(32, 32, &aom_sad32x32_sse2, -1),
+  make_tuple(32, 16, &aom_sad32x16_sse2, -1),
+  make_tuple(16, 32, &aom_sad16x32_sse2, -1),
+  make_tuple(16, 16, &aom_sad16x16_sse2, -1),
+  make_tuple(16, 8, &aom_sad16x8_sse2, -1),
+  make_tuple(8, 16, &aom_sad8x16_sse2, -1),
+  make_tuple(8, 8, &aom_sad8x8_sse2, -1),
+  make_tuple(8, 4, &aom_sad8x4_sse2, -1),
+  make_tuple(4, 8, &aom_sad4x8_sse2, -1),
+  make_tuple(4, 4, &aom_sad4x4_sse2, -1),
+#if CONFIG_HIGHBITDEPTH
+  make_tuple(64, 64, &aom_highbd_sad64x64_sse2, 8),
+  make_tuple(64, 32, &aom_highbd_sad64x32_sse2, 8),
+  make_tuple(32, 64, &aom_highbd_sad32x64_sse2, 8),
+  make_tuple(32, 32, &aom_highbd_sad32x32_sse2, 8),
+  make_tuple(32, 16, &aom_highbd_sad32x16_sse2, 8),
+  make_tuple(16, 32, &aom_highbd_sad16x32_sse2, 8),
+  make_tuple(16, 16, &aom_highbd_sad16x16_sse2, 8),
+  make_tuple(16, 8, &aom_highbd_sad16x8_sse2, 8),
+  make_tuple(8, 16, &aom_highbd_sad8x16_sse2, 8),
+  make_tuple(8, 8, &aom_highbd_sad8x8_sse2, 8),
+  make_tuple(8, 4, &aom_highbd_sad8x4_sse2, 8),
+  make_tuple(64, 64, &aom_highbd_sad64x64_sse2, 10),
+  make_tuple(64, 32, &aom_highbd_sad64x32_sse2, 10),
+  make_tuple(32, 64, &aom_highbd_sad32x64_sse2, 10),
+  make_tuple(32, 32, &aom_highbd_sad32x32_sse2, 10),
+  make_tuple(32, 16, &aom_highbd_sad32x16_sse2, 10),
+  make_tuple(16, 32, &aom_highbd_sad16x32_sse2, 10),
+  make_tuple(16, 16, &aom_highbd_sad16x16_sse2, 10),
+  make_tuple(16, 8, &aom_highbd_sad16x8_sse2, 10),
+  make_tuple(8, 16, &aom_highbd_sad8x16_sse2, 10),
+  make_tuple(8, 8, &aom_highbd_sad8x8_sse2, 10),
+  make_tuple(8, 4, &aom_highbd_sad8x4_sse2, 10),
+  make_tuple(64, 64, &aom_highbd_sad64x64_sse2, 12),
+  make_tuple(64, 32, &aom_highbd_sad64x32_sse2, 12),
+  make_tuple(32, 64, &aom_highbd_sad32x64_sse2, 12),
+  make_tuple(32, 32, &aom_highbd_sad32x32_sse2, 12),
+  make_tuple(32, 16, &aom_highbd_sad32x16_sse2, 12),
+  make_tuple(16, 32, &aom_highbd_sad16x32_sse2, 12),
+  make_tuple(16, 16, &aom_highbd_sad16x16_sse2, 12),
+  make_tuple(16, 8, &aom_highbd_sad16x8_sse2, 12),
+  make_tuple(8, 16, &aom_highbd_sad8x16_sse2, 12),
+  make_tuple(8, 8, &aom_highbd_sad8x8_sse2, 12),
+  make_tuple(8, 4, &aom_highbd_sad8x4_sse2, 12),
+#endif  // CONFIG_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(SSE2, SADTest, ::testing::ValuesIn(sse2_tests));
+
+const SadMxNAvgParam avg_sse2_tests[] = {
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_sad128x128_avg_sse2, -1),
+  make_tuple(128, 64, &aom_sad128x64_avg_sse2, -1),
+  make_tuple(64, 128, &aom_sad64x128_avg_sse2, -1),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_sad64x64_avg_sse2, -1),
+  make_tuple(64, 32, &aom_sad64x32_avg_sse2, -1),
+  make_tuple(32, 64, &aom_sad32x64_avg_sse2, -1),
+  make_tuple(32, 32, &aom_sad32x32_avg_sse2, -1),
+  make_tuple(32, 16, &aom_sad32x16_avg_sse2, -1),
+  make_tuple(16, 32, &aom_sad16x32_avg_sse2, -1),
+  make_tuple(16, 16, &aom_sad16x16_avg_sse2, -1),
+  make_tuple(16, 8, &aom_sad16x8_avg_sse2, -1),
+  make_tuple(8, 16, &aom_sad8x16_avg_sse2, -1),
+  make_tuple(8, 8, &aom_sad8x8_avg_sse2, -1),
+  make_tuple(8, 4, &aom_sad8x4_avg_sse2, -1),
+  make_tuple(4, 8, &aom_sad4x8_avg_sse2, -1),
+  make_tuple(4, 4, &aom_sad4x4_avg_sse2, -1),
+#if CONFIG_HIGHBITDEPTH
+  make_tuple(64, 64, &aom_highbd_sad64x64_avg_sse2, 8),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avg_sse2, 8),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avg_sse2, 8),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avg_sse2, 8),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avg_sse2, 8),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avg_sse2, 8),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avg_sse2, 8),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avg_sse2, 8),
+  make_tuple(8, 16, &aom_highbd_sad8x16_avg_sse2, 8),
+  make_tuple(8, 8, &aom_highbd_sad8x8_avg_sse2, 8),
+  make_tuple(8, 4, &aom_highbd_sad8x4_avg_sse2, 8),
+  make_tuple(64, 64, &aom_highbd_sad64x64_avg_sse2, 10),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avg_sse2, 10),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avg_sse2, 10),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avg_sse2, 10),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avg_sse2, 10),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avg_sse2, 10),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avg_sse2, 10),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avg_sse2, 10),
+  make_tuple(8, 16, &aom_highbd_sad8x16_avg_sse2, 10),
+  make_tuple(8, 8, &aom_highbd_sad8x8_avg_sse2, 10),
+  make_tuple(8, 4, &aom_highbd_sad8x4_avg_sse2, 10),
+  make_tuple(64, 64, &aom_highbd_sad64x64_avg_sse2, 12),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avg_sse2, 12),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avg_sse2, 12),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avg_sse2, 12),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avg_sse2, 12),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avg_sse2, 12),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avg_sse2, 12),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avg_sse2, 12),
+  make_tuple(8, 16, &aom_highbd_sad8x16_avg_sse2, 12),
+  make_tuple(8, 8, &aom_highbd_sad8x8_avg_sse2, 12),
+  make_tuple(8, 4, &aom_highbd_sad8x4_avg_sse2, 12),
+#endif  // CONFIG_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(SSE2, SADavgTest, ::testing::ValuesIn(avg_sse2_tests));
+
+const SadMxNx4Param x4d_sse2_tests[] = {
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_sad128x128x4d_sse2, -1),
+  make_tuple(128, 64, &aom_sad128x64x4d_sse2, -1),
+  make_tuple(64, 128, &aom_sad64x128x4d_sse2, -1),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(64, 64, &aom_sad64x64x4d_sse2, -1),
+  make_tuple(64, 32, &aom_sad64x32x4d_sse2, -1),
+  make_tuple(32, 64, &aom_sad32x64x4d_sse2, -1),
+  make_tuple(32, 32, &aom_sad32x32x4d_sse2, -1),
+  make_tuple(32, 16, &aom_sad32x16x4d_sse2, -1),
+  make_tuple(16, 32, &aom_sad16x32x4d_sse2, -1),
+  make_tuple(16, 16, &aom_sad16x16x4d_sse2, -1),
+  make_tuple(16, 8, &aom_sad16x8x4d_sse2, -1),
+  make_tuple(8, 16, &aom_sad8x16x4d_sse2, -1),
+  make_tuple(8, 8, &aom_sad8x8x4d_sse2, -1),
+  make_tuple(8, 4, &aom_sad8x4x4d_sse2, -1),
+  make_tuple(4, 8, &aom_sad4x8x4d_sse2, -1),
+  make_tuple(4, 4, &aom_sad4x4x4d_sse2, -1),
+#if CONFIG_HIGHBITDEPTH
+  make_tuple(64, 64, &aom_highbd_sad64x64x4d_sse2, 8),
+  make_tuple(64, 32, &aom_highbd_sad64x32x4d_sse2, 8),
+  make_tuple(32, 64, &aom_highbd_sad32x64x4d_sse2, 8),
+  make_tuple(32, 32, &aom_highbd_sad32x32x4d_sse2, 8),
+  make_tuple(32, 16, &aom_highbd_sad32x16x4d_sse2, 8),
+  make_tuple(16, 32, &aom_highbd_sad16x32x4d_sse2, 8),
+  make_tuple(16, 16, &aom_highbd_sad16x16x4d_sse2, 8),
+  make_tuple(16, 8, &aom_highbd_sad16x8x4d_sse2, 8),
+  make_tuple(8, 16, &aom_highbd_sad8x16x4d_sse2, 8),
+  make_tuple(8, 8, &aom_highbd_sad8x8x4d_sse2, 8),
+  make_tuple(8, 4, &aom_highbd_sad8x4x4d_sse2, 8),
+  make_tuple(4, 8, &aom_highbd_sad4x8x4d_sse2, 8),
+  make_tuple(4, 4, &aom_highbd_sad4x4x4d_sse2, 8),
+  make_tuple(64, 64, &aom_highbd_sad64x64x4d_sse2, 10),
+  make_tuple(64, 32, &aom_highbd_sad64x32x4d_sse2, 10),
+  make_tuple(32, 64, &aom_highbd_sad32x64x4d_sse2, 10),
+  make_tuple(32, 32, &aom_highbd_sad32x32x4d_sse2, 10),
+  make_tuple(32, 16, &aom_highbd_sad32x16x4d_sse2, 10),
+  make_tuple(16, 32, &aom_highbd_sad16x32x4d_sse2, 10),
+  make_tuple(16, 16, &aom_highbd_sad16x16x4d_sse2, 10),
+  make_tuple(16, 8, &aom_highbd_sad16x8x4d_sse2, 10),
+  make_tuple(8, 16, &aom_highbd_sad8x16x4d_sse2, 10),
+  make_tuple(8, 8, &aom_highbd_sad8x8x4d_sse2, 10),
+  make_tuple(8, 4, &aom_highbd_sad8x4x4d_sse2, 10),
+  make_tuple(4, 8, &aom_highbd_sad4x8x4d_sse2, 10),
+  make_tuple(4, 4, &aom_highbd_sad4x4x4d_sse2, 10),
+  make_tuple(64, 64, &aom_highbd_sad64x64x4d_sse2, 12),
+  make_tuple(64, 32, &aom_highbd_sad64x32x4d_sse2, 12),
+  make_tuple(32, 64, &aom_highbd_sad32x64x4d_sse2, 12),
+  make_tuple(32, 32, &aom_highbd_sad32x32x4d_sse2, 12),
+  make_tuple(32, 16, &aom_highbd_sad32x16x4d_sse2, 12),
+  make_tuple(16, 32, &aom_highbd_sad16x32x4d_sse2, 12),
+  make_tuple(16, 16, &aom_highbd_sad16x16x4d_sse2, 12),
+  make_tuple(16, 8, &aom_highbd_sad16x8x4d_sse2, 12),
+  make_tuple(8, 16, &aom_highbd_sad8x16x4d_sse2, 12),
+  make_tuple(8, 8, &aom_highbd_sad8x8x4d_sse2, 12),
+  make_tuple(8, 4, &aom_highbd_sad8x4x4d_sse2, 12),
+  make_tuple(4, 8, &aom_highbd_sad4x8x4d_sse2, 12),
+  make_tuple(4, 4, &aom_highbd_sad4x4x4d_sse2, 12),
+#endif  // CONFIG_HIGHBITDEPTH
+};
+INSTANTIATE_TEST_CASE_P(SSE2, SADx4Test, ::testing::ValuesIn(x4d_sse2_tests));
+#endif  // HAVE_SSE2
+
+#if HAVE_SSE3
+// Only functions are x3, which do not have tests.
+#endif  // HAVE_SSE3
+
+#if HAVE_SSSE3
+// Only functions are x3, which do not have tests.
+#endif  // HAVE_SSSE3
+
+#if HAVE_SSE4_1
+// Only functions are x8, which do not have tests.
+#endif  // HAVE_SSE4_1
+
+#if HAVE_AVX2
+const SadMxNParam avx2_tests[] = {
+#if CONFIG_EXT_PARTITION
+  make_tuple(64, 128, &aom_sad64x128_avx2, -1),
+  make_tuple(128, 64, &aom_sad128x64_avx2, -1),
+  make_tuple(128, 128, &aom_sad128x128_avx2, -1),
+#endif
+  make_tuple(64, 64, &aom_sad64x64_avx2, -1),
+  make_tuple(64, 32, &aom_sad64x32_avx2, -1),
+  make_tuple(32, 64, &aom_sad32x64_avx2, -1),
+  make_tuple(32, 32, &aom_sad32x32_avx2, -1),
+  make_tuple(32, 16, &aom_sad32x16_avx2, -1),
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128_avx2, 8),
+  make_tuple(128, 128, &aom_highbd_sad128x128_avx2, 10),
+  make_tuple(128, 128, &aom_highbd_sad128x128_avx2, 12),
+  make_tuple(128, 64, &aom_highbd_sad128x64_avx2, 8),
+  make_tuple(128, 64, &aom_highbd_sad128x64_avx2, 10),
+  make_tuple(128, 64, &aom_highbd_sad128x64_avx2, 12),
+  make_tuple(64, 128, &aom_highbd_sad64x128_avx2, 8),
+  make_tuple(64, 128, &aom_highbd_sad64x128_avx2, 10),
+  make_tuple(64, 128, &aom_highbd_sad64x128_avx2, 12),
+#endif
+  make_tuple(64, 64, &aom_highbd_sad64x64_avx2, 8),
+  make_tuple(64, 64, &aom_highbd_sad64x64_avx2, 10),
+  make_tuple(64, 64, &aom_highbd_sad64x64_avx2, 12),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avx2, 8),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avx2, 10),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avx2, 12),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avx2, 8),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avx2, 10),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avx2, 12),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avx2, 8),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avx2, 10),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avx2, 12),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avx2, 8),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avx2, 10),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avx2, 12),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avx2, 8),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avx2, 10),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avx2, 12),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avx2, 8),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avx2, 10),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avx2, 12),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avx2, 8),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avx2, 10),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avx2, 12),
+#endif
+};
+INSTANTIATE_TEST_CASE_P(AVX2, SADTest, ::testing::ValuesIn(avx2_tests));
+
+const SadMxNAvgParam avg_avx2_tests[] = {
+#if CONFIG_EXT_PARTITION
+  make_tuple(64, 128, &aom_sad64x128_avg_avx2, -1),
+  make_tuple(128, 64, &aom_sad128x64_avg_avx2, -1),
+  make_tuple(128, 128, &aom_sad128x128_avg_avx2, -1),
+#endif
+  make_tuple(64, 64, &aom_sad64x64_avg_avx2, -1),
+  make_tuple(64, 32, &aom_sad64x32_avg_avx2, -1),
+  make_tuple(32, 64, &aom_sad32x64_avg_avx2, -1),
+  make_tuple(32, 32, &aom_sad32x32_avg_avx2, -1),
+  make_tuple(32, 16, &aom_sad32x16_avg_avx2, -1),
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128_avg_avx2, 8),
+  make_tuple(128, 128, &aom_highbd_sad128x128_avg_avx2, 10),
+  make_tuple(128, 128, &aom_highbd_sad128x128_avg_avx2, 12),
+  make_tuple(128, 64, &aom_highbd_sad128x64_avg_avx2, 8),
+  make_tuple(128, 64, &aom_highbd_sad128x64_avg_avx2, 10),
+  make_tuple(128, 64, &aom_highbd_sad128x64_avg_avx2, 12),
+  make_tuple(64, 128, &aom_highbd_sad64x128_avg_avx2, 8),
+  make_tuple(64, 128, &aom_highbd_sad64x128_avg_avx2, 10),
+  make_tuple(64, 128, &aom_highbd_sad64x128_avg_avx2, 12),
+#endif
+  make_tuple(64, 64, &aom_highbd_sad64x64_avg_avx2, 8),
+  make_tuple(64, 64, &aom_highbd_sad64x64_avg_avx2, 10),
+  make_tuple(64, 64, &aom_highbd_sad64x64_avg_avx2, 12),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avg_avx2, 8),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avg_avx2, 10),
+  make_tuple(64, 32, &aom_highbd_sad64x32_avg_avx2, 12),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avg_avx2, 8),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avg_avx2, 10),
+  make_tuple(32, 64, &aom_highbd_sad32x64_avg_avx2, 12),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avg_avx2, 8),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avg_avx2, 10),
+  make_tuple(32, 32, &aom_highbd_sad32x32_avg_avx2, 12),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avg_avx2, 8),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avg_avx2, 10),
+  make_tuple(32, 16, &aom_highbd_sad32x16_avg_avx2, 12),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avg_avx2, 8),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avg_avx2, 10),
+  make_tuple(16, 32, &aom_highbd_sad16x32_avg_avx2, 12),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avg_avx2, 8),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avg_avx2, 10),
+  make_tuple(16, 16, &aom_highbd_sad16x16_avg_avx2, 12),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avg_avx2, 8),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avg_avx2, 10),
+  make_tuple(16, 8, &aom_highbd_sad16x8_avg_avx2, 12),
+#endif
+};
+INSTANTIATE_TEST_CASE_P(AVX2, SADavgTest, ::testing::ValuesIn(avg_avx2_tests));
+
+const SadMxNx4Param x4d_avx2_tests[] = {
+#if CONFIG_EXT_PARTITION
+  make_tuple(64, 128, &aom_sad64x128x4d_avx2, -1),
+  make_tuple(128, 64, &aom_sad128x64x4d_avx2, -1),
+  make_tuple(128, 128, &aom_sad128x128x4d_avx2, -1),
+#endif
+  make_tuple(64, 64, &aom_sad64x64x4d_avx2, -1),
+  make_tuple(32, 64, &aom_sad32x64x4d_avx2, -1),
+  make_tuple(64, 32, &aom_sad64x32x4d_avx2, -1),
+  make_tuple(32, 32, &aom_sad32x32x4d_avx2, -1),
+#if CONFIG_HIGHBITDEPTH
+#if CONFIG_EXT_PARTITION
+  make_tuple(128, 128, &aom_highbd_sad128x128x4d_avx2, 8),
+  make_tuple(128, 128, &aom_highbd_sad128x128x4d_avx2, 10),
+  make_tuple(128, 128, &aom_highbd_sad128x128x4d_avx2, 12),
+  make_tuple(128, 64, &aom_highbd_sad128x64x4d_avx2, 8),
+  make_tuple(128, 64, &aom_highbd_sad128x64x4d_avx2, 10),
+  make_tuple(128, 64, &aom_highbd_sad128x64x4d_avx2, 12),
+  make_tuple(64, 128, &aom_highbd_sad64x128x4d_avx2, 8),
+  make_tuple(64, 128, &aom_highbd_sad64x128x4d_avx2, 10),
+  make_tuple(64, 128, &aom_highbd_sad64x128x4d_avx2, 12),
+#endif
+  make_tuple(64, 64, &aom_highbd_sad64x64x4d_avx2, 8),
+  make_tuple(64, 64, &aom_highbd_sad64x64x4d_avx2, 10),
+  make_tuple(64, 64, &aom_highbd_sad64x64x4d_avx2, 12),
+  make_tuple(64, 32, &aom_highbd_sad64x32x4d_avx2, 8),
+  make_tuple(64, 32, &aom_highbd_sad64x32x4d_avx2, 10),
+  make_tuple(64, 32, &aom_highbd_sad64x32x4d_avx2, 12),
+  make_tuple(32, 64, &aom_highbd_sad32x64x4d_avx2, 8),
+  make_tuple(32, 64, &aom_highbd_sad32x64x4d_avx2, 10),
+  make_tuple(32, 64, &aom_highbd_sad32x64x4d_avx2, 12),
+  make_tuple(32, 32, &aom_highbd_sad32x32x4d_avx2, 8),
+  make_tuple(32, 32, &aom_highbd_sad32x32x4d_avx2, 10),
+  make_tuple(32, 32, &aom_highbd_sad32x32x4d_avx2, 12),
+  make_tuple(32, 16, &aom_highbd_sad32x16x4d_avx2, 8),
+  make_tuple(32, 16, &aom_highbd_sad32x16x4d_avx2, 10),
+  make_tuple(32, 16, &aom_highbd_sad32x16x4d_avx2, 12),
+  make_tuple(16, 32, &aom_highbd_sad16x32x4d_avx2, 8),
+  make_tuple(16, 32, &aom_highbd_sad16x32x4d_avx2, 10),
+  make_tuple(16, 32, &aom_highbd_sad16x32x4d_avx2, 12),
+  make_tuple(16, 16, &aom_highbd_sad16x16x4d_avx2, 8),
+  make_tuple(16, 16, &aom_highbd_sad16x16x4d_avx2, 10),
+  make_tuple(16, 16, &aom_highbd_sad16x16x4d_avx2, 12),
+  make_tuple(16, 8, &aom_highbd_sad16x8x4d_avx2, 8),
+  make_tuple(16, 8, &aom_highbd_sad16x8x4d_avx2, 10),
+  make_tuple(16, 8, &aom_highbd_sad16x8x4d_avx2, 12),
+#endif
+};
+INSTANTIATE_TEST_CASE_P(AVX2, SADx4Test, ::testing::ValuesIn(x4d_avx2_tests));
+#endif  // HAVE_AVX2
+
+//------------------------------------------------------------------------------
+// MIPS functions
+#if HAVE_MSA
+const SadMxNParam msa_tests[] = {
+  make_tuple(64, 64, &aom_sad64x64_msa, -1),
+  make_tuple(64, 32, &aom_sad64x32_msa, -1),
+  make_tuple(32, 64, &aom_sad32x64_msa, -1),
+  make_tuple(32, 32, &aom_sad32x32_msa, -1),
+  make_tuple(32, 16, &aom_sad32x16_msa, -1),
+  make_tuple(16, 32, &aom_sad16x32_msa, -1),
+  make_tuple(16, 16, &aom_sad16x16_msa, -1),
+  make_tuple(16, 8, &aom_sad16x8_msa, -1),
+  make_tuple(8, 16, &aom_sad8x16_msa, -1),
+  make_tuple(8, 8, &aom_sad8x8_msa, -1),
+  make_tuple(8, 4, &aom_sad8x4_msa, -1),
+  make_tuple(4, 8, &aom_sad4x8_msa, -1),
+  make_tuple(4, 4, &aom_sad4x4_msa, -1),
+};
+INSTANTIATE_TEST_CASE_P(MSA, SADTest, ::testing::ValuesIn(msa_tests));
+
+const SadMxNAvgParam avg_msa_tests[] = {
+  make_tuple(64, 64, &aom_sad64x64_avg_msa, -1),
+  make_tuple(64, 32, &aom_sad64x32_avg_msa, -1),
+  make_tuple(32, 64, &aom_sad32x64_avg_msa, -1),
+  make_tuple(32, 32, &aom_sad32x32_avg_msa, -1),
+  make_tuple(32, 16, &aom_sad32x16_avg_msa, -1),
+  make_tuple(16, 32, &aom_sad16x32_avg_msa, -1),
+  make_tuple(16, 16, &aom_sad16x16_avg_msa, -1),
+  make_tuple(16, 8, &aom_sad16x8_avg_msa, -1),
+  make_tuple(8, 16, &aom_sad8x16_avg_msa, -1),
+  make_tuple(8, 8, &aom_sad8x8_avg_msa, -1),
+  make_tuple(8, 4, &aom_sad8x4_avg_msa, -1),
+  make_tuple(4, 8, &aom_sad4x8_avg_msa, -1),
+  make_tuple(4, 4, &aom_sad4x4_avg_msa, -1),
+};
+INSTANTIATE_TEST_CASE_P(MSA, SADavgTest, ::testing::ValuesIn(avg_msa_tests));
+
+const SadMxNx4Param x4d_msa_tests[] = {
+  make_tuple(64, 64, &aom_sad64x64x4d_msa, -1),
+  make_tuple(64, 32, &aom_sad64x32x4d_msa, -1),
+  make_tuple(32, 64, &aom_sad32x64x4d_msa, -1),
+  make_tuple(32, 32, &aom_sad32x32x4d_msa, -1),
+  make_tuple(32, 16, &aom_sad32x16x4d_msa, -1),
+  make_tuple(16, 32, &aom_sad16x32x4d_msa, -1),
+  make_tuple(16, 16, &aom_sad16x16x4d_msa, -1),
+  make_tuple(16, 8, &aom_sad16x8x4d_msa, -1),
+  make_tuple(8, 16, &aom_sad8x16x4d_msa, -1),
+  make_tuple(8, 8, &aom_sad8x8x4d_msa, -1),
+  make_tuple(8, 4, &aom_sad8x4x4d_msa, -1),
+  make_tuple(4, 8, &aom_sad4x8x4d_msa, -1),
+  make_tuple(4, 4, &aom_sad4x4x4d_msa, -1),
+};
+INSTANTIATE_TEST_CASE_P(MSA, SADx4Test, ::testing::ValuesIn(x4d_msa_tests));
+#endif  // HAVE_MSA
+
+}  // namespace
diff --git a/third_party/aom/test/scan_test.cc b/third_party/aom/test/scan_test.cc
new file mode 100644
index 0000000000..16c831c8ea
--- /dev/null
+++ b/third_party/aom/test/scan_test.cc
@@ -0,0 +1,97 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "av1/common/common_data.h"
+#include "av1/common/scan.h"
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+namespace {
+
+TEST(ScanTest, av1_augment_prob) {
+  const TX_SIZE tx_size = TX_4X4;
+  const TX_TYPE tx_type = DCT_DCT;
+  const int tx1d_size = tx_size_wide[tx_size];
+  uint32_t prob[16] = { 8, 8, 7, 7, 8, 8, 4, 2, 3, 3, 2, 2, 2, 2, 2, 2 };
+  const uint32_t ref_prob[16] = {
+    8, 8, 7, 7, 8, 8, 4, 2, 3, 3, 2, 2, 2, 2, 2, 2
+  };
+  av1_augment_prob(tx_size, tx_type, prob);
+  for (int r = 0; r < tx1d_size; ++r) {
+    for (int c = 0; c < tx1d_size; ++c) {
+      const uint32_t idx = r * tx1d_size + c;
+      EXPECT_EQ(ref_prob[idx], prob[idx] >> 16);
+    }
+  }
+
+  const SCAN_ORDER *sc = get_default_scan(tx_size, tx_type, 0);
+  const uint32_t mask = (1 << 16) - 1;
+  for (int r = 0; r < tx1d_size; ++r) {
+    for (int c = 0; c < tx1d_size; ++c) {
+      const uint32_t ref_idx = r * tx1d_size + c;
+      const uint32_t scan_idx = mask ^ (prob[r * tx1d_size + c] & mask);
+      const uint32_t idx = sc->scan[scan_idx];
+      EXPECT_EQ(ref_idx, idx);
+    }
+  }
+}
+
+TEST(ScanTest, av1_update_sort_order) {
+  const TX_SIZE tx_size = TX_4X4;
+  const TX_TYPE tx_type = DCT_DCT;
+  const uint32_t prob[16] = { 15, 14, 11, 10, 13, 12, 9, 5,
+                              8,  7,  4,  2,  6,  3,  1, 0 };
+  const int16_t ref_sort_order[16] = { 0, 1,  4, 5,  2,  3,  6,  8,
+                                       9, 12, 7, 10, 13, 11, 14, 15 };
+  int16_t sort_order[16];
+  av1_update_sort_order(tx_size, tx_type, prob, sort_order);
+  for (int i = 0; i < 16; ++i) EXPECT_EQ(ref_sort_order[i], sort_order[i]);
+}
+
+TEST(ScanTest, av1_update_scan_order) {
+  TX_SIZE tx_size = TX_4X4;
+  const TX_TYPE tx_type = DCT_DCT;
+  const uint32_t prob[16] = { 10, 12, 14, 9, 11, 13, 15, 5,
+                              8,  7,  4,  2, 6,  3,  1,  0 };
+  int16_t sort_order[16];
+  int16_t scan[16];
+  int16_t iscan[16];
+  const int16_t ref_iscan[16] = { 0, 1, 2,  6,  3, 4,  5,  10,
+                                  7, 8, 11, 13, 9, 12, 14, 15 };
+
+  av1_update_sort_order(tx_size, tx_type, prob, sort_order);
+  av1_update_scan_order(tx_size, sort_order, scan, iscan);
+
+  for (int i = 0; i < 16; ++i) {
+    EXPECT_EQ(ref_iscan[i], iscan[i]);
+    EXPECT_EQ(i, scan[ref_iscan[i]]);
+  }
+}
+
+TEST(ScanTest, av1_update_neighbors) {
+  TX_SIZE tx_size = TX_4X4;
+  // raster order
+  const int16_t scan[16] = { 0, 1, 2,  3,  4,  5,  6,  7,
+                             8, 9, 10, 11, 12, 13, 14, 15 };
+  int16_t nb[(16 + 1) * 2];
+  const int16_t ref_nb[(16 + 1) * 2] = { 0,  0,  0,  0,  1,  1,  2, 2, 0,
+                                         1,  1,  4,  2,  5,  3,  6, 4, 5,
+                                         5,  8,  6,  9,  7,  10, 8, 9, 9,
+                                         12, 10, 13, 11, 14, 0,  0 };
+
+  // raster order's scan and iscan are the same
+  av1_update_neighbors(tx_size, scan, scan, nb);
+
+  for (int i = 0; i < (16 + 1) * 2; ++i) {
+    EXPECT_EQ(ref_nb[i], nb[i]);
+  }
+}
+
+}  // namespace
diff --git a/third_party/aom/test/selfguided_filter_test.cc b/third_party/aom/test/selfguided_filter_test.cc
new file mode 100644
index 0000000000..e87fe339a3
--- /dev/null
+++ b/third_party/aom/test/selfguided_filter_test.cc
@@ -0,0 +1,274 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <ctime>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./av1_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+
+#include "av1/common/mv.h"
+#include "av1/common/restoration.h"
+
+namespace {
+
+using std::tr1::tuple;
+using std::tr1::make_tuple;
+using libaom_test::ACMRandom;
+
+typedef tuple<> FilterTestParam;
+
+class AV1SelfguidedFilterTest
+    : public ::testing::TestWithParam<FilterTestParam> {
+ public:
+  virtual ~AV1SelfguidedFilterTest() {}
+  virtual void SetUp() {}
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunSpeedTest() {
+    const int w = 256, h = 256;
+    const int NUM_ITERS = 2000;
+    int i, j;
+
+    uint8_t *input = (uint8_t *)aom_memalign(16, w * h * sizeof(uint8_t));
+    uint8_t *output = (uint8_t *)aom_memalign(16, w * h * sizeof(uint8_t));
+    int32_t *tmpbuf = (int32_t *)aom_memalign(16, RESTORATION_TMPBUF_SIZE);
+    memset(tmpbuf, 0, RESTORATION_TMPBUF_SIZE);
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+    for (i = 0; i < h; ++i)
+      for (j = 0; j < w; ++j) input[i * w + j] = rnd.Rand16() & 0xFF;
+
+    int xqd[2] = {
+      SGRPROJ_PRJ_MIN0 +
+          rnd.PseudoUniform(SGRPROJ_PRJ_MAX0 + 1 - SGRPROJ_PRJ_MIN0),
+      SGRPROJ_PRJ_MIN1 +
+          rnd.PseudoUniform(SGRPROJ_PRJ_MAX1 + 1 - SGRPROJ_PRJ_MIN1)
+    };
+    // Fix a parameter set, since the speed depends slightly on r.
+    // Change this to test different combinations of values of r.
+    int eps = 15;
+
+    av1_loop_restoration_precal();
+
+    std::clock_t start = std::clock();
+    for (i = 0; i < NUM_ITERS; ++i) {
+      apply_selfguided_restoration(input, w, h, w, eps, xqd, output, w, tmpbuf);
+    }
+    std::clock_t end = std::clock();
+    double elapsed = ((end - start) / (double)CLOCKS_PER_SEC);
+
+    printf("%5d %dx%d blocks in %7.3fs = %7.3fus/block\n", NUM_ITERS, w, h,
+           elapsed, elapsed * 1000000. / NUM_ITERS);
+
+    aom_free(input);
+    aom_free(output);
+    aom_free(tmpbuf);
+  }
+
+  void RunCorrectnessTest() {
+    // Set the maximum width/height to test here. We actually test a small
+    // range of sizes *up to* this size, so that we can check, eg.,
+    // the behaviour on tiles which are not a multiple of 4 wide.
+    const int max_w = 260, max_h = 260, stride = 672, out_stride = 672;
+    const int NUM_ITERS = 81;
+    int i, j, k;
+
+    uint8_t *input =
+        (uint8_t *)aom_memalign(16, stride * max_h * sizeof(uint8_t));
+    uint8_t *output =
+        (uint8_t *)aom_memalign(16, out_stride * max_h * sizeof(uint8_t));
+    uint8_t *output2 =
+        (uint8_t *)aom_memalign(16, out_stride * max_h * sizeof(uint8_t));
+    int32_t *tmpbuf = (int32_t *)aom_memalign(16, RESTORATION_TMPBUF_SIZE);
+    memset(tmpbuf, 0, RESTORATION_TMPBUF_SIZE);
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+    av1_loop_restoration_precal();
+
+    for (i = 0; i < NUM_ITERS; ++i) {
+      for (j = 0; j < max_h; ++j)
+        for (k = 0; k < max_w; ++k) input[j * stride + k] = rnd.Rand16() & 0xFF;
+
+      int xqd[2] = {
+        SGRPROJ_PRJ_MIN0 +
+            rnd.PseudoUniform(SGRPROJ_PRJ_MAX0 + 1 - SGRPROJ_PRJ_MIN0),
+        SGRPROJ_PRJ_MIN1 +
+            rnd.PseudoUniform(SGRPROJ_PRJ_MAX1 + 1 - SGRPROJ_PRJ_MIN1)
+      };
+      int eps = rnd.PseudoUniform(1 << SGRPROJ_PARAMS_BITS);
+
+      // Test various tile sizes around 256x256
+      int test_w = max_w - (i / 9);
+      int test_h = max_h - (i % 9);
+
+      apply_selfguided_restoration(input, test_w, test_h, stride, eps, xqd,
+                                   output, out_stride, tmpbuf);
+      apply_selfguided_restoration_c(input, test_w, test_h, stride, eps, xqd,
+                                     output2, out_stride, tmpbuf);
+      for (j = 0; j < test_h; ++j)
+        for (k = 0; k < test_w; ++k)
+          ASSERT_EQ(output[j * out_stride + k], output2[j * out_stride + k]);
+    }
+
+    aom_free(input);
+    aom_free(output);
+    aom_free(output2);
+    aom_free(tmpbuf);
+  }
+};
+
+TEST_P(AV1SelfguidedFilterTest, SpeedTest) { RunSpeedTest(); }
+TEST_P(AV1SelfguidedFilterTest, CorrectnessTest) { RunCorrectnessTest(); }
+
+const FilterTestParam params[] = { make_tuple() };
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(SSE4_1, AV1SelfguidedFilterTest,
+                        ::testing::ValuesIn(params));
+#endif
+
+#if CONFIG_HIGHBITDEPTH
+
+typedef tuple<int> HighbdFilterTestParam;
+
+class AV1HighbdSelfguidedFilterTest
+    : public ::testing::TestWithParam<HighbdFilterTestParam> {
+ public:
+  virtual ~AV1HighbdSelfguidedFilterTest() {}
+  virtual void SetUp() {}
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void RunSpeedTest() {
+    const int w = 256, h = 256;
+    const int NUM_ITERS = 2000;
+    int i, j;
+    int bit_depth = GET_PARAM(0);
+    int mask = (1 << bit_depth) - 1;
+
+    uint16_t *input = (uint16_t *)aom_memalign(16, w * h * sizeof(uint16_t));
+    uint16_t *output = (uint16_t *)aom_memalign(16, w * h * sizeof(uint16_t));
+    int32_t *tmpbuf = (int32_t *)aom_memalign(16, RESTORATION_TMPBUF_SIZE);
+    memset(tmpbuf, 0, RESTORATION_TMPBUF_SIZE);
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+    for (i = 0; i < h; ++i)
+      for (j = 0; j < w; ++j) input[i * w + j] = rnd.Rand16() & mask;
+
+    int xqd[2] = {
+      SGRPROJ_PRJ_MIN0 +
+          rnd.PseudoUniform(SGRPROJ_PRJ_MAX0 + 1 - SGRPROJ_PRJ_MIN0),
+      SGRPROJ_PRJ_MIN1 +
+          rnd.PseudoUniform(SGRPROJ_PRJ_MAX1 + 1 - SGRPROJ_PRJ_MIN1)
+    };
+    // Fix a parameter set, since the speed depends slightly on r.
+    // Change this to test different combinations of values of r.
+    int eps = 15;
+
+    av1_loop_restoration_precal();
+
+    std::clock_t start = std::clock();
+    for (i = 0; i < NUM_ITERS; ++i) {
+      apply_selfguided_restoration_highbd(input, w, h, w, bit_depth, eps, xqd,
+                                          output, w, tmpbuf);
+    }
+    std::clock_t end = std::clock();
+    double elapsed = ((end - start) / (double)CLOCKS_PER_SEC);
+
+    printf("%5d %dx%d blocks in %7.3fs = %7.3fus/block\n", NUM_ITERS, w, h,
+           elapsed, elapsed * 1000000. / NUM_ITERS);
+
+    aom_free(input);
+    aom_free(output);
+    aom_free(tmpbuf);
+  }
+
+  void RunCorrectnessTest() {
+    // Set the maximum width/height to test here. We actually test a small
+    // range of sizes *up to* this size, so that we can check, eg.,
+    // the behaviour on tiles which are not a multiple of 4 wide.
+    const int max_w = 260, max_h = 260, stride = 672, out_stride = 672;
+    const int NUM_ITERS = 81;
+    int i, j, k;
+    int bit_depth = GET_PARAM(0);
+    int mask = (1 << bit_depth) - 1;
+
+    uint16_t *input =
+        (uint16_t *)aom_memalign(16, stride * max_h * sizeof(uint16_t));
+    uint16_t *output =
+        (uint16_t *)aom_memalign(16, out_stride * max_h * sizeof(uint16_t));
+    uint16_t *output2 =
+        (uint16_t *)aom_memalign(16, out_stride * max_h * sizeof(uint16_t));
+    int32_t *tmpbuf = (int32_t *)aom_memalign(16, RESTORATION_TMPBUF_SIZE);
+    memset(tmpbuf, 0, RESTORATION_TMPBUF_SIZE);
+
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+    av1_loop_restoration_precal();
+
+    for (i = 0; i < NUM_ITERS; ++i) {
+      for (j = 0; j < max_h; ++j)
+        for (k = 0; k < max_w; ++k) input[j * stride + k] = rnd.Rand16() & mask;
+
+      int xqd[2] = {
+        SGRPROJ_PRJ_MIN0 +
+            rnd.PseudoUniform(SGRPROJ_PRJ_MAX0 + 1 - SGRPROJ_PRJ_MIN0),
+        SGRPROJ_PRJ_MIN1 +
+            rnd.PseudoUniform(SGRPROJ_PRJ_MAX1 + 1 - SGRPROJ_PRJ_MIN1)
+      };
+      int eps = rnd.PseudoUniform(1 << SGRPROJ_PARAMS_BITS);
+
+      // Test various tile sizes around 256x256
+      int test_w = max_w - (i / 9);
+      int test_h = max_h - (i % 9);
+
+      apply_selfguided_restoration_highbd(input, test_w, test_h, stride,
+                                          bit_depth, eps, xqd, output,
+                                          out_stride, tmpbuf);
+      apply_selfguided_restoration_highbd_c(input, test_w, test_h, stride,
+                                            bit_depth, eps, xqd, output2,
+                                            out_stride, tmpbuf);
+      for (j = 0; j < test_h; ++j)
+        for (k = 0; k < test_w; ++k)
+          ASSERT_EQ(output[j * out_stride + k], output2[j * out_stride + k]);
+    }
+
+    aom_free(input);
+    aom_free(output);
+    aom_free(output2);
+    aom_free(tmpbuf);
+  }
+};
+
+TEST_P(AV1HighbdSelfguidedFilterTest, SpeedTest) { RunSpeedTest(); }
+TEST_P(AV1HighbdSelfguidedFilterTest, CorrectnessTest) { RunCorrectnessTest(); }
+
+const HighbdFilterTestParam highbd_params[] = { make_tuple(8), make_tuple(10),
+                                                make_tuple(12) };
+
+#if HAVE_SSE4_1
+INSTANTIATE_TEST_CASE_P(SSE4_1, AV1HighbdSelfguidedFilterTest,
+                        ::testing::ValuesIn(highbd_params));
+#endif
+#endif
+
+}  // namespace
diff --git a/third_party/aom/test/set_maps.sh b/third_party/aom/test/set_maps.sh
new file mode 100755
index 0000000000..4f59b06d69
--- /dev/null
+++ b/third_party/aom/test/set_maps.sh
@@ -0,0 +1,52 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This file tests the libaom set_maps example. To add new tests to this file,
+## do the following:
+##   1. Write a shell function (this is your test).
+##   2. Add the function to set_maps_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required, and set_maps must exist in
+# $LIBAOM_BIN_PATH.
+set_maps_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libaom test data must exist in LIBAOM_TEST_DATA_PATH."
+    return 1
+  fi
+  if [ -z "$(aom_tool_path set_maps)" ]; then
+    elog "set_maps not found. It must exist in LIBAOM_BIN_PATH or its parent."
+    return 1
+  fi
+}
+
+# Runs set_maps using the codec specified by $1.
+set_maps() {
+  local encoder="$(aom_tool_path set_maps)"
+  local codec="$1"
+  local output_file="${AOM_TEST_OUTPUT_DIR}/set_maps_${codec}.ivf"
+
+  eval "${AOM_TEST_PREFIX}" "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+set_maps_av1() {
+  if [ "$(av1_encode_available)" = "yes" ]; then
+    set_maps av1 || return 1
+  fi
+}
+
+set_maps_tests="set_maps_av1"
+
+run_tests set_maps_verify_environment "${set_maps_tests}"
diff --git a/third_party/aom/test/simd_cmp_impl.h b/third_party/aom/test/simd_cmp_impl.h
new file mode 100644
index 0000000000..28bd64a5bc
--- /dev/null
+++ b/third_party/aom/test/simd_cmp_impl.h
@@ -0,0 +1,1212 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <assert.h>
+#include <string>
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "aom_dsp/aom_simd.h"
+#undef SIMD_INLINE
+#define SIMD_INLINE static  // Don't enforce inlining
+#include "aom_dsp/simd/v128_intrinsics_c.h"
+
+// Machine tuned code goes into this file. This file is included from
+// simd_cmp_sse2.cc, simd_cmp_ssse3.cc etc which define the macros
+// ARCH (=neon, sse2, ssse3, etc), SIMD_NAMESPACE and ARCH_POSTFIX().
+
+using libaom_test::ACMRandom;
+
+namespace SIMD_NAMESPACE {
+
+// Wrap templates around intrinsics using immediate values
+template <int shift>
+v64 imm_v64_shl_n_byte(v64 a) {
+  return v64_shl_n_byte(a, shift);
+}
+template <int shift>
+v64 imm_v64_shr_n_byte(v64 a) {
+  return v64_shr_n_byte(a, shift);
+}
+template <int shift>
+v64 imm_v64_shl_n_8(v64 a) {
+  return v64_shl_n_8(a, shift);
+}
+template <int shift>
+v64 imm_v64_shr_n_u8(v64 a) {
+  return v64_shr_n_u8(a, shift);
+}
+template <int shift>
+v64 imm_v64_shr_n_s8(v64 a) {
+  return v64_shr_n_s8(a, shift);
+}
+template <int shift>
+v64 imm_v64_shl_n_16(v64 a) {
+  return v64_shl_n_16(a, shift);
+}
+template <int shift>
+v64 imm_v64_shr_n_u16(v64 a) {
+  return v64_shr_n_u16(a, shift);
+}
+template <int shift>
+v64 imm_v64_shr_n_s16(v64 a) {
+  return v64_shr_n_s16(a, shift);
+}
+template <int shift>
+v64 imm_v64_shl_n_32(v64 a) {
+  return v64_shl_n_32(a, shift);
+}
+template <int shift>
+v64 imm_v64_shr_n_u32(v64 a) {
+  return v64_shr_n_u32(a, shift);
+}
+template <int shift>
+v64 imm_v64_shr_n_s32(v64 a) {
+  return v64_shr_n_s32(a, shift);
+}
+template <int shift>
+v64 imm_v64_align(v64 a, v64 b) {
+  return v64_align(a, b, shift);
+}
+
+// Wrap templates around corresponding C implementations of the above
+template <int shift>
+c_v64 c_imm_v64_shl_n_byte(c_v64 a) {
+  return c_v64_shl_n_byte(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_shr_n_byte(c_v64 a) {
+  return c_v64_shr_n_byte(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_shl_n_8(c_v64 a) {
+  return c_v64_shl_n_8(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_shr_n_u8(c_v64 a) {
+  return c_v64_shr_n_u8(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_shr_n_s8(c_v64 a) {
+  return c_v64_shr_n_s8(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_shl_n_16(c_v64 a) {
+  return c_v64_shl_n_16(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_shr_n_u16(c_v64 a) {
+  return c_v64_shr_n_u16(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_shr_n_s16(c_v64 a) {
+  return c_v64_shr_n_s16(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_shl_n_32(c_v64 a) {
+  return c_v64_shl_n_32(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_shr_n_u32(c_v64 a) {
+  return c_v64_shr_n_u32(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_shr_n_s32(c_v64 a) {
+  return c_v64_shr_n_s32(a, shift);
+}
+template <int shift>
+c_v64 c_imm_v64_align(c_v64 a, c_v64 b) {
+  return c_v64_align(a, b, shift);
+}
+
+template <int shift>
+v128 imm_v128_shl_n_byte(v128 a) {
+  return v128_shl_n_byte(a, shift);
+}
+template <int shift>
+v128 imm_v128_shr_n_byte(v128 a) {
+  return v128_shr_n_byte(a, shift);
+}
+template <int shift>
+v128 imm_v128_shl_n_8(v128 a) {
+  return v128_shl_n_8(a, shift);
+}
+template <int shift>
+v128 imm_v128_shr_n_u8(v128 a) {
+  return v128_shr_n_u8(a, shift);
+}
+template <int shift>
+v128 imm_v128_shr_n_s8(v128 a) {
+  return v128_shr_n_s8(a, shift);
+}
+template <int shift>
+v128 imm_v128_shl_n_16(v128 a) {
+  return v128_shl_n_16(a, shift);
+}
+template <int shift>
+v128 imm_v128_shr_n_u16(v128 a) {
+  return v128_shr_n_u16(a, shift);
+}
+template <int shift>
+v128 imm_v128_shr_n_s16(v128 a) {
+  return v128_shr_n_s16(a, shift);
+}
+template <int shift>
+v128 imm_v128_shl_n_32(v128 a) {
+  return v128_shl_n_32(a, shift);
+}
+template <int shift>
+v128 imm_v128_shr_n_u32(v128 a) {
+  return v128_shr_n_u32(a, shift);
+}
+template <int shift>
+v128 imm_v128_shr_n_s32(v128 a) {
+  return v128_shr_n_s32(a, shift);
+}
+template <int shift>
+v128 imm_v128_align(v128 a, v128 b) {
+  return v128_align(a, b, shift);
+}
+
+template <int shift>
+c_v128 c_imm_v128_shl_n_byte(c_v128 a) {
+  return c_v128_shl_n_byte(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_shr_n_byte(c_v128 a) {
+  return c_v128_shr_n_byte(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_shl_n_8(c_v128 a) {
+  return c_v128_shl_n_8(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_shr_n_u8(c_v128 a) {
+  return c_v128_shr_n_u8(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_shr_n_s8(c_v128 a) {
+  return c_v128_shr_n_s8(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_shl_n_16(c_v128 a) {
+  return c_v128_shl_n_16(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_shr_n_u16(c_v128 a) {
+  return c_v128_shr_n_u16(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_shr_n_s16(c_v128 a) {
+  return c_v128_shr_n_s16(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_shl_n_32(c_v128 a) {
+  return c_v128_shl_n_32(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_shr_n_u32(c_v128 a) {
+  return c_v128_shr_n_u32(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_shr_n_s32(c_v128 a) {
+  return c_v128_shr_n_s32(a, shift);
+}
+template <int shift>
+c_v128 c_imm_v128_align(c_v128 a, c_v128 b) {
+  return c_v128_align(a, b, shift);
+}
+
+// Wrappers around the the SAD and SSD functions
+uint32_t v64_sad_u8(v64 a, v64 b) {
+  return v64_sad_u8_sum(::v64_sad_u8(v64_sad_u8_init(), a, b));
+}
+uint32_t v64_ssd_u8(v64 a, v64 b) {
+  return v64_ssd_u8_sum(::v64_ssd_u8(v64_ssd_u8_init(), a, b));
+}
+
+uint32_t c_v64_sad_u8(c_v64 a, c_v64 b) {
+  return c_v64_sad_u8_sum(::c_v64_sad_u8(c_v64_sad_u8_init(), a, b));
+}
+uint32_t c_v64_ssd_u8(c_v64 a, c_v64 b) {
+  return c_v64_ssd_u8_sum(::c_v64_ssd_u8(c_v64_ssd_u8_init(), a, b));
+}
+uint32_t v128_sad_u8(v128 a, v128 b) {
+  return v128_sad_u8_sum(::v128_sad_u8(v128_sad_u8_init(), a, b));
+}
+uint32_t v128_ssd_u8(v128 a, v128 b) {
+  return v128_ssd_u8_sum(::v128_ssd_u8(v128_ssd_u8_init(), a, b));
+}
+uint32_t c_v128_sad_u8(c_v128 a, c_v128 b) {
+  return c_v128_sad_u8_sum(::c_v128_sad_u8(c_v128_sad_u8_init(), a, b));
+}
+uint32_t c_v128_ssd_u8(c_v128 a, c_v128 b) {
+  return c_v128_ssd_u8_sum(::c_v128_ssd_u8(c_v128_ssd_u8_init(), a, b));
+}
+
+namespace {
+
+typedef void (*fptr)();
+
+typedef struct {
+  const char *name;
+  fptr ref;
+  fptr simd;
+} mapping;
+
+#define MAP(name)                                                      \
+  {                                                                    \
+    #name, reinterpret_cast < fptr > (c_##name),                       \
+                                      reinterpret_cast < fptr > (name) \
+  }
+
+const mapping m[] = { MAP(v64_sad_u8),
+                      MAP(v64_ssd_u8),
+                      MAP(v64_add_8),
+                      MAP(v64_add_16),
+                      MAP(v64_sadd_s16),
+                      MAP(v64_add_32),
+                      MAP(v64_sub_8),
+                      MAP(v64_ssub_u8),
+                      MAP(v64_ssub_s8),
+                      MAP(v64_sub_16),
+                      MAP(v64_ssub_s16),
+                      MAP(v64_ssub_u16),
+                      MAP(v64_sub_32),
+                      MAP(v64_ziplo_8),
+                      MAP(v64_ziphi_8),
+                      MAP(v64_ziplo_16),
+                      MAP(v64_ziphi_16),
+                      MAP(v64_ziplo_32),
+                      MAP(v64_ziphi_32),
+                      MAP(v64_pack_s32_s16),
+                      MAP(v64_pack_s16_u8),
+                      MAP(v64_pack_s16_s8),
+                      MAP(v64_unziphi_8),
+                      MAP(v64_unziplo_8),
+                      MAP(v64_unziphi_16),
+                      MAP(v64_unziplo_16),
+                      MAP(v64_or),
+                      MAP(v64_xor),
+                      MAP(v64_and),
+                      MAP(v64_andn),
+                      MAP(v64_mullo_s16),
+                      MAP(v64_mulhi_s16),
+                      MAP(v64_mullo_s32),
+                      MAP(v64_madd_s16),
+                      MAP(v64_madd_us8),
+                      MAP(v64_avg_u8),
+                      MAP(v64_rdavg_u8),
+                      MAP(v64_avg_u16),
+                      MAP(v64_min_u8),
+                      MAP(v64_max_u8),
+                      MAP(v64_min_s8),
+                      MAP(v64_max_s8),
+                      MAP(v64_min_s16),
+                      MAP(v64_max_s16),
+                      MAP(v64_cmpgt_s8),
+                      MAP(v64_cmplt_s8),
+                      MAP(v64_cmpeq_8),
+                      MAP(v64_cmpgt_s16),
+                      MAP(v64_cmplt_s16),
+                      MAP(v64_cmpeq_16),
+                      MAP(v64_shuffle_8),
+                      MAP(imm_v64_align<1>),
+                      MAP(imm_v64_align<2>),
+                      MAP(imm_v64_align<3>),
+                      MAP(imm_v64_align<4>),
+                      MAP(imm_v64_align<5>),
+                      MAP(imm_v64_align<6>),
+                      MAP(imm_v64_align<7>),
+                      MAP(v64_abs_s8),
+                      MAP(v64_abs_s16),
+                      MAP(v64_unpacklo_u8_s16),
+                      MAP(v64_unpackhi_u8_s16),
+                      MAP(v64_unpacklo_s8_s16),
+                      MAP(v64_unpackhi_s8_s16),
+                      MAP(v64_unpacklo_u16_s32),
+                      MAP(v64_unpacklo_s16_s32),
+                      MAP(v64_unpackhi_u16_s32),
+                      MAP(v64_unpackhi_s16_s32),
+                      MAP(imm_v64_shr_n_byte<1>),
+                      MAP(imm_v64_shr_n_byte<2>),
+                      MAP(imm_v64_shr_n_byte<3>),
+                      MAP(imm_v64_shr_n_byte<4>),
+                      MAP(imm_v64_shr_n_byte<5>),
+                      MAP(imm_v64_shr_n_byte<6>),
+                      MAP(imm_v64_shr_n_byte<7>),
+                      MAP(imm_v64_shl_n_byte<1>),
+                      MAP(imm_v64_shl_n_byte<2>),
+                      MAP(imm_v64_shl_n_byte<3>),
+                      MAP(imm_v64_shl_n_byte<4>),
+                      MAP(imm_v64_shl_n_byte<5>),
+                      MAP(imm_v64_shl_n_byte<6>),
+                      MAP(imm_v64_shl_n_byte<7>),
+                      MAP(imm_v64_shl_n_8<1>),
+                      MAP(imm_v64_shl_n_8<2>),
+                      MAP(imm_v64_shl_n_8<3>),
+                      MAP(imm_v64_shl_n_8<4>),
+                      MAP(imm_v64_shl_n_8<5>),
+                      MAP(imm_v64_shl_n_8<6>),
+                      MAP(imm_v64_shl_n_8<7>),
+                      MAP(imm_v64_shr_n_u8<1>),
+                      MAP(imm_v64_shr_n_u8<2>),
+                      MAP(imm_v64_shr_n_u8<3>),
+                      MAP(imm_v64_shr_n_u8<4>),
+                      MAP(imm_v64_shr_n_u8<5>),
+                      MAP(imm_v64_shr_n_u8<6>),
+                      MAP(imm_v64_shr_n_u8<7>),
+                      MAP(imm_v64_shr_n_s8<1>),
+                      MAP(imm_v64_shr_n_s8<2>),
+                      MAP(imm_v64_shr_n_s8<3>),
+                      MAP(imm_v64_shr_n_s8<4>),
+                      MAP(imm_v64_shr_n_s8<5>),
+                      MAP(imm_v64_shr_n_s8<6>),
+                      MAP(imm_v64_shr_n_s8<7>),
+                      MAP(imm_v64_shl_n_16<1>),
+                      MAP(imm_v64_shl_n_16<2>),
+                      MAP(imm_v64_shl_n_16<4>),
+                      MAP(imm_v64_shl_n_16<6>),
+                      MAP(imm_v64_shl_n_16<8>),
+                      MAP(imm_v64_shl_n_16<10>),
+                      MAP(imm_v64_shl_n_16<12>),
+                      MAP(imm_v64_shl_n_16<14>),
+                      MAP(imm_v64_shr_n_u16<1>),
+                      MAP(imm_v64_shr_n_u16<2>),
+                      MAP(imm_v64_shr_n_u16<4>),
+                      MAP(imm_v64_shr_n_u16<6>),
+                      MAP(imm_v64_shr_n_u16<8>),
+                      MAP(imm_v64_shr_n_u16<10>),
+                      MAP(imm_v64_shr_n_u16<12>),
+                      MAP(imm_v64_shr_n_u16<14>),
+                      MAP(imm_v64_shr_n_s16<1>),
+                      MAP(imm_v64_shr_n_s16<2>),
+                      MAP(imm_v64_shr_n_s16<4>),
+                      MAP(imm_v64_shr_n_s16<6>),
+                      MAP(imm_v64_shr_n_s16<8>),
+                      MAP(imm_v64_shr_n_s16<10>),
+                      MAP(imm_v64_shr_n_s16<12>),
+                      MAP(imm_v64_shr_n_s16<14>),
+                      MAP(imm_v64_shl_n_32<1>),
+                      MAP(imm_v64_shl_n_32<4>),
+                      MAP(imm_v64_shl_n_32<8>),
+                      MAP(imm_v64_shl_n_32<12>),
+                      MAP(imm_v64_shl_n_32<16>),
+                      MAP(imm_v64_shl_n_32<20>),
+                      MAP(imm_v64_shl_n_32<24>),
+                      MAP(imm_v64_shl_n_32<28>),
+                      MAP(imm_v64_shr_n_u32<1>),
+                      MAP(imm_v64_shr_n_u32<4>),
+                      MAP(imm_v64_shr_n_u32<8>),
+                      MAP(imm_v64_shr_n_u32<12>),
+                      MAP(imm_v64_shr_n_u32<16>),
+                      MAP(imm_v64_shr_n_u32<20>),
+                      MAP(imm_v64_shr_n_u32<24>),
+                      MAP(imm_v64_shr_n_u32<28>),
+                      MAP(imm_v64_shr_n_s32<1>),
+                      MAP(imm_v64_shr_n_s32<4>),
+                      MAP(imm_v64_shr_n_s32<8>),
+                      MAP(imm_v64_shr_n_s32<12>),
+                      MAP(imm_v64_shr_n_s32<16>),
+                      MAP(imm_v64_shr_n_s32<20>),
+                      MAP(imm_v64_shr_n_s32<24>),
+                      MAP(imm_v64_shr_n_s32<28>),
+                      MAP(v64_shl_8),
+                      MAP(v64_shr_u8),
+                      MAP(v64_shr_s8),
+                      MAP(v64_shl_16),
+                      MAP(v64_shr_u16),
+                      MAP(v64_shr_s16),
+                      MAP(v64_shl_32),
+                      MAP(v64_shr_u32),
+                      MAP(v64_shr_s32),
+                      MAP(v64_hadd_u8),
+                      MAP(v64_hadd_s16),
+                      MAP(v64_dotp_s16),
+                      MAP(v64_dotp_su8),
+                      MAP(v64_u64),
+                      MAP(v64_low_u32),
+                      MAP(v64_high_u32),
+                      MAP(v64_low_s32),
+                      MAP(v64_high_s32),
+                      MAP(v64_dup_8),
+                      MAP(v64_dup_16),
+                      MAP(v64_dup_32),
+                      MAP(v64_from_32),
+                      MAP(v64_zero),
+                      MAP(v64_from_16),
+                      MAP(v128_sad_u8),
+                      MAP(v128_ssd_u8),
+                      MAP(v128_add_8),
+                      MAP(v128_add_16),
+                      MAP(v128_sadd_s16),
+                      MAP(v128_add_32),
+                      MAP(v128_sub_8),
+                      MAP(v128_ssub_u8),
+                      MAP(v128_ssub_s8),
+                      MAP(v128_sub_16),
+                      MAP(v128_ssub_s16),
+                      MAP(v128_ssub_u16),
+                      MAP(v128_sub_32),
+                      MAP(v128_ziplo_8),
+                      MAP(v128_ziphi_8),
+                      MAP(v128_ziplo_16),
+                      MAP(v128_ziphi_16),
+                      MAP(v128_ziplo_32),
+                      MAP(v128_ziphi_32),
+                      MAP(v128_ziplo_64),
+                      MAP(v128_ziphi_64),
+                      MAP(v128_unziphi_8),
+                      MAP(v128_unziplo_8),
+                      MAP(v128_unziphi_16),
+                      MAP(v128_unziplo_16),
+                      MAP(v128_unziphi_32),
+                      MAP(v128_unziplo_32),
+                      MAP(v128_pack_s32_s16),
+                      MAP(v128_pack_s16_u8),
+                      MAP(v128_pack_s16_s8),
+                      MAP(v128_or),
+                      MAP(v128_xor),
+                      MAP(v128_and),
+                      MAP(v128_andn),
+                      MAP(v128_mullo_s16),
+                      MAP(v128_mulhi_s16),
+                      MAP(v128_mullo_s32),
+                      MAP(v128_madd_s16),
+                      MAP(v128_madd_us8),
+                      MAP(v128_avg_u8),
+                      MAP(v128_rdavg_u8),
+                      MAP(v128_avg_u16),
+                      MAP(v128_min_u8),
+                      MAP(v128_max_u8),
+                      MAP(v128_min_s8),
+                      MAP(v128_max_s8),
+                      MAP(v128_min_s16),
+                      MAP(v128_max_s16),
+                      MAP(v128_cmpgt_s8),
+                      MAP(v128_cmplt_s8),
+                      MAP(v128_cmpeq_8),
+                      MAP(v128_cmpgt_s16),
+                      MAP(v128_cmpeq_16),
+                      MAP(v128_cmplt_s16),
+                      MAP(v128_shuffle_8),
+                      MAP(imm_v128_align<1>),
+                      MAP(imm_v128_align<2>),
+                      MAP(imm_v128_align<3>),
+                      MAP(imm_v128_align<4>),
+                      MAP(imm_v128_align<5>),
+                      MAP(imm_v128_align<6>),
+                      MAP(imm_v128_align<7>),
+                      MAP(imm_v128_align<8>),
+                      MAP(imm_v128_align<9>),
+                      MAP(imm_v128_align<10>),
+                      MAP(imm_v128_align<11>),
+                      MAP(imm_v128_align<12>),
+                      MAP(imm_v128_align<13>),
+                      MAP(imm_v128_align<14>),
+                      MAP(imm_v128_align<15>),
+                      MAP(v128_abs_s8),
+                      MAP(v128_abs_s16),
+                      MAP(v128_padd_s16),
+                      MAP(v128_unpacklo_u16_s32),
+                      MAP(v128_unpacklo_s16_s32),
+                      MAP(v128_unpackhi_u16_s32),
+                      MAP(v128_unpackhi_s16_s32),
+                      MAP(imm_v128_shr_n_byte<1>),
+                      MAP(imm_v128_shr_n_byte<2>),
+                      MAP(imm_v128_shr_n_byte<3>),
+                      MAP(imm_v128_shr_n_byte<4>),
+                      MAP(imm_v128_shr_n_byte<5>),
+                      MAP(imm_v128_shr_n_byte<6>),
+                      MAP(imm_v128_shr_n_byte<7>),
+                      MAP(imm_v128_shr_n_byte<8>),
+                      MAP(imm_v128_shr_n_byte<9>),
+                      MAP(imm_v128_shr_n_byte<10>),
+                      MAP(imm_v128_shr_n_byte<11>),
+                      MAP(imm_v128_shr_n_byte<12>),
+                      MAP(imm_v128_shr_n_byte<13>),
+                      MAP(imm_v128_shr_n_byte<14>),
+                      MAP(imm_v128_shr_n_byte<15>),
+                      MAP(imm_v128_shl_n_byte<1>),
+                      MAP(imm_v128_shl_n_byte<2>),
+                      MAP(imm_v128_shl_n_byte<3>),
+                      MAP(imm_v128_shl_n_byte<4>),
+                      MAP(imm_v128_shl_n_byte<5>),
+                      MAP(imm_v128_shl_n_byte<6>),
+                      MAP(imm_v128_shl_n_byte<7>),
+                      MAP(imm_v128_shl_n_byte<8>),
+                      MAP(imm_v128_shl_n_byte<9>),
+                      MAP(imm_v128_shl_n_byte<10>),
+                      MAP(imm_v128_shl_n_byte<11>),
+                      MAP(imm_v128_shl_n_byte<12>),
+                      MAP(imm_v128_shl_n_byte<13>),
+                      MAP(imm_v128_shl_n_byte<14>),
+                      MAP(imm_v128_shl_n_byte<15>),
+                      MAP(imm_v128_shl_n_8<1>),
+                      MAP(imm_v128_shl_n_8<2>),
+                      MAP(imm_v128_shl_n_8<3>),
+                      MAP(imm_v128_shl_n_8<4>),
+                      MAP(imm_v128_shl_n_8<5>),
+                      MAP(imm_v128_shl_n_8<6>),
+                      MAP(imm_v128_shl_n_8<7>),
+                      MAP(imm_v128_shr_n_u8<1>),
+                      MAP(imm_v128_shr_n_u8<2>),
+                      MAP(imm_v128_shr_n_u8<3>),
+                      MAP(imm_v128_shr_n_u8<4>),
+                      MAP(imm_v128_shr_n_u8<5>),
+                      MAP(imm_v128_shr_n_u8<6>),
+                      MAP(imm_v128_shr_n_u8<7>),
+                      MAP(imm_v128_shr_n_s8<1>),
+                      MAP(imm_v128_shr_n_s8<2>),
+                      MAP(imm_v128_shr_n_s8<3>),
+                      MAP(imm_v128_shr_n_s8<4>),
+                      MAP(imm_v128_shr_n_s8<5>),
+                      MAP(imm_v128_shr_n_s8<6>),
+                      MAP(imm_v128_shr_n_s8<7>),
+                      MAP(imm_v128_shl_n_16<1>),
+                      MAP(imm_v128_shl_n_16<2>),
+                      MAP(imm_v128_shl_n_16<4>),
+                      MAP(imm_v128_shl_n_16<6>),
+                      MAP(imm_v128_shl_n_16<8>),
+                      MAP(imm_v128_shl_n_16<10>),
+                      MAP(imm_v128_shl_n_16<12>),
+                      MAP(imm_v128_shl_n_16<14>),
+                      MAP(imm_v128_shr_n_u16<1>),
+                      MAP(imm_v128_shr_n_u16<2>),
+                      MAP(imm_v128_shr_n_u16<4>),
+                      MAP(imm_v128_shr_n_u16<6>),
+                      MAP(imm_v128_shr_n_u16<8>),
+                      MAP(imm_v128_shr_n_u16<10>),
+                      MAP(imm_v128_shr_n_u16<12>),
+                      MAP(imm_v128_shr_n_u16<14>),
+                      MAP(imm_v128_shr_n_s16<1>),
+                      MAP(imm_v128_shr_n_s16<2>),
+                      MAP(imm_v128_shr_n_s16<4>),
+                      MAP(imm_v128_shr_n_s16<6>),
+                      MAP(imm_v128_shr_n_s16<8>),
+                      MAP(imm_v128_shr_n_s16<10>),
+                      MAP(imm_v128_shr_n_s16<12>),
+                      MAP(imm_v128_shr_n_s16<14>),
+                      MAP(imm_v128_shl_n_32<1>),
+                      MAP(imm_v128_shl_n_32<4>),
+                      MAP(imm_v128_shl_n_32<8>),
+                      MAP(imm_v128_shl_n_32<12>),
+                      MAP(imm_v128_shl_n_32<16>),
+                      MAP(imm_v128_shl_n_32<20>),
+                      MAP(imm_v128_shl_n_32<24>),
+                      MAP(imm_v128_shl_n_32<28>),
+                      MAP(imm_v128_shr_n_u32<1>),
+                      MAP(imm_v128_shr_n_u32<4>),
+                      MAP(imm_v128_shr_n_u32<8>),
+                      MAP(imm_v128_shr_n_u32<12>),
+                      MAP(imm_v128_shr_n_u32<16>),
+                      MAP(imm_v128_shr_n_u32<20>),
+                      MAP(imm_v128_shr_n_u32<24>),
+                      MAP(imm_v128_shr_n_u32<28>),
+                      MAP(imm_v128_shr_n_s32<1>),
+                      MAP(imm_v128_shr_n_s32<4>),
+                      MAP(imm_v128_shr_n_s32<8>),
+                      MAP(imm_v128_shr_n_s32<12>),
+                      MAP(imm_v128_shr_n_s32<16>),
+                      MAP(imm_v128_shr_n_s32<20>),
+                      MAP(imm_v128_shr_n_s32<24>),
+                      MAP(imm_v128_shr_n_s32<28>),
+                      MAP(v128_from_v64),
+                      MAP(v128_zip_8),
+                      MAP(v128_zip_16),
+                      MAP(v128_zip_32),
+                      MAP(v128_mul_s16),
+                      MAP(v128_unpack_u8_s16),
+                      MAP(v128_unpack_s8_s16),
+                      MAP(v128_unpack_u16_s32),
+                      MAP(v128_unpack_s16_s32),
+                      MAP(v128_shl_8),
+                      MAP(v128_shr_u8),
+                      MAP(v128_shr_s8),
+                      MAP(v128_shl_16),
+                      MAP(v128_shr_u16),
+                      MAP(v128_shr_s16),
+                      MAP(v128_shl_32),
+                      MAP(v128_shr_u32),
+                      MAP(v128_shr_s32),
+                      MAP(v128_hadd_u8),
+                      MAP(v128_dotp_s16),
+                      MAP(v128_low_u32),
+                      MAP(v128_low_v64),
+                      MAP(v128_high_v64),
+                      MAP(v128_from_64),
+                      MAP(v128_from_32),
+                      MAP(v128_zero),
+                      MAP(v128_dup_8),
+                      MAP(v128_dup_16),
+                      MAP(v128_dup_32),
+                      MAP(v128_unpacklo_u8_s16),
+                      MAP(v128_unpackhi_u8_s16),
+                      MAP(v128_unpacklo_s8_s16),
+                      MAP(v128_unpackhi_s8_s16),
+                      MAP(u32_load_unaligned),
+                      MAP(u32_store_unaligned),
+                      MAP(v64_load_unaligned),
+                      MAP(v64_store_unaligned),
+                      MAP(v128_load_unaligned),
+                      MAP(v128_store_unaligned),
+                      { NULL, NULL, NULL } };
+#undef MAP
+
+// Map reference functions to machine tuned functions. Since the
+// functions depend on machine tuned types, the non-machine tuned
+// instantiations of the test can't refer to these functions directly,
+// so we refer to them by name and do the mapping here.
+void Map(const char *name, fptr *ref, fptr *simd) {
+  unsigned int i;
+  for (i = 0; m[i].name && strcmp(name, m[i].name); i++) {
+  }
+
+  *ref = m[i].ref;
+  *simd = m[i].simd;
+}
+
+// Used for printing errors in TestSimd1Arg and TestSimd2Args
+std::string Print(const uint8_t *a, int size) {
+  std::string text = "0x";
+  for (int i = 0; i < size; i++) {
+    const uint8_t c = a[!CONFIG_BIG_ENDIAN ? size - 1 - i : i];
+    // Same as snprintf(..., ..., "%02x", c)
+    text += (c >> 4) + '0' + ((c >> 4) > 9) * ('a' - '0' - 10);
+    text += (c & 15) + '0' + ((c & 15) > 9) * ('a' - '0' - 10);
+  }
+
+  return text;
+}
+
+// Used in TestSimd1Arg and TestSimd2Args to restrict argument ranges
+void SetMask(uint8_t *s, int size, uint32_t mask, uint32_t maskwidth) {
+  switch (maskwidth) {
+    case 0: {
+      break;
+    }
+    case 8: {
+      for (int i = 0; i < size; i++) s[i] &= mask;
+      break;
+    }
+    case 16: {
+      uint16_t *t = reinterpret_cast<uint16_t *>(s);
+      assert(!(reinterpret_cast<uintptr_t>(s) & 1));
+      for (int i = 0; i < size / 2; i++) t[i] &= mask;
+      break;
+    }
+    case 32: {
+      uint32_t *t = reinterpret_cast<uint32_t *>(s);
+      assert(!(reinterpret_cast<uintptr_t>(s) & 3));
+      for (int i = 0; i < size / 4; i++) t[i] &= mask;
+      break;
+    }
+    case 64: {
+      uint64_t *t = reinterpret_cast<uint64_t *>(s);
+      assert(!(reinterpret_cast<uintptr_t>(s) & 7));
+      for (int i = 0; i < size / 8; i++) t[i] &= mask;
+      break;
+    }
+    default: {
+      FAIL() << "Unsupported mask width";
+      break;
+    }
+  }
+}
+
+// We need some extra load/store functions
+void u64_store_aligned(void *p, uint64_t a) {
+  v64_store_aligned(p, v64_from_64(a));
+}
+void s32_store_aligned(void *p, int32_t a) {
+  u32_store_aligned(p, static_cast<uint32_t>(a));
+}
+void s64_store_aligned(void *p, int64_t a) {
+  v64_store_aligned(p, v64_from_64(static_cast<uint64_t>(a)));
+}
+
+void c_u64_store_aligned(void *p, uint64_t a) {
+  c_v64_store_aligned(p, c_v64_from_64(a));
+}
+
+void c_s32_store_aligned(void *p, int32_t a) {
+  c_u32_store_aligned(p, static_cast<uint32_t>(a));
+}
+
+void c_s64_store_aligned(void *p, int64_t a) {
+  c_v64_store_aligned(p, c_v64_from_64(static_cast<uint64_t>(a)));
+}
+
+uint64_t u64_load_aligned(const void *p) {
+  return v64_u64(v64_load_aligned(p));
+}
+uint16_t u16_load_aligned(const void *p) {
+  return *(reinterpret_cast<const uint16_t *>(p));
+}
+uint8_t u8_load_aligned(const void *p) {
+  return *(reinterpret_cast<const uint8_t *>(p));
+}
+
+uint64_t c_u64_load_aligned(const void *p) {
+  return c_v64_u64(c_v64_load_aligned(p));
+}
+uint16_t c_u16_load_aligned(const void *p) {
+  return *(reinterpret_cast<const uint16_t *>(p));
+}
+uint8_t c_u8_load_aligned(const void *p) {
+  return *(reinterpret_cast<const uint8_t *>(p));
+}
+
+// CompareSimd1Arg and CompareSimd2Args compare intrinsics taking 1 or
+// 2 arguments respectively with their corresponding C reference.
+// Ideally, the loads and stores should have gone into the template
+// parameter list, but v64 and v128 could be typedef'ed to the same
+// type (which is the case on x86) and then we can't instantiate both
+// v64 and v128, so the function return and argument types, including
+// the always differing types in the C equivalent are used instead.
+// The function arguments must be void pointers and then go through a
+// cast to avoid matching errors in the branches eliminated by the
+// typeid tests in the calling function.
+template <typename Ret, typename Arg, typename CRet, typename CArg>
+int CompareSimd1Arg(fptr store, fptr load, fptr simd, void *d, fptr c_store,
+                    fptr c_load, fptr c_simd, void *ref_d, const void *a) {
+  void (*const my_store)(void *, Ret) = (void (*const)(void *, Ret))store;
+  Arg (*const my_load)(const void *) = (Arg(*const)(const void *))load;
+  Ret (*const my_simd)(Arg) = (Ret(*const)(Arg))simd;
+  void (*const my_c_store)(void *, CRet) = (void (*const)(void *, CRet))c_store;
+  CArg (*const my_c_load)(const void *) = (CArg(*const)(const void *))c_load;
+  CRet (*const my_c_simd)(CArg) = (CRet(*const)(CArg))c_simd;
+
+  // Call reference and intrinsic
+  my_c_store(ref_d, my_c_simd(my_c_load(a)));
+  my_store(d, my_simd(my_load(a)));
+
+  // Compare results
+  return memcmp(ref_d, d, sizeof(CRet));
+}
+
+template <typename Ret, typename Arg1, typename Arg2, typename CRet,
+          typename CArg1, typename CArg2>
+int CompareSimd2Args(fptr store, fptr load1, fptr load2, fptr simd, void *d,
+                     fptr c_store, fptr c_load1, fptr c_load2, fptr c_simd,
+                     void *ref_d, const void *a, const void *b) {
+  void (*const my_store)(void *, Ret) = (void (*const)(void *, Ret))store;
+  Arg1 (*const my_load1)(const void *) = (Arg1(*const)(const void *))load1;
+  Arg2 (*const my_load2)(const void *) = (Arg2(*const)(const void *))load2;
+  Ret (*const my_simd)(Arg1, Arg2) = (Ret(*const)(Arg1, Arg2))simd;
+  void (*const my_c_store)(void *, CRet) = (void (*const)(void *, CRet))c_store;
+  CArg1 (*const my_c_load1)(const void *) =
+      (CArg1(*const)(const void *))c_load1;
+  CArg2 (*const my_c_load2)(const void *) =
+      (CArg2(*const)(const void *))c_load2;
+  CRet (*const my_c_simd)(CArg1, CArg2) = (CRet(*const)(CArg1, CArg2))c_simd;
+
+  // Call reference and intrinsic
+  my_c_store(ref_d, my_c_simd(my_c_load1(a), my_c_load2(b)));
+  my_store(d, my_simd(my_load1(a), my_load2(b)));
+
+  // Compare results
+  return memcmp(ref_d, d, sizeof(CRet));
+}
+
+}  // namespace
+
+template <typename CRet, typename CArg>
+void TestSimd1Arg(uint32_t iterations, uint32_t mask, uint32_t maskwidth,
+                  const char *name) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  fptr ref_simd;
+  fptr simd;
+  int error = 0;
+  DECLARE_ALIGNED(32, uint8_t, s[sizeof(CArg)]);
+  DECLARE_ALIGNED(32, uint8_t, d[sizeof(CRet)]);
+  DECLARE_ALIGNED(32, uint8_t, ref_d[sizeof(CRet)]);
+  memset(ref_d, 0, sizeof(ref_d));
+  memset(d, 0, sizeof(d));
+
+  Map(name, &ref_simd, &simd);
+  if (simd == NULL || ref_simd == NULL) {
+    FAIL() << "Internal error: Unknown intrinsic function " << name;
+  }
+  for (unsigned int count = 0;
+       count < iterations && !error && !testing::Test::HasFailure(); count++) {
+    for (unsigned int c = 0; c < sizeof(CArg); c++) s[c] = rnd.Rand8();
+
+    if (maskwidth) {
+      SetMask(s, sizeof(CArg), mask, maskwidth);
+    }
+
+    if (typeid(CRet) == typeid(c_v64) && typeid(CArg) == typeid(c_v64)) {
+      // V64_V64
+      error = CompareSimd1Arg<v64, v64, CRet, CArg>(
+          reinterpret_cast<fptr>(v64_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v64_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(c_v64) &&
+               typeid(CArg) == typeid(uint8_t)) {
+      // V64_U8
+      error = CompareSimd1Arg<v64, uint8_t, CRet, CArg>(
+          reinterpret_cast<fptr>(v64_store_aligned),
+          reinterpret_cast<fptr>(u8_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v64_store_aligned),
+          reinterpret_cast<fptr>(c_u8_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(c_v64) &&
+               typeid(CArg) == typeid(uint16_t)) {
+      // V64_U16
+      error = CompareSimd1Arg<v64, uint16_t, CRet, CArg>(
+          reinterpret_cast<fptr>(v64_store_aligned),
+          reinterpret_cast<fptr>(u16_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v64_store_aligned),
+          reinterpret_cast<fptr>(c_u16_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(c_v64) &&
+               typeid(CArg) == typeid(uint32_t)) {
+      // V64_U32
+      error = CompareSimd1Arg<v64, uint32_t, CRet, CArg>(
+          reinterpret_cast<fptr>(v64_store_aligned),
+          reinterpret_cast<fptr>(u32_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v64_store_aligned),
+          reinterpret_cast<fptr>(c_u32_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(uint64_t) &&
+               typeid(CArg) == typeid(c_v64)) {
+      // U64_V64
+      error = CompareSimd1Arg<uint64_t, v64, CRet, CArg>(
+          reinterpret_cast<fptr>(u64_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_u64_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(int64_t) &&
+               typeid(CArg) == typeid(c_v64)) {
+      // S64_V64
+      error = CompareSimd1Arg<int64_t, v64, CRet, CArg>(
+          reinterpret_cast<fptr>(s64_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_s64_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(uint32_t) &&
+               typeid(CArg) == typeid(c_v64)) {
+      // U32_V64
+      error = CompareSimd1Arg<uint32_t, v64, CRet, CArg>(
+          reinterpret_cast<fptr>(u32_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_u32_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(int32_t) &&
+               typeid(CArg) == typeid(c_v64)) {
+      // S32_V64
+      error = CompareSimd1Arg<int32_t, v64, CRet, CArg>(
+          reinterpret_cast<fptr>(s32_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_s32_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(uint32_t) &&
+               typeid(CArg) == typeid(c_v128)) {
+      // U32_V128
+      error = CompareSimd1Arg<uint32_t, v128, CRet, CArg>(
+          reinterpret_cast<fptr>(u32_store_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_u32_store_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(uint64_t) &&
+               typeid(CArg) == typeid(c_v128)) {
+      // U64_V128
+      error = CompareSimd1Arg<uint64_t, v128, CRet, CArg>(
+          reinterpret_cast<fptr>(u64_store_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_u64_store_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(c_v64) &&
+               typeid(CArg) == typeid(c_v128)) {
+      // V64_V128
+      error = CompareSimd1Arg<v64, v128, CRet, CArg>(
+          reinterpret_cast<fptr>(v64_store_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v64_store_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(c_v128) &&
+               typeid(CArg) == typeid(c_v128)) {
+      // V128_V128
+      error = CompareSimd1Arg<v128, v128, CRet, CArg>(
+          reinterpret_cast<fptr>(v128_store_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v128_store_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(c_v128) &&
+               typeid(CArg) == typeid(c_v64)) {
+      // V128_V64
+      error = CompareSimd1Arg<v128, v64, CRet, CArg>(
+          reinterpret_cast<fptr>(v128_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v128_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(c_v128) &&
+               typeid(CArg) == typeid(uint8_t)) {
+      // V128_U8
+      error = CompareSimd1Arg<v128, uint8_t, CRet, CArg>(
+          reinterpret_cast<fptr>(v128_store_aligned),
+          reinterpret_cast<fptr>(u8_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v128_store_aligned),
+          reinterpret_cast<fptr>(c_u8_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(c_v128) &&
+               typeid(CArg) == typeid(uint16_t)) {
+      // V128_U16
+      error = CompareSimd1Arg<v128, uint16_t, CRet, CArg>(
+          reinterpret_cast<fptr>(v128_store_aligned),
+          reinterpret_cast<fptr>(u16_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v128_store_aligned),
+          reinterpret_cast<fptr>(c_u16_load_aligned), ref_simd, ref_d, s);
+    } else if (typeid(CRet) == typeid(c_v128) &&
+               typeid(CArg) == typeid(uint32_t)) {
+      // V128_U32
+      error = CompareSimd1Arg<v128, uint32_t, CRet, CArg>(
+          reinterpret_cast<fptr>(v128_store_aligned),
+          reinterpret_cast<fptr>(u32_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v128_store_aligned),
+          reinterpret_cast<fptr>(c_u32_load_aligned), ref_simd, ref_d, s);
+    } else {
+      FAIL() << "Internal error: Unknown intrinsic function "
+             << typeid(CRet).name() << " " << name << "(" << typeid(CArg).name()
+             << ")";
+    }
+  }
+
+  EXPECT_EQ(0, error) << "Error: mismatch for " << name << "("
+                      << Print(s, sizeof(s)) << ") -> " << Print(d, sizeof(d))
+                      << " (simd), " << Print(ref_d, sizeof(ref_d)) << " (ref)";
+}
+
+template <typename CRet, typename CArg1, typename CArg2>
+void TestSimd2Args(uint32_t iterations, uint32_t mask, uint32_t maskwidth,
+                   const char *name) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  fptr ref_simd;
+  fptr simd;
+  int error = 0;
+  DECLARE_ALIGNED(32, uint8_t, s1[sizeof(CArg1)]);
+  DECLARE_ALIGNED(32, uint8_t, s2[sizeof(CArg2)]);
+  DECLARE_ALIGNED(32, uint8_t, d[sizeof(CRet)]);
+  DECLARE_ALIGNED(32, uint8_t, ref_d[sizeof(CRet)]);
+  memset(ref_d, 0, sizeof(ref_d));
+  memset(d, 0, sizeof(d));
+
+  Map(name, &ref_simd, &simd);
+  if (simd == NULL || ref_simd == NULL) {
+    FAIL() << "Internal error: Unknown intrinsic function " << name;
+  }
+
+  for (unsigned int count = 0;
+       count < iterations && !error && !testing::Test::HasFailure(); count++) {
+    for (unsigned int c = 0; c < sizeof(CArg1); c++) s1[c] = rnd.Rand8();
+
+    for (unsigned int c = 0; c < sizeof(CArg2); c++) s2[c] = rnd.Rand8();
+
+    if (maskwidth) SetMask(s2, sizeof(CArg2), mask, maskwidth);
+
+    if (typeid(CRet) == typeid(c_v64) && typeid(CArg1) == typeid(c_v64) &&
+        typeid(CArg2) == typeid(c_v64)) {
+      // V64_V64V64
+      error = CompareSimd2Args<v64, v64, v64, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(v64_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v64_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else if (typeid(CRet) == typeid(c_v64) &&
+               typeid(CArg1) == typeid(uint32_t) &&
+               typeid(CArg2) == typeid(uint32_t)) {
+      // V64_U32U32
+      error = CompareSimd2Args<v64, uint32_t, uint32_t, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(v64_store_aligned),
+          reinterpret_cast<fptr>(u32_load_aligned),
+          reinterpret_cast<fptr>(u32_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v64_store_aligned),
+          reinterpret_cast<fptr>(c_u32_load_aligned),
+          reinterpret_cast<fptr>(c_u32_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else if (typeid(CRet) == typeid(uint32_t) &&
+               typeid(CArg1) == typeid(c_v64) &&
+               typeid(CArg2) == typeid(c_v64)) {
+      // U32_V64V64
+      error = CompareSimd2Args<uint32_t, v64, v64, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(u32_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_u32_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else if (typeid(CRet) == typeid(int64_t) &&
+               typeid(CArg1) == typeid(c_v64) &&
+               typeid(CArg2) == typeid(c_v64)) {
+      // S64_V64V64
+      error = CompareSimd2Args<int64_t, v64, v64, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(s64_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_s64_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else if (typeid(CRet) == typeid(c_v64) &&
+               typeid(CArg1) == typeid(c_v64) &&
+               typeid(CArg2) == typeid(uint32_t)) {
+      // V64_V64U32
+      error = CompareSimd2Args<v64, v64, uint32_t, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(v64_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned),
+          reinterpret_cast<fptr>(u32_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v64_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned),
+          reinterpret_cast<fptr>(c_u32_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else if (typeid(CRet) == typeid(c_v128) &&
+               typeid(CArg1) == typeid(c_v128) &&
+               typeid(CArg2) == typeid(c_v128)) {
+      // V128_V128V128
+      error = CompareSimd2Args<v128, v128, v128, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(v128_store_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v128_store_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else if (typeid(CRet) == typeid(uint32_t) &&
+               typeid(CArg1) == typeid(c_v128) &&
+               typeid(CArg2) == typeid(c_v128)) {
+      // U32_V128V128
+      error = CompareSimd2Args<uint32_t, v128, v128, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(u32_store_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_u32_store_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else if (typeid(CRet) == typeid(int64_t) &&
+               typeid(CArg1) == typeid(c_v128) &&
+               typeid(CArg2) == typeid(c_v128)) {
+      // S64_V128V128
+      error = CompareSimd2Args<int64_t, v128, v128, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(s64_store_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_s64_store_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else if (typeid(CRet) == typeid(c_v128) &&
+               typeid(CArg1) == typeid(uint64_t) &&
+               typeid(CArg2) == typeid(uint64_t)) {
+      // V128_U64U64
+      error = CompareSimd2Args<v128, uint64_t, uint64_t, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(v128_store_aligned),
+          reinterpret_cast<fptr>(u64_load_aligned),
+          reinterpret_cast<fptr>(u64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v128_store_aligned),
+          reinterpret_cast<fptr>(c_u64_load_aligned),
+          reinterpret_cast<fptr>(c_u64_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else if (typeid(CRet) == typeid(c_v128) &&
+               typeid(CArg1) == typeid(c_v64) &&
+               typeid(CArg2) == typeid(c_v64)) {
+      // V128_V64V64
+      error = CompareSimd2Args<v128, v64, v64, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(v128_store_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned),
+          reinterpret_cast<fptr>(v64_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v128_store_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned),
+          reinterpret_cast<fptr>(c_v64_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else if (typeid(CRet) == typeid(c_v128) &&
+               typeid(CArg1) == typeid(c_v128) &&
+               typeid(CArg2) == typeid(uint32_t)) {
+      // V128_V128U32
+      error = CompareSimd2Args<v128, v128, uint32_t, CRet, CArg1, CArg2>(
+          reinterpret_cast<fptr>(v128_store_aligned),
+          reinterpret_cast<fptr>(v128_load_aligned),
+          reinterpret_cast<fptr>(u32_load_aligned), simd, d,
+          reinterpret_cast<fptr>(c_v128_store_aligned),
+          reinterpret_cast<fptr>(c_v128_load_aligned),
+          reinterpret_cast<fptr>(c_u32_load_aligned),
+          reinterpret_cast<fptr>(ref_simd), ref_d, s1, s2);
+    } else {
+      FAIL() << "Internal error: Unknown intrinsic function "
+             << typeid(CRet).name() << " " << name << "("
+             << typeid(CArg1).name() << ", " << typeid(CArg2).name() << ")";
+    }
+  }
+
+  EXPECT_EQ(0, error) << "Error: mismatch for " << name << "("
+                      << Print(s1, sizeof(s1)) << ", " << Print(s2, sizeof(s2))
+                      << ") -> " << Print(d, sizeof(d)) << " (simd), "
+                      << Print(ref_d, sizeof(ref_d)) << " (ref)";
+}
+
+// Instantiations to make the functions callable from another files
+template void TestSimd1Arg<c_v64, uint8_t>(uint32_t, uint32_t, uint32_t,
+                                           const char *);
+template void TestSimd1Arg<c_v64, uint16_t>(uint32_t, uint32_t, uint32_t,
+                                            const char *);
+template void TestSimd1Arg<c_v64, uint32_t>(uint32_t, uint32_t, uint32_t,
+                                            const char *);
+template void TestSimd1Arg<c_v64, c_v64>(uint32_t, uint32_t, uint32_t,
+                                         const char *);
+template void TestSimd1Arg<uint32_t, c_v64>(uint32_t, uint32_t, uint32_t,
+                                            const char *);
+template void TestSimd1Arg<int32_t, c_v64>(uint32_t, uint32_t, uint32_t,
+                                           const char *);
+template void TestSimd1Arg<uint64_t, c_v64>(uint32_t, uint32_t, uint32_t,
+                                            const char *);
+template void TestSimd1Arg<int64_t, c_v64>(uint32_t, uint32_t, uint32_t,
+                                           const char *);
+template void TestSimd2Args<c_v64, uint32_t, uint32_t>(uint32_t, uint32_t,
+                                                       uint32_t, const char *);
+template void TestSimd2Args<c_v64, c_v64, c_v64>(uint32_t, uint32_t, uint32_t,
+                                                 const char *);
+template void TestSimd2Args<c_v64, c_v64, uint32_t>(uint32_t, uint32_t,
+                                                    uint32_t, const char *);
+template void TestSimd2Args<int64_t, c_v64, c_v64>(uint32_t, uint32_t, uint32_t,
+                                                   const char *);
+template void TestSimd2Args<uint32_t, c_v64, c_v64>(uint32_t, uint32_t,
+                                                    uint32_t, const char *);
+template void TestSimd1Arg<c_v128, c_v128>(uint32_t, uint32_t, uint32_t,
+                                           const char *);
+template void TestSimd1Arg<c_v128, uint8_t>(uint32_t, uint32_t, uint32_t,
+                                            const char *);
+template void TestSimd1Arg<c_v128, uint16_t>(uint32_t, uint32_t, uint32_t,
+                                             const char *);
+template void TestSimd1Arg<c_v128, uint32_t>(uint32_t, uint32_t, uint32_t,
+                                             const char *);
+template void TestSimd1Arg<c_v128, c_v64>(uint32_t, uint32_t, uint32_t,
+                                          const char *);
+template void TestSimd1Arg<uint32_t, c_v128>(uint32_t, uint32_t, uint32_t,
+                                             const char *);
+template void TestSimd1Arg<uint64_t, c_v128>(uint32_t, uint32_t, uint32_t,
+                                             const char *);
+template void TestSimd1Arg<c_v64, c_v128>(uint32_t, uint32_t, uint32_t,
+                                          const char *);
+template void TestSimd2Args<c_v128, c_v128, c_v128>(uint32_t, uint32_t,
+                                                    uint32_t, const char *);
+template void TestSimd2Args<c_v128, c_v128, uint32_t>(uint32_t, uint32_t,
+                                                      uint32_t, const char *);
+template void TestSimd2Args<c_v128, uint64_t, uint64_t>(uint32_t, uint32_t,
+                                                        uint32_t, const char *);
+template void TestSimd2Args<c_v128, c_v64, c_v64>(uint32_t, uint32_t, uint32_t,
+                                                  const char *);
+template void TestSimd2Args<int64_t, c_v128, c_v128>(uint32_t, uint32_t,
+                                                     uint32_t, const char *);
+template void TestSimd2Args<uint32_t, c_v128, c_v128>(uint32_t, uint32_t,
+                                                      uint32_t, const char *);
+
+}  // namespace SIMD_NAMESPACE
diff --git a/third_party/aom/test/simd_cmp_neon.cc b/third_party/aom/test/simd_cmp_neon.cc
new file mode 100644
index 0000000000..c8004cc8b1
--- /dev/null
+++ b/third_party/aom/test/simd_cmp_neon.cc
@@ -0,0 +1,17 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if defined(__OPTIMIZE__) && __OPTIMIZE__
+#define ARCH NEON
+#define ARCH_POSTFIX(name) name##_neon
+#define SIMD_NAMESPACE simd_test_neon
+#include "./simd_cmp_impl.h"
+#endif
diff --git a/third_party/aom/test/simd_cmp_sse2.cc b/third_party/aom/test/simd_cmp_sse2.cc
new file mode 100644
index 0000000000..67cb43c101
--- /dev/null
+++ b/third_party/aom/test/simd_cmp_sse2.cc
@@ -0,0 +1,18 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if (defined(__OPTIMIZE__) && __OPTIMIZE__) || \
+    (!defined(__GNUC__) && !defined(_DEBUG))
+#define ARCH SSE2
+#define ARCH_POSTFIX(name) name##_sse2
+#define SIMD_NAMESPACE simd_test_sse2
+#include "./simd_cmp_impl.h"
+#endif
diff --git a/third_party/aom/test/simd_cmp_sse4.cc b/third_party/aom/test/simd_cmp_sse4.cc
new file mode 100644
index 0000000000..ba826d8983
--- /dev/null
+++ b/third_party/aom/test/simd_cmp_sse4.cc
@@ -0,0 +1,18 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if (defined(__OPTIMIZE__) && __OPTIMIZE__) || \
+    (!defined(__GNUC__) && !defined(_DEBUG))
+#define ARCH SSE4_1
+#define ARCH_POSTFIX(name) name##_sse4_1
+#define SIMD_NAMESPACE simd_test_sse4_1
+#include "./simd_cmp_impl.h"
+#endif
diff --git a/third_party/aom/test/simd_cmp_ssse3.cc b/third_party/aom/test/simd_cmp_ssse3.cc
new file mode 100644
index 0000000000..a6c7000fd3
--- /dev/null
+++ b/third_party/aom/test/simd_cmp_ssse3.cc
@@ -0,0 +1,18 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if (defined(__OPTIMIZE__) && __OPTIMIZE__) || \
+    (!defined(__GNUC__) && !defined(_DEBUG))
+#define ARCH SSSE3
+#define ARCH_POSTFIX(name) name##_ssse3
+#define SIMD_NAMESPACE simd_test_ssse3
+#include "./simd_cmp_impl.h"
+#endif
diff --git a/third_party/aom/test/simd_impl.h b/third_party/aom/test/simd_impl.h
new file mode 100644
index 0000000000..5cfda675d3
--- /dev/null
+++ b/third_party/aom/test/simd_impl.h
@@ -0,0 +1,594 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#define SIMD_CHECK 1
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "aom_dsp/aom_simd_inline.h"
+#include "aom_dsp/simd/v128_intrinsics_c.h"
+
+namespace SIMD_NAMESPACE {
+
+template <typename param_signature>
+class TestIntrinsic : public ::testing::TestWithParam<param_signature> {
+ public:
+  virtual ~TestIntrinsic() {}
+  virtual void SetUp() {
+    mask = std::tr1::get<0>(this->GetParam());
+    maskwidth = std::tr1::get<1>(this->GetParam());
+    name = std::tr1::get<2>(this->GetParam());
+  }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  uint32_t mask, maskwidth;
+  const char *name;
+};
+
+// Create one typedef for each function signature
+#define TYPEDEF_SIMD(name)                                                  \
+  typedef TestIntrinsic<std::tr1::tuple<uint32_t, uint32_t, const char *> > \
+      ARCH_POSTFIX(name)
+
+TYPEDEF_SIMD(V64_U8);
+TYPEDEF_SIMD(V64_U16);
+TYPEDEF_SIMD(V64_U32);
+TYPEDEF_SIMD(V64_V64);
+TYPEDEF_SIMD(U32_V64);
+TYPEDEF_SIMD(S32_V64);
+TYPEDEF_SIMD(U64_V64);
+TYPEDEF_SIMD(S64_V64);
+TYPEDEF_SIMD(V64_U32U32);
+TYPEDEF_SIMD(V64_V64V64);
+TYPEDEF_SIMD(S64_V64V64);
+TYPEDEF_SIMD(V64_V64U32);
+TYPEDEF_SIMD(U32_V64V64);
+TYPEDEF_SIMD(V128_V64);
+TYPEDEF_SIMD(V128_V128);
+TYPEDEF_SIMD(U32_V128);
+TYPEDEF_SIMD(U64_V128);
+TYPEDEF_SIMD(V64_V128);
+TYPEDEF_SIMD(V128_U8);
+TYPEDEF_SIMD(V128_U16);
+TYPEDEF_SIMD(V128_U32);
+TYPEDEF_SIMD(V128_U64U64);
+TYPEDEF_SIMD(V128_V64V64);
+TYPEDEF_SIMD(V128_V128V128);
+TYPEDEF_SIMD(S64_V128V128);
+TYPEDEF_SIMD(V128_V128U32);
+TYPEDEF_SIMD(U32_V128V128);
+
+// Google Test allows up to 50 tests per case, so split the largest
+typedef ARCH_POSTFIX(V64_V64) ARCH_POSTFIX(V64_V64_Part2);
+typedef ARCH_POSTFIX(V64_V64V64) ARCH_POSTFIX(V64_V64V64_Part2);
+typedef ARCH_POSTFIX(V128_V128) ARCH_POSTFIX(V128_V128_Part2);
+typedef ARCH_POSTFIX(V128_V128) ARCH_POSTFIX(V128_V128_Part3);
+typedef ARCH_POSTFIX(V128_V128V128) ARCH_POSTFIX(V128_V128V128_Part2);
+
+// These functions are machine tuned located elsewhere
+template <typename c_ret, typename c_arg>
+void TestSimd1Arg(uint32_t iterations, uint32_t mask, uint32_t maskwidth,
+                  const char *name);
+
+template <typename c_ret, typename c_arg1, typename c_arg2>
+void TestSimd2Args(uint32_t iterations, uint32_t mask, uint32_t maskwidth,
+                   const char *name);
+
+const int kIterations = 65536;
+
+// Add a macro layer since TEST_P will quote the name so we need to
+// expand it first with the prefix.
+#define MY_TEST_P(name, test) TEST_P(name, test)
+
+MY_TEST_P(ARCH_POSTFIX(V64_U8), TestIntrinsics) {
+  TestSimd1Arg<c_v64, uint8_t>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V64_U16), TestIntrinsics) {
+  TestSimd1Arg<c_v64, uint16_t>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V64_U32), TestIntrinsics) {
+  TestSimd1Arg<c_v64, uint32_t>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V64_V64), TestIntrinsics) {
+  TestSimd1Arg<c_v64, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(U64_V64), TestIntrinsics) {
+  TestSimd1Arg<uint64_t, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(S64_V64), TestIntrinsics) {
+  TestSimd1Arg<int64_t, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(U32_V64), TestIntrinsics) {
+  TestSimd1Arg<uint32_t, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(S32_V64), TestIntrinsics) {
+  TestSimd1Arg<int32_t, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V64_U32U32), TestIntrinsics) {
+  TestSimd2Args<c_v64, uint32_t, uint32_t>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V64_V64V64), TestIntrinsics) {
+  TestSimd2Args<c_v64, c_v64, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(S64_V64V64), TestIntrinsics) {
+  TestSimd2Args<int64_t, c_v64, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(U32_V64V64), TestIntrinsics) {
+  TestSimd2Args<uint32_t, c_v64, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V64_V64U32), TestIntrinsics) {
+  TestSimd2Args<c_v64, c_v64, uint32_t>(kIterations, mask, maskwidth, name);
+}
+
+// Google Test allows up to 50 tests per case, so split the largest
+MY_TEST_P(ARCH_POSTFIX(V64_V64_Part2), TestIntrinsics) {
+  TestSimd1Arg<c_v64, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V64_V64V64_Part2), TestIntrinsics) {
+  TestSimd2Args<c_v64, c_v64, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(U32_V128), TestIntrinsics) {
+  TestSimd1Arg<uint32_t, c_v128>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(U64_V128), TestIntrinsics) {
+  TestSimd1Arg<uint64_t, c_v128>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V64_V128), TestIntrinsics) {
+  TestSimd1Arg<c_v64, c_v128>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_V128), TestIntrinsics) {
+  TestSimd1Arg<c_v128, c_v128>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_U8), TestIntrinsics) {
+  TestSimd1Arg<c_v128, uint8_t>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_U16), TestIntrinsics) {
+  TestSimd1Arg<c_v128, uint16_t>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_U32), TestIntrinsics) {
+  TestSimd1Arg<c_v128, uint32_t>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_V64), TestIntrinsics) {
+  TestSimd1Arg<c_v128, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_V128V128), TestIntrinsics) {
+  TestSimd2Args<c_v128, c_v128, c_v128>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(U32_V128V128), TestIntrinsics) {
+  TestSimd2Args<uint32_t, c_v128, c_v128>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(S64_V128V128), TestIntrinsics) {
+  TestSimd2Args<int64_t, c_v128, c_v128>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_U64U64), TestIntrinsics) {
+  TestSimd2Args<c_v128, uint64_t, uint64_t>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_V64V64), TestIntrinsics) {
+  TestSimd2Args<c_v128, c_v64, c_v64>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_V128U32), TestIntrinsics) {
+  TestSimd2Args<c_v128, c_v128, uint32_t>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_V128V128_Part2), TestIntrinsics) {
+  TestSimd2Args<c_v128, c_v128, c_v128>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_V128_Part2), TestIntrinsics) {
+  TestSimd1Arg<c_v128, c_v128>(kIterations, mask, maskwidth, name);
+}
+
+MY_TEST_P(ARCH_POSTFIX(V128_V128_Part3), TestIntrinsics) {
+  TestSimd1Arg<c_v128, c_v128>(kIterations, mask, maskwidth, name);
+}
+
+// Add a macro layer since INSTANTIATE_TEST_CASE_P will quote the name
+// so we need to expand it first with the prefix
+#define INSTANTIATE(name, type, ...) \
+  INSTANTIATE_TEST_CASE_P(name, type, ::testing::Values(__VA_ARGS__))
+
+#define SIMD_TUPLE(name, mask, maskwidth) \
+  std::tr1::make_tuple(mask, maskwidth, static_cast<const char *>(#name))
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(U32_V64V64),
+            (SIMD_TUPLE(v64_sad_u8, 0U, 0U), SIMD_TUPLE(v64_ssd_u8, 0U, 0U)));
+
+INSTANTIATE(
+    ARCH, ARCH_POSTFIX(V64_V64V64), SIMD_TUPLE(v64_add_8, 0U, 0U),
+    SIMD_TUPLE(v64_add_16, 0U, 0U), SIMD_TUPLE(v64_sadd_s16, 0U, 0U),
+    SIMD_TUPLE(v64_add_32, 0U, 0U), SIMD_TUPLE(v64_sub_8, 0U, 0U),
+    SIMD_TUPLE(v64_ssub_u8, 0U, 0U), SIMD_TUPLE(v64_ssub_s8, 0U, 0U),
+    SIMD_TUPLE(v64_sub_16, 0U, 0U), SIMD_TUPLE(v64_ssub_s16, 0U, 0U),
+    SIMD_TUPLE(v64_ssub_u16, 0U, 0U), SIMD_TUPLE(v64_sub_32, 0U, 0U),
+    SIMD_TUPLE(v64_ziplo_8, 0U, 0U), SIMD_TUPLE(v64_ziphi_8, 0U, 0U),
+    SIMD_TUPLE(v64_ziplo_16, 0U, 0U), SIMD_TUPLE(v64_ziphi_16, 0U, 0U),
+    SIMD_TUPLE(v64_ziplo_32, 0U, 0U), SIMD_TUPLE(v64_ziphi_32, 0U, 0U),
+    SIMD_TUPLE(v64_pack_s32_s16, 0U, 0U), SIMD_TUPLE(v64_pack_s16_u8, 0U, 0U),
+    SIMD_TUPLE(v64_pack_s16_s8, 0U, 0U), SIMD_TUPLE(v64_unziphi_8, 0U, 0U),
+    SIMD_TUPLE(v64_unziplo_8, 0U, 0U), SIMD_TUPLE(v64_unziphi_16, 0U, 0U),
+    SIMD_TUPLE(v64_unziplo_16, 0U, 0U), SIMD_TUPLE(v64_or, 0U, 0U),
+    SIMD_TUPLE(v64_xor, 0U, 0U), SIMD_TUPLE(v64_and, 0U, 0U),
+    SIMD_TUPLE(v64_andn, 0U, 0U), SIMD_TUPLE(v64_mullo_s16, 0U, 0U),
+    SIMD_TUPLE(v64_mulhi_s16, 0U, 0U), SIMD_TUPLE(v64_mullo_s32, 0U, 0U),
+    SIMD_TUPLE(v64_madd_s16, 0U, 0U), SIMD_TUPLE(v64_madd_us8, 0U, 0U),
+    SIMD_TUPLE(v64_avg_u8, 0U, 0U), SIMD_TUPLE(v64_rdavg_u8, 0U, 0U),
+    SIMD_TUPLE(v64_avg_u16, 0U, 0U), SIMD_TUPLE(v64_min_u8, 0U, 0U),
+    SIMD_TUPLE(v64_max_u8, 0U, 0U), SIMD_TUPLE(v64_min_s8, 0U, 0U),
+    SIMD_TUPLE(v64_max_s8, 0U, 0U), SIMD_TUPLE(v64_min_s16, 0U, 0U),
+    SIMD_TUPLE(v64_max_s16, 0U, 0U), SIMD_TUPLE(v64_cmpgt_s8, 0U, 0U),
+    SIMD_TUPLE(v64_cmplt_s8, 0U, 0U), SIMD_TUPLE(v64_cmpeq_8, 0U, 0U),
+    SIMD_TUPLE(v64_cmpgt_s16, 0U, 0U), SIMD_TUPLE(v64_cmplt_s16, 0U, 0U),
+    SIMD_TUPLE(v64_cmpeq_16, 0U, 0U));
+
+INSTANTIATE(
+    ARCH, ARCH_POSTFIX(V64_V64V64_Part2), SIMD_TUPLE(v64_shuffle_8, 7U, 8U),
+    SIMD_TUPLE(imm_v64_align<1>, 0U, 0U), SIMD_TUPLE(imm_v64_align<2>, 0U, 0U),
+    SIMD_TUPLE(imm_v64_align<3>, 0U, 0U), SIMD_TUPLE(imm_v64_align<4>, 0U, 0U),
+    SIMD_TUPLE(imm_v64_align<5>, 0U, 0U), SIMD_TUPLE(imm_v64_align<6>, 0U, 0U),
+    SIMD_TUPLE(imm_v64_align<7>, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V64_V64), SIMD_TUPLE(v64_abs_s8, 0U, 0U),
+            SIMD_TUPLE(v64_abs_s16, 0U, 0U),
+            SIMD_TUPLE(v64_unpacklo_u8_s16, 0U, 0U),
+            SIMD_TUPLE(v64_unpackhi_u8_s16, 0U, 0U),
+            SIMD_TUPLE(v64_unpacklo_s8_s16, 0U, 0U),
+            SIMD_TUPLE(v64_unpackhi_s8_s16, 0U, 0U),
+            SIMD_TUPLE(v64_unpacklo_u16_s32, 0U, 0U),
+            SIMD_TUPLE(v64_unpacklo_s16_s32, 0U, 0U),
+            SIMD_TUPLE(v64_unpackhi_u16_s32, 0U, 0U),
+            SIMD_TUPLE(v64_unpackhi_s16_s32, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_byte<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_byte<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_byte<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_byte<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_byte<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_byte<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_byte<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_byte<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_byte<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_byte<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_byte<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_byte<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_byte<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_byte<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_8<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_8<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_8<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_8<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_8<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_8<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_8<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u8<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u8<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u8<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u8<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u8<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u8<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u8<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s8<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s8<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s8<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s8<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s8<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s8<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s8<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_16<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_16<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_16<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_16<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_16<8>, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V64_V64_Part2),
+            SIMD_TUPLE(imm_v64_shl_n_16<10>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_16<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_16<14>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u16<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u16<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u16<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u16<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u16<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u16<10>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u16<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u16<14>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s16<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s16<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s16<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s16<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s16<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s16<10>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s16<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s16<14>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_32<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_32<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_32<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_32<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_32<16>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_32<20>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_32<24>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shl_n_32<28>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u32<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u32<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u32<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u32<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u32<16>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u32<20>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u32<24>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_u32<28>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s32<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s32<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s32<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s32<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s32<16>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s32<20>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s32<24>, 0U, 0U),
+            SIMD_TUPLE(imm_v64_shr_n_s32<28>, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V64_V64U32), SIMD_TUPLE(v64_shl_8, 7U, 32U),
+            SIMD_TUPLE(v64_shr_u8, 7U, 32U), SIMD_TUPLE(v64_shr_s8, 7U, 32U),
+            SIMD_TUPLE(v64_shl_16, 15U, 32U), SIMD_TUPLE(v64_shr_u16, 15U, 32U),
+            SIMD_TUPLE(v64_shr_s16, 15U, 32U), SIMD_TUPLE(v64_shl_32, 31U, 32U),
+            SIMD_TUPLE(v64_shr_u32, 31U, 32U),
+            SIMD_TUPLE(v64_shr_s32, 31U, 32U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(U64_V64), SIMD_TUPLE(v64_hadd_u8, 0U, 0U),
+            SIMD_TUPLE(v64_u64, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(S64_V64), SIMD_TUPLE(v64_hadd_s16, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(U32_V64), SIMD_TUPLE(v64_low_u32, 0U, 0U),
+            SIMD_TUPLE(v64_high_u32, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(S32_V64), SIMD_TUPLE(v64_low_s32, 0U, 0U),
+            SIMD_TUPLE(v64_high_s32, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(S64_V64V64), SIMD_TUPLE(v64_dotp_s16, 0U, 0U),
+            SIMD_TUPLE(v64_dotp_su8, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V64_U8), SIMD_TUPLE(v64_dup_8, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V64_U16), SIMD_TUPLE(v64_dup_16, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V64_U32), SIMD_TUPLE(v64_dup_32, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V64_U32U32), SIMD_TUPLE(v64_from_32, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(U32_V128V128), SIMD_TUPLE(v128_sad_u8, 0U, 0U),
+            SIMD_TUPLE(v128_ssd_u8, 0U, 0U));
+
+INSTANTIATE(
+    ARCH, ARCH_POSTFIX(V128_V128V128), SIMD_TUPLE(v128_add_8, 0U, 0U),
+    SIMD_TUPLE(v128_add_16, 0U, 0U), SIMD_TUPLE(v128_sadd_s16, 0U, 0U),
+    SIMD_TUPLE(v128_add_32, 0U, 0U), SIMD_TUPLE(v128_sub_8, 0U, 0U),
+    SIMD_TUPLE(v128_ssub_u8, 0U, 0U), SIMD_TUPLE(v128_ssub_s8, 0U, 0U),
+    SIMD_TUPLE(v128_sub_16, 0U, 0U), SIMD_TUPLE(v128_ssub_s16, 0U, 0U),
+    SIMD_TUPLE(v128_ssub_u16, 0U, 0U), SIMD_TUPLE(v128_sub_32, 0U, 0U),
+    SIMD_TUPLE(v128_ziplo_8, 0U, 0U), SIMD_TUPLE(v128_ziphi_8, 0U, 0U),
+    SIMD_TUPLE(v128_ziplo_16, 0U, 0U), SIMD_TUPLE(v128_ziphi_16, 0U, 0U),
+    SIMD_TUPLE(v128_ziplo_32, 0U, 0U), SIMD_TUPLE(v128_ziphi_32, 0U, 0U),
+    SIMD_TUPLE(v128_ziplo_64, 0U, 0U), SIMD_TUPLE(v128_ziphi_64, 0U, 0U),
+    SIMD_TUPLE(v128_unziphi_8, 0U, 0U), SIMD_TUPLE(v128_unziplo_8, 0U, 0U),
+    SIMD_TUPLE(v128_unziphi_16, 0U, 0U), SIMD_TUPLE(v128_unziplo_16, 0U, 0U),
+    SIMD_TUPLE(v128_unziphi_32, 0U, 0U), SIMD_TUPLE(v128_unziplo_32, 0U, 0U),
+    SIMD_TUPLE(v128_pack_s32_s16, 0U, 0U), SIMD_TUPLE(v128_pack_s16_u8, 0U, 0U),
+    SIMD_TUPLE(v128_pack_s16_s8, 0U, 0U), SIMD_TUPLE(v128_or, 0U, 0U),
+    SIMD_TUPLE(v128_xor, 0U, 0U), SIMD_TUPLE(v128_and, 0U, 0U),
+    SIMD_TUPLE(v128_andn, 0U, 0U), SIMD_TUPLE(v128_mullo_s16, 0U, 0U),
+    SIMD_TUPLE(v128_mulhi_s16, 0U, 0U), SIMD_TUPLE(v128_mullo_s32, 0U, 0U),
+    SIMD_TUPLE(v128_madd_s16, 0U, 0U), SIMD_TUPLE(v128_madd_us8, 0U, 0U),
+    SIMD_TUPLE(v128_avg_u8, 0U, 0U), SIMD_TUPLE(v128_rdavg_u8, 0U, 0U),
+    SIMD_TUPLE(v128_avg_u16, 0U, 0U), SIMD_TUPLE(v128_min_u8, 0U, 0U),
+    SIMD_TUPLE(v128_max_u8, 0U, 0U), SIMD_TUPLE(v128_min_s8, 0U, 0U),
+    SIMD_TUPLE(v128_max_s8, 0U, 0U), SIMD_TUPLE(v128_min_s16, 0U, 0U),
+    SIMD_TUPLE(v128_max_s16, 0U, 0U), SIMD_TUPLE(v128_cmpgt_s8, 0U, 0U),
+    SIMD_TUPLE(v128_cmplt_s8, 0U, 0U), SIMD_TUPLE(v128_cmpeq_8, 0U, 0U),
+    SIMD_TUPLE(v128_cmpgt_s16, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128V128_Part2),
+            SIMD_TUPLE(v128_cmpeq_16, 0U, 0U),
+            SIMD_TUPLE(v128_cmplt_s16, 0U, 0U),
+            SIMD_TUPLE(v128_shuffle_8, 15U, 8U),
+            SIMD_TUPLE(imm_v128_align<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<9>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<10>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<11>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<13>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<14>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_align<15>, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128), SIMD_TUPLE(v128_abs_s8, 0U, 0U),
+            SIMD_TUPLE(v128_abs_s16, 0U, 0U), SIMD_TUPLE(v128_padd_s16, 0U, 0U),
+            SIMD_TUPLE(v128_unpacklo_u8_s16, 0U, 0U),
+            SIMD_TUPLE(v128_unpacklo_s8_s16, 0U, 0U),
+            SIMD_TUPLE(v128_unpacklo_u16_s32, 0U, 0U),
+            SIMD_TUPLE(v128_unpacklo_s16_s32, 0U, 0U),
+            SIMD_TUPLE(v128_unpackhi_u8_s16, 0U, 0U),
+            SIMD_TUPLE(v128_unpackhi_s8_s16, 0U, 0U),
+            SIMD_TUPLE(v128_unpackhi_u16_s32, 0U, 0U),
+            SIMD_TUPLE(v128_unpackhi_s16_s32, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<9>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<10>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<11>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<13>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<14>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_byte<15>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<9>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<10>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<11>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<13>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<14>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_byte<15>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_8<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_8<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_8<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_8<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_8<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_8<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_8<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u8<1>, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128_Part2),
+            SIMD_TUPLE(imm_v128_shr_n_u8<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u8<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u8<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u8<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u8<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u8<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s8<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s8<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s8<3>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s8<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s8<5>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s8<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s8<7>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_16<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_16<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_16<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_16<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_16<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_16<10>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_16<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_16<14>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u16<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u16<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u16<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u16<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u16<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u16<10>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u16<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u16<14>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s16<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s16<2>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s16<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s16<6>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s16<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s16<10>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s16<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s16<14>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_32<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_32<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_32<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_32<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_32<16>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_32<20>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_32<24>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shl_n_32<28>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u32<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u32<4>, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128_Part3),
+            SIMD_TUPLE(imm_v128_shr_n_u32<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u32<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u32<16>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u32<20>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u32<24>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_u32<28>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s32<1>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s32<4>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s32<8>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s32<12>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s32<16>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s32<20>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s32<24>, 0U, 0U),
+            SIMD_TUPLE(imm_v128_shr_n_s32<28>, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V64V64), SIMD_TUPLE(v128_from_v64, 0U, 0U),
+            SIMD_TUPLE(v128_zip_8, 0U, 0U), SIMD_TUPLE(v128_zip_16, 0U, 0U),
+            SIMD_TUPLE(v128_zip_32, 0U, 0U), SIMD_TUPLE(v128_mul_s16, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_U64U64), SIMD_TUPLE(v128_from_64, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V64),
+            SIMD_TUPLE(v128_unpack_u8_s16, 0U, 0U),
+            SIMD_TUPLE(v128_unpack_s8_s16, 0U, 0U),
+            SIMD_TUPLE(v128_unpack_u16_s32, 0U, 0U),
+            SIMD_TUPLE(v128_unpack_s16_s32, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_V128U32), SIMD_TUPLE(v128_shl_8, 7U, 32U),
+            SIMD_TUPLE(v128_shr_u8, 7U, 32U), SIMD_TUPLE(v128_shr_s8, 7U, 32U),
+            SIMD_TUPLE(v128_shl_16, 15U, 32U),
+            SIMD_TUPLE(v128_shr_u16, 15U, 32U),
+            SIMD_TUPLE(v128_shr_s16, 15U, 32U),
+            SIMD_TUPLE(v128_shl_32, 31U, 32U),
+            SIMD_TUPLE(v128_shr_u32, 31U, 32U),
+            SIMD_TUPLE(v128_shr_s32, 31U, 32U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(U32_V128), SIMD_TUPLE(v128_low_u32, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(U64_V128), SIMD_TUPLE(v128_hadd_u8, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V64_V128), SIMD_TUPLE(v128_low_v64, 0U, 0U),
+            SIMD_TUPLE(v128_high_v64, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_U8), SIMD_TUPLE(v128_dup_8, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_U16), SIMD_TUPLE(v128_dup_16, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(V128_U32), SIMD_TUPLE(v128_dup_32, 0U, 0U));
+
+INSTANTIATE(ARCH, ARCH_POSTFIX(S64_V128V128),
+            SIMD_TUPLE(v128_dotp_s16, 0U, 0U));
+
+}  // namespace SIMD_NAMESPACE
diff --git a/third_party/aom/test/simd_neon_test.cc b/third_party/aom/test/simd_neon_test.cc
new file mode 100644
index 0000000000..0565fb4e2a
--- /dev/null
+++ b/third_party/aom/test/simd_neon_test.cc
@@ -0,0 +1,17 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if defined(__OPTIMIZE__) && __OPTIMIZE__
+#define ARCH NEON
+#define ARCH_POSTFIX(name) name##_neon
+#define SIMD_NAMESPACE simd_test_neon
+#include "./simd_impl.h"
+#endif
diff --git a/third_party/aom/test/simd_sse2_test.cc b/third_party/aom/test/simd_sse2_test.cc
new file mode 100644
index 0000000000..a0b49d77e6
--- /dev/null
+++ b/third_party/aom/test/simd_sse2_test.cc
@@ -0,0 +1,18 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if (defined(__OPTIMIZE__) && __OPTIMIZE__) || \
+    (!defined(__GNUC__) && !defined(_DEBUG))
+#define ARCH SSE2
+#define ARCH_POSTFIX(name) name##_sse2
+#define SIMD_NAMESPACE simd_test_sse2
+#include "./simd_impl.h"
+#endif
diff --git a/third_party/aom/test/simd_sse4_test.cc b/third_party/aom/test/simd_sse4_test.cc
new file mode 100644
index 0000000000..73c96427f2
--- /dev/null
+++ b/third_party/aom/test/simd_sse4_test.cc
@@ -0,0 +1,18 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if (defined(__OPTIMIZE__) && __OPTIMIZE__) || \
+    (!defined(__GNUC__) && !defined(_DEBUG))
+#define ARCH SSE4_1
+#define ARCH_POSTFIX(name) name##_sse4_1
+#define SIMD_NAMESPACE simd_test_sse4_1
+#include "./simd_impl.h"
+#endif
diff --git a/third_party/aom/test/simd_ssse3_test.cc b/third_party/aom/test/simd_ssse3_test.cc
new file mode 100644
index 0000000000..9ebeeef1b5
--- /dev/null
+++ b/third_party/aom/test/simd_ssse3_test.cc
@@ -0,0 +1,18 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#if (defined(__OPTIMIZE__) && __OPTIMIZE__) || \
+    (!defined(__GNUC__) && !defined(_DEBUG))
+#define ARCH SSSE3
+#define ARCH_POSTFIX(name) name##_ssse3
+#define SIMD_NAMESPACE simd_test_ssse3
+#include "./simd_impl.h"
+#endif
diff --git a/third_party/aom/test/simple_decoder.sh b/third_party/aom/test/simple_decoder.sh
new file mode 100755
index 0000000000..ac3a07b189
--- /dev/null
+++ b/third_party/aom/test/simple_decoder.sh
@@ -0,0 +1,58 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This file tests the libaom simple_decoder example code. To add new tests to
+## this file, do the following:
+##   1. Write a shell function (this is your test).
+##   2. Add the function to simple_decoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: Make sure input is available:
+simple_decoder_verify_environment() {
+  if [ ! "$(av1_encode_available)" = "yes" ] && [ ! -e "${AV1_IVF_FILE}" ]; then
+    return 1
+  fi
+}
+
+# Runs simple_decoder using $1 as input file. $2 is the codec name, and is used
+# solely to name the output file.
+simple_decoder() {
+  local decoder="${LIBAOM_BIN_PATH}/simple_decoder${AOM_TEST_EXE_SUFFIX}"
+  local input_file="$1"
+  local codec="$2"
+  local output_file="${AOM_TEST_OUTPUT_DIR}/simple_decoder_${codec}.raw"
+
+  if [ ! -x "${decoder}" ]; then
+    elog "${decoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${AOM_TEST_PREFIX}" "${decoder}" "${input_file}" "${output_file}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+simple_decoder_av1() {
+  if [ "$(av1_decode_available)" = "yes" ]; then
+    if [ ! -e "${AV1_IVF_FILE}" ]; then
+      local file="${AOM_TEST_OUTPUT_DIR}/test_encode.ivf"
+      encode_yuv_raw_input_av1 "${file}" --ivf
+      simple_decoder "${file}" av1 || return 1
+    else
+      simple_decoder "${AV1_IVF_FILE}" av1 || return 1
+    fi
+  fi
+}
+
+simple_decoder_tests="simple_decoder_av1"
+
+run_tests simple_decoder_verify_environment "${simple_decoder_tests}"
diff --git a/third_party/aom/test/simple_encoder.sh b/third_party/aom/test/simple_encoder.sh
new file mode 100755
index 0000000000..5cd6b46a10
--- /dev/null
+++ b/third_party/aom/test/simple_encoder.sh
@@ -0,0 +1,53 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This file tests the libaom simple_encoder example. To add new tests to this
+## file, do the following:
+##   1. Write a shell function (this is your test).
+##   2. Add the function to simple_encoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+simple_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libaom test data must exist in LIBAOM_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs simple_encoder using the codec specified by $1 with a frame limit of 100.
+simple_encoder() {
+  local encoder="${LIBAOM_BIN_PATH}/simple_encoder${AOM_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file="${AOM_TEST_OUTPUT_DIR}/simple_encoder_${codec}.ivf"
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${AOM_TEST_PREFIX}" "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" 9999 0 5 \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+
+simple_encoder_av1() {
+  if [ "$(av1_encode_available)" = "yes" ]; then
+    simple_encoder av1 || return 1
+  fi
+}
+
+simple_encoder_tests="simple_encoder_av1"
+
+run_tests simple_encoder_verify_environment "${simple_encoder_tests}"
diff --git a/third_party/aom/test/subtract_test.cc b/third_party/aom/test/subtract_test.cc
new file mode 100644
index 0000000000..c90ca8d56c
--- /dev/null
+++ b/third_party/aom/test/subtract_test.cc
@@ -0,0 +1,252 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#if CONFIG_AV1
+#include "av1/common/blockd.h"
+#endif
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+#define USE_SPEED_TEST (0)
+
+typedef void (*SubtractFunc)(int rows, int cols, int16_t *diff_ptr,
+                             ptrdiff_t diff_stride, const uint8_t *src_ptr,
+                             ptrdiff_t src_stride, const uint8_t *pred_ptr,
+                             ptrdiff_t pred_stride);
+
+namespace {
+
+class AV1SubtractBlockTest : public ::testing::TestWithParam<SubtractFunc> {
+ public:
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+};
+
+using libaom_test::ACMRandom;
+
+TEST_P(AV1SubtractBlockTest, SimpleSubtract) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+
+  // FIXME(rbultje) split in its own file
+  for (BLOCK_SIZE bsize = BLOCK_4X4; bsize < BLOCK_SIZES;
+       bsize = static_cast<BLOCK_SIZE>(static_cast<int>(bsize) + 1)) {
+    const int block_width = block_size_wide[bsize];
+    const int block_height = block_size_high[bsize];
+    int16_t *diff = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(*diff) * block_width * block_height * 2));
+    uint8_t *pred = reinterpret_cast<uint8_t *>(
+        aom_memalign(16, block_width * block_height * 2));
+    uint8_t *src = reinterpret_cast<uint8_t *>(
+        aom_memalign(16, block_width * block_height * 2));
+
+    for (int n = 0; n < 100; n++) {
+      for (int r = 0; r < block_height; ++r) {
+        for (int c = 0; c < block_width * 2; ++c) {
+          src[r * block_width * 2 + c] = rnd.Rand8();
+          pred[r * block_width * 2 + c] = rnd.Rand8();
+        }
+      }
+
+      GetParam()(block_height, block_width, diff, block_width, src, block_width,
+                 pred, block_width);
+
+      for (int r = 0; r < block_height; ++r) {
+        for (int c = 0; c < block_width; ++c) {
+          EXPECT_EQ(diff[r * block_width + c],
+                    (src[r * block_width + c] - pred[r * block_width + c]))
+              << "r = " << r << ", c = " << c << ", bs = " << bsize;
+        }
+      }
+
+      GetParam()(block_height, block_width, diff, block_width * 2, src,
+                 block_width * 2, pred, block_width * 2);
+
+      for (int r = 0; r < block_height; ++r) {
+        for (int c = 0; c < block_width; ++c) {
+          EXPECT_EQ(
+              diff[r * block_width * 2 + c],
+              (src[r * block_width * 2 + c] - pred[r * block_width * 2 + c]))
+              << "r = " << r << ", c = " << c << ", bs = " << bsize;
+        }
+      }
+    }
+    aom_free(diff);
+    aom_free(pred);
+    aom_free(src);
+  }
+}
+
+INSTANTIATE_TEST_CASE_P(C, AV1SubtractBlockTest,
+                        ::testing::Values(aom_subtract_block_c));
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, AV1SubtractBlockTest,
+                        ::testing::Values(aom_subtract_block_sse2));
+#endif
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, AV1SubtractBlockTest,
+                        ::testing::Values(aom_subtract_block_neon));
+#endif
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(MSA, AV1SubtractBlockTest,
+                        ::testing::Values(aom_subtract_block_msa));
+#endif
+
+typedef void (*HBDSubtractFunc)(int rows, int cols, int16_t *diff_ptr,
+                                ptrdiff_t diff_stride, const uint8_t *src_ptr,
+                                ptrdiff_t src_stride, const uint8_t *pred_ptr,
+                                ptrdiff_t pred_stride, int bd);
+
+using ::std::tr1::get;
+using ::std::tr1::make_tuple;
+using ::std::tr1::tuple;
+
+// <width, height, bit_dpeth, subtract>
+typedef tuple<int, int, int, HBDSubtractFunc> Params;
+
+#if CONFIG_HIGHBITDEPTH
+class AV1HBDSubtractBlockTest : public ::testing::TestWithParam<Params> {
+ public:
+  virtual void SetUp() {
+    block_width_ = GET_PARAM(0);
+    block_height_ = GET_PARAM(1);
+    bit_depth_ = static_cast<aom_bit_depth_t>(GET_PARAM(2));
+    func_ = GET_PARAM(3);
+
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+
+    const size_t max_width = 128;
+    const size_t max_block_size = max_width * max_width;
+    src_ = CONVERT_TO_BYTEPTR(reinterpret_cast<uint16_t *>(
+        aom_memalign(16, max_block_size * sizeof(uint16_t))));
+    pred_ = CONVERT_TO_BYTEPTR(reinterpret_cast<uint16_t *>(
+        aom_memalign(16, max_block_size * sizeof(uint16_t))));
+    diff_ = reinterpret_cast<int16_t *>(
+        aom_memalign(16, max_block_size * sizeof(int16_t)));
+  }
+
+  virtual void TearDown() {
+    aom_free(CONVERT_TO_SHORTPTR(src_));
+    aom_free(CONVERT_TO_SHORTPTR(pred_));
+    aom_free(diff_);
+  }
+
+ protected:
+  void RunForSpeed();
+  void CheckResult();
+
+ private:
+  ACMRandom rnd_;
+  int block_height_;
+  int block_width_;
+  aom_bit_depth_t bit_depth_;
+  HBDSubtractFunc func_;
+  uint8_t *src_;
+  uint8_t *pred_;
+  int16_t *diff_;
+};
+
+void AV1HBDSubtractBlockTest::RunForSpeed() {
+  const int test_num = 200000;
+  const int max_width = 128;
+  const int max_block_size = max_width * max_width;
+  const int mask = (1 << bit_depth_) - 1;
+  int i, j;
+
+  for (j = 0; j < max_block_size; ++j) {
+    CONVERT_TO_SHORTPTR(src_)[j] = rnd_.Rand16() & mask;
+    CONVERT_TO_SHORTPTR(pred_)[j] = rnd_.Rand16() & mask;
+  }
+
+  for (i = 0; i < test_num; ++i) {
+    func_(block_height_, block_width_, diff_, block_width_, src_, block_width_,
+          pred_, block_width_, bit_depth_);
+  }
+}
+
+void AV1HBDSubtractBlockTest::CheckResult() {
+  const int test_num = 100;
+  const int max_width = 128;
+  const int max_block_size = max_width * max_width;
+  const int mask = (1 << bit_depth_) - 1;
+  int i, j;
+
+  for (i = 0; i < test_num; ++i) {
+    for (j = 0; j < max_block_size; ++j) {
+      CONVERT_TO_SHORTPTR(src_)[j] = rnd_.Rand16() & mask;
+      CONVERT_TO_SHORTPTR(pred_)[j] = rnd_.Rand16() & mask;
+    }
+
+    func_(block_height_, block_width_, diff_, block_width_, src_, block_width_,
+          pred_, block_width_, bit_depth_);
+
+    for (int r = 0; r < block_height_; ++r) {
+      for (int c = 0; c < block_width_; ++c) {
+        EXPECT_EQ(diff_[r * block_width_ + c],
+                  (CONVERT_TO_SHORTPTR(src_)[r * block_width_ + c] -
+                   CONVERT_TO_SHORTPTR(pred_)[r * block_width_ + c]))
+            << "r = " << r << ", c = " << c << ", test: " << i;
+      }
+    }
+  }
+}
+
+TEST_P(AV1HBDSubtractBlockTest, CheckResult) { CheckResult(); }
+
+#if USE_SPEED_TEST
+TEST_P(AV1HBDSubtractBlockTest, CheckSpeed) { RunForSpeed(); }
+#endif  // USE_SPEED_TEST
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(
+    SSE2, AV1HBDSubtractBlockTest,
+    ::testing::Values(make_tuple(4, 4, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(4, 4, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(4, 8, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(4, 8, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(8, 4, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(8, 4, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(8, 8, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(8, 8, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(8, 16, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(8, 16, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(16, 8, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(16, 8, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(16, 16, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(16, 16, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(16, 32, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(16, 32, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(32, 16, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(32, 16, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(32, 32, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(32, 32, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(32, 64, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(32, 64, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(64, 32, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(64, 32, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(64, 64, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(64, 64, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(64, 128, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(64, 128, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(128, 64, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(128, 64, 12, &aom_highbd_subtract_block_c),
+                      make_tuple(128, 128, 12, &aom_highbd_subtract_block_sse2),
+                      make_tuple(128, 128, 12, &aom_highbd_subtract_block_c)));
+#endif  // HAVE_SSE2
+#endif  // CONFIG_HIGHBITDEPTH
+}  // namespace
diff --git a/third_party/aom/test/sum_squares_test.cc b/third_party/aom/test/sum_squares_test.cc
new file mode 100644
index 0000000000..b8701c1964
--- /dev/null
+++ b/third_party/aom/test/sum_squares_test.cc
@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <cmath>
+#include <cstdlib>
+#include <string>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "aom_ports/mem.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "test/util.h"
+#include "test/function_equivalence_test.h"
+
+using libaom_test::ACMRandom;
+using libaom_test::FunctionEquivalenceTest;
+
+namespace {
+const int kNumIterations = 10000;
+
+static const int16_t kInt13Max = (1 << 12) - 1;
+
+typedef uint64_t (*SSI16Func)(const int16_t *src, int stride, int width,
+                              int height);
+typedef libaom_test::FuncParam<SSI16Func> TestFuncs;
+
+class SumSquaresTest : public ::testing::TestWithParam<TestFuncs> {
+ public:
+  virtual ~SumSquaresTest() {}
+  virtual void SetUp() { params_ = this->GetParam(); }
+
+  virtual void TearDown() { libaom_test::ClearSystemState(); }
+
+ protected:
+  TestFuncs params_;
+};
+
+TEST_P(SumSquaresTest, OperationCheck) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, int16_t, src[256 * 256]);
+
+  int failed = 0;
+
+  const int msb = 11;  // Up to 12 bit input
+  const int limit = 1 << (msb + 1);
+
+  for (int k = 0; k < kNumIterations; k++) {
+    int width = 4 * rnd(32);   // Up to 128x128
+    int height = 4 * rnd(32);  // Up to 128x128
+    int stride = 4 << rnd(7);  // Up to 256 stride
+    while (stride < width) {   // Make sure it's valid
+      stride = 4 << rnd(7);
+    }
+
+    for (int ii = 0; ii < height; ii++) {
+      for (int jj = 0; jj < width; jj++) {
+        src[ii * stride + jj] = rnd(2) ? rnd(limit) : -rnd(limit);
+      }
+    }
+
+    const uint64_t res_ref = params_.ref_func(src, stride, width, height);
+    uint64_t res_tst;
+    ASM_REGISTER_STATE_CHECK(res_tst =
+                                 params_.tst_func(src, stride, width, height));
+
+    if (!failed) {
+      failed = res_ref != res_tst;
+      EXPECT_EQ(res_ref, res_tst)
+          << "Error: Sum Squares Test"
+          << " C output does not match optimized output.";
+    }
+  }
+}
+
+TEST_P(SumSquaresTest, ExtremeValues) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  DECLARE_ALIGNED(16, int16_t, src[256 * 256]);
+
+  int failed = 0;
+
+  const int msb = 11;  // Up to 12 bit input
+  const int limit = 1 << (msb + 1);
+
+  for (int k = 0; k < kNumIterations; k++) {
+    int width = 4 * rnd(32);   // Up to 128x128
+    int height = 4 * rnd(32);  // Up to 128x128
+    int stride = 4 << rnd(7);  // Up to 256 stride
+    while (stride < width) {   // Make sure it's valid
+      stride = 4 << rnd(7);
+    }
+
+    int val = rnd(2) ? limit - 1 : -(limit - 1);
+    for (int ii = 0; ii < height; ii++) {
+      for (int jj = 0; jj < width; jj++) {
+        src[ii * stride + jj] = val;
+      }
+    }
+
+    const uint64_t res_ref = params_.ref_func(src, stride, width, height);
+    uint64_t res_tst;
+    ASM_REGISTER_STATE_CHECK(res_tst =
+                                 params_.tst_func(src, stride, width, height));
+
+    if (!failed) {
+      failed = res_ref != res_tst;
+      EXPECT_EQ(res_ref, res_tst)
+          << "Error: Sum Squares Test"
+          << " C output does not match optimized output.";
+    }
+  }
+}
+
+#if HAVE_SSE2
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, SumSquaresTest,
+    ::testing::Values(TestFuncs(&aom_sum_squares_2d_i16_c,
+                                &aom_sum_squares_2d_i16_sse2)));
+
+#endif  // HAVE_SSE2
+
+//////////////////////////////////////////////////////////////////////////////
+// 1D version
+//////////////////////////////////////////////////////////////////////////////
+
+typedef uint64_t (*F1D)(const int16_t *src, uint32_t N);
+typedef libaom_test::FuncParam<F1D> TestFuncs1D;
+
+class SumSquares1DTest : public FunctionEquivalenceTest<F1D> {
+ protected:
+  static const int kIterations = 1000;
+  static const int kMaxSize = 256;
+};
+
+TEST_P(SumSquares1DTest, RandomValues) {
+  DECLARE_ALIGNED(16, int16_t, src[kMaxSize * kMaxSize]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    for (int i = 0; i < kMaxSize * kMaxSize; ++i)
+      src[i] = rng_(kInt13Max * 2 + 1) - kInt13Max;
+
+    const int N = rng_(2) ? rng_(kMaxSize * kMaxSize + 1 - kMaxSize) + kMaxSize
+                          : rng_(kMaxSize) + 1;
+
+    const uint64_t ref_res = params_.ref_func(src, N);
+    uint64_t tst_res;
+    ASM_REGISTER_STATE_CHECK(tst_res = params_.tst_func(src, N));
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+TEST_P(SumSquares1DTest, ExtremeValues) {
+  DECLARE_ALIGNED(16, int16_t, src[kMaxSize * kMaxSize]);
+
+  for (int iter = 0; iter < kIterations && !HasFatalFailure(); ++iter) {
+    if (rng_(2)) {
+      for (int i = 0; i < kMaxSize * kMaxSize; ++i) src[i] = kInt13Max;
+    } else {
+      for (int i = 0; i < kMaxSize * kMaxSize; ++i) src[i] = -kInt13Max;
+    }
+
+    const int N = rng_(2) ? rng_(kMaxSize * kMaxSize + 1 - kMaxSize) + kMaxSize
+                          : rng_(kMaxSize) + 1;
+
+    const uint64_t ref_res = params_.ref_func(src, N);
+    uint64_t tst_res;
+    ASM_REGISTER_STATE_CHECK(tst_res = params_.tst_func(src, N));
+
+    ASSERT_EQ(ref_res, tst_res);
+  }
+}
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, SumSquares1DTest,
+                        ::testing::Values(TestFuncs1D(
+                            aom_sum_squares_i16_c, aom_sum_squares_i16_sse2)));
+
+#endif  // HAVE_SSE2
+}  // namespace
diff --git a/third_party/aom/test/superframe_test.cc b/third_party/aom/test/superframe_test.cc
new file mode 100644
index 0000000000..0f54baeaf4
--- /dev/null
+++ b/third_party/aom/test/superframe_test.cc
@@ -0,0 +1,141 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <climits>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+
+namespace {
+
+const int kTestMode = 0;
+const int kTileCols = 1;
+const int kTileRows = 2;
+
+typedef std::tr1::tuple<libaom_test::TestMode, int, int> SuperframeTestParam;
+
+class SuperframeTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWithParam<SuperframeTestParam> {
+ protected:
+  SuperframeTest()
+      : EncoderTest(GET_PARAM(0)), modified_buf_(NULL), last_sf_pts_(0) {}
+  virtual ~SuperframeTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    const SuperframeTestParam input = GET_PARAM(1);
+    const libaom_test::TestMode mode = std::tr1::get<kTestMode>(input);
+    SetMode(mode);
+    sf_count_ = 0;
+    sf_count_max_ = INT_MAX;
+    n_tile_cols_ = std::tr1::get<kTileCols>(input);
+    n_tile_rows_ = std::tr1::get<kTileRows>(input);
+  }
+
+  virtual void TearDown() { delete[] modified_buf_; }
+
+  virtual void PreEncodeFrameHook(libaom_test::VideoSource *video,
+                                  libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AOME_SET_ENABLEAUTOALTREF, 1);
+      encoder->Control(AOME_SET_CPUUSED, 2);
+      encoder->Control(AV1E_SET_TILE_COLUMNS, n_tile_cols_);
+      encoder->Control(AV1E_SET_TILE_ROWS, n_tile_rows_);
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+      encoder->Control(AV1E_SET_TILE_LOOPFILTER, 0);
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+    }
+  }
+
+  virtual const aom_codec_cx_pkt_t *MutateEncoderOutputHook(
+      const aom_codec_cx_pkt_t *pkt) {
+    if (pkt->kind != AOM_CODEC_CX_FRAME_PKT) return pkt;
+
+    const uint8_t *buffer = reinterpret_cast<uint8_t *>(pkt->data.frame.buf);
+    const uint8_t marker = buffer[pkt->data.frame.sz - 1];
+    const int frames = (marker & 0x7) + 1;
+    const int mag = ((marker >> 3) & 3) + 1;
+    const unsigned int index_sz = 2 + mag * (frames - 1);
+    if ((marker & 0xe0) == 0xc0 && pkt->data.frame.sz >= index_sz &&
+        buffer[pkt->data.frame.sz - index_sz] == marker) {
+      // frame is a superframe. strip off the index.
+      if (modified_buf_) delete[] modified_buf_;
+      modified_buf_ = new uint8_t[pkt->data.frame.sz - index_sz];
+      memcpy(modified_buf_, pkt->data.frame.buf, pkt->data.frame.sz - index_sz);
+      modified_pkt_ = *pkt;
+      modified_pkt_.data.frame.buf = modified_buf_;
+      modified_pkt_.data.frame.sz -= index_sz;
+
+      sf_count_++;
+      last_sf_pts_ = pkt->data.frame.pts;
+      return &modified_pkt_;
+    }
+
+    // Make sure we do a few frames after the last SF
+    abort_ |=
+        sf_count_ > sf_count_max_ && pkt->data.frame.pts - last_sf_pts_ >= 5;
+    return pkt;
+  }
+
+  int sf_count_;
+  int sf_count_max_;
+  aom_codec_cx_pkt_t modified_pkt_;
+  uint8_t *modified_buf_;
+  aom_codec_pts_t last_sf_pts_;
+
+ private:
+  int n_tile_cols_;
+  int n_tile_rows_;
+};
+
+TEST_P(SuperframeTest, TestSuperframeIndexIsOptional) {
+  sf_count_max_ = 0;  // early exit on successful test.
+  cfg_.g_lag_in_frames = 25;
+
+  ::libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
+                                       30, 1, 0, 40);
+  ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+#if CONFIG_EXT_REFS
+  // NOTE: The use of BWDREF_FRAME will enable the coding of more non-show
+  //       frames besides ALTREF_FRAME.
+  EXPECT_GE(sf_count_, 1);
+#else
+  EXPECT_EQ(sf_count_, 1);
+#endif  // CONFIG_EXT_REFS
+}
+
+// The superframe index is currently mandatory with both ANS and DAALA_EC due
+// to the decoder starting at the end of the buffer.
+#if CONFIG_EXT_TILE
+// Single tile does not work with ANS (see comment above).
+#if CONFIG_ANS || CONFIG_DAALA_EC
+const int tile_col_values[] = { 1, 2 };
+#else
+const int tile_col_values[] = { 1, 2, 32 };
+#endif
+const int tile_row_values[] = { 1, 2, 32 };
+AV1_INSTANTIATE_TEST_CASE(
+    SuperframeTest,
+    ::testing::Combine(::testing::Values(::libaom_test::kTwoPassGood),
+                       ::testing::ValuesIn(tile_col_values),
+                       ::testing::ValuesIn(tile_row_values)));
+#else
+#if !CONFIG_ANS && !CONFIG_DAALA_EC
+AV1_INSTANTIATE_TEST_CASE(
+    SuperframeTest,
+    ::testing::Combine(::testing::Values(::libaom_test::kTwoPassGood),
+                       ::testing::Values(0), ::testing::Values(0)));
+#endif  // !CONFIG_ANS
+#endif  // CONFIG_EXT_TILE
+}  // namespace
diff --git a/third_party/aom/test/test-data.mk b/third_party/aom/test/test-data.mk
new file mode 100644
index 0000000000..168144a00f
--- /dev/null
+++ b/third_party/aom/test/test-data.mk
@@ -0,0 +1,45 @@
+LIBAOM_TEST_SRCS-yes += test-data.mk
+
+# Encoder test source
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += hantro_collage_w352h288.yuv
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += hantro_odd.yuv
+
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_420.y4m
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_422.y4m
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_444.y4m
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_10_440.yuv
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_420.y4m
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_422.y4m
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_444.y4m
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_12_440.yuv
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_420_a10-1.y4m
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_420.y4m
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_422.y4m
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_444.y4m
+LIBAOM_TEST_DATA-$(CONFIG_ENCODERS) += park_joy_90p_8_440.yuv
+
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += desktop_credits.y4m
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += niklas_1280_720_30.y4m
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += rush_hour_444.y4m
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += screendata.y4m
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += niklas_640_480_30.yuv
+
+ifeq ($(CONFIG_DECODE_PERF_TESTS),yes)
+# Encode / Decode test
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += niklas_1280_720_30.yuv
+endif  # CONFIG_DECODE_PERF_TESTS
+
+ifeq ($(CONFIG_ENCODE_PERF_TESTS),yes)
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += desktop_640_360_30.yuv
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += kirland_640_480_30.yuv
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += macmarcomoving_640_480_30.yuv
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += macmarcostationary_640_480_30.yuv
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += niklas_1280_720_30.yuv
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += tacomanarrows_640_480_30.yuv
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += tacomasmallcameramovement_640_480_30.yuv
+LIBAOM_TEST_DATA-$(CONFIG_AV1_ENCODER) += thaloundeskmtg_640_480_30.yuv
+endif  # CONFIG_ENCODE_PERF_TESTS
+
+# sort and remove duplicates
+LIBAOM_TEST_DATA-yes := $(sort $(LIBAOM_TEST_DATA-yes))
+
diff --git a/third_party/aom/test/test-data.sha1 b/third_party/aom/test/test-data.sha1
new file mode 100644
index 0000000000..3d9bfc7c4a
--- /dev/null
+++ b/third_party/aom/test/test-data.sha1
@@ -0,0 +1,28 @@
+d5dfb0151c9051f8c85999255645d7a23916d3c0 *hantro_collage_w352h288.yuv
+b87815bf86020c592ccc7a846ba2e28ec8043902 *hantro_odd.yuv
+a432f96ff0a787268e2f94a8092ab161a18d1b06 *park_joy_90p_10_420.y4m
+0b194cc312c3a2e84d156a221b0a5eb615dfddc5 *park_joy_90p_10_422.y4m
+ff0e0a21dc2adc95b8c1b37902713700655ced17 *park_joy_90p_10_444.y4m
+c934da6fb8cc54ee2a8c17c54cf6076dac37ead0 *park_joy_90p_10_440.yuv
+614c32ae1eca391e867c70d19974f0d62664dd99 *park_joy_90p_12_420.y4m
+c92825f1ea25c5c37855083a69faac6ac4641a9e *park_joy_90p_12_422.y4m
+b592189b885b6cc85db55cc98512a197d73d3b34 *park_joy_90p_12_444.y4m
+82c1bfcca368c2f22bad7d693d690d5499ecdd11 *park_joy_90p_12_440.yuv
+b9e1e90aece2be6e2c90d89e6ab2372d5f8c792d *park_joy_90p_8_420_a10-1.y4m
+4e0eb61e76f0684188d9bc9f3ce61f6b6b77bb2c *park_joy_90p_8_420.y4m
+7a193ff7dfeb96ba5f82b2afd7afa9e1fe83d947 *park_joy_90p_8_422.y4m
+bdb7856e6bc93599bdda05c2e773a9f22b6c6d03 *park_joy_90p_8_444.y4m
+81e1f3843748438b8f2e71db484eb22daf72e939 *park_joy_90p_8_440.yuv
+b1f1c3ec79114b9a0651af24ce634afb44a9a419 *rush_hour_444.y4m
+eb438c6540eb429f74404eedfa3228d409c57874 *desktop_640_360_30.yuv
+89e70ebd22c27d275fe14dc2f1a41841a6d8b9ab *kirland_640_480_30.yuv
+33c533192759e5bb4f07abfbac389dc259db4686 *macmarcomoving_640_480_30.yuv
+8bfaab121080821b8f03b23467911e59ec59b8fe *macmarcostationary_640_480_30.yuv
+70894878d916a599842d9ad0dcd24e10c13e5467 *niklas_640_480_30.yuv
+8784b6df2d8cc946195a90ac00540500d2e522e4 *tacomanarrows_640_480_30.yuv
+edd86a1f5e62fd9da9a9d46078247759c2638009 *tacomasmallcameramovement_640_480_30.yuv
+9a70e8b7d14fba9234d0e51dce876635413ce444 *thaloundeskmtg_640_480_30.yuv
+e7d315dbf4f3928779e0dc624311196d44491d32 *niklas_1280_720_30.yuv
+717da707afcaa1f692ff1946f291054eb75a4f06 *screendata.y4m
+9cfc855459e7549fd015c79e8eca512b2f2cb7e3 *niklas_1280_720_30.y4m
+5b5763b388b1b52a81bb82b39f7ec25c4bd3d0e1 *desktop_credits.y4m
diff --git a/third_party/aom/test/test.cmake b/third_party/aom/test/test.cmake
new file mode 100644
index 0000000000..8d3ab7059c
--- /dev/null
+++ b/third_party/aom/test/test.cmake
@@ -0,0 +1,315 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+include("${AOM_ROOT}/test/test_data_util.cmake")
+
+set(AOM_UNIT_TEST_WRAPPER_SOURCES
+    "${AOM_CONFIG_DIR}/usage_exit.c"
+    "${AOM_ROOT}/test/test_libaom.cc")
+
+set(AOM_UNIT_TEST_COMMON_SOURCES
+    "${AOM_ROOT}/test/acm_random.h"
+    "${AOM_ROOT}/test/clear_system_state.h"
+    "${AOM_ROOT}/test/codec_factory.h"
+    "${AOM_ROOT}/test/convolve_test.cc"
+    "${AOM_ROOT}/test/function_equivalence_test.h"
+    "${AOM_ROOT}/test/md5_helper.h"
+    "${AOM_ROOT}/test/register_state_check.h"
+    "${AOM_ROOT}/test/transform_test_base.h"
+    "${AOM_ROOT}/test/util.h"
+    "${AOM_ROOT}/test/video_source.h")
+
+if (CONFIG_ACCOUNTING)
+  set(AOM_UNIT_TEST_COMMON_SOURCES
+      ${AOM_UNIT_TEST_COMMON_SOURCES}
+      "${AOM_ROOT}/test/accounting_test.cc")
+endif ()
+
+if (CONFIG_ADAPT_SCAN)
+  set(AOM_UNIT_TEST_COMMON_SOURCES
+      ${AOM_UNIT_TEST_COMMON_SOURCES}
+      "${AOM_ROOT}/test/scan_test.cc")
+endif ()
+
+if (CONFIG_GLOBAL_MOTION OR CONFIG_WARPED_MOTION)
+  if (HAVE_SSE2)
+    set(AOM_UNIT_TEST_COMMON_SOURCES
+        ${AOM_UNIT_TEST_COMMON_SOURCES}
+        "${AOM_ROOT}/test/warp_filter_test.cc"
+        "${AOM_ROOT}/test/warp_filter_test_util.cc"
+        "${AOM_ROOT}/test/warp_filter_test_util.h")
+  endif ()
+endif ()
+
+set(AOM_UNIT_TEST_DECODER_SOURCES
+    "${AOM_ROOT}/test/decode_api_test.cc"
+    "${AOM_ROOT}/test/decode_test_driver.cc"
+    "${AOM_ROOT}/test/decode_test_driver.h"
+    "${AOM_ROOT}/test/ivf_video_source.h")
+
+set(AOM_UNIT_TEST_ENCODER_SOURCES
+    "${AOM_ROOT}/test/altref_test.cc"
+    "${AOM_ROOT}/test/aq_segment_test.cc"
+    "${AOM_ROOT}/test/datarate_test.cc"
+    "${AOM_ROOT}/test/dct16x16_test.cc"
+    "${AOM_ROOT}/test/dct32x32_test.cc"
+    "${AOM_ROOT}/test/encode_api_test.cc"
+    "${AOM_ROOT}/test/encode_test_driver.cc"
+    "${AOM_ROOT}/test/encode_test_driver.h"
+    "${AOM_ROOT}/test/error_resilience_test.cc"
+    "${AOM_ROOT}/test/i420_video_source.h"
+    "${AOM_ROOT}/test/sad_test.cc"
+    "${AOM_ROOT}/test/y4m_test.cc"
+    "${AOM_ROOT}/test/y4m_video_source.h"
+    "${AOM_ROOT}/test/yuv_video_source.h")
+
+set(AOM_DECODE_PERF_TEST_SOURCES "${AOM_ROOT}/test/decode_perf_test.cc")
+set(AOM_ENCODE_PERF_TEST_SOURCES "${AOM_ROOT}/test/encode_perf_test.cc")
+set(AOM_UNIT_TEST_WEBM_SOURCES "${AOM_ROOT}/test/webm_video_source.h")
+
+set(AOM_TEST_INTRA_PRED_SPEED_SOURCES
+    "${AOM_CONFIG_DIR}/usage_exit.c"
+    "${AOM_ROOT}/test/test_intra_pred_speed.cc")
+
+if (CONFIG_AV1)
+  set(AOM_UNIT_TEST_COMMON_SOURCES
+      ${AOM_UNIT_TEST_COMMON_SOURCES}
+      "${AOM_ROOT}/test/av1_convolve_optimz_test.cc"
+      "${AOM_ROOT}/test/av1_convolve_test.cc"
+      "${AOM_ROOT}/test/av1_fwd_txfm1d_test.cc"
+      "${AOM_ROOT}/test/av1_fwd_txfm2d_test.cc"
+      "${AOM_ROOT}/test/av1_inv_txfm1d_test.cc"
+      "${AOM_ROOT}/test/av1_inv_txfm2d_test.cc"
+      "${AOM_ROOT}/test/av1_txfm_test.cc"
+      "${AOM_ROOT}/test/av1_txfm_test.h"
+      "${AOM_ROOT}/test/intrapred_test.cc"
+      "${AOM_ROOT}/test/lpf_8_test.cc"
+      "${AOM_ROOT}/test/simd_cmp_impl.h")
+
+  if (CONFIG_CDEF)
+    set(AOM_UNIT_TEST_COMMON_SOURCES
+        ${AOM_UNIT_TEST_COMMON_SOURCES}
+        "${AOM_ROOT}/test/clpf_test.cc")
+  endif ()
+
+  if (CONFIG_FILTER_INTRA)
+    if (HAVE_SSE4_1)
+      set(AOM_UNIT_TEST_COMMON_SOURCES
+          ${AOM_UNIT_TEST_COMMON_SOURCES}
+          # TODO: not sure if this intrinsics or a wrapper calling intrin/asm.
+          #"${AOM_ROOT}/test/filterintra_predictors_test.cc")
+          )
+    endif ()
+  endif ()
+
+  set(AOM_UNIT_TEST_COMMON_INTRIN_NEON
+    ${AOM_UNIT_TEST_COMMON_INTRIN_NEON}
+      "${AOM_ROOT}/test/simd_cmp_neon.cc"
+      "${AOM_ROOT}/test/simd_neon_test.cc")
+  set(AOM_UNIT_TEST_COMMON_INTRIN_SSE2
+      ${AOM_UNIT_TEST_COMMON_INTRIN_SSE2}
+      "${AOM_ROOT}/test/simd_cmp_sse2.cc")
+  set(AOM_UNIT_TEST_COMMON_INTRIN_SSSE3
+      ${AOM_UNIT_TEST_COMMON_INTRIN_SSSE3}
+      "${AOM_ROOT}/test/simd_cmp_ssse3.cc")
+  set(AOM_UNIT_TEST_COMMON_INTRIN_SSE4_1
+      ${AOM_UNIT_TEST_COMMON_INTRIN_SSE4_1}
+      "${AOM_ROOT}/test/simd_cmp_sse4.cc")
+endif ()
+
+if (CONFIG_AV1_ENCODER)
+  set(AOM_UNIT_TEST_ENCODER_SOURCES
+      ${AOM_UNIT_TEST_ENCODER_SOURCES}
+      "${AOM_ROOT}/test/active_map_test.cc"
+      "${AOM_ROOT}/test/arf_freq_test.cc"
+      "${AOM_ROOT}/test/av1_dct_test.cc"
+      "${AOM_ROOT}/test/av1_fht16x16_test.cc"
+      "${AOM_ROOT}/test/av1_fht8x8_test.cc"
+      "${AOM_ROOT}/test/av1_inv_txfm_test.cc"
+      "${AOM_ROOT}/test/avg_test.cc"
+      "${AOM_ROOT}/test/blend_a64_mask_1d_test.cc"
+      "${AOM_ROOT}/test/blend_a64_mask_test.cc"
+      "${AOM_ROOT}/test/borders_test.cc"
+      "${AOM_ROOT}/test/cpu_speed_test.cc"
+      "${AOM_ROOT}/test/end_to_end_test.cc"
+      "${AOM_ROOT}/test/error_block_test.cc"
+      "${AOM_ROOT}/test/fdct4x4_test.cc"
+      "${AOM_ROOT}/test/fdct8x8_test.cc"
+      "${AOM_ROOT}/test/frame_size_tests.cc"
+      "${AOM_ROOT}/test/hadamard_test.cc"
+      "${AOM_ROOT}/test/lossless_test.cc"
+      "${AOM_ROOT}/test/minmax_test.cc"
+      "${AOM_ROOT}/test/subtract_test.cc"
+      "${AOM_ROOT}/test/sum_squares_test.cc"
+      "${AOM_ROOT}/test/variance_test.cc")
+
+  if (CONFIG_EXT_INTER)
+    set(AOM_UNIT_TEST_ENCODER_SOURCES
+        ${AOM_UNIT_TEST_ENCODER_SOURCES}
+        "${AOM_ROOT}/test/av1_wedge_utils_test.cc"
+        "${AOM_ROOT}/test/masked_sad_test.cc"
+        "${AOM_ROOT}/test/masked_variance_test.cc")
+  endif ()
+
+  if (CONFIG_EXT_TX)
+    set(AOM_UNIT_TEST_ENCODER_SOURCES
+        ${AOM_UNIT_TEST_ENCODER_SOURCES}
+        "${AOM_ROOT}/test/av1_fht16x32_test.cc"
+        "${AOM_ROOT}/test/av1_fht16x8_test.cc"
+        "${AOM_ROOT}/test/av1_fht32x16_test.cc"
+        "${AOM_ROOT}/test/av1_fht4x4_test.cc"
+        "${AOM_ROOT}/test/av1_fht4x8_test.cc"
+        "${AOM_ROOT}/test/av1_fht8x16_test.cc"
+        "${AOM_ROOT}/test/av1_fht8x4_test.cc"
+        "${AOM_ROOT}/test/fht32x32_test.cc")
+  endif ()
+
+  if (CONFIG_MOTION_VAR)
+    set(AOM_UNIT_TEST_ENCODER_SOURCES
+        ${AOM_UNIT_TEST_ENCODER_SOURCES}
+        "${AOM_ROOT}/test/obmc_sad_test.cc"
+        "${AOM_ROOT}/test/obmc_variance_test.cc")
+  endif ()
+endif ()
+
+if (CONFIG_AV1_DECODER AND CONFIG_AV1_ENCODER)
+  set(AOM_UNIT_TEST_COMMON_SOURCES
+      ${AOM_UNIT_TEST_COMMON_SOURCES}
+      "${AOM_ROOT}/test/divu_small_test.cc"
+      "${AOM_ROOT}/test/ethread_test.cc"
+      "${AOM_ROOT}/test/idct8x8_test.cc"
+      "${AOM_ROOT}/test/partial_idct_test.cc"
+      "${AOM_ROOT}/test/superframe_test.cc"
+      "${AOM_ROOT}/test/binary_codes_test.cc"
+      "${AOM_ROOT}/test/tile_independence_test.cc")
+
+  if (CONFIG_ANS)
+    set(AOM_UNIT_TEST_COMMON_SOURCES
+        ${AOM_UNIT_TEST_COMMON_SOURCES}
+        "${AOM_ROOT}/test/ans_codec_test.cc"
+        "${AOM_ROOT}/test/ans_test.cc")
+  else ()
+    set(AOM_UNIT_TEST_COMMON_SOURCES
+        ${AOM_UNIT_TEST_COMMON_SOURCES}
+        "${AOM_ROOT}/test/boolcoder_test.cc")
+  endif ()
+
+  if (CONFIG_EXT_TILE)
+    set(AOM_UNIT_TEST_COMMON_SOURCES
+        ${AOM_UNIT_TEST_COMMON_SOURCES}
+        "${AOM_ROOT}/test/av1_ext_tile_test.cc")
+  endif ()
+endif ()
+
+if (CONFIG_HIGHBITDEPTH)
+  if (CONFIG_AV1)
+    set(AOM_UNIT_TEST_COMMON_INTRIN_SSE4_1
+        ${AOM_UNIT_TEST_COMMON_INTRIN_SSE4_1}
+        "${AOM_ROOT}/test/av1_highbd_iht_test.cc"
+        "${AOM_ROOT}/test/av1_quantize_test.cc")
+  endif ()
+
+  if (CONFIG_INTERNAL_STATS)
+    set(AOM_UNIT_TEST_COMMON_SOURCES
+        ${AOM_UNIT_TEST_COMMON_SOURCES}
+        "${AOM_ROOT}/test/hbd_metrics_test.cc")
+  endif ()
+endif ()
+
+if (CONFIG_UNIT_TESTS)
+  if (MSVC)
+    # Force static run time to avoid collisions with googletest.
+    include("${AOM_ROOT}/build/cmake/msvc_runtime.cmake")
+  endif ()
+  include_directories(
+    "${AOM_ROOT}/third_party/googletest/src/googletest/src"
+    "${AOM_ROOT}/third_party/googletest/src/googletest/include")
+  add_subdirectory("${AOM_ROOT}/third_party/googletest/src/googletest"
+                   EXCLUDE_FROM_ALL)
+
+  # Generate a stub file containing the C function usage_exit(); this is
+  # required because of the test dependency on aom_common_app_util.
+  # Specifically, the function die() in tools_common.c calls usage_exit() to
+  # terminate the program on the caller's behalf.
+  file(WRITE "${AOM_CONFIG_DIR}/usage_exit.c" "void usage_exit(void) {}")
+endif ()
+
+# Setup the targets for CONFIG_UNIT_TESTS. The libaom and app util targets must
+# exist before this function is called.
+function (setup_aom_test_targets)
+  add_library(test_aom_common OBJECT ${AOM_UNIT_TEST_COMMON_SOURCES})
+  add_library(test_aom_decoder OBJECT ${AOM_UNIT_TEST_DECODER_SOURCES})
+  add_library(test_aom_encoder OBJECT ${AOM_UNIT_TEST_ENCODER_SOURCES})
+
+  set(AOM_LIB_TARGETS ${AOM_LIB_TARGETS} test_aom_common test_aom_decoder
+      test_aom_encoder PARENT_SCOPE)
+
+  add_executable(test_libaom ${AOM_UNIT_TEST_WRAPPER_SOURCES}
+                 $<TARGET_OBJECTS:aom_common_app_util>
+                 $<TARGET_OBJECTS:test_aom_common>)
+
+  if (CONFIG_DECODERS)
+    target_sources(test_libaom PUBLIC
+                   $<TARGET_OBJECTS:aom_decoder_app_util>
+                   $<TARGET_OBJECTS:test_aom_decoder>)
+
+    if (CONFIG_DECODE_PERF_TESTS AND CONFIG_WEBM_IO)
+      target_sources(test_libaom PUBLIC ${AOM_DECODE_PERF_TEST_SOURCES})
+    endif ()
+  endif ()
+
+  if (CONFIG_ENCODERS)
+    target_sources(test_libaom PUBLIC
+                   $<TARGET_OBJECTS:test_aom_encoder>
+                   $<TARGET_OBJECTS:aom_encoder_app_util>)
+
+    if (CONFIG_ENCODE_PERF_TESTS)
+      target_sources(test_libaom PUBLIC ${AOM_ENCODE_PERF_TEST_SOURCES})
+    endif ()
+  endif ()
+
+  target_link_libraries(test_libaom PUBLIC aom gtest)
+
+  add_executable(test_intra_pred_speed
+                 ${AOM_TEST_INTRA_PRED_SPEED_SOURCES}
+                 $<TARGET_OBJECTS:aom_common_app_util>)
+  target_link_libraries(test_intra_pred_speed PUBLIC aom gtest)
+
+  if (CONFIG_LIBYUV)
+    target_sources(test_libaom PUBLIC $<TARGET_OBJECTS:yuv>)
+  endif ()
+  if (CONFIG_WEBM_IO)
+    target_sources(test_libaom PUBLIC ${AOM_UNIT_TEST_WEBM_SOURCES}
+                   $<TARGET_OBJECTS:webm>)
+  endif ()
+  if (HAVE_SSE2)
+    add_intrinsics_source_to_target("-msse2" "test_libaom"
+                                    "AOM_UNIT_TEST_COMMON_INTRIN_SSE2")
+  endif ()
+  if (HAVE_SSSE3)
+    add_intrinsics_source_to_target("-mssse3" "test_libaom"
+                                    "AOM_UNIT_TEST_COMMON_INTRIN_SSSE3")
+  endif ()
+  if (HAVE_SSE4_1)
+    add_intrinsics_source_to_target("-msse4.1" "test_libaom"
+                                    "AOM_UNIT_TEST_COMMON_INTRIN_SSE4_1")
+  endif ()
+  if (HAVE_NEON)
+    add_intrinsics_source_to_target("${AOM_NEON_INTRIN_FLAG}" "test_libaom"
+                                    "AOM_UNIT_TEST_COMMON_INTRIN_NEON")
+  endif ()
+
+  add_custom_target(testdata
+                    COMMAND ${CMAKE_COMMAND}
+                      -DAOM_CONFIG_DIR="${AOM_CONFIG_DIR}"
+                      -DAOM_ROOT="${AOM_ROOT}"
+                      -P "${AOM_ROOT}/test/test_worker.cmake"
+                    SOURCES ${AOM_TEST_DATA_LIST})
+endfunction ()
diff --git a/third_party/aom/test/test.mk b/third_party/aom/test/test.mk
new file mode 100644
index 0000000000..fb0ab371e3
--- /dev/null
+++ b/third_party/aom/test/test.mk
@@ -0,0 +1,241 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+LIBAOM_TEST_SRCS-yes += acm_random.h
+LIBAOM_TEST_SRCS-yes += clear_system_state.h
+LIBAOM_TEST_SRCS-yes += codec_factory.h
+LIBAOM_TEST_SRCS-yes += md5_helper.h
+LIBAOM_TEST_SRCS-yes += register_state_check.h
+LIBAOM_TEST_SRCS-yes += test.mk
+LIBAOM_TEST_SRCS-yes += test_libaom.cc
+LIBAOM_TEST_SRCS-yes += util.h
+LIBAOM_TEST_SRCS-yes += video_source.h
+LIBAOM_TEST_SRCS-yes += transform_test_base.h
+LIBAOM_TEST_SRCS-yes += function_equivalence_test.h
+LIBAOM_TEST_SRCS-yes += warp_filter_test_util.h
+
+##
+## BLACK BOX TESTS
+##
+## Black box tests only use the public API.
+##
+LIBAOM_TEST_SRCS-yes                   += ../md5_utils.h ../md5_utils.c
+LIBAOM_TEST_SRCS-$(CONFIG_DECODERS)    += ivf_video_source.h
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += ../y4minput.h ../y4minput.c
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += altref_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += aq_segment_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += datarate_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += encode_api_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += error_resilience_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += i420_video_source.h
+#LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += realtime_test.cc
+#LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += resize_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += y4m_video_source.h
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += yuv_video_source.h
+
+#LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += level_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += active_map_refresh_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += active_map_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += borders_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += cpu_speed_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += frame_size_tests.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += lossless_test.cc
+
+LIBAOM_TEST_SRCS-yes                   += decode_test_driver.cc
+LIBAOM_TEST_SRCS-yes                   += decode_test_driver.h
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += encode_test_driver.cc
+LIBAOM_TEST_SRCS-yes                   += encode_test_driver.h
+
+## IVF writing.
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += ../ivfenc.c ../ivfenc.h
+
+## Y4m parsing.
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS)    += y4m_test.cc ../y4menc.c ../y4menc.h
+
+## WebM Parsing
+ifeq ($(CONFIG_WEBM_IO), yes)
+LIBWEBM_PARSER_SRCS += ../third_party/libwebm/mkvparser/mkvparser.cc
+LIBWEBM_PARSER_SRCS += ../third_party/libwebm/mkvparser/mkvreader.cc
+LIBWEBM_PARSER_SRCS += ../third_party/libwebm/mkvparser/mkvparser.h
+LIBWEBM_PARSER_SRCS += ../third_party/libwebm/mkvparser/mkvreader.h
+LIBAOM_TEST_SRCS-$(CONFIG_DECODERS)    += $(LIBWEBM_PARSER_SRCS)
+LIBAOM_TEST_SRCS-$(CONFIG_DECODERS)    += ../tools_common.h
+LIBAOM_TEST_SRCS-$(CONFIG_DECODERS)    += ../webmdec.cc
+LIBAOM_TEST_SRCS-$(CONFIG_DECODERS)    += ../webmdec.h
+LIBAOM_TEST_SRCS-$(CONFIG_DECODERS)    += webm_video_source.h
+endif
+
+LIBAOM_TEST_SRCS-$(CONFIG_DECODERS)    += decode_api_test.cc
+
+# Currently we only support decoder perf tests for av1. Also they read from WebM
+# files, so WebM IO is required.
+ifeq ($(CONFIG_DECODE_PERF_TESTS)$(CONFIG_AV1_DECODER)$(CONFIG_WEBM_IO), \
+      yesyesyes)
+LIBAOM_TEST_SRCS-yes                   += decode_perf_test.cc
+endif
+
+# encode perf tests are av1 only
+ifeq ($(CONFIG_ENCODE_PERF_TESTS)$(CONFIG_AV1_ENCODER), yesyes)
+LIBAOM_TEST_SRCS-yes += encode_perf_test.cc
+endif
+
+## Multi-codec / unconditional black box tests.
+ifeq ($(findstring yes,$(CONFIG_AV1_ENCODER)),yes)
+LIBAOM_TEST_SRCS-yes += active_map_refresh_test.cc
+LIBAOM_TEST_SRCS-yes += active_map_test.cc
+LIBAOM_TEST_SRCS-yes += end_to_end_test.cc
+endif
+
+##
+## WHITE BOX TESTS
+##
+## Whitebox tests invoke functions not exposed via the public API. Certain
+## shared library builds don't make these functions accessible.
+##
+ifeq ($(CONFIG_SHARED),)
+
+## AV1
+ifeq ($(CONFIG_AV1),yes)
+
+# These tests require both the encoder and decoder to be built.
+ifeq ($(CONFIG_AV1_ENCODER)$(CONFIG_AV1_DECODER),yesyes)
+# IDCT test currently depends on FDCT function
+LIBAOM_TEST_SRCS-yes                   += idct8x8_test.cc
+LIBAOM_TEST_SRCS-yes                   += partial_idct_test.cc
+LIBAOM_TEST_SRCS-yes                   += superframe_test.cc
+LIBAOM_TEST_SRCS-yes                   += tile_independence_test.cc
+LIBAOM_TEST_SRCS-yes                   += ethread_test.cc
+LIBAOM_TEST_SRCS-yes                   += motion_vector_test.cc
+ifneq ($(CONFIG_ANS),yes)
+LIBAOM_TEST_SRCS-yes                   += binary_codes_test.cc
+endif
+ifeq ($(CONFIG_EXT_TILE),yes)
+LIBAOM_TEST_SRCS-yes                   += av1_ext_tile_test.cc
+endif
+ifeq ($(CONFIG_ANS),yes)
+LIBAOM_TEST_SRCS-yes                   += ans_test.cc
+LIBAOM_TEST_SRCS-yes                   += ans_codec_test.cc
+else
+LIBAOM_TEST_SRCS-yes                   += boolcoder_test.cc
+ifeq ($(CONFIG_ACCOUNTING),yes)
+LIBAOM_TEST_SRCS-yes                   += accounting_test.cc
+endif
+endif
+LIBAOM_TEST_SRCS-yes                   += divu_small_test.cc
+#LIBAOM_TEST_SRCS-yes                   += encoder_parms_get_to_decoder.cc
+endif
+
+LIBAOM_TEST_SRCS-$(CONFIG_ADAPT_SCAN)  += scan_test.cc
+LIBAOM_TEST_SRCS-yes                   += convolve_test.cc
+LIBAOM_TEST_SRCS-yes                   += lpf_8_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_CDEF)        += dering_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_CDEF)        += clpf_test.cc
+LIBAOM_TEST_SRCS-yes                   += simd_cmp_impl.h
+LIBAOM_TEST_SRCS-$(HAVE_SSE2)          += simd_cmp_sse2.cc
+LIBAOM_TEST_SRCS-$(HAVE_SSSE3)         += simd_cmp_ssse3.cc
+LIBAOM_TEST_SRCS-$(HAVE_SSE4_1)        += simd_cmp_sse4.cc
+LIBAOM_TEST_SRCS-$(HAVE_NEON)          += simd_cmp_neon.cc
+LIBAOM_TEST_SRCS-yes                   += simd_impl.h
+LIBAOM_TEST_SRCS-$(HAVE_SSE2)          += simd_sse2_test.cc
+LIBAOM_TEST_SRCS-$(HAVE_SSSE3)         += simd_ssse3_test.cc
+LIBAOM_TEST_SRCS-$(HAVE_SSE4_1)        += simd_sse4_test.cc
+LIBAOM_TEST_SRCS-$(HAVE_NEON)          += simd_neon_test.cc
+LIBAOM_TEST_SRCS-yes                   += intrapred_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_INTRABC)     += intrabc_test.cc
+#LIBAOM_TEST_SRCS-$(CONFIG_AV1_DECODER) += av1_thread_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += dct16x16_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += dct32x32_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += fdct4x4_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += fdct8x8_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += hadamard_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += minmax_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += variance_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += error_block_test.cc
+#LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_quantize_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += subtract_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += arf_freq_test.cc
+
+
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_inv_txfm_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_dct_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fht4x4_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fht8x8_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fht16x16_test.cc
+ifeq ($(CONFIG_EXT_TX),yes)
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fht4x8_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fht8x4_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fht8x16_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fht16x8_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fht16x32_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fht32x16_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += fht32x32_test.cc
+endif
+
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += sum_squares_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += subtract_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += blend_a64_mask_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += blend_a64_mask_1d_test.cc
+
+ifeq ($(CONFIG_EXT_INTER),yes)
+LIBAOM_TEST_SRCS-$(HAVE_SSSE3) += masked_variance_test.cc
+LIBAOM_TEST_SRCS-$(HAVE_SSSE3) += masked_sad_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_wedge_utils_test.cc
+endif
+
+## Skip the unit test written for 4-tap filter intra predictor, because we
+## revert to 3-tap filter.
+## ifeq ($(CONFIG_FILTER_INTRA),yes)
+## LIBAOM_TEST_SRCS-$(HAVE_SSE4_1) += filterintra_predictors_test.cc
+## endif
+
+ifeq ($(CONFIG_MOTION_VAR),yes)
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += obmc_sad_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += obmc_variance_test.cc
+endif
+
+ifeq ($(CONFIG_HIGHBITDEPTH),yes)
+ifeq ($(CONFIG_AV1_ENCODER),yes)
+LIBAOM_TEST_SRCS-$(HAVE_SSE4_1) += av1_quantize_test.cc
+LIBAOM_TEST_SRCS-$(HAVE_SSE4_1) += av1_highbd_iht_test.cc
+endif
+endif # CONFIG_HIGHBITDEPTH
+endif # AV1
+
+## Multi-codec / unconditional whitebox tests.
+
+ifeq ($(CONFIG_AV1_ENCODER),yes)
+LIBAOM_TEST_SRCS-yes += avg_test.cc
+endif
+ifeq ($(CONFIG_INTERNAL_STATS),yes)
+LIBAOM_TEST_SRCS-$(CONFIG_HIGHBITDEPTH) += hbd_metrics_test.cc
+endif
+LIBAOM_TEST_SRCS-$(CONFIG_ENCODERS) += sad_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1) += av1_txfm_test.h
+LIBAOM_TEST_SRCS-$(CONFIG_AV1) += av1_txfm_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fwd_txfm1d_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_inv_txfm1d_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_fwd_txfm2d_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1_ENCODER) += av1_inv_txfm2d_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1) += av1_convolve_test.cc
+LIBAOM_TEST_SRCS-$(CONFIG_AV1) += av1_convolve_optimz_test.cc
+ifneq ($(findstring yes,$(CONFIG_GLOBAL_MOTION) $(CONFIG_WARPED_MOTION)),)
+LIBAOM_TEST_SRCS-$(HAVE_SSE2) += warp_filter_test.cc warp_filter_test_util.cc
+endif
+ifeq ($(CONFIG_LOOP_RESTORATION),yes)
+LIBAOM_TEST_SRCS-$(HAVE_SSE4_1) += selfguided_filter_test.cc
+endif
+
+TEST_INTRA_PRED_SPEED_SRCS-yes := test_intra_pred_speed.cc
+TEST_INTRA_PRED_SPEED_SRCS-yes += ../md5_utils.h ../md5_utils.c
+
+endif # CONFIG_SHARED
+
+include $(SRC_PATH_BARE)/test/test-data.mk
diff --git a/third_party/aom/test/test_data_util.cmake b/third_party/aom/test/test_data_util.cmake
new file mode 100644
index 0000000000..f096e4e12d
--- /dev/null
+++ b/third_party/aom/test/test_data_util.cmake
@@ -0,0 +1,76 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+
+# Parses test/test-data.sha1 and writes captured file names and checksums to
+# $out_files and $out_checksums as lists.
+function (make_test_data_lists out_files out_checksums)
+  if (NOT AOM_TEST_DATA_LIST OR NOT EXISTS "${AOM_TEST_DATA_LIST}")
+    message(FATAL_ERROR "AOM_TEST_DATA_LIST (${AOM_TEST_DATA_LIST}) missing or "
+            "variable empty.")
+  endif ()
+
+  # Read test-data.sha1 into $files_and_checksums. $files_and_checksums becomes
+  # a list with an entry for each line from $AOM_TEST_DATA_LIST.
+  file(STRINGS "${AOM_TEST_DATA_LIST}" files_and_checksums)
+
+  # Iterate over the list of lines and split it into $checksums and $filenames.
+  foreach (line ${files_and_checksums})
+    string(FIND "${line}" " *" delim_pos)
+
+    math(EXPR filename_pos "${delim_pos} + 2")
+    string(SUBSTRING "${line}" 0 ${delim_pos} checksum)
+    string(SUBSTRING "${line}" ${filename_pos} -1 filename)
+
+    set(checksums ${checksums} ${checksum})
+    set(filenames ${filenames} ${filename})
+  endforeach ()
+
+  if (NOT checksums OR NOT filenames)
+    message(FATAL_ERROR "Parsing of ${AOM_TEST_DATA_LIST} failed.")
+  endif ()
+
+  set(${out_checksums} ${checksums} PARENT_SCOPE)
+  set(${out_files} ${filenames} PARENT_SCOPE)
+endfunction ()
+
+# Appends each file name in $test_files to $test_dir and adds the result path to
+# $out_path_list.
+function (expand_test_file_paths test_files test_dir out_path_list)
+  foreach (filename ${${test_files}})
+    set(path_list ${path_list} "${test_dir}/${filename}")
+  endforeach ()
+  set(${out_path_list} ${path_list} PARENT_SCOPE)
+endfunction ()
+
+function (check_file local_path expected_checksum out_needs_update)
+  if (EXISTS "${local_path}")
+    file(SHA1 "${local_path}" file_checksum)
+  else ()
+    set(${out_needs_update} 1 PARENT_SCOPE)
+    return ()
+  endif ()
+
+  if ("${file_checksum}" STREQUAL "${expected_checksum}")
+    unset(${out_needs_update} PARENT_SCOPE)
+  else ()
+    set(${out_needs_update} 1 PARENT_SCOPE)
+  endif ()
+endfunction ()
+
+# Downloads data from $file_url, confirms that $file_checksum matches, and
+# writes it to $local_path.
+function (download_test_file file_url file_checksum local_path)
+  message("Downloading ${file_url} ...")
+  file(DOWNLOAD "${file_url}" "${local_path}"
+       SHOW_PROGRESS
+       EXPECTED_HASH SHA1=${file_checksum})
+  message("Download of ${file_url} complete.")
+endfunction ()
diff --git a/third_party/aom/test/test_intra_pred_speed.cc b/third_party/aom/test/test_intra_pred_speed.cc
new file mode 100644
index 0000000000..c4253628e7
--- /dev/null
+++ b/third_party/aom/test/test_intra_pred_speed.cc
@@ -0,0 +1,515 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+//  Test and time AOM intra-predictor functions
+
+#include <stdio.h>
+#include <string.h>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/md5_helper.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/aom_timer.h"
+
+// -----------------------------------------------------------------------------
+
+namespace {
+
+typedef void (*AvxPredFunc)(uint8_t *dst, ptrdiff_t y_stride,
+                            const uint8_t *above, const uint8_t *left);
+
+#if CONFIG_ALT_INTRA
+const int kNumAv1IntraFuncs = 14;
+#else
+const int kNumAv1IntraFuncs = 13;
+#endif  // CONFIG_ALT_INTRA
+const char *kAv1IntraPredNames[kNumAv1IntraFuncs] = {
+  "DC_PRED",    "DC_LEFT_PRED", "DC_TOP_PRED", "DC_128_PRED", "V_PRED",
+  "H_PRED",     "D45_PRED",     "D135_PRED",   "D117_PRED",   "D153_PRED",
+  "D207_PRED",  "D63_PRED",     "TM_PRED",
+#if CONFIG_ALT_INTRA
+  "SMOOTH_PRED"
+#endif  // CONFIG_ALT_INTRA
+};
+
+void TestIntraPred(const char name[], AvxPredFunc const *pred_funcs,
+                   const char *const pred_func_names[], int num_funcs,
+                   const char *const signatures[], int /*block_size*/,
+                   int num_pixels_per_test) {
+  libaom_test::ACMRandom rnd(libaom_test::ACMRandom::DeterministicSeed());
+  const int kBPS = 32;
+  const int kTotalPixels = 32 * kBPS;
+  DECLARE_ALIGNED(16, uint8_t, src[kTotalPixels]);
+  DECLARE_ALIGNED(16, uint8_t, ref_src[kTotalPixels]);
+  DECLARE_ALIGNED(16, uint8_t, left[2 * kBPS]);
+  DECLARE_ALIGNED(16, uint8_t, above_mem[2 * kBPS + 16]);
+  uint8_t *const above = above_mem + 16;
+  for (int i = 0; i < kTotalPixels; ++i) ref_src[i] = rnd.Rand8();
+  for (int i = 0; i < kBPS; ++i) left[i] = rnd.Rand8();
+  for (int i = -1; i < kBPS; ++i) above[i] = rnd.Rand8();
+  const int kNumTests = static_cast<int>(2.e10 / num_pixels_per_test);
+
+  // Fill up bottom-left and top-right pixels.
+  for (int i = kBPS; i < 2 * kBPS; ++i) {
+    left[i] = rnd.Rand8();
+    above[i] = rnd.Rand8();
+  }
+
+  for (int k = 0; k < num_funcs; ++k) {
+    if (pred_funcs[k] == NULL) continue;
+    memcpy(src, ref_src, sizeof(src));
+    aom_usec_timer timer;
+    aom_usec_timer_start(&timer);
+    for (int num_tests = 0; num_tests < kNumTests; ++num_tests) {
+      pred_funcs[k](src, kBPS, above, left);
+    }
+    libaom_test::ClearSystemState();
+    aom_usec_timer_mark(&timer);
+    const int elapsed_time =
+        static_cast<int>(aom_usec_timer_elapsed(&timer) / 1000);
+    libaom_test::MD5 md5;
+    md5.Add(src, sizeof(src));
+    printf("Mode %s[%12s]: %5d ms     MD5: %s\n", name, pred_func_names[k],
+           elapsed_time, md5.Get());
+    EXPECT_STREQ(signatures[k], md5.Get());
+  }
+}
+
+void TestIntraPred4(AvxPredFunc const *pred_funcs) {
+  static const char *const kSignatures[kNumAv1IntraFuncs] = {
+    "4334156168b34ab599d9b5b30f522fe9",
+    "bc4649d5ba47c7ff178d92e475960fb0",
+    "8d316e5933326dcac24e1064794b5d12",
+    "a27270fed024eafd762c95de85f4da51",
+    "c33dff000d4256c2b8f3bf9e9bab14d2",
+    "44d8cddc2ad8f79b8ed3306051722b4f",
+    "df62e96dfcb25d8a435482756a6fa990",
+    "ecb0d56ae5f677ea45127ce9d5c058e4",
+    "0b7936841f6813da818275944895b574",
+    "9117972ef64f91a58ff73e1731c81db2",
+    "46d493dccf6e5356c6f3c0c73b7dd141",
+    "b852f42e6c4991d415400332d567872f",
+#if CONFIG_ALT_INTRA
+    "828c49a4248993cce4876fa26eab697f",
+    "718c8cee9011f92ef31f77a9a7560010"
+#else
+    "309a618577b27c648f9c5ee45252bc8f",
+#endif  // CONFIG_ALT_INTRA
+  };
+  TestIntraPred("Intra4", pred_funcs, kAv1IntraPredNames, kNumAv1IntraFuncs,
+                kSignatures, 4, 4 * 4 * kNumAv1IntraFuncs);
+}
+
+void TestIntraPred8(AvxPredFunc const *pred_funcs) {
+  static const char *const kSignatures[kNumAv1IntraFuncs] = {
+    "7694ddeeefed887faf9d339d18850928",
+    "7d726b1213591b99f736be6dec65065b",
+    "19c5711281357a485591aaf9c96c0a67",
+    "ba6b66877a089e71cd938e3b8c40caac",
+    "802440c93317e0f8ba93fab02ef74265",
+    "9e09a47a15deb0b9d8372824f9805080",
+    "a2fd4b66e1a667a3e582588934a7e4bd",
+    "78339c1c60bb1d67d248ab8c4da08b7f",
+    "5c97d70f7d47de1882a6cd86c165c8a9",
+    "8182bf60688b42205acd95e59e967157",
+    "9d69fcaf12398e67242d3fcf5cf2267e",
+    "7a09adb0fa6c2bf889a99dd816622feb",
+#if CONFIG_ALT_INTRA
+    "f6ade499c626d38eb70661184b79bc57",
+    "1ad5b106c79b792e514ba25e87139b5e"
+#else
+    "815b75c8e0d91cc1ae766dc5d3e445a3",
+#endif  // CONFIG_ALT_INTRA
+  };
+  TestIntraPred("Intra8", pred_funcs, kAv1IntraPredNames, kNumAv1IntraFuncs,
+                kSignatures, 8, 8 * 8 * kNumAv1IntraFuncs);
+}
+
+void TestIntraPred16(AvxPredFunc const *pred_funcs) {
+  static const char *const kSignatures[kNumAv1IntraFuncs] = {
+    "b40dbb555d5d16a043dc361e6694fe53",
+    "fb08118cee3b6405d64c1fd68be878c6",
+    "6c190f341475c837cc38c2e566b64875",
+    "db5c34ccbe2c7f595d9b08b0dc2c698c",
+    "a62cbfd153a1f0b9fed13e62b8408a7a",
+    "143df5b4c89335e281103f610f5052e4",
+    "404944b521d16f6edd160feeeb31ff35",
+    "7841fae7d4d47b519322e6a03eeed9dc",
+    "f6ebed3f71cbcf8d6d0516ce87e11093",
+    "3cc480297dbfeed01a1c2d78dd03d0c5",
+    "fbd607f15da218c5390a5b183b634a10",
+    "f7063ccbc29f87303d5c3d0555b08944",
+#if CONFIG_ALT_INTRA
+    "7adcaaa3554eb71a81fc48cb9043984b",
+    "c0acea4397c1b4d54a21bbcec5731dff"
+#else
+    "b8a41aa968ec108af447af4217cba91b",
+#endif  // CONFIG_ALT_INTRA
+  };
+  TestIntraPred("Intra16", pred_funcs, kAv1IntraPredNames, kNumAv1IntraFuncs,
+                kSignatures, 16, 16 * 16 * kNumAv1IntraFuncs);
+}
+
+void TestIntraPred32(AvxPredFunc const *pred_funcs) {
+  static const char *const kSignatures[kNumAv1IntraFuncs] = {
+    "558541656d84f9ae7896db655826febe",
+    "b3587a1f9a01495fa38c8cd3c8e2a1bf",
+    "4c6501e64f25aacc55a2a16c7e8f0255",
+    "b3b01379ba08916ef6b1b35f7d9ad51c",
+    "0f1eb38b6cbddb3d496199ef9f329071",
+    "911c06efb9ed1c3b4c104b232b55812f",
+    "b4f9f177a8a259514f039cfb403a61e3",
+    "0a6d584a44f8db9aa7ade2e2fdb9fc9e",
+    "b01c9076525216925f3456f034fb6eee",
+    "d267e20ad9e5cd2915d1a47254d3d149",
+    "3c45418137114cb6cef4c7a7baf4855c",
+    "d520125ebd512c63c301bf67fea8e059",
+#if CONFIG_ALT_INTRA
+    "297e8fbb5d33c29b12b228fa9d7c40a4",
+    "31b9296d70dd82238c87173e6d5e65fd"
+#else
+    "9e1370c6d42e08d357d9612c93a71cfc",
+#endif  // CONFIG_ALT_INTRA
+  };
+  TestIntraPred("Intra32", pred_funcs, kAv1IntraPredNames, kNumAv1IntraFuncs,
+                kSignatures, 32, 32 * 32 * kNumAv1IntraFuncs);
+}
+
+}  // namespace
+
+// Defines a test case for |arch| (e.g., C, SSE2, ...) passing the predictors
+// to |test_func|. The test name is 'arch.test_func', e.g., C.TestIntraPred4.
+#define INTRA_PRED_TEST(arch, test_func, dc, dc_left, dc_top, dc_128, v, h, \
+                        d45e, d135, d117, d153, d207e, d63e, tm, smooth)    \
+  TEST(arch, test_func) {                                                   \
+    static const AvxPredFunc aom_intra_pred[] = {                           \
+      dc,   dc_left, dc_top, dc_128, v,    h,  d45e,                        \
+      d135, d117,    d153,   d207e,  d63e, tm, smooth                       \
+    };                                                                      \
+    test_func(aom_intra_pred);                                              \
+  }
+
+// -----------------------------------------------------------------------------
+// 4x4
+
+#if CONFIG_ALT_INTRA
+#define tm_pred_func aom_paeth_predictor_4x4_c
+#define smooth_pred_func aom_smooth_predictor_4x4_c
+#else
+#define tm_pred_func aom_tm_predictor_4x4_c
+#define smooth_pred_func NULL
+#endif  // CONFIG_ALT_INTRA
+
+INTRA_PRED_TEST(C, TestIntraPred4, aom_dc_predictor_4x4_c,
+                aom_dc_left_predictor_4x4_c, aom_dc_top_predictor_4x4_c,
+                aom_dc_128_predictor_4x4_c, aom_v_predictor_4x4_c,
+                aom_h_predictor_4x4_c, aom_d45e_predictor_4x4_c,
+                aom_d135_predictor_4x4_c, aom_d117_predictor_4x4_c,
+                aom_d153_predictor_4x4_c, aom_d207e_predictor_4x4_c,
+                aom_d63e_predictor_4x4_c, tm_pred_func, smooth_pred_func)
+
+#undef tm_pred_func
+#undef smooth_pred_func
+
+#if HAVE_SSE2
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_4x4_sse2
+#endif  // CONFIG_ALT_INTRA
+
+INTRA_PRED_TEST(SSE2, TestIntraPred4, aom_dc_predictor_4x4_sse2,
+                aom_dc_left_predictor_4x4_sse2, aom_dc_top_predictor_4x4_sse2,
+                aom_dc_128_predictor_4x4_sse2, aom_v_predictor_4x4_sse2,
+                aom_h_predictor_4x4_sse2, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+
+#undef tm_pred_func
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3
+INTRA_PRED_TEST(SSSE3, TestIntraPred4, NULL, NULL, NULL, NULL, NULL, NULL, NULL,
+                NULL, NULL, aom_d153_predictor_4x4_ssse3, NULL,
+                aom_d63e_predictor_4x4_ssse3, NULL, NULL)
+#endif  // HAVE_SSSE3
+
+#if HAVE_DSPR2
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_4x4_dspr2
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(DSPR2, TestIntraPred4, aom_dc_predictor_4x4_dspr2, NULL, NULL,
+                NULL, NULL, aom_h_predictor_4x4_dspr2, NULL, NULL, NULL, NULL,
+                NULL, NULL, tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_DSPR2
+
+#if HAVE_NEON
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_4x4_neon
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(NEON, TestIntraPred4, aom_dc_predictor_4x4_neon,
+                aom_dc_left_predictor_4x4_neon, aom_dc_top_predictor_4x4_neon,
+                aom_dc_128_predictor_4x4_neon, aom_v_predictor_4x4_neon,
+                aom_h_predictor_4x4_neon, NULL, aom_d135_predictor_4x4_neon,
+                NULL, NULL, NULL, NULL, tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_4x4_msa
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(MSA, TestIntraPred4, aom_dc_predictor_4x4_msa,
+                aom_dc_left_predictor_4x4_msa, aom_dc_top_predictor_4x4_msa,
+                aom_dc_128_predictor_4x4_msa, aom_v_predictor_4x4_msa,
+                aom_h_predictor_4x4_msa, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_MSA
+
+// -----------------------------------------------------------------------------
+// 8x8
+
+#if CONFIG_ALT_INTRA
+#define tm_pred_func aom_paeth_predictor_8x8_c
+#define smooth_pred_func aom_smooth_predictor_8x8_c
+#else
+#define tm_pred_func aom_tm_predictor_8x8_c
+#define smooth_pred_func NULL
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(C, TestIntraPred8, aom_dc_predictor_8x8_c,
+                aom_dc_left_predictor_8x8_c, aom_dc_top_predictor_8x8_c,
+                aom_dc_128_predictor_8x8_c, aom_v_predictor_8x8_c,
+                aom_h_predictor_8x8_c, aom_d45e_predictor_8x8_c,
+                aom_d135_predictor_8x8_c, aom_d117_predictor_8x8_c,
+                aom_d153_predictor_8x8_c, aom_d207e_predictor_8x8_c,
+                aom_d63e_predictor_8x8_c, tm_pred_func, smooth_pred_func)
+#undef tm_pred_func
+#undef smooth_pred_func
+
+#if HAVE_SSE2
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_8x8_sse2
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(SSE2, TestIntraPred8, aom_dc_predictor_8x8_sse2,
+                aom_dc_left_predictor_8x8_sse2, aom_dc_top_predictor_8x8_sse2,
+                aom_dc_128_predictor_8x8_sse2, aom_v_predictor_8x8_sse2,
+                aom_h_predictor_8x8_sse2, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3
+INTRA_PRED_TEST(SSSE3, TestIntraPred8, NULL, NULL, NULL, NULL, NULL, NULL, NULL,
+                NULL, NULL, aom_d153_predictor_8x8_ssse3, NULL, NULL, NULL,
+                NULL)
+#endif  // HAVE_SSSE3
+
+#if HAVE_DSPR2
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_8x8_dspr2
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(DSPR2, TestIntraPred8, aom_dc_predictor_8x8_dspr2, NULL, NULL,
+                NULL, NULL, aom_h_predictor_8x8_dspr2, NULL, NULL, NULL, NULL,
+                NULL, NULL, tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_DSPR2
+
+#if HAVE_NEON
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_8x8_neon
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(NEON, TestIntraPred8, aom_dc_predictor_8x8_neon,
+                aom_dc_left_predictor_8x8_neon, aom_dc_top_predictor_8x8_neon,
+                aom_dc_128_predictor_8x8_neon, aom_v_predictor_8x8_neon,
+                aom_h_predictor_8x8_neon, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_8x8_msa
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(MSA, TestIntraPred8, aom_dc_predictor_8x8_msa,
+                aom_dc_left_predictor_8x8_msa, aom_dc_top_predictor_8x8_msa,
+                aom_dc_128_predictor_8x8_msa, aom_v_predictor_8x8_msa,
+                aom_h_predictor_8x8_msa, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_MSA
+
+// -----------------------------------------------------------------------------
+// 16x16
+
+#if CONFIG_ALT_INTRA
+#define tm_pred_func aom_paeth_predictor_16x16_c
+#define smooth_pred_func aom_smooth_predictor_16x16_c
+#else
+#define tm_pred_func aom_tm_predictor_16x16_c
+#define smooth_pred_func NULL
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(C, TestIntraPred16, aom_dc_predictor_16x16_c,
+                aom_dc_left_predictor_16x16_c, aom_dc_top_predictor_16x16_c,
+                aom_dc_128_predictor_16x16_c, aom_v_predictor_16x16_c,
+                aom_h_predictor_16x16_c, aom_d45e_predictor_16x16_c,
+                aom_d135_predictor_16x16_c, aom_d117_predictor_16x16_c,
+                aom_d153_predictor_16x16_c, aom_d207e_predictor_16x16_c,
+                aom_d63e_predictor_16x16_c, tm_pred_func, smooth_pred_func)
+#undef tm_pred_func
+#undef smooth_pred_func
+
+#if HAVE_SSE2
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_16x16_sse2
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(SSE2, TestIntraPred16, aom_dc_predictor_16x16_sse2,
+                aom_dc_left_predictor_16x16_sse2,
+                aom_dc_top_predictor_16x16_sse2,
+                aom_dc_128_predictor_16x16_sse2, aom_v_predictor_16x16_sse2,
+                aom_h_predictor_16x16_sse2, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3
+INTRA_PRED_TEST(SSSE3, TestIntraPred16, NULL, NULL, NULL, NULL, NULL, NULL,
+                NULL, NULL, NULL, aom_d153_predictor_16x16_ssse3, NULL, NULL,
+                NULL, NULL)
+#endif  // HAVE_SSSE3
+
+#if HAVE_DSPR2
+INTRA_PRED_TEST(DSPR2, TestIntraPred16, aom_dc_predictor_16x16_dspr2, NULL,
+                NULL, NULL, NULL, aom_h_predictor_16x16_dspr2, NULL, NULL, NULL,
+                NULL, NULL, NULL, NULL, NULL)
+#endif  // HAVE_DSPR2
+
+#if HAVE_NEON
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_16x16_neon
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(NEON, TestIntraPred16, aom_dc_predictor_16x16_neon,
+                aom_dc_left_predictor_16x16_neon,
+                aom_dc_top_predictor_16x16_neon,
+                aom_dc_128_predictor_16x16_neon, aom_v_predictor_16x16_neon,
+                aom_h_predictor_16x16_neon, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_16x16_msa
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(MSA, TestIntraPred16, aom_dc_predictor_16x16_msa,
+                aom_dc_left_predictor_16x16_msa, aom_dc_top_predictor_16x16_msa,
+                aom_dc_128_predictor_16x16_msa, aom_v_predictor_16x16_msa,
+                aom_h_predictor_16x16_msa, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_MSA
+
+// -----------------------------------------------------------------------------
+// 32x32
+
+#if CONFIG_ALT_INTRA
+#define tm_pred_func aom_paeth_predictor_32x32_c
+#define smooth_pred_func aom_smooth_predictor_32x32_c
+#else
+#define tm_pred_func aom_tm_predictor_32x32_c
+#define smooth_pred_func NULL
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(C, TestIntraPred32, aom_dc_predictor_32x32_c,
+                aom_dc_left_predictor_32x32_c, aom_dc_top_predictor_32x32_c,
+                aom_dc_128_predictor_32x32_c, aom_v_predictor_32x32_c,
+                aom_h_predictor_32x32_c, aom_d45e_predictor_32x32_c,
+                aom_d135_predictor_32x32_c, aom_d117_predictor_32x32_c,
+                aom_d153_predictor_32x32_c, aom_d207e_predictor_32x32_c,
+                aom_d63e_predictor_32x32_c, tm_pred_func, smooth_pred_func)
+#undef tm_pred_func
+#undef smooth_pred_func
+
+#if HAVE_SSE2
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_32x32_sse2
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(SSE2, TestIntraPred32, aom_dc_predictor_32x32_sse2,
+                aom_dc_left_predictor_32x32_sse2,
+                aom_dc_top_predictor_32x32_sse2,
+                aom_dc_128_predictor_32x32_sse2, aom_v_predictor_32x32_sse2,
+                aom_h_predictor_32x32_sse2, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3
+INTRA_PRED_TEST(SSSE3, TestIntraPred32, NULL, NULL, NULL, NULL, NULL, NULL,
+                NULL, NULL, NULL, aom_d153_predictor_32x32_ssse3, NULL, NULL,
+                NULL, NULL)
+#endif  // HAVE_SSSE3
+
+#if HAVE_NEON
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_32x32_neon
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(NEON, TestIntraPred32, aom_dc_predictor_32x32_neon,
+                aom_dc_left_predictor_32x32_neon,
+                aom_dc_top_predictor_32x32_neon,
+                aom_dc_128_predictor_32x32_neon, aom_v_predictor_32x32_neon,
+                aom_h_predictor_32x32_neon, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+#if CONFIG_ALT_INTRA
+#define tm_pred_func NULL
+#else
+#define tm_pred_func aom_tm_predictor_32x32_msa
+#endif  // CONFIG_ALT_INTRA
+INTRA_PRED_TEST(MSA, TestIntraPred32, aom_dc_predictor_32x32_msa,
+                aom_dc_left_predictor_32x32_msa, aom_dc_top_predictor_32x32_msa,
+                aom_dc_128_predictor_32x32_msa, aom_v_predictor_32x32_msa,
+                aom_h_predictor_32x32_msa, NULL, NULL, NULL, NULL, NULL, NULL,
+                tm_pred_func, NULL)
+#undef tm_pred_func
+#endif  // HAVE_MSA
+
+#include "test/test_libaom.cc"
diff --git a/third_party/aom/test/test_libaom.cc b/third_party/aom/test/test_libaom.cc
new file mode 100644
index 0000000000..6d83ce66e7
--- /dev/null
+++ b/third_party/aom/test/test_libaom.cc
@@ -0,0 +1,67 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <string>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#if ARCH_X86 || ARCH_X86_64
+#include "aom_ports/x86.h"
+#endif
+extern "C" {
+#if CONFIG_AV1
+extern void av1_rtcd();
+#endif  // CONFIG_AV1
+extern void aom_dsp_rtcd();
+extern void aom_scale_rtcd();
+}
+
+#if ARCH_X86 || ARCH_X86_64
+static void append_negative_gtest_filter(const char *str) {
+  std::string filter = ::testing::FLAGS_gtest_filter;
+  // Negative patterns begin with one '-' followed by a ':' separated list.
+  if (filter.find('-') == std::string::npos) filter += '-';
+  filter += str;
+  ::testing::FLAGS_gtest_filter = filter;
+}
+#endif  // ARCH_X86 || ARCH_X86_64
+
+int main(int argc, char **argv) {
+  ::testing::InitGoogleTest(&argc, argv);
+
+#if ARCH_X86 || ARCH_X86_64
+  const int simd_caps = x86_simd_caps();
+  if (!(simd_caps & HAS_MMX)) append_negative_gtest_filter(":MMX.*:MMX/*");
+  if (!(simd_caps & HAS_SSE)) append_negative_gtest_filter(":SSE.*:SSE/*");
+  if (!(simd_caps & HAS_SSE2)) append_negative_gtest_filter(":SSE2.*:SSE2/*");
+  if (!(simd_caps & HAS_SSE3)) append_negative_gtest_filter(":SSE3.*:SSE3/*");
+  if (!(simd_caps & HAS_SSSE3))
+    append_negative_gtest_filter(":SSSE3.*:SSSE3/*");
+  if (!(simd_caps & HAS_SSE4_1))
+    append_negative_gtest_filter(":SSE4_1.*:SSE4_1/*");
+  if (!(simd_caps & HAS_AVX)) append_negative_gtest_filter(":AVX.*:AVX/*");
+  if (!(simd_caps & HAS_AVX2)) append_negative_gtest_filter(":AVX2.*:AVX2/*");
+#endif  // ARCH_X86 || ARCH_X86_64
+
+#if !CONFIG_SHARED
+// Shared library builds don't support whitebox tests
+// that exercise internal symbols.
+
+#if CONFIG_AV1
+  av1_rtcd();
+#endif  // CONFIG_AV1
+  aom_dsp_rtcd();
+  aom_scale_rtcd();
+#endif  // !CONFIG_SHARED
+
+  return RUN_ALL_TESTS();
+}
diff --git a/third_party/aom/test/test_worker.cmake b/third_party/aom/test/test_worker.cmake
new file mode 100644
index 0000000000..fa1d581306
--- /dev/null
+++ b/third_party/aom/test/test_worker.cmake
@@ -0,0 +1,49 @@
+##
+## Copyright (c) 2017, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+if (NOT AOM_ROOT OR NOT AOM_CONFIG_DIR)
+  message(FATAL_ERROR "AOM_ROOT AND AOM_CONFIG_DIR must be defined.")
+endif ()
+
+set(AOM_TEST_DATA_LIST "${AOM_ROOT}/test/test-data.sha1")
+set(AOM_TEST_DATA_URL "http://downloads.webmproject.org/test_data/libvpx")
+set(AOM_TEST_DATA_PATH "$ENV{LIBAOM_TEST_DATA_PATH}")
+
+include("${AOM_ROOT}/test/test_data_util.cmake")
+
+if (${AOM_TEST_DATA_PATH} STREQUAL "")
+  message(WARNING "Writing test data to ${AOM_CONFIG_DIR}, set "
+          "$LIBAOM_TEST_DATA_PATH in your environment to avoid this warning.")
+  set(AOM_TEST_DATA_PATH "${AOM_CONFIG_DIR}")
+endif ()
+
+if (NOT EXISTS "${AOM_TEST_DATA_PATH}")
+  file(MAKE_DIRECTORY "${AOM_TEST_DATA_PATH}")
+endif ()
+
+make_test_data_lists("AOM_TEST_DATA_FILES" "AOM_TEST_DATA_CHECKSUMS")
+expand_test_file_paths("AOM_TEST_DATA_FILES" "${AOM_TEST_DATA_PATH}"
+                       "AOM_TEST_DATA_FILE_PATHS")
+expand_test_file_paths("AOM_TEST_DATA_FILES" "${AOM_TEST_DATA_URL}"
+                       "AOM_TEST_DATA_URLS")
+list(LENGTH AOM_TEST_DATA_FILES num_files)
+math(EXPR num_files "${num_files} - 1")
+
+foreach (file_num RANGE ${num_files})
+  list(GET AOM_TEST_DATA_FILES ${file_num} filename)
+  list(GET AOM_TEST_DATA_CHECKSUMS ${file_num} checksum)
+  list(GET AOM_TEST_DATA_FILE_PATHS ${file_num} filepath)
+  list(GET AOM_TEST_DATA_URLS ${file_num} url)
+
+  check_file("${filepath}" "${checksum}" "needs_download")
+  if (needs_download)
+    download_test_file("${url}" "${checksum}" "${filepath}")
+  endif ()
+endforeach ()
diff --git a/third_party/aom/test/tile_independence_test.cc b/third_party/aom/test/tile_independence_test.cc
new file mode 100644
index 0000000000..a29051f2f7
--- /dev/null
+++ b/third_party/aom/test/tile_independence_test.cc
@@ -0,0 +1,147 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/i420_video_source.h"
+#include "test/util.h"
+#include "test/md5_helper.h"
+#include "aom_mem/aom_mem.h"
+
+namespace {
+class TileIndependenceTest
+    : public ::libaom_test::EncoderTest,
+      public ::libaom_test::CodecTestWith2Params<int, int> {
+ protected:
+  TileIndependenceTest()
+      : EncoderTest(GET_PARAM(0)), md5_fw_order_(), md5_inv_order_(),
+        n_tile_cols_(GET_PARAM(1)), n_tile_rows_(GET_PARAM(2)) {
+    init_flags_ = AOM_CODEC_USE_PSNR;
+    aom_codec_dec_cfg_t cfg = aom_codec_dec_cfg_t();
+    cfg.w = 704;
+    cfg.h = 144;
+    cfg.threads = 1;
+    fw_dec_ = codec_->CreateDecoder(cfg, 0);
+    inv_dec_ = codec_->CreateDecoder(cfg, 0);
+    inv_dec_->Control(AV1_INVERT_TILE_DECODE_ORDER, 1);
+
+#if CONFIG_AV1 && CONFIG_EXT_TILE
+    if (fw_dec_->IsAV1() && inv_dec_->IsAV1()) {
+      fw_dec_->Control(AV1_SET_DECODE_TILE_ROW, -1);
+      fw_dec_->Control(AV1_SET_DECODE_TILE_COL, -1);
+      inv_dec_->Control(AV1_SET_DECODE_TILE_ROW, -1);
+      inv_dec_->Control(AV1_SET_DECODE_TILE_COL, -1);
+    }
+#endif
+  }
+
+  virtual ~TileIndependenceTest() {
+    delete fw_dec_;
+    delete inv_dec_;
+  }
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(libaom_test::kTwoPassGood);
+  }
+
+  virtual void PreEncodeFrameHook(libaom_test::VideoSource *video,
+                                  libaom_test::Encoder *encoder) {
+    if (video->frame() == 1) {
+      encoder->Control(AV1E_SET_TILE_COLUMNS, n_tile_cols_);
+      encoder->Control(AV1E_SET_TILE_ROWS, n_tile_rows_);
+#if CONFIG_EXT_TILE
+      encoder->Control(AV1E_SET_TILE_ENCODING_MODE, 0);  // TILE_NORMAL
+#endif                                                   // CONFIG_EXT_TILE
+#if CONFIG_LOOPFILTERING_ACROSS_TILES
+      encoder->Control(AV1E_SET_TILE_LOOPFILTER, 0);
+#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
+      SetCpuUsed(encoder);
+    }
+  }
+
+  virtual void SetCpuUsed(libaom_test::Encoder *encoder) {
+    static const int kCpuUsed = 3;
+    encoder->Control(AOME_SET_CPUUSED, kCpuUsed);
+  }
+
+  void UpdateMD5(::libaom_test::Decoder *dec, const aom_codec_cx_pkt_t *pkt,
+                 ::libaom_test::MD5 *md5) {
+    const aom_codec_err_t res = dec->DecodeFrame(
+        reinterpret_cast<uint8_t *>(pkt->data.frame.buf), pkt->data.frame.sz);
+    if (res != AOM_CODEC_OK) {
+      abort_ = true;
+      ASSERT_EQ(AOM_CODEC_OK, res);
+    }
+    const aom_image_t *img = dec->GetDxData().Next();
+    md5->Add(img);
+  }
+
+  virtual void FramePktHook(const aom_codec_cx_pkt_t *pkt) {
+    UpdateMD5(fw_dec_, pkt, &md5_fw_order_);
+    UpdateMD5(inv_dec_, pkt, &md5_inv_order_);
+  }
+
+  void DoTest() {
+    const aom_rational timebase = { 33333333, 1000000000 };
+    cfg_.g_timebase = timebase;
+    cfg_.rc_target_bitrate = 500;
+    cfg_.g_lag_in_frames = 12;
+    cfg_.rc_end_usage = AOM_VBR;
+
+    libaom_test::I420VideoSource video("hantro_collage_w352h288.yuv", 704, 576,
+                                       timebase.den, timebase.num, 0, 5);
+    ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+
+    const char *md5_fw_str = md5_fw_order_.Get();
+    const char *md5_inv_str = md5_inv_order_.Get();
+    ASSERT_STREQ(md5_fw_str, md5_inv_str);
+  }
+
+  ::libaom_test::MD5 md5_fw_order_, md5_inv_order_;
+  ::libaom_test::Decoder *fw_dec_, *inv_dec_;
+
+ private:
+  int n_tile_cols_;
+  int n_tile_rows_;
+};
+
+// run an encode with 2 or 4 tiles, and do the decode both in normal and
+// inverted tile ordering. Ensure that the MD5 of the output in both cases
+// is identical. If so, tiles are considered independent and the test passes.
+TEST_P(TileIndependenceTest, MD5Match) { DoTest(); }
+
+class TileIndependenceTestLarge : public TileIndependenceTest {
+  virtual void SetCpuUsed(libaom_test::Encoder *encoder) {
+    static const int kCpuUsed = 0;
+    encoder->Control(AOME_SET_CPUUSED, kCpuUsed);
+  }
+};
+
+TEST_P(TileIndependenceTestLarge, MD5Match) { DoTest(); }
+
+#if CONFIG_EXT_TILE
+AV1_INSTANTIATE_TEST_CASE(TileIndependenceTest, ::testing::Values(1, 2, 32),
+                          ::testing::Values(1, 2, 32));
+AV1_INSTANTIATE_TEST_CASE(TileIndependenceTestLarge,
+                          ::testing::Values(1, 2, 32),
+                          ::testing::Values(1, 2, 32));
+#else
+AV1_INSTANTIATE_TEST_CASE(TileIndependenceTest, ::testing::Values(0, 1),
+                          ::testing::Values(0, 1));
+AV1_INSTANTIATE_TEST_CASE(TileIndependenceTestLarge, ::testing::Values(0, 1),
+                          ::testing::Values(0, 1));
+#endif  // CONFIG_EXT_TILE
+}  // namespace
diff --git a/third_party/aom/test/tools_common.sh b/third_party/aom/test/tools_common.sh
new file mode 100755
index 0000000000..254e6b2960
--- /dev/null
+++ b/third_party/aom/test/tools_common.sh
@@ -0,0 +1,454 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+##  This file contains shell code shared by test scripts for libaom tools.
+
+# Use $AOM_TEST_TOOLS_COMMON_SH as a pseudo include guard.
+if [ -z "${AOM_TEST_TOOLS_COMMON_SH}" ]; then
+AOM_TEST_TOOLS_COMMON_SH=included
+
+set -e
+devnull='> /dev/null 2>&1'
+AOM_TEST_PREFIX=""
+
+elog() {
+  echo "$@" 1>&2
+}
+
+vlog() {
+  if [ "${AOM_TEST_VERBOSE_OUTPUT}" = "yes" ]; then
+    echo "$@"
+  fi
+}
+
+# Sets $AOM_TOOL_TEST to the name specified by positional parameter one.
+test_begin() {
+  AOM_TOOL_TEST="${1}"
+}
+
+# Clears the AOM_TOOL_TEST variable after confirming that $AOM_TOOL_TEST matches
+# positional parameter one.
+test_end() {
+  if [ "$1" != "${AOM_TOOL_TEST}" ]; then
+    echo "FAIL completed test mismatch!."
+    echo "  completed test: ${1}"
+    echo "  active test: ${AOM_TOOL_TEST}."
+    return 1
+  fi
+  AOM_TOOL_TEST='<unset>'
+}
+
+# Echoes the target configuration being tested.
+test_configuration_target() {
+  aom_config_mk="${LIBAOM_CONFIG_PATH}/config.mk"
+  # Find the TOOLCHAIN line, split it using ':=' as the field separator, and
+  # print the last field to get the value. Then pipe the value to tr to consume
+  # any leading/trailing spaces while allowing tr to echo the output to stdout.
+  awk -F ':=' '/TOOLCHAIN/ { print $NF }' "${aom_config_mk}" | tr -d ' '
+}
+
+# Trap function used for failure reports and tool output directory removal.
+# When the contents of $AOM_TOOL_TEST do not match the string '<unset>', reports
+# failure of test stored in $AOM_TOOL_TEST.
+cleanup() {
+  if [ -n "${AOM_TOOL_TEST}" ] && [ "${AOM_TOOL_TEST}" != '<unset>' ]; then
+    echo "FAIL: $AOM_TOOL_TEST"
+  fi
+  if [ -n "${AOM_TEST_OUTPUT_DIR}" ] && [ -d "${AOM_TEST_OUTPUT_DIR}" ]; then
+    rm -rf "${AOM_TEST_OUTPUT_DIR}"
+  fi
+}
+
+# Echoes the git hash portion of the VERSION_STRING variable defined in
+# $LIBAOM_CONFIG_PATH/config.mk to stdout, or the version number string when
+# no git hash is contained in VERSION_STRING.
+config_hash() {
+  aom_config_mk="${LIBAOM_CONFIG_PATH}/config.mk"
+  # Find VERSION_STRING line, split it with "-g" and print the last field to
+  # output the git hash to stdout.
+  aom_version=$(awk -F -g '/VERSION_STRING/ {print $NF}' "${aom_config_mk}")
+  # Handle two situations here:
+  # 1. The default case: $aom_version is a git hash, so echo it unchanged.
+  # 2. When being run a non-dev tree, the -g portion is not present in the
+  #    version string: It's only the version number.
+  #    In this case $aom_version is something like 'VERSION_STRING=v1.3.0', so
+  #    we echo only what is after the '='.
+  echo "${aom_version##*=}"
+}
+
+# Echoes the short form of the current git hash.
+current_hash() {
+  if git --version > /dev/null 2>&1; then
+    (cd "$(dirname "${0}")"
+    git rev-parse --short HEAD)
+  else
+    # Return the config hash if git is unavailable: Fail silently, git hashes
+    # are used only for warnings.
+    config_hash
+  fi
+}
+
+# Echoes warnings to stdout when git hash in aom_config.h does not match the
+# current git hash.
+check_git_hashes() {
+  hash_at_configure_time=$(config_hash)
+  hash_now=$(current_hash)
+
+  if [ "${hash_at_configure_time}" != "${hash_now}" ]; then
+    echo "Warning: git hash has changed since last configure."
+  fi
+}
+
+# $1 is the name of an environment variable containing a directory name to
+# test.
+test_env_var_dir() {
+  local dir=$(eval echo "\${$1}")
+  if [ ! -d "${dir}" ]; then
+    elog "'${dir}': No such directory"
+    elog "The $1 environment variable must be set to a valid directory."
+    return 1
+  fi
+}
+
+# This script requires that the LIBAOM_BIN_PATH, LIBAOM_CONFIG_PATH, and
+# LIBAOM_TEST_DATA_PATH variables are in the environment: Confirm that
+# the variables are set and that they all evaluate to directory paths.
+verify_aom_test_environment() {
+  test_env_var_dir "LIBAOM_BIN_PATH" \
+    && test_env_var_dir "LIBAOM_CONFIG_PATH" \
+    && test_env_var_dir "LIBAOM_TEST_DATA_PATH"
+}
+
+# Greps aom_config.h in LIBAOM_CONFIG_PATH for positional parameter one, which
+# should be a LIBAOM preprocessor flag. Echoes yes to stdout when the feature
+# is available.
+aom_config_option_enabled() {
+  aom_config_option="${1}"
+  aom_config_file="${LIBAOM_CONFIG_PATH}/aom_config.h"
+  config_line=$(grep "${aom_config_option}" "${aom_config_file}")
+  if echo "${config_line}" | egrep -q '1$'; then
+    echo yes
+  fi
+}
+
+# Echoes yes when output of test_configuration_target() contains win32 or win64.
+is_windows_target() {
+  if test_configuration_target \
+     | grep -q -e win32 -e win64 > /dev/null 2>&1; then
+    echo yes
+  fi
+}
+
+# Echoes path to $1 when it's executable and exists in ${LIBAOM_BIN_PATH}, or an
+# empty string. Caller is responsible for testing the string once the function
+# returns.
+aom_tool_path() {
+  local readonly tool_name="$1"
+  local tool_path="${LIBAOM_BIN_PATH}/${tool_name}${AOM_TEST_EXE_SUFFIX}"
+  if [ ! -x "${tool_path}" ]; then
+    # Try one directory up: when running via examples.sh the tool could be in
+    # the parent directory of $LIBAOM_BIN_PATH.
+    tool_path="${LIBAOM_BIN_PATH}/../${tool_name}${AOM_TEST_EXE_SUFFIX}"
+  fi
+
+  if [ ! -x "${tool_path}" ]; then
+    tool_path=""
+  fi
+  echo "${tool_path}"
+}
+
+# Echoes yes to stdout when the file named by positional parameter one exists
+# in LIBAOM_BIN_PATH, and is executable.
+aom_tool_available() {
+  local tool_name="$1"
+  local tool="${LIBAOM_BIN_PATH}/${tool_name}${AOM_TEST_EXE_SUFFIX}"
+  [ -x "${tool}" ] && echo yes
+}
+
+# Echoes yes to stdout when aom_config_option_enabled() reports yes for
+# CONFIG_AV1_DECODER.
+av1_decode_available() {
+  [ "$(aom_config_option_enabled CONFIG_AV1_DECODER)" = "yes" ] && echo yes
+}
+
+# Echoes yes to stdout when aom_config_option_enabled() reports yes for
+# CONFIG_AV1_ENCODER.
+av1_encode_available() {
+  [ "$(aom_config_option_enabled CONFIG_AV1_ENCODER)" = "yes" ] && echo yes
+}
+
+# Echoes yes to stdout when aom_config_option_enabled() reports yes for
+# CONFIG_WEBM_IO.
+webm_io_available() {
+  [ "$(aom_config_option_enabled CONFIG_WEBM_IO)" = "yes" ] && echo yes
+}
+
+# Filters strings from $1 using the filter specified by $2. Filter behavior
+# depends on the presence of $3. When $3 is present, strings that match the
+# filter are excluded. When $3 is omitted, strings matching the filter are
+# included.
+# The filtered result is echoed to stdout.
+filter_strings() {
+  strings=${1}
+  filter=${2}
+  exclude=${3}
+
+  if [ -n "${exclude}" ]; then
+    # When positional parameter three exists the caller wants to remove strings.
+    # Tell grep to invert matches using the -v argument.
+    exclude='-v'
+  else
+    unset exclude
+  fi
+
+  if [ -n "${filter}" ]; then
+    for s in ${strings}; do
+      if echo "${s}" | egrep -q ${exclude} "${filter}" > /dev/null 2>&1; then
+        filtered_strings="${filtered_strings} ${s}"
+      fi
+    done
+  else
+    filtered_strings="${strings}"
+  fi
+  echo "${filtered_strings}"
+}
+
+# Runs user test functions passed via positional parameters one and two.
+# Functions in positional parameter one are treated as environment verification
+# functions and are run unconditionally. Functions in positional parameter two
+# are run according to the rules specified in aom_test_usage().
+run_tests() {
+  local env_tests="verify_aom_test_environment $1"
+  local tests_to_filter="$2"
+  local test_name="${AOM_TEST_NAME}"
+
+  if [ -z "${test_name}" ]; then
+    test_name="$(basename "${0%.*}")"
+  fi
+
+  if [ "${AOM_TEST_RUN_DISABLED_TESTS}" != "yes" ]; then
+    # Filter out DISABLED tests.
+    tests_to_filter=$(filter_strings "${tests_to_filter}" ^DISABLED exclude)
+  fi
+
+  if [ -n "${AOM_TEST_FILTER}" ]; then
+    # Remove tests not matching the user's filter.
+    tests_to_filter=$(filter_strings "${tests_to_filter}" ${AOM_TEST_FILTER})
+  fi
+
+  # User requested test listing: Dump test names and return.
+  if [ "${AOM_TEST_LIST_TESTS}" = "yes" ]; then
+    for test_name in $tests_to_filter; do
+      echo ${test_name}
+    done
+    return
+  fi
+
+  # Don't bother with the environment tests if everything else was disabled.
+  [ -z "${tests_to_filter}" ] && return
+
+  # Combine environment and actual tests.
+  local tests_to_run="${env_tests} ${tests_to_filter}"
+
+  check_git_hashes
+
+  # Run tests.
+  for test in ${tests_to_run}; do
+    test_begin "${test}"
+    vlog "  RUN  ${test}"
+    "${test}"
+    vlog "  PASS ${test}"
+    test_end "${test}"
+  done
+
+  local tested_config="$(test_configuration_target) @ $(current_hash)"
+  echo "${test_name}: Done, all tests pass for ${tested_config}."
+}
+
+aom_test_usage() {
+cat << EOF
+  Usage: ${0##*/} [arguments]
+    --bin-path <path to libaom binaries directory>
+    --config-path <path to libaom config directory>
+    --filter <filter>: User test filter. Only tests matching filter are run.
+    --run-disabled-tests: Run disabled tests.
+    --help: Display this message and exit.
+    --test-data-path <path to libaom test data directory>
+    --show-program-output: Shows output from all programs being tested.
+    --prefix: Allows for a user specified prefix to be inserted before all test
+              programs. Grants the ability, for example, to run test programs
+              within valgrind.
+    --list-tests: List all test names and exit without actually running tests.
+    --verbose: Verbose output.
+
+    When the --bin-path option is not specified the script attempts to use
+    \$LIBAOM_BIN_PATH and then the current directory.
+
+    When the --config-path option is not specified the script attempts to use
+    \$LIBAOM_CONFIG_PATH and then the current directory.
+
+    When the -test-data-path option is not specified the script attempts to use
+    \$LIBAOM_TEST_DATA_PATH and then the current directory.
+EOF
+}
+
+# Returns non-zero (failure) when required environment variables are empty
+# strings.
+aom_test_check_environment() {
+  if [ -z "${LIBAOM_BIN_PATH}" ] || \
+     [ -z "${LIBAOM_CONFIG_PATH}" ] || \
+     [ -z "${LIBAOM_TEST_DATA_PATH}" ]; then
+    return 1
+  fi
+}
+
+# Echo aomenc command line parameters allowing use of a raw yuv file as
+# input to aomenc.
+yuv_raw_input() {
+  echo ""${YUV_RAW_INPUT}"
+       --width="${YUV_RAW_INPUT_WIDTH}"
+       --height="${YUV_RAW_INPUT_HEIGHT}""
+}
+
+# Do a small encode for testing decoders.
+encode_yuv_raw_input_av1() {
+  if [ "$(av1_encode_available)" = "yes" ]; then
+    local readonly output="$1"
+    local readonly encoder="$(aom_tool_path aomenc)"
+    shift
+    eval "${encoder}" $(yuv_raw_input) \
+      --codec=av1 \
+      $@ \
+      --limit=5 \
+      --output="${output}" \
+      ${devnull}
+
+    if [ ! -e "${output}" ]; then
+      elog "Output file does not exist."
+      return 1
+    fi
+  fi
+}
+
+# Parse the command line.
+while [ -n "$1" ]; do
+  case "$1" in
+    --bin-path)
+      LIBAOM_BIN_PATH="$2"
+      shift
+      ;;
+    --config-path)
+      LIBAOM_CONFIG_PATH="$2"
+      shift
+      ;;
+    --filter)
+      AOM_TEST_FILTER="$2"
+      shift
+      ;;
+    --run-disabled-tests)
+      AOM_TEST_RUN_DISABLED_TESTS=yes
+      ;;
+    --help)
+      aom_test_usage
+      exit
+      ;;
+    --test-data-path)
+      LIBAOM_TEST_DATA_PATH="$2"
+      shift
+      ;;
+    --prefix)
+      AOM_TEST_PREFIX="$2"
+      shift
+      ;;
+    --verbose)
+      AOM_TEST_VERBOSE_OUTPUT=yes
+      ;;
+    --show-program-output)
+      devnull=
+      ;;
+    --list-tests)
+      AOM_TEST_LIST_TESTS=yes
+      ;;
+    *)
+      aom_test_usage
+      exit 1
+      ;;
+  esac
+  shift
+done
+
+# Handle running the tests from a build directory without arguments when running
+# the tests on *nix/macosx.
+LIBAOM_BIN_PATH="${LIBAOM_BIN_PATH:-.}"
+LIBAOM_CONFIG_PATH="${LIBAOM_CONFIG_PATH:-.}"
+LIBAOM_TEST_DATA_PATH="${LIBAOM_TEST_DATA_PATH:-.}"
+
+# Create a temporary directory for output files, and a trap to clean it up.
+if [ -n "${TMPDIR}" ]; then
+  AOM_TEST_TEMP_ROOT="${TMPDIR}"
+elif [ -n "${TEMPDIR}" ]; then
+  AOM_TEST_TEMP_ROOT="${TEMPDIR}"
+else
+  AOM_TEST_TEMP_ROOT=/tmp
+fi
+
+AOM_TEST_OUTPUT_DIR="${AOM_TEST_TEMP_ROOT}/aom_test_$$"
+
+if ! mkdir -p "${AOM_TEST_OUTPUT_DIR}" || \
+   [ ! -d "${AOM_TEST_OUTPUT_DIR}" ]; then
+  echo "${0##*/}: Cannot create output directory, giving up."
+  echo "${0##*/}:   AOM_TEST_OUTPUT_DIR=${AOM_TEST_OUTPUT_DIR}"
+  exit 1
+fi
+
+if [ "$(is_windows_target)" = "yes" ]; then
+  AOM_TEST_EXE_SUFFIX=".exe"
+fi
+
+# Variables shared by tests.
+VP8_IVF_FILE="${LIBAOM_TEST_DATA_PATH}/vp80-00-comprehensive-001.ivf"
+AV1_IVF_FILE="${LIBAOM_TEST_DATA_PATH}/vp90-2-09-subpixel-00.ivf"
+
+AV1_WEBM_FILE="${LIBAOM_TEST_DATA_PATH}/vp90-2-00-quantizer-00.webm"
+AV1_FPM_WEBM_FILE="${LIBAOM_TEST_DATA_PATH}/vp90-2-07-frame_parallel-1.webm"
+AV1_LT_50_FRAMES_WEBM_FILE="${LIBAOM_TEST_DATA_PATH}/vp90-2-02-size-32x08.webm"
+
+YUV_RAW_INPUT="${LIBAOM_TEST_DATA_PATH}/hantro_collage_w352h288.yuv"
+YUV_RAW_INPUT_WIDTH=352
+YUV_RAW_INPUT_HEIGHT=288
+
+Y4M_NOSQ_PAR_INPUT="${LIBAOM_TEST_DATA_PATH}/park_joy_90p_8_420_a10-1.y4m"
+Y4M_720P_INPUT="${LIBAOM_TEST_DATA_PATH}/niklas_1280_720_30.y4m"
+
+# Setup a trap function to clean up after tests complete.
+trap cleanup EXIT
+
+vlog "$(basename "${0%.*}") test configuration:
+  LIBAOM_BIN_PATH=${LIBAOM_BIN_PATH}
+  LIBAOM_CONFIG_PATH=${LIBAOM_CONFIG_PATH}
+  LIBAOM_TEST_DATA_PATH=${LIBAOM_TEST_DATA_PATH}
+  AOM_IVF_FILE=${AOM_IVF_FILE}
+  AV1_IVF_FILE=${AV1_IVF_FILE}
+  AV1_WEBM_FILE=${AV1_WEBM_FILE}
+  AOM_TEST_EXE_SUFFIX=${AOM_TEST_EXE_SUFFIX}
+  AOM_TEST_FILTER=${AOM_TEST_FILTER}
+  AOM_TEST_LIST_TESTS=${AOM_TEST_LIST_TESTS}
+  AOM_TEST_OUTPUT_DIR=${AOM_TEST_OUTPUT_DIR}
+  AOM_TEST_PREFIX=${AOM_TEST_PREFIX}
+  AOM_TEST_RUN_DISABLED_TESTS=${AOM_TEST_RUN_DISABLED_TESTS}
+  AOM_TEST_SHOW_PROGRAM_OUTPUT=${AOM_TEST_SHOW_PROGRAM_OUTPUT}
+  AOM_TEST_TEMP_ROOT=${AOM_TEST_TEMP_ROOT}
+  AOM_TEST_VERBOSE_OUTPUT=${AOM_TEST_VERBOSE_OUTPUT}
+  YUV_RAW_INPUT=${YUV_RAW_INPUT}
+  YUV_RAW_INPUT_WIDTH=${YUV_RAW_INPUT_WIDTH}
+  YUV_RAW_INPUT_HEIGHT=${YUV_RAW_INPUT_HEIGHT}
+  Y4M_NOSQ_PAR_INPUT=${Y4M_NOSQ_PAR_INPUT}"
+
+fi  # End $AOM_TEST_TOOLS_COMMON_SH pseudo include guard.
diff --git a/third_party/aom/test/transform_test_base.h b/third_party/aom/test/transform_test_base.h
new file mode 100644
index 0000000000..4c1a554968
--- /dev/null
+++ b/third_party/aom/test/transform_test_base.h
@@ -0,0 +1,367 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef TEST_TRANSFORM_TEST_BASE_H_
+#define TEST_TRANSFORM_TEST_BASE_H_
+
+#include "./aom_config.h"
+#include "aom_mem/aom_mem.h"
+#include "aom/aom_codec.h"
+
+namespace libaom_test {
+
+//  Note:
+//   Same constant are defined in av1/common/av1_entropy.h and
+//   av1/common/entropy.h.  Goal is to make this base class
+//   to use for future codec transform testing.  But including
+//   either of them would lead to compiling error when we do
+//   unit test for another codec. Suggest to move the definition
+//   to a aom header file.
+const int kDctMaxValue = 16384;
+
+typedef void (*FhtFunc)(const int16_t *in, tran_low_t *out, int stride,
+                        int tx_type);
+
+typedef void (*IhtFunc)(const tran_low_t *in, uint8_t *out, int stride,
+                        int tx_type);
+
+class TransformTestBase {
+ public:
+  virtual ~TransformTestBase() {}
+
+ protected:
+  virtual void RunFwdTxfm(const int16_t *in, tran_low_t *out, int stride) = 0;
+
+  virtual void RunInvTxfm(const tran_low_t *out, uint8_t *dst, int stride) = 0;
+
+  void RunAccuracyCheck(uint32_t ref_max_error, double ref_avg_error) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    uint32_t max_error = 0;
+    int64_t total_error = 0;
+    const int count_test_block = 10000;
+
+    int16_t *test_input_block = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(int16_t) * num_coeffs_));
+    tran_low_t *test_temp_block = reinterpret_cast<tran_low_t *>(
+        aom_memalign(16, sizeof(tran_low_t) * num_coeffs_));
+    uint8_t *dst = reinterpret_cast<uint8_t *>(
+        aom_memalign(16, sizeof(uint8_t) * num_coeffs_));
+    uint8_t *src = reinterpret_cast<uint8_t *>(
+        aom_memalign(16, sizeof(uint8_t) * num_coeffs_));
+#if CONFIG_HIGHBITDEPTH
+    uint16_t *dst16 = reinterpret_cast<uint16_t *>(
+        aom_memalign(16, sizeof(uint16_t) * num_coeffs_));
+    uint16_t *src16 = reinterpret_cast<uint16_t *>(
+        aom_memalign(16, sizeof(uint16_t) * num_coeffs_));
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-255, 255].
+      for (int j = 0; j < num_coeffs_; ++j) {
+        if (bit_depth_ == AOM_BITS_8) {
+          src[j] = rnd.Rand8();
+          dst[j] = rnd.Rand8();
+          test_input_block[j] = src[j] - dst[j];
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand16() & mask_;
+          dst16[j] = rnd.Rand16() & mask_;
+          test_input_block[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(test_input_block, test_temp_block, pitch_));
+      if (bit_depth_ == AOM_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(test_temp_block, dst, pitch_));
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(test_temp_block, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < num_coeffs_; ++j) {
+#if CONFIG_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == AOM_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        ASSERT_EQ(AOM_BITS_8, bit_depth_);
+        const int diff = dst[j] - src[j];
+#endif
+        const uint32_t error = diff * diff;
+        if (max_error < error) max_error = error;
+        total_error += error;
+      }
+    }
+
+    double avg_error = total_error * 1. / count_test_block / num_coeffs_;
+
+    EXPECT_GE(ref_max_error, max_error)
+        << "Error: FHT/IHT has an individual round trip error > "
+        << ref_max_error;
+
+    EXPECT_GE(ref_avg_error, avg_error)
+        << "Error: FHT/IHT has average round trip error > " << ref_avg_error
+        << " per block";
+
+    aom_free(test_input_block);
+    aom_free(test_temp_block);
+    aom_free(dst);
+    aom_free(src);
+#if CONFIG_HIGHBITDEPTH
+    aom_free(dst16);
+    aom_free(src16);
+#endif
+  }
+
+  void RunCoeffCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 5000;
+
+    // Use a stride value which is not the width of any transform, to catch
+    // cases where the transforms use the stride incorrectly.
+    int stride = 96;
+
+    int16_t *input_block = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(int16_t) * stride * height_));
+    tran_low_t *output_ref_block = reinterpret_cast<tran_low_t *>(
+        aom_memalign(16, sizeof(tran_low_t) * num_coeffs_));
+    tran_low_t *output_block = reinterpret_cast<tran_low_t *>(
+        aom_memalign(16, sizeof(tran_low_t) * num_coeffs_));
+
+    for (int i = 0; i < count_test_block; ++i) {
+      int j, k;
+      for (j = 0; j < height_; ++j) {
+        for (k = 0; k < pitch_; ++k) {
+          int in_idx = j * stride + k;
+          int out_idx = j * pitch_ + k;
+          input_block[in_idx] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+          if (bit_depth_ == AOM_BITS_8) {
+            output_block[out_idx] = output_ref_block[out_idx] = rnd.Rand8();
+#if CONFIG_HIGHBITDEPTH
+          } else {
+            output_block[out_idx] = output_ref_block[out_idx] =
+                rnd.Rand16() & mask_;
+#endif
+          }
+        }
+      }
+
+      fwd_txfm_ref(input_block, output_ref_block, stride, tx_type_);
+      ASM_REGISTER_STATE_CHECK(RunFwdTxfm(input_block, output_block, stride));
+
+      // The minimum quant value is 4.
+      for (j = 0; j < height_; ++j) {
+        for (k = 0; k < pitch_; ++k) {
+          int out_idx = j * pitch_ + k;
+          ASSERT_EQ(output_block[out_idx], output_ref_block[out_idx])
+              << "Error: not bit-exact result at index: " << out_idx
+              << " at test block: " << i;
+        }
+      }
+    }
+    aom_free(input_block);
+    aom_free(output_ref_block);
+    aom_free(output_block);
+  }
+
+  void RunInvCoeffCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 5000;
+
+    // Use a stride value which is not the width of any transform, to catch
+    // cases where the transforms use the stride incorrectly.
+    int stride = 96;
+
+    int16_t *input_block = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(int16_t) * num_coeffs_));
+    tran_low_t *trans_block = reinterpret_cast<tran_low_t *>(
+        aom_memalign(16, sizeof(tran_low_t) * num_coeffs_));
+    uint8_t *output_block = reinterpret_cast<uint8_t *>(
+        aom_memalign(16, sizeof(uint8_t) * stride * height_));
+    uint8_t *output_ref_block = reinterpret_cast<uint8_t *>(
+        aom_memalign(16, sizeof(uint8_t) * stride * height_));
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      int j, k;
+      for (j = 0; j < height_; ++j) {
+        for (k = 0; k < pitch_; ++k) {
+          int in_idx = j * pitch_ + k;
+          int out_idx = j * stride + k;
+          input_block[in_idx] = (rnd.Rand16() & mask_) - (rnd.Rand16() & mask_);
+          output_ref_block[out_idx] = rnd.Rand16() & mask_;
+          output_block[out_idx] = output_ref_block[out_idx];
+        }
+      }
+
+      fwd_txfm_ref(input_block, trans_block, pitch_, tx_type_);
+
+      inv_txfm_ref(trans_block, output_ref_block, stride, tx_type_);
+      ASM_REGISTER_STATE_CHECK(RunInvTxfm(trans_block, output_block, stride));
+
+      for (j = 0; j < height_; ++j) {
+        for (k = 0; k < pitch_; ++k) {
+          int out_idx = j * stride + k;
+          ASSERT_EQ(output_block[out_idx], output_ref_block[out_idx])
+              << "Error: not bit-exact result at index: " << out_idx
+              << " j = " << j << " k = " << k << " at test block: " << i;
+        }
+      }
+    }
+    aom_free(input_block);
+    aom_free(trans_block);
+    aom_free(output_ref_block);
+    aom_free(output_block);
+  }
+
+  void RunMemCheck() {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 5000;
+
+    int16_t *input_extreme_block = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(int16_t) * num_coeffs_));
+    tran_low_t *output_ref_block = reinterpret_cast<tran_low_t *>(
+        aom_memalign(16, sizeof(tran_low_t) * num_coeffs_));
+    tran_low_t *output_block = reinterpret_cast<tran_low_t *>(
+        aom_memalign(16, sizeof(tran_low_t) * num_coeffs_));
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < num_coeffs_; ++j) {
+        input_extreme_block[j] = rnd.Rand8() % 2 ? mask_ : -mask_;
+      }
+      if (i == 0) {
+        for (int j = 0; j < num_coeffs_; ++j) input_extreme_block[j] = mask_;
+      } else if (i == 1) {
+        for (int j = 0; j < num_coeffs_; ++j) input_extreme_block[j] = -mask_;
+      }
+
+      fwd_txfm_ref(input_extreme_block, output_ref_block, pitch_, tx_type_);
+      ASM_REGISTER_STATE_CHECK(
+          RunFwdTxfm(input_extreme_block, output_block, pitch_));
+
+      int row_length = FindRowLength();
+      // The minimum quant value is 4.
+      for (int j = 0; j < num_coeffs_; ++j) {
+        EXPECT_EQ(output_block[j], output_ref_block[j])
+            << "Not bit-exact at test index: " << i << ", "
+            << "j = " << j << std::endl;
+        EXPECT_GE(row_length * kDctMaxValue << (bit_depth_ - 8),
+                  abs(output_block[j]))
+            << "Error: NxN FDCT has coefficient larger than N*DCT_MAX_VALUE";
+      }
+    }
+    aom_free(input_extreme_block);
+    aom_free(output_ref_block);
+    aom_free(output_block);
+  }
+
+  void RunInvAccuracyCheck(int limit) {
+    ACMRandom rnd(ACMRandom::DeterministicSeed());
+    const int count_test_block = 1000;
+
+    int16_t *in = reinterpret_cast<int16_t *>(
+        aom_memalign(16, sizeof(int16_t) * num_coeffs_));
+    tran_low_t *coeff = reinterpret_cast<tran_low_t *>(
+        aom_memalign(16, sizeof(tran_low_t) * num_coeffs_));
+    uint8_t *dst = reinterpret_cast<uint8_t *>(
+        aom_memalign(16, sizeof(uint8_t) * num_coeffs_));
+    uint8_t *src = reinterpret_cast<uint8_t *>(
+        aom_memalign(16, sizeof(uint8_t) * num_coeffs_));
+
+#if CONFIG_HIGHBITDEPTH
+    uint16_t *dst16 = reinterpret_cast<uint16_t *>(
+        aom_memalign(16, sizeof(uint16_t) * num_coeffs_));
+    uint16_t *src16 = reinterpret_cast<uint16_t *>(
+        aom_memalign(16, sizeof(uint16_t) * num_coeffs_));
+#endif
+
+    for (int i = 0; i < count_test_block; ++i) {
+      // Initialize a test block with input range [-mask_, mask_].
+      for (int j = 0; j < num_coeffs_; ++j) {
+        if (bit_depth_ == AOM_BITS_8) {
+          src[j] = rnd.Rand8();
+          dst[j] = rnd.Rand8();
+          in[j] = src[j] - dst[j];
+#if CONFIG_HIGHBITDEPTH
+        } else {
+          src16[j] = rnd.Rand16() & mask_;
+          dst16[j] = rnd.Rand16() & mask_;
+          in[j] = src16[j] - dst16[j];
+#endif
+        }
+      }
+
+      fwd_txfm_ref(in, coeff, pitch_, tx_type_);
+
+      if (bit_depth_ == AOM_BITS_8) {
+        ASM_REGISTER_STATE_CHECK(RunInvTxfm(coeff, dst, pitch_));
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        ASM_REGISTER_STATE_CHECK(
+            RunInvTxfm(coeff, CONVERT_TO_BYTEPTR(dst16), pitch_));
+#endif
+      }
+
+      for (int j = 0; j < num_coeffs_; ++j) {
+#if CONFIG_HIGHBITDEPTH
+        const int diff =
+            bit_depth_ == AOM_BITS_8 ? dst[j] - src[j] : dst16[j] - src16[j];
+#else
+        const int diff = dst[j] - src[j];
+#endif
+        const uint32_t error = diff * diff;
+        EXPECT_GE(static_cast<uint32_t>(limit), error)
+            << "Error: 4x4 IDCT has error " << error << " at index " << j;
+      }
+    }
+    aom_free(in);
+    aom_free(coeff);
+    aom_free(dst);
+    aom_free(src);
+#if CONFIG_HIGHBITDEPTH
+    aom_free(src16);
+    aom_free(dst16);
+#endif
+  }
+
+  int pitch_;
+  int height_;
+  int tx_type_;
+  FhtFunc fwd_txfm_ref;
+  IhtFunc inv_txfm_ref;
+  aom_bit_depth_t bit_depth_;
+  int mask_;
+  int num_coeffs_;
+
+ private:
+  //  Assume transform size is 4x4, 8x8, 16x16,...
+  int FindRowLength() const {
+    int row = 4;
+    if (16 == num_coeffs_) {
+      row = 4;
+    } else if (64 == num_coeffs_) {
+      row = 8;
+    } else if (256 == num_coeffs_) {
+      row = 16;
+    } else if (1024 == num_coeffs_) {
+      row = 32;
+    }
+    return row;
+  }
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_TRANSFORM_TEST_BASE_H_
diff --git a/third_party/aom/test/twopass_encoder.sh b/third_party/aom/test/twopass_encoder.sh
new file mode 100755
index 0000000000..3abb7628b3
--- /dev/null
+++ b/third_party/aom/test/twopass_encoder.sh
@@ -0,0 +1,54 @@
+#!/bin/sh
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+## This file tests the libaom twopass_encoder example. To add new tests to this
+## file, do the following:
+##   1. Write a shell function (this is your test).
+##   2. Add the function to twopass_encoder_tests (on a new line).
+##
+. $(dirname $0)/tools_common.sh
+
+# Environment check: $YUV_RAW_INPUT is required.
+twopass_encoder_verify_environment() {
+  if [ ! -e "${YUV_RAW_INPUT}" ]; then
+    echo "Libaom test data must exist in LIBAOM_TEST_DATA_PATH."
+    return 1
+  fi
+}
+
+# Runs twopass_encoder using the codec specified by $1 with a frame limit of
+# 100.
+twopass_encoder() {
+  local encoder="${LIBAOM_BIN_PATH}/twopass_encoder${AOM_TEST_EXE_SUFFIX}"
+  local codec="$1"
+  local output_file="${AOM_TEST_OUTPUT_DIR}/twopass_encoder_${codec}.ivf"
+  local limit=7
+
+  if [ ! -x "${encoder}" ]; then
+    elog "${encoder} does not exist or is not executable."
+    return 1
+  fi
+
+  eval "${AOM_TEST_PREFIX}" "${encoder}" "${codec}" "${YUV_RAW_INPUT_WIDTH}" \
+      "${YUV_RAW_INPUT_HEIGHT}" "${YUV_RAW_INPUT}" "${output_file}" "${limit}" \
+      ${devnull}
+
+  [ -e "${output_file}" ] || return 1
+}
+
+twopass_encoder_av1() {
+  if [ "$(av1_encode_available)" = "yes" ]; then
+    twopass_encoder av1 || return 1
+  fi
+}
+
+twopass_encoder_tests="twopass_encoder_av1"
+
+run_tests twopass_encoder_verify_environment "${twopass_encoder_tests}"
diff --git a/third_party/aom/test/user_priv_test.cc b/third_party/aom/test/user_priv_test.cc
new file mode 100644
index 0000000000..3052b27b1c
--- /dev/null
+++ b/third_party/aom/test/user_priv_test.cc
@@ -0,0 +1,101 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "./aom_config.h"
+#include "test/acm_random.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "aom_mem/aom_mem.h"
+#include "aom/aom.h"
+
+namespace {
+
+using std::string;
+using libaom_test::ACMRandom;
+
+#if CONFIG_WEBM_IO
+
+void CheckUserPrivateData(void *user_priv, int *target) {
+  // actual pointer value should be the same as expected.
+  EXPECT_EQ(reinterpret_cast<void *>(target), user_priv)
+      << "user_priv pointer value does not match.";
+}
+
+// Decodes |filename|. Passes in user_priv data when calling DecodeFrame and
+// compares the user_priv from return img with the original user_priv to see if
+// they match. Both the pointer values and the values inside the addresses
+// should match.
+string DecodeFile(const string &filename) {
+  ACMRandom rnd(ACMRandom::DeterministicSeed());
+  libaom_test::WebMVideoSource video(filename);
+  video.Init();
+
+  aom_codec_dec_cfg_t cfg = aom_codec_dec_cfg_t();
+  libaom_test::AV1Decoder decoder(cfg, 0);
+
+  libaom_test::MD5 md5;
+  int frame_num = 0;
+  for (video.Begin(); !::testing::Test::HasFailure() && video.cxdata();
+       video.Next()) {
+    void *user_priv = reinterpret_cast<void *>(&frame_num);
+    const aom_codec_err_t res =
+        decoder.DecodeFrame(video.cxdata(), video.frame_size(),
+                            (frame_num == 0) ? NULL : user_priv);
+    if (res != AOM_CODEC_OK) {
+      EXPECT_EQ(AOM_CODEC_OK, res) << decoder.DecodeError();
+      break;
+    }
+    libaom_test::DxDataIterator dec_iter = decoder.GetDxData();
+    const aom_image_t *img = NULL;
+
+    // Get decompressed data.
+    while ((img = dec_iter.Next())) {
+      if (frame_num == 0) {
+        CheckUserPrivateData(img->user_priv, NULL);
+      } else {
+        CheckUserPrivateData(img->user_priv, &frame_num);
+
+        // Also test ctrl_get_reference api.
+        struct av1_ref_frame ref;
+        // Randomly fetch a reference frame.
+        ref.idx = rnd.Rand8() % 3;
+        decoder.Control(AV1_GET_REFERENCE, &ref);
+
+        CheckUserPrivateData(ref.img.user_priv, NULL);
+      }
+      md5.Add(img);
+    }
+
+    frame_num++;
+  }
+  return string(md5.Get());
+}
+
+TEST(UserPrivTest, VideoDecode) {
+  // no tiles or frame parallel; this exercises the decoding to test the
+  // user_priv.
+  EXPECT_STREQ("b35a1b707b28e82be025d960aba039bc",
+               DecodeFile("av10-2-03-size-226x226.webm").c_str());
+}
+
+#endif  // CONFIG_WEBM_IO
+
+}  // namespace
diff --git a/third_party/aom/test/util.h b/third_party/aom/test/util.h
new file mode 100644
index 0000000000..a20fab65cc
--- /dev/null
+++ b/third_party/aom/test/util.h
@@ -0,0 +1,46 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef TEST_UTIL_H_
+#define TEST_UTIL_H_
+
+#include <stdio.h>
+#include <math.h>
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "aom/aom_image.h"
+
+// Macros
+#define GET_PARAM(k) std::tr1::get<k>(GetParam())
+
+inline double compute_psnr(const aom_image_t *img1, const aom_image_t *img2) {
+  assert((img1->fmt == img2->fmt) && (img1->d_w == img2->d_w) &&
+         (img1->d_h == img2->d_h));
+
+  const unsigned int width_y = img1->d_w;
+  const unsigned int height_y = img1->d_h;
+  unsigned int i, j;
+
+  int64_t sqrerr = 0;
+  for (i = 0; i < height_y; ++i)
+    for (j = 0; j < width_y; ++j) {
+      int64_t d = img1->planes[AOM_PLANE_Y][i * img1->stride[AOM_PLANE_Y] + j] -
+                  img2->planes[AOM_PLANE_Y][i * img2->stride[AOM_PLANE_Y] + j];
+      sqrerr += d * d;
+    }
+  double mse = static_cast<double>(sqrerr) / (width_y * height_y);
+  double psnr = 100.0;
+  if (mse > 0.0) {
+    psnr = 10 * log10(255.0 * 255.0 / mse);
+  }
+  return psnr;
+}
+
+#endif  // TEST_UTIL_H_
diff --git a/third_party/aom/test/variance_test.cc b/third_party/aom/test/variance_test.cc
new file mode 100644
index 0000000000..5b1003ca7b
--- /dev/null
+++ b/third_party/aom/test/variance_test.cc
@@ -0,0 +1,1385 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <cstdlib>
+#include <new>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/acm_random.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+#include "aom/aom_codec.h"
+#include "aom/aom_integer.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+
+namespace {
+
+typedef unsigned int (*VarianceMxNFunc)(const uint8_t *a, int a_stride,
+                                        const uint8_t *b, int b_stride,
+                                        unsigned int *sse);
+typedef unsigned int (*SubpixVarMxNFunc)(const uint8_t *a, int a_stride,
+                                         int xoffset, int yoffset,
+                                         const uint8_t *b, int b_stride,
+                                         unsigned int *sse);
+typedef unsigned int (*SubpixAvgVarMxNFunc)(const uint8_t *a, int a_stride,
+                                            int xoffset, int yoffset,
+                                            const uint8_t *b, int b_stride,
+                                            uint32_t *sse,
+                                            const uint8_t *second_pred);
+typedef unsigned int (*Get4x4SseFunc)(const uint8_t *a, int a_stride,
+                                      const uint8_t *b, int b_stride);
+typedef unsigned int (*SumOfSquaresFunction)(const int16_t *src);
+
+using libaom_test::ACMRandom;
+
+// Truncate high bit depth results by downshifting (with rounding) by:
+// 2 * (bit_depth - 8) for sse
+// (bit_depth - 8) for se
+static void RoundHighBitDepth(int bit_depth, int64_t *se, uint64_t *sse) {
+  switch (bit_depth) {
+    case AOM_BITS_12:
+      *sse = (*sse + 128) >> 8;
+      *se = (*se + 8) >> 4;
+      break;
+    case AOM_BITS_10:
+      *sse = (*sse + 8) >> 4;
+      *se = (*se + 2) >> 2;
+      break;
+    case AOM_BITS_8:
+    default: break;
+  }
+}
+
+static unsigned int mb_ss_ref(const int16_t *src) {
+  unsigned int res = 0;
+  for (int i = 0; i < 256; ++i) {
+    res += src[i] * src[i];
+  }
+  return res;
+}
+
+/* Note:
+ *  Our codebase calculates the "diff" value in the variance algorithm by
+ *  (src - ref).
+ */
+static uint32_t variance_ref(const uint8_t *src, const uint8_t *ref, int l2w,
+                             int l2h, int src_stride, int ref_stride,
+                             uint32_t *sse_ptr, bool use_high_bit_depth_,
+                             aom_bit_depth_t bit_depth) {
+  int64_t se = 0;
+  uint64_t sse = 0;
+  const int w = 1 << l2w;
+  const int h = 1 << l2h;
+  for (int y = 0; y < h; y++) {
+    for (int x = 0; x < w; x++) {
+      int diff;
+      if (!use_high_bit_depth_) {
+        diff = src[y * src_stride + x] - ref[y * ref_stride + x];
+        se += diff;
+        sse += diff * diff;
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        diff = CONVERT_TO_SHORTPTR(src)[y * src_stride + x] -
+               CONVERT_TO_SHORTPTR(ref)[y * ref_stride + x];
+        se += diff;
+        sse += diff * diff;
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+    }
+  }
+  RoundHighBitDepth(bit_depth, &se, &sse);
+  *sse_ptr = static_cast<uint32_t>(sse);
+  return static_cast<uint32_t>(sse - ((se * se) >> (l2w + l2h)));
+}
+
+/* The subpel reference functions differ from the codec version in one aspect:
+ * they calculate the bilinear factors directly instead of using a lookup table
+ * and therefore upshift xoff and yoff by 1. Only every other calculated value
+ * is used so the codec version shrinks the table to save space and maintain
+ * compatibility with vp8.
+ */
+static uint32_t subpel_variance_ref(const uint8_t *ref, const uint8_t *src,
+                                    int l2w, int l2h, int xoff, int yoff,
+                                    uint32_t *sse_ptr, bool use_high_bit_depth_,
+                                    aom_bit_depth_t bit_depth) {
+  int64_t se = 0;
+  uint64_t sse = 0;
+  const int w = 1 << l2w;
+  const int h = 1 << l2h;
+
+  xoff <<= 1;
+  yoff <<= 1;
+
+  for (int y = 0; y < h; y++) {
+    for (int x = 0; x < w; x++) {
+      // Bilinear interpolation at a 16th pel step.
+      if (!use_high_bit_depth_) {
+        const int a1 = ref[(w + 1) * (y + 0) + x + 0];
+        const int a2 = ref[(w + 1) * (y + 0) + x + 1];
+        const int b1 = ref[(w + 1) * (y + 1) + x + 0];
+        const int b2 = ref[(w + 1) * (y + 1) + x + 1];
+        const int a = a1 + (((a2 - a1) * xoff + 8) >> 4);
+        const int b = b1 + (((b2 - b1) * xoff + 8) >> 4);
+        const int r = a + (((b - a) * yoff + 8) >> 4);
+        const int diff = r - src[w * y + x];
+        se += diff;
+        sse += diff * diff;
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        uint16_t *ref16 = CONVERT_TO_SHORTPTR(ref);
+        uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+        const int a1 = ref16[(w + 1) * (y + 0) + x + 0];
+        const int a2 = ref16[(w + 1) * (y + 0) + x + 1];
+        const int b1 = ref16[(w + 1) * (y + 1) + x + 0];
+        const int b2 = ref16[(w + 1) * (y + 1) + x + 1];
+        const int a = a1 + (((a2 - a1) * xoff + 8) >> 4);
+        const int b = b1 + (((b2 - b1) * xoff + 8) >> 4);
+        const int r = a + (((b - a) * yoff + 8) >> 4);
+        const int diff = r - src16[w * y + x];
+        se += diff;
+        sse += diff * diff;
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+    }
+  }
+  RoundHighBitDepth(bit_depth, &se, &sse);
+  *sse_ptr = static_cast<uint32_t>(sse);
+  return static_cast<uint32_t>(sse - ((se * se) >> (l2w + l2h)));
+}
+
+static uint32_t subpel_avg_variance_ref(const uint8_t *ref, const uint8_t *src,
+                                        const uint8_t *second_pred, int l2w,
+                                        int l2h, int xoff, int yoff,
+                                        uint32_t *sse_ptr,
+                                        bool use_high_bit_depth,
+                                        aom_bit_depth_t bit_depth) {
+  int64_t se = 0;
+  uint64_t sse = 0;
+  const int w = 1 << l2w;
+  const int h = 1 << l2h;
+
+  xoff <<= 1;
+  yoff <<= 1;
+
+  for (int y = 0; y < h; y++) {
+    for (int x = 0; x < w; x++) {
+      // bilinear interpolation at a 16th pel step
+      if (!use_high_bit_depth) {
+        const int a1 = ref[(w + 1) * (y + 0) + x + 0];
+        const int a2 = ref[(w + 1) * (y + 0) + x + 1];
+        const int b1 = ref[(w + 1) * (y + 1) + x + 0];
+        const int b2 = ref[(w + 1) * (y + 1) + x + 1];
+        const int a = a1 + (((a2 - a1) * xoff + 8) >> 4);
+        const int b = b1 + (((b2 - b1) * xoff + 8) >> 4);
+        const int r = a + (((b - a) * yoff + 8) >> 4);
+        const int diff =
+            ((r + second_pred[w * y + x] + 1) >> 1) - src[w * y + x];
+        se += diff;
+        sse += diff * diff;
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        const uint16_t *ref16 = CONVERT_TO_SHORTPTR(ref);
+        const uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
+        const uint16_t *sec16 = CONVERT_TO_SHORTPTR(second_pred);
+        const int a1 = ref16[(w + 1) * (y + 0) + x + 0];
+        const int a2 = ref16[(w + 1) * (y + 0) + x + 1];
+        const int b1 = ref16[(w + 1) * (y + 1) + x + 0];
+        const int b2 = ref16[(w + 1) * (y + 1) + x + 1];
+        const int a = a1 + (((a2 - a1) * xoff + 8) >> 4);
+        const int b = b1 + (((b2 - b1) * xoff + 8) >> 4);
+        const int r = a + (((b - a) * yoff + 8) >> 4);
+        const int diff = ((r + sec16[w * y + x] + 1) >> 1) - src16[w * y + x];
+        se += diff;
+        sse += diff * diff;
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+    }
+  }
+  RoundHighBitDepth(bit_depth, &se, &sse);
+  *sse_ptr = static_cast<uint32_t>(sse);
+  return static_cast<uint32_t>(sse - ((se * se) >> (l2w + l2h)));
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+class SumOfSquaresTest : public ::testing::TestWithParam<SumOfSquaresFunction> {
+ public:
+  SumOfSquaresTest() : func_(GetParam()) {}
+
+  virtual ~SumOfSquaresTest() { libaom_test::ClearSystemState(); }
+
+ protected:
+  void ConstTest();
+  void RefTest();
+
+  SumOfSquaresFunction func_;
+  ACMRandom rnd_;
+};
+
+void SumOfSquaresTest::ConstTest() {
+  int16_t mem[256];
+  unsigned int res;
+  for (int v = 0; v < 256; ++v) {
+    for (int i = 0; i < 256; ++i) {
+      mem[i] = v;
+    }
+    ASM_REGISTER_STATE_CHECK(res = func_(mem));
+    EXPECT_EQ(256u * (v * v), res);
+  }
+}
+
+void SumOfSquaresTest::RefTest() {
+  int16_t mem[256];
+  for (int i = 0; i < 100; ++i) {
+    for (int j = 0; j < 256; ++j) {
+      mem[j] = rnd_.Rand8() - rnd_.Rand8();
+    }
+
+    const unsigned int expected = mb_ss_ref(mem);
+    unsigned int res;
+    ASM_REGISTER_STATE_CHECK(res = func_(mem));
+    EXPECT_EQ(expected, res);
+  }
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// Encapsulating struct to store the function to test along with
+// some testing context.
+// Can be used for MSE, SSE, Variance, etc.
+
+template <typename Func>
+struct TestParams {
+  TestParams(int log2w = 0, int log2h = 0, Func function = NULL,
+             int bit_depth_value = 0)
+      : log2width(log2w), log2height(log2h), func(function) {
+    use_high_bit_depth = (bit_depth_value > 0);
+    if (use_high_bit_depth) {
+      bit_depth = static_cast<aom_bit_depth_t>(bit_depth_value);
+    } else {
+      bit_depth = AOM_BITS_8;
+    }
+    width = 1 << log2width;
+    height = 1 << log2height;
+    block_size = width * height;
+    mask = (1u << bit_depth) - 1;
+  }
+
+  int log2width, log2height;
+  int width, height;
+  int block_size;
+  Func func;
+  aom_bit_depth_t bit_depth;
+  bool use_high_bit_depth;
+  uint32_t mask;
+};
+
+template <typename Func>
+std::ostream &operator<<(std::ostream &os, const TestParams<Func> &p) {
+  return os << "log2width/height:" << p.log2width << "/" << p.log2height
+            << " function:" << reinterpret_cast<const void *>(p.func)
+            << " bit-depth:" << p.bit_depth;
+}
+
+// Main class for testing a function type
+template <typename FunctionType>
+class MainTestClass
+    : public ::testing::TestWithParam<TestParams<FunctionType> > {
+ public:
+  virtual void SetUp() {
+    params_ = this->GetParam();
+
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+    const size_t unit =
+        use_high_bit_depth() ? sizeof(uint16_t) : sizeof(uint8_t);
+    src_ = reinterpret_cast<uint8_t *>(aom_memalign(16, block_size() * unit));
+    ref_ = new uint8_t[block_size() * unit];
+    ASSERT_TRUE(src_ != NULL);
+    ASSERT_TRUE(ref_ != NULL);
+#if CONFIG_HIGHBITDEPTH
+    if (use_high_bit_depth()) {
+      // TODO(skal): remove!
+      src_ = CONVERT_TO_BYTEPTR(src_);
+      ref_ = CONVERT_TO_BYTEPTR(ref_);
+    }
+#endif
+  }
+
+  virtual void TearDown() {
+#if CONFIG_HIGHBITDEPTH
+    if (use_high_bit_depth()) {
+      // TODO(skal): remove!
+      src_ = reinterpret_cast<uint8_t *>(CONVERT_TO_SHORTPTR(src_));
+      ref_ = reinterpret_cast<uint8_t *>(CONVERT_TO_SHORTPTR(ref_));
+    }
+#endif
+
+    aom_free(src_);
+    delete[] ref_;
+    src_ = NULL;
+    ref_ = NULL;
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  // We could sub-class MainTestClass into dedicated class for Variance
+  // and MSE/SSE, but it involves a lot of 'this->xxx' dereferencing
+  // to access top class fields xxx. That's cumbersome, so for now we'll just
+  // implement the testing methods here:
+
+  // Variance tests
+  void ZeroTest();
+  void RefTest();
+  void RefStrideTest();
+  void OneQuarterTest();
+
+  // MSE/SSE tests
+  void RefTestMse();
+  void RefTestSse();
+  void MaxTestMse();
+  void MaxTestSse();
+
+ protected:
+  ACMRandom rnd_;
+  uint8_t *src_;
+  uint8_t *ref_;
+  TestParams<FunctionType> params_;
+
+  // some relay helpers
+  bool use_high_bit_depth() const { return params_.use_high_bit_depth; }
+  int byte_shift() const { return params_.bit_depth - 8; }
+  int block_size() const { return params_.block_size; }
+  int width() const { return params_.width; }
+  uint32_t mask() const { return params_.mask; }
+};
+
+////////////////////////////////////////////////////////////////////////////////
+// Tests related to variance.
+
+template <typename VarianceFunctionType>
+void MainTestClass<VarianceFunctionType>::ZeroTest() {
+  for (int i = 0; i <= 255; ++i) {
+    if (!use_high_bit_depth()) {
+      memset(src_, i, block_size());
+    } else {
+      uint16_t *const src16 = CONVERT_TO_SHORTPTR(src_);
+      for (int k = 0; k < block_size(); ++k) src16[k] = i << byte_shift();
+    }
+    for (int j = 0; j <= 255; ++j) {
+      if (!use_high_bit_depth()) {
+        memset(ref_, j, block_size());
+      } else {
+        uint16_t *const ref16 = CONVERT_TO_SHORTPTR(ref_);
+        for (int k = 0; k < block_size(); ++k) ref16[k] = j << byte_shift();
+      }
+      unsigned int sse, var;
+      ASM_REGISTER_STATE_CHECK(
+          var = params_.func(src_, width(), ref_, width(), &sse));
+      EXPECT_EQ(0u, var) << "src values: " << i << " ref values: " << j;
+    }
+  }
+}
+
+template <typename VarianceFunctionType>
+void MainTestClass<VarianceFunctionType>::RefTest() {
+  for (int i = 0; i < 10; ++i) {
+    for (int j = 0; j < block_size(); j++) {
+      if (!use_high_bit_depth()) {
+        src_[j] = rnd_.Rand8();
+        ref_[j] = rnd_.Rand8();
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        CONVERT_TO_SHORTPTR(src_)[j] = rnd_.Rand16() & mask();
+        CONVERT_TO_SHORTPTR(ref_)[j] = rnd_.Rand16() & mask();
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+    }
+    unsigned int sse1, sse2, var1, var2;
+    const int stride = width();
+    ASM_REGISTER_STATE_CHECK(
+        var1 = params_.func(src_, stride, ref_, stride, &sse1));
+    var2 =
+        variance_ref(src_, ref_, params_.log2width, params_.log2height, stride,
+                     stride, &sse2, use_high_bit_depth(), params_.bit_depth);
+    EXPECT_EQ(sse1, sse2) << "Error at test index: " << i;
+    EXPECT_EQ(var1, var2) << "Error at test index: " << i;
+  }
+}
+
+template <typename VarianceFunctionType>
+void MainTestClass<VarianceFunctionType>::RefStrideTest() {
+  for (int i = 0; i < 10; ++i) {
+    const int ref_stride = (i & 1) * width();
+    const int src_stride = ((i >> 1) & 1) * width();
+    for (int j = 0; j < block_size(); j++) {
+      const int ref_ind = (j / width()) * ref_stride + j % width();
+      const int src_ind = (j / width()) * src_stride + j % width();
+      if (!use_high_bit_depth()) {
+        src_[src_ind] = rnd_.Rand8();
+        ref_[ref_ind] = rnd_.Rand8();
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        CONVERT_TO_SHORTPTR(src_)[src_ind] = rnd_.Rand16() & mask();
+        CONVERT_TO_SHORTPTR(ref_)[ref_ind] = rnd_.Rand16() & mask();
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+    }
+    unsigned int sse1, sse2;
+    unsigned int var1, var2;
+
+    ASM_REGISTER_STATE_CHECK(
+        var1 = params_.func(src_, src_stride, ref_, ref_stride, &sse1));
+    var2 = variance_ref(src_, ref_, params_.log2width, params_.log2height,
+                        src_stride, ref_stride, &sse2, use_high_bit_depth(),
+                        params_.bit_depth);
+    EXPECT_EQ(sse1, sse2) << "Error at test index: " << i;
+    EXPECT_EQ(var1, var2) << "Error at test index: " << i;
+  }
+}
+
+template <typename VarianceFunctionType>
+void MainTestClass<VarianceFunctionType>::OneQuarterTest() {
+  const int half = block_size() / 2;
+  if (!use_high_bit_depth()) {
+    memset(src_, 255, block_size());
+    memset(ref_, 255, half);
+    memset(ref_ + half, 0, half);
+#if CONFIG_HIGHBITDEPTH
+  } else {
+    aom_memset16(CONVERT_TO_SHORTPTR(src_), 255 << byte_shift(), block_size());
+    aom_memset16(CONVERT_TO_SHORTPTR(ref_), 255 << byte_shift(), half);
+    aom_memset16(CONVERT_TO_SHORTPTR(ref_) + half, 0, half);
+#endif  // CONFIG_HIGHBITDEPTH
+  }
+  unsigned int sse, var, expected;
+  ASM_REGISTER_STATE_CHECK(
+      var = params_.func(src_, width(), ref_, width(), &sse));
+  expected = block_size() * 255 * 255 / 4;
+  EXPECT_EQ(expected, var);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+// Tests related to MSE / SSE.
+
+template <typename FunctionType>
+void MainTestClass<FunctionType>::RefTestMse() {
+  for (int i = 0; i < 10; ++i) {
+    for (int j = 0; j < block_size(); ++j) {
+      src_[j] = rnd_.Rand8();
+      ref_[j] = rnd_.Rand8();
+    }
+    unsigned int sse1, sse2;
+    const int stride = width();
+    ASM_REGISTER_STATE_CHECK(params_.func(src_, stride, ref_, stride, &sse1));
+    variance_ref(src_, ref_, params_.log2width, params_.log2height, stride,
+                 stride, &sse2, false, AOM_BITS_8);
+    EXPECT_EQ(sse1, sse2);
+  }
+}
+
+template <typename FunctionType>
+void MainTestClass<FunctionType>::RefTestSse() {
+  for (int i = 0; i < 10; ++i) {
+    for (int j = 0; j < block_size(); ++j) {
+      src_[j] = rnd_.Rand8();
+      ref_[j] = rnd_.Rand8();
+    }
+    unsigned int sse2;
+    unsigned int var1;
+    const int stride = width();
+    ASM_REGISTER_STATE_CHECK(var1 = params_.func(src_, stride, ref_, stride));
+    variance_ref(src_, ref_, params_.log2width, params_.log2height, stride,
+                 stride, &sse2, false, AOM_BITS_8);
+    EXPECT_EQ(var1, sse2);
+  }
+}
+
+template <typename FunctionType>
+void MainTestClass<FunctionType>::MaxTestMse() {
+  memset(src_, 255, block_size());
+  memset(ref_, 0, block_size());
+  unsigned int sse;
+  ASM_REGISTER_STATE_CHECK(params_.func(src_, width(), ref_, width(), &sse));
+  const unsigned int expected = block_size() * 255 * 255;
+  EXPECT_EQ(expected, sse);
+}
+
+template <typename FunctionType>
+void MainTestClass<FunctionType>::MaxTestSse() {
+  memset(src_, 255, block_size());
+  memset(ref_, 0, block_size());
+  unsigned int var;
+  ASM_REGISTER_STATE_CHECK(var = params_.func(src_, width(), ref_, width()));
+  const unsigned int expected = block_size() * 255 * 255;
+  EXPECT_EQ(expected, var);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+using ::std::tr1::get;
+using ::std::tr1::make_tuple;
+using ::std::tr1::tuple;
+
+template <typename SubpelVarianceFunctionType>
+class SubpelVarianceTest
+    : public ::testing::TestWithParam<
+          tuple<int, int, SubpelVarianceFunctionType, int> > {
+ public:
+  virtual void SetUp() {
+    const tuple<int, int, SubpelVarianceFunctionType, int> &params =
+        this->GetParam();
+    log2width_ = get<0>(params);
+    width_ = 1 << log2width_;
+    log2height_ = get<1>(params);
+    height_ = 1 << log2height_;
+    subpel_variance_ = get<2>(params);
+    if (get<3>(params)) {
+      bit_depth_ = (aom_bit_depth_t)get<3>(params);
+      use_high_bit_depth_ = true;
+    } else {
+      bit_depth_ = AOM_BITS_8;
+      use_high_bit_depth_ = false;
+    }
+    mask_ = (1 << bit_depth_) - 1;
+
+    rnd_.Reset(ACMRandom::DeterministicSeed());
+    block_size_ = width_ * height_;
+    if (!use_high_bit_depth_) {
+      src_ = reinterpret_cast<uint8_t *>(aom_memalign(16, block_size_));
+      sec_ = reinterpret_cast<uint8_t *>(aom_memalign(16, block_size_));
+      ref_ = new uint8_t[block_size_ + width_ + height_ + 1];
+#if CONFIG_HIGHBITDEPTH
+    } else {
+      src_ = CONVERT_TO_BYTEPTR(reinterpret_cast<uint16_t *>(
+          aom_memalign(16, block_size_ * sizeof(uint16_t))));
+      sec_ = CONVERT_TO_BYTEPTR(reinterpret_cast<uint16_t *>(
+          aom_memalign(16, block_size_ * sizeof(uint16_t))));
+      ref_ =
+          CONVERT_TO_BYTEPTR(new uint16_t[block_size_ + width_ + height_ + 1]);
+#endif  // CONFIG_HIGHBITDEPTH
+    }
+    ASSERT_TRUE(src_ != NULL);
+    ASSERT_TRUE(sec_ != NULL);
+    ASSERT_TRUE(ref_ != NULL);
+  }
+
+  virtual void TearDown() {
+    if (!use_high_bit_depth_) {
+      aom_free(src_);
+      delete[] ref_;
+      aom_free(sec_);
+#if CONFIG_HIGHBITDEPTH
+    } else {
+      aom_free(CONVERT_TO_SHORTPTR(src_));
+      delete[] CONVERT_TO_SHORTPTR(ref_);
+      aom_free(CONVERT_TO_SHORTPTR(sec_));
+#endif  // CONFIG_HIGHBITDEPTH
+    }
+    libaom_test::ClearSystemState();
+  }
+
+ protected:
+  void RefTest();
+  void ExtremeRefTest();
+
+  ACMRandom rnd_;
+  uint8_t *src_;
+  uint8_t *ref_;
+  uint8_t *sec_;
+  bool use_high_bit_depth_;
+  aom_bit_depth_t bit_depth_;
+  int width_, log2width_;
+  int height_, log2height_;
+  int block_size_, mask_;
+  SubpelVarianceFunctionType subpel_variance_;
+};
+
+template <typename SubpelVarianceFunctionType>
+void SubpelVarianceTest<SubpelVarianceFunctionType>::RefTest() {
+  for (int x = 0; x < 8; ++x) {
+    for (int y = 0; y < 8; ++y) {
+      if (!use_high_bit_depth_) {
+        for (int j = 0; j < block_size_; j++) {
+          src_[j] = rnd_.Rand8();
+        }
+        for (int j = 0; j < block_size_ + width_ + height_ + 1; j++) {
+          ref_[j] = rnd_.Rand8();
+        }
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        for (int j = 0; j < block_size_; j++) {
+          CONVERT_TO_SHORTPTR(src_)[j] = rnd_.Rand16() & mask_;
+        }
+        for (int j = 0; j < block_size_ + width_ + height_ + 1; j++) {
+          CONVERT_TO_SHORTPTR(ref_)[j] = rnd_.Rand16() & mask_;
+        }
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+      unsigned int sse1, sse2;
+      unsigned int var1;
+      ASM_REGISTER_STATE_CHECK(
+          var1 = subpel_variance_(ref_, width_ + 1, x, y, src_, width_, &sse1));
+      const unsigned int var2 =
+          subpel_variance_ref(ref_, src_, log2width_, log2height_, x, y, &sse2,
+                              use_high_bit_depth_, bit_depth_);
+      EXPECT_EQ(sse1, sse2) << "at position " << x << ", " << y;
+      EXPECT_EQ(var1, var2) << "at position " << x << ", " << y;
+    }
+  }
+}
+
+template <typename SubpelVarianceFunctionType>
+void SubpelVarianceTest<SubpelVarianceFunctionType>::ExtremeRefTest() {
+  // Compare against reference.
+  // Src: Set the first half of values to 0, the second half to the maximum.
+  // Ref: Set the first half of values to the maximum, the second half to 0.
+  for (int x = 0; x < 8; ++x) {
+    for (int y = 0; y < 8; ++y) {
+      const int half = block_size_ / 2;
+      if (!use_high_bit_depth_) {
+        memset(src_, 0, half);
+        memset(src_ + half, 255, half);
+        memset(ref_, 255, half);
+        memset(ref_ + half, 0, half + width_ + height_ + 1);
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        aom_memset16(CONVERT_TO_SHORTPTR(src_), mask_, half);
+        aom_memset16(CONVERT_TO_SHORTPTR(src_) + half, 0, half);
+        aom_memset16(CONVERT_TO_SHORTPTR(ref_), 0, half);
+        aom_memset16(CONVERT_TO_SHORTPTR(ref_) + half, mask_,
+                     half + width_ + height_ + 1);
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+      unsigned int sse1, sse2;
+      unsigned int var1;
+      ASM_REGISTER_STATE_CHECK(
+          var1 = subpel_variance_(ref_, width_ + 1, x, y, src_, width_, &sse1));
+      const unsigned int var2 =
+          subpel_variance_ref(ref_, src_, log2width_, log2height_, x, y, &sse2,
+                              use_high_bit_depth_, bit_depth_);
+      EXPECT_EQ(sse1, sse2) << "for xoffset " << x << " and yoffset " << y;
+      EXPECT_EQ(var1, var2) << "for xoffset " << x << " and yoffset " << y;
+    }
+  }
+}
+
+template <>
+void SubpelVarianceTest<SubpixAvgVarMxNFunc>::RefTest() {
+  for (int x = 0; x < 8; ++x) {
+    for (int y = 0; y < 8; ++y) {
+      if (!use_high_bit_depth_) {
+        for (int j = 0; j < block_size_; j++) {
+          src_[j] = rnd_.Rand8();
+          sec_[j] = rnd_.Rand8();
+        }
+        for (int j = 0; j < block_size_ + width_ + height_ + 1; j++) {
+          ref_[j] = rnd_.Rand8();
+        }
+#if CONFIG_HIGHBITDEPTH
+      } else {
+        for (int j = 0; j < block_size_; j++) {
+          CONVERT_TO_SHORTPTR(src_)[j] = rnd_.Rand16() & mask_;
+          CONVERT_TO_SHORTPTR(sec_)[j] = rnd_.Rand16() & mask_;
+        }
+        for (int j = 0; j < block_size_ + width_ + height_ + 1; j++) {
+          CONVERT_TO_SHORTPTR(ref_)[j] = rnd_.Rand16() & mask_;
+        }
+#endif  // CONFIG_HIGHBITDEPTH
+      }
+      uint32_t sse1, sse2;
+      uint32_t var1, var2;
+      ASM_REGISTER_STATE_CHECK(var1 =
+                                   subpel_variance_(ref_, width_ + 1, x, y,
+                                                    src_, width_, &sse1, sec_));
+      var2 = subpel_avg_variance_ref(ref_, src_, sec_, log2width_, log2height_,
+                                     x, y, &sse2, use_high_bit_depth_,
+                                     static_cast<aom_bit_depth_t>(bit_depth_));
+      EXPECT_EQ(sse1, sse2) << "at position " << x << ", " << y;
+      EXPECT_EQ(var1, var2) << "at position " << x << ", " << y;
+    }
+  }
+}
+
+typedef MainTestClass<Get4x4SseFunc> AvxSseTest;
+typedef MainTestClass<VarianceMxNFunc> AvxMseTest;
+typedef MainTestClass<VarianceMxNFunc> AvxVarianceTest;
+typedef SubpelVarianceTest<SubpixVarMxNFunc> AvxSubpelVarianceTest;
+typedef SubpelVarianceTest<SubpixAvgVarMxNFunc> AvxSubpelAvgVarianceTest;
+
+TEST_P(AvxSseTest, RefSse) { RefTestSse(); }
+TEST_P(AvxSseTest, MaxSse) { MaxTestSse(); }
+TEST_P(AvxMseTest, RefMse) { RefTestMse(); }
+TEST_P(AvxMseTest, MaxMse) { MaxTestMse(); }
+TEST_P(AvxVarianceTest, Zero) { ZeroTest(); }
+TEST_P(AvxVarianceTest, Ref) { RefTest(); }
+TEST_P(AvxVarianceTest, RefStride) { RefStrideTest(); }
+TEST_P(AvxVarianceTest, OneQuarter) { OneQuarterTest(); }
+TEST_P(SumOfSquaresTest, Const) { ConstTest(); }
+TEST_P(SumOfSquaresTest, Ref) { RefTest(); }
+TEST_P(AvxSubpelVarianceTest, Ref) { RefTest(); }
+TEST_P(AvxSubpelVarianceTest, ExtremeRef) { ExtremeRefTest(); }
+TEST_P(AvxSubpelAvgVarianceTest, Ref) { RefTest(); }
+
+INSTANTIATE_TEST_CASE_P(C, SumOfSquaresTest,
+                        ::testing::Values(aom_get_mb_ss_c));
+
+typedef TestParams<Get4x4SseFunc> SseParams;
+INSTANTIATE_TEST_CASE_P(C, AvxSseTest,
+                        ::testing::Values(SseParams(2, 2,
+                                                    &aom_get4x4sse_cs_c)));
+
+typedef TestParams<VarianceMxNFunc> MseParams;
+INSTANTIATE_TEST_CASE_P(C, AvxMseTest,
+                        ::testing::Values(MseParams(4, 4, &aom_mse16x16_c),
+                                          MseParams(4, 3, &aom_mse16x8_c),
+                                          MseParams(3, 4, &aom_mse8x16_c),
+                                          MseParams(3, 3, &aom_mse8x8_c)));
+
+typedef TestParams<VarianceMxNFunc> VarianceParams;
+INSTANTIATE_TEST_CASE_P(
+    C, AvxVarianceTest,
+    ::testing::Values(VarianceParams(6, 6, &aom_variance64x64_c),
+                      VarianceParams(6, 5, &aom_variance64x32_c),
+                      VarianceParams(5, 6, &aom_variance32x64_c),
+                      VarianceParams(5, 5, &aom_variance32x32_c),
+                      VarianceParams(5, 4, &aom_variance32x16_c),
+                      VarianceParams(4, 5, &aom_variance16x32_c),
+                      VarianceParams(4, 4, &aom_variance16x16_c),
+                      VarianceParams(4, 3, &aom_variance16x8_c),
+                      VarianceParams(3, 4, &aom_variance8x16_c),
+                      VarianceParams(3, 3, &aom_variance8x8_c),
+                      VarianceParams(3, 2, &aom_variance8x4_c),
+                      VarianceParams(2, 3, &aom_variance4x8_c),
+                      VarianceParams(2, 2, &aom_variance4x4_c)));
+
+INSTANTIATE_TEST_CASE_P(
+    C, AvxSubpelVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &aom_sub_pixel_variance64x64_c, 0),
+                      make_tuple(6, 5, &aom_sub_pixel_variance64x32_c, 0),
+                      make_tuple(5, 6, &aom_sub_pixel_variance32x64_c, 0),
+                      make_tuple(5, 5, &aom_sub_pixel_variance32x32_c, 0),
+                      make_tuple(5, 4, &aom_sub_pixel_variance32x16_c, 0),
+                      make_tuple(4, 5, &aom_sub_pixel_variance16x32_c, 0),
+                      make_tuple(4, 4, &aom_sub_pixel_variance16x16_c, 0),
+                      make_tuple(4, 3, &aom_sub_pixel_variance16x8_c, 0),
+                      make_tuple(3, 4, &aom_sub_pixel_variance8x16_c, 0),
+                      make_tuple(3, 3, &aom_sub_pixel_variance8x8_c, 0),
+                      make_tuple(3, 2, &aom_sub_pixel_variance8x4_c, 0),
+                      make_tuple(2, 3, &aom_sub_pixel_variance4x8_c, 0),
+                      make_tuple(2, 2, &aom_sub_pixel_variance4x4_c, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    C, AvxSubpelAvgVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &aom_sub_pixel_avg_variance64x64_c, 0),
+                      make_tuple(6, 5, &aom_sub_pixel_avg_variance64x32_c, 0),
+                      make_tuple(5, 6, &aom_sub_pixel_avg_variance32x64_c, 0),
+                      make_tuple(5, 5, &aom_sub_pixel_avg_variance32x32_c, 0),
+                      make_tuple(5, 4, &aom_sub_pixel_avg_variance32x16_c, 0),
+                      make_tuple(4, 5, &aom_sub_pixel_avg_variance16x32_c, 0),
+                      make_tuple(4, 4, &aom_sub_pixel_avg_variance16x16_c, 0),
+                      make_tuple(4, 3, &aom_sub_pixel_avg_variance16x8_c, 0),
+                      make_tuple(3, 4, &aom_sub_pixel_avg_variance8x16_c, 0),
+                      make_tuple(3, 3, &aom_sub_pixel_avg_variance8x8_c, 0),
+                      make_tuple(3, 2, &aom_sub_pixel_avg_variance8x4_c, 0),
+                      make_tuple(2, 3, &aom_sub_pixel_avg_variance4x8_c, 0),
+                      make_tuple(2, 2, &aom_sub_pixel_avg_variance4x4_c, 0)));
+
+#if CONFIG_HIGHBITDEPTH
+typedef MainTestClass<VarianceMxNFunc> AvxHBDMseTest;
+typedef MainTestClass<VarianceMxNFunc> AvxHBDVarianceTest;
+typedef SubpelVarianceTest<SubpixVarMxNFunc> AvxHBDSubpelVarianceTest;
+typedef SubpelVarianceTest<SubpixAvgVarMxNFunc> AvxHBDSubpelAvgVarianceTest;
+
+TEST_P(AvxHBDMseTest, RefMse) { RefTestMse(); }
+TEST_P(AvxHBDMseTest, MaxMse) { MaxTestMse(); }
+TEST_P(AvxHBDVarianceTest, Zero) { ZeroTest(); }
+TEST_P(AvxHBDVarianceTest, Ref) { RefTest(); }
+TEST_P(AvxHBDVarianceTest, RefStride) { RefStrideTest(); }
+TEST_P(AvxHBDVarianceTest, OneQuarter) { OneQuarterTest(); }
+TEST_P(AvxHBDSubpelVarianceTest, Ref) { RefTest(); }
+TEST_P(AvxHBDSubpelVarianceTest, ExtremeRef) { ExtremeRefTest(); }
+TEST_P(AvxHBDSubpelAvgVarianceTest, Ref) { RefTest(); }
+
+/* TODO(debargha): This test does not support the highbd version
+INSTANTIATE_TEST_CASE_P(
+    C, AvxHBDMseTest,
+    ::testing::Values(make_tuple(4, 4, &aom_highbd_12_mse16x16_c),
+                      make_tuple(4, 4, &aom_highbd_12_mse16x8_c),
+                      make_tuple(4, 4, &aom_highbd_12_mse8x16_c),
+                      make_tuple(4, 4, &aom_highbd_12_mse8x8_c),
+                      make_tuple(4, 4, &aom_highbd_10_mse16x16_c),
+                      make_tuple(4, 4, &aom_highbd_10_mse16x8_c),
+                      make_tuple(4, 4, &aom_highbd_10_mse8x16_c),
+                      make_tuple(4, 4, &aom_highbd_10_mse8x8_c),
+                      make_tuple(4, 4, &aom_highbd_8_mse16x16_c),
+                      make_tuple(4, 4, &aom_highbd_8_mse16x8_c),
+                      make_tuple(4, 4, &aom_highbd_8_mse8x16_c),
+                      make_tuple(4, 4, &aom_highbd_8_mse8x8_c)));
+*/
+
+const VarianceParams kArrayHBDVariance_c[] = {
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  VarianceParams(7, 7, &aom_highbd_12_variance128x128_c, 12),
+  VarianceParams(7, 6, &aom_highbd_12_variance128x64_c, 12),
+  VarianceParams(6, 7, &aom_highbd_12_variance64x128_c, 12),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  VarianceParams(6, 6, &aom_highbd_12_variance64x64_c, 12),
+  VarianceParams(6, 5, &aom_highbd_12_variance64x32_c, 12),
+  VarianceParams(5, 6, &aom_highbd_12_variance32x64_c, 12),
+  VarianceParams(5, 5, &aom_highbd_12_variance32x32_c, 12),
+  VarianceParams(5, 4, &aom_highbd_12_variance32x16_c, 12),
+  VarianceParams(4, 5, &aom_highbd_12_variance16x32_c, 12),
+  VarianceParams(4, 4, &aom_highbd_12_variance16x16_c, 12),
+  VarianceParams(4, 3, &aom_highbd_12_variance16x8_c, 12),
+  VarianceParams(3, 4, &aom_highbd_12_variance8x16_c, 12),
+  VarianceParams(3, 3, &aom_highbd_12_variance8x8_c, 12),
+  VarianceParams(3, 2, &aom_highbd_12_variance8x4_c, 12),
+  VarianceParams(2, 3, &aom_highbd_12_variance4x8_c, 12),
+  VarianceParams(2, 2, &aom_highbd_12_variance4x4_c, 12),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  VarianceParams(7, 7, &aom_highbd_10_variance128x128_c, 10),
+  VarianceParams(7, 6, &aom_highbd_10_variance128x64_c, 10),
+  VarianceParams(6, 7, &aom_highbd_10_variance64x128_c, 10),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  VarianceParams(6, 6, &aom_highbd_10_variance64x64_c, 10),
+  VarianceParams(6, 5, &aom_highbd_10_variance64x32_c, 10),
+  VarianceParams(5, 6, &aom_highbd_10_variance32x64_c, 10),
+  VarianceParams(5, 5, &aom_highbd_10_variance32x32_c, 10),
+  VarianceParams(5, 4, &aom_highbd_10_variance32x16_c, 10),
+  VarianceParams(4, 5, &aom_highbd_10_variance16x32_c, 10),
+  VarianceParams(4, 4, &aom_highbd_10_variance16x16_c, 10),
+  VarianceParams(4, 3, &aom_highbd_10_variance16x8_c, 10),
+  VarianceParams(3, 4, &aom_highbd_10_variance8x16_c, 10),
+  VarianceParams(3, 3, &aom_highbd_10_variance8x8_c, 10),
+  VarianceParams(3, 2, &aom_highbd_10_variance8x4_c, 10),
+  VarianceParams(2, 3, &aom_highbd_10_variance4x8_c, 10),
+  VarianceParams(2, 2, &aom_highbd_10_variance4x4_c, 10),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  VarianceParams(7, 7, &aom_highbd_8_variance128x128_c, 8),
+  VarianceParams(7, 6, &aom_highbd_8_variance128x64_c, 8),
+  VarianceParams(6, 7, &aom_highbd_8_variance64x128_c, 8),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  VarianceParams(6, 6, &aom_highbd_8_variance64x64_c, 8),
+  VarianceParams(6, 5, &aom_highbd_8_variance64x32_c, 8),
+  VarianceParams(5, 6, &aom_highbd_8_variance32x64_c, 8),
+  VarianceParams(5, 5, &aom_highbd_8_variance32x32_c, 8),
+  VarianceParams(5, 4, &aom_highbd_8_variance32x16_c, 8),
+  VarianceParams(4, 5, &aom_highbd_8_variance16x32_c, 8),
+  VarianceParams(4, 4, &aom_highbd_8_variance16x16_c, 8),
+  VarianceParams(4, 3, &aom_highbd_8_variance16x8_c, 8),
+  VarianceParams(3, 4, &aom_highbd_8_variance8x16_c, 8),
+  VarianceParams(3, 3, &aom_highbd_8_variance8x8_c, 8),
+  VarianceParams(3, 2, &aom_highbd_8_variance8x4_c, 8),
+  VarianceParams(2, 3, &aom_highbd_8_variance4x8_c, 8),
+  VarianceParams(2, 2, &aom_highbd_8_variance4x4_c, 8)
+};
+INSTANTIATE_TEST_CASE_P(C, AvxHBDVarianceTest,
+                        ::testing::ValuesIn(kArrayHBDVariance_c));
+
+#if HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, AvxHBDVarianceTest,
+    ::testing::Values(
+        VarianceParams(2, 2, &aom_highbd_8_variance4x4_sse4_1, 8),
+        VarianceParams(2, 2, &aom_highbd_10_variance4x4_sse4_1, 10),
+        VarianceParams(2, 2, &aom_highbd_12_variance4x4_sse4_1, 12)));
+#endif  // HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+
+const AvxHBDSubpelVarianceTest::ParamType kArrayHBDSubpelVariance_c[] = {
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(7, 7, &aom_highbd_8_sub_pixel_variance128x128_c, 8),
+  make_tuple(7, 6, &aom_highbd_8_sub_pixel_variance128x64_c, 8),
+  make_tuple(6, 7, &aom_highbd_8_sub_pixel_variance64x128_c, 8),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(6, 6, &aom_highbd_8_sub_pixel_variance64x64_c, 8),
+  make_tuple(6, 5, &aom_highbd_8_sub_pixel_variance64x32_c, 8),
+  make_tuple(5, 6, &aom_highbd_8_sub_pixel_variance32x64_c, 8),
+  make_tuple(5, 5, &aom_highbd_8_sub_pixel_variance32x32_c, 8),
+  make_tuple(5, 4, &aom_highbd_8_sub_pixel_variance32x16_c, 8),
+  make_tuple(4, 5, &aom_highbd_8_sub_pixel_variance16x32_c, 8),
+  make_tuple(4, 4, &aom_highbd_8_sub_pixel_variance16x16_c, 8),
+  make_tuple(4, 3, &aom_highbd_8_sub_pixel_variance16x8_c, 8),
+  make_tuple(3, 4, &aom_highbd_8_sub_pixel_variance8x16_c, 8),
+  make_tuple(3, 3, &aom_highbd_8_sub_pixel_variance8x8_c, 8),
+  make_tuple(3, 2, &aom_highbd_8_sub_pixel_variance8x4_c, 8),
+  make_tuple(2, 3, &aom_highbd_8_sub_pixel_variance4x8_c, 8),
+  make_tuple(2, 2, &aom_highbd_8_sub_pixel_variance4x4_c, 8),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(7, 7, &aom_highbd_10_sub_pixel_variance128x128_c, 10),
+  make_tuple(7, 6, &aom_highbd_10_sub_pixel_variance128x64_c, 10),
+  make_tuple(6, 7, &aom_highbd_10_sub_pixel_variance64x128_c, 10),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(6, 6, &aom_highbd_10_sub_pixel_variance64x64_c, 10),
+  make_tuple(6, 5, &aom_highbd_10_sub_pixel_variance64x32_c, 10),
+  make_tuple(5, 6, &aom_highbd_10_sub_pixel_variance32x64_c, 10),
+  make_tuple(5, 5, &aom_highbd_10_sub_pixel_variance32x32_c, 10),
+  make_tuple(5, 4, &aom_highbd_10_sub_pixel_variance32x16_c, 10),
+  make_tuple(4, 5, &aom_highbd_10_sub_pixel_variance16x32_c, 10),
+  make_tuple(4, 4, &aom_highbd_10_sub_pixel_variance16x16_c, 10),
+  make_tuple(4, 3, &aom_highbd_10_sub_pixel_variance16x8_c, 10),
+  make_tuple(3, 4, &aom_highbd_10_sub_pixel_variance8x16_c, 10),
+  make_tuple(3, 3, &aom_highbd_10_sub_pixel_variance8x8_c, 10),
+  make_tuple(3, 2, &aom_highbd_10_sub_pixel_variance8x4_c, 10),
+  make_tuple(2, 3, &aom_highbd_10_sub_pixel_variance4x8_c, 10),
+  make_tuple(2, 2, &aom_highbd_10_sub_pixel_variance4x4_c, 10),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(7, 7, &aom_highbd_12_sub_pixel_variance128x128_c, 12),
+  make_tuple(7, 6, &aom_highbd_12_sub_pixel_variance128x64_c, 12),
+  make_tuple(6, 7, &aom_highbd_12_sub_pixel_variance64x128_c, 12),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(6, 6, &aom_highbd_12_sub_pixel_variance64x64_c, 12),
+  make_tuple(6, 5, &aom_highbd_12_sub_pixel_variance64x32_c, 12),
+  make_tuple(5, 6, &aom_highbd_12_sub_pixel_variance32x64_c, 12),
+  make_tuple(5, 5, &aom_highbd_12_sub_pixel_variance32x32_c, 12),
+  make_tuple(5, 4, &aom_highbd_12_sub_pixel_variance32x16_c, 12),
+  make_tuple(4, 5, &aom_highbd_12_sub_pixel_variance16x32_c, 12),
+  make_tuple(4, 4, &aom_highbd_12_sub_pixel_variance16x16_c, 12),
+  make_tuple(4, 3, &aom_highbd_12_sub_pixel_variance16x8_c, 12),
+  make_tuple(3, 4, &aom_highbd_12_sub_pixel_variance8x16_c, 12),
+  make_tuple(3, 3, &aom_highbd_12_sub_pixel_variance8x8_c, 12),
+  make_tuple(3, 2, &aom_highbd_12_sub_pixel_variance8x4_c, 12),
+  make_tuple(2, 3, &aom_highbd_12_sub_pixel_variance4x8_c, 12),
+  make_tuple(2, 2, &aom_highbd_12_sub_pixel_variance4x4_c, 12),
+};
+INSTANTIATE_TEST_CASE_P(C, AvxHBDSubpelVarianceTest,
+                        ::testing::ValuesIn(kArrayHBDSubpelVariance_c));
+
+const AvxHBDSubpelAvgVarianceTest::ParamType kArrayHBDSubpelAvgVariance_c[] = {
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(7, 7, &aom_highbd_8_sub_pixel_avg_variance128x128_c, 8),
+  make_tuple(7, 6, &aom_highbd_8_sub_pixel_avg_variance128x64_c, 8),
+  make_tuple(6, 7, &aom_highbd_8_sub_pixel_avg_variance64x128_c, 8),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(6, 6, &aom_highbd_8_sub_pixel_avg_variance64x64_c, 8),
+  make_tuple(6, 5, &aom_highbd_8_sub_pixel_avg_variance64x32_c, 8),
+  make_tuple(5, 6, &aom_highbd_8_sub_pixel_avg_variance32x64_c, 8),
+  make_tuple(5, 5, &aom_highbd_8_sub_pixel_avg_variance32x32_c, 8),
+  make_tuple(5, 4, &aom_highbd_8_sub_pixel_avg_variance32x16_c, 8),
+  make_tuple(4, 5, &aom_highbd_8_sub_pixel_avg_variance16x32_c, 8),
+  make_tuple(4, 4, &aom_highbd_8_sub_pixel_avg_variance16x16_c, 8),
+  make_tuple(4, 3, &aom_highbd_8_sub_pixel_avg_variance16x8_c, 8),
+  make_tuple(3, 4, &aom_highbd_8_sub_pixel_avg_variance8x16_c, 8),
+  make_tuple(3, 3, &aom_highbd_8_sub_pixel_avg_variance8x8_c, 8),
+  make_tuple(3, 2, &aom_highbd_8_sub_pixel_avg_variance8x4_c, 8),
+  make_tuple(2, 3, &aom_highbd_8_sub_pixel_avg_variance4x8_c, 8),
+  make_tuple(2, 2, &aom_highbd_8_sub_pixel_avg_variance4x4_c, 8),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(7, 7, &aom_highbd_10_sub_pixel_avg_variance128x128_c, 10),
+  make_tuple(7, 6, &aom_highbd_10_sub_pixel_avg_variance128x64_c, 10),
+  make_tuple(6, 7, &aom_highbd_10_sub_pixel_avg_variance64x128_c, 10),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(6, 6, &aom_highbd_10_sub_pixel_avg_variance64x64_c, 10),
+  make_tuple(6, 5, &aom_highbd_10_sub_pixel_avg_variance64x32_c, 10),
+  make_tuple(5, 6, &aom_highbd_10_sub_pixel_avg_variance32x64_c, 10),
+  make_tuple(5, 5, &aom_highbd_10_sub_pixel_avg_variance32x32_c, 10),
+  make_tuple(5, 4, &aom_highbd_10_sub_pixel_avg_variance32x16_c, 10),
+  make_tuple(4, 5, &aom_highbd_10_sub_pixel_avg_variance16x32_c, 10),
+  make_tuple(4, 4, &aom_highbd_10_sub_pixel_avg_variance16x16_c, 10),
+  make_tuple(4, 3, &aom_highbd_10_sub_pixel_avg_variance16x8_c, 10),
+  make_tuple(3, 4, &aom_highbd_10_sub_pixel_avg_variance8x16_c, 10),
+  make_tuple(3, 3, &aom_highbd_10_sub_pixel_avg_variance8x8_c, 10),
+  make_tuple(3, 2, &aom_highbd_10_sub_pixel_avg_variance8x4_c, 10),
+  make_tuple(2, 3, &aom_highbd_10_sub_pixel_avg_variance4x8_c, 10),
+  make_tuple(2, 2, &aom_highbd_10_sub_pixel_avg_variance4x4_c, 10),
+#if CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(7, 7, &aom_highbd_12_sub_pixel_avg_variance128x128_c, 12),
+  make_tuple(7, 6, &aom_highbd_12_sub_pixel_avg_variance128x64_c, 12),
+  make_tuple(6, 7, &aom_highbd_12_sub_pixel_avg_variance64x128_c, 12),
+#endif  // CONFIG_AV1 && CONFIG_EXT_PARTITION
+  make_tuple(6, 6, &aom_highbd_12_sub_pixel_avg_variance64x64_c, 12),
+  make_tuple(6, 5, &aom_highbd_12_sub_pixel_avg_variance64x32_c, 12),
+  make_tuple(5, 6, &aom_highbd_12_sub_pixel_avg_variance32x64_c, 12),
+  make_tuple(5, 5, &aom_highbd_12_sub_pixel_avg_variance32x32_c, 12),
+  make_tuple(5, 4, &aom_highbd_12_sub_pixel_avg_variance32x16_c, 12),
+  make_tuple(4, 5, &aom_highbd_12_sub_pixel_avg_variance16x32_c, 12),
+  make_tuple(4, 4, &aom_highbd_12_sub_pixel_avg_variance16x16_c, 12),
+  make_tuple(4, 3, &aom_highbd_12_sub_pixel_avg_variance16x8_c, 12),
+  make_tuple(3, 4, &aom_highbd_12_sub_pixel_avg_variance8x16_c, 12),
+  make_tuple(3, 3, &aom_highbd_12_sub_pixel_avg_variance8x8_c, 12),
+  make_tuple(3, 2, &aom_highbd_12_sub_pixel_avg_variance8x4_c, 12),
+  make_tuple(2, 3, &aom_highbd_12_sub_pixel_avg_variance4x8_c, 12),
+  make_tuple(2, 2, &aom_highbd_12_sub_pixel_avg_variance4x4_c, 12)
+};
+INSTANTIATE_TEST_CASE_P(C, AvxHBDSubpelAvgVarianceTest,
+                        ::testing::ValuesIn(kArrayHBDSubpelAvgVariance_c));
+#endif  // CONFIG_HIGHBITDEPTH
+
+#if HAVE_SSE2
+INSTANTIATE_TEST_CASE_P(SSE2, SumOfSquaresTest,
+                        ::testing::Values(aom_get_mb_ss_sse2));
+
+INSTANTIATE_TEST_CASE_P(SSE2, AvxMseTest,
+                        ::testing::Values(MseParams(4, 4, &aom_mse16x16_sse2),
+                                          MseParams(4, 3, &aom_mse16x8_sse2),
+                                          MseParams(3, 4, &aom_mse8x16_sse2),
+                                          MseParams(3, 3, &aom_mse8x8_sse2)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, AvxVarianceTest,
+    ::testing::Values(VarianceParams(6, 6, &aom_variance64x64_sse2),
+                      VarianceParams(6, 5, &aom_variance64x32_sse2),
+                      VarianceParams(5, 6, &aom_variance32x64_sse2),
+                      VarianceParams(5, 5, &aom_variance32x32_sse2),
+                      VarianceParams(5, 4, &aom_variance32x16_sse2),
+                      VarianceParams(4, 5, &aom_variance16x32_sse2),
+                      VarianceParams(4, 4, &aom_variance16x16_sse2),
+                      VarianceParams(4, 3, &aom_variance16x8_sse2),
+                      VarianceParams(3, 4, &aom_variance8x16_sse2),
+                      VarianceParams(3, 3, &aom_variance8x8_sse2),
+                      VarianceParams(3, 2, &aom_variance8x4_sse2),
+                      VarianceParams(2, 3, &aom_variance4x8_sse2),
+                      VarianceParams(2, 2, &aom_variance4x4_sse2)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, AvxSubpelVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &aom_sub_pixel_variance64x64_sse2, 0),
+                      make_tuple(6, 5, &aom_sub_pixel_variance64x32_sse2, 0),
+                      make_tuple(5, 6, &aom_sub_pixel_variance32x64_sse2, 0),
+                      make_tuple(5, 5, &aom_sub_pixel_variance32x32_sse2, 0),
+                      make_tuple(5, 4, &aom_sub_pixel_variance32x16_sse2, 0),
+                      make_tuple(4, 5, &aom_sub_pixel_variance16x32_sse2, 0),
+                      make_tuple(4, 4, &aom_sub_pixel_variance16x16_sse2, 0),
+                      make_tuple(4, 3, &aom_sub_pixel_variance16x8_sse2, 0),
+                      make_tuple(3, 4, &aom_sub_pixel_variance8x16_sse2, 0),
+                      make_tuple(3, 3, &aom_sub_pixel_variance8x8_sse2, 0),
+                      make_tuple(3, 2, &aom_sub_pixel_variance8x4_sse2, 0),
+                      make_tuple(2, 3, &aom_sub_pixel_variance4x8_sse2, 0),
+                      make_tuple(2, 2, &aom_sub_pixel_variance4x4_sse2, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, AvxSubpelAvgVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &aom_sub_pixel_avg_variance64x64_sse2, 0),
+        make_tuple(6, 5, &aom_sub_pixel_avg_variance64x32_sse2, 0),
+        make_tuple(5, 6, &aom_sub_pixel_avg_variance32x64_sse2, 0),
+        make_tuple(5, 5, &aom_sub_pixel_avg_variance32x32_sse2, 0),
+        make_tuple(5, 4, &aom_sub_pixel_avg_variance32x16_sse2, 0),
+        make_tuple(4, 5, &aom_sub_pixel_avg_variance16x32_sse2, 0),
+        make_tuple(4, 4, &aom_sub_pixel_avg_variance16x16_sse2, 0),
+        make_tuple(4, 3, &aom_sub_pixel_avg_variance16x8_sse2, 0),
+        make_tuple(3, 4, &aom_sub_pixel_avg_variance8x16_sse2, 0),
+        make_tuple(3, 3, &aom_sub_pixel_avg_variance8x8_sse2, 0),
+        make_tuple(3, 2, &aom_sub_pixel_avg_variance8x4_sse2, 0),
+        make_tuple(2, 3, &aom_sub_pixel_avg_variance4x8_sse2, 0),
+        make_tuple(2, 2, &aom_sub_pixel_avg_variance4x4_sse2, 0)));
+
+#if HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, AvxSubpelVarianceTest,
+    ::testing::Values(
+        make_tuple(2, 2, &aom_highbd_8_sub_pixel_variance4x4_sse4_1, 8),
+        make_tuple(2, 2, &aom_highbd_10_sub_pixel_variance4x4_sse4_1, 10),
+        make_tuple(2, 2, &aom_highbd_12_sub_pixel_variance4x4_sse4_1, 12)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE4_1, AvxSubpelAvgVarianceTest,
+    ::testing::Values(
+        make_tuple(2, 2, &aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1, 8),
+        make_tuple(2, 2, &aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1, 10),
+        make_tuple(2, 2, &aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1, 12)));
+#endif  // HAVE_SSE4_1 && CONFIG_HIGHBITDEPTH
+
+#if CONFIG_HIGHBITDEPTH
+/* TODO(debargha): This test does not support the highbd version
+INSTANTIATE_TEST_CASE_P(
+    SSE2, AvxHBDMseTest,
+    ::testing::Values(MseParams(4, 4, &aom_highbd_12_mse16x16_sse2),
+                      MseParams(4, 3, &aom_highbd_12_mse16x8_sse2),
+                      MseParams(3, 4, &aom_highbd_12_mse8x16_sse2),
+                      MseParams(3, 3, &aom_highbd_12_mse8x8_sse2),
+                      MseParams(4, 4, &aom_highbd_10_mse16x16_sse2),
+                      MseParams(4, 3, &aom_highbd_10_mse16x8_sse2),
+                      MseParams(3, 4, &aom_highbd_10_mse8x16_sse2),
+                      MseParams(3, 3, &aom_highbd_10_mse8x8_sse2),
+                      MseParams(4, 4, &aom_highbd_8_mse16x16_sse2),
+                      MseParams(4, 3, &aom_highbd_8_mse16x8_sse2),
+                      MseParams(3, 4, &aom_highbd_8_mse8x16_sse2),
+                      MseParams(3, 3, &aom_highbd_8_mse8x8_sse2)));
+*/
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, AvxHBDVarianceTest,
+    ::testing::Values(
+        VarianceParams(6, 6, &aom_highbd_12_variance64x64_sse2, 12),
+        VarianceParams(6, 5, &aom_highbd_12_variance64x32_sse2, 12),
+        VarianceParams(5, 6, &aom_highbd_12_variance32x64_sse2, 12),
+        VarianceParams(5, 5, &aom_highbd_12_variance32x32_sse2, 12),
+        VarianceParams(5, 4, &aom_highbd_12_variance32x16_sse2, 12),
+        VarianceParams(4, 5, &aom_highbd_12_variance16x32_sse2, 12),
+        VarianceParams(4, 4, &aom_highbd_12_variance16x16_sse2, 12),
+        VarianceParams(4, 3, &aom_highbd_12_variance16x8_sse2, 12),
+        VarianceParams(3, 4, &aom_highbd_12_variance8x16_sse2, 12),
+        VarianceParams(3, 3, &aom_highbd_12_variance8x8_sse2, 12),
+        VarianceParams(6, 6, &aom_highbd_10_variance64x64_sse2, 10),
+        VarianceParams(6, 5, &aom_highbd_10_variance64x32_sse2, 10),
+        VarianceParams(5, 6, &aom_highbd_10_variance32x64_sse2, 10),
+        VarianceParams(5, 5, &aom_highbd_10_variance32x32_sse2, 10),
+        VarianceParams(5, 4, &aom_highbd_10_variance32x16_sse2, 10),
+        VarianceParams(4, 5, &aom_highbd_10_variance16x32_sse2, 10),
+        VarianceParams(4, 4, &aom_highbd_10_variance16x16_sse2, 10),
+        VarianceParams(4, 3, &aom_highbd_10_variance16x8_sse2, 10),
+        VarianceParams(3, 4, &aom_highbd_10_variance8x16_sse2, 10),
+        VarianceParams(3, 3, &aom_highbd_10_variance8x8_sse2, 10),
+        VarianceParams(6, 6, &aom_highbd_8_variance64x64_sse2, 8),
+        VarianceParams(6, 5, &aom_highbd_8_variance64x32_sse2, 8),
+        VarianceParams(5, 6, &aom_highbd_8_variance32x64_sse2, 8),
+        VarianceParams(5, 5, &aom_highbd_8_variance32x32_sse2, 8),
+        VarianceParams(5, 4, &aom_highbd_8_variance32x16_sse2, 8),
+        VarianceParams(4, 5, &aom_highbd_8_variance16x32_sse2, 8),
+        VarianceParams(4, 4, &aom_highbd_8_variance16x16_sse2, 8),
+        VarianceParams(4, 3, &aom_highbd_8_variance16x8_sse2, 8),
+        VarianceParams(3, 4, &aom_highbd_8_variance8x16_sse2, 8),
+        VarianceParams(3, 3, &aom_highbd_8_variance8x8_sse2, 8)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, AvxHBDSubpelVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &aom_highbd_12_sub_pixel_variance64x64_sse2, 12),
+        make_tuple(6, 5, &aom_highbd_12_sub_pixel_variance64x32_sse2, 12),
+        make_tuple(5, 6, &aom_highbd_12_sub_pixel_variance32x64_sse2, 12),
+        make_tuple(5, 5, &aom_highbd_12_sub_pixel_variance32x32_sse2, 12),
+        make_tuple(5, 4, &aom_highbd_12_sub_pixel_variance32x16_sse2, 12),
+        make_tuple(4, 5, &aom_highbd_12_sub_pixel_variance16x32_sse2, 12),
+        make_tuple(4, 4, &aom_highbd_12_sub_pixel_variance16x16_sse2, 12),
+        make_tuple(4, 3, &aom_highbd_12_sub_pixel_variance16x8_sse2, 12),
+        make_tuple(3, 4, &aom_highbd_12_sub_pixel_variance8x16_sse2, 12),
+        make_tuple(3, 3, &aom_highbd_12_sub_pixel_variance8x8_sse2, 12),
+        make_tuple(3, 2, &aom_highbd_12_sub_pixel_variance8x4_sse2, 12),
+        make_tuple(6, 6, &aom_highbd_10_sub_pixel_variance64x64_sse2, 10),
+        make_tuple(6, 5, &aom_highbd_10_sub_pixel_variance64x32_sse2, 10),
+        make_tuple(5, 6, &aom_highbd_10_sub_pixel_variance32x64_sse2, 10),
+        make_tuple(5, 5, &aom_highbd_10_sub_pixel_variance32x32_sse2, 10),
+        make_tuple(5, 4, &aom_highbd_10_sub_pixel_variance32x16_sse2, 10),
+        make_tuple(4, 5, &aom_highbd_10_sub_pixel_variance16x32_sse2, 10),
+        make_tuple(4, 4, &aom_highbd_10_sub_pixel_variance16x16_sse2, 10),
+        make_tuple(4, 3, &aom_highbd_10_sub_pixel_variance16x8_sse2, 10),
+        make_tuple(3, 4, &aom_highbd_10_sub_pixel_variance8x16_sse2, 10),
+        make_tuple(3, 3, &aom_highbd_10_sub_pixel_variance8x8_sse2, 10),
+        make_tuple(3, 2, &aom_highbd_10_sub_pixel_variance8x4_sse2, 10),
+        make_tuple(6, 6, &aom_highbd_8_sub_pixel_variance64x64_sse2, 8),
+        make_tuple(6, 5, &aom_highbd_8_sub_pixel_variance64x32_sse2, 8),
+        make_tuple(5, 6, &aom_highbd_8_sub_pixel_variance32x64_sse2, 8),
+        make_tuple(5, 5, &aom_highbd_8_sub_pixel_variance32x32_sse2, 8),
+        make_tuple(5, 4, &aom_highbd_8_sub_pixel_variance32x16_sse2, 8),
+        make_tuple(4, 5, &aom_highbd_8_sub_pixel_variance16x32_sse2, 8),
+        make_tuple(4, 4, &aom_highbd_8_sub_pixel_variance16x16_sse2, 8),
+        make_tuple(4, 3, &aom_highbd_8_sub_pixel_variance16x8_sse2, 8),
+        make_tuple(3, 4, &aom_highbd_8_sub_pixel_variance8x16_sse2, 8),
+        make_tuple(3, 3, &aom_highbd_8_sub_pixel_variance8x8_sse2, 8),
+        make_tuple(3, 2, &aom_highbd_8_sub_pixel_variance8x4_sse2, 8)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSE2, AvxHBDSubpelAvgVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &aom_highbd_12_sub_pixel_avg_variance64x64_sse2, 12),
+        make_tuple(6, 5, &aom_highbd_12_sub_pixel_avg_variance64x32_sse2, 12),
+        make_tuple(5, 6, &aom_highbd_12_sub_pixel_avg_variance32x64_sse2, 12),
+        make_tuple(5, 5, &aom_highbd_12_sub_pixel_avg_variance32x32_sse2, 12),
+        make_tuple(5, 4, &aom_highbd_12_sub_pixel_avg_variance32x16_sse2, 12),
+        make_tuple(4, 5, &aom_highbd_12_sub_pixel_avg_variance16x32_sse2, 12),
+        make_tuple(4, 4, &aom_highbd_12_sub_pixel_avg_variance16x16_sse2, 12),
+        make_tuple(4, 3, &aom_highbd_12_sub_pixel_avg_variance16x8_sse2, 12),
+        make_tuple(3, 4, &aom_highbd_12_sub_pixel_avg_variance8x16_sse2, 12),
+        make_tuple(3, 3, &aom_highbd_12_sub_pixel_avg_variance8x8_sse2, 12),
+        make_tuple(3, 2, &aom_highbd_12_sub_pixel_avg_variance8x4_sse2, 12),
+        make_tuple(6, 6, &aom_highbd_10_sub_pixel_avg_variance64x64_sse2, 10),
+        make_tuple(6, 5, &aom_highbd_10_sub_pixel_avg_variance64x32_sse2, 10),
+        make_tuple(5, 6, &aom_highbd_10_sub_pixel_avg_variance32x64_sse2, 10),
+        make_tuple(5, 5, &aom_highbd_10_sub_pixel_avg_variance32x32_sse2, 10),
+        make_tuple(5, 4, &aom_highbd_10_sub_pixel_avg_variance32x16_sse2, 10),
+        make_tuple(4, 5, &aom_highbd_10_sub_pixel_avg_variance16x32_sse2, 10),
+        make_tuple(4, 4, &aom_highbd_10_sub_pixel_avg_variance16x16_sse2, 10),
+        make_tuple(4, 3, &aom_highbd_10_sub_pixel_avg_variance16x8_sse2, 10),
+        make_tuple(3, 4, &aom_highbd_10_sub_pixel_avg_variance8x16_sse2, 10),
+        make_tuple(3, 3, &aom_highbd_10_sub_pixel_avg_variance8x8_sse2, 10),
+        make_tuple(3, 2, &aom_highbd_10_sub_pixel_avg_variance8x4_sse2, 10),
+        make_tuple(6, 6, &aom_highbd_8_sub_pixel_avg_variance64x64_sse2, 8),
+        make_tuple(6, 5, &aom_highbd_8_sub_pixel_avg_variance64x32_sse2, 8),
+        make_tuple(5, 6, &aom_highbd_8_sub_pixel_avg_variance32x64_sse2, 8),
+        make_tuple(5, 5, &aom_highbd_8_sub_pixel_avg_variance32x32_sse2, 8),
+        make_tuple(5, 4, &aom_highbd_8_sub_pixel_avg_variance32x16_sse2, 8),
+        make_tuple(4, 5, &aom_highbd_8_sub_pixel_avg_variance16x32_sse2, 8),
+        make_tuple(4, 4, &aom_highbd_8_sub_pixel_avg_variance16x16_sse2, 8),
+        make_tuple(4, 3, &aom_highbd_8_sub_pixel_avg_variance16x8_sse2, 8),
+        make_tuple(3, 4, &aom_highbd_8_sub_pixel_avg_variance8x16_sse2, 8),
+        make_tuple(3, 3, &aom_highbd_8_sub_pixel_avg_variance8x8_sse2, 8),
+        make_tuple(3, 2, &aom_highbd_8_sub_pixel_avg_variance8x4_sse2, 8)));
+#endif  // CONFIG_HIGHBITDEPTH
+#endif  // HAVE_SSE2
+
+#if HAVE_SSSE3
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, AvxSubpelVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &aom_sub_pixel_variance64x64_ssse3, 0),
+                      make_tuple(6, 5, &aom_sub_pixel_variance64x32_ssse3, 0),
+                      make_tuple(5, 6, &aom_sub_pixel_variance32x64_ssse3, 0),
+                      make_tuple(5, 5, &aom_sub_pixel_variance32x32_ssse3, 0),
+                      make_tuple(5, 4, &aom_sub_pixel_variance32x16_ssse3, 0),
+                      make_tuple(4, 5, &aom_sub_pixel_variance16x32_ssse3, 0),
+                      make_tuple(4, 4, &aom_sub_pixel_variance16x16_ssse3, 0),
+                      make_tuple(4, 3, &aom_sub_pixel_variance16x8_ssse3, 0),
+                      make_tuple(3, 4, &aom_sub_pixel_variance8x16_ssse3, 0),
+                      make_tuple(3, 3, &aom_sub_pixel_variance8x8_ssse3, 0),
+                      make_tuple(3, 2, &aom_sub_pixel_variance8x4_ssse3, 0),
+                      make_tuple(2, 3, &aom_sub_pixel_variance4x8_ssse3, 0),
+                      make_tuple(2, 2, &aom_sub_pixel_variance4x4_ssse3, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    SSSE3, AvxSubpelAvgVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &aom_sub_pixel_avg_variance64x64_ssse3, 0),
+        make_tuple(6, 5, &aom_sub_pixel_avg_variance64x32_ssse3, 0),
+        make_tuple(5, 6, &aom_sub_pixel_avg_variance32x64_ssse3, 0),
+        make_tuple(5, 5, &aom_sub_pixel_avg_variance32x32_ssse3, 0),
+        make_tuple(5, 4, &aom_sub_pixel_avg_variance32x16_ssse3, 0),
+        make_tuple(4, 5, &aom_sub_pixel_avg_variance16x32_ssse3, 0),
+        make_tuple(4, 4, &aom_sub_pixel_avg_variance16x16_ssse3, 0),
+        make_tuple(4, 3, &aom_sub_pixel_avg_variance16x8_ssse3, 0),
+        make_tuple(3, 4, &aom_sub_pixel_avg_variance8x16_ssse3, 0),
+        make_tuple(3, 3, &aom_sub_pixel_avg_variance8x8_ssse3, 0),
+        make_tuple(3, 2, &aom_sub_pixel_avg_variance8x4_ssse3, 0),
+        make_tuple(2, 3, &aom_sub_pixel_avg_variance4x8_ssse3, 0),
+        make_tuple(2, 2, &aom_sub_pixel_avg_variance4x4_ssse3, 0)));
+#endif  // HAVE_SSSE3
+
+#if HAVE_AVX2
+INSTANTIATE_TEST_CASE_P(AVX2, AvxMseTest,
+                        ::testing::Values(MseParams(4, 4, &aom_mse16x16_avx2)));
+
+INSTANTIATE_TEST_CASE_P(
+    AVX2, AvxVarianceTest,
+    ::testing::Values(VarianceParams(6, 6, &aom_variance64x64_avx2),
+                      VarianceParams(6, 5, &aom_variance64x32_avx2),
+                      VarianceParams(5, 5, &aom_variance32x32_avx2),
+                      VarianceParams(5, 4, &aom_variance32x16_avx2),
+                      VarianceParams(4, 4, &aom_variance16x16_avx2)));
+
+INSTANTIATE_TEST_CASE_P(
+    AVX2, AvxSubpelVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &aom_sub_pixel_variance64x64_avx2, 0),
+                      make_tuple(5, 5, &aom_sub_pixel_variance32x32_avx2, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    AVX2, AvxSubpelAvgVarianceTest,
+    ::testing::Values(
+        make_tuple(6, 6, &aom_sub_pixel_avg_variance64x64_avx2, 0),
+        make_tuple(5, 5, &aom_sub_pixel_avg_variance32x32_avx2, 0)));
+#endif  // HAVE_AVX2
+
+#if HAVE_MEDIA
+INSTANTIATE_TEST_CASE_P(MEDIA, AvxMseTest,
+                        ::testing::Values(MseParams(4, 4,
+                                                    &aom_mse16x16_media)));
+
+INSTANTIATE_TEST_CASE_P(
+    MEDIA, AvxVarianceTest,
+    ::testing::Values(VarianceParams(4, 4, &aom_variance16x16_media),
+                      VarianceParams(3, 3, &aom_variance8x8_media)));
+
+INSTANTIATE_TEST_CASE_P(
+    MEDIA, AvxSubpelVarianceTest,
+    ::testing::Values(make_tuple(4, 4, &aom_sub_pixel_variance16x16_media, 0),
+                      make_tuple(3, 3, &aom_sub_pixel_variance8x8_media, 0)));
+#endif  // HAVE_MEDIA
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(NEON, AvxSseTest,
+                        ::testing::Values(SseParams(2, 2,
+                                                    &aom_get4x4sse_cs_neon)));
+
+INSTANTIATE_TEST_CASE_P(NEON, AvxMseTest,
+                        ::testing::Values(MseParams(4, 4, &aom_mse16x16_neon)));
+
+INSTANTIATE_TEST_CASE_P(
+    NEON, AvxVarianceTest,
+    ::testing::Values(VarianceParams(6, 6, &aom_variance64x64_neon),
+                      VarianceParams(6, 5, &aom_variance64x32_neon),
+                      VarianceParams(5, 6, &aom_variance32x64_neon),
+                      VarianceParams(5, 5, &aom_variance32x32_neon),
+                      VarianceParams(4, 4, &aom_variance16x16_neon),
+                      VarianceParams(4, 3, &aom_variance16x8_neon),
+                      VarianceParams(3, 4, &aom_variance8x16_neon),
+                      VarianceParams(3, 3, &aom_variance8x8_neon)));
+
+INSTANTIATE_TEST_CASE_P(
+    NEON, AvxSubpelVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &aom_sub_pixel_variance64x64_neon, 0),
+                      make_tuple(5, 5, &aom_sub_pixel_variance32x32_neon, 0),
+                      make_tuple(4, 4, &aom_sub_pixel_variance16x16_neon, 0),
+                      make_tuple(3, 3, &aom_sub_pixel_variance8x8_neon, 0)));
+#endif  // HAVE_NEON
+
+#if HAVE_MSA
+INSTANTIATE_TEST_CASE_P(MSA, SumOfSquaresTest,
+                        ::testing::Values(aom_get_mb_ss_msa));
+
+INSTANTIATE_TEST_CASE_P(MSA, AvxSseTest,
+                        ::testing::Values(SseParams(2, 2,
+                                                    &aom_get4x4sse_cs_msa)));
+
+INSTANTIATE_TEST_CASE_P(MSA, AvxMseTest,
+                        ::testing::Values(MseParams(4, 4, &aom_mse16x16_msa),
+                                          MseParams(4, 3, &aom_mse16x8_msa),
+                                          MseParams(3, 4, &aom_mse8x16_msa),
+                                          MseParams(3, 3, &aom_mse8x8_msa)));
+
+INSTANTIATE_TEST_CASE_P(
+    MSA, AvxVarianceTest,
+    ::testing::Values(VarianceParams(6, 6, &aom_variance64x64_msa),
+                      VarianceParams(6, 5, &aom_variance64x32_msa),
+                      VarianceParams(5, 6, &aom_variance32x64_msa),
+                      VarianceParams(5, 5, &aom_variance32x32_msa),
+                      VarianceParams(5, 4, &aom_variance32x16_msa),
+                      VarianceParams(4, 5, &aom_variance16x32_msa),
+                      VarianceParams(4, 4, &aom_variance16x16_msa),
+                      VarianceParams(4, 3, &aom_variance16x8_msa),
+                      VarianceParams(3, 4, &aom_variance8x16_msa),
+                      VarianceParams(3, 3, &aom_variance8x8_msa),
+                      VarianceParams(3, 2, &aom_variance8x4_msa),
+                      VarianceParams(2, 3, &aom_variance4x8_msa),
+                      VarianceParams(2, 2, &aom_variance4x4_msa)));
+
+INSTANTIATE_TEST_CASE_P(
+    MSA, AvxSubpelVarianceTest,
+    ::testing::Values(make_tuple(2, 2, &aom_sub_pixel_variance4x4_msa, 0),
+                      make_tuple(2, 3, &aom_sub_pixel_variance4x8_msa, 0),
+                      make_tuple(3, 2, &aom_sub_pixel_variance8x4_msa, 0),
+                      make_tuple(3, 3, &aom_sub_pixel_variance8x8_msa, 0),
+                      make_tuple(3, 4, &aom_sub_pixel_variance8x16_msa, 0),
+                      make_tuple(4, 3, &aom_sub_pixel_variance16x8_msa, 0),
+                      make_tuple(4, 4, &aom_sub_pixel_variance16x16_msa, 0),
+                      make_tuple(4, 5, &aom_sub_pixel_variance16x32_msa, 0),
+                      make_tuple(5, 4, &aom_sub_pixel_variance32x16_msa, 0),
+                      make_tuple(5, 5, &aom_sub_pixel_variance32x32_msa, 0),
+                      make_tuple(5, 6, &aom_sub_pixel_variance32x64_msa, 0),
+                      make_tuple(6, 5, &aom_sub_pixel_variance64x32_msa, 0),
+                      make_tuple(6, 6, &aom_sub_pixel_variance64x64_msa, 0)));
+
+INSTANTIATE_TEST_CASE_P(
+    MSA, AvxSubpelAvgVarianceTest,
+    ::testing::Values(make_tuple(6, 6, &aom_sub_pixel_avg_variance64x64_msa, 0),
+                      make_tuple(6, 5, &aom_sub_pixel_avg_variance64x32_msa, 0),
+                      make_tuple(5, 6, &aom_sub_pixel_avg_variance32x64_msa, 0),
+                      make_tuple(5, 5, &aom_sub_pixel_avg_variance32x32_msa, 0),
+                      make_tuple(5, 4, &aom_sub_pixel_avg_variance32x16_msa, 0),
+                      make_tuple(4, 5, &aom_sub_pixel_avg_variance16x32_msa, 0),
+                      make_tuple(4, 4, &aom_sub_pixel_avg_variance16x16_msa, 0),
+                      make_tuple(4, 3, &aom_sub_pixel_avg_variance16x8_msa, 0),
+                      make_tuple(3, 4, &aom_sub_pixel_avg_variance8x16_msa, 0),
+                      make_tuple(3, 3, &aom_sub_pixel_avg_variance8x8_msa, 0),
+                      make_tuple(3, 2, &aom_sub_pixel_avg_variance8x4_msa, 0),
+                      make_tuple(2, 3, &aom_sub_pixel_avg_variance4x8_msa, 0),
+                      make_tuple(2, 2, &aom_sub_pixel_avg_variance4x4_msa, 0)));
+#endif  // HAVE_MSA
+}  // namespace
diff --git a/third_party/aom/test/video_source.h b/third_party/aom/test/video_source.h
new file mode 100644
index 0000000000..e986ffb373
--- /dev/null
+++ b/third_party/aom/test/video_source.h
@@ -0,0 +1,256 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef TEST_VIDEO_SOURCE_H_
+#define TEST_VIDEO_SOURCE_H_
+
+#if defined(_WIN32)
+#undef NOMINMAX
+#define NOMINMAX
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#endif
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+#include "test/acm_random.h"
+#include "aom/aom_encoder.h"
+
+namespace libaom_test {
+
+// Helper macros to ensure LIBAOM_TEST_DATA_PATH is a quoted string.
+// These are undefined right below GetDataPath
+// NOTE: LIBAOM_TEST_DATA_PATH MUST NOT be a quoted string before
+// Stringification or the GetDataPath will fail at runtime
+#define TO_STRING(S) #S
+#define STRINGIFY(S) TO_STRING(S)
+
+// A simple function to encapsulate cross platform retrieval of test data path
+static std::string GetDataPath() {
+  const char *const data_path = getenv("LIBAOM_TEST_DATA_PATH");
+  if (data_path == NULL) {
+#ifdef LIBAOM_TEST_DATA_PATH
+    // In some environments, we cannot set environment variables
+    // Instead, we set the data path by using a preprocessor symbol
+    // which can be set from make files
+    return STRINGIFY(LIBAOM_TEST_DATA_PATH);
+#else
+    return ".";
+#endif
+  }
+  return data_path;
+}
+
+// Undefining stringification macros because they are not used elsewhere
+#undef TO_STRING
+#undef STRINGIFY
+
+inline FILE *OpenTestDataFile(const std::string &file_name) {
+  const std::string path_to_source = GetDataPath() + "/" + file_name;
+  return fopen(path_to_source.c_str(), "rb");
+}
+
+static FILE *GetTempOutFile(std::string *file_name) {
+  file_name->clear();
+#if defined(_WIN32)
+  char fname[MAX_PATH];
+  char tmppath[MAX_PATH];
+  if (GetTempPathA(MAX_PATH, tmppath)) {
+    // Assume for now that the filename generated is unique per process
+    if (GetTempFileNameA(tmppath, "lvx", 0, fname)) {
+      file_name->assign(fname);
+      return fopen(fname, "wb+");
+    }
+  }
+  return NULL;
+#else
+  return tmpfile();
+#endif
+}
+
+class TempOutFile {
+ public:
+  TempOutFile() { file_ = GetTempOutFile(&file_name_); }
+  ~TempOutFile() {
+    CloseFile();
+    if (!file_name_.empty()) {
+      EXPECT_EQ(0, remove(file_name_.c_str()));
+    }
+  }
+  FILE *file() { return file_; }
+  const std::string &file_name() { return file_name_; }
+
+ protected:
+  void CloseFile() {
+    if (file_) {
+      fclose(file_);
+      file_ = NULL;
+    }
+  }
+  FILE *file_;
+  std::string file_name_;
+};
+
+// Abstract base class for test video sources, which provide a stream of
+// aom_image_t images with associated timestamps and duration.
+class VideoSource {
+ public:
+  virtual ~VideoSource() {}
+
+  // Prepare the stream for reading, rewind/open as necessary.
+  virtual void Begin() = 0;
+
+  // Advance the cursor to the next frame
+  virtual void Next() = 0;
+
+  // Get the current video frame, or NULL on End-Of-Stream.
+  virtual aom_image_t *img() const = 0;
+
+  // Get the presentation timestamp of the current frame.
+  virtual aom_codec_pts_t pts() const = 0;
+
+  // Get the current frame's duration
+  virtual unsigned long duration() const = 0;
+
+  // Get the timebase for the stream
+  virtual aom_rational_t timebase() const = 0;
+
+  // Get the current frame counter, starting at 0.
+  virtual unsigned int frame() const = 0;
+
+  // Get the current file limit.
+  virtual unsigned int limit() const = 0;
+};
+
+class DummyVideoSource : public VideoSource {
+ public:
+  DummyVideoSource()
+      : img_(NULL), limit_(100), width_(80), height_(64),
+        format_(AOM_IMG_FMT_I420) {
+    ReallocImage();
+  }
+
+  virtual ~DummyVideoSource() { aom_img_free(img_); }
+
+  virtual void Begin() {
+    frame_ = 0;
+    FillFrame();
+  }
+
+  virtual void Next() {
+    ++frame_;
+    FillFrame();
+  }
+
+  virtual aom_image_t *img() const { return (frame_ < limit_) ? img_ : NULL; }
+
+  // Models a stream where Timebase = 1/FPS, so pts == frame.
+  virtual aom_codec_pts_t pts() const { return frame_; }
+
+  virtual unsigned long duration() const { return 1; }
+
+  virtual aom_rational_t timebase() const {
+    const aom_rational_t t = { 1, 30 };
+    return t;
+  }
+
+  virtual unsigned int frame() const { return frame_; }
+
+  virtual unsigned int limit() const { return limit_; }
+
+  void set_limit(unsigned int limit) { limit_ = limit; }
+
+  void SetSize(unsigned int width, unsigned int height) {
+    if (width != width_ || height != height_) {
+      width_ = width;
+      height_ = height;
+      ReallocImage();
+    }
+  }
+
+  void SetImageFormat(aom_img_fmt_t format) {
+    if (format_ != format) {
+      format_ = format;
+      ReallocImage();
+    }
+  }
+
+ protected:
+  virtual void FillFrame() {
+    if (img_) memset(img_->img_data, 0, raw_sz_);
+  }
+
+  void ReallocImage() {
+    aom_img_free(img_);
+    img_ = aom_img_alloc(NULL, format_, width_, height_, 32);
+    raw_sz_ = ((img_->w + 31) & ~31) * img_->h * img_->bps / 8;
+  }
+
+  aom_image_t *img_;
+  size_t raw_sz_;
+  unsigned int limit_;
+  unsigned int frame_;
+  unsigned int width_;
+  unsigned int height_;
+  aom_img_fmt_t format_;
+};
+
+class RandomVideoSource : public DummyVideoSource {
+ public:
+  RandomVideoSource(int seed = ACMRandom::DeterministicSeed())
+      : rnd_(seed), seed_(seed) {}
+
+ protected:
+  // Reset the RNG to get a matching stream for the second pass
+  virtual void Begin() {
+    frame_ = 0;
+    rnd_.Reset(seed_);
+    FillFrame();
+  }
+
+  // 15 frames of noise, followed by 15 static frames. Reset to 0 rather
+  // than holding previous frames to encourage keyframes to be thrown.
+  virtual void FillFrame() {
+    if (img_) {
+      if (frame_ % 30 < 15)
+        for (size_t i = 0; i < raw_sz_; ++i) img_->img_data[i] = rnd_.Rand8();
+      else
+        memset(img_->img_data, 0, raw_sz_);
+    }
+  }
+
+  ACMRandom rnd_;
+  int seed_;
+};
+
+// Abstract base class for test video sources, which provide a stream of
+// decompressed images to the decoder.
+class CompressedVideoSource {
+ public:
+  virtual ~CompressedVideoSource() {}
+
+  virtual void Init() = 0;
+
+  // Prepare the stream for reading, rewind/open as necessary.
+  virtual void Begin() = 0;
+
+  // Advance the cursor to the next frame
+  virtual void Next() = 0;
+
+  virtual const uint8_t *cxdata() const = 0;
+
+  virtual size_t frame_size() const = 0;
+
+  virtual unsigned int frame_number() const = 0;
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_VIDEO_SOURCE_H_
diff --git a/third_party/aom/test/warp_filter_test.cc b/third_party/aom/test/warp_filter_test.cc
new file mode 100644
index 0000000000..fd6608bfc1
--- /dev/null
+++ b/third_party/aom/test/warp_filter_test.cc
@@ -0,0 +1,39 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/warp_filter_test_util.h"
+
+using std::tr1::tuple;
+using std::tr1::make_tuple;
+using libaom_test::ACMRandom;
+using libaom_test::AV1WarpFilter::AV1WarpFilterTest;
+#if CONFIG_HIGHBITDEPTH
+using libaom_test::AV1HighbdWarpFilter::AV1HighbdWarpFilterTest;
+#endif
+
+namespace {
+
+TEST_P(AV1WarpFilterTest, CheckOutput) { RunCheckOutput(av1_warp_affine_sse2); }
+
+INSTANTIATE_TEST_CASE_P(SSE2, AV1WarpFilterTest,
+                        libaom_test::AV1WarpFilter::GetDefaultParams());
+
+#if CONFIG_HIGHBITDEPTH
+TEST_P(AV1HighbdWarpFilterTest, CheckOutput) {
+  RunCheckOutput(av1_highbd_warp_affine_ssse3);
+}
+
+INSTANTIATE_TEST_CASE_P(SSSE3, AV1HighbdWarpFilterTest,
+                        libaom_test::AV1HighbdWarpFilter::GetDefaultParams());
+#endif
+
+}  // namespace
diff --git a/third_party/aom/test/warp_filter_test_util.cc b/third_party/aom/test/warp_filter_test_util.cc
new file mode 100644
index 0000000000..1ce265b609
--- /dev/null
+++ b/third_party/aom/test/warp_filter_test_util.cc
@@ -0,0 +1,268 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "test/warp_filter_test_util.h"
+
+using std::tr1::tuple;
+using std::tr1::make_tuple;
+using std::vector;
+using libaom_test::ACMRandom;
+using libaom_test::AV1WarpFilter::AV1WarpFilterTest;
+using libaom_test::AV1WarpFilter::WarpTestParam;
+#if CONFIG_HIGHBITDEPTH
+using libaom_test::AV1HighbdWarpFilter::AV1HighbdWarpFilterTest;
+using libaom_test::AV1HighbdWarpFilter::HighbdWarpTestParam;
+#endif
+
+::testing::internal::ParamGenerator<WarpTestParam>
+libaom_test::AV1WarpFilter::GetDefaultParams() {
+  const WarpTestParam defaultParams[] = {
+    make_tuple(4, 4, 50000),  make_tuple(8, 8, 50000),
+    make_tuple(64, 64, 1000), make_tuple(4, 16, 20000),
+    make_tuple(32, 8, 10000),
+  };
+  return ::testing::ValuesIn(defaultParams);
+}
+
+AV1WarpFilterTest::~AV1WarpFilterTest() {}
+void AV1WarpFilterTest::SetUp() { rnd_.Reset(ACMRandom::DeterministicSeed()); }
+
+void AV1WarpFilterTest::TearDown() { libaom_test::ClearSystemState(); }
+
+int32_t AV1WarpFilterTest::random_param(int bits) {
+  // 1 in 8 chance of generating zero (arbitrarily chosen)
+  if (((rnd_.Rand8()) & 7) == 0) return 0;
+  // Otherwise, enerate uniform values in the range
+  // [-(1 << bits), 1] U [1, 1<<bits]
+  int32_t v = 1 + (rnd_.Rand16() & ((1 << bits) - 1));
+  if ((rnd_.Rand8()) & 1) return -v;
+  return v;
+}
+
+void AV1WarpFilterTest::generate_model(int32_t *mat, int16_t *alpha,
+                                       int16_t *beta, int16_t *gamma,
+                                       int16_t *delta) {
+  while (1) {
+    mat[0] = random_param(WARPEDMODEL_PREC_BITS + 6);
+    mat[1] = random_param(WARPEDMODEL_PREC_BITS + 6);
+    mat[2] = (random_param(WARPEDMODEL_PREC_BITS - 3)) +
+             (1 << WARPEDMODEL_PREC_BITS);
+    mat[3] = random_param(WARPEDMODEL_PREC_BITS - 3);
+    // 50/50 chance of generating ROTZOOM vs. AFFINE models
+    if (rnd_.Rand8() & 1) {
+      // AFFINE
+      mat[4] = random_param(WARPEDMODEL_PREC_BITS - 3);
+      mat[5] = (random_param(WARPEDMODEL_PREC_BITS - 3)) +
+               (1 << WARPEDMODEL_PREC_BITS);
+    } else {
+      mat[4] = -mat[3];
+      mat[5] = mat[2];
+    }
+
+    // Calculate the derived parameters and check that they are suitable
+    // for the warp filter.
+    assert(mat[2] != 0);
+
+    *alpha = clamp(mat[2] - (1 << WARPEDMODEL_PREC_BITS), INT16_MIN, INT16_MAX);
+    *beta = clamp(mat[3], INT16_MIN, INT16_MAX);
+    *gamma = clamp(((int64_t)mat[4] * (1 << WARPEDMODEL_PREC_BITS)) / mat[2],
+                   INT16_MIN, INT16_MAX);
+    *delta =
+        clamp(mat[5] - (((int64_t)mat[3] * mat[4] + (mat[2] / 2)) / mat[2]) -
+                  (1 << WARPEDMODEL_PREC_BITS),
+              INT16_MIN, INT16_MAX);
+
+    if ((4 * abs(*alpha) + 7 * abs(*beta) >= (1 << WARPEDMODEL_PREC_BITS)) ||
+        (4 * abs(*gamma) + 4 * abs(*delta) >= (1 << WARPEDMODEL_PREC_BITS)))
+      continue;
+
+    // We have a valid model, so finish
+    return;
+  }
+}
+
+void AV1WarpFilterTest::RunCheckOutput(warp_affine_func test_impl) {
+  const int w = 128, h = 128;
+  const int border = 16;
+  const int stride = w + 2 * border;
+  const int out_w = GET_PARAM(0), out_h = GET_PARAM(1);
+  const int num_iters = GET_PARAM(2);
+  int i, j, sub_x, sub_y;
+
+  uint8_t *input_ = new uint8_t[h * stride];
+  uint8_t *input = input_ + border;
+
+  // The warp functions always write rows with widths that are multiples of 8.
+  // So to avoid a buffer overflow, we may need to pad rows to a multiple of 8.
+  int output_n = ((out_w + 7) & ~7) * out_h;
+  uint8_t *output = new uint8_t[output_n];
+  uint8_t *output2 = new uint8_t[output_n];
+  int32_t mat[8];
+  int16_t alpha, beta, gamma, delta;
+
+  // Generate an input block and extend its borders horizontally
+  for (i = 0; i < h; ++i)
+    for (j = 0; j < w; ++j) input[i * stride + j] = rnd_.Rand8();
+  for (i = 0; i < h; ++i) {
+    memset(input + i * stride - border, input[i * stride], border);
+    memset(input + i * stride + w, input[i * stride + (w - 1)], border);
+  }
+
+  for (i = 0; i < num_iters; ++i) {
+    for (sub_x = 0; sub_x < 2; ++sub_x)
+      for (sub_y = 0; sub_y < 2; ++sub_y) {
+        generate_model(mat, &alpha, &beta, &gamma, &delta);
+        av1_warp_affine_c(mat, input, w, h, stride, output, 32, 32, out_w,
+                          out_h, out_w, sub_x, sub_y, 0, alpha, beta, gamma,
+                          delta);
+        test_impl(mat, input, w, h, stride, output2, 32, 32, out_w, out_h,
+                  out_w, sub_x, sub_y, 0, alpha, beta, gamma, delta);
+
+        for (j = 0; j < out_w * out_h; ++j)
+          ASSERT_EQ(output[j], output2[j])
+              << "Pixel mismatch at index " << j << " = (" << (j % out_w)
+              << ", " << (j / out_w) << ") on iteration " << i;
+      }
+  }
+  delete[] input_;
+  delete[] output;
+  delete[] output2;
+}
+
+#if CONFIG_HIGHBITDEPTH
+::testing::internal::ParamGenerator<HighbdWarpTestParam>
+libaom_test::AV1HighbdWarpFilter::GetDefaultParams() {
+  const HighbdWarpTestParam defaultParams[] = {
+    make_tuple(4, 4, 50000, 8),   make_tuple(8, 8, 50000, 8),
+    make_tuple(64, 64, 1000, 8),  make_tuple(4, 16, 20000, 8),
+    make_tuple(32, 8, 10000, 8),  make_tuple(4, 4, 50000, 10),
+    make_tuple(8, 8, 50000, 10),  make_tuple(64, 64, 1000, 10),
+    make_tuple(4, 16, 20000, 10), make_tuple(32, 8, 10000, 10),
+    make_tuple(4, 4, 50000, 12),  make_tuple(8, 8, 50000, 12),
+    make_tuple(64, 64, 1000, 12), make_tuple(4, 16, 20000, 12),
+    make_tuple(32, 8, 10000, 12),
+  };
+  return ::testing::ValuesIn(defaultParams);
+}
+
+AV1HighbdWarpFilterTest::~AV1HighbdWarpFilterTest() {}
+void AV1HighbdWarpFilterTest::SetUp() {
+  rnd_.Reset(ACMRandom::DeterministicSeed());
+}
+
+void AV1HighbdWarpFilterTest::TearDown() { libaom_test::ClearSystemState(); }
+
+int32_t AV1HighbdWarpFilterTest::random_param(int bits) {
+  // 1 in 8 chance of generating zero (arbitrarily chosen)
+  if (((rnd_.Rand8()) & 7) == 0) return 0;
+  // Otherwise, enerate uniform values in the range
+  // [-(1 << bits), 1] U [1, 1<<bits]
+  int32_t v = 1 + (rnd_.Rand16() & ((1 << bits) - 1));
+  if ((rnd_.Rand8()) & 1) return -v;
+  return v;
+}
+
+void AV1HighbdWarpFilterTest::generate_model(int32_t *mat, int16_t *alpha,
+                                             int16_t *beta, int16_t *gamma,
+                                             int16_t *delta) {
+  while (1) {
+    mat[0] = random_param(WARPEDMODEL_PREC_BITS + 6);
+    mat[1] = random_param(WARPEDMODEL_PREC_BITS + 6);
+    mat[2] = (random_param(WARPEDMODEL_PREC_BITS - 3)) +
+             (1 << WARPEDMODEL_PREC_BITS);
+    mat[3] = random_param(WARPEDMODEL_PREC_BITS - 3);
+    // 50/50 chance of generating ROTZOOM vs. AFFINE models
+    if (rnd_.Rand8() & 1) {
+      // AFFINE
+      mat[4] = random_param(WARPEDMODEL_PREC_BITS - 3);
+      mat[5] = (random_param(WARPEDMODEL_PREC_BITS - 3)) +
+               (1 << WARPEDMODEL_PREC_BITS);
+    } else {
+      mat[4] = -mat[3];
+      mat[5] = mat[2];
+    }
+
+    // Calculate the derived parameters and check that they are suitable
+    // for the warp filter.
+    assert(mat[2] != 0);
+
+    *alpha = clamp(mat[2] - (1 << WARPEDMODEL_PREC_BITS), INT16_MIN, INT16_MAX);
+    *beta = clamp(mat[3], INT16_MIN, INT16_MAX);
+    *gamma = clamp(((int64_t)mat[4] * (1 << WARPEDMODEL_PREC_BITS)) / mat[2],
+                   INT16_MIN, INT16_MAX);
+    *delta =
+        clamp(mat[5] - (((int64_t)mat[3] * mat[4] + (mat[2] / 2)) / mat[2]) -
+                  (1 << WARPEDMODEL_PREC_BITS),
+              INT16_MIN, INT16_MAX);
+
+    if ((4 * abs(*alpha) + 7 * abs(*beta) >= (1 << WARPEDMODEL_PREC_BITS)) ||
+        (4 * abs(*gamma) + 4 * abs(*delta) >= (1 << WARPEDMODEL_PREC_BITS)))
+      continue;
+
+    // We have a valid model, so finish
+    return;
+  }
+}
+
+void AV1HighbdWarpFilterTest::RunCheckOutput(
+    highbd_warp_affine_func test_impl) {
+  const int w = 128, h = 128;
+  const int border = 16;
+  const int stride = w + 2 * border;
+  const int out_w = GET_PARAM(0), out_h = GET_PARAM(1);
+  const int num_iters = GET_PARAM(2);
+  const int bd = GET_PARAM(3);
+  const int mask = (1 << bd) - 1;
+  int i, j, sub_x, sub_y;
+
+  // The warp functions always write rows with widths that are multiples of 8.
+  // So to avoid a buffer overflow, we may need to pad rows to a multiple of 8.
+  int output_n = ((out_w + 7) & ~7) * out_h;
+  uint16_t *input_ = new uint16_t[h * stride];
+  uint16_t *input = input_ + border;
+  uint16_t *output = new uint16_t[output_n];
+  uint16_t *output2 = new uint16_t[output_n];
+  int32_t mat[8];
+  int16_t alpha, beta, gamma, delta;
+
+  // Generate an input block and extend its borders horizontally
+  for (i = 0; i < h; ++i)
+    for (j = 0; j < w; ++j) input[i * stride + j] = rnd_.Rand16() & mask;
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < border; ++j) {
+      input[i * stride - border + j] = input[i * stride];
+      input[i * stride + w + j] = input[i * stride + (w - 1)];
+    }
+  }
+
+  for (i = 0; i < num_iters; ++i) {
+    for (sub_x = 0; sub_x < 2; ++sub_x)
+      for (sub_y = 0; sub_y < 2; ++sub_y) {
+        generate_model(mat, &alpha, &beta, &gamma, &delta);
+
+        av1_highbd_warp_affine_c(mat, input, w, h, stride, output, 32, 32,
+                                 out_w, out_h, out_w, sub_x, sub_y, bd, 0,
+                                 alpha, beta, gamma, delta);
+        test_impl(mat, input, w, h, stride, output2, 32, 32, out_w, out_h,
+                  out_w, sub_x, sub_y, bd, 0, alpha, beta, gamma, delta);
+
+        for (j = 0; j < out_w * out_h; ++j)
+          ASSERT_EQ(output[j], output2[j])
+              << "Pixel mismatch at index " << j << " = (" << (j % out_w)
+              << ", " << (j / out_w) << ") on iteration " << i;
+      }
+  }
+
+  delete[] input_;
+  delete[] output;
+  delete[] output2;
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/test/warp_filter_test_util.h b/third_party/aom/test/warp_filter_test_util.h
new file mode 100644
index 0000000000..6a87e46d05
--- /dev/null
+++ b/third_party/aom/test/warp_filter_test_util.h
@@ -0,0 +1,94 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef TEST_WARP_FILTER_TEST_UTIL_H_
+#define TEST_WARP_FILTER_TEST_UTIL_H_
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+#include "test/acm_random.h"
+#include "test/util.h"
+#include "./av1_rtcd.h"
+#include "./aom_dsp_rtcd.h"
+#include "test/clear_system_state.h"
+#include "test/register_state_check.h"
+
+#include "av1/common/mv.h"
+
+namespace libaom_test {
+
+namespace AV1WarpFilter {
+
+typedef void (*warp_affine_func)(int32_t *mat, uint8_t *ref, int width,
+                                 int height, int stride, uint8_t *pred,
+                                 int p_col, int p_row, int p_width,
+                                 int p_height, int p_stride, int subsampling_x,
+                                 int subsampling_y, int ref_frm, int16_t alpha,
+                                 int16_t beta, int16_t gamma, int16_t delta);
+
+typedef std::tr1::tuple<int, int, int> WarpTestParam;
+
+::testing::internal::ParamGenerator<WarpTestParam> GetDefaultParams();
+
+class AV1WarpFilterTest : public ::testing::TestWithParam<WarpTestParam> {
+ public:
+  virtual ~AV1WarpFilterTest();
+  virtual void SetUp();
+
+  virtual void TearDown();
+
+ protected:
+  int32_t random_param(int bits);
+  void generate_model(int32_t *mat, int16_t *alpha, int16_t *beta,
+                      int16_t *gamma, int16_t *delta);
+
+  void RunCheckOutput(warp_affine_func test_impl);
+
+  libaom_test::ACMRandom rnd_;
+};
+
+}  // namespace AV1WarpFilter
+
+#if CONFIG_HIGHBITDEPTH
+namespace AV1HighbdWarpFilter {
+typedef void (*highbd_warp_affine_func)(
+    int32_t *mat, uint16_t *ref, int width, int height, int stride,
+    uint16_t *pred, int p_col, int p_row, int p_width, int p_height,
+    int p_stride, int subsampling_x, int subsampling_y, int bd, int ref_frm,
+    int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+
+typedef std::tr1::tuple<int, int, int, int> HighbdWarpTestParam;
+
+::testing::internal::ParamGenerator<HighbdWarpTestParam> GetDefaultParams();
+
+class AV1HighbdWarpFilterTest
+    : public ::testing::TestWithParam<HighbdWarpTestParam> {
+ public:
+  virtual ~AV1HighbdWarpFilterTest();
+  virtual void SetUp();
+
+  virtual void TearDown();
+
+ protected:
+  int32_t random_param(int bits);
+  void generate_model(int32_t *mat, int16_t *alpha, int16_t *beta,
+                      int16_t *gamma, int16_t *delta);
+
+  void RunCheckOutput(highbd_warp_affine_func test_impl);
+
+  libaom_test::ACMRandom rnd_;
+};
+
+}  // namespace AV1HighbdWarpFilter
+#endif  // CONFIG_HIGHBITDEPTH
+
+}  // namespace libaom_test
+
+#endif  // TEST_WARP_FILTER_TEST_UTIL_H_
diff --git a/third_party/aom/test/webm_video_source.h b/third_party/aom/test/webm_video_source.h
new file mode 100644
index 0000000000..286f69cbf1
--- /dev/null
+++ b/third_party/aom/test/webm_video_source.h
@@ -0,0 +1,94 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef TEST_WEBM_VIDEO_SOURCE_H_
+#define TEST_WEBM_VIDEO_SOURCE_H_
+#include <cstdarg>
+#include <cstdio>
+#include <cstdlib>
+#include <new>
+#include <string>
+#include "../tools_common.h"
+#include "../webmdec.h"
+#include "test/video_source.h"
+
+namespace libaom_test {
+
+// This class extends VideoSource to allow parsing of WebM files,
+// so that we can do actual file decodes.
+class WebMVideoSource : public CompressedVideoSource {
+ public:
+  explicit WebMVideoSource(const std::string &file_name)
+      : file_name_(file_name), aom_ctx_(new AvxInputContext()),
+        webm_ctx_(new WebmInputContext()), buf_(NULL), buf_sz_(0), frame_(0),
+        end_of_file_(false) {}
+
+  virtual ~WebMVideoSource() {
+    if (aom_ctx_->file != NULL) fclose(aom_ctx_->file);
+    webm_free(webm_ctx_);
+    delete aom_ctx_;
+    delete webm_ctx_;
+  }
+
+  virtual void Init() {}
+
+  virtual void Begin() {
+    aom_ctx_->file = OpenTestDataFile(file_name_);
+    ASSERT_TRUE(aom_ctx_->file != NULL) << "Input file open failed. Filename: "
+                                        << file_name_;
+
+    ASSERT_EQ(file_is_webm(webm_ctx_, aom_ctx_), 1) << "file is not WebM";
+
+    FillFrame();
+  }
+
+  virtual void Next() {
+    ++frame_;
+    FillFrame();
+  }
+
+  void FillFrame() {
+    ASSERT_TRUE(aom_ctx_->file != NULL);
+    const int status = webm_read_frame(webm_ctx_, &buf_, &buf_sz_);
+    ASSERT_GE(status, 0) << "webm_read_frame failed";
+    if (status == 1) {
+      end_of_file_ = true;
+    }
+  }
+
+  void SeekToNextKeyFrame() {
+    ASSERT_TRUE(aom_ctx_->file != NULL);
+    do {
+      const int status = webm_read_frame(webm_ctx_, &buf_, &buf_sz_);
+      ASSERT_GE(status, 0) << "webm_read_frame failed";
+      ++frame_;
+      if (status == 1) {
+        end_of_file_ = true;
+      }
+    } while (!webm_ctx_->is_key_frame && !end_of_file_);
+  }
+
+  virtual const uint8_t *cxdata() const { return end_of_file_ ? NULL : buf_; }
+  virtual size_t frame_size() const { return buf_sz_; }
+  virtual unsigned int frame_number() const { return frame_; }
+
+ protected:
+  std::string file_name_;
+  AvxInputContext *aom_ctx_;
+  WebmInputContext *webm_ctx_;
+  uint8_t *buf_;
+  size_t buf_sz_;
+  unsigned int frame_;
+  bool end_of_file_;
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_WEBM_VIDEO_SOURCE_H_
diff --git a/third_party/aom/test/y4m_test.cc b/third_party/aom/test/y4m_test.cc
new file mode 100644
index 0000000000..fc9fff5142
--- /dev/null
+++ b/third_party/aom/test/y4m_test.cc
@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include <string>
+
+#include "third_party/googletest/src/googletest/include/gtest/gtest.h"
+
+#include "./aom_config.h"
+#include "./y4menc.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#include "test/y4m_video_source.h"
+
+namespace {
+
+using std::string;
+
+static const unsigned int kWidth = 160;
+static const unsigned int kHeight = 90;
+static const unsigned int kFrames = 10;
+
+struct Y4mTestParam {
+  const char *filename;
+  unsigned int bit_depth;
+  aom_img_fmt format;
+  const char *md5raw;
+};
+
+const Y4mTestParam kY4mTestVectors[] = {
+  { "park_joy_90p_8_420.y4m", 8, AOM_IMG_FMT_I420,
+    "e5406275b9fc6bb3436c31d4a05c1cab" },
+  { "park_joy_90p_8_422.y4m", 8, AOM_IMG_FMT_I422,
+    "284a47a47133b12884ec3a14e959a0b6" },
+  { "park_joy_90p_8_444.y4m", 8, AOM_IMG_FMT_I444,
+    "90517ff33843d85de712fd4fe60dbed0" },
+  { "park_joy_90p_10_420.y4m", 10, AOM_IMG_FMT_I42016,
+    "63f21f9f717d8b8631bd2288ee87137b" },
+  { "park_joy_90p_10_422.y4m", 10, AOM_IMG_FMT_I42216,
+    "48ab51fb540aed07f7ff5af130c9b605" },
+  { "park_joy_90p_10_444.y4m", 10, AOM_IMG_FMT_I44416,
+    "067bfd75aa85ff9bae91fa3e0edd1e3e" },
+  { "park_joy_90p_12_420.y4m", 12, AOM_IMG_FMT_I42016,
+    "9e6d8f6508c6e55625f6b697bc461cef" },
+  { "park_joy_90p_12_422.y4m", 12, AOM_IMG_FMT_I42216,
+    "b239c6b301c0b835485be349ca83a7e3" },
+  { "park_joy_90p_12_444.y4m", 12, AOM_IMG_FMT_I44416,
+    "5a6481a550821dab6d0192f5c63845e9" },
+};
+
+static void write_image_file(const aom_image_t *img, FILE *file) {
+  int plane, y;
+  for (plane = 0; plane < 3; ++plane) {
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int bytes_per_sample = (img->fmt & AOM_IMG_FMT_HIGHBITDEPTH) ? 2 : 1;
+    const int h =
+        (plane ? (img->d_h + img->y_chroma_shift) >> img->y_chroma_shift
+               : img->d_h);
+    const int w =
+        (plane ? (img->d_w + img->x_chroma_shift) >> img->x_chroma_shift
+               : img->d_w);
+    for (y = 0; y < h; ++y) {
+      fwrite(buf, bytes_per_sample, w, file);
+      buf += stride;
+    }
+  }
+}
+
+class Y4mVideoSourceTest : public ::testing::TestWithParam<Y4mTestParam>,
+                           public ::libaom_test::Y4mVideoSource {
+ protected:
+  Y4mVideoSourceTest() : Y4mVideoSource("", 0, 0) {}
+
+  virtual ~Y4mVideoSourceTest() { CloseSource(); }
+
+  virtual void Init(const std::string &file_name, int limit) {
+    file_name_ = file_name;
+    start_ = 0;
+    limit_ = limit;
+    frame_ = 0;
+    Begin();
+  }
+
+  // Checks y4m header information
+  void HeaderChecks(unsigned int bit_depth, aom_img_fmt_t fmt) {
+    ASSERT_TRUE(input_file_ != NULL);
+    ASSERT_EQ(y4m_.pic_w, (int)kWidth);
+    ASSERT_EQ(y4m_.pic_h, (int)kHeight);
+    ASSERT_EQ(img()->d_w, kWidth);
+    ASSERT_EQ(img()->d_h, kHeight);
+    ASSERT_EQ(y4m_.bit_depth, bit_depth);
+    ASSERT_EQ(y4m_.aom_fmt, fmt);
+    if (fmt == AOM_IMG_FMT_I420 || fmt == AOM_IMG_FMT_I42016) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 3 / 2);
+      ASSERT_EQ(img()->x_chroma_shift, 1U);
+      ASSERT_EQ(img()->y_chroma_shift, 1U);
+    }
+    if (fmt == AOM_IMG_FMT_I422 || fmt == AOM_IMG_FMT_I42216) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 2);
+      ASSERT_EQ(img()->x_chroma_shift, 1U);
+      ASSERT_EQ(img()->y_chroma_shift, 0U);
+    }
+    if (fmt == AOM_IMG_FMT_I444 || fmt == AOM_IMG_FMT_I44416) {
+      ASSERT_EQ(y4m_.bps, (int)y4m_.bit_depth * 3);
+      ASSERT_EQ(img()->x_chroma_shift, 0U);
+      ASSERT_EQ(img()->y_chroma_shift, 0U);
+    }
+  }
+
+  // Checks MD5 of the raw frame data
+  void Md5Check(const string &expected_md5) {
+    ASSERT_TRUE(input_file_ != NULL);
+    libaom_test::MD5 md5;
+    for (unsigned int i = start_; i < limit_; i++) {
+      md5.Add(img());
+      Next();
+    }
+    ASSERT_EQ(string(md5.Get()), expected_md5);
+  }
+};
+
+TEST_P(Y4mVideoSourceTest, SourceTest) {
+  const Y4mTestParam t = GetParam();
+  Init(t.filename, kFrames);
+  HeaderChecks(t.bit_depth, t.format);
+  Md5Check(t.md5raw);
+}
+
+INSTANTIATE_TEST_CASE_P(C, Y4mVideoSourceTest,
+                        ::testing::ValuesIn(kY4mTestVectors));
+
+class Y4mVideoWriteTest : public Y4mVideoSourceTest {
+ protected:
+  Y4mVideoWriteTest() : tmpfile_(NULL) {}
+
+  virtual ~Y4mVideoWriteTest() {
+    delete tmpfile_;
+    input_file_ = NULL;
+  }
+
+  void ReplaceInputFile(FILE *input_file) {
+    CloseSource();
+    frame_ = 0;
+    input_file_ = input_file;
+    rewind(input_file_);
+    ReadSourceToStart();
+  }
+
+  // Writes out a y4m file and then reads it back
+  void WriteY4mAndReadBack() {
+    ASSERT_TRUE(input_file_ != NULL);
+    char buf[Y4M_BUFFER_SIZE] = { 0 };
+    const struct AvxRational framerate = { y4m_.fps_n, y4m_.fps_d };
+    tmpfile_ = new libaom_test::TempOutFile;
+    ASSERT_TRUE(tmpfile_->file() != NULL);
+    y4m_write_file_header(buf, sizeof(buf), kWidth, kHeight, &framerate,
+                          y4m_.aom_fmt, y4m_.bit_depth);
+    fputs(buf, tmpfile_->file());
+    for (unsigned int i = start_; i < limit_; i++) {
+      y4m_write_frame_header(buf, sizeof(buf));
+      fputs(buf, tmpfile_->file());
+      write_image_file(img(), tmpfile_->file());
+      Next();
+    }
+    ReplaceInputFile(tmpfile_->file());
+  }
+
+  virtual void Init(const std::string &file_name, int limit) {
+    Y4mVideoSourceTest::Init(file_name, limit);
+    WriteY4mAndReadBack();
+  }
+  libaom_test::TempOutFile *tmpfile_;
+};
+
+TEST_P(Y4mVideoWriteTest, WriteTest) {
+  const Y4mTestParam t = GetParam();
+  Init(t.filename, kFrames);
+  HeaderChecks(t.bit_depth, t.format);
+  Md5Check(t.md5raw);
+}
+
+INSTANTIATE_TEST_CASE_P(C, Y4mVideoWriteTest,
+                        ::testing::ValuesIn(kY4mTestVectors));
+}  // namespace
diff --git a/third_party/aom/test/y4m_video_source.h b/third_party/aom/test/y4m_video_source.h
new file mode 100644
index 0000000000..2279d7970e
--- /dev/null
+++ b/third_party/aom/test/y4m_video_source.h
@@ -0,0 +1,123 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef TEST_Y4M_VIDEO_SOURCE_H_
+#define TEST_Y4M_VIDEO_SOURCE_H_
+#include <algorithm>
+#include <string>
+
+#include "test/video_source.h"
+#include "./y4minput.h"
+
+namespace libaom_test {
+
+// This class extends VideoSource to allow parsing of raw yv12
+// so that we can do actual file encodes.
+class Y4mVideoSource : public VideoSource {
+ public:
+  Y4mVideoSource(const std::string &file_name, unsigned int start, int limit)
+      : file_name_(file_name), input_file_(NULL), img_(new aom_image_t()),
+        start_(start), limit_(limit), frame_(0), framerate_numerator_(0),
+        framerate_denominator_(0), y4m_() {}
+
+  virtual ~Y4mVideoSource() {
+    aom_img_free(img_.get());
+    CloseSource();
+  }
+
+  virtual void OpenSource() {
+    CloseSource();
+    input_file_ = OpenTestDataFile(file_name_);
+    ASSERT_TRUE(input_file_ != NULL) << "Input file open failed. Filename: "
+                                     << file_name_;
+  }
+
+  virtual void ReadSourceToStart() {
+    ASSERT_TRUE(input_file_ != NULL);
+    ASSERT_FALSE(y4m_input_open(&y4m_, input_file_, NULL, 0, 0));
+    framerate_numerator_ = y4m_.fps_n;
+    framerate_denominator_ = y4m_.fps_d;
+    frame_ = 0;
+    for (unsigned int i = 0; i < start_; i++) {
+      Next();
+    }
+    FillFrame();
+  }
+
+  virtual void Begin() {
+    OpenSource();
+    ReadSourceToStart();
+  }
+
+  virtual void Next() {
+    ++frame_;
+    FillFrame();
+  }
+
+  virtual aom_image_t *img() const {
+    return (frame_ < limit_) ? img_.get() : NULL;
+  }
+
+  // Models a stream where Timebase = 1/FPS, so pts == frame.
+  virtual aom_codec_pts_t pts() const { return frame_; }
+
+  virtual unsigned long duration() const { return 1; }
+
+  virtual aom_rational_t timebase() const {
+    const aom_rational_t t = { framerate_denominator_, framerate_numerator_ };
+    return t;
+  }
+
+  virtual unsigned int frame() const { return frame_; }
+
+  virtual unsigned int limit() const { return limit_; }
+
+  virtual void FillFrame() {
+    ASSERT_TRUE(input_file_ != NULL);
+    // Read a frame from input_file.
+    y4m_input_fetch_frame(&y4m_, input_file_, img_.get());
+  }
+
+  // Swap buffers with another y4m source. This allows reading a new frame
+  // while keeping the old frame around. A whole Y4mSource is required and
+  // not just a aom_image_t because of how the y4m reader manipulates
+  // aom_image_t internals,
+  void SwapBuffers(Y4mVideoSource *other) {
+    std::swap(other->y4m_.dst_buf, y4m_.dst_buf);
+    aom_image_t *tmp;
+    tmp = other->img_.release();
+    other->img_.reset(img_.release());
+    img_.reset(tmp);
+  }
+
+ protected:
+  void CloseSource() {
+    y4m_input_close(&y4m_);
+    y4m_ = y4m_input();
+    if (input_file_ != NULL) {
+      fclose(input_file_);
+      input_file_ = NULL;
+    }
+  }
+
+  std::string file_name_;
+  FILE *input_file_;
+  testing::internal::scoped_ptr<aom_image_t> img_;
+  unsigned int start_;
+  unsigned int limit_;
+  unsigned int frame_;
+  int framerate_numerator_;
+  int framerate_denominator_;
+  y4m_input y4m_;
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_Y4M_VIDEO_SOURCE_H_
diff --git a/third_party/aom/test/yuv_video_source.h b/third_party/aom/test/yuv_video_source.h
new file mode 100644
index 0000000000..9ff76a8d8f
--- /dev/null
+++ b/third_party/aom/test/yuv_video_source.h
@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef TEST_YUV_VIDEO_SOURCE_H_
+#define TEST_YUV_VIDEO_SOURCE_H_
+
+#include <cstdio>
+#include <cstdlib>
+#include <string>
+
+#include "test/video_source.h"
+#include "aom/aom_image.h"
+
+namespace libaom_test {
+
+// This class extends VideoSource to allow parsing of raw YUV
+// formats of various color sampling and bit-depths so that we can
+// do actual file encodes.
+class YUVVideoSource : public VideoSource {
+ public:
+  YUVVideoSource(const std::string &file_name, aom_img_fmt format,
+                 unsigned int width, unsigned int height, int rate_numerator,
+                 int rate_denominator, unsigned int start, int limit)
+      : file_name_(file_name), input_file_(NULL), img_(NULL), start_(start),
+        limit_(limit), frame_(0), width_(0), height_(0),
+        format_(AOM_IMG_FMT_NONE), framerate_numerator_(rate_numerator),
+        framerate_denominator_(rate_denominator) {
+    // This initializes format_, raw_size_, width_, height_ and allocates img.
+    SetSize(width, height, format);
+  }
+
+  virtual ~YUVVideoSource() {
+    aom_img_free(img_);
+    if (input_file_) fclose(input_file_);
+  }
+
+  virtual void Begin() {
+    if (input_file_) fclose(input_file_);
+    input_file_ = OpenTestDataFile(file_name_);
+    ASSERT_TRUE(input_file_ != NULL) << "Input file open failed. Filename: "
+                                     << file_name_;
+    if (start_)
+      fseek(input_file_, static_cast<unsigned>(raw_size_) * start_, SEEK_SET);
+
+    frame_ = start_;
+    FillFrame();
+  }
+
+  virtual void Next() {
+    ++frame_;
+    FillFrame();
+  }
+
+  virtual aom_image_t *img() const { return (frame_ < limit_) ? img_ : NULL; }
+
+  // Models a stream where Timebase = 1/FPS, so pts == frame.
+  virtual aom_codec_pts_t pts() const { return frame_; }
+
+  virtual unsigned long duration() const { return 1; }
+
+  virtual aom_rational_t timebase() const {
+    const aom_rational_t t = { framerate_denominator_, framerate_numerator_ };
+    return t;
+  }
+
+  virtual unsigned int frame() const { return frame_; }
+
+  virtual unsigned int limit() const { return limit_; }
+
+  virtual void SetSize(unsigned int width, unsigned int height,
+                       aom_img_fmt format) {
+    if (width != width_ || height != height_ || format != format_) {
+      aom_img_free(img_);
+      img_ = aom_img_alloc(NULL, format, width, height, 1);
+      ASSERT_TRUE(img_ != NULL);
+      width_ = width;
+      height_ = height;
+      format_ = format;
+      switch (format) {
+        case AOM_IMG_FMT_I420: raw_size_ = width * height * 3 / 2; break;
+        case AOM_IMG_FMT_I422: raw_size_ = width * height * 2; break;
+        case AOM_IMG_FMT_I440: raw_size_ = width * height * 2; break;
+        case AOM_IMG_FMT_I444: raw_size_ = width * height * 3; break;
+        case AOM_IMG_FMT_I42016: raw_size_ = width * height * 3; break;
+        case AOM_IMG_FMT_I42216: raw_size_ = width * height * 4; break;
+        case AOM_IMG_FMT_I44016: raw_size_ = width * height * 4; break;
+        case AOM_IMG_FMT_I44416: raw_size_ = width * height * 6; break;
+        default: ASSERT_TRUE(0);
+      }
+    }
+  }
+
+  virtual void FillFrame() {
+    ASSERT_TRUE(input_file_ != NULL);
+    // Read a frame from input_file.
+    if (fread(img_->img_data, raw_size_, 1, input_file_) == 0) {
+      limit_ = frame_;
+    }
+  }
+
+ protected:
+  std::string file_name_;
+  FILE *input_file_;
+  aom_image_t *img_;
+  size_t raw_size_;
+  unsigned int start_;
+  unsigned int limit_;
+  unsigned int frame_;
+  unsigned int width_;
+  unsigned int height_;
+  aom_img_fmt format_;
+  int framerate_numerator_;
+  int framerate_denominator_;
+};
+
+}  // namespace libaom_test
+
+#endif  // TEST_YUV_VIDEO_SOURCE_H_
diff --git a/third_party/aom/third_party/fastfeat/LICENSE b/third_party/aom/third_party/fastfeat/LICENSE
new file mode 100644
index 0000000000..f347008d6e
--- /dev/null
+++ b/third_party/aom/third_party/fastfeat/LICENSE
@@ -0,0 +1,30 @@
+Copyright (c) 2006, 2008 Edward Rosten
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+
+
+	*Redistributions of source code must retain the above copyright
+	 notice, this list of conditions and the following disclaimer.
+
+	*Redistributions in binary form must reproduce the above copyright
+	 notice, this list of conditions and the following disclaimer in the
+	 documentation and/or other materials provided with the distribution.
+
+	*Neither the name of the University of Cambridge nor the names of 
+	 its contributors may be used to endorse or promote products derived 
+	 from this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/third_party/aom/third_party/fastfeat/README.libvpx b/third_party/aom/third_party/fastfeat/README.libvpx
new file mode 100644
index 0000000000..2edd6e7bfb
--- /dev/null
+++ b/third_party/aom/third_party/fastfeat/README.libvpx
@@ -0,0 +1,38 @@
+URL: https://github.com/edrosten/fast-C-src
+Version: 391d5e939eb1545d24c10533d7de424db8d9c191
+License: BSD
+License File: LICENSE
+
+Description:
+Library to compute FAST features with non-maximum suppression.
+
+The files are valid C and C++ code, and have no special requirements for
+compiling, and they do not depend on any libraries. Just compile them along with
+the rest of your project.
+
+To use the functions, #include "fast.h"
+
+The corner detectors have the following prototype (where X is 9, 10, 11 or 12):
+
+xy* fastX_detect_nonmax(const unsigned char * data, int xsize, int ysize, int stride, int threshold, int* numcorners)
+
+Where xy is the following simple struct typedef:
+
+typedef struct
+{
+	int x, y;
+} xy;
+
+The image is passed in as a block of data and dimensions, and the list of
+corners is returned as an array of xy structs, and an integer (numcorners)
+with the number of corners returned.  The data can be deallocated with free().
+Nonmaximal suppression is performed on the corners. Note that the stride
+is the number of bytes between rows. If your image has no padding, then this
+is the same as xsize.
+
+The detection, scoring and nonmaximal suppression are available as individual
+functions.  To see how to use the individual functions, see fast.c
+
+Local Modifications:
+Add lines to turn off clang formatting for these files
+Remove Fast 10, 11 and 12
diff --git a/third_party/aom/third_party/fastfeat/fast.c b/third_party/aom/third_party/fastfeat/fast.c
new file mode 100644
index 0000000000..0d7efc154d
--- /dev/null
+++ b/third_party/aom/third_party/fastfeat/fast.c
@@ -0,0 +1,22 @@
+// clang-format off
+#include <stdlib.h>
+#include "fast.h"
+
+
+xy* fast9_detect_nonmax(const byte* im, int xsize, int ysize, int stride, int b, int* ret_num_corners)
+{
+	xy* corners;
+	int num_corners;
+	int* scores;
+	xy* nonmax;
+
+	corners = fast9_detect(im, xsize, ysize, stride, b, &num_corners);
+	scores = fast9_score(im, stride, corners, num_corners, b);
+	nonmax = nonmax_suppression(corners, scores, num_corners, ret_num_corners);
+
+	free(corners);
+	free(scores);
+
+	return nonmax;
+}
+// clang-format on
diff --git a/third_party/aom/third_party/fastfeat/fast.h b/third_party/aom/third_party/fastfeat/fast.h
new file mode 100644
index 0000000000..a00730e3d3
--- /dev/null
+++ b/third_party/aom/third_party/fastfeat/fast.h
@@ -0,0 +1,20 @@
+// clang-format off
+#ifndef FAST_H
+#define FAST_H
+
+typedef struct { int x, y; } xy;
+typedef unsigned char byte;
+
+int fast9_corner_score(const byte* p, const int pixel[], int bstart);
+
+xy* fast9_detect(const byte* im, int xsize, int ysize, int stride, int b, int* ret_num_corners);
+
+int* fast9_score(const byte* i, int stride, xy* corners, int num_corners, int b);
+
+xy* fast9_detect_nonmax(const byte* im, int xsize, int ysize, int stride, int b, int* ret_num_corners);
+
+xy* nonmax_suppression(const xy* corners, const int* scores, int num_corners, int* ret_num_nonmax);
+
+
+#endif
+// clang-format on
diff --git a/third_party/aom/third_party/fastfeat/fast_9.c b/third_party/aom/third_party/fastfeat/fast_9.c
new file mode 100644
index 0000000000..36aee19fed
--- /dev/null
+++ b/third_party/aom/third_party/fastfeat/fast_9.c
@@ -0,0 +1,5911 @@
+// clang-format off
+/*This is mechanically generated code*/
+#include <stdlib.h>
+
+typedef struct { int x, y; } xy;
+typedef unsigned char byte;
+
+int fast9_corner_score(const byte* p, const int pixel[], int bstart)
+{
+    int bmin = bstart;
+    int bmax = 255;
+    int b = (bmax + bmin)/2;
+
+    /*Compute the score using binary search*/
+	for(;;)
+    {
+		int cb = *p + b;
+		int c_b= *p - b;
+
+
+        if( p[pixel[0]] > cb)
+         if( p[pixel[1]] > cb)
+          if( p[pixel[2]] > cb)
+           if( p[pixel[3]] > cb)
+            if( p[pixel[4]] > cb)
+             if( p[pixel[5]] > cb)
+              if( p[pixel[6]] > cb)
+               if( p[pixel[7]] > cb)
+                if( p[pixel[8]] > cb)
+                 goto is_a_corner;
+                else
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else if( p[pixel[7]] < c_b)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else if( p[pixel[14]] < c_b)
+                 if( p[pixel[8]] < c_b)
+                  if( p[pixel[9]] < c_b)
+                   if( p[pixel[10]] < c_b)
+                    if( p[pixel[11]] < c_b)
+                     if( p[pixel[12]] < c_b)
+                      if( p[pixel[13]] < c_b)
+                       if( p[pixel[15]] < c_b)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else if( p[pixel[6]] < c_b)
+               if( p[pixel[15]] > cb)
+                if( p[pixel[13]] > cb)
+                 if( p[pixel[14]] > cb)
+                  goto is_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else if( p[pixel[13]] < c_b)
+                 if( p[pixel[7]] < c_b)
+                  if( p[pixel[8]] < c_b)
+                   if( p[pixel[9]] < c_b)
+                    if( p[pixel[10]] < c_b)
+                     if( p[pixel[11]] < c_b)
+                      if( p[pixel[12]] < c_b)
+                       if( p[pixel[14]] < c_b)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                if( p[pixel[7]] < c_b)
+                 if( p[pixel[8]] < c_b)
+                  if( p[pixel[9]] < c_b)
+                   if( p[pixel[10]] < c_b)
+                    if( p[pixel[11]] < c_b)
+                     if( p[pixel[12]] < c_b)
+                      if( p[pixel[13]] < c_b)
+                       if( p[pixel[14]] < c_b)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[13]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else if( p[pixel[13]] < c_b)
+                if( p[pixel[7]] < c_b)
+                 if( p[pixel[8]] < c_b)
+                  if( p[pixel[9]] < c_b)
+                   if( p[pixel[10]] < c_b)
+                    if( p[pixel[11]] < c_b)
+                     if( p[pixel[12]] < c_b)
+                      if( p[pixel[14]] < c_b)
+                       if( p[pixel[15]] < c_b)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else if( p[pixel[5]] < c_b)
+              if( p[pixel[14]] > cb)
+               if( p[pixel[12]] > cb)
+                if( p[pixel[13]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      if( p[pixel[10]] > cb)
+                       if( p[pixel[11]] > cb)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else if( p[pixel[12]] < c_b)
+                if( p[pixel[6]] < c_b)
+                 if( p[pixel[7]] < c_b)
+                  if( p[pixel[8]] < c_b)
+                   if( p[pixel[9]] < c_b)
+                    if( p[pixel[10]] < c_b)
+                     if( p[pixel[11]] < c_b)
+                      if( p[pixel[13]] < c_b)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else if( p[pixel[14]] < c_b)
+               if( p[pixel[7]] < c_b)
+                if( p[pixel[8]] < c_b)
+                 if( p[pixel[9]] < c_b)
+                  if( p[pixel[10]] < c_b)
+                   if( p[pixel[11]] < c_b)
+                    if( p[pixel[12]] < c_b)
+                     if( p[pixel[13]] < c_b)
+                      if( p[pixel[6]] < c_b)
+                       goto is_a_corner;
+                      else
+                       if( p[pixel[15]] < c_b)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               if( p[pixel[6]] < c_b)
+                if( p[pixel[7]] < c_b)
+                 if( p[pixel[8]] < c_b)
+                  if( p[pixel[9]] < c_b)
+                   if( p[pixel[10]] < c_b)
+                    if( p[pixel[11]] < c_b)
+                     if( p[pixel[12]] < c_b)
+                      if( p[pixel[13]] < c_b)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              if( p[pixel[12]] > cb)
+               if( p[pixel[13]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      if( p[pixel[10]] > cb)
+                       if( p[pixel[11]] > cb)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else if( p[pixel[12]] < c_b)
+               if( p[pixel[7]] < c_b)
+                if( p[pixel[8]] < c_b)
+                 if( p[pixel[9]] < c_b)
+                  if( p[pixel[10]] < c_b)
+                   if( p[pixel[11]] < c_b)
+                    if( p[pixel[13]] < c_b)
+                     if( p[pixel[14]] < c_b)
+                      if( p[pixel[6]] < c_b)
+                       goto is_a_corner;
+                      else
+                       if( p[pixel[15]] < c_b)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else if( p[pixel[4]] < c_b)
+             if( p[pixel[13]] > cb)
+              if( p[pixel[11]] > cb)
+               if( p[pixel[12]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      if( p[pixel[10]] > cb)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      if( p[pixel[10]] > cb)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else if( p[pixel[11]] < c_b)
+               if( p[pixel[5]] < c_b)
+                if( p[pixel[6]] < c_b)
+                 if( p[pixel[7]] < c_b)
+                  if( p[pixel[8]] < c_b)
+                   if( p[pixel[9]] < c_b)
+                    if( p[pixel[10]] < c_b)
+                     if( p[pixel[12]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else if( p[pixel[13]] < c_b)
+              if( p[pixel[7]] < c_b)
+               if( p[pixel[8]] < c_b)
+                if( p[pixel[9]] < c_b)
+                 if( p[pixel[10]] < c_b)
+                  if( p[pixel[11]] < c_b)
+                   if( p[pixel[12]] < c_b)
+                    if( p[pixel[6]] < c_b)
+                     if( p[pixel[5]] < c_b)
+                      goto is_a_corner;
+                     else
+                      if( p[pixel[14]] < c_b)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                    else
+                     if( p[pixel[14]] < c_b)
+                      if( p[pixel[15]] < c_b)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              if( p[pixel[5]] < c_b)
+               if( p[pixel[6]] < c_b)
+                if( p[pixel[7]] < c_b)
+                 if( p[pixel[8]] < c_b)
+                  if( p[pixel[9]] < c_b)
+                   if( p[pixel[10]] < c_b)
+                    if( p[pixel[11]] < c_b)
+                     if( p[pixel[12]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else
+             if( p[pixel[11]] > cb)
+              if( p[pixel[12]] > cb)
+               if( p[pixel[13]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      if( p[pixel[10]] > cb)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      if( p[pixel[10]] > cb)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else if( p[pixel[11]] < c_b)
+              if( p[pixel[7]] < c_b)
+               if( p[pixel[8]] < c_b)
+                if( p[pixel[9]] < c_b)
+                 if( p[pixel[10]] < c_b)
+                  if( p[pixel[12]] < c_b)
+                   if( p[pixel[13]] < c_b)
+                    if( p[pixel[6]] < c_b)
+                     if( p[pixel[5]] < c_b)
+                      goto is_a_corner;
+                     else
+                      if( p[pixel[14]] < c_b)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                    else
+                     if( p[pixel[14]] < c_b)
+                      if( p[pixel[15]] < c_b)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+           else if( p[pixel[3]] < c_b)
+            if( p[pixel[10]] > cb)
+             if( p[pixel[11]] > cb)
+              if( p[pixel[12]] > cb)
+               if( p[pixel[13]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else if( p[pixel[10]] < c_b)
+             if( p[pixel[7]] < c_b)
+              if( p[pixel[8]] < c_b)
+               if( p[pixel[9]] < c_b)
+                if( p[pixel[11]] < c_b)
+                 if( p[pixel[6]] < c_b)
+                  if( p[pixel[5]] < c_b)
+                   if( p[pixel[4]] < c_b)
+                    goto is_a_corner;
+                   else
+                    if( p[pixel[12]] < c_b)
+                     if( p[pixel[13]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                  else
+                   if( p[pixel[12]] < c_b)
+                    if( p[pixel[13]] < c_b)
+                     if( p[pixel[14]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                 else
+                  if( p[pixel[12]] < c_b)
+                   if( p[pixel[13]] < c_b)
+                    if( p[pixel[14]] < c_b)
+                     if( p[pixel[15]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            if( p[pixel[10]] > cb)
+             if( p[pixel[11]] > cb)
+              if( p[pixel[12]] > cb)
+               if( p[pixel[13]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     if( p[pixel[9]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else if( p[pixel[10]] < c_b)
+             if( p[pixel[7]] < c_b)
+              if( p[pixel[8]] < c_b)
+               if( p[pixel[9]] < c_b)
+                if( p[pixel[11]] < c_b)
+                 if( p[pixel[12]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[5]] < c_b)
+                    if( p[pixel[4]] < c_b)
+                     goto is_a_corner;
+                    else
+                     if( p[pixel[13]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                   else
+                    if( p[pixel[13]] < c_b)
+                     if( p[pixel[14]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                  else
+                   if( p[pixel[13]] < c_b)
+                    if( p[pixel[14]] < c_b)
+                     if( p[pixel[15]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+          else if( p[pixel[2]] < c_b)
+           if( p[pixel[9]] > cb)
+            if( p[pixel[10]] > cb)
+             if( p[pixel[11]] > cb)
+              if( p[pixel[12]] > cb)
+               if( p[pixel[13]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[3]] > cb)
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else if( p[pixel[9]] < c_b)
+            if( p[pixel[7]] < c_b)
+             if( p[pixel[8]] < c_b)
+              if( p[pixel[10]] < c_b)
+               if( p[pixel[6]] < c_b)
+                if( p[pixel[5]] < c_b)
+                 if( p[pixel[4]] < c_b)
+                  if( p[pixel[3]] < c_b)
+                   goto is_a_corner;
+                  else
+                   if( p[pixel[11]] < c_b)
+                    if( p[pixel[12]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                 else
+                  if( p[pixel[11]] < c_b)
+                   if( p[pixel[12]] < c_b)
+                    if( p[pixel[13]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[11]] < c_b)
+                  if( p[pixel[12]] < c_b)
+                   if( p[pixel[13]] < c_b)
+                    if( p[pixel[14]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[11]] < c_b)
+                 if( p[pixel[12]] < c_b)
+                  if( p[pixel[13]] < c_b)
+                   if( p[pixel[14]] < c_b)
+                    if( p[pixel[15]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else
+           if( p[pixel[9]] > cb)
+            if( p[pixel[10]] > cb)
+             if( p[pixel[11]] > cb)
+              if( p[pixel[12]] > cb)
+               if( p[pixel[13]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[3]] > cb)
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    if( p[pixel[8]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else if( p[pixel[9]] < c_b)
+            if( p[pixel[7]] < c_b)
+             if( p[pixel[8]] < c_b)
+              if( p[pixel[10]] < c_b)
+               if( p[pixel[11]] < c_b)
+                if( p[pixel[6]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[4]] < c_b)
+                   if( p[pixel[3]] < c_b)
+                    goto is_a_corner;
+                   else
+                    if( p[pixel[12]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                  else
+                   if( p[pixel[12]] < c_b)
+                    if( p[pixel[13]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                 else
+                  if( p[pixel[12]] < c_b)
+                   if( p[pixel[13]] < c_b)
+                    if( p[pixel[14]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[12]] < c_b)
+                  if( p[pixel[13]] < c_b)
+                   if( p[pixel[14]] < c_b)
+                    if( p[pixel[15]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+         else if( p[pixel[1]] < c_b)
+          if( p[pixel[8]] > cb)
+           if( p[pixel[9]] > cb)
+            if( p[pixel[10]] > cb)
+             if( p[pixel[11]] > cb)
+              if( p[pixel[12]] > cb)
+               if( p[pixel[13]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[3]] > cb)
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              if( p[pixel[2]] > cb)
+               if( p[pixel[3]] > cb)
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else if( p[pixel[8]] < c_b)
+           if( p[pixel[7]] < c_b)
+            if( p[pixel[9]] < c_b)
+             if( p[pixel[6]] < c_b)
+              if( p[pixel[5]] < c_b)
+               if( p[pixel[4]] < c_b)
+                if( p[pixel[3]] < c_b)
+                 if( p[pixel[2]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[10]] < c_b)
+                   if( p[pixel[11]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[10]] < c_b)
+                  if( p[pixel[11]] < c_b)
+                   if( p[pixel[12]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[10]] < c_b)
+                 if( p[pixel[11]] < c_b)
+                  if( p[pixel[12]] < c_b)
+                   if( p[pixel[13]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[10]] < c_b)
+                if( p[pixel[11]] < c_b)
+                 if( p[pixel[12]] < c_b)
+                  if( p[pixel[13]] < c_b)
+                   if( p[pixel[14]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              if( p[pixel[10]] < c_b)
+               if( p[pixel[11]] < c_b)
+                if( p[pixel[12]] < c_b)
+                 if( p[pixel[13]] < c_b)
+                  if( p[pixel[14]] < c_b)
+                   if( p[pixel[15]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else
+           goto is_not_a_corner;
+         else
+          if( p[pixel[8]] > cb)
+           if( p[pixel[9]] > cb)
+            if( p[pixel[10]] > cb)
+             if( p[pixel[11]] > cb)
+              if( p[pixel[12]] > cb)
+               if( p[pixel[13]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[15]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[3]] > cb)
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              if( p[pixel[2]] > cb)
+               if( p[pixel[3]] > cb)
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[7]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else if( p[pixel[8]] < c_b)
+           if( p[pixel[7]] < c_b)
+            if( p[pixel[9]] < c_b)
+             if( p[pixel[10]] < c_b)
+              if( p[pixel[6]] < c_b)
+               if( p[pixel[5]] < c_b)
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[3]] < c_b)
+                  if( p[pixel[2]] < c_b)
+                   goto is_a_corner;
+                  else
+                   if( p[pixel[11]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                 else
+                  if( p[pixel[11]] < c_b)
+                   if( p[pixel[12]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[11]] < c_b)
+                  if( p[pixel[12]] < c_b)
+                   if( p[pixel[13]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[11]] < c_b)
+                 if( p[pixel[12]] < c_b)
+                  if( p[pixel[13]] < c_b)
+                   if( p[pixel[14]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[11]] < c_b)
+                if( p[pixel[12]] < c_b)
+                 if( p[pixel[13]] < c_b)
+                  if( p[pixel[14]] < c_b)
+                   if( p[pixel[15]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else
+           goto is_not_a_corner;
+        else if( p[pixel[0]] < c_b)
+         if( p[pixel[1]] > cb)
+          if( p[pixel[8]] > cb)
+           if( p[pixel[7]] > cb)
+            if( p[pixel[9]] > cb)
+             if( p[pixel[6]] > cb)
+              if( p[pixel[5]] > cb)
+               if( p[pixel[4]] > cb)
+                if( p[pixel[3]] > cb)
+                 if( p[pixel[2]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[10]] > cb)
+                   if( p[pixel[11]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[10]] > cb)
+                  if( p[pixel[11]] > cb)
+                   if( p[pixel[12]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[10]] > cb)
+                 if( p[pixel[11]] > cb)
+                  if( p[pixel[12]] > cb)
+                   if( p[pixel[13]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[10]] > cb)
+                if( p[pixel[11]] > cb)
+                 if( p[pixel[12]] > cb)
+                  if( p[pixel[13]] > cb)
+                   if( p[pixel[14]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              if( p[pixel[10]] > cb)
+               if( p[pixel[11]] > cb)
+                if( p[pixel[12]] > cb)
+                 if( p[pixel[13]] > cb)
+                  if( p[pixel[14]] > cb)
+                   if( p[pixel[15]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else if( p[pixel[8]] < c_b)
+           if( p[pixel[9]] < c_b)
+            if( p[pixel[10]] < c_b)
+             if( p[pixel[11]] < c_b)
+              if( p[pixel[12]] < c_b)
+               if( p[pixel[13]] < c_b)
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[3]] < c_b)
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              if( p[pixel[2]] < c_b)
+               if( p[pixel[3]] < c_b)
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else
+           goto is_not_a_corner;
+         else if( p[pixel[1]] < c_b)
+          if( p[pixel[2]] > cb)
+           if( p[pixel[9]] > cb)
+            if( p[pixel[7]] > cb)
+             if( p[pixel[8]] > cb)
+              if( p[pixel[10]] > cb)
+               if( p[pixel[6]] > cb)
+                if( p[pixel[5]] > cb)
+                 if( p[pixel[4]] > cb)
+                  if( p[pixel[3]] > cb)
+                   goto is_a_corner;
+                  else
+                   if( p[pixel[11]] > cb)
+                    if( p[pixel[12]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                 else
+                  if( p[pixel[11]] > cb)
+                   if( p[pixel[12]] > cb)
+                    if( p[pixel[13]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[11]] > cb)
+                  if( p[pixel[12]] > cb)
+                   if( p[pixel[13]] > cb)
+                    if( p[pixel[14]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[11]] > cb)
+                 if( p[pixel[12]] > cb)
+                  if( p[pixel[13]] > cb)
+                   if( p[pixel[14]] > cb)
+                    if( p[pixel[15]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else if( p[pixel[9]] < c_b)
+            if( p[pixel[10]] < c_b)
+             if( p[pixel[11]] < c_b)
+              if( p[pixel[12]] < c_b)
+               if( p[pixel[13]] < c_b)
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[3]] < c_b)
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else if( p[pixel[2]] < c_b)
+           if( p[pixel[3]] > cb)
+            if( p[pixel[10]] > cb)
+             if( p[pixel[7]] > cb)
+              if( p[pixel[8]] > cb)
+               if( p[pixel[9]] > cb)
+                if( p[pixel[11]] > cb)
+                 if( p[pixel[6]] > cb)
+                  if( p[pixel[5]] > cb)
+                   if( p[pixel[4]] > cb)
+                    goto is_a_corner;
+                   else
+                    if( p[pixel[12]] > cb)
+                     if( p[pixel[13]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                  else
+                   if( p[pixel[12]] > cb)
+                    if( p[pixel[13]] > cb)
+                     if( p[pixel[14]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                 else
+                  if( p[pixel[12]] > cb)
+                   if( p[pixel[13]] > cb)
+                    if( p[pixel[14]] > cb)
+                     if( p[pixel[15]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else if( p[pixel[10]] < c_b)
+             if( p[pixel[11]] < c_b)
+              if( p[pixel[12]] < c_b)
+               if( p[pixel[13]] < c_b)
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else if( p[pixel[3]] < c_b)
+            if( p[pixel[4]] > cb)
+             if( p[pixel[13]] > cb)
+              if( p[pixel[7]] > cb)
+               if( p[pixel[8]] > cb)
+                if( p[pixel[9]] > cb)
+                 if( p[pixel[10]] > cb)
+                  if( p[pixel[11]] > cb)
+                   if( p[pixel[12]] > cb)
+                    if( p[pixel[6]] > cb)
+                     if( p[pixel[5]] > cb)
+                      goto is_a_corner;
+                     else
+                      if( p[pixel[14]] > cb)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                    else
+                     if( p[pixel[14]] > cb)
+                      if( p[pixel[15]] > cb)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else if( p[pixel[13]] < c_b)
+              if( p[pixel[11]] > cb)
+               if( p[pixel[5]] > cb)
+                if( p[pixel[6]] > cb)
+                 if( p[pixel[7]] > cb)
+                  if( p[pixel[8]] > cb)
+                   if( p[pixel[9]] > cb)
+                    if( p[pixel[10]] > cb)
+                     if( p[pixel[12]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else if( p[pixel[11]] < c_b)
+               if( p[pixel[12]] < c_b)
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      if( p[pixel[10]] < c_b)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      if( p[pixel[10]] < c_b)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              if( p[pixel[5]] > cb)
+               if( p[pixel[6]] > cb)
+                if( p[pixel[7]] > cb)
+                 if( p[pixel[8]] > cb)
+                  if( p[pixel[9]] > cb)
+                   if( p[pixel[10]] > cb)
+                    if( p[pixel[11]] > cb)
+                     if( p[pixel[12]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else if( p[pixel[4]] < c_b)
+             if( p[pixel[5]] > cb)
+              if( p[pixel[14]] > cb)
+               if( p[pixel[7]] > cb)
+                if( p[pixel[8]] > cb)
+                 if( p[pixel[9]] > cb)
+                  if( p[pixel[10]] > cb)
+                   if( p[pixel[11]] > cb)
+                    if( p[pixel[12]] > cb)
+                     if( p[pixel[13]] > cb)
+                      if( p[pixel[6]] > cb)
+                       goto is_a_corner;
+                      else
+                       if( p[pixel[15]] > cb)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else if( p[pixel[14]] < c_b)
+               if( p[pixel[12]] > cb)
+                if( p[pixel[6]] > cb)
+                 if( p[pixel[7]] > cb)
+                  if( p[pixel[8]] > cb)
+                   if( p[pixel[9]] > cb)
+                    if( p[pixel[10]] > cb)
+                     if( p[pixel[11]] > cb)
+                      if( p[pixel[13]] > cb)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else if( p[pixel[12]] < c_b)
+                if( p[pixel[13]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      if( p[pixel[10]] < c_b)
+                       if( p[pixel[11]] < c_b)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               if( p[pixel[6]] > cb)
+                if( p[pixel[7]] > cb)
+                 if( p[pixel[8]] > cb)
+                  if( p[pixel[9]] > cb)
+                   if( p[pixel[10]] > cb)
+                    if( p[pixel[11]] > cb)
+                     if( p[pixel[12]] > cb)
+                      if( p[pixel[13]] > cb)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else if( p[pixel[5]] < c_b)
+              if( p[pixel[6]] > cb)
+               if( p[pixel[15]] < c_b)
+                if( p[pixel[13]] > cb)
+                 if( p[pixel[7]] > cb)
+                  if( p[pixel[8]] > cb)
+                   if( p[pixel[9]] > cb)
+                    if( p[pixel[10]] > cb)
+                     if( p[pixel[11]] > cb)
+                      if( p[pixel[12]] > cb)
+                       if( p[pixel[14]] > cb)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else if( p[pixel[13]] < c_b)
+                 if( p[pixel[14]] < c_b)
+                  goto is_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                if( p[pixel[7]] > cb)
+                 if( p[pixel[8]] > cb)
+                  if( p[pixel[9]] > cb)
+                   if( p[pixel[10]] > cb)
+                    if( p[pixel[11]] > cb)
+                     if( p[pixel[12]] > cb)
+                      if( p[pixel[13]] > cb)
+                       if( p[pixel[14]] > cb)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else if( p[pixel[6]] < c_b)
+               if( p[pixel[7]] > cb)
+                if( p[pixel[14]] > cb)
+                 if( p[pixel[8]] > cb)
+                  if( p[pixel[9]] > cb)
+                   if( p[pixel[10]] > cb)
+                    if( p[pixel[11]] > cb)
+                     if( p[pixel[12]] > cb)
+                      if( p[pixel[13]] > cb)
+                       if( p[pixel[15]] > cb)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else if( p[pixel[7]] < c_b)
+                if( p[pixel[8]] < c_b)
+                 goto is_a_corner;
+                else
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[13]] > cb)
+                if( p[pixel[7]] > cb)
+                 if( p[pixel[8]] > cb)
+                  if( p[pixel[9]] > cb)
+                   if( p[pixel[10]] > cb)
+                    if( p[pixel[11]] > cb)
+                     if( p[pixel[12]] > cb)
+                      if( p[pixel[14]] > cb)
+                       if( p[pixel[15]] > cb)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else if( p[pixel[13]] < c_b)
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              if( p[pixel[12]] > cb)
+               if( p[pixel[7]] > cb)
+                if( p[pixel[8]] > cb)
+                 if( p[pixel[9]] > cb)
+                  if( p[pixel[10]] > cb)
+                   if( p[pixel[11]] > cb)
+                    if( p[pixel[13]] > cb)
+                     if( p[pixel[14]] > cb)
+                      if( p[pixel[6]] > cb)
+                       goto is_a_corner;
+                      else
+                       if( p[pixel[15]] > cb)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else if( p[pixel[12]] < c_b)
+               if( p[pixel[13]] < c_b)
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      if( p[pixel[10]] < c_b)
+                       if( p[pixel[11]] < c_b)
+                        goto is_a_corner;
+                       else
+                        goto is_not_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else
+             if( p[pixel[11]] > cb)
+              if( p[pixel[7]] > cb)
+               if( p[pixel[8]] > cb)
+                if( p[pixel[9]] > cb)
+                 if( p[pixel[10]] > cb)
+                  if( p[pixel[12]] > cb)
+                   if( p[pixel[13]] > cb)
+                    if( p[pixel[6]] > cb)
+                     if( p[pixel[5]] > cb)
+                      goto is_a_corner;
+                     else
+                      if( p[pixel[14]] > cb)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                    else
+                     if( p[pixel[14]] > cb)
+                      if( p[pixel[15]] > cb)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else if( p[pixel[11]] < c_b)
+              if( p[pixel[12]] < c_b)
+               if( p[pixel[13]] < c_b)
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      if( p[pixel[10]] < c_b)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      if( p[pixel[10]] < c_b)
+                       goto is_a_corner;
+                      else
+                       goto is_not_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+           else
+            if( p[pixel[10]] > cb)
+             if( p[pixel[7]] > cb)
+              if( p[pixel[8]] > cb)
+               if( p[pixel[9]] > cb)
+                if( p[pixel[11]] > cb)
+                 if( p[pixel[12]] > cb)
+                  if( p[pixel[6]] > cb)
+                   if( p[pixel[5]] > cb)
+                    if( p[pixel[4]] > cb)
+                     goto is_a_corner;
+                    else
+                     if( p[pixel[13]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                   else
+                    if( p[pixel[13]] > cb)
+                     if( p[pixel[14]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                  else
+                   if( p[pixel[13]] > cb)
+                    if( p[pixel[14]] > cb)
+                     if( p[pixel[15]] > cb)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else if( p[pixel[10]] < c_b)
+             if( p[pixel[11]] < c_b)
+              if( p[pixel[12]] < c_b)
+               if( p[pixel[13]] < c_b)
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     if( p[pixel[9]] < c_b)
+                      goto is_a_corner;
+                     else
+                      goto is_not_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+          else
+           if( p[pixel[9]] > cb)
+            if( p[pixel[7]] > cb)
+             if( p[pixel[8]] > cb)
+              if( p[pixel[10]] > cb)
+               if( p[pixel[11]] > cb)
+                if( p[pixel[6]] > cb)
+                 if( p[pixel[5]] > cb)
+                  if( p[pixel[4]] > cb)
+                   if( p[pixel[3]] > cb)
+                    goto is_a_corner;
+                   else
+                    if( p[pixel[12]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                  else
+                   if( p[pixel[12]] > cb)
+                    if( p[pixel[13]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                 else
+                  if( p[pixel[12]] > cb)
+                   if( p[pixel[13]] > cb)
+                    if( p[pixel[14]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[12]] > cb)
+                  if( p[pixel[13]] > cb)
+                   if( p[pixel[14]] > cb)
+                    if( p[pixel[15]] > cb)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else if( p[pixel[9]] < c_b)
+            if( p[pixel[10]] < c_b)
+             if( p[pixel[11]] < c_b)
+              if( p[pixel[12]] < c_b)
+               if( p[pixel[13]] < c_b)
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[3]] < c_b)
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    if( p[pixel[8]] < c_b)
+                     goto is_a_corner;
+                    else
+                     goto is_not_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+         else
+          if( p[pixel[8]] > cb)
+           if( p[pixel[7]] > cb)
+            if( p[pixel[9]] > cb)
+             if( p[pixel[10]] > cb)
+              if( p[pixel[6]] > cb)
+               if( p[pixel[5]] > cb)
+                if( p[pixel[4]] > cb)
+                 if( p[pixel[3]] > cb)
+                  if( p[pixel[2]] > cb)
+                   goto is_a_corner;
+                  else
+                   if( p[pixel[11]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                 else
+                  if( p[pixel[11]] > cb)
+                   if( p[pixel[12]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[11]] > cb)
+                  if( p[pixel[12]] > cb)
+                   if( p[pixel[13]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[11]] > cb)
+                 if( p[pixel[12]] > cb)
+                  if( p[pixel[13]] > cb)
+                   if( p[pixel[14]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[11]] > cb)
+                if( p[pixel[12]] > cb)
+                 if( p[pixel[13]] > cb)
+                  if( p[pixel[14]] > cb)
+                   if( p[pixel[15]] > cb)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else if( p[pixel[8]] < c_b)
+           if( p[pixel[9]] < c_b)
+            if( p[pixel[10]] < c_b)
+             if( p[pixel[11]] < c_b)
+              if( p[pixel[12]] < c_b)
+               if( p[pixel[13]] < c_b)
+                if( p[pixel[14]] < c_b)
+                 if( p[pixel[15]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[3]] < c_b)
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              if( p[pixel[2]] < c_b)
+               if( p[pixel[3]] < c_b)
+                if( p[pixel[4]] < c_b)
+                 if( p[pixel[5]] < c_b)
+                  if( p[pixel[6]] < c_b)
+                   if( p[pixel[7]] < c_b)
+                    goto is_a_corner;
+                   else
+                    goto is_not_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else
+             goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else
+           goto is_not_a_corner;
+        else
+         if( p[pixel[7]] > cb)
+          if( p[pixel[8]] > cb)
+           if( p[pixel[9]] > cb)
+            if( p[pixel[6]] > cb)
+             if( p[pixel[5]] > cb)
+              if( p[pixel[4]] > cb)
+               if( p[pixel[3]] > cb)
+                if( p[pixel[2]] > cb)
+                 if( p[pixel[1]] > cb)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[10]] > cb)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[10]] > cb)
+                  if( p[pixel[11]] > cb)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[10]] > cb)
+                 if( p[pixel[11]] > cb)
+                  if( p[pixel[12]] > cb)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[10]] > cb)
+                if( p[pixel[11]] > cb)
+                 if( p[pixel[12]] > cb)
+                  if( p[pixel[13]] > cb)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              if( p[pixel[10]] > cb)
+               if( p[pixel[11]] > cb)
+                if( p[pixel[12]] > cb)
+                 if( p[pixel[13]] > cb)
+                  if( p[pixel[14]] > cb)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else
+             if( p[pixel[10]] > cb)
+              if( p[pixel[11]] > cb)
+               if( p[pixel[12]] > cb)
+                if( p[pixel[13]] > cb)
+                 if( p[pixel[14]] > cb)
+                  if( p[pixel[15]] > cb)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else
+           goto is_not_a_corner;
+         else if( p[pixel[7]] < c_b)
+          if( p[pixel[8]] < c_b)
+           if( p[pixel[9]] < c_b)
+            if( p[pixel[6]] < c_b)
+             if( p[pixel[5]] < c_b)
+              if( p[pixel[4]] < c_b)
+               if( p[pixel[3]] < c_b)
+                if( p[pixel[2]] < c_b)
+                 if( p[pixel[1]] < c_b)
+                  goto is_a_corner;
+                 else
+                  if( p[pixel[10]] < c_b)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                else
+                 if( p[pixel[10]] < c_b)
+                  if( p[pixel[11]] < c_b)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+               else
+                if( p[pixel[10]] < c_b)
+                 if( p[pixel[11]] < c_b)
+                  if( p[pixel[12]] < c_b)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+              else
+               if( p[pixel[10]] < c_b)
+                if( p[pixel[11]] < c_b)
+                 if( p[pixel[12]] < c_b)
+                  if( p[pixel[13]] < c_b)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+             else
+              if( p[pixel[10]] < c_b)
+               if( p[pixel[11]] < c_b)
+                if( p[pixel[12]] < c_b)
+                 if( p[pixel[13]] < c_b)
+                  if( p[pixel[14]] < c_b)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+            else
+             if( p[pixel[10]] < c_b)
+              if( p[pixel[11]] < c_b)
+               if( p[pixel[12]] < c_b)
+                if( p[pixel[13]] < c_b)
+                 if( p[pixel[14]] < c_b)
+                  if( p[pixel[15]] < c_b)
+                   goto is_a_corner;
+                  else
+                   goto is_not_a_corner;
+                 else
+                  goto is_not_a_corner;
+                else
+                 goto is_not_a_corner;
+               else
+                goto is_not_a_corner;
+              else
+               goto is_not_a_corner;
+             else
+              goto is_not_a_corner;
+           else
+            goto is_not_a_corner;
+          else
+           goto is_not_a_corner;
+         else
+          goto is_not_a_corner;
+
+		is_a_corner:
+			bmin=b;
+			goto end_if;
+
+		is_not_a_corner:
+			bmax=b;
+			goto end_if;
+
+		end_if:
+
+		if(bmin == bmax - 1 || bmin == bmax)
+			return bmin;
+		b = (bmin + bmax) / 2;
+    }
+}
+
+static void make_offsets(int pixel[], int row_stride)
+{
+        pixel[0] = 0 + row_stride * 3;
+        pixel[1] = 1 + row_stride * 3;
+        pixel[2] = 2 + row_stride * 2;
+        pixel[3] = 3 + row_stride * 1;
+        pixel[4] = 3 + row_stride * 0;
+        pixel[5] = 3 + row_stride * -1;
+        pixel[6] = 2 + row_stride * -2;
+        pixel[7] = 1 + row_stride * -3;
+        pixel[8] = 0 + row_stride * -3;
+        pixel[9] = -1 + row_stride * -3;
+        pixel[10] = -2 + row_stride * -2;
+        pixel[11] = -3 + row_stride * -1;
+        pixel[12] = -3 + row_stride * 0;
+        pixel[13] = -3 + row_stride * 1;
+        pixel[14] = -2 + row_stride * 2;
+        pixel[15] = -1 + row_stride * 3;
+}
+
+
+
+int* fast9_score(const byte* i, int stride, xy* corners, int num_corners, int b)
+{
+	int* scores = (int*)malloc(sizeof(int)* num_corners);
+	int n;
+
+	int pixel[16];
+	make_offsets(pixel, stride);
+
+    for(n=0; n < num_corners; n++)
+        scores[n] = fast9_corner_score(i + corners[n].y*stride + corners[n].x, pixel, b);
+
+	return scores;
+}
+
+
+xy* fast9_detect(const byte* im, int xsize, int ysize, int stride, int b, int* ret_num_corners)
+{
+	int num_corners=0;
+	xy* ret_corners;
+	int rsize=512;
+	int pixel[16];
+	int x, y;
+
+	ret_corners = (xy*)malloc(sizeof(xy)*rsize);
+	make_offsets(pixel, stride);
+
+	for(y=3; y < ysize - 3; y++)
+		for(x=3; x < xsize - 3; x++)
+		{
+			const byte* p = im + y*stride + x;
+
+			int cb = *p + b;
+			int c_b= *p - b;
+        if(p[pixel[0]] > cb)
+         if(p[pixel[1]] > cb)
+          if(p[pixel[2]] > cb)
+           if(p[pixel[3]] > cb)
+            if(p[pixel[4]] > cb)
+             if(p[pixel[5]] > cb)
+              if(p[pixel[6]] > cb)
+               if(p[pixel[7]] > cb)
+                if(p[pixel[8]] > cb)
+                 {}
+                else
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  continue;
+               else if(p[pixel[7]] < c_b)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  continue;
+                else if(p[pixel[14]] < c_b)
+                 if(p[pixel[8]] < c_b)
+                  if(p[pixel[9]] < c_b)
+                   if(p[pixel[10]] < c_b)
+                    if(p[pixel[11]] < c_b)
+                     if(p[pixel[12]] < c_b)
+                      if(p[pixel[13]] < c_b)
+                       if(p[pixel[15]] < c_b)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  continue;
+                else
+                 continue;
+              else if(p[pixel[6]] < c_b)
+               if(p[pixel[15]] > cb)
+                if(p[pixel[13]] > cb)
+                 if(p[pixel[14]] > cb)
+                  {}
+                 else
+                  continue;
+                else if(p[pixel[13]] < c_b)
+                 if(p[pixel[7]] < c_b)
+                  if(p[pixel[8]] < c_b)
+                   if(p[pixel[9]] < c_b)
+                    if(p[pixel[10]] < c_b)
+                     if(p[pixel[11]] < c_b)
+                      if(p[pixel[12]] < c_b)
+                       if(p[pixel[14]] < c_b)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                if(p[pixel[7]] < c_b)
+                 if(p[pixel[8]] < c_b)
+                  if(p[pixel[9]] < c_b)
+                   if(p[pixel[10]] < c_b)
+                    if(p[pixel[11]] < c_b)
+                     if(p[pixel[12]] < c_b)
+                      if(p[pixel[13]] < c_b)
+                       if(p[pixel[14]] < c_b)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[13]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  continue;
+                else
+                 continue;
+               else if(p[pixel[13]] < c_b)
+                if(p[pixel[7]] < c_b)
+                 if(p[pixel[8]] < c_b)
+                  if(p[pixel[9]] < c_b)
+                   if(p[pixel[10]] < c_b)
+                    if(p[pixel[11]] < c_b)
+                     if(p[pixel[12]] < c_b)
+                      if(p[pixel[14]] < c_b)
+                       if(p[pixel[15]] < c_b)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else if(p[pixel[5]] < c_b)
+              if(p[pixel[14]] > cb)
+               if(p[pixel[12]] > cb)
+                if(p[pixel[13]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      if(p[pixel[10]] > cb)
+                       if(p[pixel[11]] > cb)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 continue;
+               else if(p[pixel[12]] < c_b)
+                if(p[pixel[6]] < c_b)
+                 if(p[pixel[7]] < c_b)
+                  if(p[pixel[8]] < c_b)
+                   if(p[pixel[9]] < c_b)
+                    if(p[pixel[10]] < c_b)
+                     if(p[pixel[11]] < c_b)
+                      if(p[pixel[13]] < c_b)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else if(p[pixel[14]] < c_b)
+               if(p[pixel[7]] < c_b)
+                if(p[pixel[8]] < c_b)
+                 if(p[pixel[9]] < c_b)
+                  if(p[pixel[10]] < c_b)
+                   if(p[pixel[11]] < c_b)
+                    if(p[pixel[12]] < c_b)
+                     if(p[pixel[13]] < c_b)
+                      if(p[pixel[6]] < c_b)
+                       {}
+                      else
+                       if(p[pixel[15]] < c_b)
+                        {}
+                       else
+                        continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               if(p[pixel[6]] < c_b)
+                if(p[pixel[7]] < c_b)
+                 if(p[pixel[8]] < c_b)
+                  if(p[pixel[9]] < c_b)
+                   if(p[pixel[10]] < c_b)
+                    if(p[pixel[11]] < c_b)
+                     if(p[pixel[12]] < c_b)
+                      if(p[pixel[13]] < c_b)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              if(p[pixel[12]] > cb)
+               if(p[pixel[13]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      if(p[pixel[10]] > cb)
+                       if(p[pixel[11]] > cb)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 continue;
+               else
+                continue;
+              else if(p[pixel[12]] < c_b)
+               if(p[pixel[7]] < c_b)
+                if(p[pixel[8]] < c_b)
+                 if(p[pixel[9]] < c_b)
+                  if(p[pixel[10]] < c_b)
+                   if(p[pixel[11]] < c_b)
+                    if(p[pixel[13]] < c_b)
+                     if(p[pixel[14]] < c_b)
+                      if(p[pixel[6]] < c_b)
+                       {}
+                      else
+                       if(p[pixel[15]] < c_b)
+                        {}
+                       else
+                        continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else if(p[pixel[4]] < c_b)
+             if(p[pixel[13]] > cb)
+              if(p[pixel[11]] > cb)
+               if(p[pixel[12]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      if(p[pixel[10]] > cb)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      if(p[pixel[10]] > cb)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                continue;
+              else if(p[pixel[11]] < c_b)
+               if(p[pixel[5]] < c_b)
+                if(p[pixel[6]] < c_b)
+                 if(p[pixel[7]] < c_b)
+                  if(p[pixel[8]] < c_b)
+                   if(p[pixel[9]] < c_b)
+                    if(p[pixel[10]] < c_b)
+                     if(p[pixel[12]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else if(p[pixel[13]] < c_b)
+              if(p[pixel[7]] < c_b)
+               if(p[pixel[8]] < c_b)
+                if(p[pixel[9]] < c_b)
+                 if(p[pixel[10]] < c_b)
+                  if(p[pixel[11]] < c_b)
+                   if(p[pixel[12]] < c_b)
+                    if(p[pixel[6]] < c_b)
+                     if(p[pixel[5]] < c_b)
+                      {}
+                     else
+                      if(p[pixel[14]] < c_b)
+                       {}
+                      else
+                       continue;
+                    else
+                     if(p[pixel[14]] < c_b)
+                      if(p[pixel[15]] < c_b)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              if(p[pixel[5]] < c_b)
+               if(p[pixel[6]] < c_b)
+                if(p[pixel[7]] < c_b)
+                 if(p[pixel[8]] < c_b)
+                  if(p[pixel[9]] < c_b)
+                   if(p[pixel[10]] < c_b)
+                    if(p[pixel[11]] < c_b)
+                     if(p[pixel[12]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else
+             if(p[pixel[11]] > cb)
+              if(p[pixel[12]] > cb)
+               if(p[pixel[13]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      if(p[pixel[10]] > cb)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      if(p[pixel[10]] > cb)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                continue;
+              else
+               continue;
+             else if(p[pixel[11]] < c_b)
+              if(p[pixel[7]] < c_b)
+               if(p[pixel[8]] < c_b)
+                if(p[pixel[9]] < c_b)
+                 if(p[pixel[10]] < c_b)
+                  if(p[pixel[12]] < c_b)
+                   if(p[pixel[13]] < c_b)
+                    if(p[pixel[6]] < c_b)
+                     if(p[pixel[5]] < c_b)
+                      {}
+                     else
+                      if(p[pixel[14]] < c_b)
+                       {}
+                      else
+                       continue;
+                    else
+                     if(p[pixel[14]] < c_b)
+                      if(p[pixel[15]] < c_b)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              continue;
+           else if(p[pixel[3]] < c_b)
+            if(p[pixel[10]] > cb)
+             if(p[pixel[11]] > cb)
+              if(p[pixel[12]] > cb)
+               if(p[pixel[13]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               continue;
+             else
+              continue;
+            else if(p[pixel[10]] < c_b)
+             if(p[pixel[7]] < c_b)
+              if(p[pixel[8]] < c_b)
+               if(p[pixel[9]] < c_b)
+                if(p[pixel[11]] < c_b)
+                 if(p[pixel[6]] < c_b)
+                  if(p[pixel[5]] < c_b)
+                   if(p[pixel[4]] < c_b)
+                    {}
+                   else
+                    if(p[pixel[12]] < c_b)
+                     if(p[pixel[13]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                  else
+                   if(p[pixel[12]] < c_b)
+                    if(p[pixel[13]] < c_b)
+                     if(p[pixel[14]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                 else
+                  if(p[pixel[12]] < c_b)
+                   if(p[pixel[13]] < c_b)
+                    if(p[pixel[14]] < c_b)
+                     if(p[pixel[15]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              continue;
+            else
+             continue;
+           else
+            if(p[pixel[10]] > cb)
+             if(p[pixel[11]] > cb)
+              if(p[pixel[12]] > cb)
+               if(p[pixel[13]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     if(p[pixel[9]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               continue;
+             else
+              continue;
+            else if(p[pixel[10]] < c_b)
+             if(p[pixel[7]] < c_b)
+              if(p[pixel[8]] < c_b)
+               if(p[pixel[9]] < c_b)
+                if(p[pixel[11]] < c_b)
+                 if(p[pixel[12]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[5]] < c_b)
+                    if(p[pixel[4]] < c_b)
+                     {}
+                    else
+                     if(p[pixel[13]] < c_b)
+                      {}
+                     else
+                      continue;
+                   else
+                    if(p[pixel[13]] < c_b)
+                     if(p[pixel[14]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                  else
+                   if(p[pixel[13]] < c_b)
+                    if(p[pixel[14]] < c_b)
+                     if(p[pixel[15]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              continue;
+            else
+             continue;
+          else if(p[pixel[2]] < c_b)
+           if(p[pixel[9]] > cb)
+            if(p[pixel[10]] > cb)
+             if(p[pixel[11]] > cb)
+              if(p[pixel[12]] > cb)
+               if(p[pixel[13]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[3]] > cb)
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              continue;
+            else
+             continue;
+           else if(p[pixel[9]] < c_b)
+            if(p[pixel[7]] < c_b)
+             if(p[pixel[8]] < c_b)
+              if(p[pixel[10]] < c_b)
+               if(p[pixel[6]] < c_b)
+                if(p[pixel[5]] < c_b)
+                 if(p[pixel[4]] < c_b)
+                  if(p[pixel[3]] < c_b)
+                   {}
+                  else
+                   if(p[pixel[11]] < c_b)
+                    if(p[pixel[12]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                 else
+                  if(p[pixel[11]] < c_b)
+                   if(p[pixel[12]] < c_b)
+                    if(p[pixel[13]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[11]] < c_b)
+                  if(p[pixel[12]] < c_b)
+                   if(p[pixel[13]] < c_b)
+                    if(p[pixel[14]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[11]] < c_b)
+                 if(p[pixel[12]] < c_b)
+                  if(p[pixel[13]] < c_b)
+                   if(p[pixel[14]] < c_b)
+                    if(p[pixel[15]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               continue;
+             else
+              continue;
+            else
+             continue;
+           else
+            continue;
+          else
+           if(p[pixel[9]] > cb)
+            if(p[pixel[10]] > cb)
+             if(p[pixel[11]] > cb)
+              if(p[pixel[12]] > cb)
+               if(p[pixel[13]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[3]] > cb)
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    if(p[pixel[8]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              continue;
+            else
+             continue;
+           else if(p[pixel[9]] < c_b)
+            if(p[pixel[7]] < c_b)
+             if(p[pixel[8]] < c_b)
+              if(p[pixel[10]] < c_b)
+               if(p[pixel[11]] < c_b)
+                if(p[pixel[6]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[4]] < c_b)
+                   if(p[pixel[3]] < c_b)
+                    {}
+                   else
+                    if(p[pixel[12]] < c_b)
+                     {}
+                    else
+                     continue;
+                  else
+                   if(p[pixel[12]] < c_b)
+                    if(p[pixel[13]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                 else
+                  if(p[pixel[12]] < c_b)
+                   if(p[pixel[13]] < c_b)
+                    if(p[pixel[14]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[12]] < c_b)
+                  if(p[pixel[13]] < c_b)
+                   if(p[pixel[14]] < c_b)
+                    if(p[pixel[15]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              continue;
+            else
+             continue;
+           else
+            continue;
+         else if(p[pixel[1]] < c_b)
+          if(p[pixel[8]] > cb)
+           if(p[pixel[9]] > cb)
+            if(p[pixel[10]] > cb)
+             if(p[pixel[11]] > cb)
+              if(p[pixel[12]] > cb)
+               if(p[pixel[13]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[3]] > cb)
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              if(p[pixel[2]] > cb)
+               if(p[pixel[3]] > cb)
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else
+             continue;
+           else
+            continue;
+          else if(p[pixel[8]] < c_b)
+           if(p[pixel[7]] < c_b)
+            if(p[pixel[9]] < c_b)
+             if(p[pixel[6]] < c_b)
+              if(p[pixel[5]] < c_b)
+               if(p[pixel[4]] < c_b)
+                if(p[pixel[3]] < c_b)
+                 if(p[pixel[2]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[10]] < c_b)
+                   if(p[pixel[11]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[10]] < c_b)
+                  if(p[pixel[11]] < c_b)
+                   if(p[pixel[12]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[10]] < c_b)
+                 if(p[pixel[11]] < c_b)
+                  if(p[pixel[12]] < c_b)
+                   if(p[pixel[13]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[10]] < c_b)
+                if(p[pixel[11]] < c_b)
+                 if(p[pixel[12]] < c_b)
+                  if(p[pixel[13]] < c_b)
+                   if(p[pixel[14]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              if(p[pixel[10]] < c_b)
+               if(p[pixel[11]] < c_b)
+                if(p[pixel[12]] < c_b)
+                 if(p[pixel[13]] < c_b)
+                  if(p[pixel[14]] < c_b)
+                   if(p[pixel[15]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else
+             continue;
+           else
+            continue;
+          else
+           continue;
+         else
+          if(p[pixel[8]] > cb)
+           if(p[pixel[9]] > cb)
+            if(p[pixel[10]] > cb)
+             if(p[pixel[11]] > cb)
+              if(p[pixel[12]] > cb)
+               if(p[pixel[13]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[15]] > cb)
+                  {}
+                 else
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[3]] > cb)
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              if(p[pixel[2]] > cb)
+               if(p[pixel[3]] > cb)
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[7]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else
+             continue;
+           else
+            continue;
+          else if(p[pixel[8]] < c_b)
+           if(p[pixel[7]] < c_b)
+            if(p[pixel[9]] < c_b)
+             if(p[pixel[10]] < c_b)
+              if(p[pixel[6]] < c_b)
+               if(p[pixel[5]] < c_b)
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[3]] < c_b)
+                  if(p[pixel[2]] < c_b)
+                   {}
+                  else
+                   if(p[pixel[11]] < c_b)
+                    {}
+                   else
+                    continue;
+                 else
+                  if(p[pixel[11]] < c_b)
+                   if(p[pixel[12]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[11]] < c_b)
+                  if(p[pixel[12]] < c_b)
+                   if(p[pixel[13]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[11]] < c_b)
+                 if(p[pixel[12]] < c_b)
+                  if(p[pixel[13]] < c_b)
+                   if(p[pixel[14]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[11]] < c_b)
+                if(p[pixel[12]] < c_b)
+                 if(p[pixel[13]] < c_b)
+                  if(p[pixel[14]] < c_b)
+                   if(p[pixel[15]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              continue;
+            else
+             continue;
+           else
+            continue;
+          else
+           continue;
+        else if(p[pixel[0]] < c_b)
+         if(p[pixel[1]] > cb)
+          if(p[pixel[8]] > cb)
+           if(p[pixel[7]] > cb)
+            if(p[pixel[9]] > cb)
+             if(p[pixel[6]] > cb)
+              if(p[pixel[5]] > cb)
+               if(p[pixel[4]] > cb)
+                if(p[pixel[3]] > cb)
+                 if(p[pixel[2]] > cb)
+                  {}
+                 else
+                  if(p[pixel[10]] > cb)
+                   if(p[pixel[11]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[10]] > cb)
+                  if(p[pixel[11]] > cb)
+                   if(p[pixel[12]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[10]] > cb)
+                 if(p[pixel[11]] > cb)
+                  if(p[pixel[12]] > cb)
+                   if(p[pixel[13]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[10]] > cb)
+                if(p[pixel[11]] > cb)
+                 if(p[pixel[12]] > cb)
+                  if(p[pixel[13]] > cb)
+                   if(p[pixel[14]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              if(p[pixel[10]] > cb)
+               if(p[pixel[11]] > cb)
+                if(p[pixel[12]] > cb)
+                 if(p[pixel[13]] > cb)
+                  if(p[pixel[14]] > cb)
+                   if(p[pixel[15]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else
+             continue;
+           else
+            continue;
+          else if(p[pixel[8]] < c_b)
+           if(p[pixel[9]] < c_b)
+            if(p[pixel[10]] < c_b)
+             if(p[pixel[11]] < c_b)
+              if(p[pixel[12]] < c_b)
+               if(p[pixel[13]] < c_b)
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[3]] < c_b)
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              if(p[pixel[2]] < c_b)
+               if(p[pixel[3]] < c_b)
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else
+             continue;
+           else
+            continue;
+          else
+           continue;
+         else if(p[pixel[1]] < c_b)
+          if(p[pixel[2]] > cb)
+           if(p[pixel[9]] > cb)
+            if(p[pixel[7]] > cb)
+             if(p[pixel[8]] > cb)
+              if(p[pixel[10]] > cb)
+               if(p[pixel[6]] > cb)
+                if(p[pixel[5]] > cb)
+                 if(p[pixel[4]] > cb)
+                  if(p[pixel[3]] > cb)
+                   {}
+                  else
+                   if(p[pixel[11]] > cb)
+                    if(p[pixel[12]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                 else
+                  if(p[pixel[11]] > cb)
+                   if(p[pixel[12]] > cb)
+                    if(p[pixel[13]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[11]] > cb)
+                  if(p[pixel[12]] > cb)
+                   if(p[pixel[13]] > cb)
+                    if(p[pixel[14]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[11]] > cb)
+                 if(p[pixel[12]] > cb)
+                  if(p[pixel[13]] > cb)
+                   if(p[pixel[14]] > cb)
+                    if(p[pixel[15]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               continue;
+             else
+              continue;
+            else
+             continue;
+           else if(p[pixel[9]] < c_b)
+            if(p[pixel[10]] < c_b)
+             if(p[pixel[11]] < c_b)
+              if(p[pixel[12]] < c_b)
+               if(p[pixel[13]] < c_b)
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[3]] < c_b)
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              continue;
+            else
+             continue;
+           else
+            continue;
+          else if(p[pixel[2]] < c_b)
+           if(p[pixel[3]] > cb)
+            if(p[pixel[10]] > cb)
+             if(p[pixel[7]] > cb)
+              if(p[pixel[8]] > cb)
+               if(p[pixel[9]] > cb)
+                if(p[pixel[11]] > cb)
+                 if(p[pixel[6]] > cb)
+                  if(p[pixel[5]] > cb)
+                   if(p[pixel[4]] > cb)
+                    {}
+                   else
+                    if(p[pixel[12]] > cb)
+                     if(p[pixel[13]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                  else
+                   if(p[pixel[12]] > cb)
+                    if(p[pixel[13]] > cb)
+                     if(p[pixel[14]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                 else
+                  if(p[pixel[12]] > cb)
+                   if(p[pixel[13]] > cb)
+                    if(p[pixel[14]] > cb)
+                     if(p[pixel[15]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              continue;
+            else if(p[pixel[10]] < c_b)
+             if(p[pixel[11]] < c_b)
+              if(p[pixel[12]] < c_b)
+               if(p[pixel[13]] < c_b)
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               continue;
+             else
+              continue;
+            else
+             continue;
+           else if(p[pixel[3]] < c_b)
+            if(p[pixel[4]] > cb)
+             if(p[pixel[13]] > cb)
+              if(p[pixel[7]] > cb)
+               if(p[pixel[8]] > cb)
+                if(p[pixel[9]] > cb)
+                 if(p[pixel[10]] > cb)
+                  if(p[pixel[11]] > cb)
+                   if(p[pixel[12]] > cb)
+                    if(p[pixel[6]] > cb)
+                     if(p[pixel[5]] > cb)
+                      {}
+                     else
+                      if(p[pixel[14]] > cb)
+                       {}
+                      else
+                       continue;
+                    else
+                     if(p[pixel[14]] > cb)
+                      if(p[pixel[15]] > cb)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else if(p[pixel[13]] < c_b)
+              if(p[pixel[11]] > cb)
+               if(p[pixel[5]] > cb)
+                if(p[pixel[6]] > cb)
+                 if(p[pixel[7]] > cb)
+                  if(p[pixel[8]] > cb)
+                   if(p[pixel[9]] > cb)
+                    if(p[pixel[10]] > cb)
+                     if(p[pixel[12]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else if(p[pixel[11]] < c_b)
+               if(p[pixel[12]] < c_b)
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      if(p[pixel[10]] < c_b)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      if(p[pixel[10]] < c_b)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              if(p[pixel[5]] > cb)
+               if(p[pixel[6]] > cb)
+                if(p[pixel[7]] > cb)
+                 if(p[pixel[8]] > cb)
+                  if(p[pixel[9]] > cb)
+                   if(p[pixel[10]] > cb)
+                    if(p[pixel[11]] > cb)
+                     if(p[pixel[12]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else if(p[pixel[4]] < c_b)
+             if(p[pixel[5]] > cb)
+              if(p[pixel[14]] > cb)
+               if(p[pixel[7]] > cb)
+                if(p[pixel[8]] > cb)
+                 if(p[pixel[9]] > cb)
+                  if(p[pixel[10]] > cb)
+                   if(p[pixel[11]] > cb)
+                    if(p[pixel[12]] > cb)
+                     if(p[pixel[13]] > cb)
+                      if(p[pixel[6]] > cb)
+                       {}
+                      else
+                       if(p[pixel[15]] > cb)
+                        {}
+                       else
+                        continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else if(p[pixel[14]] < c_b)
+               if(p[pixel[12]] > cb)
+                if(p[pixel[6]] > cb)
+                 if(p[pixel[7]] > cb)
+                  if(p[pixel[8]] > cb)
+                   if(p[pixel[9]] > cb)
+                    if(p[pixel[10]] > cb)
+                     if(p[pixel[11]] > cb)
+                      if(p[pixel[13]] > cb)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else if(p[pixel[12]] < c_b)
+                if(p[pixel[13]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      if(p[pixel[10]] < c_b)
+                       if(p[pixel[11]] < c_b)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               if(p[pixel[6]] > cb)
+                if(p[pixel[7]] > cb)
+                 if(p[pixel[8]] > cb)
+                  if(p[pixel[9]] > cb)
+                   if(p[pixel[10]] > cb)
+                    if(p[pixel[11]] > cb)
+                     if(p[pixel[12]] > cb)
+                      if(p[pixel[13]] > cb)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else if(p[pixel[5]] < c_b)
+              if(p[pixel[6]] > cb)
+               if(p[pixel[15]] < c_b)
+                if(p[pixel[13]] > cb)
+                 if(p[pixel[7]] > cb)
+                  if(p[pixel[8]] > cb)
+                   if(p[pixel[9]] > cb)
+                    if(p[pixel[10]] > cb)
+                     if(p[pixel[11]] > cb)
+                      if(p[pixel[12]] > cb)
+                       if(p[pixel[14]] > cb)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else if(p[pixel[13]] < c_b)
+                 if(p[pixel[14]] < c_b)
+                  {}
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                if(p[pixel[7]] > cb)
+                 if(p[pixel[8]] > cb)
+                  if(p[pixel[9]] > cb)
+                   if(p[pixel[10]] > cb)
+                    if(p[pixel[11]] > cb)
+                     if(p[pixel[12]] > cb)
+                      if(p[pixel[13]] > cb)
+                       if(p[pixel[14]] > cb)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else if(p[pixel[6]] < c_b)
+               if(p[pixel[7]] > cb)
+                if(p[pixel[14]] > cb)
+                 if(p[pixel[8]] > cb)
+                  if(p[pixel[9]] > cb)
+                   if(p[pixel[10]] > cb)
+                    if(p[pixel[11]] > cb)
+                     if(p[pixel[12]] > cb)
+                      if(p[pixel[13]] > cb)
+                       if(p[pixel[15]] > cb)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  continue;
+                else
+                 continue;
+               else if(p[pixel[7]] < c_b)
+                if(p[pixel[8]] < c_b)
+                 {}
+                else
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  continue;
+               else
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[13]] > cb)
+                if(p[pixel[7]] > cb)
+                 if(p[pixel[8]] > cb)
+                  if(p[pixel[9]] > cb)
+                   if(p[pixel[10]] > cb)
+                    if(p[pixel[11]] > cb)
+                     if(p[pixel[12]] > cb)
+                      if(p[pixel[14]] > cb)
+                       if(p[pixel[15]] > cb)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else if(p[pixel[13]] < c_b)
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              if(p[pixel[12]] > cb)
+               if(p[pixel[7]] > cb)
+                if(p[pixel[8]] > cb)
+                 if(p[pixel[9]] > cb)
+                  if(p[pixel[10]] > cb)
+                   if(p[pixel[11]] > cb)
+                    if(p[pixel[13]] > cb)
+                     if(p[pixel[14]] > cb)
+                      if(p[pixel[6]] > cb)
+                       {}
+                      else
+                       if(p[pixel[15]] > cb)
+                        {}
+                       else
+                        continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else if(p[pixel[12]] < c_b)
+               if(p[pixel[13]] < c_b)
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      if(p[pixel[10]] < c_b)
+                       if(p[pixel[11]] < c_b)
+                        {}
+                       else
+                        continue;
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else
+             if(p[pixel[11]] > cb)
+              if(p[pixel[7]] > cb)
+               if(p[pixel[8]] > cb)
+                if(p[pixel[9]] > cb)
+                 if(p[pixel[10]] > cb)
+                  if(p[pixel[12]] > cb)
+                   if(p[pixel[13]] > cb)
+                    if(p[pixel[6]] > cb)
+                     if(p[pixel[5]] > cb)
+                      {}
+                     else
+                      if(p[pixel[14]] > cb)
+                       {}
+                      else
+                       continue;
+                    else
+                     if(p[pixel[14]] > cb)
+                      if(p[pixel[15]] > cb)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else if(p[pixel[11]] < c_b)
+              if(p[pixel[12]] < c_b)
+               if(p[pixel[13]] < c_b)
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      if(p[pixel[10]] < c_b)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      if(p[pixel[10]] < c_b)
+                       {}
+                      else
+                       continue;
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              continue;
+           else
+            if(p[pixel[10]] > cb)
+             if(p[pixel[7]] > cb)
+              if(p[pixel[8]] > cb)
+               if(p[pixel[9]] > cb)
+                if(p[pixel[11]] > cb)
+                 if(p[pixel[12]] > cb)
+                  if(p[pixel[6]] > cb)
+                   if(p[pixel[5]] > cb)
+                    if(p[pixel[4]] > cb)
+                     {}
+                    else
+                     if(p[pixel[13]] > cb)
+                      {}
+                     else
+                      continue;
+                   else
+                    if(p[pixel[13]] > cb)
+                     if(p[pixel[14]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                  else
+                   if(p[pixel[13]] > cb)
+                    if(p[pixel[14]] > cb)
+                     if(p[pixel[15]] > cb)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              continue;
+            else if(p[pixel[10]] < c_b)
+             if(p[pixel[11]] < c_b)
+              if(p[pixel[12]] < c_b)
+               if(p[pixel[13]] < c_b)
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     if(p[pixel[9]] < c_b)
+                      {}
+                     else
+                      continue;
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               continue;
+             else
+              continue;
+            else
+             continue;
+          else
+           if(p[pixel[9]] > cb)
+            if(p[pixel[7]] > cb)
+             if(p[pixel[8]] > cb)
+              if(p[pixel[10]] > cb)
+               if(p[pixel[11]] > cb)
+                if(p[pixel[6]] > cb)
+                 if(p[pixel[5]] > cb)
+                  if(p[pixel[4]] > cb)
+                   if(p[pixel[3]] > cb)
+                    {}
+                   else
+                    if(p[pixel[12]] > cb)
+                     {}
+                    else
+                     continue;
+                  else
+                   if(p[pixel[12]] > cb)
+                    if(p[pixel[13]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                 else
+                  if(p[pixel[12]] > cb)
+                   if(p[pixel[13]] > cb)
+                    if(p[pixel[14]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[12]] > cb)
+                  if(p[pixel[13]] > cb)
+                   if(p[pixel[14]] > cb)
+                    if(p[pixel[15]] > cb)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              continue;
+            else
+             continue;
+           else if(p[pixel[9]] < c_b)
+            if(p[pixel[10]] < c_b)
+             if(p[pixel[11]] < c_b)
+              if(p[pixel[12]] < c_b)
+               if(p[pixel[13]] < c_b)
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[3]] < c_b)
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    if(p[pixel[8]] < c_b)
+                     {}
+                    else
+                     continue;
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              continue;
+            else
+             continue;
+           else
+            continue;
+         else
+          if(p[pixel[8]] > cb)
+           if(p[pixel[7]] > cb)
+            if(p[pixel[9]] > cb)
+             if(p[pixel[10]] > cb)
+              if(p[pixel[6]] > cb)
+               if(p[pixel[5]] > cb)
+                if(p[pixel[4]] > cb)
+                 if(p[pixel[3]] > cb)
+                  if(p[pixel[2]] > cb)
+                   {}
+                  else
+                   if(p[pixel[11]] > cb)
+                    {}
+                   else
+                    continue;
+                 else
+                  if(p[pixel[11]] > cb)
+                   if(p[pixel[12]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[11]] > cb)
+                  if(p[pixel[12]] > cb)
+                   if(p[pixel[13]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[11]] > cb)
+                 if(p[pixel[12]] > cb)
+                  if(p[pixel[13]] > cb)
+                   if(p[pixel[14]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[11]] > cb)
+                if(p[pixel[12]] > cb)
+                 if(p[pixel[13]] > cb)
+                  if(p[pixel[14]] > cb)
+                   if(p[pixel[15]] > cb)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              continue;
+            else
+             continue;
+           else
+            continue;
+          else if(p[pixel[8]] < c_b)
+           if(p[pixel[9]] < c_b)
+            if(p[pixel[10]] < c_b)
+             if(p[pixel[11]] < c_b)
+              if(p[pixel[12]] < c_b)
+               if(p[pixel[13]] < c_b)
+                if(p[pixel[14]] < c_b)
+                 if(p[pixel[15]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                else
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[3]] < c_b)
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              if(p[pixel[2]] < c_b)
+               if(p[pixel[3]] < c_b)
+                if(p[pixel[4]] < c_b)
+                 if(p[pixel[5]] < c_b)
+                  if(p[pixel[6]] < c_b)
+                   if(p[pixel[7]] < c_b)
+                    {}
+                   else
+                    continue;
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else
+             continue;
+           else
+            continue;
+          else
+           continue;
+        else
+         if(p[pixel[7]] > cb)
+          if(p[pixel[8]] > cb)
+           if(p[pixel[9]] > cb)
+            if(p[pixel[6]] > cb)
+             if(p[pixel[5]] > cb)
+              if(p[pixel[4]] > cb)
+               if(p[pixel[3]] > cb)
+                if(p[pixel[2]] > cb)
+                 if(p[pixel[1]] > cb)
+                  {}
+                 else
+                  if(p[pixel[10]] > cb)
+                   {}
+                  else
+                   continue;
+                else
+                 if(p[pixel[10]] > cb)
+                  if(p[pixel[11]] > cb)
+                   {}
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[10]] > cb)
+                 if(p[pixel[11]] > cb)
+                  if(p[pixel[12]] > cb)
+                   {}
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[10]] > cb)
+                if(p[pixel[11]] > cb)
+                 if(p[pixel[12]] > cb)
+                  if(p[pixel[13]] > cb)
+                   {}
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              if(p[pixel[10]] > cb)
+               if(p[pixel[11]] > cb)
+                if(p[pixel[12]] > cb)
+                 if(p[pixel[13]] > cb)
+                  if(p[pixel[14]] > cb)
+                   {}
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else
+             if(p[pixel[10]] > cb)
+              if(p[pixel[11]] > cb)
+               if(p[pixel[12]] > cb)
+                if(p[pixel[13]] > cb)
+                 if(p[pixel[14]] > cb)
+                  if(p[pixel[15]] > cb)
+                   {}
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              continue;
+           else
+            continue;
+          else
+           continue;
+         else if(p[pixel[7]] < c_b)
+          if(p[pixel[8]] < c_b)
+           if(p[pixel[9]] < c_b)
+            if(p[pixel[6]] < c_b)
+             if(p[pixel[5]] < c_b)
+              if(p[pixel[4]] < c_b)
+               if(p[pixel[3]] < c_b)
+                if(p[pixel[2]] < c_b)
+                 if(p[pixel[1]] < c_b)
+                  {}
+                 else
+                  if(p[pixel[10]] < c_b)
+                   {}
+                  else
+                   continue;
+                else
+                 if(p[pixel[10]] < c_b)
+                  if(p[pixel[11]] < c_b)
+                   {}
+                  else
+                   continue;
+                 else
+                  continue;
+               else
+                if(p[pixel[10]] < c_b)
+                 if(p[pixel[11]] < c_b)
+                  if(p[pixel[12]] < c_b)
+                   {}
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+              else
+               if(p[pixel[10]] < c_b)
+                if(p[pixel[11]] < c_b)
+                 if(p[pixel[12]] < c_b)
+                  if(p[pixel[13]] < c_b)
+                   {}
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+             else
+              if(p[pixel[10]] < c_b)
+               if(p[pixel[11]] < c_b)
+                if(p[pixel[12]] < c_b)
+                 if(p[pixel[13]] < c_b)
+                  if(p[pixel[14]] < c_b)
+                   {}
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+            else
+             if(p[pixel[10]] < c_b)
+              if(p[pixel[11]] < c_b)
+               if(p[pixel[12]] < c_b)
+                if(p[pixel[13]] < c_b)
+                 if(p[pixel[14]] < c_b)
+                  if(p[pixel[15]] < c_b)
+                   {}
+                  else
+                   continue;
+                 else
+                  continue;
+                else
+                 continue;
+               else
+                continue;
+              else
+               continue;
+             else
+              continue;
+           else
+            continue;
+          else
+           continue;
+         else
+          continue;
+			if(num_corners == rsize)
+			{
+				rsize*=2;
+				ret_corners = (xy*)realloc(ret_corners, sizeof(xy)*rsize);
+			}
+			ret_corners[num_corners].x = x;
+			ret_corners[num_corners].y = y;
+			num_corners++;
+
+		}
+
+	*ret_num_corners = num_corners;
+	return ret_corners;
+
+}
+
+// clang-format on
diff --git a/third_party/aom/third_party/fastfeat/nonmax.c b/third_party/aom/third_party/fastfeat/nonmax.c
new file mode 100644
index 0000000000..0438c4dc18
--- /dev/null
+++ b/third_party/aom/third_party/fastfeat/nonmax.c
@@ -0,0 +1,121 @@
+// clang-format off
+#include <stdlib.h>
+#include "fast.h"
+
+
+#define Compare(X, Y) ((X)>=(Y))
+
+xy* nonmax_suppression(const xy* corners, const int* scores, int num_corners, int* ret_num_nonmax)
+{
+  int num_nonmax=0;
+  int last_row;
+  int* row_start;
+  int i, j;
+  xy* ret_nonmax;
+  const int sz = (int)num_corners;
+
+  /*Point above points (roughly) to the pixel above the one of interest, if there
+    is a feature there.*/
+  int point_above = 0;
+  int point_below = 0;
+
+
+  if(num_corners < 1)
+  {
+    *ret_num_nonmax = 0;
+    return 0;
+  }
+
+  ret_nonmax = (xy*)malloc(num_corners * sizeof(xy));
+
+  /* Find where each row begins
+     (the corners are output in raster scan order). A beginning of -1 signifies
+     that there are no corners on that row. */
+  last_row = corners[num_corners-1].y;
+  row_start = (int*)malloc((last_row+1)*sizeof(int));
+
+  for(i=0; i < last_row+1; i++)
+    row_start[i] = -1;
+
+  {
+    int prev_row = -1;
+    for(i=0; i< num_corners; i++)
+      if(corners[i].y != prev_row)
+      {
+        row_start[corners[i].y] = i;
+        prev_row = corners[i].y;
+      }
+  }
+
+
+
+  for(i=0; i < sz; i++)
+  {
+    int score = scores[i];
+    xy pos = corners[i];
+
+    /*Check left */
+    if(i > 0)
+      if(corners[i-1].x == pos.x-1 && corners[i-1].y == pos.y && Compare(scores[i-1], score))
+        continue;
+
+    /*Check right*/
+    if(i < (sz - 1))
+      if(corners[i+1].x == pos.x+1 && corners[i+1].y == pos.y && Compare(scores[i+1], score))
+        continue;
+
+    /*Check above (if there is a valid row above)*/
+    if(pos.y > 0)
+      if (row_start[pos.y - 1] != -1)
+      {
+        /*Make sure that current point_above is one
+          row above.*/
+        if(corners[point_above].y < pos.y - 1)
+          point_above = row_start[pos.y-1];
+
+        /*Make point_above point to the first of the pixels above the current point,
+          if it exists.*/
+        for(; corners[point_above].y < pos.y && corners[point_above].x < pos.x - 1; point_above++)
+        {}
+
+
+        for(j=point_above; corners[j].y < pos.y && corners[j].x <= pos.x + 1; j++)
+        {
+          int x = corners[j].x;
+          if( (x == pos.x - 1 || x ==pos.x || x == pos.x+1) && Compare(scores[j], score))
+            goto cont;
+        }
+
+      }
+
+    /*Check below (if there is anything below)*/
+    if(pos.y >= 0)
+      if (pos.y != last_row && row_start[pos.y + 1] != -1 && point_below < sz) /*Nothing below*/
+      {
+        if(corners[point_below].y < pos.y + 1)
+          point_below = row_start[pos.y+1];
+
+        /* Make point below point to one of the pixels belowthe current point, if it
+           exists.*/
+        for(; point_below < sz && corners[point_below].y == pos.y+1 && corners[point_below].x < pos.x - 1; point_below++)
+        {}
+
+        for(j=point_below; j < sz && corners[j].y == pos.y+1 && corners[j].x <= pos.x + 1; j++)
+        {
+          int x = corners[j].x;
+          if( (x == pos.x - 1 || x ==pos.x || x == pos.x+1) && Compare(scores[j],score))
+            goto cont;
+        }
+      }
+
+    ret_nonmax[num_nonmax++] = corners[i];
+cont:
+    ;
+  }
+
+  free(row_start);
+  *ret_num_nonmax = num_nonmax;
+  return ret_nonmax;
+}
+
+// clang-format on
diff --git a/third_party/aom/third_party/googletest/README.libaom b/third_party/aom/third_party/googletest/README.libaom
new file mode 100644
index 0000000000..a53d7e0085
--- /dev/null
+++ b/third_party/aom/third_party/googletest/README.libaom
@@ -0,0 +1,24 @@
+URL: https://github.com/google/googletest
+Version: 1.8.0
+License: BSD
+License File: LICENSE
+
+Description:
+Google's framework for writing C++ tests on a variety of platforms
+(Linux, Mac OS X, Windows, Windows CE, Symbian, etc).  Based on the
+xUnit architecture.  Supports automatic test discovery, a rich set of
+assertions, user-defined assertions, death tests, fatal and non-fatal
+failures, various options for running the tests, and XML test report
+generation.
+
+Local Modifications:
+- Remove everything but:
+  googletest-release-1.8.0/googletest/
+    cmake/
+    include/
+    src/
+    CHANGES
+    CMakelists.txt
+    CONTRIBUTORS
+    LICENSE
+    README.md
diff --git a/third_party/aom/third_party/googletest/gtest.mk b/third_party/aom/third_party/googletest/gtest.mk
new file mode 100644
index 0000000000..fc4dbdc240
--- /dev/null
+++ b/third_party/aom/third_party/googletest/gtest.mk
@@ -0,0 +1 @@
+GTEST_SRCS-yes += googletest/src/googletest/src/gtest-all.cc
diff --git a/third_party/aom/third_party/googletest/src/googletest/CHANGES b/third_party/aom/third_party/googletest/src/googletest/CHANGES
new file mode 100644
index 0000000000..0552132421
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/CHANGES
@@ -0,0 +1,157 @@
+Changes for 1.7.0:
+
+* New feature: death tests are supported on OpenBSD and in iOS
+  simulator now.
+* New feature: Google Test now implements a protocol to allow
+  a test runner to detect that a test program has exited
+  prematurely and report it as a failure (before it would be
+  falsely reported as a success if the exit code is 0).
+* New feature: Test::RecordProperty() can now be used outside of the
+  lifespan of a test method, in which case it will be attributed to
+  the current test case or the test program in the XML report.
+* New feature (potentially breaking): --gtest_list_tests now prints
+  the type parameters and value parameters for each test.
+* Improvement: char pointers and char arrays are now escaped properly
+  in failure messages.
+* Improvement: failure summary in XML reports now includes file and
+  line information.
+* Improvement: the <testsuites> XML element now has a timestamp attribute.
+* Improvement: When --gtest_filter is specified, XML report now doesn't
+  contain information about tests that are filtered out.
+* Fixed the bug where long --gtest_filter flag values are truncated in
+  death tests.
+* Potentially breaking change: RUN_ALL_TESTS() is now implemented as a
+  function instead of a macro in order to work better with Clang.
+* Compatibility fixes with C++ 11 and various platforms.
+* Bug/warning fixes.
+
+Changes for 1.6.0:
+
+* New feature: ADD_FAILURE_AT() for reporting a test failure at the
+  given source location -- useful for writing testing utilities.
+* New feature: the universal value printer is moved from Google Mock
+  to Google Test.
+* New feature: type parameters and value parameters are reported in
+  the XML report now.
+* A gtest_disable_pthreads CMake option.
+* Colored output works in GNU Screen sessions now.
+* Parameters of value-parameterized tests are now printed in the
+  textual output.
+* Failures from ad hoc test assertions run before RUN_ALL_TESTS() are
+  now correctly reported.
+* Arguments of ASSERT_XY and EXPECT_XY no longer need to support << to
+  ostream.
+* More complete handling of exceptions.
+* GTEST_ASSERT_XY can be used instead of ASSERT_XY in case the latter
+  name is already used by another library.
+* --gtest_catch_exceptions is now true by default, allowing a test
+  program to continue after an exception is thrown.
+* Value-parameterized test fixtures can now derive from Test and
+  WithParamInterface<T> separately, easing conversion of legacy tests.
+* Death test messages are clearly marked to make them more
+  distinguishable from other messages.
+* Compatibility fixes for Android, Google Native Client, MinGW, HP UX,
+  PowerPC, Lucid autotools, libCStd, Sun C++, Borland C++ Builder (Code Gear),
+  IBM XL C++ (Visual Age C++), and C++0x.
+* Bug fixes and implementation clean-ups.
+* Potentially incompatible changes: disables the harmful 'make install'
+  command in autotools.
+
+Changes for 1.5.0:
+
+ * New feature: assertions can be safely called in multiple threads
+   where the pthreads library is available.
+ * New feature: predicates used inside EXPECT_TRUE() and friends
+   can now generate custom failure messages.
+ * New feature: Google Test can now be compiled as a DLL.
+ * New feature: fused source files are included.
+ * New feature: prints help when encountering unrecognized Google Test flags.
+ * Experimental feature: CMake build script (requires CMake 2.6.4+).
+ * Experimental feature: the Pump script for meta programming.
+ * double values streamed to an assertion are printed with enough precision
+   to differentiate any two different values.
+ * Google Test now works on Solaris and AIX.
+ * Build and test script improvements.
+ * Bug fixes and implementation clean-ups.
+
+ Potentially breaking changes:
+
+ * Stopped supporting VC++ 7.1 with exceptions disabled.
+ * Dropped support for 'make install'.
+
+Changes for 1.4.0:
+
+ * New feature: the event listener API
+ * New feature: test shuffling
+ * New feature: the XML report format is closer to junitreport and can
+   be parsed by Hudson now.
+ * New feature: when a test runs under Visual Studio, its failures are
+   integrated in the IDE.
+ * New feature: /MD(d) versions of VC++ projects.
+ * New feature: elapsed time for the tests is printed by default.
+ * New feature: comes with a TR1 tuple implementation such that Boost
+   is no longer needed for Combine().
+ * New feature: EXPECT_DEATH_IF_SUPPORTED macro and friends.
+ * New feature: the Xcode project can now produce static gtest
+   libraries in addition to a framework.
+ * Compatibility fixes for Solaris, Cygwin, minGW, Windows Mobile,
+   Symbian, gcc, and C++Builder.
+ * Bug fixes and implementation clean-ups.
+
+Changes for 1.3.0:
+
+ * New feature: death tests on Windows, Cygwin, and Mac.
+ * New feature: ability to use Google Test assertions in other testing
+   frameworks.
+ * New feature: ability to run disabled test via
+   --gtest_also_run_disabled_tests.
+ * New feature: the --help flag for printing the usage.
+ * New feature: access to Google Test flag values in user code.
+ * New feature: a script that packs Google Test into one .h and one
+   .cc file for easy deployment.
+ * New feature: support for distributing test functions to multiple
+   machines (requires support from the test runner).
+ * Bug fixes and implementation clean-ups.
+
+Changes for 1.2.1:
+
+ * Compatibility fixes for Linux IA-64 and IBM z/OS.
+ * Added support for using Boost and other TR1 implementations.
+ * Changes to the build scripts to support upcoming release of Google C++
+   Mocking Framework.
+ * Added Makefile to the distribution package.
+ * Improved build instructions in README.
+
+Changes for 1.2.0:
+
+ * New feature: value-parameterized tests.
+ * New feature: the ASSERT/EXPECT_(NON)FATAL_FAILURE(_ON_ALL_THREADS)
+   macros.
+ * Changed the XML report format to match JUnit/Ant's.
+ * Added tests to the Xcode project.
+ * Added scons/SConscript for building with SCons.
+ * Added src/gtest-all.cc for building Google Test from a single file.
+ * Fixed compatibility with Solaris and z/OS.
+ * Enabled running Python tests on systems with python 2.3 installed,
+   e.g. Mac OS X 10.4.
+ * Bug fixes.
+
+Changes for 1.1.0:
+
+ * New feature: type-parameterized tests.
+ * New feature: exception assertions.
+ * New feature: printing elapsed time of tests.
+ * Improved the robustness of death tests.
+ * Added an Xcode project and samples.
+ * Adjusted the output format on Windows to be understandable by Visual Studio.
+ * Minor bug fixes.
+
+Changes for 1.0.1:
+
+ * Added project files for Visual Studio 7.1.
+ * Fixed issues with compiling on Mac OS X.
+ * Fixed issues with compiling on Cygwin.
+
+Changes for 1.0.0:
+
+ * Initial Open Source release of Google Test
diff --git a/third_party/aom/third_party/googletest/src/googletest/CMakeLists.txt b/third_party/aom/third_party/googletest/src/googletest/CMakeLists.txt
new file mode 100644
index 0000000000..621d0f0421
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/CMakeLists.txt
@@ -0,0 +1,286 @@
+########################################################################
+# CMake build script for Google Test.
+#
+# To run the tests for Google Test itself on Linux, use 'make test' or
+# ctest.  You can select which tests to run using 'ctest -R regex'.
+# For more options, run 'ctest --help'.
+
+# BUILD_SHARED_LIBS is a standard CMake variable, but we declare it here to
+# make it prominent in the GUI.
+option(BUILD_SHARED_LIBS "Build shared libraries (DLLs)." OFF)
+
+# When other libraries are using a shared version of runtime libraries,
+# Google Test also has to use one.
+option(
+  gtest_force_shared_crt
+  "Use shared (DLL) run-time lib even when Google Test is built as static lib."
+  OFF)
+
+option(gtest_build_tests "Build all of gtest's own tests." OFF)
+
+option(gtest_build_samples "Build gtest's sample programs." OFF)
+
+option(gtest_disable_pthreads "Disable uses of pthreads in gtest." OFF)
+
+option(
+  gtest_hide_internal_symbols
+  "Build gtest with internal symbols hidden in shared libraries."
+  OFF)
+
+# Defines pre_project_set_up_hermetic_build() and set_up_hermetic_build().
+include(cmake/hermetic_build.cmake OPTIONAL)
+
+if (COMMAND pre_project_set_up_hermetic_build)
+  pre_project_set_up_hermetic_build()
+endif()
+
+########################################################################
+#
+# Project-wide settings
+
+# Name of the project.
+#
+# CMake files in this project can refer to the root source directory
+# as ${gtest_SOURCE_DIR} and to the root binary directory as
+# ${gtest_BINARY_DIR}.
+# Language "C" is required for find_package(Threads).
+project(gtest CXX C)
+cmake_minimum_required(VERSION 2.6.2)
+
+if (COMMAND set_up_hermetic_build)
+  set_up_hermetic_build()
+endif()
+
+if (gtest_hide_internal_symbols)
+  set(CMAKE_CXX_VISIBILITY_PRESET hidden)
+  set(CMAKE_VISIBILITY_INLINES_HIDDEN 1)
+endif()
+
+# Define helper functions and macros used by Google Test.
+include(cmake/internal_utils.cmake)
+
+config_compiler_and_linker()  # Defined in internal_utils.cmake.
+
+# Where Google Test's .h files can be found.
+include_directories(
+  ${gtest_SOURCE_DIR}/include
+  ${gtest_SOURCE_DIR})
+
+# Where Google Test's libraries can be found.
+link_directories(${gtest_BINARY_DIR}/src)
+
+# Summary of tuple support for Microsoft Visual Studio:
+# Compiler    version(MS)  version(cmake)  Support
+# ----------  -----------  --------------  -----------------------------
+# <= VS 2010  <= 10        <= 1600         Use Google Tests's own tuple.
+# VS 2012     11           1700            std::tr1::tuple + _VARIADIC_MAX=10
+# VS 2013     12           1800            std::tr1::tuple
+if (MSVC AND MSVC_VERSION EQUAL 1700)
+  add_definitions(/D _VARIADIC_MAX=10)
+endif()
+
+########################################################################
+#
+# Defines the gtest & gtest_main libraries.  User tests should link
+# with one of them.
+
+# Google Test libraries.  We build them using more strict warnings than what
+# are used for other targets, to ensure that gtest can be compiled by a user
+# aggressive about warnings.
+cxx_library(gtest "${cxx_strict}" src/gtest-all.cc)
+cxx_library(gtest_main "${cxx_strict}" src/gtest_main.cc)
+target_link_libraries(gtest_main gtest)
+
+# If the CMake version supports it, attach header directory information
+# to the targets for when we are part of a parent build (ie being pulled
+# in via add_subdirectory() rather than being a standalone build).
+if (DEFINED CMAKE_VERSION AND NOT "${CMAKE_VERSION}" VERSION_LESS "2.8.11")
+  target_include_directories(gtest      INTERFACE "${gtest_SOURCE_DIR}/include")
+  target_include_directories(gtest_main INTERFACE "${gtest_SOURCE_DIR}/include")
+endif()
+
+########################################################################
+#
+# Install rules
+install(TARGETS gtest gtest_main
+  DESTINATION lib)
+install(DIRECTORY ${gtest_SOURCE_DIR}/include/gtest
+  DESTINATION include)
+
+########################################################################
+#
+# Samples on how to link user tests with gtest or gtest_main.
+#
+# They are not built by default.  To build them, set the
+# gtest_build_samples option to ON.  You can do it by running ccmake
+# or specifying the -Dgtest_build_samples=ON flag when running cmake.
+
+if (gtest_build_samples)
+  cxx_executable(sample1_unittest samples gtest_main samples/sample1.cc)
+  cxx_executable(sample2_unittest samples gtest_main samples/sample2.cc)
+  cxx_executable(sample3_unittest samples gtest_main)
+  cxx_executable(sample4_unittest samples gtest_main samples/sample4.cc)
+  cxx_executable(sample5_unittest samples gtest_main samples/sample1.cc)
+  cxx_executable(sample6_unittest samples gtest_main)
+  cxx_executable(sample7_unittest samples gtest_main)
+  cxx_executable(sample8_unittest samples gtest_main)
+  cxx_executable(sample9_unittest samples gtest)
+  cxx_executable(sample10_unittest samples gtest)
+endif()
+
+########################################################################
+#
+# Google Test's own tests.
+#
+# You can skip this section if you aren't interested in testing
+# Google Test itself.
+#
+# The tests are not built by default.  To build them, set the
+# gtest_build_tests option to ON.  You can do it by running ccmake
+# or specifying the -Dgtest_build_tests=ON flag when running cmake.
+
+if (gtest_build_tests)
+  # This must be set in the root directory for the tests to be run by
+  # 'make test' or ctest.
+  enable_testing()
+
+  ############################################################
+  # C++ tests built with standard compiler flags.
+
+  cxx_test(gtest-death-test_test gtest_main)
+  cxx_test(gtest_environment_test gtest)
+  cxx_test(gtest-filepath_test gtest_main)
+  cxx_test(gtest-linked_ptr_test gtest_main)
+  cxx_test(gtest-listener_test gtest_main)
+  cxx_test(gtest_main_unittest gtest_main)
+  cxx_test(gtest-message_test gtest_main)
+  cxx_test(gtest_no_test_unittest gtest)
+  cxx_test(gtest-options_test gtest_main)
+  cxx_test(gtest-param-test_test gtest
+    test/gtest-param-test2_test.cc)
+  cxx_test(gtest-port_test gtest_main)
+  cxx_test(gtest_pred_impl_unittest gtest_main)
+  cxx_test(gtest_premature_exit_test gtest
+    test/gtest_premature_exit_test.cc)
+  cxx_test(gtest-printers_test gtest_main)
+  cxx_test(gtest_prod_test gtest_main
+    test/production.cc)
+  cxx_test(gtest_repeat_test gtest)
+  cxx_test(gtest_sole_header_test gtest_main)
+  cxx_test(gtest_stress_test gtest)
+  cxx_test(gtest-test-part_test gtest_main)
+  cxx_test(gtest_throw_on_failure_ex_test gtest)
+  cxx_test(gtest-typed-test_test gtest_main
+    test/gtest-typed-test2_test.cc)
+  cxx_test(gtest_unittest gtest_main)
+  cxx_test(gtest-unittest-api_test gtest)
+
+  ############################################################
+  # C++ tests built with non-standard compiler flags.
+
+  # MSVC 7.1 does not support STL with exceptions disabled.
+  if (NOT MSVC OR MSVC_VERSION GREATER 1310)
+    cxx_library(gtest_no_exception "${cxx_no_exception}"
+      src/gtest-all.cc)
+    cxx_library(gtest_main_no_exception "${cxx_no_exception}"
+      src/gtest-all.cc src/gtest_main.cc)
+  endif()
+  cxx_library(gtest_main_no_rtti "${cxx_no_rtti}"
+    src/gtest-all.cc src/gtest_main.cc)
+
+  cxx_test_with_flags(gtest-death-test_ex_nocatch_test
+    "${cxx_exception} -DGTEST_ENABLE_CATCH_EXCEPTIONS_=0"
+    gtest test/gtest-death-test_ex_test.cc)
+  cxx_test_with_flags(gtest-death-test_ex_catch_test
+    "${cxx_exception} -DGTEST_ENABLE_CATCH_EXCEPTIONS_=1"
+    gtest test/gtest-death-test_ex_test.cc)
+
+  cxx_test_with_flags(gtest_no_rtti_unittest "${cxx_no_rtti}"
+    gtest_main_no_rtti test/gtest_unittest.cc)
+
+  cxx_shared_library(gtest_dll "${cxx_default}"
+    src/gtest-all.cc src/gtest_main.cc)
+
+  cxx_executable_with_flags(gtest_dll_test_ "${cxx_default}"
+    gtest_dll test/gtest_all_test.cc)
+  set_target_properties(gtest_dll_test_
+                        PROPERTIES
+                        COMPILE_DEFINITIONS "GTEST_LINKED_AS_SHARED_LIBRARY=1")
+
+  if (NOT MSVC OR MSVC_VERSION LESS 1600)  # 1600 is Visual Studio 2010.
+    # Visual Studio 2010, 2012, and 2013 define symbols in std::tr1 that
+    # conflict with our own definitions. Therefore using our own tuple does not
+    # work on those compilers.
+    cxx_library(gtest_main_use_own_tuple "${cxx_use_own_tuple}"
+      src/gtest-all.cc src/gtest_main.cc)
+
+    cxx_test_with_flags(gtest-tuple_test "${cxx_use_own_tuple}"
+      gtest_main_use_own_tuple test/gtest-tuple_test.cc)
+
+    cxx_test_with_flags(gtest_use_own_tuple_test "${cxx_use_own_tuple}"
+      gtest_main_use_own_tuple
+      test/gtest-param-test_test.cc test/gtest-param-test2_test.cc)
+  endif()
+
+  ############################################################
+  # Python tests.
+
+  cxx_executable(gtest_break_on_failure_unittest_ test gtest)
+  py_test(gtest_break_on_failure_unittest)
+
+  # Visual Studio .NET 2003 does not support STL with exceptions disabled.
+  if (NOT MSVC OR MSVC_VERSION GREATER 1310)  # 1310 is Visual Studio .NET 2003
+    cxx_executable_with_flags(
+      gtest_catch_exceptions_no_ex_test_
+      "${cxx_no_exception}"
+      gtest_main_no_exception
+      test/gtest_catch_exceptions_test_.cc)
+  endif()
+
+  cxx_executable_with_flags(
+    gtest_catch_exceptions_ex_test_
+    "${cxx_exception}"
+    gtest_main
+    test/gtest_catch_exceptions_test_.cc)
+  py_test(gtest_catch_exceptions_test)
+
+  cxx_executable(gtest_color_test_ test gtest)
+  py_test(gtest_color_test)
+
+  cxx_executable(gtest_env_var_test_ test gtest)
+  py_test(gtest_env_var_test)
+
+  cxx_executable(gtest_filter_unittest_ test gtest)
+  py_test(gtest_filter_unittest)
+
+  cxx_executable(gtest_help_test_ test gtest_main)
+  py_test(gtest_help_test)
+
+  cxx_executable(gtest_list_tests_unittest_ test gtest)
+  py_test(gtest_list_tests_unittest)
+
+  cxx_executable(gtest_output_test_ test gtest)
+  py_test(gtest_output_test)
+
+  cxx_executable(gtest_shuffle_test_ test gtest)
+  py_test(gtest_shuffle_test)
+
+  # MSVC 7.1 does not support STL with exceptions disabled.
+  if (NOT MSVC OR MSVC_VERSION GREATER 1310)
+    cxx_executable(gtest_throw_on_failure_test_ test gtest_no_exception)
+    set_target_properties(gtest_throw_on_failure_test_
+      PROPERTIES
+      COMPILE_FLAGS "${cxx_no_exception}")
+    py_test(gtest_throw_on_failure_test)
+  endif()
+
+  cxx_executable(gtest_uninitialized_test_ test gtest)
+  py_test(gtest_uninitialized_test)
+
+  cxx_executable(gtest_xml_outfile1_test_ test gtest_main)
+  cxx_executable(gtest_xml_outfile2_test_ test gtest_main)
+  py_test(gtest_xml_outfiles_test)
+
+  cxx_executable(gtest_xml_output_unittest_ test gtest)
+  py_test(gtest_xml_output_unittest)
+endif()
diff --git a/third_party/aom/third_party/googletest/src/googletest/CONTRIBUTORS b/third_party/aom/third_party/googletest/src/googletest/CONTRIBUTORS
new file mode 100644
index 0000000000..feae2fc044
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/CONTRIBUTORS
@@ -0,0 +1,37 @@
+# This file contains a list of people who've made non-trivial
+# contribution to the Google C++ Testing Framework project.  People
+# who commit code to the project are encouraged to add their names
+# here.  Please keep the list sorted by first names.
+
+Ajay Joshi <jaj@google.com>
+Balázs Dán <balazs.dan@gmail.com>
+Bharat Mediratta <bharat@menalto.com>
+Chandler Carruth <chandlerc@google.com>
+Chris Prince <cprince@google.com>
+Chris Taylor <taylorc@google.com>
+Dan Egnor <egnor@google.com>
+Eric Roman <eroman@chromium.org>
+Hady Zalek <hady.zalek@gmail.com>
+Jeffrey Yasskin <jyasskin@google.com>
+Jói Sigurðsson <joi@google.com>
+Keir Mierle <mierle@gmail.com>
+Keith Ray <keith.ray@gmail.com>
+Kenton Varda <kenton@google.com>
+Manuel Klimek <klimek@google.com>
+Markus Heule <markus.heule@gmail.com>
+Mika Raento <mikie@iki.fi>
+Miklós Fazekas <mfazekas@szemafor.com>
+Pasi Valminen <pasi.valminen@gmail.com>
+Patrick Hanna <phanna@google.com>
+Patrick Riley <pfr@google.com>
+Peter Kaminski <piotrk@google.com>
+Preston Jackson <preston.a.jackson@gmail.com>
+Rainer Klaffenboeck <rainer.klaffenboeck@dynatrace.com>
+Russ Cox <rsc@google.com>
+Russ Rufer <russ@pentad.com>
+Sean Mcafee <eefacm@gmail.com>
+Sigurður Ásgeirsson <siggi@google.com>
+Tracy Bialik <tracy@pentad.com>
+Vadim Berman <vadimb@google.com>
+Vlad Losev <vladl@google.com>
+Zhanyong Wan <wan@google.com>
diff --git a/third_party/aom/third_party/googletest/src/googletest/LICENSE b/third_party/aom/third_party/googletest/src/googletest/LICENSE
new file mode 100644
index 0000000000..1941a11f8c
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/LICENSE
@@ -0,0 +1,28 @@
+Copyright 2008, Google Inc.
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+    * Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above
+copyright notice, this list of conditions and the following disclaimer
+in the documentation and/or other materials provided with the
+distribution.
+    * Neither the name of Google Inc. nor the names of its
+contributors may be used to endorse or promote products derived from
+this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
diff --git a/third_party/aom/third_party/googletest/src/googletest/README.md b/third_party/aom/third_party/googletest/src/googletest/README.md
new file mode 100644
index 0000000000..edd4408054
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/README.md
@@ -0,0 +1,280 @@
+
+### Generic Build Instructions ###
+
+#### Setup ####
+
+To build Google Test and your tests that use it, you need to tell your
+build system where to find its headers and source files.  The exact
+way to do it depends on which build system you use, and is usually
+straightforward.
+
+#### Build ####
+
+Suppose you put Google Test in directory `${GTEST_DIR}`.  To build it,
+create a library build target (or a project as called by Visual Studio
+and Xcode) to compile
+
+    ${GTEST_DIR}/src/gtest-all.cc
+
+with `${GTEST_DIR}/include` in the system header search path and `${GTEST_DIR}`
+in the normal header search path.  Assuming a Linux-like system and gcc,
+something like the following will do:
+
+    g++ -isystem ${GTEST_DIR}/include -I${GTEST_DIR} \
+        -pthread -c ${GTEST_DIR}/src/gtest-all.cc
+    ar -rv libgtest.a gtest-all.o
+
+(We need `-pthread` as Google Test uses threads.)
+
+Next, you should compile your test source file with
+`${GTEST_DIR}/include` in the system header search path, and link it
+with gtest and any other necessary libraries:
+
+    g++ -isystem ${GTEST_DIR}/include -pthread path/to/your_test.cc libgtest.a \
+        -o your_test
+
+As an example, the make/ directory contains a Makefile that you can
+use to build Google Test on systems where GNU make is available
+(e.g. Linux, Mac OS X, and Cygwin).  It doesn't try to build Google
+Test's own tests.  Instead, it just builds the Google Test library and
+a sample test.  You can use it as a starting point for your own build
+script.
+
+If the default settings are correct for your environment, the
+following commands should succeed:
+
+    cd ${GTEST_DIR}/make
+    make
+    ./sample1_unittest
+
+If you see errors, try to tweak the contents of `make/Makefile` to make
+them go away.  There are instructions in `make/Makefile` on how to do
+it.
+
+### Using CMake ###
+
+Google Test comes with a CMake build script (
+[CMakeLists.txt](CMakeLists.txt)) that can be used on a wide range of platforms ("C" stands for
+cross-platform.). If you don't have CMake installed already, you can
+download it for free from <http://www.cmake.org/>.
+
+CMake works by generating native makefiles or build projects that can
+be used in the compiler environment of your choice.  The typical
+workflow starts with:
+
+    mkdir mybuild       # Create a directory to hold the build output.
+    cd mybuild
+    cmake ${GTEST_DIR}  # Generate native build scripts.
+
+If you want to build Google Test's samples, you should replace the
+last command with
+
+    cmake -Dgtest_build_samples=ON ${GTEST_DIR}
+
+If you are on a \*nix system, you should now see a Makefile in the
+current directory.  Just type 'make' to build gtest.
+
+If you use Windows and have Visual Studio installed, a `gtest.sln` file
+and several `.vcproj` files will be created.  You can then build them
+using Visual Studio.
+
+On Mac OS X with Xcode installed, a `.xcodeproj` file will be generated.
+
+### Legacy Build Scripts ###
+
+Before settling on CMake, we have been providing hand-maintained build
+projects/scripts for Visual Studio, Xcode, and Autotools.  While we
+continue to provide them for convenience, they are not actively
+maintained any more.  We highly recommend that you follow the
+instructions in the previous two sections to integrate Google Test
+with your existing build system.
+
+If you still need to use the legacy build scripts, here's how:
+
+The msvc\ folder contains two solutions with Visual C++ projects.
+Open the `gtest.sln` or `gtest-md.sln` file using Visual Studio, and you
+are ready to build Google Test the same way you build any Visual
+Studio project.  Files that have names ending with -md use DLL
+versions of Microsoft runtime libraries (the /MD or the /MDd compiler
+option).  Files without that suffix use static versions of the runtime
+libraries (the /MT or the /MTd option).  Please note that one must use
+the same option to compile both gtest and the test code.  If you use
+Visual Studio 2005 or above, we recommend the -md version as /MD is
+the default for new projects in these versions of Visual Studio.
+
+On Mac OS X, open the `gtest.xcodeproj` in the `xcode/` folder using
+Xcode.  Build the "gtest" target.  The universal binary framework will
+end up in your selected build directory (selected in the Xcode
+"Preferences..." -> "Building" pane and defaults to xcode/build).
+Alternatively, at the command line, enter:
+
+    xcodebuild
+
+This will build the "Release" configuration of gtest.framework in your
+default build location.  See the "xcodebuild" man page for more
+information about building different configurations and building in
+different locations.
+
+If you wish to use the Google Test Xcode project with Xcode 4.x and
+above, you need to either:
+
+ * update the SDK configuration options in xcode/Config/General.xconfig.
+   Comment options `SDKROOT`, `MACOS_DEPLOYMENT_TARGET`, and `GCC_VERSION`. If
+   you choose this route you lose the ability to target earlier versions
+   of MacOS X.
+ * Install an SDK for an earlier version. This doesn't appear to be
+   supported by Apple, but has been reported to work
+   (http://stackoverflow.com/questions/5378518).
+
+### Tweaking Google Test ###
+
+Google Test can be used in diverse environments.  The default
+configuration may not work (or may not work well) out of the box in
+some environments.  However, you can easily tweak Google Test by
+defining control macros on the compiler command line.  Generally,
+these macros are named like `GTEST_XYZ` and you define them to either 1
+or 0 to enable or disable a certain feature.
+
+We list the most frequently used macros below.  For a complete list,
+see file [include/gtest/internal/gtest-port.h](include/gtest/internal/gtest-port.h).
+
+### Choosing a TR1 Tuple Library ###
+
+Some Google Test features require the C++ Technical Report 1 (TR1)
+tuple library, which is not yet available with all compilers.  The
+good news is that Google Test implements a subset of TR1 tuple that's
+enough for its own need, and will automatically use this when the
+compiler doesn't provide TR1 tuple.
+
+Usually you don't need to care about which tuple library Google Test
+uses.  However, if your project already uses TR1 tuple, you need to
+tell Google Test to use the same TR1 tuple library the rest of your
+project uses, or the two tuple implementations will clash.  To do
+that, add
+
+    -DGTEST_USE_OWN_TR1_TUPLE=0
+
+to the compiler flags while compiling Google Test and your tests.  If
+you want to force Google Test to use its own tuple library, just add
+
+    -DGTEST_USE_OWN_TR1_TUPLE=1
+
+to the compiler flags instead.
+
+If you don't want Google Test to use tuple at all, add
+
+    -DGTEST_HAS_TR1_TUPLE=0
+
+and all features using tuple will be disabled.
+
+### Multi-threaded Tests ###
+
+Google Test is thread-safe where the pthread library is available.
+After `#include "gtest/gtest.h"`, you can check the `GTEST_IS_THREADSAFE`
+macro to see whether this is the case (yes if the macro is `#defined` to
+1, no if it's undefined.).
+
+If Google Test doesn't correctly detect whether pthread is available
+in your environment, you can force it with
+
+    -DGTEST_HAS_PTHREAD=1
+
+or
+
+    -DGTEST_HAS_PTHREAD=0
+
+When Google Test uses pthread, you may need to add flags to your
+compiler and/or linker to select the pthread library, or you'll get
+link errors.  If you use the CMake script or the deprecated Autotools
+script, this is taken care of for you.  If you use your own build
+script, you'll need to read your compiler and linker's manual to
+figure out what flags to add.
+
+### As a Shared Library (DLL) ###
+
+Google Test is compact, so most users can build and link it as a
+static library for the simplicity.  You can choose to use Google Test
+as a shared library (known as a DLL on Windows) if you prefer.
+
+To compile *gtest* as a shared library, add
+
+    -DGTEST_CREATE_SHARED_LIBRARY=1
+
+to the compiler flags.  You'll also need to tell the linker to produce
+a shared library instead - consult your linker's manual for how to do
+it.
+
+To compile your *tests* that use the gtest shared library, add
+
+    -DGTEST_LINKED_AS_SHARED_LIBRARY=1
+
+to the compiler flags.
+
+Note: while the above steps aren't technically necessary today when
+using some compilers (e.g. GCC), they may become necessary in the
+future, if we decide to improve the speed of loading the library (see
+<http://gcc.gnu.org/wiki/Visibility> for details).  Therefore you are
+recommended to always add the above flags when using Google Test as a
+shared library.  Otherwise a future release of Google Test may break
+your build script.
+
+### Avoiding Macro Name Clashes ###
+
+In C++, macros don't obey namespaces.  Therefore two libraries that
+both define a macro of the same name will clash if you `#include` both
+definitions.  In case a Google Test macro clashes with another
+library, you can force Google Test to rename its macro to avoid the
+conflict.
+
+Specifically, if both Google Test and some other code define macro
+FOO, you can add
+
+    -DGTEST_DONT_DEFINE_FOO=1
+
+to the compiler flags to tell Google Test to change the macro's name
+from `FOO` to `GTEST_FOO`.  Currently `FOO` can be `FAIL`, `SUCCEED`,
+or `TEST`.  For example, with `-DGTEST_DONT_DEFINE_TEST=1`, you'll
+need to write
+
+    GTEST_TEST(SomeTest, DoesThis) { ... }
+
+instead of
+
+    TEST(SomeTest, DoesThis) { ... }
+
+in order to define a test.
+
+## Developing Google Test ##
+
+This section discusses how to make your own changes to Google Test.
+
+### Testing Google Test Itself ###
+
+To make sure your changes work as intended and don't break existing
+functionality, you'll want to compile and run Google Test's own tests.
+For that you can use CMake:
+
+    mkdir mybuild
+    cd mybuild
+    cmake -Dgtest_build_tests=ON ${GTEST_DIR}
+
+Make sure you have Python installed, as some of Google Test's tests
+are written in Python.  If the cmake command complains about not being
+able to find Python (`Could NOT find PythonInterp (missing:
+PYTHON_EXECUTABLE)`), try telling it explicitly where your Python
+executable can be found:
+
+    cmake -DPYTHON_EXECUTABLE=path/to/python -Dgtest_build_tests=ON ${GTEST_DIR}
+
+Next, you can build Google Test and all of its own tests.  On \*nix,
+this is usually done by 'make'.  To run the tests, do
+
+    make test
+
+All tests should pass.
+
+Normally you don't need to worry about regenerating the source files,
+unless you need to modify them.  In that case, you should modify the
+corresponding .pump files instead and run the pump.py Python script to
+regenerate them.  You can find pump.py in the [scripts/](scripts/) directory.
+Read the [Pump manual](docs/PumpManual.md) for how to use it.
diff --git a/third_party/aom/third_party/googletest/src/googletest/cmake/internal_utils.cmake b/third_party/aom/third_party/googletest/src/googletest/cmake/internal_utils.cmake
new file mode 100644
index 0000000000..777b91ed4b
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/cmake/internal_utils.cmake
@@ -0,0 +1,254 @@
+# Defines functions and macros useful for building Google Test and
+# Google Mock.
+#
+# Note:
+#
+# - This file will be run twice when building Google Mock (once via
+#   Google Test's CMakeLists.txt, and once via Google Mock's).
+#   Therefore it shouldn't have any side effects other than defining
+#   the functions and macros.
+#
+# - The functions/macros defined in this file may depend on Google
+#   Test and Google Mock's option() definitions, and thus must be
+#   called *after* the options have been defined.
+
+# Tweaks CMake's default compiler/linker settings to suit Google Test's needs.
+#
+# This must be a macro(), as inside a function string() can only
+# update variables in the function scope.
+macro(fix_default_compiler_settings_)
+  if (MSVC)
+    # For MSVC, CMake sets certain flags to defaults we want to override.
+    # This replacement code is taken from sample in the CMake Wiki at
+    # http://www.cmake.org/Wiki/CMake_FAQ#Dynamic_Replace.
+    foreach (flag_var
+             CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
+             CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
+      if (NOT BUILD_SHARED_LIBS AND NOT gtest_force_shared_crt)
+        # When Google Test is built as a shared library, it should also use
+        # shared runtime libraries.  Otherwise, it may end up with multiple
+        # copies of runtime library data in different modules, resulting in
+        # hard-to-find crashes. When it is built as a static library, it is
+        # preferable to use CRT as static libraries, as we don't have to rely
+        # on CRT DLLs being available. CMake always defaults to using shared
+        # CRT libraries, so we override that default here.
+        string(REPLACE "/MD" "-MT" ${flag_var} "${${flag_var}}")
+      endif()
+
+      # We prefer more strict warning checking for building Google Test.
+      # Replaces /W3 with /W4 in defaults.
+      string(REPLACE "/W3" "/W4" ${flag_var} "${${flag_var}}")
+    endforeach()
+  endif()
+endmacro()
+
+# Defines the compiler/linker flags used to build Google Test and
+# Google Mock.  You can tweak these definitions to suit your need.  A
+# variable's value is empty before it's explicitly assigned to.
+macro(config_compiler_and_linker)
+  if (NOT gtest_disable_pthreads)
+    # Defines CMAKE_USE_PTHREADS_INIT and CMAKE_THREAD_LIBS_INIT.
+    find_package(Threads)
+  endif()
+
+  fix_default_compiler_settings_()
+  if (MSVC)
+    # Newlines inside flags variables break CMake's NMake generator.
+    # TODO(vladl@google.com): Add -RTCs and -RTCu to debug builds.
+    set(cxx_base_flags "-GS -W4 -WX -wd4251 -wd4275 -nologo -J -Zi")
+    if (MSVC_VERSION LESS 1400)  # 1400 is Visual Studio 2005
+      # Suppress spurious warnings MSVC 7.1 sometimes issues.
+      # Forcing value to bool.
+      set(cxx_base_flags "${cxx_base_flags} -wd4800")
+      # Copy constructor and assignment operator could not be generated.
+      set(cxx_base_flags "${cxx_base_flags} -wd4511 -wd4512")
+      # Compatibility warnings not applicable to Google Test.
+      # Resolved overload was found by argument-dependent lookup.
+      set(cxx_base_flags "${cxx_base_flags} -wd4675")
+    endif()
+    if (MSVC_VERSION LESS 1500)  # 1500 is Visual Studio 2008
+      # Conditional expression is constant.
+      # When compiling with /W4, we get several instances of C4127
+      # (Conditional expression is constant). In our code, we disable that
+      # warning on a case-by-case basis. However, on Visual Studio 2005,
+      # the warning fires on std::list. Therefore on that compiler and earlier,
+      # we disable the warning project-wide.
+      set(cxx_base_flags "${cxx_base_flags} -wd4127")
+    endif()
+    if (NOT (MSVC_VERSION LESS 1700))  # 1700 is Visual Studio 2012.
+      # Suppress "unreachable code" warning on VS 2012 and later.
+      # http://stackoverflow.com/questions/3232669 explains the issue.
+      set(cxx_base_flags "${cxx_base_flags} -wd4702")
+    endif()
+    if (NOT (MSVC_VERSION GREATER 1900))  # 1900 is Visual Studio 2015
+      # BigObj required for tests.
+      set(cxx_base_flags "${cxx_base_flags} -bigobj")
+    endif()
+
+    set(cxx_base_flags "${cxx_base_flags} -D_UNICODE -DUNICODE -DWIN32 -D_WIN32")
+    set(cxx_base_flags "${cxx_base_flags} -DSTRICT -DWIN32_LEAN_AND_MEAN")
+    set(cxx_exception_flags "-EHsc -D_HAS_EXCEPTIONS=1")
+    set(cxx_no_exception_flags "-D_HAS_EXCEPTIONS=0")
+    set(cxx_no_rtti_flags "-GR-")
+  elseif (CMAKE_COMPILER_IS_GNUCXX)
+    set(cxx_base_flags "-Wall -Wshadow")
+    set(cxx_exception_flags "-fexceptions")
+    set(cxx_no_exception_flags "-fno-exceptions")
+    # Until version 4.3.2, GCC doesn't define a macro to indicate
+    # whether RTTI is enabled.  Therefore we define GTEST_HAS_RTTI
+    # explicitly.
+    set(cxx_no_rtti_flags "-fno-rtti -DGTEST_HAS_RTTI=0")
+    set(cxx_strict_flags
+      "-Wextra -Wno-unused-parameter -Wno-missing-field-initializers")
+  elseif (CMAKE_CXX_COMPILER_ID STREQUAL "SunPro")
+    set(cxx_exception_flags "-features=except")
+    # Sun Pro doesn't provide macros to indicate whether exceptions and
+    # RTTI are enabled, so we define GTEST_HAS_* explicitly.
+    set(cxx_no_exception_flags "-features=no%except -DGTEST_HAS_EXCEPTIONS=0")
+    set(cxx_no_rtti_flags "-features=no%rtti -DGTEST_HAS_RTTI=0")
+  elseif (CMAKE_CXX_COMPILER_ID STREQUAL "VisualAge" OR
+      CMAKE_CXX_COMPILER_ID STREQUAL "XL")
+    # CMake 2.8 changes Visual Age's compiler ID to "XL".
+    set(cxx_exception_flags "-qeh")
+    set(cxx_no_exception_flags "-qnoeh")
+    # Until version 9.0, Visual Age doesn't define a macro to indicate
+    # whether RTTI is enabled.  Therefore we define GTEST_HAS_RTTI
+    # explicitly.
+    set(cxx_no_rtti_flags "-qnortti -DGTEST_HAS_RTTI=0")
+  elseif (CMAKE_CXX_COMPILER_ID STREQUAL "HP")
+    set(cxx_base_flags "-AA -mt")
+    set(cxx_exception_flags "-DGTEST_HAS_EXCEPTIONS=1")
+    set(cxx_no_exception_flags "+noeh -DGTEST_HAS_EXCEPTIONS=0")
+    # RTTI can not be disabled in HP aCC compiler.
+    set(cxx_no_rtti_flags "")
+  endif()
+
+  if (CMAKE_USE_PTHREADS_INIT)  # The pthreads library is available and allowed.
+    set(cxx_base_flags "${cxx_base_flags} -DGTEST_HAS_PTHREAD=1")
+  else()
+    set(cxx_base_flags "${cxx_base_flags} -DGTEST_HAS_PTHREAD=0")
+  endif()
+
+  # For building gtest's own tests and samples.
+  set(cxx_exception "${CMAKE_CXX_FLAGS} ${cxx_base_flags} ${cxx_exception_flags}")
+  set(cxx_no_exception
+    "${CMAKE_CXX_FLAGS} ${cxx_base_flags} ${cxx_no_exception_flags}")
+  set(cxx_default "${cxx_exception}")
+  set(cxx_no_rtti "${cxx_default} ${cxx_no_rtti_flags}")
+  set(cxx_use_own_tuple "${cxx_default} -DGTEST_USE_OWN_TR1_TUPLE=1")
+
+  # For building the gtest libraries.
+  set(cxx_strict "${cxx_default} ${cxx_strict_flags}")
+endmacro()
+
+# Defines the gtest & gtest_main libraries.  User tests should link
+# with one of them.
+function(cxx_library_with_type name type cxx_flags)
+  # type can be either STATIC or SHARED to denote a static or shared library.
+  # ARGN refers to additional arguments after 'cxx_flags'.
+  add_library(${name} ${type} ${ARGN})
+  set_target_properties(${name}
+    PROPERTIES
+    COMPILE_FLAGS "${cxx_flags}")
+  if (BUILD_SHARED_LIBS OR type STREQUAL "SHARED")
+    set_target_properties(${name}
+      PROPERTIES
+      COMPILE_DEFINITIONS "GTEST_CREATE_SHARED_LIBRARY=1")
+  endif()
+  if (CMAKE_USE_PTHREADS_INIT)
+    target_link_libraries(${name} ${CMAKE_THREAD_LIBS_INIT})
+  endif()
+endfunction()
+
+########################################################################
+#
+# Helper functions for creating build targets.
+
+function(cxx_shared_library name cxx_flags)
+  cxx_library_with_type(${name} SHARED "${cxx_flags}" ${ARGN})
+endfunction()
+
+function(cxx_library name cxx_flags)
+  cxx_library_with_type(${name} "" "${cxx_flags}" ${ARGN})
+endfunction()
+
+# cxx_executable_with_flags(name cxx_flags libs srcs...)
+#
+# creates a named C++ executable that depends on the given libraries and
+# is built from the given source files with the given compiler flags.
+function(cxx_executable_with_flags name cxx_flags libs)
+  add_executable(${name} ${ARGN})
+  if (cxx_flags)
+    set_target_properties(${name}
+      PROPERTIES
+      COMPILE_FLAGS "${cxx_flags}")
+  endif()
+  if (BUILD_SHARED_LIBS)
+    set_target_properties(${name}
+      PROPERTIES
+      COMPILE_DEFINITIONS "GTEST_LINKED_AS_SHARED_LIBRARY=1")
+  endif()
+  # To support mixing linking in static and dynamic libraries, link each
+  # library in with an extra call to target_link_libraries.
+  foreach (lib "${libs}")
+    target_link_libraries(${name} ${lib})
+  endforeach()
+endfunction()
+
+# cxx_executable(name dir lib srcs...)
+#
+# creates a named target that depends on the given libs and is built
+# from the given source files.  dir/name.cc is implicitly included in
+# the source file list.
+function(cxx_executable name dir libs)
+  cxx_executable_with_flags(
+    ${name} "${cxx_default}" "${libs}" "${dir}/${name}.cc" ${ARGN})
+endfunction()
+
+# Sets PYTHONINTERP_FOUND and PYTHON_EXECUTABLE.
+find_package(PythonInterp)
+
+# cxx_test_with_flags(name cxx_flags libs srcs...)
+#
+# creates a named C++ test that depends on the given libs and is built
+# from the given source files with the given compiler flags.
+function(cxx_test_with_flags name cxx_flags libs)
+  cxx_executable_with_flags(${name} "${cxx_flags}" "${libs}" ${ARGN})
+  add_test(${name} ${name})
+endfunction()
+
+# cxx_test(name libs srcs...)
+#
+# creates a named test target that depends on the given libs and is
+# built from the given source files.  Unlike cxx_test_with_flags,
+# test/name.cc is already implicitly included in the source file list.
+function(cxx_test name libs)
+  cxx_test_with_flags("${name}" "${cxx_default}" "${libs}"
+    "test/${name}.cc" ${ARGN})
+endfunction()
+
+# py_test(name)
+#
+# creates a Python test with the given name whose main module is in
+# test/name.py.  It does nothing if Python is not installed.
+function(py_test name)
+  # We are not supporting Python tests on Linux yet as they consider
+  # all Linux environments to be google3 and try to use google3 features.
+  if (PYTHONINTERP_FOUND)
+    # ${CMAKE_BINARY_DIR} is known at configuration time, so we can
+    # directly bind it from cmake. ${CTEST_CONFIGURATION_TYPE} is known
+    # only at ctest runtime (by calling ctest -c <Configuration>), so
+    # we have to escape $ to delay variable substitution here.
+    if (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION} GREATER 3.1)
+      add_test(
+        NAME ${name}
+        COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/test/${name}.py
+            --build_dir=${CMAKE_CURRENT_BINARY_DIR}/$<CONFIGURATION>)
+    else (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION} GREATER 3.1)
+      add_test(
+        ${name}
+        ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/test/${name}.py
+          --build_dir=${CMAKE_CURRENT_BINARY_DIR}/\${CTEST_CONFIGURATION_TYPE})
+    endif (${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION} GREATER 3.1)
+  endif()
+endfunction()
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-death-test.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-death-test.h
new file mode 100644
index 0000000000..957a69c6a9
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-death-test.h
@@ -0,0 +1,294 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines the public API for death tests.  It is
+// #included by gtest.h so a user doesn't need to include this
+// directly.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_DEATH_TEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_DEATH_TEST_H_
+
+#include "gtest/internal/gtest-death-test-internal.h"
+
+namespace testing {
+
+// This flag controls the style of death tests.  Valid values are "threadsafe",
+// meaning that the death test child process will re-execute the test binary
+// from the start, running only a single death test, or "fast",
+// meaning that the child process will execute the test logic immediately
+// after forking.
+GTEST_DECLARE_string_(death_test_style);
+
+#if GTEST_HAS_DEATH_TEST
+
+namespace internal {
+
+// Returns a Boolean value indicating whether the caller is currently
+// executing in the context of the death test child process.  Tools such as
+// Valgrind heap checkers may need this to modify their behavior in death
+// tests.  IMPORTANT: This is an internal utility.  Using it may break the
+// implementation of death tests.  User code MUST NOT use it.
+GTEST_API_ bool InDeathTestChild();
+
+}  // namespace internal
+
+// The following macros are useful for writing death tests.
+
+// Here's what happens when an ASSERT_DEATH* or EXPECT_DEATH* is
+// executed:
+//
+//   1. It generates a warning if there is more than one active
+//   thread.  This is because it's safe to fork() or clone() only
+//   when there is a single thread.
+//
+//   2. The parent process clone()s a sub-process and runs the death
+//   test in it; the sub-process exits with code 0 at the end of the
+//   death test, if it hasn't exited already.
+//
+//   3. The parent process waits for the sub-process to terminate.
+//
+//   4. The parent process checks the exit code and error message of
+//   the sub-process.
+//
+// Examples:
+//
+//   ASSERT_DEATH(server.SendMessage(56, "Hello"), "Invalid port number");
+//   for (int i = 0; i < 5; i++) {
+//     EXPECT_DEATH(server.ProcessRequest(i),
+//                  "Invalid request .* in ProcessRequest()")
+//                  << "Failed to die on request " << i;
+//   }
+//
+//   ASSERT_EXIT(server.ExitNow(), ::testing::ExitedWithCode(0), "Exiting");
+//
+//   bool KilledBySIGHUP(int exit_code) {
+//     return WIFSIGNALED(exit_code) && WTERMSIG(exit_code) == SIGHUP;
+//   }
+//
+//   ASSERT_EXIT(client.HangUpServer(), KilledBySIGHUP, "Hanging up!");
+//
+// On the regular expressions used in death tests:
+//
+//   On POSIX-compliant systems (*nix), we use the <regex.h> library,
+//   which uses the POSIX extended regex syntax.
+//
+//   On other platforms (e.g. Windows), we only support a simple regex
+//   syntax implemented as part of Google Test.  This limited
+//   implementation should be enough most of the time when writing
+//   death tests; though it lacks many features you can find in PCRE
+//   or POSIX extended regex syntax.  For example, we don't support
+//   union ("x|y"), grouping ("(xy)"), brackets ("[xy]"), and
+//   repetition count ("x{5,7}"), among others.
+//
+//   Below is the syntax that we do support.  We chose it to be a
+//   subset of both PCRE and POSIX extended regex, so it's easy to
+//   learn wherever you come from.  In the following: 'A' denotes a
+//   literal character, period (.), or a single \\ escape sequence;
+//   'x' and 'y' denote regular expressions; 'm' and 'n' are for
+//   natural numbers.
+//
+//     c     matches any literal character c
+//     \\d   matches any decimal digit
+//     \\D   matches any character that's not a decimal digit
+//     \\f   matches \f
+//     \\n   matches \n
+//     \\r   matches \r
+//     \\s   matches any ASCII whitespace, including \n
+//     \\S   matches any character that's not a whitespace
+//     \\t   matches \t
+//     \\v   matches \v
+//     \\w   matches any letter, _, or decimal digit
+//     \\W   matches any character that \\w doesn't match
+//     \\c   matches any literal character c, which must be a punctuation
+//     .     matches any single character except \n
+//     A?    matches 0 or 1 occurrences of A
+//     A*    matches 0 or many occurrences of A
+//     A+    matches 1 or many occurrences of A
+//     ^     matches the beginning of a string (not that of each line)
+//     $     matches the end of a string (not that of each line)
+//     xy    matches x followed by y
+//
+//   If you accidentally use PCRE or POSIX extended regex features
+//   not implemented by us, you will get a run-time failure.  In that
+//   case, please try to rewrite your regular expression within the
+//   above syntax.
+//
+//   This implementation is *not* meant to be as highly tuned or robust
+//   as a compiled regex library, but should perform well enough for a
+//   death test, which already incurs significant overhead by launching
+//   a child process.
+//
+// Known caveats:
+//
+//   A "threadsafe" style death test obtains the path to the test
+//   program from argv[0] and re-executes it in the sub-process.  For
+//   simplicity, the current implementation doesn't search the PATH
+//   when launching the sub-process.  This means that the user must
+//   invoke the test program via a path that contains at least one
+//   path separator (e.g. path/to/foo_test and
+//   /absolute/path/to/bar_test are fine, but foo_test is not).  This
+//   is rarely a problem as people usually don't put the test binary
+//   directory in PATH.
+//
+// TODO(wan@google.com): make thread-safe death tests search the PATH.
+
+// Asserts that a given statement causes the program to exit, with an
+// integer exit status that satisfies predicate, and emitting error output
+// that matches regex.
+# define ASSERT_EXIT(statement, predicate, regex) \
+    GTEST_DEATH_TEST_(statement, predicate, regex, GTEST_FATAL_FAILURE_)
+
+// Like ASSERT_EXIT, but continues on to successive tests in the
+// test case, if any:
+# define EXPECT_EXIT(statement, predicate, regex) \
+    GTEST_DEATH_TEST_(statement, predicate, regex, GTEST_NONFATAL_FAILURE_)
+
+// Asserts that a given statement causes the program to exit, either by
+// explicitly exiting with a nonzero exit code or being killed by a
+// signal, and emitting error output that matches regex.
+# define ASSERT_DEATH(statement, regex) \
+    ASSERT_EXIT(statement, ::testing::internal::ExitedUnsuccessfully, regex)
+
+// Like ASSERT_DEATH, but continues on to successive tests in the
+// test case, if any:
+# define EXPECT_DEATH(statement, regex) \
+    EXPECT_EXIT(statement, ::testing::internal::ExitedUnsuccessfully, regex)
+
+// Two predicate classes that can be used in {ASSERT,EXPECT}_EXIT*:
+
+// Tests that an exit code describes a normal exit with a given exit code.
+class GTEST_API_ ExitedWithCode {
+ public:
+  explicit ExitedWithCode(int exit_code);
+  bool operator()(int exit_status) const;
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ExitedWithCode& other);
+
+  const int exit_code_;
+};
+
+# if !GTEST_OS_WINDOWS
+// Tests that an exit code describes an exit due to termination by a
+// given signal.
+class GTEST_API_ KilledBySignal {
+ public:
+  explicit KilledBySignal(int signum);
+  bool operator()(int exit_status) const;
+ private:
+  const int signum_;
+};
+# endif  // !GTEST_OS_WINDOWS
+
+// EXPECT_DEBUG_DEATH asserts that the given statements die in debug mode.
+// The death testing framework causes this to have interesting semantics,
+// since the sideeffects of the call are only visible in opt mode, and not
+// in debug mode.
+//
+// In practice, this can be used to test functions that utilize the
+// LOG(DFATAL) macro using the following style:
+//
+// int DieInDebugOr12(int* sideeffect) {
+//   if (sideeffect) {
+//     *sideeffect = 12;
+//   }
+//   LOG(DFATAL) << "death";
+//   return 12;
+// }
+//
+// TEST(TestCase, TestDieOr12WorksInDgbAndOpt) {
+//   int sideeffect = 0;
+//   // Only asserts in dbg.
+//   EXPECT_DEBUG_DEATH(DieInDebugOr12(&sideeffect), "death");
+//
+// #ifdef NDEBUG
+//   // opt-mode has sideeffect visible.
+//   EXPECT_EQ(12, sideeffect);
+// #else
+//   // dbg-mode no visible sideeffect.
+//   EXPECT_EQ(0, sideeffect);
+// #endif
+// }
+//
+// This will assert that DieInDebugReturn12InOpt() crashes in debug
+// mode, usually due to a DCHECK or LOG(DFATAL), but returns the
+// appropriate fallback value (12 in this case) in opt mode. If you
+// need to test that a function has appropriate side-effects in opt
+// mode, include assertions against the side-effects.  A general
+// pattern for this is:
+//
+// EXPECT_DEBUG_DEATH({
+//   // Side-effects here will have an effect after this statement in
+//   // opt mode, but none in debug mode.
+//   EXPECT_EQ(12, DieInDebugOr12(&sideeffect));
+// }, "death");
+//
+# ifdef NDEBUG
+
+#  define EXPECT_DEBUG_DEATH(statement, regex) \
+  GTEST_EXECUTE_STATEMENT_(statement, regex)
+
+#  define ASSERT_DEBUG_DEATH(statement, regex) \
+  GTEST_EXECUTE_STATEMENT_(statement, regex)
+
+# else
+
+#  define EXPECT_DEBUG_DEATH(statement, regex) \
+  EXPECT_DEATH(statement, regex)
+
+#  define ASSERT_DEBUG_DEATH(statement, regex) \
+  ASSERT_DEATH(statement, regex)
+
+# endif  // NDEBUG for EXPECT_DEBUG_DEATH
+#endif  // GTEST_HAS_DEATH_TEST
+
+// EXPECT_DEATH_IF_SUPPORTED(statement, regex) and
+// ASSERT_DEATH_IF_SUPPORTED(statement, regex) expand to real death tests if
+// death tests are supported; otherwise they just issue a warning.  This is
+// useful when you are combining death test assertions with normal test
+// assertions in one test.
+#if GTEST_HAS_DEATH_TEST
+# define EXPECT_DEATH_IF_SUPPORTED(statement, regex) \
+    EXPECT_DEATH(statement, regex)
+# define ASSERT_DEATH_IF_SUPPORTED(statement, regex) \
+    ASSERT_DEATH(statement, regex)
+#else
+# define EXPECT_DEATH_IF_SUPPORTED(statement, regex) \
+    GTEST_UNSUPPORTED_DEATH_TEST_(statement, regex, )
+# define ASSERT_DEATH_IF_SUPPORTED(statement, regex) \
+    GTEST_UNSUPPORTED_DEATH_TEST_(statement, regex, return)
+#endif
+
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_DEATH_TEST_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-message.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-message.h
new file mode 100644
index 0000000000..fe879bca79
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-message.h
@@ -0,0 +1,250 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines the Message class.
+//
+// IMPORTANT NOTE: Due to limitation of the C++ language, we have to
+// leave some internal implementation details in this header file.
+// They are clearly marked by comments like this:
+//
+//   // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+//
+// Such code is NOT meant to be used by a user directly, and is subject
+// to CHANGE WITHOUT NOTICE.  Therefore DO NOT DEPEND ON IT in a user
+// program!
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_MESSAGE_H_
+#define GTEST_INCLUDE_GTEST_GTEST_MESSAGE_H_
+
+#include <limits>
+
+#include "gtest/internal/gtest-port.h"
+
+// Ensures that there is at least one operator<< in the global namespace.
+// See Message& operator<<(...) below for why.
+void operator<<(const testing::internal::Secret&, int);
+
+namespace testing {
+
+// The Message class works like an ostream repeater.
+//
+// Typical usage:
+//
+//   1. You stream a bunch of values to a Message object.
+//      It will remember the text in a stringstream.
+//   2. Then you stream the Message object to an ostream.
+//      This causes the text in the Message to be streamed
+//      to the ostream.
+//
+// For example;
+//
+//   testing::Message foo;
+//   foo << 1 << " != " << 2;
+//   std::cout << foo;
+//
+// will print "1 != 2".
+//
+// Message is not intended to be inherited from.  In particular, its
+// destructor is not virtual.
+//
+// Note that stringstream behaves differently in gcc and in MSVC.  You
+// can stream a NULL char pointer to it in the former, but not in the
+// latter (it causes an access violation if you do).  The Message
+// class hides this difference by treating a NULL char pointer as
+// "(null)".
+class GTEST_API_ Message {
+ private:
+  // The type of basic IO manipulators (endl, ends, and flush) for
+  // narrow streams.
+  typedef std::ostream& (*BasicNarrowIoManip)(std::ostream&);
+
+ public:
+  // Constructs an empty Message.
+  Message();
+
+  // Copy constructor.
+  Message(const Message& msg) : ss_(new ::std::stringstream) {  // NOLINT
+    *ss_ << msg.GetString();
+  }
+
+  // Constructs a Message from a C-string.
+  explicit Message(const char* str) : ss_(new ::std::stringstream) {
+    *ss_ << str;
+  }
+
+#if GTEST_OS_SYMBIAN
+  // Streams a value (either a pointer or not) to this object.
+  template <typename T>
+  inline Message& operator <<(const T& value) {
+    StreamHelper(typename internal::is_pointer<T>::type(), value);
+    return *this;
+  }
+#else
+  // Streams a non-pointer value to this object.
+  template <typename T>
+  inline Message& operator <<(const T& val) {
+    // Some libraries overload << for STL containers.  These
+    // overloads are defined in the global namespace instead of ::std.
+    //
+    // C++'s symbol lookup rule (i.e. Koenig lookup) says that these
+    // overloads are visible in either the std namespace or the global
+    // namespace, but not other namespaces, including the testing
+    // namespace which Google Test's Message class is in.
+    //
+    // To allow STL containers (and other types that has a << operator
+    // defined in the global namespace) to be used in Google Test
+    // assertions, testing::Message must access the custom << operator
+    // from the global namespace.  With this using declaration,
+    // overloads of << defined in the global namespace and those
+    // visible via Koenig lookup are both exposed in this function.
+    using ::operator <<;
+    *ss_ << val;
+    return *this;
+  }
+
+  // Streams a pointer value to this object.
+  //
+  // This function is an overload of the previous one.  When you
+  // stream a pointer to a Message, this definition will be used as it
+  // is more specialized.  (The C++ Standard, section
+  // [temp.func.order].)  If you stream a non-pointer, then the
+  // previous definition will be used.
+  //
+  // The reason for this overload is that streaming a NULL pointer to
+  // ostream is undefined behavior.  Depending on the compiler, you
+  // may get "0", "(nil)", "(null)", or an access violation.  To
+  // ensure consistent result across compilers, we always treat NULL
+  // as "(null)".
+  template <typename T>
+  inline Message& operator <<(T* const& pointer) {  // NOLINT
+    if (pointer == NULL) {
+      *ss_ << "(null)";
+    } else {
+      *ss_ << pointer;
+    }
+    return *this;
+  }
+#endif  // GTEST_OS_SYMBIAN
+
+  // Since the basic IO manipulators are overloaded for both narrow
+  // and wide streams, we have to provide this specialized definition
+  // of operator <<, even though its body is the same as the
+  // templatized version above.  Without this definition, streaming
+  // endl or other basic IO manipulators to Message will confuse the
+  // compiler.
+  Message& operator <<(BasicNarrowIoManip val) {
+    *ss_ << val;
+    return *this;
+  }
+
+  // Instead of 1/0, we want to see true/false for bool values.
+  Message& operator <<(bool b) {
+    return *this << (b ? "true" : "false");
+  }
+
+  // These two overloads allow streaming a wide C string to a Message
+  // using the UTF-8 encoding.
+  Message& operator <<(const wchar_t* wide_c_str);
+  Message& operator <<(wchar_t* wide_c_str);
+
+#if GTEST_HAS_STD_WSTRING
+  // Converts the given wide string to a narrow string using the UTF-8
+  // encoding, and streams the result to this Message object.
+  Message& operator <<(const ::std::wstring& wstr);
+#endif  // GTEST_HAS_STD_WSTRING
+
+#if GTEST_HAS_GLOBAL_WSTRING
+  // Converts the given wide string to a narrow string using the UTF-8
+  // encoding, and streams the result to this Message object.
+  Message& operator <<(const ::wstring& wstr);
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+  // Gets the text streamed to this object so far as an std::string.
+  // Each '\0' character in the buffer is replaced with "\\0".
+  //
+  // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+  std::string GetString() const;
+
+ private:
+
+#if GTEST_OS_SYMBIAN
+  // These are needed as the Nokia Symbian Compiler cannot decide between
+  // const T& and const T* in a function template. The Nokia compiler _can_
+  // decide between class template specializations for T and T*, so a
+  // tr1::type_traits-like is_pointer works, and we can overload on that.
+  template <typename T>
+  inline void StreamHelper(internal::true_type /*is_pointer*/, T* pointer) {
+    if (pointer == NULL) {
+      *ss_ << "(null)";
+    } else {
+      *ss_ << pointer;
+    }
+  }
+  template <typename T>
+  inline void StreamHelper(internal::false_type /*is_pointer*/,
+                           const T& value) {
+    // See the comments in Message& operator <<(const T&) above for why
+    // we need this using statement.
+    using ::operator <<;
+    *ss_ << value;
+  }
+#endif  // GTEST_OS_SYMBIAN
+
+  // We'll hold the text streamed to this object here.
+  const internal::scoped_ptr< ::std::stringstream> ss_;
+
+  // We declare (but don't implement) this to prevent the compiler
+  // from implementing the assignment operator.
+  void operator=(const Message&);
+};
+
+// Streams a Message to an ostream.
+inline std::ostream& operator <<(std::ostream& os, const Message& sb) {
+  return os << sb.GetString();
+}
+
+namespace internal {
+
+// Converts a streamable value to an std::string.  A NULL pointer is
+// converted to "(null)".  When the input value is a ::string,
+// ::std::string, ::wstring, or ::std::wstring object, each NUL
+// character in it is replaced with "\\0".
+template <typename T>
+std::string StreamableToString(const T& streamable) {
+  return (Message() << streamable).GetString();
+}
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_MESSAGE_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-param-test.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-param-test.h
new file mode 100644
index 0000000000..038f9ba79e
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-param-test.h
@@ -0,0 +1,1444 @@
+// This file was GENERATED by command:
+//     pump.py gtest-param-test.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: vladl@google.com (Vlad Losev)
+//
+// Macros and functions for implementing parameterized tests
+// in Google C++ Testing Framework (Google Test)
+//
+// This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
+
+
+// Value-parameterized tests allow you to test your code with different
+// parameters without writing multiple copies of the same test.
+//
+// Here is how you use value-parameterized tests:
+
+#if 0
+
+// To write value-parameterized tests, first you should define a fixture
+// class. It is usually derived from testing::TestWithParam<T> (see below for
+// another inheritance scheme that's sometimes useful in more complicated
+// class hierarchies), where the type of your parameter values.
+// TestWithParam<T> is itself derived from testing::Test. T can be any
+// copyable type. If it's a raw pointer, you are responsible for managing the
+// lifespan of the pointed values.
+
+class FooTest : public ::testing::TestWithParam<const char*> {
+  // You can implement all the usual class fixture members here.
+};
+
+// Then, use the TEST_P macro to define as many parameterized tests
+// for this fixture as you want. The _P suffix is for "parameterized"
+// or "pattern", whichever you prefer to think.
+
+TEST_P(FooTest, DoesBlah) {
+  // Inside a test, access the test parameter with the GetParam() method
+  // of the TestWithParam<T> class:
+  EXPECT_TRUE(foo.Blah(GetParam()));
+  ...
+}
+
+TEST_P(FooTest, HasBlahBlah) {
+  ...
+}
+
+// Finally, you can use INSTANTIATE_TEST_CASE_P to instantiate the test
+// case with any set of parameters you want. Google Test defines a number
+// of functions for generating test parameters. They return what we call
+// (surprise!) parameter generators. Here is a  summary of them, which
+// are all in the testing namespace:
+//
+//
+//  Range(begin, end [, step]) - Yields values {begin, begin+step,
+//                               begin+step+step, ...}. The values do not
+//                               include end. step defaults to 1.
+//  Values(v1, v2, ..., vN)    - Yields values {v1, v2, ..., vN}.
+//  ValuesIn(container)        - Yields values from a C-style array, an STL
+//  ValuesIn(begin,end)          container, or an iterator range [begin, end).
+//  Bool()                     - Yields sequence {false, true}.
+//  Combine(g1, g2, ..., gN)   - Yields all combinations (the Cartesian product
+//                               for the math savvy) of the values generated
+//                               by the N generators.
+//
+// For more details, see comments at the definitions of these functions below
+// in this file.
+//
+// The following statement will instantiate tests from the FooTest test case
+// each with parameter values "meeny", "miny", and "moe".
+
+INSTANTIATE_TEST_CASE_P(InstantiationName,
+                        FooTest,
+                        Values("meeny", "miny", "moe"));
+
+// To distinguish different instances of the pattern, (yes, you
+// can instantiate it more then once) the first argument to the
+// INSTANTIATE_TEST_CASE_P macro is a prefix that will be added to the
+// actual test case name. Remember to pick unique prefixes for different
+// instantiations. The tests from the instantiation above will have
+// these names:
+//
+//    * InstantiationName/FooTest.DoesBlah/0 for "meeny"
+//    * InstantiationName/FooTest.DoesBlah/1 for "miny"
+//    * InstantiationName/FooTest.DoesBlah/2 for "moe"
+//    * InstantiationName/FooTest.HasBlahBlah/0 for "meeny"
+//    * InstantiationName/FooTest.HasBlahBlah/1 for "miny"
+//    * InstantiationName/FooTest.HasBlahBlah/2 for "moe"
+//
+// You can use these names in --gtest_filter.
+//
+// This statement will instantiate all tests from FooTest again, each
+// with parameter values "cat" and "dog":
+
+const char* pets[] = {"cat", "dog"};
+INSTANTIATE_TEST_CASE_P(AnotherInstantiationName, FooTest, ValuesIn(pets));
+
+// The tests from the instantiation above will have these names:
+//
+//    * AnotherInstantiationName/FooTest.DoesBlah/0 for "cat"
+//    * AnotherInstantiationName/FooTest.DoesBlah/1 for "dog"
+//    * AnotherInstantiationName/FooTest.HasBlahBlah/0 for "cat"
+//    * AnotherInstantiationName/FooTest.HasBlahBlah/1 for "dog"
+//
+// Please note that INSTANTIATE_TEST_CASE_P will instantiate all tests
+// in the given test case, whether their definitions come before or
+// AFTER the INSTANTIATE_TEST_CASE_P statement.
+//
+// Please also note that generator expressions (including parameters to the
+// generators) are evaluated in InitGoogleTest(), after main() has started.
+// This allows the user on one hand, to adjust generator parameters in order
+// to dynamically determine a set of tests to run and on the other hand,
+// give the user a chance to inspect the generated tests with Google Test
+// reflection API before RUN_ALL_TESTS() is executed.
+//
+// You can see samples/sample7_unittest.cc and samples/sample8_unittest.cc
+// for more examples.
+//
+// In the future, we plan to publish the API for defining new parameter
+// generators. But for now this interface remains part of the internal
+// implementation and is subject to change.
+//
+//
+// A parameterized test fixture must be derived from testing::Test and from
+// testing::WithParamInterface<T>, where T is the type of the parameter
+// values. Inheriting from TestWithParam<T> satisfies that requirement because
+// TestWithParam<T> inherits from both Test and WithParamInterface. In more
+// complicated hierarchies, however, it is occasionally useful to inherit
+// separately from Test and WithParamInterface. For example:
+
+class BaseTest : public ::testing::Test {
+  // You can inherit all the usual members for a non-parameterized test
+  // fixture here.
+};
+
+class DerivedTest : public BaseTest, public ::testing::WithParamInterface<int> {
+  // The usual test fixture members go here too.
+};
+
+TEST_F(BaseTest, HasFoo) {
+  // This is an ordinary non-parameterized test.
+}
+
+TEST_P(DerivedTest, DoesBlah) {
+  // GetParam works just the same here as if you inherit from TestWithParam.
+  EXPECT_TRUE(foo.Blah(GetParam()));
+}
+
+#endif  // 0
+
+#include "gtest/internal/gtest-port.h"
+
+#if !GTEST_OS_SYMBIAN
+# include <utility>
+#endif
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+#include "gtest/internal/gtest-internal.h"
+#include "gtest/internal/gtest-param-util.h"
+#include "gtest/internal/gtest-param-util-generated.h"
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+
+// Functions producing parameter generators.
+//
+// Google Test uses these generators to produce parameters for value-
+// parameterized tests. When a parameterized test case is instantiated
+// with a particular generator, Google Test creates and runs tests
+// for each element in the sequence produced by the generator.
+//
+// In the following sample, tests from test case FooTest are instantiated
+// each three times with parameter values 3, 5, and 8:
+//
+// class FooTest : public TestWithParam<int> { ... };
+//
+// TEST_P(FooTest, TestThis) {
+// }
+// TEST_P(FooTest, TestThat) {
+// }
+// INSTANTIATE_TEST_CASE_P(TestSequence, FooTest, Values(3, 5, 8));
+//
+
+// Range() returns generators providing sequences of values in a range.
+//
+// Synopsis:
+// Range(start, end)
+//   - returns a generator producing a sequence of values {start, start+1,
+//     start+2, ..., }.
+// Range(start, end, step)
+//   - returns a generator producing a sequence of values {start, start+step,
+//     start+step+step, ..., }.
+// Notes:
+//   * The generated sequences never include end. For example, Range(1, 5)
+//     returns a generator producing a sequence {1, 2, 3, 4}. Range(1, 9, 2)
+//     returns a generator producing {1, 3, 5, 7}.
+//   * start and end must have the same type. That type may be any integral or
+//     floating-point type or a user defined type satisfying these conditions:
+//     * It must be assignable (have operator=() defined).
+//     * It must have operator+() (operator+(int-compatible type) for
+//       two-operand version).
+//     * It must have operator<() defined.
+//     Elements in the resulting sequences will also have that type.
+//   * Condition start < end must be satisfied in order for resulting sequences
+//     to contain any elements.
+//
+template <typename T, typename IncrementT>
+internal::ParamGenerator<T> Range(T start, T end, IncrementT step) {
+  return internal::ParamGenerator<T>(
+      new internal::RangeGenerator<T, IncrementT>(start, end, step));
+}
+
+template <typename T>
+internal::ParamGenerator<T> Range(T start, T end) {
+  return Range(start, end, 1);
+}
+
+// ValuesIn() function allows generation of tests with parameters coming from
+// a container.
+//
+// Synopsis:
+// ValuesIn(const T (&array)[N])
+//   - returns a generator producing sequences with elements from
+//     a C-style array.
+// ValuesIn(const Container& container)
+//   - returns a generator producing sequences with elements from
+//     an STL-style container.
+// ValuesIn(Iterator begin, Iterator end)
+//   - returns a generator producing sequences with elements from
+//     a range [begin, end) defined by a pair of STL-style iterators. These
+//     iterators can also be plain C pointers.
+//
+// Please note that ValuesIn copies the values from the containers
+// passed in and keeps them to generate tests in RUN_ALL_TESTS().
+//
+// Examples:
+//
+// This instantiates tests from test case StringTest
+// each with C-string values of "foo", "bar", and "baz":
+//
+// const char* strings[] = {"foo", "bar", "baz"};
+// INSTANTIATE_TEST_CASE_P(StringSequence, SrtingTest, ValuesIn(strings));
+//
+// This instantiates tests from test case StlStringTest
+// each with STL strings with values "a" and "b":
+//
+// ::std::vector< ::std::string> GetParameterStrings() {
+//   ::std::vector< ::std::string> v;
+//   v.push_back("a");
+//   v.push_back("b");
+//   return v;
+// }
+//
+// INSTANTIATE_TEST_CASE_P(CharSequence,
+//                         StlStringTest,
+//                         ValuesIn(GetParameterStrings()));
+//
+//
+// This will also instantiate tests from CharTest
+// each with parameter values 'a' and 'b':
+//
+// ::std::list<char> GetParameterChars() {
+//   ::std::list<char> list;
+//   list.push_back('a');
+//   list.push_back('b');
+//   return list;
+// }
+// ::std::list<char> l = GetParameterChars();
+// INSTANTIATE_TEST_CASE_P(CharSequence2,
+//                         CharTest,
+//                         ValuesIn(l.begin(), l.end()));
+//
+template <typename ForwardIterator>
+internal::ParamGenerator<
+  typename ::testing::internal::IteratorTraits<ForwardIterator>::value_type>
+ValuesIn(ForwardIterator begin, ForwardIterator end) {
+  typedef typename ::testing::internal::IteratorTraits<ForwardIterator>
+      ::value_type ParamType;
+  return internal::ParamGenerator<ParamType>(
+      new internal::ValuesInIteratorRangeGenerator<ParamType>(begin, end));
+}
+
+template <typename T, size_t N>
+internal::ParamGenerator<T> ValuesIn(const T (&array)[N]) {
+  return ValuesIn(array, array + N);
+}
+
+template <class Container>
+internal::ParamGenerator<typename Container::value_type> ValuesIn(
+    const Container& container) {
+  return ValuesIn(container.begin(), container.end());
+}
+
+// Values() allows generating tests from explicitly specified list of
+// parameters.
+//
+// Synopsis:
+// Values(T v1, T v2, ..., T vN)
+//   - returns a generator producing sequences with elements v1, v2, ..., vN.
+//
+// For example, this instantiates tests from test case BarTest each
+// with values "one", "two", and "three":
+//
+// INSTANTIATE_TEST_CASE_P(NumSequence, BarTest, Values("one", "two", "three"));
+//
+// This instantiates tests from test case BazTest each with values 1, 2, 3.5.
+// The exact type of values will depend on the type of parameter in BazTest.
+//
+// INSTANTIATE_TEST_CASE_P(FloatingNumbers, BazTest, Values(1, 2, 3.5));
+//
+// Currently, Values() supports from 1 to 50 parameters.
+//
+template <typename T1>
+internal::ValueArray1<T1> Values(T1 v1) {
+  return internal::ValueArray1<T1>(v1);
+}
+
+template <typename T1, typename T2>
+internal::ValueArray2<T1, T2> Values(T1 v1, T2 v2) {
+  return internal::ValueArray2<T1, T2>(v1, v2);
+}
+
+template <typename T1, typename T2, typename T3>
+internal::ValueArray3<T1, T2, T3> Values(T1 v1, T2 v2, T3 v3) {
+  return internal::ValueArray3<T1, T2, T3>(v1, v2, v3);
+}
+
+template <typename T1, typename T2, typename T3, typename T4>
+internal::ValueArray4<T1, T2, T3, T4> Values(T1 v1, T2 v2, T3 v3, T4 v4) {
+  return internal::ValueArray4<T1, T2, T3, T4>(v1, v2, v3, v4);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+internal::ValueArray5<T1, T2, T3, T4, T5> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5) {
+  return internal::ValueArray5<T1, T2, T3, T4, T5>(v1, v2, v3, v4, v5);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6>
+internal::ValueArray6<T1, T2, T3, T4, T5, T6> Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6) {
+  return internal::ValueArray6<T1, T2, T3, T4, T5, T6>(v1, v2, v3, v4, v5, v6);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7>
+internal::ValueArray7<T1, T2, T3, T4, T5, T6, T7> Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6, T7 v7) {
+  return internal::ValueArray7<T1, T2, T3, T4, T5, T6, T7>(v1, v2, v3, v4, v5,
+      v6, v7);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8>
+internal::ValueArray8<T1, T2, T3, T4, T5, T6, T7, T8> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8) {
+  return internal::ValueArray8<T1, T2, T3, T4, T5, T6, T7, T8>(v1, v2, v3, v4,
+      v5, v6, v7, v8);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9>
+internal::ValueArray9<T1, T2, T3, T4, T5, T6, T7, T8, T9> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9) {
+  return internal::ValueArray9<T1, T2, T3, T4, T5, T6, T7, T8, T9>(v1, v2, v3,
+      v4, v5, v6, v7, v8, v9);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10>
+internal::ValueArray10<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10> Values(T1 v1,
+    T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10) {
+  return internal::ValueArray10<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10>(v1,
+      v2, v3, v4, v5, v6, v7, v8, v9, v10);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11>
+internal::ValueArray11<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10,
+    T11> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11) {
+  return internal::ValueArray11<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10,
+      T11>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12>
+internal::ValueArray12<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+    T12> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12) {
+  return internal::ValueArray12<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13>
+internal::ValueArray13<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+    T13> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13) {
+  return internal::ValueArray13<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14>
+internal::ValueArray14<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14) {
+  return internal::ValueArray14<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+      v14);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15>
+internal::ValueArray15<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+    T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15) {
+  return internal::ValueArray15<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+      v13, v14, v15);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16>
+internal::ValueArray16<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16) {
+  return internal::ValueArray16<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+      v12, v13, v14, v15, v16);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17>
+internal::ValueArray17<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17) {
+  return internal::ValueArray17<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10,
+      v11, v12, v13, v14, v15, v16, v17);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18>
+internal::ValueArray18<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6,
+    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18) {
+  return internal::ValueArray18<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18>(v1, v2, v3, v4, v5, v6, v7, v8, v9,
+      v10, v11, v12, v13, v14, v15, v16, v17, v18);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19>
+internal::ValueArray19<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5,
+    T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14,
+    T15 v15, T16 v16, T17 v17, T18 v18, T19 v19) {
+  return internal::ValueArray19<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19>(v1, v2, v3, v4, v5, v6, v7, v8,
+      v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20>
+internal::ValueArray20<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20) {
+  return internal::ValueArray20<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20>(v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21>
+internal::ValueArray21<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21) {
+  return internal::ValueArray21<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21>(v1, v2, v3, v4, v5, v6,
+      v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22>
+internal::ValueArray22<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22> Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22) {
+  return internal::ValueArray22<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22>(v1, v2, v3, v4,
+      v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23>
+internal::ValueArray23<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23) {
+  return internal::ValueArray23<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23>(v1, v2, v3,
+      v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24>
+internal::ValueArray24<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24) {
+  return internal::ValueArray24<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24>(v1, v2,
+      v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18,
+      v19, v20, v21, v22, v23, v24);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25>
+internal::ValueArray25<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25> Values(T1 v1,
+    T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11,
+    T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19,
+    T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25) {
+  return internal::ValueArray25<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25>(v1,
+      v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17,
+      v18, v19, v20, v21, v22, v23, v24, v25);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26>
+internal::ValueArray26<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+    T26> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26) {
+  return internal::ValueArray26<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15,
+      v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27>
+internal::ValueArray27<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+    T27> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27) {
+  return internal::ValueArray27<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14,
+      v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28>
+internal::ValueArray28<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+    T28> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28) {
+  return internal::ValueArray28<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+      v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27,
+      v28);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29>
+internal::ValueArray29<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29) {
+  return internal::ValueArray29<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+      v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26,
+      v27, v28, v29);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30>
+internal::ValueArray30<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+    T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16,
+    T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24,
+    T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30) {
+  return internal::ValueArray30<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+      v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25,
+      v26, v27, v28, v29, v30);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31>
+internal::ValueArray31<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31) {
+  return internal::ValueArray31<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10,
+      v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24,
+      v25, v26, v27, v28, v29, v30, v31);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32>
+internal::ValueArray32<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32) {
+  return internal::ValueArray32<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32>(v1, v2, v3, v4, v5, v6, v7, v8, v9,
+      v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33>
+internal::ValueArray33<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6,
+    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33) {
+  return internal::ValueArray33<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33>(v1, v2, v3, v4, v5, v6, v7, v8,
+      v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32, v33);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34>
+internal::ValueArray34<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5,
+    T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14,
+    T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22,
+    T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30,
+    T31 v31, T32 v32, T33 v33, T34 v34) {
+  return internal::ValueArray34<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34>(v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22,
+      v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35>
+internal::ValueArray35<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21,
+    T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29,
+    T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35) {
+  return internal::ValueArray35<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35>(v1, v2, v3, v4, v5, v6,
+      v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21,
+      v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36>
+internal::ValueArray36<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21,
+    T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29,
+    T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36) {
+  return internal::ValueArray36<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36>(v1, v2, v3, v4,
+      v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33,
+      v34, v35, v36);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37>
+internal::ValueArray37<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37> Values(T1 v1, T2 v2, T3 v3,
+    T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28,
+    T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36,
+    T37 v37) {
+  return internal::ValueArray37<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37>(v1, v2, v3,
+      v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33,
+      v34, v35, v36, v37);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38>
+internal::ValueArray38<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28,
+    T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36,
+    T37 v37, T38 v38) {
+  return internal::ValueArray38<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38>(v1, v2,
+      v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18,
+      v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32,
+      v33, v34, v35, v36, v37, v38);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39>
+internal::ValueArray39<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39> Values(T1 v1, T2 v2,
+    T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12,
+    T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20,
+    T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28,
+    T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36,
+    T37 v37, T38 v38, T39 v39) {
+  return internal::ValueArray39<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39>(v1,
+      v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17,
+      v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31,
+      v32, v33, v34, v35, v36, v37, v38, v39);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40>
+internal::ValueArray40<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40> Values(T1 v1,
+    T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11,
+    T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19,
+    T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27,
+    T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35,
+    T36 v36, T37 v37, T38 v38, T39 v39, T40 v40) {
+  return internal::ValueArray40<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15,
+      v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29,
+      v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41>
+internal::ValueArray41<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+    T41> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41) {
+  return internal::ValueArray41<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14,
+      v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28,
+      v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42>
+internal::ValueArray42<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+    T42> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+    T42 v42) {
+  return internal::ValueArray42<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+      v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27,
+      v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41,
+      v42);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43>
+internal::ValueArray43<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+    T43> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+    T42 v42, T43 v43) {
+  return internal::ValueArray43<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+      v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26,
+      v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40,
+      v41, v42, v43);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44>
+internal::ValueArray44<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+    T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+    T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+    T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+    T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+    T42 v42, T43 v43, T44 v44) {
+  return internal::ValueArray44<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+      v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25,
+      v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39,
+      v40, v41, v42, v43, v44);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45>
+internal::ValueArray45<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+    T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16,
+    T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24,
+    T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32,
+    T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40,
+    T41 v41, T42 v42, T43 v43, T44 v44, T45 v45) {
+  return internal::ValueArray45<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45>(v1, v2, v3, v4, v5, v6, v7, v8, v9, v10,
+      v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24,
+      v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38,
+      v39, v40, v41, v42, v43, v44, v45);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46>
+internal::ValueArray46<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39,
+    T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46) {
+  return internal::ValueArray46<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45, T46>(v1, v2, v3, v4, v5, v6, v7, v8, v9,
+      v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37,
+      v38, v39, v40, v41, v42, v43, v44, v45, v46);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47>
+internal::ValueArray47<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46, T47> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+    T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39,
+    T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47) {
+  return internal::ValueArray47<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45, T46, T47>(v1, v2, v3, v4, v5, v6, v7, v8,
+      v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23,
+      v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37,
+      v38, v39, v40, v41, v42, v43, v44, v45, v46, v47);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48>
+internal::ValueArray48<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46, T47, T48> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6,
+    T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15,
+    T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23,
+    T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31,
+    T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39,
+    T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47,
+    T48 v48) {
+  return internal::ValueArray48<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45, T46, T47, T48>(v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22,
+      v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36,
+      v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49>
+internal::ValueArray49<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46, T47, T48, T49> Values(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5,
+    T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13, T14 v14,
+    T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21, T22 v22,
+    T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29, T30 v30,
+    T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37, T38 v38,
+    T39 v39, T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45, T46 v46,
+    T47 v47, T48 v48, T49 v49) {
+  return internal::ValueArray49<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45, T46, T47, T48, T49>(v1, v2, v3, v4, v5, v6,
+      v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21,
+      v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35,
+      v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49, typename T50>
+internal::ValueArray50<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46, T47, T48, T49, T50> Values(T1 v1, T2 v2, T3 v3, T4 v4,
+    T5 v5, T6 v6, T7 v7, T8 v8, T9 v9, T10 v10, T11 v11, T12 v12, T13 v13,
+    T14 v14, T15 v15, T16 v16, T17 v17, T18 v18, T19 v19, T20 v20, T21 v21,
+    T22 v22, T23 v23, T24 v24, T25 v25, T26 v26, T27 v27, T28 v28, T29 v29,
+    T30 v30, T31 v31, T32 v32, T33 v33, T34 v34, T35 v35, T36 v36, T37 v37,
+    T38 v38, T39 v39, T40 v40, T41 v41, T42 v42, T43 v43, T44 v44, T45 v45,
+    T46 v46, T47 v47, T48 v48, T49 v49, T50 v50) {
+  return internal::ValueArray50<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12, T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26, T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40, T41, T42, T43, T44, T45, T46, T47, T48, T49, T50>(v1, v2, v3, v4,
+      v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19,
+      v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33,
+      v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47,
+      v48, v49, v50);
+}
+
+// Bool() allows generating tests with parameters in a set of (false, true).
+//
+// Synopsis:
+// Bool()
+//   - returns a generator producing sequences with elements {false, true}.
+//
+// It is useful when testing code that depends on Boolean flags. Combinations
+// of multiple flags can be tested when several Bool()'s are combined using
+// Combine() function.
+//
+// In the following example all tests in the test case FlagDependentTest
+// will be instantiated twice with parameters false and true.
+//
+// class FlagDependentTest : public testing::TestWithParam<bool> {
+//   virtual void SetUp() {
+//     external_flag = GetParam();
+//   }
+// }
+// INSTANTIATE_TEST_CASE_P(BoolSequence, FlagDependentTest, Bool());
+//
+inline internal::ParamGenerator<bool> Bool() {
+  return Values(false, true);
+}
+
+# if GTEST_HAS_COMBINE
+// Combine() allows the user to combine two or more sequences to produce
+// values of a Cartesian product of those sequences' elements.
+//
+// Synopsis:
+// Combine(gen1, gen2, ..., genN)
+//   - returns a generator producing sequences with elements coming from
+//     the Cartesian product of elements from the sequences generated by
+//     gen1, gen2, ..., genN. The sequence elements will have a type of
+//     tuple<T1, T2, ..., TN> where T1, T2, ..., TN are the types
+//     of elements from sequences produces by gen1, gen2, ..., genN.
+//
+// Combine can have up to 10 arguments. This number is currently limited
+// by the maximum number of elements in the tuple implementation used by Google
+// Test.
+//
+// Example:
+//
+// This will instantiate tests in test case AnimalTest each one with
+// the parameter values tuple("cat", BLACK), tuple("cat", WHITE),
+// tuple("dog", BLACK), and tuple("dog", WHITE):
+//
+// enum Color { BLACK, GRAY, WHITE };
+// class AnimalTest
+//     : public testing::TestWithParam<tuple<const char*, Color> > {...};
+//
+// TEST_P(AnimalTest, AnimalLooksNice) {...}
+//
+// INSTANTIATE_TEST_CASE_P(AnimalVariations, AnimalTest,
+//                         Combine(Values("cat", "dog"),
+//                                 Values(BLACK, WHITE)));
+//
+// This will instantiate tests in FlagDependentTest with all variations of two
+// Boolean flags:
+//
+// class FlagDependentTest
+//     : public testing::TestWithParam<tuple<bool, bool> > {
+//   virtual void SetUp() {
+//     // Assigns external_flag_1 and external_flag_2 values from the tuple.
+//     tie(external_flag_1, external_flag_2) = GetParam();
+//   }
+// };
+//
+// TEST_P(FlagDependentTest, TestFeature1) {
+//   // Test your code using external_flag_1 and external_flag_2 here.
+// }
+// INSTANTIATE_TEST_CASE_P(TwoBoolSequence, FlagDependentTest,
+//                         Combine(Bool(), Bool()));
+//
+template <typename Generator1, typename Generator2>
+internal::CartesianProductHolder2<Generator1, Generator2> Combine(
+    const Generator1& g1, const Generator2& g2) {
+  return internal::CartesianProductHolder2<Generator1, Generator2>(
+      g1, g2);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3>
+internal::CartesianProductHolder3<Generator1, Generator2, Generator3> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3) {
+  return internal::CartesianProductHolder3<Generator1, Generator2, Generator3>(
+      g1, g2, g3);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4>
+internal::CartesianProductHolder4<Generator1, Generator2, Generator3,
+    Generator4> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4) {
+  return internal::CartesianProductHolder4<Generator1, Generator2, Generator3,
+      Generator4>(
+      g1, g2, g3, g4);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5>
+internal::CartesianProductHolder5<Generator1, Generator2, Generator3,
+    Generator4, Generator5> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5) {
+  return internal::CartesianProductHolder5<Generator1, Generator2, Generator3,
+      Generator4, Generator5>(
+      g1, g2, g3, g4, g5);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5, typename Generator6>
+internal::CartesianProductHolder6<Generator1, Generator2, Generator3,
+    Generator4, Generator5, Generator6> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6) {
+  return internal::CartesianProductHolder6<Generator1, Generator2, Generator3,
+      Generator4, Generator5, Generator6>(
+      g1, g2, g3, g4, g5, g6);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5, typename Generator6,
+    typename Generator7>
+internal::CartesianProductHolder7<Generator1, Generator2, Generator3,
+    Generator4, Generator5, Generator6, Generator7> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7) {
+  return internal::CartesianProductHolder7<Generator1, Generator2, Generator3,
+      Generator4, Generator5, Generator6, Generator7>(
+      g1, g2, g3, g4, g5, g6, g7);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5, typename Generator6,
+    typename Generator7, typename Generator8>
+internal::CartesianProductHolder8<Generator1, Generator2, Generator3,
+    Generator4, Generator5, Generator6, Generator7, Generator8> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7, const Generator8& g8) {
+  return internal::CartesianProductHolder8<Generator1, Generator2, Generator3,
+      Generator4, Generator5, Generator6, Generator7, Generator8>(
+      g1, g2, g3, g4, g5, g6, g7, g8);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5, typename Generator6,
+    typename Generator7, typename Generator8, typename Generator9>
+internal::CartesianProductHolder9<Generator1, Generator2, Generator3,
+    Generator4, Generator5, Generator6, Generator7, Generator8,
+    Generator9> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7, const Generator8& g8, const Generator9& g9) {
+  return internal::CartesianProductHolder9<Generator1, Generator2, Generator3,
+      Generator4, Generator5, Generator6, Generator7, Generator8, Generator9>(
+      g1, g2, g3, g4, g5, g6, g7, g8, g9);
+}
+
+template <typename Generator1, typename Generator2, typename Generator3,
+    typename Generator4, typename Generator5, typename Generator6,
+    typename Generator7, typename Generator8, typename Generator9,
+    typename Generator10>
+internal::CartesianProductHolder10<Generator1, Generator2, Generator3,
+    Generator4, Generator5, Generator6, Generator7, Generator8, Generator9,
+    Generator10> Combine(
+    const Generator1& g1, const Generator2& g2, const Generator3& g3,
+        const Generator4& g4, const Generator5& g5, const Generator6& g6,
+        const Generator7& g7, const Generator8& g8, const Generator9& g9,
+        const Generator10& g10) {
+  return internal::CartesianProductHolder10<Generator1, Generator2, Generator3,
+      Generator4, Generator5, Generator6, Generator7, Generator8, Generator9,
+      Generator10>(
+      g1, g2, g3, g4, g5, g6, g7, g8, g9, g10);
+}
+# endif  // GTEST_HAS_COMBINE
+
+
+
+# define TEST_P(test_case_name, test_name) \
+  class GTEST_TEST_CLASS_NAME_(test_case_name, test_name) \
+      : public test_case_name { \
+   public: \
+    GTEST_TEST_CLASS_NAME_(test_case_name, test_name)() {} \
+    virtual void TestBody(); \
+   private: \
+    static int AddToRegistry() { \
+      ::testing::UnitTest::GetInstance()->parameterized_test_registry(). \
+          GetTestCasePatternHolder<test_case_name>(\
+              #test_case_name, \
+              ::testing::internal::CodeLocation(\
+                  __FILE__, __LINE__))->AddTestPattern(\
+                      #test_case_name, \
+                      #test_name, \
+                      new ::testing::internal::TestMetaFactory< \
+                          GTEST_TEST_CLASS_NAME_(\
+                              test_case_name, test_name)>()); \
+      return 0; \
+    } \
+    static int gtest_registering_dummy_ GTEST_ATTRIBUTE_UNUSED_; \
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(\
+        GTEST_TEST_CLASS_NAME_(test_case_name, test_name)); \
+  }; \
+  int GTEST_TEST_CLASS_NAME_(test_case_name, \
+                             test_name)::gtest_registering_dummy_ = \
+      GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::AddToRegistry(); \
+  void GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::TestBody()
+
+// The optional last argument to INSTANTIATE_TEST_CASE_P allows the user
+// to specify a function or functor that generates custom test name suffixes
+// based on the test parameters. The function should accept one argument of
+// type testing::TestParamInfo<class ParamType>, and return std::string.
+//
+// testing::PrintToStringParamName is a builtin test suffix generator that
+// returns the value of testing::PrintToString(GetParam()). It does not work
+// for std::string or C strings.
+//
+// Note: test names must be non-empty, unique, and may only contain ASCII
+// alphanumeric characters or underscore.
+
+# define INSTANTIATE_TEST_CASE_P(prefix, test_case_name, generator, ...) \
+  ::testing::internal::ParamGenerator<test_case_name::ParamType> \
+      gtest_##prefix##test_case_name##_EvalGenerator_() { return generator; } \
+  ::std::string gtest_##prefix##test_case_name##_EvalGenerateName_( \
+      const ::testing::TestParamInfo<test_case_name::ParamType>& info) { \
+    return ::testing::internal::GetParamNameGen<test_case_name::ParamType> \
+        (__VA_ARGS__)(info); \
+  } \
+  int gtest_##prefix##test_case_name##_dummy_ GTEST_ATTRIBUTE_UNUSED_ = \
+      ::testing::UnitTest::GetInstance()->parameterized_test_registry(). \
+          GetTestCasePatternHolder<test_case_name>(\
+              #test_case_name, \
+              ::testing::internal::CodeLocation(\
+                  __FILE__, __LINE__))->AddTestCaseInstantiation(\
+                      #prefix, \
+                      &gtest_##prefix##test_case_name##_EvalGenerator_, \
+                      &gtest_##prefix##test_case_name##_EvalGenerateName_, \
+                      __FILE__, __LINE__)
+
+}  // namespace testing
+
+#endif  // GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-param-test.h.pump b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-param-test.h.pump
new file mode 100644
index 0000000000..3078d6d2a1
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-param-test.h.pump
@@ -0,0 +1,510 @@
+$$ -*- mode: c++; -*-
+$var n = 50  $$ Maximum length of Values arguments we want to support.
+$var maxtuple = 10  $$ Maximum number of Combine arguments we want to support.
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: vladl@google.com (Vlad Losev)
+//
+// Macros and functions for implementing parameterized tests
+// in Google C++ Testing Framework (Google Test)
+//
+// This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
+
+
+// Value-parameterized tests allow you to test your code with different
+// parameters without writing multiple copies of the same test.
+//
+// Here is how you use value-parameterized tests:
+
+#if 0
+
+// To write value-parameterized tests, first you should define a fixture
+// class. It is usually derived from testing::TestWithParam<T> (see below for
+// another inheritance scheme that's sometimes useful in more complicated
+// class hierarchies), where the type of your parameter values.
+// TestWithParam<T> is itself derived from testing::Test. T can be any
+// copyable type. If it's a raw pointer, you are responsible for managing the
+// lifespan of the pointed values.
+
+class FooTest : public ::testing::TestWithParam<const char*> {
+  // You can implement all the usual class fixture members here.
+};
+
+// Then, use the TEST_P macro to define as many parameterized tests
+// for this fixture as you want. The _P suffix is for "parameterized"
+// or "pattern", whichever you prefer to think.
+
+TEST_P(FooTest, DoesBlah) {
+  // Inside a test, access the test parameter with the GetParam() method
+  // of the TestWithParam<T> class:
+  EXPECT_TRUE(foo.Blah(GetParam()));
+  ...
+}
+
+TEST_P(FooTest, HasBlahBlah) {
+  ...
+}
+
+// Finally, you can use INSTANTIATE_TEST_CASE_P to instantiate the test
+// case with any set of parameters you want. Google Test defines a number
+// of functions for generating test parameters. They return what we call
+// (surprise!) parameter generators. Here is a  summary of them, which
+// are all in the testing namespace:
+//
+//
+//  Range(begin, end [, step]) - Yields values {begin, begin+step,
+//                               begin+step+step, ...}. The values do not
+//                               include end. step defaults to 1.
+//  Values(v1, v2, ..., vN)    - Yields values {v1, v2, ..., vN}.
+//  ValuesIn(container)        - Yields values from a C-style array, an STL
+//  ValuesIn(begin,end)          container, or an iterator range [begin, end).
+//  Bool()                     - Yields sequence {false, true}.
+//  Combine(g1, g2, ..., gN)   - Yields all combinations (the Cartesian product
+//                               for the math savvy) of the values generated
+//                               by the N generators.
+//
+// For more details, see comments at the definitions of these functions below
+// in this file.
+//
+// The following statement will instantiate tests from the FooTest test case
+// each with parameter values "meeny", "miny", and "moe".
+
+INSTANTIATE_TEST_CASE_P(InstantiationName,
+                        FooTest,
+                        Values("meeny", "miny", "moe"));
+
+// To distinguish different instances of the pattern, (yes, you
+// can instantiate it more then once) the first argument to the
+// INSTANTIATE_TEST_CASE_P macro is a prefix that will be added to the
+// actual test case name. Remember to pick unique prefixes for different
+// instantiations. The tests from the instantiation above will have
+// these names:
+//
+//    * InstantiationName/FooTest.DoesBlah/0 for "meeny"
+//    * InstantiationName/FooTest.DoesBlah/1 for "miny"
+//    * InstantiationName/FooTest.DoesBlah/2 for "moe"
+//    * InstantiationName/FooTest.HasBlahBlah/0 for "meeny"
+//    * InstantiationName/FooTest.HasBlahBlah/1 for "miny"
+//    * InstantiationName/FooTest.HasBlahBlah/2 for "moe"
+//
+// You can use these names in --gtest_filter.
+//
+// This statement will instantiate all tests from FooTest again, each
+// with parameter values "cat" and "dog":
+
+const char* pets[] = {"cat", "dog"};
+INSTANTIATE_TEST_CASE_P(AnotherInstantiationName, FooTest, ValuesIn(pets));
+
+// The tests from the instantiation above will have these names:
+//
+//    * AnotherInstantiationName/FooTest.DoesBlah/0 for "cat"
+//    * AnotherInstantiationName/FooTest.DoesBlah/1 for "dog"
+//    * AnotherInstantiationName/FooTest.HasBlahBlah/0 for "cat"
+//    * AnotherInstantiationName/FooTest.HasBlahBlah/1 for "dog"
+//
+// Please note that INSTANTIATE_TEST_CASE_P will instantiate all tests
+// in the given test case, whether their definitions come before or
+// AFTER the INSTANTIATE_TEST_CASE_P statement.
+//
+// Please also note that generator expressions (including parameters to the
+// generators) are evaluated in InitGoogleTest(), after main() has started.
+// This allows the user on one hand, to adjust generator parameters in order
+// to dynamically determine a set of tests to run and on the other hand,
+// give the user a chance to inspect the generated tests with Google Test
+// reflection API before RUN_ALL_TESTS() is executed.
+//
+// You can see samples/sample7_unittest.cc and samples/sample8_unittest.cc
+// for more examples.
+//
+// In the future, we plan to publish the API for defining new parameter
+// generators. But for now this interface remains part of the internal
+// implementation and is subject to change.
+//
+//
+// A parameterized test fixture must be derived from testing::Test and from
+// testing::WithParamInterface<T>, where T is the type of the parameter
+// values. Inheriting from TestWithParam<T> satisfies that requirement because
+// TestWithParam<T> inherits from both Test and WithParamInterface. In more
+// complicated hierarchies, however, it is occasionally useful to inherit
+// separately from Test and WithParamInterface. For example:
+
+class BaseTest : public ::testing::Test {
+  // You can inherit all the usual members for a non-parameterized test
+  // fixture here.
+};
+
+class DerivedTest : public BaseTest, public ::testing::WithParamInterface<int> {
+  // The usual test fixture members go here too.
+};
+
+TEST_F(BaseTest, HasFoo) {
+  // This is an ordinary non-parameterized test.
+}
+
+TEST_P(DerivedTest, DoesBlah) {
+  // GetParam works just the same here as if you inherit from TestWithParam.
+  EXPECT_TRUE(foo.Blah(GetParam()));
+}
+
+#endif  // 0
+
+#include "gtest/internal/gtest-port.h"
+
+#if !GTEST_OS_SYMBIAN
+# include <utility>
+#endif
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+#include "gtest/internal/gtest-internal.h"
+#include "gtest/internal/gtest-param-util.h"
+#include "gtest/internal/gtest-param-util-generated.h"
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+
+// Functions producing parameter generators.
+//
+// Google Test uses these generators to produce parameters for value-
+// parameterized tests. When a parameterized test case is instantiated
+// with a particular generator, Google Test creates and runs tests
+// for each element in the sequence produced by the generator.
+//
+// In the following sample, tests from test case FooTest are instantiated
+// each three times with parameter values 3, 5, and 8:
+//
+// class FooTest : public TestWithParam<int> { ... };
+//
+// TEST_P(FooTest, TestThis) {
+// }
+// TEST_P(FooTest, TestThat) {
+// }
+// INSTANTIATE_TEST_CASE_P(TestSequence, FooTest, Values(3, 5, 8));
+//
+
+// Range() returns generators providing sequences of values in a range.
+//
+// Synopsis:
+// Range(start, end)
+//   - returns a generator producing a sequence of values {start, start+1,
+//     start+2, ..., }.
+// Range(start, end, step)
+//   - returns a generator producing a sequence of values {start, start+step,
+//     start+step+step, ..., }.
+// Notes:
+//   * The generated sequences never include end. For example, Range(1, 5)
+//     returns a generator producing a sequence {1, 2, 3, 4}. Range(1, 9, 2)
+//     returns a generator producing {1, 3, 5, 7}.
+//   * start and end must have the same type. That type may be any integral or
+//     floating-point type or a user defined type satisfying these conditions:
+//     * It must be assignable (have operator=() defined).
+//     * It must have operator+() (operator+(int-compatible type) for
+//       two-operand version).
+//     * It must have operator<() defined.
+//     Elements in the resulting sequences will also have that type.
+//   * Condition start < end must be satisfied in order for resulting sequences
+//     to contain any elements.
+//
+template <typename T, typename IncrementT>
+internal::ParamGenerator<T> Range(T start, T end, IncrementT step) {
+  return internal::ParamGenerator<T>(
+      new internal::RangeGenerator<T, IncrementT>(start, end, step));
+}
+
+template <typename T>
+internal::ParamGenerator<T> Range(T start, T end) {
+  return Range(start, end, 1);
+}
+
+// ValuesIn() function allows generation of tests with parameters coming from
+// a container.
+//
+// Synopsis:
+// ValuesIn(const T (&array)[N])
+//   - returns a generator producing sequences with elements from
+//     a C-style array.
+// ValuesIn(const Container& container)
+//   - returns a generator producing sequences with elements from
+//     an STL-style container.
+// ValuesIn(Iterator begin, Iterator end)
+//   - returns a generator producing sequences with elements from
+//     a range [begin, end) defined by a pair of STL-style iterators. These
+//     iterators can also be plain C pointers.
+//
+// Please note that ValuesIn copies the values from the containers
+// passed in and keeps them to generate tests in RUN_ALL_TESTS().
+//
+// Examples:
+//
+// This instantiates tests from test case StringTest
+// each with C-string values of "foo", "bar", and "baz":
+//
+// const char* strings[] = {"foo", "bar", "baz"};
+// INSTANTIATE_TEST_CASE_P(StringSequence, SrtingTest, ValuesIn(strings));
+//
+// This instantiates tests from test case StlStringTest
+// each with STL strings with values "a" and "b":
+//
+// ::std::vector< ::std::string> GetParameterStrings() {
+//   ::std::vector< ::std::string> v;
+//   v.push_back("a");
+//   v.push_back("b");
+//   return v;
+// }
+//
+// INSTANTIATE_TEST_CASE_P(CharSequence,
+//                         StlStringTest,
+//                         ValuesIn(GetParameterStrings()));
+//
+//
+// This will also instantiate tests from CharTest
+// each with parameter values 'a' and 'b':
+//
+// ::std::list<char> GetParameterChars() {
+//   ::std::list<char> list;
+//   list.push_back('a');
+//   list.push_back('b');
+//   return list;
+// }
+// ::std::list<char> l = GetParameterChars();
+// INSTANTIATE_TEST_CASE_P(CharSequence2,
+//                         CharTest,
+//                         ValuesIn(l.begin(), l.end()));
+//
+template <typename ForwardIterator>
+internal::ParamGenerator<
+  typename ::testing::internal::IteratorTraits<ForwardIterator>::value_type>
+ValuesIn(ForwardIterator begin, ForwardIterator end) {
+  typedef typename ::testing::internal::IteratorTraits<ForwardIterator>
+      ::value_type ParamType;
+  return internal::ParamGenerator<ParamType>(
+      new internal::ValuesInIteratorRangeGenerator<ParamType>(begin, end));
+}
+
+template <typename T, size_t N>
+internal::ParamGenerator<T> ValuesIn(const T (&array)[N]) {
+  return ValuesIn(array, array + N);
+}
+
+template <class Container>
+internal::ParamGenerator<typename Container::value_type> ValuesIn(
+    const Container& container) {
+  return ValuesIn(container.begin(), container.end());
+}
+
+// Values() allows generating tests from explicitly specified list of
+// parameters.
+//
+// Synopsis:
+// Values(T v1, T v2, ..., T vN)
+//   - returns a generator producing sequences with elements v1, v2, ..., vN.
+//
+// For example, this instantiates tests from test case BarTest each
+// with values "one", "two", and "three":
+//
+// INSTANTIATE_TEST_CASE_P(NumSequence, BarTest, Values("one", "two", "three"));
+//
+// This instantiates tests from test case BazTest each with values 1, 2, 3.5.
+// The exact type of values will depend on the type of parameter in BazTest.
+//
+// INSTANTIATE_TEST_CASE_P(FloatingNumbers, BazTest, Values(1, 2, 3.5));
+//
+// Currently, Values() supports from 1 to $n parameters.
+//
+$range i 1..n
+$for i [[
+$range j 1..i
+
+template <$for j, [[typename T$j]]>
+internal::ValueArray$i<$for j, [[T$j]]> Values($for j, [[T$j v$j]]) {
+  return internal::ValueArray$i<$for j, [[T$j]]>($for j, [[v$j]]);
+}
+
+]]
+
+// Bool() allows generating tests with parameters in a set of (false, true).
+//
+// Synopsis:
+// Bool()
+//   - returns a generator producing sequences with elements {false, true}.
+//
+// It is useful when testing code that depends on Boolean flags. Combinations
+// of multiple flags can be tested when several Bool()'s are combined using
+// Combine() function.
+//
+// In the following example all tests in the test case FlagDependentTest
+// will be instantiated twice with parameters false and true.
+//
+// class FlagDependentTest : public testing::TestWithParam<bool> {
+//   virtual void SetUp() {
+//     external_flag = GetParam();
+//   }
+// }
+// INSTANTIATE_TEST_CASE_P(BoolSequence, FlagDependentTest, Bool());
+//
+inline internal::ParamGenerator<bool> Bool() {
+  return Values(false, true);
+}
+
+# if GTEST_HAS_COMBINE
+// Combine() allows the user to combine two or more sequences to produce
+// values of a Cartesian product of those sequences' elements.
+//
+// Synopsis:
+// Combine(gen1, gen2, ..., genN)
+//   - returns a generator producing sequences with elements coming from
+//     the Cartesian product of elements from the sequences generated by
+//     gen1, gen2, ..., genN. The sequence elements will have a type of
+//     tuple<T1, T2, ..., TN> where T1, T2, ..., TN are the types
+//     of elements from sequences produces by gen1, gen2, ..., genN.
+//
+// Combine can have up to $maxtuple arguments. This number is currently limited
+// by the maximum number of elements in the tuple implementation used by Google
+// Test.
+//
+// Example:
+//
+// This will instantiate tests in test case AnimalTest each one with
+// the parameter values tuple("cat", BLACK), tuple("cat", WHITE),
+// tuple("dog", BLACK), and tuple("dog", WHITE):
+//
+// enum Color { BLACK, GRAY, WHITE };
+// class AnimalTest
+//     : public testing::TestWithParam<tuple<const char*, Color> > {...};
+//
+// TEST_P(AnimalTest, AnimalLooksNice) {...}
+//
+// INSTANTIATE_TEST_CASE_P(AnimalVariations, AnimalTest,
+//                         Combine(Values("cat", "dog"),
+//                                 Values(BLACK, WHITE)));
+//
+// This will instantiate tests in FlagDependentTest with all variations of two
+// Boolean flags:
+//
+// class FlagDependentTest
+//     : public testing::TestWithParam<tuple<bool, bool> > {
+//   virtual void SetUp() {
+//     // Assigns external_flag_1 and external_flag_2 values from the tuple.
+//     tie(external_flag_1, external_flag_2) = GetParam();
+//   }
+// };
+//
+// TEST_P(FlagDependentTest, TestFeature1) {
+//   // Test your code using external_flag_1 and external_flag_2 here.
+// }
+// INSTANTIATE_TEST_CASE_P(TwoBoolSequence, FlagDependentTest,
+//                         Combine(Bool(), Bool()));
+//
+$range i 2..maxtuple
+$for i [[
+$range j 1..i
+
+template <$for j, [[typename Generator$j]]>
+internal::CartesianProductHolder$i<$for j, [[Generator$j]]> Combine(
+    $for j, [[const Generator$j& g$j]]) {
+  return internal::CartesianProductHolder$i<$for j, [[Generator$j]]>(
+      $for j, [[g$j]]);
+}
+
+]]
+# endif  // GTEST_HAS_COMBINE
+
+
+
+# define TEST_P(test_case_name, test_name) \
+  class GTEST_TEST_CLASS_NAME_(test_case_name, test_name) \
+      : public test_case_name { \
+   public: \
+    GTEST_TEST_CLASS_NAME_(test_case_name, test_name)() {} \
+    virtual void TestBody(); \
+   private: \
+    static int AddToRegistry() { \
+      ::testing::UnitTest::GetInstance()->parameterized_test_registry(). \
+          GetTestCasePatternHolder<test_case_name>(\
+              #test_case_name, \
+              ::testing::internal::CodeLocation(\
+                  __FILE__, __LINE__))->AddTestPattern(\
+                      #test_case_name, \
+                      #test_name, \
+                      new ::testing::internal::TestMetaFactory< \
+                          GTEST_TEST_CLASS_NAME_(\
+                              test_case_name, test_name)>()); \
+      return 0; \
+    } \
+    static int gtest_registering_dummy_ GTEST_ATTRIBUTE_UNUSED_; \
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(\
+        GTEST_TEST_CLASS_NAME_(test_case_name, test_name)); \
+  }; \
+  int GTEST_TEST_CLASS_NAME_(test_case_name, \
+                             test_name)::gtest_registering_dummy_ = \
+      GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::AddToRegistry(); \
+  void GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::TestBody()
+
+// The optional last argument to INSTANTIATE_TEST_CASE_P allows the user
+// to specify a function or functor that generates custom test name suffixes
+// based on the test parameters. The function should accept one argument of
+// type testing::TestParamInfo<class ParamType>, and return std::string.
+//
+// testing::PrintToStringParamName is a builtin test suffix generator that
+// returns the value of testing::PrintToString(GetParam()).
+//
+// Note: test names must be non-empty, unique, and may only contain ASCII
+// alphanumeric characters or underscore. Because PrintToString adds quotes
+// to std::string and C strings, it won't work for these types.
+
+# define INSTANTIATE_TEST_CASE_P(prefix, test_case_name, generator, ...) \
+  ::testing::internal::ParamGenerator<test_case_name::ParamType> \
+      gtest_##prefix##test_case_name##_EvalGenerator_() { return generator; } \
+  ::std::string gtest_##prefix##test_case_name##_EvalGenerateName_( \
+      const ::testing::TestParamInfo<test_case_name::ParamType>& info) { \
+    return ::testing::internal::GetParamNameGen<test_case_name::ParamType> \
+        (__VA_ARGS__)(info); \
+  } \
+  int gtest_##prefix##test_case_name##_dummy_ GTEST_ATTRIBUTE_UNUSED_ = \
+      ::testing::UnitTest::GetInstance()->parameterized_test_registry(). \
+          GetTestCasePatternHolder<test_case_name>(\
+              #test_case_name, \
+              ::testing::internal::CodeLocation(\
+                  __FILE__, __LINE__))->AddTestCaseInstantiation(\
+                      #prefix, \
+                      &gtest_##prefix##test_case_name##_EvalGenerator_, \
+                      &gtest_##prefix##test_case_name##_EvalGenerateName_, \
+                      __FILE__, __LINE__)
+
+}  // namespace testing
+
+#endif  // GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PARAM_TEST_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-printers.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-printers.h
new file mode 100644
index 0000000000..8a33164cb3
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-printers.h
@@ -0,0 +1,993 @@
+// Copyright 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Google Test - The Google C++ Testing Framework
+//
+// This file implements a universal value printer that can print a
+// value of any type T:
+//
+//   void ::testing::internal::UniversalPrinter<T>::Print(value, ostream_ptr);
+//
+// A user can teach this function how to print a class type T by
+// defining either operator<<() or PrintTo() in the namespace that
+// defines T.  More specifically, the FIRST defined function in the
+// following list will be used (assuming T is defined in namespace
+// foo):
+//
+//   1. foo::PrintTo(const T&, ostream*)
+//   2. operator<<(ostream&, const T&) defined in either foo or the
+//      global namespace.
+//
+// If none of the above is defined, it will print the debug string of
+// the value if it is a protocol buffer, or print the raw bytes in the
+// value otherwise.
+//
+// To aid debugging: when T is a reference type, the address of the
+// value is also printed; when T is a (const) char pointer, both the
+// pointer value and the NUL-terminated string it points to are
+// printed.
+//
+// We also provide some convenient wrappers:
+//
+//   // Prints a value to a string.  For a (const or not) char
+//   // pointer, the NUL-terminated string (but not the pointer) is
+//   // printed.
+//   std::string ::testing::PrintToString(const T& value);
+//
+//   // Prints a value tersely: for a reference type, the referenced
+//   // value (but not the address) is printed; for a (const or not) char
+//   // pointer, the NUL-terminated string (but not the pointer) is
+//   // printed.
+//   void ::testing::internal::UniversalTersePrint(const T& value, ostream*);
+//
+//   // Prints value using the type inferred by the compiler.  The difference
+//   // from UniversalTersePrint() is that this function prints both the
+//   // pointer and the NUL-terminated string for a (const or not) char pointer.
+//   void ::testing::internal::UniversalPrint(const T& value, ostream*);
+//
+//   // Prints the fields of a tuple tersely to a string vector, one
+//   // element for each field. Tuple support must be enabled in
+//   // gtest-port.h.
+//   std::vector<string> UniversalTersePrintTupleFieldsToStrings(
+//       const Tuple& value);
+//
+// Known limitation:
+//
+// The print primitives print the elements of an STL-style container
+// using the compiler-inferred type of *iter where iter is a
+// const_iterator of the container.  When const_iterator is an input
+// iterator but not a forward iterator, this inferred type may not
+// match value_type, and the print output may be incorrect.  In
+// practice, this is rarely a problem as for most containers
+// const_iterator is a forward iterator.  We'll fix this if there's an
+// actual need for it.  Note that this fix cannot rely on value_type
+// being defined as many user-defined container types don't have
+// value_type.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PRINTERS_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PRINTERS_H_
+
+#include <ostream>  // NOLINT
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+#include "gtest/internal/gtest-port.h"
+#include "gtest/internal/gtest-internal.h"
+
+#if GTEST_HAS_STD_TUPLE_
+# include <tuple>
+#endif
+
+namespace testing {
+
+// Definitions in the 'internal' and 'internal2' name spaces are
+// subject to change without notice.  DO NOT USE THEM IN USER CODE!
+namespace internal2 {
+
+// Prints the given number of bytes in the given object to the given
+// ostream.
+GTEST_API_ void PrintBytesInObjectTo(const unsigned char* obj_bytes,
+                                     size_t count,
+                                     ::std::ostream* os);
+
+// For selecting which printer to use when a given type has neither <<
+// nor PrintTo().
+enum TypeKind {
+  kProtobuf,              // a protobuf type
+  kConvertibleToInteger,  // a type implicitly convertible to BiggestInt
+                          // (e.g. a named or unnamed enum type)
+  kOtherType              // anything else
+};
+
+// TypeWithoutFormatter<T, kTypeKind>::PrintValue(value, os) is called
+// by the universal printer to print a value of type T when neither
+// operator<< nor PrintTo() is defined for T, where kTypeKind is the
+// "kind" of T as defined by enum TypeKind.
+template <typename T, TypeKind kTypeKind>
+class TypeWithoutFormatter {
+ public:
+  // This default version is called when kTypeKind is kOtherType.
+  static void PrintValue(const T& value, ::std::ostream* os) {
+    PrintBytesInObjectTo(reinterpret_cast<const unsigned char*>(&value),
+                         sizeof(value), os);
+  }
+};
+
+// We print a protobuf using its ShortDebugString() when the string
+// doesn't exceed this many characters; otherwise we print it using
+// DebugString() for better readability.
+const size_t kProtobufOneLinerMaxLength = 50;
+
+template <typename T>
+class TypeWithoutFormatter<T, kProtobuf> {
+ public:
+  static void PrintValue(const T& value, ::std::ostream* os) {
+    const ::testing::internal::string short_str = value.ShortDebugString();
+    const ::testing::internal::string pretty_str =
+        short_str.length() <= kProtobufOneLinerMaxLength ?
+        short_str : ("\n" + value.DebugString());
+    *os << ("<" + pretty_str + ">");
+  }
+};
+
+template <typename T>
+class TypeWithoutFormatter<T, kConvertibleToInteger> {
+ public:
+  // Since T has no << operator or PrintTo() but can be implicitly
+  // converted to BiggestInt, we print it as a BiggestInt.
+  //
+  // Most likely T is an enum type (either named or unnamed), in which
+  // case printing it as an integer is the desired behavior.  In case
+  // T is not an enum, printing it as an integer is the best we can do
+  // given that it has no user-defined printer.
+  static void PrintValue(const T& value, ::std::ostream* os) {
+    const internal::BiggestInt kBigInt = value;
+    *os << kBigInt;
+  }
+};
+
+// Prints the given value to the given ostream.  If the value is a
+// protocol message, its debug string is printed; if it's an enum or
+// of a type implicitly convertible to BiggestInt, it's printed as an
+// integer; otherwise the bytes in the value are printed.  This is
+// what UniversalPrinter<T>::Print() does when it knows nothing about
+// type T and T has neither << operator nor PrintTo().
+//
+// A user can override this behavior for a class type Foo by defining
+// a << operator in the namespace where Foo is defined.
+//
+// We put this operator in namespace 'internal2' instead of 'internal'
+// to simplify the implementation, as much code in 'internal' needs to
+// use << in STL, which would conflict with our own << were it defined
+// in 'internal'.
+//
+// Note that this operator<< takes a generic std::basic_ostream<Char,
+// CharTraits> type instead of the more restricted std::ostream.  If
+// we define it to take an std::ostream instead, we'll get an
+// "ambiguous overloads" compiler error when trying to print a type
+// Foo that supports streaming to std::basic_ostream<Char,
+// CharTraits>, as the compiler cannot tell whether
+// operator<<(std::ostream&, const T&) or
+// operator<<(std::basic_stream<Char, CharTraits>, const Foo&) is more
+// specific.
+template <typename Char, typename CharTraits, typename T>
+::std::basic_ostream<Char, CharTraits>& operator<<(
+    ::std::basic_ostream<Char, CharTraits>& os, const T& x) {
+  TypeWithoutFormatter<T,
+      (internal::IsAProtocolMessage<T>::value ? kProtobuf :
+       internal::ImplicitlyConvertible<const T&, internal::BiggestInt>::value ?
+       kConvertibleToInteger : kOtherType)>::PrintValue(x, &os);
+  return os;
+}
+
+}  // namespace internal2
+}  // namespace testing
+
+// This namespace MUST NOT BE NESTED IN ::testing, or the name look-up
+// magic needed for implementing UniversalPrinter won't work.
+namespace testing_internal {
+
+// Used to print a value that is not an STL-style container when the
+// user doesn't define PrintTo() for it.
+template <typename T>
+void DefaultPrintNonContainerTo(const T& value, ::std::ostream* os) {
+  // With the following statement, during unqualified name lookup,
+  // testing::internal2::operator<< appears as if it was declared in
+  // the nearest enclosing namespace that contains both
+  // ::testing_internal and ::testing::internal2, i.e. the global
+  // namespace.  For more details, refer to the C++ Standard section
+  // 7.3.4-1 [namespace.udir].  This allows us to fall back onto
+  // testing::internal2::operator<< in case T doesn't come with a <<
+  // operator.
+  //
+  // We cannot write 'using ::testing::internal2::operator<<;', which
+  // gcc 3.3 fails to compile due to a compiler bug.
+  using namespace ::testing::internal2;  // NOLINT
+
+  // Assuming T is defined in namespace foo, in the next statement,
+  // the compiler will consider all of:
+  //
+  //   1. foo::operator<< (thanks to Koenig look-up),
+  //   2. ::operator<< (as the current namespace is enclosed in ::),
+  //   3. testing::internal2::operator<< (thanks to the using statement above).
+  //
+  // The operator<< whose type matches T best will be picked.
+  //
+  // We deliberately allow #2 to be a candidate, as sometimes it's
+  // impossible to define #1 (e.g. when foo is ::std, defining
+  // anything in it is undefined behavior unless you are a compiler
+  // vendor.).
+  *os << value;
+}
+
+}  // namespace testing_internal
+
+namespace testing {
+namespace internal {
+
+// FormatForComparison<ToPrint, OtherOperand>::Format(value) formats a
+// value of type ToPrint that is an operand of a comparison assertion
+// (e.g. ASSERT_EQ).  OtherOperand is the type of the other operand in
+// the comparison, and is used to help determine the best way to
+// format the value.  In particular, when the value is a C string
+// (char pointer) and the other operand is an STL string object, we
+// want to format the C string as a string, since we know it is
+// compared by value with the string object.  If the value is a char
+// pointer but the other operand is not an STL string object, we don't
+// know whether the pointer is supposed to point to a NUL-terminated
+// string, and thus want to print it as a pointer to be safe.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+
+// The default case.
+template <typename ToPrint, typename OtherOperand>
+class FormatForComparison {
+ public:
+  static ::std::string Format(const ToPrint& value) {
+    return ::testing::PrintToString(value);
+  }
+};
+
+// Array.
+template <typename ToPrint, size_t N, typename OtherOperand>
+class FormatForComparison<ToPrint[N], OtherOperand> {
+ public:
+  static ::std::string Format(const ToPrint* value) {
+    return FormatForComparison<const ToPrint*, OtherOperand>::Format(value);
+  }
+};
+
+// By default, print C string as pointers to be safe, as we don't know
+// whether they actually point to a NUL-terminated string.
+
+#define GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(CharType)                \
+  template <typename OtherOperand>                                      \
+  class FormatForComparison<CharType*, OtherOperand> {                  \
+   public:                                                              \
+    static ::std::string Format(CharType* value) {                      \
+      return ::testing::PrintToString(static_cast<const void*>(value)); \
+    }                                                                   \
+  }
+
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(char);
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(const char);
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(wchar_t);
+GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_(const wchar_t);
+
+#undef GTEST_IMPL_FORMAT_C_STRING_AS_POINTER_
+
+// If a C string is compared with an STL string object, we know it's meant
+// to point to a NUL-terminated string, and thus can print it as a string.
+
+#define GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(CharType, OtherStringType) \
+  template <>                                                           \
+  class FormatForComparison<CharType*, OtherStringType> {               \
+   public:                                                              \
+    static ::std::string Format(CharType* value) {                      \
+      return ::testing::PrintToString(value);                           \
+    }                                                                   \
+  }
+
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(char, ::std::string);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const char, ::std::string);
+
+#if GTEST_HAS_GLOBAL_STRING
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(char, ::string);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const char, ::string);
+#endif
+
+#if GTEST_HAS_GLOBAL_WSTRING
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(wchar_t, ::wstring);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const wchar_t, ::wstring);
+#endif
+
+#if GTEST_HAS_STD_WSTRING
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(wchar_t, ::std::wstring);
+GTEST_IMPL_FORMAT_C_STRING_AS_STRING_(const wchar_t, ::std::wstring);
+#endif
+
+#undef GTEST_IMPL_FORMAT_C_STRING_AS_STRING_
+
+// Formats a comparison assertion (e.g. ASSERT_EQ, EXPECT_LT, and etc)
+// operand to be used in a failure message.  The type (but not value)
+// of the other operand may affect the format.  This allows us to
+// print a char* as a raw pointer when it is compared against another
+// char* or void*, and print it as a C string when it is compared
+// against an std::string object, for example.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+template <typename T1, typename T2>
+std::string FormatForComparisonFailureMessage(
+    const T1& value, const T2& /* other_operand */) {
+  return FormatForComparison<T1, T2>::Format(value);
+}
+
+// UniversalPrinter<T>::Print(value, ostream_ptr) prints the given
+// value to the given ostream.  The caller must ensure that
+// 'ostream_ptr' is not NULL, or the behavior is undefined.
+//
+// We define UniversalPrinter as a class template (as opposed to a
+// function template), as we need to partially specialize it for
+// reference types, which cannot be done with function templates.
+template <typename T>
+class UniversalPrinter;
+
+template <typename T>
+void UniversalPrint(const T& value, ::std::ostream* os);
+
+// Used to print an STL-style container when the user doesn't define
+// a PrintTo() for it.
+template <typename C>
+void DefaultPrintTo(IsContainer /* dummy */,
+                    false_type /* is not a pointer */,
+                    const C& container, ::std::ostream* os) {
+  const size_t kMaxCount = 32;  // The maximum number of elements to print.
+  *os << '{';
+  size_t count = 0;
+  for (typename C::const_iterator it = container.begin();
+       it != container.end(); ++it, ++count) {
+    if (count > 0) {
+      *os << ',';
+      if (count == kMaxCount) {  // Enough has been printed.
+        *os << " ...";
+        break;
+      }
+    }
+    *os << ' ';
+    // We cannot call PrintTo(*it, os) here as PrintTo() doesn't
+    // handle *it being a native array.
+    internal::UniversalPrint(*it, os);
+  }
+
+  if (count > 0) {
+    *os << ' ';
+  }
+  *os << '}';
+}
+
+// Used to print a pointer that is neither a char pointer nor a member
+// pointer, when the user doesn't define PrintTo() for it.  (A member
+// variable pointer or member function pointer doesn't really point to
+// a location in the address space.  Their representation is
+// implementation-defined.  Therefore they will be printed as raw
+// bytes.)
+template <typename T>
+void DefaultPrintTo(IsNotContainer /* dummy */,
+                    true_type /* is a pointer */,
+                    T* p, ::std::ostream* os) {
+  if (p == NULL) {
+    *os << "NULL";
+  } else {
+    // C++ doesn't allow casting from a function pointer to any object
+    // pointer.
+    //
+    // IsTrue() silences warnings: "Condition is always true",
+    // "unreachable code".
+    if (IsTrue(ImplicitlyConvertible<T*, const void*>::value)) {
+      // T is not a function type.  We just call << to print p,
+      // relying on ADL to pick up user-defined << for their pointer
+      // types, if any.
+      *os << p;
+    } else {
+      // T is a function type, so '*os << p' doesn't do what we want
+      // (it just prints p as bool).  We want to print p as a const
+      // void*.  However, we cannot cast it to const void* directly,
+      // even using reinterpret_cast, as earlier versions of gcc
+      // (e.g. 3.4.5) cannot compile the cast when p is a function
+      // pointer.  Casting to UInt64 first solves the problem.
+      *os << reinterpret_cast<const void*>(
+          reinterpret_cast<internal::UInt64>(p));
+    }
+  }
+}
+
+// Used to print a non-container, non-pointer value when the user
+// doesn't define PrintTo() for it.
+template <typename T>
+void DefaultPrintTo(IsNotContainer /* dummy */,
+                    false_type /* is not a pointer */,
+                    const T& value, ::std::ostream* os) {
+  ::testing_internal::DefaultPrintNonContainerTo(value, os);
+}
+
+// Prints the given value using the << operator if it has one;
+// otherwise prints the bytes in it.  This is what
+// UniversalPrinter<T>::Print() does when PrintTo() is not specialized
+// or overloaded for type T.
+//
+// A user can override this behavior for a class type Foo by defining
+// an overload of PrintTo() in the namespace where Foo is defined.  We
+// give the user this option as sometimes defining a << operator for
+// Foo is not desirable (e.g. the coding style may prevent doing it,
+// or there is already a << operator but it doesn't do what the user
+// wants).
+template <typename T>
+void PrintTo(const T& value, ::std::ostream* os) {
+  // DefaultPrintTo() is overloaded.  The type of its first two
+  // arguments determine which version will be picked.  If T is an
+  // STL-style container, the version for container will be called; if
+  // T is a pointer, the pointer version will be called; otherwise the
+  // generic version will be called.
+  //
+  // Note that we check for container types here, prior to we check
+  // for protocol message types in our operator<<.  The rationale is:
+  //
+  // For protocol messages, we want to give people a chance to
+  // override Google Mock's format by defining a PrintTo() or
+  // operator<<.  For STL containers, other formats can be
+  // incompatible with Google Mock's format for the container
+  // elements; therefore we check for container types here to ensure
+  // that our format is used.
+  //
+  // The second argument of DefaultPrintTo() is needed to bypass a bug
+  // in Symbian's C++ compiler that prevents it from picking the right
+  // overload between:
+  //
+  //   PrintTo(const T& x, ...);
+  //   PrintTo(T* x, ...);
+  DefaultPrintTo(IsContainerTest<T>(0), is_pointer<T>(), value, os);
+}
+
+// The following list of PrintTo() overloads tells
+// UniversalPrinter<T>::Print() how to print standard types (built-in
+// types, strings, plain arrays, and pointers).
+
+// Overloads for various char types.
+GTEST_API_ void PrintTo(unsigned char c, ::std::ostream* os);
+GTEST_API_ void PrintTo(signed char c, ::std::ostream* os);
+inline void PrintTo(char c, ::std::ostream* os) {
+  // When printing a plain char, we always treat it as unsigned.  This
+  // way, the output won't be affected by whether the compiler thinks
+  // char is signed or not.
+  PrintTo(static_cast<unsigned char>(c), os);
+}
+
+// Overloads for other simple built-in types.
+inline void PrintTo(bool x, ::std::ostream* os) {
+  *os << (x ? "true" : "false");
+}
+
+// Overload for wchar_t type.
+// Prints a wchar_t as a symbol if it is printable or as its internal
+// code otherwise and also as its decimal code (except for L'\0').
+// The L'\0' char is printed as "L'\\0'". The decimal code is printed
+// as signed integer when wchar_t is implemented by the compiler
+// as a signed type and is printed as an unsigned integer when wchar_t
+// is implemented as an unsigned type.
+GTEST_API_ void PrintTo(wchar_t wc, ::std::ostream* os);
+
+// Overloads for C strings.
+GTEST_API_ void PrintTo(const char* s, ::std::ostream* os);
+inline void PrintTo(char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const char*>(s), os);
+}
+
+// signed/unsigned char is often used for representing binary data, so
+// we print pointers to it as void* to be safe.
+inline void PrintTo(const signed char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const void*>(s), os);
+}
+inline void PrintTo(signed char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const void*>(s), os);
+}
+inline void PrintTo(const unsigned char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const void*>(s), os);
+}
+inline void PrintTo(unsigned char* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const void*>(s), os);
+}
+
+// MSVC can be configured to define wchar_t as a typedef of unsigned
+// short.  It defines _NATIVE_WCHAR_T_DEFINED when wchar_t is a native
+// type.  When wchar_t is a typedef, defining an overload for const
+// wchar_t* would cause unsigned short* be printed as a wide string,
+// possibly causing invalid memory accesses.
+#if !defined(_MSC_VER) || defined(_NATIVE_WCHAR_T_DEFINED)
+// Overloads for wide C strings
+GTEST_API_ void PrintTo(const wchar_t* s, ::std::ostream* os);
+inline void PrintTo(wchar_t* s, ::std::ostream* os) {
+  PrintTo(ImplicitCast_<const wchar_t*>(s), os);
+}
+#endif
+
+// Overload for C arrays.  Multi-dimensional arrays are printed
+// properly.
+
+// Prints the given number of elements in an array, without printing
+// the curly braces.
+template <typename T>
+void PrintRawArrayTo(const T a[], size_t count, ::std::ostream* os) {
+  UniversalPrint(a[0], os);
+  for (size_t i = 1; i != count; i++) {
+    *os << ", ";
+    UniversalPrint(a[i], os);
+  }
+}
+
+// Overloads for ::string and ::std::string.
+#if GTEST_HAS_GLOBAL_STRING
+GTEST_API_ void PrintStringTo(const ::string&s, ::std::ostream* os);
+inline void PrintTo(const ::string& s, ::std::ostream* os) {
+  PrintStringTo(s, os);
+}
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+GTEST_API_ void PrintStringTo(const ::std::string&s, ::std::ostream* os);
+inline void PrintTo(const ::std::string& s, ::std::ostream* os) {
+  PrintStringTo(s, os);
+}
+
+// Overloads for ::wstring and ::std::wstring.
+#if GTEST_HAS_GLOBAL_WSTRING
+GTEST_API_ void PrintWideStringTo(const ::wstring&s, ::std::ostream* os);
+inline void PrintTo(const ::wstring& s, ::std::ostream* os) {
+  PrintWideStringTo(s, os);
+}
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+#if GTEST_HAS_STD_WSTRING
+GTEST_API_ void PrintWideStringTo(const ::std::wstring&s, ::std::ostream* os);
+inline void PrintTo(const ::std::wstring& s, ::std::ostream* os) {
+  PrintWideStringTo(s, os);
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+#if GTEST_HAS_TR1_TUPLE || GTEST_HAS_STD_TUPLE_
+// Helper function for printing a tuple.  T must be instantiated with
+// a tuple type.
+template <typename T>
+void PrintTupleTo(const T& t, ::std::ostream* os);
+#endif  // GTEST_HAS_TR1_TUPLE || GTEST_HAS_STD_TUPLE_
+
+#if GTEST_HAS_TR1_TUPLE
+// Overload for ::std::tr1::tuple.  Needed for printing function arguments,
+// which are packed as tuples.
+
+// Overloaded PrintTo() for tuples of various arities.  We support
+// tuples of up-to 10 fields.  The following implementation works
+// regardless of whether tr1::tuple is implemented using the
+// non-standard variadic template feature or not.
+
+inline void PrintTo(const ::std::tr1::tuple<>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1>
+void PrintTo(const ::std::tr1::tuple<T1>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2>
+void PrintTo(const ::std::tr1::tuple<T1, T2>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4, T5>& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+          typename T6>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4, T5, T6>& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+          typename T6, typename T7>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7>& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+          typename T6, typename T7, typename T8>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8>& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+          typename T6, typename T7, typename T8, typename T9>
+void PrintTo(const ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8, T9>& t,
+             ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+          typename T6, typename T7, typename T8, typename T9, typename T10>
+void PrintTo(
+    const ::std::tr1::tuple<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10>& t,
+    ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+#endif  // GTEST_HAS_TR1_TUPLE
+
+#if GTEST_HAS_STD_TUPLE_
+template <typename... Types>
+void PrintTo(const ::std::tuple<Types...>& t, ::std::ostream* os) {
+  PrintTupleTo(t, os);
+}
+#endif  // GTEST_HAS_STD_TUPLE_
+
+// Overload for std::pair.
+template <typename T1, typename T2>
+void PrintTo(const ::std::pair<T1, T2>& value, ::std::ostream* os) {
+  *os << '(';
+  // We cannot use UniversalPrint(value.first, os) here, as T1 may be
+  // a reference type.  The same for printing value.second.
+  UniversalPrinter<T1>::Print(value.first, os);
+  *os << ", ";
+  UniversalPrinter<T2>::Print(value.second, os);
+  *os << ')';
+}
+
+// Implements printing a non-reference type T by letting the compiler
+// pick the right overload of PrintTo() for T.
+template <typename T>
+class UniversalPrinter {
+ public:
+  // MSVC warns about adding const to a function type, so we want to
+  // disable the warning.
+  GTEST_DISABLE_MSC_WARNINGS_PUSH_(4180)
+
+  // Note: we deliberately don't call this PrintTo(), as that name
+  // conflicts with ::testing::internal::PrintTo in the body of the
+  // function.
+  static void Print(const T& value, ::std::ostream* os) {
+    // By default, ::testing::internal::PrintTo() is used for printing
+    // the value.
+    //
+    // Thanks to Koenig look-up, if T is a class and has its own
+    // PrintTo() function defined in its namespace, that function will
+    // be visible here.  Since it is more specific than the generic ones
+    // in ::testing::internal, it will be picked by the compiler in the
+    // following statement - exactly what we want.
+    PrintTo(value, os);
+  }
+
+  GTEST_DISABLE_MSC_WARNINGS_POP_()
+};
+
+// UniversalPrintArray(begin, len, os) prints an array of 'len'
+// elements, starting at address 'begin'.
+template <typename T>
+void UniversalPrintArray(const T* begin, size_t len, ::std::ostream* os) {
+  if (len == 0) {
+    *os << "{}";
+  } else {
+    *os << "{ ";
+    const size_t kThreshold = 18;
+    const size_t kChunkSize = 8;
+    // If the array has more than kThreshold elements, we'll have to
+    // omit some details by printing only the first and the last
+    // kChunkSize elements.
+    // TODO(wan@google.com): let the user control the threshold using a flag.
+    if (len <= kThreshold) {
+      PrintRawArrayTo(begin, len, os);
+    } else {
+      PrintRawArrayTo(begin, kChunkSize, os);
+      *os << ", ..., ";
+      PrintRawArrayTo(begin + len - kChunkSize, kChunkSize, os);
+    }
+    *os << " }";
+  }
+}
+// This overload prints a (const) char array compactly.
+GTEST_API_ void UniversalPrintArray(
+    const char* begin, size_t len, ::std::ostream* os);
+
+// This overload prints a (const) wchar_t array compactly.
+GTEST_API_ void UniversalPrintArray(
+    const wchar_t* begin, size_t len, ::std::ostream* os);
+
+// Implements printing an array type T[N].
+template <typename T, size_t N>
+class UniversalPrinter<T[N]> {
+ public:
+  // Prints the given array, omitting some elements when there are too
+  // many.
+  static void Print(const T (&a)[N], ::std::ostream* os) {
+    UniversalPrintArray(a, N, os);
+  }
+};
+
+// Implements printing a reference type T&.
+template <typename T>
+class UniversalPrinter<T&> {
+ public:
+  // MSVC warns about adding const to a function type, so we want to
+  // disable the warning.
+  GTEST_DISABLE_MSC_WARNINGS_PUSH_(4180)
+
+  static void Print(const T& value, ::std::ostream* os) {
+    // Prints the address of the value.  We use reinterpret_cast here
+    // as static_cast doesn't compile when T is a function type.
+    *os << "@" << reinterpret_cast<const void*>(&value) << " ";
+
+    // Then prints the value itself.
+    UniversalPrint(value, os);
+  }
+
+  GTEST_DISABLE_MSC_WARNINGS_POP_()
+};
+
+// Prints a value tersely: for a reference type, the referenced value
+// (but not the address) is printed; for a (const) char pointer, the
+// NUL-terminated string (but not the pointer) is printed.
+
+template <typename T>
+class UniversalTersePrinter {
+ public:
+  static void Print(const T& value, ::std::ostream* os) {
+    UniversalPrint(value, os);
+  }
+};
+template <typename T>
+class UniversalTersePrinter<T&> {
+ public:
+  static void Print(const T& value, ::std::ostream* os) {
+    UniversalPrint(value, os);
+  }
+};
+template <typename T, size_t N>
+class UniversalTersePrinter<T[N]> {
+ public:
+  static void Print(const T (&value)[N], ::std::ostream* os) {
+    UniversalPrinter<T[N]>::Print(value, os);
+  }
+};
+template <>
+class UniversalTersePrinter<const char*> {
+ public:
+  static void Print(const char* str, ::std::ostream* os) {
+    if (str == NULL) {
+      *os << "NULL";
+    } else {
+      UniversalPrint(string(str), os);
+    }
+  }
+};
+template <>
+class UniversalTersePrinter<char*> {
+ public:
+  static void Print(char* str, ::std::ostream* os) {
+    UniversalTersePrinter<const char*>::Print(str, os);
+  }
+};
+
+#if GTEST_HAS_STD_WSTRING
+template <>
+class UniversalTersePrinter<const wchar_t*> {
+ public:
+  static void Print(const wchar_t* str, ::std::ostream* os) {
+    if (str == NULL) {
+      *os << "NULL";
+    } else {
+      UniversalPrint(::std::wstring(str), os);
+    }
+  }
+};
+#endif
+
+template <>
+class UniversalTersePrinter<wchar_t*> {
+ public:
+  static void Print(wchar_t* str, ::std::ostream* os) {
+    UniversalTersePrinter<const wchar_t*>::Print(str, os);
+  }
+};
+
+template <typename T>
+void UniversalTersePrint(const T& value, ::std::ostream* os) {
+  UniversalTersePrinter<T>::Print(value, os);
+}
+
+// Prints a value using the type inferred by the compiler.  The
+// difference between this and UniversalTersePrint() is that for a
+// (const) char pointer, this prints both the pointer and the
+// NUL-terminated string.
+template <typename T>
+void UniversalPrint(const T& value, ::std::ostream* os) {
+  // A workarond for the bug in VC++ 7.1 that prevents us from instantiating
+  // UniversalPrinter with T directly.
+  typedef T T1;
+  UniversalPrinter<T1>::Print(value, os);
+}
+
+typedef ::std::vector<string> Strings;
+
+// TuplePolicy<TupleT> must provide:
+// - tuple_size
+//     size of tuple TupleT.
+// - get<size_t I>(const TupleT& t)
+//     static function extracting element I of tuple TupleT.
+// - tuple_element<size_t I>::type
+//     type of element I of tuple TupleT.
+template <typename TupleT>
+struct TuplePolicy;
+
+#if GTEST_HAS_TR1_TUPLE
+template <typename TupleT>
+struct TuplePolicy {
+  typedef TupleT Tuple;
+  static const size_t tuple_size = ::std::tr1::tuple_size<Tuple>::value;
+
+  template <size_t I>
+  struct tuple_element : ::std::tr1::tuple_element<I, Tuple> {};
+
+  template <size_t I>
+  static typename AddReference<
+      const typename ::std::tr1::tuple_element<I, Tuple>::type>::type get(
+      const Tuple& tuple) {
+    return ::std::tr1::get<I>(tuple);
+  }
+};
+template <typename TupleT>
+const size_t TuplePolicy<TupleT>::tuple_size;
+#endif  // GTEST_HAS_TR1_TUPLE
+
+#if GTEST_HAS_STD_TUPLE_
+template <typename... Types>
+struct TuplePolicy< ::std::tuple<Types...> > {
+  typedef ::std::tuple<Types...> Tuple;
+  static const size_t tuple_size = ::std::tuple_size<Tuple>::value;
+
+  template <size_t I>
+  struct tuple_element : ::std::tuple_element<I, Tuple> {};
+
+  template <size_t I>
+  static const typename ::std::tuple_element<I, Tuple>::type& get(
+      const Tuple& tuple) {
+    return ::std::get<I>(tuple);
+  }
+};
+template <typename... Types>
+const size_t TuplePolicy< ::std::tuple<Types...> >::tuple_size;
+#endif  // GTEST_HAS_STD_TUPLE_
+
+#if GTEST_HAS_TR1_TUPLE || GTEST_HAS_STD_TUPLE_
+// This helper template allows PrintTo() for tuples and
+// UniversalTersePrintTupleFieldsToStrings() to be defined by
+// induction on the number of tuple fields.  The idea is that
+// TuplePrefixPrinter<N>::PrintPrefixTo(t, os) prints the first N
+// fields in tuple t, and can be defined in terms of
+// TuplePrefixPrinter<N - 1>.
+//
+// The inductive case.
+template <size_t N>
+struct TuplePrefixPrinter {
+  // Prints the first N fields of a tuple.
+  template <typename Tuple>
+  static void PrintPrefixTo(const Tuple& t, ::std::ostream* os) {
+    TuplePrefixPrinter<N - 1>::PrintPrefixTo(t, os);
+    GTEST_INTENTIONAL_CONST_COND_PUSH_()
+    if (N > 1) {
+    GTEST_INTENTIONAL_CONST_COND_POP_()
+      *os << ", ";
+    }
+    UniversalPrinter<
+        typename TuplePolicy<Tuple>::template tuple_element<N - 1>::type>
+        ::Print(TuplePolicy<Tuple>::template get<N - 1>(t), os);
+  }
+
+  // Tersely prints the first N fields of a tuple to a string vector,
+  // one element for each field.
+  template <typename Tuple>
+  static void TersePrintPrefixToStrings(const Tuple& t, Strings* strings) {
+    TuplePrefixPrinter<N - 1>::TersePrintPrefixToStrings(t, strings);
+    ::std::stringstream ss;
+    UniversalTersePrint(TuplePolicy<Tuple>::template get<N - 1>(t), &ss);
+    strings->push_back(ss.str());
+  }
+};
+
+// Base case.
+template <>
+struct TuplePrefixPrinter<0> {
+  template <typename Tuple>
+  static void PrintPrefixTo(const Tuple&, ::std::ostream*) {}
+
+  template <typename Tuple>
+  static void TersePrintPrefixToStrings(const Tuple&, Strings*) {}
+};
+
+// Helper function for printing a tuple.
+// Tuple must be either std::tr1::tuple or std::tuple type.
+template <typename Tuple>
+void PrintTupleTo(const Tuple& t, ::std::ostream* os) {
+  *os << "(";
+  TuplePrefixPrinter<TuplePolicy<Tuple>::tuple_size>::PrintPrefixTo(t, os);
+  *os << ")";
+}
+
+// Prints the fields of a tuple tersely to a string vector, one
+// element for each field.  See the comment before
+// UniversalTersePrint() for how we define "tersely".
+template <typename Tuple>
+Strings UniversalTersePrintTupleFieldsToStrings(const Tuple& value) {
+  Strings result;
+  TuplePrefixPrinter<TuplePolicy<Tuple>::tuple_size>::
+      TersePrintPrefixToStrings(value, &result);
+  return result;
+}
+#endif  // GTEST_HAS_TR1_TUPLE || GTEST_HAS_STD_TUPLE_
+
+}  // namespace internal
+
+template <typename T>
+::std::string PrintToString(const T& value) {
+  ::std::stringstream ss;
+  internal::UniversalTersePrinter<T>::Print(value, &ss);
+  return ss.str();
+}
+
+}  // namespace testing
+
+// Include any custom printer added by the local installation.
+// We must include this header at the end to make sure it can use the
+// declarations from this file.
+#include "gtest/internal/custom/gtest-printers.h"
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PRINTERS_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-spi.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-spi.h
new file mode 100644
index 0000000000..f63fa9a1b2
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-spi.h
@@ -0,0 +1,232 @@
+// Copyright 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// Utilities for testing Google Test itself and code that uses Google Test
+// (e.g. frameworks built on top of Google Test).
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_SPI_H_
+#define GTEST_INCLUDE_GTEST_GTEST_SPI_H_
+
+#include "gtest/gtest.h"
+
+namespace testing {
+
+// This helper class can be used to mock out Google Test failure reporting
+// so that we can test Google Test or code that builds on Google Test.
+//
+// An object of this class appends a TestPartResult object to the
+// TestPartResultArray object given in the constructor whenever a Google Test
+// failure is reported. It can either intercept only failures that are
+// generated in the same thread that created this object or it can intercept
+// all generated failures. The scope of this mock object can be controlled with
+// the second argument to the two arguments constructor.
+class GTEST_API_ ScopedFakeTestPartResultReporter
+    : public TestPartResultReporterInterface {
+ public:
+  // The two possible mocking modes of this object.
+  enum InterceptMode {
+    INTERCEPT_ONLY_CURRENT_THREAD,  // Intercepts only thread local failures.
+    INTERCEPT_ALL_THREADS           // Intercepts all failures.
+  };
+
+  // The c'tor sets this object as the test part result reporter used
+  // by Google Test.  The 'result' parameter specifies where to report the
+  // results. This reporter will only catch failures generated in the current
+  // thread. DEPRECATED
+  explicit ScopedFakeTestPartResultReporter(TestPartResultArray* result);
+
+  // Same as above, but you can choose the interception scope of this object.
+  ScopedFakeTestPartResultReporter(InterceptMode intercept_mode,
+                                   TestPartResultArray* result);
+
+  // The d'tor restores the previous test part result reporter.
+  virtual ~ScopedFakeTestPartResultReporter();
+
+  // Appends the TestPartResult object to the TestPartResultArray
+  // received in the constructor.
+  //
+  // This method is from the TestPartResultReporterInterface
+  // interface.
+  virtual void ReportTestPartResult(const TestPartResult& result);
+ private:
+  void Init();
+
+  const InterceptMode intercept_mode_;
+  TestPartResultReporterInterface* old_reporter_;
+  TestPartResultArray* const result_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ScopedFakeTestPartResultReporter);
+};
+
+namespace internal {
+
+// A helper class for implementing EXPECT_FATAL_FAILURE() and
+// EXPECT_NONFATAL_FAILURE().  Its destructor verifies that the given
+// TestPartResultArray contains exactly one failure that has the given
+// type and contains the given substring.  If that's not the case, a
+// non-fatal failure will be generated.
+class GTEST_API_ SingleFailureChecker {
+ public:
+  // The constructor remembers the arguments.
+  SingleFailureChecker(const TestPartResultArray* results,
+                       TestPartResult::Type type,
+                       const string& substr);
+  ~SingleFailureChecker();
+ private:
+  const TestPartResultArray* const results_;
+  const TestPartResult::Type type_;
+  const string substr_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(SingleFailureChecker);
+};
+
+}  // namespace internal
+
+}  // namespace testing
+
+// A set of macros for testing Google Test assertions or code that's expected
+// to generate Google Test fatal failures.  It verifies that the given
+// statement will cause exactly one fatal Google Test failure with 'substr'
+// being part of the failure message.
+//
+// There are two different versions of this macro. EXPECT_FATAL_FAILURE only
+// affects and considers failures generated in the current thread and
+// EXPECT_FATAL_FAILURE_ON_ALL_THREADS does the same but for all threads.
+//
+// The verification of the assertion is done correctly even when the statement
+// throws an exception or aborts the current function.
+//
+// Known restrictions:
+//   - 'statement' cannot reference local non-static variables or
+//     non-static members of the current object.
+//   - 'statement' cannot return a value.
+//   - You cannot stream a failure message to this macro.
+//
+// Note that even though the implementations of the following two
+// macros are much alike, we cannot refactor them to use a common
+// helper macro, due to some peculiarity in how the preprocessor
+// works.  The AcceptsMacroThatExpandsToUnprotectedComma test in
+// gtest_unittest.cc will fail to compile if we do that.
+#define EXPECT_FATAL_FAILURE(statement, substr) \
+  do { \
+    class GTestExpectFatalFailureHelper {\
+     public:\
+      static void Execute() { statement; }\
+    };\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        &gtest_failures, ::testing::TestPartResult::kFatalFailure, (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter:: \
+          INTERCEPT_ONLY_CURRENT_THREAD, &gtest_failures);\
+      GTestExpectFatalFailureHelper::Execute();\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+#define EXPECT_FATAL_FAILURE_ON_ALL_THREADS(statement, substr) \
+  do { \
+    class GTestExpectFatalFailureHelper {\
+     public:\
+      static void Execute() { statement; }\
+    };\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        &gtest_failures, ::testing::TestPartResult::kFatalFailure, (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter:: \
+          INTERCEPT_ALL_THREADS, &gtest_failures);\
+      GTestExpectFatalFailureHelper::Execute();\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+// A macro for testing Google Test assertions or code that's expected to
+// generate Google Test non-fatal failures.  It asserts that the given
+// statement will cause exactly one non-fatal Google Test failure with 'substr'
+// being part of the failure message.
+//
+// There are two different versions of this macro. EXPECT_NONFATAL_FAILURE only
+// affects and considers failures generated in the current thread and
+// EXPECT_NONFATAL_FAILURE_ON_ALL_THREADS does the same but for all threads.
+//
+// 'statement' is allowed to reference local variables and members of
+// the current object.
+//
+// The verification of the assertion is done correctly even when the statement
+// throws an exception or aborts the current function.
+//
+// Known restrictions:
+//   - You cannot stream a failure message to this macro.
+//
+// Note that even though the implementations of the following two
+// macros are much alike, we cannot refactor them to use a common
+// helper macro, due to some peculiarity in how the preprocessor
+// works.  If we do that, the code won't compile when the user gives
+// EXPECT_NONFATAL_FAILURE() a statement that contains a macro that
+// expands to code containing an unprotected comma.  The
+// AcceptsMacroThatExpandsToUnprotectedComma test in gtest_unittest.cc
+// catches that.
+//
+// For the same reason, we have to write
+//   if (::testing::internal::AlwaysTrue()) { statement; }
+// instead of
+//   GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement)
+// to avoid an MSVC warning on unreachable code.
+#define EXPECT_NONFATAL_FAILURE(statement, substr) \
+  do {\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        &gtest_failures, ::testing::TestPartResult::kNonFatalFailure, \
+        (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter:: \
+          INTERCEPT_ONLY_CURRENT_THREAD, &gtest_failures);\
+      if (::testing::internal::AlwaysTrue()) { statement; }\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+#define EXPECT_NONFATAL_FAILURE_ON_ALL_THREADS(statement, substr) \
+  do {\
+    ::testing::TestPartResultArray gtest_failures;\
+    ::testing::internal::SingleFailureChecker gtest_checker(\
+        &gtest_failures, ::testing::TestPartResult::kNonFatalFailure, \
+        (substr));\
+    {\
+      ::testing::ScopedFakeTestPartResultReporter gtest_reporter(\
+          ::testing::ScopedFakeTestPartResultReporter::INTERCEPT_ALL_THREADS, \
+          &gtest_failures);\
+      if (::testing::internal::AlwaysTrue()) { statement; }\
+    }\
+  } while (::testing::internal::AlwaysFalse())
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_SPI_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-test-part.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-test-part.h
new file mode 100644
index 0000000000..77eb844839
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-test-part.h
@@ -0,0 +1,179 @@
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: mheule@google.com (Markus Heule)
+//
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_TEST_PART_H_
+#define GTEST_INCLUDE_GTEST_GTEST_TEST_PART_H_
+
+#include <iosfwd>
+#include <vector>
+#include "gtest/internal/gtest-internal.h"
+#include "gtest/internal/gtest-string.h"
+
+namespace testing {
+
+// A copyable object representing the result of a test part (i.e. an
+// assertion or an explicit FAIL(), ADD_FAILURE(), or SUCCESS()).
+//
+// Don't inherit from TestPartResult as its destructor is not virtual.
+class GTEST_API_ TestPartResult {
+ public:
+  // The possible outcomes of a test part (i.e. an assertion or an
+  // explicit SUCCEED(), FAIL(), or ADD_FAILURE()).
+  enum Type {
+    kSuccess,          // Succeeded.
+    kNonFatalFailure,  // Failed but the test can continue.
+    kFatalFailure      // Failed and the test should be terminated.
+  };
+
+  // C'tor.  TestPartResult does NOT have a default constructor.
+  // Always use this constructor (with parameters) to create a
+  // TestPartResult object.
+  TestPartResult(Type a_type,
+                 const char* a_file_name,
+                 int a_line_number,
+                 const char* a_message)
+      : type_(a_type),
+        file_name_(a_file_name == NULL ? "" : a_file_name),
+        line_number_(a_line_number),
+        summary_(ExtractSummary(a_message)),
+        message_(a_message) {
+  }
+
+  // Gets the outcome of the test part.
+  Type type() const { return type_; }
+
+  // Gets the name of the source file where the test part took place, or
+  // NULL if it's unknown.
+  const char* file_name() const {
+    return file_name_.empty() ? NULL : file_name_.c_str();
+  }
+
+  // Gets the line in the source file where the test part took place,
+  // or -1 if it's unknown.
+  int line_number() const { return line_number_; }
+
+  // Gets the summary of the failure message.
+  const char* summary() const { return summary_.c_str(); }
+
+  // Gets the message associated with the test part.
+  const char* message() const { return message_.c_str(); }
+
+  // Returns true iff the test part passed.
+  bool passed() const { return type_ == kSuccess; }
+
+  // Returns true iff the test part failed.
+  bool failed() const { return type_ != kSuccess; }
+
+  // Returns true iff the test part non-fatally failed.
+  bool nonfatally_failed() const { return type_ == kNonFatalFailure; }
+
+  // Returns true iff the test part fatally failed.
+  bool fatally_failed() const { return type_ == kFatalFailure; }
+
+ private:
+  Type type_;
+
+  // Gets the summary of the failure message by omitting the stack
+  // trace in it.
+  static std::string ExtractSummary(const char* message);
+
+  // The name of the source file where the test part took place, or
+  // "" if the source file is unknown.
+  std::string file_name_;
+  // The line in the source file where the test part took place, or -1
+  // if the line number is unknown.
+  int line_number_;
+  std::string summary_;  // The test failure summary.
+  std::string message_;  // The test failure message.
+};
+
+// Prints a TestPartResult object.
+std::ostream& operator<<(std::ostream& os, const TestPartResult& result);
+
+// An array of TestPartResult objects.
+//
+// Don't inherit from TestPartResultArray as its destructor is not
+// virtual.
+class GTEST_API_ TestPartResultArray {
+ public:
+  TestPartResultArray() {}
+
+  // Appends the given TestPartResult to the array.
+  void Append(const TestPartResult& result);
+
+  // Returns the TestPartResult at the given index (0-based).
+  const TestPartResult& GetTestPartResult(int index) const;
+
+  // Returns the number of TestPartResult objects in the array.
+  int size() const;
+
+ private:
+  std::vector<TestPartResult> array_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestPartResultArray);
+};
+
+// This interface knows how to report a test part result.
+class TestPartResultReporterInterface {
+ public:
+  virtual ~TestPartResultReporterInterface() {}
+
+  virtual void ReportTestPartResult(const TestPartResult& result) = 0;
+};
+
+namespace internal {
+
+// This helper class is used by {ASSERT|EXPECT}_NO_FATAL_FAILURE to check if a
+// statement generates new fatal failures. To do so it registers itself as the
+// current test part result reporter. Besides checking if fatal failures were
+// reported, it only delegates the reporting to the former result reporter.
+// The original result reporter is restored in the destructor.
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+class GTEST_API_ HasNewFatalFailureHelper
+    : public TestPartResultReporterInterface {
+ public:
+  HasNewFatalFailureHelper();
+  virtual ~HasNewFatalFailureHelper();
+  virtual void ReportTestPartResult(const TestPartResult& result);
+  bool has_new_fatal_failure() const { return has_new_fatal_failure_; }
+ private:
+  bool has_new_fatal_failure_;
+  TestPartResultReporterInterface* original_reporter_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(HasNewFatalFailureHelper);
+};
+
+}  // namespace internal
+
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_TEST_PART_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-typed-test.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-typed-test.h
new file mode 100644
index 0000000000..5f69d5678e
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest-typed-test.h
@@ -0,0 +1,263 @@
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_TYPED_TEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_TYPED_TEST_H_
+
+// This header implements typed tests and type-parameterized tests.
+
+// Typed (aka type-driven) tests repeat the same test for types in a
+// list.  You must know which types you want to test with when writing
+// typed tests. Here's how you do it:
+
+#if 0
+
+// First, define a fixture class template.  It should be parameterized
+// by a type.  Remember to derive it from testing::Test.
+template <typename T>
+class FooTest : public testing::Test {
+ public:
+  ...
+  typedef std::list<T> List;
+  static T shared_;
+  T value_;
+};
+
+// Next, associate a list of types with the test case, which will be
+// repeated for each type in the list.  The typedef is necessary for
+// the macro to parse correctly.
+typedef testing::Types<char, int, unsigned int> MyTypes;
+TYPED_TEST_CASE(FooTest, MyTypes);
+
+// If the type list contains only one type, you can write that type
+// directly without Types<...>:
+//   TYPED_TEST_CASE(FooTest, int);
+
+// Then, use TYPED_TEST() instead of TEST_F() to define as many typed
+// tests for this test case as you want.
+TYPED_TEST(FooTest, DoesBlah) {
+  // Inside a test, refer to TypeParam to get the type parameter.
+  // Since we are inside a derived class template, C++ requires use to
+  // visit the members of FooTest via 'this'.
+  TypeParam n = this->value_;
+
+  // To visit static members of the fixture, add the TestFixture::
+  // prefix.
+  n += TestFixture::shared_;
+
+  // To refer to typedefs in the fixture, add the "typename
+  // TestFixture::" prefix.
+  typename TestFixture::List values;
+  values.push_back(n);
+  ...
+}
+
+TYPED_TEST(FooTest, HasPropertyA) { ... }
+
+#endif  // 0
+
+// Type-parameterized tests are abstract test patterns parameterized
+// by a type.  Compared with typed tests, type-parameterized tests
+// allow you to define the test pattern without knowing what the type
+// parameters are.  The defined pattern can be instantiated with
+// different types any number of times, in any number of translation
+// units.
+//
+// If you are designing an interface or concept, you can define a
+// suite of type-parameterized tests to verify properties that any
+// valid implementation of the interface/concept should have.  Then,
+// each implementation can easily instantiate the test suite to verify
+// that it conforms to the requirements, without having to write
+// similar tests repeatedly.  Here's an example:
+
+#if 0
+
+// First, define a fixture class template.  It should be parameterized
+// by a type.  Remember to derive it from testing::Test.
+template <typename T>
+class FooTest : public testing::Test {
+  ...
+};
+
+// Next, declare that you will define a type-parameterized test case
+// (the _P suffix is for "parameterized" or "pattern", whichever you
+// prefer):
+TYPED_TEST_CASE_P(FooTest);
+
+// Then, use TYPED_TEST_P() to define as many type-parameterized tests
+// for this type-parameterized test case as you want.
+TYPED_TEST_P(FooTest, DoesBlah) {
+  // Inside a test, refer to TypeParam to get the type parameter.
+  TypeParam n = 0;
+  ...
+}
+
+TYPED_TEST_P(FooTest, HasPropertyA) { ... }
+
+// Now the tricky part: you need to register all test patterns before
+// you can instantiate them.  The first argument of the macro is the
+// test case name; the rest are the names of the tests in this test
+// case.
+REGISTER_TYPED_TEST_CASE_P(FooTest,
+                           DoesBlah, HasPropertyA);
+
+// Finally, you are free to instantiate the pattern with the types you
+// want.  If you put the above code in a header file, you can #include
+// it in multiple C++ source files and instantiate it multiple times.
+//
+// To distinguish different instances of the pattern, the first
+// argument to the INSTANTIATE_* macro is a prefix that will be added
+// to the actual test case name.  Remember to pick unique prefixes for
+// different instances.
+typedef testing::Types<char, int, unsigned int> MyTypes;
+INSTANTIATE_TYPED_TEST_CASE_P(My, FooTest, MyTypes);
+
+// If the type list contains only one type, you can write that type
+// directly without Types<...>:
+//   INSTANTIATE_TYPED_TEST_CASE_P(My, FooTest, int);
+
+#endif  // 0
+
+#include "gtest/internal/gtest-port.h"
+#include "gtest/internal/gtest-type-util.h"
+
+// Implements typed tests.
+
+#if GTEST_HAS_TYPED_TEST
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Expands to the name of the typedef for the type parameters of the
+// given test case.
+# define GTEST_TYPE_PARAMS_(TestCaseName) gtest_type_params_##TestCaseName##_
+
+// The 'Types' template argument below must have spaces around it
+// since some compilers may choke on '>>' when passing a template
+// instance (e.g. Types<int>)
+# define TYPED_TEST_CASE(CaseName, Types) \
+  typedef ::testing::internal::TypeList< Types >::type \
+      GTEST_TYPE_PARAMS_(CaseName)
+
+# define TYPED_TEST(CaseName, TestName) \
+  template <typename gtest_TypeParam_> \
+  class GTEST_TEST_CLASS_NAME_(CaseName, TestName) \
+      : public CaseName<gtest_TypeParam_> { \
+   private: \
+    typedef CaseName<gtest_TypeParam_> TestFixture; \
+    typedef gtest_TypeParam_ TypeParam; \
+    virtual void TestBody(); \
+  }; \
+  bool gtest_##CaseName##_##TestName##_registered_ GTEST_ATTRIBUTE_UNUSED_ = \
+      ::testing::internal::TypeParameterizedTest< \
+          CaseName, \
+          ::testing::internal::TemplateSel< \
+              GTEST_TEST_CLASS_NAME_(CaseName, TestName)>, \
+          GTEST_TYPE_PARAMS_(CaseName)>::Register(\
+              "", ::testing::internal::CodeLocation(__FILE__, __LINE__), \
+              #CaseName, #TestName, 0); \
+  template <typename gtest_TypeParam_> \
+  void GTEST_TEST_CLASS_NAME_(CaseName, TestName)<gtest_TypeParam_>::TestBody()
+
+#endif  // GTEST_HAS_TYPED_TEST
+
+// Implements type-parameterized tests.
+
+#if GTEST_HAS_TYPED_TEST_P
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Expands to the namespace name that the type-parameterized tests for
+// the given type-parameterized test case are defined in.  The exact
+// name of the namespace is subject to change without notice.
+# define GTEST_CASE_NAMESPACE_(TestCaseName) \
+  gtest_case_##TestCaseName##_
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Expands to the name of the variable used to remember the names of
+// the defined tests in the given test case.
+# define GTEST_TYPED_TEST_CASE_P_STATE_(TestCaseName) \
+  gtest_typed_test_case_p_state_##TestCaseName##_
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE DIRECTLY.
+//
+// Expands to the name of the variable used to remember the names of
+// the registered tests in the given test case.
+# define GTEST_REGISTERED_TEST_NAMES_(TestCaseName) \
+  gtest_registered_test_names_##TestCaseName##_
+
+// The variables defined in the type-parameterized test macros are
+// static as typically these macros are used in a .h file that can be
+// #included in multiple translation units linked together.
+# define TYPED_TEST_CASE_P(CaseName) \
+  static ::testing::internal::TypedTestCasePState \
+      GTEST_TYPED_TEST_CASE_P_STATE_(CaseName)
+
+# define TYPED_TEST_P(CaseName, TestName) \
+  namespace GTEST_CASE_NAMESPACE_(CaseName) { \
+  template <typename gtest_TypeParam_> \
+  class TestName : public CaseName<gtest_TypeParam_> { \
+   private: \
+    typedef CaseName<gtest_TypeParam_> TestFixture; \
+    typedef gtest_TypeParam_ TypeParam; \
+    virtual void TestBody(); \
+  }; \
+  static bool gtest_##TestName##_defined_ GTEST_ATTRIBUTE_UNUSED_ = \
+      GTEST_TYPED_TEST_CASE_P_STATE_(CaseName).AddTestName(\
+          __FILE__, __LINE__, #CaseName, #TestName); \
+  } \
+  template <typename gtest_TypeParam_> \
+  void GTEST_CASE_NAMESPACE_(CaseName)::TestName<gtest_TypeParam_>::TestBody()
+
+# define REGISTER_TYPED_TEST_CASE_P(CaseName, ...) \
+  namespace GTEST_CASE_NAMESPACE_(CaseName) { \
+  typedef ::testing::internal::Templates<__VA_ARGS__>::type gtest_AllTests_; \
+  } \
+  static const char* const GTEST_REGISTERED_TEST_NAMES_(CaseName) = \
+      GTEST_TYPED_TEST_CASE_P_STATE_(CaseName).VerifyRegisteredTestNames(\
+          __FILE__, __LINE__, #__VA_ARGS__)
+
+// The 'Types' template argument below must have spaces around it
+// since some compilers may choke on '>>' when passing a template
+// instance (e.g. Types<int>)
+# define INSTANTIATE_TYPED_TEST_CASE_P(Prefix, CaseName, Types) \
+  bool gtest_##Prefix##_##CaseName GTEST_ATTRIBUTE_UNUSED_ = \
+      ::testing::internal::TypeParameterizedTestCase<CaseName, \
+          GTEST_CASE_NAMESPACE_(CaseName)::gtest_AllTests_, \
+          ::testing::internal::TypeList< Types >::type>::Register(\
+              #Prefix, \
+              ::testing::internal::CodeLocation(__FILE__, __LINE__), \
+              &GTEST_TYPED_TEST_CASE_P_STATE_(CaseName), \
+              #CaseName, GTEST_REGISTERED_TEST_NAMES_(CaseName))
+
+#endif  // GTEST_HAS_TYPED_TEST_P
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_TYPED_TEST_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest.h
new file mode 100644
index 0000000000..f846c5bd66
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest.h
@@ -0,0 +1,2236 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines the public API for Google Test.  It should be
+// included by any test program that uses Google Test.
+//
+// IMPORTANT NOTE: Due to limitation of the C++ language, we have to
+// leave some internal implementation details in this header file.
+// They are clearly marked by comments like this:
+//
+//   // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+//
+// Such code is NOT meant to be used by a user directly, and is subject
+// to CHANGE WITHOUT NOTICE.  Therefore DO NOT DEPEND ON IT in a user
+// program!
+//
+// Acknowledgment: Google Test borrowed the idea of automatic test
+// registration from Barthelemy Dagenais' (barthelemy@prologique.com)
+// easyUnit framework.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_H_
+#define GTEST_INCLUDE_GTEST_GTEST_H_
+
+#include <limits>
+#include <ostream>
+#include <vector>
+
+#include "gtest/internal/gtest-internal.h"
+#include "gtest/internal/gtest-string.h"
+#include "gtest/gtest-death-test.h"
+#include "gtest/gtest-message.h"
+#include "gtest/gtest-param-test.h"
+#include "gtest/gtest-printers.h"
+#include "gtest/gtest_prod.h"
+#include "gtest/gtest-test-part.h"
+#include "gtest/gtest-typed-test.h"
+
+// Depending on the platform, different string classes are available.
+// On Linux, in addition to ::std::string, Google also makes use of
+// class ::string, which has the same interface as ::std::string, but
+// has a different implementation.
+//
+// You can define GTEST_HAS_GLOBAL_STRING to 1 to indicate that
+// ::string is available AND is a distinct type to ::std::string, or
+// define it to 0 to indicate otherwise.
+//
+// If ::std::string and ::string are the same class on your platform
+// due to aliasing, you should define GTEST_HAS_GLOBAL_STRING to 0.
+//
+// If you do not define GTEST_HAS_GLOBAL_STRING, it is defined
+// heuristically.
+
+namespace testing {
+
+// Declares the flags.
+
+// This flag temporary enables the disabled tests.
+GTEST_DECLARE_bool_(also_run_disabled_tests);
+
+// This flag brings the debugger on an assertion failure.
+GTEST_DECLARE_bool_(break_on_failure);
+
+// This flag controls whether Google Test catches all test-thrown exceptions
+// and logs them as failures.
+GTEST_DECLARE_bool_(catch_exceptions);
+
+// This flag enables using colors in terminal output. Available values are
+// "yes" to enable colors, "no" (disable colors), or "auto" (the default)
+// to let Google Test decide.
+GTEST_DECLARE_string_(color);
+
+// This flag sets up the filter to select by name using a glob pattern
+// the tests to run. If the filter is not given all tests are executed.
+GTEST_DECLARE_string_(filter);
+
+// This flag causes the Google Test to list tests. None of the tests listed
+// are actually run if the flag is provided.
+GTEST_DECLARE_bool_(list_tests);
+
+// This flag controls whether Google Test emits a detailed XML report to a file
+// in addition to its normal textual output.
+GTEST_DECLARE_string_(output);
+
+// This flags control whether Google Test prints the elapsed time for each
+// test.
+GTEST_DECLARE_bool_(print_time);
+
+// This flag specifies the random number seed.
+GTEST_DECLARE_int32_(random_seed);
+
+// This flag sets how many times the tests are repeated. The default value
+// is 1. If the value is -1 the tests are repeating forever.
+GTEST_DECLARE_int32_(repeat);
+
+// This flag controls whether Google Test includes Google Test internal
+// stack frames in failure stack traces.
+GTEST_DECLARE_bool_(show_internal_stack_frames);
+
+// When this flag is specified, tests' order is randomized on every iteration.
+GTEST_DECLARE_bool_(shuffle);
+
+// This flag specifies the maximum number of stack frames to be
+// printed in a failure message.
+GTEST_DECLARE_int32_(stack_trace_depth);
+
+// When this flag is specified, a failed assertion will throw an
+// exception if exceptions are enabled, or exit the program with a
+// non-zero code otherwise.
+GTEST_DECLARE_bool_(throw_on_failure);
+
+// When this flag is set with a "host:port" string, on supported
+// platforms test results are streamed to the specified port on
+// the specified host machine.
+GTEST_DECLARE_string_(stream_result_to);
+
+// The upper limit for valid stack trace depths.
+const int kMaxStackTraceDepth = 100;
+
+namespace internal {
+
+class AssertHelper;
+class DefaultGlobalTestPartResultReporter;
+class ExecDeathTest;
+class NoExecDeathTest;
+class FinalSuccessChecker;
+class GTestFlagSaver;
+class StreamingListenerTest;
+class TestResultAccessor;
+class TestEventListenersAccessor;
+class TestEventRepeater;
+class UnitTestRecordPropertyTestHelper;
+class WindowsDeathTest;
+class UnitTestImpl* GetUnitTestImpl();
+void ReportFailureInUnknownLocation(TestPartResult::Type result_type,
+                                    const std::string& message);
+
+}  // namespace internal
+
+// The friend relationship of some of these classes is cyclic.
+// If we don't forward declare them the compiler might confuse the classes
+// in friendship clauses with same named classes on the scope.
+class Test;
+class TestCase;
+class TestInfo;
+class UnitTest;
+
+// A class for indicating whether an assertion was successful.  When
+// the assertion wasn't successful, the AssertionResult object
+// remembers a non-empty message that describes how it failed.
+//
+// To create an instance of this class, use one of the factory functions
+// (AssertionSuccess() and AssertionFailure()).
+//
+// This class is useful for two purposes:
+//   1. Defining predicate functions to be used with Boolean test assertions
+//      EXPECT_TRUE/EXPECT_FALSE and their ASSERT_ counterparts
+//   2. Defining predicate-format functions to be
+//      used with predicate assertions (ASSERT_PRED_FORMAT*, etc).
+//
+// For example, if you define IsEven predicate:
+//
+//   testing::AssertionResult IsEven(int n) {
+//     if ((n % 2) == 0)
+//       return testing::AssertionSuccess();
+//     else
+//       return testing::AssertionFailure() << n << " is odd";
+//   }
+//
+// Then the failed expectation EXPECT_TRUE(IsEven(Fib(5)))
+// will print the message
+//
+//   Value of: IsEven(Fib(5))
+//     Actual: false (5 is odd)
+//   Expected: true
+//
+// instead of a more opaque
+//
+//   Value of: IsEven(Fib(5))
+//     Actual: false
+//   Expected: true
+//
+// in case IsEven is a simple Boolean predicate.
+//
+// If you expect your predicate to be reused and want to support informative
+// messages in EXPECT_FALSE and ASSERT_FALSE (negative assertions show up
+// about half as often as positive ones in our tests), supply messages for
+// both success and failure cases:
+//
+//   testing::AssertionResult IsEven(int n) {
+//     if ((n % 2) == 0)
+//       return testing::AssertionSuccess() << n << " is even";
+//     else
+//       return testing::AssertionFailure() << n << " is odd";
+//   }
+//
+// Then a statement EXPECT_FALSE(IsEven(Fib(6))) will print
+//
+//   Value of: IsEven(Fib(6))
+//     Actual: true (8 is even)
+//   Expected: false
+//
+// NB: Predicates that support negative Boolean assertions have reduced
+// performance in positive ones so be careful not to use them in tests
+// that have lots (tens of thousands) of positive Boolean assertions.
+//
+// To use this class with EXPECT_PRED_FORMAT assertions such as:
+//
+//   // Verifies that Foo() returns an even number.
+//   EXPECT_PRED_FORMAT1(IsEven, Foo());
+//
+// you need to define:
+//
+//   testing::AssertionResult IsEven(const char* expr, int n) {
+//     if ((n % 2) == 0)
+//       return testing::AssertionSuccess();
+//     else
+//       return testing::AssertionFailure()
+//         << "Expected: " << expr << " is even\n  Actual: it's " << n;
+//   }
+//
+// If Foo() returns 5, you will see the following message:
+//
+//   Expected: Foo() is even
+//     Actual: it's 5
+//
+class GTEST_API_ AssertionResult {
+ public:
+  // Copy constructor.
+  // Used in EXPECT_TRUE/FALSE(assertion_result).
+  AssertionResult(const AssertionResult& other);
+
+  GTEST_DISABLE_MSC_WARNINGS_PUSH_(4800 /* forcing value to bool */)
+
+  // Used in the EXPECT_TRUE/FALSE(bool_expression).
+  //
+  // T must be contextually convertible to bool.
+  //
+  // The second parameter prevents this overload from being considered if
+  // the argument is implicitly convertible to AssertionResult. In that case
+  // we want AssertionResult's copy constructor to be used.
+  template <typename T>
+  explicit AssertionResult(
+      const T& success,
+      typename internal::EnableIf<
+          !internal::ImplicitlyConvertible<T, AssertionResult>::value>::type*
+          /*enabler*/ = NULL)
+      : success_(success) {}
+
+  GTEST_DISABLE_MSC_WARNINGS_POP_()
+
+  // Assignment operator.
+  AssertionResult& operator=(AssertionResult other) {
+    swap(other);
+    return *this;
+  }
+
+  // Returns true iff the assertion succeeded.
+  operator bool() const { return success_; }  // NOLINT
+
+  // Returns the assertion's negation. Used with EXPECT/ASSERT_FALSE.
+  AssertionResult operator!() const;
+
+  // Returns the text streamed into this AssertionResult. Test assertions
+  // use it when they fail (i.e., the predicate's outcome doesn't match the
+  // assertion's expectation). When nothing has been streamed into the
+  // object, returns an empty string.
+  const char* message() const {
+    return message_.get() != NULL ?  message_->c_str() : "";
+  }
+  // TODO(vladl@google.com): Remove this after making sure no clients use it.
+  // Deprecated; please use message() instead.
+  const char* failure_message() const { return message(); }
+
+  // Streams a custom failure message into this object.
+  template <typename T> AssertionResult& operator<<(const T& value) {
+    AppendMessage(Message() << value);
+    return *this;
+  }
+
+  // Allows streaming basic output manipulators such as endl or flush into
+  // this object.
+  AssertionResult& operator<<(
+      ::std::ostream& (*basic_manipulator)(::std::ostream& stream)) {
+    AppendMessage(Message() << basic_manipulator);
+    return *this;
+  }
+
+ private:
+  // Appends the contents of message to message_.
+  void AppendMessage(const Message& a_message) {
+    if (message_.get() == NULL)
+      message_.reset(new ::std::string);
+    message_->append(a_message.GetString().c_str());
+  }
+
+  // Swap the contents of this AssertionResult with other.
+  void swap(AssertionResult& other);
+
+  // Stores result of the assertion predicate.
+  bool success_;
+  // Stores the message describing the condition in case the expectation
+  // construct is not satisfied with the predicate's outcome.
+  // Referenced via a pointer to avoid taking too much stack frame space
+  // with test assertions.
+  internal::scoped_ptr< ::std::string> message_;
+};
+
+// Makes a successful assertion result.
+GTEST_API_ AssertionResult AssertionSuccess();
+
+// Makes a failed assertion result.
+GTEST_API_ AssertionResult AssertionFailure();
+
+// Makes a failed assertion result with the given failure message.
+// Deprecated; use AssertionFailure() << msg.
+GTEST_API_ AssertionResult AssertionFailure(const Message& msg);
+
+// The abstract class that all tests inherit from.
+//
+// In Google Test, a unit test program contains one or many TestCases, and
+// each TestCase contains one or many Tests.
+//
+// When you define a test using the TEST macro, you don't need to
+// explicitly derive from Test - the TEST macro automatically does
+// this for you.
+//
+// The only time you derive from Test is when defining a test fixture
+// to be used a TEST_F.  For example:
+//
+//   class FooTest : public testing::Test {
+//    protected:
+//     void SetUp() override { ... }
+//     void TearDown() override { ... }
+//     ...
+//   };
+//
+//   TEST_F(FooTest, Bar) { ... }
+//   TEST_F(FooTest, Baz) { ... }
+//
+// Test is not copyable.
+class GTEST_API_ Test {
+ public:
+  friend class TestInfo;
+
+  // Defines types for pointers to functions that set up and tear down
+  // a test case.
+  typedef internal::SetUpTestCaseFunc SetUpTestCaseFunc;
+  typedef internal::TearDownTestCaseFunc TearDownTestCaseFunc;
+
+  // The d'tor is virtual as we intend to inherit from Test.
+  virtual ~Test();
+
+  // Sets up the stuff shared by all tests in this test case.
+  //
+  // Google Test will call Foo::SetUpTestCase() before running the first
+  // test in test case Foo.  Hence a sub-class can define its own
+  // SetUpTestCase() method to shadow the one defined in the super
+  // class.
+  static void SetUpTestCase() {}
+
+  // Tears down the stuff shared by all tests in this test case.
+  //
+  // Google Test will call Foo::TearDownTestCase() after running the last
+  // test in test case Foo.  Hence a sub-class can define its own
+  // TearDownTestCase() method to shadow the one defined in the super
+  // class.
+  static void TearDownTestCase() {}
+
+  // Returns true iff the current test has a fatal failure.
+  static bool HasFatalFailure();
+
+  // Returns true iff the current test has a non-fatal failure.
+  static bool HasNonfatalFailure();
+
+  // Returns true iff the current test has a (either fatal or
+  // non-fatal) failure.
+  static bool HasFailure() { return HasFatalFailure() || HasNonfatalFailure(); }
+
+  // Logs a property for the current test, test case, or for the entire
+  // invocation of the test program when used outside of the context of a
+  // test case.  Only the last value for a given key is remembered.  These
+  // are public static so they can be called from utility functions that are
+  // not members of the test fixture.  Calls to RecordProperty made during
+  // lifespan of the test (from the moment its constructor starts to the
+  // moment its destructor finishes) will be output in XML as attributes of
+  // the <testcase> element.  Properties recorded from fixture's
+  // SetUpTestCase or TearDownTestCase are logged as attributes of the
+  // corresponding <testsuite> element.  Calls to RecordProperty made in the
+  // global context (before or after invocation of RUN_ALL_TESTS and from
+  // SetUp/TearDown method of Environment objects registered with Google
+  // Test) will be output as attributes of the <testsuites> element.
+  static void RecordProperty(const std::string& key, const std::string& value);
+  static void RecordProperty(const std::string& key, int value);
+
+ protected:
+  // Creates a Test object.
+  Test();
+
+  // Sets up the test fixture.
+  virtual void SetUp();
+
+  // Tears down the test fixture.
+  virtual void TearDown();
+
+ private:
+  // Returns true iff the current test has the same fixture class as
+  // the first test in the current test case.
+  static bool HasSameFixtureClass();
+
+  // Runs the test after the test fixture has been set up.
+  //
+  // A sub-class must implement this to define the test logic.
+  //
+  // DO NOT OVERRIDE THIS FUNCTION DIRECTLY IN A USER PROGRAM.
+  // Instead, use the TEST or TEST_F macro.
+  virtual void TestBody() = 0;
+
+  // Sets up, executes, and tears down the test.
+  void Run();
+
+  // Deletes self.  We deliberately pick an unusual name for this
+  // internal method to avoid clashing with names used in user TESTs.
+  void DeleteSelf_() { delete this; }
+
+  const internal::scoped_ptr< GTEST_FLAG_SAVER_ > gtest_flag_saver_;
+
+  // Often a user misspells SetUp() as Setup() and spends a long time
+  // wondering why it is never called by Google Test.  The declaration of
+  // the following method is solely for catching such an error at
+  // compile time:
+  //
+  //   - The return type is deliberately chosen to be not void, so it
+  //   will be a conflict if void Setup() is declared in the user's
+  //   test fixture.
+  //
+  //   - This method is private, so it will be another compiler error
+  //   if the method is called from the user's test fixture.
+  //
+  // DO NOT OVERRIDE THIS FUNCTION.
+  //
+  // If you see an error about overriding the following function or
+  // about it being private, you have mis-spelled SetUp() as Setup().
+  struct Setup_should_be_spelled_SetUp {};
+  virtual Setup_should_be_spelled_SetUp* Setup() { return NULL; }
+
+  // We disallow copying Tests.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Test);
+};
+
+typedef internal::TimeInMillis TimeInMillis;
+
+// A copyable object representing a user specified test property which can be
+// output as a key/value string pair.
+//
+// Don't inherit from TestProperty as its destructor is not virtual.
+class TestProperty {
+ public:
+  // C'tor.  TestProperty does NOT have a default constructor.
+  // Always use this constructor (with parameters) to create a
+  // TestProperty object.
+  TestProperty(const std::string& a_key, const std::string& a_value) :
+    key_(a_key), value_(a_value) {
+  }
+
+  // Gets the user supplied key.
+  const char* key() const {
+    return key_.c_str();
+  }
+
+  // Gets the user supplied value.
+  const char* value() const {
+    return value_.c_str();
+  }
+
+  // Sets a new value, overriding the one supplied in the constructor.
+  void SetValue(const std::string& new_value) {
+    value_ = new_value;
+  }
+
+ private:
+  // The key supplied by the user.
+  std::string key_;
+  // The value supplied by the user.
+  std::string value_;
+};
+
+// The result of a single Test.  This includes a list of
+// TestPartResults, a list of TestProperties, a count of how many
+// death tests there are in the Test, and how much time it took to run
+// the Test.
+//
+// TestResult is not copyable.
+class GTEST_API_ TestResult {
+ public:
+  // Creates an empty TestResult.
+  TestResult();
+
+  // D'tor.  Do not inherit from TestResult.
+  ~TestResult();
+
+  // Gets the number of all test parts.  This is the sum of the number
+  // of successful test parts and the number of failed test parts.
+  int total_part_count() const;
+
+  // Returns the number of the test properties.
+  int test_property_count() const;
+
+  // Returns true iff the test passed (i.e. no test part failed).
+  bool Passed() const { return !Failed(); }
+
+  // Returns true iff the test failed.
+  bool Failed() const;
+
+  // Returns true iff the test fatally failed.
+  bool HasFatalFailure() const;
+
+  // Returns true iff the test has a non-fatal failure.
+  bool HasNonfatalFailure() const;
+
+  // Returns the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const { return elapsed_time_; }
+
+  // Returns the i-th test part result among all the results. i can range
+  // from 0 to test_property_count() - 1. If i is not in that range, aborts
+  // the program.
+  const TestPartResult& GetTestPartResult(int i) const;
+
+  // Returns the i-th test property. i can range from 0 to
+  // test_property_count() - 1. If i is not in that range, aborts the
+  // program.
+  const TestProperty& GetTestProperty(int i) const;
+
+ private:
+  friend class TestInfo;
+  friend class TestCase;
+  friend class UnitTest;
+  friend class internal::DefaultGlobalTestPartResultReporter;
+  friend class internal::ExecDeathTest;
+  friend class internal::TestResultAccessor;
+  friend class internal::UnitTestImpl;
+  friend class internal::WindowsDeathTest;
+
+  // Gets the vector of TestPartResults.
+  const std::vector<TestPartResult>& test_part_results() const {
+    return test_part_results_;
+  }
+
+  // Gets the vector of TestProperties.
+  const std::vector<TestProperty>& test_properties() const {
+    return test_properties_;
+  }
+
+  // Sets the elapsed time.
+  void set_elapsed_time(TimeInMillis elapsed) { elapsed_time_ = elapsed; }
+
+  // Adds a test property to the list. The property is validated and may add
+  // a non-fatal failure if invalid (e.g., if it conflicts with reserved
+  // key names). If a property is already recorded for the same key, the
+  // value will be updated, rather than storing multiple values for the same
+  // key.  xml_element specifies the element for which the property is being
+  // recorded and is used for validation.
+  void RecordProperty(const std::string& xml_element,
+                      const TestProperty& test_property);
+
+  // Adds a failure if the key is a reserved attribute of Google Test
+  // testcase tags.  Returns true if the property is valid.
+  // TODO(russr): Validate attribute names are legal and human readable.
+  static bool ValidateTestProperty(const std::string& xml_element,
+                                   const TestProperty& test_property);
+
+  // Adds a test part result to the list.
+  void AddTestPartResult(const TestPartResult& test_part_result);
+
+  // Returns the death test count.
+  int death_test_count() const { return death_test_count_; }
+
+  // Increments the death test count, returning the new count.
+  int increment_death_test_count() { return ++death_test_count_; }
+
+  // Clears the test part results.
+  void ClearTestPartResults();
+
+  // Clears the object.
+  void Clear();
+
+  // Protects mutable state of the property vector and of owned
+  // properties, whose values may be updated.
+  internal::Mutex test_properites_mutex_;
+
+  // The vector of TestPartResults
+  std::vector<TestPartResult> test_part_results_;
+  // The vector of TestProperties
+  std::vector<TestProperty> test_properties_;
+  // Running count of death tests.
+  int death_test_count_;
+  // The elapsed time, in milliseconds.
+  TimeInMillis elapsed_time_;
+
+  // We disallow copying TestResult.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestResult);
+};  // class TestResult
+
+// A TestInfo object stores the following information about a test:
+//
+//   Test case name
+//   Test name
+//   Whether the test should be run
+//   A function pointer that creates the test object when invoked
+//   Test result
+//
+// The constructor of TestInfo registers itself with the UnitTest
+// singleton such that the RUN_ALL_TESTS() macro knows which tests to
+// run.
+class GTEST_API_ TestInfo {
+ public:
+  // Destructs a TestInfo object.  This function is not virtual, so
+  // don't inherit from TestInfo.
+  ~TestInfo();
+
+  // Returns the test case name.
+  const char* test_case_name() const { return test_case_name_.c_str(); }
+
+  // Returns the test name.
+  const char* name() const { return name_.c_str(); }
+
+  // Returns the name of the parameter type, or NULL if this is not a typed
+  // or a type-parameterized test.
+  const char* type_param() const {
+    if (type_param_.get() != NULL)
+      return type_param_->c_str();
+    return NULL;
+  }
+
+  // Returns the text representation of the value parameter, or NULL if this
+  // is not a value-parameterized test.
+  const char* value_param() const {
+    if (value_param_.get() != NULL)
+      return value_param_->c_str();
+    return NULL;
+  }
+
+  // Returns the file name where this test is defined.
+  const char* file() const { return location_.file.c_str(); }
+
+  // Returns the line where this test is defined.
+  int line() const { return location_.line; }
+
+  // Returns true if this test should run, that is if the test is not
+  // disabled (or it is disabled but the also_run_disabled_tests flag has
+  // been specified) and its full name matches the user-specified filter.
+  //
+  // Google Test allows the user to filter the tests by their full names.
+  // The full name of a test Bar in test case Foo is defined as
+  // "Foo.Bar".  Only the tests that match the filter will run.
+  //
+  // A filter is a colon-separated list of glob (not regex) patterns,
+  // optionally followed by a '-' and a colon-separated list of
+  // negative patterns (tests to exclude).  A test is run if it
+  // matches one of the positive patterns and does not match any of
+  // the negative patterns.
+  //
+  // For example, *A*:Foo.* is a filter that matches any string that
+  // contains the character 'A' or starts with "Foo.".
+  bool should_run() const { return should_run_; }
+
+  // Returns true iff this test will appear in the XML report.
+  bool is_reportable() const {
+    // For now, the XML report includes all tests matching the filter.
+    // In the future, we may trim tests that are excluded because of
+    // sharding.
+    return matches_filter_;
+  }
+
+  // Returns the result of the test.
+  const TestResult* result() const { return &result_; }
+
+ private:
+#if GTEST_HAS_DEATH_TEST
+  friend class internal::DefaultDeathTestFactory;
+#endif  // GTEST_HAS_DEATH_TEST
+  friend class Test;
+  friend class TestCase;
+  friend class internal::UnitTestImpl;
+  friend class internal::StreamingListenerTest;
+  friend TestInfo* internal::MakeAndRegisterTestInfo(
+      const char* test_case_name,
+      const char* name,
+      const char* type_param,
+      const char* value_param,
+      internal::CodeLocation code_location,
+      internal::TypeId fixture_class_id,
+      Test::SetUpTestCaseFunc set_up_tc,
+      Test::TearDownTestCaseFunc tear_down_tc,
+      internal::TestFactoryBase* factory);
+
+  // Constructs a TestInfo object. The newly constructed instance assumes
+  // ownership of the factory object.
+  TestInfo(const std::string& test_case_name,
+           const std::string& name,
+           const char* a_type_param,   // NULL if not a type-parameterized test
+           const char* a_value_param,  // NULL if not a value-parameterized test
+           internal::CodeLocation a_code_location,
+           internal::TypeId fixture_class_id,
+           internal::TestFactoryBase* factory);
+
+  // Increments the number of death tests encountered in this test so
+  // far.
+  int increment_death_test_count() {
+    return result_.increment_death_test_count();
+  }
+
+  // Creates the test object, runs it, records its result, and then
+  // deletes it.
+  void Run();
+
+  static void ClearTestResult(TestInfo* test_info) {
+    test_info->result_.Clear();
+  }
+
+  // These fields are immutable properties of the test.
+  const std::string test_case_name_;     // Test case name
+  const std::string name_;               // Test name
+  // Name of the parameter type, or NULL if this is not a typed or a
+  // type-parameterized test.
+  const internal::scoped_ptr<const ::std::string> type_param_;
+  // Text representation of the value parameter, or NULL if this is not a
+  // value-parameterized test.
+  const internal::scoped_ptr<const ::std::string> value_param_;
+  internal::CodeLocation location_;
+  const internal::TypeId fixture_class_id_;   // ID of the test fixture class
+  bool should_run_;                 // True iff this test should run
+  bool is_disabled_;                // True iff this test is disabled
+  bool matches_filter_;             // True if this test matches the
+                                    // user-specified filter.
+  internal::TestFactoryBase* const factory_;  // The factory that creates
+                                              // the test object
+
+  // This field is mutable and needs to be reset before running the
+  // test for the second time.
+  TestResult result_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestInfo);
+};
+
+// A test case, which consists of a vector of TestInfos.
+//
+// TestCase is not copyable.
+class GTEST_API_ TestCase {
+ public:
+  // Creates a TestCase with the given name.
+  //
+  // TestCase does NOT have a default constructor.  Always use this
+  // constructor to create a TestCase object.
+  //
+  // Arguments:
+  //
+  //   name:         name of the test case
+  //   a_type_param: the name of the test's type parameter, or NULL if
+  //                 this is not a type-parameterized test.
+  //   set_up_tc:    pointer to the function that sets up the test case
+  //   tear_down_tc: pointer to the function that tears down the test case
+  TestCase(const char* name, const char* a_type_param,
+           Test::SetUpTestCaseFunc set_up_tc,
+           Test::TearDownTestCaseFunc tear_down_tc);
+
+  // Destructor of TestCase.
+  virtual ~TestCase();
+
+  // Gets the name of the TestCase.
+  const char* name() const { return name_.c_str(); }
+
+  // Returns the name of the parameter type, or NULL if this is not a
+  // type-parameterized test case.
+  const char* type_param() const {
+    if (type_param_.get() != NULL)
+      return type_param_->c_str();
+    return NULL;
+  }
+
+  // Returns true if any test in this test case should run.
+  bool should_run() const { return should_run_; }
+
+  // Gets the number of successful tests in this test case.
+  int successful_test_count() const;
+
+  // Gets the number of failed tests in this test case.
+  int failed_test_count() const;
+
+  // Gets the number of disabled tests that will be reported in the XML report.
+  int reportable_disabled_test_count() const;
+
+  // Gets the number of disabled tests in this test case.
+  int disabled_test_count() const;
+
+  // Gets the number of tests to be printed in the XML report.
+  int reportable_test_count() const;
+
+  // Get the number of tests in this test case that should run.
+  int test_to_run_count() const;
+
+  // Gets the number of all tests in this test case.
+  int total_test_count() const;
+
+  // Returns true iff the test case passed.
+  bool Passed() const { return !Failed(); }
+
+  // Returns true iff the test case failed.
+  bool Failed() const { return failed_test_count() > 0; }
+
+  // Returns the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const { return elapsed_time_; }
+
+  // Returns the i-th test among all the tests. i can range from 0 to
+  // total_test_count() - 1. If i is not in that range, returns NULL.
+  const TestInfo* GetTestInfo(int i) const;
+
+  // Returns the TestResult that holds test properties recorded during
+  // execution of SetUpTestCase and TearDownTestCase.
+  const TestResult& ad_hoc_test_result() const { return ad_hoc_test_result_; }
+
+ private:
+  friend class Test;
+  friend class internal::UnitTestImpl;
+
+  // Gets the (mutable) vector of TestInfos in this TestCase.
+  std::vector<TestInfo*>& test_info_list() { return test_info_list_; }
+
+  // Gets the (immutable) vector of TestInfos in this TestCase.
+  const std::vector<TestInfo*>& test_info_list() const {
+    return test_info_list_;
+  }
+
+  // Returns the i-th test among all the tests. i can range from 0 to
+  // total_test_count() - 1. If i is not in that range, returns NULL.
+  TestInfo* GetMutableTestInfo(int i);
+
+  // Sets the should_run member.
+  void set_should_run(bool should) { should_run_ = should; }
+
+  // Adds a TestInfo to this test case.  Will delete the TestInfo upon
+  // destruction of the TestCase object.
+  void AddTestInfo(TestInfo * test_info);
+
+  // Clears the results of all tests in this test case.
+  void ClearResult();
+
+  // Clears the results of all tests in the given test case.
+  static void ClearTestCaseResult(TestCase* test_case) {
+    test_case->ClearResult();
+  }
+
+  // Runs every test in this TestCase.
+  void Run();
+
+  // Runs SetUpTestCase() for this TestCase.  This wrapper is needed
+  // for catching exceptions thrown from SetUpTestCase().
+  void RunSetUpTestCase() { (*set_up_tc_)(); }
+
+  // Runs TearDownTestCase() for this TestCase.  This wrapper is
+  // needed for catching exceptions thrown from TearDownTestCase().
+  void RunTearDownTestCase() { (*tear_down_tc_)(); }
+
+  // Returns true iff test passed.
+  static bool TestPassed(const TestInfo* test_info) {
+    return test_info->should_run() && test_info->result()->Passed();
+  }
+
+  // Returns true iff test failed.
+  static bool TestFailed(const TestInfo* test_info) {
+    return test_info->should_run() && test_info->result()->Failed();
+  }
+
+  // Returns true iff the test is disabled and will be reported in the XML
+  // report.
+  static bool TestReportableDisabled(const TestInfo* test_info) {
+    return test_info->is_reportable() && test_info->is_disabled_;
+  }
+
+  // Returns true iff test is disabled.
+  static bool TestDisabled(const TestInfo* test_info) {
+    return test_info->is_disabled_;
+  }
+
+  // Returns true iff this test will appear in the XML report.
+  static bool TestReportable(const TestInfo* test_info) {
+    return test_info->is_reportable();
+  }
+
+  // Returns true if the given test should run.
+  static bool ShouldRunTest(const TestInfo* test_info) {
+    return test_info->should_run();
+  }
+
+  // Shuffles the tests in this test case.
+  void ShuffleTests(internal::Random* random);
+
+  // Restores the test order to before the first shuffle.
+  void UnshuffleTests();
+
+  // Name of the test case.
+  std::string name_;
+  // Name of the parameter type, or NULL if this is not a typed or a
+  // type-parameterized test.
+  const internal::scoped_ptr<const ::std::string> type_param_;
+  // The vector of TestInfos in their original order.  It owns the
+  // elements in the vector.
+  std::vector<TestInfo*> test_info_list_;
+  // Provides a level of indirection for the test list to allow easy
+  // shuffling and restoring the test order.  The i-th element in this
+  // vector is the index of the i-th test in the shuffled test list.
+  std::vector<int> test_indices_;
+  // Pointer to the function that sets up the test case.
+  Test::SetUpTestCaseFunc set_up_tc_;
+  // Pointer to the function that tears down the test case.
+  Test::TearDownTestCaseFunc tear_down_tc_;
+  // True iff any test in this test case should run.
+  bool should_run_;
+  // Elapsed time, in milliseconds.
+  TimeInMillis elapsed_time_;
+  // Holds test properties recorded during execution of SetUpTestCase and
+  // TearDownTestCase.
+  TestResult ad_hoc_test_result_;
+
+  // We disallow copying TestCases.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestCase);
+};
+
+// An Environment object is capable of setting up and tearing down an
+// environment.  You should subclass this to define your own
+// environment(s).
+//
+// An Environment object does the set-up and tear-down in virtual
+// methods SetUp() and TearDown() instead of the constructor and the
+// destructor, as:
+//
+//   1. You cannot safely throw from a destructor.  This is a problem
+//      as in some cases Google Test is used where exceptions are enabled, and
+//      we may want to implement ASSERT_* using exceptions where they are
+//      available.
+//   2. You cannot use ASSERT_* directly in a constructor or
+//      destructor.
+class Environment {
+ public:
+  // The d'tor is virtual as we need to subclass Environment.
+  virtual ~Environment() {}
+
+  // Override this to define how to set up the environment.
+  virtual void SetUp() {}
+
+  // Override this to define how to tear down the environment.
+  virtual void TearDown() {}
+ private:
+  // If you see an error about overriding the following function or
+  // about it being private, you have mis-spelled SetUp() as Setup().
+  struct Setup_should_be_spelled_SetUp {};
+  virtual Setup_should_be_spelled_SetUp* Setup() { return NULL; }
+};
+
+// The interface for tracing execution of tests. The methods are organized in
+// the order the corresponding events are fired.
+class TestEventListener {
+ public:
+  virtual ~TestEventListener() {}
+
+  // Fired before any test activity starts.
+  virtual void OnTestProgramStart(const UnitTest& unit_test) = 0;
+
+  // Fired before each iteration of tests starts.  There may be more than
+  // one iteration if GTEST_FLAG(repeat) is set. iteration is the iteration
+  // index, starting from 0.
+  virtual void OnTestIterationStart(const UnitTest& unit_test,
+                                    int iteration) = 0;
+
+  // Fired before environment set-up for each iteration of tests starts.
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& unit_test) = 0;
+
+  // Fired after environment set-up for each iteration of tests ends.
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& unit_test) = 0;
+
+  // Fired before the test case starts.
+  virtual void OnTestCaseStart(const TestCase& test_case) = 0;
+
+  // Fired before the test starts.
+  virtual void OnTestStart(const TestInfo& test_info) = 0;
+
+  // Fired after a failed assertion or a SUCCEED() invocation.
+  virtual void OnTestPartResult(const TestPartResult& test_part_result) = 0;
+
+  // Fired after the test ends.
+  virtual void OnTestEnd(const TestInfo& test_info) = 0;
+
+  // Fired after the test case ends.
+  virtual void OnTestCaseEnd(const TestCase& test_case) = 0;
+
+  // Fired before environment tear-down for each iteration of tests starts.
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& unit_test) = 0;
+
+  // Fired after environment tear-down for each iteration of tests ends.
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& unit_test) = 0;
+
+  // Fired after each iteration of tests finishes.
+  virtual void OnTestIterationEnd(const UnitTest& unit_test,
+                                  int iteration) = 0;
+
+  // Fired after all test activities have ended.
+  virtual void OnTestProgramEnd(const UnitTest& unit_test) = 0;
+};
+
+// The convenience class for users who need to override just one or two
+// methods and are not concerned that a possible change to a signature of
+// the methods they override will not be caught during the build.  For
+// comments about each method please see the definition of TestEventListener
+// above.
+class EmptyTestEventListener : public TestEventListener {
+ public:
+  virtual void OnTestProgramStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationStart(const UnitTest& /*unit_test*/,
+                                    int /*iteration*/) {}
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestCaseStart(const TestCase& /*test_case*/) {}
+  virtual void OnTestStart(const TestInfo& /*test_info*/) {}
+  virtual void OnTestPartResult(const TestPartResult& /*test_part_result*/) {}
+  virtual void OnTestEnd(const TestInfo& /*test_info*/) {}
+  virtual void OnTestCaseEnd(const TestCase& /*test_case*/) {}
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationEnd(const UnitTest& /*unit_test*/,
+                                  int /*iteration*/) {}
+  virtual void OnTestProgramEnd(const UnitTest& /*unit_test*/) {}
+};
+
+// TestEventListeners lets users add listeners to track events in Google Test.
+class GTEST_API_ TestEventListeners {
+ public:
+  TestEventListeners();
+  ~TestEventListeners();
+
+  // Appends an event listener to the end of the list. Google Test assumes
+  // the ownership of the listener (i.e. it will delete the listener when
+  // the test program finishes).
+  void Append(TestEventListener* listener);
+
+  // Removes the given event listener from the list and returns it.  It then
+  // becomes the caller's responsibility to delete the listener. Returns
+  // NULL if the listener is not found in the list.
+  TestEventListener* Release(TestEventListener* listener);
+
+  // Returns the standard listener responsible for the default console
+  // output.  Can be removed from the listeners list to shut down default
+  // console output.  Note that removing this object from the listener list
+  // with Release transfers its ownership to the caller and makes this
+  // function return NULL the next time.
+  TestEventListener* default_result_printer() const {
+    return default_result_printer_;
+  }
+
+  // Returns the standard listener responsible for the default XML output
+  // controlled by the --gtest_output=xml flag.  Can be removed from the
+  // listeners list by users who want to shut down the default XML output
+  // controlled by this flag and substitute it with custom one.  Note that
+  // removing this object from the listener list with Release transfers its
+  // ownership to the caller and makes this function return NULL the next
+  // time.
+  TestEventListener* default_xml_generator() const {
+    return default_xml_generator_;
+  }
+
+ private:
+  friend class TestCase;
+  friend class TestInfo;
+  friend class internal::DefaultGlobalTestPartResultReporter;
+  friend class internal::NoExecDeathTest;
+  friend class internal::TestEventListenersAccessor;
+  friend class internal::UnitTestImpl;
+
+  // Returns repeater that broadcasts the TestEventListener events to all
+  // subscribers.
+  TestEventListener* repeater();
+
+  // Sets the default_result_printer attribute to the provided listener.
+  // The listener is also added to the listener list and previous
+  // default_result_printer is removed from it and deleted. The listener can
+  // also be NULL in which case it will not be added to the list. Does
+  // nothing if the previous and the current listener objects are the same.
+  void SetDefaultResultPrinter(TestEventListener* listener);
+
+  // Sets the default_xml_generator attribute to the provided listener.  The
+  // listener is also added to the listener list and previous
+  // default_xml_generator is removed from it and deleted. The listener can
+  // also be NULL in which case it will not be added to the list. Does
+  // nothing if the previous and the current listener objects are the same.
+  void SetDefaultXmlGenerator(TestEventListener* listener);
+
+  // Controls whether events will be forwarded by the repeater to the
+  // listeners in the list.
+  bool EventForwardingEnabled() const;
+  void SuppressEventForwarding();
+
+  // The actual list of listeners.
+  internal::TestEventRepeater* repeater_;
+  // Listener responsible for the standard result output.
+  TestEventListener* default_result_printer_;
+  // Listener responsible for the creation of the XML output file.
+  TestEventListener* default_xml_generator_;
+
+  // We disallow copying TestEventListeners.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestEventListeners);
+};
+
+// A UnitTest consists of a vector of TestCases.
+//
+// This is a singleton class.  The only instance of UnitTest is
+// created when UnitTest::GetInstance() is first called.  This
+// instance is never deleted.
+//
+// UnitTest is not copyable.
+//
+// This class is thread-safe as long as the methods are called
+// according to their specification.
+class GTEST_API_ UnitTest {
+ public:
+  // Gets the singleton UnitTest object.  The first time this method
+  // is called, a UnitTest object is constructed and returned.
+  // Consecutive calls will return the same object.
+  static UnitTest* GetInstance();
+
+  // Runs all tests in this UnitTest object and prints the result.
+  // Returns 0 if successful, or 1 otherwise.
+  //
+  // This method can only be called from the main thread.
+  //
+  // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+  int Run() GTEST_MUST_USE_RESULT_;
+
+  // Returns the working directory when the first TEST() or TEST_F()
+  // was executed.  The UnitTest object owns the string.
+  const char* original_working_dir() const;
+
+  // Returns the TestCase object for the test that's currently running,
+  // or NULL if no test is running.
+  const TestCase* current_test_case() const
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Returns the TestInfo object for the test that's currently running,
+  // or NULL if no test is running.
+  const TestInfo* current_test_info() const
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Returns the random seed used at the start of the current test run.
+  int random_seed() const;
+
+#if GTEST_HAS_PARAM_TEST
+  // Returns the ParameterizedTestCaseRegistry object used to keep track of
+  // value-parameterized tests and instantiate and register them.
+  //
+  // INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+  internal::ParameterizedTestCaseRegistry& parameterized_test_registry()
+      GTEST_LOCK_EXCLUDED_(mutex_);
+#endif  // GTEST_HAS_PARAM_TEST
+
+  // Gets the number of successful test cases.
+  int successful_test_case_count() const;
+
+  // Gets the number of failed test cases.
+  int failed_test_case_count() const;
+
+  // Gets the number of all test cases.
+  int total_test_case_count() const;
+
+  // Gets the number of all test cases that contain at least one test
+  // that should run.
+  int test_case_to_run_count() const;
+
+  // Gets the number of successful tests.
+  int successful_test_count() const;
+
+  // Gets the number of failed tests.
+  int failed_test_count() const;
+
+  // Gets the number of disabled tests that will be reported in the XML report.
+  int reportable_disabled_test_count() const;
+
+  // Gets the number of disabled tests.
+  int disabled_test_count() const;
+
+  // Gets the number of tests to be printed in the XML report.
+  int reportable_test_count() const;
+
+  // Gets the number of all tests.
+  int total_test_count() const;
+
+  // Gets the number of tests that should run.
+  int test_to_run_count() const;
+
+  // Gets the time of the test program start, in ms from the start of the
+  // UNIX epoch.
+  TimeInMillis start_timestamp() const;
+
+  // Gets the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const;
+
+  // Returns true iff the unit test passed (i.e. all test cases passed).
+  bool Passed() const;
+
+  // Returns true iff the unit test failed (i.e. some test case failed
+  // or something outside of all tests failed).
+  bool Failed() const;
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  const TestCase* GetTestCase(int i) const;
+
+  // Returns the TestResult containing information on test failures and
+  // properties logged outside of individual test cases.
+  const TestResult& ad_hoc_test_result() const;
+
+  // Returns the list of event listeners that can be used to track events
+  // inside Google Test.
+  TestEventListeners& listeners();
+
+ private:
+  // Registers and returns a global test environment.  When a test
+  // program is run, all global test environments will be set-up in
+  // the order they were registered.  After all tests in the program
+  // have finished, all global test environments will be torn-down in
+  // the *reverse* order they were registered.
+  //
+  // The UnitTest object takes ownership of the given environment.
+  //
+  // This method can only be called from the main thread.
+  Environment* AddEnvironment(Environment* env);
+
+  // Adds a TestPartResult to the current TestResult object.  All
+  // Google Test assertion macros (e.g. ASSERT_TRUE, EXPECT_EQ, etc)
+  // eventually call this to report their results.  The user code
+  // should use the assertion macros instead of calling this directly.
+  void AddTestPartResult(TestPartResult::Type result_type,
+                         const char* file_name,
+                         int line_number,
+                         const std::string& message,
+                         const std::string& os_stack_trace)
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Adds a TestProperty to the current TestResult object when invoked from
+  // inside a test, to current TestCase's ad_hoc_test_result_ when invoked
+  // from SetUpTestCase or TearDownTestCase, or to the global property set
+  // when invoked elsewhere.  If the result already contains a property with
+  // the same key, the value will be updated.
+  void RecordProperty(const std::string& key, const std::string& value);
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  TestCase* GetMutableTestCase(int i);
+
+  // Accessors for the implementation object.
+  internal::UnitTestImpl* impl() { return impl_; }
+  const internal::UnitTestImpl* impl() const { return impl_; }
+
+  // These classes and funcions are friends as they need to access private
+  // members of UnitTest.
+  friend class Test;
+  friend class internal::AssertHelper;
+  friend class internal::ScopedTrace;
+  friend class internal::StreamingListenerTest;
+  friend class internal::UnitTestRecordPropertyTestHelper;
+  friend Environment* AddGlobalTestEnvironment(Environment* env);
+  friend internal::UnitTestImpl* internal::GetUnitTestImpl();
+  friend void internal::ReportFailureInUnknownLocation(
+      TestPartResult::Type result_type,
+      const std::string& message);
+
+  // Creates an empty UnitTest.
+  UnitTest();
+
+  // D'tor
+  virtual ~UnitTest();
+
+  // Pushes a trace defined by SCOPED_TRACE() on to the per-thread
+  // Google Test trace stack.
+  void PushGTestTrace(const internal::TraceInfo& trace)
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Pops a trace from the per-thread Google Test trace stack.
+  void PopGTestTrace()
+      GTEST_LOCK_EXCLUDED_(mutex_);
+
+  // Protects mutable state in *impl_.  This is mutable as some const
+  // methods need to lock it too.
+  mutable internal::Mutex mutex_;
+
+  // Opaque implementation object.  This field is never changed once
+  // the object is constructed.  We don't mark it as const here, as
+  // doing so will cause a warning in the constructor of UnitTest.
+  // Mutable state in *impl_ is protected by mutex_.
+  internal::UnitTestImpl* impl_;
+
+  // We disallow copying UnitTest.
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(UnitTest);
+};
+
+// A convenient wrapper for adding an environment for the test
+// program.
+//
+// You should call this before RUN_ALL_TESTS() is called, probably in
+// main().  If you use gtest_main, you need to call this before main()
+// starts for it to take effect.  For example, you can define a global
+// variable like this:
+//
+//   testing::Environment* const foo_env =
+//       testing::AddGlobalTestEnvironment(new FooEnvironment);
+//
+// However, we strongly recommend you to write your own main() and
+// call AddGlobalTestEnvironment() there, as relying on initialization
+// of global variables makes the code harder to read and may cause
+// problems when you register multiple environments from different
+// translation units and the environments have dependencies among them
+// (remember that the compiler doesn't guarantee the order in which
+// global variables from different translation units are initialized).
+inline Environment* AddGlobalTestEnvironment(Environment* env) {
+  return UnitTest::GetInstance()->AddEnvironment(env);
+}
+
+// Initializes Google Test.  This must be called before calling
+// RUN_ALL_TESTS().  In particular, it parses a command line for the
+// flags that Google Test recognizes.  Whenever a Google Test flag is
+// seen, it is removed from argv, and *argc is decremented.
+//
+// No value is returned.  Instead, the Google Test flag variables are
+// updated.
+//
+// Calling the function for the second time has no user-visible effect.
+GTEST_API_ void InitGoogleTest(int* argc, char** argv);
+
+// This overloaded version can be used in Windows programs compiled in
+// UNICODE mode.
+GTEST_API_ void InitGoogleTest(int* argc, wchar_t** argv);
+
+namespace internal {
+
+// Separate the error generating code from the code path to reduce the stack
+// frame size of CmpHelperEQ. This helps reduce the overhead of some sanitizers
+// when calling EXPECT_* in a tight loop.
+template <typename T1, typename T2>
+AssertionResult CmpHelperEQFailure(const char* lhs_expression,
+                                   const char* rhs_expression,
+                                   const T1& lhs, const T2& rhs) {
+  return EqFailure(lhs_expression,
+                   rhs_expression,
+                   FormatForComparisonFailureMessage(lhs, rhs),
+                   FormatForComparisonFailureMessage(rhs, lhs),
+                   false);
+}
+
+// The helper function for {ASSERT|EXPECT}_EQ.
+template <typename T1, typename T2>
+AssertionResult CmpHelperEQ(const char* lhs_expression,
+                            const char* rhs_expression,
+                            const T1& lhs,
+                            const T2& rhs) {
+GTEST_DISABLE_MSC_WARNINGS_PUSH_(4389 /* signed/unsigned mismatch */)
+  if (lhs == rhs) {
+    return AssertionSuccess();
+  }
+GTEST_DISABLE_MSC_WARNINGS_POP_()
+
+  return CmpHelperEQFailure(lhs_expression, rhs_expression, lhs, rhs);
+}
+
+// With this overloaded version, we allow anonymous enums to be used
+// in {ASSERT|EXPECT}_EQ when compiled with gcc 4, as anonymous enums
+// can be implicitly cast to BiggestInt.
+GTEST_API_ AssertionResult CmpHelperEQ(const char* lhs_expression,
+                                       const char* rhs_expression,
+                                       BiggestInt lhs,
+                                       BiggestInt rhs);
+
+// The helper class for {ASSERT|EXPECT}_EQ.  The template argument
+// lhs_is_null_literal is true iff the first argument to ASSERT_EQ()
+// is a null pointer literal.  The following default implementation is
+// for lhs_is_null_literal being false.
+template <bool lhs_is_null_literal>
+class EqHelper {
+ public:
+  // This templatized version is for the general case.
+  template <typename T1, typename T2>
+  static AssertionResult Compare(const char* lhs_expression,
+                                 const char* rhs_expression,
+                                 const T1& lhs,
+                                 const T2& rhs) {
+    return CmpHelperEQ(lhs_expression, rhs_expression, lhs, rhs);
+  }
+
+  // With this overloaded version, we allow anonymous enums to be used
+  // in {ASSERT|EXPECT}_EQ when compiled with gcc 4, as anonymous
+  // enums can be implicitly cast to BiggestInt.
+  //
+  // Even though its body looks the same as the above version, we
+  // cannot merge the two, as it will make anonymous enums unhappy.
+  static AssertionResult Compare(const char* lhs_expression,
+                                 const char* rhs_expression,
+                                 BiggestInt lhs,
+                                 BiggestInt rhs) {
+    return CmpHelperEQ(lhs_expression, rhs_expression, lhs, rhs);
+  }
+};
+
+// This specialization is used when the first argument to ASSERT_EQ()
+// is a null pointer literal, like NULL, false, or 0.
+template <>
+class EqHelper<true> {
+ public:
+  // We define two overloaded versions of Compare().  The first
+  // version will be picked when the second argument to ASSERT_EQ() is
+  // NOT a pointer, e.g. ASSERT_EQ(0, AnIntFunction()) or
+  // EXPECT_EQ(false, a_bool).
+  template <typename T1, typename T2>
+  static AssertionResult Compare(
+      const char* lhs_expression,
+      const char* rhs_expression,
+      const T1& lhs,
+      const T2& rhs,
+      // The following line prevents this overload from being considered if T2
+      // is not a pointer type.  We need this because ASSERT_EQ(NULL, my_ptr)
+      // expands to Compare("", "", NULL, my_ptr), which requires a conversion
+      // to match the Secret* in the other overload, which would otherwise make
+      // this template match better.
+      typename EnableIf<!is_pointer<T2>::value>::type* = 0) {
+    return CmpHelperEQ(lhs_expression, rhs_expression, lhs, rhs);
+  }
+
+  // This version will be picked when the second argument to ASSERT_EQ() is a
+  // pointer, e.g. ASSERT_EQ(NULL, a_pointer).
+  template <typename T>
+  static AssertionResult Compare(
+      const char* lhs_expression,
+      const char* rhs_expression,
+      // We used to have a second template parameter instead of Secret*.  That
+      // template parameter would deduce to 'long', making this a better match
+      // than the first overload even without the first overload's EnableIf.
+      // Unfortunately, gcc with -Wconversion-null warns when "passing NULL to
+      // non-pointer argument" (even a deduced integral argument), so the old
+      // implementation caused warnings in user code.
+      Secret* /* lhs (NULL) */,
+      T* rhs) {
+    // We already know that 'lhs' is a null pointer.
+    return CmpHelperEQ(lhs_expression, rhs_expression,
+                       static_cast<T*>(NULL), rhs);
+  }
+};
+
+// Separate the error generating code from the code path to reduce the stack
+// frame size of CmpHelperOP. This helps reduce the overhead of some sanitizers
+// when calling EXPECT_OP in a tight loop.
+template <typename T1, typename T2>
+AssertionResult CmpHelperOpFailure(const char* expr1, const char* expr2,
+                                   const T1& val1, const T2& val2,
+                                   const char* op) {
+  return AssertionFailure()
+         << "Expected: (" << expr1 << ") " << op << " (" << expr2
+         << "), actual: " << FormatForComparisonFailureMessage(val1, val2)
+         << " vs " << FormatForComparisonFailureMessage(val2, val1);
+}
+
+// A macro for implementing the helper functions needed to implement
+// ASSERT_?? and EXPECT_??.  It is here just to avoid copy-and-paste
+// of similar code.
+//
+// For each templatized helper function, we also define an overloaded
+// version for BiggestInt in order to reduce code bloat and allow
+// anonymous enums to be used with {ASSERT|EXPECT}_?? when compiled
+// with gcc 4.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+
+#define GTEST_IMPL_CMP_HELPER_(op_name, op)\
+template <typename T1, typename T2>\
+AssertionResult CmpHelper##op_name(const char* expr1, const char* expr2, \
+                                   const T1& val1, const T2& val2) {\
+  if (val1 op val2) {\
+    return AssertionSuccess();\
+  } else {\
+    return CmpHelperOpFailure(expr1, expr2, val1, val2, #op);\
+  }\
+}\
+GTEST_API_ AssertionResult CmpHelper##op_name(\
+    const char* expr1, const char* expr2, BiggestInt val1, BiggestInt val2)
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+
+// Implements the helper function for {ASSERT|EXPECT}_NE
+GTEST_IMPL_CMP_HELPER_(NE, !=);
+// Implements the helper function for {ASSERT|EXPECT}_LE
+GTEST_IMPL_CMP_HELPER_(LE, <=);
+// Implements the helper function for {ASSERT|EXPECT}_LT
+GTEST_IMPL_CMP_HELPER_(LT, <);
+// Implements the helper function for {ASSERT|EXPECT}_GE
+GTEST_IMPL_CMP_HELPER_(GE, >=);
+// Implements the helper function for {ASSERT|EXPECT}_GT
+GTEST_IMPL_CMP_HELPER_(GT, >);
+
+#undef GTEST_IMPL_CMP_HELPER_
+
+// The helper function for {ASSERT|EXPECT}_STREQ.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTREQ(const char* s1_expression,
+                                          const char* s2_expression,
+                                          const char* s1,
+                                          const char* s2);
+
+// The helper function for {ASSERT|EXPECT}_STRCASEEQ.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRCASEEQ(const char* s1_expression,
+                                              const char* s2_expression,
+                                              const char* s1,
+                                              const char* s2);
+
+// The helper function for {ASSERT|EXPECT}_STRNE.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                                          const char* s2_expression,
+                                          const char* s1,
+                                          const char* s2);
+
+// The helper function for {ASSERT|EXPECT}_STRCASENE.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRCASENE(const char* s1_expression,
+                                              const char* s2_expression,
+                                              const char* s1,
+                                              const char* s2);
+
+
+// Helper function for *_STREQ on wide strings.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTREQ(const char* s1_expression,
+                                          const char* s2_expression,
+                                          const wchar_t* s1,
+                                          const wchar_t* s2);
+
+// Helper function for *_STRNE on wide strings.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                                          const char* s2_expression,
+                                          const wchar_t* s1,
+                                          const wchar_t* s2);
+
+}  // namespace internal
+
+// IsSubstring() and IsNotSubstring() are intended to be used as the
+// first argument to {EXPECT,ASSERT}_PRED_FORMAT2(), not by
+// themselves.  They check whether needle is a substring of haystack
+// (NULL is considered a substring of itself only), and return an
+// appropriate error message when they fail.
+//
+// The {needle,haystack}_expr arguments are the stringified
+// expressions that generated the two real arguments.
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack);
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack);
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack);
+
+#if GTEST_HAS_STD_WSTRING
+GTEST_API_ AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack);
+GTEST_API_ AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack);
+#endif  // GTEST_HAS_STD_WSTRING
+
+namespace internal {
+
+// Helper template function for comparing floating-points.
+//
+// Template parameter:
+//
+//   RawType: the raw floating-point type (either float or double)
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+template <typename RawType>
+AssertionResult CmpHelperFloatingPointEQ(const char* lhs_expression,
+                                         const char* rhs_expression,
+                                         RawType lhs_value,
+                                         RawType rhs_value) {
+  const FloatingPoint<RawType> lhs(lhs_value), rhs(rhs_value);
+
+  if (lhs.AlmostEquals(rhs)) {
+    return AssertionSuccess();
+  }
+
+  ::std::stringstream lhs_ss;
+  lhs_ss << std::setprecision(std::numeric_limits<RawType>::digits10 + 2)
+         << lhs_value;
+
+  ::std::stringstream rhs_ss;
+  rhs_ss << std::setprecision(std::numeric_limits<RawType>::digits10 + 2)
+         << rhs_value;
+
+  return EqFailure(lhs_expression,
+                   rhs_expression,
+                   StringStreamToString(&lhs_ss),
+                   StringStreamToString(&rhs_ss),
+                   false);
+}
+
+// Helper function for implementing ASSERT_NEAR.
+//
+// INTERNAL IMPLEMENTATION - DO NOT USE IN A USER PROGRAM.
+GTEST_API_ AssertionResult DoubleNearPredFormat(const char* expr1,
+                                                const char* expr2,
+                                                const char* abs_error_expr,
+                                                double val1,
+                                                double val2,
+                                                double abs_error);
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+// A class that enables one to stream messages to assertion macros
+class GTEST_API_ AssertHelper {
+ public:
+  // Constructor.
+  AssertHelper(TestPartResult::Type type,
+               const char* file,
+               int line,
+               const char* message);
+  ~AssertHelper();
+
+  // Message assignment is a semantic trick to enable assertion
+  // streaming; see the GTEST_MESSAGE_ macro below.
+  void operator=(const Message& message) const;
+
+ private:
+  // We put our data in a struct so that the size of the AssertHelper class can
+  // be as small as possible.  This is important because gcc is incapable of
+  // re-using stack space even for temporary variables, so every EXPECT_EQ
+  // reserves stack space for another AssertHelper.
+  struct AssertHelperData {
+    AssertHelperData(TestPartResult::Type t,
+                     const char* srcfile,
+                     int line_num,
+                     const char* msg)
+        : type(t), file(srcfile), line(line_num), message(msg) { }
+
+    TestPartResult::Type const type;
+    const char* const file;
+    int const line;
+    std::string const message;
+
+   private:
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(AssertHelperData);
+  };
+
+  AssertHelperData* const data_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(AssertHelper);
+};
+
+}  // namespace internal
+
+#if GTEST_HAS_PARAM_TEST
+// The pure interface class that all value-parameterized tests inherit from.
+// A value-parameterized class must inherit from both ::testing::Test and
+// ::testing::WithParamInterface. In most cases that just means inheriting
+// from ::testing::TestWithParam, but more complicated test hierarchies
+// may need to inherit from Test and WithParamInterface at different levels.
+//
+// This interface has support for accessing the test parameter value via
+// the GetParam() method.
+//
+// Use it with one of the parameter generator defining functions, like Range(),
+// Values(), ValuesIn(), Bool(), and Combine().
+//
+// class FooTest : public ::testing::TestWithParam<int> {
+//  protected:
+//   FooTest() {
+//     // Can use GetParam() here.
+//   }
+//   virtual ~FooTest() {
+//     // Can use GetParam() here.
+//   }
+//   virtual void SetUp() {
+//     // Can use GetParam() here.
+//   }
+//   virtual void TearDown {
+//     // Can use GetParam() here.
+//   }
+// };
+// TEST_P(FooTest, DoesBar) {
+//   // Can use GetParam() method here.
+//   Foo foo;
+//   ASSERT_TRUE(foo.DoesBar(GetParam()));
+// }
+// INSTANTIATE_TEST_CASE_P(OneToTenRange, FooTest, ::testing::Range(1, 10));
+
+template <typename T>
+class WithParamInterface {
+ public:
+  typedef T ParamType;
+  virtual ~WithParamInterface() {}
+
+  // The current parameter value. Is also available in the test fixture's
+  // constructor. This member function is non-static, even though it only
+  // references static data, to reduce the opportunity for incorrect uses
+  // like writing 'WithParamInterface<bool>::GetParam()' for a test that
+  // uses a fixture whose parameter type is int.
+  const ParamType& GetParam() const {
+    GTEST_CHECK_(parameter_ != NULL)
+        << "GetParam() can only be called inside a value-parameterized test "
+        << "-- did you intend to write TEST_P instead of TEST_F?";
+    return *parameter_;
+  }
+
+ private:
+  // Sets parameter value. The caller is responsible for making sure the value
+  // remains alive and unchanged throughout the current test.
+  static void SetParam(const ParamType* parameter) {
+    parameter_ = parameter;
+  }
+
+  // Static value used for accessing parameter during a test lifetime.
+  static const ParamType* parameter_;
+
+  // TestClass must be a subclass of WithParamInterface<T> and Test.
+  template <class TestClass> friend class internal::ParameterizedTestFactory;
+};
+
+template <typename T>
+const T* WithParamInterface<T>::parameter_ = NULL;
+
+// Most value-parameterized classes can ignore the existence of
+// WithParamInterface, and can just inherit from ::testing::TestWithParam.
+
+template <typename T>
+class TestWithParam : public Test, public WithParamInterface<T> {
+};
+
+#endif  // GTEST_HAS_PARAM_TEST
+
+// Macros for indicating success/failure in test code.
+
+// ADD_FAILURE unconditionally adds a failure to the current test.
+// SUCCEED generates a success - it doesn't automatically make the
+// current test successful, as a test is only successful when it has
+// no failure.
+//
+// EXPECT_* verifies that a certain condition is satisfied.  If not,
+// it behaves like ADD_FAILURE.  In particular:
+//
+//   EXPECT_TRUE  verifies that a Boolean condition is true.
+//   EXPECT_FALSE verifies that a Boolean condition is false.
+//
+// FAIL and ASSERT_* are similar to ADD_FAILURE and EXPECT_*, except
+// that they will also abort the current function on failure.  People
+// usually want the fail-fast behavior of FAIL and ASSERT_*, but those
+// writing data-driven tests often find themselves using ADD_FAILURE
+// and EXPECT_* more.
+
+// Generates a nonfatal failure with a generic message.
+#define ADD_FAILURE() GTEST_NONFATAL_FAILURE_("Failed")
+
+// Generates a nonfatal failure at the given source file location with
+// a generic message.
+#define ADD_FAILURE_AT(file, line) \
+  GTEST_MESSAGE_AT_(file, line, "Failed", \
+                    ::testing::TestPartResult::kNonFatalFailure)
+
+// Generates a fatal failure with a generic message.
+#define GTEST_FAIL() GTEST_FATAL_FAILURE_("Failed")
+
+// Define this macro to 1 to omit the definition of FAIL(), which is a
+// generic name and clashes with some other libraries.
+#if !GTEST_DONT_DEFINE_FAIL
+# define FAIL() GTEST_FAIL()
+#endif
+
+// Generates a success with a generic message.
+#define GTEST_SUCCEED() GTEST_SUCCESS_("Succeeded")
+
+// Define this macro to 1 to omit the definition of SUCCEED(), which
+// is a generic name and clashes with some other libraries.
+#if !GTEST_DONT_DEFINE_SUCCEED
+# define SUCCEED() GTEST_SUCCEED()
+#endif
+
+// Macros for testing exceptions.
+//
+//    * {ASSERT|EXPECT}_THROW(statement, expected_exception):
+//         Tests that the statement throws the expected exception.
+//    * {ASSERT|EXPECT}_NO_THROW(statement):
+//         Tests that the statement doesn't throw any exception.
+//    * {ASSERT|EXPECT}_ANY_THROW(statement):
+//         Tests that the statement throws an exception.
+
+#define EXPECT_THROW(statement, expected_exception) \
+  GTEST_TEST_THROW_(statement, expected_exception, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_NO_THROW(statement) \
+  GTEST_TEST_NO_THROW_(statement, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_ANY_THROW(statement) \
+  GTEST_TEST_ANY_THROW_(statement, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_THROW(statement, expected_exception) \
+  GTEST_TEST_THROW_(statement, expected_exception, GTEST_FATAL_FAILURE_)
+#define ASSERT_NO_THROW(statement) \
+  GTEST_TEST_NO_THROW_(statement, GTEST_FATAL_FAILURE_)
+#define ASSERT_ANY_THROW(statement) \
+  GTEST_TEST_ANY_THROW_(statement, GTEST_FATAL_FAILURE_)
+
+// Boolean assertions. Condition can be either a Boolean expression or an
+// AssertionResult. For more information on how to use AssertionResult with
+// these macros see comments on that class.
+#define EXPECT_TRUE(condition) \
+  GTEST_TEST_BOOLEAN_((condition), #condition, false, true, \
+                      GTEST_NONFATAL_FAILURE_)
+#define EXPECT_FALSE(condition) \
+  GTEST_TEST_BOOLEAN_(!(condition), #condition, true, false, \
+                      GTEST_NONFATAL_FAILURE_)
+#define ASSERT_TRUE(condition) \
+  GTEST_TEST_BOOLEAN_((condition), #condition, false, true, \
+                      GTEST_FATAL_FAILURE_)
+#define ASSERT_FALSE(condition) \
+  GTEST_TEST_BOOLEAN_(!(condition), #condition, true, false, \
+                      GTEST_FATAL_FAILURE_)
+
+// Includes the auto-generated header that implements a family of
+// generic predicate assertion macros.
+#include "gtest/gtest_pred_impl.h"
+
+// Macros for testing equalities and inequalities.
+//
+//    * {ASSERT|EXPECT}_EQ(v1, v2): Tests that v1 == v2
+//    * {ASSERT|EXPECT}_NE(v1, v2): Tests that v1 != v2
+//    * {ASSERT|EXPECT}_LT(v1, v2): Tests that v1 < v2
+//    * {ASSERT|EXPECT}_LE(v1, v2): Tests that v1 <= v2
+//    * {ASSERT|EXPECT}_GT(v1, v2): Tests that v1 > v2
+//    * {ASSERT|EXPECT}_GE(v1, v2): Tests that v1 >= v2
+//
+// When they are not, Google Test prints both the tested expressions and
+// their actual values.  The values must be compatible built-in types,
+// or you will get a compiler error.  By "compatible" we mean that the
+// values can be compared by the respective operator.
+//
+// Note:
+//
+//   1. It is possible to make a user-defined type work with
+//   {ASSERT|EXPECT}_??(), but that requires overloading the
+//   comparison operators and is thus discouraged by the Google C++
+//   Usage Guide.  Therefore, you are advised to use the
+//   {ASSERT|EXPECT}_TRUE() macro to assert that two objects are
+//   equal.
+//
+//   2. The {ASSERT|EXPECT}_??() macros do pointer comparisons on
+//   pointers (in particular, C strings).  Therefore, if you use it
+//   with two C strings, you are testing how their locations in memory
+//   are related, not how their content is related.  To compare two C
+//   strings by content, use {ASSERT|EXPECT}_STR*().
+//
+//   3. {ASSERT|EXPECT}_EQ(v1, v2) is preferred to
+//   {ASSERT|EXPECT}_TRUE(v1 == v2), as the former tells you
+//   what the actual value is when it fails, and similarly for the
+//   other comparisons.
+//
+//   4. Do not depend on the order in which {ASSERT|EXPECT}_??()
+//   evaluate their arguments, which is undefined.
+//
+//   5. These macros evaluate their arguments exactly once.
+//
+// Examples:
+//
+//   EXPECT_NE(5, Foo());
+//   EXPECT_EQ(NULL, a_pointer);
+//   ASSERT_LT(i, array_size);
+//   ASSERT_GT(records.size(), 0) << "There is no record left.";
+
+#define EXPECT_EQ(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal:: \
+                      EqHelper<GTEST_IS_NULL_LITERAL_(val1)>::Compare, \
+                      val1, val2)
+#define EXPECT_NE(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperNE, val1, val2)
+#define EXPECT_LE(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperLE, val1, val2)
+#define EXPECT_LT(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperLT, val1, val2)
+#define EXPECT_GE(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperGE, val1, val2)
+#define EXPECT_GT(val1, val2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperGT, val1, val2)
+
+#define GTEST_ASSERT_EQ(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal:: \
+                      EqHelper<GTEST_IS_NULL_LITERAL_(val1)>::Compare, \
+                      val1, val2)
+#define GTEST_ASSERT_NE(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperNE, val1, val2)
+#define GTEST_ASSERT_LE(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperLE, val1, val2)
+#define GTEST_ASSERT_LT(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperLT, val1, val2)
+#define GTEST_ASSERT_GE(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperGE, val1, val2)
+#define GTEST_ASSERT_GT(val1, val2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperGT, val1, val2)
+
+// Define macro GTEST_DONT_DEFINE_ASSERT_XY to 1 to omit the definition of
+// ASSERT_XY(), which clashes with some users' own code.
+
+#if !GTEST_DONT_DEFINE_ASSERT_EQ
+# define ASSERT_EQ(val1, val2) GTEST_ASSERT_EQ(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_NE
+# define ASSERT_NE(val1, val2) GTEST_ASSERT_NE(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_LE
+# define ASSERT_LE(val1, val2) GTEST_ASSERT_LE(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_LT
+# define ASSERT_LT(val1, val2) GTEST_ASSERT_LT(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_GE
+# define ASSERT_GE(val1, val2) GTEST_ASSERT_GE(val1, val2)
+#endif
+
+#if !GTEST_DONT_DEFINE_ASSERT_GT
+# define ASSERT_GT(val1, val2) GTEST_ASSERT_GT(val1, val2)
+#endif
+
+// C-string Comparisons.  All tests treat NULL and any non-NULL string
+// as different.  Two NULLs are equal.
+//
+//    * {ASSERT|EXPECT}_STREQ(s1, s2):     Tests that s1 == s2
+//    * {ASSERT|EXPECT}_STRNE(s1, s2):     Tests that s1 != s2
+//    * {ASSERT|EXPECT}_STRCASEEQ(s1, s2): Tests that s1 == s2, ignoring case
+//    * {ASSERT|EXPECT}_STRCASENE(s1, s2): Tests that s1 != s2, ignoring case
+//
+// For wide or narrow string objects, you can use the
+// {ASSERT|EXPECT}_??() macros.
+//
+// Don't depend on the order in which the arguments are evaluated,
+// which is undefined.
+//
+// These macros evaluate their arguments exactly once.
+
+#define EXPECT_STREQ(s1, s2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTREQ, s1, s2)
+#define EXPECT_STRNE(s1, s2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTRNE, s1, s2)
+#define EXPECT_STRCASEEQ(s1, s2) \
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASEEQ, s1, s2)
+#define EXPECT_STRCASENE(s1, s2)\
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASENE, s1, s2)
+
+#define ASSERT_STREQ(s1, s2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTREQ, s1, s2)
+#define ASSERT_STRNE(s1, s2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTRNE, s1, s2)
+#define ASSERT_STRCASEEQ(s1, s2) \
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASEEQ, s1, s2)
+#define ASSERT_STRCASENE(s1, s2)\
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperSTRCASENE, s1, s2)
+
+// Macros for comparing floating-point numbers.
+//
+//    * {ASSERT|EXPECT}_FLOAT_EQ(val1, val2):
+//         Tests that two float values are almost equal.
+//    * {ASSERT|EXPECT}_DOUBLE_EQ(val1, val2):
+//         Tests that two double values are almost equal.
+//    * {ASSERT|EXPECT}_NEAR(v1, v2, abs_error):
+//         Tests that v1 and v2 are within the given distance to each other.
+//
+// Google Test uses ULP-based comparison to automatically pick a default
+// error bound that is appropriate for the operands.  See the
+// FloatingPoint template class in gtest-internal.h if you are
+// interested in the implementation details.
+
+#define EXPECT_FLOAT_EQ(val1, val2)\
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ<float>, \
+                      val1, val2)
+
+#define EXPECT_DOUBLE_EQ(val1, val2)\
+  EXPECT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ<double>, \
+                      val1, val2)
+
+#define ASSERT_FLOAT_EQ(val1, val2)\
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ<float>, \
+                      val1, val2)
+
+#define ASSERT_DOUBLE_EQ(val1, val2)\
+  ASSERT_PRED_FORMAT2(::testing::internal::CmpHelperFloatingPointEQ<double>, \
+                      val1, val2)
+
+#define EXPECT_NEAR(val1, val2, abs_error)\
+  EXPECT_PRED_FORMAT3(::testing::internal::DoubleNearPredFormat, \
+                      val1, val2, abs_error)
+
+#define ASSERT_NEAR(val1, val2, abs_error)\
+  ASSERT_PRED_FORMAT3(::testing::internal::DoubleNearPredFormat, \
+                      val1, val2, abs_error)
+
+// These predicate format functions work on floating-point values, and
+// can be used in {ASSERT|EXPECT}_PRED_FORMAT2*(), e.g.
+//
+//   EXPECT_PRED_FORMAT2(testing::DoubleLE, Foo(), 5.0);
+
+// Asserts that val1 is less than, or almost equal to, val2.  Fails
+// otherwise.  In particular, it fails if either val1 or val2 is NaN.
+GTEST_API_ AssertionResult FloatLE(const char* expr1, const char* expr2,
+                                   float val1, float val2);
+GTEST_API_ AssertionResult DoubleLE(const char* expr1, const char* expr2,
+                                    double val1, double val2);
+
+
+#if GTEST_OS_WINDOWS
+
+// Macros that test for HRESULT failure and success, these are only useful
+// on Windows, and rely on Windows SDK macros and APIs to compile.
+//
+//    * {ASSERT|EXPECT}_HRESULT_{SUCCEEDED|FAILED}(expr)
+//
+// When expr unexpectedly fails or succeeds, Google Test prints the
+// expected result and the actual result with both a human-readable
+// string representation of the error, if available, as well as the
+// hex result code.
+# define EXPECT_HRESULT_SUCCEEDED(expr) \
+    EXPECT_PRED_FORMAT1(::testing::internal::IsHRESULTSuccess, (expr))
+
+# define ASSERT_HRESULT_SUCCEEDED(expr) \
+    ASSERT_PRED_FORMAT1(::testing::internal::IsHRESULTSuccess, (expr))
+
+# define EXPECT_HRESULT_FAILED(expr) \
+    EXPECT_PRED_FORMAT1(::testing::internal::IsHRESULTFailure, (expr))
+
+# define ASSERT_HRESULT_FAILED(expr) \
+    ASSERT_PRED_FORMAT1(::testing::internal::IsHRESULTFailure, (expr))
+
+#endif  // GTEST_OS_WINDOWS
+
+// Macros that execute statement and check that it doesn't generate new fatal
+// failures in the current thread.
+//
+//   * {ASSERT|EXPECT}_NO_FATAL_FAILURE(statement);
+//
+// Examples:
+//
+//   EXPECT_NO_FATAL_FAILURE(Process());
+//   ASSERT_NO_FATAL_FAILURE(Process()) << "Process() failed";
+//
+#define ASSERT_NO_FATAL_FAILURE(statement) \
+    GTEST_TEST_NO_FATAL_FAILURE_(statement, GTEST_FATAL_FAILURE_)
+#define EXPECT_NO_FATAL_FAILURE(statement) \
+    GTEST_TEST_NO_FATAL_FAILURE_(statement, GTEST_NONFATAL_FAILURE_)
+
+// Causes a trace (including the source file path, the current line
+// number, and the given message) to be included in every test failure
+// message generated by code in the current scope.  The effect is
+// undone when the control leaves the current scope.
+//
+// The message argument can be anything streamable to std::ostream.
+//
+// In the implementation, we include the current line number as part
+// of the dummy variable name, thus allowing multiple SCOPED_TRACE()s
+// to appear in the same block - as long as they are on different
+// lines.
+#define SCOPED_TRACE(message) \
+  ::testing::internal::ScopedTrace GTEST_CONCAT_TOKEN_(gtest_trace_, __LINE__)(\
+    __FILE__, __LINE__, ::testing::Message() << (message))
+
+// Compile-time assertion for type equality.
+// StaticAssertTypeEq<type1, type2>() compiles iff type1 and type2 are
+// the same type.  The value it returns is not interesting.
+//
+// Instead of making StaticAssertTypeEq a class template, we make it a
+// function template that invokes a helper class template.  This
+// prevents a user from misusing StaticAssertTypeEq<T1, T2> by
+// defining objects of that type.
+//
+// CAVEAT:
+//
+// When used inside a method of a class template,
+// StaticAssertTypeEq<T1, T2>() is effective ONLY IF the method is
+// instantiated.  For example, given:
+//
+//   template <typename T> class Foo {
+//    public:
+//     void Bar() { testing::StaticAssertTypeEq<int, T>(); }
+//   };
+//
+// the code:
+//
+//   void Test1() { Foo<bool> foo; }
+//
+// will NOT generate a compiler error, as Foo<bool>::Bar() is never
+// actually instantiated.  Instead, you need:
+//
+//   void Test2() { Foo<bool> foo; foo.Bar(); }
+//
+// to cause a compiler error.
+template <typename T1, typename T2>
+bool StaticAssertTypeEq() {
+  (void)internal::StaticAssertTypeEqHelper<T1, T2>();
+  return true;
+}
+
+// Defines a test.
+//
+// The first parameter is the name of the test case, and the second
+// parameter is the name of the test within the test case.
+//
+// The convention is to end the test case name with "Test".  For
+// example, a test case for the Foo class can be named FooTest.
+//
+// Test code should appear between braces after an invocation of
+// this macro.  Example:
+//
+//   TEST(FooTest, InitializesCorrectly) {
+//     Foo foo;
+//     EXPECT_TRUE(foo.StatusIsOK());
+//   }
+
+// Note that we call GetTestTypeId() instead of GetTypeId<
+// ::testing::Test>() here to get the type ID of testing::Test.  This
+// is to work around a suspected linker bug when using Google Test as
+// a framework on Mac OS X.  The bug causes GetTypeId<
+// ::testing::Test>() to return different values depending on whether
+// the call is from the Google Test framework itself or from user test
+// code.  GetTestTypeId() is guaranteed to always return the same
+// value, as it always calls GetTypeId<>() from the Google Test
+// framework.
+#define GTEST_TEST(test_case_name, test_name)\
+  GTEST_TEST_(test_case_name, test_name, \
+              ::testing::Test, ::testing::internal::GetTestTypeId())
+
+// Define this macro to 1 to omit the definition of TEST(), which
+// is a generic name and clashes with some other libraries.
+#if !GTEST_DONT_DEFINE_TEST
+# define TEST(test_case_name, test_name) GTEST_TEST(test_case_name, test_name)
+#endif
+
+// Defines a test that uses a test fixture.
+//
+// The first parameter is the name of the test fixture class, which
+// also doubles as the test case name.  The second parameter is the
+// name of the test within the test case.
+//
+// A test fixture class must be declared earlier.  The user should put
+// his test code between braces after using this macro.  Example:
+//
+//   class FooTest : public testing::Test {
+//    protected:
+//     virtual void SetUp() { b_.AddElement(3); }
+//
+//     Foo a_;
+//     Foo b_;
+//   };
+//
+//   TEST_F(FooTest, InitializesCorrectly) {
+//     EXPECT_TRUE(a_.StatusIsOK());
+//   }
+//
+//   TEST_F(FooTest, ReturnsElementCountCorrectly) {
+//     EXPECT_EQ(0, a_.size());
+//     EXPECT_EQ(1, b_.size());
+//   }
+
+#define TEST_F(test_fixture, test_name)\
+  GTEST_TEST_(test_fixture, test_name, test_fixture, \
+              ::testing::internal::GetTypeId<test_fixture>())
+
+}  // namespace testing
+
+// Use this function in main() to run all tests.  It returns 0 if all
+// tests are successful, or 1 otherwise.
+//
+// RUN_ALL_TESTS() should be invoked after the command line has been
+// parsed by InitGoogleTest().
+//
+// This function was formerly a macro; thus, it is in the global
+// namespace and has an all-caps name.
+int RUN_ALL_TESTS() GTEST_MUST_USE_RESULT_;
+
+inline int RUN_ALL_TESTS() {
+  return ::testing::UnitTest::GetInstance()->Run();
+}
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest_pred_impl.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest_pred_impl.h
new file mode 100644
index 0000000000..30ae712f50
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest_pred_impl.h
@@ -0,0 +1,358 @@
+// Copyright 2006, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// This file is AUTOMATICALLY GENERATED on 10/31/2011 by command
+// 'gen_gtest_pred_impl.py 5'.  DO NOT EDIT BY HAND!
+//
+// Implements a family of generic predicate assertion macros.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PRED_IMPL_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PRED_IMPL_H_
+
+// Makes sure this header is not included before gtest.h.
+#ifndef GTEST_INCLUDE_GTEST_GTEST_H_
+# error Do not include gtest_pred_impl.h directly.  Include gtest.h instead.
+#endif  // GTEST_INCLUDE_GTEST_GTEST_H_
+
+// This header implements a family of generic predicate assertion
+// macros:
+//
+//   ASSERT_PRED_FORMAT1(pred_format, v1)
+//   ASSERT_PRED_FORMAT2(pred_format, v1, v2)
+//   ...
+//
+// where pred_format is a function or functor that takes n (in the
+// case of ASSERT_PRED_FORMATn) values and their source expression
+// text, and returns a testing::AssertionResult.  See the definition
+// of ASSERT_EQ in gtest.h for an example.
+//
+// If you don't care about formatting, you can use the more
+// restrictive version:
+//
+//   ASSERT_PRED1(pred, v1)
+//   ASSERT_PRED2(pred, v1, v2)
+//   ...
+//
+// where pred is an n-ary function or functor that returns bool,
+// and the values v1, v2, ..., must support the << operator for
+// streaming to std::ostream.
+//
+// We also define the EXPECT_* variations.
+//
+// For now we only support predicates whose arity is at most 5.
+// Please email googletestframework@googlegroups.com if you need
+// support for higher arities.
+
+// GTEST_ASSERT_ is the basic statement to which all of the assertions
+// in this file reduce.  Don't use this in your code.
+
+#define GTEST_ASSERT_(expression, on_failure) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (const ::testing::AssertionResult gtest_ar = (expression)) \
+    ; \
+  else \
+    on_failure(gtest_ar.failure_message())
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED1.  Don't use
+// this in your code.
+template <typename Pred,
+          typename T1>
+AssertionResult AssertPred1Helper(const char* pred_text,
+                                  const char* e1,
+                                  Pred pred,
+                                  const T1& v1) {
+  if (pred(v1)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT1.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT1_(pred_format, v1, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, v1), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED1.  Don't use
+// this in your code.
+#define GTEST_PRED1_(pred, v1, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred1Helper(#pred, \
+                                             #v1, \
+                                             pred, \
+                                             v1), on_failure)
+
+// Unary predicate assertion macros.
+#define EXPECT_PRED_FORMAT1(pred_format, v1) \
+  GTEST_PRED_FORMAT1_(pred_format, v1, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED1(pred, v1) \
+  GTEST_PRED1_(pred, v1, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT1(pred_format, v1) \
+  GTEST_PRED_FORMAT1_(pred_format, v1, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED1(pred, v1) \
+  GTEST_PRED1_(pred, v1, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED2.  Don't use
+// this in your code.
+template <typename Pred,
+          typename T1,
+          typename T2>
+AssertionResult AssertPred2Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2) {
+  if (pred(v1, v2)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT2.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT2_(pred_format, v1, v2, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, v1, v2), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED2.  Don't use
+// this in your code.
+#define GTEST_PRED2_(pred, v1, v2, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred2Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             pred, \
+                                             v1, \
+                                             v2), on_failure)
+
+// Binary predicate assertion macros.
+#define EXPECT_PRED_FORMAT2(pred_format, v1, v2) \
+  GTEST_PRED_FORMAT2_(pred_format, v1, v2, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED2(pred, v1, v2) \
+  GTEST_PRED2_(pred, v1, v2, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT2(pred_format, v1, v2) \
+  GTEST_PRED_FORMAT2_(pred_format, v1, v2, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED2(pred, v1, v2) \
+  GTEST_PRED2_(pred, v1, v2, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED3.  Don't use
+// this in your code.
+template <typename Pred,
+          typename T1,
+          typename T2,
+          typename T3>
+AssertionResult AssertPred3Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  const char* e3,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2,
+                                  const T3& v3) {
+  if (pred(v1, v2, v3)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ", "
+                            << e3 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2
+                            << "\n" << e3 << " evaluates to " << v3;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT3.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT3_(pred_format, v1, v2, v3, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, #v3, v1, v2, v3), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED3.  Don't use
+// this in your code.
+#define GTEST_PRED3_(pred, v1, v2, v3, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred3Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             #v3, \
+                                             pred, \
+                                             v1, \
+                                             v2, \
+                                             v3), on_failure)
+
+// Ternary predicate assertion macros.
+#define EXPECT_PRED_FORMAT3(pred_format, v1, v2, v3) \
+  GTEST_PRED_FORMAT3_(pred_format, v1, v2, v3, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED3(pred, v1, v2, v3) \
+  GTEST_PRED3_(pred, v1, v2, v3, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT3(pred_format, v1, v2, v3) \
+  GTEST_PRED_FORMAT3_(pred_format, v1, v2, v3, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED3(pred, v1, v2, v3) \
+  GTEST_PRED3_(pred, v1, v2, v3, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED4.  Don't use
+// this in your code.
+template <typename Pred,
+          typename T1,
+          typename T2,
+          typename T3,
+          typename T4>
+AssertionResult AssertPred4Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  const char* e3,
+                                  const char* e4,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2,
+                                  const T3& v3,
+                                  const T4& v4) {
+  if (pred(v1, v2, v3, v4)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ", "
+                            << e3 << ", "
+                            << e4 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2
+                            << "\n" << e3 << " evaluates to " << v3
+                            << "\n" << e4 << " evaluates to " << v4;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT4.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT4_(pred_format, v1, v2, v3, v4, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, #v3, #v4, v1, v2, v3, v4), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED4.  Don't use
+// this in your code.
+#define GTEST_PRED4_(pred, v1, v2, v3, v4, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred4Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             #v3, \
+                                             #v4, \
+                                             pred, \
+                                             v1, \
+                                             v2, \
+                                             v3, \
+                                             v4), on_failure)
+
+// 4-ary predicate assertion macros.
+#define EXPECT_PRED_FORMAT4(pred_format, v1, v2, v3, v4) \
+  GTEST_PRED_FORMAT4_(pred_format, v1, v2, v3, v4, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED4(pred, v1, v2, v3, v4) \
+  GTEST_PRED4_(pred, v1, v2, v3, v4, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT4(pred_format, v1, v2, v3, v4) \
+  GTEST_PRED_FORMAT4_(pred_format, v1, v2, v3, v4, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED4(pred, v1, v2, v3, v4) \
+  GTEST_PRED4_(pred, v1, v2, v3, v4, GTEST_FATAL_FAILURE_)
+
+
+
+// Helper function for implementing {EXPECT|ASSERT}_PRED5.  Don't use
+// this in your code.
+template <typename Pred,
+          typename T1,
+          typename T2,
+          typename T3,
+          typename T4,
+          typename T5>
+AssertionResult AssertPred5Helper(const char* pred_text,
+                                  const char* e1,
+                                  const char* e2,
+                                  const char* e3,
+                                  const char* e4,
+                                  const char* e5,
+                                  Pred pred,
+                                  const T1& v1,
+                                  const T2& v2,
+                                  const T3& v3,
+                                  const T4& v4,
+                                  const T5& v5) {
+  if (pred(v1, v2, v3, v4, v5)) return AssertionSuccess();
+
+  return AssertionFailure() << pred_text << "("
+                            << e1 << ", "
+                            << e2 << ", "
+                            << e3 << ", "
+                            << e4 << ", "
+                            << e5 << ") evaluates to false, where"
+                            << "\n" << e1 << " evaluates to " << v1
+                            << "\n" << e2 << " evaluates to " << v2
+                            << "\n" << e3 << " evaluates to " << v3
+                            << "\n" << e4 << " evaluates to " << v4
+                            << "\n" << e5 << " evaluates to " << v5;
+}
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED_FORMAT5.
+// Don't use this in your code.
+#define GTEST_PRED_FORMAT5_(pred_format, v1, v2, v3, v4, v5, on_failure)\
+  GTEST_ASSERT_(pred_format(#v1, #v2, #v3, #v4, #v5, v1, v2, v3, v4, v5), \
+                on_failure)
+
+// Internal macro for implementing {EXPECT|ASSERT}_PRED5.  Don't use
+// this in your code.
+#define GTEST_PRED5_(pred, v1, v2, v3, v4, v5, on_failure)\
+  GTEST_ASSERT_(::testing::AssertPred5Helper(#pred, \
+                                             #v1, \
+                                             #v2, \
+                                             #v3, \
+                                             #v4, \
+                                             #v5, \
+                                             pred, \
+                                             v1, \
+                                             v2, \
+                                             v3, \
+                                             v4, \
+                                             v5), on_failure)
+
+// 5-ary predicate assertion macros.
+#define EXPECT_PRED_FORMAT5(pred_format, v1, v2, v3, v4, v5) \
+  GTEST_PRED_FORMAT5_(pred_format, v1, v2, v3, v4, v5, GTEST_NONFATAL_FAILURE_)
+#define EXPECT_PRED5(pred, v1, v2, v3, v4, v5) \
+  GTEST_PRED5_(pred, v1, v2, v3, v4, v5, GTEST_NONFATAL_FAILURE_)
+#define ASSERT_PRED_FORMAT5(pred_format, v1, v2, v3, v4, v5) \
+  GTEST_PRED_FORMAT5_(pred_format, v1, v2, v3, v4, v5, GTEST_FATAL_FAILURE_)
+#define ASSERT_PRED5(pred, v1, v2, v3, v4, v5) \
+  GTEST_PRED5_(pred, v1, v2, v3, v4, v5, GTEST_FATAL_FAILURE_)
+
+
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PRED_IMPL_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest_prod.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest_prod.h
new file mode 100644
index 0000000000..da80ddc6c7
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/gtest_prod.h
@@ -0,0 +1,58 @@
+// Copyright 2006, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// Google C++ Testing Framework definitions useful in production code.
+
+#ifndef GTEST_INCLUDE_GTEST_GTEST_PROD_H_
+#define GTEST_INCLUDE_GTEST_GTEST_PROD_H_
+
+// When you need to test the private or protected members of a class,
+// use the FRIEND_TEST macro to declare your tests as friends of the
+// class.  For example:
+//
+// class MyClass {
+//  private:
+//   void MyMethod();
+//   FRIEND_TEST(MyClassTest, MyMethod);
+// };
+//
+// class MyClassTest : public testing::Test {
+//   // ...
+// };
+//
+// TEST_F(MyClassTest, MyMethod) {
+//   // Can call MyClass::MyMethod() here.
+// }
+
+#define FRIEND_TEST(test_case_name, test_name)\
+friend class test_case_name##_##test_name##_Test
+
+#endif  // GTEST_INCLUDE_GTEST_GTEST_PROD_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/custom/gtest-port.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/custom/gtest-port.h
new file mode 100644
index 0000000000..7e744bd3bb
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/custom/gtest-port.h
@@ -0,0 +1,69 @@
+// Copyright 2015, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Injection point for custom user configurations.
+// The following macros can be defined:
+//
+//   Flag related macros:
+//     GTEST_FLAG(flag_name)
+//     GTEST_USE_OWN_FLAGFILE_FLAG_  - Define to 0 when the system provides its
+//                                     own flagfile flag parsing.
+//     GTEST_DECLARE_bool_(name)
+//     GTEST_DECLARE_int32_(name)
+//     GTEST_DECLARE_string_(name)
+//     GTEST_DEFINE_bool_(name, default_val, doc)
+//     GTEST_DEFINE_int32_(name, default_val, doc)
+//     GTEST_DEFINE_string_(name, default_val, doc)
+//
+//   Test filtering:
+//     GTEST_TEST_FILTER_ENV_VAR_ - The name of an environment variable that
+//                                  will be used if --GTEST_FLAG(test_filter)
+//                                  is not provided.
+//
+//   Logging:
+//     GTEST_LOG_(severity)
+//     GTEST_CHECK_(condition)
+//     Functions LogToStderr() and FlushInfoLog() have to be provided too.
+//
+//   Threading:
+//     GTEST_HAS_NOTIFICATION_ - Enabled if Notification is already provided.
+//     GTEST_HAS_MUTEX_AND_THREAD_LOCAL_ - Enabled if Mutex and ThreadLocal are
+//                                         already provided.
+//     Must also provide GTEST_DECLARE_STATIC_MUTEX_(mutex) and
+//     GTEST_DEFINE_STATIC_MUTEX_(mutex)
+//
+//     GTEST_EXCLUSIVE_LOCK_REQUIRED_(locks)
+//     GTEST_LOCK_EXCLUDED_(locks)
+//
+// ** Custom implementation starts here **
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_CUSTOM_GTEST_PORT_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_CUSTOM_GTEST_PORT_H_
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_CUSTOM_GTEST_PORT_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/custom/gtest-printers.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/custom/gtest-printers.h
new file mode 100644
index 0000000000..60c1ea050b
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/custom/gtest-printers.h
@@ -0,0 +1,42 @@
+// Copyright 2015, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// This file provides an injection point for custom printers in a local
+// installation of gTest.
+// It will be included from gtest-printers.h and the overrides in this file
+// will be visible to everyone.
+// See documentation at gtest/gtest-printers.h for details on how to define a
+// custom printer.
+//
+// ** Custom implementation starts here **
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_CUSTOM_GTEST_PRINTERS_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_CUSTOM_GTEST_PRINTERS_H_
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_CUSTOM_GTEST_PRINTERS_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/custom/gtest.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/custom/gtest.h
new file mode 100644
index 0000000000..c27412a898
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/custom/gtest.h
@@ -0,0 +1,41 @@
+// Copyright 2015, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Injection point for custom user configurations.
+// The following macros can be defined:
+//
+// GTEST_OS_STACK_TRACE_GETTER_  - The name of an implementation of
+//                                 OsStackTraceGetterInterface.
+//
+// ** Custom implementation starts here **
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_CUSTOM_GTEST_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_CUSTOM_GTEST_H_
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_CUSTOM_GTEST_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-death-test-internal.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-death-test-internal.h
new file mode 100644
index 0000000000..2b3a78f5bf
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-death-test-internal.h
@@ -0,0 +1,319 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan), eefacm@gmail.com (Sean Mcafee)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines internal utilities needed for implementing
+// death tests.  They are subject to change without notice.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_DEATH_TEST_INTERNAL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_DEATH_TEST_INTERNAL_H_
+
+#include "gtest/internal/gtest-internal.h"
+
+#include <stdio.h>
+
+namespace testing {
+namespace internal {
+
+GTEST_DECLARE_string_(internal_run_death_test);
+
+// Names of the flags (needed for parsing Google Test flags).
+const char kDeathTestStyleFlag[] = "death_test_style";
+const char kDeathTestUseFork[] = "death_test_use_fork";
+const char kInternalRunDeathTestFlag[] = "internal_run_death_test";
+
+#if GTEST_HAS_DEATH_TEST
+
+// DeathTest is a class that hides much of the complexity of the
+// GTEST_DEATH_TEST_ macro.  It is abstract; its static Create method
+// returns a concrete class that depends on the prevailing death test
+// style, as defined by the --gtest_death_test_style and/or
+// --gtest_internal_run_death_test flags.
+
+// In describing the results of death tests, these terms are used with
+// the corresponding definitions:
+//
+// exit status:  The integer exit information in the format specified
+//               by wait(2)
+// exit code:    The integer code passed to exit(3), _exit(2), or
+//               returned from main()
+class GTEST_API_ DeathTest {
+ public:
+  // Create returns false if there was an error determining the
+  // appropriate action to take for the current death test; for example,
+  // if the gtest_death_test_style flag is set to an invalid value.
+  // The LastMessage method will return a more detailed message in that
+  // case.  Otherwise, the DeathTest pointer pointed to by the "test"
+  // argument is set.  If the death test should be skipped, the pointer
+  // is set to NULL; otherwise, it is set to the address of a new concrete
+  // DeathTest object that controls the execution of the current test.
+  static bool Create(const char* statement, const RE* regex,
+                     const char* file, int line, DeathTest** test);
+  DeathTest();
+  virtual ~DeathTest() { }
+
+  // A helper class that aborts a death test when it's deleted.
+  class ReturnSentinel {
+   public:
+    explicit ReturnSentinel(DeathTest* test) : test_(test) { }
+    ~ReturnSentinel() { test_->Abort(TEST_ENCOUNTERED_RETURN_STATEMENT); }
+   private:
+    DeathTest* const test_;
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(ReturnSentinel);
+  } GTEST_ATTRIBUTE_UNUSED_;
+
+  // An enumeration of possible roles that may be taken when a death
+  // test is encountered.  EXECUTE means that the death test logic should
+  // be executed immediately.  OVERSEE means that the program should prepare
+  // the appropriate environment for a child process to execute the death
+  // test, then wait for it to complete.
+  enum TestRole { OVERSEE_TEST, EXECUTE_TEST };
+
+  // An enumeration of the three reasons that a test might be aborted.
+  enum AbortReason {
+    TEST_ENCOUNTERED_RETURN_STATEMENT,
+    TEST_THREW_EXCEPTION,
+    TEST_DID_NOT_DIE
+  };
+
+  // Assumes one of the above roles.
+  virtual TestRole AssumeRole() = 0;
+
+  // Waits for the death test to finish and returns its status.
+  virtual int Wait() = 0;
+
+  // Returns true if the death test passed; that is, the test process
+  // exited during the test, its exit status matches a user-supplied
+  // predicate, and its stderr output matches a user-supplied regular
+  // expression.
+  // The user-supplied predicate may be a macro expression rather
+  // than a function pointer or functor, or else Wait and Passed could
+  // be combined.
+  virtual bool Passed(bool exit_status_ok) = 0;
+
+  // Signals that the death test did not die as expected.
+  virtual void Abort(AbortReason reason) = 0;
+
+  // Returns a human-readable outcome message regarding the outcome of
+  // the last death test.
+  static const char* LastMessage();
+
+  static void set_last_death_test_message(const std::string& message);
+
+ private:
+  // A string containing a description of the outcome of the last death test.
+  static std::string last_death_test_message_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(DeathTest);
+};
+
+// Factory interface for death tests.  May be mocked out for testing.
+class DeathTestFactory {
+ public:
+  virtual ~DeathTestFactory() { }
+  virtual bool Create(const char* statement, const RE* regex,
+                      const char* file, int line, DeathTest** test) = 0;
+};
+
+// A concrete DeathTestFactory implementation for normal use.
+class DefaultDeathTestFactory : public DeathTestFactory {
+ public:
+  virtual bool Create(const char* statement, const RE* regex,
+                      const char* file, int line, DeathTest** test);
+};
+
+// Returns true if exit_status describes a process that was terminated
+// by a signal, or exited normally with a nonzero exit code.
+GTEST_API_ bool ExitedUnsuccessfully(int exit_status);
+
+// Traps C++ exceptions escaping statement and reports them as test
+// failures. Note that trapping SEH exceptions is not implemented here.
+# if GTEST_HAS_EXCEPTIONS
+#  define GTEST_EXECUTE_DEATH_TEST_STATEMENT_(statement, death_test) \
+  try { \
+    GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+  } catch (const ::std::exception& gtest_exception) { \
+    fprintf(\
+        stderr, \
+        "\n%s: Caught std::exception-derived exception escaping the " \
+        "death test statement. Exception message: %s\n", \
+        ::testing::internal::FormatFileLocation(__FILE__, __LINE__).c_str(), \
+        gtest_exception.what()); \
+    fflush(stderr); \
+    death_test->Abort(::testing::internal::DeathTest::TEST_THREW_EXCEPTION); \
+  } catch (...) { \
+    death_test->Abort(::testing::internal::DeathTest::TEST_THREW_EXCEPTION); \
+  }
+
+# else
+#  define GTEST_EXECUTE_DEATH_TEST_STATEMENT_(statement, death_test) \
+  GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement)
+
+# endif
+
+// This macro is for implementing ASSERT_DEATH*, EXPECT_DEATH*,
+// ASSERT_EXIT*, and EXPECT_EXIT*.
+# define GTEST_DEATH_TEST_(statement, predicate, regex, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    const ::testing::internal::RE& gtest_regex = (regex); \
+    ::testing::internal::DeathTest* gtest_dt; \
+    if (!::testing::internal::DeathTest::Create(#statement, &gtest_regex, \
+        __FILE__, __LINE__, &gtest_dt)) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_, __LINE__); \
+    } \
+    if (gtest_dt != NULL) { \
+      ::testing::internal::scoped_ptr< ::testing::internal::DeathTest> \
+          gtest_dt_ptr(gtest_dt); \
+      switch (gtest_dt->AssumeRole()) { \
+        case ::testing::internal::DeathTest::OVERSEE_TEST: \
+          if (!gtest_dt->Passed(predicate(gtest_dt->Wait()))) { \
+            goto GTEST_CONCAT_TOKEN_(gtest_label_, __LINE__); \
+          } \
+          break; \
+        case ::testing::internal::DeathTest::EXECUTE_TEST: { \
+          ::testing::internal::DeathTest::ReturnSentinel \
+              gtest_sentinel(gtest_dt); \
+          GTEST_EXECUTE_DEATH_TEST_STATEMENT_(statement, gtest_dt); \
+          gtest_dt->Abort(::testing::internal::DeathTest::TEST_DID_NOT_DIE); \
+          break; \
+        } \
+        default: \
+          break; \
+      } \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_, __LINE__): \
+      fail(::testing::internal::DeathTest::LastMessage())
+// The symbol "fail" here expands to something into which a message
+// can be streamed.
+
+// This macro is for implementing ASSERT/EXPECT_DEBUG_DEATH when compiled in
+// NDEBUG mode. In this case we need the statements to be executed, the regex is
+// ignored, and the macro must accept a streamed message even though the message
+// is never printed.
+# define GTEST_EXECUTE_STATEMENT_(statement, regex) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+     GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+  } else \
+    ::testing::Message()
+
+// A class representing the parsed contents of the
+// --gtest_internal_run_death_test flag, as it existed when
+// RUN_ALL_TESTS was called.
+class InternalRunDeathTestFlag {
+ public:
+  InternalRunDeathTestFlag(const std::string& a_file,
+                           int a_line,
+                           int an_index,
+                           int a_write_fd)
+      : file_(a_file), line_(a_line), index_(an_index),
+        write_fd_(a_write_fd) {}
+
+  ~InternalRunDeathTestFlag() {
+    if (write_fd_ >= 0)
+      posix::Close(write_fd_);
+  }
+
+  const std::string& file() const { return file_; }
+  int line() const { return line_; }
+  int index() const { return index_; }
+  int write_fd() const { return write_fd_; }
+
+ private:
+  std::string file_;
+  int line_;
+  int index_;
+  int write_fd_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(InternalRunDeathTestFlag);
+};
+
+// Returns a newly created InternalRunDeathTestFlag object with fields
+// initialized from the GTEST_FLAG(internal_run_death_test) flag if
+// the flag is specified; otherwise returns NULL.
+InternalRunDeathTestFlag* ParseInternalRunDeathTestFlag();
+
+#else  // GTEST_HAS_DEATH_TEST
+
+// This macro is used for implementing macros such as
+// EXPECT_DEATH_IF_SUPPORTED and ASSERT_DEATH_IF_SUPPORTED on systems where
+// death tests are not supported. Those macros must compile on such systems
+// iff EXPECT_DEATH and ASSERT_DEATH compile with the same parameters on
+// systems that support death tests. This allows one to write such a macro
+// on a system that does not support death tests and be sure that it will
+// compile on a death-test supporting system.
+//
+// Parameters:
+//   statement -  A statement that a macro such as EXPECT_DEATH would test
+//                for program termination. This macro has to make sure this
+//                statement is compiled but not executed, to ensure that
+//                EXPECT_DEATH_IF_SUPPORTED compiles with a certain
+//                parameter iff EXPECT_DEATH compiles with it.
+//   regex     -  A regex that a macro such as EXPECT_DEATH would use to test
+//                the output of statement.  This parameter has to be
+//                compiled but not evaluated by this macro, to ensure that
+//                this macro only accepts expressions that a macro such as
+//                EXPECT_DEATH would accept.
+//   terminator - Must be an empty statement for EXPECT_DEATH_IF_SUPPORTED
+//                and a return statement for ASSERT_DEATH_IF_SUPPORTED.
+//                This ensures that ASSERT_DEATH_IF_SUPPORTED will not
+//                compile inside functions where ASSERT_DEATH doesn't
+//                compile.
+//
+//  The branch that has an always false condition is used to ensure that
+//  statement and regex are compiled (and thus syntactically correct) but
+//  never executed. The unreachable code macro protects the terminator
+//  statement from generating an 'unreachable code' warning in case
+//  statement unconditionally returns or throws. The Message constructor at
+//  the end allows the syntax of streaming additional messages into the
+//  macro, for compilational compatibility with EXPECT_DEATH/ASSERT_DEATH.
+# define GTEST_UNSUPPORTED_DEATH_TEST_(statement, regex, terminator) \
+    GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+    if (::testing::internal::AlwaysTrue()) { \
+      GTEST_LOG_(WARNING) \
+          << "Death tests are not supported on this platform.\n" \
+          << "Statement '" #statement "' cannot be verified."; \
+    } else if (::testing::internal::AlwaysFalse()) { \
+      ::testing::internal::RE::PartialMatch(".*", (regex)); \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+      terminator; \
+    } else \
+      ::testing::Message()
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_DEATH_TEST_INTERNAL_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-filepath.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-filepath.h
new file mode 100644
index 0000000000..7a13b4b0de
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-filepath.h
@@ -0,0 +1,206 @@
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: keith.ray@gmail.com (Keith Ray)
+//
+// Google Test filepath utilities
+//
+// This header file declares classes and functions used internally by
+// Google Test.  They are subject to change without notice.
+//
+// This file is #included in <gtest/internal/gtest-internal.h>.
+// Do not include this header file separately!
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_FILEPATH_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_FILEPATH_H_
+
+#include "gtest/internal/gtest-string.h"
+
+namespace testing {
+namespace internal {
+
+// FilePath - a class for file and directory pathname manipulation which
+// handles platform-specific conventions (like the pathname separator).
+// Used for helper functions for naming files in a directory for xml output.
+// Except for Set methods, all methods are const or static, which provides an
+// "immutable value object" -- useful for peace of mind.
+// A FilePath with a value ending in a path separator ("like/this/") represents
+// a directory, otherwise it is assumed to represent a file. In either case,
+// it may or may not represent an actual file or directory in the file system.
+// Names are NOT checked for syntax correctness -- no checking for illegal
+// characters, malformed paths, etc.
+
+class GTEST_API_ FilePath {
+ public:
+  FilePath() : pathname_("") { }
+  FilePath(const FilePath& rhs) : pathname_(rhs.pathname_) { }
+
+  explicit FilePath(const std::string& pathname) : pathname_(pathname) {
+    Normalize();
+  }
+
+  FilePath& operator=(const FilePath& rhs) {
+    Set(rhs);
+    return *this;
+  }
+
+  void Set(const FilePath& rhs) {
+    pathname_ = rhs.pathname_;
+  }
+
+  const std::string& string() const { return pathname_; }
+  const char* c_str() const { return pathname_.c_str(); }
+
+  // Returns the current working directory, or "" if unsuccessful.
+  static FilePath GetCurrentDir();
+
+  // Given directory = "dir", base_name = "test", number = 0,
+  // extension = "xml", returns "dir/test.xml". If number is greater
+  // than zero (e.g., 12), returns "dir/test_12.xml".
+  // On Windows platform, uses \ as the separator rather than /.
+  static FilePath MakeFileName(const FilePath& directory,
+                               const FilePath& base_name,
+                               int number,
+                               const char* extension);
+
+  // Given directory = "dir", relative_path = "test.xml",
+  // returns "dir/test.xml".
+  // On Windows, uses \ as the separator rather than /.
+  static FilePath ConcatPaths(const FilePath& directory,
+                              const FilePath& relative_path);
+
+  // Returns a pathname for a file that does not currently exist. The pathname
+  // will be directory/base_name.extension or
+  // directory/base_name_<number>.extension if directory/base_name.extension
+  // already exists. The number will be incremented until a pathname is found
+  // that does not already exist.
+  // Examples: 'dir/foo_test.xml' or 'dir/foo_test_1.xml'.
+  // There could be a race condition if two or more processes are calling this
+  // function at the same time -- they could both pick the same filename.
+  static FilePath GenerateUniqueFileName(const FilePath& directory,
+                                         const FilePath& base_name,
+                                         const char* extension);
+
+  // Returns true iff the path is "".
+  bool IsEmpty() const { return pathname_.empty(); }
+
+  // If input name has a trailing separator character, removes it and returns
+  // the name, otherwise return the name string unmodified.
+  // On Windows platform, uses \ as the separator, other platforms use /.
+  FilePath RemoveTrailingPathSeparator() const;
+
+  // Returns a copy of the FilePath with the directory part removed.
+  // Example: FilePath("path/to/file").RemoveDirectoryName() returns
+  // FilePath("file"). If there is no directory part ("just_a_file"), it returns
+  // the FilePath unmodified. If there is no file part ("just_a_dir/") it
+  // returns an empty FilePath ("").
+  // On Windows platform, '\' is the path separator, otherwise it is '/'.
+  FilePath RemoveDirectoryName() const;
+
+  // RemoveFileName returns the directory path with the filename removed.
+  // Example: FilePath("path/to/file").RemoveFileName() returns "path/to/".
+  // If the FilePath is "a_file" or "/a_file", RemoveFileName returns
+  // FilePath("./") or, on Windows, FilePath(".\\"). If the filepath does
+  // not have a file, like "just/a/dir/", it returns the FilePath unmodified.
+  // On Windows platform, '\' is the path separator, otherwise it is '/'.
+  FilePath RemoveFileName() const;
+
+  // Returns a copy of the FilePath with the case-insensitive extension removed.
+  // Example: FilePath("dir/file.exe").RemoveExtension("EXE") returns
+  // FilePath("dir/file"). If a case-insensitive extension is not
+  // found, returns a copy of the original FilePath.
+  FilePath RemoveExtension(const char* extension) const;
+
+  // Creates directories so that path exists. Returns true if successful or if
+  // the directories already exist; returns false if unable to create
+  // directories for any reason. Will also return false if the FilePath does
+  // not represent a directory (that is, it doesn't end with a path separator).
+  bool CreateDirectoriesRecursively() const;
+
+  // Create the directory so that path exists. Returns true if successful or
+  // if the directory already exists; returns false if unable to create the
+  // directory for any reason, including if the parent directory does not
+  // exist. Not named "CreateDirectory" because that's a macro on Windows.
+  bool CreateFolder() const;
+
+  // Returns true if FilePath describes something in the file-system,
+  // either a file, directory, or whatever, and that something exists.
+  bool FileOrDirectoryExists() const;
+
+  // Returns true if pathname describes a directory in the file-system
+  // that exists.
+  bool DirectoryExists() const;
+
+  // Returns true if FilePath ends with a path separator, which indicates that
+  // it is intended to represent a directory. Returns false otherwise.
+  // This does NOT check that a directory (or file) actually exists.
+  bool IsDirectory() const;
+
+  // Returns true if pathname describes a root directory. (Windows has one
+  // root directory per disk drive.)
+  bool IsRootDirectory() const;
+
+  // Returns true if pathname describes an absolute path.
+  bool IsAbsolutePath() const;
+
+ private:
+  // Replaces multiple consecutive separators with a single separator.
+  // For example, "bar///foo" becomes "bar/foo". Does not eliminate other
+  // redundancies that might be in a pathname involving "." or "..".
+  //
+  // A pathname with multiple consecutive separators may occur either through
+  // user error or as a result of some scripts or APIs that generate a pathname
+  // with a trailing separator. On other platforms the same API or script
+  // may NOT generate a pathname with a trailing "/". Then elsewhere that
+  // pathname may have another "/" and pathname components added to it,
+  // without checking for the separator already being there.
+  // The script language and operating system may allow paths like "foo//bar"
+  // but some of the functions in FilePath will not handle that correctly. In
+  // particular, RemoveTrailingPathSeparator() only removes one separator, and
+  // it is called in CreateDirectoriesRecursively() assuming that it will change
+  // a pathname from directory syntax (trailing separator) to filename syntax.
+  //
+  // On Windows this method also replaces the alternate path separator '/' with
+  // the primary path separator '\\', so that for example "bar\\/\\foo" becomes
+  // "bar\\foo".
+
+  void Normalize();
+
+  // Returns a pointer to the last occurence of a valid path separator in
+  // the FilePath. On Windows, for example, both '/' and '\' are valid path
+  // separators. Returns NULL if no path separator was found.
+  const char* FindLastPathSeparator() const;
+
+  std::string pathname_;
+};  // class FilePath
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_FILEPATH_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-internal.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-internal.h
new file mode 100644
index 0000000000..ebd1cf615d
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-internal.h
@@ -0,0 +1,1238 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan), eefacm@gmail.com (Sean Mcafee)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file declares functions and macros used internally by
+// Google Test.  They are subject to change without notice.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_INTERNAL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_INTERNAL_H_
+
+#include "gtest/internal/gtest-port.h"
+
+#if GTEST_OS_LINUX
+# include <stdlib.h>
+# include <sys/types.h>
+# include <sys/wait.h>
+# include <unistd.h>
+#endif  // GTEST_OS_LINUX
+
+#if GTEST_HAS_EXCEPTIONS
+# include <stdexcept>
+#endif
+
+#include <ctype.h>
+#include <float.h>
+#include <string.h>
+#include <iomanip>
+#include <limits>
+#include <map>
+#include <set>
+#include <string>
+#include <vector>
+
+#include "gtest/gtest-message.h"
+#include "gtest/internal/gtest-string.h"
+#include "gtest/internal/gtest-filepath.h"
+#include "gtest/internal/gtest-type-util.h"
+
+// Due to C++ preprocessor weirdness, we need double indirection to
+// concatenate two tokens when one of them is __LINE__.  Writing
+//
+//   foo ## __LINE__
+//
+// will result in the token foo__LINE__, instead of foo followed by
+// the current line number.  For more details, see
+// http://www.parashift.com/c++-faq-lite/misc-technical-issues.html#faq-39.6
+#define GTEST_CONCAT_TOKEN_(foo, bar) GTEST_CONCAT_TOKEN_IMPL_(foo, bar)
+#define GTEST_CONCAT_TOKEN_IMPL_(foo, bar) foo ## bar
+
+class ProtocolMessage;
+namespace proto2 { class Message; }
+
+namespace testing {
+
+// Forward declarations.
+
+class AssertionResult;                 // Result of an assertion.
+class Message;                         // Represents a failure message.
+class Test;                            // Represents a test.
+class TestInfo;                        // Information about a test.
+class TestPartResult;                  // Result of a test part.
+class UnitTest;                        // A collection of test cases.
+
+template <typename T>
+::std::string PrintToString(const T& value);
+
+namespace internal {
+
+struct TraceInfo;                      // Information about a trace point.
+class ScopedTrace;                     // Implements scoped trace.
+class TestInfoImpl;                    // Opaque implementation of TestInfo
+class UnitTestImpl;                    // Opaque implementation of UnitTest
+
+// The text used in failure messages to indicate the start of the
+// stack trace.
+GTEST_API_ extern const char kStackTraceMarker[];
+
+// Two overloaded helpers for checking at compile time whether an
+// expression is a null pointer literal (i.e. NULL or any 0-valued
+// compile-time integral constant).  Their return values have
+// different sizes, so we can use sizeof() to test which version is
+// picked by the compiler.  These helpers have no implementations, as
+// we only need their signatures.
+//
+// Given IsNullLiteralHelper(x), the compiler will pick the first
+// version if x can be implicitly converted to Secret*, and pick the
+// second version otherwise.  Since Secret is a secret and incomplete
+// type, the only expression a user can write that has type Secret* is
+// a null pointer literal.  Therefore, we know that x is a null
+// pointer literal if and only if the first version is picked by the
+// compiler.
+char IsNullLiteralHelper(Secret* p);
+char (&IsNullLiteralHelper(...))[2];  // NOLINT
+
+// A compile-time bool constant that is true if and only if x is a
+// null pointer literal (i.e. NULL or any 0-valued compile-time
+// integral constant).
+#ifdef GTEST_ELLIPSIS_NEEDS_POD_
+// We lose support for NULL detection where the compiler doesn't like
+// passing non-POD classes through ellipsis (...).
+# define GTEST_IS_NULL_LITERAL_(x) false
+#else
+# define GTEST_IS_NULL_LITERAL_(x) \
+    (sizeof(::testing::internal::IsNullLiteralHelper(x)) == 1)
+#endif  // GTEST_ELLIPSIS_NEEDS_POD_
+
+// Appends the user-supplied message to the Google-Test-generated message.
+GTEST_API_ std::string AppendUserMessage(
+    const std::string& gtest_msg, const Message& user_msg);
+
+#if GTEST_HAS_EXCEPTIONS
+
+// This exception is thrown by (and only by) a failed Google Test
+// assertion when GTEST_FLAG(throw_on_failure) is true (if exceptions
+// are enabled).  We derive it from std::runtime_error, which is for
+// errors presumably detectable only at run time.  Since
+// std::runtime_error inherits from std::exception, many testing
+// frameworks know how to extract and print the message inside it.
+class GTEST_API_ GoogleTestFailureException : public ::std::runtime_error {
+ public:
+  explicit GoogleTestFailureException(const TestPartResult& failure);
+};
+
+#endif  // GTEST_HAS_EXCEPTIONS
+
+// A helper class for creating scoped traces in user programs.
+class GTEST_API_ ScopedTrace {
+ public:
+  // The c'tor pushes the given source file location and message onto
+  // a trace stack maintained by Google Test.
+  ScopedTrace(const char* file, int line, const Message& message);
+
+  // The d'tor pops the info pushed by the c'tor.
+  //
+  // Note that the d'tor is not virtual in order to be efficient.
+  // Don't inherit from ScopedTrace!
+  ~ScopedTrace();
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ScopedTrace);
+} GTEST_ATTRIBUTE_UNUSED_;  // A ScopedTrace object does its job in its
+                            // c'tor and d'tor.  Therefore it doesn't
+                            // need to be used otherwise.
+
+namespace edit_distance {
+// Returns the optimal edits to go from 'left' to 'right'.
+// All edits cost the same, with replace having lower priority than
+// add/remove.
+// Simple implementation of the Wagner–Fischer algorithm.
+// See http://en.wikipedia.org/wiki/Wagner-Fischer_algorithm
+enum EditType { kMatch, kAdd, kRemove, kReplace };
+GTEST_API_ std::vector<EditType> CalculateOptimalEdits(
+    const std::vector<size_t>& left, const std::vector<size_t>& right);
+
+// Same as above, but the input is represented as strings.
+GTEST_API_ std::vector<EditType> CalculateOptimalEdits(
+    const std::vector<std::string>& left,
+    const std::vector<std::string>& right);
+
+// Create a diff of the input strings in Unified diff format.
+GTEST_API_ std::string CreateUnifiedDiff(const std::vector<std::string>& left,
+                                         const std::vector<std::string>& right,
+                                         size_t context = 2);
+
+}  // namespace edit_distance
+
+// Calculate the diff between 'left' and 'right' and return it in unified diff
+// format.
+// If not null, stores in 'total_line_count' the total number of lines found
+// in left + right.
+GTEST_API_ std::string DiffStrings(const std::string& left,
+                                   const std::string& right,
+                                   size_t* total_line_count);
+
+// Constructs and returns the message for an equality assertion
+// (e.g. ASSERT_EQ, EXPECT_STREQ, etc) failure.
+//
+// The first four parameters are the expressions used in the assertion
+// and their values, as strings.  For example, for ASSERT_EQ(foo, bar)
+// where foo is 5 and bar is 6, we have:
+//
+//   expected_expression: "foo"
+//   actual_expression:   "bar"
+//   expected_value:      "5"
+//   actual_value:        "6"
+//
+// The ignoring_case parameter is true iff the assertion is a
+// *_STRCASEEQ*.  When it's true, the string " (ignoring case)" will
+// be inserted into the message.
+GTEST_API_ AssertionResult EqFailure(const char* expected_expression,
+                                     const char* actual_expression,
+                                     const std::string& expected_value,
+                                     const std::string& actual_value,
+                                     bool ignoring_case);
+
+// Constructs a failure message for Boolean assertions such as EXPECT_TRUE.
+GTEST_API_ std::string GetBoolAssertionFailureMessage(
+    const AssertionResult& assertion_result,
+    const char* expression_text,
+    const char* actual_predicate_value,
+    const char* expected_predicate_value);
+
+// This template class represents an IEEE floating-point number
+// (either single-precision or double-precision, depending on the
+// template parameters).
+//
+// The purpose of this class is to do more sophisticated number
+// comparison.  (Due to round-off error, etc, it's very unlikely that
+// two floating-points will be equal exactly.  Hence a naive
+// comparison by the == operation often doesn't work.)
+//
+// Format of IEEE floating-point:
+//
+//   The most-significant bit being the leftmost, an IEEE
+//   floating-point looks like
+//
+//     sign_bit exponent_bits fraction_bits
+//
+//   Here, sign_bit is a single bit that designates the sign of the
+//   number.
+//
+//   For float, there are 8 exponent bits and 23 fraction bits.
+//
+//   For double, there are 11 exponent bits and 52 fraction bits.
+//
+//   More details can be found at
+//   http://en.wikipedia.org/wiki/IEEE_floating-point_standard.
+//
+// Template parameter:
+//
+//   RawType: the raw floating-point type (either float or double)
+template <typename RawType>
+class FloatingPoint {
+ public:
+  // Defines the unsigned integer type that has the same size as the
+  // floating point number.
+  typedef typename TypeWithSize<sizeof(RawType)>::UInt Bits;
+
+  // Constants.
+
+  // # of bits in a number.
+  static const size_t kBitCount = 8*sizeof(RawType);
+
+  // # of fraction bits in a number.
+  static const size_t kFractionBitCount =
+    std::numeric_limits<RawType>::digits - 1;
+
+  // # of exponent bits in a number.
+  static const size_t kExponentBitCount = kBitCount - 1 - kFractionBitCount;
+
+  // The mask for the sign bit.
+  static const Bits kSignBitMask = static_cast<Bits>(1) << (kBitCount - 1);
+
+  // The mask for the fraction bits.
+  static const Bits kFractionBitMask =
+    ~static_cast<Bits>(0) >> (kExponentBitCount + 1);
+
+  // The mask for the exponent bits.
+  static const Bits kExponentBitMask = ~(kSignBitMask | kFractionBitMask);
+
+  // How many ULP's (Units in the Last Place) we want to tolerate when
+  // comparing two numbers.  The larger the value, the more error we
+  // allow.  A 0 value means that two numbers must be exactly the same
+  // to be considered equal.
+  //
+  // The maximum error of a single floating-point operation is 0.5
+  // units in the last place.  On Intel CPU's, all floating-point
+  // calculations are done with 80-bit precision, while double has 64
+  // bits.  Therefore, 4 should be enough for ordinary use.
+  //
+  // See the following article for more details on ULP:
+  // http://randomascii.wordpress.com/2012/02/25/comparing-floating-point-numbers-2012-edition/
+  static const size_t kMaxUlps = 4;
+
+  // Constructs a FloatingPoint from a raw floating-point number.
+  //
+  // On an Intel CPU, passing a non-normalized NAN (Not a Number)
+  // around may change its bits, although the new value is guaranteed
+  // to be also a NAN.  Therefore, don't expect this constructor to
+  // preserve the bits in x when x is a NAN.
+  explicit FloatingPoint(const RawType& x) { u_.value_ = x; }
+
+  // Static methods
+
+  // Reinterprets a bit pattern as a floating-point number.
+  //
+  // This function is needed to test the AlmostEquals() method.
+  static RawType ReinterpretBits(const Bits bits) {
+    FloatingPoint fp(0);
+    fp.u_.bits_ = bits;
+    return fp.u_.value_;
+  }
+
+  // Returns the floating-point number that represent positive infinity.
+  static RawType Infinity() {
+    return ReinterpretBits(kExponentBitMask);
+  }
+
+  // Returns the maximum representable finite floating-point number.
+  static RawType Max();
+
+  // Non-static methods
+
+  // Returns the bits that represents this number.
+  const Bits &bits() const { return u_.bits_; }
+
+  // Returns the exponent bits of this number.
+  Bits exponent_bits() const { return kExponentBitMask & u_.bits_; }
+
+  // Returns the fraction bits of this number.
+  Bits fraction_bits() const { return kFractionBitMask & u_.bits_; }
+
+  // Returns the sign bit of this number.
+  Bits sign_bit() const { return kSignBitMask & u_.bits_; }
+
+  // Returns true iff this is NAN (not a number).
+  bool is_nan() const {
+    // It's a NAN if the exponent bits are all ones and the fraction
+    // bits are not entirely zeros.
+    return (exponent_bits() == kExponentBitMask) && (fraction_bits() != 0);
+  }
+
+  // Returns true iff this number is at most kMaxUlps ULP's away from
+  // rhs.  In particular, this function:
+  //
+  //   - returns false if either number is (or both are) NAN.
+  //   - treats really large numbers as almost equal to infinity.
+  //   - thinks +0.0 and -0.0 are 0 DLP's apart.
+  bool AlmostEquals(const FloatingPoint& rhs) const {
+    // The IEEE standard says that any comparison operation involving
+    // a NAN must return false.
+    if (is_nan() || rhs.is_nan()) return false;
+
+    return DistanceBetweenSignAndMagnitudeNumbers(u_.bits_, rhs.u_.bits_)
+        <= kMaxUlps;
+  }
+
+ private:
+  // The data type used to store the actual floating-point number.
+  union FloatingPointUnion {
+    RawType value_;  // The raw floating-point number.
+    Bits bits_;      // The bits that represent the number.
+  };
+
+  // Converts an integer from the sign-and-magnitude representation to
+  // the biased representation.  More precisely, let N be 2 to the
+  // power of (kBitCount - 1), an integer x is represented by the
+  // unsigned number x + N.
+  //
+  // For instance,
+  //
+  //   -N + 1 (the most negative number representable using
+  //          sign-and-magnitude) is represented by 1;
+  //   0      is represented by N; and
+  //   N - 1  (the biggest number representable using
+  //          sign-and-magnitude) is represented by 2N - 1.
+  //
+  // Read http://en.wikipedia.org/wiki/Signed_number_representations
+  // for more details on signed number representations.
+  static Bits SignAndMagnitudeToBiased(const Bits &sam) {
+    if (kSignBitMask & sam) {
+      // sam represents a negative number.
+      return ~sam + 1;
+    } else {
+      // sam represents a positive number.
+      return kSignBitMask | sam;
+    }
+  }
+
+  // Given two numbers in the sign-and-magnitude representation,
+  // returns the distance between them as an unsigned number.
+  static Bits DistanceBetweenSignAndMagnitudeNumbers(const Bits &sam1,
+                                                     const Bits &sam2) {
+    const Bits biased1 = SignAndMagnitudeToBiased(sam1);
+    const Bits biased2 = SignAndMagnitudeToBiased(sam2);
+    return (biased1 >= biased2) ? (biased1 - biased2) : (biased2 - biased1);
+  }
+
+  FloatingPointUnion u_;
+};
+
+// We cannot use std::numeric_limits<T>::max() as it clashes with the max()
+// macro defined by <windows.h>.
+template <>
+inline float FloatingPoint<float>::Max() { return FLT_MAX; }
+template <>
+inline double FloatingPoint<double>::Max() { return DBL_MAX; }
+
+// Typedefs the instances of the FloatingPoint template class that we
+// care to use.
+typedef FloatingPoint<float> Float;
+typedef FloatingPoint<double> Double;
+
+// In order to catch the mistake of putting tests that use different
+// test fixture classes in the same test case, we need to assign
+// unique IDs to fixture classes and compare them.  The TypeId type is
+// used to hold such IDs.  The user should treat TypeId as an opaque
+// type: the only operation allowed on TypeId values is to compare
+// them for equality using the == operator.
+typedef const void* TypeId;
+
+template <typename T>
+class TypeIdHelper {
+ public:
+  // dummy_ must not have a const type.  Otherwise an overly eager
+  // compiler (e.g. MSVC 7.1 & 8.0) may try to merge
+  // TypeIdHelper<T>::dummy_ for different Ts as an "optimization".
+  static bool dummy_;
+};
+
+template <typename T>
+bool TypeIdHelper<T>::dummy_ = false;
+
+// GetTypeId<T>() returns the ID of type T.  Different values will be
+// returned for different types.  Calling the function twice with the
+// same type argument is guaranteed to return the same ID.
+template <typename T>
+TypeId GetTypeId() {
+  // The compiler is required to allocate a different
+  // TypeIdHelper<T>::dummy_ variable for each T used to instantiate
+  // the template.  Therefore, the address of dummy_ is guaranteed to
+  // be unique.
+  return &(TypeIdHelper<T>::dummy_);
+}
+
+// Returns the type ID of ::testing::Test.  Always call this instead
+// of GetTypeId< ::testing::Test>() to get the type ID of
+// ::testing::Test, as the latter may give the wrong result due to a
+// suspected linker bug when compiling Google Test as a Mac OS X
+// framework.
+GTEST_API_ TypeId GetTestTypeId();
+
+// Defines the abstract factory interface that creates instances
+// of a Test object.
+class TestFactoryBase {
+ public:
+  virtual ~TestFactoryBase() {}
+
+  // Creates a test instance to run. The instance is both created and destroyed
+  // within TestInfoImpl::Run()
+  virtual Test* CreateTest() = 0;
+
+ protected:
+  TestFactoryBase() {}
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestFactoryBase);
+};
+
+// This class provides implementation of TeastFactoryBase interface.
+// It is used in TEST and TEST_F macros.
+template <class TestClass>
+class TestFactoryImpl : public TestFactoryBase {
+ public:
+  virtual Test* CreateTest() { return new TestClass; }
+};
+
+#if GTEST_OS_WINDOWS
+
+// Predicate-formatters for implementing the HRESULT checking macros
+// {ASSERT|EXPECT}_HRESULT_{SUCCEEDED|FAILED}
+// We pass a long instead of HRESULT to avoid causing an
+// include dependency for the HRESULT type.
+GTEST_API_ AssertionResult IsHRESULTSuccess(const char* expr,
+                                            long hr);  // NOLINT
+GTEST_API_ AssertionResult IsHRESULTFailure(const char* expr,
+                                            long hr);  // NOLINT
+
+#endif  // GTEST_OS_WINDOWS
+
+// Types of SetUpTestCase() and TearDownTestCase() functions.
+typedef void (*SetUpTestCaseFunc)();
+typedef void (*TearDownTestCaseFunc)();
+
+struct CodeLocation {
+  CodeLocation(const string& a_file, int a_line) : file(a_file), line(a_line) {}
+
+  string file;
+  int line;
+};
+
+// Creates a new TestInfo object and registers it with Google Test;
+// returns the created object.
+//
+// Arguments:
+//
+//   test_case_name:   name of the test case
+//   name:             name of the test
+//   type_param        the name of the test's type parameter, or NULL if
+//                     this is not a typed or a type-parameterized test.
+//   value_param       text representation of the test's value parameter,
+//                     or NULL if this is not a type-parameterized test.
+//   code_location:    code location where the test is defined
+//   fixture_class_id: ID of the test fixture class
+//   set_up_tc:        pointer to the function that sets up the test case
+//   tear_down_tc:     pointer to the function that tears down the test case
+//   factory:          pointer to the factory that creates a test object.
+//                     The newly created TestInfo instance will assume
+//                     ownership of the factory object.
+GTEST_API_ TestInfo* MakeAndRegisterTestInfo(
+    const char* test_case_name,
+    const char* name,
+    const char* type_param,
+    const char* value_param,
+    CodeLocation code_location,
+    TypeId fixture_class_id,
+    SetUpTestCaseFunc set_up_tc,
+    TearDownTestCaseFunc tear_down_tc,
+    TestFactoryBase* factory);
+
+// If *pstr starts with the given prefix, modifies *pstr to be right
+// past the prefix and returns true; otherwise leaves *pstr unchanged
+// and returns false.  None of pstr, *pstr, and prefix can be NULL.
+GTEST_API_ bool SkipPrefix(const char* prefix, const char** pstr);
+
+#if GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+// State of the definition of a type-parameterized test case.
+class GTEST_API_ TypedTestCasePState {
+ public:
+  TypedTestCasePState() : registered_(false) {}
+
+  // Adds the given test name to defined_test_names_ and return true
+  // if the test case hasn't been registered; otherwise aborts the
+  // program.
+  bool AddTestName(const char* file, int line, const char* case_name,
+                   const char* test_name) {
+    if (registered_) {
+      fprintf(stderr, "%s Test %s must be defined before "
+              "REGISTER_TYPED_TEST_CASE_P(%s, ...).\n",
+              FormatFileLocation(file, line).c_str(), test_name, case_name);
+      fflush(stderr);
+      posix::Abort();
+    }
+    registered_tests_.insert(
+        ::std::make_pair(test_name, CodeLocation(file, line)));
+    return true;
+  }
+
+  bool TestExists(const std::string& test_name) const {
+    return registered_tests_.count(test_name) > 0;
+  }
+
+  const CodeLocation& GetCodeLocation(const std::string& test_name) const {
+    RegisteredTestsMap::const_iterator it = registered_tests_.find(test_name);
+    GTEST_CHECK_(it != registered_tests_.end());
+    return it->second;
+  }
+
+  // Verifies that registered_tests match the test names in
+  // defined_test_names_; returns registered_tests if successful, or
+  // aborts the program otherwise.
+  const char* VerifyRegisteredTestNames(
+      const char* file, int line, const char* registered_tests);
+
+ private:
+  typedef ::std::map<std::string, CodeLocation> RegisteredTestsMap;
+
+  bool registered_;
+  RegisteredTestsMap registered_tests_;
+};
+
+// Skips to the first non-space char after the first comma in 'str';
+// returns NULL if no comma is found in 'str'.
+inline const char* SkipComma(const char* str) {
+  const char* comma = strchr(str, ',');
+  if (comma == NULL) {
+    return NULL;
+  }
+  while (IsSpace(*(++comma))) {}
+  return comma;
+}
+
+// Returns the prefix of 'str' before the first comma in it; returns
+// the entire string if it contains no comma.
+inline std::string GetPrefixUntilComma(const char* str) {
+  const char* comma = strchr(str, ',');
+  return comma == NULL ? str : std::string(str, comma);
+}
+
+// Splits a given string on a given delimiter, populating a given
+// vector with the fields.
+void SplitString(const ::std::string& str, char delimiter,
+                 ::std::vector< ::std::string>* dest);
+
+// TypeParameterizedTest<Fixture, TestSel, Types>::Register()
+// registers a list of type-parameterized tests with Google Test.  The
+// return value is insignificant - we just need to return something
+// such that we can call this function in a namespace scope.
+//
+// Implementation note: The GTEST_TEMPLATE_ macro declares a template
+// template parameter.  It's defined in gtest-type-util.h.
+template <GTEST_TEMPLATE_ Fixture, class TestSel, typename Types>
+class TypeParameterizedTest {
+ public:
+  // 'index' is the index of the test in the type list 'Types'
+  // specified in INSTANTIATE_TYPED_TEST_CASE_P(Prefix, TestCase,
+  // Types).  Valid values for 'index' are [0, N - 1] where N is the
+  // length of Types.
+  static bool Register(const char* prefix,
+                       CodeLocation code_location,
+                       const char* case_name, const char* test_names,
+                       int index) {
+    typedef typename Types::Head Type;
+    typedef Fixture<Type> FixtureClass;
+    typedef typename GTEST_BIND_(TestSel, Type) TestClass;
+
+    // First, registers the first type-parameterized test in the type
+    // list.
+    MakeAndRegisterTestInfo(
+        (std::string(prefix) + (prefix[0] == '\0' ? "" : "/") + case_name + "/"
+         + StreamableToString(index)).c_str(),
+        StripTrailingSpaces(GetPrefixUntilComma(test_names)).c_str(),
+        GetTypeName<Type>().c_str(),
+        NULL,  // No value parameter.
+        code_location,
+        GetTypeId<FixtureClass>(),
+        TestClass::SetUpTestCase,
+        TestClass::TearDownTestCase,
+        new TestFactoryImpl<TestClass>);
+
+    // Next, recurses (at compile time) with the tail of the type list.
+    return TypeParameterizedTest<Fixture, TestSel, typename Types::Tail>
+        ::Register(prefix, code_location, case_name, test_names, index + 1);
+  }
+};
+
+// The base case for the compile time recursion.
+template <GTEST_TEMPLATE_ Fixture, class TestSel>
+class TypeParameterizedTest<Fixture, TestSel, Types0> {
+ public:
+  static bool Register(const char* /*prefix*/, CodeLocation,
+                       const char* /*case_name*/, const char* /*test_names*/,
+                       int /*index*/) {
+    return true;
+  }
+};
+
+// TypeParameterizedTestCase<Fixture, Tests, Types>::Register()
+// registers *all combinations* of 'Tests' and 'Types' with Google
+// Test.  The return value is insignificant - we just need to return
+// something such that we can call this function in a namespace scope.
+template <GTEST_TEMPLATE_ Fixture, typename Tests, typename Types>
+class TypeParameterizedTestCase {
+ public:
+  static bool Register(const char* prefix, CodeLocation code_location,
+                       const TypedTestCasePState* state,
+                       const char* case_name, const char* test_names) {
+    std::string test_name = StripTrailingSpaces(
+        GetPrefixUntilComma(test_names));
+    if (!state->TestExists(test_name)) {
+      fprintf(stderr, "Failed to get code location for test %s.%s at %s.",
+              case_name, test_name.c_str(),
+              FormatFileLocation(code_location.file.c_str(),
+                                 code_location.line).c_str());
+      fflush(stderr);
+      posix::Abort();
+    }
+    const CodeLocation& test_location = state->GetCodeLocation(test_name);
+
+    typedef typename Tests::Head Head;
+
+    // First, register the first test in 'Test' for each type in 'Types'.
+    TypeParameterizedTest<Fixture, Head, Types>::Register(
+        prefix, test_location, case_name, test_names, 0);
+
+    // Next, recurses (at compile time) with the tail of the test list.
+    return TypeParameterizedTestCase<Fixture, typename Tests::Tail, Types>
+        ::Register(prefix, code_location, state,
+                   case_name, SkipComma(test_names));
+  }
+};
+
+// The base case for the compile time recursion.
+template <GTEST_TEMPLATE_ Fixture, typename Types>
+class TypeParameterizedTestCase<Fixture, Templates0, Types> {
+ public:
+  static bool Register(const char* /*prefix*/, CodeLocation,
+                       const TypedTestCasePState* /*state*/,
+                       const char* /*case_name*/, const char* /*test_names*/) {
+    return true;
+  }
+};
+
+#endif  // GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+// Returns the current OS stack trace as an std::string.
+//
+// The maximum number of stack frames to be included is specified by
+// the gtest_stack_trace_depth flag.  The skip_count parameter
+// specifies the number of top frames to be skipped, which doesn't
+// count against the number of frames to be included.
+//
+// For example, if Foo() calls Bar(), which in turn calls
+// GetCurrentOsStackTraceExceptTop(..., 1), Foo() will be included in
+// the trace but Bar() and GetCurrentOsStackTraceExceptTop() won't.
+GTEST_API_ std::string GetCurrentOsStackTraceExceptTop(
+    UnitTest* unit_test, int skip_count);
+
+// Helpers for suppressing warnings on unreachable code or constant
+// condition.
+
+// Always returns true.
+GTEST_API_ bool AlwaysTrue();
+
+// Always returns false.
+inline bool AlwaysFalse() { return !AlwaysTrue(); }
+
+// Helper for suppressing false warning from Clang on a const char*
+// variable declared in a conditional expression always being NULL in
+// the else branch.
+struct GTEST_API_ ConstCharPtr {
+  ConstCharPtr(const char* str) : value(str) {}
+  operator bool() const { return true; }
+  const char* value;
+};
+
+// A simple Linear Congruential Generator for generating random
+// numbers with a uniform distribution.  Unlike rand() and srand(), it
+// doesn't use global state (and therefore can't interfere with user
+// code).  Unlike rand_r(), it's portable.  An LCG isn't very random,
+// but it's good enough for our purposes.
+class GTEST_API_ Random {
+ public:
+  static const UInt32 kMaxRange = 1u << 31;
+
+  explicit Random(UInt32 seed) : state_(seed) {}
+
+  void Reseed(UInt32 seed) { state_ = seed; }
+
+  // Generates a random number from [0, range).  Crashes if 'range' is
+  // 0 or greater than kMaxRange.
+  UInt32 Generate(UInt32 range);
+
+ private:
+  UInt32 state_;
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Random);
+};
+
+// Defining a variable of type CompileAssertTypesEqual<T1, T2> will cause a
+// compiler error iff T1 and T2 are different types.
+template <typename T1, typename T2>
+struct CompileAssertTypesEqual;
+
+template <typename T>
+struct CompileAssertTypesEqual<T, T> {
+};
+
+// Removes the reference from a type if it is a reference type,
+// otherwise leaves it unchanged.  This is the same as
+// tr1::remove_reference, which is not widely available yet.
+template <typename T>
+struct RemoveReference { typedef T type; };  // NOLINT
+template <typename T>
+struct RemoveReference<T&> { typedef T type; };  // NOLINT
+
+// A handy wrapper around RemoveReference that works when the argument
+// T depends on template parameters.
+#define GTEST_REMOVE_REFERENCE_(T) \
+    typename ::testing::internal::RemoveReference<T>::type
+
+// Removes const from a type if it is a const type, otherwise leaves
+// it unchanged.  This is the same as tr1::remove_const, which is not
+// widely available yet.
+template <typename T>
+struct RemoveConst { typedef T type; };  // NOLINT
+template <typename T>
+struct RemoveConst<const T> { typedef T type; };  // NOLINT
+
+// MSVC 8.0, Sun C++, and IBM XL C++ have a bug which causes the above
+// definition to fail to remove the const in 'const int[3]' and 'const
+// char[3][4]'.  The following specialization works around the bug.
+template <typename T, size_t N>
+struct RemoveConst<const T[N]> {
+  typedef typename RemoveConst<T>::type type[N];
+};
+
+#if defined(_MSC_VER) && _MSC_VER < 1400
+// This is the only specialization that allows VC++ 7.1 to remove const in
+// 'const int[3] and 'const int[3][4]'.  However, it causes trouble with GCC
+// and thus needs to be conditionally compiled.
+template <typename T, size_t N>
+struct RemoveConst<T[N]> {
+  typedef typename RemoveConst<T>::type type[N];
+};
+#endif
+
+// A handy wrapper around RemoveConst that works when the argument
+// T depends on template parameters.
+#define GTEST_REMOVE_CONST_(T) \
+    typename ::testing::internal::RemoveConst<T>::type
+
+// Turns const U&, U&, const U, and U all into U.
+#define GTEST_REMOVE_REFERENCE_AND_CONST_(T) \
+    GTEST_REMOVE_CONST_(GTEST_REMOVE_REFERENCE_(T))
+
+// Adds reference to a type if it is not a reference type,
+// otherwise leaves it unchanged.  This is the same as
+// tr1::add_reference, which is not widely available yet.
+template <typename T>
+struct AddReference { typedef T& type; };  // NOLINT
+template <typename T>
+struct AddReference<T&> { typedef T& type; };  // NOLINT
+
+// A handy wrapper around AddReference that works when the argument T
+// depends on template parameters.
+#define GTEST_ADD_REFERENCE_(T) \
+    typename ::testing::internal::AddReference<T>::type
+
+// Adds a reference to const on top of T as necessary.  For example,
+// it transforms
+//
+//   char         ==> const char&
+//   const char   ==> const char&
+//   char&        ==> const char&
+//   const char&  ==> const char&
+//
+// The argument T must depend on some template parameters.
+#define GTEST_REFERENCE_TO_CONST_(T) \
+    GTEST_ADD_REFERENCE_(const GTEST_REMOVE_REFERENCE_(T))
+
+// ImplicitlyConvertible<From, To>::value is a compile-time bool
+// constant that's true iff type From can be implicitly converted to
+// type To.
+template <typename From, typename To>
+class ImplicitlyConvertible {
+ private:
+  // We need the following helper functions only for their types.
+  // They have no implementations.
+
+  // MakeFrom() is an expression whose type is From.  We cannot simply
+  // use From(), as the type From may not have a public default
+  // constructor.
+  static typename AddReference<From>::type MakeFrom();
+
+  // These two functions are overloaded.  Given an expression
+  // Helper(x), the compiler will pick the first version if x can be
+  // implicitly converted to type To; otherwise it will pick the
+  // second version.
+  //
+  // The first version returns a value of size 1, and the second
+  // version returns a value of size 2.  Therefore, by checking the
+  // size of Helper(x), which can be done at compile time, we can tell
+  // which version of Helper() is used, and hence whether x can be
+  // implicitly converted to type To.
+  static char Helper(To);
+  static char (&Helper(...))[2];  // NOLINT
+
+  // We have to put the 'public' section after the 'private' section,
+  // or MSVC refuses to compile the code.
+ public:
+#if defined(__BORLANDC__)
+  // C++Builder cannot use member overload resolution during template
+  // instantiation.  The simplest workaround is to use its C++0x type traits
+  // functions (C++Builder 2009 and above only).
+  static const bool value = __is_convertible(From, To);
+#else
+  // MSVC warns about implicitly converting from double to int for
+  // possible loss of data, so we need to temporarily disable the
+  // warning.
+  GTEST_DISABLE_MSC_WARNINGS_PUSH_(4244)
+  static const bool value =
+      sizeof(Helper(ImplicitlyConvertible::MakeFrom())) == 1;
+  GTEST_DISABLE_MSC_WARNINGS_POP_()
+#endif  // __BORLANDC__
+};
+template <typename From, typename To>
+const bool ImplicitlyConvertible<From, To>::value;
+
+// IsAProtocolMessage<T>::value is a compile-time bool constant that's
+// true iff T is type ProtocolMessage, proto2::Message, or a subclass
+// of those.
+template <typename T>
+struct IsAProtocolMessage
+    : public bool_constant<
+  ImplicitlyConvertible<const T*, const ::ProtocolMessage*>::value ||
+  ImplicitlyConvertible<const T*, const ::proto2::Message*>::value> {
+};
+
+// When the compiler sees expression IsContainerTest<C>(0), if C is an
+// STL-style container class, the first overload of IsContainerTest
+// will be viable (since both C::iterator* and C::const_iterator* are
+// valid types and NULL can be implicitly converted to them).  It will
+// be picked over the second overload as 'int' is a perfect match for
+// the type of argument 0.  If C::iterator or C::const_iterator is not
+// a valid type, the first overload is not viable, and the second
+// overload will be picked.  Therefore, we can determine whether C is
+// a container class by checking the type of IsContainerTest<C>(0).
+// The value of the expression is insignificant.
+//
+// Note that we look for both C::iterator and C::const_iterator.  The
+// reason is that C++ injects the name of a class as a member of the
+// class itself (e.g. you can refer to class iterator as either
+// 'iterator' or 'iterator::iterator').  If we look for C::iterator
+// only, for example, we would mistakenly think that a class named
+// iterator is an STL container.
+//
+// Also note that the simpler approach of overloading
+// IsContainerTest(typename C::const_iterator*) and
+// IsContainerTest(...) doesn't work with Visual Age C++ and Sun C++.
+typedef int IsContainer;
+template <class C>
+IsContainer IsContainerTest(int /* dummy */,
+                            typename C::iterator* /* it */ = NULL,
+                            typename C::const_iterator* /* const_it */ = NULL) {
+  return 0;
+}
+
+typedef char IsNotContainer;
+template <class C>
+IsNotContainer IsContainerTest(long /* dummy */) { return '\0'; }
+
+// EnableIf<condition>::type is void when 'Cond' is true, and
+// undefined when 'Cond' is false.  To use SFINAE to make a function
+// overload only apply when a particular expression is true, add
+// "typename EnableIf<expression>::type* = 0" as the last parameter.
+template<bool> struct EnableIf;
+template<> struct EnableIf<true> { typedef void type; };  // NOLINT
+
+// Utilities for native arrays.
+
+// ArrayEq() compares two k-dimensional native arrays using the
+// elements' operator==, where k can be any integer >= 0.  When k is
+// 0, ArrayEq() degenerates into comparing a single pair of values.
+
+template <typename T, typename U>
+bool ArrayEq(const T* lhs, size_t size, const U* rhs);
+
+// This generic version is used when k is 0.
+template <typename T, typename U>
+inline bool ArrayEq(const T& lhs, const U& rhs) { return lhs == rhs; }
+
+// This overload is used when k >= 1.
+template <typename T, typename U, size_t N>
+inline bool ArrayEq(const T(&lhs)[N], const U(&rhs)[N]) {
+  return internal::ArrayEq(lhs, N, rhs);
+}
+
+// This helper reduces code bloat.  If we instead put its logic inside
+// the previous ArrayEq() function, arrays with different sizes would
+// lead to different copies of the template code.
+template <typename T, typename U>
+bool ArrayEq(const T* lhs, size_t size, const U* rhs) {
+  for (size_t i = 0; i != size; i++) {
+    if (!internal::ArrayEq(lhs[i], rhs[i]))
+      return false;
+  }
+  return true;
+}
+
+// Finds the first element in the iterator range [begin, end) that
+// equals elem.  Element may be a native array type itself.
+template <typename Iter, typename Element>
+Iter ArrayAwareFind(Iter begin, Iter end, const Element& elem) {
+  for (Iter it = begin; it != end; ++it) {
+    if (internal::ArrayEq(*it, elem))
+      return it;
+  }
+  return end;
+}
+
+// CopyArray() copies a k-dimensional native array using the elements'
+// operator=, where k can be any integer >= 0.  When k is 0,
+// CopyArray() degenerates into copying a single value.
+
+template <typename T, typename U>
+void CopyArray(const T* from, size_t size, U* to);
+
+// This generic version is used when k is 0.
+template <typename T, typename U>
+inline void CopyArray(const T& from, U* to) { *to = from; }
+
+// This overload is used when k >= 1.
+template <typename T, typename U, size_t N>
+inline void CopyArray(const T(&from)[N], U(*to)[N]) {
+  internal::CopyArray(from, N, *to);
+}
+
+// This helper reduces code bloat.  If we instead put its logic inside
+// the previous CopyArray() function, arrays with different sizes
+// would lead to different copies of the template code.
+template <typename T, typename U>
+void CopyArray(const T* from, size_t size, U* to) {
+  for (size_t i = 0; i != size; i++) {
+    internal::CopyArray(from[i], to + i);
+  }
+}
+
+// The relation between an NativeArray object (see below) and the
+// native array it represents.
+// We use 2 different structs to allow non-copyable types to be used, as long
+// as RelationToSourceReference() is passed.
+struct RelationToSourceReference {};
+struct RelationToSourceCopy {};
+
+// Adapts a native array to a read-only STL-style container.  Instead
+// of the complete STL container concept, this adaptor only implements
+// members useful for Google Mock's container matchers.  New members
+// should be added as needed.  To simplify the implementation, we only
+// support Element being a raw type (i.e. having no top-level const or
+// reference modifier).  It's the client's responsibility to satisfy
+// this requirement.  Element can be an array type itself (hence
+// multi-dimensional arrays are supported).
+template <typename Element>
+class NativeArray {
+ public:
+  // STL-style container typedefs.
+  typedef Element value_type;
+  typedef Element* iterator;
+  typedef const Element* const_iterator;
+
+  // Constructs from a native array. References the source.
+  NativeArray(const Element* array, size_t count, RelationToSourceReference) {
+    InitRef(array, count);
+  }
+
+  // Constructs from a native array. Copies the source.
+  NativeArray(const Element* array, size_t count, RelationToSourceCopy) {
+    InitCopy(array, count);
+  }
+
+  // Copy constructor.
+  NativeArray(const NativeArray& rhs) {
+    (this->*rhs.clone_)(rhs.array_, rhs.size_);
+  }
+
+  ~NativeArray() {
+    if (clone_ != &NativeArray::InitRef)
+      delete[] array_;
+  }
+
+  // STL-style container methods.
+  size_t size() const { return size_; }
+  const_iterator begin() const { return array_; }
+  const_iterator end() const { return array_ + size_; }
+  bool operator==(const NativeArray& rhs) const {
+    return size() == rhs.size() &&
+        ArrayEq(begin(), size(), rhs.begin());
+  }
+
+ private:
+  enum {
+    kCheckTypeIsNotConstOrAReference = StaticAssertTypeEqHelper<
+        Element, GTEST_REMOVE_REFERENCE_AND_CONST_(Element)>::value,
+  };
+
+  // Initializes this object with a copy of the input.
+  void InitCopy(const Element* array, size_t a_size) {
+    Element* const copy = new Element[a_size];
+    CopyArray(array, a_size, copy);
+    array_ = copy;
+    size_ = a_size;
+    clone_ = &NativeArray::InitCopy;
+  }
+
+  // Initializes this object with a reference of the input.
+  void InitRef(const Element* array, size_t a_size) {
+    array_ = array;
+    size_ = a_size;
+    clone_ = &NativeArray::InitRef;
+  }
+
+  const Element* array_;
+  size_t size_;
+  void (NativeArray::*clone_)(const Element*, size_t);
+
+  GTEST_DISALLOW_ASSIGN_(NativeArray);
+};
+
+}  // namespace internal
+}  // namespace testing
+
+#define GTEST_MESSAGE_AT_(file, line, message, result_type) \
+  ::testing::internal::AssertHelper(result_type, file, line, message) \
+    = ::testing::Message()
+
+#define GTEST_MESSAGE_(message, result_type) \
+  GTEST_MESSAGE_AT_(__FILE__, __LINE__, message, result_type)
+
+#define GTEST_FATAL_FAILURE_(message) \
+  return GTEST_MESSAGE_(message, ::testing::TestPartResult::kFatalFailure)
+
+#define GTEST_NONFATAL_FAILURE_(message) \
+  GTEST_MESSAGE_(message, ::testing::TestPartResult::kNonFatalFailure)
+
+#define GTEST_SUCCESS_(message) \
+  GTEST_MESSAGE_(message, ::testing::TestPartResult::kSuccess)
+
+// Suppresses MSVC warnings 4072 (unreachable code) for the code following
+// statement if it returns or throws (or doesn't return or throw in some
+// situations).
+#define GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement) \
+  if (::testing::internal::AlwaysTrue()) { statement; }
+
+#define GTEST_TEST_THROW_(statement, expected_exception, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::ConstCharPtr gtest_msg = "") { \
+    bool gtest_caught_expected = false; \
+    try { \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    } \
+    catch (expected_exception const&) { \
+      gtest_caught_expected = true; \
+    } \
+    catch (...) { \
+      gtest_msg.value = \
+          "Expected: " #statement " throws an exception of type " \
+          #expected_exception ".\n  Actual: it throws a different type."; \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testthrow_, __LINE__); \
+    } \
+    if (!gtest_caught_expected) { \
+      gtest_msg.value = \
+          "Expected: " #statement " throws an exception of type " \
+          #expected_exception ".\n  Actual: it throws nothing."; \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testthrow_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testthrow_, __LINE__): \
+      fail(gtest_msg.value)
+
+#define GTEST_TEST_NO_THROW_(statement, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    try { \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    } \
+    catch (...) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testnothrow_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testnothrow_, __LINE__): \
+      fail("Expected: " #statement " doesn't throw an exception.\n" \
+           "  Actual: it throws.")
+
+#define GTEST_TEST_ANY_THROW_(statement, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    bool gtest_caught_any = false; \
+    try { \
+      GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    } \
+    catch (...) { \
+      gtest_caught_any = true; \
+    } \
+    if (!gtest_caught_any) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testanythrow_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testanythrow_, __LINE__): \
+      fail("Expected: " #statement " throws an exception.\n" \
+           "  Actual: it doesn't.")
+
+
+// Implements Boolean test assertions such as EXPECT_TRUE. expression can be
+// either a boolean expression or an AssertionResult. text is a textual
+// represenation of expression as it was passed into the EXPECT_TRUE.
+#define GTEST_TEST_BOOLEAN_(expression, text, actual, expected, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (const ::testing::AssertionResult gtest_ar_ = \
+      ::testing::AssertionResult(expression)) \
+    ; \
+  else \
+    fail(::testing::internal::GetBoolAssertionFailureMessage(\
+        gtest_ar_, text, #actual, #expected).c_str())
+
+#define GTEST_TEST_NO_FATAL_FAILURE_(statement, fail) \
+  GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+  if (::testing::internal::AlwaysTrue()) { \
+    ::testing::internal::HasNewFatalFailureHelper gtest_fatal_failure_checker; \
+    GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_(statement); \
+    if (gtest_fatal_failure_checker.has_new_fatal_failure()) { \
+      goto GTEST_CONCAT_TOKEN_(gtest_label_testnofatal_, __LINE__); \
+    } \
+  } else \
+    GTEST_CONCAT_TOKEN_(gtest_label_testnofatal_, __LINE__): \
+      fail("Expected: " #statement " doesn't generate new fatal " \
+           "failures in the current thread.\n" \
+           "  Actual: it does.")
+
+// Expands to the name of the class that implements the given test.
+#define GTEST_TEST_CLASS_NAME_(test_case_name, test_name) \
+  test_case_name##_##test_name##_Test
+
+// Helper macro for defining tests.
+#define GTEST_TEST_(test_case_name, test_name, parent_class, parent_id)\
+class GTEST_TEST_CLASS_NAME_(test_case_name, test_name) : public parent_class {\
+ public:\
+  GTEST_TEST_CLASS_NAME_(test_case_name, test_name)() {}\
+ private:\
+  virtual void TestBody();\
+  static ::testing::TestInfo* const test_info_ GTEST_ATTRIBUTE_UNUSED_;\
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(\
+      GTEST_TEST_CLASS_NAME_(test_case_name, test_name));\
+};\
+\
+::testing::TestInfo* const GTEST_TEST_CLASS_NAME_(test_case_name, test_name)\
+  ::test_info_ =\
+    ::testing::internal::MakeAndRegisterTestInfo(\
+        #test_case_name, #test_name, NULL, NULL, \
+        ::testing::internal::CodeLocation(__FILE__, __LINE__), \
+        (parent_id), \
+        parent_class::SetUpTestCase, \
+        parent_class::TearDownTestCase, \
+        new ::testing::internal::TestFactoryImpl<\
+            GTEST_TEST_CLASS_NAME_(test_case_name, test_name)>);\
+void GTEST_TEST_CLASS_NAME_(test_case_name, test_name)::TestBody()
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_INTERNAL_H_
+
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-linked_ptr.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-linked_ptr.h
new file mode 100644
index 0000000000..3602942217
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-linked_ptr.h
@@ -0,0 +1,243 @@
+// Copyright 2003 Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: Dan Egnor (egnor@google.com)
+//
+// A "smart" pointer type with reference tracking.  Every pointer to a
+// particular object is kept on a circular linked list.  When the last pointer
+// to an object is destroyed or reassigned, the object is deleted.
+//
+// Used properly, this deletes the object when the last reference goes away.
+// There are several caveats:
+// - Like all reference counting schemes, cycles lead to leaks.
+// - Each smart pointer is actually two pointers (8 bytes instead of 4).
+// - Every time a pointer is assigned, the entire list of pointers to that
+//   object is traversed.  This class is therefore NOT SUITABLE when there
+//   will often be more than two or three pointers to a particular object.
+// - References are only tracked as long as linked_ptr<> objects are copied.
+//   If a linked_ptr<> is converted to a raw pointer and back, BAD THINGS
+//   will happen (double deletion).
+//
+// A good use of this class is storing object references in STL containers.
+// You can safely put linked_ptr<> in a vector<>.
+// Other uses may not be as good.
+//
+// Note: If you use an incomplete type with linked_ptr<>, the class
+// *containing* linked_ptr<> must have a constructor and destructor (even
+// if they do nothing!).
+//
+// Bill Gibbons suggested we use something like this.
+//
+// Thread Safety:
+//   Unlike other linked_ptr implementations, in this implementation
+//   a linked_ptr object is thread-safe in the sense that:
+//     - it's safe to copy linked_ptr objects concurrently,
+//     - it's safe to copy *from* a linked_ptr and read its underlying
+//       raw pointer (e.g. via get()) concurrently, and
+//     - it's safe to write to two linked_ptrs that point to the same
+//       shared object concurrently.
+// TODO(wan@google.com): rename this to safe_linked_ptr to avoid
+// confusion with normal linked_ptr.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_LINKED_PTR_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_LINKED_PTR_H_
+
+#include <stdlib.h>
+#include <assert.h>
+
+#include "gtest/internal/gtest-port.h"
+
+namespace testing {
+namespace internal {
+
+// Protects copying of all linked_ptr objects.
+GTEST_API_ GTEST_DECLARE_STATIC_MUTEX_(g_linked_ptr_mutex);
+
+// This is used internally by all instances of linked_ptr<>.  It needs to be
+// a non-template class because different types of linked_ptr<> can refer to
+// the same object (linked_ptr<Superclass>(obj) vs linked_ptr<Subclass>(obj)).
+// So, it needs to be possible for different types of linked_ptr to participate
+// in the same circular linked list, so we need a single class type here.
+//
+// DO NOT USE THIS CLASS DIRECTLY YOURSELF.  Use linked_ptr<T>.
+class linked_ptr_internal {
+ public:
+  // Create a new circle that includes only this instance.
+  void join_new() {
+    next_ = this;
+  }
+
+  // Many linked_ptr operations may change p.link_ for some linked_ptr
+  // variable p in the same circle as this object.  Therefore we need
+  // to prevent two such operations from occurring concurrently.
+  //
+  // Note that different types of linked_ptr objects can coexist in a
+  // circle (e.g. linked_ptr<Base>, linked_ptr<Derived1>, and
+  // linked_ptr<Derived2>).  Therefore we must use a single mutex to
+  // protect all linked_ptr objects.  This can create serious
+  // contention in production code, but is acceptable in a testing
+  // framework.
+
+  // Join an existing circle.
+  void join(linked_ptr_internal const* ptr)
+      GTEST_LOCK_EXCLUDED_(g_linked_ptr_mutex) {
+    MutexLock lock(&g_linked_ptr_mutex);
+
+    linked_ptr_internal const* p = ptr;
+    while (p->next_ != ptr) {
+      assert(p->next_ != this &&
+             "Trying to join() a linked ring we are already in. "
+             "Is GMock thread safety enabled?");
+      p = p->next_;
+    }
+    p->next_ = this;
+    next_ = ptr;
+  }
+
+  // Leave whatever circle we're part of.  Returns true if we were the
+  // last member of the circle.  Once this is done, you can join() another.
+  bool depart()
+      GTEST_LOCK_EXCLUDED_(g_linked_ptr_mutex) {
+    MutexLock lock(&g_linked_ptr_mutex);
+
+    if (next_ == this) return true;
+    linked_ptr_internal const* p = next_;
+    while (p->next_ != this) {
+      assert(p->next_ != next_ &&
+             "Trying to depart() a linked ring we are not in. "
+             "Is GMock thread safety enabled?");
+      p = p->next_;
+    }
+    p->next_ = next_;
+    return false;
+  }
+
+ private:
+  mutable linked_ptr_internal const* next_;
+};
+
+template <typename T>
+class linked_ptr {
+ public:
+  typedef T element_type;
+
+  // Take over ownership of a raw pointer.  This should happen as soon as
+  // possible after the object is created.
+  explicit linked_ptr(T* ptr = NULL) { capture(ptr); }
+  ~linked_ptr() { depart(); }
+
+  // Copy an existing linked_ptr<>, adding ourselves to the list of references.
+  template <typename U> linked_ptr(linked_ptr<U> const& ptr) { copy(&ptr); }
+  linked_ptr(linked_ptr const& ptr) {  // NOLINT
+    assert(&ptr != this);
+    copy(&ptr);
+  }
+
+  // Assignment releases the old value and acquires the new.
+  template <typename U> linked_ptr& operator=(linked_ptr<U> const& ptr) {
+    depart();
+    copy(&ptr);
+    return *this;
+  }
+
+  linked_ptr& operator=(linked_ptr const& ptr) {
+    if (&ptr != this) {
+      depart();
+      copy(&ptr);
+    }
+    return *this;
+  }
+
+  // Smart pointer members.
+  void reset(T* ptr = NULL) {
+    depart();
+    capture(ptr);
+  }
+  T* get() const { return value_; }
+  T* operator->() const { return value_; }
+  T& operator*() const { return *value_; }
+
+  bool operator==(T* p) const { return value_ == p; }
+  bool operator!=(T* p) const { return value_ != p; }
+  template <typename U>
+  bool operator==(linked_ptr<U> const& ptr) const {
+    return value_ == ptr.get();
+  }
+  template <typename U>
+  bool operator!=(linked_ptr<U> const& ptr) const {
+    return value_ != ptr.get();
+  }
+
+ private:
+  template <typename U>
+  friend class linked_ptr;
+
+  T* value_;
+  linked_ptr_internal link_;
+
+  void depart() {
+    if (link_.depart()) delete value_;
+  }
+
+  void capture(T* ptr) {
+    value_ = ptr;
+    link_.join_new();
+  }
+
+  template <typename U> void copy(linked_ptr<U> const* ptr) {
+    value_ = ptr->get();
+    if (value_)
+      link_.join(&ptr->link_);
+    else
+      link_.join_new();
+  }
+};
+
+template<typename T> inline
+bool operator==(T* ptr, const linked_ptr<T>& x) {
+  return ptr == x.get();
+}
+
+template<typename T> inline
+bool operator!=(T* ptr, const linked_ptr<T>& x) {
+  return ptr != x.get();
+}
+
+// A function to convert T* into linked_ptr<T>
+// Doing e.g. make_linked_ptr(new FooBarBaz<type>(arg)) is a shorter notation
+// for linked_ptr<FooBarBaz<type> >(new FooBarBaz<type>(arg))
+template <typename T>
+linked_ptr<T> make_linked_ptr(T* ptr) {
+  return linked_ptr<T>(ptr);
+}
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_LINKED_PTR_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-param-util-generated.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-param-util-generated.h
new file mode 100644
index 0000000000..4d1d81d20f
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-param-util-generated.h
@@ -0,0 +1,5146 @@
+// This file was GENERATED by command:
+//     pump.py gtest-param-util-generated.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: vladl@google.com (Vlad Losev)
+
+// Type and function utilities for implementing parameterized tests.
+// This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+// Currently Google Test supports at most 50 arguments in Values,
+// and at most 10 arguments in Combine. Please contact
+// googletestframework@googlegroups.com if you need more.
+// Please note that the number of arguments to Combine is limited
+// by the maximum arity of the implementation of tuple which is
+// currently set at 10.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+#include "gtest/internal/gtest-param-util.h"
+#include "gtest/internal/gtest-port.h"
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+
+// Forward declarations of ValuesIn(), which is implemented in
+// include/gtest/gtest-param-test.h.
+template <typename ForwardIterator>
+internal::ParamGenerator<
+  typename ::testing::internal::IteratorTraits<ForwardIterator>::value_type>
+ValuesIn(ForwardIterator begin, ForwardIterator end);
+
+template <typename T, size_t N>
+internal::ParamGenerator<T> ValuesIn(const T (&array)[N]);
+
+template <class Container>
+internal::ParamGenerator<typename Container::value_type> ValuesIn(
+    const Container& container);
+
+namespace internal {
+
+// Used in the Values() function to provide polymorphic capabilities.
+template <typename T1>
+class ValueArray1 {
+ public:
+  explicit ValueArray1(T1 v1) : v1_(v1) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray1& other);
+
+  const T1 v1_;
+};
+
+template <typename T1, typename T2>
+class ValueArray2 {
+ public:
+  ValueArray2(T1 v1, T2 v2) : v1_(v1), v2_(v2) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray2& other);
+
+  const T1 v1_;
+  const T2 v2_;
+};
+
+template <typename T1, typename T2, typename T3>
+class ValueArray3 {
+ public:
+  ValueArray3(T1 v1, T2 v2, T3 v3) : v1_(v1), v2_(v2), v3_(v3) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray3& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4>
+class ValueArray4 {
+ public:
+  ValueArray4(T1 v1, T2 v2, T3 v3, T4 v4) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray4& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+class ValueArray5 {
+ public:
+  ValueArray5(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray5& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6>
+class ValueArray6 {
+ public:
+  ValueArray6(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray6& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7>
+class ValueArray7 {
+ public:
+  ValueArray7(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray7& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8>
+class ValueArray8 {
+ public:
+  ValueArray8(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7,
+      T8 v8) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray8& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9>
+class ValueArray9 {
+ public:
+  ValueArray9(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8,
+      T9 v9) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray9& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10>
+class ValueArray10 {
+ public:
+  ValueArray10(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray10& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11>
+class ValueArray11 {
+ public:
+  ValueArray11(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray11& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12>
+class ValueArray12 {
+ public:
+  ValueArray12(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray12& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13>
+class ValueArray13 {
+ public:
+  ValueArray13(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray13& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14>
+class ValueArray14 {
+ public:
+  ValueArray14(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray14& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15>
+class ValueArray15 {
+ public:
+  ValueArray15(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray15& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16>
+class ValueArray16 {
+ public:
+  ValueArray16(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray16& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17>
+class ValueArray17 {
+ public:
+  ValueArray17(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16,
+      T17 v17) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray17& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18>
+class ValueArray18 {
+ public:
+  ValueArray18(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray18& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19>
+class ValueArray19 {
+ public:
+  ValueArray19(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray19& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20>
+class ValueArray20 {
+ public:
+  ValueArray20(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray20& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21>
+class ValueArray21 {
+ public:
+  ValueArray21(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray21& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22>
+class ValueArray22 {
+ public:
+  ValueArray22(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray22& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23>
+class ValueArray23 {
+ public:
+  ValueArray23(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray23& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24>
+class ValueArray24 {
+ public:
+  ValueArray24(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray24& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25>
+class ValueArray25 {
+ public:
+  ValueArray25(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24,
+      T25 v25) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray25& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26>
+class ValueArray26 {
+ public:
+  ValueArray26(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray26& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27>
+class ValueArray27 {
+ public:
+  ValueArray27(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19),
+      v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25),
+      v26_(v26), v27_(v27) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray27& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28>
+class ValueArray28 {
+ public:
+  ValueArray28(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24),
+      v25_(v25), v26_(v26), v27_(v27), v28_(v28) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray28& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29>
+class ValueArray29 {
+ public:
+  ValueArray29(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23),
+      v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray29& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30>
+class ValueArray30 {
+ public:
+  ValueArray30(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray30& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31>
+class ValueArray31 {
+ public:
+  ValueArray31(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray31& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32>
+class ValueArray32 {
+ public:
+  ValueArray32(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27),
+      v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray32& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33>
+class ValueArray33 {
+ public:
+  ValueArray33(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32,
+      T33 v33) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray33& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34>
+class ValueArray34 {
+ public:
+  ValueArray34(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray34& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35>
+class ValueArray35 {
+ public:
+  ValueArray35(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19),
+      v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25),
+      v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31),
+      v32_(v32), v33_(v33), v34_(v34), v35_(v35) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray35& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36>
+class ValueArray36 {
+ public:
+  ValueArray36(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24),
+      v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30),
+      v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35), v36_(v36) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray36& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37>
+class ValueArray37 {
+ public:
+  ValueArray37(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23),
+      v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29),
+      v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35),
+      v36_(v36), v37_(v37) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray37& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38>
+class ValueArray38 {
+ public:
+  ValueArray38(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray38& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39>
+class ValueArray39 {
+ public:
+  ValueArray39(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray39& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40>
+class ValueArray40 {
+ public:
+  ValueArray40(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27),
+      v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33),
+      v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39),
+      v40_(v40) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray40& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41>
+class ValueArray41 {
+ public:
+  ValueArray41(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40,
+      T41 v41) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray41& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42>
+class ValueArray42 {
+ public:
+  ValueArray42(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41), v42_(v42) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray42& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43>
+class ValueArray43 {
+ public:
+  ValueArray43(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6),
+      v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13),
+      v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19),
+      v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25),
+      v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31),
+      v32_(v32), v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37),
+      v38_(v38), v39_(v39), v40_(v40), v41_(v41), v42_(v42), v43_(v43) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray43& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44>
+class ValueArray44 {
+ public:
+  ValueArray44(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5),
+      v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12),
+      v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17), v18_(v18),
+      v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23), v24_(v24),
+      v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29), v30_(v30),
+      v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35), v36_(v36),
+      v37_(v37), v38_(v38), v39_(v39), v40_(v40), v41_(v41), v42_(v42),
+      v43_(v43), v44_(v44) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray44& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45>
+class ValueArray45 {
+ public:
+  ValueArray45(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45) : v1_(v1), v2_(v2), v3_(v3), v4_(v4),
+      v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10), v11_(v11),
+      v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16), v17_(v17),
+      v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22), v23_(v23),
+      v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28), v29_(v29),
+      v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34), v35_(v35),
+      v36_(v36), v37_(v37), v38_(v38), v39_(v39), v40_(v40), v41_(v41),
+      v42_(v42), v43_(v43), v44_(v44), v45_(v45) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray45& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46>
+class ValueArray46 {
+ public:
+  ValueArray46(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46) : v1_(v1), v2_(v2), v3_(v3),
+      v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39), v40_(v40),
+      v41_(v41), v42_(v42), v43_(v43), v44_(v44), v45_(v45), v46_(v46) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_), static_cast<T>(v46_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray46& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47>
+class ValueArray47 {
+ public:
+  ValueArray47(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47) : v1_(v1), v2_(v2),
+      v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9), v10_(v10),
+      v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15), v16_(v16),
+      v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21), v22_(v22),
+      v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27), v28_(v28),
+      v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33), v34_(v34),
+      v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39), v40_(v40),
+      v41_(v41), v42_(v42), v43_(v43), v44_(v44), v45_(v45), v46_(v46),
+      v47_(v47) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_), static_cast<T>(v46_), static_cast<T>(v47_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray47& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48>
+class ValueArray48 {
+ public:
+  ValueArray48(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47, T48 v48) : v1_(v1),
+      v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7), v8_(v8), v9_(v9),
+      v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14), v15_(v15),
+      v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20), v21_(v21),
+      v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26), v27_(v27),
+      v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32), v33_(v33),
+      v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38), v39_(v39),
+      v40_(v40), v41_(v41), v42_(v42), v43_(v43), v44_(v44), v45_(v45),
+      v46_(v46), v47_(v47), v48_(v48) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_), static_cast<T>(v46_), static_cast<T>(v47_),
+        static_cast<T>(v48_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray48& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+  const T48 v48_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49>
+class ValueArray49 {
+ public:
+  ValueArray49(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47, T48 v48,
+      T49 v49) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41), v42_(v42), v43_(v43), v44_(v44),
+      v45_(v45), v46_(v46), v47_(v47), v48_(v48), v49_(v49) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_), static_cast<T>(v46_), static_cast<T>(v47_),
+        static_cast<T>(v48_), static_cast<T>(v49_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray49& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+  const T48 v48_;
+  const T49 v49_;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49, typename T50>
+class ValueArray50 {
+ public:
+  ValueArray50(T1 v1, T2 v2, T3 v3, T4 v4, T5 v5, T6 v6, T7 v7, T8 v8, T9 v9,
+      T10 v10, T11 v11, T12 v12, T13 v13, T14 v14, T15 v15, T16 v16, T17 v17,
+      T18 v18, T19 v19, T20 v20, T21 v21, T22 v22, T23 v23, T24 v24, T25 v25,
+      T26 v26, T27 v27, T28 v28, T29 v29, T30 v30, T31 v31, T32 v32, T33 v33,
+      T34 v34, T35 v35, T36 v36, T37 v37, T38 v38, T39 v39, T40 v40, T41 v41,
+      T42 v42, T43 v43, T44 v44, T45 v45, T46 v46, T47 v47, T48 v48, T49 v49,
+      T50 v50) : v1_(v1), v2_(v2), v3_(v3), v4_(v4), v5_(v5), v6_(v6), v7_(v7),
+      v8_(v8), v9_(v9), v10_(v10), v11_(v11), v12_(v12), v13_(v13), v14_(v14),
+      v15_(v15), v16_(v16), v17_(v17), v18_(v18), v19_(v19), v20_(v20),
+      v21_(v21), v22_(v22), v23_(v23), v24_(v24), v25_(v25), v26_(v26),
+      v27_(v27), v28_(v28), v29_(v29), v30_(v30), v31_(v31), v32_(v32),
+      v33_(v33), v34_(v34), v35_(v35), v36_(v36), v37_(v37), v38_(v38),
+      v39_(v39), v40_(v40), v41_(v41), v42_(v42), v43_(v43), v44_(v44),
+      v45_(v45), v46_(v46), v47_(v47), v48_(v48), v49_(v49), v50_(v50) {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {static_cast<T>(v1_), static_cast<T>(v2_),
+        static_cast<T>(v3_), static_cast<T>(v4_), static_cast<T>(v5_),
+        static_cast<T>(v6_), static_cast<T>(v7_), static_cast<T>(v8_),
+        static_cast<T>(v9_), static_cast<T>(v10_), static_cast<T>(v11_),
+        static_cast<T>(v12_), static_cast<T>(v13_), static_cast<T>(v14_),
+        static_cast<T>(v15_), static_cast<T>(v16_), static_cast<T>(v17_),
+        static_cast<T>(v18_), static_cast<T>(v19_), static_cast<T>(v20_),
+        static_cast<T>(v21_), static_cast<T>(v22_), static_cast<T>(v23_),
+        static_cast<T>(v24_), static_cast<T>(v25_), static_cast<T>(v26_),
+        static_cast<T>(v27_), static_cast<T>(v28_), static_cast<T>(v29_),
+        static_cast<T>(v30_), static_cast<T>(v31_), static_cast<T>(v32_),
+        static_cast<T>(v33_), static_cast<T>(v34_), static_cast<T>(v35_),
+        static_cast<T>(v36_), static_cast<T>(v37_), static_cast<T>(v38_),
+        static_cast<T>(v39_), static_cast<T>(v40_), static_cast<T>(v41_),
+        static_cast<T>(v42_), static_cast<T>(v43_), static_cast<T>(v44_),
+        static_cast<T>(v45_), static_cast<T>(v46_), static_cast<T>(v47_),
+        static_cast<T>(v48_), static_cast<T>(v49_), static_cast<T>(v50_)};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray50& other);
+
+  const T1 v1_;
+  const T2 v2_;
+  const T3 v3_;
+  const T4 v4_;
+  const T5 v5_;
+  const T6 v6_;
+  const T7 v7_;
+  const T8 v8_;
+  const T9 v9_;
+  const T10 v10_;
+  const T11 v11_;
+  const T12 v12_;
+  const T13 v13_;
+  const T14 v14_;
+  const T15 v15_;
+  const T16 v16_;
+  const T17 v17_;
+  const T18 v18_;
+  const T19 v19_;
+  const T20 v20_;
+  const T21 v21_;
+  const T22 v22_;
+  const T23 v23_;
+  const T24 v24_;
+  const T25 v25_;
+  const T26 v26_;
+  const T27 v27_;
+  const T28 v28_;
+  const T29 v29_;
+  const T30 v30_;
+  const T31 v31_;
+  const T32 v32_;
+  const T33 v33_;
+  const T34 v34_;
+  const T35 v35_;
+  const T36 v36_;
+  const T37 v37_;
+  const T38 v38_;
+  const T39 v39_;
+  const T40 v40_;
+  const T41 v41_;
+  const T42 v42_;
+  const T43 v43_;
+  const T44 v44_;
+  const T45 v45_;
+  const T46 v46_;
+  const T47 v47_;
+  const T48 v48_;
+  const T49 v49_;
+  const T50 v50_;
+};
+
+# if GTEST_HAS_COMBINE
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Generates values from the Cartesian product of values produced
+// by the argument generators.
+//
+template <typename T1, typename T2>
+class CartesianProductGenerator2
+    : public ParamGeneratorInterface< ::testing::tuple<T1, T2> > {
+ public:
+  typedef ::testing::tuple<T1, T2> ParamType;
+
+  CartesianProductGenerator2(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2)
+      : g1_(g1), g2_(g2) {}
+  virtual ~CartesianProductGenerator2() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current2_;
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator2::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator2& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+};  // class CartesianProductGenerator2
+
+
+template <typename T1, typename T2, typename T3>
+class CartesianProductGenerator3
+    : public ParamGeneratorInterface< ::testing::tuple<T1, T2, T3> > {
+ public:
+  typedef ::testing::tuple<T1, T2, T3> ParamType;
+
+  CartesianProductGenerator3(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3)
+      : g1_(g1), g2_(g2), g3_(g3) {}
+  virtual ~CartesianProductGenerator3() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current3_;
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator3::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator3& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+};  // class CartesianProductGenerator3
+
+
+template <typename T1, typename T2, typename T3, typename T4>
+class CartesianProductGenerator4
+    : public ParamGeneratorInterface< ::testing::tuple<T1, T2, T3, T4> > {
+ public:
+  typedef ::testing::tuple<T1, T2, T3, T4> ParamType;
+
+  CartesianProductGenerator4(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4) {}
+  virtual ~CartesianProductGenerator4() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current4_;
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator4::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator4& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+};  // class CartesianProductGenerator4
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+class CartesianProductGenerator5
+    : public ParamGeneratorInterface< ::testing::tuple<T1, T2, T3, T4, T5> > {
+ public:
+  typedef ::testing::tuple<T1, T2, T3, T4, T5> ParamType;
+
+  CartesianProductGenerator5(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5) {}
+  virtual ~CartesianProductGenerator5() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current5_;
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator5::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator5& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+};  // class CartesianProductGenerator5
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6>
+class CartesianProductGenerator6
+    : public ParamGeneratorInterface< ::testing::tuple<T1, T2, T3, T4, T5,
+        T6> > {
+ public:
+  typedef ::testing::tuple<T1, T2, T3, T4, T5, T6> ParamType;
+
+  CartesianProductGenerator6(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5,
+      const ParamGenerator<T6>& g6)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6) {}
+  virtual ~CartesianProductGenerator6() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5,
+      const ParamGenerator<T6>& g6,
+      const typename ParamGenerator<T6>::iterator& current6)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current6_;
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    const typename ParamGenerator<T6>::iterator begin6_;
+    const typename ParamGenerator<T6>::iterator end6_;
+    typename ParamGenerator<T6>::iterator current6_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator6::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator6& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+  const ParamGenerator<T6> g6_;
+};  // class CartesianProductGenerator6
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7>
+class CartesianProductGenerator7
+    : public ParamGeneratorInterface< ::testing::tuple<T1, T2, T3, T4, T5, T6,
+        T7> > {
+ public:
+  typedef ::testing::tuple<T1, T2, T3, T4, T5, T6, T7> ParamType;
+
+  CartesianProductGenerator7(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5,
+      const ParamGenerator<T6>& g6, const ParamGenerator<T7>& g7)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7) {}
+  virtual ~CartesianProductGenerator7() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5,
+      const ParamGenerator<T6>& g6,
+      const typename ParamGenerator<T6>::iterator& current6,
+      const ParamGenerator<T7>& g7,
+      const typename ParamGenerator<T7>::iterator& current7)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current7_;
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    const typename ParamGenerator<T6>::iterator begin6_;
+    const typename ParamGenerator<T6>::iterator end6_;
+    typename ParamGenerator<T6>::iterator current6_;
+    const typename ParamGenerator<T7>::iterator begin7_;
+    const typename ParamGenerator<T7>::iterator end7_;
+    typename ParamGenerator<T7>::iterator current7_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator7::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator7& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+  const ParamGenerator<T6> g6_;
+  const ParamGenerator<T7> g7_;
+};  // class CartesianProductGenerator7
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8>
+class CartesianProductGenerator8
+    : public ParamGeneratorInterface< ::testing::tuple<T1, T2, T3, T4, T5, T6,
+        T7, T8> > {
+ public:
+  typedef ::testing::tuple<T1, T2, T3, T4, T5, T6, T7, T8> ParamType;
+
+  CartesianProductGenerator8(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5,
+      const ParamGenerator<T6>& g6, const ParamGenerator<T7>& g7,
+      const ParamGenerator<T8>& g8)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7),
+          g8_(g8) {}
+  virtual ~CartesianProductGenerator8() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin(), g8_, g8_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end(), g8_,
+        g8_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5,
+      const ParamGenerator<T6>& g6,
+      const typename ParamGenerator<T6>::iterator& current6,
+      const ParamGenerator<T7>& g7,
+      const typename ParamGenerator<T7>::iterator& current7,
+      const ParamGenerator<T8>& g8,
+      const typename ParamGenerator<T8>::iterator& current8)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7),
+          begin8_(g8.begin()), end8_(g8.end()), current8_(current8)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current8_;
+      if (current8_ == end8_) {
+        current8_ = begin8_;
+        ++current7_;
+      }
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_ &&
+          current8_ == typed_other->current8_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_),
+        begin8_(other.begin8_),
+        end8_(other.end8_),
+        current8_(other.current8_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_, *current8_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_ ||
+          current8_ == end8_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    const typename ParamGenerator<T6>::iterator begin6_;
+    const typename ParamGenerator<T6>::iterator end6_;
+    typename ParamGenerator<T6>::iterator current6_;
+    const typename ParamGenerator<T7>::iterator begin7_;
+    const typename ParamGenerator<T7>::iterator end7_;
+    typename ParamGenerator<T7>::iterator current7_;
+    const typename ParamGenerator<T8>::iterator begin8_;
+    const typename ParamGenerator<T8>::iterator end8_;
+    typename ParamGenerator<T8>::iterator current8_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator8::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator8& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+  const ParamGenerator<T6> g6_;
+  const ParamGenerator<T7> g7_;
+  const ParamGenerator<T8> g8_;
+};  // class CartesianProductGenerator8
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9>
+class CartesianProductGenerator9
+    : public ParamGeneratorInterface< ::testing::tuple<T1, T2, T3, T4, T5, T6,
+        T7, T8, T9> > {
+ public:
+  typedef ::testing::tuple<T1, T2, T3, T4, T5, T6, T7, T8, T9> ParamType;
+
+  CartesianProductGenerator9(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5,
+      const ParamGenerator<T6>& g6, const ParamGenerator<T7>& g7,
+      const ParamGenerator<T8>& g8, const ParamGenerator<T9>& g9)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9) {}
+  virtual ~CartesianProductGenerator9() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin(), g8_, g8_.begin(), g9_, g9_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end(), g8_,
+        g8_.end(), g9_, g9_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5,
+      const ParamGenerator<T6>& g6,
+      const typename ParamGenerator<T6>::iterator& current6,
+      const ParamGenerator<T7>& g7,
+      const typename ParamGenerator<T7>::iterator& current7,
+      const ParamGenerator<T8>& g8,
+      const typename ParamGenerator<T8>::iterator& current8,
+      const ParamGenerator<T9>& g9,
+      const typename ParamGenerator<T9>::iterator& current9)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7),
+          begin8_(g8.begin()), end8_(g8.end()), current8_(current8),
+          begin9_(g9.begin()), end9_(g9.end()), current9_(current9)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current9_;
+      if (current9_ == end9_) {
+        current9_ = begin9_;
+        ++current8_;
+      }
+      if (current8_ == end8_) {
+        current8_ = begin8_;
+        ++current7_;
+      }
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_ &&
+          current8_ == typed_other->current8_ &&
+          current9_ == typed_other->current9_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_),
+        begin8_(other.begin8_),
+        end8_(other.end8_),
+        current8_(other.current8_),
+        begin9_(other.begin9_),
+        end9_(other.end9_),
+        current9_(other.current9_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_, *current8_,
+            *current9_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_ ||
+          current8_ == end8_ ||
+          current9_ == end9_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    const typename ParamGenerator<T6>::iterator begin6_;
+    const typename ParamGenerator<T6>::iterator end6_;
+    typename ParamGenerator<T6>::iterator current6_;
+    const typename ParamGenerator<T7>::iterator begin7_;
+    const typename ParamGenerator<T7>::iterator end7_;
+    typename ParamGenerator<T7>::iterator current7_;
+    const typename ParamGenerator<T8>::iterator begin8_;
+    const typename ParamGenerator<T8>::iterator end8_;
+    typename ParamGenerator<T8>::iterator current8_;
+    const typename ParamGenerator<T9>::iterator begin9_;
+    const typename ParamGenerator<T9>::iterator end9_;
+    typename ParamGenerator<T9>::iterator current9_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator9::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator9& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+  const ParamGenerator<T6> g6_;
+  const ParamGenerator<T7> g7_;
+  const ParamGenerator<T8> g8_;
+  const ParamGenerator<T9> g9_;
+};  // class CartesianProductGenerator9
+
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10>
+class CartesianProductGenerator10
+    : public ParamGeneratorInterface< ::testing::tuple<T1, T2, T3, T4, T5, T6,
+        T7, T8, T9, T10> > {
+ public:
+  typedef ::testing::tuple<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10> ParamType;
+
+  CartesianProductGenerator10(const ParamGenerator<T1>& g1,
+      const ParamGenerator<T2>& g2, const ParamGenerator<T3>& g3,
+      const ParamGenerator<T4>& g4, const ParamGenerator<T5>& g5,
+      const ParamGenerator<T6>& g6, const ParamGenerator<T7>& g7,
+      const ParamGenerator<T8>& g8, const ParamGenerator<T9>& g9,
+      const ParamGenerator<T10>& g10)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9), g10_(g10) {}
+  virtual ~CartesianProductGenerator10() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, g1_, g1_.begin(), g2_, g2_.begin(), g3_,
+        g3_.begin(), g4_, g4_.begin(), g5_, g5_.begin(), g6_, g6_.begin(), g7_,
+        g7_.begin(), g8_, g8_.begin(), g9_, g9_.begin(), g10_, g10_.begin());
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, g1_, g1_.end(), g2_, g2_.end(), g3_, g3_.end(),
+        g4_, g4_.end(), g5_, g5_.end(), g6_, g6_.end(), g7_, g7_.end(), g8_,
+        g8_.end(), g9_, g9_.end(), g10_, g10_.end());
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base,
+      const ParamGenerator<T1>& g1,
+      const typename ParamGenerator<T1>::iterator& current1,
+      const ParamGenerator<T2>& g2,
+      const typename ParamGenerator<T2>::iterator& current2,
+      const ParamGenerator<T3>& g3,
+      const typename ParamGenerator<T3>::iterator& current3,
+      const ParamGenerator<T4>& g4,
+      const typename ParamGenerator<T4>::iterator& current4,
+      const ParamGenerator<T5>& g5,
+      const typename ParamGenerator<T5>::iterator& current5,
+      const ParamGenerator<T6>& g6,
+      const typename ParamGenerator<T6>::iterator& current6,
+      const ParamGenerator<T7>& g7,
+      const typename ParamGenerator<T7>::iterator& current7,
+      const ParamGenerator<T8>& g8,
+      const typename ParamGenerator<T8>::iterator& current8,
+      const ParamGenerator<T9>& g9,
+      const typename ParamGenerator<T9>::iterator& current9,
+      const ParamGenerator<T10>& g10,
+      const typename ParamGenerator<T10>::iterator& current10)
+        : base_(base),
+          begin1_(g1.begin()), end1_(g1.end()), current1_(current1),
+          begin2_(g2.begin()), end2_(g2.end()), current2_(current2),
+          begin3_(g3.begin()), end3_(g3.end()), current3_(current3),
+          begin4_(g4.begin()), end4_(g4.end()), current4_(current4),
+          begin5_(g5.begin()), end5_(g5.end()), current5_(current5),
+          begin6_(g6.begin()), end6_(g6.end()), current6_(current6),
+          begin7_(g7.begin()), end7_(g7.end()), current7_(current7),
+          begin8_(g8.begin()), end8_(g8.end()), current8_(current8),
+          begin9_(g9.begin()), end9_(g9.end()), current9_(current9),
+          begin10_(g10.begin()), end10_(g10.end()), current10_(current10)    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current10_;
+      if (current10_ == end10_) {
+        current10_ = begin10_;
+        ++current9_;
+      }
+      if (current9_ == end9_) {
+        current9_ = begin9_;
+        ++current8_;
+      }
+      if (current8_ == end8_) {
+        current8_ = begin8_;
+        ++current7_;
+      }
+      if (current7_ == end7_) {
+        current7_ = begin7_;
+        ++current6_;
+      }
+      if (current6_ == end6_) {
+        current6_ = begin6_;
+        ++current5_;
+      }
+      if (current5_ == end5_) {
+        current5_ = begin5_;
+        ++current4_;
+      }
+      if (current4_ == end4_) {
+        current4_ = begin4_;
+        ++current3_;
+      }
+      if (current3_ == end3_) {
+        current3_ = begin3_;
+        ++current2_;
+      }
+      if (current2_ == end2_) {
+        current2_ = begin2_;
+        ++current1_;
+      }
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         (
+          current1_ == typed_other->current1_ &&
+          current2_ == typed_other->current2_ &&
+          current3_ == typed_other->current3_ &&
+          current4_ == typed_other->current4_ &&
+          current5_ == typed_other->current5_ &&
+          current6_ == typed_other->current6_ &&
+          current7_ == typed_other->current7_ &&
+          current8_ == typed_other->current8_ &&
+          current9_ == typed_other->current9_ &&
+          current10_ == typed_other->current10_);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_),
+        begin1_(other.begin1_),
+        end1_(other.end1_),
+        current1_(other.current1_),
+        begin2_(other.begin2_),
+        end2_(other.end2_),
+        current2_(other.current2_),
+        begin3_(other.begin3_),
+        end3_(other.end3_),
+        current3_(other.current3_),
+        begin4_(other.begin4_),
+        end4_(other.end4_),
+        current4_(other.current4_),
+        begin5_(other.begin5_),
+        end5_(other.end5_),
+        current5_(other.current5_),
+        begin6_(other.begin6_),
+        end6_(other.end6_),
+        current6_(other.current6_),
+        begin7_(other.begin7_),
+        end7_(other.end7_),
+        current7_(other.current7_),
+        begin8_(other.begin8_),
+        end8_(other.end8_),
+        current8_(other.current8_),
+        begin9_(other.begin9_),
+        end9_(other.end9_),
+        current9_(other.current9_),
+        begin10_(other.begin10_),
+        end10_(other.end10_),
+        current10_(other.current10_) {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType(*current1_, *current2_, *current3_,
+            *current4_, *current5_, *current6_, *current7_, *current8_,
+            *current9_, *current10_);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+          current1_ == end1_ ||
+          current2_ == end2_ ||
+          current3_ == end3_ ||
+          current4_ == end4_ ||
+          current5_ == end5_ ||
+          current6_ == end6_ ||
+          current7_ == end7_ ||
+          current8_ == end8_ ||
+          current9_ == end9_ ||
+          current10_ == end10_;
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+    const typename ParamGenerator<T1>::iterator begin1_;
+    const typename ParamGenerator<T1>::iterator end1_;
+    typename ParamGenerator<T1>::iterator current1_;
+    const typename ParamGenerator<T2>::iterator begin2_;
+    const typename ParamGenerator<T2>::iterator end2_;
+    typename ParamGenerator<T2>::iterator current2_;
+    const typename ParamGenerator<T3>::iterator begin3_;
+    const typename ParamGenerator<T3>::iterator end3_;
+    typename ParamGenerator<T3>::iterator current3_;
+    const typename ParamGenerator<T4>::iterator begin4_;
+    const typename ParamGenerator<T4>::iterator end4_;
+    typename ParamGenerator<T4>::iterator current4_;
+    const typename ParamGenerator<T5>::iterator begin5_;
+    const typename ParamGenerator<T5>::iterator end5_;
+    typename ParamGenerator<T5>::iterator current5_;
+    const typename ParamGenerator<T6>::iterator begin6_;
+    const typename ParamGenerator<T6>::iterator end6_;
+    typename ParamGenerator<T6>::iterator current6_;
+    const typename ParamGenerator<T7>::iterator begin7_;
+    const typename ParamGenerator<T7>::iterator end7_;
+    typename ParamGenerator<T7>::iterator current7_;
+    const typename ParamGenerator<T8>::iterator begin8_;
+    const typename ParamGenerator<T8>::iterator end8_;
+    typename ParamGenerator<T8>::iterator current8_;
+    const typename ParamGenerator<T9>::iterator begin9_;
+    const typename ParamGenerator<T9>::iterator end9_;
+    typename ParamGenerator<T9>::iterator current9_;
+    const typename ParamGenerator<T10>::iterator begin10_;
+    const typename ParamGenerator<T10>::iterator end10_;
+    typename ParamGenerator<T10>::iterator current10_;
+    ParamType current_value_;
+  };  // class CartesianProductGenerator10::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator10& other);
+
+  const ParamGenerator<T1> g1_;
+  const ParamGenerator<T2> g2_;
+  const ParamGenerator<T3> g3_;
+  const ParamGenerator<T4> g4_;
+  const ParamGenerator<T5> g5_;
+  const ParamGenerator<T6> g6_;
+  const ParamGenerator<T7> g7_;
+  const ParamGenerator<T8> g8_;
+  const ParamGenerator<T9> g9_;
+  const ParamGenerator<T10> g10_;
+};  // class CartesianProductGenerator10
+
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Helper classes providing Combine() with polymorphic features. They allow
+// casting CartesianProductGeneratorN<T> to ParamGenerator<U> if T is
+// convertible to U.
+//
+template <class Generator1, class Generator2>
+class CartesianProductHolder2 {
+ public:
+CartesianProductHolder2(const Generator1& g1, const Generator2& g2)
+      : g1_(g1), g2_(g2) {}
+  template <typename T1, typename T2>
+  operator ParamGenerator< ::testing::tuple<T1, T2> >() const {
+    return ParamGenerator< ::testing::tuple<T1, T2> >(
+        new CartesianProductGenerator2<T1, T2>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder2& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+};  // class CartesianProductHolder2
+
+template <class Generator1, class Generator2, class Generator3>
+class CartesianProductHolder3 {
+ public:
+CartesianProductHolder3(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3)
+      : g1_(g1), g2_(g2), g3_(g3) {}
+  template <typename T1, typename T2, typename T3>
+  operator ParamGenerator< ::testing::tuple<T1, T2, T3> >() const {
+    return ParamGenerator< ::testing::tuple<T1, T2, T3> >(
+        new CartesianProductGenerator3<T1, T2, T3>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder3& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+};  // class CartesianProductHolder3
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4>
+class CartesianProductHolder4 {
+ public:
+CartesianProductHolder4(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4) {}
+  template <typename T1, typename T2, typename T3, typename T4>
+  operator ParamGenerator< ::testing::tuple<T1, T2, T3, T4> >() const {
+    return ParamGenerator< ::testing::tuple<T1, T2, T3, T4> >(
+        new CartesianProductGenerator4<T1, T2, T3, T4>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder4& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+};  // class CartesianProductHolder4
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5>
+class CartesianProductHolder5 {
+ public:
+CartesianProductHolder5(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5>
+  operator ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5> >() const {
+    return ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5> >(
+        new CartesianProductGenerator5<T1, T2, T3, T4, T5>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder5& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+};  // class CartesianProductHolder5
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5, class Generator6>
+class CartesianProductHolder6 {
+ public:
+CartesianProductHolder6(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5,
+      typename T6>
+  operator ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5, T6> >() const {
+    return ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5, T6> >(
+        new CartesianProductGenerator6<T1, T2, T3, T4, T5, T6>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_),
+        static_cast<ParamGenerator<T6> >(g6_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder6& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+};  // class CartesianProductHolder6
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5, class Generator6, class Generator7>
+class CartesianProductHolder7 {
+ public:
+CartesianProductHolder7(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5,
+      typename T6, typename T7>
+  operator ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5, T6,
+      T7> >() const {
+    return ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5, T6, T7> >(
+        new CartesianProductGenerator7<T1, T2, T3, T4, T5, T6, T7>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_),
+        static_cast<ParamGenerator<T6> >(g6_),
+        static_cast<ParamGenerator<T7> >(g7_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder7& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+};  // class CartesianProductHolder7
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5, class Generator6, class Generator7,
+    class Generator8>
+class CartesianProductHolder8 {
+ public:
+CartesianProductHolder8(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7, const Generator8& g8)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7),
+          g8_(g8) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5,
+      typename T6, typename T7, typename T8>
+  operator ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5, T6, T7,
+      T8> >() const {
+    return ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5, T6, T7, T8> >(
+        new CartesianProductGenerator8<T1, T2, T3, T4, T5, T6, T7, T8>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_),
+        static_cast<ParamGenerator<T6> >(g6_),
+        static_cast<ParamGenerator<T7> >(g7_),
+        static_cast<ParamGenerator<T8> >(g8_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder8& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+  const Generator8 g8_;
+};  // class CartesianProductHolder8
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5, class Generator6, class Generator7,
+    class Generator8, class Generator9>
+class CartesianProductHolder9 {
+ public:
+CartesianProductHolder9(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7, const Generator8& g8,
+    const Generator9& g9)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5,
+      typename T6, typename T7, typename T8, typename T9>
+  operator ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5, T6, T7, T8,
+      T9> >() const {
+    return ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5, T6, T7, T8,
+        T9> >(
+        new CartesianProductGenerator9<T1, T2, T3, T4, T5, T6, T7, T8, T9>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_),
+        static_cast<ParamGenerator<T6> >(g6_),
+        static_cast<ParamGenerator<T7> >(g7_),
+        static_cast<ParamGenerator<T8> >(g8_),
+        static_cast<ParamGenerator<T9> >(g9_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder9& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+  const Generator8 g8_;
+  const Generator9 g9_;
+};  // class CartesianProductHolder9
+
+template <class Generator1, class Generator2, class Generator3,
+    class Generator4, class Generator5, class Generator6, class Generator7,
+    class Generator8, class Generator9, class Generator10>
+class CartesianProductHolder10 {
+ public:
+CartesianProductHolder10(const Generator1& g1, const Generator2& g2,
+    const Generator3& g3, const Generator4& g4, const Generator5& g5,
+    const Generator6& g6, const Generator7& g7, const Generator8& g8,
+    const Generator9& g9, const Generator10& g10)
+      : g1_(g1), g2_(g2), g3_(g3), g4_(g4), g5_(g5), g6_(g6), g7_(g7), g8_(g8),
+          g9_(g9), g10_(g10) {}
+  template <typename T1, typename T2, typename T3, typename T4, typename T5,
+      typename T6, typename T7, typename T8, typename T9, typename T10>
+  operator ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5, T6, T7, T8, T9,
+      T10> >() const {
+    return ParamGenerator< ::testing::tuple<T1, T2, T3, T4, T5, T6, T7, T8, T9,
+        T10> >(
+        new CartesianProductGenerator10<T1, T2, T3, T4, T5, T6, T7, T8, T9,
+            T10>(
+        static_cast<ParamGenerator<T1> >(g1_),
+        static_cast<ParamGenerator<T2> >(g2_),
+        static_cast<ParamGenerator<T3> >(g3_),
+        static_cast<ParamGenerator<T4> >(g4_),
+        static_cast<ParamGenerator<T5> >(g5_),
+        static_cast<ParamGenerator<T6> >(g6_),
+        static_cast<ParamGenerator<T7> >(g7_),
+        static_cast<ParamGenerator<T8> >(g8_),
+        static_cast<ParamGenerator<T9> >(g9_),
+        static_cast<ParamGenerator<T10> >(g10_)));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder10& other);
+
+  const Generator1 g1_;
+  const Generator2 g2_;
+  const Generator3 g3_;
+  const Generator4 g4_;
+  const Generator5 g5_;
+  const Generator6 g6_;
+  const Generator7 g7_;
+  const Generator8 g8_;
+  const Generator9 g9_;
+  const Generator10 g10_;
+};  // class CartesianProductHolder10
+
+# endif  // GTEST_HAS_COMBINE
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  //  GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-param-util-generated.h.pump b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-param-util-generated.h.pump
new file mode 100644
index 0000000000..5c7c47af0b
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-param-util-generated.h.pump
@@ -0,0 +1,286 @@
+$$ -*- mode: c++; -*-
+$var n = 50  $$ Maximum length of Values arguments we want to support.
+$var maxtuple = 10  $$ Maximum number of Combine arguments we want to support.
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: vladl@google.com (Vlad Losev)
+
+// Type and function utilities for implementing parameterized tests.
+// This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+// Currently Google Test supports at most $n arguments in Values,
+// and at most $maxtuple arguments in Combine. Please contact
+// googletestframework@googlegroups.com if you need more.
+// Please note that the number of arguments to Combine is limited
+// by the maximum arity of the implementation of tuple which is
+// currently set at $maxtuple.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+#include "gtest/internal/gtest-param-util.h"
+#include "gtest/internal/gtest-port.h"
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+
+// Forward declarations of ValuesIn(), which is implemented in
+// include/gtest/gtest-param-test.h.
+template <typename ForwardIterator>
+internal::ParamGenerator<
+  typename ::testing::internal::IteratorTraits<ForwardIterator>::value_type>
+ValuesIn(ForwardIterator begin, ForwardIterator end);
+
+template <typename T, size_t N>
+internal::ParamGenerator<T> ValuesIn(const T (&array)[N]);
+
+template <class Container>
+internal::ParamGenerator<typename Container::value_type> ValuesIn(
+    const Container& container);
+
+namespace internal {
+
+// Used in the Values() function to provide polymorphic capabilities.
+$range i 1..n
+$for i [[
+$range j 1..i
+
+template <$for j, [[typename T$j]]>
+class ValueArray$i {
+ public:
+  $if i==1 [[explicit ]]ValueArray$i($for j, [[T$j v$j]]) : $for j, [[v$(j)_(v$j)]] {}
+
+  template <typename T>
+  operator ParamGenerator<T>() const {
+    const T array[] = {$for j, [[static_cast<T>(v$(j)_)]]};
+    return ValuesIn(array);
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const ValueArray$i& other);
+
+$for j [[
+
+  const T$j v$(j)_;
+]]
+
+};
+
+]]
+
+# if GTEST_HAS_COMBINE
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Generates values from the Cartesian product of values produced
+// by the argument generators.
+//
+$range i 2..maxtuple
+$for i [[
+$range j 1..i
+$range k 2..i
+
+template <$for j, [[typename T$j]]>
+class CartesianProductGenerator$i
+    : public ParamGeneratorInterface< ::testing::tuple<$for j, [[T$j]]> > {
+ public:
+  typedef ::testing::tuple<$for j, [[T$j]]> ParamType;
+
+  CartesianProductGenerator$i($for j, [[const ParamGenerator<T$j>& g$j]])
+      : $for j, [[g$(j)_(g$j)]] {}
+  virtual ~CartesianProductGenerator$i() {}
+
+  virtual ParamIteratorInterface<ParamType>* Begin() const {
+    return new Iterator(this, $for j, [[g$(j)_, g$(j)_.begin()]]);
+  }
+  virtual ParamIteratorInterface<ParamType>* End() const {
+    return new Iterator(this, $for j, [[g$(j)_, g$(j)_.end()]]);
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<ParamType> {
+   public:
+    Iterator(const ParamGeneratorInterface<ParamType>* base, $for j, [[
+
+      const ParamGenerator<T$j>& g$j,
+      const typename ParamGenerator<T$j>::iterator& current$(j)]])
+        : base_(base),
+$for j, [[
+
+          begin$(j)_(g$j.begin()), end$(j)_(g$j.end()), current$(j)_(current$j)
+]]    {
+      ComputeCurrentValue();
+    }
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<ParamType>* BaseGenerator() const {
+      return base_;
+    }
+    // Advance should not be called on beyond-of-range iterators
+    // so no component iterators must be beyond end of range, either.
+    virtual void Advance() {
+      assert(!AtEnd());
+      ++current$(i)_;
+
+$for k [[
+      if (current$(i+2-k)_ == end$(i+2-k)_) {
+        current$(i+2-k)_ = begin$(i+2-k)_;
+        ++current$(i+2-k-1)_;
+      }
+
+]]
+      ComputeCurrentValue();
+    }
+    virtual ParamIteratorInterface<ParamType>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const ParamType* Current() const { return &current_value_; }
+    virtual bool Equals(const ParamIteratorInterface<ParamType>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const Iterator* typed_other =
+          CheckedDowncastToActualType<const Iterator>(&other);
+      // We must report iterators equal if they both point beyond their
+      // respective ranges. That can happen in a variety of fashions,
+      // so we have to consult AtEnd().
+      return (AtEnd() && typed_other->AtEnd()) ||
+         ($for j  && [[
+
+          current$(j)_ == typed_other->current$(j)_
+]]);
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : base_(other.base_), $for j, [[
+
+        begin$(j)_(other.begin$(j)_),
+        end$(j)_(other.end$(j)_),
+        current$(j)_(other.current$(j)_)
+]] {
+      ComputeCurrentValue();
+    }
+
+    void ComputeCurrentValue() {
+      if (!AtEnd())
+        current_value_ = ParamType($for j, [[*current$(j)_]]);
+    }
+    bool AtEnd() const {
+      // We must report iterator past the end of the range when either of the
+      // component iterators has reached the end of its range.
+      return
+$for j  || [[
+
+          current$(j)_ == end$(j)_
+]];
+    }
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<ParamType>* const base_;
+    // begin[i]_ and end[i]_ define the i-th range that Iterator traverses.
+    // current[i]_ is the actual traversing iterator.
+$for j [[
+
+    const typename ParamGenerator<T$j>::iterator begin$(j)_;
+    const typename ParamGenerator<T$j>::iterator end$(j)_;
+    typename ParamGenerator<T$j>::iterator current$(j)_;
+]]
+
+    ParamType current_value_;
+  };  // class CartesianProductGenerator$i::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductGenerator$i& other);
+
+
+$for j [[
+  const ParamGenerator<T$j> g$(j)_;
+
+]]
+};  // class CartesianProductGenerator$i
+
+
+]]
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Helper classes providing Combine() with polymorphic features. They allow
+// casting CartesianProductGeneratorN<T> to ParamGenerator<U> if T is
+// convertible to U.
+//
+$range i 2..maxtuple
+$for i [[
+$range j 1..i
+
+template <$for j, [[class Generator$j]]>
+class CartesianProductHolder$i {
+ public:
+CartesianProductHolder$i($for j, [[const Generator$j& g$j]])
+      : $for j, [[g$(j)_(g$j)]] {}
+  template <$for j, [[typename T$j]]>
+  operator ParamGenerator< ::testing::tuple<$for j, [[T$j]]> >() const {
+    return ParamGenerator< ::testing::tuple<$for j, [[T$j]]> >(
+        new CartesianProductGenerator$i<$for j, [[T$j]]>(
+$for j,[[
+
+        static_cast<ParamGenerator<T$j> >(g$(j)_)
+]]));
+  }
+
+ private:
+  // No implementation - assignment is unsupported.
+  void operator=(const CartesianProductHolder$i& other);
+
+
+$for j [[
+  const Generator$j g$(j)_;
+
+]]
+};  // class CartesianProductHolder$i
+
+]]
+
+# endif  // GTEST_HAS_COMBINE
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  //  GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_GENERATED_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-param-util.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-param-util.h
new file mode 100644
index 0000000000..82cab9b020
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-param-util.h
@@ -0,0 +1,731 @@
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: vladl@google.com (Vlad Losev)
+
+// Type and function utilities for implementing parameterized tests.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_H_
+
+#include <ctype.h>
+
+#include <iterator>
+#include <set>
+#include <utility>
+#include <vector>
+
+// scripts/fuse_gtest.py depends on gtest's own header being #included
+// *unconditionally*.  Therefore these #includes cannot be moved
+// inside #if GTEST_HAS_PARAM_TEST.
+#include "gtest/internal/gtest-internal.h"
+#include "gtest/internal/gtest-linked_ptr.h"
+#include "gtest/internal/gtest-port.h"
+#include "gtest/gtest-printers.h"
+
+#if GTEST_HAS_PARAM_TEST
+
+namespace testing {
+
+// Input to a parameterized test name generator, describing a test parameter.
+// Consists of the parameter value and the integer parameter index.
+template <class ParamType>
+struct TestParamInfo {
+  TestParamInfo(const ParamType& a_param, size_t an_index) :
+    param(a_param),
+    index(an_index) {}
+  ParamType param;
+  size_t index;
+};
+
+// A builtin parameterized test name generator which returns the result of
+// testing::PrintToString.
+struct PrintToStringParamName {
+  template <class ParamType>
+  std::string operator()(const TestParamInfo<ParamType>& info) const {
+    return PrintToString(info.param);
+  }
+};
+
+namespace internal {
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Outputs a message explaining invalid registration of different
+// fixture class for the same test case. This may happen when
+// TEST_P macro is used to define two tests with the same name
+// but in different namespaces.
+GTEST_API_ void ReportInvalidTestCaseType(const char* test_case_name,
+                                          CodeLocation code_location);
+
+template <typename> class ParamGeneratorInterface;
+template <typename> class ParamGenerator;
+
+// Interface for iterating over elements provided by an implementation
+// of ParamGeneratorInterface<T>.
+template <typename T>
+class ParamIteratorInterface {
+ public:
+  virtual ~ParamIteratorInterface() {}
+  // A pointer to the base generator instance.
+  // Used only for the purposes of iterator comparison
+  // to make sure that two iterators belong to the same generator.
+  virtual const ParamGeneratorInterface<T>* BaseGenerator() const = 0;
+  // Advances iterator to point to the next element
+  // provided by the generator. The caller is responsible
+  // for not calling Advance() on an iterator equal to
+  // BaseGenerator()->End().
+  virtual void Advance() = 0;
+  // Clones the iterator object. Used for implementing copy semantics
+  // of ParamIterator<T>.
+  virtual ParamIteratorInterface* Clone() const = 0;
+  // Dereferences the current iterator and provides (read-only) access
+  // to the pointed value. It is the caller's responsibility not to call
+  // Current() on an iterator equal to BaseGenerator()->End().
+  // Used for implementing ParamGenerator<T>::operator*().
+  virtual const T* Current() const = 0;
+  // Determines whether the given iterator and other point to the same
+  // element in the sequence generated by the generator.
+  // Used for implementing ParamGenerator<T>::operator==().
+  virtual bool Equals(const ParamIteratorInterface& other) const = 0;
+};
+
+// Class iterating over elements provided by an implementation of
+// ParamGeneratorInterface<T>. It wraps ParamIteratorInterface<T>
+// and implements the const forward iterator concept.
+template <typename T>
+class ParamIterator {
+ public:
+  typedef T value_type;
+  typedef const T& reference;
+  typedef ptrdiff_t difference_type;
+
+  // ParamIterator assumes ownership of the impl_ pointer.
+  ParamIterator(const ParamIterator& other) : impl_(other.impl_->Clone()) {}
+  ParamIterator& operator=(const ParamIterator& other) {
+    if (this != &other)
+      impl_.reset(other.impl_->Clone());
+    return *this;
+  }
+
+  const T& operator*() const { return *impl_->Current(); }
+  const T* operator->() const { return impl_->Current(); }
+  // Prefix version of operator++.
+  ParamIterator& operator++() {
+    impl_->Advance();
+    return *this;
+  }
+  // Postfix version of operator++.
+  ParamIterator operator++(int /*unused*/) {
+    ParamIteratorInterface<T>* clone = impl_->Clone();
+    impl_->Advance();
+    return ParamIterator(clone);
+  }
+  bool operator==(const ParamIterator& other) const {
+    return impl_.get() == other.impl_.get() || impl_->Equals(*other.impl_);
+  }
+  bool operator!=(const ParamIterator& other) const {
+    return !(*this == other);
+  }
+
+ private:
+  friend class ParamGenerator<T>;
+  explicit ParamIterator(ParamIteratorInterface<T>* impl) : impl_(impl) {}
+  scoped_ptr<ParamIteratorInterface<T> > impl_;
+};
+
+// ParamGeneratorInterface<T> is the binary interface to access generators
+// defined in other translation units.
+template <typename T>
+class ParamGeneratorInterface {
+ public:
+  typedef T ParamType;
+
+  virtual ~ParamGeneratorInterface() {}
+
+  // Generator interface definition
+  virtual ParamIteratorInterface<T>* Begin() const = 0;
+  virtual ParamIteratorInterface<T>* End() const = 0;
+};
+
+// Wraps ParamGeneratorInterface<T> and provides general generator syntax
+// compatible with the STL Container concept.
+// This class implements copy initialization semantics and the contained
+// ParamGeneratorInterface<T> instance is shared among all copies
+// of the original object. This is possible because that instance is immutable.
+template<typename T>
+class ParamGenerator {
+ public:
+  typedef ParamIterator<T> iterator;
+
+  explicit ParamGenerator(ParamGeneratorInterface<T>* impl) : impl_(impl) {}
+  ParamGenerator(const ParamGenerator& other) : impl_(other.impl_) {}
+
+  ParamGenerator& operator=(const ParamGenerator& other) {
+    impl_ = other.impl_;
+    return *this;
+  }
+
+  iterator begin() const { return iterator(impl_->Begin()); }
+  iterator end() const { return iterator(impl_->End()); }
+
+ private:
+  linked_ptr<const ParamGeneratorInterface<T> > impl_;
+};
+
+// Generates values from a range of two comparable values. Can be used to
+// generate sequences of user-defined types that implement operator+() and
+// operator<().
+// This class is used in the Range() function.
+template <typename T, typename IncrementT>
+class RangeGenerator : public ParamGeneratorInterface<T> {
+ public:
+  RangeGenerator(T begin, T end, IncrementT step)
+      : begin_(begin), end_(end),
+        step_(step), end_index_(CalculateEndIndex(begin, end, step)) {}
+  virtual ~RangeGenerator() {}
+
+  virtual ParamIteratorInterface<T>* Begin() const {
+    return new Iterator(this, begin_, 0, step_);
+  }
+  virtual ParamIteratorInterface<T>* End() const {
+    return new Iterator(this, end_, end_index_, step_);
+  }
+
+ private:
+  class Iterator : public ParamIteratorInterface<T> {
+   public:
+    Iterator(const ParamGeneratorInterface<T>* base, T value, int index,
+             IncrementT step)
+        : base_(base), value_(value), index_(index), step_(step) {}
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<T>* BaseGenerator() const {
+      return base_;
+    }
+    virtual void Advance() {
+      value_ = static_cast<T>(value_ + step_);
+      index_++;
+    }
+    virtual ParamIteratorInterface<T>* Clone() const {
+      return new Iterator(*this);
+    }
+    virtual const T* Current() const { return &value_; }
+    virtual bool Equals(const ParamIteratorInterface<T>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      const int other_index =
+          CheckedDowncastToActualType<const Iterator>(&other)->index_;
+      return index_ == other_index;
+    }
+
+   private:
+    Iterator(const Iterator& other)
+        : ParamIteratorInterface<T>(),
+          base_(other.base_), value_(other.value_), index_(other.index_),
+          step_(other.step_) {}
+
+    // No implementation - assignment is unsupported.
+    void operator=(const Iterator& other);
+
+    const ParamGeneratorInterface<T>* const base_;
+    T value_;
+    int index_;
+    const IncrementT step_;
+  };  // class RangeGenerator::Iterator
+
+  static int CalculateEndIndex(const T& begin,
+                               const T& end,
+                               const IncrementT& step) {
+    int end_index = 0;
+    for (T i = begin; i < end; i = static_cast<T>(i + step))
+      end_index++;
+    return end_index;
+  }
+
+  // No implementation - assignment is unsupported.
+  void operator=(const RangeGenerator& other);
+
+  const T begin_;
+  const T end_;
+  const IncrementT step_;
+  // The index for the end() iterator. All the elements in the generated
+  // sequence are indexed (0-based) to aid iterator comparison.
+  const int end_index_;
+};  // class RangeGenerator
+
+
+// Generates values from a pair of STL-style iterators. Used in the
+// ValuesIn() function. The elements are copied from the source range
+// since the source can be located on the stack, and the generator
+// is likely to persist beyond that stack frame.
+template <typename T>
+class ValuesInIteratorRangeGenerator : public ParamGeneratorInterface<T> {
+ public:
+  template <typename ForwardIterator>
+  ValuesInIteratorRangeGenerator(ForwardIterator begin, ForwardIterator end)
+      : container_(begin, end) {}
+  virtual ~ValuesInIteratorRangeGenerator() {}
+
+  virtual ParamIteratorInterface<T>* Begin() const {
+    return new Iterator(this, container_.begin());
+  }
+  virtual ParamIteratorInterface<T>* End() const {
+    return new Iterator(this, container_.end());
+  }
+
+ private:
+  typedef typename ::std::vector<T> ContainerType;
+
+  class Iterator : public ParamIteratorInterface<T> {
+   public:
+    Iterator(const ParamGeneratorInterface<T>* base,
+             typename ContainerType::const_iterator iterator)
+        : base_(base), iterator_(iterator) {}
+    virtual ~Iterator() {}
+
+    virtual const ParamGeneratorInterface<T>* BaseGenerator() const {
+      return base_;
+    }
+    virtual void Advance() {
+      ++iterator_;
+      value_.reset();
+    }
+    virtual ParamIteratorInterface<T>* Clone() const {
+      return new Iterator(*this);
+    }
+    // We need to use cached value referenced by iterator_ because *iterator_
+    // can return a temporary object (and of type other then T), so just
+    // having "return &*iterator_;" doesn't work.
+    // value_ is updated here and not in Advance() because Advance()
+    // can advance iterator_ beyond the end of the range, and we cannot
+    // detect that fact. The client code, on the other hand, is
+    // responsible for not calling Current() on an out-of-range iterator.
+    virtual const T* Current() const {
+      if (value_.get() == NULL)
+        value_.reset(new T(*iterator_));
+      return value_.get();
+    }
+    virtual bool Equals(const ParamIteratorInterface<T>& other) const {
+      // Having the same base generator guarantees that the other
+      // iterator is of the same type and we can downcast.
+      GTEST_CHECK_(BaseGenerator() == other.BaseGenerator())
+          << "The program attempted to compare iterators "
+          << "from different generators." << std::endl;
+      return iterator_ ==
+          CheckedDowncastToActualType<const Iterator>(&other)->iterator_;
+    }
+
+   private:
+    Iterator(const Iterator& other)
+          // The explicit constructor call suppresses a false warning
+          // emitted by gcc when supplied with the -Wextra option.
+        : ParamIteratorInterface<T>(),
+          base_(other.base_),
+          iterator_(other.iterator_) {}
+
+    const ParamGeneratorInterface<T>* const base_;
+    typename ContainerType::const_iterator iterator_;
+    // A cached value of *iterator_. We keep it here to allow access by
+    // pointer in the wrapping iterator's operator->().
+    // value_ needs to be mutable to be accessed in Current().
+    // Use of scoped_ptr helps manage cached value's lifetime,
+    // which is bound by the lifespan of the iterator itself.
+    mutable scoped_ptr<const T> value_;
+  };  // class ValuesInIteratorRangeGenerator::Iterator
+
+  // No implementation - assignment is unsupported.
+  void operator=(const ValuesInIteratorRangeGenerator& other);
+
+  const ContainerType container_;
+};  // class ValuesInIteratorRangeGenerator
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Default parameterized test name generator, returns a string containing the
+// integer test parameter index.
+template <class ParamType>
+std::string DefaultParamName(const TestParamInfo<ParamType>& info) {
+  Message name_stream;
+  name_stream << info.index;
+  return name_stream.GetString();
+}
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Parameterized test name overload helpers, which help the
+// INSTANTIATE_TEST_CASE_P macro choose between the default parameterized
+// test name generator and user param name generator.
+template <class ParamType, class ParamNameGenFunctor>
+ParamNameGenFunctor GetParamNameGen(ParamNameGenFunctor func) {
+  return func;
+}
+
+template <class ParamType>
+struct ParamNameGenFunc {
+  typedef std::string Type(const TestParamInfo<ParamType>&);
+};
+
+template <class ParamType>
+typename ParamNameGenFunc<ParamType>::Type *GetParamNameGen() {
+  return DefaultParamName;
+}
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Stores a parameter value and later creates tests parameterized with that
+// value.
+template <class TestClass>
+class ParameterizedTestFactory : public TestFactoryBase {
+ public:
+  typedef typename TestClass::ParamType ParamType;
+  explicit ParameterizedTestFactory(ParamType parameter) :
+      parameter_(parameter) {}
+  virtual Test* CreateTest() {
+    TestClass::SetParam(&parameter_);
+    return new TestClass();
+  }
+
+ private:
+  const ParamType parameter_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestFactory);
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// TestMetaFactoryBase is a base class for meta-factories that create
+// test factories for passing into MakeAndRegisterTestInfo function.
+template <class ParamType>
+class TestMetaFactoryBase {
+ public:
+  virtual ~TestMetaFactoryBase() {}
+
+  virtual TestFactoryBase* CreateTestFactory(ParamType parameter) = 0;
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// TestMetaFactory creates test factories for passing into
+// MakeAndRegisterTestInfo function. Since MakeAndRegisterTestInfo receives
+// ownership of test factory pointer, same factory object cannot be passed
+// into that method twice. But ParameterizedTestCaseInfo is going to call
+// it for each Test/Parameter value combination. Thus it needs meta factory
+// creator class.
+template <class TestCase>
+class TestMetaFactory
+    : public TestMetaFactoryBase<typename TestCase::ParamType> {
+ public:
+  typedef typename TestCase::ParamType ParamType;
+
+  TestMetaFactory() {}
+
+  virtual TestFactoryBase* CreateTestFactory(ParamType parameter) {
+    return new ParameterizedTestFactory<TestCase>(parameter);
+  }
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestMetaFactory);
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// ParameterizedTestCaseInfoBase is a generic interface
+// to ParameterizedTestCaseInfo classes. ParameterizedTestCaseInfoBase
+// accumulates test information provided by TEST_P macro invocations
+// and generators provided by INSTANTIATE_TEST_CASE_P macro invocations
+// and uses that information to register all resulting test instances
+// in RegisterTests method. The ParameterizeTestCaseRegistry class holds
+// a collection of pointers to the ParameterizedTestCaseInfo objects
+// and calls RegisterTests() on each of them when asked.
+class ParameterizedTestCaseInfoBase {
+ public:
+  virtual ~ParameterizedTestCaseInfoBase() {}
+
+  // Base part of test case name for display purposes.
+  virtual const string& GetTestCaseName() const = 0;
+  // Test case id to verify identity.
+  virtual TypeId GetTestCaseTypeId() const = 0;
+  // UnitTest class invokes this method to register tests in this
+  // test case right before running them in RUN_ALL_TESTS macro.
+  // This method should not be called more then once on any single
+  // instance of a ParameterizedTestCaseInfoBase derived class.
+  virtual void RegisterTests() = 0;
+
+ protected:
+  ParameterizedTestCaseInfoBase() {}
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestCaseInfoBase);
+};
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// ParameterizedTestCaseInfo accumulates tests obtained from TEST_P
+// macro invocations for a particular test case and generators
+// obtained from INSTANTIATE_TEST_CASE_P macro invocations for that
+// test case. It registers tests with all values generated by all
+// generators when asked.
+template <class TestCase>
+class ParameterizedTestCaseInfo : public ParameterizedTestCaseInfoBase {
+ public:
+  // ParamType and GeneratorCreationFunc are private types but are required
+  // for declarations of public methods AddTestPattern() and
+  // AddTestCaseInstantiation().
+  typedef typename TestCase::ParamType ParamType;
+  // A function that returns an instance of appropriate generator type.
+  typedef ParamGenerator<ParamType>(GeneratorCreationFunc)();
+  typedef typename ParamNameGenFunc<ParamType>::Type ParamNameGeneratorFunc;
+
+  explicit ParameterizedTestCaseInfo(
+      const char* name, CodeLocation code_location)
+      : test_case_name_(name), code_location_(code_location) {}
+
+  // Test case base name for display purposes.
+  virtual const string& GetTestCaseName() const { return test_case_name_; }
+  // Test case id to verify identity.
+  virtual TypeId GetTestCaseTypeId() const { return GetTypeId<TestCase>(); }
+  // TEST_P macro uses AddTestPattern() to record information
+  // about a single test in a LocalTestInfo structure.
+  // test_case_name is the base name of the test case (without invocation
+  // prefix). test_base_name is the name of an individual test without
+  // parameter index. For the test SequenceA/FooTest.DoBar/1 FooTest is
+  // test case base name and DoBar is test base name.
+  void AddTestPattern(const char* test_case_name,
+                      const char* test_base_name,
+                      TestMetaFactoryBase<ParamType>* meta_factory) {
+    tests_.push_back(linked_ptr<TestInfo>(new TestInfo(test_case_name,
+                                                       test_base_name,
+                                                       meta_factory)));
+  }
+  // INSTANTIATE_TEST_CASE_P macro uses AddGenerator() to record information
+  // about a generator.
+  int AddTestCaseInstantiation(const string& instantiation_name,
+                               GeneratorCreationFunc* func,
+                               ParamNameGeneratorFunc* name_func,
+                               const char* file,
+                               int line) {
+    instantiations_.push_back(
+        InstantiationInfo(instantiation_name, func, name_func, file, line));
+    return 0;  // Return value used only to run this method in namespace scope.
+  }
+  // UnitTest class invokes this method to register tests in this test case
+  // test cases right before running tests in RUN_ALL_TESTS macro.
+  // This method should not be called more then once on any single
+  // instance of a ParameterizedTestCaseInfoBase derived class.
+  // UnitTest has a guard to prevent from calling this method more then once.
+  virtual void RegisterTests() {
+    for (typename TestInfoContainer::iterator test_it = tests_.begin();
+         test_it != tests_.end(); ++test_it) {
+      linked_ptr<TestInfo> test_info = *test_it;
+      for (typename InstantiationContainer::iterator gen_it =
+               instantiations_.begin(); gen_it != instantiations_.end();
+               ++gen_it) {
+        const string& instantiation_name = gen_it->name;
+        ParamGenerator<ParamType> generator((*gen_it->generator)());
+        ParamNameGeneratorFunc* name_func = gen_it->name_func;
+        const char* file = gen_it->file;
+        int line = gen_it->line;
+
+        string test_case_name;
+        if ( !instantiation_name.empty() )
+          test_case_name = instantiation_name + "/";
+        test_case_name += test_info->test_case_base_name;
+
+        size_t i = 0;
+        std::set<std::string> test_param_names;
+        for (typename ParamGenerator<ParamType>::iterator param_it =
+                 generator.begin();
+             param_it != generator.end(); ++param_it, ++i) {
+          Message test_name_stream;
+
+          std::string param_name = name_func(
+              TestParamInfo<ParamType>(*param_it, i));
+
+          GTEST_CHECK_(IsValidParamName(param_name))
+              << "Parameterized test name '" << param_name
+              << "' is invalid, in " << file
+              << " line " << line << std::endl;
+
+          GTEST_CHECK_(test_param_names.count(param_name) == 0)
+              << "Duplicate parameterized test name '" << param_name
+              << "', in " << file << " line " << line << std::endl;
+
+          test_param_names.insert(param_name);
+
+          test_name_stream << test_info->test_base_name << "/" << param_name;
+          MakeAndRegisterTestInfo(
+              test_case_name.c_str(),
+              test_name_stream.GetString().c_str(),
+              NULL,  // No type parameter.
+              PrintToString(*param_it).c_str(),
+              code_location_,
+              GetTestCaseTypeId(),
+              TestCase::SetUpTestCase,
+              TestCase::TearDownTestCase,
+              test_info->test_meta_factory->CreateTestFactory(*param_it));
+        }  // for param_it
+      }  // for gen_it
+    }  // for test_it
+  }  // RegisterTests
+
+ private:
+  // LocalTestInfo structure keeps information about a single test registered
+  // with TEST_P macro.
+  struct TestInfo {
+    TestInfo(const char* a_test_case_base_name,
+             const char* a_test_base_name,
+             TestMetaFactoryBase<ParamType>* a_test_meta_factory) :
+        test_case_base_name(a_test_case_base_name),
+        test_base_name(a_test_base_name),
+        test_meta_factory(a_test_meta_factory) {}
+
+    const string test_case_base_name;
+    const string test_base_name;
+    const scoped_ptr<TestMetaFactoryBase<ParamType> > test_meta_factory;
+  };
+  typedef ::std::vector<linked_ptr<TestInfo> > TestInfoContainer;
+  // Records data received from INSTANTIATE_TEST_CASE_P macros:
+  //  <Instantiation name, Sequence generator creation function,
+  //     Name generator function, Source file, Source line>
+  struct InstantiationInfo {
+      InstantiationInfo(const std::string &name_in,
+                        GeneratorCreationFunc* generator_in,
+                        ParamNameGeneratorFunc* name_func_in,
+                        const char* file_in,
+                        int line_in)
+          : name(name_in),
+            generator(generator_in),
+            name_func(name_func_in),
+            file(file_in),
+            line(line_in) {}
+
+      std::string name;
+      GeneratorCreationFunc* generator;
+      ParamNameGeneratorFunc* name_func;
+      const char* file;
+      int line;
+  };
+  typedef ::std::vector<InstantiationInfo> InstantiationContainer;
+
+  static bool IsValidParamName(const std::string& name) {
+    // Check for empty string
+    if (name.empty())
+      return false;
+
+    // Check for invalid characters
+    for (std::string::size_type index = 0; index < name.size(); ++index) {
+      if (!isalnum(name[index]) && name[index] != '_')
+        return false;
+    }
+
+    return true;
+  }
+
+  const string test_case_name_;
+  CodeLocation code_location_;
+  TestInfoContainer tests_;
+  InstantiationContainer instantiations_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestCaseInfo);
+};  // class ParameterizedTestCaseInfo
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// ParameterizedTestCaseRegistry contains a map of ParameterizedTestCaseInfoBase
+// classes accessed by test case names. TEST_P and INSTANTIATE_TEST_CASE_P
+// macros use it to locate their corresponding ParameterizedTestCaseInfo
+// descriptors.
+class ParameterizedTestCaseRegistry {
+ public:
+  ParameterizedTestCaseRegistry() {}
+  ~ParameterizedTestCaseRegistry() {
+    for (TestCaseInfoContainer::iterator it = test_case_infos_.begin();
+         it != test_case_infos_.end(); ++it) {
+      delete *it;
+    }
+  }
+
+  // Looks up or creates and returns a structure containing information about
+  // tests and instantiations of a particular test case.
+  template <class TestCase>
+  ParameterizedTestCaseInfo<TestCase>* GetTestCasePatternHolder(
+      const char* test_case_name,
+      CodeLocation code_location) {
+    ParameterizedTestCaseInfo<TestCase>* typed_test_info = NULL;
+    for (TestCaseInfoContainer::iterator it = test_case_infos_.begin();
+         it != test_case_infos_.end(); ++it) {
+      if ((*it)->GetTestCaseName() == test_case_name) {
+        if ((*it)->GetTestCaseTypeId() != GetTypeId<TestCase>()) {
+          // Complain about incorrect usage of Google Test facilities
+          // and terminate the program since we cannot guaranty correct
+          // test case setup and tear-down in this case.
+          ReportInvalidTestCaseType(test_case_name, code_location);
+          posix::Abort();
+        } else {
+          // At this point we are sure that the object we found is of the same
+          // type we are looking for, so we downcast it to that type
+          // without further checks.
+          typed_test_info = CheckedDowncastToActualType<
+              ParameterizedTestCaseInfo<TestCase> >(*it);
+        }
+        break;
+      }
+    }
+    if (typed_test_info == NULL) {
+      typed_test_info = new ParameterizedTestCaseInfo<TestCase>(
+          test_case_name, code_location);
+      test_case_infos_.push_back(typed_test_info);
+    }
+    return typed_test_info;
+  }
+  void RegisterTests() {
+    for (TestCaseInfoContainer::iterator it = test_case_infos_.begin();
+         it != test_case_infos_.end(); ++it) {
+      (*it)->RegisterTests();
+    }
+  }
+
+ private:
+  typedef ::std::vector<ParameterizedTestCaseInfoBase*> TestCaseInfoContainer;
+
+  TestCaseInfoContainer test_case_infos_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ParameterizedTestCaseRegistry);
+};
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  //  GTEST_HAS_PARAM_TEST
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PARAM_UTIL_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-port-arch.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-port-arch.h
new file mode 100644
index 0000000000..74ab949057
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-port-arch.h
@@ -0,0 +1,93 @@
+// Copyright 2015, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file defines the GTEST_OS_* macro.
+// It is separate from gtest-port.h so that custom/gtest-port.h can include it.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_ARCH_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_ARCH_H_
+
+// Determines the platform on which Google Test is compiled.
+#ifdef __CYGWIN__
+# define GTEST_OS_CYGWIN 1
+#elif defined __SYMBIAN32__
+# define GTEST_OS_SYMBIAN 1
+#elif defined _WIN32
+# define GTEST_OS_WINDOWS 1
+# ifdef _WIN32_WCE
+#  define GTEST_OS_WINDOWS_MOBILE 1
+# elif defined(__MINGW__) || defined(__MINGW32__)
+#  define GTEST_OS_WINDOWS_MINGW 1
+# elif defined(WINAPI_FAMILY)
+#  include <winapifamily.h>
+#  if WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
+#   define GTEST_OS_WINDOWS_DESKTOP 1
+#  elif WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_PHONE_APP)
+#   define GTEST_OS_WINDOWS_PHONE 1
+#  elif WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_APP)
+#   define GTEST_OS_WINDOWS_RT 1
+#  else
+    // WINAPI_FAMILY defined but no known partition matched.
+    // Default to desktop.
+#   define GTEST_OS_WINDOWS_DESKTOP 1
+#  endif
+# else
+#  define GTEST_OS_WINDOWS_DESKTOP 1
+# endif  // _WIN32_WCE
+#elif defined __APPLE__
+# define GTEST_OS_MAC 1
+# if TARGET_OS_IPHONE
+#  define GTEST_OS_IOS 1
+# endif
+#elif defined __FreeBSD__
+# define GTEST_OS_FREEBSD 1
+#elif defined __linux__
+# define GTEST_OS_LINUX 1
+# if defined __ANDROID__
+#  define GTEST_OS_LINUX_ANDROID 1
+# endif
+#elif defined __MVS__
+# define GTEST_OS_ZOS 1
+#elif defined(__sun) && defined(__SVR4)
+# define GTEST_OS_SOLARIS 1
+#elif defined(_AIX)
+# define GTEST_OS_AIX 1
+#elif defined(__hpux)
+# define GTEST_OS_HPUX 1
+#elif defined __native_client__
+# define GTEST_OS_NACL 1
+#elif defined __OpenBSD__
+# define GTEST_OS_OPENBSD 1
+#elif defined __QNX__
+# define GTEST_OS_QNX 1
+#endif  // __CYGWIN__
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_ARCH_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-port.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-port.h
new file mode 100644
index 0000000000..0094ed5077
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-port.h
@@ -0,0 +1,2554 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan)
+//
+// Low-level types and utilities for porting Google Test to various
+// platforms.  All macros ending with _ and symbols defined in an
+// internal namespace are subject to change without notice.  Code
+// outside Google Test MUST NOT USE THEM DIRECTLY.  Macros that don't
+// end with _ are part of Google Test's public API and can be used by
+// code outside Google Test.
+//
+// This file is fundamental to Google Test.  All other Google Test source
+// files are expected to #include this.  Therefore, it cannot #include
+// any other Google Test header.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_H_
+
+// Environment-describing macros
+// -----------------------------
+//
+// Google Test can be used in many different environments.  Macros in
+// this section tell Google Test what kind of environment it is being
+// used in, such that Google Test can provide environment-specific
+// features and implementations.
+//
+// Google Test tries to automatically detect the properties of its
+// environment, so users usually don't need to worry about these
+// macros.  However, the automatic detection is not perfect.
+// Sometimes it's necessary for a user to define some of the following
+// macros in the build script to override Google Test's decisions.
+//
+// If the user doesn't define a macro in the list, Google Test will
+// provide a default definition.  After this header is #included, all
+// macros in this list will be defined to either 1 or 0.
+//
+// Notes to maintainers:
+//   - Each macro here is a user-tweakable knob; do not grow the list
+//     lightly.
+//   - Use #if to key off these macros.  Don't use #ifdef or "#if
+//     defined(...)", which will not work as these macros are ALWAYS
+//     defined.
+//
+//   GTEST_HAS_CLONE          - Define it to 1/0 to indicate that clone(2)
+//                              is/isn't available.
+//   GTEST_HAS_EXCEPTIONS     - Define it to 1/0 to indicate that exceptions
+//                              are enabled.
+//   GTEST_HAS_GLOBAL_STRING  - Define it to 1/0 to indicate that ::string
+//                              is/isn't available (some systems define
+//                              ::string, which is different to std::string).
+//   GTEST_HAS_GLOBAL_WSTRING - Define it to 1/0 to indicate that ::string
+//                              is/isn't available (some systems define
+//                              ::wstring, which is different to std::wstring).
+//   GTEST_HAS_POSIX_RE       - Define it to 1/0 to indicate that POSIX regular
+//                              expressions are/aren't available.
+//   GTEST_HAS_PTHREAD        - Define it to 1/0 to indicate that <pthread.h>
+//                              is/isn't available.
+//   GTEST_HAS_RTTI           - Define it to 1/0 to indicate that RTTI is/isn't
+//                              enabled.
+//   GTEST_HAS_STD_WSTRING    - Define it to 1/0 to indicate that
+//                              std::wstring does/doesn't work (Google Test can
+//                              be used where std::wstring is unavailable).
+//   GTEST_HAS_TR1_TUPLE      - Define it to 1/0 to indicate tr1::tuple
+//                              is/isn't available.
+//   GTEST_HAS_SEH            - Define it to 1/0 to indicate whether the
+//                              compiler supports Microsoft's "Structured
+//                              Exception Handling".
+//   GTEST_HAS_STREAM_REDIRECTION
+//                            - Define it to 1/0 to indicate whether the
+//                              platform supports I/O stream redirection using
+//                              dup() and dup2().
+//   GTEST_USE_OWN_TR1_TUPLE  - Define it to 1/0 to indicate whether Google
+//                              Test's own tr1 tuple implementation should be
+//                              used.  Unused when the user sets
+//                              GTEST_HAS_TR1_TUPLE to 0.
+//   GTEST_LANG_CXX11         - Define it to 1/0 to indicate that Google Test
+//                              is building in C++11/C++98 mode.
+//   GTEST_LINKED_AS_SHARED_LIBRARY
+//                            - Define to 1 when compiling tests that use
+//                              Google Test as a shared library (known as
+//                              DLL on Windows).
+//   GTEST_CREATE_SHARED_LIBRARY
+//                            - Define to 1 when compiling Google Test itself
+//                              as a shared library.
+
+// Platform-indicating macros
+// --------------------------
+//
+// Macros indicating the platform on which Google Test is being used
+// (a macro is defined to 1 if compiled on the given platform;
+// otherwise UNDEFINED -- it's never defined to 0.).  Google Test
+// defines these macros automatically.  Code outside Google Test MUST
+// NOT define them.
+//
+//   GTEST_OS_AIX      - IBM AIX
+//   GTEST_OS_CYGWIN   - Cygwin
+//   GTEST_OS_FREEBSD  - FreeBSD
+//   GTEST_OS_HPUX     - HP-UX
+//   GTEST_OS_LINUX    - Linux
+//     GTEST_OS_LINUX_ANDROID - Google Android
+//   GTEST_OS_MAC      - Mac OS X
+//     GTEST_OS_IOS    - iOS
+//   GTEST_OS_NACL     - Google Native Client (NaCl)
+//   GTEST_OS_OPENBSD  - OpenBSD
+//   GTEST_OS_QNX      - QNX
+//   GTEST_OS_SOLARIS  - Sun Solaris
+//   GTEST_OS_SYMBIAN  - Symbian
+//   GTEST_OS_WINDOWS  - Windows (Desktop, MinGW, or Mobile)
+//     GTEST_OS_WINDOWS_DESKTOP  - Windows Desktop
+//     GTEST_OS_WINDOWS_MINGW    - MinGW
+//     GTEST_OS_WINDOWS_MOBILE   - Windows Mobile
+//     GTEST_OS_WINDOWS_PHONE    - Windows Phone
+//     GTEST_OS_WINDOWS_RT       - Windows Store App/WinRT
+//   GTEST_OS_ZOS      - z/OS
+//
+// Among the platforms, Cygwin, Linux, Max OS X, and Windows have the
+// most stable support.  Since core members of the Google Test project
+// don't have access to other platforms, support for them may be less
+// stable.  If you notice any problems on your platform, please notify
+// googletestframework@googlegroups.com (patches for fixing them are
+// even more welcome!).
+//
+// It is possible that none of the GTEST_OS_* macros are defined.
+
+// Feature-indicating macros
+// -------------------------
+//
+// Macros indicating which Google Test features are available (a macro
+// is defined to 1 if the corresponding feature is supported;
+// otherwise UNDEFINED -- it's never defined to 0.).  Google Test
+// defines these macros automatically.  Code outside Google Test MUST
+// NOT define them.
+//
+// These macros are public so that portable tests can be written.
+// Such tests typically surround code using a feature with an #if
+// which controls that code.  For example:
+//
+// #if GTEST_HAS_DEATH_TEST
+//   EXPECT_DEATH(DoSomethingDeadly());
+// #endif
+//
+//   GTEST_HAS_COMBINE      - the Combine() function (for value-parameterized
+//                            tests)
+//   GTEST_HAS_DEATH_TEST   - death tests
+//   GTEST_HAS_PARAM_TEST   - value-parameterized tests
+//   GTEST_HAS_TYPED_TEST   - typed tests
+//   GTEST_HAS_TYPED_TEST_P - type-parameterized tests
+//   GTEST_IS_THREADSAFE    - Google Test is thread-safe.
+//   GTEST_USES_POSIX_RE    - enhanced POSIX regex is used. Do not confuse with
+//                            GTEST_HAS_POSIX_RE (see above) which users can
+//                            define themselves.
+//   GTEST_USES_SIMPLE_RE   - our own simple regex is used;
+//                            the above two are mutually exclusive.
+//   GTEST_CAN_COMPARE_NULL - accepts untyped NULL in EXPECT_EQ().
+
+// Misc public macros
+// ------------------
+//
+//   GTEST_FLAG(flag_name)  - references the variable corresponding to
+//                            the given Google Test flag.
+
+// Internal utilities
+// ------------------
+//
+// The following macros and utilities are for Google Test's INTERNAL
+// use only.  Code outside Google Test MUST NOT USE THEM DIRECTLY.
+//
+// Macros for basic C++ coding:
+//   GTEST_AMBIGUOUS_ELSE_BLOCKER_ - for disabling a gcc warning.
+//   GTEST_ATTRIBUTE_UNUSED_  - declares that a class' instances or a
+//                              variable don't have to be used.
+//   GTEST_DISALLOW_ASSIGN_   - disables operator=.
+//   GTEST_DISALLOW_COPY_AND_ASSIGN_ - disables copy ctor and operator=.
+//   GTEST_MUST_USE_RESULT_   - declares that a function's result must be used.
+//   GTEST_INTENTIONAL_CONST_COND_PUSH_ - start code section where MSVC C4127 is
+//                                        suppressed (constant conditional).
+//   GTEST_INTENTIONAL_CONST_COND_POP_  - finish code section where MSVC C4127
+//                                        is suppressed.
+//
+// C++11 feature wrappers:
+//
+//   testing::internal::move  - portability wrapper for std::move.
+//
+// Synchronization:
+//   Mutex, MutexLock, ThreadLocal, GetThreadCount()
+//                            - synchronization primitives.
+//
+// Template meta programming:
+//   is_pointer     - as in TR1; needed on Symbian and IBM XL C/C++ only.
+//   IteratorTraits - partial implementation of std::iterator_traits, which
+//                    is not available in libCstd when compiled with Sun C++.
+//
+// Smart pointers:
+//   scoped_ptr     - as in TR2.
+//
+// Regular expressions:
+//   RE             - a simple regular expression class using the POSIX
+//                    Extended Regular Expression syntax on UNIX-like
+//                    platforms, or a reduced regular exception syntax on
+//                    other platforms, including Windows.
+//
+// Logging:
+//   GTEST_LOG_()   - logs messages at the specified severity level.
+//   LogToStderr()  - directs all log messages to stderr.
+//   FlushInfoLog() - flushes informational log messages.
+//
+// Stdout and stderr capturing:
+//   CaptureStdout()     - starts capturing stdout.
+//   GetCapturedStdout() - stops capturing stdout and returns the captured
+//                         string.
+//   CaptureStderr()     - starts capturing stderr.
+//   GetCapturedStderr() - stops capturing stderr and returns the captured
+//                         string.
+//
+// Integer types:
+//   TypeWithSize   - maps an integer to a int type.
+//   Int32, UInt32, Int64, UInt64, TimeInMillis
+//                  - integers of known sizes.
+//   BiggestInt     - the biggest signed integer type.
+//
+// Command-line utilities:
+//   GTEST_DECLARE_*()  - declares a flag.
+//   GTEST_DEFINE_*()   - defines a flag.
+//   GetInjectableArgvs() - returns the command line as a vector of strings.
+//
+// Environment variable utilities:
+//   GetEnv()             - gets the value of an environment variable.
+//   BoolFromGTestEnv()   - parses a bool environment variable.
+//   Int32FromGTestEnv()  - parses an Int32 environment variable.
+//   StringFromGTestEnv() - parses a string environment variable.
+
+#include <ctype.h>   // for isspace, etc
+#include <stddef.h>  // for ptrdiff_t
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#ifndef _WIN32_WCE
+# include <sys/types.h>
+# include <sys/stat.h>
+#endif  // !_WIN32_WCE
+
+#if defined __APPLE__
+# include <AvailabilityMacros.h>
+# include <TargetConditionals.h>
+#endif
+
+#include <algorithm>  // NOLINT
+#include <iostream>  // NOLINT
+#include <sstream>  // NOLINT
+#include <string>  // NOLINT
+#include <utility>
+#include <vector>  // NOLINT
+
+#include "gtest/internal/gtest-port-arch.h"
+#include "gtest/internal/custom/gtest-port.h"
+
+#if !defined(GTEST_DEV_EMAIL_)
+# define GTEST_DEV_EMAIL_ "googletestframework@@googlegroups.com"
+# define GTEST_FLAG_PREFIX_ "gtest_"
+# define GTEST_FLAG_PREFIX_DASH_ "gtest-"
+# define GTEST_FLAG_PREFIX_UPPER_ "GTEST_"
+# define GTEST_NAME_ "Google Test"
+# define GTEST_PROJECT_URL_ "https://github.com/google/googletest/"
+#endif  // !defined(GTEST_DEV_EMAIL_)
+
+#if !defined(GTEST_INIT_GOOGLE_TEST_NAME_)
+# define GTEST_INIT_GOOGLE_TEST_NAME_ "testing::InitGoogleTest"
+#endif  // !defined(GTEST_INIT_GOOGLE_TEST_NAME_)
+
+// Determines the version of gcc that is used to compile this.
+#ifdef __GNUC__
+// 40302 means version 4.3.2.
+# define GTEST_GCC_VER_ \
+    (__GNUC__*10000 + __GNUC_MINOR__*100 + __GNUC_PATCHLEVEL__)
+#endif  // __GNUC__
+
+// Macros for disabling Microsoft Visual C++ warnings.
+//
+//   GTEST_DISABLE_MSC_WARNINGS_PUSH_(4800 4385)
+//   /* code that triggers warnings C4800 and C4385 */
+//   GTEST_DISABLE_MSC_WARNINGS_POP_()
+#if _MSC_VER >= 1500
+# define GTEST_DISABLE_MSC_WARNINGS_PUSH_(warnings) \
+    __pragma(warning(push))                        \
+    __pragma(warning(disable: warnings))
+# define GTEST_DISABLE_MSC_WARNINGS_POP_()          \
+    __pragma(warning(pop))
+#else
+// Older versions of MSVC don't have __pragma.
+# define GTEST_DISABLE_MSC_WARNINGS_PUSH_(warnings)
+# define GTEST_DISABLE_MSC_WARNINGS_POP_()
+#endif
+
+#ifndef GTEST_LANG_CXX11
+// gcc and clang define __GXX_EXPERIMENTAL_CXX0X__ when
+// -std={c,gnu}++{0x,11} is passed.  The C++11 standard specifies a
+// value for __cplusplus, and recent versions of clang, gcc, and
+// probably other compilers set that too in C++11 mode.
+# if __GXX_EXPERIMENTAL_CXX0X__ || __cplusplus >= 201103L
+// Compiling in at least C++11 mode.
+#  define GTEST_LANG_CXX11 1
+# else
+#  define GTEST_LANG_CXX11 0
+# endif
+#endif
+
+// Distinct from C++11 language support, some environments don't provide
+// proper C++11 library support. Notably, it's possible to build in
+// C++11 mode when targeting Mac OS X 10.6, which has an old libstdc++
+// with no C++11 support.
+//
+// libstdc++ has sufficient C++11 support as of GCC 4.6.0, __GLIBCXX__
+// 20110325, but maintenance releases in the 4.4 and 4.5 series followed
+// this date, so check for those versions by their date stamps.
+// https://gcc.gnu.org/onlinedocs/libstdc++/manual/abi.html#abi.versioning
+#if GTEST_LANG_CXX11 && \
+    (!defined(__GLIBCXX__) || ( \
+        __GLIBCXX__ >= 20110325ul &&  /* GCC >= 4.6.0 */ \
+        /* Blacklist of patch releases of older branches: */ \
+        __GLIBCXX__ != 20110416ul &&  /* GCC 4.4.6 */ \
+        __GLIBCXX__ != 20120313ul &&  /* GCC 4.4.7 */ \
+        __GLIBCXX__ != 20110428ul &&  /* GCC 4.5.3 */ \
+        __GLIBCXX__ != 20120702ul))   /* GCC 4.5.4 */
+# define GTEST_STDLIB_CXX11 1
+#endif
+
+// Only use C++11 library features if the library provides them.
+#if GTEST_STDLIB_CXX11
+# define GTEST_HAS_STD_BEGIN_AND_END_ 1
+# define GTEST_HAS_STD_FORWARD_LIST_ 1
+# define GTEST_HAS_STD_FUNCTION_ 1
+# define GTEST_HAS_STD_INITIALIZER_LIST_ 1
+# define GTEST_HAS_STD_MOVE_ 1
+# define GTEST_HAS_STD_SHARED_PTR_ 1
+# define GTEST_HAS_STD_TYPE_TRAITS_ 1
+# define GTEST_HAS_STD_UNIQUE_PTR_ 1
+#endif
+
+// C++11 specifies that <tuple> provides std::tuple.
+// Some platforms still might not have it, however.
+#if GTEST_LANG_CXX11
+# define GTEST_HAS_STD_TUPLE_ 1
+# if defined(__clang__)
+// Inspired by http://clang.llvm.org/docs/LanguageExtensions.html#__has_include
+#  if defined(__has_include) && !__has_include(<tuple>)
+#   undef GTEST_HAS_STD_TUPLE_
+#  endif
+# elif defined(_MSC_VER)
+// Inspired by boost/config/stdlib/dinkumware.hpp
+#  if defined(_CPPLIB_VER) && _CPPLIB_VER < 520
+#   undef GTEST_HAS_STD_TUPLE_
+#  endif
+# elif defined(__GLIBCXX__)
+// Inspired by boost/config/stdlib/libstdcpp3.hpp,
+// http://gcc.gnu.org/gcc-4.2/changes.html and
+// http://gcc.gnu.org/onlinedocs/libstdc++/manual/bk01pt01ch01.html#manual.intro.status.standard.200x
+#  if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ < 2)
+#   undef GTEST_HAS_STD_TUPLE_
+#  endif
+# endif
+#endif
+
+// Brings in definitions for functions used in the testing::internal::posix
+// namespace (read, write, close, chdir, isatty, stat). We do not currently
+// use them on Windows Mobile.
+#if GTEST_OS_WINDOWS
+# if !GTEST_OS_WINDOWS_MOBILE
+#  include <direct.h>
+#  include <io.h>
+# endif
+// In order to avoid having to include <windows.h>, use forward declaration
+// assuming CRITICAL_SECTION is a typedef of _RTL_CRITICAL_SECTION.
+// This assumption is verified by
+// WindowsTypesTest.CRITICAL_SECTIONIs_RTL_CRITICAL_SECTION.
+struct _RTL_CRITICAL_SECTION;
+#else
+// This assumes that non-Windows OSes provide unistd.h. For OSes where this
+// is not the case, we need to include headers that provide the functions
+// mentioned above.
+# include <unistd.h>
+# include <strings.h>
+#endif  // GTEST_OS_WINDOWS
+
+#if GTEST_OS_LINUX_ANDROID
+// Used to define __ANDROID_API__ matching the target NDK API level.
+#  include <android/api-level.h>  // NOLINT
+#endif
+
+// Defines this to true iff Google Test can use POSIX regular expressions.
+#ifndef GTEST_HAS_POSIX_RE
+# if GTEST_OS_LINUX_ANDROID
+// On Android, <regex.h> is only available starting with Gingerbread.
+#  define GTEST_HAS_POSIX_RE (__ANDROID_API__ >= 9)
+# else
+#  define GTEST_HAS_POSIX_RE (!GTEST_OS_WINDOWS)
+# endif
+#endif
+
+#if GTEST_USES_PCRE
+// The appropriate headers have already been included.
+
+#elif GTEST_HAS_POSIX_RE
+
+// On some platforms, <regex.h> needs someone to define size_t, and
+// won't compile otherwise.  We can #include it here as we already
+// included <stdlib.h>, which is guaranteed to define size_t through
+// <stddef.h>.
+# include <regex.h>  // NOLINT
+
+# define GTEST_USES_POSIX_RE 1
+
+#elif GTEST_OS_WINDOWS
+
+// <regex.h> is not available on Windows.  Use our own simple regex
+// implementation instead.
+# define GTEST_USES_SIMPLE_RE 1
+
+#else
+
+// <regex.h> may not be available on this platform.  Use our own
+// simple regex implementation instead.
+# define GTEST_USES_SIMPLE_RE 1
+
+#endif  // GTEST_USES_PCRE
+
+#ifndef GTEST_HAS_EXCEPTIONS
+// The user didn't tell us whether exceptions are enabled, so we need
+// to figure it out.
+# if defined(_MSC_VER) || defined(__BORLANDC__)
+// MSVC's and C++Builder's implementations of the STL use the _HAS_EXCEPTIONS
+// macro to enable exceptions, so we'll do the same.
+// Assumes that exceptions are enabled by default.
+#  ifndef _HAS_EXCEPTIONS
+#   define _HAS_EXCEPTIONS 1
+#  endif  // _HAS_EXCEPTIONS
+#  define GTEST_HAS_EXCEPTIONS _HAS_EXCEPTIONS
+# elif defined(__clang__)
+// clang defines __EXCEPTIONS iff exceptions are enabled before clang 220714,
+// but iff cleanups are enabled after that. In Obj-C++ files, there can be
+// cleanups for ObjC exceptions which also need cleanups, even if C++ exceptions
+// are disabled. clang has __has_feature(cxx_exceptions) which checks for C++
+// exceptions starting at clang r206352, but which checked for cleanups prior to
+// that. To reliably check for C++ exception availability with clang, check for
+// __EXCEPTIONS && __has_feature(cxx_exceptions).
+#  define GTEST_HAS_EXCEPTIONS (__EXCEPTIONS && __has_feature(cxx_exceptions))
+# elif defined(__GNUC__) && __EXCEPTIONS
+// gcc defines __EXCEPTIONS to 1 iff exceptions are enabled.
+#  define GTEST_HAS_EXCEPTIONS 1
+# elif defined(__SUNPRO_CC)
+// Sun Pro CC supports exceptions.  However, there is no compile-time way of
+// detecting whether they are enabled or not.  Therefore, we assume that
+// they are enabled unless the user tells us otherwise.
+#  define GTEST_HAS_EXCEPTIONS 1
+# elif defined(__IBMCPP__) && __EXCEPTIONS
+// xlC defines __EXCEPTIONS to 1 iff exceptions are enabled.
+#  define GTEST_HAS_EXCEPTIONS 1
+# elif defined(__HP_aCC)
+// Exception handling is in effect by default in HP aCC compiler. It has to
+// be turned of by +noeh compiler option if desired.
+#  define GTEST_HAS_EXCEPTIONS 1
+# else
+// For other compilers, we assume exceptions are disabled to be
+// conservative.
+#  define GTEST_HAS_EXCEPTIONS 0
+# endif  // defined(_MSC_VER) || defined(__BORLANDC__)
+#endif  // GTEST_HAS_EXCEPTIONS
+
+#if !defined(GTEST_HAS_STD_STRING)
+// Even though we don't use this macro any longer, we keep it in case
+// some clients still depend on it.
+# define GTEST_HAS_STD_STRING 1
+#elif !GTEST_HAS_STD_STRING
+// The user told us that ::std::string isn't available.
+# error "Google Test cannot be used where ::std::string isn't available."
+#endif  // !defined(GTEST_HAS_STD_STRING)
+
+#ifndef GTEST_HAS_GLOBAL_STRING
+// The user didn't tell us whether ::string is available, so we need
+// to figure it out.
+
+# define GTEST_HAS_GLOBAL_STRING 0
+
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+#ifndef GTEST_HAS_STD_WSTRING
+// The user didn't tell us whether ::std::wstring is available, so we need
+// to figure it out.
+// TODO(wan@google.com): uses autoconf to detect whether ::std::wstring
+//   is available.
+
+// Cygwin 1.7 and below doesn't support ::std::wstring.
+// Solaris' libc++ doesn't support it either.  Android has
+// no support for it at least as recent as Froyo (2.2).
+# define GTEST_HAS_STD_WSTRING \
+    (!(GTEST_OS_LINUX_ANDROID || GTEST_OS_CYGWIN || GTEST_OS_SOLARIS))
+
+#endif  // GTEST_HAS_STD_WSTRING
+
+#ifndef GTEST_HAS_GLOBAL_WSTRING
+// The user didn't tell us whether ::wstring is available, so we need
+// to figure it out.
+# define GTEST_HAS_GLOBAL_WSTRING \
+    (GTEST_HAS_STD_WSTRING && GTEST_HAS_GLOBAL_STRING)
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+// Determines whether RTTI is available.
+#ifndef GTEST_HAS_RTTI
+// The user didn't tell us whether RTTI is enabled, so we need to
+// figure it out.
+
+# ifdef _MSC_VER
+
+#  ifdef _CPPRTTI  // MSVC defines this macro iff RTTI is enabled.
+#   define GTEST_HAS_RTTI 1
+#  else
+#   define GTEST_HAS_RTTI 0
+#  endif
+
+// Starting with version 4.3.2, gcc defines __GXX_RTTI iff RTTI is enabled.
+# elif defined(__GNUC__) && (GTEST_GCC_VER_ >= 40302)
+
+#  ifdef __GXX_RTTI
+// When building against STLport with the Android NDK and with
+// -frtti -fno-exceptions, the build fails at link time with undefined
+// references to __cxa_bad_typeid. Note sure if STL or toolchain bug,
+// so disable RTTI when detected.
+#   if GTEST_OS_LINUX_ANDROID && defined(_STLPORT_MAJOR) && \
+       !defined(__EXCEPTIONS)
+#    define GTEST_HAS_RTTI 0
+#   else
+#    define GTEST_HAS_RTTI 1
+#   endif  // GTEST_OS_LINUX_ANDROID && __STLPORT_MAJOR && !__EXCEPTIONS
+#  else
+#   define GTEST_HAS_RTTI 0
+#  endif  // __GXX_RTTI
+
+// Clang defines __GXX_RTTI starting with version 3.0, but its manual recommends
+// using has_feature instead. has_feature(cxx_rtti) is supported since 2.7, the
+// first version with C++ support.
+# elif defined(__clang__)
+
+#  define GTEST_HAS_RTTI __has_feature(cxx_rtti)
+
+// Starting with version 9.0 IBM Visual Age defines __RTTI_ALL__ to 1 if
+// both the typeid and dynamic_cast features are present.
+# elif defined(__IBMCPP__) && (__IBMCPP__ >= 900)
+
+#  ifdef __RTTI_ALL__
+#   define GTEST_HAS_RTTI 1
+#  else
+#   define GTEST_HAS_RTTI 0
+#  endif
+
+# else
+
+// For all other compilers, we assume RTTI is enabled.
+#  define GTEST_HAS_RTTI 1
+
+# endif  // _MSC_VER
+
+#endif  // GTEST_HAS_RTTI
+
+// It's this header's responsibility to #include <typeinfo> when RTTI
+// is enabled.
+#if GTEST_HAS_RTTI
+# include <typeinfo>
+#endif
+
+// Determines whether Google Test can use the pthreads library.
+#ifndef GTEST_HAS_PTHREAD
+// The user didn't tell us explicitly, so we make reasonable assumptions about
+// which platforms have pthreads support.
+//
+// To disable threading support in Google Test, add -DGTEST_HAS_PTHREAD=0
+// to your compiler flags.
+# define GTEST_HAS_PTHREAD (GTEST_OS_LINUX || GTEST_OS_MAC || GTEST_OS_HPUX \
+    || GTEST_OS_QNX || GTEST_OS_FREEBSD || GTEST_OS_NACL)
+#endif  // GTEST_HAS_PTHREAD
+
+#if GTEST_HAS_PTHREAD
+// gtest-port.h guarantees to #include <pthread.h> when GTEST_HAS_PTHREAD is
+// true.
+# include <pthread.h>  // NOLINT
+
+// For timespec and nanosleep, used below.
+# include <time.h>  // NOLINT
+#endif
+
+// Determines if hash_map/hash_set are available.
+// Only used for testing against those containers.
+#if !defined(GTEST_HAS_HASH_MAP_)
+# if _MSC_VER
+#  define GTEST_HAS_HASH_MAP_ 1  // Indicates that hash_map is available.
+#  define GTEST_HAS_HASH_SET_ 1  // Indicates that hash_set is available.
+# endif  // _MSC_VER
+#endif  // !defined(GTEST_HAS_HASH_MAP_)
+
+// Determines whether Google Test can use tr1/tuple.  You can define
+// this macro to 0 to prevent Google Test from using tuple (any
+// feature depending on tuple with be disabled in this mode).
+#ifndef GTEST_HAS_TR1_TUPLE
+# if GTEST_OS_LINUX_ANDROID && defined(_STLPORT_MAJOR)
+// STLport, provided with the Android NDK, has neither <tr1/tuple> or <tuple>.
+#  define GTEST_HAS_TR1_TUPLE 0
+# else
+// The user didn't tell us not to do it, so we assume it's OK.
+#  define GTEST_HAS_TR1_TUPLE 1
+# endif
+#endif  // GTEST_HAS_TR1_TUPLE
+
+// Determines whether Google Test's own tr1 tuple implementation
+// should be used.
+#ifndef GTEST_USE_OWN_TR1_TUPLE
+// The user didn't tell us, so we need to figure it out.
+
+// We use our own TR1 tuple if we aren't sure the user has an
+// implementation of it already.  At this time, libstdc++ 4.0.0+ and
+// MSVC 2010 are the only mainstream standard libraries that come
+// with a TR1 tuple implementation.  NVIDIA's CUDA NVCC compiler
+// pretends to be GCC by defining __GNUC__ and friends, but cannot
+// compile GCC's tuple implementation.  MSVC 2008 (9.0) provides TR1
+// tuple in a 323 MB Feature Pack download, which we cannot assume the
+// user has.  QNX's QCC compiler is a modified GCC but it doesn't
+// support TR1 tuple.  libc++ only provides std::tuple, in C++11 mode,
+// and it can be used with some compilers that define __GNUC__.
+# if (defined(__GNUC__) && !defined(__CUDACC__) && (GTEST_GCC_VER_ >= 40000) \
+      && !GTEST_OS_QNX && !defined(_LIBCPP_VERSION)) || _MSC_VER >= 1600
+#  define GTEST_ENV_HAS_TR1_TUPLE_ 1
+# endif
+
+// C++11 specifies that <tuple> provides std::tuple. Use that if gtest is used
+// in C++11 mode and libstdc++ isn't very old (binaries targeting OS X 10.6
+// can build with clang but need to use gcc4.2's libstdc++).
+# if GTEST_LANG_CXX11 && (!defined(__GLIBCXX__) || __GLIBCXX__ > 20110325)
+#  define GTEST_ENV_HAS_STD_TUPLE_ 1
+# endif
+
+# if GTEST_ENV_HAS_TR1_TUPLE_ || GTEST_ENV_HAS_STD_TUPLE_
+#  define GTEST_USE_OWN_TR1_TUPLE 0
+# else
+#  define GTEST_USE_OWN_TR1_TUPLE 1
+# endif
+
+#endif  // GTEST_USE_OWN_TR1_TUPLE
+
+// To avoid conditional compilation everywhere, we make it
+// gtest-port.h's responsibility to #include the header implementing
+// tuple.
+#if GTEST_HAS_STD_TUPLE_
+# include <tuple>  // IWYU pragma: export
+# define GTEST_TUPLE_NAMESPACE_ ::std
+#endif  // GTEST_HAS_STD_TUPLE_
+
+// We include tr1::tuple even if std::tuple is available to define printers for
+// them.
+#if GTEST_HAS_TR1_TUPLE
+# ifndef GTEST_TUPLE_NAMESPACE_
+#  define GTEST_TUPLE_NAMESPACE_ ::std::tr1
+# endif  // GTEST_TUPLE_NAMESPACE_
+
+# if GTEST_USE_OWN_TR1_TUPLE
+#  include "gtest/internal/gtest-tuple.h"  // IWYU pragma: export  // NOLINT
+# elif GTEST_ENV_HAS_STD_TUPLE_
+#  include <tuple>
+// C++11 puts its tuple into the ::std namespace rather than
+// ::std::tr1.  gtest expects tuple to live in ::std::tr1, so put it there.
+// This causes undefined behavior, but supported compilers react in
+// the way we intend.
+namespace std {
+namespace tr1 {
+using ::std::get;
+using ::std::make_tuple;
+using ::std::tuple;
+using ::std::tuple_element;
+using ::std::tuple_size;
+}
+}
+
+# elif GTEST_OS_SYMBIAN
+
+// On Symbian, BOOST_HAS_TR1_TUPLE causes Boost's TR1 tuple library to
+// use STLport's tuple implementation, which unfortunately doesn't
+// work as the copy of STLport distributed with Symbian is incomplete.
+// By making sure BOOST_HAS_TR1_TUPLE is undefined, we force Boost to
+// use its own tuple implementation.
+#  ifdef BOOST_HAS_TR1_TUPLE
+#   undef BOOST_HAS_TR1_TUPLE
+#  endif  // BOOST_HAS_TR1_TUPLE
+
+// This prevents <boost/tr1/detail/config.hpp>, which defines
+// BOOST_HAS_TR1_TUPLE, from being #included by Boost's <tuple>.
+#  define BOOST_TR1_DETAIL_CONFIG_HPP_INCLUDED
+#  include <tuple>  // IWYU pragma: export  // NOLINT
+
+# elif defined(__GNUC__) && (GTEST_GCC_VER_ >= 40000)
+// GCC 4.0+ implements tr1/tuple in the <tr1/tuple> header.  This does
+// not conform to the TR1 spec, which requires the header to be <tuple>.
+
+#  if !GTEST_HAS_RTTI && GTEST_GCC_VER_ < 40302
+// Until version 4.3.2, gcc has a bug that causes <tr1/functional>,
+// which is #included by <tr1/tuple>, to not compile when RTTI is
+// disabled.  _TR1_FUNCTIONAL is the header guard for
+// <tr1/functional>.  Hence the following #define is a hack to prevent
+// <tr1/functional> from being included.
+#   define _TR1_FUNCTIONAL 1
+#   include <tr1/tuple>
+#   undef _TR1_FUNCTIONAL  // Allows the user to #include
+                        // <tr1/functional> if he chooses to.
+#  else
+#   include <tr1/tuple>  // NOLINT
+#  endif  // !GTEST_HAS_RTTI && GTEST_GCC_VER_ < 40302
+
+# else
+// If the compiler is not GCC 4.0+, we assume the user is using a
+// spec-conforming TR1 implementation.
+#  include <tuple>  // IWYU pragma: export  // NOLINT
+# endif  // GTEST_USE_OWN_TR1_TUPLE
+
+#endif  // GTEST_HAS_TR1_TUPLE
+
+// Determines whether clone(2) is supported.
+// Usually it will only be available on Linux, excluding
+// Linux on the Itanium architecture.
+// Also see http://linux.die.net/man/2/clone.
+#ifndef GTEST_HAS_CLONE
+// The user didn't tell us, so we need to figure it out.
+
+# if GTEST_OS_LINUX && !defined(__ia64__)
+#  if GTEST_OS_LINUX_ANDROID
+// On Android, clone() is only available on ARM starting with Gingerbread.
+#    if defined(__arm__) && __ANDROID_API__ >= 9
+#     define GTEST_HAS_CLONE 1
+#    else
+#     define GTEST_HAS_CLONE 0
+#    endif
+#  else
+#   define GTEST_HAS_CLONE 1
+#  endif
+# else
+#  define GTEST_HAS_CLONE 0
+# endif  // GTEST_OS_LINUX && !defined(__ia64__)
+
+#endif  // GTEST_HAS_CLONE
+
+// Determines whether to support stream redirection. This is used to test
+// output correctness and to implement death tests.
+#ifndef GTEST_HAS_STREAM_REDIRECTION
+// By default, we assume that stream redirection is supported on all
+// platforms except known mobile ones.
+# if GTEST_OS_WINDOWS_MOBILE || GTEST_OS_SYMBIAN || \
+    GTEST_OS_WINDOWS_PHONE || GTEST_OS_WINDOWS_RT
+#  define GTEST_HAS_STREAM_REDIRECTION 0
+# else
+#  define GTEST_HAS_STREAM_REDIRECTION 1
+# endif  // !GTEST_OS_WINDOWS_MOBILE && !GTEST_OS_SYMBIAN
+#endif  // GTEST_HAS_STREAM_REDIRECTION
+
+// Determines whether to support death tests.
+// Google Test does not support death tests for VC 7.1 and earlier as
+// abort() in a VC 7.1 application compiled as GUI in debug config
+// pops up a dialog window that cannot be suppressed programmatically.
+#if (GTEST_OS_LINUX || GTEST_OS_CYGWIN || GTEST_OS_SOLARIS || \
+     (GTEST_OS_MAC && !GTEST_OS_IOS) || \
+     (GTEST_OS_WINDOWS_DESKTOP && _MSC_VER >= 1400) || \
+     GTEST_OS_WINDOWS_MINGW || GTEST_OS_AIX || GTEST_OS_HPUX || \
+     GTEST_OS_OPENBSD || GTEST_OS_QNX || GTEST_OS_FREEBSD)
+# define GTEST_HAS_DEATH_TEST 1
+#endif
+
+// We don't support MSVC 7.1 with exceptions disabled now.  Therefore
+// all the compilers we care about are adequate for supporting
+// value-parameterized tests.
+#define GTEST_HAS_PARAM_TEST 1
+
+// Determines whether to support type-driven tests.
+
+// Typed tests need <typeinfo> and variadic macros, which GCC, VC++ 8.0,
+// Sun Pro CC, IBM Visual Age, and HP aCC support.
+#if defined(__GNUC__) || (_MSC_VER >= 1400) || defined(__SUNPRO_CC) || \
+    defined(__IBMCPP__) || defined(__HP_aCC)
+# define GTEST_HAS_TYPED_TEST 1
+# define GTEST_HAS_TYPED_TEST_P 1
+#endif
+
+// Determines whether to support Combine(). This only makes sense when
+// value-parameterized tests are enabled.  The implementation doesn't
+// work on Sun Studio since it doesn't understand templated conversion
+// operators.
+#if GTEST_HAS_PARAM_TEST && GTEST_HAS_TR1_TUPLE && !defined(__SUNPRO_CC)
+# define GTEST_HAS_COMBINE 1
+#endif
+
+// Determines whether the system compiler uses UTF-16 for encoding wide strings.
+#define GTEST_WIDE_STRING_USES_UTF16_ \
+    (GTEST_OS_WINDOWS || GTEST_OS_CYGWIN || GTEST_OS_SYMBIAN || GTEST_OS_AIX)
+
+// Determines whether test results can be streamed to a socket.
+#if GTEST_OS_LINUX
+# define GTEST_CAN_STREAM_RESULTS_ 1
+#endif
+
+// Defines some utility macros.
+
+// The GNU compiler emits a warning if nested "if" statements are followed by
+// an "else" statement and braces are not used to explicitly disambiguate the
+// "else" binding.  This leads to problems with code like:
+//
+//   if (gate)
+//     ASSERT_*(condition) << "Some message";
+//
+// The "switch (0) case 0:" idiom is used to suppress this.
+#ifdef __INTEL_COMPILER
+# define GTEST_AMBIGUOUS_ELSE_BLOCKER_
+#else
+# define GTEST_AMBIGUOUS_ELSE_BLOCKER_ switch (0) case 0: default:  // NOLINT
+#endif
+
+// Use this annotation at the end of a struct/class definition to
+// prevent the compiler from optimizing away instances that are never
+// used.  This is useful when all interesting logic happens inside the
+// c'tor and / or d'tor.  Example:
+//
+//   struct Foo {
+//     Foo() { ... }
+//   } GTEST_ATTRIBUTE_UNUSED_;
+//
+// Also use it after a variable or parameter declaration to tell the
+// compiler the variable/parameter does not have to be used.
+#if defined(__GNUC__) && !defined(COMPILER_ICC)
+# define GTEST_ATTRIBUTE_UNUSED_ __attribute__ ((unused))
+#elif defined(__clang__)
+# if __has_attribute(unused)
+#  define GTEST_ATTRIBUTE_UNUSED_ __attribute__ ((unused))
+# endif
+#endif
+#ifndef GTEST_ATTRIBUTE_UNUSED_
+# define GTEST_ATTRIBUTE_UNUSED_
+#endif
+
+// A macro to disallow operator=
+// This should be used in the private: declarations for a class.
+#define GTEST_DISALLOW_ASSIGN_(type)\
+  void operator=(type const &)
+
+// A macro to disallow copy constructor and operator=
+// This should be used in the private: declarations for a class.
+#define GTEST_DISALLOW_COPY_AND_ASSIGN_(type)\
+  type(type const &);\
+  GTEST_DISALLOW_ASSIGN_(type)
+
+// Tell the compiler to warn about unused return values for functions declared
+// with this macro.  The macro should be used on function declarations
+// following the argument list:
+//
+//   Sprocket* AllocateSprocket() GTEST_MUST_USE_RESULT_;
+#if defined(__GNUC__) && (GTEST_GCC_VER_ >= 30400) && !defined(COMPILER_ICC)
+# define GTEST_MUST_USE_RESULT_ __attribute__ ((warn_unused_result))
+#else
+# define GTEST_MUST_USE_RESULT_
+#endif  // __GNUC__ && (GTEST_GCC_VER_ >= 30400) && !COMPILER_ICC
+
+// MS C++ compiler emits warning when a conditional expression is compile time
+// constant. In some contexts this warning is false positive and needs to be
+// suppressed. Use the following two macros in such cases:
+//
+// GTEST_INTENTIONAL_CONST_COND_PUSH_()
+// while (true) {
+// GTEST_INTENTIONAL_CONST_COND_POP_()
+// }
+# define GTEST_INTENTIONAL_CONST_COND_PUSH_() \
+    GTEST_DISABLE_MSC_WARNINGS_PUSH_(4127)
+# define GTEST_INTENTIONAL_CONST_COND_POP_() \
+    GTEST_DISABLE_MSC_WARNINGS_POP_()
+
+// Determine whether the compiler supports Microsoft's Structured Exception
+// Handling.  This is supported by several Windows compilers but generally
+// does not exist on any other system.
+#ifndef GTEST_HAS_SEH
+// The user didn't tell us, so we need to figure it out.
+
+# if defined(_MSC_VER) || defined(__BORLANDC__)
+// These two compilers are known to support SEH.
+#  define GTEST_HAS_SEH 1
+# else
+// Assume no SEH.
+#  define GTEST_HAS_SEH 0
+# endif
+
+#define GTEST_IS_THREADSAFE \
+    (GTEST_HAS_MUTEX_AND_THREAD_LOCAL_ \
+     || (GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_PHONE && !GTEST_OS_WINDOWS_RT) \
+     || GTEST_HAS_PTHREAD)
+
+#endif  // GTEST_HAS_SEH
+
+#ifdef _MSC_VER
+# if GTEST_LINKED_AS_SHARED_LIBRARY
+#  define GTEST_API_ __declspec(dllimport)
+# elif GTEST_CREATE_SHARED_LIBRARY
+#  define GTEST_API_ __declspec(dllexport)
+# endif
+#elif __GNUC__ >= 4 || defined(__clang__)
+# define GTEST_API_ __attribute__((visibility ("default")))
+#endif // _MSC_VER
+
+#ifndef GTEST_API_
+# define GTEST_API_
+#endif
+
+#ifdef __GNUC__
+// Ask the compiler to never inline a given function.
+# define GTEST_NO_INLINE_ __attribute__((noinline))
+#else
+# define GTEST_NO_INLINE_
+#endif
+
+// _LIBCPP_VERSION is defined by the libc++ library from the LLVM project.
+#if defined(__GLIBCXX__) || defined(_LIBCPP_VERSION)
+# define GTEST_HAS_CXXABI_H_ 1
+#else
+# define GTEST_HAS_CXXABI_H_ 0
+#endif
+
+// A function level attribute to disable checking for use of uninitialized
+// memory when built with MemorySanitizer.
+#if defined(__clang__)
+# if __has_feature(memory_sanitizer)
+#  define GTEST_ATTRIBUTE_NO_SANITIZE_MEMORY_ \
+       __attribute__((no_sanitize_memory))
+# else
+#  define GTEST_ATTRIBUTE_NO_SANITIZE_MEMORY_
+# endif  // __has_feature(memory_sanitizer)
+#else
+# define GTEST_ATTRIBUTE_NO_SANITIZE_MEMORY_
+#endif  // __clang__
+
+// A function level attribute to disable AddressSanitizer instrumentation.
+#if defined(__clang__)
+# if __has_feature(address_sanitizer)
+#  define GTEST_ATTRIBUTE_NO_SANITIZE_ADDRESS_ \
+       __attribute__((no_sanitize_address))
+# else
+#  define GTEST_ATTRIBUTE_NO_SANITIZE_ADDRESS_
+# endif  // __has_feature(address_sanitizer)
+#else
+# define GTEST_ATTRIBUTE_NO_SANITIZE_ADDRESS_
+#endif  // __clang__
+
+// A function level attribute to disable ThreadSanitizer instrumentation.
+#if defined(__clang__)
+# if __has_feature(thread_sanitizer)
+#  define GTEST_ATTRIBUTE_NO_SANITIZE_THREAD_ \
+       __attribute__((no_sanitize_thread))
+# else
+#  define GTEST_ATTRIBUTE_NO_SANITIZE_THREAD_
+# endif  // __has_feature(thread_sanitizer)
+#else
+# define GTEST_ATTRIBUTE_NO_SANITIZE_THREAD_
+#endif  // __clang__
+
+namespace testing {
+
+class Message;
+
+#if defined(GTEST_TUPLE_NAMESPACE_)
+// Import tuple and friends into the ::testing namespace.
+// It is part of our interface, having them in ::testing allows us to change
+// their types as needed.
+using GTEST_TUPLE_NAMESPACE_::get;
+using GTEST_TUPLE_NAMESPACE_::make_tuple;
+using GTEST_TUPLE_NAMESPACE_::tuple;
+using GTEST_TUPLE_NAMESPACE_::tuple_size;
+using GTEST_TUPLE_NAMESPACE_::tuple_element;
+#endif  // defined(GTEST_TUPLE_NAMESPACE_)
+
+namespace internal {
+
+// A secret type that Google Test users don't know about.  It has no
+// definition on purpose.  Therefore it's impossible to create a
+// Secret object, which is what we want.
+class Secret;
+
+// The GTEST_COMPILE_ASSERT_ macro can be used to verify that a compile time
+// expression is true. For example, you could use it to verify the
+// size of a static array:
+//
+//   GTEST_COMPILE_ASSERT_(GTEST_ARRAY_SIZE_(names) == NUM_NAMES,
+//                         names_incorrect_size);
+//
+// or to make sure a struct is smaller than a certain size:
+//
+//   GTEST_COMPILE_ASSERT_(sizeof(foo) < 128, foo_too_large);
+//
+// The second argument to the macro is the name of the variable. If
+// the expression is false, most compilers will issue a warning/error
+// containing the name of the variable.
+
+#if GTEST_LANG_CXX11
+# define GTEST_COMPILE_ASSERT_(expr, msg) static_assert(expr, #msg)
+#else  // !GTEST_LANG_CXX11
+template <bool>
+  struct CompileAssert {
+};
+
+# define GTEST_COMPILE_ASSERT_(expr, msg) \
+  typedef ::testing::internal::CompileAssert<(static_cast<bool>(expr))> \
+      msg[static_cast<bool>(expr) ? 1 : -1] GTEST_ATTRIBUTE_UNUSED_
+#endif  // !GTEST_LANG_CXX11
+
+// Implementation details of GTEST_COMPILE_ASSERT_:
+//
+// (In C++11, we simply use static_assert instead of the following)
+//
+// - GTEST_COMPILE_ASSERT_ works by defining an array type that has -1
+//   elements (and thus is invalid) when the expression is false.
+//
+// - The simpler definition
+//
+//    #define GTEST_COMPILE_ASSERT_(expr, msg) typedef char msg[(expr) ? 1 : -1]
+//
+//   does not work, as gcc supports variable-length arrays whose sizes
+//   are determined at run-time (this is gcc's extension and not part
+//   of the C++ standard).  As a result, gcc fails to reject the
+//   following code with the simple definition:
+//
+//     int foo;
+//     GTEST_COMPILE_ASSERT_(foo, msg); // not supposed to compile as foo is
+//                                      // not a compile-time constant.
+//
+// - By using the type CompileAssert<(bool(expr))>, we ensures that
+//   expr is a compile-time constant.  (Template arguments must be
+//   determined at compile-time.)
+//
+// - The outter parentheses in CompileAssert<(bool(expr))> are necessary
+//   to work around a bug in gcc 3.4.4 and 4.0.1.  If we had written
+//
+//     CompileAssert<bool(expr)>
+//
+//   instead, these compilers will refuse to compile
+//
+//     GTEST_COMPILE_ASSERT_(5 > 0, some_message);
+//
+//   (They seem to think the ">" in "5 > 0" marks the end of the
+//   template argument list.)
+//
+// - The array size is (bool(expr) ? 1 : -1), instead of simply
+//
+//     ((expr) ? 1 : -1).
+//
+//   This is to avoid running into a bug in MS VC 7.1, which
+//   causes ((0.0) ? 1 : -1) to incorrectly evaluate to 1.
+
+// StaticAssertTypeEqHelper is used by StaticAssertTypeEq defined in gtest.h.
+//
+// This template is declared, but intentionally undefined.
+template <typename T1, typename T2>
+struct StaticAssertTypeEqHelper;
+
+template <typename T>
+struct StaticAssertTypeEqHelper<T, T> {
+  enum { value = true };
+};
+
+// Evaluates to the number of elements in 'array'.
+#define GTEST_ARRAY_SIZE_(array) (sizeof(array) / sizeof(array[0]))
+
+#if GTEST_HAS_GLOBAL_STRING
+typedef ::string string;
+#else
+typedef ::std::string string;
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+#if GTEST_HAS_GLOBAL_WSTRING
+typedef ::wstring wstring;
+#elif GTEST_HAS_STD_WSTRING
+typedef ::std::wstring wstring;
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+// A helper for suppressing warnings on constant condition.  It just
+// returns 'condition'.
+GTEST_API_ bool IsTrue(bool condition);
+
+// Defines scoped_ptr.
+
+// This implementation of scoped_ptr is PARTIAL - it only contains
+// enough stuff to satisfy Google Test's need.
+template <typename T>
+class scoped_ptr {
+ public:
+  typedef T element_type;
+
+  explicit scoped_ptr(T* p = NULL) : ptr_(p) {}
+  ~scoped_ptr() { reset(); }
+
+  T& operator*() const { return *ptr_; }
+  T* operator->() const { return ptr_; }
+  T* get() const { return ptr_; }
+
+  T* release() {
+    T* const ptr = ptr_;
+    ptr_ = NULL;
+    return ptr;
+  }
+
+  void reset(T* p = NULL) {
+    if (p != ptr_) {
+      if (IsTrue(sizeof(T) > 0)) {  // Makes sure T is a complete type.
+        delete ptr_;
+      }
+      ptr_ = p;
+    }
+  }
+
+  friend void swap(scoped_ptr& a, scoped_ptr& b) {
+    using std::swap;
+    swap(a.ptr_, b.ptr_);
+  }
+
+ private:
+  T* ptr_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(scoped_ptr);
+};
+
+// Defines RE.
+
+// A simple C++ wrapper for <regex.h>.  It uses the POSIX Extended
+// Regular Expression syntax.
+class GTEST_API_ RE {
+ public:
+  // A copy constructor is required by the Standard to initialize object
+  // references from r-values.
+  RE(const RE& other) { Init(other.pattern()); }
+
+  // Constructs an RE from a string.
+  RE(const ::std::string& regex) { Init(regex.c_str()); }  // NOLINT
+
+#if GTEST_HAS_GLOBAL_STRING
+
+  RE(const ::string& regex) { Init(regex.c_str()); }  // NOLINT
+
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+  RE(const char* regex) { Init(regex); }  // NOLINT
+  ~RE();
+
+  // Returns the string representation of the regex.
+  const char* pattern() const { return pattern_; }
+
+  // FullMatch(str, re) returns true iff regular expression re matches
+  // the entire str.
+  // PartialMatch(str, re) returns true iff regular expression re
+  // matches a substring of str (including str itself).
+  //
+  // TODO(wan@google.com): make FullMatch() and PartialMatch() work
+  // when str contains NUL characters.
+  static bool FullMatch(const ::std::string& str, const RE& re) {
+    return FullMatch(str.c_str(), re);
+  }
+  static bool PartialMatch(const ::std::string& str, const RE& re) {
+    return PartialMatch(str.c_str(), re);
+  }
+
+#if GTEST_HAS_GLOBAL_STRING
+
+  static bool FullMatch(const ::string& str, const RE& re) {
+    return FullMatch(str.c_str(), re);
+  }
+  static bool PartialMatch(const ::string& str, const RE& re) {
+    return PartialMatch(str.c_str(), re);
+  }
+
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+  static bool FullMatch(const char* str, const RE& re);
+  static bool PartialMatch(const char* str, const RE& re);
+
+ private:
+  void Init(const char* regex);
+
+  // We use a const char* instead of an std::string, as Google Test used to be
+  // used where std::string is not available.  TODO(wan@google.com): change to
+  // std::string.
+  const char* pattern_;
+  bool is_valid_;
+
+#if GTEST_USES_POSIX_RE
+
+  regex_t full_regex_;     // For FullMatch().
+  regex_t partial_regex_;  // For PartialMatch().
+
+#else  // GTEST_USES_SIMPLE_RE
+
+  const char* full_pattern_;  // For FullMatch();
+
+#endif
+
+  GTEST_DISALLOW_ASSIGN_(RE);
+};
+
+// Formats a source file path and a line number as they would appear
+// in an error message from the compiler used to compile this code.
+GTEST_API_ ::std::string FormatFileLocation(const char* file, int line);
+
+// Formats a file location for compiler-independent XML output.
+// Although this function is not platform dependent, we put it next to
+// FormatFileLocation in order to contrast the two functions.
+GTEST_API_ ::std::string FormatCompilerIndependentFileLocation(const char* file,
+                                                               int line);
+
+// Defines logging utilities:
+//   GTEST_LOG_(severity) - logs messages at the specified severity level. The
+//                          message itself is streamed into the macro.
+//   LogToStderr()  - directs all log messages to stderr.
+//   FlushInfoLog() - flushes informational log messages.
+
+enum GTestLogSeverity {
+  GTEST_INFO,
+  GTEST_WARNING,
+  GTEST_ERROR,
+  GTEST_FATAL
+};
+
+// Formats log entry severity, provides a stream object for streaming the
+// log message, and terminates the message with a newline when going out of
+// scope.
+class GTEST_API_ GTestLog {
+ public:
+  GTestLog(GTestLogSeverity severity, const char* file, int line);
+
+  // Flushes the buffers and, if severity is GTEST_FATAL, aborts the program.
+  ~GTestLog();
+
+  ::std::ostream& GetStream() { return ::std::cerr; }
+
+ private:
+  const GTestLogSeverity severity_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(GTestLog);
+};
+
+#if !defined(GTEST_LOG_)
+
+# define GTEST_LOG_(severity) \
+    ::testing::internal::GTestLog(::testing::internal::GTEST_##severity, \
+                                  __FILE__, __LINE__).GetStream()
+
+inline void LogToStderr() {}
+inline void FlushInfoLog() { fflush(NULL); }
+
+#endif  // !defined(GTEST_LOG_)
+
+#if !defined(GTEST_CHECK_)
+// INTERNAL IMPLEMENTATION - DO NOT USE.
+//
+// GTEST_CHECK_ is an all-mode assert. It aborts the program if the condition
+// is not satisfied.
+//  Synopsys:
+//    GTEST_CHECK_(boolean_condition);
+//     or
+//    GTEST_CHECK_(boolean_condition) << "Additional message";
+//
+//    This checks the condition and if the condition is not satisfied
+//    it prints message about the condition violation, including the
+//    condition itself, plus additional message streamed into it, if any,
+//    and then it aborts the program. It aborts the program irrespective of
+//    whether it is built in the debug mode or not.
+# define GTEST_CHECK_(condition) \
+    GTEST_AMBIGUOUS_ELSE_BLOCKER_ \
+    if (::testing::internal::IsTrue(condition)) \
+      ; \
+    else \
+      GTEST_LOG_(FATAL) << "Condition " #condition " failed. "
+#endif  // !defined(GTEST_CHECK_)
+
+// An all-mode assert to verify that the given POSIX-style function
+// call returns 0 (indicating success).  Known limitation: this
+// doesn't expand to a balanced 'if' statement, so enclose the macro
+// in {} if you need to use it as the only statement in an 'if'
+// branch.
+#define GTEST_CHECK_POSIX_SUCCESS_(posix_call) \
+  if (const int gtest_error = (posix_call)) \
+    GTEST_LOG_(FATAL) << #posix_call << "failed with error " \
+                      << gtest_error
+
+#if GTEST_HAS_STD_MOVE_
+using std::move;
+#else  // GTEST_HAS_STD_MOVE_
+template <typename T>
+const T& move(const T& t) {
+  return t;
+}
+#endif  // GTEST_HAS_STD_MOVE_
+
+// INTERNAL IMPLEMENTATION - DO NOT USE IN USER CODE.
+//
+// Use ImplicitCast_ as a safe version of static_cast for upcasting in
+// the type hierarchy (e.g. casting a Foo* to a SuperclassOfFoo* or a
+// const Foo*).  When you use ImplicitCast_, the compiler checks that
+// the cast is safe.  Such explicit ImplicitCast_s are necessary in
+// surprisingly many situations where C++ demands an exact type match
+// instead of an argument type convertable to a target type.
+//
+// The syntax for using ImplicitCast_ is the same as for static_cast:
+//
+//   ImplicitCast_<ToType>(expr)
+//
+// ImplicitCast_ would have been part of the C++ standard library,
+// but the proposal was submitted too late.  It will probably make
+// its way into the language in the future.
+//
+// This relatively ugly name is intentional. It prevents clashes with
+// similar functions users may have (e.g., implicit_cast). The internal
+// namespace alone is not enough because the function can be found by ADL.
+template<typename To>
+inline To ImplicitCast_(To x) { return x; }
+
+// When you upcast (that is, cast a pointer from type Foo to type
+// SuperclassOfFoo), it's fine to use ImplicitCast_<>, since upcasts
+// always succeed.  When you downcast (that is, cast a pointer from
+// type Foo to type SubclassOfFoo), static_cast<> isn't safe, because
+// how do you know the pointer is really of type SubclassOfFoo?  It
+// could be a bare Foo, or of type DifferentSubclassOfFoo.  Thus,
+// when you downcast, you should use this macro.  In debug mode, we
+// use dynamic_cast<> to double-check the downcast is legal (we die
+// if it's not).  In normal mode, we do the efficient static_cast<>
+// instead.  Thus, it's important to test in debug mode to make sure
+// the cast is legal!
+//    This is the only place in the code we should use dynamic_cast<>.
+// In particular, you SHOULDN'T be using dynamic_cast<> in order to
+// do RTTI (eg code like this:
+//    if (dynamic_cast<Subclass1>(foo)) HandleASubclass1Object(foo);
+//    if (dynamic_cast<Subclass2>(foo)) HandleASubclass2Object(foo);
+// You should design the code some other way not to need this.
+//
+// This relatively ugly name is intentional. It prevents clashes with
+// similar functions users may have (e.g., down_cast). The internal
+// namespace alone is not enough because the function can be found by ADL.
+template<typename To, typename From>  // use like this: DownCast_<T*>(foo);
+inline To DownCast_(From* f) {  // so we only accept pointers
+  // Ensures that To is a sub-type of From *.  This test is here only
+  // for compile-time type checking, and has no overhead in an
+  // optimized build at run-time, as it will be optimized away
+  // completely.
+  GTEST_INTENTIONAL_CONST_COND_PUSH_()
+  if (false) {
+  GTEST_INTENTIONAL_CONST_COND_POP_()
+    const To to = NULL;
+    ::testing::internal::ImplicitCast_<From*>(to);
+  }
+
+#if GTEST_HAS_RTTI
+  // RTTI: debug mode only!
+  GTEST_CHECK_(f == NULL || dynamic_cast<To>(f) != NULL);
+#endif
+  return static_cast<To>(f);
+}
+
+// Downcasts the pointer of type Base to Derived.
+// Derived must be a subclass of Base. The parameter MUST
+// point to a class of type Derived, not any subclass of it.
+// When RTTI is available, the function performs a runtime
+// check to enforce this.
+template <class Derived, class Base>
+Derived* CheckedDowncastToActualType(Base* base) {
+#if GTEST_HAS_RTTI
+  GTEST_CHECK_(typeid(*base) == typeid(Derived));
+#endif
+
+#if GTEST_HAS_DOWNCAST_
+  return ::down_cast<Derived*>(base);
+#elif GTEST_HAS_RTTI
+  return dynamic_cast<Derived*>(base);  // NOLINT
+#else
+  return static_cast<Derived*>(base);  // Poor man's downcast.
+#endif
+}
+
+#if GTEST_HAS_STREAM_REDIRECTION
+
+// Defines the stderr capturer:
+//   CaptureStdout     - starts capturing stdout.
+//   GetCapturedStdout - stops capturing stdout and returns the captured string.
+//   CaptureStderr     - starts capturing stderr.
+//   GetCapturedStderr - stops capturing stderr and returns the captured string.
+//
+GTEST_API_ void CaptureStdout();
+GTEST_API_ std::string GetCapturedStdout();
+GTEST_API_ void CaptureStderr();
+GTEST_API_ std::string GetCapturedStderr();
+
+#endif  // GTEST_HAS_STREAM_REDIRECTION
+
+// Returns a path to temporary directory.
+GTEST_API_ std::string TempDir();
+
+// Returns the size (in bytes) of a file.
+GTEST_API_ size_t GetFileSize(FILE* file);
+
+// Reads the entire content of a file as a string.
+GTEST_API_ std::string ReadEntireFile(FILE* file);
+
+// All command line arguments.
+GTEST_API_ const ::std::vector<testing::internal::string>& GetArgvs();
+
+#if GTEST_HAS_DEATH_TEST
+
+const ::std::vector<testing::internal::string>& GetInjectableArgvs();
+void SetInjectableArgvs(const ::std::vector<testing::internal::string>*
+                             new_argvs);
+
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+// Defines synchronization primitives.
+#if GTEST_IS_THREADSAFE
+# if GTEST_HAS_PTHREAD
+// Sleeps for (roughly) n milliseconds.  This function is only for testing
+// Google Test's own constructs.  Don't use it in user tests, either
+// directly or indirectly.
+inline void SleepMilliseconds(int n) {
+  const timespec time = {
+    0,                  // 0 seconds.
+    n * 1000L * 1000L,  // And n ms.
+  };
+  nanosleep(&time, NULL);
+}
+# endif  // GTEST_HAS_PTHREAD
+
+# if GTEST_HAS_NOTIFICATION_
+// Notification has already been imported into the namespace.
+// Nothing to do here.
+
+# elif GTEST_HAS_PTHREAD
+// Allows a controller thread to pause execution of newly created
+// threads until notified.  Instances of this class must be created
+// and destroyed in the controller thread.
+//
+// This class is only for testing Google Test's own constructs. Do not
+// use it in user tests, either directly or indirectly.
+class Notification {
+ public:
+  Notification() : notified_(false) {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_init(&mutex_, NULL));
+  }
+  ~Notification() {
+    pthread_mutex_destroy(&mutex_);
+  }
+
+  // Notifies all threads created with this notification to start. Must
+  // be called from the controller thread.
+  void Notify() {
+    pthread_mutex_lock(&mutex_);
+    notified_ = true;
+    pthread_mutex_unlock(&mutex_);
+  }
+
+  // Blocks until the controller thread notifies. Must be called from a test
+  // thread.
+  void WaitForNotification() {
+    for (;;) {
+      pthread_mutex_lock(&mutex_);
+      const bool notified = notified_;
+      pthread_mutex_unlock(&mutex_);
+      if (notified)
+        break;
+      SleepMilliseconds(10);
+    }
+  }
+
+ private:
+  pthread_mutex_t mutex_;
+  bool notified_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Notification);
+};
+
+# elif GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_PHONE && !GTEST_OS_WINDOWS_RT
+
+GTEST_API_ void SleepMilliseconds(int n);
+
+// Provides leak-safe Windows kernel handle ownership.
+// Used in death tests and in threading support.
+class GTEST_API_ AutoHandle {
+ public:
+  // Assume that Win32 HANDLE type is equivalent to void*. Doing so allows us to
+  // avoid including <windows.h> in this header file. Including <windows.h> is
+  // undesirable because it defines a lot of symbols and macros that tend to
+  // conflict with client code. This assumption is verified by
+  // WindowsTypesTest.HANDLEIsVoidStar.
+  typedef void* Handle;
+  AutoHandle();
+  explicit AutoHandle(Handle handle);
+
+  ~AutoHandle();
+
+  Handle Get() const;
+  void Reset();
+  void Reset(Handle handle);
+
+ private:
+  // Returns true iff the handle is a valid handle object that can be closed.
+  bool IsCloseable() const;
+
+  Handle handle_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(AutoHandle);
+};
+
+// Allows a controller thread to pause execution of newly created
+// threads until notified.  Instances of this class must be created
+// and destroyed in the controller thread.
+//
+// This class is only for testing Google Test's own constructs. Do not
+// use it in user tests, either directly or indirectly.
+class GTEST_API_ Notification {
+ public:
+  Notification();
+  void Notify();
+  void WaitForNotification();
+
+ private:
+  AutoHandle event_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Notification);
+};
+# endif  // GTEST_HAS_NOTIFICATION_
+
+// On MinGW, we can have both GTEST_OS_WINDOWS and GTEST_HAS_PTHREAD
+// defined, but we don't want to use MinGW's pthreads implementation, which
+// has conformance problems with some versions of the POSIX standard.
+# if GTEST_HAS_PTHREAD && !GTEST_OS_WINDOWS_MINGW
+
+// As a C-function, ThreadFuncWithCLinkage cannot be templated itself.
+// Consequently, it cannot select a correct instantiation of ThreadWithParam
+// in order to call its Run(). Introducing ThreadWithParamBase as a
+// non-templated base class for ThreadWithParam allows us to bypass this
+// problem.
+class ThreadWithParamBase {
+ public:
+  virtual ~ThreadWithParamBase() {}
+  virtual void Run() = 0;
+};
+
+// pthread_create() accepts a pointer to a function type with the C linkage.
+// According to the Standard (7.5/1), function types with different linkages
+// are different even if they are otherwise identical.  Some compilers (for
+// example, SunStudio) treat them as different types.  Since class methods
+// cannot be defined with C-linkage we need to define a free C-function to
+// pass into pthread_create().
+extern "C" inline void* ThreadFuncWithCLinkage(void* thread) {
+  static_cast<ThreadWithParamBase*>(thread)->Run();
+  return NULL;
+}
+
+// Helper class for testing Google Test's multi-threading constructs.
+// To use it, write:
+//
+//   void ThreadFunc(int param) { /* Do things with param */ }
+//   Notification thread_can_start;
+//   ...
+//   // The thread_can_start parameter is optional; you can supply NULL.
+//   ThreadWithParam<int> thread(&ThreadFunc, 5, &thread_can_start);
+//   thread_can_start.Notify();
+//
+// These classes are only for testing Google Test's own constructs. Do
+// not use them in user tests, either directly or indirectly.
+template <typename T>
+class ThreadWithParam : public ThreadWithParamBase {
+ public:
+  typedef void UserThreadFunc(T);
+
+  ThreadWithParam(UserThreadFunc* func, T param, Notification* thread_can_start)
+      : func_(func),
+        param_(param),
+        thread_can_start_(thread_can_start),
+        finished_(false) {
+    ThreadWithParamBase* const base = this;
+    // The thread can be created only after all fields except thread_
+    // have been initialized.
+    GTEST_CHECK_POSIX_SUCCESS_(
+        pthread_create(&thread_, 0, &ThreadFuncWithCLinkage, base));
+  }
+  ~ThreadWithParam() { Join(); }
+
+  void Join() {
+    if (!finished_) {
+      GTEST_CHECK_POSIX_SUCCESS_(pthread_join(thread_, 0));
+      finished_ = true;
+    }
+  }
+
+  virtual void Run() {
+    if (thread_can_start_ != NULL)
+      thread_can_start_->WaitForNotification();
+    func_(param_);
+  }
+
+ private:
+  UserThreadFunc* const func_;  // User-supplied thread function.
+  const T param_;  // User-supplied parameter to the thread function.
+  // When non-NULL, used to block execution until the controller thread
+  // notifies.
+  Notification* const thread_can_start_;
+  bool finished_;  // true iff we know that the thread function has finished.
+  pthread_t thread_;  // The native thread object.
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ThreadWithParam);
+};
+# endif  // !GTEST_OS_WINDOWS && GTEST_HAS_PTHREAD ||
+         // GTEST_HAS_MUTEX_AND_THREAD_LOCAL_
+
+# if GTEST_HAS_MUTEX_AND_THREAD_LOCAL_
+// Mutex and ThreadLocal have already been imported into the namespace.
+// Nothing to do here.
+
+# elif GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_PHONE && !GTEST_OS_WINDOWS_RT
+
+// Mutex implements mutex on Windows platforms.  It is used in conjunction
+// with class MutexLock:
+//
+//   Mutex mutex;
+//   ...
+//   MutexLock lock(&mutex);  // Acquires the mutex and releases it at the
+//                            // end of the current scope.
+//
+// A static Mutex *must* be defined or declared using one of the following
+// macros:
+//   GTEST_DEFINE_STATIC_MUTEX_(g_some_mutex);
+//   GTEST_DECLARE_STATIC_MUTEX_(g_some_mutex);
+//
+// (A non-static Mutex is defined/declared in the usual way).
+class GTEST_API_ Mutex {
+ public:
+  enum MutexType { kStatic = 0, kDynamic = 1 };
+  // We rely on kStaticMutex being 0 as it is to what the linker initializes
+  // type_ in static mutexes.  critical_section_ will be initialized lazily
+  // in ThreadSafeLazyInit().
+  enum StaticConstructorSelector { kStaticMutex = 0 };
+
+  // This constructor intentionally does nothing.  It relies on type_ being
+  // statically initialized to 0 (effectively setting it to kStatic) and on
+  // ThreadSafeLazyInit() to lazily initialize the rest of the members.
+  explicit Mutex(StaticConstructorSelector /*dummy*/) {}
+
+  Mutex();
+  ~Mutex();
+
+  void Lock();
+
+  void Unlock();
+
+  // Does nothing if the current thread holds the mutex. Otherwise, crashes
+  // with high probability.
+  void AssertHeld();
+
+ private:
+  // Initializes owner_thread_id_ and critical_section_ in static mutexes.
+  void ThreadSafeLazyInit();
+
+  // Per http://blogs.msdn.com/b/oldnewthing/archive/2004/02/23/78395.aspx,
+  // we assume that 0 is an invalid value for thread IDs.
+  unsigned int owner_thread_id_;
+
+  // For static mutexes, we rely on these members being initialized to zeros
+  // by the linker.
+  MutexType type_;
+  long critical_section_init_phase_;  // NOLINT
+  _RTL_CRITICAL_SECTION* critical_section_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Mutex);
+};
+
+# define GTEST_DECLARE_STATIC_MUTEX_(mutex) \
+    extern ::testing::internal::Mutex mutex
+
+# define GTEST_DEFINE_STATIC_MUTEX_(mutex) \
+    ::testing::internal::Mutex mutex(::testing::internal::Mutex::kStaticMutex)
+
+// We cannot name this class MutexLock because the ctor declaration would
+// conflict with a macro named MutexLock, which is defined on some
+// platforms. That macro is used as a defensive measure to prevent against
+// inadvertent misuses of MutexLock like "MutexLock(&mu)" rather than
+// "MutexLock l(&mu)".  Hence the typedef trick below.
+class GTestMutexLock {
+ public:
+  explicit GTestMutexLock(Mutex* mutex)
+      : mutex_(mutex) { mutex_->Lock(); }
+
+  ~GTestMutexLock() { mutex_->Unlock(); }
+
+ private:
+  Mutex* const mutex_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(GTestMutexLock);
+};
+
+typedef GTestMutexLock MutexLock;
+
+// Base class for ValueHolder<T>.  Allows a caller to hold and delete a value
+// without knowing its type.
+class ThreadLocalValueHolderBase {
+ public:
+  virtual ~ThreadLocalValueHolderBase() {}
+};
+
+// Provides a way for a thread to send notifications to a ThreadLocal
+// regardless of its parameter type.
+class ThreadLocalBase {
+ public:
+  // Creates a new ValueHolder<T> object holding a default value passed to
+  // this ThreadLocal<T>'s constructor and returns it.  It is the caller's
+  // responsibility not to call this when the ThreadLocal<T> instance already
+  // has a value on the current thread.
+  virtual ThreadLocalValueHolderBase* NewValueForCurrentThread() const = 0;
+
+ protected:
+  ThreadLocalBase() {}
+  virtual ~ThreadLocalBase() {}
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ThreadLocalBase);
+};
+
+// Maps a thread to a set of ThreadLocals that have values instantiated on that
+// thread and notifies them when the thread exits.  A ThreadLocal instance is
+// expected to persist until all threads it has values on have terminated.
+class GTEST_API_ ThreadLocalRegistry {
+ public:
+  // Registers thread_local_instance as having value on the current thread.
+  // Returns a value that can be used to identify the thread from other threads.
+  static ThreadLocalValueHolderBase* GetValueOnCurrentThread(
+      const ThreadLocalBase* thread_local_instance);
+
+  // Invoked when a ThreadLocal instance is destroyed.
+  static void OnThreadLocalDestroyed(
+      const ThreadLocalBase* thread_local_instance);
+};
+
+class GTEST_API_ ThreadWithParamBase {
+ public:
+  void Join();
+
+ protected:
+  class Runnable {
+   public:
+    virtual ~Runnable() {}
+    virtual void Run() = 0;
+  };
+
+  ThreadWithParamBase(Runnable *runnable, Notification* thread_can_start);
+  virtual ~ThreadWithParamBase();
+
+ private:
+  AutoHandle thread_;
+};
+
+// Helper class for testing Google Test's multi-threading constructs.
+template <typename T>
+class ThreadWithParam : public ThreadWithParamBase {
+ public:
+  typedef void UserThreadFunc(T);
+
+  ThreadWithParam(UserThreadFunc* func, T param, Notification* thread_can_start)
+      : ThreadWithParamBase(new RunnableImpl(func, param), thread_can_start) {
+  }
+  virtual ~ThreadWithParam() {}
+
+ private:
+  class RunnableImpl : public Runnable {
+   public:
+    RunnableImpl(UserThreadFunc* func, T param)
+        : func_(func),
+          param_(param) {
+    }
+    virtual ~RunnableImpl() {}
+    virtual void Run() {
+      func_(param_);
+    }
+
+   private:
+    UserThreadFunc* const func_;
+    const T param_;
+
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(RunnableImpl);
+  };
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ThreadWithParam);
+};
+
+// Implements thread-local storage on Windows systems.
+//
+//   // Thread 1
+//   ThreadLocal<int> tl(100);  // 100 is the default value for each thread.
+//
+//   // Thread 2
+//   tl.set(150);  // Changes the value for thread 2 only.
+//   EXPECT_EQ(150, tl.get());
+//
+//   // Thread 1
+//   EXPECT_EQ(100, tl.get());  // In thread 1, tl has the original value.
+//   tl.set(200);
+//   EXPECT_EQ(200, tl.get());
+//
+// The template type argument T must have a public copy constructor.
+// In addition, the default ThreadLocal constructor requires T to have
+// a public default constructor.
+//
+// The users of a TheadLocal instance have to make sure that all but one
+// threads (including the main one) using that instance have exited before
+// destroying it. Otherwise, the per-thread objects managed for them by the
+// ThreadLocal instance are not guaranteed to be destroyed on all platforms.
+//
+// Google Test only uses global ThreadLocal objects.  That means they
+// will die after main() has returned.  Therefore, no per-thread
+// object managed by Google Test will be leaked as long as all threads
+// using Google Test have exited when main() returns.
+template <typename T>
+class ThreadLocal : public ThreadLocalBase {
+ public:
+  ThreadLocal() : default_factory_(new DefaultValueHolderFactory()) {}
+  explicit ThreadLocal(const T& value)
+      : default_factory_(new InstanceValueHolderFactory(value)) {}
+
+  ~ThreadLocal() { ThreadLocalRegistry::OnThreadLocalDestroyed(this); }
+
+  T* pointer() { return GetOrCreateValue(); }
+  const T* pointer() const { return GetOrCreateValue(); }
+  const T& get() const { return *pointer(); }
+  void set(const T& value) { *pointer() = value; }
+
+ private:
+  // Holds a value of T.  Can be deleted via its base class without the caller
+  // knowing the type of T.
+  class ValueHolder : public ThreadLocalValueHolderBase {
+   public:
+    ValueHolder() : value_() {}
+    explicit ValueHolder(const T& value) : value_(value) {}
+
+    T* pointer() { return &value_; }
+
+   private:
+    T value_;
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(ValueHolder);
+  };
+
+
+  T* GetOrCreateValue() const {
+    return static_cast<ValueHolder*>(
+        ThreadLocalRegistry::GetValueOnCurrentThread(this))->pointer();
+  }
+
+  virtual ThreadLocalValueHolderBase* NewValueForCurrentThread() const {
+    return default_factory_->MakeNewHolder();
+  }
+
+  class ValueHolderFactory {
+   public:
+    ValueHolderFactory() {}
+    virtual ~ValueHolderFactory() {}
+    virtual ValueHolder* MakeNewHolder() const = 0;
+
+   private:
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(ValueHolderFactory);
+  };
+
+  class DefaultValueHolderFactory : public ValueHolderFactory {
+   public:
+    DefaultValueHolderFactory() {}
+    virtual ValueHolder* MakeNewHolder() const { return new ValueHolder(); }
+
+   private:
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(DefaultValueHolderFactory);
+  };
+
+  class InstanceValueHolderFactory : public ValueHolderFactory {
+   public:
+    explicit InstanceValueHolderFactory(const T& value) : value_(value) {}
+    virtual ValueHolder* MakeNewHolder() const {
+      return new ValueHolder(value_);
+    }
+
+   private:
+    const T value_;  // The value for each thread.
+
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(InstanceValueHolderFactory);
+  };
+
+  scoped_ptr<ValueHolderFactory> default_factory_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ThreadLocal);
+};
+
+# elif GTEST_HAS_PTHREAD
+
+// MutexBase and Mutex implement mutex on pthreads-based platforms.
+class MutexBase {
+ public:
+  // Acquires this mutex.
+  void Lock() {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_lock(&mutex_));
+    owner_ = pthread_self();
+    has_owner_ = true;
+  }
+
+  // Releases this mutex.
+  void Unlock() {
+    // Since the lock is being released the owner_ field should no longer be
+    // considered valid. We don't protect writing to has_owner_ here, as it's
+    // the caller's responsibility to ensure that the current thread holds the
+    // mutex when this is called.
+    has_owner_ = false;
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_unlock(&mutex_));
+  }
+
+  // Does nothing if the current thread holds the mutex. Otherwise, crashes
+  // with high probability.
+  void AssertHeld() const {
+    GTEST_CHECK_(has_owner_ && pthread_equal(owner_, pthread_self()))
+        << "The current thread is not holding the mutex @" << this;
+  }
+
+  // A static mutex may be used before main() is entered.  It may even
+  // be used before the dynamic initialization stage.  Therefore we
+  // must be able to initialize a static mutex object at link time.
+  // This means MutexBase has to be a POD and its member variables
+  // have to be public.
+ public:
+  pthread_mutex_t mutex_;  // The underlying pthread mutex.
+  // has_owner_ indicates whether the owner_ field below contains a valid thread
+  // ID and is therefore safe to inspect (e.g., to use in pthread_equal()). All
+  // accesses to the owner_ field should be protected by a check of this field.
+  // An alternative might be to memset() owner_ to all zeros, but there's no
+  // guarantee that a zero'd pthread_t is necessarily invalid or even different
+  // from pthread_self().
+  bool has_owner_;
+  pthread_t owner_;  // The thread holding the mutex.
+};
+
+// Forward-declares a static mutex.
+#  define GTEST_DECLARE_STATIC_MUTEX_(mutex) \
+     extern ::testing::internal::MutexBase mutex
+
+// Defines and statically (i.e. at link time) initializes a static mutex.
+#  define GTEST_DEFINE_STATIC_MUTEX_(mutex) \
+     ::testing::internal::MutexBase mutex = { PTHREAD_MUTEX_INITIALIZER, false, pthread_t() }
+
+// The Mutex class can only be used for mutexes created at runtime. It
+// shares its API with MutexBase otherwise.
+class Mutex : public MutexBase {
+ public:
+  Mutex() {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_init(&mutex_, NULL));
+    has_owner_ = false;
+  }
+  ~Mutex() {
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_mutex_destroy(&mutex_));
+  }
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(Mutex);
+};
+
+// We cannot name this class MutexLock because the ctor declaration would
+// conflict with a macro named MutexLock, which is defined on some
+// platforms. That macro is used as a defensive measure to prevent against
+// inadvertent misuses of MutexLock like "MutexLock(&mu)" rather than
+// "MutexLock l(&mu)".  Hence the typedef trick below.
+class GTestMutexLock {
+ public:
+  explicit GTestMutexLock(MutexBase* mutex)
+      : mutex_(mutex) { mutex_->Lock(); }
+
+  ~GTestMutexLock() { mutex_->Unlock(); }
+
+ private:
+  MutexBase* const mutex_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(GTestMutexLock);
+};
+
+typedef GTestMutexLock MutexLock;
+
+// Helpers for ThreadLocal.
+
+// pthread_key_create() requires DeleteThreadLocalValue() to have
+// C-linkage.  Therefore it cannot be templatized to access
+// ThreadLocal<T>.  Hence the need for class
+// ThreadLocalValueHolderBase.
+class ThreadLocalValueHolderBase {
+ public:
+  virtual ~ThreadLocalValueHolderBase() {}
+};
+
+// Called by pthread to delete thread-local data stored by
+// pthread_setspecific().
+extern "C" inline void DeleteThreadLocalValue(void* value_holder) {
+  delete static_cast<ThreadLocalValueHolderBase*>(value_holder);
+}
+
+// Implements thread-local storage on pthreads-based systems.
+template <typename T>
+class ThreadLocal {
+ public:
+  ThreadLocal()
+      : key_(CreateKey()), default_factory_(new DefaultValueHolderFactory()) {}
+  explicit ThreadLocal(const T& value)
+      : key_(CreateKey()),
+        default_factory_(new InstanceValueHolderFactory(value)) {}
+
+  ~ThreadLocal() {
+    // Destroys the managed object for the current thread, if any.
+    DeleteThreadLocalValue(pthread_getspecific(key_));
+
+    // Releases resources associated with the key.  This will *not*
+    // delete managed objects for other threads.
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_key_delete(key_));
+  }
+
+  T* pointer() { return GetOrCreateValue(); }
+  const T* pointer() const { return GetOrCreateValue(); }
+  const T& get() const { return *pointer(); }
+  void set(const T& value) { *pointer() = value; }
+
+ private:
+  // Holds a value of type T.
+  class ValueHolder : public ThreadLocalValueHolderBase {
+   public:
+    ValueHolder() : value_() {}
+    explicit ValueHolder(const T& value) : value_(value) {}
+
+    T* pointer() { return &value_; }
+
+   private:
+    T value_;
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(ValueHolder);
+  };
+
+  static pthread_key_t CreateKey() {
+    pthread_key_t key;
+    // When a thread exits, DeleteThreadLocalValue() will be called on
+    // the object managed for that thread.
+    GTEST_CHECK_POSIX_SUCCESS_(
+        pthread_key_create(&key, &DeleteThreadLocalValue));
+    return key;
+  }
+
+  T* GetOrCreateValue() const {
+    ThreadLocalValueHolderBase* const holder =
+        static_cast<ThreadLocalValueHolderBase*>(pthread_getspecific(key_));
+    if (holder != NULL) {
+      return CheckedDowncastToActualType<ValueHolder>(holder)->pointer();
+    }
+
+    ValueHolder* const new_holder = default_factory_->MakeNewHolder();
+    ThreadLocalValueHolderBase* const holder_base = new_holder;
+    GTEST_CHECK_POSIX_SUCCESS_(pthread_setspecific(key_, holder_base));
+    return new_holder->pointer();
+  }
+
+  class ValueHolderFactory {
+   public:
+    ValueHolderFactory() {}
+    virtual ~ValueHolderFactory() {}
+    virtual ValueHolder* MakeNewHolder() const = 0;
+
+   private:
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(ValueHolderFactory);
+  };
+
+  class DefaultValueHolderFactory : public ValueHolderFactory {
+   public:
+    DefaultValueHolderFactory() {}
+    virtual ValueHolder* MakeNewHolder() const { return new ValueHolder(); }
+
+   private:
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(DefaultValueHolderFactory);
+  };
+
+  class InstanceValueHolderFactory : public ValueHolderFactory {
+   public:
+    explicit InstanceValueHolderFactory(const T& value) : value_(value) {}
+    virtual ValueHolder* MakeNewHolder() const {
+      return new ValueHolder(value_);
+    }
+
+   private:
+    const T value_;  // The value for each thread.
+
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(InstanceValueHolderFactory);
+  };
+
+  // A key pthreads uses for looking up per-thread values.
+  const pthread_key_t key_;
+  scoped_ptr<ValueHolderFactory> default_factory_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ThreadLocal);
+};
+
+# endif  // GTEST_HAS_MUTEX_AND_THREAD_LOCAL_
+
+#else  // GTEST_IS_THREADSAFE
+
+// A dummy implementation of synchronization primitives (mutex, lock,
+// and thread-local variable).  Necessary for compiling Google Test where
+// mutex is not supported - using Google Test in multiple threads is not
+// supported on such platforms.
+
+class Mutex {
+ public:
+  Mutex() {}
+  void Lock() {}
+  void Unlock() {}
+  void AssertHeld() const {}
+};
+
+# define GTEST_DECLARE_STATIC_MUTEX_(mutex) \
+  extern ::testing::internal::Mutex mutex
+
+# define GTEST_DEFINE_STATIC_MUTEX_(mutex) ::testing::internal::Mutex mutex
+
+// We cannot name this class MutexLock because the ctor declaration would
+// conflict with a macro named MutexLock, which is defined on some
+// platforms. That macro is used as a defensive measure to prevent against
+// inadvertent misuses of MutexLock like "MutexLock(&mu)" rather than
+// "MutexLock l(&mu)".  Hence the typedef trick below.
+class GTestMutexLock {
+ public:
+  explicit GTestMutexLock(Mutex*) {}  // NOLINT
+};
+
+typedef GTestMutexLock MutexLock;
+
+template <typename T>
+class ThreadLocal {
+ public:
+  ThreadLocal() : value_() {}
+  explicit ThreadLocal(const T& value) : value_(value) {}
+  T* pointer() { return &value_; }
+  const T* pointer() const { return &value_; }
+  const T& get() const { return value_; }
+  void set(const T& value) { value_ = value; }
+ private:
+  T value_;
+};
+
+#endif  // GTEST_IS_THREADSAFE
+
+// Returns the number of threads running in the process, or 0 to indicate that
+// we cannot detect it.
+GTEST_API_ size_t GetThreadCount();
+
+// Passing non-POD classes through ellipsis (...) crashes the ARM
+// compiler and generates a warning in Sun Studio.  The Nokia Symbian
+// and the IBM XL C/C++ compiler try to instantiate a copy constructor
+// for objects passed through ellipsis (...), failing for uncopyable
+// objects.  We define this to ensure that only POD is passed through
+// ellipsis on these systems.
+#if defined(__SYMBIAN32__) || defined(__IBMCPP__) || defined(__SUNPRO_CC)
+// We lose support for NULL detection where the compiler doesn't like
+// passing non-POD classes through ellipsis (...).
+# define GTEST_ELLIPSIS_NEEDS_POD_ 1
+#else
+# define GTEST_CAN_COMPARE_NULL 1
+#endif
+
+// The Nokia Symbian and IBM XL C/C++ compilers cannot decide between
+// const T& and const T* in a function template.  These compilers
+// _can_ decide between class template specializations for T and T*,
+// so a tr1::type_traits-like is_pointer works.
+#if defined(__SYMBIAN32__) || defined(__IBMCPP__)
+# define GTEST_NEEDS_IS_POINTER_ 1
+#endif
+
+template <bool bool_value>
+struct bool_constant {
+  typedef bool_constant<bool_value> type;
+  static const bool value = bool_value;
+};
+template <bool bool_value> const bool bool_constant<bool_value>::value;
+
+typedef bool_constant<false> false_type;
+typedef bool_constant<true> true_type;
+
+template <typename T>
+struct is_pointer : public false_type {};
+
+template <typename T>
+struct is_pointer<T*> : public true_type {};
+
+template <typename Iterator>
+struct IteratorTraits {
+  typedef typename Iterator::value_type value_type;
+};
+
+template <typename T>
+struct IteratorTraits<T*> {
+  typedef T value_type;
+};
+
+template <typename T>
+struct IteratorTraits<const T*> {
+  typedef T value_type;
+};
+
+#if GTEST_OS_WINDOWS
+# define GTEST_PATH_SEP_ "\\"
+# define GTEST_HAS_ALT_PATH_SEP_ 1
+// The biggest signed integer type the compiler supports.
+typedef __int64 BiggestInt;
+#else
+# define GTEST_PATH_SEP_ "/"
+# define GTEST_HAS_ALT_PATH_SEP_ 0
+typedef long long BiggestInt;  // NOLINT
+#endif  // GTEST_OS_WINDOWS
+
+// Utilities for char.
+
+// isspace(int ch) and friends accept an unsigned char or EOF.  char
+// may be signed, depending on the compiler (or compiler flags).
+// Therefore we need to cast a char to unsigned char before calling
+// isspace(), etc.
+
+inline bool IsAlpha(char ch) {
+  return isalpha(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsAlNum(char ch) {
+  return isalnum(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsDigit(char ch) {
+  return isdigit(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsLower(char ch) {
+  return islower(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsSpace(char ch) {
+  return isspace(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsUpper(char ch) {
+  return isupper(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsXDigit(char ch) {
+  return isxdigit(static_cast<unsigned char>(ch)) != 0;
+}
+inline bool IsXDigit(wchar_t ch) {
+  const unsigned char low_byte = static_cast<unsigned char>(ch);
+  return ch == low_byte && isxdigit(low_byte) != 0;
+}
+
+inline char ToLower(char ch) {
+  return static_cast<char>(tolower(static_cast<unsigned char>(ch)));
+}
+inline char ToUpper(char ch) {
+  return static_cast<char>(toupper(static_cast<unsigned char>(ch)));
+}
+
+inline std::string StripTrailingSpaces(std::string str) {
+  std::string::iterator it = str.end();
+  while (it != str.begin() && IsSpace(*--it))
+    it = str.erase(it);
+  return str;
+}
+
+// The testing::internal::posix namespace holds wrappers for common
+// POSIX functions.  These wrappers hide the differences between
+// Windows/MSVC and POSIX systems.  Since some compilers define these
+// standard functions as macros, the wrapper cannot have the same name
+// as the wrapped function.
+
+namespace posix {
+
+// Functions with a different name on Windows.
+
+#if GTEST_OS_WINDOWS
+
+typedef struct _stat StatStruct;
+
+# ifdef __BORLANDC__
+inline int IsATTY(int fd) { return isatty(fd); }
+inline int StrCaseCmp(const char* s1, const char* s2) {
+  return stricmp(s1, s2);
+}
+inline char* StrDup(const char* src) { return strdup(src); }
+# else  // !__BORLANDC__
+#  if GTEST_OS_WINDOWS_MOBILE
+inline int IsATTY(int /* fd */) { return 0; }
+#  else
+inline int IsATTY(int fd) { return _isatty(fd); }
+#  endif  // GTEST_OS_WINDOWS_MOBILE
+inline int StrCaseCmp(const char* s1, const char* s2) {
+  return _stricmp(s1, s2);
+}
+inline char* StrDup(const char* src) { return _strdup(src); }
+# endif  // __BORLANDC__
+
+# if GTEST_OS_WINDOWS_MOBILE
+inline int FileNo(FILE* file) { return reinterpret_cast<int>(_fileno(file)); }
+// Stat(), RmDir(), and IsDir() are not needed on Windows CE at this
+// time and thus not defined there.
+# else
+inline int FileNo(FILE* file) { return _fileno(file); }
+inline int Stat(const char* path, StatStruct* buf) { return _stat(path, buf); }
+inline int RmDir(const char* dir) { return _rmdir(dir); }
+inline bool IsDir(const StatStruct& st) {
+  return (_S_IFDIR & st.st_mode) != 0;
+}
+# endif  // GTEST_OS_WINDOWS_MOBILE
+
+#else
+
+typedef struct stat StatStruct;
+
+inline int FileNo(FILE* file) { return fileno(file); }
+inline int IsATTY(int fd) { return isatty(fd); }
+inline int Stat(const char* path, StatStruct* buf) { return stat(path, buf); }
+inline int StrCaseCmp(const char* s1, const char* s2) {
+  return strcasecmp(s1, s2);
+}
+inline char* StrDup(const char* src) { return strdup(src); }
+inline int RmDir(const char* dir) { return rmdir(dir); }
+inline bool IsDir(const StatStruct& st) { return S_ISDIR(st.st_mode); }
+
+#endif  // GTEST_OS_WINDOWS
+
+// Functions deprecated by MSVC 8.0.
+
+GTEST_DISABLE_MSC_WARNINGS_PUSH_(4996 /* deprecated function */)
+
+inline const char* StrNCpy(char* dest, const char* src, size_t n) {
+  return strncpy(dest, src, n);
+}
+
+// ChDir(), FReopen(), FDOpen(), Read(), Write(), Close(), and
+// StrError() aren't needed on Windows CE at this time and thus not
+// defined there.
+
+#if !GTEST_OS_WINDOWS_MOBILE && !GTEST_OS_WINDOWS_PHONE && !GTEST_OS_WINDOWS_RT
+inline int ChDir(const char* dir) { return chdir(dir); }
+#endif
+inline FILE* FOpen(const char* path, const char* mode) {
+  return fopen(path, mode);
+}
+#if !GTEST_OS_WINDOWS_MOBILE
+inline FILE *FReopen(const char* path, const char* mode, FILE* stream) {
+  return freopen(path, mode, stream);
+}
+inline FILE* FDOpen(int fd, const char* mode) { return fdopen(fd, mode); }
+#endif
+inline int FClose(FILE* fp) { return fclose(fp); }
+#if !GTEST_OS_WINDOWS_MOBILE
+inline int Read(int fd, void* buf, unsigned int count) {
+  return static_cast<int>(read(fd, buf, count));
+}
+inline int Write(int fd, const void* buf, unsigned int count) {
+  return static_cast<int>(write(fd, buf, count));
+}
+inline int Close(int fd) { return close(fd); }
+inline const char* StrError(int errnum) { return strerror(errnum); }
+#endif
+inline const char* GetEnv(const char* name) {
+#if GTEST_OS_WINDOWS_MOBILE || GTEST_OS_WINDOWS_PHONE | GTEST_OS_WINDOWS_RT
+  // We are on Windows CE, which has no environment variables.
+  static_cast<void>(name);  // To prevent 'unused argument' warning.
+  return NULL;
+#elif defined(__BORLANDC__) || defined(__SunOS_5_8) || defined(__SunOS_5_9)
+  // Environment variables which we programmatically clear will be set to the
+  // empty string rather than unset (NULL).  Handle that case.
+  const char* const env = getenv(name);
+  return (env != NULL && env[0] != '\0') ? env : NULL;
+#else
+  return getenv(name);
+#endif
+}
+
+GTEST_DISABLE_MSC_WARNINGS_POP_()
+
+#if GTEST_OS_WINDOWS_MOBILE
+// Windows CE has no C library. The abort() function is used in
+// several places in Google Test. This implementation provides a reasonable
+// imitation of standard behaviour.
+void Abort();
+#else
+inline void Abort() { abort(); }
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+}  // namespace posix
+
+// MSVC "deprecates" snprintf and issues warnings wherever it is used.  In
+// order to avoid these warnings, we need to use _snprintf or _snprintf_s on
+// MSVC-based platforms.  We map the GTEST_SNPRINTF_ macro to the appropriate
+// function in order to achieve that.  We use macro definition here because
+// snprintf is a variadic function.
+#if _MSC_VER >= 1400 && !GTEST_OS_WINDOWS_MOBILE
+// MSVC 2005 and above support variadic macros.
+# define GTEST_SNPRINTF_(buffer, size, format, ...) \
+     _snprintf_s(buffer, size, size, format, __VA_ARGS__)
+#elif defined(_MSC_VER)
+// Windows CE does not define _snprintf_s and MSVC prior to 2005 doesn't
+// complain about _snprintf.
+# define GTEST_SNPRINTF_ _snprintf
+#else
+# define GTEST_SNPRINTF_ snprintf
+#endif
+
+// The maximum number a BiggestInt can represent.  This definition
+// works no matter BiggestInt is represented in one's complement or
+// two's complement.
+//
+// We cannot rely on numeric_limits in STL, as __int64 and long long
+// are not part of standard C++ and numeric_limits doesn't need to be
+// defined for them.
+const BiggestInt kMaxBiggestInt =
+    ~(static_cast<BiggestInt>(1) << (8*sizeof(BiggestInt) - 1));
+
+// This template class serves as a compile-time function from size to
+// type.  It maps a size in bytes to a primitive type with that
+// size. e.g.
+//
+//   TypeWithSize<4>::UInt
+//
+// is typedef-ed to be unsigned int (unsigned integer made up of 4
+// bytes).
+//
+// Such functionality should belong to STL, but I cannot find it
+// there.
+//
+// Google Test uses this class in the implementation of floating-point
+// comparison.
+//
+// For now it only handles UInt (unsigned int) as that's all Google Test
+// needs.  Other types can be easily added in the future if need
+// arises.
+template <size_t size>
+class TypeWithSize {
+ public:
+  // This prevents the user from using TypeWithSize<N> with incorrect
+  // values of N.
+  typedef void UInt;
+};
+
+// The specialization for size 4.
+template <>
+class TypeWithSize<4> {
+ public:
+  // unsigned int has size 4 in both gcc and MSVC.
+  //
+  // As base/basictypes.h doesn't compile on Windows, we cannot use
+  // uint32, uint64, and etc here.
+  typedef int Int;
+  typedef unsigned int UInt;
+};
+
+// The specialization for size 8.
+template <>
+class TypeWithSize<8> {
+ public:
+#if GTEST_OS_WINDOWS
+  typedef __int64 Int;
+  typedef unsigned __int64 UInt;
+#else
+  typedef long long Int;  // NOLINT
+  typedef unsigned long long UInt;  // NOLINT
+#endif  // GTEST_OS_WINDOWS
+};
+
+// Integer types of known sizes.
+typedef TypeWithSize<4>::Int Int32;
+typedef TypeWithSize<4>::UInt UInt32;
+typedef TypeWithSize<8>::Int Int64;
+typedef TypeWithSize<8>::UInt UInt64;
+typedef TypeWithSize<8>::Int TimeInMillis;  // Represents time in milliseconds.
+
+// Utilities for command line flags and environment variables.
+
+// Macro for referencing flags.
+#if !defined(GTEST_FLAG)
+# define GTEST_FLAG(name) FLAGS_gtest_##name
+#endif  // !defined(GTEST_FLAG)
+
+#if !defined(GTEST_USE_OWN_FLAGFILE_FLAG_)
+# define GTEST_USE_OWN_FLAGFILE_FLAG_ 1
+#endif  // !defined(GTEST_USE_OWN_FLAGFILE_FLAG_)
+
+#if !defined(GTEST_DECLARE_bool_)
+# define GTEST_FLAG_SAVER_ ::testing::internal::GTestFlagSaver
+
+// Macros for declaring flags.
+# define GTEST_DECLARE_bool_(name) GTEST_API_ extern bool GTEST_FLAG(name)
+# define GTEST_DECLARE_int32_(name) \
+    GTEST_API_ extern ::testing::internal::Int32 GTEST_FLAG(name)
+#define GTEST_DECLARE_string_(name) \
+    GTEST_API_ extern ::std::string GTEST_FLAG(name)
+
+// Macros for defining flags.
+#define GTEST_DEFINE_bool_(name, default_val, doc) \
+    GTEST_API_ bool GTEST_FLAG(name) = (default_val)
+#define GTEST_DEFINE_int32_(name, default_val, doc) \
+    GTEST_API_ ::testing::internal::Int32 GTEST_FLAG(name) = (default_val)
+#define GTEST_DEFINE_string_(name, default_val, doc) \
+    GTEST_API_ ::std::string GTEST_FLAG(name) = (default_val)
+
+#endif  // !defined(GTEST_DECLARE_bool_)
+
+// Thread annotations
+#if !defined(GTEST_EXCLUSIVE_LOCK_REQUIRED_)
+# define GTEST_EXCLUSIVE_LOCK_REQUIRED_(locks)
+# define GTEST_LOCK_EXCLUDED_(locks)
+#endif  // !defined(GTEST_EXCLUSIVE_LOCK_REQUIRED_)
+
+// Parses 'str' for a 32-bit signed integer.  If successful, writes the result
+// to *value and returns true; otherwise leaves *value unchanged and returns
+// false.
+// TODO(chandlerc): Find a better way to refactor flag and environment parsing
+// out of both gtest-port.cc and gtest.cc to avoid exporting this utility
+// function.
+bool ParseInt32(const Message& src_text, const char* str, Int32* value);
+
+// Parses a bool/Int32/string from the environment variable
+// corresponding to the given Google Test flag.
+bool BoolFromGTestEnv(const char* flag, bool default_val);
+GTEST_API_ Int32 Int32FromGTestEnv(const char* flag, Int32 default_val);
+std::string StringFromGTestEnv(const char* flag, const char* default_val);
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_PORT_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-string.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-string.h
new file mode 100644
index 0000000000..97f1a7fdd2
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-string.h
@@ -0,0 +1,167 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: wan@google.com (Zhanyong Wan), eefacm@gmail.com (Sean Mcafee)
+//
+// The Google C++ Testing Framework (Google Test)
+//
+// This header file declares the String class and functions used internally by
+// Google Test.  They are subject to change without notice. They should not used
+// by code external to Google Test.
+//
+// This header file is #included by <gtest/internal/gtest-internal.h>.
+// It should not be #included by other files.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_STRING_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_STRING_H_
+
+#ifdef __BORLANDC__
+// string.h is not guaranteed to provide strcpy on C++ Builder.
+# include <mem.h>
+#endif
+
+#include <string.h>
+#include <string>
+
+#include "gtest/internal/gtest-port.h"
+
+namespace testing {
+namespace internal {
+
+// String - an abstract class holding static string utilities.
+class GTEST_API_ String {
+ public:
+  // Static utility methods
+
+  // Clones a 0-terminated C string, allocating memory using new.  The
+  // caller is responsible for deleting the return value using
+  // delete[].  Returns the cloned string, or NULL if the input is
+  // NULL.
+  //
+  // This is different from strdup() in string.h, which allocates
+  // memory using malloc().
+  static const char* CloneCString(const char* c_str);
+
+#if GTEST_OS_WINDOWS_MOBILE
+  // Windows CE does not have the 'ANSI' versions of Win32 APIs. To be
+  // able to pass strings to Win32 APIs on CE we need to convert them
+  // to 'Unicode', UTF-16.
+
+  // Creates a UTF-16 wide string from the given ANSI string, allocating
+  // memory using new. The caller is responsible for deleting the return
+  // value using delete[]. Returns the wide string, or NULL if the
+  // input is NULL.
+  //
+  // The wide string is created using the ANSI codepage (CP_ACP) to
+  // match the behaviour of the ANSI versions of Win32 calls and the
+  // C runtime.
+  static LPCWSTR AnsiToUtf16(const char* c_str);
+
+  // Creates an ANSI string from the given wide string, allocating
+  // memory using new. The caller is responsible for deleting the return
+  // value using delete[]. Returns the ANSI string, or NULL if the
+  // input is NULL.
+  //
+  // The returned string is created using the ANSI codepage (CP_ACP) to
+  // match the behaviour of the ANSI versions of Win32 calls and the
+  // C runtime.
+  static const char* Utf16ToAnsi(LPCWSTR utf16_str);
+#endif
+
+  // Compares two C strings.  Returns true iff they have the same content.
+  //
+  // Unlike strcmp(), this function can handle NULL argument(s).  A
+  // NULL C string is considered different to any non-NULL C string,
+  // including the empty string.
+  static bool CStringEquals(const char* lhs, const char* rhs);
+
+  // Converts a wide C string to a String using the UTF-8 encoding.
+  // NULL will be converted to "(null)".  If an error occurred during
+  // the conversion, "(failed to convert from wide string)" is
+  // returned.
+  static std::string ShowWideCString(const wchar_t* wide_c_str);
+
+  // Compares two wide C strings.  Returns true iff they have the same
+  // content.
+  //
+  // Unlike wcscmp(), this function can handle NULL argument(s).  A
+  // NULL C string is considered different to any non-NULL C string,
+  // including the empty string.
+  static bool WideCStringEquals(const wchar_t* lhs, const wchar_t* rhs);
+
+  // Compares two C strings, ignoring case.  Returns true iff they
+  // have the same content.
+  //
+  // Unlike strcasecmp(), this function can handle NULL argument(s).
+  // A NULL C string is considered different to any non-NULL C string,
+  // including the empty string.
+  static bool CaseInsensitiveCStringEquals(const char* lhs,
+                                           const char* rhs);
+
+  // Compares two wide C strings, ignoring case.  Returns true iff they
+  // have the same content.
+  //
+  // Unlike wcscasecmp(), this function can handle NULL argument(s).
+  // A NULL C string is considered different to any non-NULL wide C string,
+  // including the empty string.
+  // NB: The implementations on different platforms slightly differ.
+  // On windows, this method uses _wcsicmp which compares according to LC_CTYPE
+  // environment variable. On GNU platform this method uses wcscasecmp
+  // which compares according to LC_CTYPE category of the current locale.
+  // On MacOS X, it uses towlower, which also uses LC_CTYPE category of the
+  // current locale.
+  static bool CaseInsensitiveWideCStringEquals(const wchar_t* lhs,
+                                               const wchar_t* rhs);
+
+  // Returns true iff the given string ends with the given suffix, ignoring
+  // case. Any string is considered to end with an empty suffix.
+  static bool EndsWithCaseInsensitive(
+      const std::string& str, const std::string& suffix);
+
+  // Formats an int value as "%02d".
+  static std::string FormatIntWidth2(int value);  // "%02d" for width == 2
+
+  // Formats an int value as "%X".
+  static std::string FormatHexInt(int value);
+
+  // Formats a byte as "%02X".
+  static std::string FormatByte(unsigned char value);
+
+ private:
+  String();  // Not meant to be instantiated.
+};  // class String
+
+// Gets the content of the stringstream's buffer as an std::string.  Each '\0'
+// character in the buffer is replaced with "\\0".
+GTEST_API_ std::string StringStreamToString(::std::stringstream* stream);
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_STRING_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-tuple.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-tuple.h
new file mode 100644
index 0000000000..e9b405340a
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-tuple.h
@@ -0,0 +1,1020 @@
+// This file was GENERATED by command:
+//     pump.py gtest-tuple.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2009 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Implements a subset of TR1 tuple needed by Google Test and Google Mock.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
+
+#include <utility>  // For ::std::pair.
+
+// The compiler used in Symbian has a bug that prevents us from declaring the
+// tuple template as a friend (it complains that tuple is redefined).  This
+// hack bypasses the bug by declaring the members that should otherwise be
+// private as public.
+// Sun Studio versions < 12 also have the above bug.
+#if defined(__SYMBIAN32__) || (defined(__SUNPRO_CC) && __SUNPRO_CC < 0x590)
+# define GTEST_DECLARE_TUPLE_AS_FRIEND_ public:
+#else
+# define GTEST_DECLARE_TUPLE_AS_FRIEND_ \
+    template <GTEST_10_TYPENAMES_(U)> friend class tuple; \
+   private:
+#endif
+
+// Visual Studio 2010, 2012, and 2013 define symbols in std::tr1 that conflict
+// with our own definitions. Therefore using our own tuple does not work on
+// those compilers.
+#if defined(_MSC_VER) && _MSC_VER >= 1600  /* 1600 is Visual Studio 2010 */
+# error "gtest's tuple doesn't compile on Visual Studio 2010 or later. \
+GTEST_USE_OWN_TR1_TUPLE must be set to 0 on those compilers."
+#endif
+
+// GTEST_n_TUPLE_(T) is the type of an n-tuple.
+#define GTEST_0_TUPLE_(T) tuple<>
+#define GTEST_1_TUPLE_(T) tuple<T##0, void, void, void, void, void, void, \
+    void, void, void>
+#define GTEST_2_TUPLE_(T) tuple<T##0, T##1, void, void, void, void, void, \
+    void, void, void>
+#define GTEST_3_TUPLE_(T) tuple<T##0, T##1, T##2, void, void, void, void, \
+    void, void, void>
+#define GTEST_4_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, void, void, void, \
+    void, void, void>
+#define GTEST_5_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, void, void, \
+    void, void, void>
+#define GTEST_6_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, T##5, void, \
+    void, void, void>
+#define GTEST_7_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, T##5, T##6, \
+    void, void, void>
+#define GTEST_8_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, T##5, T##6, \
+    T##7, void, void>
+#define GTEST_9_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, T##5, T##6, \
+    T##7, T##8, void>
+#define GTEST_10_TUPLE_(T) tuple<T##0, T##1, T##2, T##3, T##4, T##5, T##6, \
+    T##7, T##8, T##9>
+
+// GTEST_n_TYPENAMES_(T) declares a list of n typenames.
+#define GTEST_0_TYPENAMES_(T)
+#define GTEST_1_TYPENAMES_(T) typename T##0
+#define GTEST_2_TYPENAMES_(T) typename T##0, typename T##1
+#define GTEST_3_TYPENAMES_(T) typename T##0, typename T##1, typename T##2
+#define GTEST_4_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3
+#define GTEST_5_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4
+#define GTEST_6_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5
+#define GTEST_7_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6
+#define GTEST_8_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6, typename T##7
+#define GTEST_9_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6, \
+    typename T##7, typename T##8
+#define GTEST_10_TYPENAMES_(T) typename T##0, typename T##1, typename T##2, \
+    typename T##3, typename T##4, typename T##5, typename T##6, \
+    typename T##7, typename T##8, typename T##9
+
+// In theory, defining stuff in the ::std namespace is undefined
+// behavior.  We can do this as we are playing the role of a standard
+// library vendor.
+namespace std {
+namespace tr1 {
+
+template <typename T0 = void, typename T1 = void, typename T2 = void,
+    typename T3 = void, typename T4 = void, typename T5 = void,
+    typename T6 = void, typename T7 = void, typename T8 = void,
+    typename T9 = void>
+class tuple;
+
+// Anything in namespace gtest_internal is Google Test's INTERNAL
+// IMPLEMENTATION DETAIL and MUST NOT BE USED DIRECTLY in user code.
+namespace gtest_internal {
+
+// ByRef<T>::type is T if T is a reference; otherwise it's const T&.
+template <typename T>
+struct ByRef { typedef const T& type; };  // NOLINT
+template <typename T>
+struct ByRef<T&> { typedef T& type; };  // NOLINT
+
+// A handy wrapper for ByRef.
+#define GTEST_BY_REF_(T) typename ::std::tr1::gtest_internal::ByRef<T>::type
+
+// AddRef<T>::type is T if T is a reference; otherwise it's T&.  This
+// is the same as tr1::add_reference<T>::type.
+template <typename T>
+struct AddRef { typedef T& type; };  // NOLINT
+template <typename T>
+struct AddRef<T&> { typedef T& type; };  // NOLINT
+
+// A handy wrapper for AddRef.
+#define GTEST_ADD_REF_(T) typename ::std::tr1::gtest_internal::AddRef<T>::type
+
+// A helper for implementing get<k>().
+template <int k> class Get;
+
+// A helper for implementing tuple_element<k, T>.  kIndexValid is true
+// iff k < the number of fields in tuple type T.
+template <bool kIndexValid, int kIndex, class Tuple>
+struct TupleElement;
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 0, GTEST_10_TUPLE_(T) > {
+  typedef T0 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 1, GTEST_10_TUPLE_(T) > {
+  typedef T1 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 2, GTEST_10_TUPLE_(T) > {
+  typedef T2 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 3, GTEST_10_TUPLE_(T) > {
+  typedef T3 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 4, GTEST_10_TUPLE_(T) > {
+  typedef T4 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 5, GTEST_10_TUPLE_(T) > {
+  typedef T5 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 6, GTEST_10_TUPLE_(T) > {
+  typedef T6 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 7, GTEST_10_TUPLE_(T) > {
+  typedef T7 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 8, GTEST_10_TUPLE_(T) > {
+  typedef T8 type;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct TupleElement<true, 9, GTEST_10_TUPLE_(T) > {
+  typedef T9 type;
+};
+
+}  // namespace gtest_internal
+
+template <>
+class tuple<> {
+ public:
+  tuple() {}
+  tuple(const tuple& /* t */)  {}
+  tuple& operator=(const tuple& /* t */) { return *this; }
+};
+
+template <GTEST_1_TYPENAMES_(T)>
+class GTEST_1_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0) : f0_(f0) {}
+
+  tuple(const tuple& t) : f0_(t.f0_) {}
+
+  template <GTEST_1_TYPENAMES_(U)>
+  tuple(const GTEST_1_TUPLE_(U)& t) : f0_(t.f0_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_1_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_1_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_1_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_1_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    return *this;
+  }
+
+  T0 f0_;
+};
+
+template <GTEST_2_TYPENAMES_(T)>
+class GTEST_2_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1) : f0_(f0),
+      f1_(f1) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_) {}
+
+  template <GTEST_2_TYPENAMES_(U)>
+  tuple(const GTEST_2_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_) {}
+  template <typename U0, typename U1>
+  tuple(const ::std::pair<U0, U1>& p) : f0_(p.first), f1_(p.second) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_2_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_2_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+  template <typename U0, typename U1>
+  tuple& operator=(const ::std::pair<U0, U1>& p) {
+    f0_ = p.first;
+    f1_ = p.second;
+    return *this;
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_2_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_2_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+};
+
+template <GTEST_3_TYPENAMES_(T)>
+class GTEST_3_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2) : f0_(f0), f1_(f1), f2_(f2) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_) {}
+
+  template <GTEST_3_TYPENAMES_(U)>
+  tuple(const GTEST_3_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_3_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_3_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_3_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_3_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+};
+
+template <GTEST_4_TYPENAMES_(T)>
+class GTEST_4_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3) : f0_(f0), f1_(f1), f2_(f2),
+      f3_(f3) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_) {}
+
+  template <GTEST_4_TYPENAMES_(U)>
+  tuple(const GTEST_4_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_4_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_4_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_4_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_4_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+};
+
+template <GTEST_5_TYPENAMES_(T)>
+class GTEST_5_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3,
+      GTEST_BY_REF_(T4) f4) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_) {}
+
+  template <GTEST_5_TYPENAMES_(U)>
+  tuple(const GTEST_5_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_5_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_5_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_5_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_5_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+};
+
+template <GTEST_6_TYPENAMES_(T)>
+class GTEST_6_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4),
+      f5_(f5) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_) {}
+
+  template <GTEST_6_TYPENAMES_(U)>
+  tuple(const GTEST_6_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_6_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_6_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_6_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_6_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+};
+
+template <GTEST_7_TYPENAMES_(T)>
+class GTEST_7_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6) : f0_(f0), f1_(f1), f2_(f2),
+      f3_(f3), f4_(f4), f5_(f5), f6_(f6) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_) {}
+
+  template <GTEST_7_TYPENAMES_(U)>
+  tuple(const GTEST_7_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_7_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_7_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_7_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_7_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+};
+
+template <GTEST_8_TYPENAMES_(T)>
+class GTEST_8_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_(), f7_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6,
+      GTEST_BY_REF_(T7) f7) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4),
+      f5_(f5), f6_(f6), f7_(f7) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_) {}
+
+  template <GTEST_8_TYPENAMES_(U)>
+  tuple(const GTEST_8_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_8_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_8_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_8_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_8_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    f7_ = t.f7_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+  T7 f7_;
+};
+
+template <GTEST_9_TYPENAMES_(T)>
+class GTEST_9_TUPLE_(T) {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_(), f7_(), f8_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6, GTEST_BY_REF_(T7) f7,
+      GTEST_BY_REF_(T8) f8) : f0_(f0), f1_(f1), f2_(f2), f3_(f3), f4_(f4),
+      f5_(f5), f6_(f6), f7_(f7), f8_(f8) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_) {}
+
+  template <GTEST_9_TYPENAMES_(U)>
+  tuple(const GTEST_9_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_9_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_9_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_9_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_9_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    f7_ = t.f7_;
+    f8_ = t.f8_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+  T7 f7_;
+  T8 f8_;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+class tuple {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : f0_(), f1_(), f2_(), f3_(), f4_(), f5_(), f6_(), f7_(), f8_(),
+      f9_() {}
+
+  explicit tuple(GTEST_BY_REF_(T0) f0, GTEST_BY_REF_(T1) f1,
+      GTEST_BY_REF_(T2) f2, GTEST_BY_REF_(T3) f3, GTEST_BY_REF_(T4) f4,
+      GTEST_BY_REF_(T5) f5, GTEST_BY_REF_(T6) f6, GTEST_BY_REF_(T7) f7,
+      GTEST_BY_REF_(T8) f8, GTEST_BY_REF_(T9) f9) : f0_(f0), f1_(f1), f2_(f2),
+      f3_(f3), f4_(f4), f5_(f5), f6_(f6), f7_(f7), f8_(f8), f9_(f9) {}
+
+  tuple(const tuple& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_), f3_(t.f3_),
+      f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_), f9_(t.f9_) {}
+
+  template <GTEST_10_TYPENAMES_(U)>
+  tuple(const GTEST_10_TUPLE_(U)& t) : f0_(t.f0_), f1_(t.f1_), f2_(t.f2_),
+      f3_(t.f3_), f4_(t.f4_), f5_(t.f5_), f6_(t.f6_), f7_(t.f7_), f8_(t.f8_),
+      f9_(t.f9_) {}
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_10_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_10_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_10_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_10_TUPLE_(U)& t) {
+    f0_ = t.f0_;
+    f1_ = t.f1_;
+    f2_ = t.f2_;
+    f3_ = t.f3_;
+    f4_ = t.f4_;
+    f5_ = t.f5_;
+    f6_ = t.f6_;
+    f7_ = t.f7_;
+    f8_ = t.f8_;
+    f9_ = t.f9_;
+    return *this;
+  }
+
+  T0 f0_;
+  T1 f1_;
+  T2 f2_;
+  T3 f3_;
+  T4 f4_;
+  T5 f5_;
+  T6 f6_;
+  T7 f7_;
+  T8 f8_;
+  T9 f9_;
+};
+
+// 6.1.3.2 Tuple creation functions.
+
+// Known limitations: we don't support passing an
+// std::tr1::reference_wrapper<T> to make_tuple().  And we don't
+// implement tie().
+
+inline tuple<> make_tuple() { return tuple<>(); }
+
+template <GTEST_1_TYPENAMES_(T)>
+inline GTEST_1_TUPLE_(T) make_tuple(const T0& f0) {
+  return GTEST_1_TUPLE_(T)(f0);
+}
+
+template <GTEST_2_TYPENAMES_(T)>
+inline GTEST_2_TUPLE_(T) make_tuple(const T0& f0, const T1& f1) {
+  return GTEST_2_TUPLE_(T)(f0, f1);
+}
+
+template <GTEST_3_TYPENAMES_(T)>
+inline GTEST_3_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2) {
+  return GTEST_3_TUPLE_(T)(f0, f1, f2);
+}
+
+template <GTEST_4_TYPENAMES_(T)>
+inline GTEST_4_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3) {
+  return GTEST_4_TUPLE_(T)(f0, f1, f2, f3);
+}
+
+template <GTEST_5_TYPENAMES_(T)>
+inline GTEST_5_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4) {
+  return GTEST_5_TUPLE_(T)(f0, f1, f2, f3, f4);
+}
+
+template <GTEST_6_TYPENAMES_(T)>
+inline GTEST_6_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5) {
+  return GTEST_6_TUPLE_(T)(f0, f1, f2, f3, f4, f5);
+}
+
+template <GTEST_7_TYPENAMES_(T)>
+inline GTEST_7_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6) {
+  return GTEST_7_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6);
+}
+
+template <GTEST_8_TYPENAMES_(T)>
+inline GTEST_8_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6, const T7& f7) {
+  return GTEST_8_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6, f7);
+}
+
+template <GTEST_9_TYPENAMES_(T)>
+inline GTEST_9_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6, const T7& f7,
+    const T8& f8) {
+  return GTEST_9_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6, f7, f8);
+}
+
+template <GTEST_10_TYPENAMES_(T)>
+inline GTEST_10_TUPLE_(T) make_tuple(const T0& f0, const T1& f1, const T2& f2,
+    const T3& f3, const T4& f4, const T5& f5, const T6& f6, const T7& f7,
+    const T8& f8, const T9& f9) {
+  return GTEST_10_TUPLE_(T)(f0, f1, f2, f3, f4, f5, f6, f7, f8, f9);
+}
+
+// 6.1.3.3 Tuple helper classes.
+
+template <typename Tuple> struct tuple_size;
+
+template <GTEST_0_TYPENAMES_(T)>
+struct tuple_size<GTEST_0_TUPLE_(T) > {
+  static const int value = 0;
+};
+
+template <GTEST_1_TYPENAMES_(T)>
+struct tuple_size<GTEST_1_TUPLE_(T) > {
+  static const int value = 1;
+};
+
+template <GTEST_2_TYPENAMES_(T)>
+struct tuple_size<GTEST_2_TUPLE_(T) > {
+  static const int value = 2;
+};
+
+template <GTEST_3_TYPENAMES_(T)>
+struct tuple_size<GTEST_3_TUPLE_(T) > {
+  static const int value = 3;
+};
+
+template <GTEST_4_TYPENAMES_(T)>
+struct tuple_size<GTEST_4_TUPLE_(T) > {
+  static const int value = 4;
+};
+
+template <GTEST_5_TYPENAMES_(T)>
+struct tuple_size<GTEST_5_TUPLE_(T) > {
+  static const int value = 5;
+};
+
+template <GTEST_6_TYPENAMES_(T)>
+struct tuple_size<GTEST_6_TUPLE_(T) > {
+  static const int value = 6;
+};
+
+template <GTEST_7_TYPENAMES_(T)>
+struct tuple_size<GTEST_7_TUPLE_(T) > {
+  static const int value = 7;
+};
+
+template <GTEST_8_TYPENAMES_(T)>
+struct tuple_size<GTEST_8_TUPLE_(T) > {
+  static const int value = 8;
+};
+
+template <GTEST_9_TYPENAMES_(T)>
+struct tuple_size<GTEST_9_TUPLE_(T) > {
+  static const int value = 9;
+};
+
+template <GTEST_10_TYPENAMES_(T)>
+struct tuple_size<GTEST_10_TUPLE_(T) > {
+  static const int value = 10;
+};
+
+template <int k, class Tuple>
+struct tuple_element {
+  typedef typename gtest_internal::TupleElement<
+      k < (tuple_size<Tuple>::value), k, Tuple>::type type;
+};
+
+#define GTEST_TUPLE_ELEMENT_(k, Tuple) typename tuple_element<k, Tuple >::type
+
+// 6.1.3.4 Element access.
+
+namespace gtest_internal {
+
+template <>
+class Get<0> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(0, Tuple))
+  Field(Tuple& t) { return t.f0_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(0, Tuple))
+  ConstField(const Tuple& t) { return t.f0_; }
+};
+
+template <>
+class Get<1> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(1, Tuple))
+  Field(Tuple& t) { return t.f1_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(1, Tuple))
+  ConstField(const Tuple& t) { return t.f1_; }
+};
+
+template <>
+class Get<2> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(2, Tuple))
+  Field(Tuple& t) { return t.f2_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(2, Tuple))
+  ConstField(const Tuple& t) { return t.f2_; }
+};
+
+template <>
+class Get<3> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(3, Tuple))
+  Field(Tuple& t) { return t.f3_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(3, Tuple))
+  ConstField(const Tuple& t) { return t.f3_; }
+};
+
+template <>
+class Get<4> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(4, Tuple))
+  Field(Tuple& t) { return t.f4_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(4, Tuple))
+  ConstField(const Tuple& t) { return t.f4_; }
+};
+
+template <>
+class Get<5> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(5, Tuple))
+  Field(Tuple& t) { return t.f5_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(5, Tuple))
+  ConstField(const Tuple& t) { return t.f5_; }
+};
+
+template <>
+class Get<6> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(6, Tuple))
+  Field(Tuple& t) { return t.f6_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(6, Tuple))
+  ConstField(const Tuple& t) { return t.f6_; }
+};
+
+template <>
+class Get<7> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(7, Tuple))
+  Field(Tuple& t) { return t.f7_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(7, Tuple))
+  ConstField(const Tuple& t) { return t.f7_; }
+};
+
+template <>
+class Get<8> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(8, Tuple))
+  Field(Tuple& t) { return t.f8_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(8, Tuple))
+  ConstField(const Tuple& t) { return t.f8_; }
+};
+
+template <>
+class Get<9> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(9, Tuple))
+  Field(Tuple& t) { return t.f9_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(9, Tuple))
+  ConstField(const Tuple& t) { return t.f9_; }
+};
+
+}  // namespace gtest_internal
+
+template <int k, GTEST_10_TYPENAMES_(T)>
+GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(k, GTEST_10_TUPLE_(T)))
+get(GTEST_10_TUPLE_(T)& t) {
+  return gtest_internal::Get<k>::Field(t);
+}
+
+template <int k, GTEST_10_TYPENAMES_(T)>
+GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(k,  GTEST_10_TUPLE_(T)))
+get(const GTEST_10_TUPLE_(T)& t) {
+  return gtest_internal::Get<k>::ConstField(t);
+}
+
+// 6.1.3.5 Relational operators
+
+// We only implement == and !=, as we don't have a need for the rest yet.
+
+namespace gtest_internal {
+
+// SameSizeTuplePrefixComparator<k, k>::Eq(t1, t2) returns true if the
+// first k fields of t1 equals the first k fields of t2.
+// SameSizeTuplePrefixComparator(k1, k2) would be a compiler error if
+// k1 != k2.
+template <int kSize1, int kSize2>
+struct SameSizeTuplePrefixComparator;
+
+template <>
+struct SameSizeTuplePrefixComparator<0, 0> {
+  template <class Tuple1, class Tuple2>
+  static bool Eq(const Tuple1& /* t1 */, const Tuple2& /* t2 */) {
+    return true;
+  }
+};
+
+template <int k>
+struct SameSizeTuplePrefixComparator<k, k> {
+  template <class Tuple1, class Tuple2>
+  static bool Eq(const Tuple1& t1, const Tuple2& t2) {
+    return SameSizeTuplePrefixComparator<k - 1, k - 1>::Eq(t1, t2) &&
+        ::std::tr1::get<k - 1>(t1) == ::std::tr1::get<k - 1>(t2);
+  }
+};
+
+}  // namespace gtest_internal
+
+template <GTEST_10_TYPENAMES_(T), GTEST_10_TYPENAMES_(U)>
+inline bool operator==(const GTEST_10_TUPLE_(T)& t,
+                       const GTEST_10_TUPLE_(U)& u) {
+  return gtest_internal::SameSizeTuplePrefixComparator<
+      tuple_size<GTEST_10_TUPLE_(T) >::value,
+      tuple_size<GTEST_10_TUPLE_(U) >::value>::Eq(t, u);
+}
+
+template <GTEST_10_TYPENAMES_(T), GTEST_10_TYPENAMES_(U)>
+inline bool operator!=(const GTEST_10_TUPLE_(T)& t,
+                       const GTEST_10_TUPLE_(U)& u) { return !(t == u); }
+
+// 6.1.4 Pairs.
+// Unimplemented.
+
+}  // namespace tr1
+}  // namespace std
+
+#undef GTEST_0_TUPLE_
+#undef GTEST_1_TUPLE_
+#undef GTEST_2_TUPLE_
+#undef GTEST_3_TUPLE_
+#undef GTEST_4_TUPLE_
+#undef GTEST_5_TUPLE_
+#undef GTEST_6_TUPLE_
+#undef GTEST_7_TUPLE_
+#undef GTEST_8_TUPLE_
+#undef GTEST_9_TUPLE_
+#undef GTEST_10_TUPLE_
+
+#undef GTEST_0_TYPENAMES_
+#undef GTEST_1_TYPENAMES_
+#undef GTEST_2_TYPENAMES_
+#undef GTEST_3_TYPENAMES_
+#undef GTEST_4_TYPENAMES_
+#undef GTEST_5_TYPENAMES_
+#undef GTEST_6_TYPENAMES_
+#undef GTEST_7_TYPENAMES_
+#undef GTEST_8_TYPENAMES_
+#undef GTEST_9_TYPENAMES_
+#undef GTEST_10_TYPENAMES_
+
+#undef GTEST_DECLARE_TUPLE_AS_FRIEND_
+#undef GTEST_BY_REF_
+#undef GTEST_ADD_REF_
+#undef GTEST_TUPLE_ELEMENT_
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-tuple.h.pump b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-tuple.h.pump
new file mode 100644
index 0000000000..429ddfeeca
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-tuple.h.pump
@@ -0,0 +1,347 @@
+$$ -*- mode: c++; -*-
+$var n = 10  $$ Maximum number of tuple fields we want to support.
+$$ This meta comment fixes auto-indentation in Emacs. }}
+// Copyright 2009 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Implements a subset of TR1 tuple needed by Google Test and Google Mock.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
+
+#include <utility>  // For ::std::pair.
+
+// The compiler used in Symbian has a bug that prevents us from declaring the
+// tuple template as a friend (it complains that tuple is redefined).  This
+// hack bypasses the bug by declaring the members that should otherwise be
+// private as public.
+// Sun Studio versions < 12 also have the above bug.
+#if defined(__SYMBIAN32__) || (defined(__SUNPRO_CC) && __SUNPRO_CC < 0x590)
+# define GTEST_DECLARE_TUPLE_AS_FRIEND_ public:
+#else
+# define GTEST_DECLARE_TUPLE_AS_FRIEND_ \
+    template <GTEST_$(n)_TYPENAMES_(U)> friend class tuple; \
+   private:
+#endif
+
+// Visual Studio 2010, 2012, and 2013 define symbols in std::tr1 that conflict
+// with our own definitions. Therefore using our own tuple does not work on
+// those compilers.
+#if defined(_MSC_VER) && _MSC_VER >= 1600  /* 1600 is Visual Studio 2010 */
+# error "gtest's tuple doesn't compile on Visual Studio 2010 or later. \
+GTEST_USE_OWN_TR1_TUPLE must be set to 0 on those compilers."
+#endif
+
+
+$range i 0..n-1
+$range j 0..n
+$range k 1..n
+// GTEST_n_TUPLE_(T) is the type of an n-tuple.
+#define GTEST_0_TUPLE_(T) tuple<>
+
+$for k [[
+$range m 0..k-1
+$range m2 k..n-1
+#define GTEST_$(k)_TUPLE_(T) tuple<$for m, [[T##$m]]$for m2 [[, void]]>
+
+]]
+
+// GTEST_n_TYPENAMES_(T) declares a list of n typenames.
+
+$for j [[
+$range m 0..j-1
+#define GTEST_$(j)_TYPENAMES_(T) $for m, [[typename T##$m]]
+
+
+]]
+
+// In theory, defining stuff in the ::std namespace is undefined
+// behavior.  We can do this as we are playing the role of a standard
+// library vendor.
+namespace std {
+namespace tr1 {
+
+template <$for i, [[typename T$i = void]]>
+class tuple;
+
+// Anything in namespace gtest_internal is Google Test's INTERNAL
+// IMPLEMENTATION DETAIL and MUST NOT BE USED DIRECTLY in user code.
+namespace gtest_internal {
+
+// ByRef<T>::type is T if T is a reference; otherwise it's const T&.
+template <typename T>
+struct ByRef { typedef const T& type; };  // NOLINT
+template <typename T>
+struct ByRef<T&> { typedef T& type; };  // NOLINT
+
+// A handy wrapper for ByRef.
+#define GTEST_BY_REF_(T) typename ::std::tr1::gtest_internal::ByRef<T>::type
+
+// AddRef<T>::type is T if T is a reference; otherwise it's T&.  This
+// is the same as tr1::add_reference<T>::type.
+template <typename T>
+struct AddRef { typedef T& type; };  // NOLINT
+template <typename T>
+struct AddRef<T&> { typedef T& type; };  // NOLINT
+
+// A handy wrapper for AddRef.
+#define GTEST_ADD_REF_(T) typename ::std::tr1::gtest_internal::AddRef<T>::type
+
+// A helper for implementing get<k>().
+template <int k> class Get;
+
+// A helper for implementing tuple_element<k, T>.  kIndexValid is true
+// iff k < the number of fields in tuple type T.
+template <bool kIndexValid, int kIndex, class Tuple>
+struct TupleElement;
+
+
+$for i [[
+template <GTEST_$(n)_TYPENAMES_(T)>
+struct TupleElement<true, $i, GTEST_$(n)_TUPLE_(T) > {
+  typedef T$i type;
+};
+
+
+]]
+}  // namespace gtest_internal
+
+template <>
+class tuple<> {
+ public:
+  tuple() {}
+  tuple(const tuple& /* t */)  {}
+  tuple& operator=(const tuple& /* t */) { return *this; }
+};
+
+
+$for k [[
+$range m 0..k-1
+template <GTEST_$(k)_TYPENAMES_(T)>
+class $if k < n [[GTEST_$(k)_TUPLE_(T)]] $else [[tuple]] {
+ public:
+  template <int k> friend class gtest_internal::Get;
+
+  tuple() : $for m, [[f$(m)_()]] {}
+
+  explicit tuple($for m, [[GTEST_BY_REF_(T$m) f$m]]) : [[]]
+$for m, [[f$(m)_(f$m)]] {}
+
+  tuple(const tuple& t) : $for m, [[f$(m)_(t.f$(m)_)]] {}
+
+  template <GTEST_$(k)_TYPENAMES_(U)>
+  tuple(const GTEST_$(k)_TUPLE_(U)& t) : $for m, [[f$(m)_(t.f$(m)_)]] {}
+
+$if k == 2 [[
+  template <typename U0, typename U1>
+  tuple(const ::std::pair<U0, U1>& p) : f0_(p.first), f1_(p.second) {}
+
+]]
+
+  tuple& operator=(const tuple& t) { return CopyFrom(t); }
+
+  template <GTEST_$(k)_TYPENAMES_(U)>
+  tuple& operator=(const GTEST_$(k)_TUPLE_(U)& t) {
+    return CopyFrom(t);
+  }
+
+$if k == 2 [[
+  template <typename U0, typename U1>
+  tuple& operator=(const ::std::pair<U0, U1>& p) {
+    f0_ = p.first;
+    f1_ = p.second;
+    return *this;
+  }
+
+]]
+
+  GTEST_DECLARE_TUPLE_AS_FRIEND_
+
+  template <GTEST_$(k)_TYPENAMES_(U)>
+  tuple& CopyFrom(const GTEST_$(k)_TUPLE_(U)& t) {
+
+$for m [[
+    f$(m)_ = t.f$(m)_;
+
+]]
+    return *this;
+  }
+
+
+$for m [[
+  T$m f$(m)_;
+
+]]
+};
+
+
+]]
+// 6.1.3.2 Tuple creation functions.
+
+// Known limitations: we don't support passing an
+// std::tr1::reference_wrapper<T> to make_tuple().  And we don't
+// implement tie().
+
+inline tuple<> make_tuple() { return tuple<>(); }
+
+$for k [[
+$range m 0..k-1
+
+template <GTEST_$(k)_TYPENAMES_(T)>
+inline GTEST_$(k)_TUPLE_(T) make_tuple($for m, [[const T$m& f$m]]) {
+  return GTEST_$(k)_TUPLE_(T)($for m, [[f$m]]);
+}
+
+]]
+
+// 6.1.3.3 Tuple helper classes.
+
+template <typename Tuple> struct tuple_size;
+
+
+$for j [[
+template <GTEST_$(j)_TYPENAMES_(T)>
+struct tuple_size<GTEST_$(j)_TUPLE_(T) > {
+  static const int value = $j;
+};
+
+
+]]
+template <int k, class Tuple>
+struct tuple_element {
+  typedef typename gtest_internal::TupleElement<
+      k < (tuple_size<Tuple>::value), k, Tuple>::type type;
+};
+
+#define GTEST_TUPLE_ELEMENT_(k, Tuple) typename tuple_element<k, Tuple >::type
+
+// 6.1.3.4 Element access.
+
+namespace gtest_internal {
+
+
+$for i [[
+template <>
+class Get<$i> {
+ public:
+  template <class Tuple>
+  static GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_($i, Tuple))
+  Field(Tuple& t) { return t.f$(i)_; }  // NOLINT
+
+  template <class Tuple>
+  static GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_($i, Tuple))
+  ConstField(const Tuple& t) { return t.f$(i)_; }
+};
+
+
+]]
+}  // namespace gtest_internal
+
+template <int k, GTEST_$(n)_TYPENAMES_(T)>
+GTEST_ADD_REF_(GTEST_TUPLE_ELEMENT_(k, GTEST_$(n)_TUPLE_(T)))
+get(GTEST_$(n)_TUPLE_(T)& t) {
+  return gtest_internal::Get<k>::Field(t);
+}
+
+template <int k, GTEST_$(n)_TYPENAMES_(T)>
+GTEST_BY_REF_(GTEST_TUPLE_ELEMENT_(k,  GTEST_$(n)_TUPLE_(T)))
+get(const GTEST_$(n)_TUPLE_(T)& t) {
+  return gtest_internal::Get<k>::ConstField(t);
+}
+
+// 6.1.3.5 Relational operators
+
+// We only implement == and !=, as we don't have a need for the rest yet.
+
+namespace gtest_internal {
+
+// SameSizeTuplePrefixComparator<k, k>::Eq(t1, t2) returns true if the
+// first k fields of t1 equals the first k fields of t2.
+// SameSizeTuplePrefixComparator(k1, k2) would be a compiler error if
+// k1 != k2.
+template <int kSize1, int kSize2>
+struct SameSizeTuplePrefixComparator;
+
+template <>
+struct SameSizeTuplePrefixComparator<0, 0> {
+  template <class Tuple1, class Tuple2>
+  static bool Eq(const Tuple1& /* t1 */, const Tuple2& /* t2 */) {
+    return true;
+  }
+};
+
+template <int k>
+struct SameSizeTuplePrefixComparator<k, k> {
+  template <class Tuple1, class Tuple2>
+  static bool Eq(const Tuple1& t1, const Tuple2& t2) {
+    return SameSizeTuplePrefixComparator<k - 1, k - 1>::Eq(t1, t2) &&
+        ::std::tr1::get<k - 1>(t1) == ::std::tr1::get<k - 1>(t2);
+  }
+};
+
+}  // namespace gtest_internal
+
+template <GTEST_$(n)_TYPENAMES_(T), GTEST_$(n)_TYPENAMES_(U)>
+inline bool operator==(const GTEST_$(n)_TUPLE_(T)& t,
+                       const GTEST_$(n)_TUPLE_(U)& u) {
+  return gtest_internal::SameSizeTuplePrefixComparator<
+      tuple_size<GTEST_$(n)_TUPLE_(T) >::value,
+      tuple_size<GTEST_$(n)_TUPLE_(U) >::value>::Eq(t, u);
+}
+
+template <GTEST_$(n)_TYPENAMES_(T), GTEST_$(n)_TYPENAMES_(U)>
+inline bool operator!=(const GTEST_$(n)_TUPLE_(T)& t,
+                       const GTEST_$(n)_TUPLE_(U)& u) { return !(t == u); }
+
+// 6.1.4 Pairs.
+// Unimplemented.
+
+}  // namespace tr1
+}  // namespace std
+
+
+$for j [[
+#undef GTEST_$(j)_TUPLE_
+
+]]
+
+
+$for j [[
+#undef GTEST_$(j)_TYPENAMES_
+
+]]
+
+#undef GTEST_DECLARE_TUPLE_AS_FRIEND_
+#undef GTEST_BY_REF_
+#undef GTEST_ADD_REF_
+#undef GTEST_TUPLE_ELEMENT_
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TUPLE_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-type-util.h b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-type-util.h
new file mode 100644
index 0000000000..e46f7cfcb4
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-type-util.h
@@ -0,0 +1,3331 @@
+// This file was GENERATED by command:
+//     pump.py gtest-type-util.h.pump
+// DO NOT EDIT BY HAND!!!
+
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Type utilities needed for implementing typed and type-parameterized
+// tests.  This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+// Currently we support at most 50 types in a list, and at most 50
+// type-parameterized tests in one type-parameterized test case.
+// Please contact googletestframework@googlegroups.com if you need
+// more.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
+
+#include "gtest/internal/gtest-port.h"
+
+// #ifdef __GNUC__ is too general here.  It is possible to use gcc without using
+// libstdc++ (which is where cxxabi.h comes from).
+# if GTEST_HAS_CXXABI_H_
+#  include <cxxabi.h>
+# elif defined(__HP_aCC)
+#  include <acxx_demangle.h>
+# endif  // GTEST_HASH_CXXABI_H_
+
+namespace testing {
+namespace internal {
+
+// GetTypeName<T>() returns a human-readable name of type T.
+// NB: This function is also used in Google Mock, so don't move it inside of
+// the typed-test-only section below.
+template <typename T>
+std::string GetTypeName() {
+# if GTEST_HAS_RTTI
+
+  const char* const name = typeid(T).name();
+#  if GTEST_HAS_CXXABI_H_ || defined(__HP_aCC)
+  int status = 0;
+  // gcc's implementation of typeid(T).name() mangles the type name,
+  // so we have to demangle it.
+#   if GTEST_HAS_CXXABI_H_
+  using abi::__cxa_demangle;
+#   endif  // GTEST_HAS_CXXABI_H_
+  char* const readable_name = __cxa_demangle(name, 0, 0, &status);
+  const std::string name_str(status == 0 ? readable_name : name);
+  free(readable_name);
+  return name_str;
+#  else
+  return name;
+#  endif  // GTEST_HAS_CXXABI_H_ || __HP_aCC
+
+# else
+
+  return "<type>";
+
+# endif  // GTEST_HAS_RTTI
+}
+
+#if GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+// AssertyTypeEq<T1, T2>::type is defined iff T1 and T2 are the same
+// type.  This can be used as a compile-time assertion to ensure that
+// two types are equal.
+
+template <typename T1, typename T2>
+struct AssertTypeEq;
+
+template <typename T>
+struct AssertTypeEq<T, T> {
+  typedef bool type;
+};
+
+// A unique type used as the default value for the arguments of class
+// template Types.  This allows us to simulate variadic templates
+// (e.g. Types<int>, Type<int, double>, and etc), which C++ doesn't
+// support directly.
+struct None {};
+
+// The following family of struct and struct templates are used to
+// represent type lists.  In particular, TypesN<T1, T2, ..., TN>
+// represents a type list with N types (T1, T2, ..., and TN) in it.
+// Except for Types0, every struct in the family has two member types:
+// Head for the first type in the list, and Tail for the rest of the
+// list.
+
+// The empty type list.
+struct Types0 {};
+
+// Type lists of length 1, 2, 3, and so on.
+
+template <typename T1>
+struct Types1 {
+  typedef T1 Head;
+  typedef Types0 Tail;
+};
+template <typename T1, typename T2>
+struct Types2 {
+  typedef T1 Head;
+  typedef Types1<T2> Tail;
+};
+
+template <typename T1, typename T2, typename T3>
+struct Types3 {
+  typedef T1 Head;
+  typedef Types2<T2, T3> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4>
+struct Types4 {
+  typedef T1 Head;
+  typedef Types3<T2, T3, T4> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+struct Types5 {
+  typedef T1 Head;
+  typedef Types4<T2, T3, T4, T5> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6>
+struct Types6 {
+  typedef T1 Head;
+  typedef Types5<T2, T3, T4, T5, T6> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7>
+struct Types7 {
+  typedef T1 Head;
+  typedef Types6<T2, T3, T4, T5, T6, T7> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8>
+struct Types8 {
+  typedef T1 Head;
+  typedef Types7<T2, T3, T4, T5, T6, T7, T8> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9>
+struct Types9 {
+  typedef T1 Head;
+  typedef Types8<T2, T3, T4, T5, T6, T7, T8, T9> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10>
+struct Types10 {
+  typedef T1 Head;
+  typedef Types9<T2, T3, T4, T5, T6, T7, T8, T9, T10> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11>
+struct Types11 {
+  typedef T1 Head;
+  typedef Types10<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12>
+struct Types12 {
+  typedef T1 Head;
+  typedef Types11<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13>
+struct Types13 {
+  typedef T1 Head;
+  typedef Types12<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14>
+struct Types14 {
+  typedef T1 Head;
+  typedef Types13<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15>
+struct Types15 {
+  typedef T1 Head;
+  typedef Types14<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16>
+struct Types16 {
+  typedef T1 Head;
+  typedef Types15<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17>
+struct Types17 {
+  typedef T1 Head;
+  typedef Types16<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18>
+struct Types18 {
+  typedef T1 Head;
+  typedef Types17<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19>
+struct Types19 {
+  typedef T1 Head;
+  typedef Types18<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20>
+struct Types20 {
+  typedef T1 Head;
+  typedef Types19<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21>
+struct Types21 {
+  typedef T1 Head;
+  typedef Types20<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22>
+struct Types22 {
+  typedef T1 Head;
+  typedef Types21<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23>
+struct Types23 {
+  typedef T1 Head;
+  typedef Types22<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24>
+struct Types24 {
+  typedef T1 Head;
+  typedef Types23<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25>
+struct Types25 {
+  typedef T1 Head;
+  typedef Types24<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26>
+struct Types26 {
+  typedef T1 Head;
+  typedef Types25<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27>
+struct Types27 {
+  typedef T1 Head;
+  typedef Types26<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28>
+struct Types28 {
+  typedef T1 Head;
+  typedef Types27<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29>
+struct Types29 {
+  typedef T1 Head;
+  typedef Types28<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30>
+struct Types30 {
+  typedef T1 Head;
+  typedef Types29<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31>
+struct Types31 {
+  typedef T1 Head;
+  typedef Types30<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32>
+struct Types32 {
+  typedef T1 Head;
+  typedef Types31<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33>
+struct Types33 {
+  typedef T1 Head;
+  typedef Types32<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34>
+struct Types34 {
+  typedef T1 Head;
+  typedef Types33<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35>
+struct Types35 {
+  typedef T1 Head;
+  typedef Types34<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36>
+struct Types36 {
+  typedef T1 Head;
+  typedef Types35<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37>
+struct Types37 {
+  typedef T1 Head;
+  typedef Types36<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38>
+struct Types38 {
+  typedef T1 Head;
+  typedef Types37<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39>
+struct Types39 {
+  typedef T1 Head;
+  typedef Types38<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40>
+struct Types40 {
+  typedef T1 Head;
+  typedef Types39<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41>
+struct Types41 {
+  typedef T1 Head;
+  typedef Types40<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42>
+struct Types42 {
+  typedef T1 Head;
+  typedef Types41<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43>
+struct Types43 {
+  typedef T1 Head;
+  typedef Types42<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44>
+struct Types44 {
+  typedef T1 Head;
+  typedef Types43<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45>
+struct Types45 {
+  typedef T1 Head;
+  typedef Types44<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46>
+struct Types46 {
+  typedef T1 Head;
+  typedef Types45<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45, T46> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47>
+struct Types47 {
+  typedef T1 Head;
+  typedef Types46<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45, T46, T47> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48>
+struct Types48 {
+  typedef T1 Head;
+  typedef Types47<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45, T46, T47, T48> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49>
+struct Types49 {
+  typedef T1 Head;
+  typedef Types48<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45, T46, T47, T48, T49> Tail;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49, typename T50>
+struct Types50 {
+  typedef T1 Head;
+  typedef Types49<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+      T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+      T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+      T44, T45, T46, T47, T48, T49, T50> Tail;
+};
+
+
+}  // namespace internal
+
+// We don't want to require the users to write TypesN<...> directly,
+// as that would require them to count the length.  Types<...> is much
+// easier to write, but generates horrible messages when there is a
+// compiler error, as gcc insists on printing out each template
+// argument, even if it has the default value (this means Types<int>
+// will appear as Types<int, None, None, ..., None> in the compiler
+// errors).
+//
+// Our solution is to combine the best part of the two approaches: a
+// user would write Types<T1, ..., TN>, and Google Test will translate
+// that to TypesN<T1, ..., TN> internally to make error messages
+// readable.  The translation is done by the 'type' member of the
+// Types template.
+template <typename T1 = internal::None, typename T2 = internal::None,
+    typename T3 = internal::None, typename T4 = internal::None,
+    typename T5 = internal::None, typename T6 = internal::None,
+    typename T7 = internal::None, typename T8 = internal::None,
+    typename T9 = internal::None, typename T10 = internal::None,
+    typename T11 = internal::None, typename T12 = internal::None,
+    typename T13 = internal::None, typename T14 = internal::None,
+    typename T15 = internal::None, typename T16 = internal::None,
+    typename T17 = internal::None, typename T18 = internal::None,
+    typename T19 = internal::None, typename T20 = internal::None,
+    typename T21 = internal::None, typename T22 = internal::None,
+    typename T23 = internal::None, typename T24 = internal::None,
+    typename T25 = internal::None, typename T26 = internal::None,
+    typename T27 = internal::None, typename T28 = internal::None,
+    typename T29 = internal::None, typename T30 = internal::None,
+    typename T31 = internal::None, typename T32 = internal::None,
+    typename T33 = internal::None, typename T34 = internal::None,
+    typename T35 = internal::None, typename T36 = internal::None,
+    typename T37 = internal::None, typename T38 = internal::None,
+    typename T39 = internal::None, typename T40 = internal::None,
+    typename T41 = internal::None, typename T42 = internal::None,
+    typename T43 = internal::None, typename T44 = internal::None,
+    typename T45 = internal::None, typename T46 = internal::None,
+    typename T47 = internal::None, typename T48 = internal::None,
+    typename T49 = internal::None, typename T50 = internal::None>
+struct Types {
+  typedef internal::Types50<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46, T47, T48, T49, T50> type;
+};
+
+template <>
+struct Types<internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types0 type;
+};
+template <typename T1>
+struct Types<T1, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types1<T1> type;
+};
+template <typename T1, typename T2>
+struct Types<T1, T2, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types2<T1, T2> type;
+};
+template <typename T1, typename T2, typename T3>
+struct Types<T1, T2, T3, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types3<T1, T2, T3> type;
+};
+template <typename T1, typename T2, typename T3, typename T4>
+struct Types<T1, T2, T3, T4, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types4<T1, T2, T3, T4> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5>
+struct Types<T1, T2, T3, T4, T5, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types5<T1, T2, T3, T4, T5> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6>
+struct Types<T1, T2, T3, T4, T5, T6, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types6<T1, T2, T3, T4, T5, T6> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7>
+struct Types<T1, T2, T3, T4, T5, T6, T7, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types7<T1, T2, T3, T4, T5, T6, T7> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types8<T1, T2, T3, T4, T5, T6, T7, T8> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types9<T1, T2, T3, T4, T5, T6, T7, T8, T9> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types10<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types11<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types12<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11,
+      T12> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types13<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types14<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types15<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types16<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types17<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types18<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types19<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types20<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types21<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types22<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types23<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types24<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types25<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types26<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25,
+      T26> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types27<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types28<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types29<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types30<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types31<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types32<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types33<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types34<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types35<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types36<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types37<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types38<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types39<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types40<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39,
+      T40> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types41<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, internal::None,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types42<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None, internal::None> {
+  typedef internal::Types43<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None, internal::None> {
+  typedef internal::Types44<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44, T45,
+    internal::None, internal::None, internal::None, internal::None,
+    internal::None> {
+  typedef internal::Types45<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44, T45,
+    T46, internal::None, internal::None, internal::None, internal::None> {
+  typedef internal::Types46<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44, T45,
+    T46, T47, internal::None, internal::None, internal::None> {
+  typedef internal::Types47<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46, T47> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44, T45,
+    T46, T47, T48, internal::None, internal::None> {
+  typedef internal::Types48<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46, T47, T48> type;
+};
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49>
+struct Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14, T15,
+    T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29, T30,
+    T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44, T45,
+    T46, T47, T48, T49, internal::None> {
+  typedef internal::Types49<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46, T47, T48, T49> type;
+};
+
+namespace internal {
+
+# define GTEST_TEMPLATE_ template <typename T> class
+
+// The template "selector" struct TemplateSel<Tmpl> is used to
+// represent Tmpl, which must be a class template with one type
+// parameter, as a type.  TemplateSel<Tmpl>::Bind<T>::type is defined
+// as the type Tmpl<T>.  This allows us to actually instantiate the
+// template "selected" by TemplateSel<Tmpl>.
+//
+// This trick is necessary for simulating typedef for class templates,
+// which C++ doesn't support directly.
+template <GTEST_TEMPLATE_ Tmpl>
+struct TemplateSel {
+  template <typename T>
+  struct Bind {
+    typedef Tmpl<T> type;
+  };
+};
+
+# define GTEST_BIND_(TmplSel, T) \
+  TmplSel::template Bind<T>::type
+
+// A unique struct template used as the default value for the
+// arguments of class template Templates.  This allows us to simulate
+// variadic templates (e.g. Templates<int>, Templates<int, double>,
+// and etc), which C++ doesn't support directly.
+template <typename T>
+struct NoneT {};
+
+// The following family of struct and struct templates are used to
+// represent template lists.  In particular, TemplatesN<T1, T2, ...,
+// TN> represents a list of N templates (T1, T2, ..., and TN).  Except
+// for Templates0, every struct in the family has two member types:
+// Head for the selector of the first template in the list, and Tail
+// for the rest of the list.
+
+// The empty template list.
+struct Templates0 {};
+
+// Template lists of length 1, 2, 3, and so on.
+
+template <GTEST_TEMPLATE_ T1>
+struct Templates1 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates0 Tail;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2>
+struct Templates2 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates1<T2> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3>
+struct Templates3 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates2<T2, T3> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4>
+struct Templates4 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates3<T2, T3, T4> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5>
+struct Templates5 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates4<T2, T3, T4, T5> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6>
+struct Templates6 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates5<T2, T3, T4, T5, T6> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7>
+struct Templates7 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates6<T2, T3, T4, T5, T6, T7> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8>
+struct Templates8 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates7<T2, T3, T4, T5, T6, T7, T8> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9>
+struct Templates9 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates8<T2, T3, T4, T5, T6, T7, T8, T9> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10>
+struct Templates10 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates9<T2, T3, T4, T5, T6, T7, T8, T9, T10> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11>
+struct Templates11 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates10<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12>
+struct Templates12 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates11<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13>
+struct Templates13 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates12<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14>
+struct Templates14 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates13<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15>
+struct Templates15 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates14<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16>
+struct Templates16 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates15<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17>
+struct Templates17 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates16<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18>
+struct Templates18 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates17<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19>
+struct Templates19 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates18<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20>
+struct Templates20 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates19<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21>
+struct Templates21 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates20<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22>
+struct Templates22 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates21<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23>
+struct Templates23 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates22<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24>
+struct Templates24 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates23<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25>
+struct Templates25 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates24<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26>
+struct Templates26 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates25<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27>
+struct Templates27 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates26<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28>
+struct Templates28 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates27<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29>
+struct Templates29 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates28<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30>
+struct Templates30 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates29<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31>
+struct Templates31 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates30<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32>
+struct Templates32 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates31<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33>
+struct Templates33 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates32<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34>
+struct Templates34 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates33<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35>
+struct Templates35 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates34<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36>
+struct Templates36 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates35<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37>
+struct Templates37 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates36<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38>
+struct Templates38 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates37<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39>
+struct Templates39 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates38<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40>
+struct Templates40 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates39<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41>
+struct Templates41 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates40<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42>
+struct Templates42 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates41<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43>
+struct Templates43 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates42<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44>
+struct Templates44 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates43<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45>
+struct Templates45 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates44<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46>
+struct Templates46 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates45<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45, T46> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47>
+struct Templates47 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates46<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45, T46, T47> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47, GTEST_TEMPLATE_ T48>
+struct Templates48 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates47<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45, T46, T47, T48> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47, GTEST_TEMPLATE_ T48,
+    GTEST_TEMPLATE_ T49>
+struct Templates49 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates48<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45, T46, T47, T48, T49> Tail;
+};
+
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47, GTEST_TEMPLATE_ T48,
+    GTEST_TEMPLATE_ T49, GTEST_TEMPLATE_ T50>
+struct Templates50 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates49<T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+      T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+      T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42,
+      T43, T44, T45, T46, T47, T48, T49, T50> Tail;
+};
+
+
+// We don't want to require the users to write TemplatesN<...> directly,
+// as that would require them to count the length.  Templates<...> is much
+// easier to write, but generates horrible messages when there is a
+// compiler error, as gcc insists on printing out each template
+// argument, even if it has the default value (this means Templates<list>
+// will appear as Templates<list, NoneT, NoneT, ..., NoneT> in the compiler
+// errors).
+//
+// Our solution is to combine the best part of the two approaches: a
+// user would write Templates<T1, ..., TN>, and Google Test will translate
+// that to TemplatesN<T1, ..., TN> internally to make error messages
+// readable.  The translation is done by the 'type' member of the
+// Templates template.
+template <GTEST_TEMPLATE_ T1 = NoneT, GTEST_TEMPLATE_ T2 = NoneT,
+    GTEST_TEMPLATE_ T3 = NoneT, GTEST_TEMPLATE_ T4 = NoneT,
+    GTEST_TEMPLATE_ T5 = NoneT, GTEST_TEMPLATE_ T6 = NoneT,
+    GTEST_TEMPLATE_ T7 = NoneT, GTEST_TEMPLATE_ T8 = NoneT,
+    GTEST_TEMPLATE_ T9 = NoneT, GTEST_TEMPLATE_ T10 = NoneT,
+    GTEST_TEMPLATE_ T11 = NoneT, GTEST_TEMPLATE_ T12 = NoneT,
+    GTEST_TEMPLATE_ T13 = NoneT, GTEST_TEMPLATE_ T14 = NoneT,
+    GTEST_TEMPLATE_ T15 = NoneT, GTEST_TEMPLATE_ T16 = NoneT,
+    GTEST_TEMPLATE_ T17 = NoneT, GTEST_TEMPLATE_ T18 = NoneT,
+    GTEST_TEMPLATE_ T19 = NoneT, GTEST_TEMPLATE_ T20 = NoneT,
+    GTEST_TEMPLATE_ T21 = NoneT, GTEST_TEMPLATE_ T22 = NoneT,
+    GTEST_TEMPLATE_ T23 = NoneT, GTEST_TEMPLATE_ T24 = NoneT,
+    GTEST_TEMPLATE_ T25 = NoneT, GTEST_TEMPLATE_ T26 = NoneT,
+    GTEST_TEMPLATE_ T27 = NoneT, GTEST_TEMPLATE_ T28 = NoneT,
+    GTEST_TEMPLATE_ T29 = NoneT, GTEST_TEMPLATE_ T30 = NoneT,
+    GTEST_TEMPLATE_ T31 = NoneT, GTEST_TEMPLATE_ T32 = NoneT,
+    GTEST_TEMPLATE_ T33 = NoneT, GTEST_TEMPLATE_ T34 = NoneT,
+    GTEST_TEMPLATE_ T35 = NoneT, GTEST_TEMPLATE_ T36 = NoneT,
+    GTEST_TEMPLATE_ T37 = NoneT, GTEST_TEMPLATE_ T38 = NoneT,
+    GTEST_TEMPLATE_ T39 = NoneT, GTEST_TEMPLATE_ T40 = NoneT,
+    GTEST_TEMPLATE_ T41 = NoneT, GTEST_TEMPLATE_ T42 = NoneT,
+    GTEST_TEMPLATE_ T43 = NoneT, GTEST_TEMPLATE_ T44 = NoneT,
+    GTEST_TEMPLATE_ T45 = NoneT, GTEST_TEMPLATE_ T46 = NoneT,
+    GTEST_TEMPLATE_ T47 = NoneT, GTEST_TEMPLATE_ T48 = NoneT,
+    GTEST_TEMPLATE_ T49 = NoneT, GTEST_TEMPLATE_ T50 = NoneT>
+struct Templates {
+  typedef Templates50<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45, T46, T47, T48, T49, T50> type;
+};
+
+template <>
+struct Templates<NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT> {
+  typedef Templates0 type;
+};
+template <GTEST_TEMPLATE_ T1>
+struct Templates<T1, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT> {
+  typedef Templates1<T1> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2>
+struct Templates<T1, T2, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT> {
+  typedef Templates2<T1, T2> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3>
+struct Templates<T1, T2, T3, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates3<T1, T2, T3> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4>
+struct Templates<T1, T2, T3, T4, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates4<T1, T2, T3, T4> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5>
+struct Templates<T1, T2, T3, T4, T5, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates5<T1, T2, T3, T4, T5> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6>
+struct Templates<T1, T2, T3, T4, T5, T6, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates6<T1, T2, T3, T4, T5, T6> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates7<T1, T2, T3, T4, T5, T6, T7> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates8<T1, T2, T3, T4, T5, T6, T7, T8> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates9<T1, T2, T3, T4, T5, T6, T7, T8, T9> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates10<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates11<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates12<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates13<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates14<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates15<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates16<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates17<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates18<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates19<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates20<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates21<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT> {
+  typedef Templates22<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT> {
+  typedef Templates23<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT> {
+  typedef Templates24<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT> {
+  typedef Templates25<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT> {
+  typedef Templates26<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT> {
+  typedef Templates27<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT> {
+  typedef Templates28<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT> {
+  typedef Templates29<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates30<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates31<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates32<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates33<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates34<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates35<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates36<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, NoneT, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates37<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, NoneT, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates38<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates39<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, NoneT, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates40<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, NoneT, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates41<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, NoneT,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates42<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates43<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    NoneT, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates44<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    T45, NoneT, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates45<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    T45, T46, NoneT, NoneT, NoneT, NoneT> {
+  typedef Templates46<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45, T46> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    T45, T46, T47, NoneT, NoneT, NoneT> {
+  typedef Templates47<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45, T46, T47> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47, GTEST_TEMPLATE_ T48>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    T45, T46, T47, T48, NoneT, NoneT> {
+  typedef Templates48<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45, T46, T47, T48> type;
+};
+template <GTEST_TEMPLATE_ T1, GTEST_TEMPLATE_ T2, GTEST_TEMPLATE_ T3,
+    GTEST_TEMPLATE_ T4, GTEST_TEMPLATE_ T5, GTEST_TEMPLATE_ T6,
+    GTEST_TEMPLATE_ T7, GTEST_TEMPLATE_ T8, GTEST_TEMPLATE_ T9,
+    GTEST_TEMPLATE_ T10, GTEST_TEMPLATE_ T11, GTEST_TEMPLATE_ T12,
+    GTEST_TEMPLATE_ T13, GTEST_TEMPLATE_ T14, GTEST_TEMPLATE_ T15,
+    GTEST_TEMPLATE_ T16, GTEST_TEMPLATE_ T17, GTEST_TEMPLATE_ T18,
+    GTEST_TEMPLATE_ T19, GTEST_TEMPLATE_ T20, GTEST_TEMPLATE_ T21,
+    GTEST_TEMPLATE_ T22, GTEST_TEMPLATE_ T23, GTEST_TEMPLATE_ T24,
+    GTEST_TEMPLATE_ T25, GTEST_TEMPLATE_ T26, GTEST_TEMPLATE_ T27,
+    GTEST_TEMPLATE_ T28, GTEST_TEMPLATE_ T29, GTEST_TEMPLATE_ T30,
+    GTEST_TEMPLATE_ T31, GTEST_TEMPLATE_ T32, GTEST_TEMPLATE_ T33,
+    GTEST_TEMPLATE_ T34, GTEST_TEMPLATE_ T35, GTEST_TEMPLATE_ T36,
+    GTEST_TEMPLATE_ T37, GTEST_TEMPLATE_ T38, GTEST_TEMPLATE_ T39,
+    GTEST_TEMPLATE_ T40, GTEST_TEMPLATE_ T41, GTEST_TEMPLATE_ T42,
+    GTEST_TEMPLATE_ T43, GTEST_TEMPLATE_ T44, GTEST_TEMPLATE_ T45,
+    GTEST_TEMPLATE_ T46, GTEST_TEMPLATE_ T47, GTEST_TEMPLATE_ T48,
+    GTEST_TEMPLATE_ T49>
+struct Templates<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13, T14,
+    T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28, T29,
+    T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43, T44,
+    T45, T46, T47, T48, T49, NoneT> {
+  typedef Templates49<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+      T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27,
+      T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41,
+      T42, T43, T44, T45, T46, T47, T48, T49> type;
+};
+
+// The TypeList template makes it possible to use either a single type
+// or a Types<...> list in TYPED_TEST_CASE() and
+// INSTANTIATE_TYPED_TEST_CASE_P().
+
+template <typename T>
+struct TypeList {
+  typedef Types1<T> type;
+};
+
+template <typename T1, typename T2, typename T3, typename T4, typename T5,
+    typename T6, typename T7, typename T8, typename T9, typename T10,
+    typename T11, typename T12, typename T13, typename T14, typename T15,
+    typename T16, typename T17, typename T18, typename T19, typename T20,
+    typename T21, typename T22, typename T23, typename T24, typename T25,
+    typename T26, typename T27, typename T28, typename T29, typename T30,
+    typename T31, typename T32, typename T33, typename T34, typename T35,
+    typename T36, typename T37, typename T38, typename T39, typename T40,
+    typename T41, typename T42, typename T43, typename T44, typename T45,
+    typename T46, typename T47, typename T48, typename T49, typename T50>
+struct TypeList<Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12, T13,
+    T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26, T27, T28,
+    T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40, T41, T42, T43,
+    T44, T45, T46, T47, T48, T49, T50> > {
+  typedef typename Types<T1, T2, T3, T4, T5, T6, T7, T8, T9, T10, T11, T12,
+      T13, T14, T15, T16, T17, T18, T19, T20, T21, T22, T23, T24, T25, T26,
+      T27, T28, T29, T30, T31, T32, T33, T34, T35, T36, T37, T38, T39, T40,
+      T41, T42, T43, T44, T45, T46, T47, T48, T49, T50>::type type;
+};
+
+#endif  // GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-type-util.h.pump b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-type-util.h.pump
new file mode 100644
index 0000000000..251fdf025b
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/include/gtest/internal/gtest-type-util.h.pump
@@ -0,0 +1,297 @@
+$$ -*- mode: c++; -*-
+$var n = 50  $$ Maximum length of type lists we want to support.
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Type utilities needed for implementing typed and type-parameterized
+// tests.  This file is generated by a SCRIPT.  DO NOT EDIT BY HAND!
+//
+// Currently we support at most $n types in a list, and at most $n
+// type-parameterized tests in one type-parameterized test case.
+// Please contact googletestframework@googlegroups.com if you need
+// more.
+
+#ifndef GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
+#define GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
+
+#include "gtest/internal/gtest-port.h"
+
+// #ifdef __GNUC__ is too general here.  It is possible to use gcc without using
+// libstdc++ (which is where cxxabi.h comes from).
+# if GTEST_HAS_CXXABI_H_
+#  include <cxxabi.h>
+# elif defined(__HP_aCC)
+#  include <acxx_demangle.h>
+# endif  // GTEST_HASH_CXXABI_H_
+
+namespace testing {
+namespace internal {
+
+// GetTypeName<T>() returns a human-readable name of type T.
+// NB: This function is also used in Google Mock, so don't move it inside of
+// the typed-test-only section below.
+template <typename T>
+std::string GetTypeName() {
+# if GTEST_HAS_RTTI
+
+  const char* const name = typeid(T).name();
+#  if GTEST_HAS_CXXABI_H_ || defined(__HP_aCC)
+  int status = 0;
+  // gcc's implementation of typeid(T).name() mangles the type name,
+  // so we have to demangle it.
+#   if GTEST_HAS_CXXABI_H_
+  using abi::__cxa_demangle;
+#   endif  // GTEST_HAS_CXXABI_H_
+  char* const readable_name = __cxa_demangle(name, 0, 0, &status);
+  const std::string name_str(status == 0 ? readable_name : name);
+  free(readable_name);
+  return name_str;
+#  else
+  return name;
+#  endif  // GTEST_HAS_CXXABI_H_ || __HP_aCC
+
+# else
+
+  return "<type>";
+
+# endif  // GTEST_HAS_RTTI
+}
+
+#if GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+// AssertyTypeEq<T1, T2>::type is defined iff T1 and T2 are the same
+// type.  This can be used as a compile-time assertion to ensure that
+// two types are equal.
+
+template <typename T1, typename T2>
+struct AssertTypeEq;
+
+template <typename T>
+struct AssertTypeEq<T, T> {
+  typedef bool type;
+};
+
+// A unique type used as the default value for the arguments of class
+// template Types.  This allows us to simulate variadic templates
+// (e.g. Types<int>, Type<int, double>, and etc), which C++ doesn't
+// support directly.
+struct None {};
+
+// The following family of struct and struct templates are used to
+// represent type lists.  In particular, TypesN<T1, T2, ..., TN>
+// represents a type list with N types (T1, T2, ..., and TN) in it.
+// Except for Types0, every struct in the family has two member types:
+// Head for the first type in the list, and Tail for the rest of the
+// list.
+
+// The empty type list.
+struct Types0 {};
+
+// Type lists of length 1, 2, 3, and so on.
+
+template <typename T1>
+struct Types1 {
+  typedef T1 Head;
+  typedef Types0 Tail;
+};
+
+$range i 2..n
+
+$for i [[
+$range j 1..i
+$range k 2..i
+template <$for j, [[typename T$j]]>
+struct Types$i {
+  typedef T1 Head;
+  typedef Types$(i-1)<$for k, [[T$k]]> Tail;
+};
+
+
+]]
+
+}  // namespace internal
+
+// We don't want to require the users to write TypesN<...> directly,
+// as that would require them to count the length.  Types<...> is much
+// easier to write, but generates horrible messages when there is a
+// compiler error, as gcc insists on printing out each template
+// argument, even if it has the default value (this means Types<int>
+// will appear as Types<int, None, None, ..., None> in the compiler
+// errors).
+//
+// Our solution is to combine the best part of the two approaches: a
+// user would write Types<T1, ..., TN>, and Google Test will translate
+// that to TypesN<T1, ..., TN> internally to make error messages
+// readable.  The translation is done by the 'type' member of the
+// Types template.
+
+$range i 1..n
+template <$for i, [[typename T$i = internal::None]]>
+struct Types {
+  typedef internal::Types$n<$for i, [[T$i]]> type;
+};
+
+template <>
+struct Types<$for i, [[internal::None]]> {
+  typedef internal::Types0 type;
+};
+
+$range i 1..n-1
+$for i [[
+$range j 1..i
+$range k i+1..n
+template <$for j, [[typename T$j]]>
+struct Types<$for j, [[T$j]]$for k[[, internal::None]]> {
+  typedef internal::Types$i<$for j, [[T$j]]> type;
+};
+
+]]
+
+namespace internal {
+
+# define GTEST_TEMPLATE_ template <typename T> class
+
+// The template "selector" struct TemplateSel<Tmpl> is used to
+// represent Tmpl, which must be a class template with one type
+// parameter, as a type.  TemplateSel<Tmpl>::Bind<T>::type is defined
+// as the type Tmpl<T>.  This allows us to actually instantiate the
+// template "selected" by TemplateSel<Tmpl>.
+//
+// This trick is necessary for simulating typedef for class templates,
+// which C++ doesn't support directly.
+template <GTEST_TEMPLATE_ Tmpl>
+struct TemplateSel {
+  template <typename T>
+  struct Bind {
+    typedef Tmpl<T> type;
+  };
+};
+
+# define GTEST_BIND_(TmplSel, T) \
+  TmplSel::template Bind<T>::type
+
+// A unique struct template used as the default value for the
+// arguments of class template Templates.  This allows us to simulate
+// variadic templates (e.g. Templates<int>, Templates<int, double>,
+// and etc), which C++ doesn't support directly.
+template <typename T>
+struct NoneT {};
+
+// The following family of struct and struct templates are used to
+// represent template lists.  In particular, TemplatesN<T1, T2, ...,
+// TN> represents a list of N templates (T1, T2, ..., and TN).  Except
+// for Templates0, every struct in the family has two member types:
+// Head for the selector of the first template in the list, and Tail
+// for the rest of the list.
+
+// The empty template list.
+struct Templates0 {};
+
+// Template lists of length 1, 2, 3, and so on.
+
+template <GTEST_TEMPLATE_ T1>
+struct Templates1 {
+  typedef TemplateSel<T1> Head;
+  typedef Templates0 Tail;
+};
+
+$range i 2..n
+
+$for i [[
+$range j 1..i
+$range k 2..i
+template <$for j, [[GTEST_TEMPLATE_ T$j]]>
+struct Templates$i {
+  typedef TemplateSel<T1> Head;
+  typedef Templates$(i-1)<$for k, [[T$k]]> Tail;
+};
+
+
+]]
+
+// We don't want to require the users to write TemplatesN<...> directly,
+// as that would require them to count the length.  Templates<...> is much
+// easier to write, but generates horrible messages when there is a
+// compiler error, as gcc insists on printing out each template
+// argument, even if it has the default value (this means Templates<list>
+// will appear as Templates<list, NoneT, NoneT, ..., NoneT> in the compiler
+// errors).
+//
+// Our solution is to combine the best part of the two approaches: a
+// user would write Templates<T1, ..., TN>, and Google Test will translate
+// that to TemplatesN<T1, ..., TN> internally to make error messages
+// readable.  The translation is done by the 'type' member of the
+// Templates template.
+
+$range i 1..n
+template <$for i, [[GTEST_TEMPLATE_ T$i = NoneT]]>
+struct Templates {
+  typedef Templates$n<$for i, [[T$i]]> type;
+};
+
+template <>
+struct Templates<$for i, [[NoneT]]> {
+  typedef Templates0 type;
+};
+
+$range i 1..n-1
+$for i [[
+$range j 1..i
+$range k i+1..n
+template <$for j, [[GTEST_TEMPLATE_ T$j]]>
+struct Templates<$for j, [[T$j]]$for k[[, NoneT]]> {
+  typedef Templates$i<$for j, [[T$j]]> type;
+};
+
+]]
+
+// The TypeList template makes it possible to use either a single type
+// or a Types<...> list in TYPED_TEST_CASE() and
+// INSTANTIATE_TYPED_TEST_CASE_P().
+
+template <typename T>
+struct TypeList {
+  typedef Types1<T> type;
+};
+
+
+$range i 1..n
+template <$for i, [[typename T$i]]>
+struct TypeList<Types<$for i, [[T$i]]> > {
+  typedef typename Types<$for i, [[T$i]]>::type type;
+};
+
+#endif  // GTEST_HAS_TYPED_TEST || GTEST_HAS_TYPED_TEST_P
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_INCLUDE_GTEST_INTERNAL_GTEST_TYPE_UTIL_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/src/gtest-all.cc b/third_party/aom/third_party/googletest/src/googletest/src/gtest-all.cc
new file mode 100644
index 0000000000..0a9cee5223
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/src/gtest-all.cc
@@ -0,0 +1,48 @@
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: mheule@google.com (Markus Heule)
+//
+// Google C++ Testing Framework (Google Test)
+//
+// Sometimes it's desirable to build Google Test by compiling a single file.
+// This file serves this purpose.
+
+// This line ensures that gtest.h can be compiled on its own, even
+// when it's fused.
+#include "gtest/gtest.h"
+
+// The following lines pull in the real gtest *.cc files.
+#include "src/gtest.cc"
+#include "src/gtest-death-test.cc"
+#include "src/gtest-filepath.cc"
+#include "src/gtest-port.cc"
+#include "src/gtest-printers.cc"
+#include "src/gtest-test-part.cc"
+#include "src/gtest-typed-test.cc"
diff --git a/third_party/aom/third_party/googletest/src/googletest/src/gtest-death-test.cc b/third_party/aom/third_party/googletest/src/googletest/src/gtest-death-test.cc
new file mode 100644
index 0000000000..a01a369830
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/src/gtest-death-test.cc
@@ -0,0 +1,1342 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan), vladl@google.com (Vlad Losev)
+//
+// This file implements death tests.
+
+#include "gtest/gtest-death-test.h"
+#include "gtest/internal/gtest-port.h"
+#include "gtest/internal/custom/gtest.h"
+
+#if GTEST_HAS_DEATH_TEST
+
+# if GTEST_OS_MAC
+#  include <crt_externs.h>
+# endif  // GTEST_OS_MAC
+
+# include <errno.h>
+# include <fcntl.h>
+# include <limits.h>
+
+# if GTEST_OS_LINUX
+#  include <signal.h>
+# endif  // GTEST_OS_LINUX
+
+# include <stdarg.h>
+
+# if GTEST_OS_WINDOWS
+#  include <windows.h>
+# else
+#  include <sys/mman.h>
+#  include <sys/wait.h>
+# endif  // GTEST_OS_WINDOWS
+
+# if GTEST_OS_QNX
+#  include <spawn.h>
+# endif  // GTEST_OS_QNX
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+#include "gtest/gtest-message.h"
+#include "gtest/internal/gtest-string.h"
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick exists to
+// prevent the accidental inclusion of gtest-internal-inl.h in the
+// user's code.
+#define GTEST_IMPLEMENTATION_ 1
+#include "src/gtest-internal-inl.h"
+#undef GTEST_IMPLEMENTATION_
+
+namespace testing {
+
+// Constants.
+
+// The default death test style.
+static const char kDefaultDeathTestStyle[] = "fast";
+
+GTEST_DEFINE_string_(
+    death_test_style,
+    internal::StringFromGTestEnv("death_test_style", kDefaultDeathTestStyle),
+    "Indicates how to run a death test in a forked child process: "
+    "\"threadsafe\" (child process re-executes the test binary "
+    "from the beginning, running only the specific death test) or "
+    "\"fast\" (child process runs the death test immediately "
+    "after forking).");
+
+GTEST_DEFINE_bool_(
+    death_test_use_fork,
+    internal::BoolFromGTestEnv("death_test_use_fork", false),
+    "Instructs to use fork()/_exit() instead of clone() in death tests. "
+    "Ignored and always uses fork() on POSIX systems where clone() is not "
+    "implemented. Useful when running under valgrind or similar tools if "
+    "those do not support clone(). Valgrind 3.3.1 will just fail if "
+    "it sees an unsupported combination of clone() flags. "
+    "It is not recommended to use this flag w/o valgrind though it will "
+    "work in 99% of the cases. Once valgrind is fixed, this flag will "
+    "most likely be removed.");
+
+namespace internal {
+GTEST_DEFINE_string_(
+    internal_run_death_test, "",
+    "Indicates the file, line number, temporal index of "
+    "the single death test to run, and a file descriptor to "
+    "which a success code may be sent, all separated by "
+    "the '|' characters.  This flag is specified if and only if the current "
+    "process is a sub-process launched for running a thread-safe "
+    "death test.  FOR INTERNAL USE ONLY.");
+}  // namespace internal
+
+#if GTEST_HAS_DEATH_TEST
+
+namespace internal {
+
+// Valid only for fast death tests. Indicates the code is running in the
+// child process of a fast style death test.
+# if !GTEST_OS_WINDOWS
+static bool g_in_fast_death_test_child = false;
+# endif
+
+// Returns a Boolean value indicating whether the caller is currently
+// executing in the context of the death test child process.  Tools such as
+// Valgrind heap checkers may need this to modify their behavior in death
+// tests.  IMPORTANT: This is an internal utility.  Using it may break the
+// implementation of death tests.  User code MUST NOT use it.
+bool InDeathTestChild() {
+# if GTEST_OS_WINDOWS
+
+  // On Windows, death tests are thread-safe regardless of the value of the
+  // death_test_style flag.
+  return !GTEST_FLAG(internal_run_death_test).empty();
+
+# else
+
+  if (GTEST_FLAG(death_test_style) == "threadsafe")
+    return !GTEST_FLAG(internal_run_death_test).empty();
+  else
+    return g_in_fast_death_test_child;
+#endif
+}
+
+}  // namespace internal
+
+// ExitedWithCode constructor.
+ExitedWithCode::ExitedWithCode(int exit_code) : exit_code_(exit_code) {
+}
+
+// ExitedWithCode function-call operator.
+bool ExitedWithCode::operator()(int exit_status) const {
+# if GTEST_OS_WINDOWS
+
+  return exit_status == exit_code_;
+
+# else
+
+  return WIFEXITED(exit_status) && WEXITSTATUS(exit_status) == exit_code_;
+
+# endif  // GTEST_OS_WINDOWS
+}
+
+# if !GTEST_OS_WINDOWS
+// KilledBySignal constructor.
+KilledBySignal::KilledBySignal(int signum) : signum_(signum) {
+}
+
+// KilledBySignal function-call operator.
+bool KilledBySignal::operator()(int exit_status) const {
+#  if defined(GTEST_KILLED_BY_SIGNAL_OVERRIDE_)
+  {
+    bool result;
+    if (GTEST_KILLED_BY_SIGNAL_OVERRIDE_(signum_, exit_status, &result)) {
+      return result;
+    }
+  }
+#  endif  // defined(GTEST_KILLED_BY_SIGNAL_OVERRIDE_)
+  return WIFSIGNALED(exit_status) && WTERMSIG(exit_status) == signum_;
+}
+# endif  // !GTEST_OS_WINDOWS
+
+namespace internal {
+
+// Utilities needed for death tests.
+
+// Generates a textual description of a given exit code, in the format
+// specified by wait(2).
+static std::string ExitSummary(int exit_code) {
+  Message m;
+
+# if GTEST_OS_WINDOWS
+
+  m << "Exited with exit status " << exit_code;
+
+# else
+
+  if (WIFEXITED(exit_code)) {
+    m << "Exited with exit status " << WEXITSTATUS(exit_code);
+  } else if (WIFSIGNALED(exit_code)) {
+    m << "Terminated by signal " << WTERMSIG(exit_code);
+  }
+#  ifdef WCOREDUMP
+  if (WCOREDUMP(exit_code)) {
+    m << " (core dumped)";
+  }
+#  endif
+# endif  // GTEST_OS_WINDOWS
+
+  return m.GetString();
+}
+
+// Returns true if exit_status describes a process that was terminated
+// by a signal, or exited normally with a nonzero exit code.
+bool ExitedUnsuccessfully(int exit_status) {
+  return !ExitedWithCode(0)(exit_status);
+}
+
+# if !GTEST_OS_WINDOWS
+// Generates a textual failure message when a death test finds more than
+// one thread running, or cannot determine the number of threads, prior
+// to executing the given statement.  It is the responsibility of the
+// caller not to pass a thread_count of 1.
+static std::string DeathTestThreadWarning(size_t thread_count) {
+  Message msg;
+  msg << "Death tests use fork(), which is unsafe particularly"
+      << " in a threaded context. For this test, " << GTEST_NAME_ << " ";
+  if (thread_count == 0)
+    msg << "couldn't detect the number of threads.";
+  else
+    msg << "detected " << thread_count << " threads.";
+  return msg.GetString();
+}
+# endif  // !GTEST_OS_WINDOWS
+
+// Flag characters for reporting a death test that did not die.
+static const char kDeathTestLived = 'L';
+static const char kDeathTestReturned = 'R';
+static const char kDeathTestThrew = 'T';
+static const char kDeathTestInternalError = 'I';
+
+// An enumeration describing all of the possible ways that a death test can
+// conclude.  DIED means that the process died while executing the test
+// code; LIVED means that process lived beyond the end of the test code;
+// RETURNED means that the test statement attempted to execute a return
+// statement, which is not allowed; THREW means that the test statement
+// returned control by throwing an exception.  IN_PROGRESS means the test
+// has not yet concluded.
+// TODO(vladl@google.com): Unify names and possibly values for
+// AbortReason, DeathTestOutcome, and flag characters above.
+enum DeathTestOutcome { IN_PROGRESS, DIED, LIVED, RETURNED, THREW };
+
+// Routine for aborting the program which is safe to call from an
+// exec-style death test child process, in which case the error
+// message is propagated back to the parent process.  Otherwise, the
+// message is simply printed to stderr.  In either case, the program
+// then exits with status 1.
+void DeathTestAbort(const std::string& message) {
+  // On a POSIX system, this function may be called from a threadsafe-style
+  // death test child process, which operates on a very small stack.  Use
+  // the heap for any additional non-minuscule memory requirements.
+  const InternalRunDeathTestFlag* const flag =
+      GetUnitTestImpl()->internal_run_death_test_flag();
+  if (flag != NULL) {
+    FILE* parent = posix::FDOpen(flag->write_fd(), "w");
+    fputc(kDeathTestInternalError, parent);
+    fprintf(parent, "%s", message.c_str());
+    fflush(parent);
+    _exit(1);
+  } else {
+    fprintf(stderr, "%s", message.c_str());
+    fflush(stderr);
+    posix::Abort();
+  }
+}
+
+// A replacement for CHECK that calls DeathTestAbort if the assertion
+// fails.
+# define GTEST_DEATH_TEST_CHECK_(expression) \
+  do { \
+    if (!::testing::internal::IsTrue(expression)) { \
+      DeathTestAbort( \
+          ::std::string("CHECK failed: File ") + __FILE__ +  ", line " \
+          + ::testing::internal::StreamableToString(__LINE__) + ": " \
+          + #expression); \
+    } \
+  } while (::testing::internal::AlwaysFalse())
+
+// This macro is similar to GTEST_DEATH_TEST_CHECK_, but it is meant for
+// evaluating any system call that fulfills two conditions: it must return
+// -1 on failure, and set errno to EINTR when it is interrupted and
+// should be tried again.  The macro expands to a loop that repeatedly
+// evaluates the expression as long as it evaluates to -1 and sets
+// errno to EINTR.  If the expression evaluates to -1 but errno is
+// something other than EINTR, DeathTestAbort is called.
+# define GTEST_DEATH_TEST_CHECK_SYSCALL_(expression) \
+  do { \
+    int gtest_retval; \
+    do { \
+      gtest_retval = (expression); \
+    } while (gtest_retval == -1 && errno == EINTR); \
+    if (gtest_retval == -1) { \
+      DeathTestAbort( \
+          ::std::string("CHECK failed: File ") + __FILE__ + ", line " \
+          + ::testing::internal::StreamableToString(__LINE__) + ": " \
+          + #expression + " != -1"); \
+    } \
+  } while (::testing::internal::AlwaysFalse())
+
+// Returns the message describing the last system error in errno.
+std::string GetLastErrnoDescription() {
+    return errno == 0 ? "" : posix::StrError(errno);
+}
+
+// This is called from a death test parent process to read a failure
+// message from the death test child process and log it with the FATAL
+// severity. On Windows, the message is read from a pipe handle. On other
+// platforms, it is read from a file descriptor.
+static void FailFromInternalError(int fd) {
+  Message error;
+  char buffer[256];
+  int num_read;
+
+  do {
+    while ((num_read = posix::Read(fd, buffer, 255)) > 0) {
+      buffer[num_read] = '\0';
+      error << buffer;
+    }
+  } while (num_read == -1 && errno == EINTR);
+
+  if (num_read == 0) {
+    GTEST_LOG_(FATAL) << error.GetString();
+  } else {
+    const int last_error = errno;
+    GTEST_LOG_(FATAL) << "Error while reading death test internal: "
+                      << GetLastErrnoDescription() << " [" << last_error << "]";
+  }
+}
+
+// Death test constructor.  Increments the running death test count
+// for the current test.
+DeathTest::DeathTest() {
+  TestInfo* const info = GetUnitTestImpl()->current_test_info();
+  if (info == NULL) {
+    DeathTestAbort("Cannot run a death test outside of a TEST or "
+                   "TEST_F construct");
+  }
+}
+
+// Creates and returns a death test by dispatching to the current
+// death test factory.
+bool DeathTest::Create(const char* statement, const RE* regex,
+                       const char* file, int line, DeathTest** test) {
+  return GetUnitTestImpl()->death_test_factory()->Create(
+      statement, regex, file, line, test);
+}
+
+const char* DeathTest::LastMessage() {
+  return last_death_test_message_.c_str();
+}
+
+void DeathTest::set_last_death_test_message(const std::string& message) {
+  last_death_test_message_ = message;
+}
+
+std::string DeathTest::last_death_test_message_;
+
+// Provides cross platform implementation for some death functionality.
+class DeathTestImpl : public DeathTest {
+ protected:
+  DeathTestImpl(const char* a_statement, const RE* a_regex)
+      : statement_(a_statement),
+        regex_(a_regex),
+        spawned_(false),
+        status_(-1),
+        outcome_(IN_PROGRESS),
+        read_fd_(-1),
+        write_fd_(-1) {}
+
+  // read_fd_ is expected to be closed and cleared by a derived class.
+  ~DeathTestImpl() { GTEST_DEATH_TEST_CHECK_(read_fd_ == -1); }
+
+  void Abort(AbortReason reason);
+  virtual bool Passed(bool status_ok);
+
+  const char* statement() const { return statement_; }
+  const RE* regex() const { return regex_; }
+  bool spawned() const { return spawned_; }
+  void set_spawned(bool is_spawned) { spawned_ = is_spawned; }
+  int status() const { return status_; }
+  void set_status(int a_status) { status_ = a_status; }
+  DeathTestOutcome outcome() const { return outcome_; }
+  void set_outcome(DeathTestOutcome an_outcome) { outcome_ = an_outcome; }
+  int read_fd() const { return read_fd_; }
+  void set_read_fd(int fd) { read_fd_ = fd; }
+  int write_fd() const { return write_fd_; }
+  void set_write_fd(int fd) { write_fd_ = fd; }
+
+  // Called in the parent process only. Reads the result code of the death
+  // test child process via a pipe, interprets it to set the outcome_
+  // member, and closes read_fd_.  Outputs diagnostics and terminates in
+  // case of unexpected codes.
+  void ReadAndInterpretStatusByte();
+
+ private:
+  // The textual content of the code this object is testing.  This class
+  // doesn't own this string and should not attempt to delete it.
+  const char* const statement_;
+  // The regular expression which test output must match.  DeathTestImpl
+  // doesn't own this object and should not attempt to delete it.
+  const RE* const regex_;
+  // True if the death test child process has been successfully spawned.
+  bool spawned_;
+  // The exit status of the child process.
+  int status_;
+  // How the death test concluded.
+  DeathTestOutcome outcome_;
+  // Descriptor to the read end of the pipe to the child process.  It is
+  // always -1 in the child process.  The child keeps its write end of the
+  // pipe in write_fd_.
+  int read_fd_;
+  // Descriptor to the child's write end of the pipe to the parent process.
+  // It is always -1 in the parent process.  The parent keeps its end of the
+  // pipe in read_fd_.
+  int write_fd_;
+};
+
+// Called in the parent process only. Reads the result code of the death
+// test child process via a pipe, interprets it to set the outcome_
+// member, and closes read_fd_.  Outputs diagnostics and terminates in
+// case of unexpected codes.
+void DeathTestImpl::ReadAndInterpretStatusByte() {
+  char flag;
+  int bytes_read;
+
+  // The read() here blocks until data is available (signifying the
+  // failure of the death test) or until the pipe is closed (signifying
+  // its success), so it's okay to call this in the parent before
+  // the child process has exited.
+  do {
+    bytes_read = posix::Read(read_fd(), &flag, 1);
+  } while (bytes_read == -1 && errno == EINTR);
+
+  if (bytes_read == 0) {
+    set_outcome(DIED);
+  } else if (bytes_read == 1) {
+    switch (flag) {
+      case kDeathTestReturned:
+        set_outcome(RETURNED);
+        break;
+      case kDeathTestThrew:
+        set_outcome(THREW);
+        break;
+      case kDeathTestLived:
+        set_outcome(LIVED);
+        break;
+      case kDeathTestInternalError:
+        FailFromInternalError(read_fd());  // Does not return.
+        break;
+      default:
+        GTEST_LOG_(FATAL) << "Death test child process reported "
+                          << "unexpected status byte ("
+                          << static_cast<unsigned int>(flag) << ")";
+    }
+  } else {
+    GTEST_LOG_(FATAL) << "Read from death test child process failed: "
+                      << GetLastErrnoDescription();
+  }
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(posix::Close(read_fd()));
+  set_read_fd(-1);
+}
+
+// Signals that the death test code which should have exited, didn't.
+// Should be called only in a death test child process.
+// Writes a status byte to the child's status file descriptor, then
+// calls _exit(1).
+void DeathTestImpl::Abort(AbortReason reason) {
+  // The parent process considers the death test to be a failure if
+  // it finds any data in our pipe.  So, here we write a single flag byte
+  // to the pipe, then exit.
+  const char status_ch =
+      reason == TEST_DID_NOT_DIE ? kDeathTestLived :
+      reason == TEST_THREW_EXCEPTION ? kDeathTestThrew : kDeathTestReturned;
+
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(posix::Write(write_fd(), &status_ch, 1));
+  // We are leaking the descriptor here because on some platforms (i.e.,
+  // when built as Windows DLL), destructors of global objects will still
+  // run after calling _exit(). On such systems, write_fd_ will be
+  // indirectly closed from the destructor of UnitTestImpl, causing double
+  // close if it is also closed here. On debug configurations, double close
+  // may assert. As there are no in-process buffers to flush here, we are
+  // relying on the OS to close the descriptor after the process terminates
+  // when the destructors are not run.
+  _exit(1);  // Exits w/o any normal exit hooks (we were supposed to crash)
+}
+
+// Returns an indented copy of stderr output for a death test.
+// This makes distinguishing death test output lines from regular log lines
+// much easier.
+static ::std::string FormatDeathTestOutput(const ::std::string& output) {
+  ::std::string ret;
+  for (size_t at = 0; ; ) {
+    const size_t line_end = output.find('\n', at);
+    ret += "[  DEATH   ] ";
+    if (line_end == ::std::string::npos) {
+      ret += output.substr(at);
+      break;
+    }
+    ret += output.substr(at, line_end + 1 - at);
+    at = line_end + 1;
+  }
+  return ret;
+}
+
+// Assesses the success or failure of a death test, using both private
+// members which have previously been set, and one argument:
+//
+// Private data members:
+//   outcome:  An enumeration describing how the death test
+//             concluded: DIED, LIVED, THREW, or RETURNED.  The death test
+//             fails in the latter three cases.
+//   status:   The exit status of the child process. On *nix, it is in the
+//             in the format specified by wait(2). On Windows, this is the
+//             value supplied to the ExitProcess() API or a numeric code
+//             of the exception that terminated the program.
+//   regex:    A regular expression object to be applied to
+//             the test's captured standard error output; the death test
+//             fails if it does not match.
+//
+// Argument:
+//   status_ok: true if exit_status is acceptable in the context of
+//              this particular death test, which fails if it is false
+//
+// Returns true iff all of the above conditions are met.  Otherwise, the
+// first failing condition, in the order given above, is the one that is
+// reported. Also sets the last death test message string.
+bool DeathTestImpl::Passed(bool status_ok) {
+  if (!spawned())
+    return false;
+
+  const std::string error_message = GetCapturedStderr();
+
+  bool success = false;
+  Message buffer;
+
+  buffer << "Death test: " << statement() << "\n";
+  switch (outcome()) {
+    case LIVED:
+      buffer << "    Result: failed to die.\n"
+             << " Error msg:\n" << FormatDeathTestOutput(error_message);
+      break;
+    case THREW:
+      buffer << "    Result: threw an exception.\n"
+             << " Error msg:\n" << FormatDeathTestOutput(error_message);
+      break;
+    case RETURNED:
+      buffer << "    Result: illegal return in test statement.\n"
+             << " Error msg:\n" << FormatDeathTestOutput(error_message);
+      break;
+    case DIED:
+      if (status_ok) {
+        const bool matched = RE::PartialMatch(error_message.c_str(), *regex());
+        if (matched) {
+          success = true;
+        } else {
+          buffer << "    Result: died but not with expected error.\n"
+                 << "  Expected: " << regex()->pattern() << "\n"
+                 << "Actual msg:\n" << FormatDeathTestOutput(error_message);
+        }
+      } else {
+        buffer << "    Result: died but not with expected exit code:\n"
+               << "            " << ExitSummary(status()) << "\n"
+               << "Actual msg:\n" << FormatDeathTestOutput(error_message);
+      }
+      break;
+    case IN_PROGRESS:
+    default:
+      GTEST_LOG_(FATAL)
+          << "DeathTest::Passed somehow called before conclusion of test";
+  }
+
+  DeathTest::set_last_death_test_message(buffer.GetString());
+  return success;
+}
+
+# if GTEST_OS_WINDOWS
+// WindowsDeathTest implements death tests on Windows. Due to the
+// specifics of starting new processes on Windows, death tests there are
+// always threadsafe, and Google Test considers the
+// --gtest_death_test_style=fast setting to be equivalent to
+// --gtest_death_test_style=threadsafe there.
+//
+// A few implementation notes:  Like the Linux version, the Windows
+// implementation uses pipes for child-to-parent communication. But due to
+// the specifics of pipes on Windows, some extra steps are required:
+//
+// 1. The parent creates a communication pipe and stores handles to both
+//    ends of it.
+// 2. The parent starts the child and provides it with the information
+//    necessary to acquire the handle to the write end of the pipe.
+// 3. The child acquires the write end of the pipe and signals the parent
+//    using a Windows event.
+// 4. Now the parent can release the write end of the pipe on its side. If
+//    this is done before step 3, the object's reference count goes down to
+//    0 and it is destroyed, preventing the child from acquiring it. The
+//    parent now has to release it, or read operations on the read end of
+//    the pipe will not return when the child terminates.
+// 5. The parent reads child's output through the pipe (outcome code and
+//    any possible error messages) from the pipe, and its stderr and then
+//    determines whether to fail the test.
+//
+// Note: to distinguish Win32 API calls from the local method and function
+// calls, the former are explicitly resolved in the global namespace.
+//
+class WindowsDeathTest : public DeathTestImpl {
+ public:
+  WindowsDeathTest(const char* a_statement,
+                   const RE* a_regex,
+                   const char* file,
+                   int line)
+      : DeathTestImpl(a_statement, a_regex), file_(file), line_(line) {}
+
+  // All of these virtual functions are inherited from DeathTest.
+  virtual int Wait();
+  virtual TestRole AssumeRole();
+
+ private:
+  // The name of the file in which the death test is located.
+  const char* const file_;
+  // The line number on which the death test is located.
+  const int line_;
+  // Handle to the write end of the pipe to the child process.
+  AutoHandle write_handle_;
+  // Child process handle.
+  AutoHandle child_handle_;
+  // Event the child process uses to signal the parent that it has
+  // acquired the handle to the write end of the pipe. After seeing this
+  // event the parent can release its own handles to make sure its
+  // ReadFile() calls return when the child terminates.
+  AutoHandle event_handle_;
+};
+
+// Waits for the child in a death test to exit, returning its exit
+// status, or 0 if no child process exists.  As a side effect, sets the
+// outcome data member.
+int WindowsDeathTest::Wait() {
+  if (!spawned())
+    return 0;
+
+  // Wait until the child either signals that it has acquired the write end
+  // of the pipe or it dies.
+  const HANDLE wait_handles[2] = { child_handle_.Get(), event_handle_.Get() };
+  switch (::WaitForMultipleObjects(2,
+                                   wait_handles,
+                                   FALSE,  // Waits for any of the handles.
+                                   INFINITE)) {
+    case WAIT_OBJECT_0:
+    case WAIT_OBJECT_0 + 1:
+      break;
+    default:
+      GTEST_DEATH_TEST_CHECK_(false);  // Should not get here.
+  }
+
+  // The child has acquired the write end of the pipe or exited.
+  // We release the handle on our side and continue.
+  write_handle_.Reset();
+  event_handle_.Reset();
+
+  ReadAndInterpretStatusByte();
+
+  // Waits for the child process to exit if it haven't already. This
+  // returns immediately if the child has already exited, regardless of
+  // whether previous calls to WaitForMultipleObjects synchronized on this
+  // handle or not.
+  GTEST_DEATH_TEST_CHECK_(
+      WAIT_OBJECT_0 == ::WaitForSingleObject(child_handle_.Get(),
+                                             INFINITE));
+  DWORD status_code;
+  GTEST_DEATH_TEST_CHECK_(
+      ::GetExitCodeProcess(child_handle_.Get(), &status_code) != FALSE);
+  child_handle_.Reset();
+  set_status(static_cast<int>(status_code));
+  return status();
+}
+
+// The AssumeRole process for a Windows death test.  It creates a child
+// process with the same executable as the current process to run the
+// death test.  The child process is given the --gtest_filter and
+// --gtest_internal_run_death_test flags such that it knows to run the
+// current death test only.
+DeathTest::TestRole WindowsDeathTest::AssumeRole() {
+  const UnitTestImpl* const impl = GetUnitTestImpl();
+  const InternalRunDeathTestFlag* const flag =
+      impl->internal_run_death_test_flag();
+  const TestInfo* const info = impl->current_test_info();
+  const int death_test_index = info->result()->death_test_count();
+
+  if (flag != NULL) {
+    // ParseInternalRunDeathTestFlag() has performed all the necessary
+    // processing.
+    set_write_fd(flag->write_fd());
+    return EXECUTE_TEST;
+  }
+
+  // WindowsDeathTest uses an anonymous pipe to communicate results of
+  // a death test.
+  SECURITY_ATTRIBUTES handles_are_inheritable = {
+    sizeof(SECURITY_ATTRIBUTES), NULL, TRUE };
+  HANDLE read_handle, write_handle;
+  GTEST_DEATH_TEST_CHECK_(
+      ::CreatePipe(&read_handle, &write_handle, &handles_are_inheritable,
+                   0)  // Default buffer size.
+      != FALSE);
+  set_read_fd(::_open_osfhandle(reinterpret_cast<intptr_t>(read_handle),
+                                O_RDONLY));
+  write_handle_.Reset(write_handle);
+  event_handle_.Reset(::CreateEvent(
+      &handles_are_inheritable,
+      TRUE,    // The event will automatically reset to non-signaled state.
+      FALSE,   // The initial state is non-signalled.
+      NULL));  // The even is unnamed.
+  GTEST_DEATH_TEST_CHECK_(event_handle_.Get() != NULL);
+  const std::string filter_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kFilterFlag + "=" +
+      info->test_case_name() + "." + info->name();
+  const std::string internal_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kInternalRunDeathTestFlag +
+      "=" + file_ + "|" + StreamableToString(line_) + "|" +
+      StreamableToString(death_test_index) + "|" +
+      StreamableToString(static_cast<unsigned int>(::GetCurrentProcessId())) +
+      // size_t has the same width as pointers on both 32-bit and 64-bit
+      // Windows platforms.
+      // See http://msdn.microsoft.com/en-us/library/tcxf1dw6.aspx.
+      "|" + StreamableToString(reinterpret_cast<size_t>(write_handle)) +
+      "|" + StreamableToString(reinterpret_cast<size_t>(event_handle_.Get()));
+
+  char executable_path[_MAX_PATH + 1];  // NOLINT
+  GTEST_DEATH_TEST_CHECK_(
+      _MAX_PATH + 1 != ::GetModuleFileNameA(NULL,
+                                            executable_path,
+                                            _MAX_PATH));
+
+  std::string command_line =
+      std::string(::GetCommandLineA()) + " " + filter_flag + " \"" +
+      internal_flag + "\"";
+
+  DeathTest::set_last_death_test_message("");
+
+  CaptureStderr();
+  // Flush the log buffers since the log streams are shared with the child.
+  FlushInfoLog();
+
+  // The child process will share the standard handles with the parent.
+  STARTUPINFOA startup_info;
+  memset(&startup_info, 0, sizeof(STARTUPINFO));
+  startup_info.dwFlags = STARTF_USESTDHANDLES;
+  startup_info.hStdInput = ::GetStdHandle(STD_INPUT_HANDLE);
+  startup_info.hStdOutput = ::GetStdHandle(STD_OUTPUT_HANDLE);
+  startup_info.hStdError = ::GetStdHandle(STD_ERROR_HANDLE);
+
+  PROCESS_INFORMATION process_info;
+  GTEST_DEATH_TEST_CHECK_(::CreateProcessA(
+      executable_path,
+      const_cast<char*>(command_line.c_str()),
+      NULL,   // Retuned process handle is not inheritable.
+      NULL,   // Retuned thread handle is not inheritable.
+      TRUE,   // Child inherits all inheritable handles (for write_handle_).
+      0x0,    // Default creation flags.
+      NULL,   // Inherit the parent's environment.
+      UnitTest::GetInstance()->original_working_dir(),
+      &startup_info,
+      &process_info) != FALSE);
+  child_handle_.Reset(process_info.hProcess);
+  ::CloseHandle(process_info.hThread);
+  set_spawned(true);
+  return OVERSEE_TEST;
+}
+# else  // We are not on Windows.
+
+// ForkingDeathTest provides implementations for most of the abstract
+// methods of the DeathTest interface.  Only the AssumeRole method is
+// left undefined.
+class ForkingDeathTest : public DeathTestImpl {
+ public:
+  ForkingDeathTest(const char* statement, const RE* regex);
+
+  // All of these virtual functions are inherited from DeathTest.
+  virtual int Wait();
+
+ protected:
+  void set_child_pid(pid_t child_pid) { child_pid_ = child_pid; }
+
+ private:
+  // PID of child process during death test; 0 in the child process itself.
+  pid_t child_pid_;
+};
+
+// Constructs a ForkingDeathTest.
+ForkingDeathTest::ForkingDeathTest(const char* a_statement, const RE* a_regex)
+    : DeathTestImpl(a_statement, a_regex),
+      child_pid_(-1) {}
+
+// Waits for the child in a death test to exit, returning its exit
+// status, or 0 if no child process exists.  As a side effect, sets the
+// outcome data member.
+int ForkingDeathTest::Wait() {
+  if (!spawned())
+    return 0;
+
+  ReadAndInterpretStatusByte();
+
+  int status_value;
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(waitpid(child_pid_, &status_value, 0));
+  set_status(status_value);
+  return status_value;
+}
+
+// A concrete death test class that forks, then immediately runs the test
+// in the child process.
+class NoExecDeathTest : public ForkingDeathTest {
+ public:
+  NoExecDeathTest(const char* a_statement, const RE* a_regex) :
+      ForkingDeathTest(a_statement, a_regex) { }
+  virtual TestRole AssumeRole();
+};
+
+// The AssumeRole process for a fork-and-run death test.  It implements a
+// straightforward fork, with a simple pipe to transmit the status byte.
+DeathTest::TestRole NoExecDeathTest::AssumeRole() {
+  const size_t thread_count = GetThreadCount();
+  if (thread_count != 1) {
+    GTEST_LOG_(WARNING) << DeathTestThreadWarning(thread_count);
+  }
+
+  int pipe_fd[2];
+  GTEST_DEATH_TEST_CHECK_(pipe(pipe_fd) != -1);
+
+  DeathTest::set_last_death_test_message("");
+  CaptureStderr();
+  // When we fork the process below, the log file buffers are copied, but the
+  // file descriptors are shared.  We flush all log files here so that closing
+  // the file descriptors in the child process doesn't throw off the
+  // synchronization between descriptors and buffers in the parent process.
+  // This is as close to the fork as possible to avoid a race condition in case
+  // there are multiple threads running before the death test, and another
+  // thread writes to the log file.
+  FlushInfoLog();
+
+  const pid_t child_pid = fork();
+  GTEST_DEATH_TEST_CHECK_(child_pid != -1);
+  set_child_pid(child_pid);
+  if (child_pid == 0) {
+    GTEST_DEATH_TEST_CHECK_SYSCALL_(close(pipe_fd[0]));
+    set_write_fd(pipe_fd[1]);
+    // Redirects all logging to stderr in the child process to prevent
+    // concurrent writes to the log files.  We capture stderr in the parent
+    // process and append the child process' output to a log.
+    LogToStderr();
+    // Event forwarding to the listeners of event listener API mush be shut
+    // down in death test subprocesses.
+    GetUnitTestImpl()->listeners()->SuppressEventForwarding();
+    g_in_fast_death_test_child = true;
+    return EXECUTE_TEST;
+  } else {
+    GTEST_DEATH_TEST_CHECK_SYSCALL_(close(pipe_fd[1]));
+    set_read_fd(pipe_fd[0]);
+    set_spawned(true);
+    return OVERSEE_TEST;
+  }
+}
+
+// A concrete death test class that forks and re-executes the main
+// program from the beginning, with command-line flags set that cause
+// only this specific death test to be run.
+class ExecDeathTest : public ForkingDeathTest {
+ public:
+  ExecDeathTest(const char* a_statement, const RE* a_regex,
+                const char* file, int line) :
+      ForkingDeathTest(a_statement, a_regex), file_(file), line_(line) { }
+  virtual TestRole AssumeRole();
+ private:
+  static ::std::vector<testing::internal::string>
+  GetArgvsForDeathTestChildProcess() {
+    ::std::vector<testing::internal::string> args = GetInjectableArgvs();
+#  if defined(GTEST_EXTRA_DEATH_TEST_COMMAND_LINE_ARGS_)
+    ::std::vector<testing::internal::string> extra_args =
+        GTEST_EXTRA_DEATH_TEST_COMMAND_LINE_ARGS_();
+    args.insert(args.end(), extra_args.begin(), extra_args.end());
+#  endif  // defined(GTEST_EXTRA_DEATH_TEST_COMMAND_LINE_ARGS_)
+    return args;
+  }
+  // The name of the file in which the death test is located.
+  const char* const file_;
+  // The line number on which the death test is located.
+  const int line_;
+};
+
+// Utility class for accumulating command-line arguments.
+class Arguments {
+ public:
+  Arguments() {
+    args_.push_back(NULL);
+  }
+
+  ~Arguments() {
+    for (std::vector<char*>::iterator i = args_.begin(); i != args_.end();
+         ++i) {
+      free(*i);
+    }
+  }
+  void AddArgument(const char* argument) {
+    args_.insert(args_.end() - 1, posix::StrDup(argument));
+  }
+
+  template <typename Str>
+  void AddArguments(const ::std::vector<Str>& arguments) {
+    for (typename ::std::vector<Str>::const_iterator i = arguments.begin();
+         i != arguments.end();
+         ++i) {
+      args_.insert(args_.end() - 1, posix::StrDup(i->c_str()));
+    }
+  }
+  char* const* Argv() {
+    return &args_[0];
+  }
+
+ private:
+  std::vector<char*> args_;
+};
+
+// A struct that encompasses the arguments to the child process of a
+// threadsafe-style death test process.
+struct ExecDeathTestArgs {
+  char* const* argv;  // Command-line arguments for the child's call to exec
+  int close_fd;       // File descriptor to close; the read end of a pipe
+};
+
+#  if GTEST_OS_MAC
+inline char** GetEnviron() {
+  // When Google Test is built as a framework on MacOS X, the environ variable
+  // is unavailable. Apple's documentation (man environ) recommends using
+  // _NSGetEnviron() instead.
+  return *_NSGetEnviron();
+}
+#  else
+// Some POSIX platforms expect you to declare environ. extern "C" makes
+// it reside in the global namespace.
+extern "C" char** environ;
+inline char** GetEnviron() { return environ; }
+#  endif  // GTEST_OS_MAC
+
+#  if !GTEST_OS_QNX
+// The main function for a threadsafe-style death test child process.
+// This function is called in a clone()-ed process and thus must avoid
+// any potentially unsafe operations like malloc or libc functions.
+static int ExecDeathTestChildMain(void* child_arg) {
+  ExecDeathTestArgs* const args = static_cast<ExecDeathTestArgs*>(child_arg);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(close(args->close_fd));
+
+  // We need to execute the test program in the same environment where
+  // it was originally invoked.  Therefore we change to the original
+  // working directory first.
+  const char* const original_dir =
+      UnitTest::GetInstance()->original_working_dir();
+  // We can safely call chdir() as it's a direct system call.
+  if (chdir(original_dir) != 0) {
+    DeathTestAbort(std::string("chdir(\"") + original_dir + "\") failed: " +
+                   GetLastErrnoDescription());
+    return EXIT_FAILURE;
+  }
+
+  // We can safely call execve() as it's a direct system call.  We
+  // cannot use execvp() as it's a libc function and thus potentially
+  // unsafe.  Since execve() doesn't search the PATH, the user must
+  // invoke the test program via a valid path that contains at least
+  // one path separator.
+  execve(args->argv[0], args->argv, GetEnviron());
+  DeathTestAbort(std::string("execve(") + args->argv[0] + ", ...) in " +
+                 original_dir + " failed: " +
+                 GetLastErrnoDescription());
+  return EXIT_FAILURE;
+}
+#  endif  // !GTEST_OS_QNX
+
+// Two utility routines that together determine the direction the stack
+// grows.
+// This could be accomplished more elegantly by a single recursive
+// function, but we want to guard against the unlikely possibility of
+// a smart compiler optimizing the recursion away.
+//
+// GTEST_NO_INLINE_ is required to prevent GCC 4.6 from inlining
+// StackLowerThanAddress into StackGrowsDown, which then doesn't give
+// correct answer.
+void StackLowerThanAddress(const void* ptr, bool* result) GTEST_NO_INLINE_;
+void StackLowerThanAddress(const void* ptr, bool* result) {
+  int dummy;
+  *result = (&dummy < ptr);
+}
+
+// Make sure AddressSanitizer does not tamper with the stack here.
+GTEST_ATTRIBUTE_NO_SANITIZE_ADDRESS_
+bool StackGrowsDown() {
+  int dummy;
+  bool result;
+  StackLowerThanAddress(&dummy, &result);
+  return result;
+}
+
+// Spawns a child process with the same executable as the current process in
+// a thread-safe manner and instructs it to run the death test.  The
+// implementation uses fork(2) + exec.  On systems where clone(2) is
+// available, it is used instead, being slightly more thread-safe.  On QNX,
+// fork supports only single-threaded environments, so this function uses
+// spawn(2) there instead.  The function dies with an error message if
+// anything goes wrong.
+static pid_t ExecDeathTestSpawnChild(char* const* argv, int close_fd) {
+  ExecDeathTestArgs args = { argv, close_fd };
+  pid_t child_pid = -1;
+
+#  if GTEST_OS_QNX
+  // Obtains the current directory and sets it to be closed in the child
+  // process.
+  const int cwd_fd = open(".", O_RDONLY);
+  GTEST_DEATH_TEST_CHECK_(cwd_fd != -1);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(fcntl(cwd_fd, F_SETFD, FD_CLOEXEC));
+  // We need to execute the test program in the same environment where
+  // it was originally invoked.  Therefore we change to the original
+  // working directory first.
+  const char* const original_dir =
+      UnitTest::GetInstance()->original_working_dir();
+  // We can safely call chdir() as it's a direct system call.
+  if (chdir(original_dir) != 0) {
+    DeathTestAbort(std::string("chdir(\"") + original_dir + "\") failed: " +
+                   GetLastErrnoDescription());
+    return EXIT_FAILURE;
+  }
+
+  int fd_flags;
+  // Set close_fd to be closed after spawn.
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(fd_flags = fcntl(close_fd, F_GETFD));
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(fcntl(close_fd, F_SETFD,
+                                        fd_flags | FD_CLOEXEC));
+  struct inheritance inherit = {0};
+  // spawn is a system call.
+  child_pid = spawn(args.argv[0], 0, NULL, &inherit, args.argv, GetEnviron());
+  // Restores the current working directory.
+  GTEST_DEATH_TEST_CHECK_(fchdir(cwd_fd) != -1);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(close(cwd_fd));
+
+#  else   // GTEST_OS_QNX
+#   if GTEST_OS_LINUX
+  // When a SIGPROF signal is received while fork() or clone() are executing,
+  // the process may hang. To avoid this, we ignore SIGPROF here and re-enable
+  // it after the call to fork()/clone() is complete.
+  struct sigaction saved_sigprof_action;
+  struct sigaction ignore_sigprof_action;
+  memset(&ignore_sigprof_action, 0, sizeof(ignore_sigprof_action));
+  sigemptyset(&ignore_sigprof_action.sa_mask);
+  ignore_sigprof_action.sa_handler = SIG_IGN;
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(sigaction(
+      SIGPROF, &ignore_sigprof_action, &saved_sigprof_action));
+#   endif  // GTEST_OS_LINUX
+
+#   if GTEST_HAS_CLONE
+  const bool use_fork = GTEST_FLAG(death_test_use_fork);
+
+  if (!use_fork) {
+    static const bool stack_grows_down = StackGrowsDown();
+    const size_t stack_size = getpagesize();
+    // MMAP_ANONYMOUS is not defined on Mac, so we use MAP_ANON instead.
+    void* const stack = mmap(NULL, stack_size, PROT_READ | PROT_WRITE,
+                             MAP_ANON | MAP_PRIVATE, -1, 0);
+    GTEST_DEATH_TEST_CHECK_(stack != MAP_FAILED);
+
+    // Maximum stack alignment in bytes:  For a downward-growing stack, this
+    // amount is subtracted from size of the stack space to get an address
+    // that is within the stack space and is aligned on all systems we care
+    // about.  As far as I know there is no ABI with stack alignment greater
+    // than 64.  We assume stack and stack_size already have alignment of
+    // kMaxStackAlignment.
+    const size_t kMaxStackAlignment = 64;
+    void* const stack_top =
+        static_cast<char*>(stack) +
+            (stack_grows_down ? stack_size - kMaxStackAlignment : 0);
+    GTEST_DEATH_TEST_CHECK_(stack_size > kMaxStackAlignment &&
+        reinterpret_cast<intptr_t>(stack_top) % kMaxStackAlignment == 0);
+
+    child_pid = clone(&ExecDeathTestChildMain, stack_top, SIGCHLD, &args);
+
+    GTEST_DEATH_TEST_CHECK_(munmap(stack, stack_size) != -1);
+  }
+#   else
+  const bool use_fork = true;
+#   endif  // GTEST_HAS_CLONE
+
+  if (use_fork && (child_pid = fork()) == 0) {
+      ExecDeathTestChildMain(&args);
+      _exit(0);
+  }
+#  endif  // GTEST_OS_QNX
+#  if GTEST_OS_LINUX
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(
+      sigaction(SIGPROF, &saved_sigprof_action, NULL));
+#  endif  // GTEST_OS_LINUX
+
+  GTEST_DEATH_TEST_CHECK_(child_pid != -1);
+  return child_pid;
+}
+
+// The AssumeRole process for a fork-and-exec death test.  It re-executes the
+// main program from the beginning, setting the --gtest_filter
+// and --gtest_internal_run_death_test flags to cause only the current
+// death test to be re-run.
+DeathTest::TestRole ExecDeathTest::AssumeRole() {
+  const UnitTestImpl* const impl = GetUnitTestImpl();
+  const InternalRunDeathTestFlag* const flag =
+      impl->internal_run_death_test_flag();
+  const TestInfo* const info = impl->current_test_info();
+  const int death_test_index = info->result()->death_test_count();
+
+  if (flag != NULL) {
+    set_write_fd(flag->write_fd());
+    return EXECUTE_TEST;
+  }
+
+  int pipe_fd[2];
+  GTEST_DEATH_TEST_CHECK_(pipe(pipe_fd) != -1);
+  // Clear the close-on-exec flag on the write end of the pipe, lest
+  // it be closed when the child process does an exec:
+  GTEST_DEATH_TEST_CHECK_(fcntl(pipe_fd[1], F_SETFD, 0) != -1);
+
+  const std::string filter_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kFilterFlag + "="
+      + info->test_case_name() + "." + info->name();
+  const std::string internal_flag =
+      std::string("--") + GTEST_FLAG_PREFIX_ + kInternalRunDeathTestFlag + "="
+      + file_ + "|" + StreamableToString(line_) + "|"
+      + StreamableToString(death_test_index) + "|"
+      + StreamableToString(pipe_fd[1]);
+  Arguments args;
+  args.AddArguments(GetArgvsForDeathTestChildProcess());
+  args.AddArgument(filter_flag.c_str());
+  args.AddArgument(internal_flag.c_str());
+
+  DeathTest::set_last_death_test_message("");
+
+  CaptureStderr();
+  // See the comment in NoExecDeathTest::AssumeRole for why the next line
+  // is necessary.
+  FlushInfoLog();
+
+  const pid_t child_pid = ExecDeathTestSpawnChild(args.Argv(), pipe_fd[0]);
+  GTEST_DEATH_TEST_CHECK_SYSCALL_(close(pipe_fd[1]));
+  set_child_pid(child_pid);
+  set_read_fd(pipe_fd[0]);
+  set_spawned(true);
+  return OVERSEE_TEST;
+}
+
+# endif  // !GTEST_OS_WINDOWS
+
+// Creates a concrete DeathTest-derived class that depends on the
+// --gtest_death_test_style flag, and sets the pointer pointed to
+// by the "test" argument to its address.  If the test should be
+// skipped, sets that pointer to NULL.  Returns true, unless the
+// flag is set to an invalid value.
+bool DefaultDeathTestFactory::Create(const char* statement, const RE* regex,
+                                     const char* file, int line,
+                                     DeathTest** test) {
+  UnitTestImpl* const impl = GetUnitTestImpl();
+  const InternalRunDeathTestFlag* const flag =
+      impl->internal_run_death_test_flag();
+  const int death_test_index = impl->current_test_info()
+      ->increment_death_test_count();
+
+  if (flag != NULL) {
+    if (death_test_index > flag->index()) {
+      DeathTest::set_last_death_test_message(
+          "Death test count (" + StreamableToString(death_test_index)
+          + ") somehow exceeded expected maximum ("
+          + StreamableToString(flag->index()) + ")");
+      return false;
+    }
+
+    if (!(flag->file() == file && flag->line() == line &&
+          flag->index() == death_test_index)) {
+      *test = NULL;
+      return true;
+    }
+  }
+
+# if GTEST_OS_WINDOWS
+
+  if (GTEST_FLAG(death_test_style) == "threadsafe" ||
+      GTEST_FLAG(death_test_style) == "fast") {
+    *test = new WindowsDeathTest(statement, regex, file, line);
+  }
+
+# else
+
+  if (GTEST_FLAG(death_test_style) == "threadsafe") {
+    *test = new ExecDeathTest(statement, regex, file, line);
+  } else if (GTEST_FLAG(death_test_style) == "fast") {
+    *test = new NoExecDeathTest(statement, regex);
+  }
+
+# endif  // GTEST_OS_WINDOWS
+
+  else {  // NOLINT - this is more readable than unbalanced brackets inside #if.
+    DeathTest::set_last_death_test_message(
+        "Unknown death test style \"" + GTEST_FLAG(death_test_style)
+        + "\" encountered");
+    return false;
+  }
+
+  return true;
+}
+
+# if GTEST_OS_WINDOWS
+// Recreates the pipe and event handles from the provided parameters,
+// signals the event, and returns a file descriptor wrapped around the pipe
+// handle. This function is called in the child process only.
+int GetStatusFileDescriptor(unsigned int parent_process_id,
+                            size_t write_handle_as_size_t,
+                            size_t event_handle_as_size_t) {
+  AutoHandle parent_process_handle(::OpenProcess(PROCESS_DUP_HANDLE,
+                                                   FALSE,  // Non-inheritable.
+                                                   parent_process_id));
+  if (parent_process_handle.Get() == INVALID_HANDLE_VALUE) {
+    DeathTestAbort("Unable to open parent process " +
+                   StreamableToString(parent_process_id));
+  }
+
+  // TODO(vladl@google.com): Replace the following check with a
+  // compile-time assertion when available.
+  GTEST_CHECK_(sizeof(HANDLE) <= sizeof(size_t));
+
+  const HANDLE write_handle =
+      reinterpret_cast<HANDLE>(write_handle_as_size_t);
+  HANDLE dup_write_handle;
+
+  // The newly initialized handle is accessible only in in the parent
+  // process. To obtain one accessible within the child, we need to use
+  // DuplicateHandle.
+  if (!::DuplicateHandle(parent_process_handle.Get(), write_handle,
+                         ::GetCurrentProcess(), &dup_write_handle,
+                         0x0,    // Requested privileges ignored since
+                                 // DUPLICATE_SAME_ACCESS is used.
+                         FALSE,  // Request non-inheritable handler.
+                         DUPLICATE_SAME_ACCESS)) {
+    DeathTestAbort("Unable to duplicate the pipe handle " +
+                   StreamableToString(write_handle_as_size_t) +
+                   " from the parent process " +
+                   StreamableToString(parent_process_id));
+  }
+
+  const HANDLE event_handle = reinterpret_cast<HANDLE>(event_handle_as_size_t);
+  HANDLE dup_event_handle;
+
+  if (!::DuplicateHandle(parent_process_handle.Get(), event_handle,
+                         ::GetCurrentProcess(), &dup_event_handle,
+                         0x0,
+                         FALSE,
+                         DUPLICATE_SAME_ACCESS)) {
+    DeathTestAbort("Unable to duplicate the event handle " +
+                   StreamableToString(event_handle_as_size_t) +
+                   " from the parent process " +
+                   StreamableToString(parent_process_id));
+  }
+
+  const int write_fd =
+      ::_open_osfhandle(reinterpret_cast<intptr_t>(dup_write_handle), O_APPEND);
+  if (write_fd == -1) {
+    DeathTestAbort("Unable to convert pipe handle " +
+                   StreamableToString(write_handle_as_size_t) +
+                   " to a file descriptor");
+  }
+
+  // Signals the parent that the write end of the pipe has been acquired
+  // so the parent can release its own write end.
+  ::SetEvent(dup_event_handle);
+
+  return write_fd;
+}
+# endif  // GTEST_OS_WINDOWS
+
+// Returns a newly created InternalRunDeathTestFlag object with fields
+// initialized from the GTEST_FLAG(internal_run_death_test) flag if
+// the flag is specified; otherwise returns NULL.
+InternalRunDeathTestFlag* ParseInternalRunDeathTestFlag() {
+  if (GTEST_FLAG(internal_run_death_test) == "") return NULL;
+
+  // GTEST_HAS_DEATH_TEST implies that we have ::std::string, so we
+  // can use it here.
+  int line = -1;
+  int index = -1;
+  ::std::vector< ::std::string> fields;
+  SplitString(GTEST_FLAG(internal_run_death_test).c_str(), '|', &fields);
+  int write_fd = -1;
+
+# if GTEST_OS_WINDOWS
+
+  unsigned int parent_process_id = 0;
+  size_t write_handle_as_size_t = 0;
+  size_t event_handle_as_size_t = 0;
+
+  if (fields.size() != 6
+      || !ParseNaturalNumber(fields[1], &line)
+      || !ParseNaturalNumber(fields[2], &index)
+      || !ParseNaturalNumber(fields[3], &parent_process_id)
+      || !ParseNaturalNumber(fields[4], &write_handle_as_size_t)
+      || !ParseNaturalNumber(fields[5], &event_handle_as_size_t)) {
+    DeathTestAbort("Bad --gtest_internal_run_death_test flag: " +
+                   GTEST_FLAG(internal_run_death_test));
+  }
+  write_fd = GetStatusFileDescriptor(parent_process_id,
+                                     write_handle_as_size_t,
+                                     event_handle_as_size_t);
+# else
+
+  if (fields.size() != 4
+      || !ParseNaturalNumber(fields[1], &line)
+      || !ParseNaturalNumber(fields[2], &index)
+      || !ParseNaturalNumber(fields[3], &write_fd)) {
+    DeathTestAbort("Bad --gtest_internal_run_death_test flag: "
+        + GTEST_FLAG(internal_run_death_test));
+  }
+
+# endif  // GTEST_OS_WINDOWS
+
+  return new InternalRunDeathTestFlag(fields[0], line, index, write_fd);
+}
+
+}  // namespace internal
+
+#endif  // GTEST_HAS_DEATH_TEST
+
+}  // namespace testing
diff --git a/third_party/aom/third_party/googletest/src/googletest/src/gtest-filepath.cc b/third_party/aom/third_party/googletest/src/googletest/src/gtest-filepath.cc
new file mode 100644
index 0000000000..0292dc1195
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/src/gtest-filepath.cc
@@ -0,0 +1,387 @@
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Authors: keith.ray@gmail.com (Keith Ray)
+
+#include "gtest/gtest-message.h"
+#include "gtest/internal/gtest-filepath.h"
+#include "gtest/internal/gtest-port.h"
+
+#include <stdlib.h>
+
+#if GTEST_OS_WINDOWS_MOBILE
+# include <windows.h>
+#elif GTEST_OS_WINDOWS
+# include <direct.h>
+# include <io.h>
+#elif GTEST_OS_SYMBIAN
+// Symbian OpenC has PATH_MAX in sys/syslimits.h
+# include <sys/syslimits.h>
+#else
+# include <limits.h>
+# include <climits>  // Some Linux distributions define PATH_MAX here.
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+#if GTEST_OS_WINDOWS
+# define GTEST_PATH_MAX_ _MAX_PATH
+#elif defined(PATH_MAX)
+# define GTEST_PATH_MAX_ PATH_MAX
+#elif defined(_XOPEN_PATH_MAX)
+# define GTEST_PATH_MAX_ _XOPEN_PATH_MAX
+#else
+# define GTEST_PATH_MAX_ _POSIX_PATH_MAX
+#endif  // GTEST_OS_WINDOWS
+
+#include "gtest/internal/gtest-string.h"
+
+namespace testing {
+namespace internal {
+
+#if GTEST_OS_WINDOWS
+// On Windows, '\\' is the standard path separator, but many tools and the
+// Windows API also accept '/' as an alternate path separator. Unless otherwise
+// noted, a file path can contain either kind of path separators, or a mixture
+// of them.
+const char kPathSeparator = '\\';
+const char kAlternatePathSeparator = '/';
+const char kAlternatePathSeparatorString[] = "/";
+# if GTEST_OS_WINDOWS_MOBILE
+// Windows CE doesn't have a current directory. You should not use
+// the current directory in tests on Windows CE, but this at least
+// provides a reasonable fallback.
+const char kCurrentDirectoryString[] = "\\";
+// Windows CE doesn't define INVALID_FILE_ATTRIBUTES
+const DWORD kInvalidFileAttributes = 0xffffffff;
+# else
+const char kCurrentDirectoryString[] = ".\\";
+# endif  // GTEST_OS_WINDOWS_MOBILE
+#else
+const char kPathSeparator = '/';
+const char kCurrentDirectoryString[] = "./";
+#endif  // GTEST_OS_WINDOWS
+
+// Returns whether the given character is a valid path separator.
+static bool IsPathSeparator(char c) {
+#if GTEST_HAS_ALT_PATH_SEP_
+  return (c == kPathSeparator) || (c == kAlternatePathSeparator);
+#else
+  return c == kPathSeparator;
+#endif
+}
+
+// Returns the current working directory, or "" if unsuccessful.
+FilePath FilePath::GetCurrentDir() {
+#if GTEST_OS_WINDOWS_MOBILE || GTEST_OS_WINDOWS_PHONE || GTEST_OS_WINDOWS_RT
+  // Windows CE doesn't have a current directory, so we just return
+  // something reasonable.
+  return FilePath(kCurrentDirectoryString);
+#elif GTEST_OS_WINDOWS
+  char cwd[GTEST_PATH_MAX_ + 1] = { '\0' };
+  return FilePath(_getcwd(cwd, sizeof(cwd)) == NULL ? "" : cwd);
+#else
+  char cwd[GTEST_PATH_MAX_ + 1] = { '\0' };
+  char* result = getcwd(cwd, sizeof(cwd));
+# if GTEST_OS_NACL
+  // getcwd will likely fail in NaCl due to the sandbox, so return something
+  // reasonable. The user may have provided a shim implementation for getcwd,
+  // however, so fallback only when failure is detected.
+  return FilePath(result == NULL ? kCurrentDirectoryString : cwd);
+# endif  // GTEST_OS_NACL
+  return FilePath(result == NULL ? "" : cwd);
+#endif  // GTEST_OS_WINDOWS_MOBILE
+}
+
+// Returns a copy of the FilePath with the case-insensitive extension removed.
+// Example: FilePath("dir/file.exe").RemoveExtension("EXE") returns
+// FilePath("dir/file"). If a case-insensitive extension is not
+// found, returns a copy of the original FilePath.
+FilePath FilePath::RemoveExtension(const char* extension) const {
+  const std::string dot_extension = std::string(".") + extension;
+  if (String::EndsWithCaseInsensitive(pathname_, dot_extension)) {
+    return FilePath(pathname_.substr(
+        0, pathname_.length() - dot_extension.length()));
+  }
+  return *this;
+}
+
+// Returns a pointer to the last occurence of a valid path separator in
+// the FilePath. On Windows, for example, both '/' and '\' are valid path
+// separators. Returns NULL if no path separator was found.
+const char* FilePath::FindLastPathSeparator() const {
+  const char* const last_sep = strrchr(c_str(), kPathSeparator);
+#if GTEST_HAS_ALT_PATH_SEP_
+  const char* const last_alt_sep = strrchr(c_str(), kAlternatePathSeparator);
+  // Comparing two pointers of which only one is NULL is undefined.
+  if (last_alt_sep != NULL &&
+      (last_sep == NULL || last_alt_sep > last_sep)) {
+    return last_alt_sep;
+  }
+#endif
+  return last_sep;
+}
+
+// Returns a copy of the FilePath with the directory part removed.
+// Example: FilePath("path/to/file").RemoveDirectoryName() returns
+// FilePath("file"). If there is no directory part ("just_a_file"), it returns
+// the FilePath unmodified. If there is no file part ("just_a_dir/") it
+// returns an empty FilePath ("").
+// On Windows platform, '\' is the path separator, otherwise it is '/'.
+FilePath FilePath::RemoveDirectoryName() const {
+  const char* const last_sep = FindLastPathSeparator();
+  return last_sep ? FilePath(last_sep + 1) : *this;
+}
+
+// RemoveFileName returns the directory path with the filename removed.
+// Example: FilePath("path/to/file").RemoveFileName() returns "path/to/".
+// If the FilePath is "a_file" or "/a_file", RemoveFileName returns
+// FilePath("./") or, on Windows, FilePath(".\\"). If the filepath does
+// not have a file, like "just/a/dir/", it returns the FilePath unmodified.
+// On Windows platform, '\' is the path separator, otherwise it is '/'.
+FilePath FilePath::RemoveFileName() const {
+  const char* const last_sep = FindLastPathSeparator();
+  std::string dir;
+  if (last_sep) {
+    dir = std::string(c_str(), last_sep + 1 - c_str());
+  } else {
+    dir = kCurrentDirectoryString;
+  }
+  return FilePath(dir);
+}
+
+// Helper functions for naming files in a directory for xml output.
+
+// Given directory = "dir", base_name = "test", number = 0,
+// extension = "xml", returns "dir/test.xml". If number is greater
+// than zero (e.g., 12), returns "dir/test_12.xml".
+// On Windows platform, uses \ as the separator rather than /.
+FilePath FilePath::MakeFileName(const FilePath& directory,
+                                const FilePath& base_name,
+                                int number,
+                                const char* extension) {
+  std::string file;
+  if (number == 0) {
+    file = base_name.string() + "." + extension;
+  } else {
+    file = base_name.string() + "_" + StreamableToString(number)
+        + "." + extension;
+  }
+  return ConcatPaths(directory, FilePath(file));
+}
+
+// Given directory = "dir", relative_path = "test.xml", returns "dir/test.xml".
+// On Windows, uses \ as the separator rather than /.
+FilePath FilePath::ConcatPaths(const FilePath& directory,
+                               const FilePath& relative_path) {
+  if (directory.IsEmpty())
+    return relative_path;
+  const FilePath dir(directory.RemoveTrailingPathSeparator());
+  return FilePath(dir.string() + kPathSeparator + relative_path.string());
+}
+
+// Returns true if pathname describes something findable in the file-system,
+// either a file, directory, or whatever.
+bool FilePath::FileOrDirectoryExists() const {
+#if GTEST_OS_WINDOWS_MOBILE
+  LPCWSTR unicode = String::AnsiToUtf16(pathname_.c_str());
+  const DWORD attributes = GetFileAttributes(unicode);
+  delete [] unicode;
+  return attributes != kInvalidFileAttributes;
+#else
+  posix::StatStruct file_stat;
+  return posix::Stat(pathname_.c_str(), &file_stat) == 0;
+#endif  // GTEST_OS_WINDOWS_MOBILE
+}
+
+// Returns true if pathname describes a directory in the file-system
+// that exists.
+bool FilePath::DirectoryExists() const {
+  bool result = false;
+#if GTEST_OS_WINDOWS
+  // Don't strip off trailing separator if path is a root directory on
+  // Windows (like "C:\\").
+  const FilePath& path(IsRootDirectory() ? *this :
+                                           RemoveTrailingPathSeparator());
+#else
+  const FilePath& path(*this);
+#endif
+
+#if GTEST_OS_WINDOWS_MOBILE
+  LPCWSTR unicode = String::AnsiToUtf16(path.c_str());
+  const DWORD attributes = GetFileAttributes(unicode);
+  delete [] unicode;
+  if ((attributes != kInvalidFileAttributes) &&
+      (attributes & FILE_ATTRIBUTE_DIRECTORY)) {
+    result = true;
+  }
+#else
+  posix::StatStruct file_stat;
+  result = posix::Stat(path.c_str(), &file_stat) == 0 &&
+      posix::IsDir(file_stat);
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+  return result;
+}
+
+// Returns true if pathname describes a root directory. (Windows has one
+// root directory per disk drive.)
+bool FilePath::IsRootDirectory() const {
+#if GTEST_OS_WINDOWS
+  // TODO(wan@google.com): on Windows a network share like
+  // \\server\share can be a root directory, although it cannot be the
+  // current directory.  Handle this properly.
+  return pathname_.length() == 3 && IsAbsolutePath();
+#else
+  return pathname_.length() == 1 && IsPathSeparator(pathname_.c_str()[0]);
+#endif
+}
+
+// Returns true if pathname describes an absolute path.
+bool FilePath::IsAbsolutePath() const {
+  const char* const name = pathname_.c_str();
+#if GTEST_OS_WINDOWS
+  return pathname_.length() >= 3 &&
+     ((name[0] >= 'a' && name[0] <= 'z') ||
+      (name[0] >= 'A' && name[0] <= 'Z')) &&
+     name[1] == ':' &&
+     IsPathSeparator(name[2]);
+#else
+  return IsPathSeparator(name[0]);
+#endif
+}
+
+// Returns a pathname for a file that does not currently exist. The pathname
+// will be directory/base_name.extension or
+// directory/base_name_<number>.extension if directory/base_name.extension
+// already exists. The number will be incremented until a pathname is found
+// that does not already exist.
+// Examples: 'dir/foo_test.xml' or 'dir/foo_test_1.xml'.
+// There could be a race condition if two or more processes are calling this
+// function at the same time -- they could both pick the same filename.
+FilePath FilePath::GenerateUniqueFileName(const FilePath& directory,
+                                          const FilePath& base_name,
+                                          const char* extension) {
+  FilePath full_pathname;
+  int number = 0;
+  do {
+    full_pathname.Set(MakeFileName(directory, base_name, number++, extension));
+  } while (full_pathname.FileOrDirectoryExists());
+  return full_pathname;
+}
+
+// Returns true if FilePath ends with a path separator, which indicates that
+// it is intended to represent a directory. Returns false otherwise.
+// This does NOT check that a directory (or file) actually exists.
+bool FilePath::IsDirectory() const {
+  return !pathname_.empty() &&
+         IsPathSeparator(pathname_.c_str()[pathname_.length() - 1]);
+}
+
+// Create directories so that path exists. Returns true if successful or if
+// the directories already exist; returns false if unable to create directories
+// for any reason.
+bool FilePath::CreateDirectoriesRecursively() const {
+  if (!this->IsDirectory()) {
+    return false;
+  }
+
+  if (pathname_.length() == 0 || this->DirectoryExists()) {
+    return true;
+  }
+
+  const FilePath parent(this->RemoveTrailingPathSeparator().RemoveFileName());
+  return parent.CreateDirectoriesRecursively() && this->CreateFolder();
+}
+
+// Create the directory so that path exists. Returns true if successful or
+// if the directory already exists; returns false if unable to create the
+// directory for any reason, including if the parent directory does not
+// exist. Not named "CreateDirectory" because that's a macro on Windows.
+bool FilePath::CreateFolder() const {
+#if GTEST_OS_WINDOWS_MOBILE
+  FilePath removed_sep(this->RemoveTrailingPathSeparator());
+  LPCWSTR unicode = String::AnsiToUtf16(removed_sep.c_str());
+  int result = CreateDirectory(unicode, NULL) ? 0 : -1;
+  delete [] unicode;
+#elif GTEST_OS_WINDOWS
+  int result = _mkdir(pathname_.c_str());
+#else
+  int result = mkdir(pathname_.c_str(), 0777);
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+  if (result == -1) {
+    return this->DirectoryExists();  // An error is OK if the directory exists.
+  }
+  return true;  // No error.
+}
+
+// If input name has a trailing separator character, remove it and return the
+// name, otherwise return the name string unmodified.
+// On Windows platform, uses \ as the separator, other platforms use /.
+FilePath FilePath::RemoveTrailingPathSeparator() const {
+  return IsDirectory()
+      ? FilePath(pathname_.substr(0, pathname_.length() - 1))
+      : *this;
+}
+
+// Removes any redundant separators that might be in the pathname.
+// For example, "bar///foo" becomes "bar/foo". Does not eliminate other
+// redundancies that might be in a pathname involving "." or "..".
+// TODO(wan@google.com): handle Windows network shares (e.g. \\server\share).
+void FilePath::Normalize() {
+  if (pathname_.c_str() == NULL) {
+    pathname_ = "";
+    return;
+  }
+  const char* src = pathname_.c_str();
+  char* const dest = new char[pathname_.length() + 1];
+  char* dest_ptr = dest;
+  memset(dest_ptr, 0, pathname_.length() + 1);
+
+  while (*src != '\0') {
+    *dest_ptr = *src;
+    if (!IsPathSeparator(*src)) {
+      src++;
+    } else {
+#if GTEST_HAS_ALT_PATH_SEP_
+      if (*dest_ptr == kAlternatePathSeparator) {
+        *dest_ptr = kPathSeparator;
+      }
+#endif
+      while (IsPathSeparator(*src))
+        src++;
+    }
+    dest_ptr++;
+  }
+  *dest_ptr = '\0';
+  pathname_ = dest;
+  delete[] dest;
+}
+
+}  // namespace internal
+}  // namespace testing
diff --git a/third_party/aom/third_party/googletest/src/googletest/src/gtest-internal-inl.h b/third_party/aom/third_party/googletest/src/googletest/src/gtest-internal-inl.h
new file mode 100644
index 0000000000..ed8a682a96
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/src/gtest-internal-inl.h
@@ -0,0 +1,1183 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Utility functions and classes used by the Google C++ testing framework.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// This file contains purely Google Test's internal implementation.  Please
+// DO NOT #INCLUDE IT IN A USER PROGRAM.
+
+#ifndef GTEST_SRC_GTEST_INTERNAL_INL_H_
+#define GTEST_SRC_GTEST_INTERNAL_INL_H_
+
+// GTEST_IMPLEMENTATION_ is defined to 1 iff the current translation unit is
+// part of Google Test's implementation; otherwise it's undefined.
+#if !GTEST_IMPLEMENTATION_
+// If this file is included from the user's code, just say no.
+# error "gtest-internal-inl.h is part of Google Test's internal implementation."
+# error "It must not be included except by Google Test itself."
+#endif  // GTEST_IMPLEMENTATION_
+
+#ifndef _WIN32_WCE
+# include <errno.h>
+#endif  // !_WIN32_WCE
+#include <stddef.h>
+#include <stdlib.h>  // For strtoll/_strtoul64/malloc/free.
+#include <string.h>  // For memmove.
+
+#include <algorithm>
+#include <string>
+#include <vector>
+
+#include "gtest/internal/gtest-port.h"
+
+#if GTEST_CAN_STREAM_RESULTS_
+# include <arpa/inet.h>  // NOLINT
+# include <netdb.h>  // NOLINT
+#endif
+
+#if GTEST_OS_WINDOWS
+# include <windows.h>  // NOLINT
+#endif  // GTEST_OS_WINDOWS
+
+#include "gtest/gtest.h"  // NOLINT
+#include "gtest/gtest-spi.h"
+
+namespace testing {
+
+// Declares the flags.
+//
+// We don't want the users to modify this flag in the code, but want
+// Google Test's own unit tests to be able to access it. Therefore we
+// declare it here as opposed to in gtest.h.
+GTEST_DECLARE_bool_(death_test_use_fork);
+
+namespace internal {
+
+// The value of GetTestTypeId() as seen from within the Google Test
+// library.  This is solely for testing GetTestTypeId().
+GTEST_API_ extern const TypeId kTestTypeIdInGoogleTest;
+
+// Names of the flags (needed for parsing Google Test flags).
+const char kAlsoRunDisabledTestsFlag[] = "also_run_disabled_tests";
+const char kBreakOnFailureFlag[] = "break_on_failure";
+const char kCatchExceptionsFlag[] = "catch_exceptions";
+const char kColorFlag[] = "color";
+const char kFilterFlag[] = "filter";
+const char kListTestsFlag[] = "list_tests";
+const char kOutputFlag[] = "output";
+const char kPrintTimeFlag[] = "print_time";
+const char kRandomSeedFlag[] = "random_seed";
+const char kRepeatFlag[] = "repeat";
+const char kShuffleFlag[] = "shuffle";
+const char kStackTraceDepthFlag[] = "stack_trace_depth";
+const char kStreamResultToFlag[] = "stream_result_to";
+const char kThrowOnFailureFlag[] = "throw_on_failure";
+const char kFlagfileFlag[] = "flagfile";
+
+// A valid random seed must be in [1, kMaxRandomSeed].
+const int kMaxRandomSeed = 99999;
+
+// g_help_flag is true iff the --help flag or an equivalent form is
+// specified on the command line.
+GTEST_API_ extern bool g_help_flag;
+
+// Returns the current time in milliseconds.
+GTEST_API_ TimeInMillis GetTimeInMillis();
+
+// Returns true iff Google Test should use colors in the output.
+GTEST_API_ bool ShouldUseColor(bool stdout_is_tty);
+
+// Formats the given time in milliseconds as seconds.
+GTEST_API_ std::string FormatTimeInMillisAsSeconds(TimeInMillis ms);
+
+// Converts the given time in milliseconds to a date string in the ISO 8601
+// format, without the timezone information.  N.B.: due to the use the
+// non-reentrant localtime() function, this function is not thread safe.  Do
+// not use it in any code that can be called from multiple threads.
+GTEST_API_ std::string FormatEpochTimeInMillisAsIso8601(TimeInMillis ms);
+
+// Parses a string for an Int32 flag, in the form of "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+GTEST_API_ bool ParseInt32Flag(
+    const char* str, const char* flag, Int32* value);
+
+// Returns a random seed in range [1, kMaxRandomSeed] based on the
+// given --gtest_random_seed flag value.
+inline int GetRandomSeedFromFlag(Int32 random_seed_flag) {
+  const unsigned int raw_seed = (random_seed_flag == 0) ?
+      static_cast<unsigned int>(GetTimeInMillis()) :
+      static_cast<unsigned int>(random_seed_flag);
+
+  // Normalizes the actual seed to range [1, kMaxRandomSeed] such that
+  // it's easy to type.
+  const int normalized_seed =
+      static_cast<int>((raw_seed - 1U) %
+                       static_cast<unsigned int>(kMaxRandomSeed)) + 1;
+  return normalized_seed;
+}
+
+// Returns the first valid random seed after 'seed'.  The behavior is
+// undefined if 'seed' is invalid.  The seed after kMaxRandomSeed is
+// considered to be 1.
+inline int GetNextRandomSeed(int seed) {
+  GTEST_CHECK_(1 <= seed && seed <= kMaxRandomSeed)
+      << "Invalid random seed " << seed << " - must be in [1, "
+      << kMaxRandomSeed << "].";
+  const int next_seed = seed + 1;
+  return (next_seed > kMaxRandomSeed) ? 1 : next_seed;
+}
+
+// This class saves the values of all Google Test flags in its c'tor, and
+// restores them in its d'tor.
+class GTestFlagSaver {
+ public:
+  // The c'tor.
+  GTestFlagSaver() {
+    also_run_disabled_tests_ = GTEST_FLAG(also_run_disabled_tests);
+    break_on_failure_ = GTEST_FLAG(break_on_failure);
+    catch_exceptions_ = GTEST_FLAG(catch_exceptions);
+    color_ = GTEST_FLAG(color);
+    death_test_style_ = GTEST_FLAG(death_test_style);
+    death_test_use_fork_ = GTEST_FLAG(death_test_use_fork);
+    filter_ = GTEST_FLAG(filter);
+    internal_run_death_test_ = GTEST_FLAG(internal_run_death_test);
+    list_tests_ = GTEST_FLAG(list_tests);
+    output_ = GTEST_FLAG(output);
+    print_time_ = GTEST_FLAG(print_time);
+    random_seed_ = GTEST_FLAG(random_seed);
+    repeat_ = GTEST_FLAG(repeat);
+    shuffle_ = GTEST_FLAG(shuffle);
+    stack_trace_depth_ = GTEST_FLAG(stack_trace_depth);
+    stream_result_to_ = GTEST_FLAG(stream_result_to);
+    throw_on_failure_ = GTEST_FLAG(throw_on_failure);
+  }
+
+  // The d'tor is not virtual.  DO NOT INHERIT FROM THIS CLASS.
+  ~GTestFlagSaver() {
+    GTEST_FLAG(also_run_disabled_tests) = also_run_disabled_tests_;
+    GTEST_FLAG(break_on_failure) = break_on_failure_;
+    GTEST_FLAG(catch_exceptions) = catch_exceptions_;
+    GTEST_FLAG(color) = color_;
+    GTEST_FLAG(death_test_style) = death_test_style_;
+    GTEST_FLAG(death_test_use_fork) = death_test_use_fork_;
+    GTEST_FLAG(filter) = filter_;
+    GTEST_FLAG(internal_run_death_test) = internal_run_death_test_;
+    GTEST_FLAG(list_tests) = list_tests_;
+    GTEST_FLAG(output) = output_;
+    GTEST_FLAG(print_time) = print_time_;
+    GTEST_FLAG(random_seed) = random_seed_;
+    GTEST_FLAG(repeat) = repeat_;
+    GTEST_FLAG(shuffle) = shuffle_;
+    GTEST_FLAG(stack_trace_depth) = stack_trace_depth_;
+    GTEST_FLAG(stream_result_to) = stream_result_to_;
+    GTEST_FLAG(throw_on_failure) = throw_on_failure_;
+  }
+
+ private:
+  // Fields for saving the original values of flags.
+  bool also_run_disabled_tests_;
+  bool break_on_failure_;
+  bool catch_exceptions_;
+  std::string color_;
+  std::string death_test_style_;
+  bool death_test_use_fork_;
+  std::string filter_;
+  std::string internal_run_death_test_;
+  bool list_tests_;
+  std::string output_;
+  bool print_time_;
+  internal::Int32 random_seed_;
+  internal::Int32 repeat_;
+  bool shuffle_;
+  internal::Int32 stack_trace_depth_;
+  std::string stream_result_to_;
+  bool throw_on_failure_;
+} GTEST_ATTRIBUTE_UNUSED_;
+
+// Converts a Unicode code point to a narrow string in UTF-8 encoding.
+// code_point parameter is of type UInt32 because wchar_t may not be
+// wide enough to contain a code point.
+// If the code_point is not a valid Unicode code point
+// (i.e. outside of Unicode range U+0 to U+10FFFF) it will be converted
+// to "(Invalid Unicode 0xXXXXXXXX)".
+GTEST_API_ std::string CodePointToUtf8(UInt32 code_point);
+
+// Converts a wide string to a narrow string in UTF-8 encoding.
+// The wide string is assumed to have the following encoding:
+//   UTF-16 if sizeof(wchar_t) == 2 (on Windows, Cygwin, Symbian OS)
+//   UTF-32 if sizeof(wchar_t) == 4 (on Linux)
+// Parameter str points to a null-terminated wide string.
+// Parameter num_chars may additionally limit the number
+// of wchar_t characters processed. -1 is used when the entire string
+// should be processed.
+// If the string contains code points that are not valid Unicode code points
+// (i.e. outside of Unicode range U+0 to U+10FFFF) they will be output
+// as '(Invalid Unicode 0xXXXXXXXX)'. If the string is in UTF16 encoding
+// and contains invalid UTF-16 surrogate pairs, values in those pairs
+// will be encoded as individual Unicode characters from Basic Normal Plane.
+GTEST_API_ std::string WideStringToUtf8(const wchar_t* str, int num_chars);
+
+// Reads the GTEST_SHARD_STATUS_FILE environment variable, and creates the file
+// if the variable is present. If a file already exists at this location, this
+// function will write over it. If the variable is present, but the file cannot
+// be created, prints an error and exits.
+void WriteToShardStatusFileIfNeeded();
+
+// Checks whether sharding is enabled by examining the relevant
+// environment variable values. If the variables are present,
+// but inconsistent (e.g., shard_index >= total_shards), prints
+// an error and exits. If in_subprocess_for_death_test, sharding is
+// disabled because it must only be applied to the original test
+// process. Otherwise, we could filter out death tests we intended to execute.
+GTEST_API_ bool ShouldShard(const char* total_shards_str,
+                            const char* shard_index_str,
+                            bool in_subprocess_for_death_test);
+
+// Parses the environment variable var as an Int32. If it is unset,
+// returns default_val. If it is not an Int32, prints an error and
+// and aborts.
+GTEST_API_ Int32 Int32FromEnvOrDie(const char* env_var, Int32 default_val);
+
+// Given the total number of shards, the shard index, and the test id,
+// returns true iff the test should be run on this shard. The test id is
+// some arbitrary but unique non-negative integer assigned to each test
+// method. Assumes that 0 <= shard_index < total_shards.
+GTEST_API_ bool ShouldRunTestOnShard(
+    int total_shards, int shard_index, int test_id);
+
+// STL container utilities.
+
+// Returns the number of elements in the given container that satisfy
+// the given predicate.
+template <class Container, typename Predicate>
+inline int CountIf(const Container& c, Predicate predicate) {
+  // Implemented as an explicit loop since std::count_if() in libCstd on
+  // Solaris has a non-standard signature.
+  int count = 0;
+  for (typename Container::const_iterator it = c.begin(); it != c.end(); ++it) {
+    if (predicate(*it))
+      ++count;
+  }
+  return count;
+}
+
+// Applies a function/functor to each element in the container.
+template <class Container, typename Functor>
+void ForEach(const Container& c, Functor functor) {
+  std::for_each(c.begin(), c.end(), functor);
+}
+
+// Returns the i-th element of the vector, or default_value if i is not
+// in range [0, v.size()).
+template <typename E>
+inline E GetElementOr(const std::vector<E>& v, int i, E default_value) {
+  return (i < 0 || i >= static_cast<int>(v.size())) ? default_value : v[i];
+}
+
+// Performs an in-place shuffle of a range of the vector's elements.
+// 'begin' and 'end' are element indices as an STL-style range;
+// i.e. [begin, end) are shuffled, where 'end' == size() means to
+// shuffle to the end of the vector.
+template <typename E>
+void ShuffleRange(internal::Random* random, int begin, int end,
+                  std::vector<E>* v) {
+  const int size = static_cast<int>(v->size());
+  GTEST_CHECK_(0 <= begin && begin <= size)
+      << "Invalid shuffle range start " << begin << ": must be in range [0, "
+      << size << "].";
+  GTEST_CHECK_(begin <= end && end <= size)
+      << "Invalid shuffle range finish " << end << ": must be in range ["
+      << begin << ", " << size << "].";
+
+  // Fisher-Yates shuffle, from
+  // http://en.wikipedia.org/wiki/Fisher-Yates_shuffle
+  for (int range_width = end - begin; range_width >= 2; range_width--) {
+    const int last_in_range = begin + range_width - 1;
+    const int selected = begin + random->Generate(range_width);
+    std::swap((*v)[selected], (*v)[last_in_range]);
+  }
+}
+
+// Performs an in-place shuffle of the vector's elements.
+template <typename E>
+inline void Shuffle(internal::Random* random, std::vector<E>* v) {
+  ShuffleRange(random, 0, static_cast<int>(v->size()), v);
+}
+
+// A function for deleting an object.  Handy for being used as a
+// functor.
+template <typename T>
+static void Delete(T* x) {
+  delete x;
+}
+
+// A predicate that checks the key of a TestProperty against a known key.
+//
+// TestPropertyKeyIs is copyable.
+class TestPropertyKeyIs {
+ public:
+  // Constructor.
+  //
+  // TestPropertyKeyIs has NO default constructor.
+  explicit TestPropertyKeyIs(const std::string& key) : key_(key) {}
+
+  // Returns true iff the test name of test property matches on key_.
+  bool operator()(const TestProperty& test_property) const {
+    return test_property.key() == key_;
+  }
+
+ private:
+  std::string key_;
+};
+
+// Class UnitTestOptions.
+//
+// This class contains functions for processing options the user
+// specifies when running the tests.  It has only static members.
+//
+// In most cases, the user can specify an option using either an
+// environment variable or a command line flag.  E.g. you can set the
+// test filter using either GTEST_FILTER or --gtest_filter.  If both
+// the variable and the flag are present, the latter overrides the
+// former.
+class GTEST_API_ UnitTestOptions {
+ public:
+  // Functions for processing the gtest_output flag.
+
+  // Returns the output format, or "" for normal printed output.
+  static std::string GetOutputFormat();
+
+  // Returns the absolute path of the requested output file, or the
+  // default (test_detail.xml in the original working directory) if
+  // none was explicitly specified.
+  static std::string GetAbsolutePathToOutputFile();
+
+  // Functions for processing the gtest_filter flag.
+
+  // Returns true iff the wildcard pattern matches the string.  The
+  // first ':' or '\0' character in pattern marks the end of it.
+  //
+  // This recursive algorithm isn't very efficient, but is clear and
+  // works well enough for matching test names, which are short.
+  static bool PatternMatchesString(const char *pattern, const char *str);
+
+  // Returns true iff the user-specified filter matches the test case
+  // name and the test name.
+  static bool FilterMatchesTest(const std::string &test_case_name,
+                                const std::string &test_name);
+
+#if GTEST_OS_WINDOWS
+  // Function for supporting the gtest_catch_exception flag.
+
+  // Returns EXCEPTION_EXECUTE_HANDLER if Google Test should handle the
+  // given SEH exception, or EXCEPTION_CONTINUE_SEARCH otherwise.
+  // This function is useful as an __except condition.
+  static int GTestShouldProcessSEH(DWORD exception_code);
+#endif  // GTEST_OS_WINDOWS
+
+  // Returns true if "name" matches the ':' separated list of glob-style
+  // filters in "filter".
+  static bool MatchesFilter(const std::string& name, const char* filter);
+};
+
+// Returns the current application's name, removing directory path if that
+// is present.  Used by UnitTestOptions::GetOutputFile.
+GTEST_API_ FilePath GetCurrentExecutableName();
+
+// The role interface for getting the OS stack trace as a string.
+class OsStackTraceGetterInterface {
+ public:
+  OsStackTraceGetterInterface() {}
+  virtual ~OsStackTraceGetterInterface() {}
+
+  // Returns the current OS stack trace as an std::string.  Parameters:
+  //
+  //   max_depth  - the maximum number of stack frames to be included
+  //                in the trace.
+  //   skip_count - the number of top frames to be skipped; doesn't count
+  //                against max_depth.
+  virtual string CurrentStackTrace(int max_depth, int skip_count) = 0;
+
+  // UponLeavingGTest() should be called immediately before Google Test calls
+  // user code. It saves some information about the current stack that
+  // CurrentStackTrace() will use to find and hide Google Test stack frames.
+  virtual void UponLeavingGTest() = 0;
+
+  // This string is inserted in place of stack frames that are part of
+  // Google Test's implementation.
+  static const char* const kElidedFramesMarker;
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(OsStackTraceGetterInterface);
+};
+
+// A working implementation of the OsStackTraceGetterInterface interface.
+class OsStackTraceGetter : public OsStackTraceGetterInterface {
+ public:
+  OsStackTraceGetter() {}
+
+  virtual string CurrentStackTrace(int max_depth, int skip_count);
+  virtual void UponLeavingGTest();
+
+ private:
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(OsStackTraceGetter);
+};
+
+// Information about a Google Test trace point.
+struct TraceInfo {
+  const char* file;
+  int line;
+  std::string message;
+};
+
+// This is the default global test part result reporter used in UnitTestImpl.
+// This class should only be used by UnitTestImpl.
+class DefaultGlobalTestPartResultReporter
+  : public TestPartResultReporterInterface {
+ public:
+  explicit DefaultGlobalTestPartResultReporter(UnitTestImpl* unit_test);
+  // Implements the TestPartResultReporterInterface. Reports the test part
+  // result in the current test.
+  virtual void ReportTestPartResult(const TestPartResult& result);
+
+ private:
+  UnitTestImpl* const unit_test_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(DefaultGlobalTestPartResultReporter);
+};
+
+// This is the default per thread test part result reporter used in
+// UnitTestImpl. This class should only be used by UnitTestImpl.
+class DefaultPerThreadTestPartResultReporter
+    : public TestPartResultReporterInterface {
+ public:
+  explicit DefaultPerThreadTestPartResultReporter(UnitTestImpl* unit_test);
+  // Implements the TestPartResultReporterInterface. The implementation just
+  // delegates to the current global test part result reporter of *unit_test_.
+  virtual void ReportTestPartResult(const TestPartResult& result);
+
+ private:
+  UnitTestImpl* const unit_test_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(DefaultPerThreadTestPartResultReporter);
+};
+
+// The private implementation of the UnitTest class.  We don't protect
+// the methods under a mutex, as this class is not accessible by a
+// user and the UnitTest class that delegates work to this class does
+// proper locking.
+class GTEST_API_ UnitTestImpl {
+ public:
+  explicit UnitTestImpl(UnitTest* parent);
+  virtual ~UnitTestImpl();
+
+  // There are two different ways to register your own TestPartResultReporter.
+  // You can register your own repoter to listen either only for test results
+  // from the current thread or for results from all threads.
+  // By default, each per-thread test result repoter just passes a new
+  // TestPartResult to the global test result reporter, which registers the
+  // test part result for the currently running test.
+
+  // Returns the global test part result reporter.
+  TestPartResultReporterInterface* GetGlobalTestPartResultReporter();
+
+  // Sets the global test part result reporter.
+  void SetGlobalTestPartResultReporter(
+      TestPartResultReporterInterface* reporter);
+
+  // Returns the test part result reporter for the current thread.
+  TestPartResultReporterInterface* GetTestPartResultReporterForCurrentThread();
+
+  // Sets the test part result reporter for the current thread.
+  void SetTestPartResultReporterForCurrentThread(
+      TestPartResultReporterInterface* reporter);
+
+  // Gets the number of successful test cases.
+  int successful_test_case_count() const;
+
+  // Gets the number of failed test cases.
+  int failed_test_case_count() const;
+
+  // Gets the number of all test cases.
+  int total_test_case_count() const;
+
+  // Gets the number of all test cases that contain at least one test
+  // that should run.
+  int test_case_to_run_count() const;
+
+  // Gets the number of successful tests.
+  int successful_test_count() const;
+
+  // Gets the number of failed tests.
+  int failed_test_count() const;
+
+  // Gets the number of disabled tests that will be reported in the XML report.
+  int reportable_disabled_test_count() const;
+
+  // Gets the number of disabled tests.
+  int disabled_test_count() const;
+
+  // Gets the number of tests to be printed in the XML report.
+  int reportable_test_count() const;
+
+  // Gets the number of all tests.
+  int total_test_count() const;
+
+  // Gets the number of tests that should run.
+  int test_to_run_count() const;
+
+  // Gets the time of the test program start, in ms from the start of the
+  // UNIX epoch.
+  TimeInMillis start_timestamp() const { return start_timestamp_; }
+
+  // Gets the elapsed time, in milliseconds.
+  TimeInMillis elapsed_time() const { return elapsed_time_; }
+
+  // Returns true iff the unit test passed (i.e. all test cases passed).
+  bool Passed() const { return !Failed(); }
+
+  // Returns true iff the unit test failed (i.e. some test case failed
+  // or something outside of all tests failed).
+  bool Failed() const {
+    return failed_test_case_count() > 0 || ad_hoc_test_result()->Failed();
+  }
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  const TestCase* GetTestCase(int i) const {
+    const int index = GetElementOr(test_case_indices_, i, -1);
+    return index < 0 ? NULL : test_cases_[i];
+  }
+
+  // Gets the i-th test case among all the test cases. i can range from 0 to
+  // total_test_case_count() - 1. If i is not in that range, returns NULL.
+  TestCase* GetMutableTestCase(int i) {
+    const int index = GetElementOr(test_case_indices_, i, -1);
+    return index < 0 ? NULL : test_cases_[index];
+  }
+
+  // Provides access to the event listener list.
+  TestEventListeners* listeners() { return &listeners_; }
+
+  // Returns the TestResult for the test that's currently running, or
+  // the TestResult for the ad hoc test if no test is running.
+  TestResult* current_test_result();
+
+  // Returns the TestResult for the ad hoc test.
+  const TestResult* ad_hoc_test_result() const { return &ad_hoc_test_result_; }
+
+  // Sets the OS stack trace getter.
+  //
+  // Does nothing if the input and the current OS stack trace getter
+  // are the same; otherwise, deletes the old getter and makes the
+  // input the current getter.
+  void set_os_stack_trace_getter(OsStackTraceGetterInterface* getter);
+
+  // Returns the current OS stack trace getter if it is not NULL;
+  // otherwise, creates an OsStackTraceGetter, makes it the current
+  // getter, and returns it.
+  OsStackTraceGetterInterface* os_stack_trace_getter();
+
+  // Returns the current OS stack trace as an std::string.
+  //
+  // The maximum number of stack frames to be included is specified by
+  // the gtest_stack_trace_depth flag.  The skip_count parameter
+  // specifies the number of top frames to be skipped, which doesn't
+  // count against the number of frames to be included.
+  //
+  // For example, if Foo() calls Bar(), which in turn calls
+  // CurrentOsStackTraceExceptTop(1), Foo() will be included in the
+  // trace but Bar() and CurrentOsStackTraceExceptTop() won't.
+  std::string CurrentOsStackTraceExceptTop(int skip_count) GTEST_NO_INLINE_;
+
+  // Finds and returns a TestCase with the given name.  If one doesn't
+  // exist, creates one and returns it.
+  //
+  // Arguments:
+  //
+  //   test_case_name: name of the test case
+  //   type_param:     the name of the test's type parameter, or NULL if
+  //                   this is not a typed or a type-parameterized test.
+  //   set_up_tc:      pointer to the function that sets up the test case
+  //   tear_down_tc:   pointer to the function that tears down the test case
+  TestCase* GetTestCase(const char* test_case_name,
+                        const char* type_param,
+                        Test::SetUpTestCaseFunc set_up_tc,
+                        Test::TearDownTestCaseFunc tear_down_tc);
+
+  // Adds a TestInfo to the unit test.
+  //
+  // Arguments:
+  //
+  //   set_up_tc:    pointer to the function that sets up the test case
+  //   tear_down_tc: pointer to the function that tears down the test case
+  //   test_info:    the TestInfo object
+  void AddTestInfo(Test::SetUpTestCaseFunc set_up_tc,
+                   Test::TearDownTestCaseFunc tear_down_tc,
+                   TestInfo* test_info) {
+    // In order to support thread-safe death tests, we need to
+    // remember the original working directory when the test program
+    // was first invoked.  We cannot do this in RUN_ALL_TESTS(), as
+    // the user may have changed the current directory before calling
+    // RUN_ALL_TESTS().  Therefore we capture the current directory in
+    // AddTestInfo(), which is called to register a TEST or TEST_F
+    // before main() is reached.
+    if (original_working_dir_.IsEmpty()) {
+      original_working_dir_.Set(FilePath::GetCurrentDir());
+      GTEST_CHECK_(!original_working_dir_.IsEmpty())
+          << "Failed to get the current working directory.";
+    }
+
+    GetTestCase(test_info->test_case_name(),
+                test_info->type_param(),
+                set_up_tc,
+                tear_down_tc)->AddTestInfo(test_info);
+  }
+
+#if GTEST_HAS_PARAM_TEST
+  // Returns ParameterizedTestCaseRegistry object used to keep track of
+  // value-parameterized tests and instantiate and register them.
+  internal::ParameterizedTestCaseRegistry& parameterized_test_registry() {
+    return parameterized_test_registry_;
+  }
+#endif  // GTEST_HAS_PARAM_TEST
+
+  // Sets the TestCase object for the test that's currently running.
+  void set_current_test_case(TestCase* a_current_test_case) {
+    current_test_case_ = a_current_test_case;
+  }
+
+  // Sets the TestInfo object for the test that's currently running.  If
+  // current_test_info is NULL, the assertion results will be stored in
+  // ad_hoc_test_result_.
+  void set_current_test_info(TestInfo* a_current_test_info) {
+    current_test_info_ = a_current_test_info;
+  }
+
+  // Registers all parameterized tests defined using TEST_P and
+  // INSTANTIATE_TEST_CASE_P, creating regular tests for each test/parameter
+  // combination. This method can be called more then once; it has guards
+  // protecting from registering the tests more then once.  If
+  // value-parameterized tests are disabled, RegisterParameterizedTests is
+  // present but does nothing.
+  void RegisterParameterizedTests();
+
+  // Runs all tests in this UnitTest object, prints the result, and
+  // returns true if all tests are successful.  If any exception is
+  // thrown during a test, this test is considered to be failed, but
+  // the rest of the tests will still be run.
+  bool RunAllTests();
+
+  // Clears the results of all tests, except the ad hoc tests.
+  void ClearNonAdHocTestResult() {
+    ForEach(test_cases_, TestCase::ClearTestCaseResult);
+  }
+
+  // Clears the results of ad-hoc test assertions.
+  void ClearAdHocTestResult() {
+    ad_hoc_test_result_.Clear();
+  }
+
+  // Adds a TestProperty to the current TestResult object when invoked in a
+  // context of a test or a test case, or to the global property set. If the
+  // result already contains a property with the same key, the value will be
+  // updated.
+  void RecordProperty(const TestProperty& test_property);
+
+  enum ReactionToSharding {
+    HONOR_SHARDING_PROTOCOL,
+    IGNORE_SHARDING_PROTOCOL
+  };
+
+  // Matches the full name of each test against the user-specified
+  // filter to decide whether the test should run, then records the
+  // result in each TestCase and TestInfo object.
+  // If shard_tests == HONOR_SHARDING_PROTOCOL, further filters tests
+  // based on sharding variables in the environment.
+  // Returns the number of tests that should run.
+  int FilterTests(ReactionToSharding shard_tests);
+
+  // Prints the names of the tests matching the user-specified filter flag.
+  void ListTestsMatchingFilter();
+
+  const TestCase* current_test_case() const { return current_test_case_; }
+  TestInfo* current_test_info() { return current_test_info_; }
+  const TestInfo* current_test_info() const { return current_test_info_; }
+
+  // Returns the vector of environments that need to be set-up/torn-down
+  // before/after the tests are run.
+  std::vector<Environment*>& environments() { return environments_; }
+
+  // Getters for the per-thread Google Test trace stack.
+  std::vector<TraceInfo>& gtest_trace_stack() {
+    return *(gtest_trace_stack_.pointer());
+  }
+  const std::vector<TraceInfo>& gtest_trace_stack() const {
+    return gtest_trace_stack_.get();
+  }
+
+#if GTEST_HAS_DEATH_TEST
+  void InitDeathTestSubprocessControlInfo() {
+    internal_run_death_test_flag_.reset(ParseInternalRunDeathTestFlag());
+  }
+  // Returns a pointer to the parsed --gtest_internal_run_death_test
+  // flag, or NULL if that flag was not specified.
+  // This information is useful only in a death test child process.
+  // Must not be called before a call to InitGoogleTest.
+  const InternalRunDeathTestFlag* internal_run_death_test_flag() const {
+    return internal_run_death_test_flag_.get();
+  }
+
+  // Returns a pointer to the current death test factory.
+  internal::DeathTestFactory* death_test_factory() {
+    return death_test_factory_.get();
+  }
+
+  void SuppressTestEventsIfInSubprocess();
+
+  friend class ReplaceDeathTestFactory;
+#endif  // GTEST_HAS_DEATH_TEST
+
+  // Initializes the event listener performing XML output as specified by
+  // UnitTestOptions. Must not be called before InitGoogleTest.
+  void ConfigureXmlOutput();
+
+#if GTEST_CAN_STREAM_RESULTS_
+  // Initializes the event listener for streaming test results to a socket.
+  // Must not be called before InitGoogleTest.
+  void ConfigureStreamingOutput();
+#endif
+
+  // Performs initialization dependent upon flag values obtained in
+  // ParseGoogleTestFlagsOnly.  Is called from InitGoogleTest after the call to
+  // ParseGoogleTestFlagsOnly.  In case a user neglects to call InitGoogleTest
+  // this function is also called from RunAllTests.  Since this function can be
+  // called more than once, it has to be idempotent.
+  void PostFlagParsingInit();
+
+  // Gets the random seed used at the start of the current test iteration.
+  int random_seed() const { return random_seed_; }
+
+  // Gets the random number generator.
+  internal::Random* random() { return &random_; }
+
+  // Shuffles all test cases, and the tests within each test case,
+  // making sure that death tests are still run first.
+  void ShuffleTests();
+
+  // Restores the test cases and tests to their order before the first shuffle.
+  void UnshuffleTests();
+
+  // Returns the value of GTEST_FLAG(catch_exceptions) at the moment
+  // UnitTest::Run() starts.
+  bool catch_exceptions() const { return catch_exceptions_; }
+
+ private:
+  friend class ::testing::UnitTest;
+
+  // Used by UnitTest::Run() to capture the state of
+  // GTEST_FLAG(catch_exceptions) at the moment it starts.
+  void set_catch_exceptions(bool value) { catch_exceptions_ = value; }
+
+  // The UnitTest object that owns this implementation object.
+  UnitTest* const parent_;
+
+  // The working directory when the first TEST() or TEST_F() was
+  // executed.
+  internal::FilePath original_working_dir_;
+
+  // The default test part result reporters.
+  DefaultGlobalTestPartResultReporter default_global_test_part_result_reporter_;
+  DefaultPerThreadTestPartResultReporter
+      default_per_thread_test_part_result_reporter_;
+
+  // Points to (but doesn't own) the global test part result reporter.
+  TestPartResultReporterInterface* global_test_part_result_repoter_;
+
+  // Protects read and write access to global_test_part_result_reporter_.
+  internal::Mutex global_test_part_result_reporter_mutex_;
+
+  // Points to (but doesn't own) the per-thread test part result reporter.
+  internal::ThreadLocal<TestPartResultReporterInterface*>
+      per_thread_test_part_result_reporter_;
+
+  // The vector of environments that need to be set-up/torn-down
+  // before/after the tests are run.
+  std::vector<Environment*> environments_;
+
+  // The vector of TestCases in their original order.  It owns the
+  // elements in the vector.
+  std::vector<TestCase*> test_cases_;
+
+  // Provides a level of indirection for the test case list to allow
+  // easy shuffling and restoring the test case order.  The i-th
+  // element of this vector is the index of the i-th test case in the
+  // shuffled order.
+  std::vector<int> test_case_indices_;
+
+#if GTEST_HAS_PARAM_TEST
+  // ParameterizedTestRegistry object used to register value-parameterized
+  // tests.
+  internal::ParameterizedTestCaseRegistry parameterized_test_registry_;
+
+  // Indicates whether RegisterParameterizedTests() has been called already.
+  bool parameterized_tests_registered_;
+#endif  // GTEST_HAS_PARAM_TEST
+
+  // Index of the last death test case registered.  Initially -1.
+  int last_death_test_case_;
+
+  // This points to the TestCase for the currently running test.  It
+  // changes as Google Test goes through one test case after another.
+  // When no test is running, this is set to NULL and Google Test
+  // stores assertion results in ad_hoc_test_result_.  Initially NULL.
+  TestCase* current_test_case_;
+
+  // This points to the TestInfo for the currently running test.  It
+  // changes as Google Test goes through one test after another.  When
+  // no test is running, this is set to NULL and Google Test stores
+  // assertion results in ad_hoc_test_result_.  Initially NULL.
+  TestInfo* current_test_info_;
+
+  // Normally, a user only writes assertions inside a TEST or TEST_F,
+  // or inside a function called by a TEST or TEST_F.  Since Google
+  // Test keeps track of which test is current running, it can
+  // associate such an assertion with the test it belongs to.
+  //
+  // If an assertion is encountered when no TEST or TEST_F is running,
+  // Google Test attributes the assertion result to an imaginary "ad hoc"
+  // test, and records the result in ad_hoc_test_result_.
+  TestResult ad_hoc_test_result_;
+
+  // The list of event listeners that can be used to track events inside
+  // Google Test.
+  TestEventListeners listeners_;
+
+  // The OS stack trace getter.  Will be deleted when the UnitTest
+  // object is destructed.  By default, an OsStackTraceGetter is used,
+  // but the user can set this field to use a custom getter if that is
+  // desired.
+  OsStackTraceGetterInterface* os_stack_trace_getter_;
+
+  // True iff PostFlagParsingInit() has been called.
+  bool post_flag_parse_init_performed_;
+
+  // The random number seed used at the beginning of the test run.
+  int random_seed_;
+
+  // Our random number generator.
+  internal::Random random_;
+
+  // The time of the test program start, in ms from the start of the
+  // UNIX epoch.
+  TimeInMillis start_timestamp_;
+
+  // How long the test took to run, in milliseconds.
+  TimeInMillis elapsed_time_;
+
+#if GTEST_HAS_DEATH_TEST
+  // The decomposed components of the gtest_internal_run_death_test flag,
+  // parsed when RUN_ALL_TESTS is called.
+  internal::scoped_ptr<InternalRunDeathTestFlag> internal_run_death_test_flag_;
+  internal::scoped_ptr<internal::DeathTestFactory> death_test_factory_;
+#endif  // GTEST_HAS_DEATH_TEST
+
+  // A per-thread stack of traces created by the SCOPED_TRACE() macro.
+  internal::ThreadLocal<std::vector<TraceInfo> > gtest_trace_stack_;
+
+  // The value of GTEST_FLAG(catch_exceptions) at the moment RunAllTests()
+  // starts.
+  bool catch_exceptions_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(UnitTestImpl);
+};  // class UnitTestImpl
+
+// Convenience function for accessing the global UnitTest
+// implementation object.
+inline UnitTestImpl* GetUnitTestImpl() {
+  return UnitTest::GetInstance()->impl();
+}
+
+#if GTEST_USES_SIMPLE_RE
+
+// Internal helper functions for implementing the simple regular
+// expression matcher.
+GTEST_API_ bool IsInSet(char ch, const char* str);
+GTEST_API_ bool IsAsciiDigit(char ch);
+GTEST_API_ bool IsAsciiPunct(char ch);
+GTEST_API_ bool IsRepeat(char ch);
+GTEST_API_ bool IsAsciiWhiteSpace(char ch);
+GTEST_API_ bool IsAsciiWordChar(char ch);
+GTEST_API_ bool IsValidEscape(char ch);
+GTEST_API_ bool AtomMatchesChar(bool escaped, char pattern, char ch);
+GTEST_API_ bool ValidateRegex(const char* regex);
+GTEST_API_ bool MatchRegexAtHead(const char* regex, const char* str);
+GTEST_API_ bool MatchRepetitionAndRegexAtHead(
+    bool escaped, char ch, char repeat, const char* regex, const char* str);
+GTEST_API_ bool MatchRegexAnywhere(const char* regex, const char* str);
+
+#endif  // GTEST_USES_SIMPLE_RE
+
+// Parses the command line for Google Test flags, without initializing
+// other parts of Google Test.
+GTEST_API_ void ParseGoogleTestFlagsOnly(int* argc, char** argv);
+GTEST_API_ void ParseGoogleTestFlagsOnly(int* argc, wchar_t** argv);
+
+#if GTEST_HAS_DEATH_TEST
+
+// Returns the message describing the last system error, regardless of the
+// platform.
+GTEST_API_ std::string GetLastErrnoDescription();
+
+// Attempts to parse a string into a positive integer pointed to by the
+// number parameter.  Returns true if that is possible.
+// GTEST_HAS_DEATH_TEST implies that we have ::std::string, so we can use
+// it here.
+template <typename Integer>
+bool ParseNaturalNumber(const ::std::string& str, Integer* number) {
+  // Fail fast if the given string does not begin with a digit;
+  // this bypasses strtoXXX's "optional leading whitespace and plus
+  // or minus sign" semantics, which are undesirable here.
+  if (str.empty() || !IsDigit(str[0])) {
+    return false;
+  }
+  errno = 0;
+
+  char* end;
+  // BiggestConvertible is the largest integer type that system-provided
+  // string-to-number conversion routines can return.
+
+# if GTEST_OS_WINDOWS && !defined(__GNUC__)
+
+  // MSVC and C++ Builder define __int64 instead of the standard long long.
+  typedef unsigned __int64 BiggestConvertible;
+  const BiggestConvertible parsed = _strtoui64(str.c_str(), &end, 10);
+
+# else
+
+  typedef unsigned long long BiggestConvertible;  // NOLINT
+  const BiggestConvertible parsed = strtoull(str.c_str(), &end, 10);
+
+# endif  // GTEST_OS_WINDOWS && !defined(__GNUC__)
+
+  const bool parse_success = *end == '\0' && errno == 0;
+
+  // TODO(vladl@google.com): Convert this to compile time assertion when it is
+  // available.
+  GTEST_CHECK_(sizeof(Integer) <= sizeof(parsed));
+
+  const Integer result = static_cast<Integer>(parsed);
+  if (parse_success && static_cast<BiggestConvertible>(result) == parsed) {
+    *number = result;
+    return true;
+  }
+  return false;
+}
+#endif  // GTEST_HAS_DEATH_TEST
+
+// TestResult contains some private methods that should be hidden from
+// Google Test user but are required for testing. This class allow our tests
+// to access them.
+//
+// This class is supplied only for the purpose of testing Google Test's own
+// constructs. Do not use it in user tests, either directly or indirectly.
+class TestResultAccessor {
+ public:
+  static void RecordProperty(TestResult* test_result,
+                             const std::string& xml_element,
+                             const TestProperty& property) {
+    test_result->RecordProperty(xml_element, property);
+  }
+
+  static void ClearTestPartResults(TestResult* test_result) {
+    test_result->ClearTestPartResults();
+  }
+
+  static const std::vector<testing::TestPartResult>& test_part_results(
+      const TestResult& test_result) {
+    return test_result.test_part_results();
+  }
+};
+
+#if GTEST_CAN_STREAM_RESULTS_
+
+// Streams test results to the given port on the given host machine.
+class GTEST_API_ StreamingListener : public EmptyTestEventListener {
+ public:
+  // Abstract base class for writing strings to a socket.
+  class AbstractSocketWriter {
+   public:
+    virtual ~AbstractSocketWriter() {}
+
+    // Sends a string to the socket.
+    virtual void Send(const string& message) = 0;
+
+    // Closes the socket.
+    virtual void CloseConnection() {}
+
+    // Sends a string and a newline to the socket.
+    void SendLn(const string& message) {
+      Send(message + "\n");
+    }
+  };
+
+  // Concrete class for actually writing strings to a socket.
+  class SocketWriter : public AbstractSocketWriter {
+   public:
+    SocketWriter(const string& host, const string& port)
+        : sockfd_(-1), host_name_(host), port_num_(port) {
+      MakeConnection();
+    }
+
+    virtual ~SocketWriter() {
+      if (sockfd_ != -1)
+        CloseConnection();
+    }
+
+    // Sends a string to the socket.
+    virtual void Send(const string& message) {
+      GTEST_CHECK_(sockfd_ != -1)
+          << "Send() can be called only when there is a connection.";
+
+      const int len = static_cast<int>(message.length());
+      if (write(sockfd_, message.c_str(), len) != len) {
+        GTEST_LOG_(WARNING)
+            << "stream_result_to: failed to stream to "
+            << host_name_ << ":" << port_num_;
+      }
+    }
+
+   private:
+    // Creates a client socket and connects to the server.
+    void MakeConnection();
+
+    // Closes the socket.
+    void CloseConnection() {
+      GTEST_CHECK_(sockfd_ != -1)
+          << "CloseConnection() can be called only when there is a connection.";
+
+      close(sockfd_);
+      sockfd_ = -1;
+    }
+
+    int sockfd_;  // socket file descriptor
+    const string host_name_;
+    const string port_num_;
+
+    GTEST_DISALLOW_COPY_AND_ASSIGN_(SocketWriter);
+  };  // class SocketWriter
+
+  // Escapes '=', '&', '%', and '\n' characters in str as "%xx".
+  static string UrlEncode(const char* str);
+
+  StreamingListener(const string& host, const string& port)
+      : socket_writer_(new SocketWriter(host, port)) { Start(); }
+
+  explicit StreamingListener(AbstractSocketWriter* socket_writer)
+      : socket_writer_(socket_writer) { Start(); }
+
+  void OnTestProgramStart(const UnitTest& /* unit_test */) {
+    SendLn("event=TestProgramStart");
+  }
+
+  void OnTestProgramEnd(const UnitTest& unit_test) {
+    // Note that Google Test current only report elapsed time for each
+    // test iteration, not for the entire test program.
+    SendLn("event=TestProgramEnd&passed=" + FormatBool(unit_test.Passed()));
+
+    // Notify the streaming server to stop.
+    socket_writer_->CloseConnection();
+  }
+
+  void OnTestIterationStart(const UnitTest& /* unit_test */, int iteration) {
+    SendLn("event=TestIterationStart&iteration=" +
+           StreamableToString(iteration));
+  }
+
+  void OnTestIterationEnd(const UnitTest& unit_test, int /* iteration */) {
+    SendLn("event=TestIterationEnd&passed=" +
+           FormatBool(unit_test.Passed()) + "&elapsed_time=" +
+           StreamableToString(unit_test.elapsed_time()) + "ms");
+  }
+
+  void OnTestCaseStart(const TestCase& test_case) {
+    SendLn(std::string("event=TestCaseStart&name=") + test_case.name());
+  }
+
+  void OnTestCaseEnd(const TestCase& test_case) {
+    SendLn("event=TestCaseEnd&passed=" + FormatBool(test_case.Passed())
+           + "&elapsed_time=" + StreamableToString(test_case.elapsed_time())
+           + "ms");
+  }
+
+  void OnTestStart(const TestInfo& test_info) {
+    SendLn(std::string("event=TestStart&name=") + test_info.name());
+  }
+
+  void OnTestEnd(const TestInfo& test_info) {
+    SendLn("event=TestEnd&passed=" +
+           FormatBool((test_info.result())->Passed()) +
+           "&elapsed_time=" +
+           StreamableToString((test_info.result())->elapsed_time()) + "ms");
+  }
+
+  void OnTestPartResult(const TestPartResult& test_part_result) {
+    const char* file_name = test_part_result.file_name();
+    if (file_name == NULL)
+      file_name = "";
+    SendLn("event=TestPartResult&file=" + UrlEncode(file_name) +
+           "&line=" + StreamableToString(test_part_result.line_number()) +
+           "&message=" + UrlEncode(test_part_result.message()));
+  }
+
+ private:
+  // Sends the given message and a newline to the socket.
+  void SendLn(const string& message) { socket_writer_->SendLn(message); }
+
+  // Called at the start of streaming to notify the receiver what
+  // protocol we are using.
+  void Start() { SendLn("gtest_streaming_protocol_version=1.0"); }
+
+  string FormatBool(bool value) { return value ? "1" : "0"; }
+
+  const scoped_ptr<AbstractSocketWriter> socket_writer_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(StreamingListener);
+};  // class StreamingListener
+
+#endif  // GTEST_CAN_STREAM_RESULTS_
+
+}  // namespace internal
+}  // namespace testing
+
+#endif  // GTEST_SRC_GTEST_INTERNAL_INL_H_
diff --git a/third_party/aom/third_party/googletest/src/googletest/src/gtest-port.cc b/third_party/aom/third_party/googletest/src/googletest/src/gtest-port.cc
new file mode 100644
index 0000000000..e5bf3dd2be
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/src/gtest-port.cc
@@ -0,0 +1,1259 @@
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+#include "gtest/internal/gtest-port.h"
+
+#include <limits.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <fstream>
+
+#if GTEST_OS_WINDOWS
+# include <windows.h>
+# include <io.h>
+# include <sys/stat.h>
+# include <map>  // Used in ThreadLocal.
+#else
+# include <unistd.h>
+#endif  // GTEST_OS_WINDOWS
+
+#if GTEST_OS_MAC
+# include <mach/mach_init.h>
+# include <mach/task.h>
+# include <mach/vm_map.h>
+#endif  // GTEST_OS_MAC
+
+#if GTEST_OS_QNX
+# include <devctl.h>
+# include <fcntl.h>
+# include <sys/procfs.h>
+#endif  // GTEST_OS_QNX
+
+#if GTEST_OS_AIX
+# include <procinfo.h>
+# include <sys/types.h>
+#endif  // GTEST_OS_AIX
+
+#include "gtest/gtest-spi.h"
+#include "gtest/gtest-message.h"
+#include "gtest/internal/gtest-internal.h"
+#include "gtest/internal/gtest-string.h"
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick exists to
+// prevent the accidental inclusion of gtest-internal-inl.h in the
+// user's code.
+#define GTEST_IMPLEMENTATION_ 1
+#include "src/gtest-internal-inl.h"
+#undef GTEST_IMPLEMENTATION_
+
+namespace testing {
+namespace internal {
+
+#if defined(_MSC_VER) || defined(__BORLANDC__)
+// MSVC and C++Builder do not provide a definition of STDERR_FILENO.
+const int kStdOutFileno = 1;
+const int kStdErrFileno = 2;
+#else
+const int kStdOutFileno = STDOUT_FILENO;
+const int kStdErrFileno = STDERR_FILENO;
+#endif  // _MSC_VER
+
+#if GTEST_OS_LINUX
+
+namespace {
+template <typename T>
+T ReadProcFileField(const string& filename, int field) {
+  std::string dummy;
+  std::ifstream file(filename.c_str());
+  while (field-- > 0) {
+    file >> dummy;
+  }
+  T output = 0;
+  file >> output;
+  return output;
+}
+}  // namespace
+
+// Returns the number of active threads, or 0 when there is an error.
+size_t GetThreadCount() {
+  const string filename =
+      (Message() << "/proc/" << getpid() << "/stat").GetString();
+  return ReadProcFileField<int>(filename, 19);
+}
+
+#elif GTEST_OS_MAC
+
+size_t GetThreadCount() {
+  const task_t task = mach_task_self();
+  mach_msg_type_number_t thread_count;
+  thread_act_array_t thread_list;
+  const kern_return_t status = task_threads(task, &thread_list, &thread_count);
+  if (status == KERN_SUCCESS) {
+    // task_threads allocates resources in thread_list and we need to free them
+    // to avoid leaks.
+    vm_deallocate(task,
+                  reinterpret_cast<vm_address_t>(thread_list),
+                  sizeof(thread_t) * thread_count);
+    return static_cast<size_t>(thread_count);
+  } else {
+    return 0;
+  }
+}
+
+#elif GTEST_OS_QNX
+
+// Returns the number of threads running in the process, or 0 to indicate that
+// we cannot detect it.
+size_t GetThreadCount() {
+  const int fd = open("/proc/self/as", O_RDONLY);
+  if (fd < 0) {
+    return 0;
+  }
+  procfs_info process_info;
+  const int status =
+      devctl(fd, DCMD_PROC_INFO, &process_info, sizeof(process_info), NULL);
+  close(fd);
+  if (status == EOK) {
+    return static_cast<size_t>(process_info.num_threads);
+  } else {
+    return 0;
+  }
+}
+
+#elif GTEST_OS_AIX
+
+size_t GetThreadCount() {
+  struct procentry64 entry;
+  pid_t pid = getpid();
+  int status = getprocs64(&entry, sizeof(entry), NULL, 0, &pid, 1);
+  if (status == 1) {
+    return entry.pi_thcount;
+  } else {
+    return 0;
+  }
+}
+
+#else
+
+size_t GetThreadCount() {
+  // There's no portable way to detect the number of threads, so we just
+  // return 0 to indicate that we cannot detect it.
+  return 0;
+}
+
+#endif  // GTEST_OS_LINUX
+
+#if GTEST_IS_THREADSAFE && GTEST_OS_WINDOWS
+
+void SleepMilliseconds(int n) {
+  ::Sleep(n);
+}
+
+AutoHandle::AutoHandle()
+    : handle_(INVALID_HANDLE_VALUE) {}
+
+AutoHandle::AutoHandle(Handle handle)
+    : handle_(handle) {}
+
+AutoHandle::~AutoHandle() {
+  Reset();
+}
+
+AutoHandle::Handle AutoHandle::Get() const {
+  return handle_;
+}
+
+void AutoHandle::Reset() {
+  Reset(INVALID_HANDLE_VALUE);
+}
+
+void AutoHandle::Reset(HANDLE handle) {
+  // Resetting with the same handle we already own is invalid.
+  if (handle_ != handle) {
+    if (IsCloseable()) {
+      ::CloseHandle(handle_);
+    }
+    handle_ = handle;
+  } else {
+    GTEST_CHECK_(!IsCloseable())
+        << "Resetting a valid handle to itself is likely a programmer error "
+            "and thus not allowed.";
+  }
+}
+
+bool AutoHandle::IsCloseable() const {
+  // Different Windows APIs may use either of these values to represent an
+  // invalid handle.
+  return handle_ != NULL && handle_ != INVALID_HANDLE_VALUE;
+}
+
+Notification::Notification()
+    : event_(::CreateEvent(NULL,   // Default security attributes.
+                           TRUE,   // Do not reset automatically.
+                           FALSE,  // Initially unset.
+                           NULL)) {  // Anonymous event.
+  GTEST_CHECK_(event_.Get() != NULL);
+}
+
+void Notification::Notify() {
+  GTEST_CHECK_(::SetEvent(event_.Get()) != FALSE);
+}
+
+void Notification::WaitForNotification() {
+  GTEST_CHECK_(
+      ::WaitForSingleObject(event_.Get(), INFINITE) == WAIT_OBJECT_0);
+}
+
+Mutex::Mutex()
+    : owner_thread_id_(0),
+      type_(kDynamic),
+      critical_section_init_phase_(0),
+      critical_section_(new CRITICAL_SECTION) {
+  ::InitializeCriticalSection(critical_section_);
+}
+
+Mutex::~Mutex() {
+  // Static mutexes are leaked intentionally. It is not thread-safe to try
+  // to clean them up.
+  // TODO(yukawa): Switch to Slim Reader/Writer (SRW) Locks, which requires
+  // nothing to clean it up but is available only on Vista and later.
+  // http://msdn.microsoft.com/en-us/library/windows/desktop/aa904937.aspx
+  if (type_ == kDynamic) {
+    ::DeleteCriticalSection(critical_section_);
+    delete critical_section_;
+    critical_section_ = NULL;
+  }
+}
+
+void Mutex::Lock() {
+  ThreadSafeLazyInit();
+  ::EnterCriticalSection(critical_section_);
+  owner_thread_id_ = ::GetCurrentThreadId();
+}
+
+void Mutex::Unlock() {
+  ThreadSafeLazyInit();
+  // We don't protect writing to owner_thread_id_ here, as it's the
+  // caller's responsibility to ensure that the current thread holds the
+  // mutex when this is called.
+  owner_thread_id_ = 0;
+  ::LeaveCriticalSection(critical_section_);
+}
+
+// Does nothing if the current thread holds the mutex. Otherwise, crashes
+// with high probability.
+void Mutex::AssertHeld() {
+  ThreadSafeLazyInit();
+  GTEST_CHECK_(owner_thread_id_ == ::GetCurrentThreadId())
+      << "The current thread is not holding the mutex @" << this;
+}
+
+// Initializes owner_thread_id_ and critical_section_ in static mutexes.
+void Mutex::ThreadSafeLazyInit() {
+  // Dynamic mutexes are initialized in the constructor.
+  if (type_ == kStatic) {
+    switch (
+        ::InterlockedCompareExchange(&critical_section_init_phase_, 1L, 0L)) {
+      case 0:
+        // If critical_section_init_phase_ was 0 before the exchange, we
+        // are the first to test it and need to perform the initialization.
+        owner_thread_id_ = 0;
+        critical_section_ = new CRITICAL_SECTION;
+        ::InitializeCriticalSection(critical_section_);
+        // Updates the critical_section_init_phase_ to 2 to signal
+        // initialization complete.
+        GTEST_CHECK_(::InterlockedCompareExchange(
+                          &critical_section_init_phase_, 2L, 1L) ==
+                      1L);
+        break;
+      case 1:
+        // Somebody else is already initializing the mutex; spin until they
+        // are done.
+        while (::InterlockedCompareExchange(&critical_section_init_phase_,
+                                            2L,
+                                            2L) != 2L) {
+          // Possibly yields the rest of the thread's time slice to other
+          // threads.
+          ::Sleep(0);
+        }
+        break;
+
+      case 2:
+        break;  // The mutex is already initialized and ready for use.
+
+      default:
+        GTEST_CHECK_(false)
+            << "Unexpected value of critical_section_init_phase_ "
+            << "while initializing a static mutex.";
+    }
+  }
+}
+
+namespace {
+
+class ThreadWithParamSupport : public ThreadWithParamBase {
+ public:
+  static HANDLE CreateThread(Runnable* runnable,
+                             Notification* thread_can_start) {
+    ThreadMainParam* param = new ThreadMainParam(runnable, thread_can_start);
+    DWORD thread_id;
+    // TODO(yukawa): Consider to use _beginthreadex instead.
+    HANDLE thread_handle = ::CreateThread(
+        NULL,    // Default security.
+        0,       // Default stack size.
+        &ThreadWithParamSupport::ThreadMain,
+        param,   // Parameter to ThreadMainStatic
+        0x0,     // Default creation flags.
+        &thread_id);  // Need a valid pointer for the call to work under Win98.
+    GTEST_CHECK_(thread_handle != NULL) << "CreateThread failed with error "
+                                        << ::GetLastError() << ".";
+    if (thread_handle == NULL) {
+      delete param;
+    }
+    return thread_handle;
+  }
+
+ private:
+  struct ThreadMainParam {
+    ThreadMainParam(Runnable* runnable, Notification* thread_can_start)
+        : runnable_(runnable),
+          thread_can_start_(thread_can_start) {
+    }
+    scoped_ptr<Runnable> runnable_;
+    // Does not own.
+    Notification* thread_can_start_;
+  };
+
+  static DWORD WINAPI ThreadMain(void* ptr) {
+    // Transfers ownership.
+    scoped_ptr<ThreadMainParam> param(static_cast<ThreadMainParam*>(ptr));
+    if (param->thread_can_start_ != NULL)
+      param->thread_can_start_->WaitForNotification();
+    param->runnable_->Run();
+    return 0;
+  }
+
+  // Prohibit instantiation.
+  ThreadWithParamSupport();
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ThreadWithParamSupport);
+};
+
+}  // namespace
+
+ThreadWithParamBase::ThreadWithParamBase(Runnable *runnable,
+                                         Notification* thread_can_start)
+      : thread_(ThreadWithParamSupport::CreateThread(runnable,
+                                                     thread_can_start)) {
+}
+
+ThreadWithParamBase::~ThreadWithParamBase() {
+  Join();
+}
+
+void ThreadWithParamBase::Join() {
+  GTEST_CHECK_(::WaitForSingleObject(thread_.Get(), INFINITE) == WAIT_OBJECT_0)
+      << "Failed to join the thread with error " << ::GetLastError() << ".";
+}
+
+// Maps a thread to a set of ThreadIdToThreadLocals that have values
+// instantiated on that thread and notifies them when the thread exits.  A
+// ThreadLocal instance is expected to persist until all threads it has
+// values on have terminated.
+class ThreadLocalRegistryImpl {
+ public:
+  // Registers thread_local_instance as having value on the current thread.
+  // Returns a value that can be used to identify the thread from other threads.
+  static ThreadLocalValueHolderBase* GetValueOnCurrentThread(
+      const ThreadLocalBase* thread_local_instance) {
+    DWORD current_thread = ::GetCurrentThreadId();
+    MutexLock lock(&mutex_);
+    ThreadIdToThreadLocals* const thread_to_thread_locals =
+        GetThreadLocalsMapLocked();
+    ThreadIdToThreadLocals::iterator thread_local_pos =
+        thread_to_thread_locals->find(current_thread);
+    if (thread_local_pos == thread_to_thread_locals->end()) {
+      thread_local_pos = thread_to_thread_locals->insert(
+          std::make_pair(current_thread, ThreadLocalValues())).first;
+      StartWatcherThreadFor(current_thread);
+    }
+    ThreadLocalValues& thread_local_values = thread_local_pos->second;
+    ThreadLocalValues::iterator value_pos =
+        thread_local_values.find(thread_local_instance);
+    if (value_pos == thread_local_values.end()) {
+      value_pos =
+          thread_local_values
+              .insert(std::make_pair(
+                  thread_local_instance,
+                  linked_ptr<ThreadLocalValueHolderBase>(
+                      thread_local_instance->NewValueForCurrentThread())))
+              .first;
+    }
+    return value_pos->second.get();
+  }
+
+  static void OnThreadLocalDestroyed(
+      const ThreadLocalBase* thread_local_instance) {
+    std::vector<linked_ptr<ThreadLocalValueHolderBase> > value_holders;
+    // Clean up the ThreadLocalValues data structure while holding the lock, but
+    // defer the destruction of the ThreadLocalValueHolderBases.
+    {
+      MutexLock lock(&mutex_);
+      ThreadIdToThreadLocals* const thread_to_thread_locals =
+          GetThreadLocalsMapLocked();
+      for (ThreadIdToThreadLocals::iterator it =
+          thread_to_thread_locals->begin();
+          it != thread_to_thread_locals->end();
+          ++it) {
+        ThreadLocalValues& thread_local_values = it->second;
+        ThreadLocalValues::iterator value_pos =
+            thread_local_values.find(thread_local_instance);
+        if (value_pos != thread_local_values.end()) {
+          value_holders.push_back(value_pos->second);
+          thread_local_values.erase(value_pos);
+          // This 'if' can only be successful at most once, so theoretically we
+          // could break out of the loop here, but we don't bother doing so.
+        }
+      }
+    }
+    // Outside the lock, let the destructor for 'value_holders' deallocate the
+    // ThreadLocalValueHolderBases.
+  }
+
+  static void OnThreadExit(DWORD thread_id) {
+    GTEST_CHECK_(thread_id != 0) << ::GetLastError();
+    std::vector<linked_ptr<ThreadLocalValueHolderBase> > value_holders;
+    // Clean up the ThreadIdToThreadLocals data structure while holding the
+    // lock, but defer the destruction of the ThreadLocalValueHolderBases.
+    {
+      MutexLock lock(&mutex_);
+      ThreadIdToThreadLocals* const thread_to_thread_locals =
+          GetThreadLocalsMapLocked();
+      ThreadIdToThreadLocals::iterator thread_local_pos =
+          thread_to_thread_locals->find(thread_id);
+      if (thread_local_pos != thread_to_thread_locals->end()) {
+        ThreadLocalValues& thread_local_values = thread_local_pos->second;
+        for (ThreadLocalValues::iterator value_pos =
+            thread_local_values.begin();
+            value_pos != thread_local_values.end();
+            ++value_pos) {
+          value_holders.push_back(value_pos->second);
+        }
+        thread_to_thread_locals->erase(thread_local_pos);
+      }
+    }
+    // Outside the lock, let the destructor for 'value_holders' deallocate the
+    // ThreadLocalValueHolderBases.
+  }
+
+ private:
+  // In a particular thread, maps a ThreadLocal object to its value.
+  typedef std::map<const ThreadLocalBase*,
+                   linked_ptr<ThreadLocalValueHolderBase> > ThreadLocalValues;
+  // Stores all ThreadIdToThreadLocals having values in a thread, indexed by
+  // thread's ID.
+  typedef std::map<DWORD, ThreadLocalValues> ThreadIdToThreadLocals;
+
+  // Holds the thread id and thread handle that we pass from
+  // StartWatcherThreadFor to WatcherThreadFunc.
+  typedef std::pair<DWORD, HANDLE> ThreadIdAndHandle;
+
+  static void StartWatcherThreadFor(DWORD thread_id) {
+    // The returned handle will be kept in thread_map and closed by
+    // watcher_thread in WatcherThreadFunc.
+    HANDLE thread = ::OpenThread(SYNCHRONIZE | THREAD_QUERY_INFORMATION,
+                                 FALSE,
+                                 thread_id);
+    GTEST_CHECK_(thread != NULL);
+    // We need to to pass a valid thread ID pointer into CreateThread for it
+    // to work correctly under Win98.
+    DWORD watcher_thread_id;
+    HANDLE watcher_thread = ::CreateThread(
+        NULL,   // Default security.
+        0,      // Default stack size
+        &ThreadLocalRegistryImpl::WatcherThreadFunc,
+        reinterpret_cast<LPVOID>(new ThreadIdAndHandle(thread_id, thread)),
+        CREATE_SUSPENDED,
+        &watcher_thread_id);
+    GTEST_CHECK_(watcher_thread != NULL);
+    // Give the watcher thread the same priority as ours to avoid being
+    // blocked by it.
+    ::SetThreadPriority(watcher_thread,
+                        ::GetThreadPriority(::GetCurrentThread()));
+    ::ResumeThread(watcher_thread);
+    ::CloseHandle(watcher_thread);
+  }
+
+  // Monitors exit from a given thread and notifies those
+  // ThreadIdToThreadLocals about thread termination.
+  static DWORD WINAPI WatcherThreadFunc(LPVOID param) {
+    const ThreadIdAndHandle* tah =
+        reinterpret_cast<const ThreadIdAndHandle*>(param);
+    GTEST_CHECK_(
+        ::WaitForSingleObject(tah->second, INFINITE) == WAIT_OBJECT_0);
+    OnThreadExit(tah->first);
+    ::CloseHandle(tah->second);
+    delete tah;
+    return 0;
+  }
+
+  // Returns map of thread local instances.
+  static ThreadIdToThreadLocals* GetThreadLocalsMapLocked() {
+    mutex_.AssertHeld();
+    static ThreadIdToThreadLocals* map = new ThreadIdToThreadLocals;
+    return map;
+  }
+
+  // Protects access to GetThreadLocalsMapLocked() and its return value.
+  static Mutex mutex_;
+  // Protects access to GetThreadMapLocked() and its return value.
+  static Mutex thread_map_mutex_;
+};
+
+Mutex ThreadLocalRegistryImpl::mutex_(Mutex::kStaticMutex);
+Mutex ThreadLocalRegistryImpl::thread_map_mutex_(Mutex::kStaticMutex);
+
+ThreadLocalValueHolderBase* ThreadLocalRegistry::GetValueOnCurrentThread(
+      const ThreadLocalBase* thread_local_instance) {
+  return ThreadLocalRegistryImpl::GetValueOnCurrentThread(
+      thread_local_instance);
+}
+
+void ThreadLocalRegistry::OnThreadLocalDestroyed(
+      const ThreadLocalBase* thread_local_instance) {
+  ThreadLocalRegistryImpl::OnThreadLocalDestroyed(thread_local_instance);
+}
+
+#endif  // GTEST_IS_THREADSAFE && GTEST_OS_WINDOWS
+
+#if GTEST_USES_POSIX_RE
+
+// Implements RE.  Currently only needed for death tests.
+
+RE::~RE() {
+  if (is_valid_) {
+    // regfree'ing an invalid regex might crash because the content
+    // of the regex is undefined. Since the regex's are essentially
+    // the same, one cannot be valid (or invalid) without the other
+    // being so too.
+    regfree(&partial_regex_);
+    regfree(&full_regex_);
+  }
+  free(const_cast<char*>(pattern_));
+}
+
+// Returns true iff regular expression re matches the entire str.
+bool RE::FullMatch(const char* str, const RE& re) {
+  if (!re.is_valid_) return false;
+
+  regmatch_t match;
+  return regexec(&re.full_regex_, str, 1, &match, 0) == 0;
+}
+
+// Returns true iff regular expression re matches a substring of str
+// (including str itself).
+bool RE::PartialMatch(const char* str, const RE& re) {
+  if (!re.is_valid_) return false;
+
+  regmatch_t match;
+  return regexec(&re.partial_regex_, str, 1, &match, 0) == 0;
+}
+
+// Initializes an RE from its string representation.
+void RE::Init(const char* regex) {
+  pattern_ = posix::StrDup(regex);
+
+  // Reserves enough bytes to hold the regular expression used for a
+  // full match.
+  const size_t full_regex_len = strlen(regex) + 10;
+  char* const full_pattern = new char[full_regex_len];
+
+  snprintf(full_pattern, full_regex_len, "^(%s)$", regex);
+  is_valid_ = regcomp(&full_regex_, full_pattern, REG_EXTENDED) == 0;
+  // We want to call regcomp(&partial_regex_, ...) even if the
+  // previous expression returns false.  Otherwise partial_regex_ may
+  // not be properly initialized can may cause trouble when it's
+  // freed.
+  //
+  // Some implementation of POSIX regex (e.g. on at least some
+  // versions of Cygwin) doesn't accept the empty string as a valid
+  // regex.  We change it to an equivalent form "()" to be safe.
+  if (is_valid_) {
+    const char* const partial_regex = (*regex == '\0') ? "()" : regex;
+    is_valid_ = regcomp(&partial_regex_, partial_regex, REG_EXTENDED) == 0;
+  }
+  EXPECT_TRUE(is_valid_)
+      << "Regular expression \"" << regex
+      << "\" is not a valid POSIX Extended regular expression.";
+
+  delete[] full_pattern;
+}
+
+#elif GTEST_USES_SIMPLE_RE
+
+// Returns true iff ch appears anywhere in str (excluding the
+// terminating '\0' character).
+bool IsInSet(char ch, const char* str) {
+  return ch != '\0' && strchr(str, ch) != NULL;
+}
+
+// Returns true iff ch belongs to the given classification.  Unlike
+// similar functions in <ctype.h>, these aren't affected by the
+// current locale.
+bool IsAsciiDigit(char ch) { return '0' <= ch && ch <= '9'; }
+bool IsAsciiPunct(char ch) {
+  return IsInSet(ch, "^-!\"#$%&'()*+,./:;<=>?@[\\]_`{|}~");
+}
+bool IsRepeat(char ch) { return IsInSet(ch, "?*+"); }
+bool IsAsciiWhiteSpace(char ch) { return IsInSet(ch, " \f\n\r\t\v"); }
+bool IsAsciiWordChar(char ch) {
+  return ('a' <= ch && ch <= 'z') || ('A' <= ch && ch <= 'Z') ||
+      ('0' <= ch && ch <= '9') || ch == '_';
+}
+
+// Returns true iff "\\c" is a supported escape sequence.
+bool IsValidEscape(char c) {
+  return (IsAsciiPunct(c) || IsInSet(c, "dDfnrsStvwW"));
+}
+
+// Returns true iff the given atom (specified by escaped and pattern)
+// matches ch.  The result is undefined if the atom is invalid.
+bool AtomMatchesChar(bool escaped, char pattern_char, char ch) {
+  if (escaped) {  // "\\p" where p is pattern_char.
+    switch (pattern_char) {
+      case 'd': return IsAsciiDigit(ch);
+      case 'D': return !IsAsciiDigit(ch);
+      case 'f': return ch == '\f';
+      case 'n': return ch == '\n';
+      case 'r': return ch == '\r';
+      case 's': return IsAsciiWhiteSpace(ch);
+      case 'S': return !IsAsciiWhiteSpace(ch);
+      case 't': return ch == '\t';
+      case 'v': return ch == '\v';
+      case 'w': return IsAsciiWordChar(ch);
+      case 'W': return !IsAsciiWordChar(ch);
+    }
+    return IsAsciiPunct(pattern_char) && pattern_char == ch;
+  }
+
+  return (pattern_char == '.' && ch != '\n') || pattern_char == ch;
+}
+
+// Helper function used by ValidateRegex() to format error messages.
+std::string FormatRegexSyntaxError(const char* regex, int index) {
+  return (Message() << "Syntax error at index " << index
+          << " in simple regular expression \"" << regex << "\": ").GetString();
+}
+
+// Generates non-fatal failures and returns false if regex is invalid;
+// otherwise returns true.
+bool ValidateRegex(const char* regex) {
+  if (regex == NULL) {
+    // TODO(wan@google.com): fix the source file location in the
+    // assertion failures to match where the regex is used in user
+    // code.
+    ADD_FAILURE() << "NULL is not a valid simple regular expression.";
+    return false;
+  }
+
+  bool is_valid = true;
+
+  // True iff ?, *, or + can follow the previous atom.
+  bool prev_repeatable = false;
+  for (int i = 0; regex[i]; i++) {
+    if (regex[i] == '\\') {  // An escape sequence
+      i++;
+      if (regex[i] == '\0') {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i - 1)
+                      << "'\\' cannot appear at the end.";
+        return false;
+      }
+
+      if (!IsValidEscape(regex[i])) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i - 1)
+                      << "invalid escape sequence \"\\" << regex[i] << "\".";
+        is_valid = false;
+      }
+      prev_repeatable = true;
+    } else {  // Not an escape sequence.
+      const char ch = regex[i];
+
+      if (ch == '^' && i > 0) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'^' can only appear at the beginning.";
+        is_valid = false;
+      } else if (ch == '$' && regex[i + 1] != '\0') {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'$' can only appear at the end.";
+        is_valid = false;
+      } else if (IsInSet(ch, "()[]{}|")) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'" << ch << "' is unsupported.";
+        is_valid = false;
+      } else if (IsRepeat(ch) && !prev_repeatable) {
+        ADD_FAILURE() << FormatRegexSyntaxError(regex, i)
+                      << "'" << ch << "' can only follow a repeatable token.";
+        is_valid = false;
+      }
+
+      prev_repeatable = !IsInSet(ch, "^$?*+");
+    }
+  }
+
+  return is_valid;
+}
+
+// Matches a repeated regex atom followed by a valid simple regular
+// expression.  The regex atom is defined as c if escaped is false,
+// or \c otherwise.  repeat is the repetition meta character (?, *,
+// or +).  The behavior is undefined if str contains too many
+// characters to be indexable by size_t, in which case the test will
+// probably time out anyway.  We are fine with this limitation as
+// std::string has it too.
+bool MatchRepetitionAndRegexAtHead(
+    bool escaped, char c, char repeat, const char* regex,
+    const char* str) {
+  const size_t min_count = (repeat == '+') ? 1 : 0;
+  const size_t max_count = (repeat == '?') ? 1 :
+      static_cast<size_t>(-1) - 1;
+  // We cannot call numeric_limits::max() as it conflicts with the
+  // max() macro on Windows.
+
+  for (size_t i = 0; i <= max_count; ++i) {
+    // We know that the atom matches each of the first i characters in str.
+    if (i >= min_count && MatchRegexAtHead(regex, str + i)) {
+      // We have enough matches at the head, and the tail matches too.
+      // Since we only care about *whether* the pattern matches str
+      // (as opposed to *how* it matches), there is no need to find a
+      // greedy match.
+      return true;
+    }
+    if (str[i] == '\0' || !AtomMatchesChar(escaped, c, str[i]))
+      return false;
+  }
+  return false;
+}
+
+// Returns true iff regex matches a prefix of str.  regex must be a
+// valid simple regular expression and not start with "^", or the
+// result is undefined.
+bool MatchRegexAtHead(const char* regex, const char* str) {
+  if (*regex == '\0')  // An empty regex matches a prefix of anything.
+    return true;
+
+  // "$" only matches the end of a string.  Note that regex being
+  // valid guarantees that there's nothing after "$" in it.
+  if (*regex == '$')
+    return *str == '\0';
+
+  // Is the first thing in regex an escape sequence?
+  const bool escaped = *regex == '\\';
+  if (escaped)
+    ++regex;
+  if (IsRepeat(regex[1])) {
+    // MatchRepetitionAndRegexAtHead() calls MatchRegexAtHead(), so
+    // here's an indirect recursion.  It terminates as the regex gets
+    // shorter in each recursion.
+    return MatchRepetitionAndRegexAtHead(
+        escaped, regex[0], regex[1], regex + 2, str);
+  } else {
+    // regex isn't empty, isn't "$", and doesn't start with a
+    // repetition.  We match the first atom of regex with the first
+    // character of str and recurse.
+    return (*str != '\0') && AtomMatchesChar(escaped, *regex, *str) &&
+        MatchRegexAtHead(regex + 1, str + 1);
+  }
+}
+
+// Returns true iff regex matches any substring of str.  regex must be
+// a valid simple regular expression, or the result is undefined.
+//
+// The algorithm is recursive, but the recursion depth doesn't exceed
+// the regex length, so we won't need to worry about running out of
+// stack space normally.  In rare cases the time complexity can be
+// exponential with respect to the regex length + the string length,
+// but usually it's must faster (often close to linear).
+bool MatchRegexAnywhere(const char* regex, const char* str) {
+  if (regex == NULL || str == NULL)
+    return false;
+
+  if (*regex == '^')
+    return MatchRegexAtHead(regex + 1, str);
+
+  // A successful match can be anywhere in str.
+  do {
+    if (MatchRegexAtHead(regex, str))
+      return true;
+  } while (*str++ != '\0');
+  return false;
+}
+
+// Implements the RE class.
+
+RE::~RE() {
+  free(const_cast<char*>(pattern_));
+  free(const_cast<char*>(full_pattern_));
+}
+
+// Returns true iff regular expression re matches the entire str.
+bool RE::FullMatch(const char* str, const RE& re) {
+  return re.is_valid_ && MatchRegexAnywhere(re.full_pattern_, str);
+}
+
+// Returns true iff regular expression re matches a substring of str
+// (including str itself).
+bool RE::PartialMatch(const char* str, const RE& re) {
+  return re.is_valid_ && MatchRegexAnywhere(re.pattern_, str);
+}
+
+// Initializes an RE from its string representation.
+void RE::Init(const char* regex) {
+  pattern_ = full_pattern_ = NULL;
+  if (regex != NULL) {
+    pattern_ = posix::StrDup(regex);
+  }
+
+  is_valid_ = ValidateRegex(regex);
+  if (!is_valid_) {
+    // No need to calculate the full pattern when the regex is invalid.
+    return;
+  }
+
+  const size_t len = strlen(regex);
+  // Reserves enough bytes to hold the regular expression used for a
+  // full match: we need space to prepend a '^', append a '$', and
+  // terminate the string with '\0'.
+  char* buffer = static_cast<char*>(malloc(len + 3));
+  full_pattern_ = buffer;
+
+  if (*regex != '^')
+    *buffer++ = '^';  // Makes sure full_pattern_ starts with '^'.
+
+  // We don't use snprintf or strncpy, as they trigger a warning when
+  // compiled with VC++ 8.0.
+  memcpy(buffer, regex, len);
+  buffer += len;
+
+  if (len == 0 || regex[len - 1] != '$')
+    *buffer++ = '$';  // Makes sure full_pattern_ ends with '$'.
+
+  *buffer = '\0';
+}
+
+#endif  // GTEST_USES_POSIX_RE
+
+const char kUnknownFile[] = "unknown file";
+
+// Formats a source file path and a line number as they would appear
+// in an error message from the compiler used to compile this code.
+GTEST_API_ ::std::string FormatFileLocation(const char* file, int line) {
+  const std::string file_name(file == NULL ? kUnknownFile : file);
+
+  if (line < 0) {
+    return file_name + ":";
+  }
+#ifdef _MSC_VER
+  return file_name + "(" + StreamableToString(line) + "):";
+#else
+  return file_name + ":" + StreamableToString(line) + ":";
+#endif  // _MSC_VER
+}
+
+// Formats a file location for compiler-independent XML output.
+// Although this function is not platform dependent, we put it next to
+// FormatFileLocation in order to contrast the two functions.
+// Note that FormatCompilerIndependentFileLocation() does NOT append colon
+// to the file location it produces, unlike FormatFileLocation().
+GTEST_API_ ::std::string FormatCompilerIndependentFileLocation(
+    const char* file, int line) {
+  const std::string file_name(file == NULL ? kUnknownFile : file);
+
+  if (line < 0)
+    return file_name;
+  else
+    return file_name + ":" + StreamableToString(line);
+}
+
+GTestLog::GTestLog(GTestLogSeverity severity, const char* file, int line)
+    : severity_(severity) {
+  const char* const marker =
+      severity == GTEST_INFO ?    "[  INFO ]" :
+      severity == GTEST_WARNING ? "[WARNING]" :
+      severity == GTEST_ERROR ?   "[ ERROR ]" : "[ FATAL ]";
+  GetStream() << ::std::endl << marker << " "
+              << FormatFileLocation(file, line).c_str() << ": ";
+}
+
+// Flushes the buffers and, if severity is GTEST_FATAL, aborts the program.
+GTestLog::~GTestLog() {
+  GetStream() << ::std::endl;
+  if (severity_ == GTEST_FATAL) {
+    fflush(stderr);
+    posix::Abort();
+  }
+}
+// Disable Microsoft deprecation warnings for POSIX functions called from
+// this class (creat, dup, dup2, and close)
+GTEST_DISABLE_MSC_WARNINGS_PUSH_(4996)
+
+#if GTEST_HAS_STREAM_REDIRECTION
+
+// Object that captures an output stream (stdout/stderr).
+class CapturedStream {
+ public:
+  // The ctor redirects the stream to a temporary file.
+  explicit CapturedStream(int fd) : fd_(fd), uncaptured_fd_(dup(fd)) {
+# if GTEST_OS_WINDOWS
+    char temp_dir_path[MAX_PATH + 1] = { '\0' };  // NOLINT
+    char temp_file_path[MAX_PATH + 1] = { '\0' };  // NOLINT
+
+    ::GetTempPathA(sizeof(temp_dir_path), temp_dir_path);
+    const UINT success = ::GetTempFileNameA(temp_dir_path,
+                                            "gtest_redir",
+                                            0,  // Generate unique file name.
+                                            temp_file_path);
+    GTEST_CHECK_(success != 0)
+        << "Unable to create a temporary file in " << temp_dir_path;
+    const int captured_fd = creat(temp_file_path, _S_IREAD | _S_IWRITE);
+    GTEST_CHECK_(captured_fd != -1) << "Unable to open temporary file "
+                                    << temp_file_path;
+    filename_ = temp_file_path;
+# else
+    // There's no guarantee that a test has write access to the current
+    // directory, so we create the temporary file in the /tmp directory
+    // instead. We use /tmp on most systems, and /sdcard on Android.
+    // That's because Android doesn't have /tmp.
+#  if GTEST_OS_LINUX_ANDROID
+    // Note: Android applications are expected to call the framework's
+    // Context.getExternalStorageDirectory() method through JNI to get
+    // the location of the world-writable SD Card directory. However,
+    // this requires a Context handle, which cannot be retrieved
+    // globally from native code. Doing so also precludes running the
+    // code as part of a regular standalone executable, which doesn't
+    // run in a Dalvik process (e.g. when running it through 'adb shell').
+    //
+    // The location /sdcard is directly accessible from native code
+    // and is the only location (unofficially) supported by the Android
+    // team. It's generally a symlink to the real SD Card mount point
+    // which can be /mnt/sdcard, /mnt/sdcard0, /system/media/sdcard, or
+    // other OEM-customized locations. Never rely on these, and always
+    // use /sdcard.
+    char name_template[] = "/sdcard/gtest_captured_stream.XXXXXX";
+#  else
+    char name_template[] = "/tmp/captured_stream.XXXXXX";
+#  endif  // GTEST_OS_LINUX_ANDROID
+    const int captured_fd = mkstemp(name_template);
+    filename_ = name_template;
+# endif  // GTEST_OS_WINDOWS
+    fflush(NULL);
+    dup2(captured_fd, fd_);
+    close(captured_fd);
+  }
+
+  ~CapturedStream() {
+    remove(filename_.c_str());
+  }
+
+  std::string GetCapturedString() {
+    if (uncaptured_fd_ != -1) {
+      // Restores the original stream.
+      fflush(NULL);
+      dup2(uncaptured_fd_, fd_);
+      close(uncaptured_fd_);
+      uncaptured_fd_ = -1;
+    }
+
+    FILE* const file = posix::FOpen(filename_.c_str(), "r");
+    const std::string content = ReadEntireFile(file);
+    posix::FClose(file);
+    return content;
+  }
+
+ private:
+  const int fd_;  // A stream to capture.
+  int uncaptured_fd_;
+  // Name of the temporary file holding the stderr output.
+  ::std::string filename_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(CapturedStream);
+};
+
+GTEST_DISABLE_MSC_WARNINGS_POP_()
+
+static CapturedStream* g_captured_stderr = NULL;
+static CapturedStream* g_captured_stdout = NULL;
+
+// Starts capturing an output stream (stdout/stderr).
+void CaptureStream(int fd, const char* stream_name, CapturedStream** stream) {
+  if (*stream != NULL) {
+    GTEST_LOG_(FATAL) << "Only one " << stream_name
+                      << " capturer can exist at a time.";
+  }
+  *stream = new CapturedStream(fd);
+}
+
+// Stops capturing the output stream and returns the captured string.
+std::string GetCapturedStream(CapturedStream** captured_stream) {
+  const std::string content = (*captured_stream)->GetCapturedString();
+
+  delete *captured_stream;
+  *captured_stream = NULL;
+
+  return content;
+}
+
+// Starts capturing stdout.
+void CaptureStdout() {
+  CaptureStream(kStdOutFileno, "stdout", &g_captured_stdout);
+}
+
+// Starts capturing stderr.
+void CaptureStderr() {
+  CaptureStream(kStdErrFileno, "stderr", &g_captured_stderr);
+}
+
+// Stops capturing stdout and returns the captured string.
+std::string GetCapturedStdout() {
+  return GetCapturedStream(&g_captured_stdout);
+}
+
+// Stops capturing stderr and returns the captured string.
+std::string GetCapturedStderr() {
+  return GetCapturedStream(&g_captured_stderr);
+}
+
+#endif  // GTEST_HAS_STREAM_REDIRECTION
+
+std::string TempDir() {
+#if GTEST_OS_WINDOWS_MOBILE
+  return "\\temp\\";
+#elif GTEST_OS_WINDOWS
+  const char* temp_dir = posix::GetEnv("TEMP");
+  if (temp_dir == NULL || temp_dir[0] == '\0')
+    return "\\temp\\";
+  else if (temp_dir[strlen(temp_dir) - 1] == '\\')
+    return temp_dir;
+  else
+    return std::string(temp_dir) + "\\";
+#elif GTEST_OS_LINUX_ANDROID
+  return "/sdcard/";
+#else
+  return "/tmp/";
+#endif  // GTEST_OS_WINDOWS_MOBILE
+}
+
+size_t GetFileSize(FILE* file) {
+  fseek(file, 0, SEEK_END);
+  return static_cast<size_t>(ftell(file));
+}
+
+std::string ReadEntireFile(FILE* file) {
+  const size_t file_size = GetFileSize(file);
+  char* const buffer = new char[file_size];
+
+  size_t bytes_last_read = 0;  // # of bytes read in the last fread()
+  size_t bytes_read = 0;       // # of bytes read so far
+
+  fseek(file, 0, SEEK_SET);
+
+  // Keeps reading the file until we cannot read further or the
+  // pre-determined file size is reached.
+  do {
+    bytes_last_read = fread(buffer+bytes_read, 1, file_size-bytes_read, file);
+    bytes_read += bytes_last_read;
+  } while (bytes_last_read > 0 && bytes_read < file_size);
+
+  const std::string content(buffer, bytes_read);
+  delete[] buffer;
+
+  return content;
+}
+
+#if GTEST_HAS_DEATH_TEST
+
+static const ::std::vector<testing::internal::string>* g_injected_test_argvs =
+                                        NULL;  // Owned.
+
+void SetInjectableArgvs(const ::std::vector<testing::internal::string>* argvs) {
+  if (g_injected_test_argvs != argvs)
+    delete g_injected_test_argvs;
+  g_injected_test_argvs = argvs;
+}
+
+const ::std::vector<testing::internal::string>& GetInjectableArgvs() {
+  if (g_injected_test_argvs != NULL) {
+    return *g_injected_test_argvs;
+  }
+  return GetArgvs();
+}
+#endif  // GTEST_HAS_DEATH_TEST
+
+#if GTEST_OS_WINDOWS_MOBILE
+namespace posix {
+void Abort() {
+  DebugBreak();
+  TerminateProcess(GetCurrentProcess(), 1);
+}
+}  // namespace posix
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+// Returns the name of the environment variable corresponding to the
+// given flag.  For example, FlagToEnvVar("foo") will return
+// "GTEST_FOO" in the open-source version.
+static std::string FlagToEnvVar(const char* flag) {
+  const std::string full_flag =
+      (Message() << GTEST_FLAG_PREFIX_ << flag).GetString();
+
+  Message env_var;
+  for (size_t i = 0; i != full_flag.length(); i++) {
+    env_var << ToUpper(full_flag.c_str()[i]);
+  }
+
+  return env_var.GetString();
+}
+
+// Parses 'str' for a 32-bit signed integer.  If successful, writes
+// the result to *value and returns true; otherwise leaves *value
+// unchanged and returns false.
+bool ParseInt32(const Message& src_text, const char* str, Int32* value) {
+  // Parses the environment variable as a decimal integer.
+  char* end = NULL;
+  const long long_value = strtol(str, &end, 10);  // NOLINT
+
+  // Has strtol() consumed all characters in the string?
+  if (*end != '\0') {
+    // No - an invalid character was encountered.
+    Message msg;
+    msg << "WARNING: " << src_text
+        << " is expected to be a 32-bit integer, but actually"
+        << " has value \"" << str << "\".\n";
+    printf("%s", msg.GetString().c_str());
+    fflush(stdout);
+    return false;
+  }
+
+  // Is the parsed value in the range of an Int32?
+  const Int32 result = static_cast<Int32>(long_value);
+  if (long_value == LONG_MAX || long_value == LONG_MIN ||
+      // The parsed value overflows as a long.  (strtol() returns
+      // LONG_MAX or LONG_MIN when the input overflows.)
+      result != long_value
+      // The parsed value overflows as an Int32.
+      ) {
+    Message msg;
+    msg << "WARNING: " << src_text
+        << " is expected to be a 32-bit integer, but actually"
+        << " has value " << str << ", which overflows.\n";
+    printf("%s", msg.GetString().c_str());
+    fflush(stdout);
+    return false;
+  }
+
+  *value = result;
+  return true;
+}
+
+// Reads and returns the Boolean environment variable corresponding to
+// the given flag; if it's not set, returns default_value.
+//
+// The value is considered true iff it's not "0".
+bool BoolFromGTestEnv(const char* flag, bool default_value) {
+#if defined(GTEST_GET_BOOL_FROM_ENV_)
+  return GTEST_GET_BOOL_FROM_ENV_(flag, default_value);
+#endif  // defined(GTEST_GET_BOOL_FROM_ENV_)
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const string_value = posix::GetEnv(env_var.c_str());
+  return string_value == NULL ?
+      default_value : strcmp(string_value, "0") != 0;
+}
+
+// Reads and returns a 32-bit integer stored in the environment
+// variable corresponding to the given flag; if it isn't set or
+// doesn't represent a valid 32-bit integer, returns default_value.
+Int32 Int32FromGTestEnv(const char* flag, Int32 default_value) {
+#if defined(GTEST_GET_INT32_FROM_ENV_)
+  return GTEST_GET_INT32_FROM_ENV_(flag, default_value);
+#endif  // defined(GTEST_GET_INT32_FROM_ENV_)
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const string_value = posix::GetEnv(env_var.c_str());
+  if (string_value == NULL) {
+    // The environment variable is not set.
+    return default_value;
+  }
+
+  Int32 result = default_value;
+  if (!ParseInt32(Message() << "Environment variable " << env_var,
+                  string_value, &result)) {
+    printf("The default value %s is used.\n",
+           (Message() << default_value).GetString().c_str());
+    fflush(stdout);
+    return default_value;
+  }
+
+  return result;
+}
+
+// Reads and returns the string environment variable corresponding to
+// the given flag; if it's not set, returns default_value.
+std::string StringFromGTestEnv(const char* flag, const char* default_value) {
+#if defined(GTEST_GET_STRING_FROM_ENV_)
+  return GTEST_GET_STRING_FROM_ENV_(flag, default_value);
+#endif  // defined(GTEST_GET_STRING_FROM_ENV_)
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* value = posix::GetEnv(env_var.c_str());
+  if (value != NULL) {
+    return value;
+  }
+
+  // As a special case for the 'output' flag, if GTEST_OUTPUT is not
+  // set, we look for XML_OUTPUT_FILE, which is set by the Bazel build
+  // system.  The value of XML_OUTPUT_FILE is a filename without the
+  // "xml:" prefix of GTEST_OUTPUT.
+  //
+  // The net priority order after flag processing is thus:
+  //   --gtest_output command line flag
+  //   GTEST_OUTPUT environment variable
+  //   XML_OUTPUT_FILE environment variable
+  //   'default_value'
+  if (strcmp(flag, "output") == 0) {
+    value = posix::GetEnv("XML_OUTPUT_FILE");
+    if (value != NULL) {
+      return std::string("xml:") + value;
+    }
+  }
+  return default_value;
+}
+
+}  // namespace internal
+}  // namespace testing
diff --git a/third_party/aom/third_party/googletest/src/googletest/src/gtest-printers.cc b/third_party/aom/third_party/googletest/src/googletest/src/gtest-printers.cc
new file mode 100644
index 0000000000..a2df412f8a
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/src/gtest-printers.cc
@@ -0,0 +1,373 @@
+// Copyright 2007, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+// Google Test - The Google C++ Testing Framework
+//
+// This file implements a universal value printer that can print a
+// value of any type T:
+//
+//   void ::testing::internal::UniversalPrinter<T>::Print(value, ostream_ptr);
+//
+// It uses the << operator when possible, and prints the bytes in the
+// object otherwise.  A user can override its behavior for a class
+// type Foo by defining either operator<<(::std::ostream&, const Foo&)
+// or void PrintTo(const Foo&, ::std::ostream*) in the namespace that
+// defines Foo.
+
+#include "gtest/gtest-printers.h"
+#include <ctype.h>
+#include <stdio.h>
+#include <cwchar>
+#include <ostream>  // NOLINT
+#include <string>
+#include "gtest/internal/gtest-port.h"
+
+namespace testing {
+
+namespace {
+
+using ::std::ostream;
+
+// Prints a segment of bytes in the given object.
+GTEST_ATTRIBUTE_NO_SANITIZE_MEMORY_
+GTEST_ATTRIBUTE_NO_SANITIZE_ADDRESS_
+GTEST_ATTRIBUTE_NO_SANITIZE_THREAD_
+void PrintByteSegmentInObjectTo(const unsigned char* obj_bytes, size_t start,
+                                size_t count, ostream* os) {
+  char text[5] = "";
+  for (size_t i = 0; i != count; i++) {
+    const size_t j = start + i;
+    if (i != 0) {
+      // Organizes the bytes into groups of 2 for easy parsing by
+      // human.
+      if ((j % 2) == 0)
+        *os << ' ';
+      else
+        *os << '-';
+    }
+    GTEST_SNPRINTF_(text, sizeof(text), "%02X", obj_bytes[j]);
+    *os << text;
+  }
+}
+
+// Prints the bytes in the given value to the given ostream.
+void PrintBytesInObjectToImpl(const unsigned char* obj_bytes, size_t count,
+                              ostream* os) {
+  // Tells the user how big the object is.
+  *os << count << "-byte object <";
+
+  const size_t kThreshold = 132;
+  const size_t kChunkSize = 64;
+  // If the object size is bigger than kThreshold, we'll have to omit
+  // some details by printing only the first and the last kChunkSize
+  // bytes.
+  // TODO(wan): let the user control the threshold using a flag.
+  if (count < kThreshold) {
+    PrintByteSegmentInObjectTo(obj_bytes, 0, count, os);
+  } else {
+    PrintByteSegmentInObjectTo(obj_bytes, 0, kChunkSize, os);
+    *os << " ... ";
+    // Rounds up to 2-byte boundary.
+    const size_t resume_pos = (count - kChunkSize + 1)/2*2;
+    PrintByteSegmentInObjectTo(obj_bytes, resume_pos, count - resume_pos, os);
+  }
+  *os << ">";
+}
+
+}  // namespace
+
+namespace internal2 {
+
+// Delegates to PrintBytesInObjectToImpl() to print the bytes in the
+// given object.  The delegation simplifies the implementation, which
+// uses the << operator and thus is easier done outside of the
+// ::testing::internal namespace, which contains a << operator that
+// sometimes conflicts with the one in STL.
+void PrintBytesInObjectTo(const unsigned char* obj_bytes, size_t count,
+                          ostream* os) {
+  PrintBytesInObjectToImpl(obj_bytes, count, os);
+}
+
+}  // namespace internal2
+
+namespace internal {
+
+// Depending on the value of a char (or wchar_t), we print it in one
+// of three formats:
+//   - as is if it's a printable ASCII (e.g. 'a', '2', ' '),
+//   - as a hexidecimal escape sequence (e.g. '\x7F'), or
+//   - as a special escape sequence (e.g. '\r', '\n').
+enum CharFormat {
+  kAsIs,
+  kHexEscape,
+  kSpecialEscape
+};
+
+// Returns true if c is a printable ASCII character.  We test the
+// value of c directly instead of calling isprint(), which is buggy on
+// Windows Mobile.
+inline bool IsPrintableAscii(wchar_t c) {
+  return 0x20 <= c && c <= 0x7E;
+}
+
+// Prints a wide or narrow char c as a character literal without the
+// quotes, escaping it when necessary; returns how c was formatted.
+// The template argument UnsignedChar is the unsigned version of Char,
+// which is the type of c.
+template <typename UnsignedChar, typename Char>
+static CharFormat PrintAsCharLiteralTo(Char c, ostream* os) {
+  switch (static_cast<wchar_t>(c)) {
+    case L'\0':
+      *os << "\\0";
+      break;
+    case L'\'':
+      *os << "\\'";
+      break;
+    case L'\\':
+      *os << "\\\\";
+      break;
+    case L'\a':
+      *os << "\\a";
+      break;
+    case L'\b':
+      *os << "\\b";
+      break;
+    case L'\f':
+      *os << "\\f";
+      break;
+    case L'\n':
+      *os << "\\n";
+      break;
+    case L'\r':
+      *os << "\\r";
+      break;
+    case L'\t':
+      *os << "\\t";
+      break;
+    case L'\v':
+      *os << "\\v";
+      break;
+    default:
+      if (IsPrintableAscii(c)) {
+        *os << static_cast<char>(c);
+        return kAsIs;
+      } else {
+        *os << "\\x" + String::FormatHexInt(static_cast<UnsignedChar>(c));
+        return kHexEscape;
+      }
+  }
+  return kSpecialEscape;
+}
+
+// Prints a wchar_t c as if it's part of a string literal, escaping it when
+// necessary; returns how c was formatted.
+static CharFormat PrintAsStringLiteralTo(wchar_t c, ostream* os) {
+  switch (c) {
+    case L'\'':
+      *os << "'";
+      return kAsIs;
+    case L'"':
+      *os << "\\\"";
+      return kSpecialEscape;
+    default:
+      return PrintAsCharLiteralTo<wchar_t>(c, os);
+  }
+}
+
+// Prints a char c as if it's part of a string literal, escaping it when
+// necessary; returns how c was formatted.
+static CharFormat PrintAsStringLiteralTo(char c, ostream* os) {
+  return PrintAsStringLiteralTo(
+      static_cast<wchar_t>(static_cast<unsigned char>(c)), os);
+}
+
+// Prints a wide or narrow character c and its code.  '\0' is printed
+// as "'\\0'", other unprintable characters are also properly escaped
+// using the standard C++ escape sequence.  The template argument
+// UnsignedChar is the unsigned version of Char, which is the type of c.
+template <typename UnsignedChar, typename Char>
+void PrintCharAndCodeTo(Char c, ostream* os) {
+  // First, print c as a literal in the most readable form we can find.
+  *os << ((sizeof(c) > 1) ? "L'" : "'");
+  const CharFormat format = PrintAsCharLiteralTo<UnsignedChar>(c, os);
+  *os << "'";
+
+  // To aid user debugging, we also print c's code in decimal, unless
+  // it's 0 (in which case c was printed as '\\0', making the code
+  // obvious).
+  if (c == 0)
+    return;
+  *os << " (" << static_cast<int>(c);
+
+  // For more convenience, we print c's code again in hexidecimal,
+  // unless c was already printed in the form '\x##' or the code is in
+  // [1, 9].
+  if (format == kHexEscape || (1 <= c && c <= 9)) {
+    // Do nothing.
+  } else {
+    *os << ", 0x" << String::FormatHexInt(static_cast<UnsignedChar>(c));
+  }
+  *os << ")";
+}
+
+void PrintTo(unsigned char c, ::std::ostream* os) {
+  PrintCharAndCodeTo<unsigned char>(c, os);
+}
+void PrintTo(signed char c, ::std::ostream* os) {
+  PrintCharAndCodeTo<unsigned char>(c, os);
+}
+
+// Prints a wchar_t as a symbol if it is printable or as its internal
+// code otherwise and also as its code.  L'\0' is printed as "L'\\0'".
+void PrintTo(wchar_t wc, ostream* os) {
+  PrintCharAndCodeTo<wchar_t>(wc, os);
+}
+
+// Prints the given array of characters to the ostream.  CharType must be either
+// char or wchar_t.
+// The array starts at begin, the length is len, it may include '\0' characters
+// and may not be NUL-terminated.
+template <typename CharType>
+GTEST_ATTRIBUTE_NO_SANITIZE_MEMORY_
+GTEST_ATTRIBUTE_NO_SANITIZE_ADDRESS_
+GTEST_ATTRIBUTE_NO_SANITIZE_THREAD_
+static void PrintCharsAsStringTo(
+    const CharType* begin, size_t len, ostream* os) {
+  const char* const kQuoteBegin = sizeof(CharType) == 1 ? "\"" : "L\"";
+  *os << kQuoteBegin;
+  bool is_previous_hex = false;
+  for (size_t index = 0; index < len; ++index) {
+    const CharType cur = begin[index];
+    if (is_previous_hex && IsXDigit(cur)) {
+      // Previous character is of '\x..' form and this character can be
+      // interpreted as another hexadecimal digit in its number. Break string to
+      // disambiguate.
+      *os << "\" " << kQuoteBegin;
+    }
+    is_previous_hex = PrintAsStringLiteralTo(cur, os) == kHexEscape;
+  }
+  *os << "\"";
+}
+
+// Prints a (const) char/wchar_t array of 'len' elements, starting at address
+// 'begin'.  CharType must be either char or wchar_t.
+template <typename CharType>
+GTEST_ATTRIBUTE_NO_SANITIZE_MEMORY_
+GTEST_ATTRIBUTE_NO_SANITIZE_ADDRESS_
+GTEST_ATTRIBUTE_NO_SANITIZE_THREAD_
+static void UniversalPrintCharArray(
+    const CharType* begin, size_t len, ostream* os) {
+  // The code
+  //   const char kFoo[] = "foo";
+  // generates an array of 4, not 3, elements, with the last one being '\0'.
+  //
+  // Therefore when printing a char array, we don't print the last element if
+  // it's '\0', such that the output matches the string literal as it's
+  // written in the source code.
+  if (len > 0 && begin[len - 1] == '\0') {
+    PrintCharsAsStringTo(begin, len - 1, os);
+    return;
+  }
+
+  // If, however, the last element in the array is not '\0', e.g.
+  //    const char kFoo[] = { 'f', 'o', 'o' };
+  // we must print the entire array.  We also print a message to indicate
+  // that the array is not NUL-terminated.
+  PrintCharsAsStringTo(begin, len, os);
+  *os << " (no terminating NUL)";
+}
+
+// Prints a (const) char array of 'len' elements, starting at address 'begin'.
+void UniversalPrintArray(const char* begin, size_t len, ostream* os) {
+  UniversalPrintCharArray(begin, len, os);
+}
+
+// Prints a (const) wchar_t array of 'len' elements, starting at address
+// 'begin'.
+void UniversalPrintArray(const wchar_t* begin, size_t len, ostream* os) {
+  UniversalPrintCharArray(begin, len, os);
+}
+
+// Prints the given C string to the ostream.
+void PrintTo(const char* s, ostream* os) {
+  if (s == NULL) {
+    *os << "NULL";
+  } else {
+    *os << ImplicitCast_<const void*>(s) << " pointing to ";
+    PrintCharsAsStringTo(s, strlen(s), os);
+  }
+}
+
+// MSVC compiler can be configured to define whar_t as a typedef
+// of unsigned short. Defining an overload for const wchar_t* in that case
+// would cause pointers to unsigned shorts be printed as wide strings,
+// possibly accessing more memory than intended and causing invalid
+// memory accesses. MSVC defines _NATIVE_WCHAR_T_DEFINED symbol when
+// wchar_t is implemented as a native type.
+#if !defined(_MSC_VER) || defined(_NATIVE_WCHAR_T_DEFINED)
+// Prints the given wide C string to the ostream.
+void PrintTo(const wchar_t* s, ostream* os) {
+  if (s == NULL) {
+    *os << "NULL";
+  } else {
+    *os << ImplicitCast_<const void*>(s) << " pointing to ";
+    PrintCharsAsStringTo(s, std::wcslen(s), os);
+  }
+}
+#endif  // wchar_t is native
+
+// Prints a ::string object.
+#if GTEST_HAS_GLOBAL_STRING
+void PrintStringTo(const ::string& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+#endif  // GTEST_HAS_GLOBAL_STRING
+
+void PrintStringTo(const ::std::string& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+
+// Prints a ::wstring object.
+#if GTEST_HAS_GLOBAL_WSTRING
+void PrintWideStringTo(const ::wstring& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+#if GTEST_HAS_STD_WSTRING
+void PrintWideStringTo(const ::std::wstring& s, ostream* os) {
+  PrintCharsAsStringTo(s.data(), s.size(), os);
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+}  // namespace internal
+
+}  // namespace testing
diff --git a/third_party/aom/third_party/googletest/src/googletest/src/gtest-test-part.cc b/third_party/aom/third_party/googletest/src/googletest/src/gtest-test-part.cc
new file mode 100644
index 0000000000..fb0e35425e
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/src/gtest-test-part.cc
@@ -0,0 +1,110 @@
+// Copyright 2008, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: mheule@google.com (Markus Heule)
+//
+// The Google C++ Testing Framework (Google Test)
+
+#include "gtest/gtest-test-part.h"
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick exists to
+// prevent the accidental inclusion of gtest-internal-inl.h in the
+// user's code.
+#define GTEST_IMPLEMENTATION_ 1
+#include "src/gtest-internal-inl.h"
+#undef GTEST_IMPLEMENTATION_
+
+namespace testing {
+
+using internal::GetUnitTestImpl;
+
+// Gets the summary of the failure message by omitting the stack trace
+// in it.
+std::string TestPartResult::ExtractSummary(const char* message) {
+  const char* const stack_trace = strstr(message, internal::kStackTraceMarker);
+  return stack_trace == NULL ? message :
+      std::string(message, stack_trace);
+}
+
+// Prints a TestPartResult object.
+std::ostream& operator<<(std::ostream& os, const TestPartResult& result) {
+  return os
+      << result.file_name() << ":" << result.line_number() << ": "
+      << (result.type() == TestPartResult::kSuccess ? "Success" :
+          result.type() == TestPartResult::kFatalFailure ? "Fatal failure" :
+          "Non-fatal failure") << ":\n"
+      << result.message() << std::endl;
+}
+
+// Appends a TestPartResult to the array.
+void TestPartResultArray::Append(const TestPartResult& result) {
+  array_.push_back(result);
+}
+
+// Returns the TestPartResult at the given index (0-based).
+const TestPartResult& TestPartResultArray::GetTestPartResult(int index) const {
+  if (index < 0 || index >= size()) {
+    printf("\nInvalid index (%d) into TestPartResultArray.\n", index);
+    internal::posix::Abort();
+  }
+
+  return array_[index];
+}
+
+// Returns the number of TestPartResult objects in the array.
+int TestPartResultArray::size() const {
+  return static_cast<int>(array_.size());
+}
+
+namespace internal {
+
+HasNewFatalFailureHelper::HasNewFatalFailureHelper()
+    : has_new_fatal_failure_(false),
+      original_reporter_(GetUnitTestImpl()->
+                         GetTestPartResultReporterForCurrentThread()) {
+  GetUnitTestImpl()->SetTestPartResultReporterForCurrentThread(this);
+}
+
+HasNewFatalFailureHelper::~HasNewFatalFailureHelper() {
+  GetUnitTestImpl()->SetTestPartResultReporterForCurrentThread(
+      original_reporter_);
+}
+
+void HasNewFatalFailureHelper::ReportTestPartResult(
+    const TestPartResult& result) {
+  if (result.fatally_failed())
+    has_new_fatal_failure_ = true;
+  original_reporter_->ReportTestPartResult(result);
+}
+
+}  // namespace internal
+
+}  // namespace testing
diff --git a/third_party/aom/third_party/googletest/src/googletest/src/gtest-typed-test.cc b/third_party/aom/third_party/googletest/src/googletest/src/gtest-typed-test.cc
new file mode 100644
index 0000000000..df1eef4754
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/src/gtest-typed-test.cc
@@ -0,0 +1,118 @@
+// Copyright 2008 Google Inc.
+// All Rights Reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+
+#include "gtest/gtest-typed-test.h"
+#include "gtest/gtest.h"
+
+namespace testing {
+namespace internal {
+
+#if GTEST_HAS_TYPED_TEST_P
+
+// Skips to the first non-space char in str. Returns an empty string if str
+// contains only whitespace characters.
+static const char* SkipSpaces(const char* str) {
+  while (IsSpace(*str))
+    str++;
+  return str;
+}
+
+static std::vector<std::string> SplitIntoTestNames(const char* src) {
+  std::vector<std::string> name_vec;
+  src = SkipSpaces(src);
+  for (; src != NULL; src = SkipComma(src)) {
+    name_vec.push_back(StripTrailingSpaces(GetPrefixUntilComma(src)));
+  }
+  return name_vec;
+}
+
+// Verifies that registered_tests match the test names in
+// registered_tests_; returns registered_tests if successful, or
+// aborts the program otherwise.
+const char* TypedTestCasePState::VerifyRegisteredTestNames(
+    const char* file, int line, const char* registered_tests) {
+  typedef RegisteredTestsMap::const_iterator RegisteredTestIter;
+  registered_ = true;
+
+  std::vector<std::string> name_vec = SplitIntoTestNames(registered_tests);
+
+  Message errors;
+
+  std::set<std::string> tests;
+  for (std::vector<std::string>::const_iterator name_it = name_vec.begin();
+       name_it != name_vec.end(); ++name_it) {
+    const std::string& name = *name_it;
+    if (tests.count(name) != 0) {
+      errors << "Test " << name << " is listed more than once.\n";
+      continue;
+    }
+
+    bool found = false;
+    for (RegisteredTestIter it = registered_tests_.begin();
+         it != registered_tests_.end();
+         ++it) {
+      if (name == it->first) {
+        found = true;
+        break;
+      }
+    }
+
+    if (found) {
+      tests.insert(name);
+    } else {
+      errors << "No test named " << name
+             << " can be found in this test case.\n";
+    }
+  }
+
+  for (RegisteredTestIter it = registered_tests_.begin();
+       it != registered_tests_.end();
+       ++it) {
+    if (tests.count(it->first) == 0) {
+      errors << "You forgot to list test " << it->first << ".\n";
+    }
+  }
+
+  const std::string& errors_str = errors.GetString();
+  if (errors_str != "") {
+    fprintf(stderr, "%s %s", FormatFileLocation(file, line).c_str(),
+            errors_str.c_str());
+    fflush(stderr);
+    posix::Abort();
+  }
+
+  return registered_tests;
+}
+
+#endif  // GTEST_HAS_TYPED_TEST_P
+
+}  // namespace internal
+}  // namespace testing
diff --git a/third_party/aom/third_party/googletest/src/googletest/src/gtest.cc b/third_party/aom/third_party/googletest/src/googletest/src/gtest.cc
new file mode 100644
index 0000000000..d882ab2e36
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/src/gtest.cc
@@ -0,0 +1,5388 @@
+// Copyright 2005, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+// Author: wan@google.com (Zhanyong Wan)
+//
+// The Google C++ Testing Framework (Google Test)
+
+#include "gtest/gtest.h"
+#include "gtest/internal/custom/gtest.h"
+#include "gtest/gtest-spi.h"
+
+#include <ctype.h>
+#include <math.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <time.h>
+#include <wchar.h>
+#include <wctype.h>
+
+#include <algorithm>
+#include <iomanip>
+#include <limits>
+#include <list>
+#include <map>
+#include <ostream>  // NOLINT
+#include <sstream>
+#include <vector>
+
+#if GTEST_OS_LINUX
+
+// TODO(kenton@google.com): Use autoconf to detect availability of
+// gettimeofday().
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+
+# include <fcntl.h>  // NOLINT
+# include <limits.h>  // NOLINT
+# include <sched.h>  // NOLINT
+// Declares vsnprintf().  This header is not available on Windows.
+# include <strings.h>  // NOLINT
+# include <sys/mman.h>  // NOLINT
+# include <sys/time.h>  // NOLINT
+# include <unistd.h>  // NOLINT
+# include <string>
+
+#elif GTEST_OS_SYMBIAN
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+# include <sys/time.h>  // NOLINT
+
+#elif GTEST_OS_ZOS
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+# include <sys/time.h>  // NOLINT
+
+// On z/OS we additionally need strings.h for strcasecmp.
+# include <strings.h>  // NOLINT
+
+#elif GTEST_OS_WINDOWS_MOBILE  // We are on Windows CE.
+
+# include <windows.h>  // NOLINT
+# undef min
+
+#elif GTEST_OS_WINDOWS  // We are on Windows proper.
+
+# include <io.h>  // NOLINT
+# include <sys/timeb.h>  // NOLINT
+# include <sys/types.h>  // NOLINT
+# include <sys/stat.h>  // NOLINT
+
+# if GTEST_OS_WINDOWS_MINGW
+// MinGW has gettimeofday() but not _ftime64().
+// TODO(kenton@google.com): Use autoconf to detect availability of
+//   gettimeofday().
+// TODO(kenton@google.com): There are other ways to get the time on
+//   Windows, like GetTickCount() or GetSystemTimeAsFileTime().  MinGW
+//   supports these.  consider using them instead.
+#  define GTEST_HAS_GETTIMEOFDAY_ 1
+#  include <sys/time.h>  // NOLINT
+# endif  // GTEST_OS_WINDOWS_MINGW
+
+// cpplint thinks that the header is already included, so we want to
+// silence it.
+# include <windows.h>  // NOLINT
+# undef min
+
+#else
+
+// Assume other platforms have gettimeofday().
+// TODO(kenton@google.com): Use autoconf to detect availability of
+//   gettimeofday().
+# define GTEST_HAS_GETTIMEOFDAY_ 1
+
+// cpplint thinks that the header is already included, so we want to
+// silence it.
+# include <sys/time.h>  // NOLINT
+# include <unistd.h>  // NOLINT
+
+#endif  // GTEST_OS_LINUX
+
+#if GTEST_HAS_EXCEPTIONS
+# include <stdexcept>
+#endif
+
+#if GTEST_CAN_STREAM_RESULTS_
+# include <arpa/inet.h>  // NOLINT
+# include <netdb.h>  // NOLINT
+# include <sys/socket.h>  // NOLINT
+# include <sys/types.h>  // NOLINT
+#endif
+
+// Indicates that this translation unit is part of Google Test's
+// implementation.  It must come before gtest-internal-inl.h is
+// included, or there will be a compiler error.  This trick is to
+// prevent a user from accidentally including gtest-internal-inl.h in
+// his code.
+#define GTEST_IMPLEMENTATION_ 1
+#include "src/gtest-internal-inl.h"
+#undef GTEST_IMPLEMENTATION_
+
+#if GTEST_OS_WINDOWS
+# define vsnprintf _vsnprintf
+#endif  // GTEST_OS_WINDOWS
+
+namespace testing {
+
+using internal::CountIf;
+using internal::ForEach;
+using internal::GetElementOr;
+using internal::Shuffle;
+
+// Constants.
+
+// A test whose test case name or test name matches this filter is
+// disabled and not run.
+static const char kDisableTestFilter[] = "DISABLED_*:*/DISABLED_*";
+
+// A test case whose name matches this filter is considered a death
+// test case and will be run before test cases whose name doesn't
+// match this filter.
+static const char kDeathTestCaseFilter[] = "*DeathTest:*DeathTest/*";
+
+// A test filter that matches everything.
+static const char kUniversalFilter[] = "*";
+
+// The default output file for XML output.
+static const char kDefaultOutputFile[] = "test_detail.xml";
+
+// The environment variable name for the test shard index.
+static const char kTestShardIndex[] = "GTEST_SHARD_INDEX";
+// The environment variable name for the total number of test shards.
+static const char kTestTotalShards[] = "GTEST_TOTAL_SHARDS";
+// The environment variable name for the test shard status file.
+static const char kTestShardStatusFile[] = "GTEST_SHARD_STATUS_FILE";
+
+namespace internal {
+
+// The text used in failure messages to indicate the start of the
+// stack trace.
+const char kStackTraceMarker[] = "\nStack trace:\n";
+
+// g_help_flag is true iff the --help flag or an equivalent form is
+// specified on the command line.
+bool g_help_flag = false;
+
+}  // namespace internal
+
+static const char* GetDefaultFilter() {
+#ifdef GTEST_TEST_FILTER_ENV_VAR_
+  const char* const testbridge_test_only = getenv(GTEST_TEST_FILTER_ENV_VAR_);
+  if (testbridge_test_only != NULL) {
+    return testbridge_test_only;
+  }
+#endif  // GTEST_TEST_FILTER_ENV_VAR_
+  return kUniversalFilter;
+}
+
+GTEST_DEFINE_bool_(
+    also_run_disabled_tests,
+    internal::BoolFromGTestEnv("also_run_disabled_tests", false),
+    "Run disabled tests too, in addition to the tests normally being run.");
+
+GTEST_DEFINE_bool_(
+    break_on_failure,
+    internal::BoolFromGTestEnv("break_on_failure", false),
+    "True iff a failed assertion should be a debugger break-point.");
+
+GTEST_DEFINE_bool_(
+    catch_exceptions,
+    internal::BoolFromGTestEnv("catch_exceptions", true),
+    "True iff " GTEST_NAME_
+    " should catch exceptions and treat them as test failures.");
+
+GTEST_DEFINE_string_(
+    color,
+    internal::StringFromGTestEnv("color", "auto"),
+    "Whether to use colors in the output.  Valid values: yes, no, "
+    "and auto.  'auto' means to use colors if the output is "
+    "being sent to a terminal and the TERM environment variable "
+    "is set to a terminal type that supports colors.");
+
+GTEST_DEFINE_string_(
+    filter,
+    internal::StringFromGTestEnv("filter", GetDefaultFilter()),
+    "A colon-separated list of glob (not regex) patterns "
+    "for filtering the tests to run, optionally followed by a "
+    "'-' and a : separated list of negative patterns (tests to "
+    "exclude).  A test is run if it matches one of the positive "
+    "patterns and does not match any of the negative patterns.");
+
+GTEST_DEFINE_bool_(list_tests, false,
+                   "List all tests without running them.");
+
+GTEST_DEFINE_string_(
+    output,
+    internal::StringFromGTestEnv("output", ""),
+    "A format (currently must be \"xml\"), optionally followed "
+    "by a colon and an output file name or directory. A directory "
+    "is indicated by a trailing pathname separator. "
+    "Examples: \"xml:filename.xml\", \"xml::directoryname/\". "
+    "If a directory is specified, output files will be created "
+    "within that directory, with file-names based on the test "
+    "executable's name and, if necessary, made unique by adding "
+    "digits.");
+
+GTEST_DEFINE_bool_(
+    print_time,
+    internal::BoolFromGTestEnv("print_time", true),
+    "True iff " GTEST_NAME_
+    " should display elapsed time in text output.");
+
+GTEST_DEFINE_int32_(
+    random_seed,
+    internal::Int32FromGTestEnv("random_seed", 0),
+    "Random number seed to use when shuffling test orders.  Must be in range "
+    "[1, 99999], or 0 to use a seed based on the current time.");
+
+GTEST_DEFINE_int32_(
+    repeat,
+    internal::Int32FromGTestEnv("repeat", 1),
+    "How many times to repeat each test.  Specify a negative number "
+    "for repeating forever.  Useful for shaking out flaky tests.");
+
+GTEST_DEFINE_bool_(
+    show_internal_stack_frames, false,
+    "True iff " GTEST_NAME_ " should include internal stack frames when "
+    "printing test failure stack traces.");
+
+GTEST_DEFINE_bool_(
+    shuffle,
+    internal::BoolFromGTestEnv("shuffle", false),
+    "True iff " GTEST_NAME_
+    " should randomize tests' order on every run.");
+
+GTEST_DEFINE_int32_(
+    stack_trace_depth,
+    internal::Int32FromGTestEnv("stack_trace_depth", kMaxStackTraceDepth),
+    "The maximum number of stack frames to print when an "
+    "assertion fails.  The valid range is 0 through 100, inclusive.");
+
+GTEST_DEFINE_string_(
+    stream_result_to,
+    internal::StringFromGTestEnv("stream_result_to", ""),
+    "This flag specifies the host name and the port number on which to stream "
+    "test results. Example: \"localhost:555\". The flag is effective only on "
+    "Linux.");
+
+GTEST_DEFINE_bool_(
+    throw_on_failure,
+    internal::BoolFromGTestEnv("throw_on_failure", false),
+    "When this flag is specified, a failed assertion will throw an exception "
+    "if exceptions are enabled or exit the program with a non-zero code "
+    "otherwise.");
+
+#if GTEST_USE_OWN_FLAGFILE_FLAG_
+GTEST_DEFINE_string_(
+    flagfile,
+    internal::StringFromGTestEnv("flagfile", ""),
+    "This flag specifies the flagfile to read command-line flags from.");
+#endif  // GTEST_USE_OWN_FLAGFILE_FLAG_
+
+namespace internal {
+
+// Generates a random number from [0, range), using a Linear
+// Congruential Generator (LCG).  Crashes if 'range' is 0 or greater
+// than kMaxRange.
+UInt32 Random::Generate(UInt32 range) {
+  // These constants are the same as are used in glibc's rand(3).
+  state_ = (1103515245U*state_ + 12345U) % kMaxRange;
+
+  GTEST_CHECK_(range > 0)
+      << "Cannot generate a number in the range [0, 0).";
+  GTEST_CHECK_(range <= kMaxRange)
+      << "Generation of a number in [0, " << range << ") was requested, "
+      << "but this can only generate numbers in [0, " << kMaxRange << ").";
+
+  // Converting via modulus introduces a bit of downward bias, but
+  // it's simple, and a linear congruential generator isn't too good
+  // to begin with.
+  return state_ % range;
+}
+
+// GTestIsInitialized() returns true iff the user has initialized
+// Google Test.  Useful for catching the user mistake of not initializing
+// Google Test before calling RUN_ALL_TESTS().
+static bool GTestIsInitialized() { return GetArgvs().size() > 0; }
+
+// Iterates over a vector of TestCases, keeping a running sum of the
+// results of calling a given int-returning method on each.
+// Returns the sum.
+static int SumOverTestCaseList(const std::vector<TestCase*>& case_list,
+                               int (TestCase::*method)() const) {
+  int sum = 0;
+  for (size_t i = 0; i < case_list.size(); i++) {
+    sum += (case_list[i]->*method)();
+  }
+  return sum;
+}
+
+// Returns true iff the test case passed.
+static bool TestCasePassed(const TestCase* test_case) {
+  return test_case->should_run() && test_case->Passed();
+}
+
+// Returns true iff the test case failed.
+static bool TestCaseFailed(const TestCase* test_case) {
+  return test_case->should_run() && test_case->Failed();
+}
+
+// Returns true iff test_case contains at least one test that should
+// run.
+static bool ShouldRunTestCase(const TestCase* test_case) {
+  return test_case->should_run();
+}
+
+// AssertHelper constructor.
+AssertHelper::AssertHelper(TestPartResult::Type type,
+                           const char* file,
+                           int line,
+                           const char* message)
+    : data_(new AssertHelperData(type, file, line, message)) {
+}
+
+AssertHelper::~AssertHelper() {
+  delete data_;
+}
+
+// Message assignment, for assertion streaming support.
+void AssertHelper::operator=(const Message& message) const {
+  UnitTest::GetInstance()->
+    AddTestPartResult(data_->type, data_->file, data_->line,
+                      AppendUserMessage(data_->message, message),
+                      UnitTest::GetInstance()->impl()
+                      ->CurrentOsStackTraceExceptTop(1)
+                      // Skips the stack frame for this function itself.
+                      );  // NOLINT
+}
+
+// Mutex for linked pointers.
+GTEST_API_ GTEST_DEFINE_STATIC_MUTEX_(g_linked_ptr_mutex);
+
+// A copy of all command line arguments.  Set by InitGoogleTest().
+::std::vector<testing::internal::string> g_argvs;
+
+const ::std::vector<testing::internal::string>& GetArgvs() {
+#if defined(GTEST_CUSTOM_GET_ARGVS_)
+  return GTEST_CUSTOM_GET_ARGVS_();
+#else  // defined(GTEST_CUSTOM_GET_ARGVS_)
+  return g_argvs;
+#endif  // defined(GTEST_CUSTOM_GET_ARGVS_)
+}
+
+// Returns the current application's name, removing directory path if that
+// is present.
+FilePath GetCurrentExecutableName() {
+  FilePath result;
+
+#if GTEST_OS_WINDOWS
+  result.Set(FilePath(GetArgvs()[0]).RemoveExtension("exe"));
+#else
+  result.Set(FilePath(GetArgvs()[0]));
+#endif  // GTEST_OS_WINDOWS
+
+  return result.RemoveDirectoryName();
+}
+
+// Functions for processing the gtest_output flag.
+
+// Returns the output format, or "" for normal printed output.
+std::string UnitTestOptions::GetOutputFormat() {
+  const char* const gtest_output_flag = GTEST_FLAG(output).c_str();
+  if (gtest_output_flag == NULL) return std::string("");
+
+  const char* const colon = strchr(gtest_output_flag, ':');
+  return (colon == NULL) ?
+      std::string(gtest_output_flag) :
+      std::string(gtest_output_flag, colon - gtest_output_flag);
+}
+
+// Returns the name of the requested output file, or the default if none
+// was explicitly specified.
+std::string UnitTestOptions::GetAbsolutePathToOutputFile() {
+  const char* const gtest_output_flag = GTEST_FLAG(output).c_str();
+  if (gtest_output_flag == NULL)
+    return "";
+
+  const char* const colon = strchr(gtest_output_flag, ':');
+  if (colon == NULL)
+    return internal::FilePath::ConcatPaths(
+        internal::FilePath(
+            UnitTest::GetInstance()->original_working_dir()),
+        internal::FilePath(kDefaultOutputFile)).string();
+
+  internal::FilePath output_name(colon + 1);
+  if (!output_name.IsAbsolutePath())
+    // TODO(wan@google.com): on Windows \some\path is not an absolute
+    // path (as its meaning depends on the current drive), yet the
+    // following logic for turning it into an absolute path is wrong.
+    // Fix it.
+    output_name = internal::FilePath::ConcatPaths(
+        internal::FilePath(UnitTest::GetInstance()->original_working_dir()),
+        internal::FilePath(colon + 1));
+
+  if (!output_name.IsDirectory())
+    return output_name.string();
+
+  internal::FilePath result(internal::FilePath::GenerateUniqueFileName(
+      output_name, internal::GetCurrentExecutableName(),
+      GetOutputFormat().c_str()));
+  return result.string();
+}
+
+// Returns true iff the wildcard pattern matches the string.  The
+// first ':' or '\0' character in pattern marks the end of it.
+//
+// This recursive algorithm isn't very efficient, but is clear and
+// works well enough for matching test names, which are short.
+bool UnitTestOptions::PatternMatchesString(const char *pattern,
+                                           const char *str) {
+  switch (*pattern) {
+    case '\0':
+    case ':':  // Either ':' or '\0' marks the end of the pattern.
+      return *str == '\0';
+    case '?':  // Matches any single character.
+      return *str != '\0' && PatternMatchesString(pattern + 1, str + 1);
+    case '*':  // Matches any string (possibly empty) of characters.
+      return (*str != '\0' && PatternMatchesString(pattern, str + 1)) ||
+          PatternMatchesString(pattern + 1, str);
+    default:  // Non-special character.  Matches itself.
+      return *pattern == *str &&
+          PatternMatchesString(pattern + 1, str + 1);
+  }
+}
+
+bool UnitTestOptions::MatchesFilter(
+    const std::string& name, const char* filter) {
+  const char *cur_pattern = filter;
+  for (;;) {
+    if (PatternMatchesString(cur_pattern, name.c_str())) {
+      return true;
+    }
+
+    // Finds the next pattern in the filter.
+    cur_pattern = strchr(cur_pattern, ':');
+
+    // Returns if no more pattern can be found.
+    if (cur_pattern == NULL) {
+      return false;
+    }
+
+    // Skips the pattern separater (the ':' character).
+    cur_pattern++;
+  }
+}
+
+// Returns true iff the user-specified filter matches the test case
+// name and the test name.
+bool UnitTestOptions::FilterMatchesTest(const std::string &test_case_name,
+                                        const std::string &test_name) {
+  const std::string& full_name = test_case_name + "." + test_name.c_str();
+
+  // Split --gtest_filter at '-', if there is one, to separate into
+  // positive filter and negative filter portions
+  const char* const p = GTEST_FLAG(filter).c_str();
+  const char* const dash = strchr(p, '-');
+  std::string positive;
+  std::string negative;
+  if (dash == NULL) {
+    positive = GTEST_FLAG(filter).c_str();  // Whole string is a positive filter
+    negative = "";
+  } else {
+    positive = std::string(p, dash);   // Everything up to the dash
+    negative = std::string(dash + 1);  // Everything after the dash
+    if (positive.empty()) {
+      // Treat '-test1' as the same as '*-test1'
+      positive = kUniversalFilter;
+    }
+  }
+
+  // A filter is a colon-separated list of patterns.  It matches a
+  // test if any pattern in it matches the test.
+  return (MatchesFilter(full_name, positive.c_str()) &&
+          !MatchesFilter(full_name, negative.c_str()));
+}
+
+#if GTEST_HAS_SEH
+// Returns EXCEPTION_EXECUTE_HANDLER if Google Test should handle the
+// given SEH exception, or EXCEPTION_CONTINUE_SEARCH otherwise.
+// This function is useful as an __except condition.
+int UnitTestOptions::GTestShouldProcessSEH(DWORD exception_code) {
+  // Google Test should handle a SEH exception if:
+  //   1. the user wants it to, AND
+  //   2. this is not a breakpoint exception, AND
+  //   3. this is not a C++ exception (VC++ implements them via SEH,
+  //      apparently).
+  //
+  // SEH exception code for C++ exceptions.
+  // (see http://support.microsoft.com/kb/185294 for more information).
+  const DWORD kCxxExceptionCode = 0xe06d7363;
+
+  bool should_handle = true;
+
+  if (!GTEST_FLAG(catch_exceptions))
+    should_handle = false;
+  else if (exception_code == EXCEPTION_BREAKPOINT)
+    should_handle = false;
+  else if (exception_code == kCxxExceptionCode)
+    should_handle = false;
+
+  return should_handle ? EXCEPTION_EXECUTE_HANDLER : EXCEPTION_CONTINUE_SEARCH;
+}
+#endif  // GTEST_HAS_SEH
+
+}  // namespace internal
+
+// The c'tor sets this object as the test part result reporter used by
+// Google Test.  The 'result' parameter specifies where to report the
+// results. Intercepts only failures from the current thread.
+ScopedFakeTestPartResultReporter::ScopedFakeTestPartResultReporter(
+    TestPartResultArray* result)
+    : intercept_mode_(INTERCEPT_ONLY_CURRENT_THREAD),
+      result_(result) {
+  Init();
+}
+
+// The c'tor sets this object as the test part result reporter used by
+// Google Test.  The 'result' parameter specifies where to report the
+// results.
+ScopedFakeTestPartResultReporter::ScopedFakeTestPartResultReporter(
+    InterceptMode intercept_mode, TestPartResultArray* result)
+    : intercept_mode_(intercept_mode),
+      result_(result) {
+  Init();
+}
+
+void ScopedFakeTestPartResultReporter::Init() {
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  if (intercept_mode_ == INTERCEPT_ALL_THREADS) {
+    old_reporter_ = impl->GetGlobalTestPartResultReporter();
+    impl->SetGlobalTestPartResultReporter(this);
+  } else {
+    old_reporter_ = impl->GetTestPartResultReporterForCurrentThread();
+    impl->SetTestPartResultReporterForCurrentThread(this);
+  }
+}
+
+// The d'tor restores the test part result reporter used by Google Test
+// before.
+ScopedFakeTestPartResultReporter::~ScopedFakeTestPartResultReporter() {
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  if (intercept_mode_ == INTERCEPT_ALL_THREADS) {
+    impl->SetGlobalTestPartResultReporter(old_reporter_);
+  } else {
+    impl->SetTestPartResultReporterForCurrentThread(old_reporter_);
+  }
+}
+
+// Increments the test part result count and remembers the result.
+// This method is from the TestPartResultReporterInterface interface.
+void ScopedFakeTestPartResultReporter::ReportTestPartResult(
+    const TestPartResult& result) {
+  result_->Append(result);
+}
+
+namespace internal {
+
+// Returns the type ID of ::testing::Test.  We should always call this
+// instead of GetTypeId< ::testing::Test>() to get the type ID of
+// testing::Test.  This is to work around a suspected linker bug when
+// using Google Test as a framework on Mac OS X.  The bug causes
+// GetTypeId< ::testing::Test>() to return different values depending
+// on whether the call is from the Google Test framework itself or
+// from user test code.  GetTestTypeId() is guaranteed to always
+// return the same value, as it always calls GetTypeId<>() from the
+// gtest.cc, which is within the Google Test framework.
+TypeId GetTestTypeId() {
+  return GetTypeId<Test>();
+}
+
+// The value of GetTestTypeId() as seen from within the Google Test
+// library.  This is solely for testing GetTestTypeId().
+extern const TypeId kTestTypeIdInGoogleTest = GetTestTypeId();
+
+// This predicate-formatter checks that 'results' contains a test part
+// failure of the given type and that the failure message contains the
+// given substring.
+AssertionResult HasOneFailure(const char* /* results_expr */,
+                              const char* /* type_expr */,
+                              const char* /* substr_expr */,
+                              const TestPartResultArray& results,
+                              TestPartResult::Type type,
+                              const string& substr) {
+  const std::string expected(type == TestPartResult::kFatalFailure ?
+                        "1 fatal failure" :
+                        "1 non-fatal failure");
+  Message msg;
+  if (results.size() != 1) {
+    msg << "Expected: " << expected << "\n"
+        << "  Actual: " << results.size() << " failures";
+    for (int i = 0; i < results.size(); i++) {
+      msg << "\n" << results.GetTestPartResult(i);
+    }
+    return AssertionFailure() << msg;
+  }
+
+  const TestPartResult& r = results.GetTestPartResult(0);
+  if (r.type() != type) {
+    return AssertionFailure() << "Expected: " << expected << "\n"
+                              << "  Actual:\n"
+                              << r;
+  }
+
+  if (strstr(r.message(), substr.c_str()) == NULL) {
+    return AssertionFailure() << "Expected: " << expected << " containing \""
+                              << substr << "\"\n"
+                              << "  Actual:\n"
+                              << r;
+  }
+
+  return AssertionSuccess();
+}
+
+// The constructor of SingleFailureChecker remembers where to look up
+// test part results, what type of failure we expect, and what
+// substring the failure message should contain.
+SingleFailureChecker:: SingleFailureChecker(
+    const TestPartResultArray* results,
+    TestPartResult::Type type,
+    const string& substr)
+    : results_(results),
+      type_(type),
+      substr_(substr) {}
+
+// The destructor of SingleFailureChecker verifies that the given
+// TestPartResultArray contains exactly one failure that has the given
+// type and contains the given substring.  If that's not the case, a
+// non-fatal failure will be generated.
+SingleFailureChecker::~SingleFailureChecker() {
+  EXPECT_PRED_FORMAT3(HasOneFailure, *results_, type_, substr_);
+}
+
+DefaultGlobalTestPartResultReporter::DefaultGlobalTestPartResultReporter(
+    UnitTestImpl* unit_test) : unit_test_(unit_test) {}
+
+void DefaultGlobalTestPartResultReporter::ReportTestPartResult(
+    const TestPartResult& result) {
+  unit_test_->current_test_result()->AddTestPartResult(result);
+  unit_test_->listeners()->repeater()->OnTestPartResult(result);
+}
+
+DefaultPerThreadTestPartResultReporter::DefaultPerThreadTestPartResultReporter(
+    UnitTestImpl* unit_test) : unit_test_(unit_test) {}
+
+void DefaultPerThreadTestPartResultReporter::ReportTestPartResult(
+    const TestPartResult& result) {
+  unit_test_->GetGlobalTestPartResultReporter()->ReportTestPartResult(result);
+}
+
+// Returns the global test part result reporter.
+TestPartResultReporterInterface*
+UnitTestImpl::GetGlobalTestPartResultReporter() {
+  internal::MutexLock lock(&global_test_part_result_reporter_mutex_);
+  return global_test_part_result_repoter_;
+}
+
+// Sets the global test part result reporter.
+void UnitTestImpl::SetGlobalTestPartResultReporter(
+    TestPartResultReporterInterface* reporter) {
+  internal::MutexLock lock(&global_test_part_result_reporter_mutex_);
+  global_test_part_result_repoter_ = reporter;
+}
+
+// Returns the test part result reporter for the current thread.
+TestPartResultReporterInterface*
+UnitTestImpl::GetTestPartResultReporterForCurrentThread() {
+  return per_thread_test_part_result_reporter_.get();
+}
+
+// Sets the test part result reporter for the current thread.
+void UnitTestImpl::SetTestPartResultReporterForCurrentThread(
+    TestPartResultReporterInterface* reporter) {
+  per_thread_test_part_result_reporter_.set(reporter);
+}
+
+// Gets the number of successful test cases.
+int UnitTestImpl::successful_test_case_count() const {
+  return CountIf(test_cases_, TestCasePassed);
+}
+
+// Gets the number of failed test cases.
+int UnitTestImpl::failed_test_case_count() const {
+  return CountIf(test_cases_, TestCaseFailed);
+}
+
+// Gets the number of all test cases.
+int UnitTestImpl::total_test_case_count() const {
+  return static_cast<int>(test_cases_.size());
+}
+
+// Gets the number of all test cases that contain at least one test
+// that should run.
+int UnitTestImpl::test_case_to_run_count() const {
+  return CountIf(test_cases_, ShouldRunTestCase);
+}
+
+// Gets the number of successful tests.
+int UnitTestImpl::successful_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::successful_test_count);
+}
+
+// Gets the number of failed tests.
+int UnitTestImpl::failed_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::failed_test_count);
+}
+
+// Gets the number of disabled tests that will be reported in the XML report.
+int UnitTestImpl::reportable_disabled_test_count() const {
+  return SumOverTestCaseList(test_cases_,
+                             &TestCase::reportable_disabled_test_count);
+}
+
+// Gets the number of disabled tests.
+int UnitTestImpl::disabled_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::disabled_test_count);
+}
+
+// Gets the number of tests to be printed in the XML report.
+int UnitTestImpl::reportable_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::reportable_test_count);
+}
+
+// Gets the number of all tests.
+int UnitTestImpl::total_test_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::total_test_count);
+}
+
+// Gets the number of tests that should run.
+int UnitTestImpl::test_to_run_count() const {
+  return SumOverTestCaseList(test_cases_, &TestCase::test_to_run_count);
+}
+
+// Returns the current OS stack trace as an std::string.
+//
+// The maximum number of stack frames to be included is specified by
+// the gtest_stack_trace_depth flag.  The skip_count parameter
+// specifies the number of top frames to be skipped, which doesn't
+// count against the number of frames to be included.
+//
+// For example, if Foo() calls Bar(), which in turn calls
+// CurrentOsStackTraceExceptTop(1), Foo() will be included in the
+// trace but Bar() and CurrentOsStackTraceExceptTop() won't.
+std::string UnitTestImpl::CurrentOsStackTraceExceptTop(int skip_count) {
+  return os_stack_trace_getter()->CurrentStackTrace(
+      static_cast<int>(GTEST_FLAG(stack_trace_depth)),
+      skip_count + 1
+      // Skips the user-specified number of frames plus this function
+      // itself.
+      );  // NOLINT
+}
+
+// Returns the current time in milliseconds.
+TimeInMillis GetTimeInMillis() {
+#if GTEST_OS_WINDOWS_MOBILE || defined(__BORLANDC__)
+  // Difference between 1970-01-01 and 1601-01-01 in milliseconds.
+  // http://analogous.blogspot.com/2005/04/epoch.html
+  const TimeInMillis kJavaEpochToWinFileTimeDelta =
+    static_cast<TimeInMillis>(116444736UL) * 100000UL;
+  const DWORD kTenthMicrosInMilliSecond = 10000;
+
+  SYSTEMTIME now_systime;
+  FILETIME now_filetime;
+  ULARGE_INTEGER now_int64;
+  // TODO(kenton@google.com): Shouldn't this just use
+  //   GetSystemTimeAsFileTime()?
+  GetSystemTime(&now_systime);
+  if (SystemTimeToFileTime(&now_systime, &now_filetime)) {
+    now_int64.LowPart = now_filetime.dwLowDateTime;
+    now_int64.HighPart = now_filetime.dwHighDateTime;
+    now_int64.QuadPart = (now_int64.QuadPart / kTenthMicrosInMilliSecond) -
+      kJavaEpochToWinFileTimeDelta;
+    return now_int64.QuadPart;
+  }
+  return 0;
+#elif GTEST_OS_WINDOWS && !GTEST_HAS_GETTIMEOFDAY_
+  __timeb64 now;
+
+  // MSVC 8 deprecates _ftime64(), so we want to suppress warning 4996
+  // (deprecated function) there.
+  // TODO(kenton@google.com): Use GetTickCount()?  Or use
+  //   SystemTimeToFileTime()
+  GTEST_DISABLE_MSC_WARNINGS_PUSH_(4996)
+  _ftime64(&now);
+  GTEST_DISABLE_MSC_WARNINGS_POP_()
+
+  return static_cast<TimeInMillis>(now.time) * 1000 + now.millitm;
+#elif GTEST_HAS_GETTIMEOFDAY_
+  struct timeval now;
+  gettimeofday(&now, NULL);
+  return static_cast<TimeInMillis>(now.tv_sec) * 1000 + now.tv_usec / 1000;
+#else
+# error "Don't know how to get the current time on your system."
+#endif
+}
+
+// Utilities
+
+// class String.
+
+#if GTEST_OS_WINDOWS_MOBILE
+// Creates a UTF-16 wide string from the given ANSI string, allocating
+// memory using new. The caller is responsible for deleting the return
+// value using delete[]. Returns the wide string, or NULL if the
+// input is NULL.
+LPCWSTR String::AnsiToUtf16(const char* ansi) {
+  if (!ansi) return NULL;
+  const int length = strlen(ansi);
+  const int unicode_length =
+      MultiByteToWideChar(CP_ACP, 0, ansi, length,
+                          NULL, 0);
+  WCHAR* unicode = new WCHAR[unicode_length + 1];
+  MultiByteToWideChar(CP_ACP, 0, ansi, length,
+                      unicode, unicode_length);
+  unicode[unicode_length] = 0;
+  return unicode;
+}
+
+// Creates an ANSI string from the given wide string, allocating
+// memory using new. The caller is responsible for deleting the return
+// value using delete[]. Returns the ANSI string, or NULL if the
+// input is NULL.
+const char* String::Utf16ToAnsi(LPCWSTR utf16_str)  {
+  if (!utf16_str) return NULL;
+  const int ansi_length =
+      WideCharToMultiByte(CP_ACP, 0, utf16_str, -1,
+                          NULL, 0, NULL, NULL);
+  char* ansi = new char[ansi_length + 1];
+  WideCharToMultiByte(CP_ACP, 0, utf16_str, -1,
+                      ansi, ansi_length, NULL, NULL);
+  ansi[ansi_length] = 0;
+  return ansi;
+}
+
+#endif  // GTEST_OS_WINDOWS_MOBILE
+
+// Compares two C strings.  Returns true iff they have the same content.
+//
+// Unlike strcmp(), this function can handle NULL argument(s).  A NULL
+// C string is considered different to any non-NULL C string,
+// including the empty string.
+bool String::CStringEquals(const char * lhs, const char * rhs) {
+  if ( lhs == NULL ) return rhs == NULL;
+
+  if ( rhs == NULL ) return false;
+
+  return strcmp(lhs, rhs) == 0;
+}
+
+#if GTEST_HAS_STD_WSTRING || GTEST_HAS_GLOBAL_WSTRING
+
+// Converts an array of wide chars to a narrow string using the UTF-8
+// encoding, and streams the result to the given Message object.
+static void StreamWideCharsToMessage(const wchar_t* wstr, size_t length,
+                                     Message* msg) {
+  for (size_t i = 0; i != length; ) {  // NOLINT
+    if (wstr[i] != L'\0') {
+      *msg << WideStringToUtf8(wstr + i, static_cast<int>(length - i));
+      while (i != length && wstr[i] != L'\0')
+        i++;
+    } else {
+      *msg << '\0';
+      i++;
+    }
+  }
+}
+
+#endif  // GTEST_HAS_STD_WSTRING || GTEST_HAS_GLOBAL_WSTRING
+
+void SplitString(const ::std::string& str, char delimiter,
+                 ::std::vector< ::std::string>* dest) {
+  ::std::vector< ::std::string> parsed;
+  ::std::string::size_type pos = 0;
+  while (::testing::internal::AlwaysTrue()) {
+    const ::std::string::size_type colon = str.find(delimiter, pos);
+    if (colon == ::std::string::npos) {
+      parsed.push_back(str.substr(pos));
+      break;
+    } else {
+      parsed.push_back(str.substr(pos, colon - pos));
+      pos = colon + 1;
+    }
+  }
+  dest->swap(parsed);
+}
+
+}  // namespace internal
+
+// Constructs an empty Message.
+// We allocate the stringstream separately because otherwise each use of
+// ASSERT/EXPECT in a procedure adds over 200 bytes to the procedure's
+// stack frame leading to huge stack frames in some cases; gcc does not reuse
+// the stack space.
+Message::Message() : ss_(new ::std::stringstream) {
+  // By default, we want there to be enough precision when printing
+  // a double to a Message.
+  *ss_ << std::setprecision(std::numeric_limits<double>::digits10 + 2);
+}
+
+// These two overloads allow streaming a wide C string to a Message
+// using the UTF-8 encoding.
+Message& Message::operator <<(const wchar_t* wide_c_str) {
+  return *this << internal::String::ShowWideCString(wide_c_str);
+}
+Message& Message::operator <<(wchar_t* wide_c_str) {
+  return *this << internal::String::ShowWideCString(wide_c_str);
+}
+
+#if GTEST_HAS_STD_WSTRING
+// Converts the given wide string to a narrow string using the UTF-8
+// encoding, and streams the result to this Message object.
+Message& Message::operator <<(const ::std::wstring& wstr) {
+  internal::StreamWideCharsToMessage(wstr.c_str(), wstr.length(), this);
+  return *this;
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+#if GTEST_HAS_GLOBAL_WSTRING
+// Converts the given wide string to a narrow string using the UTF-8
+// encoding, and streams the result to this Message object.
+Message& Message::operator <<(const ::wstring& wstr) {
+  internal::StreamWideCharsToMessage(wstr.c_str(), wstr.length(), this);
+  return *this;
+}
+#endif  // GTEST_HAS_GLOBAL_WSTRING
+
+// Gets the text streamed to this object so far as an std::string.
+// Each '\0' character in the buffer is replaced with "\\0".
+std::string Message::GetString() const {
+  return internal::StringStreamToString(ss_.get());
+}
+
+// AssertionResult constructors.
+// Used in EXPECT_TRUE/FALSE(assertion_result).
+AssertionResult::AssertionResult(const AssertionResult& other)
+    : success_(other.success_),
+      message_(other.message_.get() != NULL ?
+               new ::std::string(*other.message_) :
+               static_cast< ::std::string*>(NULL)) {
+}
+
+// Swaps two AssertionResults.
+void AssertionResult::swap(AssertionResult& other) {
+  using std::swap;
+  swap(success_, other.success_);
+  swap(message_, other.message_);
+}
+
+// Returns the assertion's negation. Used with EXPECT/ASSERT_FALSE.
+AssertionResult AssertionResult::operator!() const {
+  AssertionResult negation(!success_);
+  if (message_.get() != NULL)
+    negation << *message_;
+  return negation;
+}
+
+// Makes a successful assertion result.
+AssertionResult AssertionSuccess() {
+  return AssertionResult(true);
+}
+
+// Makes a failed assertion result.
+AssertionResult AssertionFailure() {
+  return AssertionResult(false);
+}
+
+// Makes a failed assertion result with the given failure message.
+// Deprecated; use AssertionFailure() << message.
+AssertionResult AssertionFailure(const Message& message) {
+  return AssertionFailure() << message;
+}
+
+namespace internal {
+
+namespace edit_distance {
+std::vector<EditType> CalculateOptimalEdits(const std::vector<size_t>& left,
+                                            const std::vector<size_t>& right) {
+  std::vector<std::vector<double> > costs(
+      left.size() + 1, std::vector<double>(right.size() + 1));
+  std::vector<std::vector<EditType> > best_move(
+      left.size() + 1, std::vector<EditType>(right.size() + 1));
+
+  // Populate for empty right.
+  for (size_t l_i = 0; l_i < costs.size(); ++l_i) {
+    costs[l_i][0] = static_cast<double>(l_i);
+    best_move[l_i][0] = kRemove;
+  }
+  // Populate for empty left.
+  for (size_t r_i = 1; r_i < costs[0].size(); ++r_i) {
+    costs[0][r_i] = static_cast<double>(r_i);
+    best_move[0][r_i] = kAdd;
+  }
+
+  for (size_t l_i = 0; l_i < left.size(); ++l_i) {
+    for (size_t r_i = 0; r_i < right.size(); ++r_i) {
+      if (left[l_i] == right[r_i]) {
+        // Found a match. Consume it.
+        costs[l_i + 1][r_i + 1] = costs[l_i][r_i];
+        best_move[l_i + 1][r_i + 1] = kMatch;
+        continue;
+      }
+
+      const double add = costs[l_i + 1][r_i];
+      const double remove = costs[l_i][r_i + 1];
+      const double replace = costs[l_i][r_i];
+      if (add < remove && add < replace) {
+        costs[l_i + 1][r_i + 1] = add + 1;
+        best_move[l_i + 1][r_i + 1] = kAdd;
+      } else if (remove < add && remove < replace) {
+        costs[l_i + 1][r_i + 1] = remove + 1;
+        best_move[l_i + 1][r_i + 1] = kRemove;
+      } else {
+        // We make replace a little more expensive than add/remove to lower
+        // their priority.
+        costs[l_i + 1][r_i + 1] = replace + 1.00001;
+        best_move[l_i + 1][r_i + 1] = kReplace;
+      }
+    }
+  }
+
+  // Reconstruct the best path. We do it in reverse order.
+  std::vector<EditType> best_path;
+  for (size_t l_i = left.size(), r_i = right.size(); l_i > 0 || r_i > 0;) {
+    EditType move = best_move[l_i][r_i];
+    best_path.push_back(move);
+    l_i -= move != kAdd;
+    r_i -= move != kRemove;
+  }
+  std::reverse(best_path.begin(), best_path.end());
+  return best_path;
+}
+
+namespace {
+
+// Helper class to convert string into ids with deduplication.
+class InternalStrings {
+ public:
+  size_t GetId(const std::string& str) {
+    IdMap::iterator it = ids_.find(str);
+    if (it != ids_.end()) return it->second;
+    size_t id = ids_.size();
+    return ids_[str] = id;
+  }
+
+ private:
+  typedef std::map<std::string, size_t> IdMap;
+  IdMap ids_;
+};
+
+}  // namespace
+
+std::vector<EditType> CalculateOptimalEdits(
+    const std::vector<std::string>& left,
+    const std::vector<std::string>& right) {
+  std::vector<size_t> left_ids, right_ids;
+  {
+    InternalStrings intern_table;
+    for (size_t i = 0; i < left.size(); ++i) {
+      left_ids.push_back(intern_table.GetId(left[i]));
+    }
+    for (size_t i = 0; i < right.size(); ++i) {
+      right_ids.push_back(intern_table.GetId(right[i]));
+    }
+  }
+  return CalculateOptimalEdits(left_ids, right_ids);
+}
+
+namespace {
+
+// Helper class that holds the state for one hunk and prints it out to the
+// stream.
+// It reorders adds/removes when possible to group all removes before all
+// adds. It also adds the hunk header before printint into the stream.
+class Hunk {
+ public:
+  Hunk(size_t left_start, size_t right_start)
+      : left_start_(left_start),
+        right_start_(right_start),
+        adds_(),
+        removes_(),
+        common_() {}
+
+  void PushLine(char edit, const char* line) {
+    switch (edit) {
+      case ' ':
+        ++common_;
+        FlushEdits();
+        hunk_.push_back(std::make_pair(' ', line));
+        break;
+      case '-':
+        ++removes_;
+        hunk_removes_.push_back(std::make_pair('-', line));
+        break;
+      case '+':
+        ++adds_;
+        hunk_adds_.push_back(std::make_pair('+', line));
+        break;
+    }
+  }
+
+  void PrintTo(std::ostream* os) {
+    PrintHeader(os);
+    FlushEdits();
+    for (std::list<std::pair<char, const char*> >::const_iterator it =
+             hunk_.begin();
+         it != hunk_.end(); ++it) {
+      *os << it->first << it->second << "\n";
+    }
+  }
+
+  bool has_edits() const { return adds_ || removes_; }
+
+ private:
+  void FlushEdits() {
+    hunk_.splice(hunk_.end(), hunk_removes_);
+    hunk_.splice(hunk_.end(), hunk_adds_);
+  }
+
+  // Print a unified diff header for one hunk.
+  // The format is
+  //   "@@ -<left_start>,<left_length> +<right_start>,<right_length> @@"
+  // where the left/right parts are ommitted if unnecessary.
+  void PrintHeader(std::ostream* ss) const {
+    *ss << "@@ ";
+    if (removes_) {
+      *ss << "-" << left_start_ << "," << (removes_ + common_);
+    }
+    if (removes_ && adds_) {
+      *ss << " ";
+    }
+    if (adds_) {
+      *ss << "+" << right_start_ << "," << (adds_ + common_);
+    }
+    *ss << " @@\n";
+  }
+
+  size_t left_start_, right_start_;
+  size_t adds_, removes_, common_;
+  std::list<std::pair<char, const char*> > hunk_, hunk_adds_, hunk_removes_;
+};
+
+}  // namespace
+
+// Create a list of diff hunks in Unified diff format.
+// Each hunk has a header generated by PrintHeader above plus a body with
+// lines prefixed with ' ' for no change, '-' for deletion and '+' for
+// addition.
+// 'context' represents the desired unchanged prefix/suffix around the diff.
+// If two hunks are close enough that their contexts overlap, then they are
+// joined into one hunk.
+std::string CreateUnifiedDiff(const std::vector<std::string>& left,
+                              const std::vector<std::string>& right,
+                              size_t context) {
+  const std::vector<EditType> edits = CalculateOptimalEdits(left, right);
+
+  size_t l_i = 0, r_i = 0, edit_i = 0;
+  std::stringstream ss;
+  while (edit_i < edits.size()) {
+    // Find first edit.
+    while (edit_i < edits.size() && edits[edit_i] == kMatch) {
+      ++l_i;
+      ++r_i;
+      ++edit_i;
+    }
+
+    // Find the first line to include in the hunk.
+    const size_t prefix_context = std::min(l_i, context);
+    Hunk hunk(l_i - prefix_context + 1, r_i - prefix_context + 1);
+    for (size_t i = prefix_context; i > 0; --i) {
+      hunk.PushLine(' ', left[l_i - i].c_str());
+    }
+
+    // Iterate the edits until we found enough suffix for the hunk or the input
+    // is over.
+    size_t n_suffix = 0;
+    for (; edit_i < edits.size(); ++edit_i) {
+      if (n_suffix >= context) {
+        // Continue only if the next hunk is very close.
+        std::vector<EditType>::const_iterator it = edits.begin() + edit_i;
+        while (it != edits.end() && *it == kMatch) ++it;
+        if (it == edits.end() || (it - edits.begin()) - edit_i >= context) {
+          // There is no next edit or it is too far away.
+          break;
+        }
+      }
+
+      EditType edit = edits[edit_i];
+      // Reset count when a non match is found.
+      n_suffix = edit == kMatch ? n_suffix + 1 : 0;
+
+      if (edit == kMatch || edit == kRemove || edit == kReplace) {
+        hunk.PushLine(edit == kMatch ? ' ' : '-', left[l_i].c_str());
+      }
+      if (edit == kAdd || edit == kReplace) {
+        hunk.PushLine('+', right[r_i].c_str());
+      }
+
+      // Advance indices, depending on edit type.
+      l_i += edit != kAdd;
+      r_i += edit != kRemove;
+    }
+
+    if (!hunk.has_edits()) {
+      // We are done. We don't want this hunk.
+      break;
+    }
+
+    hunk.PrintTo(&ss);
+  }
+  return ss.str();
+}
+
+}  // namespace edit_distance
+
+namespace {
+
+// The string representation of the values received in EqFailure() are already
+// escaped. Split them on escaped '\n' boundaries. Leave all other escaped
+// characters the same.
+std::vector<std::string> SplitEscapedString(const std::string& str) {
+  std::vector<std::string> lines;
+  size_t start = 0, end = str.size();
+  if (end > 2 && str[0] == '"' && str[end - 1] == '"') {
+    ++start;
+    --end;
+  }
+  bool escaped = false;
+  for (size_t i = start; i + 1 < end; ++i) {
+    if (escaped) {
+      escaped = false;
+      if (str[i] == 'n') {
+        lines.push_back(str.substr(start, i - start - 1));
+        start = i + 1;
+      }
+    } else {
+      escaped = str[i] == '\\';
+    }
+  }
+  lines.push_back(str.substr(start, end - start));
+  return lines;
+}
+
+}  // namespace
+
+// Constructs and returns the message for an equality assertion
+// (e.g. ASSERT_EQ, EXPECT_STREQ, etc) failure.
+//
+// The first four parameters are the expressions used in the assertion
+// and their values, as strings.  For example, for ASSERT_EQ(foo, bar)
+// where foo is 5 and bar is 6, we have:
+//
+//   lhs_expression: "foo"
+//   rhs_expression: "bar"
+//   lhs_value:      "5"
+//   rhs_value:      "6"
+//
+// The ignoring_case parameter is true iff the assertion is a
+// *_STRCASEEQ*.  When it's true, the string "Ignoring case" will
+// be inserted into the message.
+AssertionResult EqFailure(const char* lhs_expression,
+                          const char* rhs_expression,
+                          const std::string& lhs_value,
+                          const std::string& rhs_value,
+                          bool ignoring_case) {
+  Message msg;
+  msg << "      Expected: " << lhs_expression;
+  if (lhs_value != lhs_expression) {
+    msg << "\n      Which is: " << lhs_value;
+  }
+  msg << "\nTo be equal to: " << rhs_expression;
+  if (rhs_value != rhs_expression) {
+    msg << "\n      Which is: " << rhs_value;
+  }
+
+  if (ignoring_case) {
+    msg << "\nIgnoring case";
+  }
+
+  if (!lhs_value.empty() && !rhs_value.empty()) {
+    const std::vector<std::string> lhs_lines =
+        SplitEscapedString(lhs_value);
+    const std::vector<std::string> rhs_lines =
+        SplitEscapedString(rhs_value);
+    if (lhs_lines.size() > 1 || rhs_lines.size() > 1) {
+      msg << "\nWith diff:\n"
+          << edit_distance::CreateUnifiedDiff(lhs_lines, rhs_lines);
+    }
+  }
+
+  return AssertionFailure() << msg;
+}
+
+// Constructs a failure message for Boolean assertions such as EXPECT_TRUE.
+std::string GetBoolAssertionFailureMessage(
+    const AssertionResult& assertion_result,
+    const char* expression_text,
+    const char* actual_predicate_value,
+    const char* expected_predicate_value) {
+  const char* actual_message = assertion_result.message();
+  Message msg;
+  msg << "Value of: " << expression_text
+      << "\n  Actual: " << actual_predicate_value;
+  if (actual_message[0] != '\0')
+    msg << " (" << actual_message << ")";
+  msg << "\nExpected: " << expected_predicate_value;
+  return msg.GetString();
+}
+
+// Helper function for implementing ASSERT_NEAR.
+AssertionResult DoubleNearPredFormat(const char* expr1,
+                                     const char* expr2,
+                                     const char* abs_error_expr,
+                                     double val1,
+                                     double val2,
+                                     double abs_error) {
+  const double diff = fabs(val1 - val2);
+  if (diff <= abs_error) return AssertionSuccess();
+
+  // TODO(wan): do not print the value of an expression if it's
+  // already a literal.
+  return AssertionFailure()
+      << "The difference between " << expr1 << " and " << expr2
+      << " is " << diff << ", which exceeds " << abs_error_expr << ", where\n"
+      << expr1 << " evaluates to " << val1 << ",\n"
+      << expr2 << " evaluates to " << val2 << ", and\n"
+      << abs_error_expr << " evaluates to " << abs_error << ".";
+}
+
+
+// Helper template for implementing FloatLE() and DoubleLE().
+template <typename RawType>
+AssertionResult FloatingPointLE(const char* expr1,
+                                const char* expr2,
+                                RawType val1,
+                                RawType val2) {
+  // Returns success if val1 is less than val2,
+  if (val1 < val2) {
+    return AssertionSuccess();
+  }
+
+  // or if val1 is almost equal to val2.
+  const FloatingPoint<RawType> lhs(val1), rhs(val2);
+  if (lhs.AlmostEquals(rhs)) {
+    return AssertionSuccess();
+  }
+
+  // Note that the above two checks will both fail if either val1 or
+  // val2 is NaN, as the IEEE floating-point standard requires that
+  // any predicate involving a NaN must return false.
+
+  ::std::stringstream val1_ss;
+  val1_ss << std::setprecision(std::numeric_limits<RawType>::digits10 + 2)
+          << val1;
+
+  ::std::stringstream val2_ss;
+  val2_ss << std::setprecision(std::numeric_limits<RawType>::digits10 + 2)
+          << val2;
+
+  return AssertionFailure()
+      << "Expected: (" << expr1 << ") <= (" << expr2 << ")\n"
+      << "  Actual: " << StringStreamToString(&val1_ss) << " vs "
+      << StringStreamToString(&val2_ss);
+}
+
+}  // namespace internal
+
+// Asserts that val1 is less than, or almost equal to, val2.  Fails
+// otherwise.  In particular, it fails if either val1 or val2 is NaN.
+AssertionResult FloatLE(const char* expr1, const char* expr2,
+                        float val1, float val2) {
+  return internal::FloatingPointLE<float>(expr1, expr2, val1, val2);
+}
+
+// Asserts that val1 is less than, or almost equal to, val2.  Fails
+// otherwise.  In particular, it fails if either val1 or val2 is NaN.
+AssertionResult DoubleLE(const char* expr1, const char* expr2,
+                         double val1, double val2) {
+  return internal::FloatingPointLE<double>(expr1, expr2, val1, val2);
+}
+
+namespace internal {
+
+// The helper function for {ASSERT|EXPECT}_EQ with int or enum
+// arguments.
+AssertionResult CmpHelperEQ(const char* lhs_expression,
+                            const char* rhs_expression,
+                            BiggestInt lhs,
+                            BiggestInt rhs) {
+  if (lhs == rhs) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(lhs_expression,
+                   rhs_expression,
+                   FormatForComparisonFailureMessage(lhs, rhs),
+                   FormatForComparisonFailureMessage(rhs, lhs),
+                   false);
+}
+
+// A macro for implementing the helper functions needed to implement
+// ASSERT_?? and EXPECT_?? with integer or enum arguments.  It is here
+// just to avoid copy-and-paste of similar code.
+#define GTEST_IMPL_CMP_HELPER_(op_name, op)\
+AssertionResult CmpHelper##op_name(const char* expr1, const char* expr2, \
+                                   BiggestInt val1, BiggestInt val2) {\
+  if (val1 op val2) {\
+    return AssertionSuccess();\
+  } else {\
+    return AssertionFailure() \
+        << "Expected: (" << expr1 << ") " #op " (" << expr2\
+        << "), actual: " << FormatForComparisonFailureMessage(val1, val2)\
+        << " vs " << FormatForComparisonFailureMessage(val2, val1);\
+  }\
+}
+
+// Implements the helper function for {ASSERT|EXPECT}_NE with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(NE, !=)
+// Implements the helper function for {ASSERT|EXPECT}_LE with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(LE, <=)
+// Implements the helper function for {ASSERT|EXPECT}_LT with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(LT, < )
+// Implements the helper function for {ASSERT|EXPECT}_GE with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(GE, >=)
+// Implements the helper function for {ASSERT|EXPECT}_GT with int or
+// enum arguments.
+GTEST_IMPL_CMP_HELPER_(GT, > )
+
+#undef GTEST_IMPL_CMP_HELPER_
+
+// The helper function for {ASSERT|EXPECT}_STREQ.
+AssertionResult CmpHelperSTREQ(const char* lhs_expression,
+                               const char* rhs_expression,
+                               const char* lhs,
+                               const char* rhs) {
+  if (String::CStringEquals(lhs, rhs)) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(lhs_expression,
+                   rhs_expression,
+                   PrintToString(lhs),
+                   PrintToString(rhs),
+                   false);
+}
+
+// The helper function for {ASSERT|EXPECT}_STRCASEEQ.
+AssertionResult CmpHelperSTRCASEEQ(const char* lhs_expression,
+                                   const char* rhs_expression,
+                                   const char* lhs,
+                                   const char* rhs) {
+  if (String::CaseInsensitiveCStringEquals(lhs, rhs)) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(lhs_expression,
+                   rhs_expression,
+                   PrintToString(lhs),
+                   PrintToString(rhs),
+                   true);
+}
+
+// The helper function for {ASSERT|EXPECT}_STRNE.
+AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                               const char* s2_expression,
+                               const char* s1,
+                               const char* s2) {
+  if (!String::CStringEquals(s1, s2)) {
+    return AssertionSuccess();
+  } else {
+    return AssertionFailure() << "Expected: (" << s1_expression << ") != ("
+                              << s2_expression << "), actual: \""
+                              << s1 << "\" vs \"" << s2 << "\"";
+  }
+}
+
+// The helper function for {ASSERT|EXPECT}_STRCASENE.
+AssertionResult CmpHelperSTRCASENE(const char* s1_expression,
+                                   const char* s2_expression,
+                                   const char* s1,
+                                   const char* s2) {
+  if (!String::CaseInsensitiveCStringEquals(s1, s2)) {
+    return AssertionSuccess();
+  } else {
+    return AssertionFailure()
+        << "Expected: (" << s1_expression << ") != ("
+        << s2_expression << ") (ignoring case), actual: \""
+        << s1 << "\" vs \"" << s2 << "\"";
+  }
+}
+
+}  // namespace internal
+
+namespace {
+
+// Helper functions for implementing IsSubString() and IsNotSubstring().
+
+// This group of overloaded functions return true iff needle is a
+// substring of haystack.  NULL is considered a substring of itself
+// only.
+
+bool IsSubstringPred(const char* needle, const char* haystack) {
+  if (needle == NULL || haystack == NULL)
+    return needle == haystack;
+
+  return strstr(haystack, needle) != NULL;
+}
+
+bool IsSubstringPred(const wchar_t* needle, const wchar_t* haystack) {
+  if (needle == NULL || haystack == NULL)
+    return needle == haystack;
+
+  return wcsstr(haystack, needle) != NULL;
+}
+
+// StringType here can be either ::std::string or ::std::wstring.
+template <typename StringType>
+bool IsSubstringPred(const StringType& needle,
+                     const StringType& haystack) {
+  return haystack.find(needle) != StringType::npos;
+}
+
+// This function implements either IsSubstring() or IsNotSubstring(),
+// depending on the value of the expected_to_be_substring parameter.
+// StringType here can be const char*, const wchar_t*, ::std::string,
+// or ::std::wstring.
+template <typename StringType>
+AssertionResult IsSubstringImpl(
+    bool expected_to_be_substring,
+    const char* needle_expr, const char* haystack_expr,
+    const StringType& needle, const StringType& haystack) {
+  if (IsSubstringPred(needle, haystack) == expected_to_be_substring)
+    return AssertionSuccess();
+
+  const bool is_wide_string = sizeof(needle[0]) > 1;
+  const char* const begin_string_quote = is_wide_string ? "L\"" : "\"";
+  return AssertionFailure()
+      << "Value of: " << needle_expr << "\n"
+      << "  Actual: " << begin_string_quote << needle << "\"\n"
+      << "Expected: " << (expected_to_be_substring ? "" : "not ")
+      << "a substring of " << haystack_expr << "\n"
+      << "Which is: " << begin_string_quote << haystack << "\"";
+}
+
+}  // namespace
+
+// IsSubstring() and IsNotSubstring() check whether needle is a
+// substring of haystack (NULL is considered a substring of itself
+// only), and return an appropriate error message when they fail.
+
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const char* needle, const char* haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const wchar_t* needle, const wchar_t* haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::string& needle, const ::std::string& haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+
+#if GTEST_HAS_STD_WSTRING
+AssertionResult IsSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack) {
+  return IsSubstringImpl(true, needle_expr, haystack_expr, needle, haystack);
+}
+
+AssertionResult IsNotSubstring(
+    const char* needle_expr, const char* haystack_expr,
+    const ::std::wstring& needle, const ::std::wstring& haystack) {
+  return IsSubstringImpl(false, needle_expr, haystack_expr, needle, haystack);
+}
+#endif  // GTEST_HAS_STD_WSTRING
+
+namespace internal {
+
+#if GTEST_OS_WINDOWS
+
+namespace {
+
+// Helper function for IsHRESULT{SuccessFailure} predicates
+AssertionResult HRESULTFailureHelper(const char* expr,
+                                     const char* expected,
+                                     long hr) {  // NOLINT
+# if GTEST_OS_WINDOWS_MOBILE
+
+  // Windows CE doesn't support FormatMessage.
+  const char error_text[] = "";
+
+# else
+
+  // Looks up the human-readable system message for the HRESULT code
+  // and since we're not passing any params to FormatMessage, we don't
+  // want inserts expanded.
+  const DWORD kFlags = FORMAT_MESSAGE_FROM_SYSTEM |
+                       FORMAT_MESSAGE_IGNORE_INSERTS;
+  const DWORD kBufSize = 4096;
+  // Gets the system's human readable message string for this HRESULT.
+  char error_text[kBufSize] = { '\0' };
+  DWORD message_length = ::FormatMessageA(kFlags,
+                                          0,  // no source, we're asking system
+                                          hr,  // the error
+                                          0,  // no line width restrictions
+                                          error_text,  // output buffer
+                                          kBufSize,  // buf size
+                                          NULL);  // no arguments for inserts
+  // Trims tailing white space (FormatMessage leaves a trailing CR-LF)
+  for (; message_length && IsSpace(error_text[message_length - 1]);
+          --message_length) {
+    error_text[message_length - 1] = '\0';
+  }
+
+# endif  // GTEST_OS_WINDOWS_MOBILE
+
+  const std::string error_hex("0x" + String::FormatHexInt(hr));
+  return ::testing::AssertionFailure()
+      << "Expected: " << expr << " " << expected << ".\n"
+      << "  Actual: " << error_hex << " " << error_text << "\n";
+}
+
+}  // namespace
+
+AssertionResult IsHRESULTSuccess(const char* expr, long hr) {  // NOLINT
+  if (SUCCEEDED(hr)) {
+    return AssertionSuccess();
+  }
+  return HRESULTFailureHelper(expr, "succeeds", hr);
+}
+
+AssertionResult IsHRESULTFailure(const char* expr, long hr) {  // NOLINT
+  if (FAILED(hr)) {
+    return AssertionSuccess();
+  }
+  return HRESULTFailureHelper(expr, "fails", hr);
+}
+
+#endif  // GTEST_OS_WINDOWS
+
+// Utility functions for encoding Unicode text (wide strings) in
+// UTF-8.
+
+// A Unicode code-point can have upto 21 bits, and is encoded in UTF-8
+// like this:
+//
+// Code-point length   Encoding
+//   0 -  7 bits       0xxxxxxx
+//   8 - 11 bits       110xxxxx 10xxxxxx
+//  12 - 16 bits       1110xxxx 10xxxxxx 10xxxxxx
+//  17 - 21 bits       11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+
+// The maximum code-point a one-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint1 = (static_cast<UInt32>(1) <<  7) - 1;
+
+// The maximum code-point a two-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint2 = (static_cast<UInt32>(1) << (5 + 6)) - 1;
+
+// The maximum code-point a three-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint3 = (static_cast<UInt32>(1) << (4 + 2*6)) - 1;
+
+// The maximum code-point a four-byte UTF-8 sequence can represent.
+const UInt32 kMaxCodePoint4 = (static_cast<UInt32>(1) << (3 + 3*6)) - 1;
+
+// Chops off the n lowest bits from a bit pattern.  Returns the n
+// lowest bits.  As a side effect, the original bit pattern will be
+// shifted to the right by n bits.
+inline UInt32 ChopLowBits(UInt32* bits, int n) {
+  const UInt32 low_bits = *bits & ((static_cast<UInt32>(1) << n) - 1);
+  *bits >>= n;
+  return low_bits;
+}
+
+// Converts a Unicode code point to a narrow string in UTF-8 encoding.
+// code_point parameter is of type UInt32 because wchar_t may not be
+// wide enough to contain a code point.
+// If the code_point is not a valid Unicode code point
+// (i.e. outside of Unicode range U+0 to U+10FFFF) it will be converted
+// to "(Invalid Unicode 0xXXXXXXXX)".
+std::string CodePointToUtf8(UInt32 code_point) {
+  if (code_point > kMaxCodePoint4) {
+    return "(Invalid Unicode 0x" + String::FormatHexInt(code_point) + ")";
+  }
+
+  char str[5];  // Big enough for the largest valid code point.
+  if (code_point <= kMaxCodePoint1) {
+    str[1] = '\0';
+    str[0] = static_cast<char>(code_point);                          // 0xxxxxxx
+  } else if (code_point <= kMaxCodePoint2) {
+    str[2] = '\0';
+    str[1] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[0] = static_cast<char>(0xC0 | code_point);                   // 110xxxxx
+  } else if (code_point <= kMaxCodePoint3) {
+    str[3] = '\0';
+    str[2] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[1] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[0] = static_cast<char>(0xE0 | code_point);                   // 1110xxxx
+  } else {  // code_point <= kMaxCodePoint4
+    str[4] = '\0';
+    str[3] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[2] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[1] = static_cast<char>(0x80 | ChopLowBits(&code_point, 6));  // 10xxxxxx
+    str[0] = static_cast<char>(0xF0 | code_point);                   // 11110xxx
+  }
+  return str;
+}
+
+// The following two functions only make sense if the the system
+// uses UTF-16 for wide string encoding. All supported systems
+// with 16 bit wchar_t (Windows, Cygwin, Symbian OS) do use UTF-16.
+
+// Determines if the arguments constitute UTF-16 surrogate pair
+// and thus should be combined into a single Unicode code point
+// using CreateCodePointFromUtf16SurrogatePair.
+inline bool IsUtf16SurrogatePair(wchar_t first, wchar_t second) {
+  return sizeof(wchar_t) == 2 &&
+      (first & 0xFC00) == 0xD800 && (second & 0xFC00) == 0xDC00;
+}
+
+// Creates a Unicode code point from UTF16 surrogate pair.
+inline UInt32 CreateCodePointFromUtf16SurrogatePair(wchar_t first,
+                                                    wchar_t second) {
+  const UInt32 mask = (1 << 10) - 1;
+  return (sizeof(wchar_t) == 2) ?
+      (((first & mask) << 10) | (second & mask)) + 0x10000 :
+      // This function should not be called when the condition is
+      // false, but we provide a sensible default in case it is.
+      static_cast<UInt32>(first);
+}
+
+// Converts a wide string to a narrow string in UTF-8 encoding.
+// The wide string is assumed to have the following encoding:
+//   UTF-16 if sizeof(wchar_t) == 2 (on Windows, Cygwin, Symbian OS)
+//   UTF-32 if sizeof(wchar_t) == 4 (on Linux)
+// Parameter str points to a null-terminated wide string.
+// Parameter num_chars may additionally limit the number
+// of wchar_t characters processed. -1 is used when the entire string
+// should be processed.
+// If the string contains code points that are not valid Unicode code points
+// (i.e. outside of Unicode range U+0 to U+10FFFF) they will be output
+// as '(Invalid Unicode 0xXXXXXXXX)'. If the string is in UTF16 encoding
+// and contains invalid UTF-16 surrogate pairs, values in those pairs
+// will be encoded as individual Unicode characters from Basic Normal Plane.
+std::string WideStringToUtf8(const wchar_t* str, int num_chars) {
+  if (num_chars == -1)
+    num_chars = static_cast<int>(wcslen(str));
+
+  ::std::stringstream stream;
+  for (int i = 0; i < num_chars; ++i) {
+    UInt32 unicode_code_point;
+
+    if (str[i] == L'\0') {
+      break;
+    } else if (i + 1 < num_chars && IsUtf16SurrogatePair(str[i], str[i + 1])) {
+      unicode_code_point = CreateCodePointFromUtf16SurrogatePair(str[i],
+                                                                 str[i + 1]);
+      i++;
+    } else {
+      unicode_code_point = static_cast<UInt32>(str[i]);
+    }
+
+    stream << CodePointToUtf8(unicode_code_point);
+  }
+  return StringStreamToString(&stream);
+}
+
+// Converts a wide C string to an std::string using the UTF-8 encoding.
+// NULL will be converted to "(null)".
+std::string String::ShowWideCString(const wchar_t * wide_c_str) {
+  if (wide_c_str == NULL)  return "(null)";
+
+  return internal::WideStringToUtf8(wide_c_str, -1);
+}
+
+// Compares two wide C strings.  Returns true iff they have the same
+// content.
+//
+// Unlike wcscmp(), this function can handle NULL argument(s).  A NULL
+// C string is considered different to any non-NULL C string,
+// including the empty string.
+bool String::WideCStringEquals(const wchar_t * lhs, const wchar_t * rhs) {
+  if (lhs == NULL) return rhs == NULL;
+
+  if (rhs == NULL) return false;
+
+  return wcscmp(lhs, rhs) == 0;
+}
+
+// Helper function for *_STREQ on wide strings.
+AssertionResult CmpHelperSTREQ(const char* lhs_expression,
+                               const char* rhs_expression,
+                               const wchar_t* lhs,
+                               const wchar_t* rhs) {
+  if (String::WideCStringEquals(lhs, rhs)) {
+    return AssertionSuccess();
+  }
+
+  return EqFailure(lhs_expression,
+                   rhs_expression,
+                   PrintToString(lhs),
+                   PrintToString(rhs),
+                   false);
+}
+
+// Helper function for *_STRNE on wide strings.
+AssertionResult CmpHelperSTRNE(const char* s1_expression,
+                               const char* s2_expression,
+                               const wchar_t* s1,
+                               const wchar_t* s2) {
+  if (!String::WideCStringEquals(s1, s2)) {
+    return AssertionSuccess();
+  }
+
+  return AssertionFailure() << "Expected: (" << s1_expression << ") != ("
+                            << s2_expression << "), actual: "
+                            << PrintToString(s1)
+                            << " vs " << PrintToString(s2);
+}
+
+// Compares two C strings, ignoring case.  Returns true iff they have
+// the same content.
+//
+// Unlike strcasecmp(), this function can handle NULL argument(s).  A
+// NULL C string is considered different to any non-NULL C string,
+// including the empty string.
+bool String::CaseInsensitiveCStringEquals(const char * lhs, const char * rhs) {
+  if (lhs == NULL)
+    return rhs == NULL;
+  if (rhs == NULL)
+    return false;
+  return posix::StrCaseCmp(lhs, rhs) == 0;
+}
+
+  // Compares two wide C strings, ignoring case.  Returns true iff they
+  // have the same content.
+  //
+  // Unlike wcscasecmp(), this function can handle NULL argument(s).
+  // A NULL C string is considered different to any non-NULL wide C string,
+  // including the empty string.
+  // NB: The implementations on different platforms slightly differ.
+  // On windows, this method uses _wcsicmp which compares according to LC_CTYPE
+  // environment variable. On GNU platform this method uses wcscasecmp
+  // which compares according to LC_CTYPE category of the current locale.
+  // On MacOS X, it uses towlower, which also uses LC_CTYPE category of the
+  // current locale.
+bool String::CaseInsensitiveWideCStringEquals(const wchar_t* lhs,
+                                              const wchar_t* rhs) {
+  if (lhs == NULL) return rhs == NULL;
+
+  if (rhs == NULL) return false;
+
+#if GTEST_OS_WINDOWS
+  return _wcsicmp(lhs, rhs) == 0;
+#elif GTEST_OS_LINUX && !GTEST_OS_LINUX_ANDROID
+  return wcscasecmp(lhs, rhs) == 0;
+#else
+  // Android, Mac OS X and Cygwin don't define wcscasecmp.
+  // Other unknown OSes may not define it either.
+  wint_t left, right;
+  do {
+    left = towlower(*lhs++);
+    right = towlower(*rhs++);
+  } while (left && left == right);
+  return left == right;
+#endif  // OS selector
+}
+
+// Returns true iff str ends with the given suffix, ignoring case.
+// Any string is considered to end with an empty suffix.
+bool String::EndsWithCaseInsensitive(
+    const std::string& str, const std::string& suffix) {
+  const size_t str_len = str.length();
+  const size_t suffix_len = suffix.length();
+  return (str_len >= suffix_len) &&
+         CaseInsensitiveCStringEquals(str.c_str() + str_len - suffix_len,
+                                      suffix.c_str());
+}
+
+// Formats an int value as "%02d".
+std::string String::FormatIntWidth2(int value) {
+  std::stringstream ss;
+  ss << std::setfill('0') << std::setw(2) << value;
+  return ss.str();
+}
+
+// Formats an int value as "%X".
+std::string String::FormatHexInt(int value) {
+  std::stringstream ss;
+  ss << std::hex << std::uppercase << value;
+  return ss.str();
+}
+
+// Formats a byte as "%02X".
+std::string String::FormatByte(unsigned char value) {
+  std::stringstream ss;
+  ss << std::setfill('0') << std::setw(2) << std::hex << std::uppercase
+     << static_cast<unsigned int>(value);
+  return ss.str();
+}
+
+// Converts the buffer in a stringstream to an std::string, converting NUL
+// bytes to "\\0" along the way.
+std::string StringStreamToString(::std::stringstream* ss) {
+  const ::std::string& str = ss->str();
+  const char* const start = str.c_str();
+  const char* const end = start + str.length();
+
+  std::string result;
+  result.reserve(2 * (end - start));
+  for (const char* ch = start; ch != end; ++ch) {
+    if (*ch == '\0') {
+      result += "\\0";  // Replaces NUL with "\\0";
+    } else {
+      result += *ch;
+    }
+  }
+
+  return result;
+}
+
+// Appends the user-supplied message to the Google-Test-generated message.
+std::string AppendUserMessage(const std::string& gtest_msg,
+                              const Message& user_msg) {
+  // Appends the user message if it's non-empty.
+  const std::string user_msg_string = user_msg.GetString();
+  if (user_msg_string.empty()) {
+    return gtest_msg;
+  }
+
+  return gtest_msg + "\n" + user_msg_string;
+}
+
+}  // namespace internal
+
+// class TestResult
+
+// Creates an empty TestResult.
+TestResult::TestResult()
+    : death_test_count_(0),
+      elapsed_time_(0) {
+}
+
+// D'tor.
+TestResult::~TestResult() {
+}
+
+// Returns the i-th test part result among all the results. i can
+// range from 0 to total_part_count() - 1. If i is not in that range,
+// aborts the program.
+const TestPartResult& TestResult::GetTestPartResult(int i) const {
+  if (i < 0 || i >= total_part_count())
+    internal::posix::Abort();
+  return test_part_results_.at(i);
+}
+
+// Returns the i-th test property. i can range from 0 to
+// test_property_count() - 1. If i is not in that range, aborts the
+// program.
+const TestProperty& TestResult::GetTestProperty(int i) const {
+  if (i < 0 || i >= test_property_count())
+    internal::posix::Abort();
+  return test_properties_.at(i);
+}
+
+// Clears the test part results.
+void TestResult::ClearTestPartResults() {
+  test_part_results_.clear();
+}
+
+// Adds a test part result to the list.
+void TestResult::AddTestPartResult(const TestPartResult& test_part_result) {
+  test_part_results_.push_back(test_part_result);
+}
+
+// Adds a test property to the list. If a property with the same key as the
+// supplied property is already represented, the value of this test_property
+// replaces the old value for that key.
+void TestResult::RecordProperty(const std::string& xml_element,
+                                const TestProperty& test_property) {
+  if (!ValidateTestProperty(xml_element, test_property)) {
+    return;
+  }
+  internal::MutexLock lock(&test_properites_mutex_);
+  const std::vector<TestProperty>::iterator property_with_matching_key =
+      std::find_if(test_properties_.begin(), test_properties_.end(),
+                   internal::TestPropertyKeyIs(test_property.key()));
+  if (property_with_matching_key == test_properties_.end()) {
+    test_properties_.push_back(test_property);
+    return;
+  }
+  property_with_matching_key->SetValue(test_property.value());
+}
+
+// The list of reserved attributes used in the <testsuites> element of XML
+// output.
+static const char* const kReservedTestSuitesAttributes[] = {
+  "disabled",
+  "errors",
+  "failures",
+  "name",
+  "random_seed",
+  "tests",
+  "time",
+  "timestamp"
+};
+
+// The list of reserved attributes used in the <testsuite> element of XML
+// output.
+static const char* const kReservedTestSuiteAttributes[] = {
+  "disabled",
+  "errors",
+  "failures",
+  "name",
+  "tests",
+  "time"
+};
+
+// The list of reserved attributes used in the <testcase> element of XML output.
+static const char* const kReservedTestCaseAttributes[] = {
+  "classname",
+  "name",
+  "status",
+  "time",
+  "type_param",
+  "value_param"
+};
+
+template <int kSize>
+std::vector<std::string> ArrayAsVector(const char* const (&array)[kSize]) {
+  return std::vector<std::string>(array, array + kSize);
+}
+
+static std::vector<std::string> GetReservedAttributesForElement(
+    const std::string& xml_element) {
+  if (xml_element == "testsuites") {
+    return ArrayAsVector(kReservedTestSuitesAttributes);
+  } else if (xml_element == "testsuite") {
+    return ArrayAsVector(kReservedTestSuiteAttributes);
+  } else if (xml_element == "testcase") {
+    return ArrayAsVector(kReservedTestCaseAttributes);
+  } else {
+    GTEST_CHECK_(false) << "Unrecognized xml_element provided: " << xml_element;
+  }
+  // This code is unreachable but some compilers may not realizes that.
+  return std::vector<std::string>();
+}
+
+static std::string FormatWordList(const std::vector<std::string>& words) {
+  Message word_list;
+  for (size_t i = 0; i < words.size(); ++i) {
+    if (i > 0 && words.size() > 2) {
+      word_list << ", ";
+    }
+    if (i == words.size() - 1) {
+      word_list << "and ";
+    }
+    word_list << "'" << words[i] << "'";
+  }
+  return word_list.GetString();
+}
+
+bool ValidateTestPropertyName(const std::string& property_name,
+                              const std::vector<std::string>& reserved_names) {
+  if (std::find(reserved_names.begin(), reserved_names.end(), property_name) !=
+          reserved_names.end()) {
+    ADD_FAILURE() << "Reserved key used in RecordProperty(): " << property_name
+                  << " (" << FormatWordList(reserved_names)
+                  << " are reserved by " << GTEST_NAME_ << ")";
+    return false;
+  }
+  return true;
+}
+
+// Adds a failure if the key is a reserved attribute of the element named
+// xml_element.  Returns true if the property is valid.
+bool TestResult::ValidateTestProperty(const std::string& xml_element,
+                                      const TestProperty& test_property) {
+  return ValidateTestPropertyName(test_property.key(),
+                                  GetReservedAttributesForElement(xml_element));
+}
+
+// Clears the object.
+void TestResult::Clear() {
+  test_part_results_.clear();
+  test_properties_.clear();
+  death_test_count_ = 0;
+  elapsed_time_ = 0;
+}
+
+// Returns true iff the test failed.
+bool TestResult::Failed() const {
+  for (int i = 0; i < total_part_count(); ++i) {
+    if (GetTestPartResult(i).failed())
+      return true;
+  }
+  return false;
+}
+
+// Returns true iff the test part fatally failed.
+static bool TestPartFatallyFailed(const TestPartResult& result) {
+  return result.fatally_failed();
+}
+
+// Returns true iff the test fatally failed.
+bool TestResult::HasFatalFailure() const {
+  return CountIf(test_part_results_, TestPartFatallyFailed) > 0;
+}
+
+// Returns true iff the test part non-fatally failed.
+static bool TestPartNonfatallyFailed(const TestPartResult& result) {
+  return result.nonfatally_failed();
+}
+
+// Returns true iff the test has a non-fatal failure.
+bool TestResult::HasNonfatalFailure() const {
+  return CountIf(test_part_results_, TestPartNonfatallyFailed) > 0;
+}
+
+// Gets the number of all test parts.  This is the sum of the number
+// of successful test parts and the number of failed test parts.
+int TestResult::total_part_count() const {
+  return static_cast<int>(test_part_results_.size());
+}
+
+// Returns the number of the test properties.
+int TestResult::test_property_count() const {
+  return static_cast<int>(test_properties_.size());
+}
+
+// class Test
+
+// Creates a Test object.
+
+// The c'tor saves the states of all flags.
+Test::Test()
+    : gtest_flag_saver_(new GTEST_FLAG_SAVER_) {
+}
+
+// The d'tor restores the states of all flags.  The actual work is
+// done by the d'tor of the gtest_flag_saver_ field, and thus not
+// visible here.
+Test::~Test() {
+}
+
+// Sets up the test fixture.
+//
+// A sub-class may override this.
+void Test::SetUp() {
+}
+
+// Tears down the test fixture.
+//
+// A sub-class may override this.
+void Test::TearDown() {
+}
+
+// Allows user supplied key value pairs to be recorded for later output.
+void Test::RecordProperty(const std::string& key, const std::string& value) {
+  UnitTest::GetInstance()->RecordProperty(key, value);
+}
+
+// Allows user supplied key value pairs to be recorded for later output.
+void Test::RecordProperty(const std::string& key, int value) {
+  Message value_message;
+  value_message << value;
+  RecordProperty(key, value_message.GetString().c_str());
+}
+
+namespace internal {
+
+void ReportFailureInUnknownLocation(TestPartResult::Type result_type,
+                                    const std::string& message) {
+  // This function is a friend of UnitTest and as such has access to
+  // AddTestPartResult.
+  UnitTest::GetInstance()->AddTestPartResult(
+      result_type,
+      NULL,  // No info about the source file where the exception occurred.
+      -1,    // We have no info on which line caused the exception.
+      message,
+      "");   // No stack trace, either.
+}
+
+}  // namespace internal
+
+// Google Test requires all tests in the same test case to use the same test
+// fixture class.  This function checks if the current test has the
+// same fixture class as the first test in the current test case.  If
+// yes, it returns true; otherwise it generates a Google Test failure and
+// returns false.
+bool Test::HasSameFixtureClass() {
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  const TestCase* const test_case = impl->current_test_case();
+
+  // Info about the first test in the current test case.
+  const TestInfo* const first_test_info = test_case->test_info_list()[0];
+  const internal::TypeId first_fixture_id = first_test_info->fixture_class_id_;
+  const char* const first_test_name = first_test_info->name();
+
+  // Info about the current test.
+  const TestInfo* const this_test_info = impl->current_test_info();
+  const internal::TypeId this_fixture_id = this_test_info->fixture_class_id_;
+  const char* const this_test_name = this_test_info->name();
+
+  if (this_fixture_id != first_fixture_id) {
+    // Is the first test defined using TEST?
+    const bool first_is_TEST = first_fixture_id == internal::GetTestTypeId();
+    // Is this test defined using TEST?
+    const bool this_is_TEST = this_fixture_id == internal::GetTestTypeId();
+
+    if (first_is_TEST || this_is_TEST) {
+      // Both TEST and TEST_F appear in same test case, which is incorrect.
+      // Tell the user how to fix this.
+
+      // Gets the name of the TEST and the name of the TEST_F.  Note
+      // that first_is_TEST and this_is_TEST cannot both be true, as
+      // the fixture IDs are different for the two tests.
+      const char* const TEST_name =
+          first_is_TEST ? first_test_name : this_test_name;
+      const char* const TEST_F_name =
+          first_is_TEST ? this_test_name : first_test_name;
+
+      ADD_FAILURE()
+          << "All tests in the same test case must use the same test fixture\n"
+          << "class, so mixing TEST_F and TEST in the same test case is\n"
+          << "illegal.  In test case " << this_test_info->test_case_name()
+          << ",\n"
+          << "test " << TEST_F_name << " is defined using TEST_F but\n"
+          << "test " << TEST_name << " is defined using TEST.  You probably\n"
+          << "want to change the TEST to TEST_F or move it to another test\n"
+          << "case.";
+    } else {
+      // Two fixture classes with the same name appear in two different
+      // namespaces, which is not allowed. Tell the user how to fix this.
+      ADD_FAILURE()
+          << "All tests in the same test case must use the same test fixture\n"
+          << "class.  However, in test case "
+          << this_test_info->test_case_name() << ",\n"
+          << "you defined test " << first_test_name
+          << " and test " << this_test_name << "\n"
+          << "using two different test fixture classes.  This can happen if\n"
+          << "the two classes are from different namespaces or translation\n"
+          << "units and have the same name.  You should probably rename one\n"
+          << "of the classes to put the tests into different test cases.";
+    }
+    return false;
+  }
+
+  return true;
+}
+
+#if GTEST_HAS_SEH
+
+// Adds an "exception thrown" fatal failure to the current test.  This
+// function returns its result via an output parameter pointer because VC++
+// prohibits creation of objects with destructors on stack in functions
+// using __try (see error C2712).
+static std::string* FormatSehExceptionMessage(DWORD exception_code,
+                                              const char* location) {
+  Message message;
+  message << "SEH exception with code 0x" << std::setbase(16) <<
+    exception_code << std::setbase(10) << " thrown in " << location << ".";
+
+  return new std::string(message.GetString());
+}
+
+#endif  // GTEST_HAS_SEH
+
+namespace internal {
+
+#if GTEST_HAS_EXCEPTIONS
+
+// Adds an "exception thrown" fatal failure to the current test.
+static std::string FormatCxxExceptionMessage(const char* description,
+                                             const char* location) {
+  Message message;
+  if (description != NULL) {
+    message << "C++ exception with description \"" << description << "\"";
+  } else {
+    message << "Unknown C++ exception";
+  }
+  message << " thrown in " << location << ".";
+
+  return message.GetString();
+}
+
+static std::string PrintTestPartResultToString(
+    const TestPartResult& test_part_result);
+
+GoogleTestFailureException::GoogleTestFailureException(
+    const TestPartResult& failure)
+    : ::std::runtime_error(PrintTestPartResultToString(failure).c_str()) {}
+
+#endif  // GTEST_HAS_EXCEPTIONS
+
+// We put these helper functions in the internal namespace as IBM's xlC
+// compiler rejects the code if they were declared static.
+
+// Runs the given method and handles SEH exceptions it throws, when
+// SEH is supported; returns the 0-value for type Result in case of an
+// SEH exception.  (Microsoft compilers cannot handle SEH and C++
+// exceptions in the same function.  Therefore, we provide a separate
+// wrapper function for handling SEH exceptions.)
+template <class T, typename Result>
+Result HandleSehExceptionsInMethodIfSupported(
+    T* object, Result (T::*method)(), const char* location) {
+#if GTEST_HAS_SEH
+  __try {
+    return (object->*method)();
+  } __except (internal::UnitTestOptions::GTestShouldProcessSEH(  // NOLINT
+      GetExceptionCode())) {
+    // We create the exception message on the heap because VC++ prohibits
+    // creation of objects with destructors on stack in functions using __try
+    // (see error C2712).
+    std::string* exception_message = FormatSehExceptionMessage(
+        GetExceptionCode(), location);
+    internal::ReportFailureInUnknownLocation(TestPartResult::kFatalFailure,
+                                             *exception_message);
+    delete exception_message;
+    return static_cast<Result>(0);
+  }
+#else
+  (void)location;
+  return (object->*method)();
+#endif  // GTEST_HAS_SEH
+}
+
+// Runs the given method and catches and reports C++ and/or SEH-style
+// exceptions, if they are supported; returns the 0-value for type
+// Result in case of an SEH exception.
+template <class T, typename Result>
+Result HandleExceptionsInMethodIfSupported(
+    T* object, Result (T::*method)(), const char* location) {
+  // NOTE: The user code can affect the way in which Google Test handles
+  // exceptions by setting GTEST_FLAG(catch_exceptions), but only before
+  // RUN_ALL_TESTS() starts. It is technically possible to check the flag
+  // after the exception is caught and either report or re-throw the
+  // exception based on the flag's value:
+  //
+  // try {
+  //   // Perform the test method.
+  // } catch (...) {
+  //   if (GTEST_FLAG(catch_exceptions))
+  //     // Report the exception as failure.
+  //   else
+  //     throw;  // Re-throws the original exception.
+  // }
+  //
+  // However, the purpose of this flag is to allow the program to drop into
+  // the debugger when the exception is thrown. On most platforms, once the
+  // control enters the catch block, the exception origin information is
+  // lost and the debugger will stop the program at the point of the
+  // re-throw in this function -- instead of at the point of the original
+  // throw statement in the code under test.  For this reason, we perform
+  // the check early, sacrificing the ability to affect Google Test's
+  // exception handling in the method where the exception is thrown.
+  if (internal::GetUnitTestImpl()->catch_exceptions()) {
+#if GTEST_HAS_EXCEPTIONS
+    try {
+      return HandleSehExceptionsInMethodIfSupported(object, method, location);
+    } catch (const internal::GoogleTestFailureException&) {  // NOLINT
+      // This exception type can only be thrown by a failed Google
+      // Test assertion with the intention of letting another testing
+      // framework catch it.  Therefore we just re-throw it.
+      throw;
+    } catch (const std::exception& e) {  // NOLINT
+      internal::ReportFailureInUnknownLocation(
+          TestPartResult::kFatalFailure,
+          FormatCxxExceptionMessage(e.what(), location));
+    } catch (...) {  // NOLINT
+      internal::ReportFailureInUnknownLocation(
+          TestPartResult::kFatalFailure,
+          FormatCxxExceptionMessage(NULL, location));
+    }
+    return static_cast<Result>(0);
+#else
+    return HandleSehExceptionsInMethodIfSupported(object, method, location);
+#endif  // GTEST_HAS_EXCEPTIONS
+  } else {
+    return (object->*method)();
+  }
+}
+
+}  // namespace internal
+
+// Runs the test and updates the test result.
+void Test::Run() {
+  if (!HasSameFixtureClass()) return;
+
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(this, &Test::SetUp, "SetUp()");
+  // We will run the test only if SetUp() was successful.
+  if (!HasFatalFailure()) {
+    impl->os_stack_trace_getter()->UponLeavingGTest();
+    internal::HandleExceptionsInMethodIfSupported(
+        this, &Test::TestBody, "the test body");
+  }
+
+  // However, we want to clean up as much as possible.  Hence we will
+  // always call TearDown(), even if SetUp() or the test body has
+  // failed.
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      this, &Test::TearDown, "TearDown()");
+}
+
+// Returns true iff the current test has a fatal failure.
+bool Test::HasFatalFailure() {
+  return internal::GetUnitTestImpl()->current_test_result()->HasFatalFailure();
+}
+
+// Returns true iff the current test has a non-fatal failure.
+bool Test::HasNonfatalFailure() {
+  return internal::GetUnitTestImpl()->current_test_result()->
+      HasNonfatalFailure();
+}
+
+// class TestInfo
+
+// Constructs a TestInfo object. It assumes ownership of the test factory
+// object.
+TestInfo::TestInfo(const std::string& a_test_case_name,
+                   const std::string& a_name,
+                   const char* a_type_param,
+                   const char* a_value_param,
+                   internal::CodeLocation a_code_location,
+                   internal::TypeId fixture_class_id,
+                   internal::TestFactoryBase* factory)
+    : test_case_name_(a_test_case_name),
+      name_(a_name),
+      type_param_(a_type_param ? new std::string(a_type_param) : NULL),
+      value_param_(a_value_param ? new std::string(a_value_param) : NULL),
+      location_(a_code_location),
+      fixture_class_id_(fixture_class_id),
+      should_run_(false),
+      is_disabled_(false),
+      matches_filter_(false),
+      factory_(factory),
+      result_() {}
+
+// Destructs a TestInfo object.
+TestInfo::~TestInfo() { delete factory_; }
+
+namespace internal {
+
+// Creates a new TestInfo object and registers it with Google Test;
+// returns the created object.
+//
+// Arguments:
+//
+//   test_case_name:   name of the test case
+//   name:             name of the test
+//   type_param:       the name of the test's type parameter, or NULL if
+//                     this is not a typed or a type-parameterized test.
+//   value_param:      text representation of the test's value parameter,
+//                     or NULL if this is not a value-parameterized test.
+//   code_location:    code location where the test is defined
+//   fixture_class_id: ID of the test fixture class
+//   set_up_tc:        pointer to the function that sets up the test case
+//   tear_down_tc:     pointer to the function that tears down the test case
+//   factory:          pointer to the factory that creates a test object.
+//                     The newly created TestInfo instance will assume
+//                     ownership of the factory object.
+TestInfo* MakeAndRegisterTestInfo(
+    const char* test_case_name,
+    const char* name,
+    const char* type_param,
+    const char* value_param,
+    CodeLocation code_location,
+    TypeId fixture_class_id,
+    SetUpTestCaseFunc set_up_tc,
+    TearDownTestCaseFunc tear_down_tc,
+    TestFactoryBase* factory) {
+  TestInfo* const test_info =
+      new TestInfo(test_case_name, name, type_param, value_param,
+                   code_location, fixture_class_id, factory);
+  GetUnitTestImpl()->AddTestInfo(set_up_tc, tear_down_tc, test_info);
+  return test_info;
+}
+
+#if GTEST_HAS_PARAM_TEST
+void ReportInvalidTestCaseType(const char* test_case_name,
+                               CodeLocation code_location) {
+  Message errors;
+  errors
+      << "Attempted redefinition of test case " << test_case_name << ".\n"
+      << "All tests in the same test case must use the same test fixture\n"
+      << "class.  However, in test case " << test_case_name << ", you tried\n"
+      << "to define a test using a fixture class different from the one\n"
+      << "used earlier. This can happen if the two fixture classes are\n"
+      << "from different namespaces and have the same name. You should\n"
+      << "probably rename one of the classes to put the tests into different\n"
+      << "test cases.";
+
+  fprintf(stderr, "%s %s",
+          FormatFileLocation(code_location.file.c_str(),
+                             code_location.line).c_str(),
+          errors.GetString().c_str());
+}
+#endif  // GTEST_HAS_PARAM_TEST
+
+}  // namespace internal
+
+namespace {
+
+// A predicate that checks the test name of a TestInfo against a known
+// value.
+//
+// This is used for implementation of the TestCase class only.  We put
+// it in the anonymous namespace to prevent polluting the outer
+// namespace.
+//
+// TestNameIs is copyable.
+class TestNameIs {
+ public:
+  // Constructor.
+  //
+  // TestNameIs has NO default constructor.
+  explicit TestNameIs(const char* name)
+      : name_(name) {}
+
+  // Returns true iff the test name of test_info matches name_.
+  bool operator()(const TestInfo * test_info) const {
+    return test_info && test_info->name() == name_;
+  }
+
+ private:
+  std::string name_;
+};
+
+}  // namespace
+
+namespace internal {
+
+// This method expands all parameterized tests registered with macros TEST_P
+// and INSTANTIATE_TEST_CASE_P into regular tests and registers those.
+// This will be done just once during the program runtime.
+void UnitTestImpl::RegisterParameterizedTests() {
+#if GTEST_HAS_PARAM_TEST
+  if (!parameterized_tests_registered_) {
+    parameterized_test_registry_.RegisterTests();
+    parameterized_tests_registered_ = true;
+  }
+#endif
+}
+
+}  // namespace internal
+
+// Creates the test object, runs it, records its result, and then
+// deletes it.
+void TestInfo::Run() {
+  if (!should_run_) return;
+
+  // Tells UnitTest where to store test result.
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  impl->set_current_test_info(this);
+
+  TestEventListener* repeater = UnitTest::GetInstance()->listeners().repeater();
+
+  // Notifies the unit test event listeners that a test is about to start.
+  repeater->OnTestStart(*this);
+
+  const TimeInMillis start = internal::GetTimeInMillis();
+
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+
+  // Creates the test object.
+  Test* const test = internal::HandleExceptionsInMethodIfSupported(
+      factory_, &internal::TestFactoryBase::CreateTest,
+      "the test fixture's constructor");
+
+  // Runs the test only if the test object was created and its
+  // constructor didn't generate a fatal failure.
+  if ((test != NULL) && !Test::HasFatalFailure()) {
+    // This doesn't throw as all user code that can throw are wrapped into
+    // exception handling code.
+    test->Run();
+  }
+
+  // Deletes the test object.
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      test, &Test::DeleteSelf_, "the test fixture's destructor");
+
+  result_.set_elapsed_time(internal::GetTimeInMillis() - start);
+
+  // Notifies the unit test event listener that a test has just finished.
+  repeater->OnTestEnd(*this);
+
+  // Tells UnitTest to stop associating assertion results to this
+  // test.
+  impl->set_current_test_info(NULL);
+}
+
+// class TestCase
+
+// Gets the number of successful tests in this test case.
+int TestCase::successful_test_count() const {
+  return CountIf(test_info_list_, TestPassed);
+}
+
+// Gets the number of failed tests in this test case.
+int TestCase::failed_test_count() const {
+  return CountIf(test_info_list_, TestFailed);
+}
+
+// Gets the number of disabled tests that will be reported in the XML report.
+int TestCase::reportable_disabled_test_count() const {
+  return CountIf(test_info_list_, TestReportableDisabled);
+}
+
+// Gets the number of disabled tests in this test case.
+int TestCase::disabled_test_count() const {
+  return CountIf(test_info_list_, TestDisabled);
+}
+
+// Gets the number of tests to be printed in the XML report.
+int TestCase::reportable_test_count() const {
+  return CountIf(test_info_list_, TestReportable);
+}
+
+// Get the number of tests in this test case that should run.
+int TestCase::test_to_run_count() const {
+  return CountIf(test_info_list_, ShouldRunTest);
+}
+
+// Gets the number of all tests.
+int TestCase::total_test_count() const {
+  return static_cast<int>(test_info_list_.size());
+}
+
+// Creates a TestCase with the given name.
+//
+// Arguments:
+//
+//   name:         name of the test case
+//   a_type_param: the name of the test case's type parameter, or NULL if
+//                 this is not a typed or a type-parameterized test case.
+//   set_up_tc:    pointer to the function that sets up the test case
+//   tear_down_tc: pointer to the function that tears down the test case
+TestCase::TestCase(const char* a_name, const char* a_type_param,
+                   Test::SetUpTestCaseFunc set_up_tc,
+                   Test::TearDownTestCaseFunc tear_down_tc)
+    : name_(a_name),
+      type_param_(a_type_param ? new std::string(a_type_param) : NULL),
+      set_up_tc_(set_up_tc),
+      tear_down_tc_(tear_down_tc),
+      should_run_(false),
+      elapsed_time_(0) {
+}
+
+// Destructor of TestCase.
+TestCase::~TestCase() {
+  // Deletes every Test in the collection.
+  ForEach(test_info_list_, internal::Delete<TestInfo>);
+}
+
+// Returns the i-th test among all the tests. i can range from 0 to
+// total_test_count() - 1. If i is not in that range, returns NULL.
+const TestInfo* TestCase::GetTestInfo(int i) const {
+  const int index = GetElementOr(test_indices_, i, -1);
+  return index < 0 ? NULL : test_info_list_[index];
+}
+
+// Returns the i-th test among all the tests. i can range from 0 to
+// total_test_count() - 1. If i is not in that range, returns NULL.
+TestInfo* TestCase::GetMutableTestInfo(int i) {
+  const int index = GetElementOr(test_indices_, i, -1);
+  return index < 0 ? NULL : test_info_list_[index];
+}
+
+// Adds a test to this test case.  Will delete the test upon
+// destruction of the TestCase object.
+void TestCase::AddTestInfo(TestInfo * test_info) {
+  test_info_list_.push_back(test_info);
+  test_indices_.push_back(static_cast<int>(test_indices_.size()));
+}
+
+// Runs every test in this TestCase.
+void TestCase::Run() {
+  if (!should_run_) return;
+
+  internal::UnitTestImpl* const impl = internal::GetUnitTestImpl();
+  impl->set_current_test_case(this);
+
+  TestEventListener* repeater = UnitTest::GetInstance()->listeners().repeater();
+
+  repeater->OnTestCaseStart(*this);
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      this, &TestCase::RunSetUpTestCase, "SetUpTestCase()");
+
+  const internal::TimeInMillis start = internal::GetTimeInMillis();
+  for (int i = 0; i < total_test_count(); i++) {
+    GetMutableTestInfo(i)->Run();
+  }
+  elapsed_time_ = internal::GetTimeInMillis() - start;
+
+  impl->os_stack_trace_getter()->UponLeavingGTest();
+  internal::HandleExceptionsInMethodIfSupported(
+      this, &TestCase::RunTearDownTestCase, "TearDownTestCase()");
+
+  repeater->OnTestCaseEnd(*this);
+  impl->set_current_test_case(NULL);
+}
+
+// Clears the results of all tests in this test case.
+void TestCase::ClearResult() {
+  ad_hoc_test_result_.Clear();
+  ForEach(test_info_list_, TestInfo::ClearTestResult);
+}
+
+// Shuffles the tests in this test case.
+void TestCase::ShuffleTests(internal::Random* random) {
+  Shuffle(random, &test_indices_);
+}
+
+// Restores the test order to before the first shuffle.
+void TestCase::UnshuffleTests() {
+  for (size_t i = 0; i < test_indices_.size(); i++) {
+    test_indices_[i] = static_cast<int>(i);
+  }
+}
+
+// Formats a countable noun.  Depending on its quantity, either the
+// singular form or the plural form is used. e.g.
+//
+// FormatCountableNoun(1, "formula", "formuli") returns "1 formula".
+// FormatCountableNoun(5, "book", "books") returns "5 books".
+static std::string FormatCountableNoun(int count,
+                                       const char * singular_form,
+                                       const char * plural_form) {
+  return internal::StreamableToString(count) + " " +
+      (count == 1 ? singular_form : plural_form);
+}
+
+// Formats the count of tests.
+static std::string FormatTestCount(int test_count) {
+  return FormatCountableNoun(test_count, "test", "tests");
+}
+
+// Formats the count of test cases.
+static std::string FormatTestCaseCount(int test_case_count) {
+  return FormatCountableNoun(test_case_count, "test case", "test cases");
+}
+
+// Converts a TestPartResult::Type enum to human-friendly string
+// representation.  Both kNonFatalFailure and kFatalFailure are translated
+// to "Failure", as the user usually doesn't care about the difference
+// between the two when viewing the test result.
+static const char * TestPartResultTypeToString(TestPartResult::Type type) {
+  switch (type) {
+    case TestPartResult::kSuccess:
+      return "Success";
+
+    case TestPartResult::kNonFatalFailure:
+    case TestPartResult::kFatalFailure:
+#ifdef _MSC_VER
+      return "error: ";
+#else
+      return "Failure\n";
+#endif
+    default:
+      return "Unknown result type";
+  }
+}
+
+namespace internal {
+
+// Prints a TestPartResult to an std::string.
+static std::string PrintTestPartResultToString(
+    const TestPartResult& test_part_result) {
+  return (Message()
+          << internal::FormatFileLocation(test_part_result.file_name(),
+                                          test_part_result.line_number())
+          << " " << TestPartResultTypeToString(test_part_result.type())
+          << test_part_result.message()).GetString();
+}
+
+// Prints a TestPartResult.
+static void PrintTestPartResult(const TestPartResult& test_part_result) {
+  const std::string& result =
+      PrintTestPartResultToString(test_part_result);
+  printf("%s\n", result.c_str());
+  fflush(stdout);
+  // If the test program runs in Visual Studio or a debugger, the
+  // following statements add the test part result message to the Output
+  // window such that the user can double-click on it to jump to the
+  // corresponding source code location; otherwise they do nothing.
+#if GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+  // We don't call OutputDebugString*() on Windows Mobile, as printing
+  // to stdout is done by OutputDebugString() there already - we don't
+  // want the same message printed twice.
+  ::OutputDebugStringA(result.c_str());
+  ::OutputDebugStringA("\n");
+#endif
+}
+
+// class PrettyUnitTestResultPrinter
+
+enum GTestColor {
+  COLOR_DEFAULT,
+  COLOR_RED,
+  COLOR_GREEN,
+  COLOR_YELLOW
+};
+
+#if GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE && \
+    !GTEST_OS_WINDOWS_PHONE && !GTEST_OS_WINDOWS_RT
+
+// Returns the character attribute for the given color.
+WORD GetColorAttribute(GTestColor color) {
+  switch (color) {
+    case COLOR_RED:    return FOREGROUND_RED;
+    case COLOR_GREEN:  return FOREGROUND_GREEN;
+    case COLOR_YELLOW: return FOREGROUND_RED | FOREGROUND_GREEN;
+    default:           return 0;
+  }
+}
+
+#else
+
+// Returns the ANSI color code for the given color.  COLOR_DEFAULT is
+// an invalid input.
+const char* GetAnsiColorCode(GTestColor color) {
+  switch (color) {
+    case COLOR_RED:     return "1";
+    case COLOR_GREEN:   return "2";
+    case COLOR_YELLOW:  return "3";
+    default:            return NULL;
+  };
+}
+
+#endif  // GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+
+// Returns true iff Google Test should use colors in the output.
+bool ShouldUseColor(bool stdout_is_tty) {
+  const char* const gtest_color = GTEST_FLAG(color).c_str();
+
+  if (String::CaseInsensitiveCStringEquals(gtest_color, "auto")) {
+#if GTEST_OS_WINDOWS
+    // On Windows the TERM variable is usually not set, but the
+    // console there does support colors.
+    return stdout_is_tty;
+#else
+    // On non-Windows platforms, we rely on the TERM variable.
+    const char* const term = posix::GetEnv("TERM");
+    const bool term_supports_color =
+        String::CStringEquals(term, "xterm") ||
+        String::CStringEquals(term, "xterm-color") ||
+        String::CStringEquals(term, "xterm-256color") ||
+        String::CStringEquals(term, "screen") ||
+        String::CStringEquals(term, "screen-256color") ||
+        String::CStringEquals(term, "tmux") ||
+        String::CStringEquals(term, "tmux-256color") ||
+        String::CStringEquals(term, "rxvt-unicode") ||
+        String::CStringEquals(term, "rxvt-unicode-256color") ||
+        String::CStringEquals(term, "linux") ||
+        String::CStringEquals(term, "cygwin");
+    return stdout_is_tty && term_supports_color;
+#endif  // GTEST_OS_WINDOWS
+  }
+
+  return String::CaseInsensitiveCStringEquals(gtest_color, "yes") ||
+      String::CaseInsensitiveCStringEquals(gtest_color, "true") ||
+      String::CaseInsensitiveCStringEquals(gtest_color, "t") ||
+      String::CStringEquals(gtest_color, "1");
+  // We take "yes", "true", "t", and "1" as meaning "yes".  If the
+  // value is neither one of these nor "auto", we treat it as "no" to
+  // be conservative.
+}
+
+// Helpers for printing colored strings to stdout. Note that on Windows, we
+// cannot simply emit special characters and have the terminal change colors.
+// This routine must actually emit the characters rather than return a string
+// that would be colored when printed, as can be done on Linux.
+void ColoredPrintf(GTestColor color, const char* fmt, ...) {
+  va_list args;
+  va_start(args, fmt);
+
+#if GTEST_OS_WINDOWS_MOBILE || GTEST_OS_SYMBIAN || GTEST_OS_ZOS || \
+    GTEST_OS_IOS || GTEST_OS_WINDOWS_PHONE || GTEST_OS_WINDOWS_RT
+  const bool use_color = AlwaysFalse();
+#else
+  static const bool in_color_mode =
+      ShouldUseColor(posix::IsATTY(posix::FileNo(stdout)) != 0);
+  const bool use_color = in_color_mode && (color != COLOR_DEFAULT);
+#endif  // GTEST_OS_WINDOWS_MOBILE || GTEST_OS_SYMBIAN || GTEST_OS_ZOS
+  // The '!= 0' comparison is necessary to satisfy MSVC 7.1.
+
+  if (!use_color) {
+    vprintf(fmt, args);
+    va_end(args);
+    return;
+  }
+
+#if GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE && \
+    !GTEST_OS_WINDOWS_PHONE && !GTEST_OS_WINDOWS_RT
+  const HANDLE stdout_handle = GetStdHandle(STD_OUTPUT_HANDLE);
+
+  // Gets the current text color.
+  CONSOLE_SCREEN_BUFFER_INFO buffer_info;
+  GetConsoleScreenBufferInfo(stdout_handle, &buffer_info);
+  const WORD old_color_attrs = buffer_info.wAttributes;
+
+  // We need to flush the stream buffers into the console before each
+  // SetConsoleTextAttribute call lest it affect the text that is already
+  // printed but has not yet reached the console.
+  fflush(stdout);
+  SetConsoleTextAttribute(stdout_handle,
+                          GetColorAttribute(color) | FOREGROUND_INTENSITY);
+  vprintf(fmt, args);
+
+  fflush(stdout);
+  // Restores the text color.
+  SetConsoleTextAttribute(stdout_handle, old_color_attrs);
+#else
+  printf("\033[0;3%sm", GetAnsiColorCode(color));
+  vprintf(fmt, args);
+  printf("\033[m");  // Resets the terminal to default.
+#endif  // GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_MOBILE
+  va_end(args);
+}
+
+// Text printed in Google Test's text output and --gunit_list_tests
+// output to label the type parameter and value parameter for a test.
+static const char kTypeParamLabel[] = "TypeParam";
+static const char kValueParamLabel[] = "GetParam()";
+
+void PrintFullTestCommentIfPresent(const TestInfo& test_info) {
+  const char* const type_param = test_info.type_param();
+  const char* const value_param = test_info.value_param();
+
+  if (type_param != NULL || value_param != NULL) {
+    printf(", where ");
+    if (type_param != NULL) {
+      printf("%s = %s", kTypeParamLabel, type_param);
+      if (value_param != NULL)
+        printf(" and ");
+    }
+    if (value_param != NULL) {
+      printf("%s = %s", kValueParamLabel, value_param);
+    }
+  }
+}
+
+// This class implements the TestEventListener interface.
+//
+// Class PrettyUnitTestResultPrinter is copyable.
+class PrettyUnitTestResultPrinter : public TestEventListener {
+ public:
+  PrettyUnitTestResultPrinter() {}
+  static void PrintTestName(const char * test_case, const char * test) {
+    printf("%s.%s", test_case, test);
+  }
+
+  // The following methods override what's in the TestEventListener class.
+  virtual void OnTestProgramStart(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationStart(const UnitTest& unit_test, int iteration);
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestCaseStart(const TestCase& test_case);
+  virtual void OnTestStart(const TestInfo& test_info);
+  virtual void OnTestPartResult(const TestPartResult& result);
+  virtual void OnTestEnd(const TestInfo& test_info);
+  virtual void OnTestCaseEnd(const TestCase& test_case);
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& /*unit_test*/) {}
+  virtual void OnTestIterationEnd(const UnitTest& unit_test, int iteration);
+  virtual void OnTestProgramEnd(const UnitTest& /*unit_test*/) {}
+
+ private:
+  static void PrintFailedTests(const UnitTest& unit_test);
+};
+
+  // Fired before each iteration of tests starts.
+void PrettyUnitTestResultPrinter::OnTestIterationStart(
+    const UnitTest& unit_test, int iteration) {
+  if (GTEST_FLAG(repeat) != 1)
+    printf("\nRepeating all tests (iteration %d) . . .\n\n", iteration + 1);
+
+  const char* const filter = GTEST_FLAG(filter).c_str();
+
+  // Prints the filter if it's not *.  This reminds the user that some
+  // tests may be skipped.
+  if (!String::CStringEquals(filter, kUniversalFilter)) {
+    ColoredPrintf(COLOR_YELLOW,
+                  "Note: %s filter = %s\n", GTEST_NAME_, filter);
+  }
+
+  if (internal::ShouldShard(kTestTotalShards, kTestShardIndex, false)) {
+    const Int32 shard_index = Int32FromEnvOrDie(kTestShardIndex, -1);
+    ColoredPrintf(COLOR_YELLOW,
+                  "Note: This is test shard %d of %s.\n",
+                  static_cast<int>(shard_index) + 1,
+                  internal::posix::GetEnv(kTestTotalShards));
+  }
+
+  if (GTEST_FLAG(shuffle)) {
+    ColoredPrintf(COLOR_YELLOW,
+                  "Note: Randomizing tests' orders with a seed of %d .\n",
+                  unit_test.random_seed());
+  }
+
+  ColoredPrintf(COLOR_GREEN,  "[==========] ");
+  printf("Running %s from %s.\n",
+         FormatTestCount(unit_test.test_to_run_count()).c_str(),
+         FormatTestCaseCount(unit_test.test_case_to_run_count()).c_str());
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnEnvironmentsSetUpStart(
+    const UnitTest& /*unit_test*/) {
+  ColoredPrintf(COLOR_GREEN,  "[----------] ");
+  printf("Global test environment set-up.\n");
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestCaseStart(const TestCase& test_case) {
+  const std::string counts =
+      FormatCountableNoun(test_case.test_to_run_count(), "test", "tests");
+  ColoredPrintf(COLOR_GREEN, "[----------] ");
+  printf("%s from %s", counts.c_str(), test_case.name());
+  if (test_case.type_param() == NULL) {
+    printf("\n");
+  } else {
+    printf(", where %s = %s\n", kTypeParamLabel, test_case.type_param());
+  }
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestStart(const TestInfo& test_info) {
+  ColoredPrintf(COLOR_GREEN,  "[ RUN      ] ");
+  PrintTestName(test_info.test_case_name(), test_info.name());
+  printf("\n");
+  fflush(stdout);
+}
+
+// Called after an assertion failure.
+void PrettyUnitTestResultPrinter::OnTestPartResult(
+    const TestPartResult& result) {
+  // If the test part succeeded, we don't need to do anything.
+  if (result.type() == TestPartResult::kSuccess)
+    return;
+
+  // Print failure message from the assertion (e.g. expected this and got that).
+  PrintTestPartResult(result);
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestEnd(const TestInfo& test_info) {
+  if (test_info.result()->Passed()) {
+    ColoredPrintf(COLOR_GREEN, "[       OK ] ");
+  } else {
+    ColoredPrintf(COLOR_RED, "[  FAILED  ] ");
+  }
+  PrintTestName(test_info.test_case_name(), test_info.name());
+  if (test_info.result()->Failed())
+    PrintFullTestCommentIfPresent(test_info);
+
+  if (GTEST_FLAG(print_time)) {
+    printf(" (%s ms)\n", internal::StreamableToString(
+           test_info.result()->elapsed_time()).c_str());
+  } else {
+    printf("\n");
+  }
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnTestCaseEnd(const TestCase& test_case) {
+  if (!GTEST_FLAG(print_time)) return;
+
+  const std::string counts =
+      FormatCountableNoun(test_case.test_to_run_count(), "test", "tests");
+  ColoredPrintf(COLOR_GREEN, "[----------] ");
+  printf("%s from %s (%s ms total)\n\n",
+         counts.c_str(), test_case.name(),
+         internal::StreamableToString(test_case.elapsed_time()).c_str());
+  fflush(stdout);
+}
+
+void PrettyUnitTestResultPrinter::OnEnvironmentsTearDownStart(
+    const UnitTest& /*unit_test*/) {
+  ColoredPrintf(COLOR_GREEN,  "[----------] ");
+  printf("Global test environment tear-down\n");
+  fflush(stdout);
+}
+
+// Internal helper for printing the list of failed tests.
+void PrettyUnitTestResultPrinter::PrintFailedTests(const UnitTest& unit_test) {
+  const int failed_test_count = unit_test.failed_test_count();
+  if (failed_test_count == 0) {
+    return;
+  }
+
+  for (int i = 0; i < unit_test.total_test_case_count(); ++i) {
+    const TestCase& test_case = *unit_test.GetTestCase(i);
+    if (!test_case.should_run() || (test_case.failed_test_count() == 0)) {
+      continue;
+    }
+    for (int j = 0; j < test_case.total_test_count(); ++j) {
+      const TestInfo& test_info = *test_case.GetTestInfo(j);
+      if (!test_info.should_run() || test_info.result()->Passed()) {
+        continue;
+      }
+      ColoredPrintf(COLOR_RED, "[  FAILED  ] ");
+      printf("%s.%s", test_case.name(), test_info.name());
+      PrintFullTestCommentIfPresent(test_info);
+      printf("\n");
+    }
+  }
+}
+
+void PrettyUnitTestResultPrinter::OnTestIterationEnd(const UnitTest& unit_test,
+                                                     int /*iteration*/) {
+  ColoredPrintf(COLOR_GREEN,  "[==========] ");
+  printf("%s from %s ran.",
+         FormatTestCount(unit_test.test_to_run_count()).c_str(),
+         FormatTestCaseCount(unit_test.test_case_to_run_count()).c_str());
+  if (GTEST_FLAG(print_time)) {
+    printf(" (%s ms total)",
+           internal::StreamableToString(unit_test.elapsed_time()).c_str());
+  }
+  printf("\n");
+  ColoredPrintf(COLOR_GREEN,  "[  PASSED  ] ");
+  printf("%s.\n", FormatTestCount(unit_test.successful_test_count()).c_str());
+
+  int num_failures = unit_test.failed_test_count();
+  if (!unit_test.Passed()) {
+    const int failed_test_count = unit_test.failed_test_count();
+    ColoredPrintf(COLOR_RED,  "[  FAILED  ] ");
+    printf("%s, listed below:\n", FormatTestCount(failed_test_count).c_str());
+    PrintFailedTests(unit_test);
+    printf("\n%2d FAILED %s\n", num_failures,
+                        num_failures == 1 ? "TEST" : "TESTS");
+  }
+
+  int num_disabled = unit_test.reportable_disabled_test_count();
+  if (num_disabled && !GTEST_FLAG(also_run_disabled_tests)) {
+    if (!num_failures) {
+      printf("\n");  // Add a spacer if no FAILURE banner is displayed.
+    }
+    ColoredPrintf(COLOR_YELLOW,
+                  "  YOU HAVE %d DISABLED %s\n\n",
+                  num_disabled,
+                  num_disabled == 1 ? "TEST" : "TESTS");
+  }
+  // Ensure that Google Test output is printed before, e.g., heapchecker output.
+  fflush(stdout);
+}
+
+// End PrettyUnitTestResultPrinter
+
+// class TestEventRepeater
+//
+// This class forwards events to other event listeners.
+class TestEventRepeater : public TestEventListener {
+ public:
+  TestEventRepeater() : forwarding_enabled_(true) {}
+  virtual ~TestEventRepeater();
+  void Append(TestEventListener *listener);
+  TestEventListener* Release(TestEventListener* listener);
+
+  // Controls whether events will be forwarded to listeners_. Set to false
+  // in death test child processes.
+  bool forwarding_enabled() const { return forwarding_enabled_; }
+  void set_forwarding_enabled(bool enable) { forwarding_enabled_ = enable; }
+
+  virtual void OnTestProgramStart(const UnitTest& unit_test);
+  virtual void OnTestIterationStart(const UnitTest& unit_test, int iteration);
+  virtual void OnEnvironmentsSetUpStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsSetUpEnd(const UnitTest& unit_test);
+  virtual void OnTestCaseStart(const TestCase& test_case);
+  virtual void OnTestStart(const TestInfo& test_info);
+  virtual void OnTestPartResult(const TestPartResult& result);
+  virtual void OnTestEnd(const TestInfo& test_info);
+  virtual void OnTestCaseEnd(const TestCase& test_case);
+  virtual void OnEnvironmentsTearDownStart(const UnitTest& unit_test);
+  virtual void OnEnvironmentsTearDownEnd(const UnitTest& unit_test);
+  virtual void OnTestIterationEnd(const UnitTest& unit_test, int iteration);
+  virtual void OnTestProgramEnd(const UnitTest& unit_test);
+
+ private:
+  // Controls whether events will be forwarded to listeners_. Set to false
+  // in death test child processes.
+  bool forwarding_enabled_;
+  // The list of listeners that receive events.
+  std::vector<TestEventListener*> listeners_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(TestEventRepeater);
+};
+
+TestEventRepeater::~TestEventRepeater() {
+  ForEach(listeners_, Delete<TestEventListener>);
+}
+
+void TestEventRepeater::Append(TestEventListener *listener) {
+  listeners_.push_back(listener);
+}
+
+// TODO(vladl@google.com): Factor the search functionality into Vector::Find.
+TestEventListener* TestEventRepeater::Release(TestEventListener *listener) {
+  for (size_t i = 0; i < listeners_.size(); ++i) {
+    if (listeners_[i] == listener) {
+      listeners_.erase(listeners_.begin() + i);
+      return listener;
+    }
+  }
+
+  return NULL;
+}
+
+// Since most methods are very similar, use macros to reduce boilerplate.
+// This defines a member that forwards the call to all listeners.
+#define GTEST_REPEATER_METHOD_(Name, Type) \
+void TestEventRepeater::Name(const Type& parameter) { \
+  if (forwarding_enabled_) { \
+    for (size_t i = 0; i < listeners_.size(); i++) { \
+      listeners_[i]->Name(parameter); \
+    } \
+  } \
+}
+// This defines a member that forwards the call to all listeners in reverse
+// order.
+#define GTEST_REVERSE_REPEATER_METHOD_(Name, Type) \
+void TestEventRepeater::Name(const Type& parameter) { \
+  if (forwarding_enabled_) { \
+    for (int i = static_cast<int>(listeners_.size()) - 1; i >= 0; i--) { \
+      listeners_[i]->Name(parameter); \
+    } \
+  } \
+}
+
+GTEST_REPEATER_METHOD_(OnTestProgramStart, UnitTest)
+GTEST_REPEATER_METHOD_(OnEnvironmentsSetUpStart, UnitTest)
+GTEST_REPEATER_METHOD_(OnTestCaseStart, TestCase)
+GTEST_REPEATER_METHOD_(OnTestStart, TestInfo)
+GTEST_REPEATER_METHOD_(OnTestPartResult, TestPartResult)
+GTEST_REPEATER_METHOD_(OnEnvironmentsTearDownStart, UnitTest)
+GTEST_REVERSE_REPEATER_METHOD_(OnEnvironmentsSetUpEnd, UnitTest)
+GTEST_REVERSE_REPEATER_METHOD_(OnEnvironmentsTearDownEnd, UnitTest)
+GTEST_REVERSE_REPEATER_METHOD_(OnTestEnd, TestInfo)
+GTEST_REVERSE_REPEATER_METHOD_(OnTestCaseEnd, TestCase)
+GTEST_REVERSE_REPEATER_METHOD_(OnTestProgramEnd, UnitTest)
+
+#undef GTEST_REPEATER_METHOD_
+#undef GTEST_REVERSE_REPEATER_METHOD_
+
+void TestEventRepeater::OnTestIterationStart(const UnitTest& unit_test,
+                                             int iteration) {
+  if (forwarding_enabled_) {
+    for (size_t i = 0; i < listeners_.size(); i++) {
+      listeners_[i]->OnTestIterationStart(unit_test, iteration);
+    }
+  }
+}
+
+void TestEventRepeater::OnTestIterationEnd(const UnitTest& unit_test,
+                                           int iteration) {
+  if (forwarding_enabled_) {
+    for (int i = static_cast<int>(listeners_.size()) - 1; i >= 0; i--) {
+      listeners_[i]->OnTestIterationEnd(unit_test, iteration);
+    }
+  }
+}
+
+// End TestEventRepeater
+
+// This class generates an XML output file.
+class XmlUnitTestResultPrinter : public EmptyTestEventListener {
+ public:
+  explicit XmlUnitTestResultPrinter(const char* output_file);
+
+  virtual void OnTestIterationEnd(const UnitTest& unit_test, int iteration);
+
+ private:
+  // Is c a whitespace character that is normalized to a space character
+  // when it appears in an XML attribute value?
+  static bool IsNormalizableWhitespace(char c) {
+    return c == 0x9 || c == 0xA || c == 0xD;
+  }
+
+  // May c appear in a well-formed XML document?
+  static bool IsValidXmlCharacter(char c) {
+    return IsNormalizableWhitespace(c) || c >= 0x20;
+  }
+
+  // Returns an XML-escaped copy of the input string str.  If
+  // is_attribute is true, the text is meant to appear as an attribute
+  // value, and normalizable whitespace is preserved by replacing it
+  // with character references.
+  static std::string EscapeXml(const std::string& str, bool is_attribute);
+
+  // Returns the given string with all characters invalid in XML removed.
+  static std::string RemoveInvalidXmlCharacters(const std::string& str);
+
+  // Convenience wrapper around EscapeXml when str is an attribute value.
+  static std::string EscapeXmlAttribute(const std::string& str) {
+    return EscapeXml(str, true);
+  }
+
+  // Convenience wrapper around EscapeXml when str is not an attribute value.
+  static std::string EscapeXmlText(const char* str) {
+    return EscapeXml(str, false);
+  }
+
+  // Verifies that the given attribute belongs to the given element and
+  // streams the attribute as XML.
+  static void OutputXmlAttribute(std::ostream* stream,
+                                 const std::string& element_name,
+                                 const std::string& name,
+                                 const std::string& value);
+
+  // Streams an XML CDATA section, escaping invalid CDATA sequences as needed.
+  static void OutputXmlCDataSection(::std::ostream* stream, const char* data);
+
+  // Streams an XML representation of a TestInfo object.
+  static void OutputXmlTestInfo(::std::ostream* stream,
+                                const char* test_case_name,
+                                const TestInfo& test_info);
+
+  // Prints an XML representation of a TestCase object
+  static void PrintXmlTestCase(::std::ostream* stream,
+                               const TestCase& test_case);
+
+  // Prints an XML summary of unit_test to output stream out.
+  static void PrintXmlUnitTest(::std::ostream* stream,
+                               const UnitTest& unit_test);
+
+  // Produces a string representing the test properties in a result as space
+  // delimited XML attributes based on the property key="value" pairs.
+  // When the std::string is not empty, it includes a space at the beginning,
+  // to delimit this attribute from prior attributes.
+  static std::string TestPropertiesAsXmlAttributes(const TestResult& result);
+
+  // The output file.
+  const std::string output_file_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(XmlUnitTestResultPrinter);
+};
+
+// Creates a new XmlUnitTestResultPrinter.
+XmlUnitTestResultPrinter::XmlUnitTestResultPrinter(const char* output_file)
+    : output_file_(output_file) {
+  if (output_file_.c_str() == NULL || output_file_.empty()) {
+    fprintf(stderr, "XML output file may not be null\n");
+    fflush(stderr);
+    exit(EXIT_FAILURE);
+  }
+}
+
+// Called after the unit test ends.
+void XmlUnitTestResultPrinter::OnTestIterationEnd(const UnitTest& unit_test,
+                                                  int /*iteration*/) {
+  FILE* xmlout = NULL;
+  FilePath output_file(output_file_);
+  FilePath output_dir(output_file.RemoveFileName());
+
+  if (output_dir.CreateDirectoriesRecursively()) {
+    xmlout = posix::FOpen(output_file_.c_str(), "w");
+  }
+  if (xmlout == NULL) {
+    // TODO(wan): report the reason of the failure.
+    //
+    // We don't do it for now as:
+    //
+    //   1. There is no urgent need for it.
+    //   2. It's a bit involved to make the errno variable thread-safe on
+    //      all three operating systems (Linux, Windows, and Mac OS).
+    //   3. To interpret the meaning of errno in a thread-safe way,
+    //      we need the strerror_r() function, which is not available on
+    //      Windows.
+    fprintf(stderr,
+            "Unable to open file \"%s\"\n",
+            output_file_.c_str());
+    fflush(stderr);
+    exit(EXIT_FAILURE);
+  }
+  std::stringstream stream;
+  PrintXmlUnitTest(&stream, unit_test);
+  fprintf(xmlout, "%s", StringStreamToString(&stream).c_str());
+  fclose(xmlout);
+}
+
+// Returns an XML-escaped copy of the input string str.  If is_attribute
+// is true, the text is meant to appear as an attribute value, and
+// normalizable whitespace is preserved by replacing it with character
+// references.
+//
+// Invalid XML characters in str, if any, are stripped from the output.
+// It is expected that most, if not all, of the text processed by this
+// module will consist of ordinary English text.
+// If this module is ever modified to produce version 1.1 XML output,
+// most invalid characters can be retained using character references.
+// TODO(wan): It might be nice to have a minimally invasive, human-readable
+// escaping scheme for invalid characters, rather than dropping them.
+std::string XmlUnitTestResultPrinter::EscapeXml(
+    const std::string& str, bool is_attribute) {
+  Message m;
+
+  for (size_t i = 0; i < str.size(); ++i) {
+    const char ch = str[i];
+    switch (ch) {
+      case '<':
+        m << "&lt;";
+        break;
+      case '>':
+        m << "&gt;";
+        break;
+      case '&':
+        m << "&amp;";
+        break;
+      case '\'':
+        if (is_attribute)
+          m << "&apos;";
+        else
+          m << '\'';
+        break;
+      case '"':
+        if (is_attribute)
+          m << "&quot;";
+        else
+          m << '"';
+        break;
+      default:
+        if (IsValidXmlCharacter(ch)) {
+          if (is_attribute && IsNormalizableWhitespace(ch))
+            m << "&#x" << String::FormatByte(static_cast<unsigned char>(ch))
+              << ";";
+          else
+            m << ch;
+        }
+        break;
+    }
+  }
+
+  return m.GetString();
+}
+
+// Returns the given string with all characters invalid in XML removed.
+// Currently invalid characters are dropped from the string. An
+// alternative is to replace them with certain characters such as . or ?.
+std::string XmlUnitTestResultPrinter::RemoveInvalidXmlCharacters(
+    const std::string& str) {
+  std::string output;
+  output.reserve(str.size());
+  for (std::string::const_iterator it = str.begin(); it != str.end(); ++it)
+    if (IsValidXmlCharacter(*it))
+      output.push_back(*it);
+
+  return output;
+}
+
+// The following routines generate an XML representation of a UnitTest
+// object.
+//
+// This is how Google Test concepts map to the DTD:
+//
+// <testsuites name="AllTests">        <-- corresponds to a UnitTest object
+//   <testsuite name="testcase-name">  <-- corresponds to a TestCase object
+//     <testcase name="test-name">     <-- corresponds to a TestInfo object
+//       <failure message="...">...</failure>
+//       <failure message="...">...</failure>
+//       <failure message="...">...</failure>
+//                                     <-- individual assertion failures
+//     </testcase>
+//   </testsuite>
+// </testsuites>
+
+// Formats the given time in milliseconds as seconds.
+std::string FormatTimeInMillisAsSeconds(TimeInMillis ms) {
+  ::std::stringstream ss;
+  ss << (static_cast<double>(ms) * 1e-3);
+  return ss.str();
+}
+
+static bool PortableLocaltime(time_t seconds, struct tm* out) {
+#if defined(_MSC_VER)
+  return localtime_s(out, &seconds) == 0;
+#elif defined(__MINGW32__) || defined(__MINGW64__)
+  // MINGW <time.h> provides neither localtime_r nor localtime_s, but uses
+  // Windows' localtime(), which has a thread-local tm buffer.
+  struct tm* tm_ptr = localtime(&seconds);  // NOLINT
+  if (tm_ptr == NULL)
+    return false;
+  *out = *tm_ptr;
+  return true;
+#else
+  return localtime_r(&seconds, out) != NULL;
+#endif
+}
+
+// Converts the given epoch time in milliseconds to a date string in the ISO
+// 8601 format, without the timezone information.
+std::string FormatEpochTimeInMillisAsIso8601(TimeInMillis ms) {
+  struct tm time_struct;
+  if (!PortableLocaltime(static_cast<time_t>(ms / 1000), &time_struct))
+    return "";
+  // YYYY-MM-DDThh:mm:ss
+  return StreamableToString(time_struct.tm_year + 1900) + "-" +
+      String::FormatIntWidth2(time_struct.tm_mon + 1) + "-" +
+      String::FormatIntWidth2(time_struct.tm_mday) + "T" +
+      String::FormatIntWidth2(time_struct.tm_hour) + ":" +
+      String::FormatIntWidth2(time_struct.tm_min) + ":" +
+      String::FormatIntWidth2(time_struct.tm_sec);
+}
+
+// Streams an XML CDATA section, escaping invalid CDATA sequences as needed.
+void XmlUnitTestResultPrinter::OutputXmlCDataSection(::std::ostream* stream,
+                                                     const char* data) {
+  const char* segment = data;
+  *stream << "<![CDATA[";
+  for (;;) {
+    const char* const next_segment = strstr(segment, "]]>");
+    if (next_segment != NULL) {
+      stream->write(
+          segment, static_cast<std::streamsize>(next_segment - segment));
+      *stream << "]]>]]&gt;<![CDATA[";
+      segment = next_segment + strlen("]]>");
+    } else {
+      *stream << segment;
+      break;
+    }
+  }
+  *stream << "]]>";
+}
+
+void XmlUnitTestResultPrinter::OutputXmlAttribute(
+    std::ostream* stream,
+    const std::string& element_name,
+    const std::string& name,
+    const std::string& value) {
+  const std::vector<std::string>& allowed_names =
+      GetReservedAttributesForElement(element_name);
+
+  GTEST_CHECK_(std::find(allowed_names.begin(), allowed_names.end(), name) !=
+                   allowed_names.end())
+      << "Attribute " << name << " is not allowed for element <" << element_name
+      << ">.";
+
+  *stream << " " << name << "=\"" << EscapeXmlAttribute(value) << "\"";
+}
+
+// Prints an XML representation of a TestInfo object.
+// TODO(wan): There is also value in printing properties with the plain printer.
+void XmlUnitTestResultPrinter::OutputXmlTestInfo(::std::ostream* stream,
+                                                 const char* test_case_name,
+                                                 const TestInfo& test_info) {
+  const TestResult& result = *test_info.result();
+  const std::string kTestcase = "testcase";
+
+  *stream << "    <testcase";
+  OutputXmlAttribute(stream, kTestcase, "name", test_info.name());
+
+  if (test_info.value_param() != NULL) {
+    OutputXmlAttribute(stream, kTestcase, "value_param",
+                       test_info.value_param());
+  }
+  if (test_info.type_param() != NULL) {
+    OutputXmlAttribute(stream, kTestcase, "type_param", test_info.type_param());
+  }
+
+  OutputXmlAttribute(stream, kTestcase, "status",
+                     test_info.should_run() ? "run" : "notrun");
+  OutputXmlAttribute(stream, kTestcase, "time",
+                     FormatTimeInMillisAsSeconds(result.elapsed_time()));
+  OutputXmlAttribute(stream, kTestcase, "classname", test_case_name);
+  *stream << TestPropertiesAsXmlAttributes(result);
+
+  int failures = 0;
+  for (int i = 0; i < result.total_part_count(); ++i) {
+    const TestPartResult& part = result.GetTestPartResult(i);
+    if (part.failed()) {
+      if (++failures == 1) {
+        *stream << ">\n";
+      }
+      const string location = internal::FormatCompilerIndependentFileLocation(
+          part.file_name(), part.line_number());
+      const string summary = location + "\n" + part.summary();
+      *stream << "      <failure message=\""
+              << EscapeXmlAttribute(summary.c_str())
+              << "\" type=\"\">";
+      const string detail = location + "\n" + part.message();
+      OutputXmlCDataSection(stream, RemoveInvalidXmlCharacters(detail).c_str());
+      *stream << "</failure>\n";
+    }
+  }
+
+  if (failures == 0)
+    *stream << " />\n";
+  else
+    *stream << "    </testcase>\n";
+}
+
+// Prints an XML representation of a TestCase object
+void XmlUnitTestResultPrinter::PrintXmlTestCase(std::ostream* stream,
+                                                const TestCase& test_case) {
+  const std::string kTestsuite = "testsuite";
+  *stream << "  <" << kTestsuite;
+  OutputXmlAttribute(stream, kTestsuite, "name", test_case.name());
+  OutputXmlAttribute(stream, kTestsuite, "tests",
+                     StreamableToString(test_case.reportable_test_count()));
+  OutputXmlAttribute(stream, kTestsuite, "failures",
+                     StreamableToString(test_case.failed_test_count()));
+  OutputXmlAttribute(
+      stream, kTestsuite, "disabled",
+      StreamableToString(test_case.reportable_disabled_test_count()));
+  OutputXmlAttribute(stream, kTestsuite, "errors", "0");
+  OutputXmlAttribute(stream, kTestsuite, "time",
+                     FormatTimeInMillisAsSeconds(test_case.elapsed_time()));
+  *stream << TestPropertiesAsXmlAttributes(test_case.ad_hoc_test_result())
+          << ">\n";
+
+  for (int i = 0; i < test_case.total_test_count(); ++i) {
+    if (test_case.GetTestInfo(i)->is_reportable())
+      OutputXmlTestInfo(stream, test_case.name(), *test_case.GetTestInfo(i));
+  }
+  *stream << "  </" << kTestsuite << ">\n";
+}
+
+// Prints an XML summary of unit_test to output stream out.
+void XmlUnitTestResultPrinter::PrintXmlUnitTest(std::ostream* stream,
+                                                const UnitTest& unit_test) {
+  const std::string kTestsuites = "testsuites";
+
+  *stream << "<?xml version=\"1.0\" encoding=\"UTF-8\"?>\n";
+  *stream << "<" << kTestsuites;
+
+  OutputXmlAttribute(stream, kTestsuites, "tests",
+                     StreamableToString(unit_test.reportable_test_count()));
+  OutputXmlAttribute(stream, kTestsuites, "failures",
+                     StreamableToString(unit_test.failed_test_count()));
+  OutputXmlAttribute(
+      stream, kTestsuites, "disabled",
+      StreamableToString(unit_test.reportable_disabled_test_count()));
+  OutputXmlAttribute(stream, kTestsuites, "errors", "0");
+  OutputXmlAttribute(
+      stream, kTestsuites, "timestamp",
+      FormatEpochTimeInMillisAsIso8601(unit_test.start_timestamp()));
+  OutputXmlAttribute(stream, kTestsuites, "time",
+                     FormatTimeInMillisAsSeconds(unit_test.elapsed_time()));
+
+  if (GTEST_FLAG(shuffle)) {
+    OutputXmlAttribute(stream, kTestsuites, "random_seed",
+                       StreamableToString(unit_test.random_seed()));
+  }
+
+  *stream << TestPropertiesAsXmlAttributes(unit_test.ad_hoc_test_result());
+
+  OutputXmlAttribute(stream, kTestsuites, "name", "AllTests");
+  *stream << ">\n";
+
+  for (int i = 0; i < unit_test.total_test_case_count(); ++i) {
+    if (unit_test.GetTestCase(i)->reportable_test_count() > 0)
+      PrintXmlTestCase(stream, *unit_test.GetTestCase(i));
+  }
+  *stream << "</" << kTestsuites << ">\n";
+}
+
+// Produces a string representing the test properties in a result as space
+// delimited XML attributes based on the property key="value" pairs.
+std::string XmlUnitTestResultPrinter::TestPropertiesAsXmlAttributes(
+    const TestResult& result) {
+  Message attributes;
+  for (int i = 0; i < result.test_property_count(); ++i) {
+    const TestProperty& property = result.GetTestProperty(i);
+    attributes << " " << property.key() << "="
+        << "\"" << EscapeXmlAttribute(property.value()) << "\"";
+  }
+  return attributes.GetString();
+}
+
+// End XmlUnitTestResultPrinter
+
+#if GTEST_CAN_STREAM_RESULTS_
+
+// Checks if str contains '=', '&', '%' or '\n' characters. If yes,
+// replaces them by "%xx" where xx is their hexadecimal value. For
+// example, replaces "=" with "%3D".  This algorithm is O(strlen(str))
+// in both time and space -- important as the input str may contain an
+// arbitrarily long test failure message and stack trace.
+string StreamingListener::UrlEncode(const char* str) {
+  string result;
+  result.reserve(strlen(str) + 1);
+  for (char ch = *str; ch != '\0'; ch = *++str) {
+    switch (ch) {
+      case '%':
+      case '=':
+      case '&':
+      case '\n':
+        result.append("%" + String::FormatByte(static_cast<unsigned char>(ch)));
+        break;
+      default:
+        result.push_back(ch);
+        break;
+    }
+  }
+  return result;
+}
+
+void StreamingListener::SocketWriter::MakeConnection() {
+  GTEST_CHECK_(sockfd_ == -1)
+      << "MakeConnection() can't be called when there is already a connection.";
+
+  addrinfo hints;
+  memset(&hints, 0, sizeof(hints));
+  hints.ai_family = AF_UNSPEC;    // To allow both IPv4 and IPv6 addresses.
+  hints.ai_socktype = SOCK_STREAM;
+  addrinfo* servinfo = NULL;
+
+  // Use the getaddrinfo() to get a linked list of IP addresses for
+  // the given host name.
+  const int error_num = getaddrinfo(
+      host_name_.c_str(), port_num_.c_str(), &hints, &servinfo);
+  if (error_num != 0) {
+    GTEST_LOG_(WARNING) << "stream_result_to: getaddrinfo() failed: "
+                        << gai_strerror(error_num);
+  }
+
+  // Loop through all the results and connect to the first we can.
+  for (addrinfo* cur_addr = servinfo; sockfd_ == -1 && cur_addr != NULL;
+       cur_addr = cur_addr->ai_next) {
+    sockfd_ = socket(
+        cur_addr->ai_family, cur_addr->ai_socktype, cur_addr->ai_protocol);
+    if (sockfd_ != -1) {
+      // Connect the client socket to the server socket.
+      if (connect(sockfd_, cur_addr->ai_addr, cur_addr->ai_addrlen) == -1) {
+        close(sockfd_);
+        sockfd_ = -1;
+      }
+    }
+  }
+
+  freeaddrinfo(servinfo);  // all done with this structure
+
+  if (sockfd_ == -1) {
+    GTEST_LOG_(WARNING) << "stream_result_to: failed to connect to "
+                        << host_name_ << ":" << port_num_;
+  }
+}
+
+// End of class Streaming Listener
+#endif  // GTEST_CAN_STREAM_RESULTS__
+
+// Class ScopedTrace
+
+// Pushes the given source file location and message onto a per-thread
+// trace stack maintained by Google Test.
+ScopedTrace::ScopedTrace(const char* file, int line, const Message& message)
+    GTEST_LOCK_EXCLUDED_(&UnitTest::mutex_) {
+  TraceInfo trace;
+  trace.file = file;
+  trace.line = line;
+  trace.message = message.GetString();
+
+  UnitTest::GetInstance()->PushGTestTrace(trace);
+}
+
+// Pops the info pushed by the c'tor.
+ScopedTrace::~ScopedTrace()
+    GTEST_LOCK_EXCLUDED_(&UnitTest::mutex_) {
+  UnitTest::GetInstance()->PopGTestTrace();
+}
+
+
+// class OsStackTraceGetter
+
+const char* const OsStackTraceGetterInterface::kElidedFramesMarker =
+    "... " GTEST_NAME_ " internal frames ...";
+
+string OsStackTraceGetter::CurrentStackTrace(int /*max_depth*/,
+                                             int /*skip_count*/) {
+  return "";
+}
+
+void OsStackTraceGetter::UponLeavingGTest() {}
+
+// A helper class that creates the premature-exit file in its
+// constructor and deletes the file in its destructor.
+class ScopedPrematureExitFile {
+ public:
+  explicit ScopedPrematureExitFile(const char* premature_exit_filepath)
+      : premature_exit_filepath_(premature_exit_filepath) {
+    // If a path to the premature-exit file is specified...
+    if (premature_exit_filepath != NULL && *premature_exit_filepath != '\0') {
+      // create the file with a single "0" character in it.  I/O
+      // errors are ignored as there's nothing better we can do and we
+      // don't want to fail the test because of this.
+      FILE* pfile = posix::FOpen(premature_exit_filepath, "w");
+      fwrite("0", 1, 1, pfile);
+      fclose(pfile);
+    }
+  }
+
+  ~ScopedPrematureExitFile() {
+    if (premature_exit_filepath_ != NULL && *premature_exit_filepath_ != '\0') {
+      remove(premature_exit_filepath_);
+    }
+  }
+
+ private:
+  const char* const premature_exit_filepath_;
+
+  GTEST_DISALLOW_COPY_AND_ASSIGN_(ScopedPrematureExitFile);
+};
+
+}  // namespace internal
+
+// class TestEventListeners
+
+TestEventListeners::TestEventListeners()
+    : repeater_(new internal::TestEventRepeater()),
+      default_result_printer_(NULL),
+      default_xml_generator_(NULL) {
+}
+
+TestEventListeners::~TestEventListeners() { delete repeater_; }
+
+// Returns the standard listener responsible for the default console
+// output.  Can be removed from the listeners list to shut down default
+// console output.  Note that removing this object from the listener list
+// with Release transfers its ownership to the user.
+void TestEventListeners::Append(TestEventListener* listener) {
+  repeater_->Append(listener);
+}
+
+// Removes the given event listener from the list and returns it.  It then
+// becomes the caller's responsibility to delete the listener. Returns
+// NULL if the listener is not found in the list.
+TestEventListener* TestEventListeners::Release(TestEventListener* listener) {
+  if (listener == default_result_printer_)
+    default_result_printer_ = NULL;
+  else if (listener == default_xml_generator_)
+    default_xml_generator_ = NULL;
+  return repeater_->Release(listener);
+}
+
+// Returns repeater that broadcasts the TestEventListener events to all
+// subscribers.
+TestEventListener* TestEventListeners::repeater() { return repeater_; }
+
+// Sets the default_result_printer attribute to the provided listener.
+// The listener is also added to the listener list and previous
+// default_result_printer is removed from it and deleted. The listener can
+// also be NULL in which case it will not be added to the list. Does
+// nothing if the previous and the current listener objects are the same.
+void TestEventListeners::SetDefaultResultPrinter(TestEventListener* listener) {
+  if (default_result_printer_ != listener) {
+    // It is an error to pass this method a listener that is already in the
+    // list.
+    delete Release(default_result_printer_);
+    default_result_printer_ = listener;
+    if (listener != NULL)
+      Append(listener);
+  }
+}
+
+// Sets the default_xml_generator attribute to the provided listener.  The
+// listener is also added to the listener list and previous
+// default_xml_generator is removed from it and deleted. The listener can
+// also be NULL in which case it will not be added to the list. Does
+// nothing if the previous and the current listener objects are the same.
+void TestEventListeners::SetDefaultXmlGenerator(TestEventListener* listener) {
+  if (default_xml_generator_ != listener) {
+    // It is an error to pass this method a listener that is already in the
+    // list.
+    delete Release(default_xml_generator_);
+    default_xml_generator_ = listener;
+    if (listener != NULL)
+      Append(listener);
+  }
+}
+
+// Controls whether events will be forwarded by the repeater to the
+// listeners in the list.
+bool TestEventListeners::EventForwardingEnabled() const {
+  return repeater_->forwarding_enabled();
+}
+
+void TestEventListeners::SuppressEventForwarding() {
+  repeater_->set_forwarding_enabled(false);
+}
+
+// class UnitTest
+
+// Gets the singleton UnitTest object.  The first time this method is
+// called, a UnitTest object is constructed and returned.  Consecutive
+// calls will return the same object.
+//
+// We don't protect this under mutex_ as a user is not supposed to
+// call this before main() starts, from which point on the return
+// value will never change.
+UnitTest* UnitTest::GetInstance() {
+  // When compiled with MSVC 7.1 in optimized mode, destroying the
+  // UnitTest object upon exiting the program messes up the exit code,
+  // causing successful tests to appear failed.  We have to use a
+  // different implementation in this case to bypass the compiler bug.
+  // This implementation makes the compiler happy, at the cost of
+  // leaking the UnitTest object.
+
+  // CodeGear C++Builder insists on a public destructor for the
+  // default implementation.  Use this implementation to keep good OO
+  // design with private destructor.
+
+#if (_MSC_VER == 1310 && !defined(_DEBUG)) || defined(__BORLANDC__)
+  static UnitTest* const instance = new UnitTest;
+  return instance;
+#else
+  static UnitTest instance;
+  return &instance;
+#endif  // (_MSC_VER == 1310 && !defined(_DEBUG)) || defined(__BORLANDC__)
+}
+
+// Gets the number of successful test cases.
+int UnitTest::successful_test_case_count() const {
+  return impl()->successful_test_case_count();
+}
+
+// Gets the number of failed test cases.
+int UnitTest::failed_test_case_count() const {
+  return impl()->failed_test_case_count();
+}
+
+// Gets the number of all test cases.
+int UnitTest::total_test_case_count() const {
+  return impl()->total_test_case_count();
+}
+
+// Gets the number of all test cases that contain at least one test
+// that should run.
+int UnitTest::test_case_to_run_count() const {
+  return impl()->test_case_to_run_count();
+}
+
+// Gets the number of successful tests.
+int UnitTest::successful_test_count() const {
+  return impl()->successful_test_count();
+}
+
+// Gets the number of failed tests.
+int UnitTest::failed_test_count() const { return impl()->failed_test_count(); }
+
+// Gets the number of disabled tests that will be reported in the XML report.
+int UnitTest::reportable_disabled_test_count() const {
+  return impl()->reportable_disabled_test_count();
+}
+
+// Gets the number of disabled tests.
+int UnitTest::disabled_test_count() const {
+  return impl()->disabled_test_count();
+}
+
+// Gets the number of tests to be printed in the XML report.
+int UnitTest::reportable_test_count() const {
+  return impl()->reportable_test_count();
+}
+
+// Gets the number of all tests.
+int UnitTest::total_test_count() const { return impl()->total_test_count(); }
+
+// Gets the number of tests that should run.
+int UnitTest::test_to_run_count() const { return impl()->test_to_run_count(); }
+
+// Gets the time of the test program start, in ms from the start of the
+// UNIX epoch.
+internal::TimeInMillis UnitTest::start_timestamp() const {
+    return impl()->start_timestamp();
+}
+
+// Gets the elapsed time, in milliseconds.
+internal::TimeInMillis UnitTest::elapsed_time() const {
+  return impl()->elapsed_time();
+}
+
+// Returns true iff the unit test passed (i.e. all test cases passed).
+bool UnitTest::Passed() const { return impl()->Passed(); }
+
+// Returns true iff the unit test failed (i.e. some test case failed
+// or something outside of all tests failed).
+bool UnitTest::Failed() const { return impl()->Failed(); }
+
+// Gets the i-th test case among all the test cases. i can range from 0 to
+// total_test_case_count() - 1. If i is not in that range, returns NULL.
+const TestCase* UnitTest::GetTestCase(int i) const {
+  return impl()->GetTestCase(i);
+}
+
+// Returns the TestResult containing information on test failures and
+// properties logged outside of individual test cases.
+const TestResult& UnitTest::ad_hoc_test_result() const {
+  return *impl()->ad_hoc_test_result();
+}
+
+// Gets the i-th test case among all the test cases. i can range from 0 to
+// total_test_case_count() - 1. If i is not in that range, returns NULL.
+TestCase* UnitTest::GetMutableTestCase(int i) {
+  return impl()->GetMutableTestCase(i);
+}
+
+// Returns the list of event listeners that can be used to track events
+// inside Google Test.
+TestEventListeners& UnitTest::listeners() {
+  return *impl()->listeners();
+}
+
+// Registers and returns a global test environment.  When a test
+// program is run, all global test environments will be set-up in the
+// order they were registered.  After all tests in the program have
+// finished, all global test environments will be torn-down in the
+// *reverse* order they were registered.
+//
+// The UnitTest object takes ownership of the given environment.
+//
+// We don't protect this under mutex_, as we only support calling it
+// from the main thread.
+Environment* UnitTest::AddEnvironment(Environment* env) {
+  if (env == NULL) {
+    return NULL;
+  }
+
+  impl_->environments().push_back(env);
+  return env;
+}
+
+// Adds a TestPartResult to the current TestResult object.  All Google Test
+// assertion macros (e.g. ASSERT_TRUE, EXPECT_EQ, etc) eventually call
+// this to report their results.  The user code should use the
+// assertion macros instead of calling this directly.
+void UnitTest::AddTestPartResult(
+    TestPartResult::Type result_type,
+    const char* file_name,
+    int line_number,
+    const std::string& message,
+    const std::string& os_stack_trace) GTEST_LOCK_EXCLUDED_(mutex_) {
+  Message msg;
+  msg << message;
+
+  internal::MutexLock lock(&mutex_);
+  if (impl_->gtest_trace_stack().size() > 0) {
+    msg << "\n" << GTEST_NAME_ << " trace:";
+
+    for (int i = static_cast<int>(impl_->gtest_trace_stack().size());
+         i > 0; --i) {
+      const internal::TraceInfo& trace = impl_->gtest_trace_stack()[i - 1];
+      msg << "\n" << internal::FormatFileLocation(trace.file, trace.line)
+          << " " << trace.message;
+    }
+  }
+
+  if (os_stack_trace.c_str() != NULL && !os_stack_trace.empty()) {
+    msg << internal::kStackTraceMarker << os_stack_trace;
+  }
+
+  const TestPartResult result =
+    TestPartResult(result_type, file_name, line_number,
+                   msg.GetString().c_str());
+  impl_->GetTestPartResultReporterForCurrentThread()->
+      ReportTestPartResult(result);
+
+  if (result_type != TestPartResult::kSuccess) {
+    // gtest_break_on_failure takes precedence over
+    // gtest_throw_on_failure.  This allows a user to set the latter
+    // in the code (perhaps in order to use Google Test assertions
+    // with another testing framework) and specify the former on the
+    // command line for debugging.
+    if (GTEST_FLAG(break_on_failure)) {
+#if GTEST_OS_WINDOWS && !GTEST_OS_WINDOWS_PHONE && !GTEST_OS_WINDOWS_RT
+      // Using DebugBreak on Windows allows gtest to still break into a debugger
+      // when a failure happens and both the --gtest_break_on_failure and
+      // the --gtest_catch_exceptions flags are specified.
+      DebugBreak();
+#else
+      // Dereference NULL through a volatile pointer to prevent the compiler
+      // from removing. We use this rather than abort() or __builtin_trap() for
+      // portability: Symbian doesn't implement abort() well, and some debuggers
+      // don't correctly trap abort().
+      *static_cast<volatile int*>(NULL) = 1;
+#endif  // GTEST_OS_WINDOWS
+    } else if (GTEST_FLAG(throw_on_failure)) {
+#if GTEST_HAS_EXCEPTIONS
+      throw internal::GoogleTestFailureException(result);
+#else
+      // We cannot call abort() as it generates a pop-up in debug mode
+      // that cannot be suppressed in VC 7.1 or below.
+      exit(1);
+#endif
+    }
+  }
+}
+
+// Adds a TestProperty to the current TestResult object when invoked from
+// inside a test, to current TestCase's ad_hoc_test_result_ when invoked
+// from SetUpTestCase or TearDownTestCase, or to the global property set
+// when invoked elsewhere.  If the result already contains a property with
+// the same key, the value will be updated.
+void UnitTest::RecordProperty(const std::string& key,
+                              const std::string& value) {
+  impl_->RecordProperty(TestProperty(key, value));
+}
+
+// Runs all tests in this UnitTest object and prints the result.
+// Returns 0 if successful, or 1 otherwise.
+//
+// We don't protect this under mutex_, as we only support calling it
+// from the main thread.
+int UnitTest::Run() {
+  const bool in_death_test_child_process =
+      internal::GTEST_FLAG(internal_run_death_test).length() > 0;
+
+  // Google Test implements this protocol for catching that a test
+  // program exits before returning control to Google Test:
+  //
+  //   1. Upon start, Google Test creates a file whose absolute path
+  //      is specified by the environment variable
+  //      TEST_PREMATURE_EXIT_FILE.
+  //   2. When Google Test has finished its work, it deletes the file.
+  //
+  // This allows a test runner to set TEST_PREMATURE_EXIT_FILE before
+  // running a Google-Test-based test program and check the existence
+  // of the file at the end of the test execution to see if it has
+  // exited prematurely.
+
+  // If we are in the child process of a death test, don't
+  // create/delete the premature exit file, as doing so is unnecessary
+  // and will confuse the parent process.  Otherwise, create/delete
+  // the file upon entering/leaving this function.  If the program
+  // somehow exits before this function has a chance to return, the
+  // premature-exit file will be left undeleted, causing a test runner
+  // that understands the premature-exit-file protocol to report the
+  // test as having failed.
+  const internal::ScopedPrematureExitFile premature_exit_file(
+      in_death_test_child_process ?
+      NULL : internal::posix::GetEnv("TEST_PREMATURE_EXIT_FILE"));
+
+  // Captures the value of GTEST_FLAG(catch_exceptions).  This value will be
+  // used for the duration of the program.
+  impl()->set_catch_exceptions(GTEST_FLAG(catch_exceptions));
+
+#if GTEST_HAS_SEH
+  // Either the user wants Google Test to catch exceptions thrown by the
+  // tests or this is executing in the context of death test child
+  // process. In either case the user does not want to see pop-up dialogs
+  // about crashes - they are expected.
+  if (impl()->catch_exceptions() || in_death_test_child_process) {
+# if !GTEST_OS_WINDOWS_MOBILE && !GTEST_OS_WINDOWS_PHONE && !GTEST_OS_WINDOWS_RT
+    // SetErrorMode doesn't exist on CE.
+    SetErrorMode(SEM_FAILCRITICALERRORS | SEM_NOALIGNMENTFAULTEXCEPT |
+                 SEM_NOGPFAULTERRORBOX | SEM_NOOPENFILEERRORBOX);
+# endif  // !GTEST_OS_WINDOWS_MOBILE
+
+# if (defined(_MSC_VER) || GTEST_OS_WINDOWS_MINGW) && !GTEST_OS_WINDOWS_MOBILE
+    // Death test children can be terminated with _abort().  On Windows,
+    // _abort() can show a dialog with a warning message.  This forces the
+    // abort message to go to stderr instead.
+    _set_error_mode(_OUT_TO_STDERR);
+# endif
+
+# if _MSC_VER >= 1400 && !GTEST_OS_WINDOWS_MOBILE
+    // In the debug version, Visual Studio pops up a separate dialog
+    // offering a choice to debug the aborted program. We need to suppress
+    // this dialog or it will pop up for every EXPECT/ASSERT_DEATH statement
+    // executed. Google Test will notify the user of any unexpected
+    // failure via stderr.
+    //
+    // VC++ doesn't define _set_abort_behavior() prior to the version 8.0.
+    // Users of prior VC versions shall suffer the agony and pain of
+    // clicking through the countless debug dialogs.
+    // TODO(vladl@google.com): find a way to suppress the abort dialog() in the
+    // debug mode when compiled with VC 7.1 or lower.
+    if (!GTEST_FLAG(break_on_failure))
+      _set_abort_behavior(
+          0x0,                                    // Clear the following flags:
+          _WRITE_ABORT_MSG | _CALL_REPORTFAULT);  // pop-up window, core dump.
+# endif
+  }
+#endif  // GTEST_HAS_SEH
+
+  return internal::HandleExceptionsInMethodIfSupported(
+      impl(),
+      &internal::UnitTestImpl::RunAllTests,
+      "auxiliary test code (environments or event listeners)") ? 0 : 1;
+}
+
+// Returns the working directory when the first TEST() or TEST_F() was
+// executed.
+const char* UnitTest::original_working_dir() const {
+  return impl_->original_working_dir_.c_str();
+}
+
+// Returns the TestCase object for the test that's currently running,
+// or NULL if no test is running.
+const TestCase* UnitTest::current_test_case() const
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  return impl_->current_test_case();
+}
+
+// Returns the TestInfo object for the test that's currently running,
+// or NULL if no test is running.
+const TestInfo* UnitTest::current_test_info() const
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  return impl_->current_test_info();
+}
+
+// Returns the random seed used at the start of the current test run.
+int UnitTest::random_seed() const { return impl_->random_seed(); }
+
+#if GTEST_HAS_PARAM_TEST
+// Returns ParameterizedTestCaseRegistry object used to keep track of
+// value-parameterized tests and instantiate and register them.
+internal::ParameterizedTestCaseRegistry&
+    UnitTest::parameterized_test_registry()
+        GTEST_LOCK_EXCLUDED_(mutex_) {
+  return impl_->parameterized_test_registry();
+}
+#endif  // GTEST_HAS_PARAM_TEST
+
+// Creates an empty UnitTest.
+UnitTest::UnitTest() {
+  impl_ = new internal::UnitTestImpl(this);
+}
+
+// Destructor of UnitTest.
+UnitTest::~UnitTest() {
+  delete impl_;
+}
+
+// Pushes a trace defined by SCOPED_TRACE() on to the per-thread
+// Google Test trace stack.
+void UnitTest::PushGTestTrace(const internal::TraceInfo& trace)
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  impl_->gtest_trace_stack().push_back(trace);
+}
+
+// Pops a trace from the per-thread Google Test trace stack.
+void UnitTest::PopGTestTrace()
+    GTEST_LOCK_EXCLUDED_(mutex_) {
+  internal::MutexLock lock(&mutex_);
+  impl_->gtest_trace_stack().pop_back();
+}
+
+namespace internal {
+
+UnitTestImpl::UnitTestImpl(UnitTest* parent)
+    : parent_(parent),
+      GTEST_DISABLE_MSC_WARNINGS_PUSH_(4355 /* using this in initializer */)
+      default_global_test_part_result_reporter_(this),
+      default_per_thread_test_part_result_reporter_(this),
+      GTEST_DISABLE_MSC_WARNINGS_POP_()
+      global_test_part_result_repoter_(
+          &default_global_test_part_result_reporter_),
+      per_thread_test_part_result_reporter_(
+          &default_per_thread_test_part_result_reporter_),
+#if GTEST_HAS_PARAM_TEST
+      parameterized_test_registry_(),
+      parameterized_tests_registered_(false),
+#endif  // GTEST_HAS_PARAM_TEST
+      last_death_test_case_(-1),
+      current_test_case_(NULL),
+      current_test_info_(NULL),
+      ad_hoc_test_result_(),
+      os_stack_trace_getter_(NULL),
+      post_flag_parse_init_performed_(false),
+      random_seed_(0),  // Will be overridden by the flag before first use.
+      random_(0),  // Will be reseeded before first use.
+      start_timestamp_(0),
+      elapsed_time_(0),
+#if GTEST_HAS_DEATH_TEST
+      death_test_factory_(new DefaultDeathTestFactory),
+#endif
+      // Will be overridden by the flag before first use.
+      catch_exceptions_(false) {
+  listeners()->SetDefaultResultPrinter(new PrettyUnitTestResultPrinter);
+}
+
+UnitTestImpl::~UnitTestImpl() {
+  // Deletes every TestCase.
+  ForEach(test_cases_, internal::Delete<TestCase>);
+
+  // Deletes every Environment.
+  ForEach(environments_, internal::Delete<Environment>);
+
+  delete os_stack_trace_getter_;
+}
+
+// Adds a TestProperty to the current TestResult object when invoked in a
+// context of a test, to current test case's ad_hoc_test_result when invoke
+// from SetUpTestCase/TearDownTestCase, or to the global property set
+// otherwise.  If the result already contains a property with the same key,
+// the value will be updated.
+void UnitTestImpl::RecordProperty(const TestProperty& test_property) {
+  std::string xml_element;
+  TestResult* test_result;  // TestResult appropriate for property recording.
+
+  if (current_test_info_ != NULL) {
+    xml_element = "testcase";
+    test_result = &(current_test_info_->result_);
+  } else if (current_test_case_ != NULL) {
+    xml_element = "testsuite";
+    test_result = &(current_test_case_->ad_hoc_test_result_);
+  } else {
+    xml_element = "testsuites";
+    test_result = &ad_hoc_test_result_;
+  }
+  test_result->RecordProperty(xml_element, test_property);
+}
+
+#if GTEST_HAS_DEATH_TEST
+// Disables event forwarding if the control is currently in a death test
+// subprocess. Must not be called before InitGoogleTest.
+void UnitTestImpl::SuppressTestEventsIfInSubprocess() {
+  if (internal_run_death_test_flag_.get() != NULL)
+    listeners()->SuppressEventForwarding();
+}
+#endif  // GTEST_HAS_DEATH_TEST
+
+// Initializes event listeners performing XML output as specified by
+// UnitTestOptions. Must not be called before InitGoogleTest.
+void UnitTestImpl::ConfigureXmlOutput() {
+  const std::string& output_format = UnitTestOptions::GetOutputFormat();
+  if (output_format == "xml") {
+    listeners()->SetDefaultXmlGenerator(new XmlUnitTestResultPrinter(
+        UnitTestOptions::GetAbsolutePathToOutputFile().c_str()));
+  } else if (output_format != "") {
+    printf("WARNING: unrecognized output format \"%s\" ignored.\n",
+           output_format.c_str());
+    fflush(stdout);
+  }
+}
+
+#if GTEST_CAN_STREAM_RESULTS_
+// Initializes event listeners for streaming test results in string form.
+// Must not be called before InitGoogleTest.
+void UnitTestImpl::ConfigureStreamingOutput() {
+  const std::string& target = GTEST_FLAG(stream_result_to);
+  if (!target.empty()) {
+    const size_t pos = target.find(':');
+    if (pos != std::string::npos) {
+      listeners()->Append(new StreamingListener(target.substr(0, pos),
+                                                target.substr(pos+1)));
+    } else {
+      printf("WARNING: unrecognized streaming target \"%s\" ignored.\n",
+             target.c_str());
+      fflush(stdout);
+    }
+  }
+}
+#endif  // GTEST_CAN_STREAM_RESULTS_
+
+// Performs initialization dependent upon flag values obtained in
+// ParseGoogleTestFlagsOnly.  Is called from InitGoogleTest after the call to
+// ParseGoogleTestFlagsOnly.  In case a user neglects to call InitGoogleTest
+// this function is also called from RunAllTests.  Since this function can be
+// called more than once, it has to be idempotent.
+void UnitTestImpl::PostFlagParsingInit() {
+  // Ensures that this function does not execute more than once.
+  if (!post_flag_parse_init_performed_) {
+    post_flag_parse_init_performed_ = true;
+
+#if defined(GTEST_CUSTOM_TEST_EVENT_LISTENER_)
+    // Register to send notifications about key process state changes.
+    listeners()->Append(new GTEST_CUSTOM_TEST_EVENT_LISTENER_());
+#endif  // defined(GTEST_CUSTOM_TEST_EVENT_LISTENER_)
+
+#if GTEST_HAS_DEATH_TEST
+    InitDeathTestSubprocessControlInfo();
+    SuppressTestEventsIfInSubprocess();
+#endif  // GTEST_HAS_DEATH_TEST
+
+    // Registers parameterized tests. This makes parameterized tests
+    // available to the UnitTest reflection API without running
+    // RUN_ALL_TESTS.
+    RegisterParameterizedTests();
+
+    // Configures listeners for XML output. This makes it possible for users
+    // to shut down the default XML output before invoking RUN_ALL_TESTS.
+    ConfigureXmlOutput();
+
+#if GTEST_CAN_STREAM_RESULTS_
+    // Configures listeners for streaming test results to the specified server.
+    ConfigureStreamingOutput();
+#endif  // GTEST_CAN_STREAM_RESULTS_
+  }
+}
+
+// A predicate that checks the name of a TestCase against a known
+// value.
+//
+// This is used for implementation of the UnitTest class only.  We put
+// it in the anonymous namespace to prevent polluting the outer
+// namespace.
+//
+// TestCaseNameIs is copyable.
+class TestCaseNameIs {
+ public:
+  // Constructor.
+  explicit TestCaseNameIs(const std::string& name)
+      : name_(name) {}
+
+  // Returns true iff the name of test_case matches name_.
+  bool operator()(const TestCase* test_case) const {
+    return test_case != NULL && strcmp(test_case->name(), name_.c_str()) == 0;
+  }
+
+ private:
+  std::string name_;
+};
+
+// Finds and returns a TestCase with the given name.  If one doesn't
+// exist, creates one and returns it.  It's the CALLER'S
+// RESPONSIBILITY to ensure that this function is only called WHEN THE
+// TESTS ARE NOT SHUFFLED.
+//
+// Arguments:
+//
+//   test_case_name: name of the test case
+//   type_param:     the name of the test case's type parameter, or NULL if
+//                   this is not a typed or a type-parameterized test case.
+//   set_up_tc:      pointer to the function that sets up the test case
+//   tear_down_tc:   pointer to the function that tears down the test case
+TestCase* UnitTestImpl::GetTestCase(const char* test_case_name,
+                                    const char* type_param,
+                                    Test::SetUpTestCaseFunc set_up_tc,
+                                    Test::TearDownTestCaseFunc tear_down_tc) {
+  // Can we find a TestCase with the given name?
+  const std::vector<TestCase*>::const_iterator test_case =
+      std::find_if(test_cases_.begin(), test_cases_.end(),
+                   TestCaseNameIs(test_case_name));
+
+  if (test_case != test_cases_.end())
+    return *test_case;
+
+  // No.  Let's create one.
+  TestCase* const new_test_case =
+      new TestCase(test_case_name, type_param, set_up_tc, tear_down_tc);
+
+  // Is this a death test case?
+  if (internal::UnitTestOptions::MatchesFilter(test_case_name,
+                                               kDeathTestCaseFilter)) {
+    // Yes.  Inserts the test case after the last death test case
+    // defined so far.  This only works when the test cases haven't
+    // been shuffled.  Otherwise we may end up running a death test
+    // after a non-death test.
+    ++last_death_test_case_;
+    test_cases_.insert(test_cases_.begin() + last_death_test_case_,
+                       new_test_case);
+  } else {
+    // No.  Appends to the end of the list.
+    test_cases_.push_back(new_test_case);
+  }
+
+  test_case_indices_.push_back(static_cast<int>(test_case_indices_.size()));
+  return new_test_case;
+}
+
+// Helpers for setting up / tearing down the given environment.  They
+// are for use in the ForEach() function.
+static void SetUpEnvironment(Environment* env) { env->SetUp(); }
+static void TearDownEnvironment(Environment* env) { env->TearDown(); }
+
+// Runs all tests in this UnitTest object, prints the result, and
+// returns true if all tests are successful.  If any exception is
+// thrown during a test, the test is considered to be failed, but the
+// rest of the tests will still be run.
+//
+// When parameterized tests are enabled, it expands and registers
+// parameterized tests first in RegisterParameterizedTests().
+// All other functions called from RunAllTests() may safely assume that
+// parameterized tests are ready to be counted and run.
+bool UnitTestImpl::RunAllTests() {
+  // Makes sure InitGoogleTest() was called.
+  if (!GTestIsInitialized()) {
+    printf("%s",
+           "\nThis test program did NOT call ::testing::InitGoogleTest "
+           "before calling RUN_ALL_TESTS().  Please fix it.\n");
+    return false;
+  }
+
+  // Do not run any test if the --help flag was specified.
+  if (g_help_flag)
+    return true;
+
+  // Repeats the call to the post-flag parsing initialization in case the
+  // user didn't call InitGoogleTest.
+  PostFlagParsingInit();
+
+  // Even if sharding is not on, test runners may want to use the
+  // GTEST_SHARD_STATUS_FILE to query whether the test supports the sharding
+  // protocol.
+  internal::WriteToShardStatusFileIfNeeded();
+
+  // True iff we are in a subprocess for running a thread-safe-style
+  // death test.
+  bool in_subprocess_for_death_test = false;
+
+#if GTEST_HAS_DEATH_TEST
+  in_subprocess_for_death_test = (internal_run_death_test_flag_.get() != NULL);
+# if defined(GTEST_EXTRA_DEATH_TEST_CHILD_SETUP_)
+  if (in_subprocess_for_death_test) {
+    GTEST_EXTRA_DEATH_TEST_CHILD_SETUP_();
+  }
+# endif  // defined(GTEST_EXTRA_DEATH_TEST_CHILD_SETUP_)
+#endif  // GTEST_HAS_DEATH_TEST
+
+  const bool should_shard = ShouldShard(kTestTotalShards, kTestShardIndex,
+                                        in_subprocess_for_death_test);
+
+  // Compares the full test names with the filter to decide which
+  // tests to run.
+  const bool has_tests_to_run = FilterTests(should_shard
+                                              ? HONOR_SHARDING_PROTOCOL
+                                              : IGNORE_SHARDING_PROTOCOL) > 0;
+
+  // Lists the tests and exits if the --gtest_list_tests flag was specified.
+  if (GTEST_FLAG(list_tests)) {
+    // This must be called *after* FilterTests() has been called.
+    ListTestsMatchingFilter();
+    return true;
+  }
+
+  random_seed_ = GTEST_FLAG(shuffle) ?
+      GetRandomSeedFromFlag(GTEST_FLAG(random_seed)) : 0;
+
+  // True iff at least one test has failed.
+  bool failed = false;
+
+  TestEventListener* repeater = listeners()->repeater();
+
+  start_timestamp_ = GetTimeInMillis();
+  repeater->OnTestProgramStart(*parent_);
+
+  // How many times to repeat the tests?  We don't want to repeat them
+  // when we are inside the subprocess of a death test.
+  const int repeat = in_subprocess_for_death_test ? 1 : GTEST_FLAG(repeat);
+  // Repeats forever if the repeat count is negative.
+  const bool forever = repeat < 0;
+  for (int i = 0; forever || i != repeat; i++) {
+    // We want to preserve failures generated by ad-hoc test
+    // assertions executed before RUN_ALL_TESTS().
+    ClearNonAdHocTestResult();
+
+    const TimeInMillis start = GetTimeInMillis();
+
+    // Shuffles test cases and tests if requested.
+    if (has_tests_to_run && GTEST_FLAG(shuffle)) {
+      random()->Reseed(random_seed_);
+      // This should be done before calling OnTestIterationStart(),
+      // such that a test event listener can see the actual test order
+      // in the event.
+      ShuffleTests();
+    }
+
+    // Tells the unit test event listeners that the tests are about to start.
+    repeater->OnTestIterationStart(*parent_, i);
+
+    // Runs each test case if there is at least one test to run.
+    if (has_tests_to_run) {
+      // Sets up all environments beforehand.
+      repeater->OnEnvironmentsSetUpStart(*parent_);
+      ForEach(environments_, SetUpEnvironment);
+      repeater->OnEnvironmentsSetUpEnd(*parent_);
+
+      // Runs the tests only if there was no fatal failure during global
+      // set-up.
+      if (!Test::HasFatalFailure()) {
+        for (int test_index = 0; test_index < total_test_case_count();
+             test_index++) {
+          GetMutableTestCase(test_index)->Run();
+        }
+      }
+
+      // Tears down all environments in reverse order afterwards.
+      repeater->OnEnvironmentsTearDownStart(*parent_);
+      std::for_each(environments_.rbegin(), environments_.rend(),
+                    TearDownEnvironment);
+      repeater->OnEnvironmentsTearDownEnd(*parent_);
+    }
+
+    elapsed_time_ = GetTimeInMillis() - start;
+
+    // Tells the unit test event listener that the tests have just finished.
+    repeater->OnTestIterationEnd(*parent_, i);
+
+    // Gets the result and clears it.
+    if (!Passed()) {
+      failed = true;
+    }
+
+    // Restores the original test order after the iteration.  This
+    // allows the user to quickly repro a failure that happens in the
+    // N-th iteration without repeating the first (N - 1) iterations.
+    // This is not enclosed in "if (GTEST_FLAG(shuffle)) { ... }", in
+    // case the user somehow changes the value of the flag somewhere
+    // (it's always safe to unshuffle the tests).
+    UnshuffleTests();
+
+    if (GTEST_FLAG(shuffle)) {
+      // Picks a new random seed for each iteration.
+      random_seed_ = GetNextRandomSeed(random_seed_);
+    }
+  }
+
+  repeater->OnTestProgramEnd(*parent_);
+
+  return !failed;
+}
+
+// Reads the GTEST_SHARD_STATUS_FILE environment variable, and creates the file
+// if the variable is present. If a file already exists at this location, this
+// function will write over it. If the variable is present, but the file cannot
+// be created, prints an error and exits.
+void WriteToShardStatusFileIfNeeded() {
+  const char* const test_shard_file = posix::GetEnv(kTestShardStatusFile);
+  if (test_shard_file != NULL) {
+    FILE* const file = posix::FOpen(test_shard_file, "w");
+    if (file == NULL) {
+      ColoredPrintf(COLOR_RED,
+                    "Could not write to the test shard status file \"%s\" "
+                    "specified by the %s environment variable.\n",
+                    test_shard_file, kTestShardStatusFile);
+      fflush(stdout);
+      exit(EXIT_FAILURE);
+    }
+    fclose(file);
+  }
+}
+
+// Checks whether sharding is enabled by examining the relevant
+// environment variable values. If the variables are present,
+// but inconsistent (i.e., shard_index >= total_shards), prints
+// an error and exits. If in_subprocess_for_death_test, sharding is
+// disabled because it must only be applied to the original test
+// process. Otherwise, we could filter out death tests we intended to execute.
+bool ShouldShard(const char* total_shards_env,
+                 const char* shard_index_env,
+                 bool in_subprocess_for_death_test) {
+  if (in_subprocess_for_death_test) {
+    return false;
+  }
+
+  const Int32 total_shards = Int32FromEnvOrDie(total_shards_env, -1);
+  const Int32 shard_index = Int32FromEnvOrDie(shard_index_env, -1);
+
+  if (total_shards == -1 && shard_index == -1) {
+    return false;
+  } else if (total_shards == -1 && shard_index != -1) {
+    const Message msg = Message()
+      << "Invalid environment variables: you have "
+      << kTestShardIndex << " = " << shard_index
+      << ", but have left " << kTestTotalShards << " unset.\n";
+    ColoredPrintf(COLOR_RED, msg.GetString().c_str());
+    fflush(stdout);
+    exit(EXIT_FAILURE);
+  } else if (total_shards != -1 && shard_index == -1) {
+    const Message msg = Message()
+      << "Invalid environment variables: you have "
+      << kTestTotalShards << " = " << total_shards
+      << ", but have left " << kTestShardIndex << " unset.\n";
+    ColoredPrintf(COLOR_RED, msg.GetString().c_str());
+    fflush(stdout);
+    exit(EXIT_FAILURE);
+  } else if (shard_index < 0 || shard_index >= total_shards) {
+    const Message msg = Message()
+      << "Invalid environment variables: we require 0 <= "
+      << kTestShardIndex << " < " << kTestTotalShards
+      << ", but you have " << kTestShardIndex << "=" << shard_index
+      << ", " << kTestTotalShards << "=" << total_shards << ".\n";
+    ColoredPrintf(COLOR_RED, msg.GetString().c_str());
+    fflush(stdout);
+    exit(EXIT_FAILURE);
+  }
+
+  return total_shards > 1;
+}
+
+// Parses the environment variable var as an Int32. If it is unset,
+// returns default_val. If it is not an Int32, prints an error
+// and aborts.
+Int32 Int32FromEnvOrDie(const char* var, Int32 default_val) {
+  const char* str_val = posix::GetEnv(var);
+  if (str_val == NULL) {
+    return default_val;
+  }
+
+  Int32 result;
+  if (!ParseInt32(Message() << "The value of environment variable " << var,
+                  str_val, &result)) {
+    exit(EXIT_FAILURE);
+  }
+  return result;
+}
+
+// Given the total number of shards, the shard index, and the test id,
+// returns true iff the test should be run on this shard. The test id is
+// some arbitrary but unique non-negative integer assigned to each test
+// method. Assumes that 0 <= shard_index < total_shards.
+bool ShouldRunTestOnShard(int total_shards, int shard_index, int test_id) {
+  return (test_id % total_shards) == shard_index;
+}
+
+// Compares the name of each test with the user-specified filter to
+// decide whether the test should be run, then records the result in
+// each TestCase and TestInfo object.
+// If shard_tests == true, further filters tests based on sharding
+// variables in the environment - see
+// http://code.google.com/p/googletest/wiki/GoogleTestAdvancedGuide.
+// Returns the number of tests that should run.
+int UnitTestImpl::FilterTests(ReactionToSharding shard_tests) {
+  const Int32 total_shards = shard_tests == HONOR_SHARDING_PROTOCOL ?
+      Int32FromEnvOrDie(kTestTotalShards, -1) : -1;
+  const Int32 shard_index = shard_tests == HONOR_SHARDING_PROTOCOL ?
+      Int32FromEnvOrDie(kTestShardIndex, -1) : -1;
+
+  // num_runnable_tests are the number of tests that will
+  // run across all shards (i.e., match filter and are not disabled).
+  // num_selected_tests are the number of tests to be run on
+  // this shard.
+  int num_runnable_tests = 0;
+  int num_selected_tests = 0;
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    TestCase* const test_case = test_cases_[i];
+    const std::string &test_case_name = test_case->name();
+    test_case->set_should_run(false);
+
+    for (size_t j = 0; j < test_case->test_info_list().size(); j++) {
+      TestInfo* const test_info = test_case->test_info_list()[j];
+      const std::string test_name(test_info->name());
+      // A test is disabled if test case name or test name matches
+      // kDisableTestFilter.
+      const bool is_disabled =
+          internal::UnitTestOptions::MatchesFilter(test_case_name,
+                                                   kDisableTestFilter) ||
+          internal::UnitTestOptions::MatchesFilter(test_name,
+                                                   kDisableTestFilter);
+      test_info->is_disabled_ = is_disabled;
+
+      const bool matches_filter =
+          internal::UnitTestOptions::FilterMatchesTest(test_case_name,
+                                                       test_name);
+      test_info->matches_filter_ = matches_filter;
+
+      const bool is_runnable =
+          (GTEST_FLAG(also_run_disabled_tests) || !is_disabled) &&
+          matches_filter;
+
+      const bool is_selected = is_runnable &&
+          (shard_tests == IGNORE_SHARDING_PROTOCOL ||
+           ShouldRunTestOnShard(total_shards, shard_index,
+                                num_runnable_tests));
+
+      num_runnable_tests += is_runnable;
+      num_selected_tests += is_selected;
+
+      test_info->should_run_ = is_selected;
+      test_case->set_should_run(test_case->should_run() || is_selected);
+    }
+  }
+  return num_selected_tests;
+}
+
+// Prints the given C-string on a single line by replacing all '\n'
+// characters with string "\\n".  If the output takes more than
+// max_length characters, only prints the first max_length characters
+// and "...".
+static void PrintOnOneLine(const char* str, int max_length) {
+  if (str != NULL) {
+    for (int i = 0; *str != '\0'; ++str) {
+      if (i >= max_length) {
+        printf("...");
+        break;
+      }
+      if (*str == '\n') {
+        printf("\\n");
+        i += 2;
+      } else {
+        printf("%c", *str);
+        ++i;
+      }
+    }
+  }
+}
+
+// Prints the names of the tests matching the user-specified filter flag.
+void UnitTestImpl::ListTestsMatchingFilter() {
+  // Print at most this many characters for each type/value parameter.
+  const int kMaxParamLength = 250;
+
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    const TestCase* const test_case = test_cases_[i];
+    bool printed_test_case_name = false;
+
+    for (size_t j = 0; j < test_case->test_info_list().size(); j++) {
+      const TestInfo* const test_info =
+          test_case->test_info_list()[j];
+      if (test_info->matches_filter_) {
+        if (!printed_test_case_name) {
+          printed_test_case_name = true;
+          printf("%s.", test_case->name());
+          if (test_case->type_param() != NULL) {
+            printf("  # %s = ", kTypeParamLabel);
+            // We print the type parameter on a single line to make
+            // the output easy to parse by a program.
+            PrintOnOneLine(test_case->type_param(), kMaxParamLength);
+          }
+          printf("\n");
+        }
+        printf("  %s", test_info->name());
+        if (test_info->value_param() != NULL) {
+          printf("  # %s = ", kValueParamLabel);
+          // We print the value parameter on a single line to make the
+          // output easy to parse by a program.
+          PrintOnOneLine(test_info->value_param(), kMaxParamLength);
+        }
+        printf("\n");
+      }
+    }
+  }
+  fflush(stdout);
+}
+
+// Sets the OS stack trace getter.
+//
+// Does nothing if the input and the current OS stack trace getter are
+// the same; otherwise, deletes the old getter and makes the input the
+// current getter.
+void UnitTestImpl::set_os_stack_trace_getter(
+    OsStackTraceGetterInterface* getter) {
+  if (os_stack_trace_getter_ != getter) {
+    delete os_stack_trace_getter_;
+    os_stack_trace_getter_ = getter;
+  }
+}
+
+// Returns the current OS stack trace getter if it is not NULL;
+// otherwise, creates an OsStackTraceGetter, makes it the current
+// getter, and returns it.
+OsStackTraceGetterInterface* UnitTestImpl::os_stack_trace_getter() {
+  if (os_stack_trace_getter_ == NULL) {
+#ifdef GTEST_OS_STACK_TRACE_GETTER_
+    os_stack_trace_getter_ = new GTEST_OS_STACK_TRACE_GETTER_;
+#else
+    os_stack_trace_getter_ = new OsStackTraceGetter;
+#endif  // GTEST_OS_STACK_TRACE_GETTER_
+  }
+
+  return os_stack_trace_getter_;
+}
+
+// Returns the TestResult for the test that's currently running, or
+// the TestResult for the ad hoc test if no test is running.
+TestResult* UnitTestImpl::current_test_result() {
+  return current_test_info_ ?
+      &(current_test_info_->result_) : &ad_hoc_test_result_;
+}
+
+// Shuffles all test cases, and the tests within each test case,
+// making sure that death tests are still run first.
+void UnitTestImpl::ShuffleTests() {
+  // Shuffles the death test cases.
+  ShuffleRange(random(), 0, last_death_test_case_ + 1, &test_case_indices_);
+
+  // Shuffles the non-death test cases.
+  ShuffleRange(random(), last_death_test_case_ + 1,
+               static_cast<int>(test_cases_.size()), &test_case_indices_);
+
+  // Shuffles the tests inside each test case.
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    test_cases_[i]->ShuffleTests(random());
+  }
+}
+
+// Restores the test cases and tests to their order before the first shuffle.
+void UnitTestImpl::UnshuffleTests() {
+  for (size_t i = 0; i < test_cases_.size(); i++) {
+    // Unshuffles the tests in each test case.
+    test_cases_[i]->UnshuffleTests();
+    // Resets the index of each test case.
+    test_case_indices_[i] = static_cast<int>(i);
+  }
+}
+
+// Returns the current OS stack trace as an std::string.
+//
+// The maximum number of stack frames to be included is specified by
+// the gtest_stack_trace_depth flag.  The skip_count parameter
+// specifies the number of top frames to be skipped, which doesn't
+// count against the number of frames to be included.
+//
+// For example, if Foo() calls Bar(), which in turn calls
+// GetCurrentOsStackTraceExceptTop(..., 1), Foo() will be included in
+// the trace but Bar() and GetCurrentOsStackTraceExceptTop() won't.
+std::string GetCurrentOsStackTraceExceptTop(UnitTest* /*unit_test*/,
+                                            int skip_count) {
+  // We pass skip_count + 1 to skip this wrapper function in addition
+  // to what the user really wants to skip.
+  return GetUnitTestImpl()->CurrentOsStackTraceExceptTop(skip_count + 1);
+}
+
+// Used by the GTEST_SUPPRESS_UNREACHABLE_CODE_WARNING_BELOW_ macro to
+// suppress unreachable code warnings.
+namespace {
+class ClassUniqueToAlwaysTrue {};
+}
+
+bool IsTrue(bool condition) { return condition; }
+
+bool AlwaysTrue() {
+#if GTEST_HAS_EXCEPTIONS
+  // This condition is always false so AlwaysTrue() never actually throws,
+  // but it makes the compiler think that it may throw.
+  if (IsTrue(false))
+    throw ClassUniqueToAlwaysTrue();
+#endif  // GTEST_HAS_EXCEPTIONS
+  return true;
+}
+
+// If *pstr starts with the given prefix, modifies *pstr to be right
+// past the prefix and returns true; otherwise leaves *pstr unchanged
+// and returns false.  None of pstr, *pstr, and prefix can be NULL.
+bool SkipPrefix(const char* prefix, const char** pstr) {
+  const size_t prefix_len = strlen(prefix);
+  if (strncmp(*pstr, prefix, prefix_len) == 0) {
+    *pstr += prefix_len;
+    return true;
+  }
+  return false;
+}
+
+// Parses a string as a command line flag.  The string should have
+// the format "--flag=value".  When def_optional is true, the "=value"
+// part can be omitted.
+//
+// Returns the value of the flag, or NULL if the parsing failed.
+const char* ParseFlagValue(const char* str,
+                           const char* flag,
+                           bool def_optional) {
+  // str and flag must not be NULL.
+  if (str == NULL || flag == NULL) return NULL;
+
+  // The flag must start with "--" followed by GTEST_FLAG_PREFIX_.
+  const std::string flag_str = std::string("--") + GTEST_FLAG_PREFIX_ + flag;
+  const size_t flag_len = flag_str.length();
+  if (strncmp(str, flag_str.c_str(), flag_len) != 0) return NULL;
+
+  // Skips the flag name.
+  const char* flag_end = str + flag_len;
+
+  // When def_optional is true, it's OK to not have a "=value" part.
+  if (def_optional && (flag_end[0] == '\0')) {
+    return flag_end;
+  }
+
+  // If def_optional is true and there are more characters after the
+  // flag name, or if def_optional is false, there must be a '=' after
+  // the flag name.
+  if (flag_end[0] != '=') return NULL;
+
+  // Returns the string after "=".
+  return flag_end + 1;
+}
+
+// Parses a string for a bool flag, in the form of either
+// "--flag=value" or "--flag".
+//
+// In the former case, the value is taken as true as long as it does
+// not start with '0', 'f', or 'F'.
+//
+// In the latter case, the value is taken as true.
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseBoolFlag(const char* str, const char* flag, bool* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, true);
+
+  // Aborts if the parsing failed.
+  if (value_str == NULL) return false;
+
+  // Converts the string value to a bool.
+  *value = !(*value_str == '0' || *value_str == 'f' || *value_str == 'F');
+  return true;
+}
+
+// Parses a string for an Int32 flag, in the form of
+// "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseInt32Flag(const char* str, const char* flag, Int32* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, false);
+
+  // Aborts if the parsing failed.
+  if (value_str == NULL) return false;
+
+  // Sets *value to the value of the flag.
+  return ParseInt32(Message() << "The value of flag --" << flag,
+                    value_str, value);
+}
+
+// Parses a string for a string flag, in the form of
+// "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseStringFlag(const char* str, const char* flag, std::string* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, false);
+
+  // Aborts if the parsing failed.
+  if (value_str == NULL) return false;
+
+  // Sets *value to the value of the flag.
+  *value = value_str;
+  return true;
+}
+
+// Determines whether a string has a prefix that Google Test uses for its
+// flags, i.e., starts with GTEST_FLAG_PREFIX_ or GTEST_FLAG_PREFIX_DASH_.
+// If Google Test detects that a command line flag has its prefix but is not
+// recognized, it will print its help message. Flags starting with
+// GTEST_INTERNAL_PREFIX_ followed by "internal_" are considered Google Test
+// internal flags and do not trigger the help message.
+static bool HasGoogleTestFlagPrefix(const char* str) {
+  return (SkipPrefix("--", &str) ||
+          SkipPrefix("-", &str) ||
+          SkipPrefix("/", &str)) &&
+         !SkipPrefix(GTEST_FLAG_PREFIX_ "internal_", &str) &&
+         (SkipPrefix(GTEST_FLAG_PREFIX_, &str) ||
+          SkipPrefix(GTEST_FLAG_PREFIX_DASH_, &str));
+}
+
+// Prints a string containing code-encoded text.  The following escape
+// sequences can be used in the string to control the text color:
+//
+//   @@    prints a single '@' character.
+//   @R    changes the color to red.
+//   @G    changes the color to green.
+//   @Y    changes the color to yellow.
+//   @D    changes to the default terminal text color.
+//
+// TODO(wan@google.com): Write tests for this once we add stdout
+// capturing to Google Test.
+static void PrintColorEncoded(const char* str) {
+  GTestColor color = COLOR_DEFAULT;  // The current color.
+
+  // Conceptually, we split the string into segments divided by escape
+  // sequences.  Then we print one segment at a time.  At the end of
+  // each iteration, the str pointer advances to the beginning of the
+  // next segment.
+  for (;;) {
+    const char* p = strchr(str, '@');
+    if (p == NULL) {
+      ColoredPrintf(color, "%s", str);
+      return;
+    }
+
+    ColoredPrintf(color, "%s", std::string(str, p).c_str());
+
+    const char ch = p[1];
+    str = p + 2;
+    if (ch == '@') {
+      ColoredPrintf(color, "@");
+    } else if (ch == 'D') {
+      color = COLOR_DEFAULT;
+    } else if (ch == 'R') {
+      color = COLOR_RED;
+    } else if (ch == 'G') {
+      color = COLOR_GREEN;
+    } else if (ch == 'Y') {
+      color = COLOR_YELLOW;
+    } else {
+      --str;
+    }
+  }
+}
+
+static const char kColorEncodedHelpMessage[] =
+"This program contains tests written using " GTEST_NAME_ ". You can use the\n"
+"following command line flags to control its behavior:\n"
+"\n"
+"Test Selection:\n"
+"  @G--" GTEST_FLAG_PREFIX_ "list_tests@D\n"
+"      List the names of all tests instead of running them. The name of\n"
+"      TEST(Foo, Bar) is \"Foo.Bar\".\n"
+"  @G--" GTEST_FLAG_PREFIX_ "filter=@YPOSTIVE_PATTERNS"
+    "[@G-@YNEGATIVE_PATTERNS]@D\n"
+"      Run only the tests whose name matches one of the positive patterns but\n"
+"      none of the negative patterns. '?' matches any single character; '*'\n"
+"      matches any substring; ':' separates two patterns.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "also_run_disabled_tests@D\n"
+"      Run all disabled tests too.\n"
+"\n"
+"Test Execution:\n"
+"  @G--" GTEST_FLAG_PREFIX_ "repeat=@Y[COUNT]@D\n"
+"      Run the tests repeatedly; use a negative count to repeat forever.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "shuffle@D\n"
+"      Randomize tests' orders on every iteration.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "random_seed=@Y[NUMBER]@D\n"
+"      Random number seed to use for shuffling test orders (between 1 and\n"
+"      99999, or 0 to use a seed based on the current time).\n"
+"\n"
+"Test Output:\n"
+"  @G--" GTEST_FLAG_PREFIX_ "color=@Y(@Gyes@Y|@Gno@Y|@Gauto@Y)@D\n"
+"      Enable/disable colored output. The default is @Gauto@D.\n"
+"  -@G-" GTEST_FLAG_PREFIX_ "print_time=0@D\n"
+"      Don't print the elapsed time of each test.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "output=xml@Y[@G:@YDIRECTORY_PATH@G"
+    GTEST_PATH_SEP_ "@Y|@G:@YFILE_PATH]@D\n"
+"      Generate an XML report in the given directory or with the given file\n"
+"      name. @YFILE_PATH@D defaults to @Gtest_details.xml@D.\n"
+#if GTEST_CAN_STREAM_RESULTS_
+"  @G--" GTEST_FLAG_PREFIX_ "stream_result_to=@YHOST@G:@YPORT@D\n"
+"      Stream test results to the given server.\n"
+#endif  // GTEST_CAN_STREAM_RESULTS_
+"\n"
+"Assertion Behavior:\n"
+#if GTEST_HAS_DEATH_TEST && !GTEST_OS_WINDOWS
+"  @G--" GTEST_FLAG_PREFIX_ "death_test_style=@Y(@Gfast@Y|@Gthreadsafe@Y)@D\n"
+"      Set the default death test style.\n"
+#endif  // GTEST_HAS_DEATH_TEST && !GTEST_OS_WINDOWS
+"  @G--" GTEST_FLAG_PREFIX_ "break_on_failure@D\n"
+"      Turn assertion failures into debugger break-points.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "throw_on_failure@D\n"
+"      Turn assertion failures into C++ exceptions.\n"
+"  @G--" GTEST_FLAG_PREFIX_ "catch_exceptions=0@D\n"
+"      Do not report exceptions as test failures. Instead, allow them\n"
+"      to crash the program or throw a pop-up (on Windows).\n"
+"\n"
+"Except for @G--" GTEST_FLAG_PREFIX_ "list_tests@D, you can alternatively set "
+    "the corresponding\n"
+"environment variable of a flag (all letters in upper-case). For example, to\n"
+"disable colored text output, you can either specify @G--" GTEST_FLAG_PREFIX_
+    "color=no@D or set\n"
+"the @G" GTEST_FLAG_PREFIX_UPPER_ "COLOR@D environment variable to @Gno@D.\n"
+"\n"
+"For more information, please read the " GTEST_NAME_ " documentation at\n"
+"@G" GTEST_PROJECT_URL_ "@D. If you find a bug in " GTEST_NAME_ "\n"
+"(not one in your own code or tests), please report it to\n"
+"@G<" GTEST_DEV_EMAIL_ ">@D.\n";
+
+bool ParseGoogleTestFlag(const char* const arg) {
+  return ParseBoolFlag(arg, kAlsoRunDisabledTestsFlag,
+                       &GTEST_FLAG(also_run_disabled_tests)) ||
+      ParseBoolFlag(arg, kBreakOnFailureFlag,
+                    &GTEST_FLAG(break_on_failure)) ||
+      ParseBoolFlag(arg, kCatchExceptionsFlag,
+                    &GTEST_FLAG(catch_exceptions)) ||
+      ParseStringFlag(arg, kColorFlag, &GTEST_FLAG(color)) ||
+      ParseStringFlag(arg, kDeathTestStyleFlag,
+                      &GTEST_FLAG(death_test_style)) ||
+      ParseBoolFlag(arg, kDeathTestUseFork,
+                    &GTEST_FLAG(death_test_use_fork)) ||
+      ParseStringFlag(arg, kFilterFlag, &GTEST_FLAG(filter)) ||
+      ParseStringFlag(arg, kInternalRunDeathTestFlag,
+                      &GTEST_FLAG(internal_run_death_test)) ||
+      ParseBoolFlag(arg, kListTestsFlag, &GTEST_FLAG(list_tests)) ||
+      ParseStringFlag(arg, kOutputFlag, &GTEST_FLAG(output)) ||
+      ParseBoolFlag(arg, kPrintTimeFlag, &GTEST_FLAG(print_time)) ||
+      ParseInt32Flag(arg, kRandomSeedFlag, &GTEST_FLAG(random_seed)) ||
+      ParseInt32Flag(arg, kRepeatFlag, &GTEST_FLAG(repeat)) ||
+      ParseBoolFlag(arg, kShuffleFlag, &GTEST_FLAG(shuffle)) ||
+      ParseInt32Flag(arg, kStackTraceDepthFlag,
+                     &GTEST_FLAG(stack_trace_depth)) ||
+      ParseStringFlag(arg, kStreamResultToFlag,
+                      &GTEST_FLAG(stream_result_to)) ||
+      ParseBoolFlag(arg, kThrowOnFailureFlag,
+                    &GTEST_FLAG(throw_on_failure));
+}
+
+#if GTEST_USE_OWN_FLAGFILE_FLAG_
+void LoadFlagsFromFile(const std::string& path) {
+  FILE* flagfile = posix::FOpen(path.c_str(), "r");
+  if (!flagfile) {
+    fprintf(stderr,
+            "Unable to open file \"%s\"\n",
+            GTEST_FLAG(flagfile).c_str());
+    fflush(stderr);
+    exit(EXIT_FAILURE);
+  }
+  std::string contents(ReadEntireFile(flagfile));
+  posix::FClose(flagfile);
+  std::vector<std::string> lines;
+  SplitString(contents, '\n', &lines);
+  for (size_t i = 0; i < lines.size(); ++i) {
+    if (lines[i].empty())
+      continue;
+    if (!ParseGoogleTestFlag(lines[i].c_str()))
+      g_help_flag = true;
+  }
+}
+#endif  // GTEST_USE_OWN_FLAGFILE_FLAG_
+
+// Parses the command line for Google Test flags, without initializing
+// other parts of Google Test.  The type parameter CharType can be
+// instantiated to either char or wchar_t.
+template <typename CharType>
+void ParseGoogleTestFlagsOnlyImpl(int* argc, CharType** argv) {
+  for (int i = 1; i < *argc; i++) {
+    const std::string arg_string = StreamableToString(argv[i]);
+    const char* const arg = arg_string.c_str();
+
+    using internal::ParseBoolFlag;
+    using internal::ParseInt32Flag;
+    using internal::ParseStringFlag;
+
+    bool remove_flag = false;
+    if (ParseGoogleTestFlag(arg)) {
+      remove_flag = true;
+#if GTEST_USE_OWN_FLAGFILE_FLAG_
+    } else if (ParseStringFlag(arg, kFlagfileFlag, &GTEST_FLAG(flagfile))) {
+      LoadFlagsFromFile(GTEST_FLAG(flagfile));
+      remove_flag = true;
+#endif  // GTEST_USE_OWN_FLAGFILE_FLAG_
+    } else if (arg_string == "--help" || arg_string == "-h" ||
+               arg_string == "-?" || arg_string == "/?" ||
+               HasGoogleTestFlagPrefix(arg)) {
+      // Both help flag and unrecognized Google Test flags (excluding
+      // internal ones) trigger help display.
+      g_help_flag = true;
+    }
+
+    if (remove_flag) {
+      // Shift the remainder of the argv list left by one.  Note
+      // that argv has (*argc + 1) elements, the last one always being
+      // NULL.  The following loop moves the trailing NULL element as
+      // well.
+      for (int j = i; j != *argc; j++) {
+        argv[j] = argv[j + 1];
+      }
+
+      // Decrements the argument count.
+      (*argc)--;
+
+      // We also need to decrement the iterator as we just removed
+      // an element.
+      i--;
+    }
+  }
+
+  if (g_help_flag) {
+    // We print the help here instead of in RUN_ALL_TESTS(), as the
+    // latter may not be called at all if the user is using Google
+    // Test with another testing framework.
+    PrintColorEncoded(kColorEncodedHelpMessage);
+  }
+}
+
+// Parses the command line for Google Test flags, without initializing
+// other parts of Google Test.
+void ParseGoogleTestFlagsOnly(int* argc, char** argv) {
+  ParseGoogleTestFlagsOnlyImpl(argc, argv);
+}
+void ParseGoogleTestFlagsOnly(int* argc, wchar_t** argv) {
+  ParseGoogleTestFlagsOnlyImpl(argc, argv);
+}
+
+// The internal implementation of InitGoogleTest().
+//
+// The type parameter CharType can be instantiated to either char or
+// wchar_t.
+template <typename CharType>
+void InitGoogleTestImpl(int* argc, CharType** argv) {
+  // We don't want to run the initialization code twice.
+  if (GTestIsInitialized()) return;
+
+  if (*argc <= 0) return;
+
+  g_argvs.clear();
+  for (int i = 0; i != *argc; i++) {
+    g_argvs.push_back(StreamableToString(argv[i]));
+  }
+
+  ParseGoogleTestFlagsOnly(argc, argv);
+  GetUnitTestImpl()->PostFlagParsingInit();
+}
+
+}  // namespace internal
+
+// Initializes Google Test.  This must be called before calling
+// RUN_ALL_TESTS().  In particular, it parses a command line for the
+// flags that Google Test recognizes.  Whenever a Google Test flag is
+// seen, it is removed from argv, and *argc is decremented.
+//
+// No value is returned.  Instead, the Google Test flag variables are
+// updated.
+//
+// Calling the function for the second time has no user-visible effect.
+void InitGoogleTest(int* argc, char** argv) {
+#if defined(GTEST_CUSTOM_INIT_GOOGLE_TEST_FUNCTION_)
+  GTEST_CUSTOM_INIT_GOOGLE_TEST_FUNCTION_(argc, argv);
+#else  // defined(GTEST_CUSTOM_INIT_GOOGLE_TEST_FUNCTION_)
+  internal::InitGoogleTestImpl(argc, argv);
+#endif  // defined(GTEST_CUSTOM_INIT_GOOGLE_TEST_FUNCTION_)
+}
+
+// This overloaded version can be used in Windows programs compiled in
+// UNICODE mode.
+void InitGoogleTest(int* argc, wchar_t** argv) {
+#if defined(GTEST_CUSTOM_INIT_GOOGLE_TEST_FUNCTION_)
+  GTEST_CUSTOM_INIT_GOOGLE_TEST_FUNCTION_(argc, argv);
+#else  // defined(GTEST_CUSTOM_INIT_GOOGLE_TEST_FUNCTION_)
+  internal::InitGoogleTestImpl(argc, argv);
+#endif  // defined(GTEST_CUSTOM_INIT_GOOGLE_TEST_FUNCTION_)
+}
+
+}  // namespace testing
diff --git a/third_party/aom/third_party/googletest/src/googletest/src/gtest_main.cc b/third_party/aom/third_party/googletest/src/googletest/src/gtest_main.cc
new file mode 100644
index 0000000000..f302822552
--- /dev/null
+++ b/third_party/aom/third_party/googletest/src/googletest/src/gtest_main.cc
@@ -0,0 +1,38 @@
+// Copyright 2006, Google Inc.
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+#include <stdio.h>
+
+#include "gtest/gtest.h"
+
+GTEST_API_ int main(int argc, char **argv) {
+  printf("Running main() from gtest_main.cc\n");
+  testing::InitGoogleTest(&argc, argv);
+  return RUN_ALL_TESTS();
+}
diff --git a/third_party/aom/third_party/libwebm/AUTHORS.TXT b/third_party/aom/third_party/libwebm/AUTHORS.TXT
new file mode 100644
index 0000000000..9686ac13eb
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/AUTHORS.TXT
@@ -0,0 +1,4 @@
+# Names should be added to this file like so:
+# Name or Organization <email address>
+
+Google Inc.
diff --git a/third_party/aom/third_party/libwebm/Android.mk b/third_party/aom/third_party/libwebm/Android.mk
new file mode 100644
index 0000000000..8149a083f4
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/Android.mk
@@ -0,0 +1,17 @@
+LOCAL_PATH:= $(call my-dir)
+
+include $(CLEAR_VARS)
+LOCAL_MODULE:= libwebm
+LOCAL_CPPFLAGS:=-D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS
+LOCAL_CPPFLAGS+=-D__STDC_LIMIT_MACROS -Wno-extern-c-compat
+LOCAL_C_INCLUDES:= $(LOCAL_PATH)
+LOCAL_EXPORT_C_INCLUDES:= $(LOCAL_PATH)
+
+LOCAL_SRC_FILES:= common/file_util.cc \
+                  common/hdr_util.cc \
+                  mkvparser/mkvparser.cc \
+                  mkvparser/mkvreader.cc \
+                  mkvmuxer/mkvmuxer.cc \
+                  mkvmuxer/mkvmuxerutil.cc \
+                  mkvmuxer/mkvwriter.cc
+include $(BUILD_STATIC_LIBRARY)
diff --git a/third_party/aom/third_party/libwebm/LICENSE.TXT b/third_party/aom/third_party/libwebm/LICENSE.TXT
new file mode 100644
index 0000000000..7a6f99547d
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/LICENSE.TXT
@@ -0,0 +1,30 @@
+Copyright (c) 2010, Google Inc. All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+
+  * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+
+  * Redistributions in binary form must reproduce the above copyright
+    notice, this list of conditions and the following disclaimer in
+    the documentation and/or other materials provided with the
+    distribution.
+
+  * Neither the name of Google nor the names of its contributors may
+    be used to endorse or promote products derived from this software
+    without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
diff --git a/third_party/aom/third_party/libwebm/PATENTS.TXT b/third_party/aom/third_party/libwebm/PATENTS.TXT
new file mode 100644
index 0000000000..caedf607e9
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/PATENTS.TXT
@@ -0,0 +1,23 @@
+Additional IP Rights Grant (Patents)
+------------------------------------
+
+"These implementations" means the copyrightable works that implement the WebM
+codecs distributed by Google as part of the WebM Project.
+
+Google hereby grants to you a perpetual, worldwide, non-exclusive, no-charge,
+royalty-free, irrevocable (except as stated in this section) patent license to
+make, have made, use, offer to sell, sell, import, transfer, and otherwise
+run, modify and propagate the contents of these implementations of WebM, where
+such license applies only to those patent claims, both currently owned by
+Google and acquired in the future, licensable by Google that are necessarily
+infringed by these implementations of WebM. This grant does not include claims
+that would be infringed only as a consequence of further modification of these
+implementations. If you or your agent or exclusive licensee institute or order
+or agree to the institution of patent litigation or any other patent
+enforcement activity against any entity (including a cross-claim or
+counterclaim in a lawsuit) alleging that any of these implementations of WebM
+or any code incorporated within any of these implementations of WebM
+constitute direct or contributory patent infringement, or inducement of
+patent infringement, then any patent rights granted to you under this License
+for these implementations of WebM shall terminate as of the date such
+litigation is filed.
diff --git a/third_party/aom/third_party/libwebm/README.libaom b/third_party/aom/third_party/libwebm/README.libaom
new file mode 100644
index 0000000000..73f8303222
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/README.libaom
@@ -0,0 +1,10 @@
+URL: https://chromium.googlesource.com/webm/libwebm
+Version: 32d5ac49414a8914ec1e1f285f3f927c6e8ec29d
+License: BSD
+License File: LICENSE.txt
+
+Description:
+libwebm is used to handle WebM container I/O.
+
+Local Changes:
+* <none>
diff --git a/third_party/aom/third_party/libwebm/common/file_util.cc b/third_party/aom/third_party/libwebm/common/file_util.cc
new file mode 100644
index 0000000000..4f91318f3e
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/common/file_util.cc
@@ -0,0 +1,67 @@
+// Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#include "common/file_util.h"
+
+#include <sys/stat.h>
+#ifndef _MSC_VER
+#include <unistd.h>  // close()
+#endif
+
+#include <cstdio>
+#include <cstdlib>
+#include <fstream>
+#include <ios>
+
+namespace libwebm {
+
+std::string GetTempFileName() {
+#if !defined _MSC_VER && !defined __MINGW32__
+  char temp_file_name_template[] = "libwebm_temp.XXXXXX";
+  int fd = mkstemp(temp_file_name_template);
+  if (fd != -1) {
+    close(fd);
+    return std::string(temp_file_name_template);
+  }
+  return std::string();
+#else
+  char tmp_file_name[_MAX_PATH];
+  errno_t err = tmpnam_s(tmp_file_name);
+  if (err == 0) {
+    return std::string(tmp_file_name);
+  }
+  return std::string();
+#endif
+}
+
+uint64_t GetFileSize(const std::string& file_name) {
+  uint64_t file_size = 0;
+#ifndef _MSC_VER
+  struct stat st;
+  st.st_size = 0;
+  if (stat(file_name.c_str(), &st) == 0) {
+#else
+  struct _stat st;
+  st.st_size = 0;
+  if (_stat(file_name.c_str(), &st) == 0) {
+#endif
+    file_size = st.st_size;
+  }
+  return file_size;
+}
+
+TempFileDeleter::TempFileDeleter() { file_name_ = GetTempFileName(); }
+
+TempFileDeleter::~TempFileDeleter() {
+  std::ifstream file(file_name_.c_str());
+  if (file.good()) {
+    file.close();
+    std::remove(file_name_.c_str());
+  }
+}
+
+}  // namespace libwebm
diff --git a/third_party/aom/third_party/libwebm/common/file_util.h b/third_party/aom/third_party/libwebm/common/file_util.h
new file mode 100644
index 0000000000..0e71eac11e
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/common/file_util.h
@@ -0,0 +1,41 @@
+// Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef LIBWEBM_COMMON_FILE_UTIL_H_
+#define LIBWEBM_COMMON_FILE_UTIL_H_
+
+#include <stdint.h>
+
+#include <string>
+
+#include "mkvmuxer/mkvmuxertypes.h"  // LIBWEBM_DISALLOW_COPY_AND_ASSIGN()
+
+namespace libwebm {
+
+// Returns a temporary file name.
+std::string GetTempFileName();
+
+// Returns size of file specified by |file_name|, or 0 upon failure.
+uint64_t GetFileSize(const std::string& file_name);
+
+// Manages life of temporary file specified at time of construction. Deletes
+// file upon destruction.
+class TempFileDeleter {
+ public:
+  TempFileDeleter();
+  explicit TempFileDeleter(std::string file_name) : file_name_(file_name) {}
+  ~TempFileDeleter();
+  const std::string& name() const { return file_name_; }
+
+ private:
+  std::string file_name_;
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(TempFileDeleter);
+};
+
+}  // namespace libwebm
+
+#endif  // LIBWEBM_COMMON_FILE_UTIL_H_
\ No newline at end of file
diff --git a/third_party/aom/third_party/libwebm/common/hdr_util.cc b/third_party/aom/third_party/libwebm/common/hdr_util.cc
new file mode 100644
index 0000000000..e1a9842fb6
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/common/hdr_util.cc
@@ -0,0 +1,182 @@
+// Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#include "hdr_util.h"
+
+#include <cstddef>
+#include <new>
+
+#include "mkvparser/mkvparser.h"
+
+namespace libwebm {
+bool CopyPrimaryChromaticity(const mkvparser::PrimaryChromaticity& parser_pc,
+                             PrimaryChromaticityPtr* muxer_pc) {
+  muxer_pc->reset(new (std::nothrow)
+                      mkvmuxer::PrimaryChromaticity(parser_pc.x, parser_pc.y));
+  if (!muxer_pc->get())
+    return false;
+  return true;
+}
+
+bool MasteringMetadataValuePresent(double value) {
+  return value != mkvparser::MasteringMetadata::kValueNotPresent;
+}
+
+bool CopyMasteringMetadata(const mkvparser::MasteringMetadata& parser_mm,
+                           mkvmuxer::MasteringMetadata* muxer_mm) {
+  if (MasteringMetadataValuePresent(parser_mm.luminance_max))
+    muxer_mm->luminance_max = parser_mm.luminance_max;
+  if (MasteringMetadataValuePresent(parser_mm.luminance_min))
+    muxer_mm->luminance_min = parser_mm.luminance_min;
+
+  PrimaryChromaticityPtr r_ptr(NULL);
+  PrimaryChromaticityPtr g_ptr(NULL);
+  PrimaryChromaticityPtr b_ptr(NULL);
+  PrimaryChromaticityPtr wp_ptr(NULL);
+
+  if (parser_mm.r) {
+    if (!CopyPrimaryChromaticity(*parser_mm.r, &r_ptr))
+      return false;
+  }
+  if (parser_mm.g) {
+    if (!CopyPrimaryChromaticity(*parser_mm.g, &g_ptr))
+      return false;
+  }
+  if (parser_mm.b) {
+    if (!CopyPrimaryChromaticity(*parser_mm.b, &b_ptr))
+      return false;
+  }
+  if (parser_mm.white_point) {
+    if (!CopyPrimaryChromaticity(*parser_mm.white_point, &wp_ptr))
+      return false;
+  }
+
+  if (!muxer_mm->SetChromaticity(r_ptr.get(), g_ptr.get(), b_ptr.get(),
+                                 wp_ptr.get())) {
+    return false;
+  }
+
+  return true;
+}
+
+bool ColourValuePresent(long long value) {
+  return value != mkvparser::Colour::kValueNotPresent;
+}
+
+bool CopyColour(const mkvparser::Colour& parser_colour,
+                mkvmuxer::Colour* muxer_colour) {
+  if (!muxer_colour)
+    return false;
+
+  if (ColourValuePresent(parser_colour.matrix_coefficients))
+    muxer_colour->matrix_coefficients = parser_colour.matrix_coefficients;
+  if (ColourValuePresent(parser_colour.bits_per_channel))
+    muxer_colour->bits_per_channel = parser_colour.bits_per_channel;
+  if (ColourValuePresent(parser_colour.chroma_subsampling_horz))
+    muxer_colour->chroma_subsampling_horz =
+        parser_colour.chroma_subsampling_horz;
+  if (ColourValuePresent(parser_colour.chroma_subsampling_vert))
+    muxer_colour->chroma_subsampling_vert =
+        parser_colour.chroma_subsampling_vert;
+  if (ColourValuePresent(parser_colour.cb_subsampling_horz))
+    muxer_colour->cb_subsampling_horz = parser_colour.cb_subsampling_horz;
+  if (ColourValuePresent(parser_colour.cb_subsampling_vert))
+    muxer_colour->cb_subsampling_vert = parser_colour.cb_subsampling_vert;
+  if (ColourValuePresent(parser_colour.chroma_siting_horz))
+    muxer_colour->chroma_siting_horz = parser_colour.chroma_siting_horz;
+  if (ColourValuePresent(parser_colour.chroma_siting_vert))
+    muxer_colour->chroma_siting_vert = parser_colour.chroma_siting_vert;
+  if (ColourValuePresent(parser_colour.range))
+    muxer_colour->range = parser_colour.range;
+  if (ColourValuePresent(parser_colour.transfer_characteristics))
+    muxer_colour->transfer_characteristics =
+        parser_colour.transfer_characteristics;
+  if (ColourValuePresent(parser_colour.primaries))
+    muxer_colour->primaries = parser_colour.primaries;
+  if (ColourValuePresent(parser_colour.max_cll))
+    muxer_colour->max_cll = parser_colour.max_cll;
+  if (ColourValuePresent(parser_colour.max_fall))
+    muxer_colour->max_fall = parser_colour.max_fall;
+
+  if (parser_colour.mastering_metadata) {
+    mkvmuxer::MasteringMetadata muxer_mm;
+    if (!CopyMasteringMetadata(*parser_colour.mastering_metadata, &muxer_mm))
+      return false;
+    if (!muxer_colour->SetMasteringMetadata(muxer_mm))
+      return false;
+  }
+  return true;
+}
+
+// Format of VPx private data:
+//
+//   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1
+//  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+//  |    ID Byte    |             Length            |               |
+//  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+               |
+//  |                                                               |
+//  :               Bytes 1..Length of Codec Feature                :
+//  |                                                               |
+//  +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
+//
+// ID Byte Format
+// ID byte is an unsigned byte.
+//   0 1 2 3 4 5 6 7
+//  +-+-+-+-+-+-+-+-+
+//  |X|    ID       |
+//  +-+-+-+-+-+-+-+-+
+//
+// The X bit is reserved.
+//
+// Currently only profile level is supported. ID byte must be set to 1, and
+// length must be 1. Supported values are:
+//
+//   10: Level 1
+//   11: Level 1.1
+//   20: Level 2
+//   21: Level 2.1
+//   30: Level 3
+//   31: Level 3.1
+//   40: Level 4
+//   41: Level 4.1
+//   50: Level 5
+//   51: Level 5.1
+//   52: Level 5.2
+//   60: Level 6
+//   61: Level 6.1
+//   62: Level 6.2
+//
+// See the following link for more information:
+// http://www.webmproject.org/vp9/profiles/
+int ParseVpxCodecPrivate(const uint8_t* private_data, int32_t length) {
+  const int kVpxCodecPrivateLength = 3;
+  if (!private_data || length != kVpxCodecPrivateLength)
+    return 0;
+
+  const uint8_t id_byte = *private_data;
+  if (id_byte != 1)
+    return 0;
+
+  const int kVpxProfileLength = 1;
+  const uint8_t length_byte = private_data[1];
+  if (length_byte != kVpxProfileLength)
+    return 0;
+
+  const int level = static_cast<int>(private_data[2]);
+
+  const int kNumLevels = 14;
+  const int levels[kNumLevels] = {10, 11, 20, 21, 30, 31, 40,
+                                  41, 50, 51, 52, 60, 61, 62};
+
+  for (int i = 0; i < kNumLevels; ++i) {
+    if (level == levels[i])
+      return level;
+  }
+
+  return 0;
+}
+}  // namespace libwebm
diff --git a/third_party/aom/third_party/libwebm/common/hdr_util.h b/third_party/aom/third_party/libwebm/common/hdr_util.h
new file mode 100644
index 0000000000..d30c2b9f2a
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/common/hdr_util.h
@@ -0,0 +1,51 @@
+// Copyright (c) 2016 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef LIBWEBM_COMMON_HDR_UTIL_H_
+#define LIBWEBM_COMMON_HDR_UTIL_H_
+
+#include <stdint.h>
+
+#include <memory>
+
+#include "mkvmuxer/mkvmuxer.h"
+
+namespace mkvparser {
+struct Colour;
+struct MasteringMetadata;
+struct PrimaryChromaticity;
+}  // namespace mkvparser
+
+namespace libwebm {
+// Utility types and functions for working with the Colour element and its
+// children. Copiers return true upon success. Presence functions return true
+// when the specified element is present.
+
+// TODO(tomfinegan): These should be moved to libwebm_utils once c++11 is
+// required by libwebm.
+
+typedef std::auto_ptr<mkvmuxer::PrimaryChromaticity> PrimaryChromaticityPtr;
+
+bool CopyPrimaryChromaticity(const mkvparser::PrimaryChromaticity& parser_pc,
+                             PrimaryChromaticityPtr* muxer_pc);
+
+bool MasteringMetadataValuePresent(double value);
+
+bool CopyMasteringMetadata(const mkvparser::MasteringMetadata& parser_mm,
+                           mkvmuxer::MasteringMetadata* muxer_mm);
+
+bool ColourValuePresent(long long value);
+
+bool CopyColour(const mkvparser::Colour& parser_colour,
+                mkvmuxer::Colour* muxer_colour);
+
+// Returns VP9 profile upon success or 0 upon failure.
+int ParseVpxCodecPrivate(const uint8_t* private_data, int32_t length);
+
+}  // namespace libwebm
+
+#endif  // LIBWEBM_COMMON_HDR_UTIL_H_
diff --git a/third_party/aom/third_party/libwebm/common/webmids.h b/third_party/aom/third_party/libwebm/common/webmids.h
new file mode 100644
index 0000000000..32a0c5fb91
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/common/webmids.h
@@ -0,0 +1,184 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#ifndef COMMON_WEBMIDS_H_
+#define COMMON_WEBMIDS_H_
+
+namespace libwebm {
+
+enum MkvId {
+  kMkvEBML = 0x1A45DFA3,
+  kMkvEBMLVersion = 0x4286,
+  kMkvEBMLReadVersion = 0x42F7,
+  kMkvEBMLMaxIDLength = 0x42F2,
+  kMkvEBMLMaxSizeLength = 0x42F3,
+  kMkvDocType = 0x4282,
+  kMkvDocTypeVersion = 0x4287,
+  kMkvDocTypeReadVersion = 0x4285,
+  kMkvVoid = 0xEC,
+  kMkvSignatureSlot = 0x1B538667,
+  kMkvSignatureAlgo = 0x7E8A,
+  kMkvSignatureHash = 0x7E9A,
+  kMkvSignaturePublicKey = 0x7EA5,
+  kMkvSignature = 0x7EB5,
+  kMkvSignatureElements = 0x7E5B,
+  kMkvSignatureElementList = 0x7E7B,
+  kMkvSignedElement = 0x6532,
+  // segment
+  kMkvSegment = 0x18538067,
+  // Meta Seek Information
+  kMkvSeekHead = 0x114D9B74,
+  kMkvSeek = 0x4DBB,
+  kMkvSeekID = 0x53AB,
+  kMkvSeekPosition = 0x53AC,
+  // Segment Information
+  kMkvInfo = 0x1549A966,
+  kMkvTimecodeScale = 0x2AD7B1,
+  kMkvDuration = 0x4489,
+  kMkvDateUTC = 0x4461,
+  kMkvTitle = 0x7BA9,
+  kMkvMuxingApp = 0x4D80,
+  kMkvWritingApp = 0x5741,
+  // Cluster
+  kMkvCluster = 0x1F43B675,
+  kMkvTimecode = 0xE7,
+  kMkvPrevSize = 0xAB,
+  kMkvBlockGroup = 0xA0,
+  kMkvBlock = 0xA1,
+  kMkvBlockDuration = 0x9B,
+  kMkvReferenceBlock = 0xFB,
+  kMkvLaceNumber = 0xCC,
+  kMkvSimpleBlock = 0xA3,
+  kMkvBlockAdditions = 0x75A1,
+  kMkvBlockMore = 0xA6,
+  kMkvBlockAddID = 0xEE,
+  kMkvBlockAdditional = 0xA5,
+  kMkvDiscardPadding = 0x75A2,
+  // Track
+  kMkvTracks = 0x1654AE6B,
+  kMkvTrackEntry = 0xAE,
+  kMkvTrackNumber = 0xD7,
+  kMkvTrackUID = 0x73C5,
+  kMkvTrackType = 0x83,
+  kMkvFlagEnabled = 0xB9,
+  kMkvFlagDefault = 0x88,
+  kMkvFlagForced = 0x55AA,
+  kMkvFlagLacing = 0x9C,
+  kMkvDefaultDuration = 0x23E383,
+  kMkvMaxBlockAdditionID = 0x55EE,
+  kMkvName = 0x536E,
+  kMkvLanguage = 0x22B59C,
+  kMkvCodecID = 0x86,
+  kMkvCodecPrivate = 0x63A2,
+  kMkvCodecName = 0x258688,
+  kMkvCodecDelay = 0x56AA,
+  kMkvSeekPreRoll = 0x56BB,
+  // video
+  kMkvVideo = 0xE0,
+  kMkvFlagInterlaced = 0x9A,
+  kMkvStereoMode = 0x53B8,
+  kMkvAlphaMode = 0x53C0,
+  kMkvPixelWidth = 0xB0,
+  kMkvPixelHeight = 0xBA,
+  kMkvPixelCropBottom = 0x54AA,
+  kMkvPixelCropTop = 0x54BB,
+  kMkvPixelCropLeft = 0x54CC,
+  kMkvPixelCropRight = 0x54DD,
+  kMkvDisplayWidth = 0x54B0,
+  kMkvDisplayHeight = 0x54BA,
+  kMkvDisplayUnit = 0x54B2,
+  kMkvAspectRatioType = 0x54B3,
+  kMkvFrameRate = 0x2383E3,
+  // end video
+  // colour
+  kMkvColour = 0x55B0,
+  kMkvMatrixCoefficients = 0x55B1,
+  kMkvBitsPerChannel = 0x55B2,
+  kMkvChromaSubsamplingHorz = 0x55B3,
+  kMkvChromaSubsamplingVert = 0x55B4,
+  kMkvCbSubsamplingHorz = 0x55B5,
+  kMkvCbSubsamplingVert = 0x55B6,
+  kMkvChromaSitingHorz = 0x55B7,
+  kMkvChromaSitingVert = 0x55B8,
+  kMkvRange = 0x55B9,
+  kMkvTransferCharacteristics = 0x55BA,
+  kMkvPrimaries = 0x55BB,
+  kMkvMaxCLL = 0x55BC,
+  kMkvMaxFALL = 0x55BD,
+  // mastering metadata
+  kMkvMasteringMetadata = 0x55D0,
+  kMkvPrimaryRChromaticityX = 0x55D1,
+  kMkvPrimaryRChromaticityY = 0x55D2,
+  kMkvPrimaryGChromaticityX = 0x55D3,
+  kMkvPrimaryGChromaticityY = 0x55D4,
+  kMkvPrimaryBChromaticityX = 0x55D5,
+  kMkvPrimaryBChromaticityY = 0x55D6,
+  kMkvWhitePointChromaticityX = 0x55D7,
+  kMkvWhitePointChromaticityY = 0x55D8,
+  kMkvLuminanceMax = 0x55D9,
+  kMkvLuminanceMin = 0x55DA,
+  // end mastering metadata
+  // end colour
+  // audio
+  kMkvAudio = 0xE1,
+  kMkvSamplingFrequency = 0xB5,
+  kMkvOutputSamplingFrequency = 0x78B5,
+  kMkvChannels = 0x9F,
+  kMkvBitDepth = 0x6264,
+  // end audio
+  // ContentEncodings
+  kMkvContentEncodings = 0x6D80,
+  kMkvContentEncoding = 0x6240,
+  kMkvContentEncodingOrder = 0x5031,
+  kMkvContentEncodingScope = 0x5032,
+  kMkvContentEncodingType = 0x5033,
+  kMkvContentCompression = 0x5034,
+  kMkvContentCompAlgo = 0x4254,
+  kMkvContentCompSettings = 0x4255,
+  kMkvContentEncryption = 0x5035,
+  kMkvContentEncAlgo = 0x47E1,
+  kMkvContentEncKeyID = 0x47E2,
+  kMkvContentSignature = 0x47E3,
+  kMkvContentSigKeyID = 0x47E4,
+  kMkvContentSigAlgo = 0x47E5,
+  kMkvContentSigHashAlgo = 0x47E6,
+  kMkvContentEncAESSettings = 0x47E7,
+  kMkvAESSettingsCipherMode = 0x47E8,
+  kMkvAESSettingsCipherInitData = 0x47E9,
+  // end ContentEncodings
+  // Cueing Data
+  kMkvCues = 0x1C53BB6B,
+  kMkvCuePoint = 0xBB,
+  kMkvCueTime = 0xB3,
+  kMkvCueTrackPositions = 0xB7,
+  kMkvCueTrack = 0xF7,
+  kMkvCueClusterPosition = 0xF1,
+  kMkvCueBlockNumber = 0x5378,
+  // Chapters
+  kMkvChapters = 0x1043A770,
+  kMkvEditionEntry = 0x45B9,
+  kMkvChapterAtom = 0xB6,
+  kMkvChapterUID = 0x73C4,
+  kMkvChapterStringUID = 0x5654,
+  kMkvChapterTimeStart = 0x91,
+  kMkvChapterTimeEnd = 0x92,
+  kMkvChapterDisplay = 0x80,
+  kMkvChapString = 0x85,
+  kMkvChapLanguage = 0x437C,
+  kMkvChapCountry = 0x437E,
+  // Tags
+  kMkvTags = 0x1254C367,
+  kMkvTag = 0x7373,
+  kMkvSimpleTag = 0x67C8,
+  kMkvTagName = 0x45A3,
+  kMkvTagString = 0x4487
+};
+
+}  // namespace libwebm
+
+#endif  // COMMON_WEBMIDS_H_
diff --git a/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxer.cc b/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxer.cc
new file mode 100644
index 0000000000..689b5ae427
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxer.cc
@@ -0,0 +1,3769 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#include "mkvmuxer/mkvmuxer.h"
+
+#include <cfloat>
+#include <climits>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <ctime>
+#include <memory>
+#include <new>
+#include <vector>
+
+#include "common/webmids.h"
+#include "mkvmuxer/mkvmuxerutil.h"
+#include "mkvmuxer/mkvwriter.h"
+#include "mkvparser/mkvparser.h"
+
+namespace mkvmuxer {
+
+const float MasteringMetadata::kValueNotPresent = FLT_MAX;
+const uint64_t Colour::kValueNotPresent = UINT64_MAX;
+
+namespace {
+// Deallocate the string designated by |dst|, and then copy the |src|
+// string to |dst|.  The caller owns both the |src| string and the
+// |dst| copy (hence the caller is responsible for eventually
+// deallocating the strings, either directly, or indirectly via
+// StrCpy).  Returns true if the source string was successfully copied
+// to the destination.
+bool StrCpy(const char* src, char** dst_ptr) {
+  if (dst_ptr == NULL)
+    return false;
+
+  char*& dst = *dst_ptr;
+
+  delete[] dst;
+  dst = NULL;
+
+  if (src == NULL)
+    return true;
+
+  const size_t size = strlen(src) + 1;
+
+  dst = new (std::nothrow) char[size];  // NOLINT
+  if (dst == NULL)
+    return false;
+
+  strcpy(dst, src);  // NOLINT
+  return true;
+}
+
+typedef std::auto_ptr<PrimaryChromaticity> PrimaryChromaticityPtr;
+bool CopyChromaticity(const PrimaryChromaticity* src,
+                      PrimaryChromaticityPtr* dst) {
+  if (!dst)
+    return false;
+
+  dst->reset(new (std::nothrow) PrimaryChromaticity(src->x, src->y));
+  if (!dst->get())
+    return false;
+
+  return true;
+}
+
+}  // namespace
+
+///////////////////////////////////////////////////////////////
+//
+// IMkvWriter Class
+
+IMkvWriter::IMkvWriter() {}
+
+IMkvWriter::~IMkvWriter() {}
+
+bool WriteEbmlHeader(IMkvWriter* writer, uint64_t doc_type_version) {
+  // Level 0
+  uint64_t size = EbmlElementSize(libwebm::kMkvEBMLVersion, UINT64_C(1));
+  size += EbmlElementSize(libwebm::kMkvEBMLReadVersion, UINT64_C(1));
+  size += EbmlElementSize(libwebm::kMkvEBMLMaxIDLength, UINT64_C(4));
+  size += EbmlElementSize(libwebm::kMkvEBMLMaxSizeLength, UINT64_C(8));
+  size += EbmlElementSize(libwebm::kMkvDocType, "webm");
+  size += EbmlElementSize(libwebm::kMkvDocTypeVersion, doc_type_version);
+  size += EbmlElementSize(libwebm::kMkvDocTypeReadVersion, UINT64_C(2));
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvEBML, size))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvEBMLVersion, UINT64_C(1)))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvEBMLReadVersion, UINT64_C(1)))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvEBMLMaxIDLength, UINT64_C(4)))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvEBMLMaxSizeLength, UINT64_C(8)))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvDocType, "webm"))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvDocTypeVersion, doc_type_version))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvDocTypeReadVersion, UINT64_C(2)))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlHeader(IMkvWriter* writer) {
+  return WriteEbmlHeader(writer, mkvmuxer::Segment::kDefaultDocTypeVersion);
+}
+
+bool ChunkedCopy(mkvparser::IMkvReader* source, mkvmuxer::IMkvWriter* dst,
+                 int64_t start, int64_t size) {
+  // TODO(vigneshv): Check if this is a reasonable value.
+  const uint32_t kBufSize = 2048;
+  uint8_t* buf = new uint8_t[kBufSize];
+  int64_t offset = start;
+  while (size > 0) {
+    const int64_t read_len = (size > kBufSize) ? kBufSize : size;
+    if (source->Read(offset, static_cast<long>(read_len), buf))
+      return false;
+    dst->Write(buf, static_cast<uint32_t>(read_len));
+    offset += read_len;
+    size -= read_len;
+  }
+  delete[] buf;
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Frame Class
+
+Frame::Frame()
+    : add_id_(0),
+      additional_(NULL),
+      additional_length_(0),
+      duration_(0),
+      duration_set_(false),
+      frame_(NULL),
+      is_key_(false),
+      length_(0),
+      track_number_(0),
+      timestamp_(0),
+      discard_padding_(0),
+      reference_block_timestamp_(0),
+      reference_block_timestamp_set_(false) {}
+
+Frame::~Frame() {
+  delete[] frame_;
+  delete[] additional_;
+}
+
+bool Frame::CopyFrom(const Frame& frame) {
+  delete[] frame_;
+  frame_ = NULL;
+  length_ = 0;
+  if (frame.length() > 0 && frame.frame() != NULL &&
+      !Init(frame.frame(), frame.length())) {
+    return false;
+  }
+  add_id_ = 0;
+  delete[] additional_;
+  additional_ = NULL;
+  additional_length_ = 0;
+  if (frame.additional_length() > 0 && frame.additional() != NULL &&
+      !AddAdditionalData(frame.additional(), frame.additional_length(),
+                         frame.add_id())) {
+    return false;
+  }
+  duration_ = frame.duration();
+  duration_set_ = frame.duration_set();
+  is_key_ = frame.is_key();
+  track_number_ = frame.track_number();
+  timestamp_ = frame.timestamp();
+  discard_padding_ = frame.discard_padding();
+  reference_block_timestamp_ = frame.reference_block_timestamp();
+  reference_block_timestamp_set_ = frame.reference_block_timestamp_set();
+  return true;
+}
+
+bool Frame::Init(const uint8_t* frame, uint64_t length) {
+  uint8_t* const data =
+      new (std::nothrow) uint8_t[static_cast<size_t>(length)];  // NOLINT
+  if (!data)
+    return false;
+
+  delete[] frame_;
+  frame_ = data;
+  length_ = length;
+
+  memcpy(frame_, frame, static_cast<size_t>(length_));
+  return true;
+}
+
+bool Frame::AddAdditionalData(const uint8_t* additional, uint64_t length,
+                              uint64_t add_id) {
+  uint8_t* const data =
+      new (std::nothrow) uint8_t[static_cast<size_t>(length)];  // NOLINT
+  if (!data)
+    return false;
+
+  delete[] additional_;
+  additional_ = data;
+  additional_length_ = length;
+  add_id_ = add_id;
+
+  memcpy(additional_, additional, static_cast<size_t>(additional_length_));
+  return true;
+}
+
+bool Frame::IsValid() const {
+  if (length_ == 0 || !frame_) {
+    return false;
+  }
+  if ((additional_length_ != 0 && !additional_) ||
+      (additional_ != NULL && additional_length_ == 0)) {
+    return false;
+  }
+  if (track_number_ == 0 || track_number_ > kMaxTrackNumber) {
+    return false;
+  }
+  if (!CanBeSimpleBlock() && !is_key_ && !reference_block_timestamp_set_) {
+    return false;
+  }
+  return true;
+}
+
+bool Frame::CanBeSimpleBlock() const {
+  return additional_ == NULL && discard_padding_ == 0 && duration_ == 0;
+}
+
+void Frame::set_duration(uint64_t duration) {
+  duration_ = duration;
+  duration_set_ = true;
+}
+
+void Frame::set_reference_block_timestamp(int64_t reference_block_timestamp) {
+  reference_block_timestamp_ = reference_block_timestamp;
+  reference_block_timestamp_set_ = true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// CuePoint Class
+
+CuePoint::CuePoint()
+    : time_(0),
+      track_(0),
+      cluster_pos_(0),
+      block_number_(1),
+      output_block_number_(true) {}
+
+CuePoint::~CuePoint() {}
+
+bool CuePoint::Write(IMkvWriter* writer) const {
+  if (!writer || track_ < 1 || cluster_pos_ < 1)
+    return false;
+
+  uint64_t size =
+      EbmlElementSize(libwebm::kMkvCueClusterPosition, cluster_pos_);
+  size += EbmlElementSize(libwebm::kMkvCueTrack, track_);
+  if (output_block_number_ && block_number_ > 1)
+    size += EbmlElementSize(libwebm::kMkvCueBlockNumber, block_number_);
+  const uint64_t track_pos_size =
+      EbmlMasterElementSize(libwebm::kMkvCueTrackPositions, size) + size;
+  const uint64_t payload_size =
+      EbmlElementSize(libwebm::kMkvCueTime, time_) + track_pos_size;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvCuePoint, payload_size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvCueTime, time_))
+    return false;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvCueTrackPositions, size))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvCueTrack, track_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvCueClusterPosition, cluster_pos_))
+    return false;
+  if (output_block_number_ && block_number_ > 1)
+    if (!WriteEbmlElement(writer, libwebm::kMkvCueBlockNumber, block_number_))
+      return false;
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0)
+    return false;
+
+  if (stop_position - payload_position != static_cast<int64_t>(payload_size))
+    return false;
+
+  return true;
+}
+
+uint64_t CuePoint::PayloadSize() const {
+  uint64_t size =
+      EbmlElementSize(libwebm::kMkvCueClusterPosition, cluster_pos_);
+  size += EbmlElementSize(libwebm::kMkvCueTrack, track_);
+  if (output_block_number_ && block_number_ > 1)
+    size += EbmlElementSize(libwebm::kMkvCueBlockNumber, block_number_);
+  const uint64_t track_pos_size =
+      EbmlMasterElementSize(libwebm::kMkvCueTrackPositions, size) + size;
+  const uint64_t payload_size =
+      EbmlElementSize(libwebm::kMkvCueTime, time_) + track_pos_size;
+
+  return payload_size;
+}
+
+uint64_t CuePoint::Size() const {
+  const uint64_t payload_size = PayloadSize();
+  return EbmlMasterElementSize(libwebm::kMkvCuePoint, payload_size) +
+         payload_size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Cues Class
+
+Cues::Cues()
+    : cue_entries_capacity_(0),
+      cue_entries_size_(0),
+      cue_entries_(NULL),
+      output_block_number_(true) {}
+
+Cues::~Cues() {
+  if (cue_entries_) {
+    for (int32_t i = 0; i < cue_entries_size_; ++i) {
+      CuePoint* const cue = cue_entries_[i];
+      delete cue;
+    }
+    delete[] cue_entries_;
+  }
+}
+
+bool Cues::AddCue(CuePoint* cue) {
+  if (!cue)
+    return false;
+
+  if ((cue_entries_size_ + 1) > cue_entries_capacity_) {
+    // Add more CuePoints.
+    const int32_t new_capacity =
+        (!cue_entries_capacity_) ? 2 : cue_entries_capacity_ * 2;
+
+    if (new_capacity < 1)
+      return false;
+
+    CuePoint** const cues =
+        new (std::nothrow) CuePoint*[new_capacity];  // NOLINT
+    if (!cues)
+      return false;
+
+    for (int32_t i = 0; i < cue_entries_size_; ++i) {
+      cues[i] = cue_entries_[i];
+    }
+
+    delete[] cue_entries_;
+
+    cue_entries_ = cues;
+    cue_entries_capacity_ = new_capacity;
+  }
+
+  cue->set_output_block_number(output_block_number_);
+  cue_entries_[cue_entries_size_++] = cue;
+  return true;
+}
+
+CuePoint* Cues::GetCueByIndex(int32_t index) const {
+  if (cue_entries_ == NULL)
+    return NULL;
+
+  if (index >= cue_entries_size_)
+    return NULL;
+
+  return cue_entries_[index];
+}
+
+uint64_t Cues::Size() {
+  uint64_t size = 0;
+  for (int32_t i = 0; i < cue_entries_size_; ++i)
+    size += GetCueByIndex(i)->Size();
+  size += EbmlMasterElementSize(libwebm::kMkvCues, size);
+  return size;
+}
+
+bool Cues::Write(IMkvWriter* writer) const {
+  if (!writer)
+    return false;
+
+  uint64_t size = 0;
+  for (int32_t i = 0; i < cue_entries_size_; ++i) {
+    const CuePoint* const cue = GetCueByIndex(i);
+
+    if (!cue)
+      return false;
+
+    size += cue->Size();
+  }
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvCues, size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  for (int32_t i = 0; i < cue_entries_size_; ++i) {
+    const CuePoint* const cue = GetCueByIndex(i);
+
+    if (!cue->Write(writer))
+      return false;
+  }
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0)
+    return false;
+
+  if (stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// ContentEncAESSettings Class
+
+ContentEncAESSettings::ContentEncAESSettings() : cipher_mode_(kCTR) {}
+
+uint64_t ContentEncAESSettings::Size() const {
+  const uint64_t payload = PayloadSize();
+  const uint64_t size =
+      EbmlMasterElementSize(libwebm::kMkvContentEncAESSettings, payload) +
+      payload;
+  return size;
+}
+
+bool ContentEncAESSettings::Write(IMkvWriter* writer) const {
+  const uint64_t payload = PayloadSize();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvContentEncAESSettings,
+                              payload))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvAESSettingsCipherMode,
+                        cipher_mode_))
+    return false;
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(payload))
+    return false;
+
+  return true;
+}
+
+uint64_t ContentEncAESSettings::PayloadSize() const {
+  uint64_t size =
+      EbmlElementSize(libwebm::kMkvAESSettingsCipherMode, cipher_mode_);
+  return size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// ContentEncoding Class
+
+ContentEncoding::ContentEncoding()
+    : enc_algo_(5),
+      enc_key_id_(NULL),
+      encoding_order_(0),
+      encoding_scope_(1),
+      encoding_type_(1),
+      enc_key_id_length_(0) {}
+
+ContentEncoding::~ContentEncoding() { delete[] enc_key_id_; }
+
+bool ContentEncoding::SetEncryptionID(const uint8_t* id, uint64_t length) {
+  if (!id || length < 1)
+    return false;
+
+  delete[] enc_key_id_;
+
+  enc_key_id_ =
+      new (std::nothrow) uint8_t[static_cast<size_t>(length)];  // NOLINT
+  if (!enc_key_id_)
+    return false;
+
+  memcpy(enc_key_id_, id, static_cast<size_t>(length));
+  enc_key_id_length_ = length;
+
+  return true;
+}
+
+uint64_t ContentEncoding::Size() const {
+  const uint64_t encryption_size = EncryptionSize();
+  const uint64_t encoding_size = EncodingSize(0, encryption_size);
+  const uint64_t encodings_size =
+      EbmlMasterElementSize(libwebm::kMkvContentEncoding, encoding_size) +
+      encoding_size;
+
+  return encodings_size;
+}
+
+bool ContentEncoding::Write(IMkvWriter* writer) const {
+  const uint64_t encryption_size = EncryptionSize();
+  const uint64_t encoding_size = EncodingSize(0, encryption_size);
+  const uint64_t size =
+      EbmlMasterElementSize(libwebm::kMkvContentEncoding, encoding_size) +
+      encoding_size;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvContentEncoding,
+                              encoding_size))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvContentEncodingOrder,
+                        encoding_order_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvContentEncodingScope,
+                        encoding_scope_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvContentEncodingType,
+                        encoding_type_))
+    return false;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvContentEncryption,
+                              encryption_size))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvContentEncAlgo, enc_algo_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvContentEncKeyID, enc_key_id_,
+                        enc_key_id_length_))
+    return false;
+
+  if (!enc_aes_settings_.Write(writer))
+    return false;
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  return true;
+}
+
+uint64_t ContentEncoding::EncodingSize(uint64_t compresion_size,
+                                       uint64_t encryption_size) const {
+  // TODO(fgalligan): Add support for compression settings.
+  if (compresion_size != 0)
+    return 0;
+
+  uint64_t encoding_size = 0;
+
+  if (encryption_size > 0) {
+    encoding_size +=
+        EbmlMasterElementSize(libwebm::kMkvContentEncryption, encryption_size) +
+        encryption_size;
+  }
+  encoding_size +=
+      EbmlElementSize(libwebm::kMkvContentEncodingType, encoding_type_);
+  encoding_size +=
+      EbmlElementSize(libwebm::kMkvContentEncodingScope, encoding_scope_);
+  encoding_size +=
+      EbmlElementSize(libwebm::kMkvContentEncodingOrder, encoding_order_);
+
+  return encoding_size;
+}
+
+uint64_t ContentEncoding::EncryptionSize() const {
+  const uint64_t aes_size = enc_aes_settings_.Size();
+
+  uint64_t encryption_size = EbmlElementSize(libwebm::kMkvContentEncKeyID,
+                                             enc_key_id_, enc_key_id_length_);
+  encryption_size += EbmlElementSize(libwebm::kMkvContentEncAlgo, enc_algo_);
+
+  return encryption_size + aes_size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Track Class
+
+Track::Track(unsigned int* seed)
+    : codec_id_(NULL),
+      codec_private_(NULL),
+      language_(NULL),
+      max_block_additional_id_(0),
+      name_(NULL),
+      number_(0),
+      type_(0),
+      uid_(MakeUID(seed)),
+      codec_delay_(0),
+      seek_pre_roll_(0),
+      default_duration_(0),
+      codec_private_length_(0),
+      content_encoding_entries_(NULL),
+      content_encoding_entries_size_(0) {}
+
+Track::~Track() {
+  delete[] codec_id_;
+  delete[] codec_private_;
+  delete[] language_;
+  delete[] name_;
+
+  if (content_encoding_entries_) {
+    for (uint32_t i = 0; i < content_encoding_entries_size_; ++i) {
+      ContentEncoding* const encoding = content_encoding_entries_[i];
+      delete encoding;
+    }
+    delete[] content_encoding_entries_;
+  }
+}
+
+bool Track::AddContentEncoding() {
+  const uint32_t count = content_encoding_entries_size_ + 1;
+
+  ContentEncoding** const content_encoding_entries =
+      new (std::nothrow) ContentEncoding*[count];  // NOLINT
+  if (!content_encoding_entries)
+    return false;
+
+  ContentEncoding* const content_encoding =
+      new (std::nothrow) ContentEncoding();  // NOLINT
+  if (!content_encoding) {
+    delete[] content_encoding_entries;
+    return false;
+  }
+
+  for (uint32_t i = 0; i < content_encoding_entries_size_; ++i) {
+    content_encoding_entries[i] = content_encoding_entries_[i];
+  }
+
+  delete[] content_encoding_entries_;
+
+  content_encoding_entries_ = content_encoding_entries;
+  content_encoding_entries_[content_encoding_entries_size_] = content_encoding;
+  content_encoding_entries_size_ = count;
+  return true;
+}
+
+ContentEncoding* Track::GetContentEncodingByIndex(uint32_t index) const {
+  if (content_encoding_entries_ == NULL)
+    return NULL;
+
+  if (index >= content_encoding_entries_size_)
+    return NULL;
+
+  return content_encoding_entries_[index];
+}
+
+uint64_t Track::PayloadSize() const {
+  uint64_t size = EbmlElementSize(libwebm::kMkvTrackNumber, number_);
+  size += EbmlElementSize(libwebm::kMkvTrackUID, uid_);
+  size += EbmlElementSize(libwebm::kMkvTrackType, type_);
+  if (codec_id_)
+    size += EbmlElementSize(libwebm::kMkvCodecID, codec_id_);
+  if (codec_private_)
+    size += EbmlElementSize(libwebm::kMkvCodecPrivate, codec_private_,
+                            codec_private_length_);
+  if (language_)
+    size += EbmlElementSize(libwebm::kMkvLanguage, language_);
+  if (name_)
+    size += EbmlElementSize(libwebm::kMkvName, name_);
+  if (max_block_additional_id_)
+    size += EbmlElementSize(libwebm::kMkvMaxBlockAdditionID,
+                            max_block_additional_id_);
+  if (codec_delay_)
+    size += EbmlElementSize(libwebm::kMkvCodecDelay, codec_delay_);
+  if (seek_pre_roll_)
+    size += EbmlElementSize(libwebm::kMkvSeekPreRoll, seek_pre_roll_);
+  if (default_duration_)
+    size += EbmlElementSize(libwebm::kMkvDefaultDuration, default_duration_);
+
+  if (content_encoding_entries_size_ > 0) {
+    uint64_t content_encodings_size = 0;
+    for (uint32_t i = 0; i < content_encoding_entries_size_; ++i) {
+      ContentEncoding* const encoding = content_encoding_entries_[i];
+      content_encodings_size += encoding->Size();
+    }
+
+    size += EbmlMasterElementSize(libwebm::kMkvContentEncodings,
+                                  content_encodings_size) +
+            content_encodings_size;
+  }
+
+  return size;
+}
+
+uint64_t Track::Size() const {
+  uint64_t size = PayloadSize();
+  size += EbmlMasterElementSize(libwebm::kMkvTrackEntry, size);
+  return size;
+}
+
+bool Track::Write(IMkvWriter* writer) const {
+  if (!writer)
+    return false;
+
+  // mandatory elements without a default value.
+  if (!type_ || !codec_id_)
+    return false;
+
+  // |size| may be bigger than what is written out in this function because
+  // derived classes may write out more data in the Track element.
+  const uint64_t payload_size = PayloadSize();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvTrackEntry, payload_size))
+    return false;
+
+  uint64_t size = EbmlElementSize(libwebm::kMkvTrackNumber, number_);
+  size += EbmlElementSize(libwebm::kMkvTrackUID, uid_);
+  size += EbmlElementSize(libwebm::kMkvTrackType, type_);
+  if (codec_id_)
+    size += EbmlElementSize(libwebm::kMkvCodecID, codec_id_);
+  if (codec_private_)
+    size += EbmlElementSize(libwebm::kMkvCodecPrivate, codec_private_,
+                            codec_private_length_);
+  if (language_)
+    size += EbmlElementSize(libwebm::kMkvLanguage, language_);
+  if (name_)
+    size += EbmlElementSize(libwebm::kMkvName, name_);
+  if (max_block_additional_id_)
+    size += EbmlElementSize(libwebm::kMkvMaxBlockAdditionID,
+                            max_block_additional_id_);
+  if (codec_delay_)
+    size += EbmlElementSize(libwebm::kMkvCodecDelay, codec_delay_);
+  if (seek_pre_roll_)
+    size += EbmlElementSize(libwebm::kMkvSeekPreRoll, seek_pre_roll_);
+  if (default_duration_)
+    size += EbmlElementSize(libwebm::kMkvDefaultDuration, default_duration_);
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvTrackNumber, number_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvTrackUID, uid_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvTrackType, type_))
+    return false;
+  if (max_block_additional_id_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvMaxBlockAdditionID,
+                          max_block_additional_id_)) {
+      return false;
+    }
+  }
+  if (codec_delay_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvCodecDelay, codec_delay_))
+      return false;
+  }
+  if (seek_pre_roll_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvSeekPreRoll, seek_pre_roll_))
+      return false;
+  }
+  if (default_duration_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvDefaultDuration,
+                          default_duration_))
+      return false;
+  }
+  if (codec_id_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvCodecID, codec_id_))
+      return false;
+  }
+  if (codec_private_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvCodecPrivate, codec_private_,
+                          codec_private_length_))
+      return false;
+  }
+  if (language_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvLanguage, language_))
+      return false;
+  }
+  if (name_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvName, name_))
+      return false;
+  }
+
+  int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  if (content_encoding_entries_size_ > 0) {
+    uint64_t content_encodings_size = 0;
+    for (uint32_t i = 0; i < content_encoding_entries_size_; ++i) {
+      ContentEncoding* const encoding = content_encoding_entries_[i];
+      content_encodings_size += encoding->Size();
+    }
+
+    if (!WriteEbmlMasterElement(writer, libwebm::kMkvContentEncodings,
+                                content_encodings_size))
+      return false;
+
+    for (uint32_t i = 0; i < content_encoding_entries_size_; ++i) {
+      ContentEncoding* const encoding = content_encoding_entries_[i];
+      if (!encoding->Write(writer))
+        return false;
+    }
+  }
+
+  stop_position = writer->Position();
+  if (stop_position < 0)
+    return false;
+  return true;
+}
+
+bool Track::SetCodecPrivate(const uint8_t* codec_private, uint64_t length) {
+  if (!codec_private || length < 1)
+    return false;
+
+  delete[] codec_private_;
+
+  codec_private_ =
+      new (std::nothrow) uint8_t[static_cast<size_t>(length)];  // NOLINT
+  if (!codec_private_)
+    return false;
+
+  memcpy(codec_private_, codec_private, static_cast<size_t>(length));
+  codec_private_length_ = length;
+
+  return true;
+}
+
+void Track::set_codec_id(const char* codec_id) {
+  if (codec_id) {
+    delete[] codec_id_;
+
+    const size_t length = strlen(codec_id) + 1;
+    codec_id_ = new (std::nothrow) char[length];  // NOLINT
+    if (codec_id_) {
+#ifdef _MSC_VER
+      strcpy_s(codec_id_, length, codec_id);
+#else
+      strcpy(codec_id_, codec_id);
+#endif
+    }
+  }
+}
+
+// TODO(fgalligan): Vet the language parameter.
+void Track::set_language(const char* language) {
+  if (language) {
+    delete[] language_;
+
+    const size_t length = strlen(language) + 1;
+    language_ = new (std::nothrow) char[length];  // NOLINT
+    if (language_) {
+#ifdef _MSC_VER
+      strcpy_s(language_, length, language);
+#else
+      strcpy(language_, language);
+#endif
+    }
+  }
+}
+
+void Track::set_name(const char* name) {
+  if (name) {
+    delete[] name_;
+
+    const size_t length = strlen(name) + 1;
+    name_ = new (std::nothrow) char[length];  // NOLINT
+    if (name_) {
+#ifdef _MSC_VER
+      strcpy_s(name_, length, name);
+#else
+      strcpy(name_, name);
+#endif
+    }
+  }
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Colour and its child elements
+
+uint64_t PrimaryChromaticity::PrimaryChromaticityPayloadSize(
+    libwebm::MkvId x_id, libwebm::MkvId y_id) const {
+  return EbmlElementSize(x_id, x) + EbmlElementSize(y_id, y);
+}
+
+bool PrimaryChromaticity::Write(IMkvWriter* writer, libwebm::MkvId x_id,
+                                libwebm::MkvId y_id) const {
+  return WriteEbmlElement(writer, x_id, x) && WriteEbmlElement(writer, y_id, y);
+}
+
+uint64_t MasteringMetadata::MasteringMetadataSize() const {
+  uint64_t size = PayloadSize();
+
+  if (size > 0)
+    size += EbmlMasterElementSize(libwebm::kMkvMasteringMetadata, size);
+
+  return size;
+}
+
+bool MasteringMetadata::Write(IMkvWriter* writer) const {
+  const uint64_t size = PayloadSize();
+
+  // Don't write an empty element.
+  if (size == 0)
+    return true;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvMasteringMetadata, size))
+    return false;
+  if (luminance_max != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvLuminanceMax, luminance_max)) {
+    return false;
+  }
+  if (luminance_min != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvLuminanceMin, luminance_min)) {
+    return false;
+  }
+  if (r_ &&
+      !r_->Write(writer, libwebm::kMkvPrimaryRChromaticityX,
+                 libwebm::kMkvPrimaryRChromaticityY)) {
+    return false;
+  }
+  if (g_ &&
+      !g_->Write(writer, libwebm::kMkvPrimaryGChromaticityX,
+                 libwebm::kMkvPrimaryGChromaticityY)) {
+    return false;
+  }
+  if (b_ &&
+      !b_->Write(writer, libwebm::kMkvPrimaryBChromaticityX,
+                 libwebm::kMkvPrimaryBChromaticityY)) {
+    return false;
+  }
+  if (white_point_ &&
+      !white_point_->Write(writer, libwebm::kMkvWhitePointChromaticityX,
+                           libwebm::kMkvWhitePointChromaticityY)) {
+    return false;
+  }
+
+  return true;
+}
+
+bool MasteringMetadata::SetChromaticity(
+    const PrimaryChromaticity* r, const PrimaryChromaticity* g,
+    const PrimaryChromaticity* b, const PrimaryChromaticity* white_point) {
+  PrimaryChromaticityPtr r_ptr(NULL);
+  if (r) {
+    if (!CopyChromaticity(r, &r_ptr))
+      return false;
+  }
+  PrimaryChromaticityPtr g_ptr(NULL);
+  if (g) {
+    if (!CopyChromaticity(g, &g_ptr))
+      return false;
+  }
+  PrimaryChromaticityPtr b_ptr(NULL);
+  if (b) {
+    if (!CopyChromaticity(b, &b_ptr))
+      return false;
+  }
+  PrimaryChromaticityPtr wp_ptr(NULL);
+  if (white_point) {
+    if (!CopyChromaticity(white_point, &wp_ptr))
+      return false;
+  }
+
+  r_ = r_ptr.release();
+  g_ = g_ptr.release();
+  b_ = b_ptr.release();
+  white_point_ = wp_ptr.release();
+  return true;
+}
+
+uint64_t MasteringMetadata::PayloadSize() const {
+  uint64_t size = 0;
+
+  if (luminance_max != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvLuminanceMax, luminance_max);
+  if (luminance_min != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvLuminanceMin, luminance_min);
+
+  if (r_) {
+    size += r_->PrimaryChromaticityPayloadSize(
+        libwebm::kMkvPrimaryRChromaticityX, libwebm::kMkvPrimaryRChromaticityY);
+  }
+  if (g_) {
+    size += g_->PrimaryChromaticityPayloadSize(
+        libwebm::kMkvPrimaryGChromaticityX, libwebm::kMkvPrimaryGChromaticityY);
+  }
+  if (b_) {
+    size += b_->PrimaryChromaticityPayloadSize(
+        libwebm::kMkvPrimaryBChromaticityX, libwebm::kMkvPrimaryBChromaticityY);
+  }
+  if (white_point_) {
+    size += white_point_->PrimaryChromaticityPayloadSize(
+        libwebm::kMkvWhitePointChromaticityX,
+        libwebm::kMkvWhitePointChromaticityY);
+  }
+
+  return size;
+}
+
+uint64_t Colour::ColourSize() const {
+  uint64_t size = PayloadSize();
+
+  if (size > 0)
+    size += EbmlMasterElementSize(libwebm::kMkvColour, size);
+
+  return size;
+}
+
+bool Colour::Write(IMkvWriter* writer) const {
+  const uint64_t size = PayloadSize();
+
+  // Don't write an empty element.
+  if (size == 0)
+    return true;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvColour, size))
+    return false;
+
+  if (matrix_coefficients != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvMatrixCoefficients,
+                        matrix_coefficients)) {
+    return false;
+  }
+  if (bits_per_channel != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvBitsPerChannel,
+                        bits_per_channel)) {
+    return false;
+  }
+  if (chroma_subsampling_horz != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvChromaSubsamplingHorz,
+                        chroma_subsampling_horz)) {
+    return false;
+  }
+  if (chroma_subsampling_vert != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvChromaSubsamplingVert,
+                        chroma_subsampling_vert)) {
+    return false;
+  }
+
+  if (cb_subsampling_horz != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvCbSubsamplingHorz,
+                        cb_subsampling_horz)) {
+    return false;
+  }
+  if (cb_subsampling_vert != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvCbSubsamplingVert,
+                        cb_subsampling_vert)) {
+    return false;
+  }
+  if (chroma_siting_horz != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvChromaSitingHorz,
+                        chroma_siting_horz)) {
+    return false;
+  }
+  if (chroma_siting_vert != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvChromaSitingVert,
+                        chroma_siting_vert)) {
+    return false;
+  }
+  if (range != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvRange, range)) {
+    return false;
+  }
+  if (transfer_characteristics != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvTransferCharacteristics,
+                        transfer_characteristics)) {
+    return false;
+  }
+  if (primaries != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvPrimaries, primaries)) {
+    return false;
+  }
+  if (max_cll != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvMaxCLL, max_cll)) {
+    return false;
+  }
+  if (max_fall != kValueNotPresent &&
+      !WriteEbmlElement(writer, libwebm::kMkvMaxFALL, max_fall)) {
+    return false;
+  }
+
+  if (mastering_metadata_ && !mastering_metadata_->Write(writer))
+    return false;
+
+  return true;
+}
+
+bool Colour::SetMasteringMetadata(const MasteringMetadata& mastering_metadata) {
+  std::auto_ptr<MasteringMetadata> mm_ptr(new MasteringMetadata());
+  if (!mm_ptr.get())
+    return false;
+
+  mm_ptr->luminance_max = mastering_metadata.luminance_max;
+  mm_ptr->luminance_min = mastering_metadata.luminance_min;
+
+  if (!mm_ptr->SetChromaticity(mastering_metadata.r(), mastering_metadata.g(),
+                               mastering_metadata.b(),
+                               mastering_metadata.white_point())) {
+    return false;
+  }
+
+  delete mastering_metadata_;
+  mastering_metadata_ = mm_ptr.release();
+  return true;
+}
+
+uint64_t Colour::PayloadSize() const {
+  uint64_t size = 0;
+
+  if (matrix_coefficients != kValueNotPresent)
+    size +=
+        EbmlElementSize(libwebm::kMkvMatrixCoefficients, matrix_coefficients);
+  if (bits_per_channel != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvBitsPerChannel, bits_per_channel);
+  if (chroma_subsampling_horz != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvChromaSubsamplingHorz,
+                            chroma_subsampling_horz);
+  if (chroma_subsampling_vert != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvChromaSubsamplingVert,
+                            chroma_subsampling_vert);
+  if (cb_subsampling_horz != kValueNotPresent)
+    size +=
+        EbmlElementSize(libwebm::kMkvCbSubsamplingHorz, cb_subsampling_horz);
+  if (cb_subsampling_vert != kValueNotPresent)
+    size +=
+        EbmlElementSize(libwebm::kMkvCbSubsamplingVert, cb_subsampling_vert);
+  if (chroma_siting_horz != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvChromaSitingHorz, chroma_siting_horz);
+  if (chroma_siting_vert != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvChromaSitingVert, chroma_siting_vert);
+  if (range != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvRange, range);
+  if (transfer_characteristics != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvTransferCharacteristics,
+                            transfer_characteristics);
+  if (primaries != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvPrimaries, primaries);
+  if (max_cll != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvMaxCLL, max_cll);
+  if (max_fall != kValueNotPresent)
+    size += EbmlElementSize(libwebm::kMkvMaxFALL, max_fall);
+
+  if (mastering_metadata_)
+    size += mastering_metadata_->MasteringMetadataSize();
+
+  return size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// VideoTrack Class
+
+VideoTrack::VideoTrack(unsigned int* seed)
+    : Track(seed),
+      display_height_(0),
+      display_width_(0),
+      crop_left_(0),
+      crop_right_(0),
+      crop_top_(0),
+      crop_bottom_(0),
+      frame_rate_(0.0),
+      height_(0),
+      stereo_mode_(0),
+      alpha_mode_(0),
+      width_(0),
+      colour_(NULL) {}
+
+VideoTrack::~VideoTrack() { delete colour_; }
+
+bool VideoTrack::SetStereoMode(uint64_t stereo_mode) {
+  if (stereo_mode != kMono && stereo_mode != kSideBySideLeftIsFirst &&
+      stereo_mode != kTopBottomRightIsFirst &&
+      stereo_mode != kTopBottomLeftIsFirst &&
+      stereo_mode != kSideBySideRightIsFirst)
+    return false;
+
+  stereo_mode_ = stereo_mode;
+  return true;
+}
+
+bool VideoTrack::SetAlphaMode(uint64_t alpha_mode) {
+  if (alpha_mode != kNoAlpha && alpha_mode != kAlpha)
+    return false;
+
+  alpha_mode_ = alpha_mode;
+  return true;
+}
+
+uint64_t VideoTrack::PayloadSize() const {
+  const uint64_t parent_size = Track::PayloadSize();
+
+  uint64_t size = VideoPayloadSize();
+  size += EbmlMasterElementSize(libwebm::kMkvVideo, size);
+
+  return parent_size + size;
+}
+
+bool VideoTrack::Write(IMkvWriter* writer) const {
+  if (!Track::Write(writer))
+    return false;
+
+  const uint64_t size = VideoPayloadSize();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvVideo, size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvPixelWidth, width_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvPixelHeight, height_))
+    return false;
+  if (display_width_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvDisplayWidth, display_width_))
+      return false;
+  }
+  if (display_height_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvDisplayHeight, display_height_))
+      return false;
+  }
+  if (crop_left_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvPixelCropLeft, crop_left_))
+      return false;
+  }
+  if (crop_right_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvPixelCropRight, crop_right_))
+      return false;
+  }
+  if (crop_top_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvPixelCropTop, crop_top_))
+      return false;
+  }
+  if (crop_bottom_ > 0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvPixelCropBottom, crop_bottom_))
+      return false;
+  }
+  if (stereo_mode_ > kMono) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvStereoMode, stereo_mode_))
+      return false;
+  }
+  if (alpha_mode_ > kNoAlpha) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvAlphaMode, alpha_mode_))
+      return false;
+  }
+  if (frame_rate_ > 0.0) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvFrameRate,
+                          static_cast<float>(frame_rate_))) {
+      return false;
+    }
+  }
+  if (colour_) {
+    if (!colour_->Write(writer))
+      return false;
+  }
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size)) {
+    return false;
+  }
+
+  return true;
+}
+
+bool VideoTrack::SetColour(const Colour& colour) {
+  std::auto_ptr<Colour> colour_ptr(new Colour());
+  if (!colour_ptr.get())
+    return false;
+
+  if (colour.mastering_metadata()) {
+    if (!colour_ptr->SetMasteringMetadata(*colour.mastering_metadata()))
+      return false;
+  }
+
+  colour_ptr->matrix_coefficients = colour.matrix_coefficients;
+  colour_ptr->bits_per_channel = colour.bits_per_channel;
+  colour_ptr->chroma_subsampling_horz = colour.chroma_subsampling_horz;
+  colour_ptr->chroma_subsampling_vert = colour.chroma_subsampling_vert;
+  colour_ptr->cb_subsampling_horz = colour.cb_subsampling_horz;
+  colour_ptr->cb_subsampling_vert = colour.cb_subsampling_vert;
+  colour_ptr->chroma_siting_horz = colour.chroma_siting_horz;
+  colour_ptr->chroma_siting_vert = colour.chroma_siting_vert;
+  colour_ptr->range = colour.range;
+  colour_ptr->transfer_characteristics = colour.transfer_characteristics;
+  colour_ptr->primaries = colour.primaries;
+  colour_ptr->max_cll = colour.max_cll;
+  colour_ptr->max_fall = colour.max_fall;
+  colour_ = colour_ptr.release();
+  return true;
+}
+
+uint64_t VideoTrack::VideoPayloadSize() const {
+  uint64_t size = EbmlElementSize(libwebm::kMkvPixelWidth, width_);
+  size += EbmlElementSize(libwebm::kMkvPixelHeight, height_);
+  if (display_width_ > 0)
+    size += EbmlElementSize(libwebm::kMkvDisplayWidth, display_width_);
+  if (display_height_ > 0)
+    size += EbmlElementSize(libwebm::kMkvDisplayHeight, display_height_);
+  if (crop_left_ > 0)
+    size += EbmlElementSize(libwebm::kMkvPixelCropLeft, crop_left_);
+  if (crop_right_ > 0)
+    size += EbmlElementSize(libwebm::kMkvPixelCropRight, crop_right_);
+  if (crop_top_ > 0)
+    size += EbmlElementSize(libwebm::kMkvPixelCropTop, crop_top_);
+  if (crop_bottom_ > 0)
+    size += EbmlElementSize(libwebm::kMkvPixelCropBottom, crop_bottom_);
+  if (stereo_mode_ > kMono)
+    size += EbmlElementSize(libwebm::kMkvStereoMode, stereo_mode_);
+  if (alpha_mode_ > kNoAlpha)
+    size += EbmlElementSize(libwebm::kMkvAlphaMode, alpha_mode_);
+  if (frame_rate_ > 0.0)
+    size += EbmlElementSize(libwebm::kMkvFrameRate,
+                            static_cast<float>(frame_rate_));
+  if (colour_)
+    size += colour_->ColourSize();
+
+  return size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// AudioTrack Class
+
+AudioTrack::AudioTrack(unsigned int* seed)
+    : Track(seed), bit_depth_(0), channels_(1), sample_rate_(0.0) {}
+
+AudioTrack::~AudioTrack() {}
+
+uint64_t AudioTrack::PayloadSize() const {
+  const uint64_t parent_size = Track::PayloadSize();
+
+  uint64_t size = EbmlElementSize(libwebm::kMkvSamplingFrequency,
+                                  static_cast<float>(sample_rate_));
+  size += EbmlElementSize(libwebm::kMkvChannels, channels_);
+  if (bit_depth_ > 0)
+    size += EbmlElementSize(libwebm::kMkvBitDepth, bit_depth_);
+  size += EbmlMasterElementSize(libwebm::kMkvAudio, size);
+
+  return parent_size + size;
+}
+
+bool AudioTrack::Write(IMkvWriter* writer) const {
+  if (!Track::Write(writer))
+    return false;
+
+  // Calculate AudioSettings size.
+  uint64_t size = EbmlElementSize(libwebm::kMkvSamplingFrequency,
+                                  static_cast<float>(sample_rate_));
+  size += EbmlElementSize(libwebm::kMkvChannels, channels_);
+  if (bit_depth_ > 0)
+    size += EbmlElementSize(libwebm::kMkvBitDepth, bit_depth_);
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvAudio, size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvSamplingFrequency,
+                        static_cast<float>(sample_rate_)))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvChannels, channels_))
+    return false;
+  if (bit_depth_ > 0)
+    if (!WriteEbmlElement(writer, libwebm::kMkvBitDepth, bit_depth_))
+      return false;
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Tracks Class
+
+const char Tracks::kOpusCodecId[] = "A_OPUS";
+const char Tracks::kVorbisCodecId[] = "A_VORBIS";
+const char Tracks::kVp8CodecId[] = "V_VP8";
+const char Tracks::kVp9CodecId[] = "V_VP9";
+const char Tracks::kAv1CodecId[] = "V_AV1";
+
+Tracks::Tracks()
+    : track_entries_(NULL), track_entries_size_(0), wrote_tracks_(false) {}
+
+Tracks::~Tracks() {
+  if (track_entries_) {
+    for (uint32_t i = 0; i < track_entries_size_; ++i) {
+      Track* const track = track_entries_[i];
+      delete track;
+    }
+    delete[] track_entries_;
+  }
+}
+
+bool Tracks::AddTrack(Track* track, int32_t number) {
+  if (number < 0 || wrote_tracks_)
+    return false;
+
+  // This muxer only supports track numbers in the range [1, 126], in
+  // order to be able (to use Matroska integer representation) to
+  // serialize the block header (of which the track number is a part)
+  // for a frame using exactly 4 bytes.
+
+  if (number > 0x7E)
+    return false;
+
+  uint32_t track_num = number;
+
+  if (track_num > 0) {
+    // Check to make sure a track does not already have |track_num|.
+    for (uint32_t i = 0; i < track_entries_size_; ++i) {
+      if (track_entries_[i]->number() == track_num)
+        return false;
+    }
+  }
+
+  const uint32_t count = track_entries_size_ + 1;
+
+  Track** const track_entries = new (std::nothrow) Track*[count];  // NOLINT
+  if (!track_entries)
+    return false;
+
+  for (uint32_t i = 0; i < track_entries_size_; ++i) {
+    track_entries[i] = track_entries_[i];
+  }
+
+  delete[] track_entries_;
+
+  // Find the lowest availible track number > 0.
+  if (track_num == 0) {
+    track_num = count;
+
+    // Check to make sure a track does not already have |track_num|.
+    bool exit = false;
+    do {
+      exit = true;
+      for (uint32_t i = 0; i < track_entries_size_; ++i) {
+        if (track_entries[i]->number() == track_num) {
+          track_num++;
+          exit = false;
+          break;
+        }
+      }
+    } while (!exit);
+  }
+  track->set_number(track_num);
+
+  track_entries_ = track_entries;
+  track_entries_[track_entries_size_] = track;
+  track_entries_size_ = count;
+  return true;
+}
+
+const Track* Tracks::GetTrackByIndex(uint32_t index) const {
+  if (track_entries_ == NULL)
+    return NULL;
+
+  if (index >= track_entries_size_)
+    return NULL;
+
+  return track_entries_[index];
+}
+
+Track* Tracks::GetTrackByNumber(uint64_t track_number) const {
+  const int32_t count = track_entries_size();
+  for (int32_t i = 0; i < count; ++i) {
+    if (track_entries_[i]->number() == track_number)
+      return track_entries_[i];
+  }
+
+  return NULL;
+}
+
+bool Tracks::TrackIsAudio(uint64_t track_number) const {
+  const Track* const track = GetTrackByNumber(track_number);
+
+  if (track->type() == kAudio)
+    return true;
+
+  return false;
+}
+
+bool Tracks::TrackIsVideo(uint64_t track_number) const {
+  const Track* const track = GetTrackByNumber(track_number);
+
+  if (track->type() == kVideo)
+    return true;
+
+  return false;
+}
+
+bool Tracks::Write(IMkvWriter* writer) const {
+  uint64_t size = 0;
+  const int32_t count = track_entries_size();
+  for (int32_t i = 0; i < count; ++i) {
+    const Track* const track = GetTrackByIndex(i);
+
+    if (!track)
+      return false;
+
+    size += track->Size();
+  }
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvTracks, size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  for (int32_t i = 0; i < count; ++i) {
+    const Track* const track = GetTrackByIndex(i);
+    if (!track->Write(writer))
+      return false;
+  }
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  wrote_tracks_ = true;
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Chapter Class
+
+bool Chapter::set_id(const char* id) { return StrCpy(id, &id_); }
+
+void Chapter::set_time(const Segment& segment, uint64_t start_ns,
+                       uint64_t end_ns) {
+  const SegmentInfo* const info = segment.GetSegmentInfo();
+  const uint64_t timecode_scale = info->timecode_scale();
+  start_timecode_ = start_ns / timecode_scale;
+  end_timecode_ = end_ns / timecode_scale;
+}
+
+bool Chapter::add_string(const char* title, const char* language,
+                         const char* country) {
+  if (!ExpandDisplaysArray())
+    return false;
+
+  Display& d = displays_[displays_count_++];
+  d.Init();
+
+  if (!d.set_title(title))
+    return false;
+
+  if (!d.set_language(language))
+    return false;
+
+  if (!d.set_country(country))
+    return false;
+
+  return true;
+}
+
+Chapter::Chapter() {
+  // This ctor only constructs the object.  Proper initialization is
+  // done in Init() (called in Chapters::AddChapter()).  The only
+  // reason we bother implementing this ctor is because we had to
+  // declare it as private (along with the dtor), in order to prevent
+  // clients from creating Chapter instances (a privelege we grant
+  // only to the Chapters class).  Doing no initialization here also
+  // means that creating arrays of chapter objects is more efficient,
+  // because we only initialize each new chapter object as it becomes
+  // active on the array.
+}
+
+Chapter::~Chapter() {}
+
+void Chapter::Init(unsigned int* seed) {
+  id_ = NULL;
+  start_timecode_ = 0;
+  end_timecode_ = 0;
+  displays_ = NULL;
+  displays_size_ = 0;
+  displays_count_ = 0;
+  uid_ = MakeUID(seed);
+}
+
+void Chapter::ShallowCopy(Chapter* dst) const {
+  dst->id_ = id_;
+  dst->start_timecode_ = start_timecode_;
+  dst->end_timecode_ = end_timecode_;
+  dst->uid_ = uid_;
+  dst->displays_ = displays_;
+  dst->displays_size_ = displays_size_;
+  dst->displays_count_ = displays_count_;
+}
+
+void Chapter::Clear() {
+  StrCpy(NULL, &id_);
+
+  while (displays_count_ > 0) {
+    Display& d = displays_[--displays_count_];
+    d.Clear();
+  }
+
+  delete[] displays_;
+  displays_ = NULL;
+
+  displays_size_ = 0;
+}
+
+bool Chapter::ExpandDisplaysArray() {
+  if (displays_size_ > displays_count_)
+    return true;  // nothing to do yet
+
+  const int size = (displays_size_ == 0) ? 1 : 2 * displays_size_;
+
+  Display* const displays = new (std::nothrow) Display[size];  // NOLINT
+  if (displays == NULL)
+    return false;
+
+  for (int idx = 0; idx < displays_count_; ++idx) {
+    displays[idx] = displays_[idx];  // shallow copy
+  }
+
+  delete[] displays_;
+
+  displays_ = displays;
+  displays_size_ = size;
+
+  return true;
+}
+
+uint64_t Chapter::WriteAtom(IMkvWriter* writer) const {
+  uint64_t payload_size =
+      EbmlElementSize(libwebm::kMkvChapterStringUID, id_) +
+      EbmlElementSize(libwebm::kMkvChapterUID, uid_) +
+      EbmlElementSize(libwebm::kMkvChapterTimeStart, start_timecode_) +
+      EbmlElementSize(libwebm::kMkvChapterTimeEnd, end_timecode_);
+
+  for (int idx = 0; idx < displays_count_; ++idx) {
+    const Display& d = displays_[idx];
+    payload_size += d.WriteDisplay(NULL);
+  }
+
+  const uint64_t atom_size =
+      EbmlMasterElementSize(libwebm::kMkvChapterAtom, payload_size) +
+      payload_size;
+
+  if (writer == NULL)
+    return atom_size;
+
+  const int64_t start = writer->Position();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvChapterAtom, payload_size))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvChapterStringUID, id_))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvChapterUID, uid_))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvChapterTimeStart, start_timecode_))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvChapterTimeEnd, end_timecode_))
+    return 0;
+
+  for (int idx = 0; idx < displays_count_; ++idx) {
+    const Display& d = displays_[idx];
+
+    if (!d.WriteDisplay(writer))
+      return 0;
+  }
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != atom_size)
+    return 0;
+
+  return atom_size;
+}
+
+void Chapter::Display::Init() {
+  title_ = NULL;
+  language_ = NULL;
+  country_ = NULL;
+}
+
+void Chapter::Display::Clear() {
+  StrCpy(NULL, &title_);
+  StrCpy(NULL, &language_);
+  StrCpy(NULL, &country_);
+}
+
+bool Chapter::Display::set_title(const char* title) {
+  return StrCpy(title, &title_);
+}
+
+bool Chapter::Display::set_language(const char* language) {
+  return StrCpy(language, &language_);
+}
+
+bool Chapter::Display::set_country(const char* country) {
+  return StrCpy(country, &country_);
+}
+
+uint64_t Chapter::Display::WriteDisplay(IMkvWriter* writer) const {
+  uint64_t payload_size = EbmlElementSize(libwebm::kMkvChapString, title_);
+
+  if (language_)
+    payload_size += EbmlElementSize(libwebm::kMkvChapLanguage, language_);
+
+  if (country_)
+    payload_size += EbmlElementSize(libwebm::kMkvChapCountry, country_);
+
+  const uint64_t display_size =
+      EbmlMasterElementSize(libwebm::kMkvChapterDisplay, payload_size) +
+      payload_size;
+
+  if (writer == NULL)
+    return display_size;
+
+  const int64_t start = writer->Position();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvChapterDisplay,
+                              payload_size))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvChapString, title_))
+    return 0;
+
+  if (language_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvChapLanguage, language_))
+      return 0;
+  }
+
+  if (country_) {
+    if (!WriteEbmlElement(writer, libwebm::kMkvChapCountry, country_))
+      return 0;
+  }
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != display_size)
+    return 0;
+
+  return display_size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Chapters Class
+
+Chapters::Chapters() : chapters_size_(0), chapters_count_(0), chapters_(NULL) {}
+
+Chapters::~Chapters() {
+  while (chapters_count_ > 0) {
+    Chapter& chapter = chapters_[--chapters_count_];
+    chapter.Clear();
+  }
+
+  delete[] chapters_;
+  chapters_ = NULL;
+}
+
+int Chapters::Count() const { return chapters_count_; }
+
+Chapter* Chapters::AddChapter(unsigned int* seed) {
+  if (!ExpandChaptersArray())
+    return NULL;
+
+  Chapter& chapter = chapters_[chapters_count_++];
+  chapter.Init(seed);
+
+  return &chapter;
+}
+
+bool Chapters::Write(IMkvWriter* writer) const {
+  if (writer == NULL)
+    return false;
+
+  const uint64_t payload_size = WriteEdition(NULL);  // return size only
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvChapters, payload_size))
+    return false;
+
+  const int64_t start = writer->Position();
+
+  if (WriteEdition(writer) == 0)  // error
+    return false;
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != payload_size)
+    return false;
+
+  return true;
+}
+
+bool Chapters::ExpandChaptersArray() {
+  if (chapters_size_ > chapters_count_)
+    return true;  // nothing to do yet
+
+  const int size = (chapters_size_ == 0) ? 1 : 2 * chapters_size_;
+
+  Chapter* const chapters = new (std::nothrow) Chapter[size];  // NOLINT
+  if (chapters == NULL)
+    return false;
+
+  for (int idx = 0; idx < chapters_count_; ++idx) {
+    const Chapter& src = chapters_[idx];
+    Chapter* const dst = chapters + idx;
+    src.ShallowCopy(dst);
+  }
+
+  delete[] chapters_;
+
+  chapters_ = chapters;
+  chapters_size_ = size;
+
+  return true;
+}
+
+uint64_t Chapters::WriteEdition(IMkvWriter* writer) const {
+  uint64_t payload_size = 0;
+
+  for (int idx = 0; idx < chapters_count_; ++idx) {
+    const Chapter& chapter = chapters_[idx];
+    payload_size += chapter.WriteAtom(NULL);
+  }
+
+  const uint64_t edition_size =
+      EbmlMasterElementSize(libwebm::kMkvEditionEntry, payload_size) +
+      payload_size;
+
+  if (writer == NULL)  // return size only
+    return edition_size;
+
+  const int64_t start = writer->Position();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvEditionEntry, payload_size))
+    return 0;  // error
+
+  for (int idx = 0; idx < chapters_count_; ++idx) {
+    const Chapter& chapter = chapters_[idx];
+
+    const uint64_t chapter_size = chapter.WriteAtom(writer);
+    if (chapter_size == 0)  // error
+      return 0;
+  }
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != edition_size)
+    return 0;
+
+  return edition_size;
+}
+
+// Tag Class
+
+bool Tag::add_simple_tag(const char* tag_name, const char* tag_string) {
+  if (!ExpandSimpleTagsArray())
+    return false;
+
+  SimpleTag& st = simple_tags_[simple_tags_count_++];
+  st.Init();
+
+  if (!st.set_tag_name(tag_name))
+    return false;
+
+  if (!st.set_tag_string(tag_string))
+    return false;
+
+  return true;
+}
+
+Tag::Tag() {
+  simple_tags_ = NULL;
+  simple_tags_size_ = 0;
+  simple_tags_count_ = 0;
+}
+
+Tag::~Tag() {}
+
+void Tag::ShallowCopy(Tag* dst) const {
+  dst->simple_tags_ = simple_tags_;
+  dst->simple_tags_size_ = simple_tags_size_;
+  dst->simple_tags_count_ = simple_tags_count_;
+}
+
+void Tag::Clear() {
+  while (simple_tags_count_ > 0) {
+    SimpleTag& st = simple_tags_[--simple_tags_count_];
+    st.Clear();
+  }
+
+  delete[] simple_tags_;
+  simple_tags_ = NULL;
+
+  simple_tags_size_ = 0;
+}
+
+bool Tag::ExpandSimpleTagsArray() {
+  if (simple_tags_size_ > simple_tags_count_)
+    return true;  // nothing to do yet
+
+  const int size = (simple_tags_size_ == 0) ? 1 : 2 * simple_tags_size_;
+
+  SimpleTag* const simple_tags = new (std::nothrow) SimpleTag[size];  // NOLINT
+  if (simple_tags == NULL)
+    return false;
+
+  for (int idx = 0; idx < simple_tags_count_; ++idx) {
+    simple_tags[idx] = simple_tags_[idx];  // shallow copy
+  }
+
+  delete[] simple_tags_;
+
+  simple_tags_ = simple_tags;
+  simple_tags_size_ = size;
+
+  return true;
+}
+
+uint64_t Tag::Write(IMkvWriter* writer) const {
+  uint64_t payload_size = 0;
+
+  for (int idx = 0; idx < simple_tags_count_; ++idx) {
+    const SimpleTag& st = simple_tags_[idx];
+    payload_size += st.Write(NULL);
+  }
+
+  const uint64_t tag_size =
+      EbmlMasterElementSize(libwebm::kMkvTag, payload_size) + payload_size;
+
+  if (writer == NULL)
+    return tag_size;
+
+  const int64_t start = writer->Position();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvTag, payload_size))
+    return 0;
+
+  for (int idx = 0; idx < simple_tags_count_; ++idx) {
+    const SimpleTag& st = simple_tags_[idx];
+
+    if (!st.Write(writer))
+      return 0;
+  }
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != tag_size)
+    return 0;
+
+  return tag_size;
+}
+
+// Tag::SimpleTag
+
+void Tag::SimpleTag::Init() {
+  tag_name_ = NULL;
+  tag_string_ = NULL;
+}
+
+void Tag::SimpleTag::Clear() {
+  StrCpy(NULL, &tag_name_);
+  StrCpy(NULL, &tag_string_);
+}
+
+bool Tag::SimpleTag::set_tag_name(const char* tag_name) {
+  return StrCpy(tag_name, &tag_name_);
+}
+
+bool Tag::SimpleTag::set_tag_string(const char* tag_string) {
+  return StrCpy(tag_string, &tag_string_);
+}
+
+uint64_t Tag::SimpleTag::Write(IMkvWriter* writer) const {
+  uint64_t payload_size = EbmlElementSize(libwebm::kMkvTagName, tag_name_);
+
+  payload_size += EbmlElementSize(libwebm::kMkvTagString, tag_string_);
+
+  const uint64_t simple_tag_size =
+      EbmlMasterElementSize(libwebm::kMkvSimpleTag, payload_size) +
+      payload_size;
+
+  if (writer == NULL)
+    return simple_tag_size;
+
+  const int64_t start = writer->Position();
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvSimpleTag, payload_size))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvTagName, tag_name_))
+    return 0;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvTagString, tag_string_))
+    return 0;
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != simple_tag_size)
+    return 0;
+
+  return simple_tag_size;
+}
+
+// Tags Class
+
+Tags::Tags() : tags_size_(0), tags_count_(0), tags_(NULL) {}
+
+Tags::~Tags() {
+  while (tags_count_ > 0) {
+    Tag& tag = tags_[--tags_count_];
+    tag.Clear();
+  }
+
+  delete[] tags_;
+  tags_ = NULL;
+}
+
+int Tags::Count() const { return tags_count_; }
+
+Tag* Tags::AddTag() {
+  if (!ExpandTagsArray())
+    return NULL;
+
+  Tag& tag = tags_[tags_count_++];
+
+  return &tag;
+}
+
+bool Tags::Write(IMkvWriter* writer) const {
+  if (writer == NULL)
+    return false;
+
+  uint64_t payload_size = 0;
+
+  for (int idx = 0; idx < tags_count_; ++idx) {
+    const Tag& tag = tags_[idx];
+    payload_size += tag.Write(NULL);
+  }
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvTags, payload_size))
+    return false;
+
+  const int64_t start = writer->Position();
+
+  for (int idx = 0; idx < tags_count_; ++idx) {
+    const Tag& tag = tags_[idx];
+
+    const uint64_t tag_size = tag.Write(writer);
+    if (tag_size == 0)  // error
+      return 0;
+  }
+
+  const int64_t stop = writer->Position();
+
+  if (stop >= start && uint64_t(stop - start) != payload_size)
+    return false;
+
+  return true;
+}
+
+bool Tags::ExpandTagsArray() {
+  if (tags_size_ > tags_count_)
+    return true;  // nothing to do yet
+
+  const int size = (tags_size_ == 0) ? 1 : 2 * tags_size_;
+
+  Tag* const tags = new (std::nothrow) Tag[size];  // NOLINT
+  if (tags == NULL)
+    return false;
+
+  for (int idx = 0; idx < tags_count_; ++idx) {
+    const Tag& src = tags_[idx];
+    Tag* const dst = tags + idx;
+    src.ShallowCopy(dst);
+  }
+
+  delete[] tags_;
+
+  tags_ = tags;
+  tags_size_ = size;
+
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Cluster class
+
+Cluster::Cluster(uint64_t timecode, int64_t cues_pos, uint64_t timecode_scale,
+                 bool write_last_frame_with_duration, bool fixed_size_timecode)
+    : blocks_added_(0),
+      finalized_(false),
+      fixed_size_timecode_(fixed_size_timecode),
+      header_written_(false),
+      payload_size_(0),
+      position_for_cues_(cues_pos),
+      size_position_(-1),
+      timecode_(timecode),
+      timecode_scale_(timecode_scale),
+      write_last_frame_with_duration_(write_last_frame_with_duration),
+      writer_(NULL) {}
+
+Cluster::~Cluster() {}
+
+bool Cluster::Init(IMkvWriter* ptr_writer) {
+  if (!ptr_writer) {
+    return false;
+  }
+  writer_ = ptr_writer;
+  return true;
+}
+
+bool Cluster::AddFrame(const Frame* const frame) {
+  return QueueOrWriteFrame(frame);
+}
+
+bool Cluster::AddFrame(const uint8_t* data, uint64_t length,
+                       uint64_t track_number, uint64_t abs_timecode,
+                       bool is_key) {
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_track_number(track_number);
+  frame.set_timestamp(abs_timecode);
+  frame.set_is_key(is_key);
+  return QueueOrWriteFrame(&frame);
+}
+
+bool Cluster::AddFrameWithAdditional(const uint8_t* data, uint64_t length,
+                                     const uint8_t* additional,
+                                     uint64_t additional_length,
+                                     uint64_t add_id, uint64_t track_number,
+                                     uint64_t abs_timecode, bool is_key) {
+  if (!additional || additional_length == 0) {
+    return false;
+  }
+  Frame frame;
+  if (!frame.Init(data, length) ||
+      !frame.AddAdditionalData(additional, additional_length, add_id)) {
+    return false;
+  }
+  frame.set_track_number(track_number);
+  frame.set_timestamp(abs_timecode);
+  frame.set_is_key(is_key);
+  return QueueOrWriteFrame(&frame);
+}
+
+bool Cluster::AddFrameWithDiscardPadding(const uint8_t* data, uint64_t length,
+                                         int64_t discard_padding,
+                                         uint64_t track_number,
+                                         uint64_t abs_timecode, bool is_key) {
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_discard_padding(discard_padding);
+  frame.set_track_number(track_number);
+  frame.set_timestamp(abs_timecode);
+  frame.set_is_key(is_key);
+  return QueueOrWriteFrame(&frame);
+}
+
+bool Cluster::AddMetadata(const uint8_t* data, uint64_t length,
+                          uint64_t track_number, uint64_t abs_timecode,
+                          uint64_t duration_timecode) {
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_track_number(track_number);
+  frame.set_timestamp(abs_timecode);
+  frame.set_duration(duration_timecode);
+  frame.set_is_key(true);  // All metadata blocks are keyframes.
+  return QueueOrWriteFrame(&frame);
+}
+
+void Cluster::AddPayloadSize(uint64_t size) { payload_size_ += size; }
+
+bool Cluster::Finalize() {
+  return !write_last_frame_with_duration_ && Finalize(false, 0);
+}
+
+bool Cluster::Finalize(bool set_last_frame_duration, uint64_t duration) {
+  if (!writer_ || finalized_)
+    return false;
+
+  if (write_last_frame_with_duration_) {
+    // Write out held back Frames. This essentially performs a k-way merge
+    // across all tracks in the increasing order of timestamps.
+    while (!stored_frames_.empty()) {
+      Frame* frame = stored_frames_.begin()->second.front();
+
+      // Get the next frame to write (frame with least timestamp across all
+      // tracks).
+      for (FrameMapIterator frames_iterator = ++stored_frames_.begin();
+           frames_iterator != stored_frames_.end(); ++frames_iterator) {
+        if (frames_iterator->second.front()->timestamp() < frame->timestamp()) {
+          frame = frames_iterator->second.front();
+        }
+      }
+
+      // Set the duration if it's the last frame for the track.
+      if (set_last_frame_duration &&
+          stored_frames_[frame->track_number()].size() == 1 &&
+          !frame->duration_set()) {
+        frame->set_duration(duration - frame->timestamp());
+        if (!frame->is_key() && !frame->reference_block_timestamp_set()) {
+          frame->set_reference_block_timestamp(
+              last_block_timestamp_[frame->track_number()]);
+        }
+      }
+
+      // Write the frame and remove it from |stored_frames_|.
+      const bool wrote_frame = DoWriteFrame(frame);
+      stored_frames_[frame->track_number()].pop_front();
+      if (stored_frames_[frame->track_number()].empty()) {
+        stored_frames_.erase(frame->track_number());
+      }
+      delete frame;
+      if (!wrote_frame)
+        return false;
+    }
+  }
+
+  if (size_position_ == -1)
+    return false;
+
+  if (writer_->Seekable()) {
+    const int64_t pos = writer_->Position();
+
+    if (writer_->Position(size_position_))
+      return false;
+
+    if (WriteUIntSize(writer_, payload_size(), 8))
+      return false;
+
+    if (writer_->Position(pos))
+      return false;
+  }
+
+  finalized_ = true;
+
+  return true;
+}
+
+uint64_t Cluster::Size() const {
+  const uint64_t element_size =
+      EbmlMasterElementSize(libwebm::kMkvCluster, 0xFFFFFFFFFFFFFFFFULL) +
+      payload_size_;
+  return element_size;
+}
+
+bool Cluster::PreWriteBlock() {
+  if (finalized_)
+    return false;
+
+  if (!header_written_) {
+    if (!WriteClusterHeader())
+      return false;
+  }
+
+  return true;
+}
+
+void Cluster::PostWriteBlock(uint64_t element_size) {
+  AddPayloadSize(element_size);
+  ++blocks_added_;
+}
+
+int64_t Cluster::GetRelativeTimecode(int64_t abs_timecode) const {
+  const int64_t cluster_timecode = this->Cluster::timecode();
+  const int64_t rel_timecode =
+      static_cast<int64_t>(abs_timecode) - cluster_timecode;
+
+  if (rel_timecode < 0 || rel_timecode > kMaxBlockTimecode)
+    return -1;
+
+  return rel_timecode;
+}
+
+bool Cluster::DoWriteFrame(const Frame* const frame) {
+  if (!frame || !frame->IsValid())
+    return false;
+
+  if (!PreWriteBlock())
+    return false;
+
+  const uint64_t element_size = WriteFrame(writer_, frame, this);
+  if (element_size == 0)
+    return false;
+
+  PostWriteBlock(element_size);
+  last_block_timestamp_[frame->track_number()] = frame->timestamp();
+  return true;
+}
+
+bool Cluster::QueueOrWriteFrame(const Frame* const frame) {
+  if (!frame || !frame->IsValid())
+    return false;
+
+  // If |write_last_frame_with_duration_| is not set, then write the frame right
+  // away.
+  if (!write_last_frame_with_duration_) {
+    return DoWriteFrame(frame);
+  }
+
+  // Queue the current frame.
+  uint64_t track_number = frame->track_number();
+  Frame* const frame_to_store = new Frame();
+  frame_to_store->CopyFrom(*frame);
+  stored_frames_[track_number].push_back(frame_to_store);
+
+  // Iterate through all queued frames in the current track except the last one
+  // and write it if it is okay to do so (i.e.) no other track has an held back
+  // frame with timestamp <= the timestamp of the frame in question.
+  std::vector<std::list<Frame*>::iterator> frames_to_erase;
+  for (std::list<Frame *>::iterator
+           current_track_iterator = stored_frames_[track_number].begin(),
+           end = --stored_frames_[track_number].end();
+       current_track_iterator != end; ++current_track_iterator) {
+    const Frame* const frame_to_write = *current_track_iterator;
+    bool okay_to_write = true;
+    for (FrameMapIterator track_iterator = stored_frames_.begin();
+         track_iterator != stored_frames_.end(); ++track_iterator) {
+      if (track_iterator->first == track_number) {
+        continue;
+      }
+      if (track_iterator->second.front()->timestamp() <
+          frame_to_write->timestamp()) {
+        okay_to_write = false;
+        break;
+      }
+    }
+    if (okay_to_write) {
+      const bool wrote_frame = DoWriteFrame(frame_to_write);
+      delete frame_to_write;
+      if (!wrote_frame)
+        return false;
+      frames_to_erase.push_back(current_track_iterator);
+    } else {
+      break;
+    }
+  }
+  for (std::vector<std::list<Frame*>::iterator>::iterator iterator =
+           frames_to_erase.begin();
+       iterator != frames_to_erase.end(); ++iterator) {
+    stored_frames_[track_number].erase(*iterator);
+  }
+  return true;
+}
+
+bool Cluster::WriteClusterHeader() {
+  if (finalized_)
+    return false;
+
+  if (WriteID(writer_, libwebm::kMkvCluster))
+    return false;
+
+  // Save for later.
+  size_position_ = writer_->Position();
+
+  // Write "unknown" (EBML coded -1) as cluster size value. We need to write 8
+  // bytes because we do not know how big our cluster will be.
+  if (SerializeInt(writer_, kEbmlUnknownValue, 8))
+    return false;
+
+  if (!WriteEbmlElement(writer_, libwebm::kMkvTimecode, timecode(),
+                        fixed_size_timecode_ ? 8 : 0)) {
+    return false;
+  }
+  AddPayloadSize(EbmlElementSize(libwebm::kMkvTimecode, timecode(),
+                                 fixed_size_timecode_ ? 8 : 0));
+  header_written_ = true;
+
+  return true;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// SeekHead Class
+
+SeekHead::SeekHead() : start_pos_(0ULL) {
+  for (int32_t i = 0; i < kSeekEntryCount; ++i) {
+    seek_entry_id_[i] = 0;
+    seek_entry_pos_[i] = 0;
+  }
+}
+
+SeekHead::~SeekHead() {}
+
+bool SeekHead::Finalize(IMkvWriter* writer) const {
+  if (writer->Seekable()) {
+    if (start_pos_ == -1)
+      return false;
+
+    uint64_t payload_size = 0;
+    uint64_t entry_size[kSeekEntryCount];
+
+    for (int32_t i = 0; i < kSeekEntryCount; ++i) {
+      if (seek_entry_id_[i] != 0) {
+        entry_size[i] = EbmlElementSize(
+            libwebm::kMkvSeekID, static_cast<uint64_t>(seek_entry_id_[i]));
+        entry_size[i] +=
+            EbmlElementSize(libwebm::kMkvSeekPosition, seek_entry_pos_[i]);
+
+        payload_size +=
+            EbmlMasterElementSize(libwebm::kMkvSeek, entry_size[i]) +
+            entry_size[i];
+      }
+    }
+
+    // No SeekHead elements
+    if (payload_size == 0)
+      return true;
+
+    const int64_t pos = writer->Position();
+    if (writer->Position(start_pos_))
+      return false;
+
+    if (!WriteEbmlMasterElement(writer, libwebm::kMkvSeekHead, payload_size))
+      return false;
+
+    for (int32_t i = 0; i < kSeekEntryCount; ++i) {
+      if (seek_entry_id_[i] != 0) {
+        if (!WriteEbmlMasterElement(writer, libwebm::kMkvSeek, entry_size[i]))
+          return false;
+
+        if (!WriteEbmlElement(writer, libwebm::kMkvSeekID,
+                              static_cast<uint64_t>(seek_entry_id_[i])))
+          return false;
+
+        if (!WriteEbmlElement(writer, libwebm::kMkvSeekPosition,
+                              seek_entry_pos_[i]))
+          return false;
+      }
+    }
+
+    const uint64_t total_entry_size = kSeekEntryCount * MaxEntrySize();
+    const uint64_t total_size =
+        EbmlMasterElementSize(libwebm::kMkvSeekHead, total_entry_size) +
+        total_entry_size;
+    const int64_t size_left = total_size - (writer->Position() - start_pos_);
+
+    const uint64_t bytes_written = WriteVoidElement(writer, size_left);
+    if (!bytes_written)
+      return false;
+
+    if (writer->Position(pos))
+      return false;
+  }
+
+  return true;
+}
+
+bool SeekHead::Write(IMkvWriter* writer) {
+  const uint64_t entry_size = kSeekEntryCount * MaxEntrySize();
+  const uint64_t size =
+      EbmlMasterElementSize(libwebm::kMkvSeekHead, entry_size);
+
+  start_pos_ = writer->Position();
+
+  const uint64_t bytes_written = WriteVoidElement(writer, size + entry_size);
+  if (!bytes_written)
+    return false;
+
+  return true;
+}
+
+bool SeekHead::AddSeekEntry(uint32_t id, uint64_t pos) {
+  for (int32_t i = 0; i < kSeekEntryCount; ++i) {
+    if (seek_entry_id_[i] == 0) {
+      seek_entry_id_[i] = id;
+      seek_entry_pos_[i] = pos;
+      return true;
+    }
+  }
+  return false;
+}
+
+uint32_t SeekHead::GetId(int index) const {
+  if (index < 0 || index >= kSeekEntryCount)
+    return UINT_MAX;
+  return seek_entry_id_[index];
+}
+
+uint64_t SeekHead::GetPosition(int index) const {
+  if (index < 0 || index >= kSeekEntryCount)
+    return ULLONG_MAX;
+  return seek_entry_pos_[index];
+}
+
+bool SeekHead::SetSeekEntry(int index, uint32_t id, uint64_t position) {
+  if (index < 0 || index >= kSeekEntryCount)
+    return false;
+  seek_entry_id_[index] = id;
+  seek_entry_pos_[index] = position;
+  return true;
+}
+
+uint64_t SeekHead::MaxEntrySize() const {
+  const uint64_t max_entry_payload_size =
+      EbmlElementSize(libwebm::kMkvSeekID, UINT64_C(0xffffffff)) +
+      EbmlElementSize(libwebm::kMkvSeekPosition, UINT64_C(0xffffffffffffffff));
+  const uint64_t max_entry_size =
+      EbmlMasterElementSize(libwebm::kMkvSeek, max_entry_payload_size) +
+      max_entry_payload_size;
+
+  return max_entry_size;
+}
+
+///////////////////////////////////////////////////////////////
+//
+// SegmentInfo Class
+
+SegmentInfo::SegmentInfo()
+    : duration_(-1.0),
+      muxing_app_(NULL),
+      timecode_scale_(1000000ULL),
+      writing_app_(NULL),
+      date_utc_(LLONG_MIN),
+      duration_pos_(-1) {}
+
+SegmentInfo::~SegmentInfo() {
+  delete[] muxing_app_;
+  delete[] writing_app_;
+}
+
+bool SegmentInfo::Init() {
+  int32_t major;
+  int32_t minor;
+  int32_t build;
+  int32_t revision;
+  GetVersion(&major, &minor, &build, &revision);
+  char temp[256];
+#ifdef _MSC_VER
+  sprintf_s(temp, sizeof(temp) / sizeof(temp[0]), "libwebm-%d.%d.%d.%d", major,
+            minor, build, revision);
+#else
+  snprintf(temp, sizeof(temp) / sizeof(temp[0]), "libwebm-%d.%d.%d.%d", major,
+           minor, build, revision);
+#endif
+
+  const size_t app_len = strlen(temp) + 1;
+
+  delete[] muxing_app_;
+
+  muxing_app_ = new (std::nothrow) char[app_len];  // NOLINT
+  if (!muxing_app_)
+    return false;
+
+#ifdef _MSC_VER
+  strcpy_s(muxing_app_, app_len, temp);
+#else
+  strcpy(muxing_app_, temp);
+#endif
+
+  set_writing_app(temp);
+  if (!writing_app_)
+    return false;
+  return true;
+}
+
+bool SegmentInfo::Finalize(IMkvWriter* writer) const {
+  if (!writer)
+    return false;
+
+  if (duration_ > 0.0) {
+    if (writer->Seekable()) {
+      if (duration_pos_ == -1)
+        return false;
+
+      const int64_t pos = writer->Position();
+
+      if (writer->Position(duration_pos_))
+        return false;
+
+      if (!WriteEbmlElement(writer, libwebm::kMkvDuration,
+                            static_cast<float>(duration_)))
+        return false;
+
+      if (writer->Position(pos))
+        return false;
+    }
+  }
+
+  return true;
+}
+
+bool SegmentInfo::Write(IMkvWriter* writer) {
+  if (!writer || !muxing_app_ || !writing_app_)
+    return false;
+
+  uint64_t size = EbmlElementSize(libwebm::kMkvTimecodeScale, timecode_scale_);
+  if (duration_ > 0.0)
+    size +=
+        EbmlElementSize(libwebm::kMkvDuration, static_cast<float>(duration_));
+  if (date_utc_ != LLONG_MIN)
+    size += EbmlDateElementSize(libwebm::kMkvDateUTC);
+  size += EbmlElementSize(libwebm::kMkvMuxingApp, muxing_app_);
+  size += EbmlElementSize(libwebm::kMkvWritingApp, writing_app_);
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvInfo, size))
+    return false;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return false;
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvTimecodeScale, timecode_scale_))
+    return false;
+
+  if (duration_ > 0.0) {
+    // Save for later
+    duration_pos_ = writer->Position();
+
+    if (!WriteEbmlElement(writer, libwebm::kMkvDuration,
+                          static_cast<float>(duration_)))
+      return false;
+  }
+
+  if (date_utc_ != LLONG_MIN)
+    WriteEbmlDateElement(writer, libwebm::kMkvDateUTC, date_utc_);
+
+  if (!WriteEbmlElement(writer, libwebm::kMkvMuxingApp, muxing_app_))
+    return false;
+  if (!WriteEbmlElement(writer, libwebm::kMkvWritingApp, writing_app_))
+    return false;
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(size))
+    return false;
+
+  return true;
+}
+
+void SegmentInfo::set_muxing_app(const char* app) {
+  if (app) {
+    const size_t length = strlen(app) + 1;
+    char* temp_str = new (std::nothrow) char[length];  // NOLINT
+    if (!temp_str)
+      return;
+
+#ifdef _MSC_VER
+    strcpy_s(temp_str, length, app);
+#else
+    strcpy(temp_str, app);
+#endif
+
+    delete[] muxing_app_;
+    muxing_app_ = temp_str;
+  }
+}
+
+void SegmentInfo::set_writing_app(const char* app) {
+  if (app) {
+    const size_t length = strlen(app) + 1;
+    char* temp_str = new (std::nothrow) char[length];  // NOLINT
+    if (!temp_str)
+      return;
+
+#ifdef _MSC_VER
+    strcpy_s(temp_str, length, app);
+#else
+    strcpy(temp_str, app);
+#endif
+
+    delete[] writing_app_;
+    writing_app_ = temp_str;
+  }
+}
+
+///////////////////////////////////////////////////////////////
+//
+// Segment Class
+
+Segment::Segment()
+    : chunk_count_(0),
+      chunk_name_(NULL),
+      chunk_writer_cluster_(NULL),
+      chunk_writer_cues_(NULL),
+      chunk_writer_header_(NULL),
+      chunking_(false),
+      chunking_base_name_(NULL),
+      cluster_list_(NULL),
+      cluster_list_capacity_(0),
+      cluster_list_size_(0),
+      cues_position_(kAfterClusters),
+      cues_track_(0),
+      force_new_cluster_(false),
+      frames_(NULL),
+      frames_capacity_(0),
+      frames_size_(0),
+      has_video_(false),
+      header_written_(false),
+      last_block_duration_(0),
+      last_timestamp_(0),
+      max_cluster_duration_(kDefaultMaxClusterDuration),
+      max_cluster_size_(0),
+      mode_(kFile),
+      new_cuepoint_(false),
+      output_cues_(true),
+      accurate_cluster_duration_(false),
+      fixed_size_cluster_timecode_(false),
+      payload_pos_(0),
+      size_position_(0),
+      doc_type_version_(kDefaultDocTypeVersion),
+      doc_type_version_written_(0),
+      writer_cluster_(NULL),
+      writer_cues_(NULL),
+      writer_header_(NULL) {
+  const time_t curr_time = time(NULL);
+  seed_ = static_cast<unsigned int>(curr_time);
+#ifdef _WIN32
+  srand(seed_);
+#endif
+}
+
+Segment::~Segment() {
+  if (cluster_list_) {
+    for (int32_t i = 0; i < cluster_list_size_; ++i) {
+      Cluster* const cluster = cluster_list_[i];
+      delete cluster;
+    }
+    delete[] cluster_list_;
+  }
+
+  if (frames_) {
+    for (int32_t i = 0; i < frames_size_; ++i) {
+      Frame* const frame = frames_[i];
+      delete frame;
+    }
+    delete[] frames_;
+  }
+
+  delete[] chunk_name_;
+  delete[] chunking_base_name_;
+
+  if (chunk_writer_cluster_) {
+    chunk_writer_cluster_->Close();
+    delete chunk_writer_cluster_;
+  }
+  if (chunk_writer_cues_) {
+    chunk_writer_cues_->Close();
+    delete chunk_writer_cues_;
+  }
+  if (chunk_writer_header_) {
+    chunk_writer_header_->Close();
+    delete chunk_writer_header_;
+  }
+}
+
+void Segment::MoveCuesBeforeClustersHelper(uint64_t diff, int32_t index,
+                                           uint64_t* cues_size) {
+  CuePoint* const cue_point = cues_.GetCueByIndex(index);
+  if (cue_point == NULL)
+    return;
+  const uint64_t old_cue_point_size = cue_point->Size();
+  const uint64_t cluster_pos = cue_point->cluster_pos() + diff;
+  cue_point->set_cluster_pos(cluster_pos);  // update the new cluster position
+  // New size of the cue is computed as follows
+  //    Let a = current sum of size of all CuePoints
+  //    Let b = Increase in Cue Point's size due to this iteration
+  //    Let c = Increase in size of Cues Element's length due to this iteration
+  //            (This is computed as CodedSize(a + b) - CodedSize(a))
+  //    Let d = b + c. Now d is the |diff| passed to the next recursive call.
+  //    Let e = a + b. Now e is the |cues_size| passed to the next recursive
+  //                   call.
+  const uint64_t cue_point_size_diff = cue_point->Size() - old_cue_point_size;
+  const uint64_t cue_size_diff =
+      GetCodedUIntSize(*cues_size + cue_point_size_diff) -
+      GetCodedUIntSize(*cues_size);
+  *cues_size += cue_point_size_diff;
+  diff = cue_size_diff + cue_point_size_diff;
+  if (diff > 0) {
+    for (int32_t i = 0; i < cues_.cue_entries_size(); ++i) {
+      MoveCuesBeforeClustersHelper(diff, i, cues_size);
+    }
+  }
+}
+
+void Segment::MoveCuesBeforeClusters() {
+  const uint64_t current_cue_size = cues_.Size();
+  uint64_t cue_size = 0;
+  for (int32_t i = 0; i < cues_.cue_entries_size(); ++i)
+    cue_size += cues_.GetCueByIndex(i)->Size();
+  for (int32_t i = 0; i < cues_.cue_entries_size(); ++i)
+    MoveCuesBeforeClustersHelper(current_cue_size, i, &cue_size);
+
+  // Adjust the Seek Entry to reflect the change in position
+  // of Cluster and Cues
+  int32_t cluster_index = 0;
+  int32_t cues_index = 0;
+  for (int32_t i = 0; i < SeekHead::kSeekEntryCount; ++i) {
+    if (seek_head_.GetId(i) == libwebm::kMkvCluster)
+      cluster_index = i;
+    if (seek_head_.GetId(i) == libwebm::kMkvCues)
+      cues_index = i;
+  }
+  seek_head_.SetSeekEntry(cues_index, libwebm::kMkvCues,
+                          seek_head_.GetPosition(cluster_index));
+  seek_head_.SetSeekEntry(cluster_index, libwebm::kMkvCluster,
+                          cues_.Size() + seek_head_.GetPosition(cues_index));
+}
+
+bool Segment::Init(IMkvWriter* ptr_writer) {
+  if (!ptr_writer) {
+    return false;
+  }
+  writer_cluster_ = ptr_writer;
+  writer_cues_ = ptr_writer;
+  writer_header_ = ptr_writer;
+  return segment_info_.Init();
+}
+
+bool Segment::CopyAndMoveCuesBeforeClusters(mkvparser::IMkvReader* reader,
+                                            IMkvWriter* writer) {
+  if (!writer->Seekable() || chunking_)
+    return false;
+  const int64_t cluster_offset =
+      cluster_list_[0]->size_position() - GetUIntSize(libwebm::kMkvCluster);
+
+  // Copy the headers.
+  if (!ChunkedCopy(reader, writer, 0, cluster_offset))
+    return false;
+
+  // Recompute cue positions and seek entries.
+  MoveCuesBeforeClusters();
+
+  // Write cues and seek entries.
+  // TODO(vigneshv): As of now, it's safe to call seek_head_.Finalize() for the
+  // second time with a different writer object. But the name Finalize() doesn't
+  // indicate something we want to call more than once. So consider renaming it
+  // to write() or some such.
+  if (!cues_.Write(writer) || !seek_head_.Finalize(writer))
+    return false;
+
+  // Copy the Clusters.
+  if (!ChunkedCopy(reader, writer, cluster_offset,
+                   cluster_end_offset_ - cluster_offset))
+    return false;
+
+  // Update the Segment size in case the Cues size has changed.
+  const int64_t pos = writer->Position();
+  const int64_t segment_size = writer->Position() - payload_pos_;
+  if (writer->Position(size_position_) ||
+      WriteUIntSize(writer, segment_size, 8) || writer->Position(pos))
+    return false;
+  return true;
+}
+
+bool Segment::Finalize() {
+  if (WriteFramesAll() < 0)
+    return false;
+
+  if (cluster_list_size_ > 0) {
+    // Update last cluster's size
+    Cluster* const old_cluster = cluster_list_[cluster_list_size_ - 1];
+
+    // For the last frame of the last Cluster, we don't write it as a BlockGroup
+    // with Duration unless the frame itself has duration set explicitly.
+    if (!old_cluster || !old_cluster->Finalize(false, 0))
+      return false;
+  }
+
+  if (mode_ == kFile) {
+    if (chunking_ && chunk_writer_cluster_) {
+      chunk_writer_cluster_->Close();
+      chunk_count_++;
+    }
+
+    const double duration =
+        (static_cast<double>(last_timestamp_) + last_block_duration_) /
+        segment_info_.timecode_scale();
+    segment_info_.set_duration(duration);
+    if (!segment_info_.Finalize(writer_header_))
+      return false;
+
+    if (output_cues_)
+      if (!seek_head_.AddSeekEntry(libwebm::kMkvCues, MaxOffset()))
+        return false;
+
+    if (chunking_) {
+      if (!chunk_writer_cues_)
+        return false;
+
+      char* name = NULL;
+      if (!UpdateChunkName("cues", &name))
+        return false;
+
+      const bool cues_open = chunk_writer_cues_->Open(name);
+      delete[] name;
+      if (!cues_open)
+        return false;
+    }
+
+    cluster_end_offset_ = writer_cluster_->Position();
+
+    // Write the seek headers and cues
+    if (output_cues_)
+      if (!cues_.Write(writer_cues_))
+        return false;
+
+    if (!seek_head_.Finalize(writer_header_))
+      return false;
+
+    if (writer_header_->Seekable()) {
+      if (size_position_ == -1)
+        return false;
+
+      const int64_t segment_size = MaxOffset();
+      if (segment_size < 1)
+        return false;
+
+      const int64_t pos = writer_header_->Position();
+      UpdateDocTypeVersion();
+      if (doc_type_version_ != doc_type_version_written_) {
+        if (writer_header_->Position(0))
+          return false;
+
+        if (!WriteEbmlHeader(writer_header_, doc_type_version_))
+          return false;
+        if (writer_header_->Position() != ebml_header_size_)
+          return false;
+
+        doc_type_version_written_ = doc_type_version_;
+      }
+
+      if (writer_header_->Position(size_position_))
+        return false;
+
+      if (WriteUIntSize(writer_header_, segment_size, 8))
+        return false;
+
+      if (writer_header_->Position(pos))
+        return false;
+    }
+
+    if (chunking_) {
+      // Do not close any writers until the segment size has been written,
+      // otherwise the size may be off.
+      if (!chunk_writer_cues_ || !chunk_writer_header_)
+        return false;
+
+      chunk_writer_cues_->Close();
+      chunk_writer_header_->Close();
+    }
+  }
+
+  return true;
+}
+
+Track* Segment::AddTrack(int32_t number) {
+  Track* const track = new (std::nothrow) Track(&seed_);  // NOLINT
+
+  if (!track)
+    return NULL;
+
+  if (!tracks_.AddTrack(track, number)) {
+    delete track;
+    return NULL;
+  }
+
+  return track;
+}
+
+Chapter* Segment::AddChapter() { return chapters_.AddChapter(&seed_); }
+
+Tag* Segment::AddTag() { return tags_.AddTag(); }
+
+uint64_t Segment::AddVideoTrack(int32_t width, int32_t height, int32_t number) {
+  VideoTrack* const track = new (std::nothrow) VideoTrack(&seed_);  // NOLINT
+  if (!track)
+    return 0;
+
+  track->set_type(Tracks::kVideo);
+  track->set_codec_id(Tracks::kVp8CodecId);
+  track->set_width(width);
+  track->set_height(height);
+
+  tracks_.AddTrack(track, number);
+  has_video_ = true;
+
+  return track->number();
+}
+
+bool Segment::AddCuePoint(uint64_t timestamp, uint64_t track) {
+  if (cluster_list_size_ < 1)
+    return false;
+
+  const Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
+  if (!cluster)
+    return false;
+
+  CuePoint* const cue = new (std::nothrow) CuePoint();  // NOLINT
+  if (!cue)
+    return false;
+
+  cue->set_time(timestamp / segment_info_.timecode_scale());
+  cue->set_block_number(cluster->blocks_added());
+  cue->set_cluster_pos(cluster->position_for_cues());
+  cue->set_track(track);
+  if (!cues_.AddCue(cue))
+    return false;
+
+  new_cuepoint_ = false;
+  return true;
+}
+
+uint64_t Segment::AddAudioTrack(int32_t sample_rate, int32_t channels,
+                                int32_t number) {
+  AudioTrack* const track = new (std::nothrow) AudioTrack(&seed_);  // NOLINT
+  if (!track)
+    return 0;
+
+  track->set_type(Tracks::kAudio);
+  track->set_codec_id(Tracks::kVorbisCodecId);
+  track->set_sample_rate(sample_rate);
+  track->set_channels(channels);
+
+  tracks_.AddTrack(track, number);
+
+  return track->number();
+}
+
+bool Segment::AddFrame(const uint8_t* data, uint64_t length,
+                       uint64_t track_number, uint64_t timestamp, bool is_key) {
+  if (!data)
+    return false;
+
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_track_number(track_number);
+  frame.set_timestamp(timestamp);
+  frame.set_is_key(is_key);
+  return AddGenericFrame(&frame);
+}
+
+bool Segment::AddFrameWithAdditional(const uint8_t* data, uint64_t length,
+                                     const uint8_t* additional,
+                                     uint64_t additional_length,
+                                     uint64_t add_id, uint64_t track_number,
+                                     uint64_t timestamp, bool is_key) {
+  if (!data || !additional)
+    return false;
+
+  Frame frame;
+  if (!frame.Init(data, length) ||
+      !frame.AddAdditionalData(additional, additional_length, add_id)) {
+    return false;
+  }
+  frame.set_track_number(track_number);
+  frame.set_timestamp(timestamp);
+  frame.set_is_key(is_key);
+  return AddGenericFrame(&frame);
+}
+
+bool Segment::AddFrameWithDiscardPadding(const uint8_t* data, uint64_t length,
+                                         int64_t discard_padding,
+                                         uint64_t track_number,
+                                         uint64_t timestamp, bool is_key) {
+  if (!data)
+    return false;
+
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_discard_padding(discard_padding);
+  frame.set_track_number(track_number);
+  frame.set_timestamp(timestamp);
+  frame.set_is_key(is_key);
+  return AddGenericFrame(&frame);
+}
+
+bool Segment::AddMetadata(const uint8_t* data, uint64_t length,
+                          uint64_t track_number, uint64_t timestamp_ns,
+                          uint64_t duration_ns) {
+  if (!data)
+    return false;
+
+  Frame frame;
+  if (!frame.Init(data, length))
+    return false;
+  frame.set_track_number(track_number);
+  frame.set_timestamp(timestamp_ns);
+  frame.set_duration(duration_ns);
+  frame.set_is_key(true);  // All metadata blocks are keyframes.
+  return AddGenericFrame(&frame);
+}
+
+bool Segment::AddGenericFrame(const Frame* frame) {
+  if (!frame)
+    return false;
+
+  if (!CheckHeaderInfo())
+    return false;
+
+  // Check for non-monotonically increasing timestamps.
+  if (frame->timestamp() < last_timestamp_)
+    return false;
+
+  // Check if the track number is valid.
+  if (!tracks_.GetTrackByNumber(frame->track_number()))
+    return false;
+
+  if (frame->discard_padding() != 0)
+    doc_type_version_ = 4;
+
+  // If the segment has a video track hold onto audio frames to make sure the
+  // audio that is associated with the start time of a video key-frame is
+  // muxed into the same cluster.
+  if (has_video_ && tracks_.TrackIsAudio(frame->track_number()) &&
+      !force_new_cluster_) {
+    Frame* const new_frame = new (std::nothrow) Frame();
+    if (!new_frame || !new_frame->CopyFrom(*frame))
+      return false;
+    return QueueFrame(new_frame);
+  }
+
+  if (!DoNewClusterProcessing(frame->track_number(), frame->timestamp(),
+                              frame->is_key())) {
+    return false;
+  }
+
+  if (cluster_list_size_ < 1)
+    return false;
+
+  Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
+  if (!cluster)
+    return false;
+
+  // If the Frame is not a SimpleBlock, then set the reference_block_timestamp
+  // if it is not set already.
+  bool frame_created = false;
+  if (!frame->CanBeSimpleBlock() && !frame->is_key() &&
+      !frame->reference_block_timestamp_set()) {
+    Frame* const new_frame = new (std::nothrow) Frame();
+    if (!new_frame->CopyFrom(*frame))
+      return false;
+    new_frame->set_reference_block_timestamp(
+        last_track_timestamp_[frame->track_number() - 1]);
+    frame = new_frame;
+    frame_created = true;
+  }
+
+  if (!cluster->AddFrame(frame))
+    return false;
+
+  if (new_cuepoint_ && cues_track_ == frame->track_number()) {
+    if (!AddCuePoint(frame->timestamp(), cues_track_))
+      return false;
+  }
+
+  last_timestamp_ = frame->timestamp();
+  last_track_timestamp_[frame->track_number() - 1] = frame->timestamp();
+  last_block_duration_ = frame->duration();
+
+  if (frame_created)
+    delete frame;
+
+  return true;
+}
+
+void Segment::OutputCues(bool output_cues) { output_cues_ = output_cues; }
+
+void Segment::AccurateClusterDuration(bool accurate_cluster_duration) {
+  accurate_cluster_duration_ = accurate_cluster_duration;
+}
+
+void Segment::UseFixedSizeClusterTimecode(bool fixed_size_cluster_timecode) {
+  fixed_size_cluster_timecode_ = fixed_size_cluster_timecode;
+}
+
+bool Segment::SetChunking(bool chunking, const char* filename) {
+  if (chunk_count_ > 0)
+    return false;
+
+  if (chunking) {
+    if (!filename)
+      return false;
+
+    // Check if we are being set to what is already set.
+    if (chunking_ && !strcmp(filename, chunking_base_name_))
+      return true;
+
+    const size_t name_length = strlen(filename) + 1;
+    char* const temp = new (std::nothrow) char[name_length];  // NOLINT
+    if (!temp)
+      return false;
+
+#ifdef _MSC_VER
+    strcpy_s(temp, name_length, filename);
+#else
+    strcpy(temp, filename);
+#endif
+
+    delete[] chunking_base_name_;
+    chunking_base_name_ = temp;
+
+    if (!UpdateChunkName("chk", &chunk_name_))
+      return false;
+
+    if (!chunk_writer_cluster_) {
+      chunk_writer_cluster_ = new (std::nothrow) MkvWriter();  // NOLINT
+      if (!chunk_writer_cluster_)
+        return false;
+    }
+
+    if (!chunk_writer_cues_) {
+      chunk_writer_cues_ = new (std::nothrow) MkvWriter();  // NOLINT
+      if (!chunk_writer_cues_)
+        return false;
+    }
+
+    if (!chunk_writer_header_) {
+      chunk_writer_header_ = new (std::nothrow) MkvWriter();  // NOLINT
+      if (!chunk_writer_header_)
+        return false;
+    }
+
+    if (!chunk_writer_cluster_->Open(chunk_name_))
+      return false;
+
+    const size_t header_length = strlen(filename) + strlen(".hdr") + 1;
+    char* const header = new (std::nothrow) char[header_length];  // NOLINT
+    if (!header)
+      return false;
+
+#ifdef _MSC_VER
+    strcpy_s(header, header_length - strlen(".hdr"), chunking_base_name_);
+    strcat_s(header, header_length, ".hdr");
+#else
+    strcpy(header, chunking_base_name_);
+    strcat(header, ".hdr");
+#endif
+    if (!chunk_writer_header_->Open(header)) {
+      delete[] header;
+      return false;
+    }
+
+    writer_cluster_ = chunk_writer_cluster_;
+    writer_cues_ = chunk_writer_cues_;
+    writer_header_ = chunk_writer_header_;
+
+    delete[] header;
+  }
+
+  chunking_ = chunking;
+
+  return true;
+}
+
+bool Segment::CuesTrack(uint64_t track_number) {
+  const Track* const track = GetTrackByNumber(track_number);
+  if (!track)
+    return false;
+
+  cues_track_ = track_number;
+  return true;
+}
+
+void Segment::ForceNewClusterOnNextFrame() { force_new_cluster_ = true; }
+
+Track* Segment::GetTrackByNumber(uint64_t track_number) const {
+  return tracks_.GetTrackByNumber(track_number);
+}
+
+bool Segment::WriteSegmentHeader() {
+  UpdateDocTypeVersion();
+
+  // TODO(fgalligan): Support more than one segment.
+  if (!WriteEbmlHeader(writer_header_, doc_type_version_))
+    return false;
+  doc_type_version_written_ = doc_type_version_;
+  ebml_header_size_ = static_cast<int32_t>(writer_header_->Position());
+
+  // Write "unknown" (-1) as segment size value. If mode is kFile, Segment
+  // will write over duration when the file is finalized.
+  if (WriteID(writer_header_, libwebm::kMkvSegment))
+    return false;
+
+  // Save for later.
+  size_position_ = writer_header_->Position();
+
+  // Write "unknown" (EBML coded -1) as segment size value. We need to write 8
+  // bytes because if we are going to overwrite the segment size later we do
+  // not know how big our segment will be.
+  if (SerializeInt(writer_header_, kEbmlUnknownValue, 8))
+    return false;
+
+  payload_pos_ = writer_header_->Position();
+
+  if (mode_ == kFile && writer_header_->Seekable()) {
+    // Set the duration > 0.0 so SegmentInfo will write out the duration. When
+    // the muxer is done writing we will set the correct duration and have
+    // SegmentInfo upadte it.
+    segment_info_.set_duration(1.0);
+
+    if (!seek_head_.Write(writer_header_))
+      return false;
+  }
+
+  if (!seek_head_.AddSeekEntry(libwebm::kMkvInfo, MaxOffset()))
+    return false;
+  if (!segment_info_.Write(writer_header_))
+    return false;
+
+  if (!seek_head_.AddSeekEntry(libwebm::kMkvTracks, MaxOffset()))
+    return false;
+  if (!tracks_.Write(writer_header_))
+    return false;
+
+  if (chapters_.Count() > 0) {
+    if (!seek_head_.AddSeekEntry(libwebm::kMkvChapters, MaxOffset()))
+      return false;
+    if (!chapters_.Write(writer_header_))
+      return false;
+  }
+
+  if (tags_.Count() > 0) {
+    if (!seek_head_.AddSeekEntry(libwebm::kMkvTags, MaxOffset()))
+      return false;
+    if (!tags_.Write(writer_header_))
+      return false;
+  }
+
+  if (chunking_ && (mode_ == kLive || !writer_header_->Seekable())) {
+    if (!chunk_writer_header_)
+      return false;
+
+    chunk_writer_header_->Close();
+  }
+
+  header_written_ = true;
+
+  return true;
+}
+
+// Here we are testing whether to create a new cluster, given a frame
+// having time frame_timestamp_ns.
+//
+int Segment::TestFrame(uint64_t track_number, uint64_t frame_timestamp_ns,
+                       bool is_key) const {
+  if (force_new_cluster_)
+    return 1;
+
+  // If no clusters have been created yet, then create a new cluster
+  // and write this frame immediately, in the new cluster.  This path
+  // should only be followed once, the first time we attempt to write
+  // a frame.
+
+  if (cluster_list_size_ <= 0)
+    return 1;
+
+  // There exists at least one cluster. We must compare the frame to
+  // the last cluster, in order to determine whether the frame is
+  // written to the existing cluster, or that a new cluster should be
+  // created.
+
+  const uint64_t timecode_scale = segment_info_.timecode_scale();
+  const uint64_t frame_timecode = frame_timestamp_ns / timecode_scale;
+
+  const Cluster* const last_cluster = cluster_list_[cluster_list_size_ - 1];
+  const uint64_t last_cluster_timecode = last_cluster->timecode();
+
+  // For completeness we test for the case when the frame's timecode
+  // is less than the cluster's timecode.  Although in principle that
+  // is allowed, this muxer doesn't actually write clusters like that,
+  // so this indicates a bug somewhere in our algorithm.
+
+  if (frame_timecode < last_cluster_timecode)  // should never happen
+    return -1;
+
+  // If the frame has a timestamp significantly larger than the last
+  // cluster (in Matroska, cluster-relative timestamps are serialized
+  // using a 16-bit signed integer), then we cannot write this frame
+  // to that cluster, and so we must create a new cluster.
+
+  const int64_t delta_timecode = frame_timecode - last_cluster_timecode;
+
+  if (delta_timecode > kMaxBlockTimecode)
+    return 2;
+
+  // We decide to create a new cluster when we have a video keyframe.
+  // This will flush queued (audio) frames, and write the keyframe
+  // immediately, in the newly-created cluster.
+
+  if (is_key && tracks_.TrackIsVideo(track_number))
+    return 1;
+
+  // Create a new cluster if we have accumulated too many frames
+  // already, where "too many" is defined as "the total time of frames
+  // in the cluster exceeds a threshold".
+
+  const uint64_t delta_ns = delta_timecode * timecode_scale;
+
+  if (max_cluster_duration_ > 0 && delta_ns >= max_cluster_duration_)
+    return 1;
+
+  // This is similar to the case above, with the difference that a new
+  // cluster is created when the size of the current cluster exceeds a
+  // threshold.
+
+  const uint64_t cluster_size = last_cluster->payload_size();
+
+  if (max_cluster_size_ > 0 && cluster_size >= max_cluster_size_)
+    return 1;
+
+  // There's no need to create a new cluster, so emit this frame now.
+
+  return 0;
+}
+
+bool Segment::MakeNewCluster(uint64_t frame_timestamp_ns) {
+  const int32_t new_size = cluster_list_size_ + 1;
+
+  if (new_size > cluster_list_capacity_) {
+    // Add more clusters.
+    const int32_t new_capacity =
+        (cluster_list_capacity_ <= 0) ? 1 : cluster_list_capacity_ * 2;
+    Cluster** const clusters =
+        new (std::nothrow) Cluster*[new_capacity];  // NOLINT
+    if (!clusters)
+      return false;
+
+    for (int32_t i = 0; i < cluster_list_size_; ++i) {
+      clusters[i] = cluster_list_[i];
+    }
+
+    delete[] cluster_list_;
+
+    cluster_list_ = clusters;
+    cluster_list_capacity_ = new_capacity;
+  }
+
+  if (!WriteFramesLessThan(frame_timestamp_ns))
+    return false;
+
+  if (cluster_list_size_ > 0) {
+    // Update old cluster's size
+    Cluster* const old_cluster = cluster_list_[cluster_list_size_ - 1];
+
+    if (!old_cluster || !old_cluster->Finalize(true, frame_timestamp_ns))
+      return false;
+  }
+
+  if (output_cues_)
+    new_cuepoint_ = true;
+
+  if (chunking_ && cluster_list_size_ > 0) {
+    chunk_writer_cluster_->Close();
+    chunk_count_++;
+
+    if (!UpdateChunkName("chk", &chunk_name_))
+      return false;
+    if (!chunk_writer_cluster_->Open(chunk_name_))
+      return false;
+  }
+
+  const uint64_t timecode_scale = segment_info_.timecode_scale();
+  const uint64_t frame_timecode = frame_timestamp_ns / timecode_scale;
+
+  uint64_t cluster_timecode = frame_timecode;
+
+  if (frames_size_ > 0) {
+    const Frame* const f = frames_[0];  // earliest queued frame
+    const uint64_t ns = f->timestamp();
+    const uint64_t tc = ns / timecode_scale;
+
+    if (tc < cluster_timecode)
+      cluster_timecode = tc;
+  }
+
+  Cluster*& cluster = cluster_list_[cluster_list_size_];
+  const int64_t offset = MaxOffset();
+  cluster = new (std::nothrow)
+      Cluster(cluster_timecode, offset, segment_info_.timecode_scale(),
+              accurate_cluster_duration_, fixed_size_cluster_timecode_);
+  if (!cluster)
+    return false;
+
+  if (!cluster->Init(writer_cluster_))
+    return false;
+
+  cluster_list_size_ = new_size;
+  return true;
+}
+
+bool Segment::DoNewClusterProcessing(uint64_t track_number,
+                                     uint64_t frame_timestamp_ns, bool is_key) {
+  for (;;) {
+    // Based on the characteristics of the current frame and current
+    // cluster, decide whether to create a new cluster.
+    const int result = TestFrame(track_number, frame_timestamp_ns, is_key);
+    if (result < 0)  // error
+      return false;
+
+    // Always set force_new_cluster_ to false after TestFrame.
+    force_new_cluster_ = false;
+
+    // A non-zero result means create a new cluster.
+    if (result > 0 && !MakeNewCluster(frame_timestamp_ns))
+      return false;
+
+    // Write queued (audio) frames.
+    const int frame_count = WriteFramesAll();
+    if (frame_count < 0)  // error
+      return false;
+
+    // Write the current frame to the current cluster (if TestFrame
+    // returns 0) or to a newly created cluster (TestFrame returns 1).
+    if (result <= 1)
+      return true;
+
+    // TestFrame returned 2, which means there was a large time
+    // difference between the cluster and the frame itself.  Do the
+    // test again, comparing the frame to the new cluster.
+  }
+}
+
+bool Segment::CheckHeaderInfo() {
+  if (!header_written_) {
+    if (!WriteSegmentHeader())
+      return false;
+
+    if (!seek_head_.AddSeekEntry(libwebm::kMkvCluster, MaxOffset()))
+      return false;
+
+    if (output_cues_ && cues_track_ == 0) {
+      // Check for a video track
+      for (uint32_t i = 0; i < tracks_.track_entries_size(); ++i) {
+        const Track* const track = tracks_.GetTrackByIndex(i);
+        if (!track)
+          return false;
+
+        if (tracks_.TrackIsVideo(track->number())) {
+          cues_track_ = track->number();
+          break;
+        }
+      }
+
+      // Set first track found
+      if (cues_track_ == 0) {
+        const Track* const track = tracks_.GetTrackByIndex(0);
+        if (!track)
+          return false;
+
+        cues_track_ = track->number();
+      }
+    }
+  }
+  return true;
+}
+
+void Segment::UpdateDocTypeVersion() {
+  for (uint32_t index = 0; index < tracks_.track_entries_size(); ++index) {
+    const Track* track = tracks_.GetTrackByIndex(index);
+    if (track == NULL)
+      break;
+    if ((track->codec_delay() || track->seek_pre_roll()) &&
+        doc_type_version_ < 4) {
+      doc_type_version_ = 4;
+      break;
+    }
+  }
+}
+
+bool Segment::UpdateChunkName(const char* ext, char** name) const {
+  if (!name || !ext)
+    return false;
+
+  char ext_chk[64];
+#ifdef _MSC_VER
+  sprintf_s(ext_chk, sizeof(ext_chk), "_%06d.%s", chunk_count_, ext);
+#else
+  snprintf(ext_chk, sizeof(ext_chk), "_%06d.%s", chunk_count_, ext);
+#endif
+
+  const size_t length = strlen(chunking_base_name_) + strlen(ext_chk) + 1;
+  char* const str = new (std::nothrow) char[length];  // NOLINT
+  if (!str)
+    return false;
+
+#ifdef _MSC_VER
+  strcpy_s(str, length - strlen(ext_chk), chunking_base_name_);
+  strcat_s(str, length, ext_chk);
+#else
+  strcpy(str, chunking_base_name_);
+  strcat(str, ext_chk);
+#endif
+
+  delete[] * name;
+  *name = str;
+
+  return true;
+}
+
+int64_t Segment::MaxOffset() {
+  if (!writer_header_)
+    return -1;
+
+  int64_t offset = writer_header_->Position() - payload_pos_;
+
+  if (chunking_) {
+    for (int32_t i = 0; i < cluster_list_size_; ++i) {
+      Cluster* const cluster = cluster_list_[i];
+      offset += cluster->Size();
+    }
+
+    if (writer_cues_)
+      offset += writer_cues_->Position();
+  }
+
+  return offset;
+}
+
+bool Segment::QueueFrame(Frame* frame) {
+  const int32_t new_size = frames_size_ + 1;
+
+  if (new_size > frames_capacity_) {
+    // Add more frames.
+    const int32_t new_capacity = (!frames_capacity_) ? 2 : frames_capacity_ * 2;
+
+    if (new_capacity < 1)
+      return false;
+
+    Frame** const frames = new (std::nothrow) Frame*[new_capacity];  // NOLINT
+    if (!frames)
+      return false;
+
+    for (int32_t i = 0; i < frames_size_; ++i) {
+      frames[i] = frames_[i];
+    }
+
+    delete[] frames_;
+    frames_ = frames;
+    frames_capacity_ = new_capacity;
+  }
+
+  frames_[frames_size_++] = frame;
+
+  return true;
+}
+
+int Segment::WriteFramesAll() {
+  if (frames_ == NULL)
+    return 0;
+
+  if (cluster_list_size_ < 1)
+    return -1;
+
+  Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
+
+  if (!cluster)
+    return -1;
+
+  for (int32_t i = 0; i < frames_size_; ++i) {
+    Frame*& frame = frames_[i];
+    // TODO(jzern/vigneshv): using Segment::AddGenericFrame here would limit the
+    // places where |doc_type_version_| needs to be updated.
+    if (frame->discard_padding() != 0)
+      doc_type_version_ = 4;
+    if (!cluster->AddFrame(frame))
+      return -1;
+
+    if (new_cuepoint_ && cues_track_ == frame->track_number()) {
+      if (!AddCuePoint(frame->timestamp(), cues_track_))
+        return -1;
+    }
+
+    if (frame->timestamp() > last_timestamp_) {
+      last_timestamp_ = frame->timestamp();
+      last_track_timestamp_[frame->track_number() - 1] = frame->timestamp();
+    }
+
+    delete frame;
+    frame = NULL;
+  }
+
+  const int result = frames_size_;
+  frames_size_ = 0;
+
+  return result;
+}
+
+bool Segment::WriteFramesLessThan(uint64_t timestamp) {
+  // Check |cluster_list_size_| to see if this is the first cluster. If it is
+  // the first cluster the audio frames that are less than the first video
+  // timesatmp will be written in a later step.
+  if (frames_size_ > 0 && cluster_list_size_ > 0) {
+    if (!frames_)
+      return false;
+
+    Cluster* const cluster = cluster_list_[cluster_list_size_ - 1];
+    if (!cluster)
+      return false;
+
+    int32_t shift_left = 0;
+
+    // TODO(fgalligan): Change this to use the durations of frames instead of
+    // the next frame's start time if the duration is accurate.
+    for (int32_t i = 1; i < frames_size_; ++i) {
+      const Frame* const frame_curr = frames_[i];
+
+      if (frame_curr->timestamp() > timestamp)
+        break;
+
+      const Frame* const frame_prev = frames_[i - 1];
+      if (frame_prev->discard_padding() != 0)
+        doc_type_version_ = 4;
+      if (!cluster->AddFrame(frame_prev))
+        return false;
+
+      if (new_cuepoint_ && cues_track_ == frame_prev->track_number()) {
+        if (!AddCuePoint(frame_prev->timestamp(), cues_track_))
+          return false;
+      }
+
+      ++shift_left;
+      if (frame_prev->timestamp() > last_timestamp_) {
+        last_timestamp_ = frame_prev->timestamp();
+        last_track_timestamp_[frame_prev->track_number() - 1] =
+            frame_prev->timestamp();
+      }
+
+      delete frame_prev;
+    }
+
+    if (shift_left > 0) {
+      if (shift_left >= frames_size_)
+        return false;
+
+      const int32_t new_frames_size = frames_size_ - shift_left;
+      for (int32_t i = 0; i < new_frames_size; ++i) {
+        frames_[i] = frames_[i + shift_left];
+      }
+
+      frames_size_ = new_frames_size;
+    }
+  }
+
+  return true;
+}
+
+}  // namespace mkvmuxer
diff --git a/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxer.h b/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxer.h
new file mode 100644
index 0000000000..01b26a212b
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxer.h
@@ -0,0 +1,1695 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#ifndef MKVMUXER_MKVMUXER_H_
+#define MKVMUXER_MKVMUXER_H_
+
+#include <stdint.h>
+
+#include <cstddef>
+#include <list>
+#include <map>
+
+#include "common/webmids.h"
+#include "mkvmuxer/mkvmuxertypes.h"
+
+// For a description of the WebM elements see
+// http://www.webmproject.org/code/specs/container/.
+
+namespace mkvparser {
+class IMkvReader;
+}  // namespace mkvparser
+
+namespace mkvmuxer {
+
+class MkvWriter;
+class Segment;
+
+const uint64_t kMaxTrackNumber = 126;
+
+///////////////////////////////////////////////////////////////
+// Interface used by the mkvmuxer to write out the Mkv data.
+class IMkvWriter {
+ public:
+  // Writes out |len| bytes of |buf|. Returns 0 on success.
+  virtual int32 Write(const void* buf, uint32 len) = 0;
+
+  // Returns the offset of the output position from the beginning of the
+  // output.
+  virtual int64 Position() const = 0;
+
+  // Set the current File position. Returns 0 on success.
+  virtual int32 Position(int64 position) = 0;
+
+  // Returns true if the writer is seekable.
+  virtual bool Seekable() const = 0;
+
+  // Element start notification. Called whenever an element identifier is about
+  // to be written to the stream. |element_id| is the element identifier, and
+  // |position| is the location in the WebM stream where the first octet of the
+  // element identifier will be written.
+  // Note: the |MkvId| enumeration in webmids.hpp defines element values.
+  virtual void ElementStartNotify(uint64 element_id, int64 position) = 0;
+
+ protected:
+  IMkvWriter();
+  virtual ~IMkvWriter();
+
+ private:
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(IMkvWriter);
+};
+
+// Writes out the EBML header for a WebM file. This function must be called
+// before any other libwebm writing functions are called.
+bool WriteEbmlHeader(IMkvWriter* writer, uint64_t doc_type_version);
+
+// Deprecated. Writes out EBML header with doc_type_version as
+// kDefaultDocTypeVersion. Exists for backward compatibility.
+bool WriteEbmlHeader(IMkvWriter* writer);
+
+// Copies in Chunk from source to destination between the given byte positions
+bool ChunkedCopy(mkvparser::IMkvReader* source, IMkvWriter* dst, int64_t start,
+                 int64_t size);
+
+///////////////////////////////////////////////////////////////
+// Class to hold data the will be written to a block.
+class Frame {
+ public:
+  Frame();
+  ~Frame();
+
+  // Sets this frame's contents based on |frame|. Returns true on success. On
+  // failure, this frame's existing contents may be lost.
+  bool CopyFrom(const Frame& frame);
+
+  // Copies |frame| data into |frame_|. Returns true on success.
+  bool Init(const uint8_t* frame, uint64_t length);
+
+  // Copies |additional| data into |additional_|. Returns true on success.
+  bool AddAdditionalData(const uint8_t* additional, uint64_t length,
+                         uint64_t add_id);
+
+  // Returns true if the frame has valid parameters.
+  bool IsValid() const;
+
+  // Returns true if the frame can be written as a SimpleBlock based on current
+  // parameters.
+  bool CanBeSimpleBlock() const;
+
+  uint64_t add_id() const { return add_id_; }
+  const uint8_t* additional() const { return additional_; }
+  uint64_t additional_length() const { return additional_length_; }
+  void set_duration(uint64_t duration);
+  uint64_t duration() const { return duration_; }
+  bool duration_set() const { return duration_set_; }
+  const uint8_t* frame() const { return frame_; }
+  void set_is_key(bool key) { is_key_ = key; }
+  bool is_key() const { return is_key_; }
+  uint64_t length() const { return length_; }
+  void set_track_number(uint64_t track_number) { track_number_ = track_number; }
+  uint64_t track_number() const { return track_number_; }
+  void set_timestamp(uint64_t timestamp) { timestamp_ = timestamp; }
+  uint64_t timestamp() const { return timestamp_; }
+  void set_discard_padding(int64_t discard_padding) {
+    discard_padding_ = discard_padding;
+  }
+  int64_t discard_padding() const { return discard_padding_; }
+  void set_reference_block_timestamp(int64_t reference_block_timestamp);
+  int64_t reference_block_timestamp() const {
+    return reference_block_timestamp_;
+  }
+  bool reference_block_timestamp_set() const {
+    return reference_block_timestamp_set_;
+  }
+
+ private:
+  // Id of the Additional data.
+  uint64_t add_id_;
+
+  // Pointer to additional data. Owned by this class.
+  uint8_t* additional_;
+
+  // Length of the additional data.
+  uint64_t additional_length_;
+
+  // Duration of the frame in nanoseconds.
+  uint64_t duration_;
+
+  // Flag indicating that |duration_| has been set. Setting duration causes the
+  // frame to be written out as a Block with BlockDuration instead of as a
+  // SimpleBlock.
+  bool duration_set_;
+
+  // Pointer to the data. Owned by this class.
+  uint8_t* frame_;
+
+  // Flag telling if the data should set the key flag of a block.
+  bool is_key_;
+
+  // Length of the data.
+  uint64_t length_;
+
+  // Mkv track number the data is associated with.
+  uint64_t track_number_;
+
+  // Timestamp of the data in nanoseconds.
+  uint64_t timestamp_;
+
+  // Discard padding for the frame.
+  int64_t discard_padding_;
+
+  // Reference block timestamp.
+  int64_t reference_block_timestamp_;
+
+  // Flag indicating if |reference_block_timestamp_| has been set.
+  bool reference_block_timestamp_set_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Frame);
+};
+
+///////////////////////////////////////////////////////////////
+// Class to hold one cue point in a Cues element.
+class CuePoint {
+ public:
+  CuePoint();
+  ~CuePoint();
+
+  // Returns the size in bytes for the entire CuePoint element.
+  uint64_t Size() const;
+
+  // Output the CuePoint element to the writer. Returns true on success.
+  bool Write(IMkvWriter* writer) const;
+
+  void set_time(uint64_t time) { time_ = time; }
+  uint64_t time() const { return time_; }
+  void set_track(uint64_t track) { track_ = track; }
+  uint64_t track() const { return track_; }
+  void set_cluster_pos(uint64_t cluster_pos) { cluster_pos_ = cluster_pos; }
+  uint64_t cluster_pos() const { return cluster_pos_; }
+  void set_block_number(uint64_t block_number) { block_number_ = block_number; }
+  uint64_t block_number() const { return block_number_; }
+  void set_output_block_number(bool output_block_number) {
+    output_block_number_ = output_block_number;
+  }
+  bool output_block_number() const { return output_block_number_; }
+
+ private:
+  // Returns the size in bytes for the payload of the CuePoint element.
+  uint64_t PayloadSize() const;
+
+  // Absolute timecode according to the segment time base.
+  uint64_t time_;
+
+  // The Track element associated with the CuePoint.
+  uint64_t track_;
+
+  // The position of the Cluster containing the Block.
+  uint64_t cluster_pos_;
+
+  // Number of the Block within the Cluster, starting from 1.
+  uint64_t block_number_;
+
+  // If true the muxer will write out the block number for the cue if the
+  // block number is different than the default of 1. Default is set to true.
+  bool output_block_number_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(CuePoint);
+};
+
+///////////////////////////////////////////////////////////////
+// Cues element.
+class Cues {
+ public:
+  Cues();
+  ~Cues();
+
+  // Adds a cue point to the Cues element. Returns true on success.
+  bool AddCue(CuePoint* cue);
+
+  // Returns the cue point by index. Returns NULL if there is no cue point
+  // match.
+  CuePoint* GetCueByIndex(int32_t index) const;
+
+  // Returns the total size of the Cues element
+  uint64_t Size();
+
+  // Output the Cues element to the writer. Returns true on success.
+  bool Write(IMkvWriter* writer) const;
+
+  int32_t cue_entries_size() const { return cue_entries_size_; }
+  void set_output_block_number(bool output_block_number) {
+    output_block_number_ = output_block_number;
+  }
+  bool output_block_number() const { return output_block_number_; }
+
+ private:
+  // Number of allocated elements in |cue_entries_|.
+  int32_t cue_entries_capacity_;
+
+  // Number of CuePoints in |cue_entries_|.
+  int32_t cue_entries_size_;
+
+  // CuePoint list.
+  CuePoint** cue_entries_;
+
+  // If true the muxer will write out the block number for the cue if the
+  // block number is different than the default of 1. Default is set to true.
+  bool output_block_number_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Cues);
+};
+
+///////////////////////////////////////////////////////////////
+// ContentEncAESSettings element
+class ContentEncAESSettings {
+ public:
+  enum { kCTR = 1 };
+
+  ContentEncAESSettings();
+  ~ContentEncAESSettings() {}
+
+  // Returns the size in bytes for the ContentEncAESSettings element.
+  uint64_t Size() const;
+
+  // Writes out the ContentEncAESSettings element to |writer|. Returns true on
+  // success.
+  bool Write(IMkvWriter* writer) const;
+
+  uint64_t cipher_mode() const { return cipher_mode_; }
+
+ private:
+  // Returns the size in bytes for the payload of the ContentEncAESSettings
+  // element.
+  uint64_t PayloadSize() const;
+
+  // Sub elements
+  uint64_t cipher_mode_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(ContentEncAESSettings);
+};
+
+///////////////////////////////////////////////////////////////
+// ContentEncoding element
+// Elements used to describe if the track data has been encrypted or
+// compressed with zlib or header stripping.
+// Currently only whole frames can be encrypted with AES. This dictates that
+// ContentEncodingOrder will be 0, ContentEncodingScope will be 1,
+// ContentEncodingType will be 1, and ContentEncAlgo will be 5.
+class ContentEncoding {
+ public:
+  ContentEncoding();
+  ~ContentEncoding();
+
+  // Sets the content encryption id. Copies |length| bytes from |id| to
+  // |enc_key_id_|. Returns true on success.
+  bool SetEncryptionID(const uint8_t* id, uint64_t length);
+
+  // Returns the size in bytes for the ContentEncoding element.
+  uint64_t Size() const;
+
+  // Writes out the ContentEncoding element to |writer|. Returns true on
+  // success.
+  bool Write(IMkvWriter* writer) const;
+
+  uint64_t enc_algo() const { return enc_algo_; }
+  uint64_t encoding_order() const { return encoding_order_; }
+  uint64_t encoding_scope() const { return encoding_scope_; }
+  uint64_t encoding_type() const { return encoding_type_; }
+  ContentEncAESSettings* enc_aes_settings() { return &enc_aes_settings_; }
+
+ private:
+  // Returns the size in bytes for the encoding elements.
+  uint64_t EncodingSize(uint64_t compresion_size,
+                        uint64_t encryption_size) const;
+
+  // Returns the size in bytes for the encryption elements.
+  uint64_t EncryptionSize() const;
+
+  // Track element names
+  uint64_t enc_algo_;
+  uint8_t* enc_key_id_;
+  uint64_t encoding_order_;
+  uint64_t encoding_scope_;
+  uint64_t encoding_type_;
+
+  // ContentEncAESSettings element.
+  ContentEncAESSettings enc_aes_settings_;
+
+  // Size of the ContentEncKeyID data in bytes.
+  uint64_t enc_key_id_length_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(ContentEncoding);
+};
+
+///////////////////////////////////////////////////////////////
+// Colour element.
+struct PrimaryChromaticity {
+  PrimaryChromaticity(float x_val, float y_val) : x(x_val), y(y_val) {}
+  PrimaryChromaticity() : x(0), y(0) {}
+  ~PrimaryChromaticity() {}
+  uint64_t PrimaryChromaticityPayloadSize(libwebm::MkvId x_id,
+                                          libwebm::MkvId y_id) const;
+  bool Write(IMkvWriter* writer, libwebm::MkvId x_id,
+             libwebm::MkvId y_id) const;
+
+  float x;
+  float y;
+};
+
+class MasteringMetadata {
+ public:
+  static const float kValueNotPresent;
+
+  MasteringMetadata()
+      : luminance_max(kValueNotPresent),
+        luminance_min(kValueNotPresent),
+        r_(NULL),
+        g_(NULL),
+        b_(NULL),
+        white_point_(NULL) {}
+  ~MasteringMetadata() {
+    delete r_;
+    delete g_;
+    delete b_;
+    delete white_point_;
+  }
+
+  // Returns total size of the MasteringMetadata element.
+  uint64_t MasteringMetadataSize() const;
+  bool Write(IMkvWriter* writer) const;
+
+  // Copies non-null chromaticity.
+  bool SetChromaticity(const PrimaryChromaticity* r,
+                       const PrimaryChromaticity* g,
+                       const PrimaryChromaticity* b,
+                       const PrimaryChromaticity* white_point);
+  const PrimaryChromaticity* r() const { return r_; }
+  const PrimaryChromaticity* g() const { return g_; }
+  const PrimaryChromaticity* b() const { return b_; }
+  const PrimaryChromaticity* white_point() const { return white_point_; }
+
+  float luminance_max;
+  float luminance_min;
+
+ private:
+  // Returns size of MasteringMetadata child elements.
+  uint64_t PayloadSize() const;
+
+  PrimaryChromaticity* r_;
+  PrimaryChromaticity* g_;
+  PrimaryChromaticity* b_;
+  PrimaryChromaticity* white_point_;
+};
+
+class Colour {
+ public:
+  static const uint64_t kValueNotPresent;
+  Colour()
+      : matrix_coefficients(kValueNotPresent),
+        bits_per_channel(kValueNotPresent),
+        chroma_subsampling_horz(kValueNotPresent),
+        chroma_subsampling_vert(kValueNotPresent),
+        cb_subsampling_horz(kValueNotPresent),
+        cb_subsampling_vert(kValueNotPresent),
+        chroma_siting_horz(kValueNotPresent),
+        chroma_siting_vert(kValueNotPresent),
+        range(kValueNotPresent),
+        transfer_characteristics(kValueNotPresent),
+        primaries(kValueNotPresent),
+        max_cll(kValueNotPresent),
+        max_fall(kValueNotPresent),
+        mastering_metadata_(NULL) {}
+  ~Colour() { delete mastering_metadata_; }
+
+  // Returns total size of the Colour element.
+  uint64_t ColourSize() const;
+  bool Write(IMkvWriter* writer) const;
+
+  // Deep copies |mastering_metadata|.
+  bool SetMasteringMetadata(const MasteringMetadata& mastering_metadata);
+
+  const MasteringMetadata* mastering_metadata() const {
+    return mastering_metadata_;
+  }
+
+  uint64_t matrix_coefficients;
+  uint64_t bits_per_channel;
+  uint64_t chroma_subsampling_horz;
+  uint64_t chroma_subsampling_vert;
+  uint64_t cb_subsampling_horz;
+  uint64_t cb_subsampling_vert;
+  uint64_t chroma_siting_horz;
+  uint64_t chroma_siting_vert;
+  uint64_t range;
+  uint64_t transfer_characteristics;
+  uint64_t primaries;
+  uint64_t max_cll;
+  uint64_t max_fall;
+
+ private:
+  // Returns size of Colour child elements.
+  uint64_t PayloadSize() const;
+
+  MasteringMetadata* mastering_metadata_;
+};
+
+///////////////////////////////////////////////////////////////
+// Track element.
+class Track {
+ public:
+  // The |seed| parameter is used to synthesize a UID for the track.
+  explicit Track(unsigned int* seed);
+  virtual ~Track();
+
+  // Adds a ContentEncoding element to the Track. Returns true on success.
+  virtual bool AddContentEncoding();
+
+  // Returns the ContentEncoding by index. Returns NULL if there is no
+  // ContentEncoding match.
+  ContentEncoding* GetContentEncodingByIndex(uint32_t index) const;
+
+  // Returns the size in bytes for the payload of the Track element.
+  virtual uint64_t PayloadSize() const;
+
+  // Returns the size in bytes of the Track element.
+  virtual uint64_t Size() const;
+
+  // Output the Track element to the writer. Returns true on success.
+  virtual bool Write(IMkvWriter* writer) const;
+
+  // Sets the CodecPrivate element of the Track element. Copies |length|
+  // bytes from |codec_private| to |codec_private_|. Returns true on success.
+  bool SetCodecPrivate(const uint8_t* codec_private, uint64_t length);
+
+  void set_codec_id(const char* codec_id);
+  const char* codec_id() const { return codec_id_; }
+  const uint8_t* codec_private() const { return codec_private_; }
+  void set_language(const char* language);
+  const char* language() const { return language_; }
+  void set_max_block_additional_id(uint64_t max_block_additional_id) {
+    max_block_additional_id_ = max_block_additional_id;
+  }
+  uint64_t max_block_additional_id() const { return max_block_additional_id_; }
+  void set_name(const char* name);
+  const char* name() const { return name_; }
+  void set_number(uint64_t number) { number_ = number; }
+  uint64_t number() const { return number_; }
+  void set_type(uint64_t type) { type_ = type; }
+  uint64_t type() const { return type_; }
+  void set_uid(uint64_t uid) { uid_ = uid; }
+  uint64_t uid() const { return uid_; }
+  void set_codec_delay(uint64_t codec_delay) { codec_delay_ = codec_delay; }
+  uint64_t codec_delay() const { return codec_delay_; }
+  void set_seek_pre_roll(uint64_t seek_pre_roll) {
+    seek_pre_roll_ = seek_pre_roll;
+  }
+  uint64_t seek_pre_roll() const { return seek_pre_roll_; }
+  void set_default_duration(uint64_t default_duration) {
+    default_duration_ = default_duration;
+  }
+  uint64_t default_duration() const { return default_duration_; }
+
+  uint64_t codec_private_length() const { return codec_private_length_; }
+  uint32_t content_encoding_entries_size() const {
+    return content_encoding_entries_size_;
+  }
+
+ private:
+  // Track element names.
+  char* codec_id_;
+  uint8_t* codec_private_;
+  char* language_;
+  uint64_t max_block_additional_id_;
+  char* name_;
+  uint64_t number_;
+  uint64_t type_;
+  uint64_t uid_;
+  uint64_t codec_delay_;
+  uint64_t seek_pre_roll_;
+  uint64_t default_duration_;
+
+  // Size of the CodecPrivate data in bytes.
+  uint64_t codec_private_length_;
+
+  // ContentEncoding element list.
+  ContentEncoding** content_encoding_entries_;
+
+  // Number of ContentEncoding elements added.
+  uint32_t content_encoding_entries_size_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Track);
+};
+
+///////////////////////////////////////////////////////////////
+// Track that has video specific elements.
+class VideoTrack : public Track {
+ public:
+  // Supported modes for stereo 3D.
+  enum StereoMode {
+    kMono = 0,
+    kSideBySideLeftIsFirst = 1,
+    kTopBottomRightIsFirst = 2,
+    kTopBottomLeftIsFirst = 3,
+    kSideBySideRightIsFirst = 11
+  };
+
+  enum AlphaMode { kNoAlpha = 0, kAlpha = 1 };
+
+  // The |seed| parameter is used to synthesize a UID for the track.
+  explicit VideoTrack(unsigned int* seed);
+  virtual ~VideoTrack();
+
+  // Returns the size in bytes for the payload of the Track element plus the
+  // video specific elements.
+  virtual uint64_t PayloadSize() const;
+
+  // Output the VideoTrack element to the writer. Returns true on success.
+  virtual bool Write(IMkvWriter* writer) const;
+
+  // Sets the video's stereo mode. Returns true on success.
+  bool SetStereoMode(uint64_t stereo_mode);
+
+  // Sets the video's alpha mode. Returns true on success.
+  bool SetAlphaMode(uint64_t alpha_mode);
+
+  void set_display_height(uint64_t height) { display_height_ = height; }
+  uint64_t display_height() const { return display_height_; }
+  void set_display_width(uint64_t width) { display_width_ = width; }
+  uint64_t display_width() const { return display_width_; }
+
+  void set_crop_left(uint64_t crop_left) { crop_left_ = crop_left; }
+  uint64_t crop_left() const { return crop_left_; }
+  void set_crop_right(uint64_t crop_right) { crop_right_ = crop_right; }
+  uint64_t crop_right() const { return crop_right_; }
+  void set_crop_top(uint64_t crop_top) { crop_top_ = crop_top; }
+  uint64_t crop_top() const { return crop_top_; }
+  void set_crop_bottom(uint64_t crop_bottom) { crop_bottom_ = crop_bottom; }
+  uint64_t crop_bottom() const { return crop_bottom_; }
+
+  void set_frame_rate(double frame_rate) { frame_rate_ = frame_rate; }
+  double frame_rate() const { return frame_rate_; }
+  void set_height(uint64_t height) { height_ = height; }
+  uint64_t height() const { return height_; }
+  uint64_t stereo_mode() { return stereo_mode_; }
+  uint64_t alpha_mode() { return alpha_mode_; }
+  void set_width(uint64_t width) { width_ = width; }
+  uint64_t width() const { return width_; }
+
+  Colour* colour() { return colour_; }
+
+  // Deep copies |colour|.
+  bool SetColour(const Colour& colour);
+
+ private:
+  // Returns the size in bytes of the Video element.
+  uint64_t VideoPayloadSize() const;
+
+  // Video track element names.
+  uint64_t display_height_;
+  uint64_t display_width_;
+  uint64_t crop_left_;
+  uint64_t crop_right_;
+  uint64_t crop_top_;
+  uint64_t crop_bottom_;
+  double frame_rate_;
+  uint64_t height_;
+  uint64_t stereo_mode_;
+  uint64_t alpha_mode_;
+  uint64_t width_;
+
+  Colour* colour_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(VideoTrack);
+};
+
+///////////////////////////////////////////////////////////////
+// Track that has audio specific elements.
+class AudioTrack : public Track {
+ public:
+  // The |seed| parameter is used to synthesize a UID for the track.
+  explicit AudioTrack(unsigned int* seed);
+  virtual ~AudioTrack();
+
+  // Returns the size in bytes for the payload of the Track element plus the
+  // audio specific elements.
+  virtual uint64_t PayloadSize() const;
+
+  // Output the AudioTrack element to the writer. Returns true on success.
+  virtual bool Write(IMkvWriter* writer) const;
+
+  void set_bit_depth(uint64_t bit_depth) { bit_depth_ = bit_depth; }
+  uint64_t bit_depth() const { return bit_depth_; }
+  void set_channels(uint64_t channels) { channels_ = channels; }
+  uint64_t channels() const { return channels_; }
+  void set_sample_rate(double sample_rate) { sample_rate_ = sample_rate; }
+  double sample_rate() const { return sample_rate_; }
+
+ private:
+  // Audio track element names.
+  uint64_t bit_depth_;
+  uint64_t channels_;
+  double sample_rate_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(AudioTrack);
+};
+
+///////////////////////////////////////////////////////////////
+// Tracks element
+class Tracks {
+ public:
+  // Audio and video type defined by the Matroska specs.
+  enum { kVideo = 0x1, kAudio = 0x2 };
+
+  static const char kOpusCodecId[];
+  static const char kVorbisCodecId[];
+  static const char kVp8CodecId[];
+  static const char kVp9CodecId[];
+  static const char kAv1CodecId[];
+
+  Tracks();
+  ~Tracks();
+
+  // Adds a Track element to the Tracks object. |track| will be owned and
+  // deleted by the Tracks object. Returns true on success. |number| is the
+  // number to use for the track. |number| must be >= 0. If |number| == 0
+  // then the muxer will decide on the track number.
+  bool AddTrack(Track* track, int32_t number);
+
+  // Returns the track by index. Returns NULL if there is no track match.
+  const Track* GetTrackByIndex(uint32_t idx) const;
+
+  // Search the Tracks and return the track that matches |tn|. Returns NULL
+  // if there is no track match.
+  Track* GetTrackByNumber(uint64_t track_number) const;
+
+  // Returns true if the track number is an audio track.
+  bool TrackIsAudio(uint64_t track_number) const;
+
+  // Returns true if the track number is a video track.
+  bool TrackIsVideo(uint64_t track_number) const;
+
+  // Output the Tracks element to the writer. Returns true on success.
+  bool Write(IMkvWriter* writer) const;
+
+  uint32_t track_entries_size() const { return track_entries_size_; }
+
+ private:
+  // Track element list.
+  Track** track_entries_;
+
+  // Number of Track elements added.
+  uint32_t track_entries_size_;
+
+  // Whether or not Tracks element has already been written via IMkvWriter.
+  mutable bool wrote_tracks_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Tracks);
+};
+
+///////////////////////////////////////////////////////////////
+// Chapter element
+//
+class Chapter {
+ public:
+  // Set the identifier for this chapter.  (This corresponds to the
+  // Cue Identifier line in WebVTT.)
+  // TODO(matthewjheaney): the actual serialization of this item in
+  // MKV is pending.
+  bool set_id(const char* id);
+
+  // Converts the nanosecond start and stop times of this chapter to
+  // their corresponding timecode values, and stores them that way.
+  void set_time(const Segment& segment, uint64_t start_time_ns,
+                uint64_t end_time_ns);
+
+  // Sets the uid for this chapter. Primarily used to enable
+  // deterministic output from the muxer.
+  void set_uid(const uint64_t uid) { uid_ = uid; }
+
+  // Add a title string to this chapter, per the semantics described
+  // here:
+  //  http://www.matroska.org/technical/specs/index.html
+  //
+  // The title ("chapter string") is a UTF-8 string.
+  //
+  // The language has ISO 639-2 representation, described here:
+  //  http://www.loc.gov/standards/iso639-2/englangn.html
+  //  http://www.loc.gov/standards/iso639-2/php/English_list.php
+  // If you specify NULL as the language value, this implies
+  // English ("eng").
+  //
+  // The country value corresponds to the codes listed here:
+  //  http://www.iana.org/domains/root/db/
+  //
+  // The function returns false if the string could not be allocated.
+  bool add_string(const char* title, const char* language, const char* country);
+
+ private:
+  friend class Chapters;
+
+  // For storage of chapter titles that differ by language.
+  class Display {
+   public:
+    // Establish representation invariant for new Display object.
+    void Init();
+
+    // Reclaim resources, in anticipation of destruction.
+    void Clear();
+
+    // Copies the title to the |title_| member.  Returns false on
+    // error.
+    bool set_title(const char* title);
+
+    // Copies the language to the |language_| member.  Returns false
+    // on error.
+    bool set_language(const char* language);
+
+    // Copies the country to the |country_| member.  Returns false on
+    // error.
+    bool set_country(const char* country);
+
+    // If |writer| is non-NULL, serialize the Display sub-element of
+    // the Atom into the stream.  Returns the Display element size on
+    // success, 0 if error.
+    uint64_t WriteDisplay(IMkvWriter* writer) const;
+
+   private:
+    char* title_;
+    char* language_;
+    char* country_;
+  };
+
+  Chapter();
+  ~Chapter();
+
+  // Establish the representation invariant for a newly-created
+  // Chapter object.  The |seed| parameter is used to create the UID
+  // for this chapter atom.
+  void Init(unsigned int* seed);
+
+  // Copies this Chapter object to a different one.  This is used when
+  // expanding a plain array of Chapter objects (see Chapters).
+  void ShallowCopy(Chapter* dst) const;
+
+  // Reclaim resources used by this Chapter object, pending its
+  // destruction.
+  void Clear();
+
+  // If there is no storage remaining on the |displays_| array for a
+  // new display object, creates a new, longer array and copies the
+  // existing Display objects to the new array.  Returns false if the
+  // array cannot be expanded.
+  bool ExpandDisplaysArray();
+
+  // If |writer| is non-NULL, serialize the Atom sub-element into the
+  // stream.  Returns the total size of the element on success, 0 if
+  // error.
+  uint64_t WriteAtom(IMkvWriter* writer) const;
+
+  // The string identifier for this chapter (corresponds to WebVTT cue
+  // identifier).
+  char* id_;
+
+  // Start timecode of the chapter.
+  uint64_t start_timecode_;
+
+  // Stop timecode of the chapter.
+  uint64_t end_timecode_;
+
+  // The binary identifier for this chapter.
+  uint64_t uid_;
+
+  // The Atom element can contain multiple Display sub-elements, as
+  // the same logical title can be rendered in different languages.
+  Display* displays_;
+
+  // The physical length (total size) of the |displays_| array.
+  int displays_size_;
+
+  // The logical length (number of active elements) on the |displays_|
+  // array.
+  int displays_count_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Chapter);
+};
+
+///////////////////////////////////////////////////////////////
+// Chapters element
+//
+class Chapters {
+ public:
+  Chapters();
+  ~Chapters();
+
+  Chapter* AddChapter(unsigned int* seed);
+
+  // Returns the number of chapters that have been added.
+  int Count() const;
+
+  // Output the Chapters element to the writer. Returns true on success.
+  bool Write(IMkvWriter* writer) const;
+
+ private:
+  // Expands the chapters_ array if there is not enough space to contain
+  // another chapter object.  Returns true on success.
+  bool ExpandChaptersArray();
+
+  // If |writer| is non-NULL, serialize the Edition sub-element of the
+  // Chapters element into the stream.  Returns the Edition element
+  // size on success, 0 if error.
+  uint64_t WriteEdition(IMkvWriter* writer) const;
+
+  // Total length of the chapters_ array.
+  int chapters_size_;
+
+  // Number of active chapters on the chapters_ array.
+  int chapters_count_;
+
+  // Array for storage of chapter objects.
+  Chapter* chapters_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Chapters);
+};
+
+///////////////////////////////////////////////////////////////
+// Tag element
+//
+class Tag {
+ public:
+  bool add_simple_tag(const char* tag_name, const char* tag_string);
+
+ private:
+  // Tags calls Clear and the destructor of Tag
+  friend class Tags;
+
+  // For storage of simple tags
+  class SimpleTag {
+   public:
+    // Establish representation invariant for new SimpleTag object.
+    void Init();
+
+    // Reclaim resources, in anticipation of destruction.
+    void Clear();
+
+    // Copies the title to the |tag_name_| member.  Returns false on
+    // error.
+    bool set_tag_name(const char* tag_name);
+
+    // Copies the language to the |tag_string_| member.  Returns false
+    // on error.
+    bool set_tag_string(const char* tag_string);
+
+    // If |writer| is non-NULL, serialize the SimpleTag sub-element of
+    // the Atom into the stream.  Returns the SimpleTag element size on
+    // success, 0 if error.
+    uint64_t Write(IMkvWriter* writer) const;
+
+   private:
+    char* tag_name_;
+    char* tag_string_;
+  };
+
+  Tag();
+  ~Tag();
+
+  // Copies this Tag object to a different one.  This is used when
+  // expanding a plain array of Tag objects (see Tags).
+  void ShallowCopy(Tag* dst) const;
+
+  // Reclaim resources used by this Tag object, pending its
+  // destruction.
+  void Clear();
+
+  // If there is no storage remaining on the |simple_tags_| array for a
+  // new display object, creates a new, longer array and copies the
+  // existing SimpleTag objects to the new array.  Returns false if the
+  // array cannot be expanded.
+  bool ExpandSimpleTagsArray();
+
+  // If |writer| is non-NULL, serialize the Tag sub-element into the
+  // stream.  Returns the total size of the element on success, 0 if
+  // error.
+  uint64_t Write(IMkvWriter* writer) const;
+
+  // The Atom element can contain multiple SimpleTag sub-elements
+  SimpleTag* simple_tags_;
+
+  // The physical length (total size) of the |simple_tags_| array.
+  int simple_tags_size_;
+
+  // The logical length (number of active elements) on the |simple_tags_|
+  // array.
+  int simple_tags_count_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Tag);
+};
+
+///////////////////////////////////////////////////////////////
+// Tags element
+//
+class Tags {
+ public:
+  Tags();
+  ~Tags();
+
+  Tag* AddTag();
+
+  // Returns the number of tags that have been added.
+  int Count() const;
+
+  // Output the Tags element to the writer. Returns true on success.
+  bool Write(IMkvWriter* writer) const;
+
+ private:
+  // Expands the tags_ array if there is not enough space to contain
+  // another tag object.  Returns true on success.
+  bool ExpandTagsArray();
+
+  // Total length of the tags_ array.
+  int tags_size_;
+
+  // Number of active tags on the tags_ array.
+  int tags_count_;
+
+  // Array for storage of tag objects.
+  Tag* tags_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Tags);
+};
+
+///////////////////////////////////////////////////////////////
+// Cluster element
+//
+// Notes:
+//  |Init| must be called before any other method in this class.
+class Cluster {
+ public:
+  // |timecode| is the absolute timecode of the cluster. |cues_pos| is the
+  // position for the cluster within the segment that should be written in
+  // the cues element. |timecode_scale| is the timecode scale of the segment.
+  Cluster(uint64_t timecode, int64_t cues_pos, uint64_t timecode_scale,
+          bool write_last_frame_with_duration = false,
+          bool fixed_size_timecode = false);
+  ~Cluster();
+
+  bool Init(IMkvWriter* ptr_writer);
+
+  // Adds a frame to be output in the file. The frame is written out through
+  // |writer_| if successful. Returns true on success.
+  bool AddFrame(const Frame* frame);
+
+  // Adds a frame to be output in the file. The frame is written out through
+  // |writer_| if successful. Returns true on success.
+  // Inputs:
+  //   data: Pointer to the data
+  //   length: Length of the data
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.  The range of allowed values is [1, 126].
+  //   timecode:     Absolute (not relative to cluster) timestamp of the
+  //                 frame, expressed in timecode units.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrame(const uint8_t* data, uint64_t length, uint64_t track_number,
+                uint64_t timecode,  // timecode units (absolute)
+                bool is_key);
+
+  // Adds a frame to be output in the file. The frame is written out through
+  // |writer_| if successful. Returns true on success.
+  // Inputs:
+  //   data: Pointer to the data
+  //   length: Length of the data
+  //   additional: Pointer to the additional data
+  //   additional_length: Length of the additional data
+  //   add_id: Value of BlockAddID element
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.  The range of allowed values is [1, 126].
+  //   abs_timecode: Absolute (not relative to cluster) timestamp of the
+  //                 frame, expressed in timecode units.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrameWithAdditional(const uint8_t* data, uint64_t length,
+                              const uint8_t* additional,
+                              uint64_t additional_length, uint64_t add_id,
+                              uint64_t track_number, uint64_t abs_timecode,
+                              bool is_key);
+
+  // Adds a frame to be output in the file. The frame is written out through
+  // |writer_| if successful. Returns true on success.
+  // Inputs:
+  //   data: Pointer to the data.
+  //   length: Length of the data.
+  //   discard_padding: DiscardPadding element value.
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.  The range of allowed values is [1, 126].
+  //   abs_timecode: Absolute (not relative to cluster) timestamp of the
+  //                 frame, expressed in timecode units.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrameWithDiscardPadding(const uint8_t* data, uint64_t length,
+                                  int64_t discard_padding,
+                                  uint64_t track_number, uint64_t abs_timecode,
+                                  bool is_key);
+
+  // Writes a frame of metadata to the output medium; returns true on
+  // success.
+  // Inputs:
+  //   data: Pointer to the data
+  //   length: Length of the data
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.  The range of allowed values is [1, 126].
+  //   timecode:     Absolute (not relative to cluster) timestamp of the
+  //                 metadata frame, expressed in timecode units.
+  //   duration:     Duration of metadata frame, in timecode units.
+  //
+  // The metadata frame is written as a block group, with a duration
+  // sub-element but no reference time sub-elements (indicating that
+  // it is considered a keyframe, per Matroska semantics).
+  bool AddMetadata(const uint8_t* data, uint64_t length, uint64_t track_number,
+                   uint64_t timecode, uint64_t duration);
+
+  // Increments the size of the cluster's data in bytes.
+  void AddPayloadSize(uint64_t size);
+
+  // Closes the cluster so no more data can be written to it. Will update the
+  // cluster's size if |writer_| is seekable. Returns true on success. This
+  // variant of Finalize() fails when |write_last_frame_with_duration_| is set
+  // to true.
+  bool Finalize();
+
+  // Closes the cluster so no more data can be written to it. Will update the
+  // cluster's size if |writer_| is seekable. Returns true on success.
+  // Inputs:
+  //   set_last_frame_duration: Boolean indicating whether or not the duration
+  //                            of the last frame should be set. If set to
+  //                            false, the |duration| value is ignored and
+  //                            |write_last_frame_with_duration_| will not be
+  //                            honored.
+  //   duration: Duration of the Cluster in timecode scale.
+  bool Finalize(bool set_last_frame_duration, uint64_t duration);
+
+  // Returns the size in bytes for the entire Cluster element.
+  uint64_t Size() const;
+
+  // Given |abs_timecode|, calculates timecode relative to most recent timecode.
+  // Returns -1 on failure, or a relative timecode.
+  int64_t GetRelativeTimecode(int64_t abs_timecode) const;
+
+  int64_t size_position() const { return size_position_; }
+  int32_t blocks_added() const { return blocks_added_; }
+  uint64_t payload_size() const { return payload_size_; }
+  int64_t position_for_cues() const { return position_for_cues_; }
+  uint64_t timecode() const { return timecode_; }
+  uint64_t timecode_scale() const { return timecode_scale_; }
+  void set_write_last_frame_with_duration(bool write_last_frame_with_duration) {
+    write_last_frame_with_duration_ = write_last_frame_with_duration;
+  }
+  bool write_last_frame_with_duration() const {
+    return write_last_frame_with_duration_;
+  }
+
+ private:
+  // Iterator type for the |stored_frames_| map.
+  typedef std::map<uint64_t, std::list<Frame*> >::iterator FrameMapIterator;
+
+  // Utility method that confirms that blocks can still be added, and that the
+  // cluster header has been written. Used by |DoWriteFrame*|. Returns true
+  // when successful.
+  bool PreWriteBlock();
+
+  // Utility method used by the |DoWriteFrame*| methods that handles the book
+  // keeping required after each block is written.
+  void PostWriteBlock(uint64_t element_size);
+
+  // Does some verification and calls WriteFrame.
+  bool DoWriteFrame(const Frame* const frame);
+
+  // Either holds back the given frame, or writes it out depending on whether or
+  // not |write_last_frame_with_duration_| is set.
+  bool QueueOrWriteFrame(const Frame* const frame);
+
+  // Outputs the Cluster header to |writer_|. Returns true on success.
+  bool WriteClusterHeader();
+
+  // Number of blocks added to the cluster.
+  int32_t blocks_added_;
+
+  // Flag telling if the cluster has been closed.
+  bool finalized_;
+
+  // Flag indicating whether the cluster's timecode will always be written out
+  // using 8 bytes.
+  bool fixed_size_timecode_;
+
+  // Flag telling if the cluster's header has been written.
+  bool header_written_;
+
+  // The size of the cluster elements in bytes.
+  uint64_t payload_size_;
+
+  // The file position used for cue points.
+  const int64_t position_for_cues_;
+
+  // The file position of the cluster's size element.
+  int64_t size_position_;
+
+  // The absolute timecode of the cluster.
+  const uint64_t timecode_;
+
+  // The timecode scale of the Segment containing the cluster.
+  const uint64_t timecode_scale_;
+
+  // Flag indicating whether the last frame of the cluster should be written as
+  // a Block with Duration. If set to true, then it will result in holding back
+  // of frames and the parameterized version of Finalize() must be called to
+  // finish writing the Cluster.
+  bool write_last_frame_with_duration_;
+
+  // Map used to hold back frames, if required. Track number is the key.
+  std::map<uint64_t, std::list<Frame*> > stored_frames_;
+
+  // Map from track number to the timestamp of the last block written for that
+  // track.
+  std::map<uint64_t, uint64_t> last_block_timestamp_;
+
+  // Pointer to the writer object. Not owned by this class.
+  IMkvWriter* writer_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Cluster);
+};
+
+///////////////////////////////////////////////////////////////
+// SeekHead element
+class SeekHead {
+ public:
+  SeekHead();
+  ~SeekHead();
+
+  // TODO(fgalligan): Change this to reserve a certain size. Then check how
+  // big the seek entry to be added is as not every seek entry will be the
+  // maximum size it could be.
+  // Adds a seek entry to be written out when the element is finalized. |id|
+  // must be the coded mkv element id. |pos| is the file position of the
+  // element. Returns true on success.
+  bool AddSeekEntry(uint32_t id, uint64_t pos);
+
+  // Writes out SeekHead and SeekEntry elements. Returns true on success.
+  bool Finalize(IMkvWriter* writer) const;
+
+  // Returns the id of the Seek Entry at the given index. Returns -1 if index is
+  // out of range.
+  uint32_t GetId(int index) const;
+
+  // Returns the position of the Seek Entry at the given index. Returns -1 if
+  // index is out of range.
+  uint64_t GetPosition(int index) const;
+
+  // Sets the Seek Entry id and position at given index.
+  // Returns true on success.
+  bool SetSeekEntry(int index, uint32_t id, uint64_t position);
+
+  // Reserves space by writing out a Void element which will be updated with
+  // a SeekHead element later. Returns true on success.
+  bool Write(IMkvWriter* writer);
+
+  // We are going to put a cap on the number of Seek Entries.
+  const static int32_t kSeekEntryCount = 5;
+
+ private:
+  // Returns the maximum size in bytes of one seek entry.
+  uint64_t MaxEntrySize() const;
+
+  // Seek entry id element list.
+  uint32_t seek_entry_id_[kSeekEntryCount];
+
+  // Seek entry pos element list.
+  uint64_t seek_entry_pos_[kSeekEntryCount];
+
+  // The file position of SeekHead element.
+  int64_t start_pos_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(SeekHead);
+};
+
+///////////////////////////////////////////////////////////////
+// Segment Information element
+class SegmentInfo {
+ public:
+  SegmentInfo();
+  ~SegmentInfo();
+
+  // Will update the duration if |duration_| is > 0.0. Returns true on success.
+  bool Finalize(IMkvWriter* writer) const;
+
+  // Sets |muxing_app_| and |writing_app_|.
+  bool Init();
+
+  // Output the Segment Information element to the writer. Returns true on
+  // success.
+  bool Write(IMkvWriter* writer);
+
+  void set_duration(double duration) { duration_ = duration; }
+  double duration() const { return duration_; }
+  void set_muxing_app(const char* app);
+  const char* muxing_app() const { return muxing_app_; }
+  void set_timecode_scale(uint64_t scale) { timecode_scale_ = scale; }
+  uint64_t timecode_scale() const { return timecode_scale_; }
+  void set_writing_app(const char* app);
+  const char* writing_app() const { return writing_app_; }
+  void set_date_utc(int64_t date_utc) { date_utc_ = date_utc; }
+  int64_t date_utc() const { return date_utc_; }
+
+ private:
+  // Segment Information element names.
+  // Initially set to -1 to signify that a duration has not been set and should
+  // not be written out.
+  double duration_;
+  // Set to libwebm-%d.%d.%d.%d, major, minor, build, revision.
+  char* muxing_app_;
+  uint64_t timecode_scale_;
+  // Initially set to libwebm-%d.%d.%d.%d, major, minor, build, revision.
+  char* writing_app_;
+  // LLONG_MIN when DateUTC is not set.
+  int64_t date_utc_;
+
+  // The file position of the duration element.
+  int64_t duration_pos_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(SegmentInfo);
+};
+
+///////////////////////////////////////////////////////////////
+// This class represents the main segment in a WebM file. Currently only
+// supports one Segment element.
+//
+// Notes:
+//  |Init| must be called before any other method in this class.
+class Segment {
+ public:
+  enum Mode { kLive = 0x1, kFile = 0x2 };
+
+  enum CuesPosition {
+    kAfterClusters = 0x0,  // Position Cues after Clusters - Default
+    kBeforeClusters = 0x1  // Position Cues before Clusters
+  };
+
+  const static uint32_t kDefaultDocTypeVersion = 2;
+  const static uint64_t kDefaultMaxClusterDuration = 30000000000ULL;
+
+  Segment();
+  ~Segment();
+
+  // Initializes |SegmentInfo| and returns result. Always returns false when
+  // |ptr_writer| is NULL.
+  bool Init(IMkvWriter* ptr_writer);
+
+  // Adds a generic track to the segment.  Returns the newly-allocated
+  // track object (which is owned by the segment) on success, NULL on
+  // error. |number| is the number to use for the track.  |number|
+  // must be >= 0. If |number| == 0 then the muxer will decide on the
+  // track number.
+  Track* AddTrack(int32_t number);
+
+  // Adds a Vorbis audio track to the segment. Returns the number of the track
+  // on success, 0 on error. |number| is the number to use for the audio track.
+  // |number| must be >= 0. If |number| == 0 then the muxer will decide on
+  // the track number.
+  uint64_t AddAudioTrack(int32_t sample_rate, int32_t channels, int32_t number);
+
+  // Adds an empty chapter to the chapters of this segment.  Returns
+  // non-NULL on success.  After adding the chapter, the caller should
+  // populate its fields via the Chapter member functions.
+  Chapter* AddChapter();
+
+  // Adds an empty tag to the tags of this segment.  Returns
+  // non-NULL on success.  After adding the tag, the caller should
+  // populate its fields via the Tag member functions.
+  Tag* AddTag();
+
+  // Adds a cue point to the Cues element. |timestamp| is the time in
+  // nanoseconds of the cue's time. |track| is the Track of the Cue. This
+  // function must be called after AddFrame to calculate the correct
+  // BlockNumber for the CuePoint. Returns true on success.
+  bool AddCuePoint(uint64_t timestamp, uint64_t track);
+
+  // Adds a frame to be output in the file. Returns true on success.
+  // Inputs:
+  //   data: Pointer to the data
+  //   length: Length of the data
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.
+  //   timestamp:    Timestamp of the frame in nanoseconds from 0.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrame(const uint8_t* data, uint64_t length, uint64_t track_number,
+                uint64_t timestamp_ns, bool is_key);
+
+  // Writes a frame of metadata to the output medium; returns true on
+  // success.
+  // Inputs:
+  //   data: Pointer to the data
+  //   length: Length of the data
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.
+  //   timecode:     Absolute timestamp of the metadata frame, expressed
+  //                 in nanosecond units.
+  //   duration:     Duration of metadata frame, in nanosecond units.
+  //
+  // The metadata frame is written as a block group, with a duration
+  // sub-element but no reference time sub-elements (indicating that
+  // it is considered a keyframe, per Matroska semantics).
+  bool AddMetadata(const uint8_t* data, uint64_t length, uint64_t track_number,
+                   uint64_t timestamp_ns, uint64_t duration_ns);
+
+  // Writes a frame with additional data to the output medium; returns true on
+  // success.
+  // Inputs:
+  //   data: Pointer to the data.
+  //   length: Length of the data.
+  //   additional: Pointer to additional data.
+  //   additional_length: Length of additional data.
+  //   add_id: Additional ID which identifies the type of additional data.
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.
+  //   timestamp:    Absolute timestamp of the frame, expressed in nanosecond
+  //                 units.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrameWithAdditional(const uint8_t* data, uint64_t length,
+                              const uint8_t* additional,
+                              uint64_t additional_length, uint64_t add_id,
+                              uint64_t track_number, uint64_t timestamp,
+                              bool is_key);
+
+  // Writes a frame with DiscardPadding to the output medium; returns true on
+  // success.
+  // Inputs:
+  //   data: Pointer to the data.
+  //   length: Length of the data.
+  //   discard_padding: DiscardPadding element value.
+  //   track_number: Track to add the data to. Value returned by Add track
+  //                 functions.
+  //   timestamp:    Absolute timestamp of the frame, expressed in nanosecond
+  //                 units.
+  //   is_key:       Flag telling whether or not this frame is a key frame.
+  bool AddFrameWithDiscardPadding(const uint8_t* data, uint64_t length,
+                                  int64_t discard_padding,
+                                  uint64_t track_number, uint64_t timestamp,
+                                  bool is_key);
+
+  // Writes a Frame to the output medium. Chooses the correct way of writing
+  // the frame (Block vs SimpleBlock) based on the parameters passed.
+  // Inputs:
+  //   frame: frame object
+  bool AddGenericFrame(const Frame* frame);
+
+  // Adds a VP8 video track to the segment. Returns the number of the track on
+  // success, 0 on error. |number| is the number to use for the video track.
+  // |number| must be >= 0. If |number| == 0 then the muxer will decide on
+  // the track number.
+  uint64_t AddVideoTrack(int32_t width, int32_t height, int32_t number);
+
+  // This function must be called after Finalize() if you need a copy of the
+  // output with Cues written before the Clusters. It will return false if the
+  // writer is not seekable of if chunking is set to true.
+  // Input parameters:
+  // reader - an IMkvReader object created with the same underlying file of the
+  //          current writer object. Make sure to close the existing writer
+  //          object before creating this so that all the data is properly
+  //          flushed and available for reading.
+  // writer - an IMkvWriter object pointing to a *different* file than the one
+  //          pointed by the current writer object. This file will contain the
+  //          Cues element before the Clusters.
+  bool CopyAndMoveCuesBeforeClusters(mkvparser::IMkvReader* reader,
+                                     IMkvWriter* writer);
+
+  // Sets which track to use for the Cues element. Must have added the track
+  // before calling this function. Returns true on success. |track_number| is
+  // returned by the Add track functions.
+  bool CuesTrack(uint64_t track_number);
+
+  // This will force the muxer to create a new Cluster when the next frame is
+  // added.
+  void ForceNewClusterOnNextFrame();
+
+  // Writes out any frames that have not been written out. Finalizes the last
+  // cluster. May update the size and duration of the segment. May output the
+  // Cues element. May finalize the SeekHead element. Returns true on success.
+  bool Finalize();
+
+  // Returns the Cues object.
+  Cues* GetCues() { return &cues_; }
+
+  // Returns the Segment Information object.
+  const SegmentInfo* GetSegmentInfo() const { return &segment_info_; }
+  SegmentInfo* GetSegmentInfo() { return &segment_info_; }
+
+  // Search the Tracks and return the track that matches |track_number|.
+  // Returns NULL if there is no track match.
+  Track* GetTrackByNumber(uint64_t track_number) const;
+
+  // Toggles whether to output a cues element.
+  void OutputCues(bool output_cues);
+
+  // Toggles whether to write the last frame in each Cluster with Duration.
+  void AccurateClusterDuration(bool accurate_cluster_duration);
+
+  // Toggles whether to write the Cluster Timecode using exactly 8 bytes.
+  void UseFixedSizeClusterTimecode(bool fixed_size_cluster_timecode);
+
+  // Sets if the muxer will output files in chunks or not. |chunking| is a
+  // flag telling whether or not to turn on chunking. |filename| is the base
+  // filename for the chunk files. The header chunk file will be named
+  // |filename|.hdr and the data chunks will be named
+  // |filename|_XXXXXX.chk. Chunking implies that the muxer will be writing
+  // to files so the muxer will use the default MkvWriter class to control
+  // what data is written to what files. Returns true on success.
+  // TODO: Should we change the IMkvWriter Interface to add Open and Close?
+  // That will force the interface to be dependent on files.
+  bool SetChunking(bool chunking, const char* filename);
+
+  bool chunking() const { return chunking_; }
+  uint64_t cues_track() const { return cues_track_; }
+  void set_max_cluster_duration(uint64_t max_cluster_duration) {
+    max_cluster_duration_ = max_cluster_duration;
+  }
+  uint64_t max_cluster_duration() const { return max_cluster_duration_; }
+  void set_max_cluster_size(uint64_t max_cluster_size) {
+    max_cluster_size_ = max_cluster_size;
+  }
+  uint64_t max_cluster_size() const { return max_cluster_size_; }
+  void set_mode(Mode mode) { mode_ = mode; }
+  Mode mode() const { return mode_; }
+  CuesPosition cues_position() const { return cues_position_; }
+  bool output_cues() const { return output_cues_; }
+  const SegmentInfo* segment_info() const { return &segment_info_; }
+
+ private:
+  // Checks if header information has been output and initialized. If not it
+  // will output the Segment element and initialize the SeekHead elment and
+  // Cues elements.
+  bool CheckHeaderInfo();
+
+  // Sets |doc_type_version_| based on the current element requirements.
+  void UpdateDocTypeVersion();
+
+  // Sets |name| according to how many chunks have been written. |ext| is the
+  // file extension. |name| must be deleted by the calling app. Returns true
+  // on success.
+  bool UpdateChunkName(const char* ext, char** name) const;
+
+  // Returns the maximum offset within the segment's payload. When chunking
+  // this function is needed to determine offsets of elements within the
+  // chunked files. Returns -1 on error.
+  int64_t MaxOffset();
+
+  // Adds the frame to our frame array.
+  bool QueueFrame(Frame* frame);
+
+  // Output all frames that are queued. Returns -1 on error, otherwise
+  // it returns the number of frames written.
+  int WriteFramesAll();
+
+  // Output all frames that are queued that have an end time that is less
+  // then |timestamp|. Returns true on success and if there are no frames
+  // queued.
+  bool WriteFramesLessThan(uint64_t timestamp);
+
+  // Outputs the segment header, Segment Information element, SeekHead element,
+  // and Tracks element to |writer_|.
+  bool WriteSegmentHeader();
+
+  // Given a frame with the specified timestamp (nanosecond units) and
+  // keyframe status, determine whether a new cluster should be
+  // created, before writing enqueued frames and the frame itself. The
+  // function returns one of the following values:
+  //  -1 = error: an out-of-order frame was detected
+  //  0 = do not create a new cluster, and write frame to the existing cluster
+  //  1 = create a new cluster, and write frame to that new cluster
+  //  2 = create a new cluster, and re-run test
+  int TestFrame(uint64_t track_num, uint64_t timestamp_ns, bool key) const;
+
+  // Create a new cluster, using the earlier of the first enqueued
+  // frame, or the indicated time. Returns true on success.
+  bool MakeNewCluster(uint64_t timestamp_ns);
+
+  // Checks whether a new cluster needs to be created, and if so
+  // creates a new cluster. Returns false if creation of a new cluster
+  // was necessary but creation was not successful.
+  bool DoNewClusterProcessing(uint64_t track_num, uint64_t timestamp_ns,
+                              bool key);
+
+  // Adjusts Cue Point values (to place Cues before Clusters) so that they
+  // reflect the correct offsets.
+  void MoveCuesBeforeClusters();
+
+  // This function recursively computes the correct cluster offsets (this is
+  // done to move the Cues before Clusters). It recursively updates the change
+  // in size (which indicates a change in cluster offset) until no sizes change.
+  // Parameters:
+  // diff - indicates the difference in size of the Cues element that needs to
+  //        accounted for.
+  // index - index in the list of Cues which is currently being adjusted.
+  // cue_size - sum of size of all the CuePoint elements.
+  void MoveCuesBeforeClustersHelper(uint64_t diff, int index,
+                                    uint64_t* cue_size);
+
+  // Seeds the random number generator used to make UIDs.
+  unsigned int seed_;
+
+  // WebM elements
+  Cues cues_;
+  SeekHead seek_head_;
+  SegmentInfo segment_info_;
+  Tracks tracks_;
+  Chapters chapters_;
+  Tags tags_;
+
+  // Number of chunks written.
+  int chunk_count_;
+
+  // Current chunk filename.
+  char* chunk_name_;
+
+  // Default MkvWriter object created by this class used for writing clusters
+  // out in separate files.
+  MkvWriter* chunk_writer_cluster_;
+
+  // Default MkvWriter object created by this class used for writing Cues
+  // element out to a file.
+  MkvWriter* chunk_writer_cues_;
+
+  // Default MkvWriter object created by this class used for writing the
+  // Matroska header out to a file.
+  MkvWriter* chunk_writer_header_;
+
+  // Flag telling whether or not the muxer is chunking output to multiple
+  // files.
+  bool chunking_;
+
+  // Base filename for the chunked files.
+  char* chunking_base_name_;
+
+  // File position offset where the Clusters end.
+  int64_t cluster_end_offset_;
+
+  // List of clusters.
+  Cluster** cluster_list_;
+
+  // Number of cluster pointers allocated in the cluster list.
+  int32_t cluster_list_capacity_;
+
+  // Number of clusters in the cluster list.
+  int32_t cluster_list_size_;
+
+  // Indicates whether Cues should be written before or after Clusters
+  CuesPosition cues_position_;
+
+  // Track number that is associated with the cues element for this segment.
+  uint64_t cues_track_;
+
+  // Tells the muxer to force a new cluster on the next Block.
+  bool force_new_cluster_;
+
+  // List of stored audio frames. These variables are used to store frames so
+  // the muxer can follow the guideline "Audio blocks that contain the video
+  // key frame's timecode should be in the same cluster as the video key frame
+  // block."
+  Frame** frames_;
+
+  // Number of frame pointers allocated in the frame list.
+  int32_t frames_capacity_;
+
+  // Number of frames in the frame list.
+  int32_t frames_size_;
+
+  // Flag telling if a video track has been added to the segment.
+  bool has_video_;
+
+  // Flag telling if the segment's header has been written.
+  bool header_written_;
+
+  // Duration of the last block in nanoseconds.
+  uint64_t last_block_duration_;
+
+  // Last timestamp in nanoseconds added to a cluster.
+  uint64_t last_timestamp_;
+
+  // Last timestamp in nanoseconds by track number added to a cluster.
+  uint64_t last_track_timestamp_[kMaxTrackNumber];
+
+  // Maximum time in nanoseconds for a cluster duration. This variable is a
+  // guideline and some clusters may have a longer duration. Default is 30
+  // seconds.
+  uint64_t max_cluster_duration_;
+
+  // Maximum size in bytes for a cluster. This variable is a guideline and
+  // some clusters may have a larger size. Default is 0 which signifies that
+  // the muxer will decide the size.
+  uint64_t max_cluster_size_;
+
+  // The mode that segment is in. If set to |kLive| the writer must not
+  // seek backwards.
+  Mode mode_;
+
+  // Flag telling the muxer that a new cue point should be added.
+  bool new_cuepoint_;
+
+  // TODO(fgalligan): Should we add support for more than one Cues element?
+  // Flag whether or not the muxer should output a Cues element.
+  bool output_cues_;
+
+  // Flag whether or not the last frame in each Cluster will have a Duration
+  // element in it.
+  bool accurate_cluster_duration_;
+
+  // Flag whether or not to write the Cluster Timecode using exactly 8 bytes.
+  bool fixed_size_cluster_timecode_;
+
+  // The size of the EBML header, used to validate the header if
+  // WriteEbmlHeader() is called more than once.
+  int32_t ebml_header_size_;
+
+  // The file position of the segment's payload.
+  int64_t payload_pos_;
+
+  // The file position of the element's size.
+  int64_t size_position_;
+
+  // Current DocTypeVersion (|doc_type_version_|) and that written in
+  // WriteSegmentHeader().
+  // WriteEbmlHeader() will be called from Finalize() if |doc_type_version_|
+  // differs from |doc_type_version_written_|.
+  uint32_t doc_type_version_;
+  uint32_t doc_type_version_written_;
+
+  // Pointer to the writer objects. Not owned by this class.
+  IMkvWriter* writer_cluster_;
+  IMkvWriter* writer_cues_;
+  IMkvWriter* writer_header_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(Segment);
+};
+
+}  // namespace mkvmuxer
+
+#endif  // MKVMUXER_MKVMUXER_H_
diff --git a/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxertypes.h b/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxertypes.h
new file mode 100644
index 0000000000..e5db121605
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxertypes.h
@@ -0,0 +1,28 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#ifndef MKVMUXER_MKVMUXERTYPES_H_
+#define MKVMUXER_MKVMUXERTYPES_H_
+
+namespace mkvmuxer {
+typedef unsigned char uint8;
+typedef short int16;
+typedef int int32;
+typedef unsigned int uint32;
+typedef long long int64;
+typedef unsigned long long uint64;
+}  // namespace mkvmuxer
+
+// Copied from Chromium basictypes.h
+// A macro to disallow the copy constructor and operator= functions
+// This should be used in the private: declarations for a class
+#define LIBWEBM_DISALLOW_COPY_AND_ASSIGN(TypeName) \
+  TypeName(const TypeName&);                       \
+  void operator=(const TypeName&)
+
+#endif  // MKVMUXER_MKVMUXERTYPES_HPP_
diff --git a/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxerutil.cc b/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxerutil.cc
new file mode 100644
index 0000000000..3562b8ab82
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxerutil.cc
@@ -0,0 +1,650 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#include "mkvmuxer/mkvmuxerutil.h"
+
+#ifdef __ANDROID__
+#include <fcntl.h>
+#endif
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <ctime>
+#include <new>
+
+#include "common/webmids.h"
+#include "mkvmuxer/mkvmuxer.h"
+#include "mkvmuxer/mkvwriter.h"
+
+namespace mkvmuxer {
+
+namespace {
+
+// Date elements are always 8 octets in size.
+const int kDateElementSize = 8;
+
+uint64_t WriteBlock(IMkvWriter* writer, const Frame* const frame,
+                    int64_t timecode, uint64_t timecode_scale) {
+  uint64_t block_additional_elem_size = 0;
+  uint64_t block_addid_elem_size = 0;
+  uint64_t block_more_payload_size = 0;
+  uint64_t block_more_elem_size = 0;
+  uint64_t block_additions_payload_size = 0;
+  uint64_t block_additions_elem_size = 0;
+  if (frame->additional()) {
+    block_additional_elem_size =
+        EbmlElementSize(libwebm::kMkvBlockAdditional, frame->additional(),
+                        frame->additional_length());
+    block_addid_elem_size =
+        EbmlElementSize(libwebm::kMkvBlockAddID, frame->add_id());
+
+    block_more_payload_size =
+        block_addid_elem_size + block_additional_elem_size;
+    block_more_elem_size =
+        EbmlMasterElementSize(libwebm::kMkvBlockMore, block_more_payload_size) +
+        block_more_payload_size;
+    block_additions_payload_size = block_more_elem_size;
+    block_additions_elem_size =
+        EbmlMasterElementSize(libwebm::kMkvBlockAdditions,
+                              block_additions_payload_size) +
+        block_additions_payload_size;
+  }
+
+  uint64_t discard_padding_elem_size = 0;
+  if (frame->discard_padding() != 0) {
+    discard_padding_elem_size =
+        EbmlElementSize(libwebm::kMkvDiscardPadding, frame->discard_padding());
+  }
+
+  const uint64_t reference_block_timestamp =
+      frame->reference_block_timestamp() / timecode_scale;
+  uint64_t reference_block_elem_size = 0;
+  if (!frame->is_key()) {
+    reference_block_elem_size =
+        EbmlElementSize(libwebm::kMkvReferenceBlock, reference_block_timestamp);
+  }
+
+  const uint64_t duration = frame->duration() / timecode_scale;
+  uint64_t block_duration_elem_size = 0;
+  if (duration > 0)
+    block_duration_elem_size =
+        EbmlElementSize(libwebm::kMkvBlockDuration, duration);
+
+  const uint64_t block_payload_size = 4 + frame->length();
+  const uint64_t block_elem_size =
+      EbmlMasterElementSize(libwebm::kMkvBlock, block_payload_size) +
+      block_payload_size;
+
+  const uint64_t block_group_payload_size =
+      block_elem_size + block_additions_elem_size + block_duration_elem_size +
+      discard_padding_elem_size + reference_block_elem_size;
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvBlockGroup,
+                              block_group_payload_size)) {
+    return 0;
+  }
+
+  if (!WriteEbmlMasterElement(writer, libwebm::kMkvBlock, block_payload_size))
+    return 0;
+
+  if (WriteUInt(writer, frame->track_number()))
+    return 0;
+
+  if (SerializeInt(writer, timecode, 2))
+    return 0;
+
+  // For a Block, flags is always 0.
+  if (SerializeInt(writer, 0, 1))
+    return 0;
+
+  if (writer->Write(frame->frame(), static_cast<uint32_t>(frame->length())))
+    return 0;
+
+  if (frame->additional()) {
+    if (!WriteEbmlMasterElement(writer, libwebm::kMkvBlockAdditions,
+                                block_additions_payload_size)) {
+      return 0;
+    }
+
+    if (!WriteEbmlMasterElement(writer, libwebm::kMkvBlockMore,
+                                block_more_payload_size))
+      return 0;
+
+    if (!WriteEbmlElement(writer, libwebm::kMkvBlockAddID, frame->add_id()))
+      return 0;
+
+    if (!WriteEbmlElement(writer, libwebm::kMkvBlockAdditional,
+                          frame->additional(), frame->additional_length())) {
+      return 0;
+    }
+  }
+
+  if (frame->discard_padding() != 0 &&
+      !WriteEbmlElement(writer, libwebm::kMkvDiscardPadding,
+                        frame->discard_padding())) {
+    return false;
+  }
+
+  if (!frame->is_key() &&
+      !WriteEbmlElement(writer, libwebm::kMkvReferenceBlock,
+                        reference_block_timestamp)) {
+    return false;
+  }
+
+  if (duration > 0 &&
+      !WriteEbmlElement(writer, libwebm::kMkvBlockDuration, duration)) {
+    return false;
+  }
+  return EbmlMasterElementSize(libwebm::kMkvBlockGroup,
+                               block_group_payload_size) +
+         block_group_payload_size;
+}
+
+uint64_t WriteSimpleBlock(IMkvWriter* writer, const Frame* const frame,
+                          int64_t timecode) {
+  if (WriteID(writer, libwebm::kMkvSimpleBlock))
+    return 0;
+
+  const int32_t size = static_cast<int32_t>(frame->length()) + 4;
+  if (WriteUInt(writer, size))
+    return 0;
+
+  if (WriteUInt(writer, static_cast<uint64_t>(frame->track_number())))
+    return 0;
+
+  if (SerializeInt(writer, timecode, 2))
+    return 0;
+
+  uint64_t flags = 0;
+  if (frame->is_key())
+    flags |= 0x80;
+
+  if (SerializeInt(writer, flags, 1))
+    return 0;
+
+  if (writer->Write(frame->frame(), static_cast<uint32_t>(frame->length())))
+    return 0;
+
+  return static_cast<uint64_t>(GetUIntSize(libwebm::kMkvSimpleBlock) +
+                               GetCodedUIntSize(size) + 4 + frame->length());
+}
+
+}  // namespace
+
+int32_t GetCodedUIntSize(uint64_t value) {
+  if (value < 0x000000000000007FULL)
+    return 1;
+  else if (value < 0x0000000000003FFFULL)
+    return 2;
+  else if (value < 0x00000000001FFFFFULL)
+    return 3;
+  else if (value < 0x000000000FFFFFFFULL)
+    return 4;
+  else if (value < 0x00000007FFFFFFFFULL)
+    return 5;
+  else if (value < 0x000003FFFFFFFFFFULL)
+    return 6;
+  else if (value < 0x0001FFFFFFFFFFFFULL)
+    return 7;
+  return 8;
+}
+
+int32_t GetUIntSize(uint64_t value) {
+  if (value < 0x0000000000000100ULL)
+    return 1;
+  else if (value < 0x0000000000010000ULL)
+    return 2;
+  else if (value < 0x0000000001000000ULL)
+    return 3;
+  else if (value < 0x0000000100000000ULL)
+    return 4;
+  else if (value < 0x0000010000000000ULL)
+    return 5;
+  else if (value < 0x0001000000000000ULL)
+    return 6;
+  else if (value < 0x0100000000000000ULL)
+    return 7;
+  return 8;
+}
+
+int32_t GetIntSize(int64_t value) {
+  // Doubling the requested value ensures positive values with their high bit
+  // set are written with 0-padding to avoid flipping the signedness.
+  const uint64_t v = (value < 0) ? value ^ -1LL : value;
+  return GetUIntSize(2 * v);
+}
+
+uint64_t EbmlMasterElementSize(uint64_t type, uint64_t value) {
+  // Size of EBML ID
+  int32_t ebml_size = GetUIntSize(type);
+
+  // Datasize
+  ebml_size += GetCodedUIntSize(value);
+
+  return static_cast<uint64_t>(ebml_size);
+}
+
+uint64_t EbmlElementSize(uint64_t type, int64_t value) {
+  // Size of EBML ID
+  int32_t ebml_size = GetUIntSize(type);
+
+  // Datasize
+  ebml_size += GetIntSize(value);
+
+  // Size of Datasize
+  ebml_size++;
+
+  return static_cast<uint64_t>(ebml_size);
+}
+
+uint64_t EbmlElementSize(uint64_t type, uint64_t value) {
+  return EbmlElementSize(type, value, 0);
+}
+
+uint64_t EbmlElementSize(uint64_t type, uint64_t value, uint64_t fixed_size) {
+  // Size of EBML ID
+  uint64_t ebml_size = static_cast<uint64_t>(GetUIntSize(type));
+
+  // Datasize
+  ebml_size +=
+      (fixed_size > 0) ? fixed_size : static_cast<uint64_t>(GetUIntSize(value));
+
+  // Size of Datasize
+  ebml_size++;
+
+  return ebml_size;
+}
+
+uint64_t EbmlElementSize(uint64_t type, float /* value */) {
+  // Size of EBML ID
+  uint64_t ebml_size = static_cast<uint64_t>(GetUIntSize(type));
+
+  // Datasize
+  ebml_size += sizeof(float);
+
+  // Size of Datasize
+  ebml_size++;
+
+  return ebml_size;
+}
+
+uint64_t EbmlElementSize(uint64_t type, const char* value) {
+  if (!value)
+    return 0;
+
+  // Size of EBML ID
+  uint64_t ebml_size = static_cast<uint64_t>(GetUIntSize(type));
+
+  // Datasize
+  ebml_size += strlen(value);
+
+  // Size of Datasize
+  ebml_size++;
+
+  return ebml_size;
+}
+
+uint64_t EbmlElementSize(uint64_t type, const uint8_t* value, uint64_t size) {
+  if (!value)
+    return 0;
+
+  // Size of EBML ID
+  uint64_t ebml_size = static_cast<uint64_t>(GetUIntSize(type));
+
+  // Datasize
+  ebml_size += size;
+
+  // Size of Datasize
+  ebml_size += GetCodedUIntSize(size);
+
+  return ebml_size;
+}
+
+uint64_t EbmlDateElementSize(uint64_t type) {
+  // Size of EBML ID
+  uint64_t ebml_size = static_cast<uint64_t>(GetUIntSize(type));
+
+  // Datasize
+  ebml_size += kDateElementSize;
+
+  // Size of Datasize
+  ebml_size++;
+
+  return ebml_size;
+}
+
+int32_t SerializeInt(IMkvWriter* writer, int64_t value, int32_t size) {
+  if (!writer || size < 1 || size > 8)
+    return -1;
+
+  for (int32_t i = 1; i <= size; ++i) {
+    const int32_t byte_count = size - i;
+    const int32_t bit_count = byte_count * 8;
+
+    const int64_t bb = value >> bit_count;
+    const uint8_t b = static_cast<uint8_t>(bb);
+
+    const int32_t status = writer->Write(&b, 1);
+
+    if (status < 0)
+      return status;
+  }
+
+  return 0;
+}
+
+int32_t SerializeFloat(IMkvWriter* writer, float f) {
+  if (!writer)
+    return -1;
+
+  assert(sizeof(uint32_t) == sizeof(float));
+  // This union is merely used to avoid a reinterpret_cast from float& to
+  // uint32& which will result in violation of strict aliasing.
+  union U32 {
+    uint32_t u32;
+    float f;
+  } value;
+  value.f = f;
+
+  for (int32_t i = 1; i <= 4; ++i) {
+    const int32_t byte_count = 4 - i;
+    const int32_t bit_count = byte_count * 8;
+
+    const uint8_t byte = static_cast<uint8_t>(value.u32 >> bit_count);
+
+    const int32_t status = writer->Write(&byte, 1);
+
+    if (status < 0)
+      return status;
+  }
+
+  return 0;
+}
+
+int32_t WriteUInt(IMkvWriter* writer, uint64_t value) {
+  if (!writer)
+    return -1;
+
+  int32_t size = GetCodedUIntSize(value);
+
+  return WriteUIntSize(writer, value, size);
+}
+
+int32_t WriteUIntSize(IMkvWriter* writer, uint64_t value, int32_t size) {
+  if (!writer || size < 0 || size > 8)
+    return -1;
+
+  if (size > 0) {
+    const uint64_t bit = 1LL << (size * 7);
+
+    if (value > (bit - 2))
+      return -1;
+
+    value |= bit;
+  } else {
+    size = 1;
+    int64_t bit;
+
+    for (;;) {
+      bit = 1LL << (size * 7);
+      const uint64_t max = bit - 2;
+
+      if (value <= max)
+        break;
+
+      ++size;
+    }
+
+    if (size > 8)
+      return false;
+
+    value |= bit;
+  }
+
+  return SerializeInt(writer, value, size);
+}
+
+int32_t WriteID(IMkvWriter* writer, uint64_t type) {
+  if (!writer)
+    return -1;
+
+  writer->ElementStartNotify(type, writer->Position());
+
+  const int32_t size = GetUIntSize(type);
+
+  return SerializeInt(writer, type, size);
+}
+
+bool WriteEbmlMasterElement(IMkvWriter* writer, uint64_t type, uint64_t size) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  if (WriteUInt(writer, size))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, uint64_t value) {
+  return WriteEbmlElement(writer, type, value, 0);
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, uint64_t value,
+                      uint64_t fixed_size) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  uint64_t size = static_cast<uint64_t>(GetUIntSize(value));
+  if (fixed_size > 0) {
+    if (size > fixed_size)
+      return false;
+    size = fixed_size;
+  }
+  if (WriteUInt(writer, size))
+    return false;
+
+  if (SerializeInt(writer, value, static_cast<int32_t>(size)))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, int64_t value) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return 0;
+
+  const uint64_t size = GetIntSize(value);
+  if (WriteUInt(writer, size))
+    return false;
+
+  if (SerializeInt(writer, value, static_cast<int32_t>(size)))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, float value) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  if (WriteUInt(writer, 4))
+    return false;
+
+  if (SerializeFloat(writer, value))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, const char* value) {
+  if (!writer || !value)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  const uint64_t length = strlen(value);
+  if (WriteUInt(writer, length))
+    return false;
+
+  if (writer->Write(value, static_cast<const uint32_t>(length)))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, const uint8_t* value,
+                      uint64_t size) {
+  if (!writer || !value || size < 1)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  if (WriteUInt(writer, size))
+    return false;
+
+  if (writer->Write(value, static_cast<uint32_t>(size)))
+    return false;
+
+  return true;
+}
+
+bool WriteEbmlDateElement(IMkvWriter* writer, uint64_t type, int64_t value) {
+  if (!writer)
+    return false;
+
+  if (WriteID(writer, type))
+    return false;
+
+  if (WriteUInt(writer, kDateElementSize))
+    return false;
+
+  if (SerializeInt(writer, value, kDateElementSize))
+    return false;
+
+  return true;
+}
+
+uint64_t WriteFrame(IMkvWriter* writer, const Frame* const frame,
+                    Cluster* cluster) {
+  if (!writer || !frame || !frame->IsValid() || !cluster ||
+      !cluster->timecode_scale())
+    return 0;
+
+  //  Technically the timecode for a block can be less than the
+  //  timecode for the cluster itself (remember that block timecode
+  //  is a signed, 16-bit integer).  However, as a simplification we
+  //  only permit non-negative cluster-relative timecodes for blocks.
+  const int64_t relative_timecode = cluster->GetRelativeTimecode(
+      frame->timestamp() / cluster->timecode_scale());
+  if (relative_timecode < 0 || relative_timecode > kMaxBlockTimecode)
+    return 0;
+
+  return frame->CanBeSimpleBlock() ?
+             WriteSimpleBlock(writer, frame, relative_timecode) :
+             WriteBlock(writer, frame, relative_timecode,
+                        cluster->timecode_scale());
+}
+
+uint64_t WriteVoidElement(IMkvWriter* writer, uint64_t size) {
+  if (!writer)
+    return false;
+
+  // Subtract one for the void ID and the coded size.
+  uint64_t void_entry_size = size - 1 - GetCodedUIntSize(size - 1);
+  uint64_t void_size =
+      EbmlMasterElementSize(libwebm::kMkvVoid, void_entry_size) +
+      void_entry_size;
+
+  if (void_size != size)
+    return 0;
+
+  const int64_t payload_position = writer->Position();
+  if (payload_position < 0)
+    return 0;
+
+  if (WriteID(writer, libwebm::kMkvVoid))
+    return 0;
+
+  if (WriteUInt(writer, void_entry_size))
+    return 0;
+
+  const uint8_t value = 0;
+  for (int32_t i = 0; i < static_cast<int32_t>(void_entry_size); ++i) {
+    if (writer->Write(&value, 1))
+      return 0;
+  }
+
+  const int64_t stop_position = writer->Position();
+  if (stop_position < 0 ||
+      stop_position - payload_position != static_cast<int64_t>(void_size))
+    return 0;
+
+  return void_size;
+}
+
+void GetVersion(int32_t* major, int32_t* minor, int32_t* build,
+                int32_t* revision) {
+  *major = 0;
+  *minor = 2;
+  *build = 1;
+  *revision = 0;
+}
+
+uint64_t MakeUID(unsigned int* seed) {
+  uint64_t uid = 0;
+
+#ifdef __MINGW32__
+  srand(*seed);
+#endif
+
+  for (int i = 0; i < 7; ++i) {  // avoid problems with 8-byte values
+    uid <<= 8;
+
+// TODO(fgalligan): Move random number generation to platform specific code.
+#ifdef _MSC_VER
+    (void)seed;
+    const int32_t nn = rand();
+#elif __ANDROID__
+    int32_t temp_num = 1;
+    int fd = open("/dev/urandom", O_RDONLY);
+    if (fd != -1) {
+      read(fd, &temp_num, sizeof(temp_num));
+      close(fd);
+    }
+    const int32_t nn = temp_num;
+#elif defined __MINGW32__
+    const int32_t nn = rand();
+#else
+    const int32_t nn = rand_r(seed);
+#endif
+    const int32_t n = 0xFF & (nn >> 4);  // throw away low-order bits
+
+    uid |= n;
+  }
+
+  return uid;
+}
+
+}  // namespace mkvmuxer
diff --git a/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxerutil.h b/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxerutil.h
new file mode 100644
index 0000000000..0e21a2dcbe
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxerutil.h
@@ -0,0 +1,95 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef MKVMUXER_MKVMUXERUTIL_H_
+#define MKVMUXER_MKVMUXERUTIL_H_
+
+#include <stdint.h>
+
+namespace mkvmuxer {
+class Cluster;
+class Frame;
+class IMkvWriter;
+
+const uint64_t kEbmlUnknownValue = 0x01FFFFFFFFFFFFFFULL;
+const int64_t kMaxBlockTimecode = 0x07FFFLL;
+
+// Writes out |value| in Big Endian order. Returns 0 on success.
+int32_t SerializeInt(IMkvWriter* writer, int64_t value, int32_t size);
+
+// Returns the size in bytes of the element.
+int32_t GetUIntSize(uint64_t value);
+int32_t GetIntSize(int64_t value);
+int32_t GetCodedUIntSize(uint64_t value);
+uint64_t EbmlMasterElementSize(uint64_t type, uint64_t value);
+uint64_t EbmlElementSize(uint64_t type, int64_t value);
+uint64_t EbmlElementSize(uint64_t type, uint64_t value);
+uint64_t EbmlElementSize(uint64_t type, float value);
+uint64_t EbmlElementSize(uint64_t type, const char* value);
+uint64_t EbmlElementSize(uint64_t type, const uint8_t* value, uint64_t size);
+uint64_t EbmlDateElementSize(uint64_t type);
+
+// Returns the size in bytes of the element assuming that the element was
+// written using |fixed_size| bytes. If |fixed_size| is set to zero, then it
+// computes the necessary number of bytes based on |value|.
+uint64_t EbmlElementSize(uint64_t type, uint64_t value, uint64_t fixed_size);
+
+// Creates an EBML coded number from |value| and writes it out. The size of
+// the coded number is determined by the value of |value|. |value| must not
+// be in a coded form. Returns 0 on success.
+int32_t WriteUInt(IMkvWriter* writer, uint64_t value);
+
+// Creates an EBML coded number from |value| and writes it out. The size of
+// the coded number is determined by the value of |size|. |value| must not
+// be in a coded form. Returns 0 on success.
+int32_t WriteUIntSize(IMkvWriter* writer, uint64_t value, int32_t size);
+
+// Output an Mkv master element. Returns true if the element was written.
+bool WriteEbmlMasterElement(IMkvWriter* writer, uint64_t value, uint64_t size);
+
+// Outputs an Mkv ID, calls |IMkvWriter::ElementStartNotify|, and passes the
+// ID to |SerializeInt|. Returns 0 on success.
+int32_t WriteID(IMkvWriter* writer, uint64_t type);
+
+// Output an Mkv non-master element. Returns true if the element was written.
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, uint64_t value);
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, int64_t value);
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, float value);
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, const char* value);
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, const uint8_t* value,
+                      uint64_t size);
+bool WriteEbmlDateElement(IMkvWriter* writer, uint64_t type, int64_t value);
+
+// Output an Mkv non-master element using fixed size. The element will be
+// written out using exactly |fixed_size| bytes. If |fixed_size| is set to zero
+// then it computes the necessary number of bytes based on |value|. Returns true
+// if the element was written.
+bool WriteEbmlElement(IMkvWriter* writer, uint64_t type, uint64_t value,
+                      uint64_t fixed_size);
+
+// Output a Mkv Frame. It decides the correct element to write (Block vs
+// SimpleBlock) based on the parameters of the Frame.
+uint64_t WriteFrame(IMkvWriter* writer, const Frame* const frame,
+                    Cluster* cluster);
+
+// Output a void element. |size| must be the entire size in bytes that will be
+// void. The function will calculate the size of the void header and subtract
+// it from |size|.
+uint64_t WriteVoidElement(IMkvWriter* writer, uint64_t size);
+
+// Returns the version number of the muxer in |major|, |minor|, |build|,
+// and |revision|.
+void GetVersion(int32_t* major, int32_t* minor, int32_t* build,
+                int32_t* revision);
+
+// Returns a random number to be used for UID, using |seed| to seed
+// the random-number generator (see POSIX rand_r() for semantics).
+uint64_t MakeUID(unsigned int* seed);
+
+}  // namespace mkvmuxer
+
+#endif  // MKVMUXER_MKVMUXERUTIL_H_
diff --git a/third_party/aom/third_party/libwebm/mkvmuxer/mkvwriter.cc b/third_party/aom/third_party/libwebm/mkvmuxer/mkvwriter.cc
new file mode 100644
index 0000000000..ca48e149c6
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvmuxer/mkvwriter.cc
@@ -0,0 +1,88 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#include "mkvmuxer/mkvwriter.h"
+
+#ifdef _MSC_VER
+#include <share.h>  // for _SH_DENYWR
+#endif
+
+namespace mkvmuxer {
+
+MkvWriter::MkvWriter() : file_(NULL), writer_owns_file_(true) {}
+
+MkvWriter::MkvWriter(FILE* fp) : file_(fp), writer_owns_file_(false) {}
+
+MkvWriter::~MkvWriter() { Close(); }
+
+int32 MkvWriter::Write(const void* buffer, uint32 length) {
+  if (!file_)
+    return -1;
+
+  if (length == 0)
+    return 0;
+
+  if (buffer == NULL)
+    return -1;
+
+  const size_t bytes_written = fwrite(buffer, 1, length, file_);
+
+  return (bytes_written == length) ? 0 : -1;
+}
+
+bool MkvWriter::Open(const char* filename) {
+  if (filename == NULL)
+    return false;
+
+  if (file_)
+    return false;
+
+#ifdef _MSC_VER
+  file_ = _fsopen(filename, "wb", _SH_DENYWR);
+#else
+  file_ = fopen(filename, "wb");
+#endif
+  if (file_ == NULL)
+    return false;
+  return true;
+}
+
+void MkvWriter::Close() {
+  if (file_ && writer_owns_file_) {
+    fclose(file_);
+  }
+  file_ = NULL;
+}
+
+int64 MkvWriter::Position() const {
+  if (!file_)
+    return 0;
+
+#ifdef _MSC_VER
+  return _ftelli64(file_);
+#else
+  return ftell(file_);
+#endif
+}
+
+int32 MkvWriter::Position(int64 position) {
+  if (!file_)
+    return -1;
+
+#ifdef _MSC_VER
+  return _fseeki64(file_, position, SEEK_SET);
+#else
+  return fseek(file_, position, SEEK_SET);
+#endif
+}
+
+bool MkvWriter::Seekable() const { return true; }
+
+void MkvWriter::ElementStartNotify(uint64, int64) {}
+
+}  // namespace mkvmuxer
diff --git a/third_party/aom/third_party/libwebm/mkvmuxer/mkvwriter.h b/third_party/aom/third_party/libwebm/mkvmuxer/mkvwriter.h
new file mode 100644
index 0000000000..4227c63748
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvmuxer/mkvwriter.h
@@ -0,0 +1,51 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+
+#ifndef MKVMUXER_MKVWRITER_H_
+#define MKVMUXER_MKVWRITER_H_
+
+#include <stdio.h>
+
+#include "mkvmuxer/mkvmuxer.h"
+#include "mkvmuxer/mkvmuxertypes.h"
+
+namespace mkvmuxer {
+
+// Default implementation of the IMkvWriter interface on Windows.
+class MkvWriter : public IMkvWriter {
+ public:
+  MkvWriter();
+  explicit MkvWriter(FILE* fp);
+  virtual ~MkvWriter();
+
+  // IMkvWriter interface
+  virtual int64 Position() const;
+  virtual int32 Position(int64 position);
+  virtual bool Seekable() const;
+  virtual int32 Write(const void* buffer, uint32 length);
+  virtual void ElementStartNotify(uint64 element_id, int64 position);
+
+  // Creates and opens a file for writing. |filename| is the name of the file
+  // to open. This function will overwrite the contents of |filename|. Returns
+  // true on success.
+  bool Open(const char* filename);
+
+  // Closes an opened file.
+  void Close();
+
+ private:
+  // File handle to output file.
+  FILE* file_;
+  bool writer_owns_file_;
+
+  LIBWEBM_DISALLOW_COPY_AND_ASSIGN(MkvWriter);
+};
+
+}  // namespace mkvmuxer
+
+#endif  // MKVMUXER_MKVWRITER_H_
diff --git a/third_party/aom/third_party/libwebm/mkvparser/mkvparser.cc b/third_party/aom/third_party/libwebm/mkvparser/mkvparser.cc
new file mode 100644
index 0000000000..21801154d9
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvparser/mkvparser.cc
@@ -0,0 +1,7940 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#include "mkvparser/mkvparser.h"
+
+#if defined(_MSC_VER) && _MSC_VER < 1800
+#include <float.h>  // _isnan() / _finite()
+#define MSC_COMPAT
+#endif
+
+#include <cassert>
+#include <cfloat>
+#include <climits>
+#include <cmath>
+#include <cstring>
+#include <memory>
+#include <new>
+
+#include "common/webmids.h"
+
+namespace mkvparser {
+const float MasteringMetadata::kValueNotPresent = FLT_MAX;
+const long long Colour::kValueNotPresent = LLONG_MAX;
+
+#ifdef MSC_COMPAT
+inline bool isnan(double val) { return !!_isnan(val); }
+inline bool isinf(double val) { return !_finite(val); }
+#else
+inline bool isnan(double val) { return std::isnan(val); }
+inline bool isinf(double val) { return std::isinf(val); }
+#endif  // MSC_COMPAT
+
+IMkvReader::~IMkvReader() {}
+
+template <typename Type>
+Type* SafeArrayAlloc(unsigned long long num_elements,
+                     unsigned long long element_size) {
+  if (num_elements == 0 || element_size == 0)
+    return NULL;
+
+  const size_t kMaxAllocSize = 0x80000000;  // 2GiB
+  const unsigned long long num_bytes = num_elements * element_size;
+  if (element_size > (kMaxAllocSize / num_elements))
+    return NULL;
+  if (num_bytes != static_cast<size_t>(num_bytes))
+    return NULL;
+
+  return new (std::nothrow) Type[static_cast<size_t>(num_bytes)];
+}
+
+void GetVersion(int& major, int& minor, int& build, int& revision) {
+  major = 1;
+  minor = 0;
+  build = 0;
+  revision = 30;
+}
+
+long long ReadUInt(IMkvReader* pReader, long long pos, long& len) {
+  if (!pReader || pos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  len = 1;
+  unsigned char b;
+  int status = pReader->Read(pos, 1, &b);
+
+  if (status < 0)  // error or underflow
+    return status;
+
+  if (status > 0)  // interpreted as "underflow"
+    return E_BUFFER_NOT_FULL;
+
+  if (b == 0)  // we can't handle u-int values larger than 8 bytes
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char m = 0x80;
+
+  while (!(b & m)) {
+    m >>= 1;
+    ++len;
+  }
+
+  long long result = b & (~m);
+  ++pos;
+
+  for (int i = 1; i < len; ++i) {
+    status = pReader->Read(pos, 1, &b);
+
+    if (status < 0) {
+      len = 1;
+      return status;
+    }
+
+    if (status > 0) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result <<= 8;
+    result |= b;
+
+    ++pos;
+  }
+
+  return result;
+}
+
+// Reads an EBML ID and returns it.
+// An ID must at least 1 byte long, cannot exceed 4, and its value must be
+// greater than 0.
+// See known EBML values and EBMLMaxIDLength:
+// http://www.matroska.org/technical/specs/index.html
+// Returns the ID, or a value less than 0 to report an error while reading the
+// ID.
+long long ReadID(IMkvReader* pReader, long long pos, long& len) {
+  if (pReader == NULL || pos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  // Read the first byte. The length in bytes of the ID is determined by
+  // finding the first set bit in the first byte of the ID.
+  unsigned char temp_byte = 0;
+  int read_status = pReader->Read(pos, 1, &temp_byte);
+
+  if (read_status < 0)
+    return E_FILE_FORMAT_INVALID;
+  else if (read_status > 0)  // No data to read.
+    return E_BUFFER_NOT_FULL;
+
+  if (temp_byte == 0)  // ID length > 8 bytes; invalid file.
+    return E_FILE_FORMAT_INVALID;
+
+  int bit_pos = 0;
+  const int kMaxIdLengthInBytes = 4;
+  const int kCheckByte = 0x80;
+
+  // Find the first bit that's set.
+  bool found_bit = false;
+  for (; bit_pos < kMaxIdLengthInBytes; ++bit_pos) {
+    if ((kCheckByte >> bit_pos) & temp_byte) {
+      found_bit = true;
+      break;
+    }
+  }
+
+  if (!found_bit) {
+    // The value is too large to be a valid ID.
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  // Read the remaining bytes of the ID (if any).
+  const int id_length = bit_pos + 1;
+  long long ebml_id = temp_byte;
+  for (int i = 1; i < id_length; ++i) {
+    ebml_id <<= 8;
+    read_status = pReader->Read(pos + i, 1, &temp_byte);
+
+    if (read_status < 0)
+      return E_FILE_FORMAT_INVALID;
+    else if (read_status > 0)
+      return E_BUFFER_NOT_FULL;
+
+    ebml_id |= temp_byte;
+  }
+
+  len = id_length;
+  return ebml_id;
+}
+
+long long GetUIntLength(IMkvReader* pReader, long long pos, long& len) {
+  if (!pReader || pos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  long long total, available;
+
+  int status = pReader->Length(&total, &available);
+  if (status < 0 || (total >= 0 && available > total))
+    return E_FILE_FORMAT_INVALID;
+
+  len = 1;
+
+  if (pos >= available)
+    return pos;  // too few bytes available
+
+  unsigned char b;
+
+  status = pReader->Read(pos, 1, &b);
+
+  if (status != 0)
+    return status;
+
+  if (b == 0)  // we can't handle u-int values larger than 8 bytes
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char m = 0x80;
+
+  while (!(b & m)) {
+    m >>= 1;
+    ++len;
+  }
+
+  return 0;  // success
+}
+
+// TODO(vigneshv): This function assumes that unsigned values never have their
+// high bit set.
+long long UnserializeUInt(IMkvReader* pReader, long long pos, long long size) {
+  if (!pReader || pos < 0 || (size <= 0) || (size > 8))
+    return E_FILE_FORMAT_INVALID;
+
+  long long result = 0;
+
+  for (long long i = 0; i < size; ++i) {
+    unsigned char b;
+
+    const long status = pReader->Read(pos, 1, &b);
+
+    if (status < 0)
+      return status;
+
+    result <<= 8;
+    result |= b;
+
+    ++pos;
+  }
+
+  return result;
+}
+
+long UnserializeFloat(IMkvReader* pReader, long long pos, long long size_,
+                      double& result) {
+  if (!pReader || pos < 0 || ((size_ != 4) && (size_ != 8)))
+    return E_FILE_FORMAT_INVALID;
+
+  const long size = static_cast<long>(size_);
+
+  unsigned char buf[8];
+
+  const int status = pReader->Read(pos, size, buf);
+
+  if (status < 0)  // error
+    return status;
+
+  if (size == 4) {
+    union {
+      float f;
+      unsigned long ff;
+    };
+
+    ff = 0;
+
+    for (int i = 0;;) {
+      ff |= buf[i];
+
+      if (++i >= 4)
+        break;
+
+      ff <<= 8;
+    }
+
+    result = f;
+  } else {
+    union {
+      double d;
+      unsigned long long dd;
+    };
+
+    dd = 0;
+
+    for (int i = 0;;) {
+      dd |= buf[i];
+
+      if (++i >= 8)
+        break;
+
+      dd <<= 8;
+    }
+
+    result = d;
+  }
+
+  if (mkvparser::isinf(result) || mkvparser::isnan(result))
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+long UnserializeInt(IMkvReader* pReader, long long pos, long long size,
+                    long long& result_ref) {
+  if (!pReader || pos < 0 || size < 1 || size > 8)
+    return E_FILE_FORMAT_INVALID;
+
+  signed char first_byte = 0;
+  const long status = pReader->Read(pos, 1, (unsigned char*)&first_byte);
+
+  if (status < 0)
+    return status;
+
+  unsigned long long result = first_byte;
+  ++pos;
+
+  for (long i = 1; i < size; ++i) {
+    unsigned char b;
+
+    const long status = pReader->Read(pos, 1, &b);
+
+    if (status < 0)
+      return status;
+
+    result <<= 8;
+    result |= b;
+
+    ++pos;
+  }
+
+  result_ref = static_cast<long long>(result);
+  return 0;
+}
+
+long UnserializeString(IMkvReader* pReader, long long pos, long long size,
+                       char*& str) {
+  delete[] str;
+  str = NULL;
+
+  if (size >= LONG_MAX || size < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  // +1 for '\0' terminator
+  const long required_size = static_cast<long>(size) + 1;
+
+  str = SafeArrayAlloc<char>(1, required_size);
+  if (str == NULL)
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char* const buf = reinterpret_cast<unsigned char*>(str);
+
+  const long status = pReader->Read(pos, static_cast<long>(size), buf);
+
+  if (status) {
+    delete[] str;
+    str = NULL;
+
+    return status;
+  }
+
+  str[required_size - 1] = '\0';
+  return 0;
+}
+
+long ParseElementHeader(IMkvReader* pReader, long long& pos, long long stop,
+                        long long& id, long long& size) {
+  if (stop >= 0 && pos >= stop)
+    return E_FILE_FORMAT_INVALID;
+
+  long len;
+
+  id = ReadID(pReader, pos, len);
+
+  if (id < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume id
+
+  if (stop >= 0 && pos >= stop)
+    return E_FILE_FORMAT_INVALID;
+
+  size = ReadUInt(pReader, pos, len);
+
+  if (size < 0 || len < 1 || len > 8) {
+    // Invalid: Negative payload size, negative or 0 length integer, or integer
+    // larger than 64 bits (libwebm cannot handle them).
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  // Avoid rolling over pos when very close to LLONG_MAX.
+  const unsigned long long rollover_check =
+      static_cast<unsigned long long>(pos) + len;
+  if (rollover_check > LLONG_MAX)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume length of size
+
+  // pos now designates payload
+
+  if (stop >= 0 && pos > stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;  // success
+}
+
+bool Match(IMkvReader* pReader, long long& pos, unsigned long expected_id,
+           long long& val) {
+  if (!pReader || pos < 0)
+    return false;
+
+  long long total = 0;
+  long long available = 0;
+
+  const long status = pReader->Length(&total, &available);
+  if (status < 0 || (total >= 0 && available > total))
+    return false;
+
+  long len = 0;
+
+  const long long id = ReadID(pReader, pos, len);
+  if (id < 0 || (available - pos) > len)
+    return false;
+
+  if (static_cast<unsigned long>(id) != expected_id)
+    return false;
+
+  pos += len;  // consume id
+
+  const long long size = ReadUInt(pReader, pos, len);
+  if (size < 0 || size > 8 || len < 1 || len > 8 || (available - pos) > len)
+    return false;
+
+  pos += len;  // consume length of size of payload
+
+  val = UnserializeUInt(pReader, pos, size);
+  if (val < 0)
+    return false;
+
+  pos += size;  // consume size of payload
+
+  return true;
+}
+
+bool Match(IMkvReader* pReader, long long& pos, unsigned long expected_id,
+           unsigned char*& buf, size_t& buflen) {
+  if (!pReader || pos < 0)
+    return false;
+
+  long long total = 0;
+  long long available = 0;
+
+  long status = pReader->Length(&total, &available);
+  if (status < 0 || (total >= 0 && available > total))
+    return false;
+
+  long len = 0;
+  const long long id = ReadID(pReader, pos, len);
+  if (id < 0 || (available - pos) > len)
+    return false;
+
+  if (static_cast<unsigned long>(id) != expected_id)
+    return false;
+
+  pos += len;  // consume id
+
+  const long long size = ReadUInt(pReader, pos, len);
+  if (size < 0 || len <= 0 || len > 8 || (available - pos) > len)
+    return false;
+
+  unsigned long long rollover_check =
+      static_cast<unsigned long long>(pos) + len;
+  if (rollover_check > LLONG_MAX)
+    return false;
+
+  pos += len;  // consume length of size of payload
+
+  rollover_check = static_cast<unsigned long long>(pos) + size;
+  if (rollover_check > LLONG_MAX)
+    return false;
+
+  if ((pos + size) > available)
+    return false;
+
+  if (size >= LONG_MAX)
+    return false;
+
+  const long buflen_ = static_cast<long>(size);
+
+  buf = SafeArrayAlloc<unsigned char>(1, buflen_);
+  if (!buf)
+    return false;
+
+  status = pReader->Read(pos, buflen_, buf);
+  if (status != 0)
+    return false;
+
+  buflen = buflen_;
+
+  pos += size;  // consume size of payload
+  return true;
+}
+
+EBMLHeader::EBMLHeader() : m_docType(NULL) { Init(); }
+
+EBMLHeader::~EBMLHeader() { delete[] m_docType; }
+
+void EBMLHeader::Init() {
+  m_version = 1;
+  m_readVersion = 1;
+  m_maxIdLength = 4;
+  m_maxSizeLength = 8;
+
+  if (m_docType) {
+    delete[] m_docType;
+    m_docType = NULL;
+  }
+
+  m_docTypeVersion = 1;
+  m_docTypeReadVersion = 1;
+}
+
+long long EBMLHeader::Parse(IMkvReader* pReader, long long& pos) {
+  if (!pReader)
+    return E_FILE_FORMAT_INVALID;
+
+  long long total, available;
+
+  long status = pReader->Length(&total, &available);
+
+  if (status < 0)  // error
+    return status;
+
+  pos = 0;
+
+  // Scan until we find what looks like the first byte of the EBML header.
+  const long long kMaxScanBytes = (available >= 1024) ? 1024 : available;
+  const unsigned char kEbmlByte0 = 0x1A;
+  unsigned char scan_byte = 0;
+
+  while (pos < kMaxScanBytes) {
+    status = pReader->Read(pos, 1, &scan_byte);
+
+    if (status < 0)  // error
+      return status;
+    else if (status > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if (scan_byte == kEbmlByte0)
+      break;
+
+    ++pos;
+  }
+
+  long len = 0;
+  const long long ebml_id = ReadID(pReader, pos, len);
+
+  if (ebml_id == E_BUFFER_NOT_FULL)
+    return E_BUFFER_NOT_FULL;
+
+  if (len != 4 || ebml_id != libwebm::kMkvEBML)
+    return E_FILE_FORMAT_INVALID;
+
+  // Move read pos forward to the EBML header size field.
+  pos += 4;
+
+  // Read length of size field.
+  long long result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error
+    return E_FILE_FORMAT_INVALID;
+  else if (result > 0)  // need more data
+    return E_BUFFER_NOT_FULL;
+
+  if (len < 1 || len > 8)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((total >= 0) && ((total - pos) < len))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((available - pos) < len)
+    return pos + len;  // try again later
+
+  // Read the EBML header size.
+  result = ReadUInt(pReader, pos, len);
+
+  if (result < 0)  // error
+    return result;
+
+  pos += len;  // consume size field
+
+  // pos now designates start of payload
+
+  if ((total >= 0) && ((total - pos) < result))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((available - pos) < result)
+    return pos + result;
+
+  const long long end = pos + result;
+
+  Init();
+
+  while (pos < end) {
+    long long id, size;
+
+    status = ParseElementHeader(pReader, pos, end, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvEBMLVersion) {
+      m_version = UnserializeUInt(pReader, pos, size);
+
+      if (m_version <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvEBMLReadVersion) {
+      m_readVersion = UnserializeUInt(pReader, pos, size);
+
+      if (m_readVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvEBMLMaxIDLength) {
+      m_maxIdLength = UnserializeUInt(pReader, pos, size);
+
+      if (m_maxIdLength <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvEBMLMaxSizeLength) {
+      m_maxSizeLength = UnserializeUInt(pReader, pos, size);
+
+      if (m_maxSizeLength <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDocType) {
+      if (m_docType)
+        return E_FILE_FORMAT_INVALID;
+
+      status = UnserializeString(pReader, pos, size, m_docType);
+
+      if (status)  // error
+        return status;
+    } else if (id == libwebm::kMkvDocTypeVersion) {
+      m_docTypeVersion = UnserializeUInt(pReader, pos, size);
+
+      if (m_docTypeVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDocTypeReadVersion) {
+      m_docTypeReadVersion = UnserializeUInt(pReader, pos, size);
+
+      if (m_docTypeReadVersion <= 0)
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;
+  }
+
+  if (pos != end)
+    return E_FILE_FORMAT_INVALID;
+
+  // Make sure DocType, DocTypeReadVersion, and DocTypeVersion are valid.
+  if (m_docType == NULL || m_docTypeReadVersion <= 0 || m_docTypeVersion <= 0)
+    return E_FILE_FORMAT_INVALID;
+
+  // Make sure EBMLMaxIDLength and EBMLMaxSizeLength are valid.
+  if (m_maxIdLength <= 0 || m_maxIdLength > 4 || m_maxSizeLength <= 0 ||
+      m_maxSizeLength > 8)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+Segment::Segment(IMkvReader* pReader, long long elem_start,
+                 // long long elem_size,
+                 long long start, long long size)
+    : m_pReader(pReader),
+      m_element_start(elem_start),
+      // m_element_size(elem_size),
+      m_start(start),
+      m_size(size),
+      m_pos(start),
+      m_pUnknownSize(0),
+      m_pSeekHead(NULL),
+      m_pInfo(NULL),
+      m_pTracks(NULL),
+      m_pCues(NULL),
+      m_pChapters(NULL),
+      m_pTags(NULL),
+      m_clusters(NULL),
+      m_clusterCount(0),
+      m_clusterPreloadCount(0),
+      m_clusterSize(0) {}
+
+Segment::~Segment() {
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  Cluster** i = m_clusters;
+  Cluster** j = m_clusters + count;
+
+  while (i != j) {
+    Cluster* const p = *i++;
+    delete p;
+  }
+
+  delete[] m_clusters;
+
+  delete m_pTracks;
+  delete m_pInfo;
+  delete m_pCues;
+  delete m_pChapters;
+  delete m_pTags;
+  delete m_pSeekHead;
+}
+
+long long Segment::CreateInstance(IMkvReader* pReader, long long pos,
+                                  Segment*& pSegment) {
+  if (pReader == NULL || pos < 0)
+    return E_PARSE_FAILED;
+
+  pSegment = NULL;
+
+  long long total, available;
+
+  const long status = pReader->Length(&total, &available);
+
+  if (status < 0)  // error
+    return status;
+
+  if (available < 0)
+    return -1;
+
+  if ((total >= 0) && (available > total))
+    return -1;
+
+  // I would assume that in practice this loop would execute
+  // exactly once, but we allow for other elements (e.g. Void)
+  // to immediately follow the EBML header.  This is fine for
+  // the source filter case (since the entire file is available),
+  // but in the splitter case over a network we should probably
+  // just give up early.  We could for example decide only to
+  // execute this loop a maximum of, say, 10 times.
+  // TODO:
+  // There is an implied "give up early" by only parsing up
+  // to the available limit.  We do do that, but only if the
+  // total file size is unknown.  We could decide to always
+  // use what's available as our limit (irrespective of whether
+  // we happen to know the total file length).  This would have
+  // as its sense "parse this much of the file before giving up",
+  // which a slightly different sense from "try to parse up to
+  // 10 EMBL elements before giving up".
+
+  for (;;) {
+    if ((total >= 0) && (pos >= total))
+      return E_FILE_FORMAT_INVALID;
+
+    // Read ID
+    long len;
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result)  // error, or too few available bytes
+      return result;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    const long long idpos = pos;
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result)  // error, or too few available bytes
+      return result;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return size;
+
+    pos += len;  // consume length of size of element
+
+    // Pos now points to start of payload
+
+    // Handle "unknown size" for live streaming of webm files.
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (id == libwebm::kMkvSegment) {
+      if (size == unknown_size)
+        size = -1;
+
+      else if (total < 0)
+        size = -1;
+
+      else if ((pos + size) > total)
+        size = -1;
+
+      pSegment = new (std::nothrow) Segment(pReader, idpos, pos, size);
+      if (pSegment == NULL)
+        return E_PARSE_FAILED;
+
+      return 0;  // success
+    }
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && ((pos + size) > total))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + size) > available)
+      return pos + size;
+
+    pos += size;  // consume payload
+  }
+}
+
+long long Segment::ParseHeaders() {
+  // Outermost (level 0) segment object has been constructed,
+  // and pos designates start of payload.  We need to find the
+  // inner (level 1) elements.
+  long long total, available;
+
+  const int status = m_pReader->Length(&total, &available);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total > 0 && available > total)
+    return E_FILE_FORMAT_INVALID;
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  if ((segment_stop >= 0 && total >= 0 && segment_stop > total) ||
+      (segment_stop >= 0 && m_pos > segment_stop)) {
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  for (;;) {
+    if ((total >= 0) && (m_pos >= total))
+      break;
+
+    if ((segment_stop >= 0) && (m_pos >= segment_stop))
+      break;
+
+    long long pos = m_pos;
+    const long long element_start = pos;
+
+    // Avoid rolling over pos when very close to LLONG_MAX.
+    unsigned long long rollover_check = pos + 1ULL;
+    if (rollover_check > LLONG_MAX)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + 1) > available)
+      return (pos + 1);
+
+    long len;
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return result;
+
+    if (result > 0) {
+      // MkvReader doesn't have enough data to satisfy this read attempt.
+      return (pos + 1);
+    }
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    const long long idpos = pos;
+    const long long id = ReadID(m_pReader, idpos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvCluster)
+      break;
+
+    pos += len;  // consume ID
+
+    if ((pos + 1) > available)
+      return (pos + 1);
+
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return result;
+
+    if (result > 0) {
+      // MkvReader doesn't have enough data to satisfy this read attempt.
+      return (pos + 1);
+    }
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > available)
+      return pos + len;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0 || len < 1 || len > 8) {
+      // TODO(tomfinegan): ReadUInt should return an error when len is < 1 or
+      // len > 8 is true instead of checking this _everywhere_.
+      return size;
+    }
+
+    pos += len;  // consume length of size of element
+
+    // Avoid rolling over pos when very close to LLONG_MAX.
+    rollover_check = static_cast<unsigned long long>(pos) + size;
+    if (rollover_check > LLONG_MAX)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long element_size = size + pos - element_start;
+
+    // Pos now points to start of payload
+
+    if ((segment_stop >= 0) && ((pos + size) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // We read EBML elements either in total or nothing at all.
+
+    if ((pos + size) > available)
+      return pos + size;
+
+    if (id == libwebm::kMkvInfo) {
+      if (m_pInfo)
+        return E_FILE_FORMAT_INVALID;
+
+      m_pInfo = new (std::nothrow)
+          SegmentInfo(this, pos, size, element_start, element_size);
+
+      if (m_pInfo == NULL)
+        return -1;
+
+      const long status = m_pInfo->Parse();
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvTracks) {
+      if (m_pTracks)
+        return E_FILE_FORMAT_INVALID;
+
+      m_pTracks = new (std::nothrow)
+          Tracks(this, pos, size, element_start, element_size);
+
+      if (m_pTracks == NULL)
+        return -1;
+
+      const long status = m_pTracks->Parse();
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvCues) {
+      if (m_pCues == NULL) {
+        m_pCues = new (std::nothrow)
+            Cues(this, pos, size, element_start, element_size);
+
+        if (m_pCues == NULL)
+          return -1;
+      }
+    } else if (id == libwebm::kMkvSeekHead) {
+      if (m_pSeekHead == NULL) {
+        m_pSeekHead = new (std::nothrow)
+            SeekHead(this, pos, size, element_start, element_size);
+
+        if (m_pSeekHead == NULL)
+          return -1;
+
+        const long status = m_pSeekHead->Parse();
+
+        if (status)
+          return status;
+      }
+    } else if (id == libwebm::kMkvChapters) {
+      if (m_pChapters == NULL) {
+        m_pChapters = new (std::nothrow)
+            Chapters(this, pos, size, element_start, element_size);
+
+        if (m_pChapters == NULL)
+          return -1;
+
+        const long status = m_pChapters->Parse();
+
+        if (status)
+          return status;
+      }
+    } else if (id == libwebm::kMkvTags) {
+      if (m_pTags == NULL) {
+        m_pTags = new (std::nothrow)
+            Tags(this, pos, size, element_start, element_size);
+
+        if (m_pTags == NULL)
+          return -1;
+
+        const long status = m_pTags->Parse();
+
+        if (status)
+          return status;
+      }
+    }
+
+    m_pos = pos + size;  // consume payload
+  }
+
+  if (segment_stop >= 0 && m_pos > segment_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (m_pInfo == NULL)  // TODO: liberalize this behavior
+    return E_FILE_FORMAT_INVALID;
+
+  if (m_pTracks == NULL)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;  // success
+}
+
+long Segment::LoadCluster(long long& pos, long& len) {
+  for (;;) {
+    const long result = DoLoadCluster(pos, len);
+
+    if (result <= 1)
+      return result;
+  }
+}
+
+long Segment::DoLoadCluster(long long& pos, long& len) {
+  if (m_pos < 0)
+    return DoLoadClusterUnknownSize(pos, len);
+
+  long long total, avail;
+
+  long status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total >= 0 && avail > total)
+    return E_FILE_FORMAT_INVALID;
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  long long cluster_off = -1;  // offset relative to start of segment
+  long long cluster_size = -1;  // size of cluster payload
+
+  for (;;) {
+    if ((total >= 0) && (m_pos >= total))
+      return 1;  // no more clusters
+
+    if ((segment_stop >= 0) && (m_pos >= segment_stop))
+      return 1;  // no more clusters
+
+    pos = m_pos;
+
+    // Read ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long idpos = pos;
+    const long long id = ReadID(m_pReader, idpos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume length of size of element
+
+    // pos now points to start of payload
+
+    if (size == 0) {
+      // Missing element payload: move on.
+      m_pos = pos;
+      continue;
+    }
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if ((segment_stop >= 0) && (size != unknown_size) &&
+        ((pos + size) > segment_stop)) {
+      return E_FILE_FORMAT_INVALID;
+    }
+
+    if (id == libwebm::kMkvCues) {
+      if (size == unknown_size) {
+        // Cues element of unknown size: Not supported.
+        return E_FILE_FORMAT_INVALID;
+      }
+
+      if (m_pCues == NULL) {
+        const long long element_size = (pos - idpos) + size;
+
+        m_pCues = new (std::nothrow) Cues(this, pos, size, idpos, element_size);
+        if (m_pCues == NULL)
+          return -1;
+      }
+
+      m_pos = pos + size;  // consume payload
+      continue;
+    }
+
+    if (id != libwebm::kMkvCluster) {
+      // Besides the Segment, Libwebm allows only cluster elements of unknown
+      // size. Fail the parse upon encountering a non-cluster element reporting
+      // unknown size.
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
+
+      m_pos = pos + size;  // consume payload
+      continue;
+    }
+
+    // We have a cluster.
+
+    cluster_off = idpos - m_start;  // relative pos
+
+    if (size != unknown_size)
+      cluster_size = size;
+
+    break;
+  }
+
+  if (cluster_off < 0) {
+    // No cluster, die.
+    return E_FILE_FORMAT_INVALID;
+  }
+
+  long long pos_;
+  long len_;
+
+  status = Cluster::HasBlockEntries(this, cluster_off, pos_, len_);
+
+  if (status < 0) {  // error, or underflow
+    pos = pos_;
+    len = len_;
+
+    return status;
+  }
+
+  // status == 0 means "no block entries found"
+  // status > 0 means "found at least one block entry"
+
+  // TODO:
+  // The issue here is that the segment increments its own
+  // pos ptr past the most recent cluster parsed, and then
+  // starts from there to parse the next cluster.  If we
+  // don't know the size of the current cluster, then we
+  // must either parse its payload (as we do below), looking
+  // for the cluster (or cues) ID to terminate the parse.
+  // This isn't really what we want: rather, we really need
+  // a way to create the curr cluster object immediately.
+  // The pity is that cluster::parse can determine its own
+  // boundary, and we largely duplicate that same logic here.
+  //
+  // Maybe we need to get rid of our look-ahead preloading
+  // in source::parse???
+  //
+  // As we're parsing the blocks in the curr cluster
+  //(in cluster::parse), we should have some way to signal
+  // to the segment that we have determined the boundary,
+  // so it can adjust its own segment::m_pos member.
+  //
+  // The problem is that we're asserting in asyncreadinit,
+  // because we adjust the pos down to the curr seek pos,
+  // and the resulting adjusted len is > 2GB.  I'm suspicious
+  // that this is even correct, but even if it is, we can't
+  // be loading that much data in the cache anyway.
+
+  const long idx = m_clusterCount;
+
+  if (m_clusterPreloadCount > 0) {
+    if (idx >= m_clusterSize)
+      return E_FILE_FORMAT_INVALID;
+
+    Cluster* const pCluster = m_clusters[idx];
+    if (pCluster == NULL || pCluster->m_index >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long off = pCluster->GetPosition();
+    if (off < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (off == cluster_off) {  // preloaded already
+      if (status == 0)  // no entries found
+        return E_FILE_FORMAT_INVALID;
+
+      if (cluster_size >= 0)
+        pos += cluster_size;
+      else {
+        const long long element_size = pCluster->GetElementSize();
+
+        if (element_size <= 0)
+          return E_FILE_FORMAT_INVALID;  // TODO: handle this case
+
+        pos = pCluster->m_element_start + element_size;
+      }
+
+      pCluster->m_index = idx;  // move from preloaded to loaded
+      ++m_clusterCount;
+      --m_clusterPreloadCount;
+
+      m_pos = pos;  // consume payload
+      if (segment_stop >= 0 && m_pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      return 0;  // success
+    }
+  }
+
+  if (status == 0) {  // no entries found
+    if (cluster_size >= 0)
+      pos += cluster_size;
+
+    if ((total >= 0) && (pos >= total)) {
+      m_pos = total;
+      return 1;  // no more clusters
+    }
+
+    if ((segment_stop >= 0) && (pos >= segment_stop)) {
+      m_pos = segment_stop;
+      return 1;  // no more clusters
+    }
+
+    m_pos = pos;
+    return 2;  // try again
+  }
+
+  // status > 0 means we have an entry
+
+  Cluster* const pCluster = Cluster::Create(this, idx, cluster_off);
+  if (pCluster == NULL)
+    return -1;
+
+  if (!AppendCluster(pCluster)) {
+    delete pCluster;
+    return -1;
+  }
+
+  if (cluster_size >= 0) {
+    pos += cluster_size;
+
+    m_pos = pos;
+
+    if (segment_stop > 0 && m_pos > segment_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    return 0;
+  }
+
+  m_pUnknownSize = pCluster;
+  m_pos = -pos;
+
+  return 0;  // partial success, since we have a new cluster
+
+  // status == 0 means "no block entries found"
+  // pos designates start of payload
+  // m_pos has NOT been adjusted yet (in case we need to come back here)
+}
+
+long Segment::DoLoadClusterUnknownSize(long long& pos, long& len) {
+  if (m_pos >= 0 || m_pUnknownSize == NULL)
+    return E_PARSE_FAILED;
+
+  const long status = m_pUnknownSize->Parse(pos, len);
+
+  if (status < 0)  // error or underflow
+    return status;
+
+  if (status == 0)  // parsed a block
+    return 2;  // continue parsing
+
+  const long long start = m_pUnknownSize->m_element_start;
+  const long long size = m_pUnknownSize->GetElementSize();
+
+  if (size < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos = start + size;
+  m_pos = pos;
+
+  m_pUnknownSize = 0;
+
+  return 2;  // continue parsing
+}
+
+bool Segment::AppendCluster(Cluster* pCluster) {
+  if (pCluster == NULL || pCluster->m_index < 0)
+    return false;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  long& size = m_clusterSize;
+  const long idx = pCluster->m_index;
+
+  if (size < count || idx != m_clusterCount)
+    return false;
+
+  if (count >= size) {
+    const long n = (size <= 0) ? 2048 : 2 * size;
+
+    Cluster** const qq = new (std::nothrow) Cluster*[n];
+    if (qq == NULL)
+      return false;
+
+    Cluster** q = qq;
+    Cluster** p = m_clusters;
+    Cluster** const pp = p + count;
+
+    while (p != pp)
+      *q++ = *p++;
+
+    delete[] m_clusters;
+
+    m_clusters = qq;
+    size = n;
+  }
+
+  if (m_clusterPreloadCount > 0) {
+    Cluster** const p = m_clusters + m_clusterCount;
+    if (*p == NULL || (*p)->m_index >= 0)
+      return false;
+
+    Cluster** q = p + m_clusterPreloadCount;
+    if (q >= (m_clusters + size))
+      return false;
+
+    for (;;) {
+      Cluster** const qq = q - 1;
+      if ((*qq)->m_index >= 0)
+        return false;
+
+      *q = *qq;
+      q = qq;
+
+      if (q == p)
+        break;
+    }
+  }
+
+  m_clusters[idx] = pCluster;
+  ++m_clusterCount;
+  return true;
+}
+
+bool Segment::PreloadCluster(Cluster* pCluster, ptrdiff_t idx) {
+  if (pCluster == NULL || pCluster->m_index >= 0 || idx < m_clusterCount)
+    return false;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  long& size = m_clusterSize;
+  if (size < count)
+    return false;
+
+  if (count >= size) {
+    const long n = (size <= 0) ? 2048 : 2 * size;
+
+    Cluster** const qq = new (std::nothrow) Cluster*[n];
+    if (qq == NULL)
+      return false;
+    Cluster** q = qq;
+
+    Cluster** p = m_clusters;
+    Cluster** const pp = p + count;
+
+    while (p != pp)
+      *q++ = *p++;
+
+    delete[] m_clusters;
+
+    m_clusters = qq;
+    size = n;
+  }
+
+  if (m_clusters == NULL)
+    return false;
+
+  Cluster** const p = m_clusters + idx;
+
+  Cluster** q = m_clusters + count;
+  if (q < p || q >= (m_clusters + size))
+    return false;
+
+  while (q > p) {
+    Cluster** const qq = q - 1;
+
+    if ((*qq)->m_index >= 0)
+      return false;
+
+    *q = *qq;
+    q = qq;
+  }
+
+  m_clusters[idx] = pCluster;
+  ++m_clusterPreloadCount;
+  return true;
+}
+
+long Segment::Load() {
+  if (m_clusters != NULL || m_clusterSize != 0 || m_clusterCount != 0)
+    return E_PARSE_FAILED;
+
+  // Outermost (level 0) segment object has been constructed,
+  // and pos designates start of payload.  We need to find the
+  // inner (level 1) elements.
+
+  const long long header_status = ParseHeaders();
+
+  if (header_status < 0)  // error
+    return static_cast<long>(header_status);
+
+  if (header_status > 0)  // underflow
+    return E_BUFFER_NOT_FULL;
+
+  if (m_pInfo == NULL || m_pTracks == NULL)
+    return E_FILE_FORMAT_INVALID;
+
+  for (;;) {
+    const long status = LoadCluster();
+
+    if (status < 0)  // error
+      return status;
+
+    if (status >= 1)  // no more clusters
+      return 0;
+  }
+}
+
+SeekHead::SeekHead(Segment* pSegment, long long start, long long size_,
+                   long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_entries(0),
+      m_entry_count(0),
+      m_void_elements(0),
+      m_void_element_count(0) {}
+
+SeekHead::~SeekHead() {
+  delete[] m_entries;
+  delete[] m_void_elements;
+}
+
+long SeekHead::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
+
+  // first count the seek head entries
+
+  int entry_count = 0;
+  int void_element_count = 0;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvSeek)
+      ++entry_count;
+    else if (id == libwebm::kMkvVoid)
+      ++void_element_count;
+
+    pos += size;  // consume payload
+
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  m_entries = new (std::nothrow) Entry[entry_count];
+
+  if (m_entries == NULL)
+    return -1;
+
+  m_void_elements = new (std::nothrow) VoidElement[void_element_count];
+
+  if (m_void_elements == NULL)
+    return -1;
+
+  // now parse the entries and void elements
+
+  Entry* pEntry = m_entries;
+  VoidElement* pVoidElement = m_void_elements;
+
+  pos = m_start;
+
+  while (pos < stop) {
+    const long long idpos = pos;
+
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvSeek) {
+      if (ParseEntry(pReader, pos, size, pEntry)) {
+        Entry& e = *pEntry++;
+
+        e.element_start = idpos;
+        e.element_size = (pos + size) - idpos;
+      }
+    } else if (id == libwebm::kMkvVoid) {
+      VoidElement& e = *pVoidElement++;
+
+      e.element_start = idpos;
+      e.element_size = (pos + size) - idpos;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  ptrdiff_t count_ = ptrdiff_t(pEntry - m_entries);
+  assert(count_ >= 0);
+  assert(count_ <= entry_count);
+
+  m_entry_count = static_cast<int>(count_);
+
+  count_ = ptrdiff_t(pVoidElement - m_void_elements);
+  assert(count_ >= 0);
+  assert(count_ <= void_element_count);
+
+  m_void_element_count = static_cast<int>(count_);
+
+  return 0;
+}
+
+int SeekHead::GetCount() const { return m_entry_count; }
+
+const SeekHead::Entry* SeekHead::GetEntry(int idx) const {
+  if (idx < 0)
+    return 0;
+
+  if (idx >= m_entry_count)
+    return 0;
+
+  return m_entries + idx;
+}
+
+int SeekHead::GetVoidElementCount() const { return m_void_element_count; }
+
+const SeekHead::VoidElement* SeekHead::GetVoidElement(int idx) const {
+  if (idx < 0)
+    return 0;
+
+  if (idx >= m_void_element_count)
+    return 0;
+
+  return m_void_elements + idx;
+}
+
+long Segment::ParseCues(long long off, long long& pos, long& len) {
+  if (m_pCues)
+    return 0;  // success
+
+  if (off < 0)
+    return -1;
+
+  long long total, avail;
+
+  const int status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  pos = m_start + off;
+
+  if ((total < 0) || (pos >= total))
+    return 1;  // don't bother parsing cues
+
+  const long long element_start = pos;
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long result = GetUIntLength(m_pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)  // underflow (weird)
+  {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long idpos = pos;
+
+  const long long id = ReadID(m_pReader, idpos, len);
+
+  if (id != libwebm::kMkvCues)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume ID
+  assert((segment_stop < 0) || (pos <= segment_stop));
+
+  // Read Size
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  result = GetUIntLength(m_pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)  // underflow (weird)
+  {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long size = ReadUInt(m_pReader, pos, len);
+
+  if (size < 0)  // error
+    return static_cast<long>(size);
+
+  if (size == 0)  // weird, although technically not illegal
+    return 1;  // done
+
+  pos += len;  // consume length of size of element
+  assert((segment_stop < 0) || (pos <= segment_stop));
+
+  // Pos now points to start of payload
+
+  const long long element_stop = pos + size;
+
+  if ((segment_stop >= 0) && (element_stop > segment_stop))
+    return E_FILE_FORMAT_INVALID;
+
+  if ((total >= 0) && (element_stop > total))
+    return 1;  // don't bother parsing anymore
+
+  len = static_cast<long>(size);
+
+  if (element_stop > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long element_size = element_stop - element_start;
+
+  m_pCues =
+      new (std::nothrow) Cues(this, pos, size, element_start, element_size);
+  if (m_pCues == NULL)
+    return -1;
+
+  return 0;  // success
+}
+
+bool SeekHead::ParseEntry(IMkvReader* pReader, long long start, long long size_,
+                          Entry* pEntry) {
+  if (size_ <= 0)
+    return false;
+
+  long long pos = start;
+  const long long stop = start + size_;
+
+  long len;
+
+  // parse the container for the level-1 element ID
+
+  const long long seekIdId = ReadID(pReader, pos, len);
+  if (seekIdId < 0)
+    return false;
+
+  if (seekIdId != libwebm::kMkvSeekID)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume SeekID id
+
+  const long long seekIdSize = ReadUInt(pReader, pos, len);
+
+  if (seekIdSize <= 0)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume size of field
+
+  if ((pos + seekIdSize) > stop)
+    return false;
+
+  // Note that the SeekId payload really is serialized
+  // as a "Matroska integer", not as a plain binary value.
+  // In fact, Matroska requires that ID values in the
+  // stream exactly match the binary representation as listed
+  // in the Matroska specification.
+  //
+  // This parser is more liberal, and permits IDs to have
+  // any width.  (This could make the representation in the stream
+  // different from what's in the spec, but it doesn't matter here,
+  // since we always normalize "Matroska integer" values.)
+
+  pEntry->id = ReadUInt(pReader, pos, len);  // payload
+
+  if (pEntry->id <= 0)
+    return false;
+
+  if (len != seekIdSize)
+    return false;
+
+  pos += seekIdSize;  // consume SeekID payload
+
+  const long long seekPosId = ReadID(pReader, pos, len);
+
+  if (seekPosId != libwebm::kMkvSeekPosition)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume id
+
+  const long long seekPosSize = ReadUInt(pReader, pos, len);
+
+  if (seekPosSize <= 0)
+    return false;
+
+  if ((pos + len) > stop)
+    return false;
+
+  pos += len;  // consume size
+
+  if ((pos + seekPosSize) > stop)
+    return false;
+
+  pEntry->pos = UnserializeUInt(pReader, pos, seekPosSize);
+
+  if (pEntry->pos < 0)
+    return false;
+
+  pos += seekPosSize;  // consume payload
+
+  if (pos != stop)
+    return false;
+
+  return true;
+}
+
+Cues::Cues(Segment* pSegment, long long start_, long long size_,
+           long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start_),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_cue_points(NULL),
+      m_count(0),
+      m_preload_count(0),
+      m_pos(start_) {}
+
+Cues::~Cues() {
+  const long n = m_count + m_preload_count;
+
+  CuePoint** p = m_cue_points;
+  CuePoint** const q = p + n;
+
+  while (p != q) {
+    CuePoint* const pCP = *p++;
+    assert(pCP);
+
+    delete pCP;
+  }
+
+  delete[] m_cue_points;
+}
+
+long Cues::GetCount() const {
+  if (m_cue_points == NULL)
+    return -1;
+
+  return m_count;  // TODO: really ignore preload count?
+}
+
+bool Cues::DoneParsing() const {
+  const long long stop = m_start + m_size;
+  return (m_pos >= stop);
+}
+
+bool Cues::Init() const {
+  if (m_cue_points)
+    return true;
+
+  if (m_count != 0 || m_preload_count != 0)
+    return false;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  const long long stop = m_start + m_size;
+  long long pos = m_start;
+
+  long cue_points_size = 0;
+
+  while (pos < stop) {
+    const long long idpos = pos;
+
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if (id < 0 || (pos + len) > stop) {
+      return false;
+    }
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    if (size < 0 || (pos + len > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume Size field
+    if (pos + size > stop) {
+      return false;
+    }
+
+    if (id == libwebm::kMkvCuePoint) {
+      if (!PreloadCuePoint(cue_points_size, idpos))
+        return false;
+    }
+
+    pos += size;  // skip payload
+  }
+  return true;
+}
+
+bool Cues::PreloadCuePoint(long& cue_points_size, long long pos) const {
+  if (m_count != 0)
+    return false;
+
+  if (m_preload_count >= cue_points_size) {
+    const long n = (cue_points_size <= 0) ? 2048 : 2 * cue_points_size;
+
+    CuePoint** const qq = new (std::nothrow) CuePoint*[n];
+    if (qq == NULL)
+      return false;
+
+    CuePoint** q = qq;  // beginning of target
+
+    CuePoint** p = m_cue_points;  // beginning of source
+    CuePoint** const pp = p + m_preload_count;  // end of source
+
+    while (p != pp)
+      *q++ = *p++;
+
+    delete[] m_cue_points;
+
+    m_cue_points = qq;
+    cue_points_size = n;
+  }
+
+  CuePoint* const pCP = new (std::nothrow) CuePoint(m_preload_count, pos);
+  if (pCP == NULL)
+    return false;
+
+  m_cue_points[m_preload_count++] = pCP;
+  return true;
+}
+
+bool Cues::LoadCuePoint() const {
+  const long long stop = m_start + m_size;
+
+  if (m_pos >= stop)
+    return false;  // nothing else to do
+
+  if (!Init()) {
+    m_pos = stop;
+    return false;
+  }
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  while (m_pos < stop) {
+    const long long idpos = m_pos;
+
+    long len;
+
+    const long long id = ReadID(pReader, m_pos, len);
+    if (id < 0 || (m_pos + len) > stop)
+      return false;
+
+    m_pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, m_pos, len);
+    if (size < 0 || (m_pos + len) > stop)
+      return false;
+
+    m_pos += len;  // consume Size field
+    if ((m_pos + size) > stop)
+      return false;
+
+    if (id != libwebm::kMkvCuePoint) {
+      m_pos += size;  // consume payload
+      if (m_pos > stop)
+        return false;
+
+      continue;
+    }
+
+    if (m_preload_count < 1)
+      return false;
+
+    CuePoint* const pCP = m_cue_points[m_count];
+    if (!pCP || (pCP->GetTimeCode() < 0 && (-pCP->GetTimeCode() != idpos)))
+      return false;
+
+    if (!pCP->Load(pReader)) {
+      m_pos = stop;
+      return false;
+    }
+    ++m_count;
+    --m_preload_count;
+
+    m_pos += size;  // consume payload
+    if (m_pos > stop)
+      return false;
+
+    return true;  // yes, we loaded a cue point
+  }
+
+  return false;  // no, we did not load a cue point
+}
+
+bool Cues::Find(long long time_ns, const Track* pTrack, const CuePoint*& pCP,
+                const CuePoint::TrackPosition*& pTP) const {
+  if (time_ns < 0 || pTrack == NULL || m_cue_points == NULL || m_count == 0)
+    return false;
+
+  CuePoint** const ii = m_cue_points;
+  CuePoint** i = ii;
+
+  CuePoint** const jj = ii + m_count;
+  CuePoint** j = jj;
+
+  pCP = *i;
+  if (pCP == NULL)
+    return false;
+
+  if (time_ns <= pCP->GetTime(m_pSegment)) {
+    pTP = pCP->Find(pTrack);
+    return (pTP != NULL);
+  }
+
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) <= time_ns
+    //[i, j)  ?
+    //[j, jj) > time_ns
+
+    CuePoint** const k = i + (j - i) / 2;
+    if (k >= jj)
+      return false;
+
+    CuePoint* const pCP = *k;
+    if (pCP == NULL)
+      return false;
+
+    const long long t = pCP->GetTime(m_pSegment);
+
+    if (t <= time_ns)
+      i = k + 1;
+    else
+      j = k;
+
+    if (i > j)
+      return false;
+  }
+
+  if (i != j || i > jj || i <= ii)
+    return false;
+
+  pCP = *--i;
+
+  if (pCP == NULL || pCP->GetTime(m_pSegment) > time_ns)
+    return false;
+
+  // TODO: here and elsewhere, it's probably not correct to search
+  // for the cue point with this time, and then search for a matching
+  // track.  In principle, the matching track could be on some earlier
+  // cue point, and with our current algorithm, we'd miss it.  To make
+  // this bullet-proof, we'd need to create a secondary structure,
+  // with a list of cue points that apply to a track, and then search
+  // that track-based structure for a matching cue point.
+
+  pTP = pCP->Find(pTrack);
+  return (pTP != NULL);
+}
+
+const CuePoint* Cues::GetFirst() const {
+  if (m_cue_points == NULL || m_count == 0)
+    return NULL;
+
+  CuePoint* const* const pp = m_cue_points;
+  if (pp == NULL)
+    return NULL;
+
+  CuePoint* const pCP = pp[0];
+  if (pCP == NULL || pCP->GetTimeCode() < 0)
+    return NULL;
+
+  return pCP;
+}
+
+const CuePoint* Cues::GetLast() const {
+  if (m_cue_points == NULL || m_count <= 0)
+    return NULL;
+
+  const long index = m_count - 1;
+
+  CuePoint* const* const pp = m_cue_points;
+  if (pp == NULL)
+    return NULL;
+
+  CuePoint* const pCP = pp[index];
+  if (pCP == NULL || pCP->GetTimeCode() < 0)
+    return NULL;
+
+  return pCP;
+}
+
+const CuePoint* Cues::GetNext(const CuePoint* pCurr) const {
+  if (pCurr == NULL || pCurr->GetTimeCode() < 0 || m_cue_points == NULL ||
+      m_count < 1) {
+    return NULL;
+  }
+
+  long index = pCurr->m_index;
+  if (index >= m_count)
+    return NULL;
+
+  CuePoint* const* const pp = m_cue_points;
+  if (pp == NULL || pp[index] != pCurr)
+    return NULL;
+
+  ++index;
+
+  if (index >= m_count)
+    return NULL;
+
+  CuePoint* const pNext = pp[index];
+
+  if (pNext == NULL || pNext->GetTimeCode() < 0)
+    return NULL;
+
+  return pNext;
+}
+
+const BlockEntry* Cues::GetBlock(const CuePoint* pCP,
+                                 const CuePoint::TrackPosition* pTP) const {
+  if (pCP == NULL || pTP == NULL)
+    return NULL;
+
+  return m_pSegment->GetBlock(*pCP, *pTP);
+}
+
+const BlockEntry* Segment::GetBlock(const CuePoint& cp,
+                                    const CuePoint::TrackPosition& tp) {
+  Cluster** const ii = m_clusters;
+  Cluster** i = ii;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  Cluster** const jj = ii + count;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) < pTP->m_pos
+    //[i, j) ?
+    //[j, jj)  > pTP->m_pos
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    Cluster* const pCluster = *k;
+    assert(pCluster);
+
+    // const long long pos_ = pCluster->m_pos;
+    // assert(pos_);
+    // const long long pos = pos_ * ((pos_ < 0) ? -1 : 1);
+
+    const long long pos = pCluster->GetPosition();
+    assert(pos >= 0);
+
+    if (pos < tp.m_pos)
+      i = k + 1;
+    else if (pos > tp.m_pos)
+      j = k;
+    else
+      return pCluster->GetEntry(cp, tp);
+  }
+
+  assert(i == j);
+  // assert(Cluster::HasBlockEntries(this, tp.m_pos));
+
+  Cluster* const pCluster = Cluster::Create(this, -1, tp.m_pos);  //, -1);
+  if (pCluster == NULL)
+    return NULL;
+
+  const ptrdiff_t idx = i - m_clusters;
+
+  if (!PreloadCluster(pCluster, idx)) {
+    delete pCluster;
+    return NULL;
+  }
+  assert(m_clusters);
+  assert(m_clusterPreloadCount > 0);
+  assert(m_clusters[idx] == pCluster);
+
+  return pCluster->GetEntry(cp, tp);
+}
+
+const Cluster* Segment::FindOrPreloadCluster(long long requested_pos) {
+  if (requested_pos < 0)
+    return 0;
+
+  Cluster** const ii = m_clusters;
+  Cluster** i = ii;
+
+  const long count = m_clusterCount + m_clusterPreloadCount;
+
+  Cluster** const jj = ii + count;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[ii, i) < pTP->m_pos
+    //[i, j) ?
+    //[j, jj)  > pTP->m_pos
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    Cluster* const pCluster = *k;
+    assert(pCluster);
+
+    // const long long pos_ = pCluster->m_pos;
+    // assert(pos_);
+    // const long long pos = pos_ * ((pos_ < 0) ? -1 : 1);
+
+    const long long pos = pCluster->GetPosition();
+    assert(pos >= 0);
+
+    if (pos < requested_pos)
+      i = k + 1;
+    else if (pos > requested_pos)
+      j = k;
+    else
+      return pCluster;
+  }
+
+  assert(i == j);
+  // assert(Cluster::HasBlockEntries(this, tp.m_pos));
+
+  Cluster* const pCluster = Cluster::Create(this, -1, requested_pos);
+  if (pCluster == NULL)
+    return NULL;
+
+  const ptrdiff_t idx = i - m_clusters;
+
+  if (!PreloadCluster(pCluster, idx)) {
+    delete pCluster;
+    return NULL;
+  }
+  assert(m_clusters);
+  assert(m_clusterPreloadCount > 0);
+  assert(m_clusters[idx] == pCluster);
+
+  return pCluster;
+}
+
+CuePoint::CuePoint(long idx, long long pos)
+    : m_element_start(0),
+      m_element_size(0),
+      m_index(idx),
+      m_timecode(-1 * pos),
+      m_track_positions(NULL),
+      m_track_positions_count(0) {
+  assert(pos > 0);
+}
+
+CuePoint::~CuePoint() { delete[] m_track_positions; }
+
+bool CuePoint::Load(IMkvReader* pReader) {
+  // odbgstream os;
+  // os << "CuePoint::Load(begin): timecode=" << m_timecode << endl;
+
+  if (m_timecode >= 0)  // already loaded
+    return true;
+
+  assert(m_track_positions == NULL);
+  assert(m_track_positions_count == 0);
+
+  long long pos_ = -m_timecode;
+  const long long element_start = pos_;
+
+  long long stop;
+
+  {
+    long len;
+
+    const long long id = ReadID(pReader, pos_, len);
+    if (id != libwebm::kMkvCuePoint)
+      return false;
+
+    pos_ += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos_, len);
+    assert(size >= 0);
+
+    pos_ += len;  // consume Size field
+    // pos_ now points to start of payload
+
+    stop = pos_ + size;
+  }
+
+  const long long element_size = stop - element_start;
+
+  long long pos = pos_;
+
+  // First count number of track positions
+
+  while (pos < stop) {
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if ((id < 0) || (pos + len > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    if ((size < 0) || (pos + len > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume Size field
+    if ((pos + size) > stop) {
+      return false;
+    }
+
+    if (id == libwebm::kMkvCueTime)
+      m_timecode = UnserializeUInt(pReader, pos, size);
+
+    else if (id == libwebm::kMkvCueTrackPositions)
+      ++m_track_positions_count;
+
+    pos += size;  // consume payload
+  }
+
+  if (m_timecode < 0 || m_track_positions_count <= 0) {
+    return false;
+  }
+
+  // os << "CuePoint::Load(cont'd): idpos=" << idpos
+  //   << " timecode=" << m_timecode
+  //   << endl;
+
+  m_track_positions = new (std::nothrow) TrackPosition[m_track_positions_count];
+  if (m_track_positions == NULL)
+    return false;
+
+  // Now parse track positions
+
+  TrackPosition* p = m_track_positions;
+  pos = pos_;
+
+  while (pos < stop) {
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if (id < 0 || (pos + len) > stop)
+      return false;
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);
+    assert((pos + len) <= stop);
+
+    pos += len;  // consume Size field
+    assert((pos + size) <= stop);
+
+    if (id == libwebm::kMkvCueTrackPositions) {
+      TrackPosition& tp = *p++;
+      if (!tp.Parse(pReader, pos, size)) {
+        return false;
+      }
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return false;
+  }
+
+  assert(size_t(p - m_track_positions) == m_track_positions_count);
+
+  m_element_start = element_start;
+  m_element_size = element_size;
+
+  return true;
+}
+
+bool CuePoint::TrackPosition::Parse(IMkvReader* pReader, long long start_,
+                                    long long size_) {
+  const long long stop = start_ + size_;
+  long long pos = start_;
+
+  m_track = -1;
+  m_pos = -1;
+  m_block = 1;  // default
+
+  while (pos < stop) {
+    long len;
+
+    const long long id = ReadID(pReader, pos, len);
+    if ((id < 0) || ((pos + len) > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    if ((size < 0) || ((pos + len) > stop)) {
+      return false;
+    }
+
+    pos += len;  // consume Size field
+    if ((pos + size) > stop) {
+      return false;
+    }
+
+    if (id == libwebm::kMkvCueTrack)
+      m_track = UnserializeUInt(pReader, pos, size);
+    else if (id == libwebm::kMkvCueClusterPosition)
+      m_pos = UnserializeUInt(pReader, pos, size);
+    else if (id == libwebm::kMkvCueBlockNumber)
+      m_block = UnserializeUInt(pReader, pos, size);
+
+    pos += size;  // consume payload
+  }
+
+  if ((m_pos < 0) || (m_track <= 0)) {
+    return false;
+  }
+
+  return true;
+}
+
+const CuePoint::TrackPosition* CuePoint::Find(const Track* pTrack) const {
+  assert(pTrack);
+
+  const long long n = pTrack->GetNumber();
+
+  const TrackPosition* i = m_track_positions;
+  const TrackPosition* const j = i + m_track_positions_count;
+
+  while (i != j) {
+    const TrackPosition& p = *i++;
+
+    if (p.m_track == n)
+      return &p;
+  }
+
+  return NULL;  // no matching track number found
+}
+
+long long CuePoint::GetTimeCode() const { return m_timecode; }
+
+long long CuePoint::GetTime(const Segment* pSegment) const {
+  assert(pSegment);
+  assert(m_timecode >= 0);
+
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+  assert(pInfo);
+
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
+
+  const long long time = scale * m_timecode;
+
+  return time;
+}
+
+bool Segment::DoneParsing() const {
+  if (m_size < 0) {
+    long long total, avail;
+
+    const int status = m_pReader->Length(&total, &avail);
+
+    if (status < 0)  // error
+      return true;  // must assume done
+
+    if (total < 0)
+      return false;  // assume live stream
+
+    return (m_pos >= total);
+  }
+
+  const long long stop = m_start + m_size;
+
+  return (m_pos >= stop);
+}
+
+const Cluster* Segment::GetFirst() const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
+
+  Cluster* const pCluster = m_clusters[0];
+  assert(pCluster);
+
+  return pCluster;
+}
+
+const Cluster* Segment::GetLast() const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
+
+  const long idx = m_clusterCount - 1;
+
+  Cluster* const pCluster = m_clusters[idx];
+  assert(pCluster);
+
+  return pCluster;
+}
+
+unsigned long Segment::GetCount() const { return m_clusterCount; }
+
+const Cluster* Segment::GetNext(const Cluster* pCurr) {
+  assert(pCurr);
+  assert(pCurr != &m_eos);
+  assert(m_clusters);
+
+  long idx = pCurr->m_index;
+
+  if (idx >= 0) {
+    assert(m_clusterCount > 0);
+    assert(idx < m_clusterCount);
+    assert(pCurr == m_clusters[idx]);
+
+    ++idx;
+
+    if (idx >= m_clusterCount)
+      return &m_eos;  // caller will LoadCluster as desired
+
+    Cluster* const pNext = m_clusters[idx];
+    assert(pNext);
+    assert(pNext->m_index >= 0);
+    assert(pNext->m_index == idx);
+
+    return pNext;
+  }
+
+  assert(m_clusterPreloadCount > 0);
+
+  long long pos = pCurr->m_element_start;
+
+  assert(m_size >= 0);  // TODO
+  const long long stop = m_start + m_size;  // end of segment
+
+  {
+    long len;
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);
+    assert((pos + len) <= stop);  // TODO
+    if (result != 0)
+      return NULL;
+
+    const long long id = ReadID(m_pReader, pos, len);
+    if (id != libwebm::kMkvCluster)
+      return NULL;
+
+    pos += len;  // consume ID
+
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);  // TODO
+    assert((pos + len) <= stop);  // TODO
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+    assert(size > 0);  // TODO
+    // assert((pCurr->m_size <= 0) || (pCurr->m_size == size));
+
+    pos += len;  // consume length of size of element
+    assert((pos + size) <= stop);  // TODO
+
+    // Pos now points to start of payload
+
+    pos += size;  // consume payload
+  }
+
+  long long off_next = 0;
+
+  while (pos < stop) {
+    long len;
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);
+    assert((pos + len) <= stop);  // TODO
+    if (result != 0)
+      return NULL;
+
+    const long long idpos = pos;  // pos of next (potential) cluster
+
+    const long long id = ReadID(m_pReader, idpos, len);
+    if (id < 0)
+      return NULL;
+
+    pos += len;  // consume ID
+
+    // Read Size
+    result = GetUIntLength(m_pReader, pos, len);
+    assert(result == 0);  // TODO
+    assert((pos + len) <= stop);  // TODO
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+    assert(size >= 0);  // TODO
+
+    pos += len;  // consume length of size of element
+    assert((pos + size) <= stop);  // TODO
+
+    // Pos now points to start of payload
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvCluster) {
+      const long long off_next_ = idpos - m_start;
+
+      long long pos_;
+      long len_;
+
+      const long status = Cluster::HasBlockEntries(this, off_next_, pos_, len_);
+
+      assert(status >= 0);
+
+      if (status > 0) {
+        off_next = off_next_;
+        break;
+      }
+    }
+
+    pos += size;  // consume payload
+  }
+
+  if (off_next <= 0)
+    return 0;
+
+  Cluster** const ii = m_clusters + m_clusterCount;
+  Cluster** i = ii;
+
+  Cluster** const jj = ii + m_clusterPreloadCount;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) < pos_next
+    //[i, j) ?
+    //[j, jj)  > pos_next
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    Cluster* const pNext = *k;
+    assert(pNext);
+    assert(pNext->m_index < 0);
+
+    // const long long pos_ = pNext->m_pos;
+    // assert(pos_);
+    // pos = pos_ * ((pos_ < 0) ? -1 : 1);
+
+    pos = pNext->GetPosition();
+
+    if (pos < off_next)
+      i = k + 1;
+    else if (pos > off_next)
+      j = k;
+    else
+      return pNext;
+  }
+
+  assert(i == j);
+
+  Cluster* const pNext = Cluster::Create(this, -1, off_next);
+  if (pNext == NULL)
+    return NULL;
+
+  const ptrdiff_t idx_next = i - m_clusters;  // insertion position
+
+  if (!PreloadCluster(pNext, idx_next)) {
+    delete pNext;
+    return NULL;
+  }
+  assert(m_clusters);
+  assert(idx_next < m_clusterSize);
+  assert(m_clusters[idx_next] == pNext);
+
+  return pNext;
+}
+
+long Segment::ParseNext(const Cluster* pCurr, const Cluster*& pResult,
+                        long long& pos, long& len) {
+  assert(pCurr);
+  assert(!pCurr->EOS());
+  assert(m_clusters);
+
+  pResult = 0;
+
+  if (pCurr->m_index >= 0) {  // loaded (not merely preloaded)
+    assert(m_clusters[pCurr->m_index] == pCurr);
+
+    const long next_idx = pCurr->m_index + 1;
+
+    if (next_idx < m_clusterCount) {
+      pResult = m_clusters[next_idx];
+      return 0;  // success
+    }
+
+    // curr cluster is last among loaded
+
+    const long result = LoadCluster(pos, len);
+
+    if (result < 0)  // error or underflow
+      return result;
+
+    if (result > 0)  // no more clusters
+    {
+      // pResult = &m_eos;
+      return 1;
+    }
+
+    pResult = GetLast();
+    return 0;  // success
+  }
+
+  assert(m_pos > 0);
+
+  long long total, avail;
+
+  long status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  // interrogate curr cluster
+
+  pos = pCurr->m_element_start;
+
+  if (pCurr->m_element_size >= 0)
+    pos += pCurr->m_element_size;
+  else {
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadUInt(m_pReader, pos, len);
+
+    if (id != libwebm::kMkvCluster)
+      return -1;
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume size field
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)  // TODO: should never happen
+      return E_FILE_FORMAT_INVALID;  // TODO: resolve this
+
+    // assert((pCurr->m_size <= 0) || (pCurr->m_size == size));
+
+    if ((segment_stop >= 0) && ((pos + size) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // Pos now points to start of payload
+
+    pos += size;  // consume payload (that is, the current cluster)
+    if (segment_stop >= 0 && pos > segment_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    // By consuming the payload, we are assuming that the curr
+    // cluster isn't interesting.  That is, we don't bother checking
+    // whether the payload of the curr cluster is less than what
+    // happens to be available (obtained via IMkvReader::Length).
+    // Presumably the caller has already dispensed with the current
+    // cluster, and really does want the next cluster.
+  }
+
+  // pos now points to just beyond the last fully-loaded cluster
+
+  for (;;) {
+    const long status = DoParseNext(pResult, pos, len);
+
+    if (status <= 1)
+      return status;
+  }
+}
+
+long Segment::DoParseNext(const Cluster*& pResult, long long& pos, long& len) {
+  long long total, avail;
+
+  long status = m_pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  const long long segment_stop = (m_size < 0) ? -1 : m_start + m_size;
+
+  // Parse next cluster.  This is strictly a parsing activity.
+  // Creation of a new cluster object happens later, after the
+  // parsing is done.
+
+  long long off_next = 0;
+  long long cluster_size = -1;
+
+  for (;;) {
+    if ((total >= 0) && (pos >= total))
+      return 1;  // EOF
+
+    if ((segment_stop >= 0) && (pos >= segment_stop))
+      return 1;  // EOF
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long idpos = pos;  // absolute
+    const long long idoff = pos - m_start;  // relative
+
+    const long long id = ReadID(m_pReader, idpos, len);  // absolute
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id == 0)  // weird
+      return -1;  // generic error
+
+    pos += len;  // consume ID
+
+    // Read Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(m_pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(m_pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume length of size of element
+
+    // Pos now points to start of payload
+
+    if (size == 0)  // weird
+      continue;
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if ((segment_stop >= 0) && (size != unknown_size) &&
+        ((pos + size) > segment_stop)) {
+      return E_FILE_FORMAT_INVALID;
+    }
+
+    if (id == libwebm::kMkvCues) {
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
+
+      const long long element_stop = pos + size;
+
+      if ((segment_stop >= 0) && (element_stop > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      const long long element_start = idpos;
+      const long long element_size = element_stop - element_start;
+
+      if (m_pCues == NULL) {
+        m_pCues = new (std::nothrow)
+            Cues(this, pos, size, element_start, element_size);
+        if (m_pCues == NULL)
+          return false;
+      }
+
+      pos += size;  // consume payload
+      if (segment_stop >= 0 && pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      continue;
+    }
+
+    if (id != libwebm::kMkvCluster) {  // not a Cluster ID
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;
+
+      pos += size;  // consume payload
+      if (segment_stop >= 0 && pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      continue;
+    }
+
+    // We have a cluster.
+    off_next = idoff;
+
+    if (size != unknown_size)
+      cluster_size = size;
+
+    break;
+  }
+
+  assert(off_next > 0);  // have cluster
+
+  // We have parsed the next cluster.
+  // We have not created a cluster object yet.  What we need
+  // to do now is determine whether it has already be preloaded
+  //(in which case, an object for this cluster has already been
+  // created), and if not, create a new cluster object.
+
+  Cluster** const ii = m_clusters + m_clusterCount;
+  Cluster** i = ii;
+
+  Cluster** const jj = ii + m_clusterPreloadCount;
+  Cluster** j = jj;
+
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) < pos_next
+    //[i, j) ?
+    //[j, jj)  > pos_next
+
+    Cluster** const k = i + (j - i) / 2;
+    assert(k < jj);
+
+    const Cluster* const pNext = *k;
+    assert(pNext);
+    assert(pNext->m_index < 0);
+
+    pos = pNext->GetPosition();
+    assert(pos >= 0);
+
+    if (pos < off_next)
+      i = k + 1;
+    else if (pos > off_next)
+      j = k;
+    else {
+      pResult = pNext;
+      return 0;  // success
+    }
+  }
+
+  assert(i == j);
+
+  long long pos_;
+  long len_;
+
+  status = Cluster::HasBlockEntries(this, off_next, pos_, len_);
+
+  if (status < 0) {  // error or underflow
+    pos = pos_;
+    len = len_;
+
+    return status;
+  }
+
+  if (status > 0) {  // means "found at least one block entry"
+    Cluster* const pNext = Cluster::Create(this,
+                                           -1,  // preloaded
+                                           off_next);
+    if (pNext == NULL)
+      return -1;
+
+    const ptrdiff_t idx_next = i - m_clusters;  // insertion position
+
+    if (!PreloadCluster(pNext, idx_next)) {
+      delete pNext;
+      return -1;
+    }
+    assert(m_clusters);
+    assert(idx_next < m_clusterSize);
+    assert(m_clusters[idx_next] == pNext);
+
+    pResult = pNext;
+    return 0;  // success
+  }
+
+  // status == 0 means "no block entries found"
+
+  if (cluster_size < 0) {  // unknown size
+    const long long payload_pos = pos;  // absolute pos of cluster payload
+
+    for (;;) {  // determine cluster size
+      if ((total >= 0) && (pos >= total))
+        break;
+
+      if ((segment_stop >= 0) && (pos >= segment_stop))
+        break;  // no more clusters
+
+      // Read ID
+
+      if ((pos + 1) > avail) {
+        len = 1;
+        return E_BUFFER_NOT_FULL;
+      }
+
+      long long result = GetUIntLength(m_pReader, pos, len);
+
+      if (result < 0)  // error
+        return static_cast<long>(result);
+
+      if (result > 0)  // weird
+        return E_BUFFER_NOT_FULL;
+
+      if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      if ((pos + len) > avail)
+        return E_BUFFER_NOT_FULL;
+
+      const long long idpos = pos;
+      const long long id = ReadID(m_pReader, idpos, len);
+
+      if (id < 0)  // error (or underflow)
+        return static_cast<long>(id);
+
+      // This is the distinguished set of ID's we use to determine
+      // that we have exhausted the sub-element's inside the cluster
+      // whose ID we parsed earlier.
+
+      if (id == libwebm::kMkvCluster || id == libwebm::kMkvCues)
+        break;
+
+      pos += len;  // consume ID (of sub-element)
+
+      // Read Size
+
+      if ((pos + 1) > avail) {
+        len = 1;
+        return E_BUFFER_NOT_FULL;
+      }
+
+      result = GetUIntLength(m_pReader, pos, len);
+
+      if (result < 0)  // error
+        return static_cast<long>(result);
+
+      if (result > 0)  // weird
+        return E_BUFFER_NOT_FULL;
+
+      if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      if ((pos + len) > avail)
+        return E_BUFFER_NOT_FULL;
+
+      const long long size = ReadUInt(m_pReader, pos, len);
+
+      if (size < 0)  // error
+        return static_cast<long>(size);
+
+      pos += len;  // consume size field of element
+
+      // pos now points to start of sub-element's payload
+
+      if (size == 0)  // weird
+        continue;
+
+      const long long unknown_size = (1LL << (7 * len)) - 1;
+
+      if (size == unknown_size)
+        return E_FILE_FORMAT_INVALID;  // not allowed for sub-elements
+
+      if ((segment_stop >= 0) && ((pos + size) > segment_stop))  // weird
+        return E_FILE_FORMAT_INVALID;
+
+      pos += size;  // consume payload of sub-element
+      if (segment_stop >= 0 && pos > segment_stop)
+        return E_FILE_FORMAT_INVALID;
+    }  // determine cluster size
+
+    cluster_size = pos - payload_pos;
+    assert(cluster_size >= 0);  // TODO: handle cluster_size = 0
+
+    pos = payload_pos;  // reset and re-parse original cluster
+  }
+
+  pos += cluster_size;  // consume payload
+  if (segment_stop >= 0 && pos > segment_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 2;  // try to find a cluster that follows next
+}
+
+const Cluster* Segment::FindCluster(long long time_ns) const {
+  if ((m_clusters == NULL) || (m_clusterCount <= 0))
+    return &m_eos;
+
+  {
+    Cluster* const pCluster = m_clusters[0];
+    assert(pCluster);
+    assert(pCluster->m_index == 0);
+
+    if (time_ns <= pCluster->GetTime())
+      return pCluster;
+  }
+
+  // Binary search of cluster array
+
+  long i = 0;
+  long j = m_clusterCount;
+
+  while (i < j) {
+    // INVARIANT:
+    //[0, i) <= time_ns
+    //[i, j) ?
+    //[j, m_clusterCount)  > time_ns
+
+    const long k = i + (j - i) / 2;
+    assert(k < m_clusterCount);
+
+    Cluster* const pCluster = m_clusters[k];
+    assert(pCluster);
+    assert(pCluster->m_index == k);
+
+    const long long t = pCluster->GetTime();
+
+    if (t <= time_ns)
+      i = k + 1;
+    else
+      j = k;
+
+    assert(i <= j);
+  }
+
+  assert(i == j);
+  assert(i > 0);
+  assert(i <= m_clusterCount);
+
+  const long k = i - 1;
+
+  Cluster* const pCluster = m_clusters[k];
+  assert(pCluster);
+  assert(pCluster->m_index == k);
+  assert(pCluster->GetTime() <= time_ns);
+
+  return pCluster;
+}
+
+const Tracks* Segment::GetTracks() const { return m_pTracks; }
+const SegmentInfo* Segment::GetInfo() const { return m_pInfo; }
+const Cues* Segment::GetCues() const { return m_pCues; }
+const Chapters* Segment::GetChapters() const { return m_pChapters; }
+const Tags* Segment::GetTags() const { return m_pTags; }
+const SeekHead* Segment::GetSeekHead() const { return m_pSeekHead; }
+
+long long Segment::GetDuration() const {
+  assert(m_pInfo);
+  return m_pInfo->GetDuration();
+}
+
+Chapters::Chapters(Segment* pSegment, long long payload_start,
+                   long long payload_size, long long element_start,
+                   long long element_size)
+    : m_pSegment(pSegment),
+      m_start(payload_start),
+      m_size(payload_size),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_editions(NULL),
+      m_editions_size(0),
+      m_editions_count(0) {}
+
+Chapters::~Chapters() {
+  while (m_editions_count > 0) {
+    Edition& e = m_editions[--m_editions_count];
+    e.Clear();
+  }
+  delete[] m_editions;
+}
+
+long Chapters::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;  // payload start
+  const long long stop = pos + m_size;  // payload stop
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvEditionEntry) {
+      status = ParseEdition(pos, size);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+int Chapters::GetEditionCount() const { return m_editions_count; }
+
+const Chapters::Edition* Chapters::GetEdition(int idx) const {
+  if (idx < 0)
+    return NULL;
+
+  if (idx >= m_editions_count)
+    return NULL;
+
+  return m_editions + idx;
+}
+
+bool Chapters::ExpandEditionsArray() {
+  if (m_editions_size > m_editions_count)
+    return true;  // nothing else to do
+
+  const int size = (m_editions_size == 0) ? 1 : 2 * m_editions_size;
+
+  Edition* const editions = new (std::nothrow) Edition[size];
+
+  if (editions == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_editions_count; ++idx) {
+    m_editions[idx].ShallowCopy(editions[idx]);
+  }
+
+  delete[] m_editions;
+  m_editions = editions;
+
+  m_editions_size = size;
+  return true;
+}
+
+long Chapters::ParseEdition(long long pos, long long size) {
+  if (!ExpandEditionsArray())
+    return -1;
+
+  Edition& e = m_editions[m_editions_count++];
+  e.Init();
+
+  return e.Parse(m_pSegment->m_pReader, pos, size);
+}
+
+Chapters::Edition::Edition() {}
+
+Chapters::Edition::~Edition() {}
+
+int Chapters::Edition::GetAtomCount() const { return m_atoms_count; }
+
+const Chapters::Atom* Chapters::Edition::GetAtom(int index) const {
+  if (index < 0)
+    return NULL;
+
+  if (index >= m_atoms_count)
+    return NULL;
+
+  return m_atoms + index;
+}
+
+void Chapters::Edition::Init() {
+  m_atoms = NULL;
+  m_atoms_size = 0;
+  m_atoms_count = 0;
+}
+
+void Chapters::Edition::ShallowCopy(Edition& rhs) const {
+  rhs.m_atoms = m_atoms;
+  rhs.m_atoms_size = m_atoms_size;
+  rhs.m_atoms_count = m_atoms_count;
+}
+
+void Chapters::Edition::Clear() {
+  while (m_atoms_count > 0) {
+    Atom& a = m_atoms[--m_atoms_count];
+    a.Clear();
+  }
+
+  delete[] m_atoms;
+  m_atoms = NULL;
+
+  m_atoms_size = 0;
+}
+
+long Chapters::Edition::Parse(IMkvReader* pReader, long long pos,
+                              long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)
+      continue;
+
+    if (id == libwebm::kMkvChapterAtom) {
+      status = ParseAtom(pReader, pos, size);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long Chapters::Edition::ParseAtom(IMkvReader* pReader, long long pos,
+                                  long long size) {
+  if (!ExpandAtomsArray())
+    return -1;
+
+  Atom& a = m_atoms[m_atoms_count++];
+  a.Init();
+
+  return a.Parse(pReader, pos, size);
+}
+
+bool Chapters::Edition::ExpandAtomsArray() {
+  if (m_atoms_size > m_atoms_count)
+    return true;  // nothing else to do
+
+  const int size = (m_atoms_size == 0) ? 1 : 2 * m_atoms_size;
+
+  Atom* const atoms = new (std::nothrow) Atom[size];
+
+  if (atoms == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_atoms_count; ++idx) {
+    m_atoms[idx].ShallowCopy(atoms[idx]);
+  }
+
+  delete[] m_atoms;
+  m_atoms = atoms;
+
+  m_atoms_size = size;
+  return true;
+}
+
+Chapters::Atom::Atom() {}
+
+Chapters::Atom::~Atom() {}
+
+unsigned long long Chapters::Atom::GetUID() const { return m_uid; }
+
+const char* Chapters::Atom::GetStringUID() const { return m_string_uid; }
+
+long long Chapters::Atom::GetStartTimecode() const { return m_start_timecode; }
+
+long long Chapters::Atom::GetStopTimecode() const { return m_stop_timecode; }
+
+long long Chapters::Atom::GetStartTime(const Chapters* pChapters) const {
+  return GetTime(pChapters, m_start_timecode);
+}
+
+long long Chapters::Atom::GetStopTime(const Chapters* pChapters) const {
+  return GetTime(pChapters, m_stop_timecode);
+}
+
+int Chapters::Atom::GetDisplayCount() const { return m_displays_count; }
+
+const Chapters::Display* Chapters::Atom::GetDisplay(int index) const {
+  if (index < 0)
+    return NULL;
+
+  if (index >= m_displays_count)
+    return NULL;
+
+  return m_displays + index;
+}
+
+void Chapters::Atom::Init() {
+  m_string_uid = NULL;
+  m_uid = 0;
+  m_start_timecode = -1;
+  m_stop_timecode = -1;
+
+  m_displays = NULL;
+  m_displays_size = 0;
+  m_displays_count = 0;
+}
+
+void Chapters::Atom::ShallowCopy(Atom& rhs) const {
+  rhs.m_string_uid = m_string_uid;
+  rhs.m_uid = m_uid;
+  rhs.m_start_timecode = m_start_timecode;
+  rhs.m_stop_timecode = m_stop_timecode;
+
+  rhs.m_displays = m_displays;
+  rhs.m_displays_size = m_displays_size;
+  rhs.m_displays_count = m_displays_count;
+}
+
+void Chapters::Atom::Clear() {
+  delete[] m_string_uid;
+  m_string_uid = NULL;
+
+  while (m_displays_count > 0) {
+    Display& d = m_displays[--m_displays_count];
+    d.Clear();
+  }
+
+  delete[] m_displays;
+  m_displays = NULL;
+
+  m_displays_size = 0;
+}
+
+long Chapters::Atom::Parse(IMkvReader* pReader, long long pos, long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // 0 length payload, skip.
+      continue;
+
+    if (id == libwebm::kMkvChapterDisplay) {
+      status = ParseDisplay(pReader, pos, size);
+
+      if (status < 0)  // error
+        return status;
+    } else if (id == libwebm::kMkvChapterStringUID) {
+      status = UnserializeString(pReader, pos, size, m_string_uid);
+
+      if (status < 0)  // error
+        return status;
+    } else if (id == libwebm::kMkvChapterUID) {
+      long long val;
+      status = UnserializeInt(pReader, pos, size, val);
+
+      if (status < 0)  // error
+        return status;
+
+      m_uid = static_cast<unsigned long long>(val);
+    } else if (id == libwebm::kMkvChapterTimeStart) {
+      const long long val = UnserializeUInt(pReader, pos, size);
+
+      if (val < 0)  // error
+        return static_cast<long>(val);
+
+      m_start_timecode = val;
+    } else if (id == libwebm::kMkvChapterTimeEnd) {
+      const long long val = UnserializeUInt(pReader, pos, size);
+
+      if (val < 0)  // error
+        return static_cast<long>(val);
+
+      m_stop_timecode = val;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long long Chapters::Atom::GetTime(const Chapters* pChapters,
+                                  long long timecode) {
+  if (pChapters == NULL)
+    return -1;
+
+  Segment* const pSegment = pChapters->m_pSegment;
+
+  if (pSegment == NULL)  // weird
+    return -1;
+
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+
+  if (pInfo == NULL)
+    return -1;
+
+  const long long timecode_scale = pInfo->GetTimeCodeScale();
+
+  if (timecode_scale < 1)  // weird
+    return -1;
+
+  if (timecode < 0)
+    return -1;
+
+  const long long result = timecode_scale * timecode;
+
+  return result;
+}
+
+long Chapters::Atom::ParseDisplay(IMkvReader* pReader, long long pos,
+                                  long long size) {
+  if (!ExpandDisplaysArray())
+    return -1;
+
+  Display& d = m_displays[m_displays_count++];
+  d.Init();
+
+  return d.Parse(pReader, pos, size);
+}
+
+bool Chapters::Atom::ExpandDisplaysArray() {
+  if (m_displays_size > m_displays_count)
+    return true;  // nothing else to do
+
+  const int size = (m_displays_size == 0) ? 1 : 2 * m_displays_size;
+
+  Display* const displays = new (std::nothrow) Display[size];
+
+  if (displays == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_displays_count; ++idx) {
+    m_displays[idx].ShallowCopy(displays[idx]);
+  }
+
+  delete[] m_displays;
+  m_displays = displays;
+
+  m_displays_size = size;
+  return true;
+}
+
+Chapters::Display::Display() {}
+
+Chapters::Display::~Display() {}
+
+const char* Chapters::Display::GetString() const { return m_string; }
+
+const char* Chapters::Display::GetLanguage() const { return m_language; }
+
+const char* Chapters::Display::GetCountry() const { return m_country; }
+
+void Chapters::Display::Init() {
+  m_string = NULL;
+  m_language = NULL;
+  m_country = NULL;
+}
+
+void Chapters::Display::ShallowCopy(Display& rhs) const {
+  rhs.m_string = m_string;
+  rhs.m_language = m_language;
+  rhs.m_country = m_country;
+}
+
+void Chapters::Display::Clear() {
+  delete[] m_string;
+  m_string = NULL;
+
+  delete[] m_language;
+  m_language = NULL;
+
+  delete[] m_country;
+  m_country = NULL;
+}
+
+long Chapters::Display::Parse(IMkvReader* pReader, long long pos,
+                              long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // No payload.
+      continue;
+
+    if (id == libwebm::kMkvChapString) {
+      status = UnserializeString(pReader, pos, size, m_string);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvChapLanguage) {
+      status = UnserializeString(pReader, pos, size, m_language);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvChapCountry) {
+      status = UnserializeString(pReader, pos, size, m_country);
+
+      if (status)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+Tags::Tags(Segment* pSegment, long long payload_start, long long payload_size,
+           long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(payload_start),
+      m_size(payload_size),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_tags(NULL),
+      m_tags_size(0),
+      m_tags_count(0) {}
+
+Tags::~Tags() {
+  while (m_tags_count > 0) {
+    Tag& t = m_tags[--m_tags_count];
+    t.Clear();
+  }
+  delete[] m_tags;
+}
+
+long Tags::Parse() {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;  // payload start
+  const long long stop = pos + m_size;  // payload stop
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)
+      return status;
+
+    if (size == 0)  // 0 length tag, read another
+      continue;
+
+    if (id == libwebm::kMkvTag) {
+      status = ParseTag(pos, size);
+
+      if (status < 0)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+int Tags::GetTagCount() const { return m_tags_count; }
+
+const Tags::Tag* Tags::GetTag(int idx) const {
+  if (idx < 0)
+    return NULL;
+
+  if (idx >= m_tags_count)
+    return NULL;
+
+  return m_tags + idx;
+}
+
+bool Tags::ExpandTagsArray() {
+  if (m_tags_size > m_tags_count)
+    return true;  // nothing else to do
+
+  const int size = (m_tags_size == 0) ? 1 : 2 * m_tags_size;
+
+  Tag* const tags = new (std::nothrow) Tag[size];
+
+  if (tags == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_tags_count; ++idx) {
+    m_tags[idx].ShallowCopy(tags[idx]);
+  }
+
+  delete[] m_tags;
+  m_tags = tags;
+
+  m_tags_size = size;
+  return true;
+}
+
+long Tags::ParseTag(long long pos, long long size) {
+  if (!ExpandTagsArray())
+    return -1;
+
+  Tag& t = m_tags[m_tags_count++];
+  t.Init();
+
+  return t.Parse(m_pSegment->m_pReader, pos, size);
+}
+
+Tags::Tag::Tag() {}
+
+Tags::Tag::~Tag() {}
+
+int Tags::Tag::GetSimpleTagCount() const { return m_simple_tags_count; }
+
+const Tags::SimpleTag* Tags::Tag::GetSimpleTag(int index) const {
+  if (index < 0)
+    return NULL;
+
+  if (index >= m_simple_tags_count)
+    return NULL;
+
+  return m_simple_tags + index;
+}
+
+void Tags::Tag::Init() {
+  m_simple_tags = NULL;
+  m_simple_tags_size = 0;
+  m_simple_tags_count = 0;
+}
+
+void Tags::Tag::ShallowCopy(Tag& rhs) const {
+  rhs.m_simple_tags = m_simple_tags;
+  rhs.m_simple_tags_size = m_simple_tags_size;
+  rhs.m_simple_tags_count = m_simple_tags_count;
+}
+
+void Tags::Tag::Clear() {
+  while (m_simple_tags_count > 0) {
+    SimpleTag& d = m_simple_tags[--m_simple_tags_count];
+    d.Clear();
+  }
+
+  delete[] m_simple_tags;
+  m_simple_tags = NULL;
+
+  m_simple_tags_size = 0;
+}
+
+long Tags::Tag::Parse(IMkvReader* pReader, long long pos, long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)
+      return status;
+
+    if (size == 0)  // 0 length tag, read another
+      continue;
+
+    if (id == libwebm::kMkvSimpleTag) {
+      status = ParseSimpleTag(pReader, pos, size);
+
+      if (status < 0)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long Tags::Tag::ParseSimpleTag(IMkvReader* pReader, long long pos,
+                               long long size) {
+  if (!ExpandSimpleTagsArray())
+    return -1;
+
+  SimpleTag& st = m_simple_tags[m_simple_tags_count++];
+  st.Init();
+
+  return st.Parse(pReader, pos, size);
+}
+
+bool Tags::Tag::ExpandSimpleTagsArray() {
+  if (m_simple_tags_size > m_simple_tags_count)
+    return true;  // nothing else to do
+
+  const int size = (m_simple_tags_size == 0) ? 1 : 2 * m_simple_tags_size;
+
+  SimpleTag* const displays = new (std::nothrow) SimpleTag[size];
+
+  if (displays == NULL)
+    return false;
+
+  for (int idx = 0; idx < m_simple_tags_count; ++idx) {
+    m_simple_tags[idx].ShallowCopy(displays[idx]);
+  }
+
+  delete[] m_simple_tags;
+  m_simple_tags = displays;
+
+  m_simple_tags_size = size;
+  return true;
+}
+
+Tags::SimpleTag::SimpleTag() {}
+
+Tags::SimpleTag::~SimpleTag() {}
+
+const char* Tags::SimpleTag::GetTagName() const { return m_tag_name; }
+
+const char* Tags::SimpleTag::GetTagString() const { return m_tag_string; }
+
+void Tags::SimpleTag::Init() {
+  m_tag_name = NULL;
+  m_tag_string = NULL;
+}
+
+void Tags::SimpleTag::ShallowCopy(SimpleTag& rhs) const {
+  rhs.m_tag_name = m_tag_name;
+  rhs.m_tag_string = m_tag_string;
+}
+
+void Tags::SimpleTag::Clear() {
+  delete[] m_tag_name;
+  m_tag_name = NULL;
+
+  delete[] m_tag_string;
+  m_tag_string = NULL;
+}
+
+long Tags::SimpleTag::Parse(IMkvReader* pReader, long long pos,
+                            long long size) {
+  const long long stop = pos + size;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvTagName) {
+      status = UnserializeString(pReader, pos, size, m_tag_name);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvTagString) {
+      status = UnserializeString(pReader, pos, size, m_tag_string);
+
+      if (status)
+        return status;
+    }
+
+    pos += size;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+SegmentInfo::SegmentInfo(Segment* pSegment, long long start, long long size_,
+                         long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_pMuxingAppAsUTF8(NULL),
+      m_pWritingAppAsUTF8(NULL),
+      m_pTitleAsUTF8(NULL) {}
+
+SegmentInfo::~SegmentInfo() {
+  delete[] m_pMuxingAppAsUTF8;
+  m_pMuxingAppAsUTF8 = NULL;
+
+  delete[] m_pWritingAppAsUTF8;
+  m_pWritingAppAsUTF8 = NULL;
+
+  delete[] m_pTitleAsUTF8;
+  m_pTitleAsUTF8 = NULL;
+}
+
+long SegmentInfo::Parse() {
+  assert(m_pMuxingAppAsUTF8 == NULL);
+  assert(m_pWritingAppAsUTF8 == NULL);
+  assert(m_pTitleAsUTF8 == NULL);
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
+
+  m_timecodeScale = 1000000;
+  m_duration = -1;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvTimecodeScale) {
+      m_timecodeScale = UnserializeUInt(pReader, pos, size);
+
+      if (m_timecodeScale <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDuration) {
+      const long status = UnserializeFloat(pReader, pos, size, m_duration);
+
+      if (status < 0)
+        return status;
+
+      if (m_duration < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvMuxingApp) {
+      const long status =
+          UnserializeString(pReader, pos, size, m_pMuxingAppAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvWritingApp) {
+      const long status =
+          UnserializeString(pReader, pos, size, m_pWritingAppAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvTitle) {
+      const long status = UnserializeString(pReader, pos, size, m_pTitleAsUTF8);
+
+      if (status)
+        return status;
+    }
+
+    pos += size;
+
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  const double rollover_check = m_duration * m_timecodeScale;
+  if (rollover_check > LLONG_MAX)
+    return E_FILE_FORMAT_INVALID;
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+long long SegmentInfo::GetTimeCodeScale() const { return m_timecodeScale; }
+
+long long SegmentInfo::GetDuration() const {
+  if (m_duration < 0)
+    return -1;
+
+  assert(m_timecodeScale >= 1);
+
+  const double dd = double(m_duration) * double(m_timecodeScale);
+  const long long d = static_cast<long long>(dd);
+
+  return d;
+}
+
+const char* SegmentInfo::GetMuxingAppAsUTF8() const {
+  return m_pMuxingAppAsUTF8;
+}
+
+const char* SegmentInfo::GetWritingAppAsUTF8() const {
+  return m_pWritingAppAsUTF8;
+}
+
+const char* SegmentInfo::GetTitleAsUTF8() const { return m_pTitleAsUTF8; }
+
+///////////////////////////////////////////////////////////////
+// ContentEncoding element
+ContentEncoding::ContentCompression::ContentCompression()
+    : algo(0), settings(NULL), settings_len(0) {}
+
+ContentEncoding::ContentCompression::~ContentCompression() {
+  delete[] settings;
+}
+
+ContentEncoding::ContentEncryption::ContentEncryption()
+    : algo(0),
+      key_id(NULL),
+      key_id_len(0),
+      signature(NULL),
+      signature_len(0),
+      sig_key_id(NULL),
+      sig_key_id_len(0),
+      sig_algo(0),
+      sig_hash_algo(0) {}
+
+ContentEncoding::ContentEncryption::~ContentEncryption() {
+  delete[] key_id;
+  delete[] signature;
+  delete[] sig_key_id;
+}
+
+ContentEncoding::ContentEncoding()
+    : compression_entries_(NULL),
+      compression_entries_end_(NULL),
+      encryption_entries_(NULL),
+      encryption_entries_end_(NULL),
+      encoding_order_(0),
+      encoding_scope_(1),
+      encoding_type_(0) {}
+
+ContentEncoding::~ContentEncoding() {
+  ContentCompression** comp_i = compression_entries_;
+  ContentCompression** const comp_j = compression_entries_end_;
+
+  while (comp_i != comp_j) {
+    ContentCompression* const comp = *comp_i++;
+    delete comp;
+  }
+
+  delete[] compression_entries_;
+
+  ContentEncryption** enc_i = encryption_entries_;
+  ContentEncryption** const enc_j = encryption_entries_end_;
+
+  while (enc_i != enc_j) {
+    ContentEncryption* const enc = *enc_i++;
+    delete enc;
+  }
+
+  delete[] encryption_entries_;
+}
+
+const ContentEncoding::ContentCompression*
+    ContentEncoding::GetCompressionByIndex(unsigned long idx) const {
+  const ptrdiff_t count = compression_entries_end_ - compression_entries_;
+  assert(count >= 0);
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return compression_entries_[idx];
+}
+
+unsigned long ContentEncoding::GetCompressionCount() const {
+  const ptrdiff_t count = compression_entries_end_ - compression_entries_;
+  assert(count >= 0);
+
+  return static_cast<unsigned long>(count);
+}
+
+const ContentEncoding::ContentEncryption* ContentEncoding::GetEncryptionByIndex(
+    unsigned long idx) const {
+  const ptrdiff_t count = encryption_entries_end_ - encryption_entries_;
+  assert(count >= 0);
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return encryption_entries_[idx];
+}
+
+unsigned long ContentEncoding::GetEncryptionCount() const {
+  const ptrdiff_t count = encryption_entries_end_ - encryption_entries_;
+  assert(count >= 0);
+
+  return static_cast<unsigned long>(count);
+}
+
+long ContentEncoding::ParseContentEncAESSettingsEntry(
+    long long start, long long size, IMkvReader* pReader,
+    ContentEncAESSettings* aes) {
+  assert(pReader);
+  assert(aes);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvAESSettingsCipherMode) {
+      aes->cipher_mode = UnserializeUInt(pReader, pos, size);
+      if (aes->cipher_mode != 1)
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  return 0;
+}
+
+long ContentEncoding::ParseContentEncodingEntry(long long start, long long size,
+                                                IMkvReader* pReader) {
+  assert(pReader);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  // Count ContentCompression and ContentEncryption elements.
+  int compression_count = 0;
+  int encryption_count = 0;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentCompression)
+      ++compression_count;
+
+    if (id == libwebm::kMkvContentEncryption)
+      ++encryption_count;
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (compression_count <= 0 && encryption_count <= 0)
+    return -1;
+
+  if (compression_count > 0) {
+    compression_entries_ =
+        new (std::nothrow) ContentCompression*[compression_count];
+    if (!compression_entries_)
+      return -1;
+    compression_entries_end_ = compression_entries_;
+  }
+
+  if (encryption_count > 0) {
+    encryption_entries_ =
+        new (std::nothrow) ContentEncryption*[encryption_count];
+    if (!encryption_entries_) {
+      delete[] compression_entries_;
+      return -1;
+    }
+    encryption_entries_end_ = encryption_entries_;
+  }
+
+  pos = start;
+  while (pos < stop) {
+    long long id, size;
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentEncodingOrder) {
+      encoding_order_ = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentEncodingScope) {
+      encoding_scope_ = UnserializeUInt(pReader, pos, size);
+      if (encoding_scope_ < 1)
+        return -1;
+    } else if (id == libwebm::kMkvContentEncodingType) {
+      encoding_type_ = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentCompression) {
+      ContentCompression* const compression =
+          new (std::nothrow) ContentCompression();
+      if (!compression)
+        return -1;
+
+      status = ParseCompressionEntry(pos, size, pReader, compression);
+      if (status) {
+        delete compression;
+        return status;
+      }
+      *compression_entries_end_++ = compression;
+    } else if (id == libwebm::kMkvContentEncryption) {
+      ContentEncryption* const encryption =
+          new (std::nothrow) ContentEncryption();
+      if (!encryption)
+        return -1;
+
+      status = ParseEncryptionEntry(pos, size, pReader, encryption);
+      if (status) {
+        delete encryption;
+        return status;
+      }
+      *encryption_entries_end_++ = encryption;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  return 0;
+}
+
+long ContentEncoding::ParseCompressionEntry(long long start, long long size,
+                                            IMkvReader* pReader,
+                                            ContentCompression* compression) {
+  assert(pReader);
+  assert(compression);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  bool valid = false;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentCompAlgo) {
+      long long algo = UnserializeUInt(pReader, pos, size);
+      if (algo < 0)
+        return E_FILE_FORMAT_INVALID;
+      compression->algo = algo;
+      valid = true;
+    } else if (id == libwebm::kMkvContentCompSettings) {
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      compression->settings = buf;
+      compression->settings_len = buflen;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  // ContentCompAlgo is mandatory
+  if (!valid)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+long ContentEncoding::ParseEncryptionEntry(long long start, long long size,
+                                           IMkvReader* pReader,
+                                           ContentEncryption* encryption) {
+  assert(pReader);
+  assert(encryption);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvContentEncAlgo) {
+      encryption->algo = UnserializeUInt(pReader, pos, size);
+      if (encryption->algo != 5)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvContentEncKeyID) {
+      delete[] encryption->key_id;
+      encryption->key_id = NULL;
+      encryption->key_id_len = 0;
+
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      encryption->key_id = buf;
+      encryption->key_id_len = buflen;
+    } else if (id == libwebm::kMkvContentSignature) {
+      delete[] encryption->signature;
+      encryption->signature = NULL;
+      encryption->signature_len = 0;
+
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      encryption->signature = buf;
+      encryption->signature_len = buflen;
+    } else if (id == libwebm::kMkvContentSigKeyID) {
+      delete[] encryption->sig_key_id;
+      encryption->sig_key_id = NULL;
+      encryption->sig_key_id_len = 0;
+
+      if (size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      const size_t buflen = static_cast<size_t>(size);
+      unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+      if (buf == NULL)
+        return -1;
+
+      const int read_status =
+          pReader->Read(pos, static_cast<long>(buflen), buf);
+      if (read_status) {
+        delete[] buf;
+        return status;
+      }
+
+      encryption->sig_key_id = buf;
+      encryption->sig_key_id_len = buflen;
+    } else if (id == libwebm::kMkvContentSigAlgo) {
+      encryption->sig_algo = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentSigHashAlgo) {
+      encryption->sig_hash_algo = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvContentEncAESSettings) {
+      const long status = ParseContentEncAESSettingsEntry(
+          pos, size, pReader, &encryption->aes_settings);
+      if (status)
+        return status;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  return 0;
+}
+
+Track::Track(Segment* pSegment, long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      content_encoding_entries_(NULL),
+      content_encoding_entries_end_(NULL) {}
+
+Track::~Track() {
+  Info& info = const_cast<Info&>(m_info);
+  info.Clear();
+
+  ContentEncoding** i = content_encoding_entries_;
+  ContentEncoding** const j = content_encoding_entries_end_;
+
+  while (i != j) {
+    ContentEncoding* const encoding = *i++;
+    delete encoding;
+  }
+
+  delete[] content_encoding_entries_;
+}
+
+long Track::Create(Segment* pSegment, const Info& info, long long element_start,
+                   long long element_size, Track*& pResult) {
+  if (pResult)
+    return -1;
+
+  Track* const pTrack =
+      new (std::nothrow) Track(pSegment, element_start, element_size);
+
+  if (pTrack == NULL)
+    return -1;  // generic error
+
+  const int status = info.Copy(pTrack->m_info);
+
+  if (status) {  // error
+    delete pTrack;
+    return status;
+  }
+
+  pResult = pTrack;
+  return 0;  // success
+}
+
+Track::Info::Info()
+    : uid(0),
+      defaultDuration(0),
+      codecDelay(0),
+      seekPreRoll(0),
+      nameAsUTF8(NULL),
+      language(NULL),
+      codecId(NULL),
+      codecNameAsUTF8(NULL),
+      codecPrivate(NULL),
+      codecPrivateSize(0),
+      lacing(false) {}
+
+Track::Info::~Info() { Clear(); }
+
+void Track::Info::Clear() {
+  delete[] nameAsUTF8;
+  nameAsUTF8 = NULL;
+
+  delete[] language;
+  language = NULL;
+
+  delete[] codecId;
+  codecId = NULL;
+
+  delete[] codecPrivate;
+  codecPrivate = NULL;
+  codecPrivateSize = 0;
+
+  delete[] codecNameAsUTF8;
+  codecNameAsUTF8 = NULL;
+}
+
+int Track::Info::CopyStr(char* Info::*str, Info& dst_) const {
+  if (str == static_cast<char * Info::*>(NULL))
+    return -1;
+
+  char*& dst = dst_.*str;
+
+  if (dst)  // should be NULL already
+    return -1;
+
+  const char* const src = this->*str;
+
+  if (src == NULL)
+    return 0;
+
+  const size_t len = strlen(src);
+
+  dst = SafeArrayAlloc<char>(1, len + 1);
+
+  if (dst == NULL)
+    return -1;
+
+  strcpy(dst, src);
+
+  return 0;
+}
+
+int Track::Info::Copy(Info& dst) const {
+  if (&dst == this)
+    return 0;
+
+  dst.type = type;
+  dst.number = number;
+  dst.defaultDuration = defaultDuration;
+  dst.codecDelay = codecDelay;
+  dst.seekPreRoll = seekPreRoll;
+  dst.uid = uid;
+  dst.lacing = lacing;
+  dst.settings = settings;
+
+  // We now copy the string member variables from src to dst.
+  // This involves memory allocation so in principle the operation
+  // can fail (indeed, that's why we have Info::Copy), so we must
+  // report this to the caller.  An error return from this function
+  // therefore implies that the copy was only partially successful.
+
+  if (int status = CopyStr(&Info::nameAsUTF8, dst))
+    return status;
+
+  if (int status = CopyStr(&Info::language, dst))
+    return status;
+
+  if (int status = CopyStr(&Info::codecId, dst))
+    return status;
+
+  if (int status = CopyStr(&Info::codecNameAsUTF8, dst))
+    return status;
+
+  if (codecPrivateSize > 0) {
+    if (codecPrivate == NULL)
+      return -1;
+
+    if (dst.codecPrivate)
+      return -1;
+
+    if (dst.codecPrivateSize != 0)
+      return -1;
+
+    dst.codecPrivate = SafeArrayAlloc<unsigned char>(1, codecPrivateSize);
+
+    if (dst.codecPrivate == NULL)
+      return -1;
+
+    memcpy(dst.codecPrivate, codecPrivate, codecPrivateSize);
+    dst.codecPrivateSize = codecPrivateSize;
+  }
+
+  return 0;
+}
+
+const BlockEntry* Track::GetEOS() const { return &m_eos; }
+
+long Track::GetType() const { return m_info.type; }
+
+long Track::GetNumber() const { return m_info.number; }
+
+unsigned long long Track::GetUid() const { return m_info.uid; }
+
+const char* Track::GetNameAsUTF8() const { return m_info.nameAsUTF8; }
+
+const char* Track::GetLanguage() const { return m_info.language; }
+
+const char* Track::GetCodecNameAsUTF8() const { return m_info.codecNameAsUTF8; }
+
+const char* Track::GetCodecId() const { return m_info.codecId; }
+
+const unsigned char* Track::GetCodecPrivate(size_t& size) const {
+  size = m_info.codecPrivateSize;
+  return m_info.codecPrivate;
+}
+
+bool Track::GetLacing() const { return m_info.lacing; }
+
+unsigned long long Track::GetDefaultDuration() const {
+  return m_info.defaultDuration;
+}
+
+unsigned long long Track::GetCodecDelay() const { return m_info.codecDelay; }
+
+unsigned long long Track::GetSeekPreRoll() const { return m_info.seekPreRoll; }
+
+long Track::GetFirst(const BlockEntry*& pBlockEntry) const {
+  const Cluster* pCluster = m_pSegment->GetFirst();
+
+  for (int i = 0;;) {
+    if (pCluster == NULL) {
+      pBlockEntry = GetEOS();
+      return 1;
+    }
+
+    if (pCluster->EOS()) {
+      if (m_pSegment->DoneParsing()) {
+        pBlockEntry = GetEOS();
+        return 1;
+      }
+
+      pBlockEntry = 0;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long status = pCluster->GetFirst(pBlockEntry);
+
+    if (status < 0)  // error
+      return status;
+
+    if (pBlockEntry == 0) {  // empty cluster
+      pCluster = m_pSegment->GetNext(pCluster);
+      continue;
+    }
+
+    for (;;) {
+      const Block* const pBlock = pBlockEntry->GetBlock();
+      assert(pBlock);
+
+      const long long tn = pBlock->GetTrackNumber();
+
+      if ((tn == m_info.number) && VetEntry(pBlockEntry))
+        return 0;
+
+      const BlockEntry* pNextEntry;
+
+      status = pCluster->GetNext(pBlockEntry, pNextEntry);
+
+      if (status < 0)  // error
+        return status;
+
+      if (pNextEntry == 0)
+        break;
+
+      pBlockEntry = pNextEntry;
+    }
+
+    ++i;
+
+    if (i >= 100)
+      break;
+
+    pCluster = m_pSegment->GetNext(pCluster);
+  }
+
+  // NOTE: if we get here, it means that we didn't find a block with
+  // a matching track number.  We interpret that as an error (which
+  // might be too conservative).
+
+  pBlockEntry = GetEOS();  // so we can return a non-NULL value
+  return 1;
+}
+
+long Track::GetNext(const BlockEntry* pCurrEntry,
+                    const BlockEntry*& pNextEntry) const {
+  assert(pCurrEntry);
+  assert(!pCurrEntry->EOS());  //?
+
+  const Block* const pCurrBlock = pCurrEntry->GetBlock();
+  assert(pCurrBlock && pCurrBlock->GetTrackNumber() == m_info.number);
+  if (!pCurrBlock || pCurrBlock->GetTrackNumber() != m_info.number)
+    return -1;
+
+  const Cluster* pCluster = pCurrEntry->GetCluster();
+  assert(pCluster);
+  assert(!pCluster->EOS());
+
+  long status = pCluster->GetNext(pCurrEntry, pNextEntry);
+
+  if (status < 0)  // error
+    return status;
+
+  for (int i = 0;;) {
+    while (pNextEntry) {
+      const Block* const pNextBlock = pNextEntry->GetBlock();
+      assert(pNextBlock);
+
+      if (pNextBlock->GetTrackNumber() == m_info.number)
+        return 0;
+
+      pCurrEntry = pNextEntry;
+
+      status = pCluster->GetNext(pCurrEntry, pNextEntry);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    pCluster = m_pSegment->GetNext(pCluster);
+
+    if (pCluster == NULL) {
+      pNextEntry = GetEOS();
+      return 1;
+    }
+
+    if (pCluster->EOS()) {
+      if (m_pSegment->DoneParsing()) {
+        pNextEntry = GetEOS();
+        return 1;
+      }
+
+      // TODO: there is a potential O(n^2) problem here: we tell the
+      // caller to (pre)load another cluster, which he does, but then he
+      // calls GetNext again, which repeats the same search.  This is
+      // a pathological case, since the only way it can happen is if
+      // there exists a long sequence of clusters none of which contain a
+      // block from this track.  One way around this problem is for the
+      // caller to be smarter when he loads another cluster: don't call
+      // us back until you have a cluster that contains a block from this
+      // track. (Of course, that's not cheap either, since our caller
+      // would have to scan the each cluster as it's loaded, so that
+      // would just push back the problem.)
+
+      pNextEntry = NULL;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    status = pCluster->GetFirst(pNextEntry);
+
+    if (status < 0)  // error
+      return status;
+
+    if (pNextEntry == NULL)  // empty cluster
+      continue;
+
+    ++i;
+
+    if (i >= 100)
+      break;
+  }
+
+  // NOTE: if we get here, it means that we didn't find a block with
+  // a matching track number after lots of searching, so we give
+  // up trying.
+
+  pNextEntry = GetEOS();  // so we can return a non-NULL value
+  return 1;
+}
+
+bool Track::VetEntry(const BlockEntry* pBlockEntry) const {
+  assert(pBlockEntry);
+  const Block* const pBlock = pBlockEntry->GetBlock();
+  assert(pBlock);
+  assert(pBlock->GetTrackNumber() == m_info.number);
+  if (!pBlock || pBlock->GetTrackNumber() != m_info.number)
+    return false;
+
+  // This function is used during a seek to determine whether the
+  // frame is a valid seek target.  This default function simply
+  // returns true, which means all frames are valid seek targets.
+  // It gets overridden by the VideoTrack class, because only video
+  // keyframes can be used as seek target.
+
+  return true;
+}
+
+long Track::Seek(long long time_ns, const BlockEntry*& pResult) const {
+  const long status = GetFirst(pResult);
+
+  if (status < 0)  // buffer underflow, etc
+    return status;
+
+  assert(pResult);
+
+  if (pResult->EOS())
+    return 0;
+
+  const Cluster* pCluster = pResult->GetCluster();
+  assert(pCluster);
+  assert(pCluster->GetIndex() >= 0);
+
+  if (time_ns <= pResult->GetBlock()->GetTime(pCluster))
+    return 0;
+
+  Cluster** const clusters = m_pSegment->m_clusters;
+  assert(clusters);
+
+  const long count = m_pSegment->GetCount();  // loaded only, not preloaded
+  assert(count > 0);
+
+  Cluster** const i = clusters + pCluster->GetIndex();
+  assert(i);
+  assert(*i == pCluster);
+  assert(pCluster->GetTime() <= time_ns);
+
+  Cluster** const j = clusters + count;
+
+  Cluster** lo = i;
+  Cluster** hi = j;
+
+  while (lo < hi) {
+    // INVARIANT:
+    //[i, lo) <= time_ns
+    //[lo, hi) ?
+    //[hi, j)  > time_ns
+
+    Cluster** const mid = lo + (hi - lo) / 2;
+    assert(mid < hi);
+
+    pCluster = *mid;
+    assert(pCluster);
+    assert(pCluster->GetIndex() >= 0);
+    assert(pCluster->GetIndex() == long(mid - m_pSegment->m_clusters));
+
+    const long long t = pCluster->GetTime();
+
+    if (t <= time_ns)
+      lo = mid + 1;
+    else
+      hi = mid;
+
+    assert(lo <= hi);
+  }
+
+  assert(lo == hi);
+  assert(lo > i);
+  assert(lo <= j);
+
+  while (lo > i) {
+    pCluster = *--lo;
+    assert(pCluster);
+    assert(pCluster->GetTime() <= time_ns);
+
+    pResult = pCluster->GetEntry(this);
+
+    if ((pResult != 0) && !pResult->EOS())
+      return 0;
+
+    // landed on empty cluster (no entries)
+  }
+
+  pResult = GetEOS();  // weird
+  return 0;
+}
+
+const ContentEncoding* Track::GetContentEncodingByIndex(
+    unsigned long idx) const {
+  const ptrdiff_t count =
+      content_encoding_entries_end_ - content_encoding_entries_;
+  assert(count >= 0);
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return content_encoding_entries_[idx];
+}
+
+unsigned long Track::GetContentEncodingCount() const {
+  const ptrdiff_t count =
+      content_encoding_entries_end_ - content_encoding_entries_;
+  assert(count >= 0);
+
+  return static_cast<unsigned long>(count);
+}
+
+long Track::ParseContentEncodingsEntry(long long start, long long size) {
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+  assert(pReader);
+
+  long long pos = start;
+  const long long stop = start + size;
+
+  // Count ContentEncoding elements.
+  int count = 0;
+  while (pos < stop) {
+    long long id, size;
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    // pos now designates start of element
+    if (id == libwebm::kMkvContentEncoding)
+      ++count;
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (count <= 0)
+    return -1;
+
+  content_encoding_entries_ = new (std::nothrow) ContentEncoding*[count];
+  if (!content_encoding_entries_)
+    return -1;
+
+  content_encoding_entries_end_ = content_encoding_entries_;
+
+  pos = start;
+  while (pos < stop) {
+    long long id, size;
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+    if (status < 0)  // error
+      return status;
+
+    // pos now designates start of element
+    if (id == libwebm::kMkvContentEncoding) {
+      ContentEncoding* const content_encoding =
+          new (std::nothrow) ContentEncoding();
+      if (!content_encoding)
+        return -1;
+
+      status = content_encoding->ParseContentEncodingEntry(pos, size, pReader);
+      if (status) {
+        delete content_encoding;
+        return status;
+      }
+
+      *content_encoding_entries_end_++ = content_encoding;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;
+}
+
+Track::EOSBlock::EOSBlock() : BlockEntry(NULL, LONG_MIN) {}
+
+BlockEntry::Kind Track::EOSBlock::GetKind() const { return kBlockEOS; }
+
+const Block* Track::EOSBlock::GetBlock() const { return NULL; }
+
+bool PrimaryChromaticity::Parse(IMkvReader* reader, long long read_pos,
+                                long long value_size, bool is_x,
+                                PrimaryChromaticity** chromaticity) {
+  if (!reader)
+    return false;
+
+  std::auto_ptr<PrimaryChromaticity> chromaticity_ptr;
+
+  if (!*chromaticity) {
+    chromaticity_ptr.reset(new PrimaryChromaticity());
+  } else {
+    chromaticity_ptr.reset(*chromaticity);
+  }
+
+  if (!chromaticity_ptr.get())
+    return false;
+
+  float* value = is_x ? &chromaticity_ptr->x : &chromaticity_ptr->y;
+
+  double parser_value = 0;
+  const long long value_parse_status =
+      UnserializeFloat(reader, read_pos, value_size, parser_value);
+
+  *value = static_cast<float>(parser_value);
+
+  if (value_parse_status < 0 || *value < 0.0 || *value > 1.0)
+    return false;
+
+  *chromaticity = chromaticity_ptr.release();
+  return true;
+}
+
+bool MasteringMetadata::Parse(IMkvReader* reader, long long mm_start,
+                              long long mm_size, MasteringMetadata** mm) {
+  if (!reader || *mm)
+    return false;
+
+  std::auto_ptr<MasteringMetadata> mm_ptr(new MasteringMetadata());
+  if (!mm_ptr.get())
+    return false;
+
+  const long long mm_end = mm_start + mm_size;
+  long long read_pos = mm_start;
+
+  while (read_pos < mm_end) {
+    long long child_id = 0;
+    long long child_size = 0;
+
+    const long long status =
+        ParseElementHeader(reader, read_pos, mm_end, child_id, child_size);
+    if (status < 0)
+      return false;
+
+    if (child_id == libwebm::kMkvLuminanceMax) {
+      double value = 0;
+      const long long value_parse_status =
+          UnserializeFloat(reader, read_pos, child_size, value);
+      mm_ptr->luminance_max = static_cast<float>(value);
+      if (value_parse_status < 0 || mm_ptr->luminance_max < 0.0 ||
+          mm_ptr->luminance_max > 9999.99) {
+        return false;
+      }
+    } else if (child_id == libwebm::kMkvLuminanceMin) {
+      double value = 0;
+      const long long value_parse_status =
+          UnserializeFloat(reader, read_pos, child_size, value);
+      mm_ptr->luminance_min = static_cast<float>(value);
+      if (value_parse_status < 0 || mm_ptr->luminance_min < 0.0 ||
+          mm_ptr->luminance_min > 999.9999) {
+        return false;
+      }
+    } else {
+      bool is_x = false;
+      PrimaryChromaticity** chromaticity;
+      switch (child_id) {
+        case libwebm::kMkvPrimaryRChromaticityX:
+        case libwebm::kMkvPrimaryRChromaticityY:
+          is_x = child_id == libwebm::kMkvPrimaryRChromaticityX;
+          chromaticity = &mm_ptr->r;
+          break;
+        case libwebm::kMkvPrimaryGChromaticityX:
+        case libwebm::kMkvPrimaryGChromaticityY:
+          is_x = child_id == libwebm::kMkvPrimaryGChromaticityX;
+          chromaticity = &mm_ptr->g;
+          break;
+        case libwebm::kMkvPrimaryBChromaticityX:
+        case libwebm::kMkvPrimaryBChromaticityY:
+          is_x = child_id == libwebm::kMkvPrimaryBChromaticityX;
+          chromaticity = &mm_ptr->b;
+          break;
+        case libwebm::kMkvWhitePointChromaticityX:
+        case libwebm::kMkvWhitePointChromaticityY:
+          is_x = child_id == libwebm::kMkvWhitePointChromaticityX;
+          chromaticity = &mm_ptr->white_point;
+          break;
+        default:
+          return false;
+      }
+      const bool value_parse_status = PrimaryChromaticity::Parse(
+          reader, read_pos, child_size, is_x, chromaticity);
+      if (!value_parse_status)
+        return false;
+    }
+
+    read_pos += child_size;
+    if (read_pos > mm_end)
+      return false;
+  }
+
+  *mm = mm_ptr.release();
+  return true;
+}
+
+bool Colour::Parse(IMkvReader* reader, long long colour_start,
+                   long long colour_size, Colour** colour) {
+  if (!reader || *colour)
+    return false;
+
+  std::auto_ptr<Colour> colour_ptr(new Colour());
+  if (!colour_ptr.get())
+    return false;
+
+  const long long colour_end = colour_start + colour_size;
+  long long read_pos = colour_start;
+
+  while (read_pos < colour_end) {
+    long long child_id = 0;
+    long long child_size = 0;
+
+    const long status =
+        ParseElementHeader(reader, read_pos, colour_end, child_id, child_size);
+    if (status < 0)
+      return false;
+
+    if (child_id == libwebm::kMkvMatrixCoefficients) {
+      colour_ptr->matrix_coefficients =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->matrix_coefficients < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvBitsPerChannel) {
+      colour_ptr->bits_per_channel =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->bits_per_channel < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSubsamplingHorz) {
+      colour_ptr->chroma_subsampling_horz =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_subsampling_horz < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSubsamplingVert) {
+      colour_ptr->chroma_subsampling_vert =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_subsampling_vert < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvCbSubsamplingHorz) {
+      colour_ptr->cb_subsampling_horz =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->cb_subsampling_horz < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvCbSubsamplingVert) {
+      colour_ptr->cb_subsampling_vert =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->cb_subsampling_vert < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSitingHorz) {
+      colour_ptr->chroma_siting_horz =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_siting_horz < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvChromaSitingVert) {
+      colour_ptr->chroma_siting_vert =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->chroma_siting_vert < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvRange) {
+      colour_ptr->range = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->range < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvTransferCharacteristics) {
+      colour_ptr->transfer_characteristics =
+          UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->transfer_characteristics < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvPrimaries) {
+      colour_ptr->primaries = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->primaries < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvMaxCLL) {
+      colour_ptr->max_cll = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->max_cll < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvMaxFALL) {
+      colour_ptr->max_fall = UnserializeUInt(reader, read_pos, child_size);
+      if (colour_ptr->max_fall < 0)
+        return false;
+    } else if (child_id == libwebm::kMkvMasteringMetadata) {
+      if (!MasteringMetadata::Parse(reader, read_pos, child_size,
+                                    &colour_ptr->mastering_metadata))
+        return false;
+    } else {
+      return false;
+    }
+
+    read_pos += child_size;
+    if (read_pos > colour_end)
+      return false;
+  }
+  *colour = colour_ptr.release();
+  return true;
+}
+
+VideoTrack::VideoTrack(Segment* pSegment, long long element_start,
+                       long long element_size)
+    : Track(pSegment, element_start, element_size), m_colour(NULL) {}
+
+VideoTrack::~VideoTrack() { delete m_colour; }
+
+long VideoTrack::Parse(Segment* pSegment, const Info& info,
+                       long long element_start, long long element_size,
+                       VideoTrack*& pResult) {
+  if (pResult)
+    return -1;
+
+  if (info.type != Track::kVideo)
+    return -1;
+
+  long long width = 0;
+  long long height = 0;
+  long long display_width = 0;
+  long long display_height = 0;
+  long long display_unit = 0;
+  long long stereo_mode = 0;
+
+  double rate = 0.0;
+
+  IMkvReader* const pReader = pSegment->m_pReader;
+
+  const Settings& s = info.settings;
+  assert(s.start >= 0);
+  assert(s.size >= 0);
+
+  long long pos = s.start;
+  assert(pos >= 0);
+
+  const long long stop = pos + s.size;
+
+  Colour* colour = NULL;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvPixelWidth) {
+      width = UnserializeUInt(pReader, pos, size);
+
+      if (width <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvPixelHeight) {
+      height = UnserializeUInt(pReader, pos, size);
+
+      if (height <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDisplayWidth) {
+      display_width = UnserializeUInt(pReader, pos, size);
+
+      if (display_width <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDisplayHeight) {
+      display_height = UnserializeUInt(pReader, pos, size);
+
+      if (display_height <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvDisplayUnit) {
+      display_unit = UnserializeUInt(pReader, pos, size);
+
+      if (display_unit < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvStereoMode) {
+      stereo_mode = UnserializeUInt(pReader, pos, size);
+
+      if (stereo_mode < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvFrameRate) {
+      const long status = UnserializeFloat(pReader, pos, size, rate);
+
+      if (status < 0)
+        return status;
+
+      if (rate <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvColour) {
+      if (!Colour::Parse(pReader, pos, size, &colour))
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  VideoTrack* const pTrack =
+      new (std::nothrow) VideoTrack(pSegment, element_start, element_size);
+
+  if (pTrack == NULL)
+    return -1;  // generic error
+
+  const int status = info.Copy(pTrack->m_info);
+
+  if (status) {  // error
+    delete pTrack;
+    return status;
+  }
+
+  pTrack->m_width = width;
+  pTrack->m_height = height;
+  pTrack->m_display_width = display_width;
+  pTrack->m_display_height = display_height;
+  pTrack->m_display_unit = display_unit;
+  pTrack->m_stereo_mode = stereo_mode;
+  pTrack->m_rate = rate;
+  pTrack->m_colour = colour;
+
+  pResult = pTrack;
+  return 0;  // success
+}
+
+bool VideoTrack::VetEntry(const BlockEntry* pBlockEntry) const {
+  return Track::VetEntry(pBlockEntry) && pBlockEntry->GetBlock()->IsKey();
+}
+
+long VideoTrack::Seek(long long time_ns, const BlockEntry*& pResult) const {
+  const long status = GetFirst(pResult);
+
+  if (status < 0)  // buffer underflow, etc
+    return status;
+
+  assert(pResult);
+
+  if (pResult->EOS())
+    return 0;
+
+  const Cluster* pCluster = pResult->GetCluster();
+  assert(pCluster);
+  assert(pCluster->GetIndex() >= 0);
+
+  if (time_ns <= pResult->GetBlock()->GetTime(pCluster))
+    return 0;
+
+  Cluster** const clusters = m_pSegment->m_clusters;
+  assert(clusters);
+
+  const long count = m_pSegment->GetCount();  // loaded only, not pre-loaded
+  assert(count > 0);
+
+  Cluster** const i = clusters + pCluster->GetIndex();
+  assert(i);
+  assert(*i == pCluster);
+  assert(pCluster->GetTime() <= time_ns);
+
+  Cluster** const j = clusters + count;
+
+  Cluster** lo = i;
+  Cluster** hi = j;
+
+  while (lo < hi) {
+    // INVARIANT:
+    //[i, lo) <= time_ns
+    //[lo, hi) ?
+    //[hi, j)  > time_ns
+
+    Cluster** const mid = lo + (hi - lo) / 2;
+    assert(mid < hi);
+
+    pCluster = *mid;
+    assert(pCluster);
+    assert(pCluster->GetIndex() >= 0);
+    assert(pCluster->GetIndex() == long(mid - m_pSegment->m_clusters));
+
+    const long long t = pCluster->GetTime();
+
+    if (t <= time_ns)
+      lo = mid + 1;
+    else
+      hi = mid;
+
+    assert(lo <= hi);
+  }
+
+  assert(lo == hi);
+  assert(lo > i);
+  assert(lo <= j);
+
+  pCluster = *--lo;
+  assert(pCluster);
+  assert(pCluster->GetTime() <= time_ns);
+
+  pResult = pCluster->GetEntry(this, time_ns);
+
+  if ((pResult != 0) && !pResult->EOS())  // found a keyframe
+    return 0;
+
+  while (lo != i) {
+    pCluster = *--lo;
+    assert(pCluster);
+    assert(pCluster->GetTime() <= time_ns);
+
+    pResult = pCluster->GetEntry(this, time_ns);
+
+    if ((pResult != 0) && !pResult->EOS())
+      return 0;
+  }
+
+  // weird: we're on the first cluster, but no keyframe found
+  // should never happen but we must return something anyway
+
+  pResult = GetEOS();
+  return 0;
+}
+
+Colour* VideoTrack::GetColour() const { return m_colour; }
+
+long long VideoTrack::GetWidth() const { return m_width; }
+
+long long VideoTrack::GetHeight() const { return m_height; }
+
+long long VideoTrack::GetDisplayWidth() const {
+  return m_display_width > 0 ? m_display_width : GetWidth();
+}
+
+long long VideoTrack::GetDisplayHeight() const {
+  return m_display_height > 0 ? m_display_height : GetHeight();
+}
+
+long long VideoTrack::GetDisplayUnit() const { return m_display_unit; }
+
+long long VideoTrack::GetStereoMode() const { return m_stereo_mode; }
+
+double VideoTrack::GetFrameRate() const { return m_rate; }
+
+AudioTrack::AudioTrack(Segment* pSegment, long long element_start,
+                       long long element_size)
+    : Track(pSegment, element_start, element_size) {}
+
+long AudioTrack::Parse(Segment* pSegment, const Info& info,
+                       long long element_start, long long element_size,
+                       AudioTrack*& pResult) {
+  if (pResult)
+    return -1;
+
+  if (info.type != Track::kAudio)
+    return -1;
+
+  IMkvReader* const pReader = pSegment->m_pReader;
+
+  const Settings& s = info.settings;
+  assert(s.start >= 0);
+  assert(s.size >= 0);
+
+  long long pos = s.start;
+  assert(pos >= 0);
+
+  const long long stop = pos + s.size;
+
+  double rate = 8000.0;  // MKV default
+  long long channels = 1;
+  long long bit_depth = 0;
+
+  while (pos < stop) {
+    long long id, size;
+
+    long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (id == libwebm::kMkvSamplingFrequency) {
+      status = UnserializeFloat(pReader, pos, size, rate);
+
+      if (status < 0)
+        return status;
+
+      if (rate <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvChannels) {
+      channels = UnserializeUInt(pReader, pos, size);
+
+      if (channels <= 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvBitDepth) {
+      bit_depth = UnserializeUInt(pReader, pos, size);
+
+      if (bit_depth <= 0)
+        return E_FILE_FORMAT_INVALID;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  AudioTrack* const pTrack =
+      new (std::nothrow) AudioTrack(pSegment, element_start, element_size);
+
+  if (pTrack == NULL)
+    return -1;  // generic error
+
+  const int status = info.Copy(pTrack->m_info);
+
+  if (status) {
+    delete pTrack;
+    return status;
+  }
+
+  pTrack->m_rate = rate;
+  pTrack->m_channels = channels;
+  pTrack->m_bitDepth = bit_depth;
+
+  pResult = pTrack;
+  return 0;  // success
+}
+
+double AudioTrack::GetSamplingRate() const { return m_rate; }
+
+long long AudioTrack::GetChannels() const { return m_channels; }
+
+long long AudioTrack::GetBitDepth() const { return m_bitDepth; }
+
+Tracks::Tracks(Segment* pSegment, long long start, long long size_,
+               long long element_start, long long element_size)
+    : m_pSegment(pSegment),
+      m_start(start),
+      m_size(size_),
+      m_element_start(element_start),
+      m_element_size(element_size),
+      m_trackEntries(NULL),
+      m_trackEntriesEnd(NULL) {}
+
+long Tracks::Parse() {
+  assert(m_trackEntries == NULL);
+  assert(m_trackEntriesEnd == NULL);
+
+  const long long stop = m_start + m_size;
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  int count = 0;
+  long long pos = m_start;
+
+  while (pos < stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size == 0)  // weird
+      continue;
+
+    if (id == libwebm::kMkvTrackEntry)
+      ++count;
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (count <= 0)
+    return 0;  // success
+
+  m_trackEntries = new (std::nothrow) Track*[count];
+
+  if (m_trackEntries == NULL)
+    return -1;
+
+  m_trackEntriesEnd = m_trackEntries;
+
+  pos = m_start;
+
+  while (pos < stop) {
+    const long long element_start = pos;
+
+    long long id, payload_size;
+
+    const long status =
+        ParseElementHeader(pReader, pos, stop, id, payload_size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (payload_size == 0)  // weird
+      continue;
+
+    const long long payload_stop = pos + payload_size;
+    assert(payload_stop <= stop);  // checked in ParseElement
+
+    const long long element_size = payload_stop - element_start;
+
+    if (id == libwebm::kMkvTrackEntry) {
+      Track*& pTrack = *m_trackEntriesEnd;
+      pTrack = NULL;
+
+      const long status = ParseTrackEntry(pos, payload_size, element_start,
+                                          element_size, pTrack);
+      if (status)
+        return status;
+
+      if (pTrack)
+        ++m_trackEntriesEnd;
+    }
+
+    pos = payload_stop;
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+
+  return 0;  // success
+}
+
+unsigned long Tracks::GetTracksCount() const {
+  const ptrdiff_t result = m_trackEntriesEnd - m_trackEntries;
+  assert(result >= 0);
+
+  return static_cast<unsigned long>(result);
+}
+
+long Tracks::ParseTrackEntry(long long track_start, long long track_size,
+                             long long element_start, long long element_size,
+                             Track*& pResult) const {
+  if (pResult)
+    return -1;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = track_start;
+  const long long track_stop = track_start + track_size;
+
+  Track::Info info;
+
+  info.type = 0;
+  info.number = 0;
+  info.uid = 0;
+  info.defaultDuration = 0;
+
+  Track::Settings v;
+  v.start = -1;
+  v.size = -1;
+
+  Track::Settings a;
+  a.start = -1;
+  a.size = -1;
+
+  Track::Settings e;  // content_encodings_settings;
+  e.start = -1;
+  e.size = -1;
+
+  long long lacing = 1;  // default is true
+
+  while (pos < track_stop) {
+    long long id, size;
+
+    const long status = ParseElementHeader(pReader, pos, track_stop, id, size);
+
+    if (status < 0)  // error
+      return status;
+
+    if (size < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long start = pos;
+
+    if (id == libwebm::kMkvVideo) {
+      v.start = start;
+      v.size = size;
+    } else if (id == libwebm::kMkvAudio) {
+      a.start = start;
+      a.size = size;
+    } else if (id == libwebm::kMkvContentEncodings) {
+      e.start = start;
+      e.size = size;
+    } else if (id == libwebm::kMkvTrackUID) {
+      if (size > 8)
+        return E_FILE_FORMAT_INVALID;
+
+      info.uid = 0;
+
+      long long pos_ = start;
+      const long long pos_end = start + size;
+
+      while (pos_ != pos_end) {
+        unsigned char b;
+
+        const int status = pReader->Read(pos_, 1, &b);
+
+        if (status)
+          return status;
+
+        info.uid <<= 8;
+        info.uid |= b;
+
+        ++pos_;
+      }
+    } else if (id == libwebm::kMkvTrackNumber) {
+      const long long num = UnserializeUInt(pReader, pos, size);
+
+      if ((num <= 0) || (num > 127))
+        return E_FILE_FORMAT_INVALID;
+
+      info.number = static_cast<long>(num);
+    } else if (id == libwebm::kMkvTrackType) {
+      const long long type = UnserializeUInt(pReader, pos, size);
+
+      if ((type <= 0) || (type > 254))
+        return E_FILE_FORMAT_INVALID;
+
+      info.type = static_cast<long>(type);
+    } else if (id == libwebm::kMkvName) {
+      const long status =
+          UnserializeString(pReader, pos, size, info.nameAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvLanguage) {
+      const long status = UnserializeString(pReader, pos, size, info.language);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvDefaultDuration) {
+      const long long duration = UnserializeUInt(pReader, pos, size);
+
+      if (duration < 0)
+        return E_FILE_FORMAT_INVALID;
+
+      info.defaultDuration = static_cast<unsigned long long>(duration);
+    } else if (id == libwebm::kMkvCodecID) {
+      const long status = UnserializeString(pReader, pos, size, info.codecId);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvFlagLacing) {
+      lacing = UnserializeUInt(pReader, pos, size);
+
+      if ((lacing < 0) || (lacing > 1))
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvCodecPrivate) {
+      delete[] info.codecPrivate;
+      info.codecPrivate = NULL;
+      info.codecPrivateSize = 0;
+
+      const size_t buflen = static_cast<size_t>(size);
+
+      if (buflen) {
+        unsigned char* buf = SafeArrayAlloc<unsigned char>(1, buflen);
+
+        if (buf == NULL)
+          return -1;
+
+        const int status = pReader->Read(pos, static_cast<long>(buflen), buf);
+
+        if (status) {
+          delete[] buf;
+          return status;
+        }
+
+        info.codecPrivate = buf;
+        info.codecPrivateSize = buflen;
+      }
+    } else if (id == libwebm::kMkvCodecName) {
+      const long status =
+          UnserializeString(pReader, pos, size, info.codecNameAsUTF8);
+
+      if (status)
+        return status;
+    } else if (id == libwebm::kMkvCodecDelay) {
+      info.codecDelay = UnserializeUInt(pReader, pos, size);
+    } else if (id == libwebm::kMkvSeekPreRoll) {
+      info.seekPreRoll = UnserializeUInt(pReader, pos, size);
+    }
+
+    pos += size;  // consume payload
+    if (pos > track_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != track_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (info.number <= 0)  // not specified
+    return E_FILE_FORMAT_INVALID;
+
+  if (GetTrackByNumber(info.number))
+    return E_FILE_FORMAT_INVALID;
+
+  if (info.type <= 0)  // not specified
+    return E_FILE_FORMAT_INVALID;
+
+  info.lacing = (lacing > 0) ? true : false;
+
+  if (info.type == Track::kVideo) {
+    if (v.start < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (a.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    info.settings = v;
+
+    VideoTrack* pTrack = NULL;
+
+    const long status = VideoTrack::Parse(m_pSegment, info, element_start,
+                                          element_size, pTrack);
+
+    if (status)
+      return status;
+
+    pResult = pTrack;
+    assert(pResult);
+
+    if (e.start >= 0)
+      pResult->ParseContentEncodingsEntry(e.start, e.size);
+  } else if (info.type == Track::kAudio) {
+    if (a.start < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (v.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    info.settings = a;
+
+    AudioTrack* pTrack = NULL;
+
+    const long status = AudioTrack::Parse(m_pSegment, info, element_start,
+                                          element_size, pTrack);
+
+    if (status)
+      return status;
+
+    pResult = pTrack;
+    assert(pResult);
+
+    if (e.start >= 0)
+      pResult->ParseContentEncodingsEntry(e.start, e.size);
+  } else {
+    // neither video nor audio - probably metadata or subtitles
+
+    if (a.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (v.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (info.type == Track::kMetadata && e.start >= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    info.settings.start = -1;
+    info.settings.size = 0;
+
+    Track* pTrack = NULL;
+
+    const long status =
+        Track::Create(m_pSegment, info, element_start, element_size, pTrack);
+
+    if (status)
+      return status;
+
+    pResult = pTrack;
+    assert(pResult);
+  }
+
+  return 0;  // success
+}
+
+Tracks::~Tracks() {
+  Track** i = m_trackEntries;
+  Track** const j = m_trackEntriesEnd;
+
+  while (i != j) {
+    Track* const pTrack = *i++;
+    delete pTrack;
+  }
+
+  delete[] m_trackEntries;
+}
+
+const Track* Tracks::GetTrackByNumber(long tn) const {
+  if (tn < 0)
+    return NULL;
+
+  Track** i = m_trackEntries;
+  Track** const j = m_trackEntriesEnd;
+
+  while (i != j) {
+    Track* const pTrack = *i++;
+
+    if (pTrack == NULL)
+      continue;
+
+    if (tn == pTrack->GetNumber())
+      return pTrack;
+  }
+
+  return NULL;  // not found
+}
+
+const Track* Tracks::GetTrackByIndex(unsigned long idx) const {
+  const ptrdiff_t count = m_trackEntriesEnd - m_trackEntries;
+
+  if (idx >= static_cast<unsigned long>(count))
+    return NULL;
+
+  return m_trackEntries[idx];
+}
+
+long Cluster::Load(long long& pos, long& len) const {
+  if (m_pSegment == NULL)
+    return E_PARSE_FAILED;
+
+  if (m_timecode >= 0)  // at least partially loaded
+    return 0;
+
+  if (m_pos != m_element_start || m_element_size >= 0)
+    return E_PARSE_FAILED;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+  long long total, avail;
+  const int status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total >= 0 && (avail > total || m_pos > total))
+    return E_FILE_FORMAT_INVALID;
+
+  pos = m_pos;
+
+  long long cluster_size = -1;
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error or underflow
+    return static_cast<long>(result);
+
+  if (result > 0)
+    return E_BUFFER_NOT_FULL;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long id_ = ReadID(pReader, pos, len);
+
+  if (id_ < 0)  // error
+    return static_cast<long>(id_);
+
+  if (id_ != libwebm::kMkvCluster)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume id
+
+  // read cluster size
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)
+    return E_BUFFER_NOT_FULL;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long size = ReadUInt(pReader, pos, len);
+
+  if (size < 0)  // error
+    return static_cast<long>(cluster_size);
+
+  if (size == 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume length of size of element
+
+  const long long unknown_size = (1LL << (7 * len)) - 1;
+
+  if (size != unknown_size)
+    cluster_size = size;
+
+  // pos points to start of payload
+  long long timecode = -1;
+  long long new_pos = -1;
+  bool bBlock = false;
+
+  long long cluster_stop = (cluster_size < 0) ? -1 : pos + cluster_size;
+
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      break;
+
+    // Parse ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id == 0)
+      return E_FILE_FORMAT_INVALID;
+
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
+
+    if (id == libwebm::kMkvCluster)
+      break;
+
+    if (id == libwebm::kMkvCues)
+      break;
+
+    pos += len;  // consume ID field
+
+    // Parse Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume size field
+
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // pos now points to start of payload
+
+    if (size == 0)
+      continue;
+
+    if ((cluster_stop >= 0) && ((pos + size) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvTimecode) {
+      len = static_cast<long>(size);
+
+      if ((pos + size) > avail)
+        return E_BUFFER_NOT_FULL;
+
+      timecode = UnserializeUInt(pReader, pos, size);
+
+      if (timecode < 0)  // error (or underflow)
+        return static_cast<long>(timecode);
+
+      new_pos = pos + size;
+
+      if (bBlock)
+        break;
+    } else if (id == libwebm::kMkvBlockGroup) {
+      bBlock = true;
+      break;
+    } else if (id == libwebm::kMkvSimpleBlock) {
+      bBlock = true;
+      break;
+    }
+
+    pos += size;  // consume payload
+    if (cluster_stop >= 0 && pos > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (cluster_stop >= 0 && pos > cluster_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (timecode < 0)  // no timecode found
+    return E_FILE_FORMAT_INVALID;
+
+  if (!bBlock)
+    return E_FILE_FORMAT_INVALID;
+
+  m_pos = new_pos;  // designates position just beyond timecode payload
+  m_timecode = timecode;  // m_timecode >= 0 means we're partially loaded
+
+  if (cluster_size >= 0)
+    m_element_size = cluster_stop - m_element_start;
+
+  return 0;
+}
+
+long Cluster::Parse(long long& pos, long& len) const {
+  long status = Load(pos, len);
+
+  if (status < 0)
+    return status;
+
+  if (m_pos < m_element_start || m_timecode < 0)
+    return E_PARSE_FAILED;
+
+  const long long cluster_stop =
+      (m_element_size < 0) ? -1 : m_element_start + m_element_size;
+
+  if ((cluster_stop >= 0) && (m_pos >= cluster_stop))
+    return 1;  // nothing else to do
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long total, avail;
+
+  status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  if (total >= 0 && avail > total)
+    return E_FILE_FORMAT_INVALID;
+
+  pos = m_pos;
+
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      break;
+
+    if ((total >= 0) && (pos >= total)) {
+      if (m_element_size < 0)
+        m_element_size = pos - m_element_start;
+
+      break;
+    }
+
+    // Parse ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)
+      return E_FILE_FORMAT_INVALID;
+
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
+
+    if ((id == libwebm::kMkvCluster) || (id == libwebm::kMkvCues)) {
+      if (m_element_size < 0)
+        m_element_size = pos - m_element_start;
+
+      break;
+    }
+
+    pos += len;  // consume ID field
+
+    // Parse Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume size field
+
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // pos now points to start of payload
+
+    if (size == 0)
+      continue;
+
+    // const long long block_start = pos;
+    const long long block_stop = pos + size;
+
+    if (cluster_stop >= 0) {
+      if (block_stop > cluster_stop) {
+        if (id == libwebm::kMkvBlockGroup || id == libwebm::kMkvSimpleBlock) {
+          return E_FILE_FORMAT_INVALID;
+        }
+
+        pos = cluster_stop;
+        break;
+      }
+    } else if ((total >= 0) && (block_stop > total)) {
+      m_element_size = total - m_element_start;
+      pos = total;
+      break;
+    } else if (block_stop > avail) {
+      len = static_cast<long>(size);
+      return E_BUFFER_NOT_FULL;
+    }
+
+    Cluster* const this_ = const_cast<Cluster*>(this);
+
+    if (id == libwebm::kMkvBlockGroup)
+      return this_->ParseBlockGroup(size, pos, len);
+
+    if (id == libwebm::kMkvSimpleBlock)
+      return this_->ParseSimpleBlock(size, pos, len);
+
+    pos += size;  // consume payload
+    if (cluster_stop >= 0 && pos > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (m_element_size < 1)
+    return E_FILE_FORMAT_INVALID;
+
+  m_pos = pos;
+  if (cluster_stop >= 0 && m_pos > cluster_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if (m_entries_count > 0) {
+    const long idx = m_entries_count - 1;
+
+    const BlockEntry* const pLast = m_entries[idx];
+    if (pLast == NULL)
+      return E_PARSE_FAILED;
+
+    const Block* const pBlock = pLast->GetBlock();
+    if (pBlock == NULL)
+      return E_PARSE_FAILED;
+
+    const long long start = pBlock->m_start;
+
+    if ((total >= 0) && (start > total))
+      return E_PARSE_FAILED;  // defend against trucated stream
+
+    const long long size = pBlock->m_size;
+
+    const long long stop = start + size;
+    if (cluster_stop >= 0 && stop > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && (stop > total))
+      return E_PARSE_FAILED;  // defend against trucated stream
+  }
+
+  return 1;  // no more entries
+}
+
+long Cluster::ParseSimpleBlock(long long block_size, long long& pos,
+                               long& len) {
+  const long long block_start = pos;
+  const long long block_stop = pos + block_size;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long total, avail;
+
+  long status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  // parse track number
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long result = GetUIntLength(pReader, pos, len);
+
+  if (result < 0)  // error
+    return static_cast<long>(result);
+
+  if (result > 0)  // weird
+    return E_BUFFER_NOT_FULL;
+
+  if ((pos + len) > block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > avail)
+    return E_BUFFER_NOT_FULL;
+
+  const long long track = ReadUInt(pReader, pos, len);
+
+  if (track < 0)  // error
+    return static_cast<long>(track);
+
+  if (track == 0)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume track number
+
+  if ((pos + 2) > block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + 2) > avail) {
+    len = 2;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  pos += 2;  // consume timecode
+
+  if ((pos + 1) > block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + 1) > avail) {
+    len = 1;
+    return E_BUFFER_NOT_FULL;
+  }
+
+  unsigned char flags;
+
+  status = pReader->Read(pos, 1, &flags);
+
+  if (status < 0) {  // error or underflow
+    len = 1;
+    return status;
+  }
+
+  ++pos;  // consume flags byte
+  assert(pos <= avail);
+
+  if (pos >= block_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  const int lacing = int(flags & 0x06) >> 1;
+
+  if ((lacing != 0) && (block_stop > avail)) {
+    len = static_cast<long>(block_stop - pos);
+    return E_BUFFER_NOT_FULL;
+  }
+
+  status = CreateBlock(libwebm::kMkvSimpleBlock, block_start, block_size,
+                       0);  // DiscardPadding
+
+  if (status != 0)
+    return status;
+
+  m_pos = block_stop;
+
+  return 0;  // success
+}
+
+long Cluster::ParseBlockGroup(long long payload_size, long long& pos,
+                              long& len) {
+  const long long payload_start = pos;
+  const long long payload_stop = pos + payload_size;
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long total, avail;
+
+  long status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  if ((total >= 0) && (payload_stop > total))
+    return E_FILE_FORMAT_INVALID;
+
+  if (payload_stop > avail) {
+    len = static_cast<long>(payload_size);
+    return E_BUFFER_NOT_FULL;
+  }
+
+  long long discard_padding = 0;
+
+  while (pos < payload_stop) {
+    // parse sub-block element ID
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((pos + len) > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id == 0)  // not a valid ID
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID field
+
+    // Parse Size
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((pos + len) > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume size field
+
+    // pos now points to start of sub-block group payload
+
+    if (pos > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if (size == 0)  // weird
+      continue;
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvDiscardPadding) {
+      status = UnserializeInt(pReader, pos, size, discard_padding);
+
+      if (status < 0)  // error
+        return status;
+    }
+
+    if (id != libwebm::kMkvBlock) {
+      pos += size;  // consume sub-part of block group
+
+      if (pos > payload_stop)
+        return E_FILE_FORMAT_INVALID;
+
+      continue;
+    }
+
+    const long long block_stop = pos + size;
+
+    if (block_stop > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    // parse track number
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((pos + len) > block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long track = ReadUInt(pReader, pos, len);
+
+    if (track < 0)  // error
+      return static_cast<long>(track);
+
+    if (track == 0)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume track number
+
+    if ((pos + 2) > block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + 2) > avail) {
+      len = 2;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    pos += 2;  // consume timecode
+
+    if ((pos + 1) > block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    unsigned char flags;
+
+    status = pReader->Read(pos, 1, &flags);
+
+    if (status < 0) {  // error or underflow
+      len = 1;
+      return status;
+    }
+
+    ++pos;  // consume flags byte
+    assert(pos <= avail);
+
+    if (pos >= block_stop)
+      return E_FILE_FORMAT_INVALID;
+
+    const int lacing = int(flags & 0x06) >> 1;
+
+    if ((lacing != 0) && (block_stop > avail)) {
+      len = static_cast<long>(block_stop - pos);
+      return E_BUFFER_NOT_FULL;
+    }
+
+    pos = block_stop;  // consume block-part of block group
+    if (pos > payload_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  if (pos != payload_stop)
+    return E_FILE_FORMAT_INVALID;
+
+  status = CreateBlock(libwebm::kMkvBlockGroup, payload_start, payload_size,
+                       discard_padding);
+  if (status != 0)
+    return status;
+
+  m_pos = payload_stop;
+
+  return 0;  // success
+}
+
+long Cluster::GetEntry(long index, const mkvparser::BlockEntry*& pEntry) const {
+  assert(m_pos >= m_element_start);
+
+  pEntry = NULL;
+
+  if (index < 0)
+    return -1;  // generic error
+
+  if (m_entries_count < 0)
+    return E_BUFFER_NOT_FULL;
+
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count <= m_entries_size);
+
+  if (index < m_entries_count) {
+    pEntry = m_entries[index];
+    assert(pEntry);
+
+    return 1;  // found entry
+  }
+
+  if (m_element_size < 0)  // we don't know cluster end yet
+    return E_BUFFER_NOT_FULL;  // underflow
+
+  const long long element_stop = m_element_start + m_element_size;
+
+  if (m_pos >= element_stop)
+    return 0;  // nothing left to parse
+
+  return E_BUFFER_NOT_FULL;  // underflow, since more remains to be parsed
+}
+
+Cluster* Cluster::Create(Segment* pSegment, long idx, long long off) {
+  if (!pSegment || off < 0)
+    return NULL;
+
+  const long long element_start = pSegment->m_start + off;
+
+  Cluster* const pCluster =
+      new (std::nothrow) Cluster(pSegment, idx, element_start);
+
+  return pCluster;
+}
+
+Cluster::Cluster()
+    : m_pSegment(NULL),
+      m_element_start(0),
+      m_index(0),
+      m_pos(0),
+      m_element_size(0),
+      m_timecode(0),
+      m_entries(NULL),
+      m_entries_size(0),
+      m_entries_count(0)  // means "no entries"
+{}
+
+Cluster::Cluster(Segment* pSegment, long idx, long long element_start
+                 /* long long element_size */)
+    : m_pSegment(pSegment),
+      m_element_start(element_start),
+      m_index(idx),
+      m_pos(element_start),
+      m_element_size(-1 /* element_size */),
+      m_timecode(-1),
+      m_entries(NULL),
+      m_entries_size(0),
+      m_entries_count(-1)  // means "has not been parsed yet"
+{}
+
+Cluster::~Cluster() {
+  if (m_entries_count <= 0)
+    return;
+
+  BlockEntry** i = m_entries;
+  BlockEntry** const j = m_entries + m_entries_count;
+
+  while (i != j) {
+    BlockEntry* p = *i++;
+    assert(p);
+
+    delete p;
+  }
+
+  delete[] m_entries;
+}
+
+bool Cluster::EOS() const { return (m_pSegment == NULL); }
+
+long Cluster::GetIndex() const { return m_index; }
+
+long long Cluster::GetPosition() const {
+  const long long pos = m_element_start - m_pSegment->m_start;
+  assert(pos >= 0);
+
+  return pos;
+}
+
+long long Cluster::GetElementSize() const { return m_element_size; }
+
+long Cluster::HasBlockEntries(
+    const Segment* pSegment,
+    long long off,  // relative to start of segment payload
+    long long& pos, long& len) {
+  assert(pSegment);
+  assert(off >= 0);  // relative to segment
+
+  IMkvReader* const pReader = pSegment->m_pReader;
+
+  long long total, avail;
+
+  long status = pReader->Length(&total, &avail);
+
+  if (status < 0)  // error
+    return status;
+
+  assert((total < 0) || (avail <= total));
+
+  pos = pSegment->m_start + off;  // absolute
+
+  if ((total >= 0) && (pos >= total))
+    return 0;  // we don't even have a complete cluster
+
+  const long long segment_stop =
+      (pSegment->m_size < 0) ? -1 : pSegment->m_start + pSegment->m_size;
+
+  long long cluster_stop = -1;  // interpreted later to mean "unknown size"
+
+  {
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // need more data
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return 0;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    if (id != libwebm::kMkvCluster)
+      return E_PARSE_FAILED;
+
+    pos += len;  // consume Cluster ID field
+
+    // read size field
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // weird
+      return E_BUFFER_NOT_FULL;
+
+    if ((segment_stop >= 0) && ((pos + len) > segment_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((total >= 0) && ((pos + len) > total))
+      return 0;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    if (size == 0)
+      return 0;  // cluster does not have entries
+
+    pos += len;  // consume size field
+
+    // pos now points to start of payload
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size != unknown_size) {
+      cluster_stop = pos + size;
+      assert(cluster_stop >= 0);
+
+      if ((segment_stop >= 0) && (cluster_stop > segment_stop))
+        return E_FILE_FORMAT_INVALID;
+
+      if ((total >= 0) && (cluster_stop > total))
+        // return E_FILE_FORMAT_INVALID;  //too conservative
+        return 0;  // cluster does not have any entries
+    }
+  }
+
+  for (;;) {
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      return 0;  // no entries detected
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    long long result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // need more data
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long id = ReadID(pReader, pos, len);
+
+    if (id < 0)  // error
+      return static_cast<long>(id);
+
+    // This is the distinguished set of ID's we use to determine
+    // that we have exhausted the sub-element's inside the cluster
+    // whose ID we parsed earlier.
+
+    if (id == libwebm::kMkvCluster)
+      return 0;  // no entries found
+
+    if (id == libwebm::kMkvCues)
+      return 0;  // no entries found
+
+    pos += len;  // consume id field
+
+    if ((cluster_stop >= 0) && (pos >= cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    // read size field
+
+    if ((pos + 1) > avail) {
+      len = 1;
+      return E_BUFFER_NOT_FULL;
+    }
+
+    result = GetUIntLength(pReader, pos, len);
+
+    if (result < 0)  // error
+      return static_cast<long>(result);
+
+    if (result > 0)  // underflow
+      return E_BUFFER_NOT_FULL;
+
+    if ((cluster_stop >= 0) && ((pos + len) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > avail)
+      return E_BUFFER_NOT_FULL;
+
+    const long long size = ReadUInt(pReader, pos, len);
+
+    if (size < 0)  // error
+      return static_cast<long>(size);
+
+    pos += len;  // consume size field
+
+    // pos now points to start of payload
+
+    if ((cluster_stop >= 0) && (pos > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if (size == 0)  // weird
+      continue;
+
+    const long long unknown_size = (1LL << (7 * len)) - 1;
+
+    if (size == unknown_size)
+      return E_FILE_FORMAT_INVALID;  // not supported inside cluster
+
+    if ((cluster_stop >= 0) && ((pos + size) > cluster_stop))
+      return E_FILE_FORMAT_INVALID;
+
+    if (id == libwebm::kMkvBlockGroup)
+      return 1;  // have at least one entry
+
+    if (id == libwebm::kMkvSimpleBlock)
+      return 1;  // have at least one entry
+
+    pos += size;  // consume payload
+    if (cluster_stop >= 0 && pos > cluster_stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+}
+
+long long Cluster::GetTimeCode() const {
+  long long pos;
+  long len;
+
+  const long status = Load(pos, len);
+
+  if (status < 0)  // error
+    return status;
+
+  return m_timecode;
+}
+
+long long Cluster::GetTime() const {
+  const long long tc = GetTimeCode();
+
+  if (tc < 0)
+    return tc;
+
+  const SegmentInfo* const pInfo = m_pSegment->GetInfo();
+  assert(pInfo);
+
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
+
+  const long long t = m_timecode * scale;
+
+  return t;
+}
+
+long long Cluster::GetFirstTime() const {
+  const BlockEntry* pEntry;
+
+  const long status = GetFirst(pEntry);
+
+  if (status < 0)  // error
+    return status;
+
+  if (pEntry == NULL)  // empty cluster
+    return GetTime();
+
+  const Block* const pBlock = pEntry->GetBlock();
+  assert(pBlock);
+
+  return pBlock->GetTime(this);
+}
+
+long long Cluster::GetLastTime() const {
+  const BlockEntry* pEntry;
+
+  const long status = GetLast(pEntry);
+
+  if (status < 0)  // error
+    return status;
+
+  if (pEntry == NULL)  // empty cluster
+    return GetTime();
+
+  const Block* const pBlock = pEntry->GetBlock();
+  assert(pBlock);
+
+  return pBlock->GetTime(this);
+}
+
+long Cluster::CreateBlock(long long id,
+                          long long pos,  // absolute pos of payload
+                          long long size, long long discard_padding) {
+  if (id != libwebm::kMkvBlockGroup && id != libwebm::kMkvSimpleBlock)
+    return E_PARSE_FAILED;
+
+  if (m_entries_count < 0) {  // haven't parsed anything yet
+    assert(m_entries == NULL);
+    assert(m_entries_size == 0);
+
+    m_entries_size = 1024;
+    m_entries = new (std::nothrow) BlockEntry*[m_entries_size];
+    if (m_entries == NULL)
+      return -1;
+
+    m_entries_count = 0;
+  } else {
+    assert(m_entries);
+    assert(m_entries_size > 0);
+    assert(m_entries_count <= m_entries_size);
+
+    if (m_entries_count >= m_entries_size) {
+      const long entries_size = 2 * m_entries_size;
+
+      BlockEntry** const entries = new (std::nothrow) BlockEntry*[entries_size];
+      if (entries == NULL)
+        return -1;
+
+      BlockEntry** src = m_entries;
+      BlockEntry** const src_end = src + m_entries_count;
+
+      BlockEntry** dst = entries;
+
+      while (src != src_end)
+        *dst++ = *src++;
+
+      delete[] m_entries;
+
+      m_entries = entries;
+      m_entries_size = entries_size;
+    }
+  }
+
+  if (id == libwebm::kMkvBlockGroup)
+    return CreateBlockGroup(pos, size, discard_padding);
+  else
+    return CreateSimpleBlock(pos, size);
+}
+
+long Cluster::CreateBlockGroup(long long start_offset, long long size,
+                               long long discard_padding) {
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count >= 0);
+  assert(m_entries_count < m_entries_size);
+
+  IMkvReader* const pReader = m_pSegment->m_pReader;
+
+  long long pos = start_offset;
+  const long long stop = start_offset + size;
+
+  // For WebM files, there is a bias towards previous reference times
+  //(in order to support alt-ref frames, which refer back to the previous
+  // keyframe).  Normally a 0 value is not possible, but here we tenatively
+  // allow 0 as the value of a reference frame, with the interpretation
+  // that this is a "previous" reference time.
+
+  long long prev = 1;  // nonce
+  long long next = 0;  // nonce
+  long long duration = -1;  // really, this is unsigned
+
+  long long bpos = -1;
+  long long bsize = -1;
+
+  while (pos < stop) {
+    long len;
+    const long long id = ReadID(pReader, pos, len);
+    if (id < 0 || (pos + len) > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume ID
+
+    const long long size = ReadUInt(pReader, pos, len);
+    assert(size >= 0);  // TODO
+    assert((pos + len) <= stop);
+
+    pos += len;  // consume size
+
+    if (id == libwebm::kMkvBlock) {
+      if (bpos < 0) {  // Block ID
+        bpos = pos;
+        bsize = size;
+      }
+    } else if (id == libwebm::kMkvBlockDuration) {
+      if (size > 8)
+        return E_FILE_FORMAT_INVALID;
+
+      duration = UnserializeUInt(pReader, pos, size);
+
+      if (duration < 0)
+        return E_FILE_FORMAT_INVALID;
+    } else if (id == libwebm::kMkvReferenceBlock) {
+      if (size > 8 || size <= 0)
+        return E_FILE_FORMAT_INVALID;
+      const long size_ = static_cast<long>(size);
+
+      long long time;
+
+      long status = UnserializeInt(pReader, pos, size_, time);
+      assert(status == 0);
+      if (status != 0)
+        return -1;
+
+      if (time <= 0)  // see note above
+        prev = time;
+      else
+        next = time;
+    }
+
+    pos += size;  // consume payload
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+  if (bpos < 0)
+    return E_FILE_FORMAT_INVALID;
+
+  if (pos != stop)
+    return E_FILE_FORMAT_INVALID;
+  assert(bsize >= 0);
+
+  const long idx = m_entries_count;
+
+  BlockEntry** const ppEntry = m_entries + idx;
+  BlockEntry*& pEntry = *ppEntry;
+
+  pEntry = new (std::nothrow)
+      BlockGroup(this, idx, bpos, bsize, prev, next, duration, discard_padding);
+
+  if (pEntry == NULL)
+    return -1;  // generic error
+
+  BlockGroup* const p = static_cast<BlockGroup*>(pEntry);
+
+  const long status = p->Parse();
+
+  if (status == 0) {  // success
+    ++m_entries_count;
+    return 0;
+  }
+
+  delete pEntry;
+  pEntry = 0;
+
+  return status;
+}
+
+long Cluster::CreateSimpleBlock(long long st, long long sz) {
+  assert(m_entries);
+  assert(m_entries_size > 0);
+  assert(m_entries_count >= 0);
+  assert(m_entries_count < m_entries_size);
+
+  const long idx = m_entries_count;
+
+  BlockEntry** const ppEntry = m_entries + idx;
+  BlockEntry*& pEntry = *ppEntry;
+
+  pEntry = new (std::nothrow) SimpleBlock(this, idx, st, sz);
+
+  if (pEntry == NULL)
+    return -1;  // generic error
+
+  SimpleBlock* const p = static_cast<SimpleBlock*>(pEntry);
+
+  const long status = p->Parse();
+
+  if (status == 0) {
+    ++m_entries_count;
+    return 0;
+  }
+
+  delete pEntry;
+  pEntry = 0;
+
+  return status;
+}
+
+long Cluster::GetFirst(const BlockEntry*& pFirst) const {
+  if (m_entries_count <= 0) {
+    long long pos;
+    long len;
+
+    const long status = Parse(pos, len);
+
+    if (status < 0) {  // error
+      pFirst = NULL;
+      return status;
+    }
+
+    if (m_entries_count <= 0) {  // empty cluster
+      pFirst = NULL;
+      return 0;
+    }
+  }
+
+  assert(m_entries);
+
+  pFirst = m_entries[0];
+  assert(pFirst);
+
+  return 0;  // success
+}
+
+long Cluster::GetLast(const BlockEntry*& pLast) const {
+  for (;;) {
+    long long pos;
+    long len;
+
+    const long status = Parse(pos, len);
+
+    if (status < 0) {  // error
+      pLast = NULL;
+      return status;
+    }
+
+    if (status > 0)  // no new block
+      break;
+  }
+
+  if (m_entries_count <= 0) {
+    pLast = NULL;
+    return 0;
+  }
+
+  assert(m_entries);
+
+  const long idx = m_entries_count - 1;
+
+  pLast = m_entries[idx];
+  assert(pLast);
+
+  return 0;
+}
+
+long Cluster::GetNext(const BlockEntry* pCurr, const BlockEntry*& pNext) const {
+  assert(pCurr);
+  assert(m_entries);
+  assert(m_entries_count > 0);
+
+  size_t idx = pCurr->GetIndex();
+  assert(idx < size_t(m_entries_count));
+  assert(m_entries[idx] == pCurr);
+
+  ++idx;
+
+  if (idx >= size_t(m_entries_count)) {
+    long long pos;
+    long len;
+
+    const long status = Parse(pos, len);
+
+    if (status < 0) {  // error
+      pNext = NULL;
+      return status;
+    }
+
+    if (status > 0) {
+      pNext = NULL;
+      return 0;
+    }
+
+    assert(m_entries);
+    assert(m_entries_count > 0);
+    assert(idx < size_t(m_entries_count));
+  }
+
+  pNext = m_entries[idx];
+  assert(pNext);
+
+  return 0;
+}
+
+long Cluster::GetEntryCount() const { return m_entries_count; }
+
+const BlockEntry* Cluster::GetEntry(const Track* pTrack,
+                                    long long time_ns) const {
+  assert(pTrack);
+
+  if (m_pSegment == NULL)  // this is the special EOS cluster
+    return pTrack->GetEOS();
+
+  const BlockEntry* pResult = pTrack->GetEOS();
+
+  long index = 0;
+
+  for (;;) {
+    if (index >= m_entries_count) {
+      long long pos;
+      long len;
+
+      const long status = Parse(pos, len);
+      assert(status >= 0);
+
+      if (status > 0)  // completely parsed, and no more entries
+        return pResult;
+
+      if (status < 0)  // should never happen
+        return 0;
+
+      assert(m_entries);
+      assert(index < m_entries_count);
+    }
+
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
+
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
+
+    if (pBlock->GetTrackNumber() != pTrack->GetNumber()) {
+      ++index;
+      continue;
+    }
+
+    if (pTrack->VetEntry(pEntry)) {
+      if (time_ns < 0)  // just want first candidate block
+        return pEntry;
+
+      const long long ns = pBlock->GetTime(this);
+
+      if (ns > time_ns)
+        return pResult;
+
+      pResult = pEntry;  // have a candidate
+    } else if (time_ns >= 0) {
+      const long long ns = pBlock->GetTime(this);
+
+      if (ns > time_ns)
+        return pResult;
+    }
+
+    ++index;
+  }
+}
+
+const BlockEntry* Cluster::GetEntry(const CuePoint& cp,
+                                    const CuePoint::TrackPosition& tp) const {
+  assert(m_pSegment);
+  const long long tc = cp.GetTimeCode();
+
+  if (tp.m_block > 0) {
+    const long block = static_cast<long>(tp.m_block);
+    const long index = block - 1;
+
+    while (index >= m_entries_count) {
+      long long pos;
+      long len;
+
+      const long status = Parse(pos, len);
+
+      if (status < 0)  // TODO: can this happen?
+        return NULL;
+
+      if (status > 0)  // nothing remains to be parsed
+        return NULL;
+    }
+
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
+
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
+
+    if ((pBlock->GetTrackNumber() == tp.m_track) &&
+        (pBlock->GetTimeCode(this) == tc)) {
+      return pEntry;
+    }
+  }
+
+  long index = 0;
+
+  for (;;) {
+    if (index >= m_entries_count) {
+      long long pos;
+      long len;
+
+      const long status = Parse(pos, len);
+
+      if (status < 0)  // TODO: can this happen?
+        return NULL;
+
+      if (status > 0)  // nothing remains to be parsed
+        return NULL;
+
+      assert(m_entries);
+      assert(index < m_entries_count);
+    }
+
+    const BlockEntry* const pEntry = m_entries[index];
+    assert(pEntry);
+    assert(!pEntry->EOS());
+
+    const Block* const pBlock = pEntry->GetBlock();
+    assert(pBlock);
+
+    if (pBlock->GetTrackNumber() != tp.m_track) {
+      ++index;
+      continue;
+    }
+
+    const long long tc_ = pBlock->GetTimeCode(this);
+
+    if (tc_ < tc) {
+      ++index;
+      continue;
+    }
+
+    if (tc_ > tc)
+      return NULL;
+
+    const Tracks* const pTracks = m_pSegment->GetTracks();
+    assert(pTracks);
+
+    const long tn = static_cast<long>(tp.m_track);
+    const Track* const pTrack = pTracks->GetTrackByNumber(tn);
+
+    if (pTrack == NULL)
+      return NULL;
+
+    const long long type = pTrack->GetType();
+
+    if (type == 2)  // audio
+      return pEntry;
+
+    if (type != 1)  // not video
+      return NULL;
+
+    if (!pBlock->IsKey())
+      return NULL;
+
+    return pEntry;
+  }
+}
+
+BlockEntry::BlockEntry(Cluster* p, long idx) : m_pCluster(p), m_index(idx) {}
+BlockEntry::~BlockEntry() {}
+const Cluster* BlockEntry::GetCluster() const { return m_pCluster; }
+long BlockEntry::GetIndex() const { return m_index; }
+
+SimpleBlock::SimpleBlock(Cluster* pCluster, long idx, long long start,
+                         long long size)
+    : BlockEntry(pCluster, idx), m_block(start, size, 0) {}
+
+long SimpleBlock::Parse() { return m_block.Parse(m_pCluster); }
+BlockEntry::Kind SimpleBlock::GetKind() const { return kBlockSimple; }
+const Block* SimpleBlock::GetBlock() const { return &m_block; }
+
+BlockGroup::BlockGroup(Cluster* pCluster, long idx, long long block_start,
+                       long long block_size, long long prev, long long next,
+                       long long duration, long long discard_padding)
+    : BlockEntry(pCluster, idx),
+      m_block(block_start, block_size, discard_padding),
+      m_prev(prev),
+      m_next(next),
+      m_duration(duration) {}
+
+long BlockGroup::Parse() {
+  const long status = m_block.Parse(m_pCluster);
+
+  if (status)
+    return status;
+
+  m_block.SetKey((m_prev > 0) && (m_next <= 0));
+
+  return 0;
+}
+
+BlockEntry::Kind BlockGroup::GetKind() const { return kBlockGroup; }
+const Block* BlockGroup::GetBlock() const { return &m_block; }
+long long BlockGroup::GetPrevTimeCode() const { return m_prev; }
+long long BlockGroup::GetNextTimeCode() const { return m_next; }
+long long BlockGroup::GetDurationTimeCode() const { return m_duration; }
+
+Block::Block(long long start, long long size_, long long discard_padding)
+    : m_start(start),
+      m_size(size_),
+      m_track(0),
+      m_timecode(-1),
+      m_flags(0),
+      m_frames(NULL),
+      m_frame_count(-1),
+      m_discard_padding(discard_padding) {}
+
+Block::~Block() { delete[] m_frames; }
+
+long Block::Parse(const Cluster* pCluster) {
+  if (pCluster == NULL)
+    return -1;
+
+  if (pCluster->m_pSegment == NULL)
+    return -1;
+
+  assert(m_start >= 0);
+  assert(m_size >= 0);
+  assert(m_track <= 0);
+  assert(m_frames == NULL);
+  assert(m_frame_count <= 0);
+
+  long long pos = m_start;
+  const long long stop = m_start + m_size;
+
+  long len;
+
+  IMkvReader* const pReader = pCluster->m_pSegment->m_pReader;
+
+  m_track = ReadUInt(pReader, pos, len);
+
+  if (m_track <= 0)
+    return E_FILE_FORMAT_INVALID;
+
+  if ((pos + len) > stop)
+    return E_FILE_FORMAT_INVALID;
+
+  pos += len;  // consume track number
+
+  if ((stop - pos) < 2)
+    return E_FILE_FORMAT_INVALID;
+
+  long status;
+  long long value;
+
+  status = UnserializeInt(pReader, pos, 2, value);
+
+  if (status)
+    return E_FILE_FORMAT_INVALID;
+
+  if (value < SHRT_MIN)
+    return E_FILE_FORMAT_INVALID;
+
+  if (value > SHRT_MAX)
+    return E_FILE_FORMAT_INVALID;
+
+  m_timecode = static_cast<short>(value);
+
+  pos += 2;
+
+  if ((stop - pos) <= 0)
+    return E_FILE_FORMAT_INVALID;
+
+  status = pReader->Read(pos, 1, &m_flags);
+
+  if (status)
+    return E_FILE_FORMAT_INVALID;
+
+  const int lacing = int(m_flags & 0x06) >> 1;
+
+  ++pos;  // consume flags byte
+
+  if (lacing == 0) {  // no lacing
+    if (pos > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    m_frame_count = 1;
+    m_frames = new (std::nothrow) Frame[m_frame_count];
+    if (m_frames == NULL)
+      return -1;
+
+    Frame& f = m_frames[0];
+    f.pos = pos;
+
+    const long long frame_size = stop - pos;
+
+    if (frame_size > LONG_MAX || frame_size <= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    f.len = static_cast<long>(frame_size);
+
+    return 0;  // success
+  }
+
+  if (pos >= stop)
+    return E_FILE_FORMAT_INVALID;
+
+  unsigned char biased_count;
+
+  status = pReader->Read(pos, 1, &biased_count);
+
+  if (status)
+    return E_FILE_FORMAT_INVALID;
+
+  ++pos;  // consume frame count
+  if (pos > stop)
+    return E_FILE_FORMAT_INVALID;
+
+  m_frame_count = int(biased_count) + 1;
+
+  m_frames = new (std::nothrow) Frame[m_frame_count];
+  if (m_frames == NULL)
+    return -1;
+
+  if (!m_frames)
+    return E_FILE_FORMAT_INVALID;
+
+  if (lacing == 1) {  // Xiph
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
+
+    long long size = 0;
+    int frame_count = m_frame_count;
+
+    while (frame_count > 1) {
+      long frame_size = 0;
+
+      for (;;) {
+        unsigned char val;
+
+        if (pos >= stop)
+          return E_FILE_FORMAT_INVALID;
+
+        status = pReader->Read(pos, 1, &val);
+
+        if (status)
+          return E_FILE_FORMAT_INVALID;
+
+        ++pos;  // consume xiph size byte
+
+        frame_size += val;
+
+        if (val < 255)
+          break;
+      }
+
+      Frame& f = *pf++;
+      assert(pf < pf_end);
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = 0;  // patch later
+
+      if (frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      f.len = frame_size;
+      size += frame_size;  // contribution of this frame
+
+      --frame_count;
+    }
+
+    if (pf >= pf_end || pos > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    {
+      Frame& f = *pf++;
+
+      if (pf != pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = 0;  // patch later
+
+      const long long total_size = stop - pos;
+
+      if (total_size < size)
+        return E_FILE_FORMAT_INVALID;
+
+      const long long frame_size = total_size - size;
+
+      if (frame_size > LONG_MAX || frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      f.len = static_cast<long>(frame_size);
+    }
+
+    pf = m_frames;
+    while (pf != pf_end) {
+      Frame& f = *pf++;
+      assert((pos + f.len) <= stop);
+
+      if ((pos + f.len) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = pos;
+      pos += f.len;
+    }
+
+    assert(pos == stop);
+    if (pos != stop)
+      return E_FILE_FORMAT_INVALID;
+
+  } else if (lacing == 2) {  // fixed-size lacing
+    if (pos >= stop)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long total_size = stop - pos;
+
+    if ((total_size % m_frame_count) != 0)
+      return E_FILE_FORMAT_INVALID;
+
+    const long long frame_size = total_size / m_frame_count;
+
+    if (frame_size > LONG_MAX || frame_size <= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
+
+    while (pf != pf_end) {
+      assert((pos + frame_size) <= stop);
+      if ((pos + frame_size) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      Frame& f = *pf++;
+
+      f.pos = pos;
+      f.len = static_cast<long>(frame_size);
+
+      pos += frame_size;
+    }
+
+    assert(pos == stop);
+    if (pos != stop)
+      return E_FILE_FORMAT_INVALID;
+
+  } else {
+    assert(lacing == 3);  // EBML lacing
+
+    if (pos >= stop)
+      return E_FILE_FORMAT_INVALID;
+
+    long long size = 0;
+    int frame_count = m_frame_count;
+
+    long long frame_size = ReadUInt(pReader, pos, len);
+
+    if (frame_size <= 0)
+      return E_FILE_FORMAT_INVALID;
+
+    if (frame_size > LONG_MAX)
+      return E_FILE_FORMAT_INVALID;
+
+    if ((pos + len) > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    pos += len;  // consume length of size of first frame
+
+    if ((pos + frame_size) > stop)
+      return E_FILE_FORMAT_INVALID;
+
+    Frame* pf = m_frames;
+    Frame* const pf_end = pf + m_frame_count;
+
+    {
+      Frame& curr = *pf;
+
+      curr.pos = 0;  // patch later
+
+      curr.len = static_cast<long>(frame_size);
+      size += curr.len;  // contribution of this frame
+    }
+
+    --frame_count;
+
+    while (frame_count > 1) {
+      if (pos >= stop)
+        return E_FILE_FORMAT_INVALID;
+
+      assert(pf < pf_end);
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      const Frame& prev = *pf++;
+      assert(prev.len == frame_size);
+      if (prev.len != frame_size)
+        return E_FILE_FORMAT_INVALID;
+
+      assert(pf < pf_end);
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      Frame& curr = *pf;
+
+      curr.pos = 0;  // patch later
+
+      const long long delta_size_ = ReadUInt(pReader, pos, len);
+
+      if (delta_size_ < 0)
+        return E_FILE_FORMAT_INVALID;
+
+      if ((pos + len) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      pos += len;  // consume length of (delta) size
+      if (pos > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      const long exp = 7 * len - 1;
+      const long long bias = (1LL << exp) - 1LL;
+      const long long delta_size = delta_size_ - bias;
+
+      frame_size += delta_size;
+
+      if (frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      if (frame_size > LONG_MAX)
+        return E_FILE_FORMAT_INVALID;
+
+      curr.len = static_cast<long>(frame_size);
+      size += curr.len;  // contribution of this frame
+
+      --frame_count;
+    }
+
+    // parse last frame
+    if (frame_count > 0) {
+      if (pos > stop || pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      const Frame& prev = *pf++;
+      assert(prev.len == frame_size);
+      if (prev.len != frame_size)
+        return E_FILE_FORMAT_INVALID;
+
+      if (pf >= pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      Frame& curr = *pf++;
+      if (pf != pf_end)
+        return E_FILE_FORMAT_INVALID;
+
+      curr.pos = 0;  // patch later
+
+      const long long total_size = stop - pos;
+
+      if (total_size < size)
+        return E_FILE_FORMAT_INVALID;
+
+      frame_size = total_size - size;
+
+      if (frame_size > LONG_MAX || frame_size <= 0)
+        return E_FILE_FORMAT_INVALID;
+
+      curr.len = static_cast<long>(frame_size);
+    }
+
+    pf = m_frames;
+    while (pf != pf_end) {
+      Frame& f = *pf++;
+      assert((pos + f.len) <= stop);
+      if ((pos + f.len) > stop)
+        return E_FILE_FORMAT_INVALID;
+
+      f.pos = pos;
+      pos += f.len;
+    }
+
+    if (pos != stop)
+      return E_FILE_FORMAT_INVALID;
+  }
+
+  return 0;  // success
+}
+
+long long Block::GetTimeCode(const Cluster* pCluster) const {
+  if (pCluster == 0)
+    return m_timecode;
+
+  const long long tc0 = pCluster->GetTimeCode();
+  assert(tc0 >= 0);
+
+  const long long tc = tc0 + m_timecode;
+
+  return tc;  // unscaled timecode units
+}
+
+long long Block::GetTime(const Cluster* pCluster) const {
+  assert(pCluster);
+
+  const long long tc = GetTimeCode(pCluster);
+
+  const Segment* const pSegment = pCluster->m_pSegment;
+  const SegmentInfo* const pInfo = pSegment->GetInfo();
+  assert(pInfo);
+
+  const long long scale = pInfo->GetTimeCodeScale();
+  assert(scale >= 1);
+
+  const long long ns = tc * scale;
+
+  return ns;
+}
+
+long long Block::GetTrackNumber() const { return m_track; }
+
+bool Block::IsKey() const {
+  return ((m_flags & static_cast<unsigned char>(1 << 7)) != 0);
+}
+
+void Block::SetKey(bool bKey) {
+  if (bKey)
+    m_flags |= static_cast<unsigned char>(1 << 7);
+  else
+    m_flags &= 0x7F;
+}
+
+bool Block::IsInvisible() const { return bool(int(m_flags & 0x08) != 0); }
+
+Block::Lacing Block::GetLacing() const {
+  const int value = int(m_flags & 0x06) >> 1;
+  return static_cast<Lacing>(value);
+}
+
+int Block::GetFrameCount() const { return m_frame_count; }
+
+const Block::Frame& Block::GetFrame(int idx) const {
+  assert(idx >= 0);
+  assert(idx < m_frame_count);
+
+  const Frame& f = m_frames[idx];
+  assert(f.pos > 0);
+  assert(f.len > 0);
+
+  return f;
+}
+
+long Block::Frame::Read(IMkvReader* pReader, unsigned char* buf) const {
+  assert(pReader);
+  assert(buf);
+
+  const long status = pReader->Read(pos, len, buf);
+  return status;
+}
+
+long long Block::GetDiscardPadding() const { return m_discard_padding; }
+
+}  // namespace mkvparser
diff --git a/third_party/aom/third_party/libwebm/mkvparser/mkvparser.h b/third_party/aom/third_party/libwebm/mkvparser/mkvparser.h
new file mode 100644
index 0000000000..42e6e88ab4
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvparser/mkvparser.h
@@ -0,0 +1,1112 @@
+// Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef MKVPARSER_MKVPARSER_H_
+#define MKVPARSER_MKVPARSER_H_
+
+#include <cstddef>
+
+namespace mkvparser {
+
+const int E_PARSE_FAILED = -1;
+const int E_FILE_FORMAT_INVALID = -2;
+const int E_BUFFER_NOT_FULL = -3;
+
+class IMkvReader {
+ public:
+  virtual int Read(long long pos, long len, unsigned char* buf) = 0;
+  virtual int Length(long long* total, long long* available) = 0;
+
+ protected:
+  virtual ~IMkvReader();
+};
+
+template <typename Type>
+Type* SafeArrayAlloc(unsigned long long num_elements,
+                     unsigned long long element_size);
+long long GetUIntLength(IMkvReader*, long long, long&);
+long long ReadUInt(IMkvReader*, long long, long&);
+long long ReadID(IMkvReader* pReader, long long pos, long& len);
+long long UnserializeUInt(IMkvReader*, long long pos, long long size);
+
+long UnserializeFloat(IMkvReader*, long long pos, long long size, double&);
+long UnserializeInt(IMkvReader*, long long pos, long long size,
+                    long long& result);
+
+long UnserializeString(IMkvReader*, long long pos, long long size, char*& str);
+
+long ParseElementHeader(IMkvReader* pReader,
+                        long long& pos,  // consume id and size fields
+                        long long stop,  // if you know size of element's parent
+                        long long& id, long long& size);
+
+bool Match(IMkvReader*, long long&, unsigned long, long long&);
+bool Match(IMkvReader*, long long&, unsigned long, unsigned char*&, size_t&);
+
+void GetVersion(int& major, int& minor, int& build, int& revision);
+
+struct EBMLHeader {
+  EBMLHeader();
+  ~EBMLHeader();
+  long long m_version;
+  long long m_readVersion;
+  long long m_maxIdLength;
+  long long m_maxSizeLength;
+  char* m_docType;
+  long long m_docTypeVersion;
+  long long m_docTypeReadVersion;
+
+  long long Parse(IMkvReader*, long long&);
+  void Init();
+};
+
+class Segment;
+class Track;
+class Cluster;
+
+class Block {
+  Block(const Block&);
+  Block& operator=(const Block&);
+
+ public:
+  const long long m_start;
+  const long long m_size;
+
+  Block(long long start, long long size, long long discard_padding);
+  ~Block();
+
+  long Parse(const Cluster*);
+
+  long long GetTrackNumber() const;
+  long long GetTimeCode(const Cluster*) const;  // absolute, but not scaled
+  long long GetTime(const Cluster*) const;  // absolute, and scaled (ns)
+  bool IsKey() const;
+  void SetKey(bool);
+  bool IsInvisible() const;
+
+  enum Lacing { kLacingNone, kLacingXiph, kLacingFixed, kLacingEbml };
+  Lacing GetLacing() const;
+
+  int GetFrameCount() const;  // to index frames: [0, count)
+
+  struct Frame {
+    long long pos;  // absolute offset
+    long len;
+
+    long Read(IMkvReader*, unsigned char*) const;
+  };
+
+  const Frame& GetFrame(int frame_index) const;
+
+  long long GetDiscardPadding() const;
+
+ private:
+  long long m_track;  // Track::Number()
+  short m_timecode;  // relative to cluster
+  unsigned char m_flags;
+
+  Frame* m_frames;
+  int m_frame_count;
+
+ protected:
+  const long long m_discard_padding;
+};
+
+class BlockEntry {
+  BlockEntry(const BlockEntry&);
+  BlockEntry& operator=(const BlockEntry&);
+
+ protected:
+  BlockEntry(Cluster*, long index);
+
+ public:
+  virtual ~BlockEntry();
+
+  bool EOS() const { return (GetKind() == kBlockEOS); }
+  const Cluster* GetCluster() const;
+  long GetIndex() const;
+  virtual const Block* GetBlock() const = 0;
+
+  enum Kind { kBlockEOS, kBlockSimple, kBlockGroup };
+  virtual Kind GetKind() const = 0;
+
+ protected:
+  Cluster* const m_pCluster;
+  const long m_index;
+};
+
+class SimpleBlock : public BlockEntry {
+  SimpleBlock(const SimpleBlock&);
+  SimpleBlock& operator=(const SimpleBlock&);
+
+ public:
+  SimpleBlock(Cluster*, long index, long long start, long long size);
+  long Parse();
+
+  Kind GetKind() const;
+  const Block* GetBlock() const;
+
+ protected:
+  Block m_block;
+};
+
+class BlockGroup : public BlockEntry {
+  BlockGroup(const BlockGroup&);
+  BlockGroup& operator=(const BlockGroup&);
+
+ public:
+  BlockGroup(Cluster*, long index,
+             long long block_start,  // absolute pos of block's payload
+             long long block_size,  // size of block's payload
+             long long prev, long long next, long long duration,
+             long long discard_padding);
+
+  long Parse();
+
+  Kind GetKind() const;
+  const Block* GetBlock() const;
+
+  long long GetPrevTimeCode() const;  // relative to block's time
+  long long GetNextTimeCode() const;  // as above
+  long long GetDurationTimeCode() const;
+
+ private:
+  Block m_block;
+  const long long m_prev;
+  const long long m_next;
+  const long long m_duration;
+};
+
+///////////////////////////////////////////////////////////////
+// ContentEncoding element
+// Elements used to describe if the track data has been encrypted or
+// compressed with zlib or header stripping.
+class ContentEncoding {
+ public:
+  enum { kCTR = 1 };
+
+  ContentEncoding();
+  ~ContentEncoding();
+
+  // ContentCompression element names
+  struct ContentCompression {
+    ContentCompression();
+    ~ContentCompression();
+
+    unsigned long long algo;
+    unsigned char* settings;
+    long long settings_len;
+  };
+
+  // ContentEncAESSettings element names
+  struct ContentEncAESSettings {
+    ContentEncAESSettings() : cipher_mode(kCTR) {}
+    ~ContentEncAESSettings() {}
+
+    unsigned long long cipher_mode;
+  };
+
+  // ContentEncryption element names
+  struct ContentEncryption {
+    ContentEncryption();
+    ~ContentEncryption();
+
+    unsigned long long algo;
+    unsigned char* key_id;
+    long long key_id_len;
+    unsigned char* signature;
+    long long signature_len;
+    unsigned char* sig_key_id;
+    long long sig_key_id_len;
+    unsigned long long sig_algo;
+    unsigned long long sig_hash_algo;
+
+    ContentEncAESSettings aes_settings;
+  };
+
+  // Returns ContentCompression represented by |idx|. Returns NULL if |idx|
+  // is out of bounds.
+  const ContentCompression* GetCompressionByIndex(unsigned long idx) const;
+
+  // Returns number of ContentCompression elements in this ContentEncoding
+  // element.
+  unsigned long GetCompressionCount() const;
+
+  // Parses the ContentCompression element from |pReader|. |start| is the
+  // starting offset of the ContentCompression payload. |size| is the size in
+  // bytes of the ContentCompression payload. |compression| is where the parsed
+  // values will be stored.
+  long ParseCompressionEntry(long long start, long long size,
+                             IMkvReader* pReader,
+                             ContentCompression* compression);
+
+  // Returns ContentEncryption represented by |idx|. Returns NULL if |idx|
+  // is out of bounds.
+  const ContentEncryption* GetEncryptionByIndex(unsigned long idx) const;
+
+  // Returns number of ContentEncryption elements in this ContentEncoding
+  // element.
+  unsigned long GetEncryptionCount() const;
+
+  // Parses the ContentEncAESSettings element from |pReader|. |start| is the
+  // starting offset of the ContentEncAESSettings payload. |size| is the
+  // size in bytes of the ContentEncAESSettings payload. |encryption| is
+  // where the parsed values will be stored.
+  long ParseContentEncAESSettingsEntry(long long start, long long size,
+                                       IMkvReader* pReader,
+                                       ContentEncAESSettings* aes);
+
+  // Parses the ContentEncoding element from |pReader|. |start| is the
+  // starting offset of the ContentEncoding payload. |size| is the size in
+  // bytes of the ContentEncoding payload. Returns true on success.
+  long ParseContentEncodingEntry(long long start, long long size,
+                                 IMkvReader* pReader);
+
+  // Parses the ContentEncryption element from |pReader|. |start| is the
+  // starting offset of the ContentEncryption payload. |size| is the size in
+  // bytes of the ContentEncryption payload. |encryption| is where the parsed
+  // values will be stored.
+  long ParseEncryptionEntry(long long start, long long size,
+                            IMkvReader* pReader, ContentEncryption* encryption);
+
+  unsigned long long encoding_order() const { return encoding_order_; }
+  unsigned long long encoding_scope() const { return encoding_scope_; }
+  unsigned long long encoding_type() const { return encoding_type_; }
+
+ private:
+  // Member variables for list of ContentCompression elements.
+  ContentCompression** compression_entries_;
+  ContentCompression** compression_entries_end_;
+
+  // Member variables for list of ContentEncryption elements.
+  ContentEncryption** encryption_entries_;
+  ContentEncryption** encryption_entries_end_;
+
+  // ContentEncoding element names
+  unsigned long long encoding_order_;
+  unsigned long long encoding_scope_;
+  unsigned long long encoding_type_;
+
+  // LIBWEBM_DISALLOW_COPY_AND_ASSIGN(ContentEncoding);
+  ContentEncoding(const ContentEncoding&);
+  ContentEncoding& operator=(const ContentEncoding&);
+};
+
+class Track {
+  Track(const Track&);
+  Track& operator=(const Track&);
+
+ public:
+  class Info;
+  static long Create(Segment*, const Info&, long long element_start,
+                     long long element_size, Track*&);
+
+  enum Type { kVideo = 1, kAudio = 2, kSubtitle = 0x11, kMetadata = 0x21 };
+
+  Segment* const m_pSegment;
+  const long long m_element_start;
+  const long long m_element_size;
+  virtual ~Track();
+
+  long GetType() const;
+  long GetNumber() const;
+  unsigned long long GetUid() const;
+  const char* GetNameAsUTF8() const;
+  const char* GetLanguage() const;
+  const char* GetCodecNameAsUTF8() const;
+  const char* GetCodecId() const;
+  const unsigned char* GetCodecPrivate(size_t&) const;
+  bool GetLacing() const;
+  unsigned long long GetDefaultDuration() const;
+  unsigned long long GetCodecDelay() const;
+  unsigned long long GetSeekPreRoll() const;
+
+  const BlockEntry* GetEOS() const;
+
+  struct Settings {
+    long long start;
+    long long size;
+  };
+
+  class Info {
+   public:
+    Info();
+    ~Info();
+    int Copy(Info&) const;
+    void Clear();
+    long type;
+    long number;
+    unsigned long long uid;
+    unsigned long long defaultDuration;
+    unsigned long long codecDelay;
+    unsigned long long seekPreRoll;
+    char* nameAsUTF8;
+    char* language;
+    char* codecId;
+    char* codecNameAsUTF8;
+    unsigned char* codecPrivate;
+    size_t codecPrivateSize;
+    bool lacing;
+    Settings settings;
+
+   private:
+    Info(const Info&);
+    Info& operator=(const Info&);
+    int CopyStr(char* Info::*str, Info&) const;
+  };
+
+  long GetFirst(const BlockEntry*&) const;
+  long GetNext(const BlockEntry* pCurr, const BlockEntry*& pNext) const;
+  virtual bool VetEntry(const BlockEntry*) const;
+  virtual long Seek(long long time_ns, const BlockEntry*&) const;
+
+  const ContentEncoding* GetContentEncodingByIndex(unsigned long idx) const;
+  unsigned long GetContentEncodingCount() const;
+
+  long ParseContentEncodingsEntry(long long start, long long size);
+
+ protected:
+  Track(Segment*, long long element_start, long long element_size);
+
+  Info m_info;
+
+  class EOSBlock : public BlockEntry {
+   public:
+    EOSBlock();
+
+    Kind GetKind() const;
+    const Block* GetBlock() const;
+  };
+
+  EOSBlock m_eos;
+
+ private:
+  ContentEncoding** content_encoding_entries_;
+  ContentEncoding** content_encoding_entries_end_;
+};
+
+struct PrimaryChromaticity {
+  PrimaryChromaticity() : x(0), y(0) {}
+  ~PrimaryChromaticity() {}
+  static bool Parse(IMkvReader* reader, long long read_pos,
+                    long long value_size, bool is_x,
+                    PrimaryChromaticity** chromaticity);
+  float x;
+  float y;
+};
+
+struct MasteringMetadata {
+  static const float kValueNotPresent;
+
+  MasteringMetadata()
+      : r(NULL),
+        g(NULL),
+        b(NULL),
+        white_point(NULL),
+        luminance_max(kValueNotPresent),
+        luminance_min(kValueNotPresent) {}
+  ~MasteringMetadata() {
+    delete r;
+    delete g;
+    delete b;
+    delete white_point;
+  }
+
+  static bool Parse(IMkvReader* reader, long long element_start,
+                    long long element_size,
+                    MasteringMetadata** mastering_metadata);
+
+  PrimaryChromaticity* r;
+  PrimaryChromaticity* g;
+  PrimaryChromaticity* b;
+  PrimaryChromaticity* white_point;
+  float luminance_max;
+  float luminance_min;
+};
+
+struct Colour {
+  static const long long kValueNotPresent;
+
+  // Unless otherwise noted all values assigned upon construction are the
+  // equivalent of unspecified/default.
+  Colour()
+      : matrix_coefficients(kValueNotPresent),
+        bits_per_channel(kValueNotPresent),
+        chroma_subsampling_horz(kValueNotPresent),
+        chroma_subsampling_vert(kValueNotPresent),
+        cb_subsampling_horz(kValueNotPresent),
+        cb_subsampling_vert(kValueNotPresent),
+        chroma_siting_horz(kValueNotPresent),
+        chroma_siting_vert(kValueNotPresent),
+        range(kValueNotPresent),
+        transfer_characteristics(kValueNotPresent),
+        primaries(kValueNotPresent),
+        max_cll(kValueNotPresent),
+        max_fall(kValueNotPresent),
+        mastering_metadata(NULL) {}
+  ~Colour() {
+    delete mastering_metadata;
+    mastering_metadata = NULL;
+  }
+
+  static bool Parse(IMkvReader* reader, long long element_start,
+                    long long element_size, Colour** colour);
+
+  long long matrix_coefficients;
+  long long bits_per_channel;
+  long long chroma_subsampling_horz;
+  long long chroma_subsampling_vert;
+  long long cb_subsampling_horz;
+  long long cb_subsampling_vert;
+  long long chroma_siting_horz;
+  long long chroma_siting_vert;
+  long long range;
+  long long transfer_characteristics;
+  long long primaries;
+  long long max_cll;
+  long long max_fall;
+
+  MasteringMetadata* mastering_metadata;
+};
+
+class VideoTrack : public Track {
+  VideoTrack(const VideoTrack&);
+  VideoTrack& operator=(const VideoTrack&);
+
+  VideoTrack(Segment*, long long element_start, long long element_size);
+
+ public:
+  virtual ~VideoTrack();
+  static long Parse(Segment*, const Info&, long long element_start,
+                    long long element_size, VideoTrack*&);
+
+  long long GetWidth() const;
+  long long GetHeight() const;
+  long long GetDisplayWidth() const;
+  long long GetDisplayHeight() const;
+  long long GetDisplayUnit() const;
+  long long GetStereoMode() const;
+  double GetFrameRate() const;
+
+  bool VetEntry(const BlockEntry*) const;
+  long Seek(long long time_ns, const BlockEntry*&) const;
+
+  Colour* GetColour() const;
+
+ private:
+  long long m_width;
+  long long m_height;
+  long long m_display_width;
+  long long m_display_height;
+  long long m_display_unit;
+  long long m_stereo_mode;
+
+  double m_rate;
+
+  Colour* m_colour;
+};
+
+class AudioTrack : public Track {
+  AudioTrack(const AudioTrack&);
+  AudioTrack& operator=(const AudioTrack&);
+
+  AudioTrack(Segment*, long long element_start, long long element_size);
+
+ public:
+  static long Parse(Segment*, const Info&, long long element_start,
+                    long long element_size, AudioTrack*&);
+
+  double GetSamplingRate() const;
+  long long GetChannels() const;
+  long long GetBitDepth() const;
+
+ private:
+  double m_rate;
+  long long m_channels;
+  long long m_bitDepth;
+};
+
+class Tracks {
+  Tracks(const Tracks&);
+  Tracks& operator=(const Tracks&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  Tracks(Segment*, long long start, long long size, long long element_start,
+         long long element_size);
+
+  ~Tracks();
+
+  long Parse();
+
+  unsigned long GetTracksCount() const;
+
+  const Track* GetTrackByNumber(long tn) const;
+  const Track* GetTrackByIndex(unsigned long idx) const;
+
+ private:
+  Track** m_trackEntries;
+  Track** m_trackEntriesEnd;
+
+  long ParseTrackEntry(long long payload_start, long long payload_size,
+                       long long element_start, long long element_size,
+                       Track*&) const;
+};
+
+class Chapters {
+  Chapters(const Chapters&);
+  Chapters& operator=(const Chapters&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  Chapters(Segment*, long long payload_start, long long payload_size,
+           long long element_start, long long element_size);
+
+  ~Chapters();
+
+  long Parse();
+
+  class Atom;
+  class Edition;
+
+  class Display {
+    friend class Atom;
+    Display();
+    Display(const Display&);
+    ~Display();
+    Display& operator=(const Display&);
+
+   public:
+    const char* GetString() const;
+    const char* GetLanguage() const;
+    const char* GetCountry() const;
+
+   private:
+    void Init();
+    void ShallowCopy(Display&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    char* m_string;
+    char* m_language;
+    char* m_country;
+  };
+
+  class Atom {
+    friend class Edition;
+    Atom();
+    Atom(const Atom&);
+    ~Atom();
+    Atom& operator=(const Atom&);
+
+   public:
+    unsigned long long GetUID() const;
+    const char* GetStringUID() const;
+
+    long long GetStartTimecode() const;
+    long long GetStopTimecode() const;
+
+    long long GetStartTime(const Chapters*) const;
+    long long GetStopTime(const Chapters*) const;
+
+    int GetDisplayCount() const;
+    const Display* GetDisplay(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Atom&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+    static long long GetTime(const Chapters*, long long timecode);
+
+    long ParseDisplay(IMkvReader*, long long pos, long long size);
+    bool ExpandDisplaysArray();
+
+    char* m_string_uid;
+    unsigned long long m_uid;
+    long long m_start_timecode;
+    long long m_stop_timecode;
+
+    Display* m_displays;
+    int m_displays_size;
+    int m_displays_count;
+  };
+
+  class Edition {
+    friend class Chapters;
+    Edition();
+    Edition(const Edition&);
+    ~Edition();
+    Edition& operator=(const Edition&);
+
+   public:
+    int GetAtomCount() const;
+    const Atom* GetAtom(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Edition&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    long ParseAtom(IMkvReader*, long long pos, long long size);
+    bool ExpandAtomsArray();
+
+    Atom* m_atoms;
+    int m_atoms_size;
+    int m_atoms_count;
+  };
+
+  int GetEditionCount() const;
+  const Edition* GetEdition(int index) const;
+
+ private:
+  long ParseEdition(long long pos, long long size);
+  bool ExpandEditionsArray();
+
+  Edition* m_editions;
+  int m_editions_size;
+  int m_editions_count;
+};
+
+class Tags {
+  Tags(const Tags&);
+  Tags& operator=(const Tags&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  Tags(Segment*, long long payload_start, long long payload_size,
+       long long element_start, long long element_size);
+
+  ~Tags();
+
+  long Parse();
+
+  class Tag;
+  class SimpleTag;
+
+  class SimpleTag {
+    friend class Tag;
+    SimpleTag();
+    SimpleTag(const SimpleTag&);
+    ~SimpleTag();
+    SimpleTag& operator=(const SimpleTag&);
+
+   public:
+    const char* GetTagName() const;
+    const char* GetTagString() const;
+
+   private:
+    void Init();
+    void ShallowCopy(SimpleTag&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    char* m_tag_name;
+    char* m_tag_string;
+  };
+
+  class Tag {
+    friend class Tags;
+    Tag();
+    Tag(const Tag&);
+    ~Tag();
+    Tag& operator=(const Tag&);
+
+   public:
+    int GetSimpleTagCount() const;
+    const SimpleTag* GetSimpleTag(int index) const;
+
+   private:
+    void Init();
+    void ShallowCopy(Tag&) const;
+    void Clear();
+    long Parse(IMkvReader*, long long pos, long long size);
+
+    long ParseSimpleTag(IMkvReader*, long long pos, long long size);
+    bool ExpandSimpleTagsArray();
+
+    SimpleTag* m_simple_tags;
+    int m_simple_tags_size;
+    int m_simple_tags_count;
+  };
+
+  int GetTagCount() const;
+  const Tag* GetTag(int index) const;
+
+ private:
+  long ParseTag(long long pos, long long size);
+  bool ExpandTagsArray();
+
+  Tag* m_tags;
+  int m_tags_size;
+  int m_tags_count;
+};
+
+class SegmentInfo {
+  SegmentInfo(const SegmentInfo&);
+  SegmentInfo& operator=(const SegmentInfo&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  SegmentInfo(Segment*, long long start, long long size,
+              long long element_start, long long element_size);
+
+  ~SegmentInfo();
+
+  long Parse();
+
+  long long GetTimeCodeScale() const;
+  long long GetDuration() const;  // scaled
+  const char* GetMuxingAppAsUTF8() const;
+  const char* GetWritingAppAsUTF8() const;
+  const char* GetTitleAsUTF8() const;
+
+ private:
+  long long m_timecodeScale;
+  double m_duration;
+  char* m_pMuxingAppAsUTF8;
+  char* m_pWritingAppAsUTF8;
+  char* m_pTitleAsUTF8;
+};
+
+class SeekHead {
+  SeekHead(const SeekHead&);
+  SeekHead& operator=(const SeekHead&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  SeekHead(Segment*, long long start, long long size, long long element_start,
+           long long element_size);
+
+  ~SeekHead();
+
+  long Parse();
+
+  struct Entry {
+    // the SeekHead entry payload
+    long long id;
+    long long pos;
+
+    // absolute pos of SeekEntry ID
+    long long element_start;
+
+    // SeekEntry ID size + size size + payload
+    long long element_size;
+  };
+
+  int GetCount() const;
+  const Entry* GetEntry(int idx) const;
+
+  struct VoidElement {
+    // absolute pos of Void ID
+    long long element_start;
+
+    // ID size + size size + payload size
+    long long element_size;
+  };
+
+  int GetVoidElementCount() const;
+  const VoidElement* GetVoidElement(int idx) const;
+
+ private:
+  Entry* m_entries;
+  int m_entry_count;
+
+  VoidElement* m_void_elements;
+  int m_void_element_count;
+
+  static bool ParseEntry(IMkvReader*,
+                         long long pos,  // payload
+                         long long size, Entry*);
+};
+
+class Cues;
+class CuePoint {
+  friend class Cues;
+
+  CuePoint(long, long long);
+  ~CuePoint();
+
+  CuePoint(const CuePoint&);
+  CuePoint& operator=(const CuePoint&);
+
+ public:
+  long long m_element_start;
+  long long m_element_size;
+
+  bool Load(IMkvReader*);
+
+  long long GetTimeCode() const;  // absolute but unscaled
+  long long GetTime(const Segment*) const;  // absolute and scaled (ns units)
+
+  struct TrackPosition {
+    long long m_track;
+    long long m_pos;  // of cluster
+    long long m_block;
+    // codec_state  //defaults to 0
+    // reference = clusters containing req'd referenced blocks
+    //  reftime = timecode of the referenced block
+
+    bool Parse(IMkvReader*, long long, long long);
+  };
+
+  const TrackPosition* Find(const Track*) const;
+
+ private:
+  const long m_index;
+  long long m_timecode;
+  TrackPosition* m_track_positions;
+  size_t m_track_positions_count;
+};
+
+class Cues {
+  friend class Segment;
+
+  Cues(Segment*, long long start, long long size, long long element_start,
+       long long element_size);
+  ~Cues();
+
+  Cues(const Cues&);
+  Cues& operator=(const Cues&);
+
+ public:
+  Segment* const m_pSegment;
+  const long long m_start;
+  const long long m_size;
+  const long long m_element_start;
+  const long long m_element_size;
+
+  bool Find(  // lower bound of time_ns
+      long long time_ns, const Track*, const CuePoint*&,
+      const CuePoint::TrackPosition*&) const;
+
+  const CuePoint* GetFirst() const;
+  const CuePoint* GetLast() const;
+  const CuePoint* GetNext(const CuePoint*) const;
+
+  const BlockEntry* GetBlock(const CuePoint*,
+                             const CuePoint::TrackPosition*) const;
+
+  bool LoadCuePoint() const;
+  long GetCount() const;  // loaded only
+  // long GetTotal() const;  //loaded + preloaded
+  bool DoneParsing() const;
+
+ private:
+  bool Init() const;
+  bool PreloadCuePoint(long&, long long) const;
+
+  mutable CuePoint** m_cue_points;
+  mutable long m_count;
+  mutable long m_preload_count;
+  mutable long long m_pos;
+};
+
+class Cluster {
+  friend class Segment;
+
+  Cluster(const Cluster&);
+  Cluster& operator=(const Cluster&);
+
+ public:
+  Segment* const m_pSegment;
+
+ public:
+  static Cluster* Create(Segment*,
+                         long index,  // index in segment
+                         long long off);  // offset relative to segment
+  // long long element_size);
+
+  Cluster();  // EndOfStream
+  ~Cluster();
+
+  bool EOS() const;
+
+  long long GetTimeCode() const;  // absolute, but not scaled
+  long long GetTime() const;  // absolute, and scaled (nanosecond units)
+  long long GetFirstTime() const;  // time (ns) of first (earliest) block
+  long long GetLastTime() const;  // time (ns) of last (latest) block
+
+  long GetFirst(const BlockEntry*&) const;
+  long GetLast(const BlockEntry*&) const;
+  long GetNext(const BlockEntry* curr, const BlockEntry*& next) const;
+
+  const BlockEntry* GetEntry(const Track*, long long ns = -1) const;
+  const BlockEntry* GetEntry(const CuePoint&,
+                             const CuePoint::TrackPosition&) const;
+  // const BlockEntry* GetMaxKey(const VideoTrack*) const;
+
+  //    static bool HasBlockEntries(const Segment*, long long);
+
+  static long HasBlockEntries(const Segment*, long long idoff, long long& pos,
+                              long& size);
+
+  long GetEntryCount() const;
+
+  long Load(long long& pos, long& size) const;
+
+  long Parse(long long& pos, long& size) const;
+  long GetEntry(long index, const mkvparser::BlockEntry*&) const;
+
+ protected:
+  Cluster(Segment*, long index, long long element_start);
+  // long long element_size);
+
+ public:
+  const long long m_element_start;
+  long long GetPosition() const;  // offset relative to segment
+
+  long GetIndex() const;
+  long long GetElementSize() const;
+  // long long GetPayloadSize() const;
+
+  // long long Unparsed() const;
+
+ private:
+  long m_index;
+  mutable long long m_pos;
+  // mutable long long m_size;
+  mutable long long m_element_size;
+  mutable long long m_timecode;
+  mutable BlockEntry** m_entries;
+  mutable long m_entries_size;
+  mutable long m_entries_count;
+
+  long ParseSimpleBlock(long long, long long&, long&);
+  long ParseBlockGroup(long long, long long&, long&);
+
+  long CreateBlock(long long id, long long pos, long long size,
+                   long long discard_padding);
+  long CreateBlockGroup(long long start_offset, long long size,
+                        long long discard_padding);
+  long CreateSimpleBlock(long long, long long);
+};
+
+class Segment {
+  friend class Cues;
+  friend class Track;
+  friend class VideoTrack;
+
+  Segment(const Segment&);
+  Segment& operator=(const Segment&);
+
+ private:
+  Segment(IMkvReader*, long long elem_start,
+          // long long elem_size,
+          long long pos, long long size);
+
+ public:
+  IMkvReader* const m_pReader;
+  const long long m_element_start;
+  // const long long m_element_size;
+  const long long m_start;  // posn of segment payload
+  const long long m_size;  // size of segment payload
+  Cluster m_eos;  // TODO: make private?
+
+  static long long CreateInstance(IMkvReader*, long long, Segment*&);
+  ~Segment();
+
+  long Load();  // loads headers and all clusters
+
+  // for incremental loading
+  // long long Unparsed() const;
+  bool DoneParsing() const;
+  long long ParseHeaders();  // stops when first cluster is found
+  // long FindNextCluster(long long& pos, long& size) const;
+  long LoadCluster(long long& pos, long& size);  // load one cluster
+  long LoadCluster();
+
+  long ParseNext(const Cluster* pCurr, const Cluster*& pNext, long long& pos,
+                 long& size);
+
+  const SeekHead* GetSeekHead() const;
+  const Tracks* GetTracks() const;
+  const SegmentInfo* GetInfo() const;
+  const Cues* GetCues() const;
+  const Chapters* GetChapters() const;
+  const Tags* GetTags() const;
+
+  long long GetDuration() const;
+
+  unsigned long GetCount() const;
+  const Cluster* GetFirst() const;
+  const Cluster* GetLast() const;
+  const Cluster* GetNext(const Cluster*);
+
+  const Cluster* FindCluster(long long time_nanoseconds) const;
+  // const BlockEntry* Seek(long long time_nanoseconds, const Track*) const;
+
+  const Cluster* FindOrPreloadCluster(long long pos);
+
+  long ParseCues(long long cues_off,  // offset relative to start of segment
+                 long long& parse_pos, long& parse_len);
+
+ private:
+  long long m_pos;  // absolute file posn; what has been consumed so far
+  Cluster* m_pUnknownSize;
+
+  SeekHead* m_pSeekHead;
+  SegmentInfo* m_pInfo;
+  Tracks* m_pTracks;
+  Cues* m_pCues;
+  Chapters* m_pChapters;
+  Tags* m_pTags;
+  Cluster** m_clusters;
+  long m_clusterCount;  // number of entries for which m_index >= 0
+  long m_clusterPreloadCount;  // number of entries for which m_index < 0
+  long m_clusterSize;  // array size
+
+  long DoLoadCluster(long long&, long&);
+  long DoLoadClusterUnknownSize(long long&, long&);
+  long DoParseNext(const Cluster*&, long long&, long&);
+
+  bool AppendCluster(Cluster*);
+  bool PreloadCluster(Cluster*, ptrdiff_t);
+
+  // void ParseSeekHead(long long pos, long long size);
+  // void ParseSeekEntry(long long pos, long long size);
+  // void ParseCues(long long);
+
+  const BlockEntry* GetBlock(const CuePoint&, const CuePoint::TrackPosition&);
+};
+
+}  // namespace mkvparser
+
+inline long mkvparser::Segment::LoadCluster() {
+  long long pos;
+  long size;
+
+  return LoadCluster(pos, size);
+}
+
+#endif  // MKVPARSER_MKVPARSER_H_
diff --git a/third_party/aom/third_party/libwebm/mkvparser/mkvreader.cc b/third_party/aom/third_party/libwebm/mkvparser/mkvreader.cc
new file mode 100644
index 0000000000..9f90d8c4f8
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvparser/mkvreader.cc
@@ -0,0 +1,131 @@
+// Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#include "mkvparser/mkvreader.h"
+
+#include <cassert>
+
+namespace mkvparser {
+
+MkvReader::MkvReader() : m_file(NULL), reader_owns_file_(true) {}
+
+MkvReader::MkvReader(FILE* fp) : m_file(fp), reader_owns_file_(false) {
+  GetFileSize();
+}
+
+MkvReader::~MkvReader() {
+  if (reader_owns_file_)
+    Close();
+  m_file = NULL;
+}
+
+int MkvReader::Open(const char* fileName) {
+  if (fileName == NULL)
+    return -1;
+
+  if (m_file)
+    return -1;
+
+#ifdef _MSC_VER
+  const errno_t e = fopen_s(&m_file, fileName, "rb");
+
+  if (e)
+    return -1;  // error
+#else
+  m_file = fopen(fileName, "rb");
+
+  if (m_file == NULL)
+    return -1;
+#endif
+  return !GetFileSize();
+}
+
+bool MkvReader::GetFileSize() {
+  if (m_file == NULL)
+    return false;
+#ifdef _MSC_VER
+  int status = _fseeki64(m_file, 0L, SEEK_END);
+
+  if (status)
+    return false;  // error
+
+  m_length = _ftelli64(m_file);
+#else
+  fseek(m_file, 0L, SEEK_END);
+  m_length = ftell(m_file);
+#endif
+  assert(m_length >= 0);
+
+  if (m_length < 0)
+    return false;
+
+#ifdef _MSC_VER
+  status = _fseeki64(m_file, 0L, SEEK_SET);
+
+  if (status)
+    return false;  // error
+#else
+  fseek(m_file, 0L, SEEK_SET);
+#endif
+
+  return true;
+}
+
+void MkvReader::Close() {
+  if (m_file != NULL) {
+    fclose(m_file);
+    m_file = NULL;
+  }
+}
+
+int MkvReader::Length(long long* total, long long* available) {
+  if (m_file == NULL)
+    return -1;
+
+  if (total)
+    *total = m_length;
+
+  if (available)
+    *available = m_length;
+
+  return 0;
+}
+
+int MkvReader::Read(long long offset, long len, unsigned char* buffer) {
+  if (m_file == NULL)
+    return -1;
+
+  if (offset < 0)
+    return -1;
+
+  if (len < 0)
+    return -1;
+
+  if (len == 0)
+    return 0;
+
+  if (offset >= m_length)
+    return -1;
+
+#ifdef _MSC_VER
+  const int status = _fseeki64(m_file, offset, SEEK_SET);
+
+  if (status)
+    return -1;  // error
+#else
+  fseek(m_file, offset, SEEK_SET);
+#endif
+
+  const size_t size = fread(buffer, 1, len, m_file);
+
+  if (size < size_t(len))
+    return -1;  // error
+
+  return 0;  // success
+}
+
+}  // namespace mkvparser
\ No newline at end of file
diff --git a/third_party/aom/third_party/libwebm/mkvparser/mkvreader.h b/third_party/aom/third_party/libwebm/mkvparser/mkvreader.h
new file mode 100644
index 0000000000..9831ecf645
--- /dev/null
+++ b/third_party/aom/third_party/libwebm/mkvparser/mkvreader.h
@@ -0,0 +1,45 @@
+// Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+//
+// Use of this source code is governed by a BSD-style license
+// that can be found in the LICENSE file in the root of the source
+// tree. An additional intellectual property rights grant can be found
+// in the file PATENTS.  All contributing project authors may
+// be found in the AUTHORS file in the root of the source tree.
+#ifndef MKVPARSER_MKVREADER_H_
+#define MKVPARSER_MKVREADER_H_
+
+#include <cstdio>
+
+#include "mkvparser/mkvparser.h"
+
+namespace mkvparser {
+
+class MkvReader : public IMkvReader {
+ public:
+  MkvReader();
+  explicit MkvReader(FILE* fp);
+  virtual ~MkvReader();
+
+  int Open(const char*);
+  void Close();
+
+  virtual int Read(long long position, long length, unsigned char* buffer);
+  virtual int Length(long long* total, long long* available);
+
+ private:
+  MkvReader(const MkvReader&);
+  MkvReader& operator=(const MkvReader&);
+
+  // Determines the size of the file. This is called either by the constructor
+  // or by the Open function depending on file ownership. Returns true on
+  // success.
+  bool GetFileSize();
+
+  long long m_length;
+  FILE* m_file;
+  bool reader_owns_file_;
+};
+
+}  // namespace mkvparser
+
+#endif  // MKVPARSER_MKVREADER_H_
diff --git a/third_party/aom/third_party/libyuv/README.libaom b/third_party/aom/third_party/libyuv/README.libaom
new file mode 100644
index 0000000000..09693c1f2c
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/README.libaom
@@ -0,0 +1,15 @@
+Name: libyuv
+URL: http://code.google.com/p/libyuv/
+Version: 1456
+License: BSD
+License File: LICENSE
+
+Description:
+libyuv is an open source project that includes YUV conversion and scaling
+functionality.
+
+The optimized scaler in libyuv is used in multiple resolution encoder example,
+which down-samples the original input video (f.g. 1280x720) a number of times
+in order to encode multiple resolution bit streams.
+
+Local Modifications:
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/basic_types.h b/third_party/aom/third_party/libyuv/include/libyuv/basic_types.h
new file mode 100644
index 0000000000..66e68536cb
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/basic_types.h
@@ -0,0 +1,119 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_BASIC_TYPES_H_  // NOLINT
+#define INCLUDE_LIBYUV_BASIC_TYPES_H_
+
+#include <stddef.h>  // for NULL, size_t
+
+#if defined(__ANDROID__) || (defined(_MSC_VER) && (_MSC_VER < 1600))
+#include <sys/types.h>  // for uintptr_t on x86
+#else
+#include <stdint.h>  // for uintptr_t
+#endif
+
+#ifndef GG_LONGLONG
+#ifndef INT_TYPES_DEFINED
+#define INT_TYPES_DEFINED
+#ifdef COMPILER_MSVC
+typedef unsigned __int64 uint64;
+typedef __int64 int64;
+#ifndef INT64_C
+#define INT64_C(x) x ## I64
+#endif
+#ifndef UINT64_C
+#define UINT64_C(x) x ## UI64
+#endif
+#define INT64_F "I64"
+#else  // COMPILER_MSVC
+#if defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long uint64;  // NOLINT
+typedef long int64;  // NOLINT
+#ifndef INT64_C
+#define INT64_C(x) x ## L
+#endif
+#ifndef UINT64_C
+#define UINT64_C(x) x ## UL
+#endif
+#define INT64_F "l"
+#else  // defined(__LP64__) && !defined(__OpenBSD__) && !defined(__APPLE__)
+typedef unsigned long long uint64;  // NOLINT
+typedef long long int64;  // NOLINT
+#ifndef INT64_C
+#define INT64_C(x) x ## LL
+#endif
+#ifndef UINT64_C
+#define UINT64_C(x) x ## ULL
+#endif
+#define INT64_F "ll"
+#endif  // __LP64__
+#endif  // COMPILER_MSVC
+typedef unsigned int uint32;
+typedef int int32;
+typedef unsigned short uint16;  // NOLINT
+typedef short int16;  // NOLINT
+typedef unsigned char uint8;
+typedef signed char int8;
+#endif  // INT_TYPES_DEFINED
+#endif  // GG_LONGLONG
+
+// Detect compiler is for x86 or x64.
+#if defined(__x86_64__) || defined(_M_X64) || \
+    defined(__i386__) || defined(_M_IX86)
+#define CPU_X86 1
+#endif
+// Detect compiler is for ARM.
+#if defined(__arm__) || defined(_M_ARM)
+#define CPU_ARM 1
+#endif
+
+#ifndef ALIGNP
+#ifdef __cplusplus
+#define ALIGNP(p, t) \
+    (reinterpret_cast<uint8*>(((reinterpret_cast<uintptr_t>(p) + \
+    ((t) - 1)) & ~((t) - 1))))
+#else
+#define ALIGNP(p, t) \
+    ((uint8*)((((uintptr_t)(p) + ((t) - 1)) & ~((t) - 1))))  /* NOLINT */
+#endif
+#endif
+
+#if !defined(LIBYUV_API)
+#if defined(_WIN32) || defined(__CYGWIN__)
+#if defined(LIBYUV_BUILDING_SHARED_LIBRARY)
+#define LIBYUV_API __declspec(dllexport)
+#elif defined(LIBYUV_USING_SHARED_LIBRARY)
+#define LIBYUV_API __declspec(dllimport)
+#else
+#define LIBYUV_API
+#endif  // LIBYUV_BUILDING_SHARED_LIBRARY
+#elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__APPLE__) && \
+    (defined(LIBYUV_BUILDING_SHARED_LIBRARY) || \
+    defined(LIBYUV_USING_SHARED_LIBRARY))
+#define LIBYUV_API __attribute__ ((visibility ("default")))
+#else
+#define LIBYUV_API
+#endif  // __GNUC__
+#endif  // LIBYUV_API
+
+#define LIBYUV_BOOL int
+#define LIBYUV_FALSE 0
+#define LIBYUV_TRUE 1
+
+// Visual C x86 or GCC little endian.
+#if defined(__x86_64__) || defined(_M_X64) || \
+  defined(__i386__) || defined(_M_IX86) || \
+  defined(__arm__) || defined(_M_ARM) || \
+  (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
+#define LIBYUV_LITTLE_ENDIAN
+#endif
+
+#endif  // INCLUDE_LIBYUV_BASIC_TYPES_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/compare.h b/third_party/aom/third_party/libyuv/include/libyuv/compare.h
new file mode 100644
index 0000000000..2a9f1560ce
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/compare.h
@@ -0,0 +1,79 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_COMPARE_H_  // NOLINT
+#define INCLUDE_LIBYUV_COMPARE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Compute a hash for specified memory. Seed of 5381 recommended.
+LIBYUV_API
+uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed);
+
+// Scan an opaque argb image and return fourcc based on alpha offset.
+// Returns FOURCC_ARGB, FOURCC_BGRA, or 0 if unknown.
+LIBYUV_API
+uint32 ARGBDetect(const uint8* argb, int stride_argb, int width, int height);
+
+// Sum Square Error - used to compute Mean Square Error or PSNR.
+LIBYUV_API
+uint64 ComputeSumSquareError(const uint8* src_a,
+                             const uint8* src_b, int count);
+
+LIBYUV_API
+uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
+                                  const uint8* src_b, int stride_b,
+                                  int width, int height);
+
+static const int kMaxPsnr = 128;
+
+LIBYUV_API
+double SumSquareErrorToPsnr(uint64 sse, uint64 count);
+
+LIBYUV_API
+double CalcFramePsnr(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height);
+
+LIBYUV_API
+double I420Psnr(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height);
+
+LIBYUV_API
+double CalcFrameSsim(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height);
+
+LIBYUV_API
+double I420Ssim(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_COMPARE_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/convert.h b/third_party/aom/third_party/libyuv/include/libyuv/convert.h
new file mode 100644
index 0000000000..d6f206c10f
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/convert.h
@@ -0,0 +1,246 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_H_
+
+#include "libyuv/basic_types.h"
+// TODO(fbarchard): Remove the following headers includes.
+#include "libyuv/convert_from.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert I444 to I420.
+LIBYUV_API
+int I444ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I422 to I420.
+LIBYUV_API
+int I422ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert I411 to I420.
+LIBYUV_API
+int I411ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Copy I420 to I420.
+#define I420ToI420 I420Copy
+LIBYUV_API
+int I420Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Convert I400 (grey) to I420.
+LIBYUV_API
+int I400ToI420(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+#define J400ToJ420 I400ToI420
+
+// Convert NV12 to I420.
+LIBYUV_API
+int NV12ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert NV21 to I420.
+LIBYUV_API
+int NV21ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert YUY2 to I420.
+LIBYUV_API
+int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert UYVY to I420.
+LIBYUV_API
+int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert M420 to I420.
+LIBYUV_API
+int M420ToI420(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// ARGB little endian (bgra in memory) to I420.
+LIBYUV_API
+int ARGBToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// BGRA little endian (argb in memory) to I420.
+LIBYUV_API
+int BGRAToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// ABGR little endian (rgba in memory) to I420.
+LIBYUV_API
+int ABGRToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// RGBA little endian (abgr in memory) to I420.
+LIBYUV_API
+int RGBAToI420(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// RGB little endian (bgr in memory) to I420.
+LIBYUV_API
+int RGB24ToI420(const uint8* src_frame, int src_stride_frame,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height);
+
+// RGB big endian (rgb in memory) to I420.
+LIBYUV_API
+int RAWToI420(const uint8* src_frame, int src_stride_frame,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int width, int height);
+
+// RGB16 (RGBP fourcc) little endian to I420.
+LIBYUV_API
+int RGB565ToI420(const uint8* src_frame, int src_stride_frame,
+                 uint8* dst_y, int dst_stride_y,
+                 uint8* dst_u, int dst_stride_u,
+                 uint8* dst_v, int dst_stride_v,
+                 int width, int height);
+
+// RGB15 (RGBO fourcc) little endian to I420.
+LIBYUV_API
+int ARGB1555ToI420(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height);
+
+// RGB12 (R444 fourcc) little endian to I420.
+LIBYUV_API
+int ARGB4444ToI420(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height);
+
+#ifdef HAVE_JPEG
+// src_width/height provided by capture.
+// dst_width/height for clipping determine final size.
+LIBYUV_API
+int MJPGToI420(const uint8* sample, size_t sample_size,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int src_width, int src_height,
+               int dst_width, int dst_height);
+
+// Query size of MJPG in pixels.
+LIBYUV_API
+int MJPGSize(const uint8* sample, size_t sample_size,
+             int* width, int* height);
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// "src_size" is needed to parse MJPG.
+// "dst_stride_y" number of bytes in a row of the dst_y plane.
+//   Normally this would be the same as dst_width, with recommended alignment
+//   to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected. The caller should
+//   allocate the I420 buffer according to rotation.
+// "dst_stride_u" number of bytes in a row of the dst_u plane.
+//   Normally this would be the same as (dst_width + 1) / 2, with
+//   recommended alignment to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected.
+// "crop_x" and "crop_y" are starting position for cropping.
+//   To center, crop_x = (src_width - dst_width) / 2
+//              crop_y = (src_height - dst_height) / 2
+// "src_width" / "src_height" is size of src_frame in pixels.
+//   "src_height" can be negative indicating a vertically flipped image source.
+// "crop_width" / "crop_height" is the size to crop the src to.
+//    Must be less than or equal to src_width/src_height
+//    Cropping parameters are pre-rotation.
+// "rotation" can be 0, 90, 180 or 270.
+// "format" is a fourcc. ie 'I420', 'YUY2'
+// Returns 0 for successful; -1 for invalid parameter. Non-zero for failure.
+LIBYUV_API
+int ConvertToI420(const uint8* src_frame, size_t src_size,
+                  uint8* dst_y, int dst_stride_y,
+                  uint8* dst_u, int dst_stride_u,
+                  uint8* dst_v, int dst_stride_v,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/convert_argb.h b/third_party/aom/third_party/libyuv/include/libyuv/convert_argb.h
new file mode 100644
index 0000000000..ea75c0b26a
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/convert_argb.h
@@ -0,0 +1,232 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_ARGB_H_
+
+#include "libyuv/basic_types.h"
+// TODO(fbarchard): Remove the following headers includes
+#include "libyuv/convert_from.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+
+// TODO(fbarchard): This set of functions should exactly match convert.h
+// TODO(fbarchard): Add tests. Create random content of right size and convert
+// with C vs Opt and or to I420 and compare.
+// TODO(fbarchard): Some of these functions lack parameter setting.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Alias.
+#define ARGBToARGB ARGBCopy
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Convert I420 to ARGB.
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I422 to ARGB.
+LIBYUV_API
+int I422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I444 to ARGB.
+LIBYUV_API
+int I444ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I411 to ARGB.
+LIBYUV_API
+int I411ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert I400 (grey) to ARGB.  Reverse of ARGBToI400.
+LIBYUV_API
+int I400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert J400 (jpeg grey) to ARGB.
+LIBYUV_API
+int J400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Alias.
+#define YToARGB I400ToARGB
+
+// Convert NV12 to ARGB.
+LIBYUV_API
+int NV12ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert NV21 to ARGB.
+LIBYUV_API
+int NV21ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert M420 to ARGB.
+LIBYUV_API
+int M420ToARGB(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert YUY2 to ARGB.
+LIBYUV_API
+int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert UYVY to ARGB.
+LIBYUV_API
+int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert J420 to ARGB.
+LIBYUV_API
+int J420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert J422 to ARGB.
+LIBYUV_API
+int J422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// BGRA little endian (argb in memory) to ARGB.
+LIBYUV_API
+int BGRAToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// ABGR little endian (rgba in memory) to ARGB.
+LIBYUV_API
+int ABGRToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// RGBA little endian (abgr in memory) to ARGB.
+LIBYUV_API
+int RGBAToARGB(const uint8* src_frame, int src_stride_frame,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Deprecated function name.
+#define BG24ToARGB RGB24ToARGB
+
+// RGB little endian (bgr in memory) to ARGB.
+LIBYUV_API
+int RGB24ToARGB(const uint8* src_frame, int src_stride_frame,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height);
+
+// RGB big endian (rgb in memory) to ARGB.
+LIBYUV_API
+int RAWToARGB(const uint8* src_frame, int src_stride_frame,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// RGB16 (RGBP fourcc) little endian to ARGB.
+LIBYUV_API
+int RGB565ToARGB(const uint8* src_frame, int src_stride_frame,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// RGB15 (RGBO fourcc) little endian to ARGB.
+LIBYUV_API
+int ARGB1555ToARGB(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height);
+
+// RGB12 (R444 fourcc) little endian to ARGB.
+LIBYUV_API
+int ARGB4444ToARGB(const uint8* src_frame, int src_stride_frame,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height);
+
+#ifdef HAVE_JPEG
+// src_width/height provided by capture
+// dst_width/height for clipping determine final size.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample, size_t sample_size,
+               uint8* dst_argb, int dst_stride_argb,
+               int src_width, int src_height,
+               int dst_width, int dst_height);
+#endif
+
+// Convert camera sample to ARGB with cropping, rotation and vertical flip.
+// "src_size" is needed to parse MJPG.
+// "dst_stride_argb" number of bytes in a row of the dst_argb plane.
+//   Normally this would be the same as dst_width, with recommended alignment
+//   to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected. The caller should
+//   allocate the I420 buffer according to rotation.
+// "dst_stride_u" number of bytes in a row of the dst_u plane.
+//   Normally this would be the same as (dst_width + 1) / 2, with
+//   recommended alignment to 16 bytes for better efficiency.
+//   If rotation of 90 or 270 is used, stride is affected.
+// "crop_x" and "crop_y" are starting position for cropping.
+//   To center, crop_x = (src_width - dst_width) / 2
+//              crop_y = (src_height - dst_height) / 2
+// "src_width" / "src_height" is size of src_frame in pixels.
+//   "src_height" can be negative indicating a vertically flipped image source.
+// "crop_width" / "crop_height" is the size to crop the src to.
+//    Must be less than or equal to src_width/src_height
+//    Cropping parameters are pre-rotation.
+// "rotation" can be 0, 90, 180 or 270.
+// "format" is a fourcc. ie 'I420', 'YUY2'
+// Returns 0 for successful; -1 for invalid parameter. Non-zero for failure.
+LIBYUV_API
+int ConvertToARGB(const uint8* src_frame, size_t src_size,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_ARGB_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/convert_from.h b/third_party/aom/third_party/libyuv/include/libyuv/convert_from.h
new file mode 100644
index 0000000000..3591b4fd6a
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/convert_from.h
@@ -0,0 +1,182 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_FROM_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_FROM_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/rotate.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// See Also convert.h for conversions from formats to I420.
+
+// I420Copy in convert to I420ToI420.
+
+LIBYUV_API
+int I420ToI422(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+LIBYUV_API
+int I420ToI444(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+LIBYUV_API
+int I420ToI411(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Copy to I400. Source can be I420, I422, I444, I400, NV12 or NV21.
+LIBYUV_API
+int I400Copy(const uint8* src_y, int src_stride_y,
+             uint8* dst_y, int dst_stride_y,
+             int width, int height);
+
+// TODO(fbarchard): I420ToM420
+
+LIBYUV_API
+int I420ToNV12(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+LIBYUV_API
+int I420ToNV21(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+LIBYUV_API
+int I420ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+LIBYUV_API
+int I420ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+LIBYUV_API
+int I420ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+LIBYUV_API
+int I420ToRGB24(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_frame, int dst_stride_frame,
+                int width, int height);
+
+LIBYUV_API
+int I420ToRAW(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              uint8* dst_frame, int dst_stride_frame,
+              int width, int height);
+
+LIBYUV_API
+int I420ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_u, int src_stride_u,
+                 const uint8* src_v, int src_stride_v,
+                 uint8* dst_frame, int dst_stride_frame,
+                 int width, int height);
+
+// Convert I420 To RGB565 with 4x4 dither matrix (16 bytes).
+// Values in dither matrix from 0 to 7 recommended.
+// The order of the dither matrix is first byte is upper left.
+
+LIBYUV_API
+int I420ToRGB565Dither(const uint8* src_y, int src_stride_y,
+                       const uint8* src_u, int src_stride_u,
+                       const uint8* src_v, int src_stride_v,
+                       uint8* dst_frame, int dst_stride_frame,
+                       const uint8* dither4x4, int width, int height);
+
+LIBYUV_API
+int I420ToARGB1555(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_frame, int dst_stride_frame,
+                   int width, int height);
+
+LIBYUV_API
+int I420ToARGB4444(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_frame, int dst_stride_frame,
+                   int width, int height);
+
+// Convert I420 to specified format.
+// "dst_sample_stride" is bytes in a row for the destination. Pass 0 if the
+//    buffer has contiguous rows. Can be negative. A multiple of 16 is optimal.
+LIBYUV_API
+int ConvertFromI420(const uint8* y, int y_stride,
+                    const uint8* u, int u_stride,
+                    const uint8* v, int v_stride,
+                    uint8* dst_sample, int dst_sample_stride,
+                    int width, int height,
+                    uint32 format);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_FROM_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/convert_from_argb.h b/third_party/aom/third_party/libyuv/include/libyuv/convert_from_argb.h
new file mode 100644
index 0000000000..4a62268138
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/convert_from_argb.h
@@ -0,0 +1,191 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy ARGB to ARGB.
+#define ARGBToARGB ARGBCopy
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Convert ARGB To BGRA.
+LIBYUV_API
+int ARGBToBGRA(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height);
+
+// Convert ARGB To ABGR.
+LIBYUV_API
+int ARGBToABGR(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height);
+
+// Convert ARGB To RGBA.
+LIBYUV_API
+int ARGBToRGBA(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+// Convert ARGB To RGB24.
+LIBYUV_API
+int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height);
+
+// Convert ARGB To RAW.
+LIBYUV_API
+int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_rgb, int dst_stride_rgb,
+              int width, int height);
+
+// Convert ARGB To RGB565.
+LIBYUV_API
+int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// Convert ARGB To RGB565 with 4x4 dither matrix (16 bytes).
+// Values in dither matrix from 0 to 7 recommended.
+// The order of the dither matrix is first byte is upper left.
+// TODO(fbarchard): Consider pointer to 2d array for dither4x4.
+// const uint8(*dither)[4][4];
+LIBYUV_API
+int ARGBToRGB565Dither(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_rgb565, int dst_stride_rgb565,
+                       const uint8* dither4x4, int width, int height);
+
+// Convert ARGB To ARGB1555.
+LIBYUV_API
+int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height);
+
+// Convert ARGB To ARGB4444.
+LIBYUV_API
+int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height);
+
+// Convert ARGB To I444.
+LIBYUV_API
+int ARGBToI444(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I422.
+LIBYUV_API
+int ARGBToI422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I420. (also in convert.h)
+LIBYUV_API
+int ARGBToI420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB to J420. (JPeg full range I420).
+LIBYUV_API
+int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB to J422.
+LIBYUV_API
+int ARGBToJ422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB To I411.
+LIBYUV_API
+int ARGBToI411(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert ARGB to J400. (JPeg full range).
+LIBYUV_API
+int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               int width, int height);
+
+// Convert ARGB to I400.
+LIBYUV_API
+int ARGBToI400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Convert ARGB to G. (Reverse of J400toARGB, which replicates G back to ARGB)
+LIBYUV_API
+int ARGBToG(const uint8* src_argb, int src_stride_argb,
+            uint8* dst_g, int dst_stride_g,
+            int width, int height);
+
+// Convert ARGB To NV12.
+LIBYUV_API
+int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+// Convert ARGB To NV21.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+// Convert ARGB To NV21.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height);
+
+// Convert ARGB To YUY2.
+LIBYUV_API
+int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height);
+
+// Convert ARGB To UYVY.
+LIBYUV_API
+int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CONVERT_FROM_ARGB_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/cpu_id.h b/third_party/aom/third_party/libyuv/include/libyuv/cpu_id.h
new file mode 100644
index 0000000000..870e94e8cd
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/cpu_id.h
@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_CPU_ID_H_  // NOLINT
+#define INCLUDE_LIBYUV_CPU_ID_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// TODO(fbarchard): Consider overlapping bits for different architectures.
+// Internal flag to indicate cpuid requires initialization.
+#define kCpuInit 0x1
+
+// These flags are only valid on ARM processors.
+static const int kCpuHasARM = 0x2;
+static const int kCpuHasNEON = 0x4;
+// 0x8 reserved for future ARM flag.
+
+// These flags are only valid on x86 processors.
+static const int kCpuHasX86 = 0x10;
+static const int kCpuHasSSE2 = 0x20;
+static const int kCpuHasSSSE3 = 0x40;
+static const int kCpuHasSSE41 = 0x80;
+static const int kCpuHasSSE42 = 0x100;
+static const int kCpuHasAVX = 0x200;
+static const int kCpuHasAVX2 = 0x400;
+static const int kCpuHasERMS = 0x800;
+static const int kCpuHasFMA3 = 0x1000;
+// 0x2000, 0x4000, 0x8000 reserved for future X86 flags.
+
+// These flags are only valid on MIPS processors.
+static const int kCpuHasMIPS = 0x10000;
+static const int kCpuHasMIPS_DSP = 0x20000;
+static const int kCpuHasMIPS_DSPR2 = 0x40000;
+
+// Internal function used to auto-init.
+LIBYUV_API
+int InitCpuFlags(void);
+
+// Internal function for parsing /proc/cpuinfo.
+LIBYUV_API
+int ArmCpuCaps(const char* cpuinfo_name);
+
+// Detect CPU has SSE2 etc.
+// Test_flag parameter should be one of kCpuHas constants above.
+// returns non-zero if instruction set is detected
+static __inline int TestCpuFlag(int test_flag) {
+  LIBYUV_API extern int cpu_info_;
+  return (cpu_info_ == kCpuInit ? InitCpuFlags() : cpu_info_) & test_flag;
+}
+
+// For testing, allow CPU flags to be disabled.
+// ie MaskCpuFlags(~kCpuHasSSSE3) to disable SSSE3.
+// MaskCpuFlags(-1) to enable all cpu specific optimizations.
+// MaskCpuFlags(0) to disable all cpu specific optimizations.
+LIBYUV_API
+void MaskCpuFlags(int enable_flags);
+
+// Low level cpuid for X86. Returns zeros on other CPUs.
+// eax is the info type that you want.
+// ecx is typically the cpu number, and should normally be zero.
+LIBYUV_API
+void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_CPU_ID_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/mjpeg_decoder.h b/third_party/aom/third_party/libyuv/include/libyuv/mjpeg_decoder.h
new file mode 100644
index 0000000000..fa1e51f9ac
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/mjpeg_decoder.h
@@ -0,0 +1,193 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_MJPEG_DECODER_H_  // NOLINT
+#define INCLUDE_LIBYUV_MJPEG_DECODER_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+// NOTE: For a simplified public API use convert.h MJPGToI420().
+
+struct jpeg_common_struct;
+struct jpeg_decompress_struct;
+struct jpeg_source_mgr;
+
+namespace libyuv {
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+static const uint32 kUnknownDataSize = 0xFFFFFFFF;
+
+enum JpegSubsamplingType {
+  kJpegYuv420,
+  kJpegYuv422,
+  kJpegYuv411,
+  kJpegYuv444,
+  kJpegYuv400,
+  kJpegUnknown
+};
+
+struct Buffer {
+  const uint8* data;
+  int len;
+};
+
+struct BufferVector {
+  Buffer* buffers;
+  int len;
+  int pos;
+};
+
+struct SetJmpErrorMgr;
+
+// MJPEG ("Motion JPEG") is a pseudo-standard video codec where the frames are
+// simply independent JPEG images with a fixed huffman table (which is omitted).
+// It is rarely used in video transmission, but is common as a camera capture
+// format, especially in Logitech devices. This class implements a decoder for
+// MJPEG frames.
+//
+// See http://tools.ietf.org/html/rfc2435
+class LIBYUV_API MJpegDecoder {
+ public:
+  typedef void (*CallbackFunction)(void* opaque,
+                                   const uint8* const* data,
+                                   const int* strides,
+                                   int rows);
+
+  static const int kColorSpaceUnknown;
+  static const int kColorSpaceGrayscale;
+  static const int kColorSpaceRgb;
+  static const int kColorSpaceYCbCr;
+  static const int kColorSpaceCMYK;
+  static const int kColorSpaceYCCK;
+
+  MJpegDecoder();
+  ~MJpegDecoder();
+
+  // Loads a new frame, reads its headers, and determines the uncompressed
+  // image format.
+  // Returns LIBYUV_TRUE if image looks valid and format is supported.
+  // If return value is LIBYUV_TRUE, then the values for all the following
+  // getters are populated.
+  // src_len is the size of the compressed mjpeg frame in bytes.
+  LIBYUV_BOOL LoadFrame(const uint8* src, size_t src_len);
+
+  // Returns width of the last loaded frame in pixels.
+  int GetWidth();
+
+  // Returns height of the last loaded frame in pixels.
+  int GetHeight();
+
+  // Returns format of the last loaded frame. The return value is one of the
+  // kColorSpace* constants.
+  int GetColorSpace();
+
+  // Number of color components in the color space.
+  int GetNumComponents();
+
+  // Sample factors of the n-th component.
+  int GetHorizSampFactor(int component);
+
+  int GetVertSampFactor(int component);
+
+  int GetHorizSubSampFactor(int component);
+
+  int GetVertSubSampFactor(int component);
+
+  // Public for testability.
+  int GetImageScanlinesPerImcuRow();
+
+  // Public for testability.
+  int GetComponentScanlinesPerImcuRow(int component);
+
+  // Width of a component in bytes.
+  int GetComponentWidth(int component);
+
+  // Height of a component.
+  int GetComponentHeight(int component);
+
+  // Width of a component in bytes with padding for DCTSIZE. Public for testing.
+  int GetComponentStride(int component);
+
+  // Size of a component in bytes.
+  int GetComponentSize(int component);
+
+  // Call this after LoadFrame() if you decide you don't want to decode it
+  // after all.
+  LIBYUV_BOOL UnloadFrame();
+
+  // Decodes the entire image into a one-buffer-per-color-component format.
+  // dst_width must match exactly. dst_height must be <= to image height; if
+  // less, the image is cropped. "planes" must have size equal to at least
+  // GetNumComponents() and they must point to non-overlapping buffers of size
+  // at least GetComponentSize(i). The pointers in planes are incremented
+  // to point to after the end of the written data.
+  // TODO(fbarchard): Add dst_x, dst_y to allow specific rect to be decoded.
+  LIBYUV_BOOL DecodeToBuffers(uint8** planes, int dst_width, int dst_height);
+
+  // Decodes the entire image and passes the data via repeated calls to a
+  // callback function. Each call will get the data for a whole number of
+  // image scanlines.
+  // TODO(fbarchard): Add dst_x, dst_y to allow specific rect to be decoded.
+  LIBYUV_BOOL DecodeToCallback(CallbackFunction fn, void* opaque,
+                        int dst_width, int dst_height);
+
+  // The helper function which recognizes the jpeg sub-sampling type.
+  static JpegSubsamplingType JpegSubsamplingTypeHelper(
+     int* subsample_x, int* subsample_y, int number_of_components);
+
+ private:
+  void AllocOutputBuffers(int num_outbufs);
+  void DestroyOutputBuffers();
+
+  LIBYUV_BOOL StartDecode();
+  LIBYUV_BOOL FinishDecode();
+
+  void SetScanlinePointers(uint8** data);
+  LIBYUV_BOOL DecodeImcuRow();
+
+  int GetComponentScanlinePadding(int component);
+
+  // A buffer holding the input data for a frame.
+  Buffer buf_;
+  BufferVector buf_vec_;
+
+  jpeg_decompress_struct* decompress_struct_;
+  jpeg_source_mgr* source_mgr_;
+  SetJmpErrorMgr* error_mgr_;
+
+  // LIBYUV_TRUE iff at least one component has scanline padding. (i.e.,
+  // GetComponentScanlinePadding() != 0.)
+  LIBYUV_BOOL has_scanline_padding_;
+
+  // Temporaries used to point to scanline outputs.
+  int num_outbufs_;  // Outermost size of all arrays below.
+  uint8*** scanlines_;
+  int* scanlines_sizes_;
+  // Temporary buffer used for decoding when we can't decode directly to the
+  // output buffers. Large enough for just one iMCU row.
+  uint8** databuf_;
+  int* databuf_strides_;
+};
+
+}  // namespace libyuv
+
+#endif  //  __cplusplus
+#endif  // INCLUDE_LIBYUV_MJPEG_DECODER_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/planar_functions.h b/third_party/aom/third_party/libyuv/include/libyuv/planar_functions.h
new file mode 100644
index 0000000000..7fe4d8eedd
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/planar_functions.h
@@ -0,0 +1,454 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  // NOLINT
+#define INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_
+
+#include "libyuv/basic_types.h"
+
+// TODO(fbarchard): Remove the following headers includes.
+#include "libyuv/convert.h"
+#include "libyuv/convert_argb.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy a plane of data.
+LIBYUV_API
+void CopyPlane(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+LIBYUV_API
+void CopyPlane_16(const uint16* src_y, int src_stride_y,
+                  uint16* dst_y, int dst_stride_y,
+                  int width, int height);
+
+// Set a plane of data to a 32 bit value.
+LIBYUV_API
+void SetPlane(uint8* dst_y, int dst_stride_y,
+              int width, int height,
+              uint32 value);
+
+// Copy I400.  Supports inverting.
+LIBYUV_API
+int I400ToI400(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+#define J400ToJ400 I400ToI400
+
+// Copy I422 to I422.
+#define I422ToI422 I422Copy
+LIBYUV_API
+int I422Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Copy I444 to I444.
+#define I444ToI444 I444Copy
+LIBYUV_API
+int I444Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height);
+
+// Convert YUY2 to I422.
+LIBYUV_API
+int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Convert UYVY to I422.
+LIBYUV_API
+int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+LIBYUV_API
+int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+LIBYUV_API
+int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height);
+
+// Convert I420 to I400. (calls CopyPlane ignoring u/v).
+LIBYUV_API
+int I420ToI400(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Alias
+#define J420ToJ400 I420ToI400
+#define I420ToI420Mirror I420Mirror
+
+// I420 mirror.
+LIBYUV_API
+int I420Mirror(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height);
+
+// Alias
+#define I400ToI400Mirror I400Mirror
+
+// I400 mirror.  A single plane is mirrored horizontally.
+// Pass negative height to achieve 180 degree rotation.
+LIBYUV_API
+int I400Mirror(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height);
+
+// Alias
+#define ARGBToARGBMirror ARGBMirror
+
+// ARGB mirror.
+LIBYUV_API
+int ARGBMirror(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// Convert NV21 to RGB565.
+LIBYUV_API
+int NV21ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height);
+
+// I422ToARGB is in convert_argb.h
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height);
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height);
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height);
+
+// Draw a rectangle into I420.
+LIBYUV_API
+int I420Rect(uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int x, int y, int width, int height,
+             int value_y, int value_u, int value_v);
+
+// Draw a rectangle into ARGB.
+LIBYUV_API
+int ARGBRect(uint8* dst_argb, int dst_stride_argb,
+             int x, int y, int width, int height, uint32 value);
+
+// Convert ARGB to gray scale ARGB.
+LIBYUV_API
+int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height);
+
+// Make a rectangle of ARGB gray scale.
+LIBYUV_API
+int ARGBGray(uint8* dst_argb, int dst_stride_argb,
+             int x, int y, int width, int height);
+
+// Make a rectangle of ARGB Sepia tone.
+LIBYUV_API
+int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
+              int x, int y, int width, int height);
+
+// Apply a matrix rotation to each ARGB pixel.
+// matrix_argb is 4 signed ARGB values. -128 to 127 representing -2 to 2.
+// The first 4 coefficients apply to B, G, R, A and produce B of the output.
+// The next 4 coefficients apply to B, G, R, A and produce G of the output.
+// The next 4 coefficients apply to B, G, R, A and produce R of the output.
+// The last 4 coefficients apply to B, G, R, A and produce A of the output.
+LIBYUV_API
+int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    const int8* matrix_argb,
+                    int width, int height);
+
+// Deprecated. Use ARGBColorMatrix instead.
+// Apply a matrix rotation to each ARGB pixel.
+// matrix_argb is 3 signed ARGB values. -128 to 127 representing -1 to 1.
+// The first 4 coefficients apply to B, G, R, A and produce B of the output.
+// The next 4 coefficients apply to B, G, R, A and produce G of the output.
+// The last 4 coefficients apply to B, G, R, A and produce R of the output.
+LIBYUV_API
+int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
+                   const int8* matrix_rgb,
+                   int x, int y, int width, int height);
+
+// Apply a color table each ARGB pixel.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                   const uint8* table_argb,
+                   int x, int y, int width, int height);
+
+// Apply a color table each ARGB pixel but preserve destination alpha.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                  const uint8* table_argb,
+                  int x, int y, int width, int height);
+
+// Apply a luma/color table each ARGB pixel but preserve destination alpha.
+// Table contains 32768 values indexed by [Y][C] where 7 it 7 bit luma from
+// RGB (YJ style) and C is an 8 bit color component (R, G or B).
+LIBYUV_API
+int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_argb, int dst_stride_argb,
+                       const uint8* luma_rgb_table,
+                       int width, int height);
+
+// Apply a 3 term polynomial to ARGB values.
+// poly points to a 4x4 matrix.  The first row is constants.  The 2nd row is
+// coefficients for b, g, r and a.  The 3rd row is coefficients for b squared,
+// g squared, r squared and a squared.  The 4rd row is coefficients for b to
+// the 3, g to the 3, r to the 3 and a to the 3.  The values are summed and
+// result clamped to 0 to 255.
+// A polynomial approximation can be dirived using software such as 'R'.
+
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height);
+
+// Quantize a rectangle of ARGB. Alpha unaffected.
+// scale is a 16 bit fractional fixed point scaler between 0 and 65535.
+// interval_size should be a value between 1 and 255.
+// interval_offset should be a value between 0 and 255.
+LIBYUV_API
+int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
+                 int scale, int interval_size, int interval_offset,
+                 int x, int y, int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
+// Copy ARGB to ARGB.
+LIBYUV_API
+int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
+                     uint8* dst_argb, int dst_stride_argb,
+                     int width, int height);
+
+typedef void (*ARGBBlendRow)(const uint8* src_argb0, const uint8* src_argb1,
+                             uint8* dst_argb, int width);
+
+// Get function to Alpha Blend ARGB pixels and store to destination.
+LIBYUV_API
+ARGBBlendRow GetARGBBlend();
+
+// Alpha Blend ARGB images and store to destination.
+// Alpha of destination is set to 255.
+LIBYUV_API
+int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
+              const uint8* src_argb1, int src_stride_argb1,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// Multiply ARGB image by ARGB image. Shifted down by 8. Saturates to 255.
+LIBYUV_API
+int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// Add ARGB image with ARGB image. Saturates to 255.
+LIBYUV_API
+int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
+            const uint8* src_argb1, int src_stride_argb1,
+            uint8* dst_argb, int dst_stride_argb,
+            int width, int height);
+
+// Subtract ARGB image (argb1) from ARGB image (argb0). Saturates to 0.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
+// Convert I422 to YUY2.
+LIBYUV_API
+int I422ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+// Convert I422 to UYVY.
+LIBYUV_API
+int I422ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_frame, int dst_stride_frame,
+               int width, int height);
+
+// Convert unattentuated ARGB to preattenuated ARGB.
+LIBYUV_API
+int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height);
+
+// Convert preattentuated ARGB to unattenuated ARGB.
+LIBYUV_API
+int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height);
+
+// Convert MJPG to ARGB.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample, size_t sample_size,
+               uint8* argb, int argb_stride,
+               int w, int h, int dw, int dh);
+
+// Internal function - do not call directly.
+// Computes table of cumulative sum for image where the value is the sum
+// of all values above and to the left of the entry. Used by ARGBBlur.
+LIBYUV_API
+int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
+                             int32* dst_cumsum, int dst_stride32_cumsum,
+                             int width, int height);
+
+// Blur ARGB image.
+// dst_cumsum table of width * (height + 1) * 16 bytes aligned to
+//   16 byte boundary.
+// dst_stride32_cumsum is number of ints in a row (width * 4).
+// radius is number of pixels around the center.  e.g. 1 = 3x3. 2=5x5.
+// Blur is optimized for radius of 5 (11x11) or less.
+LIBYUV_API
+int ARGBBlur(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int32* dst_cumsum, int dst_stride32_cumsum,
+             int width, int height, int radius);
+
+// Multiply ARGB image by ARGB value.
+LIBYUV_API
+int ARGBShade(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height, uint32 value);
+
+// Interpolate between two ARGB images using specified amount of interpolation
+// (0 to 255) and store to destination.
+// 'interpolation' is specified as 8 bit fraction where 0 means 100% src_argb0
+// and 255 means 1% src_argb0 and 99% src_argb1.
+// Internally uses ARGBScale bilinear filtering.
+// Caveat: This function will write up to 16 bytes beyond the end of dst_argb.
+LIBYUV_API
+int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
+                    const uint8* src_argb1, int src_stride_argb1,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height, int interpolation);
+
+#if defined(__pnacl__) || defined(__CLR_VER) || \
+    (defined(__i386__) && !defined(__SSE2__))
+#define LIBYUV_DISABLE_X86
+#endif
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_ARGBAFFINEROW_SSE2
+#endif
+
+// Row function for copying pixels from a source with a slope to a row
+// of destination. Useful for scaling, rotation, mirror, texture mapping.
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width);
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width);
+
+// Shuffle ARGB channel order.  e.g. BGRA to ARGB.
+// shuffler is 16 bytes and must be aligned.
+LIBYUV_API
+int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
+                uint8* dst_argb, int dst_stride_argb,
+                const uint8* shuffler, int width, int height);
+
+// Sobel ARGB effect with planar output.
+LIBYUV_API
+int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_y, int dst_stride_y,
+                     int width, int height);
+
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height);
+
+// Sobel ARGB effect w/ Sobel X, Sobel, Sobel Y in ARGB.
+LIBYUV_API
+int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_PLANAR_FUNCTIONS_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/rotate.h b/third_party/aom/third_party/libyuv/include/libyuv/rotate.h
new file mode 100644
index 0000000000..8a9673f280
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/rotate.h
@@ -0,0 +1,118 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROTATE_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROTATE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Supported rotation.
+typedef enum RotationMode {
+  kRotate0 = 0,  // No rotation.
+  kRotate90 = 90,  // Rotate 90 degrees clockwise.
+  kRotate180 = 180,  // Rotate 180 degrees.
+  kRotate270 = 270,  // Rotate 270 degrees clockwise.
+
+  // Deprecated.
+  kRotateNone = 0,
+  kRotateClockwise = 90,
+  kRotateCounterClockwise = 270,
+} RotationModeEnum;
+
+// Rotate I420 frame.
+LIBYUV_API
+int I420Rotate(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int src_width, int src_height, enum RotationMode mode);
+
+// Rotate NV12 input and store in I420.
+LIBYUV_API
+int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
+                     const uint8* src_uv, int src_stride_uv,
+                     uint8* dst_y, int dst_stride_y,
+                     uint8* dst_u, int dst_stride_u,
+                     uint8* dst_v, int dst_stride_v,
+                     int src_width, int src_height, enum RotationMode mode);
+
+// Rotate a plane by 0, 90, 180, or 270.
+LIBYUV_API
+int RotatePlane(const uint8* src, int src_stride,
+                uint8* dst, int dst_stride,
+                int src_width, int src_height, enum RotationMode mode);
+
+// Rotate planes by 90, 180, 270. Deprecated.
+LIBYUV_API
+void RotatePlane90(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride,
+                   int width, int height);
+
+LIBYUV_API
+void RotatePlane180(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void RotatePlane270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void RotateUV90(const uint8* src, int src_stride,
+                uint8* dst_a, int dst_stride_a,
+                uint8* dst_b, int dst_stride_b,
+                int width, int height);
+
+// Rotations for when U and V are interleaved.
+// These functions take one input pointer and
+// split the data into two buffers while
+// rotating them. Deprecated.
+LIBYUV_API
+void RotateUV180(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+LIBYUV_API
+void RotateUV270(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+// The 90 and 270 functions are based on transposes.
+// Doing a transpose with reversing the read/write
+// order will result in a rotation by +- 90 degrees.
+// Deprecated.
+LIBYUV_API
+void TransposePlane(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height);
+
+LIBYUV_API
+void TransposeUV(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROTATE_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/rotate_argb.h b/third_party/aom/third_party/libyuv/include/libyuv/rotate_argb.h
new file mode 100644
index 0000000000..2bdc8ec6b4
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/rotate_argb.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROTATE_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROTATE_ARGB_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/rotate.h"  // For RotationMode.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Rotate ARGB frame
+LIBYUV_API
+int ARGBRotate(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int src_width, int src_height, enum RotationMode mode);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROTATE_ARGB_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/rotate_row.h b/third_party/aom/third_party/libyuv/include/libyuv/rotate_row.h
new file mode 100644
index 0000000000..d0bfbdd2b0
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/rotate_row.h
@@ -0,0 +1,139 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROTATE_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROTATE_ROW_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if defined(__pnacl__) || defined(__CLR_VER) || \
+    (defined(__i386__) && !defined(__SSE2__))
+#define LIBYUV_DISABLE_X86
+#endif
+
+// Visual C 2012 required for AVX2.
+#if defined(_M_IX86) && !defined(__clang__) && \
+    defined(_MSC_VER) && _MSC_VER >= 1700
+#define VISUALC_HAS_AVX2 1
+#endif  // VisualStudio >= 2012
+
+// TODO(fbarchard): switch to standard form of inline; fails on clangcl.
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#if defined(__APPLE__) && defined(__i386__)
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".private_extern _" #name "                \n"                             \
+    ".align 4,0x90                             \n"                             \
+"_" #name ":                                   \n"
+#elif defined(__MINGW32__) || defined(__CYGWIN__) && defined(__i386__)
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".align 4,0x90                             \n"                             \
+"_" #name ":                                   \n"
+#else
+#define DECLARE_FUNCTION(name)                                                 \
+    ".text                                     \n"                             \
+    ".align 4,0x90                             \n"                             \
+#name ":                                       \n"
+#endif
+#endif
+
+// The following are available for Visual C:
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
+    defined(_MSC_VER) && !defined(__clang__)
+#define HAS_TRANSPOSEWX8_SSSE3
+#define HAS_TRANSPOSEUVWX8_SSE2
+#endif
+
+// The following are available for GCC but not NaCL:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__i386__) || (defined(__x86_64__) && !defined(__native_client__)))
+#define HAS_TRANSPOSEWX8_SSSE3
+#endif
+
+// The following are available for 32 bit GCC:
+#if !defined(LIBYUV_DISABLE_X86) && defined(__i386__)  && !defined(__clang__)
+#define HAS_TRANSPOSEUVWX8_SSE2
+#endif
+
+// The following are available for 64 bit GCC but not NaCL:
+#if !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \
+    defined(__x86_64__)
+#define HAS_TRANSPOSEWX8_FAST_SSSE3
+#define HAS_TRANSPOSEUVWX8_SSE2
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
+#define HAS_TRANSPOSEWX8_NEON
+#define HAS_TRANSPOSEUVWX8_NEON
+#endif
+
+#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
+    defined(__mips__) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_TRANSPOSEWX8_MIPS_DSPR2
+#define HAS_TRANSPOSEUVWx8_MIPS_DSPR2
+#endif  // defined(__mips__)
+
+void TransposeWxH_C(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride, int width, int height);
+
+void TransposeWx8_C(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride, int width);
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride, int width);
+void TransposeWx8_SSSE3(const uint8* src, int src_stride,
+                        uint8* dst, int dst_stride, int width);
+void TransposeWx8_Fast_SSSE3(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride, int width);
+void TransposeWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride, int width);
+
+void TransposeWx8_Any_NEON(const uint8* src, int src_stride,
+                           uint8* dst, int dst_stride, int width);
+void TransposeWx8_Any_SSSE3(const uint8* src, int src_stride,
+                            uint8* dst, int dst_stride, int width);
+void TransposeWx8_Fast_Any_SSSE3(const uint8* src, int src_stride,
+                                 uint8* dst, int dst_stride, int width);
+void TransposeWx8_Any_MIPS_DSPR2(const uint8* src, int src_stride,
+                                 uint8* dst, int dst_stride, int width);
+
+void TransposeUVWxH_C(const uint8* src, int src_stride,
+                      uint8* dst_a, int dst_stride_a,
+                      uint8* dst_b, int dst_stride_b,
+                      int width, int height);
+
+void TransposeUVWx8_C(const uint8* src, int src_stride,
+                      uint8* dst_a, int dst_stride_a,
+                      uint8* dst_b, int dst_stride_b, int width);
+void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b, int width);
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b, int width);
+void TransposeUVWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                               uint8* dst_a, int dst_stride_a,
+                               uint8* dst_b, int dst_stride_b, int width);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROTATE_ROW_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/row.h b/third_party/aom/third_party/libyuv/include/libyuv/row.h
new file mode 100644
index 0000000000..5c3187ef79
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/row.h
@@ -0,0 +1,1857 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_ROW_H_
+
+#include <stdlib.h>  // For malloc.
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a) - 1)))
+
+#ifdef __cplusplus
+#define align_buffer_64(var, size)                                             \
+  uint8* var##_mem = reinterpret_cast<uint8*>(malloc((size) + 63));            \
+  uint8* var = reinterpret_cast<uint8*>                                        \
+      ((reinterpret_cast<intptr_t>(var##_mem) + 63) & ~63)
+#else
+#define align_buffer_64(var, size)                                             \
+  uint8* var##_mem = (uint8*)(malloc((size) + 63));               /* NOLINT */ \
+  uint8* var = (uint8*)(((intptr_t)(var##_mem) + 63) & ~63)       /* NOLINT */
+#endif
+
+#define free_aligned_buffer_64(var) \
+  free(var##_mem);  \
+  var = 0
+
+#if defined(__pnacl__) || defined(__CLR_VER) || \
+    (defined(__i386__) && !defined(__SSE2__))
+#define LIBYUV_DISABLE_X86
+#endif
+// True if compiling for SSSE3 as a requirement.
+#if defined(__SSSE3__) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 3))
+#define LIBYUV_SSSE3_ONLY
+#endif
+
+#if defined(__native_client__)
+#define LIBYUV_DISABLE_NEON
+#endif
+// clang >= 3.5.0 required for Arm64.
+#if defined(__clang__) && defined(__aarch64__) && !defined(LIBYUV_DISABLE_NEON)
+#if (__clang_major__ < 3) || (__clang_major__ == 3 && (__clang_minor__ < 5))
+#define LIBYUV_DISABLE_NEON
+#endif  // clang >= 3.5
+#endif  // __clang__
+
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+// Conversions:
+#define HAS_ABGRTOUVROW_SSSE3
+#define HAS_ABGRTOYROW_SSSE3
+#define HAS_ARGB1555TOARGBROW_SSE2
+#define HAS_ARGB4444TOARGBROW_SSE2
+#define HAS_ARGBSETROW_X86
+#define HAS_ARGBSHUFFLEROW_SSE2
+#define HAS_ARGBSHUFFLEROW_SSSE3
+#define HAS_ARGBTOARGB1555ROW_SSE2
+#define HAS_ARGBTOARGB4444ROW_SSE2
+#define HAS_ARGBTORAWROW_SSSE3
+#define HAS_ARGBTORGB24ROW_SSSE3
+#define HAS_ARGBTORGB565ROW_SSE2
+#define HAS_ARGBTOUV422ROW_SSSE3
+#define HAS_ARGBTOUV444ROW_SSSE3
+#define HAS_ARGBTOUVJROW_SSSE3
+#define HAS_ARGBTOUVROW_SSSE3
+#define HAS_ARGBTOYJROW_SSSE3
+#define HAS_ARGBTOYROW_SSSE3
+#define HAS_BGRATOUVROW_SSSE3
+#define HAS_BGRATOYROW_SSSE3
+#define HAS_COPYROW_ERMS
+#define HAS_COPYROW_SSE2
+#define HAS_I400TOARGBROW_SSE2
+#define HAS_I411TOARGBROW_SSSE3
+#define HAS_I422TOABGRROW_SSSE3
+#define HAS_I422TOARGB1555ROW_SSSE3
+#define HAS_I422TOARGB4444ROW_SSSE3
+#define HAS_I422TOARGBROW_SSSE3
+#define HAS_I422TOBGRAROW_SSSE3
+#define HAS_I422TORAWROW_SSSE3
+#define HAS_I422TORGB24ROW_SSSE3
+#define HAS_I422TORGB565ROW_SSSE3
+#define HAS_I422TORGBAROW_SSSE3
+#define HAS_I422TOUYVYROW_SSE2
+#define HAS_I422TOYUY2ROW_SSE2
+#define HAS_I444TOARGBROW_SSSE3
+#define HAS_J400TOARGBROW_SSE2
+#define HAS_J422TOARGBROW_SSSE3
+#define HAS_MERGEUVROW_SSE2
+#define HAS_MIRRORROW_SSE2
+#define HAS_MIRRORROW_SSSE3
+#define HAS_MIRRORROW_UV_SSSE3
+#define HAS_MIRRORUVROW_SSSE3
+#define HAS_NV12TOARGBROW_SSSE3
+#define HAS_NV12TORGB565ROW_SSSE3
+#define HAS_NV21TOARGBROW_SSSE3
+#define HAS_NV21TORGB565ROW_SSSE3
+#define HAS_RAWTOARGBROW_SSSE3
+#define HAS_RAWTOYROW_SSSE3
+#define HAS_RGB24TOARGBROW_SSSE3
+#define HAS_RGB24TOYROW_SSSE3
+#define HAS_RGB565TOARGBROW_SSE2
+#define HAS_RGBATOUVROW_SSSE3
+#define HAS_RGBATOYROW_SSSE3
+#define HAS_SETROW_ERMS
+#define HAS_SETROW_X86
+#define HAS_SPLITUVROW_SSE2
+#define HAS_UYVYTOARGBROW_SSSE3
+#define HAS_UYVYTOUV422ROW_SSE2
+#define HAS_UYVYTOUVROW_SSE2
+#define HAS_UYVYTOYROW_SSE2
+#define HAS_YUY2TOARGBROW_SSSE3
+#define HAS_YUY2TOUV422ROW_SSE2
+#define HAS_YUY2TOUVROW_SSE2
+#define HAS_YUY2TOYROW_SSE2
+
+// Effects:
+#define HAS_ARGBADDROW_SSE2
+#define HAS_ARGBAFFINEROW_SSE2
+#define HAS_ARGBATTENUATEROW_SSSE3
+#define HAS_ARGBBLENDROW_SSSE3
+#define HAS_ARGBCOLORMATRIXROW_SSSE3
+#define HAS_ARGBCOLORTABLEROW_X86
+#define HAS_ARGBCOPYALPHAROW_SSE2
+#define HAS_ARGBCOPYYTOALPHAROW_SSE2
+#define HAS_ARGBGRAYROW_SSSE3
+#define HAS_ARGBLUMACOLORTABLEROW_SSSE3
+#define HAS_ARGBMIRRORROW_SSE2
+#define HAS_ARGBMULTIPLYROW_SSE2
+#define HAS_ARGBPOLYNOMIALROW_SSE2
+#define HAS_ARGBQUANTIZEROW_SSE2
+#define HAS_ARGBSEPIAROW_SSSE3
+#define HAS_ARGBSHADEROW_SSE2
+#define HAS_ARGBSUBTRACTROW_SSE2
+#define HAS_ARGBUNATTENUATEROW_SSE2
+#define HAS_COMPUTECUMULATIVESUMROW_SSE2
+#define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+#define HAS_INTERPOLATEROW_SSE2
+#define HAS_INTERPOLATEROW_SSSE3
+#define HAS_RGBCOLORTABLEROW_X86
+#define HAS_SOBELROW_SSE2
+#define HAS_SOBELTOPLANEROW_SSE2
+#define HAS_SOBELXROW_SSE2
+#define HAS_SOBELXYROW_SSE2
+#define HAS_SOBELYROW_SSE2
+#endif
+
+// The following are available on x64 Visual C and clangcl.
+#if !defined(LIBYUV_DISABLE_X86) && defined (_M_X64) && \
+    (!defined(__clang__) || defined(__SSSE3__))
+#define HAS_I422TOARGBROW_SSSE3
+#endif
+
+// GCC >= 4.7.0 required for AVX2.
+#if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
+#if (__GNUC__ > 4) || (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7))
+#define GCC_HAS_AVX2 1
+#endif  // GNUC >= 4.7
+#endif  // __GNUC__
+
+// clang >= 3.4.0 required for AVX2.
+#if defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
+#if (__clang_major__ > 3) || (__clang_major__ == 3 && (__clang_minor__ >= 4))
+#define CLANG_HAS_AVX2 1
+#endif  // clang >= 3.4
+#endif  // __clang__
+
+// Visual C 2012 required for AVX2.
+#if defined(_M_IX86) && !defined(__clang__) && \
+    defined(_MSC_VER) && _MSC_VER >= 1700
+#define VISUALC_HAS_AVX2 1
+#endif  // VisualStudio >= 2012
+
+// The following are available require VS2012.  Port to GCC.
+#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
+#define HAS_ARGB1555TOARGBROW_AVX2
+#define HAS_ARGB4444TOARGBROW_AVX2
+#define HAS_ARGBTOARGB1555ROW_AVX2
+#define HAS_ARGBTOARGB4444ROW_AVX2
+#define HAS_ARGBTORGB565DITHERROW_AVX2
+#define HAS_ARGBTORGB565DITHERROW_SSE2
+#define HAS_ARGBTORGB565ROW_AVX2
+#define HAS_I411TOARGBROW_AVX2
+#define HAS_I422TOARGB1555ROW_AVX2
+#define HAS_I422TOARGB4444ROW_AVX2
+#define HAS_I422TORGB565ROW_AVX2
+#define HAS_I444TOARGBROW_AVX2
+#define HAS_J400TOARGBROW_AVX2
+#define HAS_NV12TOARGBROW_AVX2
+#define HAS_NV12TORGB565ROW_AVX2
+#define HAS_NV21TOARGBROW_AVX2
+#define HAS_NV21TORGB565ROW_AVX2
+#define HAS_RGB565TOARGBROW_AVX2
+#endif
+
+// The following are available on all x86 platforms, but
+// require VS2012, clang 3.4 or gcc 4.7.
+// The code supports NaCL but requires a new compiler and validator.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(VISUALC_HAS_AVX2) || \
+    defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
+#define HAS_ARGBCOPYALPHAROW_AVX2
+#define HAS_ARGBCOPYYTOALPHAROW_AVX2
+#define HAS_ARGBMIRRORROW_AVX2
+#define HAS_ARGBPOLYNOMIALROW_AVX2
+#define HAS_ARGBSHUFFLEROW_AVX2
+#define HAS_ARGBTOUVROW_AVX2
+#define HAS_ARGBTOYJROW_AVX2
+#define HAS_ARGBTOYROW_AVX2
+#define HAS_COPYROW_AVX
+#define HAS_I400TOARGBROW_AVX2
+#define HAS_I422TOABGRROW_AVX2
+#define HAS_I422TOARGBROW_AVX2
+#define HAS_I422TOBGRAROW_AVX2
+#define HAS_I422TORAWROW_AVX2
+#define HAS_I422TORGB24ROW_AVX2
+#define HAS_I422TORGBAROW_AVX2
+#define HAS_INTERPOLATEROW_AVX2
+#define HAS_J422TOARGBROW_AVX2
+#define HAS_MERGEUVROW_AVX2
+#define HAS_MIRRORROW_AVX2
+#define HAS_SPLITUVROW_AVX2
+#define HAS_UYVYTOARGBROW_AVX2
+#define HAS_UYVYTOUV422ROW_AVX2
+#define HAS_UYVYTOUVROW_AVX2
+#define HAS_UYVYTOYROW_AVX2
+#define HAS_YUY2TOARGBROW_AVX2
+#define HAS_YUY2TOUV422ROW_AVX2
+#define HAS_YUY2TOUVROW_AVX2
+#define HAS_YUY2TOYROW_AVX2
+
+// Effects:
+#define HAS_ARGBADDROW_AVX2
+#define HAS_ARGBATTENUATEROW_AVX2
+#define HAS_ARGBMULTIPLYROW_AVX2
+#define HAS_ARGBSUBTRACTROW_AVX2
+#define HAS_ARGBUNATTENUATEROW_AVX2
+#endif
+
+// The following are disabled when SSSE3 is available:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) && \
+    !defined(LIBYUV_SSSE3_ONLY)
+#define HAS_ARGBATTENUATEROW_SSE2
+#define HAS_ARGBBLENDROW_SSE2
+#define HAS_MIRRORROW_SSE2
+#endif
+
+// The following are available on Neon platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON))
+#define HAS_ABGRTOUVROW_NEON
+#define HAS_ABGRTOYROW_NEON
+#define HAS_ARGB1555TOARGBROW_NEON
+#define HAS_ARGB1555TOUVROW_NEON
+#define HAS_ARGB1555TOYROW_NEON
+#define HAS_ARGB4444TOARGBROW_NEON
+#define HAS_ARGB4444TOUVROW_NEON
+#define HAS_ARGB4444TOYROW_NEON
+#define HAS_ARGBTOARGB1555ROW_NEON
+#define HAS_ARGBTOARGB4444ROW_NEON
+#define HAS_ARGBTORAWROW_NEON
+#define HAS_ARGBTORGB24ROW_NEON
+#define HAS_ARGBTORGB565ROW_NEON
+#define HAS_ARGBTOUV411ROW_NEON
+#define HAS_ARGBTOUV422ROW_NEON
+#define HAS_ARGBTOUV444ROW_NEON
+#define HAS_ARGBTOUVJROW_NEON
+#define HAS_ARGBTOUVROW_NEON
+#define HAS_ARGBTOYJROW_NEON
+#define HAS_ARGBTOYROW_NEON
+#define HAS_BGRATOUVROW_NEON
+#define HAS_BGRATOYROW_NEON
+#define HAS_COPYROW_NEON
+#define HAS_J400TOARGBROW_NEON
+#define HAS_I411TOARGBROW_NEON
+#define HAS_I422TOABGRROW_NEON
+#define HAS_I422TOARGB1555ROW_NEON
+#define HAS_I422TOARGB4444ROW_NEON
+#define HAS_I422TOARGBROW_NEON
+#define HAS_I422TOBGRAROW_NEON
+#define HAS_I422TORAWROW_NEON
+#define HAS_I422TORGB24ROW_NEON
+#define HAS_I422TORGB565ROW_NEON
+#define HAS_I422TORGBAROW_NEON
+#define HAS_I422TOUYVYROW_NEON
+#define HAS_I422TOYUY2ROW_NEON
+#define HAS_I444TOARGBROW_NEON
+#define HAS_MERGEUVROW_NEON
+#define HAS_MIRRORROW_NEON
+#define HAS_MIRRORUVROW_NEON
+#define HAS_NV12TOARGBROW_NEON
+#define HAS_NV12TORGB565ROW_NEON
+#define HAS_NV21TOARGBROW_NEON
+#define HAS_NV21TORGB565ROW_NEON
+#define HAS_RAWTOARGBROW_NEON
+#define HAS_RAWTOUVROW_NEON
+#define HAS_RAWTOYROW_NEON
+#define HAS_RGB24TOARGBROW_NEON
+#define HAS_RGB24TOUVROW_NEON
+#define HAS_RGB24TOYROW_NEON
+#define HAS_RGB565TOARGBROW_NEON
+#define HAS_RGB565TOUVROW_NEON
+#define HAS_RGB565TOYROW_NEON
+#define HAS_RGBATOUVROW_NEON
+#define HAS_RGBATOYROW_NEON
+#define HAS_SETROW_NEON
+#define HAS_ARGBSETROW_NEON
+#define HAS_SPLITUVROW_NEON
+#define HAS_UYVYTOARGBROW_NEON
+#define HAS_UYVYTOUV422ROW_NEON
+#define HAS_UYVYTOUVROW_NEON
+#define HAS_UYVYTOYROW_NEON
+#define HAS_I400TOARGBROW_NEON
+#define HAS_YUY2TOARGBROW_NEON
+#define HAS_YUY2TOUV422ROW_NEON
+#define HAS_YUY2TOUVROW_NEON
+#define HAS_YUY2TOYROW_NEON
+#define HAS_ARGBTORGB565DITHERROW_NEON
+
+// Effects:
+#define HAS_ARGBADDROW_NEON
+#define HAS_ARGBATTENUATEROW_NEON
+#define HAS_ARGBBLENDROW_NEON
+#define HAS_ARGBGRAYROW_NEON
+#define HAS_ARGBMIRRORROW_NEON
+#define HAS_ARGBMULTIPLYROW_NEON
+#define HAS_ARGBQUANTIZEROW_NEON
+#define HAS_ARGBSEPIAROW_NEON
+#define HAS_ARGBSHADEROW_NEON
+#define HAS_ARGBSUBTRACTROW_NEON
+#define HAS_INTERPOLATEROW_NEON
+#define HAS_SOBELROW_NEON
+#define HAS_SOBELTOPLANEROW_NEON
+#define HAS_SOBELXROW_NEON
+#define HAS_SOBELXYROW_NEON
+#define HAS_SOBELYROW_NEON
+#define HAS_ARGBCOLORMATRIXROW_NEON
+#define HAS_ARGBSHUFFLEROW_NEON
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32) && (__mips_isa_rev < 6)
+#define HAS_COPYROW_MIPS
+#if defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_I422TOABGRROW_MIPS_DSPR2
+#define HAS_I422TOARGBROW_MIPS_DSPR2
+#define HAS_I422TOBGRAROW_MIPS_DSPR2
+#define HAS_INTERPOLATEROW_MIPS_DSPR2
+#define HAS_MIRRORROW_MIPS_DSPR2
+#define HAS_MIRRORUVROW_MIPS_DSPR2
+#define HAS_SPLITUVROW_MIPS_DSPR2
+#endif
+#endif
+
+#if defined(_MSC_VER) && !defined(__CLR_VER)
+#define SIMD_ALIGNED(var) __declspec(align(16)) var
+#define SIMD_ALIGNED32(var) __declspec(align(64)) var
+typedef __declspec(align(16)) int16 vec16[8];
+typedef __declspec(align(16)) int32 vec32[4];
+typedef __declspec(align(16)) int8 vec8[16];
+typedef __declspec(align(16)) uint16 uvec16[8];
+typedef __declspec(align(16)) uint32 uvec32[4];
+typedef __declspec(align(16)) uint8 uvec8[16];
+typedef __declspec(align(32)) int16 lvec16[16];
+typedef __declspec(align(32)) int32 lvec32[8];
+typedef __declspec(align(32)) int8 lvec8[32];
+typedef __declspec(align(32)) uint16 ulvec16[16];
+typedef __declspec(align(32)) uint32 ulvec32[8];
+typedef __declspec(align(32)) uint8 ulvec8[32];
+#elif defined(__GNUC__)
+// Caveat GCC 4.2 to 4.7 have a known issue using vectors with const.
+#define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
+#define SIMD_ALIGNED32(var) var __attribute__((aligned(64)))
+typedef int16 __attribute__((vector_size(16))) vec16;
+typedef int32 __attribute__((vector_size(16))) vec32;
+typedef int8 __attribute__((vector_size(16))) vec8;
+typedef uint16 __attribute__((vector_size(16))) uvec16;
+typedef uint32 __attribute__((vector_size(16))) uvec32;
+typedef uint8 __attribute__((vector_size(16))) uvec8;
+typedef int16 __attribute__((vector_size(32))) lvec16;
+typedef int32 __attribute__((vector_size(32))) lvec32;
+typedef int8 __attribute__((vector_size(32))) lvec8;
+typedef uint16 __attribute__((vector_size(32))) ulvec16;
+typedef uint32 __attribute__((vector_size(32))) ulvec32;
+typedef uint8 __attribute__((vector_size(32))) ulvec8;
+#else
+#define SIMD_ALIGNED(var) var
+#define SIMD_ALIGNED32(var) var
+typedef int16 vec16[8];
+typedef int32 vec32[4];
+typedef int8 vec8[16];
+typedef uint16 uvec16[8];
+typedef uint32 uvec32[4];
+typedef uint8 uvec8[16];
+typedef int16 lvec16[16];
+typedef int32 lvec32[8];
+typedef int8 lvec8[32];
+typedef uint16 ulvec16[16];
+typedef uint32 ulvec32[8];
+typedef uint8 ulvec8[32];
+#endif
+
+#if defined(__APPLE__) || defined(__x86_64__) || defined(__llvm__)
+#define OMITFP
+#else
+#define OMITFP __attribute__((optimize("omit-frame-pointer")))
+#endif
+
+// NaCL macros for GCC x86 and x64.
+#if defined(__native_client__)
+#define LABELALIGN ".p2align 5\n"
+#else
+#define LABELALIGN
+#endif
+#if defined(__native_client__) && defined(__x86_64__)
+// r14 is used for MEMOP macros.
+#define NACL_R14 "r14",
+#define BUNDLELOCK ".bundle_lock\n"
+#define BUNDLEUNLOCK ".bundle_unlock\n"
+#define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
+#define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
+#define MEMLEA(offset, base) #offset "(%q" #base ")"
+#define MEMLEA3(offset, index, scale) \
+    #offset "(,%q" #index "," #scale ")"
+#define MEMLEA4(offset, base, index, scale) \
+    #offset "(%q" #base ",%q" #index "," #scale ")"
+#define MEMMOVESTRING(s, d) "%%nacl:(%q" #s "),%%nacl:(%q" #d "), %%r15"
+#define MEMSTORESTRING(reg, d) "%%" #reg ",%%nacl:(%q" #d "), %%r15"
+#define MEMOPREG(opcode, offset, base, index, scale, reg) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%%" #reg "\n" \
+    BUNDLEUNLOCK
+#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " %%" #reg ",(%%r15,%%r14)\n" \
+    BUNDLEUNLOCK
+#define MEMOPARG(opcode, offset, base, index, scale, arg) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%" #arg "\n" \
+    BUNDLEUNLOCK
+#define VMEMOPREG(opcode, offset, base, index, scale, reg1, reg2) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #opcode " (%%r15,%%r14),%%" #reg1 ",%%" #reg2 "\n" \
+    BUNDLEUNLOCK
+#define VEXTOPMEM(op, sel, reg, offset, base, index, scale) \
+    BUNDLELOCK \
+    "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
+    #op " $" #sel ",%%" #reg ",(%%r15,%%r14)\n" \
+    BUNDLEUNLOCK
+#else  // defined(__native_client__) && defined(__x86_64__)
+#define NACL_R14
+#define BUNDLEALIGN
+#define MEMACCESS(base) "(%" #base ")"
+#define MEMACCESS2(offset, base) #offset "(%" #base ")"
+#define MEMLEA(offset, base) #offset "(%" #base ")"
+#define MEMLEA3(offset, index, scale) \
+    #offset "(,%" #index "," #scale ")"
+#define MEMLEA4(offset, base, index, scale) \
+    #offset "(%" #base ",%" #index "," #scale ")"
+#define MEMMOVESTRING(s, d)
+#define MEMSTORESTRING(reg, d)
+#define MEMOPREG(opcode, offset, base, index, scale, reg) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
+#define MEMOPMEM(opcode, reg, offset, base, index, scale) \
+    #opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
+#define MEMOPARG(opcode, offset, base, index, scale, arg) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%" #arg "\n"
+#define VMEMOPREG(opcode, offset, base, index, scale, reg1, reg2) \
+    #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg1 ",%%" \
+    #reg2 "\n"
+#define VEXTOPMEM(op, sel, reg, offset, base, index, scale) \
+    #op " $" #sel ",%%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
+#endif  // defined(__native_client__) && defined(__x86_64__)
+
+#if defined(__arm__) || defined(__aarch64__)
+#undef MEMACCESS
+#if defined(__native_client__)
+#define MEMACCESS(base) ".p2align 3\nbic %" #base ", #0xc0000000\n"
+#else
+#define MEMACCESS(base)
+#endif
+#endif
+
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width);
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width);
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width);
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width);
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width);
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width);
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width);
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width);
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width);
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_vu,
+                        uint8* dst_argb,
+                        int width);
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width);
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_vu,
+                          uint8* dst_rgb565,
+                          int width);
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width);
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width);
+
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix);
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix);
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix);
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
+void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_C(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_C(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_C(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_C(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_C(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_C(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int pix);
+void ARGBToYRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Any_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Any_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Any_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Any_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
+void ARGBToYRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void ARGBToYJRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
+void BGRAToYRow_Any_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
+void ABGRToYRow_Any_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
+void RGBAToYRow_Any_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
+void RGB24ToYRow_Any_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
+void RAWToYRow_Any_NEON(const uint8* src_raw, uint8* dst_y, int pix);
+void RGB565ToYRow_Any_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
+void ARGB1555ToYRow_Any_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
+void ARGB4444ToYRow_Any_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
+
+void ARGBToUVRow_AVX2(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Any_AVX2(const uint8* src_argb, int src_stride_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_SSSE3(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_SSSE3(const uint8* src_argb, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                       uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
+                            uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_Any_SSSE3(const uint8* src_bgra, int src_stride_bgra,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_Any_SSSE3(const uint8* src_abgr, int src_stride_abgr,
+                           uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_Any_SSSE3(const uint8* src_rgba, int src_stride_rgba,
+                           uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUV422Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUV411Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                             int pix);
+void ARGBToUVRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVJRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
+                           uint8* dst_u, uint8* dst_v, int pix);
+void BGRAToUVRow_Any_NEON(const uint8* src_bgra, int src_stride_bgra,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void ABGRToUVRow_Any_NEON(const uint8* src_abgr, int src_stride_abgr,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void RGBAToUVRow_Any_NEON(const uint8* src_rgba, int src_stride_rgba,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void RGB24ToUVRow_Any_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                           uint8* dst_u, uint8* dst_v, int pix);
+void RAWToUVRow_Any_NEON(const uint8* src_raw, int src_stride_raw,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void RGB565ToUVRow_Any_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                            uint8* dst_u, uint8* dst_v, int pix);
+void ARGB1555ToUVRow_Any_NEON(const uint8* src_argb1555,
+                              int src_stride_argb1555,
+                              uint8* dst_u, uint8* dst_v, int pix);
+void ARGB4444ToUVRow_Any_NEON(const uint8* src_argb4444,
+                              int src_stride_argb4444,
+                              uint8* dst_u, uint8* dst_v, int pix);
+void ARGBToUVRow_C(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJRow_C(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_u, uint8* dst_v, int width);
+void BGRAToUVRow_C(const uint8* src_bgra, int src_stride_bgra,
+                   uint8* dst_u, uint8* dst_v, int width);
+void ABGRToUVRow_C(const uint8* src_abgr, int src_stride_abgr,
+                   uint8* dst_u, uint8* dst_v, int width);
+void RGBAToUVRow_C(const uint8* src_rgba, int src_stride_rgba,
+                   uint8* dst_u, uint8* dst_v, int width);
+void RGB24ToUVRow_C(const uint8* src_rgb24, int src_stride_rgb24,
+                    uint8* dst_u, uint8* dst_v, int width);
+void RAWToUVRow_C(const uint8* src_raw, int src_stride_raw,
+                  uint8* dst_u, uint8* dst_v, int width);
+void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
+                     uint8* dst_u, uint8* dst_v, int width);
+void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
+                       uint8* dst_u, uint8* dst_v, int width);
+void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
+                       uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV444Row_SSSE3(const uint8* src_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV444Row_Any_SSSE3(const uint8* src_argb,
+                              uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV422Row_SSSE3(const uint8* src_argb,
+                          uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_Any_SSSE3(const uint8* src_argb,
+                              uint8* dst_u, uint8* dst_v, int width);
+
+void ARGBToUV444Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV422Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUV411Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width);
+void ARGBToUVJ422Row_C(const uint8* src_argb,
+                       uint8* dst_u, uint8* dst_v, int width);
+
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width);
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width);
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
+void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width);
+void MirrorRow_C(const uint8* src, uint8* dst, int width);
+void MirrorRow_Any_AVX2(const uint8* src, uint8* dst, int width);
+void MirrorRow_Any_SSSE3(const uint8* src, uint8* dst, int width);
+void MirrorRow_Any_SSE2(const uint8* src, uint8* dst, int width);
+void MirrorRow_Any_NEON(const uint8* src, uint8* dst, int width);
+
+void MirrorUVRow_SSSE3(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                       int width);
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width);
+void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                            int width);
+void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                   int width);
+
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_Any_AVX2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_Any_SSE2(const uint8* src, uint8* dst, int width);
+void ARGBMirrorRow_Any_NEON(const uint8* src, uint8* dst, int width);
+
+void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
+void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                           int pix);
+void SplitUVRow_Any_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                         int pix);
+void SplitUVRow_Any_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                               int pix);
+
+void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                  int width);
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width);
+void MergeUVRow_Any_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+void MergeUVRow_Any_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+void MergeUVRow_Any_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                         int width);
+
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count);
+void CopyRow_AVX(const uint8* src, uint8* dst, int count);
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count);
+void CopyRow_NEON(const uint8* src, uint8* dst, int count);
+void CopyRow_MIPS(const uint8* src, uint8* dst, int count);
+void CopyRow_C(const uint8* src, uint8* dst, int count);
+void CopyRow_Any_SSE2(const uint8* src, uint8* dst, int count);
+void CopyRow_Any_AVX(const uint8* src, uint8* dst, int count);
+void CopyRow_Any_NEON(const uint8* src, uint8* dst, int count);
+
+void CopyRow_16_C(const uint16* src, uint16* dst, int count);
+
+void ARGBCopyAlphaRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBCopyAlphaRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBCopyAlphaRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+
+void ARGBCopyYToAlphaRow_C(const uint8* src_y, uint8* dst_argb, int width);
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
+
+void SetRow_C(uint8* dst, uint8 v8, int count);
+void SetRow_X86(uint8* dst, uint8 v8, int count);
+void SetRow_ERMS(uint8* dst, uint8 v8, int count);
+void SetRow_NEON(uint8* dst, uint8 v8, int count);
+void SetRow_Any_X86(uint8* dst, uint8 v8, int count);
+void SetRow_Any_NEON(uint8* dst, uint8 v8, int count);
+
+void ARGBSetRow_C(uint8* dst_argb, uint32 v32, int count);
+void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count);
+void ARGBSetRow_NEON(uint8* dst_argb, uint32 v32, int count);
+void ARGBSetRow_Any_NEON(uint8* dst_argb, uint32 v32, int count);
+
+// ARGBShufflers for BGRAToARGB etc.
+void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
+                      const uint8* shuffler, int pix);
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix);
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+void ARGBShuffleRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const uint8* shuffler, int pix);
+
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+void RGB565ToARGBRow_AVX2(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_AVX2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_AVX2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix);
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix);
+void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_C(const uint8* src_rgb, uint8* dst_argb, int pix);
+void ARGB1555ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
+void ARGB4444ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
+void RGB24ToARGBRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_Any_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
+
+void RGB565ToARGBRow_Any_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+void RGB565ToARGBRow_Any_AVX2(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_AVX2(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_AVX2(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+
+void RGB24ToARGBRow_Any_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
+void RAWToARGBRow_Any_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
+void RGB565ToARGBRow_Any_NEON(const uint8* src_rgb565, uint8* dst_argb,
+                              int pix);
+void ARGB1555ToARGBRow_Any_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                                int pix);
+void ARGB4444ToARGBRow_Any_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                                int pix);
+
+void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB565DitherRow_C(const uint8* src_argb, uint8* dst_rgb,
+                             const uint32 dither4, int pix);
+void ARGBToRGB565DitherRow_SSE2(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int pix);
+void ARGBToRGB565DitherRow_AVX2(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int pix);
+
+void ARGBToRGB565Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int width);
+
+void ARGBToRGBARow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_Any_AVX2(const uint8* src_y, uint8* dst_argb, int pix);
+void J400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int pix);
+
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I411ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void NV12ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_uv,
+                     uint8* dst_argb,
+                     int width);
+void NV21ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_vu,
+                       uint8* dst_argb,
+                       int width);
+void NV12ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_uv,
+                       uint8* dst_argb,
+                       int width);
+void NV21ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_vu,
+                     uint8* dst_argb,
+                     int width);
+void YUY2ToARGBRow_C(const uint8* src_yuy2,
+                     uint8* dst_argb,
+                     int width);
+void UYVYToARGBRow_C(const uint8* src_uyvy,
+                     uint8* dst_argb,
+                     int width);
+void J422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_argb,
+                     int width);
+void I422ToBGRARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_bgra,
+                     int width);
+void I422ToABGRRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_abgr,
+                     int width);
+void I422ToRGBARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_rgba,
+                     int width);
+void I422ToRGB24Row_C(const uint8* src_y,
+                      const uint8* src_u,
+                      const uint8* src_v,
+                      uint8* dst_rgb24,
+                      int width);
+void I422ToRAWRow_C(const uint8* src_y,
+                    const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* dst_raw,
+                    int width);
+void I422ToARGB4444Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width);
+void I422ToARGB1555Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width);
+void I422ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_rgb565,
+                       int width);
+void I422ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToBGRARow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToRGBARow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToABGRRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I444ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I444ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I411ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void I411ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void NV12ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_uv,
+                         uint8* dst_argb,
+                         int width);
+void NV21ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_vu,
+                         uint8* dst_argb,
+                         int width);
+void NV12ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width);
+void NV21ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_vu,
+                        uint8* dst_argb,
+                        int width);
+void NV12ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_uv,
+                           uint8* dst_argb,
+                           int width);
+void NV21ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_vu,
+                           uint8* dst_argb,
+                           int width);
+void NV12ToRGB565Row_AVX2(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_argb,
+                          int width);
+void NV21ToRGB565Row_AVX2(const uint8* src_y,
+                          const uint8* src_vu,
+                          uint8* dst_argb,
+                          int width);
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
+                         uint8* dst_argb,
+                         int width);
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
+                         uint8* dst_argb,
+                         int width);
+void YUY2ToARGBRow_AVX2(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width);
+void UYVYToARGBRow_AVX2(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width);
+void J422ToARGBRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb,
+                         int width);
+void J422ToARGBRow_AVX2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width);
+void I422ToBGRARow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_bgra,
+                         int width);
+void I422ToABGRRow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_abgr,
+                         int width);
+void I422ToRGBARow_SSSE3(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgba,
+                         int width);
+void I422ToARGB4444Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToARGB4444Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGB1555Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToARGB1555Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+void I422ToRGB565Row_AVX2(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_argb,
+                          int width);
+void I422ToRGB24Row_SSSE3(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb24,
+                          int width);
+void I422ToRGB24Row_AVX2(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width);
+void I422ToRAWRow_SSSE3(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_raw,
+                        int width);
+void I422ToRAWRow_AVX2(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width);
+void I422ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToBGRARow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGBARow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToABGRRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I444ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I444ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I411ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I411ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void NV12ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_uv,
+                             uint8* dst_argb,
+                             int width);
+void NV21ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_vu,
+                             uint8* dst_argb,
+                             int width);
+void NV12ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV21ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_vu,
+                            uint8* dst_argb,
+                            int width);
+void NV12ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_uv,
+                               uint8* dst_argb,
+                               int width);
+void NV21ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_vu,
+                               uint8* dst_argb,
+                               int width);
+void NV12ToRGB565Row_Any_AVX2(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void NV21ToRGB565Row_Any_AVX2(const uint8* src_y,
+                              const uint8* src_vu,
+                              uint8* dst_argb,
+                              int width);
+void YUY2ToARGBRow_Any_SSSE3(const uint8* src_yuy2,
+                             uint8* dst_argb,
+                             int width);
+void UYVYToARGBRow_Any_SSSE3(const uint8* src_uyvy,
+                             uint8* dst_argb,
+                             int width);
+void YUY2ToARGBRow_Any_AVX2(const uint8* src_yuy2,
+                            uint8* dst_argb,
+                            int width);
+void UYVYToARGBRow_Any_AVX2(const uint8* src_uyvy,
+                            uint8* dst_argb,
+                            int width);
+void J422ToARGBRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void J422ToARGBRow_Any_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToBGRARow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_bgra,
+                             int width);
+void I422ToABGRRow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_abgr,
+                             int width);
+void I422ToRGBARow_Any_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_rgba,
+                             int width);
+void I422ToARGB4444Row_Any_SSSE3(const uint8* src_y,
+                                 const uint8* src_u,
+                                 const uint8* src_v,
+                                 uint8* dst_rgba,
+                                 int width);
+void I422ToARGB4444Row_Any_AVX2(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_rgba,
+                                int width);
+void I422ToARGB1555Row_Any_SSSE3(const uint8* src_y,
+                                 const uint8* src_u,
+                                 const uint8* src_v,
+                                 uint8* dst_rgba,
+                                 int width);
+void I422ToARGB1555Row_Any_AVX2(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_rgba,
+                                int width);
+void I422ToRGB565Row_Any_SSSE3(const uint8* src_y,
+                               const uint8* src_u,
+                               const uint8* src_v,
+                               uint8* dst_rgba,
+                               int width);
+void I422ToRGB565Row_Any_AVX2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_rgba,
+                              int width);
+void I422ToRGB24Row_Any_SSSE3(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToRGB24Row_Any_AVX2(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToRAWRow_Any_SSSE3(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRAWRow_Any_AVX2(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+
+void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_Any_AVX2(const uint8* src_y, uint8* dst_argb, int width);
+void I400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int width);
+
+// ARGB preattenuated alpha blend.
+void ARGBBlendRow_SSSE3(const uint8* src_argb, const uint8* src_argb1,
+                        uint8* dst_argb, int width);
+void ARGBBlendRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBBlendRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1,
+                    uint8* dst_argb, int width);
+
+// ARGB multiply images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBMultiplyRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBMultiplyRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBMultiplyRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBMultiplyRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
+// ARGB add images.
+void ARGBAddRow_C(const uint8* src_argb, const uint8* src_argb1,
+                  uint8* dst_argb, int width);
+void ARGBAddRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+void ARGBAddRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+void ARGBAddRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                     uint8* dst_argb, int width);
+void ARGBAddRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                         uint8* dst_argb, int width);
+
+// ARGB subtract images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBSubtractRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBSubtractRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
+void ARGBToRGB24Row_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB565DitherRow_Any_SSE2(const uint8* src_argb, uint8* dst_rgb,
+                                    const uint32 dither4, int pix);
+void ARGBToRGB565DitherRow_Any_AVX2(const uint8* src_argb, uint8* dst_rgb,
+                                    const uint32 dither4, int pix);
+
+void ARGBToRGB565Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix);
+
+void ARGBToRGB24Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRAWRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB1555Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToARGB4444Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
+void ARGBToRGB565DitherRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb,
+                                    const uint32 dither4, int width);
+
+void I444ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I411ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToBGRARow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToABGRRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGBARow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb,
+                            int width);
+void I422ToRGB24Row_Any_NEON(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb,
+                             int width);
+void I422ToRAWRow_Any_NEON(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_argb,
+                           int width);
+void I422ToARGB4444Row_Any_NEON(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_argb,
+                                int width);
+void I422ToARGB1555Row_Any_NEON(const uint8* src_y,
+                                const uint8* src_u,
+                                const uint8* src_v,
+                                uint8* dst_argb,
+                                int width);
+void I422ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void NV12ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV21ToARGBRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_uv,
+                            uint8* dst_argb,
+                            int width);
+void NV12ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void NV21ToRGB565Row_Any_NEON(const uint8* src_y,
+                              const uint8* src_uv,
+                              uint8* dst_argb,
+                              int width);
+void YUY2ToARGBRow_Any_NEON(const uint8* src_yuy2,
+                            uint8* dst_argb,
+                            int width);
+void UYVYToARGBRow_Any_NEON(const uint8* src_uyvy,
+                            uint8* dst_argb,
+                            int width);
+void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
+                              const uint8* src_u,
+                              const uint8* src_v,
+                              uint8* dst_argb,
+                              int width);
+
+void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_C(const uint8* src_yuy2, int stride_yuy2,
+                   uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_C(const uint8* src_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_AVX2(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_SSE2(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToYRow_Any_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
+void YUY2ToUVRow_Any_NEON(const uint8* src_yuy2, int stride_yuy2,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void YUY2ToUV422Row_Any_NEON(const uint8* src_yuy2,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_NEON(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix);
+
+void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_C(const uint8* src_uyvy, int stride_uyvy,
+                   uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_C(const uint8* src_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_AVX2(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_SSE2(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToYRow_Any_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
+void UYVYToUVRow_Any_NEON(const uint8* src_uyvy, int stride_uyvy,
+                          uint8* dst_u, uint8* dst_v, int pix);
+void UYVYToUV422Row_Any_NEON(const uint8* src_uyvy,
+                             uint8* dst_u, uint8* dst_v, int pix);
+
+void I422ToYUY2Row_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_yuy2, int width);
+void I422ToUYVYRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_uyvy, int width);
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width);
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width);
+void I422ToYUY2Row_Any_SSE2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_yuy2, int width);
+void I422ToUYVYRow_Any_SSE2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_uyvy, int width);
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width);
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width);
+void I422ToYUY2Row_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_yuy2, int width);
+void I422ToUYVYRow_Any_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_uyvy, int width);
+
+// Effects related row functions.
+void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBAttenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+void ARGBAttenuateRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                int width);
+void ARGBAttenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+void ARGBAttenuateRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
+                               int width);
+
+// Inverse table for unattenuate, shared by C and SSE2.
+extern const uint32 fixed_invtbl8[256];
+void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+void ARGBUnattenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+
+void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
+
+void ARGBSepiaRow_C(uint8* dst_argb, int width);
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width);
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width);
+
+void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
+                          const int8* matrix_argb, int width);
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width);
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width);
+
+void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
+
+void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
+
+void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
+                       int interval_offset, int width);
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width);
+
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value);
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value);
+
+// Used for blur.
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst, int count);
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width);
+
+void CumulativeSumToAverageRow_C(const int32* topleft, const int32* botleft,
+                                 int width, int area, uint8* dst, int count);
+void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
+                               const int32* previous_cumsum, int width);
+
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width);
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width);
+
+// Used for I420Scale, ARGBScale, and ARGBInterpolate.
+void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+                      ptrdiff_t src_stride_ptr,
+                      int width, int source_y_fraction);
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride_ptr, int width,
+                          int source_y_fraction);
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride_ptr, int width,
+                         int source_y_fraction);
+void InterpolateRow_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                               ptrdiff_t src_stride_ptr, int width,
+                               int source_y_fraction);
+void InterpolateRow_Any_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                              ptrdiff_t src_stride_ptr, int width,
+                              int source_y_fraction);
+void InterpolateRow_Any_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                             ptrdiff_t src_stride_ptr, int width,
+                             int source_y_fraction);
+void InterpolateRow_Any_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                                   ptrdiff_t src_stride_ptr, int width,
+                                   int source_y_fraction);
+
+void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         ptrdiff_t src_stride_ptr,
+                         int width, int source_y_fraction);
+
+// Sobel images.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width);
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width);
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width);
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width);
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width);
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width);
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width);
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width);
+void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_y, int width);
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width);
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width);
+void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                  uint8* dst_argb, int width);
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width);
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width);
+void SobelRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_argb, int width);
+void SobelRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_argb, int width);
+void SobelToPlaneRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                              uint8* dst_y, int width);
+void SobelToPlaneRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                              uint8* dst_y, int width);
+void SobelXYRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                         uint8* dst_argb, int width);
+void SobelXYRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                         uint8* dst_argb, int width);
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width);
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+
+void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                             const uint8* luma, uint32 lumacoeff);
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_ROW_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/scale.h b/third_party/aom/third_party/libyuv/include/libyuv/scale.h
new file mode 100644
index 0000000000..3974aba34e
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/scale.h
@@ -0,0 +1,104 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Supported filtering.
+typedef enum FilterMode {
+  kFilterNone = 0,  // Point sample; Fastest.
+  kFilterLinear = 1,  // Filter horizontally only.
+  kFilterBilinear = 2,  // Faster than box, but lower quality scaling down.
+  kFilterBox = 3  // Highest quality.
+} FilterModeEnum;
+
+// Scale a YUV plane.
+LIBYUV_API
+void ScalePlane(const uint8* src, int src_stride,
+                int src_width, int src_height,
+                uint8* dst, int dst_stride,
+                int dst_width, int dst_height,
+                enum FilterMode filtering);
+
+LIBYUV_API
+void ScalePlane_16(const uint16* src, int src_stride,
+                   int src_width, int src_height,
+                   uint16* dst, int dst_stride,
+                   int dst_width, int dst_height,
+                   enum FilterMode filtering);
+
+// Scales a YUV 4:2:0 image from the src width and height to the
+// dst width and height.
+// If filtering is kFilterNone, a simple nearest-neighbor algorithm is
+// used. This produces basic (blocky) quality at the fastest speed.
+// If filtering is kFilterBilinear, interpolation is used to produce a better
+// quality image, at the expense of speed.
+// If filtering is kFilterBox, averaging is used to produce ever better
+// quality image, at further expense of speed.
+// Returns 0 if successful.
+
+LIBYUV_API
+int I420Scale(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              int src_width, int src_height,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int dst_width, int dst_height,
+              enum FilterMode filtering);
+
+LIBYUV_API
+int I420Scale_16(const uint16* src_y, int src_stride_y,
+                 const uint16* src_u, int src_stride_u,
+                 const uint16* src_v, int src_stride_v,
+                 int src_width, int src_height,
+                 uint16* dst_y, int dst_stride_y,
+                 uint16* dst_u, int dst_stride_u,
+                 uint16* dst_v, int dst_stride_v,
+                 int dst_width, int dst_height,
+                 enum FilterMode filtering);
+
+#ifdef __cplusplus
+// Legacy API.  Deprecated.
+LIBYUV_API
+int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
+          int src_stride_y, int src_stride_u, int src_stride_v,
+          int src_width, int src_height,
+          uint8* dst_y, uint8* dst_u, uint8* dst_v,
+          int dst_stride_y, int dst_stride_u, int dst_stride_v,
+          int dst_width, int dst_height,
+          LIBYUV_BOOL interpolate);
+
+// Legacy API.  Deprecated.
+LIBYUV_API
+int ScaleOffset(const uint8* src_i420, int src_width, int src_height,
+                uint8* dst_i420, int dst_width, int dst_height, int dst_yoffset,
+                LIBYUV_BOOL interpolate);
+
+// For testing, allow disabling of specialized scalers.
+LIBYUV_API
+void SetUseReferenceImpl(LIBYUV_BOOL use);
+#endif  // __cplusplus
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/scale_argb.h b/third_party/aom/third_party/libyuv/include/libyuv/scale_argb.h
new file mode 100644
index 0000000000..22563837dd
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/scale_argb.h
@@ -0,0 +1,58 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_ARGB_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_ARGB_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/scale.h"  // For FilterMode
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+LIBYUV_API
+int ARGBScale(const uint8* src_argb, int src_stride_argb,
+              int src_width, int src_height,
+              uint8* dst_argb, int dst_stride_argb,
+              int dst_width, int dst_height,
+              enum FilterMode filtering);
+
+// Clipped scale takes destination rectangle coordinates for clip values.
+LIBYUV_API
+int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
+                  int src_width, int src_height,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int dst_width, int dst_height,
+                  int clip_x, int clip_y, int clip_width, int clip_height,
+                  enum FilterMode filtering);
+
+// TODO(fbarchard): Implement this.
+// Scale with YUV conversion to ARGB and clipping.
+LIBYUV_API
+int YUVToARGBScaleClip(const uint8* src_y, int src_stride_y,
+                       const uint8* src_u, int src_stride_u,
+                       const uint8* src_v, int src_stride_v,
+                       uint32 src_fourcc,
+                       int src_width, int src_height,
+                       uint8* dst_argb, int dst_stride_argb,
+                       uint32 dst_fourcc,
+                       int dst_width, int dst_height,
+                       int clip_x, int clip_y, int clip_width, int clip_height,
+                       enum FilterMode filtering);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_ARGB_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/scale_row.h b/third_party/aom/third_party/libyuv/include/libyuv/scale_row.h
new file mode 100644
index 0000000000..a46b5ce692
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/scale_row.h
@@ -0,0 +1,479 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_SCALE_ROW_H_  // NOLINT
+#define INCLUDE_LIBYUV_SCALE_ROW_H_
+
+#include "libyuv/basic_types.h"
+#include "libyuv/scale.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if defined(__pnacl__) || defined(__CLR_VER) || \
+    (defined(__i386__) && !defined(__SSE2__))
+#define LIBYUV_DISABLE_X86
+#endif
+
+// Visual C 2012 required for AVX2.
+#if defined(_M_IX86) && !defined(__clang__) && \
+    defined(_MSC_VER) && _MSC_VER >= 1700
+#define VISUALC_HAS_AVX2 1
+#endif  // VisualStudio >= 2012
+
+// The following are available on all x86 platforms:
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_FIXEDDIV1_X86
+#define HAS_FIXEDDIV_X86
+#define HAS_SCALEARGBCOLS_SSE2
+#define HAS_SCALEARGBCOLSUP2_SSE2
+#define HAS_SCALEARGBFILTERCOLS_SSSE3
+#define HAS_SCALEARGBROWDOWN2_SSE2
+#define HAS_SCALEARGBROWDOWNEVEN_SSE2
+#define HAS_SCALECOLSUP2_SSE2
+#define HAS_SCALEFILTERCOLS_SSSE3
+#define HAS_SCALEROWDOWN2_SSE2
+#define HAS_SCALEROWDOWN34_SSSE3
+#define HAS_SCALEROWDOWN38_SSSE3
+#define HAS_SCALEROWDOWN4_SSE2
+#endif
+
+// The following are available on VS2012:
+#if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
+#define HAS_SCALEADDROW_AVX2
+#define HAS_SCALEROWDOWN2_AVX2
+#define HAS_SCALEROWDOWN4_AVX2
+#endif
+
+// The following are available on Visual C:
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && !defined(__clang__)
+#define HAS_SCALEADDROW_SSE2
+#endif
+
+// The following are available on Neon platforms:
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
+#define HAS_SCALEARGBCOLS_NEON
+#define HAS_SCALEARGBROWDOWN2_NEON
+#define HAS_SCALEARGBROWDOWNEVEN_NEON
+#define HAS_SCALEFILTERCOLS_NEON
+#define HAS_SCALEROWDOWN2_NEON
+#define HAS_SCALEROWDOWN34_NEON
+#define HAS_SCALEROWDOWN38_NEON
+#define HAS_SCALEROWDOWN4_NEON
+#define HAS_SCALEARGBFILTERCOLS_NEON
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && !defined(__native_client__) && \
+    defined(__mips__) && defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+#define HAS_SCALEROWDOWN2_MIPS_DSPR2
+#define HAS_SCALEROWDOWN4_MIPS_DSPR2
+#define HAS_SCALEROWDOWN34_MIPS_DSPR2
+#define HAS_SCALEROWDOWN38_MIPS_DSPR2
+#endif
+
+// Scale ARGB vertically with bilinear interpolation.
+void ScalePlaneVertical(int src_height,
+                        int dst_width, int dst_height,
+                        int src_stride, int dst_stride,
+                        const uint8* src_argb, uint8* dst_argb,
+                        int x, int y, int dy,
+                        int bpp, enum FilterMode filtering);
+
+void ScalePlaneVertical_16(int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint16* src_argb, uint16* dst_argb,
+                           int x, int y, int dy,
+                           int wpp, enum FilterMode filtering);
+
+// Simplify the filtering based on scale factors.
+enum FilterMode ScaleFilterReduce(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  enum FilterMode filtering);
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_C(int num, int div);
+int FixedDiv_X86(int num, int div);
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
+int FixedDiv1_C(int num, int div);
+int FixedDiv1_X86(int num, int div);
+#ifdef HAS_FIXEDDIV_X86
+#define FixedDiv FixedDiv_X86
+#define FixedDiv1 FixedDiv1_X86
+#else
+#define FixedDiv FixedDiv_C
+#define FixedDiv1 FixedDiv1_C
+#endif
+
+// Compute slope values for stepping.
+void ScaleSlope(int src_width, int src_height,
+                int dst_width, int dst_height,
+                enum FilterMode filtering,
+                int* x, int* y, int* dx, int* dy);
+
+void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width);
+void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width);
+void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width);
+void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                              uint16* dst, int dst_width);
+void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width);
+void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width);
+void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width);
+void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width);
+void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width);
+void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width);
+void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width);
+void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width);
+void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width);
+void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width);
+void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                 int dst_width, int x, int dx);
+void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                    int dst_width, int x, int dx);
+void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
+                    int dst_width, int, int);
+void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int, int);
+void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx);
+void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                          int dst_width, int x, int dx);
+void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
+                         int dst_width, int x, int dx);
+void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                            int dst_width, int x, int dx);
+void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width);
+void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width);
+void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width);
+void ScaleAddRow_C(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+void ScaleAddRow_16_C(const uint16* src_ptr, uint32* dst_ptr, int src_width);
+void ScaleARGBRowDown2_C(const uint8* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            int src_stepx,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
+                     int dst_width, int x, int dx);
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x, int dx);
+void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int, int);
+void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx);
+void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
+                             int dst_width, int x, int dx);
+
+// Specialized scalers for x86.
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown2_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Linear_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_Any_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown4_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+void ScaleRowDown34_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_2_Box_Any_SSSE3(const uint8* src_ptr,
+                                    ptrdiff_t src_stride,
+                                    uint8* dst_ptr, int dst_width);
+
+void ScaleAddRow_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+void ScaleAddRow_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+void ScaleAddRow_Any_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+void ScaleAddRow_Any_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx);
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx);
+
+
+// ARGB Column functions
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx);
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx);
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx);
+void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                              int dst_width, int x, int dx);
+void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx);
+void ScaleARGBFilterCols_Any_NEON(uint8* dst_argb, const uint8* src_argb,
+                                  int dst_width, int x, int dx);
+void ScaleARGBCols_Any_NEON(uint8* dst_argb, const uint8* src_argb,
+                            int dst_width, int x, int dx);
+
+// ARGB Row functions
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width);
+void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleARGBRowDown2_Any_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Linear_Any_SSE2(const uint8* src_argb,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_Any_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                   uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                uint8* dst, int dst_width);
+void ScaleARGBRowDown2Linear_Any_NEON(const uint8* src_argb,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_argb, int dst_width);
+void ScaleARGBRowDown2Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                   uint8* dst, int dst_width);
+
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_Any_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                                   int src_stepx,
+                                   uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_Any_SSE2(const uint8* src_argb,
+                                      ptrdiff_t src_stride,
+                                      int src_stepx,
+                                      uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEven_Any_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                   int src_stepx,
+                                   uint8* dst_argb, int dst_width);
+void ScaleARGBRowDownEvenBox_Any_NEON(const uint8* src_argb,
+                                      ptrdiff_t src_stride,
+                                      int src_stepx,
+                                      uint8* dst_argb, int dst_width);
+
+// ScaleRowDown2Box also used by planar functions
+// NEON downscalers with interpolation.
+
+// Note - not static due to reuse in convert for 444 to 420.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width);
+void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width);
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+//  to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width);
+// 32x3 -> 12x1
+void ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+void ScaleRowDown2_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width);
+void ScaleRowDown2Linear_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                  uint8* dst, int dst_width);
+void ScaleRowDown2Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown4_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width);
+void ScaleRowDown4Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                             uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_0_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                   uint8* dst_ptr, int dst_width);
+void ScaleRowDown34_1_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                                   uint8* dst_ptr, int dst_width);
+// 32 -> 12
+void ScaleRowDown38_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                             uint8* dst_ptr, int dst_width);
+// 32x3 -> 12x1
+void ScaleRowDown38_3_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
+
+void ScaleAddRow_NEON(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+void ScaleAddRow_Any_NEON(const uint8* src_ptr, uint16* dst_ptr, int src_width);
+
+void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                          int dst_width, int x, int dx);
+
+void ScaleFilterCols_Any_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                              int dst_width, int x, int dx);
+
+
+void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width);
+void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width);
+void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width);
+void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width);
+void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
+void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_SCALE_ROW_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/version.h b/third_party/aom/third_party/libyuv/include/libyuv/version.h
new file mode 100644
index 0000000000..287b98ebf2
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/version.h
@@ -0,0 +1,17 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
+#define INCLUDE_LIBYUV_VERSION_H_
+
+#define LIBYUV_VERSION 1456
+
+#endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/include/libyuv/video_common.h b/third_party/aom/third_party/libyuv/include/libyuv/video_common.h
new file mode 100644
index 0000000000..7b0a19cc90
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/include/libyuv/video_common.h
@@ -0,0 +1,183 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+// Common definitions for video, including fourcc and VideoFormat.
+
+#ifndef INCLUDE_LIBYUV_VIDEO_COMMON_H_  // NOLINT
+#define INCLUDE_LIBYUV_VIDEO_COMMON_H_
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+// Definition of FourCC codes
+//////////////////////////////////////////////////////////////////////////////
+
+// Convert four characters to a FourCC code.
+// Needs to be a macro otherwise the OS X compiler complains when the kFormat*
+// constants are used in a switch.
+#ifdef __cplusplus
+#define FOURCC(a, b, c, d) ( \
+    (static_cast<uint32>(a)) | (static_cast<uint32>(b) << 8) | \
+    (static_cast<uint32>(c) << 16) | (static_cast<uint32>(d) << 24))
+#else
+#define FOURCC(a, b, c, d) ( \
+    ((uint32)(a)) | ((uint32)(b) << 8) | /* NOLINT */ \
+    ((uint32)(c) << 16) | ((uint32)(d) << 24))  /* NOLINT */
+#endif
+
+// Some pages discussing FourCC codes:
+//   http://www.fourcc.org/yuv.php
+//   http://v4l2spec.bytesex.org/spec/book1.htm
+//   http://developer.apple.com/quicktime/icefloe/dispatch020.html
+//   http://msdn.microsoft.com/library/windows/desktop/dd206750.aspx#nv12
+//   http://people.xiph.org/~xiphmont/containers/nut/nut4cc.txt
+
+// FourCC codes grouped according to implementation efficiency.
+// Primary formats should convert in 1 efficient step.
+// Secondary formats are converted in 2 steps.
+// Auxilliary formats call primary converters.
+enum FourCC {
+  // 9 Primary YUV formats: 5 planar, 2 biplanar, 2 packed.
+  FOURCC_I420 = FOURCC('I', '4', '2', '0'),
+  FOURCC_I422 = FOURCC('I', '4', '2', '2'),
+  FOURCC_I444 = FOURCC('I', '4', '4', '4'),
+  FOURCC_I411 = FOURCC('I', '4', '1', '1'),
+  FOURCC_I400 = FOURCC('I', '4', '0', '0'),
+  FOURCC_NV21 = FOURCC('N', 'V', '2', '1'),
+  FOURCC_NV12 = FOURCC('N', 'V', '1', '2'),
+  FOURCC_YUY2 = FOURCC('Y', 'U', 'Y', '2'),
+  FOURCC_UYVY = FOURCC('U', 'Y', 'V', 'Y'),
+
+  // 2 Secondary YUV formats: row biplanar.
+  FOURCC_M420 = FOURCC('M', '4', '2', '0'),
+  FOURCC_Q420 = FOURCC('Q', '4', '2', '0'), // deprecated.
+
+  // 9 Primary RGB formats: 4 32 bpp, 2 24 bpp, 3 16 bpp.
+  FOURCC_ARGB = FOURCC('A', 'R', 'G', 'B'),
+  FOURCC_BGRA = FOURCC('B', 'G', 'R', 'A'),
+  FOURCC_ABGR = FOURCC('A', 'B', 'G', 'R'),
+  FOURCC_24BG = FOURCC('2', '4', 'B', 'G'),
+  FOURCC_RAW  = FOURCC('r', 'a', 'w', ' '),
+  FOURCC_RGBA = FOURCC('R', 'G', 'B', 'A'),
+  FOURCC_RGBP = FOURCC('R', 'G', 'B', 'P'),  // rgb565 LE.
+  FOURCC_RGBO = FOURCC('R', 'G', 'B', 'O'),  // argb1555 LE.
+  FOURCC_R444 = FOURCC('R', '4', '4', '4'),  // argb4444 LE.
+
+  // 4 Secondary RGB formats: 4 Bayer Patterns. deprecated.
+  FOURCC_RGGB = FOURCC('R', 'G', 'G', 'B'),
+  FOURCC_BGGR = FOURCC('B', 'G', 'G', 'R'),
+  FOURCC_GRBG = FOURCC('G', 'R', 'B', 'G'),
+  FOURCC_GBRG = FOURCC('G', 'B', 'R', 'G'),
+
+  // 1 Primary Compressed YUV format.
+  FOURCC_MJPG = FOURCC('M', 'J', 'P', 'G'),
+
+  // 5 Auxiliary YUV variations: 3 with U and V planes are swapped, 1 Alias.
+  FOURCC_YV12 = FOURCC('Y', 'V', '1', '2'),
+  FOURCC_YV16 = FOURCC('Y', 'V', '1', '6'),
+  FOURCC_YV24 = FOURCC('Y', 'V', '2', '4'),
+  FOURCC_YU12 = FOURCC('Y', 'U', '1', '2'),  // Linux version of I420.
+  FOURCC_J420 = FOURCC('J', '4', '2', '0'),
+  FOURCC_J400 = FOURCC('J', '4', '0', '0'),
+
+  // 14 Auxiliary aliases.  CanonicalFourCC() maps these to canonical fourcc.
+  FOURCC_IYUV = FOURCC('I', 'Y', 'U', 'V'),  // Alias for I420.
+  FOURCC_YU16 = FOURCC('Y', 'U', '1', '6'),  // Alias for I422.
+  FOURCC_YU24 = FOURCC('Y', 'U', '2', '4'),  // Alias for I444.
+  FOURCC_YUYV = FOURCC('Y', 'U', 'Y', 'V'),  // Alias for YUY2.
+  FOURCC_YUVS = FOURCC('y', 'u', 'v', 's'),  // Alias for YUY2 on Mac.
+  FOURCC_HDYC = FOURCC('H', 'D', 'Y', 'C'),  // Alias for UYVY.
+  FOURCC_2VUY = FOURCC('2', 'v', 'u', 'y'),  // Alias for UYVY on Mac.
+  FOURCC_JPEG = FOURCC('J', 'P', 'E', 'G'),  // Alias for MJPG.
+  FOURCC_DMB1 = FOURCC('d', 'm', 'b', '1'),  // Alias for MJPG on Mac.
+  FOURCC_BA81 = FOURCC('B', 'A', '8', '1'),  // Alias for BGGR.
+  FOURCC_RGB3 = FOURCC('R', 'G', 'B', '3'),  // Alias for RAW.
+  FOURCC_BGR3 = FOURCC('B', 'G', 'R', '3'),  // Alias for 24BG.
+  FOURCC_CM32 = FOURCC(0, 0, 0, 32),  // Alias for BGRA kCMPixelFormat_32ARGB
+  FOURCC_CM24 = FOURCC(0, 0, 0, 24),  // Alias for RAW kCMPixelFormat_24RGB
+  FOURCC_L555 = FOURCC('L', '5', '5', '5'),  // Alias for RGBO.
+  FOURCC_L565 = FOURCC('L', '5', '6', '5'),  // Alias for RGBP.
+  FOURCC_5551 = FOURCC('5', '5', '5', '1'),  // Alias for RGBO.
+
+  // 1 Auxiliary compressed YUV format set aside for capturer.
+  FOURCC_H264 = FOURCC('H', '2', '6', '4'),
+
+  // Match any fourcc.
+  FOURCC_ANY = -1,
+};
+
+enum FourCCBpp {
+  // Canonical fourcc codes used in our code.
+  FOURCC_BPP_I420 = 12,
+  FOURCC_BPP_I422 = 16,
+  FOURCC_BPP_I444 = 24,
+  FOURCC_BPP_I411 = 12,
+  FOURCC_BPP_I400 = 8,
+  FOURCC_BPP_NV21 = 12,
+  FOURCC_BPP_NV12 = 12,
+  FOURCC_BPP_YUY2 = 16,
+  FOURCC_BPP_UYVY = 16,
+  FOURCC_BPP_M420 = 12,
+  FOURCC_BPP_Q420 = 12,
+  FOURCC_BPP_ARGB = 32,
+  FOURCC_BPP_BGRA = 32,
+  FOURCC_BPP_ABGR = 32,
+  FOURCC_BPP_RGBA = 32,
+  FOURCC_BPP_24BG = 24,
+  FOURCC_BPP_RAW  = 24,
+  FOURCC_BPP_RGBP = 16,
+  FOURCC_BPP_RGBO = 16,
+  FOURCC_BPP_R444 = 16,
+  FOURCC_BPP_RGGB = 8,
+  FOURCC_BPP_BGGR = 8,
+  FOURCC_BPP_GRBG = 8,
+  FOURCC_BPP_GBRG = 8,
+  FOURCC_BPP_YV12 = 12,
+  FOURCC_BPP_YV16 = 16,
+  FOURCC_BPP_YV24 = 24,
+  FOURCC_BPP_YU12 = 12,
+  FOURCC_BPP_J420 = 12,
+  FOURCC_BPP_J400 = 8,
+  FOURCC_BPP_MJPG = 0,  // 0 means unknown.
+  FOURCC_BPP_H264 = 0,
+  FOURCC_BPP_IYUV = 12,
+  FOURCC_BPP_YU16 = 16,
+  FOURCC_BPP_YU24 = 24,
+  FOURCC_BPP_YUYV = 16,
+  FOURCC_BPP_YUVS = 16,
+  FOURCC_BPP_HDYC = 16,
+  FOURCC_BPP_2VUY = 16,
+  FOURCC_BPP_JPEG = 1,
+  FOURCC_BPP_DMB1 = 1,
+  FOURCC_BPP_BA81 = 8,
+  FOURCC_BPP_RGB3 = 24,
+  FOURCC_BPP_BGR3 = 24,
+  FOURCC_BPP_CM32 = 32,
+  FOURCC_BPP_CM24 = 24,
+
+  // Match any fourcc.
+  FOURCC_BPP_ANY  = 0,  // 0 means unknown.
+};
+
+// Converts fourcc aliases into canonical ones.
+LIBYUV_API uint32 CanonicalFourCC(uint32 fourcc);
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+#endif  // INCLUDE_LIBYUV_VIDEO_COMMON_H_  NOLINT
diff --git a/third_party/aom/third_party/libyuv/source/compare.cc b/third_party/aom/third_party/libyuv/source/compare.cc
new file mode 100644
index 0000000000..46aa8473d2
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/compare.cc
@@ -0,0 +1,373 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/compare.h"
+
+#include <float.h>
+#include <math.h>
+#ifdef _OPENMP
+#include <omp.h>
+#endif
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// hash seed of 5381 recommended.
+// Internal C version of HashDjb2 with int sized count for efficiency.
+uint32 HashDjb2_C(const uint8* src, int count, uint32 seed);
+
+// This module is for Visual C x86
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__))))
+#define HAS_HASHDJB2_SSE41
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed);
+
+#ifdef VISUALC_HAS_AVX2
+#define HAS_HASHDJB2_AVX2
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed);
+#endif
+
+#endif  // HAS_HASHDJB2_SSE41
+
+// hash seed of 5381 recommended.
+LIBYUV_API
+uint32 HashDjb2(const uint8* src, uint64 count, uint32 seed) {
+  const int kBlockSize = 1 << 15;  // 32768;
+  int remainder;
+  uint32 (*HashDjb2_SSE)(const uint8* src, int count, uint32 seed) = HashDjb2_C;
+#if defined(HAS_HASHDJB2_SSE41)
+  if (TestCpuFlag(kCpuHasSSE41)) {
+    HashDjb2_SSE = HashDjb2_SSE41;
+  }
+#endif
+#if defined(HAS_HASHDJB2_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    HashDjb2_SSE = HashDjb2_AVX2;
+  }
+#endif
+
+  while (count >= (uint64)(kBlockSize)) {
+    seed = HashDjb2_SSE(src, kBlockSize, seed);
+    src += kBlockSize;
+    count -= kBlockSize;
+  }
+  remainder = (int)(count) & ~15;
+  if (remainder) {
+    seed = HashDjb2_SSE(src, remainder, seed);
+    src += remainder;
+    count -= remainder;
+  }
+  remainder = (int)(count) & 15;
+  if (remainder) {
+    seed = HashDjb2_C(src, remainder, seed);
+  }
+  return seed;
+}
+
+static uint32 ARGBDetectRow_C(const uint8* argb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    if (argb[0] != 255) {  // First byte is not Alpha of 255, so not ARGB.
+      return FOURCC_BGRA;
+    }
+    if (argb[3] != 255) {  // 4th byte is not Alpha of 255, so not BGRA.
+      return FOURCC_ARGB;
+    }
+    if (argb[4] != 255) {  // Second pixel first byte is not Alpha of 255.
+      return FOURCC_BGRA;
+    }
+    if (argb[7] != 255) {  // Second pixel 4th byte is not Alpha of 255.
+      return FOURCC_ARGB;
+    }
+    argb += 8;
+  }
+  if (width & 1) {
+    if (argb[0] != 255) {  // First byte is not Alpha of 255, so not ARGB.
+      return FOURCC_BGRA;
+    }
+    if (argb[3] != 255) {  // 4th byte is not Alpha of 255, so not BGRA.
+      return FOURCC_ARGB;
+    }
+  }
+  return 0;
+}
+
+// Scan an opaque argb image and return fourcc based on alpha offset.
+// Returns FOURCC_ARGB, FOURCC_BGRA, or 0 if unknown.
+LIBYUV_API
+uint32 ARGBDetect(const uint8* argb, int stride_argb, int width, int height) {
+  uint32 fourcc = 0;
+  int h;
+
+  // Coalesce rows.
+  if (stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    stride_argb = 0;
+  }
+  for (h = 0; h < height && fourcc == 0; ++h) {
+    fourcc = ARGBDetectRow_C(argb, width);
+    argb += stride_argb;
+  }
+  return fourcc;
+}
+
+uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count);
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
+#define HAS_SUMSQUAREERROR_NEON
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count);
+#endif
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
+#define HAS_SUMSQUAREERROR_SSE2
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count);
+#endif
+
+#ifdef VISUALC_HAS_AVX2
+#define HAS_SUMSQUAREERROR_AVX2
+uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count);
+#endif
+
+// TODO(fbarchard): Refactor into row function.
+LIBYUV_API
+uint64 ComputeSumSquareError(const uint8* src_a, const uint8* src_b,
+                             int count) {
+  // SumSquareError returns values 0 to 65535 for each squared difference.
+  // Up to 65536 of those can be summed and remain within a uint32.
+  // After each block of 65536 pixels, accumulate into a uint64.
+  const int kBlockSize = 65536;
+  int remainder = count & (kBlockSize - 1) & ~31;
+  uint64 sse = 0;
+  int i;
+  uint32 (*SumSquareError)(const uint8* src_a, const uint8* src_b, int count) =
+      SumSquareError_C;
+#if defined(HAS_SUMSQUAREERROR_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SumSquareError = SumSquareError_NEON;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    // Note only used for multiples of 16 so count is not checked.
+    SumSquareError = SumSquareError_SSE2;
+  }
+#endif
+#if defined(HAS_SUMSQUAREERROR_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    // Note only used for multiples of 32 so count is not checked.
+    SumSquareError = SumSquareError_AVX2;
+  }
+#endif
+#ifdef _OPENMP
+#pragma omp parallel for reduction(+: sse)
+#endif
+  for (i = 0; i < (count - (kBlockSize - 1)); i += kBlockSize) {
+    sse += SumSquareError(src_a + i, src_b + i, kBlockSize);
+  }
+  src_a += count & ~(kBlockSize - 1);
+  src_b += count & ~(kBlockSize - 1);
+  if (remainder) {
+    sse += SumSquareError(src_a, src_b, remainder);
+    src_a += remainder;
+    src_b += remainder;
+  }
+  remainder = count & 31;
+  if (remainder) {
+    sse += SumSquareError_C(src_a, src_b, remainder);
+  }
+  return sse;
+}
+
+LIBYUV_API
+uint64 ComputeSumSquareErrorPlane(const uint8* src_a, int stride_a,
+                                  const uint8* src_b, int stride_b,
+                                  int width, int height) {
+  uint64 sse = 0;
+  int h;
+  // Coalesce rows.
+  if (stride_a == width &&
+      stride_b == width) {
+    width *= height;
+    height = 1;
+    stride_a = stride_b = 0;
+  }
+  for (h = 0; h < height; ++h) {
+    sse += ComputeSumSquareError(src_a, src_b, width);
+    src_a += stride_a;
+    src_b += stride_b;
+  }
+  return sse;
+}
+
+LIBYUV_API
+double SumSquareErrorToPsnr(uint64 sse, uint64 count) {
+  double psnr;
+  if (sse > 0) {
+    double mse = (double)(count) / (double)(sse);
+    psnr = 10.0 * log10(255.0 * 255.0 * mse);
+  } else {
+    psnr = kMaxPsnr;      // Limit to prevent divide by 0
+  }
+
+  if (psnr > kMaxPsnr)
+    psnr = kMaxPsnr;
+
+  return psnr;
+}
+
+LIBYUV_API
+double CalcFramePsnr(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height) {
+  const uint64 samples = width * height;
+  const uint64 sse = ComputeSumSquareErrorPlane(src_a, stride_a,
+                                                src_b, stride_b,
+                                                width, height);
+  return SumSquareErrorToPsnr(sse, samples);
+}
+
+LIBYUV_API
+double I420Psnr(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height) {
+  const uint64 sse_y = ComputeSumSquareErrorPlane(src_y_a, stride_y_a,
+                                                  src_y_b, stride_y_b,
+                                                  width, height);
+  const int width_uv = (width + 1) >> 1;
+  const int height_uv = (height + 1) >> 1;
+  const uint64 sse_u = ComputeSumSquareErrorPlane(src_u_a, stride_u_a,
+                                                  src_u_b, stride_u_b,
+                                                  width_uv, height_uv);
+  const uint64 sse_v = ComputeSumSquareErrorPlane(src_v_a, stride_v_a,
+                                                  src_v_b, stride_v_b,
+                                                  width_uv, height_uv);
+  const uint64 samples = width * height + 2 * (width_uv * height_uv);
+  const uint64 sse = sse_y + sse_u + sse_v;
+  return SumSquareErrorToPsnr(sse, samples);
+}
+
+static const int64 cc1 =  26634;  // (64^2*(.01*255)^2
+static const int64 cc2 = 239708;  // (64^2*(.03*255)^2
+
+static double Ssim8x8_C(const uint8* src_a, int stride_a,
+                        const uint8* src_b, int stride_b) {
+  int64 sum_a = 0;
+  int64 sum_b = 0;
+  int64 sum_sq_a = 0;
+  int64 sum_sq_b = 0;
+  int64 sum_axb = 0;
+
+  int i;
+  for (i = 0; i < 8; ++i) {
+    int j;
+    for (j = 0; j < 8; ++j) {
+      sum_a += src_a[j];
+      sum_b += src_b[j];
+      sum_sq_a += src_a[j] * src_a[j];
+      sum_sq_b += src_b[j] * src_b[j];
+      sum_axb += src_a[j] * src_b[j];
+    }
+
+    src_a += stride_a;
+    src_b += stride_b;
+  }
+
+  {
+    const int64 count = 64;
+    // scale the constants by number of pixels
+    const int64 c1 = (cc1 * count * count) >> 12;
+    const int64 c2 = (cc2 * count * count) >> 12;
+
+    const int64 sum_a_x_sum_b = sum_a * sum_b;
+
+    const int64 ssim_n = (2 * sum_a_x_sum_b + c1) *
+                         (2 * count * sum_axb - 2 * sum_a_x_sum_b + c2);
+
+    const int64 sum_a_sq = sum_a*sum_a;
+    const int64 sum_b_sq = sum_b*sum_b;
+
+    const int64 ssim_d = (sum_a_sq + sum_b_sq + c1) *
+                         (count * sum_sq_a - sum_a_sq +
+                          count * sum_sq_b - sum_b_sq + c2);
+
+    if (ssim_d == 0.0) {
+      return DBL_MAX;
+    }
+    return ssim_n * 1.0 / ssim_d;
+  }
+}
+
+// We are using a 8x8 moving window with starting location of each 8x8 window
+// on the 4x4 pixel grid. Such arrangement allows the windows to overlap
+// block boundaries to penalize blocking artifacts.
+LIBYUV_API
+double CalcFrameSsim(const uint8* src_a, int stride_a,
+                     const uint8* src_b, int stride_b,
+                     int width, int height) {
+  int samples = 0;
+  double ssim_total = 0;
+  double (*Ssim8x8)(const uint8* src_a, int stride_a,
+                    const uint8* src_b, int stride_b) = Ssim8x8_C;
+
+  // sample point start with each 4x4 location
+  int i;
+  for (i = 0; i < height - 8; i += 4) {
+    int j;
+    for (j = 0; j < width - 8; j += 4) {
+      ssim_total += Ssim8x8(src_a + j, stride_a, src_b + j, stride_b);
+      samples++;
+    }
+
+    src_a += stride_a * 4;
+    src_b += stride_b * 4;
+  }
+
+  ssim_total /= samples;
+  return ssim_total;
+}
+
+LIBYUV_API
+double I420Ssim(const uint8* src_y_a, int stride_y_a,
+                const uint8* src_u_a, int stride_u_a,
+                const uint8* src_v_a, int stride_v_a,
+                const uint8* src_y_b, int stride_y_b,
+                const uint8* src_u_b, int stride_u_b,
+                const uint8* src_v_b, int stride_v_b,
+                int width, int height) {
+  const double ssim_y = CalcFrameSsim(src_y_a, stride_y_a,
+                                      src_y_b, stride_y_b, width, height);
+  const int width_uv = (width + 1) >> 1;
+  const int height_uv = (height + 1) >> 1;
+  const double ssim_u = CalcFrameSsim(src_u_a, stride_u_a,
+                                      src_u_b, stride_u_b,
+                                      width_uv, height_uv);
+  const double ssim_v = CalcFrameSsim(src_v_a, stride_v_a,
+                                      src_v_b, stride_v_b,
+                                      width_uv, height_uv);
+  return ssim_y * 0.8 + 0.1 * (ssim_u + ssim_v);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/compare_common.cc b/third_party/aom/third_party/libyuv/source/compare_common.cc
new file mode 100644
index 0000000000..c546b51829
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/compare_common.cc
@@ -0,0 +1,42 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+uint32 SumSquareError_C(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse = 0u;
+  int i;
+  for (i = 0; i < count; ++i) {
+    int diff = src_a[i] - src_b[i];
+    sse += (uint32)(diff * diff);
+  }
+  return sse;
+}
+
+// hash seed of 5381 recommended.
+// Internal C version of HashDjb2 with int sized count for efficiency.
+uint32 HashDjb2_C(const uint8* src, int count, uint32 seed) {
+  uint32 hash = seed;
+  int i;
+  for (i = 0; i < count; ++i) {
+    hash += (hash << 5) + src[i];
+  }
+  return hash;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/compare_gcc.cc b/third_party/aom/third_party/libyuv/source/compare_gcc.cc
new file mode 100644
index 0000000000..247cb33bba
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/compare_gcc.cc
@@ -0,0 +1,152 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+  uint32 sse;
+  asm volatile (  // NOLINT
+    "pxor      %%xmm0,%%xmm0                   \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10, 0) ",%0          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10, 1) ",%1          \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psubusb   %%xmm2,%%xmm1                   \n"
+    "psubusb   %%xmm3,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpckhbw %%xmm5,%%xmm2                   \n"
+    "pmaddwd   %%xmm1,%%xmm1                   \n"
+    "pmaddwd   %%xmm2,%%xmm2                   \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+
+    "pshufd    $0xee,%%xmm0,%%xmm1             \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "pshufd    $0x1,%%xmm0,%%xmm1              \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "movd      %%xmm0,%3                       \n"
+
+  : "+r"(src_a),      // %0
+    "+r"(src_b),      // %1
+    "+r"(count),      // %2
+    "=g"(sse)         // %3
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );  // NOLINT
+  return sse;
+}
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
+#define HAS_HASHDJB2_SSE41
+static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
+static uvec32 kHashMul0 = {
+  0x0c3525e1,  // 33 ^ 15
+  0xa3476dc1,  // 33 ^ 14
+  0x3b4039a1,  // 33 ^ 13
+  0x4f5f0981,  // 33 ^ 12
+};
+static uvec32 kHashMul1 = {
+  0x30f35d61,  // 33 ^ 11
+  0x855cb541,  // 33 ^ 10
+  0x040a9121,  // 33 ^ 9
+  0x747c7101,  // 33 ^ 8
+};
+static uvec32 kHashMul2 = {
+  0xec41d4e1,  // 33 ^ 7
+  0x4cfa3cc1,  // 33 ^ 6
+  0x025528a1,  // 33 ^ 5
+  0x00121881,  // 33 ^ 4
+};
+static uvec32 kHashMul3 = {
+  0x00008c61,  // 33 ^ 3
+  0x00000441,  // 33 ^ 2
+  0x00000021,  // 33 ^ 1
+  0x00000001,  // 33 ^ 0
+};
+
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+  uint32 hash;
+  asm volatile (  // NOLINT
+    "movd      %2,%%xmm0                       \n"
+    "pxor      %%xmm7,%%xmm7                   \n"
+    "movdqa    %4,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10, 0) ",%0          \n"
+    "pmulld    %%xmm6,%%xmm0                   \n"
+    "movdqa    %5,%%xmm5                       \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm7,%%xmm3                   \n"
+    "pmulld    %%xmm5,%%xmm3                   \n"
+    "movdqa    %6,%%xmm5                       \n"
+    "movdqa    %%xmm2,%%xmm4                   \n"
+    "punpckhwd %%xmm7,%%xmm4                   \n"
+    "pmulld    %%xmm5,%%xmm4                   \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "punpckhbw %%xmm7,%%xmm1                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm7,%%xmm2                   \n"
+    "pmulld    %%xmm5,%%xmm2                   \n"
+    "movdqa    %8,%%xmm5                       \n"
+    "punpckhwd %%xmm7,%%xmm1                   \n"
+    "pmulld    %%xmm5,%%xmm1                   \n"
+    "paddd     %%xmm4,%%xmm3                   \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm1                   \n"
+    "pshufd    $0xe,%%xmm1,%%xmm2              \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "pshufd    $0x1,%%xmm1,%%xmm2              \n"
+    "paddd     %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm1,%%xmm0                   \n"
+    "sub       $0x10,%1                        \n"
+    "jg        1b                              \n"
+    "movd      %%xmm0,%3                       \n"
+  : "+r"(src),        // %0
+    "+r"(count),      // %1
+    "+rm"(seed),      // %2
+    "=g"(hash)        // %3
+  : "m"(kHash16x33),  // %4
+    "m"(kHashMul0),   // %5
+    "m"(kHashMul1),   // %6
+    "m"(kHashMul2),   // %7
+    "m"(kHashMul3)    // %8
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );  // NOLINT
+  return hash;
+}
+#endif  // defined(__x86_64__) || (defined(__i386__) && !defined(__pic__)))
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/third_party/aom/third_party/libyuv/source/compare_neon.cc b/third_party/aom/third_party/libyuv/source/compare_neon.cc
new file mode 100644
index 0000000000..ef006ec41c
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/compare_neon.cc
@@ -0,0 +1,65 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  volatile uint32 sse;
+  asm volatile (
+    "vmov.u8    q8, #0                         \n"
+    "vmov.u8    q10, #0                        \n"
+    "vmov.u8    q9, #0                         \n"
+    "vmov.u8    q11, #0                        \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"
+    "subs       %2, %2, #16                    \n"
+    "vsubl.u8   q2, d0, d2                     \n"
+    "vsubl.u8   q3, d1, d3                     \n"
+    "vmlal.s16  q8, d4, d4                     \n"
+    "vmlal.s16  q9, d6, d6                     \n"
+    "vmlal.s16  q10, d5, d5                    \n"
+    "vmlal.s16  q11, d7, d7                    \n"
+    "bgt        1b                             \n"
+
+    "vadd.u32   q8, q8, q9                     \n"
+    "vadd.u32   q10, q10, q11                  \n"
+    "vadd.u32   q11, q8, q10                   \n"
+    "vpaddl.u32 q1, q11                        \n"
+    "vadd.u64   d0, d2, d3                     \n"
+    "vmov.32    %3, d0[0]                      \n"
+    : "+r"(src_a),
+      "+r"(src_b),
+      "+r"(count),
+      "=r"(sse)
+    :
+    : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11");
+  return sse;
+}
+
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/compare_neon64.cc b/third_party/aom/third_party/libyuv/source/compare_neon64.cc
new file mode 100644
index 0000000000..6d1e5e1bc9
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/compare_neon64.cc
@@ -0,0 +1,63 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+uint32 SumSquareError_NEON(const uint8* src_a, const uint8* src_b, int count) {
+  volatile uint32 sse;
+  asm volatile (
+    "eor        v16.16b, v16.16b, v16.16b      \n"
+    "eor        v18.16b, v18.16b, v18.16b      \n"
+    "eor        v17.16b, v17.16b, v17.16b      \n"
+    "eor        v19.16b, v19.16b, v19.16b      \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"
+    MEMACCESS(1)
+    "ld1        {v1.16b}, [%1], #16            \n"
+    "subs       %w2, %w2, #16                  \n"
+    "usubl      v2.8h, v0.8b, v1.8b            \n"
+    "usubl2     v3.8h, v0.16b, v1.16b          \n"
+    "smlal      v16.4s, v2.4h, v2.4h           \n"
+    "smlal      v17.4s, v3.4h, v3.4h           \n"
+    "smlal2     v18.4s, v2.8h, v2.8h           \n"
+    "smlal2     v19.4s, v3.8h, v3.8h           \n"
+    "b.gt       1b                             \n"
+
+    "add        v16.4s, v16.4s, v17.4s         \n"
+    "add        v18.4s, v18.4s, v19.4s         \n"
+    "add        v19.4s, v16.4s, v18.4s         \n"
+    "addv       s0, v19.4s                     \n"
+    "fmov       %w3, s0                        \n"
+    : "+r"(src_a),
+      "+r"(src_b),
+      "+r"(count),
+      "=r"(sse)
+    :
+    : "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
+  return sse;
+}
+
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/compare_win.cc b/third_party/aom/third_party/libyuv/source/compare_win.cc
new file mode 100644
index 0000000000..19806f2750
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/compare_win.cc
@@ -0,0 +1,229 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
+    defined(_MSC_VER) && !defined(__clang__)
+
+__declspec(naked)
+uint32 SumSquareError_SSE2(const uint8* src_a, const uint8* src_b, int count) {
+  __asm {
+    mov        eax, [esp + 4]    // src_a
+    mov        edx, [esp + 8]    // src_b
+    mov        ecx, [esp + 12]   // count
+    pxor       xmm0, xmm0
+    pxor       xmm5, xmm5
+
+  wloop:
+    movdqu     xmm1, [eax]
+    lea        eax,  [eax + 16]
+    movdqu     xmm2, [edx]
+    lea        edx,  [edx + 16]
+    movdqa     xmm3, xmm1  // abs trick
+    psubusb    xmm1, xmm2
+    psubusb    xmm2, xmm3
+    por        xmm1, xmm2
+    movdqa     xmm2, xmm1
+    punpcklbw  xmm1, xmm5
+    punpckhbw  xmm2, xmm5
+    pmaddwd    xmm1, xmm1
+    pmaddwd    xmm2, xmm2
+    paddd      xmm0, xmm1
+    paddd      xmm0, xmm2
+    sub        ecx, 16
+    jg         wloop
+
+    pshufd     xmm1, xmm0, 0xee
+    paddd      xmm0, xmm1
+    pshufd     xmm1, xmm0, 0x01
+    paddd      xmm0, xmm1
+    movd       eax, xmm0
+    ret
+  }
+}
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+// C4752: found Intel(R) Advanced Vector Extensions; consider using /arch:AVX.
+#pragma warning(disable: 4752)
+__declspec(naked)
+uint32 SumSquareError_AVX2(const uint8* src_a, const uint8* src_b, int count) {
+  __asm {
+    mov        eax, [esp + 4]    // src_a
+    mov        edx, [esp + 8]    // src_b
+    mov        ecx, [esp + 12]   // count
+    vpxor      ymm0, ymm0, ymm0  // sum
+    vpxor      ymm5, ymm5, ymm5  // constant 0 for unpck
+    sub        edx, eax
+
+  wloop:
+    vmovdqu    ymm1, [eax]
+    vmovdqu    ymm2, [eax + edx]
+    lea        eax,  [eax + 32]
+    vpsubusb   ymm3, ymm1, ymm2  // abs difference trick
+    vpsubusb   ymm2, ymm2, ymm1
+    vpor       ymm1, ymm2, ymm3
+    vpunpcklbw ymm2, ymm1, ymm5  // u16.  mutates order.
+    vpunpckhbw ymm1, ymm1, ymm5
+    vpmaddwd   ymm2, ymm2, ymm2  // square + hadd to u32.
+    vpmaddwd   ymm1, ymm1, ymm1
+    vpaddd     ymm0, ymm0, ymm1
+    vpaddd     ymm0, ymm0, ymm2
+    sub        ecx, 32
+    jg         wloop
+
+    vpshufd    ymm1, ymm0, 0xee  // 3, 2 + 1, 0 both lanes.
+    vpaddd     ymm0, ymm0, ymm1
+    vpshufd    ymm1, ymm0, 0x01  // 1 + 0 both lanes.
+    vpaddd     ymm0, ymm0, ymm1
+    vpermq     ymm1, ymm0, 0x02  // high + low lane.
+    vpaddd     ymm0, ymm0, ymm1
+    vmovd      eax, xmm0
+    vzeroupper
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+
+#define HAS_HASHDJB2_SSE41
+static uvec32 kHash16x33 = { 0x92d9e201, 0, 0, 0 };  // 33 ^ 16
+static uvec32 kHashMul0 = {
+  0x0c3525e1,  // 33 ^ 15
+  0xa3476dc1,  // 33 ^ 14
+  0x3b4039a1,  // 33 ^ 13
+  0x4f5f0981,  // 33 ^ 12
+};
+static uvec32 kHashMul1 = {
+  0x30f35d61,  // 33 ^ 11
+  0x855cb541,  // 33 ^ 10
+  0x040a9121,  // 33 ^ 9
+  0x747c7101,  // 33 ^ 8
+};
+static uvec32 kHashMul2 = {
+  0xec41d4e1,  // 33 ^ 7
+  0x4cfa3cc1,  // 33 ^ 6
+  0x025528a1,  // 33 ^ 5
+  0x00121881,  // 33 ^ 4
+};
+static uvec32 kHashMul3 = {
+  0x00008c61,  // 33 ^ 3
+  0x00000441,  // 33 ^ 2
+  0x00000021,  // 33 ^ 1
+  0x00000001,  // 33 ^ 0
+};
+
+// 27: 66 0F 38 40 C6     pmulld      xmm0,xmm6
+// 44: 66 0F 38 40 DD     pmulld      xmm3,xmm5
+// 59: 66 0F 38 40 E5     pmulld      xmm4,xmm5
+// 72: 66 0F 38 40 D5     pmulld      xmm2,xmm5
+// 83: 66 0F 38 40 CD     pmulld      xmm1,xmm5
+#define pmulld(reg) _asm _emit 0x66 _asm _emit 0x0F _asm _emit 0x38 \
+    _asm _emit 0x40 _asm _emit reg
+
+__declspec(naked)
+uint32 HashDjb2_SSE41(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+
+    pxor       xmm7, xmm7        // constant 0 for unpck
+    movdqa     xmm6, kHash16x33
+
+  wloop:
+    movdqu     xmm1, [eax]       // src[0-15]
+    lea        eax, [eax + 16]
+    pmulld(0xc6)                 // pmulld      xmm0,xmm6  hash *= 33 ^ 16
+    movdqa     xmm5, kHashMul0
+    movdqa     xmm2, xmm1
+    punpcklbw  xmm2, xmm7        // src[0-7]
+    movdqa     xmm3, xmm2
+    punpcklwd  xmm3, xmm7        // src[0-3]
+    pmulld(0xdd)                 // pmulld     xmm3, xmm5
+    movdqa     xmm5, kHashMul1
+    movdqa     xmm4, xmm2
+    punpckhwd  xmm4, xmm7        // src[4-7]
+    pmulld(0xe5)                 // pmulld     xmm4, xmm5
+    movdqa     xmm5, kHashMul2
+    punpckhbw  xmm1, xmm7        // src[8-15]
+    movdqa     xmm2, xmm1
+    punpcklwd  xmm2, xmm7        // src[8-11]
+    pmulld(0xd5)                 // pmulld     xmm2, xmm5
+    movdqa     xmm5, kHashMul3
+    punpckhwd  xmm1, xmm7        // src[12-15]
+    pmulld(0xcd)                 // pmulld     xmm1, xmm5
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    paddd      xmm1, xmm3
+
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    sub        ecx, 16
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+
+// Visual C 2012 required for AVX2.
+#if _MSC_VER >= 1700
+__declspec(naked)
+uint32 HashDjb2_AVX2(const uint8* src, int count, uint32 seed) {
+  __asm {
+    mov        eax, [esp + 4]    // src
+    mov        ecx, [esp + 8]    // count
+    movd       xmm0, [esp + 12]  // seed
+    movdqa     xmm6, kHash16x33
+
+  wloop:
+    vpmovzxbd  xmm3, dword ptr [eax]  // src[0-3]
+    pmulld     xmm0, xmm6  // hash *= 33 ^ 16
+    vpmovzxbd  xmm4, dword ptr [eax + 4]  // src[4-7]
+    pmulld     xmm3, kHashMul0
+    vpmovzxbd  xmm2, dword ptr [eax + 8]  // src[8-11]
+    pmulld     xmm4, kHashMul1
+    vpmovzxbd  xmm1, dword ptr [eax + 12]  // src[12-15]
+    pmulld     xmm2, kHashMul2
+    lea        eax, [eax + 16]
+    pmulld     xmm1, kHashMul3
+    paddd      xmm3, xmm4        // add 16 results
+    paddd      xmm1, xmm2
+    paddd      xmm1, xmm3
+    pshufd     xmm2, xmm1, 0x0e  // upper 2 dwords
+    paddd      xmm1, xmm2
+    pshufd     xmm2, xmm1, 0x01
+    paddd      xmm1, xmm2
+    paddd      xmm0, xmm1
+    sub        ecx, 16
+    jg         wloop
+
+    movd       eax, xmm0         // return hash
+    ret
+  }
+}
+#endif  // _MSC_VER >= 1700
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/convert.cc b/third_party/aom/third_party/libyuv/source/convert.cc
new file mode 100644
index 0000000000..3ad6bd7a4b
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/convert.cc
@@ -0,0 +1,1389 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+#include "libyuv/scale.h"  // For ScalePlane()
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// Any I4xx To I420 format with mirroring.
+static int I4xxToI420(const uint8* src_y, int src_stride_y,
+                      const uint8* src_u, int src_stride_u,
+                      const uint8* src_v, int src_stride_v,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int src_y_width, int src_y_height,
+                      int src_uv_width, int src_uv_height) {
+  const int dst_y_width = Abs(src_y_width);
+  const int dst_y_height = Abs(src_y_height);
+  const int dst_uv_width = SUBSAMPLE(dst_y_width, 1, 1);
+  const int dst_uv_height = SUBSAMPLE(dst_y_height, 1, 1);
+  if (src_y_width == 0 || src_y_height == 0 ||
+      src_uv_width == 0 || src_uv_height == 0) {
+    return -1;
+  }
+  ScalePlane(src_y, src_stride_y, src_y_width, src_y_height,
+             dst_y, dst_stride_y, dst_y_width, dst_y_height,
+             kFilterBilinear);
+  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height,
+             dst_u, dst_stride_u, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height,
+             dst_v, dst_stride_v, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  return 0;
+}
+
+// Copy I420 with optional flipping
+// TODO(fbarchard): Use Scale plane which supports mirroring, but ensure
+// is does row coalescing.
+LIBYUV_API
+int I420Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  // Copy UV planes.
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
+  return 0;
+}
+
+// 422 chroma is 1/2 width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I422ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int src_uv_width = SUBSAMPLE(width, 1, 1);
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    src_uv_width, height);
+}
+
+// 444 chroma is 1x width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I444ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    width, height);
+}
+
+// 411 chroma is 1/4 width, 1x height
+// 420 chroma is 1/2 width, 1/2 height
+LIBYUV_API
+int I411ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int src_uv_width = SUBSAMPLE(width, 3, 2);
+  return I4xxToI420(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    src_uv_width, height);
+}
+
+// I400 is greyscale typically used in MJPG
+LIBYUV_API
+int I400ToI420(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  SetPlane(dst_u, dst_stride_u, halfwidth, halfheight, 128);
+  SetPlane(dst_v, dst_stride_v, halfwidth, halfheight, 128);
+  return 0;
+}
+
+static void CopyPlane2(const uint8* src, int src_stride_0, int src_stride_1,
+                       uint8* dst, int dst_stride,
+                       int width, int height) {
+  int y;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height - 1; y += 2) {
+    CopyRow(src, dst, width);
+    CopyRow(src + src_stride_0, dst + dst_stride, width);
+    src += src_stride_0 + src_stride_1;
+    dst += dst_stride * 2;
+  }
+  if (height & 1) {
+    CopyRow(src, dst, width);
+  }
+}
+
+// Support converting from FOURCC_M420
+// Useful for bandwidth constrained transports like USB 1.0 and 2.0 and for
+// easy conversion to I420.
+// M420 format description:
+// M420 is row biplanar 420: 2 rows of Y and 1 row of UV.
+// Chroma is half width / half height. (420)
+// src_stride_m420 is row planar. Normally this will be the width in pixels.
+//   The UV plane is half width, but 2 values, so src_stride_m420 applies to
+//   this as well as the two Y planes.
+static int X420ToI420(const uint8* src_y,
+                      int src_stride_y0, int src_stride_y1,
+                      const uint8* src_uv, int src_stride_uv,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) =
+      SplitUVRow_C;
+  if (!src_y || !src_uv ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_u = dst_u + (halfheight - 1) * dst_stride_u;
+    dst_v = dst_v + (halfheight - 1) * dst_stride_v;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_u = -dst_stride_u;
+    dst_stride_v = -dst_stride_v;
+  }
+  // Coalesce rows.
+  if (src_stride_y0 == width &&
+      src_stride_y1 == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y0 = src_stride_y1 = dst_stride_y = 0;
+  }
+  // Coalesce rows.
+  if (src_stride_uv == halfwidth * 2 &&
+      dst_stride_u == halfwidth &&
+      dst_stride_v == halfwidth) {
+    halfwidth *= halfheight;
+    halfheight = 1;
+    src_stride_uv = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_SPLITUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitUVRow = SplitUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitUVRow = SplitUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      SplitUVRow = SplitUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitUVRow = SplitUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_uv, 4) && IS_ALIGNED(src_stride_uv, 4) &&
+      IS_ALIGNED(dst_u, 4) && IS_ALIGNED(dst_stride_u, 4) &&
+      IS_ALIGNED(dst_v, 4) && IS_ALIGNED(dst_stride_v, 4)) {
+    SplitUVRow = SplitUVRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      SplitUVRow = SplitUVRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (dst_y) {
+    if (src_stride_y0 == src_stride_y1) {
+      CopyPlane(src_y, src_stride_y0, dst_y, dst_stride_y, width, height);
+    } else {
+      CopyPlane2(src_y, src_stride_y0, src_stride_y1, dst_y, dst_stride_y,
+                 width, height);
+    }
+  }
+
+  for (y = 0; y < halfheight; ++y) {
+    // Copy a row of UV.
+    SplitUVRow(src_uv, dst_u, dst_v, halfwidth);
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+    src_uv += src_stride_uv;
+  }
+  return 0;
+}
+
+// Convert NV12 to I420.
+LIBYUV_API
+int NV12ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_y, src_stride_y, src_stride_y,
+                    src_uv, src_stride_uv,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height);
+}
+
+// Convert NV21 to I420.  Same as NV12 but u and v pointers swapped.
+LIBYUV_API
+int NV21ToI420(const uint8* src_y, int src_stride_y,
+               const uint8* src_vu, int src_stride_vu,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_y, src_stride_y, src_stride_y,
+                    src_vu, src_stride_vu,
+                    dst_y, dst_stride_y,
+                    dst_v, dst_stride_v,
+                    dst_u, dst_stride_u,
+                    width, height);
+}
+
+// Convert M420 to I420.
+LIBYUV_API
+int M420ToI420(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  return X420ToI420(src_m420, src_stride_m420, src_stride_m420 * 2,
+                    src_m420 + src_stride_m420 * 2, src_stride_m420 * 3,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height);
+}
+
+// Convert YUY2 to I420.
+LIBYUV_API
+int YUY2ToI420(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*YUY2ToUVRow)(const uint8* src_yuy2, int src_stride_yuy2,
+      uint8* dst_u, uint8* dst_v, int pix) = YUY2ToUVRow_C;
+  void (*YUY2ToYRow)(const uint8* src_yuy2,
+      uint8* dst_y, int pix) = YUY2ToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    YUY2ToUVRow = YUY2ToUVRow_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUVRow = YUY2ToUVRow_SSE2;
+      YUY2ToYRow = YUY2ToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YUY2ToUVRow = YUY2ToUVRow_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUVRow = YUY2ToUVRow_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    YUY2ToUVRow = YUY2ToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUVRow = YUY2ToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    YUY2ToUVRow(src_yuy2, src_stride_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    YUY2ToYRow(src_yuy2 + src_stride_yuy2, dst_y + dst_stride_y, width);
+    src_yuy2 += src_stride_yuy2 * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    YUY2ToUVRow(src_yuy2, 0, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert UYVY to I420.
+LIBYUV_API
+int UYVYToI420(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*UYVYToUVRow)(const uint8* src_uyvy, int src_stride_uyvy,
+      uint8* dst_u, uint8* dst_v, int pix) = UYVYToUVRow_C;
+  void (*UYVYToYRow)(const uint8* src_uyvy,
+      uint8* dst_y, int pix) = UYVYToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    UYVYToUVRow = UYVYToUVRow_Any_SSE2;
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToUVRow = UYVYToUVRow_SSE2;
+      UYVYToYRow = UYVYToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    UYVYToUVRow = UYVYToUVRow_Any_AVX2;
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToUVRow = UYVYToUVRow_AVX2;
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    UYVYToUVRow = UYVYToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+      UYVYToUVRow = UYVYToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    UYVYToUVRow(src_uyvy, src_stride_uyvy, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+    UYVYToYRow(src_uyvy + src_stride_uyvy, dst_y + dst_stride_y, width);
+    src_uyvy += src_stride_uyvy * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    UYVYToUVRow(src_uyvy, 0, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert ARGB to I420.
+LIBYUV_API
+int ARGBToI420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ARGBToUVRow(src_argb, src_stride_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+    src_argb += src_stride_argb * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ARGBToUVRow(src_argb, 0, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert BGRA to I420.
+LIBYUV_API
+int BGRAToI420(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*BGRAToUVRow)(const uint8* src_bgra0, int src_stride_bgra,
+      uint8* dst_u, uint8* dst_v, int width) = BGRAToUVRow_C;
+  void (*BGRAToYRow)(const uint8* src_bgra, uint8* dst_y, int pix) =
+      BGRAToYRow_C;
+  if (!src_bgra ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_bgra = src_bgra + (height - 1) * src_stride_bgra;
+    src_stride_bgra = -src_stride_bgra;
+  }
+#if defined(HAS_BGRATOYROW_SSSE3) && defined(HAS_BGRATOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    BGRAToUVRow = BGRAToUVRow_Any_SSSE3;
+    BGRAToYRow = BGRAToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      BGRAToUVRow = BGRAToUVRow_SSSE3;
+      BGRAToYRow = BGRAToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_BGRATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    BGRAToYRow = BGRAToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      BGRAToYRow = BGRAToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_BGRATOUVROW_NEON)
+    if (TestCpuFlag(kCpuHasNEON)) {
+      BGRAToUVRow = BGRAToUVRow_Any_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        BGRAToUVRow = BGRAToUVRow_NEON;
+      }
+    }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    BGRAToUVRow(src_bgra, src_stride_bgra, dst_u, dst_v, width);
+    BGRAToYRow(src_bgra, dst_y, width);
+    BGRAToYRow(src_bgra + src_stride_bgra, dst_y + dst_stride_y, width);
+    src_bgra += src_stride_bgra * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    BGRAToUVRow(src_bgra, 0, dst_u, dst_v, width);
+    BGRAToYRow(src_bgra, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert ABGR to I420.
+LIBYUV_API
+int ABGRToI420(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ABGRToUVRow)(const uint8* src_abgr0, int src_stride_abgr,
+      uint8* dst_u, uint8* dst_v, int width) = ABGRToUVRow_C;
+  void (*ABGRToYRow)(const uint8* src_abgr, uint8* dst_y, int pix) =
+      ABGRToYRow_C;
+  if (!src_abgr ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_abgr = src_abgr + (height - 1) * src_stride_abgr;
+    src_stride_abgr = -src_stride_abgr;
+  }
+#if defined(HAS_ABGRTOYROW_SSSE3) && defined(HAS_ABGRTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ABGRToUVRow = ABGRToUVRow_Any_SSSE3;
+    ABGRToYRow = ABGRToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_SSSE3;
+      ABGRToYRow = ABGRToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToYRow = ABGRToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ABGRToYRow = ABGRToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ABGRToUVRow = ABGRToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToUVRow = ABGRToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ABGRToUVRow(src_abgr, src_stride_abgr, dst_u, dst_v, width);
+    ABGRToYRow(src_abgr, dst_y, width);
+    ABGRToYRow(src_abgr + src_stride_abgr, dst_y + dst_stride_y, width);
+    src_abgr += src_stride_abgr * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ABGRToUVRow(src_abgr, 0, dst_u, dst_v, width);
+    ABGRToYRow(src_abgr, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert RGBA to I420.
+LIBYUV_API
+int RGBAToI420(const uint8* src_rgba, int src_stride_rgba,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*RGBAToUVRow)(const uint8* src_rgba0, int src_stride_rgba,
+      uint8* dst_u, uint8* dst_v, int width) = RGBAToUVRow_C;
+  void (*RGBAToYRow)(const uint8* src_rgba, uint8* dst_y, int pix) =
+      RGBAToYRow_C;
+  if (!src_rgba ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgba = src_rgba + (height - 1) * src_stride_rgba;
+    src_stride_rgba = -src_stride_rgba;
+  }
+#if defined(HAS_RGBATOYROW_SSSE3) && defined(HAS_RGBATOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGBAToUVRow = RGBAToUVRow_Any_SSSE3;
+    RGBAToYRow = RGBAToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToUVRow = RGBAToUVRow_SSSE3;
+      RGBAToYRow = RGBAToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGBAToYRow = RGBAToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGBAToYRow = RGBAToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGBAToUVRow = RGBAToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToUVRow = RGBAToUVRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    RGBAToUVRow(src_rgba, src_stride_rgba, dst_u, dst_v, width);
+    RGBAToYRow(src_rgba, dst_y, width);
+    RGBAToYRow(src_rgba + src_stride_rgba, dst_y + dst_stride_y, width);
+    src_rgba += src_stride_rgba * 2;
+    dst_y += dst_stride_y * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    RGBAToUVRow(src_rgba, 0, dst_u, dst_v, width);
+    RGBAToYRow(src_rgba, dst_y, width);
+  }
+  return 0;
+}
+
+// Convert RGB24 to I420.
+LIBYUV_API
+int RGB24ToI420(const uint8* src_rgb24, int src_stride_rgb24,
+                uint8* dst_y, int dst_stride_y,
+                uint8* dst_u, int dst_stride_u,
+                uint8* dst_v, int dst_stride_v,
+                int width, int height) {
+  int y;
+#if defined(HAS_RGB24TOYROW_NEON)
+  void (*RGB24ToUVRow)(const uint8* src_rgb24, int src_stride_rgb24,
+      uint8* dst_u, uint8* dst_v, int width) = RGB24ToUVRow_C;
+  void (*RGB24ToYRow)(const uint8* src_rgb24, uint8* dst_y, int pix) =
+      RGB24ToYRow_C;
+#else
+  void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB24ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#endif
+  if (!src_rgb24 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+
+// Neon version does direct RGB24 to YUV.
+#if defined(HAS_RGB24TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB24ToUVRow = RGB24ToUVRow_Any_NEON;
+    RGB24ToYRow = RGB24ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB24ToYRow = RGB24ToYRow_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RGB24ToUVRow = RGB24ToUVRow_NEON;
+      }
+    }
+  }
+// Other platforms do intermediate conversion from RGB24 to ARGB.
+#else
+#if defined(HAS_RGB24TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RGB24TOYROW_NEON)
+      RGB24ToUVRow(src_rgb24, src_stride_rgb24, dst_u, dst_v, width);
+      RGB24ToYRow(src_rgb24, dst_y, width);
+      RGB24ToYRow(src_rgb24 + src_stride_rgb24, dst_y + dst_stride_y, width);
+#else
+      RGB24ToARGBRow(src_rgb24, row, width);
+      RGB24ToARGBRow(src_rgb24 + src_stride_rgb24, row + kRowSize, width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_rgb24 += src_stride_rgb24 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_RGB24TOYROW_NEON)
+      RGB24ToUVRow(src_rgb24, 0, dst_u, dst_v, width);
+      RGB24ToYRow(src_rgb24, dst_y, width);
+#else
+      RGB24ToARGBRow(src_rgb24, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RGB24TOYROW_NEON)
+    free_aligned_buffer_64(row);
+  }
+#endif
+  return 0;
+}
+
+// Convert RAW to I420.
+LIBYUV_API
+int RAWToI420(const uint8* src_raw, int src_stride_raw,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int width, int height) {
+  int y;
+#if defined(HAS_RAWTOYROW_NEON)
+  void (*RAWToUVRow)(const uint8* src_raw, int src_stride_raw,
+      uint8* dst_u, uint8* dst_v, int width) = RAWToUVRow_C;
+  void (*RAWToYRow)(const uint8* src_raw, uint8* dst_y, int pix) =
+      RAWToYRow_C;
+#else
+  void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RAWToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#endif
+  if (!src_raw || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+
+// Neon version does direct RAW to YUV.
+#if defined(HAS_RAWTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToUVRow = RAWToUVRow_Any_NEON;
+    RAWToYRow = RAWToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RAWToYRow = RAWToYRow_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RAWToUVRow = RAWToUVRow_NEON;
+      }
+    }
+  }
+// Other platforms do intermediate conversion from RAW to ARGB.
+#else
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RAWTOYROW_NEON)
+      RAWToUVRow(src_raw, src_stride_raw, dst_u, dst_v, width);
+      RAWToYRow(src_raw, dst_y, width);
+      RAWToYRow(src_raw + src_stride_raw, dst_y + dst_stride_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      RAWToARGBRow(src_raw + src_stride_raw, row + kRowSize, width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_raw += src_stride_raw * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_RAWTOYROW_NEON)
+      RAWToUVRow(src_raw, 0, dst_u, dst_v, width);
+      RAWToYRow(src_raw, dst_y, width);
+#else
+      RAWToARGBRow(src_raw, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RAWTOYROW_NEON)
+    free_aligned_buffer_64(row);
+  }
+#endif
+  return 0;
+}
+
+// Convert RGB565 to I420.
+LIBYUV_API
+int RGB565ToI420(const uint8* src_rgb565, int src_stride_rgb565,
+                 uint8* dst_y, int dst_stride_y,
+                 uint8* dst_u, int dst_stride_u,
+                 uint8* dst_v, int dst_stride_v,
+                 int width, int height) {
+  int y;
+#if defined(HAS_RGB565TOYROW_NEON)
+  void (*RGB565ToUVRow)(const uint8* src_rgb565, int src_stride_rgb565,
+      uint8* dst_u, uint8* dst_v, int width) = RGB565ToUVRow_C;
+  void (*RGB565ToYRow)(const uint8* src_rgb565, uint8* dst_y, int pix) =
+      RGB565ToYRow_C;
+#else
+  void (*RGB565ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB565ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#endif
+  if (!src_rgb565 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb565 = src_rgb565 + (height - 1) * src_stride_rgb565;
+    src_stride_rgb565 = -src_stride_rgb565;
+  }
+
+// Neon version does direct RGB565 to YUV.
+#if defined(HAS_RGB565TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB565ToUVRow = RGB565ToUVRow_Any_NEON;
+    RGB565ToYRow = RGB565ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToYRow = RGB565ToYRow_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        RGB565ToUVRow = RGB565ToUVRow_NEON;
+      }
+    }
+  }
+// Other platforms do intermediate conversion from RGB565 to ARGB.
+#else
+#if defined(HAS_RGB565TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_RGB565TOYROW_NEON)
+      RGB565ToUVRow(src_rgb565, src_stride_rgb565, dst_u, dst_v, width);
+      RGB565ToYRow(src_rgb565, dst_y, width);
+      RGB565ToYRow(src_rgb565 + src_stride_rgb565, dst_y + dst_stride_y, width);
+#else
+      RGB565ToARGBRow(src_rgb565, row, width);
+      RGB565ToARGBRow(src_rgb565 + src_stride_rgb565, row + kRowSize, width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_rgb565 += src_stride_rgb565 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_RGB565TOYROW_NEON)
+      RGB565ToUVRow(src_rgb565, 0, dst_u, dst_v, width);
+      RGB565ToYRow(src_rgb565, dst_y, width);
+#else
+      RGB565ToARGBRow(src_rgb565, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_RGB565TOYROW_NEON)
+    free_aligned_buffer_64(row);
+  }
+#endif
+  return 0;
+}
+
+// Convert ARGB1555 to I420.
+LIBYUV_API
+int ARGB1555ToI420(const uint8* src_argb1555, int src_stride_argb1555,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height) {
+  int y;
+#if defined(HAS_ARGB1555TOYROW_NEON)
+  void (*ARGB1555ToUVRow)(const uint8* src_argb1555, int src_stride_argb1555,
+      uint8* dst_u, uint8* dst_v, int width) = ARGB1555ToUVRow_C;
+  void (*ARGB1555ToYRow)(const uint8* src_argb1555, uint8* dst_y, int pix) =
+      ARGB1555ToYRow_C;
+#else
+  void (*ARGB1555ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      ARGB1555ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#endif
+  if (!src_argb1555 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb1555 = src_argb1555 + (height - 1) * src_stride_argb1555;
+    src_stride_argb1555 = -src_stride_argb1555;
+  }
+
+// Neon version does direct ARGB1555 to YUV.
+#if defined(HAS_ARGB1555TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGB1555ToUVRow = ARGB1555ToUVRow_Any_NEON;
+    ARGB1555ToYRow = ARGB1555ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToYRow = ARGB1555ToYRow_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGB1555ToUVRow = ARGB1555ToUVRow_NEON;
+      }
+    }
+  }
+// Other platforms do intermediate conversion from ARGB1555 to ARGB.
+#else
+#if defined(HAS_ARGB1555TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_ARGB1555TOYROW_NEON)
+      ARGB1555ToUVRow(src_argb1555, src_stride_argb1555, dst_u, dst_v, width);
+      ARGB1555ToYRow(src_argb1555, dst_y, width);
+      ARGB1555ToYRow(src_argb1555 + src_stride_argb1555, dst_y + dst_stride_y,
+                     width);
+#else
+      ARGB1555ToARGBRow(src_argb1555, row, width);
+      ARGB1555ToARGBRow(src_argb1555 + src_stride_argb1555, row + kRowSize,
+                        width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_argb1555 += src_stride_argb1555 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_ARGB1555TOYROW_NEON)
+      ARGB1555ToUVRow(src_argb1555, 0, dst_u, dst_v, width);
+      ARGB1555ToYRow(src_argb1555, dst_y, width);
+#else
+      ARGB1555ToARGBRow(src_argb1555, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_ARGB1555TOYROW_NEON)
+    free_aligned_buffer_64(row);
+  }
+#endif
+  return 0;
+}
+
+// Convert ARGB4444 to I420.
+LIBYUV_API
+int ARGB4444ToI420(const uint8* src_argb4444, int src_stride_argb4444,
+                   uint8* dst_y, int dst_stride_y,
+                   uint8* dst_u, int dst_stride_u,
+                   uint8* dst_v, int dst_stride_v,
+                   int width, int height) {
+  int y;
+#if defined(HAS_ARGB4444TOYROW_NEON)
+  void (*ARGB4444ToUVRow)(const uint8* src_argb4444, int src_stride_argb4444,
+      uint8* dst_u, uint8* dst_v, int width) = ARGB4444ToUVRow_C;
+  void (*ARGB4444ToYRow)(const uint8* src_argb4444, uint8* dst_y, int pix) =
+      ARGB4444ToYRow_C;
+#else
+  void (*ARGB4444ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      ARGB4444ToARGBRow_C;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+#endif
+  if (!src_argb4444 || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb4444 = src_argb4444 + (height - 1) * src_stride_argb4444;
+    src_stride_argb4444 = -src_stride_argb4444;
+  }
+
+// Neon version does direct ARGB4444 to YUV.
+#if defined(HAS_ARGB4444TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGB4444ToUVRow = ARGB4444ToUVRow_Any_NEON;
+    ARGB4444ToYRow = ARGB4444ToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToYRow = ARGB4444ToYRow_NEON;
+      if (IS_ALIGNED(width, 16)) {
+        ARGB4444ToUVRow = ARGB4444ToUVRow_NEON;
+      }
+    }
+  }
+// Other platforms do intermediate conversion from ARGB4444 to ARGB.
+#else
+#if defined(HAS_ARGB4444TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+  {
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+#endif
+
+    for (y = 0; y < height - 1; y += 2) {
+#if defined(HAS_ARGB4444TOYROW_NEON)
+      ARGB4444ToUVRow(src_argb4444, src_stride_argb4444, dst_u, dst_v, width);
+      ARGB4444ToYRow(src_argb4444, dst_y, width);
+      ARGB4444ToYRow(src_argb4444 + src_stride_argb4444, dst_y + dst_stride_y,
+                     width);
+#else
+      ARGB4444ToARGBRow(src_argb4444, row, width);
+      ARGB4444ToARGBRow(src_argb4444 + src_stride_argb4444, row + kRowSize,
+                        width);
+      ARGBToUVRow(row, kRowSize, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+      ARGBToYRow(row + kRowSize, dst_y + dst_stride_y, width);
+#endif
+      src_argb4444 += src_stride_argb4444 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_u += dst_stride_u;
+      dst_v += dst_stride_v;
+    }
+    if (height & 1) {
+#if defined(HAS_ARGB4444TOYROW_NEON)
+      ARGB4444ToUVRow(src_argb4444, 0, dst_u, dst_v, width);
+      ARGB4444ToYRow(src_argb4444, dst_y, width);
+#else
+      ARGB4444ToARGBRow(src_argb4444, row, width);
+      ARGBToUVRow(row, 0, dst_u, dst_v, width);
+      ARGBToYRow(row, dst_y, width);
+#endif
+    }
+#if !defined(HAS_ARGB4444TOYROW_NEON)
+    free_aligned_buffer_64(row);
+  }
+#endif
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/convert_argb.cc b/third_party/aom/third_party/libyuv/source/convert_argb.cc
new file mode 100644
index 0000000000..44756bc41c
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/convert_argb.cc
@@ -0,0 +1,1155 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_argb.h"
+
+#include "libyuv/cpu_id.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/rotate_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy ARGB with optional flipping
+LIBYUV_API
+int ARGBCopy(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int width, int height) {
+  if (!src_argb || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+  CopyPlane(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+            width * 4, height);
+  return 0;
+}
+
+// Convert I444 to ARGB.
+LIBYUV_API
+int I444ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I444ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I444ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u == width &&
+      src_stride_v == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_I444TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I444ToARGBRow = I444ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I444ToARGBRow = I444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToARGBRow = I444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I444ToARGBRow = I444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I444ToARGBRow = I444ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I444ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to ARGB.
+LIBYUV_API
+int I422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I411 to ARGB.
+LIBYUV_API
+int I411ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I411ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I411ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 4 == width &&
+      src_stride_v * 4 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_I411TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I411ToARGBRow = I411ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I411ToARGBRow = I411ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I411TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I411ToARGBRow = I411ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I411ToARGBRow = I411ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I411TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I411ToARGBRow = I411ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I411ToARGBRow = I411ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I411ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I400 to ARGB.
+LIBYUV_API
+int I400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I400ToARGBRow)(const uint8* y_buf,
+                     uint8* rgb_buf,
+                     int width) = I400ToARGBRow_C;
+  if (!src_y || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_I400TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I400ToARGBRow = I400ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      I400ToARGBRow = I400ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I400TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I400ToARGBRow = I400ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I400ToARGBRow = I400ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I400TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I400ToARGBRow = I400ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I400ToARGBRow = I400ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I400ToARGBRow(src_y, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+  }
+  return 0;
+}
+
+// Convert J400 to ARGB.
+LIBYUV_API
+int J400ToARGB(const uint8* src_y, int src_stride_y,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*J400ToARGBRow)(const uint8* src_y, uint8* dst_argb, int pix) =
+      J400ToARGBRow_C;
+  if (!src_y || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_J400TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    J400ToARGBRow = J400ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      J400ToARGBRow = J400ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_J400TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    J400ToARGBRow = J400ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      J400ToARGBRow = J400ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_J400TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    J400ToARGBRow = J400ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      J400ToARGBRow = J400ToARGBRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    J400ToARGBRow(src_y, dst_argb, width);
+    src_y += src_stride_y;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Shuffle table for converting BGRA to ARGB.
+static uvec8 kShuffleMaskBGRAToARGB = {
+  3u, 2u, 1u, 0u, 7u, 6u, 5u, 4u, 11u, 10u, 9u, 8u, 15u, 14u, 13u, 12u
+};
+
+// Shuffle table for converting ABGR to ARGB.
+static uvec8 kShuffleMaskABGRToARGB = {
+  2u, 1u, 0u, 3u, 6u, 5u, 4u, 7u, 10u, 9u, 8u, 11u, 14u, 13u, 12u, 15u
+};
+
+// Shuffle table for converting RGBA to ARGB.
+static uvec8 kShuffleMaskRGBAToARGB = {
+  1u, 2u, 3u, 0u, 5u, 6u, 7u, 4u, 9u, 10u, 11u, 8u, 13u, 14u, 15u, 12u
+};
+
+// Convert BGRA to ARGB.
+LIBYUV_API
+int BGRAToARGB(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_bgra, src_stride_bgra,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskBGRAToARGB),
+                     width, height);
+}
+
+// Convert ARGB to BGRA (same as BGRAToARGB).
+LIBYUV_API
+int ARGBToBGRA(const uint8* src_bgra, int src_stride_bgra,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_bgra, src_stride_bgra,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskBGRAToARGB),
+                     width, height);
+}
+
+// Convert ABGR to ARGB.
+LIBYUV_API
+int ABGRToARGB(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_abgr, src_stride_abgr,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskABGRToARGB),
+                     width, height);
+}
+
+// Convert ARGB to ABGR to (same as ABGRToARGB).
+LIBYUV_API
+int ARGBToABGR(const uint8* src_abgr, int src_stride_abgr,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_abgr, src_stride_abgr,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskABGRToARGB),
+                     width, height);
+}
+
+// Convert RGBA to ARGB.
+LIBYUV_API
+int RGBAToARGB(const uint8* src_rgba, int src_stride_rgba,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  return ARGBShuffle(src_rgba, src_stride_rgba,
+                     dst_argb, dst_stride_argb,
+                     (const uint8*)(&kShuffleMaskRGBAToARGB),
+                     width, height);
+}
+
+// Convert RGB24 to ARGB.
+LIBYUV_API
+int RGB24ToARGB(const uint8* src_rgb24, int src_stride_rgb24,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height) {
+  int y;
+  void (*RGB24ToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RGB24ToARGBRow_C;
+  if (!src_rgb24 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb24 = src_rgb24 + (height - 1) * src_stride_rgb24;
+    src_stride_rgb24 = -src_stride_rgb24;
+  }
+  // Coalesce rows.
+  if (src_stride_rgb24 == width * 3 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_rgb24 = dst_stride_argb = 0;
+  }
+#if defined(HAS_RGB24TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RGB24TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RGB24ToARGBRow(src_rgb24, dst_argb, width);
+    src_rgb24 += src_stride_rgb24;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert RAW to ARGB.
+LIBYUV_API
+int RAWToARGB(const uint8* src_raw, int src_stride_raw,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  int y;
+  void (*RAWToARGBRow)(const uint8* src_rgb, uint8* dst_argb, int pix) =
+      RAWToARGBRow_C;
+  if (!src_raw || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_raw = src_raw + (height - 1) * src_stride_raw;
+    src_stride_raw = -src_stride_raw;
+  }
+  // Coalesce rows.
+  if (src_stride_raw == width * 3 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_raw = dst_stride_argb = 0;
+  }
+#if defined(HAS_RAWTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    RAWToARGBRow = RAWToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RAWToARGBRow = RAWToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RAWToARGBRow = RAWToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RAWToARGBRow(src_raw, dst_argb, width);
+    src_raw += src_stride_raw;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert RGB565 to ARGB.
+LIBYUV_API
+int RGB565ToARGB(const uint8* src_rgb565, int src_stride_rgb565,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*RGB565ToARGBRow)(const uint8* src_rgb565, uint8* dst_argb, int pix) =
+      RGB565ToARGBRow_C;
+  if (!src_rgb565 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_rgb565 = src_rgb565 + (height - 1) * src_stride_rgb565;
+    src_stride_rgb565 = -src_stride_rgb565;
+  }
+  // Coalesce rows.
+  if (src_stride_rgb565 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_rgb565 = dst_stride_argb = 0;
+  }
+#if defined(HAS_RGB565TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_RGB565TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    RGB565ToARGBRow(src_rgb565, dst_argb, width);
+    src_rgb565 += src_stride_rgb565;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB1555 to ARGB.
+LIBYUV_API
+int ARGB1555ToARGB(const uint8* src_argb1555, int src_stride_argb1555,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height) {
+  int y;
+  void (*ARGB1555ToARGBRow)(const uint8* src_argb1555, uint8* dst_argb,
+      int pix) = ARGB1555ToARGBRow_C;
+  if (!src_argb1555 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb1555 = src_argb1555 + (height - 1) * src_stride_argb1555;
+    src_stride_argb1555 = -src_stride_argb1555;
+  }
+  // Coalesce rows.
+  if (src_stride_argb1555 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb1555 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGB1555TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGB1555ToARGBRow(src_argb1555, dst_argb, width);
+    src_argb1555 += src_stride_argb1555;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB4444 to ARGB.
+LIBYUV_API
+int ARGB4444ToARGB(const uint8* src_argb4444, int src_stride_argb4444,
+                   uint8* dst_argb, int dst_stride_argb,
+                   int width, int height) {
+  int y;
+  void (*ARGB4444ToARGBRow)(const uint8* src_argb4444, uint8* dst_argb,
+      int pix) = ARGB4444ToARGBRow_C;
+  if (!src_argb4444 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb4444 = src_argb4444 + (height - 1) * src_stride_argb4444;
+    src_stride_argb4444 = -src_stride_argb4444;
+  }
+  // Coalesce rows.
+  if (src_stride_argb4444 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb4444 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGB4444TOARGBROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_SSE2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGB4444ToARGBRow(src_argb4444, dst_argb, width);
+    src_argb4444 += src_stride_argb4444;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert NV12 to ARGB.
+LIBYUV_API
+int NV12ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*NV12ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV12ToARGBRow_C;
+  if (!src_y || !src_uv || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_NV12TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToARGBRow = NV12ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToARGBRow(src_y, src_uv, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV21 to ARGB.
+LIBYUV_API
+int NV21ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_uv, int src_stride_uv,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*NV21ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV21ToARGBRow_C;
+  if (!src_y || !src_uv || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_NV21TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV21TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToARGBRow = NV21ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV21TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV21ToARGBRow(src_y, src_uv, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert M420 to ARGB.
+LIBYUV_API
+int M420ToARGB(const uint8* src_m420, int src_stride_m420,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*NV12ToARGBRow)(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* rgb_buf,
+                        int width) = NV12ToARGBRow_C;
+  if (!src_m420 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_NV12TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToARGBRow = NV12ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb, width);
+    NV12ToARGBRow(src_m420 + src_stride_m420, src_m420 + src_stride_m420 * 2,
+                  dst_argb + dst_stride_argb, width);
+    dst_argb += dst_stride_argb * 2;
+    src_m420 += src_stride_m420 * 3;
+  }
+  if (height & 1) {
+    NV12ToARGBRow(src_m420, src_m420 + src_stride_m420 * 2, dst_argb, width);
+  }
+  return 0;
+}
+
+// Convert YUY2 to ARGB.
+LIBYUV_API
+int YUY2ToARGB(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*YUY2ToARGBRow)(const uint8* src_yuy2, uint8* dst_argb, int pix) =
+      YUY2ToARGBRow_C;
+  if (!src_yuy2 || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_yuy2 == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_yuy2 = dst_stride_argb = 0;
+  }
+#if defined(HAS_YUY2TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    YUY2ToARGBRow(src_yuy2, dst_argb, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert UYVY to ARGB.
+LIBYUV_API
+int UYVYToARGB(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*UYVYToARGBRow)(const uint8* src_uyvy, uint8* dst_argb, int pix) =
+      UYVYToARGBRow_C;
+  if (!src_uyvy || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_argb = 0;
+  }
+#if defined(HAS_UYVYTOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToARGBRow = UYVYToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToARGBRow = UYVYToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      UYVYToARGBRow = UYVYToARGBRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    UYVYToARGBRow(src_uyvy, dst_argb, width);
+    src_uyvy += src_stride_uyvy;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert J420 to ARGB.
+LIBYUV_API
+int J420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*J422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = J422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_J422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    J422ToARGBRow = J422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    J422ToARGBRow = J422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      J422ToARGBRow = J422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    J422ToARGBRow = J422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    J422ToARGBRow = J422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    J422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert J422 to ARGB.
+LIBYUV_API
+int J422ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*J422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = J422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_argb = 0;
+  }
+#if defined(HAS_J422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    J422ToARGBRow = J422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    J422ToARGBRow = J422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      J422ToARGBRow = J422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    J422ToARGBRow = J422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      J422ToARGBRow = J422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_J422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    J422ToARGBRow = J422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    J422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/convert_from.cc b/third_party/aom/third_party/libyuv/source/convert_from.cc
new file mode 100644
index 0000000000..31f1ac992a
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/convert_from.cc
@@ -0,0 +1,1348 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_from.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/convert.h"  // For I420Copy
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate.h"
+#include "libyuv/scale.h"  // For ScalePlane()
+#include "libyuv/video_common.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// I420 To any I4xx YUV format with mirroring.
+static int I420ToI4xx(const uint8* src_y, int src_stride_y,
+                      const uint8* src_u, int src_stride_u,
+                      const uint8* src_v, int src_stride_v,
+                      uint8* dst_y, int dst_stride_y,
+                      uint8* dst_u, int dst_stride_u,
+                      uint8* dst_v, int dst_stride_v,
+                      int src_y_width, int src_y_height,
+                      int dst_uv_width, int dst_uv_height) {
+  const int dst_y_width = Abs(src_y_width);
+  const int dst_y_height = Abs(src_y_height);
+  const int src_uv_width = SUBSAMPLE(src_y_width, 1, 1);
+  const int src_uv_height = SUBSAMPLE(src_y_height, 1, 1);
+  if (src_y_width == 0 || src_y_height == 0 ||
+      dst_uv_width <= 0 || dst_uv_height <= 0) {
+    return -1;
+  }
+  ScalePlane(src_y, src_stride_y, src_y_width, src_y_height,
+             dst_y, dst_stride_y, dst_y_width, dst_y_height,
+             kFilterBilinear);
+  ScalePlane(src_u, src_stride_u, src_uv_width, src_uv_height,
+             dst_u, dst_stride_u, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  ScalePlane(src_v, src_stride_v, src_uv_width, src_uv_height,
+             dst_v, dst_stride_v, dst_uv_width, dst_uv_height,
+             kFilterBilinear);
+  return 0;
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 422 chroma is 1/2 width, 1x height
+LIBYUV_API
+int I420ToI422(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = (Abs(width) + 1) >> 1;
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 444 chroma is 1x width, 1x height
+LIBYUV_API
+int I420ToI444(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = Abs(width);
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// 420 chroma is 1/2 width, 1/2 height
+// 411 chroma is 1/4 width, 1x height
+LIBYUV_API
+int I420ToI411(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  const int dst_uv_width = (Abs(width) + 3) >> 2;
+  const int dst_uv_height = Abs(height);
+  return I420ToI4xx(src_y, src_stride_y,
+                    src_u, src_stride_u,
+                    src_v, src_stride_v,
+                    dst_y, dst_stride_y,
+                    dst_u, dst_stride_u,
+                    dst_v, dst_stride_v,
+                    width, height,
+                    dst_uv_width, dst_uv_height);
+}
+
+// Copy to I400. Source can be I420,422,444,400,NV12,NV21
+LIBYUV_API
+int I400Copy(const uint8* src_y, int src_stride_y,
+             uint8* dst_y, int dst_stride_y,
+             int width, int height) {
+  if (!src_y || !dst_y ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+LIBYUV_API
+int I422ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  int y;
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_yuy2, int width) =
+      I422ToYUY2Row_C;
+  if (!src_y || !src_u || !src_v || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_yuy2 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_yuy2 = 0;
+  }
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_yuy2 += dst_stride_yuy2;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToYUY2(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  int y;
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_yuy2, int width) =
+      I422ToYUY2Row_C;
+  if (!src_y || !src_u || !src_v || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+    I422ToYUY2Row(src_y + src_stride_y, src_u, src_v,
+                  dst_yuy2 + dst_stride_yuy2, width);
+    src_y += src_stride_y * 2;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_yuy2 += dst_stride_yuy2 * 2;
+  }
+  if (height & 1) {
+    I422ToYUY2Row(src_y, src_u, src_v, dst_yuy2, width);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I422ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  int y;
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_uyvy, int width) =
+      I422ToUYVYRow_C;
+  if (!src_y || !src_u || !src_v || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_uyvy == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_uyvy = 0;
+  }
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uyvy += dst_stride_uyvy;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToUYVY(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  int y;
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+                        const uint8* src_v, uint8* dst_uyvy, int width) =
+      I422ToUYVYRow_C;
+  if (!src_y || !src_u || !src_v || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+    I422ToUYVYRow(src_y + src_stride_y, src_u, src_v,
+                  dst_uyvy + dst_stride_uyvy, width);
+    src_y += src_stride_y * 2;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uyvy += dst_stride_uyvy * 2;
+  }
+  if (height & 1) {
+    I422ToUYVYRow(src_y, src_u, src_v, dst_uyvy, width);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToNV12(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+      int width) = MergeUVRow_C;
+  // Coalesce rows.
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_uv = dst_uv + (halfheight - 1) * dst_stride_uv;
+    dst_stride_y = -dst_stride_y;
+    dst_stride_uv = -dst_stride_uv;
+  }
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+  // Coalesce rows.
+  if (src_stride_u == halfwidth &&
+      src_stride_v == halfwidth &&
+      dst_stride_uv == halfwidth * 2) {
+    halfwidth *= halfheight;
+    halfheight = 1;
+    src_stride_u = src_stride_v = dst_stride_uv = 0;
+  }
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  for (y = 0; y < halfheight; ++y) {
+    // Merge a row of U and V into a row of UV.
+    MergeUVRow_(src_u, src_v, dst_uv, halfwidth);
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+    dst_uv += dst_stride_uv;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int I420ToNV21(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_vu, int dst_stride_vu,
+               int width, int height) {
+  return I420ToNV12(src_y, src_stride_y,
+                    src_v, src_stride_v,
+                    src_u, src_stride_u,
+                    dst_y, src_stride_y,
+                    dst_vu, dst_stride_vu,
+                    width, height);
+}
+
+// Convert I420 to ARGB.
+LIBYUV_API
+int I420ToARGB(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+  if (!src_y || !src_u || !src_v || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGBRow(src_y, src_u, src_v, dst_argb, width);
+    dst_argb += dst_stride_argb;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to BGRA.
+LIBYUV_API
+int I420ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height) {
+  int y;
+  void (*I422ToBGRARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToBGRARow_C;
+  if (!src_y || !src_u || !src_v || !dst_bgra ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra;
+    dst_stride_bgra = -dst_stride_bgra;
+  }
+#if defined(HAS_I422TOBGRAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToBGRARow = I422ToBGRARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToBGRARow = I422ToBGRARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) {
+    I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width);
+    dst_bgra += dst_stride_bgra;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to ABGR.
+LIBYUV_API
+int I420ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height) {
+  int y;
+  void (*I422ToABGRRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToABGRRow_C;
+  if (!src_y || !src_u || !src_v || !dst_abgr ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr;
+    dst_stride_abgr = -dst_stride_abgr;
+  }
+#if defined(HAS_I422TOABGRROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOABGRROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToABGRRow = I422ToABGRRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOABGRROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToABGRRow = I422ToABGRRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width);
+    dst_abgr += dst_stride_abgr;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGBA.
+LIBYUV_API
+int I420ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRGBARow_C;
+  if (!src_y || !src_u || !src_v || !dst_rgba ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB24.
+LIBYUV_API
+int I420ToRGB24(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height) {
+  int y;
+  void (*I422ToRGB24Row)(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* rgb_buf,
+                         int width) = I422ToRGB24Row_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb24 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb24 = dst_rgb24 + (height - 1) * dst_stride_rgb24;
+    dst_stride_rgb24 = -dst_stride_rgb24;
+  }
+#if defined(HAS_I422TORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB24Row = I422ToRGB24Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB24Row = I422ToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB24Row = I422ToRGB24Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB24Row(src_y, src_u, src_v, dst_rgb24, width);
+    dst_rgb24 += dst_stride_rgb24;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RAW.
+LIBYUV_API
+int I420ToRAW(const uint8* src_y, int src_stride_y,
+                const uint8* src_u, int src_stride_u,
+                const uint8* src_v, int src_stride_v,
+                uint8* dst_raw, int dst_stride_raw,
+                int width, int height) {
+  int y;
+  void (*I422ToRAWRow)(const uint8* y_buf,
+                       const uint8* u_buf,
+                       const uint8* v_buf,
+                       uint8* rgb_buf,
+                       int width) = I422ToRAWRow_C;
+  if (!src_y || !src_u || !src_v || !dst_raw ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_raw = dst_raw + (height - 1) * dst_stride_raw;
+    dst_stride_raw = -dst_stride_raw;
+  }
+#if defined(HAS_I422TORAWROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRAWRow = I422ToRAWRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRAWRow = I422ToRAWRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORAWROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRAWRow = I422ToRAWRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRAWRow = I422ToRAWRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORAWROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRAWRow = I422ToRAWRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRAWRow = I422ToRAWRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRAWRow(src_y, src_u, src_v, dst_raw, width);
+    dst_raw += dst_stride_raw;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to ARGB1555.
+LIBYUV_API
+int I420ToARGB1555(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height) {
+  int y;
+  void (*I422ToARGB1555Row)(const uint8* y_buf,
+                            const uint8* u_buf,
+                            const uint8* v_buf,
+                            uint8* rgb_buf,
+                            int width) = I422ToARGB1555Row_C;
+  if (!src_y || !src_u || !src_v || !dst_argb1555 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb1555 = dst_argb1555 + (height - 1) * dst_stride_argb1555;
+    dst_stride_argb1555 = -dst_stride_argb1555;
+  }
+#if defined(HAS_I422TOARGB1555ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGB1555Row = I422ToARGB1555Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB1555Row = I422ToARGB1555Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGB1555Row(src_y, src_u, src_v, dst_argb1555, width);
+    dst_argb1555 += dst_stride_argb1555;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+
+// Convert I420 to ARGB4444.
+LIBYUV_API
+int I420ToARGB4444(const uint8* src_y, int src_stride_y,
+                   const uint8* src_u, int src_stride_u,
+                   const uint8* src_v, int src_stride_v,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height) {
+  int y;
+  void (*I422ToARGB4444Row)(const uint8* y_buf,
+                            const uint8* u_buf,
+                            const uint8* v_buf,
+                            uint8* rgb_buf,
+                            int width) = I422ToARGB4444Row_C;
+  if (!src_y || !src_u || !src_v || !dst_argb4444 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb4444 = dst_argb4444 + (height - 1) * dst_stride_argb4444;
+    dst_stride_argb4444 = -dst_stride_argb4444;
+  }
+#if defined(HAS_I422TOARGB4444ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGB4444Row = I422ToARGB4444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGB4444Row = I422ToARGB4444Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToARGB4444Row(src_y, src_u, src_v, dst_argb4444, width);
+    dst_argb4444 += dst_stride_argb4444;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Convert I420 to RGB565.
+LIBYUV_API
+int I420ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_u, int src_stride_u,
+                 const uint8* src_v, int src_stride_v,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*I422ToRGB565Row)(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* rgb_buf,
+                          int width) = I422ToRGB565Row_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_I422TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGB565Row = I422ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToRGB565Row = I422ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGB565Row = I422ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGB565Row(src_y, src_u, src_v, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_u += src_stride_u;
+      src_v += src_stride_v;
+    }
+  }
+  return 0;
+}
+
+// Ordered 8x8 dither for 888 to 565.  Values from 0 to 7.
+static const uint8 kDither565_4x4[16] = {
+  0, 4, 1, 5,
+  6, 2, 7, 3,
+  1, 5, 0, 4,
+  7, 3, 6, 2,
+};
+
+// Convert I420 to RGB565 with dithering.
+LIBYUV_API
+int I420ToRGB565Dither(const uint8* src_y, int src_stride_y,
+                       const uint8* src_u, int src_stride_u,
+                       const uint8* src_v, int src_stride_v,
+                       uint8* dst_rgb565, int dst_stride_rgb565,
+                       const uint8* dither4x4, int width, int height) {
+  int y;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+  void (*ARGBToRGB565DitherRow)(const uint8* src_argb, uint8* dst_rgb,
+      const uint32 dither4, int pix) = ARGBToRGB565DitherRow_C;
+  if (!src_y || !src_u || !src_v || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+  if (!dither4x4) {
+    dither4x4 = kDither565_4x4;
+  }
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_NEON;
+    }
+  }
+#endif
+  {
+    // Allocate a row of argb.
+    align_buffer_64(row_argb, width * 4);
+    for (y = 0; y < height; ++y) {
+      I422ToARGBRow(src_y, src_u, src_v, row_argb, width);
+      ARGBToRGB565DitherRow(row_argb, dst_rgb565,
+                            *(uint32*)(dither4x4 + ((y & 3) << 2)), width);
+      dst_rgb565 += dst_stride_rgb565;
+      src_y += src_stride_y;
+      if (y & 1) {
+        src_u += src_stride_u;
+        src_v += src_stride_v;
+      }
+    }
+    free_aligned_buffer_64(row_argb);
+  }
+  return 0;
+}
+
+// Convert I420 to specified format
+LIBYUV_API
+int ConvertFromI420(const uint8* y, int y_stride,
+                    const uint8* u, int u_stride,
+                    const uint8* v, int v_stride,
+                    uint8* dst_sample, int dst_sample_stride,
+                    int width, int height,
+                    uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int r = 0;
+  if (!y || !u|| !v || !dst_sample ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      r = I420ToYUY2(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 2,
+                     width, height);
+      break;
+    case FOURCC_UYVY:
+      r = I420ToUYVY(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 2,
+                     width, height);
+      break;
+    case FOURCC_RGBP:
+      r = I420ToRGB565(y, y_stride,
+                       u, u_stride,
+                       v, v_stride,
+                       dst_sample,
+                       dst_sample_stride ? dst_sample_stride : width * 2,
+                       width, height);
+      break;
+    case FOURCC_RGBO:
+      r = I420ToARGB1555(y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         dst_sample,
+                         dst_sample_stride ? dst_sample_stride : width * 2,
+                         width, height);
+      break;
+    case FOURCC_R444:
+      r = I420ToARGB4444(y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         dst_sample,
+                         dst_sample_stride ? dst_sample_stride : width * 2,
+                         width, height);
+      break;
+    case FOURCC_24BG:
+      r = I420ToRGB24(y, y_stride,
+                      u, u_stride,
+                      v, v_stride,
+                      dst_sample,
+                      dst_sample_stride ? dst_sample_stride : width * 3,
+                      width, height);
+      break;
+    case FOURCC_RAW:
+      r = I420ToRAW(y, y_stride,
+                    u, u_stride,
+                    v, v_stride,
+                    dst_sample,
+                    dst_sample_stride ? dst_sample_stride : width * 3,
+                    width, height);
+      break;
+    case FOURCC_ARGB:
+      r = I420ToARGB(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_BGRA:
+      r = I420ToBGRA(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_ABGR:
+      r = I420ToABGR(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_RGBA:
+      r = I420ToRGBA(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width * 4,
+                     width, height);
+      break;
+    case FOURCC_I400:
+      r = I400Copy(y, y_stride,
+                   dst_sample,
+                   dst_sample_stride ? dst_sample_stride : width,
+                   width, height);
+      break;
+    case FOURCC_NV12: {
+      uint8* dst_uv = dst_sample + width * height;
+      r = I420ToNV12(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     dst_uv,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     width, height);
+      break;
+    }
+    case FOURCC_NV21: {
+      uint8* dst_vu = dst_sample + width * height;
+      r = I420ToNV21(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     dst_vu,
+                     dst_sample_stride ? dst_sample_stride : width,
+                     width, height);
+      break;
+    }
+    // TODO(fbarchard): Add M420.
+    // Triplanar formats
+    // TODO(fbarchard): halfstride instead of halfwidth
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      int halfwidth = (width + 1) / 2;
+      int halfheight = (height + 1) / 2;
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV12) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + halfwidth * halfheight;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + halfwidth * halfheight;
+      }
+      r = I420Copy(y, y_stride,
+                   u, u_stride,
+                   v, v_stride,
+                   dst_sample, width,
+                   dst_u, halfwidth,
+                   dst_v, halfwidth,
+                   width, height);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      int halfwidth = (width + 1) / 2;
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV16) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + halfwidth * height;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + halfwidth * height;
+      }
+      r = I420ToI422(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, halfwidth,
+                     dst_v, halfwidth,
+                     width, height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      uint8* dst_u;
+      uint8* dst_v;
+      if (format == FOURCC_YV24) {
+        dst_v = dst_sample + width * height;
+        dst_u = dst_v + width * height;
+      } else {
+        dst_u = dst_sample + width * height;
+        dst_v = dst_u + width * height;
+      }
+      r = I420ToI444(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, width,
+                     dst_v, width,
+                     width, height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (width + 3) / 4;
+      uint8* dst_u = dst_sample + width * height;
+      uint8* dst_v = dst_u + quarterwidth * height;
+      r = I420ToI411(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     dst_sample, width,
+                     dst_u, quarterwidth,
+                     dst_v, quarterwidth,
+                     width, height);
+      break;
+    }
+
+    // Formats not supported - MJPG, biplanar, some rgb formats.
+    default:
+      return -1;  // unknown fourcc - return failure code.
+  }
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/convert_from_argb.cc b/third_party/aom/third_party/libyuv/source/convert_from_argb.cc
new file mode 100644
index 0000000000..8d1e97aec2
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/convert_from_argb.cc
@@ -0,0 +1,1301 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_from_argb.h"
+
+#include "libyuv/basic_types.h"
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// ARGB little endian (bgra in memory) to I444
+LIBYUV_API
+int ARGBToI444(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*ARGBToUV444Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV444Row_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u == width &&
+      dst_stride_v == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOUV444ROW_SSSE3)
+    if (TestCpuFlag(kCpuHasSSSE3)) {
+      ARGBToUV444Row = ARGBToUV444Row_Any_SSSE3;
+      if (IS_ALIGNED(width, 16)) {
+        ARGBToUV444Row = ARGBToUV444Row_SSSE3;
+      }
+  }
+#endif
+#if defined(HAS_ARGBTOUV444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUV444Row = ARGBToUV444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToUV444Row = ARGBToUV444Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUV444Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// ARGB little endian (bgra in memory) to I422
+LIBYUV_API
+int ARGBToI422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV422Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUV422Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// ARGB little endian (bgra in memory) to I411
+LIBYUV_API
+int ARGBToI411(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV411Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV411Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 4 == width &&
+      dst_stride_v * 4 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUV411ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUV411Row = ARGBToUV411Row_Any_NEON;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUV411Row = ARGBToUV411Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUV411Row(src_argb, dst_u, dst_v, width);
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                      int width) = MergeUVRow_C;
+  if (!src_argb ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+  {
+    // Allocate a rows of uv.
+    align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2);
+    uint8* row_v = row_u + ((halfwidth + 31) & ~31);
+
+    for (y = 0; y < height - 1; y += 2) {
+      ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
+      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      ARGBToYRow(src_argb, dst_y, width);
+      ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+      src_argb += src_stride_argb * 2;
+      dst_y += dst_stride_y * 2;
+      dst_uv += dst_stride_uv;
+    }
+    if (height & 1) {
+      ARGBToUVRow(src_argb, 0, row_u, row_v, width);
+      MergeUVRow_(row_u, row_v, dst_uv, halfwidth);
+      ARGBToYRow(src_argb, dst_y, width);
+    }
+    free_aligned_buffer_64(row_u);
+  }
+  return 0;
+}
+
+// Same as NV12 but U and V swapped.
+LIBYUV_API
+int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*ARGBToUVRow)(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) = ARGBToUVRow_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*MergeUVRow_)(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                      int width) = MergeUVRow_C;
+  if (!src_argb ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3) && defined(HAS_ARGBTOUVROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVRow = ARGBToUVRow_Any_SSSE3;
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_SSSE3;
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2) && defined(HAS_ARGBTOUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToUVRow = ARGBToUVRow_Any_AVX2;
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToUVRow = ARGBToUVRow_AVX2;
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVRow = ARGBToUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVRow = ARGBToUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MergeUVRow_ = MergeUVRow_Any_SSE2;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MergeUVRow_ = MergeUVRow_Any_AVX2;
+    if (IS_ALIGNED(halfwidth, 32)) {
+      MergeUVRow_ = MergeUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_MERGEUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MergeUVRow_ = MergeUVRow_Any_NEON;
+    if (IS_ALIGNED(halfwidth, 16)) {
+      MergeUVRow_ = MergeUVRow_NEON;
+    }
+  }
+#endif
+  {
+    // Allocate a rows of uv.
+    align_buffer_64(row_u, ((halfwidth + 31) & ~31) * 2);
+    uint8* row_v = row_u + ((halfwidth + 31) & ~31);
+
+    for (y = 0; y < height - 1; y += 2) {
+      ARGBToUVRow(src_argb, src_stride_argb, row_u, row_v, width);
+      MergeUVRow_(row_v, row_u, dst_uv, halfwidth);
+      ARGBToYRow(src_argb, dst_y, width);
+      ARGBToYRow(src_argb + src_stride_argb, dst_y + dst_stride_y, width);
+      src_argb += src_stride_argb * 2;
+      dst_y += dst_stride_y * 2;
+      dst_uv += dst_stride_uv;
+    }
+    if (height & 1) {
+      ARGBToUVRow(src_argb, 0, row_u, row_v, width);
+      MergeUVRow_(row_v, row_u, dst_uv, halfwidth);
+      ARGBToYRow(src_argb, dst_y, width);
+    }
+    free_aligned_buffer_64(row_u);
+  }
+  return 0;
+}
+
+// Convert ARGB to YUY2.
+LIBYUV_API
+int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yuy2, int dst_stride_yuy2,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV422Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*I422ToYUY2Row)(const uint8* src_y, const uint8* src_u,
+      const uint8* src_v, uint8* dst_yuy2, int width) = I422ToYUY2Row_C;
+
+  if (!src_argb || !dst_yuy2 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_yuy2 = dst_yuy2 + (height - 1) * dst_stride_yuy2;
+    dst_stride_yuy2 = -dst_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_yuy2 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_yuy2 = 0;
+  }
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+#if defined(HAS_I422TOYUY2ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOYUY2ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToYUY2Row = I422ToYUY2Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToYUY2Row = I422ToYUY2Row_NEON;
+    }
+  }
+#endif
+
+  {
+    // Allocate a rows of yuv.
+    align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+    uint8* row_u = row_y + ((width + 63) & ~63);
+    uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+
+    for (y = 0; y < height; ++y) {
+      ARGBToUV422Row(src_argb, row_u, row_v, width);
+      ARGBToYRow(src_argb, row_y, width);
+      I422ToYUY2Row(row_y, row_u, row_v, dst_yuy2, width);
+      src_argb += src_stride_argb;
+      dst_yuy2 += dst_stride_yuy2;
+    }
+
+    free_aligned_buffer_64(row_y);
+  }
+  return 0;
+}
+
+// Convert ARGB to UYVY.
+LIBYUV_API
+int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_uyvy, int dst_stride_uyvy,
+               int width, int height) {
+  int y;
+  void (*ARGBToUV422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUV422Row_C;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  void (*I422ToUYVYRow)(const uint8* src_y, const uint8* src_u,
+      const uint8* src_v, uint8* dst_uyvy, int width) = I422ToUYVYRow_C;
+
+  if (!src_argb || !dst_uyvy ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_uyvy = dst_uyvy + (height - 1) * dst_stride_uyvy;
+    dst_stride_uyvy = -dst_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_uyvy == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_uyvy = 0;
+  }
+#if defined(HAS_ARGBTOUV422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+#if defined(HAS_I422TOUYVYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOUYVYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToUYVYRow = I422ToUYVYRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToUYVYRow = I422ToUYVYRow_NEON;
+    }
+  }
+#endif
+
+  {
+    // Allocate a rows of yuv.
+    align_buffer_64(row_y, ((width + 63) & ~63) * 2);
+    uint8* row_u = row_y + ((width + 63) & ~63);
+    uint8* row_v = row_u + ((width + 63) & ~63) / 2;
+
+    for (y = 0; y < height; ++y) {
+      ARGBToUV422Row(src_argb, row_u, row_v, width);
+      ARGBToYRow(src_argb, row_y, width);
+      I422ToUYVYRow(row_y, row_u, row_v, dst_uyvy, width);
+      src_argb += src_stride_argb;
+      dst_uyvy += dst_stride_uyvy;
+    }
+
+    free_aligned_buffer_64(row_y);
+  }
+  return 0;
+}
+
+// Convert ARGB to I400.
+LIBYUV_API
+int ARGBToI400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  int y;
+  void (*ARGBToYRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYRow_C;
+  if (!src_argb || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = 0;
+  }
+#if defined(HAS_ARGBTOYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYRow = ARGBToYRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYRow = ARGBToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYRow = ARGBToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYRow = ARGBToYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYRow = ARGBToYRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToYRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+  }
+  return 0;
+}
+
+// Shuffle table for converting ARGB to RGBA.
+static uvec8 kShuffleMaskARGBToRGBA = {
+  3u, 0u, 1u, 2u, 7u, 4u, 5u, 6u, 11u, 8u, 9u, 10u, 15u, 12u, 13u, 14u
+};
+
+// Convert ARGB to RGBA.
+LIBYUV_API
+int ARGBToRGBA(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  return ARGBShuffle(src_argb, src_stride_argb,
+                     dst_rgba, dst_stride_rgba,
+                     (const uint8*)(&kShuffleMaskARGBToRGBA),
+                     width, height);
+}
+
+// Convert ARGB To RGB24.
+LIBYUV_API
+int ARGBToRGB24(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_rgb24, int dst_stride_rgb24,
+                int width, int height) {
+  int y;
+  void (*ARGBToRGB24Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRGB24Row_C;
+  if (!src_argb || !dst_rgb24 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_rgb24 == width * 3) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_rgb24 = 0;
+  }
+#if defined(HAS_ARGBTORGB24ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB24ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRGB24Row = ARGBToRGB24Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB24Row = ARGBToRGB24Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToRGB24Row(src_argb, dst_rgb24, width);
+    src_argb += src_stride_argb;
+    dst_rgb24 += dst_stride_rgb24;
+  }
+  return 0;
+}
+
+// Convert ARGB To RAW.
+LIBYUV_API
+int ARGBToRAW(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_raw, int dst_stride_raw,
+              int width, int height) {
+  int y;
+  void (*ARGBToRAWRow)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRAWRow_C;
+  if (!src_argb || !dst_raw || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_raw == width * 3) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_raw = 0;
+  }
+#if defined(HAS_ARGBTORAWROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToRAWRow = ARGBToRAWRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORAWROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRAWRow = ARGBToRAWRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRAWRow = ARGBToRAWRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToRAWRow(src_argb, dst_raw, width);
+    src_argb += src_stride_argb;
+    dst_raw += dst_stride_raw;
+  }
+  return 0;
+}
+
+// Ordered 8x8 dither for 888 to 565.  Values from 0 to 7.
+static const uint8 kDither565_4x4[16] = {
+  0, 4, 1, 5,
+  6, 2, 7, 3,
+  1, 5, 0, 4,
+  7, 3, 6, 2,
+};
+
+// Convert ARGB To RGB565 with 4x4 dither matrix (16 bytes).
+LIBYUV_API
+int ARGBToRGB565Dither(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_rgb565, int dst_stride_rgb565,
+                       const uint8* dither4x4, int width, int height) {
+  int y;
+  void (*ARGBToRGB565DitherRow)(const uint8* src_argb, uint8* dst_rgb,
+      const uint32 dither4, int pix) = ARGBToRGB565DitherRow_C;
+  if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  if (!dither4x4) {
+    dither4x4 = kDither565_4x4;
+  }
+#if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565DitherRow = ARGBToRGB565DitherRow_NEON;
+    }
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBToRGB565DitherRow(src_argb, dst_rgb565,
+                          *(uint32*)(dither4x4 + ((y & 3) << 2)), width);
+    src_argb += src_stride_argb;
+    dst_rgb565 += dst_stride_rgb565;
+  }
+  return 0;
+}
+
+// Convert ARGB To RGB565.
+// TODO(fbarchard): Consider using dither function low level with zeros.
+LIBYUV_API
+int ARGBToRGB565(const uint8* src_argb, int src_stride_argb,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*ARGBToRGB565Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToRGB565Row_C;
+  if (!src_argb || !dst_rgb565 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_rgb565 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_rgb565 = 0;
+  }
+#if defined(HAS_ARGBTORGB565ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToRGB565Row(src_argb, dst_rgb565, width);
+    src_argb += src_stride_argb;
+    dst_rgb565 += dst_stride_rgb565;
+  }
+  return 0;
+}
+
+// Convert ARGB To ARGB1555.
+LIBYUV_API
+int ARGBToARGB1555(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb1555, int dst_stride_argb1555,
+                   int width, int height) {
+  int y;
+  void (*ARGBToARGB1555Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToARGB1555Row_C;
+  if (!src_argb || !dst_argb1555 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb1555 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb1555 = 0;
+  }
+#if defined(HAS_ARGBTOARGB1555ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB1555ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB1555ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToARGB1555Row(src_argb, dst_argb1555, width);
+    src_argb += src_stride_argb;
+    dst_argb1555 += dst_stride_argb1555;
+  }
+  return 0;
+}
+
+// Convert ARGB To ARGB4444.
+LIBYUV_API
+int ARGBToARGB4444(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb4444, int dst_stride_argb4444,
+                   int width, int height) {
+  int y;
+  void (*ARGBToARGB4444Row)(const uint8* src_argb, uint8* dst_rgb, int pix) =
+      ARGBToARGB4444Row_C;
+  if (!src_argb || !dst_argb4444 || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb4444 == width * 2) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb4444 = 0;
+  }
+#if defined(HAS_ARGBTOARGB4444ROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB4444ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOARGB4444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToARGB4444Row(src_argb, dst_argb4444, width);
+    src_argb += src_stride_argb;
+    dst_argb4444 += dst_stride_argb4444;
+  }
+  return 0;
+}
+
+// Convert ARGB to J420. (JPeg full range I420).
+LIBYUV_API
+int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUVJRow)(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) = ARGBToUVJRow_C;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int pix) =
+      ARGBToYJRow_C;
+  if (!src_argb ||
+      !dst_yj || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBTOYJROW_SSSE3) && defined(HAS_ARGBTOUVJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_SSSE3;
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_SSSE3;
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVJRow = ARGBToUVJRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJRow = ARGBToUVJRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height - 1; y += 2) {
+    ARGBToUVJRow(src_argb, src_stride_argb, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_yj, width);
+    ARGBToYJRow(src_argb + src_stride_argb, dst_yj + dst_stride_yj, width);
+    src_argb += src_stride_argb * 2;
+    dst_yj += dst_stride_yj * 2;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  if (height & 1) {
+    ARGBToUVJRow(src_argb, 0, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_yj, width);
+  }
+  return 0;
+}
+
+// ARGB little endian (bgra in memory) to J422
+LIBYUV_API
+int ARGBToJ422(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*ARGBToUVJ422Row)(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+      int pix) = ARGBToUVJ422Row_C;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_y, int pix) =
+      ARGBToYJRow_C;
+  if (!src_argb || !dst_y || !dst_u || !dst_v || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_ARGBTOUVJ422ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToUVJ422Row = ARGBToUVJ422Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJ422Row = ARGBToUVJ422Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOUVJ422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToUVJ422Row = ARGBToUVJ422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUVJ422Row = ARGBToUVJ422Row_NEON;
+    }
+  }
+#endif
+
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToUVJ422Row(src_argb, dst_u, dst_v, width);
+    ARGBToYJRow(src_argb, dst_y, width);
+    src_argb += src_stride_argb;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Convert ARGB to J400.
+LIBYUV_API
+int ARGBToJ400(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_yj, int dst_stride_yj,
+               int width, int height) {
+  int y;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_yj, int pix) =
+      ARGBToYJRow_C;
+  if (!src_argb || !dst_yj || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_yj == width) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_yj = 0;
+  }
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBToYJRow(src_argb, dst_yj, width);
+    src_argb += src_stride_argb;
+    dst_yj += dst_stride_yj;
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/convert_jpeg.cc b/third_party/aom/third_party/libyuv/source/convert_jpeg.cc
new file mode 100644
index 0000000000..bcb980f7f1
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/convert_jpeg.cc
@@ -0,0 +1,392 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert.h"
+
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#ifdef HAVE_JPEG
+struct I420Buffers {
+  uint8* y;
+  int y_stride;
+  uint8* u;
+  int u_stride;
+  uint8* v;
+  int v_stride;
+  int w;
+  int h;
+};
+
+static void JpegCopyI420(void* opaque,
+                         const uint8* const* data,
+                         const int* strides,
+                         int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I420Copy(data[0], strides[0],
+           data[1], strides[1],
+           data[2], strides[2],
+           dest->y, dest->y_stride,
+           dest->u, dest->u_stride,
+           dest->v, dest->v_stride,
+           dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I422ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I444ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI411ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I411ToI420(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToI420(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  I420Buffers* dest = (I420Buffers*)(opaque);
+  I400ToI420(data[0], strides[0],
+             dest->y, dest->y_stride,
+             dest->u, dest->u_stride,
+             dest->v, dest->v_stride,
+             dest->w, rows);
+  dest->y += rows * dest->y_stride;
+  dest->u += ((rows + 1) >> 1) * dest->u_stride;
+  dest->v += ((rows + 1) >> 1) * dest->v_stride;
+  dest->h -= rows;
+}
+
+// Query size of MJPG in pixels.
+LIBYUV_API
+int MJPGSize(const uint8* sample, size_t sample_size,
+             int* width, int* height) {
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret) {
+    *width = mjpeg_decoder.GetWidth();
+    *height = mjpeg_decoder.GetHeight();
+  }
+  mjpeg_decoder.UnloadFrame();
+  return ret ? 0 : -1;  // -1 for runtime failure.
+}
+
+// MJPG (Motion JPeg) to I420
+// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+LIBYUV_API
+int MJPGToI420(const uint8* sample,
+               size_t sample_size,
+               uint8* y, int y_stride,
+               uint8* u, int u_stride,
+               uint8* v, int v_stride,
+               int w, int h,
+               int dw, int dh) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret && (mjpeg_decoder.GetWidth() != w ||
+              mjpeg_decoder.GetHeight() != h)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    I420Buffers bufs = { y, y_stride, u, u_stride, v, v_stride, dw, dh };
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() ==
+            MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegCopyI420, &bufs, dw, dh);
+    // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToI420, &bufs, dw, dh);
+    // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToI420, &bufs, dw, dh);
+    // YUV411
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 4 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToI420, &bufs, dw, dh);
+    // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToI420, &bufs, dw, dh);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other colorspace/sample
+      // factors that occur in practice. 411 is supported by libjpeg
+      // ERROR: Unable to convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+
+#ifdef HAVE_JPEG
+struct ARGBBuffers {
+  uint8* argb;
+  int argb_stride;
+  int w;
+  int h;
+};
+
+static void JpegI420ToARGB(void* opaque,
+                         const uint8* const* data,
+                         const int* strides,
+                         int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I420ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI422ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I422ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI444ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I444ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI411ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I411ToARGB(data[0], strides[0],
+             data[1], strides[1],
+             data[2], strides[2],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+static void JpegI400ToARGB(void* opaque,
+                           const uint8* const* data,
+                           const int* strides,
+                           int rows) {
+  ARGBBuffers* dest = (ARGBBuffers*)(opaque);
+  I400ToARGB(data[0], strides[0],
+             dest->argb, dest->argb_stride,
+             dest->w, rows);
+  dest->argb += rows * dest->argb_stride;
+  dest->h -= rows;
+}
+
+// MJPG (Motion JPeg) to ARGB
+// TODO(fbarchard): review w and h requirement. dw and dh may be enough.
+LIBYUV_API
+int MJPGToARGB(const uint8* sample,
+               size_t sample_size,
+               uint8* argb, int argb_stride,
+               int w, int h,
+               int dw, int dh) {
+  if (sample_size == kUnknownDataSize) {
+    // ERROR: MJPEG frame size unknown
+    return -1;
+  }
+
+  // TODO(fbarchard): Port MJpeg to C.
+  MJpegDecoder mjpeg_decoder;
+  LIBYUV_BOOL ret = mjpeg_decoder.LoadFrame(sample, sample_size);
+  if (ret && (mjpeg_decoder.GetWidth() != w ||
+              mjpeg_decoder.GetHeight() != h)) {
+    // ERROR: MJPEG frame has unexpected dimensions
+    mjpeg_decoder.UnloadFrame();
+    return 1;  // runtime failure
+  }
+  if (ret) {
+    ARGBBuffers bufs = { argb, argb_stride, dw, dh };
+    // YUV420
+    if (mjpeg_decoder.GetColorSpace() ==
+            MJpegDecoder::kColorSpaceYCbCr &&
+        mjpeg_decoder.GetNumComponents() == 3 &&
+        mjpeg_decoder.GetVertSampFactor(0) == 2 &&
+        mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+        mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+        mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+        mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI420ToARGB, &bufs, dw, dh);
+    // YUV422
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 2 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI422ToARGB, &bufs, dw, dh);
+    // YUV444
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI444ToARGB, &bufs, dw, dh);
+    // YUV411
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceYCbCr &&
+               mjpeg_decoder.GetNumComponents() == 3 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 4 &&
+               mjpeg_decoder.GetVertSampFactor(1) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(1) == 1 &&
+               mjpeg_decoder.GetVertSampFactor(2) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(2) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI411ToARGB, &bufs, dw, dh);
+    // YUV400
+    } else if (mjpeg_decoder.GetColorSpace() ==
+                   MJpegDecoder::kColorSpaceGrayscale &&
+               mjpeg_decoder.GetNumComponents() == 1 &&
+               mjpeg_decoder.GetVertSampFactor(0) == 1 &&
+               mjpeg_decoder.GetHorizSampFactor(0) == 1) {
+      ret = mjpeg_decoder.DecodeToCallback(&JpegI400ToARGB, &bufs, dw, dh);
+    } else {
+      // TODO(fbarchard): Implement conversion for any other colorspace/sample
+      // factors that occur in practice. 411 is supported by libjpeg
+      // ERROR: Unable to convert MJPEG frame because format is not supported
+      mjpeg_decoder.UnloadFrame();
+      return 1;
+    }
+  }
+  return ret ? 0 : 1;
+}
+#endif
+
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/convert_to_argb.cc b/third_party/aom/third_party/libyuv/source/convert_to_argb.cc
new file mode 100644
index 0000000000..af829fbd32
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/convert_to_argb.cc
@@ -0,0 +1,306 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/convert_argb.h"
+
+#include "libyuv/cpu_id.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/rotate_argb.h"
+#include "libyuv/row.h"
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// src_width is used for source stride computation
+// src_height is used to compute location of planes, and indicate inversion
+// sample_size is measured in bytes and is the size of the frame.
+//   With MJPEG it is the compressed size of the frame.
+LIBYUV_API
+int ConvertToARGB(const uint8* sample, size_t sample_size,
+                  uint8* crop_argb, int argb_stride,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int aligned_src_width = (src_width + 1) & ~1;
+  const uint8* src;
+  const uint8* src_uv;
+  int abs_src_height = (src_height < 0) ? -src_height : src_height;
+  int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+  int r = 0;
+
+  // One pass rotation is available for some formats. For the rest, convert
+  // to I420 (with optional vertical flipping) into a temporary I420 buffer,
+  // and then rotate the I420 to the final destination buffer.
+  // For in-place conversion, if destination crop_argb is same as source sample,
+  // also enable temporary buffer.
+  LIBYUV_BOOL need_buf = (rotation && format != FOURCC_ARGB) ||
+      crop_argb == sample;
+  uint8* tmp_argb = crop_argb;
+  int tmp_argb_stride = argb_stride;
+  uint8* rotate_buffer = NULL;
+  int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+
+  if (crop_argb == NULL || sample == NULL ||
+      src_width <= 0 || crop_width <= 0 ||
+      src_height == 0 || crop_height == 0) {
+    return -1;
+  }
+  if (src_height < 0) {
+    inv_crop_height = -inv_crop_height;
+  }
+
+  if (need_buf) {
+    int argb_size = crop_width * abs_crop_height * 4;
+    rotate_buffer = (uint8*)malloc(argb_size);
+    if (!rotate_buffer) {
+      return 1;  // Out of memory runtime error.
+    }
+    crop_argb = rotate_buffer;
+    argb_stride = crop_width;
+  }
+
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = YUY2ToARGB(src, aligned_src_width * 2,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_UYVY:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = UYVYToARGB(src, aligned_src_width * 2,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_24BG:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RGB24ToARGB(src, src_width * 3,
+                      crop_argb, argb_stride,
+                      crop_width, inv_crop_height);
+      break;
+    case FOURCC_RAW:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RAWToARGB(src, src_width * 3,
+                    crop_argb, argb_stride,
+                    crop_width, inv_crop_height);
+      break;
+    case FOURCC_ARGB:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ARGBToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_BGRA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = BGRAToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_ABGR:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ABGRToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = RGBAToARGB(src, src_width * 4,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBP:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = RGB565ToARGB(src, src_width * 2,
+                       crop_argb, argb_stride,
+                       crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBO:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB1555ToARGB(src, src_width * 2,
+                         crop_argb, argb_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_R444:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB4444ToARGB(src, src_width * 2,
+                         crop_argb, argb_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_I400:
+      src = sample + src_width * crop_y + crop_x;
+      r = I400ToARGB(src, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+
+    // Biplanar formats
+    case FOURCC_NV12:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      r = NV12ToARGB(src, src_width,
+                     src_uv, aligned_src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_NV21:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + aligned_src_width * (src_height + crop_y / 2) + crop_x;
+      // Call NV12 but with u and v parameters swapped.
+      r = NV21ToARGB(src, src_width,
+                     src_uv, aligned_src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_M420:
+      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
+      r = M420ToARGB(src, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // Triplanar formats
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      if (format == FOURCC_YV12) {
+        src_v = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      }
+      r = I420ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_J420: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      src_u = sample + src_width * abs_src_height +
+          (halfwidth * crop_y + crop_x) / 2;
+      src_v = sample + src_width * abs_src_height +
+          halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      r = J420ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      if (format == FOURCC_YV16) {
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      }
+      r = I422ToARGB(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      if (format == FOURCC_YV24) {
+        src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      } else {
+        src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      }
+      r = I444ToARGB(src_y, src_width,
+                     src_u, src_width,
+                     src_v, src_width,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (src_width + 3) / 4;
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u = sample + src_width * abs_src_height +
+          quarterwidth * crop_y + crop_x / 4;
+      const uint8* src_v = sample + src_width * abs_src_height +
+          quarterwidth * (abs_src_height + crop_y) + crop_x / 4;
+      r = I411ToARGB(src_y, src_width,
+                     src_u, quarterwidth,
+                     src_v, quarterwidth,
+                     crop_argb, argb_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+#ifdef HAVE_JPEG
+    case FOURCC_MJPG:
+      r = MJPGToARGB(sample, sample_size,
+                     crop_argb, argb_stride,
+                     src_width, abs_src_height, crop_width, inv_crop_height);
+      break;
+#endif
+    default:
+      r = -1;  // unknown fourcc - return failure code.
+  }
+
+  if (need_buf) {
+    if (!r) {
+      r = ARGBRotate(crop_argb, argb_stride,
+                     tmp_argb, tmp_argb_stride,
+                     crop_width, abs_crop_height, rotation);
+    }
+    free(rotate_buffer);
+  }
+
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/convert_to_i420.cc b/third_party/aom/third_party/libyuv/source/convert_to_i420.cc
new file mode 100644
index 0000000000..5e75369b55
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/convert_to_i420.cc
@@ -0,0 +1,339 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <stdlib.h>
+
+#include "libyuv/convert.h"
+
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Convert camera sample to I420 with cropping, rotation and vertical flip.
+// src_width is used for source stride computation
+// src_height is used to compute location of planes, and indicate inversion
+// sample_size is measured in bytes and is the size of the frame.
+//   With MJPEG it is the compressed size of the frame.
+LIBYUV_API
+int ConvertToI420(const uint8* sample,
+                  size_t sample_size,
+                  uint8* y, int y_stride,
+                  uint8* u, int u_stride,
+                  uint8* v, int v_stride,
+                  int crop_x, int crop_y,
+                  int src_width, int src_height,
+                  int crop_width, int crop_height,
+                  enum RotationMode rotation,
+                  uint32 fourcc) {
+  uint32 format = CanonicalFourCC(fourcc);
+  int aligned_src_width = (src_width + 1) & ~1;
+  const uint8* src;
+  const uint8* src_uv;
+  int abs_src_height = (src_height < 0) ? -src_height : src_height;
+  int inv_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+  int r = 0;
+  LIBYUV_BOOL need_buf = (rotation && format != FOURCC_I420 &&
+      format != FOURCC_NV12 && format != FOURCC_NV21 &&
+      format != FOURCC_YU12 && format != FOURCC_YV12) || y == sample;
+  uint8* tmp_y = y;
+  uint8* tmp_u = u;
+  uint8* tmp_v = v;
+  int tmp_y_stride = y_stride;
+  int tmp_u_stride = u_stride;
+  int tmp_v_stride = v_stride;
+  uint8* rotate_buffer = NULL;
+  int abs_crop_height = (crop_height < 0) ? -crop_height : crop_height;
+
+  if (!y || !u || !v || !sample ||
+      src_width <= 0 || crop_width <= 0  ||
+      src_height == 0 || crop_height == 0) {
+    return -1;
+  }
+  if (src_height < 0) {
+    inv_crop_height = -inv_crop_height;
+  }
+
+  // One pass rotation is available for some formats. For the rest, convert
+  // to I420 (with optional vertical flipping) into a temporary I420 buffer,
+  // and then rotate the I420 to the final destination buffer.
+  // For in-place conversion, if destination y is same as source sample,
+  // also enable temporary buffer.
+  if (need_buf) {
+    int y_size = crop_width * abs_crop_height;
+    int uv_size = ((crop_width + 1) / 2) * ((abs_crop_height + 1) / 2);
+    rotate_buffer = (uint8*)malloc(y_size + uv_size * 2);
+    if (!rotate_buffer) {
+      return 1;  // Out of memory runtime error.
+    }
+    y = rotate_buffer;
+    u = y + y_size;
+    v = u + uv_size;
+    y_stride = crop_width;
+    u_stride = v_stride = ((crop_width + 1) / 2);
+  }
+
+  switch (format) {
+    // Single plane formats
+    case FOURCC_YUY2:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = YUY2ToI420(src, aligned_src_width * 2,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_UYVY:
+      src = sample + (aligned_src_width * crop_y + crop_x) * 2;
+      r = UYVYToI420(src, aligned_src_width * 2,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBP:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = RGB565ToI420(src, src_width * 2,
+                       y, y_stride,
+                       u, u_stride,
+                       v, v_stride,
+                       crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBO:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB1555ToI420(src, src_width * 2,
+                         y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_R444:
+      src = sample + (src_width * crop_y + crop_x) * 2;
+      r = ARGB4444ToI420(src, src_width * 2,
+                         y, y_stride,
+                         u, u_stride,
+                         v, v_stride,
+                         crop_width, inv_crop_height);
+      break;
+    case FOURCC_24BG:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RGB24ToI420(src, src_width * 3,
+                      y, y_stride,
+                      u, u_stride,
+                      v, v_stride,
+                      crop_width, inv_crop_height);
+      break;
+    case FOURCC_RAW:
+      src = sample + (src_width * crop_y + crop_x) * 3;
+      r = RAWToI420(src, src_width * 3,
+                    y, y_stride,
+                    u, u_stride,
+                    v, v_stride,
+                    crop_width, inv_crop_height);
+      break;
+    case FOURCC_ARGB:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ARGBToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_BGRA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = BGRAToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_ABGR:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = ABGRToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_RGBA:
+      src = sample + (src_width * crop_y + crop_x) * 4;
+      r = RGBAToI420(src, src_width * 4,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    case FOURCC_I400:
+      src = sample + src_width * crop_y + crop_x;
+      r = I400ToI420(src, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // Biplanar formats
+    case FOURCC_NV12:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + (src_width * src_height) +
+        ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
+      r = NV12ToI420Rotate(src, src_width,
+                           src_uv, aligned_src_width,
+                           y, y_stride,
+                           u, u_stride,
+                           v, v_stride,
+                           crop_width, inv_crop_height, rotation);
+      break;
+    case FOURCC_NV21:
+      src = sample + (src_width * crop_y + crop_x);
+      src_uv = sample + (src_width * src_height) +
+        ((crop_y / 2) * aligned_src_width) + ((crop_x / 2) * 2);
+      // Call NV12 but with u and v parameters swapped.
+      r = NV12ToI420Rotate(src, src_width,
+                           src_uv, aligned_src_width,
+                           y, y_stride,
+                           v, v_stride,
+                           u, u_stride,
+                           crop_width, inv_crop_height, rotation);
+      break;
+    case FOURCC_M420:
+      src = sample + (src_width * crop_y) * 12 / 8 + crop_x;
+      r = M420ToI420(src, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    // Triplanar formats
+    case FOURCC_I420:
+    case FOURCC_YU12:
+    case FOURCC_YV12: {
+      const uint8* src_y = sample + (src_width * crop_y + crop_x);
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      int halfheight = (abs_src_height + 1) / 2;
+      if (format == FOURCC_YV12) {
+        src_v = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            (halfwidth * crop_y + crop_x) / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (halfheight + crop_y / 2) + crop_x / 2;
+      }
+      r = I420Rotate(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height, rotation);
+      break;
+    }
+    case FOURCC_I422:
+    case FOURCC_YV16: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      int halfwidth = (src_width + 1) / 2;
+      if (format == FOURCC_YV16) {
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      } else {
+        src_u = sample + src_width * abs_src_height +
+            halfwidth * crop_y + crop_x / 2;
+        src_v = sample + src_width * abs_src_height +
+            halfwidth * (abs_src_height + crop_y) + crop_x / 2;
+      }
+      r = I422ToI420(src_y, src_width,
+                     src_u, halfwidth,
+                     src_v, halfwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I444:
+    case FOURCC_YV24: {
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u;
+      const uint8* src_v;
+      if (format == FOURCC_YV24) {
+        src_v = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_u = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      } else {
+        src_u = sample + src_width * (abs_src_height + crop_y) + crop_x;
+        src_v = sample + src_width * (abs_src_height * 2 + crop_y) + crop_x;
+      }
+      r = I444ToI420(src_y, src_width,
+                     src_u, src_width,
+                     src_v, src_width,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+    case FOURCC_I411: {
+      int quarterwidth = (src_width + 3) / 4;
+      const uint8* src_y = sample + src_width * crop_y + crop_x;
+      const uint8* src_u = sample + src_width * abs_src_height +
+          quarterwidth * crop_y + crop_x / 4;
+      const uint8* src_v = sample + src_width * abs_src_height +
+          quarterwidth * (abs_src_height + crop_y) + crop_x / 4;
+      r = I411ToI420(src_y, src_width,
+                     src_u, quarterwidth,
+                     src_v, quarterwidth,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     crop_width, inv_crop_height);
+      break;
+    }
+#ifdef HAVE_JPEG
+    case FOURCC_MJPG:
+      r = MJPGToI420(sample, sample_size,
+                     y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     src_width, abs_src_height, crop_width, inv_crop_height);
+      break;
+#endif
+    default:
+      r = -1;  // unknown fourcc - return failure code.
+  }
+
+  if (need_buf) {
+    if (!r) {
+      r = I420Rotate(y, y_stride,
+                     u, u_stride,
+                     v, v_stride,
+                     tmp_y, tmp_y_stride,
+                     tmp_u, tmp_u_stride,
+                     tmp_v, tmp_v_stride,
+                     crop_width, abs_crop_height, rotation);
+    }
+    free(rotate_buffer);
+  }
+
+  return r;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/cpu_id.cc b/third_party/aom/third_party/libyuv/source/cpu_id.cc
new file mode 100644
index 0000000000..72f686e3b3
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/cpu_id.cc
@@ -0,0 +1,307 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/cpu_id.h"
+
+#if (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+#include <intrin.h>  // For __cpuidex()
+#endif
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    !defined(__native_client__) && (defined(_M_IX86) || defined(_M_X64)) && \
+    defined(_MSC_VER) && !defined(__clang__) && (_MSC_FULL_VER >= 160040219)
+#include <immintrin.h>  // For _xgetbv()
+#endif
+
+#if !defined(__native_client__)
+#include <stdlib.h>  // For getenv()
+#endif
+
+// For ArmCpuCaps() but unittested on all platforms
+#include <stdio.h>
+#include <string.h>
+
+#include "libyuv/basic_types.h"  // For CPU_X86
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// For functions that use the stack and have runtime checks for overflow,
+// use SAFEBUFFERS to avoid additional check.
+#if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219)
+#define SAFEBUFFERS __declspec(safebuffers)
+#else
+#define SAFEBUFFERS
+#endif
+
+// Low level cpuid for X86.
+#if (defined(_M_IX86) || defined(_M_X64) || \
+    defined(__i386__) || defined(__x86_64__)) && \
+    !defined(__pnacl__) && !defined(__CLR_VER)
+LIBYUV_API
+void CpuId(uint32 info_eax, uint32 info_ecx, uint32* cpu_info) {
+#if (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+// Visual C version uses intrinsic or inline x86 assembly.
+#if (_MSC_FULL_VER >= 160040219)
+  __cpuidex((int*)(cpu_info), info_eax, info_ecx);
+#elif defined(_M_IX86)
+  __asm {
+    mov        eax, info_eax
+    mov        ecx, info_ecx
+    mov        edi, cpu_info
+    cpuid
+    mov        [edi], eax
+    mov        [edi + 4], ebx
+    mov        [edi + 8], ecx
+    mov        [edi + 12], edx
+  }
+#else
+  if (info_ecx == 0) {
+    __cpuid((int*)(cpu_info), info_eax);
+  } else {
+    cpu_info[3] = cpu_info[2] = cpu_info[1] = cpu_info[0] = 0;
+  }
+#endif
+// GCC version uses inline x86 assembly.
+#else  // (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+  uint32 info_ebx, info_edx;
+  asm volatile (  // NOLINT
+#if defined( __i386__) && defined(__PIC__)
+    // Preserve ebx for fpic 32 bit.
+    "mov %%ebx, %%edi                          \n"
+    "cpuid                                     \n"
+    "xchg %%edi, %%ebx                         \n"
+    : "=D" (info_ebx),
+#else
+    "cpuid                                     \n"
+    : "=b" (info_ebx),
+#endif  //  defined( __i386__) && defined(__PIC__)
+      "+a" (info_eax), "+c" (info_ecx), "=d" (info_edx));
+  cpu_info[0] = info_eax;
+  cpu_info[1] = info_ebx;
+  cpu_info[2] = info_ecx;
+  cpu_info[3] = info_edx;
+#endif  // (defined(_MSC_VER) && !defined(__clang__)) && !defined(__clang__)
+}
+#else  // (defined(_M_IX86) || defined(_M_X64) ...
+LIBYUV_API
+void CpuId(uint32 eax, uint32 ecx, uint32* cpu_info) {
+  cpu_info[0] = cpu_info[1] = cpu_info[2] = cpu_info[3] = 0;
+}
+#endif
+
+// TODO(fbarchard): Enable xgetbv when validator supports it.
+#if (defined(_M_IX86) || defined(_M_X64) || \
+    defined(__i386__) || defined(__x86_64__)) && \
+    !defined(__pnacl__) && !defined(__CLR_VER) && !defined(__native_client__)
+#define HAS_XGETBV
+// X86 CPUs have xgetbv to detect OS saves high parts of ymm registers.
+int TestOsSaveYmm() {
+  uint32 xcr0 = 0u;
+#if (defined(_MSC_VER) && !defined(__clang__)) && (_MSC_FULL_VER >= 160040219)
+  xcr0 = (uint32)(_xgetbv(0));  // VS2010 SP1 required.
+#elif defined(_M_IX86) && defined(_MSC_VER) && !defined(__clang__)
+  __asm {
+    xor        ecx, ecx    // xcr 0
+    _asm _emit 0x0f _asm _emit 0x01 _asm _emit 0xd0  // For VS2010 and earlier.
+    mov        xcr0, eax
+  }
+#elif defined(__i386__) || defined(__x86_64__)
+  asm(".byte 0x0f, 0x01, 0xd0" : "=a" (xcr0) : "c" (0) : "%edx");
+#endif  // defined(__i386__) || defined(__x86_64__)
+  return((xcr0 & 6) == 6);  // Is ymm saved?
+}
+#endif  // defined(_M_IX86) || defined(_M_X64) ..
+
+// based on libaom arm_cpudetect.c
+// For Arm, but public to allow testing on any CPU
+LIBYUV_API SAFEBUFFERS
+int ArmCpuCaps(const char* cpuinfo_name) {
+  char cpuinfo_line[512];
+  FILE* f = fopen(cpuinfo_name, "r");
+  if (!f) {
+    // Assume Neon if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasNEON;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f)) {
+    if (memcmp(cpuinfo_line, "Features", 8) == 0) {
+      char* p = strstr(cpuinfo_line, " neon");
+      if (p && (p[5] == ' ' || p[5] == '\n')) {
+        fclose(f);
+        return kCpuHasNEON;
+      }
+      // aarch64 uses asimd for Neon.
+      p = strstr(cpuinfo_line, " asimd");
+      if (p && (p[6] == ' ' || p[6] == '\n')) {
+        fclose(f);
+        return kCpuHasNEON;
+      }
+    }
+  }
+  fclose(f);
+  return 0;
+}
+
+#if defined(__mips__) && defined(__linux__)
+static int MipsCpuCaps(const char* search_string) {
+  char cpuinfo_line[512];
+  const char* file_name = "/proc/cpuinfo";
+  FILE* f = fopen(file_name, "r");
+  if (!f) {
+    // Assume DSP if /proc/cpuinfo is unavailable.
+    // This will occur for Chrome sandbox for Pepper or Render process.
+    return kCpuHasMIPS_DSP;
+  }
+  while (fgets(cpuinfo_line, sizeof(cpuinfo_line) - 1, f) != NULL) {
+    if (strstr(cpuinfo_line, search_string) != NULL) {
+      fclose(f);
+      return kCpuHasMIPS_DSP;
+    }
+  }
+  fclose(f);
+  return 0;
+}
+#endif
+
+// CPU detect function for SIMD instruction sets.
+LIBYUV_API
+int cpu_info_ = kCpuInit;  // cpu_info is not initialized yet.
+
+// Test environment variable for disabling CPU features. Any non-zero value
+// to disable. Zero ignored to make it easy to set the variable on/off.
+#if !defined(__native_client__) && !defined(_M_ARM)
+
+static LIBYUV_BOOL TestEnv(const char* name) {
+  const char* var = getenv(name);
+  if (var) {
+    if (var[0] != '0') {
+      return LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_FALSE;
+}
+#else  // nacl does not support getenv().
+static LIBYUV_BOOL TestEnv(const char*) {
+  return LIBYUV_FALSE;
+}
+#endif
+
+LIBYUV_API SAFEBUFFERS
+int InitCpuFlags(void) {
+#if !defined(__pnacl__) && !defined(__CLR_VER) && defined(CPU_X86)
+
+  uint32 cpu_info0[4] = { 0, 0, 0, 0 };
+  uint32 cpu_info1[4] = { 0, 0, 0, 0 };
+  uint32 cpu_info7[4] = { 0, 0, 0, 0 };
+  CpuId(0, 0, cpu_info0);
+  CpuId(1, 0, cpu_info1);
+  if (cpu_info0[0] >= 7) {
+    CpuId(7, 0, cpu_info7);
+  }
+  cpu_info_ = ((cpu_info1[3] & 0x04000000) ? kCpuHasSSE2 : 0) |
+              ((cpu_info1[2] & 0x00000200) ? kCpuHasSSSE3 : 0) |
+              ((cpu_info1[2] & 0x00080000) ? kCpuHasSSE41 : 0) |
+              ((cpu_info1[2] & 0x00100000) ? kCpuHasSSE42 : 0) |
+              ((cpu_info7[1] & 0x00000200) ? kCpuHasERMS : 0) |
+              ((cpu_info1[2] & 0x00001000) ? kCpuHasFMA3 : 0) |
+              kCpuHasX86;
+
+#ifdef HAS_XGETBV
+  if ((cpu_info1[2] & 0x18000000) == 0x18000000 &&  // AVX and OSSave
+      TestOsSaveYmm()) {  // Saves YMM.
+    cpu_info_ |= ((cpu_info7[1] & 0x00000020) ? kCpuHasAVX2 : 0) |
+                 kCpuHasAVX;
+  }
+#endif
+  // Environment variable overrides for testing.
+  if (TestEnv("LIBYUV_DISABLE_X86")) {
+    cpu_info_ &= ~kCpuHasX86;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE2")) {
+    cpu_info_ &= ~kCpuHasSSE2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSSE3")) {
+    cpu_info_ &= ~kCpuHasSSSE3;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE41")) {
+    cpu_info_ &= ~kCpuHasSSE41;
+  }
+  if (TestEnv("LIBYUV_DISABLE_SSE42")) {
+    cpu_info_ &= ~kCpuHasSSE42;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX")) {
+    cpu_info_ &= ~kCpuHasAVX;
+  }
+  if (TestEnv("LIBYUV_DISABLE_AVX2")) {
+    cpu_info_ &= ~kCpuHasAVX2;
+  }
+  if (TestEnv("LIBYUV_DISABLE_ERMS")) {
+    cpu_info_ &= ~kCpuHasERMS;
+  }
+  if (TestEnv("LIBYUV_DISABLE_FMA3")) {
+    cpu_info_ &= ~kCpuHasFMA3;
+  }
+#endif
+#if defined(__mips__) && defined(__linux__)
+  // Linux mips parse text file for dsp detect.
+  cpu_info_ = MipsCpuCaps("dsp");  // set kCpuHasMIPS_DSP.
+#if defined(__mips_dspr2)
+  cpu_info_ |= kCpuHasMIPS_DSPR2;
+#endif
+  cpu_info_ |= kCpuHasMIPS;
+
+  if (getenv("LIBYUV_DISABLE_MIPS")) {
+    cpu_info_ &= ~kCpuHasMIPS;
+  }
+  if (getenv("LIBYUV_DISABLE_MIPS_DSP")) {
+    cpu_info_ &= ~kCpuHasMIPS_DSP;
+  }
+  if (getenv("LIBYUV_DISABLE_MIPS_DSPR2")) {
+    cpu_info_ &= ~kCpuHasMIPS_DSPR2;
+  }
+#endif
+#if defined(__arm__) || defined(__aarch64__)
+// gcc -mfpu=neon defines __ARM_NEON__
+// __ARM_NEON__ generates code that requires Neon.  NaCL also requires Neon.
+// For Linux, /proc/cpuinfo can be tested but without that assume Neon.
+#if defined(__ARM_NEON__) || defined(__native_client__) || !defined(__linux__)
+  cpu_info_ = kCpuHasNEON;
+// For aarch64(arm64), /proc/cpuinfo's feature is not complete, e.g. no neon
+// flag in it.
+// So for aarch64, neon enabling is hard coded here.
+#endif
+#if defined(__aarch64__)
+  cpu_info_ = kCpuHasNEON;
+#else
+  // Linux arm parse text file for neon detect.
+  cpu_info_ = ArmCpuCaps("/proc/cpuinfo");
+#endif
+  cpu_info_ |= kCpuHasARM;
+  if (TestEnv("LIBYUV_DISABLE_NEON")) {
+    cpu_info_ &= ~kCpuHasNEON;
+  }
+#endif  // __arm__
+  if (TestEnv("LIBYUV_DISABLE_ASM")) {
+    cpu_info_ = 0;
+  }
+  return cpu_info_;
+}
+
+LIBYUV_API
+void MaskCpuFlags(int enable_flags) {
+  cpu_info_ = InitCpuFlags() & enable_flags;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/mjpeg_decoder.cc b/third_party/aom/third_party/libyuv/source/mjpeg_decoder.cc
new file mode 100644
index 0000000000..75f8a610e3
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/mjpeg_decoder.cc
@@ -0,0 +1,572 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/mjpeg_decoder.h"
+
+#ifdef HAVE_JPEG
+#include <assert.h>
+
+#if !defined(__pnacl__) && !defined(__CLR_VER) && \
+    !defined(COVERAGE_ENABLED) && !defined(TARGET_IPHONE_SIMULATOR)
+// Must be included before jpeglib.
+#include <setjmp.h>
+#define HAVE_SETJMP
+
+#if defined(_MSC_VER)
+// disable warning 4324: structure was padded due to __declspec(align())
+#pragma warning(disable:4324)
+#endif
+
+#endif
+struct FILE;  // For jpeglib.h.
+
+// C++ build requires extern C for jpeg internals.
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include <jpeglib.h>
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#include "libyuv/planar_functions.h"  // For CopyPlane().
+
+namespace libyuv {
+
+#ifdef HAVE_SETJMP
+struct SetJmpErrorMgr {
+  jpeg_error_mgr base;  // Must be at the top
+  jmp_buf setjmp_buffer;
+};
+#endif
+
+const int MJpegDecoder::kColorSpaceUnknown = JCS_UNKNOWN;
+const int MJpegDecoder::kColorSpaceGrayscale = JCS_GRAYSCALE;
+const int MJpegDecoder::kColorSpaceRgb = JCS_RGB;
+const int MJpegDecoder::kColorSpaceYCbCr = JCS_YCbCr;
+const int MJpegDecoder::kColorSpaceCMYK = JCS_CMYK;
+const int MJpegDecoder::kColorSpaceYCCK = JCS_YCCK;
+
+// Methods that are passed to jpeglib.
+boolean fill_input_buffer(jpeg_decompress_struct* cinfo);
+void init_source(jpeg_decompress_struct* cinfo);
+void skip_input_data(jpeg_decompress_struct* cinfo,
+                     long num_bytes);  // NOLINT
+void term_source(jpeg_decompress_struct* cinfo);
+void ErrorHandler(jpeg_common_struct* cinfo);
+
+MJpegDecoder::MJpegDecoder()
+    : has_scanline_padding_(LIBYUV_FALSE),
+      num_outbufs_(0),
+      scanlines_(NULL),
+      scanlines_sizes_(NULL),
+      databuf_(NULL),
+      databuf_strides_(NULL) {
+  decompress_struct_ = new jpeg_decompress_struct;
+  source_mgr_ = new jpeg_source_mgr;
+#ifdef HAVE_SETJMP
+  error_mgr_ = new SetJmpErrorMgr;
+  decompress_struct_->err = jpeg_std_error(&error_mgr_->base);
+  // Override standard exit()-based error handler.
+  error_mgr_->base.error_exit = &ErrorHandler;
+#endif
+  decompress_struct_->client_data = NULL;
+  source_mgr_->init_source = &init_source;
+  source_mgr_->fill_input_buffer = &fill_input_buffer;
+  source_mgr_->skip_input_data = &skip_input_data;
+  source_mgr_->resync_to_restart = &jpeg_resync_to_restart;
+  source_mgr_->term_source = &term_source;
+  jpeg_create_decompress(decompress_struct_);
+  decompress_struct_->src = source_mgr_;
+  buf_vec_.buffers = &buf_;
+  buf_vec_.len = 1;
+}
+
+MJpegDecoder::~MJpegDecoder() {
+  jpeg_destroy_decompress(decompress_struct_);
+  delete decompress_struct_;
+  delete source_mgr_;
+#ifdef HAVE_SETJMP
+  delete error_mgr_;
+#endif
+  DestroyOutputBuffers();
+}
+
+LIBYUV_BOOL MJpegDecoder::LoadFrame(const uint8* src, size_t src_len) {
+  if (!ValidateJpeg(src, src_len)) {
+    return LIBYUV_FALSE;
+  }
+
+  buf_.data = src;
+  buf_.len = static_cast<int>(src_len);
+  buf_vec_.pos = 0;
+  decompress_struct_->client_data = &buf_vec_;
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called jpeg_read_header, it experienced an error, and we called
+    // longjmp() and rewound the stack to here. Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (jpeg_read_header(decompress_struct_, TRUE) != JPEG_HEADER_OK) {
+    // ERROR: Bad MJPEG header
+    return LIBYUV_FALSE;
+  }
+  AllocOutputBuffers(GetNumComponents());
+  for (int i = 0; i < num_outbufs_; ++i) {
+    int scanlines_size = GetComponentScanlinesPerImcuRow(i);
+    if (scanlines_sizes_[i] != scanlines_size) {
+      if (scanlines_[i]) {
+        delete scanlines_[i];
+      }
+      scanlines_[i] = new uint8* [scanlines_size];
+      scanlines_sizes_[i] = scanlines_size;
+    }
+
+    // We allocate padding for the final scanline to pad it up to DCTSIZE bytes
+    // to avoid memory errors, since jpeglib only reads full MCUs blocks. For
+    // the preceding scanlines, the padding is not needed/wanted because the
+    // following addresses will already be valid (they are the initial bytes of
+    // the next scanline) and will be overwritten when jpeglib writes out that
+    // next scanline.
+    int databuf_stride = GetComponentStride(i);
+    int databuf_size = scanlines_size * databuf_stride;
+    if (databuf_strides_[i] != databuf_stride) {
+      if (databuf_[i]) {
+        delete databuf_[i];
+      }
+      databuf_[i] = new uint8[databuf_size];
+      databuf_strides_[i] = databuf_stride;
+    }
+
+    if (GetComponentStride(i) != GetComponentWidth(i)) {
+      has_scanline_padding_ = LIBYUV_TRUE;
+    }
+  }
+  return LIBYUV_TRUE;
+}
+
+static int DivideAndRoundUp(int numerator, int denominator) {
+  return (numerator + denominator - 1) / denominator;
+}
+
+static int DivideAndRoundDown(int numerator, int denominator) {
+  return numerator / denominator;
+}
+
+// Returns width of the last loaded frame.
+int MJpegDecoder::GetWidth() {
+  return decompress_struct_->image_width;
+}
+
+// Returns height of the last loaded frame.
+int MJpegDecoder::GetHeight() {
+  return decompress_struct_->image_height;
+}
+
+// Returns format of the last loaded frame. The return value is one of the
+// kColorSpace* constants.
+int MJpegDecoder::GetColorSpace() {
+  return decompress_struct_->jpeg_color_space;
+}
+
+// Number of color components in the color space.
+int MJpegDecoder::GetNumComponents() {
+  return decompress_struct_->num_components;
+}
+
+// Sample factors of the n-th component.
+int MJpegDecoder::GetHorizSampFactor(int component) {
+  return decompress_struct_->comp_info[component].h_samp_factor;
+}
+
+int MJpegDecoder::GetVertSampFactor(int component) {
+  return decompress_struct_->comp_info[component].v_samp_factor;
+}
+
+int MJpegDecoder::GetHorizSubSampFactor(int component) {
+  return decompress_struct_->max_h_samp_factor /
+      GetHorizSampFactor(component);
+}
+
+int MJpegDecoder::GetVertSubSampFactor(int component) {
+  return decompress_struct_->max_v_samp_factor /
+      GetVertSampFactor(component);
+}
+
+int MJpegDecoder::GetImageScanlinesPerImcuRow() {
+  return decompress_struct_->max_v_samp_factor * DCTSIZE;
+}
+
+int MJpegDecoder::GetComponentScanlinesPerImcuRow(int component) {
+  int vs = GetVertSubSampFactor(component);
+  return DivideAndRoundUp(GetImageScanlinesPerImcuRow(), vs);
+}
+
+int MJpegDecoder::GetComponentWidth(int component) {
+  int hs = GetHorizSubSampFactor(component);
+  return DivideAndRoundUp(GetWidth(), hs);
+}
+
+int MJpegDecoder::GetComponentHeight(int component) {
+  int vs = GetVertSubSampFactor(component);
+  return DivideAndRoundUp(GetHeight(), vs);
+}
+
+// Get width in bytes padded out to a multiple of DCTSIZE
+int MJpegDecoder::GetComponentStride(int component) {
+  return (GetComponentWidth(component) + DCTSIZE - 1) & ~(DCTSIZE - 1);
+}
+
+int MJpegDecoder::GetComponentSize(int component) {
+  return GetComponentWidth(component) * GetComponentHeight(component);
+}
+
+LIBYUV_BOOL MJpegDecoder::UnloadFrame() {
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called jpeg_abort_decompress, it experienced an error, and we called
+    // longjmp() and rewound the stack to here. Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  jpeg_abort_decompress(decompress_struct_);
+  return LIBYUV_TRUE;
+}
+
+// TODO(fbarchard): Allow rectangle to be specified: x, y, width, height.
+LIBYUV_BOOL MJpegDecoder::DecodeToBuffers(
+    uint8** planes, int dst_width, int dst_height) {
+  if (dst_width != GetWidth() ||
+      dst_height > GetHeight()) {
+    // ERROR: Bad dimensions
+    return LIBYUV_FALSE;
+  }
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called into jpeglib, it experienced an error sometime during this
+    // function call, and we called longjmp() and rewound the stack to here.
+    // Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (!StartDecode()) {
+    return LIBYUV_FALSE;
+  }
+  SetScanlinePointers(databuf_);
+  int lines_left = dst_height;
+  // Compute amount of lines to skip to implement vertical crop.
+  // TODO(fbarchard): Ensure skip is a multiple of maximum component
+  // subsample. ie 2
+  int skip = (GetHeight() - dst_height) / 2;
+  if (skip > 0) {
+    // There is no API to skip lines in the output data, so we read them
+    // into the temp buffer.
+    while (skip >= GetImageScanlinesPerImcuRow()) {
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      skip -= GetImageScanlinesPerImcuRow();
+    }
+    if (skip > 0) {
+      // Have a partial iMCU row left over to skip. Must read it and then
+      // copy the parts we want into the destination.
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      for (int i = 0; i < num_outbufs_; ++i) {
+        // TODO(fbarchard): Compute skip to avoid this
+        assert(skip % GetVertSubSampFactor(i) == 0);
+        int rows_to_skip =
+            DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i) -
+                                rows_to_skip;
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        CopyPlane(databuf_[i] + data_to_skip, GetComponentStride(i),
+                  planes[i], GetComponentWidth(i),
+                  GetComponentWidth(i), scanlines_to_copy);
+        planes[i] += scanlines_to_copy * GetComponentWidth(i);
+      }
+      lines_left -= (GetImageScanlinesPerImcuRow() - skip);
+    }
+  }
+
+  // Read full MCUs but cropped horizontally
+  for (; lines_left > GetImageScanlinesPerImcuRow();
+         lines_left -= GetImageScanlinesPerImcuRow()) {
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    for (int i = 0; i < num_outbufs_; ++i) {
+      int scanlines_to_copy = GetComponentScanlinesPerImcuRow(i);
+      CopyPlane(databuf_[i], GetComponentStride(i),
+                planes[i], GetComponentWidth(i),
+                GetComponentWidth(i), scanlines_to_copy);
+      planes[i] += scanlines_to_copy * GetComponentWidth(i);
+    }
+  }
+
+  if (lines_left > 0) {
+    // Have a partial iMCU row left over to decode.
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    for (int i = 0; i < num_outbufs_; ++i) {
+      int scanlines_to_copy =
+          DivideAndRoundUp(lines_left, GetVertSubSampFactor(i));
+      CopyPlane(databuf_[i], GetComponentStride(i),
+                planes[i], GetComponentWidth(i),
+                GetComponentWidth(i), scanlines_to_copy);
+      planes[i] += scanlines_to_copy * GetComponentWidth(i);
+    }
+  }
+  return FinishDecode();
+}
+
+LIBYUV_BOOL MJpegDecoder::DecodeToCallback(CallbackFunction fn, void* opaque,
+    int dst_width, int dst_height) {
+  if (dst_width != GetWidth() ||
+      dst_height > GetHeight()) {
+    // ERROR: Bad dimensions
+    return LIBYUV_FALSE;
+  }
+#ifdef HAVE_SETJMP
+  if (setjmp(error_mgr_->setjmp_buffer)) {
+    // We called into jpeglib, it experienced an error sometime during this
+    // function call, and we called longjmp() and rewound the stack to here.
+    // Return error.
+    return LIBYUV_FALSE;
+  }
+#endif
+  if (!StartDecode()) {
+    return LIBYUV_FALSE;
+  }
+  SetScanlinePointers(databuf_);
+  int lines_left = dst_height;
+  // TODO(fbarchard): Compute amount of lines to skip to implement vertical crop
+  int skip = (GetHeight() - dst_height) / 2;
+  if (skip > 0) {
+    while (skip >= GetImageScanlinesPerImcuRow()) {
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      skip -= GetImageScanlinesPerImcuRow();
+    }
+    if (skip > 0) {
+      // Have a partial iMCU row left over to skip.
+      if (!DecodeImcuRow()) {
+        FinishDecode();
+        return LIBYUV_FALSE;
+      }
+      for (int i = 0; i < num_outbufs_; ++i) {
+        // TODO(fbarchard): Compute skip to avoid this
+        assert(skip % GetVertSubSampFactor(i) == 0);
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        // Change our own data buffer pointers so we can pass them to the
+        // callback.
+        databuf_[i] += data_to_skip;
+      }
+      int scanlines_to_copy = GetImageScanlinesPerImcuRow() - skip;
+      (*fn)(opaque, databuf_, databuf_strides_, scanlines_to_copy);
+      // Now change them back.
+      for (int i = 0; i < num_outbufs_; ++i) {
+        int rows_to_skip = DivideAndRoundDown(skip, GetVertSubSampFactor(i));
+        int data_to_skip = rows_to_skip * GetComponentStride(i);
+        databuf_[i] -= data_to_skip;
+      }
+      lines_left -= scanlines_to_copy;
+    }
+  }
+  // Read full MCUs until we get to the crop point.
+  for (; lines_left >= GetImageScanlinesPerImcuRow();
+         lines_left -= GetImageScanlinesPerImcuRow()) {
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    (*fn)(opaque, databuf_, databuf_strides_, GetImageScanlinesPerImcuRow());
+  }
+  if (lines_left > 0) {
+    // Have a partial iMCU row left over to decode.
+    if (!DecodeImcuRow()) {
+      FinishDecode();
+      return LIBYUV_FALSE;
+    }
+    (*fn)(opaque, databuf_, databuf_strides_, lines_left);
+  }
+  return FinishDecode();
+}
+
+void init_source(j_decompress_ptr cinfo) {
+  fill_input_buffer(cinfo);
+}
+
+boolean fill_input_buffer(j_decompress_ptr cinfo) {
+  BufferVector* buf_vec = reinterpret_cast<BufferVector*>(cinfo->client_data);
+  if (buf_vec->pos >= buf_vec->len) {
+    assert(0 && "No more data");
+    // ERROR: No more data
+    return FALSE;
+  }
+  cinfo->src->next_input_byte = buf_vec->buffers[buf_vec->pos].data;
+  cinfo->src->bytes_in_buffer = buf_vec->buffers[buf_vec->pos].len;
+  ++buf_vec->pos;
+  return TRUE;
+}
+
+void skip_input_data(j_decompress_ptr cinfo,
+                     long num_bytes) {  // NOLINT
+  cinfo->src->next_input_byte += num_bytes;
+}
+
+void term_source(j_decompress_ptr cinfo) {
+  // Nothing to do.
+}
+
+#ifdef HAVE_SETJMP
+void ErrorHandler(j_common_ptr cinfo) {
+  // This is called when a jpeglib command experiences an error. Unfortunately
+  // jpeglib's error handling model is not very flexible, because it expects the
+  // error handler to not return--i.e., it wants the program to terminate. To
+  // recover from errors we use setjmp() as shown in their example. setjmp() is
+  // C's implementation for the "call with current continuation" functionality
+  // seen in some functional programming languages.
+  // A formatted message can be output, but is unsafe for release.
+#ifdef DEBUG
+  char buf[JMSG_LENGTH_MAX];
+  (*cinfo->err->format_message)(cinfo, buf);
+  // ERROR: Error in jpeglib: buf
+#endif
+
+  SetJmpErrorMgr* mgr = reinterpret_cast<SetJmpErrorMgr*>(cinfo->err);
+  // This rewinds the call stack to the point of the corresponding setjmp()
+  // and causes it to return (for a second time) with value 1.
+  longjmp(mgr->setjmp_buffer, 1);
+}
+#endif
+
+void MJpegDecoder::AllocOutputBuffers(int num_outbufs) {
+  if (num_outbufs != num_outbufs_) {
+    // We could perhaps optimize this case to resize the output buffers without
+    // necessarily having to delete and recreate each one, but it's not worth
+    // it.
+    DestroyOutputBuffers();
+
+    scanlines_ = new uint8** [num_outbufs];
+    scanlines_sizes_ = new int[num_outbufs];
+    databuf_ = new uint8* [num_outbufs];
+    databuf_strides_ = new int[num_outbufs];
+
+    for (int i = 0; i < num_outbufs; ++i) {
+      scanlines_[i] = NULL;
+      scanlines_sizes_[i] = 0;
+      databuf_[i] = NULL;
+      databuf_strides_[i] = 0;
+    }
+
+    num_outbufs_ = num_outbufs;
+  }
+}
+
+void MJpegDecoder::DestroyOutputBuffers() {
+  for (int i = 0; i < num_outbufs_; ++i) {
+    delete [] scanlines_[i];
+    delete [] databuf_[i];
+  }
+  delete [] scanlines_;
+  delete [] databuf_;
+  delete [] scanlines_sizes_;
+  delete [] databuf_strides_;
+  scanlines_ = NULL;
+  databuf_ = NULL;
+  scanlines_sizes_ = NULL;
+  databuf_strides_ = NULL;
+  num_outbufs_ = 0;
+}
+
+// JDCT_IFAST and do_block_smoothing improve performance substantially.
+LIBYUV_BOOL MJpegDecoder::StartDecode() {
+  decompress_struct_->raw_data_out = TRUE;
+  decompress_struct_->dct_method = JDCT_IFAST;  // JDCT_ISLOW is default
+  decompress_struct_->dither_mode = JDITHER_NONE;
+  // Not applicable to 'raw':
+  decompress_struct_->do_fancy_upsampling = (boolean)(LIBYUV_FALSE);
+  // Only for buffered mode:
+  decompress_struct_->enable_2pass_quant = (boolean)(LIBYUV_FALSE);
+  // Blocky but fast:
+  decompress_struct_->do_block_smoothing = (boolean)(LIBYUV_FALSE);
+
+  if (!jpeg_start_decompress(decompress_struct_)) {
+    // ERROR: Couldn't start JPEG decompressor";
+    return LIBYUV_FALSE;
+  }
+  return LIBYUV_TRUE;
+}
+
+LIBYUV_BOOL MJpegDecoder::FinishDecode() {
+  // jpeglib considers it an error if we finish without decoding the whole
+  // image, so we call "abort" rather than "finish".
+  jpeg_abort_decompress(decompress_struct_);
+  return LIBYUV_TRUE;
+}
+
+void MJpegDecoder::SetScanlinePointers(uint8** data) {
+  for (int i = 0; i < num_outbufs_; ++i) {
+    uint8* data_i = data[i];
+    for (int j = 0; j < scanlines_sizes_[i]; ++j) {
+      scanlines_[i][j] = data_i;
+      data_i += GetComponentStride(i);
+    }
+  }
+}
+
+inline LIBYUV_BOOL MJpegDecoder::DecodeImcuRow() {
+  return (unsigned int)(GetImageScanlinesPerImcuRow()) ==
+      jpeg_read_raw_data(decompress_struct_,
+                         scanlines_,
+                         GetImageScanlinesPerImcuRow());
+}
+
+// The helper function which recognizes the jpeg sub-sampling type.
+JpegSubsamplingType MJpegDecoder::JpegSubsamplingTypeHelper(
+    int* subsample_x, int* subsample_y, int number_of_components) {
+  if (number_of_components == 3) {  // Color images.
+    if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 2 && subsample_y[1] == 2 &&
+        subsample_x[2] == 2 && subsample_y[2] == 2) {
+      return kJpegYuv420;
+    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 2 && subsample_y[1] == 1 &&
+        subsample_x[2] == 2 && subsample_y[2] == 1) {
+      return kJpegYuv422;
+    } else if (subsample_x[0] == 1 && subsample_y[0] == 1 &&
+        subsample_x[1] == 1 && subsample_y[1] == 1 &&
+        subsample_x[2] == 1 && subsample_y[2] == 1) {
+      return kJpegYuv444;
+    }
+  } else if (number_of_components == 1) {  // Grey-scale images.
+    if (subsample_x[0] == 1 && subsample_y[0] == 1) {
+      return kJpegYuv400;
+    }
+  }
+  return kJpegUnknown;
+}
+
+}  // namespace libyuv
+#endif  // HAVE_JPEG
+
diff --git a/third_party/aom/third_party/libyuv/source/mjpeg_validate.cc b/third_party/aom/third_party/libyuv/source/mjpeg_validate.cc
new file mode 100644
index 0000000000..8edfbe1e74
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/mjpeg_validate.cc
@@ -0,0 +1,101 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/mjpeg_decoder.h"
+
+#include <string.h>  // For memchr.
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Enable this to try scasb implementation.
+// #define ENABLE_SCASB 1
+
+#ifdef ENABLE_SCASB
+
+// Multiple of 1.
+__declspec(naked)
+const uint8* ScanRow_ERMS(const uint8* src, uint32 val, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // src
+    mov        eax, [esp + 8]   // val
+    mov        ecx, [esp + 12]  // count
+    repne scasb
+    jne        sr99
+    mov        eax, edi
+    sub        eax, 1
+    mov        edi, edx
+    ret
+
+  sr99:
+    mov        eax, 0
+    mov        edi, edx
+    ret
+  }
+}
+#endif
+
+// Helper function to scan for EOI marker.
+static LIBYUV_BOOL ScanEOI(const uint8* sample, size_t sample_size) {
+  const uint8* end = sample + sample_size - 1;
+  const uint8* it = sample;
+  for (;;) {
+#ifdef ENABLE_SCASB
+    it = ScanRow_ERMS(it, 0xff, end - it);
+#else
+    it = static_cast<const uint8*>(memchr(it, 0xff, end - it));
+#endif
+    if (it == NULL) {
+      break;
+    }
+    if (it[1] == 0xd9) {
+      return LIBYUV_TRUE;  // Success: Valid jpeg.
+    }
+    ++it;  // Skip over current 0xff.
+  }
+  // ERROR: Invalid jpeg end code not found. Size sample_size
+  return LIBYUV_FALSE;
+}
+
+// Helper function to validate the jpeg appears intact.
+LIBYUV_BOOL ValidateJpeg(const uint8* sample, size_t sample_size) {
+  const size_t kBackSearchSize = 1024;
+  if (sample_size < 64) {
+    // ERROR: Invalid jpeg size: sample_size
+    return LIBYUV_FALSE;
+  }
+  if (sample[0] != 0xff || sample[1] != 0xd8) {  // Start Of Image
+    // ERROR: Invalid jpeg initial start code
+    return LIBYUV_FALSE;
+  }
+  // Step over SOI marker.
+  sample += 2;
+  sample_size -= 2;
+
+  // Look for the End Of Image (EOI) marker in the end kilobyte of the buffer.
+  if (sample_size > kBackSearchSize) {
+    if (ScanEOI(sample + sample_size - kBackSearchSize, kBackSearchSize)) {
+      return LIBYUV_TRUE;  // Success: Valid jpeg.
+    }
+    // Reduce search size for forward search.
+    sample_size = sample_size - kBackSearchSize + 1;
+  }
+  return ScanEOI(sample, sample_size);
+
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/third_party/aom/third_party/libyuv/source/planar_functions.cc b/third_party/aom/third_party/libyuv/source/planar_functions.cc
new file mode 100644
index 0000000000..b96bd50206
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/planar_functions.cc
@@ -0,0 +1,2555 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/planar_functions.h"
+
+#include <string.h>  // for memset()
+
+#include "libyuv/cpu_id.h"
+#ifdef HAVE_JPEG
+#include "libyuv/mjpeg_decoder.h"
+#endif
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Copy a plane of data
+LIBYUV_API
+void CopyPlane(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  int y;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+  // Nothing to do.
+  if (src_y == dst_y && src_stride_y == dst_stride_y) {
+    return;
+  }
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height; ++y) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+LIBYUV_API
+void CopyPlane_16(const uint16* src_y, int src_stride_y,
+                  uint16* dst_y, int dst_stride_y,
+                  int width, int height) {
+  int y;
+  void (*CopyRow)(const uint16* src, uint16* dst, int width) = CopyRow_16_C;
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_y = 0;
+  }
+#if defined(HAS_COPYROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_16_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_16_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_16_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 32)) {
+    CopyRow = CopyRow_16_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_16_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_16_MIPS;
+  }
+#endif
+
+  // Copy plane
+  for (y = 0; y < height; ++y) {
+    CopyRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Copy I422.
+LIBYUV_API
+int I422Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, height);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, height);
+  return 0;
+}
+
+// Copy I444.
+LIBYUV_API
+int I444Copy(const uint8* src_y, int src_stride_y,
+             const uint8* src_u, int src_stride_u,
+             const uint8* src_v, int src_stride_v,
+             uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int width, int height) {
+  if (!src_y || !src_u || !src_v ||
+      !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (height - 1) * src_stride_u;
+    src_v = src_v + (height - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  CopyPlane(src_u, src_stride_u, dst_u, dst_stride_u, width, height);
+  CopyPlane(src_v, src_stride_v, dst_v, dst_stride_v, width, height);
+  return 0;
+}
+
+// Copy I400.
+LIBYUV_API
+int I400ToI400(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Convert I420 to I400.
+LIBYUV_API
+int I420ToI400(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  CopyPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Mirror a plane of data.
+void MirrorPlane(const uint8* src_y, int src_stride_y,
+                 uint8* dst_y, int dst_stride_y,
+                 int width, int height) {
+  int y;
+  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorRow = MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MirrorRow = MirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorRow = MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorRow = MirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Mirror on mips handle unaligned memory.
+#if defined(HAS_MIRRORROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(dst_y, 4) && IS_ALIGNED(dst_stride_y, 4)) {
+    MirrorRow = MirrorRow_MIPS_DSPR2;
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    MirrorRow(src_y, dst_y, width);
+    src_y += src_stride_y;
+    dst_y += dst_stride_y;
+  }
+}
+
+// Convert YUY2 to I422.
+LIBYUV_API
+int YUY2ToI422(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*YUY2ToUV422Row)(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) =
+      YUY2ToUV422Row_C;
+  void (*YUY2ToYRow)(const uint8* src_yuy2, uint8* dst_y, int pix) =
+      YUY2ToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+  // Coalesce rows.
+  if (src_stride_yuy2 == width * 2 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_yuy2 = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_YUY2TOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_SSE2;
+    YUY2ToYRow = YUY2ToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_SSE2;
+      YUY2ToYRow = YUY2ToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    YUY2ToUV422Row = YUY2ToUV422Row_Any_AVX2;
+    YUY2ToYRow = YUY2ToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      YUY2ToUV422Row = YUY2ToUV422Row_AVX2;
+      YUY2ToYRow = YUY2ToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_YUY2TOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    YUY2ToYRow = YUY2ToYRow_Any_NEON;
+    if (width >= 16) {
+      YUY2ToUV422Row = YUY2ToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      YUY2ToYRow = YUY2ToYRow_NEON;
+      YUY2ToUV422Row = YUY2ToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    YUY2ToUV422Row(src_yuy2, dst_u, dst_v, width);
+    YUY2ToYRow(src_yuy2, dst_y, width);
+    src_yuy2 += src_stride_yuy2;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Convert UYVY to I422.
+LIBYUV_API
+int UYVYToI422(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int y;
+  void (*UYVYToUV422Row)(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) =
+      UYVYToUV422Row_C;
+  void (*UYVYToYRow)(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) = UYVYToYRow_C;
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+  // Coalesce rows.
+  if (src_stride_uyvy == width * 2 &&
+      dst_stride_y == width &&
+      dst_stride_u * 2 == width &&
+      dst_stride_v * 2 == width) {
+    width *= height;
+    height = 1;
+    src_stride_uyvy = dst_stride_y = dst_stride_u = dst_stride_v = 0;
+  }
+#if defined(HAS_UYVYTOYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    UYVYToUV422Row = UYVYToUV422Row_Any_SSE2;
+    UYVYToYRow = UYVYToYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToUV422Row = UYVYToUV422Row_SSE2;
+      UYVYToYRow = UYVYToYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    UYVYToUV422Row = UYVYToUV422Row_Any_AVX2;
+    UYVYToYRow = UYVYToYRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      UYVYToUV422Row = UYVYToUV422Row_AVX2;
+      UYVYToYRow = UYVYToYRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_UYVYTOYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    UYVYToYRow = UYVYToYRow_Any_NEON;
+    if (width >= 16) {
+      UYVYToUV422Row = UYVYToUV422Row_Any_NEON;
+    }
+    if (IS_ALIGNED(width, 16)) {
+      UYVYToYRow = UYVYToYRow_NEON;
+      UYVYToUV422Row = UYVYToUV422Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    UYVYToUV422Row(src_uyvy, dst_u, dst_v, width);
+    UYVYToYRow(src_uyvy, dst_y, width);
+    src_uyvy += src_stride_uyvy;
+    dst_y += dst_stride_y;
+    dst_u += dst_stride_u;
+    dst_v += dst_stride_v;
+  }
+  return 0;
+}
+
+// Mirror I400 with optional flipping
+LIBYUV_API
+int I400Mirror(const uint8* src_y, int src_stride_y,
+               uint8* dst_y, int dst_stride_y,
+               int width, int height) {
+  if (!src_y || !dst_y ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+
+  MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  return 0;
+}
+
+// Mirror I420 with optional flipping
+LIBYUV_API
+int I420Mirror(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || !dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  if (dst_y) {
+    MirrorPlane(src_y, src_stride_y, dst_y, dst_stride_y, width, height);
+  }
+  MirrorPlane(src_u, src_stride_u, dst_u, dst_stride_u, halfwidth, halfheight);
+  MirrorPlane(src_v, src_stride_v, dst_v, dst_stride_v, halfwidth, halfheight);
+  return 0;
+}
+
+// ARGB mirror.
+LIBYUV_API
+int ARGBMirror(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+      ARGBMirrorRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMirrorRow = ARGBMirrorRow_AVX2;
+    }
+  }
+#endif
+
+  // Mirror plane
+  for (y = 0; y < height; ++y) {
+    ARGBMirrorRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Get a blender that optimized for the CPU and pixel count.
+// As there are 6 blenders to choose from, the caller should try to use
+// the same blend function for all pixels if possible.
+LIBYUV_API
+ARGBBlendRow GetARGBBlend() {
+  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width) = ARGBBlendRow_C;
+#if defined(HAS_ARGBBLENDROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBBlendRow = ARGBBlendRow_SSSE3;
+    return ARGBBlendRow;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBBlendRow = ARGBBlendRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBBLENDROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBBlendRow = ARGBBlendRow_NEON;
+  }
+#endif
+  return ARGBBlendRow;
+}
+
+// Alpha Blend 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
+              const uint8* src_argb1, int src_stride_argb1,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  int y;
+  void (*ARGBBlendRow)(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width) = GetARGBBlend();
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+
+  for (y = 0; y < height; ++y) {
+    ARGBBlendRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Multiply 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*ARGBMultiplyRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBMultiplyRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBMULTIPLYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMULTIPLYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBMultiplyRow = ARGBMultiplyRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMultiplyRow = ARGBMultiplyRow_NEON;
+    }
+  }
+#endif
+
+  // Multiply plane
+  for (y = 0; y < height; ++y) {
+    ARGBMultiplyRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Add 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
+            const uint8* src_argb1, int src_stride_argb1,
+            uint8* dst_argb, int dst_stride_argb,
+            int width, int height) {
+  int y;
+  void (*ARGBAddRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                     int width) = ARGBAddRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBADDROW_SSE2) && (defined(_MSC_VER) && !defined(__clang__))
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBAddRow = ARGBAddRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBADDROW_SSE2) && !(defined(_MSC_VER) && !defined(__clang__))
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBAddRow = ARGBAddRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAddRow = ARGBAddRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAddRow = ARGBAddRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBADDROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAddRow = ARGBAddRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAddRow = ARGBAddRow_NEON;
+    }
+  }
+#endif
+
+  // Add plane
+  for (y = 0; y < height; ++y) {
+    ARGBAddRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Subtract 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  int y;
+  void (*ARGBSubtractRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBSubtractRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSUBTRACTROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSubtractRow = ARGBSubtractRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSUBTRACTROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSUBTRACTROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_NEON;
+    }
+  }
+#endif
+
+  // Subtract plane
+  for (y = 0; y < height; ++y) {
+    ARGBSubtractRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert I422 to BGRA.
+LIBYUV_API
+int I422ToBGRA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_bgra, int dst_stride_bgra,
+               int width, int height) {
+  int y;
+  void (*I422ToBGRARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToBGRARow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_bgra ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_bgra = dst_bgra + (height - 1) * dst_stride_bgra;
+    dst_stride_bgra = -dst_stride_bgra;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_bgra == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_bgra = 0;
+  }
+#if defined(HAS_I422TOBGRAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToBGRARow = I422ToBGRARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToBGRARow = I422ToBGRARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToBGRARow = I422ToBGRARow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToBGRARow = I422ToBGRARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToBGRARow = I422ToBGRARow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOBGRAROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_bgra, 4) && IS_ALIGNED(dst_stride_bgra, 4)) {
+    I422ToBGRARow = I422ToBGRARow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToBGRARow(src_y, src_u, src_v, dst_bgra, width);
+    dst_bgra += dst_stride_bgra;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to ABGR.
+LIBYUV_API
+int I422ToABGR(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_abgr, int dst_stride_abgr,
+               int width, int height) {
+  int y;
+  void (*I422ToABGRRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToABGRRow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_abgr ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_abgr = dst_abgr + (height - 1) * dst_stride_abgr;
+    dst_stride_abgr = -dst_stride_abgr;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_abgr == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_abgr = 0;
+  }
+#if defined(HAS_I422TOABGRROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToABGRRow = I422ToABGRRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOABGRROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToABGRRow = I422ToABGRRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToABGRRow = I422ToABGRRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOABGRROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToABGRRow = I422ToABGRRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToABGRRow = I422ToABGRRow_AVX2;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToABGRRow(src_y, src_u, src_v, dst_abgr, width);
+    dst_abgr += dst_stride_abgr;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert I422 to RGBA.
+LIBYUV_API
+int I422ToRGBA(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_rgba, int dst_stride_rgba,
+               int width, int height) {
+  int y;
+  void (*I422ToRGBARow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToRGBARow_C;
+  if (!src_y || !src_u || !src_v ||
+      !dst_rgba ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgba = dst_rgba + (height - 1) * dst_stride_rgba;
+    dst_stride_rgba = -dst_stride_rgba;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      src_stride_u * 2 == width &&
+      src_stride_v * 2 == width &&
+      dst_stride_rgba == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = src_stride_u = src_stride_v = dst_stride_rgba = 0;
+  }
+#if defined(HAS_I422TORGBAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    I422ToRGBARow = I422ToRGBARow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToRGBARow = I422ToRGBARow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      I422ToRGBARow = I422ToRGBARow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TORGBAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToRGBARow = I422ToRGBARow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      I422ToRGBARow = I422ToRGBARow_AVX2;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    I422ToRGBARow(src_y, src_u, src_v, dst_rgba, width);
+    dst_rgba += dst_stride_rgba;
+    src_y += src_stride_y;
+    src_u += src_stride_u;
+    src_v += src_stride_v;
+  }
+  return 0;
+}
+
+// Convert NV12 to RGB565.
+LIBYUV_API
+int NV12ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_uv, int src_stride_uv,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*NV12ToRGB565Row)(const uint8* y_buf,
+                          const uint8* uv_buf,
+                          uint8* rgb_buf,
+                          int width) = NV12ToRGB565Row_C;
+  if (!src_y || !src_uv || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV12TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV12TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV12ToRGB565Row(src_y, src_uv, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_uv += src_stride_uv;
+    }
+  }
+  return 0;
+}
+
+// Convert NV21 to RGB565.
+LIBYUV_API
+int NV21ToRGB565(const uint8* src_y, int src_stride_y,
+                 const uint8* src_vu, int src_stride_vu,
+                 uint8* dst_rgb565, int dst_stride_rgb565,
+                 int width, int height) {
+  int y;
+  void (*NV21ToRGB565Row)(const uint8* y_buf,
+                          const uint8* src_vu,
+                          uint8* rgb_buf,
+                          int width) = NV21ToRGB565Row_C;
+  if (!src_y || !src_vu || !dst_rgb565 ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_rgb565 = dst_rgb565 + (height - 1) * dst_stride_rgb565;
+    dst_stride_rgb565 = -dst_stride_rgb565;
+  }
+#if defined(HAS_NV21TORGB565ROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_NV21TORGB565ROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_NV21TORGB565ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    NV21ToRGB565Row = NV21ToRGB565Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToRGB565Row = NV21ToRGB565Row_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    NV21ToRGB565Row(src_y, src_vu, dst_rgb565, width);
+    dst_rgb565 += dst_stride_rgb565;
+    src_y += src_stride_y;
+    if (y & 1) {
+      src_vu += src_stride_vu;
+    }
+  }
+  return 0;
+}
+
+LIBYUV_API
+void SetPlane(uint8* dst_y, int dst_stride_y,
+              int width, int height,
+              uint32 value) {
+  int y;
+  void (*SetRow)(uint8* dst, uint8 value, int pix) = SetRow_C;
+  if (height < 0) {
+    height = -height;
+    dst_y = dst_y + (height - 1) * dst_stride_y;
+    dst_stride_y = -dst_stride_y;
+  }
+  // Coalesce rows.
+  if (dst_stride_y == width) {
+    width *= height;
+    height = 1;
+    dst_stride_y = 0;
+  }
+#if defined(HAS_SETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SetRow = SetRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SetRow = SetRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SETROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    SetRow = SetRow_Any_X86;
+    if (IS_ALIGNED(width, 4)) {
+      SetRow = SetRow_X86;
+    }
+  }
+#endif
+#if defined(HAS_SETROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    SetRow = SetRow_ERMS;
+  }
+#endif
+
+  // Set plane
+  for (y = 0; y < height; ++y) {
+    SetRow(dst_y, value, width);
+    dst_y += dst_stride_y;
+  }
+}
+
+// Draw a rectangle into I420
+LIBYUV_API
+int I420Rect(uint8* dst_y, int dst_stride_y,
+             uint8* dst_u, int dst_stride_u,
+             uint8* dst_v, int dst_stride_v,
+             int x, int y,
+             int width, int height,
+             int value_y, int value_u, int value_v) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  uint8* start_y = dst_y + y * dst_stride_y + x;
+  uint8* start_u = dst_u + (y / 2) * dst_stride_u + (x / 2);
+  uint8* start_v = dst_v + (y / 2) * dst_stride_v + (x / 2);
+  if (!dst_y || !dst_u || !dst_v ||
+      width <= 0 || height == 0 ||
+      x < 0 || y < 0 ||
+      value_y < 0 || value_y > 255 ||
+      value_u < 0 || value_u > 255 ||
+      value_v < 0 || value_v > 255) {
+    return -1;
+  }
+
+  SetPlane(start_y, dst_stride_y, width, height, value_y);
+  SetPlane(start_u, dst_stride_u, halfwidth, halfheight, value_u);
+  SetPlane(start_v, dst_stride_v, halfwidth, halfheight, value_v);
+  return 0;
+}
+
+// Draw a rectangle into ARGB
+LIBYUV_API
+int ARGBRect(uint8* dst_argb, int dst_stride_argb,
+             int dst_x, int dst_y,
+             int width, int height,
+             uint32 value) {
+  int y;
+  void (*ARGBSetRow)(uint8* dst_argb, uint32 value, int pix) = ARGBSetRow_C;
+  if (!dst_argb ||
+      width <= 0 || height == 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  dst_argb += dst_y * dst_stride_argb + dst_x * 4;
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+
+#if defined(HAS_ARGBSETROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBSetRow = ARGBSetRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSetRow = ARGBSetRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSETROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    ARGBSetRow = ARGBSetRow_X86;
+  }
+#endif
+
+  // Set plane
+  for (y = 0; y < height; ++y) {
+    ARGBSetRow(dst_argb, value, width);
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert unattentuated ARGB to preattenuated ARGB.
+// An unattenutated ARGB alpha blend uses the formula
+// p = a * f + (1 - a) * b
+// where
+//   p is output pixel
+//   f is foreground pixel
+//   b is background pixel
+//   a is alpha value from foreground pixel
+// An preattenutated ARGB alpha blend uses the formula
+// p = f + (1 - a) * b
+// where
+//   f is foreground pixel premultiplied by alpha
+
+LIBYUV_API
+int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  int y;
+  void (*ARGBAttenuateRow)(const uint8* src_argb, uint8* dst_argb,
+                           int width) = ARGBAttenuateRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBATTENUATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBATTENUATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBAttenuateRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert preattentuated ARGB to unattenuated ARGB.
+LIBYUV_API
+int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height) {
+  int y;
+  void (*ARGBUnattenuateRow)(const uint8* src_argb, uint8* dst_argb,
+                             int width) = ARGBUnattenuateRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBUNATTENUATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBUNATTENUATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Neon version.
+
+  for (y = 0; y < height; ++y) {
+    ARGBUnattenuateRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Convert ARGB to Grayed ARGB.
+LIBYUV_API
+int ARGBGrayTo(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb,
+               int width, int height) {
+  int y;
+  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
+                      int width) = ARGBGrayRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBGRAYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBGrayRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Make a rectangle of ARGB gray scale.
+LIBYUV_API
+int ARGBGray(uint8* dst_argb, int dst_stride_argb,
+             int dst_x, int dst_y,
+             int width, int height) {
+  int y;
+  void (*ARGBGrayRow)(const uint8* src_argb, uint8* dst_argb,
+                      int width) = ARGBGrayRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBGRAYROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBGRAYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBGrayRow = ARGBGrayRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBGrayRow(dst, dst, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Make a rectangle of ARGB Sepia tone.
+LIBYUV_API
+int ARGBSepia(uint8* dst_argb, int dst_stride_argb,
+              int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBSepiaRow)(uint8* dst_argb, int width) = ARGBSepiaRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSEPIAROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
+    ARGBSepiaRow = ARGBSepiaRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBSEPIAROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBSepiaRow = ARGBSepiaRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBSepiaRow(dst, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a 4x4 matrix to each ARGB pixel.
+// Note: Normally for shading, but can be used to swizzle or invert.
+LIBYUV_API
+int ARGBColorMatrix(const uint8* src_argb, int src_stride_argb,
+                    uint8* dst_argb, int dst_stride_argb,
+                    const int8* matrix_argb,
+                    int width, int height) {
+  int y;
+  void (*ARGBColorMatrixRow)(const uint8* src_argb, uint8* dst_argb,
+      const int8* matrix_argb, int width) = ARGBColorMatrixRow_C;
+  if (!src_argb || !dst_argb || !matrix_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOLORMATRIXROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 8)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_SSSE3;
+  }
+#endif
+#if defined(HAS_ARGBCOLORMATRIXROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBColorMatrixRow = ARGBColorMatrixRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBColorMatrixRow(src_argb, dst_argb, matrix_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a 4x3 matrix to each ARGB pixel.
+// Deprecated.
+LIBYUV_API
+int RGBColorMatrix(uint8* dst_argb, int dst_stride_argb,
+                   const int8* matrix_rgb,
+                   int dst_x, int dst_y, int width, int height) {
+  SIMD_ALIGNED(int8 matrix_argb[16]);
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !matrix_rgb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+
+  // Convert 4x3 7 bit matrix to 4x4 6 bit matrix.
+  matrix_argb[0] = matrix_rgb[0] / 2;
+  matrix_argb[1] = matrix_rgb[1] / 2;
+  matrix_argb[2] = matrix_rgb[2] / 2;
+  matrix_argb[3] = matrix_rgb[3] / 2;
+  matrix_argb[4] = matrix_rgb[4] / 2;
+  matrix_argb[5] = matrix_rgb[5] / 2;
+  matrix_argb[6] = matrix_rgb[6] / 2;
+  matrix_argb[7] = matrix_rgb[7] / 2;
+  matrix_argb[8] = matrix_rgb[8] / 2;
+  matrix_argb[9] = matrix_rgb[9] / 2;
+  matrix_argb[10] = matrix_rgb[10] / 2;
+  matrix_argb[11] = matrix_rgb[11] / 2;
+  matrix_argb[14] = matrix_argb[13] = matrix_argb[12] = 0;
+  matrix_argb[15] = 64;  // 1.0
+
+  return ARGBColorMatrix((const uint8*)(dst), dst_stride_argb,
+                         dst, dst_stride_argb,
+                         &matrix_argb[0], width, height);
+}
+
+// Apply a color table each ARGB pixel.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int ARGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                   const uint8* table_argb,
+                   int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+                            int width) = ARGBColorTableRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOLORTABLEROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    ARGBColorTableRow = ARGBColorTableRow_X86;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBColorTableRow(dst, table_argb, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a color table each ARGB pixel but preserve destination alpha.
+// Table contains 256 ARGB values.
+LIBYUV_API
+int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
+                  const uint8* table_argb,
+                  int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*RGBColorTableRow)(uint8* dst_argb, const uint8* table_argb,
+                           int width) = RGBColorTableRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || !table_argb || width <= 0 || height <= 0 ||
+      dst_x < 0 || dst_y < 0) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_RGBCOLORTABLEROW_X86)
+  if (TestCpuFlag(kCpuHasX86)) {
+    RGBColorTableRow = RGBColorTableRow_X86;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    RGBColorTableRow(dst, table_argb, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// ARGBQuantize is used to posterize art.
+// e.g. rgb / qvalue * qvalue + qvalue / 2
+// But the low levels implement efficiently with 3 parameters, and could be
+// used for other high level operations.
+// dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
+// where scale is 1 / interval_size as a fixed point value.
+// The divide is replaces with a multiply by reciprocal fixed point multiply.
+// Caveat - although SSE2 saturates, the C function does not and should be used
+// with care if doing anything but quantization.
+LIBYUV_API
+int ARGBQuantize(uint8* dst_argb, int dst_stride_argb,
+                 int scale, int interval_size, int interval_offset,
+                 int dst_x, int dst_y, int width, int height) {
+  int y;
+  void (*ARGBQuantizeRow)(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) = ARGBQuantizeRow_C;
+  uint8* dst = dst_argb + dst_y * dst_stride_argb + dst_x * 4;
+  if (!dst_argb || width <= 0 || height <= 0 || dst_x < 0 || dst_y < 0 ||
+      interval_size < 1 || interval_size > 255) {
+    return -1;
+  }
+  // Coalesce rows.
+  if (dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBQUANTIZEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBQUANTIZEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBQuantizeRow = ARGBQuantizeRow_NEON;
+  }
+#endif
+  for (y = 0; y < height; ++y) {
+    ARGBQuantizeRow(dst, scale, interval_size, interval_offset, width);
+    dst += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Computes table of cumulative sum for image where the value is the sum
+// of all values above and to the left of the entry. Used by ARGBBlur.
+LIBYUV_API
+int ARGBComputeCumulativeSum(const uint8* src_argb, int src_stride_argb,
+                             int32* dst_cumsum, int dst_stride32_cumsum,
+                             int width, int height) {
+  int y;
+  void (*ComputeCumulativeSumRow)(const uint8* row, int32* cumsum,
+      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
+  int32* previous_cumsum = dst_cumsum;
+  if (!dst_cumsum || !src_argb || width <= 0 || height <= 0) {
+    return -1;
+  }
+#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
+  }
+#endif
+  memset(dst_cumsum, 0, width * sizeof(dst_cumsum[0]) * 4);  // 4 int per pixel.
+  for (y = 0; y < height; ++y) {
+    ComputeCumulativeSumRow(src_argb, dst_cumsum, previous_cumsum, width);
+    previous_cumsum = dst_cumsum;
+    dst_cumsum += dst_stride32_cumsum;
+    src_argb += src_stride_argb;
+  }
+  return 0;
+}
+
+// Blur ARGB image.
+// Caller should allocate CumulativeSum table of width * height * 16 bytes
+// aligned to 16 byte boundary. height can be radius * 2 + 2 to save memory
+// as the buffer is treated as circular.
+LIBYUV_API
+int ARGBBlur(const uint8* src_argb, int src_stride_argb,
+             uint8* dst_argb, int dst_stride_argb,
+             int32* dst_cumsum, int dst_stride32_cumsum,
+             int width, int height, int radius) {
+  int y;
+  void (*ComputeCumulativeSumRow)(const uint8 *row, int32 *cumsum,
+      const int32* previous_cumsum, int width) = ComputeCumulativeSumRow_C;
+  void (*CumulativeSumToAverageRow)(const int32* topleft, const int32* botleft,
+      int width, int area, uint8* dst, int count) = CumulativeSumToAverageRow_C;
+  int32* cumsum_bot_row;
+  int32* max_cumsum_bot_row;
+  int32* cumsum_top_row;
+
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  if (radius > height) {
+    radius = height;
+  }
+  if (radius > (width / 2 - 1)) {
+    radius = width / 2 - 1;
+  }
+  if (radius <= 0) {
+    return -1;
+  }
+#if defined(HAS_CUMULATIVESUMTOAVERAGEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ComputeCumulativeSumRow = ComputeCumulativeSumRow_SSE2;
+    CumulativeSumToAverageRow = CumulativeSumToAverageRow_SSE2;
+  }
+#endif
+  // Compute enough CumulativeSum for first row to be blurred. After this
+  // one row of CumulativeSum is updated at a time.
+  ARGBComputeCumulativeSum(src_argb, src_stride_argb,
+                           dst_cumsum, dst_stride32_cumsum,
+                           width, radius);
+
+  src_argb = src_argb + radius * src_stride_argb;
+  cumsum_bot_row = &dst_cumsum[(radius - 1) * dst_stride32_cumsum];
+
+  max_cumsum_bot_row = &dst_cumsum[(radius * 2 + 2) * dst_stride32_cumsum];
+  cumsum_top_row = &dst_cumsum[0];
+
+  for (y = 0; y < height; ++y) {
+    int top_y = ((y - radius - 1) >= 0) ? (y - radius - 1) : 0;
+    int bot_y = ((y + radius) < height) ? (y + radius) : (height - 1);
+    int area = radius * (bot_y - top_y);
+    int boxwidth = radius * 4;
+    int x;
+    int n;
+
+    // Increment cumsum_top_row pointer with circular buffer wrap around.
+    if (top_y) {
+      cumsum_top_row += dst_stride32_cumsum;
+      if (cumsum_top_row >= max_cumsum_bot_row) {
+        cumsum_top_row = dst_cumsum;
+      }
+    }
+    // Increment cumsum_bot_row pointer with circular buffer wrap around and
+    // then fill in a row of CumulativeSum.
+    if ((y + radius) < height) {
+      const int32* prev_cumsum_bot_row = cumsum_bot_row;
+      cumsum_bot_row += dst_stride32_cumsum;
+      if (cumsum_bot_row >= max_cumsum_bot_row) {
+        cumsum_bot_row = dst_cumsum;
+      }
+      ComputeCumulativeSumRow(src_argb, cumsum_bot_row, prev_cumsum_bot_row,
+                              width);
+      src_argb += src_stride_argb;
+    }
+
+    // Left clipped.
+    for (x = 0; x < radius + 1; ++x) {
+      CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
+                                boxwidth, area, &dst_argb[x * 4], 1);
+      area += (bot_y - top_y);
+      boxwidth += 4;
+    }
+
+    // Middle unclipped.
+    n = (width - 1) - radius - x + 1;
+    CumulativeSumToAverageRow(cumsum_top_row, cumsum_bot_row,
+                              boxwidth, area, &dst_argb[x * 4], n);
+
+    // Right clipped.
+    for (x += n; x <= width - 1; ++x) {
+      area -= (bot_y - top_y);
+      boxwidth -= 4;
+      CumulativeSumToAverageRow(cumsum_top_row + (x - radius - 1) * 4,
+                                cumsum_bot_row + (x - radius - 1) * 4,
+                                boxwidth, area, &dst_argb[x * 4], 1);
+    }
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Multiply ARGB image by a specified ARGB value.
+LIBYUV_API
+int ARGBShade(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height, uint32 value) {
+  int y;
+  void (*ARGBShadeRow)(const uint8* src_argb, uint8* dst_argb,
+                       int width, uint32 value) = ARGBShadeRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0 || value == 0u) {
+    return -1;
+  }
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSHADEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4)) {
+    ARGBShadeRow = ARGBShadeRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBSHADEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    ARGBShadeRow = ARGBShadeRow_NEON;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBShadeRow(src_argb, dst_argb, width, value);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Interpolate 2 ARGB images by specified amount (0 to 255).
+LIBYUV_API
+int ARGBInterpolate(const uint8* src_argb0, int src_stride_argb0,
+                    const uint8* src_argb1, int src_stride_argb1,
+                    uint8* dst_argb, int dst_stride_argb,
+                    int width, int height, int interpolation) {
+  int y;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb0 == width * 4 &&
+      src_stride_argb1 == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb0 = src_stride_argb1 = dst_stride_argb = 0;
+  }
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_argb0, 4) && IS_ALIGNED(src_stride_argb0, 4) &&
+      IS_ALIGNED(src_argb1, 4) && IS_ALIGNED(src_stride_argb1, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    InterpolateRow(dst_argb, src_argb0, src_argb1 - src_argb0,
+                   width * 4, interpolation);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Shuffle ARGB channel order.  e.g. BGRA to ARGB.
+LIBYUV_API
+int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
+                uint8* dst_argb, int dst_stride_argb,
+                const uint8* shuffler, int width, int height) {
+  int y;
+  void (*ARGBShuffleRow)(const uint8* src_bgra, uint8* dst_argb,
+                         const uint8* shuffler, int pix) = ARGBShuffleRow_C;
+  if (!src_bgra || !dst_argb ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_bgra = src_bgra + (height - 1) * src_stride_bgra;
+    src_stride_bgra = -src_stride_bgra;
+  }
+  // Coalesce rows.
+  if (src_stride_bgra == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_bgra = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBSHUFFLEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBShuffleRow = ARGBShuffleRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBShuffleRow = ARGBShuffleRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_AVX2;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBShuffleRow = ARGBShuffleRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBSHUFFLEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBShuffleRow = ARGBShuffleRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBShuffleRow = ARGBShuffleRow_NEON;
+    }
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBShuffleRow(src_bgra, dst_argb, shuffler, width);
+    src_bgra += src_stride_bgra;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Sobel ARGB effect.
+static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
+                        uint8* dst_argb, int dst_stride_argb,
+                        int width, int height,
+                        void (*SobelRow)(const uint8* src_sobelx,
+                                         const uint8* src_sobely,
+                                         uint8* dst, int width)) {
+  int y;
+  void (*ARGBToYJRow)(const uint8* src_argb, uint8* dst_g, int pix) =
+      ARGBToYJRow_C;
+  void (*SobelYRow)(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) = SobelYRow_C;
+  void (*SobelXRow)(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobely, int width) =
+      SobelXRow_C;
+  const int kEdge = 16;  // Extra pixels at start of row for extrude/align.
+  if (!src_argb  || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+#if defined(HAS_ARGBTOYJROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    ARGBToYJRow = ARGBToYJRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBToYJRow = ARGBToYJRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      ARGBToYJRow = ARGBToYJRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBTOYJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBToYJRow = ARGBToYJRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToYJRow = ARGBToYJRow_NEON;
+    }
+  }
+#endif
+
+#if defined(HAS_SOBELYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelYRow = SobelYRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelYRow = SobelYRow_NEON;
+  }
+#endif
+#if defined(HAS_SOBELXROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelXRow = SobelXRow_SSE2;
+  }
+#endif
+#if defined(HAS_SOBELXROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelXRow = SobelXRow_NEON;
+  }
+#endif
+  {
+    // 3 rows with edges before/after.
+    const int kRowSize = (width + kEdge + 31) & ~31;
+    align_buffer_64(rows, kRowSize * 2 + (kEdge + kRowSize * 3 + kEdge));
+    uint8* row_sobelx = rows;
+    uint8* row_sobely = rows + kRowSize;
+    uint8* row_y = rows + kRowSize * 2;
+
+    // Convert first row.
+    uint8* row_y0 = row_y + kEdge;
+    uint8* row_y1 = row_y0 + kRowSize;
+    uint8* row_y2 = row_y1 + kRowSize;
+    ARGBToYJRow(src_argb, row_y0, width);
+    row_y0[-1] = row_y0[0];
+    memset(row_y0 + width, row_y0[width - 1], 16);  // Extrude 16 for valgrind.
+    ARGBToYJRow(src_argb, row_y1, width);
+    row_y1[-1] = row_y1[0];
+    memset(row_y1 + width, row_y1[width - 1], 16);
+    memset(row_y2 + width, 0, 16);
+
+    for (y = 0; y < height; ++y) {
+      // Convert next row of ARGB to G.
+      if (y < (height - 1)) {
+        src_argb += src_stride_argb;
+      }
+      ARGBToYJRow(src_argb, row_y2, width);
+      row_y2[-1] = row_y2[0];
+      row_y2[width] = row_y2[width - 1];
+
+      SobelXRow(row_y0 - 1, row_y1 - 1, row_y2 - 1, row_sobelx, width);
+      SobelYRow(row_y0 - 1, row_y2 - 1, row_sobely, width);
+      SobelRow(row_sobelx, row_sobely, dst_argb, width);
+
+      // Cycle thru circular queue of 3 row_y buffers.
+      {
+        uint8* row_yt = row_y0;
+        row_y0 = row_y1;
+        row_y1 = row_y2;
+        row_y2 = row_yt;
+      }
+
+      dst_argb += dst_stride_argb;
+    }
+    free_aligned_buffer_64(rows);
+  }
+  return 0;
+}
+
+// Sobel ARGB effect.
+LIBYUV_API
+int ARGBSobel(const uint8* src_argb, int src_stride_argb,
+              uint8* dst_argb, int dst_stride_argb,
+              int width, int height) {
+  void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) = SobelRow_C;
+#if defined(HAS_SOBELROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelRow = SobelRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SobelRow = SobelRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SOBELROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelRow = SobelRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      SobelRow = SobelRow_NEON;
+    }
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height, SobelRow);
+}
+
+// Sobel ARGB effect with planar output.
+LIBYUV_API
+int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
+                     uint8* dst_y, int dst_stride_y,
+                     int width, int height) {
+  void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_, int width) = SobelToPlaneRow_C;
+#if defined(HAS_SOBELTOPLANEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelToPlaneRow = SobelToPlaneRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SobelToPlaneRow = SobelToPlaneRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SOBELTOPLANEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelToPlaneRow = SobelToPlaneRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SobelToPlaneRow = SobelToPlaneRow_NEON;
+    }
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_y, dst_stride_y,
+                      width, height, SobelToPlaneRow);
+}
+
+// SobelXY ARGB effect.
+// Similar to Sobel, but also stores Sobel X in R and Sobel Y in B.  G = Sobel.
+LIBYUV_API
+int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
+                uint8* dst_argb, int dst_stride_argb,
+                int width, int height) {
+  void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) = SobelXYRow_C;
+#if defined(HAS_SOBELXYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SobelXYRow = SobelXYRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SobelXYRow = SobelXYRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SOBELXYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SobelXYRow = SobelXYRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      SobelXYRow = SobelXYRow_NEON;
+    }
+  }
+#endif
+  return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
+                      width, height, SobelXYRow);
+}
+
+// Apply a 4x4 polynomial to each ARGB pixel.
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height) {
+  int y;
+  void (*ARGBPolynomialRow)(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) = ARGBPolynomialRow_C;
+  if (!src_argb || !dst_argb || !poly || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBPOLYNOMIALROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBPOLYNOMIALROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && TestCpuFlag(kCpuHasFMA3) &&
+      IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBPolynomialRow(src_argb, dst_argb, poly, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Apply a lumacolortable to each ARGB pixel.
+LIBYUV_API
+int ARGBLumaColorTable(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_argb, int dst_stride_argb,
+                       const uint8* luma,
+                       int width, int height) {
+  int y;
+  void (*ARGBLumaColorTableRow)(const uint8* src_argb, uint8* dst_argb,
+      int width, const uint8* luma, const uint32 lumacoeff) =
+      ARGBLumaColorTableRow_C;
+  if (!src_argb || !dst_argb || !luma || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb  = src_argb  + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBLUMACOLORTABLEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 4)) {
+    ARGBLumaColorTableRow = ARGBLumaColorTableRow_SSSE3;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBLumaColorTableRow(src_argb, dst_argb, width, luma, 0x00264b0f);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Copy Alpha from one ARGB image to another.
+LIBYUV_API
+int ARGBCopyAlpha(const uint8* src_argb, int src_stride_argb,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int width, int height) {
+  int y;
+  void (*ARGBCopyAlphaRow)(const uint8* src_argb, uint8* dst_argb, int width) =
+      ARGBCopyAlphaRow_C;
+  if (!src_argb || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+  // Coalesce rows.
+  if (src_stride_argb == width * 4 &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_argb = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOPYALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBCOPYALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
+    ARGBCopyAlphaRow = ARGBCopyAlphaRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBCopyAlphaRow(src_argb, dst_argb, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+// Copy a planar Y channel to the alpha channel of a destination ARGB image.
+LIBYUV_API
+int ARGBCopyYToAlpha(const uint8* src_y, int src_stride_y,
+                     uint8* dst_argb, int dst_stride_argb,
+                     int width, int height) {
+  int y;
+  void (*ARGBCopyYToAlphaRow)(const uint8* src_y, uint8* dst_argb, int width) =
+      ARGBCopyYToAlphaRow_C;
+  if (!src_y || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_stride_y = -src_stride_y;
+  }
+  // Coalesce rows.
+  if (src_stride_y == width &&
+      dst_stride_argb == width * 4) {
+    width *= height;
+    height = 1;
+    src_stride_y = dst_stride_argb = 0;
+  }
+#if defined(HAS_ARGBCOPYYTOALPHAROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
+    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_SSE2;
+  }
+#endif
+#if defined(HAS_ARGBCOPYYTOALPHAROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 16)) {
+    ARGBCopyYToAlphaRow = ARGBCopyYToAlphaRow_AVX2;
+  }
+#endif
+
+  for (y = 0; y < height; ++y) {
+    ARGBCopyYToAlphaRow(src_y, dst_argb, width);
+    src_y += src_stride_y;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
+LIBYUV_API
+int YUY2ToNV12(const uint8* src_yuy2, int src_stride_yuy2,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) =
+      SplitUVRow_C;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  if (!src_yuy2 ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_yuy2 = src_yuy2 + (height - 1) * src_stride_yuy2;
+    src_stride_yuy2 = -src_stride_yuy2;
+  }
+#if defined(HAS_SPLITUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitUVRow = SplitUVRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow = SplitUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitUVRow = SplitUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitUVRow = SplitUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow = SplitUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+
+  {
+    int awidth = halfwidth * 2;
+    // 2 rows of uv
+    align_buffer_64(rows, awidth * 2);
+
+    for (y = 0; y < height - 1; y += 2) {
+      // Split Y from UV.
+      SplitUVRow(src_yuy2, dst_y, rows, awidth);
+      SplitUVRow(src_yuy2 + src_stride_yuy2, dst_y + dst_stride_y,
+                 rows + awidth, awidth);
+      InterpolateRow(dst_uv, rows, awidth, awidth, 128);
+      src_yuy2 += src_stride_yuy2 * 2;
+      dst_y += dst_stride_y * 2;
+      dst_uv += dst_stride_uv;
+    }
+    if (height & 1) {
+      // Split Y from UV.
+      SplitUVRow(src_yuy2, dst_y, dst_uv, width);
+    }
+    free_aligned_buffer_64(rows);
+  }
+  return 0;
+}
+
+LIBYUV_API
+int UYVYToNV12(const uint8* src_uyvy, int src_stride_uyvy,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_uv, int dst_stride_uv,
+               int width, int height) {
+  int y;
+  int halfwidth = (width + 1) >> 1;
+  void (*SplitUVRow)(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) =
+      SplitUVRow_C;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) = InterpolateRow_C;
+  if (!src_uyvy ||
+      !dst_y || !dst_uv ||
+      width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_uyvy = src_uyvy + (height - 1) * src_stride_uyvy;
+    src_stride_uyvy = -src_stride_uyvy;
+  }
+#if defined(HAS_SPLITUVROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    SplitUVRow = SplitUVRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow = SplitUVRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    SplitUVRow = SplitUVRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      SplitUVRow = SplitUVRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SPLITUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    SplitUVRow = SplitUVRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      SplitUVRow = SplitUVRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+
+  {
+    int awidth = halfwidth * 2;
+    // 2 rows of uv
+    align_buffer_64(rows, awidth * 2);
+
+    for (y = 0; y < height - 1; y += 2) {
+      // Split Y from UV.
+      SplitUVRow(src_uyvy, rows, dst_y, awidth);
+      SplitUVRow(src_uyvy + src_stride_uyvy, rows + awidth,
+                 dst_y + dst_stride_y, awidth);
+      InterpolateRow(dst_uv, rows, awidth, awidth, 128);
+      src_uyvy += src_stride_uyvy * 2;
+      dst_y += dst_stride_y * 2;
+      dst_uv += dst_stride_uv;
+    }
+    if (height & 1) {
+      // Split Y from UV.
+      SplitUVRow(src_uyvy, dst_y, dst_uv, width);
+    }
+    free_aligned_buffer_64(rows);
+  }
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/rotate.cc b/third_party/aom/third_party/libyuv/source/rotate.cc
new file mode 100644
index 0000000000..be3d589207
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/rotate.cc
@@ -0,0 +1,496 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/convert.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/rotate_row.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+LIBYUV_API
+void TransposePlane(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  int i = height;
+  void (*TransposeWx8)(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride, int width) = TransposeWx8_C;
+#if defined(HAS_TRANSPOSEWX8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    TransposeWx8 = TransposeWx8_NEON;
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX8_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    TransposeWx8 = TransposeWx8_Any_SSSE3;
+    if (IS_ALIGNED(width, 8)) {
+      TransposeWx8 = TransposeWx8_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX8_FAST_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    TransposeWx8 = TransposeWx8_Fast_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      TransposeWx8 = TransposeWx8_Fast_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_TRANSPOSEWX8_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    if (IS_ALIGNED(width, 4) &&
+        IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+      TransposeWx8 = TransposeWx8_Fast_MIPS_DSPR2;
+    } else {
+      TransposeWx8 = TransposeWx8_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  // Work across the source in 8x8 tiles
+  while (i >= 8) {
+    TransposeWx8(src, src_stride, dst, dst_stride, width);
+    src += 8 * src_stride;    // Go down 8 rows.
+    dst += 8;                 // Move over 8 columns.
+    i -= 8;
+  }
+
+  if (i > 0) {
+    TransposeWxH_C(src, src_stride, dst, dst_stride, width, i);
+  }
+}
+
+LIBYUV_API
+void RotatePlane90(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride,
+                   int width, int height) {
+  // Rotate by 90 is a transpose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  TransposePlane(src, src_stride, dst, dst_stride, width, height);
+}
+
+LIBYUV_API
+void RotatePlane270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  // Rotate by 270 is a transpose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  TransposePlane(src, src_stride, dst, dst_stride, width, height);
+}
+
+LIBYUV_API
+void RotatePlane180(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64(row, width);
+  const uint8* src_bot = src + src_stride * (height - 1);
+  uint8* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+  void (*MirrorRow)(const uint8* src, uint8* dst, int width) = MirrorRow_C;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_MIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    MirrorRow = MirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    MirrorRow = MirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    MirrorRow = MirrorRow_Any_SSSE3;
+    if (IS_ALIGNED(width, 16)) {
+      MirrorRow = MirrorRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_MIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    MirrorRow = MirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 32)) {
+      MirrorRow = MirrorRow_AVX2;
+    }
+  }
+#endif
+// TODO(fbarchard): Mirror on mips handle unaligned memory.
+#if defined(HAS_MIRRORROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst, 4) && IS_ALIGNED(dst_stride, 4)) {
+    MirrorRow = MirrorRow_MIPS_DSPR2;
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    MirrorRow(src, row, width);  // Mirror first row into a buffer
+    src += src_stride;
+    MirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    dst += dst_stride;
+    CopyRow(row, dst_bot, width);  // Copy first mirrored row into last
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+LIBYUV_API
+void TransposeUV(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  int i = height;
+  void (*TransposeUVWx8)(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) = TransposeUVWx8_C;
+#if defined(HAS_TRANSPOSEUVWX8_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    TransposeUVWx8 = TransposeUVWx8_NEON;
+  }
+#endif
+#if defined(HAS_TRANSPOSEUVWX8_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 8)) {
+    TransposeUVWx8 = TransposeUVWx8_SSE2;
+  }
+#endif
+#if defined(HAS_TRANSPOSEUVWx8_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(width, 2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+    TransposeUVWx8 = TransposeUVWx8_MIPS_DSPR2;
+  }
+#endif
+
+  // Work through the source in 8x8 tiles.
+  while (i >= 8) {
+    TransposeUVWx8(src, src_stride,
+                   dst_a, dst_stride_a,
+                   dst_b, dst_stride_b,
+                   width);
+    src += 8 * src_stride;    // Go down 8 rows.
+    dst_a += 8;               // Move over 8 columns.
+    dst_b += 8;               // Move over 8 columns.
+    i -= 8;
+  }
+
+  if (i > 0) {
+    TransposeUVWxH_C(src, src_stride,
+                     dst_a, dst_stride_a,
+                     dst_b, dst_stride_b,
+                     width, i);
+  }
+}
+
+LIBYUV_API
+void RotateUV90(const uint8* src, int src_stride,
+                uint8* dst_a, int dst_stride_a,
+                uint8* dst_b, int dst_stride_b,
+                int width, int height) {
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+
+  TransposeUV(src, src_stride,
+              dst_a, dst_stride_a,
+              dst_b, dst_stride_b,
+              width, height);
+}
+
+LIBYUV_API
+void RotateUV270(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  dst_a += dst_stride_a * (width - 1);
+  dst_b += dst_stride_b * (width - 1);
+  dst_stride_a = -dst_stride_a;
+  dst_stride_b = -dst_stride_b;
+
+  TransposeUV(src, src_stride,
+              dst_a, dst_stride_a,
+              dst_b, dst_stride_b,
+              width, height);
+}
+
+// Rotate 180 is a horizontal and vertical flip.
+LIBYUV_API
+void RotateUV180(const uint8* src, int src_stride,
+                 uint8* dst_a, int dst_stride_a,
+                 uint8* dst_b, int dst_stride_b,
+                 int width, int height) {
+  int i;
+  void (*MirrorRowUV)(const uint8* src, uint8* dst_u, uint8* dst_v, int width) =
+      MirrorUVRow_C;
+#if defined(HAS_MIRRORUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
+    MirrorRowUV = MirrorUVRow_NEON;
+  }
+#endif
+#if defined(HAS_MIRRORROW_UV_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && IS_ALIGNED(width, 16)) {
+    MirrorRowUV = MirrorUVRow_SSSE3;
+  }
+#endif
+#if defined(HAS_MIRRORUVROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src, 4) && IS_ALIGNED(src_stride, 4)) {
+    MirrorRowUV = MirrorUVRow_MIPS_DSPR2;
+  }
+#endif
+
+  dst_a += dst_stride_a * (height - 1);
+  dst_b += dst_stride_b * (height - 1);
+
+  for (i = 0; i < height; ++i) {
+    MirrorRowUV(src, dst_a, dst_b, width);
+    src += src_stride;
+    dst_a -= dst_stride_a;
+    dst_b -= dst_stride_b;
+  }
+}
+
+LIBYUV_API
+int RotatePlane(const uint8* src, int src_stride,
+                uint8* dst, int dst_stride,
+                int width, int height,
+                enum RotationMode mode) {
+  if (!src || width <= 0 || height == 0 || !dst) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src = src + (height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      CopyPlane(src, src_stride,
+                dst, dst_stride,
+                width, height);
+      return 0;
+    case kRotate90:
+      RotatePlane90(src, src_stride,
+                    dst, dst_stride,
+                    width, height);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src, src_stride,
+                     dst, dst_stride,
+                     width, height);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src, src_stride,
+                     dst, dst_stride,
+                     width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int I420Rotate(const uint8* src_y, int src_stride_y,
+               const uint8* src_u, int src_stride_u,
+               const uint8* src_v, int src_stride_v,
+               uint8* dst_y, int dst_stride_y,
+               uint8* dst_u, int dst_stride_u,
+               uint8* dst_v, int dst_stride_v,
+               int width, int height,
+               enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_u || !src_v || width <= 0 || height == 0 ||
+      !dst_y || !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_u = src_u + (halfheight - 1) * src_stride_u;
+    src_v = src_v + (halfheight - 1) * src_stride_v;
+    src_stride_y = -src_stride_y;
+    src_stride_u = -src_stride_u;
+    src_stride_v = -src_stride_v;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return I420Copy(src_y, src_stride_y,
+                      src_u, src_stride_u,
+                      src_v, src_stride_v,
+                      dst_y, dst_stride_y,
+                      dst_u, dst_stride_u,
+                      dst_v, dst_stride_v,
+                      width, height);
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y,
+                    dst_y, dst_stride_y,
+                    width, height);
+      RotatePlane90(src_u, src_stride_u,
+                    dst_u, dst_stride_u,
+                    halfwidth, halfheight);
+      RotatePlane90(src_v, src_stride_v,
+                    dst_v, dst_stride_v,
+                    halfwidth, halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotatePlane270(src_u, src_stride_u,
+                     dst_u, dst_stride_u,
+                     halfwidth, halfheight);
+      RotatePlane270(src_v, src_stride_v,
+                     dst_v, dst_stride_v,
+                     halfwidth, halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotatePlane180(src_u, src_stride_u,
+                     dst_u, dst_stride_u,
+                     halfwidth, halfheight);
+      RotatePlane180(src_v, src_stride_v,
+                     dst_v, dst_stride_v,
+                     halfwidth, halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+LIBYUV_API
+int NV12ToI420Rotate(const uint8* src_y, int src_stride_y,
+                     const uint8* src_uv, int src_stride_uv,
+                     uint8* dst_y, int dst_stride_y,
+                     uint8* dst_u, int dst_stride_u,
+                     uint8* dst_v, int dst_stride_v,
+                     int width, int height,
+                     enum RotationMode mode) {
+  int halfwidth = (width + 1) >> 1;
+  int halfheight = (height + 1) >> 1;
+  if (!src_y || !src_uv || width <= 0 || height == 0 ||
+      !dst_y || !dst_u || !dst_v) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    halfheight = (height + 1) >> 1;
+    src_y = src_y + (height - 1) * src_stride_y;
+    src_uv = src_uv + (halfheight - 1) * src_stride_uv;
+    src_stride_y = -src_stride_y;
+    src_stride_uv = -src_stride_uv;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return NV12ToI420(src_y, src_stride_y,
+                        src_uv, src_stride_uv,
+                        dst_y, dst_stride_y,
+                        dst_u, dst_stride_u,
+                        dst_v, dst_stride_v,
+                        width, height);
+    case kRotate90:
+      RotatePlane90(src_y, src_stride_y,
+                    dst_y, dst_stride_y,
+                    width, height);
+      RotateUV90(src_uv, src_stride_uv,
+                 dst_u, dst_stride_u,
+                 dst_v, dst_stride_v,
+                 halfwidth, halfheight);
+      return 0;
+    case kRotate270:
+      RotatePlane270(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotateUV270(src_uv, src_stride_uv,
+                  dst_u, dst_stride_u,
+                  dst_v, dst_stride_v,
+                  halfwidth, halfheight);
+      return 0;
+    case kRotate180:
+      RotatePlane180(src_y, src_stride_y,
+                     dst_y, dst_stride_y,
+                     width, height);
+      RotateUV180(src_uv, src_stride_uv,
+                  dst_u, dst_stride_u,
+                  dst_v, dst_stride_v,
+                  halfwidth, halfheight);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/rotate_any.cc b/third_party/aom/third_party/libyuv/source/rotate_any.cc
new file mode 100644
index 0000000000..4d6eb34e18
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/rotate_any.cc
@@ -0,0 +1,55 @@
+/*
+ *  Copyright 2015 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+#include "libyuv/rotate_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define TANY(NAMEANY, TPOS_SIMD, TPOS_C, MASK)                                 \
+    void NAMEANY(const uint8* src, int src_stride,                             \
+                 uint8* dst, int dst_stride, int width) {                      \
+      int r = width & MASK;                                                    \
+      int n = width - r;                                                       \
+      if (n > 0) {                                                             \
+        TPOS_SIMD(src, src_stride, dst, dst_stride, n);                        \
+      }                                                                        \
+      TPOS_C(src + n, src_stride, dst + n * dst_stride, dst_stride, r);        \
+    }
+
+#ifdef HAS_TRANSPOSEWX8_NEON
+TANY(TransposeWx8_Any_NEON, TransposeWx8_NEON, TransposeWx8_C, 7)
+#endif
+#ifdef HAS_TRANSPOSEWX8_SSSE3
+TANY(TransposeWx8_Any_SSSE3, TransposeWx8_SSSE3, TransposeWx8_C, 7)
+#endif
+#ifdef HAS_TRANSPOSEWX8_FAST_SSSE3
+TANY(TransposeWx8_Fast_Any_SSSE3, TransposeWx8_Fast_SSSE3, TransposeWx8_C, 15)
+#endif
+#ifdef HAS_TRANSPOSEWX8_MIPS_DSPR2
+TANY(TransposeWx8_Any_MIPS_DSPR2, TransposeWx8_MIPS_DSPR2, TransposeWx8_C, 7)
+#endif
+
+#undef TANY
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+
+
+
+
diff --git a/third_party/aom/third_party/libyuv/source/rotate_argb.cc b/third_party/aom/third_party/libyuv/source/rotate_argb.cc
new file mode 100644
index 0000000000..787c0ad1be
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/rotate_argb.cc
@@ -0,0 +1,205 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/rotate.h"
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/convert.h"
+#include "libyuv/planar_functions.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// ARGBScale has a function to copy pixels to a row, striding each source
+// pixel by a constant.
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(_M_IX86) || \
+    (defined(__x86_64__) && !defined(__native_client__)) || defined(__i386__))
+#define HAS_SCALEARGBROWDOWNEVEN_SSE2
+void ScaleARGBRowDownEven_SSE2(const uint8* src_ptr, int src_stride,
+                               int src_stepx, uint8* dst_ptr, int dst_width);
+#endif
+#if !defined(LIBYUV_DISABLE_NEON) && !defined(__native_client__) && \
+    (defined(__ARM_NEON__) || defined(LIBYUV_NEON) || defined(__aarch64__))
+#define HAS_SCALEARGBROWDOWNEVEN_NEON
+void ScaleARGBRowDownEven_NEON(const uint8* src_ptr, int src_stride,
+                               int src_stepx, uint8* dst_ptr, int dst_width);
+#endif
+
+void ScaleARGBRowDownEven_C(const uint8* src_ptr, int,
+                            int src_stepx, uint8* dst_ptr, int dst_width);
+
+static void ARGBTranspose(const uint8* src, int src_stride,
+                          uint8* dst, int dst_stride, int width, int height) {
+  int i;
+  int src_pixel_step = src_stride >> 2;
+  void (*ScaleARGBRowDownEven)(const uint8* src_ptr, int src_stride,
+      int src_step, uint8* dst_ptr, int dst_width) = ScaleARGBRowDownEven_C;
+#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(height, 4)) {  // Width of dest.
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(height, 4)) {  // Width of dest.
+    ScaleARGBRowDownEven = ScaleARGBRowDownEven_NEON;
+  }
+#endif
+
+  for (i = 0; i < width; ++i) {  // column of source to row of dest.
+    ScaleARGBRowDownEven(src, 0, src_pixel_step, dst, height);
+    dst += dst_stride;
+    src += 4;
+  }
+}
+
+void ARGBRotate90(const uint8* src, int src_stride,
+                  uint8* dst, int dst_stride, int width, int height) {
+  // Rotate by 90 is a ARGBTranspose with the source read
+  // from bottom to top. So set the source pointer to the end
+  // of the buffer and flip the sign of the source stride.
+  src += src_stride * (height - 1);
+  src_stride = -src_stride;
+  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+}
+
+void ARGBRotate270(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride, int width, int height) {
+  // Rotate by 270 is a ARGBTranspose with the destination written
+  // from bottom to top. So set the destination pointer to the end
+  // of the buffer and flip the sign of the destination stride.
+  dst += dst_stride * (width - 1);
+  dst_stride = -dst_stride;
+  ARGBTranspose(src, src_stride, dst, dst_stride, width, height);
+}
+
+void ARGBRotate180(const uint8* src, int src_stride,
+                   uint8* dst, int dst_stride, int width, int height) {
+  // Swap first and last row and mirror the content. Uses a temporary row.
+  align_buffer_64(row, width * 4);
+  const uint8* src_bot = src + src_stride * (height - 1);
+  uint8* dst_bot = dst + dst_stride * (height - 1);
+  int half_height = (height + 1) >> 1;
+  int y;
+  void (*ARGBMirrorRow)(const uint8* src, uint8* dst, int width) =
+      ARGBMirrorRow_C;
+  void (*CopyRow)(const uint8* src, uint8* dst, int width) = CopyRow_C;
+#if defined(HAS_ARGBMIRRORROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_NEON;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBMirrorRow = ARGBMirrorRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_ARGBMIRRORROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ARGBMirrorRow = ARGBMirrorRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBMirrorRow = ARGBMirrorRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_COPYROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_SSE2 : CopyRow_Any_SSE2;
+  }
+#endif
+#if defined(HAS_COPYROW_AVX)
+  if (TestCpuFlag(kCpuHasAVX)) {
+    CopyRow = IS_ALIGNED(width * 4, 64) ? CopyRow_AVX : CopyRow_Any_AVX;
+  }
+#endif
+#if defined(HAS_COPYROW_ERMS)
+  if (TestCpuFlag(kCpuHasERMS)) {
+    CopyRow = CopyRow_ERMS;
+  }
+#endif
+#if defined(HAS_COPYROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    CopyRow = IS_ALIGNED(width * 4, 32) ? CopyRow_NEON : CopyRow_Any_NEON;
+  }
+#endif
+#if defined(HAS_COPYROW_MIPS)
+  if (TestCpuFlag(kCpuHasMIPS)) {
+    CopyRow = CopyRow_MIPS;
+  }
+#endif
+
+  // Odd height will harmlessly mirror the middle row twice.
+  for (y = 0; y < half_height; ++y) {
+    ARGBMirrorRow(src, row, width);  // Mirror first row into a buffer
+    ARGBMirrorRow(src_bot, dst, width);  // Mirror last row into first row
+    CopyRow(row, dst_bot, width * 4);  // Copy first mirrored row into last
+    src += src_stride;
+    dst += dst_stride;
+    src_bot -= src_stride;
+    dst_bot -= dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+LIBYUV_API
+int ARGBRotate(const uint8* src_argb, int src_stride_argb,
+               uint8* dst_argb, int dst_stride_argb, int width, int height,
+               enum RotationMode mode) {
+  if (!src_argb || width <= 0 || height == 0 || !dst_argb) {
+    return -1;
+  }
+
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    src_argb = src_argb + (height - 1) * src_stride_argb;
+    src_stride_argb = -src_stride_argb;
+  }
+
+  switch (mode) {
+    case kRotate0:
+      // copy frame
+      return ARGBCopy(src_argb, src_stride_argb,
+                      dst_argb, dst_stride_argb,
+                      width, height);
+    case kRotate90:
+      ARGBRotate90(src_argb, src_stride_argb,
+                   dst_argb, dst_stride_argb,
+                   width, height);
+      return 0;
+    case kRotate270:
+      ARGBRotate270(src_argb, src_stride_argb,
+                    dst_argb, dst_stride_argb,
+                    width, height);
+      return 0;
+    case kRotate180:
+      ARGBRotate180(src_argb, src_stride_argb,
+                    dst_argb, dst_stride_argb,
+                    width, height);
+      return 0;
+    default:
+      break;
+  }
+  return -1;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/rotate_common.cc b/third_party/aom/third_party/libyuv/source/rotate_common.cc
new file mode 100644
index 0000000000..b33a9a0c6e
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/rotate_common.cc
@@ -0,0 +1,92 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+void TransposeWx8_C(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst[0] = src[0 * src_stride];
+    dst[1] = src[1 * src_stride];
+    dst[2] = src[2 * src_stride];
+    dst[3] = src[3 * src_stride];
+    dst[4] = src[4 * src_stride];
+    dst[5] = src[5 * src_stride];
+    dst[6] = src[6 * src_stride];
+    dst[7] = src[7 * src_stride];
+    ++src;
+    dst += dst_stride;
+  }
+}
+
+void TransposeUVWx8_C(const uint8* src, int src_stride,
+                      uint8* dst_a, int dst_stride_a,
+                      uint8* dst_b, int dst_stride_b, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    dst_a[0] = src[0 * src_stride + 0];
+    dst_b[0] = src[0 * src_stride + 1];
+    dst_a[1] = src[1 * src_stride + 0];
+    dst_b[1] = src[1 * src_stride + 1];
+    dst_a[2] = src[2 * src_stride + 0];
+    dst_b[2] = src[2 * src_stride + 1];
+    dst_a[3] = src[3 * src_stride + 0];
+    dst_b[3] = src[3 * src_stride + 1];
+    dst_a[4] = src[4 * src_stride + 0];
+    dst_b[4] = src[4 * src_stride + 1];
+    dst_a[5] = src[5 * src_stride + 0];
+    dst_b[5] = src[5 * src_stride + 1];
+    dst_a[6] = src[6 * src_stride + 0];
+    dst_b[6] = src[6 * src_stride + 1];
+    dst_a[7] = src[7 * src_stride + 0];
+    dst_b[7] = src[7 * src_stride + 1];
+    src += 2;
+    dst_a += dst_stride_a;
+    dst_b += dst_stride_b;
+  }
+}
+
+void TransposeWxH_C(const uint8* src, int src_stride,
+                    uint8* dst, int dst_stride,
+                    int width, int height) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst[i * dst_stride + j] = src[j * src_stride + i];
+    }
+  }
+}
+
+void TransposeUVWxH_C(const uint8* src, int src_stride,
+                      uint8* dst_a, int dst_stride_a,
+                      uint8* dst_b, int dst_stride_b,
+                      int width, int height) {
+  int i;
+  for (i = 0; i < width * 2; i += 2) {
+    int j;
+    for (j = 0; j < height; ++j) {
+      dst_a[j + ((i >> 1) * dst_stride_a)] = src[i + (j * src_stride)];
+      dst_b[j + ((i >> 1) * dst_stride_b)] = src[i + (j * src_stride) + 1];
+    }
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/rotate_gcc.cc b/third_party/aom/third_party/libyuv/source/rotate_gcc.cc
new file mode 100644
index 0000000000..fd385bcd30
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/rotate_gcc.cc
@@ -0,0 +1,493 @@
+/*
+ *  Copyright 2015 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+#if !defined(LIBYUV_DISABLE_X86) && \
+    (defined(__i386__) || (defined(__x86_64__) && !defined(__native_client__)))
+void TransposeWx8_SSSE3(const uint8* src, int src_stride,
+                        uint8* dst, int dst_stride, int width) {
+  asm volatile (
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    ".p2align  2                                 \n"
+  "1:                                            \n"
+    "movq       (%0),%%xmm0                      \n"
+    "movq       (%0,%3),%%xmm1                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm1,%%xmm0                    \n"
+    "movq       (%0),%%xmm2                      \n"
+    "movdqa     %%xmm0,%%xmm1                    \n"
+    "palignr    $0x8,%%xmm1,%%xmm1               \n"
+    "movq       (%0,%3),%%xmm3                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm3,%%xmm2                    \n"
+    "movdqa     %%xmm2,%%xmm3                    \n"
+    "movq       (%0),%%xmm4                      \n"
+    "palignr    $0x8,%%xmm3,%%xmm3               \n"
+    "movq       (%0,%3),%%xmm5                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm5,%%xmm4                    \n"
+    "movdqa     %%xmm4,%%xmm5                    \n"
+    "movq       (%0),%%xmm6                      \n"
+    "palignr    $0x8,%%xmm5,%%xmm5               \n"
+    "movq       (%0,%3),%%xmm7                   \n"
+    "lea        (%0,%3,2),%0                     \n"
+    "punpcklbw  %%xmm7,%%xmm6                    \n"
+    "neg        %3                               \n"
+    "movdqa     %%xmm6,%%xmm7                    \n"
+    "lea        0x8(%0,%3,8),%0                  \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    "neg        %3                               \n"
+     // Second round of bit swap.
+    "punpcklwd  %%xmm2,%%xmm0                    \n"
+    "punpcklwd  %%xmm3,%%xmm1                    \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "palignr    $0x8,%%xmm2,%%xmm2               \n"
+    "palignr    $0x8,%%xmm3,%%xmm3               \n"
+    "punpcklwd  %%xmm6,%%xmm4                    \n"
+    "punpcklwd  %%xmm7,%%xmm5                    \n"
+    "movdqa     %%xmm4,%%xmm6                    \n"
+    "movdqa     %%xmm5,%%xmm7                    \n"
+    "palignr    $0x8,%%xmm6,%%xmm6               \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    "punpckldq  %%xmm4,%%xmm0                    \n"
+    "movq       %%xmm0,(%1)                      \n"
+    "movdqa     %%xmm0,%%xmm4                    \n"
+    "palignr    $0x8,%%xmm4,%%xmm4               \n"
+    "movq       %%xmm4,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "punpckldq  %%xmm6,%%xmm2                    \n"
+    "movdqa     %%xmm2,%%xmm6                    \n"
+    "movq       %%xmm2,(%1)                      \n"
+    "palignr    $0x8,%%xmm6,%%xmm6               \n"
+    "punpckldq  %%xmm5,%%xmm1                    \n"
+    "movq       %%xmm6,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "movdqa     %%xmm1,%%xmm5                    \n"
+    "movq       %%xmm1,(%1)                      \n"
+    "palignr    $0x8,%%xmm5,%%xmm5               \n"
+    "movq       %%xmm5,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "punpckldq  %%xmm7,%%xmm3                    \n"
+    "movq       %%xmm3,(%1)                      \n"
+    "movdqa     %%xmm3,%%xmm7                    \n"
+    "palignr    $0x8,%%xmm7,%%xmm7               \n"
+    "sub        $0x8,%2                          \n"
+    "movq       %%xmm7,(%1,%4)                   \n"
+    "lea        (%1,%4,2),%1                     \n"
+    "jg         1b                               \n"
+    : "+r"(src),    // %0
+      "+r"(dst),    // %1
+      "+r"(width)   // %2
+    : "r"((intptr_t)(src_stride)),  // %3
+      "r"((intptr_t)(dst_stride))   // %4
+    : "memory", "cc",
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(__i386__)  && !defined(__clang__)
+void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b, int width);
+  asm (
+    DECLARE_FUNCTION(TransposeUVWx8_SSE2)
+    "push   %ebx                               \n"
+    "push   %esi                               \n"
+    "push   %edi                               \n"
+    "push   %ebp                               \n"
+    "mov    0x14(%esp),%eax                    \n"
+    "mov    0x18(%esp),%edi                    \n"
+    "mov    0x1c(%esp),%edx                    \n"
+    "mov    0x20(%esp),%esi                    \n"
+    "mov    0x24(%esp),%ebx                    \n"
+    "mov    0x28(%esp),%ebp                    \n"
+    "mov    %esp,%ecx                          \n"
+    "sub    $0x14,%esp                         \n"
+    "and    $0xfffffff0,%esp                   \n"
+    "mov    %ecx,0x10(%esp)                    \n"
+    "mov    0x2c(%ecx),%ecx                    \n"
+
+"1:                                            \n"
+    "movdqu (%eax),%xmm0                       \n"
+    "movdqu (%eax,%edi,1),%xmm1                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm0,%xmm7                        \n"
+    "punpcklbw %xmm1,%xmm0                     \n"
+    "punpckhbw %xmm1,%xmm7                     \n"
+    "movdqa %xmm7,%xmm1                        \n"
+    "movdqu (%eax),%xmm2                       \n"
+    "movdqu (%eax,%edi,1),%xmm3                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm2,%xmm7                        \n"
+    "punpcklbw %xmm3,%xmm2                     \n"
+    "punpckhbw %xmm3,%xmm7                     \n"
+    "movdqa %xmm7,%xmm3                        \n"
+    "movdqu (%eax),%xmm4                       \n"
+    "movdqu (%eax,%edi,1),%xmm5                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqa %xmm4,%xmm7                        \n"
+    "punpcklbw %xmm5,%xmm4                     \n"
+    "punpckhbw %xmm5,%xmm7                     \n"
+    "movdqa %xmm7,%xmm5                        \n"
+    "movdqu (%eax),%xmm6                       \n"
+    "movdqu (%eax,%edi,1),%xmm7                \n"
+    "lea    (%eax,%edi,2),%eax                 \n"
+    "movdqu %xmm5,(%esp)                       \n"
+    "neg    %edi                               \n"
+    "movdqa %xmm6,%xmm5                        \n"
+    "punpcklbw %xmm7,%xmm6                     \n"
+    "punpckhbw %xmm7,%xmm5                     \n"
+    "movdqa %xmm5,%xmm7                        \n"
+    "lea    0x10(%eax,%edi,8),%eax             \n"
+    "neg    %edi                               \n"
+    "movdqa %xmm0,%xmm5                        \n"
+    "punpcklwd %xmm2,%xmm0                     \n"
+    "punpckhwd %xmm2,%xmm5                     \n"
+    "movdqa %xmm5,%xmm2                        \n"
+    "movdqa %xmm1,%xmm5                        \n"
+    "punpcklwd %xmm3,%xmm1                     \n"
+    "punpckhwd %xmm3,%xmm5                     \n"
+    "movdqa %xmm5,%xmm3                        \n"
+    "movdqa %xmm4,%xmm5                        \n"
+    "punpcklwd %xmm6,%xmm4                     \n"
+    "punpckhwd %xmm6,%xmm5                     \n"
+    "movdqa %xmm5,%xmm6                        \n"
+    "movdqu (%esp),%xmm5                       \n"
+    "movdqu %xmm6,(%esp)                       \n"
+    "movdqa %xmm5,%xmm6                        \n"
+    "punpcklwd %xmm7,%xmm5                     \n"
+    "punpckhwd %xmm7,%xmm6                     \n"
+    "movdqa %xmm6,%xmm7                        \n"
+    "movdqa %xmm0,%xmm6                        \n"
+    "punpckldq %xmm4,%xmm0                     \n"
+    "punpckhdq %xmm4,%xmm6                     \n"
+    "movdqa %xmm6,%xmm4                        \n"
+    "movdqu (%esp),%xmm6                       \n"
+    "movlpd %xmm0,(%edx)                       \n"
+    "movhpd %xmm0,(%ebx)                       \n"
+    "movlpd %xmm4,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm4,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm2,%xmm0                        \n"
+    "punpckldq %xmm6,%xmm2                     \n"
+    "movlpd %xmm2,(%edx)                       \n"
+    "movhpd %xmm2,(%ebx)                       \n"
+    "punpckhdq %xmm6,%xmm0                     \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm1,%xmm0                        \n"
+    "punpckldq %xmm5,%xmm1                     \n"
+    "movlpd %xmm1,(%edx)                       \n"
+    "movhpd %xmm1,(%ebx)                       \n"
+    "punpckhdq %xmm5,%xmm0                     \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "movdqa %xmm3,%xmm0                        \n"
+    "punpckldq %xmm7,%xmm3                     \n"
+    "movlpd %xmm3,(%edx)                       \n"
+    "movhpd %xmm3,(%ebx)                       \n"
+    "punpckhdq %xmm7,%xmm0                     \n"
+    "sub    $0x8,%ecx                          \n"
+    "movlpd %xmm0,(%edx,%esi,1)                \n"
+    "lea    (%edx,%esi,2),%edx                 \n"
+    "movhpd %xmm0,(%ebx,%ebp,1)                \n"
+    "lea    (%ebx,%ebp,2),%ebx                 \n"
+    "jg     1b                                 \n"
+    "mov    0x10(%esp),%esp                    \n"
+    "pop    %ebp                               \n"
+    "pop    %edi                               \n"
+    "pop    %esi                               \n"
+    "pop    %ebx                               \n"
+#if defined(__native_client__)
+    "pop    %ecx                               \n"
+    "and    $0xffffffe0,%ecx                   \n"
+    "jmp    *%ecx                              \n"
+#else
+    "ret                                       \n"
+#endif
+);
+#endif
+#if !defined(LIBYUV_DISABLE_X86) && !defined(__native_client__) && \
+    defined(__x86_64__)
+// 64 bit version has enough registers to do 16x8 to 8x16 at a time.
+void TransposeWx8_Fast_SSSE3(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride, int width) {
+  asm volatile (
+  // Read in the data from the source pointer.
+  // First round of bit swap.
+  ".p2align  2                                 \n"
+"1:                                            \n"
+  "movdqu     (%0),%%xmm0                      \n"
+  "movdqu     (%0,%3),%%xmm1                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpcklbw  %%xmm1,%%xmm0                    \n"
+  "punpckhbw  %%xmm1,%%xmm8                    \n"
+  "movdqu     (%0),%%xmm2                      \n"
+  "movdqa     %%xmm0,%%xmm1                    \n"
+  "movdqa     %%xmm8,%%xmm9                    \n"
+  "palignr    $0x8,%%xmm1,%%xmm1               \n"
+  "palignr    $0x8,%%xmm9,%%xmm9               \n"
+  "movdqu     (%0,%3),%%xmm3                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm2,%%xmm10                   \n"
+  "punpcklbw  %%xmm3,%%xmm2                    \n"
+  "punpckhbw  %%xmm3,%%xmm10                   \n"
+  "movdqa     %%xmm2,%%xmm3                    \n"
+  "movdqa     %%xmm10,%%xmm11                  \n"
+  "movdqu     (%0),%%xmm4                      \n"
+  "palignr    $0x8,%%xmm3,%%xmm3               \n"
+  "palignr    $0x8,%%xmm11,%%xmm11             \n"
+  "movdqu     (%0,%3),%%xmm5                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm4,%%xmm12                   \n"
+  "punpcklbw  %%xmm5,%%xmm4                    \n"
+  "punpckhbw  %%xmm5,%%xmm12                   \n"
+  "movdqa     %%xmm4,%%xmm5                    \n"
+  "movdqa     %%xmm12,%%xmm13                  \n"
+  "movdqu     (%0),%%xmm6                      \n"
+  "palignr    $0x8,%%xmm5,%%xmm5               \n"
+  "palignr    $0x8,%%xmm13,%%xmm13             \n"
+  "movdqu     (%0,%3),%%xmm7                   \n"
+  "lea        (%0,%3,2),%0                     \n"
+  "movdqa     %%xmm6,%%xmm14                   \n"
+  "punpcklbw  %%xmm7,%%xmm6                    \n"
+  "punpckhbw  %%xmm7,%%xmm14                   \n"
+  "neg        %3                               \n"
+  "movdqa     %%xmm6,%%xmm7                    \n"
+  "movdqa     %%xmm14,%%xmm15                  \n"
+  "lea        0x10(%0,%3,8),%0                 \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  "neg        %3                               \n"
+   // Second round of bit swap.
+  "punpcklwd  %%xmm2,%%xmm0                    \n"
+  "punpcklwd  %%xmm3,%%xmm1                    \n"
+  "movdqa     %%xmm0,%%xmm2                    \n"
+  "movdqa     %%xmm1,%%xmm3                    \n"
+  "palignr    $0x8,%%xmm2,%%xmm2               \n"
+  "palignr    $0x8,%%xmm3,%%xmm3               \n"
+  "punpcklwd  %%xmm6,%%xmm4                    \n"
+  "punpcklwd  %%xmm7,%%xmm5                    \n"
+  "movdqa     %%xmm4,%%xmm6                    \n"
+  "movdqa     %%xmm5,%%xmm7                    \n"
+  "palignr    $0x8,%%xmm6,%%xmm6               \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "punpcklwd  %%xmm10,%%xmm8                   \n"
+  "punpcklwd  %%xmm11,%%xmm9                   \n"
+  "movdqa     %%xmm8,%%xmm10                   \n"
+  "movdqa     %%xmm9,%%xmm11                   \n"
+  "palignr    $0x8,%%xmm10,%%xmm10             \n"
+  "palignr    $0x8,%%xmm11,%%xmm11             \n"
+  "punpcklwd  %%xmm14,%%xmm12                  \n"
+  "punpcklwd  %%xmm15,%%xmm13                  \n"
+  "movdqa     %%xmm12,%%xmm14                  \n"
+  "movdqa     %%xmm13,%%xmm15                  \n"
+  "palignr    $0x8,%%xmm14,%%xmm14             \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  // Third round of bit swap.
+  // Write to the destination pointer.
+  "punpckldq  %%xmm4,%%xmm0                    \n"
+  "movq       %%xmm0,(%1)                      \n"
+  "movdqa     %%xmm0,%%xmm4                    \n"
+  "palignr    $0x8,%%xmm4,%%xmm4               \n"
+  "movq       %%xmm4,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm6,%%xmm2                    \n"
+  "movdqa     %%xmm2,%%xmm6                    \n"
+  "movq       %%xmm2,(%1)                      \n"
+  "palignr    $0x8,%%xmm6,%%xmm6               \n"
+  "punpckldq  %%xmm5,%%xmm1                    \n"
+  "movq       %%xmm6,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "movdqa     %%xmm1,%%xmm5                    \n"
+  "movq       %%xmm1,(%1)                      \n"
+  "palignr    $0x8,%%xmm5,%%xmm5               \n"
+  "movq       %%xmm5,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm7,%%xmm3                    \n"
+  "movq       %%xmm3,(%1)                      \n"
+  "movdqa     %%xmm3,%%xmm7                    \n"
+  "palignr    $0x8,%%xmm7,%%xmm7               \n"
+  "movq       %%xmm7,(%1,%4)                   \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm12,%%xmm8                   \n"
+  "movq       %%xmm8,(%1)                      \n"
+  "movdqa     %%xmm8,%%xmm12                   \n"
+  "palignr    $0x8,%%xmm12,%%xmm12             \n"
+  "movq       %%xmm12,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm14,%%xmm10                  \n"
+  "movdqa     %%xmm10,%%xmm14                  \n"
+  "movq       %%xmm10,(%1)                     \n"
+  "palignr    $0x8,%%xmm14,%%xmm14             \n"
+  "punpckldq  %%xmm13,%%xmm9                   \n"
+  "movq       %%xmm14,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "movdqa     %%xmm9,%%xmm13                   \n"
+  "movq       %%xmm9,(%1)                      \n"
+  "palignr    $0x8,%%xmm13,%%xmm13             \n"
+  "movq       %%xmm13,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "punpckldq  %%xmm15,%%xmm11                  \n"
+  "movq       %%xmm11,(%1)                     \n"
+  "movdqa     %%xmm11,%%xmm15                  \n"
+  "palignr    $0x8,%%xmm15,%%xmm15             \n"
+  "sub        $0x10,%2                         \n"
+  "movq       %%xmm15,(%1,%4)                  \n"
+  "lea        (%1,%4,2),%1                     \n"
+  "jg         1b                               \n"
+  : "+r"(src),    // %0
+    "+r"(dst),    // %1
+    "+r"(width)   // %2
+  : "r"((intptr_t)(src_stride)),  // %3
+    "r"((intptr_t)(dst_stride))   // %4
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
+    "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13",  "xmm14",  "xmm15"
+);
+}
+
+void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b, int width) {
+  asm volatile (
+  // Read in the data from the source pointer.
+  // First round of bit swap.
+  ".p2align  2                                 \n"
+"1:                                            \n"
+  "movdqu     (%0),%%xmm0                      \n"
+  "movdqu     (%0,%4),%%xmm1                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpcklbw  %%xmm1,%%xmm0                    \n"
+  "punpckhbw  %%xmm1,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm1                    \n"
+  "movdqu     (%0),%%xmm2                      \n"
+  "movdqu     (%0,%4),%%xmm3                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm2,%%xmm8                    \n"
+  "punpcklbw  %%xmm3,%%xmm2                    \n"
+  "punpckhbw  %%xmm3,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm3                    \n"
+  "movdqu     (%0),%%xmm4                      \n"
+  "movdqu     (%0,%4),%%xmm5                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm4,%%xmm8                    \n"
+  "punpcklbw  %%xmm5,%%xmm4                    \n"
+  "punpckhbw  %%xmm5,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm5                    \n"
+  "movdqu     (%0),%%xmm6                      \n"
+  "movdqu     (%0,%4),%%xmm7                   \n"
+  "lea        (%0,%4,2),%0                     \n"
+  "movdqa     %%xmm6,%%xmm8                    \n"
+  "punpcklbw  %%xmm7,%%xmm6                    \n"
+  "neg        %4                               \n"
+  "lea        0x10(%0,%4,8),%0                 \n"
+  "punpckhbw  %%xmm7,%%xmm8                    \n"
+  "movdqa     %%xmm8,%%xmm7                    \n"
+  "neg        %4                               \n"
+   // Second round of bit swap.
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "movdqa     %%xmm1,%%xmm9                    \n"
+  "punpckhwd  %%xmm2,%%xmm8                    \n"
+  "punpckhwd  %%xmm3,%%xmm9                    \n"
+  "punpcklwd  %%xmm2,%%xmm0                    \n"
+  "punpcklwd  %%xmm3,%%xmm1                    \n"
+  "movdqa     %%xmm8,%%xmm2                    \n"
+  "movdqa     %%xmm9,%%xmm3                    \n"
+  "movdqa     %%xmm4,%%xmm8                    \n"
+  "movdqa     %%xmm5,%%xmm9                    \n"
+  "punpckhwd  %%xmm6,%%xmm8                    \n"
+  "punpckhwd  %%xmm7,%%xmm9                    \n"
+  "punpcklwd  %%xmm6,%%xmm4                    \n"
+  "punpcklwd  %%xmm7,%%xmm5                    \n"
+  "movdqa     %%xmm8,%%xmm6                    \n"
+  "movdqa     %%xmm9,%%xmm7                    \n"
+  // Third round of bit swap.
+  // Write to the destination pointer.
+  "movdqa     %%xmm0,%%xmm8                    \n"
+  "punpckldq  %%xmm4,%%xmm0                    \n"
+  "movlpd     %%xmm0,(%1)                      \n"  // Write back U channel
+  "movhpd     %%xmm0,(%2)                      \n"  // Write back V channel
+  "punpckhdq  %%xmm4,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm2,%%xmm8                    \n"
+  "punpckldq  %%xmm6,%%xmm2                    \n"
+  "movlpd     %%xmm2,(%1)                      \n"
+  "movhpd     %%xmm2,(%2)                      \n"
+  "punpckhdq  %%xmm6,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm1,%%xmm8                    \n"
+  "punpckldq  %%xmm5,%%xmm1                    \n"
+  "movlpd     %%xmm1,(%1)                      \n"
+  "movhpd     %%xmm1,(%2)                      \n"
+  "punpckhdq  %%xmm5,%%xmm8                    \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "movdqa     %%xmm3,%%xmm8                    \n"
+  "punpckldq  %%xmm7,%%xmm3                    \n"
+  "movlpd     %%xmm3,(%1)                      \n"
+  "movhpd     %%xmm3,(%2)                      \n"
+  "punpckhdq  %%xmm7,%%xmm8                    \n"
+  "sub        $0x8,%3                          \n"
+  "movlpd     %%xmm8,(%1,%5)                   \n"
+  "lea        (%1,%5,2),%1                     \n"
+  "movhpd     %%xmm8,(%2,%6)                   \n"
+  "lea        (%2,%6,2),%2                     \n"
+  "jg         1b                               \n"
+  : "+r"(src),    // %0
+    "+r"(dst_a),  // %1
+    "+r"(dst_b),  // %2
+    "+r"(width)   // %3
+  : "r"((intptr_t)(src_stride)),    // %4
+    "r"((intptr_t)(dst_stride_a)),  // %5
+    "r"((intptr_t)(dst_stride_b))   // %6
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
+    "xmm8", "xmm9"
+);
+}
+#endif
+#endif
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/rotate_mips.cc b/third_party/aom/third_party/libyuv/source/rotate_mips.cc
new file mode 100644
index 0000000000..efe6bd909e
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/rotate_mips.cc
@@ -0,0 +1,484 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_MIPS) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+void TransposeWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                             uint8* dst, int dst_stride, int width) {
+   __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "sll              $t2, %[src_stride], 0x1          \n" // src_stride x 2
+      "sll              $t4, %[src_stride], 0x2          \n" // src_stride x 4
+      "sll              $t9, %[src_stride], 0x3          \n" // src_stride x 8
+      "addu             $t3, $t2, %[src_stride]          \n"
+      "addu             $t5, $t4, %[src_stride]          \n"
+      "addu             $t6, $t2, $t4                    \n"
+      "andi             $t0, %[dst], 0x3                 \n"
+      "andi             $t1, %[dst_stride], 0x3          \n"
+      "or               $t0, $t0, $t1                    \n"
+      "bnez             $t0, 11f                         \n"
+      " subu            $t7, $t9, %[src_stride]          \n"
+//dst + dst_stride word aligned
+    "1:                                                  \n"
+      "lbu              $t0, 0(%[src])                   \n"
+      "lbux             $t1, %[src_stride](%[src])       \n"
+      "lbux             $t8, $t2(%[src])                 \n"
+      "lbux             $t9, $t3(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s0, $t8, $t0                    \n"
+      "lbux             $t0, $t4(%[src])                 \n"
+      "lbux             $t1, $t5(%[src])                 \n"
+      "lbux             $t8, $t6(%[src])                 \n"
+      "lbux             $t9, $t7(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s1, $t8, $t0                    \n"
+      "sw               $s0, 0(%[dst])                   \n"
+      "addiu            %[width], -1                     \n"
+      "addiu            %[src], 1                        \n"
+      "sw               $s1, 4(%[dst])                   \n"
+      "bnez             %[width], 1b                     \n"
+      " addu            %[dst], %[dst], %[dst_stride]    \n"
+      "b                2f                               \n"
+//dst + dst_stride unaligned
+   "11:                                                  \n"
+      "lbu              $t0, 0(%[src])                   \n"
+      "lbux             $t1, %[src_stride](%[src])       \n"
+      "lbux             $t8, $t2(%[src])                 \n"
+      "lbux             $t9, $t3(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s0, $t8, $t0                    \n"
+      "lbux             $t0, $t4(%[src])                 \n"
+      "lbux             $t1, $t5(%[src])                 \n"
+      "lbux             $t8, $t6(%[src])                 \n"
+      "lbux             $t9, $t7(%[src])                 \n"
+      "sll              $t1, $t1, 16                     \n"
+      "sll              $t9, $t9, 16                     \n"
+      "or               $t0, $t0, $t1                    \n"
+      "or               $t8, $t8, $t9                    \n"
+      "precr.qb.ph      $s1, $t8, $t0                    \n"
+      "swr              $s0, 0(%[dst])                   \n"
+      "swl              $s0, 3(%[dst])                   \n"
+      "addiu            %[width], -1                     \n"
+      "addiu            %[src], 1                        \n"
+      "swr              $s1, 4(%[dst])                   \n"
+      "swl              $s1, 7(%[dst])                   \n"
+      "bnez             %[width], 11b                    \n"
+       "addu             %[dst], %[dst], %[dst_stride]   \n"
+    "2:                                                  \n"
+      ".set pop                                          \n"
+      :[src] "+r" (src),
+       [dst] "+r" (dst),
+       [width] "+r" (width)
+      :[src_stride] "r" (src_stride),
+       [dst_stride] "r" (dst_stride)
+      : "t0", "t1",  "t2", "t3", "t4", "t5",
+        "t6", "t7", "t8", "t9",
+        "s0", "s1"
+  );
+}
+
+void TransposeWx8_Fast_MIPS_DSPR2(const uint8* src, int src_stride,
+                                  uint8* dst, int dst_stride, int width) {
+  __asm__ __volatile__ (
+      ".set noat                                         \n"
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "beqz             %[width], 2f                     \n"
+      " sll             $t2, %[src_stride], 0x1          \n"  // src_stride x 2
+      "sll              $t4, %[src_stride], 0x2          \n"  // src_stride x 4
+      "sll              $t9, %[src_stride], 0x3          \n"  // src_stride x 8
+      "addu             $t3, $t2, %[src_stride]          \n"
+      "addu             $t5, $t4, %[src_stride]          \n"
+      "addu             $t6, $t2, $t4                    \n"
+
+      "srl              $AT, %[width], 0x2               \n"
+      "andi             $t0, %[dst], 0x3                 \n"
+      "andi             $t1, %[dst_stride], 0x3          \n"
+      "or               $t0, $t0, $t1                    \n"
+      "bnez             $t0, 11f                         \n"
+      " subu            $t7, $t9, %[src_stride]          \n"
+//dst + dst_stride word aligned
+      "1:                                                \n"
+      "lw               $t0, 0(%[src])                   \n"
+      "lwx              $t1, %[src_stride](%[src])       \n"
+      "lwx              $t8, $t2(%[src])                 \n"
+      "lwx              $t9, $t3(%[src])                 \n"
+
+// t0 = | 30 | 20 | 10 | 00 |
+// t1 = | 31 | 21 | 11 | 01 |
+// t8 = | 32 | 22 | 12 | 02 |
+// t9 = | 33 | 23 | 13 | 03 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 21 | 01 | 20 | 00 |
+  // s1 = | 23 | 03 | 22 | 02 |
+  // s2 = | 31 | 11 | 30 | 10 |
+  // s3 = | 33 | 13 | 32 | 12 |
+
+      "precr.qb.ph     $s4, $s1, $s0                     \n"
+      "precrq.qb.ph    $s5, $s1, $s0                     \n"
+      "precr.qb.ph     $s6, $s3, $s2                     \n"
+      "precrq.qb.ph    $s7, $s3, $s2                     \n"
+
+  // s4 = | 03 | 02 | 01 | 00 |
+  // s5 = | 23 | 22 | 21 | 20 |
+  // s6 = | 13 | 12 | 11 | 10 |
+  // s7 = | 33 | 32 | 31 | 30 |
+
+      "lwx              $t0, $t4(%[src])                 \n"
+      "lwx              $t1, $t5(%[src])                 \n"
+      "lwx              $t8, $t6(%[src])                 \n"
+      "lwx              $t9, $t7(%[src])                 \n"
+
+// t0 = | 34 | 24 | 14 | 04 |
+// t1 = | 35 | 25 | 15 | 05 |
+// t8 = | 36 | 26 | 16 | 06 |
+// t9 = | 37 | 27 | 17 | 07 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 25 | 05 | 24 | 04 |
+  // s1 = | 27 | 07 | 26 | 06 |
+  // s2 = | 35 | 15 | 34 | 14 |
+  // s3 = | 37 | 17 | 36 | 16 |
+
+      "precr.qb.ph     $t0, $s1, $s0                     \n"
+      "precrq.qb.ph    $t1, $s1, $s0                     \n"
+      "precr.qb.ph     $t8, $s3, $s2                     \n"
+      "precrq.qb.ph    $t9, $s3, $s2                     \n"
+
+  // t0 = | 07 | 06 | 05 | 04 |
+  // t1 = | 27 | 26 | 25 | 24 |
+  // t8 = | 17 | 16 | 15 | 14 |
+  // t9 = | 37 | 36 | 35 | 34 |
+
+      "addu            $s0, %[dst], %[dst_stride]        \n"
+      "addu            $s1, $s0, %[dst_stride]           \n"
+      "addu            $s2, $s1, %[dst_stride]           \n"
+
+      "sw              $s4, 0(%[dst])                    \n"
+      "sw              $t0, 4(%[dst])                    \n"
+      "sw              $s6, 0($s0)                       \n"
+      "sw              $t8, 4($s0)                       \n"
+      "sw              $s5, 0($s1)                       \n"
+      "sw              $t1, 4($s1)                       \n"
+      "sw              $s7, 0($s2)                       \n"
+      "sw              $t9, 4($s2)                       \n"
+
+      "addiu            $AT, -1                          \n"
+      "addiu            %[src], 4                        \n"
+
+      "bnez             $AT, 1b                          \n"
+      " addu            %[dst], $s2, %[dst_stride]       \n"
+      "b                2f                               \n"
+//dst + dst_stride unaligned
+      "11:                                               \n"
+      "lw               $t0, 0(%[src])                   \n"
+      "lwx              $t1, %[src_stride](%[src])       \n"
+      "lwx              $t8, $t2(%[src])                 \n"
+      "lwx              $t9, $t3(%[src])                 \n"
+
+// t0 = | 30 | 20 | 10 | 00 |
+// t1 = | 31 | 21 | 11 | 01 |
+// t8 = | 32 | 22 | 12 | 02 |
+// t9 = | 33 | 23 | 13 | 03 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 21 | 01 | 20 | 00 |
+  // s1 = | 23 | 03 | 22 | 02 |
+  // s2 = | 31 | 11 | 30 | 10 |
+  // s3 = | 33 | 13 | 32 | 12 |
+
+      "precr.qb.ph     $s4, $s1, $s0                     \n"
+      "precrq.qb.ph    $s5, $s1, $s0                     \n"
+      "precr.qb.ph     $s6, $s3, $s2                     \n"
+      "precrq.qb.ph    $s7, $s3, $s2                     \n"
+
+  // s4 = | 03 | 02 | 01 | 00 |
+  // s5 = | 23 | 22 | 21 | 20 |
+  // s6 = | 13 | 12 | 11 | 10 |
+  // s7 = | 33 | 32 | 31 | 30 |
+
+      "lwx              $t0, $t4(%[src])                 \n"
+      "lwx              $t1, $t5(%[src])                 \n"
+      "lwx              $t8, $t6(%[src])                 \n"
+      "lwx              $t9, $t7(%[src])                 \n"
+
+// t0 = | 34 | 24 | 14 | 04 |
+// t1 = | 35 | 25 | 15 | 05 |
+// t8 = | 36 | 26 | 16 | 06 |
+// t9 = | 37 | 27 | 17 | 07 |
+
+      "precr.qb.ph     $s0, $t1, $t0                     \n"
+      "precr.qb.ph     $s1, $t9, $t8                     \n"
+      "precrq.qb.ph    $s2, $t1, $t0                     \n"
+      "precrq.qb.ph    $s3, $t9, $t8                     \n"
+
+  // s0 = | 25 | 05 | 24 | 04 |
+  // s1 = | 27 | 07 | 26 | 06 |
+  // s2 = | 35 | 15 | 34 | 14 |
+  // s3 = | 37 | 17 | 36 | 16 |
+
+      "precr.qb.ph     $t0, $s1, $s0                     \n"
+      "precrq.qb.ph    $t1, $s1, $s0                     \n"
+      "precr.qb.ph     $t8, $s3, $s2                     \n"
+      "precrq.qb.ph    $t9, $s3, $s2                     \n"
+
+  // t0 = | 07 | 06 | 05 | 04 |
+  // t1 = | 27 | 26 | 25 | 24 |
+  // t8 = | 17 | 16 | 15 | 14 |
+  // t9 = | 37 | 36 | 35 | 34 |
+
+      "addu            $s0, %[dst], %[dst_stride]        \n"
+      "addu            $s1, $s0, %[dst_stride]           \n"
+      "addu            $s2, $s1, %[dst_stride]           \n"
+
+      "swr              $s4, 0(%[dst])                   \n"
+      "swl              $s4, 3(%[dst])                   \n"
+      "swr              $t0, 4(%[dst])                   \n"
+      "swl              $t0, 7(%[dst])                   \n"
+      "swr              $s6, 0($s0)                      \n"
+      "swl              $s6, 3($s0)                      \n"
+      "swr              $t8, 4($s0)                      \n"
+      "swl              $t8, 7($s0)                      \n"
+      "swr              $s5, 0($s1)                      \n"
+      "swl              $s5, 3($s1)                      \n"
+      "swr              $t1, 4($s1)                      \n"
+      "swl              $t1, 7($s1)                      \n"
+      "swr              $s7, 0($s2)                      \n"
+      "swl              $s7, 3($s2)                      \n"
+      "swr              $t9, 4($s2)                      \n"
+      "swl              $t9, 7($s2)                      \n"
+
+      "addiu            $AT, -1                          \n"
+      "addiu            %[src], 4                        \n"
+
+      "bnez             $AT, 11b                         \n"
+      " addu            %[dst], $s2, %[dst_stride]       \n"
+      "2:                                                \n"
+      ".set pop                                          \n"
+      ".set at                                           \n"
+      :[src] "+r" (src),
+       [dst] "+r" (dst),
+       [width] "+r" (width)
+      :[src_stride] "r" (src_stride),
+       [dst_stride] "r" (dst_stride)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7", "t8", "t9",
+        "s0", "s1", "s2", "s3", "s4", "s5", "s6", "s7"
+  );
+}
+
+void TransposeUVWx8_MIPS_DSPR2(const uint8* src, int src_stride,
+                               uint8* dst_a, int dst_stride_a,
+                               uint8* dst_b, int dst_stride_b,
+                               int width) {
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "beqz            %[width], 2f                      \n"
+      " sll            $t2, %[src_stride], 0x1           \n" // src_stride x 2
+      "sll             $t4, %[src_stride], 0x2           \n" // src_stride x 4
+      "sll             $t9, %[src_stride], 0x3           \n" // src_stride x 8
+      "addu            $t3, $t2, %[src_stride]           \n"
+      "addu            $t5, $t4, %[src_stride]           \n"
+      "addu            $t6, $t2, $t4                     \n"
+      "subu            $t7, $t9, %[src_stride]           \n"
+      "srl             $t1, %[width], 1                  \n"
+
+// check word aligment for dst_a, dst_b, dst_stride_a and dst_stride_b
+      "andi            $t0, %[dst_a], 0x3                \n"
+      "andi            $t8, %[dst_b], 0x3                \n"
+      "or              $t0, $t0, $t8                     \n"
+      "andi            $t8, %[dst_stride_a], 0x3         \n"
+      "andi            $s5, %[dst_stride_b], 0x3         \n"
+      "or              $t8, $t8, $s5                     \n"
+      "or              $t0, $t0, $t8                     \n"
+      "bnez            $t0, 11f                          \n"
+      " nop                                              \n"
+// dst + dst_stride word aligned (both, a & b dst addresses)
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src])                    \n" // |B0|A0|b0|a0|
+      "lwx             $t8, %[src_stride](%[src])        \n" // |B1|A1|b1|a1|
+      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
+      "lwx             $t9, $t2(%[src])                  \n" // |B2|A2|b2|a2|
+      "lwx             $s0, $t3(%[src])                  \n" // |B3|A3|b3|a3|
+      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B1|A1|B0|A0|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B3|A3|B2|A2|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A3|A2|A1|A0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B3|B2|B1|B0|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b1|a1|b0|a0|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b3|a3|b2|a2|
+
+      "sw              $s3, 0($s5)                       \n"
+      "sw              $s4, 0($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a3|a2|a1|a0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b3|b2|b1|b0|
+
+      "lwx             $t0, $t4(%[src])                  \n" // |B4|A4|b4|a4|
+      "lwx             $t8, $t5(%[src])                  \n" // |B5|A5|b5|a5|
+      "lwx             $t9, $t6(%[src])                  \n" // |B6|A6|b6|a6|
+      "lwx             $s0, $t7(%[src])                  \n" // |B7|A7|b7|a7|
+      "sw              $s3, 0(%[dst_a])                  \n"
+      "sw              $s4, 0(%[dst_b])                  \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B5|A5|B4|A4|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B6|A6|B7|A7|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A7|A6|A5|A4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B7|B6|B5|B4|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b5|a5|b4|a4|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b7|a7|b6|a6|
+      "sw              $s3, 4($s5)                       \n"
+      "sw              $s4, 4($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a7|a6|a5|a4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b7|b6|b5|b4|
+
+      "addiu           %[src], 4                         \n"
+      "addiu           $t1, -1                           \n"
+      "sll             $t0, %[dst_stride_a], 1           \n"
+      "sll             $t8, %[dst_stride_b], 1           \n"
+      "sw              $s3, 4(%[dst_a])                  \n"
+      "sw              $s4, 4(%[dst_b])                  \n"
+      "addu            %[dst_a], %[dst_a], $t0           \n"
+      "bnez            $t1, 1b                           \n"
+      " addu           %[dst_b], %[dst_b], $t8           \n"
+      "b               2f                                \n"
+      " nop                                              \n"
+
+// dst_a or dst_b or dst_stride_a or dst_stride_b not word aligned
+   "11:                                                  \n"
+      "lw              $t0, 0(%[src])                    \n" // |B0|A0|b0|a0|
+      "lwx             $t8, %[src_stride](%[src])        \n" // |B1|A1|b1|a1|
+      "addu            $s5, %[dst_a], %[dst_stride_a]    \n"
+      "lwx             $t9, $t2(%[src])                  \n" // |B2|A2|b2|a2|
+      "lwx             $s0, $t3(%[src])                  \n" // |B3|A3|b3|a3|
+      "addu            $s6, %[dst_b], %[dst_stride_b]    \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B1|A1|B0|A0|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B3|A3|B2|A2|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A3|A2|A1|A0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B3|B2|B1|B0|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b1|a1|b0|a0|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b3|a3|b2|a2|
+
+      "swr             $s3, 0($s5)                       \n"
+      "swl             $s3, 3($s5)                       \n"
+      "swr             $s4, 0($s6)                       \n"
+      "swl             $s4, 3($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a3|a2|a1|a0|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b3|b2|b1|b0|
+
+      "lwx             $t0, $t4(%[src])                  \n" // |B4|A4|b4|a4|
+      "lwx             $t8, $t5(%[src])                  \n" // |B5|A5|b5|a5|
+      "lwx             $t9, $t6(%[src])                  \n" // |B6|A6|b6|a6|
+      "lwx             $s0, $t7(%[src])                  \n" // |B7|A7|b7|a7|
+      "swr             $s3, 0(%[dst_a])                  \n"
+      "swl             $s3, 3(%[dst_a])                  \n"
+      "swr             $s4, 0(%[dst_b])                  \n"
+      "swl             $s4, 3(%[dst_b])                  \n"
+
+      "precrq.ph.w     $s1, $t8, $t0                     \n" // |B5|A5|B4|A4|
+      "precrq.ph.w     $s2, $s0, $t9                     \n" // |B6|A6|B7|A7|
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |A7|A6|A5|A4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |B7|B6|B5|B4|
+
+      "sll             $t0, $t0, 16                      \n"
+      "packrl.ph       $s1, $t8, $t0                     \n" // |b5|a5|b4|a4|
+      "sll             $t9, $t9, 16                      \n"
+      "packrl.ph       $s2, $s0, $t9                     \n" // |b7|a7|b6|a6|
+
+      "swr             $s3, 4($s5)                       \n"
+      "swl             $s3, 7($s5)                       \n"
+      "swr             $s4, 4($s6)                       \n"
+      "swl             $s4, 7($s6)                       \n"
+
+      "precr.qb.ph     $s3, $s2, $s1                     \n" // |a7|a6|a5|a4|
+      "precrq.qb.ph    $s4, $s2, $s1                     \n" // |b7|b6|b5|b4|
+
+      "addiu           %[src], 4                         \n"
+      "addiu           $t1, -1                           \n"
+      "sll             $t0, %[dst_stride_a], 1           \n"
+      "sll             $t8, %[dst_stride_b], 1           \n"
+      "swr             $s3, 4(%[dst_a])                  \n"
+      "swl             $s3, 7(%[dst_a])                  \n"
+      "swr             $s4, 4(%[dst_b])                  \n"
+      "swl             $s4, 7(%[dst_b])                  \n"
+      "addu            %[dst_a], %[dst_a], $t0           \n"
+      "bnez            $t1, 11b                          \n"
+      " addu           %[dst_b], %[dst_b], $t8           \n"
+
+      "2:                                                \n"
+      ".set pop                                          \n"
+      : [src] "+r" (src),
+        [dst_a] "+r" (dst_a),
+        [dst_b] "+r" (dst_b),
+        [width] "+r" (width),
+        [src_stride] "+r" (src_stride)
+      : [dst_stride_a] "r" (dst_stride_a),
+        [dst_stride_b] "r" (dst_stride_b)
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+        "t6", "t7", "t8", "t9",
+        "s0", "s1", "s2", "s3",
+        "s4", "s5", "s6"
+  );
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/rotate_neon.cc b/third_party/aom/third_party/libyuv/source/rotate_neon.cc
new file mode 100644
index 0000000000..76043b3b3c
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/rotate_neon.cc
@@ -0,0 +1,535 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+static uvec8 kVTbl4x4Transpose =
+  { 0,  4,  8, 12,  1,  5,  9, 13,  2,  6, 10, 14,  3,  7, 11, 15 };
+
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride,
+                       int width) {
+  const uint8* src_temp = NULL;
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub         %5, #8                        \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    ".p2align  2                               \n"
+    "1:                                        \n"
+      "mov         %0, %1                      \n"
+
+      MEMACCESS(0)
+      "vld1.8      {d0}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d1}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d2}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d3}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d4}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d5}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d6}, [%0], %2              \n"
+      MEMACCESS(0)
+      "vld1.8      {d7}, [%0]                  \n"
+
+      "vtrn.8      d1, d0                      \n"
+      "vtrn.8      d3, d2                      \n"
+      "vtrn.8      d5, d4                      \n"
+      "vtrn.8      d7, d6                      \n"
+
+      "vtrn.16     d1, d3                      \n"
+      "vtrn.16     d0, d2                      \n"
+      "vtrn.16     d5, d7                      \n"
+      "vtrn.16     d4, d6                      \n"
+
+      "vtrn.32     d1, d5                      \n"
+      "vtrn.32     d0, d4                      \n"
+      "vtrn.32     d3, d7                      \n"
+      "vtrn.32     d2, d6                      \n"
+
+      "vrev16.8    q0, q0                      \n"
+      "vrev16.8    q1, q1                      \n"
+      "vrev16.8    q2, q2                      \n"
+      "vrev16.8    q3, q3                      \n"
+
+      "mov         %0, %3                      \n"
+
+    MEMACCESS(0)
+      "vst1.8      {d1}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d0}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d3}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d2}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d5}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d4}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d7}, [%0], %4              \n"
+    MEMACCESS(0)
+      "vst1.8      {d6}, [%0]                  \n"
+
+      "add         %1, #8                      \n"  // src += 8
+      "add         %3, %3, %4, lsl #3          \n"  // dst += 8 * dst_stride
+      "subs        %5,  #8                     \n"  // w   -= 8
+      "bge         1b                          \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %5, #8                        \n"
+    "beq         4f                            \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %5, #2                        \n"
+    "blt         3f                            \n"
+
+    "cmp         %5, #4                        \n"
+    "blt         2f                            \n"
+
+    // 4x8 block
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld1.32     {d0[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d0[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d1[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d1[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d2[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d2[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d3[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.32     {d3[1]}, [%0]                 \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(6)
+    "vld1.8      {q3}, [%6]                    \n"
+
+    "vtbl.8      d4, {d0, d1}, d6              \n"
+    "vtbl.8      d5, {d0, d1}, d7              \n"
+    "vtbl.8      d0, {d2, d3}, d6              \n"
+    "vtbl.8      d1, {d2, d3}, d7              \n"
+
+    // TODO(frkoenig): Rework shuffle above to
+    // write out with 4 instead of 8 writes.
+    MEMACCESS(0)
+    "vst1.32     {d4[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d4[1]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d5[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d5[1]}, [%0]                 \n"
+
+    "add         %0, %3, #4                    \n"
+    MEMACCESS(0)
+    "vst1.32     {d0[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d0[1]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d1[0]}, [%0], %4             \n"
+    MEMACCESS(0)
+    "vst1.32     {d1[1]}, [%0]                 \n"
+
+    "add         %1, #4                        \n"  // src += 4
+    "add         %3, %3, %4, lsl #2            \n"  // dst += 4 * dst_stride
+    "subs        %5,  #4                       \n"  // w   -= 4
+    "beq         4f                            \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %5, #2                        \n"
+    "blt         3f                            \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[0]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[1]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[2]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[2]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d0[3]}, [%0], %2             \n"
+    MEMACCESS(0)
+    "vld1.16     {d1[3]}, [%0]                 \n"
+
+    "vtrn.8      d0, d1                        \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(0)
+    "vst1.64     {d0}, [%0], %4                \n"
+    MEMACCESS(0)
+    "vst1.64     {d1}, [%0]                    \n"
+
+    "add         %1, #2                        \n"  // src += 2
+    "add         %3, %3, %4, lsl #1            \n"  // dst += 2 * dst_stride
+    "subs        %5,  #2                       \n"  // w   -= 2
+    "beq         4f                            \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[0]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[1]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[2]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[3]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[4]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[5]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[6]}, [%1], %2             \n"
+    MEMACCESS(1)
+    "vld1.8      {d0[7]}, [%1]                 \n"
+
+    MEMACCESS(3)
+    "vst1.64     {d0}, [%3]                    \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),          // %0
+      "+r"(src),               // %1
+      "+r"(src_stride),        // %2
+      "+r"(dst),               // %3
+      "+r"(dst_stride),        // %4
+      "+r"(width)              // %5
+    : "r"(&kVTbl4x4Transpose)  // %6
+    : "memory", "cc", "q0", "q1", "q2", "q3"
+  );
+}
+
+static uvec8 kVTbl4x4TransposeDi =
+  { 0,  8,  1,  9,  2, 10,  3, 11,  4, 12,  5, 13,  6, 14,  7, 15 };
+
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) {
+  const uint8* src_temp = NULL;
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub         %7, #8                        \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    ".p2align  2                               \n"
+    "1:                                        \n"
+      "mov         %0, %1                      \n"
+
+      MEMACCESS(0)
+      "vld2.8      {d0,  d1},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d2,  d3},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d4,  d5},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d6,  d7},  [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d16, d17}, [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d18, d19}, [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d20, d21}, [%0], %2        \n"
+      MEMACCESS(0)
+      "vld2.8      {d22, d23}, [%0]            \n"
+
+      "vtrn.8      q1, q0                      \n"
+      "vtrn.8      q3, q2                      \n"
+      "vtrn.8      q9, q8                      \n"
+      "vtrn.8      q11, q10                    \n"
+
+      "vtrn.16     q1, q3                      \n"
+      "vtrn.16     q0, q2                      \n"
+      "vtrn.16     q9, q11                     \n"
+      "vtrn.16     q8, q10                     \n"
+
+      "vtrn.32     q1, q9                      \n"
+      "vtrn.32     q0, q8                      \n"
+      "vtrn.32     q3, q11                     \n"
+      "vtrn.32     q2, q10                     \n"
+
+      "vrev16.8    q0, q0                      \n"
+      "vrev16.8    q1, q1                      \n"
+      "vrev16.8    q2, q2                      \n"
+      "vrev16.8    q3, q3                      \n"
+      "vrev16.8    q8, q8                      \n"
+      "vrev16.8    q9, q9                      \n"
+      "vrev16.8    q10, q10                    \n"
+      "vrev16.8    q11, q11                    \n"
+
+      "mov         %0, %3                      \n"
+
+    MEMACCESS(0)
+      "vst1.8      {d2},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d0},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d6},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d4},  [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d18}, [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d16}, [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d22}, [%0], %4             \n"
+    MEMACCESS(0)
+      "vst1.8      {d20}, [%0]                 \n"
+
+      "mov         %0, %5                      \n"
+
+    MEMACCESS(0)
+      "vst1.8      {d3},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d1},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d7},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d5},  [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d19}, [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d17}, [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d23}, [%0], %6             \n"
+    MEMACCESS(0)
+      "vst1.8      {d21}, [%0]                 \n"
+
+      "add         %1, #8*2                    \n"  // src   += 8*2
+      "add         %3, %3, %4, lsl #3          \n"  // dst_a += 8 * dst_stride_a
+      "add         %5, %5, %6, lsl #3          \n"  // dst_b += 8 * dst_stride_b
+      "subs        %7,  #8                     \n"  // w     -= 8
+      "bge         1b                          \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %7, #8                        \n"
+    "beq         4f                            \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %7, #2                        \n"
+    "blt         3f                            \n"
+
+    "cmp         %7, #4                        \n"
+    "blt         2f                            \n"
+
+    // TODO(frkoenig): Clean this up
+    // 4x8 block
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld1.64     {d0}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d1}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d2}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d3}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d4}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d5}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d6}, [%0], %2                \n"
+    MEMACCESS(0)
+    "vld1.64     {d7}, [%0]                    \n"
+
+    MEMACCESS(8)
+    "vld1.8      {q15}, [%8]                   \n"
+
+    "vtrn.8      q0, q1                        \n"
+    "vtrn.8      q2, q3                        \n"
+
+    "vtbl.8      d16, {d0, d1}, d30            \n"
+    "vtbl.8      d17, {d0, d1}, d31            \n"
+    "vtbl.8      d18, {d2, d3}, d30            \n"
+    "vtbl.8      d19, {d2, d3}, d31            \n"
+    "vtbl.8      d20, {d4, d5}, d30            \n"
+    "vtbl.8      d21, {d4, d5}, d31            \n"
+    "vtbl.8      d22, {d6, d7}, d30            \n"
+    "vtbl.8      d23, {d6, d7}, d31            \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(0)
+    "vst1.32     {d16[0]},  [%0], %4           \n"
+    MEMACCESS(0)
+    "vst1.32     {d16[1]},  [%0], %4           \n"
+    MEMACCESS(0)
+    "vst1.32     {d17[0]},  [%0], %4           \n"
+    MEMACCESS(0)
+    "vst1.32     {d17[1]},  [%0], %4           \n"
+
+    "add         %0, %3, #4                    \n"
+    MEMACCESS(0)
+    "vst1.32     {d20[0]}, [%0], %4            \n"
+    MEMACCESS(0)
+    "vst1.32     {d20[1]}, [%0], %4            \n"
+    MEMACCESS(0)
+    "vst1.32     {d21[0]}, [%0], %4            \n"
+    MEMACCESS(0)
+    "vst1.32     {d21[1]}, [%0]                \n"
+
+    "mov         %0, %5                        \n"
+
+    MEMACCESS(0)
+    "vst1.32     {d18[0]}, [%0], %6            \n"
+    MEMACCESS(0)
+    "vst1.32     {d18[1]}, [%0], %6            \n"
+    MEMACCESS(0)
+    "vst1.32     {d19[0]}, [%0], %6            \n"
+    MEMACCESS(0)
+    "vst1.32     {d19[1]}, [%0], %6            \n"
+
+    "add         %0, %5, #4                    \n"
+    MEMACCESS(0)
+    "vst1.32     {d22[0]},  [%0], %6           \n"
+    MEMACCESS(0)
+    "vst1.32     {d22[1]},  [%0], %6           \n"
+    MEMACCESS(0)
+    "vst1.32     {d23[0]},  [%0], %6           \n"
+    MEMACCESS(0)
+    "vst1.32     {d23[1]},  [%0]               \n"
+
+    "add         %1, #4*2                      \n"  // src   += 4 * 2
+    "add         %3, %3, %4, lsl #2            \n"  // dst_a += 4 * dst_stride_a
+    "add         %5, %5, %6, lsl #2            \n"  // dst_b += 4 * dst_stride_b
+    "subs        %7,  #4                       \n"  // w     -= 4
+    "beq         4f                            \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %7, #2                        \n"
+    "blt         3f                            \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov         %0, %1                        \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[0], d2[0]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[0], d3[0]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[1], d2[1]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[1], d3[1]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[2], d2[2]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[2], d3[2]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d0[3], d2[3]}, [%0], %2      \n"
+    MEMACCESS(0)
+    "vld2.16     {d1[3], d3[3]}, [%0]          \n"
+
+    "vtrn.8      d0, d1                        \n"
+    "vtrn.8      d2, d3                        \n"
+
+    "mov         %0, %3                        \n"
+
+    MEMACCESS(0)
+    "vst1.64     {d0}, [%0], %4                \n"
+    MEMACCESS(0)
+    "vst1.64     {d2}, [%0]                    \n"
+
+    "mov         %0, %5                        \n"
+
+    MEMACCESS(0)
+    "vst1.64     {d1}, [%0], %6                \n"
+    MEMACCESS(0)
+    "vst1.64     {d3}, [%0]                    \n"
+
+    "add         %1, #2*2                      \n"  // src   += 2 * 2
+    "add         %3, %3, %4, lsl #1            \n"  // dst_a += 2 * dst_stride_a
+    "add         %5, %5, %6, lsl #1            \n"  // dst_b += 2 * dst_stride_b
+    "subs        %7,  #2                       \n"  // w     -= 2
+    "beq         4f                            \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[0], d1[0]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[1], d1[1]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[2], d1[2]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[3], d1[3]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[4], d1[4]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[5], d1[5]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[6], d1[6]}, [%1], %2      \n"
+    MEMACCESS(1)
+    "vld2.8      {d0[7], d1[7]}, [%1]          \n"
+
+    MEMACCESS(3)
+    "vst1.64     {d0}, [%3]                    \n"
+    MEMACCESS(5)
+    "vst1.64     {d1}, [%5]                    \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),            // %0
+      "+r"(src),                 // %1
+      "+r"(src_stride),          // %2
+      "+r"(dst_a),               // %3
+      "+r"(dst_stride_a),        // %4
+      "+r"(dst_b),               // %5
+      "+r"(dst_stride_b),        // %6
+      "+r"(width)                // %7
+    : "r"(&kVTbl4x4TransposeDi)  // %8
+    : "memory", "cc",
+      "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  );
+}
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/rotate_neon64.cc b/third_party/aom/third_party/libyuv/source/rotate_neon64.cc
new file mode 100644
index 0000000000..f52c082b3f
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/rotate_neon64.cc
@@ -0,0 +1,543 @@
+/*
+ *  Copyright 2014 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon armv8 64 bit.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+static uvec8 kVTbl4x4Transpose =
+  { 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15 };
+
+void TransposeWx8_NEON(const uint8* src, int src_stride,
+                       uint8* dst, int dst_stride, int width) {
+  const uint8* src_temp = NULL;
+  int64 width64 = (int64) width;  // Work around clang 3.4 warning.
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub         %3, %3, #8                      \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    "1:                                          \n"
+      "mov         %0, %1                        \n"
+
+      MEMACCESS(0)
+      "ld1        {v0.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v1.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v2.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v3.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v4.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v5.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v6.8b}, [%0], %5              \n"
+      MEMACCESS(0)
+      "ld1        {v7.8b}, [%0]                  \n"
+
+      "trn2     v16.8b, v0.8b, v1.8b             \n"
+      "trn1     v17.8b, v0.8b, v1.8b             \n"
+      "trn2     v18.8b, v2.8b, v3.8b             \n"
+      "trn1     v19.8b, v2.8b, v3.8b             \n"
+      "trn2     v20.8b, v4.8b, v5.8b             \n"
+      "trn1     v21.8b, v4.8b, v5.8b             \n"
+      "trn2     v22.8b, v6.8b, v7.8b             \n"
+      "trn1     v23.8b, v6.8b, v7.8b             \n"
+
+      "trn2     v3.4h, v17.4h, v19.4h            \n"
+      "trn1     v1.4h, v17.4h, v19.4h            \n"
+      "trn2     v2.4h, v16.4h, v18.4h            \n"
+      "trn1     v0.4h, v16.4h, v18.4h            \n"
+      "trn2     v7.4h, v21.4h, v23.4h            \n"
+      "trn1     v5.4h, v21.4h, v23.4h            \n"
+      "trn2     v6.4h, v20.4h, v22.4h            \n"
+      "trn1     v4.4h, v20.4h, v22.4h            \n"
+
+      "trn2     v21.2s, v1.2s, v5.2s             \n"
+      "trn1     v17.2s, v1.2s, v5.2s             \n"
+      "trn2     v20.2s, v0.2s, v4.2s             \n"
+      "trn1     v16.2s, v0.2s, v4.2s             \n"
+      "trn2     v23.2s, v3.2s, v7.2s             \n"
+      "trn1     v19.2s, v3.2s, v7.2s             \n"
+      "trn2     v22.2s, v2.2s, v6.2s             \n"
+      "trn1     v18.2s, v2.2s, v6.2s             \n"
+
+      "mov         %0, %2                        \n"
+
+    MEMACCESS(0)
+      "st1      {v17.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v16.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v19.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v18.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v21.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v20.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v23.8b}, [%0], %6               \n"
+    MEMACCESS(0)
+      "st1      {v22.8b}, [%0]                   \n"
+
+      "add         %1, %1, #8                    \n"  // src += 8
+      "add         %2, %2, %6, lsl #3            \n"  // dst += 8 * dst_stride
+      "subs        %3, %3, #8                    \n"  // w   -= 8
+      "b.ge        1b                            \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds        %3, %3, #8                      \n"
+    "b.eq        4f                              \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp         %3, #2                          \n"
+    "b.lt        3f                              \n"
+
+    "cmp         %3, #4                          \n"
+    "b.lt        2f                              \n"
+
+    // 4x8 block
+    "mov         %0, %1                          \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.s}[3], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.s}[3], [%0]                     \n"
+
+    "mov         %0, %2                          \n"
+
+    MEMACCESS(4)
+    "ld1      {v2.16b}, [%4]                     \n"
+
+    "tbl      v3.16b, {v0.16b}, v2.16b           \n"
+    "tbl      v0.16b, {v1.16b}, v2.16b           \n"
+
+    // TODO(frkoenig): Rework shuffle above to
+    // write out with 4 instead of 8 writes.
+    MEMACCESS(0)
+    "st1 {v3.s}[0], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v3.s}[1], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v3.s}[2], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v3.s}[3], [%0]                         \n"
+
+    "add         %0, %2, #4                      \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[0], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[1], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[2], [%0], %6                     \n"
+    MEMACCESS(0)
+    "st1 {v0.s}[3], [%0]                         \n"
+
+    "add         %1, %1, #4                      \n"  // src += 4
+    "add         %2, %2, %6, lsl #2              \n"  // dst += 4 * dst_stride
+    "subs        %3, %3, #4                      \n"  // w   -= 4
+    "b.eq        4f                              \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp         %3, #2                          \n"
+    "b.lt        3f                              \n"
+
+    // 2x8 block
+    "2:                                          \n"
+    "mov         %0, %1                          \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[0], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[1], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[2], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v0.h}[3], [%0], %5                 \n"
+    MEMACCESS(0)
+    "ld1     {v1.h}[3], [%0]                     \n"
+
+    "trn2    v2.8b, v0.8b, v1.8b                 \n"
+    "trn1    v3.8b, v0.8b, v1.8b                 \n"
+
+    "mov         %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1     {v3.8b}, [%0], %6                   \n"
+    MEMACCESS(0)
+    "st1     {v2.8b}, [%0]                       \n"
+
+    "add         %1, %1, #2                      \n"  // src += 2
+    "add         %2, %2, %6, lsl #1              \n"  // dst += 2 * dst_stride
+    "subs        %3, %3,  #2                     \n"  // w   -= 2
+    "b.eq        4f                              \n"
+
+    // 1x8 block
+    "3:                                          \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[0], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[1], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[2], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[3], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[4], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[5], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[6], [%1], %5             \n"
+    MEMACCESS(1)
+    "ld1         {v0.b}[7], [%1]                 \n"
+
+    MEMACCESS(2)
+    "st1         {v0.8b}, [%2]                   \n"
+
+    "4:                                          \n"
+
+    : "+r"(src_temp),                             // %0
+      "+r"(src),                                  // %1
+      "+r"(dst),                                  // %2
+      "+r"(width64)                               // %3
+    : "r"(&kVTbl4x4Transpose),                    // %4
+      "r"(static_cast<ptrdiff_t>(src_stride)),    // %5
+      "r"(static_cast<ptrdiff_t>(dst_stride))     // %6
+    : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16",
+      "v17", "v18", "v19", "v20", "v21", "v22", "v23"
+  );
+}
+
+static uint8 kVTbl4x4TransposeDi[32] =
+  { 0,  16, 32, 48,  2, 18, 34, 50,  4, 20, 36, 52,  6, 22, 38, 54,
+    1,  17, 33, 49,  3, 19, 35, 51,  5, 21, 37, 53,  7, 23, 39, 55};
+
+void TransposeUVWx8_NEON(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int width) {
+  const uint8* src_temp = NULL;
+  int64 width64 = (int64) width;  // Work around clang 3.4 warning.
+  asm volatile (
+    // loops are on blocks of 8. loop will stop when
+    // counter gets to or below 0. starting the counter
+    // at w-8 allow for this
+    "sub       %4, %4, #8                      \n"
+
+    // handle 8x8 blocks. this should be the majority of the plane
+    "1:                                        \n"
+    "mov       %0, %1                          \n"
+
+    MEMACCESS(0)
+    "ld1       {v0.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v1.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v2.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v3.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v4.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v5.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v6.16b}, [%0], %5              \n"
+    MEMACCESS(0)
+    "ld1       {v7.16b}, [%0]                  \n"
+
+    "trn1      v16.16b, v0.16b, v1.16b         \n"
+    "trn2      v17.16b, v0.16b, v1.16b         \n"
+    "trn1      v18.16b, v2.16b, v3.16b         \n"
+    "trn2      v19.16b, v2.16b, v3.16b         \n"
+    "trn1      v20.16b, v4.16b, v5.16b         \n"
+    "trn2      v21.16b, v4.16b, v5.16b         \n"
+    "trn1      v22.16b, v6.16b, v7.16b         \n"
+    "trn2      v23.16b, v6.16b, v7.16b         \n"
+
+    "trn1      v0.8h, v16.8h, v18.8h           \n"
+    "trn2      v1.8h, v16.8h, v18.8h           \n"
+    "trn1      v2.8h, v20.8h, v22.8h           \n"
+    "trn2      v3.8h, v20.8h, v22.8h           \n"
+    "trn1      v4.8h, v17.8h, v19.8h           \n"
+    "trn2      v5.8h, v17.8h, v19.8h           \n"
+    "trn1      v6.8h, v21.8h, v23.8h           \n"
+    "trn2      v7.8h, v21.8h, v23.8h           \n"
+
+    "trn1      v16.4s, v0.4s, v2.4s            \n"
+    "trn2      v17.4s, v0.4s, v2.4s            \n"
+    "trn1      v18.4s, v1.4s, v3.4s            \n"
+    "trn2      v19.4s, v1.4s, v3.4s            \n"
+    "trn1      v20.4s, v4.4s, v6.4s            \n"
+    "trn2      v21.4s, v4.4s, v6.4s            \n"
+    "trn1      v22.4s, v5.4s, v7.4s            \n"
+    "trn2      v23.4s, v5.4s, v7.4s            \n"
+
+    "mov       %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1       {v16.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v17.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v19.d}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v16.d}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.d}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v17.d}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v19.d}[1], [%0]                \n"
+
+    "mov       %0, %3                          \n"
+
+    MEMACCESS(0)
+    "st1       {v20.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v22.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v21.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v23.d}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v20.d}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v22.d}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v21.d}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v23.d}[1], [%0]                \n"
+
+    "add       %1, %1, #16                     \n"  // src   += 8*2
+    "add       %2, %2, %6, lsl #3              \n"  // dst_a += 8 * dst_stride_a
+    "add       %3, %3, %7, lsl #3              \n"  // dst_b += 8 * dst_stride_b
+    "subs      %4, %4,  #8                     \n"  // w     -= 8
+    "b.ge      1b                              \n"
+
+    // add 8 back to counter. if the result is 0 there are
+    // no residuals.
+    "adds      %4, %4, #8                      \n"
+    "b.eq      4f                              \n"
+
+    // some residual, so between 1 and 7 lines left to transpose
+    "cmp       %4, #2                          \n"
+    "b.lt      3f                              \n"
+
+    "cmp       %4, #4                          \n"
+    "b.lt      2f                              \n"
+
+    // TODO(frkoenig): Clean this up
+    // 4x8 block
+    "mov       %0, %1                          \n"
+    MEMACCESS(0)
+    "ld1       {v0.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v1.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v2.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v3.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v4.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v5.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v6.8b}, [%0], %5               \n"
+    MEMACCESS(0)
+    "ld1       {v7.8b}, [%0]                   \n"
+
+    MEMACCESS(8)
+    "ld1       {v30.16b}, [%8], #16            \n"
+    "ld1       {v31.16b}, [%8]                 \n"
+
+    "tbl       v16.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v30.16b  \n"
+    "tbl       v17.16b, {v0.16b, v1.16b, v2.16b, v3.16b}, v31.16b  \n"
+    "tbl       v18.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v30.16b  \n"
+    "tbl       v19.16b, {v4.16b, v5.16b, v6.16b, v7.16b}, v31.16b  \n"
+
+    "mov       %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1       {v16.s}[0],  [%0], %6           \n"
+    MEMACCESS(0)
+    "st1       {v16.s}[1],  [%0], %6           \n"
+    MEMACCESS(0)
+    "st1       {v16.s}[2],  [%0], %6           \n"
+    MEMACCESS(0)
+    "st1       {v16.s}[3],  [%0], %6           \n"
+
+    "add       %0, %2, #4                      \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[0], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[1], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[2], [%0], %6            \n"
+    MEMACCESS(0)
+    "st1       {v18.s}[3], [%0]                \n"
+
+    "mov       %0, %3                          \n"
+
+    MEMACCESS(0)
+    "st1       {v17.s}[0], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v17.s}[1], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v17.s}[2], [%0], %7            \n"
+    MEMACCESS(0)
+    "st1       {v17.s}[3], [%0], %7            \n"
+
+    "add       %0, %3, #4                      \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[0],  [%0], %7           \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[1],  [%0], %7           \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[2],  [%0], %7           \n"
+    MEMACCESS(0)
+    "st1       {v19.s}[3],  [%0]               \n"
+
+    "add       %1, %1, #8                      \n"  // src   += 4 * 2
+    "add       %2, %2, %6, lsl #2              \n"  // dst_a += 4 * dst_stride_a
+    "add       %3, %3, %7, lsl #2              \n"  // dst_b += 4 * dst_stride_b
+    "subs      %4,  %4,  #4                    \n"  // w     -= 4
+    "b.eq      4f                              \n"
+
+    // some residual, check to see if it includes a 2x8 block,
+    // or less
+    "cmp       %4, #2                          \n"
+    "b.lt      3f                              \n"
+
+    // 2x8 block
+    "2:                                        \n"
+    "mov       %0, %1                          \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[0], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[0], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[1], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[1], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[2], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[2], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v0.h, v1.h}[3], [%0], %5       \n"
+    MEMACCESS(0)
+    "ld2       {v2.h, v3.h}[3], [%0]           \n"
+
+    "trn1      v4.8b, v0.8b, v2.8b             \n"
+    "trn2      v5.8b, v0.8b, v2.8b             \n"
+    "trn1      v6.8b, v1.8b, v3.8b             \n"
+    "trn2      v7.8b, v1.8b, v3.8b             \n"
+
+    "mov       %0, %2                          \n"
+
+    MEMACCESS(0)
+    "st1       {v4.d}[0], [%0], %6             \n"
+    MEMACCESS(0)
+    "st1       {v6.d}[0], [%0]                 \n"
+
+    "mov       %0, %3                          \n"
+
+    MEMACCESS(0)
+    "st1       {v5.d}[0], [%0], %7             \n"
+    MEMACCESS(0)
+    "st1       {v7.d}[0], [%0]                 \n"
+
+    "add       %1, %1, #4                      \n"  // src   += 2 * 2
+    "add       %2, %2, %6, lsl #1              \n"  // dst_a += 2 * dst_stride_a
+    "add       %3, %3, %7, lsl #1              \n"  // dst_b += 2 * dst_stride_b
+    "subs      %4,  %4,  #2                    \n"  // w     -= 2
+    "b.eq      4f                              \n"
+
+    // 1x8 block
+    "3:                                        \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[0], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[1], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[2], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[3], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[4], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[5], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[6], [%1], %5       \n"
+    MEMACCESS(1)
+    "ld2       {v0.b, v1.b}[7], [%1]           \n"
+
+    MEMACCESS(2)
+    "st1       {v0.d}[0], [%2]                 \n"
+    MEMACCESS(3)
+    "st1       {v1.d}[0], [%3]                 \n"
+
+    "4:                                        \n"
+
+    : "+r"(src_temp),                             // %0
+      "+r"(src),                                  // %1
+      "+r"(dst_a),                                // %2
+      "+r"(dst_b),                                // %3
+      "+r"(width64)                               // %4
+    : "r"(static_cast<ptrdiff_t>(src_stride)),    // %5
+      "r"(static_cast<ptrdiff_t>(dst_stride_a)),  // %6
+      "r"(static_cast<ptrdiff_t>(dst_stride_b)),  // %7
+      "r"(&kVTbl4x4TransposeDi)                   // %8
+    : "memory", "cc",
+      "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+      "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23",
+      "v30", "v31"
+  );
+}
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/rotate_win.cc b/third_party/aom/third_party/libyuv/source/rotate_win.cc
new file mode 100644
index 0000000000..2760066dfd
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/rotate_win.cc
@@ -0,0 +1,248 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/rotate_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
+    defined(_MSC_VER) && !defined(__clang__)
+
+__declspec(naked)
+void TransposeWx8_SSSE3(const uint8* src, int src_stride,
+                        uint8* dst, int dst_stride, int width) {
+  __asm {
+    push      edi
+    push      esi
+    push      ebp
+    mov       eax, [esp + 12 + 4]   // src
+    mov       edi, [esp + 12 + 8]   // src_stride
+    mov       edx, [esp + 12 + 12]  // dst
+    mov       esi, [esp + 12 + 16]  // dst_stride
+    mov       ecx, [esp + 12 + 20]  // width
+
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    align      4
+ convertloop:
+    movq      xmm0, qword ptr [eax]
+    lea       ebp, [eax + 8]
+    movq      xmm1, qword ptr [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    punpcklbw xmm0, xmm1
+    movq      xmm2, qword ptr [eax]
+    movdqa    xmm1, xmm0
+    palignr   xmm1, xmm1, 8
+    movq      xmm3, qword ptr [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    punpcklbw xmm2, xmm3
+    movdqa    xmm3, xmm2
+    movq      xmm4, qword ptr [eax]
+    palignr   xmm3, xmm3, 8
+    movq      xmm5, qword ptr [eax + edi]
+    punpcklbw xmm4, xmm5
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm5, xmm4
+    movq      xmm6, qword ptr [eax]
+    palignr   xmm5, xmm5, 8
+    movq      xmm7, qword ptr [eax + edi]
+    punpcklbw xmm6, xmm7
+    mov       eax, ebp
+    movdqa    xmm7, xmm6
+    palignr   xmm7, xmm7, 8
+    // Second round of bit swap.
+    punpcklwd xmm0, xmm2
+    punpcklwd xmm1, xmm3
+    movdqa    xmm2, xmm0
+    movdqa    xmm3, xmm1
+    palignr   xmm2, xmm2, 8
+    palignr   xmm3, xmm3, 8
+    punpcklwd xmm4, xmm6
+    punpcklwd xmm5, xmm7
+    movdqa    xmm6, xmm4
+    movdqa    xmm7, xmm5
+    palignr   xmm6, xmm6, 8
+    palignr   xmm7, xmm7, 8
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    punpckldq xmm0, xmm4
+    movq      qword ptr [edx], xmm0
+    movdqa    xmm4, xmm0
+    palignr   xmm4, xmm4, 8
+    movq      qword ptr [edx + esi], xmm4
+    lea       edx, [edx + 2 * esi]
+    punpckldq xmm2, xmm6
+    movdqa    xmm6, xmm2
+    palignr   xmm6, xmm6, 8
+    movq      qword ptr [edx], xmm2
+    punpckldq xmm1, xmm5
+    movq      qword ptr [edx + esi], xmm6
+    lea       edx, [edx + 2 * esi]
+    movdqa    xmm5, xmm1
+    movq      qword ptr [edx], xmm1
+    palignr   xmm5, xmm5, 8
+    punpckldq xmm3, xmm7
+    movq      qword ptr [edx + esi], xmm5
+    lea       edx, [edx + 2 * esi]
+    movq      qword ptr [edx], xmm3
+    movdqa    xmm7, xmm3
+    palignr   xmm7, xmm7, 8
+    sub       ecx, 8
+    movq      qword ptr [edx + esi], xmm7
+    lea       edx, [edx + 2 * esi]
+    jg        convertloop
+
+    pop       ebp
+    pop       esi
+    pop       edi
+    ret
+  }
+}
+
+__declspec(naked)
+void TransposeUVWx8_SSE2(const uint8* src, int src_stride,
+                         uint8* dst_a, int dst_stride_a,
+                         uint8* dst_b, int dst_stride_b,
+                         int w) {
+  __asm {
+    push      ebx
+    push      esi
+    push      edi
+    push      ebp
+    mov       eax, [esp + 16 + 4]   // src
+    mov       edi, [esp + 16 + 8]   // src_stride
+    mov       edx, [esp + 16 + 12]  // dst_a
+    mov       esi, [esp + 16 + 16]  // dst_stride_a
+    mov       ebx, [esp + 16 + 20]  // dst_b
+    mov       ebp, [esp + 16 + 24]  // dst_stride_b
+    mov       ecx, esp
+    sub       esp, 4 + 16
+    and       esp, ~15
+    mov       [esp + 16], ecx
+    mov       ecx, [ecx + 16 + 28]  // w
+
+    align      4
+ convertloop:
+    // Read in the data from the source pointer.
+    // First round of bit swap.
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm0  // use xmm7 as temp register.
+    punpcklbw xmm0, xmm1
+    punpckhbw xmm7, xmm1
+    movdqa    xmm1, xmm7
+    movdqu    xmm2, [eax]
+    movdqu    xmm3, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm2
+    punpcklbw xmm2, xmm3
+    punpckhbw xmm7, xmm3
+    movdqa    xmm3, xmm7
+    movdqu    xmm4, [eax]
+    movdqu    xmm5, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqa    xmm7, xmm4
+    punpcklbw xmm4, xmm5
+    punpckhbw xmm7, xmm5
+    movdqa    xmm5, xmm7
+    movdqu    xmm6, [eax]
+    movdqu    xmm7, [eax + edi]
+    lea       eax, [eax + 2 * edi]
+    movdqu    [esp], xmm5  // backup xmm5
+    neg       edi
+    movdqa    xmm5, xmm6   // use xmm5 as temp register.
+    punpcklbw xmm6, xmm7
+    punpckhbw xmm5, xmm7
+    movdqa    xmm7, xmm5
+    lea       eax, [eax + 8 * edi + 16]
+    neg       edi
+    // Second round of bit swap.
+    movdqa    xmm5, xmm0
+    punpcklwd xmm0, xmm2
+    punpckhwd xmm5, xmm2
+    movdqa    xmm2, xmm5
+    movdqa    xmm5, xmm1
+    punpcklwd xmm1, xmm3
+    punpckhwd xmm5, xmm3
+    movdqa    xmm3, xmm5
+    movdqa    xmm5, xmm4
+    punpcklwd xmm4, xmm6
+    punpckhwd xmm5, xmm6
+    movdqa    xmm6, xmm5
+    movdqu    xmm5, [esp]  // restore xmm5
+    movdqu    [esp], xmm6  // backup xmm6
+    movdqa    xmm6, xmm5    // use xmm6 as temp register.
+    punpcklwd xmm5, xmm7
+    punpckhwd xmm6, xmm7
+    movdqa    xmm7, xmm6
+    // Third round of bit swap.
+    // Write to the destination pointer.
+    movdqa    xmm6, xmm0
+    punpckldq xmm0, xmm4
+    punpckhdq xmm6, xmm4
+    movdqa    xmm4, xmm6
+    movdqu    xmm6, [esp]  // restore xmm6
+    movlpd    qword ptr [edx], xmm0
+    movhpd    qword ptr [ebx], xmm0
+    movlpd    qword ptr [edx + esi], xmm4
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm4
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm2   // use xmm0 as the temp register.
+    punpckldq xmm2, xmm6
+    movlpd    qword ptr [edx], xmm2
+    movhpd    qword ptr [ebx], xmm2
+    punpckhdq xmm0, xmm6
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm1   // use xmm0 as the temp register.
+    punpckldq xmm1, xmm5
+    movlpd    qword ptr [edx], xmm1
+    movhpd    qword ptr [ebx], xmm1
+    punpckhdq xmm0, xmm5
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    movdqa    xmm0, xmm3   // use xmm0 as the temp register.
+    punpckldq xmm3, xmm7
+    movlpd    qword ptr [edx], xmm3
+    movhpd    qword ptr [ebx], xmm3
+    punpckhdq xmm0, xmm7
+    sub       ecx, 8
+    movlpd    qword ptr [edx + esi], xmm0
+    lea       edx, [edx + 2 * esi]
+    movhpd    qword ptr [ebx + ebp], xmm0
+    lea       ebx, [ebx + 2 * ebp]
+    jg        convertloop
+
+    mov       esp, [esp + 16]
+    pop       ebp
+    pop       edi
+    pop       esi
+    pop       ebx
+    ret
+  }
+}
+
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/row_any.cc b/third_party/aom/third_party/libyuv/source/row_any.cc
new file mode 100644
index 0000000000..1cb1f6b930
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/row_any.cc
@@ -0,0 +1,680 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include <string.h>  // For memset.
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Subsampled source needs to be increase by 1 of not even.
+#define SS(width, shift) (((width) + (1 << (shift)) - 1) >> (shift))
+
+// Any 3 planes to 1.
+#define ANY31(NAMEANY, ANY_SIMD, UVSHIFT, DUVSHIFT, BPP, MASK)                 \
+    void NAMEANY(const uint8* y_buf, const uint8* u_buf, const uint8* v_buf,   \
+                 uint8* dst_ptr, int width) {                                  \
+      SIMD_ALIGNED(uint8 temp[64 * 4]);                                        \
+      memset(temp, 0, 64 * 3);  /* for YUY2 and msan */                        \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(y_buf, u_buf, v_buf, dst_ptr, n);                             \
+      }                                                                        \
+      memcpy(temp, y_buf + n, r);                                              \
+      memcpy(temp + 64, u_buf + (n >> UVSHIFT), SS(r, UVSHIFT));               \
+      memcpy(temp + 128, v_buf + (n >> UVSHIFT), SS(r, UVSHIFT));              \
+      ANY_SIMD(temp, temp + 64, temp + 128, temp + 192, MASK + 1);             \
+      memcpy(dst_ptr + (n >> DUVSHIFT) * BPP, temp + 192,                      \
+             SS(r, DUVSHIFT) * BPP);                                           \
+    }
+
+#ifdef HAS_I422TOARGBROW_SSSE3
+ANY31(I422ToARGBRow_Any_SSSE3, I422ToARGBRow_SSSE3, 1, 0, 4, 7)
+#endif
+#ifdef HAS_I444TOARGBROW_SSSE3
+ANY31(I444ToARGBRow_Any_SSSE3, I444ToARGBRow_SSSE3, 0, 0, 4, 7)
+ANY31(I411ToARGBRow_Any_SSSE3, I411ToARGBRow_SSSE3, 2, 0, 4, 7)
+ANY31(I422ToBGRARow_Any_SSSE3, I422ToBGRARow_SSSE3, 1, 0, 4, 7)
+ANY31(I422ToABGRRow_Any_SSSE3, I422ToABGRRow_SSSE3, 1, 0, 4, 7)
+ANY31(I422ToRGBARow_Any_SSSE3, I422ToRGBARow_SSSE3, 1, 0, 4, 7)
+ANY31(I422ToARGB4444Row_Any_SSSE3, I422ToARGB4444Row_SSSE3, 1, 0, 2, 7)
+ANY31(I422ToARGB1555Row_Any_SSSE3, I422ToARGB1555Row_SSSE3, 1, 0, 2, 7)
+ANY31(I422ToRGB565Row_Any_SSSE3, I422ToRGB565Row_SSSE3, 1, 0, 2, 7)
+ANY31(I422ToRGB24Row_Any_SSSE3, I422ToRGB24Row_SSSE3, 1, 0, 3, 7)
+ANY31(I422ToRAWRow_Any_SSSE3, I422ToRAWRow_SSSE3, 1, 0, 3, 7)
+ANY31(I422ToYUY2Row_Any_SSE2, I422ToYUY2Row_SSE2, 1, 1, 4, 15)
+ANY31(I422ToUYVYRow_Any_SSE2, I422ToUYVYRow_SSE2, 1, 1, 4, 15)
+#endif  // HAS_I444TOARGBROW_SSSE3
+#ifdef HAS_I422TORGB24ROW_AVX2
+ANY31(I422ToRGB24Row_Any_AVX2, I422ToRGB24Row_AVX2, 1, 0, 3, 15)
+#endif
+#ifdef HAS_I422TORAWROW_AVX2
+ANY31(I422ToRAWRow_Any_AVX2, I422ToRAWRow_AVX2, 1, 0, 3, 15)
+#endif
+#ifdef HAS_J422TOARGBROW_SSSE3
+ANY31(J422ToARGBRow_Any_SSSE3, J422ToARGBRow_SSSE3, 1, 0, 4, 7)
+#endif
+#ifdef HAS_J422TOARGBROW_AVX2
+ANY31(J422ToARGBRow_Any_AVX2, J422ToARGBRow_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I422TOARGBROW_AVX2
+ANY31(I422ToARGBRow_Any_AVX2, I422ToARGBRow_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I422TOBGRAROW_AVX2
+ANY31(I422ToBGRARow_Any_AVX2, I422ToBGRARow_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I422TORGBAROW_AVX2
+ANY31(I422ToRGBARow_Any_AVX2, I422ToRGBARow_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I422TOABGRROW_AVX2
+ANY31(I422ToABGRRow_Any_AVX2, I422ToABGRRow_AVX2, 1, 0, 4, 15)
+#endif
+#ifdef HAS_I444TOARGBROW_AVX2
+ANY31(I444ToARGBRow_Any_AVX2, I444ToARGBRow_AVX2, 0, 0, 4, 15)
+#endif
+#ifdef HAS_I411TOARGBROW_AVX2
+ANY31(I411ToARGBRow_Any_AVX2, I411ToARGBRow_AVX2, 2, 0, 4, 15)
+#endif
+#ifdef HAS_I422TOARGB4444ROW_AVX2
+ANY31(I422ToARGB4444Row_Any_AVX2, I422ToARGB4444Row_AVX2, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TOARGB1555ROW_AVX2
+ANY31(I422ToARGB1555Row_Any_AVX2, I422ToARGB1555Row_AVX2, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TORGB565ROW_AVX2
+ANY31(I422ToRGB565Row_Any_AVX2, I422ToRGB565Row_AVX2, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TOARGBROW_NEON
+ANY31(I444ToARGBRow_Any_NEON, I444ToARGBRow_NEON, 0, 0, 4, 7)
+ANY31(I422ToARGBRow_Any_NEON, I422ToARGBRow_NEON, 1, 0, 4, 7)
+ANY31(I411ToARGBRow_Any_NEON, I411ToARGBRow_NEON, 2, 0, 4, 7)
+ANY31(I422ToBGRARow_Any_NEON, I422ToBGRARow_NEON, 1, 0, 4, 7)
+ANY31(I422ToABGRRow_Any_NEON, I422ToABGRRow_NEON, 1, 0, 4, 7)
+ANY31(I422ToRGBARow_Any_NEON, I422ToRGBARow_NEON, 1, 0, 4, 7)
+ANY31(I422ToRGB24Row_Any_NEON, I422ToRGB24Row_NEON, 1, 0, 3, 7)
+ANY31(I422ToRAWRow_Any_NEON, I422ToRAWRow_NEON, 1, 0, 3, 7)
+ANY31(I422ToARGB4444Row_Any_NEON, I422ToARGB4444Row_NEON, 1, 0, 2, 7)
+ANY31(I422ToARGB1555Row_Any_NEON, I422ToARGB1555Row_NEON, 1, 0, 2, 7)
+ANY31(I422ToRGB565Row_Any_NEON, I422ToRGB565Row_NEON, 1, 0, 2, 7)
+#endif
+#ifdef HAS_I422TOYUY2ROW_NEON
+ANY31(I422ToYUY2Row_Any_NEON, I422ToYUY2Row_NEON, 1, 1, 4, 15)
+#endif
+#ifdef HAS_I422TOUYVYROW_NEON
+ANY31(I422ToUYVYRow_Any_NEON, I422ToUYVYRow_NEON, 1, 1, 4, 15)
+#endif
+#undef ANY31
+
+// Any 2 planes to 1.
+#define ANY21(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, SBPP2, BPP, MASK)              \
+    void NAMEANY(const uint8* y_buf, const uint8* uv_buf,                      \
+                 uint8* dst_ptr, int width) {                                  \
+      SIMD_ALIGNED(uint8 temp[64 * 3]);                                        \
+      memset(temp, 0, 64 * 2);  /* for msan */                                 \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(y_buf, uv_buf, dst_ptr, n);                                   \
+      }                                                                        \
+      memcpy(temp, y_buf + n * SBPP, r * SBPP);                                \
+      memcpy(temp + 64, uv_buf + (n >> UVSHIFT) * SBPP2,                       \
+             SS(r, UVSHIFT) * SBPP2);                                          \
+      ANY_SIMD(temp, temp + 64, temp + 128, MASK + 1);                         \
+      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
+    }
+
+// Biplanar to RGB.
+#ifdef HAS_NV12TOARGBROW_SSSE3
+ANY21(NV12ToARGBRow_Any_SSSE3, NV12ToARGBRow_SSSE3, 1, 1, 2, 4, 7)
+ANY21(NV21ToARGBRow_Any_SSSE3, NV21ToARGBRow_SSSE3, 1, 1, 2, 4, 7)
+#endif
+#ifdef HAS_NV12TOARGBROW_AVX2
+ANY21(NV12ToARGBRow_Any_AVX2, NV12ToARGBRow_AVX2, 1, 1, 2, 4, 15)
+ANY21(NV21ToARGBRow_Any_AVX2, NV21ToARGBRow_AVX2, 1, 1, 2, 4, 15)
+#endif
+#ifdef HAS_NV12TOARGBROW_NEON
+ANY21(NV12ToARGBRow_Any_NEON, NV12ToARGBRow_NEON, 1, 1, 2, 4, 7)
+ANY21(NV21ToARGBRow_Any_NEON, NV21ToARGBRow_NEON, 1, 1, 2, 4, 7)
+#endif
+#ifdef HAS_NV12TORGB565ROW_SSSE3
+ANY21(NV12ToRGB565Row_Any_SSSE3, NV12ToRGB565Row_SSSE3, 1, 1, 2, 2, 7)
+ANY21(NV21ToRGB565Row_Any_SSSE3, NV21ToRGB565Row_SSSE3, 1, 1, 2, 2, 7)
+#endif
+#ifdef HAS_NV12TORGB565ROW_AVX2
+ANY21(NV12ToRGB565Row_Any_AVX2, NV12ToRGB565Row_AVX2, 1, 1, 2, 2, 15)
+ANY21(NV21ToRGB565Row_Any_AVX2, NV21ToRGB565Row_AVX2, 1, 1, 2, 2, 15)
+#endif
+#ifdef HAS_NV12TORGB565ROW_NEON
+ANY21(NV12ToRGB565Row_Any_NEON, NV12ToRGB565Row_NEON, 1, 1, 2, 2, 7)
+ANY21(NV21ToRGB565Row_Any_NEON, NV21ToRGB565Row_NEON, 1, 1, 2, 2, 7)
+#endif
+
+// Merge functions.
+#ifdef HAS_MERGEUVROW_SSE2
+ANY21(MergeUVRow_Any_SSE2, MergeUVRow_SSE2, 0, 1, 1, 2, 15)
+#endif
+#ifdef HAS_MERGEUVROW_AVX2
+ANY21(MergeUVRow_Any_AVX2, MergeUVRow_AVX2, 0, 1, 1, 2, 31)
+#endif
+#ifdef HAS_MERGEUVROW_NEON
+ANY21(MergeUVRow_Any_NEON, MergeUVRow_NEON, 0, 1, 1, 2, 15)
+#endif
+
+// Math functions.
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+ANY21(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, 0, 4, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBADDROW_SSE2
+ANY21(ARGBAddRow_Any_SSE2, ARGBAddRow_SSE2, 0, 4, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+ANY21(ARGBSubtractRow_Any_SSE2, ARGBSubtractRow_SSE2, 0, 4, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+ANY21(ARGBMultiplyRow_Any_AVX2, ARGBMultiplyRow_AVX2, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBADDROW_AVX2
+ANY21(ARGBAddRow_Any_AVX2, ARGBAddRow_AVX2, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+ANY21(ARGBSubtractRow_Any_AVX2, ARGBSubtractRow_AVX2, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+ANY21(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBADDROW_NEON
+ANY21(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSUBTRACTROW_NEON
+ANY21(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, 0, 4, 4, 4, 7)
+#endif
+#ifdef HAS_SOBELROW_SSE2
+ANY21(SobelRow_Any_SSE2, SobelRow_SSE2, 0, 1, 1, 4, 15)
+#endif
+#ifdef HAS_SOBELROW_NEON
+ANY21(SobelRow_Any_NEON, SobelRow_NEON, 0, 1, 1, 4, 7)
+#endif
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+ANY21(SobelToPlaneRow_Any_SSE2, SobelToPlaneRow_SSE2, 0, 1, 1, 1, 15)
+#endif
+#ifdef HAS_SOBELTOPLANEROW_NEON
+ANY21(SobelToPlaneRow_Any_NEON, SobelToPlaneRow_NEON, 0, 1, 1, 1, 15)
+#endif
+#ifdef HAS_SOBELXYROW_SSE2
+ANY21(SobelXYRow_Any_SSE2, SobelXYRow_SSE2, 0, 1, 1, 4, 15)
+#endif
+#ifdef HAS_SOBELXYROW_NEON
+ANY21(SobelXYRow_Any_NEON, SobelXYRow_NEON, 0, 1, 1, 4, 7)
+#endif
+#undef ANY21
+
+// Any 1 to 1.
+#define ANY11(NAMEANY, ANY_SIMD, UVSHIFT, SBPP, BPP, MASK)                     \
+    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr, int width) {            \
+      SIMD_ALIGNED(uint8 temp[128 * 2]);                                       \
+      memset(temp, 0, 128);  /* for YUY2 and msan */                           \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(src_ptr, dst_ptr, n);                                         \
+      }                                                                        \
+      memcpy(temp, src_ptr + (n >> UVSHIFT) * SBPP, SS(r, UVSHIFT) * SBPP);    \
+      ANY_SIMD(temp, temp + 128, MASK + 1);                                    \
+      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
+    }
+
+#ifdef HAS_COPYROW_AVX
+ANY11(CopyRow_Any_AVX, CopyRow_AVX, 0, 1, 1, 63)
+#endif
+#ifdef HAS_COPYROW_SSE2
+ANY11(CopyRow_Any_SSE2, CopyRow_SSE2, 0, 1, 1, 31)
+#endif
+#ifdef HAS_COPYROW_NEON
+ANY11(CopyRow_Any_NEON, CopyRow_NEON, 0, 1, 1, 31)
+#endif
+#if defined(HAS_ARGBTORGB24ROW_SSSE3)
+ANY11(ARGBToRGB24Row_Any_SSSE3, ARGBToRGB24Row_SSSE3, 0, 4, 3, 15)
+ANY11(ARGBToRAWRow_Any_SSSE3, ARGBToRAWRow_SSSE3, 0, 4, 3, 15)
+ANY11(ARGBToRGB565Row_Any_SSE2, ARGBToRGB565Row_SSE2, 0, 4, 2, 3)
+ANY11(ARGBToARGB1555Row_Any_SSE2, ARGBToARGB1555Row_SSE2, 0, 4, 2, 3)
+ANY11(ARGBToARGB4444Row_Any_SSE2, ARGBToARGB4444Row_SSE2, 0, 4, 2, 3)
+#endif
+#if defined(HAS_ARGBTOARGB4444ROW_AVX2)
+ANY11(ARGBToRGB565Row_Any_AVX2, ARGBToRGB565Row_AVX2, 0, 4, 2, 7)
+ANY11(ARGBToARGB1555Row_Any_AVX2, ARGBToARGB1555Row_AVX2, 0, 4, 2, 7)
+ANY11(ARGBToARGB4444Row_Any_AVX2, ARGBToARGB4444Row_AVX2, 0, 4, 2, 7)
+#endif
+#if defined(HAS_J400TOARGBROW_SSE2)
+ANY11(J400ToARGBRow_Any_SSE2, J400ToARGBRow_SSE2, 0, 1, 4, 7)
+#endif
+#if defined(HAS_J400TOARGBROW_AVX2)
+ANY11(J400ToARGBRow_Any_AVX2, J400ToARGBRow_AVX2, 0, 1, 4, 15)
+#endif
+#if defined(HAS_I400TOARGBROW_SSE2)
+ANY11(I400ToARGBRow_Any_SSE2, I400ToARGBRow_SSE2, 0, 1, 4, 7)
+#endif
+#if defined(HAS_I400TOARGBROW_AVX2)
+ANY11(I400ToARGBRow_Any_AVX2, I400ToARGBRow_AVX2, 0, 1, 4, 15)
+#endif
+#if defined(HAS_YUY2TOARGBROW_SSSE3)
+ANY11(YUY2ToARGBRow_Any_SSSE3, YUY2ToARGBRow_SSSE3, 1, 4, 4, 15)
+ANY11(UYVYToARGBRow_Any_SSSE3, UYVYToARGBRow_SSSE3, 1, 4, 4, 15)
+ANY11(RGB24ToARGBRow_Any_SSSE3, RGB24ToARGBRow_SSSE3, 0, 3, 4, 15)
+ANY11(RAWToARGBRow_Any_SSSE3, RAWToARGBRow_SSSE3, 0, 3, 4, 15)
+ANY11(RGB565ToARGBRow_Any_SSE2, RGB565ToARGBRow_SSE2, 0, 2, 4, 7)
+ANY11(ARGB1555ToARGBRow_Any_SSE2, ARGB1555ToARGBRow_SSE2, 0, 2, 4, 7)
+ANY11(ARGB4444ToARGBRow_Any_SSE2, ARGB4444ToARGBRow_SSE2, 0, 2, 4, 7)
+#endif
+#if defined(HAS_RGB565TOARGBROW_AVX2)
+ANY11(RGB565ToARGBRow_Any_AVX2, RGB565ToARGBRow_AVX2, 0, 2, 4, 15)
+#endif
+#if defined(HAS_ARGB1555TOARGBROW_AVX2)
+ANY11(ARGB1555ToARGBRow_Any_AVX2, ARGB1555ToARGBRow_AVX2, 0, 2, 4, 15)
+#endif
+#if defined(HAS_ARGB4444TOARGBROW_AVX2)
+ANY11(ARGB4444ToARGBRow_Any_AVX2, ARGB4444ToARGBRow_AVX2, 0, 2, 4, 15)
+#endif
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+ANY11(YUY2ToARGBRow_Any_AVX2, YUY2ToARGBRow_AVX2, 1, 4, 4, 31)
+ANY11(UYVYToARGBRow_Any_AVX2, UYVYToARGBRow_AVX2, 1, 4, 4, 31)
+#endif
+#if defined(HAS_ARGBTORGB24ROW_NEON)
+ANY11(ARGBToRGB24Row_Any_NEON, ARGBToRGB24Row_NEON, 0, 4, 3, 7)
+ANY11(ARGBToRAWRow_Any_NEON, ARGBToRAWRow_NEON, 0, 4, 3, 7)
+ANY11(ARGBToRGB565Row_Any_NEON, ARGBToRGB565Row_NEON, 0, 4, 2, 7)
+ANY11(ARGBToARGB1555Row_Any_NEON, ARGBToARGB1555Row_NEON, 0, 4, 2, 7)
+ANY11(ARGBToARGB4444Row_Any_NEON, ARGBToARGB4444Row_NEON, 0, 4, 2, 7)
+ANY11(J400ToARGBRow_Any_NEON, J400ToARGBRow_NEON, 0, 1, 4, 7)
+ANY11(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, 0, 1, 4, 7)
+ANY11(YUY2ToARGBRow_Any_NEON, YUY2ToARGBRow_NEON, 1, 4, 4, 7)
+ANY11(UYVYToARGBRow_Any_NEON, UYVYToARGBRow_NEON, 1, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBTOYROW_AVX2
+ANY11(ARGBToYRow_Any_AVX2, ARGBToYRow_AVX2, 0, 4, 1, 31)
+#endif
+#ifdef HAS_ARGBTOYJROW_AVX2
+ANY11(ARGBToYJRow_Any_AVX2, ARGBToYJRow_AVX2, 0, 4, 1, 31)
+#endif
+#ifdef HAS_UYVYTOYROW_AVX2
+ANY11(UYVYToYRow_Any_AVX2, UYVYToYRow_AVX2, 0, 2, 1, 31)
+#endif
+#ifdef HAS_YUY2TOYROW_AVX2
+ANY11(YUY2ToYRow_Any_AVX2, YUY2ToYRow_AVX2, 1, 4, 1, 31)
+#endif
+#ifdef HAS_ARGBTOYROW_SSSE3
+ANY11(ARGBToYRow_Any_SSSE3, ARGBToYRow_SSSE3, 0, 4, 1, 15)
+#endif
+#ifdef HAS_BGRATOYROW_SSSE3
+ANY11(BGRAToYRow_Any_SSSE3, BGRAToYRow_SSSE3, 0, 4, 1, 15)
+ANY11(ABGRToYRow_Any_SSSE3, ABGRToYRow_SSSE3, 0, 4, 1, 15)
+ANY11(RGBAToYRow_Any_SSSE3, RGBAToYRow_SSSE3, 0, 4, 1, 15)
+ANY11(YUY2ToYRow_Any_SSE2, YUY2ToYRow_SSE2, 1, 4, 1, 15)
+ANY11(UYVYToYRow_Any_SSE2, UYVYToYRow_SSE2, 1, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBTOYJROW_SSSE3
+ANY11(ARGBToYJRow_Any_SSSE3, ARGBToYJRow_SSSE3, 0, 4, 1, 15)
+#endif
+#ifdef HAS_ARGBTOYROW_NEON
+ANY11(ARGBToYRow_Any_NEON, ARGBToYRow_NEON, 0, 4, 1, 7)
+#endif
+#ifdef HAS_ARGBTOYJROW_NEON
+ANY11(ARGBToYJRow_Any_NEON, ARGBToYJRow_NEON, 0, 4, 1, 7)
+#endif
+#ifdef HAS_BGRATOYROW_NEON
+ANY11(BGRAToYRow_Any_NEON, BGRAToYRow_NEON, 0, 4, 1, 7)
+#endif
+#ifdef HAS_ABGRTOYROW_NEON
+ANY11(ABGRToYRow_Any_NEON, ABGRToYRow_NEON, 0, 4, 1, 7)
+#endif
+#ifdef HAS_RGBATOYROW_NEON
+ANY11(RGBAToYRow_Any_NEON, RGBAToYRow_NEON, 0, 4, 1, 7)
+#endif
+#ifdef HAS_RGB24TOYROW_NEON
+ANY11(RGB24ToYRow_Any_NEON, RGB24ToYRow_NEON, 0, 3, 1, 7)
+#endif
+#ifdef HAS_RAWTOYROW_NEON
+ANY11(RAWToYRow_Any_NEON, RAWToYRow_NEON, 0, 3, 1, 7)
+#endif
+#ifdef HAS_RGB565TOYROW_NEON
+ANY11(RGB565ToYRow_Any_NEON, RGB565ToYRow_NEON, 0, 2, 1, 7)
+#endif
+#ifdef HAS_ARGB1555TOYROW_NEON
+ANY11(ARGB1555ToYRow_Any_NEON, ARGB1555ToYRow_NEON, 0, 2, 1, 7)
+#endif
+#ifdef HAS_ARGB4444TOYROW_NEON
+ANY11(ARGB4444ToYRow_Any_NEON, ARGB4444ToYRow_NEON, 0, 2, 1, 7)
+#endif
+#ifdef HAS_YUY2TOYROW_NEON
+ANY11(YUY2ToYRow_Any_NEON, YUY2ToYRow_NEON, 1, 4, 1, 15)
+#endif
+#ifdef HAS_UYVYTOYROW_NEON
+ANY11(UYVYToYRow_Any_NEON, UYVYToYRow_NEON, 0, 2, 1, 15)
+#endif
+#ifdef HAS_RGB24TOARGBROW_NEON
+ANY11(RGB24ToARGBRow_Any_NEON, RGB24ToARGBRow_NEON, 0, 3, 4, 7)
+#endif
+#ifdef HAS_RAWTOARGBROW_NEON
+ANY11(RAWToARGBRow_Any_NEON, RAWToARGBRow_NEON, 0, 3, 4, 7)
+#endif
+#ifdef HAS_RGB565TOARGBROW_NEON
+ANY11(RGB565ToARGBRow_Any_NEON, RGB565ToARGBRow_NEON, 0, 2, 4, 7)
+#endif
+#ifdef HAS_ARGB1555TOARGBROW_NEON
+ANY11(ARGB1555ToARGBRow_Any_NEON, ARGB1555ToARGBRow_NEON, 0, 2, 4, 7)
+#endif
+#ifdef HAS_ARGB4444TOARGBROW_NEON
+ANY11(ARGB4444ToARGBRow_Any_NEON, ARGB4444ToARGBRow_NEON, 0, 2, 4, 7)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+ANY11(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, 0, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+ANY11(ARGBAttenuateRow_Any_SSE2, ARGBAttenuateRow_SSE2, 0, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+ANY11(ARGBUnattenuateRow_Any_SSE2, ARGBUnattenuateRow_SSE2, 0, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+ANY11(ARGBAttenuateRow_Any_AVX2, ARGBAttenuateRow_AVX2, 0, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+ANY11(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, 0, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_NEON
+ANY11(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, 0, 4, 4, 7)
+#endif
+#undef ANY11
+
+// Any 1 to 1 with parameter.
+#define ANY11P(NAMEANY, ANY_SIMD, T, SBPP, BPP, MASK)                          \
+    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr,                         \
+                 T shuffler, int width) {                                      \
+      SIMD_ALIGNED(uint8 temp[64 * 2]);                                        \
+      memset(temp, 0, 64);  /* for msan */                                     \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(src_ptr, dst_ptr, shuffler, n);                               \
+      }                                                                        \
+      memcpy(temp, src_ptr + n * SBPP, r * SBPP);                              \
+      ANY_SIMD(temp, temp + 64, shuffler, MASK + 1);                           \
+      memcpy(dst_ptr + n * BPP, temp + 64, r * BPP);                           \
+    }
+
+#if defined(HAS_ARGBTORGB565DITHERROW_SSE2)
+ANY11P(ARGBToRGB565DitherRow_Any_SSE2, ARGBToRGB565DitherRow_SSE2,
+       const uint32, 4, 2, 3)
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_AVX2)
+ANY11P(ARGBToRGB565DitherRow_Any_AVX2, ARGBToRGB565DitherRow_AVX2,
+       const uint32, 4, 2, 7)
+#endif
+#if defined(HAS_ARGBTORGB565DITHERROW_NEON)
+ANY11P(ARGBToRGB565DitherRow_Any_NEON, ARGBToRGB565DitherRow_NEON,
+       const uint32, 4, 2, 7)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_SSE2
+ANY11P(ARGBShuffleRow_Any_SSE2, ARGBShuffleRow_SSE2, const uint8*, 4, 4, 3)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_SSSE3
+ANY11P(ARGBShuffleRow_Any_SSSE3, ARGBShuffleRow_SSSE3, const uint8*, 4, 4, 7)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+ANY11P(ARGBShuffleRow_Any_AVX2, ARGBShuffleRow_AVX2, const uint8*, 4, 4, 15)
+#endif
+#ifdef HAS_ARGBSHUFFLEROW_NEON
+ANY11P(ARGBShuffleRow_Any_NEON, ARGBShuffleRow_NEON, const uint8*, 4, 4, 3)
+#endif
+#undef ANY11P
+
+// Any 1 to 1 interpolate.  Takes 2 rows of source via stride.
+#define ANY11T(NAMEANY, ANY_SIMD, SBPP, BPP, MASK)                             \
+    void NAMEANY(uint8* dst_ptr, const uint8* src_ptr,                         \
+                 ptrdiff_t src_stride_ptr, int width,                          \
+                 int source_y_fraction) {                                      \
+      SIMD_ALIGNED(uint8 temp[64 * 3]);                                        \
+      memset(temp, 0, 64 * 2);  /* for msan */                                 \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(dst_ptr, src_ptr, src_stride_ptr, n, source_y_fraction);      \
+      }                                                                        \
+      memcpy(temp, src_ptr + n * SBPP, r * SBPP);                              \
+      memcpy(temp + 64, src_ptr + src_stride_ptr + n * SBPP, r * SBPP);        \
+      ANY_SIMD(temp + 128, temp, 64, MASK + 1, source_y_fraction);             \
+      memcpy(dst_ptr + n * BPP, temp + 128, r * BPP);                          \
+    }
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+ANY11T(InterpolateRow_Any_AVX2, InterpolateRow_AVX2, 1, 1, 31)
+#endif
+#ifdef HAS_INTERPOLATEROW_SSSE3
+ANY11T(InterpolateRow_Any_SSSE3, InterpolateRow_SSSE3, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_SSE2
+ANY11T(InterpolateRow_Any_SSE2, InterpolateRow_SSE2, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_NEON
+ANY11T(InterpolateRow_Any_NEON, InterpolateRow_NEON, 1, 1, 15)
+#endif
+#ifdef HAS_INTERPOLATEROW_MIPS_DSPR2
+ANY11T(InterpolateRow_Any_MIPS_DSPR2, InterpolateRow_MIPS_DSPR2, 1, 1, 3)
+#endif
+#undef ANY11T
+
+// Any 1 to 1 mirror.
+#define ANY11M(NAMEANY, ANY_SIMD, BPP, MASK)                                   \
+    void NAMEANY(const uint8* src_ptr, uint8* dst_ptr, int width) {            \
+      SIMD_ALIGNED(uint8 temp[64 * 2]);                                        \
+      memset(temp, 0, 64);  /* for msan */                                     \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(src_ptr + r * BPP, dst_ptr, n);                               \
+      }                                                                        \
+      memcpy(temp, src_ptr, r * BPP);                                          \
+      ANY_SIMD(temp, temp + 64, MASK + 1);                                     \
+      memcpy(dst_ptr + n * BPP, temp + 64 + (MASK + 1 - r) * BPP, r * BPP);    \
+    }
+
+#ifdef HAS_MIRRORROW_AVX2
+ANY11M(MirrorRow_Any_AVX2, MirrorRow_AVX2, 1, 31)
+#endif
+#ifdef HAS_MIRRORROW_SSSE3
+ANY11M(MirrorRow_Any_SSSE3, MirrorRow_SSSE3, 1, 15)
+#endif
+#ifdef HAS_MIRRORROW_SSE2
+ANY11M(MirrorRow_Any_SSE2, MirrorRow_SSE2, 1, 15)
+#endif
+#ifdef HAS_MIRRORROW_NEON
+ANY11M(MirrorRow_Any_NEON, MirrorRow_NEON, 1, 15)
+#endif
+#ifdef HAS_ARGBMIRRORROW_AVX2
+ANY11M(ARGBMirrorRow_Any_AVX2, ARGBMirrorRow_AVX2, 4, 7)
+#endif
+#ifdef HAS_ARGBMIRRORROW_SSE2
+ANY11M(ARGBMirrorRow_Any_SSE2, ARGBMirrorRow_SSE2, 4, 3)
+#endif
+#ifdef HAS_ARGBMIRRORROW_NEON
+ANY11M(ARGBMirrorRow_Any_NEON, ARGBMirrorRow_NEON, 4, 3)
+#endif
+#undef ANY11M
+
+// Any 1 plane. (memset)
+#define ANY1(NAMEANY, ANY_SIMD, T, BPP, MASK)                                  \
+    void NAMEANY(uint8* dst_ptr, T v32, int width) {                           \
+      SIMD_ALIGNED(uint8 temp[64]);                                            \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(dst_ptr, v32, n);                                             \
+      }                                                                        \
+      ANY_SIMD(temp, v32, MASK + 1);                                           \
+      memcpy(dst_ptr + n * BPP, temp, r * BPP);                                \
+    }
+
+#ifdef HAS_SETROW_X86
+ANY1(SetRow_Any_X86, SetRow_X86, uint8, 1, 3)
+#endif
+#ifdef HAS_SETROW_NEON
+ANY1(SetRow_Any_NEON, SetRow_NEON, uint8, 1, 15)
+#endif
+#ifdef HAS_ARGBSETROW_NEON
+ANY1(ARGBSetRow_Any_NEON, ARGBSetRow_NEON, uint32, 4, 3)
+#endif
+#undef ANY1
+
+// Any 1 to 2.  Outputs UV planes.
+#define ANY12(NAMEANY, ANY_SIMD, UVSHIFT, BPP, DUVSHIFT, MASK)                 \
+    void NAMEANY(const uint8* src_ptr, uint8* dst_u, uint8* dst_v, int width) {\
+      SIMD_ALIGNED(uint8 temp[128 * 3]);                                       \
+      memset(temp, 0, 128);  /* for msan */                                    \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(src_ptr, dst_u, dst_v, n);                                    \
+      }                                                                        \
+      memcpy(temp, src_ptr  + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP);     \
+      if ((width & 1) && BPP == 4) {  /* repeat last 4 bytes for subsampler */ \
+        memcpy(temp + SS(r, UVSHIFT) * BPP,                                    \
+               temp + SS(r, UVSHIFT) * BPP - BPP, 4);                          \
+      }                                                                        \
+      ANY_SIMD(temp, temp + 128, temp + 256, MASK + 1);                        \
+      memcpy(dst_u + (n >> DUVSHIFT), temp + 128, SS(r, DUVSHIFT));            \
+      memcpy(dst_v + (n >> DUVSHIFT), temp + 256, SS(r, DUVSHIFT));            \
+    }
+
+#ifdef HAS_SPLITUVROW_SSE2
+ANY12(SplitUVRow_Any_SSE2, SplitUVRow_SSE2, 0, 2, 0, 15)
+#endif
+#ifdef HAS_SPLITUVROW_AVX2
+ANY12(SplitUVRow_Any_AVX2, SplitUVRow_AVX2, 0, 2, 0, 31)
+#endif
+#ifdef HAS_SPLITUVROW_NEON
+ANY12(SplitUVRow_Any_NEON, SplitUVRow_NEON, 0, 2, 0, 15)
+#endif
+#ifdef HAS_SPLITUVROW_MIPS_DSPR2
+ANY12(SplitUVRow_Any_MIPS_DSPR2, SplitUVRow_MIPS_DSPR2, 0, 2, 0, 15)
+#endif
+#ifdef HAS_ARGBTOUV444ROW_SSSE3
+ANY12(ARGBToUV444Row_Any_SSSE3, ARGBToUV444Row_SSSE3, 0, 4, 0, 15)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_AVX2
+ANY12(YUY2ToUV422Row_Any_AVX2, YUY2ToUV422Row_AVX2, 1, 4, 1, 31)
+ANY12(UYVYToUV422Row_Any_AVX2, UYVYToUV422Row_AVX2, 1, 4, 1, 31)
+#endif
+#ifdef HAS_ARGBTOUV422ROW_SSSE3
+ANY12(ARGBToUV422Row_Any_SSSE3, ARGBToUV422Row_SSSE3, 0, 4, 1, 15)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_SSE2
+ANY12(YUY2ToUV422Row_Any_SSE2, YUY2ToUV422Row_SSE2, 1, 4, 1, 15)
+ANY12(UYVYToUV422Row_Any_SSE2, UYVYToUV422Row_SSE2, 1, 4, 1, 15)
+#endif
+#ifdef HAS_YUY2TOUV422ROW_NEON
+ANY12(ARGBToUV444Row_Any_NEON, ARGBToUV444Row_NEON, 0, 4, 0, 7)
+ANY12(ARGBToUV422Row_Any_NEON, ARGBToUV422Row_NEON, 0, 4, 1, 15)
+ANY12(ARGBToUV411Row_Any_NEON, ARGBToUV411Row_NEON, 0, 4, 2, 31)
+ANY12(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON, 1, 4, 1, 15)
+ANY12(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON, 1, 4, 1, 15)
+#endif
+#undef ANY12
+
+// Any 1 to 2 with source stride (2 rows of source).  Outputs UV planes.
+// 128 byte row allows for 32 avx ARGB pixels.
+#define ANY12S(NAMEANY, ANY_SIMD, UVSHIFT, BPP, MASK)                          \
+    void NAMEANY(const uint8* src_ptr, int src_stride_ptr,                     \
+                 uint8* dst_u, uint8* dst_v, int width) {                      \
+      SIMD_ALIGNED(uint8 temp[128 * 4]);                                       \
+      memset(temp, 0, 128 * 2);  /* for msan */                                \
+      int r = width & MASK;                                                    \
+      int n = width & ~MASK;                                                   \
+      if (n > 0) {                                                             \
+        ANY_SIMD(src_ptr, src_stride_ptr, dst_u, dst_v, n);                    \
+      }                                                                        \
+      memcpy(temp, src_ptr  + (n >> UVSHIFT) * BPP, SS(r, UVSHIFT) * BPP);     \
+      memcpy(temp + 128, src_ptr  + src_stride_ptr + (n >> UVSHIFT) * BPP,     \
+             SS(r, UVSHIFT) * BPP);                                            \
+      if ((width & 1) && BPP == 4) {  /* repeat last 4 bytes for subsampler */ \
+        memcpy(temp + SS(r, UVSHIFT) * BPP,                                    \
+               temp + SS(r, UVSHIFT) * BPP - BPP, 4);                          \
+        memcpy(temp + 128 + SS(r, UVSHIFT) * BPP,                              \
+               temp + 128 + SS(r, UVSHIFT) * BPP - BPP, 4);                    \
+      }                                                                        \
+      ANY_SIMD(temp, 128, temp + 256, temp + 384, MASK + 1);                   \
+      memcpy(dst_u + (n >> 1), temp + 256, SS(r, 1));                          \
+      memcpy(dst_v + (n >> 1), temp + 384, SS(r, 1));                          \
+    }
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+ANY12S(ARGBToUVRow_Any_AVX2, ARGBToUVRow_AVX2, 0, 4, 31)
+#endif
+#ifdef HAS_ARGBTOUVROW_SSSE3
+ANY12S(ARGBToUVRow_Any_SSSE3, ARGBToUVRow_SSSE3, 0, 4, 15)
+ANY12S(ARGBToUVJRow_Any_SSSE3, ARGBToUVJRow_SSSE3, 0, 4, 15)
+ANY12S(BGRAToUVRow_Any_SSSE3, BGRAToUVRow_SSSE3, 0, 4, 15)
+ANY12S(ABGRToUVRow_Any_SSSE3, ABGRToUVRow_SSSE3, 0, 4, 15)
+ANY12S(RGBAToUVRow_Any_SSSE3, RGBAToUVRow_SSSE3, 0, 4, 15)
+#endif
+#ifdef HAS_YUY2TOUVROW_AVX2
+ANY12S(YUY2ToUVRow_Any_AVX2, YUY2ToUVRow_AVX2, 1, 4, 31)
+ANY12S(UYVYToUVRow_Any_AVX2, UYVYToUVRow_AVX2, 1, 4, 31)
+#endif
+#ifdef HAS_YUY2TOUVROW_SSE2
+ANY12S(YUY2ToUVRow_Any_SSE2, YUY2ToUVRow_SSE2, 1, 4, 15)
+ANY12S(UYVYToUVRow_Any_SSE2, UYVYToUVRow_SSE2, 1, 4, 15)
+#endif
+#ifdef HAS_ARGBTOUVROW_NEON
+ANY12S(ARGBToUVRow_Any_NEON, ARGBToUVRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_ARGBTOUVJROW_NEON
+ANY12S(ARGBToUVJRow_Any_NEON, ARGBToUVJRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_BGRATOUVROW_NEON
+ANY12S(BGRAToUVRow_Any_NEON, BGRAToUVRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_ABGRTOUVROW_NEON
+ANY12S(ABGRToUVRow_Any_NEON, ABGRToUVRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_RGBATOUVROW_NEON
+ANY12S(RGBAToUVRow_Any_NEON, RGBAToUVRow_NEON, 0, 4, 15)
+#endif
+#ifdef HAS_RGB24TOUVROW_NEON
+ANY12S(RGB24ToUVRow_Any_NEON, RGB24ToUVRow_NEON, 0, 3, 15)
+#endif
+#ifdef HAS_RAWTOUVROW_NEON
+ANY12S(RAWToUVRow_Any_NEON, RAWToUVRow_NEON, 0, 3, 15)
+#endif
+#ifdef HAS_RGB565TOUVROW_NEON
+ANY12S(RGB565ToUVRow_Any_NEON, RGB565ToUVRow_NEON, 0, 2, 15)
+#endif
+#ifdef HAS_ARGB1555TOUVROW_NEON
+ANY12S(ARGB1555ToUVRow_Any_NEON, ARGB1555ToUVRow_NEON, 0, 2, 15)
+#endif
+#ifdef HAS_ARGB4444TOUVROW_NEON
+ANY12S(ARGB4444ToUVRow_Any_NEON, ARGB4444ToUVRow_NEON, 0, 2, 15)
+#endif
+#ifdef HAS_YUY2TOUVROW_NEON
+ANY12S(YUY2ToUVRow_Any_NEON, YUY2ToUVRow_NEON, 1, 4, 15)
+#endif
+#ifdef HAS_UYVYTOUVROW_NEON
+ANY12S(UYVYToUVRow_Any_NEON, UYVYToUVRow_NEON, 1, 4, 15)
+#endif
+#undef ANY12S
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/row_common.cc b/third_party/aom/third_party/libyuv/source/row_common.cc
new file mode 100644
index 0000000000..49875894fe
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/row_common.cc
@@ -0,0 +1,2576 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#include <string.h>  // For memcpy and memset.
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// llvm x86 is poor at ternary operator, so use branchless min/max.
+
+#define USE_BRANCHLESS 1
+#if USE_BRANCHLESS
+static __inline int32 clamp0(int32 v) {
+  return ((-(v) >> 31) & (v));
+}
+
+static __inline int32 clamp255(int32 v) {
+  return (((255 - (v)) >> 31) | (v)) & 255;
+}
+
+static __inline uint32 Clamp(int32 val) {
+  int v = clamp0(val);
+  return (uint32)(clamp255(v));
+}
+
+static __inline uint32 Abs(int32 v) {
+  int m = v >> 31;
+  return (v + m) ^ m;
+}
+#else  // USE_BRANCHLESS
+static __inline int32 clamp0(int32 v) {
+  return (v < 0) ? 0 : v;
+}
+
+static __inline int32 clamp255(int32 v) {
+  return (v > 255) ? 255 : v;
+}
+
+static __inline uint32 Clamp(int32 val) {
+  int v = clamp0(val);
+  return (uint32)(clamp255(v));
+}
+
+static __inline uint32 Abs(int32 v) {
+  return (v < 0) ? -v : v;
+}
+#endif  // USE_BRANCHLESS
+
+#ifdef LIBYUV_LITTLE_ENDIAN
+#define WRITEWORD(p, v) *(uint32*)(p) = v
+#else
+static inline void WRITEWORD(uint8* p, uint32 v) {
+  p[0] = (uint8)(v & 255);
+  p[1] = (uint8)((v >> 8) & 255);
+  p[2] = (uint8)((v >> 16) & 255);
+  p[3] = (uint8)((v >> 24) & 255);
+}
+#endif
+
+void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb24[0];
+    uint8 g = src_rgb24[1];
+    uint8 r = src_rgb24[2];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_rgb24 += 3;
+  }
+}
+
+void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 r = src_raw[0];
+    uint8 g = src_raw[1];
+    uint8 b = src_raw[2];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_raw += 3;
+  }
+}
+
+void RGB565ToARGBRow_C(const uint8* src_rgb565, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb565[0] & 0x1f;
+    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r = src_rgb565[1] >> 3;
+    dst_argb[0] = (b << 3) | (b >> 2);
+    dst_argb[1] = (g << 2) | (g >> 4);
+    dst_argb[2] = (r << 3) | (r >> 2);
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    src_rgb565 += 2;
+  }
+}
+
+void ARGB1555ToARGBRow_C(const uint8* src_argb1555, uint8* dst_argb,
+                         int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb1555[0] & 0x1f;
+    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 a = src_argb1555[1] >> 7;
+    dst_argb[0] = (b << 3) | (b >> 2);
+    dst_argb[1] = (g << 3) | (g >> 2);
+    dst_argb[2] = (r << 3) | (r >> 2);
+    dst_argb[3] = -a;
+    dst_argb += 4;
+    src_argb1555 += 2;
+  }
+}
+
+void ARGB4444ToARGBRow_C(const uint8* src_argb4444, uint8* dst_argb,
+                         int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb4444[0] & 0x0f;
+    uint8 g = src_argb4444[0] >> 4;
+    uint8 r = src_argb4444[1] & 0x0f;
+    uint8 a = src_argb4444[1] >> 4;
+    dst_argb[0] = (b << 4) | b;
+    dst_argb[1] = (g << 4) | g;
+    dst_argb[2] = (r << 4) | r;
+    dst_argb[3] = (a << 4) | a;
+    dst_argb += 4;
+    src_argb4444 += 2;
+  }
+}
+
+void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb[0];
+    uint8 g = src_argb[1];
+    uint8 r = src_argb[2];
+    dst_rgb[0] = b;
+    dst_rgb[1] = g;
+    dst_rgb[2] = r;
+    dst_rgb += 3;
+    src_argb += 4;
+  }
+}
+
+void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb[0];
+    uint8 g = src_argb[1];
+    uint8 r = src_argb[2];
+    dst_rgb[0] = r;
+    dst_rgb[1] = g;
+    dst_rgb[2] = b;
+    dst_rgb += 3;
+    src_argb += 4;
+  }
+}
+
+void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 2;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 b1 = src_argb[4] >> 3;
+    uint8 g1 = src_argb[5] >> 2;
+    uint8 r1 = src_argb[6] >> 3;
+    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) |
+              (b1 << 16) | (g1 << 21) | (r1 << 27));
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 2;
+    uint8 r0 = src_argb[2] >> 3;
+    *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+// dither4 is a row of 4 values from 4x4 dither matrix.
+// The 4x4 matrix contains values to increase RGB.  When converting to
+// fewer bits (565) this provides an ordered dither.
+// The order in the 4x4 matrix in first byte is upper left.
+// The 4 values are passed as an int, then referenced as an array, so
+// endian will not affect order of the original matrix.  But the dither4
+// will containing the first pixel in the lower byte for little endian
+// or the upper byte for big endian.
+void ARGBToRGB565DitherRow_C(const uint8* src_argb, uint8* dst_rgb,
+                             const uint32 dither4, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    int dither0 = ((const unsigned char*)(&dither4))[x & 3];
+    int dither1 = ((const unsigned char*)(&dither4))[(x + 1) & 3];
+    uint8 b0 = clamp255(src_argb[0] + dither0) >> 3;
+    uint8 g0 = clamp255(src_argb[1] + dither0) >> 2;
+    uint8 r0 = clamp255(src_argb[2] + dither0) >> 3;
+    uint8 b1 = clamp255(src_argb[4] + dither1) >> 3;
+    uint8 g1 = clamp255(src_argb[5] + dither1) >> 2;
+    uint8 r1 = clamp255(src_argb[6] + dither1) >> 3;
+    WRITEWORD(dst_rgb, b0 | (g0 << 5) | (r0 << 11) |
+              (b1 << 16) | (g1 << 21) | (r1 << 27));
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    int dither0 = ((const unsigned char*)(&dither4))[(width - 1) & 3];
+    uint8 b0 = clamp255(src_argb[0] + dither0) >> 3;
+    uint8 g0 = clamp255(src_argb[1] + dither0) >> 2;
+    uint8 r0 = clamp255(src_argb[2] + dither0) >> 3;
+    *(uint16*)(dst_rgb) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 3;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 a0 = src_argb[3] >> 7;
+    uint8 b1 = src_argb[4] >> 3;
+    uint8 g1 = src_argb[5] >> 3;
+    uint8 r1 = src_argb[6] >> 3;
+    uint8 a1 = src_argb[7] >> 7;
+    *(uint32*)(dst_rgb) =
+        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15) |
+        (b1 << 16) | (g1 << 21) | (r1 << 26) | (a1 << 31);
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 3;
+    uint8 g0 = src_argb[1] >> 3;
+    uint8 r0 = src_argb[2] >> 3;
+    uint8 a0 = src_argb[3] >> 7;
+    *(uint16*)(dst_rgb) =
+        b0 | (g0 << 5) | (r0 << 10) | (a0 << 15);
+  }
+}
+
+void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb[0] >> 4;
+    uint8 g0 = src_argb[1] >> 4;
+    uint8 r0 = src_argb[2] >> 4;
+    uint8 a0 = src_argb[3] >> 4;
+    uint8 b1 = src_argb[4] >> 4;
+    uint8 g1 = src_argb[5] >> 4;
+    uint8 r1 = src_argb[6] >> 4;
+    uint8 a1 = src_argb[7] >> 4;
+    *(uint32*)(dst_rgb) =
+        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12) |
+        (b1 << 16) | (g1 << 20) | (r1 << 24) | (a1 << 28);
+    dst_rgb += 4;
+    src_argb += 8;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb[0] >> 4;
+    uint8 g0 = src_argb[1] >> 4;
+    uint8 r0 = src_argb[2] >> 4;
+    uint8 a0 = src_argb[3] >> 4;
+    *(uint16*)(dst_rgb) =
+        b0 | (g0 << 4) | (r0 << 8) | (a0 << 12);
+  }
+}
+
+static __inline int RGBToY(uint8 r, uint8 g, uint8 b) {
+  return (66 * r + 129 * g +  25 * b + 0x1080) >> 8;
+}
+
+static __inline int RGBToU(uint8 r, uint8 g, uint8 b) {
+  return (112 * b - 74 * g - 38 * r + 0x8080) >> 8;
+}
+static __inline int RGBToV(uint8 r, uint8 g, uint8 b) {
+  return (112 * r - 94 * g - 18 * b + 0x8080) >> 8;
+}
+
+#define MAKEROWY(NAME, R, G, B, BPP) \
+void NAME ## ToYRow_C(const uint8* src_argb0, uint8* dst_y, int width) {       \
+  int x;                                                                       \
+  for (x = 0; x < width; ++x) {                                                \
+    dst_y[0] = RGBToY(src_argb0[R], src_argb0[G], src_argb0[B]);               \
+    src_argb0 += BPP;                                                          \
+    dst_y += 1;                                                                \
+  }                                                                            \
+}                                                                              \
+void NAME ## ToUVRow_C(const uint8* src_rgb0, int src_stride_rgb,              \
+                       uint8* dst_u, uint8* dst_v, int width) {                \
+  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
+  int x;                                                                       \
+  for (x = 0; x < width - 1; x += 2) {                                         \
+    uint8 ab = (src_rgb0[B] + src_rgb0[B + BPP] +                              \
+               src_rgb1[B] + src_rgb1[B + BPP]) >> 2;                          \
+    uint8 ag = (src_rgb0[G] + src_rgb0[G + BPP] +                              \
+               src_rgb1[G] + src_rgb1[G + BPP]) >> 2;                          \
+    uint8 ar = (src_rgb0[R] + src_rgb0[R + BPP] +                              \
+               src_rgb1[R] + src_rgb1[R + BPP]) >> 2;                          \
+    dst_u[0] = RGBToU(ar, ag, ab);                                             \
+    dst_v[0] = RGBToV(ar, ag, ab);                                             \
+    src_rgb0 += BPP * 2;                                                       \
+    src_rgb1 += BPP * 2;                                                       \
+    dst_u += 1;                                                                \
+    dst_v += 1;                                                                \
+  }                                                                            \
+  if (width & 1) {                                                             \
+    uint8 ab = (src_rgb0[B] + src_rgb1[B]) >> 1;                               \
+    uint8 ag = (src_rgb0[G] + src_rgb1[G]) >> 1;                               \
+    uint8 ar = (src_rgb0[R] + src_rgb1[R]) >> 1;                               \
+    dst_u[0] = RGBToU(ar, ag, ab);                                             \
+    dst_v[0] = RGBToV(ar, ag, ab);                                             \
+  }                                                                            \
+}
+
+MAKEROWY(ARGB, 2, 1, 0, 4)
+MAKEROWY(BGRA, 1, 2, 3, 4)
+MAKEROWY(ABGR, 0, 1, 2, 4)
+MAKEROWY(RGBA, 3, 2, 1, 4)
+MAKEROWY(RGB24, 2, 1, 0, 3)
+MAKEROWY(RAW, 0, 1, 2, 3)
+#undef MAKEROWY
+
+// JPeg uses a variation on BT.601-1 full range
+// y =  0.29900 * r + 0.58700 * g + 0.11400 * b
+// u = -0.16874 * r - 0.33126 * g + 0.50000 * b  + center
+// v =  0.50000 * r - 0.41869 * g - 0.08131 * b  + center
+// BT.601 Mpeg range uses:
+// b 0.1016 * 255 = 25.908 = 25
+// g 0.5078 * 255 = 129.489 = 129
+// r 0.2578 * 255 = 65.739 = 66
+// JPeg 8 bit Y (not used):
+// b 0.11400 * 256 = 29.184 = 29
+// g 0.58700 * 256 = 150.272 = 150
+// r 0.29900 * 256 = 76.544 = 77
+// JPeg 7 bit Y:
+// b 0.11400 * 128 = 14.592 = 15
+// g 0.58700 * 128 = 75.136 = 75
+// r 0.29900 * 128 = 38.272 = 38
+// JPeg 8 bit U:
+// b  0.50000 * 255 = 127.5 = 127
+// g -0.33126 * 255 = -84.4713 = -84
+// r -0.16874 * 255 = -43.0287 = -43
+// JPeg 8 bit V:
+// b -0.08131 * 255 = -20.73405 = -20
+// g -0.41869 * 255 = -106.76595 = -107
+// r  0.50000 * 255 = 127.5 = 127
+
+static __inline int RGBToYJ(uint8 r, uint8 g, uint8 b) {
+  return (38 * r + 75 * g +  15 * b + 64) >> 7;
+}
+
+static __inline int RGBToUJ(uint8 r, uint8 g, uint8 b) {
+  return (127 * b - 84 * g - 43 * r + 0x8080) >> 8;
+}
+static __inline int RGBToVJ(uint8 r, uint8 g, uint8 b) {
+  return (127 * r - 107 * g - 20 * b + 0x8080) >> 8;
+}
+
+#define AVGB(a, b) (((a) + (b) + 1) >> 1)
+
+#define MAKEROWYJ(NAME, R, G, B, BPP) \
+void NAME ## ToYJRow_C(const uint8* src_argb0, uint8* dst_y, int width) {      \
+  int x;                                                                       \
+  for (x = 0; x < width; ++x) {                                                \
+    dst_y[0] = RGBToYJ(src_argb0[R], src_argb0[G], src_argb0[B]);              \
+    src_argb0 += BPP;                                                          \
+    dst_y += 1;                                                                \
+  }                                                                            \
+}                                                                              \
+void NAME ## ToUVJRow_C(const uint8* src_rgb0, int src_stride_rgb,             \
+                        uint8* dst_u, uint8* dst_v, int width) {               \
+  const uint8* src_rgb1 = src_rgb0 + src_stride_rgb;                           \
+  int x;                                                                       \
+  for (x = 0; x < width - 1; x += 2) {                                         \
+    uint8 ab = AVGB(AVGB(src_rgb0[B], src_rgb1[B]),                            \
+                    AVGB(src_rgb0[B + BPP], src_rgb1[B + BPP]));               \
+    uint8 ag = AVGB(AVGB(src_rgb0[G], src_rgb1[G]),                            \
+                    AVGB(src_rgb0[G + BPP], src_rgb1[G + BPP]));               \
+    uint8 ar = AVGB(AVGB(src_rgb0[R], src_rgb1[R]),                            \
+                    AVGB(src_rgb0[R + BPP], src_rgb1[R + BPP]));               \
+    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
+    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
+    src_rgb0 += BPP * 2;                                                       \
+    src_rgb1 += BPP * 2;                                                       \
+    dst_u += 1;                                                                \
+    dst_v += 1;                                                                \
+  }                                                                            \
+  if (width & 1) {                                                             \
+    uint8 ab = AVGB(src_rgb0[B], src_rgb1[B]);                                 \
+    uint8 ag = AVGB(src_rgb0[G], src_rgb1[G]);                                 \
+    uint8 ar = AVGB(src_rgb0[R], src_rgb1[R]);                                 \
+    dst_u[0] = RGBToUJ(ar, ag, ab);                                            \
+    dst_v[0] = RGBToVJ(ar, ag, ab);                                            \
+  }                                                                            \
+}
+
+MAKEROWYJ(ARGB, 2, 1, 0, 4)
+#undef MAKEROWYJ
+
+void ARGBToUVJ422Row_C(const uint8* src_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToUJ(ar, ag, ab);
+    dst_v[0] = RGBToVJ(ar, ag, ab);
+    src_argb += 8;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToUJ(ar, ag, ab);
+    dst_v[0] = RGBToVJ(ar, ag, ab);
+  }
+}
+
+void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_rgb565[0] & 0x1f;
+    uint8 g = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r = src_rgb565[1] >> 3;
+    b = (b << 3) | (b >> 2);
+    g = (g << 2) | (g >> 4);
+    r = (r << 3) | (r >> 2);
+    dst_y[0] = RGBToY(r, g, b);
+    src_rgb565 += 2;
+    dst_y += 1;
+  }
+}
+
+void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb1555[0] & 0x1f;
+    uint8 g = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r = (src_argb1555[1] & 0x7c) >> 2;
+    b = (b << 3) | (b >> 2);
+    g = (g << 3) | (g >> 2);
+    r = (r << 3) | (r >> 2);
+    dst_y[0] = RGBToY(r, g, b);
+    src_argb1555 += 2;
+    dst_y += 1;
+  }
+}
+
+void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 b = src_argb4444[0] & 0x0f;
+    uint8 g = src_argb4444[0] >> 4;
+    uint8 r = src_argb4444[1] & 0x0f;
+    b = (b << 4) | b;
+    g = (g << 4) | g;
+    r = (r << 4) | r;
+    dst_y[0] = RGBToY(r, g, b);
+    src_argb4444 += 2;
+    dst_y += 1;
+  }
+}
+
+void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
+                     uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_rgb565 = src_rgb565 + src_stride_rgb565;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_rgb565[0] & 0x1f;
+    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r0 = src_rgb565[1] >> 3;
+    uint8 b1 = src_rgb565[2] & 0x1f;
+    uint8 g1 = (src_rgb565[2] >> 5) | ((src_rgb565[3] & 0x07) << 3);
+    uint8 r1 = src_rgb565[3] >> 3;
+    uint8 b2 = next_rgb565[0] & 0x1f;
+    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
+    uint8 r2 = next_rgb565[1] >> 3;
+    uint8 b3 = next_rgb565[2] & 0x1f;
+    uint8 g3 = (next_rgb565[2] >> 5) | ((next_rgb565[3] & 0x07) << 3);
+    uint8 r3 = next_rgb565[3] >> 3;
+    uint8 b = (b0 + b1 + b2 + b3);  // 565 * 4 = 787.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 1) | (b >> 6);  // 787 -> 888.
+    r = (r << 1) | (r >> 6);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_rgb565 += 4;
+    next_rgb565 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_rgb565[0] & 0x1f;
+    uint8 g0 = (src_rgb565[0] >> 5) | ((src_rgb565[1] & 0x07) << 3);
+    uint8 r0 = src_rgb565[1] >> 3;
+    uint8 b2 = next_rgb565[0] & 0x1f;
+    uint8 g2 = (next_rgb565[0] >> 5) | ((next_rgb565[1] & 0x07) << 3);
+    uint8 r2 = next_rgb565[1] >> 3;
+    uint8 b = (b0 + b2);  // 565 * 2 = 676.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 2) | (b >> 4);  // 676 -> 888
+    g = (g << 1) | (g >> 6);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_argb1555 = src_argb1555 + src_stride_argb1555;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb1555[0] & 0x1f;
+    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 b1 = src_argb1555[2] & 0x1f;
+    uint8 g1 = (src_argb1555[2] >> 5) | ((src_argb1555[3] & 0x03) << 3);
+    uint8 r1 = (src_argb1555[3] & 0x7c) >> 2;
+    uint8 b2 = next_argb1555[0] & 0x1f;
+    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
+    uint8 r2 = (next_argb1555[1] & 0x7c) >> 2;
+    uint8 b3 = next_argb1555[2] & 0x1f;
+    uint8 g3 = (next_argb1555[2] >> 5) | ((next_argb1555[3] & 0x03) << 3);
+    uint8 r3 = (next_argb1555[3] & 0x7c) >> 2;
+    uint8 b = (b0 + b1 + b2 + b3);  // 555 * 4 = 777.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 1) | (b >> 6);  // 777 -> 888.
+    g = (g << 1) | (g >> 6);
+    r = (r << 1) | (r >> 6);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_argb1555 += 4;
+    next_argb1555 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb1555[0] & 0x1f;
+    uint8 g0 = (src_argb1555[0] >> 5) | ((src_argb1555[1] & 0x03) << 3);
+    uint8 r0 = (src_argb1555[1] & 0x7c) >> 2;
+    uint8 b2 = next_argb1555[0] & 0x1f;
+    uint8 g2 = (next_argb1555[0] >> 5) | ((next_argb1555[1] & 0x03) << 3);
+    uint8 r2 = next_argb1555[1] >> 3;
+    uint8 b = (b0 + b2);  // 555 * 2 = 666.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 2) | (b >> 4);  // 666 -> 888.
+    g = (g << 2) | (g >> 4);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  const uint8* next_argb4444 = src_argb4444 + src_stride_argb4444;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 b0 = src_argb4444[0] & 0x0f;
+    uint8 g0 = src_argb4444[0] >> 4;
+    uint8 r0 = src_argb4444[1] & 0x0f;
+    uint8 b1 = src_argb4444[2] & 0x0f;
+    uint8 g1 = src_argb4444[2] >> 4;
+    uint8 r1 = src_argb4444[3] & 0x0f;
+    uint8 b2 = next_argb4444[0] & 0x0f;
+    uint8 g2 = next_argb4444[0] >> 4;
+    uint8 r2 = next_argb4444[1] & 0x0f;
+    uint8 b3 = next_argb4444[2] & 0x0f;
+    uint8 g3 = next_argb4444[2] >> 4;
+    uint8 r3 = next_argb4444[3] & 0x0f;
+    uint8 b = (b0 + b1 + b2 + b3);  // 444 * 4 = 666.
+    uint8 g = (g0 + g1 + g2 + g3);
+    uint8 r = (r0 + r1 + r2 + r3);
+    b = (b << 2) | (b >> 4);  // 666 -> 888.
+    g = (g << 2) | (g >> 4);
+    r = (r << 2) | (r >> 4);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+    src_argb4444 += 4;
+    next_argb4444 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 b0 = src_argb4444[0] & 0x0f;
+    uint8 g0 = src_argb4444[0] >> 4;
+    uint8 r0 = src_argb4444[1] & 0x0f;
+    uint8 b2 = next_argb4444[0] & 0x0f;
+    uint8 g2 = next_argb4444[0] >> 4;
+    uint8 r2 = next_argb4444[1] & 0x0f;
+    uint8 b = (b0 + b2);  // 444 * 2 = 555.
+    uint8 g = (g0 + g2);
+    uint8 r = (r0 + r2);
+    b = (b << 3) | (b >> 2);  // 555 -> 888.
+    g = (g << 3) | (g >> 2);
+    r = (r << 3) | (r >> 2);
+    dst_u[0] = RGBToU(r, g, b);
+    dst_v[0] = RGBToV(r, g, b);
+  }
+}
+
+void ARGBToUV444Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+void ARGBToUV422Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 8;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if (width & 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  }
+}
+
+void ARGBToUV411Row_C(const uint8* src_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 3; x += 4) {
+    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8] + src_argb[12]) >> 2;
+    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9] + src_argb[13]) >> 2;
+    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10] + src_argb[14]) >> 2;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+    src_argb += 16;
+    dst_u += 1;
+    dst_v += 1;
+  }
+  if ((width & 3) == 3) {
+    uint8 ab = (src_argb[0] + src_argb[4] + src_argb[8]) / 3;
+    uint8 ag = (src_argb[1] + src_argb[5] + src_argb[9]) / 3;
+    uint8 ar = (src_argb[2] + src_argb[6] + src_argb[10]) / 3;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  } else if ((width & 3) == 2) {
+    uint8 ab = (src_argb[0] + src_argb[4]) >> 1;
+    uint8 ag = (src_argb[1] + src_argb[5]) >> 1;
+    uint8 ar = (src_argb[2] + src_argb[6]) >> 1;
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  } else if ((width & 3) == 1) {
+    uint8 ab = src_argb[0];
+    uint8 ag = src_argb[1];
+    uint8 ar = src_argb[2];
+    dst_u[0] = RGBToU(ar, ag, ab);
+    dst_v[0] = RGBToV(ar, ag, ab);
+  }
+}
+
+void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 y = RGBToYJ(src_argb[2], src_argb[1], src_argb[0]);
+    dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
+    dst_argb[3] = src_argb[3];
+    dst_argb += 4;
+    src_argb += 4;
+  }
+}
+
+// Convert a row of image to Sepia tone.
+void ARGBSepiaRow_C(uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    int sb = (b * 17 + g * 68 + r * 35) >> 7;
+    int sg = (b * 22 + g * 88 + r * 45) >> 7;
+    int sr = (b * 24 + g * 98 + r * 50) >> 7;
+    // b does not over flow. a is preserved from original.
+    dst_argb[0] = sb;
+    dst_argb[1] = clamp255(sg);
+    dst_argb[2] = clamp255(sr);
+    dst_argb += 4;
+  }
+}
+
+// Apply color matrix to a row of image. Matrix is signed.
+// TODO(fbarchard): Consider adding rounding (+32).
+void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
+                          const int8* matrix_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = src_argb[0];
+    int g = src_argb[1];
+    int r = src_argb[2];
+    int a = src_argb[3];
+    int sb = (b * matrix_argb[0] + g * matrix_argb[1] +
+              r * matrix_argb[2] + a * matrix_argb[3]) >> 6;
+    int sg = (b * matrix_argb[4] + g * matrix_argb[5] +
+              r * matrix_argb[6] + a * matrix_argb[7]) >> 6;
+    int sr = (b * matrix_argb[8] + g * matrix_argb[9] +
+              r * matrix_argb[10] + a * matrix_argb[11]) >> 6;
+    int sa = (b * matrix_argb[12] + g * matrix_argb[13] +
+              r * matrix_argb[14] + a * matrix_argb[15]) >> 6;
+    dst_argb[0] = Clamp(sb);
+    dst_argb[1] = Clamp(sg);
+    dst_argb[2] = Clamp(sr);
+    dst_argb[3] = Clamp(sa);
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+// Apply color table to a row of image.
+void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    int a = dst_argb[3];
+    dst_argb[0] = table_argb[b * 4 + 0];
+    dst_argb[1] = table_argb[g * 4 + 1];
+    dst_argb[2] = table_argb[r * 4 + 2];
+    dst_argb[3] = table_argb[a * 4 + 3];
+    dst_argb += 4;
+  }
+}
+
+// Apply color table to a row of image.
+void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    dst_argb[0] = table_argb[b * 4 + 0];
+    dst_argb[1] = table_argb[g * 4 + 1];
+    dst_argb[2] = table_argb[r * 4 + 2];
+    dst_argb += 4;
+  }
+}
+
+void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
+                       int interval_offset, int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    int b = dst_argb[0];
+    int g = dst_argb[1];
+    int r = dst_argb[2];
+    dst_argb[0] = (b * scale >> 16) * interval_size + interval_offset;
+    dst_argb[1] = (g * scale >> 16) * interval_size + interval_offset;
+    dst_argb[2] = (r * scale >> 16) * interval_size + interval_offset;
+    dst_argb += 4;
+  }
+}
+
+#define REPEAT8(v) (v) | ((v) << 8)
+#define SHADE(f, v) v * f >> 24
+
+void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                    uint32 value) {
+  const uint32 b_scale = REPEAT8(value & 0xff);
+  const uint32 g_scale = REPEAT8((value >> 8) & 0xff);
+  const uint32 r_scale = REPEAT8((value >> 16) & 0xff);
+  const uint32 a_scale = REPEAT8(value >> 24);
+
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32 b = REPEAT8(src_argb[0]);
+    const uint32 g = REPEAT8(src_argb[1]);
+    const uint32 r = REPEAT8(src_argb[2]);
+    const uint32 a = REPEAT8(src_argb[3]);
+    dst_argb[0] = SHADE(b, b_scale);
+    dst_argb[1] = SHADE(g, g_scale);
+    dst_argb[2] = SHADE(r, r_scale);
+    dst_argb[3] = SHADE(a, a_scale);
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+#undef REPEAT8
+#undef SHADE
+
+#define REPEAT8(v) (v) | ((v) << 8)
+#define SHADE(f, v) v * f >> 16
+
+void ARGBMultiplyRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const uint32 b = REPEAT8(src_argb0[0]);
+    const uint32 g = REPEAT8(src_argb0[1]);
+    const uint32 r = REPEAT8(src_argb0[2]);
+    const uint32 a = REPEAT8(src_argb0[3]);
+    const uint32 b_scale = src_argb1[0];
+    const uint32 g_scale = src_argb1[1];
+    const uint32 r_scale = src_argb1[2];
+    const uint32 a_scale = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_scale);
+    dst_argb[1] = SHADE(g, g_scale);
+    dst_argb[2] = SHADE(r, r_scale);
+    dst_argb[3] = SHADE(a, a_scale);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef REPEAT8
+#undef SHADE
+
+#define SHADE(f, v) clamp255(v + f)
+
+void ARGBAddRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                  uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const int b = src_argb0[0];
+    const int g = src_argb0[1];
+    const int r = src_argb0[2];
+    const int a = src_argb0[3];
+    const int b_add = src_argb1[0];
+    const int g_add = src_argb1[1];
+    const int r_add = src_argb1[2];
+    const int a_add = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_add);
+    dst_argb[1] = SHADE(g, g_add);
+    dst_argb[2] = SHADE(r, r_add);
+    dst_argb[3] = SHADE(a, a_add);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
+#define SHADE(f, v) clamp0(f - v)
+
+void ARGBSubtractRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    const int b = src_argb0[0];
+    const int g = src_argb0[1];
+    const int r = src_argb0[2];
+    const int a = src_argb0[3];
+    const int b_sub = src_argb1[0];
+    const int g_sub = src_argb1[1];
+    const int r_sub = src_argb1[2];
+    const int a_sub = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_sub);
+    dst_argb[1] = SHADE(g, g_sub);
+    dst_argb[2] = SHADE(r, r_sub);
+    dst_argb[3] = SHADE(a, a_sub);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
+// Sobel functions which mimics SSSE3.
+void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
+                 uint8* dst_sobelx, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int a = src_y0[i];
+    int b = src_y1[i];
+    int c = src_y2[i];
+    int a_sub = src_y0[i + 2];
+    int b_sub = src_y1[i + 2];
+    int c_sub = src_y2[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = Abs(a_diff + b_diff * 2 + c_diff);
+    dst_sobelx[i] = (uint8)(clamp255(sobel));
+  }
+}
+
+void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
+                 uint8* dst_sobely, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int a = src_y0[i + 0];
+    int b = src_y0[i + 1];
+    int c = src_y0[i + 2];
+    int a_sub = src_y1[i + 0];
+    int b_sub = src_y1[i + 1];
+    int c_sub = src_y1[i + 2];
+    int a_diff = a - a_sub;
+    int b_diff = b - b_sub;
+    int c_diff = c - c_sub;
+    int sobel = Abs(a_diff + b_diff * 2 + c_diff);
+    dst_sobely[i] = (uint8)(clamp255(sobel));
+  }
+}
+
+void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = clamp255(r + b);
+    dst_argb[0] = (uint8)(s);
+    dst_argb[1] = (uint8)(s);
+    dst_argb[2] = (uint8)(s);
+    dst_argb[3] = (uint8)(255u);
+    dst_argb += 4;
+  }
+}
+
+void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                       uint8* dst_y, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int s = clamp255(r + b);
+    dst_y[i] = (uint8)(s);
+  }
+}
+
+void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
+                  uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    int r = src_sobelx[i];
+    int b = src_sobely[i];
+    int g = clamp255(r + b);
+    dst_argb[0] = (uint8)(b);
+    dst_argb[1] = (uint8)(g);
+    dst_argb[2] = (uint8)(r);
+    dst_argb[3] = (uint8)(255u);
+    dst_argb += 4;
+  }
+}
+
+void J400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) {
+  // Copy a Y to RGB.
+  int x;
+  for (x = 0; x < width; ++x) {
+    uint8 y = src_y[0];
+    dst_argb[2] = dst_argb[1] = dst_argb[0] = y;
+    dst_argb[3] = 255u;
+    dst_argb += 4;
+    ++src_y;
+  }
+}
+
+// BT.601 YUV to RGB reference
+//  R = (Y - 16) * 1.164              - V * -1.596
+//  G = (Y - 16) * 1.164 - U *  0.391 - V *  0.813
+//  B = (Y - 16) * 1.164 - U * -2.018
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* max(-128, round(-2.018 * 64)) */
+#define UG 25 /* round(0.391 * 64) */
+#define VG 52 /* round(0.813 * 64) */
+#define VR -102 /* round(-1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128 + YGB)
+#define BG (UG * 128 + VG * 128 + YGB)
+#define BR (VR * 128 + YGB)
+
+// C reference code that mimics the YUV assembly.
+static __inline void YuvPixel(uint8 y, uint8 u, uint8 v,
+                              uint8* b, uint8* g, uint8* r) {
+  uint32 y1 = (uint32)(y * 0x0101 * YG) >> 16;
+  *b = Clamp((int32)(-(u * UB) + y1 + BB) >> 6);
+  *g = Clamp((int32)(-(v * VG + u * UG) + y1 + BG) >> 6);
+  *r = Clamp((int32)(-(v * VR)+ y1 + BR) >> 6);
+}
+
+// C reference code that mimics the YUV assembly.
+static __inline void YPixel(uint8 y, uint8* b, uint8* g, uint8* r) {
+  uint32 y1 = (uint32)(y * 0x0101 * YG) >> 16;
+  *b = Clamp((int32)(y1 + YGB) >> 6);
+  *g = Clamp((int32)(y1 + YGB) >> 6);
+  *r = Clamp((int32)(y1 + YGB) >> 6);
+}
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+// JPEG YUV to RGB reference
+// *  R = Y                - V * -1.40200
+// *  G = Y - U *  0.34414 - V *  0.71414
+// *  B = Y - U * -1.77200
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YGJ 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+#define YGBJ 32  /* 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UBJ -113 /* round(-1.77200 * 64) */
+#define UGJ 22 /* round(0.34414 * 64) */
+#define VGJ 46 /* round(0.71414  * 64) */
+#define VRJ -90 /* round(-1.40200 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BBJ (UBJ * 128 + YGBJ)
+#define BGJ (UGJ * 128 + VGJ * 128 + YGBJ)
+#define BRJ (VRJ * 128 + YGBJ)
+
+// C reference code that mimics the YUV assembly.
+static __inline void YuvJPixel(uint8 y, uint8 u, uint8 v,
+                               uint8* b, uint8* g, uint8* r) {
+  uint32 y1 = (uint32)(y * 0x0101 * YGJ) >> 16;
+  *b = Clamp((int32)(-(u * UBJ) + y1 + BBJ) >> 6);
+  *g = Clamp((int32)(-(v * VGJ + u * UGJ) + y1 + BGJ) >> 6);
+  *r = Clamp((int32)(-(v * VRJ) + y1 + BRJ) >> 6);
+}
+
+#undef YGJ
+#undef YGBJ
+#undef UBJ
+#undef UGJ
+#undef VGJ
+#undef VRJ
+#undef BBJ
+#undef BGJ
+#undef BRJ
+
+#if !defined(LIBYUV_DISABLE_NEON) && \
+    (defined(__ARM_NEON__) || defined(__aarch64__) || defined(LIBYUV_NEON))
+// C mimic assembly.
+// TODO(fbarchard): Remove subsampling from Neon.
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint8 u = (src_u[0] + src_u[1] + 1) >> 1;
+    uint8 v = (src_v[0] + src_v[1] + 1) >> 1;
+    YuvPixel(src_y[0], u, v, rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], u, v, rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 2;
+    src_v += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+  }
+}
+#else
+void I444ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width; ++x) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    src_y += 1;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 4;  // Advance 1 pixel.
+  }
+}
+#endif
+
+// Also used for 420
+void I422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void J422ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvJPixel(src_y[0], src_u[0], src_v[0],
+              rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvJPixel(src_y[1], src_u[0], src_v[0],
+              rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvJPixel(src_y[0], src_u[0], src_v[0],
+              rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToRGB24Row_C(const uint8* src_y,
+                      const uint8* src_u,
+                      const uint8* src_v,
+                      uint8* rgb_buf,
+                      int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 4, rgb_buf + 5);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+  }
+}
+
+void I422ToRAWRow_C(const uint8* src_y,
+                    const uint8* src_u,
+                    const uint8* src_v,
+                    uint8* rgb_buf,
+                    int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 5, rgb_buf + 4, rgb_buf + 3);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 6;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+  }
+}
+
+void I422ToARGB4444Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb4444,
+                         int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 4;
+    g0 = g0 >> 4;
+    r0 = r0 >> 4;
+    b1 = b1 >> 4;
+    g1 = g1 >> 4;
+    r1 = r1 >> 4;
+    *(uint32*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
+        (b1 << 16) | (g1 << 20) | (r1 << 24) | 0xf000f000;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_argb4444 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 4;
+    g0 = g0 >> 4;
+    r0 = r0 >> 4;
+    *(uint16*)(dst_argb4444) = b0 | (g0 << 4) | (r0 << 8) |
+        0xf000;
+  }
+}
+
+void I422ToARGB1555Row_C(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_argb1555,
+                         int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 3;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 3;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
+        (b1 << 16) | (g1 << 21) | (r1 << 26) | 0x80008000;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_argb1555 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 3;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_argb1555) = b0 | (g0 << 5) | (r0 << 10) |
+        0x8000;
+  }
+}
+
+void I422ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_u[0], src_v[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void I411ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 3; x += 4) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    YuvPixel(src_y[2], src_u[0], src_v[0],
+             rgb_buf + 8, rgb_buf + 9, rgb_buf + 10);
+    rgb_buf[11] = 255;
+    YuvPixel(src_y[3], src_u[0], src_v[0],
+             rgb_buf + 12, rgb_buf + 13, rgb_buf + 14);
+    rgb_buf[15] = 255;
+    src_y += 4;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 16;  // Advance 4 pixels.
+  }
+  if (width & 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV12ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_uv,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_uv[0], src_uv[1],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_uv[0], src_uv[1],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_uv += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_uv[0], src_uv[1],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV21ToARGBRow_C(const uint8* src_y,
+                     const uint8* src_vu,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+
+    YuvPixel(src_y[1], src_vu[1], src_vu[0],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+
+    src_y += 2;
+    src_vu += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_vu[1], src_vu[0],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void NV12ToRGB565Row_C(const uint8* src_y,
+                       const uint8* src_uv,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_uv[0], src_uv[1], &b0, &g0, &r0);
+    YuvPixel(src_y[1], src_uv[0], src_uv[1], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    src_uv += 2;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_uv[0], src_uv[1], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void NV21ToRGB565Row_C(const uint8* src_y,
+                       const uint8* vsrc_u,
+                       uint8* dst_rgb565,
+                       int width) {
+  uint8 b0;
+  uint8 g0;
+  uint8 r0;
+  uint8 b1;
+  uint8 g1;
+  uint8 r1;
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0);
+    YuvPixel(src_y[1], vsrc_u[1], vsrc_u[0], &b1, &g1, &r1);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    b1 = b1 >> 3;
+    g1 = g1 >> 2;
+    r1 = r1 >> 3;
+    *(uint32*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11) |
+        (b1 << 16) | (g1 << 21) | (r1 << 27);
+    src_y += 2;
+    vsrc_u += 2;
+    dst_rgb565 += 4;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], vsrc_u[1], vsrc_u[0], &b0, &g0, &r0);
+    b0 = b0 >> 3;
+    g0 = g0 >> 2;
+    r0 = r0 >> 3;
+    *(uint16*)(dst_rgb565) = b0 | (g0 << 5) | (r0 << 11);
+  }
+}
+
+void YUY2ToARGBRow_C(const uint8* src_yuy2,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_yuy2[2], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_yuy2 += 4;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_yuy2[0], src_yuy2[1], src_yuy2[3],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void UYVYToARGBRow_C(const uint8* src_uyvy,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YuvPixel(src_uyvy[3], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_uyvy += 4;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_uyvy[1], src_uyvy[0], src_uyvy[2],
+             rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToBGRARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 2, rgb_buf + 1);
+    rgb_buf[0] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 7, rgb_buf + 6, rgb_buf + 5);
+    rgb_buf[4] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 3, rgb_buf + 2, rgb_buf + 1);
+    rgb_buf[0] = 255;
+  }
+}
+
+void I422ToABGRRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    rgb_buf[3] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 6, rgb_buf + 5, rgb_buf + 4);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 2, rgb_buf + 1, rgb_buf + 0);
+    rgb_buf[3] = 255;
+  }
+}
+
+void I422ToRGBARow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* rgb_buf,
+                     int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3);
+    rgb_buf[0] = 255;
+    YuvPixel(src_y[1], src_u[0], src_v[0],
+             rgb_buf + 5, rgb_buf + 6, rgb_buf + 7);
+    rgb_buf[4] = 255;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YuvPixel(src_y[0], src_u[0], src_v[0],
+             rgb_buf + 1, rgb_buf + 2, rgb_buf + 3);
+    rgb_buf[0] = 255;
+  }
+}
+
+void I400ToARGBRow_C(const uint8* src_y, uint8* rgb_buf, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+    YPixel(src_y[1], rgb_buf + 4, rgb_buf + 5, rgb_buf + 6);
+    rgb_buf[7] = 255;
+    src_y += 2;
+    rgb_buf += 8;  // Advance 2 pixels.
+  }
+  if (width & 1) {
+    YPixel(src_y[0], rgb_buf + 0, rgb_buf + 1, rgb_buf + 2);
+    rgb_buf[3] = 255;
+  }
+}
+
+void MirrorRow_C(const uint8* src, uint8* dst, int width) {
+  int x;
+  src += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst[x] = src[0];
+    dst[x + 1] = src[-1];
+    src -= 2;
+  }
+  if (width & 1) {
+    dst[width - 1] = src[0];
+  }
+}
+
+void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  src_uv += (width - 1) << 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_u[x] = src_uv[0];
+    dst_u[x + 1] = src_uv[-2];
+    dst_v[x] = src_uv[1];
+    dst_v[x + 1] = src_uv[-2 + 1];
+    src_uv -= 4;
+  }
+  if (width & 1) {
+    dst_u[width - 1] = src_uv[0];
+    dst_v[width - 1] = src_uv[1];
+  }
+}
+
+void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width) {
+  int x;
+  const uint32* src32 = (const uint32*)(src);
+  uint32* dst32 = (uint32*)(dst);
+  src32 += width - 1;
+  for (x = 0; x < width - 1; x += 2) {
+    dst32[x] = src32[0];
+    dst32[x + 1] = src32[-1];
+    src32 -= 2;
+  }
+  if (width & 1) {
+    dst32[width - 1] = src32[0];
+  }
+}
+
+void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_u[x] = src_uv[0];
+    dst_u[x + 1] = src_uv[2];
+    dst_v[x] = src_uv[1];
+    dst_v[x + 1] = src_uv[3];
+    src_uv += 4;
+  }
+  if (width & 1) {
+    dst_u[width - 1] = src_uv[0];
+    dst_v[width - 1] = src_uv[1];
+  }
+}
+
+void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                  int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_uv[0] = src_u[x];
+    dst_uv[1] = src_v[x];
+    dst_uv[2] = src_u[x + 1];
+    dst_uv[3] = src_v[x + 1];
+    dst_uv += 4;
+  }
+  if (width & 1) {
+    dst_uv[0] = src_u[width - 1];
+    dst_uv[1] = src_v[width - 1];
+  }
+}
+
+void CopyRow_C(const uint8* src, uint8* dst, int count) {
+  memcpy(dst, src, count);
+}
+
+void CopyRow_16_C(const uint16* src, uint16* dst, int count) {
+  memcpy(dst, src, count * 2);
+}
+
+void SetRow_C(uint8* dst, uint8 v8, int width) {
+  memset(dst, v8, width);
+}
+
+void ARGBSetRow_C(uint8* dst_argb, uint32 v32, int width) {
+  uint32* d = (uint32*)(dst_argb);
+  int x;
+  for (x = 0; x < width; ++x) {
+    d[x] = v32;
+  }
+}
+
+// Filter 2 rows of YUY2 UV's (422) into U and V (420).
+void YUY2ToUVRow_C(const uint8* src_yuy2, int src_stride_yuy2,
+                   uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values, filtering 2 rows of YUY2.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = (src_yuy2[1] + src_yuy2[src_stride_yuy2 + 1] + 1) >> 1;
+    dst_v[0] = (src_yuy2[3] + src_yuy2[src_stride_yuy2 + 3] + 1) >> 1;
+    src_yuy2 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of YUY2 UV's (422) into U and V (422).
+void YUY2ToUV422Row_C(const uint8* src_yuy2,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = src_yuy2[1];
+    dst_v[0] = src_yuy2[3];
+    src_yuy2 += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of YUY2 Y's (422) into Y (420/422).
+void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_y[x] = src_yuy2[0];
+    dst_y[x + 1] = src_yuy2[2];
+    src_yuy2 += 4;
+  }
+  if (width & 1) {
+    dst_y[width - 1] = src_yuy2[0];
+  }
+}
+
+// Filter 2 rows of UYVY UV's (422) into U and V (420).
+void UYVYToUVRow_C(const uint8* src_uyvy, int src_stride_uyvy,
+                   uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = (src_uyvy[0] + src_uyvy[src_stride_uyvy + 0] + 1) >> 1;
+    dst_v[0] = (src_uyvy[2] + src_uyvy[src_stride_uyvy + 2] + 1) >> 1;
+    src_uyvy += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of UYVY UV's (422) into U and V (422).
+void UYVYToUV422Row_C(const uint8* src_uyvy,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  // Output a row of UV values.
+  int x;
+  for (x = 0; x < width; x += 2) {
+    dst_u[0] = src_uyvy[0];
+    dst_v[0] = src_uyvy[2];
+    src_uyvy += 4;
+    dst_u += 1;
+    dst_v += 1;
+  }
+}
+
+// Copy row of UYVY Y's (422) into Y (420/422).
+void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int width) {
+  // Output a row of Y values.
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_y[x] = src_uyvy[1];
+    dst_y[x + 1] = src_uyvy[3];
+    src_uyvy += 4;
+  }
+  if (width & 1) {
+    dst_y[width - 1] = src_uyvy[1];
+  }
+}
+
+#define BLEND(f, b, a) (((256 - a) * b) >> 8) + f
+
+// Blend src_argb0 over src_argb1 and store to dst_argb.
+// dst_argb may be src_argb0 or src_argb1.
+// This code mimics the SSSE3 version for better testability.
+void ARGBBlendRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                    uint8* dst_argb, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    uint32 fb = src_argb0[0];
+    uint32 fg = src_argb0[1];
+    uint32 fr = src_argb0[2];
+    uint32 a = src_argb0[3];
+    uint32 bb = src_argb1[0];
+    uint32 bg = src_argb1[1];
+    uint32 br = src_argb1[2];
+    dst_argb[0] = BLEND(fb, bb, a);
+    dst_argb[1] = BLEND(fg, bg, a);
+    dst_argb[2] = BLEND(fr, br, a);
+    dst_argb[3] = 255u;
+
+    fb = src_argb0[4 + 0];
+    fg = src_argb0[4 + 1];
+    fr = src_argb0[4 + 2];
+    a = src_argb0[4 + 3];
+    bb = src_argb1[4 + 0];
+    bg = src_argb1[4 + 1];
+    br = src_argb1[4 + 2];
+    dst_argb[4 + 0] = BLEND(fb, bb, a);
+    dst_argb[4 + 1] = BLEND(fg, bg, a);
+    dst_argb[4 + 2] = BLEND(fr, br, a);
+    dst_argb[4 + 3] = 255u;
+    src_argb0 += 8;
+    src_argb1 += 8;
+    dst_argb += 8;
+  }
+
+  if (width & 1) {
+    uint32 fb = src_argb0[0];
+    uint32 fg = src_argb0[1];
+    uint32 fr = src_argb0[2];
+    uint32 a = src_argb0[3];
+    uint32 bb = src_argb1[0];
+    uint32 bg = src_argb1[1];
+    uint32 br = src_argb1[2];
+    dst_argb[0] = BLEND(fb, bb, a);
+    dst_argb[1] = BLEND(fg, bg, a);
+    dst_argb[2] = BLEND(fr, br, a);
+    dst_argb[3] = 255u;
+  }
+}
+#undef BLEND
+#define ATTENUATE(f, a) (a | (a << 8)) * (f | (f << 8)) >> 24
+
+// Multiply source RGB by alpha and store to destination.
+// This code mimics the SSSE3 version for better testability.
+void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    uint32 b = src_argb[0];
+    uint32 g = src_argb[1];
+    uint32 r = src_argb[2];
+    uint32 a = src_argb[3];
+    dst_argb[0] = ATTENUATE(b, a);
+    dst_argb[1] = ATTENUATE(g, a);
+    dst_argb[2] = ATTENUATE(r, a);
+    dst_argb[3] = a;
+    b = src_argb[4];
+    g = src_argb[5];
+    r = src_argb[6];
+    a = src_argb[7];
+    dst_argb[4] = ATTENUATE(b, a);
+    dst_argb[5] = ATTENUATE(g, a);
+    dst_argb[6] = ATTENUATE(r, a);
+    dst_argb[7] = a;
+    src_argb += 8;
+    dst_argb += 8;
+  }
+
+  if (width & 1) {
+    const uint32 b = src_argb[0];
+    const uint32 g = src_argb[1];
+    const uint32 r = src_argb[2];
+    const uint32 a = src_argb[3];
+    dst_argb[0] = ATTENUATE(b, a);
+    dst_argb[1] = ATTENUATE(g, a);
+    dst_argb[2] = ATTENUATE(r, a);
+    dst_argb[3] = a;
+  }
+}
+#undef ATTENUATE
+
+// Divide source RGB by alpha and store to destination.
+// b = (b * 255 + (a / 2)) / a;
+// g = (g * 255 + (a / 2)) / a;
+// r = (r * 255 + (a / 2)) / a;
+// Reciprocal method is off by 1 on some values. ie 125
+// 8.8 fixed point inverse table with 1.0 in upper short and 1 / a in lower.
+#define T(a) 0x01000000 + (0x10000 / a)
+const uint32 fixed_invtbl8[256] = {
+  0x01000000, 0x0100ffff, T(0x02), T(0x03), T(0x04), T(0x05), T(0x06), T(0x07),
+  T(0x08), T(0x09), T(0x0a), T(0x0b), T(0x0c), T(0x0d), T(0x0e), T(0x0f),
+  T(0x10), T(0x11), T(0x12), T(0x13), T(0x14), T(0x15), T(0x16), T(0x17),
+  T(0x18), T(0x19), T(0x1a), T(0x1b), T(0x1c), T(0x1d), T(0x1e), T(0x1f),
+  T(0x20), T(0x21), T(0x22), T(0x23), T(0x24), T(0x25), T(0x26), T(0x27),
+  T(0x28), T(0x29), T(0x2a), T(0x2b), T(0x2c), T(0x2d), T(0x2e), T(0x2f),
+  T(0x30), T(0x31), T(0x32), T(0x33), T(0x34), T(0x35), T(0x36), T(0x37),
+  T(0x38), T(0x39), T(0x3a), T(0x3b), T(0x3c), T(0x3d), T(0x3e), T(0x3f),
+  T(0x40), T(0x41), T(0x42), T(0x43), T(0x44), T(0x45), T(0x46), T(0x47),
+  T(0x48), T(0x49), T(0x4a), T(0x4b), T(0x4c), T(0x4d), T(0x4e), T(0x4f),
+  T(0x50), T(0x51), T(0x52), T(0x53), T(0x54), T(0x55), T(0x56), T(0x57),
+  T(0x58), T(0x59), T(0x5a), T(0x5b), T(0x5c), T(0x5d), T(0x5e), T(0x5f),
+  T(0x60), T(0x61), T(0x62), T(0x63), T(0x64), T(0x65), T(0x66), T(0x67),
+  T(0x68), T(0x69), T(0x6a), T(0x6b), T(0x6c), T(0x6d), T(0x6e), T(0x6f),
+  T(0x70), T(0x71), T(0x72), T(0x73), T(0x74), T(0x75), T(0x76), T(0x77),
+  T(0x78), T(0x79), T(0x7a), T(0x7b), T(0x7c), T(0x7d), T(0x7e), T(0x7f),
+  T(0x80), T(0x81), T(0x82), T(0x83), T(0x84), T(0x85), T(0x86), T(0x87),
+  T(0x88), T(0x89), T(0x8a), T(0x8b), T(0x8c), T(0x8d), T(0x8e), T(0x8f),
+  T(0x90), T(0x91), T(0x92), T(0x93), T(0x94), T(0x95), T(0x96), T(0x97),
+  T(0x98), T(0x99), T(0x9a), T(0x9b), T(0x9c), T(0x9d), T(0x9e), T(0x9f),
+  T(0xa0), T(0xa1), T(0xa2), T(0xa3), T(0xa4), T(0xa5), T(0xa6), T(0xa7),
+  T(0xa8), T(0xa9), T(0xaa), T(0xab), T(0xac), T(0xad), T(0xae), T(0xaf),
+  T(0xb0), T(0xb1), T(0xb2), T(0xb3), T(0xb4), T(0xb5), T(0xb6), T(0xb7),
+  T(0xb8), T(0xb9), T(0xba), T(0xbb), T(0xbc), T(0xbd), T(0xbe), T(0xbf),
+  T(0xc0), T(0xc1), T(0xc2), T(0xc3), T(0xc4), T(0xc5), T(0xc6), T(0xc7),
+  T(0xc8), T(0xc9), T(0xca), T(0xcb), T(0xcc), T(0xcd), T(0xce), T(0xcf),
+  T(0xd0), T(0xd1), T(0xd2), T(0xd3), T(0xd4), T(0xd5), T(0xd6), T(0xd7),
+  T(0xd8), T(0xd9), T(0xda), T(0xdb), T(0xdc), T(0xdd), T(0xde), T(0xdf),
+  T(0xe0), T(0xe1), T(0xe2), T(0xe3), T(0xe4), T(0xe5), T(0xe6), T(0xe7),
+  T(0xe8), T(0xe9), T(0xea), T(0xeb), T(0xec), T(0xed), T(0xee), T(0xef),
+  T(0xf0), T(0xf1), T(0xf2), T(0xf3), T(0xf4), T(0xf5), T(0xf6), T(0xf7),
+  T(0xf8), T(0xf9), T(0xfa), T(0xfb), T(0xfc), T(0xfd), T(0xfe), 0x01000100 };
+#undef T
+
+void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    uint32 b = src_argb[0];
+    uint32 g = src_argb[1];
+    uint32 r = src_argb[2];
+    const uint32 a = src_argb[3];
+    const uint32 ia = fixed_invtbl8[a] & 0xffff;  // 8.8 fixed point
+    b = (b * ia) >> 8;
+    g = (g * ia) >> 8;
+    r = (r * ia) >> 8;
+    // Clamping should not be necessary but is free in assembly.
+    dst_argb[0] = clamp255(b);
+    dst_argb[1] = clamp255(g);
+    dst_argb[2] = clamp255(r);
+    dst_argb[3] = a;
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
+                               const int32* previous_cumsum, int width) {
+  int32 row_sum[4] = {0, 0, 0, 0};
+  int x;
+  for (x = 0; x < width; ++x) {
+    row_sum[0] += row[x * 4 + 0];
+    row_sum[1] += row[x * 4 + 1];
+    row_sum[2] += row[x * 4 + 2];
+    row_sum[3] += row[x * 4 + 3];
+    cumsum[x * 4 + 0] = row_sum[0]  + previous_cumsum[x * 4 + 0];
+    cumsum[x * 4 + 1] = row_sum[1]  + previous_cumsum[x * 4 + 1];
+    cumsum[x * 4 + 2] = row_sum[2]  + previous_cumsum[x * 4 + 2];
+    cumsum[x * 4 + 3] = row_sum[3]  + previous_cumsum[x * 4 + 3];
+  }
+}
+
+void CumulativeSumToAverageRow_C(const int32* tl, const int32* bl,
+                                int w, int area, uint8* dst, int count) {
+  float ooa = 1.0f / area;
+  int i;
+  for (i = 0; i < count; ++i) {
+    dst[0] = (uint8)((bl[w + 0] + tl[0] - bl[0] - tl[w + 0]) * ooa);
+    dst[1] = (uint8)((bl[w + 1] + tl[1] - bl[1] - tl[w + 1]) * ooa);
+    dst[2] = (uint8)((bl[w + 2] + tl[2] - bl[2] - tl[w + 2]) * ooa);
+    dst[3] = (uint8)((bl[w + 3] + tl[3] - bl[3] - tl[w + 3]) * ooa);
+    dst += 4;
+    tl += 4;
+    bl += 4;
+  }
+}
+
+// Copy pixels from rotated source to destination row with a slope.
+LIBYUV_API
+void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
+                     uint8* dst_argb, const float* uv_dudv, int width) {
+  int i;
+  // Render a row of pixels from source into a buffer.
+  float uv[2];
+  uv[0] = uv_dudv[0];
+  uv[1] = uv_dudv[1];
+  for (i = 0; i < width; ++i) {
+    int x = (int)(uv[0]);
+    int y = (int)(uv[1]);
+    *(uint32*)(dst_argb) =
+        *(const uint32*)(src_argb + y * src_argb_stride +
+                                         x * 4);
+    dst_argb += 4;
+    uv[0] += uv_dudv[2];
+    uv[1] += uv_dudv[3];
+  }
+}
+
+// Blend 2 rows into 1.
+static void HalfRow_C(const uint8* src_uv, int src_uv_stride,
+                      uint8* dst_uv, int pix) {
+  int x;
+  for (x = 0; x < pix; ++x) {
+    dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
+  }
+}
+
+static void HalfRow_16_C(const uint16* src_uv, int src_uv_stride,
+                         uint16* dst_uv, int pix) {
+  int x;
+  for (x = 0; x < pix; ++x) {
+    dst_uv[x] = (src_uv[x] + src_uv[src_uv_stride + x] + 1) >> 1;
+  }
+}
+
+// C version 2x2 -> 2x1.
+void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
+                      ptrdiff_t src_stride,
+                      int width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  int x;
+  if (source_y_fraction == 0) {
+    memcpy(dst_ptr, src_ptr, width);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_C(src_ptr, (int)(src_stride), dst_ptr, width);
+    return;
+  }
+  for (x = 0; x < width - 1; x += 2) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
+    src_ptr += 2;
+    src_ptr1 += 2;
+    dst_ptr += 2;
+  }
+  if (width & 1) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+  }
+}
+
+void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         ptrdiff_t src_stride,
+                         int width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint16* src_ptr1 = src_ptr + src_stride;
+  int x;
+  if (source_y_fraction == 0) {
+    memcpy(dst_ptr, src_ptr, width * 2);
+    return;
+  }
+  if (source_y_fraction == 128) {
+    HalfRow_16_C(src_ptr, (int)(src_stride), dst_ptr, width);
+    return;
+  }
+  for (x = 0; x < width - 1; x += 2) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+    dst_ptr[1] = (src_ptr[1] * y0_fraction + src_ptr1[1] * y1_fraction) >> 8;
+    src_ptr += 2;
+    src_ptr1 += 2;
+    dst_ptr += 2;
+  }
+  if (width & 1) {
+    dst_ptr[0] = (src_ptr[0] * y0_fraction + src_ptr1[0] * y1_fraction) >> 8;
+  }
+}
+
+// Use first 4 shuffler values to reorder ARGB channels.
+void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
+                      const uint8* shuffler, int pix) {
+  int index0 = shuffler[0];
+  int index1 = shuffler[1];
+  int index2 = shuffler[2];
+  int index3 = shuffler[3];
+  // Shuffle a row of ARGB.
+  int x;
+  for (x = 0; x < pix; ++x) {
+    // To support in-place conversion.
+    uint8 b = src_argb[index0];
+    uint8 g = src_argb[index1];
+    uint8 r = src_argb[index2];
+    uint8 a = src_argb[index3];
+    dst_argb[0] = b;
+    dst_argb[1] = g;
+    dst_argb[2] = r;
+    dst_argb[3] = a;
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void I422ToYUY2Row_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_frame, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_frame[0] = src_y[0];
+    dst_frame[1] = src_u[0];
+    dst_frame[2] = src_y[1];
+    dst_frame[3] = src_v[0];
+    dst_frame += 4;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+  }
+  if (width & 1) {
+    dst_frame[0] = src_y[0];
+    dst_frame[1] = src_u[0];
+    dst_frame[2] = 0;
+    dst_frame[3] = src_v[0];
+  }
+}
+
+void I422ToUYVYRow_C(const uint8* src_y,
+                     const uint8* src_u,
+                     const uint8* src_v,
+                     uint8* dst_frame, int width) {
+  int x;
+  for (x = 0; x < width - 1; x += 2) {
+    dst_frame[0] = src_u[0];
+    dst_frame[1] = src_y[0];
+    dst_frame[2] = src_v[0];
+    dst_frame[3] = src_y[1];
+    dst_frame += 4;
+    src_y += 2;
+    src_u += 1;
+    src_v += 1;
+  }
+  if (width & 1) {
+    dst_frame[0] = src_u[0];
+    dst_frame[1] = src_y[0];
+    dst_frame[2] = src_v[0];
+    dst_frame[3] = 0;
+  }
+}
+
+// Maximum temporary width for wrappers to process at a time, in pixels.
+#define MAXTWIDTH 2048
+
+#if !(defined(_MSC_VER) && !defined(__clang__)) && \
+    defined(HAS_I422TORGB565ROW_SSSE3)
+// row_win.cc has asm version, but GCC uses 2 step wrapper.
+void I422ToRGB565Row_SSSE3(const uint8* src_y,
+                           const uint8* src_u,
+                           const uint8* src_v,
+                           uint8* dst_rgb565,
+                           int width) {
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth);
+    ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TOARGB1555ROW_SSSE3)
+void I422ToARGB1555Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb1555,
+                             int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB1555Row_SSE2(row, dst_argb1555, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb1555 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TOARGB4444ROW_SSSE3)
+void I422ToARGB4444Row_SSSE3(const uint8* src_y,
+                             const uint8* src_u,
+                             const uint8* src_v,
+                             uint8* dst_argb4444,
+                             int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_SSSE3(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB4444Row_SSE2(row, dst_argb4444, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb4444 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV12TORGB565ROW_SSSE3)
+void NV12ToRGB565Row_SSSE3(const uint8* src_y, const uint8* src_uv,
+                           uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV12ToARGBRow_SSSE3(src_y, src_uv, row, twidth);
+    ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_uv += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV21TORGB565ROW_SSSE3)
+void NV21ToRGB565Row_SSSE3(const uint8* src_y, const uint8* src_vu,
+                           uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV21ToARGBRow_SSSE3(src_y, src_vu, row, twidth);
+    ARGBToRGB565Row_SSE2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_vu += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_YUY2TOARGBROW_SSSE3)
+void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    YUY2ToUV422Row_SSE2(src_yuy2, row_u, row_v, twidth);
+    YUY2ToYRow_SSE2(src_yuy2, row_y, twidth);
+    I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, twidth);
+    src_yuy2 += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_UYVYTOARGBROW_SSSE3)
+void UYVYToARGBRow_SSSE3(const uint8* src_uyvy, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    UYVYToUV422Row_SSE2(src_uyvy, row_u, row_v, twidth);
+    UYVYToYRow_SSE2(src_uyvy, row_y, twidth);
+    I422ToARGBRow_SSSE3(row_y, row_u, row_v, dst_argb, twidth);
+    src_uyvy += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif  // !defined(LIBYUV_DISABLE_X86)
+
+#if defined(HAS_I422TORGB565ROW_AVX2)
+void I422ToRGB565Row_AVX2(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TOARGB1555ROW_AVX2)
+void I422ToARGB1555Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB1555Row_AVX2(row, dst_argb1555, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb1555 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TOARGB4444ROW_AVX2)
+void I422ToARGB4444Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    ARGBToARGB4444Row_AVX2(row, dst_argb4444, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_argb4444 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TORGB24ROW_AVX2)
+void I422ToRGB24Row_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_rgb24,
+                            int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    // TODO(fbarchard): ARGBToRGB24Row_AVX2
+    ARGBToRGB24Row_SSSE3(row, dst_rgb24, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_rgb24 += twidth * 3;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_I422TORAWROW_AVX2)
+void I422ToRAWRow_AVX2(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_raw,
+                            int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    I422ToARGBRow_AVX2(src_y, src_u, src_v, row, twidth);
+    // TODO(fbarchard): ARGBToRAWRow_AVX2
+    ARGBToRAWRow_SSSE3(row, dst_raw, twidth);
+    src_y += twidth;
+    src_u += twidth / 2;
+    src_v += twidth / 2;
+    dst_raw += twidth * 3;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV12TORGB565ROW_AVX2)
+void NV12ToRGB565Row_AVX2(const uint8* src_y, const uint8* src_uv,
+                          uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV12ToARGBRow_AVX2(src_y, src_uv, row, twidth);
+    ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_uv += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_NV21TORGB565ROW_AVX2)
+void NV21ToRGB565Row_AVX2(const uint8* src_y, const uint8* src_vu,
+                          uint8* dst_rgb565, int width) {
+  // Row buffer for intermediate ARGB pixels.
+  SIMD_ALIGNED32(uint8 row[MAXTWIDTH * 4]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    NV21ToARGBRow_AVX2(src_y, src_vu, row, twidth);
+    ARGBToRGB565Row_AVX2(row, dst_rgb565, twidth);
+    src_y += twidth;
+    src_vu += twidth;
+    dst_rgb565 += twidth * 2;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_YUY2TOARGBROW_AVX2)
+void YUY2ToARGBRow_AVX2(const uint8* src_yuy2, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED32(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED32(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED32(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    YUY2ToUV422Row_AVX2(src_yuy2, row_u, row_v, twidth);
+    YUY2ToYRow_AVX2(src_yuy2, row_y, twidth);
+    I422ToARGBRow_AVX2(row_y, row_u, row_v, dst_argb, twidth);
+    src_yuy2 += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif
+
+#if defined(HAS_UYVYTOARGBROW_AVX2)
+void UYVYToARGBRow_AVX2(const uint8* src_uyvy, uint8* dst_argb, int width) {
+  // Row buffers for intermediate YUV pixels.
+  SIMD_ALIGNED32(uint8 row_y[MAXTWIDTH]);
+  SIMD_ALIGNED32(uint8 row_u[MAXTWIDTH / 2]);
+  SIMD_ALIGNED32(uint8 row_v[MAXTWIDTH / 2]);
+  while (width > 0) {
+    int twidth = width > MAXTWIDTH ? MAXTWIDTH : width;
+    UYVYToUV422Row_AVX2(src_uyvy, row_u, row_v, twidth);
+    UYVYToYRow_AVX2(src_uyvy, row_y, twidth);
+    I422ToARGBRow_AVX2(row_y, row_u, row_v, dst_argb, twidth);
+    src_uyvy += twidth * 2;
+    dst_argb += twidth * 4;
+    width -= twidth;
+  }
+}
+#endif  // !defined(LIBYUV_DISABLE_X86)
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width) {
+  int i;
+  for (i = 0; i < width; ++i) {
+    float b = (float)(src_argb[0]);
+    float g = (float)(src_argb[1]);
+    float r = (float)(src_argb[2]);
+    float a = (float)(src_argb[3]);
+    float b2 = b * b;
+    float g2 = g * g;
+    float r2 = r * r;
+    float a2 = a * a;
+    float db = poly[0] + poly[4] * b;
+    float dg = poly[1] + poly[5] * g;
+    float dr = poly[2] + poly[6] * r;
+    float da = poly[3] + poly[7] * a;
+    float b3 = b2 * b;
+    float g3 = g2 * g;
+    float r3 = r2 * r;
+    float a3 = a2 * a;
+    db += poly[8] * b2;
+    dg += poly[9] * g2;
+    dr += poly[10] * r2;
+    da += poly[11] * a2;
+    db += poly[12] * b3;
+    dg += poly[13] * g3;
+    dr += poly[14] * r3;
+    da += poly[15] * a3;
+
+    dst_argb[0] = Clamp((int32)(db));
+    dst_argb[1] = Clamp((int32)(dg));
+    dst_argb[2] = Clamp((int32)(dr));
+    dst_argb[3] = Clamp((int32)(da));
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
+void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
+                             const uint8* luma, uint32 lumacoeff) {
+  uint32 bc = lumacoeff & 0xff;
+  uint32 gc = (lumacoeff >> 8) & 0xff;
+  uint32 rc = (lumacoeff >> 16) & 0xff;
+
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
+                           src_argb[2] * rc) & 0x7F00u) + luma;
+    const uint8* luma1;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+    luma1 = ((src_argb[4] * bc + src_argb[5] * gc +
+              src_argb[6] * rc) & 0x7F00u) + luma;
+    dst_argb[4] = luma1[src_argb[4]];
+    dst_argb[5] = luma1[src_argb[5]];
+    dst_argb[6] = luma1[src_argb[6]];
+    dst_argb[7] = src_argb[7];
+    src_argb += 8;
+    dst_argb += 8;
+  }
+  if (width & 1) {
+    // Luminance in rows, color values in columns.
+    const uint8* luma0 = ((src_argb[0] * bc + src_argb[1] * gc +
+                           src_argb[2] * rc) & 0x7F00u) + luma;
+    dst_argb[0] = luma0[src_argb[0]];
+    dst_argb[1] = luma0[src_argb[1]];
+    dst_argb[2] = luma0[src_argb[2]];
+    dst_argb[3] = src_argb[3];
+  }
+}
+
+void ARGBCopyAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst[3] = src[3];
+    dst[7] = src[7];
+    dst += 8;
+    src += 8;
+  }
+  if (width & 1) {
+    dst[3] = src[3];
+  }
+}
+
+void ARGBCopyYToAlphaRow_C(const uint8* src, uint8* dst, int width) {
+  int i;
+  for (i = 0; i < width - 1; i += 2) {
+    dst[3] = src[0];
+    dst[7] = src[1];
+    dst += 8;
+    src += 2;
+  }
+  if (width & 1) {
+    dst[3] = src[0];
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/row_gcc.cc b/third_party/aom/third_party/libyuv/source/row_gcc.cc
new file mode 100644
index 0000000000..820de0a1c6
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/row_gcc.cc
@@ -0,0 +1,5475 @@
+// VERSION 2
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
+
+// Constants for ARGB
+static vec8 kARGBToY = {
+  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
+};
+
+// JPeg full range.
+static vec8 kARGBToYJ = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+#endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3)
+
+#if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
+
+static vec8 kARGBToU = {
+  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
+};
+
+static vec8 kARGBToUJ = {
+  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
+};
+
+static vec8 kARGBToV = {
+  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+};
+
+static vec8 kARGBToVJ = {
+  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
+};
+
+// Constants for BGRA
+static vec8 kBGRAToY = {
+  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
+};
+
+static vec8 kBGRAToU = {
+  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
+};
+
+static vec8 kBGRAToV = {
+  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
+};
+
+// Constants for ABGR
+static vec8 kABGRToY = {
+  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
+};
+
+static vec8 kABGRToU = {
+  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
+};
+
+static vec8 kABGRToV = {
+  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
+};
+
+// Constants for RGBA.
+static vec8 kRGBAToY = {
+  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
+};
+
+static vec8 kRGBAToU = {
+  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
+};
+
+static vec8 kRGBAToV = {
+  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
+};
+
+static uvec8 kAddY16 = {
+  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
+};
+
+// 7 bit fixed point 0.5.
+static vec16 kAddYJ64 = {
+  64, 64, 64, 64, 64, 64, 64, 64
+};
+
+static uvec8 kAddUV128 = {
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
+};
+
+static uvec16 kAddUVJ128 = {
+  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
+};
+#endif  // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3)
+
+#ifdef HAS_RGB24TOARGBROW_SSSE3
+
+// Shuffle table for converting RGB24 to ARGB.
+static uvec8 kShuffleMaskRGB24ToARGB = {
+  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
+};
+
+// Shuffle table for converting RAW to ARGB.
+static uvec8 kShuffleMaskRAWToARGB = {
+  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
+};
+
+// Shuffle table for converting ARGB to RGB24.
+static uvec8 kShuffleMaskARGBToRGB24 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static uvec8 kShuffleMaskARGBToRAW = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
+static uvec8 kShuffleMaskARGBToRGB24_0 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static uvec8 kShuffleMaskARGBToRAW_0 = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u
+};
+#endif  // HAS_RGB24TOARGBROW_SSSE3
+
+#if defined(TESTING) && defined(__x86_64__)
+void TestRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  asm volatile (
+    ".p2align  5                               \n"
+    "mov       %%eax,%%eax                     \n"
+    "mov       %%ebx,%%ebx                     \n"
+    "mov       %%ecx,%%ecx                     \n"
+    "mov       %%edx,%%edx                     \n"
+    "mov       %%esi,%%esi                     \n"
+    "mov       %%edi,%%edi                     \n"
+    "mov       %%ebp,%%ebp                     \n"
+    "mov       %%esp,%%esp                     \n"
+    ".p2align  5                               \n"
+    "mov       %%r8d,%%r8d                     \n"
+    "mov       %%r9d,%%r9d                     \n"
+    "mov       %%r10d,%%r10d                   \n"
+    "mov       %%r11d,%%r11d                   \n"
+    "mov       %%r12d,%%r12d                   \n"
+    "mov       %%r13d,%%r13d                   \n"
+    "mov       %%r14d,%%r14d                   \n"
+    "mov       %%r15d,%%r15d                   \n"
+    ".p2align  5                               \n"
+    "lea       (%%rax),%%eax                   \n"
+    "lea       (%%rbx),%%ebx                   \n"
+    "lea       (%%rcx),%%ecx                   \n"
+    "lea       (%%rdx),%%edx                   \n"
+    "lea       (%%rsi),%%esi                   \n"
+    "lea       (%%rdi),%%edi                   \n"
+    "lea       (%%rbp),%%ebp                   \n"
+    "lea       (%%rsp),%%esp                   \n"
+    ".p2align  5                               \n"
+    "lea       (%%r8),%%r8d                    \n"
+    "lea       (%%r9),%%r9d                    \n"
+    "lea       (%%r10),%%r10d                  \n"
+    "lea       (%%r11),%%r11d                  \n"
+    "lea       (%%r12),%%r12d                  \n"
+    "lea       (%%r13),%%r13d                  \n"
+    "lea       (%%r14),%%r14d                  \n"
+    "lea       (%%r15),%%r15d                  \n"
+
+    ".p2align  5                               \n"
+    "lea       0x10(%%rax),%%eax               \n"
+    "lea       0x10(%%rbx),%%ebx               \n"
+    "lea       0x10(%%rcx),%%ecx               \n"
+    "lea       0x10(%%rdx),%%edx               \n"
+    "lea       0x10(%%rsi),%%esi               \n"
+    "lea       0x10(%%rdi),%%edi               \n"
+    "lea       0x10(%%rbp),%%ebp               \n"
+    "lea       0x10(%%rsp),%%esp               \n"
+    ".p2align  5                               \n"
+    "lea       0x10(%%r8),%%r8d                \n"
+    "lea       0x10(%%r9),%%r9d                \n"
+    "lea       0x10(%%r10),%%r10d              \n"
+    "lea       0x10(%%r11),%%r11d              \n"
+    "lea       0x10(%%r12),%%r12d              \n"
+    "lea       0x10(%%r13),%%r13d              \n"
+    "lea       0x10(%%r14),%%r14d              \n"
+    "lea       0x10(%%r15),%%r15d              \n"
+
+    ".p2align  5                               \n"
+    "add       0x10,%%eax                      \n"
+    "add       0x10,%%ebx                      \n"
+    "add       0x10,%%ecx                      \n"
+    "add       0x10,%%edx                      \n"
+    "add       0x10,%%esi                      \n"
+    "add       0x10,%%edi                      \n"
+    "add       0x10,%%ebp                      \n"
+    "add       0x10,%%esp                      \n"
+    ".p2align  5                               \n"
+    "add       0x10,%%r8d                      \n"
+    "add       0x10,%%r9d                      \n"
+    "add       0x10,%%r10d                     \n"
+    "add       0x10,%%r11d                     \n"
+    "add       0x10,%%r12d                     \n"
+    "add       0x10,%%r13d                     \n"
+    "add       0x10,%%r14d                     \n"
+    "add       0x10,%%r15d                     \n"
+
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc", "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // TESTING
+
+#ifdef HAS_J400TOARGBROW_SSE2
+void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y),     // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_J400TOARGBROW_SSE2
+
+#ifdef HAS_RGB24TOARGBROW_SSSE3
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
+    "pslld     $0x18,%%xmm5                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x30,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm2                   \n"
+    "palignr   $0x8,%%xmm1,%%xmm2              \n"
+    "pshufb    %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "palignr   $0x4,%%xmm3,%%xmm3              \n"
+    "pshufb    %%xmm4,%%xmm3                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "m"(kShuffleMaskRGB24ToARGB)  // %3
+  : "memory", "cc" , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"  // generate mask 0xff000000
+    "pslld     $0x18,%%xmm5                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x30,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm2                   \n"
+    "palignr   $0x8,%%xmm1,%%xmm2              \n"
+    "pshufb    %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "palignr   $0x4,%%xmm3,%%xmm3              \n"
+    "pshufb    %%xmm4,%%xmm3                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "m"(kShuffleMaskRAWToARGB)  // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void RGB565ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0x1080108,%%eax                \n"
+    "movd      %%eax,%%xmm5                    \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "mov       $0x20802080,%%eax               \n"
+    "movd      %%eax,%%xmm6                    \n"
+    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psllw     $0xb,%%xmm3                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0xa,%%xmm4                     \n"
+    "psrlw     $0x5,%%xmm4                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psllw     $0x8,%%xmm7                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "psllw     $0xb,%%xmm2                     \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "psllw     $0x8,%%xmm1                     \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "pmulhuw   %%xmm6,%%xmm0                   \n"
+    "por       %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpckhbw %%xmm0,%%xmm2                   \n"
+    MEMOPMEM(movdqu,xmm1,0x00,1,0,2)           //  movdqu  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqu,xmm2,0x10,1,0,2)           //  movdqu  %%xmm2,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void ARGB1555ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0x1080108,%%eax                \n"
+    "movd      %%eax,%%xmm5                    \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "mov       $0x42004200,%%eax               \n"
+    "movd      %%eax,%%xmm6                    \n"
+    "pshufd    $0x0,%%xmm6,%%xmm6              \n"
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psllw     $0xb,%%xmm3                     \n"
+    "movdqa    %%xmm3,%%xmm4                   \n"
+    "psrlw     $0x6,%%xmm4                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psllw     $0x8,%%xmm7                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psllw     $0x1,%%xmm1                     \n"
+    "psllw     $0xb,%%xmm2                     \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "psllw     $0x8,%%xmm1                     \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "pmulhuw   %%xmm6,%%xmm0                   \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm1                   \n"
+    "punpckhbw %%xmm0,%%xmm2                   \n"
+    MEMOPMEM(movdqu,xmm1,0x00,1,0,2)           //  movdqu  %%xmm1,(%1,%0,2)
+    MEMOPMEM(movdqu,xmm2,0x10,1,0,2)           //  movdqu  %%xmm2,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void ARGB4444ToARGBRow_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "mov       $0xf0f0f0f,%%eax                \n"
+    "movd      %%eax,%%xmm4                    \n"
+    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "pslld     $0x4,%%xmm5                     \n"
+    "sub       %0,%1                           \n"
+    "sub       %0,%1                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "psllw     $0x4,%%xmm1                     \n"
+    "psrlw     $0x4,%%xmm3                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,2)           //  movdqu  %%xmm0,(%1,%0,2)
+    MEMOPMEM(movdqu,xmm1,0x10,1,0,2)           //  movdqu  %%xmm1,0x10(%1,%0,2)
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :
+  : "memory", "cc", "eax", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void ARGBToRGB24Row_SSSE3(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "pshufb    %%xmm6,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm6,%%xmm2                   \n"
+    "pshufb    %%xmm6,%%xmm3                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "psrldq    $0x4,%%xmm1                     \n"
+    "pslldq    $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm2,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pslldq    $0x8,%%xmm5                     \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "psrldq    $0x8,%%xmm2                     \n"
+    "pslldq    $0x4,%%xmm3                     \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x30,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "m"(kShuffleMaskARGBToRGB24)  // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+void ARGBToRAWRow_SSSE3(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm6                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "pshufb    %%xmm6,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm6,%%xmm2                   \n"
+    "pshufb    %%xmm6,%%xmm3                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "psrldq    $0x4,%%xmm1                     \n"
+    "pslldq    $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm2,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pslldq    $0x8,%%xmm5                     \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "psrldq    $0x8,%%xmm2                     \n"
+    "pslldq    $0x4,%%xmm3                     \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x30,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  : "m"(kShuffleMaskARGBToRAW)  // %3
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+void ARGBToRGB565Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "psrld     $0x1b,%%xmm3                    \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psrld     $0x1a,%%xmm4                    \n"
+    "pslld     $0x5,%%xmm4                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0xb,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "pslld     $0x8,%%xmm0                     \n"
+    "psrld     $0x3,%%xmm1                     \n"
+    "psrld     $0x5,%%xmm2                     \n"
+    "psrad     $0x10,%%xmm0                    \n"
+    "pand      %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm4,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm1                   \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void ARGBToARGB1555Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psrld     $0x1b,%%xmm4                    \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "pslld     $0x5,%%xmm5                     \n"
+    "movdqa    %%xmm4,%%xmm6                   \n"
+    "pslld     $0xa,%%xmm6                     \n"
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "pslld     $0xf,%%xmm7                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "psrad     $0x10,%%xmm0                    \n"
+    "psrld     $0x3,%%xmm1                     \n"
+    "psrld     $0x6,%%xmm2                     \n"
+    "psrld     $0x9,%%xmm3                     \n"
+    "pand      %%xmm7,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm6,%%xmm3                   \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "por       %%xmm3,%%xmm2                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :: "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void ARGBToARGB4444Row_SSE2(const uint8* src, uint8* dst, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0xc,%%xmm4                     \n"
+    "movdqa    %%xmm4,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm3,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm1                   \n"
+    "psrlq     $0x4,%%xmm0                     \n"
+    "psrlq     $0x8,%%xmm1                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(pix)   // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+#endif  // HAS_RGB24TOARGBROW_SSSE3
+
+#ifdef HAS_ARGBTOYROW_SSSE3
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBTOYROW_SSSE3
+
+#ifdef HAS_ARGBTOYJROW_SSSE3
+// Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
+// Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),  // %3
+    "m"(kAddYJ64)    // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBTOYJROW_SSSE3
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// vpermd for vphaddw + vpackuswb vpermd.
+static const lvec32 kPermdARGBToY_AVX = {
+  0, 4, 1, 5, 2, 6, 3, 7
+};
+
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vbroadcastf128 %3,%%ymm4                  \n"
+    "vbroadcastf128 %4,%%ymm5                  \n"
+    "vmovdqu    %5,%%ymm6                      \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
+    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
+    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
+    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
+    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
+    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
+    "lea       " MEMLEA(0x80,0) ",%0           \n"
+    "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"  // mutates.
+    "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"
+    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x7,%%ymm2,%%ymm2             \n"
+    "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
+    "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
+    "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"  // add 16 for Y
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToY),   // %3
+    "m"(kAddY16),    // %4
+    "m"(kPermdARGBToY_AVX)  // %5
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBTOYROW_AVX2
+
+#ifdef HAS_ARGBTOYJROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vbroadcastf128 %3,%%ymm4                  \n"
+    "vbroadcastf128 %4,%%ymm5                  \n"
+    "vmovdqu    %5,%%ymm6                      \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
+    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
+    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
+    "vpmaddubsw %%ymm4,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm4,%%ymm1,%%ymm1           \n"
+    "vpmaddubsw %%ymm4,%%ymm2,%%ymm2           \n"
+    "vpmaddubsw %%ymm4,%%ymm3,%%ymm3           \n"
+    "lea       " MEMLEA(0x80,0) ",%0           \n"
+    "vphaddw    %%ymm1,%%ymm0,%%ymm0           \n"  // mutates.
+    "vphaddw    %%ymm3,%%ymm2,%%ymm2           \n"
+    "vpaddw     %%ymm5,%%ymm0,%%ymm0           \n"  // Add .5 for rounding.
+    "vpaddw     %%ymm5,%%ymm2,%%ymm2           \n"
+    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x7,%%ymm2,%%ymm2             \n"
+    "vpackuswb  %%ymm2,%%ymm0,%%ymm0           \n"  // mutates.
+    "vpermd     %%ymm0,%%ymm6,%%ymm0           \n"  // unmutate.
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kARGBToYJ),   // %3
+    "m"(kAddYJ64),    // %4
+    "m"(kPermdARGBToY_AVX)  // %5
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBTOYJROW_AVX2
+
+#ifdef HAS_ARGBTOUVROW_SSSE3
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps    %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)), // %4
+    "m"(kARGBToV),  // %5
+    "m"(kARGBToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBTOUVROW_SSSE3
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+// vpshufb for vphaddw + vpackuswb packed to shorts.
+static const lvec8 kShufARGBToUV_AVX = {
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15
+};
+void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "vbroadcastf128 %5,%%ymm5                  \n"
+    "vbroadcastf128 %6,%%ymm6                  \n"
+    "vbroadcastf128 %7,%%ymm7                  \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1  \n"
+    "vmovdqu    " MEMACCESS2(0x40,0) ",%%ymm2  \n"
+    "vmovdqu    " MEMACCESS2(0x60,0) ",%%ymm3  \n"
+    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
+    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
+    VMEMOPREG(vpavgb,0x40,0,4,1,ymm2,ymm2)
+    VMEMOPREG(vpavgb,0x60,0,4,1,ymm3,ymm3)
+    "lea       " MEMLEA(0x80,0) ",%0           \n"
+    "vshufps    $0x88,%%ymm1,%%ymm0,%%ymm4     \n"
+    "vshufps    $0xdd,%%ymm1,%%ymm0,%%ymm0     \n"
+    "vpavgb     %%ymm4,%%ymm0,%%ymm0           \n"
+    "vshufps    $0x88,%%ymm3,%%ymm2,%%ymm4     \n"
+    "vshufps    $0xdd,%%ymm3,%%ymm2,%%ymm2     \n"
+    "vpavgb     %%ymm4,%%ymm2,%%ymm2           \n"
+
+    "vpmaddubsw %%ymm7,%%ymm0,%%ymm1           \n"
+    "vpmaddubsw %%ymm7,%%ymm2,%%ymm3           \n"
+    "vpmaddubsw %%ymm6,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm6,%%ymm2,%%ymm2           \n"
+    "vphaddw    %%ymm3,%%ymm1,%%ymm1           \n"
+    "vphaddw    %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpsraw     $0x8,%%ymm1,%%ymm1             \n"
+    "vpsraw     $0x8,%%ymm0,%%ymm0             \n"
+    "vpacksswb  %%ymm0,%%ymm1,%%ymm0           \n"
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpshufb    %8,%%ymm0,%%ymm0               \n"
+    "vpaddb     %%ymm5,%%ymm0,%%ymm0           \n"
+
+    "vextractf128 $0x0,%%ymm0," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,1,ymm0,0x0,1,2,1) // vextractf128 $1,%%ymm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x20,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)), // %4
+    "m"(kAddUV128),  // %5
+    "m"(kARGBToV),   // %6
+    "m"(kARGBToU),   // %7
+    "m"(kShufARGBToUV_AVX)  // %8
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBTOUVROW_AVX2
+
+#ifdef HAS_ARGBTOUVJROW_SSSE3
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_argb)), // %4
+    "m"(kARGBToVJ),  // %5
+    "m"(kARGBToUJ),  // %6
+    "m"(kAddUVJ128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBTOUVJROW_SSSE3
+
+#ifdef HAS_ARGBTOUV444ROW_SSSE3
+void ARGBToUV444Row_SSSE3(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                          int width) {
+  asm volatile (
+    "movdqa    %4,%%xmm3                       \n"
+    "movdqa    %5,%%xmm4                       \n"
+    "movdqa    %6,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm2                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm2                     \n"
+    "packsswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,2,1)           //  movdqu  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),        // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "m"(kARGBToV),  // %4
+    "m"(kARGBToU),  // %5
+    "m"(kAddUV128)  // %6
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6"
+  );
+}
+#endif  // HAS_ARGBTOUV444ROW_SSSE3
+
+#ifdef HAS_ARGBTOUV422ROW_SSSE3
+void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %4,%%xmm3                       \n"
+    "movdqa    %5,%%xmm4                       \n"
+    "movdqa    %6,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "m"(kARGBToV),  // %4
+    "m"(kARGBToU),  // %5
+    "m"(kAddUV128)  // %6
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBTOUV422ROW_SSSE3
+
+void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kBGRAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void BGRAToUVRow_SSSE3(const uint8* src_bgra0, int src_stride_bgra,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_bgra0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_bgra)), // %4
+    "m"(kBGRAToV),  // %5
+    "m"(kBGRAToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+
+void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kABGRToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "movdqa    %4,%%xmm5                       \n"
+    "movdqa    %3,%%xmm4                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm4,%%xmm3                   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "phaddw    %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm2                     \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  : "m"(kRGBAToY),   // %3
+    "m"(kAddY16)     // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void ABGRToUVRow_SSSE3(const uint8* src_abgr0, int src_stride_abgr,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_abgr0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_abgr)), // %4
+    "m"(kABGRToV),  // %5
+    "m"(kABGRToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+
+void RGBAToUVRow_SSSE3(const uint8* src_rgba0, int src_stride_rgba,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  asm volatile (
+    "movdqa    %5,%%xmm3                       \n"
+    "movdqa    %6,%%xmm4                       \n"
+    "movdqa    %7,%%xmm5                       \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm7)            //  movdqu (%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x10,0,4,1,xmm7)            //  movdqu 0x10(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    MEMOPREG(movdqu,0x20,0,4,1,xmm7)            //  movdqu 0x20(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x30,0,4,1,xmm7)            //  movdqu 0x30(%0,%4,1),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm7                   \n"
+    "shufps    $0x88,%%xmm6,%%xmm2             \n"
+    "shufps    $0xdd,%%xmm6,%%xmm7             \n"
+    "pavgb     %%xmm7,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm2                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "phaddw    %%xmm2,%%xmm0                   \n"
+    "phaddw    %%xmm6,%%xmm1                   \n"
+    "psraw     $0x8,%%xmm0                     \n"
+    "psraw     $0x8,%%xmm1                     \n"
+    "packsswb  %%xmm1,%%xmm0                   \n"
+    "paddb     %%xmm5,%%xmm0                   \n"
+    "movlps    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhps,xmm0,0x00,1,2,1)           //  movhps  %%xmm0,(%1,%2,1)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_rgba0),       // %0
+    "+r"(dst_u),           // %1
+    "+r"(dst_v),           // %2
+    "+rm"(width)           // %3
+  : "r"((intptr_t)(src_stride_rgba)), // %4
+    "m"(kRGBAToV),  // %5
+    "m"(kRGBAToU),  // %6
+    "m"(kAddUV128)  // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm6", "xmm7"
+  );
+}
+
+#if defined(HAS_I422TOARGBROW_SSSE3) || defined(HAS_I422TOARGBROW_AVX2)
+
+struct YuvConstants {
+  lvec8 kUVToB;     // 0
+  lvec8 kUVToG;     // 32
+  lvec8 kUVToR;     // 64
+  lvec16 kUVBiasB;  // 96
+  lvec16 kUVBiasG;  // 128
+  lvec16 kUVBiasR;  // 160
+  lvec16 kYToRgb;   // 192
+};
+
+// BT.601 YUV to RGB reference
+//  R = (Y - 16) * 1.164              - V * -1.596
+//  G = (Y - 16) * 1.164 - U *  0.391 - V *  0.813
+//  B = (Y - 16) * 1.164 - U * -2.018
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* max(-128, round(-2.018 * 64)) */
+#define UG 25 /* round(0.391 * 64) */
+#define VG 52 /* round(0.813 * 64) */
+#define VR -102 /* round(-1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            + YGB)
+#define BG (UG * 128 + VG * 128 + YGB)
+#define BR            (VR * 128 + YGB)
+
+// BT601 constants for YUV to RGB.
+static YuvConstants SIMD_ALIGNED(kYuvConstants) = {
+  { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
+    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 },
+  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
+    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
+  { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
+    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
+};
+
+// BT601 constants for NV21 where chroma plane is VU instead of UV.
+static YuvConstants SIMD_ALIGNED(kYvuConstants) = {
+  { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
+    0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB },
+  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
+    VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
+  { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
+    VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
+};
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+// JPEG YUV to RGB reference
+// *  R = Y                - V * -1.40200
+// *  G = Y - U *  0.34414 - V *  0.71414
+// *  B = Y - U * -1.77200
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YGJ 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+#define YGBJ 32  /* 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UBJ -113 /* round(-1.77200 * 64) */
+#define UGJ 22 /* round(0.34414 * 64) */
+#define VGJ 46 /* round(0.71414  * 64) */
+#define VRJ -90 /* round(-1.40200 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BBJ (UBJ * 128             + YGBJ)
+#define BGJ (UGJ * 128 + VGJ * 128 + YGBJ)
+#define BRJ             (VRJ * 128 + YGBJ)
+
+// JPEG constants for YUV to RGB.
+YuvConstants SIMD_ALIGNED(kYuvJConstants) = {
+  { UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0,
+    UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0 },
+  { UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ },
+  { 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ,
+    0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ },
+  { BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ,
+    BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ },
+  { BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ,
+    BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ },
+  { BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ,
+    BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ },
+  { YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ,
+    YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ }
+};
+
+#undef YGJ
+#undef YGBJ
+#undef UBJ
+#undef UGJ
+#undef VGJ
+#undef VRJ
+#undef BBJ
+#undef BGJ
+#undef BRJ
+
+// Read 8 UV from 411
+#define READYUV444                                                             \
+    "movq       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    MEMOPREG(movq, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"
+
+// Read 4 UV from 422, upsample to 8 UV
+#define READYUV422                                                             \
+    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x4, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"
+
+// Read 2 UV from 411, upsample to 8 UV
+#define READYUV411                                                             \
+    "movd       " MEMACCESS([u_buf]) ",%%xmm0                   \n"            \
+    MEMOPREG(movd, 0x00, [u_buf], [v_buf], 1, xmm1)                            \
+    "lea        " MEMLEA(0x2, [u_buf]) ",%[u_buf]               \n"            \
+    "punpcklbw  %%xmm1,%%xmm0                                   \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"            \
+    "punpckldq  %%xmm0,%%xmm0                                   \n"
+
+// Read 4 UV from NV12, upsample to 8 UV
+#define READNV12                                                               \
+    "movq       " MEMACCESS([uv_buf]) ",%%xmm0                  \n"            \
+    "lea        " MEMLEA(0x8, [uv_buf]) ",%[uv_buf]             \n"            \
+    "punpcklwd  %%xmm0,%%xmm0                                   \n"
+
+// Convert 8 pixels: 8 UV and 8 Y
+#define YUVTORGB(YuvConstants)                                                 \
+    "movdqa     %%xmm0,%%xmm1                                   \n"            \
+    "movdqa     %%xmm0,%%xmm2                                   \n"            \
+    "movdqa     %%xmm0,%%xmm3                                   \n"            \
+    "movdqa     " MEMACCESS2(96, [YuvConstants]) ",%%xmm0       \n"            \
+    "pmaddubsw  " MEMACCESS([YuvConstants]) ",%%xmm1            \n"            \
+    "psubw      %%xmm1,%%xmm0                                   \n"            \
+    "movdqa     " MEMACCESS2(128, [YuvConstants]) ",%%xmm1      \n"            \
+    "pmaddubsw  " MEMACCESS2(32, [YuvConstants]) ",%%xmm2       \n"            \
+    "psubw      %%xmm2,%%xmm1                                   \n"            \
+    "movdqa     " MEMACCESS2(160, [YuvConstants]) ",%%xmm2      \n"            \
+    "pmaddubsw  " MEMACCESS2(64, [YuvConstants]) ",%%xmm3       \n"            \
+    "psubw      %%xmm3,%%xmm2                                   \n"            \
+    "movq       " MEMACCESS([y_buf]) ",%%xmm3                   \n"            \
+    "lea        " MEMLEA(0x8, [y_buf]) ",%[y_buf]               \n"            \
+    "punpcklbw  %%xmm3,%%xmm3                                   \n"            \
+    "pmulhuw    " MEMACCESS2(192, [YuvConstants]) ",%%xmm3      \n"            \
+    "paddsw     %%xmm3,%%xmm0                                   \n"            \
+    "paddsw     %%xmm3,%%xmm1                                   \n"            \
+    "paddsw     %%xmm3,%%xmm2                                   \n"            \
+    "psraw      $0x6,%%xmm0                                     \n"            \
+    "psraw      $0x6,%%xmm1                                     \n"            \
+    "psraw      $0x6,%%xmm2                                     \n"            \
+    "packuswb   %%xmm0,%%xmm0                                   \n"            \
+    "packuswb   %%xmm1,%%xmm1                                   \n"            \
+    "packuswb   %%xmm2,%%xmm2                                   \n"
+
+// Store 8 ARGB values. Assumes XMM5 is zero.
+#define STOREARGB                                                              \
+    "punpcklbw  %%xmm1,%%xmm0                                    \n"           \
+    "punpcklbw  %%xmm5,%%xmm2                                    \n"           \
+    "movdqa     %%xmm0,%%xmm1                                    \n"           \
+    "punpcklwd  %%xmm2,%%xmm0                                    \n"           \
+    "punpckhwd  %%xmm2,%%xmm1                                    \n"           \
+    "movdqu     %%xmm0," MEMACCESS([dst_argb]) "                 \n"           \
+    "movdqu     %%xmm1," MEMACCESS2(0x10, [dst_argb]) "          \n"           \
+    "lea        " MEMLEA(0x20, [dst_argb]) ", %[dst_argb]        \n"
+
+// Store 8 BGRA values. Assumes XMM5 is zero.
+#define STOREBGRA                                                              \
+    "pcmpeqb   %%xmm5,%%xmm5                                     \n"           \
+    "punpcklbw %%xmm0,%%xmm1                                     \n"           \
+    "punpcklbw %%xmm2,%%xmm5                                     \n"           \
+    "movdqa    %%xmm5,%%xmm0                                     \n"           \
+    "punpcklwd %%xmm1,%%xmm5                                     \n"           \
+    "punpckhwd %%xmm1,%%xmm0                                     \n"           \
+    "movdqu    %%xmm5," MEMACCESS([dst_bgra]) "                  \n"           \
+    "movdqu    %%xmm0," MEMACCESS2(0x10, [dst_bgra]) "           \n"           \
+    "lea       " MEMLEA(0x20, [dst_bgra]) ", %[dst_bgra]         \n"
+
+// Store 8 ABGR values. Assumes XMM5 is zero.
+#define STOREABGR                                                              \
+    "punpcklbw %%xmm1,%%xmm2                                     \n"           \
+    "punpcklbw %%xmm5,%%xmm0                                     \n"           \
+    "movdqa    %%xmm2,%%xmm1                                     \n"           \
+    "punpcklwd %%xmm0,%%xmm2                                     \n"           \
+    "punpckhwd %%xmm0,%%xmm1                                     \n"           \
+    "movdqu    %%xmm2," MEMACCESS([dst_abgr]) "                  \n"           \
+    "movdqu    %%xmm1," MEMACCESS2(0x10, [dst_abgr]) "           \n"           \
+    "lea       " MEMLEA(0x20, [dst_abgr]) ", %[dst_abgr]         \n"
+
+// Store 8 RGBA values. Assumes XMM5 is zero.
+#define STORERGBA                                                              \
+    "pcmpeqb   %%xmm5,%%xmm5                                     \n"           \
+    "punpcklbw %%xmm2,%%xmm1                                     \n"           \
+    "punpcklbw %%xmm0,%%xmm5                                     \n"           \
+    "movdqa    %%xmm5,%%xmm0                                     \n"           \
+    "punpcklwd %%xmm1,%%xmm5                                     \n"           \
+    "punpckhwd %%xmm1,%%xmm0                                     \n"           \
+    "movdqu    %%xmm5," MEMACCESS([dst_rgba]) "                  \n"           \
+    "movdqu    %%xmm0," MEMACCESS2(0x10, [dst_rgba]) "           \n"           \
+    "lea       " MEMLEA(0x20, [dst_rgba]) ",%[dst_rgba]          \n"
+
+void OMITFP I444ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV444
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+// TODO(fbarchard): Consider putting masks into constants.
+void OMITFP I422ToRGB24Row_SSSE3(const uint8* y_buf,
+                                 const uint8* u_buf,
+                                 const uint8* v_buf,
+                                 uint8* dst_rgb24,
+                                 int width) {
+  asm volatile (
+    "movdqa    %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRGB24],%%xmm6   \n"
+    "sub       %[u_buf],%[v_buf]               \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "movq      %%xmm0," MEMACCESS([dst_rgb24]) "\n"
+    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_rgb24]) "\n"
+    "lea       " MEMLEA(0x18,[dst_rgb24]) ",%[dst_rgb24] \n"
+    "subl      $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgb24]"+r"(dst_rgb24),  // %[dst_rgb24]
+// TODO(fbarchard): Make width a register for 32 bit.
+#if defined(__i386__) && defined(__pic__)
+    [width]"+m"(width)     // %[width]
+#else
+    [width]"+rm"(width)    // %[width]
+#endif
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB),
+    [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0),
+    [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24)
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6"
+  );
+}
+
+void OMITFP I422ToRAWRow_SSSE3(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_raw,
+                               int width) {
+  asm volatile (
+    "movdqa    %[kShuffleMaskARGBToRAW_0],%%xmm5 \n"
+    "movdqa    %[kShuffleMaskARGBToRAW],%%xmm6   \n"
+    "sub       %[u_buf],%[v_buf]               \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    "punpcklbw %%xmm1,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm0                   \n"
+    "punpckhwd %%xmm2,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm6,%%xmm1                   \n"
+    "palignr   $0xc,%%xmm0,%%xmm1              \n"
+    "movq      %%xmm0," MEMACCESS([dst_raw]) " \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x8,[dst_raw]) "\n"
+    "lea       " MEMLEA(0x18,[dst_raw]) ",%[dst_raw] \n"
+    "subl      $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_raw]"+r"(dst_raw),  // %[dst_raw]
+// TODO(fbarchard): Make width a register for 32 bit.
+#if defined(__i386__) && defined(__pic__)
+    [width]"+m"(width)    // %[width]
+#else
+    [width]"+rm"(width)    // %[width]
+#endif
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB),
+    [kShuffleMaskARGBToRAW_0]"m"(kShuffleMaskARGBToRAW_0),
+    [kShuffleMaskARGBToRAW]"m"(kShuffleMaskARGBToRAW)
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6"
+  );
+}
+
+void OMITFP I422ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP J422ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvJConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP I411ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV411
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP NV12ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* uv_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  // Does not use r14.
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP NV21ToARGBRow_SSSE3(const uint8* y_buf,
+                                const uint8* uv_buf,
+                                uint8* dst_argb,
+                                int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READNV12
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [uv_buf]"+r"(uv_buf),    // %[uv_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYvuConstants.kUVToB) // %[kYuvConstants]
+  // Does not use r14.
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP I422ToBGRARow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_bgra,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREBGRA
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP I422ToABGRRow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_abgr,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREABGR
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_abgr]"+r"(dst_abgr),  // %[dst_abgr]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void OMITFP I422ToRGBARow_SSSE3(const uint8* y_buf,
+                                const uint8* u_buf,
+                                const uint8* v_buf,
+                                uint8* dst_rgba,
+                                int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STORERGBA
+    "sub       $0x8,%[width]                   \n"
+    "jg        1b                              \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_rgba]"+r"(dst_rgba),  // %[dst_rgba]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB) // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+// Read 8 UV from 422, upsample to 16 UV.
+#define READYUV422_AVX2                                                        \
+    "vmovq       " MEMACCESS([u_buf]) ",%%xmm0                      \n"        \
+    MEMOPREG(vmovq, 0x00, [u_buf], [v_buf], 1, xmm1)                           \
+    "lea        " MEMLEA(0x8, [u_buf]) ",%[u_buf]                   \n"        \
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0                                \n"        \
+    "vpermq     $0xd8,%%ymm0,%%ymm0                                 \n"        \
+    "vpunpcklwd %%ymm0,%%ymm0,%%ymm0                                \n"
+
+// Convert 16 pixels: 16 UV and 16 Y.
+#define YUVTORGB_AVX2(YuvConstants)                                            \
+    "vpmaddubsw  " MEMACCESS2(64, [YuvConstants]) ",%%ymm0,%%ymm2   \n"        \
+    "vpmaddubsw  " MEMACCESS2(32, [YuvConstants]) ",%%ymm0,%%ymm1   \n"        \
+    "vpmaddubsw  " MEMACCESS([YuvConstants]) ",%%ymm0,%%ymm0        \n"        \
+    "vmovdqu     " MEMACCESS2(160, [YuvConstants]) ",%%ymm3         \n"        \
+    "vpsubw      %%ymm2,%%ymm3,%%ymm2                               \n"        \
+    "vmovdqu     " MEMACCESS2(128, [YuvConstants]) ",%%ymm3         \n"        \
+    "vpsubw      %%ymm1,%%ymm3,%%ymm1                               \n"        \
+    "vmovdqu     " MEMACCESS2(96, [YuvConstants]) ",%%ymm3          \n"        \
+    "vpsubw      %%ymm0,%%ymm3,%%ymm0                               \n"        \
+    "vmovdqu     " MEMACCESS([y_buf]) ",%%xmm3                      \n"        \
+    "lea         " MEMLEA(0x10, [y_buf]) ",%[y_buf]                 \n"        \
+    "vpermq      $0xd8,%%ymm3,%%ymm3                                \n"        \
+    "vpunpcklbw  %%ymm3,%%ymm3,%%ymm3                               \n"        \
+    "vpmulhuw    " MEMACCESS2(192, [YuvConstants]) ",%%ymm3,%%ymm3  \n"        \
+    "vpaddsw     %%ymm3,%%ymm0,%%ymm0           \n"                            \
+    "vpaddsw     %%ymm3,%%ymm1,%%ymm1           \n"                            \
+    "vpaddsw     %%ymm3,%%ymm2,%%ymm2           \n"                            \
+    "vpsraw      $0x6,%%ymm0,%%ymm0             \n"                            \
+    "vpsraw      $0x6,%%ymm1,%%ymm1             \n"                            \
+    "vpsraw      $0x6,%%ymm2,%%ymm2             \n"                            \
+    "vpackuswb   %%ymm0,%%ymm0,%%ymm0           \n"                            \
+    "vpackuswb   %%ymm1,%%ymm1,%%ymm1           \n"                            \
+    "vpackuswb   %%ymm2,%%ymm2,%%ymm2           \n"
+
+#if defined(HAS_I422TOBGRAROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 BGRA (64 bytes).
+void OMITFP I422ToBGRARow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_bgra,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into BGRA
+    "vpunpcklbw %%ymm0,%%ymm1,%%ymm1           \n"  // GB
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklbw %%ymm2,%%ymm5,%%ymm2           \n"  // AR
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm1,%%ymm2,%%ymm0           \n"  // ARGB first 8 pixels
+    "vpunpckhwd %%ymm1,%%ymm2,%%ymm2           \n"  // ARGB next 8 pixels
+
+    "vmovdqu    %%ymm0," MEMACCESS([dst_bgra]) "\n"
+    "vmovdqu    %%ymm2," MEMACCESS2(0x20,[dst_bgra]) "\n"
+    "lea       " MEMLEA(0x40,[dst_bgra]) ",%[dst_bgra] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_bgra]"+r"(dst_bgra),  // %[dst_bgra]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_I422TOBGRAROW_AVX2
+
+#if defined(HAS_I422TOARGBROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP I422ToARGBRow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ARGB
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0           \n"  // BG
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpunpcklbw %%ymm5,%%ymm2,%%ymm2           \n"  // RA
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm2,%%ymm0,%%ymm1           \n"  // BGRA first 8 pixels
+    "vpunpckhwd %%ymm2,%%ymm0,%%ymm0           \n"  // BGRA next 8 pixels
+
+    "vmovdqu    %%ymm1," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm0," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_I422TOARGBROW_AVX2
+
+#if defined(HAS_J422TOARGBROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+void OMITFP J422ToARGBRow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ARGB
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm0           \n"  // BG
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpunpcklbw %%ymm5,%%ymm2,%%ymm2           \n"  // RA
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm2,%%ymm0,%%ymm1           \n"  // BGRA first 8 pixels
+    "vpunpckhwd %%ymm2,%%ymm0,%%ymm0           \n"  // BGRA next 8 pixels
+
+    "vmovdqu    %%ymm1," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm0," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvJConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_J422TOARGBROW_AVX2
+
+#if defined(HAS_I422TOABGRROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ABGR (64 bytes).
+void OMITFP I422ToABGRRow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ABGR
+    "vpunpcklbw %%ymm1,%%ymm2,%%ymm1           \n"  // RG
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklbw %%ymm5,%%ymm0,%%ymm2           \n"  // BA
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm2,%%ymm1,%%ymm0           \n"  // RGBA first 8 pixels
+    "vpunpckhwd %%ymm2,%%ymm1,%%ymm1           \n"  // RGBA next 8 pixels
+    "vmovdqu    %%ymm0," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm1," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_I422TOABGRROW_AVX2
+
+#if defined(HAS_I422TORGBAROW_AVX2)
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
+void OMITFP I422ToRGBARow_AVX2(const uint8* y_buf,
+                               const uint8* u_buf,
+                               const uint8* v_buf,
+                               uint8* dst_argb,
+                               int width) {
+  asm volatile (
+    "sub       %[u_buf],%[v_buf]               \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    LABELALIGN
+  "1:                                          \n"
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into RGBA
+    "vpunpcklbw %%ymm2,%%ymm1,%%ymm1           \n"
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklbw %%ymm0,%%ymm5,%%ymm2           \n"
+    "vpermq     $0xd8,%%ymm2,%%ymm2            \n"
+    "vpunpcklwd %%ymm1,%%ymm2,%%ymm0           \n"
+    "vpunpckhwd %%ymm1,%%ymm2,%%ymm1           \n"
+    "vmovdqu    %%ymm0," MEMACCESS([dst_argb]) "\n"
+    "vmovdqu    %%ymm1," MEMACCESS2(0x20,[dst_argb]) "\n"
+    "lea       " MEMLEA(0x40,[dst_argb]) ",%[dst_argb] \n"
+    "sub       $0x10,%[width]                  \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : [y_buf]"+r"(y_buf),    // %[y_buf]
+    [u_buf]"+r"(u_buf),    // %[u_buf]
+    [v_buf]"+r"(v_buf),    // %[v_buf]
+    [dst_argb]"+r"(dst_argb),  // %[dst_argb]
+    [width]"+rm"(width)    // %[width]
+  : [kYuvConstants]"r"(&kYuvConstants.kUVToB)  // %[kYuvConstants]
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_I422TORGBAROW_AVX2
+
+#ifdef HAS_I400TOARGBROW_SSE2
+void I400ToARGBRow_SSE2(const uint8* y_buf, uint8* dst_argb, int width) {
+  asm volatile (
+    "mov       $0x4a354a35,%%eax               \n"  // 4a35 = 18997 = 1.164
+    "movd      %%eax,%%xmm2                    \n"
+    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    "mov       $0x04880488,%%eax               \n"  // 0488 = 1160 = 1.164 * 16
+    "movd      %%eax,%%xmm3                    \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    LABELALIGN
+  "1:                                          \n"
+    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "psubusw   %%xmm3,%%xmm0                   \n"
+    "psrlw     $6, %%xmm0                      \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+
+    // Step 2: Weave into ARGB
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm0,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm1                   \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "por       %%xmm4,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(y_buf),     // %0
+    "+r"(dst_argb),  // %1
+    "+rm"(width)     // %2
+  :
+  : "memory", "cc", "eax"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+#endif  // HAS_I400TOARGBROW_SSE2
+
+#ifdef HAS_I400TOARGBROW_AVX2
+// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
+// note: vpunpcklbw mutates and vpackuswb unmutates.
+void I400ToARGBRow_AVX2(const uint8* y_buf, uint8* dst_argb, int width) {
+  asm volatile (
+    "mov        $0x4a354a35,%%eax              \n" // 0488 = 1160 = 1.164 * 16
+    "vmovd      %%eax,%%xmm2                   \n"
+    "vbroadcastss %%xmm2,%%ymm2                \n"
+    "mov        $0x4880488,%%eax               \n" // 4a35 = 18997 = 1.164
+    "vmovd      %%eax,%%xmm3                   \n"
+    "vbroadcastss %%xmm3,%%ymm3                \n"
+    "vpcmpeqb   %%ymm4,%%ymm4,%%ymm4           \n"
+    "vpslld     $0x18,%%ymm4,%%ymm4            \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    // Step 1: Scale Y contribution to 16 G values. G = (y - 16) * 1.164
+    "vmovdqu    " MEMACCESS(0) ",%%xmm0        \n"
+    "lea        " MEMLEA(0x10,0) ",%0          \n"
+    "vpermq     $0xd8,%%ymm0,%%ymm0            \n"
+    "vpunpcklbw %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpsubusw   %%ymm3,%%ymm0,%%ymm0           \n"
+    "vpsrlw     $0x6,%%ymm0,%%ymm0             \n"
+    "vpackuswb  %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpunpcklbw %%ymm0,%%ymm0,%%ymm1           \n"
+    "vpermq     $0xd8,%%ymm1,%%ymm1            \n"
+    "vpunpcklwd %%ymm1,%%ymm1,%%ymm0           \n"
+    "vpunpckhwd %%ymm1,%%ymm1,%%ymm1           \n"
+    "vpor       %%ymm4,%%ymm0,%%ymm0           \n"
+    "vpor       %%ymm4,%%ymm1,%%ymm1           \n"
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "vmovdqu    %%ymm1," MEMACCESS2(0x20,1) "  \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub        $0x10,%2                       \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(y_buf),     // %0
+    "+r"(dst_argb),  // %1
+    "+rm"(width)     // %2
+  :
+  : "memory", "cc", "eax"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+#endif  // HAS_I400TOARGBROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static uvec8 kShuffleMirror = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "movdqa    %3,%%xmm5                       \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(movdqu,-0x10,0,2,1,xmm0)          //  movdqu -0x10(%0,%2),%%xmm0
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kShuffleMirror) // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm5"
+  );
+}
+#endif  // HAS_MIRRORROW_SSSE3
+
+#ifdef HAS_MIRRORROW_AVX2
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "vbroadcastf128 %3,%%ymm5                  \n"
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(vmovdqu,-0x20,0,2,1,ymm0)         //  vmovdqu -0x20(%0,%2),%%ymm0
+    "vpshufb    %%ymm5,%%ymm0,%%ymm0           \n"
+    "vpermq     $0x4e,%%ymm0,%%ymm0            \n"
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kShuffleMirror) // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm5"
+  );
+}
+#endif  // HAS_MIRRORROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSE2
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    MEMOPREG(movdqu,-0x10,0,2,1,xmm0)          //  movdqu -0x10(%0,%2),%%xmm0
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "psllw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm1,%%xmm0                   \n"
+    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufd    $0x4e,%%xmm0,%%xmm0             \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1)",%1            \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_MIRRORROW_SSE2
+
+#ifdef HAS_MIRRORROW_UV_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static uvec8 kShuffleMirrorUV = {
+  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
+};
+void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
+                       int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "movdqa    %4,%%xmm1                       \n"
+    "lea       " MEMLEA4(-0x10,0,3,2) ",%0     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    "pshufb    %%xmm1,%%xmm0                   \n"
+    "movlpd    %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movhpd,xmm0,0x00,1,2,1)           //  movhpd    %%xmm0,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $8,%3                           \n"
+    "jg        1b                              \n"
+  : "+r"(src),      // %0
+    "+r"(dst_u),    // %1
+    "+r"(dst_v),    // %2
+    "+r"(temp_width)  // %3
+  : "m"(kShuffleMirrorUV)  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_MIRRORROW_UV_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_SSE2
+
+void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "lea       " MEMLEA4(-0x10,0,2,4) ",%0     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufd    $0x1b,%%xmm0,%%xmm0             \n"
+    "lea       " MEMLEA(-0x10,0) ",%0          \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0"
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_SSE2
+
+#ifdef HAS_ARGBMIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec32 kARGBShuffleMirror_AVX2 = {
+  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  intptr_t temp_width = (intptr_t)(width);
+  asm volatile (
+    "vmovdqu    %3,%%ymm5                      \n"
+    LABELALIGN
+  "1:                                          \n"
+    VMEMOPREG(vpermd,-0x20,0,2,4,ymm5,ymm0) // vpermd -0x20(%0,%2,4),ymm5,ymm0
+    "vmovdqu    %%ymm0," MEMACCESS(1) "        \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "sub        $0x8,%2                        \n"
+    "jg         1b                             \n"
+    "vzeroupper                                \n"
+  : "+r"(src),  // %0
+    "+r"(dst),  // %1
+    "+r"(temp_width)  // %2
+  : "m"(kARGBShuffleMirror_AVX2) // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm5"
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_AVX2
+
+#ifdef HAS_SPLITUVROW_AVX2
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5             \n"
+    "vpsrlw     $0x8,%%ymm5,%%ymm5               \n"
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0          \n"
+    "vmovdqu    " MEMACCESS2(0x20,0) ",%%ymm1    \n"
+    "lea        " MEMLEA(0x40,0) ",%0            \n"
+    "vpsrlw     $0x8,%%ymm0,%%ymm2               \n"
+    "vpsrlw     $0x8,%%ymm1,%%ymm3               \n"
+    "vpand      %%ymm5,%%ymm0,%%ymm0             \n"
+    "vpand      %%ymm5,%%ymm1,%%ymm1             \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0             \n"
+    "vpackuswb  %%ymm3,%%ymm2,%%ymm2             \n"
+    "vpermq     $0xd8,%%ymm0,%%ymm0              \n"
+    "vpermq     $0xd8,%%ymm2,%%ymm2              \n"
+    "vmovdqu    %%ymm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(vmovdqu,ymm2,0x00,1,2,1)             //  vmovdqu %%ymm2,(%1,%2)
+    "lea        " MEMLEA(0x20,1) ",%1            \n"
+    "sub        $0x20,%3                         \n"
+    "jg         1b                               \n"
+    "vzeroupper                                  \n"
+  : "+r"(src_uv),     // %0
+    "+r"(dst_u),      // %1
+    "+r"(dst_v),      // %2
+    "+r"(pix)         // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_SPLITUVROW_AVX2
+
+#ifdef HAS_SPLITUVROW_SSE2
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb    %%xmm5,%%xmm5                    \n"
+    "psrlw      $0x8,%%xmm5                      \n"
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqu     " MEMACCESS(0) ",%%xmm0          \n"
+    "movdqu     " MEMACCESS2(0x10,0) ",%%xmm1    \n"
+    "lea        " MEMLEA(0x20,0) ",%0            \n"
+    "movdqa     %%xmm0,%%xmm2                    \n"
+    "movdqa     %%xmm1,%%xmm3                    \n"
+    "pand       %%xmm5,%%xmm0                    \n"
+    "pand       %%xmm5,%%xmm1                    \n"
+    "packuswb   %%xmm1,%%xmm0                    \n"
+    "psrlw      $0x8,%%xmm2                      \n"
+    "psrlw      $0x8,%%xmm3                      \n"
+    "packuswb   %%xmm3,%%xmm2                    \n"
+    "movdqu     %%xmm0," MEMACCESS(1) "          \n"
+    MEMOPMEM(movdqu,xmm2,0x00,1,2,1)             //  movdqu     %%xmm2,(%1,%2)
+    "lea        " MEMLEA(0x10,1) ",%1            \n"
+    "sub        $0x10,%3                         \n"
+    "jg         1b                               \n"
+  : "+r"(src_uv),     // %0
+    "+r"(dst_u),      // %1
+    "+r"(dst_v),      // %2
+    "+r"(pix)         // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_SPLITUVROW_SSE2
+
+#ifdef HAS_MERGEUVROW_AVX2
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    "sub       %0,%1                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0           \n"
+    MEMOPREG(vmovdqu,0x00,0,1,1,ymm1)             //  vmovdqu (%0,%1,1),%%ymm1
+    "lea       " MEMLEA(0x20,0) ",%0             \n"
+    "vpunpcklbw %%ymm1,%%ymm0,%%ymm2             \n"
+    "vpunpckhbw %%ymm1,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm2," MEMACCESS(2) "   \n"
+    "vextractf128 $0x0,%%ymm0," MEMACCESS2(0x10,2) "\n"
+    "vextractf128 $0x1,%%ymm2," MEMACCESS2(0x20,2) "\n"
+    "vextractf128 $0x1,%%ymm0," MEMACCESS2(0x30,2) "\n"
+    "lea       " MEMLEA(0x40,2) ",%2             \n"
+    "sub       $0x20,%3                          \n"
+    "jg        1b                                \n"
+    "vzeroupper                                  \n"
+  : "+r"(src_u),     // %0
+    "+r"(src_v),     // %1
+    "+r"(dst_uv),    // %2
+    "+r"(width)      // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2"
+  );
+}
+#endif  // HAS_MERGEUVROW_AVX2
+
+#ifdef HAS_MERGEUVROW_SSE2
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    "sub       %0,%1                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)             //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm2                     \n"
+    "punpcklbw %%xmm1,%%xmm0                     \n"
+    "punpckhbw %%xmm1,%%xmm2                     \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "           \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,2) "     \n"
+    "lea       " MEMLEA(0x20,2) ",%2             \n"
+    "sub       $0x10,%3                          \n"
+    "jg        1b                                \n"
+  : "+r"(src_u),     // %0
+    "+r"(src_v),     // %1
+    "+r"(dst_uv),    // %2
+    "+r"(width)      // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2"
+  );
+}
+#endif  // HAS_MERGEUVROW_SSE2
+
+#ifdef HAS_COPYROW_SSE2
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_COPYROW_SSE2
+
+#ifdef HAS_COPYROW_AVX
+void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x40,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_COPYROW_AVX
+
+#ifdef HAS_COPYROW_ERMS
+// Multiple of 1.
+void CopyRow_ERMS(const uint8* src, uint8* dst, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep movsb " MEMMOVESTRING(0,1) "          \n"
+  : "+S"(src),  // %0
+    "+D"(dst),  // %1
+    "+c"(width_tmp) // %2
+  :
+  : "memory", "cc"
+  );
+}
+#endif  // HAS_COPYROW_ERMS
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm0,%%xmm0                   \n"
+    "pslld     $0x18,%%xmm0                    \n"
+    "pcmpeqb   %%xmm1,%%xmm1                   \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqu    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "pand      %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm0,%%xmm3                   \n"
+    "pand      %%xmm1,%%xmm4                   \n"
+    "pand      %%xmm1,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYALPHAROW_AVX2
+// width in pixels
+void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm1         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm2   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
+    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
+    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2"
+  );
+}
+#endif  // HAS_ARGBCOPYALPHAROW_AVX2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
+// width in pixels
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm0,%%xmm0                   \n"
+    "pslld     $0x18,%%xmm0                    \n"
+    "pcmpeqb   %%xmm1,%%xmm1                   \n"
+    "psrld     $0x8,%%xmm1                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpckhwd %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm2,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm4         \n"
+    "movdqu    " MEMACCESS2(0x10,1) ",%%xmm5   \n"
+    "pand      %%xmm0,%%xmm2                   \n"
+    "pand      %%xmm0,%%xmm3                   \n"
+    "pand      %%xmm1,%%xmm4                   \n"
+    "pand      %%xmm1,%%xmm5                   \n"
+    "por       %%xmm4,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
+// width in pixels
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    "vpcmpeqb  %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpsrld    $0x8,%%ymm0,%%ymm0              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vpmovzxbd " MEMACCESS(0) ",%%ymm1         \n"
+    "vpmovzxbd " MEMACCESS2(0x8,0) ",%%ymm2    \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "vpslld    $0x18,%%ymm1,%%ymm1             \n"
+    "vpslld    $0x18,%%ymm2,%%ymm2             \n"
+    "vpblendvb %%ymm0," MEMACCESS(1) ",%%ymm1,%%ymm1        \n"
+    "vpblendvb %%ymm0," MEMACCESS2(0x20,1) ",%%ymm2,%%ymm2  \n"
+    "vmovdqu   %%ymm1," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm2," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2"
+  );
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
+
+#ifdef HAS_SETROW_X86
+void SetRow_X86(uint8* dst, uint8 v8, int width) {
+  size_t width_tmp = (size_t)(width >> 2);
+  const uint32 v32 = v8 * 0x01010101;  // Duplicate byte to all bytes.
+  asm volatile (
+    "rep stosl " MEMSTORESTRING(eax,0) "       \n"
+    : "+D"(dst),       // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v32)         // %2
+    : "memory", "cc");
+}
+
+void SetRow_ERMS(uint8* dst, uint8 v8, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep stosb " MEMSTORESTRING(al,0) "        \n"
+    : "+D"(dst),       // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v8)          // %2
+    : "memory", "cc");
+}
+
+void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int width) {
+  size_t width_tmp = (size_t)(width);
+  asm volatile (
+    "rep stosl " MEMSTORESTRING(eax,0) "       \n"
+    : "+D"(dst_argb),  // %0
+      "+c"(width_tmp)  // %1
+    : "a"(v32)         // %2
+    : "memory", "cc");
+}
+#endif  // HAS_SETROW_X86
+
+#ifdef HAS_YUY2TOYROW_SSE2
+void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    MEMOPMEM(movq,xmm1,0x00,1,2,1)             //  movq    %%xmm1,(%1,%2)
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_YUY2TOYROW_SSE2
+
+#ifdef HAS_YUY2TOYROW_AVX2
+void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "lea      " MEMLEA(0x20,1) ",%1            \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
+    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_yuy2))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_yuy2),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpsrlw    $0x8,%%ymm1,%%ymm1              \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "lea      " MEMLEA(0x20,1) ",%1            \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm5"
+  );
+}
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb  %%ymm5,%%ymm5,%%ymm5            \n"
+    "vpsrlw    $0x8,%%ymm5,%%ymm5              \n"
+    "sub       %1,%2                           \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    VMEMOPREG(vpavgb,0x00,0,4,1,ymm0,ymm0)     // vpavgb (%0,%4,1),%%ymm0,%%ymm0
+    VMEMOPREG(vpavgb,0x20,0,4,1,ymm1,ymm1)
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  : "r"((intptr_t)(stride_uyvy))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    "vpsrlw     $0x8,%%ymm5,%%ymm5             \n"
+    "sub       %1,%2                           \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpand     %%ymm5,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm1,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vpand     %%ymm5,%%ymm0,%%ymm1            \n"
+    "vpsrlw    $0x8,%%ymm0,%%ymm0              \n"
+    "vpackuswb %%ymm1,%%ymm1,%%ymm1            \n"
+    "vpackuswb %%ymm0,%%ymm0,%%ymm0            \n"
+    "vpermq    $0xd8,%%ymm1,%%ymm1             \n"
+    "vpermq    $0xd8,%%ymm0,%%ymm0             \n"
+    "vextractf128 $0x0,%%ymm1," MEMACCESS(1) " \n"
+    VEXTOPMEM(vextractf128,0,ymm0,0x00,1,2,1) // vextractf128 $0x0,%%ymm0,(%1,%2,1)
+    "lea      " MEMLEA(0x10,1) ",%1            \n"
+    "sub       $0x20,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_uyvy),    // %0
+    "+r"(dst_u),       // %1
+    "+r"(dst_v),       // %2
+    "+r"(pix)          // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_YUY2TOYROW_AVX2
+
+#ifdef HAS_ARGBBLENDROW_SSE2
+// Blend 8 pixels at a time.
+void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0xf,%%xmm7                     \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x8,%%xmm6                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psllw     $0x8,%%xmm5                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "41:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       41b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop.
+  "91:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "psrlw     $0x8,%%xmm3                     \n"
+    "pshufhw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0xf5,%%xmm3,%%xmm3             \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
+    "jge       91b                             \n"
+  "99:                                         \n"
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBBLENDROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSSE3
+// Shuffle table for isolating alpha.
+static uvec8 kShuffleAlpha = {
+  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
+  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
+};
+
+// Blend 8 pixels at a time
+// Shuffle table for reversing the bytes.
+
+// Same as SSE2, but replaces
+//    psrlw      xmm3, 8          // alpha
+//    pshufhw    xmm3, xmm3,0F5h  // 8 alpha words
+//    pshuflw    xmm3, xmm3,0F5h
+// with..
+//    pshufb     xmm3, kShuffleAlpha // alpha
+
+void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
+                        uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0xf,%%xmm7                     \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x8,%%xmm6                     \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psllw     $0x8,%%xmm5                     \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "40:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        99f                             \n"
+
+    // 1 pixel loop.
+  "91:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm3         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "movdqa    %%xmm3,%%xmm0                   \n"
+    "pxor      %%xmm4,%%xmm3                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm2         \n"
+    "pshufb    %4,%%xmm3                       \n"
+    "pand      %%xmm6,%%xmm2                   \n"
+    "paddw     %%xmm7,%%xmm3                   \n"
+    "pmullw    %%xmm3,%%xmm2                   \n"
+    "movd      " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "por       %%xmm4,%%xmm0                   \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm2                     \n"
+    "paddusb   %%xmm2,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
+    "jge       91b                             \n"
+  "99:                                         \n"
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  : "m"(kShuffleAlpha)  // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBBLENDROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+// Attenuate 4 pixels at a time.
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "pslld     $0x18,%%xmm4                    \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x8,%%xmm5                     \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "pshufhw   $0xff,%%xmm0,%%xmm2             \n"
+    "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pshufhw   $0xff,%%xmm1,%%xmm2             \n"
+    "pshuflw   $0xff,%%xmm2,%%xmm2             \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "pand      %%xmm4,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+// Shuffle table duplicating alpha
+static uvec8 kShuffleAlpha0 = {
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u
+};
+static uvec8 kShuffleAlpha1 = {
+  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u
+};
+// Attenuate 4 pixels at a time.
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "pcmpeqb   %%xmm3,%%xmm3                   \n"
+    "pslld     $0x18,%%xmm3                    \n"
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpcklbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm1,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "punpckhbw %%xmm2,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "pand      %%xmm3,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "por       %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  : "m"(kShuffleAlpha0),  // %3
+    "m"(kShuffleAlpha1)  // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha_AVX2 = {
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
+};
+// Attenuate 8 pixels at a time.
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "vbroadcastf128 %3,%%ymm4                  \n"
+    "vpcmpeqb   %%ymm5,%%ymm5,%%ymm5           \n"
+    "vpslld     $0x18,%%ymm5,%%ymm5            \n"
+    "sub        %0,%1                          \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm6        \n"
+    "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
+    "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
+    "vpshufb    %%ymm4,%%ymm0,%%ymm2           \n"
+    "vpshufb    %%ymm4,%%ymm1,%%ymm3           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpand      %%ymm5,%%ymm6,%%ymm6           \n"
+    "vpsrlw     $0x8,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x8,%%ymm1,%%ymm1             \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpor       %%ymm6,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1)          //  vmovdqu %%ymm0,(%0,%1)
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub        $0x8,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width)        // %2
+  : "m"(kShuffleAlpha_AVX2)  // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+// Unattenuate 4 pixels at a time.
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  uintptr_t alpha = 0;
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "movlhps   %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm2)             //  movd      0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
+    MEMOPREG(movd,0x00,4,3,4,xmm3)             //  movd      0x0(%4,%3,4),%%xmm3
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "movlhps   %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width),       // %2
+    "+r"(alpha)        // %3
+  : "r"(fixed_invtbl8)  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBUNATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kUnattenShuffleAlpha_AVX2 = {
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u
+};
+// Unattenuate 8 pixels at a time.
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  uintptr_t alpha = 0;
+  asm volatile (
+    "sub        %0,%1                          \n"
+    "vbroadcastf128 %5,%%ymm5                  \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    // replace VPGATHER
+    "movzb     " MEMACCESS2(0x03,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm0)             //  vmovd 0x0(%4,%3,4),%%xmm0
+    "movzb     " MEMACCESS2(0x07,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm1)             //  vmovd 0x0(%4,%3,4),%%xmm1
+    "movzb     " MEMACCESS2(0x0b,0) ",%3       \n"
+    "vpunpckldq %%xmm1,%%xmm0,%%xmm6           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm2)             //  vmovd 0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x0f,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm3)             //  vmovd 0x0(%4,%3,4),%%xmm3
+    "movzb     " MEMACCESS2(0x13,0) ",%3       \n"
+    "vpunpckldq %%xmm3,%%xmm2,%%xmm7           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm0)             //  vmovd 0x0(%4,%3,4),%%xmm0
+    "movzb     " MEMACCESS2(0x17,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm1)             //  vmovd 0x0(%4,%3,4),%%xmm1
+    "movzb     " MEMACCESS2(0x1b,0) ",%3       \n"
+    "vpunpckldq %%xmm1,%%xmm0,%%xmm0           \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm2)             //  vmovd 0x0(%4,%3,4),%%xmm2
+    "movzb     " MEMACCESS2(0x1f,0) ",%3       \n"
+    MEMOPREG(vmovd,0x00,4,3,4,xmm3)             //  vmovd 0x0(%4,%3,4),%%xmm3
+    "vpunpckldq %%xmm3,%%xmm2,%%xmm2           \n"
+    "vpunpcklqdq %%xmm7,%%xmm6,%%xmm3          \n"
+    "vpunpcklqdq %%xmm2,%%xmm0,%%xmm0          \n"
+    "vinserti128 $0x1,%%xmm0,%%ymm3,%%ymm3     \n"
+    // end of VPGATHER
+
+    "vmovdqu    " MEMACCESS(0) ",%%ymm6        \n"
+    "vpunpcklbw %%ymm6,%%ymm6,%%ymm0           \n"
+    "vpunpckhbw %%ymm6,%%ymm6,%%ymm1           \n"
+    "vpunpcklwd %%ymm3,%%ymm3,%%ymm2           \n"
+    "vpunpckhwd %%ymm3,%%ymm3,%%ymm3           \n"
+    "vpshufb    %%ymm5,%%ymm2,%%ymm2           \n"
+    "vpshufb    %%ymm5,%%ymm3,%%ymm3           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,0,1,1)          //  vmovdqu %%ymm0,(%0,%1)
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub        $0x8,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(width),       // %2
+    "+r"(alpha)        // %3
+  : "r"(fixed_invtbl8),  // %4
+    "m"(kUnattenShuffleAlpha_AVX2)  // %5
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBUNATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBGRAYROW_SSSE3
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm4,%%xmm0                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm5,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm3   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrld     $0x18,%%xmm2                    \n"
+    "psrld     $0x18,%%xmm3                    \n"
+    "packuswb  %%xmm3,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpcklbw %%xmm2,%%xmm3                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "m"(kARGBToYJ),   // %3
+    "m"(kAddYJ64)     // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBGRAYROW_SSSE3
+
+#ifdef HAS_ARGBSEPIAROW_SSSE3
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+// Constant for ARGB color to sepia tone
+static vec8 kARGBToSepiaB = {
+  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
+};
+
+static vec8 kARGBToSepiaG = {
+  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
+};
+
+static vec8 kARGBToSepiaR = {
+  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
+};
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+  asm volatile (
+    "movdqa    %2,%%xmm2                       \n"
+    "movdqa    %3,%%xmm3                       \n"
+    "movdqa    %4,%%xmm4                       \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "pmaddubsw %%xmm2,%%xmm6                   \n"
+    "phaddw    %%xmm6,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm3,%%xmm5                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm5                   \n"
+    "psrlw     $0x7,%%xmm5                     \n"
+    "packuswb  %%xmm5,%%xmm5                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm5         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm4,%%xmm5                   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "phaddw    %%xmm1,%%xmm5                   \n"
+    "psrlw     $0x7,%%xmm5                     \n"
+    "packuswb  %%xmm5,%%xmm5                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "psrld     $0x18,%%xmm6                    \n"
+    "psrld     $0x18,%%xmm1                    \n"
+    "packuswb  %%xmm1,%%xmm6                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm5                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklwd %%xmm5,%%xmm0                   \n"
+    "punpckhwd %%xmm5,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub       $0x8,%1                         \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),      // %0
+    "+r"(width)          // %1
+  : "m"(kARGBToSepiaB),  // %2
+    "m"(kARGBToSepiaG),  // %3
+    "m"(kARGBToSepiaR)   // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBSEPIAROW_SSSE3
+
+#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// Same as Sepia except matrix is provided.
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width) {
+  asm volatile (
+    "movdqu    " MEMACCESS(3) ",%%xmm5         \n"
+    "pshufd    $0x00,%%xmm5,%%xmm2             \n"
+    "pshufd    $0x55,%%xmm5,%%xmm3             \n"
+    "pshufd    $0xaa,%%xmm5,%%xmm4             \n"
+    "pshufd    $0xff,%%xmm5,%%xmm5             \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "pmaddubsw %%xmm2,%%xmm7                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "pmaddubsw %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm3,%%xmm1                   \n"
+    "phaddsw   %%xmm7,%%xmm0                   \n"
+    "phaddsw   %%xmm1,%%xmm6                   \n"
+    "psraw     $0x6,%%xmm0                     \n"
+    "psraw     $0x6,%%xmm6                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm4,%%xmm1                   \n"
+    "pmaddubsw %%xmm4,%%xmm7                   \n"
+    "phaddsw   %%xmm7,%%xmm1                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm7   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm7                   \n"
+    "phaddsw   %%xmm7,%%xmm6                   \n"
+    "psraw     $0x6,%%xmm1                     \n"
+    "psraw     $0x6,%%xmm6                     \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "punpcklbw %%xmm6,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "punpcklwd %%xmm1,%%xmm0                   \n"
+    "punpckhwd %%xmm1,%%xmm6                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb),      // %1
+    "+r"(width)          // %2
+  : "r"(matrix_argb)     // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
+
+#ifdef HAS_ARGBQUANTIZEROW_SSE2
+// Quantize 4 ARGB pixels (16 bytes).
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "movd      %2,%%xmm2                       \n"
+    "movd      %3,%%xmm3                       \n"
+    "movd      %4,%%xmm4                       \n"
+    "pshuflw   $0x40,%%xmm2,%%xmm2             \n"
+    "pshufd    $0x44,%%xmm2,%%xmm2             \n"
+    "pshuflw   $0x40,%%xmm3,%%xmm3             \n"
+    "pshufd    $0x44,%%xmm3,%%xmm3             \n"
+    "pshuflw   $0x40,%%xmm4,%%xmm4             \n"
+    "pshufd    $0x44,%%xmm4,%%xmm4             \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "pslld     $0x18,%%xmm6                    \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm1         \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "pmullw    %%xmm3,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm7         \n"
+    "pmullw    %%xmm3,%%xmm1                   \n"
+    "pand      %%xmm6,%%xmm7                   \n"
+    "paddw     %%xmm4,%%xmm0                   \n"
+    "paddw     %%xmm4,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "por       %%xmm7,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "sub       $0x4,%1                         \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBQUANTIZEROW_SSE2
+
+#ifdef HAS_ARGBSHADEROW_SSE2
+// Shade 4 pixels at a time by specified value.
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "movd      %3,%%xmm2                       \n"
+    "punpcklbw %%xmm2,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm2                  \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm2,%%xmm1                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(value)       // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2"
+  );
+}
+#endif  // HAS_ARGBSHADEROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+// Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                  \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqu    %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpckhbw %%xmm5,%%xmm3                   \n"
+    "pmulhuw   %%xmm2,%%xmm0                   \n"
+    "pmulhuw   %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    "vpxor      %%ymm5,%%ymm5,%%ymm5           \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm1        \n"
+    "lea        " MEMLEA(0x20,0) ",%0          \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm3        \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "vpunpcklbw %%ymm1,%%ymm1,%%ymm0           \n"
+    "vpunpckhbw %%ymm1,%%ymm1,%%ymm1           \n"
+    "vpunpcklbw %%ymm5,%%ymm3,%%ymm2           \n"
+    "vpunpckhbw %%ymm5,%%ymm3,%%ymm3           \n"
+    "vpmulhuw   %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmulhuw   %%ymm3,%%ymm1,%%ymm1           \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
+    "lea       " MEMLEA(0x20,2) ",%2           \n"
+    "sub        $0x8,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__AVX2__)
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+#endif
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_AVX2
+
+#ifdef HAS_ARGBADDROW_SSE2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_ARGBADDROW_SSE2
+
+#ifdef HAS_ARGBADDROW_AVX2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "lea        " MEMLEA(0x20,0) ",%0          \n"
+    "vpaddusb   " MEMACCESS(1) ",%%ymm0,%%ymm0 \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
+    "lea        " MEMLEA(0x20,2) ",%2          \n"
+    "sub        $0x8,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0"
+  );
+}
+#endif  // HAS_ARGBADDROW_AVX2
+
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels, 4 pixels at a time.
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "psubusb   %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(0) ",%%ymm0        \n"
+    "lea        " MEMLEA(0x20,0) ",%0          \n"
+    "vpsubusb   " MEMACCESS(1) ",%%ymm0,%%ymm0 \n"
+    "lea        " MEMLEA(0x20,1) ",%1          \n"
+    "vmovdqu    %%ymm0," MEMACCESS(2) "        \n"
+    "lea        " MEMLEA(0x20,2) ",%2          \n"
+    "sub        $0x8,%3                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+    , "xmm0"
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_AVX2
+
+#ifdef HAS_SOBELXROW_SSE2
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+    "sub       %0,%3                           \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "movq      " MEMACCESS2(0x2,0) ",%%xmm1    \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "psubw     %%xmm1,%%xmm0                   \n"
+    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
+    MEMOPREG(movq,0x02,0,1,1,xmm2)             //  movq      0x2(%0,%1,1),%%xmm2
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "psubw     %%xmm2,%%xmm1                   \n"
+    MEMOPREG(movq,0x00,0,2,1,xmm2)             //  movq      (%0,%2,1),%%xmm2
+    MEMOPREG(movq,0x02,0,2,1,xmm3)             //  movq      0x2(%0,%2,1),%%xmm3
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "psubw     %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm1                   \n"
+    "pmaxsw    %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    MEMOPMEM(movq,xmm0,0x00,0,3,1)             //  movq      %%xmm0,(%0,%3,1)
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x8,%4                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_SOBELXROW_SSE2
+
+#ifdef HAS_SOBELYROW_SSE2
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movq,0x00,0,1,1,xmm1)             //  movq      (%0,%1,1),%%xmm1
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "psubw     %%xmm1,%%xmm0                   \n"
+    "movq      " MEMACCESS2(0x1,0) ",%%xmm1    \n"
+    MEMOPREG(movq,0x01,0,1,1,xmm2)             //  movq      0x1(%0,%1,1),%%xmm2
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "psubw     %%xmm2,%%xmm1                   \n"
+    "movq      " MEMACCESS2(0x2,0) ",%%xmm2    \n"
+    MEMOPREG(movq,0x02,0,1,1,xmm3)             //  movq      0x2(%0,%1,1),%%xmm3
+    "punpcklbw %%xmm5,%%xmm2                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "psubw     %%xmm3,%%xmm2                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "paddw     %%xmm1,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm1                   \n"
+    "pmaxsw    %%xmm1,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    MEMOPMEM(movq,xmm0,0x00,0,2,1)             //  movq      %%xmm0,(%0,%2,1)
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x8,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_SOBELYROW_SSE2
+
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "punpcklbw %%xmm0,%%xmm2                   \n"
+    "punpckhbw %%xmm0,%%xmm0                   \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "punpcklwd %%xmm2,%%xmm1                   \n"
+    "punpckhwd %%xmm2,%%xmm2                   \n"
+    "por       %%xmm5,%%xmm1                   \n"
+    "por       %%xmm5,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklwd %%xmm0,%%xmm3                   \n"
+    "punpckhwd %%xmm0,%%xmm0                   \n"
+    "por       %%xmm5,%%xmm3                   \n"
+    "por       %%xmm5,%%xmm0                   \n"
+    "movdqu    %%xmm1," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,2) "   \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x20,2) "   \n"
+    "movdqu    %%xmm0," MEMACCESS2(0x30,2) "   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+#endif  // HAS_SOBELROW_SSE2
+
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into a plane.
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "pslld     $0x18,%%xmm5                    \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
+  );
+}
+#endif  // HAS_SOBELTOPLANEROW_SSE2
+
+#ifdef HAS_SOBELXYROW_SSE2
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+
+    // 8 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,1,1,xmm1)           //  movdqu    (%0,%1,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "paddusb   %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm0,%%xmm3                   \n"
+    "punpcklbw %%xmm5,%%xmm3                   \n"
+    "punpckhbw %%xmm5,%%xmm0                   \n"
+    "movdqa    %%xmm1,%%xmm4                   \n"
+    "punpcklbw %%xmm2,%%xmm4                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm6                   \n"
+    "punpcklwd %%xmm3,%%xmm6                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "movdqa    %%xmm1,%%xmm7                   \n"
+    "punpcklwd %%xmm0,%%xmm7                   \n"
+    "punpckhwd %%xmm0,%%xmm1                   \n"
+    "movdqu    %%xmm6," MEMACCESS(2) "         \n"
+    "movdqu    %%xmm4," MEMACCESS2(0x10,2) "   \n"
+    "movdqu    %%xmm7," MEMACCESS2(0x20,2) "   \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x30,2) "   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "sub       $0x10,%3                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_SOBELXYROW_SSE2
+
+#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
+// Creates a table of cumulative sums where each value is a sum of all values
+// above and to the left of the value, inclusive of the value.
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width) {
+  asm volatile (
+    "pxor      %%xmm0,%%xmm0                   \n"
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+    "test      $0xf,%1                         \n"
+    "jne       49f                             \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm2,%%xmm4                   \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
+    "punpckhwd %%xmm1,%%xmm3                   \n"
+    "punpckhbw %%xmm1,%%xmm4                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "punpcklwd %%xmm1,%%xmm4                   \n"
+    "punpckhwd %%xmm1,%%xmm5                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm2         \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x10,2) ",%%xmm3   \n"
+    "paddd     %%xmm0,%%xmm3                   \n"
+    "paddd     %%xmm4,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x20,2) ",%%xmm4   \n"
+    "paddd     %%xmm0,%%xmm4                   \n"
+    "paddd     %%xmm5,%%xmm0                   \n"
+    "movdqu    " MEMACCESS2(0x30,2) ",%%xmm5   \n"
+    "lea       " MEMLEA(0x40,2) ",%2           \n"
+    "paddd     %%xmm0,%%xmm5                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm3," MEMACCESS2(0x10,1) "   \n"
+    "movdqu    %%xmm4," MEMACCESS2(0x20,1) "   \n"
+    "movdqu    %%xmm5," MEMACCESS2(0x30,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "movd      " MEMACCESS(0) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "punpcklbw %%xmm1,%%xmm2                   \n"
+    "punpcklwd %%xmm1,%%xmm2                   \n"
+    "paddd     %%xmm2,%%xmm0                   \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm2         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "movdqu    %%xmm2," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x1,%3                         \n"
+    "jge       10b                             \n"
+
+  "19:                                         \n"
+  : "+r"(row),  // %0
+    "+r"(cumsum),  // %1
+    "+r"(previous_cumsum),  // %2
+    "+r"(width)  // %3
+  :
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
+
+#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst,
+                                    int count) {
+  asm volatile (
+    "movd      %5,%%xmm5                       \n"
+    "cvtdq2ps  %%xmm5,%%xmm5                   \n"
+    "rcpss     %%xmm5,%%xmm4                   \n"
+    "pshufd    $0x0,%%xmm4,%%xmm4              \n"
+    "sub       $0x4,%3                         \n"
+    "jl        49f                             \n"
+    "cmpl      $0x80,%5                        \n"
+    "ja        40f                             \n"
+
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrld     $0x10,%%xmm6                    \n"
+    "cvtdq2ps  %%xmm6,%%xmm6                   \n"
+    "addps     %%xmm6,%%xmm5                   \n"
+    "mulps     %%xmm4,%%xmm5                   \n"
+    "cvtps2dq  %%xmm5,%%xmm5                   \n"
+    "packssdw  %%xmm5,%%xmm5                   \n"
+
+  // 4 pixel small loop                        \n"
+    LABELALIGN
+  "4:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
+    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
+    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
+    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
+    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
+    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
+    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm3,%%xmm2                   \n"
+    "pmulhuw   %%xmm5,%%xmm0                   \n"
+    "pmulhuw   %%xmm5,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       4b                              \n"
+    "jmp       49f                             \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "movdqu    " MEMACCESS2(0x20,0) ",%%xmm2   \n"
+    "movdqu    " MEMACCESS2(0x30,0) ",%%xmm3   \n"
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    MEMOPREG(psubd,0x10,0,4,4,xmm1)            // psubd    0x10(%0,%4,4),%%xmm1
+    MEMOPREG(psubd,0x20,0,4,4,xmm2)            // psubd    0x20(%0,%4,4),%%xmm2
+    MEMOPREG(psubd,0x30,0,4,4,xmm3)            // psubd    0x30(%0,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    "psubd     " MEMACCESS2(0x10,1) ",%%xmm1   \n"
+    "psubd     " MEMACCESS2(0x20,1) ",%%xmm2   \n"
+    "psubd     " MEMACCESS2(0x30,1) ",%%xmm3   \n"
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    MEMOPREG(paddd,0x10,1,4,4,xmm1)            // paddd    0x10(%1,%4,4),%%xmm1
+    MEMOPREG(paddd,0x20,1,4,4,xmm2)            // paddd    0x20(%1,%4,4),%%xmm2
+    MEMOPREG(paddd,0x30,1,4,4,xmm3)            // paddd    0x30(%1,%4,4),%%xmm3
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm1,%%xmm1                   \n"
+    "mulps     %%xmm4,%%xmm0                   \n"
+    "mulps     %%xmm4,%%xmm1                   \n"
+    "cvtdq2ps  %%xmm2,%%xmm2                   \n"
+    "cvtdq2ps  %%xmm3,%%xmm3                   \n"
+    "mulps     %%xmm4,%%xmm2                   \n"
+    "mulps     %%xmm4,%%xmm3                   \n"
+    "cvtps2dq  %%xmm0,%%xmm0                   \n"
+    "cvtps2dq  %%xmm1,%%xmm1                   \n"
+    "cvtps2dq  %%xmm2,%%xmm2                   \n"
+    "cvtps2dq  %%xmm3,%%xmm3                   \n"
+    "packssdw  %%xmm1,%%xmm0                   \n"
+    "packssdw  %%xmm3,%%xmm2                   \n"
+    "packuswb  %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%3                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(psubd,0x00,0,4,4,xmm0)            // psubd    0x00(%0,%4,4),%%xmm0
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "psubd     " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(paddd,0x00,1,4,4,xmm0)            // paddd    0x00(%1,%4,4),%%xmm0
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "mulps     %%xmm4,%%xmm0                   \n"
+    "cvtps2dq  %%xmm0,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x4,2) ",%2            \n"
+    "sub       $0x1,%3                         \n"
+    "jge       10b                             \n"
+  "19:                                         \n"
+  : "+r"(topleft),  // %0
+    "+r"(botleft),  // %1
+    "+r"(dst),      // %2
+    "+rm"(count)    // %3
+  : "r"((intptr_t)(width)),  // %4
+    "rm"(area)     // %5
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+
+#ifdef HAS_ARGBAFFINEROW_SSE2
+// Copy ARGB pixels from source image with slope to a row of destination.
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* src_dudv, int width) {
+  intptr_t src_argb_stride_temp = src_argb_stride;
+  intptr_t temp = 0;
+  asm volatile (
+    "movq      " MEMACCESS(3) ",%%xmm2         \n"
+    "movq      " MEMACCESS2(0x08,3) ",%%xmm7   \n"
+    "shl       $0x10,%1                        \n"
+    "add       $0x4,%1                         \n"
+    "movd      %1,%%xmm5                       \n"
+    "sub       $0x4,%4                         \n"
+    "jl        49f                             \n"
+
+    "pshufd    $0x44,%%xmm7,%%xmm7             \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm7,%%xmm0                   \n"
+    "movlhps   %%xmm0,%%xmm2                   \n"
+    "movdqa    %%xmm7,%%xmm4                   \n"
+    "addps     %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "addps     %%xmm4,%%xmm3                   \n"
+    "addps     %%xmm4,%%xmm4                   \n"
+
+  // 4 pixel loop                              \n"
+    LABELALIGN
+  "40:                                         \n"
+    "cvttps2dq %%xmm2,%%xmm0                   \n"  // x, y float to int first 2
+    "cvttps2dq %%xmm3,%%xmm1                   \n"  // x, y float to int next 2
+    "packssdw  %%xmm1,%%xmm0                   \n"  // x, y as 8 shorts
+    "pmaddwd   %%xmm5,%%xmm0                   \n"  // off = x * 4 + y * stride
+    "movd      %%xmm0,%k1                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    "movd      %%xmm0,%k5                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
+    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
+    "punpckldq %%xmm6,%%xmm1                   \n"
+    "addps     %%xmm4,%%xmm2                   \n"
+    "movq      %%xmm1," MEMACCESS(2) "         \n"
+    "movd      %%xmm0,%k1                      \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+    "movd      %%xmm0,%k5                      \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
+    MEMOPREG(movd,0x00,0,5,1,xmm6)             //  movd      (%0,%5,1),%%xmm6
+    "punpckldq %%xmm6,%%xmm0                   \n"
+    "addps     %%xmm4,%%xmm3                   \n"
+    "movq      %%xmm0," MEMACCESS2(0x08,2) "   \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%4                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "add       $0x3,%4                         \n"
+    "jl        19f                             \n"
+
+  // 1 pixel loop                              \n"
+    LABELALIGN
+  "10:                                         \n"
+    "cvttps2dq %%xmm2,%%xmm0                   \n"
+    "packssdw  %%xmm0,%%xmm0                   \n"
+    "pmaddwd   %%xmm5,%%xmm0                   \n"
+    "addps     %%xmm7,%%xmm2                   \n"
+    "movd      %%xmm0,%k1                      \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm0)             //  movd      (%0,%1,1),%%xmm0
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x04,2) ",%2           \n"
+    "sub       $0x1,%4                         \n"
+    "jge       10b                             \n"
+  "19:                                         \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_argb_stride_temp),  // %1
+    "+r"(dst_argb),  // %2
+    "+r"(src_dudv),  // %3
+    "+rm"(width),    // %4
+    "+r"(temp)   // %5
+  :
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBAFFINEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_SSSE3
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm2,%%xmm0                   \n"
+    "punpckhbw %%xmm2,%%xmm1                   \n"
+    "pmaddubsw %%xmm5,%%xmm0                   \n"
+    "pmaddubsw %%xmm5,%%xmm1                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "psrlw     $0x7,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm5"
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSSE3
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+// Bilinear filter 32x2 -> 32x1
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  asm volatile (
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "sub       %1,%0                           \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "vmovd      %3,%%xmm0                      \n"
+    "neg        %3                             \n"
+    "add        $0x80,%3                       \n"
+    "vmovd      %3,%%xmm5                      \n"
+    "vpunpcklbw %%xmm0,%%xmm5,%%xmm5           \n"
+    "vpunpcklwd %%xmm5,%%xmm5,%%xmm5           \n"
+    "vpxor      %%ymm0,%%ymm0,%%ymm0           \n"
+    "vpermd     %%ymm5,%%ymm0,%%ymm5           \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
+    MEMOPREG(vmovdqu,0x00,1,4,1,ymm2)
+    "vpunpckhbw %%ymm2,%%ymm0,%%ymm1           \n"
+    "vpunpcklbw %%ymm2,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm5,%%ymm0,%%ymm0           \n"
+    "vpmaddubsw %%ymm5,%%ymm1,%%ymm1           \n"
+    "vpsrlw     $0x7,%%ymm0,%%ymm0             \n"
+    "vpsrlw     $0x7,%%ymm1,%%ymm1             \n"
+    "vpackuswb  %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
+    MEMOPREG(vmovdqu,0x00,1,4,1,ymm1)
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm0        \n"
+    VMEMOPREG(vpavgb,0x00,1,4,1,ymm0,ymm0)     // vpavgb (%1,%4,1),%%ymm0,%%ymm0
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "vmovdqu    " MEMACCESS(1) ",%%ymm1        \n"
+    MEMOPREG(vmovdqu,0x00,1,4,1,ymm0)
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    "vpavgb     %%ymm1,%%ymm0,%%ymm0           \n"
+    MEMOPMEM(vmovdqu,ymm0,0x00,1,0,1)
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x20,%2                        \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "rep movsb " MEMMOVESTRING(1,0) "          \n"
+    "jmp       999f                            \n"
+
+  "99:                                         \n"
+    "vzeroupper                                \n"
+  "999:                                        \n"
+  : "+D"(dst_ptr),    // %0
+    "+S"(src_ptr),    // %1
+    "+c"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm5"
+  );
+}
+#endif  // HAS_INTERPOLATEROW_AVX2
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  asm volatile (
+    "sub       %1,%0                           \n"
+    "shr       %3                              \n"
+    "cmp       $0x0,%3                         \n"
+    "je        100f                            \n"
+    "cmp       $0x20,%3                        \n"
+    "je        75f                             \n"
+    "cmp       $0x40,%3                        \n"
+    "je        50f                             \n"
+    "cmp       $0x60,%3                        \n"
+    "je        25f                             \n"
+
+    "movd      %3,%%xmm0                       \n"
+    "neg       %3                              \n"
+    "add       $0x80,%3                        \n"
+    "movd      %3,%%xmm5                       \n"
+    "punpcklbw %%xmm0,%%xmm5                   \n"
+    "punpcklwd %%xmm5,%%xmm5                   \n"
+    "pshufd    $0x0,%%xmm5,%%xmm5              \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+
+    // General purpose row blend.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm2)           //  movdqu    (%1,%4,1),%%xmm2
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm0                   \n"
+    "punpckhbw %%xmm4,%%xmm1                   \n"
+    "psubw     %%xmm0,%%xmm2                   \n"
+    "psubw     %%xmm1,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm2                   \n"
+    "paddw     %%xmm3,%%xmm3                   \n"
+    "pmulhw    %%xmm5,%%xmm2                   \n"
+    "pmulhw    %%xmm5,%%xmm3                   \n"
+    "paddw     %%xmm2,%%xmm0                   \n"
+    "paddw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    // Blend 25 / 75.
+    LABELALIGN
+  "25:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        25b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 50 / 50.
+    LABELALIGN
+  "50:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm1)           //  movdqu    (%1,%4,1),%%xmm1
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        50b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 75 / 25.
+    LABELALIGN
+  "75:                                         \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm1         \n"
+    MEMOPREG(movdqu,0x00,1,4,1,xmm0)           //  movdqu    (%1,%4,1),%%xmm0
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        75b                             \n"
+    "jmp       99f                             \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+    LABELALIGN
+  "100:                                        \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    MEMOPMEM(movdqu,xmm0,0x00,1,0,1)           //  movdqu    %%xmm0,(%1,%0,1)
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        100b                            \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),    // %0
+    "+r"(src_ptr),    // %1
+    "+r"(dst_width),  // %2
+    "+r"(source_y_fraction)  // %3
+  : "r"((intptr_t)(src_stride))  // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+#ifdef HAS_ARGBSHUFFLEROW_SSSE3
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix) {
+  asm volatile (
+    "movdqu    " MEMACCESS(3) ",%%xmm5         \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm5,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_SSSE3
+
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    "vbroadcastf128 " MEMACCESS(3) ",%%ymm5    \n"
+    LABELALIGN
+  "1:                                          \n"
+    "vmovdqu   " MEMACCESS(0) ",%%ymm0         \n"
+    "vmovdqu   " MEMACCESS2(0x20,0) ",%%ymm1   \n"
+    "lea       " MEMLEA(0x40,0) ",%0           \n"
+    "vpshufb   %%ymm5,%%ymm0,%%ymm0            \n"
+    "vpshufb   %%ymm5,%%ymm1,%%ymm1            \n"
+    "vmovdqu   %%ymm0," MEMACCESS(1) "         \n"
+    "vmovdqu   %%ymm1," MEMACCESS2(0x20,1) "   \n"
+    "lea       " MEMLEA(0x40,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_AVX2
+
+#ifdef HAS_ARGBSHUFFLEROW_SSE2
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    "pxor      %%xmm5,%%xmm5                   \n"
+    "mov       " MEMACCESS(4) ",%k2            \n"
+    "cmp       $0x3000102,%k2                  \n"
+    "je        3012f                           \n"
+    "cmp       $0x10203,%k2                    \n"
+    "je        123f                            \n"
+    "cmp       $0x30201,%k2                    \n"
+    "je        321f                            \n"
+    "cmp       $0x2010003,%k2                  \n"
+    "je        2103f                           \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(4) ",%2             \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS(1) "            \n"
+    "movzb     " MEMACCESS2(0x1,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x1,1) "       \n"
+    "movzb     " MEMACCESS2(0x2,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x2,1) "       \n"
+    "movzb     " MEMACCESS2(0x3,4) ",%2        \n"
+    MEMOPARG(movzb,0x00,0,2,1,2) "             \n"  //  movzb     (%0,%2,1),%2
+    "mov       %b2," MEMACCESS2(0x3,1) "       \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    "lea       " MEMLEA(0x4,1) ",%1            \n"
+    "sub       $0x1,%3                         \n"
+    "jg        1b                              \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "123:                                        \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x1b,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x1b,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x1b,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        123b                            \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "321:                                        \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x39,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x39,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x39,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x39,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        321b                            \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "2103:                                       \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0x93,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0x93,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0x93,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0x93,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        2103b                           \n"
+    "jmp       99f                             \n"
+
+    LABELALIGN
+  "3012:                                       \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpckhbw %%xmm5,%%xmm1                   \n"
+    "pshufhw   $0xc6,%%xmm0,%%xmm0             \n"
+    "pshuflw   $0xc6,%%xmm0,%%xmm0             \n"
+    "pshufhw   $0xc6,%%xmm1,%%xmm1             \n"
+    "pshuflw   $0xc6,%%xmm1,%%xmm1             \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        3012b                           \n"
+
+  "99:                                         \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+d"(pixel_temp),  // %2
+    "+r"(pix)         // %3
+  : "r"(shuffler)      // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm5"
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_SSE2
+
+#ifdef HAS_I422TOYUY2ROW_SSE2
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+ asm volatile (
+    "sub       %1,%2                             \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movq      " MEMACCESS(1) ",%%xmm2           \n"
+    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
+    "lea       " MEMLEA(0x8,1) ",%1              \n"
+    "punpcklbw %%xmm3,%%xmm2                     \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "movdqa    %%xmm0,%%xmm1                     \n"
+    "punpcklbw %%xmm2,%%xmm0                     \n"
+    "punpckhbw %%xmm2,%%xmm1                     \n"
+    "movdqu    %%xmm0," MEMACCESS(3) "           \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,3) "     \n"
+    "lea       " MEMLEA(0x20,3) ",%3             \n"
+    "sub       $0x10,%4                          \n"
+    "jg         1b                               \n"
+    : "+r"(src_y),  // %0
+      "+r"(src_u),  // %1
+      "+r"(src_v),  // %2
+      "+r"(dst_frame),  // %3
+      "+rm"(width)  // %4
+    :
+    : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
+  );
+}
+#endif  // HAS_I422TOYUY2ROW_SSE2
+
+#ifdef HAS_I422TOUYVYROW_SSE2
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+ asm volatile (
+    "sub        %1,%2                            \n"
+    LABELALIGN
+  "1:                                            \n"
+    "movq      " MEMACCESS(1) ",%%xmm2           \n"
+    MEMOPREG(movq,0x00,1,2,1,xmm3)               //  movq    (%1,%2,1),%%xmm3
+    "lea       " MEMLEA(0x8,1) ",%1              \n"
+    "punpcklbw %%xmm3,%%xmm2                     \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0           \n"
+    "movdqa    %%xmm2,%%xmm1                     \n"
+    "lea       " MEMLEA(0x10,0) ",%0             \n"
+    "punpcklbw %%xmm0,%%xmm1                     \n"
+    "punpckhbw %%xmm0,%%xmm2                     \n"
+    "movdqu    %%xmm1," MEMACCESS(3) "           \n"
+    "movdqu    %%xmm2," MEMACCESS2(0x10,3) "     \n"
+    "lea       " MEMLEA(0x20,3) ",%3             \n"
+    "sub       $0x10,%4                          \n"
+    "jg         1b                               \n"
+    : "+r"(src_y),  // %0
+      "+r"(src_u),  // %1
+      "+r"(src_v),  // %2
+      "+r"(dst_frame),  // %3
+      "+rm"(width)  // %4
+    :
+    : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
+  );
+}
+#endif  // HAS_I422TOUYVYROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "pxor      %%xmm3,%%xmm3                   \n"
+
+    // 2 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "punpcklbw %%xmm3,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm4                   \n"
+    "punpcklwd %%xmm3,%%xmm0                   \n"
+    "punpckhwd %%xmm3,%%xmm4                   \n"
+    "cvtdq2ps  %%xmm0,%%xmm0                   \n"
+    "cvtdq2ps  %%xmm4,%%xmm4                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "movdqa    %%xmm4,%%xmm5                   \n"
+    "mulps     " MEMACCESS2(0x10,3) ",%%xmm0   \n"
+    "mulps     " MEMACCESS2(0x10,3) ",%%xmm4   \n"
+    "addps     " MEMACCESS(3) ",%%xmm0         \n"
+    "addps     " MEMACCESS(3) ",%%xmm4         \n"
+    "movdqa    %%xmm1,%%xmm2                   \n"
+    "movdqa    %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm1,%%xmm2                   \n"
+    "mulps     %%xmm5,%%xmm6                   \n"
+    "mulps     %%xmm2,%%xmm1                   \n"
+    "mulps     %%xmm6,%%xmm5                   \n"
+    "mulps     " MEMACCESS2(0x20,3) ",%%xmm2   \n"
+    "mulps     " MEMACCESS2(0x20,3) ",%%xmm6   \n"
+    "mulps     " MEMACCESS2(0x30,3) ",%%xmm1   \n"
+    "mulps     " MEMACCESS2(0x30,3) ",%%xmm5   \n"
+    "addps     %%xmm2,%%xmm0                   \n"
+    "addps     %%xmm6,%%xmm4                   \n"
+    "addps     %%xmm1,%%xmm0                   \n"
+    "addps     %%xmm5,%%xmm4                   \n"
+    "cvttps2dq %%xmm0,%%xmm0                   \n"
+    "cvttps2dq %%xmm4,%%xmm4                   \n"
+    "packuswb  %%xmm4,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x2,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc"
+    , "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  asm volatile (
+    "vbroadcastf128 " MEMACCESS(3) ",%%ymm4     \n"
+    "vbroadcastf128 " MEMACCESS2(0x10,3) ",%%ymm5 \n"
+    "vbroadcastf128 " MEMACCESS2(0x20,3) ",%%ymm6 \n"
+    "vbroadcastf128 " MEMACCESS2(0x30,3) ",%%ymm7 \n"
+
+    // 2 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "vpmovzxbd   " MEMACCESS(0) ",%%ymm0       \n"  // 2 ARGB pixels
+    "lea         " MEMLEA(0x8,0) ",%0          \n"
+    "vcvtdq2ps   %%ymm0,%%ymm0                 \n"  // X 8 floats
+    "vmulps      %%ymm0,%%ymm0,%%ymm2          \n"  // X * X
+    "vmulps      %%ymm7,%%ymm0,%%ymm3          \n"  // C3 * X
+    "vfmadd132ps %%ymm5,%%ymm4,%%ymm0          \n"  // result = C0 + C1 * X
+    "vfmadd231ps %%ymm6,%%ymm2,%%ymm0          \n"  // result += C2 * X * X
+    "vfmadd231ps %%ymm3,%%ymm2,%%ymm0          \n"  // result += C3 * X * X * X
+    "vcvttps2dq  %%ymm0,%%ymm0                 \n"
+    "vpackusdw   %%ymm0,%%ymm0,%%ymm0          \n"
+    "vpermq      $0xd8,%%ymm0,%%ymm0           \n"
+    "vpackuswb   %%xmm0,%%xmm0,%%xmm0          \n"
+    "vmovq       %%xmm0," MEMACCESS(1) "       \n"
+    "lea         " MEMLEA(0x8,1) ",%1          \n"
+    "sub         $0x2,%2                       \n"
+    "jg          1b                            \n"
+    "vzeroupper                                \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  : "r"(poly)        // %3
+  : "memory", "cc",
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
+#ifdef HAS_ARGBCOLORTABLEROW_X86
+// Tranform ARGB pixels with color table.
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+                           int width) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    // 1 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(0) ",%1             \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
+    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
+    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
+    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
+    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
+    "movzb     " MEMACCESS2(-0x1,0) ",%1       \n"
+    MEMOPARG(movzb,0x03,3,1,4,1) "             \n"  // movzb 0x3(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x1,0) "      \n"
+    "dec       %2                              \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),   // %0
+    "+d"(pixel_temp), // %1
+    "+r"(width)       // %2
+  : "r"(table_argb)   // %3
+  : "memory", "cc");
+}
+#endif  // HAS_ARGBCOLORTABLEROW_X86
+
+#ifdef HAS_RGBCOLORTABLEROW_X86
+// Tranform RGB pixels with color table.
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+  uintptr_t pixel_temp = 0u;
+  asm volatile (
+    // 1 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movzb     " MEMACCESS(0) ",%1             \n"
+    "lea       " MEMLEA(0x4,0) ",%0            \n"
+    MEMOPARG(movzb,0x00,3,1,4,1) "             \n"  // movzb (%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x4,0) "      \n"
+    "movzb     " MEMACCESS2(-0x3,0) ",%1       \n"
+    MEMOPARG(movzb,0x01,3,1,4,1) "             \n"  // movzb 0x1(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x3,0) "      \n"
+    "movzb     " MEMACCESS2(-0x2,0) ",%1       \n"
+    MEMOPARG(movzb,0x02,3,1,4,1) "             \n"  // movzb 0x2(%3,%1,4),%1
+    "mov       %b1," MEMACCESS2(-0x2,0) "      \n"
+    "dec       %2                              \n"
+    "jg        1b                              \n"
+  : "+r"(dst_argb),   // %0
+    "+d"(pixel_temp), // %1
+    "+r"(width)       // %2
+  : "r"(table_argb)   // %3
+  : "memory", "cc");
+}
+#endif  // HAS_RGBCOLORTABLEROW_X86
+
+#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// Tranform RGB pixels with luma table.
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff) {
+  uintptr_t pixel_temp = 0u;
+  uintptr_t table_temp = 0u;
+  asm volatile (
+    "movd      %6,%%xmm3                       \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pcmpeqb   %%xmm4,%%xmm4                   \n"
+    "psllw     $0x8,%%xmm4                     \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+
+    // 4 pixel loop.
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(2) ",%%xmm0         \n"
+    "pmaddubsw %%xmm3,%%xmm0                   \n"
+    "phaddw    %%xmm0,%%xmm0                   \n"
+    "pand      %%xmm4,%%xmm0                   \n"
+    "punpcklwd %%xmm5,%%xmm0                   \n"
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS(2) ",%0             \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS(3) "            \n"
+    "movzb     " MEMACCESS2(0x1,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x1,3) "       \n"
+    "movzb     " MEMACCESS2(0x2,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x2,3) "       \n"
+    "movzb     " MEMACCESS2(0x3,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0x3,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS2(0x4,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x4,3) "       \n"
+    "movzb     " MEMACCESS2(0x5,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x5,3) "       \n"
+    "movzb     " MEMACCESS2(0x6,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x6,3) "       \n"
+    "movzb     " MEMACCESS2(0x7,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0x7,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+    "pshufd    $0x39,%%xmm0,%%xmm0             \n"
+
+    "movzb     " MEMACCESS2(0x8,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x8,3) "       \n"
+    "movzb     " MEMACCESS2(0x9,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0x9,3) "       \n"
+    "movzb     " MEMACCESS2(0xa,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xa,3) "       \n"
+    "movzb     " MEMACCESS2(0xb,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0xb,3) "       \n"
+
+    "movd      %%xmm0,%k1                      \n"  // 32 bit offset
+    "add       %5,%1                           \n"
+
+    "movzb     " MEMACCESS2(0xc,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xc,3) "       \n"
+    "movzb     " MEMACCESS2(0xd,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xd,3) "       \n"
+    "movzb     " MEMACCESS2(0xe,2) ",%0        \n"
+    MEMOPARG(movzb,0x00,1,0,1,0) "             \n"  // movzb     (%1,%0,1),%0
+    "mov       %b0," MEMACCESS2(0xe,3) "       \n"
+    "movzb     " MEMACCESS2(0xf,2) ",%0        \n"
+    "mov       %b0," MEMACCESS2(0xf,3) "       \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "lea       " MEMLEA(0x10,3) ",%3           \n"
+    "sub       $0x4,%4                         \n"
+    "jg        1b                              \n"
+  : "+d"(pixel_temp),  // %0
+    "+a"(table_temp),  // %1
+    "+r"(src_argb),    // %2
+    "+r"(dst_argb),    // %3
+    "+rm"(width)       // %4
+  : "r"(luma),         // %5
+    "rm"(lumacoeff)    // %6
+  : "memory", "cc", "xmm0", "xmm3", "xmm4", "xmm5"
+  );
+}
+#endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/row_mips.cc b/third_party/aom/third_party/libyuv/source/row_mips.cc
new file mode 100644
index 0000000000..cfc9ffe036
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/row_mips.cc
@@ -0,0 +1,911 @@
+/*
+ *  Copyright (c) 2012 The LibYuv project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// The following are available on Mips platforms:
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+#ifdef HAS_COPYROW_MIPS
+void CopyRow_MIPS(const uint8* src, uint8* dst, int count) {
+  __asm__ __volatile__ (
+    ".set      noreorder                         \n"
+    ".set      noat                              \n"
+    "slti      $at, %[count], 8                  \n"
+    "bne       $at ,$zero, $last8                \n"
+    "xor       $t8, %[src], %[dst]               \n"
+    "andi      $t8, $t8, 0x3                     \n"
+
+    "bne       $t8, $zero, unaligned             \n"
+    "negu      $a3, %[dst]                       \n"
+    // make dst/src aligned
+    "andi      $a3, $a3, 0x3                     \n"
+    "beq       $a3, $zero, $chk16w               \n"
+    // word-aligned now count is the remining bytes count
+    "subu     %[count], %[count], $a3            \n"
+
+    "lwr       $t8, 0(%[src])                    \n"
+    "addu      %[src], %[src], $a3               \n"
+    "swr       $t8, 0(%[dst])                    \n"
+    "addu      %[dst], %[dst], $a3               \n"
+
+    // Now the dst/src are mutually word-aligned with word-aligned addresses
+    "$chk16w:                                    \n"
+    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
+    // t8 is the byte count after 64-byte chunks
+    "beq       %[count], $t8, chk8w              \n"
+    // There will be at most 1 32-byte chunk after it
+    "subu      $a3, %[count], $t8                \n"  // the reminder
+    // Here a3 counts bytes in 16w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // Now a3 is the final dst after 64-byte chunks
+    "addu      $t0, %[dst], %[count]             \n"
+    // t0 is the "past the end" address
+
+    // When in the loop we exercise "pref 30,x(a1)", the a1+x should not be past
+    // the "t0-32" address
+    // This means: for x=128 the last "safe" a1 address is "t0-160"
+    // Alternatively, for x=64 the last "safe" a1 address is "t0-96"
+    // we will use "pref 30,128(a1)", so "t0-160" is the limit
+    "subu      $t9, $t0, 160                     \n"
+    // t9 is the "last safe pref 30,128(a1)" address
+    "pref      0, 0(%[src])                      \n"  // first line of src
+    "pref      0, 32(%[src])                     \n"  // second line of src
+    "pref      0, 64(%[src])                     \n"
+    "pref      30, 32(%[dst])                    \n"
+    // In case the a1 > t9 don't use "pref 30" at all
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bgtz      $v1, $loop16w                     \n"
+    "nop                                         \n"
+    // otherwise, start with using pref30
+    "pref      30, 64(%[dst])                    \n"
+    "$loop16w:                                    \n"
+    "pref      0, 96(%[src])                     \n"
+    "lw        $t0, 0(%[src])                    \n"
+    "bgtz      $v1, $skip_pref30_96              \n"  // skip
+    "lw        $t1, 4(%[src])                    \n"
+    "pref      30, 96(%[dst])                    \n"  // continue
+    "$skip_pref30_96:                            \n"
+    "lw        $t2, 8(%[src])                    \n"
+    "lw        $t3, 12(%[src])                   \n"
+    "lw        $t4, 16(%[src])                   \n"
+    "lw        $t5, 20(%[src])                   \n"
+    "lw        $t6, 24(%[src])                   \n"
+    "lw        $t7, 28(%[src])                   \n"
+    "pref      0, 128(%[src])                    \n"
+    //  bring the next lines of src, addr 128
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "lw        $t0, 32(%[src])                   \n"
+    "bgtz      $v1, $skip_pref30_128             \n"  // skip pref 30,128(a1)
+    "lw        $t1, 36(%[src])                   \n"
+    "pref      30, 128(%[dst])                   \n"  // set dest, addr 128
+    "$skip_pref30_128:                           \n"
+    "lw        $t2, 40(%[src])                   \n"
+    "lw        $t3, 44(%[src])                   \n"
+    "lw        $t4, 48(%[src])                   \n"
+    "lw        $t5, 52(%[src])                   \n"
+    "lw        $t6, 56(%[src])                   \n"
+    "lw        $t7, 60(%[src])                   \n"
+    "pref      0, 160(%[src])                    \n"
+    // bring the next lines of src, addr 160
+    "sw        $t0, 32(%[dst])                   \n"
+    "sw        $t1, 36(%[dst])                   \n"
+    "sw        $t2, 40(%[dst])                   \n"
+    "sw        $t3, 44(%[dst])                   \n"
+    "sw        $t4, 48(%[dst])                   \n"
+    "sw        $t5, 52(%[dst])                   \n"
+    "sw        $t6, 56(%[dst])                   \n"
+    "sw        $t7, 60(%[dst])                   \n"
+
+    "addiu     %[dst], %[dst], 64                \n"  // adding 64 to dest
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bne       %[dst], $a3, $loop16w             \n"
+    " addiu    %[src], %[src], 64                \n"  // adding 64 to src
+    "move      %[count], $t8                     \n"
+
+    // Here we have src and dest word-aligned but less than 64-bytes to go
+
+    "chk8w:                                      \n"
+    "pref      0, 0x0(%[src])                    \n"
+    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
+    // the t8 is the reminder count past 32-bytes
+    "beq       %[count], $t8, chk1w              \n"
+    // count=t8,no 32-byte chunk
+    " nop                                        \n"
+
+    "lw        $t0, 0(%[src])                    \n"
+    "lw        $t1, 4(%[src])                    \n"
+    "lw        $t2, 8(%[src])                    \n"
+    "lw        $t3, 12(%[src])                   \n"
+    "lw        $t4, 16(%[src])                   \n"
+    "lw        $t5, 20(%[src])                   \n"
+    "lw        $t6, 24(%[src])                   \n"
+    "lw        $t7, 28(%[src])                   \n"
+    "addiu     %[src], %[src], 32                \n"
+
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "addiu     %[dst], %[dst], 32                \n"
+
+    "chk1w:                                      \n"
+    "andi      %[count], $t8, 0x3                \n"
+    // now count is the reminder past 1w chunks
+    "beq       %[count], $t8, $last8             \n"
+    " subu     $a3, $t8, %[count]                \n"
+    // a3 is count of bytes in 1w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // now a3 is the dst address past the 1w chunks
+    // copying in words (4-byte chunks)
+    "$wordCopy_loop:                             \n"
+    "lw        $t3, 0(%[src])                    \n"
+    // the first t3 may be equal t0 ... optimize?
+    "addiu     %[src], %[src],4                  \n"
+    "addiu     %[dst], %[dst],4                  \n"
+    "bne       %[dst], $a3,$wordCopy_loop        \n"
+    " sw       $t3, -4(%[dst])                   \n"
+
+    // For the last (<8) bytes
+    "$last8:                                     \n"
+    "blez      %[count], leave                   \n"
+    " addu     $a3, %[dst], %[count]             \n"  // a3 -last dst address
+    "$last8loop:                                 \n"
+    "lb        $v1, 0(%[src])                    \n"
+    "addiu     %[src], %[src], 1                 \n"
+    "addiu     %[dst], %[dst], 1                 \n"
+    "bne       %[dst], $a3, $last8loop           \n"
+    " sb       $v1, -1(%[dst])                   \n"
+
+    "leave:                                      \n"
+    "  j       $ra                               \n"
+    "  nop                                       \n"
+
+    //
+    // UNALIGNED case
+    //
+
+    "unaligned:                                  \n"
+    // got here with a3="negu a1"
+    "andi      $a3, $a3, 0x3                     \n"  // a1 is word aligned?
+    "beqz      $a3, $ua_chk16w                   \n"
+    " subu     %[count], %[count], $a3           \n"
+    // bytes left after initial a3 bytes
+    "lwr       $v1, 0(%[src])                    \n"
+    "lwl       $v1, 3(%[src])                    \n"
+    "addu      %[src], %[src], $a3               \n"  // a3 may be 1, 2 or 3
+    "swr       $v1, 0(%[dst])                    \n"
+    "addu      %[dst], %[dst], $a3               \n"
+    // below the dst will be word aligned (NOTE1)
+    "$ua_chk16w:                                 \n"
+    "andi      $t8, %[count], 0x3f               \n"  // whole 64-B chunks?
+    // t8 is the byte count after 64-byte chunks
+    "beq       %[count], $t8, ua_chk8w           \n"
+    // if a2==t8, no 64-byte chunks
+    // There will be at most 1 32-byte chunk after it
+    "subu      $a3, %[count], $t8                \n"  // the reminder
+    // Here a3 counts bytes in 16w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // Now a3 is the final dst after 64-byte chunks
+    "addu      $t0, %[dst], %[count]             \n"  // t0 "past the end"
+    "subu      $t9, $t0, 160                     \n"
+    // t9 is the "last safe pref 30,128(a1)" address
+    "pref      0, 0(%[src])                      \n"  // first line of src
+    "pref      0, 32(%[src])                     \n"  // second line  addr 32
+    "pref      0, 64(%[src])                     \n"
+    "pref      30, 32(%[dst])                    \n"
+    // safe, as we have at least 64 bytes ahead
+    // In case the a1 > t9 don't use "pref 30" at all
+    "sgtu      $v1, %[dst], $t9                  \n"
+    "bgtz      $v1, $ua_loop16w                  \n"
+    // skip "pref 30,64(a1)" for too short arrays
+    " nop                                        \n"
+    // otherwise, start with using pref30
+    "pref      30, 64(%[dst])                    \n"
+    "$ua_loop16w:                                \n"
+    "pref      0, 96(%[src])                     \n"
+    "lwr       $t0, 0(%[src])                    \n"
+    "lwl       $t0, 3(%[src])                    \n"
+    "lwr       $t1, 4(%[src])                    \n"
+    "bgtz      $v1, $ua_skip_pref30_96           \n"
+    " lwl      $t1, 7(%[src])                    \n"
+    "pref      30, 96(%[dst])                    \n"
+    // continue setting up the dest, addr 96
+    "$ua_skip_pref30_96:                         \n"
+    "lwr       $t2, 8(%[src])                    \n"
+    "lwl       $t2, 11(%[src])                   \n"
+    "lwr       $t3, 12(%[src])                   \n"
+    "lwl       $t3, 15(%[src])                   \n"
+    "lwr       $t4, 16(%[src])                   \n"
+    "lwl       $t4, 19(%[src])                   \n"
+    "lwr       $t5, 20(%[src])                   \n"
+    "lwl       $t5, 23(%[src])                   \n"
+    "lwr       $t6, 24(%[src])                   \n"
+    "lwl       $t6, 27(%[src])                   \n"
+    "lwr       $t7, 28(%[src])                   \n"
+    "lwl       $t7, 31(%[src])                   \n"
+    "pref      0, 128(%[src])                    \n"
+    // bring the next lines of src, addr 128
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "lwr       $t0, 32(%[src])                   \n"
+    "lwl       $t0, 35(%[src])                   \n"
+    "lwr       $t1, 36(%[src])                   \n"
+    "bgtz      $v1, ua_skip_pref30_128           \n"
+    " lwl      $t1, 39(%[src])                   \n"
+    "pref      30, 128(%[dst])                   \n"
+    // continue setting up the dest, addr 128
+    "ua_skip_pref30_128:                         \n"
+
+    "lwr       $t2, 40(%[src])                   \n"
+    "lwl       $t2, 43(%[src])                   \n"
+    "lwr       $t3, 44(%[src])                   \n"
+    "lwl       $t3, 47(%[src])                   \n"
+    "lwr       $t4, 48(%[src])                   \n"
+    "lwl       $t4, 51(%[src])                   \n"
+    "lwr       $t5, 52(%[src])                   \n"
+    "lwl       $t5, 55(%[src])                   \n"
+    "lwr       $t6, 56(%[src])                   \n"
+    "lwl       $t6, 59(%[src])                   \n"
+    "lwr       $t7, 60(%[src])                   \n"
+    "lwl       $t7, 63(%[src])                   \n"
+    "pref      0, 160(%[src])                    \n"
+    // bring the next lines of src, addr 160
+    "sw        $t0, 32(%[dst])                   \n"
+    "sw        $t1, 36(%[dst])                   \n"
+    "sw        $t2, 40(%[dst])                   \n"
+    "sw        $t3, 44(%[dst])                   \n"
+    "sw        $t4, 48(%[dst])                   \n"
+    "sw        $t5, 52(%[dst])                   \n"
+    "sw        $t6, 56(%[dst])                   \n"
+    "sw        $t7, 60(%[dst])                   \n"
+
+    "addiu     %[dst],%[dst],64                  \n"  // adding 64 to dest
+    "sgtu      $v1,%[dst],$t9                    \n"
+    "bne       %[dst],$a3,$ua_loop16w            \n"
+    " addiu    %[src],%[src],64                  \n"  // adding 64 to src
+    "move      %[count],$t8                      \n"
+
+    // Here we have src and dest word-aligned but less than 64-bytes to go
+
+    "ua_chk8w:                                   \n"
+    "pref      0, 0x0(%[src])                    \n"
+    "andi      $t8, %[count], 0x1f               \n"  // 32-byte chunk?
+    // the t8 is the reminder count
+    "beq       %[count], $t8, $ua_chk1w          \n"
+    // when count==t8, no 32-byte chunk
+
+    "lwr       $t0, 0(%[src])                    \n"
+    "lwl       $t0, 3(%[src])                    \n"
+    "lwr       $t1, 4(%[src])                    \n"
+    "lwl       $t1, 7(%[src])                    \n"
+    "lwr       $t2, 8(%[src])                    \n"
+    "lwl       $t2, 11(%[src])                   \n"
+    "lwr       $t3, 12(%[src])                   \n"
+    "lwl       $t3, 15(%[src])                   \n"
+    "lwr       $t4, 16(%[src])                   \n"
+    "lwl       $t4, 19(%[src])                   \n"
+    "lwr       $t5, 20(%[src])                   \n"
+    "lwl       $t5, 23(%[src])                   \n"
+    "lwr       $t6, 24(%[src])                   \n"
+    "lwl       $t6, 27(%[src])                   \n"
+    "lwr       $t7, 28(%[src])                   \n"
+    "lwl       $t7, 31(%[src])                   \n"
+    "addiu     %[src], %[src], 32                \n"
+
+    "sw        $t0, 0(%[dst])                    \n"
+    "sw        $t1, 4(%[dst])                    \n"
+    "sw        $t2, 8(%[dst])                    \n"
+    "sw        $t3, 12(%[dst])                   \n"
+    "sw        $t4, 16(%[dst])                   \n"
+    "sw        $t5, 20(%[dst])                   \n"
+    "sw        $t6, 24(%[dst])                   \n"
+    "sw        $t7, 28(%[dst])                   \n"
+    "addiu     %[dst], %[dst], 32                \n"
+
+    "$ua_chk1w:                                  \n"
+    "andi      %[count], $t8, 0x3                \n"
+    // now count is the reminder past 1w chunks
+    "beq       %[count], $t8, ua_smallCopy       \n"
+    "subu      $a3, $t8, %[count]                \n"
+    // a3 is count of bytes in 1w chunks
+    "addu      $a3, %[dst], $a3                  \n"
+    // now a3 is the dst address past the 1w chunks
+
+    // copying in words (4-byte chunks)
+    "$ua_wordCopy_loop:                          \n"
+    "lwr       $v1, 0(%[src])                    \n"
+    "lwl       $v1, 3(%[src])                    \n"
+    "addiu     %[src], %[src], 4                 \n"
+    "addiu     %[dst], %[dst], 4                 \n"
+    // note: dst=a1 is word aligned here, see NOTE1
+    "bne       %[dst], $a3, $ua_wordCopy_loop    \n"
+    " sw       $v1,-4(%[dst])                    \n"
+
+    // Now less than 4 bytes (value in count) left to copy
+    "ua_smallCopy:                               \n"
+    "beqz      %[count], leave                   \n"
+    " addu     $a3, %[dst], %[count]             \n" // a3 = last dst address
+    "$ua_smallCopy_loop:                         \n"
+    "lb        $v1, 0(%[src])                    \n"
+    "addiu     %[src], %[src], 1                 \n"
+    "addiu     %[dst], %[dst], 1                 \n"
+    "bne       %[dst],$a3,$ua_smallCopy_loop     \n"
+    " sb       $v1, -1(%[dst])                   \n"
+
+    "j         $ra                               \n"
+    " nop                                        \n"
+    ".set      at                                \n"
+    ".set      reorder                           \n"
+       : [dst] "+r" (dst), [src] "+r" (src)
+       : [count] "r" (count)
+       : "t0", "t1", "t2", "t3", "t4", "t5", "t6", "t7",
+       "t8", "t9", "a3", "v1", "at"
+  );
+}
+#endif  // HAS_COPYROW_MIPS
+
+// MIPS DSPR2 functions
+#if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips_dsp) && \
+    (__mips_dsp_rev >= 2) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32) && (__mips_isa_rev < 6)
+
+void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                           int width) {
+  __asm__ __volatile__ (
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+    "srl             $t4, %[width], 4              \n"  // multiplies of 16
+    "blez            $t4, 2f                       \n"
+    " andi           %[width], %[width], 0xf       \n"  // residual
+
+    ".p2align        2                             \n"
+  "1:                                              \n"
+    "addiu           $t4, $t4, -1                  \n"
+    "lw              $t0, 0(%[src_uv])             \n"  // V1 | U1 | V0 | U0
+    "lw              $t1, 4(%[src_uv])             \n"  // V3 | U3 | V2 | U2
+    "lw              $t2, 8(%[src_uv])             \n"  // V5 | U5 | V4 | U4
+    "lw              $t3, 12(%[src_uv])            \n"  // V7 | U7 | V6 | U6
+    "lw              $t5, 16(%[src_uv])            \n"  // V9 | U9 | V8 | U8
+    "lw              $t6, 20(%[src_uv])            \n"  // V11 | U11 | V10 | U10
+    "lw              $t7, 24(%[src_uv])            \n"  // V13 | U13 | V12 | U12
+    "lw              $t8, 28(%[src_uv])            \n"  // V15 | U15 | V14 | U14
+    "addiu           %[src_uv], %[src_uv], 32      \n"
+    "precrq.qb.ph    $t9, $t1, $t0                 \n"  // V3 | V2 | V1 | V0
+    "precr.qb.ph     $t0, $t1, $t0                 \n"  // U3 | U2 | U1 | U0
+    "precrq.qb.ph    $t1, $t3, $t2                 \n"  // V7 | V6 | V5 | V4
+    "precr.qb.ph     $t2, $t3, $t2                 \n"  // U7 | U6 | U5 | U4
+    "precrq.qb.ph    $t3, $t6, $t5                 \n"  // V11 | V10 | V9 | V8
+    "precr.qb.ph     $t5, $t6, $t5                 \n"  // U11 | U10 | U9 | U8
+    "precrq.qb.ph    $t6, $t8, $t7                 \n"  // V15 | V14 | V13 | V12
+    "precr.qb.ph     $t7, $t8, $t7                 \n"  // U15 | U14 | U13 | U12
+    "sw              $t9, 0(%[dst_v])              \n"
+    "sw              $t0, 0(%[dst_u])              \n"
+    "sw              $t1, 4(%[dst_v])              \n"
+    "sw              $t2, 4(%[dst_u])              \n"
+    "sw              $t3, 8(%[dst_v])              \n"
+    "sw              $t5, 8(%[dst_u])              \n"
+    "sw              $t6, 12(%[dst_v])             \n"
+    "sw              $t7, 12(%[dst_u])             \n"
+    "addiu           %[dst_v], %[dst_v], 16        \n"
+    "bgtz            $t4, 1b                       \n"
+    " addiu          %[dst_u], %[dst_u], 16        \n"
+
+    "beqz            %[width], 3f                  \n"
+    " nop                                          \n"
+
+  "2:                                              \n"
+    "lbu             $t0, 0(%[src_uv])             \n"
+    "lbu             $t1, 1(%[src_uv])             \n"
+    "addiu           %[src_uv], %[src_uv], 2       \n"
+    "addiu           %[width], %[width], -1        \n"
+    "sb              $t0, 0(%[dst_u])              \n"
+    "sb              $t1, 0(%[dst_v])              \n"
+    "addiu           %[dst_u], %[dst_u], 1         \n"
+    "bgtz            %[width], 2b                  \n"
+    " addiu          %[dst_v], %[dst_v], 1         \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+     : [src_uv] "+r" (src_uv),
+       [width] "+r" (width),
+       [dst_u] "+r" (dst_u),
+       [dst_v] "+r" (dst_v)
+     :
+     : "t0", "t1", "t2", "t3",
+     "t4", "t5", "t6", "t7", "t8", "t9"
+  );
+}
+
+void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width) {
+  __asm__ __volatile__ (
+    ".set push                             \n"
+    ".set noreorder                        \n"
+
+    "srl       $t4, %[width], 4            \n"  // multiplies of 16
+    "andi      $t5, %[width], 0xf          \n"
+    "blez      $t4, 2f                     \n"
+    " addu     %[src], %[src], %[width]    \n"  // src += width
+
+    ".p2align  2                           \n"
+   "1:                                     \n"
+    "lw        $t0, -16(%[src])            \n"  // |3|2|1|0|
+    "lw        $t1, -12(%[src])            \n"  // |7|6|5|4|
+    "lw        $t2, -8(%[src])             \n"  // |11|10|9|8|
+    "lw        $t3, -4(%[src])             \n"  // |15|14|13|12|
+    "wsbh      $t0, $t0                    \n"  // |2|3|0|1|
+    "wsbh      $t1, $t1                    \n"  // |6|7|4|5|
+    "wsbh      $t2, $t2                    \n"  // |10|11|8|9|
+    "wsbh      $t3, $t3                    \n"  // |14|15|12|13|
+    "rotr      $t0, $t0, 16                \n"  // |0|1|2|3|
+    "rotr      $t1, $t1, 16                \n"  // |4|5|6|7|
+    "rotr      $t2, $t2, 16                \n"  // |8|9|10|11|
+    "rotr      $t3, $t3, 16                \n"  // |12|13|14|15|
+    "addiu     %[src], %[src], -16         \n"
+    "addiu     $t4, $t4, -1                \n"
+    "sw        $t3, 0(%[dst])              \n"  // |15|14|13|12|
+    "sw        $t2, 4(%[dst])              \n"  // |11|10|9|8|
+    "sw        $t1, 8(%[dst])              \n"  // |7|6|5|4|
+    "sw        $t0, 12(%[dst])             \n"  // |3|2|1|0|
+    "bgtz      $t4, 1b                     \n"
+    " addiu    %[dst], %[dst], 16          \n"
+    "beqz      $t5, 3f                     \n"
+    " nop                                  \n"
+
+   "2:                                     \n"
+    "lbu       $t0, -1(%[src])             \n"
+    "addiu     $t5, $t5, -1                \n"
+    "addiu     %[src], %[src], -1          \n"
+    "sb        $t0, 0(%[dst])              \n"
+    "bgez      $t5, 2b                     \n"
+    " addiu    %[dst], %[dst], 1           \n"
+
+   "3:                                     \n"
+    ".set pop                              \n"
+      : [src] "+r" (src), [dst] "+r" (dst)
+      : [width] "r" (width)
+      : "t0", "t1", "t2", "t3", "t4", "t5"
+  );
+}
+
+void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                            int width) {
+  int x = 0;
+  int y = 0;
+  __asm__ __volatile__ (
+    ".set push                                    \n"
+    ".set noreorder                               \n"
+
+    "addu            $t4, %[width], %[width]      \n"
+    "srl             %[x], %[width], 4            \n"
+    "andi            %[y], %[width], 0xf          \n"
+    "blez            %[x], 2f                     \n"
+    " addu           %[src_uv], %[src_uv], $t4    \n"
+
+    ".p2align        2                            \n"
+   "1:                                            \n"
+    "lw              $t0, -32(%[src_uv])          \n"  // |3|2|1|0|
+    "lw              $t1, -28(%[src_uv])          \n"  // |7|6|5|4|
+    "lw              $t2, -24(%[src_uv])          \n"  // |11|10|9|8|
+    "lw              $t3, -20(%[src_uv])          \n"  // |15|14|13|12|
+    "lw              $t4, -16(%[src_uv])          \n"  // |19|18|17|16|
+    "lw              $t6, -12(%[src_uv])          \n"  // |23|22|21|20|
+    "lw              $t7, -8(%[src_uv])           \n"  // |27|26|25|24|
+    "lw              $t8, -4(%[src_uv])           \n"  // |31|30|29|28|
+
+    "rotr            $t0, $t0, 16                 \n"  // |1|0|3|2|
+    "rotr            $t1, $t1, 16                 \n"  // |5|4|7|6|
+    "rotr            $t2, $t2, 16                 \n"  // |9|8|11|10|
+    "rotr            $t3, $t3, 16                 \n"  // |13|12|15|14|
+    "rotr            $t4, $t4, 16                 \n"  // |17|16|19|18|
+    "rotr            $t6, $t6, 16                 \n"  // |21|20|23|22|
+    "rotr            $t7, $t7, 16                 \n"  // |25|24|27|26|
+    "rotr            $t8, $t8, 16                 \n"  // |29|28|31|30|
+    "precr.qb.ph     $t9, $t0, $t1                \n"  // |0|2|4|6|
+    "precrq.qb.ph    $t5, $t0, $t1                \n"  // |1|3|5|7|
+    "precr.qb.ph     $t0, $t2, $t3                \n"  // |8|10|12|14|
+    "precrq.qb.ph    $t1, $t2, $t3                \n"  // |9|11|13|15|
+    "precr.qb.ph     $t2, $t4, $t6                \n"  // |16|18|20|22|
+    "precrq.qb.ph    $t3, $t4, $t6                \n"  // |17|19|21|23|
+    "precr.qb.ph     $t4, $t7, $t8                \n"  // |24|26|28|30|
+    "precrq.qb.ph    $t6, $t7, $t8                \n"  // |25|27|29|31|
+    "addiu           %[src_uv], %[src_uv], -32    \n"
+    "addiu           %[x], %[x], -1               \n"
+    "swr             $t4, 0(%[dst_u])             \n"
+    "swl             $t4, 3(%[dst_u])             \n"  // |30|28|26|24|
+    "swr             $t6, 0(%[dst_v])             \n"
+    "swl             $t6, 3(%[dst_v])             \n"  // |31|29|27|25|
+    "swr             $t2, 4(%[dst_u])             \n"
+    "swl             $t2, 7(%[dst_u])             \n"  // |22|20|18|16|
+    "swr             $t3, 4(%[dst_v])             \n"
+    "swl             $t3, 7(%[dst_v])             \n"  // |23|21|19|17|
+    "swr             $t0, 8(%[dst_u])             \n"
+    "swl             $t0, 11(%[dst_u])            \n"  // |14|12|10|8|
+    "swr             $t1, 8(%[dst_v])             \n"
+    "swl             $t1, 11(%[dst_v])            \n"  // |15|13|11|9|
+    "swr             $t9, 12(%[dst_u])            \n"
+    "swl             $t9, 15(%[dst_u])            \n"  // |6|4|2|0|
+    "swr             $t5, 12(%[dst_v])            \n"
+    "swl             $t5, 15(%[dst_v])            \n"  // |7|5|3|1|
+    "addiu           %[dst_v], %[dst_v], 16       \n"
+    "bgtz            %[x], 1b                     \n"
+    " addiu          %[dst_u], %[dst_u], 16       \n"
+    "beqz            %[y], 3f                     \n"
+    " nop                                         \n"
+    "b               2f                           \n"
+    " nop                                         \n"
+
+   "2:                                            \n"
+    "lbu             $t0, -2(%[src_uv])           \n"
+    "lbu             $t1, -1(%[src_uv])           \n"
+    "addiu           %[src_uv], %[src_uv], -2     \n"
+    "addiu           %[y], %[y], -1               \n"
+    "sb              $t0, 0(%[dst_u])             \n"
+    "sb              $t1, 0(%[dst_v])             \n"
+    "addiu           %[dst_u], %[dst_u], 1        \n"
+    "bgtz            %[y], 2b                     \n"
+    " addiu          %[dst_v], %[dst_v], 1        \n"
+
+   "3:                                            \n"
+    ".set pop                                     \n"
+      : [src_uv] "+r" (src_uv),
+        [dst_u] "+r" (dst_u),
+        [dst_v] "+r" (dst_v),
+        [x] "=&r" (x),
+        [y] "+r" (y)
+      : [width] "r" (width)
+      : "t0", "t1", "t2", "t3", "t4",
+      "t5", "t7", "t8", "t9"
+  );
+}
+
+// Convert (4 Y and 2 VU) I422 and arrange RGB values into
+// t5 = | 0 | B0 | 0 | b0 |
+// t4 = | 0 | B1 | 0 | b1 |
+// t9 = | 0 | G0 | 0 | g0 |
+// t8 = | 0 | G1 | 0 | g1 |
+// t2 = | 0 | R0 | 0 | r0 |
+// t1 = | 0 | R1 | 0 | r1 |
+#define I422ToTransientMipsRGB                                                 \
+      "lw                $t0, 0(%[y_buf])       \n"                            \
+      "lhu               $t1, 0(%[u_buf])       \n"                            \
+      "lhu               $t2, 0(%[v_buf])       \n"                            \
+      "preceu.ph.qbr     $t1, $t1               \n"                            \
+      "preceu.ph.qbr     $t2, $t2               \n"                            \
+      "preceu.ph.qbra    $t3, $t0               \n"                            \
+      "preceu.ph.qbla    $t0, $t0               \n"                            \
+      "subu.ph           $t1, $t1, $s5          \n"                            \
+      "subu.ph           $t2, $t2, $s5          \n"                            \
+      "subu.ph           $t3, $t3, $s4          \n"                            \
+      "subu.ph           $t0, $t0, $s4          \n"                            \
+      "mul.ph            $t3, $t3, $s0          \n"                            \
+      "mul.ph            $t0, $t0, $s0          \n"                            \
+      "shll.ph           $t4, $t1, 0x7          \n"                            \
+      "subu.ph           $t4, $t4, $t1          \n"                            \
+      "mul.ph            $t6, $t1, $s1          \n"                            \
+      "mul.ph            $t1, $t2, $s2          \n"                            \
+      "addq_s.ph         $t5, $t4, $t3          \n"                            \
+      "addq_s.ph         $t4, $t4, $t0          \n"                            \
+      "shra.ph           $t5, $t5, 6            \n"                            \
+      "shra.ph           $t4, $t4, 6            \n"                            \
+      "addiu             %[u_buf], 2            \n"                            \
+      "addiu             %[v_buf], 2            \n"                            \
+      "addu.ph           $t6, $t6, $t1          \n"                            \
+      "mul.ph            $t1, $t2, $s3          \n"                            \
+      "addu.ph           $t9, $t6, $t3          \n"                            \
+      "addu.ph           $t8, $t6, $t0          \n"                            \
+      "shra.ph           $t9, $t9, 6            \n"                            \
+      "shra.ph           $t8, $t8, 6            \n"                            \
+      "addu.ph           $t2, $t1, $t3          \n"                            \
+      "addu.ph           $t1, $t1, $t0          \n"                            \
+      "shra.ph           $t2, $t2, 6            \n"                            \
+      "shra.ph           $t1, $t1, 6            \n"                            \
+      "subu.ph           $t5, $t5, $s5          \n"                            \
+      "subu.ph           $t4, $t4, $s5          \n"                            \
+      "subu.ph           $t9, $t9, $s5          \n"                            \
+      "subu.ph           $t8, $t8, $s5          \n"                            \
+      "subu.ph           $t2, $t2, $s5          \n"                            \
+      "subu.ph           $t1, $t1, $s5          \n"                            \
+      "shll_s.ph         $t5, $t5, 8            \n"                            \
+      "shll_s.ph         $t4, $t4, 8            \n"                            \
+      "shll_s.ph         $t9, $t9, 8            \n"                            \
+      "shll_s.ph         $t8, $t8, 8            \n"                            \
+      "shll_s.ph         $t2, $t2, 8            \n"                            \
+      "shll_s.ph         $t1, $t1, 8            \n"                            \
+      "shra.ph           $t5, $t5, 8            \n"                            \
+      "shra.ph           $t4, $t4, 8            \n"                            \
+      "shra.ph           $t9, $t9, 8            \n"                            \
+      "shra.ph           $t8, $t8, 8            \n"                            \
+      "shra.ph           $t2, $t2, 8            \n"                            \
+      "shra.ph           $t1, $t1, 8            \n"                            \
+      "addu.ph           $t5, $t5, $s5          \n"                            \
+      "addu.ph           $t4, $t4, $s5          \n"                            \
+      "addu.ph           $t9, $t9, $s5          \n"                            \
+      "addu.ph           $t8, $t8, $s5          \n"                            \
+      "addu.ph           $t2, $t2, $s5          \n"                            \
+      "addu.ph           $t1, $t1, $s5          \n"
+
+void I422ToARGBRow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128| // clipping
+    "lui               $s6, 0xff00            \n"
+    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|ff|
+
+    ".p2align          2                      \n"
+   "1:                                        \n"
+      I422ToTransientMipsRGB
+// Arranging into argb format
+    "precr.qb.ph       $t4, $t8, $t4          \n"  // |G1|g1|B1|b1|
+    "precr.qb.ph       $t5, $t9, $t5          \n"  // |G0|g0|B0|b0|
+    "addiu             %[width], -4           \n"
+    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |G1|B1|G0|B0|
+    "precr.qb.ph       $t9, $t4, $t5          \n"  // |g1|b1|g0|b0|
+    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
+
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
+    "or                $t1, $t1, $s6          \n"  // |ff|R1|ff|R0|
+    "or                $t2, $t2, $s6          \n"  // |ff|r1|ff|r0|
+    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|r1|g1|b1|
+    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|R1|G1|B1|
+    "sll               $t9, $t9, 16           \n"
+    "sll               $t8, $t8, 16           \n"
+    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|r0|g0|b0|
+    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|R0|G0|B0|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+void I422ToABGRRow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74|74|
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128|
+    "lui               $s6, 0xff00            \n"
+    "ori               $s6, 0xff00            \n"  // |ff|00|ff|00|
+
+    ".p2align          2                       \n"
+   "1:                                         \n"
+      I422ToTransientMipsRGB
+// Arranging into abgr format
+    "precr.qb.ph      $t0, $t8, $t1           \n"  // |G1|g1|R1|r1|
+    "precr.qb.ph      $t3, $t9, $t2           \n"  // |G0|g0|R0|r0|
+    "precrq.qb.ph     $t8, $t0, $t3           \n"  // |G1|R1|G0|R0|
+    "precr.qb.ph      $t9, $t0, $t3           \n"  // |g1|r1|g0|r0|
+
+    "precr.qb.ph       $t2, $t4, $t5          \n"  // |B1|b1|B0|b0|
+    "addiu             %[width], -4           \n"
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |B1|0 |B0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |b1|0 |b0|
+    "or                $t1, $t1, $s6          \n"  // |ff|B1|ff|B0|
+    "or                $t2, $t2, $s6          \n"  // |ff|b1|ff|b0|
+    "precrq.ph.w       $t0, $t2, $t9          \n"  // |ff|b1|g1|r1|
+    "precrq.ph.w       $t3, $t1, $t8          \n"  // |ff|B1|G1|R1|
+    "sll               $t9, $t9, 16           \n"
+    "sll               $t8, $t8, 16           \n"
+    "packrl.ph         $t2, $t2, $t9          \n"  // |ff|b0|g0|r0|
+    "packrl.ph         $t1, $t1, $t8          \n"  // |ff|B0|G0|R0|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+void I422ToBGRARow_MIPS_DSPR2(const uint8* y_buf,
+                              const uint8* u_buf,
+                              const uint8* v_buf,
+                              uint8* rgb_buf,
+                              int width) {
+  __asm__ __volatile__ (
+    ".set push                                \n"
+    ".set noreorder                           \n"
+    "beqz              %[width], 2f           \n"
+    " repl.ph          $s0, 74                \n"  // |YG|YG| = |74 |74 |
+    "repl.ph           $s1, -25               \n"  // |UG|UG| = |-25|-25|
+    "repl.ph           $s2, -52               \n"  // |VG|VG| = |-52|-52|
+    "repl.ph           $s3, 102               \n"  // |VR|VR| = |102|102|
+    "repl.ph           $s4, 16                \n"  // |0|16|0|16|
+    "repl.ph           $s5, 128               \n"  // |128|128|
+    "lui               $s6, 0xff              \n"
+    "ori               $s6, 0xff              \n"  // |00|ff|00|ff|
+
+    ".p2align          2                      \n"
+   "1:                                        \n"
+      I422ToTransientMipsRGB
+      // Arranging into bgra format
+    "precr.qb.ph       $t4, $t4, $t8          \n"  // |B1|b1|G1|g1|
+    "precr.qb.ph       $t5, $t5, $t9          \n"  // |B0|b0|G0|g0|
+    "precrq.qb.ph      $t8, $t4, $t5          \n"  // |B1|G1|B0|G0|
+    "precr.qb.ph       $t9, $t4, $t5          \n"  // |b1|g1|b0|g0|
+
+    "precr.qb.ph       $t2, $t1, $t2          \n"  // |R1|r1|R0|r0|
+    "addiu             %[width], -4           \n"
+    "addiu             %[y_buf], 4            \n"
+    "preceu.ph.qbla    $t1, $t2               \n"  // |0 |R1|0 |R0|
+    "preceu.ph.qbra    $t2, $t2               \n"  // |0 |r1|0 |r0|
+    "sll               $t1, $t1, 8            \n"  // |R1|0 |R0|0 |
+    "sll               $t2, $t2, 8            \n"  // |r1|0 |r0|0 |
+    "or                $t1, $t1, $s6          \n"  // |R1|ff|R0|ff|
+    "or                $t2, $t2, $s6          \n"  // |r1|ff|r0|ff|
+    "precrq.ph.w       $t0, $t9, $t2          \n"  // |b1|g1|r1|ff|
+    "precrq.ph.w       $t3, $t8, $t1          \n"  // |B1|G1|R1|ff|
+    "sll               $t1, $t1, 16           \n"
+    "sll               $t2, $t2, 16           \n"
+    "packrl.ph         $t2, $t9, $t2          \n"  // |b0|g0|r0|ff|
+    "packrl.ph         $t1, $t8, $t1          \n"  // |B0|G0|R0|ff|
+// Store results.
+    "sw                $t2, 0(%[rgb_buf])     \n"
+    "sw                $t0, 4(%[rgb_buf])     \n"
+    "sw                $t1, 8(%[rgb_buf])     \n"
+    "sw                $t3, 12(%[rgb_buf])    \n"
+    "bnez              %[width], 1b           \n"
+    " addiu            %[rgb_buf], 16         \n"
+   "2:                                        \n"
+    ".set pop                                 \n"
+      :[y_buf] "+r" (y_buf),
+       [u_buf] "+r" (u_buf),
+       [v_buf] "+r" (v_buf),
+       [width] "+r" (width),
+       [rgb_buf] "+r" (rgb_buf)
+      :
+      : "t0", "t1",  "t2", "t3",  "t4", "t5",
+      "t6", "t7", "t8", "t9",
+      "s0", "s1", "s2", "s3",
+      "s4", "s5", "s6"
+  );
+}
+
+// Bilinear filter 8x2 -> 8x1
+void InterpolateRow_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
+                               ptrdiff_t src_stride, int dst_width,
+                               int source_y_fraction) {
+    int y0_fraction = 256 - source_y_fraction;
+    const uint8* src_ptr1 = src_ptr + src_stride;
+
+  __asm__ __volatile__ (
+     ".set push                                           \n"
+     ".set noreorder                                      \n"
+
+     "replv.ph          $t0, %[y0_fraction]               \n"
+     "replv.ph          $t1, %[source_y_fraction]         \n"
+
+    ".p2align           2                                 \n"
+   "1:                                                    \n"
+     "lw                $t2, 0(%[src_ptr])                \n"
+     "lw                $t3, 0(%[src_ptr1])               \n"
+     "lw                $t4, 4(%[src_ptr])                \n"
+     "lw                $t5, 4(%[src_ptr1])               \n"
+     "muleu_s.ph.qbl    $t6, $t2, $t0                     \n"
+     "muleu_s.ph.qbr    $t7, $t2, $t0                     \n"
+     "muleu_s.ph.qbl    $t8, $t3, $t1                     \n"
+     "muleu_s.ph.qbr    $t9, $t3, $t1                     \n"
+     "muleu_s.ph.qbl    $t2, $t4, $t0                     \n"
+     "muleu_s.ph.qbr    $t3, $t4, $t0                     \n"
+     "muleu_s.ph.qbl    $t4, $t5, $t1                     \n"
+     "muleu_s.ph.qbr    $t5, $t5, $t1                     \n"
+     "addq.ph           $t6, $t6, $t8                     \n"
+     "addq.ph           $t7, $t7, $t9                     \n"
+     "addq.ph           $t2, $t2, $t4                     \n"
+     "addq.ph           $t3, $t3, $t5                     \n"
+     "shra.ph           $t6, $t6, 8                       \n"
+     "shra.ph           $t7, $t7, 8                       \n"
+     "shra.ph           $t2, $t2, 8                       \n"
+     "shra.ph           $t3, $t3, 8                       \n"
+     "precr.qb.ph       $t6, $t6, $t7                     \n"
+     "precr.qb.ph       $t2, $t2, $t3                     \n"
+     "addiu             %[src_ptr], %[src_ptr], 8         \n"
+     "addiu             %[src_ptr1], %[src_ptr1], 8       \n"
+     "addiu             %[dst_width], %[dst_width], -8    \n"
+     "sw                $t6, 0(%[dst_ptr])                \n"
+     "sw                $t2, 4(%[dst_ptr])                \n"
+     "bgtz              %[dst_width], 1b                  \n"
+     " addiu            %[dst_ptr], %[dst_ptr], 8         \n"
+
+     ".set pop                                            \n"
+  : [dst_ptr] "+r" (dst_ptr),
+    [src_ptr1] "+r" (src_ptr1),
+    [src_ptr] "+r" (src_ptr),
+    [dst_width] "+r" (dst_width)
+  : [source_y_fraction] "r" (source_y_fraction),
+    [y0_fraction] "r" (y0_fraction),
+    [src_stride] "r" (src_stride)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+#endif  // __mips_dsp_rev >= 2
+
+#endif  // defined(__mips__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/row_neon.cc b/third_party/aom/third_party/libyuv/source/row_neon.cc
new file mode 100644
index 0000000000..1a72eb9039
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/row_neon.cc
@@ -0,0 +1,3084 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+// Read 8 Y, 4 U and 4 V from 422
+#define READYUV422                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.32    {d2[0]}, [%1]!                 \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.32    {d2[1]}, [%2]!                 \n"
+
+// Read 8 Y, 2 U and 2 V from 422
+#define READYUV411                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.16    {d2[0]}, [%1]!                 \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.16    {d2[1]}, [%2]!                 \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vzip.u8    d2, d3                         \n"
+
+// Read 8 Y, 8 U and 8 V from 444
+#define READYUV444                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    MEMACCESS(2)                                                               \
+    "vld1.8     {d3}, [%2]!                    \n"                             \
+    "vpaddl.u8  q1, q1                         \n"                             \
+    "vrshrn.u16 d2, q1, #1                     \n"
+
+// Read 8 Y, and set 4 U and 4 V to 128
+#define READYUV400                                                             \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    "vmov.u8    d2, #128                       \n"
+
+// Read 8 Y and 4 UV from NV12
+#define READNV12                                                               \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 Y and 4 VU from NV21
+#define READNV21                                                               \
+    MEMACCESS(0)                                                               \
+    "vld1.8     {d0}, [%0]!                    \n"                             \
+    MEMACCESS(1)                                                               \
+    "vld1.8     {d2}, [%1]!                    \n"                             \
+    "vmov.u8    d3, d2                         \n"/* split odd/even uv apart */\
+    "vuzp.u8    d3, d2                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 YUY2
+#define READYUY2                                                               \
+    MEMACCESS(0)                                                               \
+    "vld2.8     {d0, d2}, [%0]!                \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+// Read 8 UYVY
+#define READUYVY                                                               \
+    MEMACCESS(0)                                                               \
+    "vld2.8     {d2, d3}, [%0]!                \n"                             \
+    "vmov.u8    d0, d3                         \n"                             \
+    "vmov.u8    d3, d2                         \n"                             \
+    "vuzp.u8    d2, d3                         \n"                             \
+    "vtrn.u32   d2, d3                         \n"
+
+#define YUV422TORGB_SETUP_REG                                                  \
+    MEMACCESS([kUVToRB])                                                       \
+    "vld1.8     {d24}, [%[kUVToRB]]            \n"                             \
+    MEMACCESS([kUVToG])                                                        \
+    "vld1.8     {d25}, [%[kUVToG]]             \n"                             \
+    MEMACCESS([kUVBiasBGR])                                                    \
+    "vld1.16    {d26[], d27[]}, [%[kUVBiasBGR]]! \n"                           \
+    MEMACCESS([kUVBiasBGR])                                                    \
+    "vld1.16    {d8[], d9[]}, [%[kUVBiasBGR]]!   \n"                           \
+    MEMACCESS([kUVBiasBGR])                                                    \
+    "vld1.16    {d28[], d29[]}, [%[kUVBiasBGR]]  \n"                           \
+    MEMACCESS([kYToRgb])                                                       \
+    "vld1.32    {d30[], d31[]}, [%[kYToRgb]]     \n"
+
+#define YUV422TORGB                                                            \
+    "vmull.u8   q8, d2, d24                    \n" /* u/v B/R component      */\
+    "vmull.u8   q9, d2, d25                    \n" /* u/v G component        */\
+    "vmovl.u8   q0, d0                         \n" /* Y                      */\
+    "vmovl.s16  q10, d1                        \n"                             \
+    "vmovl.s16  q0, d0                         \n"                             \
+    "vmul.s32   q10, q10, q15                  \n"                             \
+    "vmul.s32   q0, q0, q15                    \n"                             \
+    "vqshrun.s32 d0, q0, #16                   \n"                             \
+    "vqshrun.s32 d1, q10, #16                  \n" /* Y                      */\
+    "vadd.s16   d18, d19                       \n"                             \
+    "vshll.u16  q1, d16, #16                   \n" /* Replicate u * UB       */\
+    "vshll.u16  q10, d17, #16                  \n" /* Replicate v * VR       */\
+    "vshll.u16  q3, d18, #16                   \n" /* Replicate (v*VG + u*UG)*/\
+    "vaddw.u16  q1, q1, d16                    \n"                             \
+    "vaddw.u16  q10, q10, d17                  \n"                             \
+    "vaddw.u16  q3, q3, d18                    \n"                             \
+    "vqadd.s16  q8, q0, q13                    \n" /* B */                     \
+    "vqadd.s16  q9, q0, q14                    \n" /* R */                     \
+    "vqadd.s16  q0, q0, q4                     \n" /* G */                     \
+    "vqadd.s16  q8, q8, q1                     \n" /* B */                     \
+    "vqadd.s16  q9, q9, q10                    \n" /* R */                     \
+    "vqsub.s16  q0, q0, q3                     \n" /* G */                     \
+    "vqshrun.s16 d20, q8, #6                   \n" /* B */                     \
+    "vqshrun.s16 d22, q9, #6                   \n" /* R */                     \
+    "vqshrun.s16 d21, q0, #6                   \n" /* G */
+
+// YUV to RGB conversion constants.
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* -min(128, round(2.018 * 64)) */
+#define UG 25 /* -round(-0.391 * 64) */
+#define VG 52 /* -round(-0.813 * 64) */
+#define VR -102 /* -round(1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            - YGB)
+#define BG (UG * 128 + VG * 128 - YGB)
+#define BR            (VR * 128 - YGB)
+
+static uvec8 kUVToRB  = { 128, 128, 128, 128, 102, 102, 102, 102,
+                          0, 0, 0, 0, 0, 0, 0, 0 };
+static uvec8 kUVToG = { 25, 25, 25, 25, 52, 52, 52, 52,
+                        0, 0, 0, 0, 0, 0, 0, 0 };
+static vec16 kUVBiasBGR = { BB, BG, BR, 0, 0, 0, 0, 0 };
+static vec32 kYToRgb = { 0x0101 * YG, 0, 0, 0 };
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV444
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV411
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d19, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_bgra),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d20, d21, d22, d23}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_abgr),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d19, #255                      \n"
+    MEMACCESS(3)
+    "vst4.8     {d19, d20, d21, d22}, [%3]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgba),  // %3
+      "+r"(width)      // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    MEMACCESS(3)
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),      // %0
+      "+r"(src_u),      // %1
+      "+r"(src_v),      // %2
+      "+r"(dst_rgb24),  // %3
+      "+r"(width)       // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vswp.u8    d20, d22                       \n"
+    MEMACCESS(3)
+    "vst3.8     {d20, d21, d22}, [%3]!         \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_raw),  // %3
+      "+r"(width)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTORGB565                                                           \
+    "vshr.u8    d20, d20, #3                   \n"  /* B                    */ \
+    "vshr.u8    d21, d21, #2                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #11                  \n"  /* R                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q0, q0, q10                    \n"  /* BGR                  */
+
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_rgb565),  // %3
+      "+r"(width)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTOARGB1555                                                         \
+    "vshr.u8    q10, q10, #3                   \n"  /* B                    */ \
+    "vshr.u8    d22, d22, #3                   \n"  /* R                    */ \
+    "vshr.u8    d23, d23, #7                   \n"  /* A                    */ \
+    "vmovl.u8   q8, d20                        \n"  /* B                    */ \
+    "vmovl.u8   q9, d21                        \n"  /* G                    */ \
+    "vmovl.u8   q10, d22                       \n"  /* R                    */ \
+    "vmovl.u8   q11, d23                       \n"  /* A                    */ \
+    "vshl.u16   q9, q9, #5                     \n"  /* G                    */ \
+    "vshl.u16   q10, q10, #10                  \n"  /* R                    */ \
+    "vshl.u16   q11, q11, #15                  \n"  /* A                    */ \
+    "vorr       q0, q8, q9                     \n"  /* BG                   */ \
+    "vorr       q1, q10, q11                   \n"  /* RA                   */ \
+    "vorr       q0, q0, q1                     \n"  /* BGRA                 */
+
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB1555
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb1555),  // %3
+      "+r"(width)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#define ARGBTOARGB4444                                                         \
+    "vshr.u8    d20, d20, #4                   \n"  /* B                    */ \
+    "vbic.32    d21, d21, d4                   \n"  /* G                    */ \
+    "vshr.u8    d22, d22, #4                   \n"  /* R                    */ \
+    "vbic.32    d23, d23, d4                   \n"  /* A                    */ \
+    "vorr       d0, d20, d21                   \n"  /* BG                   */ \
+    "vorr       d1, d22, d23                   \n"  /* RA                   */ \
+    "vzip.u8    d0, d1                         \n"  /* BGRA                 */
+
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB
+    "subs       %4, %4, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    ARGBTOARGB4444
+    MEMACCESS(3)
+    "vst1.8     {q0}, [%3]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb4444),  // %3
+      "+r"(width)     // %4
+    : [kUVToRB]"r"(&kUVToRB),   // %5
+      [kUVToG]"r"(&kUVToG),     // %6
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void I400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUV400
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : [kUVToRB]"r"(&kUVToRB),   // %3
+      [kUVToG]"r"(&kUVToG),     // %4
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void J400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "vmov.u8    d23, #255                      \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d20}, [%0]!                   \n"
+    "vmov       d21, d20                       \n"
+    "vmov       d22, d20                       \n"
+    "subs       %2, %2, #8                     \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    :
+    : "cc", "memory", "d20", "d21", "d22", "d23"
+  );
+}
+
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(2)
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(2)
+    "vst4.8     {d20, d21, d22, d23}, [%2]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB
+    "subs       %3, %3, #8                     \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : [kUVToRB]"r"(&kUVToRB),   // %4
+      [kUVToG]"r"(&kUVToG),     // %5
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READYUY2
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_yuy2),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : [kUVToRB]"r"(&kUVToRB),   // %3
+      [kUVToG]"r"(&kUVToG),     // %4
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    READUYVY
+    YUV422TORGB
+    "subs       %2, %2, #8                     \n"
+    "vmov.u8    d23, #255                      \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d21, d22, d23}, [%1]!    \n"
+    "bgt        1b                             \n"
+    : "+r"(src_uyvy),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    : [kUVToRB]"r"(&kUVToRB),   // %3
+      [kUVToG]"r"(&kUVToG),     // %4
+      [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "q0", "q1", "q2", "q3", "q4",
+      "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pairs of UV
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store U
+    MEMACCESS(2)
+    "vst1.8     {q1}, [%2]!                    \n"  // store V
+    "bgt        1b                             \n"
+    : "+r"(src_uv),  // %0
+      "+r"(dst_u),   // %1
+      "+r"(dst_v),   // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load U
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"  // load V
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    MEMACCESS(2)
+    "vst2.u8    {q0, q1}, [%2]!                \n"  // store 16 pairs of UV
+    "bgt        1b                             \n"
+    :
+      "+r"(src_u),   // %0
+      "+r"(src_v),   // %1
+      "+r"(dst_uv),  // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
+void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 32
+    "subs       %2, %2, #32                    \n"  // 32 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 32
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2  // Output registers
+  :                     // Input registers
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SetRow writes 'count' bytes using an 8 bit value repeated.
+void SetRow_NEON(uint8* dst, uint8 v8, int count) {
+  asm volatile (
+    "vdup.8    q0, %2                          \n"  // duplicate 16 bytes
+  "1:                                          \n"
+    "subs      %1, %1, #16                     \n"  // 16 bytes per loop
+    MEMACCESS(0)
+    "vst1.8    {q0}, [%0]!                     \n"  // store
+    "bgt       1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v8)      // %2
+  : "cc", "memory", "q0"
+  );
+}
+
+// ARGBSetRow writes 'count' pixels using an 32 bit value repeated.
+void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "vdup.u32  q0, %2                          \n"  // duplicate 4 ints
+  "1:                                          \n"
+    "subs      %1, %1, #4                      \n"  // 4 pixels per loop
+    MEMACCESS(0)
+    "vst1.8    {q0}, [%0]!                     \n"  // store
+    "bgt       1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "q0"
+  );
+}
+
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2                     \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #16                        \n"  // 16 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r12, #-16                      \n"
+    "add        %0, %0, %3, lsl #1             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d0, d1}, [%0], r12            \n"  // src -= 16
+    "subs       %3, #8                         \n"  // 8 pixels per loop.
+    "vrev64.8   q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // dst += 8
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_u),   // %1
+    "+r"(dst_v),   // %2
+    "+r"(width)    // %3
+  :
+  : "cc", "memory", "r12", "q0"
+  );
+}
+
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  asm volatile (
+    // Start at end of source row.
+    "mov        r3, #-16                       \n"
+    "add        %0, %0, %2, lsl #2             \n"
+    "sub        %0, #16                        \n"
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], r3                 \n"  // src -= 16
+    "subs       %2, #4                         \n"  // 4 pixels per loop.
+    "vrev64.32  q0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // dst += 16
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width)  // %2
+  :
+  : "cc", "memory", "r3", "q0"
+  );
+}
+
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),   // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d1, d2, d3}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    MEMACCESS(1)
+    "vst4.8     {d1, d2, d3, d4}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+#define RGB565TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxGGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB RRRRRxxx */ \
+    "vshl.u8    d6, d6, #2                     \n"  /* G GGGGGG00 upper 6   */ \
+    "vshr.u8    d1, d1, #3                     \n"  /* R 000RRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #6                     \n"  /* G 000000GG lower 2   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB1555TOARGB                                                         \
+    "vshrn.u16  d7, q0, #8                     \n"  /* A Arrrrrxx           */ \
+    "vshr.u8    d6, d7, #2                     \n"  /* R xxxRRRRR           */ \
+    "vshrn.u16  d5, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vmovn.u16  d4, q0                         \n"  /* B xxxBBBBB           */ \
+    "vshr.u8    d7, d7, #7                     \n"  /* A 0000000A           */ \
+    "vneg.s8    d7, d7                         \n"  /* A AAAAAAAA upper 8   */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* R RRRRR000 upper 5   */ \
+    "vshr.u8    q1, q3, #5                     \n"  /* R,A 00000RRR lower 3 */ \
+    "vshl.u8    q0, q2, #3                     \n"  /* B,G BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,G 00000BBB lower 3 */ \
+    "vorr.u8    q1, q1, q3                     \n"  /* R,A                  */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,G                  */ \
+
+// RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
+#define RGB555TOARGB                                                           \
+    "vshrn.u16  d6, q0, #5                     \n"  /* G xxxGGGGG           */ \
+    "vuzp.u8    d0, d1                         \n"  /* d0 xxxBBBBB xRRRRRxx */ \
+    "vshl.u8    d6, d6, #3                     \n"  /* G GGGGG000 upper 5   */ \
+    "vshr.u8    d1, d1, #2                     \n"  /* R 00xRRRRR lower 5   */ \
+    "vshl.u8    q0, q0, #3                     \n"  /* B,R BBBBB000 upper 5 */ \
+    "vshr.u8    q2, q0, #5                     \n"  /* B,R 00000BBB lower 3 */ \
+    "vorr.u8    d0, d0, d4                     \n"  /* B                    */ \
+    "vshr.u8    d4, d6, #5                     \n"  /* G 00000GGG lower 3   */ \
+    "vorr.u8    d2, d1, d5                     \n"  /* R                    */ \
+    "vorr.u8    d1, d4, d6                     \n"  /* G                    */
+
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+#define ARGB4444TOARGB                                                         \
+    "vuzp.u8    d0, d1                         \n"  /* d0 BG, d1 RA         */ \
+    "vshl.u8    q2, q0, #4                     \n"  /* B,R BBBB0000         */ \
+    "vshr.u8    q1, q0, #4                     \n"  /* G,A 0000GGGG         */ \
+    "vshr.u8    q0, q2, #4                     \n"  /* B,R 0000BBBB         */ \
+    "vorr.u8    q0, q0, q2                     \n"  /* B,R BBBBBBBB         */ \
+    "vshl.u8    q2, q1, #4                     \n"  /* G,A GGGG0000         */ \
+    "vorr.u8    q1, q1, q2                     \n"  /* G,A GGGGGGGG         */ \
+    "vswp.u8    d1, d2                         \n"  /* B,R,G,A -> B,G,R,A   */
+
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // Alpha
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RGB24.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_rgb24),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d1, d2, d3, d4}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vswp.u8    d1, d3                         \n"  // swap R, B
+    MEMACCESS(1)
+    "vst3.8     {d1, d2, d3}, [%1]!            \n"  // store 8 pixels of RAW.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_raw),   // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d1", "d2", "d3", "d4"  // Clobber List
+  );
+}
+
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of YUY2.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"  // load 16 pixels of UYVY.
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store 16 pixels of Y.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vst1.8     {d1}, [%1]!                    \n"  // store 8 U.
+    MEMACCESS(2)
+    "vst1.8     {d3}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %3, %3, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 U.
+    MEMACCESS(2)
+    "vst1.8     {d2}, [%2]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  );
+}
+
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_yuy2
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of YUY2.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row YUY2.
+    "vrhadd.u8  d1, d1, d5                     \n"  // average rows of U
+    "vrhadd.u8  d3, d3, d7                     \n"  // average rows of V
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 U.
+    MEMACCESS(3)
+    "vst1.8     {d3}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_yuy2),     // %0
+    "+r"(stride_yuy2),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // stride + src_uyvy
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 16 pixels of UYVY.
+    "subs       %4, %4, #16                    \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load next row UYVY.
+    "vrhadd.u8  d0, d0, d4                     \n"  // average rows of U
+    "vrhadd.u8  d2, d2, d6                     \n"  // average rows of V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 U.
+    MEMACCESS(3)
+    "vst1.8     {d2}, [%3]!                    \n"  // store 8 V.
+    "bgt        1b                             \n"
+  : "+r"(src_uyvy),     // %0
+    "+r"(stride_uyvy),  // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"  // Clobber List
+  );
+}
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q2}, [%3]                     \n"  // shuffler
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 4 pixels.
+    "subs       %2, %2, #4                     \n"  // 4 processed per loop
+    "vtbl.8     d2, {d0, d1}, d4               \n"  // look up 2 first pixels
+    "vtbl.8     d3, {d0, d1}, d5               \n"  // look up 2 next pixels
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store 4.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "cc", "memory", "q0", "q1", "q2"  // Clobber List
+  );
+}
+
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d0, d2}, [%0]!                \n"  // load 16 Ys
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 Us
+    MEMACCESS(2)
+    "vld1.8     {d3}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    MEMACCESS(3)
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 YUY2/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_yuy2),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld2.8     {d1, d3}, [%0]!                \n"  // load 16 Ys
+    MEMACCESS(1)
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 Us
+    MEMACCESS(2)
+    "vld1.8     {d2}, [%2]!                    \n"  // load 8 Vs
+    "subs       %4, %4, #16                    \n"  // 16 pixels
+    MEMACCESS(3)
+    "vst4.8     {d0, d1, d2, d3}, [%3]!        \n"  // Store 8 UYVY/16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_uyvy),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3"
+  );
+}
+
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTORGB565
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_rgb565),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+    "vdup.32    d2, %2                         \n"  // dither4
+  "1:                                          \n"
+    MEMACCESS(1)
+    "vld4.8     {d20, d21, d22, d23}, [%1]!    \n"  // load 8 pixels of ARGB.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d20, d20, d2                   \n"
+    "vqadd.u8   d21, d21, d2                   \n"
+    "vqadd.u8   d22, d22, d2                   \n"
+    ARGBTORGB565
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"  // store 8 pixels RGB565.
+    "bgt        1b                             \n"
+  : "+r"(dst_rgb)    // %0
+  : "r"(src_argb),   // %1
+    "r"(dither4),    // %2
+    "r"(width)       // %3
+  : "cc", "memory", "q0", "q1", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
+                            int pix) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB1555
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB1555.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb1555),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
+                            int pix) {
+  asm volatile (
+    "vmov.u8    d4, #0x0f                      \n"  // bits to clear with vbic.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d21, d22, d23}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGBTOARGB4444
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"  // store 8 pixels ARGB4444.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb4444),  // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q8", "q9", "q10", "q11"
+  );
+}
+
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit Y
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+// 8x1 pixels.
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.u8    d24, #112                      \n"  // UB / VR 0.875 coefficient
+    "vmov.u8    d25, #74                       \n"  // UG -0.5781 coefficient
+    "vmov.u8    d26, #38                       \n"  // UR -0.2969 coefficient
+    "vmov.u8    d27, #18                       \n"  // VB -0.1406 coefficient
+    "vmov.u8    d28, #94                       \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlsl.u8   q2, d1, d25                    \n"  // G
+    "vmlsl.u8   q2, d2, d26                    \n"  // R
+    "vadd.u16   q2, q2, q15                    \n"  // +128 -> unsigned
+
+    "vmull.u8   q3, d2, d24                    \n"  // R
+    "vmlsl.u8   q3, d1, d28                    \n"  // G
+    "vmlsl.u8   q3, d0, d27                    \n"  // B
+    "vadd.u16   q3, q3, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q2, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q3, #8                    \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 32x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 32.
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(0)
+    "vld4.8     {d8, d10, d12, d14}, [%0]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d9, d11, d13, d15}, [%0]!     \n"  // load last 8 ARGB pixels.
+    "vpaddl.u8  q4, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q5, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q6, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vpadd.u16  d0, d0, d1                     \n"  // B 16 shorts -> 8 shorts.
+    "vpadd.u16  d1, d8, d9                     \n"  // B
+    "vpadd.u16  d2, d2, d3                     \n"  // G 16 shorts -> 8 shorts.
+    "vpadd.u16  d3, d10, d11                   \n"  // G
+    "vpadd.u16  d4, d4, d5                     \n"  // R 16 shorts -> 8 shorts.
+    "vpadd.u16  d5, d12, d13                   \n"  // R
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %3, %3, #32                    \n"  // 32 processed per loop.
+    "vmul.s16   q8, q0, q10                    \n"  // B
+    "vmls.s16   q8, q1, q11                    \n"  // G
+    "vmls.s16   q8, q2, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q2, q10                    \n"  // R
+    "vmls.s16   q9, q1, q14                    \n"  // G
+    "vmls.s16   q9, q0, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "vst1.8     {d1}, [%2]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#define RGBTOUV(QB, QG, QR) \
+    "vmul.s16   q8, " #QB ", q10               \n"  /* B                    */ \
+    "vmls.s16   q8, " #QG ", q11               \n"  /* G                    */ \
+    "vmls.s16   q8, " #QR ", q12               \n"  /* R                    */ \
+    "vadd.u16   q8, q8, q15                    \n"  /* +128 -> unsigned     */ \
+    "vmul.s16   q9, " #QR ", q10               \n"  /* R                    */ \
+    "vmls.s16   q9, " #QG ", q14               \n"  /* G                    */ \
+    "vmls.s16   q9, " #QB ", q13               \n"  /* B                    */ \
+    "vadd.u16   q9, q9, q15                    \n"  /* +128 -> unsigned     */ \
+    "vqshrn.u16  d0, q8, #8                    \n"  /* 16 bit to 8 bit U    */ \
+    "vqshrn.u16  d1, q9, #8                    \n"  /* 16 bit to 8 bit V    */
+
+// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): Subsample match C code.
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #127 / 2                  \n"  // UB / VR 0.500 coefficient
+    "vmov.s16   q11, #84 / 2                   \n"  // UG -0.33126 coefficient
+    "vmov.s16   q12, #43 / 2                   \n"  // UR -0.16874 coefficient
+    "vmov.s16   q13, #20 / 2                   \n"  // VB -0.08131 coefficient
+    "vmov.s16   q14, #107 / 2                  \n"  // VG -0.41869 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_stride_argb),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_bgra
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 BGRA pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 BGRA pixels.
+    "vpaddl.u8  q3, q3                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more BGRA pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 BGRA pixels.
+    "vpadal.u8  q3, q7                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q1, q1, #1                     \n"  // 2x average
+    "vrshr.u16  q2, q2, #1                     \n"
+    "vrshr.u16  q3, q3, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q3, q2, q1)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(src_stride_bgra),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_abgr
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ABGR pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ABGR pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more ABGR pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 ABGR pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_stride_abgr),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgba
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 RGBA pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 RGBA pixels.
+    "vpaddl.u8  q0, q1                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q2                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q3                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d8, d10, d12, d14}, [%1]!     \n"  // load 8 more RGBA pixels.
+    MEMACCESS(1)
+    "vld4.8     {d9, d11, d13, d15}, [%1]!     \n"  // load last 8 RGBA pixels.
+    "vpadal.u8  q0, q5                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q6                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q7                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(src_stride_rgba),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_rgb24
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RGB24 pixels.
+    MEMACCESS(0)
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RGB24 pixels.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RGB24 pixels.
+    MEMACCESS(1)
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RGB24 pixels.
+    "vpadal.u8  q0, q4                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q6                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q0, q1, q2)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_stride_rgb24),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_raw
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d2, d4}, [%0]!            \n"  // load 8 RAW pixels.
+    MEMACCESS(0)
+    "vld3.8     {d1, d3, d5}, [%0]!            \n"  // load next 8 RAW pixels.
+    "vpaddl.u8  q2, q2                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q0, q0                         \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld3.8     {d8, d10, d12}, [%1]!          \n"  // load 8 more RAW pixels.
+    MEMACCESS(1)
+    "vld3.8     {d9, d11, d13}, [%1]!          \n"  // load last 8 RAW pixels.
+    "vpadal.u8  q2, q6                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q5                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q0, q4                         \n"  // R 16 bytes -> 8 shorts.
+
+    "vrshr.u16  q0, q0, #1                     \n"  // 2x average
+    "vrshr.u16  q1, q1, #1                     \n"
+    "vrshr.u16  q2, q2, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 32 processed per loop.
+    RGBTOUV(q2, q1, q0)
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_stride_raw),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(src_stride_rgb565),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(src_stride_argb1555),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix) {
+  asm volatile (
+    "add        %1, %0, %1                     \n"  // src_stride + src_argb
+    "vmov.s16   q10, #112 / 2                  \n"  // UB / VR 0.875 coefficient
+    "vmov.s16   q11, #74 / 2                   \n"  // UG -0.5781 coefficient
+    "vmov.s16   q12, #38 / 2                   \n"  // UR -0.2969 coefficient
+    "vmov.s16   q13, #18 / 2                   \n"  // VB -0.1406 coefficient
+    "vmov.s16   q14, #94 / 2                   \n"  // VG -0.7344 coefficient
+    "vmov.u16   q15, #0x8080                   \n"  // 128.5
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpaddl.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpaddl.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpaddl.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d8, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d10, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d12, d2                        \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "vpadal.u8  d9, d0                         \n"  // B 8 bytes -> 4 shorts.
+    "vpadal.u8  d11, d1                        \n"  // G 8 bytes -> 4 shorts.
+    "vpadal.u8  d13, d2                        \n"  // R 8 bytes -> 4 shorts.
+
+    "vrshr.u16  q4, q4, #1                     \n"  // 2x average
+    "vrshr.u16  q5, q5, #1                     \n"
+    "vrshr.u16  q6, q6, #1                     \n"
+
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop.
+    "vmul.s16   q8, q4, q10                    \n"  // B
+    "vmls.s16   q8, q5, q11                    \n"  // G
+    "vmls.s16   q8, q6, q12                    \n"  // R
+    "vadd.u16   q8, q8, q15                    \n"  // +128 -> unsigned
+    "vmul.s16   q9, q6, q10                    \n"  // R
+    "vmls.s16   q9, q5, q14                    \n"  // G
+    "vmls.s16   q9, q4, q13                    \n"  // B
+    "vadd.u16   q9, q9, q15                    \n"  // +128 -> unsigned
+    "vqshrn.u16  d0, q8, #8                    \n"  // 16 bit to 8 bit U
+    "vqshrn.u16  d1, q9, #8                    \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "vst1.8     {d1}, [%3]!                    \n"  // store 8 pixels V.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(src_stride_argb4444),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7",
+    "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 RGB565 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    RGB565TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_y),       // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB1555 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
+    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
+    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
+    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 8 ARGB4444 pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d27                        \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q12", "q13"
+  );
+}
+
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of BGRA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // R
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ABGR.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // R
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // B
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of RGBA.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d1, d4                     \n"  // B
+    "vmlal.u8   q8, d2, d5                     \n"  // G
+    "vmlal.u8   q8, d3, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RGB24.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) {
+  asm volatile (
+    "vmov.u8    d4, #33                        \n"  // R * 0.2578 coefficient
+    "vmov.u8    d5, #65                        \n"  // G * 0.5078 coefficient
+    "vmov.u8    d6, #13                        \n"  // B * 0.1016 coefficient
+    "vmov.u8    d7, #16                        \n"  // Add 16 constant
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld3.8     {d0, d1, d2}, [%0]!            \n"  // load 8 pixels of RAW.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q8, d0, d4                     \n"  // B
+    "vmlal.u8   q8, d1, d5                     \n"  // G
+    "vmlal.u8   q8, d2, d6                     \n"  // R
+    "vqrshrun.s16 d0, q8, #7                   \n"  // 16 bit to 8 bit Y
+    "vqadd.u8   d0, d7                         \n"
+    MEMACCESS(1)
+    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "bgt        1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(dst_y),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "q8"
+  );
+}
+
+// Bilinear filter 16x2 -> 16x1
+void InterpolateRow_NEON(uint8* dst_ptr,
+                         const uint8* src_ptr, ptrdiff_t src_stride,
+                         int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp        %4, #0                         \n"
+    "beq        100f                           \n"
+    "add        %2, %1                         \n"
+    "cmp        %4, #64                        \n"
+    "beq        75f                            \n"
+    "cmp        %4, #128                       \n"
+    "beq        50f                            \n"
+    "cmp        %4, #192                       \n"
+    "beq        25f                            \n"
+
+    "vdup.8     d5, %4                         \n"
+    "rsb        %4, #256                       \n"
+    "vdup.8     d4, %4                         \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vmull.u8   q13, d0, d4                    \n"
+    "vmull.u8   q14, d1, d4                    \n"
+    "vmlal.u8   q13, d2, d5                    \n"
+    "vmlal.u8   q14, d3, d5                    \n"
+    "vrshrn.u16 d0, q13, #8                    \n"
+    "vrshrn.u16 d1, q14, #8                    \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        1b                             \n"
+    "b          99f                            \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        25b                            \n"
+    "b          99f                            \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q1}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        50b                            \n"
+    "b          99f                            \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"
+    MEMACCESS(2)
+    "vld1.8     {q0}, [%2]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    "vrhadd.u8  q0, q1                         \n"
+    "vrhadd.u8  q0, q1                         \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        75b                            \n"
+    "b          99f                            \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "vld1.8     {q0}, [%1]!                    \n"
+    "subs       %3, %3, #16                    \n"
+    MEMACCESS(0)
+    "vst1.8     {q0}, [%0]!                    \n"
+    "bgt        100b                           \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "cc", "memory", "q0", "q1", "d4", "d5", "q13", "q14"
+  );
+}
+
+// dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
+void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "subs       %3, #8                         \n"
+    "blt        89f                            \n"
+    // Blend 8 pixels.
+  "8:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB0.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 pixels of ARGB1.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 pixels of ARGB.
+    "bge        8b                             \n"
+
+  "89:                                         \n"
+    "adds       %3, #8-1                       \n"
+    "blt        99f                            \n"
+
+    // Blend 1 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0[0],d1[0],d2[0],d3[0]}, [%0]! \n"  // load 1 pixel ARGB0.
+    MEMACCESS(1)
+    "vld4.8     {d4[0],d5[0],d6[0],d7[0]}, [%1]! \n"  // load 1 pixel ARGB1.
+    "subs       %3, %3, #1                     \n"  // 1 processed per loop.
+    "vmull.u8   q10, d4, d3                    \n"  // db * a
+    "vmull.u8   q11, d5, d3                    \n"  // dg * a
+    "vmull.u8   q12, d6, d3                    \n"  // dr * a
+    "vqrshrn.u16 d20, q10, #8                  \n"  // db >>= 8
+    "vqrshrn.u16 d21, q11, #8                  \n"  // dg >>= 8
+    "vqrshrn.u16 d22, q12, #8                  \n"  // dr >>= 8
+    "vqsub.u8   q2, q2, q10                    \n"  // dbg - dbg * a / 256
+    "vqsub.u8   d6, d6, d22                    \n"  // dr - dr * a / 256
+    "vqadd.u8   q0, q0, q2                     \n"  // + sbg
+    "vqadd.u8   d2, d2, d6                     \n"  // + sr
+    "vmov.u8    d3, #255                       \n"  // a = 255
+    MEMACCESS(2)
+    "vst4.8     {d0[0],d1[0],d2[0],d3[0]}, [%2]! \n"  // store 1 pixel.
+    "bge        1b                             \n"
+
+  "99:                                         \n"
+
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q10", "q11", "q12"
+  );
+}
+
+// Attenuate 8 pixels at a time.
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    // Attenuate 8 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q10, d0, d3                    \n"  // b * a
+    "vmull.u8   q11, d1, d3                    \n"  // g * a
+    "vmull.u8   q12, d2, d3                    \n"  // r * a
+    "vqrshrn.u16 d0, q10, #8                   \n"  // b >>= 8
+    "vqrshrn.u16 d1, q11, #8                   \n"  // g >>= 8
+    "vqrshrn.u16 d2, q12, #8                   \n"  // r >>= 8
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  :
+  : "cc", "memory", "q0", "q1", "q10", "q11", "q12"
+  );
+}
+
+// Quantize 8 ARGB pixels (32 bytes).
+// dst = (dst * scale >> 16) * interval_size + interval_offset;
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "vdup.u16   q8, %2                         \n"
+    "vshr.u16   q8, q8, #1                     \n"  // scale >>= 1
+    "vdup.u16   q9, %3                         \n"  // interval multiply.
+    "vdup.u16   q10, %4                        \n"  // interval add
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]         \n"  // load 8 pixels of ARGB.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q0, d0                         \n"  // b (0 .. 255)
+    "vmovl.u8   q1, d2                         \n"
+    "vmovl.u8   q2, d4                         \n"
+    "vqdmulh.s16 q0, q0, q8                    \n"  // b * scale
+    "vqdmulh.s16 q1, q1, q8                    \n"  // g
+    "vqdmulh.s16 q2, q2, q8                    \n"  // r
+    "vmul.u16   q0, q0, q9                     \n"  // b * interval_size
+    "vmul.u16   q1, q1, q9                     \n"  // g
+    "vmul.u16   q2, q2, q9                     \n"  // r
+    "vadd.u16   q0, q0, q10                    \n"  // b + interval_offset
+    "vadd.u16   q1, q1, q10                    \n"  // g
+    "vadd.u16   q2, q2, q10                    \n"  // r
+    "vqmovn.u16 d0, q0                         \n"
+    "vqmovn.u16 d2, q1                         \n"
+    "vqmovn.u16 d4, q2                         \n"
+    MEMACCESS(0)
+    "vst4.8     {d0, d2, d4, d6}, [%0]!        \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "cc", "memory", "q0", "q1", "q2", "q3", "q8", "q9", "q10"
+  );
+}
+
+// Shade 8 pixels at a time by specified value.
+// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
+// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "vdup.u32   q0, %3                         \n"  // duplicate scale value.
+    "vzip.u8    d0, d1                         \n"  // d0 aarrggbb.
+    "vshr.u16   q0, q0, #1                     \n"  // scale / 2.
+
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d20, d22, d24, d26}, [%0]!    \n"  // load 8 pixels of ARGB.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q10, d20                       \n"  // b (0 .. 255)
+    "vmovl.u8   q11, d22                       \n"
+    "vmovl.u8   q12, d24                       \n"
+    "vmovl.u8   q13, d26                       \n"
+    "vqrdmulh.s16 q10, q10, d0[0]              \n"  // b * scale * 2
+    "vqrdmulh.s16 q11, q11, d0[1]              \n"  // g
+    "vqrdmulh.s16 q12, q12, d0[2]              \n"  // r
+    "vqrdmulh.s16 q13, q13, d0[3]              \n"  // a
+    "vqmovn.u16 d20, q10                       \n"
+    "vqmovn.u16 d22, q11                       \n"
+    "vqmovn.u16 d24, q12                       \n"
+    "vqmovn.u16 d26, q13                       \n"
+    MEMACCESS(1)
+    "vst4.8     {d20, d22, d24, d26}, [%1]!    \n"  // store 8 pixels of ARGB.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),       // %0
+    "+r"(dst_argb),       // %1
+    "+r"(width)           // %2
+  : "r"(value)            // %3
+  : "cc", "memory", "q0", "q10", "q11", "q12", "q13"
+  );
+}
+
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+// Similar to ARGBToYJ but stores ARGB.
+// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
+    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
+    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d24                    \n"  // B
+    "vmlal.u8   q2, d1, d25                    \n"  // G
+    "vmlal.u8   q2, d2, d26                    \n"  // R
+    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit B
+    "vmov       d1, d0                         \n"  // G
+    "vmov       d2, d0                         \n"  // R
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  );
+}
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d20, #17                       \n"  // BB coefficient
+    "vmov.u8    d21, #68                       \n"  // BG coefficient
+    "vmov.u8    d22, #35                       \n"  // BR coefficient
+    "vmov.u8    d24, #22                       \n"  // GB coefficient
+    "vmov.u8    d25, #88                       \n"  // GG coefficient
+    "vmov.u8    d26, #45                       \n"  // GR coefficient
+    "vmov.u8    d28, #24                       \n"  // BB coefficient
+    "vmov.u8    d29, #98                       \n"  // BG coefficient
+    "vmov.u8    d30, #50                       \n"  // BR coefficient
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]         \n"  // load 8 ARGB pixels.
+    "subs       %1, %1, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q2, d0, d20                    \n"  // B to Sepia B
+    "vmlal.u8   q2, d1, d21                    \n"  // G
+    "vmlal.u8   q2, d2, d22                    \n"  // R
+    "vmull.u8   q3, d0, d24                    \n"  // B to Sepia G
+    "vmlal.u8   q3, d1, d25                    \n"  // G
+    "vmlal.u8   q3, d2, d26                    \n"  // R
+    "vmull.u8   q8, d0, d28                    \n"  // B to Sepia R
+    "vmlal.u8   q8, d1, d29                    \n"  // G
+    "vmlal.u8   q8, d2, d30                    \n"  // R
+    "vqshrn.u16 d0, q2, #7                     \n"  // 16 bit to 8 bit B
+    "vqshrn.u16 d1, q3, #7                     \n"  // 16 bit to 8 bit G
+    "vqshrn.u16 d2, q8, #7                     \n"  // 16 bit to 8 bit R
+    MEMACCESS(0)
+    "vst4.8     {d0, d1, d2, d3}, [%0]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),  // %0
+    "+r"(width)      // %1
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// TODO(fbarchard): Was same as Sepia except matrix is provided.  This function
+// needs to saturate.  Consider doing a non-saturating version.
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q2}, [%3]                     \n"  // load 3 ARGB vectors.
+    "vmovl.s8   q0, d4                         \n"  // B,G coefficients s16.
+    "vmovl.s8   q1, d5                         \n"  // R,A coefficients s16.
+
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d16, d18, d20, d22}, [%0]!    \n"  // load 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop.
+    "vmovl.u8   q8, d16                        \n"  // b (0 .. 255) 16 bit
+    "vmovl.u8   q9, d18                        \n"  // g
+    "vmovl.u8   q10, d20                       \n"  // r
+    "vmovl.u8   q11, d22                       \n"  // a
+    "vmul.s16   q12, q8, d0[0]                 \n"  // B = B * Matrix B
+    "vmul.s16   q13, q8, d1[0]                 \n"  // G = B * Matrix G
+    "vmul.s16   q14, q8, d2[0]                 \n"  // R = B * Matrix R
+    "vmul.s16   q15, q8, d3[0]                 \n"  // A = B * Matrix A
+    "vmul.s16   q4, q9, d0[1]                  \n"  // B += G * Matrix B
+    "vmul.s16   q5, q9, d1[1]                  \n"  // G += G * Matrix G
+    "vmul.s16   q6, q9, d2[1]                  \n"  // R += G * Matrix R
+    "vmul.s16   q7, q9, d3[1]                  \n"  // A += G * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q10, d0[2]                 \n"  // B += R * Matrix B
+    "vmul.s16   q5, q10, d1[2]                 \n"  // G += R * Matrix G
+    "vmul.s16   q6, q10, d2[2]                 \n"  // R += R * Matrix R
+    "vmul.s16   q7, q10, d3[2]                 \n"  // A += R * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vmul.s16   q4, q11, d0[3]                 \n"  // B += A * Matrix B
+    "vmul.s16   q5, q11, d1[3]                 \n"  // G += A * Matrix G
+    "vmul.s16   q6, q11, d2[3]                 \n"  // R += A * Matrix R
+    "vmul.s16   q7, q11, d3[3]                 \n"  // A += A * Matrix A
+    "vqadd.s16  q12, q12, q4                   \n"  // Accumulate B
+    "vqadd.s16  q13, q13, q5                   \n"  // Accumulate G
+    "vqadd.s16  q14, q14, q6                   \n"  // Accumulate R
+    "vqadd.s16  q15, q15, q7                   \n"  // Accumulate A
+    "vqshrun.s16 d16, q12, #6                  \n"  // 16 bit to 8 bit B
+    "vqshrun.s16 d18, q13, #6                  \n"  // 16 bit to 8 bit G
+    "vqshrun.s16 d20, q14, #6                  \n"  // 16 bit to 8 bit R
+    "vqshrun.s16 d22, q15, #6                  \n"  // 16 bit to 8 bit A
+    MEMACCESS(1)
+    "vst4.8     {d16, d18, d20, d22}, [%1]!    \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "r"(matrix_argb)  // %3
+  : "cc", "memory", "q0", "q1", "q2", "q4", "q5", "q6", "q7", "q8", "q9",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+// TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable.
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d1, d3, d5, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vmull.u8   q0, d0, d1                     \n"  // multiply B
+    "vmull.u8   q1, d2, d3                     \n"  // multiply G
+    "vmull.u8   q2, d4, d5                     \n"  // multiply R
+    "vmull.u8   q3, d6, d7                     \n"  // multiply A
+    "vrshrn.u16 d0, q0, #8                     \n"  // 16 bit to 8 bit B
+    "vrshrn.u16 d1, q1, #8                     \n"  // 16 bit to 8 bit G
+    "vrshrn.u16 d2, q2, #8                     \n"  // 16 bit to 8 bit R
+    "vrshrn.u16 d3, q3, #8                     \n"  // 16 bit to 8 bit A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_NEON
+
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   q0, q0, q2                     \n"  // add B, G
+    "vqadd.u8   q1, q1, q3                     \n"  // add R, A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqsub.u8   q0, q0, q2                     \n"  // subtract B, G
+    "vqsub.u8   q1, q1, q3                     \n"  // subtract R, A
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0]!                    \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d0, d0, d1                     \n"  // add
+    "vmov.u8    d1, d0                         \n"
+    "vmov.u8    d2, d0                         \n"
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// Adds Sobel X and Sobel Y and stores Sobel into plane.
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    // 16 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"  // load 16 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {q1}, [%1]!                    \n"  // load 16 sobely.
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop.
+    "vqadd.u8   q0, q0, q1                     \n"  // add
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"  // store 16 pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "vmov.u8    d3, #255                       \n"  // alpha
+    // 8 pixel loop.
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0]!                    \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "vld1.8     {d0}, [%1]!                    \n"  // load 8 sobely.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqadd.u8   d1, d0, d2                     \n"  // add
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "q0", "q1"
+  );
+}
+
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0],%5                  \n"  // top
+    MEMACCESS(0)
+    "vld1.8     {d1}, [%0],%6                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    MEMACCESS(1)
+    "vld1.8     {d2}, [%1],%5                  \n"  // center * 2
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%6                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    MEMACCESS(2)
+    "vld1.8     {d2}, [%2],%5                  \n"  // bottom
+    MEMACCESS(2)
+    "vld1.8     {d3}, [%2],%6                  \n"
+    "subs       %4, %4, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    MEMACCESS(3)
+    "vst1.8     {d0}, [%3]!                    \n"  // store 8 sobelx
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  : "r"(2),            // %5
+    "r"(6)             // %6
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0],%4                  \n"  // left
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1],%4                  \n"
+    "vsubl.u8   q0, d0, d1                     \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0],%4                  \n"  // center * 2
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%4                  \n"
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0],%5                  \n"  // right
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1],%5                  \n"
+    "subs       %3, %3, #8                     \n"  // 8 pixels
+    "vsubl.u8   q1, d2, d3                     \n"
+    "vadd.s16   q0, q0, q1                     \n"
+    "vabs.s16   q0, q0                         \n"
+    "vqmovn.u16 d0, q0                         \n"
+    MEMACCESS(2)
+    "vst1.8     {d0}, [%2]!                    \n"  // store 8 sobely
+    "bgt        1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  : "r"(1),            // %4
+    "r"(6)             // %5
+  : "cc", "memory", "q0", "q1"  // Clobber List
+  );
+}
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/row_neon64.cc b/third_party/aom/third_party/libyuv/source/row_neon64.cc
new file mode 100644
index 0000000000..5d015454b0
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/row_neon64.cc
@@ -0,0 +1,3087 @@
+/*
+ *  Copyright 2014 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon armv8 64 bit.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+// Read 8 Y, 4 U and 4 V from 422
+#define READYUV422                                                             \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    MEMACCESS(1)                                                               \
+    "ld1        {v1.s}[0], [%1], #4            \n"                             \
+    MEMACCESS(2)                                                               \
+    "ld1        {v1.s}[1], [%2], #4            \n"
+
+// Read 8 Y, 2 U and 2 V from 422
+#define READYUV411                                                             \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    MEMACCESS(1)                                                               \
+    "ld1        {v2.h}[0], [%1], #2            \n"                             \
+    MEMACCESS(2)                                                               \
+    "ld1        {v2.h}[1], [%2], #2            \n"                             \
+    "zip1       v1.8b, v2.8b, v2.8b            \n"
+
+// Read 8 Y, 8 U and 8 V from 444
+#define READYUV444                                                             \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    MEMACCESS(1)                                                               \
+    "ld1        {v1.d}[0], [%1], #8            \n"                             \
+    MEMACCESS(2)                                                               \
+    "ld1        {v1.d}[1], [%2], #8            \n"                             \
+    "uaddlp     v1.8h, v1.16b                  \n"                             \
+    "rshrn      v1.8b, v1.8h, #1               \n"
+
+// Read 8 Y, and set 4 U and 4 V to 128
+#define READYUV400                                                             \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    "movi       v1.8b , #128                   \n"
+
+// Read 8 Y and 4 UV from NV12
+#define READNV12                                                               \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    MEMACCESS(1)                                                               \
+    "ld1        {v2.8b}, [%1], #8              \n"                             \
+    "uzp1       v1.8b, v2.8b, v2.8b            \n"                             \
+    "uzp2       v3.8b, v2.8b, v2.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
+
+// Read 8 Y and 4 VU from NV21
+#define READNV21                                                               \
+    MEMACCESS(0)                                                               \
+    "ld1        {v0.8b}, [%0], #8              \n"                             \
+    MEMACCESS(1)                                                               \
+    "ld1        {v2.8b}, [%1], #8              \n"                             \
+    "uzp1       v3.8b, v2.8b, v2.8b            \n"                             \
+    "uzp2       v1.8b, v2.8b, v2.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
+
+// Read 8 YUY2
+#define READYUY2                                                               \
+    MEMACCESS(0)                                                               \
+    "ld2        {v0.8b, v1.8b}, [%0], #16      \n"                             \
+    "uzp2       v3.8b, v1.8b, v1.8b            \n"                             \
+    "uzp1       v1.8b, v1.8b, v1.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
+
+// Read 8 UYVY
+#define READUYVY                                                               \
+    MEMACCESS(0)                                                               \
+    "ld2        {v2.8b, v3.8b}, [%0], #16      \n"                             \
+    "orr        v0.8b, v3.8b, v3.8b            \n"                             \
+    "uzp1       v1.8b, v2.8b, v2.8b            \n"                             \
+    "uzp2       v3.8b, v2.8b, v2.8b            \n"                             \
+    "ins        v1.s[1], v3.s[0]               \n"
+
+#define YUV422TORGB_SETUP_REG                                                  \
+    "ld1r       {v24.8h}, [%[kUVBiasBGR]], #2  \n"                             \
+    "ld1r       {v25.8h}, [%[kUVBiasBGR]], #2  \n"                             \
+    "ld1r       {v26.8h}, [%[kUVBiasBGR]]      \n"                             \
+    "ld1r       {v31.4s}, [%[kYToRgb]]         \n"                             \
+    "movi       v27.8h, #128                   \n"                             \
+    "movi       v28.8h, #102                   \n"                             \
+    "movi       v29.8h, #25                    \n"                             \
+    "movi       v30.8h, #52                    \n"
+
+#define YUV422TORGB(vR, vG, vB)                                                \
+    "uxtl       v0.8h, v0.8b                   \n" /* Extract Y    */          \
+    "shll       v2.8h, v1.8b, #8               \n" /* Replicate UV */          \
+    "ushll2     v3.4s, v0.8h, #0               \n" /* Y */                     \
+    "ushll      v0.4s, v0.4h, #0               \n"                             \
+    "mul        v3.4s, v3.4s, v31.4s           \n"                             \
+    "mul        v0.4s, v0.4s, v31.4s           \n"                             \
+    "sqshrun    v0.4h, v0.4s, #16              \n"                             \
+    "sqshrun2   v0.8h, v3.4s, #16              \n" /* Y */                     \
+    "uaddw      v1.8h, v2.8h, v1.8b            \n" /* Replicate UV */          \
+    "mov        v2.d[0], v1.d[1]               \n" /* Extract V */             \
+    "uxtl       v2.8h, v2.8b                   \n"                             \
+    "uxtl       v1.8h, v1.8b                   \n" /* Extract U */             \
+    "mul        v3.8h, v1.8h, v27.8h           \n"                             \
+    "mul        v5.8h, v1.8h, v29.8h           \n"                             \
+    "mul        v6.8h, v2.8h, v30.8h           \n"                             \
+    "mul        v7.8h, v2.8h, v28.8h           \n"                             \
+    "sqadd      v6.8h, v6.8h, v5.8h            \n"                             \
+    "sqadd      " #vB ".8h, v24.8h, v0.8h      \n" /* B */                     \
+    "sqadd      " #vG ".8h, v25.8h, v0.8h      \n" /* G */                     \
+    "sqadd      " #vR ".8h, v26.8h, v0.8h      \n" /* R */                     \
+    "sqadd      " #vB ".8h, " #vB ".8h, v3.8h  \n" /* B */                     \
+    "sqsub      " #vG ".8h, " #vG ".8h, v6.8h  \n" /* G */                     \
+    "sqadd      " #vR ".8h, " #vR ".8h, v7.8h  \n" /* R */                     \
+    "sqshrun    " #vB ".8b, " #vB ".8h, #6     \n" /* B */                     \
+    "sqshrun    " #vG ".8b, " #vG ".8h, #6     \n" /* G */                     \
+    "sqshrun    " #vR ".8b, " #vR ".8h, #6     \n" /* R */                     \
+
+// YUV to RGB conversion constants.
+// Y contribution to R,G,B.  Scale and bias.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB 1160 /* 1.164 * 64 * 16 - adjusted for even error distribution */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* -min(128, round(2.018 * 64)) */
+#define UG 25 /* -round(-0.391 * 64) */
+#define VG 52 /* -round(-0.813 * 64) */
+#define VR -102 /* -round(1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            - YGB)
+#define BG (UG * 128 + VG * 128 - YGB)
+#define BR            (VR * 128 - YGB)
+
+static vec16 kUVBiasBGR = { BB, BG, BR, 0, 0, 0, 0, 0 };
+static vec32 kYToRgb = { 0x0101 * YG, 0, 0, 0 };
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+#define RGBTOUV_SETUP_REG                                                      \
+    "movi       v20.8h, #56, lsl #0  \n"  /* UB/VR coefficient (0.875) / 2 */  \
+    "movi       v21.8h, #37, lsl #0  \n"  /* UG coefficient (-0.5781) / 2  */  \
+    "movi       v22.8h, #19, lsl #0  \n"  /* UR coefficient (-0.2969) / 2  */  \
+    "movi       v23.8h, #9,  lsl #0  \n"  /* VB coefficient (-0.1406) / 2  */  \
+    "movi       v24.8h, #47, lsl #0  \n"  /* VG coefficient (-0.7344) / 2  */  \
+    "movi       v25.16b, #0x80       \n"  /* 128.5 (0x8080 in 16-bit)      */
+
+
+#ifdef HAS_I444TOARGBROW_NEON
+void I444ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV444
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                 \n"
+    "movi       v23.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32 \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I444TOARGBROW_NEON
+
+#ifdef HAS_I422TOARGBROW_NEON
+void I422ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v23.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TOARGBROW_NEON
+
+#ifdef HAS_I411TOARGBROW_NEON
+void I411ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV411
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v23.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_argb),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I411TOARGBROW_NEON
+
+#ifdef HAS_I422TOBGRAROW_NEON
+void I422ToBGRARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_bgra,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v21, v22, v23)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v20.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_bgra),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TOBGRAROW_NEON
+
+#ifdef HAS_I422TOABGRROW_NEON
+void I422ToABGRRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_abgr,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v20, v21, v22)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v23.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_abgr),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TOABGRROW_NEON
+
+#ifdef HAS_I422TORGBAROW_NEON
+void I422ToRGBARow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_rgba,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v23, v22, v21)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v20.8b, #255                   \n" /* A */
+    MEMACCESS(3)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%3], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgba),  // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TORGBAROW_NEON
+
+#ifdef HAS_I422TORGB24ROW_NEON
+void I422ToRGB24Row_NEON(const uint8* src_y,
+                         const uint8* src_u,
+                         const uint8* src_v,
+                         uint8* dst_rgb24,
+                         int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    MEMACCESS(3)
+    "st3        {v20.8b,v21.8b,v22.8b}, [%3], #24     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_rgb24), // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TORGB24ROW_NEON
+
+#ifdef HAS_I422TORAWROW_NEON
+void I422ToRAWRow_NEON(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_raw,
+                       int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v20, v21, v22)
+    "subs       %w4, %w4, #8                   \n"
+    MEMACCESS(3)
+    "st3        {v20.8b,v21.8b,v22.8b}, [%3], #24     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_u),     // %1
+      "+r"(src_v),     // %2
+      "+r"(dst_raw),   // %3
+      "+r"(width)      // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TORAWROW_NEON
+
+#define ARGBTORGB565                                                           \
+    "shll       v0.8h,  v22.8b, #8             \n"  /* R                    */ \
+    "shll       v20.8h, v20.8b, #8             \n"  /* B                    */ \
+    "shll       v21.8h, v21.8b, #8             \n"  /* G                    */ \
+    "sri        v0.8h,  v21.8h, #5             \n"  /* RG                   */ \
+    "sri        v0.8h,  v20.8h, #11            \n"  /* RGB                  */
+
+#ifdef HAS_I422TORGB565ROW_NEON
+void I422ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_u,
+                          const uint8* src_v,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    ARGBTORGB565
+    MEMACCESS(3)
+    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_rgb565),  // %3
+      "+r"(width)     // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TORGB565ROW_NEON
+
+#define ARGBTOARGB1555                                                         \
+    "shll       v0.8h,  v23.8b, #8             \n"  /* A                    */ \
+    "shll       v22.8h, v22.8b, #8             \n"  /* R                    */ \
+    "shll       v20.8h, v20.8b, #8             \n"  /* B                    */ \
+    "shll       v21.8h, v21.8b, #8             \n"  /* G                    */ \
+    "sri        v0.8h,  v22.8h, #1             \n"  /* AR                   */ \
+    "sri        v0.8h,  v21.8h, #6             \n"  /* ARG                  */ \
+    "sri        v0.8h,  v20.8h, #11            \n"  /* ARGB                 */
+
+#ifdef HAS_I422TOARGB1555ROW_NEON
+void I422ToARGB1555Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb1555,
+                            int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    ARGBTOARGB1555
+    MEMACCESS(3)
+    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb1555),  // %3
+      "+r"(width)     // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TOARGB1555ROW_NEON
+
+#define ARGBTOARGB4444                                                         \
+    /* Input v20.8b<=B, v21.8b<=G, v22.8b<=R, v23.8b<=A, v4.8b<=0x0f        */ \
+    "ushr       v20.8b, v20.8b, #4             \n"  /* B                    */ \
+    "bic        v21.8b, v21.8b, v4.8b          \n"  /* G                    */ \
+    "ushr       v22.8b, v22.8b, #4             \n"  /* R                    */ \
+    "bic        v23.8b, v23.8b, v4.8b          \n"  /* A                    */ \
+    "orr        v0.8b,  v20.8b, v21.8b         \n"  /* BG                   */ \
+    "orr        v1.8b,  v22.8b, v23.8b         \n"  /* RA                   */ \
+    "zip1       v0.16b, v0.16b, v1.16b         \n"  /* BGRA                 */
+
+#ifdef HAS_I422TOARGB4444ROW_NEON
+void I422ToARGB4444Row_NEON(const uint8* src_y,
+                            const uint8* src_u,
+                            const uint8* src_v,
+                            uint8* dst_argb4444,
+                            int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+    "movi       v4.16b, #0x0f                  \n"  // bits to clear with vbic.
+  "1:                                          \n"
+    READYUV422
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w4, %w4, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    ARGBTOARGB4444
+    MEMACCESS(3)
+    "st1        {v0.8h}, [%3], #16             \n"  // store 8 pixels ARGB4444.
+    "b.gt       1b                             \n"
+    : "+r"(src_y),    // %0
+      "+r"(src_u),    // %1
+      "+r"(src_v),    // %2
+      "+r"(dst_argb4444),  // %3
+      "+r"(width)     // %4
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I422TOARGB4444ROW_NEON
+
+#ifdef HAS_I400TOARGBROW_NEON
+void I400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  int64 width64 = (int64)(width);
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUV400
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w2, %w2, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    MEMACCESS(1)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width64)    // %2
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_I400TOARGBROW_NEON
+
+#ifdef HAS_J400TOARGBROW_NEON
+void J400ToARGBRow_NEON(const uint8* src_y,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    "movi       v23.8b, #255                   \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v20.8b}, [%0], #8             \n"
+    "orr        v21.8b, v20.8b, v20.8b         \n"
+    "orr        v22.8b, v20.8b, v20.8b         \n"
+    "subs       %w2, %w2, #8                   \n"
+    MEMACCESS(1)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width)      // %2
+    :
+    : "cc", "memory", "v20", "v21", "v22", "v23"
+  );
+}
+#endif  // HAS_J400TOARGBROW_NEON
+
+#ifdef HAS_NV12TOARGBROW_NEON
+void NV12ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w3, %w3, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    MEMACCESS(2)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_NV12TOARGBROW_NEON
+
+#ifdef HAS_NV21TOARGBROW_NEON
+void NV21ToARGBRow_NEON(const uint8* src_y,
+                        const uint8* src_uv,
+                        uint8* dst_argb,
+                        int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w3, %w3, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    MEMACCESS(2)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%2], #32     \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_argb),  // %2
+      "+r"(width)      // %3
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_NV21TOARGBROW_NEON
+
+#ifdef HAS_NV12TORGB565ROW_NEON
+void NV12ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READNV12
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w3, %w3, #8                   \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "st1        {v0.8h}, [%2], 16              \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_NV12TORGB565ROW_NEON
+
+#ifdef HAS_NV21TORGB565ROW_NEON
+void NV21ToRGB565Row_NEON(const uint8* src_y,
+                          const uint8* src_uv,
+                          uint8* dst_rgb565,
+                          int width) {
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READNV21
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w3, %w3, #8                   \n"
+    ARGBTORGB565
+    MEMACCESS(2)
+    "st1        {v0.8h}, [%2], 16              \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+    : "+r"(src_y),     // %0
+      "+r"(src_uv),    // %1
+      "+r"(dst_rgb565),  // %2
+      "+r"(width)      // %3
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_NV21TORGB565ROW_NEON
+
+#ifdef HAS_YUY2TOARGBROW_NEON
+void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
+                        uint8* dst_argb,
+                        int width) {
+  int64 width64 = (int64)(width);
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READYUY2
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w2, %w2, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    MEMACCESS(1)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32      \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_yuy2),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width64)    // %2
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_YUY2TOARGBROW_NEON
+
+#ifdef HAS_UYVYTOARGBROW_NEON
+void UYVYToARGBRow_NEON(const uint8* src_uyvy,
+                        uint8* dst_argb,
+                        int width) {
+  int64 width64 = (int64)(width);
+  asm volatile (
+    YUV422TORGB_SETUP_REG
+  "1:                                          \n"
+    READUYVY
+    YUV422TORGB(v22, v21, v20)
+    "subs       %w2, %w2, #8                   \n"
+    "movi       v23.8b, #255                   \n"
+    MEMACCESS(1)
+    "st4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], 32      \n"
+    "b.gt       1b                             \n"
+    : "+r"(src_uyvy),  // %0
+      "+r"(dst_argb),  // %1
+      "+r"(width64)    // %2
+    : [kUVBiasBGR]"r"(&kUVBiasBGR),
+      [kYToRgb]"r"(&kYToRgb)
+    : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20",
+      "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_UYVYTOARGBROW_NEON
+
+// Reads 16 pairs of UV and write even values to dst_u and odd to dst_v.
+#ifdef HAS_SPLITUVROW_NEON
+void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                     int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pairs of UV
+    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"  // store U
+    MEMACCESS(2)
+    "st1        {v1.16b}, [%2], #16            \n"  // store V
+    "b.gt       1b                             \n"
+    : "+r"(src_uv),  // %0
+      "+r"(dst_u),   // %1
+      "+r"(dst_v),   // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+#endif  // HAS_SPLITUVROW_NEON
+
+// Reads 16 U's and V's and writes out 16 pairs of UV.
+#ifdef HAS_MERGEUVROW_NEON
+void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load U
+    MEMACCESS(1)
+    "ld1        {v1.16b}, [%1], #16            \n"  // load V
+    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
+    MEMACCESS(2)
+    "st2        {v0.16b,v1.16b}, [%2], #32     \n"  // store 16 pairs of UV
+    "b.gt       1b                             \n"
+    :
+      "+r"(src_u),   // %0
+      "+r"(src_v),   // %1
+      "+r"(dst_uv),  // %2
+      "+r"(width)    // %3  // Output registers
+    :                       // Input registers
+    : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+#endif  // HAS_MERGEUVROW_NEON
+
+// Copy multiple of 32.  vld4.8  allow unaligned and is fastest on a15.
+#ifdef HAS_COPYROW_NEON
+void CopyRow_NEON(const uint8* src, uint8* dst, int count) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32       \n"  // load 32
+    "subs       %w2, %w2, #32                  \n"  // 32 processed per loop
+    MEMACCESS(1)
+    "st1        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32       \n"  // store 32
+    "b.gt       1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(count)  // %2  // Output registers
+  :                     // Input registers
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_COPYROW_NEON
+
+// SetRow writes 'count' bytes using an 8 bit value repeated.
+void SetRow_NEON(uint8* dst, uint8 v8, int count) {
+  asm volatile (
+    "dup        v0.16b, %w2                    \n"  // duplicate 16 bytes
+  "1:                                          \n"
+    "subs      %w1, %w1, #16                   \n"  // 16 bytes per loop
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"  // store
+    "b.gt      1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v8)      // %2
+  : "cc", "memory", "v0"
+  );
+}
+
+void ARGBSetRow_NEON(uint8* dst, uint32 v32, int count) {
+  asm volatile (
+    "dup        v0.4s, %w2                     \n"  // duplicate 4 ints
+  "1:                                          \n"
+    "subs      %w1, %w1, #4                    \n"  // 4 ints per loop
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"  // store
+    "b.gt      1b                              \n"
+  : "+r"(dst),   // %0
+    "+r"(count)  // %1
+  : "r"(v32)     // %2
+  : "cc", "memory", "v0"
+  );
+}
+
+#ifdef HAS_MIRRORROW_NEON
+void MirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  int64 width64 = (int64) width;
+  asm volatile (
+    // Start at end of source row.
+    "add        %0, %0, %2                     \n"
+    "sub        %0, %0, #16                    \n"
+
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], %3             \n"  // src -= 16
+    "subs       %2, %2, #16                   \n"  // 16 pixels per loop.
+    "rev64      v0.16b, v0.16b                 \n"
+    MEMACCESS(1)
+    "st1        {v0.D}[1], [%1], #8            \n"  // dst += 16
+    MEMACCESS(1)
+    "st1        {v0.D}[0], [%1], #8            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width64)  // %2
+  : "r"((ptrdiff_t)-16)    // %3
+  : "cc", "memory", "v0"
+  );
+}
+#endif  // HAS_MIRRORROW_NEON
+
+#ifdef HAS_MIRRORUVROW_NEON
+void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
+                      int width) {
+  int64 width64 = (int64) width;
+  asm volatile (
+    // Start at end of source row.
+    "add        %0, %0, %3, lsl #1             \n"
+    "sub        %0, %0, #16                    \n"
+
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v0.8b, v1.8b}, [%0], %4       \n"  // src -= 16
+    "subs       %3, %3, #8                     \n"  // 8 pixels per loop.
+    "rev64      v0.8b, v0.8b                   \n"
+    "rev64      v1.8b, v1.8b                   \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // dst += 8
+    MEMACCESS(2)
+    "st1        {v1.8b}, [%2], #8              \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_uv),  // %0
+    "+r"(dst_u),   // %1
+    "+r"(dst_v),   // %2
+    "+r"(width64)    // %3
+  : "r"((ptrdiff_t)-16)      // %4
+  : "cc", "memory", "v0", "v1"
+  );
+}
+#endif  // HAS_MIRRORUVROW_NEON
+
+#ifdef HAS_ARGBMIRRORROW_NEON
+void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width) {
+  int64 width64 = (int64) width;
+  asm volatile (
+    // Start at end of source row.
+    "add        %0, %0, %2, lsl #2             \n"
+    "sub        %0, %0, #16                    \n"
+
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], %3             \n"  // src -= 16
+    "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
+    "rev64      v0.4s, v0.4s                   \n"
+    MEMACCESS(1)
+    "st1        {v0.D}[1], [%1], #8            \n"  // dst += 16
+    MEMACCESS(1)
+    "st1        {v0.D}[0], [%1], #8            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src),   // %0
+    "+r"(dst),   // %1
+    "+r"(width64)  // %2
+  : "r"((ptrdiff_t)-16)    // %3
+  : "cc", "memory", "v0"
+  );
+}
+#endif  // HAS_ARGBMIRRORROW_NEON
+
+#ifdef HAS_RGB24TOARGBROW_NEON
+void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  asm volatile (
+    "movi       v4.8b, #255                    \n"  // Alpha
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v1.8b,v2.8b,v3.8b}, [%0], #24 \n"  // load 8 pixels of RGB24.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "st4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_argb),   // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "v1", "v2", "v3", "v4"  // Clobber List
+  );
+}
+#endif  // HAS_RGB24TOARGBROW_NEON
+
+#ifdef HAS_RAWTOARGBROW_NEON
+void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix) {
+  asm volatile (
+    "movi       v5.8b, #255                    \n"  // Alpha
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // read r g b
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "orr        v3.8b, v1.8b, v1.8b            \n"  // move g
+    "orr        v4.8b, v0.8b, v0.8b            \n"  // move r
+    MEMACCESS(1)
+    "st4        {v2.8b,v3.8b,v4.8b,v5.8b}, [%1], #32 \n"  // store b g r a
+    "b.gt       1b                             \n"
+  : "+r"(src_raw),   // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5"  // Clobber List
+  );
+}
+#endif  // HAS_RAWTOARGBROW_NEON
+
+#define RGB565TOARGB                                                           \
+    "shrn       v6.8b, v0.8h, #5               \n"  /* G xxGGGGGG           */ \
+    "shl        v6.8b, v6.8b, #2               \n"  /* G GGGGGG00 upper 6   */ \
+    "ushr       v4.8b, v6.8b, #6               \n"  /* G 000000GG lower 2   */ \
+    "orr        v1.8b, v4.8b, v6.8b            \n"  /* G                    */ \
+    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
+    "ushr       v0.8h, v0.8h, #11              \n"  /* R 000RRRRR           */ \
+    "xtn2       v2.16b,v0.8h                   \n"  /* R in upper part      */ \
+    "shl        v2.16b, v2.16b, #3             \n"  /* R,B BBBBB000 upper 5 */ \
+    "ushr       v0.16b, v2.16b, #5             \n"  /* R,B 00000BBB lower 3 */ \
+    "orr        v0.16b, v0.16b, v2.16b         \n"  /* R,B                  */ \
+    "dup        v2.2D, v0.D[1]                 \n"  /* R                    */
+
+#ifdef HAS_RGB565TOARGBROW_NEON
+void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix) {
+  asm volatile (
+    "movi       v3.8b, #255                    \n"  // Alpha
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    RGB565TOARGB
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6"  // Clobber List
+  );
+}
+#endif  // HAS_RGB565TOARGBROW_NEON
+
+#define ARGB1555TOARGB                                                         \
+    "ushr       v2.8h, v0.8h, #10              \n"  /* R xxxRRRRR           */ \
+    "shl        v2.8h, v2.8h, #3               \n"  /* R RRRRR000 upper 5   */ \
+    "xtn        v3.8b, v2.8h                   \n"  /* RRRRR000 AAAAAAAA    */ \
+                                                                               \
+    "sshr       v2.8h, v0.8h, #15              \n"  /* A AAAAAAAA           */ \
+    "xtn2       v3.16b, v2.8h                  \n"                             \
+                                                                               \
+    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
+    "shrn2      v2.16b,v0.8h, #5               \n"  /* G xxxGGGGG           */ \
+                                                                               \
+    "ushr       v1.16b, v3.16b, #5             \n"  /* R,A 00000RRR lower 3 */ \
+    "shl        v0.16b, v2.16b, #3             \n"  /* B,G BBBBB000 upper 5 */ \
+    "ushr       v2.16b, v0.16b, #5             \n"  /* B,G 00000BBB lower 3 */ \
+                                                                               \
+    "orr        v0.16b, v0.16b, v2.16b         \n"  /* B,G                  */ \
+    "orr        v2.16b, v1.16b, v3.16b         \n"  /* R,A                  */ \
+    "dup        v1.2D, v0.D[1]                 \n"                             \
+    "dup        v3.2D, v2.D[1]                 \n"
+
+// RGB555TOARGB is same as ARGB1555TOARGB but ignores alpha.
+#define RGB555TOARGB                                                           \
+    "ushr       v2.8h, v0.8h, #10              \n"  /* R xxxRRRRR           */ \
+    "shl        v2.8h, v2.8h, #3               \n"  /* R RRRRR000 upper 5   */ \
+    "xtn        v3.8b, v2.8h                   \n"  /* RRRRR000             */ \
+                                                                               \
+    "xtn        v2.8b, v0.8h                   \n"  /* B xxxBBBBB           */ \
+    "shrn2      v2.16b,v0.8h, #5               \n"  /* G xxxGGGGG           */ \
+                                                                               \
+    "ushr       v1.16b, v3.16b, #5             \n"  /* R   00000RRR lower 3 */ \
+    "shl        v0.16b, v2.16b, #3             \n"  /* B,G BBBBB000 upper 5 */ \
+    "ushr       v2.16b, v0.16b, #5             \n"  /* B,G 00000BBB lower 3 */ \
+                                                                               \
+    "orr        v0.16b, v0.16b, v2.16b         \n"  /* B,G                  */ \
+    "orr        v2.16b, v1.16b, v3.16b         \n"  /* R                    */ \
+    "dup        v1.2D, v0.D[1]                 \n"  /* G */                    \
+
+#ifdef HAS_ARGB1555TOARGBROW_NEON
+void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+    "movi       v3.8b, #255                    \n"  // Alpha
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_ARGB1555TOARGBROW_NEON
+
+#define ARGB4444TOARGB                                                         \
+    "shrn       v1.8b,  v0.8h, #8              \n"  /* v1(l) AR             */ \
+    "xtn2       v1.16b, v0.8h                  \n"  /* v1(h) GB             */ \
+    "shl        v2.16b, v1.16b, #4             \n"  /* B,R BBBB0000         */ \
+    "ushr       v3.16b, v1.16b, #4             \n"  /* G,A 0000GGGG         */ \
+    "ushr       v0.16b, v2.16b, #4             \n"  /* B,R 0000BBBB         */ \
+    "shl        v1.16b, v3.16b, #4             \n"  /* G,A GGGG0000         */ \
+    "orr        v2.16b, v0.16b, v2.16b         \n"  /* B,R BBBBBBBB         */ \
+    "orr        v3.16b, v1.16b, v3.16b         \n"  /* G,A GGGGGGGG         */ \
+    "dup        v0.2D, v2.D[1]                 \n"                             \
+    "dup        v1.2D, v3.D[1]                 \n"
+
+#ifdef HAS_ARGB4444TOARGBROW_NEON
+void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_argb),    // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4"  // Clobber List
+  );
+}
+#endif  // HAS_ARGB4444TOARGBROW_NEON
+
+#ifdef HAS_ARGBTORGB24ROW_NEON
+void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb24, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load 8 ARGB pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    MEMACCESS(1)
+    "st3        {v1.8b,v2.8b,v3.8b}, [%1], #24 \n"  // store 8 pixels of RGB24.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_rgb24),  // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "v1", "v2", "v3", "v4"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBTORGB24ROW_NEON
+
+#ifdef HAS_ARGBTORAWROW_NEON
+void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_raw, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v1.8b,v2.8b,v3.8b,v4.8b}, [%0], #32 \n"  // load b g r a
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "orr        v4.8b, v2.8b, v2.8b            \n"  // mov g
+    "orr        v5.8b, v1.8b, v1.8b            \n"  // mov b
+    MEMACCESS(1)
+    "st3        {v3.8b,v4.8b,v5.8b}, [%1], #24 \n"  // store r g b
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_raw),   // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v1", "v2", "v3", "v4", "v5"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBTORAWROW_NEON
+
+#ifdef HAS_YUY2TOYROW_NEON
+void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pixels of YUY2.
+    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"  // store 16 pixels of Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+#endif  // HAS_YUY2TOYROW_NEON
+
+#ifdef HAS_UYVYTOYROW_NEON
+void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"  // load 16 pixels of UYVY.
+    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop.
+    MEMACCESS(1)
+    "st1        {v1.16b}, [%1], #16            \n"  // store 16 pixels of Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1"  // Clobber List
+  );
+}
+#endif  // HAS_UYVYTOYROW_NEON
+
+#ifdef HAS_YUY2TOUV422ROW_NEON
+void YUY2ToUV422Row_NEON(const uint8* src_yuy2, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 YUY2 pixels
+    "subs       %w3, %w3, #16                  \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "st1        {v1.8b}, [%1], #8              \n"  // store 8 U.
+    MEMACCESS(2)
+    "st1        {v3.8b}, [%2], #8              \n"  // store 8 V.
+    "b.gt       1b                             \n"
+  : "+r"(src_yuy2),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_YUY2TOUV422ROW_NEON
+
+#ifdef HAS_UYVYTOUV422ROW_NEON
+void UYVYToUV422Row_NEON(const uint8* src_uyvy, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 UYVY pixels
+    "subs       %w3, %w3, #16                  \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 U.
+    MEMACCESS(2)
+    "st1        {v2.8b}, [%2], #8              \n"  // store 8 V.
+    "b.gt       1b                             \n"
+  : "+r"(src_uyvy),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_UYVYTOUV422ROW_NEON
+
+#ifdef HAS_YUY2TOUVROW_NEON
+void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_yuy2b = src_yuy2 + stride_yuy2;
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
+    "subs       %w4, %w4, #16                  \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
+    "urhadd     v1.8b, v1.8b, v5.8b            \n"  // average rows of U
+    "urhadd     v3.8b, v3.8b, v7.8b            \n"  // average rows of V
+    MEMACCESS(2)
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 U.
+    MEMACCESS(3)
+    "st1        {v3.8b}, [%3], #8              \n"  // store 8 V.
+    "b.gt       1b                             \n"
+  : "+r"(src_yuy2),     // %0
+    "+r"(src_yuy2b),    // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
+    "v5", "v6", "v7"  // Clobber List
+  );
+}
+#endif  // HAS_YUY2TOUVROW_NEON
+
+#ifdef HAS_UYVYTOUVROW_NEON
+void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_uyvyb = src_uyvy + stride_uyvy;
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 16 pixels
+    "subs       %w4, %w4, #16                  \n"  // 16 pixels = 8 UVs.
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load next row
+    "urhadd     v0.8b, v0.8b, v4.8b            \n"  // average rows of U
+    "urhadd     v2.8b, v2.8b, v6.8b            \n"  // average rows of V
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 U.
+    MEMACCESS(3)
+    "st1        {v2.8b}, [%3], #8              \n"  // store 8 V.
+    "b.gt       1b                             \n"
+  : "+r"(src_uyvy),     // %0
+    "+r"(src_uyvyb),    // %1
+    "+r"(dst_u),        // %2
+    "+r"(dst_v),        // %3
+    "+r"(pix)           // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
+    "v5", "v6", "v7"  // Clobber List
+  );
+}
+#endif  // HAS_UYVYTOUVROW_NEON
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+#ifdef HAS_ARGBSHUFFLEROW_NEON
+void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  asm volatile (
+    MEMACCESS(3)
+    "ld1        {v2.16b}, [%3]                 \n"  // shuffler
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 4 pixels.
+    "subs       %w2, %w2, #4                   \n"  // 4 processed per loop
+    "tbl        v1.16b, {v0.16b}, v2.16b       \n"  // look up 4 pixels
+    MEMACCESS(1)
+    "st1        {v1.16b}, [%1], #16            \n"  // store 4.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(pix)        // %2
+  : "r"(shuffler)    // %3
+  : "cc", "memory", "v0", "v1", "v2"  // Clobber List
+  );
+}
+#endif  // HAS_ARGBSHUFFLEROW_NEON
+
+#ifdef HAS_I422TOYUY2ROW_NEON
+void I422ToYUY2Row_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_yuy2, int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v0.8b, v1.8b}, [%0], #16      \n"  // load 16 Ys
+    "orr        v2.8b, v1.8b, v1.8b            \n"
+    MEMACCESS(1)
+    "ld1        {v1.8b}, [%1], #8              \n"  // load 8 Us
+    MEMACCESS(2)
+    "ld1        {v3.8b}, [%2], #8              \n"  // load 8 Vs
+    "subs       %w4, %w4, #16                  \n"  // 16 pixels
+    MEMACCESS(3)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_yuy2),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"
+  );
+}
+#endif  // HAS_I422TOYUY2ROW_NEON
+
+#ifdef HAS_I422TOUYVYROW_NEON
+void I422ToUYVYRow_NEON(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_uyvy, int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld2        {v1.8b,v2.8b}, [%0], #16       \n"  // load 16 Ys
+    "orr        v3.8b, v2.8b, v2.8b            \n"
+    MEMACCESS(1)
+    "ld1        {v0.8b}, [%1], #8              \n"  // load 8 Us
+    MEMACCESS(2)
+    "ld1        {v2.8b}, [%2], #8              \n"  // load 8 Vs
+    "subs       %w4, %w4, #16                  \n"  // 16 pixels
+    MEMACCESS(3)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%3], #32 \n"  // Store 16 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(src_y),     // %0
+    "+r"(src_u),     // %1
+    "+r"(src_v),     // %2
+    "+r"(dst_uyvy),  // %3
+    "+r"(width)      // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"
+  );
+}
+#endif  // HAS_I422TOUYVYROW_NEON
+
+#ifdef HAS_ARGBTORGB565ROW_NEON
+void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb565, int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGBTORGB565
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_rgb565),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v20", "v21", "v22", "v23"
+  );
+}
+#endif  // HAS_ARGBTORGB565ROW_NEON
+
+#ifdef HAS_ARGBTORGB565DITHERROW_NEON
+void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int width) {
+  asm volatile (
+    "dup        v1.4s, %w2                     \n"  // dither4
+  "1:                                          \n"
+    MEMACCESS(1)
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%1], #32 \n"  // load 8 pixels
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uqadd      v20.8b, v20.8b, v1.8b          \n"
+    "uqadd      v21.8b, v21.8b, v1.8b          \n"
+    "uqadd      v22.8b, v22.8b, v1.8b          \n"
+    ARGBTORGB565
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"  // store 8 pixels RGB565.
+    "b.gt       1b                             \n"
+  : "+r"(dst_rgb)    // %0
+  : "r"(src_argb),   // %1
+    "r"(dither4),    // %2
+    "r"(width)       // %3
+  : "cc", "memory", "v0", "v1", "v20", "v21", "v22", "v23"
+  );
+}
+#endif  // HAS_ARGBTORGB565ROW_NEON
+
+#ifdef HAS_ARGBTOARGB1555ROW_NEON
+void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_argb1555,
+                            int pix) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGBTOARGB1555
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels ARGB1555.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb1555),  // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v20", "v21", "v22", "v23"
+  );
+}
+#endif  // HAS_ARGBTOARGB1555ROW_NEON
+
+#ifdef HAS_ARGBTOARGB4444ROW_NEON
+void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
+                            int pix) {
+  asm volatile (
+    "movi       v4.16b, #0x0f                  \n"  // bits to clear with vbic.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v20.8b,v21.8b,v22.8b,v23.8b}, [%0], #32 \n"  // load 8 pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGBTOARGB4444
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"  // store 8 pixels ARGB4444.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),      // %0
+    "+r"(dst_argb4444),  // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "v0", "v1", "v4", "v20", "v21", "v22", "v23"
+  );
+}
+#endif  // HAS_ARGBTOARGB4444ROW_NEON
+
+#ifdef HAS_ARGBTOYROW_NEON
+void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
+    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
+    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGBTOYROW_NEON
+
+#ifdef HAS_ARGBTOYJROW_NEON
+void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #15                     \n"  // B * 0.11400 coefficient
+    "movi       v5.8b, #75                     \n"  // G * 0.58700 coefficient
+    "movi       v6.8b, #38                     \n"  // R * 0.29900 coefficient
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
+    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
+    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 15 bit to 8 bit Y
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
+  );
+}
+#endif  // HAS_ARGBTOYJROW_NEON
+
+// 8x1 pixels.
+#ifdef HAS_ARGBTOUV444ROW_NEON
+void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    "movi       v24.8b, #112                   \n"  // UB / VR 0.875 coefficient
+    "movi       v25.8b, #74                    \n"  // UG -0.5781 coefficient
+    "movi       v26.8b, #38                    \n"  // UR -0.2969 coefficient
+    "movi       v27.8b, #18                    \n"  // VB -0.1406 coefficient
+    "movi       v28.8b, #94                    \n"  // VG -0.7344 coefficient
+    "movi       v29.16b,#0x80                  \n"  // 128.5
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "umull      v4.8h, v0.8b, v24.8b           \n"  // B
+    "umlsl      v4.8h, v1.8b, v25.8b           \n"  // G
+    "umlsl      v4.8h, v2.8b, v26.8b           \n"  // R
+    "add        v4.8h, v4.8h, v29.8h           \n"  // +128 -> unsigned
+
+    "umull      v3.8h, v2.8b, v24.8b           \n"  // R
+    "umlsl      v3.8h, v1.8b, v28.8b           \n"  // G
+    "umlsl      v3.8h, v0.8b, v27.8b           \n"  // B
+    "add        v3.8h, v3.8h, v29.8h           \n"  // +128 -> unsigned
+
+    "uqshrn     v0.8b, v4.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4",
+    "v24", "v25", "v26", "v27", "v28", "v29"
+  );
+}
+#endif  // HAS_ARGBTOUV444ROW_NEON
+
+// 16x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_ARGBTOUV422ROW_NEON
+void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop.
+    "mul        v3.8h, v0.8h, v20.8h           \n"  // B
+    "mls        v3.8h, v1.8h, v21.8h           \n"  // G
+    "mls        v3.8h, v2.8h, v22.8h           \n"  // R
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+
+    "mul        v4.8h, v2.8h, v20.8h           \n"  // R
+    "mls        v4.8h, v1.8h, v24.8h           \n"  // G
+    "mls        v4.8h, v0.8h, v23.8h           \n"  // B
+    "add        v4.8h, v4.8h, v25.8h           \n"  // +128 -> unsigned
+
+    "uqshrn     v0.8b, v3.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v4.8h, #8               \n"  // 16 bit to 8 bit V
+
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ARGBTOUV422ROW_NEON
+
+// 32x1 pixels -> 8x1.  pix is number of argb pixels. e.g. 32.
+#ifdef HAS_ARGBTOUV411ROW_NEON
+void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
+                         int pix) {
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(0)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%0], #64 \n"  // load next 16.
+    "uaddlp     v4.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v5.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v6.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "addp       v0.8h, v0.8h, v4.8h            \n"  // B 16 shorts -> 8 shorts.
+    "addp       v1.8h, v1.8h, v5.8h            \n"  // G 16 shorts -> 8 shorts.
+    "addp       v2.8h, v2.8h, v6.8h            \n"  // R 16 shorts -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w3, %w3, #32                  \n"  // 32 processed per loop.
+    "mul        v3.8h, v0.8h, v20.8h           \n"  // B
+    "mls        v3.8h, v1.8h, v21.8h           \n"  // G
+    "mls        v3.8h, v2.8h, v22.8h           \n"  // R
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+    "mul        v4.8h, v2.8h, v20.8h           \n"  // R
+    "mls        v4.8h, v1.8h, v24.8h           \n"  // G
+    "mls        v4.8h, v0.8h, v23.8h           \n"  // B
+    "add        v4.8h, v4.8h, v25.8h           \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v3.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v4.8h, #8               \n"  // 16 bit to 8 bit V
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels U.
+    MEMACCESS(2)
+    "st1        {v1.8b}, [%2], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_u),     // %1
+    "+r"(dst_v),     // %2
+    "+r"(pix)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ARGBTOUV411ROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#define RGBTOUV(QB, QG, QR) \
+    "mul        v3.8h, " #QB ",v20.8h          \n"  /* B                    */ \
+    "mul        v4.8h, " #QR ",v20.8h          \n"  /* R                    */ \
+    "mls        v3.8h, " #QG ",v21.8h          \n"  /* G                    */ \
+    "mls        v4.8h, " #QG ",v24.8h          \n"  /* G                    */ \
+    "mls        v3.8h, " #QR ",v22.8h          \n"  /* R                    */ \
+    "mls        v4.8h, " #QB ",v23.8h          \n"  /* B                    */ \
+    "add        v3.8h, v3.8h, v25.8h           \n"  /* +128 -> unsigned     */ \
+    "add        v4.8h, v4.8h, v25.8h           \n"  /* +128 -> unsigned     */ \
+    "uqshrn     v0.8b, v3.8h, #8               \n"  /* 16 bit to 8 bit U    */ \
+    "uqshrn     v1.8b, v4.8h, #8               \n"  /* 16 bit to 8 bit V    */
+
+// TODO(fbarchard): Consider vhadd vertical, then vpaddl horizontal, avoid shr.
+// TODO(fbarchard): consider ptrdiff_t for all strides.
+
+#ifdef HAS_ARGBTOUVROW_NEON
+void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb_1 = src_argb + src_stride_argb;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    MEMACCESS(1)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load next 16
+    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_argb_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ARGBTOUVROW_NEON
+
+// TODO(fbarchard): Subsample match C code.
+#ifdef HAS_ARGBTOUVJROW_NEON
+void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb_1 = src_argb + src_stride_argb;
+  asm volatile (
+    "movi       v20.8h, #63, lsl #0            \n"  // UB/VR coeff (0.500) / 2
+    "movi       v21.8h, #42, lsl #0            \n"  // UG coeff (-0.33126) / 2
+    "movi       v22.8h, #21, lsl #0            \n"  // UR coeff (-0.16874) / 2
+    "movi       v23.8h, #10, lsl #0            \n"  // VB coeff (-0.08131) / 2
+    "movi       v24.8h, #53, lsl #0            \n"  // VG coeff (-0.41869) / 2
+    "movi       v25.16b, #0x80                 \n"  // 128.5 (0x8080 in 16-bit)
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64  \n"  // load next 16
+    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(src_argb_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ARGBTOUVJROW_NEON
+
+#ifdef HAS_BGRATOUVROW_NEON
+void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_bgra_1 = src_bgra + src_stride_bgra;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v3.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v3.8h, v2.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v1.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more
+    "uadalp     v0.8h, v7.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v3.8h, v6.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v5.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v3.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(src_bgra_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_BGRATOUVROW_NEON
+
+#ifdef HAS_ABGRTOUVROW_NEON
+void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_abgr_1 = src_abgr + src_stride_abgr;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v3.8h, v2.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v0.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
+    "uadalp     v3.8h, v6.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v4.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v3.8h, #1               \n"  // 2x average
+    "urshr      v2.8h, v2.8h, #1               \n"
+    "urshr      v1.8h, v1.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v2.8h, v1.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(src_abgr_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ABGRTOUVROW_NEON
+
+#ifdef HAS_RGBATOUVROW_NEON
+void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_rgba_1 = src_rgba + src_stride_rgba;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v1.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v2.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v3.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld4        {v4.16b,v5.16b,v6.16b,v7.16b}, [%1], #64 \n"  // load 16 more.
+    "uadalp     v0.8h, v5.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v6.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v7.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(src_rgba_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_RGBATOUVROW_NEON
+
+#ifdef HAS_RGB24TOUVROW_NEON
+void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
+                       uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_rgb24_1 = src_rgb24 + src_stride_rgb24;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 16 pixels.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld3        {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 16 more.
+    "uadalp     v0.8h, v4.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v6.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v0.8h, v0.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v2.8h, v2.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v0.8h, v1.8h, v2.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(src_rgb24_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_RGB24TOUVROW_NEON
+
+#ifdef HAS_RAWTOUVROW_NEON
+void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
+                     uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_raw_1 = src_raw + src_stride_raw;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v0.16b,v1.16b,v2.16b}, [%0], #48 \n"  // load 8 RAW pixels.
+    "uaddlp     v2.8h, v2.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v0.8h, v0.16b                  \n"  // R 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "ld3        {v4.16b,v5.16b,v6.16b}, [%1], #48 \n"  // load 8 more RAW pixels
+    "uadalp     v2.8h, v6.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v5.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v0.8h, v4.16b                  \n"  // R 16 bytes -> 8 shorts.
+
+    "urshr      v2.8h, v2.8h, #1               \n"  // 2x average
+    "urshr      v1.8h, v1.8h, #1               \n"
+    "urshr      v0.8h, v0.8h, #1               \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 32 processed per loop.
+    RGBTOUV(v2.8h, v1.8h, v0.8h)
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(src_raw_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_RAWTOUVROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_RGB565TOUVROW_NEON
+void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_rgb565_1 = src_rgb565 + src_stride_rgb565;
+  asm volatile (
+    "movi       v22.8h, #56, lsl #0            \n"  // UB / VR coeff (0.875) / 2
+    "movi       v23.8h, #37, lsl #0            \n"  // UG coeff (-0.5781) / 2
+    "movi       v24.8h, #19, lsl #0            \n"  // UR coeff (-0.2969) / 2
+    "movi       v25.8h, #9 , lsl #0            \n"  // VB coeff (-0.1406) / 2
+    "movi       v26.8h, #47, lsl #0            \n"  // VG coeff (-0.7344) / 2
+    "movi       v27.16b, #0x80                 \n"  // 128.5 (0x8080 in 16-bit)
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v18.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v20.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "uaddlp     v17.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v19.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v21.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 RGB565 pixels.
+    RGB565TOARGB
+    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v18.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v20.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 RGB565 pixels.
+    RGB565TOARGB
+    "uadalp     v17.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v19.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v21.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    "ins        v16.D[1], v17.D[0]             \n"
+    "ins        v18.D[1], v19.D[0]             \n"
+    "ins        v20.D[1], v21.D[0]             \n"
+
+    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
+    "urshr      v5.8h, v18.8h, #1              \n"
+    "urshr      v6.8h, v20.8h, #1              \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
+    "mul        v16.8h, v4.8h, v22.8h          \n"  // B
+    "mls        v16.8h, v5.8h, v23.8h          \n"  // G
+    "mls        v16.8h, v6.8h, v24.8h          \n"  // R
+    "add        v16.8h, v16.8h, v27.8h         \n"  // +128 -> unsigned
+    "mul        v17.8h, v6.8h, v22.8h          \n"  // R
+    "mls        v17.8h, v5.8h, v26.8h          \n"  // G
+    "mls        v17.8h, v4.8h, v25.8h          \n"  // B
+    "add        v17.8h, v17.8h, v27.8h         \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v16.8h, #8              \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v17.8h, #8              \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(src_rgb565_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24",
+    "v25", "v26", "v27"
+  );
+}
+#endif  // HAS_RGB565TOUVROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_ARGB1555TOUVROW_NEON
+void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
+                        uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb1555_1 = src_argb1555 + src_stride_argb1555;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "uaddlp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 ARGB1555 pixels.
+    RGB555TOARGB
+    "uadalp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    "ins        v16.D[1], v26.D[0]             \n"
+    "ins        v17.D[1], v27.D[0]             \n"
+    "ins        v18.D[1], v28.D[0]             \n"
+
+    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
+    "urshr      v5.8h, v17.8h, #1              \n"
+    "urshr      v6.8h, v18.8h, #1              \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
+    "mul        v2.8h, v4.8h, v20.8h           \n"  // B
+    "mls        v2.8h, v5.8h, v21.8h           \n"  // G
+    "mls        v2.8h, v6.8h, v22.8h           \n"  // R
+    "add        v2.8h, v2.8h, v25.8h           \n"  // +128 -> unsigned
+    "mul        v3.8h, v6.8h, v20.8h           \n"  // R
+    "mls        v3.8h, v5.8h, v24.8h           \n"  // G
+    "mls        v3.8h, v4.8h, v23.8h           \n"  // B
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v2.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(src_argb1555_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25",
+    "v26", "v27", "v28"
+  );
+}
+#endif  // HAS_ARGB1555TOUVROW_NEON
+
+// 16x2 pixels -> 8x1.  pix is number of argb pixels. e.g. 16.
+#ifdef HAS_ARGB4444TOUVROW_NEON
+void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
+                          uint8* dst_u, uint8* dst_v, int pix) {
+  const uint8* src_argb4444_1 = src_argb4444 + src_stride_argb4444;
+  asm volatile (
+    RGBTOUV_SETUP_REG
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "uaddlp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "uaddlp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uaddlp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uaddlp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // load 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "uadalp     v16.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v17.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v18.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"  // next 8 ARGB4444 pixels.
+    ARGB4444TOARGB
+    "uadalp     v26.4h, v0.8b                  \n"  // B 8 bytes -> 4 shorts.
+    "uadalp     v27.4h, v1.8b                  \n"  // G 8 bytes -> 4 shorts.
+    "uadalp     v28.4h, v2.8b                  \n"  // R 8 bytes -> 4 shorts.
+
+    "ins        v16.D[1], v26.D[0]             \n"
+    "ins        v17.D[1], v27.D[0]             \n"
+    "ins        v18.D[1], v28.D[0]             \n"
+
+    "urshr      v4.8h, v16.8h, #1              \n"  // 2x average
+    "urshr      v5.8h, v17.8h, #1              \n"
+    "urshr      v6.8h, v18.8h, #1              \n"
+
+    "subs       %w4, %w4, #16                  \n"  // 16 processed per loop.
+    "mul        v2.8h, v4.8h, v20.8h           \n"  // B
+    "mls        v2.8h, v5.8h, v21.8h           \n"  // G
+    "mls        v2.8h, v6.8h, v22.8h           \n"  // R
+    "add        v2.8h, v2.8h, v25.8h           \n"  // +128 -> unsigned
+    "mul        v3.8h, v6.8h, v20.8h           \n"  // R
+    "mls        v3.8h, v5.8h, v24.8h           \n"  // G
+    "mls        v3.8h, v4.8h, v23.8h           \n"  // B
+    "add        v3.8h, v3.8h, v25.8h           \n"  // +128 -> unsigned
+    "uqshrn     v0.8b, v2.8h, #8               \n"  // 16 bit to 8 bit U
+    "uqshrn     v1.8b, v3.8h, #8               \n"  // 16 bit to 8 bit V
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 pixels U.
+    MEMACCESS(3)
+    "st1        {v1.8b}, [%3], #8              \n"  // store 8 pixels V.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(src_argb4444_1),  // %1
+    "+r"(dst_u),     // %2
+    "+r"(dst_v),     // %3
+    "+r"(pix)        // %4
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+    "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25",
+    "v26", "v27", "v28"
+
+  );
+}
+#endif  // HAS_ARGB4444TOUVROW_NEON
+
+#ifdef HAS_RGB565TOYROW_NEON
+void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v24.8b, #13                    \n"  // B * 0.1016 coefficient
+    "movi       v25.8b, #65                    \n"  // G * 0.5078 coefficient
+    "movi       v26.8b, #33                    \n"  // R * 0.2578 coefficient
+    "movi       v27.8b, #16                    \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 RGB565 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    RGB565TOARGB
+    "umull      v3.8h, v0.8b, v24.8b           \n"  // B
+    "umlal      v3.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v3.8h, v2.8b, v26.8b           \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v27.8b           \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb565),  // %0
+    "+r"(dst_y),       // %1
+    "+r"(pix)          // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v6",
+    "v24", "v25", "v26", "v27"
+  );
+}
+#endif  // HAS_RGB565TOYROW_NEON
+
+#ifdef HAS_ARGB1555TOYROW_NEON
+void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB1555 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGB1555TOARGB
+    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
+    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
+    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb1555),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGB1555TOYROW_NEON
+
+#ifdef HAS_ARGB4444TOYROW_NEON
+void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v24.8b, #13                    \n"  // B * 0.1016 coefficient
+    "movi       v25.8b, #65                    \n"  // G * 0.5078 coefficient
+    "movi       v26.8b, #33                    \n"  // R * 0.2578 coefficient
+    "movi       v27.8b, #16                    \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 8 ARGB4444 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    ARGB4444TOARGB
+    "umull      v3.8h, v0.8b, v24.8b           \n"  // B
+    "umlal      v3.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v3.8h, v2.8b, v26.8b           \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v27.8b           \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb4444),  // %0
+    "+r"(dst_y),         // %1
+    "+r"(pix)            // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v24", "v25", "v26", "v27"
+  );
+}
+#endif  // HAS_ARGB4444TOYROW_NEON
+
+#ifdef HAS_BGRATOYROW_NEON
+void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v1.8b, v4.8b           \n"  // R
+    "umlal      v16.8h, v2.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v3.8b, v6.8b           \n"  // B
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_bgra),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+#endif  // HAS_BGRATOYROW_NEON
+
+#ifdef HAS_ABGRTOYROW_NEON
+void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v0.8b, v4.8b           \n"  // R
+    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v2.8b, v6.8b           \n"  // B
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_abgr),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+#endif  // HAS_ABGRTOYROW_NEON
+
+#ifdef HAS_RGBATOYROW_NEON
+void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v1.8b, v4.8b           \n"  // B
+    "umlal      v16.8h, v2.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v3.8b, v6.8b           \n"  // R
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgba),  // %0
+    "+r"(dst_y),     // %1
+    "+r"(pix)        // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+#endif  // HAS_RGBATOYROW_NEON
+
+#ifdef HAS_RGB24TOYROW_NEON
+void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v0.8b, v4.8b           \n"  // B
+    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v2.8b, v6.8b           \n"  // R
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_rgb24),  // %0
+    "+r"(dst_y),      // %1
+    "+r"(pix)         // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+#endif  // HAS_RGB24TOYROW_NEON
+
+#ifdef HAS_RAWTOYROW_NEON
+void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix) {
+  asm volatile (
+    "movi       v4.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld3        {v0.8b,v1.8b,v2.8b}, [%0], #24 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v0.8b, v4.8b           \n"  // B
+    "umlal      v16.8h, v1.8b, v5.8b           \n"  // G
+    "umlal      v16.8h, v2.8b, v6.8b           \n"  // R
+    "sqrshrun   v0.8b, v16.8h, #7              \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
+    MEMACCESS(1)
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
+    "b.gt       1b                             \n"
+  : "+r"(src_raw),  // %0
+    "+r"(dst_y),    // %1
+    "+r"(pix)       // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+#endif  // HAS_RAWTOYROW_NEON
+
+// Bilinear filter 16x2 -> 16x1
+#ifdef HAS_INTERPOLATEROW_NEON
+void InterpolateRow_NEON(uint8* dst_ptr,
+                         const uint8* src_ptr, ptrdiff_t src_stride,
+                         int dst_width, int source_y_fraction) {
+  int y1_fraction = source_y_fraction;
+  int y0_fraction = 256 - y1_fraction;
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  asm volatile (
+    "cmp        %w4, #0                        \n"
+    "b.eq       100f                           \n"
+    "cmp        %w4, #64                       \n"
+    "b.eq       75f                            \n"
+    "cmp        %w4, #128                      \n"
+    "b.eq       50f                            \n"
+    "cmp        %w4, #192                      \n"
+    "b.eq       25f                            \n"
+
+    "dup        v5.16b, %w4                    \n"
+    "dup        v4.16b, %w5                    \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"
+    MEMACCESS(2)
+    "ld1        {v1.16b}, [%2], #16            \n"
+    "subs       %w3, %w3, #16                  \n"
+    "umull      v2.8h, v0.8b,  v4.8b           \n"
+    "umull2     v3.8h, v0.16b, v4.16b          \n"
+    "umlal      v2.8h, v1.8b,  v5.8b           \n"
+    "umlal2     v3.8h, v1.16b, v5.16b          \n"
+    "rshrn      v0.8b,  v2.8h, #8              \n"
+    "rshrn2     v0.16b, v3.8h, #8              \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       1b                             \n"
+    "b          99f                            \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"
+    MEMACCESS(2)
+    "ld1        {v1.16b}, [%2], #16            \n"
+    "subs       %w3, %w3, #16                  \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       25b                            \n"
+    "b          99f                            \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"
+    MEMACCESS(2)
+    "ld1        {v1.16b}, [%2], #16            \n"
+    "subs       %w3, %w3, #16                  \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       50b                            \n"
+    "b          99f                            \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "ld1        {v1.16b}, [%1], #16            \n"
+    MEMACCESS(2)
+    "ld1        {v0.16b}, [%2], #16            \n"
+    "subs       %w3, %w3, #16                  \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    "urhadd     v0.16b, v0.16b, v1.16b         \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       75b                            \n"
+    "b          99f                            \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "ld1        {v0.16b}, [%1], #16            \n"
+    "subs       %w3, %w3, #16                  \n"
+    MEMACCESS(0)
+    "st1        {v0.16b}, [%0], #16            \n"
+    "b.gt       100b                           \n"
+
+  "99:                                         \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_ptr1),         // %2
+    "+r"(dst_width),        // %3
+    "+r"(y1_fraction),      // %4
+    "+r"(y0_fraction)       // %5
+  :
+  : "cc", "memory", "v0", "v1", "v3", "v4", "v5"
+  );
+}
+#endif  // HAS_INTERPOLATEROW_NEON
+
+// dr * (256 - sa) / 256 + sr = dr - dr * sa / 256 + sr
+#ifdef HAS_ARGBBLENDROW_NEON
+void ARGBBlendRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  asm volatile (
+    "subs       %w3, %w3, #8                   \n"
+    "b.lt       89f                            \n"
+    // Blend 8 pixels.
+  "8:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB0 pixels
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 ARGB1 pixels
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "umull      v16.8h, v4.8b, v3.8b           \n"  // db * a
+    "umull      v17.8h, v5.8b, v3.8b           \n"  // dg * a
+    "umull      v18.8h, v6.8b, v3.8b           \n"  // dr * a
+    "uqrshrn    v16.8b, v16.8h, #8             \n"  // db >>= 8
+    "uqrshrn    v17.8b, v17.8h, #8             \n"  // dg >>= 8
+    "uqrshrn    v18.8b, v18.8h, #8             \n"  // dr >>= 8
+    "uqsub      v4.8b, v4.8b, v16.8b           \n"  // db - (db * a / 256)
+    "uqsub      v5.8b, v5.8b, v17.8b           \n"  // dg - (dg * a / 256)
+    "uqsub      v6.8b, v6.8b, v18.8b           \n"  // dr - (dr * a / 256)
+    "uqadd      v0.8b, v0.8b, v4.8b            \n"  // + sb
+    "uqadd      v1.8b, v1.8b, v5.8b            \n"  // + sg
+    "uqadd      v2.8b, v2.8b, v6.8b            \n"  // + sr
+    "movi       v3.8b, #255                    \n"  // a = 255
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.ge       8b                             \n"
+
+  "89:                                         \n"
+    "adds       %w3, %w3, #8-1                 \n"
+    "b.lt       99f                            \n"
+
+    // Blend 1 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.b,v1.b,v2.b,v3.b}[0], [%0], #4 \n"  // load 1 pixel ARGB0.
+    MEMACCESS(1)
+    "ld4        {v4.b,v5.b,v6.b,v7.b}[0], [%1], #4 \n"  // load 1 pixel ARGB1.
+    "subs       %w3, %w3, #1                   \n"  // 1 processed per loop.
+    "umull      v16.8h, v4.8b, v3.8b           \n"  // db * a
+    "umull      v17.8h, v5.8b, v3.8b           \n"  // dg * a
+    "umull      v18.8h, v6.8b, v3.8b           \n"  // dr * a
+    "uqrshrn    v16.8b, v16.8h, #8             \n"  // db >>= 8
+    "uqrshrn    v17.8b, v17.8h, #8             \n"  // dg >>= 8
+    "uqrshrn    v18.8b, v18.8h, #8             \n"  // dr >>= 8
+    "uqsub      v4.8b, v4.8b, v16.8b           \n"  // db - (db * a / 256)
+    "uqsub      v5.8b, v5.8b, v17.8b           \n"  // dg - (dg * a / 256)
+    "uqsub      v6.8b, v6.8b, v18.8b           \n"  // dr - (dr * a / 256)
+    "uqadd      v0.8b, v0.8b, v4.8b            \n"  // + sb
+    "uqadd      v1.8b, v1.8b, v5.8b            \n"  // + sg
+    "uqadd      v2.8b, v2.8b, v6.8b            \n"  // + sr
+    "movi       v3.8b, #255                    \n"  // a = 255
+    MEMACCESS(2)
+    "st4        {v0.b,v1.b,v2.b,v3.b}[0], [%2], #4 \n"  // store 1 pixel.
+    "b.ge       1b                             \n"
+
+  "99:                                         \n"
+
+  : "+r"(src_argb0),    // %0
+    "+r"(src_argb1),    // %1
+    "+r"(dst_argb),     // %2
+    "+r"(width)         // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v16", "v17", "v18"
+  );
+}
+#endif  // HAS_ARGBBLENDROW_NEON
+
+// Attenuate 8 pixels at a time.
+#ifdef HAS_ARGBATTENUATEROW_NEON
+void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    // Attenuate 8 pixels.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v4.8h, v0.8b, v3.8b            \n"  // b * a
+    "umull      v5.8h, v1.8b, v3.8b            \n"  // g * a
+    "umull      v6.8h, v2.8b, v3.8b            \n"  // r * a
+    "uqrshrn    v0.8b, v4.8h, #8               \n"  // b >>= 8
+    "uqrshrn    v1.8b, v5.8h, #8               \n"  // g >>= 8
+    "uqrshrn    v2.8b, v6.8h, #8               \n"  // r >>= 8
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
+  );
+}
+#endif  // HAS_ARGBATTENUATEROW_NEON
+
+// Quantize 8 ARGB pixels (32 bytes).
+// dst = (dst * scale >> 16) * interval_size + interval_offset;
+#ifdef HAS_ARGBQUANTIZEROW_NEON
+void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  asm volatile (
+    "dup        v4.8h, %w2                     \n"
+    "ushr       v4.8h, v4.8h, #1               \n"  // scale >>= 1
+    "dup        v5.8h, %w3                     \n"  // interval multiply.
+    "dup        v6.8h, %w4                     \n"  // interval add
+
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0]  \n"  // load 8 pixels of ARGB.
+    "subs       %w1, %w1, #8                   \n"  // 8 processed per loop.
+    "uxtl       v0.8h, v0.8b                   \n"  // b (0 .. 255)
+    "uxtl       v1.8h, v1.8b                   \n"
+    "uxtl       v2.8h, v2.8b                   \n"
+    "sqdmulh    v0.8h, v0.8h, v4.8h            \n"  // b * scale
+    "sqdmulh    v1.8h, v1.8h, v4.8h            \n"  // g
+    "sqdmulh    v2.8h, v2.8h, v4.8h            \n"  // r
+    "mul        v0.8h, v0.8h, v5.8h            \n"  // b * interval_size
+    "mul        v1.8h, v1.8h, v5.8h            \n"  // g
+    "mul        v2.8h, v2.8h, v5.8h            \n"  // r
+    "add        v0.8h, v0.8h, v6.8h            \n"  // b + interval_offset
+    "add        v1.8h, v1.8h, v6.8h            \n"  // g
+    "add        v2.8h, v2.8h, v6.8h            \n"  // r
+    "uqxtn      v0.8b, v0.8h                   \n"
+    "uqxtn      v1.8b, v1.8h                   \n"
+    "uqxtn      v2.8b, v2.8h                   \n"
+    MEMACCESS(0)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(dst_argb),       // %0
+    "+r"(width)           // %1
+  : "r"(scale),           // %2
+    "r"(interval_size),   // %3
+    "r"(interval_offset)  // %4
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
+  );
+}
+#endif  // HAS_ARGBQUANTIZEROW_NEON
+
+// Shade 8 pixels at a time by specified value.
+// NOTE vqrdmulh.s16 q10, q10, d0[0] must use a scaler register from 0 to 8.
+// Rounding in vqrdmulh does +1 to high if high bit of low s16 is set.
+#ifdef HAS_ARGBSHADEROW_NEON
+void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  asm volatile (
+    "dup        v0.4s, %w3                     \n"  // duplicate scale value.
+    "zip1       v0.8b, v0.8b, v0.8b            \n"  // v0.8b aarrggbb.
+    "ushr       v0.8h, v0.8h, #1               \n"  // scale / 2.
+
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "uxtl       v4.8h, v4.8b                   \n"  // b (0 .. 255)
+    "uxtl       v5.8h, v5.8b                   \n"
+    "uxtl       v6.8h, v6.8b                   \n"
+    "uxtl       v7.8h, v7.8b                   \n"
+    "sqrdmulh   v4.8h, v4.8h, v0.h[0]          \n"  // b * scale * 2
+    "sqrdmulh   v5.8h, v5.8h, v0.h[1]          \n"  // g
+    "sqrdmulh   v6.8h, v6.8h, v0.h[2]          \n"  // r
+    "sqrdmulh   v7.8h, v7.8h, v0.h[3]          \n"  // a
+    "uqxtn      v4.8b, v4.8h                   \n"
+    "uqxtn      v5.8b, v5.8h                   \n"
+    "uqxtn      v6.8b, v6.8h                   \n"
+    "uqxtn      v7.8b, v7.8h                   \n"
+    MEMACCESS(1)
+    "st4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),       // %0
+    "+r"(dst_argb),       // %1
+    "+r"(width)           // %2
+  : "r"(value)            // %3
+  : "cc", "memory", "v0", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGBSHADEROW_NEON
+
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels
+// Similar to ARGBToYJ but stores ARGB.
+// C code is (15 * b + 75 * g + 38 * r + 64) >> 7;
+#ifdef HAS_ARGBGRAYROW_NEON
+void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width) {
+  asm volatile (
+    "movi       v24.8b, #15                    \n"  // B * 0.11400 coefficient
+    "movi       v25.8b, #75                    \n"  // G * 0.58700 coefficient
+    "movi       v26.8b, #38                    \n"  // R * 0.29900 coefficient
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "umull      v4.8h, v0.8b, v24.8b           \n"  // B
+    "umlal      v4.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v4.8h, v2.8b, v26.8b           \n"  // R
+    "sqrshrun   v0.8b, v4.8h, #7               \n"  // 15 bit to 8 bit B
+    "orr        v1.8b, v0.8b, v0.8b            \n"  // G
+    "orr        v2.8b, v0.8b, v0.8b            \n"  // R
+    MEMACCESS(1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32 \n"  // store 8 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(width)      // %2
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v24", "v25", "v26"
+  );
+}
+#endif  // HAS_ARGBGRAYROW_NEON
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+
+#ifdef HAS_ARGBSEPIAROW_NEON
+void ARGBSepiaRow_NEON(uint8* dst_argb, int width) {
+  asm volatile (
+    "movi       v20.8b, #17                    \n"  // BB coefficient
+    "movi       v21.8b, #68                    \n"  // BG coefficient
+    "movi       v22.8b, #35                    \n"  // BR coefficient
+    "movi       v24.8b, #22                    \n"  // GB coefficient
+    "movi       v25.8b, #88                    \n"  // GG coefficient
+    "movi       v26.8b, #45                    \n"  // GR coefficient
+    "movi       v28.8b, #24                    \n"  // BB coefficient
+    "movi       v29.8b, #98                    \n"  // BG coefficient
+    "movi       v30.8b, #50                    \n"  // BR coefficient
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0] \n"  // load 8 ARGB pixels.
+    "subs       %w1, %w1, #8                   \n"  // 8 processed per loop.
+    "umull      v4.8h, v0.8b, v20.8b           \n"  // B to Sepia B
+    "umlal      v4.8h, v1.8b, v21.8b           \n"  // G
+    "umlal      v4.8h, v2.8b, v22.8b           \n"  // R
+    "umull      v5.8h, v0.8b, v24.8b           \n"  // B to Sepia G
+    "umlal      v5.8h, v1.8b, v25.8b           \n"  // G
+    "umlal      v5.8h, v2.8b, v26.8b           \n"  // R
+    "umull      v6.8h, v0.8b, v28.8b           \n"  // B to Sepia R
+    "umlal      v6.8h, v1.8b, v29.8b           \n"  // G
+    "umlal      v6.8h, v2.8b, v30.8b           \n"  // R
+    "uqshrn     v0.8b, v4.8h, #7               \n"  // 16 bit to 8 bit B
+    "uqshrn     v1.8b, v5.8h, #7               \n"  // 16 bit to 8 bit G
+    "uqshrn     v2.8b, v6.8h, #7               \n"  // 16 bit to 8 bit R
+    MEMACCESS(0)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // store 8 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(dst_argb),  // %0
+    "+r"(width)      // %1
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+    "v20", "v21", "v22", "v24", "v25", "v26", "v28", "v29", "v30"
+  );
+}
+#endif  // HAS_ARGBSEPIAROW_NEON
+
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// TODO(fbarchard): Was same as Sepia except matrix is provided.  This function
+// needs to saturate.  Consider doing a non-saturating version.
+#ifdef HAS_ARGBCOLORMATRIXROW_NEON
+void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
+                             const int8* matrix_argb, int width) {
+  asm volatile (
+    MEMACCESS(3)
+    "ld1        {v2.16b}, [%3]                 \n"  // load 3 ARGB vectors.
+    "sxtl       v0.8h, v2.8b                   \n"  // B,G coefficients s16.
+    "sxtl2      v1.8h, v2.16b                  \n"  // R,A coefficients s16.
+
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v16.8b,v17.8b,v18.8b,v19.8b}, [%0], #32 \n"  // load 8 pixels.
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "uxtl       v16.8h, v16.8b                 \n"  // b (0 .. 255) 16 bit
+    "uxtl       v17.8h, v17.8b                 \n"  // g
+    "uxtl       v18.8h, v18.8b                 \n"  // r
+    "uxtl       v19.8h, v19.8b                 \n"  // a
+    "mul        v22.8h, v16.8h, v0.h[0]        \n"  // B = B * Matrix B
+    "mul        v23.8h, v16.8h, v0.h[4]        \n"  // G = B * Matrix G
+    "mul        v24.8h, v16.8h, v1.h[0]        \n"  // R = B * Matrix R
+    "mul        v25.8h, v16.8h, v1.h[4]        \n"  // A = B * Matrix A
+    "mul        v4.8h, v17.8h, v0.h[1]         \n"  // B += G * Matrix B
+    "mul        v5.8h, v17.8h, v0.h[5]         \n"  // G += G * Matrix G
+    "mul        v6.8h, v17.8h, v1.h[1]         \n"  // R += G * Matrix R
+    "mul        v7.8h, v17.8h, v1.h[5]         \n"  // A += G * Matrix A
+    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
+    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
+    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
+    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
+    "mul        v4.8h, v18.8h, v0.h[2]         \n"  // B += R * Matrix B
+    "mul        v5.8h, v18.8h, v0.h[6]         \n"  // G += R * Matrix G
+    "mul        v6.8h, v18.8h, v1.h[2]         \n"  // R += R * Matrix R
+    "mul        v7.8h, v18.8h, v1.h[6]         \n"  // A += R * Matrix A
+    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
+    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
+    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
+    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
+    "mul        v4.8h, v19.8h, v0.h[3]         \n"  // B += A * Matrix B
+    "mul        v5.8h, v19.8h, v0.h[7]         \n"  // G += A * Matrix G
+    "mul        v6.8h, v19.8h, v1.h[3]         \n"  // R += A * Matrix R
+    "mul        v7.8h, v19.8h, v1.h[7]         \n"  // A += A * Matrix A
+    "sqadd      v22.8h, v22.8h, v4.8h          \n"  // Accumulate B
+    "sqadd      v23.8h, v23.8h, v5.8h          \n"  // Accumulate G
+    "sqadd      v24.8h, v24.8h, v6.8h          \n"  // Accumulate R
+    "sqadd      v25.8h, v25.8h, v7.8h          \n"  // Accumulate A
+    "sqshrun    v16.8b, v22.8h, #6             \n"  // 16 bit to 8 bit B
+    "sqshrun    v17.8b, v23.8h, #6             \n"  // 16 bit to 8 bit G
+    "sqshrun    v18.8b, v24.8h, #6             \n"  // 16 bit to 8 bit R
+    "sqshrun    v19.8b, v25.8h, #6             \n"  // 16 bit to 8 bit A
+    MEMACCESS(1)
+    "st4        {v16.8b,v17.8b,v18.8b,v19.8b}, [%1], #32 \n"  // store 8 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(width)       // %2
+  : "r"(matrix_argb)  // %3
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
+    "v18", "v19", "v22", "v23", "v24", "v25"
+  );
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_NEON
+
+// TODO(fbarchard): fix vqshrun in ARGBMultiplyRow_NEON and reenable.
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+#ifdef HAS_ARGBMULTIPLYROW_NEON
+void ARGBMultiplyRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "umull      v0.8h, v0.8b, v4.8b            \n"  // multiply B
+    "umull      v1.8h, v1.8b, v5.8b            \n"  // multiply G
+    "umull      v2.8h, v2.8b, v6.8b            \n"  // multiply R
+    "umull      v3.8h, v3.8b, v7.8b            \n"  // multiply A
+    "rshrn      v0.8b, v0.8h, #8               \n"  // 16 bit to 8 bit B
+    "rshrn      v1.8b, v1.8h, #8               \n"  // 16 bit to 8 bit G
+    "rshrn      v2.8b, v2.8h, #8               \n"  // 16 bit to 8 bit R
+    "rshrn      v3.8b, v3.8h, #8               \n"  // 16 bit to 8 bit A
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGBMULTIPLYROW_NEON
+
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+#ifdef HAS_ARGBADDROW_NEON
+void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uqadd      v0.8b, v0.8b, v4.8b            \n"
+    "uqadd      v1.8b, v1.8b, v5.8b            \n"
+    "uqadd      v2.8b, v2.8b, v6.8b            \n"
+    "uqadd      v3.8b, v3.8b, v7.8b            \n"
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGBADDROW_NEON
+
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+#ifdef HAS_ARGBSUBTRACTROW_NEON
+void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load 8 ARGB pixels.
+    MEMACCESS(1)
+    "ld4        {v4.8b,v5.8b,v6.8b,v7.8b}, [%1], #32 \n"  // load 8 more pixels.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uqsub      v0.8b, v0.8b, v4.8b            \n"
+    "uqsub      v1.8b, v1.8b, v5.8b            \n"
+    "uqsub      v2.8b, v2.8b, v6.8b            \n"
+    "uqsub      v3.8b, v3.8b, v7.8b            \n"
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_NEON
+
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+#ifdef HAS_SOBELROW_NEON
+void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "movi       v3.8b, #255                    \n"  // alpha
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.8b}, [%0], #8              \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "ld1        {v1.8b}, [%1], #8              \n"  // load 8 sobely.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uqadd      v0.8b, v0.8b, v1.8b            \n"  // add
+    "orr        v1.8b, v0.8b, v0.8b            \n"
+    "orr        v2.8b, v0.8b, v0.8b            \n"
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"
+  );
+}
+#endif  // HAS_SOBELROW_NEON
+
+// Adds Sobel X and Sobel Y and stores Sobel into plane.
+#ifdef HAS_SOBELTOPLANEROW_NEON
+void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  asm volatile (
+    // 16 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b}, [%0], #16            \n"  // load 16 sobelx.
+    MEMACCESS(1)
+    "ld1        {v1.16b}, [%1], #16            \n"  // load 16 sobely.
+    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop.
+    "uqadd      v0.16b, v0.16b, v1.16b         \n"  // add
+    MEMACCESS(2)
+    "st1        {v0.16b}, [%2], #16            \n"  // store 16 pixels.
+    "b.gt       1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_y),       // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "v0", "v1"
+  );
+}
+#endif  // HAS_SOBELTOPLANEROW_NEON
+
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+#ifdef HAS_SOBELXYROW_NEON
+void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  asm volatile (
+    "movi       v3.8b, #255                    \n"  // alpha
+    // 8 pixel loop.
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v2.8b}, [%0], #8              \n"  // load 8 sobelx.
+    MEMACCESS(1)
+    "ld1        {v0.8b}, [%1], #8              \n"  // load 8 sobely.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uqadd      v1.8b, v0.8b, v2.8b            \n"  // add
+    MEMACCESS(2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32 \n"  // store 8 ARGB pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_sobelx),  // %0
+    "+r"(src_sobely),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(width)        // %3
+  :
+  : "cc", "memory", "v0", "v1", "v2", "v3"
+  );
+}
+#endif  // HAS_SOBELXYROW_NEON
+
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+#ifdef HAS_SOBELXROW_NEON
+void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.8b}, [%0],%5               \n"  // top
+    MEMACCESS(0)
+    "ld1        {v1.8b}, [%0],%6               \n"
+    "usubl      v0.8h, v0.8b, v1.8b            \n"
+    MEMACCESS(1)
+    "ld1        {v2.8b}, [%1],%5               \n"  // center * 2
+    MEMACCESS(1)
+    "ld1        {v3.8b}, [%1],%6               \n"
+    "usubl      v1.8h, v2.8b, v3.8b            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    MEMACCESS(2)
+    "ld1        {v2.8b}, [%2],%5               \n"  // bottom
+    MEMACCESS(2)
+    "ld1        {v3.8b}, [%2],%6               \n"
+    "subs       %w4, %w4, #8                   \n"  // 8 pixels
+    "usubl      v1.8h, v2.8b, v3.8b            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    "abs        v0.8h, v0.8h                   \n"
+    "uqxtn      v0.8b, v0.8h                   \n"
+    MEMACCESS(3)
+    "st1        {v0.8b}, [%3], #8              \n"  // store 8 sobelx
+    "b.gt       1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(src_y2),      // %2
+    "+r"(dst_sobelx),  // %3
+    "+r"(width)        // %4
+  : "r"(2LL),          // %5
+    "r"(6LL)           // %6
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_SOBELXROW_NEON
+
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+#ifdef HAS_SOBELYROW_NEON
+void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.8b}, [%0],%4               \n"  // left
+    MEMACCESS(1)
+    "ld1        {v1.8b}, [%1],%4               \n"
+    "usubl      v0.8h, v0.8b, v1.8b            \n"
+    MEMACCESS(0)
+    "ld1        {v2.8b}, [%0],%4               \n"  // center * 2
+    MEMACCESS(1)
+    "ld1        {v3.8b}, [%1],%4               \n"
+    "usubl      v1.8h, v2.8b, v3.8b            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    MEMACCESS(0)
+    "ld1        {v2.8b}, [%0],%5               \n"  // right
+    MEMACCESS(1)
+    "ld1        {v3.8b}, [%1],%5               \n"
+    "subs       %w3, %w3, #8                   \n"  // 8 pixels
+    "usubl      v1.8h, v2.8b, v3.8b            \n"
+    "add        v0.8h, v0.8h, v1.8h            \n"
+    "abs        v0.8h, v0.8h                   \n"
+    "uqxtn      v0.8b, v0.8h                   \n"
+    MEMACCESS(2)
+    "st1        {v0.8b}, [%2], #8              \n"  // store 8 sobely
+    "b.gt       1b                             \n"
+  : "+r"(src_y0),      // %0
+    "+r"(src_y1),      // %1
+    "+r"(dst_sobely),  // %2
+    "+r"(width)        // %3
+  : "r"(1LL),          // %4
+    "r"(6LL)           // %5
+  : "cc", "memory", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+#endif  // HAS_SOBELYROW_NEON
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/row_win.cc b/third_party/aom/third_party/libyuv/source/row_win.cc
new file mode 100644
index 0000000000..71be268b47
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/row_win.cc
@@ -0,0 +1,6331 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_X64) && \
+    defined(_MSC_VER) && !defined(__clang__)
+#include <emmintrin.h>
+#include <tmmintrin.h>  // For _mm_maddubs_epi16
+#endif
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(_M_IX86) || defined(_M_X64)) && \
+    defined(_MSC_VER) && !defined(__clang__)
+
+struct YuvConstants {
+  lvec8 kUVToB;     // 0
+  lvec8 kUVToG;     // 32
+  lvec8 kUVToR;     // 64
+  lvec16 kUVBiasB;  // 96
+  lvec16 kUVBiasG;  // 128
+  lvec16 kUVBiasR;  // 160
+  lvec16 kYToRgb;   // 192
+};
+
+// BT.601 YUV to RGB reference
+//  R = (Y - 16) * 1.164              - V * -1.596
+//  G = (Y - 16) * 1.164 - U *  0.391 - V *  0.813
+//  B = (Y - 16) * 1.164 - U * -2.018
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YG 18997 /* round(1.164 * 64 * 256 * 256 / 257) */
+#define YGB -1160 /* 1.164 * 64 * -16 + 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UB -128 /* max(-128, round(-2.018 * 64)) */
+#define UG 25 /* round(0.391 * 64) */
+#define VG 52 /* round(0.813 * 64) */
+#define VR -102 /* round(-1.596 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BB (UB * 128            + YGB)
+#define BG (UG * 128 + VG * 128 + YGB)
+#define BR            (VR * 128 + YGB)
+
+// BT601 constants for YUV to RGB.
+static YuvConstants SIMD_ALIGNED(kYuvConstants) = {
+  { UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0,
+    UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0 },
+  { UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG,
+    UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG },
+  { 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR,
+    0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
+};
+
+// BT601 constants for NV21 where chroma plane is VU instead of UV.
+static YuvConstants SIMD_ALIGNED(kYvuConstants) = {
+  { 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB,
+    0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB, 0, UB },
+  { VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG,
+    VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG, VG, UG },
+  { VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0,
+    VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0, VR, 0 },
+  { BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB, BB },
+  { BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG, BG },
+  { BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR, BR },
+  { YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG, YG }
+};
+
+#undef YG
+#undef YGB
+#undef UB
+#undef UG
+#undef VG
+#undef VR
+#undef BB
+#undef BG
+#undef BR
+
+// JPEG YUV to RGB reference
+// *  R = Y                - V * -1.40200
+// *  G = Y - U *  0.34414 - V *  0.71414
+// *  B = Y - U * -1.77200
+
+// Y contribution to R,G,B.  Scale and bias.
+// TODO(fbarchard): Consider moving constants into a common header.
+#define YGJ 16320 /* round(1.000 * 64 * 256 * 256 / 257) */
+#define YGBJ 32  /* 64 / 2 */
+
+// U and V contributions to R,G,B.
+#define UBJ -113 /* round(-1.77200 * 64) */
+#define UGJ 22 /* round(0.34414 * 64) */
+#define VGJ 46 /* round(0.71414  * 64) */
+#define VRJ -90 /* round(-1.40200 * 64) */
+
+// Bias values to subtract 16 from Y and 128 from U and V.
+#define BBJ (UBJ * 128             + YGBJ)
+#define BGJ (UGJ * 128 + VGJ * 128 + YGBJ)
+#define BRJ             (VRJ * 128 + YGBJ)
+
+// JPEG constants for YUV to RGB.
+static YuvConstants SIMD_ALIGNED(kYuvJConstants) = {
+  { UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0,
+    UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0, UBJ, 0 },
+  { UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ,
+    UGJ, VGJ, UGJ, VGJ, UGJ, VGJ, UGJ, VGJ },
+  { 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ,
+    0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ, 0, VRJ },
+  { BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ,
+    BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ, BBJ },
+  { BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ,
+    BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ, BGJ },
+  { BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ,
+    BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ, BRJ },
+  { YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ,
+    YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ, YGJ }
+};
+
+#undef YGJ
+#undef YGBJ
+#undef UBJ
+#undef UGJ
+#undef VGJ
+#undef VRJ
+#undef BBJ
+#undef BGJ
+#undef BRJ
+
+// 64 bit
+#if defined(_M_X64)
+#if defined(HAS_I422TOARGBROW_SSSE3)
+void I422ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __m128i xmm0, xmm1, xmm2, xmm3;
+  const __m128i xmm5 = _mm_set1_epi8(-1);
+  const ptrdiff_t offset = (uint8*)v_buf - (uint8*)u_buf;
+
+  while (width > 0) {
+    xmm0 = _mm_cvtsi32_si128(*(uint32*)u_buf);
+    xmm1 = _mm_cvtsi32_si128(*(uint32*)(u_buf + offset));
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
+    xmm0 = _mm_unpacklo_epi16(xmm0, xmm0);
+    xmm1 = _mm_loadu_si128(&xmm0);
+    xmm2 = _mm_loadu_si128(&xmm0);
+    xmm0 = _mm_maddubs_epi16(xmm0, *(__m128i*)kYuvConstants.kUVToB);
+    xmm1 = _mm_maddubs_epi16(xmm1, *(__m128i*)kYuvConstants.kUVToG);
+    xmm2 = _mm_maddubs_epi16(xmm2, *(__m128i*)kYuvConstants.kUVToR);
+    xmm0 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasB, xmm0);
+    xmm1 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasG, xmm1);
+    xmm2 = _mm_sub_epi16(*(__m128i*)kYuvConstants.kUVBiasR, xmm2);
+    xmm3 = _mm_loadl_epi64((__m128i*)y_buf);
+    xmm3 = _mm_unpacklo_epi8(xmm3, xmm3);
+    xmm3 = _mm_mulhi_epu16(xmm3, *(__m128i*)kYuvConstants.kYToRgb);
+    xmm0 = _mm_adds_epi16(xmm0, xmm3);
+    xmm1 = _mm_adds_epi16(xmm1, xmm3);
+    xmm2 = _mm_adds_epi16(xmm2, xmm3);
+    xmm0 = _mm_srai_epi16(xmm0, 6);
+    xmm1 = _mm_srai_epi16(xmm1, 6);
+    xmm2 = _mm_srai_epi16(xmm2, 6);
+    xmm0 = _mm_packus_epi16(xmm0, xmm0);
+    xmm1 = _mm_packus_epi16(xmm1, xmm1);
+    xmm2 = _mm_packus_epi16(xmm2, xmm2);
+    xmm0 = _mm_unpacklo_epi8(xmm0, xmm1);
+    xmm2 = _mm_unpacklo_epi8(xmm2, xmm5);
+    xmm1 = _mm_loadu_si128(&xmm0);
+    xmm0 = _mm_unpacklo_epi16(xmm0, xmm2);
+    xmm1 = _mm_unpackhi_epi16(xmm1, xmm2);
+
+    _mm_storeu_si128((__m128i *)dst_argb, xmm0);
+    _mm_storeu_si128((__m128i *)(dst_argb + 16), xmm1);
+
+    y_buf += 8;
+    u_buf += 4;
+    dst_argb += 32;
+    width -= 8;
+  }
+}
+#endif
+// 32 bit
+#else  // defined(_M_X64)
+#ifdef HAS_ARGBTOYROW_SSSE3
+
+// Constants for ARGB.
+static const vec8 kARGBToY = {
+  13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0
+};
+
+// JPeg full range.
+static const vec8 kARGBToYJ = {
+  15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0, 15, 75, 38, 0
+};
+
+static const vec8 kARGBToU = {
+  112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0
+};
+
+static const vec8 kARGBToUJ = {
+  127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0, 127, -84, -43, 0
+};
+
+static const vec8 kARGBToV = {
+  -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0,
+};
+
+static const vec8 kARGBToVJ = {
+  -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0, -20, -107, 127, 0
+};
+
+// vpshufb for vphaddw + vpackuswb packed to shorts.
+static const lvec8 kShufARGBToUV_AVX = {
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15,
+  0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15
+};
+
+// Constants for BGRA.
+static const vec8 kBGRAToY = {
+  0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13
+};
+
+static const vec8 kBGRAToU = {
+  0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112
+};
+
+static const vec8 kBGRAToV = {
+  0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18
+};
+
+// Constants for ABGR.
+static const vec8 kABGRToY = {
+  33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0, 33, 65, 13, 0
+};
+
+static const vec8 kABGRToU = {
+  -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0, -38, -74, 112, 0
+};
+
+static const vec8 kABGRToV = {
+  112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0, 112, -94, -18, 0
+};
+
+// Constants for RGBA.
+static const vec8 kRGBAToY = {
+  0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33, 0, 13, 65, 33
+};
+
+static const vec8 kRGBAToU = {
+  0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38, 0, 112, -74, -38
+};
+
+static const vec8 kRGBAToV = {
+  0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112, 0, -18, -94, 112
+};
+
+static const uvec8 kAddY16 = {
+  16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u, 16u
+};
+
+// 7 bit fixed point 0.5.
+static const vec16 kAddYJ64 = {
+  64, 64, 64, 64, 64, 64, 64, 64
+};
+
+static const uvec8 kAddUV128 = {
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u,
+  128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u
+};
+
+static const uvec16 kAddUVJ128 = {
+  0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u, 0x8080u
+};
+
+// Shuffle table for converting RGB24 to ARGB.
+static const uvec8 kShuffleMaskRGB24ToARGB = {
+  0u, 1u, 2u, 12u, 3u, 4u, 5u, 13u, 6u, 7u, 8u, 14u, 9u, 10u, 11u, 15u
+};
+
+// Shuffle table for converting RAW to ARGB.
+static const uvec8 kShuffleMaskRAWToARGB = {
+  2u, 1u, 0u, 12u, 5u, 4u, 3u, 13u, 8u, 7u, 6u, 14u, 11u, 10u, 9u, 15u
+};
+
+// Shuffle table for converting ARGB to RGB24.
+static const uvec8 kShuffleMaskARGBToRGB24 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static const uvec8 kShuffleMaskARGBToRAW = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u
+};
+
+// Shuffle table for converting ARGBToRGB24 for I422ToRGB24.  First 8 + next 4
+static const uvec8 kShuffleMaskARGBToRGB24_0 = {
+  0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u
+};
+
+// Shuffle table for converting ARGB to RAW.
+static const uvec8 kShuffleMaskARGBToRAW_0 = {
+  2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 128u, 128u, 128u, 128u, 8u, 14u, 13u, 12u
+};
+
+// Duplicates gray value 3 times and fills in alpha opaque.
+__declspec(naked)
+void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]        // src_y
+    mov        edx, [esp + 8]        // dst_argb
+    mov        ecx, [esp + 12]       // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0xff000000
+    pslld      xmm5, 24
+
+  convertloop:
+    movq       xmm0, qword ptr [eax]
+    lea        eax,  [eax + 8]
+    punpcklbw  xmm0, xmm0
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0
+    punpckhwd  xmm1, xmm1
+    por        xmm0, xmm5
+    por        xmm1, xmm5
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+
+#ifdef HAS_J400TOARGBROW_AVX2
+// Duplicates gray value 3 times and fills in alpha opaque.
+__declspec(naked)
+void J400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int pix) {
+  __asm {
+    mov         eax, [esp + 4]        // src_y
+    mov         edx, [esp + 8]        // dst_argb
+    mov         ecx, [esp + 12]       // pix
+    vpcmpeqb    ymm5, ymm5, ymm5      // generate mask 0xff000000
+    vpslld      ymm5, ymm5, 24
+
+  convertloop:
+    vmovdqu     xmm0, [eax]
+    lea         eax,  [eax + 16]
+    vpermq      ymm0, ymm0, 0xd8
+    vpunpcklbw  ymm0, ymm0, ymm0
+    vpermq      ymm0, ymm0, 0xd8
+    vpunpckhwd  ymm1, ymm0, ymm0
+    vpunpcklwd  ymm0, ymm0, ymm0
+    vpor        ymm0, ymm0, ymm5
+    vpor        ymm1, ymm1, ymm5
+    vmovdqu     [edx], ymm0
+    vmovdqu     [edx + 32], ymm1
+    lea         edx, [edx + 64]
+    sub         ecx, 16
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_J400TOARGBROW_AVX2
+
+__declspec(naked)
+void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_rgb24
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pslld     xmm5, 24
+    movdqa    xmm4, kShuffleMaskRGB24ToARGB
+
+ convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm3, [eax + 32]
+    lea       eax, [eax + 48]
+    movdqa    xmm2, xmm3
+    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    pshufb    xmm2, xmm4
+    por       xmm2, xmm5
+    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    pshufb    xmm0, xmm4
+    movdqu    [edx + 32], xmm2
+    por       xmm0, xmm5
+    pshufb    xmm1, xmm4
+    movdqu    [edx], xmm0
+    por       xmm1, xmm5
+    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    pshufb    xmm3, xmm4
+    movdqu    [edx + 16], xmm1
+    por       xmm3, xmm5
+    movdqu    [edx + 48], xmm3
+    lea       edx, [edx + 64]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb,
+                        int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_raw
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm5, xmm5       // generate mask 0xff000000
+    pslld     xmm5, 24
+    movdqa    xmm4, kShuffleMaskRAWToARGB
+
+ convertloop:
+    movdqu    xmm0, [eax]
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm3, [eax + 32]
+    lea       eax, [eax + 48]
+    movdqa    xmm2, xmm3
+    palignr   xmm2, xmm1, 8    // xmm2 = { xmm3[0:3] xmm1[8:15]}
+    pshufb    xmm2, xmm4
+    por       xmm2, xmm5
+    palignr   xmm1, xmm0, 12   // xmm1 = { xmm3[0:7] xmm0[12:15]}
+    pshufb    xmm0, xmm4
+    movdqu    [edx + 32], xmm2
+    por       xmm0, xmm5
+    pshufb    xmm1, xmm4
+    movdqu    [edx], xmm0
+    por       xmm1, xmm5
+    palignr   xmm3, xmm3, 4    // xmm3 = { xmm3[4:15]}
+    pshufb    xmm3, xmm4
+    movdqu    [edx + 16], xmm1
+    por       xmm3, xmm5
+    movdqu    [edx + 48], xmm3
+    lea       edx, [edx + 64]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+// pmul method to replicate bits.
+// Math to replicate bits:
+// (v << 8) | (v << 3)
+// v * 256 + v * 8
+// v * (256 + 8)
+// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
+// 20 instructions.
+__declspec(naked)
+void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb,
+                          int pix) {
+  __asm {
+    mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
+    movd      xmm5, eax
+    pshufd    xmm5, xmm5, 0
+    mov       eax, 0x20802080  // multiplier shift by 5 and then repeat 6 bits
+    movd      xmm6, eax
+    pshufd    xmm6, xmm6, 0
+    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    psllw     xmm3, 11
+    pcmpeqb   xmm4, xmm4       // generate mask 0x07e007e0 for Green
+    psllw     xmm4, 10
+    psrlw     xmm4, 5
+    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    psllw     xmm7, 8
+
+    mov       eax, [esp + 4]   // src_rgb565
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of bgr565
+    movdqa    xmm1, xmm0
+    movdqa    xmm2, xmm0
+    pand      xmm1, xmm3    // R in upper 5 bits
+    psllw     xmm2, 11      // B in upper 5 bits
+    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    psllw     xmm1, 8
+    por       xmm1, xmm2    // RB
+    pand      xmm0, xmm4    // G in middle 6 bits
+    pmulhuw   xmm0, xmm6    // << 5 * (256 + 4)
+    por       xmm0, xmm7    // AG
+    movdqa    xmm2, xmm1
+    punpcklbw xmm1, xmm0
+    punpckhbw xmm2, xmm0
+    movdqu    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqu    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+#ifdef HAS_RGB565TOARGBROW_AVX2
+// pmul method to replicate bits.
+// Math to replicate bits:
+// (v << 8) | (v << 3)
+// v * 256 + v * 8
+// v * (256 + 8)
+// G shift of 5 is incorporated, so shift is 5 + 8 and 5 + 3
+__declspec(naked)
+void RGB565ToARGBRow_AVX2(const uint8* src_rgb565, uint8* dst_argb,
+                          int pix) {
+  __asm {
+    mov        eax, 0x01080108  // generate multiplier to repeat 5 bits
+    vmovd      xmm5, eax
+    vbroadcastss ymm5, xmm5
+    mov        eax, 0x20802080  // multiplier shift by 5 and then repeat 6 bits
+    movd       xmm6, eax
+    vbroadcastss ymm6, xmm6
+    vpcmpeqb   ymm3, ymm3, ymm3       // generate mask 0xf800f800 for Red
+    vpsllw     ymm3, ymm3, 11
+    vpcmpeqb   ymm4, ymm4, ymm4       // generate mask 0x07e007e0 for Green
+    vpsllw     ymm4, ymm4, 10
+    vpsrlw     ymm4, ymm4, 5
+    vpcmpeqb   ymm7, ymm7, ymm7       // generate mask 0xff00ff00 for Alpha
+    vpsllw     ymm7, ymm7, 8
+
+    mov        eax, [esp + 4]   // src_rgb565
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // pix
+    sub        edx, eax
+    sub        edx, eax
+
+ convertloop:
+    vmovdqu    ymm0, [eax]   // fetch 16 pixels of bgr565
+    vpand      ymm1, ymm0, ymm3    // R in upper 5 bits
+    vpsllw     ymm2, ymm0, 11      // B in upper 5 bits
+    vpmulhuw   ymm1, ymm1, ymm5    // * (256 + 8)
+    vpmulhuw   ymm2, ymm2, ymm5    // * (256 + 8)
+    vpsllw     ymm1, ymm1, 8
+    vpor       ymm1, ymm1, ymm2    // RB
+    vpand      ymm0, ymm0, ymm4    // G in middle 6 bits
+    vpmulhuw   ymm0, ymm0, ymm6    // << 5 * (256 + 4)
+    vpor       ymm0, ymm0, ymm7    // AG
+    vpermq     ymm0, ymm0, 0xd8    // mutate for unpack
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpckhbw ymm2, ymm1, ymm0
+    vpunpcklbw ymm1, ymm1, ymm0
+    vmovdqu    [eax * 2 + edx], ymm1  // store 4 pixels of ARGB
+    vmovdqu    [eax * 2 + edx + 32], ymm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 32]
+    sub       ecx, 16
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_RGB565TOARGBROW_AVX2
+
+#ifdef HAS_ARGB1555TOARGBROW_AVX2
+__declspec(naked)
+void ARGB1555ToARGBRow_AVX2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov        eax, 0x01080108  // generate multiplier to repeat 5 bits
+    vmovd      xmm5, eax
+    vbroadcastss ymm5, xmm5
+    mov        eax, 0x42004200  // multiplier shift by 6 and then repeat 5 bits
+    movd       xmm6, eax
+    vbroadcastss ymm6, xmm6
+    vpcmpeqb   ymm3, ymm3, ymm3 // generate mask 0xf800f800 for Red
+    vpsllw     ymm3, ymm3, 11
+    vpsrlw     ymm4, ymm3, 6    // generate mask 0x03e003e0 for Green
+    vpcmpeqb   ymm7, ymm7, ymm7 // generate mask 0xff00ff00 for Alpha
+    vpsllw     ymm7, ymm7, 8
+
+    mov        eax,  [esp + 4]   // src_argb1555
+    mov        edx,  [esp + 8]   // dst_argb
+    mov        ecx,  [esp + 12]  // pix
+    sub        edx,  eax
+    sub        edx,  eax
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 16 pixels of 1555
+    vpsllw     ymm1, ymm0, 1       // R in upper 5 bits
+    vpsllw     ymm2, ymm0, 11      // B in upper 5 bits
+    vpand      ymm1, ymm1, ymm3
+    vpmulhuw   ymm2, ymm2, ymm5    // * (256 + 8)
+    vpmulhuw   ymm1, ymm1, ymm5    // * (256 + 8)
+    vpsllw     ymm1, ymm1, 8
+    vpor       ymm1, ymm1, ymm2    // RB
+    vpsraw     ymm2, ymm0, 8       // A
+    vpand      ymm0, ymm0, ymm4    // G in middle 5 bits
+    vpmulhuw   ymm0, ymm0, ymm6    // << 6 * (256 + 8)
+    vpand      ymm2, ymm2, ymm7
+    vpor       ymm0, ymm0, ymm2    // AG
+    vpermq     ymm0, ymm0, 0xd8    // mutate for unpack
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpckhbw ymm2, ymm1, ymm0
+    vpunpcklbw ymm1, ymm1, ymm0
+    vmovdqu    [eax * 2 + edx], ymm1  // store 8 pixels of ARGB
+    vmovdqu    [eax * 2 + edx + 32], ymm2  // store next 8 pixels of ARGB
+    lea       eax, [eax + 32]
+    sub       ecx, 16
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGB1555TOARGBROW_AVX2
+
+#ifdef HAS_ARGB4444TOARGBROW_AVX2
+__declspec(naked)
+void ARGB4444ToARGBRow_AVX2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax,  0x0f0f0f0f  // generate mask 0x0f0f0f0f
+    vmovd     xmm4, eax
+    vbroadcastss ymm4, xmm4
+    vpslld    ymm5, ymm4, 4     // 0xf0f0f0f0 for high nibbles
+    mov       eax,  [esp + 4]   // src_argb4444
+    mov       edx,  [esp + 8]   // dst_argb
+    mov       ecx,  [esp + 12]  // pix
+    sub       edx,  eax
+    sub       edx,  eax
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 16 pixels of bgra4444
+    vpand      ymm2, ymm0, ymm5    // mask high nibbles
+    vpand      ymm0, ymm0, ymm4    // mask low nibbles
+    vpsrlw     ymm3, ymm2, 4
+    vpsllw     ymm1, ymm0, 4
+    vpor       ymm2, ymm2, ymm3
+    vpor       ymm0, ymm0, ymm1
+    vpermq     ymm0, ymm0, 0xd8    // mutate for unpack
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpckhbw ymm1, ymm0, ymm2
+    vpunpcklbw ymm0, ymm0, ymm2
+    vmovdqu    [eax * 2 + edx], ymm0  // store 8 pixels of ARGB
+    vmovdqu    [eax * 2 + edx + 32], ymm1  // store next 8 pixels of ARGB
+    lea       eax, [eax + 32]
+    sub       ecx, 16
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGB4444TOARGBROW_AVX2
+
+// 24 instructions
+__declspec(naked)
+void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax, 0x01080108  // generate multiplier to repeat 5 bits
+    movd      xmm5, eax
+    pshufd    xmm5, xmm5, 0
+    mov       eax, 0x42004200  // multiplier shift by 6 and then repeat 5 bits
+    movd      xmm6, eax
+    pshufd    xmm6, xmm6, 0
+    pcmpeqb   xmm3, xmm3       // generate mask 0xf800f800 for Red
+    psllw     xmm3, 11
+    movdqa    xmm4, xmm3       // generate mask 0x03e003e0 for Green
+    psrlw     xmm4, 6
+    pcmpeqb   xmm7, xmm7       // generate mask 0xff00ff00 for Alpha
+    psllw     xmm7, 8
+
+    mov       eax, [esp + 4]   // src_argb1555
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of 1555
+    movdqa    xmm1, xmm0
+    movdqa    xmm2, xmm0
+    psllw     xmm1, 1       // R in upper 5 bits
+    psllw     xmm2, 11      // B in upper 5 bits
+    pand      xmm1, xmm3
+    pmulhuw   xmm2, xmm5    // * (256 + 8)
+    pmulhuw   xmm1, xmm5    // * (256 + 8)
+    psllw     xmm1, 8
+    por       xmm1, xmm2    // RB
+    movdqa    xmm2, xmm0
+    pand      xmm0, xmm4    // G in middle 5 bits
+    psraw     xmm2, 8       // A
+    pmulhuw   xmm0, xmm6    // << 6 * (256 + 8)
+    pand      xmm2, xmm7
+    por       xmm0, xmm2    // AG
+    movdqa    xmm2, xmm1
+    punpcklbw xmm1, xmm0
+    punpckhbw xmm2, xmm0
+    movdqu    [eax * 2 + edx], xmm1  // store 4 pixels of ARGB
+    movdqu    [eax * 2 + edx + 16], xmm2  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+// 18 instructions.
+__declspec(naked)
+void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
+                            int pix) {
+  __asm {
+    mov       eax, 0x0f0f0f0f  // generate mask 0x0f0f0f0f
+    movd      xmm4, eax
+    pshufd    xmm4, xmm4, 0
+    movdqa    xmm5, xmm4       // 0xf0f0f0f0 for high nibbles
+    pslld     xmm5, 4
+    mov       eax, [esp + 4]   // src_argb4444
+    mov       edx, [esp + 8]   // dst_argb
+    mov       ecx, [esp + 12]  // pix
+    sub       edx, eax
+    sub       edx, eax
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 8 pixels of bgra4444
+    movdqa    xmm2, xmm0
+    pand      xmm0, xmm4    // mask low nibbles
+    pand      xmm2, xmm5    // mask high nibbles
+    movdqa    xmm1, xmm0
+    movdqa    xmm3, xmm2
+    psllw     xmm1, 4
+    psrlw     xmm3, 4
+    por       xmm0, xmm1
+    por       xmm2, xmm3
+    movdqa    xmm1, xmm0
+    punpcklbw xmm0, xmm2
+    punpckhbw xmm1, xmm2
+    movdqu    [eax * 2 + edx], xmm0  // store 4 pixels of ARGB
+    movdqu    [eax * 2 + edx + 16], xmm1  // store next 4 pixels of ARGB
+    lea       eax, [eax + 16]
+    sub       ecx, 8
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    movdqa    xmm6, kShuffleMaskARGBToRGB24
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm2, [eax + 32]
+    movdqu    xmm3, [eax + 48]
+    lea       eax, [eax + 64]
+    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm1, xmm6
+    pshufb    xmm2, xmm6
+    pshufb    xmm3, xmm6
+    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
+    psrldq    xmm1, 4      // 8 bytes from 1
+    pslldq    xmm4, 12     // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
+    por       xmm0, xmm4   // 4 bytes from 1 for 0
+    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqu    [edx], xmm0  // store 0
+    por       xmm1, xmm5   // 8 bytes from 2 for 1
+    psrldq    xmm2, 8      // 4 bytes from 2
+    pslldq    xmm3, 4      // 12 bytes from 3 for 2
+    por       xmm2, xmm3   // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1   // store 1
+    movdqu    [edx + 32], xmm2   // store 2
+    lea       edx, [edx + 48]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    movdqa    xmm6, kShuffleMaskARGBToRAW
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 16 pixels of argb
+    movdqu    xmm1, [eax + 16]
+    movdqu    xmm2, [eax + 32]
+    movdqu    xmm3, [eax + 48]
+    lea       eax, [eax + 64]
+    pshufb    xmm0, xmm6    // pack 16 bytes of ARGB to 12 bytes of RGB
+    pshufb    xmm1, xmm6
+    pshufb    xmm2, xmm6
+    pshufb    xmm3, xmm6
+    movdqa    xmm4, xmm1   // 4 bytes from 1 for 0
+    psrldq    xmm1, 4      // 8 bytes from 1
+    pslldq    xmm4, 12     // 4 bytes from 1 for 0
+    movdqa    xmm5, xmm2   // 8 bytes from 2 for 1
+    por       xmm0, xmm4   // 4 bytes from 1 for 0
+    pslldq    xmm5, 8      // 8 bytes from 2 for 1
+    movdqu    [edx], xmm0  // store 0
+    por       xmm1, xmm5   // 8 bytes from 2 for 1
+    psrldq    xmm2, 8      // 4 bytes from 2
+    pslldq    xmm3, 4      // 12 bytes from 3 for 2
+    por       xmm2, xmm3   // 12 bytes from 3 for 2
+    movdqu    [edx + 16], xmm1   // store 1
+    movdqu    [edx + 32], xmm2   // store 2
+    lea       edx, [edx + 48]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+
+// 4 pixels
+__declspec(naked)
+void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm3, xmm3       // generate mask 0x0000001f
+    psrld     xmm3, 27
+    pcmpeqb   xmm4, xmm4       // generate mask 0x000007e0
+    psrld     xmm4, 26
+    pslld     xmm4, 5
+    pcmpeqb   xmm5, xmm5       // generate mask 0xfffff800
+    pslld     xmm5, 11
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    pslld     xmm0, 8       // R
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 5       // G
+    psrad     xmm0, 16      // R
+    pand      xmm1, xmm3    // B
+    pand      xmm2, xmm4    // G
+    pand      xmm0, xmm5    // R
+    por       xmm1, xmm2    // BG
+    por       xmm0, xmm1    // BGR
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of RGB565
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+// 8 pixels
+__declspec(naked)
+void ARGBToRGB565DitherRow_SSE2(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int pix) {
+  __asm {
+
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    movd      xmm6, [esp + 12] // dither4
+    mov       ecx, [esp + 16]  // pix
+    punpcklbw xmm6, xmm6       // make dither 16 bytes
+    movdqa    xmm7, xmm6
+    punpcklwd xmm6, xmm6
+    punpckhwd xmm7, xmm7
+    pcmpeqb   xmm3, xmm3       // generate mask 0x0000001f
+    psrld     xmm3, 27
+    pcmpeqb   xmm4, xmm4       // generate mask 0x000007e0
+    psrld     xmm4, 26
+    pslld     xmm4, 5
+    pcmpeqb   xmm5, xmm5       // generate mask 0xfffff800
+    pslld     xmm5, 11
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
+    paddusb   xmm0, xmm6    // add dither
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    pslld     xmm0, 8       // R
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 5       // G
+    psrad     xmm0, 16      // R
+    pand      xmm1, xmm3    // B
+    pand      xmm2, xmm4    // G
+    pand      xmm0, xmm5    // R
+    por       xmm1, xmm2    // BG
+    por       xmm0, xmm1    // BGR
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of RGB565
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTORGB565DITHERROW_AVX2
+__declspec(naked)
+void ARGBToRGB565DitherRow_AVX2(const uint8* src_argb, uint8* dst_rgb,
+                                const uint32 dither4, int pix) {
+  __asm {
+    mov        eax, [esp + 4]      // src_argb
+    mov        edx, [esp + 8]      // dst_rgb
+    vbroadcastss xmm6, [esp + 12]  // dither4
+    mov        ecx, [esp + 16]     // pix
+    vpunpcklbw xmm6, xmm6, xmm6    // make dither 32 bytes
+    vpermq     ymm6, ymm6, 0xd8
+    vpunpcklwd ymm6, ymm6, ymm6
+    vpcmpeqb   ymm3, ymm3, ymm3    // generate mask 0x0000001f
+    vpsrld     ymm3, ymm3, 27
+    vpcmpeqb   ymm4, ymm4, ymm4    // generate mask 0x000007e0
+    vpsrld     ymm4, ymm4, 26
+    vpslld     ymm4, ymm4, 5
+    vpslld     ymm5, ymm3, 11      // generate mask 0x0000f800
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpaddusb   ymm0, ymm0, ymm6    // add dither
+    vpsrld     ymm2, ymm0, 5       // G
+    vpsrld     ymm1, ymm0, 3       // B
+    vpsrld     ymm0, ymm0, 8       // R
+    vpand      ymm2, ymm2, ymm4    // G
+    vpand      ymm1, ymm1, ymm3    // B
+    vpand      ymm0, ymm0, ymm5    // R
+    vpor       ymm1, ymm1, ymm2    // BG
+    vpor       ymm0, ymm0, ymm1    // BGR
+    vpackusdw  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of RGB565
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTORGB565DITHERROW_AVX2
+
+// TODO(fbarchard): Improve sign extension/packing.
+__declspec(naked)
+void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm4, xmm4       // generate mask 0x0000001f
+    psrld     xmm4, 27
+    movdqa    xmm5, xmm4       // generate mask 0x000003e0
+    pslld     xmm5, 5
+    movdqa    xmm6, xmm4       // generate mask 0x00007c00
+    pslld     xmm6, 10
+    pcmpeqb   xmm7, xmm7       // generate mask 0xffff8000
+    pslld     xmm7, 15
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0    // B
+    movdqa    xmm2, xmm0    // G
+    movdqa    xmm3, xmm0    // R
+    psrad     xmm0, 16      // A
+    psrld     xmm1, 3       // B
+    psrld     xmm2, 6       // G
+    psrld     xmm3, 9       // R
+    pand      xmm0, xmm7    // A
+    pand      xmm1, xmm4    // B
+    pand      xmm2, xmm5    // G
+    pand      xmm3, xmm6    // R
+    por       xmm0, xmm1    // BA
+    por       xmm2, xmm3    // GR
+    por       xmm0, xmm2    // BGRA
+    packssdw  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB1555
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov       eax, [esp + 4]   // src_argb
+    mov       edx, [esp + 8]   // dst_rgb
+    mov       ecx, [esp + 12]  // pix
+    pcmpeqb   xmm4, xmm4       // generate mask 0xf000f000
+    psllw     xmm4, 12
+    movdqa    xmm3, xmm4       // generate mask 0x00f000f0
+    psrlw     xmm3, 8
+
+ convertloop:
+    movdqu    xmm0, [eax]   // fetch 4 pixels of argb
+    movdqa    xmm1, xmm0
+    pand      xmm0, xmm3    // low nibble
+    pand      xmm1, xmm4    // high nibble
+    psrld     xmm0, 4
+    psrld     xmm1, 8
+    por       xmm0, xmm1
+    packuswb  xmm0, xmm0
+    lea       eax, [eax + 16]
+    movq      qword ptr [edx], xmm0  // store 4 pixels of ARGB4444
+    lea       edx, [edx + 8]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTORGB565ROW_AVX2
+__declspec(naked)
+void ARGBToRGB565Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]      // src_argb
+    mov        edx, [esp + 8]      // dst_rgb
+    mov        ecx, [esp + 12]     // pix
+    vpcmpeqb   ymm3, ymm3, ymm3    // generate mask 0x0000001f
+    vpsrld     ymm3, ymm3, 27
+    vpcmpeqb   ymm4, ymm4, ymm4    // generate mask 0x000007e0
+    vpsrld     ymm4, ymm4, 26
+    vpslld     ymm4, ymm4, 5
+    vpslld     ymm5, ymm3, 11      // generate mask 0x0000f800
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpsrld     ymm2, ymm0, 5       // G
+    vpsrld     ymm1, ymm0, 3       // B
+    vpsrld     ymm0, ymm0, 8       // R
+    vpand      ymm2, ymm2, ymm4    // G
+    vpand      ymm1, ymm1, ymm3    // B
+    vpand      ymm0, ymm0, ymm5    // R
+    vpor       ymm1, ymm1, ymm2    // BG
+    vpor       ymm0, ymm0, ymm1    // BGR
+    vpackusdw  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of RGB565
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTORGB565ROW_AVX2
+
+#ifdef HAS_ARGBTOARGB1555ROW_AVX2
+__declspec(naked)
+void ARGBToARGB1555Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]      // src_argb
+    mov        edx, [esp + 8]      // dst_rgb
+    mov        ecx, [esp + 12]     // pix
+    vpcmpeqb   ymm4, ymm4, ymm4
+    vpsrld     ymm4, ymm4, 27      // generate mask 0x0000001f
+    vpslld     ymm5, ymm4, 5       // generate mask 0x000003e0
+    vpslld     ymm6, ymm4, 10      // generate mask 0x00007c00
+    vpcmpeqb   ymm7, ymm7, ymm7    // generate mask 0xffff8000
+    vpslld     ymm7, ymm7, 15
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpsrld     ymm3, ymm0, 9       // R
+    vpsrld     ymm2, ymm0, 6       // G
+    vpsrld     ymm1, ymm0, 3       // B
+    vpsrad     ymm0, ymm0, 16      // A
+    vpand      ymm3, ymm3, ymm6    // R
+    vpand      ymm2, ymm2, ymm5    // G
+    vpand      ymm1, ymm1, ymm4    // B
+    vpand      ymm0, ymm0, ymm7    // A
+    vpor       ymm0, ymm0, ymm1    // BA
+    vpor       ymm2, ymm2, ymm3    // GR
+    vpor       ymm0, ymm0, ymm2    // BGRA
+    vpackssdw  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of ARGB1555
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOARGB1555ROW_AVX2
+
+#ifdef HAS_ARGBTOARGB4444ROW_AVX2
+__declspec(naked)
+void ARGBToARGB4444Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb
+    mov        edx, [esp + 8]   // dst_rgb
+    mov        ecx, [esp + 12]  // pix
+    vpcmpeqb   ymm4, ymm4, ymm4   // generate mask 0xf000f000
+    vpsllw     ymm4, ymm4, 12
+    vpsrlw     ymm3, ymm4, 8      // generate mask 0x00f000f0
+
+ convertloop:
+    vmovdqu    ymm0, [eax]         // fetch 8 pixels of argb
+    vpand      ymm1, ymm0, ymm4    // high nibble
+    vpand      ymm0, ymm0, ymm3    // low nibble
+    vpsrld     ymm1, ymm1, 8
+    vpsrld     ymm0, ymm0, 4
+    vpor       ymm0, ymm0, ymm1
+    vpackuswb  ymm0, ymm0, ymm0
+    vpermq     ymm0, ymm0, 0xd8
+    lea        eax, [eax + 32]
+    vmovdqu    [edx], xmm0         // store 8 pixels of ARGB4444
+    lea        edx, [edx + 16]
+    sub        ecx, 8
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOARGB4444ROW_AVX2
+
+// Convert 16 ARGB pixels (64 bytes) to 16 Y values.
+__declspec(naked)
+void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kARGBToY
+    movdqa     xmm5, kAddY16
+
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+// Convert 16 ARGB pixels (64 bytes) to 16 YJ values.
+// Same as ARGBToYRow but different coefficients, no add 16, but do rounding.
+__declspec(naked)
+void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    paddw      xmm0, xmm5  // Add .5 for rounding.
+    paddw      xmm2, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTOYROW_AVX2
+// vpermd for vphaddw + vpackuswb vpermd.
+static const lvec32 kPermdARGBToY_AVX = {
+  0, 4, 1, 5, 2, 6, 3, 7
+};
+
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked)
+void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vbroadcastf128 ymm4, kARGBToY
+    vbroadcastf128 ymm5, kAddY16
+    vmovdqu    ymm6, kPermdARGBToY_AVX
+
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    vpaddb     ymm0, ymm0, ymm5  // add 16 for Y
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYROW_AVX2
+
+#ifdef HAS_ARGBTOYJROW_AVX2
+// Convert 32 ARGB pixels (128 bytes) to 32 Y values.
+__declspec(naked)
+void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    vbroadcastf128 ymm4, kARGBToYJ
+    vbroadcastf128 ymm5, kAddYJ64
+    vmovdqu    ymm6, kPermdARGBToY_AVX
+
+ convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpmaddubsw ymm0, ymm0, ymm4
+    vpmaddubsw ymm1, ymm1, ymm4
+    vpmaddubsw ymm2, ymm2, ymm4
+    vpmaddubsw ymm3, ymm3, ymm4
+    lea        eax, [eax + 128]
+    vphaddw    ymm0, ymm0, ymm1  // mutates.
+    vphaddw    ymm2, ymm2, ymm3
+    vpaddw     ymm0, ymm0, ymm5  // Add .5 for rounding.
+    vpaddw     ymm2, ymm2, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm2, ymm2, 7
+    vpackuswb  ymm0, ymm0, ymm2  // mutates.
+    vpermd     ymm0, ymm6, ymm0  // For vphaddw + vpackuswb mutation.
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_ARGBTOYJROW_AVX2
+
+__declspec(naked)
+void BGRAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kBGRAToY
+    movdqa     xmm5, kAddY16
+
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void ABGRToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kABGRToY
+    movdqa     xmm5, kAddY16
+
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void RGBAToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_y */
+    mov        ecx, [esp + 12]  /* pix */
+    movdqa     xmm4, kRGBAToY
+    movdqa     xmm5, kAddY16
+
+ convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm2, xmm4
+    pmaddubsw  xmm3, xmm4
+    lea        eax, [eax + 64]
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psrlw      xmm0, 7
+    psrlw      xmm2, 7
+    packuswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToUVJRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                        uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm5, kAddUVJ128
+    movdqa     xmm6, kARGBToVJ
+    movdqa     xmm7, kARGBToUJ
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    paddw      xmm0, xmm5            // +.5 rounding -> unsigned
+    paddw      xmm1, xmm5
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_ARGBTOUVROW_AVX2
+__declspec(naked)
+void ARGBToUVRow_AVX2(const uint8* src_argb0, int src_stride_argb,
+                      uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    vbroadcastf128 ymm5, kAddUV128
+    vbroadcastf128 ymm6, kARGBToV
+    vbroadcastf128 ymm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 32x2 argb pixels to 16x1 */
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vmovdqu    ymm2, [eax + 64]
+    vmovdqu    ymm3, [eax + 96]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    vpavgb     ymm2, ymm2, [eax + esi + 64]
+    vpavgb     ymm3, ymm3, [eax + esi + 96]
+    lea        eax,  [eax + 128]
+    vshufps    ymm4, ymm0, ymm1, 0x88
+    vshufps    ymm0, ymm0, ymm1, 0xdd
+    vpavgb     ymm0, ymm0, ymm4  // mutated by vshufps
+    vshufps    ymm4, ymm2, ymm3, 0x88
+    vshufps    ymm2, ymm2, ymm3, 0xdd
+    vpavgb     ymm2, ymm2, ymm4  // mutated by vshufps
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 32 different pixels, its 16 pixels of U and 16 of V
+    vpmaddubsw ymm1, ymm0, ymm7  // U
+    vpmaddubsw ymm3, ymm2, ymm7
+    vpmaddubsw ymm0, ymm0, ymm6  // V
+    vpmaddubsw ymm2, ymm2, ymm6
+    vphaddw    ymm1, ymm1, ymm3  // mutates
+    vphaddw    ymm0, ymm0, ymm2
+    vpsraw     ymm1, ymm1, 8
+    vpsraw     ymm0, ymm0, 8
+    vpacksswb  ymm0, ymm1, ymm0  // mutates
+    vpermq     ymm0, ymm0, 0xd8  // For vpacksswb
+    vpshufb    ymm0, ymm0, kShufARGBToUV_AVX  // For vshufps + vphaddw
+    vpaddb     ymm0, ymm0, ymm5  // -> unsigned
+
+    // step 3 - store 16 U and 16 V values
+    vextractf128 [edx], ymm0, 0 // U
+    vextractf128 [edx + edi], ymm0, 1 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBTOUVROW_AVX2
+
+__declspec(naked)
+void ARGBToUV444Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* convert to U and V */
+    movdqu     xmm0, [eax]          // U
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm7
+    pmaddubsw  xmm1, xmm7
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm3, xmm7
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    movdqu     [edx], xmm0
+
+    movdqu     xmm0, [eax]          // V
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    pmaddubsw  xmm0, xmm6
+    pmaddubsw  xmm1, xmm6
+    pmaddubsw  xmm2, xmm6
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm1
+    phaddw     xmm2, xmm3
+    psraw      xmm0, 8
+    psraw      xmm2, 8
+    packsswb   xmm0, xmm2
+    paddb      xmm0, xmm5
+    lea        eax,  [eax + 64]
+    movdqu     [edx + edi], xmm0
+    lea        edx,  [edx + 16]
+    sub        ecx,  16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked)
+void ARGBToUV422Row_SSSE3(const uint8* src_argb0,
+                          uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]   // src_argb
+    mov        edx, [esp + 4 + 8]   // dst_u
+    mov        edi, [esp + 4 + 12]  // dst_v
+    mov        ecx, [esp + 4 + 16]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kARGBToV
+    movdqa     xmm7, kARGBToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked)
+void BGRAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kBGRAToV
+    movdqa     xmm7, kBGRAToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void ABGRToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kABGRToV
+    movdqa     xmm7, kABGRToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void RGBAToUVRow_SSSE3(const uint8* src_argb0, int src_stride_argb,
+                       uint8* dst_u, uint8* dst_v, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb
+    mov        esi, [esp + 8 + 8]   // src_stride_argb
+    mov        edx, [esp + 8 + 12]  // dst_u
+    mov        edi, [esp + 8 + 16]  // dst_v
+    mov        ecx, [esp + 8 + 20]  // pix
+    movdqa     xmm5, kAddUV128
+    movdqa     xmm6, kRGBAToV
+    movdqa     xmm7, kRGBAToU
+    sub        edi, edx             // stride from u to v
+
+ convertloop:
+    /* step 1 - subsample 16x2 argb pixels to 8x1 */
+    movdqu     xmm0, [eax]
+    movdqu     xmm4, [eax + esi]
+    pavgb      xmm0, xmm4
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm4, [eax + esi + 16]
+    pavgb      xmm1, xmm4
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm4, [eax + esi + 32]
+    pavgb      xmm2, xmm4
+    movdqu     xmm3, [eax + 48]
+    movdqu     xmm4, [eax + esi + 48]
+    pavgb      xmm3, xmm4
+
+    lea        eax,  [eax + 64]
+    movdqa     xmm4, xmm0
+    shufps     xmm0, xmm1, 0x88
+    shufps     xmm4, xmm1, 0xdd
+    pavgb      xmm0, xmm4
+    movdqa     xmm4, xmm2
+    shufps     xmm2, xmm3, 0x88
+    shufps     xmm4, xmm3, 0xdd
+    pavgb      xmm2, xmm4
+
+    // step 2 - convert to U and V
+    // from here down is very similar to Y code except
+    // instead of 16 different pixels, its 8 pixels of U and 8 of V
+    movdqa     xmm1, xmm0
+    movdqa     xmm3, xmm2
+    pmaddubsw  xmm0, xmm7  // U
+    pmaddubsw  xmm2, xmm7
+    pmaddubsw  xmm1, xmm6  // V
+    pmaddubsw  xmm3, xmm6
+    phaddw     xmm0, xmm2
+    phaddw     xmm1, xmm3
+    psraw      xmm0, 8
+    psraw      xmm1, 8
+    packsswb   xmm0, xmm1
+    paddb      xmm0, xmm5            // -> unsigned
+
+    // step 3 - store 8 U and 8 V values
+    movlps     qword ptr [edx], xmm0 // U
+    movhps     qword ptr [edx + edi], xmm0 // V
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBTOYROW_SSSE3
+
+// Read 16 UV from 444
+#define READYUV444_AVX2 __asm {                                                \
+    __asm vmovdqu    xmm0, [esi]                  /* U */         /* NOLINT */ \
+    __asm vmovdqu    xmm1, [esi + edi]            /* V */         /* NOLINT */ \
+    __asm lea        esi,  [esi + 16]                                          \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpermq     ymm1, ymm1, 0xd8                                          \
+    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
+  }
+
+// Read 8 UV from 422, upsample to 16 UV.
+#define READYUV422_AVX2 __asm {                                                \
+    __asm vmovq      xmm0, qword ptr [esi]        /* U */         /* NOLINT */ \
+    __asm vmovq      xmm1, qword ptr [esi + edi]  /* V */         /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
+  }
+
+// Read 4 UV from 411, upsample to 16 UV.
+#define READYUV411_AVX2 __asm {                                                \
+    __asm vmovd      xmm0, dword ptr [esi]        /* U */         /* NOLINT */ \
+    __asm vmovd      xmm1, dword ptr [esi + edi]  /* V */         /* NOLINT */ \
+    __asm lea        esi,  [esi + 4]                                           \
+    __asm vpunpcklbw ymm0, ymm0, ymm1             /* UV */                     \
+    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpckldq ymm0, ymm0, ymm0             /* UVUVUVUV (upsample) */    \
+  }
+
+// Read 8 UV from NV12, upsample to 16 UV.
+#define READNV12_AVX2 __asm {                                                  \
+    __asm vmovdqu    xmm0, [esi]                  /* UV */                     \
+    __asm lea        esi,  [esi + 16]                                          \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpcklwd ymm0, ymm0, ymm0             /* UVUV (upsample) */        \
+  }
+
+// Convert 16 pixels: 16 UV and 16 Y.
+#define YUVTORGB_AVX2(YuvConstants) __asm {                                    \
+    /* Step 1: Find 8 UV contributions to 16 R,G,B values */                   \
+    __asm vpmaddubsw ymm2, ymm0, YuvConstants.kUVToR        /* scale R UV */   \
+    __asm vpmaddubsw ymm1, ymm0, YuvConstants.kUVToG        /* scale G UV */   \
+    __asm vpmaddubsw ymm0, ymm0, YuvConstants.kUVToB        /* scale B UV */   \
+    __asm vmovdqu    ymm3, YuvConstants.kUVBiasR                               \
+    __asm vpsubw     ymm2, ymm3, ymm2                                          \
+    __asm vmovdqu    ymm3, YuvConstants.kUVBiasG                               \
+    __asm vpsubw     ymm1, ymm3, ymm1                                          \
+    __asm vmovdqu    ymm3, YuvConstants.kUVBiasB                               \
+    __asm vpsubw     ymm0, ymm3, ymm0                                          \
+    /* Step 2: Find Y contribution to 16 R,G,B values */                       \
+    __asm vmovdqu    xmm3, [eax]                  /* NOLINT */                 \
+    __asm lea        eax, [eax + 16]                                           \
+    __asm vpermq     ymm3, ymm3, 0xd8                                          \
+    __asm vpunpcklbw ymm3, ymm3, ymm3                                          \
+    __asm vpmulhuw   ymm3, ymm3, YuvConstants.kYToRgb                          \
+    __asm vpaddsw    ymm0, ymm0, ymm3           /* B += Y */                   \
+    __asm vpaddsw    ymm1, ymm1, ymm3           /* G += Y */                   \
+    __asm vpaddsw    ymm2, ymm2, ymm3           /* R += Y */                   \
+    __asm vpsraw     ymm0, ymm0, 6                                             \
+    __asm vpsraw     ymm1, ymm1, 6                                             \
+    __asm vpsraw     ymm2, ymm2, 6                                             \
+    __asm vpackuswb  ymm0, ymm0, ymm0           /* B */                        \
+    __asm vpackuswb  ymm1, ymm1, ymm1           /* G */                        \
+    __asm vpackuswb  ymm2, ymm2, ymm2           /* R */                        \
+  }
+
+// Store 16 ARGB values.
+#define STOREARGB_AVX2 __asm {                                                 \
+    /* Step 3: Weave into ARGB */                                              \
+    __asm vpunpcklbw ymm0, ymm0, ymm1           /* BG */                       \
+    __asm vpermq     ymm0, ymm0, 0xd8                                          \
+    __asm vpunpcklbw ymm2, ymm2, ymm5           /* RA */                       \
+    __asm vpermq     ymm2, ymm2, 0xd8                                          \
+    __asm vpunpcklwd ymm1, ymm0, ymm2           /* BGRA first 8 pixels */      \
+    __asm vpunpckhwd ymm0, ymm0, ymm2           /* BGRA next 8 pixels */       \
+    __asm vmovdqu    0[edx], ymm1                                              \
+    __asm vmovdqu    32[edx], ymm0                                             \
+    __asm lea        edx,  [edx + 64]                                          \
+  }
+
+#ifdef HAS_I422TOARGBROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void I422ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I422TOARGBROW_AVX2
+
+#ifdef HAS_J422TOARGBROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void J422ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvJConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_J422TOARGBROW_AVX2
+
+#ifdef HAS_I444TOARGBROW_AVX2
+// 16 pixels
+// 16 UV values with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void I444ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV444_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I444TOARGBROW_AVX2
+
+#ifdef HAS_I411TOARGBROW_AVX2
+// 16 pixels
+// 4 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void I411ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV411_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I411TOARGBROW_AVX2
+
+#ifdef HAS_NV12TOARGBROW_AVX2
+// 16 pixels.
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void NV12ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READNV12_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_NV12TOARGBROW_AVX2
+
+#ifdef HAS_NV21TOARGBROW_AVX2
+// 16 pixels.
+// 8 VU values upsampled to 16 VU, mixed with 16 Y producing 16 ARGB (64 bytes).
+__declspec(naked)
+void NV21ToARGBRow_AVX2(const uint8* y_buf,
+                        const uint8* uv_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READNV12_AVX2
+    YUVTORGB_AVX2(kYvuConstants)
+    STOREARGB_AVX2
+
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_NV21TOARGBROW_AVX2
+
+#ifdef HAS_I422TOBGRAROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 BGRA (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
+__declspec(naked)
+void I422ToBGRARow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into BGRA
+    vpunpcklbw ymm1, ymm1, ymm0           // GB
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm2, ymm5, ymm2           // AR
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklwd ymm0, ymm2, ymm1           // ARGB first 8 pixels
+    vpunpckhwd ymm2, ymm2, ymm1           // ARGB next 8 pixels
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm2
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I422TOBGRAROW_AVX2
+
+#ifdef HAS_I422TORGBAROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
+__declspec(naked)
+void I422ToRGBARow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into RGBA
+    vpunpcklbw ymm1, ymm1, ymm2           // GR
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm2, ymm5, ymm0           // AB
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklwd ymm0, ymm2, ymm1           // ABGR first 8 pixels
+    vpunpckhwd ymm1, ymm2, ymm1           // ABGR next 8 pixels
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I422TORGBAROW_AVX2
+
+#ifdef HAS_I422TOABGRROW_AVX2
+// 16 pixels
+// 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ABGR (64 bytes).
+// TODO(fbarchard): Use macros to reduce duplicate code.  See SSSE3.
+__declspec(naked)
+void I422ToABGRRow_AVX2(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_argb,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    vpcmpeqb   ymm5, ymm5, ymm5     // generate 0xffffffffffffffff for alpha
+
+ convertloop:
+    READYUV422_AVX2
+    YUVTORGB_AVX2(kYuvConstants)
+
+    // Step 3: Weave into ABGR
+    vpunpcklbw ymm1, ymm2, ymm1           // RG
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklbw ymm2, ymm0, ymm5           // BA
+    vpermq     ymm2, ymm2, 0xd8
+    vpunpcklwd ymm0, ymm1, ymm2           // RGBA first 8 pixels
+    vpunpckhwd ymm1, ymm1, ymm2           // RGBA next 8 pixels
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I422TOABGRROW_AVX2
+
+#if defined(HAS_I422TOARGBROW_SSSE3)
+// TODO(fbarchard): Read that does half size on Y and treats 420 as 444.
+
+// Read 8 UV from 444.
+#define READYUV444 __asm {                                                     \
+    __asm movq       xmm0, qword ptr [esi] /* U */                /* NOLINT */ \
+    __asm movq       xmm1, qword ptr [esi + edi] /* V */          /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+  }
+
+// Read 4 UV from 422, upsample to 8 UV.
+#define READYUV422 __asm {                                                     \
+    __asm movd       xmm0, [esi]          /* U */                              \
+    __asm movd       xmm1, [esi + edi]    /* V */                              \
+    __asm lea        esi,  [esi + 4]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Read 2 UV from 411, upsample to 8 UV.
+#define READYUV411 __asm {                                                     \
+    __asm movzx      ebx, word ptr [esi]        /* U */           /* NOLINT */ \
+    __asm movd       xmm0, ebx                                                 \
+    __asm movzx      ebx, word ptr [esi + edi]  /* V */           /* NOLINT */ \
+    __asm movd       xmm1, ebx                                                 \
+    __asm lea        esi,  [esi + 2]                                           \
+    __asm punpcklbw  xmm0, xmm1           /* UV */                             \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+    __asm punpckldq  xmm0, xmm0           /* UVUVUVUV (upsample) */            \
+  }
+
+// Read 4 UV from NV12, upsample to 8 UV.
+#define READNV12 __asm {                                                       \
+    __asm movq       xmm0, qword ptr [esi] /* UV */               /* NOLINT */ \
+    __asm lea        esi,  [esi + 8]                                           \
+    __asm punpcklwd  xmm0, xmm0           /* UVUV (upsample) */                \
+  }
+
+// Convert 8 pixels: 8 UV and 8 Y.
+#define YUVTORGB(YuvConstants) __asm {                                         \
+    /* Step 1: Find 4 UV contributions to 8 R,G,B values */                    \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm movdqa     xmm2, xmm0                                                \
+    __asm movdqa     xmm3, xmm0                                                \
+    __asm movdqa     xmm0, YuvConstants.kUVBiasB /* unbias back to signed */   \
+    __asm pmaddubsw  xmm1, YuvConstants.kUVToB   /* scale B UV */              \
+    __asm psubw      xmm0, xmm1                                                \
+    __asm movdqa     xmm1, YuvConstants.kUVBiasG                               \
+    __asm pmaddubsw  xmm2, YuvConstants.kUVToG   /* scale G UV */              \
+    __asm psubw      xmm1, xmm2                                                \
+    __asm movdqa     xmm2, YuvConstants.kUVBiasR                               \
+    __asm pmaddubsw  xmm3, YuvConstants.kUVToR   /* scale R UV */              \
+    __asm psubw      xmm2, xmm3                                                \
+    /* Step 2: Find Y contribution to 8 R,G,B values */                        \
+    __asm movq       xmm3, qword ptr [eax]                        /* NOLINT */ \
+    __asm lea        eax, [eax + 8]                                            \
+    __asm punpcklbw  xmm3, xmm3                                                \
+    __asm pmulhuw    xmm3, YuvConstants.kYToRgb                                \
+    __asm paddsw     xmm0, xmm3           /* B += Y */                         \
+    __asm paddsw     xmm1, xmm3           /* G += Y */                         \
+    __asm paddsw     xmm2, xmm3           /* R += Y */                         \
+    __asm psraw      xmm0, 6                                                   \
+    __asm psraw      xmm1, 6                                                   \
+    __asm psraw      xmm2, 6                                                   \
+    __asm packuswb   xmm0, xmm0           /* B */                              \
+    __asm packuswb   xmm1, xmm1           /* G */                              \
+    __asm packuswb   xmm2, xmm2           /* R */                              \
+  }
+
+// Store 8 ARGB values.
+#define STOREARGB __asm {                                                      \
+    /* Step 3: Weave into ARGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm5           /* RA */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRA first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRA next 4 pixels */             \
+    __asm movdqu     0[edx], xmm0                                              \
+    __asm movdqu     16[edx], xmm1                                             \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 BGRA values.
+#define STOREBGRA __asm {                                                      \
+    /* Step 3: Weave into BGRA */                                              \
+    __asm pcmpeqb    xmm5, xmm5           /* generate 0xffffffff for alpha */  \
+    __asm punpcklbw  xmm1, xmm0           /* GB */                             \
+    __asm punpcklbw  xmm5, xmm2           /* AR */                             \
+    __asm movdqa     xmm0, xmm5                                                \
+    __asm punpcklwd  xmm5, xmm1           /* BGRA first 4 pixels */            \
+    __asm punpckhwd  xmm0, xmm1           /* BGRA next 4 pixels */             \
+    __asm movdqu     0[edx], xmm5                                              \
+    __asm movdqu     16[edx], xmm0                                             \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 ABGR values.
+#define STOREABGR __asm {                                                      \
+    /* Step 3: Weave into ABGR */                                              \
+    __asm punpcklbw  xmm2, xmm1           /* RG */                             \
+    __asm punpcklbw  xmm0, xmm5           /* BA */                             \
+    __asm movdqa     xmm1, xmm2                                                \
+    __asm punpcklwd  xmm2, xmm0           /* RGBA first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm0           /* RGBA next 4 pixels */             \
+    __asm movdqu     0[edx], xmm2                                              \
+    __asm movdqu     16[edx], xmm1                                             \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 RGBA values.
+#define STORERGBA __asm {                                                      \
+    /* Step 3: Weave into RGBA */                                              \
+    __asm pcmpeqb    xmm5, xmm5           /* generate 0xffffffff for alpha */  \
+    __asm punpcklbw  xmm1, xmm2           /* GR */                             \
+    __asm punpcklbw  xmm5, xmm0           /* AB */                             \
+    __asm movdqa     xmm0, xmm5                                                \
+    __asm punpcklwd  xmm5, xmm1           /* RGBA first 4 pixels */            \
+    __asm punpckhwd  xmm0, xmm1           /* RGBA next 4 pixels */             \
+    __asm movdqu     0[edx], xmm5                                              \
+    __asm movdqu     16[edx], xmm0                                             \
+    __asm lea        edx,  [edx + 32]                                          \
+  }
+
+// Store 8 RGB24 values.
+#define STORERGB24 __asm {                                                     \
+    /* Step 3: Weave into RRGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
+    /* Step 4: RRGB -> RGB24 */                                                \
+    __asm pshufb     xmm0, xmm5           /* Pack first 8 and last 4 bytes. */ \
+    __asm pshufb     xmm1, xmm6           /* Pack first 12 bytes. */           \
+    __asm palignr    xmm1, xmm0, 12       /* last 4 bytes of xmm0 + 12 xmm1 */ \
+    __asm movq       qword ptr 0[edx], xmm0  /* First 8 bytes */               \
+    __asm movdqu     8[edx], xmm1         /* Last 16 bytes */                  \
+    __asm lea        edx,  [edx + 24]                                          \
+  }
+
+// Store 8 RAW values.
+#define STORERAW __asm {                                                       \
+    /* Step 3: Weave into RRGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
+    /* Step 4: RRGB -> RAW */                                                  \
+    __asm pshufb     xmm0, xmm5           /* Pack first 8 and last 4 bytes. */ \
+    __asm pshufb     xmm1, xmm6           /* Pack first 12 bytes. */           \
+    __asm palignr    xmm1, xmm0, 12       /* last 4 bytes of xmm0 + 12 xmm1 */ \
+    __asm movq       qword ptr 0[edx], xmm0  /* First 8 bytes */               \
+    __asm movdqu     8[edx], xmm1         /* Last 16 bytes */                  \
+    __asm lea        edx,  [edx + 24]                                          \
+  }
+
+// Store 8 RGB565 values.
+#define STORERGB565 __asm {                                                    \
+    /* Step 3: Weave into RRGB */                                              \
+    __asm punpcklbw  xmm0, xmm1           /* BG */                             \
+    __asm punpcklbw  xmm2, xmm2           /* RR */                             \
+    __asm movdqa     xmm1, xmm0                                                \
+    __asm punpcklwd  xmm0, xmm2           /* BGRR first 4 pixels */            \
+    __asm punpckhwd  xmm1, xmm2           /* BGRR next 4 pixels */             \
+    /* Step 4: RRGB -> RGB565 */                                               \
+    __asm movdqa     xmm3, xmm0    /* B  first 4 pixels of argb */             \
+    __asm movdqa     xmm2, xmm0    /* G */                                     \
+    __asm pslld      xmm0, 8       /* R */                                     \
+    __asm psrld      xmm3, 3       /* B */                                     \
+    __asm psrld      xmm2, 5       /* G */                                     \
+    __asm psrad      xmm0, 16      /* R */                                     \
+    __asm pand       xmm3, xmm5    /* B */                                     \
+    __asm pand       xmm2, xmm6    /* G */                                     \
+    __asm pand       xmm0, xmm7    /* R */                                     \
+    __asm por        xmm3, xmm2    /* BG */                                    \
+    __asm por        xmm0, xmm3    /* BGR */                                   \
+    __asm movdqa     xmm3, xmm1    /* B  next 4 pixels of argb */              \
+    __asm movdqa     xmm2, xmm1    /* G */                                     \
+    __asm pslld      xmm1, 8       /* R */                                     \
+    __asm psrld      xmm3, 3       /* B */                                     \
+    __asm psrld      xmm2, 5       /* G */                                     \
+    __asm psrad      xmm1, 16      /* R */                                     \
+    __asm pand       xmm3, xmm5    /* B */                                     \
+    __asm pand       xmm2, xmm6    /* G */                                     \
+    __asm pand       xmm1, xmm7    /* R */                                     \
+    __asm por        xmm3, xmm2    /* BG */                                    \
+    __asm por        xmm1, xmm3    /* BGR */                                   \
+    __asm packssdw   xmm0, xmm1                                                \
+    __asm movdqu     0[edx], xmm0  /* store 8 pixels of RGB565 */              \
+    __asm lea        edx, [edx + 16]                                           \
+  }
+
+// 8 pixels.
+// 8 UV values, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked)
+void I444ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READYUV444
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB24 (24 bytes).
+__declspec(naked)
+void I422ToRGB24Row_SSSE3(const uint8* y_buf,
+                          const uint8* u_buf,
+                          const uint8* v_buf,
+                          uint8* dst_rgb24,
+                          int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgb24
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    movdqa     xmm5, kShuffleMaskARGBToRGB24_0
+    movdqa     xmm6, kShuffleMaskARGBToRGB24
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STORERGB24
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RAW (24 bytes).
+__declspec(naked)
+void I422ToRAWRow_SSSE3(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* dst_raw,
+                        int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // raw
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    movdqa     xmm5, kShuffleMaskARGBToRAW_0
+    movdqa     xmm6, kShuffleMaskARGBToRAW
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STORERAW
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 RGB565 (16 bytes).
+__declspec(naked)
+void I422ToRGB565Row_SSSE3(const uint8* y_buf,
+                           const uint8* u_buf,
+                           const uint8* v_buf,
+                           uint8* rgb565_buf,
+                           int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgb565
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5       // generate mask 0x0000001f
+    psrld      xmm5, 27
+    pcmpeqb    xmm6, xmm6       // generate mask 0x000007e0
+    psrld      xmm6, 26
+    pslld      xmm6, 5
+    pcmpeqb    xmm7, xmm7       // generate mask 0xfffff800
+    pslld      xmm7, 11
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STORERGB565
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked)
+void I422ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// JPeg color space version of I422ToARGB
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked)
+void J422ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // argb
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvJConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// 2 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+// Similar to I420 but duplicate UV once more.
+__declspec(naked)
+void I411ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]   // Y
+    mov        esi, [esp + 12 + 8]   // U
+    mov        edi, [esp + 12 + 12]  // V
+    mov        edx, [esp + 12 + 16]  // argb
+    mov        ecx, [esp + 12 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5            // generate 0xffffffff for alpha
+
+ convertloop:
+    READYUV411  // modifies EBX
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// 8 pixels.
+// 4 UV values upsampled to 8 UV, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked)
+void NV12ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* uv_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READNV12
+    YUVTORGB(kYuvConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 8 pixels.
+// 4 VU values upsampled to 8 VU, mixed with 8 Y producing 8 ARGB (32 bytes).
+__declspec(naked)
+void NV21ToARGBRow_SSSE3(const uint8* y_buf,
+                         const uint8* uv_buf,
+                         uint8* dst_argb,
+                         int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // Y
+    mov        esi, [esp + 4 + 8]   // UV
+    mov        edx, [esp + 4 + 12]  // argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READNV12
+    YUVTORGB(kYvuConstants)
+    STOREARGB
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void I422ToBGRARow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_bgra,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // bgra
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREBGRA
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void I422ToABGRRow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_abgr,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // abgr
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+    pcmpeqb    xmm5, xmm5           // generate 0xffffffff for alpha
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STOREABGR
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void I422ToRGBARow_SSSE3(const uint8* y_buf,
+                         const uint8* u_buf,
+                         const uint8* v_buf,
+                         uint8* dst_rgba,
+                         int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // Y
+    mov        esi, [esp + 8 + 8]   // U
+    mov        edi, [esp + 8 + 12]  // V
+    mov        edx, [esp + 8 + 16]  // rgba
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        edi, esi
+
+ convertloop:
+    READYUV422
+    YUVTORGB(kYuvConstants)
+    STORERGBA
+
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#endif  // HAS_I422TOARGBROW_SSSE3
+
+#ifdef HAS_I400TOARGBROW_SSE2
+// 8 pixels of Y converted to 8 pixels of ARGB (32 bytes).
+__declspec(naked)
+void I400ToARGBRow_SSE2(const uint8* y_buf,
+                        uint8* rgb_buf,
+                        int width) {
+  __asm {
+    mov        eax, 0x4a354a35      // 4a35 = 18997 = round(1.164 * 64 * 256)
+    movd       xmm2, eax
+    pshufd     xmm2, xmm2,0
+    mov        eax, 0x04880488      // 0488 = 1160 = round(1.164 * 64 * 16)
+    movd       xmm3, eax
+    pshufd     xmm3, xmm3, 0
+    pcmpeqb    xmm4, xmm4           // generate mask 0xff000000
+    pslld      xmm4, 24
+
+    mov        eax, [esp + 4]       // Y
+    mov        edx, [esp + 8]       // rgb
+    mov        ecx, [esp + 12]      // width
+
+ convertloop:
+    // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164
+    movq       xmm0, qword ptr [eax]
+    lea        eax, [eax + 8]
+    punpcklbw  xmm0, xmm0           // Y.Y
+    pmulhuw    xmm0, xmm2
+    psubusw    xmm0, xmm3
+    psrlw      xmm0, 6
+    packuswb   xmm0, xmm0           // G
+
+    // Step 2: Weave into ARGB
+    punpcklbw  xmm0, xmm0           // GG
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm0           // BGRA first 4 pixels
+    punpckhwd  xmm1, xmm1           // BGRA next 4 pixels
+    por        xmm0, xmm4
+    por        xmm1, xmm4
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx,  [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_I400TOARGBROW_SSE2
+
+#ifdef HAS_I400TOARGBROW_AVX2
+// 16 pixels of Y converted to 16 pixels of ARGB (64 bytes).
+// note: vpunpcklbw mutates and vpackuswb unmutates.
+__declspec(naked)
+void I400ToARGBRow_AVX2(const uint8* y_buf,
+                        uint8* rgb_buf,
+                        int width) {
+  __asm {
+    mov        eax, 0x4a354a35      // 4a35 = 18997 = round(1.164 * 64 * 256)
+    vmovd      xmm2, eax
+    vbroadcastss ymm2, xmm2
+    mov        eax, 0x04880488      // 0488 = 1160 = round(1.164 * 64 * 16)
+    vmovd      xmm3, eax
+    vbroadcastss ymm3, xmm3
+    vpcmpeqb   ymm4, ymm4, ymm4     // generate mask 0xff000000
+    vpslld     ymm4, ymm4, 24
+
+    mov        eax, [esp + 4]       // Y
+    mov        edx, [esp + 8]       // rgb
+    mov        ecx, [esp + 12]      // width
+
+ convertloop:
+    // Step 1: Scale Y contriportbution to 16 G values. G = (y - 16) * 1.164
+    vmovdqu    xmm0, [eax]
+    lea        eax, [eax + 16]
+    vpermq     ymm0, ymm0, 0xd8           // vpunpcklbw mutates
+    vpunpcklbw ymm0, ymm0, ymm0           // Y.Y
+    vpmulhuw   ymm0, ymm0, ymm2
+    vpsubusw   ymm0, ymm0, ymm3
+    vpsrlw     ymm0, ymm0, 6
+    vpackuswb  ymm0, ymm0, ymm0           // G.  still mutated: 3120
+
+    // TODO(fbarchard): Weave alpha with unpack.
+    // Step 2: Weave into ARGB
+    vpunpcklbw ymm1, ymm0, ymm0           // GG - mutates
+    vpermq     ymm1, ymm1, 0xd8
+    vpunpcklwd ymm0, ymm1, ymm1           // GGGG first 8 pixels
+    vpunpckhwd ymm1, ymm1, ymm1           // GGGG next 8 pixels
+    vpor       ymm0, ymm0, ymm4
+    vpor       ymm1, ymm1, ymm4
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx,  [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_I400TOARGBROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSSE3
+// Shuffle table for reversing the bytes.
+static const uvec8 kShuffleMirror = {
+  15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+// TODO(fbarchard): Replace lea with -16 offset.
+__declspec(naked)
+void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    movdqa    xmm5, kShuffleMirror
+
+ convertloop:
+    movdqu    xmm0, [eax - 16 + ecx]
+    pshufb    xmm0, xmm5
+    movdqu    [edx], xmm0
+    lea       edx, [edx + 16]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_SSSE3
+
+#ifdef HAS_MIRRORROW_AVX2
+__declspec(naked)
+void MirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    vbroadcastf128 ymm5, kShuffleMirror
+
+ convertloop:
+    vmovdqu   ymm0, [eax - 32 + ecx]
+    vpshufb   ymm0, ymm0, ymm5
+    vpermq    ymm0, ymm0, 0x4e  // swap high and low halfs
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    sub       ecx, 32
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_AVX2
+
+#ifdef HAS_MIRRORROW_SSE2
+__declspec(naked)
+void MirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+
+ convertloop:
+    movdqu    xmm0, [eax - 16 + ecx]
+    movdqa    xmm1, xmm0        // swap bytes
+    psllw     xmm0, 8
+    psrlw     xmm1, 8
+    por       xmm0, xmm1
+    pshuflw   xmm0, xmm0, 0x1b  // swap words
+    pshufhw   xmm0, xmm0, 0x1b
+    pshufd    xmm0, xmm0, 0x4e  // swap qwords
+    movdqu    [edx], xmm0
+    lea       edx, [edx + 16]
+    sub       ecx, 16
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_SSE2
+
+#ifdef HAS_MIRRORROW_UV_SSSE3
+// Shuffle table for reversing the bytes of UV channels.
+static const uvec8 kShuffleMirrorUV = {
+  14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u
+};
+
+__declspec(naked)
+void MirrorUVRow_SSSE3(const uint8* src, uint8* dst_u, uint8* dst_v,
+                       int width) {
+  __asm {
+    push      edi
+    mov       eax, [esp + 4 + 4]   // src
+    mov       edx, [esp + 4 + 8]   // dst_u
+    mov       edi, [esp + 4 + 12]  // dst_v
+    mov       ecx, [esp + 4 + 16]  // width
+    movdqa    xmm1, kShuffleMirrorUV
+    lea       eax, [eax + ecx * 2 - 16]
+    sub       edi, edx
+
+ convertloop:
+    movdqu    xmm0, [eax]
+    lea       eax, [eax - 16]
+    pshufb    xmm0, xmm1
+    movlpd    qword ptr [edx], xmm0
+    movhpd    qword ptr [edx + edi], xmm0
+    lea       edx, [edx + 8]
+    sub       ecx, 8
+    jg        convertloop
+
+    pop       edi
+    ret
+  }
+}
+#endif  // HAS_MIRRORROW_UV_SSSE3
+
+#ifdef HAS_ARGBMIRRORROW_SSE2
+__declspec(naked)
+void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    lea       eax, [eax - 16 + ecx * 4]  // last 4 pixels.
+
+ convertloop:
+    movdqu    xmm0, [eax]
+    lea       eax, [eax - 16]
+    pshufd    xmm0, xmm0, 0x1b
+    movdqu    [edx], xmm0
+    lea       edx, [edx + 16]
+    sub       ecx, 4
+    jg        convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_SSE2
+
+#ifdef HAS_ARGBMIRRORROW_AVX2
+// Shuffle table for reversing the bytes.
+static const ulvec32 kARGBShuffleMirror_AVX2 = {
+  7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u
+};
+
+__declspec(naked)
+void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov       eax, [esp + 4]   // src
+    mov       edx, [esp + 8]   // dst
+    mov       ecx, [esp + 12]  // width
+    vmovdqu   ymm5, kARGBShuffleMirror_AVX2
+
+ convertloop:
+    vpermd    ymm0, ymm5, [eax - 32 + ecx * 4]  // permute dword order
+    vmovdqu   [edx], ymm0
+    lea       edx, [edx + 32]
+    sub       ecx, 8
+    jg        convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBMIRRORROW_AVX2
+
+#ifdef HAS_SPLITUVROW_SSE2
+__declspec(naked)
+void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    movdqa     xmm3, xmm1
+    pand       xmm0, xmm5   // even bytes
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm2, 8      // odd bytes
+    psrlw      xmm3, 8
+    packuswb   xmm2, xmm3
+    movdqu     [edx], xmm0
+    movdqu     [edx + edi], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+#endif  // HAS_SPLITUVROW_SSE2
+
+#ifdef HAS_SPLITUVROW_AVX2
+__declspec(naked)
+void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_uv
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm2, ymm0, 8      // odd bytes
+    vpsrlw     ymm3, ymm1, 8
+    vpand      ymm0, ymm0, ymm5   // even bytes
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1
+    vpackuswb  ymm2, ymm2, ymm3
+    vpermq     ymm0, ymm0, 0xd8
+    vpermq     ymm2, ymm2, 0xd8
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + edi], ymm2
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SPLITUVROW_AVX2
+
+#ifdef HAS_MERGEUVROW_SSE2
+__declspec(naked)
+void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+  convertloop:
+    movdqu     xmm0, [eax]      // read 16 U's
+    movdqu     xmm1, [eax + edx]  // and 16 V's
+    lea        eax,  [eax + 16]
+    movdqa     xmm2, xmm0
+    punpcklbw  xmm0, xmm1       // first 8 UV pairs
+    punpckhbw  xmm2, xmm1       // next 8 UV pairs
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  //  HAS_MERGEUVROW_SSE2
+
+#ifdef HAS_MERGEUVROW_AVX2
+__declspec(naked)
+void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+                     int width) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_u
+    mov        edx, [esp + 4 + 8]    // src_v
+    mov        edi, [esp + 4 + 12]   // dst_uv
+    mov        ecx, [esp + 4 + 16]   // width
+    sub        edx, eax
+
+  convertloop:
+    vmovdqu    ymm0, [eax]           // read 32 U's
+    vmovdqu    ymm1, [eax + edx]     // and 32 V's
+    lea        eax,  [eax + 32]
+    vpunpcklbw ymm2, ymm0, ymm1      // low 16 UV pairs. mutated qqword 0,2
+    vpunpckhbw ymm0, ymm0, ymm1      // high 16 UV pairs. mutated qqword 1,3
+    vextractf128 [edi], ymm2, 0       // bytes 0..15
+    vextractf128 [edi + 16], ymm0, 0  // bytes 16..31
+    vextractf128 [edi + 32], ymm2, 1  // bytes 32..47
+    vextractf128 [edi + 48], ymm0, 1  // bytes 47..63
+    lea        edi, [edi + 64]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  //  HAS_MERGEUVROW_AVX2
+
+#ifdef HAS_COPYROW_SSE2
+// CopyRow copys 'count' bytes using a 16 byte load/store, 32 bytes at time.
+__declspec(naked)
+void CopyRow_SSE2(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_COPYROW_SSE2
+
+#ifdef HAS_COPYROW_AVX
+// CopyRow copys 'count' bytes using a 32 byte load/store, 64 bytes at time.
+__declspec(naked)
+void CopyRow_AVX(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax, [eax + 64]
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx, [edx + 64]
+    sub        ecx, 64
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_COPYROW_AVX
+
+// Multiple of 1.
+__declspec(naked)
+void CopyRow_ERMS(const uint8* src, uint8* dst, int count) {
+  __asm {
+    mov        eax, esi
+    mov        edx, edi
+    mov        esi, [esp + 4]   // src
+    mov        edi, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    rep movsb
+    mov        edi, edx
+    mov        esi, eax
+    ret
+  }
+}
+
+#ifdef HAS_ARGBCOPYALPHAROW_SSE2
+// width in pixels
+__declspec(naked)
+void ARGBCopyAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    pslld      xmm0, 24
+    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    psrld      xmm1, 8
+
+  convertloop:
+    movdqu     xmm2, [eax]
+    movdqu     xmm3, [eax + 16]
+    lea        eax, [eax + 32]
+    movdqu     xmm4, [edx]
+    movdqu     xmm5, [edx + 16]
+    pand       xmm2, xmm0
+    pand       xmm3, xmm0
+    pand       xmm4, xmm1
+    pand       xmm5, xmm1
+    por        xmm2, xmm4
+    por        xmm3, xmm5
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm3
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYALPHAROW_AVX2
+// width in pixels
+__declspec(naked)
+void ARGBCopyAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+
+  convertloop:
+    vmovdqu    ymm1, [eax]
+    vmovdqu    ymm2, [eax + 32]
+    lea        eax, [eax + 64]
+    vpblendvb  ymm1, ymm1, [edx], ymm0
+    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYALPHAROW_AVX2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2
+// width in pixels
+__declspec(naked)
+void ARGBCopyYToAlphaRow_SSE2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    pcmpeqb    xmm0, xmm0       // generate mask 0xff000000
+    pslld      xmm0, 24
+    pcmpeqb    xmm1, xmm1       // generate mask 0x00ffffff
+    psrld      xmm1, 8
+
+  convertloop:
+    movq       xmm2, qword ptr [eax]  // 8 Y's
+    lea        eax, [eax + 8]
+    punpcklbw  xmm2, xmm2
+    punpckhwd  xmm3, xmm2
+    punpcklwd  xmm2, xmm2
+    movdqu     xmm4, [edx]
+    movdqu     xmm5, [edx + 16]
+    pand       xmm2, xmm0
+    pand       xmm3, xmm0
+    pand       xmm4, xmm1
+    pand       xmm5, xmm1
+    por        xmm2, xmm4
+    por        xmm3, xmm5
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm3
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_SSE2
+
+#ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2
+// width in pixels
+__declspec(naked)
+void ARGBCopyYToAlphaRow_AVX2(const uint8* src, uint8* dst, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src
+    mov        edx, [esp + 8]   // dst
+    mov        ecx, [esp + 12]  // count
+    vpcmpeqb   ymm0, ymm0, ymm0
+    vpsrld     ymm0, ymm0, 8    // generate mask 0x00ffffff
+
+  convertloop:
+    vpmovzxbd  ymm1, qword ptr [eax]
+    vpmovzxbd  ymm2, qword ptr [eax + 8]
+    lea        eax, [eax + 16]
+    vpslld     ymm1, ymm1, 24
+    vpslld     ymm2, ymm2, 24
+    vpblendvb  ymm1, ymm1, [edx], ymm0
+    vpblendvb  ymm2, ymm2, [edx + 32], ymm0
+    vmovdqu    [edx], ymm1
+    vmovdqu    [edx + 32], ymm2
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBCOPYYTOALPHAROW_AVX2
+
+#ifdef HAS_SETROW_X86
+// Write 'count' bytes using an 8 bit value repeated.
+// Count should be multiple of 4.
+__declspec(naked)
+void SetRow_X86(uint8* dst, uint8 v8, int count) {
+  __asm {
+    movzx      eax, byte ptr [esp + 8]    // v8
+    mov        edx, 0x01010101  // Duplicate byte to all bytes.
+    mul        edx              // overwrites edx with upper part of result.
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        ecx, [esp + 12]  // count
+    shr        ecx, 2
+    rep stosd
+    mov        edi, edx
+    ret
+  }
+}
+
+// Write 'count' bytes using an 8 bit value repeated.
+__declspec(naked)
+void SetRow_ERMS(uint8* dst, uint8 v8, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        eax, [esp + 8]   // v8
+    mov        ecx, [esp + 12]  // count
+    rep stosb
+    mov        edi, edx
+    ret
+  }
+}
+
+// Write 'count' 32 bit values.
+__declspec(naked)
+void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count) {
+  __asm {
+    mov        edx, edi
+    mov        edi, [esp + 4]   // dst
+    mov        eax, [esp + 8]   // v32
+    mov        ecx, [esp + 12]  // count
+    rep stosd
+    mov        edi, edx
+    ret
+  }
+}
+#endif  // HAS_SETROW_X86
+
+#ifdef HAS_YUY2TOYROW_AVX2
+__declspec(naked)
+void YUY2ToYRow_AVX2(const uint8* src_yuy2,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // even bytes are Y
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked)
+void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked)
+void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // YUYV -> UVUV
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToYRow_AVX2(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpsrlw     ymm0, ymm0, 8      // odd bytes are Y
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    vpavgb     ymm0, ymm0, [eax + esi]
+    vpavgb     ymm1, ymm1, [eax + esi + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    vpcmpeqb   ymm5, ymm5, ymm5      // generate mask 0x00ff00ff
+    vpsrlw     ymm5, ymm5, 8
+    sub        edi, edx
+
+  convertloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax,  [eax + 64]
+    vpand      ymm0, ymm0, ymm5   // UYVY -> UVUV
+    vpand      ymm1, ymm1, ymm5
+    vpackuswb  ymm0, ymm0, ymm1   // mutates.
+    vpermq     ymm0, ymm0, 0xd8
+    vpand      ymm1, ymm0, ymm5  // U
+    vpsrlw     ymm0, ymm0, 8     // V
+    vpackuswb  ymm1, ymm1, ymm1  // mutates.
+    vpackuswb  ymm0, ymm0, ymm0  // mutates.
+    vpermq     ymm1, ymm1, 0xd8
+    vpermq     ymm0, ymm0, 0xd8
+    vextractf128 [edx], ymm1, 0  // U
+    vextractf128 [edx + edi], ymm0, 0 // V
+    lea        edx, [edx + 16]
+    sub        ecx, 32
+    jg         convertloop
+
+    pop        edi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_YUY2TOYROW_AVX2
+
+#ifdef HAS_YUY2TOYROW_SSE2
+__declspec(naked)
+void YUY2ToYRow_SSE2(const uint8* src_yuy2,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_yuy2
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+    pcmpeqb    xmm5, xmm5        // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // even bytes are Y
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8      // YUYV -> UVUV
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToYRow_SSE2(const uint8* src_uyvy,
+                     uint8* dst_y, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_uyvy
+    mov        edx, [esp + 8]    // dst_y
+    mov        ecx, [esp + 12]   // pix
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8    // odd bytes are Y
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
+                      uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_yuy2
+    mov        esi, [esp + 8 + 8]    // stride_yuy2
+    mov        edx, [esp + 8 + 12]   // dst_u
+    mov        edi, [esp + 8 + 16]   // dst_v
+    mov        ecx, [esp + 8 + 20]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
+                         uint8* dst_u, uint8* dst_v, int pix) {
+  __asm {
+    push       edi
+    mov        eax, [esp + 4 + 4]    // src_yuy2
+    mov        edx, [esp + 4 + 8]    // dst_u
+    mov        edi, [esp + 4 + 12]   // dst_v
+    mov        ecx, [esp + 4 + 16]   // pix
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+    sub        edi, edx
+
+  convertloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5   // UYVY -> UVUV
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    movdqa     xmm1, xmm0
+    pand       xmm0, xmm5  // U
+    packuswb   xmm0, xmm0
+    psrlw      xmm1, 8     // V
+    packuswb   xmm1, xmm1
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + edi], xmm1
+    lea        edx, [edx + 8]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    ret
+  }
+}
+#endif  // HAS_YUY2TOYROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSE2
+// Blend 8 pixels at a time.
+__declspec(naked)
+void ARGBBlendRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm7, xmm7       // generate constant 1
+    psrlw      xmm7, 15
+    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    psrlw      xmm6, 8
+    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    psllw      xmm5, 8
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+    sub        ecx, 4
+    jl         convertloop4b    // less than 4 pixels?
+
+    // 4 pixel loop.
+  convertloop4:
+    movdqu     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqu     xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqu     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jge        convertloop4
+
+  convertloop4b:
+    add        ecx, 4 - 1
+    jl         convertloop1b
+
+    // 1 pixel loop.
+  convertloop1:
+    movd       xmm3, [eax]      // src argb
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    psrlw      xmm3, 8          // alpha
+    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+    pshuflw    xmm3, xmm3, 0F5h
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jge        convertloop1
+
+  convertloop1b:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBBLENDROW_SSE2
+
+#ifdef HAS_ARGBBLENDROW_SSSE3
+// Shuffle table for isolating alpha.
+static const uvec8 kShuffleAlpha = {
+  3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80,
+  11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80
+};
+// Same as SSE2, but replaces:
+//    psrlw      xmm3, 8          // alpha
+//    pshufhw    xmm3, xmm3, 0F5h // 8 alpha words
+//    pshuflw    xmm3, xmm3, 0F5h
+// with..
+//    pshufb     xmm3, kShuffleAlpha // alpha
+// Blend 8 pixels at a time.
+
+__declspec(naked)
+void ARGBBlendRow_SSSE3(const uint8* src_argb0, const uint8* src_argb1,
+                        uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pcmpeqb    xmm7, xmm7       // generate constant 0x0001
+    psrlw      xmm7, 15
+    pcmpeqb    xmm6, xmm6       // generate mask 0x00ff00ff
+    psrlw      xmm6, 8
+    pcmpeqb    xmm5, xmm5       // generate mask 0xff00ff00
+    psllw      xmm5, 8
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+    sub        ecx, 4
+    jl         convertloop4b    // less than 4 pixels?
+
+    // 4 pixel loop.
+  convertloop4:
+    movdqu     xmm3, [eax]      // src argb
+    lea        eax, [eax + 16]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movdqu     xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movdqu     xmm1, [esi]      // _a_g
+    lea        esi, [esi + 16]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jge        convertloop4
+
+  convertloop4b:
+    add        ecx, 4 - 1
+    jl         convertloop1b
+
+    // 1 pixel loop.
+  convertloop1:
+    movd       xmm3, [eax]      // src argb
+    lea        eax, [eax + 4]
+    movdqa     xmm0, xmm3       // src argb
+    pxor       xmm3, xmm4       // ~alpha
+    movd       xmm2, [esi]      // _r_b
+    pshufb     xmm3, kShuffleAlpha // alpha
+    pand       xmm2, xmm6       // _r_b
+    paddw      xmm3, xmm7       // 256 - alpha
+    pmullw     xmm2, xmm3       // _r_b * alpha
+    movd       xmm1, [esi]      // _a_g
+    lea        esi, [esi + 4]
+    psrlw      xmm1, 8          // _a_g
+    por        xmm0, xmm4       // set alpha to 255
+    pmullw     xmm1, xmm3       // _a_g * alpha
+    psrlw      xmm2, 8          // _r_b convert to 8 bits again
+    paddusb    xmm0, xmm2       // + src argb
+    pand       xmm1, xmm5       // a_g_ convert to 8 bits again
+    paddusb    xmm0, xmm1       // + src argb
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jge        convertloop1
+
+  convertloop1b:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBBLENDROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_SSE2
+// Attenuate 4 pixels at a time.
+__declspec(naked)
+void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm4, xmm4       // generate mask 0xff000000
+    pslld      xmm4, 24
+    pcmpeqb    xmm5, xmm5       // generate mask 0x00ffffff
+    psrld      xmm5, 8
+
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    punpcklbw  xmm0, xmm0       // first 2
+    pshufhw    xmm2, xmm0, 0FFh // 8 alpha words
+    pshuflw    xmm2, xmm2, 0FFh
+    pmulhuw    xmm0, xmm2       // rgb * a
+    movdqu     xmm1, [eax]      // read 4 pixels
+    punpckhbw  xmm1, xmm1       // next 2 pixels
+    pshufhw    xmm2, xmm1, 0FFh // 8 alpha words
+    pshuflw    xmm2, xmm2, 0FFh
+    pmulhuw    xmm1, xmm2       // rgb * a
+    movdqu     xmm2, [eax]      // alphas
+    lea        eax, [eax + 16]
+    psrlw      xmm0, 8
+    pand       xmm2, xmm4
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    pand       xmm0, xmm5       // keep original alphas
+    por        xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBATTENUATEROW_SSSE3
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha0 = {
+  3u, 3u, 3u, 3u, 3u, 3u, 128u, 128u, 7u, 7u, 7u, 7u, 7u, 7u, 128u, 128u,
+};
+static const uvec8 kShuffleAlpha1 = {
+  11u, 11u, 11u, 11u, 11u, 11u, 128u, 128u,
+  15u, 15u, 15u, 15u, 15u, 15u, 128u, 128u,
+};
+__declspec(naked)
+void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    pcmpeqb    xmm3, xmm3       // generate mask 0xff000000
+    pslld      xmm3, 24
+    movdqa     xmm4, kShuffleAlpha0
+    movdqa     xmm5, kShuffleAlpha1
+
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    pshufb     xmm0, xmm4       // isolate first 2 alphas
+    movdqu     xmm1, [eax]      // read 4 pixels
+    punpcklbw  xmm1, xmm1       // first 2 pixel rgbs
+    pmulhuw    xmm0, xmm1       // rgb * a
+    movdqu     xmm1, [eax]      // read 4 pixels
+    pshufb     xmm1, xmm5       // isolate next 2 alphas
+    movdqu     xmm2, [eax]      // read 4 pixels
+    punpckhbw  xmm2, xmm2       // next 2 pixel rgbs
+    pmulhuw    xmm1, xmm2       // rgb * a
+    movdqu     xmm2, [eax]      // mask original alpha
+    lea        eax, [eax + 16]
+    pand       xmm2, xmm3
+    psrlw      xmm0, 8
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    por        xmm0, xmm2       // copy original alpha
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_SSSE3
+
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kShuffleAlpha_AVX2 = {
+  6u, 7u, 6u, 7u, 6u, 7u, 128u, 128u, 14u, 15u, 14u, 15u, 14u, 15u, 128u, 128u
+};
+__declspec(naked)
+void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vbroadcastf128 ymm4,kShuffleAlpha_AVX2
+    vpcmpeqb   ymm5, ymm5, ymm5 // generate mask 0xff000000
+    vpslld     ymm5, ymm5, 24
+
+ convertloop:
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpshufb    ymm2, ymm0, ymm4  // low 4 alphas
+    vpshufb    ymm3, ymm1, ymm4  // high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * a
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * a
+    vpand      ymm6, ymm6, ymm5  // isolate alpha
+    vpsrlw     ymm0, ymm0, 8
+    vpsrlw     ymm1, ymm1, 8
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    vpor       ymm0, ymm0, ymm6  // copy original alpha
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+// Unattenuate 4 pixels at a time.
+__declspec(naked)
+void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_argb0
+    mov        edx, [esp + 8 + 8]   // dst_argb
+    mov        ecx, [esp + 8 + 12]  // width
+
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    movzx      esi, byte ptr [eax + 3]  // first alpha
+    movzx      edi, byte ptr [eax + 7]  // second alpha
+    punpcklbw  xmm0, xmm0       // first 2
+    movd       xmm2, dword ptr fixed_invtbl8[esi * 4]
+    movd       xmm3, dword ptr fixed_invtbl8[edi * 4]
+    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words.  1, a, a, a
+    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    movlhps    xmm2, xmm3
+    pmulhuw    xmm0, xmm2       // rgb * a
+
+    movdqu     xmm1, [eax]      // read 4 pixels
+    movzx      esi, byte ptr [eax + 11]  // third alpha
+    movzx      edi, byte ptr [eax + 15]  // forth alpha
+    punpckhbw  xmm1, xmm1       // next 2
+    movd       xmm2, dword ptr fixed_invtbl8[esi * 4]
+    movd       xmm3, dword ptr fixed_invtbl8[edi * 4]
+    pshuflw    xmm2, xmm2, 040h // first 4 inv_alpha words
+    pshuflw    xmm3, xmm3, 040h // next 4 inv_alpha words
+    movlhps    xmm2, xmm3
+    pmulhuw    xmm1, xmm2       // rgb * a
+    lea        eax, [eax + 16]
+
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBUNATTENUATEROW_SSE2
+
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+// Shuffle table duplicating alpha.
+static const uvec8 kUnattenShuffleAlpha_AVX2 = {
+  0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u
+};
+// TODO(fbarchard): Enable USE_GATHER for future hardware if faster.
+// USE_GATHER is not on by default, due to being a slow instruction.
+#ifdef USE_GATHER
+__declspec(naked)
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vbroadcastf128 ymm4, kUnattenShuffleAlpha_AVX2
+
+ convertloop:
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpcmpeqb   ymm5, ymm5, ymm5  // generate mask 0xffffffff for gather.
+    vpsrld     ymm2, ymm6, 24    // alpha in low 8 bits.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpgatherdd ymm3, [ymm2 * 4 + fixed_invtbl8], ymm5  // ymm5 cleared.  1, a
+    vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
+    vpunpckhwd ymm3, ymm3, ymm3  // high 4 inverted alphas. mutated.
+    vpshufb    ymm2, ymm2, ymm4  // replicate low 4 alphas. 1, a, a, a
+    vpshufb    ymm3, ymm3, ymm4  // replicate high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    vzeroupper
+    ret
+  }
+}
+#else  // USE_GATHER
+__declspec(naked)
+void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                             int width) {
+  __asm {
+
+    mov        eax, [esp + 4]   // src_argb0
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    sub        edx, eax
+    vbroadcastf128 ymm5, kUnattenShuffleAlpha_AVX2
+
+    push       esi
+    push       edi
+
+ convertloop:
+    // replace VPGATHER
+    movzx      esi, byte ptr [eax + 3]                 // alpha0
+    movzx      edi, byte ptr [eax + 7]                 // alpha1
+    vmovd      xmm0, dword ptr fixed_invtbl8[esi * 4]  // [1,a0]
+    vmovd      xmm1, dword ptr fixed_invtbl8[edi * 4]  // [1,a1]
+    movzx      esi, byte ptr [eax + 11]                // alpha2
+    movzx      edi, byte ptr [eax + 15]                // alpha3
+    vpunpckldq xmm6, xmm0, xmm1                        // [1,a1,1,a0]
+    vmovd      xmm2, dword ptr fixed_invtbl8[esi * 4]  // [1,a2]
+    vmovd      xmm3, dword ptr fixed_invtbl8[edi * 4]  // [1,a3]
+    movzx      esi, byte ptr [eax + 19]                // alpha4
+    movzx      edi, byte ptr [eax + 23]                // alpha5
+    vpunpckldq xmm7, xmm2, xmm3                        // [1,a3,1,a2]
+    vmovd      xmm0, dword ptr fixed_invtbl8[esi * 4]  // [1,a4]
+    vmovd      xmm1, dword ptr fixed_invtbl8[edi * 4]  // [1,a5]
+    movzx      esi, byte ptr [eax + 27]                // alpha6
+    movzx      edi, byte ptr [eax + 31]                // alpha7
+    vpunpckldq xmm0, xmm0, xmm1                        // [1,a5,1,a4]
+    vmovd      xmm2, dword ptr fixed_invtbl8[esi * 4]  // [1,a6]
+    vmovd      xmm3, dword ptr fixed_invtbl8[edi * 4]  // [1,a7]
+    vpunpckldq xmm2, xmm2, xmm3                        // [1,a7,1,a6]
+    vpunpcklqdq xmm3, xmm6, xmm7                       // [1,a3,1,a2,1,a1,1,a0]
+    vpunpcklqdq xmm0, xmm0, xmm2                       // [1,a7,1,a6,1,a5,1,a4]
+    vinserti128 ymm3, ymm3, xmm0, 1 // [1,a7,1,a6,1,a5,1,a4,1,a3,1,a2,1,a1,1,a0]
+    // end of VPGATHER
+
+    vmovdqu    ymm6, [eax]       // read 8 pixels.
+    vpunpcklbw ymm0, ymm6, ymm6  // low 4 pixels. mutated.
+    vpunpckhbw ymm1, ymm6, ymm6  // high 4 pixels. mutated.
+    vpunpcklwd ymm2, ymm3, ymm3  // low 4 inverted alphas. mutated. 1, 1, a, a
+    vpunpckhwd ymm3, ymm3, ymm3  // high 4 inverted alphas. mutated.
+    vpshufb    ymm2, ymm2, ymm5  // replicate low 4 alphas. 1, a, a, a
+    vpshufb    ymm3, ymm3, ymm5  // replicate high 4 alphas
+    vpmulhuw   ymm0, ymm0, ymm2  // rgb * ia
+    vpmulhuw   ymm1, ymm1, ymm3  // rgb * ia
+    vpackuswb  ymm0, ymm0, ymm1  // unmutated.
+    vmovdqu    [eax + edx], ymm0
+    lea        eax, [eax + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // USE_GATHER
+#endif  // HAS_ARGBATTENUATEROW_AVX2
+
+#ifdef HAS_ARGBGRAYROW_SSSE3
+// Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels.
+__declspec(naked)
+void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]  /* width */
+    movdqa     xmm4, kARGBToYJ
+    movdqa     xmm5, kAddYJ64
+
+ convertloop:
+    movdqu     xmm0, [eax]  // G
+    movdqu     xmm1, [eax + 16]
+    pmaddubsw  xmm0, xmm4
+    pmaddubsw  xmm1, xmm4
+    phaddw     xmm0, xmm1
+    paddw      xmm0, xmm5  // Add .5 for rounding.
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0   // 8 G bytes
+    movdqu     xmm2, [eax]  // A
+    movdqu     xmm3, [eax + 16]
+    lea        eax, [eax + 32]
+    psrld      xmm2, 24
+    psrld      xmm3, 24
+    packuswb   xmm2, xmm3
+    packuswb   xmm2, xmm2   // 8 A bytes
+    movdqa     xmm3, xmm0   // Weave into GG, GA, then GGGA
+    punpcklbw  xmm0, xmm0   // 8 GG words
+    punpcklbw  xmm3, xmm2   // 8 GA words
+    movdqa     xmm1, xmm0
+    punpcklwd  xmm0, xmm3   // GGGA first 4
+    punpckhwd  xmm1, xmm3   // GGGA next 4
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBGRAYROW_SSSE3
+
+#ifdef HAS_ARGBSEPIAROW_SSSE3
+//    b = (r * 35 + g * 68 + b * 17) >> 7
+//    g = (r * 45 + g * 88 + b * 22) >> 7
+//    r = (r * 50 + g * 98 + b * 24) >> 7
+// Constant for ARGB color to sepia tone.
+static const vec8 kARGBToSepiaB = {
+  17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0, 17, 68, 35, 0
+};
+
+static const vec8 kARGBToSepiaG = {
+  22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0, 22, 88, 45, 0
+};
+
+static const vec8 kARGBToSepiaR = {
+  24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0, 24, 98, 50, 0
+};
+
+// Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels.
+__declspec(naked)
+void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* dst_argb */
+    mov        ecx, [esp + 8]   /* width */
+    movdqa     xmm2, kARGBToSepiaB
+    movdqa     xmm3, kARGBToSepiaG
+    movdqa     xmm4, kARGBToSepiaR
+
+ convertloop:
+    movdqu     xmm0, [eax]  // B
+    movdqu     xmm6, [eax + 16]
+    pmaddubsw  xmm0, xmm2
+    pmaddubsw  xmm6, xmm2
+    phaddw     xmm0, xmm6
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0   // 8 B values
+    movdqu     xmm5, [eax]  // G
+    movdqu     xmm1, [eax + 16]
+    pmaddubsw  xmm5, xmm3
+    pmaddubsw  xmm1, xmm3
+    phaddw     xmm5, xmm1
+    psrlw      xmm5, 7
+    packuswb   xmm5, xmm5   // 8 G values
+    punpcklbw  xmm0, xmm5   // 8 BG values
+    movdqu     xmm5, [eax]  // R
+    movdqu     xmm1, [eax + 16]
+    pmaddubsw  xmm5, xmm4
+    pmaddubsw  xmm1, xmm4
+    phaddw     xmm5, xmm1
+    psrlw      xmm5, 7
+    packuswb   xmm5, xmm5   // 8 R values
+    movdqu     xmm6, [eax]  // A
+    movdqu     xmm1, [eax + 16]
+    psrld      xmm6, 24
+    psrld      xmm1, 24
+    packuswb   xmm6, xmm1
+    packuswb   xmm6, xmm6   // 8 A values
+    punpcklbw  xmm5, xmm6   // 8 RA values
+    movdqa     xmm1, xmm0   // Weave BG, RA together
+    punpcklwd  xmm0, xmm5   // BGRA first 4
+    punpckhwd  xmm1, xmm5   // BGRA next 4
+    movdqu     [eax], xmm0
+    movdqu     [eax + 16], xmm1
+    lea        eax, [eax + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBSEPIAROW_SSSE3
+
+#ifdef HAS_ARGBCOLORMATRIXROW_SSSE3
+// Tranform 8 ARGB pixels (32 bytes) with color matrix.
+// Same as Sepia except matrix is provided.
+// TODO(fbarchard): packuswbs only use half of the reg. To make RGBA, combine R
+// and B into a high and low, then G/A, unpackl/hbw and then unpckl/hwd.
+__declspec(naked)
+void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                              const int8* matrix_argb, int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]  /* matrix_argb */
+    movdqu     xmm5, [ecx]
+    pshufd     xmm2, xmm5, 0x00
+    pshufd     xmm3, xmm5, 0x55
+    pshufd     xmm4, xmm5, 0xaa
+    pshufd     xmm5, xmm5, 0xff
+    mov        ecx, [esp + 16]  /* width */
+
+ convertloop:
+    movdqu     xmm0, [eax]  // B
+    movdqu     xmm7, [eax + 16]
+    pmaddubsw  xmm0, xmm2
+    pmaddubsw  xmm7, xmm2
+    movdqu     xmm6, [eax]  // G
+    movdqu     xmm1, [eax + 16]
+    pmaddubsw  xmm6, xmm3
+    pmaddubsw  xmm1, xmm3
+    phaddsw    xmm0, xmm7   // B
+    phaddsw    xmm6, xmm1   // G
+    psraw      xmm0, 6      // B
+    psraw      xmm6, 6      // G
+    packuswb   xmm0, xmm0   // 8 B values
+    packuswb   xmm6, xmm6   // 8 G values
+    punpcklbw  xmm0, xmm6   // 8 BG values
+    movdqu     xmm1, [eax]  // R
+    movdqu     xmm7, [eax + 16]
+    pmaddubsw  xmm1, xmm4
+    pmaddubsw  xmm7, xmm4
+    phaddsw    xmm1, xmm7   // R
+    movdqu     xmm6, [eax]  // A
+    movdqu     xmm7, [eax + 16]
+    pmaddubsw  xmm6, xmm5
+    pmaddubsw  xmm7, xmm5
+    phaddsw    xmm6, xmm7   // A
+    psraw      xmm1, 6      // R
+    psraw      xmm6, 6      // A
+    packuswb   xmm1, xmm1   // 8 R values
+    packuswb   xmm6, xmm6   // 8 A values
+    punpcklbw  xmm1, xmm6   // 8 RA values
+    movdqa     xmm6, xmm0   // Weave BG, RA together
+    punpcklwd  xmm0, xmm1   // BGRA first 4
+    punpckhwd  xmm6, xmm1   // BGRA next 4
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm6
+    lea        eax, [eax + 32]
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBCOLORMATRIXROW_SSSE3
+
+#ifdef HAS_ARGBQUANTIZEROW_SSE2
+// Quantize 4 ARGB pixels (16 bytes).
+__declspec(naked)
+void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
+                          int interval_offset, int width) {
+  __asm {
+    mov        eax, [esp + 4]    /* dst_argb */
+    movd       xmm2, [esp + 8]   /* scale */
+    movd       xmm3, [esp + 12]  /* interval_size */
+    movd       xmm4, [esp + 16]  /* interval_offset */
+    mov        ecx, [esp + 20]   /* width */
+    pshuflw    xmm2, xmm2, 040h
+    pshufd     xmm2, xmm2, 044h
+    pshuflw    xmm3, xmm3, 040h
+    pshufd     xmm3, xmm3, 044h
+    pshuflw    xmm4, xmm4, 040h
+    pshufd     xmm4, xmm4, 044h
+    pxor       xmm5, xmm5  // constant 0
+    pcmpeqb    xmm6, xmm6  // generate mask 0xff000000
+    pslld      xmm6, 24
+
+ convertloop:
+    movdqu     xmm0, [eax]  // read 4 pixels
+    punpcklbw  xmm0, xmm5   // first 2 pixels
+    pmulhuw    xmm0, xmm2   // pixel * scale >> 16
+    movdqu     xmm1, [eax]  // read 4 pixels
+    punpckhbw  xmm1, xmm5   // next 2 pixels
+    pmulhuw    xmm1, xmm2
+    pmullw     xmm0, xmm3   // * interval_size
+    movdqu     xmm7, [eax]  // read 4 pixels
+    pmullw     xmm1, xmm3
+    pand       xmm7, xmm6   // mask alpha
+    paddw      xmm0, xmm4   // + interval_size / 2
+    paddw      xmm1, xmm4
+    packuswb   xmm0, xmm1
+    por        xmm0, xmm7
+    movdqu     [eax], xmm0
+    lea        eax, [eax + 16]
+    sub        ecx, 4
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBQUANTIZEROW_SSE2
+
+#ifdef HAS_ARGBSHADEROW_SSE2
+// Shade 4 pixels at a time by specified value.
+__declspec(naked)
+void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
+                       uint32 value) {
+  __asm {
+    mov        eax, [esp + 4]   // src_argb
+    mov        edx, [esp + 8]   // dst_argb
+    mov        ecx, [esp + 12]  // width
+    movd       xmm2, [esp + 16]  // value
+    punpcklbw  xmm2, xmm2
+    punpcklqdq xmm2, xmm2
+
+ convertloop:
+    movdqu     xmm0, [eax]      // read 4 pixels
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm0       // first 2
+    punpckhbw  xmm1, xmm1       // next 2
+    pmulhuw    xmm0, xmm2       // argb * value
+    pmulhuw    xmm1, xmm2       // argb * value
+    psrlw      xmm0, 8
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    ret
+  }
+}
+#endif  // HAS_ARGBSHADEROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_SSE2
+// Multiply 2 rows of ARGB pixels together, 4 pixels at a time.
+__declspec(naked)
+void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    pxor       xmm5, xmm5  // constant 0
+
+ convertloop:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    movdqu     xmm2, [esi]        // read 4 pixels from src_argb1
+    movdqu     xmm1, xmm0
+    movdqu     xmm3, xmm2
+    punpcklbw  xmm0, xmm0         // first 2
+    punpckhbw  xmm1, xmm1         // next 2
+    punpcklbw  xmm2, xmm5         // first 2
+    punpckhbw  xmm3, xmm5         // next 2
+    pmulhuw    xmm0, xmm2         // src_argb0 * src_argb1 first 2
+    pmulhuw    xmm1, xmm3         // src_argb0 * src_argb1 next 2
+    lea        eax, [eax + 16]
+    lea        esi, [esi + 16]
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBMULTIPLYROW_SSE2
+
+#ifdef HAS_ARGBADDROW_SSE2
+// Add 2 rows of ARGB pixels together, 4 pixels at a time.
+// TODO(fbarchard): Port this to posix, neon and other math functions.
+__declspec(naked)
+void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+    sub        ecx, 4
+    jl         convertloop49
+
+ convertloop4:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    lea        eax, [eax + 16]
+    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    lea        esi, [esi + 16]
+    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jge        convertloop4
+
+ convertloop49:
+    add        ecx, 4 - 1
+    jl         convertloop19
+
+ convertloop1:
+    movd       xmm0, [eax]        // read 1 pixels from src_argb0
+    lea        eax, [eax + 4]
+    movd       xmm1, [esi]        // read 1 pixels from src_argb1
+    lea        esi, [esi + 4]
+    paddusb    xmm0, xmm1         // src_argb0 + src_argb1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jge        convertloop1
+
+ convertloop19:
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBADDROW_SSE2
+
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels together, 4 pixels at a time.
+__declspec(naked)
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+ convertloop:
+    movdqu     xmm0, [eax]        // read 4 pixels from src_argb0
+    lea        eax, [eax + 16]
+    movdqu     xmm1, [esi]        // read 4 pixels from src_argb1
+    lea        esi, [esi + 16]
+    psubusb    xmm0, xmm1         // src_argb0 - src_argb1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
+#ifdef HAS_ARGBMULTIPLYROW_AVX2
+// Multiply 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked)
+void ARGBMultiplyRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    vpxor      ymm5, ymm5, ymm5     // constant 0
+
+ convertloop:
+    vmovdqu    ymm1, [eax]        // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vmovdqu    ymm3, [esi]        // read 8 pixels from src_argb1
+    lea        esi, [esi + 32]
+    vpunpcklbw ymm0, ymm1, ymm1   // low 4
+    vpunpckhbw ymm1, ymm1, ymm1   // high 4
+    vpunpcklbw ymm2, ymm3, ymm5   // low 4
+    vpunpckhbw ymm3, ymm3, ymm5   // high 4
+    vpmulhuw   ymm0, ymm0, ymm2   // src_argb0 * src_argb1 low 4
+    vpmulhuw   ymm1, ymm1, ymm3   // src_argb0 * src_argb1 high 4
+    vpackuswb  ymm0, ymm0, ymm1
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBMULTIPLYROW_AVX2
+
+#ifdef HAS_ARGBADDROW_AVX2
+// Add 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked)
+void ARGBAddRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+ convertloop:
+    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vpaddusb   ymm0, ymm0, [esi]        // add 8 pixels from src_argb1
+    lea        esi, [esi + 32]
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBADDROW_AVX2
+
+#ifdef HAS_ARGBSUBTRACTROW_AVX2
+// Subtract 2 rows of ARGB pixels together, 8 pixels at a time.
+__declspec(naked)
+void ARGBSubtractRow_AVX2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+
+ convertloop:
+    vmovdqu    ymm0, [eax]              // read 8 pixels from src_argb0
+    lea        eax, [eax + 32]
+    vpsubusb   ymm0, ymm0, [esi]        // src_argb0 - src_argb1
+    lea        esi, [esi + 32]
+    vmovdqu    [edx], ymm0
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_AVX2
+
+#ifdef HAS_SOBELXROW_SSE2
+// SobelX as a matrix is
+// -1  0  1
+// -2  0  2
+// -1  0  1
+__declspec(naked)
+void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    const uint8* src_y2, uint8* dst_sobelx, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   // src_y0
+    mov        esi, [esp + 8 + 8]   // src_y1
+    mov        edi, [esp + 8 + 12]  // src_y2
+    mov        edx, [esp + 8 + 16]  // dst_sobelx
+    mov        ecx, [esp + 8 + 20]  // width
+    sub        esi, eax
+    sub        edi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    movq       xmm2, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + edi]      // read 8 pixels from src_y2[0]
+    movq       xmm3, qword ptr [eax + edi + 2]  // read 8 pixels from src_y2[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    psubw      xmm1, xmm0
+    pmaxsw     xmm0, xmm1
+    packuswb   xmm0, xmm0
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXROW_SSE2
+
+#ifdef HAS_SOBELYROW_SSE2
+// SobelY as a matrix is
+// -1 -2 -1
+//  0  0  0
+//  1  2  1
+__declspec(naked)
+void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
+                    uint8* dst_sobely, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_y0
+    mov        esi, [esp + 4 + 8]   // src_y1
+    mov        edx, [esp + 4 + 12]  // dst_sobely
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    sub        edx, eax
+    pxor       xmm5, xmm5  // constant 0
+
+ convertloop:
+    movq       xmm0, qword ptr [eax]            // read 8 pixels from src_y0[0]
+    movq       xmm1, qword ptr [eax + esi]      // read 8 pixels from src_y1[0]
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    psubw      xmm0, xmm1
+    movq       xmm1, qword ptr [eax + 1]        // read 8 pixels from src_y0[1]
+    movq       xmm2, qword ptr [eax + esi + 1]  // read 8 pixels from src_y1[1]
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm2, xmm5
+    psubw      xmm1, xmm2
+    movq       xmm2, qword ptr [eax + 2]        // read 8 pixels from src_y0[2]
+    movq       xmm3, qword ptr [eax + esi + 2]  // read 8 pixels from src_y1[2]
+    punpcklbw  xmm2, xmm5
+    punpcklbw  xmm3, xmm5
+    psubw      xmm2, xmm3
+    paddw      xmm0, xmm2
+    paddw      xmm0, xmm1
+    paddw      xmm0, xmm1
+    pxor       xmm1, xmm1   // abs = max(xmm0, -xmm0).  SSSE3 could use pabsw
+    psubw      xmm1, xmm0
+    pmaxsw     xmm0, xmm1
+    packuswb   xmm0, xmm0
+    movq       qword ptr [eax + edx], xmm0
+    lea        eax, [eax + 8]
+    sub        ecx, 8
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELYROW_SSE2
+
+#ifdef HAS_SOBELROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into ARGB.
+// A = 255
+// R = Sobel
+// G = Sobel
+// B = Sobel
+__declspec(naked)
+void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                   uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+    pslld      xmm5, 24             // 0xff000000
+
+ convertloop:
+    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm2, xmm0             // GG
+    punpcklbw  xmm2, xmm0             // First 8
+    punpckhbw  xmm0, xmm0             // Next 8
+    movdqa     xmm1, xmm2             // GGGG
+    punpcklwd  xmm1, xmm2             // First 4
+    punpckhwd  xmm2, xmm2             // Next 4
+    por        xmm1, xmm5             // GGGA
+    por        xmm2, xmm5
+    movdqa     xmm3, xmm0             // GGGG
+    punpcklwd  xmm3, xmm0             // Next 4
+    punpckhwd  xmm0, xmm0             // Last 4
+    por        xmm3, xmm5             // GGGA
+    por        xmm0, xmm5
+    movdqu     [edx], xmm1
+    movdqu     [edx + 16], xmm2
+    movdqu     [edx + 32], xmm3
+    movdqu     [edx + 48], xmm0
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELROW_SSE2
+
+#ifdef HAS_SOBELTOPLANEROW_SSE2
+// Adds Sobel X and Sobel Y and stores Sobel into a plane.
+__declspec(naked)
+void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                          uint8* dst_y, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+
+ convertloop:
+    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    paddusb    xmm0, xmm1             // sobel = sobelx + sobely
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELTOPLANEROW_SSE2
+
+#ifdef HAS_SOBELXYROW_SSE2
+// Mixes Sobel X, Sobel Y and Sobel into ARGB.
+// A = 255
+// R = Sobel X
+// G = Sobel
+// B = Sobel Y
+__declspec(naked)
+void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
+                     uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_sobelx
+    mov        esi, [esp + 4 + 8]   // src_sobely
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    pcmpeqb    xmm5, xmm5           // alpha 255
+
+ convertloop:
+    movdqu     xmm0, [eax]            // read 16 pixels src_sobelx
+    movdqu     xmm1, [eax + esi]      // read 16 pixels src_sobely
+    lea        eax, [eax + 16]
+    movdqa     xmm2, xmm0
+    paddusb    xmm2, xmm1             // sobel = sobelx + sobely
+    movdqa     xmm3, xmm0             // XA
+    punpcklbw  xmm3, xmm5
+    punpckhbw  xmm0, xmm5
+    movdqa     xmm4, xmm1             // YS
+    punpcklbw  xmm4, xmm2
+    punpckhbw  xmm1, xmm2
+    movdqa     xmm6, xmm4             // YSXA
+    punpcklwd  xmm6, xmm3             // First 4
+    punpckhwd  xmm4, xmm3             // Next 4
+    movdqa     xmm7, xmm1             // YSXA
+    punpcklwd  xmm7, xmm0             // Next 4
+    punpckhwd  xmm1, xmm0             // Last 4
+    movdqu     [edx], xmm6
+    movdqu     [edx + 16], xmm4
+    movdqu     [edx + 32], xmm7
+    movdqu     [edx + 48], xmm1
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_SOBELXYROW_SSE2
+
+#ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+// Consider float CumulativeSum.
+// Consider calling CumulativeSum one row at time as needed.
+// Consider circular CumulativeSum buffer of radius * 2 + 1 height.
+// Convert cumulative sum for an area to an average for 1 pixel.
+// topleft is pointer to top left of CumulativeSum buffer for area.
+// botleft is pointer to bottom left of CumulativeSum buffer.
+// width is offset from left to right of area in CumulativeSum buffer measured
+//   in number of ints.
+// area is the number of pixels in the area being averaged.
+// dst points to pixel to store result to.
+// count is number of averaged pixels to produce.
+// Does 4 pixels at a time.
+void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
+                                    int width, int area, uint8* dst,
+                                    int count) {
+  __asm {
+    mov        eax, topleft  // eax topleft
+    mov        esi, botleft  // esi botleft
+    mov        edx, width
+    movd       xmm5, area
+    mov        edi, dst
+    mov        ecx, count
+    cvtdq2ps   xmm5, xmm5
+    rcpss      xmm4, xmm5  // 1.0f / area
+    pshufd     xmm4, xmm4, 0
+    sub        ecx, 4
+    jl         l4b
+
+    cmp        area, 128  // 128 pixels will not overflow 15 bits.
+    ja         l4
+
+    pshufd     xmm5, xmm5, 0        // area
+    pcmpeqb    xmm6, xmm6           // constant of 65536.0 - 1 = 65535.0
+    psrld      xmm6, 16
+    cvtdq2ps   xmm6, xmm6
+    addps      xmm5, xmm6           // (65536.0 + area - 1)
+    mulps      xmm5, xmm4           // (65536.0 + area - 1) * 1 / area
+    cvtps2dq   xmm5, xmm5           // 0.16 fixed point
+    packssdw   xmm5, xmm5           // 16 bit shorts
+
+    // 4 pixel loop small blocks.
+  s4:
+    // top left
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+
+    // - top right
+    psubd      xmm0, [eax + edx * 4]
+    psubd      xmm1, [eax + edx * 4 + 16]
+    psubd      xmm2, [eax + edx * 4 + 32]
+    psubd      xmm3, [eax + edx * 4 + 48]
+    lea        eax, [eax + 64]
+
+    // - bottom left
+    psubd      xmm0, [esi]
+    psubd      xmm1, [esi + 16]
+    psubd      xmm2, [esi + 32]
+    psubd      xmm3, [esi + 48]
+
+    // + bottom right
+    paddd      xmm0, [esi + edx * 4]
+    paddd      xmm1, [esi + edx * 4 + 16]
+    paddd      xmm2, [esi + edx * 4 + 32]
+    paddd      xmm3, [esi + edx * 4 + 48]
+    lea        esi, [esi + 64]
+
+    packssdw   xmm0, xmm1  // pack 4 pixels into 2 registers
+    packssdw   xmm2, xmm3
+
+    pmulhuw    xmm0, xmm5
+    pmulhuw    xmm2, xmm5
+
+    packuswb   xmm0, xmm2
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jge        s4
+
+    jmp        l4b
+
+    // 4 pixel loop
+  l4:
+    // top left
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + 32]
+    movdqu     xmm3, [eax + 48]
+
+    // - top right
+    psubd      xmm0, [eax + edx * 4]
+    psubd      xmm1, [eax + edx * 4 + 16]
+    psubd      xmm2, [eax + edx * 4 + 32]
+    psubd      xmm3, [eax + edx * 4 + 48]
+    lea        eax, [eax + 64]
+
+    // - bottom left
+    psubd      xmm0, [esi]
+    psubd      xmm1, [esi + 16]
+    psubd      xmm2, [esi + 32]
+    psubd      xmm3, [esi + 48]
+
+    // + bottom right
+    paddd      xmm0, [esi + edx * 4]
+    paddd      xmm1, [esi + edx * 4 + 16]
+    paddd      xmm2, [esi + edx * 4 + 32]
+    paddd      xmm3, [esi + edx * 4 + 48]
+    lea        esi, [esi + 64]
+
+    cvtdq2ps   xmm0, xmm0   // Average = Sum * 1 / Area
+    cvtdq2ps   xmm1, xmm1
+    mulps      xmm0, xmm4
+    mulps      xmm1, xmm4
+    cvtdq2ps   xmm2, xmm2
+    cvtdq2ps   xmm3, xmm3
+    mulps      xmm2, xmm4
+    mulps      xmm3, xmm4
+    cvtps2dq   xmm0, xmm0
+    cvtps2dq   xmm1, xmm1
+    cvtps2dq   xmm2, xmm2
+    cvtps2dq   xmm3, xmm3
+    packssdw   xmm0, xmm1
+    packssdw   xmm2, xmm3
+    packuswb   xmm0, xmm2
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+  l1:
+    movdqu     xmm0, [eax]
+    psubd      xmm0, [eax + edx * 4]
+    lea        eax, [eax + 16]
+    psubd      xmm0, [esi]
+    paddd      xmm0, [esi + edx * 4]
+    lea        esi, [esi + 16]
+    cvtdq2ps   xmm0, xmm0
+    mulps      xmm0, xmm4
+    cvtps2dq   xmm0, xmm0
+    packssdw   xmm0, xmm0
+    packuswb   xmm0, xmm0
+    movd       dword ptr [edi], xmm0
+    lea        edi, [edi + 4]
+    sub        ecx, 1
+    jge        l1
+  l1b:
+  }
+}
+#endif  // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
+
+#ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
+// Creates a table of cumulative sums where each value is a sum of all values
+// above and to the left of the value.
+void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
+                                  const int32* previous_cumsum, int width) {
+  __asm {
+    mov        eax, row
+    mov        edx, cumsum
+    mov        esi, previous_cumsum
+    mov        ecx, width
+    pxor       xmm0, xmm0
+    pxor       xmm1, xmm1
+
+    sub        ecx, 4
+    jl         l4b
+    test       edx, 15
+    jne        l4b
+
+    // 4 pixel loop
+  l4:
+    movdqu     xmm2, [eax]  // 4 argb pixels 16 bytes.
+    lea        eax, [eax + 16]
+    movdqa     xmm4, xmm2
+
+    punpcklbw  xmm2, xmm1
+    movdqa     xmm3, xmm2
+    punpcklwd  xmm2, xmm1
+    punpckhwd  xmm3, xmm1
+
+    punpckhbw  xmm4, xmm1
+    movdqa     xmm5, xmm4
+    punpcklwd  xmm4, xmm1
+    punpckhwd  xmm5, xmm1
+
+    paddd      xmm0, xmm2
+    movdqu     xmm2, [esi]  // previous row above.
+    paddd      xmm2, xmm0
+
+    paddd      xmm0, xmm3
+    movdqu     xmm3, [esi + 16]
+    paddd      xmm3, xmm0
+
+    paddd      xmm0, xmm4
+    movdqu     xmm4, [esi + 32]
+    paddd      xmm4, xmm0
+
+    paddd      xmm0, xmm5
+    movdqu     xmm5, [esi + 48]
+    lea        esi, [esi + 64]
+    paddd      xmm5, xmm0
+
+    movdqu     [edx], xmm2
+    movdqu     [edx + 16], xmm3
+    movdqu     [edx + 32], xmm4
+    movdqu     [edx + 48], xmm5
+
+    lea        edx, [edx + 64]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+  l1:
+    movd       xmm2, dword ptr [eax]  // 1 argb pixel 4 bytes.
+    lea        eax, [eax + 4]
+    punpcklbw  xmm2, xmm1
+    punpcklwd  xmm2, xmm1
+    paddd      xmm0, xmm2
+    movdqu     xmm2, [esi]
+    lea        esi, [esi + 16]
+    paddd      xmm2, xmm0
+    movdqu     [edx], xmm2
+    lea        edx, [edx + 16]
+    sub        ecx, 1
+    jge        l1
+
+ l1b:
+  }
+}
+#endif  // HAS_COMPUTECUMULATIVESUMROW_SSE2
+
+#ifdef HAS_ARGBAFFINEROW_SSE2
+// Copy ARGB pixels from source image with slope to a row of destination.
+__declspec(naked)
+LIBYUV_API
+void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
+                        uint8* dst_argb, const float* uv_dudv, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 12]  // src_argb
+    mov        esi, [esp + 16]  // stride
+    mov        edx, [esp + 20]  // dst_argb
+    mov        ecx, [esp + 24]  // pointer to uv_dudv
+    movq       xmm2, qword ptr [ecx]  // uv
+    movq       xmm7, qword ptr [ecx + 8]  // dudv
+    mov        ecx, [esp + 28]  // width
+    shl        esi, 16          // 4, stride
+    add        esi, 4
+    movd       xmm5, esi
+    sub        ecx, 4
+    jl         l4b
+
+    // setup for 4 pixel loop
+    pshufd     xmm7, xmm7, 0x44  // dup dudv
+    pshufd     xmm5, xmm5, 0  // dup 4, stride
+    movdqa     xmm0, xmm2    // x0, y0, x1, y1
+    addps      xmm0, xmm7
+    movlhps    xmm2, xmm0
+    movdqa     xmm4, xmm7
+    addps      xmm4, xmm4    // dudv *= 2
+    movdqa     xmm3, xmm2    // x2, y2, x3, y3
+    addps      xmm3, xmm4
+    addps      xmm4, xmm4    // dudv *= 4
+
+    // 4 pixel loop
+  l4:
+    cvttps2dq  xmm0, xmm2    // x, y float to int first 2
+    cvttps2dq  xmm1, xmm3    // x, y float to int next 2
+    packssdw   xmm0, xmm1    // x, y as 8 shorts
+    pmaddwd    xmm0, xmm5    // offsets = x * 4 + y * stride.
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       edi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       xmm1, [eax + esi]  // read pixel 0
+    movd       xmm6, [eax + edi]  // read pixel 1
+    punpckldq  xmm1, xmm6     // combine pixel 0 and 1
+    addps      xmm2, xmm4    // x, y += dx, dy first 2
+    movq       qword ptr [edx], xmm1
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // shift right
+    movd       edi, xmm0
+    movd       xmm6, [eax + esi]  // read pixel 2
+    movd       xmm0, [eax + edi]  // read pixel 3
+    punpckldq  xmm6, xmm0     // combine pixel 2 and 3
+    addps      xmm3, xmm4    // x, y += dx, dy next 2
+    movq       qword ptr 8[edx], xmm6
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jge        l4
+
+  l4b:
+    add        ecx, 4 - 1
+    jl         l1b
+
+    // 1 pixel loop
+  l1:
+    cvttps2dq  xmm0, xmm2    // x, y float to int
+    packssdw   xmm0, xmm0    // x, y as shorts
+    pmaddwd    xmm0, xmm5    // offset = x * 4 + y * stride
+    addps      xmm2, xmm7    // x, y += dx, dy
+    movd       esi, xmm0
+    movd       xmm0, [eax + esi]  // copy a pixel
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jge        l1
+  l1b:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBAFFINEROW_SSE2
+
+#ifdef HAS_INTERPOLATEROW_AVX2
+// Bilinear filter 32x2 -> 32x1
+__declspec(naked)
+void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    sub        edi, esi
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    vmovd      xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    vmovd      xmm5, eax  // low fraction 128..1
+    vpunpcklbw xmm5, xmm5, xmm0
+    vpunpcklwd xmm5, xmm5, xmm5
+    vpxor      ymm0, ymm0, ymm0
+    vpermd     ymm5, ymm0, ymm5
+
+  xloop:
+    vmovdqu    ymm0, [esi]
+    vmovdqu    ymm2, [esi + edx]
+    vpunpckhbw ymm1, ymm0, ymm2  // mutates
+    vpunpcklbw ymm0, ymm0, ymm2  // mutates
+    vpmaddubsw ymm0, ymm0, ymm5
+    vpmaddubsw ymm1, ymm1, ymm5
+    vpsrlw     ymm0, ymm0, 7
+    vpsrlw     ymm1, ymm1, 7
+    vpackuswb  ymm0, ymm0, ymm1  // unmutates
+    vmovdqu    [esi + edi], ymm0
+    lea        esi, [esi + 32]
+    sub        ecx, 32
+    jg         xloop
+    jmp        xloop99
+
+   // Blend 25 / 75.
+ xloop25:
+   vmovdqu    ymm0, [esi]
+   vmovdqu    ymm1, [esi + edx]
+   vpavgb     ymm0, ymm0, ymm1
+   vpavgb     ymm0, ymm0, ymm1
+   vmovdqu    [esi + edi], ymm0
+   lea        esi, [esi + 32]
+   sub        ecx, 32
+   jg         xloop25
+   jmp        xloop99
+
+   // Blend 50 / 50.
+ xloop50:
+   vmovdqu    ymm0, [esi]
+   vpavgb     ymm0, ymm0, [esi + edx]
+   vmovdqu    [esi + edi], ymm0
+   lea        esi, [esi + 32]
+   sub        ecx, 32
+   jg         xloop50
+   jmp        xloop99
+
+   // Blend 75 / 25.
+ xloop75:
+   vmovdqu    ymm1, [esi]
+   vmovdqu    ymm0, [esi + edx]
+   vpavgb     ymm0, ymm0, ymm1
+   vpavgb     ymm0, ymm0, ymm1
+   vmovdqu    [esi + edi], ymm0
+   lea        esi, [esi + 32]
+   sub        ecx, 32
+   jg         xloop75
+   jmp        xloop99
+
+   // Blend 100 / 0 - Copy row unchanged.
+ xloop100:
+   rep movsb
+
+  xloop99:
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_AVX2
+
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked)
+void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                          ptrdiff_t src_stride, int dst_width,
+                          int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    shr        eax, 1
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 128.  Blend 100 / 0.
+    cmp        eax, 32
+    je         xloop75   // 32 / 128 is 0.25.  Blend 75 / 25.
+    cmp        eax, 64
+    je         xloop50   // 64 / 128 is 0.50.  Blend 50 / 50.
+    cmp        eax, 96
+    je         xloop25   // 96 / 128 is 0.75.  Blend 25 / 75.
+
+    movd       xmm0, eax  // high fraction 0..127
+    neg        eax
+    add        eax, 128
+    movd       xmm5, eax  // low fraction 128..1
+    punpcklbw  xmm5, xmm0
+    punpcklwd  xmm5, xmm5
+    pshufd     xmm5, xmm5, 0
+
+  xloop:
+    movdqu     xmm0, [esi]
+    movdqu     xmm2, [esi + edx]
+    movdqu     xmm1, xmm0
+    punpcklbw  xmm0, xmm2
+    punpckhbw  xmm1, xmm2
+    pmaddubsw  xmm0, xmm5
+    pmaddubsw  xmm1, xmm5
+    psrlw      xmm0, 7
+    psrlw      xmm1, 7
+    packuswb   xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+  xloop25:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+  xloop50:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+  xloop75:
+    movdqu     xmm1, [esi]
+    movdqu     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+  xloop100:
+    movdqu     xmm0, [esi]
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_INTERPOLATEROW_SSE2
+// Bilinear filter 16x2 -> 16x1
+__declspec(naked)
+void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                         ptrdiff_t src_stride, int dst_width,
+                         int source_y_fraction) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]   // dst_ptr
+    mov        esi, [esp + 8 + 8]   // src_ptr
+    mov        edx, [esp + 8 + 12]  // src_stride
+    mov        ecx, [esp + 8 + 16]  // dst_width
+    mov        eax, [esp + 8 + 20]  // source_y_fraction (0..255)
+    sub        edi, esi
+    // Dispatch to specialized filters if applicable.
+    cmp        eax, 0
+    je         xloop100  // 0 / 256.  Blend 100 / 0.
+    cmp        eax, 64
+    je         xloop75   // 64 / 256 is 0.25.  Blend 75 / 25.
+    cmp        eax, 128
+    je         xloop50   // 128 / 256 is 0.50.  Blend 50 / 50.
+    cmp        eax, 192
+    je         xloop25   // 192 / 256 is 0.75.  Blend 25 / 75.
+
+    movd       xmm5, eax            // xmm5 = y fraction
+    punpcklbw  xmm5, xmm5
+    psrlw      xmm5, 1
+    punpcklwd  xmm5, xmm5
+    punpckldq  xmm5, xmm5
+    punpcklqdq xmm5, xmm5
+    pxor       xmm4, xmm4
+
+  xloop:
+    movdqu     xmm0, [esi]  // row0
+    movdqu     xmm2, [esi + edx]  // row1
+    movdqu     xmm1, xmm0
+    movdqu     xmm3, xmm2
+    punpcklbw  xmm2, xmm4
+    punpckhbw  xmm3, xmm4
+    punpcklbw  xmm0, xmm4
+    punpckhbw  xmm1, xmm4
+    psubw      xmm2, xmm0  // row1 - row0
+    psubw      xmm3, xmm1
+    paddw      xmm2, xmm2  // 9 bits * 15 bits = 8.16
+    paddw      xmm3, xmm3
+    pmulhw     xmm2, xmm5  // scale diff
+    pmulhw     xmm3, xmm5
+    paddw      xmm0, xmm2  // sum rows
+    paddw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop
+    jmp        xloop99
+
+    // Blend 25 / 75.
+  xloop25:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop25
+    jmp        xloop99
+
+    // Blend 50 / 50.
+  xloop50:
+    movdqu     xmm0, [esi]
+    movdqu     xmm1, [esi + edx]
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop50
+    jmp        xloop99
+
+    // Blend 75 / 25.
+  xloop75:
+    movdqu     xmm1, [esi]
+    movdqu     xmm0, [esi + edx]
+    pavgb      xmm0, xmm1
+    pavgb      xmm0, xmm1
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop75
+    jmp        xloop99
+
+    // Blend 100 / 0 - Copy row unchanged.
+  xloop100:
+    movdqu     xmm0, [esi]
+    movdqu     [esi + edi], xmm0
+    lea        esi, [esi + 16]
+    sub        ecx, 16
+    jg         xloop100
+
+  xloop99:
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_INTERPOLATEROW_SSE2
+
+// For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA.
+__declspec(naked)
+void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                          const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]    // src_argb
+    mov        edx, [esp + 8]    // dst_argb
+    mov        ecx, [esp + 12]   // shuffler
+    movdqu     xmm5, [ecx]
+    mov        ecx, [esp + 16]   // pix
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm5
+    pshufb     xmm1, xmm5
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         wloop
+    ret
+  }
+}
+
+#ifdef HAS_ARGBSHUFFLEROW_AVX2
+__declspec(naked)
+void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  __asm {
+    mov        eax, [esp + 4]     // src_argb
+    mov        edx, [esp + 8]     // dst_argb
+    mov        ecx, [esp + 12]    // shuffler
+    vbroadcastf128 ymm5, [ecx]    // same shuffle in high as low.
+    mov        ecx, [esp + 16]    // pix
+
+  wloop:
+    vmovdqu    ymm0, [eax]
+    vmovdqu    ymm1, [eax + 32]
+    lea        eax, [eax + 64]
+    vpshufb    ymm0, ymm0, ymm5
+    vpshufb    ymm1, ymm1, ymm5
+    vmovdqu    [edx], ymm0
+    vmovdqu    [edx + 32], ymm1
+    lea        edx, [edx + 64]
+    sub        ecx, 16
+    jg         wloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBSHUFFLEROW_AVX2
+
+__declspec(naked)
+void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
+                         const uint8* shuffler, int pix) {
+  __asm {
+    push       ebx
+    push       esi
+    mov        eax, [esp + 8 + 4]    // src_argb
+    mov        edx, [esp + 8 + 8]    // dst_argb
+    mov        esi, [esp + 8 + 12]   // shuffler
+    mov        ecx, [esp + 8 + 16]   // pix
+    pxor       xmm5, xmm5
+
+    mov        ebx, [esi]   // shuffler
+    cmp        ebx, 0x03000102
+    je         shuf_3012
+    cmp        ebx, 0x00010203
+    je         shuf_0123
+    cmp        ebx, 0x00030201
+    je         shuf_0321
+    cmp        ebx, 0x02010003
+    je         shuf_2103
+
+  // TODO(fbarchard): Use one source pointer and 3 offsets.
+  shuf_any1:
+    movzx      ebx, byte ptr [esi]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx], bl
+    movzx      ebx, byte ptr [esi + 1]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 1], bl
+    movzx      ebx, byte ptr [esi + 2]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 2], bl
+    movzx      ebx, byte ptr [esi + 3]
+    movzx      ebx, byte ptr [eax + ebx]
+    mov        [edx + 3], bl
+    lea        eax, [eax + 4]
+    lea        edx, [edx + 4]
+    sub        ecx, 1
+    jg         shuf_any1
+    jmp        shuf99
+
+  shuf_0123:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 01Bh   // 1B = 00011011 = 0x0123 = BGRAToARGB
+    pshuflw    xmm0, xmm0, 01Bh
+    pshufhw    xmm1, xmm1, 01Bh
+    pshuflw    xmm1, xmm1, 01Bh
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         shuf_0123
+    jmp        shuf99
+
+  shuf_0321:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 039h   // 39 = 00111001 = 0x0321 = RGBAToARGB
+    pshuflw    xmm0, xmm0, 039h
+    pshufhw    xmm1, xmm1, 039h
+    pshuflw    xmm1, xmm1, 039h
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         shuf_0321
+    jmp        shuf99
+
+  shuf_2103:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 093h   // 93 = 10010011 = 0x2103 = ARGBToRGBA
+    pshuflw    xmm0, xmm0, 093h
+    pshufhw    xmm1, xmm1, 093h
+    pshuflw    xmm1, xmm1, 093h
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         shuf_2103
+    jmp        shuf99
+
+  shuf_3012:
+    movdqu     xmm0, [eax]
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm5
+    punpckhbw  xmm1, xmm5
+    pshufhw    xmm0, xmm0, 0C6h   // C6 = 11000110 = 0x3012 = ABGRToARGB
+    pshuflw    xmm0, xmm0, 0C6h
+    pshufhw    xmm1, xmm1, 0C6h
+    pshuflw    xmm1, xmm1, 0C6h
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         shuf_3012
+
+  shuf99:
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// YUY2 - Macro-pixel = 2 image pixels
+// Y0U0Y1V0....Y2U2Y3V2...Y4U4Y5V4....
+
+// UYVY - Macro-pixel = 2 image pixels
+// U0Y0V0Y1
+
+__declspec(naked)
+void I422ToYUY2Row_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_y
+    mov        esi, [esp + 8 + 8]    // src_u
+    mov        edx, [esp + 8 + 12]   // src_v
+    mov        edi, [esp + 8 + 16]   // dst_frame
+    mov        ecx, [esp + 8 + 20]   // width
+    sub        edx, esi
+
+  convertloop:
+    movq       xmm2, qword ptr [esi] // U
+    movq       xmm3, qword ptr [esi + edx] // V
+    lea        esi, [esi + 8]
+    punpcklbw  xmm2, xmm3 // UV
+    movdqu     xmm0, [eax] // Y
+    lea        eax, [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm2 // YUYV
+    punpckhbw  xmm1, xmm2
+    movdqu     [edi], xmm0
+    movdqu     [edi + 16], xmm1
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+__declspec(naked)
+void I422ToUYVYRow_SSE2(const uint8* src_y,
+                        const uint8* src_u,
+                        const uint8* src_v,
+                        uint8* dst_frame, int width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_y
+    mov        esi, [esp + 8 + 8]    // src_u
+    mov        edx, [esp + 8 + 12]   // src_v
+    mov        edi, [esp + 8 + 16]   // dst_frame
+    mov        ecx, [esp + 8 + 20]   // width
+    sub        edx, esi
+
+  convertloop:
+    movq       xmm2, qword ptr [esi] // U
+    movq       xmm3, qword ptr [esi + edx] // V
+    lea        esi, [esi + 8]
+    punpcklbw  xmm2, xmm3 // UV
+    movdqu     xmm0, [eax] // Y
+    movdqa     xmm1, xmm2
+    lea        eax, [eax + 16]
+    punpcklbw  xmm1, xmm0 // UYVY
+    punpckhbw  xmm2, xmm0
+    movdqu     [edi], xmm1
+    movdqu     [edi + 16], xmm2
+    lea        edi, [edi + 32]
+    sub        ecx, 16
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+__declspec(naked)
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* src_argb */
+    mov        edx, [esp + 4 + 8]   /* dst_argb */
+    mov        esi, [esp + 4 + 12]  /* poly */
+    mov        ecx, [esp + 4 + 16]  /* width */
+    pxor       xmm3, xmm3  // 0 constant for zero extending bytes to ints.
+
+    // 2 pixel loop.
+ convertloop:
+//    pmovzxbd  xmm0, dword ptr [eax]  // BGRA pixel
+//    pmovzxbd  xmm4, dword ptr [eax + 4]  // BGRA pixel
+    movq       xmm0, qword ptr [eax]  // BGRABGRA
+    lea        eax, [eax + 8]
+    punpcklbw  xmm0, xmm3
+    movdqa     xmm4, xmm0
+    punpcklwd  xmm0, xmm3  // pixel 0
+    punpckhwd  xmm4, xmm3  // pixel 1
+    cvtdq2ps   xmm0, xmm0  // 4 floats
+    cvtdq2ps   xmm4, xmm4
+    movdqa     xmm1, xmm0  // X
+    movdqa     xmm5, xmm4
+    mulps      xmm0, [esi + 16]  // C1 * X
+    mulps      xmm4, [esi + 16]
+    addps      xmm0, [esi]  // result = C0 + C1 * X
+    addps      xmm4, [esi]
+    movdqa     xmm2, xmm1
+    movdqa     xmm6, xmm5
+    mulps      xmm2, xmm1  // X * X
+    mulps      xmm6, xmm5
+    mulps      xmm1, xmm2  // X * X * X
+    mulps      xmm5, xmm6
+    mulps      xmm2, [esi + 32]  // C2 * X * X
+    mulps      xmm6, [esi + 32]
+    mulps      xmm1, [esi + 48]  // C3 * X * X * X
+    mulps      xmm5, [esi + 48]
+    addps      xmm0, xmm2  // result += C2 * X * X
+    addps      xmm4, xmm6
+    addps      xmm0, xmm1  // result += C3 * X * X * X
+    addps      xmm4, xmm5
+    cvttps2dq  xmm0, xmm0
+    cvttps2dq  xmm4, xmm4
+    packuswb   xmm0, xmm4
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    sub        ecx, 2
+    jg         convertloop
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+__declspec(naked)
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 12]   /* poly */
+    vbroadcastf128 ymm4, [ecx]       // C0
+    vbroadcastf128 ymm5, [ecx + 16]  // C1
+    vbroadcastf128 ymm6, [ecx + 32]  // C2
+    vbroadcastf128 ymm7, [ecx + 48]  // C3
+    mov        ecx, [esp + 16]  /* width */
+
+    // 2 pixel loop.
+ convertloop:
+    vpmovzxbd   ymm0, qword ptr [eax]  // 2 BGRA pixels
+    lea         eax, [eax + 8]
+    vcvtdq2ps   ymm0, ymm0        // X 8 floats
+    vmulps      ymm2, ymm0, ymm0  // X * X
+    vmulps      ymm3, ymm0, ymm7  // C3 * X
+    vfmadd132ps ymm0, ymm4, ymm5  // result = C0 + C1 * X
+    vfmadd231ps ymm0, ymm2, ymm6  // result += C2 * X * X
+    vfmadd231ps ymm0, ymm2, ymm3  // result += C3 * X * X * X
+    vcvttps2dq  ymm0, ymm0
+    vpackusdw   ymm0, ymm0, ymm0  // b0g0r0a0_00000000_b0g0r0a0_00000000
+    vpermq      ymm0, ymm0, 0xd8  // b0g0r0a0_b0g0r0a0_00000000_00000000
+    vpackuswb   xmm0, xmm0, xmm0  // bgrabgra_00000000_00000000_00000000
+    vmovq       qword ptr [edx], xmm0
+    lea         edx, [edx + 8]
+    sub         ecx, 2
+    jg          convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
+#ifdef HAS_ARGBCOLORTABLEROW_X86
+// Tranform ARGB pixels with color table.
+__declspec(naked)
+void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb,
+                           int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* dst_argb */
+    mov        esi, [esp + 4 + 8]   /* table_argb */
+    mov        ecx, [esp + 4 + 12]  /* width */
+
+    // 1 pixel loop.
+  convertloop:
+    movzx      edx, byte ptr [eax]
+    lea        eax, [eax + 4]
+    movzx      edx, byte ptr [esi + edx * 4]
+    mov        byte ptr [eax - 4], dl
+    movzx      edx, byte ptr [eax - 4 + 1]
+    movzx      edx, byte ptr [esi + edx * 4 + 1]
+    mov        byte ptr [eax - 4 + 1], dl
+    movzx      edx, byte ptr [eax - 4 + 2]
+    movzx      edx, byte ptr [esi + edx * 4 + 2]
+    mov        byte ptr [eax - 4 + 2], dl
+    movzx      edx, byte ptr [eax - 4 + 3]
+    movzx      edx, byte ptr [esi + edx * 4 + 3]
+    mov        byte ptr [eax - 4 + 3], dl
+    dec        ecx
+    jg         convertloop
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBCOLORTABLEROW_X86
+
+#ifdef HAS_RGBCOLORTABLEROW_X86
+// Tranform RGB pixels with color table.
+__declspec(naked)
+void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   /* dst_argb */
+    mov        esi, [esp + 4 + 8]   /* table_argb */
+    mov        ecx, [esp + 4 + 12]  /* width */
+
+    // 1 pixel loop.
+  convertloop:
+    movzx      edx, byte ptr [eax]
+    lea        eax, [eax + 4]
+    movzx      edx, byte ptr [esi + edx * 4]
+    mov        byte ptr [eax - 4], dl
+    movzx      edx, byte ptr [eax - 4 + 1]
+    movzx      edx, byte ptr [esi + edx * 4 + 1]
+    mov        byte ptr [eax - 4 + 1], dl
+    movzx      edx, byte ptr [eax - 4 + 2]
+    movzx      edx, byte ptr [esi + edx * 4 + 2]
+    mov        byte ptr [eax - 4 + 2], dl
+    dec        ecx
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_RGBCOLORTABLEROW_X86
+
+#ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3
+// Tranform RGB pixels with luma table.
+__declspec(naked)
+void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
+                                 int width,
+                                 const uint8* luma, uint32 lumacoeff) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]   /* src_argb */
+    mov        edi, [esp + 8 + 8]   /* dst_argb */
+    mov        ecx, [esp + 8 + 12]  /* width */
+    movd       xmm2, dword ptr [esp + 8 + 16]  // luma table
+    movd       xmm3, dword ptr [esp + 8 + 20]  // lumacoeff
+    pshufd     xmm2, xmm2, 0
+    pshufd     xmm3, xmm3, 0
+    pcmpeqb    xmm4, xmm4        // generate mask 0xff00ff00
+    psllw      xmm4, 8
+    pxor       xmm5, xmm5
+
+    // 4 pixel loop.
+  convertloop:
+    movdqu     xmm0, qword ptr [eax]      // generate luma ptr
+    pmaddubsw  xmm0, xmm3
+    phaddw     xmm0, xmm0
+    pand       xmm0, xmm4  // mask out low bits
+    punpcklwd  xmm0, xmm5
+    paddd      xmm0, xmm2  // add table base
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi], dl
+    movzx      edx, byte ptr [eax + 1]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 1], dl
+    movzx      edx, byte ptr [eax + 2]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 2], dl
+    movzx      edx, byte ptr [eax + 3]  // copy alpha.
+    mov        byte ptr [edi + 3], dl
+
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax + 4]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 4], dl
+    movzx      edx, byte ptr [eax + 5]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 5], dl
+    movzx      edx, byte ptr [eax + 6]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 6], dl
+    movzx      edx, byte ptr [eax + 7]  // copy alpha.
+    mov        byte ptr [edi + 7], dl
+
+    movd       esi, xmm0
+    pshufd     xmm0, xmm0, 0x39  // 00111001 to rotate right 32
+
+    movzx      edx, byte ptr [eax + 8]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 8], dl
+    movzx      edx, byte ptr [eax + 9]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 9], dl
+    movzx      edx, byte ptr [eax + 10]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 10], dl
+    movzx      edx, byte ptr [eax + 11]  // copy alpha.
+    mov        byte ptr [edi + 11], dl
+
+    movd       esi, xmm0
+
+    movzx      edx, byte ptr [eax + 12]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 12], dl
+    movzx      edx, byte ptr [eax + 13]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 13], dl
+    movzx      edx, byte ptr [eax + 14]
+    movzx      edx, byte ptr [esi + edx]
+    mov        byte ptr [edi + 14], dl
+    movzx      edx, byte ptr [eax + 15]  // copy alpha.
+    mov        byte ptr [edi + 15], dl
+
+    lea        eax, [eax + 16]
+    lea        edi, [edi + 16]
+    sub        ecx, 4
+    jg         convertloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBLUMACOLORTABLEROW_SSSE3
+
+#endif  // defined(_M_X64)
+#endif  // !defined(LIBYUV_DISABLE_X86) && (defined(_M_IX86) || defined(_M_X64))
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/row_x86.asm b/third_party/aom/third_party/libyuv/source/row_x86.asm
new file mode 100644
index 0000000000..0cb326f8e5
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/row_x86.asm
@@ -0,0 +1,146 @@
+;
+; Copyright 2012 The LibYuv Project Authors. All rights reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+%ifdef __YASM_VERSION_ID__
+%if __YASM_VERSION_ID__ < 01020000h
+%error AVX2 is supported only by yasm 1.2.0 or later.
+%endif
+%endif
+%include "x86inc.asm"
+
+SECTION .text
+
+; cglobal numeric constants are parameters, gpr regs, mm regs
+
+; void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix)
+
+%macro YUY2TOYROW 2-3
+cglobal %1ToYRow%3, 3, 3, 3, src_yuy2, dst_y, pix
+%ifidn %1,YUY2
+    pcmpeqb    m2, m2, m2        ; generate mask 0x00ff00ff
+    psrlw      m2, m2, 8
+%endif
+
+    ALIGN      4
+.convertloop:
+    mov%2      m0, [src_yuy2q]
+    mov%2      m1, [src_yuy2q + mmsize]
+    lea        src_yuy2q, [src_yuy2q + mmsize * 2]
+%ifidn %1,YUY2
+    pand       m0, m0, m2   ; YUY2 even bytes are Y
+    pand       m1, m1, m2
+%else
+    psrlw      m0, m0, 8    ; UYVY odd bytes are Y
+    psrlw      m1, m1, 8
+%endif
+    packuswb   m0, m0, m1
+%if cpuflag(AVX2)
+    vpermq     m0, m0, 0xd8
+%endif
+    sub        pixd, mmsize
+    mov%2      [dst_yq], m0
+    lea        dst_yq, [dst_yq + mmsize]
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+; TODO(fbarchard): Remove MMX.  Add SSSE3 pshufb version.
+INIT_MMX MMX
+YUY2TOYROW YUY2,a,
+YUY2TOYROW YUY2,u,_Unaligned
+YUY2TOYROW UYVY,a,
+YUY2TOYROW UYVY,u,_Unaligned
+INIT_XMM SSE2
+YUY2TOYROW YUY2,a,
+YUY2TOYROW YUY2,u,_Unaligned
+YUY2TOYROW UYVY,a,
+YUY2TOYROW UYVY,u,_Unaligned
+INIT_YMM AVX2
+YUY2TOYROW YUY2,a,
+YUY2TOYROW UYVY,a,
+
+; void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix)
+
+%macro SplitUVRow 1-2
+cglobal SplitUVRow%2, 4, 4, 5, src_uv, dst_u, dst_v, pix
+    pcmpeqb    m4, m4, m4        ; generate mask 0x00ff00ff
+    psrlw      m4, m4, 8
+    sub        dst_vq, dst_uq
+
+    ALIGN      4
+.convertloop:
+    mov%1      m0, [src_uvq]
+    mov%1      m1, [src_uvq + mmsize]
+    lea        src_uvq, [src_uvq + mmsize * 2]
+    psrlw      m2, m0, 8         ; odd bytes
+    psrlw      m3, m1, 8
+    pand       m0, m0, m4        ; even bytes
+    pand       m1, m1, m4
+    packuswb   m0, m0, m1
+    packuswb   m2, m2, m3
+%if cpuflag(AVX2)
+    vpermq     m0, m0, 0xd8
+    vpermq     m2, m2, 0xd8
+%endif
+    mov%1      [dst_uq], m0
+    mov%1      [dst_uq + dst_vq], m2
+    lea        dst_uq, [dst_uq + mmsize]
+    sub        pixd, mmsize
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+INIT_MMX MMX
+SplitUVRow a,
+SplitUVRow u,_Unaligned
+INIT_XMM SSE2
+SplitUVRow a,
+SplitUVRow u,_Unaligned
+INIT_YMM AVX2
+SplitUVRow a,
+
+; void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
+;                      int width);
+
+%macro MergeUVRow_ 1-2
+cglobal MergeUVRow_%2, 4, 4, 3, src_u, src_v, dst_uv, pix
+    sub        src_vq, src_uq
+
+    ALIGN      4
+.convertloop:
+    mov%1      m0, [src_uq]
+    mov%1      m1, [src_vq]
+    lea        src_uq, [src_uq + mmsize]
+    punpcklbw  m2, m0, m1       // first 8 UV pairs
+    punpckhbw  m0, m0, m1       // next 8 UV pairs
+%if cpuflag(AVX2)
+    vperm2i128 m1, m2, m0, 0x20  // low 128 of ymm2 and low 128 of ymm0
+    vperm2i128 m2, m2, m0, 0x31  // high 128 of ymm2 and high 128 of ymm0
+    mov%1      [dst_uvq], m1
+    mov%1      [dst_uvq + mmsize], m2
+%else
+    mov%1      [dst_uvq], m2
+    mov%1      [dst_uvq + mmsize], m0
+%endif
+    lea        dst_uvq, [dst_uvq + mmsize * 2]
+    sub        pixd, mmsize
+    jg         .convertloop
+    REP_RET
+%endmacro
+
+INIT_MMX MMX
+MergeUVRow_ a,
+MergeUVRow_ u,_Unaligned
+INIT_XMM SSE2
+MergeUVRow_ a,
+MergeUVRow_ u,_Unaligned
+INIT_YMM AVX2
+MergeUVRow_ a,
+
diff --git a/third_party/aom/third_party/libyuv/source/scale.cc b/third_party/aom/third_party/libyuv/source/scale.cc
new file mode 100644
index 0000000000..0a01304c41
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/scale.cc
@@ -0,0 +1,1689 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyPlane
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+#define SUBSAMPLE(v, a, s) (v < 0) ? (-((-v + a) >> s)) : ((v + a) >> s)
+
+// Scale plane, 1/2
+// This is an optimized version for scaling down a plane to 1/2 of
+// its original size.
+
+static void ScalePlaneDown2(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown2)(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) =
+      filtering == kFilterNone ? ScaleRowDown2_C :
+      (filtering == kFilterLinear ? ScaleRowDown2Linear_C : ScaleRowDown2Box_C);
+  int row_stride = src_stride << 1;
+  if (!filtering) {
+    src_ptr += src_stride;  // Point to odd rows.
+    src_stride = 0;
+  }
+
+#if defined(HAS_SCALEROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_NEON :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_NEON :
+        ScaleRowDown2Box_Any_NEON);
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_NEON :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_NEON :
+          ScaleRowDown2Box_NEON);
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_SSE2 :
+        ScaleRowDown2Box_Any_SSE2);
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_SSE2 :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_SSE2 :
+          ScaleRowDown2Box_SSE2);
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_Any_AVX2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_Any_AVX2 :
+        ScaleRowDown2Box_Any_AVX2);
+    if (IS_ALIGNED(dst_width, 32)) {
+      ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_AVX2 :
+          (filtering == kFilterLinear ? ScaleRowDown2Linear_AVX2 :
+          ScaleRowDown2Box_AVX2);
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
+      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown2 = filtering ?
+        ScaleRowDown2Box_MIPS_DSPR2 : ScaleRowDown2_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  // TODO(fbarchard): Loop through source height to allow odd height.
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void ScalePlaneDown2_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown2)(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst_ptr, int dst_width) =
+    filtering == kFilterNone ? ScaleRowDown2_16_C :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_16_C :
+        ScaleRowDown2Box_16_C);
+  int row_stride = src_stride << 1;
+  if (!filtering) {
+    src_ptr += src_stride;  // Point to odd rows.
+    src_stride = 0;
+  }
+
+#if defined(HAS_SCALEROWDOWN2_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering ? ScaleRowDown2Box_16_NEON :
+        ScaleRowDown2_16_NEON;
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 16)) {
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_16_SSE2 :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_16_SSE2 :
+        ScaleRowDown2Box_16_SSE2);
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN2_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_ptr, 4) &&
+      IS_ALIGNED(src_stride, 4) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown2 = filtering ?
+        ScaleRowDown2Box_16_MIPS_DSPR2 : ScaleRowDown2_16_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  // TODO(fbarchard): Loop through source height to allow odd height.
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown2(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+// Scale plane, 1/4
+// This is an optimized version for scaling down a plane to 1/4 of
+// its original size.
+
+static void ScalePlaneDown4(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown4)(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) =
+      filtering ? ScaleRowDown4Box_C : ScaleRowDown4_C;
+  int row_stride = src_stride << 2;
+  if (!filtering) {
+    src_ptr += src_stride * 2;  // Point to row 2.
+    src_stride = 0;
+  }
+#if defined(HAS_SCALEROWDOWN4_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_Any_NEON : ScaleRowDown4_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleRowDown4 = filtering ? ScaleRowDown4Box_NEON : ScaleRowDown4_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_Any_SSE2 : ScaleRowDown4_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleRowDown4 = filtering ? ScaleRowDown4Box_SSE2 : ScaleRowDown4_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_Any_AVX2 : ScaleRowDown4_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      ScaleRowDown4 = filtering ? ScaleRowDown4Box_AVX2 : ScaleRowDown4_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_MIPS_DSPR2 : ScaleRowDown4_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+static void ScalePlaneDown4_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown4)(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst_ptr, int dst_width) =
+      filtering ? ScaleRowDown4Box_16_C : ScaleRowDown4_16_C;
+  int row_stride = src_stride << 2;
+  if (!filtering) {
+    src_ptr += src_stride * 2;  // Point to row 2.
+    src_stride = 0;
+  }
+#if defined(HAS_SCALEROWDOWN4_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 8)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_NEON :
+        ScaleRowDown4_16_NEON;
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+    ScaleRowDown4 = filtering ? ScaleRowDown4Box_16_SSE2 :
+        ScaleRowDown4_16_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN4_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(row_stride, 4) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    ScaleRowDown4 = filtering ?
+        ScaleRowDown4Box_16_MIPS_DSPR2 : ScaleRowDown4_16_MIPS_DSPR2;
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (y = 0; y < dst_height; ++y) {
+    ScaleRowDown4(src_ptr, src_stride, dst_ptr, dst_width);
+    src_ptr += row_stride;
+    dst_ptr += dst_stride;
+  }
+}
+
+// Scale plane down, 3/4
+
+static void ScalePlaneDown34(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr,
+                             enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown34_0)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown34_0 = ScaleRowDown34_C;
+    ScaleRowDown34_1 = ScaleRowDown34_C;
+  } else {
+    ScaleRowDown34_0 = ScaleRowDown34_0_Box_C;
+    ScaleRowDown34_1 = ScaleRowDown34_1_Box_C;
+  }
+#if defined(HAS_SCALEROWDOWN34_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_Any_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_Any_NEON;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_NEON;
+    }
+    if (dst_width % 24 == 0) {
+      if (!filtering) {
+        ScaleRowDown34_0 = ScaleRowDown34_NEON;
+        ScaleRowDown34_1 = ScaleRowDown34_NEON;
+      } else {
+        ScaleRowDown34_0 = ScaleRowDown34_0_Box_NEON;
+        ScaleRowDown34_1 = ScaleRowDown34_1_Box_NEON;
+      }
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_Any_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_Any_SSSE3;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_Any_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_Any_SSSE3;
+    }
+    if (dst_width % 24 == 0) {
+      if (!filtering) {
+        ScaleRowDown34_0 = ScaleRowDown34_SSSE3;
+        ScaleRowDown34_1 = ScaleRowDown34_SSSE3;
+      } else {
+        ScaleRowDown34_0 = ScaleRowDown34_0_Box_SSSE3;
+        ScaleRowDown34_1 = ScaleRowDown34_1_Box_SSSE3;
+      }
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_MIPS_DSPR2;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
+                     dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static void ScalePlaneDown34_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr,
+                                enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown34_0)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  void (*ScaleRowDown34_1)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown34_0 = ScaleRowDown34_16_C;
+    ScaleRowDown34_1 = ScaleRowDown34_16_C;
+  } else {
+    ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_C;
+    ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_C;
+  }
+#if defined(HAS_SCALEROWDOWN34_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_16_NEON;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_NEON;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_16_SSSE3;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_SSSE3;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN34_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 24 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown34_0 = ScaleRowDown34_16_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_16_MIPS_DSPR2;
+    } else {
+      ScaleRowDown34_0 = ScaleRowDown34_0_Box_16_MIPS_DSPR2;
+      ScaleRowDown34_1 = ScaleRowDown34_1_Box_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_0(src_ptr + src_stride, -filter_stride,
+                     dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown34_0(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride;
+    dst_ptr += dst_stride;
+    ScaleRowDown34_1(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown34_0(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+
+// Scale plane, 3/8
+// This is an optimized version for scaling down a plane to 3/8
+// of its original size.
+//
+// Uses box filter arranges like this
+// aaabbbcc -> abc
+// aaabbbcc    def
+// aaabbbcc    ghi
+// dddeeeff
+// dddeeeff
+// dddeeeff
+// ggghhhii
+// ggghhhii
+// Boxes are 3x3, 2x3, 3x2 and 2x2
+
+static void ScalePlaneDown38(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr,
+                             enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown38_3)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown38_3 = ScaleRowDown38_C;
+    ScaleRowDown38_2 = ScaleRowDown38_C;
+  } else {
+    ScaleRowDown38_3 = ScaleRowDown38_3_Box_C;
+    ScaleRowDown38_2 = ScaleRowDown38_2_Box_C;
+  }
+
+#if defined(HAS_SCALEROWDOWN38_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_Any_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_Any_NEON;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_NEON;
+    }
+    if (dst_width % 12 == 0) {
+      if (!filtering) {
+        ScaleRowDown38_3 = ScaleRowDown38_NEON;
+        ScaleRowDown38_2 = ScaleRowDown38_NEON;
+      } else {
+        ScaleRowDown38_3 = ScaleRowDown38_3_Box_NEON;
+        ScaleRowDown38_2 = ScaleRowDown38_2_Box_NEON;
+      }
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN38_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_Any_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_Any_SSSE3;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_Any_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_Any_SSSE3;
+    }
+    if (dst_width % 12 == 0 && !filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_SSSE3;
+    }
+    if (dst_width % 6 == 0 && filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN38_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_MIPS_DSPR2;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+static void ScalePlaneDown38_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr,
+                                enum FilterMode filtering) {
+  int y;
+  void (*ScaleRowDown38_3)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  void (*ScaleRowDown38_2)(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst_ptr, int dst_width);
+  const int filter_stride = (filtering == kFilterLinear) ? 0 : src_stride;
+  assert(dst_width % 3 == 0);
+  if (!filtering) {
+    ScaleRowDown38_3 = ScaleRowDown38_16_C;
+    ScaleRowDown38_2 = ScaleRowDown38_16_C;
+  } else {
+    ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_C;
+    ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_C;
+  }
+#if defined(HAS_SCALEROWDOWN38_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && (dst_width % 12 == 0)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_16_NEON;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_NEON;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN38_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && (dst_width % 24 == 0)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_16_SSSE3;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_SSSE3;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_SCALEROWDOWN38_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && (dst_width % 12 == 0) &&
+      IS_ALIGNED(src_ptr, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_ptr, 4) && IS_ALIGNED(dst_stride, 4)) {
+    if (!filtering) {
+      ScaleRowDown38_3 = ScaleRowDown38_16_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_16_MIPS_DSPR2;
+    } else {
+      ScaleRowDown38_3 = ScaleRowDown38_3_Box_16_MIPS_DSPR2;
+      ScaleRowDown38_2 = ScaleRowDown38_2_Box_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  for (y = 0; y < dst_height - 2; y += 3) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_2(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 2;
+    dst_ptr += dst_stride;
+  }
+
+  // Remainder 1 or 2 rows with last row vertically unfiltered
+  if ((dst_height % 3) == 2) {
+    ScaleRowDown38_3(src_ptr, filter_stride, dst_ptr, dst_width);
+    src_ptr += src_stride * 3;
+    dst_ptr += dst_stride;
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  } else if ((dst_height % 3) == 1) {
+    ScaleRowDown38_3(src_ptr, 0, dst_ptr, dst_width);
+  }
+}
+
+#define MIN1(x) ((x) < 1 ? 1 : (x))
+
+static __inline uint32 SumPixels(int iboxwidth, const uint16* src_ptr) {
+  uint32 sum = 0u;
+  int x;
+  assert(iboxwidth > 0);
+  for (x = 0; x < iboxwidth; ++x) {
+    sum += src_ptr[x];
+  }
+  return sum;
+}
+
+static __inline uint32 SumPixels_16(int iboxwidth, const uint32* src_ptr) {
+  uint32 sum = 0u;
+  int x;
+  assert(iboxwidth > 0);
+  for (x = 0; x < iboxwidth; ++x) {
+    sum += src_ptr[x];
+  }
+  return sum;
+}
+
+static void ScaleAddCols2_C(int dst_width, int boxheight, int x, int dx,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int i;
+  int scaletbl[2];
+  int minboxwidth = dx >> 16;
+  int* scaleptr = scaletbl - minboxwidth;
+  int boxwidth;
+  scaletbl[0] = 65536 / (MIN1(minboxwidth) * boxheight);
+  scaletbl[1] = 65536 / (MIN1(minboxwidth + 1) * boxheight);
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = MIN1((x >> 16) - ix);
+    *dst_ptr++ = SumPixels(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16;
+  }
+}
+
+static void ScaleAddCols2_16_C(int dst_width, int boxheight, int x, int dx,
+                               const uint32* src_ptr, uint16* dst_ptr) {
+  int i;
+  int scaletbl[2];
+  int minboxwidth = dx >> 16;
+  int* scaleptr = scaletbl - minboxwidth;
+  int boxwidth;
+  scaletbl[0] = 65536 / (MIN1(minboxwidth) * boxheight);
+  scaletbl[1] = 65536 / (MIN1(minboxwidth + 1) * boxheight);
+  for (i = 0; i < dst_width; ++i) {
+    int ix = x >> 16;
+    x += dx;
+    boxwidth = MIN1((x >> 16) - ix);
+    *dst_ptr++ =
+        SumPixels_16(boxwidth, src_ptr + ix) * scaleptr[boxwidth] >> 16;
+  }
+}
+
+static void ScaleAddCols0_C(int dst_width, int boxheight, int x, int,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int scaleval = 65536 / boxheight;
+  int i;
+  src_ptr += (x >> 16);
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = src_ptr[i] * scaleval >> 16;
+  }
+}
+
+static void ScaleAddCols1_C(int dst_width, int boxheight, int x, int dx,
+                            const uint16* src_ptr, uint8* dst_ptr) {
+  int boxwidth = MIN1(dx >> 16);
+  int scaleval = 65536 / (boxwidth * boxheight);
+  int i;
+  x >>= 16;
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = SumPixels(boxwidth, src_ptr + x) * scaleval >> 16;
+    x += boxwidth;
+  }
+}
+
+static void ScaleAddCols1_16_C(int dst_width, int boxheight, int x, int dx,
+                               const uint32* src_ptr, uint16* dst_ptr) {
+  int boxwidth = MIN1(dx >> 16);
+  int scaleval = 65536 / (boxwidth * boxheight);
+  int i;
+  for (i = 0; i < dst_width; ++i) {
+    *dst_ptr++ = SumPixels_16(boxwidth, src_ptr + x) * scaleval >> 16;
+    x += boxwidth;
+  }
+}
+
+// Scale plane down to any dimensions, with interpolation.
+// (boxfilter).
+//
+// Same method as SimpleScale, which is fixed point, outputting
+// one pixel of destination using fixed point (16.16) to step
+// through source, sampling a box of pixel with simple
+// averaging.
+static void ScalePlaneBox(int src_width, int src_height,
+                          int dst_width, int dst_height,
+                          int src_stride, int dst_stride,
+                          const uint8* src_ptr, uint8* dst_ptr) {
+  int j, k;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height << 16);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  {
+    // Allocate a row buffer of uint16.
+    align_buffer_64(row16, src_width * 2);
+    void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
+        const uint16* src_ptr, uint8* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_C:
+        ((dx != 0x10000) ? ScaleAddCols1_C : ScaleAddCols0_C);
+    void (*ScaleAddRow)(const uint8* src_ptr, uint16* dst_ptr, int src_width) =
+        ScaleAddRow_C;
+#if defined(HAS_SCALEADDROW_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2)) {
+      ScaleAddRow = ScaleAddRow_Any_SSE2;
+      if (IS_ALIGNED(src_width, 16)) {
+        ScaleAddRow = ScaleAddRow_SSE2;
+      }
+    }
+#endif
+#if defined(HAS_SCALEADDROW_AVX2)
+    if (TestCpuFlag(kCpuHasAVX2)) {
+      ScaleAddRow = ScaleAddRow_Any_AVX2;
+      if (IS_ALIGNED(src_width, 32)) {
+        ScaleAddRow = ScaleAddRow_AVX2;
+      }
+    }
+#endif
+#if defined(HAS_SCALEADDROW_NEON)
+    if (TestCpuFlag(kCpuHasNEON)) {
+      ScaleAddRow = ScaleAddRow_Any_NEON;
+      if (IS_ALIGNED(src_width, 16)) {
+        ScaleAddRow = ScaleAddRow_NEON;
+      }
+    }
+#endif
+
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint8* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+      boxheight = MIN1((y >> 16) - iy);
+      memset(row16, 0, src_width * 2);
+      for (k = 0; k < boxheight; ++k) {
+        ScaleAddRow(src, (uint16 *)(row16), src_width);
+        src += src_stride;
+      }
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint16*)(row16), dst_ptr);
+      dst_ptr += dst_stride;
+    }
+    free_aligned_buffer_64(row16);
+  }
+}
+
+static void ScalePlaneBox_16(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint16* src_ptr, uint16* dst_ptr) {
+  int j, k;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height << 16);
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterBox,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  {
+    // Allocate a row buffer of uint32.
+    align_buffer_64(row32, src_width * 4);
+    void (*ScaleAddCols)(int dst_width, int boxheight, int x, int dx,
+        const uint32* src_ptr, uint16* dst_ptr) =
+        (dx & 0xffff) ? ScaleAddCols2_16_C: ScaleAddCols1_16_C;
+    void (*ScaleAddRow)(const uint16* src_ptr, uint32* dst_ptr, int src_width) =
+        ScaleAddRow_16_C;
+
+#if defined(HAS_SCALEADDROW_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(src_width, 16)) {
+      ScaleAddRow = ScaleAddRow_16_SSE2;
+    }
+#endif
+
+    for (j = 0; j < dst_height; ++j) {
+      int boxheight;
+      int iy = y >> 16;
+      const uint16* src = src_ptr + iy * src_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+      boxheight = MIN1((y >> 16) - iy);
+      memset(row32, 0, src_width * 4);
+      for (k = 0; k < boxheight; ++k) {
+        ScaleAddRow(src, (uint32 *)(row32), src_width);
+        src += src_stride;
+      }
+      ScaleAddCols(dst_width, boxheight, x, dx, (uint32*)(row32), dst_ptr);
+      dst_ptr += dst_stride;
+    }
+    free_aligned_buffer_64(row32);
+  }
+}
+
+// Scale plane down with bilinear interpolation.
+void ScalePlaneBilinearDown(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_ptr, uint8* dst_ptr,
+                            enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row buffer.
+  align_buffer_64(row, src_width);
+
+  const int max_y = (src_height - 1) << 16;
+  int j;
+  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleFilterCols64_C : ScaleFilterCols_C;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(src_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+
+#if defined(HAS_SCALEFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEFILTERCOLS_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleFilterCols = ScaleFilterCols_NEON;
+    }
+  }
+#endif
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    int yi = y >> 16;
+    const uint8* src = src_ptr + yi * src_stride;
+    if (filtering == kFilterLinear) {
+      ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow(row, src, src_stride, src_width, yf);
+      ScaleFilterCols(dst_ptr, row, dst_width, x, dx);
+    }
+    dst_ptr += dst_stride;
+    y += dy;
+    if (y > max_y) {
+      y = max_y;
+    }
+  }
+  free_aligned_buffer_64(row);
+}
+
+void ScalePlaneBilinearDown_16(int src_width, int src_height,
+                               int dst_width, int dst_height,
+                               int src_stride, int dst_stride,
+                               const uint16* src_ptr, uint16* dst_ptr,
+                               enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row buffer.
+  align_buffer_64(row, src_width * 2);
+
+  const int max_y = (src_height - 1) << 16;
+  int j;
+  void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleFilterCols64_16_C : ScaleFilterCols_16_C;
+  void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_16_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(src_width, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(src_width, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(src_width, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+
+#if defined(HAS_SCALEFILTERCOLS_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_16_SSSE3;
+  }
+#endif
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    int yi = y >> 16;
+    const uint16* src = src_ptr + yi * src_stride;
+    if (filtering == kFilterLinear) {
+      ScaleFilterCols(dst_ptr, src, dst_width, x, dx);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow((uint16*)row, src, src_stride, src_width, yf);
+      ScaleFilterCols(dst_ptr, (uint16*)row, dst_width, x, dx);
+    }
+    dst_ptr += dst_stride;
+    y += dy;
+    if (y > max_y) {
+      y = max_y;
+    }
+  }
+  free_aligned_buffer_64(row);
+}
+
+// Scale up down with bilinear interpolation.
+void ScalePlaneBilinearUp(int src_width, int src_height,
+                          int dst_width, int dst_height,
+                          int src_stride, int dst_stride,
+                          const uint8* src_ptr, uint8* dst_ptr,
+                          enum FilterMode filtering) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height - 1) << 16;
+  void (*InterpolateRow)(uint8* dst_ptr, const uint8* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleFilterCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleFilterCols_C : ScaleCols_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (filtering && src_width >= 32768) {
+    ScaleFilterCols = ScaleFilterCols64_C;
+  }
+#if defined(HAS_SCALEFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEFILTERCOLS_NEON)
+  if (filtering && TestCpuFlag(kCpuHasNEON) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleFilterCols = ScaleFilterCols_NEON;
+    }
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleFilterCols = ScaleColsUp2_C;
+#if defined(HAS_SCALECOLS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleFilterCols = ScaleColsUp2_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+  {
+    int yi = y >> 16;
+    const uint8* src = src_ptr + yi * src_stride;
+
+    // Allocate 2 row buffers.
+    const int kRowSize = (dst_width + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+
+    uint8* rowptr = row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_ptr + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf);
+      }
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+void ScalePlaneBilinearUp_16(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint16* src_ptr, uint16* dst_ptr,
+                             enum FilterMode filtering) {
+  int j;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  const int max_y = (src_height - 1) << 16;
+  void (*InterpolateRow)(uint16* dst_ptr, const uint16* src_ptr,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  void (*ScaleFilterCols)(uint16* dst_ptr, const uint16* src_ptr,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleFilterCols_16_C : ScaleCols_16_C;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_16_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(dst_width, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(dst_width, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2)) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+
+  if (filtering && src_width >= 32768) {
+    ScaleFilterCols = ScaleFilterCols64_16_C;
+  }
+#if defined(HAS_SCALEFILTERCOLS_16_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleFilterCols = ScaleFilterCols_16_SSSE3;
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleFilterCols = ScaleColsUp2_16_C;
+#if defined(HAS_SCALECOLS_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleFilterCols = ScaleColsUp2_16_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+  {
+    int yi = y >> 16;
+    const uint16* src = src_ptr + yi * src_stride;
+
+    // Allocate 2 row buffers.
+    const int kRowSize = (dst_width + 31) & ~31;
+    align_buffer_64(row, kRowSize * 4);
+
+    uint16* rowptr = (uint16*)row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_ptr + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_ptr, rowptr, 0, dst_width, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_ptr, rowptr, rowstride, dst_width, yf);
+      }
+      dst_ptr += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+// Scale Plane to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScalePlaneSimple(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             int src_stride, int dst_stride,
+                             const uint8* src_ptr, uint8* dst_ptr) {
+  int i;
+  void (*ScaleCols)(uint8* dst_ptr, const uint8* src_ptr,
+      int dst_width, int x, int dx) = ScaleCols_C;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleCols = ScaleColsUp2_C;
+#if defined(HAS_SCALECOLS_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleCols = ScaleColsUp2_SSE2;
+    }
+#endif
+  }
+
+  for (i = 0; i < dst_height; ++i) {
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride, dst_width, x, dx);
+    dst_ptr += dst_stride;
+    y += dy;
+  }
+}
+
+static void ScalePlaneSimple_16(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint16* src_ptr, uint16* dst_ptr) {
+  int i;
+  void (*ScaleCols)(uint16* dst_ptr, const uint16* src_ptr,
+      int dst_width, int x, int dx) = ScaleCols_16_C;
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  ScaleSlope(src_width, src_height, dst_width, dst_height, kFilterNone,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleCols = ScaleColsUp2_16_C;
+#if defined(HAS_SCALECOLS_16_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleCols = ScaleColsUp2_16_SSE2;
+    }
+#endif
+  }
+
+  for (i = 0; i < dst_height; ++i) {
+    ScaleCols(dst_ptr, src_ptr + (y >> 16) * src_stride,
+              dst_width, x, dx);
+    dst_ptr += dst_stride;
+    y += dy;
+  }
+}
+
+// Scale a plane.
+// This function dispatches to a specialized scaler based on scale factor.
+
+LIBYUV_API
+void ScalePlane(const uint8* src, int src_stride,
+                int src_width, int src_height,
+                uint8* dst, int dst_stride,
+                int dst_width, int dst_height,
+                enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height, filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  // Use specialized scales to improve performance for common resolutions.
+  // For example, all the 1/2 scalings will use ScalePlaneDown2()
+  if (dst_width == src_width && dst_height == src_height) {
+    // Straight copy.
+    CopyPlane(src, src_stride, dst, dst_stride, dst_width, dst_height);
+    return;
+  }
+  if (dst_width == src_width && filtering != kFilterBox) {
+    int dy = FixedDiv(src_height, dst_height);
+    // Arbitrary scale vertically, but unscaled horizontally.
+    ScalePlaneVertical(src_height,
+                       dst_width, dst_height,
+                       src_stride, dst_stride, src, dst,
+                       0, 0, dy, 1, filtering);
+    return;
+  }
+  if (dst_width <= Abs(src_width) && dst_height <= src_height) {
+    // Scale down.
+    if (4 * dst_width == 3 * src_width &&
+        4 * dst_height == 3 * src_height) {
+      // optimized, 3/4
+      ScalePlaneDown34(src_width, src_height, dst_width, dst_height,
+                       src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (2 * dst_width == src_width && 2 * dst_height == src_height) {
+      // optimized, 1/2
+      ScalePlaneDown2(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    // 3/8 rounded up for odd sized chroma height.
+    if (8 * dst_width == 3 * src_width &&
+        dst_height == ((src_height * 3 + 7) / 8)) {
+      // optimized, 3/8
+      ScalePlaneDown38(src_width, src_height, dst_width, dst_height,
+                       src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (4 * dst_width == src_width && 4 * dst_height == src_height &&
+        (filtering == kFilterBox || filtering == kFilterNone)) {
+      // optimized, 1/4
+      ScalePlaneDown4(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+  }
+  if (filtering == kFilterBox && dst_height * 2 < src_height) {
+    ScalePlaneBox(src_width, src_height, dst_width, dst_height,
+                  src_stride, dst_stride, src, dst);
+    return;
+  }
+  if (filtering && dst_height > src_height) {
+    ScalePlaneBilinearUp(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  if (filtering) {
+    ScalePlaneBilinearDown(src_width, src_height, dst_width, dst_height,
+                           src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  ScalePlaneSimple(src_width, src_height, dst_width, dst_height,
+                   src_stride, dst_stride, src, dst);
+}
+
+LIBYUV_API
+void ScalePlane_16(const uint16* src, int src_stride,
+                  int src_width, int src_height,
+                  uint16* dst, int dst_stride,
+                  int dst_width, int dst_height,
+                  enum FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height, filtering);
+
+  // Negative height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+
+  // Use specialized scales to improve performance for common resolutions.
+  // For example, all the 1/2 scalings will use ScalePlaneDown2()
+  if (dst_width == src_width && dst_height == src_height) {
+    // Straight copy.
+    CopyPlane_16(src, src_stride, dst, dst_stride, dst_width, dst_height);
+    return;
+  }
+  if (dst_width == src_width) {
+    int dy = FixedDiv(src_height, dst_height);
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical_16(src_height,
+                          dst_width, dst_height,
+                          src_stride, dst_stride, src, dst,
+                          0, 0, dy, 1, filtering);
+    return;
+  }
+  if (dst_width <= Abs(src_width) && dst_height <= src_height) {
+    // Scale down.
+    if (4 * dst_width == 3 * src_width &&
+        4 * dst_height == 3 * src_height) {
+      // optimized, 3/4
+      ScalePlaneDown34_16(src_width, src_height, dst_width, dst_height,
+                          src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (2 * dst_width == src_width && 2 * dst_height == src_height) {
+      // optimized, 1/2
+      ScalePlaneDown2_16(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    // 3/8 rounded up for odd sized chroma height.
+    if (8 * dst_width == 3 * src_width &&
+        dst_height == ((src_height * 3 + 7) / 8)) {
+      // optimized, 3/8
+      ScalePlaneDown38_16(src_width, src_height, dst_width, dst_height,
+                          src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+    if (4 * dst_width == src_width && 4 * dst_height == src_height &&
+               filtering != kFilterBilinear) {
+      // optimized, 1/4
+      ScalePlaneDown4_16(src_width, src_height, dst_width, dst_height,
+                         src_stride, dst_stride, src, dst, filtering);
+      return;
+    }
+  }
+  if (filtering == kFilterBox && dst_height * 2 < src_height) {
+    ScalePlaneBox_16(src_width, src_height, dst_width, dst_height,
+                     src_stride, dst_stride, src, dst);
+    return;
+  }
+  if (filtering && dst_height > src_height) {
+    ScalePlaneBilinearUp_16(src_width, src_height, dst_width, dst_height,
+                            src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  if (filtering) {
+    ScalePlaneBilinearDown_16(src_width, src_height, dst_width, dst_height,
+                              src_stride, dst_stride, src, dst, filtering);
+    return;
+  }
+  ScalePlaneSimple_16(src_width, src_height, dst_width, dst_height,
+                      src_stride, dst_stride, src, dst);
+}
+
+// Scale an I420 image.
+// This function in turn calls a scaling function for each plane.
+
+LIBYUV_API
+int I420Scale(const uint8* src_y, int src_stride_y,
+              const uint8* src_u, int src_stride_u,
+              const uint8* src_v, int src_stride_v,
+              int src_width, int src_height,
+              uint8* dst_y, int dst_stride_y,
+              uint8* dst_u, int dst_stride_u,
+              uint8* dst_v, int dst_stride_v,
+              int dst_width, int dst_height,
+              enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 ||
+      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane(src_y, src_stride_y, src_width, src_height,
+             dst_y, dst_stride_y, dst_width, dst_height,
+             filtering);
+  ScalePlane(src_u, src_stride_u, src_halfwidth, src_halfheight,
+             dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
+             filtering);
+  ScalePlane(src_v, src_stride_v, src_halfwidth, src_halfheight,
+             dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
+             filtering);
+  return 0;
+}
+
+LIBYUV_API
+int I420Scale_16(const uint16* src_y, int src_stride_y,
+                 const uint16* src_u, int src_stride_u,
+                 const uint16* src_v, int src_stride_v,
+                 int src_width, int src_height,
+                 uint16* dst_y, int dst_stride_y,
+                 uint16* dst_u, int dst_stride_u,
+                 uint16* dst_v, int dst_stride_v,
+                 int dst_width, int dst_height,
+                 enum FilterMode filtering) {
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  if (!src_y || !src_u || !src_v || src_width == 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 ||
+      !dst_y || !dst_u || !dst_v || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+
+  ScalePlane_16(src_y, src_stride_y, src_width, src_height,
+                dst_y, dst_stride_y, dst_width, dst_height,
+                filtering);
+  ScalePlane_16(src_u, src_stride_u, src_halfwidth, src_halfheight,
+                dst_u, dst_stride_u, dst_halfwidth, dst_halfheight,
+                filtering);
+  ScalePlane_16(src_v, src_stride_v, src_halfwidth, src_halfheight,
+                dst_v, dst_stride_v, dst_halfwidth, dst_halfheight,
+                filtering);
+  return 0;
+}
+
+// Deprecated api
+LIBYUV_API
+int Scale(const uint8* src_y, const uint8* src_u, const uint8* src_v,
+          int src_stride_y, int src_stride_u, int src_stride_v,
+          int src_width, int src_height,
+          uint8* dst_y, uint8* dst_u, uint8* dst_v,
+          int dst_stride_y, int dst_stride_u, int dst_stride_v,
+          int dst_width, int dst_height,
+          LIBYUV_BOOL interpolate) {
+  return I420Scale(src_y, src_stride_y,
+                   src_u, src_stride_u,
+                   src_v, src_stride_v,
+                   src_width, src_height,
+                   dst_y, dst_stride_y,
+                   dst_u, dst_stride_u,
+                   dst_v, dst_stride_v,
+                   dst_width, dst_height,
+                   interpolate ? kFilterBox : kFilterNone);
+}
+
+// Deprecated api
+LIBYUV_API
+int ScaleOffset(const uint8* src, int src_width, int src_height,
+                uint8* dst, int dst_width, int dst_height, int dst_yoffset,
+                LIBYUV_BOOL interpolate) {
+  // Chroma requires offset to multiple of 2.
+  int dst_yoffset_even = dst_yoffset & ~1;
+  int src_halfwidth = SUBSAMPLE(src_width, 1, 1);
+  int src_halfheight = SUBSAMPLE(src_height, 1, 1);
+  int dst_halfwidth = SUBSAMPLE(dst_width, 1, 1);
+  int dst_halfheight = SUBSAMPLE(dst_height, 1, 1);
+  int aheight = dst_height - dst_yoffset_even * 2;  // actual output height
+  const uint8* src_y = src;
+  const uint8* src_u = src + src_width * src_height;
+  const uint8* src_v = src + src_width * src_height +
+                             src_halfwidth * src_halfheight;
+  uint8* dst_y = dst + dst_yoffset_even * dst_width;
+  uint8* dst_u = dst + dst_width * dst_height +
+                 (dst_yoffset_even >> 1) * dst_halfwidth;
+  uint8* dst_v = dst + dst_width * dst_height + dst_halfwidth * dst_halfheight +
+                 (dst_yoffset_even >> 1) * dst_halfwidth;
+  if (!src || src_width <= 0 || src_height <= 0 ||
+      !dst || dst_width <= 0 || dst_height <= 0 || dst_yoffset_even < 0 ||
+      dst_yoffset_even >= dst_height) {
+    return -1;
+  }
+  return I420Scale(src_y, src_width,
+                   src_u, src_halfwidth,
+                   src_v, src_halfwidth,
+                   src_width, src_height,
+                   dst_y, dst_width,
+                   dst_u, dst_halfwidth,
+                   dst_v, dst_halfwidth,
+                   dst_width, aheight,
+                   interpolate ? kFilterBox : kFilterNone);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/scale_any.cc b/third_party/aom/third_party/libyuv/source/scale_any.cc
new file mode 100644
index 0000000000..2f6a2c8baf
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/scale_any.cc
@@ -0,0 +1,200 @@
+/*
+ *  Copyright 2015 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+#include "libyuv/scale_row.h"
+
+#include "libyuv/basic_types.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// Definition for ScaleFilterCols, ScaleARGBCols and ScaleARGBFilterCols
+#define CANY(NAMEANY, TERP_SIMD, TERP_C, BPP, MASK)                            \
+    void NAMEANY(uint8* dst_ptr, const uint8* src_ptr,                         \
+                 int dst_width, int x, int dx) {                               \
+      int n = dst_width & ~MASK;                                               \
+      if (n > 0) {                                                             \
+        TERP_SIMD(dst_ptr, src_ptr, n, x, dx);                                 \
+      }                                                                        \
+      TERP_C(dst_ptr + n * BPP, src_ptr,                                       \
+             dst_width & MASK, x + n * dx, dx);                                \
+    }
+
+#ifdef HAS_SCALEFILTERCOLS_NEON
+CANY(ScaleFilterCols_Any_NEON, ScaleFilterCols_NEON, ScaleFilterCols_C, 1, 7)
+#endif
+#ifdef HAS_SCALEARGBCOLS_NEON
+CANY(ScaleARGBCols_Any_NEON, ScaleARGBCols_NEON, ScaleARGBCols_C, 4, 7)
+#endif
+#ifdef HAS_SCALEARGBFILTERCOLS_NEON
+CANY(ScaleARGBFilterCols_Any_NEON, ScaleARGBFilterCols_NEON,
+     ScaleARGBFilterCols_C, 4, 3)
+#endif
+#undef CANY
+
+// Fixed scale down.
+#define SDANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK)   \
+    void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride,                   \
+                 uint8* dst_ptr, int dst_width) {                              \
+      int r = (int)((unsigned int)dst_width % (MASK + 1));                     \
+      int n = dst_width - r;                                                   \
+      if (n > 0) {                                                             \
+        SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                    \
+      }                                                                        \
+      SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                 \
+                     dst_ptr + n * BPP, r);                                    \
+    }
+
+#ifdef HAS_SCALEROWDOWN2_SSE2
+SDANY(ScaleRowDown2_Any_SSE2, ScaleRowDown2_SSE2, ScaleRowDown2_C, 2, 1, 15)
+SDANY(ScaleRowDown2Linear_Any_SSE2, ScaleRowDown2Linear_SSE2,
+      ScaleRowDown2Linear_C, 2, 1, 15)
+SDANY(ScaleRowDown2Box_Any_SSE2, ScaleRowDown2Box_SSE2, ScaleRowDown2Box_C,
+      2, 1, 15)
+#endif
+#ifdef HAS_SCALEROWDOWN2_AVX2
+SDANY(ScaleRowDown2_Any_AVX2, ScaleRowDown2_AVX2, ScaleRowDown2_C, 2, 1, 31)
+SDANY(ScaleRowDown2Linear_Any_AVX2, ScaleRowDown2Linear_AVX2,
+      ScaleRowDown2Linear_C, 2, 1, 31)
+SDANY(ScaleRowDown2Box_Any_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_C,
+      2, 1, 31)
+#endif
+#ifdef HAS_SCALEROWDOWN2_NEON
+SDANY(ScaleRowDown2_Any_NEON, ScaleRowDown2_NEON, ScaleRowDown2_C, 2, 1, 15)
+SDANY(ScaleRowDown2Linear_Any_NEON, ScaleRowDown2Linear_NEON,
+      ScaleRowDown2Linear_C, 2, 1, 15)
+SDANY(ScaleRowDown2Box_Any_NEON, ScaleRowDown2Box_NEON,
+      ScaleRowDown2Box_C, 2, 1, 15)
+#endif
+#ifdef HAS_SCALEROWDOWN4_SSE2
+SDANY(ScaleRowDown4_Any_SSE2, ScaleRowDown4_SSE2, ScaleRowDown4_C, 4, 1, 7)
+SDANY(ScaleRowDown4Box_Any_SSE2, ScaleRowDown4Box_SSE2, ScaleRowDown4Box_C,
+      4, 1, 7)
+#endif
+#ifdef HAS_SCALEROWDOWN4_AVX2
+SDANY(ScaleRowDown4_Any_AVX2, ScaleRowDown4_AVX2, ScaleRowDown4_C, 4, 1, 15)
+SDANY(ScaleRowDown4Box_Any_AVX2, ScaleRowDown4Box_AVX2, ScaleRowDown4Box_C,
+      4, 1, 15)
+#endif
+#ifdef HAS_SCALEROWDOWN4_NEON
+SDANY(ScaleRowDown4_Any_NEON, ScaleRowDown4_NEON, ScaleRowDown4_C, 4, 1, 7)
+SDANY(ScaleRowDown4Box_Any_NEON, ScaleRowDown4Box_NEON, ScaleRowDown4Box_C,
+      4, 1, 7)
+#endif
+#ifdef HAS_SCALEROWDOWN34_SSSE3
+SDANY(ScaleRowDown34_Any_SSSE3, ScaleRowDown34_SSSE3,
+      ScaleRowDown34_C, 4 / 3, 1, 23)
+SDANY(ScaleRowDown34_0_Box_Any_SSSE3, ScaleRowDown34_0_Box_SSSE3,
+      ScaleRowDown34_0_Box_C, 4 / 3, 1, 23)
+SDANY(ScaleRowDown34_1_Box_Any_SSSE3, ScaleRowDown34_1_Box_SSSE3,
+      ScaleRowDown34_1_Box_C, 4 / 3, 1, 23)
+#endif
+#ifdef HAS_SCALEROWDOWN34_NEON
+SDANY(ScaleRowDown34_Any_NEON, ScaleRowDown34_NEON,
+      ScaleRowDown34_C, 4 / 3, 1, 23)
+SDANY(ScaleRowDown34_0_Box_Any_NEON, ScaleRowDown34_0_Box_NEON,
+      ScaleRowDown34_0_Box_C, 4 / 3, 1, 23)
+SDANY(ScaleRowDown34_1_Box_Any_NEON, ScaleRowDown34_1_Box_NEON,
+      ScaleRowDown34_1_Box_C, 4 / 3, 1, 23)
+#endif
+#ifdef HAS_SCALEROWDOWN38_SSSE3
+SDANY(ScaleRowDown38_Any_SSSE3, ScaleRowDown38_SSSE3,
+      ScaleRowDown38_C, 8 / 3, 1, 11)
+SDANY(ScaleRowDown38_3_Box_Any_SSSE3, ScaleRowDown38_3_Box_SSSE3,
+      ScaleRowDown38_3_Box_C, 8 / 3, 1, 5)
+SDANY(ScaleRowDown38_2_Box_Any_SSSE3, ScaleRowDown38_2_Box_SSSE3,
+      ScaleRowDown38_2_Box_C, 8 / 3, 1, 5)
+#endif
+#ifdef HAS_SCALEROWDOWN38_NEON
+SDANY(ScaleRowDown38_Any_NEON, ScaleRowDown38_NEON,
+      ScaleRowDown38_C, 8 / 3, 1, 11)
+SDANY(ScaleRowDown38_3_Box_Any_NEON, ScaleRowDown38_3_Box_NEON,
+      ScaleRowDown38_3_Box_C, 8 / 3, 1, 11)
+SDANY(ScaleRowDown38_2_Box_Any_NEON, ScaleRowDown38_2_Box_NEON,
+      ScaleRowDown38_2_Box_C, 8 / 3, 1, 11)
+#endif
+
+#ifdef HAS_SCALEARGBROWDOWN2_SSE2
+SDANY(ScaleARGBRowDown2_Any_SSE2, ScaleARGBRowDown2_SSE2,
+      ScaleARGBRowDown2_C, 2, 4, 3)
+SDANY(ScaleARGBRowDown2Linear_Any_SSE2, ScaleARGBRowDown2Linear_SSE2,
+      ScaleARGBRowDown2Linear_C, 2, 4, 3)
+SDANY(ScaleARGBRowDown2Box_Any_SSE2, ScaleARGBRowDown2Box_SSE2,
+      ScaleARGBRowDown2Box_C, 2, 4, 3)
+#endif
+#ifdef HAS_SCALEARGBROWDOWN2_NEON
+SDANY(ScaleARGBRowDown2_Any_NEON, ScaleARGBRowDown2_NEON,
+      ScaleARGBRowDown2_C, 2, 4, 7)
+SDANY(ScaleARGBRowDown2Linear_Any_NEON, ScaleARGBRowDown2Linear_NEON,
+      ScaleARGBRowDown2Linear_C, 2, 4, 7)
+SDANY(ScaleARGBRowDown2Box_Any_NEON, ScaleARGBRowDown2Box_NEON,
+      ScaleARGBRowDown2Box_C, 2, 4, 7)
+#endif
+#undef SDANY
+
+// Scale down by even scale factor.
+#define SDAANY(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, BPP, MASK)          \
+    void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride, int src_stepx,    \
+                 uint8* dst_ptr, int dst_width) {                              \
+      int r = (int)((unsigned int)dst_width % (MASK + 1));                     \
+      int n = dst_width - r;                                                   \
+      if (n > 0) {                                                             \
+        SCALEROWDOWN_SIMD(src_ptr, src_stride, src_stepx, dst_ptr, n);         \
+      }                                                                        \
+      SCALEROWDOWN_C(src_ptr + (n * src_stepx) * BPP, src_stride,              \
+                     src_stepx, dst_ptr + n * BPP, r);                         \
+    }
+
+#ifdef HAS_SCALEARGBROWDOWNEVEN_SSE2
+SDAANY(ScaleARGBRowDownEven_Any_SSE2, ScaleARGBRowDownEven_SSE2,
+       ScaleARGBRowDownEven_C, 4, 3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_SSE2, ScaleARGBRowDownEvenBox_SSE2,
+       ScaleARGBRowDownEvenBox_C, 4, 3)
+#endif
+#ifdef HAS_SCALEARGBROWDOWNEVEN_NEON
+SDAANY(ScaleARGBRowDownEven_Any_NEON, ScaleARGBRowDownEven_NEON,
+       ScaleARGBRowDownEven_C, 4, 3)
+SDAANY(ScaleARGBRowDownEvenBox_Any_NEON, ScaleARGBRowDownEvenBox_NEON,
+       ScaleARGBRowDownEvenBox_C, 4, 3)
+#endif
+
+// Add rows box filter scale down.
+#define SAANY(NAMEANY, SCALEADDROW_SIMD, SCALEADDROW_C, MASK)                  \
+  void NAMEANY(const uint8* src_ptr, uint16* dst_ptr, int src_width) {         \
+      int n = src_width & ~MASK;                                               \
+      if (n > 0) {                                                             \
+        SCALEADDROW_SIMD(src_ptr, dst_ptr, n);                                 \
+      }                                                                        \
+      SCALEADDROW_C(src_ptr + n, dst_ptr + n, src_width & MASK);               \
+    }
+
+#ifdef HAS_SCALEADDROW_SSE2
+SAANY(ScaleAddRow_Any_SSE2, ScaleAddRow_SSE2, ScaleAddRow_C, 15)
+#endif
+#ifdef HAS_SCALEADDROW_AVX2
+SAANY(ScaleAddRow_Any_AVX2, ScaleAddRow_AVX2, ScaleAddRow_C, 31)
+#endif
+#ifdef HAS_SCALEADDROW_NEON
+SAANY(ScaleAddRow_Any_NEON, ScaleAddRow_NEON, ScaleAddRow_C, 15)
+#endif
+#undef SAANY
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
+
+
+
+
diff --git a/third_party/aom/third_party/libyuv/source/scale_argb.cc b/third_party/aom/third_party/libyuv/source/scale_argb.cc
new file mode 100644
index 0000000000..40a2d1ab20
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/scale_argb.cc
@@ -0,0 +1,853 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyARGB
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// ScaleARGB ARGB, 1/2
+// This is an optimized version for scaling down a ARGB to 1/2 of
+// its original size.
+static void ScaleARGBDown2(int src_width, int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint8* src_argb, uint8* dst_argb,
+                           int x, int dx, int y, int dy,
+                           enum FilterMode filtering) {
+  int j;
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) =
+    filtering == kFilterNone ? ScaleARGBRowDown2_C :
+        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_C :
+        ScaleARGBRowDown2Box_C);
+  assert(dx == 65536 * 2);  // Test scale factor of 2.
+  assert((dy & 0x1ffff) == 0);  // Test vertical scale is multiple of 2.
+  // Advance to odd row, even column.
+  if (filtering == kFilterBilinear) {
+    src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  } else {
+    src_argb += (y >> 16) * src_stride + ((x >> 16) - 1) * 4;
+  }
+
+#if defined(HAS_SCALEARGBROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_Any_SSE2 :
+        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_SSE2 :
+        ScaleARGBRowDown2Box_Any_SSE2);
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_SSE2 :
+          (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_SSE2 :
+          ScaleARGBRowDown2Box_SSE2);
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_Any_NEON :
+        (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_Any_NEON :
+        ScaleARGBRowDown2Box_Any_NEON);
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBRowDown2 = filtering == kFilterNone ? ScaleARGBRowDown2_NEON :
+          (filtering == kFilterLinear ? ScaleARGBRowDown2Linear_NEON :
+          ScaleARGBRowDown2Box_NEON);
+    }
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDown2(src_argb, src_stride, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+}
+
+// ScaleARGB ARGB, 1/4
+// This is an optimized version for scaling down a ARGB to 1/4 of
+// its original size.
+static void ScaleARGBDown4Box(int src_width, int src_height,
+                              int dst_width, int dst_height,
+                              int src_stride, int dst_stride,
+                              const uint8* src_argb, uint8* dst_argb,
+                              int x, int dx, int y, int dy) {
+  int j;
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (dst_width * 2 * 4 + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2);
+  int row_stride = src_stride * (dy >> 16);
+  void (*ScaleARGBRowDown2)(const uint8* src_argb, ptrdiff_t src_stride,
+    uint8* dst_argb, int dst_width) = ScaleARGBRowDown2Box_C;
+  // Advance to odd row, even column.
+  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+  assert(dx == 65536 * 4);  // Test scale factor of 4.
+  assert((dy & 0x3ffff) == 0);  // Test vertical scale is multiple of 4.
+#if defined(HAS_SCALEARGBROWDOWN2_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleARGBRowDown2 = ScaleARGBRowDown2Box_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDown2 = ScaleARGBRowDown2Box_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWN2_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBRowDown2 = ScaleARGBRowDown2Box_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBRowDown2 = ScaleARGBRowDown2Box_NEON;
+    }
+  }
+#endif
+
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDown2(src_argb, src_stride, row, dst_width * 2);
+    ScaleARGBRowDown2(src_argb + src_stride * 2, src_stride,
+                      row + kRowSize, dst_width * 2);
+    ScaleARGBRowDown2(row, kRowSize, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+  free_aligned_buffer_64(row);
+}
+
+// ScaleARGB ARGB Even
+// This is an optimized version for scaling down a ARGB to even
+// multiple of its original size.
+static void ScaleARGBDownEven(int src_width, int src_height,
+                              int dst_width, int dst_height,
+                              int src_stride, int dst_stride,
+                              const uint8* src_argb, uint8* dst_argb,
+                              int x, int dx, int y, int dy,
+                              enum FilterMode filtering) {
+  int j;
+  int col_step = dx >> 16;
+  int row_stride = (dy >> 16) * src_stride;
+  void (*ScaleARGBRowDownEven)(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_step, uint8* dst_argb, int dst_width) =
+      filtering ? ScaleARGBRowDownEvenBox_C : ScaleARGBRowDownEven_C;
+  assert(IS_ALIGNED(src_width, 2));
+  assert(IS_ALIGNED(src_height, 2));
+  src_argb += (y >> 16) * src_stride + (x >> 16) * 4;
+#if defined(HAS_SCALEARGBROWDOWNEVEN_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_SSE2 :
+        ScaleARGBRowDownEven_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_SSE2 :
+          ScaleARGBRowDownEven_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBROWDOWNEVEN_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_Any_NEON :
+        ScaleARGBRowDownEven_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBRowDownEven = filtering ? ScaleARGBRowDownEvenBox_NEON :
+          ScaleARGBRowDownEven_NEON;
+    }
+  }
+#endif
+
+  if (filtering == kFilterLinear) {
+    src_stride = 0;
+  }
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBRowDownEven(src_argb, src_stride, col_step, dst_argb, dst_width);
+    src_argb += row_stride;
+    dst_argb += dst_stride;
+  }
+}
+
+// Scale ARGB down with bilinear interpolation.
+static void ScaleARGBBilinearDown(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  int src_stride, int dst_stride,
+                                  const uint8* src_argb, uint8* dst_argb,
+                                  int x, int dx, int y, int dy,
+                                  enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleARGBFilterCols64_C : ScaleARGBFilterCols_C;
+  int64 xlast = x + (int64)(dst_width - 1) * dx;
+  int64 xl = (dx >= 0) ? x : xlast;
+  int64 xr = (dx >= 0) ? xlast : x;
+  int clip_src_width;
+  xl = (xl >> 16) & ~3;  // Left edge aligned.
+  xr = (xr >> 16) + 1;  // Right most pixel used.  Bilinear uses 2 pixels.
+  xr = (xr + 1 + 3) & ~3;  // 1 beyond 4 pixel aligned right most pixel.
+  if (xr > src_width) {
+    xr = src_width;
+  }
+  clip_src_width = (int)(xr - xl) * 4;  // Width aligned to 4.
+  src_argb += xl * 4;
+  x -= (int)(xl << 16);
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(clip_src_width, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(clip_src_width, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(clip_src_width, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_NEON;
+    }
+  }
+#endif
+  // TODO(fbarchard): Consider not allocating row buffer for kFilterLinear.
+  // Allocate a row of ARGB.
+  {
+    align_buffer_64(row, clip_src_width * 4);
+
+    const int max_y = (src_height - 1) << 16;
+    if (y > max_y) {
+      y = max_y;
+    }
+    for (j = 0; j < dst_height; ++j) {
+      int yi = y >> 16;
+      const uint8* src = src_argb + yi * src_stride;
+      if (filtering == kFilterLinear) {
+        ScaleARGBFilterCols(dst_argb, src, dst_width, x, dx);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(row, src, src_stride, clip_src_width, yf);
+        ScaleARGBFilterCols(dst_argb, row, dst_width, x, dx);
+      }
+      dst_argb += dst_stride;
+      y += dy;
+      if (y > max_y) {
+        y = max_y;
+      }
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+// Scale ARGB up with bilinear interpolation.
+static void ScaleARGBBilinearUp(int src_width, int src_height,
+                                int dst_width, int dst_height,
+                                int src_stride, int dst_stride,
+                                const uint8* src_argb, uint8* dst_argb,
+                                int x, int dx, int y, int dy,
+                                enum FilterMode filtering) {
+  int j;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
+  const int max_y = (src_height - 1) << 16;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+  if (src_width >= 32768) {
+    ScaleARGBFilterCols = filtering ?
+        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
+  }
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_NEON)
+  if (filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBCols_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_NEON)
+  if (!filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBCols_NEON;
+    }
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBFilterCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  if (y > max_y) {
+    y = max_y;
+  }
+
+  {
+    int yi = y >> 16;
+    const uint8* src = src_argb + yi * src_stride;
+
+    // Allocate 2 rows of ARGB.
+    const int kRowSize = (dst_width * 4 + 31) & ~31;
+    align_buffer_64(row, kRowSize * 2);
+
+    uint8* rowptr = row;
+    int rowstride = kRowSize;
+    int lasty = yi;
+
+    ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
+    if (src_height > 1) {
+      src += src_stride;
+    }
+    ScaleARGBFilterCols(rowptr + rowstride, src, dst_width, x, dx);
+    src += src_stride;
+
+    for (j = 0; j < dst_height; ++j) {
+      yi = y >> 16;
+      if (yi != lasty) {
+        if (y > max_y) {
+          y = max_y;
+          yi = y >> 16;
+          src = src_argb + yi * src_stride;
+        }
+        if (yi != lasty) {
+          ScaleARGBFilterCols(rowptr, src, dst_width, x, dx);
+          rowptr += rowstride;
+          rowstride = -rowstride;
+          lasty = yi;
+          src += src_stride;
+        }
+      }
+      if (filtering == kFilterLinear) {
+        InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
+      } else {
+        int yf = (y >> 8) & 255;
+        InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
+      }
+      dst_argb += dst_stride;
+      y += dy;
+    }
+    free_aligned_buffer_64(row);
+  }
+}
+
+#ifdef YUVSCALEUP
+// Scale YUV to ARGB up with bilinear interpolation.
+static void ScaleYUVToARGBBilinearUp(int src_width, int src_height,
+                                     int dst_width, int dst_height,
+                                     int src_stride_y,
+                                     int src_stride_u,
+                                     int src_stride_v,
+                                     int dst_stride_argb,
+                                     const uint8* src_y,
+                                     const uint8* src_u,
+                                     const uint8* src_v,
+                                     uint8* dst_argb,
+                                     int x, int dx, int y, int dy,
+                                     enum FilterMode filtering) {
+  int j;
+  void (*I422ToARGBRow)(const uint8* y_buf,
+                        const uint8* u_buf,
+                        const uint8* v_buf,
+                        uint8* rgb_buf,
+                        int width) = I422ToARGBRow_C;
+#if defined(HAS_I422TOARGBROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    I422ToARGBRow = I422ToARGBRow_Any_SSSE3;
+    if (IS_ALIGNED(src_width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    I422ToARGBRow = I422ToARGBRow_Any_AVX2;
+    if (IS_ALIGNED(src_width, 16)) {
+      I422ToARGBRow = I422ToARGBRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    I422ToARGBRow = I422ToARGBRow_Any_NEON;
+    if (IS_ALIGNED(src_width, 8)) {
+      I422ToARGBRow = I422ToARGBRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_I422TOARGBROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && IS_ALIGNED(src_width, 4) &&
+      IS_ALIGNED(src_y, 4) && IS_ALIGNED(src_stride_y, 4) &&
+      IS_ALIGNED(src_u, 2) && IS_ALIGNED(src_stride_u, 2) &&
+      IS_ALIGNED(src_v, 2) && IS_ALIGNED(src_stride_v, 2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    I422ToARGBRow = I422ToARGBRow_MIPS_DSPR2;
+  }
+#endif
+
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width, 8)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride_argb, 4)) {
+    InterpolateRow = InterpolateRow_MIPS_DSPR2;
+  }
+#endif
+
+  void (*ScaleARGBFilterCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      filtering ? ScaleARGBFilterCols_C : ScaleARGBCols_C;
+  if (src_width >= 32768) {
+    ScaleARGBFilterCols = filtering ?
+        ScaleARGBFilterCols64_C : ScaleARGBCols64_C;
+  }
+#if defined(HAS_SCALEARGBFILTERCOLS_SSSE3)
+  if (filtering && TestCpuFlag(kCpuHasSSSE3) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_SSSE3;
+  }
+#endif
+#if defined(HAS_SCALEARGBFILTERCOLS_NEON)
+  if (filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBFilterCols = ScaleARGBFilterCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 4)) {
+      ScaleARGBFilterCols = ScaleARGBFilterCols_NEON;
+    }
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (!filtering && TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBFilterCols = ScaleARGBCols_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_NEON)
+  if (!filtering && TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBFilterCols = ScaleARGBCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBCols_NEON;
+    }
+  }
+#endif
+  if (!filtering && src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBFilterCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBFilterCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  const int max_y = (src_height - 1) << 16;
+  if (y > max_y) {
+    y = max_y;
+  }
+  const int kYShift = 1;  // Shift Y by 1 to convert Y plane to UV coordinate.
+  int yi = y >> 16;
+  int uv_yi = yi >> kYShift;
+  const uint8* src_row_y = src_y + yi * src_stride_y;
+  const uint8* src_row_u = src_u + uv_yi * src_stride_u;
+  const uint8* src_row_v = src_v + uv_yi * src_stride_v;
+
+  // Allocate 2 rows of ARGB.
+  const int kRowSize = (dst_width * 4 + 31) & ~31;
+  align_buffer_64(row, kRowSize * 2);
+
+  // Allocate 1 row of ARGB for source conversion.
+  align_buffer_64(argb_row, src_width * 4);
+
+  uint8* rowptr = row;
+  int rowstride = kRowSize;
+  int lasty = yi;
+
+  // TODO(fbarchard): Convert first 2 rows of YUV to ARGB.
+  ScaleARGBFilterCols(rowptr, src_row_y, dst_width, x, dx);
+  if (src_height > 1) {
+    src_row_y += src_stride_y;
+    if (yi & 1) {
+      src_row_u += src_stride_u;
+      src_row_v += src_stride_v;
+    }
+  }
+  ScaleARGBFilterCols(rowptr + rowstride, src_row_y, dst_width, x, dx);
+  if (src_height > 2) {
+    src_row_y += src_stride_y;
+    if (!(yi & 1)) {
+      src_row_u += src_stride_u;
+      src_row_v += src_stride_v;
+    }
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    yi = y >> 16;
+    if (yi != lasty) {
+      if (y > max_y) {
+        y = max_y;
+        yi = y >> 16;
+        uv_yi = yi >> kYShift;
+        src_row_y = src_y + yi * src_stride_y;
+        src_row_u = src_u + uv_yi * src_stride_u;
+        src_row_v = src_v + uv_yi * src_stride_v;
+      }
+      if (yi != lasty) {
+        // TODO(fbarchard): Convert the clipped region of row.
+        I422ToARGBRow(src_row_y, src_row_u, src_row_v, argb_row, src_width);
+        ScaleARGBFilterCols(rowptr, argb_row, dst_width, x, dx);
+        rowptr += rowstride;
+        rowstride = -rowstride;
+        lasty = yi;
+        src_row_y += src_stride_y;
+        if (yi & 1) {
+          src_row_u += src_stride_u;
+          src_row_v += src_stride_v;
+        }
+      }
+    }
+    if (filtering == kFilterLinear) {
+      InterpolateRow(dst_argb, rowptr, 0, dst_width * 4, 0);
+    } else {
+      int yf = (y >> 8) & 255;
+      InterpolateRow(dst_argb, rowptr, rowstride, dst_width * 4, yf);
+    }
+    dst_argb += dst_stride_argb;
+    y += dy;
+  }
+  free_aligned_buffer_64(row);
+  free_aligned_buffer_64(row_argb);
+}
+#endif
+
+// Scale ARGB to/from any dimensions, without interpolation.
+// Fixed point math is used for performance: The upper 16 bits
+// of x and dx is the integer part of the source position and
+// the lower 16 bits are the fixed decimal part.
+
+static void ScaleARGBSimple(int src_width, int src_height,
+                            int dst_width, int dst_height,
+                            int src_stride, int dst_stride,
+                            const uint8* src_argb, uint8* dst_argb,
+                            int x, int dx, int y, int dy) {
+  int j;
+  void (*ScaleARGBCols)(uint8* dst_argb, const uint8* src_argb,
+      int dst_width, int x, int dx) =
+      (src_width >= 32768) ? ScaleARGBCols64_C : ScaleARGBCols_C;
+#if defined(HAS_SCALEARGBCOLS_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && src_width < 32768) {
+    ScaleARGBCols = ScaleARGBCols_SSE2;
+  }
+#endif
+#if defined(HAS_SCALEARGBCOLS_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleARGBCols = ScaleARGBCols_Any_NEON;
+    if (IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBCols = ScaleARGBCols_NEON;
+    }
+  }
+#endif
+  if (src_width * 2 == dst_width && x < 0x8000) {
+    ScaleARGBCols = ScaleARGBColsUp2_C;
+#if defined(HAS_SCALEARGBCOLSUP2_SSE2)
+    if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(dst_width, 8)) {
+      ScaleARGBCols = ScaleARGBColsUp2_SSE2;
+    }
+#endif
+  }
+
+  for (j = 0; j < dst_height; ++j) {
+    ScaleARGBCols(dst_argb, src_argb + (y >> 16) * src_stride,
+                  dst_width, x, dx);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
+// ScaleARGB a ARGB.
+// This function in turn calls a scaling function
+// suitable for handling the desired resolutions.
+static void ScaleARGB(const uint8* src, int src_stride,
+                      int src_width, int src_height,
+                      uint8* dst, int dst_stride,
+                      int dst_width, int dst_height,
+                      int clip_x, int clip_y, int clip_width, int clip_height,
+                      enum FilterMode filtering) {
+  // Initial source x/y coordinate and step values as 16.16 fixed point.
+  int x = 0;
+  int y = 0;
+  int dx = 0;
+  int dy = 0;
+  // ARGB does not support box filter yet, but allow the user to pass it.
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
+  // Negative src_height means invert the image.
+  if (src_height < 0) {
+    src_height = -src_height;
+    src = src + (src_height - 1) * src_stride;
+    src_stride = -src_stride;
+  }
+  ScaleSlope(src_width, src_height, dst_width, dst_height, filtering,
+             &x, &y, &dx, &dy);
+  src_width = Abs(src_width);
+  if (clip_x) {
+    int64 clipf = (int64)(clip_x) * dx;
+    x += (clipf & 0xffff);
+    src += (clipf >> 16) * 4;
+    dst += clip_x * 4;
+  }
+  if (clip_y) {
+    int64 clipf = (int64)(clip_y) * dy;
+    y += (clipf & 0xffff);
+    src += (clipf >> 16) * src_stride;
+    dst += clip_y * dst_stride;
+  }
+
+  // Special case for integer step values.
+  if (((dx | dy) & 0xffff) == 0) {
+    if (!dx || !dy) {  // 1 pixel wide and/or tall.
+      filtering = kFilterNone;
+    } else {
+      // Optimized even scale down. ie 2, 4, 6, 8, 10x.
+      if (!(dx & 0x10000) && !(dy & 0x10000)) {
+        if (dx == 0x20000) {
+          // Optimized 1/2 downsample.
+          ScaleARGBDown2(src_width, src_height,
+                         clip_width, clip_height,
+                         src_stride, dst_stride, src, dst,
+                         x, dx, y, dy, filtering);
+          return;
+        }
+        if (dx == 0x40000 && filtering == kFilterBox) {
+          // Optimized 1/4 box downsample.
+          ScaleARGBDown4Box(src_width, src_height,
+                            clip_width, clip_height,
+                            src_stride, dst_stride, src, dst,
+                            x, dx, y, dy);
+          return;
+        }
+        ScaleARGBDownEven(src_width, src_height,
+                          clip_width, clip_height,
+                          src_stride, dst_stride, src, dst,
+                          x, dx, y, dy, filtering);
+        return;
+      }
+      // Optimized odd scale down. ie 3, 5, 7, 9x.
+      if ((dx & 0x10000) && (dy & 0x10000)) {
+        filtering = kFilterNone;
+        if (dx == 0x10000 && dy == 0x10000) {
+          // Straight copy.
+          ARGBCopy(src + (y >> 16) * src_stride + (x >> 16) * 4, src_stride,
+                   dst, dst_stride, clip_width, clip_height);
+          return;
+        }
+      }
+    }
+  }
+  if (dx == 0x10000 && (x & 0xffff) == 0) {
+    // Arbitrary scale vertically, but unscaled vertically.
+    ScalePlaneVertical(src_height,
+                       clip_width, clip_height,
+                       src_stride, dst_stride, src, dst,
+                       x, y, dy, 4, filtering);
+    return;
+  }
+  if (filtering && dy < 65536) {
+    ScaleARGBBilinearUp(src_width, src_height,
+                        clip_width, clip_height,
+                        src_stride, dst_stride, src, dst,
+                        x, dx, y, dy, filtering);
+    return;
+  }
+  if (filtering) {
+    ScaleARGBBilinearDown(src_width, src_height,
+                          clip_width, clip_height,
+                          src_stride, dst_stride, src, dst,
+                          x, dx, y, dy, filtering);
+    return;
+  }
+  ScaleARGBSimple(src_width, src_height, clip_width, clip_height,
+                  src_stride, dst_stride, src, dst,
+                  x, dx, y, dy);
+}
+
+LIBYUV_API
+int ARGBScaleClip(const uint8* src_argb, int src_stride_argb,
+                  int src_width, int src_height,
+                  uint8* dst_argb, int dst_stride_argb,
+                  int dst_width, int dst_height,
+                  int clip_x, int clip_y, int clip_width, int clip_height,
+                  enum FilterMode filtering) {
+  if (!src_argb || src_width == 0 || src_height == 0 ||
+      !dst_argb || dst_width <= 0 || dst_height <= 0 ||
+      clip_x < 0 || clip_y < 0 ||
+      clip_width > 32768 || clip_height > 32768 ||
+      (clip_x + clip_width) > dst_width ||
+      (clip_y + clip_height) > dst_height) {
+    return -1;
+  }
+  ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
+            dst_argb, dst_stride_argb, dst_width, dst_height,
+            clip_x, clip_y, clip_width, clip_height, filtering);
+  return 0;
+}
+
+// Scale an ARGB image.
+LIBYUV_API
+int ARGBScale(const uint8* src_argb, int src_stride_argb,
+              int src_width, int src_height,
+              uint8* dst_argb, int dst_stride_argb,
+              int dst_width, int dst_height,
+              enum FilterMode filtering) {
+  if (!src_argb || src_width == 0 || src_height == 0 ||
+      src_width > 32768 || src_height > 32768 ||
+      !dst_argb || dst_width <= 0 || dst_height <= 0) {
+    return -1;
+  }
+  ScaleARGB(src_argb, src_stride_argb, src_width, src_height,
+            dst_argb, dst_stride_argb, dst_width, dst_height,
+            0, 0, dst_width, dst_height, filtering);
+  return 0;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/scale_common.cc b/third_party/aom/third_party/libyuv/source/scale_common.cc
new file mode 100644
index 0000000000..1711f3d54c
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/scale_common.cc
@@ -0,0 +1,1137 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+
+#include <assert.h>
+#include <string.h>
+
+#include "libyuv/cpu_id.h"
+#include "libyuv/planar_functions.h"  // For CopyARGB
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+static __inline int Abs(int v) {
+  return v >= 0 ? v : -v;
+}
+
+// CPU agnostic row functions
+void ScaleRowDown2_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[1];
+    dst[1] = src_ptr[3];
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[1];
+  }
+}
+
+void ScaleRowDown2_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[1];
+    dst[1] = src_ptr[3];
+    dst += 2;
+    src_ptr += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[1];
+  }
+}
+
+void ScaleRowDown2Linear_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+    dst[1] = (s[2] + s[3] + 1) >> 1;
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+  }
+}
+
+void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                              uint16* dst, int dst_width) {
+  const uint16* s = src_ptr;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+    dst[1] = (s[2] + s[3] + 1) >> 1;
+    dst += 2;
+    s += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + 1) >> 1;
+  }
+}
+
+void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+  }
+}
+
+void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+  }
+}
+
+void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                     uint8* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[2];
+    dst[1] = src_ptr[6];
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[2];
+  }
+}
+
+void ScaleRowDown4_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                        uint16* dst, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src_ptr[2];
+    dst[1] = src_ptr[6];
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = src_ptr[2];
+  }
+}
+
+void ScaleRowDown4Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+    dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
+             src_ptr[stride + 4] + src_ptr[stride + 5] +
+             src_ptr[stride + 6] + src_ptr[stride + 7] +
+             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
+             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
+             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
+             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
+             8) >> 4;
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+  }
+}
+
+void ScaleRowDown4Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                           uint16* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+    dst[1] = (src_ptr[4] + src_ptr[5] + src_ptr[6] + src_ptr[7] +
+             src_ptr[stride + 4] + src_ptr[stride + 5] +
+             src_ptr[stride + 6] + src_ptr[stride + 7] +
+             src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5] +
+             src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7] +
+             src_ptr[stride * 3 + 4] + src_ptr[stride * 3 + 5] +
+             src_ptr[stride * 3 + 6] + src_ptr[stride * 3 + 7] +
+             8) >> 4;
+    dst += 2;
+    src_ptr += 8;
+  }
+  if (dst_width & 1) {
+    dst[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] + src_ptr[3] +
+             src_ptr[stride + 0] + src_ptr[stride + 1] +
+             src_ptr[stride + 2] + src_ptr[stride + 3] +
+             src_ptr[stride * 2 + 0] + src_ptr[stride * 2 + 1] +
+             src_ptr[stride * 2 + 2] + src_ptr[stride * 2 + 3] +
+             src_ptr[stride * 3 + 0] + src_ptr[stride * 3 + 1] +
+             src_ptr[stride * 3 + 2] + src_ptr[stride * 3 + 3] +
+             8) >> 4;
+  }
+}
+
+void ScaleRowDown34_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width) {
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[1];
+    dst[2] = src_ptr[3];
+    dst += 3;
+    src_ptr += 4;
+  }
+}
+
+void ScaleRowDown34_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width) {
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[1];
+    dst[2] = src_ptr[3];
+    dst += 3;
+    src_ptr += 4;
+  }
+}
+
+// Filter rows 0 and 1 together, 3 : 1
+void ScaleRowDown34_0_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 * 3 + b0 + 2) >> 2;
+    d[1] = (a1 * 3 + b1 + 2) >> 2;
+    d[2] = (a2 * 3 + b2 + 2) >> 2;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+void ScaleRowDown34_0_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 * 3 + b0 + 2) >> 2;
+    d[1] = (a1 * 3 + b1 + 2) >> 2;
+    d[2] = (a2 * 3 + b2 + 2) >> 2;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+// Filter rows 1 and 2 together, 1 : 1
+void ScaleRowDown34_1_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* d, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint8 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint8 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint8 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint8 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint8 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint8 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 + b0 + 1) >> 1;
+    d[1] = (a1 + b1 + 1) >> 1;
+    d[2] = (a2 + b2 + 1) >> 1;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+void ScaleRowDown34_1_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* d, int dst_width) {
+  const uint16* s = src_ptr;
+  const uint16* t = src_ptr + src_stride;
+  int x;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (x = 0; x < dst_width; x += 3) {
+    uint16 a0 = (s[0] * 3 + s[1] * 1 + 2) >> 2;
+    uint16 a1 = (s[1] * 1 + s[2] * 1 + 1) >> 1;
+    uint16 a2 = (s[2] * 1 + s[3] * 3 + 2) >> 2;
+    uint16 b0 = (t[0] * 3 + t[1] * 1 + 2) >> 2;
+    uint16 b1 = (t[1] * 1 + t[2] * 1 + 1) >> 1;
+    uint16 b2 = (t[2] * 1 + t[3] * 3 + 2) >> 2;
+    d[0] = (a0 + b0 + 1) >> 1;
+    d[1] = (a1 + b1 + 1) >> 1;
+    d[2] = (a2 + b2 + 1) >> 1;
+    d += 3;
+    s += 4;
+    t += 4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                 int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[0] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr[1] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[x >> 16];
+  }
+}
+
+void ScaleCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                    int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[0] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr[1] = src_ptr[x >> 16];
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleColsUp2_C(uint8* dst_ptr, const uint8* src_ptr,
+                    int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[1] = dst_ptr[0] = src_ptr[0];
+    src_ptr += 1;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[0];
+  }
+}
+
+void ScaleColsUp2_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst_ptr[1] = dst_ptr[0] = src_ptr[0];
+    src_ptr += 1;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    dst_ptr[0] = src_ptr[0];
+  }
+}
+
+// (1-f)a + fb can be replaced with a + f(b-a)
+#define BLENDER(a, b, f) (uint8)((int)(a) + \
+    ((int)(f) * ((int)(b) - (int)(a)) >> 16))
+
+void ScaleFilterCols_C(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+
+void ScaleFilterCols64_C(uint8* dst_ptr, const uint8* src_ptr,
+                         int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+#undef BLENDER
+
+#define BLENDER(a, b, f) (uint16)((int)(a) + \
+    ((int)(f) * ((int)(b) - (int)(a)) >> 16))
+
+void ScaleFilterCols_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                       int dst_width, int x, int dx) {
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+
+void ScaleFilterCols64_16_C(uint16* dst_ptr, const uint16* src_ptr,
+                         int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    xi = x >> 16;
+    a = src_ptr[xi];
+    b = src_ptr[xi + 1];
+    dst_ptr[1] = BLENDER(a, b, x & 0xffff);
+    x += dx;
+    dst_ptr += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int a = src_ptr[xi];
+    int b = src_ptr[xi + 1];
+    dst_ptr[0] = BLENDER(a, b, x & 0xffff);
+  }
+}
+#undef BLENDER
+
+void ScaleRowDown38_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                      uint8* dst, int dst_width) {
+  int x;
+  assert(dst_width % 3 == 0);
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[3];
+    dst[2] = src_ptr[6];
+    dst += 3;
+    src_ptr += 8;
+  }
+}
+
+void ScaleRowDown38_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                         uint16* dst, int dst_width) {
+  int x;
+  assert(dst_width % 3 == 0);
+  for (x = 0; x < dst_width; x += 3) {
+    dst[0] = src_ptr[0];
+    dst[1] = src_ptr[3];
+    dst[2] = src_ptr[6];
+    dst += 3;
+    src_ptr += 8;
+  }
+}
+
+// 8x3 -> 3x1
+void ScaleRowDown38_3_Box_C(const uint8* src_ptr,
+                            ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+        (65536 / 9) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+        (65536 / 9) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7] +
+        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+        (65536 / 6) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleRowDown38_3_Box_16_C(const uint16* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2] + src_ptr[stride * 2 + 0] +
+        src_ptr[stride * 2 + 1] + src_ptr[stride * 2 + 2]) *
+        (65536 / 9) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5] + src_ptr[stride * 2 + 3] +
+        src_ptr[stride * 2 + 4] + src_ptr[stride * 2 + 5]) *
+        (65536 / 9) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7] +
+        src_ptr[stride * 2 + 6] + src_ptr[stride * 2 + 7]) *
+        (65536 / 6) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+// 8x2 -> 3x1
+void ScaleRowDown38_2_Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2]) * (65536 / 6) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5]) * (65536 / 6) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7]) *
+        (65536 / 4) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleRowDown38_2_Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
+                               uint16* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  int i;
+  assert((dst_width % 3 == 0) && (dst_width > 0));
+  for (i = 0; i < dst_width; i += 3) {
+    dst_ptr[0] = (src_ptr[0] + src_ptr[1] + src_ptr[2] +
+        src_ptr[stride + 0] + src_ptr[stride + 1] +
+        src_ptr[stride + 2]) * (65536 / 6) >> 16;
+    dst_ptr[1] = (src_ptr[3] + src_ptr[4] + src_ptr[5] +
+        src_ptr[stride + 3] + src_ptr[stride + 4] +
+        src_ptr[stride + 5]) * (65536 / 6) >> 16;
+    dst_ptr[2] = (src_ptr[6] + src_ptr[7] +
+        src_ptr[stride + 6] + src_ptr[stride + 7]) *
+        (65536 / 4) >> 16;
+    src_ptr += 8;
+    dst_ptr += 3;
+  }
+}
+
+void ScaleAddRow_C(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
+  int x;
+  assert(src_width > 0);
+  for (x = 0; x < src_width - 1; x += 2) {
+    dst_ptr[0] += src_ptr[0];
+    dst_ptr[1] += src_ptr[1];
+    src_ptr += 2;
+    dst_ptr += 2;
+  }
+  if (src_width & 1) {
+    dst_ptr[0] += src_ptr[0];
+  }
+}
+
+void ScaleAddRow_16_C(const uint16* src_ptr, uint32* dst_ptr, int src_width) {
+  int x;
+  assert(src_width > 0);
+  for (x = 0; x < src_width - 1; x += 2) {
+    dst_ptr[0] += src_ptr[0];
+    dst_ptr[1] += src_ptr[1];
+    src_ptr += 2;
+    dst_ptr += 2;
+  }
+  if (src_width & 1) {
+    dst_ptr[0] += src_ptr[0];
+  }
+}
+
+void ScaleARGBRowDown2_C(const uint8* src_argb,
+                         ptrdiff_t src_stride,
+                         uint8* dst_argb, int dst_width) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[1];
+    dst[1] = src[3];
+    src += 4;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[1];
+  }
+}
+
+void ScaleARGBRowDown2Linear_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] + 1) >> 1;
+    dst_argb[1] = (src_argb[1] + src_argb[5] + 1) >> 1;
+    dst_argb[2] = (src_argb[2] + src_argb[6] + 1) >> 1;
+    dst_argb[3] = (src_argb[3] + src_argb[7] + 1) >> 1;
+    src_argb += 8;
+    dst_argb += 4;
+  }
+}
+
+void ScaleARGBRowDown2Box_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] +
+                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] +
+                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] +
+                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] +
+                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    src_argb += 8;
+    dst_argb += 4;
+  }
+}
+
+void ScaleARGBRowDownEven_C(const uint8* src_argb, ptrdiff_t src_stride,
+                            int src_stepx,
+                            uint8* dst_argb, int dst_width) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+
+  int x;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = src[0];
+    dst[1] = src[src_stepx];
+    src += src_stepx * 2;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+void ScaleARGBRowDownEvenBox_C(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  int x;
+  for (x = 0; x < dst_width; ++x) {
+    dst_argb[0] = (src_argb[0] + src_argb[4] +
+                  src_argb[src_stride] + src_argb[src_stride + 4] + 2) >> 2;
+    dst_argb[1] = (src_argb[1] + src_argb[5] +
+                  src_argb[src_stride + 1] + src_argb[src_stride + 5] + 2) >> 2;
+    dst_argb[2] = (src_argb[2] + src_argb[6] +
+                  src_argb[src_stride + 2] + src_argb[src_stride + 6] + 2) >> 2;
+    dst_argb[3] = (src_argb[3] + src_argb[7] +
+                  src_argb[src_stride + 3] + src_argb[src_stride + 7] + 2) >> 2;
+    src_argb += src_stepx * 4;
+    dst_argb += 4;
+  }
+}
+
+// Scales a single row of pixels using point sampling.
+void ScaleARGBCols_C(uint8* dst_argb, const uint8* src_argb,
+                     int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+void ScaleARGBCols64_C(uint8* dst_argb, const uint8* src_argb,
+                       int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[0] = src[x >> 16];
+    x += dx;
+    dst[1] = src[x >> 16];
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[x >> 16];
+  }
+}
+
+// Scales a single row of pixels up by 2x using point sampling.
+void ScaleARGBColsUp2_C(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    dst[1] = dst[0] = src[0];
+    src += 1;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    dst[0] = src[0];
+  }
+}
+
+// Mimics SSSE3 blender
+#define BLENDER1(a, b, f) ((a) * (0x7f ^ f) + (b) * f) >> 7
+#define BLENDERC(a, b, f, s) (uint32)( \
+    BLENDER1(((a) >> s) & 255, ((b) >> s) & 255, f) << s)
+#define BLENDER(a, b, f) \
+    BLENDERC(a, b, f, 24) | BLENDERC(a, b, f, 16) | \
+    BLENDERC(a, b, f, 8) | BLENDERC(a, b, f, 0)
+
+void ScaleARGBFilterCols_C(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+
+void ScaleARGBFilterCols64_C(uint8* dst_argb, const uint8* src_argb,
+                             int dst_width, int x32, int dx) {
+  int64 x = (int64)(x32);
+  const uint32* src = (const uint32*)(src_argb);
+  uint32* dst = (uint32*)(dst_argb);
+  int j;
+  for (j = 0; j < dst_width - 1; j += 2) {
+    int64 xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+    x += dx;
+    xi = x >> 16;
+    xf = (x >> 9) & 0x7f;
+    a = src[xi];
+    b = src[xi + 1];
+    dst[1] = BLENDER(a, b, xf);
+    x += dx;
+    dst += 2;
+  }
+  if (dst_width & 1) {
+    int64 xi = x >> 16;
+    int xf = (x >> 9) & 0x7f;
+    uint32 a = src[xi];
+    uint32 b = src[xi + 1];
+    dst[0] = BLENDER(a, b, xf);
+  }
+}
+#undef BLENDER1
+#undef BLENDERC
+#undef BLENDER
+
+// Scale plane vertically with bilinear interpolation.
+void ScalePlaneVertical(int src_height,
+                        int dst_width, int dst_height,
+                        int src_stride, int dst_stride,
+                        const uint8* src_argb, uint8* dst_argb,
+                        int x, int y, int dy,
+                        int bpp, enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher bpp.
+  int dst_width_bytes = dst_width * bpp;
+  void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(bpp >= 1 && bpp <= 4);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * bpp;
+#if defined(HAS_INTERPOLATEROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_SSE2;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_SSSE3;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_AVX2;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_NEON;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width_bytes, 4)) {
+      InterpolateRow = InterpolateRow_MIPS_DSPR2;
+    }
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow(dst_argb, src_argb + yi * src_stride,
+                   src_stride, dst_width_bytes, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+void ScalePlaneVertical_16(int src_height,
+                           int dst_width, int dst_height,
+                           int src_stride, int dst_stride,
+                           const uint16* src_argb, uint16* dst_argb,
+                           int x, int y, int dy,
+                           int wpp, enum FilterMode filtering) {
+  // TODO(fbarchard): Allow higher wpp.
+  int dst_width_words = dst_width * wpp;
+  void (*InterpolateRow)(uint16* dst_argb, const uint16* src_argb,
+      ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
+      InterpolateRow_16_C;
+  const int max_y = (src_height > 1) ? ((src_height - 1) << 16) - 1 : 0;
+  int j;
+  assert(wpp >= 1 && wpp <= 2);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  src_argb += (x >> 16) * wpp;
+#if defined(HAS_INTERPOLATEROW_16_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    InterpolateRow = InterpolateRow_Any_16_SSE2;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_16_SSE2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_SSSE3)
+  if (TestCpuFlag(kCpuHasSSSE3)) {
+    InterpolateRow = InterpolateRow_Any_16_SSSE3;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_16_SSSE3;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2)) {
+    InterpolateRow = InterpolateRow_Any_16_AVX2;
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
+      InterpolateRow = InterpolateRow_16_AVX2;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_NEON)
+  if (TestCpuFlag(kCpuHasNEON)) {
+    InterpolateRow = InterpolateRow_Any_16_NEON;
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
+      InterpolateRow = InterpolateRow_16_NEON;
+    }
+  }
+#endif
+#if defined(HAS_INTERPOLATEROW_16_MIPS_DSPR2)
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) &&
+      IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
+      IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
+    InterpolateRow = InterpolateRow_Any_16_MIPS_DSPR2;
+    if (IS_ALIGNED(dst_width_bytes, 4)) {
+      InterpolateRow = InterpolateRow_16_MIPS_DSPR2;
+    }
+  }
+#endif
+  for (j = 0; j < dst_height; ++j) {
+    int yi;
+    int yf;
+    if (y > max_y) {
+      y = max_y;
+    }
+    yi = y >> 16;
+    yf = filtering ? ((y >> 8) & 255) : 0;
+    InterpolateRow(dst_argb, src_argb + yi * src_stride,
+                   src_stride, dst_width_words, yf);
+    dst_argb += dst_stride;
+    y += dy;
+  }
+}
+
+// Simplify the filtering based on scale factors.
+enum FilterMode ScaleFilterReduce(int src_width, int src_height,
+                                  int dst_width, int dst_height,
+                                  enum FilterMode filtering) {
+  if (src_width < 0) {
+    src_width = -src_width;
+  }
+  if (src_height < 0) {
+    src_height = -src_height;
+  }
+  if (filtering == kFilterBox) {
+    // If scaling both axis to 0.5 or larger, switch from Box to Bilinear.
+    if (dst_width * 2 >= src_width && dst_height * 2 >= src_height) {
+      filtering = kFilterBilinear;
+    }
+  }
+  if (filtering == kFilterBilinear) {
+    if (src_height == 1) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to Linear.
+    if (dst_height == src_height || dst_height * 3 == src_height) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Remove 1 pixel wide filter restriction, which is to
+    // avoid reading 2 pixels horizontally that causes memory exception.
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+  }
+  if (filtering == kFilterLinear) {
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to None.
+    if (dst_width == src_width || dst_width * 3 == src_width) {
+      filtering = kFilterNone;
+    }
+  }
+  return filtering;
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_C(int num, int div) {
+  return (int)(((int64)(num) << 16) / div);
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv1_C(int num, int div) {
+  return (int)((((int64)(num) << 16) - 0x00010001) /
+                          (div - 1));
+}
+
+#define CENTERSTART(dx, s) (dx < 0) ? -((-dx >> 1) + s) : ((dx >> 1) + s)
+
+// Compute slope values for stepping.
+void ScaleSlope(int src_width, int src_height,
+                int dst_width, int dst_height,
+                enum FilterMode filtering,
+                int* x, int* y, int* dx, int* dy) {
+  assert(x != NULL);
+  assert(y != NULL);
+  assert(dx != NULL);
+  assert(dy != NULL);
+  assert(src_width != 0);
+  assert(src_height != 0);
+  assert(dst_width > 0);
+  assert(dst_height > 0);
+  // Check for 1 pixel and avoid FixedDiv overflow.
+  if (dst_width == 1 && src_width >= 32768) {
+    dst_width = src_width;
+  }
+  if (dst_height == 1 && src_height >= 32768) {
+    dst_height = src_height;
+  }
+  if (filtering == kFilterBox) {
+    // Scale step for point sampling duplicates all pixels equally.
+    *dx = FixedDiv(Abs(src_width), dst_width);
+    *dy = FixedDiv(src_height, dst_height);
+    *x = 0;
+    *y = 0;
+  } else if (filtering == kFilterBilinear) {
+    // Scale step for bilinear sampling renders last pixel once for upsample.
+    if (dst_width <= Abs(src_width)) {
+      *dx = FixedDiv(Abs(src_width), dst_width);
+      *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_width > 1) {
+      *dx = FixedDiv1(Abs(src_width), dst_width);
+      *x = 0;
+    }
+    if (dst_height <= src_height) {
+      *dy = FixedDiv(src_height,  dst_height);
+      *y = CENTERSTART(*dy, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_height > 1) {
+      *dy = FixedDiv1(src_height, dst_height);
+      *y = 0;
+    }
+  } else if (filtering == kFilterLinear) {
+    // Scale step for bilinear sampling renders last pixel once for upsample.
+    if (dst_width <= Abs(src_width)) {
+      *dx = FixedDiv(Abs(src_width), dst_width);
+      *x = CENTERSTART(*dx, -32768);  // Subtract 0.5 (32768) to center filter.
+    } else if (dst_width > 1) {
+      *dx = FixedDiv1(Abs(src_width), dst_width);
+      *x = 0;
+    }
+    *dy = FixedDiv(src_height, dst_height);
+    *y = *dy >> 1;
+  } else {
+    // Scale step for point sampling duplicates all pixels equally.
+    *dx = FixedDiv(Abs(src_width), dst_width);
+    *dy = FixedDiv(src_height, dst_height);
+    *x = CENTERSTART(*dx, 0);
+    *y = CENTERSTART(*dy, 0);
+  }
+  // Negative src_width means horizontally mirror.
+  if (src_width < 0) {
+    *x += (dst_width - 1) * *dx;
+    *dx = -*dx;
+    // src_width = -src_width;   // Caller must do this.
+  }
+}
+#undef CENTERSTART
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/scale_gcc.cc b/third_party/aom/third_party/libyuv/source/scale_gcc.cc
new file mode 100644
index 0000000000..8a6ac54592
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/scale_gcc.cc
@@ -0,0 +1,1089 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC x86 and x64.
+#if !defined(LIBYUV_DISABLE_X86) && (defined(__x86_64__) || defined(__i386__))
+
+// Offsets for source bytes 0 to 9
+static uvec8 kShuf0 =
+  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
+static uvec8 kShuf1 =
+  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf2 =
+  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 0 to 10
+static uvec8 kShuf01 =
+  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+
+// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
+static uvec8 kShuf11 =
+  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf21 =
+  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+
+// Coefficients for source bytes 0 to 10
+static uvec8 kMadd01 =
+  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+
+// Coefficients for source bytes 10 to 21
+static uvec8 kMadd11 =
+  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+
+// Coefficients for source bytes 21 to 31
+static uvec8 kMadd21 =
+  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+
+// Coefficients for source bytes 21 to 31
+static vec16 kRound34 =
+  { 2, 2, 2, 2, 2, 2, 2, 2 };
+
+static uvec8 kShuf38a =
+  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+static uvec8 kShuf38b =
+  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 0,1,2
+static uvec8 kShufAc =
+  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 3,4,5
+static uvec8 kShufAc3 =
+  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x3 and 2x3
+static uvec16 kScaleAc33 =
+  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+
+// Arrange first value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb0 =
+  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+
+// Arrange second value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb1 =
+  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+
+// Arrange third value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb2 =
+  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x2 and 2x2
+static uvec16 kScaleAb2 =
+  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+
+// GCC versions of row functions are verbatim conversions from Visual C.
+// Generated using gcc disassembly on Visual C object file:
+// objdump -D yuvscaler.obj >yuvscaler.txt
+
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :: "memory", "cc", "xmm0", "xmm1"
+  );
+}
+
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10, 0) ",%%xmm1  \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrlw     $0x8,%%xmm5                     \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu  (%0,%3,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu  0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm5,%%xmm2                   \n"
+    "pand      %%xmm5,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x10,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"
+  );
+}
+
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "pcmpeqb   %%xmm5,%%xmm5                   \n"
+    "psrld     $0x18,%%xmm5                    \n"
+    "pslld     $0x10,%%xmm5                    \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pand      %%xmm5,%%xmm0                   \n"
+    "pand      %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm5"
+  );
+}
+
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  intptr_t stridex3 = 0;
+  asm volatile (
+    "pcmpeqb   %%xmm7,%%xmm7                   \n"
+    "psrlw     $0x8,%%xmm7                     \n"
+    "lea       " MEMLEA4(0x00,4,4,2) ",%3      \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,4,1,xmm2)           //  movdqu  (%0,%4,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,1,xmm3)           //  movdqu  0x10(%0,%4,1),%%xmm3
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x00,0,4,2,xmm2)           //  movdqu  (%0,%4,2),%%xmm2
+    MEMOPREG(movdqu,0x10,0,4,2,xmm3)           //  movdqu  0x10(%0,%4,2),%%xmm3
+    MEMOPREG(movdqu,0x00,0,3,1,xmm4)           //  movdqu  (%0,%3,1),%%xmm4
+    MEMOPREG(movdqu,0x10,0,3,1,xmm5)           //  movdqu  0x10(%0,%3,1),%%xmm5
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm4,%%xmm2                   \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm5,%%xmm3                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "movdqa    %%xmm1,%%xmm3                   \n"
+    "psrlw     $0x8,%%xmm1                     \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "pand      %%xmm7,%%xmm3                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "pavgw     %%xmm3,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "psrlw     $0x8,%%xmm0                     \n"
+    "pand      %%xmm7,%%xmm2                   \n"
+    "pavgw     %%xmm2,%%xmm0                   \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x8,1) ",%1            \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(dst_width),   // %2
+    "+r"(stridex3)     // %3
+  : "r"((intptr_t)(src_stride))    // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm7"
+  );
+}
+
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm3                       \n"
+    "movdqa    %1,%%xmm4                       \n"
+    "movdqa    %2,%%xmm5                       \n"
+  :
+  : "m"(kShuf0),  // %0
+    "m"(kShuf1),  // %1
+    "m"(kShuf2)   // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm2   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "palignr   $0x8,%%xmm0,%%xmm1              \n"
+    "pshufb    %%xmm3,%%xmm0                   \n"
+    "pshufb    %%xmm4,%%xmm1                   \n"
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "movq      %%xmm1," MEMACCESS2(0x8,1) "    \n"
+    "movq      %%xmm2," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  :: "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"
+  );
+}
+
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"  // kShuf01
+    "movdqa    %1,%%xmm3                       \n"  // kShuf11
+    "movdqa    %2,%%xmm4                       \n"  // kShuf21
+  :
+  : "m"(kShuf01),  // %0
+    "m"(kShuf11),  // %1
+    "m"(kShuf21)   // %2
+  );
+  asm volatile (
+    "movdqa    %0,%%xmm5                       \n"  // kMadd01
+    "movdqa    %1,%%xmm0                       \n"  // kMadd11
+    "movdqa    %2,%%xmm1                       \n"  // kRound34
+  :
+  : "m"(kMadd01),  // %0
+    "m"(kMadd11),  // %1
+    "m"(kRound34)  // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm7)           //  movdqu  (%0,%3),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
+    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3),%%xmm7
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm0,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x10,0,3,1,xmm7)           //  movdqu  0x10(%0,%3),%%xmm7
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm4,%%xmm6                   \n"
+    "pmaddubsw %4,%%xmm6                       \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  : "r"((intptr_t)(src_stride)),  // %3
+    "m"(kMadd21)     // %4
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"  // kShuf01
+    "movdqa    %1,%%xmm3                       \n"  // kShuf11
+    "movdqa    %2,%%xmm4                       \n"  // kShuf21
+  :
+  : "m"(kShuf01),  // %0
+    "m"(kShuf11),  // %1
+    "m"(kShuf21)   // %2
+  );
+  asm volatile (
+    "movdqa    %0,%%xmm5                       \n"  // kMadd01
+    "movdqa    %1,%%xmm0                       \n"  // kMadd11
+    "movdqa    %2,%%xmm1                       \n"  // kRound34
+  :
+  : "m"(kMadd01),  // %0
+    "m"(kMadd11),  // %1
+    "m"(kRound34)  // %2
+  );
+
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm6         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm7)           //  movdqu  (%0,%3,1),%%xmm7
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "pmaddubsw %%xmm5,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS(1) "         \n"
+    "movdqu    " MEMACCESS2(0x8,0) ",%%xmm6    \n"
+    MEMOPREG(movdqu,0x8,0,3,1,xmm7)            //  movdqu  0x8(%0,%3,1),%%xmm7
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "pmaddubsw %%xmm0,%%xmm6                   \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x8,1) "    \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm6   \n"
+    MEMOPREG(movdqu,0x10,0,3,1,xmm7)           //  movdqu  0x10(%0,%3,1),%%xmm7
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm6,%%xmm7                   \n"
+    "pavgb     %%xmm7,%%xmm6                   \n"
+    "pshufb    %%xmm4,%%xmm6                   \n"
+    "pmaddubsw %4,%%xmm6                       \n"
+    "paddsw    %%xmm1,%%xmm6                   \n"
+    "psrlw     $0x2,%%xmm6                     \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movq      %%xmm6," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x18,1) ",%1           \n"
+    "sub       $0x18,%2                        \n"
+    "jg        1b                              \n"
+    : "+r"(src_ptr),   // %0
+      "+r"(dst_ptr),   // %1
+      "+r"(dst_width)  // %2
+    : "r"((intptr_t)(src_stride)),  // %3
+      "m"(kMadd21)     // %4
+    : "memory", "cc", NACL_R14
+      "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %3,%%xmm4                       \n"
+    "movdqa    %4,%%xmm5                       \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "paddusb   %%xmm1,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(1) "         \n"
+    "movhlps   %%xmm0,%%xmm1                   \n"
+    "movd      %%xmm1," MEMACCESS2(0x8,1) "    \n"
+    "lea       " MEMLEA(0xc,1) ",%1            \n"
+    "sub       $0xc,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width)  // %2
+  : "m"(kShuf38a),   // %3
+    "m"(kShuf38b)    // %4
+  : "memory", "cc", "xmm0", "xmm1", "xmm4", "xmm5"
+  );
+}
+
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm4                       \n"
+    "movdqa    %3,%%xmm5                       \n"
+  :
+  : "m"(kShufAb0),   // %0
+    "m"(kShufAb1),   // %1
+    "m"(kShufAb2),   // %2
+    "m"(kScaleAb2)   // %3
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm1)           //  movdqu  (%0,%3,1),%%xmm1
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "pavgb     %%xmm1,%%xmm0                   \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "pshufb    %%xmm2,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "pshufb    %%xmm3,%%xmm6                   \n"
+    "paddusw   %%xmm6,%%xmm1                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "paddusw   %%xmm0,%%xmm1                   \n"
+    "pmulhuw   %%xmm5,%%xmm1                   \n"
+    "packuswb  %%xmm1,%%xmm1                   \n"
+    "movd      %%xmm1," MEMACCESS(1) "         \n"
+    "psrlq     $0x10,%%xmm1                    \n"
+    "movd      %%xmm1," MEMACCESS2(0x2,1) "    \n"
+    "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "sub       $0x6,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(dst_width)    // %2
+  : "r"((intptr_t)(src_stride))  // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movdqa    %0,%%xmm2                       \n"
+    "movdqa    %1,%%xmm3                       \n"
+    "movdqa    %2,%%xmm4                       \n"
+    "pxor      %%xmm5,%%xmm5                   \n"
+  :
+  : "m"(kShufAc),    // %0
+    "m"(kShufAc3),   // %1
+    "m"(kScaleAc33)  // %2
+  );
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm6)           //  movdqu  (%0,%3,1),%%xmm6
+    "movhlps   %%xmm0,%%xmm1                   \n"
+    "movhlps   %%xmm6,%%xmm7                   \n"
+    "punpcklbw %%xmm5,%%xmm0                   \n"
+    "punpcklbw %%xmm5,%%xmm1                   \n"
+    "punpcklbw %%xmm5,%%xmm6                   \n"
+    "punpcklbw %%xmm5,%%xmm7                   \n"
+    "paddusw   %%xmm6,%%xmm0                   \n"
+    "paddusw   %%xmm7,%%xmm1                   \n"
+    MEMOPREG(movdqu,0x00,0,3,2,xmm6)           //  movdqu  (%0,%3,2),%%xmm6
+    "lea       " MEMLEA(0x10,0) ",%0           \n"
+    "movhlps   %%xmm6,%%xmm7                   \n"
+    "punpcklbw %%xmm5,%%xmm6                   \n"
+    "punpcklbw %%xmm5,%%xmm7                   \n"
+    "paddusw   %%xmm6,%%xmm0                   \n"
+    "paddusw   %%xmm7,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm6                   \n"
+    "psrldq    $0x2,%%xmm0                     \n"
+    "paddusw   %%xmm0,%%xmm6                   \n"
+    "psrldq    $0x2,%%xmm0                     \n"
+    "paddusw   %%xmm0,%%xmm6                   \n"
+    "pshufb    %%xmm2,%%xmm6                   \n"
+    "movdqa    %%xmm1,%%xmm7                   \n"
+    "psrldq    $0x2,%%xmm1                     \n"
+    "paddusw   %%xmm1,%%xmm7                   \n"
+    "psrldq    $0x2,%%xmm1                     \n"
+    "paddusw   %%xmm1,%%xmm7                   \n"
+    "pshufb    %%xmm3,%%xmm7                   \n"
+    "paddusw   %%xmm7,%%xmm6                   \n"
+    "pmulhuw   %%xmm4,%%xmm6                   \n"
+    "packuswb  %%xmm6,%%xmm6                   \n"
+    "movd      %%xmm6," MEMACCESS(1) "         \n"
+    "psrlq     $0x10,%%xmm6                    \n"
+    "movd      %%xmm6," MEMACCESS2(0x2,1) "    \n"
+    "lea       " MEMLEA(0x6,1) ",%1            \n"
+    "sub       $0x6,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),    // %0
+    "+r"(dst_ptr),    // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"
+  );
+}
+
+// Reads 16xN bytes and produces 16 shorts at a time.
+void ScaleAddRows_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width, int src_height) {
+  int tmp_height = 0;
+  intptr_t tmp_src = 0;
+  asm volatile (
+    "mov       %0,%3                           \n"  // row pointer
+    "mov       %5,%2                           \n"  // height
+    "pxor      %%xmm0,%%xmm0                   \n"  // clear accumulators
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "pxor      %%xmm4,%%xmm4                   \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(3) ",%%xmm2         \n"
+    "add       %6,%3                           \n"
+    "movdqa    %%xmm2,%%xmm3                   \n"
+    "punpcklbw %%xmm4,%%xmm2                   \n"
+    "punpckhbw %%xmm4,%%xmm3                   \n"
+    "paddusw   %%xmm2,%%xmm0                   \n"
+    "paddusw   %%xmm3,%%xmm1                   \n"
+    "sub       $0x1,%2                         \n"
+    "jg        1b                              \n"
+
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,1) "   \n"
+    "lea       " MEMLEA(0x20,1) ",%1           \n"
+    "lea       " MEMLEA(0x10,0) ",%0           \n"  // src_ptr += 16
+    "mov       %0,%3                           \n"  // row pointer
+    "mov       %5,%2                           \n"  // height
+    "pxor      %%xmm0,%%xmm0                   \n"  // clear accumulators
+    "pxor      %%xmm1,%%xmm1                   \n"
+    "sub       $0x10,%4                        \n"
+    "jg        1b                              \n"
+  : "+r"(src_ptr),     // %0
+    "+r"(dst_ptr),     // %1
+    "+r"(tmp_height),  // %2
+    "+r"(tmp_src),     // %3
+    "+r"(src_width),   // %4
+    "+rm"(src_height)  // %5
+  : "rm"((intptr_t)(src_stride))  // %6
+  : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+
+// Bilinear column filtering. SSSE3 version.
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0, temp_pixel = 0;
+  asm volatile (
+    "movd      %6,%%xmm2                       \n"
+    "movd      %7,%%xmm3                       \n"
+    "movl      $0x04040000,%k2                 \n"
+    "movd      %k2,%%xmm5                      \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x9,%%xmm6                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "subl      $0x2,%5                         \n"
+    "jl        29f                             \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "punpckldq %%xmm0,%%xmm2                   \n"
+    "punpckldq %%xmm3,%%xmm3                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
+    "movd      %k2,%%xmm0                      \n"
+    "psrlw     $0x9,%%xmm1                     \n"
+    MEMOPARG(movzwl,0x00,1,4,1,k2)             //  movzwl  (%1,%4,1),%k2
+    "movd      %k2,%%xmm4                      \n"
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "punpcklwd %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm1                   \n"
+    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0,%k2                      \n"
+    "mov       %w2," MEMACCESS(0) "            \n"
+    "lea       " MEMLEA(0x2,0) ",%0            \n"
+    "sub       $0x2,%5                         \n"
+    "jge       2b                              \n"
+
+    LABELALIGN
+  "29:                                         \n"
+    "addl      $0x1,%5                         \n"
+    "jl        99f                             \n"
+    MEMOPARG(movzwl,0x00,1,3,1,k2)             //  movzwl  (%1,%3,1),%k2
+    "movd      %k2,%%xmm0                      \n"
+    "psrlw     $0x9,%%xmm2                     \n"
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "pxor      %%xmm6,%%xmm2                   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0,%k2                      \n"
+    "mov       %b2," MEMACCESS(0) "            \n"
+  "99:                                         \n"
+  : "+r"(dst_ptr),     // %0
+    "+r"(src_ptr),     // %1
+    "+a"(temp_pixel),  // %2
+    "+r"(x0),          // %3
+    "+r"(x1),          // %4
+    "+rm"(dst_width)   // %5
+  : "rm"(x),           // %6
+    "rm"(dx)           // %7
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpcklbw %%xmm0,%%xmm0                   \n"
+    "punpckhbw %%xmm1,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub       $0x20,%2                         \n"
+    "jg        1b                              \n"
+
+  : "+r"(dst_ptr),     // %0
+    "+r"(src_ptr),     // %1
+    "+r"(dst_width)    // %2
+  :: "memory", "cc", "xmm0", "xmm1"
+  );
+}
+
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "shufps    $0xdd,%%xmm1,%%xmm0             \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(dst_width)  // %2
+  :: "memory", "cc", "xmm0", "xmm1"
+  );
+}
+
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),  // %0
+    "+r"(dst_argb),  // %1
+    "+r"(dst_width)  // %2
+  :: "memory", "cc", "xmm0", "xmm1"
+  );
+}
+
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(0) ",%%xmm0         \n"
+    "movdqu    " MEMACCESS2(0x10,0) ",%%xmm1   \n"
+    MEMOPREG(movdqu,0x00,0,3,1,xmm2)           //  movdqu   (%0,%3,1),%%xmm2
+    MEMOPREG(movdqu,0x10,0,3,1,xmm3)           //  movdqu   0x10(%0,%3,1),%%xmm3
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(1) "         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "sub       $0x4,%2                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),   // %0
+    "+r"(dst_argb),   // %1
+    "+r"(dst_width)   // %2
+  : "r"((intptr_t)(src_stride))   // %3
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: dst_argb 16 byte aligned.
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width) {
+  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
+  intptr_t src_stepx_x12 = 0;
+  asm volatile (
+    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
+    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
+    LABELALIGN
+  "1:                                          \n"
+    "movd      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movd,0x00,0,1,1,xmm1)             //  movd      (%0,%1,1),%%xmm1
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,0,1,2,xmm2)             //  movd      (%0,%1,2),%%xmm2
+    MEMOPREG(movd,0x00,0,4,1,xmm3)             //  movd      (%0,%4,1),%%xmm3
+    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
+    "punpckldq %%xmm3,%%xmm2                   \n"
+    "punpcklqdq %%xmm2,%%xmm0                  \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),      // %0
+    "+r"(src_stepx_x4),  // %1
+    "+r"(dst_argb),      // %2
+    "+r"(dst_width),     // %3
+    "+r"(src_stepx_x12)  // %4
+  :: "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
+  );
+}
+
+// Blends four 2x2 to 4x1.
+// Alignment requirement: dst_argb 16 byte aligned.
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride, int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  intptr_t src_stepx_x4 = (intptr_t)(src_stepx);
+  intptr_t src_stepx_x12 = 0;
+  intptr_t row1 = (intptr_t)(src_stride);
+  asm volatile (
+    "lea       " MEMLEA3(0x00,1,4) ",%1        \n"
+    "lea       " MEMLEA4(0x00,1,1,2) ",%4      \n"
+    "lea       " MEMLEA4(0x00,0,5,1) ",%5      \n"
+
+    LABELALIGN
+  "1:                                          \n"
+    "movq      " MEMACCESS(0) ",%%xmm0         \n"
+    MEMOPREG(movhps,0x00,0,1,1,xmm0)           //  movhps    (%0,%1,1),%%xmm0
+    MEMOPREG(movq,0x00,0,1,2,xmm1)             //  movq      (%0,%1,2),%%xmm1
+    MEMOPREG(movhps,0x00,0,4,1,xmm1)           //  movhps    (%0,%4,1),%%xmm1
+    "lea       " MEMLEA4(0x00,0,1,4) ",%0      \n"
+    "movq      " MEMACCESS(5) ",%%xmm2         \n"
+    MEMOPREG(movhps,0x00,5,1,1,xmm2)           //  movhps    (%5,%1,1),%%xmm2
+    MEMOPREG(movq,0x00,5,1,2,xmm3)             //  movq      (%5,%1,2),%%xmm3
+    MEMOPREG(movhps,0x00,5,4,1,xmm3)           //  movhps    (%5,%4,1),%%xmm3
+    "lea       " MEMLEA4(0x00,5,1,4) ",%5      \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "pavgb     %%xmm3,%%xmm1                   \n"
+    "movdqa    %%xmm0,%%xmm2                   \n"
+    "shufps    $0x88,%%xmm1,%%xmm0             \n"
+    "shufps    $0xdd,%%xmm1,%%xmm2             \n"
+    "pavgb     %%xmm2,%%xmm0                   \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%3                         \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb),       // %0
+    "+r"(src_stepx_x4),   // %1
+    "+r"(dst_argb),       // %2
+    "+rm"(dst_width),     // %3
+    "+r"(src_stepx_x12),  // %4
+    "+r"(row1)            // %5
+  :: "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3"
+  );
+}
+
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0;
+  asm volatile (
+    "movd      %5,%%xmm2                       \n"
+    "movd      %6,%%xmm3                       \n"
+    "pshufd    $0x0,%%xmm2,%%xmm2              \n"
+    "pshufd    $0x11,%%xmm3,%%xmm0             \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pshufd    $0x5,%%xmm3,%%xmm0              \n"
+    "paddd     %%xmm0,%%xmm2                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pshufd    $0x0,%%xmm3,%%xmm3              \n"
+    "pextrw    $0x1,%%xmm2,%k0                 \n"
+    "pextrw    $0x3,%%xmm2,%k1                 \n"
+    "cmp       $0x0,%4                         \n"
+    "jl        99f                             \n"
+    "sub       $0x4,%4                         \n"
+    "jl        49f                             \n"
+
+    LABELALIGN
+  "40:                                         \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
+    "pextrw    $0x5,%%xmm2,%k0                 \n"
+    "pextrw    $0x7,%%xmm2,%k1                 \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm1)             //  movd      (%3,%0,4),%%xmm1
+    MEMOPREG(movd,0x00,3,1,4,xmm4)             //  movd      (%3,%1,4),%%xmm4
+    "pextrw    $0x1,%%xmm2,%k0                 \n"
+    "pextrw    $0x3,%%xmm2,%k1                 \n"
+    "punpckldq %%xmm4,%%xmm1                   \n"
+    "punpcklqdq %%xmm1,%%xmm0                  \n"
+    "movdqu    %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x10,2) ",%2           \n"
+    "sub       $0x4,%4                         \n"
+    "jge       40b                             \n"
+
+  "49:                                         \n"
+    "test      $0x2,%4                         \n"
+    "je        29f                             \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    MEMOPREG(movd,0x00,3,1,4,xmm1)             //  movd      (%3,%1,4),%%xmm1
+    "pextrw    $0x5,%%xmm2,%k0                 \n"
+    "punpckldq %%xmm1,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(2) "         \n"
+    "lea       " MEMLEA(0x8,2) ",%2            \n"
+  "29:                                         \n"
+    "test      $0x1,%4                         \n"
+    "je        99f                             \n"
+    MEMOPREG(movd,0x00,3,0,4,xmm0)             //  movd      (%3,%0,4),%%xmm0
+    "movd      %%xmm0," MEMACCESS(2) "         \n"
+  "99:                                         \n"
+  : "+a"(x0),          // %0
+    "+d"(x1),          // %1
+    "+r"(dst_argb),    // %2
+    "+r"(src_argb),    // %3
+    "+r"(dst_width)    // %4
+  : "rm"(x),           // %5
+    "rm"(dx)           // %6
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"
+  );
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+// Alignment requirement: src_argb 16 byte aligned, dst_argb 16 byte aligned.
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  asm volatile (
+    LABELALIGN
+  "1:                                          \n"
+    "movdqu    " MEMACCESS(1) ",%%xmm0         \n"
+    "lea       " MEMLEA(0x10,1) ",%1           \n"
+    "movdqa    %%xmm0,%%xmm1                   \n"
+    "punpckldq %%xmm0,%%xmm0                   \n"
+    "punpckhdq %%xmm1,%%xmm1                   \n"
+    "movdqu    %%xmm0," MEMACCESS(0) "         \n"
+    "movdqu    %%xmm1," MEMACCESS2(0x10,0) "   \n"
+    "lea       " MEMLEA(0x20,0) ",%0           \n"
+    "sub       $0x8,%2                         \n"
+    "jg        1b                              \n"
+
+  : "+r"(dst_argb),    // %0
+    "+r"(src_argb),    // %1
+    "+r"(dst_width)    // %2
+  :: "memory", "cc", NACL_R14
+    "xmm0", "xmm1"
+  );
+}
+
+// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
+static uvec8 kShuffleColARGB = {
+  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
+  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+};
+
+// Shuffle table for duplicating 2 fractions into 8 bytes each
+static uvec8 kShuffleFractions = {
+  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+};
+
+// Bilinear row filtering combines 4x2 -> 4x1. SSSE3 version
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx) {
+  intptr_t x0 = 0, x1 = 0;
+  asm volatile (
+    "movdqa    %0,%%xmm4                       \n"
+    "movdqa    %1,%%xmm5                       \n"
+  :
+  : "m"(kShuffleColARGB),  // %0
+    "m"(kShuffleFractions)  // %1
+  );
+
+  asm volatile (
+    "movd      %5,%%xmm2                       \n"
+    "movd      %6,%%xmm3                       \n"
+    "pcmpeqb   %%xmm6,%%xmm6                   \n"
+    "psrlw     $0x9,%%xmm6                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "sub       $0x2,%2                         \n"
+    "jl        29f                             \n"
+    "movdqa    %%xmm2,%%xmm0                   \n"
+    "paddd     %%xmm3,%%xmm0                   \n"
+    "punpckldq %%xmm0,%%xmm2                   \n"
+    "punpckldq %%xmm3,%%xmm3                   \n"
+    "paddd     %%xmm3,%%xmm3                   \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+
+    LABELALIGN
+  "2:                                          \n"
+    "movdqa    %%xmm2,%%xmm1                   \n"
+    "paddd     %%xmm3,%%xmm2                   \n"
+    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
+    "psrlw     $0x9,%%xmm1                     \n"
+    MEMOPREG(movhps,0x00,1,4,4,xmm0)           //  movhps    (%1,%4,4),%%xmm0
+    "pshufb    %%xmm5,%%xmm1                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm1                   \n"
+    "pmaddubsw %%xmm1,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "pextrw    $0x1,%%xmm2,%k3                 \n"
+    "pextrw    $0x3,%%xmm2,%k4                 \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movq      %%xmm0," MEMACCESS(0) "         \n"
+    "lea       " MEMLEA(0x8,0) ",%0            \n"
+    "sub       $0x2,%2                         \n"
+    "jge       2b                              \n"
+
+    LABELALIGN
+  "29:                                         \n"
+    "add       $0x1,%2                         \n"
+    "jl        99f                             \n"
+    "psrlw     $0x9,%%xmm2                     \n"
+    MEMOPREG(movq,0x00,1,3,4,xmm0)             //  movq      (%1,%3,4),%%xmm0
+    "pshufb    %%xmm5,%%xmm2                   \n"
+    "pshufb    %%xmm4,%%xmm0                   \n"
+    "pxor      %%xmm6,%%xmm2                   \n"
+    "pmaddubsw %%xmm2,%%xmm0                   \n"
+    "psrlw     $0x7,%%xmm0                     \n"
+    "packuswb  %%xmm0,%%xmm0                   \n"
+    "movd      %%xmm0," MEMACCESS(0) "         \n"
+
+    LABELALIGN
+  "99:                                         \n"
+  : "+r"(dst_argb),    // %0
+    "+r"(src_argb),    // %1
+    "+rm"(dst_width),  // %2
+    "+r"(x0),          // %3
+    "+r"(x1)           // %4
+  : "rm"(x),           // %5
+    "rm"(dx)           // %6
+  : "memory", "cc", NACL_R14
+    "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"
+  );
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+int FixedDiv_X86(int num, int div) {
+  asm volatile (
+    "cdq                                       \n"
+    "shld      $0x10,%%eax,%%edx               \n"
+    "shl       $0x10,%%eax                     \n"
+    "idiv      %1                              \n"
+    "mov       %0, %%eax                       \n"
+    : "+a"(num)  // %0
+    : "c"(div)   // %1
+    : "memory", "cc", "edx"
+  );
+  return num;
+}
+
+// Divide num - 1 by div - 1 and return as 16.16 fixed point result.
+int FixedDiv1_X86(int num, int div) {
+  asm volatile (
+    "cdq                                       \n"
+    "shld      $0x10,%%eax,%%edx               \n"
+    "shl       $0x10,%%eax                     \n"
+    "sub       $0x10001,%%eax                  \n"
+    "sbb       $0x0,%%edx                      \n"
+    "sub       $0x1,%1                         \n"
+    "idiv      %1                              \n"
+    "mov       %0, %%eax                       \n"
+    : "+a"(num)  // %0
+    : "c"(div)   // %1
+    : "memory", "cc", "edx"
+  );
+  return num;
+}
+
+#endif  // defined(__x86_64__) || defined(__i386__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/scale_mips.cc b/third_party/aom/third_party/libyuv/source/scale_mips.cc
new file mode 100644
index 0000000000..3eb4f27c45
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/scale_mips.cc
@@ -0,0 +1,654 @@
+/*
+ *  Copyright 2012 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/basic_types.h"
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC MIPS DSPR2
+#if !defined(LIBYUV_DISABLE_MIPS) && \
+    defined(__mips_dsp) && (__mips_dsp_rev >= 2) && \
+    (_MIPS_SIM == _MIPS_SIM_ABI32)
+
+void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width) {
+  __asm__ __volatile__(
+    ".set push                                     \n"
+    ".set noreorder                                \n"
+
+    "srl            $t9, %[dst_width], 4           \n"  // iterations -> by 16
+    "beqz           $t9, 2f                        \n"
+    " nop                                          \n"
+
+    ".p2align       2                              \n"
+  "1:                                              \n"
+    "lw             $t0, 0(%[src_ptr])             \n"  // |3|2|1|0|
+    "lw             $t1, 4(%[src_ptr])             \n"  // |7|6|5|4|
+    "lw             $t2, 8(%[src_ptr])             \n"  // |11|10|9|8|
+    "lw             $t3, 12(%[src_ptr])            \n"  // |15|14|13|12|
+    "lw             $t4, 16(%[src_ptr])            \n"  // |19|18|17|16|
+    "lw             $t5, 20(%[src_ptr])            \n"  // |23|22|21|20|
+    "lw             $t6, 24(%[src_ptr])            \n"  // |27|26|25|24|
+    "lw             $t7, 28(%[src_ptr])            \n"  // |31|30|29|28|
+    // TODO(fbarchard): Use odd pixels instead of even.
+    "precr.qb.ph    $t8, $t1, $t0                  \n"  // |6|4|2|0|
+    "precr.qb.ph    $t0, $t3, $t2                  \n"  // |14|12|10|8|
+    "precr.qb.ph    $t1, $t5, $t4                  \n"  // |22|20|18|16|
+    "precr.qb.ph    $t2, $t7, $t6                  \n"  // |30|28|26|24|
+    "addiu          %[src_ptr], %[src_ptr], 32     \n"
+    "addiu          $t9, $t9, -1                   \n"
+    "sw             $t8, 0(%[dst])                 \n"
+    "sw             $t0, 4(%[dst])                 \n"
+    "sw             $t1, 8(%[dst])                 \n"
+    "sw             $t2, 12(%[dst])                \n"
+    "bgtz           $t9, 1b                        \n"
+    " addiu         %[dst], %[dst], 16             \n"
+
+  "2:                                              \n"
+    "andi           $t9, %[dst_width], 0xf         \n"  // residue
+    "beqz           $t9, 3f                        \n"
+    " nop                                          \n"
+
+  "21:                                             \n"
+    "lbu            $t0, 0(%[src_ptr])             \n"
+    "addiu          %[src_ptr], %[src_ptr], 2      \n"
+    "addiu          $t9, $t9, -1                   \n"
+    "sb             $t0, 0(%[dst])                 \n"
+    "bgtz           $t9, 21b                       \n"
+    " addiu         %[dst], %[dst], 1              \n"
+
+  "3:                                              \n"
+    ".set pop                                      \n"
+  : [src_ptr] "+r" (src_ptr),
+    [dst] "+r" (dst)
+  : [dst_width] "r" (dst_width)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width) {
+  const uint8* t = src_ptr + src_stride;
+
+  __asm__ __volatile__ (
+    ".set push                                    \n"
+    ".set noreorder                               \n"
+
+    "srl            $t9, %[dst_width], 3          \n"  // iterations -> step 8
+    "bltz           $t9, 2f                       \n"
+    " nop                                         \n"
+
+    ".p2align       2                             \n"
+  "1:                                             \n"
+    "lw             $t0, 0(%[src_ptr])            \n"  // |3|2|1|0|
+    "lw             $t1, 4(%[src_ptr])            \n"  // |7|6|5|4|
+    "lw             $t2, 8(%[src_ptr])            \n"  // |11|10|9|8|
+    "lw             $t3, 12(%[src_ptr])           \n"  // |15|14|13|12|
+    "lw             $t4, 0(%[t])                  \n"  // |19|18|17|16|
+    "lw             $t5, 4(%[t])                  \n"  // |23|22|21|20|
+    "lw             $t6, 8(%[t])                  \n"  // |27|26|25|24|
+    "lw             $t7, 12(%[t])                 \n"  // |31|30|29|28|
+    "addiu          $t9, $t9, -1                  \n"
+    "srl            $t8, $t0, 16                  \n"  // |X|X|3|2|
+    "ins            $t0, $t4, 16, 16              \n"  // |17|16|1|0|
+    "ins            $t4, $t8, 0, 16               \n"  // |19|18|3|2|
+    "raddu.w.qb     $t0, $t0                      \n"  // |17+16+1+0|
+    "raddu.w.qb     $t4, $t4                      \n"  // |19+18+3+2|
+    "shra_r.w       $t0, $t0, 2                   \n"  // |t0+2|>>2
+    "shra_r.w       $t4, $t4, 2                   \n"  // |t4+2|>>2
+    "srl            $t8, $t1, 16                  \n"  // |X|X|7|6|
+    "ins            $t1, $t5, 16, 16              \n"  // |21|20|5|4|
+    "ins            $t5, $t8, 0, 16               \n"  // |22|23|7|6|
+    "raddu.w.qb     $t1, $t1                      \n"  // |21+20+5+4|
+    "raddu.w.qb     $t5, $t5                      \n"  // |23+22+7+6|
+    "shra_r.w       $t1, $t1, 2                   \n"  // |t1+2|>>2
+    "shra_r.w       $t5, $t5, 2                   \n"  // |t5+2|>>2
+    "srl            $t8, $t2, 16                  \n"  // |X|X|11|10|
+    "ins            $t2, $t6, 16, 16              \n"  // |25|24|9|8|
+    "ins            $t6, $t8, 0, 16               \n"  // |27|26|11|10|
+    "raddu.w.qb     $t2, $t2                      \n"  // |25+24+9+8|
+    "raddu.w.qb     $t6, $t6                      \n"  // |27+26+11+10|
+    "shra_r.w       $t2, $t2, 2                   \n"  // |t2+2|>>2
+    "shra_r.w       $t6, $t6, 2                   \n"  // |t5+2|>>2
+    "srl            $t8, $t3, 16                  \n"  // |X|X|15|14|
+    "ins            $t3, $t7, 16, 16              \n"  // |29|28|13|12|
+    "ins            $t7, $t8, 0, 16               \n"  // |31|30|15|14|
+    "raddu.w.qb     $t3, $t3                      \n"  // |29+28+13+12|
+    "raddu.w.qb     $t7, $t7                      \n"  // |31+30+15+14|
+    "shra_r.w       $t3, $t3, 2                   \n"  // |t3+2|>>2
+    "shra_r.w       $t7, $t7, 2                   \n"  // |t7+2|>>2
+    "addiu          %[src_ptr], %[src_ptr], 16    \n"
+    "addiu          %[t], %[t], 16                \n"
+    "sb             $t0, 0(%[dst])                \n"
+    "sb             $t4, 1(%[dst])                \n"
+    "sb             $t1, 2(%[dst])                \n"
+    "sb             $t5, 3(%[dst])                \n"
+    "sb             $t2, 4(%[dst])                \n"
+    "sb             $t6, 5(%[dst])                \n"
+    "sb             $t3, 6(%[dst])                \n"
+    "sb             $t7, 7(%[dst])                \n"
+    "bgtz           $t9, 1b                       \n"
+    " addiu         %[dst], %[dst], 8             \n"
+
+  "2:                                             \n"
+    "andi           $t9, %[dst_width], 0x7        \n"  // x = residue
+    "beqz           $t9, 3f                       \n"
+    " nop                                         \n"
+
+    "21:                                          \n"
+    "lwr            $t1, 0(%[src_ptr])            \n"
+    "lwl            $t1, 3(%[src_ptr])            \n"
+    "lwr            $t2, 0(%[t])                  \n"
+    "lwl            $t2, 3(%[t])                  \n"
+    "srl            $t8, $t1, 16                  \n"
+    "ins            $t1, $t2, 16, 16              \n"
+    "ins            $t2, $t8, 0, 16               \n"
+    "raddu.w.qb     $t1, $t1                      \n"
+    "raddu.w.qb     $t2, $t2                      \n"
+    "shra_r.w       $t1, $t1, 2                   \n"
+    "shra_r.w       $t2, $t2, 2                   \n"
+    "sb             $t1, 0(%[dst])                \n"
+    "sb             $t2, 1(%[dst])                \n"
+    "addiu          %[src_ptr], %[src_ptr], 4     \n"
+    "addiu          $t9, $t9, -2                  \n"
+    "addiu          %[t], %[t], 4                 \n"
+    "bgtz           $t9, 21b                      \n"
+    " addiu         %[dst], %[dst], 2             \n"
+
+  "3:                                             \n"
+    ".set pop                                     \n"
+
+  : [src_ptr] "+r" (src_ptr),
+    [dst] "+r" (dst), [t] "+r" (t)
+  : [dst_width] "r" (dst_width)
+  : "t0", "t1", "t2", "t3", "t4", "t5",
+    "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown4_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                    \n"
+      ".set noreorder                               \n"
+
+      "srl            $t9, %[dst_width], 3          \n"
+      "beqz           $t9, 2f                       \n"
+      " nop                                         \n"
+
+      ".p2align       2                             \n"
+     "1:                                            \n"
+      "lw             $t1, 0(%[src_ptr])            \n"  // |3|2|1|0|
+      "lw             $t2, 4(%[src_ptr])            \n"  // |7|6|5|4|
+      "lw             $t3, 8(%[src_ptr])            \n"  // |11|10|9|8|
+      "lw             $t4, 12(%[src_ptr])           \n"  // |15|14|13|12|
+      "lw             $t5, 16(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw             $t6, 20(%[src_ptr])           \n"  // |23|22|21|20|
+      "lw             $t7, 24(%[src_ptr])           \n"  // |27|26|25|24|
+      "lw             $t8, 28(%[src_ptr])           \n"  // |31|30|29|28|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |6|4|2|0|
+      "precr.qb.ph    $t2, $t4, $t3                 \n"  // |14|12|10|8|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |22|20|18|16|
+      "precr.qb.ph    $t6, $t8, $t7                 \n"  // |30|28|26|24|
+      "precr.qb.ph    $t1, $t2, $t1                 \n"  // |12|8|4|0|
+      "precr.qb.ph    $t5, $t6, $t5                 \n"  // |28|24|20|16|
+      "addiu          %[src_ptr], %[src_ptr], 32    \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sw             $t1, 0(%[dst])                \n"
+      "sw             $t5, 4(%[dst])                \n"
+      "bgtz           $t9, 1b                       \n"
+      " addiu         %[dst], %[dst], 8             \n"
+
+    "2:                                             \n"
+      "andi           $t9, %[dst_width], 7          \n"  // residue
+      "beqz           $t9, 3f                       \n"
+      " nop                                         \n"
+
+    "21:                                            \n"
+      "lbu            $t1, 0(%[src_ptr])            \n"
+      "addiu          %[src_ptr], %[src_ptr], 4     \n"
+      "addiu          $t9, $t9, -1                  \n"
+      "sb             $t1, 0(%[dst])                \n"
+      "bgtz           $t9, 21b                      \n"
+      " addiu         %[dst], %[dst], 1             \n"
+
+    "3:                                             \n"
+      ".set pop                                     \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst)
+      : [dst_width] "r" (dst_width)
+      : "t1", "t2", "t3", "t4", "t5",
+        "t6", "t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown4Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                 uint8* dst, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  const uint8* s2 = s1 + stride;
+  const uint8* s3 = s2 + stride;
+
+  __asm__ __volatile__ (
+      ".set push                                  \n"
+      ".set noreorder                             \n"
+
+      "srl           $t9, %[dst_width], 1         \n"
+      "andi          $t8, %[dst_width], 1         \n"
+
+      ".p2align      2                            \n"
+     "1:                                          \n"
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "lw            $t4, 4(%[src_ptr])           \n"  // |19|18|17|16|
+      "lw            $t5, 4(%[s1])                \n"  // |23|22|21|20|
+      "lw            $t6, 4(%[s2])                \n"  // |27|26|25|24|
+      "lw            $t7, 4(%[s3])                \n"  // |31|30|29|28|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "raddu.w.qb    $t4, $t4                     \n"  // |19 + 18 + 17 + 16|
+      "raddu.w.qb    $t5, $t5                     \n"  // |23 + 22 + 21 + 20|
+      "raddu.w.qb    $t6, $t6                     \n"  // |27 + 26 + 25 + 24|
+      "raddu.w.qb    $t7, $t7                     \n"  // |31 + 30 + 29 + 28|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "add           $t4, $t4, $t5                \n"
+      "add           $t6, $t6, $t7                \n"
+      "add           $t4, $t4, $t6                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "shra_r.w      $t4, $t4, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+      "sb            $t4, 1(%[dst])               \n"
+      "addiu         %[src_ptr], %[src_ptr], 8    \n"
+      "addiu         %[s1], %[s1], 8              \n"
+      "addiu         %[s2], %[s2], 8              \n"
+      "addiu         %[s3], %[s3], 8              \n"
+      "addiu         $t9, $t9, -1                 \n"
+      "bgtz          $t9, 1b                      \n"
+      " addiu        %[dst], %[dst], 2            \n"
+      "beqz          $t8, 2f                      \n"
+      " nop                                       \n"
+
+      "lw            $t0, 0(%[src_ptr])           \n"  // |3|2|1|0|
+      "lw            $t1, 0(%[s1])                \n"  // |7|6|5|4|
+      "lw            $t2, 0(%[s2])                \n"  // |11|10|9|8|
+      "lw            $t3, 0(%[s3])                \n"  // |15|14|13|12|
+      "raddu.w.qb    $t0, $t0                     \n"  // |3 + 2 + 1 + 0|
+      "raddu.w.qb    $t1, $t1                     \n"  // |7 + 6 + 5 + 4|
+      "raddu.w.qb    $t2, $t2                     \n"  // |11 + 10 + 9 + 8|
+      "raddu.w.qb    $t3, $t3                     \n"  // |15 + 14 + 13 + 12|
+      "add           $t0, $t0, $t1                \n"
+      "add           $t1, $t2, $t3                \n"
+      "add           $t0, $t0, $t1                \n"
+      "shra_r.w      $t0, $t0, 4                  \n"
+      "sb            $t0, 0(%[dst])               \n"
+
+      "2:                                         \n"
+      ".set pop                                   \n"
+
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [s1] "+r" (s1),
+        [s2] "+r" (s2),
+        [s3] "+r" (s3)
+      : [dst_width] "r" (dst_width)
+      : "t0", "t1", "t2", "t3", "t4", "t5",
+        "t6","t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown34_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                          \n"
+      ".set noreorder                                     \n"
+      ".p2align        2                                  \n"
+    "1:                                                   \n"
+      "lw              $t1, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw              $t2, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw              $t3, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw              $t4, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw              $t5, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw              $t6, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw              $t7, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw              $t8, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "precrq.qb.ph    $t0, $t2, $t4                      \n"  // |7|5|15|13|
+      "precrq.qb.ph    $t9, $t6, $t8                      \n"  // |23|21|31|30|
+      "addiu           %[dst_width], %[dst_width], -24    \n"
+      "ins             $t1, $t1, 8, 16                    \n"  // |3|1|0|X|
+      "ins             $t4, $t0, 8, 16                    \n"  // |X|15|13|12|
+      "ins             $t5, $t5, 8, 16                    \n"  // |19|17|16|X|
+      "ins             $t8, $t9, 8, 16                    \n"  // |X|31|29|28|
+      "addiu           %[src_ptr], %[src_ptr], 32         \n"
+      "packrl.ph       $t0, $t3, $t0                      \n"  // |9|8|7|5|
+      "packrl.ph       $t9, $t7, $t9                      \n"  // |25|24|23|21|
+      "prepend         $t1, $t2, 8                        \n"  // |4|3|1|0|
+      "prepend         $t3, $t4, 24                       \n"  // |15|13|12|11|
+      "prepend         $t5, $t6, 8                        \n"  // |20|19|17|16|
+      "prepend         $t7, $t8, 24                       \n"  // |31|29|28|27|
+      "sw              $t1, 0(%[dst])                     \n"
+      "sw              $t0, 4(%[dst])                     \n"
+      "sw              $t3, 8(%[dst])                     \n"
+      "sw              $t5, 12(%[dst])                    \n"
+      "sw              $t9, 16(%[dst])                    \n"
+      "sw              $t7, 20(%[dst])                    \n"
+      "bnez            %[dst_width], 1b                   \n"
+      " addiu          %[dst], %[dst], 24                 \n"
+      ".set pop                                           \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4", "t5",
+        "t6","t7", "t8", "t9"
+  );
+}
+
+void ScaleRowDown34_0_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+      "repl.ph           $t3, 3                          \n"  // 0x00030003
+
+     ".p2align           2                               \n"
+    "1:                                                  \n"
+      "lw                $t0, 0(%[src_ptr])              \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])  \n"  // |T3|T2|T1|T0|
+      "rotr              $t2, $t0, 8                     \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                     \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t4, $t2, $t3                   \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t3                   \n"  // |T0*3|T3*3|
+      "andi              $t0, $t2, 0xFFFF                \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                        \n"
+      "raddu.w.qb        $t1, $t1                        \n"
+      "shra_r.w          $t0, $t0, 1                     \n"
+      "shra_r.w          $t1, $t1, 1                     \n"
+      "preceu.ph.qbr     $t2, $t2                        \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                        \n"  // |0|T2|0|T1|
+      "rotr              $t2, $t2, 16                    \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                    \n"  // |0|T1|0|T2|
+      "addu.ph           $t2, $t2, $t4                   \n"
+      "addu.ph           $t6, $t6, $t5                   \n"
+      "sll               $t5, $t0, 1                     \n"
+      "add               $t0, $t5, $t0                   \n"
+      "shra_r.ph         $t2, $t2, 2                     \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "shll.ph           $t4, $t2, 1                     \n"
+      "addq.ph           $t4, $t4, $t2                   \n"
+      "addu              $t0, $t0, $t1                   \n"
+      "addiu             %[src_ptr], %[src_ptr], 4       \n"
+      "shra_r.w          $t0, $t0, 2                     \n"
+      "addu.ph           $t6, $t6, $t4                   \n"
+      "shra_r.ph         $t6, $t6, 2                     \n"
+      "srl               $t1, $t6, 16                    \n"
+      "addiu             %[dst_width], %[dst_width], -3  \n"
+      "sb                $t1, 0(%[d])                    \n"
+      "sb                $t0, 1(%[d])                    \n"
+      "sb                $t6, 2(%[d])                    \n"
+      "bgtz              %[dst_width], 1b                \n"
+      " addiu            %[d], %[d], 3                   \n"
+    "3:                                                  \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [src_stride] "+r" (src_stride),
+        [d] "+r" (d),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3",
+        "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown34_1_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* d, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                           \n"
+      ".set noreorder                                      \n"
+      "repl.ph           $t2, 3                            \n"  // 0x00030003
+
+      ".p2align          2                                 \n"
+    "1:                                                    \n"
+      "lw                $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lwx               $t1, %[src_stride](%[src_ptr])    \n"  // |T3|T2|T1|T0|
+      "rotr              $t4, $t0, 8                       \n"  // |S0|S3|S2|S1|
+      "rotr              $t6, $t1, 8                       \n"  // |T0|T3|T2|T1|
+      "muleu_s.ph.qbl    $t3, $t4, $t2                     \n"  // |S0*3|S3*3|
+      "muleu_s.ph.qbl    $t5, $t6, $t2                     \n"  // |T0*3|T3*3|
+      "andi              $t0, $t4, 0xFFFF                  \n"  // |0|0|S2|S1|
+      "andi              $t1, $t6, 0xFFFF                  \n"  // |0|0|T2|T1|
+      "raddu.w.qb        $t0, $t0                          \n"
+      "raddu.w.qb        $t1, $t1                          \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "shra_r.w          $t1, $t1, 1                       \n"
+      "preceu.ph.qbr     $t4, $t4                          \n"  // |0|S2|0|S1|
+      "preceu.ph.qbr     $t6, $t6                          \n"  // |0|T2|0|T1|
+      "rotr              $t4, $t4, 16                      \n"  // |0|S1|0|S2|
+      "rotr              $t6, $t6, 16                      \n"  // |0|T1|0|T2|
+      "addu.ph           $t4, $t4, $t3                     \n"
+      "addu.ph           $t6, $t6, $t5                     \n"
+      "shra_r.ph         $t6, $t6, 2                       \n"
+      "shra_r.ph         $t4, $t4, 2                       \n"
+      "addu.ph           $t6, $t6, $t4                     \n"
+      "addiu             %[src_ptr], %[src_ptr], 4         \n"
+      "shra_r.ph         $t6, $t6, 1                       \n"
+      "addu              $t0, $t0, $t1                     \n"
+      "addiu             %[dst_width], %[dst_width], -3    \n"
+      "shra_r.w          $t0, $t0, 1                       \n"
+      "srl               $t1, $t6, 16                      \n"
+      "sb                $t1, 0(%[d])                      \n"
+      "sb                $t0, 1(%[d])                      \n"
+      "sb                $t6, 2(%[d])                      \n"
+      "bgtz              %[dst_width], 1b                  \n"
+      " addiu            %[d], %[d], 3                     \n"
+    "3:                                                    \n"
+      ".set pop                                            \n"
+      : [src_ptr] "+r" (src_ptr),
+        [src_stride] "+r" (src_stride),
+        [d] "+r" (d),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3",
+        "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown38_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  __asm__ __volatile__ (
+      ".set push                                     \n"
+      ".set noreorder                                \n"
+
+      ".p2align   2                                  \n"
+    "1:                                              \n"
+      "lw         $t0, 0(%[src_ptr])                 \n"  // |3|2|1|0|
+      "lw         $t1, 4(%[src_ptr])                 \n"  // |7|6|5|4|
+      "lw         $t2, 8(%[src_ptr])                 \n"  // |11|10|9|8|
+      "lw         $t3, 12(%[src_ptr])                \n"  // |15|14|13|12|
+      "lw         $t4, 16(%[src_ptr])                \n"  // |19|18|17|16|
+      "lw         $t5, 20(%[src_ptr])                \n"  // |23|22|21|20|
+      "lw         $t6, 24(%[src_ptr])                \n"  // |27|26|25|24|
+      "lw         $t7, 28(%[src_ptr])                \n"  // |31|30|29|28|
+      "wsbh       $t0, $t0                           \n"  // |2|3|0|1|
+      "wsbh       $t6, $t6                           \n"  // |26|27|24|25|
+      "srl        $t0, $t0, 8                        \n"  // |X|2|3|0|
+      "srl        $t3, $t3, 16                       \n"  // |X|X|15|14|
+      "srl        $t5, $t5, 16                       \n"  // |X|X|23|22|
+      "srl        $t7, $t7, 16                       \n"  // |X|X|31|30|
+      "ins        $t1, $t2, 24, 8                    \n"  // |8|6|5|4|
+      "ins        $t6, $t5, 0, 8                     \n"  // |26|27|24|22|
+      "ins        $t1, $t0, 0, 16                    \n"  // |8|6|3|0|
+      "ins        $t6, $t7, 24, 8                    \n"  // |30|27|24|22|
+      "prepend    $t2, $t3, 24                       \n"  // |X|15|14|11|
+      "ins        $t4, $t4, 16, 8                    \n"  // |19|16|17|X|
+      "ins        $t4, $t2, 0, 16                    \n"  // |19|16|14|11|
+      "addiu      %[src_ptr], %[src_ptr], 32         \n"
+      "addiu      %[dst_width], %[dst_width], -12    \n"
+      "addiu      $t8,%[dst_width], -12              \n"
+      "sw         $t1, 0(%[dst])                     \n"
+      "sw         $t4, 4(%[dst])                     \n"
+      "sw         $t6, 8(%[dst])                     \n"
+      "bgez       $t8, 1b                            \n"
+      " addiu     %[dst], %[dst], 12                 \n"
+      ".set pop                                      \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst] "+r" (dst),
+        [dst_width] "+r" (dst_width)
+      :
+      : "t0", "t1", "t2", "t3", "t4",
+        "t5", "t6", "t7", "t8"
+  );
+}
+
+void ScaleRowDown38_2_Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* t = src_ptr + stride;
+  const int c = 0x2AAA;
+
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      ".p2align        2                                 \n"
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[t])                      \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[t])                      \n"  // |T7|T6|T5|T4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t4, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "packrl.ph       $t5, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t4, $t4                          \n"  // S7+S6+T7+T6
+      "raddu.w.qb      $t5, $t5                          \n"  // T5+T4+S5+S4
+      "precrq.qb.ph    $t6, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t6, $t6, $t6                     \n"  // |S3|T3|S3|T3|
+      "srl             $t4, $t4, 2                       \n"  // t4 / 4
+      "srl             $t6, $t6, 16                      \n"  // |0|0|S3|T3|
+      "raddu.w.qb      $t6, $t6                          \n"  // 0+0+S3+T3
+      "addu            $t6, $t5, $t6                     \n"
+      "mul             $t6, $t6, %[c]                    \n"  // t6 * 0x2AAA
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "raddu.w.qb      $t0, $t0                          \n"  // S2+S1+S0+0
+      "raddu.w.qb      $t2, $t2                          \n"  // T2+T1+T0+0
+      "addu            $t0, $t0, $t2                     \n"
+      "mul             $t0, $t0, %[c]                    \n"  // t0 * 0x2AAA
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[t], %[t], 8                     \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t4, -1(%[dst_ptr])               \n"
+      "sb              $t6, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst_ptr] "+r" (dst_ptr),
+        [t] "+r" (t),
+        [dst_width] "+r" (dst_width)
+      : [c] "r" (c)
+      : "t0", "t1", "t2", "t3", "t4", "t5", "t6"
+  );
+}
+
+void ScaleRowDown38_3_Box_MIPS_DSPR2(const uint8* src_ptr,
+                                     ptrdiff_t src_stride,
+                                     uint8* dst_ptr, int dst_width) {
+  intptr_t stride = src_stride;
+  const uint8* s1 = src_ptr + stride;
+  stride += stride;
+  const uint8* s2 = src_ptr + stride;
+  const int c1 = 0x1C71;
+  const int c2 = 0x2AAA;
+
+  __asm__ __volatile__ (
+      ".set push                                         \n"
+      ".set noreorder                                    \n"
+
+      ".p2align        2                                 \n"
+    "1:                                                  \n"
+      "lw              $t0, 0(%[src_ptr])                \n"  // |S3|S2|S1|S0|
+      "lw              $t1, 4(%[src_ptr])                \n"  // |S7|S6|S5|S4|
+      "lw              $t2, 0(%[s1])                     \n"  // |T3|T2|T1|T0|
+      "lw              $t3, 4(%[s1])                     \n"  // |T7|T6|T5|T4|
+      "lw              $t4, 0(%[s2])                     \n"  // |R3|R2|R1|R0|
+      "lw              $t5, 4(%[s2])                     \n"  // |R7|R6|R5|R4|
+      "rotr            $t1, $t1, 16                      \n"  // |S5|S4|S7|S6|
+      "packrl.ph       $t6, $t1, $t3                     \n"  // |S7|S6|T7|T6|
+      "raddu.w.qb      $t6, $t6                          \n"  // S7+S6+T7+T6
+      "packrl.ph       $t7, $t3, $t1                     \n"  // |T5|T4|S5|S4|
+      "raddu.w.qb      $t7, $t7                          \n"  // T5+T4+S5+S4
+      "sll             $t8, $t5, 16                      \n"  // |R5|R4|0|0|
+      "raddu.w.qb      $t8, $t8                          \n"  // R5+R4
+      "addu            $t7, $t7, $t8                     \n"
+      "srl             $t8, $t5, 16                      \n"  // |0|0|R7|R6|
+      "raddu.w.qb      $t8, $t8                          \n"  // R7 + R6
+      "addu            $t6, $t6, $t8                     \n"
+      "mul             $t6, $t6, %[c2]                   \n"  // t6 * 0x2AAA
+      "precrq.qb.ph    $t8, $t0, $t2                     \n"  // |S3|S1|T3|T1|
+      "precrq.qb.ph    $t8, $t8, $t4                     \n"  // |S3|T3|R3|R1|
+      "srl             $t8, $t8, 8                       \n"  // |0|S3|T3|R3|
+      "raddu.w.qb      $t8, $t8                          \n"  // S3 + T3 + R3
+      "addu            $t7, $t7, $t8                     \n"
+      "mul             $t7, $t7, %[c1]                   \n"  // t7 * 0x1C71
+      "sll             $t0, $t0, 8                       \n"  // |S2|S1|S0|0|
+      "sll             $t2, $t2, 8                       \n"  // |T2|T1|T0|0|
+      "sll             $t4, $t4, 8                       \n"  // |R2|R1|R0|0|
+      "raddu.w.qb      $t0, $t0                          \n"
+      "raddu.w.qb      $t2, $t2                          \n"
+      "raddu.w.qb      $t4, $t4                          \n"
+      "addu            $t0, $t0, $t2                     \n"
+      "addu            $t0, $t0, $t4                     \n"
+      "mul             $t0, $t0, %[c1]                   \n"  // t0 * 0x1C71
+      "addiu           %[src_ptr], %[src_ptr], 8         \n"
+      "addiu           %[s1], %[s1], 8                   \n"
+      "addiu           %[s2], %[s2], 8                   \n"
+      "addiu           %[dst_width], %[dst_width], -3    \n"
+      "addiu           %[dst_ptr], %[dst_ptr], 3         \n"
+      "srl             $t6, $t6, 16                      \n"
+      "srl             $t7, $t7, 16                      \n"
+      "srl             $t0, $t0, 16                      \n"
+      "sb              $t6, -1(%[dst_ptr])               \n"
+      "sb              $t7, -2(%[dst_ptr])               \n"
+      "bgtz            %[dst_width], 1b                  \n"
+      " sb             $t0, -3(%[dst_ptr])               \n"
+      ".set pop                                          \n"
+      : [src_ptr] "+r" (src_ptr),
+        [dst_ptr] "+r" (dst_ptr),
+        [s1] "+r" (s1),
+        [s2] "+r" (s2),
+        [dst_width] "+r" (dst_width)
+      : [c1] "r" (c1), [c2] "r" (c2)
+      : "t0", "t1", "t2", "t3", "t4",
+        "t5", "t6", "t7", "t8"
+  );
+}
+
+#endif  // defined(__mips_dsp) && (__mips_dsp_rev >= 2)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/third_party/aom/third_party/libyuv/source/scale_neon.cc b/third_party/aom/third_party/libyuv/source/scale_neon.cc
new file mode 100644
index 0000000000..7825878e98
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/scale_neon.cc
@@ -0,0 +1,1037 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__ARM_NEON__) && \
+    !defined(__aarch64__)
+
+// NEON downscalers with interpolation.
+// Provided by Fritz Koenig
+
+// Read 32x1 throw away even pixels, and write 16x1.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    MEMACCESS(0)
+    "vld2.8     {q0, q1}, [%0]!                \n"
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1"              // Clobber List
+  );
+}
+
+// Read 32x1 average down and write 16x1.
+void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load pixels and post inc
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop
+    "vpaddl.u8  q0, q0                         \n"  // add adjacent
+    "vpaddl.u8  q1, q1                         \n"
+    "vrshrn.u16 d0, q0, #1                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #1                     \n"
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1"     // Clobber List
+  );
+}
+
+// Read 32x2 average down and write 16x1.
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load row 1 and post inc
+    MEMACCESS(1)
+    "vld1.8     {q2, q3}, [%1]!                \n"  // load row 2 and post inc
+    "subs       %3, %3, #16                    \n"  // 16 processed per loop
+    "vpaddl.u8  q0, q0                         \n"  // row 1 add adjacent
+    "vpaddl.u8  q1, q1                         \n"
+    "vpadal.u8  q0, q2                         \n"  // row 2 add adjacent + row1
+    "vpadal.u8  q1, q3                         \n"
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "q0", "q1", "q2", "q3"     // Clobber List
+  );
+}
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n" // src line 0
+    "subs       %2, %2, #8                     \n" // 8 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {d2}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1", "memory", "cc"
+  );
+}
+
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  const uint8* src_ptr2 = src_ptr + src_stride * 2;
+  const uint8* src_ptr3 = src_ptr + src_stride * 3;
+asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0]!                    \n"   // load up 16x4
+    MEMACCESS(3)
+    "vld1.8     {q1}, [%3]!                    \n"
+    MEMACCESS(4)
+    "vld1.8     {q2}, [%4]!                    \n"
+    MEMACCESS(5)
+    "vld1.8     {q3}, [%5]!                    \n"
+    "subs       %2, %2, #4                     \n"
+    "vpaddl.u8  q0, q0                         \n"
+    "vpadal.u8  q0, q1                         \n"
+    "vpadal.u8  q0, q2                         \n"
+    "vpadal.u8  q0, q3                         \n"
+    "vpaddl.u16 q0, q0                         \n"
+    "vrshrn.u32 d0, q0, #4                     \n"   // divide by 16 w/rounding
+    "vmovn.u16  d0, q0                         \n"
+    MEMACCESS(1)
+    "vst1.32    {d0[0]}, [%1]!                 \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(dst_width), // %2
+    "+r"(src_ptr1),  // %3
+    "+r"(src_ptr2),  // %4
+    "+r"(src_ptr3)   // %5
+  :
+  : "q0", "q1", "q2", "q3", "memory", "cc"
+  );
+}
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+// to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    "subs       %2, %2, #24                  \n"
+    "vmov       d2, d3                       \n" // order d0, d1, d2
+    MEMACCESS(1)
+    "vst3.8     {d0, d1, d2}, [%1]!          \n"
+    "bgt        1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "d0", "d1", "d2", "d3", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vmov.u8    d24, #3                        \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
+    "subs         %2, %2, #24                  \n"
+
+    // filter src line 0 with src line 1
+    // expand chars to shorts to allow for room
+    // when adding lines together
+    "vmovl.u8     q8, d4                       \n"
+    "vmovl.u8     q9, d5                       \n"
+    "vmovl.u8     q10, d6                      \n"
+    "vmovl.u8     q11, d7                      \n"
+
+    // 3 * line_0 + line_1
+    "vmlal.u8     q8, d0, d24                  \n"
+    "vmlal.u8     q9, d1, d24                  \n"
+    "vmlal.u8     q10, d2, d24                 \n"
+    "vmlal.u8     q11, d3, d24                 \n"
+
+    // (3 * line_0 + line_1) >> 2
+    "vqrshrn.u16  d0, q8, #2                   \n"
+    "vqrshrn.u16  d1, q9, #2                   \n"
+    "vqrshrn.u16  d2, q10, #2                  \n"
+    "vqrshrn.u16  d3, q11, #2                  \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "vmovl.u8     q8, d1                       \n"
+    "vmlal.u8     q8, d0, d24                  \n"
+    "vqrshrn.u16  d0, q8, #2                   \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "vrhadd.u8    d1, d1, d2                   \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "vmovl.u8     q8, d2                       \n"
+    "vmlal.u8     q8, d3, d24                  \n"
+    "vqrshrn.u16  d2, q8, #2                   \n"
+
+    MEMACCESS(1)
+    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "d24", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "vmov.u8    d24, #3                        \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n" // src line 0
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n" // src line 1
+    "subs         %2, %2, #24                  \n"
+    // average src line 0 with src line 1
+    "vrhadd.u8    q0, q0, q2                   \n"
+    "vrhadd.u8    q1, q1, q3                   \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "vmovl.u8     q3, d1                       \n"
+    "vmlal.u8     q3, d0, d24                  \n"
+    "vqrshrn.u16  d0, q3, #2                   \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "vrhadd.u8    d1, d1, d2                   \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "vmovl.u8     q3, d2                       \n"
+    "vmlal.u8     q3, d3, d24                  \n"
+    "vqrshrn.u16  d2, q3, #2                   \n"
+
+    MEMACCESS(1)
+    "vst3.8       {d0, d1, d2}, [%1]!          \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "r4", "q0", "q1", "q2", "q3", "d24", "memory", "cc"
+  );
+}
+
+#define HAS_SCALEROWDOWN38_NEON
+static uvec8 kShuf38 =
+  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
+static uvec8 kShuf38_2 =
+  { 0, 8, 16, 2, 10, 17, 4, 12, 18, 6, 14, 19, 0, 0, 0, 0 };
+static vec16 kMult38_Div6 =
+  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
+    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
+static vec16 kMult38_Div9 =
+  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
+    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    MEMACCESS(3)
+    "vld1.8     {q3}, [%3]                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0, d1, d2, d3}, [%0]!        \n"
+    "subs       %2, %2, #12                    \n"
+    "vtbl.u8    d4, {d0, d1, d2, d3}, d6       \n"
+    "vtbl.u8    d5, {d0, d1, d2, d3}, d7       \n"
+    MEMACCESS(1)
+    "vst1.8     {d4}, [%1]!                    \n"
+    MEMACCESS(1)
+    "vst1.32    {d5[0]}, [%1]!                 \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  : "r"(&kShuf38)           // %3
+  : "d0", "d1", "d2", "d3", "d4", "d5", "memory", "cc"
+  );
+}
+
+// 32x3 -> 12x1
+void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_ptr, int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride * 2;
+
+  asm volatile (
+    MEMACCESS(5)
+    "vld1.16    {q13}, [%5]                    \n"
+    MEMACCESS(6)
+    "vld1.8     {q14}, [%6]                    \n"
+    MEMACCESS(7)
+    "vld1.8     {q15}, [%7]                    \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+
+    // d0 = 00 40 01 41 02 42 03 43
+    // d1 = 10 50 11 51 12 52 13 53
+    // d2 = 20 60 21 61 22 62 23 63
+    // d3 = 30 70 31 71 32 72 33 73
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    MEMACCESS(4)
+    "vld4.8       {d16, d17, d18, d19}, [%4]!  \n"
+    "subs         %2, %2, #12                  \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // d0 = 00 10 01 11 02 12 03 13
+    // d1 = 40 50 41 51 42 52 43 53
+    "vtrn.u8      d0, d1                       \n"
+    "vtrn.u8      d4, d5                       \n"
+    "vtrn.u8      d16, d17                     \n"
+
+    // d2 = 20 30 21 31 22 32 23 33
+    // d3 = 60 70 61 71 62 72 63 73
+    "vtrn.u8      d2, d3                       \n"
+    "vtrn.u8      d6, d7                       \n"
+    "vtrn.u8      d18, d19                     \n"
+
+    // d0 = 00+10 01+11 02+12 03+13
+    // d2 = 40+50 41+51 42+52 43+53
+    "vpaddl.u8    q0, q0                       \n"
+    "vpaddl.u8    q2, q2                       \n"
+    "vpaddl.u8    q8, q8                       \n"
+
+    // d3 = 60+70 61+71 62+72 63+73
+    "vpaddl.u8    d3, d3                       \n"
+    "vpaddl.u8    d7, d7                       \n"
+    "vpaddl.u8    d19, d19                     \n"
+
+    // combine source lines
+    "vadd.u16     q0, q2                       \n"
+    "vadd.u16     q0, q8                       \n"
+    "vadd.u16     d4, d3, d7                   \n"
+    "vadd.u16     d4, d19                      \n"
+
+    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
+    //             + s[6 + st * 1] + s[7 + st * 1]
+    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
+    "vqrdmulh.s16 q2, q2, q13                  \n"
+    "vmovn.u16    d4, q2                       \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "vmovl.u8     q1, d2                       \n"
+    "vmovl.u8     q3, d6                       \n"
+    "vmovl.u8     q9, d18                      \n"
+
+    // combine source lines
+    "vadd.u16     q1, q3                       \n"
+    "vadd.u16     q1, q9                       \n"
+
+    // d4 = xx 20 xx 30 xx 22 xx 32
+    // d5 = xx 21 xx 31 xx 23 xx 33
+    "vtrn.u32     d2, d3                       \n"
+
+    // d4 = xx 20 xx 21 xx 22 xx 23
+    // d5 = xx 30 xx 31 xx 32 xx 33
+    "vtrn.u16     d2, d3                       \n"
+
+    // 0+1+2, 3+4+5
+    "vadd.u16     q0, q1                       \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "vqrdmulh.s16 q0, q0, q15                  \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "vmov.u8      d2, d4                       \n"
+
+    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+    MEMACCESS(1)
+    "vst1.8       {d3}, [%1]!                  \n"
+    MEMACCESS(1)
+    "vst1.32      {d4[0]}, [%1]!               \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride),       // %3
+    "+r"(src_ptr1)          // %4
+  : "r"(&kMult38_Div6),     // %5
+    "r"(&kShuf38_2),        // %6
+    "r"(&kMult38_Div9)      // %7
+  : "q0", "q1", "q2", "q3", "q8", "q9", "q13", "q14", "q15", "memory", "cc"
+  );
+}
+
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    MEMACCESS(4)
+    "vld1.16    {q13}, [%4]                    \n"
+    MEMACCESS(5)
+    "vld1.8     {q14}, [%5]                    \n"
+    "add        %3, %0                         \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+
+    // d0 = 00 40 01 41 02 42 03 43
+    // d1 = 10 50 11 51 12 52 13 53
+    // d2 = 20 60 21 61 22 62 23 63
+    // d3 = 30 70 31 71 32 72 33 73
+    MEMACCESS(0)
+    "vld4.8       {d0, d1, d2, d3}, [%0]!      \n"
+    MEMACCESS(3)
+    "vld4.8       {d4, d5, d6, d7}, [%3]!      \n"
+    "subs         %2, %2, #12                  \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // d0 = 00 10 01 11 02 12 03 13
+    // d1 = 40 50 41 51 42 52 43 53
+    "vtrn.u8      d0, d1                       \n"
+    "vtrn.u8      d4, d5                       \n"
+
+    // d2 = 20 30 21 31 22 32 23 33
+    // d3 = 60 70 61 71 62 72 63 73
+    "vtrn.u8      d2, d3                       \n"
+    "vtrn.u8      d6, d7                       \n"
+
+    // d0 = 00+10 01+11 02+12 03+13
+    // d2 = 40+50 41+51 42+52 43+53
+    "vpaddl.u8    q0, q0                       \n"
+    "vpaddl.u8    q2, q2                       \n"
+
+    // d3 = 60+70 61+71 62+72 63+73
+    "vpaddl.u8    d3, d3                       \n"
+    "vpaddl.u8    d7, d7                       \n"
+
+    // combine source lines
+    "vadd.u16     q0, q2                       \n"
+    "vadd.u16     d4, d3, d7                   \n"
+
+    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
+    "vqrshrn.u16  d4, q2, #2                   \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // q2 = xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "vmovl.u8     q1, d2                       \n"
+    "vmovl.u8     q3, d6                       \n"
+
+    // combine source lines
+    "vadd.u16     q1, q3                       \n"
+
+    // d4 = xx 20 xx 30 xx 22 xx 32
+    // d5 = xx 21 xx 31 xx 23 xx 33
+    "vtrn.u32     d2, d3                       \n"
+
+    // d4 = xx 20 xx 21 xx 22 xx 23
+    // d5 = xx 30 xx 31 xx 32 xx 33
+    "vtrn.u16     d2, d3                       \n"
+
+    // 0+1+2, 3+4+5
+    "vadd.u16     q0, q1                       \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "vqrdmulh.s16 q0, q0, q13                  \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "vmov.u8      d2, d4                       \n"
+
+    "vtbl.u8      d3, {d0, d1, d2}, d28        \n"
+    "vtbl.u8      d4, {d0, d1, d2}, d29        \n"
+
+    MEMACCESS(1)
+    "vst1.8       {d3}, [%1]!                  \n"
+    MEMACCESS(1)
+    "vst1.32      {d4[0]}, [%1]!               \n"
+    "bgt          1b                           \n"
+  : "+r"(src_ptr),       // %0
+    "+r"(dst_ptr),       // %1
+    "+r"(dst_width),     // %2
+    "+r"(src_stride)     // %3
+  : "r"(&kMult38_Div6),  // %4
+    "r"(&kShuf38_2)      // %5
+  : "q0", "q1", "q2", "q3", "q13", "q14", "memory", "cc"
+  );
+}
+
+void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height) {
+  const uint8* src_tmp = NULL;
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "mov       %0, %1                          \n"
+    "mov       r12, %5                         \n"
+    "veor      q2, q2, q2                      \n"
+    "veor      q3, q3, q3                      \n"
+  "2:                                          \n"
+    // load 16 pixels into q0
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], %3                 \n"
+    "vaddw.u8   q3, q3, d1                     \n"
+    "vaddw.u8   q2, q2, d0                     \n"
+    "subs       r12, r12, #1                   \n"
+    "bgt        2b                             \n"
+    MEMACCESS(2)
+    "vst1.16    {q2, q3}, [%2]!                \n"  // store pixels
+    "add        %1, %1, #16                    \n"
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop
+    "bgt        1b                             \n"
+  : "+r"(src_tmp),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_ptr),          // %2
+    "+r"(src_stride),       // %3
+    "+r"(src_width),        // %4
+    "+r"(src_height)        // %5
+  :
+  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD2_DATA8_LANE(n)                                    \
+    "lsr        %5, %3, #16                    \n"             \
+    "add        %6, %1, %5                     \n"             \
+    "add        %3, %3, %4                     \n"             \
+    MEMACCESS(6)                                               \
+    "vld2.8     {d6["#n"], d7["#n"]}, [%6]     \n"
+
+void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                          int dst_width, int x, int dx) {
+  int dx_offset[4] = {0, 1, 2, 3};
+  int* tmp = dx_offset;
+  const uint8* src_tmp = src_ptr;
+  asm volatile (
+    ".p2align   2                              \n"
+    "vdup.32    q0, %3                         \n"  // x
+    "vdup.32    q1, %4                         \n"  // dx
+    "vld1.32    {q2}, [%5]                     \n"  // 0 1 2 3
+    "vshl.i32   q3, q1, #2                     \n"  // 4 * dx
+    "vmul.s32   q1, q1, q2                     \n"
+    // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
+    "vadd.s32   q1, q1, q0                     \n"
+    // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx
+    "vadd.s32   q2, q1, q3                     \n"
+    "vshl.i32   q0, q3, #1                     \n"  // 8 * dx
+  "1:                                          \n"
+    LOAD2_DATA8_LANE(0)
+    LOAD2_DATA8_LANE(1)
+    LOAD2_DATA8_LANE(2)
+    LOAD2_DATA8_LANE(3)
+    LOAD2_DATA8_LANE(4)
+    LOAD2_DATA8_LANE(5)
+    LOAD2_DATA8_LANE(6)
+    LOAD2_DATA8_LANE(7)
+    "vmov       q10, q1                        \n"
+    "vmov       q11, q2                        \n"
+    "vuzp.16    q10, q11                       \n"
+    "vmovl.u8   q8, d6                         \n"
+    "vmovl.u8   q9, d7                         \n"
+    "vsubl.s16  q11, d18, d16                  \n"
+    "vsubl.s16  q12, d19, d17                  \n"
+    "vmovl.u16  q13, d20                       \n"
+    "vmovl.u16  q10, d21                       \n"
+    "vmul.s32   q11, q11, q13                  \n"
+    "vmul.s32   q12, q12, q10                  \n"
+    "vshrn.s32  d18, q11, #16                  \n"
+    "vshrn.s32  d19, q12, #16                  \n"
+    "vadd.s16   q8, q8, q9                     \n"
+    "vmovn.s16  d6, q8                         \n"
+
+    MEMACCESS(0)
+    "vst1.8     {d6}, [%0]!                    \n"  // store pixels
+    "vadd.s32   q1, q1, q0                     \n"
+    "vadd.s32   q2, q2, q0                     \n"
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "bgt        1b                             \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(x),                // %3
+    "+r"(dx),               // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3",
+    "q8", "q9", "q10", "q11", "q12", "q13"
+  );
+}
+
+#undef LOAD2_DATA8_LANE
+
+// 16x2 -> 16x1
+void ScaleFilterRows_NEON(uint8* dst_ptr,
+                          const uint8* src_ptr, ptrdiff_t src_stride,
+                          int dst_width, int source_y_fraction) {
+  asm volatile (
+    "cmp          %4, #0                       \n"
+    "beq          100f                         \n"
+    "add          %2, %1                       \n"
+    "cmp          %4, #64                      \n"
+    "beq          75f                          \n"
+    "cmp          %4, #128                     \n"
+    "beq          50f                          \n"
+    "cmp          %4, #192                     \n"
+    "beq          25f                          \n"
+
+    "vdup.8       d5, %4                       \n"
+    "rsb          %4, #256                     \n"
+    "vdup.8       d4, %4                       \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vmull.u8     q13, d0, d4                  \n"
+    "vmull.u8     q14, d1, d4                  \n"
+    "vmlal.u8     q13, d2, d5                  \n"
+    "vmlal.u8     q14, d3, d5                  \n"
+    "vrshrn.u16   d0, q13, #8                  \n"
+    "vrshrn.u16   d1, q14, #8                  \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          1b                           \n"
+    "b            99f                          \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vrhadd.u8    q0, q1                       \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          25b                          \n"
+    "b            99f                          \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q1}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          50b                          \n"
+    "b            99f                          \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "vld1.8       {q1}, [%1]!                  \n"
+    MEMACCESS(2)
+    "vld1.8       {q0}, [%2]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    "vrhadd.u8    q0, q1                       \n"
+    "vrhadd.u8    q0, q1                       \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          75b                          \n"
+    "b            99f                          \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "vld1.8       {q0}, [%1]!                  \n"
+    "subs         %3, %3, #16                  \n"
+    MEMACCESS(0)
+    "vst1.8       {q0}, [%0]!                  \n"
+    "bgt          100b                         \n"
+
+  "99:                                         \n"
+    MEMACCESS(0)
+    "vst1.8       {d1[7]}, [%0]                \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction) // %4
+  :
+  : "q0", "q1", "d4", "d5", "q13", "q14", "memory", "cc"
+  );
+}
+
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    MEMACCESS(0)
+    "vld2.32    {q0, q1}, [%0]!                \n"
+    MEMACCESS(0)
+    "vld2.32    {q2, q3}, [%0]!                \n"
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    MEMACCESS(1)
+    "vst1.8     {q1}, [%1]!                    \n"  // store odd pixels
+    MEMACCESS(1)
+    "vst1.8     {q3}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
+    "vrshrn.u16 d0, q0, #1                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #1                     \n"
+    "vrshrn.u16 d2, q2, #1                     \n"
+    "vrshrn.u16 d3, q3, #1                     \n"
+    MEMACCESS(1)
+    "vst4.8     {d0, d1, d2, d3}, [%1]!        \n"
+    "bgt       1b                              \n"
+  : "+r"(src_argb),         // %0
+    "+r"(dst_argb),         // %1
+    "+r"(dst_width)         // %2
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3"     // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %1, %0                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels.
+    MEMACCESS(0)
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vpaddl.u8  q0, q0                         \n"  // B 16 bytes -> 8 shorts.
+    "vpaddl.u8  q1, q1                         \n"  // G 16 bytes -> 8 shorts.
+    "vpaddl.u8  q2, q2                         \n"  // R 16 bytes -> 8 shorts.
+    "vpaddl.u8  q3, q3                         \n"  // A 16 bytes -> 8 shorts.
+    MEMACCESS(1)
+    "vld4.8     {d16, d18, d20, d22}, [%1]!    \n"  // load 8 more ARGB pixels.
+    MEMACCESS(1)
+    "vld4.8     {d17, d19, d21, d23}, [%1]!    \n"  // load last 8 ARGB pixels.
+    "vpadal.u8  q0, q8                         \n"  // B 16 bytes -> 8 shorts.
+    "vpadal.u8  q1, q9                         \n"  // G 16 bytes -> 8 shorts.
+    "vpadal.u8  q2, q10                        \n"  // R 16 bytes -> 8 shorts.
+    "vpadal.u8  q3, q11                        \n"  // A 16 bytes -> 8 shorts.
+    "vrshrn.u16 d0, q0, #2                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #2                     \n"
+    "vrshrn.u16 d2, q2, #2                     \n"
+    "vrshrn.u16 d3, q3, #2                     \n"
+    MEMACCESS(2)
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %3, lsl #2                \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.32    {d0[0]}, [%0], r12             \n"
+    MEMACCESS(0)
+    "vld1.32    {d0[1]}, [%0], r12             \n"
+    MEMACCESS(0)
+    "vld1.32    {d1[0]}, [%0], r12             \n"
+    MEMACCESS(0)
+    "vld1.32    {d1[1]}, [%0], r12             \n"
+    "subs       %2, %2, #4                     \n"  // 4 pixels per loop.
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(dst_width)    // %2
+  : "r"(src_stepx)     // %3
+  : "memory", "cc", "r12", "q0"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "mov        r12, %4, lsl #2                \n"
+    "add        %1, %1, %0                     \n"
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {d0}, [%0], r12                \n"  // Read 4 2x2 blocks -> 2x1
+    MEMACCESS(1)
+    "vld1.8     {d1}, [%1], r12                \n"
+    MEMACCESS(0)
+    "vld1.8     {d2}, [%0], r12                \n"
+    MEMACCESS(1)
+    "vld1.8     {d3}, [%1], r12                \n"
+    MEMACCESS(0)
+    "vld1.8     {d4}, [%0], r12                \n"
+    MEMACCESS(1)
+    "vld1.8     {d5}, [%1], r12                \n"
+    MEMACCESS(0)
+    "vld1.8     {d6}, [%0], r12                \n"
+    MEMACCESS(1)
+    "vld1.8     {d7}, [%1], r12                \n"
+    "vaddl.u8   q0, d0, d1                     \n"
+    "vaddl.u8   q1, d2, d3                     \n"
+    "vaddl.u8   q2, d4, d5                     \n"
+    "vaddl.u8   q3, d6, d7                     \n"
+    "vswp.8     d1, d2                         \n"  // ab_cd -> ac_bd
+    "vswp.8     d5, d6                         \n"  // ef_gh -> eg_fh
+    "vadd.u16   q0, q0, q1                     \n"  // (a+b)_(c+d)
+    "vadd.u16   q2, q2, q3                     \n"  // (e+f)_(g+h)
+    "vrshrn.u16 d0, q0, #2                     \n"  // first 2 pixels.
+    "vrshrn.u16 d1, q2, #2                     \n"  // next 2 pixels.
+    "subs       %3, %3, #4                     \n"  // 4 pixels per loop.
+    MEMACCESS(2)
+    "vst1.8     {q0}, [%2]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(src_stride),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(dst_width)    // %3
+  : "r"(src_stepx)     // %4
+  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"
+  );
+}
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD1_DATA32_LANE(dn, n)                               \
+    "lsr        %5, %3, #16                    \n"             \
+    "add        %6, %1, %5, lsl #2             \n"             \
+    "add        %3, %3, %4                     \n"             \
+    MEMACCESS(6)                                               \
+    "vld1.32    {"#dn"["#n"]}, [%6]            \n"
+
+void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  int tmp = 0;
+  const uint8* src_tmp = src_argb;
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    LOAD1_DATA32_LANE(d0, 0)
+    LOAD1_DATA32_LANE(d0, 1)
+    LOAD1_DATA32_LANE(d1, 0)
+    LOAD1_DATA32_LANE(d1, 1)
+    LOAD1_DATA32_LANE(d2, 0)
+    LOAD1_DATA32_LANE(d2, 1)
+    LOAD1_DATA32_LANE(d3, 0)
+    LOAD1_DATA32_LANE(d3, 1)
+
+    MEMACCESS(0)
+    "vst1.32     {q0, q1}, [%0]!               \n"  // store pixels
+    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "bgt        1b                             \n"
+  : "+r"(dst_argb),         // %0
+    "+r"(src_argb),         // %1
+    "+r"(dst_width),        // %2
+    "+r"(x),                // %3
+    "+r"(dx),               // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "q0", "q1"
+  );
+}
+
+#undef LOAD1_DATA32_LANE
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD2_DATA32_LANE(dn1, dn2, n)                         \
+    "lsr        %5, %3, #16                           \n"      \
+    "add        %6, %1, %5, lsl #2                    \n"      \
+    "add        %3, %3, %4                            \n"      \
+    MEMACCESS(6)                                               \
+    "vld2.32    {"#dn1"["#n"], "#dn2"["#n"]}, [%6]    \n"
+
+void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                              int dst_width, int x, int dx) {
+  int dx_offset[4] = {0, 1, 2, 3};
+  int* tmp = dx_offset;
+  const uint8* src_tmp = src_argb;
+  asm volatile (
+    ".p2align   2                              \n"
+    "vdup.32    q0, %3                         \n"  // x
+    "vdup.32    q1, %4                         \n"  // dx
+    "vld1.32    {q2}, [%5]                     \n"  // 0 1 2 3
+    "vshl.i32   q9, q1, #2                     \n"  // 4 * dx
+    "vmul.s32   q1, q1, q2                     \n"
+    "vmov.i8    q3, #0x7f                      \n"  // 0x7F
+    "vmov.i16   q15, #0x7f                     \n"  // 0x7F
+    // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
+    "vadd.s32   q8, q1, q0                     \n"
+  "1:                                          \n"
+    // d0, d1: a
+    // d2, d3: b
+    LOAD2_DATA32_LANE(d0, d2, 0)
+    LOAD2_DATA32_LANE(d0, d2, 1)
+    LOAD2_DATA32_LANE(d1, d3, 0)
+    LOAD2_DATA32_LANE(d1, d3, 1)
+    "vshrn.i32   d22, q8, #9                   \n"
+    "vand.16     d22, d22, d30                 \n"
+    "vdup.8      d24, d22[0]                   \n"
+    "vdup.8      d25, d22[2]                   \n"
+    "vdup.8      d26, d22[4]                   \n"
+    "vdup.8      d27, d22[6]                   \n"
+    "vext.8      d4, d24, d25, #4              \n"
+    "vext.8      d5, d26, d27, #4              \n"  // f
+    "veor.8      q10, q2, q3                   \n"  // 0x7f ^ f
+    "vmull.u8    q11, d0, d20                  \n"
+    "vmull.u8    q12, d1, d21                  \n"
+    "vmull.u8    q13, d2, d4                   \n"
+    "vmull.u8    q14, d3, d5                   \n"
+    "vadd.i16    q11, q11, q13                 \n"
+    "vadd.i16    q12, q12, q14                 \n"
+    "vshrn.i16   d0, q11, #7                   \n"
+    "vshrn.i16   d1, q12, #7                   \n"
+
+    MEMACCESS(0)
+    "vst1.32     {d0, d1}, [%0]!               \n"  // store pixels
+    "vadd.s32    q8, q8, q9                    \n"
+    "subs        %2, %2, #4                    \n"  // 4 processed per loop
+    "bgt         1b                            \n"
+  : "+r"(dst_argb),         // %0
+    "+r"(src_argb),         // %1
+    "+r"(dst_width),        // %2
+    "+r"(x),                // %3
+    "+r"(dx),               // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "q0", "q1", "q2", "q3", "q8", "q9",
+    "q10", "q11", "q12", "q13", "q14", "q15"
+  );
+}
+
+#undef LOAD2_DATA32_LANE
+
+#endif  // defined(__ARM_NEON__) && !defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/scale_neon64.cc b/third_party/aom/third_party/libyuv/source/scale_neon64.cc
new file mode 100644
index 0000000000..1d55193579
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/scale_neon64.cc
@@ -0,0 +1,1042 @@
+/*
+ *  Copyright 2014 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/scale.h"
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for GCC Neon armv8 64 bit.
+#if !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+// Read 32x1 throw away even pixels, and write 16x1.
+void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    // load even pixels into v0, odd into v1
+    MEMACCESS(0)
+    "ld2        {v0.16b,v1.16b}, [%0], #32     \n"
+    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
+    MEMACCESS(1)
+    "st1        {v1.16b}, [%1], #16            \n"  // store odd pixels
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "v0", "v1"              // Clobber List
+  );
+}
+
+// Read 32x1 average down and write 16x1.
+void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b,v1.16b}, [%0], #32     \n"  // load pixels and post inc
+    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
+    "uaddlp     v0.8h, v0.16b                  \n"  // add adjacent
+    "uaddlp     v1.8h, v1.16b                  \n"
+    "rshrn      v0.8b, v0.8h, #1               \n"  // downshift, round and pack
+    "rshrn2     v0.16b, v1.8h, #1              \n"
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "v0", "v1"     // Clobber List
+  );
+}
+
+// Read 32x2 average down and write 16x1.
+void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %1, %0                     \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b,v1.16b}, [%0], #32    \n"  // load row 1 and post inc
+    MEMACCESS(1)
+    "ld1        {v2.16b, v3.16b}, [%1], #32    \n"  // load row 2 and post inc
+    "subs       %w3, %w3, #16                  \n"  // 16 processed per loop
+    "uaddlp     v0.8h, v0.16b                  \n"  // row 1 add adjacent
+    "uaddlp     v1.8h, v1.16b                  \n"
+    "uadalp     v0.8h, v2.16b                  \n"  // row 2 add adjacent + row1
+    "uadalp     v1.8h, v3.16b                  \n"
+    "rshrn      v0.8b, v0.8h, #2               \n"  // downshift, round and pack
+    "rshrn2     v0.16b, v1.8h, #2              \n"
+    MEMACCESS(2)
+    "st1        {v0.16b}, [%2], #16            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(src_stride),       // %1
+    "+r"(dst),              // %2
+    "+r"(dst_width)         // %3
+  :
+  : "v0", "v1", "v2", "v3"     // Clobber List
+  );
+}
+
+void ScaleRowDown4_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld4     {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32          \n"  // src line 0
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
+    MEMACCESS(1)
+    "st1     {v2.8b}, [%1], #8                 \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "v0", "v1", "v2", "v3", "memory", "cc"
+  );
+}
+
+void ScaleRowDown4Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride;
+  const uint8* src_ptr2 = src_ptr + src_stride * 2;
+  const uint8* src_ptr3 = src_ptr + src_stride * 3;
+asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1     {v0.16b}, [%0], #16               \n"   // load up 16x4
+    MEMACCESS(3)
+    "ld1     {v1.16b}, [%2], #16               \n"
+    MEMACCESS(4)
+    "ld1     {v2.16b}, [%3], #16               \n"
+    MEMACCESS(5)
+    "ld1     {v3.16b}, [%4], #16               \n"
+    "subs    %w5, %w5, #4                      \n"
+    "uaddlp  v0.8h, v0.16b                     \n"
+    "uadalp  v0.8h, v1.16b                     \n"
+    "uadalp  v0.8h, v2.16b                     \n"
+    "uadalp  v0.8h, v3.16b                     \n"
+    "addp    v0.8h, v0.8h, v0.8h               \n"
+    "rshrn   v0.8b, v0.8h, #4                  \n"   // divide by 16 w/rounding
+    MEMACCESS(1)
+    "st1    {v0.s}[0], [%1], #4                \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),   // %0
+    "+r"(dst_ptr),   // %1
+    "+r"(src_ptr1),  // %2
+    "+r"(src_ptr2),  // %3
+    "+r"(src_ptr3),  // %4
+    "+r"(dst_width)  // %5
+  :
+  : "v0", "v1", "v2", "v3", "memory", "cc"
+  );
+}
+
+// Down scale from 4 to 3 pixels. Use the neon multilane read/write
+// to load up the every 4th pixel into a 4 different registers.
+// Point samples 32 pixels to 24 pixels.
+void ScaleRowDown34_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+  "1:                                                  \n"
+    MEMACCESS(0)
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
+    "subs      %w2, %w2, #24                           \n"
+    "orr       v2.16b, v3.16b, v3.16b                  \n"  // order v0, v1, v2
+    MEMACCESS(1)
+    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  :
+  : "v0", "v1", "v2", "v3", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_0_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movi      v20.8b, #3                              \n"
+    "add       %3, %3, %0                              \n"
+  "1:                                                  \n"
+    MEMACCESS(0)
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
+    MEMACCESS(3)
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32                \n"  // src line 1
+    "subs         %w2, %w2, #24                        \n"
+
+    // filter src line 0 with src line 1
+    // expand chars to shorts to allow for room
+    // when adding lines together
+    "ushll     v16.8h, v4.8b, #0                       \n"
+    "ushll     v17.8h, v5.8b, #0                       \n"
+    "ushll     v18.8h, v6.8b, #0                       \n"
+    "ushll     v19.8h, v7.8b, #0                       \n"
+
+    // 3 * line_0 + line_1
+    "umlal     v16.8h, v0.8b, v20.8b                   \n"
+    "umlal     v17.8h, v1.8b, v20.8b                   \n"
+    "umlal     v18.8h, v2.8b, v20.8b                   \n"
+    "umlal     v19.8h, v3.8b, v20.8b                   \n"
+
+    // (3 * line_0 + line_1) >> 2
+    "uqrshrn   v0.8b, v16.8h, #2                       \n"
+    "uqrshrn   v1.8b, v17.8h, #2                       \n"
+    "uqrshrn   v2.8b, v18.8h, #2                       \n"
+    "uqrshrn   v3.8b, v19.8h, #2                       \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "ushll     v16.8h, v1.8b, #0                       \n"
+    "umlal     v16.8h, v0.8b, v20.8b                   \n"
+    "uqrshrn   v0.8b, v16.8h, #2                       \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "urhadd    v1.8b, v1.8b, v2.8b                     \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "ushll     v16.8h, v2.8b, #0                       \n"
+    "umlal     v16.8h, v3.8b, v20.8b                   \n"
+    "uqrshrn   v2.8b, v16.8h, #2                       \n"
+
+    MEMACCESS(1)
+    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
+
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19",
+    "v20", "memory", "cc"
+  );
+}
+
+void ScaleRowDown34_1_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    "movi      v20.8b, #3                              \n"
+    "add       %3, %3, %0                              \n"
+  "1:                                                  \n"
+    MEMACCESS(0)
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"  // src line 0
+    MEMACCESS(3)
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%3], #32                \n"  // src line 1
+    "subs         %w2, %w2, #24                        \n"
+    // average src line 0 with src line 1
+    "urhadd    v0.8b, v0.8b, v4.8b                     \n"
+    "urhadd    v1.8b, v1.8b, v5.8b                     \n"
+    "urhadd    v2.8b, v2.8b, v6.8b                     \n"
+    "urhadd    v3.8b, v3.8b, v7.8b                     \n"
+
+    // a0 = (src[0] * 3 + s[1] * 1) >> 2
+    "ushll     v4.8h, v1.8b, #0                        \n"
+    "umlal     v4.8h, v0.8b, v20.8b                    \n"
+    "uqrshrn   v0.8b, v4.8h, #2                        \n"
+
+    // a1 = (src[1] * 1 + s[2] * 1) >> 1
+    "urhadd    v1.8b, v1.8b, v2.8b                     \n"
+
+    // a2 = (src[2] * 1 + s[3] * 3) >> 2
+    "ushll     v4.8h, v2.8b, #0                        \n"
+    "umlal     v4.8h, v3.8b, v20.8b                    \n"
+    "uqrshrn   v2.8b, v4.8h, #2                        \n"
+
+    MEMACCESS(1)
+    "st3       {v0.8b,v1.8b,v2.8b}, [%1], #24                \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width),        // %2
+    "+r"(src_stride)        // %3
+  :
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v20", "memory", "cc"
+  );
+}
+
+static uvec8 kShuf38 =
+  { 0, 3, 6, 8, 11, 14, 16, 19, 22, 24, 27, 30, 0, 0, 0, 0 };
+static uvec8 kShuf38_2 =
+  { 0, 16, 32, 2, 18, 33, 4, 20, 34, 6, 22, 35, 0, 0, 0, 0 };
+static vec16 kMult38_Div6 =
+  { 65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12,
+    65536 / 12, 65536 / 12, 65536 / 12, 65536 / 12 };
+static vec16 kMult38_Div9 =
+  { 65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18,
+    65536 / 18, 65536 / 18, 65536 / 18, 65536 / 18 };
+
+// 32 -> 12
+void ScaleRowDown38_NEON(const uint8* src_ptr,
+                         ptrdiff_t src_stride,
+                         uint8* dst_ptr, int dst_width) {
+  asm volatile (
+    MEMACCESS(3)
+    "ld1       {v3.16b}, [%3]                          \n"
+  "1:                                                  \n"
+    MEMACCESS(0)
+    "ld1       {v0.16b,v1.16b}, [%0], #32             \n"
+    "subs      %w2, %w2, #12                           \n"
+    "tbl       v2.16b, {v0.16b,v1.16b}, v3.16b        \n"
+    MEMACCESS(1)
+    "st1       {v2.8b}, [%1], #8                       \n"
+    MEMACCESS(1)
+    "st1       {v2.s}[2], [%1], #4                     \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(dst_width)         // %2
+  : "r"(&kShuf38)           // %3
+  : "v0", "v1", "v2", "v3", "memory", "cc"
+  );
+}
+
+// 32x3 -> 12x1
+void OMITFP ScaleRowDown38_3_Box_NEON(const uint8* src_ptr,
+                                      ptrdiff_t src_stride,
+                                      uint8* dst_ptr, int dst_width) {
+  const uint8* src_ptr1 = src_ptr + src_stride * 2;
+  ptrdiff_t tmp_src_stride = src_stride;
+
+  asm volatile (
+    MEMACCESS(5)
+    "ld1       {v29.8h}, [%5]                          \n"
+    MEMACCESS(6)
+    "ld1       {v30.16b}, [%6]                         \n"
+    MEMACCESS(7)
+    "ld1       {v31.8h}, [%7]                          \n"
+    "add       %2, %2, %0                              \n"
+  "1:                                                  \n"
+
+    // 00 40 01 41 02 42 03 43
+    // 10 50 11 51 12 52 13 53
+    // 20 60 21 61 22 62 23 63
+    // 30 70 31 71 32 72 33 73
+    MEMACCESS(0)
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"
+    MEMACCESS(3)
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32                \n"
+    MEMACCESS(4)
+    "ld4       {v16.8b,v17.8b,v18.8b,v19.8b}, [%3], #32              \n"
+    "subs      %w4, %w4, #12                           \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // 00 10 01 11 02 12 03 13
+    // 40 50 41 51 42 52 43 53
+    "trn1      v20.8b, v0.8b, v1.8b                    \n"
+    "trn2      v21.8b, v0.8b, v1.8b                    \n"
+    "trn1      v22.8b, v4.8b, v5.8b                    \n"
+    "trn2      v23.8b, v4.8b, v5.8b                    \n"
+    "trn1      v24.8b, v16.8b, v17.8b                  \n"
+    "trn2      v25.8b, v16.8b, v17.8b                  \n"
+
+    // 20 30 21 31 22 32 23 33
+    // 60 70 61 71 62 72 63 73
+    "trn1      v0.8b, v2.8b, v3.8b                     \n"
+    "trn2      v1.8b, v2.8b, v3.8b                     \n"
+    "trn1      v4.8b, v6.8b, v7.8b                     \n"
+    "trn2      v5.8b, v6.8b, v7.8b                     \n"
+    "trn1      v16.8b, v18.8b, v19.8b                  \n"
+    "trn2      v17.8b, v18.8b, v19.8b                  \n"
+
+    // 00+10 01+11 02+12 03+13
+    // 40+50 41+51 42+52 43+53
+    "uaddlp    v20.4h, v20.8b                          \n"
+    "uaddlp    v21.4h, v21.8b                          \n"
+    "uaddlp    v22.4h, v22.8b                          \n"
+    "uaddlp    v23.4h, v23.8b                          \n"
+    "uaddlp    v24.4h, v24.8b                          \n"
+    "uaddlp    v25.4h, v25.8b                          \n"
+
+    // 60+70 61+71 62+72 63+73
+    "uaddlp    v1.4h, v1.8b                            \n"
+    "uaddlp    v5.4h, v5.8b                            \n"
+    "uaddlp    v17.4h, v17.8b                          \n"
+
+    // combine source lines
+    "add       v20.4h, v20.4h, v22.4h                  \n"
+    "add       v21.4h, v21.4h, v23.4h                  \n"
+    "add       v20.4h, v20.4h, v24.4h                  \n"
+    "add       v21.4h, v21.4h, v25.4h                  \n"
+    "add       v2.4h, v1.4h, v5.4h                     \n"
+    "add       v2.4h, v2.4h, v17.4h                    \n"
+
+    // dst_ptr[3] = (s[6 + st * 0] + s[7 + st * 0]
+    //             + s[6 + st * 1] + s[7 + st * 1]
+    //             + s[6 + st * 2] + s[7 + st * 2]) / 6
+    "sqrdmulh  v2.8h, v2.8h, v29.8h                    \n"
+    "xtn       v2.8b,  v2.8h                           \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+    "ushll     v16.8h, v16.8b, #0                      \n"
+    "uaddl     v0.8h, v0.8b, v4.8b                     \n"
+
+    // combine source lines
+    "add       v0.8h, v0.8h, v16.8h                    \n"
+
+    // xx 20 xx 21 xx 22 xx 23
+    // xx 30 xx 31 xx 32 xx 33
+    "trn1      v1.8h, v0.8h, v0.8h                     \n"
+    "trn2      v4.8h, v0.8h, v0.8h                     \n"
+    "xtn       v0.4h, v1.4s                            \n"
+    "xtn       v4.4h, v4.4s                            \n"
+
+    // 0+1+2, 3+4+5
+    "add       v20.8h, v20.8h, v0.8h                   \n"
+    "add       v21.8h, v21.8h, v4.8h                   \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "sqrdmulh  v0.8h, v20.8h, v31.8h                   \n"
+    "sqrdmulh  v1.8h, v21.8h, v31.8h                   \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+    "tbl       v3.16b, {v0.16b, v1.16b, v2.16b}, v30.16b \n"
+
+    MEMACCESS(1)
+    "st1       {v3.8b}, [%1], #8                       \n"
+    MEMACCESS(1)
+    "st1       {v3.s}[2], [%1], #4                     \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst_ptr),          // %1
+    "+r"(tmp_src_stride),   // %2
+    "+r"(src_ptr1),         // %3
+    "+r"(dst_width)         // %4
+  : "r"(&kMult38_Div6),     // %5
+    "r"(&kShuf38_2),        // %6
+    "r"(&kMult38_Div9)      // %7
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
+    "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v29",
+    "v30", "v31", "memory", "cc"
+  );
+}
+
+// 32x2 -> 12x1
+void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
+                               ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width) {
+  // TODO(fbarchard): use src_stride directly for clang 3.5+.
+  ptrdiff_t tmp_src_stride = src_stride;
+  asm volatile (
+    MEMACCESS(4)
+    "ld1       {v30.8h}, [%4]                          \n"
+    MEMACCESS(5)
+    "ld1       {v31.16b}, [%5]                         \n"
+    "add       %2, %2, %0                              \n"
+  "1:                                                  \n"
+
+    // 00 40 01 41 02 42 03 43
+    // 10 50 11 51 12 52 13 53
+    // 20 60 21 61 22 62 23 63
+    // 30 70 31 71 32 72 33 73
+    MEMACCESS(0)
+    "ld4       {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32                \n"
+    MEMACCESS(3)
+    "ld4       {v4.8b,v5.8b,v6.8b,v7.8b}, [%2], #32                \n"
+    "subs      %w3, %w3, #12                           \n"
+
+    // Shuffle the input data around to get align the data
+    //  so adjacent data can be added. 0,1 - 2,3 - 4,5 - 6,7
+    // 00 10 01 11 02 12 03 13
+    // 40 50 41 51 42 52 43 53
+    "trn1      v16.8b, v0.8b, v1.8b                    \n"
+    "trn2      v17.8b, v0.8b, v1.8b                    \n"
+    "trn1      v18.8b, v4.8b, v5.8b                    \n"
+    "trn2      v19.8b, v4.8b, v5.8b                    \n"
+
+    // 20 30 21 31 22 32 23 33
+    // 60 70 61 71 62 72 63 73
+    "trn1      v0.8b, v2.8b, v3.8b                     \n"
+    "trn2      v1.8b, v2.8b, v3.8b                     \n"
+    "trn1      v4.8b, v6.8b, v7.8b                     \n"
+    "trn2      v5.8b, v6.8b, v7.8b                     \n"
+
+    // 00+10 01+11 02+12 03+13
+    // 40+50 41+51 42+52 43+53
+    "uaddlp    v16.4h, v16.8b                          \n"
+    "uaddlp    v17.4h, v17.8b                          \n"
+    "uaddlp    v18.4h, v18.8b                          \n"
+    "uaddlp    v19.4h, v19.8b                          \n"
+
+    // 60+70 61+71 62+72 63+73
+    "uaddlp    v1.4h, v1.8b                            \n"
+    "uaddlp    v5.4h, v5.8b                            \n"
+
+    // combine source lines
+    "add       v16.4h, v16.4h, v18.4h                  \n"
+    "add       v17.4h, v17.4h, v19.4h                  \n"
+    "add       v2.4h, v1.4h, v5.4h                     \n"
+
+    // dst_ptr[3] = (s[6] + s[7] + s[6+st] + s[7+st]) / 4
+    "uqrshrn   v2.8b, v2.8h, #2                        \n"
+
+    // Shuffle 2,3 reg around so that 2 can be added to the
+    //  0,1 reg and 3 can be added to the 4,5 reg. This
+    //  requires expanding from u8 to u16 as the 0,1 and 4,5
+    //  registers are already expanded. Then do transposes
+    //  to get aligned.
+    // xx 20 xx 30 xx 21 xx 31 xx 22 xx 32 xx 23 xx 33
+
+    // combine source lines
+    "uaddl     v0.8h, v0.8b, v4.8b                     \n"
+
+    // xx 20 xx 21 xx 22 xx 23
+    // xx 30 xx 31 xx 32 xx 33
+    "trn1      v1.8h, v0.8h, v0.8h                     \n"
+    "trn2      v4.8h, v0.8h, v0.8h                     \n"
+    "xtn       v0.4h, v1.4s                            \n"
+    "xtn       v4.4h, v4.4s                            \n"
+
+    // 0+1+2, 3+4+5
+    "add       v16.8h, v16.8h, v0.8h                   \n"
+    "add       v17.8h, v17.8h, v4.8h                   \n"
+
+    // Need to divide, but can't downshift as the the value
+    //  isn't a power of 2. So multiply by 65536 / n
+    //  and take the upper 16 bits.
+    "sqrdmulh  v0.8h, v16.8h, v30.8h                   \n"
+    "sqrdmulh  v1.8h, v17.8h, v30.8h                   \n"
+
+    // Align for table lookup, vtbl requires registers to
+    //  be adjacent
+
+    "tbl       v3.16b, {v0.16b, v1.16b, v2.16b}, v31.16b \n"
+
+    MEMACCESS(1)
+    "st1       {v3.8b}, [%1], #8                       \n"
+    MEMACCESS(1)
+    "st1       {v3.s}[2], [%1], #4                     \n"
+    "b.gt      1b                                      \n"
+  : "+r"(src_ptr),         // %0
+    "+r"(dst_ptr),         // %1
+    "+r"(tmp_src_stride),  // %2
+    "+r"(dst_width)        // %3
+  : "r"(&kMult38_Div6),    // %4
+    "r"(&kShuf38_2)        // %5
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17",
+    "v18", "v19", "v30", "v31", "memory", "cc"
+  );
+}
+
+void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height) {
+  const uint8* src_tmp = NULL;
+  asm volatile (
+  "1:                                          \n"
+    "mov       %0, %1                          \n"
+    "mov       w12, %w5                        \n"
+    "eor       v2.16b, v2.16b, v2.16b          \n"
+    "eor       v3.16b, v3.16b, v3.16b          \n"
+  "2:                                          \n"
+    // load 16 pixels into q0
+    MEMACCESS(0)
+    "ld1       {v0.16b}, [%0], %3              \n"
+    "uaddw2    v3.8h, v3.8h, v0.16b            \n"
+    "uaddw     v2.8h, v2.8h, v0.8b             \n"
+    "subs      w12, w12, #1                    \n"
+    "b.gt      2b                              \n"
+    MEMACCESS(2)
+    "st1      {v2.8h, v3.8h}, [%2], #32        \n"  // store pixels
+    "add      %1, %1, #16                      \n"
+    "subs     %w4, %w4, #16                    \n"  // 16 processed per loop
+    "b.gt     1b                               \n"
+  : "+r"(src_tmp),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_ptr),          // %2
+    "+r"(src_stride),       // %3
+    "+r"(src_width),        // %4
+    "+r"(src_height)        // %5
+  :
+  : "memory", "cc", "w12", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD2_DATA8_LANE(n)                                    \
+    "lsr        %5, %3, #16                    \n"             \
+    "add        %6, %1, %5                    \n"              \
+    "add        %3, %3, %4                     \n"             \
+    MEMACCESS(6)                                               \
+    "ld2        {v4.b, v5.b}["#n"], [%6]      \n"
+
+void ScaleFilterCols_NEON(uint8* dst_ptr, const uint8* src_ptr,
+                          int dst_width, int x, int dx) {
+  int dx_offset[4] = {0, 1, 2, 3};
+  int* tmp = dx_offset;
+  const uint8* src_tmp = src_ptr;
+  int64 dst_width64 = (int64) dst_width;  // Work around ios 64 bit warning.
+  int64 x64 = (int64) x;
+  int64 dx64 = (int64) dx;
+  asm volatile (
+    "dup        v0.4s, %w3                     \n"  // x
+    "dup        v1.4s, %w4                     \n"  // dx
+    "ld1        {v2.4s}, [%5]                  \n"  // 0 1 2 3
+    "shl        v3.4s, v1.4s, #2               \n"  // 4 * dx
+    "mul        v1.4s, v1.4s, v2.4s            \n"
+    // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
+    "add        v1.4s, v1.4s, v0.4s            \n"
+    // x + 4 * dx, x + 5 * dx, x + 6 * dx, x + 7 * dx
+    "add        v2.4s, v1.4s, v3.4s            \n"
+    "shl        v0.4s, v3.4s, #1               \n"  // 8 * dx
+  "1:                                          \n"
+    LOAD2_DATA8_LANE(0)
+    LOAD2_DATA8_LANE(1)
+    LOAD2_DATA8_LANE(2)
+    LOAD2_DATA8_LANE(3)
+    LOAD2_DATA8_LANE(4)
+    LOAD2_DATA8_LANE(5)
+    LOAD2_DATA8_LANE(6)
+    LOAD2_DATA8_LANE(7)
+    "mov       v6.16b, v1.16b                  \n"
+    "mov       v7.16b, v2.16b                  \n"
+    "uzp1      v6.8h, v6.8h, v7.8h             \n"
+    "ushll     v4.8h, v4.8b, #0                \n"
+    "ushll     v5.8h, v5.8b, #0                \n"
+    "ssubl     v16.4s, v5.4h, v4.4h            \n"
+    "ssubl2    v17.4s, v5.8h, v4.8h            \n"
+    "ushll     v7.4s, v6.4h, #0                \n"
+    "ushll2    v6.4s, v6.8h, #0                \n"
+    "mul       v16.4s, v16.4s, v7.4s           \n"
+    "mul       v17.4s, v17.4s, v6.4s           \n"
+    "shrn      v6.4h, v16.4s, #16              \n"
+    "shrn2     v6.8h, v17.4s, #16              \n"
+    "add       v4.8h, v4.8h, v6.8h             \n"
+    "xtn       v4.8b, v4.8h                    \n"
+
+    MEMACCESS(0)
+    "st1       {v4.8b}, [%0], #8               \n"  // store pixels
+    "add       v1.4s, v1.4s, v0.4s             \n"
+    "add       v2.4s, v2.4s, v0.4s             \n"
+    "subs      %w2, %w2, #8                    \n"  // 8 processed per loop
+    "b.gt      1b                              \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_width64),      // %2
+    "+r"(x64),              // %3
+    "+r"(dx64),             // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "v0", "v1", "v2", "v3",
+    "v4", "v5", "v6", "v7", "v16", "v17"
+  );
+}
+
+#undef LOAD2_DATA8_LANE
+
+// 16x2 -> 16x1
+void ScaleFilterRows_NEON(uint8* dst_ptr,
+                          const uint8* src_ptr, ptrdiff_t src_stride,
+                          int dst_width, int source_y_fraction) {
+    int y_fraction = 256 - source_y_fraction;
+  asm volatile (
+    "cmp          %w4, #0                      \n"
+    "b.eq         100f                         \n"
+    "add          %2, %2, %1                   \n"
+    "cmp          %w4, #64                     \n"
+    "b.eq         75f                          \n"
+    "cmp          %w4, #128                    \n"
+    "b.eq         50f                          \n"
+    "cmp          %w4, #192                    \n"
+    "b.eq         25f                          \n"
+
+    "dup          v5.8b, %w4                   \n"
+    "dup          v4.8b, %w5                   \n"
+    // General purpose row blend.
+  "1:                                          \n"
+    MEMACCESS(1)
+    "ld1          {v0.16b}, [%1], #16          \n"
+    MEMACCESS(2)
+    "ld1          {v1.16b}, [%2], #16          \n"
+    "subs         %w3, %w3, #16                \n"
+    "umull        v6.8h, v0.8b, v4.8b          \n"
+    "umull2       v7.8h, v0.16b, v4.16b        \n"
+    "umlal        v6.8h, v1.8b, v5.8b          \n"
+    "umlal2       v7.8h, v1.16b, v5.16b        \n"
+    "rshrn        v0.8b, v6.8h, #8             \n"
+    "rshrn2       v0.16b, v7.8h, #8            \n"
+    MEMACCESS(0)
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         1b                           \n"
+    "b            99f                          \n"
+
+    // Blend 25 / 75.
+  "25:                                         \n"
+    MEMACCESS(1)
+    "ld1          {v0.16b}, [%1], #16          \n"
+    MEMACCESS(2)
+    "ld1          {v1.16b}, [%2], #16          \n"
+    "subs         %w3, %w3, #16                \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    MEMACCESS(0)
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         25b                          \n"
+    "b            99f                          \n"
+
+    // Blend 50 / 50.
+  "50:                                         \n"
+    MEMACCESS(1)
+    "ld1          {v0.16b}, [%1], #16          \n"
+    MEMACCESS(2)
+    "ld1          {v1.16b}, [%2], #16          \n"
+    "subs         %w3, %w3, #16                \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    MEMACCESS(0)
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         50b                          \n"
+    "b            99f                          \n"
+
+    // Blend 75 / 25.
+  "75:                                         \n"
+    MEMACCESS(1)
+    "ld1          {v1.16b}, [%1], #16          \n"
+    MEMACCESS(2)
+    "ld1          {v0.16b}, [%2], #16          \n"
+    "subs         %w3, %w3, #16                \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    "urhadd       v0.16b, v0.16b, v1.16b       \n"
+    MEMACCESS(0)
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         75b                          \n"
+    "b            99f                          \n"
+
+    // Blend 100 / 0 - Copy row unchanged.
+  "100:                                        \n"
+    MEMACCESS(1)
+    "ld1          {v0.16b}, [%1], #16          \n"
+    "subs         %w3, %w3, #16                \n"
+    MEMACCESS(0)
+    "st1          {v0.16b}, [%0], #16          \n"
+    "b.gt         100b                         \n"
+
+  "99:                                         \n"
+    MEMACCESS(0)
+    "st1          {v0.b}[15], [%0]             \n"
+  : "+r"(dst_ptr),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(src_stride),       // %2
+    "+r"(dst_width),        // %3
+    "+r"(source_y_fraction),// %4
+    "+r"(y_fraction)        // %5
+  :
+  : "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "memory", "cc"
+  );
+}
+
+void ScaleARGBRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    // load even pixels into q0, odd into q1
+    MEMACCESS (0)
+    "ld2        {v0.4s, v1.4s}, [%0], #32      \n"
+    MEMACCESS (0)
+    "ld2        {v2.4s, v3.4s}, [%0], #32      \n"
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
+    MEMACCESS (1)
+    "st1        {v1.16b}, [%1], #16            \n"  // store odd pixels
+    MEMACCESS (1)
+    "st1        {v3.16b}, [%1], #16            \n"
+    "b.gt       1b                             \n"
+  : "+r" (src_ptr),          // %0
+    "+r" (dst),              // %1
+    "+r" (dst_width)         // %2
+  :
+  : "memory", "cc", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Linear_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS (0)
+    // load 8 ARGB pixels.
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64   \n"
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    "uaddlp     v3.8h, v3.16b                  \n"  // A 16 bytes -> 8 shorts.
+    "rshrn      v0.8b, v0.8h, #1               \n"  // downshift, round and pack
+    "rshrn      v1.8b, v1.8h, #1               \n"
+    "rshrn      v2.8b, v2.8h, #1               \n"
+    "rshrn      v3.8b, v3.8h, #1               \n"
+    MEMACCESS (1)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%1], #32     \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),         // %0
+    "+r"(dst_argb),         // %1
+    "+r"(dst_width)         // %2
+  :
+  : "memory", "cc", "v0", "v1", "v2", "v3"    // Clobber List
+  );
+}
+
+void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width) {
+  asm volatile (
+    // change the stride to row 2 pointer
+    "add        %1, %1, %0                     \n"
+  "1:                                          \n"
+    MEMACCESS (0)
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64   \n"  // load 8 ARGB pixels.
+    "subs       %w3, %w3, #8                   \n"  // 8 processed per loop.
+    "uaddlp     v0.8h, v0.16b                  \n"  // B 16 bytes -> 8 shorts.
+    "uaddlp     v1.8h, v1.16b                  \n"  // G 16 bytes -> 8 shorts.
+    "uaddlp     v2.8h, v2.16b                  \n"  // R 16 bytes -> 8 shorts.
+    "uaddlp     v3.8h, v3.16b                  \n"  // A 16 bytes -> 8 shorts.
+    MEMACCESS (1)
+    "ld4        {v16.16b,v17.16b,v18.16b,v19.16b}, [%1], #64 \n"  // load 8 more ARGB pixels.
+    "uadalp     v0.8h, v16.16b                 \n"  // B 16 bytes -> 8 shorts.
+    "uadalp     v1.8h, v17.16b                 \n"  // G 16 bytes -> 8 shorts.
+    "uadalp     v2.8h, v18.16b                 \n"  // R 16 bytes -> 8 shorts.
+    "uadalp     v3.8h, v19.16b                 \n"  // A 16 bytes -> 8 shorts.
+    "rshrn      v0.8b, v0.8h, #2               \n"  // downshift, round and pack
+    "rshrn      v1.8b, v1.8h, #2               \n"
+    "rshrn      v2.8b, v2.8h, #2               \n"
+    "rshrn      v3.8b, v3.8h, #2               \n"
+    MEMACCESS (2)
+    "st4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%2], #32     \n"
+    "b.gt       1b                             \n"
+  : "+r" (src_ptr),          // %0
+    "+r" (src_stride),       // %1
+    "+r" (dst),              // %2
+    "+r" (dst_width)         // %3
+  :
+  : "memory", "cc", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+void ScaleARGBRowDownEven_NEON(const uint8* src_argb,  ptrdiff_t src_stride,
+                               int src_stepx, uint8* dst_argb, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.s}[0], [%0], %3            \n"
+    MEMACCESS(0)
+    "ld1        {v0.s}[1], [%0], %3            \n"
+    MEMACCESS(0)
+    "ld1        {v0.s}[2], [%0], %3            \n"
+    MEMACCESS(0)
+    "ld1        {v0.s}[3], [%0], %3            \n"
+    "subs       %w2, %w2, #4                   \n"  // 4 pixels per loop.
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(dst_argb),    // %1
+    "+r"(dst_width)    // %2
+  : "r"((int64)(src_stepx * 4)) // %3
+  : "memory", "cc", "v0"
+  );
+}
+
+// Reads 4 pixels at a time.
+// Alignment requirement: src_argb 4 byte aligned.
+// TODO(Yang Zhang): Might be worth another optimization pass in future.
+// It could be upgraded to 8 pixels at a time to start with.
+void ScaleARGBRowDownEvenBox_NEON(const uint8* src_argb, ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  asm volatile (
+    "add        %1, %1, %0                     \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.8b}, [%0], %4              \n"  // Read 4 2x2 blocks -> 2x1
+    MEMACCESS(1)
+    "ld1        {v1.8b}, [%1], %4              \n"
+    MEMACCESS(0)
+    "ld1        {v2.8b}, [%0], %4              \n"
+    MEMACCESS(1)
+    "ld1        {v3.8b}, [%1], %4              \n"
+    MEMACCESS(0)
+    "ld1        {v4.8b}, [%0], %4              \n"
+    MEMACCESS(1)
+    "ld1        {v5.8b}, [%1], %4              \n"
+    MEMACCESS(0)
+    "ld1        {v6.8b}, [%0], %4              \n"
+    MEMACCESS(1)
+    "ld1        {v7.8b}, [%1], %4              \n"
+    "uaddl      v0.8h, v0.8b, v1.8b            \n"
+    "uaddl      v2.8h, v2.8b, v3.8b            \n"
+    "uaddl      v4.8h, v4.8b, v5.8b            \n"
+    "uaddl      v6.8h, v6.8b, v7.8b            \n"
+    "mov        v16.d[1], v0.d[1]              \n"  // ab_cd -> ac_bd
+    "mov        v0.d[1], v2.d[0]               \n"
+    "mov        v2.d[0], v16.d[1]              \n"
+    "mov        v16.d[1], v4.d[1]              \n"  // ef_gh -> eg_fh
+    "mov        v4.d[1], v6.d[0]               \n"
+    "mov        v6.d[0], v16.d[1]              \n"
+    "add        v0.8h, v0.8h, v2.8h            \n"  // (a+b)_(c+d)
+    "add        v4.8h, v4.8h, v6.8h            \n"  // (e+f)_(g+h)
+    "rshrn      v0.8b, v0.8h, #2               \n"  // first 2 pixels.
+    "rshrn2     v0.16b, v4.8h, #2              \n"  // next 2 pixels.
+    "subs       %w3, %w3, #4                   \n"  // 4 pixels per loop.
+    MEMACCESS(2)
+    "st1     {v0.16b}, [%2], #16               \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_argb),    // %0
+    "+r"(src_stride),  // %1
+    "+r"(dst_argb),    // %2
+    "+r"(dst_width)    // %3
+  : "r"((int64)(src_stepx * 4)) // %4
+  : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16"
+  );
+}
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD1_DATA32_LANE(vn, n)                               \
+    "lsr        %5, %3, #16                    \n"             \
+    "add        %6, %1, %5, lsl #2             \n"             \
+    "add        %3, %3, %4                     \n"             \
+    MEMACCESS(6)                                               \
+    "ld1        {"#vn".s}["#n"], [%6]          \n"
+
+void ScaleARGBCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  const uint8* src_tmp = src_argb;
+  int64 dst_width64 = (int64) dst_width;  // Work around ios 64 bit warning.
+  int64 x64 = (int64) x;
+  int64 dx64 = (int64) dx;
+  int64 tmp64 = 0;
+  asm volatile (
+  "1:                                          \n"
+    LOAD1_DATA32_LANE(v0, 0)
+    LOAD1_DATA32_LANE(v0, 1)
+    LOAD1_DATA32_LANE(v0, 2)
+    LOAD1_DATA32_LANE(v0, 3)
+    LOAD1_DATA32_LANE(v1, 0)
+    LOAD1_DATA32_LANE(v1, 1)
+    LOAD1_DATA32_LANE(v1, 2)
+    LOAD1_DATA32_LANE(v1, 3)
+
+    MEMACCESS(0)
+    "st1        {v0.4s, v1.4s}, [%0], #32      \n"  // store pixels
+    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
+    "b.gt        1b                            \n"
+  : "+r"(dst_argb),         // %0
+    "+r"(src_argb),         // %1
+    "+r"(dst_width64),      // %2
+    "+r"(x64),              // %3
+    "+r"(dx64),             // %4
+    "+r"(tmp64),            // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "v0", "v1"
+  );
+}
+
+#undef LOAD1_DATA32_LANE
+
+// TODO(Yang Zhang): Investigate less load instructions for
+// the x/dx stepping
+#define LOAD2_DATA32_LANE(vn1, vn2, n)                         \
+    "lsr        %5, %3, #16                           \n"      \
+    "add        %6, %1, %5, lsl #2                    \n"      \
+    "add        %3, %3, %4                            \n"      \
+    MEMACCESS(6)                                               \
+    "ld2        {"#vn1".s, "#vn2".s}["#n"], [%6]      \n"
+
+void ScaleARGBFilterCols_NEON(uint8* dst_argb, const uint8* src_argb,
+                              int dst_width, int x, int dx) {
+  int dx_offset[4] = {0, 1, 2, 3};
+  int* tmp = dx_offset;
+  const uint8* src_tmp = src_argb;
+  int64 dst_width64 = (int64) dst_width;  // Work around ios 64 bit warning.
+  int64 x64 = (int64) x;
+  int64 dx64 = (int64) dx;
+  asm volatile (
+    "dup        v0.4s, %w3                     \n"  // x
+    "dup        v1.4s, %w4                     \n"  // dx
+    "ld1        {v2.4s}, [%5]                  \n"  // 0 1 2 3
+    "shl        v6.4s, v1.4s, #2               \n"  // 4 * dx
+    "mul        v1.4s, v1.4s, v2.4s            \n"
+    "movi       v3.16b, #0x7f                  \n"  // 0x7F
+    "movi       v4.8h, #0x7f                   \n"  // 0x7F
+    // x         , x + 1 * dx, x + 2 * dx, x + 3 * dx
+    "add        v5.4s, v1.4s, v0.4s            \n"
+  "1:                                          \n"
+    // d0, d1: a
+    // d2, d3: b
+    LOAD2_DATA32_LANE(v0, v1, 0)
+    LOAD2_DATA32_LANE(v0, v1, 1)
+    LOAD2_DATA32_LANE(v0, v1, 2)
+    LOAD2_DATA32_LANE(v0, v1, 3)
+    "shrn       v2.4h, v5.4s, #9               \n"
+    "and        v2.8b, v2.8b, v4.8b            \n"
+    "dup        v16.8b, v2.b[0]                \n"
+    "dup        v17.8b, v2.b[2]                \n"
+    "dup        v18.8b, v2.b[4]                \n"
+    "dup        v19.8b, v2.b[6]                \n"
+    "ext        v2.8b, v16.8b, v17.8b, #4      \n"
+    "ext        v17.8b, v18.8b, v19.8b, #4     \n"
+    "ins        v2.d[1], v17.d[0]              \n"  // f
+    "eor        v7.16b, v2.16b, v3.16b         \n"  // 0x7f ^ f
+    "umull      v16.8h, v0.8b, v7.8b           \n"
+    "umull2     v17.8h, v0.16b, v7.16b         \n"
+    "umull      v18.8h, v1.8b, v2.8b           \n"
+    "umull2     v19.8h, v1.16b, v2.16b         \n"
+    "add        v16.8h, v16.8h, v18.8h         \n"
+    "add        v17.8h, v17.8h, v19.8h         \n"
+    "shrn       v0.8b, v16.8h, #7              \n"
+    "shrn2      v0.16b, v17.8h, #7             \n"
+
+    MEMACCESS(0)
+    "st1     {v0.4s}, [%0], #16                \n"  // store pixels
+    "add     v5.4s, v5.4s, v6.4s               \n"
+    "subs    %w2, %w2, #4                      \n"  // 4 processed per loop
+    "b.gt    1b                                \n"
+  : "+r"(dst_argb),         // %0
+    "+r"(src_argb),         // %1
+    "+r"(dst_width64),      // %2
+    "+r"(x64),              // %3
+    "+r"(dx64),             // %4
+    "+r"(tmp),              // %5
+    "+r"(src_tmp)           // %6
+  :
+  : "memory", "cc", "v0", "v1", "v2", "v3", "v4", "v5",
+    "v6", "v7", "v16", "v17", "v18", "v19"
+  );
+}
+
+#undef LOAD2_DATA32_LANE
+
+#endif  // !defined(LIBYUV_DISABLE_NEON) && defined(__aarch64__)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/scale_win.cc b/third_party/aom/third_party/libyuv/source/scale_win.cc
new file mode 100644
index 0000000000..c3896ebad2
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/scale_win.cc
@@ -0,0 +1,1354 @@
+/*
+ *  Copyright 2013 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "libyuv/row.h"
+#include "libyuv/scale_row.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+// This module is for Visual C x86.
+#if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \
+    defined(_MSC_VER) && !defined(__clang__)
+
+// Offsets for source bytes 0 to 9
+static uvec8 kShuf0 =
+  { 0, 1, 3, 4, 5, 7, 8, 9, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 11 to 20 with 8 subtracted = 3 to 12.
+static uvec8 kShuf1 =
+  { 3, 4, 5, 7, 8, 9, 11, 12, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf2 =
+  { 5, 7, 8, 9, 11, 12, 13, 15, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Offsets for source bytes 0 to 10
+static uvec8 kShuf01 =
+  { 0, 1, 1, 2, 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10 };
+
+// Offsets for source bytes 10 to 21 with 8 subtracted = 3 to 13.
+static uvec8 kShuf11 =
+  { 2, 3, 4, 5, 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13 };
+
+// Offsets for source bytes 21 to 31 with 16 subtracted = 5 to 31.
+static uvec8 kShuf21 =
+  { 5, 6, 6, 7, 8, 9, 9, 10, 10, 11, 12, 13, 13, 14, 14, 15 };
+
+// Coefficients for source bytes 0 to 10
+static uvec8 kMadd01 =
+  { 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2 };
+
+// Coefficients for source bytes 10 to 21
+static uvec8 kMadd11 =
+  { 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1 };
+
+// Coefficients for source bytes 21 to 31
+static uvec8 kMadd21 =
+  { 2, 2, 1, 3, 3, 1, 2, 2, 1, 3, 3, 1, 2, 2, 1, 3 };
+
+// Coefficients for source bytes 21 to 31
+static vec16 kRound34 =
+  { 2, 2, 2, 2, 2, 2, 2, 2 };
+
+static uvec8 kShuf38a =
+  { 0, 3, 6, 8, 11, 14, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+static uvec8 kShuf38b =
+  { 128, 128, 128, 128, 128, 128, 0, 3, 6, 8, 11, 14, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 0,1,2
+static uvec8 kShufAc =
+  { 0, 1, 6, 7, 12, 13, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128 };
+
+// Arrange words 0,3,6 into 3,4,5
+static uvec8 kShufAc3 =
+  { 128, 128, 128, 128, 128, 128, 0, 1, 6, 7, 12, 13, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x3 and 2x3
+static uvec16 kScaleAc33 =
+  { 65536 / 9, 65536 / 9, 65536 / 6, 65536 / 9, 65536 / 9, 65536 / 6, 0, 0 };
+
+// Arrange first value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb0 =
+  { 0, 128, 3, 128, 6, 128, 8, 128, 11, 128, 14, 128, 128, 128, 128, 128 };
+
+// Arrange second value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb1 =
+  { 1, 128, 4, 128, 7, 128, 9, 128, 12, 128, 15, 128, 128, 128, 128, 128 };
+
+// Arrange third value for pixels 0,1,2,3,4,5
+static uvec8 kShufAb2 =
+  { 2, 128, 5, 128, 128, 128, 10, 128, 13, 128, 128, 128, 128, 128, 128, 128 };
+
+// Scaling values for boxes of 3x2 and 2x2
+static uvec16 kScaleAb2 =
+  { 65536 / 3, 65536 / 3, 65536 / 2, 65536 / 3, 65536 / 3, 65536 / 2, 0, 0 };
+
+// Reads 32 pixels, throws half away and writes 16 pixels.
+__declspec(naked)
+void ScaleRowDown2_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    psrlw      xmm0, 8               // isolate odd pixels.
+    psrlw      xmm1, 8
+    packuswb   xmm0, xmm1
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x1 rectangle to 16x1.
+__declspec(naked)
+void ScaleRowDown2Linear_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 16x1.
+__declspec(naked)
+void ScaleRowDown2Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff00ff
+    psrlw      xmm5, 8
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm5
+    pand       xmm3, xmm5
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 16
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_SCALEROWDOWN2_AVX2
+// Reads 64 pixels, throws half away and writes 32 pixels.
+__declspec(naked)
+void ScaleRowDown2_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+
+  wloop:
+    vmovdqu     ymm0, [eax]
+    vmovdqu     ymm1, [eax + 32]
+    lea         eax,  [eax + 64]
+    vpsrlw      ymm0, ymm0, 8        // isolate odd pixels.
+    vpsrlw      ymm1, ymm1, 8
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8     // unmutate vpackuswb
+    vmovdqu     [edx], ymm0
+    lea         edx, [edx + 32]
+    sub         ecx, 32
+    jg          wloop
+
+    vzeroupper
+    ret
+  }
+}
+
+// Blends 64x1 rectangle to 32x1.
+__declspec(naked)
+void ScaleRowDown2Linear_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov         eax, [esp + 4]        // src_ptr
+                                      // src_stride
+    mov         edx, [esp + 12]       // dst_ptr
+    mov         ecx, [esp + 16]       // dst_width
+
+    vpcmpeqb    ymm4, ymm4, ymm4      // '1' constant, 8b
+    vpsrlw      ymm4, ymm4, 15
+    vpackuswb   ymm4, ymm4, ymm4
+    vpxor       ymm5, ymm5, ymm5      // constant 0
+
+  wloop:
+    vmovdqu     ymm0, [eax]
+    vmovdqu     ymm1, [eax + 32]
+    lea         eax,  [eax + 64]
+
+    vpmaddubsw  ymm0, ymm0, ymm4      // average horizontally
+    vpmaddubsw  ymm1, ymm1, ymm4
+    vpavgw      ymm0, ymm0, ymm5      // (x + 1) / 2
+    vpavgw      ymm1, ymm1, ymm5
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+
+    vmovdqu     [edx], ymm0
+    lea         edx, [edx + 32]
+    sub         ecx, 32
+    jg          wloop
+
+    vzeroupper
+    ret
+  }
+}
+
+// Blends 64x2 rectangle to 32x1.
+__declspec(naked)
+void ScaleRowDown2Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push        esi
+    mov         eax, [esp + 4 + 4]    // src_ptr
+    mov         esi, [esp + 4 + 8]    // src_stride
+    mov         edx, [esp + 4 + 12]   // dst_ptr
+    mov         ecx, [esp + 4 + 16]   // dst_width
+
+    vpcmpeqb    ymm4, ymm4, ymm4      // '1' constant, 8b
+    vpsrlw      ymm4, ymm4, 15
+    vpackuswb   ymm4, ymm4, ymm4
+    vpxor       ymm5, ymm5, ymm5      // constant 0
+
+  wloop:
+    vmovdqu     ymm0, [eax]           // average rows
+    vmovdqu     ymm1, [eax + 32]
+    vpavgb      ymm0, ymm0, [eax + esi]
+    vpavgb      ymm1, ymm1, [eax + esi + 32]
+    lea         eax,  [eax + 64]
+
+    vpmaddubsw  ymm0, ymm0, ymm4      // average horizontally
+    vpmaddubsw  ymm1, ymm1, ymm4
+    vpavgw      ymm0, ymm0, ymm5      // (x + 1) / 2
+    vpavgw      ymm1, ymm1, ymm5
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+
+    vmovdqu     [edx], ymm0
+    lea         edx, [edx + 32]
+    sub         ecx, 32
+    jg          wloop
+
+    pop         esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SCALEROWDOWN2_AVX2
+
+// Point samples 32 pixels to 8 pixels.
+__declspec(naked)
+void ScaleRowDown4_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    pcmpeqb    xmm5, xmm5            // generate mask 0x00ff0000
+    psrld      xmm5, 24
+    pslld      xmm5, 16
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    pand       xmm0, xmm5
+    pand       xmm1, xmm5
+    packuswb   xmm0, xmm1
+    psrlw      xmm0, 8
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    sub        ecx, 8
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x4 rectangle to 8x1.
+__declspec(naked)
+void ScaleRowDown4Box_SSE2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_ptr
+    mov        esi, [esp + 8 + 8]    // src_stride
+    mov        edx, [esp + 8 + 12]   // dst_ptr
+    mov        ecx, [esp + 8 + 16]   // dst_width
+    lea        edi, [esi + esi * 2]  // src_stride * 3
+    pcmpeqb    xmm7, xmm7            // generate mask 0x00ff00ff
+    psrlw      xmm7, 8
+
+  wloop:
+    movdqu     xmm0, [eax]           // average rows
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+    movdqu     xmm2, [eax + esi * 2]
+    movdqu     xmm3, [eax + esi * 2 + 16]
+    movdqu     xmm4, [eax + edi]
+    movdqu     xmm5, [eax + edi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm2, xmm4
+    pavgb      xmm3, xmm5
+    pavgb      xmm0, xmm2
+    pavgb      xmm1, xmm3
+
+    movdqa     xmm2, xmm0            // average columns (32 to 16 pixels)
+    psrlw      xmm0, 8
+    movdqa     xmm3, xmm1
+    psrlw      xmm1, 8
+    pand       xmm2, xmm7
+    pand       xmm3, xmm7
+    pavgw      xmm0, xmm2
+    pavgw      xmm1, xmm3
+    packuswb   xmm0, xmm1
+
+    movdqa     xmm2, xmm0            // average columns (16 to 8 pixels)
+    psrlw      xmm0, 8
+    pand       xmm2, xmm7
+    pavgw      xmm0, xmm2
+    packuswb   xmm0, xmm0
+
+    movq       qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    sub        ecx, 8
+    jg         wloop
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+#ifdef HAS_SCALEROWDOWN4_AVX2
+// Point samples 64 pixels to 16 pixels.
+__declspec(naked)
+void ScaleRowDown4_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                        uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov         eax, [esp + 4]        // src_ptr
+                                      // src_stride ignored
+    mov         edx, [esp + 12]       // dst_ptr
+    mov         ecx, [esp + 16]       // dst_width
+    vpcmpeqb    ymm5, ymm5, ymm5      // generate mask 0x00ff0000
+    vpsrld      ymm5, ymm5, 24
+    vpslld      ymm5, ymm5, 16
+
+  wloop:
+    vmovdqu     ymm0, [eax]
+    vmovdqu     ymm1, [eax + 32]
+    lea         eax,  [eax + 64]
+    vpand       ymm0, ymm0, ymm5
+    vpand       ymm1, ymm1, ymm5
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+    vpsrlw      ymm0, ymm0, 8
+    vpackuswb   ymm0, ymm0, ymm0
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+    vmovdqu     [edx], xmm0
+    lea         edx, [edx + 16]
+    sub         ecx, 16
+    jg          wloop
+
+    vzeroupper
+    ret
+  }
+}
+
+// Blends 64x4 rectangle to 16x1.
+__declspec(naked)
+void ScaleRowDown4Box_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst_ptr, int dst_width) {
+  __asm {
+    push        esi
+    push        edi
+    mov         eax, [esp + 8 + 4]    // src_ptr
+    mov         esi, [esp + 8 + 8]    // src_stride
+    mov         edx, [esp + 8 + 12]   // dst_ptr
+    mov         ecx, [esp + 8 + 16]   // dst_width
+    lea         edi, [esi + esi * 2]  // src_stride * 3
+    vpcmpeqb    ymm7, ymm7, ymm7      // generate mask 0x00ff00ff
+    vpsrlw      ymm7, ymm7, 8
+
+  wloop:
+    vmovdqu     ymm0, [eax]           // average rows
+    vmovdqu     ymm1, [eax + 32]
+    vpavgb      ymm0, ymm0, [eax + esi]
+    vpavgb      ymm1, ymm1, [eax + esi + 32]
+    vmovdqu     ymm2, [eax + esi * 2]
+    vmovdqu     ymm3, [eax + esi * 2 + 32]
+    vpavgb      ymm2, ymm2, [eax + edi]
+    vpavgb      ymm3, ymm3, [eax + edi + 32]
+    lea         eax, [eax + 64]
+    vpavgb      ymm0, ymm0, ymm2
+    vpavgb      ymm1, ymm1, ymm3
+
+    vpand       ymm2, ymm0, ymm7      // average columns (64 to 32 pixels)
+    vpand       ymm3, ymm1, ymm7
+    vpsrlw      ymm0, ymm0, 8
+    vpsrlw      ymm1, ymm1, 8
+    vpavgw      ymm0, ymm0, ymm2
+    vpavgw      ymm1, ymm1, ymm3
+    vpackuswb   ymm0, ymm0, ymm1
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+
+    vpand       ymm2, ymm0, ymm7      // average columns (32 to 16 pixels)
+    vpsrlw      ymm0, ymm0, 8
+    vpavgw      ymm0, ymm0, ymm2
+    vpackuswb   ymm0, ymm0, ymm0
+    vpermq      ymm0, ymm0, 0xd8      // unmutate vpackuswb
+
+    vmovdqu     [edx], xmm0
+    lea         edx, [edx + 16]
+    sub         ecx, 16
+    jg          wloop
+
+    pop        edi
+    pop        esi
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SCALEROWDOWN4_AVX2
+
+// Point samples 32 pixels to 24 pixels.
+// Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
+// Then shuffled to do the scaling.
+
+__declspec(naked)
+void ScaleRowDown34_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    movdqa     xmm3, kShuf0
+    movdqa     xmm4, kShuf1
+    movdqa     xmm5, kShuf2
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm1
+    palignr    xmm1, xmm0, 8
+    pshufb     xmm0, xmm3
+    pshufb     xmm1, xmm4
+    pshufb     xmm2, xmm5
+    movq       qword ptr [edx], xmm0
+    movq       qword ptr [edx + 8], xmm1
+    movq       qword ptr [edx + 16], xmm2
+    lea        edx, [edx + 24]
+    sub        ecx, 24
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 32x2 rectangle to 24x1
+// Produces three 8 byte values. For each 8 bytes, 16 bytes are read.
+// Then shuffled to do the scaling.
+
+// Register usage:
+// xmm0 src_row 0
+// xmm1 src_row 1
+// xmm2 shuf 0
+// xmm3 shuf 1
+// xmm4 shuf 2
+// xmm5 madd 0
+// xmm6 madd 1
+// xmm7 kRound34
+
+// Note that movdqa+palign may be better than movdqu.
+__declspec(naked)
+void ScaleRowDown34_1_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShuf01
+    movdqa     xmm3, kShuf11
+    movdqa     xmm4, kShuf21
+    movdqa     xmm5, kMadd01
+    movdqa     xmm6, kMadd11
+    movdqa     xmm7, kRound34
+
+  wloop:
+    movdqu     xmm0, [eax]           // pixels 0..7
+    movdqu     xmm1, [eax + esi]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm2
+    pmaddubsw  xmm0, xmm5
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm1, [eax + esi + 8]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm3
+    pmaddubsw  xmm0, xmm6
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 8], xmm0
+    movdqu     xmm0, [eax + 16]      // pixels 16..23
+    movdqu     xmm1, [eax + esi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm4
+    movdqa     xmm1, kMadd21
+    pmaddubsw  xmm0, xmm1
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 16], xmm0
+    lea        edx, [edx + 24]
+    sub        ecx, 24
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Note that movdqa+palign may be better than movdqu.
+__declspec(naked)
+void ScaleRowDown34_0_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShuf01
+    movdqa     xmm3, kShuf11
+    movdqa     xmm4, kShuf21
+    movdqa     xmm5, kMadd01
+    movdqa     xmm6, kMadd11
+    movdqa     xmm7, kRound34
+
+  wloop:
+    movdqu     xmm0, [eax]           // pixels 0..7
+    movdqu     xmm1, [eax + esi]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm2
+    pmaddubsw  xmm0, xmm5
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx], xmm0
+    movdqu     xmm0, [eax + 8]       // pixels 8..15
+    movdqu     xmm1, [eax + esi + 8]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm3
+    pmaddubsw  xmm0, xmm6
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 8], xmm0
+    movdqu     xmm0, [eax + 16]      // pixels 16..23
+    movdqu     xmm1, [eax + esi + 16]
+    lea        eax, [eax + 32]
+    pavgb      xmm1, xmm0
+    pavgb      xmm0, xmm1
+    pshufb     xmm0, xmm4
+    movdqa     xmm1, kMadd21
+    pmaddubsw  xmm0, xmm1
+    paddsw     xmm0, xmm7
+    psrlw      xmm0, 2
+    packuswb   xmm0, xmm0
+    movq       qword ptr [edx + 16], xmm0
+    lea        edx, [edx+24]
+    sub        ecx, 24
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// 3/8 point sampler
+
+// Scale 32 pixels to 12
+__declspec(naked)
+void ScaleRowDown38_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                          uint8* dst_ptr, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_ptr
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_ptr
+    mov        ecx, [esp + 16]       // dst_width
+    movdqa     xmm4, kShuf38a
+    movdqa     xmm5, kShuf38b
+
+  xloop:
+    movdqu     xmm0, [eax]           // 16 pixels -> 0,1,2,3,4,5
+    movdqu     xmm1, [eax + 16]      // 16 pixels -> 6,7,8,9,10,11
+    lea        eax, [eax + 32]
+    pshufb     xmm0, xmm4
+    pshufb     xmm1, xmm5
+    paddusb    xmm0, xmm1
+
+    movq       qword ptr [edx], xmm0  // write 12 pixels
+    movhlps    xmm1, xmm0
+    movd       [edx + 8], xmm1
+    lea        edx, [edx + 12]
+    sub        ecx, 12
+    jg         xloop
+
+    ret
+  }
+}
+
+// Scale 16x3 pixels to 6x1 with interpolation
+__declspec(naked)
+void ScaleRowDown38_3_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShufAc
+    movdqa     xmm3, kShufAc3
+    movdqa     xmm4, kScaleAc33
+    pxor       xmm5, xmm5
+
+  xloop:
+    movdqu     xmm0, [eax]           // sum up 3 rows into xmm0/1
+    movdqu     xmm6, [eax + esi]
+    movhlps    xmm1, xmm0
+    movhlps    xmm7, xmm6
+    punpcklbw  xmm0, xmm5
+    punpcklbw  xmm1, xmm5
+    punpcklbw  xmm6, xmm5
+    punpcklbw  xmm7, xmm5
+    paddusw    xmm0, xmm6
+    paddusw    xmm1, xmm7
+    movdqu     xmm6, [eax + esi * 2]
+    lea        eax, [eax + 16]
+    movhlps    xmm7, xmm6
+    punpcklbw  xmm6, xmm5
+    punpcklbw  xmm7, xmm5
+    paddusw    xmm0, xmm6
+    paddusw    xmm1, xmm7
+
+    movdqa     xmm6, xmm0            // 8 pixels -> 0,1,2 of xmm6
+    psrldq     xmm0, 2
+    paddusw    xmm6, xmm0
+    psrldq     xmm0, 2
+    paddusw    xmm6, xmm0
+    pshufb     xmm6, xmm2
+
+    movdqa     xmm7, xmm1            // 8 pixels -> 3,4,5 of xmm6
+    psrldq     xmm1, 2
+    paddusw    xmm7, xmm1
+    psrldq     xmm1, 2
+    paddusw    xmm7, xmm1
+    pshufb     xmm7, xmm3
+    paddusw    xmm6, xmm7
+
+    pmulhuw    xmm6, xmm4            // divide by 9,9,6, 9,9,6
+    packuswb   xmm6, xmm6
+
+    movd       [edx], xmm6           // write 6 pixels
+    psrlq      xmm6, 16
+    movd       [edx + 2], xmm6
+    lea        edx, [edx + 6]
+    sub        ecx, 6
+    jg         xloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Scale 16x2 pixels to 6x1 with interpolation
+__declspec(naked)
+void ScaleRowDown38_2_Box_SSSE3(const uint8* src_ptr,
+                                ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_ptr
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_ptr
+    mov        ecx, [esp + 4 + 16]   // dst_width
+    movdqa     xmm2, kShufAb0
+    movdqa     xmm3, kShufAb1
+    movdqa     xmm4, kShufAb2
+    movdqa     xmm5, kScaleAb2
+
+  xloop:
+    movdqu     xmm0, [eax]           // average 2 rows into xmm0
+    movdqu     xmm1, [eax + esi]
+    lea        eax, [eax + 16]
+    pavgb      xmm0, xmm1
+
+    movdqa     xmm1, xmm0            // 16 pixels -> 0,1,2,3,4,5 of xmm1
+    pshufb     xmm1, xmm2
+    movdqa     xmm6, xmm0
+    pshufb     xmm6, xmm3
+    paddusw    xmm1, xmm6
+    pshufb     xmm0, xmm4
+    paddusw    xmm1, xmm0
+
+    pmulhuw    xmm1, xmm5            // divide by 3,3,2, 3,3,2
+    packuswb   xmm1, xmm1
+
+    movd       [edx], xmm1           // write 6 pixels
+    psrlq      xmm1, 16
+    movd       [edx + 2], xmm1
+    lea        edx, [edx + 6]
+    sub        ecx, 6
+    jg         xloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 16 bytes and accumulates to 16 shorts at a time.
+__declspec(naked)
+void ScaleAddRow_SSE2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
+  __asm {
+    mov        eax, [esp + 4]   // src_ptr
+    mov        edx, [esp + 8]   // dst_ptr
+    mov        ecx, [esp + 12]  // src_width
+    pxor       xmm5, xmm5
+
+  // sum rows
+  xloop:
+    movdqu     xmm3, [eax]       // read 16 bytes
+    lea        eax, [eax + 16]
+    movdqu     xmm0, [edx]       // read 16 words from destination
+    movdqu     xmm1, [edx + 16]
+    movdqa     xmm2, xmm3
+    punpcklbw  xmm2, xmm5
+    punpckhbw  xmm3, xmm5
+    paddusw    xmm0, xmm2        // sum 16 words
+    paddusw    xmm1, xmm3
+    movdqu     [edx], xmm0       // write 16 words to destination
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 16
+    jg         xloop
+    ret
+  }
+}
+
+#ifdef HAS_SCALEADDROW_AVX2
+// Reads 32 bytes and accumulates to 32 shorts at a time.
+__declspec(naked)
+void ScaleAddRow_AVX2(const uint8* src_ptr, uint16* dst_ptr, int src_width) {
+  __asm {
+    mov         eax, [esp + 4]   // src_ptr
+    mov         edx, [esp + 8]   // dst_ptr
+    mov         ecx, [esp + 12]  // src_width
+    vpxor       ymm5, ymm5, ymm5
+
+  // sum rows
+  xloop:
+    vmovdqu     ymm3, [eax]       // read 32 bytes
+    lea         eax, [eax + 32]
+    vpermq      ymm3, ymm3, 0xd8  // unmutate for vpunpck
+    vpunpcklbw  ymm2, ymm3, ymm5
+    vpunpckhbw  ymm3, ymm3, ymm5
+    vpaddusw    ymm0, ymm2, [edx] // sum 16 words
+    vpaddusw    ymm1, ymm3, [edx + 32]
+    vmovdqu     [edx], ymm0       // write 32 words to destination
+    vmovdqu     [edx + 32], ymm1
+    lea         edx, [edx + 64]
+    sub         ecx, 32
+    jg          xloop
+
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_SCALEADDROW_AVX2
+
+// Bilinear column filtering. SSSE3 version.
+__declspec(naked)
+void ScaleFilterCols_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
+                           int dst_width, int x, int dx) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        edi, [esp + 12 + 4]    // dst_ptr
+    mov        esi, [esp + 12 + 8]    // src_ptr
+    mov        ecx, [esp + 12 + 12]   // dst_width
+    movd       xmm2, [esp + 12 + 16]  // x
+    movd       xmm3, [esp + 12 + 20]  // dx
+    mov        eax, 0x04040000      // shuffle to line up fractions with pixel.
+    movd       xmm5, eax
+    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    psrlw      xmm6, 9
+    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    sub        ecx, 2
+    jl         xloop29
+
+    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    paddd      xmm0, xmm3
+    punpckldq  xmm2, xmm0           // x0 x1
+    punpckldq  xmm3, xmm3           // dx dx
+    paddd      xmm3, xmm3           // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+
+    // 2 Pixel loop.
+  xloop2:
+    movdqa     xmm1, xmm2           // x0, x1 fractions.
+    paddd      xmm2, xmm3           // x += dx
+    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+    movd       xmm0, ebx
+    psrlw      xmm1, 9              // 7 bit fractions.
+    movzx      ebx, word ptr [esi + edx]  // 2 source x1 pixels
+    movd       xmm4, ebx
+    pshufb     xmm1, xmm5           // 0011
+    punpcklwd  xmm0, xmm4
+    pxor       xmm1, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1           // 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    psrlw      xmm0, 7              // 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // 8 bits, 2 pixels.
+    movd       ebx, xmm0
+    mov        [edi], bx
+    lea        edi, [edi + 2]
+    sub        ecx, 2               // 2 pixels
+    jge        xloop2
+
+ xloop29:
+
+    add        ecx, 2 - 1
+    jl         xloop99
+
+    // 1 pixel remainder
+    movzx      ebx, word ptr [esi + eax]  // 2 source x0 pixels
+    movd       xmm0, ebx
+    psrlw      xmm2, 9              // 7 bit fractions.
+    pshufb     xmm2, xmm5           // 0011
+    pxor       xmm2, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2           // 16 bit
+    psrlw      xmm0, 7              // 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // 8 bits
+    movd       ebx, xmm0
+    mov        [edi], bl
+
+ xloop99:
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// Reads 16 pixels, duplicates them and writes 32 pixels.
+__declspec(naked)
+void ScaleColsUp2_SSE2(uint8* dst_ptr, const uint8* src_ptr,
+                       int dst_width, int x, int dx) {
+  __asm {
+    mov        edx, [esp + 4]    // dst_ptr
+    mov        eax, [esp + 8]    // src_ptr
+    mov        ecx, [esp + 12]   // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm1, xmm0
+    punpcklbw  xmm0, xmm0
+    punpckhbw  xmm1, xmm1
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 32
+    jg         wloop
+
+    ret
+  }
+}
+
+// Reads 8 pixels, throws half away and writes 4 even pixels (0, 2, 4, 6)
+__declspec(naked)
+void ScaleARGBRowDown2_SSE2(const uint8* src_argb,
+                            ptrdiff_t src_stride,
+                            uint8* dst_argb, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_argb
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_argb
+    mov        ecx, [esp + 16]       // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    shufps     xmm0, xmm1, 0xdd
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 8x1 rectangle to 4x1.
+__declspec(naked)
+void ScaleARGBRowDown2Linear_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  uint8* dst_argb, int dst_width) {
+  __asm {
+    mov        eax, [esp + 4]        // src_argb
+                                     // src_stride ignored
+    mov        edx, [esp + 12]       // dst_argb
+    mov        ecx, [esp + 16]       // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    lea        eax,  [eax + 32]
+    movdqa     xmm2, xmm0
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         wloop
+
+    ret
+  }
+}
+
+// Blends 8x2 rectangle to 4x1.
+__declspec(naked)
+void ScaleARGBRowDown2Box_SSE2(const uint8* src_argb,
+                               ptrdiff_t src_stride,
+                               uint8* dst_argb, int dst_width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]    // src_argb
+    mov        esi, [esp + 4 + 8]    // src_stride
+    mov        edx, [esp + 4 + 12]   // dst_argb
+    mov        ecx, [esp + 4 + 16]   // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    movdqu     xmm1, [eax + 16]
+    movdqu     xmm2, [eax + esi]
+    movdqu     xmm3, [eax + esi + 16]
+    lea        eax,  [eax + 32]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         wloop
+
+    pop        esi
+    ret
+  }
+}
+
+// Reads 4 pixels at a time.
+__declspec(naked)
+void ScaleARGBRowDownEven_SSE2(const uint8* src_argb, ptrdiff_t src_stride,
+                               int src_stepx,
+                               uint8* dst_argb, int dst_width) {
+  __asm {
+    push       ebx
+    push       edi
+    mov        eax, [esp + 8 + 4]    // src_argb
+                                     // src_stride ignored
+    mov        ebx, [esp + 8 + 12]   // src_stepx
+    mov        edx, [esp + 8 + 16]   // dst_argb
+    mov        ecx, [esp + 8 + 20]   // dst_width
+    lea        ebx, [ebx * 4]
+    lea        edi, [ebx + ebx * 2]
+
+  wloop:
+    movd       xmm0, [eax]
+    movd       xmm1, [eax + ebx]
+    punpckldq  xmm0, xmm1
+    movd       xmm2, [eax + ebx * 2]
+    movd       xmm3, [eax + edi]
+    lea        eax,  [eax + ebx * 4]
+    punpckldq  xmm2, xmm3
+    punpcklqdq xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         wloop
+
+    pop        edi
+    pop        ebx
+    ret
+  }
+}
+
+// Blends four 2x2 to 4x1.
+__declspec(naked)
+void ScaleARGBRowDownEvenBox_SSE2(const uint8* src_argb,
+                                  ptrdiff_t src_stride,
+                                  int src_stepx,
+                                  uint8* dst_argb, int dst_width) {
+  __asm {
+    push       ebx
+    push       esi
+    push       edi
+    mov        eax, [esp + 12 + 4]    // src_argb
+    mov        esi, [esp + 12 + 8]    // src_stride
+    mov        ebx, [esp + 12 + 12]   // src_stepx
+    mov        edx, [esp + 12 + 16]   // dst_argb
+    mov        ecx, [esp + 12 + 20]   // dst_width
+    lea        esi, [eax + esi]       // row1 pointer
+    lea        ebx, [ebx * 4]
+    lea        edi, [ebx + ebx * 2]
+
+  wloop:
+    movq       xmm0, qword ptr [eax]  // row0 4 pairs
+    movhps     xmm0, qword ptr [eax + ebx]
+    movq       xmm1, qword ptr [eax + ebx * 2]
+    movhps     xmm1, qword ptr [eax + edi]
+    lea        eax,  [eax + ebx * 4]
+    movq       xmm2, qword ptr [esi]  // row1 4 pairs
+    movhps     xmm2, qword ptr [esi + ebx]
+    movq       xmm3, qword ptr [esi + ebx * 2]
+    movhps     xmm3, qword ptr [esi + edi]
+    lea        esi,  [esi + ebx * 4]
+    pavgb      xmm0, xmm2            // average rows
+    pavgb      xmm1, xmm3
+    movdqa     xmm2, xmm0            // average columns (8 to 4 pixels)
+    shufps     xmm0, xmm1, 0x88      // even pixels
+    shufps     xmm2, xmm1, 0xdd      // odd pixels
+    pavgb      xmm0, xmm2
+    movdqu     [edx], xmm0
+    lea        edx, [edx + 16]
+    sub        ecx, 4
+    jg         wloop
+
+    pop        edi
+    pop        esi
+    pop        ebx
+    ret
+  }
+}
+
+// Column scaling unfiltered. SSE2 version.
+__declspec(naked)
+void ScaleARGBCols_SSE2(uint8* dst_argb, const uint8* src_argb,
+                        int dst_width, int x, int dx) {
+  __asm {
+    push       edi
+    push       esi
+    mov        edi, [esp + 8 + 4]    // dst_argb
+    mov        esi, [esp + 8 + 8]    // src_argb
+    mov        ecx, [esp + 8 + 12]   // dst_width
+    movd       xmm2, [esp + 8 + 16]  // x
+    movd       xmm3, [esp + 8 + 20]  // dx
+
+    pshufd     xmm2, xmm2, 0         // x0 x0 x0 x0
+    pshufd     xmm0, xmm3, 0x11      // dx  0 dx  0
+    paddd      xmm2, xmm0
+    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 2
+    pshufd     xmm0, xmm3, 0x05      // dx * 2, dx * 2, 0, 0
+    paddd      xmm2, xmm0            // x3 x2 x1 x0
+    paddd      xmm3, xmm3            // 0, 0, 0,  dx * 4
+    pshufd     xmm3, xmm3, 0         // dx * 4, dx * 4, dx * 4, dx * 4
+
+    pextrw     eax, xmm2, 1          // get x0 integer.
+    pextrw     edx, xmm2, 3          // get x1 integer.
+
+    cmp        ecx, 0
+    jle        xloop99
+    sub        ecx, 4
+    jl         xloop49
+
+    // 4 Pixel loop.
+ xloop4:
+    movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
+    movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
+    pextrw     eax, xmm2, 5           // get x2 integer.
+    pextrw     edx, xmm2, 7           // get x3 integer.
+    paddd      xmm2, xmm3             // x += dx
+    punpckldq  xmm0, xmm1             // x0 x1
+
+    movd       xmm1, [esi + eax * 4]  // 1 source x2 pixels
+    movd       xmm4, [esi + edx * 4]  // 1 source x3 pixels
+    pextrw     eax, xmm2, 1           // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3           // get x1 integer. next iteration.
+    punpckldq  xmm1, xmm4             // x2 x3
+    punpcklqdq xmm0, xmm1             // x0 x1 x2 x3
+    movdqu     [edi], xmm0
+    lea        edi, [edi + 16]
+    sub        ecx, 4                 // 4 pixels
+    jge        xloop4
+
+ xloop49:
+    test       ecx, 2
+    je         xloop29
+
+    // 2 Pixels.
+    movd       xmm0, [esi + eax * 4]  // 1 source x0 pixels
+    movd       xmm1, [esi + edx * 4]  // 1 source x1 pixels
+    pextrw     eax, xmm2, 5           // get x2 integer.
+    punpckldq  xmm0, xmm1             // x0 x1
+
+    movq       qword ptr [edi], xmm0
+    lea        edi, [edi + 8]
+
+ xloop29:
+    test       ecx, 1
+    je         xloop99
+
+    // 1 Pixels.
+    movd       xmm0, [esi + eax * 4]  // 1 source x2 pixels
+    movd       dword ptr [edi], xmm0
+ xloop99:
+
+    pop        esi
+    pop        edi
+    ret
+  }
+}
+
+// Bilinear row filtering combines 2x1 -> 1x1. SSSE3 version.
+// TODO(fbarchard): Port to Neon
+
+// Shuffle table for arranging 2 pixels into pairs for pmaddubsw
+static uvec8 kShuffleColARGB = {
+  0u, 4u, 1u, 5u, 2u, 6u, 3u, 7u,  // bbggrraa 1st pixel
+  8u, 12u, 9u, 13u, 10u, 14u, 11u, 15u  // bbggrraa 2nd pixel
+};
+
+// Shuffle table for duplicating 2 fractions into 8 bytes each
+static uvec8 kShuffleFractions = {
+  0u, 0u, 0u, 0u, 0u, 0u, 0u, 0u, 4u, 4u, 4u, 4u, 4u, 4u, 4u, 4u,
+};
+
+__declspec(naked)
+void ScaleARGBFilterCols_SSSE3(uint8* dst_argb, const uint8* src_argb,
+                               int dst_width, int x, int dx) {
+  __asm {
+    push       esi
+    push       edi
+    mov        edi, [esp + 8 + 4]    // dst_argb
+    mov        esi, [esp + 8 + 8]    // src_argb
+    mov        ecx, [esp + 8 + 12]   // dst_width
+    movd       xmm2, [esp + 8 + 16]  // x
+    movd       xmm3, [esp + 8 + 20]  // dx
+    movdqa     xmm4, kShuffleColARGB
+    movdqa     xmm5, kShuffleFractions
+    pcmpeqb    xmm6, xmm6           // generate 0x007f for inverting fraction.
+    psrlw      xmm6, 9
+    pextrw     eax, xmm2, 1         // get x0 integer. preroll
+    sub        ecx, 2
+    jl         xloop29
+
+    movdqa     xmm0, xmm2           // x1 = x0 + dx
+    paddd      xmm0, xmm3
+    punpckldq  xmm2, xmm0           // x0 x1
+    punpckldq  xmm3, xmm3           // dx dx
+    paddd      xmm3, xmm3           // dx * 2, dx * 2
+    pextrw     edx, xmm2, 3         // get x1 integer. preroll
+
+    // 2 Pixel loop.
+  xloop2:
+    movdqa     xmm1, xmm2           // x0, x1 fractions.
+    paddd      xmm2, xmm3           // x += dx
+    movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
+    psrlw      xmm1, 9              // 7 bit fractions.
+    movhps     xmm0, qword ptr [esi + edx * 4]  // 2 source x1 pixels
+    pshufb     xmm1, xmm5           // 0000000011111111
+    pshufb     xmm0, xmm4           // arrange pixels into pairs
+    pxor       xmm1, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm1           // argb_argb 16 bit, 2 pixels.
+    pextrw     eax, xmm2, 1         // get x0 integer. next iteration.
+    pextrw     edx, xmm2, 3         // get x1 integer. next iteration.
+    psrlw      xmm0, 7              // argb 8.7 fixed point to low 8 bits.
+    packuswb   xmm0, xmm0           // argb_argb 8 bits, 2 pixels.
+    movq       qword ptr [edi], xmm0
+    lea        edi, [edi + 8]
+    sub        ecx, 2               // 2 pixels
+    jge        xloop2
+
+ xloop29:
+
+    add        ecx, 2 - 1
+    jl         xloop99
+
+    // 1 pixel remainder
+    psrlw      xmm2, 9              // 7 bit fractions.
+    movq       xmm0, qword ptr [esi + eax * 4]  // 2 source x0 pixels
+    pshufb     xmm2, xmm5           // 00000000
+    pshufb     xmm0, xmm4           // arrange pixels into pairs
+    pxor       xmm2, xmm6           // 0..7f and 7f..0
+    pmaddubsw  xmm0, xmm2           // argb 16 bit, 1 pixel.
+    psrlw      xmm0, 7
+    packuswb   xmm0, xmm0           // argb 8 bits, 1 pixel.
+    movd       [edi], xmm0
+
+ xloop99:
+
+    pop        edi
+    pop        esi
+    ret
+  }
+}
+
+// Reads 4 pixels, duplicates them and writes 8 pixels.
+__declspec(naked)
+void ScaleARGBColsUp2_SSE2(uint8* dst_argb, const uint8* src_argb,
+                           int dst_width, int x, int dx) {
+  __asm {
+    mov        edx, [esp + 4]    // dst_argb
+    mov        eax, [esp + 8]    // src_argb
+    mov        ecx, [esp + 12]   // dst_width
+
+  wloop:
+    movdqu     xmm0, [eax]
+    lea        eax,  [eax + 16]
+    movdqa     xmm1, xmm0
+    punpckldq  xmm0, xmm0
+    punpckhdq  xmm1, xmm1
+    movdqu     [edx], xmm0
+    movdqu     [edx + 16], xmm1
+    lea        edx, [edx + 32]
+    sub        ecx, 8
+    jg         wloop
+
+    ret
+  }
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+__declspec(naked)
+int FixedDiv_X86(int num, int div) {
+  __asm {
+    mov        eax, [esp + 4]    // num
+    cdq                          // extend num to 64 bits
+    shld       edx, eax, 16      // 32.16
+    shl        eax, 16
+    idiv       dword ptr [esp + 8]
+    ret
+  }
+}
+
+// Divide num by div and return as 16.16 fixed point result.
+__declspec(naked)
+int FixedDiv1_X86(int num, int div) {
+  __asm {
+    mov        eax, [esp + 4]    // num
+    mov        ecx, [esp + 8]    // denom
+    cdq                          // extend num to 64 bits
+    shld       edx, eax, 16      // 32.16
+    shl        eax, 16
+    sub        eax, 0x00010001
+    sbb        edx, 0
+    sub        ecx, 1
+    idiv       ecx
+    ret
+  }
+}
+#endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86)
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
diff --git a/third_party/aom/third_party/libyuv/source/video_common.cc b/third_party/aom/third_party/libyuv/source/video_common.cc
new file mode 100644
index 0000000000..379a0669ae
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/video_common.cc
@@ -0,0 +1,64 @@
+/*
+ *  Copyright 2011 The LibYuv Project Authors. All rights reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS. All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+
+#include "libyuv/video_common.h"
+
+#ifdef __cplusplus
+namespace libyuv {
+extern "C" {
+#endif
+
+#define ARRAY_SIZE(x) (int)(sizeof(x) / sizeof(x[0]))
+
+struct FourCCAliasEntry {
+  uint32 alias;
+  uint32 canonical;
+};
+
+static const struct FourCCAliasEntry kFourCCAliases[] = {
+  {FOURCC_IYUV, FOURCC_I420},
+  {FOURCC_YU16, FOURCC_I422},
+  {FOURCC_YU24, FOURCC_I444},
+  {FOURCC_YUYV, FOURCC_YUY2},
+  {FOURCC_YUVS, FOURCC_YUY2},  // kCMPixelFormat_422YpCbCr8_yuvs
+  {FOURCC_HDYC, FOURCC_UYVY},
+  {FOURCC_2VUY, FOURCC_UYVY},  // kCMPixelFormat_422YpCbCr8
+  {FOURCC_JPEG, FOURCC_MJPG},  // Note: JPEG has DHT while MJPG does not.
+  {FOURCC_DMB1, FOURCC_MJPG},
+  {FOURCC_BA81, FOURCC_BGGR},  // deprecated.
+  {FOURCC_RGB3, FOURCC_RAW },
+  {FOURCC_BGR3, FOURCC_24BG},
+  {FOURCC_CM32, FOURCC_BGRA},  // kCMPixelFormat_32ARGB
+  {FOURCC_CM24, FOURCC_RAW },  // kCMPixelFormat_24RGB
+  {FOURCC_L555, FOURCC_RGBO},  // kCMPixelFormat_16LE555
+  {FOURCC_L565, FOURCC_RGBP},  // kCMPixelFormat_16LE565
+  {FOURCC_5551, FOURCC_RGBO},  // kCMPixelFormat_16LE5551
+};
+// TODO(fbarchard): Consider mapping kCMPixelFormat_32BGRA to FOURCC_ARGB.
+//  {FOURCC_BGRA, FOURCC_ARGB},  // kCMPixelFormat_32BGRA
+
+LIBYUV_API
+uint32 CanonicalFourCC(uint32 fourcc) {
+  int i;
+  for (i = 0; i < ARRAY_SIZE(kFourCCAliases); ++i) {
+    if (kFourCCAliases[i].alias == fourcc) {
+      return kFourCCAliases[i].canonical;
+    }
+  }
+  // Not an alias, so return it as-is.
+  return fourcc;
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+}  // namespace libyuv
+#endif
+
diff --git a/third_party/aom/third_party/libyuv/source/x86inc.asm b/third_party/aom/third_party/libyuv/source/x86inc.asm
new file mode 100644
index 0000000000..cb5c32df3a
--- /dev/null
+++ b/third_party/aom/third_party/libyuv/source/x86inc.asm
@@ -0,0 +1,1136 @@
+;*****************************************************************************
+;* x86inc.asm: x264asm abstraction layer
+;*****************************************************************************
+;* Copyright (C) 2005-2012 x264 project
+;*
+;* Authors: Loren Merritt <lorenm@u.washington.edu>
+;*          Anton Mitrofanov <BugMaster@narod.ru>
+;*          Jason Garrett-Glaser <darkshikari@gmail.com>
+;*          Henrik Gramner <hengar-6@student.ltu.se>
+;*
+;* Permission to use, copy, modify, and/or distribute this software for any
+;* purpose with or without fee is hereby granted, provided that the above
+;* copyright notice and this permission notice appear in all copies.
+;*
+;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+;*****************************************************************************
+
+; This is a header file for the x264ASM assembly language, which uses
+; NASM/YASM syntax combined with a large number of macros to provide easy
+; abstraction between different calling conventions (x86_32, win64, linux64).
+; It also has various other useful features to simplify writing the kind of
+; DSP functions that are most often used in x264.
+
+; Unlike the rest of x264, this file is available under an ISC license, as it
+; has significant usefulness outside of x264 and we want it to be available
+; to the largest audience possible.  Of course, if you modify it for your own
+; purposes to add a new feature, we strongly encourage contributing a patch
+; as this feature might be useful for others as well.  Send patches or ideas
+; to x264-devel@videolan.org .
+
+; Local changes for libyuv:
+; remove %define program_name and references in labels
+; rename cpus to uppercase
+
+%define WIN64  0
+%define UNIX64 0
+%if ARCH_X86_64
+    %ifidn __OUTPUT_FORMAT__,win32
+        %define WIN64  1
+    %elifidn __OUTPUT_FORMAT__,win64
+        %define WIN64  1
+    %else
+        %define UNIX64 1
+    %endif
+%endif
+
+%ifdef PREFIX
+    %define mangle(x) _ %+ x
+%else
+    %define mangle(x) x
+%endif
+
+; Name of the .rodata section.
+; Kludge: Something on OS X fails to align .rodata even given an align attribute,
+; so use a different read-only section.
+%macro SECTION_RODATA 0-1 16
+    %ifidn __OUTPUT_FORMAT__,macho64
+        SECTION .text align=%1
+    %elifidn __OUTPUT_FORMAT__,macho
+        SECTION .text align=%1
+        fakegot:
+    %elifidn __OUTPUT_FORMAT__,aout
+        section .text
+    %else
+        SECTION .rodata align=%1
+    %endif
+%endmacro
+
+; aout does not support align=
+%macro SECTION_TEXT 0-1 16
+    %ifidn __OUTPUT_FORMAT__,aout
+        SECTION .text
+    %else
+        SECTION .text align=%1
+    %endif
+%endmacro
+
+%if WIN64
+    %define PIC
+%elif ARCH_X86_64 == 0
+; x86_32 doesn't require PIC.
+; Some distros prefer shared objects to be PIC, but nothing breaks if
+; the code contains a few textrels, so we'll skip that complexity.
+    %undef PIC
+%endif
+%ifdef PIC
+    default rel
+%endif
+
+; Always use long nops (reduces 0x90 spam in disassembly on x86_32)
+CPU amdnop
+
+; Macros to eliminate most code duplication between x86_32 and x86_64:
+; Currently this works only for leaf functions which load all their arguments
+; into registers at the start, and make no other use of the stack. Luckily that
+; covers most of x264's asm.
+
+; PROLOGUE:
+; %1 = number of arguments. loads them from stack if needed.
+; %2 = number of registers used. pushes callee-saved regs if needed.
+; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
+; %4 = list of names to define to registers
+; PROLOGUE can also be invoked by adding the same options to cglobal
+
+; e.g.
+; cglobal foo, 2,3,0, dst, src, tmp
+; declares a function (foo), taking two args (dst and src) and one local variable (tmp)
+
+; TODO Some functions can use some args directly from the stack. If they're the
+; last args then you can just not declare them, but if they're in the middle
+; we need more flexible macro.
+
+; RET:
+; Pops anything that was pushed by PROLOGUE, and returns.
+
+; REP_RET:
+; Same, but if it doesn't pop anything it becomes a 2-byte ret, for athlons
+; which are slow when a normal ret follows a branch.
+
+; registers:
+; rN and rNq are the native-size register holding function argument N
+; rNd, rNw, rNb are dword, word, and byte size
+; rNh is the high 8 bits of the word size
+; rNm is the original location of arg N (a register or on the stack), dword
+; rNmp is native size
+
+%macro DECLARE_REG 2-3
+    %define r%1q %2
+    %define r%1d %2d
+    %define r%1w %2w
+    %define r%1b %2b
+    %define r%1h %2h
+    %if %0 == 2
+        %define r%1m  %2d
+        %define r%1mp %2
+    %elif ARCH_X86_64 ; memory
+        %define r%1m [rsp + stack_offset + %3]
+        %define r%1mp qword r %+ %1m
+    %else
+        %define r%1m [esp + stack_offset + %3]
+        %define r%1mp dword r %+ %1m
+    %endif
+    %define r%1  %2
+%endmacro
+
+%macro DECLARE_REG_SIZE 3
+    %define r%1q r%1
+    %define e%1q r%1
+    %define r%1d e%1
+    %define e%1d e%1
+    %define r%1w %1
+    %define e%1w %1
+    %define r%1h %3
+    %define e%1h %3
+    %define r%1b %2
+    %define e%1b %2
+%if ARCH_X86_64 == 0
+    %define r%1  e%1
+%endif
+%endmacro
+
+DECLARE_REG_SIZE ax, al, ah
+DECLARE_REG_SIZE bx, bl, bh
+DECLARE_REG_SIZE cx, cl, ch
+DECLARE_REG_SIZE dx, dl, dh
+DECLARE_REG_SIZE si, sil, null
+DECLARE_REG_SIZE di, dil, null
+DECLARE_REG_SIZE bp, bpl, null
+
+; t# defines for when per-arch register allocation is more complex than just function arguments
+
+%macro DECLARE_REG_TMP 1-*
+    %assign %%i 0
+    %rep %0
+        CAT_XDEFINE t, %%i, r%1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro DECLARE_REG_TMP_SIZE 0-*
+    %rep %0
+        %define t%1q t%1 %+ q
+        %define t%1d t%1 %+ d
+        %define t%1w t%1 %+ w
+        %define t%1h t%1 %+ h
+        %define t%1b t%1 %+ b
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
+
+%if ARCH_X86_64
+    %define gprsize 8
+%else
+    %define gprsize 4
+%endif
+
+%macro PUSH 1
+    push %1
+    %assign stack_offset stack_offset+gprsize
+%endmacro
+
+%macro POP 1
+    pop %1
+    %assign stack_offset stack_offset-gprsize
+%endmacro
+
+%macro PUSH_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            PUSH r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro POP_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            pop r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro LOAD_IF_USED 1-*
+    %rep %0
+        %if %1 < num_args
+            mov r%1, r %+ %1 %+ mp
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro SUB 2
+    sub %1, %2
+    %ifidn %1, rsp
+        %assign stack_offset stack_offset+(%2)
+    %endif
+%endmacro
+
+%macro ADD 2
+    add %1, %2
+    %ifidn %1, rsp
+        %assign stack_offset stack_offset-(%2)
+    %endif
+%endmacro
+
+%macro movifnidn 2
+    %ifnidn %1, %2
+        mov %1, %2
+    %endif
+%endmacro
+
+%macro movsxdifnidn 2
+    %ifnidn %1, %2
+        movsxd %1, %2
+    %endif
+%endmacro
+
+%macro ASSERT 1
+    %if (%1) == 0
+        %error assert failed
+    %endif
+%endmacro
+
+%macro DEFINE_ARGS 0-*
+    %ifdef n_arg_names
+        %assign %%i 0
+        %rep n_arg_names
+            CAT_UNDEF arg_name %+ %%i, q
+            CAT_UNDEF arg_name %+ %%i, d
+            CAT_UNDEF arg_name %+ %%i, w
+            CAT_UNDEF arg_name %+ %%i, h
+            CAT_UNDEF arg_name %+ %%i, b
+            CAT_UNDEF arg_name %+ %%i, m
+            CAT_UNDEF arg_name %+ %%i, mp
+            CAT_UNDEF arg_name, %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+
+    %xdefine %%stack_offset stack_offset
+    %undef stack_offset ; so that the current value of stack_offset doesn't get baked in by xdefine
+    %assign %%i 0
+    %rep %0
+        %xdefine %1q r %+ %%i %+ q
+        %xdefine %1d r %+ %%i %+ d
+        %xdefine %1w r %+ %%i %+ w
+        %xdefine %1h r %+ %%i %+ h
+        %xdefine %1b r %+ %%i %+ b
+        %xdefine %1m r %+ %%i %+ m
+        %xdefine %1mp r %+ %%i %+ mp
+        CAT_XDEFINE arg_name, %%i, %1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+    %xdefine stack_offset %%stack_offset
+    %assign n_arg_names %0
+%endmacro
+
+%if WIN64 ; Windows x64 ;=================================================
+
+DECLARE_REG 0,  rcx
+DECLARE_REG 1,  rdx
+DECLARE_REG 2,  R8
+DECLARE_REG 3,  R9
+DECLARE_REG 4,  R10, 40
+DECLARE_REG 5,  R11, 48
+DECLARE_REG 6,  rax, 56
+DECLARE_REG 7,  rdi, 64
+DECLARE_REG 8,  rsi, 72
+DECLARE_REG 9,  rbx, 80
+DECLARE_REG 10, rbp, 88
+DECLARE_REG 11, R12, 96
+DECLARE_REG 12, R13, 104
+DECLARE_REG 13, R14, 112
+DECLARE_REG 14, R15, 120
+
+%macro PROLOGUE 2-4+ 0 ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 7, 8, 9, 10, 11, 12, 13, 14
+    %if mmsize == 8
+        %assign xmm_regs_used 0
+    %else
+        WIN64_SPILL_XMM %3
+    %endif
+    LOAD_IF_USED 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS %4
+%endmacro
+
+%macro WIN64_SPILL_XMM 1
+    %assign xmm_regs_used %1
+    ASSERT xmm_regs_used <= 16
+    %if xmm_regs_used > 6
+        SUB rsp, (xmm_regs_used-6)*16+16
+        %assign %%i xmm_regs_used
+        %rep (xmm_regs_used-6)
+            %assign %%i %%i-1
+            movdqa [rsp + (%%i-6)*16+(~stack_offset&8)], xmm %+ %%i
+        %endrep
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM_INTERNAL 1
+    %if xmm_regs_used > 6
+        %assign %%i xmm_regs_used
+        %rep (xmm_regs_used-6)
+            %assign %%i %%i-1
+            movdqa xmm %+ %%i, [%1 + (%%i-6)*16+(~stack_offset&8)]
+        %endrep
+        add %1, (xmm_regs_used-6)*16+16
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM 1
+    WIN64_RESTORE_XMM_INTERNAL %1
+    %assign stack_offset stack_offset-(xmm_regs_used-6)*16+16
+    %assign xmm_regs_used 0
+%endmacro
+
+%define has_epilogue regs_used > 7 || xmm_regs_used > 6 || mmsize == 32
+
+%macro RET 0
+    WIN64_RESTORE_XMM_INTERNAL rsp
+    POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%elif ARCH_X86_64 ; *nix x64 ;=============================================
+
+DECLARE_REG 0,  rdi
+DECLARE_REG 1,  rsi
+DECLARE_REG 2,  rdx
+DECLARE_REG 3,  rcx
+DECLARE_REG 4,  R8
+DECLARE_REG 5,  R9
+DECLARE_REG 6,  rax, 8
+DECLARE_REG 7,  R10, 16
+DECLARE_REG 8,  R11, 24
+DECLARE_REG 9,  rbx, 32
+DECLARE_REG 10, rbp, 40
+DECLARE_REG 11, R12, 48
+DECLARE_REG 12, R13, 56
+DECLARE_REG 13, R14, 64
+DECLARE_REG 14, R15, 72
+
+%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 9, 10, 11, 12, 13, 14
+    LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS %4
+%endmacro
+
+%define has_epilogue regs_used > 9 || mmsize == 32
+
+%macro RET 0
+    POP_IF_USED 14, 13, 12, 11, 10, 9
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%else ; X86_32 ;==============================================================
+
+DECLARE_REG 0, eax, 4
+DECLARE_REG 1, ecx, 8
+DECLARE_REG 2, edx, 12
+DECLARE_REG 3, ebx, 16
+DECLARE_REG 4, esi, 20
+DECLARE_REG 5, edi, 24
+DECLARE_REG 6, ebp, 28
+%define rsp esp
+
+%macro DECLARE_ARG 1-*
+    %rep %0
+        %define r%1m [esp + stack_offset + 4*%1 + 4]
+        %define r%1mp dword r%1m
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
+
+%macro PROLOGUE 2-4+ ; #args, #regs, #xmm_regs, arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    %if regs_used > 7
+        %assign regs_used 7
+    %endif
+    ASSERT regs_used >= num_args
+    PUSH_IF_USED 3, 4, 5, 6
+    LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6
+    DEFINE_ARGS %4
+%endmacro
+
+%define has_epilogue regs_used > 3 || mmsize == 32
+
+%macro RET 0
+    POP_IF_USED 6, 5, 4, 3
+%if mmsize == 32
+    vzeroupper
+%endif
+    ret
+%endmacro
+
+%endif ;======================================================================
+
+%if WIN64 == 0
+%macro WIN64_SPILL_XMM 1
+%endmacro
+%macro WIN64_RESTORE_XMM 1
+%endmacro
+%endif
+
+%macro REP_RET 0
+    %if has_epilogue
+        RET
+    %else
+        rep ret
+    %endif
+%endmacro
+
+%macro TAIL_CALL 2 ; callee, is_nonadjacent
+    %if has_epilogue
+        call %1
+        RET
+    %elif %2
+        jmp %1
+    %endif
+%endmacro
+
+;=============================================================================
+; arch-independent part
+;=============================================================================
+
+%assign function_align 16
+
+; Begin a function.
+; Applies any symbol mangling needed for C linkage, and sets up a define such that
+; subsequent uses of the function name automatically refer to the mangled version.
+; Appends cpuflags to the function name if cpuflags has been specified.
+%macro cglobal 1-2+ ; name, [PROLOGUE args]
+%if %0 == 1
+    cglobal_internal %1 %+ SUFFIX
+%else
+    cglobal_internal %1 %+ SUFFIX, %2
+%endif
+%endmacro
+%macro cglobal_internal 1-2+
+    %ifndef cglobaled_%1
+        %xdefine %1 mangle(%1)
+        %xdefine %1.skip_prologue %1 %+ .skip_prologue
+        CAT_XDEFINE cglobaled_, %1, 1
+    %endif
+    %xdefine current_function %1
+    %ifidn __OUTPUT_FORMAT__,elf
+        global %1:function hidden
+    %else
+        global %1
+    %endif
+    align function_align
+    %1:
+    RESET_MM_PERMUTATION ; not really needed, but makes disassembly somewhat nicer
+    %assign stack_offset 0
+    %if %0 > 1
+        PROLOGUE %2
+    %endif
+%endmacro
+
+%macro cextern 1
+    %xdefine %1 mangle(%1)
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+; like cextern, but without the prefix
+%macro cextern_naked 1
+    %xdefine %1 mangle(%1)
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+%macro const 2+
+    %xdefine %1 mangle(%1)
+    global %1
+    %1: %2
+%endmacro
+
+; This is needed for ELF, otherwise the GNU linker assumes the stack is
+; executable by default.
+%ifidn __OUTPUT_FORMAT__,elf
+SECTION .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+%ifidn __OUTPUT_FORMAT__,elf32
+section .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+%ifidn __OUTPUT_FORMAT__,elf64
+section .note.GNU-stack noalloc noexec nowrite progbits
+%endif
+
+; cpuflags
+
+%assign cpuflags_MMX      (1<<0)
+%assign cpuflags_MMX2     (1<<1) | cpuflags_MMX
+%assign cpuflags_3dnow    (1<<2) | cpuflags_MMX
+%assign cpuflags_3dnow2   (1<<3) | cpuflags_3dnow
+%assign cpuflags_SSE      (1<<4) | cpuflags_MMX2
+%assign cpuflags_SSE2     (1<<5) | cpuflags_SSE
+%assign cpuflags_SSE2slow (1<<6) | cpuflags_SSE2
+%assign cpuflags_SSE3     (1<<7) | cpuflags_SSE2
+%assign cpuflags_SSSE3    (1<<8) | cpuflags_SSE3
+%assign cpuflags_SSE4     (1<<9) | cpuflags_SSSE3
+%assign cpuflags_SSE42    (1<<10)| cpuflags_SSE4
+%assign cpuflags_AVX      (1<<11)| cpuflags_SSE42
+%assign cpuflags_xop      (1<<12)| cpuflags_AVX
+%assign cpuflags_fma4     (1<<13)| cpuflags_AVX
+%assign cpuflags_AVX2     (1<<14)| cpuflags_AVX
+%assign cpuflags_fma3     (1<<15)| cpuflags_AVX
+
+%assign cpuflags_cache32  (1<<16)
+%assign cpuflags_cache64  (1<<17)
+%assign cpuflags_slowctz  (1<<18)
+%assign cpuflags_lzcnt    (1<<19)
+%assign cpuflags_misalign (1<<20)
+%assign cpuflags_aligned  (1<<21) ; not a cpu feature, but a function variant
+%assign cpuflags_atom     (1<<22)
+%assign cpuflags_bmi1     (1<<23)
+%assign cpuflags_bmi2     (1<<24)|cpuflags_bmi1
+%assign cpuflags_tbm      (1<<25)|cpuflags_bmi1
+
+%define    cpuflag(x) ((cpuflags & (cpuflags_ %+ x)) == (cpuflags_ %+ x))
+%define notcpuflag(x) ((cpuflags & (cpuflags_ %+ x)) != (cpuflags_ %+ x))
+
+; Takes up to 2 cpuflags from the above list.
+; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu.
+; You shouldn't need to invoke this macro directly, it's a subroutine for INIT_MMX &co.
+%macro INIT_CPUFLAGS 0-2
+    %if %0 >= 1
+        %xdefine cpuname %1
+        %assign cpuflags cpuflags_%1
+        %if %0 >= 2
+            %xdefine cpuname %1_%2
+            %assign cpuflags cpuflags | cpuflags_%2
+        %endif
+        %xdefine SUFFIX _ %+ cpuname
+        %if cpuflag(AVX)
+            %assign AVX_enabled 1
+        %endif
+        %if mmsize == 16 && notcpuflag(SSE2)
+            %define mova movaps
+            %define movu movups
+            %define movnta movntps
+        %endif
+        %if cpuflag(aligned)
+            %define movu mova
+        %elifidn %1, SSE3
+            %define movu lddqu
+        %endif
+    %else
+        %xdefine SUFFIX
+        %undef cpuname
+        %undef cpuflags
+    %endif
+%endmacro
+
+; merge MMX and SSE*
+
+%macro CAT_XDEFINE 3
+    %xdefine %1%2 %3
+%endmacro
+
+%macro CAT_UNDEF 2
+    %undef %1%2
+%endmacro
+
+%macro INIT_MMX 0-1+
+    %assign AVX_enabled 0
+    %define RESET_MM_PERMUTATION INIT_MMX %1
+    %define mmsize 8
+    %define num_mmregs 8
+    %define mova movq
+    %define movu movq
+    %define movh movd
+    %define movnta movntq
+    %assign %%i 0
+    %rep 8
+    CAT_XDEFINE m, %%i, mm %+ %%i
+    CAT_XDEFINE nmm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    %rep 8
+    CAT_UNDEF m, %%i
+    CAT_UNDEF nmm, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_XMM 0-1+
+    %assign AVX_enabled 0
+    %define RESET_MM_PERMUTATION INIT_XMM %1
+    %define mmsize 16
+    %define num_mmregs 8
+    %if ARCH_X86_64
+    %define num_mmregs 16
+    %endif
+    %define mova movdqa
+    %define movu movdqu
+    %define movh movq
+    %define movnta movntdq
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, xmm %+ %%i
+    CAT_XDEFINE nxmm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_YMM 0-1+
+    %assign AVX_enabled 1
+    %define RESET_MM_PERMUTATION INIT_YMM %1
+    %define mmsize 32
+    %define num_mmregs 8
+    %if ARCH_X86_64
+    %define num_mmregs 16
+    %endif
+    %define mova vmovaps
+    %define movu vmovups
+    %undef movh
+    %define movnta vmovntps
+    %assign %%i 0
+    %rep num_mmregs
+    CAT_XDEFINE m, %%i, ymm %+ %%i
+    CAT_XDEFINE nymm, %%i, %%i
+    %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+INIT_XMM
+
+; I often want to use macros that permute their arguments. e.g. there's no
+; efficient way to implement butterfly or transpose or dct without swapping some
+; arguments.
+;
+; I would like to not have to manually keep track of the permutations:
+; If I insert a permutation in the middle of a function, it should automatically
+; change everything that follows. For more complex macros I may also have multiple
+; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
+;
+; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
+; permutes its arguments. It's equivalent to exchanging the contents of the
+; registers, except that this way you exchange the register names instead, so it
+; doesn't cost any cycles.
+
+%macro PERMUTE 2-* ; takes a list of pairs to swap
+%rep %0/2
+    %xdefine tmp%2 m%2
+    %xdefine ntmp%2 nm%2
+    %rotate 2
+%endrep
+%rep %0/2
+    %xdefine m%1 tmp%2
+    %xdefine nm%1 ntmp%2
+    %undef tmp%2
+    %undef ntmp%2
+    %rotate 2
+%endrep
+%endmacro
+
+%macro SWAP 2-* ; swaps a single chain (sometimes more concise than pairs)
+%rep %0-1
+%ifdef m%1
+    %xdefine tmp m%1
+    %xdefine m%1 m%2
+    %xdefine m%2 tmp
+    CAT_XDEFINE n, m%1, %1
+    CAT_XDEFINE n, m%2, %2
+%else
+    ; If we were called as "SWAP m0,m1" rather than "SWAP 0,1" infer the original numbers here.
+    ; Be careful using this mode in nested macros though, as in some cases there may be
+    ; other copies of m# that have already been dereferenced and don't get updated correctly.
+    %xdefine %%n1 n %+ %1
+    %xdefine %%n2 n %+ %2
+    %xdefine tmp m %+ %%n1
+    CAT_XDEFINE m, %%n1, m %+ %%n2
+    CAT_XDEFINE m, %%n2, tmp
+    CAT_XDEFINE n, m %+ %%n1, %%n1
+    CAT_XDEFINE n, m %+ %%n2, %%n2
+%endif
+    %undef tmp
+    %rotate 1
+%endrep
+%endmacro
+
+; If SAVE_MM_PERMUTATION is placed at the end of a function, then any later
+; calls to that function will automatically load the permutation, so values can
+; be returned in mmregs.
+%macro SAVE_MM_PERMUTATION 0-1
+    %if %0
+        %xdefine %%f %1_m
+    %else
+        %xdefine %%f current_function %+ _m
+    %endif
+    %assign %%i 0
+    %rep num_mmregs
+        CAT_XDEFINE %%f, %%i, m %+ %%i
+    %assign %%i %%i+1
+    %endrep
+%endmacro
+
+%macro LOAD_MM_PERMUTATION 1 ; name to load from
+    %ifdef %1_m0
+        %assign %%i 0
+        %rep num_mmregs
+            CAT_XDEFINE m, %%i, %1_m %+ %%i
+            CAT_XDEFINE n, m %+ %%i, %%i
+        %assign %%i %%i+1
+        %endrep
+    %endif
+%endmacro
+
+; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't
+%macro call 1
+    call_internal %1, %1 %+ SUFFIX
+%endmacro
+%macro call_internal 2
+    %xdefine %%i %1
+    %ifndef cglobaled_%1
+        %ifdef cglobaled_%2
+            %xdefine %%i %2
+        %endif
+    %endif
+    call %%i
+    LOAD_MM_PERMUTATION %%i
+%endmacro
+
+; Substitutions that reduce instruction size but are functionally equivalent
+%macro add 2
+    %ifnum %2
+        %if %2==128
+            sub %1, -128
+        %else
+            add %1, %2
+        %endif
+    %else
+        add %1, %2
+    %endif
+%endmacro
+
+%macro sub 2
+    %ifnum %2
+        %if %2==128
+            add %1, -128
+        %else
+            sub %1, %2
+        %endif
+    %else
+        sub %1, %2
+    %endif
+%endmacro
+
+;=============================================================================
+; AVX abstraction layer
+;=============================================================================
+
+%assign i 0
+%rep 16
+    %if i < 8
+        CAT_XDEFINE sizeofmm, i, 8
+    %endif
+    CAT_XDEFINE sizeofxmm, i, 16
+    CAT_XDEFINE sizeofymm, i, 32
+%assign i i+1
+%endrep
+%undef i
+
+%macro CHECK_AVX_INSTR_EMU 3-*
+    %xdefine %%opcode %1
+    %xdefine %%dst %2
+    %rep %0-2
+        %ifidn %%dst, %3
+            %error non-AVX emulation of ``%%opcode'' is not supported
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+;%1 == instruction
+;%2 == 1 if float, 0 if int
+;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm)
+;%4 == number of operands given
+;%5+: operands
+%macro RUN_AVX_INSTR 6-7+
+    %ifid %6
+        %define %%sizeofreg sizeof%6
+    %elifid %5
+        %define %%sizeofreg sizeof%5
+    %else
+        %define %%sizeofreg mmsize
+    %endif
+    %if %%sizeofreg==32
+        %if %4>=3
+            v%1 %5, %6, %7
+        %else
+            v%1 %5, %6
+        %endif
+    %else
+        %if %%sizeofreg==8
+            %define %%regmov movq
+        %elif %2
+            %define %%regmov movaps
+        %else
+            %define %%regmov movdqa
+        %endif
+
+        %if %4>=3+%3
+            %ifnidn %5, %6
+                %if AVX_enabled && %%sizeofreg==16
+                    v%1 %5, %6, %7
+                %else
+                    CHECK_AVX_INSTR_EMU {%1 %5, %6, %7}, %5, %7
+                    %%regmov %5, %6
+                    %1 %5, %7
+                %endif
+            %else
+                %1 %5, %7
+            %endif
+        %elif %4>=3
+            %1 %5, %6, %7
+        %else
+            %1 %5, %6
+        %endif
+    %endif
+%endmacro
+
+; 3arg AVX ops with a memory arg can only have it in src2,
+; whereas SSE emulation of 3arg prefers to have it in src1 (i.e. the mov).
+; So, if the op is symmetric and the wrong one is memory, swap them.
+%macro RUN_AVX_INSTR1 8
+    %assign %%swap 0
+    %if AVX_enabled
+        %ifnid %6
+            %assign %%swap 1
+        %endif
+    %elifnidn %5, %6
+        %ifnid %7
+            %assign %%swap 1
+        %endif
+    %endif
+    %if %%swap && %3 == 0 && %8 == 1
+        RUN_AVX_INSTR %1, %2, %3, %4, %5, %7, %6
+    %else
+        RUN_AVX_INSTR %1, %2, %3, %4, %5, %6, %7
+    %endif
+%endmacro
+
+;%1 == instruction
+;%2 == 1 if float, 0 if int
+;%3 == 1 if 4-operand (xmm, xmm, xmm, imm), 0 if 2- or 3-operand (xmm, xmm, xmm)
+;%4 == 1 if symmetric (i.e. doesn't matter which src arg is which), 0 if not
+%macro AVX_INSTR 4
+    %macro %1 2-9 fnord, fnord, fnord, %1, %2, %3, %4
+        %ifidn %3, fnord
+            RUN_AVX_INSTR %6, %7, %8, 2, %1, %2
+        %elifidn %4, fnord
+            RUN_AVX_INSTR1 %6, %7, %8, 3, %1, %2, %3, %9
+        %elifidn %5, fnord
+            RUN_AVX_INSTR %6, %7, %8, 4, %1, %2, %3, %4
+        %else
+            RUN_AVX_INSTR %6, %7, %8, 5, %1, %2, %3, %4, %5
+        %endif
+    %endmacro
+%endmacro
+
+AVX_INSTR addpd, 1, 0, 1
+AVX_INSTR addps, 1, 0, 1
+AVX_INSTR addsd, 1, 0, 1
+AVX_INSTR addss, 1, 0, 1
+AVX_INSTR addsubpd, 1, 0, 0
+AVX_INSTR addsubps, 1, 0, 0
+AVX_INSTR andpd, 1, 0, 1
+AVX_INSTR andps, 1, 0, 1
+AVX_INSTR andnpd, 1, 0, 0
+AVX_INSTR andnps, 1, 0, 0
+AVX_INSTR blendpd, 1, 0, 0
+AVX_INSTR blendps, 1, 0, 0
+AVX_INSTR blendvpd, 1, 0, 0
+AVX_INSTR blendvps, 1, 0, 0
+AVX_INSTR cmppd, 1, 0, 0
+AVX_INSTR cmpps, 1, 0, 0
+AVX_INSTR cmpsd, 1, 0, 0
+AVX_INSTR cmpss, 1, 0, 0
+AVX_INSTR cvtdq2ps, 1, 0, 0
+AVX_INSTR cvtps2dq, 1, 0, 0
+AVX_INSTR divpd, 1, 0, 0
+AVX_INSTR divps, 1, 0, 0
+AVX_INSTR divsd, 1, 0, 0
+AVX_INSTR divss, 1, 0, 0
+AVX_INSTR dppd, 1, 1, 0
+AVX_INSTR dpps, 1, 1, 0
+AVX_INSTR haddpd, 1, 0, 0
+AVX_INSTR haddps, 1, 0, 0
+AVX_INSTR hsubpd, 1, 0, 0
+AVX_INSTR hsubps, 1, 0, 0
+AVX_INSTR maxpd, 1, 0, 1
+AVX_INSTR maxps, 1, 0, 1
+AVX_INSTR maxsd, 1, 0, 1
+AVX_INSTR maxss, 1, 0, 1
+AVX_INSTR minpd, 1, 0, 1
+AVX_INSTR minps, 1, 0, 1
+AVX_INSTR minsd, 1, 0, 1
+AVX_INSTR minss, 1, 0, 1
+AVX_INSTR movhlps, 1, 0, 0
+AVX_INSTR movlhps, 1, 0, 0
+AVX_INSTR movsd, 1, 0, 0
+AVX_INSTR movss, 1, 0, 0
+AVX_INSTR mpsadbw, 0, 1, 0
+AVX_INSTR mulpd, 1, 0, 1
+AVX_INSTR mulps, 1, 0, 1
+AVX_INSTR mulsd, 1, 0, 1
+AVX_INSTR mulss, 1, 0, 1
+AVX_INSTR orpd, 1, 0, 1
+AVX_INSTR orps, 1, 0, 1
+AVX_INSTR pabsb, 0, 0, 0
+AVX_INSTR pabsw, 0, 0, 0
+AVX_INSTR pabsd, 0, 0, 0
+AVX_INSTR packsswb, 0, 0, 0
+AVX_INSTR packssdw, 0, 0, 0
+AVX_INSTR packuswb, 0, 0, 0
+AVX_INSTR packusdw, 0, 0, 0
+AVX_INSTR paddb, 0, 0, 1
+AVX_INSTR paddw, 0, 0, 1
+AVX_INSTR paddd, 0, 0, 1
+AVX_INSTR paddq, 0, 0, 1
+AVX_INSTR paddsb, 0, 0, 1
+AVX_INSTR paddsw, 0, 0, 1
+AVX_INSTR paddusb, 0, 0, 1
+AVX_INSTR paddusw, 0, 0, 1
+AVX_INSTR palignr, 0, 1, 0
+AVX_INSTR pand, 0, 0, 1
+AVX_INSTR pandn, 0, 0, 0
+AVX_INSTR pavgb, 0, 0, 1
+AVX_INSTR pavgw, 0, 0, 1
+AVX_INSTR pblendvb, 0, 0, 0
+AVX_INSTR pblendw, 0, 1, 0
+AVX_INSTR pcmpestri, 0, 0, 0
+AVX_INSTR pcmpestrm, 0, 0, 0
+AVX_INSTR pcmpistri, 0, 0, 0
+AVX_INSTR pcmpistrm, 0, 0, 0
+AVX_INSTR pcmpeqb, 0, 0, 1
+AVX_INSTR pcmpeqw, 0, 0, 1
+AVX_INSTR pcmpeqd, 0, 0, 1
+AVX_INSTR pcmpeqq, 0, 0, 1
+AVX_INSTR pcmpgtb, 0, 0, 0
+AVX_INSTR pcmpgtw, 0, 0, 0
+AVX_INSTR pcmpgtd, 0, 0, 0
+AVX_INSTR pcmpgtq, 0, 0, 0
+AVX_INSTR phaddw, 0, 0, 0
+AVX_INSTR phaddd, 0, 0, 0
+AVX_INSTR phaddsw, 0, 0, 0
+AVX_INSTR phsubw, 0, 0, 0
+AVX_INSTR phsubd, 0, 0, 0
+AVX_INSTR phsubsw, 0, 0, 0
+AVX_INSTR pmaddwd, 0, 0, 1
+AVX_INSTR pmaddubsw, 0, 0, 0
+AVX_INSTR pmaxsb, 0, 0, 1
+AVX_INSTR pmaxsw, 0, 0, 1
+AVX_INSTR pmaxsd, 0, 0, 1
+AVX_INSTR pmaxub, 0, 0, 1
+AVX_INSTR pmaxuw, 0, 0, 1
+AVX_INSTR pmaxud, 0, 0, 1
+AVX_INSTR pminsb, 0, 0, 1
+AVX_INSTR pminsw, 0, 0, 1
+AVX_INSTR pminsd, 0, 0, 1
+AVX_INSTR pminub, 0, 0, 1
+AVX_INSTR pminuw, 0, 0, 1
+AVX_INSTR pminud, 0, 0, 1
+AVX_INSTR pmovmskb, 0, 0, 0
+AVX_INSTR pmulhuw, 0, 0, 1
+AVX_INSTR pmulhrsw, 0, 0, 1
+AVX_INSTR pmulhw, 0, 0, 1
+AVX_INSTR pmullw, 0, 0, 1
+AVX_INSTR pmulld, 0, 0, 1
+AVX_INSTR pmuludq, 0, 0, 1
+AVX_INSTR pmuldq, 0, 0, 1
+AVX_INSTR por, 0, 0, 1
+AVX_INSTR psadbw, 0, 0, 1
+AVX_INSTR pshufb, 0, 0, 0
+AVX_INSTR pshufd, 0, 1, 0
+AVX_INSTR pshufhw, 0, 1, 0
+AVX_INSTR pshuflw, 0, 1, 0
+AVX_INSTR psignb, 0, 0, 0
+AVX_INSTR psignw, 0, 0, 0
+AVX_INSTR psignd, 0, 0, 0
+AVX_INSTR psllw, 0, 0, 0
+AVX_INSTR pslld, 0, 0, 0
+AVX_INSTR psllq, 0, 0, 0
+AVX_INSTR pslldq, 0, 0, 0
+AVX_INSTR psraw, 0, 0, 0
+AVX_INSTR psrad, 0, 0, 0
+AVX_INSTR psrlw, 0, 0, 0
+AVX_INSTR psrld, 0, 0, 0
+AVX_INSTR psrlq, 0, 0, 0
+AVX_INSTR psrldq, 0, 0, 0
+AVX_INSTR psubb, 0, 0, 0
+AVX_INSTR psubw, 0, 0, 0
+AVX_INSTR psubd, 0, 0, 0
+AVX_INSTR psubq, 0, 0, 0
+AVX_INSTR psubsb, 0, 0, 0
+AVX_INSTR psubsw, 0, 0, 0
+AVX_INSTR psubusb, 0, 0, 0
+AVX_INSTR psubusw, 0, 0, 0
+AVX_INSTR ptest, 0, 0, 0
+AVX_INSTR punpckhbw, 0, 0, 0
+AVX_INSTR punpckhwd, 0, 0, 0
+AVX_INSTR punpckhdq, 0, 0, 0
+AVX_INSTR punpckhqdq, 0, 0, 0
+AVX_INSTR punpcklbw, 0, 0, 0
+AVX_INSTR punpcklwd, 0, 0, 0
+AVX_INSTR punpckldq, 0, 0, 0
+AVX_INSTR punpcklqdq, 0, 0, 0
+AVX_INSTR pxor, 0, 0, 1
+AVX_INSTR shufps, 1, 1, 0
+AVX_INSTR subpd, 1, 0, 0
+AVX_INSTR subps, 1, 0, 0
+AVX_INSTR subsd, 1, 0, 0
+AVX_INSTR subss, 1, 0, 0
+AVX_INSTR unpckhpd, 1, 0, 0
+AVX_INSTR unpckhps, 1, 0, 0
+AVX_INSTR unpcklpd, 1, 0, 0
+AVX_INSTR unpcklps, 1, 0, 0
+AVX_INSTR xorpd, 1, 0, 1
+AVX_INSTR xorps, 1, 0, 1
+
+; 3DNow instructions, for sharing code between AVX, SSE and 3DN
+AVX_INSTR pfadd, 1, 0, 1
+AVX_INSTR pfsub, 1, 0, 0
+AVX_INSTR pfmul, 1, 0, 1
+
+; base-4 constants for shuffles
+%assign i 0
+%rep 256
+    %assign j ((i>>6)&3)*1000 + ((i>>4)&3)*100 + ((i>>2)&3)*10 + (i&3)
+    %if j < 10
+        CAT_XDEFINE q000, j, i
+    %elif j < 100
+        CAT_XDEFINE q00, j, i
+    %elif j < 1000
+        CAT_XDEFINE q0, j, i
+    %else
+        CAT_XDEFINE q, j, i
+    %endif
+%assign i i+1
+%endrep
+%undef i
+%undef j
+
+%macro FMA_INSTR 3
+    %macro %1 4-7 %1, %2, %3
+        %if cpuflag(xop)
+            v%5 %1, %2, %3, %4
+        %else
+            %6 %1, %2, %3
+            %7 %1, %4
+        %endif
+    %endmacro
+%endmacro
+
+FMA_INSTR  pmacsdd,  pmulld, paddd
+FMA_INSTR  pmacsww,  pmullw, paddw
+FMA_INSTR pmadcswd, pmaddwd, paddd
+
+; tzcnt is equivalent to "rep bsf" and is backwards-compatible with bsf.
+; This lets us use tzcnt without bumping the yasm version requirement yet.
+%define tzcnt rep bsf
diff --git a/third_party/aom/third_party/x86inc/LICENSE b/third_party/aom/third_party/x86inc/LICENSE
new file mode 100644
index 0000000000..7d07645a17
--- /dev/null
+++ b/third_party/aom/third_party/x86inc/LICENSE
@@ -0,0 +1,18 @@
+Copyright (C) 2005-2012 x264 project
+
+Authors: Loren Merritt <lorenm@u.washington.edu>
+         Anton Mitrofanov <BugMaster@narod.ru>
+         Jason Garrett-Glaser <darkshikari@gmail.com>
+         Henrik Gramner <hengar-6@student.ltu.se>
+
+Permission to use, copy, modify, and/or distribute this software for any
+purpose with or without fee is hereby granted, provided that the above
+copyright notice and this permission notice appear in all copies.
+
+THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
diff --git a/third_party/aom/third_party/x86inc/README.libaom b/third_party/aom/third_party/x86inc/README.libaom
new file mode 100644
index 0000000000..07c4dad20a
--- /dev/null
+++ b/third_party/aom/third_party/x86inc/README.libaom
@@ -0,0 +1,20 @@
+URL: https://git.videolan.org/git/x264.git
+Version: d23d18655249944c1ca894b451e2c82c7a584c62
+License: ISC
+License File: LICENSE
+
+Description:
+x264/libav's framework for x86 assembly. Contains a variety of macros and
+defines that help automatically allow assembly to work cross-platform.
+
+Local Modifications:
+Get configuration from aom_config.asm.
+Prefix functions with aom by default.
+Manage name mangling (prefixing with '_') manually because 'PREFIX' does not
+  exist in libaom.
+Expand PIC default to macho64 and respect CONFIG_PIC from libaom
+Set 'private_extern' visibility for macho targets.
+Copy PIC 'GLOBAL' macros from x86_abi_support.asm
+Use .text instead of .rodata on macho to avoid broken tables in PIC mode.
+Use .text with no alignment for aout
+Only use 'hidden' visibility with Chromium
diff --git a/third_party/aom/third_party/x86inc/x86inc.asm b/third_party/aom/third_party/x86inc/x86inc.asm
new file mode 100644
index 0000000000..cfee99c788
--- /dev/null
+++ b/third_party/aom/third_party/x86inc/x86inc.asm
@@ -0,0 +1,1649 @@
+;*****************************************************************************
+;* x86inc.asm: x264asm abstraction layer
+;*****************************************************************************
+;* Copyright (C) 2005-2016 x264 project
+;*
+;* Authors: Loren Merritt <lorenm@u.washington.edu>
+;*          Anton Mitrofanov <BugMaster@narod.ru>
+;*          Fiona Glaser <fiona@x264.com>
+;*          Henrik Gramner <henrik@gramner.com>
+;*
+;* Permission to use, copy, modify, and/or distribute this software for any
+;* purpose with or without fee is hereby granted, provided that the above
+;* copyright notice and this permission notice appear in all copies.
+;*
+;* THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
+;* WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
+;* MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
+;* ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
+;* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
+;* ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
+;* OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
+;*****************************************************************************
+
+; This is a header file for the x264ASM assembly language, which uses
+; NASM/YASM syntax combined with a large number of macros to provide easy
+; abstraction between different calling conventions (x86_32, win64, linux64).
+; It also has various other useful features to simplify writing the kind of
+; DSP functions that are most often used in x264.
+
+; Unlike the rest of x264, this file is available under an ISC license, as it
+; has significant usefulness outside of x264 and we want it to be available
+; to the largest audience possible.  Of course, if you modify it for your own
+; purposes to add a new feature, we strongly encourage contributing a patch
+; as this feature might be useful for others as well.  Send patches or ideas
+; to x264-devel@videolan.org .
+
+%include "aom_config.asm"
+
+%ifndef private_prefix
+    %define private_prefix aom
+%endif
+
+%ifndef public_prefix
+    %define public_prefix private_prefix
+%endif
+
+%ifndef STACK_ALIGNMENT
+    %if ARCH_X86_64
+        %define STACK_ALIGNMENT 16
+    %else
+        %define STACK_ALIGNMENT 4
+    %endif
+%endif
+
+%define WIN64  0
+%define UNIX64 0
+%if ARCH_X86_64
+    %ifidn __OUTPUT_FORMAT__,win32
+        %define WIN64  1
+    %elifidn __OUTPUT_FORMAT__,win64
+        %define WIN64  1
+    %elifidn __OUTPUT_FORMAT__,x64
+        %define WIN64  1
+    %else
+        %define UNIX64 1
+    %endif
+%endif
+
+%define FORMAT_ELF 0
+%ifidn __OUTPUT_FORMAT__,elf
+    %define FORMAT_ELF 1
+%elifidn __OUTPUT_FORMAT__,elf32
+    %define FORMAT_ELF 1
+%elifidn __OUTPUT_FORMAT__,elf64
+    %define FORMAT_ELF 1
+%endif
+
+%define FORMAT_MACHO 0
+%ifidn __OUTPUT_FORMAT__,macho32
+     %define FORMAT_MACHO 1
+%elifidn __OUTPUT_FORMAT__,macho64
+     %define FORMAT_MACHO 1
+%endif
+
+; Set PREFIX for libaom builds.
+%if FORMAT_ELF
+    %undef PREFIX
+%elif WIN64
+    %undef PREFIX
+%else
+    %define PREFIX
+%endif
+
+%ifdef PREFIX
+    %define mangle(x) _ %+ x
+%else
+    %define mangle(x) x
+%endif
+
+; In some instances macho32 tables get misaligned when using .rodata.
+; When looking at the disassembly it appears that the offset is either
+; correct or consistently off by 90. Placing them in the .text section
+; works around the issue. It appears to be specific to the way libaom
+; handles the tables.
+%macro SECTION_RODATA 0-1 16
+    %ifidn __OUTPUT_FORMAT__,macho32
+        SECTION .text align=%1
+        fakegot:
+    %elifidn __OUTPUT_FORMAT__,aout
+        SECTION .text
+    %else
+        SECTION .rodata align=%1
+    %endif
+%endmacro
+
+; PIC macros are copied from aom_ports/x86_abi_support.asm. The "define PIC"
+; from original code is added in for 64bit.
+%ifidn __OUTPUT_FORMAT__,elf32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,macho32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,win32
+%define ABI_IS_32BIT 1
+%elifidn __OUTPUT_FORMAT__,aout
+%define ABI_IS_32BIT 1
+%else
+%define ABI_IS_32BIT 0
+%endif
+
+%if ABI_IS_32BIT
+    %if CONFIG_PIC=1
+        %ifidn __OUTPUT_FORMAT__,elf32
+            %define GET_GOT_DEFINED 1
+            %define WRT_PLT wrt ..plt
+            %macro GET_GOT 1
+                extern _GLOBAL_OFFSET_TABLE_
+                push %1
+                call %%get_got
+                %%sub_offset:
+                jmp %%exitGG
+                %%get_got:
+                mov %1, [esp]
+                add %1, _GLOBAL_OFFSET_TABLE_ + $$ - %%sub_offset wrt ..gotpc
+                ret
+                %%exitGG:
+                %undef GLOBAL
+                %define GLOBAL(x) x + %1 wrt ..gotoff
+                %undef RESTORE_GOT
+                %define RESTORE_GOT pop %1
+            %endmacro
+        %elifidn __OUTPUT_FORMAT__,macho32
+            %define GET_GOT_DEFINED 1
+            %macro GET_GOT 1
+                push %1
+                call %%get_got
+                %%get_got:
+                pop  %1
+                %undef GLOBAL
+                %define GLOBAL(x) x + %1 - %%get_got
+                %undef RESTORE_GOT
+                %define RESTORE_GOT pop %1
+            %endmacro
+        %else
+            %define GET_GOT_DEFINED 0
+        %endif
+    %endif
+
+    %if ARCH_X86_64 == 0
+        %undef PIC
+    %endif
+
+%else
+    %macro GET_GOT 1
+    %endmacro
+    %define GLOBAL(x) rel x
+    %define WRT_PLT wrt ..plt
+
+    %if WIN64
+        %define PIC
+    %elifidn __OUTPUT_FORMAT__,macho64
+        %define PIC
+    %elif CONFIG_PIC
+        %define PIC
+    %endif
+%endif
+
+%ifnmacro GET_GOT
+    %macro GET_GOT 1
+    %endmacro
+    %define GLOBAL(x) x
+%endif
+%ifndef RESTORE_GOT
+    %define RESTORE_GOT
+%endif
+%ifndef WRT_PLT
+    %define WRT_PLT
+%endif
+
+%ifdef PIC
+    default rel
+%endif
+
+%ifndef GET_GOT_DEFINED
+    %define GET_GOT_DEFINED 0
+%endif
+; Done with PIC macros
+
+%ifdef __NASM_VER__
+    %use smartalign
+%endif
+
+; Macros to eliminate most code duplication between x86_32 and x86_64:
+; Currently this works only for leaf functions which load all their arguments
+; into registers at the start, and make no other use of the stack. Luckily that
+; covers most of x264's asm.
+
+; PROLOGUE:
+; %1 = number of arguments. loads them from stack if needed.
+; %2 = number of registers used. pushes callee-saved regs if needed.
+; %3 = number of xmm registers used. pushes callee-saved xmm regs if needed.
+; %4 = (optional) stack size to be allocated. The stack will be aligned before
+;      allocating the specified stack size. If the required stack alignment is
+;      larger than the known stack alignment the stack will be manually aligned
+;      and an extra register will be allocated to hold the original stack
+;      pointer (to not invalidate r0m etc.). To prevent the use of an extra
+;      register as stack pointer, request a negative stack size.
+; %4+/%5+ = list of names to define to registers
+; PROLOGUE can also be invoked by adding the same options to cglobal
+
+; e.g.
+; cglobal foo, 2,3,7,0x40, dst, src, tmp
+; declares a function (foo) that automatically loads two arguments (dst and
+; src) into registers, uses one additional register (tmp) plus 7 vector
+; registers (m0-m6) and allocates 0x40 bytes of stack space.
+
+; TODO Some functions can use some args directly from the stack. If they're the
+; last args then you can just not declare them, but if they're in the middle
+; we need more flexible macro.
+
+; RET:
+; Pops anything that was pushed by PROLOGUE, and returns.
+
+; REP_RET:
+; Use this instead of RET if it's a branch target.
+
+; registers:
+; rN and rNq are the native-size register holding function argument N
+; rNd, rNw, rNb are dword, word, and byte size
+; rNh is the high 8 bits of the word size
+; rNm is the original location of arg N (a register or on the stack), dword
+; rNmp is native size
+
+%macro DECLARE_REG 2-3
+    %define r%1q %2
+    %define r%1d %2d
+    %define r%1w %2w
+    %define r%1b %2b
+    %define r%1h %2h
+    %define %2q %2
+    %if %0 == 2
+        %define r%1m  %2d
+        %define r%1mp %2
+    %elif ARCH_X86_64 ; memory
+        %define r%1m [rstk + stack_offset + %3]
+        %define r%1mp qword r %+ %1 %+ m
+    %else
+        %define r%1m [rstk + stack_offset + %3]
+        %define r%1mp dword r %+ %1 %+ m
+    %endif
+    %define r%1  %2
+%endmacro
+
+%macro DECLARE_REG_SIZE 3
+    %define r%1q r%1
+    %define e%1q r%1
+    %define r%1d e%1
+    %define e%1d e%1
+    %define r%1w %1
+    %define e%1w %1
+    %define r%1h %3
+    %define e%1h %3
+    %define r%1b %2
+    %define e%1b %2
+    %if ARCH_X86_64 == 0
+        %define r%1 e%1
+    %endif
+%endmacro
+
+DECLARE_REG_SIZE ax, al, ah
+DECLARE_REG_SIZE bx, bl, bh
+DECLARE_REG_SIZE cx, cl, ch
+DECLARE_REG_SIZE dx, dl, dh
+DECLARE_REG_SIZE si, sil, null
+DECLARE_REG_SIZE di, dil, null
+DECLARE_REG_SIZE bp, bpl, null
+
+; t# defines for when per-arch register allocation is more complex than just function arguments
+
+%macro DECLARE_REG_TMP 1-*
+    %assign %%i 0
+    %rep %0
+        CAT_XDEFINE t, %%i, r%1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro DECLARE_REG_TMP_SIZE 0-*
+    %rep %0
+        %define t%1q t%1 %+ q
+        %define t%1d t%1 %+ d
+        %define t%1w t%1 %+ w
+        %define t%1h t%1 %+ h
+        %define t%1b t%1 %+ b
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_REG_TMP_SIZE 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14
+
+%if ARCH_X86_64
+    %define gprsize 8
+%else
+    %define gprsize 4
+%endif
+
+%macro PUSH 1
+    push %1
+    %ifidn rstk, rsp
+        %assign stack_offset stack_offset+gprsize
+    %endif
+%endmacro
+
+%macro POP 1
+    pop %1
+    %ifidn rstk, rsp
+        %assign stack_offset stack_offset-gprsize
+    %endif
+%endmacro
+
+%macro PUSH_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            PUSH r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro POP_IF_USED 1-*
+    %rep %0
+        %if %1 < regs_used
+            pop r%1
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro LOAD_IF_USED 1-*
+    %rep %0
+        %if %1 < num_args
+            mov r%1, r %+ %1 %+ mp
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro SUB 2
+    sub %1, %2
+    %ifidn %1, rstk
+        %assign stack_offset stack_offset+(%2)
+    %endif
+%endmacro
+
+%macro ADD 2
+    add %1, %2
+    %ifidn %1, rstk
+        %assign stack_offset stack_offset-(%2)
+    %endif
+%endmacro
+
+%macro movifnidn 2
+    %ifnidn %1, %2
+        mov %1, %2
+    %endif
+%endmacro
+
+%macro movsxdifnidn 2
+    %ifnidn %1, %2
+        movsxd %1, %2
+    %endif
+%endmacro
+
+%macro ASSERT 1
+    %if (%1) == 0
+        %error assertion ``%1'' failed
+    %endif
+%endmacro
+
+%macro DEFINE_ARGS 0-*
+    %ifdef n_arg_names
+        %assign %%i 0
+        %rep n_arg_names
+            CAT_UNDEF arg_name %+ %%i, q
+            CAT_UNDEF arg_name %+ %%i, d
+            CAT_UNDEF arg_name %+ %%i, w
+            CAT_UNDEF arg_name %+ %%i, h
+            CAT_UNDEF arg_name %+ %%i, b
+            CAT_UNDEF arg_name %+ %%i, m
+            CAT_UNDEF arg_name %+ %%i, mp
+            CAT_UNDEF arg_name, %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+
+    %xdefine %%stack_offset stack_offset
+    %undef stack_offset ; so that the current value of stack_offset doesn't get baked in by xdefine
+    %assign %%i 0
+    %rep %0
+        %xdefine %1q r %+ %%i %+ q
+        %xdefine %1d r %+ %%i %+ d
+        %xdefine %1w r %+ %%i %+ w
+        %xdefine %1h r %+ %%i %+ h
+        %xdefine %1b r %+ %%i %+ b
+        %xdefine %1m r %+ %%i %+ m
+        %xdefine %1mp r %+ %%i %+ mp
+        CAT_XDEFINE arg_name, %%i, %1
+        %assign %%i %%i+1
+        %rotate 1
+    %endrep
+    %xdefine stack_offset %%stack_offset
+    %assign n_arg_names %0
+%endmacro
+
+%define required_stack_alignment ((mmsize + 15) & ~15)
+
+%macro ALLOC_STACK 1-2 0 ; stack_size, n_xmm_regs (for win64 only)
+    %ifnum %1
+        %if %1 != 0
+            %assign %%pad 0
+            %assign stack_size %1
+            %if stack_size < 0
+                %assign stack_size -stack_size
+            %endif
+            %if WIN64
+                %assign %%pad %%pad + 32 ; shadow space
+                %if mmsize != 8
+                    %assign xmm_regs_used %2
+                    %if xmm_regs_used > 8
+                        %assign %%pad %%pad + (xmm_regs_used-8)*16 ; callee-saved xmm registers
+                    %endif
+                %endif
+            %endif
+            %if required_stack_alignment <= STACK_ALIGNMENT
+                ; maintain the current stack alignment
+                %assign stack_size_padded stack_size + %%pad + ((-%%pad-stack_offset-gprsize) & (STACK_ALIGNMENT-1))
+                SUB rsp, stack_size_padded
+            %else
+                %assign %%reg_num (regs_used - 1)
+                %xdefine rstk r %+ %%reg_num
+                ; align stack, and save original stack location directly above
+                ; it, i.e. in [rsp+stack_size_padded], so we can restore the
+                ; stack in a single instruction (i.e. mov rsp, rstk or mov
+                ; rsp, [rsp+stack_size_padded])
+                %if %1 < 0 ; need to store rsp on stack
+                    %xdefine rstkm [rsp + stack_size + %%pad]
+                    %assign %%pad %%pad + gprsize
+                %else ; can keep rsp in rstk during whole function
+                    %xdefine rstkm rstk
+                %endif
+                %assign stack_size_padded stack_size + ((%%pad + required_stack_alignment-1) & ~(required_stack_alignment-1))
+                mov rstk, rsp
+                and rsp, ~(required_stack_alignment-1)
+                sub rsp, stack_size_padded
+                movifnidn rstkm, rstk
+            %endif
+            WIN64_PUSH_XMM
+        %endif
+    %endif
+%endmacro
+
+%macro SETUP_STACK_POINTER 1
+    %ifnum %1
+        %if %1 != 0 && required_stack_alignment > STACK_ALIGNMENT
+            %if %1 > 0
+                %assign regs_used (regs_used + 1)
+            %endif
+            %if ARCH_X86_64 && regs_used < 5 + UNIX64 * 3
+                ; Ensure that we don't clobber any registers containing arguments
+                %assign regs_used 5 + UNIX64 * 3
+            %endif
+        %endif
+    %endif
+%endmacro
+
+%macro DEFINE_ARGS_INTERNAL 3+
+    %ifnum %2
+        DEFINE_ARGS %3
+    %elif %1 == 4
+        DEFINE_ARGS %2
+    %elif %1 > 4
+        DEFINE_ARGS %2, %3
+    %endif
+%endmacro
+
+%if WIN64 ; Windows x64 ;=================================================
+
+DECLARE_REG 0,  rcx
+DECLARE_REG 1,  rdx
+DECLARE_REG 2,  R8
+DECLARE_REG 3,  R9
+DECLARE_REG 4,  R10, 40
+DECLARE_REG 5,  R11, 48
+DECLARE_REG 6,  rax, 56
+DECLARE_REG 7,  rdi, 64
+DECLARE_REG 8,  rsi, 72
+DECLARE_REG 9,  rbx, 80
+DECLARE_REG 10, rbp, 88
+DECLARE_REG 11, R12, 96
+DECLARE_REG 12, R13, 104
+DECLARE_REG 13, R14, 112
+DECLARE_REG 14, R15, 120
+
+%macro PROLOGUE 2-5+ 0 ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    SETUP_STACK_POINTER %4
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 7, 8, 9, 10, 11, 12, 13, 14
+    ALLOC_STACK %4, %3
+    %if mmsize != 8 && stack_size == 0
+        WIN64_SPILL_XMM %3
+    %endif
+    LOAD_IF_USED 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS_INTERNAL %0, %4, %5
+%endmacro
+
+%macro WIN64_PUSH_XMM 0
+    ; Use the shadow space to store XMM6 and XMM7, the rest needs stack space allocated.
+    %if xmm_regs_used > 6
+        movaps [rstk + stack_offset +  8], xmm6
+    %endif
+    %if xmm_regs_used > 7
+        movaps [rstk + stack_offset + 24], xmm7
+    %endif
+    %if xmm_regs_used > 8
+        %assign %%i 8
+        %rep xmm_regs_used-8
+            movaps [rsp + (%%i-8)*16 + stack_size + 32], xmm %+ %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+%endmacro
+
+%macro WIN64_SPILL_XMM 1
+    %assign xmm_regs_used %1
+    ASSERT xmm_regs_used <= 16
+    %if xmm_regs_used > 8
+        ; Allocate stack space for callee-saved xmm registers plus shadow space and align the stack.
+        %assign %%pad (xmm_regs_used-8)*16 + 32
+        %assign stack_size_padded %%pad + ((-%%pad-stack_offset-gprsize) & (STACK_ALIGNMENT-1))
+        SUB rsp, stack_size_padded
+    %endif
+    WIN64_PUSH_XMM
+%endmacro
+
+%macro WIN64_RESTORE_XMM_INTERNAL 1
+    %assign %%pad_size 0
+    %if xmm_regs_used > 8
+        %assign %%i xmm_regs_used
+        %rep xmm_regs_used-8
+            %assign %%i %%i-1
+            movaps xmm %+ %%i, [%1 + (%%i-8)*16 + stack_size + 32]
+        %endrep
+    %endif
+    %if stack_size_padded > 0
+        %if stack_size > 0 && required_stack_alignment > STACK_ALIGNMENT
+            mov rsp, rstkm
+        %else
+            add %1, stack_size_padded
+            %assign %%pad_size stack_size_padded
+        %endif
+    %endif
+    %if xmm_regs_used > 7
+        movaps xmm7, [%1 + stack_offset - %%pad_size + 24]
+    %endif
+    %if xmm_regs_used > 6
+        movaps xmm6, [%1 + stack_offset - %%pad_size +  8]
+    %endif
+%endmacro
+
+%macro WIN64_RESTORE_XMM 1
+    WIN64_RESTORE_XMM_INTERNAL %1
+    %assign stack_offset (stack_offset-stack_size_padded)
+    %assign xmm_regs_used 0
+%endmacro
+
+%define has_epilogue regs_used > 7 || xmm_regs_used > 6 || mmsize == 32 || stack_size > 0
+
+%macro RET 0
+    WIN64_RESTORE_XMM_INTERNAL rsp
+    POP_IF_USED 14, 13, 12, 11, 10, 9, 8, 7
+    %if mmsize == 32
+        vzeroupper
+    %endif
+    AUTO_REP_RET
+%endmacro
+
+%elif ARCH_X86_64 ; *nix x64 ;=============================================
+
+DECLARE_REG 0,  rdi
+DECLARE_REG 1,  rsi
+DECLARE_REG 2,  rdx
+DECLARE_REG 3,  rcx
+DECLARE_REG 4,  R8
+DECLARE_REG 5,  R9
+DECLARE_REG 6,  rax, 8
+DECLARE_REG 7,  R10, 16
+DECLARE_REG 8,  R11, 24
+DECLARE_REG 9,  rbx, 32
+DECLARE_REG 10, rbp, 40
+DECLARE_REG 11, R12, 48
+DECLARE_REG 12, R13, 56
+DECLARE_REG 13, R14, 64
+DECLARE_REG 14, R15, 72
+
+%macro PROLOGUE 2-5+ ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    SETUP_STACK_POINTER %4
+    ASSERT regs_used <= 15
+    PUSH_IF_USED 9, 10, 11, 12, 13, 14
+    ALLOC_STACK %4
+    LOAD_IF_USED 6, 7, 8, 9, 10, 11, 12, 13, 14
+    DEFINE_ARGS_INTERNAL %0, %4, %5
+%endmacro
+
+%define has_epilogue regs_used > 9 || mmsize == 32 || stack_size > 0
+
+%macro RET 0
+    %if stack_size_padded > 0
+        %if required_stack_alignment > STACK_ALIGNMENT
+            mov rsp, rstkm
+        %else
+            add rsp, stack_size_padded
+        %endif
+    %endif
+    POP_IF_USED 14, 13, 12, 11, 10, 9
+    %if mmsize == 32
+        vzeroupper
+    %endif
+    AUTO_REP_RET
+%endmacro
+
+%else ; X86_32 ;==============================================================
+
+DECLARE_REG 0, eax, 4
+DECLARE_REG 1, ecx, 8
+DECLARE_REG 2, edx, 12
+DECLARE_REG 3, ebx, 16
+DECLARE_REG 4, esi, 20
+DECLARE_REG 5, edi, 24
+DECLARE_REG 6, ebp, 28
+%define rsp esp
+
+%macro DECLARE_ARG 1-*
+    %rep %0
+        %define r%1m [rstk + stack_offset + 4*%1 + 4]
+        %define r%1mp dword r%1m
+        %rotate 1
+    %endrep
+%endmacro
+
+DECLARE_ARG 7, 8, 9, 10, 11, 12, 13, 14
+
+%macro PROLOGUE 2-5+ ; #args, #regs, #xmm_regs, [stack_size,] arg_names...
+    %assign num_args %1
+    %assign regs_used %2
+    ASSERT regs_used >= num_args
+    %if num_args > 7
+        %assign num_args 7
+    %endif
+    %if regs_used > 7
+        %assign regs_used 7
+    %endif
+    SETUP_STACK_POINTER %4
+    ASSERT regs_used <= 7
+    PUSH_IF_USED 3, 4, 5, 6
+    ALLOC_STACK %4
+    LOAD_IF_USED 0, 1, 2, 3, 4, 5, 6
+    DEFINE_ARGS_INTERNAL %0, %4, %5
+%endmacro
+
+%define has_epilogue regs_used > 3 || mmsize == 32 || stack_size > 0
+
+%macro RET 0
+    %if stack_size_padded > 0
+        %if required_stack_alignment > STACK_ALIGNMENT
+            mov rsp, rstkm
+        %else
+            add rsp, stack_size_padded
+        %endif
+    %endif
+    POP_IF_USED 6, 5, 4, 3
+    %if mmsize == 32
+        vzeroupper
+    %endif
+    AUTO_REP_RET
+%endmacro
+
+%endif ;======================================================================
+
+%if WIN64 == 0
+    %macro WIN64_SPILL_XMM 1
+    %endmacro
+    %macro WIN64_RESTORE_XMM 1
+    %endmacro
+    %macro WIN64_PUSH_XMM 0
+    %endmacro
+%endif
+
+; On AMD cpus <=K10, an ordinary ret is slow if it immediately follows either
+; a branch or a branch target. So switch to a 2-byte form of ret in that case.
+; We can automatically detect "follows a branch", but not a branch target.
+; (SSSE3 is a sufficient condition to know that your cpu doesn't have this problem.)
+%macro REP_RET 0
+    %if has_epilogue
+        RET
+    %else
+        rep ret
+    %endif
+    annotate_function_size
+%endmacro
+
+%define last_branch_adr $$
+%macro AUTO_REP_RET 0
+    %if notcpuflag(ssse3)
+        times ((last_branch_adr-$)>>31)+1 rep ; times 1 iff $ == last_branch_adr.
+    %endif
+    ret
+    annotate_function_size
+%endmacro
+
+%macro BRANCH_INSTR 0-*
+    %rep %0
+        %macro %1 1-2 %1
+            %2 %1
+            %if notcpuflag(ssse3)
+                %%branch_instr equ $
+                %xdefine last_branch_adr %%branch_instr
+            %endif
+        %endmacro
+        %rotate 1
+    %endrep
+%endmacro
+
+BRANCH_INSTR jz, je, jnz, jne, jl, jle, jnl, jnle, jg, jge, jng, jnge, ja, jae, jna, jnae, jb, jbe, jnb, jnbe, jc, jnc, js, jns, jo, jno, jp, jnp
+
+%macro TAIL_CALL 2 ; callee, is_nonadjacent
+    %if has_epilogue
+        call %1
+        RET
+    %elif %2
+        jmp %1
+    %endif
+    annotate_function_size
+%endmacro
+
+;=============================================================================
+; arch-independent part
+;=============================================================================
+
+%assign function_align 16
+
+; Begin a function.
+; Applies any symbol mangling needed for C linkage, and sets up a define such that
+; subsequent uses of the function name automatically refer to the mangled version.
+; Appends cpuflags to the function name if cpuflags has been specified.
+; The "" empty default parameter is a workaround for nasm, which fails if SUFFIX
+; is empty and we call cglobal_internal with just %1 %+ SUFFIX (without %2).
+%macro cglobal 1-2+ "" ; name, [PROLOGUE args]
+    cglobal_internal 1, %1 %+ SUFFIX, %2
+%endmacro
+%macro cvisible 1-2+ "" ; name, [PROLOGUE args]
+    cglobal_internal 0, %1 %+ SUFFIX, %2
+%endmacro
+%macro cglobal_internal 2-3+
+    annotate_function_size
+    %if %1
+        %xdefine %%FUNCTION_PREFIX private_prefix
+        ; libaom explicitly sets visibility in shared object builds. Avoid
+        ; setting visibility to hidden as it may break builds that split
+        ; sources on e.g., directory boundaries.
+        %ifdef CHROMIUM
+            %xdefine %%VISIBILITY hidden
+        %else
+            %xdefine %%VISIBILITY
+        %endif
+    %else
+        %xdefine %%FUNCTION_PREFIX public_prefix
+        %xdefine %%VISIBILITY
+    %endif
+    %ifndef cglobaled_%2
+        %xdefine %2 mangle(%%FUNCTION_PREFIX %+ _ %+ %2)
+        %xdefine %2.skip_prologue %2 %+ .skip_prologue
+        CAT_XDEFINE cglobaled_, %2, 1
+    %endif
+    %xdefine current_function %2
+    %xdefine current_function_section __SECT__
+    %if FORMAT_ELF
+        global %2:function %%VISIBILITY
+    %elif FORMAT_MACHO
+        %ifdef __NASM_VER__
+            global %2
+        %else
+            global %2:private_extern
+        %endif
+    %else
+        global %2
+    %endif
+    align function_align
+    %2:
+    RESET_MM_PERMUTATION        ; needed for x86-64, also makes disassembly somewhat nicer
+    %xdefine rstk rsp           ; copy of the original stack pointer, used when greater alignment than the known stack alignment is required
+    %assign stack_offset 0      ; stack pointer offset relative to the return address
+    %assign stack_size 0        ; amount of stack space that can be freely used inside a function
+    %assign stack_size_padded 0 ; total amount of allocated stack space, including space for callee-saved xmm registers on WIN64 and alignment padding
+    %assign xmm_regs_used 0     ; number of XMM registers requested, used for dealing with callee-saved registers on WIN64
+    %ifnidn %3, ""
+        PROLOGUE %3
+    %endif
+%endmacro
+
+%macro cextern 1
+    %xdefine %1 mangle(private_prefix %+ _ %+ %1)
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+; like cextern, but without the prefix
+%macro cextern_naked 1
+    %ifdef PREFIX
+        %xdefine %1 mangle(%1)
+    %endif
+    CAT_XDEFINE cglobaled_, %1, 1
+    extern %1
+%endmacro
+
+%macro const 1-2+
+    %xdefine %1 mangle(private_prefix %+ _ %+ %1)
+    %if FORMAT_ELF
+        global %1:data hidden
+    %else
+        global %1
+    %endif
+    %1: %2
+%endmacro
+
+; This is needed for ELF, otherwise the GNU linker assumes the stack is executable by default.
+%if FORMAT_ELF
+    [SECTION .note.GNU-stack noalloc noexec nowrite progbits]
+%endif
+
+; Tell debuggers how large the function was.
+; This may be invoked multiple times per function; we rely on later instances overriding earlier ones.
+; This is invoked by RET and similar macros, and also cglobal does it for the previous function,
+; but if the last function in a source file doesn't use any of the standard macros for its epilogue,
+; then its size might be unspecified.
+%macro annotate_function_size 0
+    %ifdef __YASM_VER__
+        %ifdef current_function
+            %if FORMAT_ELF
+                current_function_section
+                %%ecf equ $
+                size current_function %%ecf - current_function
+                __SECT__
+            %endif
+        %endif
+    %endif
+%endmacro
+
+; cpuflags
+
+%assign cpuflags_mmx      (1<<0)
+%assign cpuflags_mmx2     (1<<1) | cpuflags_mmx
+%assign cpuflags_3dnow    (1<<2) | cpuflags_mmx
+%assign cpuflags_3dnowext (1<<3) | cpuflags_3dnow
+%assign cpuflags_sse      (1<<4) | cpuflags_mmx2
+%assign cpuflags_sse2     (1<<5) | cpuflags_sse
+%assign cpuflags_sse2slow (1<<6) | cpuflags_sse2
+%assign cpuflags_sse3     (1<<7) | cpuflags_sse2
+%assign cpuflags_ssse3    (1<<8) | cpuflags_sse3
+%assign cpuflags_sse4     (1<<9) | cpuflags_ssse3
+%assign cpuflags_sse42    (1<<10)| cpuflags_sse4
+%assign cpuflags_avx      (1<<11)| cpuflags_sse42
+%assign cpuflags_xop      (1<<12)| cpuflags_avx
+%assign cpuflags_fma4     (1<<13)| cpuflags_avx
+%assign cpuflags_fma3     (1<<14)| cpuflags_avx
+%assign cpuflags_avx2     (1<<15)| cpuflags_fma3
+
+%assign cpuflags_cache32  (1<<16)
+%assign cpuflags_cache64  (1<<17)
+%assign cpuflags_slowctz  (1<<18)
+%assign cpuflags_lzcnt    (1<<19)
+%assign cpuflags_aligned  (1<<20) ; not a cpu feature, but a function variant
+%assign cpuflags_atom     (1<<21)
+%assign cpuflags_bmi1     (1<<22)|cpuflags_lzcnt
+%assign cpuflags_bmi2     (1<<23)|cpuflags_bmi1
+
+; Returns a boolean value expressing whether or not the specified cpuflag is enabled.
+%define    cpuflag(x) (((((cpuflags & (cpuflags_ %+ x)) ^ (cpuflags_ %+ x)) - 1) >> 31) & 1)
+%define notcpuflag(x) (cpuflag(x) ^ 1)
+
+; Takes an arbitrary number of cpuflags from the above list.
+; All subsequent functions (up to the next INIT_CPUFLAGS) is built for the specified cpu.
+; You shouldn't need to invoke this macro directly, it's a subroutine for INIT_MMX &co.
+%macro INIT_CPUFLAGS 0-*
+    %xdefine SUFFIX
+    %undef cpuname
+    %assign cpuflags 0
+
+    %if %0 >= 1
+        %rep %0
+            %ifdef cpuname
+                %xdefine cpuname cpuname %+ _%1
+            %else
+                %xdefine cpuname %1
+            %endif
+            %assign cpuflags cpuflags | cpuflags_%1
+            %rotate 1
+        %endrep
+        %xdefine SUFFIX _ %+ cpuname
+
+        %if cpuflag(avx)
+            %assign avx_enabled 1
+        %endif
+        %if (mmsize == 16 && notcpuflag(sse2)) || (mmsize == 32 && notcpuflag(avx2))
+            %define mova movaps
+            %define movu movups
+            %define movnta movntps
+        %endif
+        %if cpuflag(aligned)
+            %define movu mova
+        %elif cpuflag(sse3) && notcpuflag(ssse3)
+            %define movu lddqu
+        %endif
+    %endif
+
+    %if ARCH_X86_64 || cpuflag(sse2)
+        %ifdef __NASM_VER__
+            ALIGNMODE k8
+        %else
+            CPU amdnop
+        %endif
+    %else
+        %ifdef __NASM_VER__
+            ALIGNMODE nop
+        %else
+            CPU basicnop
+        %endif
+    %endif
+%endmacro
+
+; Merge mmx and sse*
+; m# is a simd register of the currently selected size
+; xm# is the corresponding xmm register if mmsize >= 16, otherwise the same as m#
+; ym# is the corresponding ymm register if mmsize >= 32, otherwise the same as m#
+; (All 3 remain in sync through SWAP.)
+
+%macro CAT_XDEFINE 3
+    %xdefine %1%2 %3
+%endmacro
+
+%macro CAT_UNDEF 2
+    %undef %1%2
+%endmacro
+
+%macro INIT_MMX 0-1+
+    %assign avx_enabled 0
+    %define RESET_MM_PERMUTATION INIT_MMX %1
+    %define mmsize 8
+    %define num_mmregs 8
+    %define mova movq
+    %define movu movq
+    %define movh movd
+    %define movnta movntq
+    %assign %%i 0
+    %rep 8
+        CAT_XDEFINE m, %%i, mm %+ %%i
+        CAT_XDEFINE nnmm, %%i, %%i
+        %assign %%i %%i+1
+    %endrep
+    %rep 8
+        CAT_UNDEF m, %%i
+        CAT_UNDEF nnmm, %%i
+        %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_XMM 0-1+
+    %assign avx_enabled 0
+    %define RESET_MM_PERMUTATION INIT_XMM %1
+    %define mmsize 16
+    %define num_mmregs 8
+    %if ARCH_X86_64
+        %define num_mmregs 16
+    %endif
+    %define mova movdqa
+    %define movu movdqu
+    %define movh movq
+    %define movnta movntdq
+    %assign %%i 0
+    %rep num_mmregs
+        CAT_XDEFINE m, %%i, xmm %+ %%i
+        CAT_XDEFINE nnxmm, %%i, %%i
+        %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+%macro INIT_YMM 0-1+
+    %assign avx_enabled 1
+    %define RESET_MM_PERMUTATION INIT_YMM %1
+    %define mmsize 32
+    %define num_mmregs 8
+    %if ARCH_X86_64
+        %define num_mmregs 16
+    %endif
+    %define mova movdqa
+    %define movu movdqu
+    %undef movh
+    %define movnta movntdq
+    %assign %%i 0
+    %rep num_mmregs
+        CAT_XDEFINE m, %%i, ymm %+ %%i
+        CAT_XDEFINE nnymm, %%i, %%i
+        %assign %%i %%i+1
+    %endrep
+    INIT_CPUFLAGS %1
+%endmacro
+
+INIT_XMM
+
+%macro DECLARE_MMCAST 1
+    %define  mmmm%1   mm%1
+    %define  mmxmm%1  mm%1
+    %define  mmymm%1  mm%1
+    %define xmmmm%1   mm%1
+    %define xmmxmm%1 xmm%1
+    %define xmmymm%1 xmm%1
+    %define ymmmm%1   mm%1
+    %define ymmxmm%1 xmm%1
+    %define ymmymm%1 ymm%1
+    %define xm%1 xmm %+ m%1
+    %define ym%1 ymm %+ m%1
+%endmacro
+
+%assign i 0
+%rep 16
+    DECLARE_MMCAST i
+    %assign i i+1
+%endrep
+
+; I often want to use macros that permute their arguments. e.g. there's no
+; efficient way to implement butterfly or transpose or dct without swapping some
+; arguments.
+;
+; I would like to not have to manually keep track of the permutations:
+; If I insert a permutation in the middle of a function, it should automatically
+; change everything that follows. For more complex macros I may also have multiple
+; implementations, e.g. the SSE2 and SSSE3 versions may have different permutations.
+;
+; Hence these macros. Insert a PERMUTE or some SWAPs at the end of a macro that
+; permutes its arguments. It's equivalent to exchanging the contents of the
+; registers, except that this way you exchange the register names instead, so it
+; doesn't cost any cycles.
+
+%macro PERMUTE 2-* ; takes a list of pairs to swap
+    %rep %0/2
+        %xdefine %%tmp%2 m%2
+        %rotate 2
+    %endrep
+    %rep %0/2
+        %xdefine m%1 %%tmp%2
+        CAT_XDEFINE nn, m%1, %1
+        %rotate 2
+    %endrep
+%endmacro
+
+%macro SWAP 2+ ; swaps a single chain (sometimes more concise than pairs)
+    %ifnum %1 ; SWAP 0, 1, ...
+        SWAP_INTERNAL_NUM %1, %2
+    %else ; SWAP m0, m1, ...
+        SWAP_INTERNAL_NAME %1, %2
+    %endif
+%endmacro
+
+%macro SWAP_INTERNAL_NUM 2-*
+    %rep %0-1
+        %xdefine %%tmp m%1
+        %xdefine m%1 m%2
+        %xdefine m%2 %%tmp
+        CAT_XDEFINE nn, m%1, %1
+        CAT_XDEFINE nn, m%2, %2
+        %rotate 1
+    %endrep
+%endmacro
+
+%macro SWAP_INTERNAL_NAME 2-*
+    %xdefine %%args nn %+ %1
+    %rep %0-1
+        %xdefine %%args %%args, nn %+ %2
+        %rotate 1
+    %endrep
+    SWAP_INTERNAL_NUM %%args
+%endmacro
+
+; If SAVE_MM_PERMUTATION is placed at the end of a function, then any later
+; calls to that function will automatically load the permutation, so values can
+; be returned in mmregs.
+%macro SAVE_MM_PERMUTATION 0-1
+    %if %0
+        %xdefine %%f %1_m
+    %else
+        %xdefine %%f current_function %+ _m
+    %endif
+    %assign %%i 0
+    %rep num_mmregs
+        CAT_XDEFINE %%f, %%i, m %+ %%i
+        %assign %%i %%i+1
+    %endrep
+%endmacro
+
+%macro LOAD_MM_PERMUTATION 1 ; name to load from
+    %ifdef %1_m0
+        %assign %%i 0
+        %rep num_mmregs
+            CAT_XDEFINE m, %%i, %1_m %+ %%i
+            CAT_XDEFINE nn, m %+ %%i, %%i
+            %assign %%i %%i+1
+        %endrep
+    %endif
+%endmacro
+
+; Append cpuflags to the callee's name iff the appended name is known and the plain name isn't
+%macro call 1
+    call_internal %1 %+ SUFFIX, %1
+%endmacro
+%macro call_internal 2
+    %xdefine %%i %2
+    %ifndef cglobaled_%2
+        %ifdef cglobaled_%1
+            %xdefine %%i %1
+        %endif
+    %endif
+    call %%i
+    LOAD_MM_PERMUTATION %%i
+%endmacro
+
+; Substitutions that reduce instruction size but are functionally equivalent
+%macro add 2
+    %ifnum %2
+        %if %2==128
+            sub %1, -128
+        %else
+            add %1, %2
+        %endif
+    %else
+        add %1, %2
+    %endif
+%endmacro
+
+%macro sub 2
+    %ifnum %2
+        %if %2==128
+            add %1, -128
+        %else
+            sub %1, %2
+        %endif
+    %else
+        sub %1, %2
+    %endif
+%endmacro
+
+;=============================================================================
+; AVX abstraction layer
+;=============================================================================
+
+%assign i 0
+%rep 16
+    %if i < 8
+        CAT_XDEFINE sizeofmm, i, 8
+    %endif
+    CAT_XDEFINE sizeofxmm, i, 16
+    CAT_XDEFINE sizeofymm, i, 32
+    %assign i i+1
+%endrep
+%undef i
+
+%macro CHECK_AVX_INSTR_EMU 3-*
+    %xdefine %%opcode %1
+    %xdefine %%dst %2
+    %rep %0-2
+        %ifidn %%dst, %3
+            %error non-avx emulation of ``%%opcode'' is not supported
+        %endif
+        %rotate 1
+    %endrep
+%endmacro
+
+;%1 == instruction
+;%2 == minimal instruction set
+;%3 == 1 if float, 0 if int
+;%4 == 1 if non-destructive or 4-operand (xmm, xmm, xmm, imm), 0 otherwise
+;%5 == 1 if commutative (i.e. doesn't matter which src arg is which), 0 if not
+;%6+: operands
+%macro RUN_AVX_INSTR 6-9+
+    %ifnum sizeof%7
+        %assign __sizeofreg sizeof%7
+    %elifnum sizeof%6
+        %assign __sizeofreg sizeof%6
+    %else
+        %assign __sizeofreg mmsize
+    %endif
+    %assign __emulate_avx 0
+    %if avx_enabled && __sizeofreg >= 16
+        %xdefine __instr v%1
+    %else
+        %xdefine __instr %1
+        %if %0 >= 8+%4
+            %assign __emulate_avx 1
+        %endif
+    %endif
+    %ifnidn %2, fnord
+        %ifdef cpuname
+            %if notcpuflag(%2)
+                %error use of ``%1'' %2 instruction in cpuname function: current_function
+            %elif cpuflags_%2 < cpuflags_sse && notcpuflag(sse2) && __sizeofreg > 8
+                %error use of ``%1'' sse2 instruction in cpuname function: current_function
+            %endif
+        %endif
+    %endif
+
+    %if __emulate_avx
+        %xdefine __src1 %7
+        %xdefine __src2 %8
+        %ifnidn %6, %7
+            %if %0 >= 9
+                CHECK_AVX_INSTR_EMU {%1 %6, %7, %8, %9}, %6, %8, %9
+            %else
+                CHECK_AVX_INSTR_EMU {%1 %6, %7, %8}, %6, %8
+            %endif
+            %if %5 && %4 == 0
+                %ifnid %8
+                    ; 3-operand AVX instructions with a memory arg can only have it in src2,
+                    ; whereas SSE emulation prefers to have it in src1 (i.e. the mov).
+                    ; So, if the instruction is commutative with a memory arg, swap them.
+                    %xdefine __src1 %8
+                    %xdefine __src2 %7
+                %endif
+            %endif
+            %if __sizeofreg == 8
+                MOVQ %6, __src1
+            %elif %3
+                MOVAPS %6, __src1
+            %else
+                MOVDQA %6, __src1
+            %endif
+        %endif
+        %if %0 >= 9
+            %1 %6, __src2, %9
+        %else
+            %1 %6, __src2
+        %endif
+    %elif %0 >= 9
+        __instr %6, %7, %8, %9
+    %elif %0 == 8
+        __instr %6, %7, %8
+    %elif %0 == 7
+        __instr %6, %7
+    %else
+        __instr %6
+    %endif
+%endmacro
+
+;%1 == instruction
+;%2 == minimal instruction set
+;%3 == 1 if float, 0 if int
+;%4 == 1 if non-destructive or 4-operand (xmm, xmm, xmm, imm), 0 otherwise
+;%5 == 1 if commutative (i.e. doesn't matter which src arg is which), 0 if not
+%macro AVX_INSTR 1-5 fnord, 0, 1, 0
+    %macro %1 1-10 fnord, fnord, fnord, fnord, %1, %2, %3, %4, %5
+        %ifidn %2, fnord
+            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1
+        %elifidn %3, fnord
+            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2
+        %elifidn %4, fnord
+            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2, %3
+        %elifidn %5, fnord
+            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2, %3, %4
+        %else
+            RUN_AVX_INSTR %6, %7, %8, %9, %10, %1, %2, %3, %4, %5
+        %endif
+    %endmacro
+%endmacro
+
+; Instructions with both VEX and non-VEX encodings
+; Non-destructive instructions are written without parameters
+AVX_INSTR addpd, sse2, 1, 0, 1
+AVX_INSTR addps, sse, 1, 0, 1
+AVX_INSTR addsd, sse2, 1, 0, 1
+AVX_INSTR addss, sse, 1, 0, 1
+AVX_INSTR addsubpd, sse3, 1, 0, 0
+AVX_INSTR addsubps, sse3, 1, 0, 0
+AVX_INSTR aesdec, fnord, 0, 0, 0
+AVX_INSTR aesdeclast, fnord, 0, 0, 0
+AVX_INSTR aesenc, fnord, 0, 0, 0
+AVX_INSTR aesenclast, fnord, 0, 0, 0
+AVX_INSTR aesimc
+AVX_INSTR aeskeygenassist
+AVX_INSTR andnpd, sse2, 1, 0, 0
+AVX_INSTR andnps, sse, 1, 0, 0
+AVX_INSTR andpd, sse2, 1, 0, 1
+AVX_INSTR andps, sse, 1, 0, 1
+AVX_INSTR blendpd, sse4, 1, 0, 0
+AVX_INSTR blendps, sse4, 1, 0, 0
+AVX_INSTR blendvpd, sse4, 1, 0, 0
+AVX_INSTR blendvps, sse4, 1, 0, 0
+AVX_INSTR cmppd, sse2, 1, 1, 0
+AVX_INSTR cmpps, sse, 1, 1, 0
+AVX_INSTR cmpsd, sse2, 1, 1, 0
+AVX_INSTR cmpss, sse, 1, 1, 0
+AVX_INSTR comisd, sse2
+AVX_INSTR comiss, sse
+AVX_INSTR cvtdq2pd, sse2
+AVX_INSTR cvtdq2ps, sse2
+AVX_INSTR cvtpd2dq, sse2
+AVX_INSTR cvtpd2ps, sse2
+AVX_INSTR cvtps2dq, sse2
+AVX_INSTR cvtps2pd, sse2
+AVX_INSTR cvtsd2si, sse2
+AVX_INSTR cvtsd2ss, sse2
+AVX_INSTR cvtsi2sd, sse2
+AVX_INSTR cvtsi2ss, sse
+AVX_INSTR cvtss2sd, sse2
+AVX_INSTR cvtss2si, sse
+AVX_INSTR cvttpd2dq, sse2
+AVX_INSTR cvttps2dq, sse2
+AVX_INSTR cvttsd2si, sse2
+AVX_INSTR cvttss2si, sse
+AVX_INSTR divpd, sse2, 1, 0, 0
+AVX_INSTR divps, sse, 1, 0, 0
+AVX_INSTR divsd, sse2, 1, 0, 0
+AVX_INSTR divss, sse, 1, 0, 0
+AVX_INSTR dppd, sse4, 1, 1, 0
+AVX_INSTR dpps, sse4, 1, 1, 0
+AVX_INSTR extractps, sse4
+AVX_INSTR haddpd, sse3, 1, 0, 0
+AVX_INSTR haddps, sse3, 1, 0, 0
+AVX_INSTR hsubpd, sse3, 1, 0, 0
+AVX_INSTR hsubps, sse3, 1, 0, 0
+AVX_INSTR insertps, sse4, 1, 1, 0
+AVX_INSTR lddqu, sse3
+AVX_INSTR ldmxcsr, sse
+AVX_INSTR maskmovdqu, sse2
+AVX_INSTR maxpd, sse2, 1, 0, 1
+AVX_INSTR maxps, sse, 1, 0, 1
+AVX_INSTR maxsd, sse2, 1, 0, 1
+AVX_INSTR maxss, sse, 1, 0, 1
+AVX_INSTR minpd, sse2, 1, 0, 1
+AVX_INSTR minps, sse, 1, 0, 1
+AVX_INSTR minsd, sse2, 1, 0, 1
+AVX_INSTR minss, sse, 1, 0, 1
+AVX_INSTR movapd, sse2
+AVX_INSTR movaps, sse
+AVX_INSTR movd, mmx
+AVX_INSTR movddup, sse3
+AVX_INSTR movdqa, sse2
+AVX_INSTR movdqu, sse2
+AVX_INSTR movhlps, sse, 1, 0, 0
+AVX_INSTR movhpd, sse2, 1, 0, 0
+AVX_INSTR movhps, sse, 1, 0, 0
+AVX_INSTR movlhps, sse, 1, 0, 0
+AVX_INSTR movlpd, sse2, 1, 0, 0
+AVX_INSTR movlps, sse, 1, 0, 0
+AVX_INSTR movmskpd, sse2
+AVX_INSTR movmskps, sse
+AVX_INSTR movntdq, sse2
+AVX_INSTR movntdqa, sse4
+AVX_INSTR movntpd, sse2
+AVX_INSTR movntps, sse
+AVX_INSTR movq, mmx
+AVX_INSTR movsd, sse2, 1, 0, 0
+AVX_INSTR movshdup, sse3
+AVX_INSTR movsldup, sse3
+AVX_INSTR movss, sse, 1, 0, 0
+AVX_INSTR movupd, sse2
+AVX_INSTR movups, sse
+AVX_INSTR mpsadbw, sse4
+AVX_INSTR mulpd, sse2, 1, 0, 1
+AVX_INSTR mulps, sse, 1, 0, 1
+AVX_INSTR mulsd, sse2, 1, 0, 1
+AVX_INSTR mulss, sse, 1, 0, 1
+AVX_INSTR orpd, sse2, 1, 0, 1
+AVX_INSTR orps, sse, 1, 0, 1
+AVX_INSTR pabsb, ssse3
+AVX_INSTR pabsd, ssse3
+AVX_INSTR pabsw, ssse3
+AVX_INSTR packsswb, mmx, 0, 0, 0
+AVX_INSTR packssdw, mmx, 0, 0, 0
+AVX_INSTR packuswb, mmx, 0, 0, 0
+AVX_INSTR packusdw, sse4, 0, 0, 0
+AVX_INSTR paddb, mmx, 0, 0, 1
+AVX_INSTR paddw, mmx, 0, 0, 1
+AVX_INSTR paddd, mmx, 0, 0, 1
+AVX_INSTR paddq, sse2, 0, 0, 1
+AVX_INSTR paddsb, mmx, 0, 0, 1
+AVX_INSTR paddsw, mmx, 0, 0, 1
+AVX_INSTR paddusb, mmx, 0, 0, 1
+AVX_INSTR paddusw, mmx, 0, 0, 1
+AVX_INSTR palignr, ssse3
+AVX_INSTR pand, mmx, 0, 0, 1
+AVX_INSTR pandn, mmx, 0, 0, 0
+AVX_INSTR pavgb, mmx2, 0, 0, 1
+AVX_INSTR pavgw, mmx2, 0, 0, 1
+AVX_INSTR pblendvb, sse4, 0, 0, 0
+AVX_INSTR pblendw, sse4
+AVX_INSTR pclmulqdq
+AVX_INSTR pcmpestri, sse42
+AVX_INSTR pcmpestrm, sse42
+AVX_INSTR pcmpistri, sse42
+AVX_INSTR pcmpistrm, sse42
+AVX_INSTR pcmpeqb, mmx, 0, 0, 1
+AVX_INSTR pcmpeqw, mmx, 0, 0, 1
+AVX_INSTR pcmpeqd, mmx, 0, 0, 1
+AVX_INSTR pcmpeqq, sse4, 0, 0, 1
+AVX_INSTR pcmpgtb, mmx, 0, 0, 0
+AVX_INSTR pcmpgtw, mmx, 0, 0, 0
+AVX_INSTR pcmpgtd, mmx, 0, 0, 0
+AVX_INSTR pcmpgtq, sse42, 0, 0, 0
+AVX_INSTR pextrb, sse4
+AVX_INSTR pextrd, sse4
+AVX_INSTR pextrq, sse4
+AVX_INSTR pextrw, mmx2
+AVX_INSTR phaddw, ssse3, 0, 0, 0
+AVX_INSTR phaddd, ssse3, 0, 0, 0
+AVX_INSTR phaddsw, ssse3, 0, 0, 0
+AVX_INSTR phminposuw, sse4
+AVX_INSTR phsubw, ssse3, 0, 0, 0
+AVX_INSTR phsubd, ssse3, 0, 0, 0
+AVX_INSTR phsubsw, ssse3, 0, 0, 0
+AVX_INSTR pinsrb, sse4
+AVX_INSTR pinsrd, sse4
+AVX_INSTR pinsrq, sse4
+AVX_INSTR pinsrw, mmx2
+AVX_INSTR pmaddwd, mmx, 0, 0, 1
+AVX_INSTR pmaddubsw, ssse3, 0, 0, 0
+AVX_INSTR pmaxsb, sse4, 0, 0, 1
+AVX_INSTR pmaxsw, mmx2, 0, 0, 1
+AVX_INSTR pmaxsd, sse4, 0, 0, 1
+AVX_INSTR pmaxub, mmx2, 0, 0, 1
+AVX_INSTR pmaxuw, sse4, 0, 0, 1
+AVX_INSTR pmaxud, sse4, 0, 0, 1
+AVX_INSTR pminsb, sse4, 0, 0, 1
+AVX_INSTR pminsw, mmx2, 0, 0, 1
+AVX_INSTR pminsd, sse4, 0, 0, 1
+AVX_INSTR pminub, mmx2, 0, 0, 1
+AVX_INSTR pminuw, sse4, 0, 0, 1
+AVX_INSTR pminud, sse4, 0, 0, 1
+AVX_INSTR pmovmskb, mmx2
+AVX_INSTR pmovsxbw, sse4
+AVX_INSTR pmovsxbd, sse4
+AVX_INSTR pmovsxbq, sse4
+AVX_INSTR pmovsxwd, sse4
+AVX_INSTR pmovsxwq, sse4
+AVX_INSTR pmovsxdq, sse4
+AVX_INSTR pmovzxbw, sse4
+AVX_INSTR pmovzxbd, sse4
+AVX_INSTR pmovzxbq, sse4
+AVX_INSTR pmovzxwd, sse4
+AVX_INSTR pmovzxwq, sse4
+AVX_INSTR pmovzxdq, sse4
+AVX_INSTR pmuldq, sse4, 0, 0, 1
+AVX_INSTR pmulhrsw, ssse3, 0, 0, 1
+AVX_INSTR pmulhuw, mmx2, 0, 0, 1
+AVX_INSTR pmulhw, mmx, 0, 0, 1
+AVX_INSTR pmullw, mmx, 0, 0, 1
+AVX_INSTR pmulld, sse4, 0, 0, 1
+AVX_INSTR pmuludq, sse2, 0, 0, 1
+AVX_INSTR por, mmx, 0, 0, 1
+AVX_INSTR psadbw, mmx2, 0, 0, 1
+AVX_INSTR pshufb, ssse3, 0, 0, 0
+AVX_INSTR pshufd, sse2
+AVX_INSTR pshufhw, sse2
+AVX_INSTR pshuflw, sse2
+AVX_INSTR psignb, ssse3, 0, 0, 0
+AVX_INSTR psignw, ssse3, 0, 0, 0
+AVX_INSTR psignd, ssse3, 0, 0, 0
+AVX_INSTR psllw, mmx, 0, 0, 0
+AVX_INSTR pslld, mmx, 0, 0, 0
+AVX_INSTR psllq, mmx, 0, 0, 0
+AVX_INSTR pslldq, sse2, 0, 0, 0
+AVX_INSTR psraw, mmx, 0, 0, 0
+AVX_INSTR psrad, mmx, 0, 0, 0
+AVX_INSTR psrlw, mmx, 0, 0, 0
+AVX_INSTR psrld, mmx, 0, 0, 0
+AVX_INSTR psrlq, mmx, 0, 0, 0
+AVX_INSTR psrldq, sse2, 0, 0, 0
+AVX_INSTR psubb, mmx, 0, 0, 0
+AVX_INSTR psubw, mmx, 0, 0, 0
+AVX_INSTR psubd, mmx, 0, 0, 0
+AVX_INSTR psubq, sse2, 0, 0, 0
+AVX_INSTR psubsb, mmx, 0, 0, 0
+AVX_INSTR psubsw, mmx, 0, 0, 0
+AVX_INSTR psubusb, mmx, 0, 0, 0
+AVX_INSTR psubusw, mmx, 0, 0, 0
+AVX_INSTR ptest, sse4
+AVX_INSTR punpckhbw, mmx, 0, 0, 0
+AVX_INSTR punpckhwd, mmx, 0, 0, 0
+AVX_INSTR punpckhdq, mmx, 0, 0, 0
+AVX_INSTR punpckhqdq, sse2, 0, 0, 0
+AVX_INSTR punpcklbw, mmx, 0, 0, 0
+AVX_INSTR punpcklwd, mmx, 0, 0, 0
+AVX_INSTR punpckldq, mmx, 0, 0, 0
+AVX_INSTR punpcklqdq, sse2, 0, 0, 0
+AVX_INSTR pxor, mmx, 0, 0, 1
+AVX_INSTR rcpps, sse, 1, 0, 0
+AVX_INSTR rcpss, sse, 1, 0, 0
+AVX_INSTR roundpd, sse4
+AVX_INSTR roundps, sse4
+AVX_INSTR roundsd, sse4
+AVX_INSTR roundss, sse4
+AVX_INSTR rsqrtps, sse, 1, 0, 0
+AVX_INSTR rsqrtss, sse, 1, 0, 0
+AVX_INSTR shufpd, sse2, 1, 1, 0
+AVX_INSTR shufps, sse, 1, 1, 0
+AVX_INSTR sqrtpd, sse2, 1, 0, 0
+AVX_INSTR sqrtps, sse, 1, 0, 0
+AVX_INSTR sqrtsd, sse2, 1, 0, 0
+AVX_INSTR sqrtss, sse, 1, 0, 0
+AVX_INSTR stmxcsr, sse
+AVX_INSTR subpd, sse2, 1, 0, 0
+AVX_INSTR subps, sse, 1, 0, 0
+AVX_INSTR subsd, sse2, 1, 0, 0
+AVX_INSTR subss, sse, 1, 0, 0
+AVX_INSTR ucomisd, sse2
+AVX_INSTR ucomiss, sse
+AVX_INSTR unpckhpd, sse2, 1, 0, 0
+AVX_INSTR unpckhps, sse, 1, 0, 0
+AVX_INSTR unpcklpd, sse2, 1, 0, 0
+AVX_INSTR unpcklps, sse, 1, 0, 0
+AVX_INSTR xorpd, sse2, 1, 0, 1
+AVX_INSTR xorps, sse, 1, 0, 1
+
+; 3DNow instructions, for sharing code between AVX, SSE and 3DN
+AVX_INSTR pfadd, 3dnow, 1, 0, 1
+AVX_INSTR pfsub, 3dnow, 1, 0, 0
+AVX_INSTR pfmul, 3dnow, 1, 0, 1
+
+; base-4 constants for shuffles
+%assign i 0
+%rep 256
+    %assign j ((i>>6)&3)*1000 + ((i>>4)&3)*100 + ((i>>2)&3)*10 + (i&3)
+    %if j < 10
+        CAT_XDEFINE q000, j, i
+    %elif j < 100
+        CAT_XDEFINE q00, j, i
+    %elif j < 1000
+        CAT_XDEFINE q0, j, i
+    %else
+        CAT_XDEFINE q, j, i
+    %endif
+    %assign i i+1
+%endrep
+%undef i
+%undef j
+
+%macro FMA_INSTR 3
+    %macro %1 4-7 %1, %2, %3
+        %if cpuflag(xop)
+            v%5 %1, %2, %3, %4
+        %elifnidn %1, %4
+            %6 %1, %2, %3
+            %7 %1, %4
+        %else
+            %error non-xop emulation of ``%5 %1, %2, %3, %4'' is not supported
+        %endif
+    %endmacro
+%endmacro
+
+FMA_INSTR  pmacsww,  pmullw, paddw
+FMA_INSTR  pmacsdd,  pmulld, paddd ; sse4 emulation
+FMA_INSTR pmacsdql,  pmuldq, paddq ; sse4 emulation
+FMA_INSTR pmadcswd, pmaddwd, paddd
+
+; Macros for consolidating FMA3 and FMA4 using 4-operand (dst, src1, src2, src3) syntax.
+; FMA3 is only possible if dst is the same as one of the src registers.
+; Either src2 or src3 can be a memory operand.
+%macro FMA4_INSTR 2-*
+    %push fma4_instr
+    %xdefine %$prefix %1
+    %rep %0 - 1
+        %macro %$prefix%2 4-6 %$prefix, %2
+            %if notcpuflag(fma3) && notcpuflag(fma4)
+                %error use of ``%5%6'' fma instruction in cpuname function: current_function
+            %elif cpuflag(fma4)
+                v%5%6 %1, %2, %3, %4
+            %elifidn %1, %2
+                ; If %3 or %4 is a memory operand it needs to be encoded as the last operand.
+                %ifid %3
+                    v%{5}213%6 %2, %3, %4
+                %else
+                    v%{5}132%6 %2, %4, %3
+                %endif
+            %elifidn %1, %3
+                v%{5}213%6 %3, %2, %4
+            %elifidn %1, %4
+                v%{5}231%6 %4, %2, %3
+            %else
+                %error fma3 emulation of ``%5%6 %1, %2, %3, %4'' is not supported
+            %endif
+        %endmacro
+        %rotate 1
+    %endrep
+    %pop
+%endmacro
+
+FMA4_INSTR fmadd,    pd, ps, sd, ss
+FMA4_INSTR fmaddsub, pd, ps
+FMA4_INSTR fmsub,    pd, ps, sd, ss
+FMA4_INSTR fmsubadd, pd, ps
+FMA4_INSTR fnmadd,   pd, ps, sd, ss
+FMA4_INSTR fnmsub,   pd, ps, sd, ss
+
+; workaround: vpbroadcastq is broken in x86_32 due to a yasm bug (fixed in 1.3.0)
+%ifdef __YASM_VER__
+    %if __YASM_VERSION_ID__ < 0x01030000 && ARCH_X86_64 == 0
+        %macro vpbroadcastq 2
+            %if sizeof%1 == 16
+                movddup %1, %2
+            %else
+                vbroadcastsd %1, %2
+            %endif
+        %endmacro
+    %endif
+%endif
diff --git a/third_party/aom/tools/all_builds.py b/third_party/aom/tools/all_builds.py
new file mode 100755
index 0000000000..d1f0c80c03
--- /dev/null
+++ b/third_party/aom/tools/all_builds.py
@@ -0,0 +1,72 @@
+#!/usr/bin/python
+
+import getopt
+import subprocess
+import sys
+
+LONG_OPTIONS = ["shard=", "shards="]
+BASE_COMMAND = "./configure --enable-internal-stats --enable-experimental"
+
+def RunCommand(command):
+  run = subprocess.Popen(command, shell=True)
+  output = run.communicate()
+  if run.returncode:
+    print "Non-zero return code: " + str(run.returncode) + " => exiting!"
+    sys.exit(1)
+
+def list_of_experiments():
+  experiments = []
+  configure_file = open("configure")
+  list_start = False
+  for line in configure_file.read().split("\n"):
+    if line == 'EXPERIMENT_LIST="':
+      list_start = True
+    elif line == '"':
+      list_start = False
+    elif list_start:
+      currently_broken = ["csm"]
+      experiment = line[4:]
+      if experiment not in currently_broken:
+        experiments.append(experiment)
+  return experiments
+
+def main(argv):
+  # Parse arguments
+  options = {"--shard": 0, "--shards": 1}
+  if "--" in argv:
+    opt_end_index = argv.index("--")
+  else:
+    opt_end_index = len(argv)
+  try:
+    o, _ = getopt.getopt(argv[1:opt_end_index], None, LONG_OPTIONS)
+  except getopt.GetoptError, err:
+    print str(err)
+    print "Usage: %s [--shard=<n> --shards=<n>] -- [configure flag ...]"%argv[0]
+    sys.exit(2)
+
+  options.update(o)
+  extra_args = argv[opt_end_index + 1:]
+
+  # Shard experiment list
+  shard = int(options["--shard"])
+  shards = int(options["--shards"])
+  experiments = list_of_experiments()
+  base_command = " ".join([BASE_COMMAND] + extra_args)
+  configs = [base_command]
+  configs += ["%s --enable-%s" % (base_command, e) for e in experiments]
+  my_configs = zip(configs, range(len(configs)))
+  my_configs = filter(lambda x: x[1] % shards == shard, my_configs)
+  my_configs = [e[0] for e in my_configs]
+
+  # Run configs for this shard
+  for config in my_configs:
+    test_build(config)
+
+def test_build(configure_command):
+  print "\033[34m\033[47mTesting %s\033[0m" % (configure_command)
+  RunCommand(configure_command)
+  RunCommand("make clean")
+  RunCommand("make")
+
+if __name__ == "__main__":
+  main(sys.argv)
diff --git a/third_party/aom/tools/author_first_release.sh b/third_party/aom/tools/author_first_release.sh
new file mode 100755
index 0000000000..7b0b797212
--- /dev/null
+++ b/third_party/aom/tools/author_first_release.sh
@@ -0,0 +1,15 @@
+#!/bin/bash
+##
+## List the release each author first contributed to.
+##
+## Usage: author_first_release.sh [TAGS]
+##
+## If the TAGS arguments are unspecified, all tags reported by `git tag`
+## will be considered.
+##
+tags=${@:-$(git tag)}
+for tag in $tags; do
+  git shortlog -n -e -s $tag |
+      cut -f2- |
+      awk "{print \"${tag#v}\t\"\$0}"
+done | sort -k2  | uniq -f2
diff --git a/third_party/aom/tools/build_inspector.sh b/third_party/aom/tools/build_inspector.sh
new file mode 100755
index 0000000000..a67aa15f6c
--- /dev/null
+++ b/third_party/aom/tools/build_inspector.sh
@@ -0,0 +1,16 @@
+if ! [ -x "$(command -v emcc)" ] || ! [ -x "$(command -v emconfigure)" ] || ! [ -x "$(command -v emmake)" ]; then
+  echo 'Emscripten SDK is not available (emcc, emconfigure or emmake is missing). Install it from https://kripken.github.io/emscripten-site/docs/getting_started/downloads.html and try again.' >&2
+  exit 1
+fi
+
+echo 'Building JS Inspector'
+if [ ! -d ".inspect" ]; then
+  mkdir .inspect
+  cd .inspect && emconfigure ../../configure --disable-multithread --disable-runtime-cpu-detect --target=generic-gnu --enable-accounting --enable-inspection --enable-aom_highbitdepth --extra-cflags="-D_POSIX_SOURCE"
+fi
+
+cd .inspect
+emmake make -j 8
+cp examples/inspect inspect.bc
+emcc -O3 inspect.bc -o inspect.js -s TOTAL_MEMORY=134217728 -s MODULARIZE=1 -s EXPORT_NAME="'DecoderModule'" --post-js "../inspect-post.js" --memory-init-file 0
+cp inspect.js ../inspect.js
diff --git a/third_party/aom/tools/cpplint.py b/third_party/aom/tools/cpplint.py
new file mode 100755
index 0000000000..25fbef73d8
--- /dev/null
+++ b/third_party/aom/tools/cpplint.py
@@ -0,0 +1,4756 @@
+#!/usr/bin/python
+#
+# Copyright (c) 2009 Google Inc. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#    * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#    * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#    * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Does google-lint on c++ files.
+
+The goal of this script is to identify places in the code that *may*
+be in non-compliance with google style.  It does not attempt to fix
+up these problems -- the point is to educate.  It does also not
+attempt to find all problems, or to ensure that everything it does
+find is legitimately a problem.
+
+In particular, we can get very confused by /* and // inside strings!
+We do a small hack, which is to ignore //'s with "'s after them on the
+same line, but it is far from perfect (in either direction).
+"""
+
+import codecs
+import copy
+import getopt
+import math  # for log
+import os
+import re
+import sre_compile
+import string
+import sys
+import unicodedata
+
+
+_USAGE = """
+Syntax: cpplint.py [--verbose=#] [--output=vs7] [--filter=-x,+y,...]
+                   [--counting=total|toplevel|detailed] [--root=subdir]
+                   [--linelength=digits]
+        <file> [file] ...
+
+  The style guidelines this tries to follow are those in
+    http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml
+
+  Every problem is given a confidence score from 1-5, with 5 meaning we are
+  certain of the problem, and 1 meaning it could be a legitimate construct.
+  This will miss some errors, and is not a substitute for a code review.
+
+  To suppress false-positive errors of a certain category, add a
+  'NOLINT(category)' comment to the line.  NOLINT or NOLINT(*)
+  suppresses errors of all categories on that line.
+
+  The files passed in will be linted; at least one file must be provided.
+  Default linted extensions are .cc, .cpp, .cu, .cuh and .h.  Change the
+  extensions with the --extensions flag.
+
+  Flags:
+
+    output=vs7
+      By default, the output is formatted to ease emacs parsing.  Visual Studio
+      compatible output (vs7) may also be used.  Other formats are unsupported.
+
+    verbose=#
+      Specify a number 0-5 to restrict errors to certain verbosity levels.
+
+    filter=-x,+y,...
+      Specify a comma-separated list of category-filters to apply: only
+      error messages whose category names pass the filters will be printed.
+      (Category names are printed with the message and look like
+      "[whitespace/indent]".)  Filters are evaluated left to right.
+      "-FOO" and "FOO" means "do not print categories that start with FOO".
+      "+FOO" means "do print categories that start with FOO".
+
+      Examples: --filter=-whitespace,+whitespace/braces
+                --filter=whitespace,runtime/printf,+runtime/printf_format
+                --filter=-,+build/include_what_you_use
+
+      To see a list of all the categories used in cpplint, pass no arg:
+         --filter=
+
+    counting=total|toplevel|detailed
+      The total number of errors found is always printed. If
+      'toplevel' is provided, then the count of errors in each of
+      the top-level categories like 'build' and 'whitespace' will
+      also be printed. If 'detailed' is provided, then a count
+      is provided for each category like 'build/class'.
+
+    root=subdir
+      The root directory used for deriving header guard CPP variable.
+      By default, the header guard CPP variable is calculated as the relative
+      path to the directory that contains .git, .hg, or .svn.  When this flag
+      is specified, the relative path is calculated from the specified
+      directory. If the specified directory does not exist, this flag is
+      ignored.
+
+      Examples:
+        Assuing that src/.git exists, the header guard CPP variables for
+        src/chrome/browser/ui/browser.h are:
+
+        No flag => CHROME_BROWSER_UI_BROWSER_H_
+        --root=chrome => BROWSER_UI_BROWSER_H_
+        --root=chrome/browser => UI_BROWSER_H_
+
+    linelength=digits
+      This is the allowed line length for the project. The default value is
+      80 characters.
+
+      Examples:
+        --linelength=120
+
+    extensions=extension,extension,...
+      The allowed file extensions that cpplint will check
+
+      Examples:
+        --extensions=hpp,cpp
+"""
+
+# We categorize each error message we print.  Here are the categories.
+# We want an explicit list so we can list them all in cpplint --filter=.
+# If you add a new error message with a new category, add it to the list
+# here!  cpplint_unittest.py should tell you if you forget to do this.
+_ERROR_CATEGORIES = [
+  'build/class',
+  'build/deprecated',
+  'build/endif_comment',
+  'build/explicit_make_pair',
+  'build/forward_decl',
+  'build/header_guard',
+  'build/include',
+  'build/include_alpha',
+  'build/include_order',
+  'build/include_what_you_use',
+  'build/namespaces',
+  'build/printf_format',
+  'build/storage_class',
+  'legal/copyright',
+  'readability/alt_tokens',
+  'readability/braces',
+  'readability/casting',
+  'readability/check',
+  'readability/constructors',
+  'readability/fn_size',
+  'readability/function',
+  'readability/multiline_comment',
+  'readability/multiline_string',
+  'readability/namespace',
+  'readability/nolint',
+  'readability/nul',
+  'readability/streams',
+  'readability/todo',
+  'readability/utf8',
+  'runtime/arrays',
+  'runtime/casting',
+  'runtime/explicit',
+  'runtime/int',
+  'runtime/init',
+  'runtime/invalid_increment',
+  'runtime/member_string_references',
+  'runtime/memset',
+  'runtime/operator',
+  'runtime/printf',
+  'runtime/printf_format',
+  'runtime/references',
+  'runtime/sizeof',
+  'runtime/string',
+  'runtime/threadsafe_fn',
+  'runtime/vlog',
+  'whitespace/blank_line',
+  'whitespace/braces',
+  'whitespace/comma',
+  'whitespace/comments',
+  'whitespace/empty_conditional_body',
+  'whitespace/empty_loop_body',
+  'whitespace/end_of_line',
+  'whitespace/ending_newline',
+  'whitespace/forcolon',
+  'whitespace/indent',
+  'whitespace/line_length',
+  'whitespace/newline',
+  'whitespace/operators',
+  'whitespace/parens',
+  'whitespace/semicolon',
+  'whitespace/tab',
+  'whitespace/todo'
+  ]
+
+# The default state of the category filter. This is overrided by the --filter=
+# flag. By default all errors are on, so only add here categories that should be
+# off by default (i.e., categories that must be enabled by the --filter= flags).
+# All entries here should start with a '-' or '+', as in the --filter= flag.
+_DEFAULT_FILTERS = ['-build/include_alpha']
+
+# We used to check for high-bit characters, but after much discussion we
+# decided those were OK, as long as they were in UTF-8 and didn't represent
+# hard-coded international strings, which belong in a separate i18n file.
+
+
+# C++ headers
+_CPP_HEADERS = frozenset([
+    # Legacy
+    'algobase.h',
+    'algo.h',
+    'alloc.h',
+    'builtinbuf.h',
+    'bvector.h',
+    'complex.h',
+    'defalloc.h',
+    'deque.h',
+    'editbuf.h',
+    'fstream.h',
+    'function.h',
+    'hash_map',
+    'hash_map.h',
+    'hash_set',
+    'hash_set.h',
+    'hashtable.h',
+    'heap.h',
+    'indstream.h',
+    'iomanip.h',
+    'iostream.h',
+    'istream.h',
+    'iterator.h',
+    'list.h',
+    'map.h',
+    'multimap.h',
+    'multiset.h',
+    'ostream.h',
+    'pair.h',
+    'parsestream.h',
+    'pfstream.h',
+    'procbuf.h',
+    'pthread_alloc',
+    'pthread_alloc.h',
+    'rope',
+    'rope.h',
+    'ropeimpl.h',
+    'set.h',
+    'slist',
+    'slist.h',
+    'stack.h',
+    'stdiostream.h',
+    'stl_alloc.h',
+    'stl_relops.h',
+    'streambuf.h',
+    'stream.h',
+    'strfile.h',
+    'strstream.h',
+    'tempbuf.h',
+    'tree.h',
+    'type_traits.h',
+    'vector.h',
+    # 17.6.1.2 C++ library headers
+    'algorithm',
+    'array',
+    'atomic',
+    'bitset',
+    'chrono',
+    'codecvt',
+    'complex',
+    'condition_variable',
+    'deque',
+    'exception',
+    'forward_list',
+    'fstream',
+    'functional',
+    'future',
+    'initializer_list',
+    'iomanip',
+    'ios',
+    'iosfwd',
+    'iostream',
+    'istream',
+    'iterator',
+    'limits',
+    'list',
+    'locale',
+    'map',
+    'memory',
+    'mutex',
+    'new',
+    'numeric',
+    'ostream',
+    'queue',
+    'random',
+    'ratio',
+    'regex',
+    'set',
+    'sstream',
+    'stack',
+    'stdexcept',
+    'streambuf',
+    'string',
+    'strstream',
+    'system_error',
+    'thread',
+    'tuple',
+    'typeindex',
+    'typeinfo',
+    'type_traits',
+    'unordered_map',
+    'unordered_set',
+    'utility',
+    'valarray',
+    'vector',
+    # 17.6.1.2 C++ headers for C library facilities
+    'cassert',
+    'ccomplex',
+    'cctype',
+    'cerrno',
+    'cfenv',
+    'cfloat',
+    'cinttypes',
+    'ciso646',
+    'climits',
+    'clocale',
+    'cmath',
+    'csetjmp',
+    'csignal',
+    'cstdalign',
+    'cstdarg',
+    'cstdbool',
+    'cstddef',
+    'cstdint',
+    'cstdio',
+    'cstdlib',
+    'cstring',
+    'ctgmath',
+    'ctime',
+    'cuchar',
+    'cwchar',
+    'cwctype',
+    ])
+
+# Assertion macros.  These are defined in base/logging.h and
+# testing/base/gunit.h.  Note that the _M versions need to come first
+# for substring matching to work.
+_CHECK_MACROS = [
+    'DCHECK', 'CHECK',
+    'EXPECT_TRUE_M', 'EXPECT_TRUE',
+    'ASSERT_TRUE_M', 'ASSERT_TRUE',
+    'EXPECT_FALSE_M', 'EXPECT_FALSE',
+    'ASSERT_FALSE_M', 'ASSERT_FALSE',
+    ]
+
+# Replacement macros for CHECK/DCHECK/EXPECT_TRUE/EXPECT_FALSE
+_CHECK_REPLACEMENT = dict([(m, {}) for m in _CHECK_MACROS])
+
+for op, replacement in [('==', 'EQ'), ('!=', 'NE'),
+                        ('>=', 'GE'), ('>', 'GT'),
+                        ('<=', 'LE'), ('<', 'LT')]:
+  _CHECK_REPLACEMENT['DCHECK'][op] = 'DCHECK_%s' % replacement
+  _CHECK_REPLACEMENT['CHECK'][op] = 'CHECK_%s' % replacement
+  _CHECK_REPLACEMENT['EXPECT_TRUE'][op] = 'EXPECT_%s' % replacement
+  _CHECK_REPLACEMENT['ASSERT_TRUE'][op] = 'ASSERT_%s' % replacement
+  _CHECK_REPLACEMENT['EXPECT_TRUE_M'][op] = 'EXPECT_%s_M' % replacement
+  _CHECK_REPLACEMENT['ASSERT_TRUE_M'][op] = 'ASSERT_%s_M' % replacement
+
+for op, inv_replacement in [('==', 'NE'), ('!=', 'EQ'),
+                            ('>=', 'LT'), ('>', 'LE'),
+                            ('<=', 'GT'), ('<', 'GE')]:
+  _CHECK_REPLACEMENT['EXPECT_FALSE'][op] = 'EXPECT_%s' % inv_replacement
+  _CHECK_REPLACEMENT['ASSERT_FALSE'][op] = 'ASSERT_%s' % inv_replacement
+  _CHECK_REPLACEMENT['EXPECT_FALSE_M'][op] = 'EXPECT_%s_M' % inv_replacement
+  _CHECK_REPLACEMENT['ASSERT_FALSE_M'][op] = 'ASSERT_%s_M' % inv_replacement
+
+# Alternative tokens and their replacements.  For full list, see section 2.5
+# Alternative tokens [lex.digraph] in the C++ standard.
+#
+# Digraphs (such as '%:') are not included here since it's a mess to
+# match those on a word boundary.
+_ALT_TOKEN_REPLACEMENT = {
+    'and': '&&',
+    'bitor': '|',
+    'or': '||',
+    'xor': '^',
+    'compl': '~',
+    'bitand': '&',
+    'and_eq': '&=',
+    'or_eq': '|=',
+    'xor_eq': '^=',
+    'not': '!',
+    'not_eq': '!='
+    }
+
+# Compile regular expression that matches all the above keywords.  The "[ =()]"
+# bit is meant to avoid matching these keywords outside of boolean expressions.
+#
+# False positives include C-style multi-line comments and multi-line strings
+# but those have always been troublesome for cpplint.
+_ALT_TOKEN_REPLACEMENT_PATTERN = re.compile(
+    r'[ =()](' + ('|'.join(_ALT_TOKEN_REPLACEMENT.keys())) + r')(?=[ (]|$)')
+
+
+# These constants define types of headers for use with
+# _IncludeState.CheckNextIncludeOrder().
+_C_SYS_HEADER = 1
+_CPP_SYS_HEADER = 2
+_LIKELY_MY_HEADER = 3
+_POSSIBLE_MY_HEADER = 4
+_OTHER_HEADER = 5
+
+# These constants define the current inline assembly state
+_NO_ASM = 0       # Outside of inline assembly block
+_INSIDE_ASM = 1   # Inside inline assembly block
+_END_ASM = 2      # Last line of inline assembly block
+_BLOCK_ASM = 3    # The whole block is an inline assembly block
+
+# Match start of assembly blocks
+_MATCH_ASM = re.compile(r'^\s*(?:asm|_asm|__asm|__asm__)'
+                        r'(?:\s+(volatile|__volatile__))?'
+                        r'\s*[{(]')
+
+
+_regexp_compile_cache = {}
+
+# Finds occurrences of NOLINT or NOLINT(...).
+_RE_SUPPRESSION = re.compile(r'\bNOLINT\b(\([^)]*\))?')
+
+# {str, set(int)}: a map from error categories to sets of linenumbers
+# on which those errors are expected and should be suppressed.
+_error_suppressions = {}
+
+# The root directory used for deriving header guard CPP variable.
+# This is set by --root flag.
+_root = None
+
+# The allowed line length of files.
+# This is set by --linelength flag.
+_line_length = 80
+
+# The allowed extensions for file names
+# This is set by --extensions flag.
+_valid_extensions = set(['cc', 'h', 'cpp', 'cu', 'cuh'])
+
+def ParseNolintSuppressions(filename, raw_line, linenum, error):
+  """Updates the global list of error-suppressions.
+
+  Parses any NOLINT comments on the current line, updating the global
+  error_suppressions store.  Reports an error if the NOLINT comment
+  was malformed.
+
+  Args:
+    filename: str, the name of the input file.
+    raw_line: str, the line of input text, with comments.
+    linenum: int, the number of the current line.
+    error: function, an error handler.
+  """
+  # FIXME(adonovan): "NOLINT(" is misparsed as NOLINT(*).
+  matched = _RE_SUPPRESSION.search(raw_line)
+  if matched:
+    category = matched.group(1)
+    if category in (None, '(*)'):  # => "suppress all"
+      _error_suppressions.setdefault(None, set()).add(linenum)
+    else:
+      if category.startswith('(') and category.endswith(')'):
+        category = category[1:-1]
+        if category in _ERROR_CATEGORIES:
+          _error_suppressions.setdefault(category, set()).add(linenum)
+        else:
+          error(filename, linenum, 'readability/nolint', 5,
+                'Unknown NOLINT error category: %s' % category)
+
+
+def ResetNolintSuppressions():
+  "Resets the set of NOLINT suppressions to empty."
+  _error_suppressions.clear()
+
+
+def IsErrorSuppressedByNolint(category, linenum):
+  """Returns true if the specified error category is suppressed on this line.
+
+  Consults the global error_suppressions map populated by
+  ParseNolintSuppressions/ResetNolintSuppressions.
+
+  Args:
+    category: str, the category of the error.
+    linenum: int, the current line number.
+  Returns:
+    bool, True iff the error should be suppressed due to a NOLINT comment.
+  """
+  return (linenum in _error_suppressions.get(category, set()) or
+          linenum in _error_suppressions.get(None, set()))
+
+def Match(pattern, s):
+  """Matches the string with the pattern, caching the compiled regexp."""
+  # The regexp compilation caching is inlined in both Match and Search for
+  # performance reasons; factoring it out into a separate function turns out
+  # to be noticeably expensive.
+  if pattern not in _regexp_compile_cache:
+    _regexp_compile_cache[pattern] = sre_compile.compile(pattern)
+  return _regexp_compile_cache[pattern].match(s)
+
+
+def ReplaceAll(pattern, rep, s):
+  """Replaces instances of pattern in a string with a replacement.
+
+  The compiled regex is kept in a cache shared by Match and Search.
+
+  Args:
+    pattern: regex pattern
+    rep: replacement text
+    s: search string
+
+  Returns:
+    string with replacements made (or original string if no replacements)
+  """
+  if pattern not in _regexp_compile_cache:
+    _regexp_compile_cache[pattern] = sre_compile.compile(pattern)
+  return _regexp_compile_cache[pattern].sub(rep, s)
+
+
+def Search(pattern, s):
+  """Searches the string for the pattern, caching the compiled regexp."""
+  if pattern not in _regexp_compile_cache:
+    _regexp_compile_cache[pattern] = sre_compile.compile(pattern)
+  return _regexp_compile_cache[pattern].search(s)
+
+
+class _IncludeState(dict):
+  """Tracks line numbers for includes, and the order in which includes appear.
+
+  As a dict, an _IncludeState object serves as a mapping between include
+  filename and line number on which that file was included.
+
+  Call CheckNextIncludeOrder() once for each header in the file, passing
+  in the type constants defined above. Calls in an illegal order will
+  raise an _IncludeError with an appropriate error message.
+
+  """
+  # self._section will move monotonically through this set. If it ever
+  # needs to move backwards, CheckNextIncludeOrder will raise an error.
+  _INITIAL_SECTION = 0
+  _MY_H_SECTION = 1
+  _C_SECTION = 2
+  _CPP_SECTION = 3
+  _OTHER_H_SECTION = 4
+
+  _TYPE_NAMES = {
+      _C_SYS_HEADER: 'C system header',
+      _CPP_SYS_HEADER: 'C++ system header',
+      _LIKELY_MY_HEADER: 'header this file implements',
+      _POSSIBLE_MY_HEADER: 'header this file may implement',
+      _OTHER_HEADER: 'other header',
+      }
+  _SECTION_NAMES = {
+      _INITIAL_SECTION: "... nothing. (This can't be an error.)",
+      _MY_H_SECTION: 'a header this file implements',
+      _C_SECTION: 'C system header',
+      _CPP_SECTION: 'C++ system header',
+      _OTHER_H_SECTION: 'other header',
+      }
+
+  def __init__(self):
+    dict.__init__(self)
+    self.ResetSection()
+
+  def ResetSection(self):
+    # The name of the current section.
+    self._section = self._INITIAL_SECTION
+    # The path of last found header.
+    self._last_header = ''
+
+  def SetLastHeader(self, header_path):
+    self._last_header = header_path
+
+  def CanonicalizeAlphabeticalOrder(self, header_path):
+    """Returns a path canonicalized for alphabetical comparison.
+
+    - replaces "-" with "_" so they both cmp the same.
+    - removes '-inl' since we don't require them to be after the main header.
+    - lowercase everything, just in case.
+
+    Args:
+      header_path: Path to be canonicalized.
+
+    Returns:
+      Canonicalized path.
+    """
+    return header_path.replace('-inl.h', '.h').replace('-', '_').lower()
+
+  def IsInAlphabeticalOrder(self, clean_lines, linenum, header_path):
+    """Check if a header is in alphabetical order with the previous header.
+
+    Args:
+      clean_lines: A CleansedLines instance containing the file.
+      linenum: The number of the line to check.
+      header_path: Canonicalized header to be checked.
+
+    Returns:
+      Returns true if the header is in alphabetical order.
+    """
+    # If previous section is different from current section, _last_header will
+    # be reset to empty string, so it's always less than current header.
+    #
+    # If previous line was a blank line, assume that the headers are
+    # intentionally sorted the way they are.
+    if (self._last_header > header_path and
+        not Match(r'^\s*$', clean_lines.elided[linenum - 1])):
+      return False
+    return True
+
+  def CheckNextIncludeOrder(self, header_type):
+    """Returns a non-empty error message if the next header is out of order.
+
+    This function also updates the internal state to be ready to check
+    the next include.
+
+    Args:
+      header_type: One of the _XXX_HEADER constants defined above.
+
+    Returns:
+      The empty string if the header is in the right order, or an
+      error message describing what's wrong.
+
+    """
+    error_message = ('Found %s after %s' %
+                     (self._TYPE_NAMES[header_type],
+                      self._SECTION_NAMES[self._section]))
+
+    last_section = self._section
+
+    if header_type == _C_SYS_HEADER:
+      if self._section <= self._C_SECTION:
+        self._section = self._C_SECTION
+      else:
+        self._last_header = ''
+        return error_message
+    elif header_type == _CPP_SYS_HEADER:
+      if self._section <= self._CPP_SECTION:
+        self._section = self._CPP_SECTION
+      else:
+        self._last_header = ''
+        return error_message
+    elif header_type == _LIKELY_MY_HEADER:
+      if self._section <= self._MY_H_SECTION:
+        self._section = self._MY_H_SECTION
+      else:
+        self._section = self._OTHER_H_SECTION
+    elif header_type == _POSSIBLE_MY_HEADER:
+      if self._section <= self._MY_H_SECTION:
+        self._section = self._MY_H_SECTION
+      else:
+        # This will always be the fallback because we're not sure
+        # enough that the header is associated with this file.
+        self._section = self._OTHER_H_SECTION
+    else:
+      assert header_type == _OTHER_HEADER
+      self._section = self._OTHER_H_SECTION
+
+    if last_section != self._section:
+      self._last_header = ''
+
+    return ''
+
+
+class _CppLintState(object):
+  """Maintains module-wide state.."""
+
+  def __init__(self):
+    self.verbose_level = 1  # global setting.
+    self.error_count = 0    # global count of reported errors
+    # filters to apply when emitting error messages
+    self.filters = _DEFAULT_FILTERS[:]
+    self.counting = 'total'  # In what way are we counting errors?
+    self.errors_by_category = {}  # string to int dict storing error counts
+
+    # output format:
+    # "emacs" - format that emacs can parse (default)
+    # "vs7" - format that Microsoft Visual Studio 7 can parse
+    self.output_format = 'emacs'
+
+  def SetOutputFormat(self, output_format):
+    """Sets the output format for errors."""
+    self.output_format = output_format
+
+  def SetVerboseLevel(self, level):
+    """Sets the module's verbosity, and returns the previous setting."""
+    last_verbose_level = self.verbose_level
+    self.verbose_level = level
+    return last_verbose_level
+
+  def SetCountingStyle(self, counting_style):
+    """Sets the module's counting options."""
+    self.counting = counting_style
+
+  def SetFilters(self, filters):
+    """Sets the error-message filters.
+
+    These filters are applied when deciding whether to emit a given
+    error message.
+
+    Args:
+      filters: A string of comma-separated filters (eg "+whitespace/indent").
+               Each filter should start with + or -; else we die.
+
+    Raises:
+      ValueError: The comma-separated filters did not all start with '+' or '-'.
+                  E.g. "-,+whitespace,-whitespace/indent,whitespace/badfilter"
+    """
+    # Default filters always have less priority than the flag ones.
+    self.filters = _DEFAULT_FILTERS[:]
+    for filt in filters.split(','):
+      clean_filt = filt.strip()
+      if clean_filt:
+        self.filters.append(clean_filt)
+    for filt in self.filters:
+      if not (filt.startswith('+') or filt.startswith('-')):
+        raise ValueError('Every filter in --filters must start with + or -'
+                         ' (%s does not)' % filt)
+
+  def ResetErrorCounts(self):
+    """Sets the module's error statistic back to zero."""
+    self.error_count = 0
+    self.errors_by_category = {}
+
+  def IncrementErrorCount(self, category):
+    """Bumps the module's error statistic."""
+    self.error_count += 1
+    if self.counting in ('toplevel', 'detailed'):
+      if self.counting != 'detailed':
+        category = category.split('/')[0]
+      if category not in self.errors_by_category:
+        self.errors_by_category[category] = 0
+      self.errors_by_category[category] += 1
+
+  def PrintErrorCounts(self):
+    """Print a summary of errors by category, and the total."""
+    for category, count in self.errors_by_category.iteritems():
+      sys.stderr.write('Category \'%s\' errors found: %d\n' %
+                       (category, count))
+    sys.stderr.write('Total errors found: %d\n' % self.error_count)
+
+_cpplint_state = _CppLintState()
+
+
+def _OutputFormat():
+  """Gets the module's output format."""
+  return _cpplint_state.output_format
+
+
+def _SetOutputFormat(output_format):
+  """Sets the module's output format."""
+  _cpplint_state.SetOutputFormat(output_format)
+
+
+def _VerboseLevel():
+  """Returns the module's verbosity setting."""
+  return _cpplint_state.verbose_level
+
+
+def _SetVerboseLevel(level):
+  """Sets the module's verbosity, and returns the previous setting."""
+  return _cpplint_state.SetVerboseLevel(level)
+
+
+def _SetCountingStyle(level):
+  """Sets the module's counting options."""
+  _cpplint_state.SetCountingStyle(level)
+
+
+def _Filters():
+  """Returns the module's list of output filters, as a list."""
+  return _cpplint_state.filters
+
+
+def _SetFilters(filters):
+  """Sets the module's error-message filters.
+
+  These filters are applied when deciding whether to emit a given
+  error message.
+
+  Args:
+    filters: A string of comma-separated filters (eg "whitespace/indent").
+             Each filter should start with + or -; else we die.
+  """
+  _cpplint_state.SetFilters(filters)
+
+
+class _FunctionState(object):
+  """Tracks current function name and the number of lines in its body."""
+
+  _NORMAL_TRIGGER = 250  # for --v=0, 500 for --v=1, etc.
+  _TEST_TRIGGER = 400    # about 50% more than _NORMAL_TRIGGER.
+
+  def __init__(self):
+    self.in_a_function = False
+    self.lines_in_function = 0
+    self.current_function = ''
+
+  def Begin(self, function_name):
+    """Start analyzing function body.
+
+    Args:
+      function_name: The name of the function being tracked.
+    """
+    self.in_a_function = True
+    self.lines_in_function = 0
+    self.current_function = function_name
+
+  def Count(self):
+    """Count line in current function body."""
+    if self.in_a_function:
+      self.lines_in_function += 1
+
+  def Check(self, error, filename, linenum):
+    """Report if too many lines in function body.
+
+    Args:
+      error: The function to call with any errors found.
+      filename: The name of the current file.
+      linenum: The number of the line to check.
+    """
+    if Match(r'T(EST|est)', self.current_function):
+      base_trigger = self._TEST_TRIGGER
+    else:
+      base_trigger = self._NORMAL_TRIGGER
+    trigger = base_trigger * 2**_VerboseLevel()
+
+    if self.lines_in_function > trigger:
+      error_level = int(math.log(self.lines_in_function / base_trigger, 2))
+      # 50 => 0, 100 => 1, 200 => 2, 400 => 3, 800 => 4, 1600 => 5, ...
+      if error_level > 5:
+        error_level = 5
+      error(filename, linenum, 'readability/fn_size', error_level,
+            'Small and focused functions are preferred:'
+            ' %s has %d non-comment lines'
+            ' (error triggered by exceeding %d lines).'  % (
+                self.current_function, self.lines_in_function, trigger))
+
+  def End(self):
+    """Stop analyzing function body."""
+    self.in_a_function = False
+
+
+class _IncludeError(Exception):
+  """Indicates a problem with the include order in a file."""
+  pass
+
+
+class FileInfo:
+  """Provides utility functions for filenames.
+
+  FileInfo provides easy access to the components of a file's path
+  relative to the project root.
+  """
+
+  def __init__(self, filename):
+    self._filename = filename
+
+  def FullName(self):
+    """Make Windows paths like Unix."""
+    return os.path.abspath(self._filename).replace('\\', '/')
+
+  def RepositoryName(self):
+    """FullName after removing the local path to the repository.
+
+    If we have a real absolute path name here we can try to do something smart:
+    detecting the root of the checkout and truncating /path/to/checkout from
+    the name so that we get header guards that don't include things like
+    "C:\Documents and Settings\..." or "/home/username/..." in them and thus
+    people on different computers who have checked the source out to different
+    locations won't see bogus errors.
+    """
+    fullname = self.FullName()
+
+    if os.path.exists(fullname):
+      project_dir = os.path.dirname(fullname)
+
+      if os.path.exists(os.path.join(project_dir, ".svn")):
+        # If there's a .svn file in the current directory, we recursively look
+        # up the directory tree for the top of the SVN checkout
+        root_dir = project_dir
+        one_up_dir = os.path.dirname(root_dir)
+        while os.path.exists(os.path.join(one_up_dir, ".svn")):
+          root_dir = os.path.dirname(root_dir)
+          one_up_dir = os.path.dirname(one_up_dir)
+
+        prefix = os.path.commonprefix([root_dir, project_dir])
+        return fullname[len(prefix) + 1:]
+
+      # Not SVN <= 1.6? Try to find a git, hg, or svn top level directory by
+      # searching up from the current path.
+      root_dir = os.path.dirname(fullname)
+      while (root_dir != os.path.dirname(root_dir) and
+             not os.path.exists(os.path.join(root_dir, ".git")) and
+             not os.path.exists(os.path.join(root_dir, ".hg")) and
+             not os.path.exists(os.path.join(root_dir, ".svn"))):
+        root_dir = os.path.dirname(root_dir)
+
+      if (os.path.exists(os.path.join(root_dir, ".git")) or
+          os.path.exists(os.path.join(root_dir, ".hg")) or
+          os.path.exists(os.path.join(root_dir, ".svn"))):
+        prefix = os.path.commonprefix([root_dir, project_dir])
+        return fullname[len(prefix) + 1:]
+
+    # Don't know what to do; header guard warnings may be wrong...
+    return fullname
+
+  def Split(self):
+    """Splits the file into the directory, basename, and extension.
+
+    For 'chrome/browser/browser.cc', Split() would
+    return ('chrome/browser', 'browser', '.cc')
+
+    Returns:
+      A tuple of (directory, basename, extension).
+    """
+
+    googlename = self.RepositoryName()
+    project, rest = os.path.split(googlename)
+    return (project,) + os.path.splitext(rest)
+
+  def BaseName(self):
+    """File base name - text after the final slash, before the final period."""
+    return self.Split()[1]
+
+  def Extension(self):
+    """File extension - text following the final period."""
+    return self.Split()[2]
+
+  def NoExtension(self):
+    """File has no source file extension."""
+    return '/'.join(self.Split()[0:2])
+
+  def IsSource(self):
+    """File has a source file extension."""
+    return self.Extension()[1:] in ('c', 'cc', 'cpp', 'cxx')
+
+
+def _ShouldPrintError(category, confidence, linenum):
+  """If confidence >= verbose, category passes filter and is not suppressed."""
+
+  # There are three ways we might decide not to print an error message:
+  # a "NOLINT(category)" comment appears in the source,
+  # the verbosity level isn't high enough, or the filters filter it out.
+  if IsErrorSuppressedByNolint(category, linenum):
+    return False
+  if confidence < _cpplint_state.verbose_level:
+    return False
+
+  is_filtered = False
+  for one_filter in _Filters():
+    if one_filter.startswith('-'):
+      if category.startswith(one_filter[1:]):
+        is_filtered = True
+    elif one_filter.startswith('+'):
+      if category.startswith(one_filter[1:]):
+        is_filtered = False
+    else:
+      assert False  # should have been checked for in SetFilter.
+  if is_filtered:
+    return False
+
+  return True
+
+
+def Error(filename, linenum, category, confidence, message):
+  """Logs the fact we've found a lint error.
+
+  We log where the error was found, and also our confidence in the error,
+  that is, how certain we are this is a legitimate style regression, and
+  not a misidentification or a use that's sometimes justified.
+
+  False positives can be suppressed by the use of
+  "cpplint(category)"  comments on the offending line.  These are
+  parsed into _error_suppressions.
+
+  Args:
+    filename: The name of the file containing the error.
+    linenum: The number of the line containing the error.
+    category: A string used to describe the "category" this bug
+      falls under: "whitespace", say, or "runtime".  Categories
+      may have a hierarchy separated by slashes: "whitespace/indent".
+    confidence: A number from 1-5 representing a confidence score for
+      the error, with 5 meaning that we are certain of the problem,
+      and 1 meaning that it could be a legitimate construct.
+    message: The error message.
+  """
+  if _ShouldPrintError(category, confidence, linenum):
+    _cpplint_state.IncrementErrorCount(category)
+    if _cpplint_state.output_format == 'vs7':
+      sys.stderr.write('%s(%s):  %s  [%s] [%d]\n' % (
+          filename, linenum, message, category, confidence))
+    elif _cpplint_state.output_format == 'eclipse':
+      sys.stderr.write('%s:%s: warning: %s  [%s] [%d]\n' % (
+          filename, linenum, message, category, confidence))
+    else:
+      sys.stderr.write('%s:%s:  %s  [%s] [%d]\n' % (
+          filename, linenum, message, category, confidence))
+
+
+# Matches standard C++ escape sequences per 2.13.2.3 of the C++ standard.
+_RE_PATTERN_CLEANSE_LINE_ESCAPES = re.compile(
+    r'\\([abfnrtv?"\\\']|\d+|x[0-9a-fA-F]+)')
+# Matches strings.  Escape codes should already be removed by ESCAPES.
+_RE_PATTERN_CLEANSE_LINE_DOUBLE_QUOTES = re.compile(r'"[^"]*"')
+# Matches characters.  Escape codes should already be removed by ESCAPES.
+_RE_PATTERN_CLEANSE_LINE_SINGLE_QUOTES = re.compile(r"'.'")
+# Matches multi-line C++ comments.
+# This RE is a little bit more complicated than one might expect, because we
+# have to take care of space removals tools so we can handle comments inside
+# statements better.
+# The current rule is: We only clear spaces from both sides when we're at the
+# end of the line. Otherwise, we try to remove spaces from the right side,
+# if this doesn't work we try on left side but only if there's a non-character
+# on the right.
+_RE_PATTERN_CLEANSE_LINE_C_COMMENTS = re.compile(
+    r"""(\s*/\*.*\*/\s*$|
+            /\*.*\*/\s+|
+         \s+/\*.*\*/(?=\W)|
+            /\*.*\*/)""", re.VERBOSE)
+
+
+def IsCppString(line):
+  """Does line terminate so, that the next symbol is in string constant.
+
+  This function does not consider single-line nor multi-line comments.
+
+  Args:
+    line: is a partial line of code starting from the 0..n.
+
+  Returns:
+    True, if next character appended to 'line' is inside a
+    string constant.
+  """
+
+  line = line.replace(r'\\', 'XX')  # after this, \\" does not match to \"
+  return ((line.count('"') - line.count(r'\"') - line.count("'\"'")) & 1) == 1
+
+
+def CleanseRawStrings(raw_lines):
+  """Removes C++11 raw strings from lines.
+
+    Before:
+      static const char kData[] = R"(
+          multi-line string
+          )";
+
+    After:
+      static const char kData[] = ""
+          (replaced by blank line)
+          "";
+
+  Args:
+    raw_lines: list of raw lines.
+
+  Returns:
+    list of lines with C++11 raw strings replaced by empty strings.
+  """
+
+  delimiter = None
+  lines_without_raw_strings = []
+  for line in raw_lines:
+    if delimiter:
+      # Inside a raw string, look for the end
+      end = line.find(delimiter)
+      if end >= 0:
+        # Found the end of the string, match leading space for this
+        # line and resume copying the original lines, and also insert
+        # a "" on the last line.
+        leading_space = Match(r'^(\s*)\S', line)
+        line = leading_space.group(1) + '""' + line[end + len(delimiter):]
+        delimiter = None
+      else:
+        # Haven't found the end yet, append a blank line.
+        line = ''
+
+    else:
+      # Look for beginning of a raw string.
+      # See 2.14.15 [lex.string] for syntax.
+      matched = Match(r'^(.*)\b(?:R|u8R|uR|UR|LR)"([^\s\\()]*)\((.*)$', line)
+      if matched:
+        delimiter = ')' + matched.group(2) + '"'
+
+        end = matched.group(3).find(delimiter)
+        if end >= 0:
+          # Raw string ended on same line
+          line = (matched.group(1) + '""' +
+                  matched.group(3)[end + len(delimiter):])
+          delimiter = None
+        else:
+          # Start of a multi-line raw string
+          line = matched.group(1) + '""'
+
+    lines_without_raw_strings.append(line)
+
+  # TODO(unknown): if delimiter is not None here, we might want to
+  # emit a warning for unterminated string.
+  return lines_without_raw_strings
+
+
+def FindNextMultiLineCommentStart(lines, lineix):
+  """Find the beginning marker for a multiline comment."""
+  while lineix < len(lines):
+    if lines[lineix].strip().startswith('/*'):
+      # Only return this marker if the comment goes beyond this line
+      if lines[lineix].strip().find('*/', 2) < 0:
+        return lineix
+    lineix += 1
+  return len(lines)
+
+
+def FindNextMultiLineCommentEnd(lines, lineix):
+  """We are inside a comment, find the end marker."""
+  while lineix < len(lines):
+    if lines[lineix].strip().endswith('*/'):
+      return lineix
+    lineix += 1
+  return len(lines)
+
+
+def RemoveMultiLineCommentsFromRange(lines, begin, end):
+  """Clears a range of lines for multi-line comments."""
+  # Having // dummy comments makes the lines non-empty, so we will not get
+  # unnecessary blank line warnings later in the code.
+  for i in range(begin, end):
+    lines[i] = '// dummy'
+
+
+def RemoveMultiLineComments(filename, lines, error):
+  """Removes multiline (c-style) comments from lines."""
+  lineix = 0
+  while lineix < len(lines):
+    lineix_begin = FindNextMultiLineCommentStart(lines, lineix)
+    if lineix_begin >= len(lines):
+      return
+    lineix_end = FindNextMultiLineCommentEnd(lines, lineix_begin)
+    if lineix_end >= len(lines):
+      error(filename, lineix_begin + 1, 'readability/multiline_comment', 5,
+            'Could not find end of multi-line comment')
+      return
+    RemoveMultiLineCommentsFromRange(lines, lineix_begin, lineix_end + 1)
+    lineix = lineix_end + 1
+
+
+def CleanseComments(line):
+  """Removes //-comments and single-line C-style /* */ comments.
+
+  Args:
+    line: A line of C++ source.
+
+  Returns:
+    The line with single-line comments removed.
+  """
+  commentpos = line.find('//')
+  if commentpos != -1 and not IsCppString(line[:commentpos]):
+    line = line[:commentpos].rstrip()
+  # get rid of /* ... */
+  return _RE_PATTERN_CLEANSE_LINE_C_COMMENTS.sub('', line)
+
+
+class CleansedLines(object):
+  """Holds 3 copies of all lines with different preprocessing applied to them.
+
+  1) elided member contains lines without strings and comments,
+  2) lines member contains lines without comments, and
+  3) raw_lines member contains all the lines without processing.
+  All these three members are of <type 'list'>, and of the same length.
+  """
+
+  def __init__(self, lines):
+    self.elided = []
+    self.lines = []
+    self.raw_lines = lines
+    self.num_lines = len(lines)
+    self.lines_without_raw_strings = CleanseRawStrings(lines)
+    for linenum in range(len(self.lines_without_raw_strings)):
+      self.lines.append(CleanseComments(
+          self.lines_without_raw_strings[linenum]))
+      elided = self._CollapseStrings(self.lines_without_raw_strings[linenum])
+      self.elided.append(CleanseComments(elided))
+
+  def NumLines(self):
+    """Returns the number of lines represented."""
+    return self.num_lines
+
+  @staticmethod
+  def _CollapseStrings(elided):
+    """Collapses strings and chars on a line to simple "" or '' blocks.
+
+    We nix strings first so we're not fooled by text like '"http://"'
+
+    Args:
+      elided: The line being processed.
+
+    Returns:
+      The line with collapsed strings.
+    """
+    if not _RE_PATTERN_INCLUDE.match(elided):
+      # Remove escaped characters first to make quote/single quote collapsing
+      # basic.  Things that look like escaped characters shouldn't occur
+      # outside of strings and chars.
+      elided = _RE_PATTERN_CLEANSE_LINE_ESCAPES.sub('', elided)
+      elided = _RE_PATTERN_CLEANSE_LINE_SINGLE_QUOTES.sub("''", elided)
+      elided = _RE_PATTERN_CLEANSE_LINE_DOUBLE_QUOTES.sub('""', elided)
+    return elided
+
+
+def FindEndOfExpressionInLine(line, startpos, depth, startchar, endchar):
+  """Find the position just after the matching endchar.
+
+  Args:
+    line: a CleansedLines line.
+    startpos: start searching at this position.
+    depth: nesting level at startpos.
+    startchar: expression opening character.
+    endchar: expression closing character.
+
+  Returns:
+    On finding matching endchar: (index just after matching endchar, 0)
+    Otherwise: (-1, new depth at end of this line)
+  """
+  for i in xrange(startpos, len(line)):
+    if line[i] == startchar:
+      depth += 1
+    elif line[i] == endchar:
+      depth -= 1
+      if depth == 0:
+        return (i + 1, 0)
+  return (-1, depth)
+
+
+def CloseExpression(clean_lines, linenum, pos):
+  """If input points to ( or { or [ or <, finds the position that closes it.
+
+  If lines[linenum][pos] points to a '(' or '{' or '[' or '<', finds the
+  linenum/pos that correspond to the closing of the expression.
+
+  Args:
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    pos: A position on the line.
+
+  Returns:
+    A tuple (line, linenum, pos) pointer *past* the closing brace, or
+    (line, len(lines), -1) if we never find a close.  Note we ignore
+    strings and comments when matching; and the line we return is the
+    'cleansed' line at linenum.
+  """
+
+  line = clean_lines.elided[linenum]
+  startchar = line[pos]
+  if startchar not in '({[<':
+    return (line, clean_lines.NumLines(), -1)
+  if startchar == '(': endchar = ')'
+  if startchar == '[': endchar = ']'
+  if startchar == '{': endchar = '}'
+  if startchar == '<': endchar = '>'
+
+  # Check first line
+  (end_pos, num_open) = FindEndOfExpressionInLine(
+      line, pos, 0, startchar, endchar)
+  if end_pos > -1:
+    return (line, linenum, end_pos)
+
+  # Continue scanning forward
+  while linenum < clean_lines.NumLines() - 1:
+    linenum += 1
+    line = clean_lines.elided[linenum]
+    (end_pos, num_open) = FindEndOfExpressionInLine(
+        line, 0, num_open, startchar, endchar)
+    if end_pos > -1:
+      return (line, linenum, end_pos)
+
+  # Did not find endchar before end of file, give up
+  return (line, clean_lines.NumLines(), -1)
+
+
+def FindStartOfExpressionInLine(line, endpos, depth, startchar, endchar):
+  """Find position at the matching startchar.
+
+  This is almost the reverse of FindEndOfExpressionInLine, but note
+  that the input position and returned position differs by 1.
+
+  Args:
+    line: a CleansedLines line.
+    endpos: start searching at this position.
+    depth: nesting level at endpos.
+    startchar: expression opening character.
+    endchar: expression closing character.
+
+  Returns:
+    On finding matching startchar: (index at matching startchar, 0)
+    Otherwise: (-1, new depth at beginning of this line)
+  """
+  for i in xrange(endpos, -1, -1):
+    if line[i] == endchar:
+      depth += 1
+    elif line[i] == startchar:
+      depth -= 1
+      if depth == 0:
+        return (i, 0)
+  return (-1, depth)
+
+
+def ReverseCloseExpression(clean_lines, linenum, pos):
+  """If input points to ) or } or ] or >, finds the position that opens it.
+
+  If lines[linenum][pos] points to a ')' or '}' or ']' or '>', finds the
+  linenum/pos that correspond to the opening of the expression.
+
+  Args:
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    pos: A position on the line.
+
+  Returns:
+    A tuple (line, linenum, pos) pointer *at* the opening brace, or
+    (line, 0, -1) if we never find the matching opening brace.  Note
+    we ignore strings and comments when matching; and the line we
+    return is the 'cleansed' line at linenum.
+  """
+  line = clean_lines.elided[linenum]
+  endchar = line[pos]
+  if endchar not in ')}]>':
+    return (line, 0, -1)
+  if endchar == ')': startchar = '('
+  if endchar == ']': startchar = '['
+  if endchar == '}': startchar = '{'
+  if endchar == '>': startchar = '<'
+
+  # Check last line
+  (start_pos, num_open) = FindStartOfExpressionInLine(
+      line, pos, 0, startchar, endchar)
+  if start_pos > -1:
+    return (line, linenum, start_pos)
+
+  # Continue scanning backward
+  while linenum > 0:
+    linenum -= 1
+    line = clean_lines.elided[linenum]
+    (start_pos, num_open) = FindStartOfExpressionInLine(
+        line, len(line) - 1, num_open, startchar, endchar)
+    if start_pos > -1:
+      return (line, linenum, start_pos)
+
+  # Did not find startchar before beginning of file, give up
+  return (line, 0, -1)
+
+
+def CheckForCopyright(filename, lines, error):
+  """Logs an error if no Copyright message appears at the top of the file."""
+
+  # We'll say it should occur by line 10. Don't forget there's a
+  # dummy line at the front.
+  for line in xrange(1, min(len(lines), 11)):
+    if re.search(r'Copyright', lines[line], re.I): break
+  else:                       # means no copyright line was found
+    error(filename, 0, 'legal/copyright', 5,
+          'No copyright message found.  '
+          'You should have a line: "Copyright [year] <Copyright Owner>"')
+
+
+def GetHeaderGuardCPPVariable(filename):
+  """Returns the CPP variable that should be used as a header guard.
+
+  Args:
+    filename: The name of a C++ header file.
+
+  Returns:
+    The CPP variable that should be used as a header guard in the
+    named file.
+
+  """
+
+  # Restores original filename in case that cpplint is invoked from Emacs's
+  # flymake.
+  filename = re.sub(r'_flymake\.h$', '.h', filename)
+  filename = re.sub(r'/\.flymake/([^/]*)$', r'/\1', filename)
+
+  fileinfo = FileInfo(filename)
+  file_path_from_root = fileinfo.RepositoryName()
+  if _root:
+    file_path_from_root = re.sub('^' + _root + os.sep, '', file_path_from_root)
+  return re.sub(r'[-./\s]', '_', file_path_from_root).upper() + '_'
+
+
+def CheckForHeaderGuard(filename, lines, error):
+  """Checks that the file contains a header guard.
+
+  Logs an error if no #ifndef header guard is present.  For other
+  headers, checks that the full pathname is used.
+
+  Args:
+    filename: The name of the C++ header file.
+    lines: An array of strings, each representing a line of the file.
+    error: The function to call with any errors found.
+  """
+
+  cppvar = GetHeaderGuardCPPVariable(filename)
+
+  ifndef = None
+  ifndef_linenum = 0
+  define = None
+  endif = None
+  endif_linenum = 0
+  for linenum, line in enumerate(lines):
+    linesplit = line.split()
+    if len(linesplit) >= 2:
+      # find the first occurrence of #ifndef and #define, save arg
+      if not ifndef and linesplit[0] == '#ifndef':
+        # set ifndef to the header guard presented on the #ifndef line.
+        ifndef = linesplit[1]
+        ifndef_linenum = linenum
+      if not define and linesplit[0] == '#define':
+        define = linesplit[1]
+    # find the last occurrence of #endif, save entire line
+    if line.startswith('#endif'):
+      endif = line
+      endif_linenum = linenum
+
+  if not ifndef:
+    error(filename, 0, 'build/header_guard', 5,
+          'No #ifndef header guard found, suggested CPP variable is: %s' %
+          cppvar)
+    return
+
+  if not define:
+    error(filename, 0, 'build/header_guard', 5,
+          'No #define header guard found, suggested CPP variable is: %s' %
+          cppvar)
+    return
+
+  # The guard should be PATH_FILE_H_, but we also allow PATH_FILE_H__
+  # for backward compatibility.
+  if ifndef != cppvar:
+    error_level = 0
+    if ifndef != cppvar + '_':
+      error_level = 5
+
+    ParseNolintSuppressions(filename, lines[ifndef_linenum], ifndef_linenum,
+                            error)
+    error(filename, ifndef_linenum, 'build/header_guard', error_level,
+          '#ifndef header guard has wrong style, please use: %s' % cppvar)
+
+  if define != ifndef:
+    error(filename, 0, 'build/header_guard', 5,
+          '#ifndef and #define don\'t match, suggested CPP variable is: %s' %
+          cppvar)
+    return
+
+  if endif != ('#endif  // %s' % cppvar):
+    error_level = 0
+    if endif != ('#endif  // %s' % (cppvar + '_')):
+      error_level = 5
+
+    ParseNolintSuppressions(filename, lines[endif_linenum], endif_linenum,
+                            error)
+    error(filename, endif_linenum, 'build/header_guard', error_level,
+          '#endif line should be "#endif  // %s"' % cppvar)
+
+
+def CheckForBadCharacters(filename, lines, error):
+  """Logs an error for each line containing bad characters.
+
+  Two kinds of bad characters:
+
+  1. Unicode replacement characters: These indicate that either the file
+  contained invalid UTF-8 (likely) or Unicode replacement characters (which
+  it shouldn't).  Note that it's possible for this to throw off line
+  numbering if the invalid UTF-8 occurred adjacent to a newline.
+
+  2. NUL bytes.  These are problematic for some tools.
+
+  Args:
+    filename: The name of the current file.
+    lines: An array of strings, each representing a line of the file.
+    error: The function to call with any errors found.
+  """
+  for linenum, line in enumerate(lines):
+    if u'\ufffd' in line:
+      error(filename, linenum, 'readability/utf8', 5,
+            'Line contains invalid UTF-8 (or Unicode replacement character).')
+    if '\0' in line:
+      error(filename, linenum, 'readability/nul', 5, 'Line contains NUL byte.')
+
+
+def CheckForNewlineAtEOF(filename, lines, error):
+  """Logs an error if there is no newline char at the end of the file.
+
+  Args:
+    filename: The name of the current file.
+    lines: An array of strings, each representing a line of the file.
+    error: The function to call with any errors found.
+  """
+
+  # The array lines() was created by adding two newlines to the
+  # original file (go figure), then splitting on \n.
+  # To verify that the file ends in \n, we just have to make sure the
+  # last-but-two element of lines() exists and is empty.
+  if len(lines) < 3 or lines[-2]:
+    error(filename, len(lines) - 2, 'whitespace/ending_newline', 5,
+          'Could not find a newline character at the end of the file.')
+
+
+def CheckForMultilineCommentsAndStrings(filename, clean_lines, linenum, error):
+  """Logs an error if we see /* ... */ or "..." that extend past one line.
+
+  /* ... */ comments are legit inside macros, for one line.
+  Otherwise, we prefer // comments, so it's ok to warn about the
+  other.  Likewise, it's ok for strings to extend across multiple
+  lines, as long as a line continuation character (backslash)
+  terminates each line. Although not currently prohibited by the C++
+  style guide, it's ugly and unnecessary. We don't do well with either
+  in this lint program, so we warn about both.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+
+  # Remove all \\ (escaped backslashes) from the line. They are OK, and the
+  # second (escaped) slash may trigger later \" detection erroneously.
+  line = line.replace('\\\\', '')
+
+  if line.count('/*') > line.count('*/'):
+    error(filename, linenum, 'readability/multiline_comment', 5,
+          'Complex multi-line /*...*/-style comment found. '
+          'Lint may give bogus warnings.  '
+          'Consider replacing these with //-style comments, '
+          'with #if 0...#endif, '
+          'or with more clearly structured multi-line comments.')
+
+  if (line.count('"') - line.count('\\"')) % 2:
+    error(filename, linenum, 'readability/multiline_string', 5,
+          'Multi-line string ("...") found.  This lint script doesn\'t '
+          'do well with such strings, and may give bogus warnings.  '
+          'Use C++11 raw strings or concatenation instead.')
+
+
+threading_list = (
+    ('asctime(', 'asctime_r('),
+    ('ctime(', 'ctime_r('),
+    ('getgrgid(', 'getgrgid_r('),
+    ('getgrnam(', 'getgrnam_r('),
+    ('getlogin(', 'getlogin_r('),
+    ('getpwnam(', 'getpwnam_r('),
+    ('getpwuid(', 'getpwuid_r('),
+    ('gmtime(', 'gmtime_r('),
+    ('localtime(', 'localtime_r('),
+    ('rand(', 'rand_r('),
+    ('strtok(', 'strtok_r('),
+    ('ttyname(', 'ttyname_r('),
+    )
+
+
+def CheckPosixThreading(filename, clean_lines, linenum, error):
+  """Checks for calls to thread-unsafe functions.
+
+  Much code has been originally written without consideration of
+  multi-threading. Also, engineers are relying on their old experience;
+  they have learned posix before threading extensions were added. These
+  tests guide the engineers to use thread-safe functions (when using
+  posix directly).
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+  for single_thread_function, multithread_safe_function in threading_list:
+    ix = line.find(single_thread_function)
+    # Comparisons made explicit for clarity -- pylint: disable=g-explicit-bool-comparison
+    if ix >= 0 and (ix == 0 or (not line[ix - 1].isalnum() and
+                                line[ix - 1] not in ('_', '.', '>'))):
+      error(filename, linenum, 'runtime/threadsafe_fn', 2,
+            'Consider using ' + multithread_safe_function +
+            '...) instead of ' + single_thread_function +
+            '...) for improved thread safety.')
+
+
+def CheckVlogArguments(filename, clean_lines, linenum, error):
+  """Checks that VLOG() is only used for defining a logging level.
+
+  For example, VLOG(2) is correct. VLOG(INFO), VLOG(WARNING), VLOG(ERROR), and
+  VLOG(FATAL) are not.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+  if Search(r'\bVLOG\((INFO|ERROR|WARNING|DFATAL|FATAL)\)', line):
+    error(filename, linenum, 'runtime/vlog', 5,
+          'VLOG() should be used with numeric verbosity level.  '
+          'Use LOG() if you want symbolic severity levels.')
+
+
+# Matches invalid increment: *count++, which moves pointer instead of
+# incrementing a value.
+_RE_PATTERN_INVALID_INCREMENT = re.compile(
+    r'^\s*\*\w+(\+\+|--);')
+
+
+def CheckInvalidIncrement(filename, clean_lines, linenum, error):
+  """Checks for invalid increment *count++.
+
+  For example following function:
+  void increment_counter(int* count) {
+    *count++;
+  }
+  is invalid, because it effectively does count++, moving pointer, and should
+  be replaced with ++*count, (*count)++ or *count += 1.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+  if _RE_PATTERN_INVALID_INCREMENT.match(line):
+    error(filename, linenum, 'runtime/invalid_increment', 5,
+          'Changing pointer instead of value (or unused value of operator*).')
+
+
+class _BlockInfo(object):
+  """Stores information about a generic block of code."""
+
+  def __init__(self, seen_open_brace):
+    self.seen_open_brace = seen_open_brace
+    self.open_parentheses = 0
+    self.inline_asm = _NO_ASM
+
+  def CheckBegin(self, filename, clean_lines, linenum, error):
+    """Run checks that applies to text up to the opening brace.
+
+    This is mostly for checking the text after the class identifier
+    and the "{", usually where the base class is specified.  For other
+    blocks, there isn't much to check, so we always pass.
+
+    Args:
+      filename: The name of the current file.
+      clean_lines: A CleansedLines instance containing the file.
+      linenum: The number of the line to check.
+      error: The function to call with any errors found.
+    """
+    pass
+
+  def CheckEnd(self, filename, clean_lines, linenum, error):
+    """Run checks that applies to text after the closing brace.
+
+    This is mostly used for checking end of namespace comments.
+
+    Args:
+      filename: The name of the current file.
+      clean_lines: A CleansedLines instance containing the file.
+      linenum: The number of the line to check.
+      error: The function to call with any errors found.
+    """
+    pass
+
+
+class _ClassInfo(_BlockInfo):
+  """Stores information about a class."""
+
+  def __init__(self, name, class_or_struct, clean_lines, linenum):
+    _BlockInfo.__init__(self, False)
+    self.name = name
+    self.starting_linenum = linenum
+    self.is_derived = False
+    if class_or_struct == 'struct':
+      self.access = 'public'
+      self.is_struct = True
+    else:
+      self.access = 'private'
+      self.is_struct = False
+
+    # Remember initial indentation level for this class.  Using raw_lines here
+    # instead of elided to account for leading comments.
+    initial_indent = Match(r'^( *)\S', clean_lines.raw_lines[linenum])
+    if initial_indent:
+      self.class_indent = len(initial_indent.group(1))
+    else:
+      self.class_indent = 0
+
+    # Try to find the end of the class.  This will be confused by things like:
+    #   class A {
+    #   } *x = { ...
+    #
+    # But it's still good enough for CheckSectionSpacing.
+    self.last_line = 0
+    depth = 0
+    for i in range(linenum, clean_lines.NumLines()):
+      line = clean_lines.elided[i]
+      depth += line.count('{') - line.count('}')
+      if not depth:
+        self.last_line = i
+        break
+
+  def CheckBegin(self, filename, clean_lines, linenum, error):
+    # Look for a bare ':'
+    if Search('(^|[^:]):($|[^:])', clean_lines.elided[linenum]):
+      self.is_derived = True
+
+  def CheckEnd(self, filename, clean_lines, linenum, error):
+    # Check that closing brace is aligned with beginning of the class.
+    # Only do this if the closing brace is indented by only whitespaces.
+    # This means we will not check single-line class definitions.
+    indent = Match(r'^( *)\}', clean_lines.elided[linenum])
+    if indent and len(indent.group(1)) != self.class_indent:
+      if self.is_struct:
+        parent = 'struct ' + self.name
+      else:
+        parent = 'class ' + self.name
+      error(filename, linenum, 'whitespace/indent', 3,
+            'Closing brace should be aligned with beginning of %s' % parent)
+
+
+class _NamespaceInfo(_BlockInfo):
+  """Stores information about a namespace."""
+
+  def __init__(self, name, linenum):
+    _BlockInfo.__init__(self, False)
+    self.name = name or ''
+    self.starting_linenum = linenum
+
+  def CheckEnd(self, filename, clean_lines, linenum, error):
+    """Check end of namespace comments."""
+    line = clean_lines.raw_lines[linenum]
+
+    # Check how many lines is enclosed in this namespace.  Don't issue
+    # warning for missing namespace comments if there aren't enough
+    # lines.  However, do apply checks if there is already an end of
+    # namespace comment and it's incorrect.
+    #
+    # TODO(unknown): We always want to check end of namespace comments
+    # if a namespace is large, but sometimes we also want to apply the
+    # check if a short namespace contained nontrivial things (something
+    # other than forward declarations).  There is currently no logic on
+    # deciding what these nontrivial things are, so this check is
+    # triggered by namespace size only, which works most of the time.
+    if (linenum - self.starting_linenum < 10
+        and not Match(r'};*\s*(//|/\*).*\bnamespace\b', line)):
+      return
+
+    # Look for matching comment at end of namespace.
+    #
+    # Note that we accept C style "/* */" comments for terminating
+    # namespaces, so that code that terminate namespaces inside
+    # preprocessor macros can be cpplint clean.
+    #
+    # We also accept stuff like "// end of namespace <name>." with the
+    # period at the end.
+    #
+    # Besides these, we don't accept anything else, otherwise we might
+    # get false negatives when existing comment is a substring of the
+    # expected namespace.
+    if self.name:
+      # Named namespace
+      if not Match((r'};*\s*(//|/\*).*\bnamespace\s+' + re.escape(self.name) +
+                    r'[\*/\.\\\s]*$'),
+                   line):
+        error(filename, linenum, 'readability/namespace', 5,
+              'Namespace should be terminated with "// namespace %s"' %
+              self.name)
+    else:
+      # Anonymous namespace
+      if not Match(r'};*\s*(//|/\*).*\bnamespace[\*/\.\\\s]*$', line):
+        error(filename, linenum, 'readability/namespace', 5,
+              'Namespace should be terminated with "// namespace"')
+
+
+class _PreprocessorInfo(object):
+  """Stores checkpoints of nesting stacks when #if/#else is seen."""
+
+  def __init__(self, stack_before_if):
+    # The entire nesting stack before #if
+    self.stack_before_if = stack_before_if
+
+    # The entire nesting stack up to #else
+    self.stack_before_else = []
+
+    # Whether we have already seen #else or #elif
+    self.seen_else = False
+
+
+class _NestingState(object):
+  """Holds states related to parsing braces."""
+
+  def __init__(self):
+    # Stack for tracking all braces.  An object is pushed whenever we
+    # see a "{", and popped when we see a "}".  Only 3 types of
+    # objects are possible:
+    # - _ClassInfo: a class or struct.
+    # - _NamespaceInfo: a namespace.
+    # - _BlockInfo: some other type of block.
+    self.stack = []
+
+    # Stack of _PreprocessorInfo objects.
+    self.pp_stack = []
+
+  def SeenOpenBrace(self):
+    """Check if we have seen the opening brace for the innermost block.
+
+    Returns:
+      True if we have seen the opening brace, False if the innermost
+      block is still expecting an opening brace.
+    """
+    return (not self.stack) or self.stack[-1].seen_open_brace
+
+  def InNamespaceBody(self):
+    """Check if we are currently one level inside a namespace body.
+
+    Returns:
+      True if top of the stack is a namespace block, False otherwise.
+    """
+    return self.stack and isinstance(self.stack[-1], _NamespaceInfo)
+
+  def UpdatePreprocessor(self, line):
+    """Update preprocessor stack.
+
+    We need to handle preprocessors due to classes like this:
+      #ifdef SWIG
+      struct ResultDetailsPageElementExtensionPoint {
+      #else
+      struct ResultDetailsPageElementExtensionPoint : public Extension {
+      #endif
+
+    We make the following assumptions (good enough for most files):
+    - Preprocessor condition evaluates to true from #if up to first
+      #else/#elif/#endif.
+
+    - Preprocessor condition evaluates to false from #else/#elif up
+      to #endif.  We still perform lint checks on these lines, but
+      these do not affect nesting stack.
+
+    Args:
+      line: current line to check.
+    """
+    if Match(r'^\s*#\s*(if|ifdef|ifndef)\b', line):
+      # Beginning of #if block, save the nesting stack here.  The saved
+      # stack will allow us to restore the parsing state in the #else case.
+      self.pp_stack.append(_PreprocessorInfo(copy.deepcopy(self.stack)))
+    elif Match(r'^\s*#\s*(else|elif)\b', line):
+      # Beginning of #else block
+      if self.pp_stack:
+        if not self.pp_stack[-1].seen_else:
+          # This is the first #else or #elif block.  Remember the
+          # whole nesting stack up to this point.  This is what we
+          # keep after the #endif.
+          self.pp_stack[-1].seen_else = True
+          self.pp_stack[-1].stack_before_else = copy.deepcopy(self.stack)
+
+        # Restore the stack to how it was before the #if
+        self.stack = copy.deepcopy(self.pp_stack[-1].stack_before_if)
+      else:
+        # TODO(unknown): unexpected #else, issue warning?
+        pass
+    elif Match(r'^\s*#\s*endif\b', line):
+      # End of #if or #else blocks.
+      if self.pp_stack:
+        # If we saw an #else, we will need to restore the nesting
+        # stack to its former state before the #else, otherwise we
+        # will just continue from where we left off.
+        if self.pp_stack[-1].seen_else:
+          # Here we can just use a shallow copy since we are the last
+          # reference to it.
+          self.stack = self.pp_stack[-1].stack_before_else
+        # Drop the corresponding #if
+        self.pp_stack.pop()
+      else:
+        # TODO(unknown): unexpected #endif, issue warning?
+        pass
+
+  def Update(self, filename, clean_lines, linenum, error):
+    """Update nesting state with current line.
+
+    Args:
+      filename: The name of the current file.
+      clean_lines: A CleansedLines instance containing the file.
+      linenum: The number of the line to check.
+      error: The function to call with any errors found.
+    """
+    line = clean_lines.elided[linenum]
+
+    # Update pp_stack first
+    self.UpdatePreprocessor(line)
+
+    # Count parentheses.  This is to avoid adding struct arguments to
+    # the nesting stack.
+    if self.stack:
+      inner_block = self.stack[-1]
+      depth_change = line.count('(') - line.count(')')
+      inner_block.open_parentheses += depth_change
+
+      # Also check if we are starting or ending an inline assembly block.
+      if inner_block.inline_asm in (_NO_ASM, _END_ASM):
+        if (depth_change != 0 and
+            inner_block.open_parentheses == 1 and
+            _MATCH_ASM.match(line)):
+          # Enter assembly block
+          inner_block.inline_asm = _INSIDE_ASM
+        else:
+          # Not entering assembly block.  If previous line was _END_ASM,
+          # we will now shift to _NO_ASM state.
+          inner_block.inline_asm = _NO_ASM
+      elif (inner_block.inline_asm == _INSIDE_ASM and
+            inner_block.open_parentheses == 0):
+        # Exit assembly block
+        inner_block.inline_asm = _END_ASM
+
+    # Consume namespace declaration at the beginning of the line.  Do
+    # this in a loop so that we catch same line declarations like this:
+    #   namespace proto2 { namespace bridge { class MessageSet; } }
+    while True:
+      # Match start of namespace.  The "\b\s*" below catches namespace
+      # declarations even if it weren't followed by a whitespace, this
+      # is so that we don't confuse our namespace checker.  The
+      # missing spaces will be flagged by CheckSpacing.
+      namespace_decl_match = Match(r'^\s*namespace\b\s*([:\w]+)?(.*)$', line)
+      if not namespace_decl_match:
+        break
+
+      new_namespace = _NamespaceInfo(namespace_decl_match.group(1), linenum)
+      self.stack.append(new_namespace)
+
+      line = namespace_decl_match.group(2)
+      if line.find('{') != -1:
+        new_namespace.seen_open_brace = True
+        line = line[line.find('{') + 1:]
+
+    # Look for a class declaration in whatever is left of the line
+    # after parsing namespaces.  The regexp accounts for decorated classes
+    # such as in:
+    #   class LOCKABLE API Object {
+    #   };
+    #
+    # Templates with class arguments may confuse the parser, for example:
+    #   template <class T
+    #             class Comparator = less<T>,
+    #             class Vector = vector<T> >
+    #   class HeapQueue {
+    #
+    # Because this parser has no nesting state about templates, by the
+    # time it saw "class Comparator", it may think that it's a new class.
+    # Nested templates have a similar problem:
+    #   template <
+    #       typename ExportedType,
+    #       typename TupleType,
+    #       template <typename, typename> class ImplTemplate>
+    #
+    # To avoid these cases, we ignore classes that are followed by '=' or '>'
+    class_decl_match = Match(
+        r'\s*(template\s*<[\w\s<>,:]*>\s*)?'
+        r'(class|struct)\s+([A-Z_]+\s+)*(\w+(?:::\w+)*)'
+        r'(([^=>]|<[^<>]*>|<[^<>]*<[^<>]*>\s*>)*)$', line)
+    if (class_decl_match and
+        (not self.stack or self.stack[-1].open_parentheses == 0)):
+      self.stack.append(_ClassInfo(
+          class_decl_match.group(4), class_decl_match.group(2),
+          clean_lines, linenum))
+      line = class_decl_match.group(5)
+
+    # If we have not yet seen the opening brace for the innermost block,
+    # run checks here.
+    if not self.SeenOpenBrace():
+      self.stack[-1].CheckBegin(filename, clean_lines, linenum, error)
+
+    # Update access control if we are inside a class/struct
+    if self.stack and isinstance(self.stack[-1], _ClassInfo):
+      classinfo = self.stack[-1]
+      access_match = Match(
+          r'^(.*)\b(public|private|protected|signals)(\s+(?:slots\s*)?)?'
+          r':(?:[^:]|$)',
+          line)
+      if access_match:
+        classinfo.access = access_match.group(2)
+
+        # Check that access keywords are indented +1 space.  Skip this
+        # check if the keywords are not preceded by whitespaces.
+        indent = access_match.group(1)
+        if (len(indent) != classinfo.class_indent + 1 and
+            Match(r'^\s*$', indent)):
+          if classinfo.is_struct:
+            parent = 'struct ' + classinfo.name
+          else:
+            parent = 'class ' + classinfo.name
+          slots = ''
+          if access_match.group(3):
+            slots = access_match.group(3)
+          error(filename, linenum, 'whitespace/indent', 3,
+                '%s%s: should be indented +1 space inside %s' % (
+                    access_match.group(2), slots, parent))
+
+    # Consume braces or semicolons from what's left of the line
+    while True:
+      # Match first brace, semicolon, or closed parenthesis.
+      matched = Match(r'^[^{;)}]*([{;)}])(.*)$', line)
+      if not matched:
+        break
+
+      token = matched.group(1)
+      if token == '{':
+        # If namespace or class hasn't seen a opening brace yet, mark
+        # namespace/class head as complete.  Push a new block onto the
+        # stack otherwise.
+        if not self.SeenOpenBrace():
+          self.stack[-1].seen_open_brace = True
+        else:
+          self.stack.append(_BlockInfo(True))
+          if _MATCH_ASM.match(line):
+            self.stack[-1].inline_asm = _BLOCK_ASM
+      elif token == ';' or token == ')':
+        # If we haven't seen an opening brace yet, but we already saw
+        # a semicolon, this is probably a forward declaration.  Pop
+        # the stack for these.
+        #
+        # Similarly, if we haven't seen an opening brace yet, but we
+        # already saw a closing parenthesis, then these are probably
+        # function arguments with extra "class" or "struct" keywords.
+        # Also pop these stack for these.
+        if not self.SeenOpenBrace():
+          self.stack.pop()
+      else:  # token == '}'
+        # Perform end of block checks and pop the stack.
+        if self.stack:
+          self.stack[-1].CheckEnd(filename, clean_lines, linenum, error)
+          self.stack.pop()
+      line = matched.group(2)
+
+  def InnermostClass(self):
+    """Get class info on the top of the stack.
+
+    Returns:
+      A _ClassInfo object if we are inside a class, or None otherwise.
+    """
+    for i in range(len(self.stack), 0, -1):
+      classinfo = self.stack[i - 1]
+      if isinstance(classinfo, _ClassInfo):
+        return classinfo
+    return None
+
+  def CheckCompletedBlocks(self, filename, error):
+    """Checks that all classes and namespaces have been completely parsed.
+
+    Call this when all lines in a file have been processed.
+    Args:
+      filename: The name of the current file.
+      error: The function to call with any errors found.
+    """
+    # Note: This test can result in false positives if #ifdef constructs
+    # get in the way of brace matching. See the testBuildClass test in
+    # cpplint_unittest.py for an example of this.
+    for obj in self.stack:
+      if isinstance(obj, _ClassInfo):
+        error(filename, obj.starting_linenum, 'build/class', 5,
+              'Failed to find complete declaration of class %s' %
+              obj.name)
+      elif isinstance(obj, _NamespaceInfo):
+        error(filename, obj.starting_linenum, 'build/namespaces', 5,
+              'Failed to find complete declaration of namespace %s' %
+              obj.name)
+
+
+def CheckForNonStandardConstructs(filename, clean_lines, linenum,
+                                  nesting_state, error):
+  r"""Logs an error if we see certain non-ANSI constructs ignored by gcc-2.
+
+  Complain about several constructs which gcc-2 accepts, but which are
+  not standard C++.  Warning about these in lint is one way to ease the
+  transition to new compilers.
+  - put storage class first (e.g. "static const" instead of "const static").
+  - "%lld" instead of %qd" in printf-type functions.
+  - "%1$d" is non-standard in printf-type functions.
+  - "\%" is an undefined character escape sequence.
+  - text after #endif is not allowed.
+  - invalid inner-style forward declaration.
+  - >? and <? operators, and their >?= and <?= cousins.
+
+  Additionally, check for constructor/destructor style violations and reference
+  members, as it is very convenient to do so while checking for
+  gcc-2 compliance.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: A callable to which errors are reported, which takes 4 arguments:
+           filename, line number, error level, and message
+  """
+
+  # Remove comments from the line, but leave in strings for now.
+  line = clean_lines.lines[linenum]
+
+  if Search(r'printf\s*\(.*".*%[-+ ]?\d*q', line):
+    error(filename, linenum, 'runtime/printf_format', 3,
+          '%q in format strings is deprecated.  Use %ll instead.')
+
+  if Search(r'printf\s*\(.*".*%\d+\$', line):
+    error(filename, linenum, 'runtime/printf_format', 2,
+          '%N$ formats are unconventional.  Try rewriting to avoid them.')
+
+  # Remove escaped backslashes before looking for undefined escapes.
+  line = line.replace('\\\\', '')
+
+  if Search(r'("|\').*\\(%|\[|\(|{)', line):
+    error(filename, linenum, 'build/printf_format', 3,
+          '%, [, (, and { are undefined character escapes.  Unescape them.')
+
+  # For the rest, work with both comments and strings removed.
+  line = clean_lines.elided[linenum]
+
+  if Search(r'\b(const|volatile|void|char|short|int|long'
+            r'|float|double|signed|unsigned'
+            r'|schar|u?int8|u?int16|u?int32|u?int64)'
+            r'\s+(register|static|extern|typedef)\b',
+            line):
+    error(filename, linenum, 'build/storage_class', 5,
+          'Storage class (static, extern, typedef, etc) should be first.')
+
+  if Match(r'\s*#\s*endif\s*[^/\s]+', line):
+    error(filename, linenum, 'build/endif_comment', 5,
+          'Uncommented text after #endif is non-standard.  Use a comment.')
+
+  if Match(r'\s*class\s+(\w+\s*::\s*)+\w+\s*;', line):
+    error(filename, linenum, 'build/forward_decl', 5,
+          'Inner-style forward declarations are invalid.  Remove this line.')
+
+  if Search(r'(\w+|[+-]?\d+(\.\d*)?)\s*(<|>)\?=?\s*(\w+|[+-]?\d+)(\.\d*)?',
+            line):
+    error(filename, linenum, 'build/deprecated', 3,
+          '>? and <? (max and min) operators are non-standard and deprecated.')
+
+  if Search(r'^\s*const\s*string\s*&\s*\w+\s*;', line):
+    # TODO(unknown): Could it be expanded safely to arbitrary references,
+    # without triggering too many false positives? The first
+    # attempt triggered 5 warnings for mostly benign code in the regtest, hence
+    # the restriction.
+    # Here's the original regexp, for the reference:
+    # type_name = r'\w+((\s*::\s*\w+)|(\s*<\s*\w+?\s*>))?'
+    # r'\s*const\s*' + type_name + '\s*&\s*\w+\s*;'
+    error(filename, linenum, 'runtime/member_string_references', 2,
+          'const string& members are dangerous. It is much better to use '
+          'alternatives, such as pointers or simple constants.')
+
+  # Everything else in this function operates on class declarations.
+  # Return early if the top of the nesting stack is not a class, or if
+  # the class head is not completed yet.
+  classinfo = nesting_state.InnermostClass()
+  if not classinfo or not classinfo.seen_open_brace:
+    return
+
+  # The class may have been declared with namespace or classname qualifiers.
+  # The constructor and destructor will not have those qualifiers.
+  base_classname = classinfo.name.split('::')[-1]
+
+  # Look for single-argument constructors that aren't marked explicit.
+  # Technically a valid construct, but against style.
+  args = Match(r'\s+(?:inline\s+)?%s\s*\(([^,()]+)\)'
+               % re.escape(base_classname),
+               line)
+  if (args and
+      args.group(1) != 'void' and
+      not Match(r'(const\s+)?%s(\s+const)?\s*(?:<\w+>\s*)?&'
+                % re.escape(base_classname), args.group(1).strip())):
+    error(filename, linenum, 'runtime/explicit', 5,
+          'Single-argument constructors should be marked explicit.')
+
+
+def CheckSpacingForFunctionCall(filename, line, linenum, error):
+  """Checks for the correctness of various spacing around function calls.
+
+  Args:
+    filename: The name of the current file.
+    line: The text of the line to check.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+
+  # Since function calls often occur inside if/for/while/switch
+  # expressions - which have their own, more liberal conventions - we
+  # first see if we should be looking inside such an expression for a
+  # function call, to which we can apply more strict standards.
+  fncall = line    # if there's no control flow construct, look at whole line
+  for pattern in (r'\bif\s*\((.*)\)\s*{',
+                  r'\bfor\s*\((.*)\)\s*{',
+                  r'\bwhile\s*\((.*)\)\s*[{;]',
+                  r'\bswitch\s*\((.*)\)\s*{'):
+    match = Search(pattern, line)
+    if match:
+      fncall = match.group(1)    # look inside the parens for function calls
+      break
+
+  # Except in if/for/while/switch, there should never be space
+  # immediately inside parens (eg "f( 3, 4 )").  We make an exception
+  # for nested parens ( (a+b) + c ).  Likewise, there should never be
+  # a space before a ( when it's a function argument.  I assume it's a
+  # function argument when the char before the whitespace is legal in
+  # a function name (alnum + _) and we're not starting a macro. Also ignore
+  # pointers and references to arrays and functions coz they're too tricky:
+  # we use a very simple way to recognize these:
+  # " (something)(maybe-something)" or
+  # " (something)(maybe-something," or
+  # " (something)[something]"
+  # Note that we assume the contents of [] to be short enough that
+  # they'll never need to wrap.
+  if (  # Ignore control structures.
+      not Search(r'\b(if|for|while|switch|return|new|delete|catch|sizeof)\b',
+                 fncall) and
+      # Ignore pointers/references to functions.
+      not Search(r' \([^)]+\)\([^)]*(\)|,$)', fncall) and
+      # Ignore pointers/references to arrays.
+      not Search(r' \([^)]+\)\[[^\]]+\]', fncall)):
+    if Search(r'\w\s*\(\s(?!\s*\\$)', fncall):      # a ( used for a fn call
+      error(filename, linenum, 'whitespace/parens', 4,
+            'Extra space after ( in function call')
+    elif Search(r'\(\s+(?!(\s*\\)|\()', fncall):
+      error(filename, linenum, 'whitespace/parens', 2,
+            'Extra space after (')
+    if (Search(r'\w\s+\(', fncall) and
+        not Search(r'#\s*define|typedef', fncall) and
+        not Search(r'\w\s+\((\w+::)*\*\w+\)\(', fncall)):
+      error(filename, linenum, 'whitespace/parens', 4,
+            'Extra space before ( in function call')
+    # If the ) is followed only by a newline or a { + newline, assume it's
+    # part of a control statement (if/while/etc), and don't complain
+    if Search(r'[^)]\s+\)\s*[^{\s]', fncall):
+      # If the closing parenthesis is preceded by only whitespaces,
+      # try to give a more descriptive error message.
+      if Search(r'^\s+\)', fncall):
+        error(filename, linenum, 'whitespace/parens', 2,
+              'Closing ) should be moved to the previous line')
+      else:
+        error(filename, linenum, 'whitespace/parens', 2,
+              'Extra space before )')
+
+
+def IsBlankLine(line):
+  """Returns true if the given line is blank.
+
+  We consider a line to be blank if the line is empty or consists of
+  only white spaces.
+
+  Args:
+    line: A line of a string.
+
+  Returns:
+    True, if the given line is blank.
+  """
+  return not line or line.isspace()
+
+
+def CheckForFunctionLengths(filename, clean_lines, linenum,
+                            function_state, error):
+  """Reports for long function bodies.
+
+  For an overview why this is done, see:
+  http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml#Write_Short_Functions
+
+  Uses a simplistic algorithm assuming other style guidelines
+  (especially spacing) are followed.
+  Only checks unindented functions, so class members are unchecked.
+  Trivial bodies are unchecked, so constructors with huge initializer lists
+  may be missed.
+  Blank/comment lines are not counted so as to avoid encouraging the removal
+  of vertical space and comments just to get through a lint check.
+  NOLINT *on the last line of a function* disables this check.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    function_state: Current function name and lines in body so far.
+    error: The function to call with any errors found.
+  """
+  lines = clean_lines.lines
+  line = lines[linenum]
+  raw = clean_lines.raw_lines
+  raw_line = raw[linenum]
+  joined_line = ''
+
+  starting_func = False
+  regexp = r'(\w(\w|::|\*|\&|\s)*)\('  # decls * & space::name( ...
+  match_result = Match(regexp, line)
+  if match_result:
+    # If the name is all caps and underscores, figure it's a macro and
+    # ignore it, unless it's TEST or TEST_F.
+    function_name = match_result.group(1).split()[-1]
+    if function_name == 'TEST' or function_name == 'TEST_F' or (
+        not Match(r'[A-Z_]+$', function_name)):
+      starting_func = True
+
+  if starting_func:
+    body_found = False
+    for start_linenum in xrange(linenum, clean_lines.NumLines()):
+      start_line = lines[start_linenum]
+      joined_line += ' ' + start_line.lstrip()
+      if Search(r'(;|})', start_line):  # Declarations and trivial functions
+        body_found = True
+        break                              # ... ignore
+      elif Search(r'{', start_line):
+        body_found = True
+        function = Search(r'((\w|:)*)\(', line).group(1)
+        if Match(r'TEST', function):    # Handle TEST... macros
+          parameter_regexp = Search(r'(\(.*\))', joined_line)
+          if parameter_regexp:             # Ignore bad syntax
+            function += parameter_regexp.group(1)
+        else:
+          function += '()'
+        function_state.Begin(function)
+        break
+    if not body_found:
+      # No body for the function (or evidence of a non-function) was found.
+      error(filename, linenum, 'readability/fn_size', 5,
+            'Lint failed to find start of function body.')
+  elif Match(r'^\}\s*$', line):  # function end
+    function_state.Check(error, filename, linenum)
+    function_state.End()
+  elif not Match(r'^\s*$', line):
+    function_state.Count()  # Count non-blank/non-comment lines.
+
+
+_RE_PATTERN_TODO = re.compile(r'^//(\s*)TODO(\(.+?\))?:?(\s|$)?')
+
+
+def CheckComment(comment, filename, linenum, error):
+  """Checks for common mistakes in TODO comments.
+
+  Args:
+    comment: The text of the comment from the line in question.
+    filename: The name of the current file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  match = _RE_PATTERN_TODO.match(comment)
+  if match:
+    # One whitespace is correct; zero whitespace is handled elsewhere.
+    leading_whitespace = match.group(1)
+    if len(leading_whitespace) > 1:
+      error(filename, linenum, 'whitespace/todo', 2,
+            'Too many spaces before TODO')
+
+    username = match.group(2)
+    if not username:
+      error(filename, linenum, 'readability/todo', 2,
+            'Missing username in TODO; it should look like '
+            '"// TODO(my_username): Stuff."')
+
+    middle_whitespace = match.group(3)
+    # Comparisons made explicit for correctness -- pylint: disable=g-explicit-bool-comparison
+    if middle_whitespace != ' ' and middle_whitespace != '':
+      error(filename, linenum, 'whitespace/todo', 2,
+            'TODO(my_username) should be followed by a space')
+
+def CheckAccess(filename, clean_lines, linenum, nesting_state, error):
+  """Checks for improper use of DISALLOW* macros.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]  # get rid of comments and strings
+
+  matched = Match((r'\s*(DISALLOW_COPY_AND_ASSIGN|'
+                   r'DISALLOW_EVIL_CONSTRUCTORS|'
+                   r'DISALLOW_IMPLICIT_CONSTRUCTORS)'), line)
+  if not matched:
+    return
+  if nesting_state.stack and isinstance(nesting_state.stack[-1], _ClassInfo):
+    if nesting_state.stack[-1].access != 'private':
+      error(filename, linenum, 'readability/constructors', 3,
+            '%s must be in the private: section' % matched.group(1))
+
+  else:
+    # Found DISALLOW* macro outside a class declaration, or perhaps it
+    # was used inside a function when it should have been part of the
+    # class declaration.  We could issue a warning here, but it
+    # probably resulted in a compiler error already.
+    pass
+
+
+def FindNextMatchingAngleBracket(clean_lines, linenum, init_suffix):
+  """Find the corresponding > to close a template.
+
+  Args:
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: Current line number.
+    init_suffix: Remainder of the current line after the initial <.
+
+  Returns:
+    True if a matching bracket exists.
+  """
+  line = init_suffix
+  nesting_stack = ['<']
+  while True:
+    # Find the next operator that can tell us whether < is used as an
+    # opening bracket or as a less-than operator.  We only want to
+    # warn on the latter case.
+    #
+    # We could also check all other operators and terminate the search
+    # early, e.g. if we got something like this "a<b+c", the "<" is
+    # most likely a less-than operator, but then we will get false
+    # positives for default arguments and other template expressions.
+    match = Search(r'^[^<>(),;\[\]]*([<>(),;\[\]])(.*)$', line)
+    if match:
+      # Found an operator, update nesting stack
+      operator = match.group(1)
+      line = match.group(2)
+
+      if nesting_stack[-1] == '<':
+        # Expecting closing angle bracket
+        if operator in ('<', '(', '['):
+          nesting_stack.append(operator)
+        elif operator == '>':
+          nesting_stack.pop()
+          if not nesting_stack:
+            # Found matching angle bracket
+            return True
+        elif operator == ',':
+          # Got a comma after a bracket, this is most likely a template
+          # argument.  We have not seen a closing angle bracket yet, but
+          # it's probably a few lines later if we look for it, so just
+          # return early here.
+          return True
+        else:
+          # Got some other operator.
+          return False
+
+      else:
+        # Expecting closing parenthesis or closing bracket
+        if operator in ('<', '(', '['):
+          nesting_stack.append(operator)
+        elif operator in (')', ']'):
+          # We don't bother checking for matching () or [].  If we got
+          # something like (] or [), it would have been a syntax error.
+          nesting_stack.pop()
+
+    else:
+      # Scan the next line
+      linenum += 1
+      if linenum >= len(clean_lines.elided):
+        break
+      line = clean_lines.elided[linenum]
+
+  # Exhausted all remaining lines and still no matching angle bracket.
+  # Most likely the input was incomplete, otherwise we should have
+  # seen a semicolon and returned early.
+  return True
+
+
+def FindPreviousMatchingAngleBracket(clean_lines, linenum, init_prefix):
+  """Find the corresponding < that started a template.
+
+  Args:
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: Current line number.
+    init_prefix: Part of the current line before the initial >.
+
+  Returns:
+    True if a matching bracket exists.
+  """
+  line = init_prefix
+  nesting_stack = ['>']
+  while True:
+    # Find the previous operator
+    match = Search(r'^(.*)([<>(),;\[\]])[^<>(),;\[\]]*$', line)
+    if match:
+      # Found an operator, update nesting stack
+      operator = match.group(2)
+      line = match.group(1)
+
+      if nesting_stack[-1] == '>':
+        # Expecting opening angle bracket
+        if operator in ('>', ')', ']'):
+          nesting_stack.append(operator)
+        elif operator == '<':
+          nesting_stack.pop()
+          if not nesting_stack:
+            # Found matching angle bracket
+            return True
+        elif operator == ',':
+          # Got a comma before a bracket, this is most likely a
+          # template argument.  The opening angle bracket is probably
+          # there if we look for it, so just return early here.
+          return True
+        else:
+          # Got some other operator.
+          return False
+
+      else:
+        # Expecting opening parenthesis or opening bracket
+        if operator in ('>', ')', ']'):
+          nesting_stack.append(operator)
+        elif operator in ('(', '['):
+          nesting_stack.pop()
+
+    else:
+      # Scan the previous line
+      linenum -= 1
+      if linenum < 0:
+        break
+      line = clean_lines.elided[linenum]
+
+  # Exhausted all earlier lines and still no matching angle bracket.
+  return False
+
+
+def CheckSpacing(filename, clean_lines, linenum, nesting_state, error):
+  """Checks for the correctness of various spacing issues in the code.
+
+  Things we check for: spaces around operators, spaces after
+  if/for/while/switch, no spaces around parens in function calls, two
+  spaces between code and comment, don't start a block with a blank
+  line, don't end a function with a blank line, don't add a blank line
+  after public/protected/private, don't have too many blank lines in a row.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: The function to call with any errors found.
+  """
+
+  # Don't use "elided" lines here, otherwise we can't check commented lines.
+  # Don't want to use "raw" either, because we don't want to check inside C++11
+  # raw strings,
+  raw = clean_lines.lines_without_raw_strings
+  line = raw[linenum]
+
+  # Before nixing comments, check if the line is blank for no good
+  # reason.  This includes the first line after a block is opened, and
+  # blank lines at the end of a function (ie, right before a line like '}'
+  #
+  # Skip all the blank line checks if we are immediately inside a
+  # namespace body.  In other words, don't issue blank line warnings
+  # for this block:
+  #   namespace {
+  #
+  #   }
+  #
+  # A warning about missing end of namespace comments will be issued instead.
+  if IsBlankLine(line) and not nesting_state.InNamespaceBody():
+    elided = clean_lines.elided
+    prev_line = elided[linenum - 1]
+    prevbrace = prev_line.rfind('{')
+    # TODO(unknown): Don't complain if line before blank line, and line after,
+    #                both start with alnums and are indented the same amount.
+    #                This ignores whitespace at the start of a namespace block
+    #                because those are not usually indented.
+    if prevbrace != -1 and prev_line[prevbrace:].find('}') == -1:
+      # OK, we have a blank line at the start of a code block.  Before we
+      # complain, we check if it is an exception to the rule: The previous
+      # non-empty line has the parameters of a function header that are indented
+      # 4 spaces (because they did not fit in a 80 column line when placed on
+      # the same line as the function name).  We also check for the case where
+      # the previous line is indented 6 spaces, which may happen when the
+      # initializers of a constructor do not fit into a 80 column line.
+      exception = False
+      if Match(r' {6}\w', prev_line):  # Initializer list?
+        # We are looking for the opening column of initializer list, which
+        # should be indented 4 spaces to cause 6 space indentation afterwards.
+        search_position = linenum-2
+        while (search_position >= 0
+               and Match(r' {6}\w', elided[search_position])):
+          search_position -= 1
+        exception = (search_position >= 0
+                     and elided[search_position][:5] == '    :')
+      else:
+        # Search for the function arguments or an initializer list.  We use a
+        # simple heuristic here: If the line is indented 4 spaces; and we have a
+        # closing paren, without the opening paren, followed by an opening brace
+        # or colon (for initializer lists) we assume that it is the last line of
+        # a function header.  If we have a colon indented 4 spaces, it is an
+        # initializer list.
+        exception = (Match(r' {4}\w[^\(]*\)\s*(const\s*)?(\{\s*$|:)',
+                           prev_line)
+                     or Match(r' {4}:', prev_line))
+
+      if not exception:
+        error(filename, linenum, 'whitespace/blank_line', 2,
+              'Redundant blank line at the start of a code block '
+              'should be deleted.')
+    # Ignore blank lines at the end of a block in a long if-else
+    # chain, like this:
+    #   if (condition1) {
+    #     // Something followed by a blank line
+    #
+    #   } else if (condition2) {
+    #     // Something else
+    #   }
+    if linenum + 1 < clean_lines.NumLines():
+      next_line = raw[linenum + 1]
+      if (next_line
+          and Match(r'\s*}', next_line)
+          and next_line.find('} else ') == -1):
+        error(filename, linenum, 'whitespace/blank_line', 3,
+              'Redundant blank line at the end of a code block '
+              'should be deleted.')
+
+    matched = Match(r'\s*(public|protected|private):', prev_line)
+    if matched:
+      error(filename, linenum, 'whitespace/blank_line', 3,
+            'Do not leave a blank line after "%s:"' % matched.group(1))
+
+  # Next, we complain if there's a comment too near the text
+  commentpos = line.find('//')
+  if commentpos != -1:
+    # Check if the // may be in quotes.  If so, ignore it
+    # Comparisons made explicit for clarity -- pylint: disable=g-explicit-bool-comparison
+    if (line.count('"', 0, commentpos) -
+        line.count('\\"', 0, commentpos)) % 2 == 0:   # not in quotes
+      # Allow one space for new scopes, two spaces otherwise:
+      if (not Match(r'^\s*{ //', line) and
+          ((commentpos >= 1 and
+            line[commentpos-1] not in string.whitespace) or
+           (commentpos >= 2 and
+            line[commentpos-2] not in string.whitespace))):
+        error(filename, linenum, 'whitespace/comments', 2,
+              'At least two spaces is best between code and comments')
+      # There should always be a space between the // and the comment
+      commentend = commentpos + 2
+      if commentend < len(line) and not line[commentend] == ' ':
+        # but some lines are exceptions -- e.g. if they're big
+        # comment delimiters like:
+        # //----------------------------------------------------------
+        # or are an empty C++ style Doxygen comment, like:
+        # ///
+        # or C++ style Doxygen comments placed after the variable:
+        # ///<  Header comment
+        # //!<  Header comment
+        # or they begin with multiple slashes followed by a space:
+        # //////// Header comment
+        match = (Search(r'[=/-]{4,}\s*$', line[commentend:]) or
+                 Search(r'^/$', line[commentend:]) or
+                 Search(r'^!< ', line[commentend:]) or
+                 Search(r'^/< ', line[commentend:]) or
+                 Search(r'^/+ ', line[commentend:]))
+        if not match:
+          error(filename, linenum, 'whitespace/comments', 4,
+                'Should have a space between // and comment')
+      CheckComment(line[commentpos:], filename, linenum, error)
+
+  line = clean_lines.elided[linenum]  # get rid of comments and strings
+
+  # Don't try to do spacing checks for operator methods
+  line = re.sub(r'operator(==|!=|<|<<|<=|>=|>>|>)\(', 'operator\(', line)
+
+  # We allow no-spaces around = within an if: "if ( (a=Foo()) == 0 )".
+  # Otherwise not.  Note we only check for non-spaces on *both* sides;
+  # sometimes people put non-spaces on one side when aligning ='s among
+  # many lines (not that this is behavior that I approve of...)
+  if Search(r'[\w.]=[\w.]', line) and not Search(r'\b(if|while) ', line):
+    error(filename, linenum, 'whitespace/operators', 4,
+          'Missing spaces around =')
+
+  # It's ok not to have spaces around binary operators like + - * /, but if
+  # there's too little whitespace, we get concerned.  It's hard to tell,
+  # though, so we punt on this one for now.  TODO.
+
+  # You should always have whitespace around binary operators.
+  #
+  # Check <= and >= first to avoid false positives with < and >, then
+  # check non-include lines for spacing around < and >.
+  match = Search(r'[^<>=!\s](==|!=|<=|>=)[^<>=!\s]', line)
+  if match:
+    error(filename, linenum, 'whitespace/operators', 3,
+          'Missing spaces around %s' % match.group(1))
+  # We allow no-spaces around << when used like this: 10<<20, but
+  # not otherwise (particularly, not when used as streams)
+  # Also ignore using ns::operator<<;
+  match = Search(r'(operator|\S)(?:L|UL|ULL|l|ul|ull)?<<(\S)', line)
+  if (match and
+      not (match.group(1).isdigit() and match.group(2).isdigit()) and
+      not (match.group(1) == 'operator' and match.group(2) == ';')):
+    error(filename, linenum, 'whitespace/operators', 3,
+          'Missing spaces around <<')
+  elif not Match(r'#.*include', line):
+    # Avoid false positives on ->
+    reduced_line = line.replace('->', '')
+
+    # Look for < that is not surrounded by spaces.  This is only
+    # triggered if both sides are missing spaces, even though
+    # technically should should flag if at least one side is missing a
+    # space.  This is done to avoid some false positives with shifts.
+    match = Search(r'[^\s<]<([^\s=<].*)', reduced_line)
+    if (match and
+        not FindNextMatchingAngleBracket(clean_lines, linenum, match.group(1))):
+      error(filename, linenum, 'whitespace/operators', 3,
+            'Missing spaces around <')
+
+    # Look for > that is not surrounded by spaces.  Similar to the
+    # above, we only trigger if both sides are missing spaces to avoid
+    # false positives with shifts.
+    match = Search(r'^(.*[^\s>])>[^\s=>]', reduced_line)
+    if (match and
+        not FindPreviousMatchingAngleBracket(clean_lines, linenum,
+                                             match.group(1))):
+      error(filename, linenum, 'whitespace/operators', 3,
+            'Missing spaces around >')
+
+  # We allow no-spaces around >> for almost anything.  This is because
+  # C++11 allows ">>" to close nested templates, which accounts for
+  # most cases when ">>" is not followed by a space.
+  #
+  # We still warn on ">>" followed by alpha character, because that is
+  # likely due to ">>" being used for right shifts, e.g.:
+  #   value >> alpha
+  #
+  # When ">>" is used to close templates, the alphanumeric letter that
+  # follows would be part of an identifier, and there should still be
+  # a space separating the template type and the identifier.
+  #   type<type<type>> alpha
+  match = Search(r'>>[a-zA-Z_]', line)
+  if match:
+    error(filename, linenum, 'whitespace/operators', 3,
+          'Missing spaces around >>')
+
+  # There shouldn't be space around unary operators
+  match = Search(r'(!\s|~\s|[\s]--[\s;]|[\s]\+\+[\s;])', line)
+  if match:
+    error(filename, linenum, 'whitespace/operators', 4,
+          'Extra space for operator %s' % match.group(1))
+
+  # A pet peeve of mine: no spaces after an if, while, switch, or for
+  match = Search(r' (if\(|for\(|while\(|switch\()', line)
+  if match:
+    error(filename, linenum, 'whitespace/parens', 5,
+          'Missing space before ( in %s' % match.group(1))
+
+  # For if/for/while/switch, the left and right parens should be
+  # consistent about how many spaces are inside the parens, and
+  # there should either be zero or one spaces inside the parens.
+  # We don't want: "if ( foo)" or "if ( foo   )".
+  # Exception: "for ( ; foo; bar)" and "for (foo; bar; )" are allowed.
+  match = Search(r'\b(if|for|while|switch)\s*'
+                 r'\(([ ]*)(.).*[^ ]+([ ]*)\)\s*{\s*$',
+                 line)
+  if match:
+    if len(match.group(2)) != len(match.group(4)):
+      if not (match.group(3) == ';' and
+              len(match.group(2)) == 1 + len(match.group(4)) or
+              not match.group(2) and Search(r'\bfor\s*\(.*; \)', line)):
+        error(filename, linenum, 'whitespace/parens', 5,
+              'Mismatching spaces inside () in %s' % match.group(1))
+    if len(match.group(2)) not in [0, 1]:
+      error(filename, linenum, 'whitespace/parens', 5,
+            'Should have zero or one spaces inside ( and ) in %s' %
+            match.group(1))
+
+  # You should always have a space after a comma (either as fn arg or operator)
+  #
+  # This does not apply when the non-space character following the
+  # comma is another comma, since the only time when that happens is
+  # for empty macro arguments.
+  #
+  # We run this check in two passes: first pass on elided lines to
+  # verify that lines contain missing whitespaces, second pass on raw
+  # lines to confirm that those missing whitespaces are not due to
+  # elided comments.
+  if Search(r',[^,\s]', line) and Search(r',[^,\s]', raw[linenum]):
+    error(filename, linenum, 'whitespace/comma', 3,
+          'Missing space after ,')
+
+  # You should always have a space after a semicolon
+  # except for few corner cases
+  # TODO(unknown): clarify if 'if (1) { return 1;}' is requires one more
+  # space after ;
+  if Search(r';[^\s};\\)/]', line):
+    error(filename, linenum, 'whitespace/semicolon', 3,
+          'Missing space after ;')
+
+  # Next we will look for issues with function calls.
+  CheckSpacingForFunctionCall(filename, line, linenum, error)
+
+  # Except after an opening paren, or after another opening brace (in case of
+  # an initializer list, for instance), you should have spaces before your
+  # braces. And since you should never have braces at the beginning of a line,
+  # this is an easy test.
+  match = Match(r'^(.*[^ ({]){', line)
+  if match:
+    # Try a bit harder to check for brace initialization.  This
+    # happens in one of the following forms:
+    #   Constructor() : initializer_list_{} { ... }
+    #   Constructor{}.MemberFunction()
+    #   Type variable{};
+    #   FunctionCall(type{}, ...);
+    #   LastArgument(..., type{});
+    #   LOG(INFO) << type{} << " ...";
+    #   map_of_type[{...}] = ...;
+    #
+    # We check for the character following the closing brace, and
+    # silence the warning if it's one of those listed above, i.e.
+    # "{.;,)<]".
+    #
+    # To account for nested initializer list, we allow any number of
+    # closing braces up to "{;,)<".  We can't simply silence the
+    # warning on first sight of closing brace, because that would
+    # cause false negatives for things that are not initializer lists.
+    #   Silence this:         But not this:
+    #     Outer{                if (...) {
+    #       Inner{...}            if (...){  // Missing space before {
+    #     };                    }
+    #
+    # There is a false negative with this approach if people inserted
+    # spurious semicolons, e.g. "if (cond){};", but we will catch the
+    # spurious semicolon with a separate check.
+    (endline, endlinenum, endpos) = CloseExpression(
+        clean_lines, linenum, len(match.group(1)))
+    trailing_text = ''
+    if endpos > -1:
+      trailing_text = endline[endpos:]
+    for offset in xrange(endlinenum + 1,
+                         min(endlinenum + 3, clean_lines.NumLines() - 1)):
+      trailing_text += clean_lines.elided[offset]
+    if not Match(r'^[\s}]*[{.;,)<\]]', trailing_text):
+      error(filename, linenum, 'whitespace/braces', 5,
+            'Missing space before {')
+
+  # Make sure '} else {' has spaces.
+  if Search(r'}else', line):
+    error(filename, linenum, 'whitespace/braces', 5,
+          'Missing space before else')
+
+  # You shouldn't have spaces before your brackets, except maybe after
+  # 'delete []' or 'new char * []'.
+  if Search(r'\w\s+\[', line) and not Search(r'delete\s+\[', line):
+    error(filename, linenum, 'whitespace/braces', 5,
+          'Extra space before [')
+
+  # You shouldn't have a space before a semicolon at the end of the line.
+  # There's a special case for "for" since the style guide allows space before
+  # the semicolon there.
+  if Search(r':\s*;\s*$', line):
+    error(filename, linenum, 'whitespace/semicolon', 5,
+          'Semicolon defining empty statement. Use {} instead.')
+  elif Search(r'^\s*;\s*$', line):
+    error(filename, linenum, 'whitespace/semicolon', 5,
+          'Line contains only semicolon. If this should be an empty statement, '
+          'use {} instead.')
+  elif (Search(r'\s+;\s*$', line) and
+        not Search(r'\bfor\b', line)):
+    error(filename, linenum, 'whitespace/semicolon', 5,
+          'Extra space before last semicolon. If this should be an empty '
+          'statement, use {} instead.')
+
+  # In range-based for, we wanted spaces before and after the colon, but
+  # not around "::" tokens that might appear.
+  if (Search('for *\(.*[^:]:[^: ]', line) or
+      Search('for *\(.*[^: ]:[^:]', line)):
+    error(filename, linenum, 'whitespace/forcolon', 2,
+          'Missing space around colon in range-based for loop')
+
+
+def CheckSectionSpacing(filename, clean_lines, class_info, linenum, error):
+  """Checks for additional blank line issues related to sections.
+
+  Currently the only thing checked here is blank line before protected/private.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    class_info: A _ClassInfo objects.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  # Skip checks if the class is small, where small means 25 lines or less.
+  # 25 lines seems like a good cutoff since that's the usual height of
+  # terminals, and any class that can't fit in one screen can't really
+  # be considered "small".
+  #
+  # Also skip checks if we are on the first line.  This accounts for
+  # classes that look like
+  #   class Foo { public: ... };
+  #
+  # If we didn't find the end of the class, last_line would be zero,
+  # and the check will be skipped by the first condition.
+  if (class_info.last_line - class_info.starting_linenum <= 24 or
+      linenum <= class_info.starting_linenum):
+    return
+
+  matched = Match(r'\s*(public|protected|private):', clean_lines.lines[linenum])
+  if matched:
+    # Issue warning if the line before public/protected/private was
+    # not a blank line, but don't do this if the previous line contains
+    # "class" or "struct".  This can happen two ways:
+    #  - We are at the beginning of the class.
+    #  - We are forward-declaring an inner class that is semantically
+    #    private, but needed to be public for implementation reasons.
+    # Also ignores cases where the previous line ends with a backslash as can be
+    # common when defining classes in C macros.
+    prev_line = clean_lines.lines[linenum - 1]
+    if (not IsBlankLine(prev_line) and
+        not Search(r'\b(class|struct)\b', prev_line) and
+        not Search(r'\\$', prev_line)):
+      # Try a bit harder to find the beginning of the class.  This is to
+      # account for multi-line base-specifier lists, e.g.:
+      #   class Derived
+      #       : public Base {
+      end_class_head = class_info.starting_linenum
+      for i in range(class_info.starting_linenum, linenum):
+        if Search(r'\{\s*$', clean_lines.lines[i]):
+          end_class_head = i
+          break
+      if end_class_head < linenum - 1:
+        error(filename, linenum, 'whitespace/blank_line', 3,
+              '"%s:" should be preceded by a blank line' % matched.group(1))
+
+
+def GetPreviousNonBlankLine(clean_lines, linenum):
+  """Return the most recent non-blank line and its line number.
+
+  Args:
+    clean_lines: A CleansedLines instance containing the file contents.
+    linenum: The number of the line to check.
+
+  Returns:
+    A tuple with two elements.  The first element is the contents of the last
+    non-blank line before the current line, or the empty string if this is the
+    first non-blank line.  The second is the line number of that line, or -1
+    if this is the first non-blank line.
+  """
+
+  prevlinenum = linenum - 1
+  while prevlinenum >= 0:
+    prevline = clean_lines.elided[prevlinenum]
+    if not IsBlankLine(prevline):     # if not a blank line...
+      return (prevline, prevlinenum)
+    prevlinenum -= 1
+  return ('', -1)
+
+
+def CheckBraces(filename, clean_lines, linenum, error):
+  """Looks for misplaced braces (e.g. at the end of line).
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+
+  line = clean_lines.elided[linenum]        # get rid of comments and strings
+
+  if Match(r'\s*{\s*$', line):
+    # We allow an open brace to start a line in the case where someone is using
+    # braces in a block to explicitly create a new scope, which is commonly used
+    # to control the lifetime of stack-allocated variables.  Braces are also
+    # used for brace initializers inside function calls.  We don't detect this
+    # perfectly: we just don't complain if the last non-whitespace character on
+    # the previous non-blank line is ',', ';', ':', '(', '{', or '}', or if the
+    # previous line starts a preprocessor block.
+    prevline = GetPreviousNonBlankLine(clean_lines, linenum)[0]
+    if (not Search(r'[,;:}{(]\s*$', prevline) and
+        not Match(r'\s*#', prevline)):
+      error(filename, linenum, 'whitespace/braces', 4,
+            '{ should almost always be at the end of the previous line')
+
+  # An else clause should be on the same line as the preceding closing brace.
+  if Match(r'\s*else\s*', line):
+    prevline = GetPreviousNonBlankLine(clean_lines, linenum)[0]
+    if Match(r'\s*}\s*$', prevline):
+      error(filename, linenum, 'whitespace/newline', 4,
+            'An else should appear on the same line as the preceding }')
+
+  # If braces come on one side of an else, they should be on both.
+  # However, we have to worry about "else if" that spans multiple lines!
+  if Search(r'}\s*else[^{]*$', line) or Match(r'[^}]*else\s*{', line):
+    if Search(r'}\s*else if([^{]*)$', line):       # could be multi-line if
+      # find the ( after the if
+      pos = line.find('else if')
+      pos = line.find('(', pos)
+      if pos > 0:
+        (endline, _, endpos) = CloseExpression(clean_lines, linenum, pos)
+        if endline[endpos:].find('{') == -1:    # must be brace after if
+          error(filename, linenum, 'readability/braces', 5,
+                'If an else has a brace on one side, it should have it on both')
+    else:            # common case: else not followed by a multi-line if
+      error(filename, linenum, 'readability/braces', 5,
+            'If an else has a brace on one side, it should have it on both')
+
+  # Likewise, an else should never have the else clause on the same line
+  if Search(r'\belse [^\s{]', line) and not Search(r'\belse if\b', line):
+    error(filename, linenum, 'whitespace/newline', 4,
+          'Else clause should never be on same line as else (use 2 lines)')
+
+  # In the same way, a do/while should never be on one line
+  if Match(r'\s*do [^\s{]', line):
+    error(filename, linenum, 'whitespace/newline', 4,
+          'do/while clauses should not be on a single line')
+
+  # Block bodies should not be followed by a semicolon.  Due to C++11
+  # brace initialization, there are more places where semicolons are
+  # required than not, so we use a whitelist approach to check these
+  # rather than a blacklist.  These are the places where "};" should
+  # be replaced by just "}":
+  # 1. Some flavor of block following closing parenthesis:
+  #    for (;;) {};
+  #    while (...) {};
+  #    switch (...) {};
+  #    Function(...) {};
+  #    if (...) {};
+  #    if (...) else if (...) {};
+  #
+  # 2. else block:
+  #    if (...) else {};
+  #
+  # 3. const member function:
+  #    Function(...) const {};
+  #
+  # 4. Block following some statement:
+  #    x = 42;
+  #    {};
+  #
+  # 5. Block at the beginning of a function:
+  #    Function(...) {
+  #      {};
+  #    }
+  #
+  #    Note that naively checking for the preceding "{" will also match
+  #    braces inside multi-dimensional arrays, but this is fine since
+  #    that expression will not contain semicolons.
+  #
+  # 6. Block following another block:
+  #    while (true) {}
+  #    {};
+  #
+  # 7. End of namespaces:
+  #    namespace {};
+  #
+  #    These semicolons seems far more common than other kinds of
+  #    redundant semicolons, possibly due to people converting classes
+  #    to namespaces.  For now we do not warn for this case.
+  #
+  # Try matching case 1 first.
+  match = Match(r'^(.*\)\s*)\{', line)
+  if match:
+    # Matched closing parenthesis (case 1).  Check the token before the
+    # matching opening parenthesis, and don't warn if it looks like a
+    # macro.  This avoids these false positives:
+    #  - macro that defines a base class
+    #  - multi-line macro that defines a base class
+    #  - macro that defines the whole class-head
+    #
+    # But we still issue warnings for macros that we know are safe to
+    # warn, specifically:
+    #  - TEST, TEST_F, TEST_P, MATCHER, MATCHER_P
+    #  - TYPED_TEST
+    #  - INTERFACE_DEF
+    #  - EXCLUSIVE_LOCKS_REQUIRED, SHARED_LOCKS_REQUIRED, LOCKS_EXCLUDED:
+    #
+    # We implement a whitelist of safe macros instead of a blacklist of
+    # unsafe macros, even though the latter appears less frequently in
+    # google code and would have been easier to implement.  This is because
+    # the downside for getting the whitelist wrong means some extra
+    # semicolons, while the downside for getting the blacklist wrong
+    # would result in compile errors.
+    #
+    # In addition to macros, we also don't want to warn on compound
+    # literals.
+    closing_brace_pos = match.group(1).rfind(')')
+    opening_parenthesis = ReverseCloseExpression(
+        clean_lines, linenum, closing_brace_pos)
+    if opening_parenthesis[2] > -1:
+      line_prefix = opening_parenthesis[0][0:opening_parenthesis[2]]
+      macro = Search(r'\b([A-Z_]+)\s*$', line_prefix)
+      if ((macro and
+           macro.group(1) not in (
+               'TEST', 'TEST_F', 'MATCHER', 'MATCHER_P', 'TYPED_TEST',
+               'EXCLUSIVE_LOCKS_REQUIRED', 'SHARED_LOCKS_REQUIRED',
+               'LOCKS_EXCLUDED', 'INTERFACE_DEF')) or
+          Search(r'\s+=\s*$', line_prefix)):
+        match = None
+
+  else:
+    # Try matching cases 2-3.
+    match = Match(r'^(.*(?:else|\)\s*const)\s*)\{', line)
+    if not match:
+      # Try matching cases 4-6.  These are always matched on separate lines.
+      #
+      # Note that we can't simply concatenate the previous line to the
+      # current line and do a single match, otherwise we may output
+      # duplicate warnings for the blank line case:
+      #   if (cond) {
+      #     // blank line
+      #   }
+      prevline = GetPreviousNonBlankLine(clean_lines, linenum)[0]
+      if prevline and Search(r'[;{}]\s*$', prevline):
+        match = Match(r'^(\s*)\{', line)
+
+  # Check matching closing brace
+  if match:
+    (endline, endlinenum, endpos) = CloseExpression(
+        clean_lines, linenum, len(match.group(1)))
+    if endpos > -1 and Match(r'^\s*;', endline[endpos:]):
+      # Current {} pair is eligible for semicolon check, and we have found
+      # the redundant semicolon, output warning here.
+      #
+      # Note: because we are scanning forward for opening braces, and
+      # outputting warnings for the matching closing brace, if there are
+      # nested blocks with trailing semicolons, we will get the error
+      # messages in reversed order.
+      error(filename, endlinenum, 'readability/braces', 4,
+            "You don't need a ; after a }")
+
+
+def CheckEmptyBlockBody(filename, clean_lines, linenum, error):
+  """Look for empty loop/conditional body with only a single semicolon.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+
+  # Search for loop keywords at the beginning of the line.  Because only
+  # whitespaces are allowed before the keywords, this will also ignore most
+  # do-while-loops, since those lines should start with closing brace.
+  #
+  # We also check "if" blocks here, since an empty conditional block
+  # is likely an error.
+  line = clean_lines.elided[linenum]
+  matched = Match(r'\s*(for|while|if)\s*\(', line)
+  if matched:
+    # Find the end of the conditional expression
+    (end_line, end_linenum, end_pos) = CloseExpression(
+        clean_lines, linenum, line.find('('))
+
+    # Output warning if what follows the condition expression is a semicolon.
+    # No warning for all other cases, including whitespace or newline, since we
+    # have a separate check for semicolons preceded by whitespace.
+    if end_pos >= 0 and Match(r';', end_line[end_pos:]):
+      if matched.group(1) == 'if':
+        error(filename, end_linenum, 'whitespace/empty_conditional_body', 5,
+              'Empty conditional bodies should use {}')
+      else:
+        error(filename, end_linenum, 'whitespace/empty_loop_body', 5,
+              'Empty loop bodies should use {} or continue')
+
+
+def CheckCheck(filename, clean_lines, linenum, error):
+  """Checks the use of CHECK and EXPECT macros.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+
+  # Decide the set of replacement macros that should be suggested
+  lines = clean_lines.elided
+  check_macro = None
+  start_pos = -1
+  for macro in _CHECK_MACROS:
+    i = lines[linenum].find(macro)
+    if i >= 0:
+      check_macro = macro
+
+      # Find opening parenthesis.  Do a regular expression match here
+      # to make sure that we are matching the expected CHECK macro, as
+      # opposed to some other macro that happens to contain the CHECK
+      # substring.
+      matched = Match(r'^(.*\b' + check_macro + r'\s*)\(', lines[linenum])
+      if not matched:
+        continue
+      start_pos = len(matched.group(1))
+      break
+  if not check_macro or start_pos < 0:
+    # Don't waste time here if line doesn't contain 'CHECK' or 'EXPECT'
+    return
+
+  # Find end of the boolean expression by matching parentheses
+  (last_line, end_line, end_pos) = CloseExpression(
+      clean_lines, linenum, start_pos)
+  if end_pos < 0:
+    return
+  if linenum == end_line:
+    expression = lines[linenum][start_pos + 1:end_pos - 1]
+  else:
+    expression = lines[linenum][start_pos + 1:]
+    for i in xrange(linenum + 1, end_line):
+      expression += lines[i]
+    expression += last_line[0:end_pos - 1]
+
+  # Parse expression so that we can take parentheses into account.
+  # This avoids false positives for inputs like "CHECK((a < 4) == b)",
+  # which is not replaceable by CHECK_LE.
+  lhs = ''
+  rhs = ''
+  operator = None
+  while expression:
+    matched = Match(r'^\s*(<<|<<=|>>|>>=|->\*|->|&&|\|\||'
+                    r'==|!=|>=|>|<=|<|\()(.*)$', expression)
+    if matched:
+      token = matched.group(1)
+      if token == '(':
+        # Parenthesized operand
+        expression = matched.group(2)
+        (end, _) = FindEndOfExpressionInLine(expression, 0, 1, '(', ')')
+        if end < 0:
+          return  # Unmatched parenthesis
+        lhs += '(' + expression[0:end]
+        expression = expression[end:]
+      elif token in ('&&', '||'):
+        # Logical and/or operators.  This means the expression
+        # contains more than one term, for example:
+        #   CHECK(42 < a && a < b);
+        #
+        # These are not replaceable with CHECK_LE, so bail out early.
+        return
+      elif token in ('<<', '<<=', '>>', '>>=', '->*', '->'):
+        # Non-relational operator
+        lhs += token
+        expression = matched.group(2)
+      else:
+        # Relational operator
+        operator = token
+        rhs = matched.group(2)
+        break
+    else:
+      # Unparenthesized operand.  Instead of appending to lhs one character
+      # at a time, we do another regular expression match to consume several
+      # characters at once if possible.  Trivial benchmark shows that this
+      # is more efficient when the operands are longer than a single
+      # character, which is generally the case.
+      matched = Match(r'^([^-=!<>()&|]+)(.*)$', expression)
+      if not matched:
+        matched = Match(r'^(\s*\S)(.*)$', expression)
+        if not matched:
+          break
+      lhs += matched.group(1)
+      expression = matched.group(2)
+
+  # Only apply checks if we got all parts of the boolean expression
+  if not (lhs and operator and rhs):
+    return
+
+  # Check that rhs do not contain logical operators.  We already know
+  # that lhs is fine since the loop above parses out && and ||.
+  if rhs.find('&&') > -1 or rhs.find('||') > -1:
+    return
+
+  # At least one of the operands must be a constant literal.  This is
+  # to avoid suggesting replacements for unprintable things like
+  # CHECK(variable != iterator)
+  #
+  # The following pattern matches decimal, hex integers, strings, and
+  # characters (in that order).
+  lhs = lhs.strip()
+  rhs = rhs.strip()
+  match_constant = r'^([-+]?(\d+|0[xX][0-9a-fA-F]+)[lLuU]{0,3}|".*"|\'.*\')$'
+  if Match(match_constant, lhs) or Match(match_constant, rhs):
+    # Note: since we know both lhs and rhs, we can provide a more
+    # descriptive error message like:
+    #   Consider using CHECK_EQ(x, 42) instead of CHECK(x == 42)
+    # Instead of:
+    #   Consider using CHECK_EQ instead of CHECK(a == b)
+    #
+    # We are still keeping the less descriptive message because if lhs
+    # or rhs gets long, the error message might become unreadable.
+    error(filename, linenum, 'readability/check', 2,
+          'Consider using %s instead of %s(a %s b)' % (
+              _CHECK_REPLACEMENT[check_macro][operator],
+              check_macro, operator))
+
+
+def CheckAltTokens(filename, clean_lines, linenum, error):
+  """Check alternative keywords being used in boolean expressions.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+
+  # Avoid preprocessor lines
+  if Match(r'^\s*#', line):
+    return
+
+  # Last ditch effort to avoid multi-line comments.  This will not help
+  # if the comment started before the current line or ended after the
+  # current line, but it catches most of the false positives.  At least,
+  # it provides a way to workaround this warning for people who use
+  # multi-line comments in preprocessor macros.
+  #
+  # TODO(unknown): remove this once cpplint has better support for
+  # multi-line comments.
+  if line.find('/*') >= 0 or line.find('*/') >= 0:
+    return
+
+  for match in _ALT_TOKEN_REPLACEMENT_PATTERN.finditer(line):
+    error(filename, linenum, 'readability/alt_tokens', 2,
+          'Use operator %s instead of %s' % (
+              _ALT_TOKEN_REPLACEMENT[match.group(1)], match.group(1)))
+
+
+def GetLineWidth(line):
+  """Determines the width of the line in column positions.
+
+  Args:
+    line: A string, which may be a Unicode string.
+
+  Returns:
+    The width of the line in column positions, accounting for Unicode
+    combining characters and wide characters.
+  """
+  if isinstance(line, unicode):
+    width = 0
+    for uc in unicodedata.normalize('NFC', line):
+      if unicodedata.east_asian_width(uc) in ('W', 'F'):
+        width += 2
+      elif not unicodedata.combining(uc):
+        width += 1
+    return width
+  else:
+    return len(line)
+
+
+def CheckStyle(filename, clean_lines, linenum, file_extension, nesting_state,
+               error):
+  """Checks rules from the 'C++ style rules' section of cppguide.html.
+
+  Most of these rules are hard to test (naming, comment style), but we
+  do what we can.  In particular we check for 2-space indents, line lengths,
+  tab usage, spaces inside code, etc.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    file_extension: The extension (without the dot) of the filename.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: The function to call with any errors found.
+  """
+
+  # Don't use "elided" lines here, otherwise we can't check commented lines.
+  # Don't want to use "raw" either, because we don't want to check inside C++11
+  # raw strings,
+  raw_lines = clean_lines.lines_without_raw_strings
+  line = raw_lines[linenum]
+
+  if line.find('\t') != -1:
+    error(filename, linenum, 'whitespace/tab', 1,
+          'Tab found; better to use spaces')
+
+  # One or three blank spaces at the beginning of the line is weird; it's
+  # hard to reconcile that with 2-space indents.
+  # NOTE: here are the conditions rob pike used for his tests.  Mine aren't
+  # as sophisticated, but it may be worth becoming so:  RLENGTH==initial_spaces
+  # if(RLENGTH > 20) complain = 0;
+  # if(match($0, " +(error|private|public|protected):")) complain = 0;
+  # if(match(prev, "&& *$")) complain = 0;
+  # if(match(prev, "\\|\\| *$")) complain = 0;
+  # if(match(prev, "[\",=><] *$")) complain = 0;
+  # if(match($0, " <<")) complain = 0;
+  # if(match(prev, " +for \\(")) complain = 0;
+  # if(prevodd && match(prevprev, " +for \\(")) complain = 0;
+  initial_spaces = 0
+  cleansed_line = clean_lines.elided[linenum]
+  while initial_spaces < len(line) and line[initial_spaces] == ' ':
+    initial_spaces += 1
+  if line and line[-1].isspace():
+    error(filename, linenum, 'whitespace/end_of_line', 4,
+          'Line ends in whitespace.  Consider deleting these extra spaces.')
+  # There are certain situations we allow one space, notably for section labels
+  elif ((initial_spaces == 1 or initial_spaces == 3) and
+        not Match(r'\s*\w+\s*:\s*$', cleansed_line)):
+    error(filename, linenum, 'whitespace/indent', 3,
+          'Weird number of spaces at line-start.  '
+          'Are you using a 2-space indent?')
+
+  # Check if the line is a header guard.
+  is_header_guard = False
+  if file_extension == 'h':
+    cppvar = GetHeaderGuardCPPVariable(filename)
+    if (line.startswith('#ifndef %s' % cppvar) or
+        line.startswith('#define %s' % cppvar) or
+        line.startswith('#endif  // %s' % cppvar)):
+      is_header_guard = True
+  # #include lines and header guards can be long, since there's no clean way to
+  # split them.
+  #
+  # URLs can be long too.  It's possible to split these, but it makes them
+  # harder to cut&paste.
+  #
+  # The "$Id:...$" comment may also get very long without it being the
+  # developers fault.
+  if (not line.startswith('#include') and not is_header_guard and
+      not Match(r'^\s*//.*http(s?)://\S*$', line) and
+      not Match(r'^// \$Id:.*#[0-9]+ \$$', line)):
+    line_width = GetLineWidth(line)
+    extended_length = int((_line_length * 1.25))
+    if line_width > extended_length:
+      error(filename, linenum, 'whitespace/line_length', 4,
+            'Lines should very rarely be longer than %i characters' %
+            extended_length)
+    elif line_width > _line_length:
+      error(filename, linenum, 'whitespace/line_length', 2,
+            'Lines should be <= %i characters long' % _line_length)
+
+  if (cleansed_line.count(';') > 1 and
+      # for loops are allowed two ;'s (and may run over two lines).
+      cleansed_line.find('for') == -1 and
+      (GetPreviousNonBlankLine(clean_lines, linenum)[0].find('for') == -1 or
+       GetPreviousNonBlankLine(clean_lines, linenum)[0].find(';') != -1) and
+      # It's ok to have many commands in a switch case that fits in 1 line
+      not ((cleansed_line.find('case ') != -1 or
+            cleansed_line.find('default:') != -1) and
+           cleansed_line.find('break;') != -1)):
+    error(filename, linenum, 'whitespace/newline', 0,
+          'More than one command on the same line')
+
+  # Some more style checks
+  CheckBraces(filename, clean_lines, linenum, error)
+  CheckEmptyBlockBody(filename, clean_lines, linenum, error)
+  CheckAccess(filename, clean_lines, linenum, nesting_state, error)
+  CheckSpacing(filename, clean_lines, linenum, nesting_state, error)
+  CheckCheck(filename, clean_lines, linenum, error)
+  CheckAltTokens(filename, clean_lines, linenum, error)
+  classinfo = nesting_state.InnermostClass()
+  if classinfo:
+    CheckSectionSpacing(filename, clean_lines, classinfo, linenum, error)
+
+
+_RE_PATTERN_INCLUDE_NEW_STYLE = re.compile(r'#include +"[^/]+\.h"')
+_RE_PATTERN_INCLUDE = re.compile(r'^\s*#\s*include\s*([<"])([^>"]*)[>"].*$')
+# Matches the first component of a filename delimited by -s and _s. That is:
+#  _RE_FIRST_COMPONENT.match('foo').group(0) == 'foo'
+#  _RE_FIRST_COMPONENT.match('foo.cc').group(0) == 'foo'
+#  _RE_FIRST_COMPONENT.match('foo-bar_baz.cc').group(0) == 'foo'
+#  _RE_FIRST_COMPONENT.match('foo_bar-baz.cc').group(0) == 'foo'
+_RE_FIRST_COMPONENT = re.compile(r'^[^-_.]+')
+
+
+def _DropCommonSuffixes(filename):
+  """Drops common suffixes like _test.cc or -inl.h from filename.
+
+  For example:
+    >>> _DropCommonSuffixes('foo/foo-inl.h')
+    'foo/foo'
+    >>> _DropCommonSuffixes('foo/bar/foo.cc')
+    'foo/bar/foo'
+    >>> _DropCommonSuffixes('foo/foo_internal.h')
+    'foo/foo'
+    >>> _DropCommonSuffixes('foo/foo_unusualinternal.h')
+    'foo/foo_unusualinternal'
+
+  Args:
+    filename: The input filename.
+
+  Returns:
+    The filename with the common suffix removed.
+  """
+  for suffix in ('test.cc', 'regtest.cc', 'unittest.cc',
+                 'inl.h', 'impl.h', 'internal.h'):
+    if (filename.endswith(suffix) and len(filename) > len(suffix) and
+        filename[-len(suffix) - 1] in ('-', '_')):
+      return filename[:-len(suffix) - 1]
+  return os.path.splitext(filename)[0]
+
+
+def _IsTestFilename(filename):
+  """Determines if the given filename has a suffix that identifies it as a test.
+
+  Args:
+    filename: The input filename.
+
+  Returns:
+    True if 'filename' looks like a test, False otherwise.
+  """
+  if (filename.endswith('_test.cc') or
+      filename.endswith('_unittest.cc') or
+      filename.endswith('_regtest.cc')):
+    return True
+  else:
+    return False
+
+
+def _ClassifyInclude(fileinfo, include, is_system):
+  """Figures out what kind of header 'include' is.
+
+  Args:
+    fileinfo: The current file cpplint is running over. A FileInfo instance.
+    include: The path to a #included file.
+    is_system: True if the #include used <> rather than "".
+
+  Returns:
+    One of the _XXX_HEADER constants.
+
+  For example:
+    >>> _ClassifyInclude(FileInfo('foo/foo.cc'), 'stdio.h', True)
+    _C_SYS_HEADER
+    >>> _ClassifyInclude(FileInfo('foo/foo.cc'), 'string', True)
+    _CPP_SYS_HEADER
+    >>> _ClassifyInclude(FileInfo('foo/foo.cc'), 'foo/foo.h', False)
+    _LIKELY_MY_HEADER
+    >>> _ClassifyInclude(FileInfo('foo/foo_unknown_extension.cc'),
+    ...                  'bar/foo_other_ext.h', False)
+    _POSSIBLE_MY_HEADER
+    >>> _ClassifyInclude(FileInfo('foo/foo.cc'), 'foo/bar.h', False)
+    _OTHER_HEADER
+  """
+  # This is a list of all standard c++ header files, except
+  # those already checked for above.
+  is_cpp_h = include in _CPP_HEADERS
+
+  if is_system:
+    if is_cpp_h:
+      return _CPP_SYS_HEADER
+    else:
+      return _C_SYS_HEADER
+
+  # If the target file and the include we're checking share a
+  # basename when we drop common extensions, and the include
+  # lives in . , then it's likely to be owned by the target file.
+  target_dir, target_base = (
+      os.path.split(_DropCommonSuffixes(fileinfo.RepositoryName())))
+  include_dir, include_base = os.path.split(_DropCommonSuffixes(include))
+  if target_base == include_base and (
+      include_dir == target_dir or
+      include_dir == os.path.normpath(target_dir + '/../public')):
+    return _LIKELY_MY_HEADER
+
+  # If the target and include share some initial basename
+  # component, it's possible the target is implementing the
+  # include, so it's allowed to be first, but we'll never
+  # complain if it's not there.
+  target_first_component = _RE_FIRST_COMPONENT.match(target_base)
+  include_first_component = _RE_FIRST_COMPONENT.match(include_base)
+  if (target_first_component and include_first_component and
+      target_first_component.group(0) ==
+      include_first_component.group(0)):
+    return _POSSIBLE_MY_HEADER
+
+  return _OTHER_HEADER
+
+
+
+def CheckIncludeLine(filename, clean_lines, linenum, include_state, error):
+  """Check rules that are applicable to #include lines.
+
+  Strings on #include lines are NOT removed from elided line, to make
+  certain tasks easier. However, to prevent false positives, checks
+  applicable to #include lines in CheckLanguage must be put here.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    include_state: An _IncludeState instance in which the headers are inserted.
+    error: The function to call with any errors found.
+  """
+  fileinfo = FileInfo(filename)
+
+  line = clean_lines.lines[linenum]
+
+  # "include" should use the new style "foo/bar.h" instead of just "bar.h"
+  if _RE_PATTERN_INCLUDE_NEW_STYLE.search(line):
+    error(filename, linenum, 'build/include', 4,
+          'Include the directory when naming .h files')
+
+  # we shouldn't include a file more than once. actually, there are a
+  # handful of instances where doing so is okay, but in general it's
+  # not.
+  match = _RE_PATTERN_INCLUDE.search(line)
+  if match:
+    include = match.group(2)
+    is_system = (match.group(1) == '<')
+    if include in include_state:
+      error(filename, linenum, 'build/include', 4,
+            '"%s" already included at %s:%s' %
+            (include, filename, include_state[include]))
+    else:
+      include_state[include] = linenum
+
+      # We want to ensure that headers appear in the right order:
+      # 1) for foo.cc, foo.h  (preferred location)
+      # 2) c system files
+      # 3) cpp system files
+      # 4) for foo.cc, foo.h  (deprecated location)
+      # 5) other google headers
+      #
+      # We classify each include statement as one of those 5 types
+      # using a number of techniques. The include_state object keeps
+      # track of the highest type seen, and complains if we see a
+      # lower type after that.
+      error_message = include_state.CheckNextIncludeOrder(
+          _ClassifyInclude(fileinfo, include, is_system))
+      if error_message:
+        error(filename, linenum, 'build/include_order', 4,
+              '%s. Should be: %s.h, c system, c++ system, other.' %
+              (error_message, fileinfo.BaseName()))
+      canonical_include = include_state.CanonicalizeAlphabeticalOrder(include)
+      if not include_state.IsInAlphabeticalOrder(
+          clean_lines, linenum, canonical_include):
+        error(filename, linenum, 'build/include_alpha', 4,
+              'Include "%s" not in alphabetical order' % include)
+      include_state.SetLastHeader(canonical_include)
+
+  # Look for any of the stream classes that are part of standard C++.
+  match = _RE_PATTERN_INCLUDE.match(line)
+  if match:
+    include = match.group(2)
+    if Match(r'(f|ind|io|i|o|parse|pf|stdio|str|)?stream$', include):
+      # Many unit tests use cout, so we exempt them.
+      if not _IsTestFilename(filename):
+        error(filename, linenum, 'readability/streams', 3,
+              'Streams are highly discouraged.')
+
+
+def _GetTextInside(text, start_pattern):
+  r"""Retrieves all the text between matching open and close parentheses.
+
+  Given a string of lines and a regular expression string, retrieve all the text
+  following the expression and between opening punctuation symbols like
+  (, [, or {, and the matching close-punctuation symbol. This properly nested
+  occurrences of the punctuations, so for the text like
+    printf(a(), b(c()));
+  a call to _GetTextInside(text, r'printf\(') will return 'a(), b(c())'.
+  start_pattern must match string having an open punctuation symbol at the end.
+
+  Args:
+    text: The lines to extract text. Its comments and strings must be elided.
+           It can be single line and can span multiple lines.
+    start_pattern: The regexp string indicating where to start extracting
+                   the text.
+  Returns:
+    The extracted text.
+    None if either the opening string or ending punctuation could not be found.
+  """
+  # TODO(sugawarayu): Audit cpplint.py to see what places could be profitably
+  # rewritten to use _GetTextInside (and use inferior regexp matching today).
+
+  # Give opening punctuations to get the matching close-punctuations.
+  matching_punctuation = {'(': ')', '{': '}', '[': ']'}
+  closing_punctuation = set(matching_punctuation.itervalues())
+
+  # Find the position to start extracting text.
+  match = re.search(start_pattern, text, re.M)
+  if not match:  # start_pattern not found in text.
+    return None
+  start_position = match.end(0)
+
+  assert start_position > 0, (
+      'start_pattern must ends with an opening punctuation.')
+  assert text[start_position - 1] in matching_punctuation, (
+      'start_pattern must ends with an opening punctuation.')
+  # Stack of closing punctuations we expect to have in text after position.
+  punctuation_stack = [matching_punctuation[text[start_position - 1]]]
+  position = start_position
+  while punctuation_stack and position < len(text):
+    if text[position] == punctuation_stack[-1]:
+      punctuation_stack.pop()
+    elif text[position] in closing_punctuation:
+      # A closing punctuation without matching opening punctuations.
+      return None
+    elif text[position] in matching_punctuation:
+      punctuation_stack.append(matching_punctuation[text[position]])
+    position += 1
+  if punctuation_stack:
+    # Opening punctuations left without matching close-punctuations.
+    return None
+  # punctuations match.
+  return text[start_position:position - 1]
+
+
+# Patterns for matching call-by-reference parameters.
+#
+# Supports nested templates up to 2 levels deep using this messy pattern:
+#   < (?: < (?: < [^<>]*
+#               >
+#           |   [^<>] )*
+#         >
+#     |   [^<>] )*
+#   >
+_RE_PATTERN_IDENT = r'[_a-zA-Z]\w*'  # =~ [[:alpha:]][[:alnum:]]*
+_RE_PATTERN_TYPE = (
+    r'(?:const\s+)?(?:typename\s+|class\s+|struct\s+|union\s+|enum\s+)?'
+    r'(?:\w|'
+    r'\s*<(?:<(?:<[^<>]*>|[^<>])*>|[^<>])*>|'
+    r'::)+')
+# A call-by-reference parameter ends with '& identifier'.
+_RE_PATTERN_REF_PARAM = re.compile(
+    r'(' + _RE_PATTERN_TYPE + r'(?:\s*(?:\bconst\b|[*]))*\s*'
+    r'&\s*' + _RE_PATTERN_IDENT + r')\s*(?:=[^,()]+)?[,)]')
+# A call-by-const-reference parameter either ends with 'const& identifier'
+# or looks like 'const type& identifier' when 'type' is atomic.
+_RE_PATTERN_CONST_REF_PARAM = (
+    r'(?:.*\s*\bconst\s*&\s*' + _RE_PATTERN_IDENT +
+    r'|const\s+' + _RE_PATTERN_TYPE + r'\s*&\s*' + _RE_PATTERN_IDENT + r')')
+
+
+def CheckLanguage(filename, clean_lines, linenum, file_extension,
+                  include_state, nesting_state, error):
+  """Checks rules from the 'C++ language rules' section of cppguide.html.
+
+  Some of these rules are hard to test (function overloading, using
+  uint32 inappropriately), but we do the best we can.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    file_extension: The extension (without the dot) of the filename.
+    include_state: An _IncludeState instance in which the headers are inserted.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: The function to call with any errors found.
+  """
+  # If the line is empty or consists of entirely a comment, no need to
+  # check it.
+  line = clean_lines.elided[linenum]
+  if not line:
+    return
+
+  match = _RE_PATTERN_INCLUDE.search(line)
+  if match:
+    CheckIncludeLine(filename, clean_lines, linenum, include_state, error)
+    return
+
+  # Reset include state across preprocessor directives.  This is meant
+  # to silence warnings for conditional includes.
+  if Match(r'^\s*#\s*(?:ifdef|elif|else|endif)\b', line):
+    include_state.ResetSection()
+
+  # Make Windows paths like Unix.
+  fullname = os.path.abspath(filename).replace('\\', '/')
+
+  # TODO(unknown): figure out if they're using default arguments in fn proto.
+
+  # Check to see if they're using an conversion function cast.
+  # I just try to capture the most common basic types, though there are more.
+  # Parameterless conversion functions, such as bool(), are allowed as they are
+  # probably a member operator declaration or default constructor.
+  match = Search(
+      r'(\bnew\s+)?\b'  # Grab 'new' operator, if it's there
+      r'(int|float|double|bool|char|int32|uint32|int64|uint64)'
+      r'(\([^)].*)', line)
+  if match:
+    matched_new = match.group(1)
+    matched_type = match.group(2)
+    matched_funcptr = match.group(3)
+
+    # gMock methods are defined using some variant of MOCK_METHODx(name, type)
+    # where type may be float(), int(string), etc.  Without context they are
+    # virtually indistinguishable from int(x) casts. Likewise, gMock's
+    # MockCallback takes a template parameter of the form return_type(arg_type),
+    # which looks much like the cast we're trying to detect.
+    #
+    # std::function<> wrapper has a similar problem.
+    #
+    # Return types for function pointers also look like casts if they
+    # don't have an extra space.
+    if (matched_new is None and  # If new operator, then this isn't a cast
+        not (Match(r'^\s*MOCK_(CONST_)?METHOD\d+(_T)?\(', line) or
+             Search(r'\bMockCallback<.*>', line) or
+             Search(r'\bstd::function<.*>', line)) and
+        not (matched_funcptr and
+             Match(r'\((?:[^() ]+::\s*\*\s*)?[^() ]+\)\s*\(',
+                   matched_funcptr))):
+      # Try a bit harder to catch gmock lines: the only place where
+      # something looks like an old-style cast is where we declare the
+      # return type of the mocked method, and the only time when we
+      # are missing context is if MOCK_METHOD was split across
+      # multiple lines.  The missing MOCK_METHOD is usually one or two
+      # lines back, so scan back one or two lines.
+      #
+      # It's not possible for gmock macros to appear in the first 2
+      # lines, since the class head + section name takes up 2 lines.
+      if (linenum < 2 or
+          not (Match(r'^\s*MOCK_(?:CONST_)?METHOD\d+(?:_T)?\((?:\S+,)?\s*$',
+                     clean_lines.elided[linenum - 1]) or
+               Match(r'^\s*MOCK_(?:CONST_)?METHOD\d+(?:_T)?\(\s*$',
+                     clean_lines.elided[linenum - 2]))):
+        error(filename, linenum, 'readability/casting', 4,
+              'Using deprecated casting style.  '
+              'Use static_cast<%s>(...) instead' %
+              matched_type)
+
+  CheckCStyleCast(filename, linenum, line, clean_lines.raw_lines[linenum],
+                  'static_cast',
+                  r'\((int|float|double|bool|char|u?int(16|32|64))\)', error)
+
+  # This doesn't catch all cases. Consider (const char * const)"hello".
+  #
+  # (char *) "foo" should always be a const_cast (reinterpret_cast won't
+  # compile).
+  if CheckCStyleCast(filename, linenum, line, clean_lines.raw_lines[linenum],
+                     'const_cast', r'\((char\s?\*+\s?)\)\s*"', error):
+    pass
+  else:
+    # Check pointer casts for other than string constants
+    CheckCStyleCast(filename, linenum, line, clean_lines.raw_lines[linenum],
+                    'reinterpret_cast', r'\((\w+\s?\*+\s?)\)', error)
+
+  # In addition, we look for people taking the address of a cast.  This
+  # is dangerous -- casts can assign to temporaries, so the pointer doesn't
+  # point where you think.
+  match = Search(
+      r'(?:&\(([^)]+)\)[\w(])|'
+      r'(?:&(static|dynamic|down|reinterpret)_cast\b)', line)
+  if match and match.group(1) != '*':
+    error(filename, linenum, 'runtime/casting', 4,
+          ('Are you taking an address of a cast?  '
+           'This is dangerous: could be a temp var.  '
+           'Take the address before doing the cast, rather than after'))
+
+  # Create an extended_line, which is the concatenation of the current and
+  # next lines, for more effective checking of code that may span more than one
+  # line.
+  if linenum + 1 < clean_lines.NumLines():
+    extended_line = line + clean_lines.elided[linenum + 1]
+  else:
+    extended_line = line
+
+  # Check for people declaring static/global STL strings at the top level.
+  # This is dangerous because the C++ language does not guarantee that
+  # globals with constructors are initialized before the first access.
+  match = Match(
+      r'((?:|static +)(?:|const +))string +([a-zA-Z0-9_:]+)\b(.*)',
+      line)
+  # Make sure it's not a function.
+  # Function template specialization looks like: "string foo<Type>(...".
+  # Class template definitions look like: "string Foo<Type>::Method(...".
+  #
+  # Also ignore things that look like operators.  These are matched separately
+  # because operator names cross non-word boundaries.  If we change the pattern
+  # above, we would decrease the accuracy of matching identifiers.
+  if (match and
+      not Search(r'\boperator\W', line) and
+      not Match(r'\s*(<.*>)?(::[a-zA-Z0-9_]+)?\s*\(([^"]|$)', match.group(3))):
+    error(filename, linenum, 'runtime/string', 4,
+          'For a static/global string constant, use a C style string instead: '
+          '"%schar %s[]".' %
+          (match.group(1), match.group(2)))
+
+  if Search(r'\b([A-Za-z0-9_]*_)\(\1\)', line):
+    error(filename, linenum, 'runtime/init', 4,
+          'You seem to be initializing a member variable with itself.')
+
+  if file_extension == 'h':
+    # TODO(unknown): check that 1-arg constructors are explicit.
+    #                How to tell it's a constructor?
+    #                (handled in CheckForNonStandardConstructs for now)
+    # TODO(unknown): check that classes have DISALLOW_EVIL_CONSTRUCTORS
+    #                (level 1 error)
+    pass
+
+  # Check if people are using the verboten C basic types.  The only exception
+  # we regularly allow is "unsigned short port" for port.
+  if Search(r'\bshort port\b', line):
+    if not Search(r'\bunsigned short port\b', line):
+      error(filename, linenum, 'runtime/int', 4,
+            'Use "unsigned short" for ports, not "short"')
+  else:
+    match = Search(r'\b(short|long(?! +double)|long long)\b', line)
+    if match:
+      error(filename, linenum, 'runtime/int', 4,
+            'Use int16/int64/etc, rather than the C type %s' % match.group(1))
+
+  # When snprintf is used, the second argument shouldn't be a literal.
+  match = Search(r'snprintf\s*\(([^,]*),\s*([0-9]*)\s*,', line)
+  if match and match.group(2) != '0':
+    # If 2nd arg is zero, snprintf is used to calculate size.
+    error(filename, linenum, 'runtime/printf', 3,
+          'If you can, use sizeof(%s) instead of %s as the 2nd arg '
+          'to snprintf.' % (match.group(1), match.group(2)))
+
+  # Check if some verboten C functions are being used.
+  if Search(r'\bsprintf\b', line):
+    error(filename, linenum, 'runtime/printf', 5,
+          'Never use sprintf.  Use snprintf instead.')
+  match = Search(r'\b(strcpy|strcat)\b', line)
+  if match:
+    error(filename, linenum, 'runtime/printf', 4,
+          'Almost always, snprintf is better than %s' % match.group(1))
+
+  # Check if some verboten operator overloading is going on
+  # TODO(unknown): catch out-of-line unary operator&:
+  #   class X {};
+  #   int operator&(const X& x) { return 42; }  // unary operator&
+  # The trick is it's hard to tell apart from binary operator&:
+  #   class Y { int operator&(const Y& x) { return 23; } }; // binary operator&
+  if Search(r'\boperator\s*&\s*\(\s*\)', line):
+    error(filename, linenum, 'runtime/operator', 4,
+          'Unary operator& is dangerous.  Do not use it.')
+
+  # Check for suspicious usage of "if" like
+  # } if (a == b) {
+  if Search(r'\}\s*if\s*\(', line):
+    error(filename, linenum, 'readability/braces', 4,
+          'Did you mean "else if"? If not, start a new line for "if".')
+
+  # Check for potential format string bugs like printf(foo).
+  # We constrain the pattern not to pick things like DocidForPrintf(foo).
+  # Not perfect but it can catch printf(foo.c_str()) and printf(foo->c_str())
+  # TODO(sugawarayu): Catch the following case. Need to change the calling
+  # convention of the whole function to process multiple line to handle it.
+  #   printf(
+  #       boy_this_is_a_really_long_variable_that_cannot_fit_on_the_prev_line);
+  printf_args = _GetTextInside(line, r'(?i)\b(string)?printf\s*\(')
+  if printf_args:
+    match = Match(r'([\w.\->()]+)$', printf_args)
+    if match and match.group(1) != '__VA_ARGS__':
+      function_name = re.search(r'\b((?:string)?printf)\s*\(',
+                                line, re.I).group(1)
+      error(filename, linenum, 'runtime/printf', 4,
+            'Potential format string bug. Do %s("%%s", %s) instead.'
+            % (function_name, match.group(1)))
+
+  # Check for potential memset bugs like memset(buf, sizeof(buf), 0).
+  match = Search(r'memset\s*\(([^,]*),\s*([^,]*),\s*0\s*\)', line)
+  if match and not Match(r"^''|-?[0-9]+|0x[0-9A-Fa-f]$", match.group(2)):
+    error(filename, linenum, 'runtime/memset', 4,
+          'Did you mean "memset(%s, 0, %s)"?'
+          % (match.group(1), match.group(2)))
+
+  if Search(r'\busing namespace\b', line):
+    error(filename, linenum, 'build/namespaces', 5,
+          'Do not use namespace using-directives.  '
+          'Use using-declarations instead.')
+
+  # Detect variable-length arrays.
+  match = Match(r'\s*(.+::)?(\w+) [a-z]\w*\[(.+)];', line)
+  if (match and match.group(2) != 'return' and match.group(2) != 'delete' and
+      match.group(3).find(']') == -1):
+    # Split the size using space and arithmetic operators as delimiters.
+    # If any of the resulting tokens are not compile time constants then
+    # report the error.
+    tokens = re.split(r'\s|\+|\-|\*|\/|<<|>>]', match.group(3))
+    is_const = True
+    skip_next = False
+    for tok in tokens:
+      if skip_next:
+        skip_next = False
+        continue
+
+      if Search(r'sizeof\(.+\)', tok): continue
+      if Search(r'arraysize\(\w+\)', tok): continue
+
+      tok = tok.lstrip('(')
+      tok = tok.rstrip(')')
+      if not tok: continue
+      if Match(r'\d+', tok): continue
+      if Match(r'0[xX][0-9a-fA-F]+', tok): continue
+      if Match(r'k[A-Z0-9]\w*', tok): continue
+      if Match(r'(.+::)?k[A-Z0-9]\w*', tok): continue
+      if Match(r'(.+::)?[A-Z][A-Z0-9_]*', tok): continue
+      # A catch all for tricky sizeof cases, including 'sizeof expression',
+      # 'sizeof(*type)', 'sizeof(const type)', 'sizeof(struct StructName)'
+      # requires skipping the next token because we split on ' ' and '*'.
+      if tok.startswith('sizeof'):
+        skip_next = True
+        continue
+      is_const = False
+      break
+    if not is_const:
+      error(filename, linenum, 'runtime/arrays', 1,
+            'Do not use variable-length arrays.  Use an appropriately named '
+            "('k' followed by CamelCase) compile-time constant for the size.")
+
+  # If DISALLOW_EVIL_CONSTRUCTORS, DISALLOW_COPY_AND_ASSIGN, or
+  # DISALLOW_IMPLICIT_CONSTRUCTORS is present, then it should be the last thing
+  # in the class declaration.
+  match = Match(
+      (r'\s*'
+       r'(DISALLOW_(EVIL_CONSTRUCTORS|COPY_AND_ASSIGN|IMPLICIT_CONSTRUCTORS))'
+       r'\(.*\);$'),
+      line)
+  if match and linenum + 1 < clean_lines.NumLines():
+    next_line = clean_lines.elided[linenum + 1]
+    # We allow some, but not all, declarations of variables to be present
+    # in the statement that defines the class.  The [\w\*,\s]* fragment of
+    # the regular expression below allows users to declare instances of
+    # the class or pointers to instances, but not less common types such
+    # as function pointers or arrays.  It's a tradeoff between allowing
+    # reasonable code and avoiding trying to parse more C++ using regexps.
+    if not Search(r'^\s*}[\w\*,\s]*;', next_line):
+      error(filename, linenum, 'readability/constructors', 3,
+            match.group(1) + ' should be the last thing in the class')
+
+  # Check for use of unnamed namespaces in header files.  Registration
+  # macros are typically OK, so we allow use of "namespace {" on lines
+  # that end with backslashes.
+  if (file_extension == 'h'
+      and Search(r'\bnamespace\s*{', line)
+      and line[-1] != '\\'):
+    error(filename, linenum, 'build/namespaces', 4,
+          'Do not use unnamed namespaces in header files.  See '
+          'http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml#Namespaces'
+          ' for more information.')
+
+def CheckForNonConstReference(filename, clean_lines, linenum,
+                              nesting_state, error):
+  """Check for non-const references.
+
+  Separate from CheckLanguage since it scans backwards from current
+  line, instead of scanning forward.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: The function to call with any errors found.
+  """
+  # Do nothing if there is no '&' on current line.
+  line = clean_lines.elided[linenum]
+  if '&' not in line:
+    return
+
+  # Long type names may be broken across multiple lines, usually in one
+  # of these forms:
+  #   LongType
+  #       ::LongTypeContinued &identifier
+  #   LongType::
+  #       LongTypeContinued &identifier
+  #   LongType<
+  #       ...>::LongTypeContinued &identifier
+  #
+  # If we detected a type split across two lines, join the previous
+  # line to current line so that we can match const references
+  # accordingly.
+  #
+  # Note that this only scans back one line, since scanning back
+  # arbitrary number of lines would be expensive.  If you have a type
+  # that spans more than 2 lines, please use a typedef.
+  if linenum > 1:
+    previous = None
+    if Match(r'\s*::(?:[\w<>]|::)+\s*&\s*\S', line):
+      # previous_line\n + ::current_line
+      previous = Search(r'\b((?:const\s*)?(?:[\w<>]|::)+[\w<>])\s*$',
+                        clean_lines.elided[linenum - 1])
+    elif Match(r'\s*[a-zA-Z_]([\w<>]|::)+\s*&\s*\S', line):
+      # previous_line::\n + current_line
+      previous = Search(r'\b((?:const\s*)?(?:[\w<>]|::)+::)\s*$',
+                        clean_lines.elided[linenum - 1])
+    if previous:
+      line = previous.group(1) + line.lstrip()
+    else:
+      # Check for templated parameter that is split across multiple lines
+      endpos = line.rfind('>')
+      if endpos > -1:
+        (_, startline, startpos) = ReverseCloseExpression(
+            clean_lines, linenum, endpos)
+        if startpos > -1 and startline < linenum:
+          # Found the matching < on an earlier line, collect all
+          # pieces up to current line.
+          line = ''
+          for i in xrange(startline, linenum + 1):
+            line += clean_lines.elided[i].strip()
+
+  # Check for non-const references in function parameters.  A single '&' may
+  # found in the following places:
+  #   inside expression: binary & for bitwise AND
+  #   inside expression: unary & for taking the address of something
+  #   inside declarators: reference parameter
+  # We will exclude the first two cases by checking that we are not inside a
+  # function body, including one that was just introduced by a trailing '{'.
+  # TODO(unknwon): Doesn't account for preprocessor directives.
+  # TODO(unknown): Doesn't account for 'catch(Exception& e)' [rare].
+  check_params = False
+  if not nesting_state.stack:
+    check_params = True  # top level
+  elif (isinstance(nesting_state.stack[-1], _ClassInfo) or
+        isinstance(nesting_state.stack[-1], _NamespaceInfo)):
+    check_params = True  # within class or namespace
+  elif Match(r'.*{\s*$', line):
+    if (len(nesting_state.stack) == 1 or
+        isinstance(nesting_state.stack[-2], _ClassInfo) or
+        isinstance(nesting_state.stack[-2], _NamespaceInfo)):
+      check_params = True  # just opened global/class/namespace block
+  # We allow non-const references in a few standard places, like functions
+  # called "swap()" or iostream operators like "<<" or ">>".  Do not check
+  # those function parameters.
+  #
+  # We also accept & in static_assert, which looks like a function but
+  # it's actually a declaration expression.
+  whitelisted_functions = (r'(?:[sS]wap(?:<\w:+>)?|'
+                           r'operator\s*[<>][<>]|'
+                           r'static_assert|COMPILE_ASSERT'
+                           r')\s*\(')
+  if Search(whitelisted_functions, line):
+    check_params = False
+  elif not Search(r'\S+\([^)]*$', line):
+    # Don't see a whitelisted function on this line.  Actually we
+    # didn't see any function name on this line, so this is likely a
+    # multi-line parameter list.  Try a bit harder to catch this case.
+    for i in xrange(2):
+      if (linenum > i and
+          Search(whitelisted_functions, clean_lines.elided[linenum - i - 1])):
+        check_params = False
+        break
+
+  if check_params:
+    decls = ReplaceAll(r'{[^}]*}', ' ', line)  # exclude function body
+    for parameter in re.findall(_RE_PATTERN_REF_PARAM, decls):
+      if not Match(_RE_PATTERN_CONST_REF_PARAM, parameter):
+        error(filename, linenum, 'runtime/references', 2,
+              'Is this a non-const reference? '
+              'If so, make const or use a pointer: ' +
+              ReplaceAll(' *<', '<', parameter))
+
+
+def CheckCStyleCast(filename, linenum, line, raw_line, cast_type, pattern,
+                    error):
+  """Checks for a C-style cast by looking for the pattern.
+
+  Args:
+    filename: The name of the current file.
+    linenum: The number of the line to check.
+    line: The line of code to check.
+    raw_line: The raw line of code to check, with comments.
+    cast_type: The string for the C++ cast to recommend.  This is either
+      reinterpret_cast, static_cast, or const_cast, depending.
+    pattern: The regular expression used to find C-style casts.
+    error: The function to call with any errors found.
+
+  Returns:
+    True if an error was emitted.
+    False otherwise.
+  """
+  match = Search(pattern, line)
+  if not match:
+    return False
+
+  # e.g., sizeof(int)
+  sizeof_match = Match(r'.*sizeof\s*$', line[0:match.start(1) - 1])
+  if sizeof_match:
+    error(filename, linenum, 'runtime/sizeof', 1,
+          'Using sizeof(type).  Use sizeof(varname) instead if possible')
+    return True
+
+  # operator++(int) and operator--(int)
+  if (line[0:match.start(1) - 1].endswith(' operator++') or
+      line[0:match.start(1) - 1].endswith(' operator--')):
+    return False
+
+  # A single unnamed argument for a function tends to look like old
+  # style cast.  If we see those, don't issue warnings for deprecated
+  # casts, instead issue warnings for unnamed arguments where
+  # appropriate.
+  #
+  # These are things that we want warnings for, since the style guide
+  # explicitly require all parameters to be named:
+  #   Function(int);
+  #   Function(int) {
+  #   ConstMember(int) const;
+  #   ConstMember(int) const {
+  #   ExceptionMember(int) throw (...);
+  #   ExceptionMember(int) throw (...) {
+  #   PureVirtual(int) = 0;
+  #
+  # These are functions of some sort, where the compiler would be fine
+  # if they had named parameters, but people often omit those
+  # identifiers to reduce clutter:
+  #   (FunctionPointer)(int);
+  #   (FunctionPointer)(int) = value;
+  #   Function((function_pointer_arg)(int))
+  #   <TemplateArgument(int)>;
+  #   <(FunctionPointerTemplateArgument)(int)>;
+  remainder = line[match.end(0):]
+  if Match(r'^\s*(?:;|const\b|throw\b|=|>|\{|\))', remainder):
+    # Looks like an unnamed parameter.
+
+    # Don't warn on any kind of template arguments.
+    if Match(r'^\s*>', remainder):
+      return False
+
+    # Don't warn on assignments to function pointers, but keep warnings for
+    # unnamed parameters to pure virtual functions.  Note that this pattern
+    # will also pass on assignments of "0" to function pointers, but the
+    # preferred values for those would be "nullptr" or "NULL".
+    matched_zero = Match(r'^\s=\s*(\S+)\s*;', remainder)
+    if matched_zero and matched_zero.group(1) != '0':
+      return False
+
+    # Don't warn on function pointer declarations.  For this we need
+    # to check what came before the "(type)" string.
+    if Match(r'.*\)\s*$', line[0:match.start(0)]):
+      return False
+
+    # Don't warn if the parameter is named with block comments, e.g.:
+    #  Function(int /*unused_param*/);
+    if '/*' in raw_line:
+      return False
+
+    # Passed all filters, issue warning here.
+    error(filename, linenum, 'readability/function', 3,
+          'All parameters should be named in a function')
+    return True
+
+  # At this point, all that should be left is actual casts.
+  error(filename, linenum, 'readability/casting', 4,
+        'Using C-style cast.  Use %s<%s>(...) instead' %
+        (cast_type, match.group(1)))
+
+  return True
+
+
+_HEADERS_CONTAINING_TEMPLATES = (
+    ('<deque>', ('deque',)),
+    ('<functional>', ('unary_function', 'binary_function',
+                      'plus', 'minus', 'multiplies', 'divides', 'modulus',
+                      'negate',
+                      'equal_to', 'not_equal_to', 'greater', 'less',
+                      'greater_equal', 'less_equal',
+                      'logical_and', 'logical_or', 'logical_not',
+                      'unary_negate', 'not1', 'binary_negate', 'not2',
+                      'bind1st', 'bind2nd',
+                      'pointer_to_unary_function',
+                      'pointer_to_binary_function',
+                      'ptr_fun',
+                      'mem_fun_t', 'mem_fun', 'mem_fun1_t', 'mem_fun1_ref_t',
+                      'mem_fun_ref_t',
+                      'const_mem_fun_t', 'const_mem_fun1_t',
+                      'const_mem_fun_ref_t', 'const_mem_fun1_ref_t',
+                      'mem_fun_ref',
+                     )),
+    ('<limits>', ('numeric_limits',)),
+    ('<list>', ('list',)),
+    ('<map>', ('map', 'multimap',)),
+    ('<memory>', ('allocator',)),
+    ('<queue>', ('queue', 'priority_queue',)),
+    ('<set>', ('set', 'multiset',)),
+    ('<stack>', ('stack',)),
+    ('<string>', ('char_traits', 'basic_string',)),
+    ('<utility>', ('pair',)),
+    ('<vector>', ('vector',)),
+
+    # gcc extensions.
+    # Note: std::hash is their hash, ::hash is our hash
+    ('<hash_map>', ('hash_map', 'hash_multimap',)),
+    ('<hash_set>', ('hash_set', 'hash_multiset',)),
+    ('<slist>', ('slist',)),
+    )
+
+_RE_PATTERN_STRING = re.compile(r'\bstring\b')
+
+_re_pattern_algorithm_header = []
+for _template in ('copy', 'max', 'min', 'min_element', 'sort', 'swap',
+                  'transform'):
+  # Match max<type>(..., ...), max(..., ...), but not foo->max, foo.max or
+  # type::max().
+  _re_pattern_algorithm_header.append(
+      (re.compile(r'[^>.]\b' + _template + r'(<.*?>)?\([^\)]'),
+       _template,
+       '<algorithm>'))
+
+_re_pattern_templates = []
+for _header, _templates in _HEADERS_CONTAINING_TEMPLATES:
+  for _template in _templates:
+    _re_pattern_templates.append(
+        (re.compile(r'(\<|\b)' + _template + r'\s*\<'),
+         _template + '<>',
+         _header))
+
+
+def FilesBelongToSameModule(filename_cc, filename_h):
+  """Check if these two filenames belong to the same module.
+
+  The concept of a 'module' here is a as follows:
+  foo.h, foo-inl.h, foo.cc, foo_test.cc and foo_unittest.cc belong to the
+  same 'module' if they are in the same directory.
+  some/path/public/xyzzy and some/path/internal/xyzzy are also considered
+  to belong to the same module here.
+
+  If the filename_cc contains a longer path than the filename_h, for example,
+  '/absolute/path/to/base/sysinfo.cc', and this file would include
+  'base/sysinfo.h', this function also produces the prefix needed to open the
+  header. This is used by the caller of this function to more robustly open the
+  header file. We don't have access to the real include paths in this context,
+  so we need this guesswork here.
+
+  Known bugs: tools/base/bar.cc and base/bar.h belong to the same module
+  according to this implementation. Because of this, this function gives
+  some false positives. This should be sufficiently rare in practice.
+
+  Args:
+    filename_cc: is the path for the .cc file
+    filename_h: is the path for the header path
+
+  Returns:
+    Tuple with a bool and a string:
+    bool: True if filename_cc and filename_h belong to the same module.
+    string: the additional prefix needed to open the header file.
+  """
+
+  if not filename_cc.endswith('.cc'):
+    return (False, '')
+  filename_cc = filename_cc[:-len('.cc')]
+  if filename_cc.endswith('_unittest'):
+    filename_cc = filename_cc[:-len('_unittest')]
+  elif filename_cc.endswith('_test'):
+    filename_cc = filename_cc[:-len('_test')]
+  filename_cc = filename_cc.replace('/public/', '/')
+  filename_cc = filename_cc.replace('/internal/', '/')
+
+  if not filename_h.endswith('.h'):
+    return (False, '')
+  filename_h = filename_h[:-len('.h')]
+  if filename_h.endswith('-inl'):
+    filename_h = filename_h[:-len('-inl')]
+  filename_h = filename_h.replace('/public/', '/')
+  filename_h = filename_h.replace('/internal/', '/')
+
+  files_belong_to_same_module = filename_cc.endswith(filename_h)
+  common_path = ''
+  if files_belong_to_same_module:
+    common_path = filename_cc[:-len(filename_h)]
+  return files_belong_to_same_module, common_path
+
+
+def UpdateIncludeState(filename, include_state, io=codecs):
+  """Fill up the include_state with new includes found from the file.
+
+  Args:
+    filename: the name of the header to read.
+    include_state: an _IncludeState instance in which the headers are inserted.
+    io: The io factory to use to read the file. Provided for testability.
+
+  Returns:
+    True if a header was succesfully added. False otherwise.
+  """
+  headerfile = None
+  try:
+    headerfile = io.open(filename, 'r', 'utf8', 'replace')
+  except IOError:
+    return False
+  linenum = 0
+  for line in headerfile:
+    linenum += 1
+    clean_line = CleanseComments(line)
+    match = _RE_PATTERN_INCLUDE.search(clean_line)
+    if match:
+      include = match.group(2)
+      # The value formatting is cute, but not really used right now.
+      # What matters here is that the key is in include_state.
+      include_state.setdefault(include, '%s:%d' % (filename, linenum))
+  return True
+
+
+def CheckForIncludeWhatYouUse(filename, clean_lines, include_state, error,
+                              io=codecs):
+  """Reports for missing stl includes.
+
+  This function will output warnings to make sure you are including the headers
+  necessary for the stl containers and functions that you use. We only give one
+  reason to include a header. For example, if you use both equal_to<> and
+  less<> in a .h file, only one (the latter in the file) of these will be
+  reported as a reason to include the <functional>.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    include_state: An _IncludeState instance.
+    error: The function to call with any errors found.
+    io: The IO factory to use to read the header file. Provided for unittest
+        injection.
+  """
+  required = {}  # A map of header name to linenumber and the template entity.
+                 # Example of required: { '<functional>': (1219, 'less<>') }
+
+  for linenum in xrange(clean_lines.NumLines()):
+    line = clean_lines.elided[linenum]
+    if not line or line[0] == '#':
+      continue
+
+    # String is special -- it is a non-templatized type in STL.
+    matched = _RE_PATTERN_STRING.search(line)
+    if matched:
+      # Don't warn about strings in non-STL namespaces:
+      # (We check only the first match per line; good enough.)
+      prefix = line[:matched.start()]
+      if prefix.endswith('std::') or not prefix.endswith('::'):
+        required['<string>'] = (linenum, 'string')
+
+    for pattern, template, header in _re_pattern_algorithm_header:
+      if pattern.search(line):
+        required[header] = (linenum, template)
+
+    # The following function is just a speed up, no semantics are changed.
+    if not '<' in line:  # Reduces the cpu time usage by skipping lines.
+      continue
+
+    for pattern, template, header in _re_pattern_templates:
+      if pattern.search(line):
+        required[header] = (linenum, template)
+
+  # The policy is that if you #include something in foo.h you don't need to
+  # include it again in foo.cc. Here, we will look at possible includes.
+  # Let's copy the include_state so it is only messed up within this function.
+  include_state = include_state.copy()
+
+  # Did we find the header for this file (if any) and succesfully load it?
+  header_found = False
+
+  # Use the absolute path so that matching works properly.
+  abs_filename = FileInfo(filename).FullName()
+
+  # For Emacs's flymake.
+  # If cpplint is invoked from Emacs's flymake, a temporary file is generated
+  # by flymake and that file name might end with '_flymake.cc'. In that case,
+  # restore original file name here so that the corresponding header file can be
+  # found.
+  # e.g. If the file name is 'foo_flymake.cc', we should search for 'foo.h'
+  # instead of 'foo_flymake.h'
+  abs_filename = re.sub(r'_flymake\.cc$', '.cc', abs_filename)
+
+  # include_state is modified during iteration, so we iterate over a copy of
+  # the keys.
+  header_keys = include_state.keys()
+  for header in header_keys:
+    (same_module, common_path) = FilesBelongToSameModule(abs_filename, header)
+    fullpath = common_path + header
+    if same_module and UpdateIncludeState(fullpath, include_state, io):
+      header_found = True
+
+  # If we can't find the header file for a .cc, assume it's because we don't
+  # know where to look. In that case we'll give up as we're not sure they
+  # didn't include it in the .h file.
+  # TODO(unknown): Do a better job of finding .h files so we are confident that
+  # not having the .h file means there isn't one.
+  if filename.endswith('.cc') and not header_found:
+    return
+
+  # All the lines have been processed, report the errors found.
+  for required_header_unstripped in required:
+    template = required[required_header_unstripped][1]
+    if required_header_unstripped.strip('<>"') not in include_state:
+      error(filename, required[required_header_unstripped][0],
+            'build/include_what_you_use', 4,
+            'Add #include ' + required_header_unstripped + ' for ' + template)
+
+
+_RE_PATTERN_EXPLICIT_MAKEPAIR = re.compile(r'\bmake_pair\s*<')
+
+
+def CheckMakePairUsesDeduction(filename, clean_lines, linenum, error):
+  """Check that make_pair's template arguments are deduced.
+
+  G++ 4.6 in C++0x mode fails badly if make_pair's template arguments are
+  specified explicitly, and such use isn't intended in any case.
+
+  Args:
+    filename: The name of the current file.
+    clean_lines: A CleansedLines instance containing the file.
+    linenum: The number of the line to check.
+    error: The function to call with any errors found.
+  """
+  line = clean_lines.elided[linenum]
+  match = _RE_PATTERN_EXPLICIT_MAKEPAIR.search(line)
+  if match:
+    error(filename, linenum, 'build/explicit_make_pair',
+          4,  # 4 = high confidence
+          'For C++11-compatibility, omit template arguments from make_pair'
+          ' OR use pair directly OR if appropriate, construct a pair directly')
+
+
+def ProcessLine(filename, file_extension, clean_lines, line,
+                include_state, function_state, nesting_state, error,
+                extra_check_functions=[]):
+  """Processes a single line in the file.
+
+  Args:
+    filename: Filename of the file that is being processed.
+    file_extension: The extension (dot not included) of the file.
+    clean_lines: An array of strings, each representing a line of the file,
+                 with comments stripped.
+    line: Number of line being processed.
+    include_state: An _IncludeState instance in which the headers are inserted.
+    function_state: A _FunctionState instance which counts function lines, etc.
+    nesting_state: A _NestingState instance which maintains information about
+                   the current stack of nested blocks being parsed.
+    error: A callable to which errors are reported, which takes 4 arguments:
+           filename, line number, error level, and message
+    extra_check_functions: An array of additional check functions that will be
+                           run on each source line. Each function takes 4
+                           arguments: filename, clean_lines, line, error
+  """
+  raw_lines = clean_lines.raw_lines
+  ParseNolintSuppressions(filename, raw_lines[line], line, error)
+  nesting_state.Update(filename, clean_lines, line, error)
+  if nesting_state.stack and nesting_state.stack[-1].inline_asm != _NO_ASM:
+    return
+  CheckForFunctionLengths(filename, clean_lines, line, function_state, error)
+  CheckForMultilineCommentsAndStrings(filename, clean_lines, line, error)
+  CheckStyle(filename, clean_lines, line, file_extension, nesting_state, error)
+  CheckLanguage(filename, clean_lines, line, file_extension, include_state,
+                nesting_state, error)
+  CheckForNonConstReference(filename, clean_lines, line, nesting_state, error)
+  CheckForNonStandardConstructs(filename, clean_lines, line,
+                                nesting_state, error)
+  CheckVlogArguments(filename, clean_lines, line, error)
+  CheckPosixThreading(filename, clean_lines, line, error)
+  CheckInvalidIncrement(filename, clean_lines, line, error)
+  CheckMakePairUsesDeduction(filename, clean_lines, line, error)
+  for check_fn in extra_check_functions:
+    check_fn(filename, clean_lines, line, error)
+
+def ProcessFileData(filename, file_extension, lines, error,
+                    extra_check_functions=[]):
+  """Performs lint checks and reports any errors to the given error function.
+
+  Args:
+    filename: Filename of the file that is being processed.
+    file_extension: The extension (dot not included) of the file.
+    lines: An array of strings, each representing a line of the file, with the
+           last element being empty if the file is terminated with a newline.
+    error: A callable to which errors are reported, which takes 4 arguments:
+           filename, line number, error level, and message
+    extra_check_functions: An array of additional check functions that will be
+                           run on each source line. Each function takes 4
+                           arguments: filename, clean_lines, line, error
+  """
+  lines = (['// marker so line numbers and indices both start at 1'] + lines +
+           ['// marker so line numbers end in a known way'])
+
+  include_state = _IncludeState()
+  function_state = _FunctionState()
+  nesting_state = _NestingState()
+
+  ResetNolintSuppressions()
+
+  CheckForCopyright(filename, lines, error)
+
+  if file_extension == 'h':
+    CheckForHeaderGuard(filename, lines, error)
+
+  RemoveMultiLineComments(filename, lines, error)
+  clean_lines = CleansedLines(lines)
+  for line in xrange(clean_lines.NumLines()):
+    ProcessLine(filename, file_extension, clean_lines, line,
+                include_state, function_state, nesting_state, error,
+                extra_check_functions)
+  nesting_state.CheckCompletedBlocks(filename, error)
+
+  CheckForIncludeWhatYouUse(filename, clean_lines, include_state, error)
+
+  # We check here rather than inside ProcessLine so that we see raw
+  # lines rather than "cleaned" lines.
+  CheckForBadCharacters(filename, lines, error)
+
+  CheckForNewlineAtEOF(filename, lines, error)
+
+def ProcessFile(filename, vlevel, extra_check_functions=[]):
+  """Does google-lint on a single file.
+
+  Args:
+    filename: The name of the file to parse.
+
+    vlevel: The level of errors to report.  Every error of confidence
+    >= verbose_level will be reported.  0 is a good default.
+
+    extra_check_functions: An array of additional check functions that will be
+                           run on each source line. Each function takes 4
+                           arguments: filename, clean_lines, line, error
+  """
+
+  _SetVerboseLevel(vlevel)
+
+  try:
+    # Support the UNIX convention of using "-" for stdin.  Note that
+    # we are not opening the file with universal newline support
+    # (which codecs doesn't support anyway), so the resulting lines do
+    # contain trailing '\r' characters if we are reading a file that
+    # has CRLF endings.
+    # If after the split a trailing '\r' is present, it is removed
+    # below. If it is not expected to be present (i.e. os.linesep !=
+    # '\r\n' as in Windows), a warning is issued below if this file
+    # is processed.
+
+    if filename == '-':
+      lines = codecs.StreamReaderWriter(sys.stdin,
+                                        codecs.getreader('utf8'),
+                                        codecs.getwriter('utf8'),
+                                        'replace').read().split('\n')
+    else:
+      lines = codecs.open(filename, 'r', 'utf8', 'replace').read().split('\n')
+
+    carriage_return_found = False
+    # Remove trailing '\r'.
+    for linenum in range(len(lines)):
+      if lines[linenum].endswith('\r'):
+        lines[linenum] = lines[linenum].rstrip('\r')
+        carriage_return_found = True
+
+  except IOError:
+    sys.stderr.write(
+        "Skipping input '%s': Can't open for reading\n" % filename)
+    return
+
+  # Note, if no dot is found, this will give the entire filename as the ext.
+  file_extension = filename[filename.rfind('.') + 1:]
+
+  # When reading from stdin, the extension is unknown, so no cpplint tests
+  # should rely on the extension.
+  if filename != '-' and file_extension not in _valid_extensions:
+    sys.stderr.write('Ignoring %s; not a valid file name '
+                     '(%s)\n' % (filename, ', '.join(_valid_extensions)))
+  else:
+    ProcessFileData(filename, file_extension, lines, Error,
+                    extra_check_functions)
+    if carriage_return_found and os.linesep != '\r\n':
+      # Use 0 for linenum since outputting only one error for potentially
+      # several lines.
+      Error(filename, 0, 'whitespace/newline', 1,
+            'One or more unexpected \\r (^M) found;'
+            'better to use only a \\n')
+
+  sys.stderr.write('Done processing %s\n' % filename)
+
+
+def PrintUsage(message):
+  """Prints a brief usage string and exits, optionally with an error message.
+
+  Args:
+    message: The optional error message.
+  """
+  sys.stderr.write(_USAGE)
+  if message:
+    sys.exit('\nFATAL ERROR: ' + message)
+  else:
+    sys.exit(1)
+
+
+def PrintCategories():
+  """Prints a list of all the error-categories used by error messages.
+
+  These are the categories used to filter messages via --filter.
+  """
+  sys.stderr.write(''.join('  %s\n' % cat for cat in _ERROR_CATEGORIES))
+  sys.exit(0)
+
+
+def ParseArguments(args):
+  """Parses the command line arguments.
+
+  This may set the output format and verbosity level as side-effects.
+
+  Args:
+    args: The command line arguments:
+
+  Returns:
+    The list of filenames to lint.
+  """
+  try:
+    (opts, filenames) = getopt.getopt(args, '', ['help', 'output=', 'verbose=',
+                                                 'counting=',
+                                                 'filter=',
+                                                 'root=',
+                                                 'linelength=',
+                                                 'extensions='])
+  except getopt.GetoptError:
+    PrintUsage('Invalid arguments.')
+
+  verbosity = _VerboseLevel()
+  output_format = _OutputFormat()
+  filters = ''
+  counting_style = ''
+
+  for (opt, val) in opts:
+    if opt == '--help':
+      PrintUsage(None)
+    elif opt == '--output':
+      if val not in ('emacs', 'vs7', 'eclipse'):
+        PrintUsage('The only allowed output formats are emacs, vs7 and eclipse.')
+      output_format = val
+    elif opt == '--verbose':
+      verbosity = int(val)
+    elif opt == '--filter':
+      filters = val
+      if not filters:
+        PrintCategories()
+    elif opt == '--counting':
+      if val not in ('total', 'toplevel', 'detailed'):
+        PrintUsage('Valid counting options are total, toplevel, and detailed')
+      counting_style = val
+    elif opt == '--root':
+      global _root
+      _root = val
+    elif opt == '--linelength':
+      global _line_length
+      try:
+          _line_length = int(val)
+      except ValueError:
+          PrintUsage('Line length must be digits.')
+    elif opt == '--extensions':
+      global _valid_extensions
+      try:
+          _valid_extensions = set(val.split(','))
+      except ValueError:
+          PrintUsage('Extensions must be comma seperated list.')
+
+  if not filenames:
+    PrintUsage('No files were specified.')
+
+  _SetOutputFormat(output_format)
+  _SetVerboseLevel(verbosity)
+  _SetFilters(filters)
+  _SetCountingStyle(counting_style)
+
+  return filenames
+
+
+def main():
+  filenames = ParseArguments(sys.argv[1:])
+
+  # Change stderr to write with replacement characters so we don't die
+  # if we try to print something containing non-ASCII characters.
+  sys.stderr = codecs.StreamReaderWriter(sys.stderr,
+                                         codecs.getreader('utf8'),
+                                         codecs.getwriter('utf8'),
+                                         'replace')
+
+  _cpplint_state.ResetErrorCounts()
+  for filename in filenames:
+    ProcessFile(filename, _cpplint_state.verbose_level)
+  _cpplint_state.PrintErrorCounts()
+
+  sys.exit(_cpplint_state.error_count > 0)
+
+
+if __name__ == '__main__':
+  main()
diff --git a/third_party/aom/tools/diff.py b/third_party/aom/tools/diff.py
new file mode 100644
index 0000000000..bac6aabdc0
--- /dev/null
+++ b/third_party/aom/tools/diff.py
@@ -0,0 +1,132 @@
+#!/usr/bin/env python
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+"""Classes for representing diff pieces."""
+
+__author__ = "jkoleszar@google.com"
+
+import re
+
+
+class DiffLines(object):
+    """A container for one half of a diff."""
+
+    def __init__(self, filename, offset, length):
+        self.filename = filename
+        self.offset = offset
+        self.length = length
+        self.lines = []
+        self.delta_line_nums = []
+
+    def Append(self, line):
+        l = len(self.lines)
+        if line[0] != " ":
+            self.delta_line_nums.append(self.offset + l)
+        self.lines.append(line[1:])
+        assert l+1 <= self.length
+
+    def Complete(self):
+        return len(self.lines) == self.length
+
+    def __contains__(self, item):
+        return item >= self.offset and item <= self.offset + self.length - 1
+
+
+class DiffHunk(object):
+    """A container for one diff hunk, consisting of two DiffLines."""
+
+    def __init__(self, header, file_a, file_b, start_a, len_a, start_b, len_b):
+        self.header = header
+        self.left = DiffLines(file_a, start_a, len_a)
+        self.right = DiffLines(file_b, start_b, len_b)
+        self.lines = []
+
+    def Append(self, line):
+        """Adds a line to the DiffHunk and its DiffLines children."""
+        if line[0] == "-":
+            self.left.Append(line)
+        elif line[0] == "+":
+            self.right.Append(line)
+        elif line[0] == " ":
+            self.left.Append(line)
+            self.right.Append(line)
+        elif line[0] == "\\":
+            # Ignore newline messages from git diff.
+            pass
+        else:
+            assert False, ("Unrecognized character at start of diff line "
+                           "%r" % line[0])
+        self.lines.append(line)
+
+    def Complete(self):
+        return self.left.Complete() and self.right.Complete()
+
+    def __repr__(self):
+        return "DiffHunk(%s, %s, len %d)" % (
+            self.left.filename, self.right.filename,
+            max(self.left.length, self.right.length))
+
+
+def ParseDiffHunks(stream):
+    """Walk a file-like object, yielding DiffHunks as they're parsed."""
+
+    file_regex = re.compile(r"(\+\+\+|---) (\S+)")
+    range_regex = re.compile(r"@@ -(\d+)(,(\d+))? \+(\d+)(,(\d+))?")
+    hunk = None
+    while True:
+        line = stream.readline()
+        if not line:
+            break
+
+        if hunk is None:
+            # Parse file names
+            diff_file = file_regex.match(line)
+            if diff_file:
+              if line.startswith("---"):
+                  a_line = line
+                  a = diff_file.group(2)
+                  continue
+              if line.startswith("+++"):
+                  b_line = line
+                  b = diff_file.group(2)
+                  continue
+
+            # Parse offset/lengths
+            diffrange = range_regex.match(line)
+            if diffrange:
+                if diffrange.group(2):
+                    start_a = int(diffrange.group(1))
+                    len_a = int(diffrange.group(3))
+                else:
+                    start_a = 1
+                    len_a = int(diffrange.group(1))
+
+                if diffrange.group(5):
+                    start_b = int(diffrange.group(4))
+                    len_b = int(diffrange.group(6))
+                else:
+                    start_b = 1
+                    len_b = int(diffrange.group(4))
+
+                header = [a_line, b_line, line]
+                hunk = DiffHunk(header, a, b, start_a, len_a, start_b, len_b)
+        else:
+            # Add the current line to the hunk
+            hunk.Append(line)
+
+            # See if the whole hunk has been parsed. If so, yield it and prepare
+            # for the next hunk.
+            if hunk.Complete():
+                yield hunk
+                hunk = None
+
+    # Partial hunks are a parse error
+    assert hunk is None
diff --git a/third_party/aom/tools/ftfy.sh b/third_party/aom/tools/ftfy.sh
new file mode 100755
index 0000000000..315da1af5d
--- /dev/null
+++ b/third_party/aom/tools/ftfy.sh
@@ -0,0 +1,158 @@
+#!/bin/sh
+self="$0"
+dirname_self=$(dirname "$self")
+
+usage() {
+  cat <<EOF >&2
+Usage: $self [option]
+
+This script applies a whitespace transformation to the commit at HEAD. If no
+options are given, then the modified files are left in the working tree.
+
+Options:
+  -h, --help     Shows this message
+  -n, --dry-run  Shows a diff of the changes to be made.
+  --amend        Squashes the changes into the commit at HEAD
+                     This option will also reformat the commit message.
+  --commit       Creates a new commit containing only the whitespace changes
+  --msg-only     Reformat the commit message only, ignore the patch itself.
+
+EOF
+  rm -f ${CLEAN_FILES}
+  exit 1
+}
+
+
+log() {
+  echo "${self##*/}: $@" >&2
+}
+
+
+aom_style() {
+  for f; do
+    case "$f" in
+      *.h|*.c|*.cc)
+        clang-format -i --style=file "$f"
+        ;;
+    esac
+  done
+}
+
+
+apply() {
+  [ $INTERSECT_RESULT -ne 0 ] && patch -p1 < "$1"
+}
+
+
+commit() {
+  LAST_CHANGEID=$(git show | awk '/Change-Id:/{print $2}')
+  if [ -z "$LAST_CHANGEID" ]; then
+    log "HEAD doesn't have a Change-Id, unable to generate a new commit"
+    exit 1
+  fi
+
+  # Build a deterministic Change-Id from the parent's
+  NEW_CHANGEID=${LAST_CHANGEID}-styled
+  NEW_CHANGEID=I$(echo $NEW_CHANGEID | git hash-object --stdin)
+
+  # Commit, preserving authorship from the parent commit.
+  git commit -a -C HEAD > /dev/null
+  git commit --amend -F- << EOF
+Cosmetic: Fix whitespace in change ${LAST_CHANGEID:0:9}
+
+Change-Id: ${NEW_CHANGEID}
+EOF
+}
+
+
+show_commit_msg_diff() {
+  if [ $DIFF_MSG_RESULT -ne 0 ]; then
+    log "Modified commit message:"
+    diff -u "$ORIG_COMMIT_MSG" "$NEW_COMMIT_MSG" | tail -n +3
+  fi
+}
+
+
+amend() {
+  show_commit_msg_diff
+  if [ $DIFF_MSG_RESULT -ne 0 ] || [ $INTERSECT_RESULT -ne 0 ]; then
+    git commit -a --amend -F "$NEW_COMMIT_MSG"
+  fi
+}
+
+
+diff_msg() {
+  git log -1 --format=%B > "$ORIG_COMMIT_MSG"
+  "${dirname_self}"/wrap-commit-msg.py \
+      < "$ORIG_COMMIT_MSG" > "$NEW_COMMIT_MSG"
+  cmp -s "$ORIG_COMMIT_MSG" "$NEW_COMMIT_MSG"
+  DIFF_MSG_RESULT=$?
+}
+
+
+# Temporary files
+ORIG_DIFF=orig.diff.$$
+MODIFIED_DIFF=modified.diff.$$
+FINAL_DIFF=final.diff.$$
+ORIG_COMMIT_MSG=orig.commit-msg.$$
+NEW_COMMIT_MSG=new.commit-msg.$$
+CLEAN_FILES="${ORIG_DIFF} ${MODIFIED_DIFF} ${FINAL_DIFF}"
+CLEAN_FILES="${CLEAN_FILES} ${ORIG_COMMIT_MSG} ${NEW_COMMIT_MSG}"
+
+# Preconditions
+[ $# -lt 2 ] || usage
+
+if ! clang-format -version >/dev/null 2>&1; then
+  log "clang-format not found"
+  exit 1
+fi
+
+if ! git diff --quiet HEAD; then
+  log "Working tree is dirty, commit your changes first"
+  exit 1
+fi
+
+# Need to be in the root
+cd "$(git rev-parse --show-toplevel)"
+
+# Collect the original diff
+git show > "${ORIG_DIFF}"
+
+# Apply the style guide on new and modified files and collect its diff
+for f in $(git diff HEAD^ --name-only -M90 --diff-filter=AM); do
+  case "$f" in
+    third_party/*) continue;;
+  esac
+  aom_style "$f"
+done
+git diff --no-color --no-ext-diff > "${MODIFIED_DIFF}"
+
+# Intersect the two diffs
+"${dirname_self}"/intersect-diffs.py \
+    "${ORIG_DIFF}" "${MODIFIED_DIFF}" > "${FINAL_DIFF}"
+INTERSECT_RESULT=$?
+git reset --hard >/dev/null
+
+# Fixup the commit message
+diff_msg
+
+# Handle options
+if [ -n "$1" ]; then
+  case "$1" in
+    -h|--help) usage;;
+    -n|--dry-run) cat "${FINAL_DIFF}"; show_commit_msg_diff;;
+    --commit) apply "${FINAL_DIFF}"; commit;;
+    --amend) apply "${FINAL_DIFF}"; amend;;
+    --msg-only) amend;;
+    *) usage;;
+  esac
+else
+  apply "${FINAL_DIFF}"
+  if ! git diff --quiet; then
+    log "Formatting changes applied, verify and commit."
+    log "See also: http://www.webmproject.org/code/contribute/conventions/"
+    git diff --stat
+  fi
+fi
+
+rm -f ${CLEAN_FILES}
diff --git a/third_party/aom/tools/gen_authors.sh b/third_party/aom/tools/gen_authors.sh
new file mode 100755
index 0000000000..5def8bc898
--- /dev/null
+++ b/third_party/aom/tools/gen_authors.sh
@@ -0,0 +1,10 @@
+#!/bin/bash
+
+# Add organization names manually.
+
+cat <<EOF
+# This file is automatically generated from the git commit history
+# by tools/gen_authors.sh.
+
+$(git log --pretty=format:"%aN <%aE>" | sort | uniq | grep -v "corp.google\|clang-format")
+EOF
diff --git a/third_party/aom/tools/gen_constrained_tokenset.py b/third_party/aom/tools/gen_constrained_tokenset.py
new file mode 100755
index 0000000000..5d12ee1ef5
--- /dev/null
+++ b/third_party/aom/tools/gen_constrained_tokenset.py
@@ -0,0 +1,120 @@
+#!/usr/bin/python
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+"""Generate the probability model for the constrained token set.
+
+Model obtained from a 2-sided zero-centered distribution derived
+from a Pareto distribution. The cdf of the distribution is:
+cdf(x) = 0.5 + 0.5 * sgn(x) * [1 - {alpha/(alpha + |x|)} ^ beta]
+
+For a given beta and a given probability of the 1-node, the alpha
+is first solved, and then the {alpha, beta} pair is used to generate
+the probabilities for the rest of the nodes.
+"""
+
+import heapq
+import sys
+import numpy as np
+import scipy.optimize
+import scipy.stats
+
+
+def cdf_spareto(x, xm, beta):
+  p = 1 - (xm / (np.abs(x) + xm))**beta
+  p = 0.5 + 0.5 * np.sign(x) * p
+  return p
+
+
+def get_spareto(p, beta):
+  cdf = cdf_spareto
+
+  def func(x):
+    return ((cdf(1.5, x, beta) - cdf(0.5, x, beta)) /
+            (1 - cdf(0.5, x, beta)) - p)**2
+
+  alpha = scipy.optimize.fminbound(func, 1e-12, 10000, xtol=1e-12)
+  parray = np.zeros(11)
+  parray[0] = 2 * (cdf(0.5, alpha, beta) - 0.5)
+  parray[1] = (2 * (cdf(1.5, alpha, beta) - cdf(0.5, alpha, beta)))
+  parray[2] = (2 * (cdf(2.5, alpha, beta) - cdf(1.5, alpha, beta)))
+  parray[3] = (2 * (cdf(3.5, alpha, beta) - cdf(2.5, alpha, beta)))
+  parray[4] = (2 * (cdf(4.5, alpha, beta) - cdf(3.5, alpha, beta)))
+  parray[5] = (2 * (cdf(6.5, alpha, beta) - cdf(4.5, alpha, beta)))
+  parray[6] = (2 * (cdf(10.5, alpha, beta) - cdf(6.5, alpha, beta)))
+  parray[7] = (2 * (cdf(18.5, alpha, beta) - cdf(10.5, alpha, beta)))
+  parray[8] = (2 * (cdf(34.5, alpha, beta) - cdf(18.5, alpha, beta)))
+  parray[9] = (2 * (cdf(66.5, alpha, beta) - cdf(34.5, alpha, beta)))
+  parray[10] = 2 * (1. - cdf(66.5, alpha, beta))
+  return parray
+
+
+def quantize_probs(p, save_first_bin, bits):
+  """Quantize probability precisely.
+
+  Quantize probabilities minimizing dH (Kullback-Leibler divergence)
+  approximated by: sum (p_i-q_i)^2/p_i.
+  References:
+  https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence
+  https://github.com/JarekDuda/AsymmetricNumeralSystemsToolkit
+  """
+  num_sym = p.size
+  p = np.clip(p, 1e-16, 1)
+  L = 2**bits
+  pL = p * L
+  ip = 1. / p  # inverse probability
+  q = np.clip(np.round(pL), 1, L + 1 - num_sym)
+  quant_err = (pL - q)**2 * ip
+  sgn = np.sign(L - q.sum())  # direction of correction
+  if sgn != 0:  # correction is needed
+    v = []  # heap of adjustment results (adjustment err, index) of each symbol
+    for i in range(1 if save_first_bin else 0, num_sym):
+      q_adj = q[i] + sgn
+      if q_adj > 0 and q_adj < L:
+        adj_err = (pL[i] - q_adj)**2 * ip[i] - quant_err[i]
+        heapq.heappush(v, (adj_err, i))
+    while q.sum() != L:
+      # apply lowest error adjustment
+      (adj_err, i) = heapq.heappop(v)
+      quant_err[i] += adj_err
+      q[i] += sgn
+      # calculate the cost of adjusting this symbol again
+      q_adj = q[i] + sgn
+      if q_adj > 0 and q_adj < L:
+        adj_err = (pL[i] - q_adj)**2 * ip[i] - quant_err[i]
+        heapq.heappush(v, (adj_err, i))
+  return q
+
+
+def get_quantized_spareto(p, beta, bits, first_token):
+  parray = get_spareto(p, beta)
+  parray = parray[1:] / (1 - parray[0])
+  # CONFIG_NEW_TOKENSET
+  if first_token > 1:
+    parray = parray[1:] / (1 - parray[0])
+  qarray = quantize_probs(parray, first_token == 1, bits)
+  return qarray.astype(np.int)
+
+
+def main(bits=15, first_token=1):
+  beta = 8
+  for q in range(1, 256):
+    parray = get_quantized_spareto(q / 256., beta, bits, first_token)
+    assert parray.sum() == 2**bits
+    print '{', ', '.join('%d' % i for i in parray), '},'
+
+
+if __name__ == '__main__':
+  if len(sys.argv) > 2:
+    main(int(sys.argv[1]), int(sys.argv[2]))
+  elif len(sys.argv) > 1:
+    main(int(sys.argv[1]))
+  else:
+    main()
diff --git a/third_party/aom/tools/inspect-cli.js b/third_party/aom/tools/inspect-cli.js
new file mode 100644
index 0000000000..a14c08111a
--- /dev/null
+++ b/third_party/aom/tools/inspect-cli.js
@@ -0,0 +1,39 @@
+/**
+ * This tool lets you test if the compiled Javascript decoder is functioning properly. You'll
+ * need to download a SpiderMonkey js-shell to run this script.
+ * https://archive.mozilla.org/pub/firefox/nightly/latest-mozilla-central/
+ *
+ * Example:
+ *   js-shell inspect-cli.js video.ivf
+ */
+load("inspect.js");
+var buffer = read(scriptArgs[0], "binary");
+var Module = {
+  noExitRuntime: true,
+  noInitialRun: true,
+  preInit: [],
+  preRun: [],
+  postRun: [function () {
+    printErr(`Loaded Javascript Decoder OK`);
+  }],
+  memoryInitializerPrefixURL: "bin/",
+  arguments: ['input.ivf', 'output.raw'],
+  on_frame_decoded_json: function (jsonString) {
+    let json = JSON.parse("[" + Module.UTF8ToString(jsonString) + "null]");
+    json.forEach(frame => {
+      if (frame) {
+        print(frame.frame);
+      }
+    });
+  }
+};
+DecoderModule(Module);
+Module.FS.writeFile("/tmp/input.ivf", buffer, { encoding: "binary" });
+Module._open_file();
+Module._set_layers(0xFFFFFFFF); // Set this to zero if you want to benchmark decoding.
+while(true) {
+  printErr("Decoding Frame ...");
+  if (Module._read_frame()) {
+    break;
+  }
+}
diff --git a/third_party/aom/tools/inspect-post.js b/third_party/aom/tools/inspect-post.js
new file mode 100644
index 0000000000..31c40bb82c
--- /dev/null
+++ b/third_party/aom/tools/inspect-post.js
@@ -0,0 +1 @@
+Module["FS"] = FS;
diff --git a/third_party/aom/tools/intersect-diffs.py b/third_party/aom/tools/intersect-diffs.py
new file mode 100755
index 0000000000..df13c4ef70
--- /dev/null
+++ b/third_party/aom/tools/intersect-diffs.py
@@ -0,0 +1,78 @@
+#!/usr/bin/env python
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+"""Calculates the "intersection" of two unified diffs.
+
+Given two diffs, A and B, it finds all hunks in B that had non-context lines
+in A and prints them to stdout. This is useful to determine the hunks in B that
+are relevant to A. The resulting file can be applied with patch(1) on top of A.
+"""
+
+__author__ = "jkoleszar@google.com"
+
+import sys
+
+import diff
+
+
+def FormatDiffHunks(hunks):
+    """Re-serialize a list of DiffHunks."""
+    r = []
+    last_header = None
+    for hunk in hunks:
+        this_header = hunk.header[0:2]
+        if last_header != this_header:
+            r.extend(hunk.header)
+            last_header = this_header
+        else:
+            r.extend(hunk.header[2])
+        r.extend(hunk.lines)
+        r.append("\n")
+    return "".join(r)
+
+
+def ZipHunks(rhs_hunks, lhs_hunks):
+    """Join two hunk lists on filename."""
+    for rhs_hunk in rhs_hunks:
+        rhs_file = rhs_hunk.right.filename.split("/")[1:]
+
+        for lhs_hunk in lhs_hunks:
+            lhs_file = lhs_hunk.left.filename.split("/")[1:]
+            if lhs_file != rhs_file:
+                continue
+            yield (rhs_hunk, lhs_hunk)
+
+
+def main():
+    old_hunks = [x for x in diff.ParseDiffHunks(open(sys.argv[1], "r"))]
+    new_hunks = [x for x in diff.ParseDiffHunks(open(sys.argv[2], "r"))]
+    out_hunks = []
+
+    # Join the right hand side of the older diff with the left hand side of the
+    # newer diff.
+    for old_hunk, new_hunk in ZipHunks(old_hunks, new_hunks):
+        if new_hunk in out_hunks:
+            continue
+        old_lines = old_hunk.right
+        new_lines = new_hunk.left
+
+        # Determine if this hunk overlaps any non-context line from the other
+        for i in old_lines.delta_line_nums:
+            if i in new_lines:
+                out_hunks.append(new_hunk)
+                break
+
+    if out_hunks:
+        print FormatDiffHunks(out_hunks)
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/third_party/aom/tools/lint-hunks.py b/third_party/aom/tools/lint-hunks.py
new file mode 100755
index 0000000000..d02bee16ce
--- /dev/null
+++ b/third_party/aom/tools/lint-hunks.py
@@ -0,0 +1,146 @@
+#!/usr/bin/python
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+"""Performs style checking on each diff hunk."""
+import getopt
+import os
+import StringIO
+import subprocess
+import sys
+
+import diff
+
+
+SHORT_OPTIONS = "h"
+LONG_OPTIONS = ["help"]
+
+TOPLEVEL_CMD = ["git", "rev-parse", "--show-toplevel"]
+DIFF_CMD = ["git", "diff"]
+DIFF_INDEX_CMD = ["git", "diff-index", "-u", "HEAD", "--"]
+SHOW_CMD = ["git", "show"]
+CPPLINT_FILTERS = ["-readability/casting"]
+
+
+class Usage(Exception):
+    pass
+
+
+class SubprocessException(Exception):
+    def __init__(self, args):
+        msg = "Failed to execute '%s'"%(" ".join(args))
+        super(SubprocessException, self).__init__(msg)
+
+
+class Subprocess(subprocess.Popen):
+    """Adds the notion of an expected returncode to Popen."""
+
+    def __init__(self, args, expected_returncode=0, **kwargs):
+        self._args = args
+        self._expected_returncode = expected_returncode
+        super(Subprocess, self).__init__(args, **kwargs)
+
+    def communicate(self, *args, **kwargs):
+        result = super(Subprocess, self).communicate(*args, **kwargs)
+        if self._expected_returncode is not None:
+            try:
+                ok = self.returncode in self._expected_returncode
+            except TypeError:
+                ok = self.returncode == self._expected_returncode
+            if not ok:
+                raise SubprocessException(self._args)
+        return result
+
+
+def main(argv=None):
+    if argv is None:
+        argv = sys.argv
+    try:
+        try:
+            opts, args = getopt.getopt(argv[1:], SHORT_OPTIONS, LONG_OPTIONS)
+        except getopt.error, msg:
+            raise Usage(msg)
+
+        # process options
+        for o, _ in opts:
+            if o in ("-h", "--help"):
+                print __doc__
+                sys.exit(0)
+
+        if args and len(args) > 1:
+            print __doc__
+            sys.exit(0)
+
+        # Find the fully qualified path to the root of the tree
+        tl = Subprocess(TOPLEVEL_CMD, stdout=subprocess.PIPE)
+        tl = tl.communicate()[0].strip()
+
+        # See if we're working on the index or not.
+        if args:
+            diff_cmd = DIFF_CMD + [args[0] + "^!"]
+        else:
+            diff_cmd = DIFF_INDEX_CMD
+
+        # Build the command line to execute cpplint
+        cpplint_cmd = [os.path.join(tl, "tools", "cpplint.py"),
+                       "--filter=" + ",".join(CPPLINT_FILTERS),
+                       "-"]
+
+        # Get a list of all affected lines
+        file_affected_line_map = {}
+        p = Subprocess(diff_cmd, stdout=subprocess.PIPE)
+        stdout = p.communicate()[0]
+        for hunk in diff.ParseDiffHunks(StringIO.StringIO(stdout)):
+            filename = hunk.right.filename[2:]
+            if filename not in file_affected_line_map:
+                file_affected_line_map[filename] = set()
+            file_affected_line_map[filename].update(hunk.right.delta_line_nums)
+
+        # Run each affected file through cpplint
+        lint_failed = False
+        for filename, affected_lines in file_affected_line_map.iteritems():
+            if filename.split(".")[-1] not in ("c", "h", "cc"):
+                continue
+
+            if args:
+                # File contents come from git
+                show_cmd = SHOW_CMD + [args[0] + ":" + filename]
+                show = Subprocess(show_cmd, stdout=subprocess.PIPE)
+                lint = Subprocess(cpplint_cmd, expected_returncode=(0, 1),
+                                  stdin=show.stdout, stderr=subprocess.PIPE)
+                lint_out = lint.communicate()[1]
+            else:
+                # File contents come from the working tree
+                lint = Subprocess(cpplint_cmd, expected_returncode=(0, 1),
+                                  stdin=subprocess.PIPE, stderr=subprocess.PIPE)
+                stdin = open(os.path.join(tl, filename)).read()
+                lint_out = lint.communicate(stdin)[1]
+
+            for line in lint_out.split("\n"):
+                fields = line.split(":")
+                if fields[0] != "-":
+                    continue
+                warning_line_num = int(fields[1])
+                if warning_line_num in affected_lines:
+                    print "%s:%d:%s"%(filename, warning_line_num,
+                                      ":".join(fields[2:]))
+                    lint_failed = True
+
+        # Set exit code if any relevant lint errors seen
+        if lint_failed:
+            return 1
+
+    except Usage, err:
+        print >>sys.stderr, err
+        print >>sys.stderr, "for help use --help"
+        return 2
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/third_party/aom/tools/wrap-commit-msg.py b/third_party/aom/tools/wrap-commit-msg.py
new file mode 100755
index 0000000000..1c78824439
--- /dev/null
+++ b/third_party/aom/tools/wrap-commit-msg.py
@@ -0,0 +1,72 @@
+#!/usr/bin/env python
+##
+## Copyright (c) 2016, Alliance for Open Media. All rights reserved
+##
+## This source code is subject to the terms of the BSD 2 Clause License and
+## the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+## was not distributed with this source code in the LICENSE file, you can
+## obtain it at www.aomedia.org/license/software. If the Alliance for Open
+## Media Patent License 1.0 was not distributed with this source code in the
+## PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+##
+"""Wraps paragraphs of text, preserving manual formatting
+
+This is like fold(1), but has the special convention of not modifying lines
+that start with whitespace. This allows you to intersperse blocks with
+special formatting, like code blocks, with written prose. The prose will
+be wordwrapped, and the manual formatting will be preserved.
+
+ * This won't handle the case of a bulleted (or ordered) list specially, so
+   manual wrapping must be done.
+
+Occasionally it's useful to put something with explicit formatting that
+doesn't look at all like a block of text inline.
+
+  indicator = has_leading_whitespace(line);
+  if (indicator)
+    preserve_formatting(line);
+
+The intent is that this docstring would make it through the transform
+and still be legible and presented as it is in the source. If additional
+cases are handled, update this doc to describe the effect.
+"""
+
+__author__ = "jkoleszar@google.com"
+import textwrap
+import sys
+
+def wrap(text):
+    if text:
+        return textwrap.fill(text, break_long_words=False) + '\n'
+    return ""
+
+
+def main(fileobj):
+    text = ""
+    output = ""
+    while True:
+        line = fileobj.readline()
+        if not line:
+            break
+
+        if line.lstrip() == line:
+            text += line
+        else:
+            output += wrap(text)
+            text=""
+            output += line
+    output += wrap(text)
+
+    # Replace the file or write to stdout.
+    if fileobj == sys.stdin:
+        fileobj = sys.stdout
+    else:
+        fileobj.seek(0)
+        fileobj.truncate(0)
+    fileobj.write(output)
+
+if __name__ == "__main__":
+    if len(sys.argv) > 1:
+        main(open(sys.argv[1], "r+"))
+    else:
+        main(sys.stdin)
diff --git a/third_party/aom/tools_common.c b/third_party/aom/tools_common.c
new file mode 100644
index 0000000000..353021093f
--- /dev/null
+++ b/third_party/aom/tools_common.c
@@ -0,0 +1,454 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <math.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "./tools_common.h"
+
+#if CONFIG_AV1_ENCODER
+#include "aom/aomcx.h"
+#endif
+
+#if CONFIG_AV1_DECODER
+#include "aom/aomdx.h"
+#endif
+
+#if defined(_WIN32) || defined(__OS2__)
+#include <io.h>
+#include <fcntl.h>
+
+#ifdef __OS2__
+#define _setmode setmode
+#define _fileno fileno
+#define _O_BINARY O_BINARY
+#endif
+#endif
+
+#define LOG_ERROR(label)               \
+  do {                                 \
+    const char *l = label;             \
+    va_list ap;                        \
+    va_start(ap, fmt);                 \
+    if (l) fprintf(stderr, "%s: ", l); \
+    vfprintf(stderr, fmt, ap);         \
+    fprintf(stderr, "\n");             \
+    va_end(ap);                        \
+  } while (0)
+
+FILE *set_binary_mode(FILE *stream) {
+  (void)stream;
+#if defined(_WIN32) || defined(__OS2__)
+  _setmode(_fileno(stream), _O_BINARY);
+#endif
+  return stream;
+}
+
+void die(const char *fmt, ...) {
+  LOG_ERROR(NULL);
+  usage_exit();
+}
+
+void fatal(const char *fmt, ...) {
+  LOG_ERROR("Fatal");
+  exit(EXIT_FAILURE);
+}
+
+void warn(const char *fmt, ...) { LOG_ERROR("Warning"); }
+
+void die_codec(aom_codec_ctx_t *ctx, const char *s) {
+  const char *detail = aom_codec_error_detail(ctx);
+
+  printf("%s: %s\n", s, aom_codec_error(ctx));
+  if (detail) printf("    %s\n", detail);
+  exit(EXIT_FAILURE);
+}
+
+int read_yuv_frame(struct AvxInputContext *input_ctx, aom_image_t *yuv_frame) {
+  FILE *f = input_ctx->file;
+  struct FileTypeDetectionBuffer *detect = &input_ctx->detect;
+  int plane = 0;
+  int shortread = 0;
+  const int bytespp = (yuv_frame->fmt & AOM_IMG_FMT_HIGHBITDEPTH) ? 2 : 1;
+
+  for (plane = 0; plane < 3; ++plane) {
+    uint8_t *ptr;
+    const int w = aom_img_plane_width(yuv_frame, plane);
+    const int h = aom_img_plane_height(yuv_frame, plane);
+    int r;
+
+    /* Determine the correct plane based on the image format. The for-loop
+     * always counts in Y,U,V order, but this may not match the order of
+     * the data on disk.
+     */
+    switch (plane) {
+      case 1:
+        ptr =
+            yuv_frame->planes[yuv_frame->fmt == AOM_IMG_FMT_YV12 ? AOM_PLANE_V
+                                                                 : AOM_PLANE_U];
+        break;
+      case 2:
+        ptr =
+            yuv_frame->planes[yuv_frame->fmt == AOM_IMG_FMT_YV12 ? AOM_PLANE_U
+                                                                 : AOM_PLANE_V];
+        break;
+      default: ptr = yuv_frame->planes[plane];
+    }
+
+    for (r = 0; r < h; ++r) {
+      size_t needed = w * bytespp;
+      size_t buf_position = 0;
+      const size_t left = detect->buf_read - detect->position;
+      if (left > 0) {
+        const size_t more = (left < needed) ? left : needed;
+        memcpy(ptr, detect->buf + detect->position, more);
+        buf_position = more;
+        needed -= more;
+        detect->position += more;
+      }
+      if (needed > 0) {
+        shortread |= (fread(ptr + buf_position, 1, needed, f) < needed);
+      }
+
+      ptr += yuv_frame->stride[plane];
+    }
+  }
+
+  return shortread;
+}
+
+#if CONFIG_ENCODERS
+
+static const AvxInterface aom_encoders[] = {
+#if CONFIG_AV1_ENCODER
+  { "av1", AV1_FOURCC, &aom_codec_av1_cx },
+#endif
+};
+
+int get_aom_encoder_count(void) {
+  return sizeof(aom_encoders) / sizeof(aom_encoders[0]);
+}
+
+const AvxInterface *get_aom_encoder_by_index(int i) { return &aom_encoders[i]; }
+
+const AvxInterface *get_aom_encoder_by_name(const char *name) {
+  int i;
+
+  for (i = 0; i < get_aom_encoder_count(); ++i) {
+    const AvxInterface *encoder = get_aom_encoder_by_index(i);
+    if (strcmp(encoder->name, name) == 0) return encoder;
+  }
+
+  return NULL;
+}
+
+#endif  // CONFIG_ENCODERS
+
+#if CONFIG_DECODERS
+
+static const AvxInterface aom_decoders[] = {
+#if CONFIG_AV1_DECODER
+  { "av1", AV1_FOURCC, &aom_codec_av1_dx },
+#endif
+};
+
+int get_aom_decoder_count(void) {
+  return sizeof(aom_decoders) / sizeof(aom_decoders[0]);
+}
+
+const AvxInterface *get_aom_decoder_by_index(int i) { return &aom_decoders[i]; }
+
+const AvxInterface *get_aom_decoder_by_name(const char *name) {
+  int i;
+
+  for (i = 0; i < get_aom_decoder_count(); ++i) {
+    const AvxInterface *const decoder = get_aom_decoder_by_index(i);
+    if (strcmp(decoder->name, name) == 0) return decoder;
+  }
+
+  return NULL;
+}
+
+const AvxInterface *get_aom_decoder_by_fourcc(uint32_t fourcc) {
+  int i;
+
+  for (i = 0; i < get_aom_decoder_count(); ++i) {
+    const AvxInterface *const decoder = get_aom_decoder_by_index(i);
+    if (decoder->fourcc == fourcc) return decoder;
+  }
+
+  return NULL;
+}
+
+#endif  // CONFIG_DECODERS
+
+// TODO(dkovalev): move this function to aom_image.{c, h}, so it will be part
+// of aom_image_t support
+int aom_img_plane_width(const aom_image_t *img, int plane) {
+  if (plane > 0 && img->x_chroma_shift > 0)
+    return (img->d_w + 1) >> img->x_chroma_shift;
+  else
+    return img->d_w;
+}
+
+int aom_img_plane_height(const aom_image_t *img, int plane) {
+  if (plane > 0 && img->y_chroma_shift > 0)
+    return (img->d_h + 1) >> img->y_chroma_shift;
+  else
+    return img->d_h;
+}
+
+void aom_img_write(const aom_image_t *img, FILE *file) {
+  int plane;
+
+  for (plane = 0; plane < 3; ++plane) {
+    const unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int w = aom_img_plane_width(img, plane) *
+                  ((img->fmt & AOM_IMG_FMT_HIGHBITDEPTH) ? 2 : 1);
+    const int h = aom_img_plane_height(img, plane);
+    int y;
+
+    for (y = 0; y < h; ++y) {
+      fwrite(buf, 1, w, file);
+      buf += stride;
+    }
+  }
+}
+
+int aom_img_read(aom_image_t *img, FILE *file) {
+  int plane;
+
+  for (plane = 0; plane < 3; ++plane) {
+    unsigned char *buf = img->planes[plane];
+    const int stride = img->stride[plane];
+    const int w = aom_img_plane_width(img, plane) *
+                  ((img->fmt & AOM_IMG_FMT_HIGHBITDEPTH) ? 2 : 1);
+    const int h = aom_img_plane_height(img, plane);
+    int y;
+
+    for (y = 0; y < h; ++y) {
+      if (fread(buf, 1, w, file) != (size_t)w) return 0;
+      buf += stride;
+    }
+  }
+
+  return 1;
+}
+
+// TODO(dkovalev) change sse_to_psnr signature: double -> int64_t
+double sse_to_psnr(double samples, double peak, double sse) {
+  static const double kMaxPSNR = 100.0;
+
+  if (sse > 0.0) {
+    const double psnr = 10.0 * log10(samples * peak * peak / sse);
+    return psnr > kMaxPSNR ? kMaxPSNR : psnr;
+  } else {
+    return kMaxPSNR;
+  }
+}
+
+// TODO(debargha): Consolidate the functions below into a separate file.
+#if CONFIG_HIGHBITDEPTH
+static void highbd_img_upshift(aom_image_t *dst, aom_image_t *src,
+                               int input_shift) {
+  // Note the offset is 1 less than half.
+  const int offset = input_shift > 0 ? (1 << (input_shift - 1)) - 1 : 0;
+  int plane;
+  if (dst->d_w != src->d_w || dst->d_h != src->d_h ||
+      dst->x_chroma_shift != src->x_chroma_shift ||
+      dst->y_chroma_shift != src->y_chroma_shift || dst->fmt != src->fmt ||
+      input_shift < 0) {
+    fatal("Unsupported image conversion");
+  }
+  switch (src->fmt) {
+    case AOM_IMG_FMT_I42016:
+    case AOM_IMG_FMT_I42216:
+    case AOM_IMG_FMT_I44416:
+    case AOM_IMG_FMT_I44016: break;
+    default: fatal("Unsupported image conversion"); break;
+  }
+  for (plane = 0; plane < 3; plane++) {
+    int w = src->d_w;
+    int h = src->d_h;
+    int x, y;
+    if (plane) {
+      w = (w + src->x_chroma_shift) >> src->x_chroma_shift;
+      h = (h + src->y_chroma_shift) >> src->y_chroma_shift;
+    }
+    for (y = 0; y < h; y++) {
+      uint16_t *p_src =
+          (uint16_t *)(src->planes[plane] + y * src->stride[plane]);
+      uint16_t *p_dst =
+          (uint16_t *)(dst->planes[plane] + y * dst->stride[plane]);
+      for (x = 0; x < w; x++) *p_dst++ = (*p_src++ << input_shift) + offset;
+    }
+  }
+}
+
+static void lowbd_img_upshift(aom_image_t *dst, aom_image_t *src,
+                              int input_shift) {
+  // Note the offset is 1 less than half.
+  const int offset = input_shift > 0 ? (1 << (input_shift - 1)) - 1 : 0;
+  int plane;
+  if (dst->d_w != src->d_w || dst->d_h != src->d_h ||
+      dst->x_chroma_shift != src->x_chroma_shift ||
+      dst->y_chroma_shift != src->y_chroma_shift ||
+      dst->fmt != src->fmt + AOM_IMG_FMT_HIGHBITDEPTH || input_shift < 0) {
+    fatal("Unsupported image conversion");
+  }
+  switch (src->fmt) {
+    case AOM_IMG_FMT_I420:
+    case AOM_IMG_FMT_I422:
+    case AOM_IMG_FMT_I444:
+    case AOM_IMG_FMT_I440: break;
+    default: fatal("Unsupported image conversion"); break;
+  }
+  for (plane = 0; plane < 3; plane++) {
+    int w = src->d_w;
+    int h = src->d_h;
+    int x, y;
+    if (plane) {
+      w = (w + src->x_chroma_shift) >> src->x_chroma_shift;
+      h = (h + src->y_chroma_shift) >> src->y_chroma_shift;
+    }
+    for (y = 0; y < h; y++) {
+      uint8_t *p_src = src->planes[plane] + y * src->stride[plane];
+      uint16_t *p_dst =
+          (uint16_t *)(dst->planes[plane] + y * dst->stride[plane]);
+      for (x = 0; x < w; x++) {
+        *p_dst++ = (*p_src++ << input_shift) + offset;
+      }
+    }
+  }
+}
+
+void aom_img_upshift(aom_image_t *dst, aom_image_t *src, int input_shift) {
+  if (src->fmt & AOM_IMG_FMT_HIGHBITDEPTH) {
+    highbd_img_upshift(dst, src, input_shift);
+  } else {
+    lowbd_img_upshift(dst, src, input_shift);
+  }
+}
+
+void aom_img_truncate_16_to_8(aom_image_t *dst, aom_image_t *src) {
+  int plane;
+  if (dst->fmt + AOM_IMG_FMT_HIGHBITDEPTH != src->fmt || dst->d_w != src->d_w ||
+      dst->d_h != src->d_h || dst->x_chroma_shift != src->x_chroma_shift ||
+      dst->y_chroma_shift != src->y_chroma_shift) {
+    fatal("Unsupported image conversion");
+  }
+  switch (dst->fmt) {
+    case AOM_IMG_FMT_I420:
+    case AOM_IMG_FMT_I422:
+    case AOM_IMG_FMT_I444:
+    case AOM_IMG_FMT_I440: break;
+    default: fatal("Unsupported image conversion"); break;
+  }
+  for (plane = 0; plane < 3; plane++) {
+    int w = src->d_w;
+    int h = src->d_h;
+    int x, y;
+    if (plane) {
+      w = (w + src->x_chroma_shift) >> src->x_chroma_shift;
+      h = (h + src->y_chroma_shift) >> src->y_chroma_shift;
+    }
+    for (y = 0; y < h; y++) {
+      uint16_t *p_src =
+          (uint16_t *)(src->planes[plane] + y * src->stride[plane]);
+      uint8_t *p_dst = dst->planes[plane] + y * dst->stride[plane];
+      for (x = 0; x < w; x++) {
+        *p_dst++ = (uint8_t)(*p_src++);
+      }
+    }
+  }
+}
+
+static void highbd_img_downshift(aom_image_t *dst, aom_image_t *src,
+                                 int down_shift) {
+  int plane;
+  if (dst->d_w != src->d_w || dst->d_h != src->d_h ||
+      dst->x_chroma_shift != src->x_chroma_shift ||
+      dst->y_chroma_shift != src->y_chroma_shift || dst->fmt != src->fmt ||
+      down_shift < 0) {
+    fatal("Unsupported image conversion");
+  }
+  switch (src->fmt) {
+    case AOM_IMG_FMT_I42016:
+    case AOM_IMG_FMT_I42216:
+    case AOM_IMG_FMT_I44416:
+    case AOM_IMG_FMT_I44016: break;
+    default: fatal("Unsupported image conversion"); break;
+  }
+  for (plane = 0; plane < 3; plane++) {
+    int w = src->d_w;
+    int h = src->d_h;
+    int x, y;
+    if (plane) {
+      w = (w + src->x_chroma_shift) >> src->x_chroma_shift;
+      h = (h + src->y_chroma_shift) >> src->y_chroma_shift;
+    }
+    for (y = 0; y < h; y++) {
+      uint16_t *p_src =
+          (uint16_t *)(src->planes[plane] + y * src->stride[plane]);
+      uint16_t *p_dst =
+          (uint16_t *)(dst->planes[plane] + y * dst->stride[plane]);
+      for (x = 0; x < w; x++) *p_dst++ = *p_src++ >> down_shift;
+    }
+  }
+}
+
+static void lowbd_img_downshift(aom_image_t *dst, aom_image_t *src,
+                                int down_shift) {
+  int plane;
+  if (dst->d_w != src->d_w || dst->d_h != src->d_h ||
+      dst->x_chroma_shift != src->x_chroma_shift ||
+      dst->y_chroma_shift != src->y_chroma_shift ||
+      src->fmt != dst->fmt + AOM_IMG_FMT_HIGHBITDEPTH || down_shift < 0) {
+    fatal("Unsupported image conversion");
+  }
+  switch (dst->fmt) {
+    case AOM_IMG_FMT_I420:
+    case AOM_IMG_FMT_I422:
+    case AOM_IMG_FMT_I444:
+    case AOM_IMG_FMT_I440: break;
+    default: fatal("Unsupported image conversion"); break;
+  }
+  for (plane = 0; plane < 3; plane++) {
+    int w = src->d_w;
+    int h = src->d_h;
+    int x, y;
+    if (plane) {
+      w = (w + src->x_chroma_shift) >> src->x_chroma_shift;
+      h = (h + src->y_chroma_shift) >> src->y_chroma_shift;
+    }
+    for (y = 0; y < h; y++) {
+      uint16_t *p_src =
+          (uint16_t *)(src->planes[plane] + y * src->stride[plane]);
+      uint8_t *p_dst = dst->planes[plane] + y * dst->stride[plane];
+      for (x = 0; x < w; x++) {
+        *p_dst++ = *p_src++ >> down_shift;
+      }
+    }
+  }
+}
+
+void aom_img_downshift(aom_image_t *dst, aom_image_t *src, int down_shift) {
+  if (dst->fmt & AOM_IMG_FMT_HIGHBITDEPTH) {
+    highbd_img_downshift(dst, src, down_shift);
+  } else {
+    lowbd_img_downshift(dst, src, down_shift);
+  }
+}
+#endif  // CONFIG_HIGHBITDEPTH
diff --git a/third_party/aom/tools_common.h b/third_party/aom/tools_common.h
new file mode 100644
index 0000000000..8acb3709f3
--- /dev/null
+++ b/third_party/aom/tools_common.h
@@ -0,0 +1,168 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef TOOLS_COMMON_H_
+#define TOOLS_COMMON_H_
+
+#include <stdio.h>
+
+#include "./aom_config.h"
+#include "aom/aom_codec.h"
+#include "aom/aom_image.h"
+#include "aom/aom_integer.h"
+#include "aom_ports/msvc.h"
+
+#if CONFIG_ENCODERS
+#include "./y4minput.h"
+#endif
+
+#if defined(_MSC_VER)
+/* MSVS uses _f{seek,tell}i64. */
+#define fseeko _fseeki64
+#define ftello _ftelli64
+typedef int64_t FileOffset;
+#elif defined(_WIN32)
+#include <sys/types.h> /* NOLINT*/
+/* MinGW uses f{seek,tell}o64 for large files. */
+#define fseeko fseeko64
+#define ftello ftello64
+typedef off64_t FileOffset;
+#elif CONFIG_OS_SUPPORT
+#include <sys/types.h> /* NOLINT*/
+typedef off_t FileOffset;
+/* Use 32-bit file operations in WebM file format when building ARM
+ * executables (.axf) with RVCT. */
+#else
+#define fseeko fseek
+#define ftello ftell
+typedef long FileOffset; /* NOLINT */
+#endif /* CONFIG_OS_SUPPORT */
+
+#if CONFIG_OS_SUPPORT
+#if defined(_MSC_VER)
+#include <io.h> /* NOLINT */
+#define isatty _isatty
+#define fileno _fileno
+#else
+#include <unistd.h> /* NOLINT */
+#endif              /* _MSC_VER */
+#endif              /* CONFIG_OS_SUPPORT */
+
+#define LITERALU64(hi, lo) ((((uint64_t)hi) << 32) | lo)
+
+#ifndef PATH_MAX
+#define PATH_MAX 512
+#endif
+
+#define IVF_FRAME_HDR_SZ (4 + 8) /* 4 byte size + 8 byte timestamp */
+#define IVF_FILE_HDR_SZ 32
+
+#define RAW_FRAME_HDR_SZ sizeof(uint32_t)
+
+#define AV1_FOURCC 0x31305641
+
+enum VideoFileType {
+  FILE_TYPE_RAW,
+  FILE_TYPE_IVF,
+  FILE_TYPE_Y4M,
+  FILE_TYPE_WEBM
+};
+
+struct FileTypeDetectionBuffer {
+  char buf[4];
+  size_t buf_read;
+  size_t position;
+};
+
+struct AvxRational {
+  int numerator;
+  int denominator;
+};
+
+struct AvxInputContext {
+  const char *filename;
+  FILE *file;
+  int64_t length;
+  struct FileTypeDetectionBuffer detect;
+  enum VideoFileType file_type;
+  uint32_t width;
+  uint32_t height;
+  struct AvxRational pixel_aspect_ratio;
+  aom_img_fmt_t fmt;
+  aom_bit_depth_t bit_depth;
+  int only_i420;
+  uint32_t fourcc;
+  struct AvxRational framerate;
+#if CONFIG_ENCODERS
+  y4m_input y4m;
+#endif
+};
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(__GNUC__)
+#define AOM_NO_RETURN __attribute__((noreturn))
+#else
+#define AOM_NO_RETURN
+#endif
+
+/* Sets a stdio stream into binary mode */
+FILE *set_binary_mode(FILE *stream);
+
+void die(const char *fmt, ...) AOM_NO_RETURN;
+void fatal(const char *fmt, ...) AOM_NO_RETURN;
+void warn(const char *fmt, ...);
+
+void die_codec(aom_codec_ctx_t *ctx, const char *s) AOM_NO_RETURN;
+
+/* The tool including this file must define usage_exit() */
+void usage_exit(void) AOM_NO_RETURN;
+
+#undef AOM_NO_RETURN
+
+int read_yuv_frame(struct AvxInputContext *input_ctx, aom_image_t *yuv_frame);
+
+typedef struct AvxInterface {
+  const char *const name;
+  const uint32_t fourcc;
+  aom_codec_iface_t *(*const codec_interface)();
+} AvxInterface;
+
+int get_aom_encoder_count(void);
+const AvxInterface *get_aom_encoder_by_index(int i);
+const AvxInterface *get_aom_encoder_by_name(const char *name);
+
+int get_aom_decoder_count(void);
+const AvxInterface *get_aom_decoder_by_index(int i);
+const AvxInterface *get_aom_decoder_by_name(const char *name);
+const AvxInterface *get_aom_decoder_by_fourcc(uint32_t fourcc);
+
+// TODO(dkovalev): move this function to aom_image.{c, h}, so it will be part
+// of aom_image_t support
+int aom_img_plane_width(const aom_image_t *img, int plane);
+int aom_img_plane_height(const aom_image_t *img, int plane);
+void aom_img_write(const aom_image_t *img, FILE *file);
+int aom_img_read(aom_image_t *img, FILE *file);
+
+double sse_to_psnr(double samples, double peak, double mse);
+
+#if CONFIG_HIGHBITDEPTH
+void aom_img_upshift(aom_image_t *dst, aom_image_t *src, int input_shift);
+void aom_img_downshift(aom_image_t *dst, aom_image_t *src, int down_shift);
+void aom_img_truncate_16_to_8(aom_image_t *dst, aom_image_t *src);
+#endif
+
+#ifdef __cplusplus
+} /* extern "C" */
+#endif
+
+#endif  // TOOLS_COMMON_H_
diff --git a/third_party/aom/usage.dox b/third_party/aom/usage.dox
new file mode 100644
index 0000000000..59239e8f18
--- /dev/null
+++ b/third_party/aom/usage.dox
@@ -0,0 +1,135 @@
+/*!\page usage Usage
+
+    The aom multi-format codec SDK provides a unified interface amongst its
+    supported codecs. This abstraction allows applications using this SDK to
+    easily support multiple video formats with minimal code duplication or
+    "special casing." This section describes the interface common to all codecs.
+    For codec-specific details, see the \ref codecs page.
+
+    The following sections are common to all codecs:
+    - \ref usage_types
+    - \ref usage_features
+    - \ref usage_init
+    - \ref usage_errors
+
+    For more information on decoder and encoder specific usage, see the
+    following pages:
+    \if decoder
+    \li \subpage usage_decode
+    \endif
+    \if encoder
+    \li \subpage usage_encode
+    \endif
+
+    \section usage_types Important Data Types
+    There are two important data structures to consider in this interface.
+
+    \subsection usage_ctxs Contexts
+    A context is a storage area allocated by the calling application that the
+    codec may write into to store details about a single instance of that codec.
+    Most of the context is implementation specific, and thus opaque to the
+    application. The context structure as seen by the application is of fixed
+    size, and thus can be allocated with automatic storage or dynamically
+    on the heap.
+
+    Most operations require an initialized codec context. Codec context
+    instances are codec specific. That is, the codec to be used for the encoded
+    video must be known at initialization time. See #aom_codec_ctx_t for further
+    information.
+
+    \subsection usage_ifaces Interfaces
+    A codec interface is an opaque structure that controls how function calls
+    into the generic interface are dispatched to their codec-specific
+    implementations. Applications \ref MUSTNOT attempt to examine or override
+    this storage, as it contains internal implementation details likely to
+    change from release to release.
+
+    Each supported codec will expose an interface structure to the application
+    as an <code>extern</code> reference to a structure of the incomplete type
+    #aom_codec_iface_t.
+
+    \section usage_features Features
+    Several "features" are defined that are optionally implemented by codec
+    algorithms. Indeed, the same algorithm may support different features on
+    different platforms. The purpose of defining these features is that when
+    they are implemented, they conform to a common interface. The features, or
+    capabilities, of an algorithm can be queried from it's interface by using
+    the aom_codec_get_caps() method. Attempts to invoke features not supported
+    by an algorithm will generally result in #AOM_CODEC_INCAPABLE.
+
+    \if decoder
+    Currently defined decoder features include:
+    - \ref usage_cb
+    \endif
+
+    \section usage_init Initialization
+    To initialize a codec instance, the address of the codec context
+    and interface structures are passed to an initialization function. Depending
+    on the \ref usage_features that the codec supports, the codec could be
+    initialized in different modes.
+
+    To prevent cases of confusion where the ABI of the library changes,
+    the ABI is versioned. The ABI version number must be passed at
+    initialization time to ensure the application is using a header file that
+    matches the library. The current ABI version number is stored in the
+    preprocessor macros #AOM_CODEC_ABI_VERSION, #AOM_ENCODER_ABI_VERSION, and
+    #AOM_DECODER_ABI_VERSION. For convenience, each initialization function has
+    a wrapper macro that inserts the correct version number. These macros are
+    named like the initialization methods, but without the _ver suffix.
+
+
+    The available initialization methods are:
+    \if encoder
+    \li #aom_codec_enc_init (calls aom_codec_enc_init_ver())
+    \li #aom_codec_enc_init_multi (calls aom_codec_enc_init_multi_ver())
+    \endif
+    \if decoder
+    \li #aom_codec_dec_init (calls aom_codec_dec_init_ver())
+    \endif
+
+
+    \section usage_errors Error Handling
+    Almost all codec functions return an error status of type #aom_codec_err_t.
+    The semantics of how each error condition should be processed is clearly
+    defined in the definitions of each enumerated value. Error values can be
+    converted into ASCII strings with the aom_codec_error() and
+    aom_codec_err_to_string() methods. The difference between these two methods is
+    that aom_codec_error() returns the error state from an initialized context,
+    whereas aom_codec_err_to_string() can be used in cases where an error occurs
+    outside any context. The enumerated value returned from the last call can be
+    retrieved from the <code>err</code> member of the decoder context as well.
+    Finally, more detailed error information may be able to be obtained by using
+    the aom_codec_error_detail() method. Not all errors produce detailed error
+    information.
+
+    In addition to error information, the codec library's build configuration
+    is available at runtime on some platforms. This information can be returned
+    by calling aom_codec_build_config(), and is formatted as a base64 coded string
+    (comprised of characters in the set [a-z_a-Z0-9+/]). This information is not
+    useful to an application at runtime, but may be of use to aom for support.
+
+
+    \section usage_deadline Deadline
+    Both the encoding and decoding functions have a <code>deadline</code>
+    parameter. This parameter indicates the amount of time, in microseconds
+    (us), that the application wants the codec to spend processing before
+    returning. This is a soft deadline -- that is, the semantics of the
+    requested operation take precedence over meeting the deadline. If, for
+    example, an application sets a <code>deadline</code> of 1000us, and the
+    frame takes 2000us to decode, the call to aom_codec_decode() will return
+    after 2000us. In this case the deadline is not met, but the semantics of the
+    function are preserved. If, for the same frame, an application instead sets
+    a <code>deadline</code> of 5000us, the decoder will see that it has 3000us
+    remaining in its time slice when decoding completes. It could then choose to
+    run a set of \ref usage_postproc filters, and perhaps would return after
+    4000us (instead of the allocated 5000us). In this case the deadline is met,
+    and the semantics of the call are preserved, as before.
+
+    The special value <code>0</code> is reserved to represent an infinite
+    deadline. In this case, the codec will perform as much processing as
+    possible to yield the highest quality frame.
+
+    By convention, the value <code>1</code> is used to mean "return as fast as
+    possible."
+
+*/
diff --git a/third_party/aom/usage_cx.dox b/third_party/aom/usage_cx.dox
new file mode 100644
index 0000000000..dcf267ce43
--- /dev/null
+++ b/third_party/aom/usage_cx.dox
@@ -0,0 +1,13 @@
+/*! \page usage_encode Encoding
+
+    The aom_codec_encode() function is at the core of the encode loop. It
+    processes raw images passed by the application, producing packets of
+    compressed data. The <code>deadline</code> parameter controls the amount
+    of time in microseconds the encoder should spend working on the frame. For
+    more information on the <code>deadline</code> parameter, see
+    \ref usage_deadline.
+
+
+    \ref samples
+
+*/
diff --git a/third_party/aom/usage_dx.dox b/third_party/aom/usage_dx.dox
new file mode 100644
index 0000000000..6b76bf7b00
--- /dev/null
+++ b/third_party/aom/usage_dx.dox
@@ -0,0 +1,62 @@
+/*! \page usage_decode Decoding
+
+    The aom_codec_decode() function is at the core of the decode loop. It
+    processes packets of compressed data passed by the application, producing
+    decoded images. The decoder expects packets to comprise exactly one image
+    frame of data. Packets \ref MUST be passed in decode order. If the
+    application wishes to associate some data with the frame, the
+    <code>user_priv</code> member may be set. The <code>deadline</code>
+    parameter controls the amount of time in microseconds the decoder should
+    spend working on the frame. This is typically used to support adaptive
+    \ref usage_postproc based on the amount of free CPU time. For more
+    information on the <code>deadline</code> parameter, see \ref usage_deadline.
+
+    \ref samples
+
+
+    \section usage_cb Callback Based Decoding
+    There are two methods for the application to access decoded frame data. Some
+    codecs support asynchronous (callback-based) decoding \ref usage_features
+    that allow the application to register a callback to be invoked by the
+    decoder when decoded data becomes available. Decoders are not required to
+    support this feature, however. Like all \ref usage_features, support can be
+    determined by calling aom_codec_get_caps(). Callbacks are available in both
+    frame-based and slice-based variants. Frame based callbacks conform to the
+    signature of #aom_codec_put_frame_cb_fn_t and are invoked once the entire
+    frame has been decoded. Slice based callbacks conform to the signature of
+    #aom_codec_put_slice_cb_fn_t and are invoked after a subsection of the frame
+    is decoded. For example, a slice callback could be issued for each
+    macroblock row. However, the number and size of slices to return is
+    implementation specific. Also, the image data passed in a slice callback is
+    not necessarily in the same memory segment as the data will be when it is
+    assembled into a full frame. For this reason, the application \ref MUST
+    examine the rectangles that describe what data is valid to access and what
+    data has been updated in this call. For all their additional complexity,
+    slice based decoding callbacks provide substantial speed gains to the
+    overall application in some cases, due to improved cache behavior.
+
+
+    \section usage_frame_iter Frame Iterator Based Decoding
+    If the codec does not support callback based decoding, or the application
+    chooses not to make use of that feature, decoded frames are made available
+    through the aom_codec_get_frame() iterator. The application initializes the
+    iterator storage (of type #aom_codec_iter_t) to NULL, then calls
+    aom_codec_get_frame repeatedly until it returns NULL, indicating that all
+    images have been returned. This process may result in zero, one, or many
+    frames that are ready for display, depending on the codec.
+
+
+    \section usage_postproc Postprocessing
+    Postprocessing is a process that is applied after a frame is decoded to
+    enhance the image's appearance by removing artifacts introduced in the
+    compression process. It is not required to properly decode the frame, and
+    is generally done only when there is enough spare CPU time to execute
+    the required filters. Codecs may support a number of different
+    postprocessing filters, and the available filters may differ from platform
+    to platform. Embedded devices often do not have enough CPU to implement
+    postprocessing in software. The filter selection is generally handled
+    automatically by the codec, depending on the amount of time remaining before
+    hitting the user-specified \ref usage_deadline after decoding the frame.
+
+
+*/
diff --git a/third_party/aom/video_common.h b/third_party/aom/video_common.h
new file mode 100644
index 0000000000..1aa82f613a
--- /dev/null
+++ b/third_party/aom/video_common.h
@@ -0,0 +1,24 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef VIDEO_COMMON_H_
+#define VIDEO_COMMON_H_
+
+#include "./tools_common.h"
+
+typedef struct {
+  uint32_t codec_fourcc;
+  int frame_width;
+  int frame_height;
+  struct AvxRational time_base;
+} AvxVideoInfo;
+
+#endif  // VIDEO_COMMON_H_
diff --git a/third_party/aom/video_reader.c b/third_party/aom/video_reader.c
new file mode 100644
index 0000000000..6a96af967e
--- /dev/null
+++ b/third_party/aom/video_reader.c
@@ -0,0 +1,78 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+
+#include "./ivfdec.h"
+#include "./video_reader.h"
+
+#include "aom_ports/mem_ops.h"
+
+static const char *const kIVFSignature = "DKIF";
+
+struct AvxVideoReaderStruct {
+  AvxVideoInfo info;
+  FILE *file;
+  uint8_t *buffer;
+  size_t buffer_size;
+  size_t frame_size;
+};
+
+AvxVideoReader *aom_video_reader_open(const char *filename) {
+  char header[32];
+  AvxVideoReader *reader = NULL;
+  FILE *const file = fopen(filename, "rb");
+  if (!file) return NULL;  // Can't open file
+
+  if (fread(header, 1, 32, file) != 32) return NULL;  // Can't read file header
+
+  if (memcmp(kIVFSignature, header, 4) != 0)
+    return NULL;  // Wrong IVF signature
+
+  if (mem_get_le16(header + 4) != 0) return NULL;  // Wrong IVF version
+
+  reader = (AvxVideoReader *)calloc(1, sizeof(*reader));
+  if (!reader) return NULL;  // Can't allocate AvxVideoReader
+
+  reader->file = file;
+  reader->info.codec_fourcc = mem_get_le32(header + 8);
+  reader->info.frame_width = mem_get_le16(header + 12);
+  reader->info.frame_height = mem_get_le16(header + 14);
+  reader->info.time_base.numerator = mem_get_le32(header + 16);
+  reader->info.time_base.denominator = mem_get_le32(header + 20);
+
+  return reader;
+}
+
+void aom_video_reader_close(AvxVideoReader *reader) {
+  if (reader) {
+    fclose(reader->file);
+    free(reader->buffer);
+    free(reader);
+  }
+}
+
+int aom_video_reader_read_frame(AvxVideoReader *reader) {
+  return !ivf_read_frame(reader->file, &reader->buffer, &reader->frame_size,
+                         &reader->buffer_size);
+}
+
+const uint8_t *aom_video_reader_get_frame(AvxVideoReader *reader,
+                                          size_t *size) {
+  if (size) *size = reader->frame_size;
+
+  return reader->buffer;
+}
+
+const AvxVideoInfo *aom_video_reader_get_info(AvxVideoReader *reader) {
+  return &reader->info;
+}
diff --git a/third_party/aom/video_reader.h b/third_party/aom/video_reader.h
new file mode 100644
index 0000000000..962c6653b5
--- /dev/null
+++ b/third_party/aom/video_reader.h
@@ -0,0 +1,52 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef VIDEO_READER_H_
+#define VIDEO_READER_H_
+
+#include "./video_common.h"
+
+// The following code is work in progress. It is going to  support transparent
+// reading of input files. Right now only IVF format is supported for
+// simplicity. The main goal the API is to be simple and easy to use in example
+// code and in aomenc/aomdec later. All low-level details like memory
+// buffer management are hidden from API users.
+struct AvxVideoReaderStruct;
+typedef struct AvxVideoReaderStruct AvxVideoReader;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Opens the input file for reading and inspects it to determine file type.
+// Returns an opaque AvxVideoReader* upon success, or NULL upon failure.
+// Right now only IVF format is supported.
+AvxVideoReader *aom_video_reader_open(const char *filename);
+
+// Frees all resources associated with AvxVideoReader* returned from
+// aom_video_reader_open() call.
+void aom_video_reader_close(AvxVideoReader *reader);
+
+// Reads frame from the file and stores it in internal buffer.
+int aom_video_reader_read_frame(AvxVideoReader *reader);
+
+// Returns the pointer to memory buffer with frame data read by last call to
+// aom_video_reader_read_frame().
+const uint8_t *aom_video_reader_get_frame(AvxVideoReader *reader, size_t *size);
+
+// Fills AvxVideoInfo with information from opened video file.
+const AvxVideoInfo *aom_video_reader_get_info(AvxVideoReader *reader);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VIDEO_READER_H_
diff --git a/third_party/aom/video_writer.c b/third_party/aom/video_writer.c
new file mode 100644
index 0000000000..4e072c7dcc
--- /dev/null
+++ b/third_party/aom/video_writer.c
@@ -0,0 +1,77 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <stdlib.h>
+
+#include "./ivfenc.h"
+#include "./video_writer.h"
+#include "aom/aom_encoder.h"
+
+struct AvxVideoWriterStruct {
+  AvxVideoInfo info;
+  FILE *file;
+  int frame_count;
+};
+
+static void write_header(FILE *file, const AvxVideoInfo *info,
+                         int frame_count) {
+  struct aom_codec_enc_cfg cfg;
+  cfg.g_w = info->frame_width;
+  cfg.g_h = info->frame_height;
+  cfg.g_timebase.num = info->time_base.numerator;
+  cfg.g_timebase.den = info->time_base.denominator;
+
+  ivf_write_file_header(file, &cfg, info->codec_fourcc, frame_count);
+}
+
+AvxVideoWriter *aom_video_writer_open(const char *filename,
+                                      AvxContainer container,
+                                      const AvxVideoInfo *info) {
+  if (container == kContainerIVF) {
+    AvxVideoWriter *writer = NULL;
+    FILE *const file = fopen(filename, "wb");
+    if (!file) return NULL;
+
+    writer = malloc(sizeof(*writer));
+    if (!writer) return NULL;
+
+    writer->frame_count = 0;
+    writer->info = *info;
+    writer->file = file;
+
+    write_header(writer->file, info, 0);
+
+    return writer;
+  }
+
+  return NULL;
+}
+
+void aom_video_writer_close(AvxVideoWriter *writer) {
+  if (writer) {
+    // Rewriting frame header with real frame count
+    rewind(writer->file);
+    write_header(writer->file, &writer->info, writer->frame_count);
+
+    fclose(writer->file);
+    free(writer);
+  }
+}
+
+int aom_video_writer_write_frame(AvxVideoWriter *writer, const uint8_t *buffer,
+                                 size_t size, int64_t pts) {
+  ivf_write_frame_header(writer->file, pts, size);
+  if (fwrite(buffer, 1, size, writer->file) != size) return 0;
+
+  ++writer->frame_count;
+
+  return 1;
+}
diff --git a/third_party/aom/video_writer.h b/third_party/aom/video_writer.h
new file mode 100644
index 0000000000..ad9e6ebb98
--- /dev/null
+++ b/third_party/aom/video_writer.h
@@ -0,0 +1,45 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef VIDEO_WRITER_H_
+#define VIDEO_WRITER_H_
+
+#include "./video_common.h"
+
+typedef enum { kContainerIVF } AvxContainer;
+
+struct AvxVideoWriterStruct;
+typedef struct AvxVideoWriterStruct AvxVideoWriter;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Finds and opens writer for specified container format.
+// Returns an opaque AvxVideoWriter* upon success, or NULL upon failure.
+// Right now only IVF format is supported.
+AvxVideoWriter *aom_video_writer_open(const char *filename,
+                                      AvxContainer container,
+                                      const AvxVideoInfo *info);
+
+// Frees all resources associated with AvxVideoWriter* returned from
+// aom_video_writer_open() call.
+void aom_video_writer_close(AvxVideoWriter *writer);
+
+// Writes frame bytes to the file.
+int aom_video_writer_write_frame(AvxVideoWriter *writer, const uint8_t *buffer,
+                                 size_t size, int64_t pts);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VIDEO_WRITER_H_
diff --git a/third_party/aom/warnings.c b/third_party/aom/warnings.c
new file mode 100644
index 0000000000..16d5c6c18a
--- /dev/null
+++ b/third_party/aom/warnings.c
@@ -0,0 +1,98 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include "./warnings.h"
+
+#include <assert.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aom_encoder.h"
+
+#include "./tools_common.h"
+#include "./aomenc.h"
+
+static const char quantizer_warning_string[] =
+    "Bad quantizer values. Quantizer values should not be equal, and should "
+    "differ by at least 8.";
+
+struct WarningListNode {
+  const char *warning_string;
+  struct WarningListNode *next_warning;
+};
+
+struct WarningList {
+  struct WarningListNode *warning_node;
+};
+
+static void add_warning(const char *warning_string,
+                        struct WarningList *warning_list) {
+  struct WarningListNode **node = &warning_list->warning_node;
+
+  struct WarningListNode *new_node = malloc(sizeof(*new_node));
+  if (new_node == NULL) {
+    fatal("Unable to allocate warning node.");
+  }
+
+  new_node->warning_string = warning_string;
+  new_node->next_warning = NULL;
+
+  while (*node != NULL) node = &(*node)->next_warning;
+
+  *node = new_node;
+}
+
+static void free_warning_list(struct WarningList *warning_list) {
+  while (warning_list->warning_node != NULL) {
+    struct WarningListNode *const node = warning_list->warning_node;
+    warning_list->warning_node = node->next_warning;
+    free(node);
+  }
+}
+
+static int continue_prompt(int num_warnings) {
+  int c;
+  fprintf(stderr,
+          "%d encoder configuration warning(s). Continue? (y to continue) ",
+          num_warnings);
+  c = getchar();
+  return c == 'y';
+}
+
+static void check_quantizer(int min_q, int max_q,
+                            struct WarningList *warning_list) {
+  const int lossless = min_q == 0 && max_q == 0;
+  if (!lossless && (min_q == max_q || abs(max_q - min_q) < 8))
+    add_warning(quantizer_warning_string, warning_list);
+}
+
+void check_encoder_config(int disable_prompt,
+                          const struct AvxEncoderConfig *global_config,
+                          const struct aom_codec_enc_cfg *stream_config) {
+  int num_warnings = 0;
+  struct WarningListNode *warning = NULL;
+  struct WarningList warning_list = { 0 };
+  (void)global_config;
+  check_quantizer(stream_config->rc_min_quantizer,
+                  stream_config->rc_max_quantizer, &warning_list);
+  /* Count and print warnings. */
+  for (warning = warning_list.warning_node; warning != NULL;
+       warning = warning->next_warning, ++num_warnings) {
+    warn(warning->warning_string);
+  }
+
+  free_warning_list(&warning_list);
+
+  if (num_warnings) {
+    if (!disable_prompt && !continue_prompt(num_warnings)) exit(EXIT_FAILURE);
+  }
+}
diff --git a/third_party/aom/warnings.h b/third_party/aom/warnings.h
new file mode 100644
index 0000000000..61db2dcf8a
--- /dev/null
+++ b/third_party/aom/warnings.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef WARNINGS_H_
+#define WARNINGS_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct aom_codec_enc_cfg;
+struct AvxEncoderConfig;
+
+/*
+ * Checks config for improperly used settings. Warns user upon encountering
+ * settings that will lead to poor output quality. Prompts user to continue
+ * when warnings are issued.
+ */
+void check_encoder_config(int disable_prompt,
+                          const struct AvxEncoderConfig *global_config,
+                          const struct aom_codec_enc_cfg *stream_config);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // WARNINGS_H_
diff --git a/third_party/aom/webmdec.cc b/third_party/aom/webmdec.cc
new file mode 100644
index 0000000000..39ecef7063
--- /dev/null
+++ b/third_party/aom/webmdec.cc
@@ -0,0 +1,223 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "./webmdec.h"
+
+#include <cstring>
+#include <cstdio>
+
+#include "third_party/libwebm/mkvparser/mkvparser.h"
+#include "third_party/libwebm/mkvparser/mkvreader.h"
+
+namespace {
+
+void reset(struct WebmInputContext *const webm_ctx) {
+  if (webm_ctx->reader != NULL) {
+    mkvparser::MkvReader *const reader =
+        reinterpret_cast<mkvparser::MkvReader *>(webm_ctx->reader);
+    delete reader;
+  }
+  if (webm_ctx->segment != NULL) {
+    mkvparser::Segment *const segment =
+        reinterpret_cast<mkvparser::Segment *>(webm_ctx->segment);
+    delete segment;
+  }
+  if (webm_ctx->buffer != NULL) {
+    delete[] webm_ctx->buffer;
+  }
+  webm_ctx->reader = NULL;
+  webm_ctx->segment = NULL;
+  webm_ctx->buffer = NULL;
+  webm_ctx->cluster = NULL;
+  webm_ctx->block_entry = NULL;
+  webm_ctx->block = NULL;
+  webm_ctx->block_frame_index = 0;
+  webm_ctx->video_track_index = 0;
+  webm_ctx->timestamp_ns = 0;
+  webm_ctx->is_key_frame = false;
+}
+
+void get_first_cluster(struct WebmInputContext *const webm_ctx) {
+  mkvparser::Segment *const segment =
+      reinterpret_cast<mkvparser::Segment *>(webm_ctx->segment);
+  const mkvparser::Cluster *const cluster = segment->GetFirst();
+  webm_ctx->cluster = cluster;
+}
+
+void rewind_and_reset(struct WebmInputContext *const webm_ctx,
+                      struct AvxInputContext *const aom_ctx) {
+  rewind(aom_ctx->file);
+  reset(webm_ctx);
+}
+
+}  // namespace
+
+int file_is_webm(struct WebmInputContext *webm_ctx,
+                 struct AvxInputContext *aom_ctx) {
+  mkvparser::MkvReader *const reader = new mkvparser::MkvReader(aom_ctx->file);
+  webm_ctx->reader = reader;
+  webm_ctx->reached_eos = 0;
+
+  mkvparser::EBMLHeader header;
+  long long pos = 0;
+  if (header.Parse(reader, pos) < 0) {
+    rewind_and_reset(webm_ctx, aom_ctx);
+    return 0;
+  }
+
+  mkvparser::Segment *segment;
+  if (mkvparser::Segment::CreateInstance(reader, pos, segment)) {
+    rewind_and_reset(webm_ctx, aom_ctx);
+    return 0;
+  }
+  webm_ctx->segment = segment;
+  if (segment->Load() < 0) {
+    rewind_and_reset(webm_ctx, aom_ctx);
+    return 0;
+  }
+
+  const mkvparser::Tracks *const tracks = segment->GetTracks();
+  const mkvparser::VideoTrack *video_track = NULL;
+  for (unsigned long i = 0; i < tracks->GetTracksCount(); ++i) {
+    const mkvparser::Track *const track = tracks->GetTrackByIndex(i);
+    if (track->GetType() == mkvparser::Track::kVideo) {
+      video_track = static_cast<const mkvparser::VideoTrack *>(track);
+      webm_ctx->video_track_index = static_cast<int>(track->GetNumber());
+      break;
+    }
+  }
+
+  if (video_track == NULL || video_track->GetCodecId() == NULL) {
+    rewind_and_reset(webm_ctx, aom_ctx);
+    return 0;
+  }
+
+  if (!strncmp(video_track->GetCodecId(), "V_AV1", 5)) {
+    aom_ctx->fourcc = AV1_FOURCC;
+  } else {
+    rewind_and_reset(webm_ctx, aom_ctx);
+    return 0;
+  }
+
+  aom_ctx->framerate.denominator = 0;
+  aom_ctx->framerate.numerator = 0;
+  aom_ctx->width = static_cast<uint32_t>(video_track->GetWidth());
+  aom_ctx->height = static_cast<uint32_t>(video_track->GetHeight());
+
+  get_first_cluster(webm_ctx);
+
+  return 1;
+}
+
+int webm_read_frame(struct WebmInputContext *webm_ctx, uint8_t **buffer,
+                    size_t *buffer_size) {
+  // This check is needed for frame parallel decoding, in which case this
+  // function could be called even after it has reached end of input stream.
+  if (webm_ctx->reached_eos) {
+    return 1;
+  }
+  mkvparser::Segment *const segment =
+      reinterpret_cast<mkvparser::Segment *>(webm_ctx->segment);
+  const mkvparser::Cluster *cluster =
+      reinterpret_cast<const mkvparser::Cluster *>(webm_ctx->cluster);
+  const mkvparser::Block *block =
+      reinterpret_cast<const mkvparser::Block *>(webm_ctx->block);
+  const mkvparser::BlockEntry *block_entry =
+      reinterpret_cast<const mkvparser::BlockEntry *>(webm_ctx->block_entry);
+  bool block_entry_eos = false;
+  do {
+    long status = 0;
+    bool get_new_block = false;
+    if (block_entry == NULL && !block_entry_eos) {
+      status = cluster->GetFirst(block_entry);
+      get_new_block = true;
+    } else if (block_entry_eos || block_entry->EOS()) {
+      cluster = segment->GetNext(cluster);
+      if (cluster == NULL || cluster->EOS()) {
+        *buffer_size = 0;
+        webm_ctx->reached_eos = 1;
+        return 1;
+      }
+      status = cluster->GetFirst(block_entry);
+      block_entry_eos = false;
+      get_new_block = true;
+    } else if (block == NULL ||
+               webm_ctx->block_frame_index == block->GetFrameCount() ||
+               block->GetTrackNumber() != webm_ctx->video_track_index) {
+      status = cluster->GetNext(block_entry, block_entry);
+      if (block_entry == NULL || block_entry->EOS()) {
+        block_entry_eos = true;
+        continue;
+      }
+      get_new_block = true;
+    }
+    if (status || block_entry == NULL) {
+      return -1;
+    }
+    if (get_new_block) {
+      block = block_entry->GetBlock();
+      if (block == NULL) return -1;
+      webm_ctx->block_frame_index = 0;
+    }
+  } while (block_entry_eos ||
+           block->GetTrackNumber() != webm_ctx->video_track_index);
+
+  webm_ctx->cluster = cluster;
+  webm_ctx->block_entry = block_entry;
+  webm_ctx->block = block;
+
+  const mkvparser::Block::Frame &frame =
+      block->GetFrame(webm_ctx->block_frame_index);
+  ++webm_ctx->block_frame_index;
+  if (frame.len > static_cast<long>(*buffer_size)) {
+    delete[] * buffer;
+    *buffer = new uint8_t[frame.len];
+    if (*buffer == NULL) {
+      return -1;
+    }
+    webm_ctx->buffer = *buffer;
+  }
+  *buffer_size = frame.len;
+  webm_ctx->timestamp_ns = block->GetTime(cluster);
+  webm_ctx->is_key_frame = block->IsKey();
+
+  mkvparser::MkvReader *const reader =
+      reinterpret_cast<mkvparser::MkvReader *>(webm_ctx->reader);
+  return frame.Read(reader, *buffer) ? -1 : 0;
+}
+
+int webm_guess_framerate(struct WebmInputContext *webm_ctx,
+                         struct AvxInputContext *aom_ctx) {
+  uint32_t i = 0;
+  uint8_t *buffer = NULL;
+  size_t buffer_size = 0;
+  while (webm_ctx->timestamp_ns < 1000000000 && i < 50) {
+    if (webm_read_frame(webm_ctx, &buffer, &buffer_size)) {
+      break;
+    }
+    ++i;
+  }
+  aom_ctx->framerate.numerator = (i - 1) * 1000000;
+  aom_ctx->framerate.denominator =
+      static_cast<int>(webm_ctx->timestamp_ns / 1000);
+  delete[] buffer;
+
+  get_first_cluster(webm_ctx);
+  webm_ctx->block = NULL;
+  webm_ctx->block_entry = NULL;
+  webm_ctx->block_frame_index = 0;
+  webm_ctx->timestamp_ns = 0;
+  webm_ctx->reached_eos = 0;
+
+  return 0;
+}
+
+void webm_free(struct WebmInputContext *webm_ctx) { reset(webm_ctx); }
diff --git a/third_party/aom/webmdec.h b/third_party/aom/webmdec.h
new file mode 100644
index 0000000000..329908eebf
--- /dev/null
+++ b/third_party/aom/webmdec.h
@@ -0,0 +1,70 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef WEBMDEC_H_
+#define WEBMDEC_H_
+
+#include "./tools_common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct AvxInputContext;
+
+struct WebmInputContext {
+  void *reader;
+  void *segment;
+  uint8_t *buffer;
+  const void *cluster;
+  const void *block_entry;
+  const void *block;
+  int block_frame_index;
+  int video_track_index;
+  uint64_t timestamp_ns;
+  int is_key_frame;
+  int reached_eos;
+};
+
+// Checks if the input is a WebM file. If so, initializes WebMInputContext so
+// that webm_read_frame can be called to retrieve a video frame.
+// Returns 1 on success and 0 on failure or input is not WebM file.
+// TODO(vigneshv): Refactor this function into two smaller functions specific
+// to their task.
+int file_is_webm(struct WebmInputContext *webm_ctx,
+                 struct AvxInputContext *aom_ctx);
+
+// Reads a WebM Video Frame. Memory for the buffer is created, owned and managed
+// by this function. For the first call, |buffer| should be NULL and
+// |*buffer_size| should be 0. Once all the frames are read and used,
+// webm_free() should be called, otherwise there will be a leak.
+// Parameters:
+//      webm_ctx - WebmInputContext object
+//      buffer - pointer where the frame data will be filled.
+//      buffer_size - pointer to buffer size.
+// Return values:
+//      0 - Success
+//      1 - End of Stream
+//     -1 - Error
+int webm_read_frame(struct WebmInputContext *webm_ctx, uint8_t **buffer,
+                    size_t *buffer_size);
+
+// Guesses the frame rate of the input file based on the container timestamps.
+int webm_guess_framerate(struct WebmInputContext *webm_ctx,
+                         struct AvxInputContext *aom_ctx);
+
+// Resets the WebMInputContext.
+void webm_free(struct WebmInputContext *webm_ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // WEBMDEC_H_
diff --git a/third_party/aom/webmenc.cc b/third_party/aom/webmenc.cc
new file mode 100644
index 0000000000..e3d209a276
--- /dev/null
+++ b/third_party/aom/webmenc.cc
@@ -0,0 +1,96 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+*/
+
+#include "./webmenc.h"
+
+#include <string>
+
+#include "third_party/libwebm/mkvmuxer/mkvmuxer.h"
+#include "third_party/libwebm/mkvmuxer/mkvmuxerutil.h"
+#include "third_party/libwebm/mkvmuxer/mkvwriter.h"
+
+namespace {
+const uint64_t kDebugTrackUid = 0xDEADBEEF;
+const int kVideoTrackNumber = 1;
+}  // namespace
+
+void write_webm_file_header(struct WebmOutputContext *webm_ctx,
+                            const aom_codec_enc_cfg_t *cfg,
+                            stereo_format_t stereo_fmt, unsigned int fourcc,
+                            const struct AvxRational *par) {
+  mkvmuxer::MkvWriter *const writer = new mkvmuxer::MkvWriter(webm_ctx->stream);
+  mkvmuxer::Segment *const segment = new mkvmuxer::Segment();
+  segment->Init(writer);
+  segment->set_mode(mkvmuxer::Segment::kFile);
+  segment->OutputCues(true);
+
+  mkvmuxer::SegmentInfo *const info = segment->GetSegmentInfo();
+  const uint64_t kTimecodeScale = 1000000;
+  info->set_timecode_scale(kTimecodeScale);
+  std::string version = "aomenc";
+  if (!webm_ctx->debug) {
+    version.append(std::string(" ") + aom_codec_version_str());
+  }
+  info->set_writing_app(version.c_str());
+
+  const uint64_t video_track_id =
+      segment->AddVideoTrack(static_cast<int>(cfg->g_w),
+                             static_cast<int>(cfg->g_h), kVideoTrackNumber);
+  mkvmuxer::VideoTrack *const video_track = static_cast<mkvmuxer::VideoTrack *>(
+      segment->GetTrackByNumber(video_track_id));
+  video_track->SetStereoMode(stereo_fmt);
+  const char *codec_id;
+  switch (fourcc) {
+    case AV1_FOURCC: codec_id = "V_AV1"; break;
+    default: codec_id = "V_AV1"; break;
+  }
+  video_track->set_codec_id(codec_id);
+  if (par->numerator > 1 || par->denominator > 1) {
+    // TODO(fgalligan): Add support of DisplayUnit, Display Aspect Ratio type
+    // to WebM format.
+    const uint64_t display_width = static_cast<uint64_t>(
+        ((cfg->g_w * par->numerator * 1.0) / par->denominator) + .5);
+    video_track->set_display_width(display_width);
+    video_track->set_display_height(cfg->g_h);
+  }
+  if (webm_ctx->debug) {
+    video_track->set_uid(kDebugTrackUid);
+  }
+  webm_ctx->writer = writer;
+  webm_ctx->segment = segment;
+}
+
+void write_webm_block(struct WebmOutputContext *webm_ctx,
+                      const aom_codec_enc_cfg_t *cfg,
+                      const aom_codec_cx_pkt_t *pkt) {
+  mkvmuxer::Segment *const segment =
+      reinterpret_cast<mkvmuxer::Segment *>(webm_ctx->segment);
+  int64_t pts_ns = pkt->data.frame.pts * 1000000000ll * cfg->g_timebase.num /
+                   cfg->g_timebase.den;
+  if (pts_ns <= webm_ctx->last_pts_ns) pts_ns = webm_ctx->last_pts_ns + 1000000;
+  webm_ctx->last_pts_ns = pts_ns;
+
+  segment->AddFrame(static_cast<uint8_t *>(pkt->data.frame.buf),
+                    pkt->data.frame.sz, kVideoTrackNumber, pts_ns,
+                    pkt->data.frame.flags & AOM_FRAME_IS_KEY);
+}
+
+void write_webm_file_footer(struct WebmOutputContext *webm_ctx) {
+  mkvmuxer::MkvWriter *const writer =
+      reinterpret_cast<mkvmuxer::MkvWriter *>(webm_ctx->writer);
+  mkvmuxer::Segment *const segment =
+      reinterpret_cast<mkvmuxer::Segment *>(webm_ctx->segment);
+  segment->Finalize();
+  delete segment;
+  delete writer;
+  webm_ctx->writer = NULL;
+  webm_ctx->segment = NULL;
+}
diff --git a/third_party/aom/webmenc.h b/third_party/aom/webmenc.h
new file mode 100644
index 0000000000..74387fb8d1
--- /dev/null
+++ b/third_party/aom/webmenc.h
@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#ifndef WEBMENC_H_
+#define WEBMENC_H_
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "tools_common.h"
+#include "aom/aom_encoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct WebmOutputContext {
+  int debug;
+  FILE *stream;
+  int64_t last_pts_ns;
+  void *writer;
+  void *segment;
+};
+
+/* Stereo 3D packed frame format */
+typedef enum stereo_format {
+  STEREO_FORMAT_MONO = 0,
+  STEREO_FORMAT_LEFT_RIGHT = 1,
+  STEREO_FORMAT_BOTTOM_TOP = 2,
+  STEREO_FORMAT_TOP_BOTTOM = 3,
+  STEREO_FORMAT_RIGHT_LEFT = 11
+} stereo_format_t;
+
+void write_webm_file_header(struct WebmOutputContext *webm_ctx,
+                            const aom_codec_enc_cfg_t *cfg,
+                            stereo_format_t stereo_fmt, unsigned int fourcc,
+                            const struct AvxRational *par);
+
+void write_webm_block(struct WebmOutputContext *webm_ctx,
+                      const aom_codec_enc_cfg_t *cfg,
+                      const aom_codec_cx_pkt_t *pkt);
+
+void write_webm_file_footer(struct WebmOutputContext *webm_ctx);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // WEBMENC_H_
diff --git a/third_party/aom/y4menc.c b/third_party/aom/y4menc.c
new file mode 100644
index 0000000000..b094f74fed
--- /dev/null
+++ b/third_party/aom/y4menc.c
@@ -0,0 +1,65 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include "./y4menc.h"
+
+int y4m_write_file_header(char *buf, size_t len, int width, int height,
+                          const struct AvxRational *framerate,
+                          aom_img_fmt_t fmt, unsigned int bit_depth) {
+  const char *color;
+  switch (bit_depth) {
+    case 8:
+      color = fmt == AOM_IMG_FMT_444A
+                  ? "C444alpha\n"
+                  : fmt == AOM_IMG_FMT_I444
+                        ? "C444\n"
+                        : fmt == AOM_IMG_FMT_I422 ? "C422\n" : "C420jpeg\n";
+      break;
+    case 9:
+      color = fmt == AOM_IMG_FMT_I44416
+                  ? "C444p9 XYSCSS=444P9\n"
+                  : fmt == AOM_IMG_FMT_I42216 ? "C422p9 XYSCSS=422P9\n"
+                                              : "C420p9 XYSCSS=420P9\n";
+      break;
+    case 10:
+      color = fmt == AOM_IMG_FMT_I44416
+                  ? "C444p10 XYSCSS=444P10\n"
+                  : fmt == AOM_IMG_FMT_I42216 ? "C422p10 XYSCSS=422P10\n"
+                                              : "C420p10 XYSCSS=420P10\n";
+      break;
+    case 12:
+      color = fmt == AOM_IMG_FMT_I44416
+                  ? "C444p12 XYSCSS=444P12\n"
+                  : fmt == AOM_IMG_FMT_I42216 ? "C422p12 XYSCSS=422P12\n"
+                                              : "C420p12 XYSCSS=420P12\n";
+      break;
+    case 14:
+      color = fmt == AOM_IMG_FMT_I44416
+                  ? "C444p14 XYSCSS=444P14\n"
+                  : fmt == AOM_IMG_FMT_I42216 ? "C422p14 XYSCSS=422P14\n"
+                                              : "C420p14 XYSCSS=420P14\n";
+      break;
+    case 16:
+      color = fmt == AOM_IMG_FMT_I44416
+                  ? "C444p16 XYSCSS=444P16\n"
+                  : fmt == AOM_IMG_FMT_I42216 ? "C422p16 XYSCSS=422P16\n"
+                                              : "C420p16 XYSCSS=420P16\n";
+      break;
+    default: color = NULL; assert(0);
+  }
+  return snprintf(buf, len, "YUV4MPEG2 W%u H%u F%u:%u I%c %s", width, height,
+                  framerate->numerator, framerate->denominator, 'p', color);
+}
+
+int y4m_write_frame_header(char *buf, size_t len) {
+  return snprintf(buf, len, "FRAME\n");
+}
diff --git a/third_party/aom/y4menc.h b/third_party/aom/y4menc.h
new file mode 100644
index 0000000000..cb75eeb42e
--- /dev/null
+++ b/third_party/aom/y4menc.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#ifndef Y4MENC_H_
+#define Y4MENC_H_
+
+#include "./tools_common.h"
+
+#include "aom/aom_decoder.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define Y4M_BUFFER_SIZE 128
+
+int y4m_write_file_header(char *buf, size_t len, int width, int height,
+                          const struct AvxRational *framerate,
+                          aom_img_fmt_t fmt, unsigned int bit_depth);
+int y4m_write_frame_header(char *buf, size_t len);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // Y4MENC_H_
diff --git a/third_party/aom/y4minput.c b/third_party/aom/y4minput.c
new file mode 100644
index 0000000000..1919189245
--- /dev/null
+++ b/third_party/aom/y4minput.c
@@ -0,0 +1,1127 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ *
+ * Based on code from the OggTheora software codec source code,
+ * Copyright (C) 2002-2010 The Xiph.Org Foundation and contributors.
+ */
+#include <errno.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include "aom/aom_integer.h"
+#include "y4minput.h"
+
+// Reads 'size' bytes from 'file' into 'buf' with some fault tolerance.
+// Returns true on success.
+static int file_read(void *buf, size_t size, FILE *file) {
+  const int kMaxRetries = 5;
+  int retry_count = 0;
+  int file_error;
+  size_t len = 0;
+  do {
+    const size_t n = fread((uint8_t *)buf + len, 1, size - len, file);
+    len += n;
+    file_error = ferror(file);
+    if (file_error) {
+      if (errno == EINTR || errno == EAGAIN) {
+        clearerr(file);
+        continue;
+      } else {
+        fprintf(stderr, "Error reading file: %u of %u bytes read, %d: %s\n",
+                (uint32_t)len, (uint32_t)size, errno, strerror(errno));
+        return 0;
+      }
+    }
+  } while (!feof(file) && len < size && ++retry_count < kMaxRetries);
+
+  if (!feof(file) && len != size) {
+    fprintf(stderr,
+            "Error reading file: %u of %u bytes read,"
+            " error: %d, retries: %d, %d: %s\n",
+            (uint32_t)len, (uint32_t)size, file_error, retry_count, errno,
+            strerror(errno));
+  }
+  return len == size;
+}
+
+static int y4m_parse_tags(y4m_input *_y4m, char *_tags) {
+  int got_w;
+  int got_h;
+  int got_fps;
+  int got_interlace;
+  int got_par;
+  int got_chroma;
+  char *p;
+  char *q;
+  got_w = got_h = got_fps = got_interlace = got_par = got_chroma = 0;
+  for (p = _tags;; p = q) {
+    /*Skip any leading spaces.*/
+    while (*p == ' ') p++;
+    /*If that's all we have, stop.*/
+    if (p[0] == '\0') break;
+    /*Find the end of this tag.*/
+    for (q = p + 1; *q != '\0' && *q != ' '; q++) {
+    }
+    /*Process the tag.*/
+    switch (p[0]) {
+      case 'W': {
+        if (sscanf(p + 1, "%d", &_y4m->pic_w) != 1) return -1;
+        got_w = 1;
+      } break;
+      case 'H': {
+        if (sscanf(p + 1, "%d", &_y4m->pic_h) != 1) return -1;
+        got_h = 1;
+      } break;
+      case 'F': {
+        if (sscanf(p + 1, "%d:%d", &_y4m->fps_n, &_y4m->fps_d) != 2) {
+          return -1;
+        }
+        got_fps = 1;
+      } break;
+      case 'I': {
+        _y4m->interlace = p[1];
+        got_interlace = 1;
+      } break;
+      case 'A': {
+        if (sscanf(p + 1, "%d:%d", &_y4m->par_n, &_y4m->par_d) != 2) {
+          return -1;
+        }
+        got_par = 1;
+      } break;
+      case 'C': {
+        if (q - p > 16) return -1;
+        memcpy(_y4m->chroma_type, p + 1, q - p - 1);
+        _y4m->chroma_type[q - p - 1] = '\0';
+        got_chroma = 1;
+      } break;
+        /*Ignore unknown tags.*/
+    }
+  }
+  if (!got_w || !got_h || !got_fps) return -1;
+  if (!got_interlace) _y4m->interlace = '?';
+  if (!got_par) _y4m->par_n = _y4m->par_d = 0;
+  /*Chroma-type is not specified in older files, e.g., those generated by
+     mplayer.*/
+  if (!got_chroma) strcpy(_y4m->chroma_type, "420");
+  return 0;
+}
+
+/*All anti-aliasing filters in the following conversion functions are based on
+   one of two window functions:
+  The 6-tap Lanczos window (for down-sampling and shifts):
+   sinc(\pi*t)*sinc(\pi*t/3), |t|<3  (sinc(t)==sin(t)/t)
+   0,                         |t|>=3
+  The 4-tap Mitchell window (for up-sampling):
+   7|t|^3-12|t|^2+16/3,             |t|<1
+   -(7/3)|x|^3+12|x|^2-20|x|+32/3,  |t|<2
+   0,                               |t|>=2
+  The number of taps is intentionally kept small to reduce computational
+   overhead and limit ringing.
+
+  The taps from these filters are scaled so that their sum is 1, and the result
+   is scaled by 128 and rounded to integers to create a filter whose
+   intermediate values fit inside 16 bits.
+  Coefficients are rounded in such a way as to ensure their sum is still 128,
+   which is usually equivalent to normal rounding.
+
+  Conversions which require both horizontal and vertical filtering could
+   have these steps pipelined, for less memory consumption and better cache
+   performance, but we do them separately for simplicity.*/
+
+#define OC_MINI(_a, _b) ((_a) > (_b) ? (_b) : (_a))
+#define OC_MAXI(_a, _b) ((_a) < (_b) ? (_b) : (_a))
+#define OC_CLAMPI(_a, _b, _c) (OC_MAXI(_a, OC_MINI(_b, _c)))
+
+/*420jpeg chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  420mpeg2 chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  BR      |       BR      |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  BR      |       BR      |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  We use a resampling filter to shift the site locations one quarter pixel (at
+   the chroma plane's resolution) to the right.
+  The 4:2:2 modes look exactly the same, except there are twice as many chroma
+   lines, and they are vertically co-sited with the luma samples in both the
+   mpeg2 and jpeg cases (thus requiring no vertical resampling).*/
+static void y4m_42xmpeg2_42xjpeg_helper(unsigned char *_dst,
+                                        const unsigned char *_src, int _c_w,
+                                        int _c_h) {
+  int y;
+  int x;
+  for (y = 0; y < _c_h; y++) {
+    /*Filter: [4 -17 114 35 -9 1]/128, derived from a 6-tap Lanczos
+       window.*/
+    for (x = 0; x < OC_MINI(_c_w, 2); x++) {
+      _dst[x] = (unsigned char)OC_CLAMPI(
+          0, (4 * _src[0] - 17 * _src[OC_MAXI(x - 1, 0)] + 114 * _src[x] +
+              35 * _src[OC_MINI(x + 1, _c_w - 1)] -
+              9 * _src[OC_MINI(x + 2, _c_w - 1)] +
+              _src[OC_MINI(x + 3, _c_w - 1)] + 64) >>
+                 7,
+          255);
+    }
+    for (; x < _c_w - 3; x++) {
+      _dst[x] = (unsigned char)OC_CLAMPI(
+          0, (4 * _src[x - 2] - 17 * _src[x - 1] + 114 * _src[x] +
+              35 * _src[x + 1] - 9 * _src[x + 2] + _src[x + 3] + 64) >>
+                 7,
+          255);
+    }
+    for (; x < _c_w; x++) {
+      _dst[x] = (unsigned char)OC_CLAMPI(
+          0, (4 * _src[x - 2] - 17 * _src[x - 1] + 114 * _src[x] +
+              35 * _src[OC_MINI(x + 1, _c_w - 1)] -
+              9 * _src[OC_MINI(x + 2, _c_w - 1)] + _src[_c_w - 1] + 64) >>
+                 7,
+          255);
+    }
+    _dst += _c_w;
+    _src += _c_w;
+  }
+}
+
+/*Handles both 422 and 420mpeg2 to 422jpeg and 420jpeg, respectively.*/
+static void y4m_convert_42xmpeg2_42xjpeg(y4m_input *_y4m, unsigned char *_dst,
+                                         unsigned char *_aux) {
+  int c_w;
+  int c_h;
+  int c_sz;
+  int pli;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  for (pli = 1; pli < 3; pli++) {
+    y4m_42xmpeg2_42xjpeg_helper(_dst, _aux, c_w, c_h);
+    _dst += c_sz;
+    _aux += c_sz;
+  }
+}
+
+/*This format is only used for interlaced content, but is included for
+   completeness.
+
+  420jpeg chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  420paldv chroma samples are sited like:
+  YR------Y-------YR------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YB------Y-------YB------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YR------Y-------YR------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YB------Y-------YB------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  We use a resampling filter to shift the site locations one quarter pixel (at
+   the chroma plane's resolution) to the right.
+  Then we use another filter to move the C_r location down one quarter pixel,
+   and the C_b location up one quarter pixel.*/
+static void y4m_convert_42xpaldv_42xjpeg(y4m_input *_y4m, unsigned char *_dst,
+                                         unsigned char *_aux) {
+  unsigned char *tmp;
+  int c_w;
+  int c_h;
+  int c_sz;
+  int pli;
+  int y;
+  int x;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + 1) / 2;
+  c_h = (_y4m->pic_h + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  c_sz = c_w * c_h;
+  tmp = _aux + 2 * c_sz;
+  for (pli = 1; pli < 3; pli++) {
+    /*First do the horizontal re-sampling.
+      This is the same as the mpeg2 case, except that after the horizontal
+       case, we need to apply a second vertical filter.*/
+    y4m_42xmpeg2_42xjpeg_helper(tmp, _aux, c_w, c_h);
+    _aux += c_sz;
+    switch (pli) {
+      case 1: {
+        /*Slide C_b up a quarter-pel.
+          This is the same filter used above, but in the other order.*/
+        for (x = 0; x < c_w; x++) {
+          for (y = 0; y < OC_MINI(c_h, 3); y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(
+                0, (tmp[0] - 9 * tmp[OC_MAXI(y - 2, 0) * c_w] +
+                    35 * tmp[OC_MAXI(y - 1, 0) * c_w] + 114 * tmp[y * c_w] -
+                    17 * tmp[OC_MINI(y + 1, c_h - 1) * c_w] +
+                    4 * tmp[OC_MINI(y + 2, c_h - 1) * c_w] + 64) >>
+                       7,
+                255);
+          }
+          for (; y < c_h - 2; y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(
+                0, (tmp[(y - 3) * c_w] - 9 * tmp[(y - 2) * c_w] +
+                    35 * tmp[(y - 1) * c_w] + 114 * tmp[y * c_w] -
+                    17 * tmp[(y + 1) * c_w] + 4 * tmp[(y + 2) * c_w] + 64) >>
+                       7,
+                255);
+          }
+          for (; y < c_h; y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(
+                0, (tmp[(y - 3) * c_w] - 9 * tmp[(y - 2) * c_w] +
+                    35 * tmp[(y - 1) * c_w] + 114 * tmp[y * c_w] -
+                    17 * tmp[OC_MINI(y + 1, c_h - 1) * c_w] +
+                    4 * tmp[(c_h - 1) * c_w] + 64) >>
+                       7,
+                255);
+          }
+          _dst++;
+          tmp++;
+        }
+        _dst += c_sz - c_w;
+        tmp -= c_w;
+      } break;
+      case 2: {
+        /*Slide C_r down a quarter-pel.
+          This is the same as the horizontal filter.*/
+        for (x = 0; x < c_w; x++) {
+          for (y = 0; y < OC_MINI(c_h, 2); y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(
+                0,
+                (4 * tmp[0] - 17 * tmp[OC_MAXI(y - 1, 0) * c_w] +
+                 114 * tmp[y * c_w] + 35 * tmp[OC_MINI(y + 1, c_h - 1) * c_w] -
+                 9 * tmp[OC_MINI(y + 2, c_h - 1) * c_w] +
+                 tmp[OC_MINI(y + 3, c_h - 1) * c_w] + 64) >>
+                    7,
+                255);
+          }
+          for (; y < c_h - 3; y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(
+                0, (4 * tmp[(y - 2) * c_w] - 17 * tmp[(y - 1) * c_w] +
+                    114 * tmp[y * c_w] + 35 * tmp[(y + 1) * c_w] -
+                    9 * tmp[(y + 2) * c_w] + tmp[(y + 3) * c_w] + 64) >>
+                       7,
+                255);
+          }
+          for (; y < c_h; y++) {
+            _dst[y * c_w] = (unsigned char)OC_CLAMPI(
+                0,
+                (4 * tmp[(y - 2) * c_w] - 17 * tmp[(y - 1) * c_w] +
+                 114 * tmp[y * c_w] + 35 * tmp[OC_MINI(y + 1, c_h - 1) * c_w] -
+                 9 * tmp[OC_MINI(y + 2, c_h - 1) * c_w] + tmp[(c_h - 1) * c_w] +
+                 64) >>
+                    7,
+                255);
+          }
+          _dst++;
+          tmp++;
+        }
+      } break;
+    }
+    /*For actual interlaced material, this would have to be done separately on
+       each field, and the shift amounts would be different.
+      C_r moves down 1/8, C_b up 3/8 in the top field, and C_r moves down 3/8,
+       C_b up 1/8 in the bottom field.
+      The corresponding filters would be:
+       Down 1/8 (reverse order for up): [3 -11 125 15 -4 0]/128
+       Down 3/8 (reverse order for up): [4 -19 98 56 -13 2]/128*/
+  }
+}
+
+/*Perform vertical filtering to reduce a single plane from 4:2:2 to 4:2:0.
+  This is used as a helper by several converation routines.*/
+static void y4m_422jpeg_420jpeg_helper(unsigned char *_dst,
+                                       const unsigned char *_src, int _c_w,
+                                       int _c_h) {
+  int y;
+  int x;
+  /*Filter: [3 -17 78 78 -17 3]/128, derived from a 6-tap Lanczos window.*/
+  for (x = 0; x < _c_w; x++) {
+    for (y = 0; y < OC_MINI(_c_h, 2); y += 2) {
+      _dst[(y >> 1) * _c_w] =
+          OC_CLAMPI(0, (64 * _src[0] + 78 * _src[OC_MINI(1, _c_h - 1) * _c_w] -
+                        17 * _src[OC_MINI(2, _c_h - 1) * _c_w] +
+                        3 * _src[OC_MINI(3, _c_h - 1) * _c_w] + 64) >>
+                           7,
+                    255);
+    }
+    for (; y < _c_h - 3; y += 2) {
+      _dst[(y >> 1) * _c_w] =
+          OC_CLAMPI(0, (3 * (_src[(y - 2) * _c_w] + _src[(y + 3) * _c_w]) -
+                        17 * (_src[(y - 1) * _c_w] + _src[(y + 2) * _c_w]) +
+                        78 * (_src[y * _c_w] + _src[(y + 1) * _c_w]) + 64) >>
+                           7,
+                    255);
+    }
+    for (; y < _c_h; y += 2) {
+      _dst[(y >> 1) * _c_w] = OC_CLAMPI(
+          0,
+          (3 * (_src[(y - 2) * _c_w] + _src[(_c_h - 1) * _c_w]) -
+           17 * (_src[(y - 1) * _c_w] + _src[OC_MINI(y + 2, _c_h - 1) * _c_w]) +
+           78 * (_src[y * _c_w] + _src[OC_MINI(y + 1, _c_h - 1) * _c_w]) +
+           64) >>
+              7,
+          255);
+    }
+    _src++;
+    _dst++;
+  }
+}
+
+/*420jpeg chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  422jpeg chroma samples are sited like:
+  Y---BR--Y-------Y---BR--Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y---BR--Y-------Y---BR--Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y---BR--Y-------Y---BR--Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y---BR--Y-------Y---BR--Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  We use a resampling filter to decimate the chroma planes by two in the
+   vertical direction.*/
+static void y4m_convert_422jpeg_420jpeg(y4m_input *_y4m, unsigned char *_dst,
+                                        unsigned char *_aux) {
+  int c_w;
+  int c_h;
+  int c_sz;
+  int dst_c_w;
+  int dst_c_h;
+  int dst_c_sz;
+  int pli;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + _y4m->src_c_dec_h - 1) / _y4m->src_c_dec_h;
+  c_h = _y4m->pic_h;
+  dst_c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  dst_c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  dst_c_sz = dst_c_w * dst_c_h;
+  for (pli = 1; pli < 3; pli++) {
+    y4m_422jpeg_420jpeg_helper(_dst, _aux, c_w, c_h);
+    _aux += c_sz;
+    _dst += dst_c_sz;
+  }
+}
+
+/*420jpeg chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  422 chroma samples are sited like:
+  YBR-----Y-------YBR-----Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------YBR-----Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------YBR-----Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------YBR-----Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  We use a resampling filter to shift the original site locations one quarter
+   pixel (at the original chroma resolution) to the right.
+  Then we use a second resampling filter to decimate the chroma planes by two
+   in the vertical direction.*/
+static void y4m_convert_422_420jpeg(y4m_input *_y4m, unsigned char *_dst,
+                                    unsigned char *_aux) {
+  unsigned char *tmp;
+  int c_w;
+  int c_h;
+  int c_sz;
+  int dst_c_h;
+  int dst_c_sz;
+  int pli;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + _y4m->src_c_dec_h - 1) / _y4m->src_c_dec_h;
+  c_h = _y4m->pic_h;
+  dst_c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  dst_c_sz = c_w * dst_c_h;
+  tmp = _aux + 2 * c_sz;
+  for (pli = 1; pli < 3; pli++) {
+    /*In reality, the horizontal and vertical steps could be pipelined, for
+       less memory consumption and better cache performance, but we do them
+       separately for simplicity.*/
+    /*First do horizontal filtering (convert to 422jpeg)*/
+    y4m_42xmpeg2_42xjpeg_helper(tmp, _aux, c_w, c_h);
+    /*Now do the vertical filtering.*/
+    y4m_422jpeg_420jpeg_helper(_dst, tmp, c_w, c_h);
+    _aux += c_sz;
+    _dst += dst_c_sz;
+  }
+}
+
+/*420jpeg chroma samples are sited like:
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |   BR  |       |   BR  |
+  |       |       |       |
+  Y-------Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  411 chroma samples are sited like:
+  YBR-----Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+  YBR-----Y-------Y-------Y-------
+  |       |       |       |
+  |       |       |       |
+  |       |       |       |
+
+  We use a filter to resample at site locations one eighth pixel (at the source
+   chroma plane's horizontal resolution) and five eighths of a pixel to the
+   right.
+  Then we use another filter to decimate the planes by 2 in the vertical
+   direction.*/
+static void y4m_convert_411_420jpeg(y4m_input *_y4m, unsigned char *_dst,
+                                    unsigned char *_aux) {
+  unsigned char *tmp;
+  int c_w;
+  int c_h;
+  int c_sz;
+  int dst_c_w;
+  int dst_c_h;
+  int dst_c_sz;
+  int tmp_sz;
+  int pli;
+  int y;
+  int x;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + _y4m->src_c_dec_h - 1) / _y4m->src_c_dec_h;
+  c_h = _y4m->pic_h;
+  dst_c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  dst_c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  dst_c_sz = dst_c_w * dst_c_h;
+  tmp_sz = dst_c_w * c_h;
+  tmp = _aux + 2 * c_sz;
+  for (pli = 1; pli < 3; pli++) {
+    /*In reality, the horizontal and vertical steps could be pipelined, for
+       less memory consumption and better cache performance, but we do them
+       separately for simplicity.*/
+    /*First do horizontal filtering (convert to 422jpeg)*/
+    for (y = 0; y < c_h; y++) {
+      /*Filters: [1 110 18 -1]/128 and [-3 50 86 -5]/128, both derived from a
+         4-tap Mitchell window.*/
+      for (x = 0; x < OC_MINI(c_w, 1); x++) {
+        tmp[x << 1] = (unsigned char)OC_CLAMPI(
+            0, (111 * _aux[0] + 18 * _aux[OC_MINI(1, c_w - 1)] -
+                _aux[OC_MINI(2, c_w - 1)] + 64) >>
+                   7,
+            255);
+        tmp[x << 1 | 1] = (unsigned char)OC_CLAMPI(
+            0, (47 * _aux[0] + 86 * _aux[OC_MINI(1, c_w - 1)] -
+                5 * _aux[OC_MINI(2, c_w - 1)] + 64) >>
+                   7,
+            255);
+      }
+      for (; x < c_w - 2; x++) {
+        tmp[x << 1] =
+            (unsigned char)OC_CLAMPI(0, (_aux[x - 1] + 110 * _aux[x] +
+                                         18 * _aux[x + 1] - _aux[x + 2] + 64) >>
+                                            7,
+                                     255);
+        tmp[x << 1 | 1] = (unsigned char)OC_CLAMPI(
+            0, (-3 * _aux[x - 1] + 50 * _aux[x] + 86 * _aux[x + 1] -
+                5 * _aux[x + 2] + 64) >>
+                   7,
+            255);
+      }
+      for (; x < c_w; x++) {
+        tmp[x << 1] = (unsigned char)OC_CLAMPI(
+            0, (_aux[x - 1] + 110 * _aux[x] +
+                18 * _aux[OC_MINI(x + 1, c_w - 1)] - _aux[c_w - 1] + 64) >>
+                   7,
+            255);
+        if ((x << 1 | 1) < dst_c_w) {
+          tmp[x << 1 | 1] = (unsigned char)OC_CLAMPI(
+              0,
+              (-3 * _aux[x - 1] + 50 * _aux[x] +
+               86 * _aux[OC_MINI(x + 1, c_w - 1)] - 5 * _aux[c_w - 1] + 64) >>
+                  7,
+              255);
+        }
+      }
+      tmp += dst_c_w;
+      _aux += c_w;
+    }
+    tmp -= tmp_sz;
+    /*Now do the vertical filtering.*/
+    y4m_422jpeg_420jpeg_helper(_dst, tmp, dst_c_w, c_h);
+    _dst += dst_c_sz;
+  }
+}
+
+/*Convert 444 to 420jpeg.*/
+static void y4m_convert_444_420jpeg(y4m_input *_y4m, unsigned char *_dst,
+                                    unsigned char *_aux) {
+  unsigned char *tmp;
+  int c_w;
+  int c_h;
+  int c_sz;
+  int dst_c_w;
+  int dst_c_h;
+  int dst_c_sz;
+  int tmp_sz;
+  int pli;
+  int y;
+  int x;
+  /*Skip past the luma data.*/
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  /*Compute the size of each chroma plane.*/
+  c_w = (_y4m->pic_w + _y4m->src_c_dec_h - 1) / _y4m->src_c_dec_h;
+  c_h = _y4m->pic_h;
+  dst_c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  dst_c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  dst_c_sz = dst_c_w * dst_c_h;
+  tmp_sz = dst_c_w * c_h;
+  tmp = _aux + 2 * c_sz;
+  for (pli = 1; pli < 3; pli++) {
+    /*Filter: [3 -17 78 78 -17 3]/128, derived from a 6-tap Lanczos window.*/
+    for (y = 0; y < c_h; y++) {
+      for (x = 0; x < OC_MINI(c_w, 2); x += 2) {
+        tmp[x >> 1] =
+            OC_CLAMPI(0, (64 * _aux[0] + 78 * _aux[OC_MINI(1, c_w - 1)] -
+                          17 * _aux[OC_MINI(2, c_w - 1)] +
+                          3 * _aux[OC_MINI(3, c_w - 1)] + 64) >>
+                             7,
+                      255);
+      }
+      for (; x < c_w - 3; x += 2) {
+        tmp[x >> 1] = OC_CLAMPI(0, (3 * (_aux[x - 2] + _aux[x + 3]) -
+                                    17 * (_aux[x - 1] + _aux[x + 2]) +
+                                    78 * (_aux[x] + _aux[x + 1]) + 64) >>
+                                       7,
+                                255);
+      }
+      for (; x < c_w; x += 2) {
+        tmp[x >> 1] = OC_CLAMPI(
+            0, (3 * (_aux[x - 2] + _aux[c_w - 1]) -
+                17 * (_aux[x - 1] + _aux[OC_MINI(x + 2, c_w - 1)]) +
+                78 * (_aux[x] + _aux[OC_MINI(x + 1, c_w - 1)]) + 64) >>
+                   7,
+            255);
+      }
+      tmp += dst_c_w;
+      _aux += c_w;
+    }
+    tmp -= tmp_sz;
+    /*Now do the vertical filtering.*/
+    y4m_422jpeg_420jpeg_helper(_dst, tmp, dst_c_w, c_h);
+    _dst += dst_c_sz;
+  }
+}
+
+/*The image is padded with empty chroma components at 4:2:0.*/
+static void y4m_convert_mono_420jpeg(y4m_input *_y4m, unsigned char *_dst,
+                                     unsigned char *_aux) {
+  int c_sz;
+  (void)_aux;
+  _dst += _y4m->pic_w * _y4m->pic_h;
+  c_sz = ((_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h) *
+         ((_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v);
+  memset(_dst, 128, c_sz * 2);
+}
+
+/*No conversion function needed.*/
+static void y4m_convert_null(y4m_input *_y4m, unsigned char *_dst,
+                             unsigned char *_aux) {
+  (void)_y4m;
+  (void)_dst;
+  (void)_aux;
+}
+
+int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
+                   int only_420) {
+  char buffer[80] = { 0 };
+  int ret;
+  int i;
+  /*Read until newline, or 80 cols, whichever happens first.*/
+  for (i = 0; i < 79; i++) {
+    if (_nskip > 0) {
+      buffer[i] = *_skip++;
+      _nskip--;
+    } else {
+      if (!file_read(buffer + i, 1, _fin)) return -1;
+    }
+    if (buffer[i] == '\n') break;
+  }
+  /*We skipped too much header data.*/
+  if (_nskip > 0) return -1;
+  if (i == 79) {
+    fprintf(stderr, "Error parsing header; not a YUV2MPEG2 file?\n");
+    return -1;
+  }
+  buffer[i] = '\0';
+  if (memcmp(buffer, "YUV4MPEG", 8)) {
+    fprintf(stderr, "Incomplete magic for YUV4MPEG file.\n");
+    return -1;
+  }
+  if (buffer[8] != '2') {
+    fprintf(stderr, "Incorrect YUV input file version; YUV4MPEG2 required.\n");
+  }
+  ret = y4m_parse_tags(_y4m, buffer + 5);
+  if (ret < 0) {
+    fprintf(stderr, "Error parsing YUV4MPEG2 header.\n");
+    return ret;
+  }
+  if (_y4m->interlace == '?') {
+    fprintf(stderr,
+            "Warning: Input video interlacing format unknown; "
+            "assuming progressive scan.\n");
+  } else if (_y4m->interlace != 'p') {
+    fprintf(stderr,
+            "Input video is interlaced; "
+            "Only progressive scan handled.\n");
+    return -1;
+  }
+  _y4m->aom_fmt = AOM_IMG_FMT_I420;
+  _y4m->bps = 12;
+  _y4m->bit_depth = 8;
+  if (strcmp(_y4m->chroma_type, "420") == 0 ||
+      strcmp(_y4m->chroma_type, "420jpeg") == 0) {
+    _y4m->src_c_dec_h = _y4m->dst_c_dec_h = _y4m->src_c_dec_v =
+        _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz =
+        _y4m->pic_w * _y4m->pic_h +
+        2 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2);
+    /* Natively supported: no conversion required. */
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+  } else if (strcmp(_y4m->chroma_type, "420p10") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->dst_c_dec_h = 2;
+    _y4m->src_c_dec_v = 2;
+    _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz =
+        2 * (_y4m->pic_w * _y4m->pic_h +
+             2 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2));
+    /* Natively supported: no conversion required. */
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    _y4m->bit_depth = 10;
+    _y4m->bps = 15;
+    _y4m->aom_fmt = AOM_IMG_FMT_I42016;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 420p10 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "420p12") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->dst_c_dec_h = 2;
+    _y4m->src_c_dec_v = 2;
+    _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz =
+        2 * (_y4m->pic_w * _y4m->pic_h +
+             2 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2));
+    /* Natively supported: no conversion required. */
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    _y4m->bit_depth = 12;
+    _y4m->bps = 18;
+    _y4m->aom_fmt = AOM_IMG_FMT_I42016;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 420p12 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "420mpeg2") == 0) {
+    _y4m->src_c_dec_h = _y4m->dst_c_dec_h = _y4m->src_c_dec_v =
+        _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+    /*Chroma filter required: read into the aux buf first.*/
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz =
+        2 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2);
+    _y4m->convert = y4m_convert_42xmpeg2_42xjpeg;
+  } else if (strcmp(_y4m->chroma_type, "420paldv") == 0) {
+    _y4m->src_c_dec_h = _y4m->dst_c_dec_h = _y4m->src_c_dec_v =
+        _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+    /*Chroma filter required: read into the aux buf first.
+      We need to make two filter passes, so we need some extra space in the
+       aux buffer.*/
+    _y4m->aux_buf_sz = 3 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2);
+    _y4m->aux_buf_read_sz =
+        2 * ((_y4m->pic_w + 1) / 2) * ((_y4m->pic_h + 1) / 2);
+    _y4m->convert = y4m_convert_42xpaldv_42xjpeg;
+  } else if (strcmp(_y4m->chroma_type, "422jpeg") == 0) {
+    _y4m->src_c_dec_h = _y4m->dst_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+    /*Chroma filter required: read into the aux buf first.*/
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz =
+        2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+    _y4m->convert = y4m_convert_422jpeg_420jpeg;
+  } else if (strcmp(_y4m->chroma_type, "422") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    if (only_420) {
+      _y4m->dst_c_dec_h = 2;
+      _y4m->dst_c_dec_v = 2;
+      _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+      /*Chroma filter required: read into the aux buf first.
+        We need to make two filter passes, so we need some extra space in the
+         aux buffer.*/
+      _y4m->aux_buf_read_sz = 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+      _y4m->aux_buf_sz =
+          _y4m->aux_buf_read_sz + ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+      _y4m->convert = y4m_convert_422_420jpeg;
+    } else {
+      _y4m->aom_fmt = AOM_IMG_FMT_I422;
+      _y4m->bps = 16;
+      _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+      _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+      _y4m->dst_buf_read_sz =
+          _y4m->pic_w * _y4m->pic_h + 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+      /*Natively supported: no conversion required.*/
+      _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+      _y4m->convert = y4m_convert_null;
+    }
+  } else if (strcmp(_y4m->chroma_type, "422p10") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    _y4m->aom_fmt = AOM_IMG_FMT_I42216;
+    _y4m->bps = 20;
+    _y4m->bit_depth = 10;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * (_y4m->pic_w * _y4m->pic_h +
+                                 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h);
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 422p10 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "422p12") == 0) {
+    _y4m->src_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    _y4m->aom_fmt = AOM_IMG_FMT_I42216;
+    _y4m->bps = 24;
+    _y4m->bit_depth = 12;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * (_y4m->pic_w * _y4m->pic_h +
+                                 2 * ((_y4m->pic_w + 1) / 2) * _y4m->pic_h);
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 422p12 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "411") == 0) {
+    _y4m->src_c_dec_h = 4;
+    _y4m->dst_c_dec_h = 2;
+    _y4m->src_c_dec_v = 1;
+    _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+    /*Chroma filter required: read into the aux buf first.
+      We need to make two filter passes, so we need some extra space in the
+       aux buffer.*/
+    _y4m->aux_buf_read_sz = 2 * ((_y4m->pic_w + 3) / 4) * _y4m->pic_h;
+    _y4m->aux_buf_sz =
+        _y4m->aux_buf_read_sz + ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+    _y4m->convert = y4m_convert_411_420jpeg;
+  } else if (strcmp(_y4m->chroma_type, "444") == 0) {
+    _y4m->src_c_dec_h = 1;
+    _y4m->src_c_dec_v = 1;
+    if (only_420) {
+      _y4m->dst_c_dec_h = 2;
+      _y4m->dst_c_dec_v = 2;
+      _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+      /*Chroma filter required: read into the aux buf first.
+        We need to make two filter passes, so we need some extra space in the
+         aux buffer.*/
+      _y4m->aux_buf_read_sz = 2 * _y4m->pic_w * _y4m->pic_h;
+      _y4m->aux_buf_sz =
+          _y4m->aux_buf_read_sz + ((_y4m->pic_w + 1) / 2) * _y4m->pic_h;
+      _y4m->convert = y4m_convert_444_420jpeg;
+    } else {
+      _y4m->aom_fmt = AOM_IMG_FMT_I444;
+      _y4m->bps = 24;
+      _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+      _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+      _y4m->dst_buf_read_sz = 3 * _y4m->pic_w * _y4m->pic_h;
+      /*Natively supported: no conversion required.*/
+      _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+      _y4m->convert = y4m_convert_null;
+    }
+  } else if (strcmp(_y4m->chroma_type, "444p10") == 0) {
+    _y4m->src_c_dec_h = 1;
+    _y4m->src_c_dec_v = 1;
+    _y4m->aom_fmt = AOM_IMG_FMT_I44416;
+    _y4m->bps = 30;
+    _y4m->bit_depth = 10;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * 3 * _y4m->pic_w * _y4m->pic_h;
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 444p10 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "444p12") == 0) {
+    _y4m->src_c_dec_h = 1;
+    _y4m->src_c_dec_v = 1;
+    _y4m->aom_fmt = AOM_IMG_FMT_I44416;
+    _y4m->bps = 36;
+    _y4m->bit_depth = 12;
+    _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+    _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+    _y4m->dst_buf_read_sz = 2 * 3 * _y4m->pic_w * _y4m->pic_h;
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_null;
+    if (only_420) {
+      fprintf(stderr, "Unsupported conversion from 444p12 to 420jpeg\n");
+      return -1;
+    }
+  } else if (strcmp(_y4m->chroma_type, "444alpha") == 0) {
+    _y4m->src_c_dec_h = 1;
+    _y4m->src_c_dec_v = 1;
+    if (only_420) {
+      _y4m->dst_c_dec_h = 2;
+      _y4m->dst_c_dec_v = 2;
+      _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+      /*Chroma filter required: read into the aux buf first.
+        We need to make two filter passes, so we need some extra space in the
+         aux buffer.
+        The extra plane also gets read into the aux buf.
+        It will be discarded.*/
+      _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 3 * _y4m->pic_w * _y4m->pic_h;
+      _y4m->convert = y4m_convert_444_420jpeg;
+    } else {
+      _y4m->aom_fmt = AOM_IMG_FMT_444A;
+      _y4m->bps = 32;
+      _y4m->dst_c_dec_h = _y4m->src_c_dec_h;
+      _y4m->dst_c_dec_v = _y4m->src_c_dec_v;
+      _y4m->dst_buf_read_sz = 4 * _y4m->pic_w * _y4m->pic_h;
+      /*Natively supported: no conversion required.*/
+      _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+      _y4m->convert = y4m_convert_null;
+    }
+  } else if (strcmp(_y4m->chroma_type, "mono") == 0) {
+    _y4m->src_c_dec_h = _y4m->src_c_dec_v = 0;
+    _y4m->dst_c_dec_h = _y4m->dst_c_dec_v = 2;
+    _y4m->dst_buf_read_sz = _y4m->pic_w * _y4m->pic_h;
+    /*No extra space required, but we need to clear the chroma planes.*/
+    _y4m->aux_buf_sz = _y4m->aux_buf_read_sz = 0;
+    _y4m->convert = y4m_convert_mono_420jpeg;
+  } else {
+    fprintf(stderr, "Unknown chroma sampling type: %s\n", _y4m->chroma_type);
+    return -1;
+  }
+  /*The size of the final frame buffers is always computed from the
+     destination chroma decimation type.*/
+  _y4m->dst_buf_sz =
+      _y4m->pic_w * _y4m->pic_h +
+      2 * ((_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h) *
+          ((_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v);
+  if (_y4m->bit_depth == 8)
+    _y4m->dst_buf = (unsigned char *)malloc(_y4m->dst_buf_sz);
+  else
+    _y4m->dst_buf = (unsigned char *)malloc(2 * _y4m->dst_buf_sz);
+
+  if (_y4m->aux_buf_sz > 0)
+    _y4m->aux_buf = (unsigned char *)malloc(_y4m->aux_buf_sz);
+  return 0;
+}
+
+void y4m_input_close(y4m_input *_y4m) {
+  free(_y4m->dst_buf);
+  free(_y4m->aux_buf);
+}
+
+int y4m_input_fetch_frame(y4m_input *_y4m, FILE *_fin, aom_image_t *_img) {
+  char frame[6];
+  int pic_sz;
+  int c_w;
+  int c_h;
+  int c_sz;
+  int bytes_per_sample = _y4m->bit_depth > 8 ? 2 : 1;
+  /*Read and skip the frame header.*/
+  if (!file_read(frame, 6, _fin)) return 0;
+  if (memcmp(frame, "FRAME", 5)) {
+    fprintf(stderr, "Loss of framing in Y4M input data\n");
+    return -1;
+  }
+  if (frame[5] != '\n') {
+    char c;
+    int j;
+    for (j = 0; j < 79 && file_read(&c, 1, _fin) && c != '\n'; j++) {
+    }
+    if (j == 79) {
+      fprintf(stderr, "Error parsing Y4M frame header\n");
+      return -1;
+    }
+  }
+  /*Read the frame data that needs no conversion.*/
+  if (!file_read(_y4m->dst_buf, _y4m->dst_buf_read_sz, _fin)) {
+    fprintf(stderr, "Error reading Y4M frame data.\n");
+    return -1;
+  }
+  /*Read the frame data that does need conversion.*/
+  if (!file_read(_y4m->aux_buf, _y4m->aux_buf_read_sz, _fin)) {
+    fprintf(stderr, "Error reading Y4M frame data.\n");
+    return -1;
+  }
+  /*Now convert the just read frame.*/
+  (*_y4m->convert)(_y4m, _y4m->dst_buf, _y4m->aux_buf);
+  /*Fill in the frame buffer pointers.
+    We don't use aom_img_wrap() because it forces padding for odd picture
+     sizes, which would require a separate fread call for every row.*/
+  memset(_img, 0, sizeof(*_img));
+  /*Y4M has the planes in Y'CbCr order, which libaom calls Y, U, and V.*/
+  _img->fmt = _y4m->aom_fmt;
+  _img->w = _img->d_w = _y4m->pic_w;
+  _img->h = _img->d_h = _y4m->pic_h;
+  _img->x_chroma_shift = _y4m->dst_c_dec_h >> 1;
+  _img->y_chroma_shift = _y4m->dst_c_dec_v >> 1;
+  _img->bps = _y4m->bps;
+
+  /*Set up the buffer pointers.*/
+  pic_sz = _y4m->pic_w * _y4m->pic_h * bytes_per_sample;
+  c_w = (_y4m->pic_w + _y4m->dst_c_dec_h - 1) / _y4m->dst_c_dec_h;
+  c_w *= bytes_per_sample;
+  c_h = (_y4m->pic_h + _y4m->dst_c_dec_v - 1) / _y4m->dst_c_dec_v;
+  c_sz = c_w * c_h;
+  _img->stride[AOM_PLANE_Y] = _img->stride[AOM_PLANE_ALPHA] =
+      _y4m->pic_w * bytes_per_sample;
+  _img->stride[AOM_PLANE_U] = _img->stride[AOM_PLANE_V] = c_w;
+  _img->planes[AOM_PLANE_Y] = _y4m->dst_buf;
+  _img->planes[AOM_PLANE_U] = _y4m->dst_buf + pic_sz;
+  _img->planes[AOM_PLANE_V] = _y4m->dst_buf + pic_sz + c_sz;
+  _img->planes[AOM_PLANE_ALPHA] = _y4m->dst_buf + pic_sz + 2 * c_sz;
+  return 1;
+}
diff --git a/third_party/aom/y4minput.h b/third_party/aom/y4minput.h
new file mode 100644
index 0000000000..db20190db6
--- /dev/null
+++ b/third_party/aom/y4minput.h
@@ -0,0 +1,69 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ *
+ * Based on code from the OggTheora software codec source code,
+ * Copyright (C) 2002-2010 The Xiph.Org Foundation and contributors.
+ */
+
+#ifndef Y4MINPUT_H_
+#define Y4MINPUT_H_
+
+#include <stdio.h>
+#include "aom/aom_image.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct y4m_input y4m_input;
+
+/*The function used to perform chroma conversion.*/
+typedef void (*y4m_convert_func)(y4m_input *_y4m, unsigned char *_dst,
+                                 unsigned char *_src);
+
+struct y4m_input {
+  int pic_w;
+  int pic_h;
+  int fps_n;
+  int fps_d;
+  int par_n;
+  int par_d;
+  char interlace;
+  int src_c_dec_h;
+  int src_c_dec_v;
+  int dst_c_dec_h;
+  int dst_c_dec_v;
+  char chroma_type[16];
+  /*The size of each converted frame buffer.*/
+  size_t dst_buf_sz;
+  /*The amount to read directly into the converted frame buffer.*/
+  size_t dst_buf_read_sz;
+  /*The size of the auxilliary buffer.*/
+  size_t aux_buf_sz;
+  /*The amount to read into the auxilliary buffer.*/
+  size_t aux_buf_read_sz;
+  y4m_convert_func convert;
+  unsigned char *dst_buf;
+  unsigned char *aux_buf;
+  enum aom_img_fmt aom_fmt;
+  int bps;
+  unsigned int bit_depth;
+};
+
+int y4m_input_open(y4m_input *_y4m, FILE *_fin, char *_skip, int _nskip,
+                   int only_420);
+void y4m_input_close(y4m_input *_y4m);
+int y4m_input_fetch_frame(y4m_input *_y4m, FILE *_fin, aom_image_t *img);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // Y4MINPUT_H_